Datasets:

timaeus
/

pile-arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 100k rows

text stringlengths 4 2.78M	meta dict
--- author: - \| H. B. Benaoum\ Prince Mohammad Bin Fahd University, Al-Khobar 31952, Saudi Arabia\ Email: [email protected] title: 'Broken $S_3$ Neutrinos ' --- \ Motivated by recent measurements which strongly support a nonzero reactor mixing angle $\theta_{13}$, we study a deviation from $S_3$ neutrino discrete symmetry by explicitly breaking the neutrino mass matrix with a general retrocirculant matrix. We show that nonzero $\theta_{13}$ and nonzero CP violation parameter $J_{CP}$ arise due to the difference between $y_2$ and $y_3$. We demonstrate that it is possible to obtain the experimentally favored results for neutrino masses and mixing angles from this mass matrix. Furthermore, we estimate the effective masses $m_{\beta}$ and $m_{\beta \beta}$ and total neutrino mass $\sum \|m_i\|$ predicted by this mass matrix.\ \ [Keywords]{}: Neutrino Physics; Flavor Symmetry; Neutrino Mass and Mixing; Discrete Symmetries.\ [PACS numbers]{}: 14.60.Pq; 12.15.Ff 0.3cm ------------------------------------------------------------------------ 0.5cm Experiments using solar, atmospheric and reactor neutrinos have made considerable progress in establishing two different mass squared differences ( $\Delta m^2_{21}$ and $\Delta m^2_{31}$ ) and two large mixing angles ( $\theta_{12}$ and $\theta_{23}$ ) in the lepton sector. Recently MINOS [@minos1; @minos2], T2K [@t2k], Double CHOOZ [@chooz], Daya Bay [@daya] and RENO [@reno] have revealed that the reactor mixing angle $\theta_{13}$ is not only nonzero but relatively large.\ The phenomenon of neutrino mixing can be simply described by the Pontecorvo-Maki-Nakagawa-Sakata ( PMNS ) neutrino mixing matrix $V_{PMNS}$ [@maki], which links the neutrino flavor eigenstates $\nu_e, \nu_{\mu}, \nu_{\tau}$ to the mass eigenstates $\nu_1, \nu_2, \nu_2$ :\ $$\begin{aligned} V & = & \left( \begin{array}{ccc} V_{e 1} & V_{e 2} & V_{e 3} \\ V_{\mu 1} & V_{\mu 2} & V_{\mu 3} \\ V_{\tau 1} & V_{\tau 2} & V_{\tau 3} \\ \end{array} \right) ~~~~~.\end{aligned}$$ In the standard parametrization used by the Particle Data Group ( PDG ), the PMSN matrix is expressed by three mixing angles $\theta_{12}, \theta_{23}$ and $\theta_{13}$ and one intrinsic CP violating phase $\delta$ for Dirac neutrinos, $$\begin{aligned} V & = & { \left( \begin{array}{ccc} c_{12} c_{13} & s_{12} c_{13} & s_{13} e^{- i \delta} \\ - s_{12} c_{23} - c_{12} s_{23} s_{13} e^{i \delta} & c_{12} c_{23} - s_{12} s_{23} s_{13} e^{i \delta} & s_{23} c_{13} \\ s_{12} s_{23} - c_{12} c_{23} s_{13} e^{i \delta} & - c_{12} s_{23} - s_{12} c_{23} s_{13} e^{i \delta} & c_{23} c_{13} \\ \end{array} \right) } . P_{Maj}\end{aligned}$$ where $c_{ij} = \cos \theta_{ij}, s_{ij} = \sin \theta_{ij}$ and $P_{Maj}$ is a diagonal matrix with Majorana CP violating phases.\ The three mixing angles are related to the moduli of the elements of the PMNS mixing matrix as : $$\begin{aligned} \sin^2 \theta_{13} & = & \| V_{e 3} \|^2 \nonumber \\ \sin^2 \theta_{12} & = & \frac{\| V_{e 2} \|^2}{1 - \| V_{e 3} \|^2} \nonumber \\ \sin^2 \theta_{23} & = & \frac{\| V_{\mu 3} \|^2}{1 - \| V_{e 3} \|^2} ~~~~~~~.\end{aligned}$$ The well-known tri-bimaximal (TBM) mixing pattern, which corresponds to $\theta_{13} = 0$, $\theta_{23} = \pm \frac{\pi}{4}$ and $\theta_{12} = \sin^{-1} \left( \frac{1}{\sqrt{3}} \right)$, has attracted a degree of attention in the literature because it suggests some underlying flavor symmetry among lepton’s generations. This flavor symmetry is expected to explain the mass spectrum and neutrino mixing pattern.\ The TBM form is : $$\begin{aligned} V_0 & = & \left( \begin{array}{ccc} \sqrt{\frac{2}{3}} & \frac{1}{\sqrt{3}} & 0 \\ - \frac{1}{\sqrt{6}} & \frac{1}{\sqrt{3}} & - \frac{1}{\sqrt{2}} \\ -\frac{1}{\sqrt{6}} & \frac{1}{\sqrt{3}} & \frac{1}{\sqrt{2}} \end{array} \right) ~~~~.\end{aligned}$$ Models based on discrete symmetries were successful in reproducing this matrix. Among a number of interesting discrete flavor symmetries discussed in the literature, the $S_3$ symmetry which is the permutation group of three objects, is the simplest [@wolfenstein]. $S_3$ is the smallest non-Abelian discrete group.\ The three-dimensional reducible representations of all $S_3$ group elements are : $$\begin{aligned} S^{(1)} & = & \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array} \right) ~~~;~~~ S^{(12)} ~~=~~ \left( \begin{array}{ccc} 0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 1 \end{array} \right)~~~;~~~ S^{(13)} ~~= ~~\left( \begin{array}{ccc} 0 & 0 & 1 \\ 0 & 1 & 0 \\ 1 & 0 & 0 \end{array} \right) \nonumber \\ S^{(23)} & = & \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & 0 & 1 \\ 0 & 1 & 0 \end{array} \right)~~~;~~~ S^{(123)} ~~=~~ \left( \begin{array}{ccc} 0 & 0 & 1 \\ 1 & 0 & 0 \\ 0 & 1 & 0 \end{array} \right)~~~;~~~ S^{(132)} ~~= ~~\left( \begin{array}{ccc} 0 & 1 & 0 \\ 0 & 0 & 1 \\ 1 & 0 & 0 \end{array} \right) ~~~.\end{aligned}$$ The most general neutrino mass matrix $M_{\nu}^0$ invariant under $S_3$ is : $$\begin{aligned} M_{\nu}^0 & = & \alpha ~S^{(1)} + \beta \left( S^{(12)} + S^{(13)} + S^{(23)} \right)\end{aligned}$$ where $\alpha$ and $\beta$ are, in general, complex numbers.\ In the basis where the charged lepton mass matrix is diagonal, the TBM mixing matrix diagonalizes the neutrino matrix $M_{\nu}^0$. $$\begin{aligned} V_0^{T} M_{\nu}^0 V_0 & = & \left( \begin{array}{ccc} \alpha & 0 & 0 \\ 0 & \alpha + \beta & 0 \\ 0 & 0 & \alpha \end{array} \right) ~~~~~~.\end{aligned}$$ This matrix leads to two degenerate masses, namely $m_1$ and $m_3$. However, this is not correct experimentally. To overcome this problem, it was suggested in [@jora] that in fact the three masses are degenerate by letting the complex number $\alpha$ lies in the third quadrant, $$\begin{aligned} \alpha & = & - i \|\alpha\| e^{-i \frac{\psi}{2}} ~~~~ \mbox{for}~~0 \leq \psi < \pi\end{aligned}$$ and taking $\beta$ as real number such that : $$\begin{aligned} \beta & = & \frac{2}{3} \|\alpha\| \sin \frac{\psi}{2} ~~~~~~~.\end{aligned}$$ In this work, we consider the neutrino matrix $M_{\nu}^0$, which is invariant under $S_3$, as zeroth order with degenerate masses and TBM mixing angles. Nondegenerate mass spectrum and nonzero $\theta_{13}$ were realized in [@jora]-[@dev] by introducing small perturbations that violate $S_3$ symmetry.\ Here, we investigate the phenomenological consequences of the deviation from an exact $S_3$ symmetry by explicitly breaking the neutrino mass matrix $M_{\nu}^0$ with a general retrocirculant matrix : $$\begin{aligned} \Delta M_{\nu} & = & - \alpha~ \left( \begin{array}{ccc} y_1 & y_2 & y_3\\ y_2 & y_3 & y_1 \\ y_3 & y_1 & y_2 \end{array} \right)\end{aligned}$$ where the dimensionless parameters $y_i = \|y_i\| e^{i \varphi_i/2}$ are complex numbers with magnitude less than one and $0 \leq \varphi_i \leq \pi$. We also consider the charged lepton to be diagonal so the leptonic mixing solely comes from the neutrino sector. It is easy to see that $\Delta M_{\nu}$ can be written as a linear combination of $S^{(23)}$, $S^{(12)}$ and $S^{(13)}$ as : $$\begin{aligned} \Delta M_{\nu} & = & - \alpha \left( y_1 ~S^{(23)} + y_2 ~S^{(12)} + y_3 ~S^{(13)} \right) ~~~~~~.\end{aligned}$$ As a result the broken neutrino matrix $M_{\nu}$ becomes : $$\begin{aligned} M_{\nu} & = & M_{\nu}^0 + \Delta M_{\nu} \nonumber \\ & = & \left( \begin{array}{ccc} \alpha + \beta - \alpha~ y_1 & \beta - \alpha~ y_2 & \beta - \alpha~ y_3 \\ \beta - \alpha~ y_2 & \alpha + \beta - \alpha~ y_3 & \beta - \alpha~ y_1 \\ \beta - \alpha~ y_3 & \beta - \alpha~ y_1 & \alpha + \beta - \alpha~ y_2 \end{array} \right) ~~~~ .\end{aligned}$$ The eigenvalues of the above matrix are : $$\begin{aligned} m_1 & = & \alpha - \alpha~ \sqrt{y_1^2 + y_2^2 + y_3^2 - y_1 y_2 - y_1 y_3 - y_2 y_3} \nonumber \\ m_2 & = & \alpha + 3 \beta - \alpha ~\left( y_1 + y_2 + y_3 \right) \nonumber \\ m_3 & = & \alpha + \alpha~ \sqrt{y_1^2 + y_2^2 + y_3^2 - y_1 y_2 - y_1 y_3 - y_2 y_3}~~~~~~.\end{aligned}$$ The matrix $M_{\nu}$ is called magic mass matrix since every row and column add up to the same value which is $m_2$ for this case. It implies that this mass matrix has a trimaximal eigenvector $\left( \frac{1}{\sqrt{3}}, \frac{1}{\sqrt{3}}, \frac{1}{\sqrt{3}} \right)^T$ [@kumar]. Moreover the $\mu-\tau$ symmetry corresponding to $\theta_{13} = 0$ and $\theta_{23} = \pm \frac{\pi}{4}$ is broken for $M_{\nu}$, $$\begin{aligned} [M_{\nu}, S^{(23)}] & = & - \alpha~ (y_2 - y_3) ~\left( S^{(123)} - S^{(132)} \right) \end{aligned}$$ due to the difference between $y_2$ and $y_3$.\ Interestingly, by rotating the above matrix by the TBM mixing matrix $V_0$, we get : $$\begin{aligned} V_0^{T} M_{\nu} V_0 & = & \left( \begin{array}{ccc} \alpha - \frac{\alpha}{2} \left(2 y_1 - y_2 - y_3 \right) & 0 & \frac{\alpha \sqrt{3}}{2} \left(y_2 - y_3 \right) \\ 0 & \alpha + 3 \beta - \alpha \left( y_1 + y_2 + y_3 \right) & 0 \\ \frac{\alpha \sqrt{3}}{2} \left(y_2 - y_3 \right) & 0 & \alpha + \frac{\alpha}{2} \left(2 y_1 - y_2 - y_3 \right) \end{array} \right) ~~.\end{aligned}$$ As a consequence, the neutrino matrix $M_{\nu}$ is diagonalized by the total unitary matrix $V = V_0 U P_{Maj}$. The mixing matrix $U$ is given by : $$\begin{aligned} U & = & \left( \begin{array}{ccc} \cos \theta & 0 & e^{- i \delta} \sin \theta \\ 0 & 1 & 0 \\ - e^{i \delta} \sin \theta & 0 & \cos \theta \end{array} \right) ~~~~.\end{aligned}$$ A straightforward calculation yields that the angle $\theta$ and the CP-phase $\delta$ are : $$\begin{aligned} \tan 2 \theta & = & \frac{\sqrt{X^2 + Y^2}}{Z} \nonumber \\ \tan \delta & = & \frac{Y}{X}\end{aligned}$$ where $$\begin{aligned} X & = & \sqrt{3} \left( \|y_2\| \cos \frac{\varphi_2}{2} - \|y_3\| \cos \frac{\varphi_3}{2} \right) \nonumber \\ Y & = & \sqrt{3} \left( \|y_1\| \|y_2\| \sin (\frac{\varphi_1 - \varphi_2}{2}) - \|y_1\| \|y_3\| \sin (\frac{\varphi_1 - \varphi_3}{2}) + \|y_2\| \|y_3\| \sin (\frac{\varphi_2 - \varphi_3}{2}) \right) \nonumber \\ Z & = & 2 \|y_1\| \cos \frac{\varphi_1}{2} - \|y_2\| \cos \frac{\varphi_2}{2}-\|y_3\| \cos \frac{\varphi_3}{2} ~~~~.\end{aligned}$$ The explicit expression of $V$ is : $$\begin{aligned} V & = & \left( \begin{array}{ccc} \sqrt{\frac{2}{3}}~\cos \theta & \frac{1}{\sqrt{3}} & \sqrt{\frac{2}{3}}~e^{-i \delta} ~\sin \theta \\ - \frac{\cos \theta}{\sqrt{6}} + \frac{e^{i \delta} ~\sin \theta}{\sqrt{2}} & \frac{1}{\sqrt{3}} & - \frac{\cos \theta}{\sqrt{2}} - \frac{e^{-i \delta} ~\sin \theta}{\sqrt{6}} \\ - \frac{\cos \theta}{\sqrt{6}} - \frac{e^{i \delta} ~\sin \theta}{\sqrt{2}} & \frac{1}{\sqrt{3}} & \frac{\cos \theta}{\sqrt{2}} - \frac{e^{i \delta} ~\sin \theta}{\sqrt{6}} \end{array} \right) ~.P_{Maj}~~~.\end{aligned}$$ We immediately obtain the mixing angles : $$\begin{aligned} \sin^2 \theta_{13} & = & \frac{2}{3}~ \sin^2 \theta \nonumber \\ \sin^2 \theta_{12} & = & \frac{1}{2 + \cos 2 \theta} \nonumber \\ \sin^2 \theta_{23} & = & \frac{1}{2} \left( 1 + \frac{\sqrt{3} \sin 2 \theta \cos \delta}{2 + \cos 2 \theta} \right) \end{aligned}$$ which for $\|y_3\| = \|y_2\| = 0$ ( i.e. $\theta_{13}=0$) give the TBM mixing angles.\ The first and second equation in (20) show that the solar and reactor neutrino mixing angles are related by : $$\begin{aligned} \sin^2 \theta_{12} & = & \frac{1}{3 \cos^2 \theta_{13}} ~~~~.\end{aligned}$$ Next by considering the third equation in (20), a simple relation between the CP-phase $\delta$ and the mixing angles can be derived. The result for $\cos \delta$ reads : $$\begin{aligned} \cos \delta & = & - \frac{1}{\sqrt{3}} ~\frac{\cos 2 \theta_{23}}{\cos \theta_{12} \sqrt{3 \sin^2 \theta_{12} -1}} ~~~~. \end{aligned}$$ This, in turn, would imply that the solar mixing angle $\theta_{12}$ has to be, $$\begin{aligned} \sin^2 \theta_{12} & > & \frac{1}{3} \end{aligned}$$ which is right on the edge from the global fits to the neutrino oscillation data. Such a constraint could be confirmed or ruled out with a little better data.\ The strength of CP violation in neutrino oscillations is described by the Jarlskog rephasing invariant parameter. It is given by : $$\begin{aligned} J_{CP} & = & Im \left( V_{e 2} V_{\mu 3} V_{e 3}^{\star} V_{\mu 2}^{\star} \right) ~=~ - \frac{1}{6 \sqrt{3}} ~\sin 2 \theta \sin \delta \nonumber \\ & = & - \frac{1}{6 \sqrt{3}} ~\frac{Y}{\sqrt{X^2 + Y^2 + Z^2}} ~~~~~.\end{aligned}$$ Similarly, the Jarlskog parameter $J_{CP}$ can be written in terms of the solar $\theta_{12}$ and the atmospheric $\theta_{23}$ mixing angles.\ It proves convenient to use the ratio $R_{\nu} = \frac{\Delta m^2_{21}}{\Delta m^2_{31}}$, to write the mass squared differences as : $$\begin{aligned} \Delta m^2_{21} & = & 2 \|\alpha\|^2 R_{\nu} \sqrt{X^2 + Y^2 + Z^2} \nonumber \\ \Delta m^2_{31} & = & 2 \|\alpha\|^2 \sqrt{X^2 + Y^2 + Z^2} \nonumber \\ \Delta m^2_{32} & = & 2 \|\alpha\|^2 (1 - R_{\nu}) ~\sqrt{X^2 + Y^2 + Z^2} ~~~~~.\end{aligned}$$ In order to confront the above neutrino matrix with the experimental observations (Table 1), we use the constraints on neutrino parameters at $2 \sigma$ and $3 \sigma$ [@valle]. Parameter Best fit $2 \sigma$ $3 \sigma$ -------------------------------------- ------------- ----------------- ----------------- $\Delta m_{21}^2 [ 10^{-5} eV^2 ]$ $7.62$ $7.27 - 8.01$ $7.12 - 8.20$ $\|\Delta m_{31}^2\| [ 10^{-3} eV^2 ]$ $2.55$ $2.38 - 2.68$ $2.31 - 2.74$ $2.43$ $2.29 - 2.58$ $2.21 - 2.64$ $\sin^2 \theta_{12}$ $0.320$ $0.29 - 0.35$ $0.27 - 0.37$ $\sin^2 \theta_{23}$ $0.613$ $0.38 - 0.66$ $0.36 - 0.68$ $0.600$ $0.39 - 0.65$ $0.37 - 0.67$ $\sin^2 \theta_{13}$ $0.0246$ $0.019 - 0.030$ $0.017 - 0.033$ $0.0250$ $0.020 - 0.030$ $0.017 - 0.033$ $\delta$ $0.80 \pi$ $0 - 2 \pi$ $0 - 2 \pi$ $-0.03 \pi$ $0 - 2 \pi$ $0 - 2 \pi$ : Global oscillation analysis with best fit for $\Delta m_{21}^2,\Delta m_{31}^2, \sin^2 \theta_{12}, \sin^2 \theta_{23}, \sin^2 \theta_{13}$ and $\delta$ the upper and/or lower corresponds to normal and/or inverted neutrino mass hierarchy. []{data-label="table1"} \ For numerical analysis, we use the $3\sigma$ ranges of neutrino oscillation parameters. The neutrino mass matrix $M_{\nu}$ depends on $\|\alpha\|, \|y_1\|, \|y_2\|, \|y_3\|$ and the phases $\psi, \varphi_1, \varphi_2$ and $\varphi_3$.\ Since there are many unknown parameters, we consider a particular set of those parameters and show how the measured values of neutrino experiments can be accommodated in our neutrino mass matrix $M_{\nu}$. For simplicity, we take $\varphi_2 = \varphi_3 = 0$, as inputs.\ To see how the nonzero value of the phase $\varphi_1$ can lift the degeneracy between $m_1$ and $m_2$, we notice that for $y_3 = y_2 = 0$, the mass squared differences become : $$\begin{aligned} \Delta m^2_{21} & = & 4 \|\alpha\|^2 \|y_1\| \sin \frac{\psi}{2} \sin \frac{\psi - \varphi_1}{2} \nonumber \\ \Delta m^2_{31} & = & 4 \|\alpha\|^2 \|y_1\| \cos \frac{\varphi_1}{2} \nonumber \\ \Delta m^2_{32} & = & 4 \|\alpha\|^2 \|y_1\| \cos \frac{\psi}{2} \cos \frac{\psi-\varphi_1}{2} ~~~~~~.\end{aligned}$$ It is worthwhile to remark that for $\varphi_1 = \psi$, the masses $m_1$ and $m_2$ are degenerate and $\|m_3\| > \|m_2\|$. Now to separately obtain a nonzero mass squared difference $\Delta m^2_{21}$ which is smaller than $\Delta m^2_{31}$, we introduce a small phase difference $\epsilon_1 = \psi - \varphi_1$ between the phases $\psi$ and $\varphi_1$. Such a small phase difference will be responsible for lifting the mass degeneracy between the first and second generation.\ To see the behavior of the mass eigenvalues with respect to $\psi$ and other observables, we take $\epsilon_1 = 6^{\circ}$ as a typical value with $\|\alpha\| =0.1~eV$, $\|y_1\| = 5 \times 10^{-2}$, $\|y_2\| =1.3 \times 10^{-3}$ and $\|y_3\| =1.2 \times 10^{-3}$. Figure 1 shows the variation of the masses $\|m_i\|$ as a function of the phase $\psi$. One clearly observes that Fig. 1 suggests a normal hierarchical ordering pattern for $\psi > 70^{\circ}$.\ Based on the expression of the neutrino mass squared differences, we numerically scan over a broader range of $\|y_1\|$ $( 0.05 \leq \|y_1\| \leq 0.4 )$ to obtain the restriction on the parameter space of $\|y_2\|$ and $\|y_3\|$. We have plotted in Fig. 2 the contour plots of the mass squared differences in the two-dimensional parameter spaces $\left( \|y_2\|, \|y_3\| \right)$ where we take $\psi = 120^{\circ}, \epsilon_1 =6^{\circ}$ and $\|\alpha\| = 0.1~eV$ as inputs.\ From Fig. 2, we obtain the allowed ranges of $\|y_2\|$ and $\|y_3\|$ for the normal mass hierarchy, $$\begin{aligned} \|y_i\| \leq 0.02~~~~~~~~~\mbox{for}~i=2,3 ~~~~~~~~.\end{aligned}$$ The ratio $R_{\nu}$ for normal mass hierarchy ordering in the $3 \sigma$ allowed range is : $$\begin{aligned} R_{\nu} & = & \left(2.99^{+0.32}_{-0.34} \right) \times 10^{-2} ~~~~~~.\end{aligned}$$ Figure 3 shows contour plot of the ratio $R_{\nu}$ in the two parameter space $\left( \psi, \epsilon_1 = \psi -\varphi_1 \right)$ where $\|y_1\| =0.05, \|y_2\|=1.3 \times 10^{-3}$ and $\|y_3\| =1.2 \times 10^{-3}$. As expected, the ratio $R_{\nu}$ at $3 \sigma$ indicates that $\psi$ has to be greater than $70^{\circ}$.\ The departure of the mixing angles from TBM mixing angles depend on the phase $\epsilon_1$ and the two parameters $\|y_2\|$ and $\|y_3\|$. Scanning over $\|y_2\|$ and $\|y_3\|$ within their allowed ranges ( $\|y2\|, \|y_3\| \leq 0.02$ ), we investigate how a nonzero $\theta_{13}$ can be obtained for normal mass hierarchy. As a result of numerical analysis, contour plots in the $\left( \psi, \|y_1\| \right)$ parameter plane of the mixing angles $\theta_{13}$, $\theta_{12}$ and $\theta_{23}$ are shown in Fig. 4 using the experimental constraints on the measured angles.\ The CP violation parameter $J_{CP}$, which is directly related to the Dirac phase $\delta$, arises due to nonzero value of the difference $\|y_3\| - \|y_2\|$ in the neutrino matrix $M_{\nu}$. We have plotted in Fig. 5, $J_{CP}$ with respect to $\psi$ using the allowed region of $\|y_i$. It leads to values of $\|J_{CP}\|$ around $10^{-3}$. Nonvanishing $\|J_{CP}\|$ will be explored by the next generation high performance long-baseline neutrino experiments.\ The absolute neutrino mass scale can be probed by nonoscillatory neutrino experiments. Cosmology is sensitive to the sum of neutrino masses $\sum \|m_i\|$. The beta decay endpoint measurements probe the so-called effective electron neutrino mass $m_{\beta}$. The rate of the neutrinoless double beta decay depends on the effective Majorana mass of the electron neutrino $m_{\beta \beta}$.\ Both $m_{\beta \beta}$ and $m_{\beta}$ and the sum of neutrino masses are given by : $$\begin{aligned} m_{\beta \beta} & = & \|\sum_i m_i V_{e i}^2\| \nonumber \\ m_{\beta} & = & \sqrt{\sum_i m_i^2 \|V_{e i}\|^2} \nonumber \\ \sum \|m_i\| & = & \|m_1\| + \|m_2\| + \|m_3\| ~~~~~~~~~.\end{aligned}$$ From the neutrino mass $M_{\nu}$, the effective Majorana masses $m_{\beta \beta}$ and $m_{\beta}$ can be written as : $$\begin{aligned} m_{\beta \beta} & = & \frac{\|\alpha\|}{3}~\sqrt{5 + 4 \cos \psi - 6 \|y_1\| \left( \sin \frac{\psi -\varphi_1}{2} + 2 \sin \frac{\varphi_1}{2} \right) + 9 \|y_1\|^2} \nonumber \\ m_{\beta} & = & \|\alpha\|~\sqrt{1 - \frac{2}{3} y_1 \left( 2 \cos \frac{\varphi_1}{2} + \cos( \psi - \frac{\varphi_1}{2}) \right) + \frac{4}{3} (y_2+y_3) \cos^2 \frac{\psi}{2} + y_1^2 + y_2^2 +y_3^2} ~~.\end{aligned}$$ Present cosmological constraints on the sum of neutrino masses $\sum \|m_i\|$ are in the range $0.44-0.76~eV$ [@raffelt]. The Mainz [@kraus] and Troitsk [@troitsk] experiments on the high precision measurement of the end-point part of the $\beta$-spectrum of $^{3}H$ decay found the $95\%$ C.L. upper bounds $m_{\beta} \leq 2.3~eV$ ( Mainz ) and $m_{\beta} \leq 2.1~eV$ ( Troitsk). Experimental bound on $m_{\beta \beta}$ is below $0.36~eV$ [@gomez].\ Figure 6 gives the effective electron neutrino mass $m_{\beta}$, the effective Majorana mass $m_{\beta \beta}$ and the sum of neutrino masses $\sum \|m_i\|$ with respect to $\psi$ for $\epsilon_1 = 6^{\circ}$. It shows the predicted $m_{\beta}$, $m_{\beta \beta}$ and $\sum \|m_i\|$ are well below the experimental bounds. The magnitude of $m_{\beta \beta}$ increases with larger $\|y_1\|$ values. Acknowledgments {#acknowledgments .unnumbered} =============== I would like to thank S. Nasri and J. Schechter for reading the manuscript and useful comments. [99]{} L. Whitehead \[MINOS Collaboration\], Recent results from MINOS, Joint Experimental- Theoretical Seminar (24 June 2011, Fermilab, USA). Websites: theory.fnal.gov/jetp, http://www-numi.fnal.gov/pr${\_}$plots/ . P. Adamson et al. \[MINOS Collaboration\], Improved search for muon-neutrino to electron- neutrino oscillations in MINOS, \[arXiv:1108.0015 \[hep-ex\]\]. The T2K Collaboration, K. Abe et al.,Phys. Rev. Lett. [107]{} (2011) 041801 \[arXiv:1106.2822\]. H. De. Kerrect, Low Nu 2011, Seoul, South Korea, http://workshop.kias.re.kr/lownu11/ . F. P. An et. al, The Daya Bay Collaboration, arXiv:1203.1669 \[hep-ex\]. J. K. Ahn et. al, RENO Collaboration, arXiv:1204.0626 \[hep-ex\]. B. Pontecorvo, J. Exptl. Theoret. Phys. [33]{} (1957) 549 \[Sov. Phys. JETP [6]{} (1958) 429\]; J. Exptl. Theoret. Phys. [34]{} (1958) 247 \[Sov. Phys. JETP [7]{} (1958) 172; Z. Maki, M. Nakagawa and S. Sakata, Prog. Theor. Phys. [28]{} (1962) 870 . L. Wolfenstein, Phys. Rev. [D 18]{} (1978) 958; S. Pakvasa and H. Sugawara, Phys. Lett. [B 73]{} (1978) 61; 82, 105 (1979); E. Derman and H.S.Tsao, Phys. Rev. [D 20]{} (1979) 1207; S.-L. Chen, M. Frigerio and E. Ma, hep-ph/0404084; M. Fukugita, M. Tanimoto and T. Yanagida, Phys. Rev. [D 57]{} (1998) 4429; E. Ma and G. Rajasekaran, Phys. Rev. [D 64]{} (2001) 113012; Z-z. Xing, D. Yang and S. Zhou, arXiv:1004.4234v2\[hep-ph\]; D. A. Dicus, S-F. Ge and W. W. Repko, arXiv:1004.3266\[hep-ph\]; H. Fritzsch and Z.-Z.Xing, Phys. Lett. [B 440]{} (1988) 313; P. F. Harrison, D. H. Perkins and W. G. Scott, Phys. Lett. [B530]{} (2002) 79; Z.-Z.Xing, Phys. Lett. [B 533]{} (2002) 85; X.G.He and A. Zee, Phys. Lett. [B 560]{} (2003) 87; P. F. Harrison and W. G. Scott, hep-ph/0302025; C.I.Low and R.R.Volkas, Phys. Rev. [D 68]{} (2003) 033007; A.Zee, Phys. Rev. [D 68]{} (2003) 093002; J.D. Bjorken, P. F. Harrison and W. G. Scott, hep-ph/0511201; R. N. Mohapatra, S. Nasri and H. B. Yu, arXiv:hep-ph/0605020; S. F. King, Nucl. Phys. [B 576]{} (2000) 85; S. F. King and N. N. Singh, Nucl. Phys. [B 591]{} (2000) 3; E. Ma, Phys. Rev. [D 70]{} (2004) 091301; M. Hirsch, A. Velanova del Morel, J.W.F. Valle and E. Ma, Phys. Rev. [D 72]{} (2005) 031901; A. Montdragon, M. Montdragon and E. Peinado, J.Phys. [A41]{} (2008) 304035; T. Fukuyama and H. Nishiura, hep-ph/9702253; R. N. Mohapatra and S. Nussinov, Phys. Rev. [D 60]{} (1999) 013002; E. Ma and M. Raidal, Phys. Rev. Lett. [87]{} (2001) 011802; C. S. Lam, hep-ph/0104116; T. Kitabayashi and M. Yasue, Phys.Rev. [D67]{} (2003) 015006; W. Grimus and L. Lavoura, hep-ph/0305046; Y. Koide, Phys.Rev. [D69]{} (2004) 093001; Y. H. Ahn, Sin Kyu Kang, C. S. Kim, Jake Lee, hep-ph/0602160; A. Ghosal, hep-ph/0304090; W. Grimus and L. Lavoura, Phys. Lett. [B 572]{} (2003) 189; W. Grimus and L. Lavoura, J. Phys. [G 30]{} (2004) 73; W. Grimus, A. S.Joshipura, S. Kaneko, L. Lavoura, H. Sawanaka, M. Tanimoto, hep-ph/0408123; R. N. Mohapatra, JHEP, [0410]{} (2004) 027; A. de Gouvea, Phys.Rev. [D69]{} (2004) 093007; R. N. Mohapatra and W. Rodejohann, Phys. Rev. [D72]{} (2005) 053001; T. Kitabayashi and M. Yasue, Phys. Lett,. [B 621]{} (2005) 133; R. N. Mohapatra and S. Nasri, Phys. Rev. [D 71]{} (2005) 033001;R. N. Mohapatra, S. Nasri and H. B. Yu, Phys. Lett. [B 615]{} (2005) 231; K. Matsuda and H. Nishiura, Phys. Rev. [D 73]{} (2006) 013008; A. Joshipura, hep-ph/0512252; R. N. Mohapatra, S. Nasri and H. B. Yu, Phys. Lett. [B 636]{} (2006) 114; T. Kaneko, H. Sugawara, Phys. Lett. [B 697]{} (2011) 329. R. Jora, S. Nasri and J. Schechter, Int. J. Mod. Phys. [A 21]{} (2006) 5875; R. Jora, J. Schechter and M. Naeem Shahid, Phys. Rev. [D 80]{} (2009) 093007, \[Erratum-ibid. 82 (2010) 079902 \]; R. Jora, J. Schechter and M. Naeem Shahid, Phys. Rev. [D 82]{} (2010) 053006; R. Jora, J. Schechter and M. Naeem Shahid, arXiv:1210.6755 \[hep-ph\] . S. Dev, S. Gupta, R. R. Gautam, Phys. Lett. [B 702]{} (2011) 28; S. Dev, S. Gupta, R. R. Gautam, Phys. Lett.[B 708]{} (2012) 284. P. F. Harrison and W. G. Scott, Phys. Lett. [B 557]{} (2003) 76; C. S. Lam, Phys. Lett. [B 640]{} (2006) 260; W. Grimus, L. Lavoura, JHEP [0809]{} (2008) 106; C. H. Albright, W. Rodejohann Eur. Phys. J. [C 62]{} (2009) 599; S. Kumar, Phys. Rev. [D 82]{} (2010) 013010. D. V. Forero, M. Tortola and J. W. F. Valle, arXiv:1205.4018 . S. Hannestad, A. Mirizzi, G.G. Raffelt and Y.Wong, JCAP [08]{} (2010) 001; M.C. Gonzalez-Garcia, M. Maltoni and J. Salvado, J. High. Ener. Phys. [1008]{} (2010) 117. C. Kraus et al., Eur. Phys. J. C40 (2005) 447. Troitsk, V. Aseev et al., Phys. Rev. D84 (2011) 112003. J. J. Gomez-Cadenas, J. Martin-Albo, M. Mezzetto, F. Monrabal and M. Sorel, Riv. Nuovo Cim. 35, 29 (2012) .	{ "pile_set_name": "ArXiv" }
--- abstract: \| Given a finite set in a metric space, the topological analysis generalizes hierarchical clustering using a 1-parameter family of homology groups to quantify connectivity in all dimensions. The connectivity is compactly described by the persistence diagram. One limitation of the current framework is the reliance on metric distances, whereas in many practical applications objects are compared by non-metric dissimilarity measures. Examples are the Kullback–Leibler divergence, which is commonly used for comparing text and images, and the Itakura–Saito divergence, popular for speech and sound. These are two members of the broad family of dissimilarities called Bregman divergences. We show that the framework of topological data analysis can be extended to general Bregman divergences, widening the scope of possible applications. In particular, we prove that appropriately generalized Čech and Delaunay (alpha) complexes capture the correct homotopy type, namely that of the corresponding union of Bregman balls. Consequently, their filtrations give the correct persistence diagram, namely the one generated by the uniformly growing Bregman balls. Moreover, we show that unlike the metric setting, the filtration of Vietoris-Rips complexes may fail to approximate the persistence diagram. We propose algorithms to compute the thus generalized Čech, Vietoris-Rips and Delaunay complexes and experimentally test their efficiency. Lastly, we explain their surprisingly good performance by making a connection with discrete Morse theory. author: - Herbert Edelsbrunner - Hubert Wagner title: 'Topological Data Analysis with Bregman Divergences[^1]' --- Introduction {#sec1} ============ The starting point for the work reported in this paper is the desire to extend the basic topological data analysis (TDA) paradigm to data measured with dissimilarities. In particular for high-dimensional data, such as discrete probability distributions, notions of dissimilarity inspired by information theory behave strikingly different from the Euclidean distance, which is the usual setting for TDA. On the practical side, the Euclidean distance is particularly ill-suited for many types of high-dimensional data; see for example [@Hua08], which provides evidence that the Euclidean distance consistently performs the worst among several dissimilarity measures across a range of text-retrieval tasks. A broad class of dissimilarities are the Bregman divergences [@Bre67]. Its most prominent members are the Kullback–Leibler divergence [@KuLe51], which is commonly used both for text documents [@Big03; @Hua08] and for images [@DoVe02], and the Itakura–Saito divergence [@ItSa68], which is popular for speech and sound data [@FBD09]. We propose a TDA framework in the setting of Bregman divergences. Since TDA and more generally computational topology are young and emerging fields, we provide some context for the reader. For more a comprehensive introduction, see the recent textbook [@EdHa10]. #### Computational topology. Computational topology is an algorithmic approach to describing shape in a coarser sense than computational geometry does. TDA utilizes such algorithms within data analysis. One usually works with a finite set of points, possibly embedded in a high-dimensional space. Such data may be viewed as a collection of balls of a radius that depends on the scale of interest. Intersections reveal the connectivity of the data. For example, the components of the intersection graph correspond to the components of the union of balls. ##### Homology groups. These are studied in the area of algebraic topology, where they are used to describe and analyze topological spaces; see e.g. [@Hat02]. The connected components of a space or, dually, the gaps between them are encoded in its zero-dimensional homology group. There is a group for each dimension. For example, the one-dimensional group encodes loops or, dually, the tunnels, and the two-dimensional group encodes closed shells or, dually, the voids. Importantly, homology provides a formalism to talk about different kinds of connectivity and holes of a space that allows for fast algorithms. ##### Nerves and simplicial complexes. The nerve of a collection of balls generalizes the intersection graph and contains a $k$-dimensional simplex for every $k+1$ balls that have a non-empty common intersection. It is a hypergraph that is closed under taking subsets, a structure known as a simplicial complex in topology. If the balls are convex, then the Nerve Theorem states that this combinatorial construction captures the topology of the union of balls. More precisely, the nerve and the union have the same homotopy type and therefore isomorphic homology groups [@Bor48; @Ler45]. This result generalizes to the case in which the balls are not necessarily convex but their common intersections of all orders are contractible. In the context in which we center a ball of some radius at each point of a given set, the nerve is referred to as the Čech complex of the points for the given radius. Its $k$-skeleton is obtained by discarding simplices of dimension greater than $k$. The practice-oriented reader will spot a flaw in this setup: fixing the radius is a serious drawback that limits data analysis applications. ##### Persistent homology. To remedy this deficiency, we study the evolution of the topology across all scales, thus developing what we refer to as persistent homology. For graphs and connected components, this idea is natural but more difficult to flesh out in full generality. In essence, one varies the radius of balls from $0$ to $\infty$, giving rise to a nested sequence of spaces, called a filtration. Topological features, namely homology classes of different dimensions, are created and destroyed along the way. In practice, one computes the persistence diagram of a filtration, which discriminates topological features based on their lifetime, or persistence. The persistence diagram serves as a compact topological descriptor of a dataset, which is provably robust against noise. Owing to its algebraic and topological foundations, the theory is very general. Importantly, the Nerve Theorem extends to filtrations [@ChOu08 Lemma 3.4], so we can often restrict our considerations to complexes for a fixed radius. Moreover, the existing algorithms for persistence diagrams can be used without modification. ##### TDA in the Bregman setting. In the light of the above, there are only two obstacles to applying topological data analysis to data measured with Bregman divergences. We need to prove that the Nerve Theorem applies also when the balls are induced by Bregman divergences, and we need to provide efficient algorithms to construct the relevant complexes. The main complication is that the balls may be nonconvex, which we overcome by combining results from convex analysis and topology. ##### Applications. Persistence is an important method within TDA, which has been successfully used in a variety of applications. In low dimensions, it was for example used to shed light on the distribution of matter in the Universe [@Sou11] and to characterize the structure of atomic configurations in silica glass [@NHHEN15]. As for high-dimensional data, Chan et al. analyze viral DNA and relate persistent cycles with recombinations [@CCR13], and Port et al. study languages leaving the interpretation of a persistent cycle in the Indo-Germanic family open [@PGGC15]. #### Related work. This paper is the first work at the intersection of topology and Bregman divergences. We list related papers in relevant fields. In machine learning, Banerjee et al. use the family of Bregman divergences as the unifying framework for clustering algorithms [@BMDG05]. The field of information geometry deals with selected Bregman divergences and related concepts from a geometric perspective [@AmNa00]. Building on the classical work of Rockafellar [@Roc70] in convex analysis, Bauschke and Borwein are the first to use the Legendre transform for analyzing Bregman divergences [@BaBo97]. Boissonnat, Nielsen and Nock [@BNN10] use similar methods to make significant contributions at the intersection of computational geometry and Bregman divergences. In particular, they study the geometry of Bregman balls and Delaunay triangulations, but not the topologically more interesting Delaunay, or alpha, complexes. In the Euclidean setting, the basic constructions are well understood [@BaEd16; @Zom10], including approximations, which are interesting and useful, but beyond the scope of this paper. #### Results. This paper provides the first general TDA framework that applies to high-dimensional data measured with non-metric dissimilarities. Indeed, prior high-dimensional applications of TDA were restricted to low-dimensional homology, required custom-made topological results, or used common metrics such as the Euclidean and the Hamming distances, which are often not good choices for such data. We list the main technical contributions: 1. \[res:contr\] We show that the balls under any Bregman divergence have common intersections that are either empty of contractible. 2. \[res:gap\] We show that the persistence diagram of the Vietoris–Rips complex can be arbitrarily far from that of the filtration of the union on Bregman balls. 3. \[res:morse\] We show that the radius functions that correspond to the Čech and Delaunay complexes for Bregman divergences are generalized discrete Morse functions. 4. \[res:algs\] We develop algorithms for computing Čech and Delaunay radius functions for Bregman divergences, which owe their speed to non-trivial structural properties implied by Result \[res:morse\]. Most fundamental of the four is Result \[res:contr\], which forms the theoretical foundation of TDA in the Bregman setting. It implies that the Čech and Delaunay complexes for a given radius have the same homotopy type as the union of Bregman balls. Combined with the Nerve Theorem for filtrations, it also implies that the filtration of Čech and Delaunay complexes have the same persistence diagram as the filtration of the unions. In the practice of TDA, the filtration of Vietoris–Rips complexes is often substituted for the filtration of Čech or Delaunay complexes. For metrics, this is justified by the small bottleneck distance between the persistence diagrams if drawn in log-log scale. Result \[res:gap\] shows that such a substitution is not generally justified for Bregman divergences. In other words, for some Bregman divergences higher order interactions have to be taken into account explicitly as they are not approximated by implications of pairwise interactions. To appreciate Results \[res:morse\] and \[res:algs\], we note that the Čech radius function maps every simplex to the smallest radius, $r$, such that the simplex belongs to the Čech complex for radius $r$, and similarly for Delaunay. Being a generalized discrete Morse function has important structural consequences that make it possible to construct Čech and Delaunay complexes in an output-sensitive manner. We support this claim with experiments. ##### Implications. Our results open up a new area of research at the intersection of geometry, topology, algorithms and data analysis. On the application side, it enables TDA for a wide variety of data. Moreover, it connects topology with information theory and statistics, where Bregman divergences play a significant role. Finally, efficient algorithms and data structures are needed to handle large datasets. Considerable progress has been made within the TDA community, but we believe a collaboration with the wider computer science community would be fruitful. ##### Scope. The aim of this paper is to show that the machinery of persistent homology is applicable to different kinds of high-dimensional data. While questions remain, we provide a solid foundation for further developments. #### Outline. Section \[sec2\] reviews the concept of Bregman divergences, including an elementary description of the Legendre transform. Section \[sec3\] proves the contractibility of common intersections of Bregman balls and introduces Čech, Delaunay, and Vietoris–Rips complexes in the Bregman setting. Section \[sec4\] introduces the Čech and Delaunay radius functions and explains algorithms for constructing them. Section \[sec5\] concludes this paper. Bregman Divergences {#sec2} =================== Bregman divergences are sometimes called distances because they measure dissimilarity. As we will see shortly, they are generally not symmetric, and they always violate the triangle inequality. So really they satisfy only the first axiom of a metric, mapping ordered pairs to non-negative numbers and to zero iff the two elements are equal. We begin with a formal introduction of the concept, which originated in the paper by Bregman [@Bre67]. Their basic properties are well known; see the recent paper by Boissonnat, Nielsen and Nock [@BNN10]. We stress that our setting is slightly different: following Bauschke and Borwein [@BaBo97], we define the divergences in terms of functions of Legendre type. The crucial benefit of this additional requirement is that the conjugate of a function of Legendre type is again a function of Legendre type, even if the domain is bounded as in the important case of the standard simplex. In contract, the conjugate of a differentiable and strictly convex function that is not of Legendre type is not necessarily again a convex function. #### Functions of Legendre type. Let ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\subseteq {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$ be an nonempty open convex set and ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ a strictly convex differentiable function. In addition, we require that the length of the gradient of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ goes to infinity whenever we approach the boundary of ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$. Following [@Roc70 page 259], we say that ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ is a function of Legendre type. As suggested by the naming convention, these conditions are crucial when we apply the Legendre transform to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. The last condition prevents us from arbitrarily restricting the domain and is vacuous whenever ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ does not have a boundary, for example when ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}= {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$. For points $x,y \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$, the Bregman divergence from $x$ to $y$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is the difference between ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and the best linear approximation of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ at $y$, both evaluated at $x$: $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) - \left[{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(y) + {{\langle \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(y) , x-y \rangle}}\right] . \label{eqn:Bdistance}\end{aligned}$$ As illustrated in Figure \[fig:Bdistance\], we get ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}}$ by first drawing the hyperplane that touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ at the point $(y, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(y))$. We then intersect the vertical line that passes through $x$ with the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and the said hyperplane: the Bregman divergence is the height difference between the two intersections. Note that it is not necessarily symmetric: ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} \neq {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({y}\\|{x})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({y}\\|{x})}\)}}\fi}}$ for most ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}, x, y$. Accordingly, we introduce two balls of radius $r \geq 0$ centered at a point $x \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$: the primal Bregman ball containing all points $y$ so that the divergence from $x$ to $y$ is at most $r$, and the dual Bregman ball containing all points $y$ so that the divergence from $y$ to $x$ as at most $r$: $$\begin{aligned} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({x};{r})}\)}}\fi}} &= \{ y \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} \leq r \}; \label{eqn:ball1} \\ {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({x};{r})}\)}}\fi}} &= \{ y \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({y}\\|{x})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({y}\\|{x})}\)}}\fi}} \leq r \}. \label{eqn:ball2}\end{aligned}$$ To construct the primal ball geometrically, we take the point $(x, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x)-r)$ at height $r$ below the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and shine light along straight half-lines emanating from this point onto the graph. The ball is the vertical projection of the illuminated portion onto ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$; see Figure \[fig:Bball\]. To construct the dual ball geometrically, we start with the hyperplane that touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ at $(x, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x))$, translating it to height $r$ above the initial position. The ball is the vertical projection of the portion of the graph below the translated hyperplane onto ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$; see again Figure \[fig:Bball\]. Since ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}\)}}\fi}}$ is not necessarily symmetric, the two Bregman balls are not necessarily the same. Indeed, the dual ball is necessarily convex while the primal ball is not. ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}\)}}\fi}} \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\times {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ is strictly convex in the first argument but not necessarily convex in the second argument. Fixing $y$, set $f(x) = {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}}$. According to , $f$ is the difference between ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and an affine function; compare with the geometric interpretation of the dual Bregman ball. The strict convexity of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ implies the strict convexity of $f$. This argument does not apply to $g(y) = {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}}$, which we obtain by fixing $x$, and it is easy to find an example in which $g$ is non-convex; see Figure \[fig:itsaballs2\]. #### Legendre transform and conjugate function. In a nutshell, the Legendre transform applies elementary polarity to the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$, giving rise to the graph of another, conjugate function, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^* \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$, that relates to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ in interesting ways. If ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is of Legendre type then so is ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$; see [@Roc70 Theorem 26.5]. The notion of polarity we use in this paper relates points in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \times {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ with affine functions ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$. Specifically, it maps a point $C = (c, \gamma)$ to the function defined by $C^ (x) = {{\langle c , x \rangle}} - \gamma$, and it maps $C^$ back to $(C^)^* = C$. We refer to Figure \[fig:conjugate\] for an illustration and to Appendix \[appA\] for more details. As a first step in constructing the conjugate function, we get ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ as the set of points $e = c^ = \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(c)$ with $c \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$. We define $h \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ by mapping $c$ to $h(c) = c^$. Note that differentiability of strictly convex functions implies continuous differentiability [@DhDu12 Theorem 2.86], hence $h$ is a homeomorphism between the two domains. The conjugate function, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^* \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$, is then defined by mapping $e$ to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (e) = \epsilon$ such that $(e, \epsilon)$ is the polar point of the affine function whose graph touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ in the point $(c, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(c))$. Writing $b = a^$ and $q = p^$, we get $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({b}\\|{q})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({b}\\|{q})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^(b) - P^(b) \geq 0 , \label{eqn:nonnegative3} \\ {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({q}\\|{b})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({q}\\|{b})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^(q) - A^(q) \geq 0 \label{eqn:nonnegative4}\end{aligned}$$ from , and , in Appendix \[appA\]; see again Figure \[fig:conjugate\]. The left-hand sides of and are both non-negative and vanish iff $b = q$. Since this is true for all points $b, q \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ is strictly convex, provided ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ is convex. Proving that this assumption is always fulfilled is more involved. We therefore resort to a classical theorem [@Roc70 Theorem 26.5], which states that ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ is again of Legendre type and, in particular, ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ is convex. Hence, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ defines a Bregman divergence and, importantly, this divergence is symmetric to the one defined by ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. Let ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ and ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^* \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ be conjugate functions of Legendre type. Then ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{p})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{p})}\)}}\fi}} = {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({p^}\\|{a^})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({p^}\\|{a^})}\)}}\fi}}$ for all $a, p \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$. Let $A = (a, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a))$ and $P = (p, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p))$ be two points on the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$, and let $B^* (x) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) + {{\langle \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) , x-a \rangle}}$ and $Q^* (x) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) + {{\langle \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) , x-p \rangle}}$ be the corresponding affine functions. Then ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{p})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{p})}\)}}\fi}} = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) - Q^* (a)$, namely the height difference between the graphs of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and $Q^$ at $a$. Let $b = \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a)$, $q = \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p)$, and $\beta, \psi \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ such that $B = (b, \beta)$ and $Q = (q, \psi)$. By construction, $\beta = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^ (b)$ and $\psi = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (q)$. The conjugate Bregman divergence is ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({q}\\|{b})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({q}\\|{b})}\)}}\fi}} = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (q) - A^* (q)$, and applying with $C = Q$ and $S = A$, we get ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) - Q^(a) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^ (q) - A^* (q)$. The claimed relation follows. In words, the Legendre transform preserves the divergences, but it does so by exchanging the arguments. This is interesting because ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}\)}}\fi}}$ is strictly convex in the first argument and so is ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}\)}}\fi}}$, only that its first argument corresponds to the second argument of ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}\)}}\fi}}$. To avoid potential confusion, we thus consider the primal and dual Bregman balls of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$: $$\begin{aligned} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}{({u};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}}{({u};{r})}\)}}\fi}} &= \{ v \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^ \mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({u}\\|{v})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({u}\\|{v})}\)}}\fi}} \leq r \} , \label{eqn:ball3} \\ {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}'{({u};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}}'{({u};{r})}\)}}\fi}} &= \{ v \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^* \mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({v}\\|{u})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({v}\\|{u})}\)}}\fi}} \leq r \} , \label{eqn:ball4}\end{aligned}$$ where we write $u = x^$ and $v = y^$ so we can compare the two balls with the ones defined in and . As mentioned earlier, both dual balls are necessarily convex while both primal balls are possibly non-convex. Recall the homeomorphism $h \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ that maps $x$ to $x^$. It also maps ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({x};{r})}\)}}\fi}}$ to ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}'{({u};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}}'{({u};{r})}\)}}\fi}}$ and ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({x};{r})}\)}}\fi}}$ to ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}{({u};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}}{({u};{r})}\)}}\fi}}$. In words, it makes the non-convex ball convex and the convex ball non-convex, and it does this while preserving the divergences. We use this property to explain the necessity on using functions of Legendre type; it also plays a crucial role later. Consider a dual Bregman ball with a non-convex conjugate image, namely the corresponding primal ball. Then the restriction of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ to this dual ball is strictly convex and differentiable. However, it is not of Legendre type and its conjugate has a non-convex domain. #### Examples. We close this section with a short list of functions, their conjugates, and the corresponding Bregman divergences. Half the squared Euclidean norm maps a point $x \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$ to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = \tfrac{1}{2} {{\ifmmode{\\|{x}\\|}\else{\mbox{$\\|{x}\\|$}}\fi}}^2$. The gradient is $\nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = x$, and the conjugate is defined by ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (x) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x)$. The divergence associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is half the squared Euclidean distance: $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} &= \tfrac{1}{2} {{\ifmmode{\\|{x}-{y}\\|}\else{\mbox{$\\|{x}-{y}\\|$}}\fi}}^2 .\end{aligned}$$ This Bregman divergence is special because it is symmetric in the two arguments. The Shannon entropy of a discrete probability distribution is $- \sum_{i=1}^n x_i \ln x_i$. To turn this into a convex function, we change the sign, and to simplify the computations, we subtract the sum of the $x_i$, defining ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = \sum_{i=1}^n [x_i \ln x_i - x_i]$ over the positive orthant, which we denote as ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n_+$. The gradient is $\nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = [\ln x_1, \ln x_2, \ldots, \ln x_n]^T$, and the conjugate is the exponential function, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (u) = \sum_{i=1}^n e^{u_i}$, with $u = x^$, defined on ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$. Associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is the Kullback–Leibler divergence* and with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ is the exponential loss: $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} &= \sum_{i=1}^ n \left[x_i \ln \tfrac{x_i}{y_i}-x_i+y_i\right], \\ {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({u}\\|{v})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({u}\\|{v})}\)}}\fi}} &= \sum_{i=1}^ n \left[e^{u_i} - (u_i-v_i+1) e^{v_i} \right].\end{aligned}$$ The Kullback–Leibler is perhaps the best known Bregman divergence; it is also referred to as the information divergence, information gain, relative entropy; see [@AmNa00 page 57]. If applied to finite distributions, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ would be defined on the standard $(n-1)$-simplex, where it measures the difference in information when we go from $y$ to $x$. It also measures the expected number of extra bits required to code samples from $x$ using a code that is optimized for $y$ instead of for $x$. Since the $(n-1)$-simplex is the intersection of ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n_+$ with a hyperplane, this restriction of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is again of Legendre type. In this particular case, we can extend the function to the closed* $(n-1)$-simplex, so that some coordinates may be zero, provided we accept infinite divergences for some pairs. In other words, the framework is also suitable for sparse data, pervasive for example in text-retrieval applications. The Burg entropy maps a point $x \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n_+$ to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = \sum_{i=1}^n [1 - \ln x_i]$. The components of the gradient are $- 1 / x_i$, for $1 \leq i \leq n$. The conjugate is the function ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* \colon {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n_- \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ defined by ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (u) = \sum_{i=1}^n \left[ 1 - \ln \|u_i\| \right]$. Associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is the Itakura–Saito divergence: $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{y})}\)}}\fi}} &= \sum_{i=1}^n \left[ \tfrac{x_i}{y_i} - \ln \tfrac{x_i}{y_i} - 1 \right] .\end{aligned}$$ We note that ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ are very similar, but their domains are diagonally opposite orthants. Indeed, the Itakura–Saito distance is not symmetric and generates non-convex primal balls; see Figure \[fig:itsaballs2\]. Proximity Complexes for Bregman divergences {#sec3} =========================================== In this section, we extend the standard constructions of topological data analysis (Čech, Vietoris–Rips, Delaunay complexes) to the setting of Bregman divergences. Importantly, we prove the contractibility of non-empty common intersections of Bregman balls and Voronoi domains. This property guarantees that the Čech and Delaunay complexes capture the correct homotopy type of the data. #### Contractibility for balls. Every non-empty convex set is contractible, which means it has the homotopy type of a point. The common intersection of two or more convex sets is either empty or again convex and therefore contractible. While primal Bregman balls are not necessarily convex, we show that their common intersections are contractible unless empty. The reason for our interest in this property is the Nerve Theorem [@Bor48; @Ler45], which asserts that the nerve of a cover with said property has the same homotopy type as the union of this cover. Let ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ be of Legendre type, $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$, and $r \geq 0$. Then $\bigcap_{x \in X} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({x};{r})}\)}}\fi}}$ is either empty or contractible. Recall the homeomorphism $h \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$ obtained as a side-effect of applying the Legendre transform to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. It maps every primal Bregman ball in ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ homeomorphically to a dual Bregman ball in ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$, which is convex. Similarly, it maps the common intersection of primal Bregman balls in ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ homeomorphically to the common intersection of dual Bregman balls in ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$: $h({{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}) = {{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}$ in which ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}= \bigcap_{x \in X} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({x};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({x};{r})}\)}}\fi}}$ and ${{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}= \bigcap_{x \in X} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}}'{({x^};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}}'{({x^};{r})}\)}}\fi}}$. Since ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}$ and ${{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}$ are homeomorphic, they have the same homotopy type. Hence, either ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}= {{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}= \emptyset$ or ${{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}$ is convex and ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}$ is contractible. #### Čech and Vietoris–Rips constructions for Bregman divergences. The contractibility of the common intersection suggests we take the nerve of the Bregman balls. Given a finite set $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ and $r \geq 0$, we call the resulting simplicial complex the Čech complex of $X$ and $r$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. Related to it is the Vietoris–Rips complex, which is the clique complex of the $1$-skeleton of the Čech complex: $$\begin{aligned} {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} &= \{ P \subseteq X \mid \bigcap_{p \in P} {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({p};{r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({p};{r})}\)}}\fi}} \neq \emptyset \} , \label{eqn:Cech} \\ {{\ifmmode{\rm Rips}\else{\mbox{$\rm Rips$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} &= \{ Q \subseteq X \mid \tbinom{Q}{2} \subseteq {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} \} . \label{eqn:VietorisRips}\end{aligned}$$ In words, the Vietoris–Rips complex contains a simplex iff all its edges belong to the Čech complex. We note that for ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = {{\ifmmode{\\|{x}\\|}\else{\mbox{$\\|{x}\\|$}}\fi}}^2$, translates to the usual Euclidean definition of the Vietoris–Rips complex. Increasing the radius from $0$ to $\infty$, we get a filtration of Čech complexes and a filtration of Vietoris–Rips complexes. By construction, the Čech complex is contained in the Vietoris–Rips complex for the same radius. If we measure distance with the Euclidean metric, this relation extends to $$\begin{aligned} {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}{({X};{r})}} \subseteq {{\ifmmode{\rm Rips}\else{\mbox{$\rm Rips$}}\fi}{({X};{r})}} \subseteq {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}{({X};{\sqrt{2} r})}} . \label{eqn:interleaving}\end{aligned}$$ Indeed, if all pairs in a set of $k+1$ balls of radius $r$ have a non-empty common intersection, then increasing the radius to $\sqrt{2} r$ guarantees that the $k+1$ balls have a non-empty intersection. This fact is often expressed by saying that the two filtrations have a small interleaving distance if indexed logarithmically. #### No interleaving. The interleaving property expressed in extends to general metrics – except that the constant factor is $2$ rather than $\sqrt{2}$ – but not to general Bregman divergences. To see that does not extend, we give an example of $3$ points whose Bregman balls overlap pairwise for a small radius but not triplewise until the radius is very large. The example uses the exponential function defined on the standard triangle, which we parametrize using barycentric coordinates. For convenience, the explanation uses the conjugate function, which is the Shannon entropy; that is: we look at dual balls in which distance is measured with the Kullback–Leibler divergence. Specifically, we use ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) = \sum_{i=1}^3 x_i \ln x_i$. The barycentric coordinates are non-negative and satisfy $\sum_{i=1}^3 x_i = 1$. We therefore get the maximum value of $0$ at the three corners, and the minimum of $- \ln 3$ at the center of the triangle; see Figure \[fig:KL\_with\_planes\]. After some calculations, we get the squared length of the gradient at $x$ as $\tfrac{1}{3} [ (\ln x_1 - \ln x_2)^2 + (\ln x_1 - \ln x_3)^2 + (\ln x_2 - \ln x_3)^2 ]$. It goes to infinity when $x$ approaches the boundary of the triangle. We construct the example using points near the midpoints of the edges. Choosing them in the interior of the triangle but close to the boundary, the corresponding three tangent planes are as steep as we like. Moving the planes upward, we get the dual balls as the vertical projections of the parts of the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ on or below the planes. Moving the planes continuously, we let $r$ be the height above the initial positions, and note that $r$ is also the radius of the dual balls. Pairwise overlap between the balls starts when the three lines at which the planes meet intersect the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. This happens at $r < \ln 3$. Triplewise overlap starts when the point common to all three planes passes through the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. This happens at a value of $r$ that we can make arbitrarily large. #### Contractibility for Voronoi domains. Čech and Vietoris–Rips complexes can be high-dimensional and of exponential size, even if the data lives in low dimensions. To remedy this shortcoming, we use the Delaunay (or alpha) complex; see [@EdHa10; @EdMu94]. It is obtained by clipping the balls before taking the nerve. We explain this by introducing the Voronoi domains of the generating points as the clipping agents. Letting $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ be finite, we define the primal and dual Voronoi domains of $x \in X$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ as the sets of points for which $x$ minimizes the Bregman divergence to or from the point: $$\begin{aligned} {{\ifmmode{{V}_{\!{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x})}}\else{\mbox{${V}_{\!{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x})}\)}}\fi}} &= \{ a \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{a})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{a})}\)}}\fi}} \leq {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({y}\\|{a})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({y}\\|{a})}\)}}\fi}}, \forall y \in X \} ; \\ {{\ifmmode{{V}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}'{({x})}}\else{\mbox{${V}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}'{({x})}\)}}\fi}} &= \{ a \in {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\mid {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{x})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{x})}\)}}\fi}} \leq {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{y})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{y})}\)}}\fi}}, \forall y \in X \} .\end{aligned}$$ An intuitive construction of the primal domains grows the primal Bregman balls around the points, stopping the growth at places where the balls meet. Similarly, we get the dual Voronoi domains by growing dual Bregman balls. Not surprisingly, the primal Voronoi domains are not necessarily convex, and the dual Voronoi cells are convex. To see the latter property, we recall that the dual ball centered at $x$ is constructed by translating the hyperplane that touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ above $x$. Specifically, ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{x})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{x})}\)}}\fi}}$ is the height at which the hyperplane passes through the point $(a, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a))$. This implies that we can construct the dual Voronoi domains as follows: - For each $x \in X$, consider the half-space of points in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \times {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ on or above the hyperplane that touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ at $(x, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x))$. - Form the intersection of these half-spaces, which is a convex polyhedron. We call its boundary the upper envelope of the hyperplanes, noting that it is the graph of a piecewise linear function from ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$ to ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$. - Project the upper envelope vertically onto ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$. Each dual Voronoi domain is the intersection of ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ with the image of an $n$-dimensional face of the upper envelope. We conclude that the dual Voronoi domains are convex and use this property to show that the primal Voronoi domains intersect contractibly. Let ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ be of Legendre type, and $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ finite. Then $\bigcap_{x \in X} {{\ifmmode{{V}_{\!{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x})}}\else{\mbox{${V}_{\!{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x})}\)}}\fi}}$ is either empty or contractible. The proof is similar to that of the Contractibility Lemma for Balls and therefore omitted. As in the proof of the Contractibility Lemma for Balls, we use the conjugate, ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^* \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$, and the homeomorphism, $h \colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}^$, obtained by applying the Legendre transform to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. By the Duality Lemma, we have ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{a})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{a})}\)}}\fi}} = {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({a^}\\|{x^})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({a^}\\|{x^})}\)}}\fi}}$. Recall that the dual Voronoi domains associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$ are convex. They are the homeomorphic images of the primal Voronoi domains associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. Similarly, ${{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}= \bigcap_{x \in X} {{\ifmmode{{V}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}'{({x^})}}\else{\mbox{${V}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}'{({x^})}\)}}\fi}}$ is the homeomorphic image of ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}= \bigcap_{x \in X} {{\ifmmode{{V}_{\!{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x})}}\else{\mbox{${V}_{\!{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x})}\)}}\fi}}$. Hence, either ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}= {{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}= \emptyset$ or ${{\ifmmode{\mathbb{Y}}\else{\mbox{$\mathbb{Y}$}}\fi}}$ is convex and ${{\ifmmode{\mathbb{X}}\else{\mbox{$\mathbb{X}$}}\fi}}$ is contractible. #### Delaunay construction for Bregman divergences. Taking the nerve of the primal Voronoi domains, we get the Delaunay triangulation of $X$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$, which we denote as ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$. Further restricting the primal Voronoi domains by primal Bregman balls of radius $r$, we get the Delaunay complex of $X$ and $r$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$: $$\begin{aligned} \!\!{{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} &= \{ P \!\subseteq\! X \mid \! \bigcap_{p \in P} \! \left[ {{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({p};{\!r})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}{({p};{\!r})}\)}}\fi}} \cap {{\ifmmode{{V}_{\!{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({p})}}\else{\mbox{${V}_{\!{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({p})}\)}}\fi}} \right] \neq \emptyset \} .\end{aligned}$$ Assuming general position of the points in $X$, the Delaunay triangulation is a simplicial complex of dimension at most $n$. We will be explicit about what we mean by general position shortly. Combining the proofs of the two Contractibility Lemmas, we see that the common intersection of any set of clipped primary balls is either empty or contractible. This together with the Nerve Theorem implies that ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$ has the same homotopy type as ${{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$, namely the homotopy type of the union of the Bregman balls that define the two complexes. Algorithms {#sec4} ========== Recall that all three proximity complexes defined in Section \[sec3\] depend on a radius parameter. In this section, we give algorithms that compute the values of this parameter beyond which the simplices belong to the complexes. By focusing on the resulting radius functions, we decouple the computation of the radius for each simplex from the technicalities of constructing the actual simplicial complex. In particular, we show that the Čech complexes can be efficiently reconstructed from the Vietoris–Rips complexes, and the Delaunay complexes from the Delaunay triangulations. We exploit a connection with discrete Morse theory to develop efficient algorithms. #### Radius functions. Let $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ be finite, write $\Delta (X)$ for the simplex whose vertices are the points in $X$, and recall that ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$ is the Delaunay triangulation of $X$ associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. The Čech, Vietoris–Rips, and Delaunay radius functions associated with ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$, $$\begin{aligned} {{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} &\colon \Delta (X) \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}, \\ {{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Rips}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Rips}\)}}\fi}}\, &\colon \Delta (X) \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}, \\ {{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}}\,\; &\colon {{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}} \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}},\end{aligned}$$ are defined such that $P \in {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$ iff ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} (P) \leq r$, and similarly for Vietoris–Rips and for Delaunay. By definition of the Čech complex, ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} (P)$ is the minimum radius at which the primal Bregman balls centered at the points of $P$ have a non-empty common intersection. We are interested in an equivalent characterization using dual Bregman balls. To this end, we say that a dual Bregman ball, ${{\ifmmode{{B}'}\else{\mbox{${B}'$}}\fi}}$, includes $P$ if $P \subseteq {{\ifmmode{{B}'}\else{\mbox{${B}'$}}\fi}}$, and we call ${{\ifmmode{{B}'}\else{\mbox{${B}'$}}\fi}}$ the smallest including dual ball if there is no other dual ball that includes $P$ and has a smaller radius. Because ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is strictly convex, the smallest including dual ball of $P$ is unique; see Figure \[fig:itsaballs2\], which shows the smallest including dual Itakura–Saito ball of a pair of points. We call ${{\ifmmode{{B}'}\else{\mbox{${B}'$}}\fi}}$ empty if no point of $X$ lies in its interior, and we call it a circumball of $P$ if all points of $P$ lie on its boundary. We observe that a simplex $P \in \Delta (X)$ belongs to ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$ iff it has an empty dual circumball. Because ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is strictly convex, the smallest empty dual circumball of a simplex is either unique or does not exist. The characterization of the radius functions in terms of dual balls is strictly analogous to the Euclidean case studied in [@BaEd16]. Let ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ be of Legendre type, $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ finite, and $\emptyset \neq P \subseteq X$. 1. ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} (P)$ is the radius of the smallest including dual ball of $P$, and ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Rips}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Rips}\)}}\fi}} (P)$ is the maximum radius of the smallest including dual balls of the pairs in $P$. 2. Assuming $P \in {{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$, ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}} (P)$ is the radius of the smallest empty dual circumball of $P$. We omit the proof, which is not difficult. Every circumball also includes, which implies ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Rips}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Rips}\)}}\fi}} (P) \leq {{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} (P) \leq {{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}} (P)$ whenever the radius functions are defined. Correspondingly, ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} \subseteq {{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}} \subseteq {{\ifmmode{\rm Rips}\else{\mbox{$\rm Rips$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$ for every value of $r$. #### General position. It is often convenient and sometimes necessary to assume that the points in $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ are in general position, for example when we require the Delaunay triangulation be a simplicial complex in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$. Here is a notion that suffices for the purposes of this paper. Let ${{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\subseteq {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$ and ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ of Legendre type. A finite set $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ is in general position with respect to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ if, for every $P \subseteq X$ of cardinality at most $n+1$, 1. the points in $P$ are affinely independent, 2. no point of $X \setminus P$ lies on the boundary of the smallest dual circumball of $P$. Let $k = {{\ifmmode{\rm dim}\else{\mbox{$\rm dim$}}\fi}\,{P}}$. Property I implies that $P$ has an $(n-k)$-parameter family of circumballs. In particular, there is at least one circumball as long as $k \leq n$. Property II implies that no two different simplices have the same smallest dual circumball. In particular, no two $n$-simplices in the Delaunay triangulation have the same circumball. #### Discrete Morse theory. For points in general position, two of the radius functions exhibit a structural property that arises in the translation of Morse theoretic ideas from the smooth category to the simplicial category. Following [@BaEd16], we extend the original formulation of discrete Morse theory given by Forman [@For98]. Letting $K$ be a simplicial complex, and $P, R \in K$ two simplices, we write $P \leq R$ if $P$ is a face of $R$. The interval of simplices between $P$ and $R$ is $[P,R] = \{ Q \in K \mid P \leq Q \leq R \}$. We call $P$ the lower bound and $R$ the upper bound of the interval. A generalized discrete vector field is a partition of $K$ into intervals. We call it a generalized discrete gradient if there exists a function $f \colon K \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ such that $f(P) \leq f(Q)$ whenever $P$ is a face of $Q$, with equality iff $P$ and $Q$ belong to a common interval. A function with this property is called a generalized discrete Morse function. To get an intuitive feeling for this concept, consider the sequence of sublevel sets of $f$. Any two contiguous sublevel sets differ by one or more intervals, and any two of these intervals are independent in the sense that neither interval contains a face of a simplex in the other interval. Indeed, this property characterizes generalized discrete Morse functions. Let ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ be of Legendre type and let $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ be finite and in general position. Then ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} \colon \Delta (X) \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ and ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}} \colon {{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}} \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ are generalized discrete Morse functions. We give the proof in Appendix \[appB\]. Observe that the Vietoris–Rips radius function is not a generalized discrete Morse function. The structural properties implied by the GDMF Theorem will be useful in the design of algorithms that compute the radius functions. The theorem should be compared with the analogous result in the Euclidean case [@BaEd16]. The arguments used there can be translated almost verbatim to prove additional structural results for Bregman divergences. Perhaps most importantly, they imply that the Wrap complex of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and $X$ is well defined – see [@Ede03] for the original paper on these complexes defined in $3$-dimensional Euclidean space – and that the Čech complex collapses to the Delaunay complex and further to the Wrap complex, all defined for the same radius. #### Bregman circumball algorithm. Depending on how the function ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is represented, there may be a numerical component to the algorithms needed to find smallest including dual balls. Consider a $k$-simplex $Q \subseteq X$ with $0 \leq k \leq n$. Assuming general position, the affine hull of the points $A = (a, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a))$ with $a \in Q$ is a $k$-dimensional plane, which we denote as ${{\ifmmode{\cal Q}\else{\mbox{$\cal Q$}}\fi}}$. We are interested in the point $(q, \psi) \in {{\ifmmode{\cal Q}\else{\mbox{$\cal Q$}}\fi}}$ that maximizes $\psi - {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(q)$, the height above the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$. The point $q$ is the center of the smallest dual circumball of $Q$, and $\psi - {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(q)$ is the radius. Interestingly, this observation implies that the point of first intersection of two primal Bregman balls lies on a line joining their centers. For later reference, we assume a routine that computes this point, possibly using a standard numerical optimization method. mmmmmmmm= $(Q)$:\ let ${{\ifmmode{\cal Q}\else{\mbox{$\cal Q$}}\fi}}$ be the affine hull of the points $(a,{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a))$, $a \in Q$;\ find $(q, \psi) \in {{\ifmmode{\cal Q}\else{\mbox{$\cal Q$}}\fi}}$ maximizing $\psi - {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(q)$;\ $(q, \psi - {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(q))$. This is an unconstrained $k$-dimensional convex optimization, and $k$ is much smaller than $n$ for high dimensional data. Indeed, the optimization can be performed in the space of affine coordinates of the plane ${{\ifmmode{\cal Q}\else{\mbox{$\cal Q$}}\fi}}$. Importantly, the Hessian is of dimension $k \times k$ and not $n \times n$, which would be prohibitive. This allows us to use second-order quasi-Newton methods, such as the fast BFGS algorithm [@NoWr06]. Note that the smallest dual circumball of $Q$ includes $Q$ but is not necessarily the smallest including dual ball. However, the latter is necessarily the smallest dual circumball of a face of $Q$. Next, we show how the [CircumBall]{} routine is used to efficiently compute the radius functions. #### Čech radius function algorithm. According to the Radius Function Lemma (i), the value of a simplex, $Q \in \Delta (X)$, under the Čech radius function is the radius of the smallest including dual ball of $Q$. To compute this value, we visit the simplices in a particular sequence. Recalling the GDMF Theorem, we note that the smallest including dual ball of a simplex $Q$ is the smallest dual circumball of the minimum face $P \subseteq Q$ in the same interval. It is therefore opportune to traverse the simplices in the order of increasing dimension. Whenever the smallest dual circumball of a simplex $Q$ is not the smallest including dual ball, we get ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} (Q)$ from one of its codimension $1$ faces. We identify such a simplex $Q$ when we come across a face whose smallest dual circumball includes $Q$, and we mark $Q$ with the center and radius of this ball. The following pseudocode computes the radius function of the Čech complex restricted to the $k$-skeleton of $\Delta (X)$ for some nonnegative integer $k$: mmmmmmmm= $i = 0$ [to]{} $k$ [do]{}\ $P \subseteq X$ [with]{} ${{\ifmmode{\rm dim}\else{\mbox{$\rm dim$}}\fi}\,{P}} = i$ [do]{}\ $P$ unmarked [then]{} $(p, r) = \mbox{\sc CircumBall} (P)$;\ $a \in X$ with ${{\ifmmode{{D}_{F}{({a}\\|{p})}}\else{\mbox{${D}_{F}{({a}\\|{p})}$}}\fi}} < r $ [do]{} mark $P \cup \{a\}$ with $(p, r)$. As in the Euclidean setting, the size of $\Delta (X)$ is exponential in the size of $X$ so that the computations are feasible only for reasonably small values of $k$ or small radius cut-offs. In practice, we would run the algorithm with a radius cut-off, or use an approximation strategy yielding a similar persistence diagram. Observe the similarity to the standard algorithm for constructing the $k$-skeleton of the Vietoris–Rips complex: after adding all edges of length at most $2r$, we add simplices of dimension $2$ and higher whenever possible. Geometric considerations are thus restricted to edges and the rest of the construction is combinatorial; see [@Zom10] for a fast implementation. Our algorithm can be interpreted as constructing the Čech complex from the Vietoris–Rips complex at the cost of at most one call to [CircumBall]{} per simplex. This is more efficient than explicitly computing the smallest including dual ball for each simplex, even if we use fast randomized algorithms as described in [@NiNo06; @Wel91]. Furthermore, the [CircumBall]{} routine is only called for the lower bounds of the intervals of the Čech radius function or, equivalently, for each subcomplex in the resulting filtration. The number of such intervals depends on the relative position of the points in $X$ and not only on the cardinality. Notwithstanding, the number of intervals is significantly smaller than the number of simplices in the Čech complex. This suggests that only a small overhead is needed to compute the Čech from the Vietoris–Rips complexes. Our preliminary experiments for the Kullback-Leibler divergence support this claim; see Table \[table:exp\]. Note that the number of calls to the [CircumBall]{} routine is between $\frac{1}{10}$ and $\frac{1}{3}$ of the number of simplices, with an average between $6$ and $15$ function evaluations per call. [A]{} (20 pts) [B]{} (256 pts) [C]{} (4,000 pts) ---------------------------------------------- -------------------- --------------------- ----------------------- [\#]{}edges 190 7,715 36,937 [\#]{}simplices 1,048,575 1,155,301 1,222,688 [\#]{}calls to [CircumBall]{} 104,030 346,475 283,622 [\#]{}function evaluations in [CircumBall]{} 1,523,295 2,904,603 1,783,474 : Experimental evaluation on three synthetic datasets: ([[A]{}]{}) Full Čech complex with 20 points in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^{20}$; ([[B]{}]{}) $3$-skeleton with 256 points in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^4$ and radius cutoff $r=0.1$; ([[C]{}]{}) $4$-skeleton with 4,000 points in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^4$ and radius cutoff $r=0.01$. []{data-label="table:exp"} #### Delaunay radius function algorithm. According to the Radius Function Lemma (ii), the value of a simplex $Q \in {{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$ under the Delaunay radius function is the radius of the smallest empty dual Bregman circumball of $Q$. mmmmmmmm= $(Q)$:\ $(q, r) = \mbox{\sc CircumBall} (Q)$;\ $a \in X \setminus Q$ [do]{}\ ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{q})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{q})}\)}}\fi}} < r$ [then]{} [return]{} [none]{};\ $r$. The [CircumBall]{} routine gives only the smallest dual circumball of $Q$, and if it is not empty, then we have to get the value of the Delaunay radius function from somewhere else. According to the GDMF Theorem, we get the value from the maximum simplex in the interval that contains $Q$. It is therefore opportune to traverse the simplices of the Delaunay triangulation in the order of decreasing dimension. Whenever the smallest dual circumball of a simplex $Q$ is non-empty, we get ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}} (Q)$ from one of the simplices that contain $Q$ as a codimension $1$ face. As already observed in [@BNN10], we can construct the full Delaunay triangulation, ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$, using existing algorithms for the Euclidean case. We get the Delaunay complexes as sublevel sets of the radius function. Specifically, we first use the polarity transform to map the points $(x, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x))$ to the corresponding affine functions; see Section \[sec2\]. We then get a geometric realization of ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$ from the vertical projection of the upper envelope of the affine functions onto ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$, which is a Euclidean weighted Voronoi diagram, also known as power diagram or Dirichlet tessellation. Its dual is the Euclidean weighted Delaunay triangulation, also known as regular or coherent triangulation. The data that defines these Euclidean diagrams are the points $x \in X$ with weights $\xi = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(x) - {{\ifmmode{\\|{x}\\|}\else{\mbox{$\\|{x}\\|$}}\fi}}^2$. Finally, after computing the radius function on all simplices in ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}}$, we get the Delaunay complexes as a filtration of this weighted Delaunay triangulation. Interestingly, this is not necessarily the filtration we obtain by simultaneously and uniformly increasing the weights of the points. Discussion {#sec5} ========== The main contribution of this paper is the extension of the mathematical and computational machinery of topological data analysis (TDA) to applications in which distance is measured with a Bregman divergence. This includes text and image data often compared with the Kullback–Leibler divergence, and speech and sound data often studied with the Itakura–Saito divergence. It is our hope that the combination of Bregman divergences and TDA technology will bring light into the generally difficult study of high-dimensional data. In support of this optimism, Rieck and Leitte [@RiLe15] provide experimental evidence that good dimension reduction methods preserve the persistent homology of the data. With our extension to Bregman divergences, such experiments can now be performed for a much wider spectrum of applications. There are specific mathematical questions whose incomplete understanding is currently an obstacle to progress in the direction suggested by this paper: - A cornerstone of TDA is the stability of its persistence diagrams, as originally proved in [@CEH07]. How does the use of Bregman divergences affect the stability of the diagrams? - Related to the question of stability is the existence of sparse complexes and filtrations for data in Bregman spaces whose persistence diagrams are close to the ones we get for the Čech and Delaunay complexes. Last but not least, we mention the urgent task to further study the related algorithmic questions and to implement software that is fast, can cope with large sets of data, and is easy to use also for non-specialists. Acknowledgements {#acknowledgements .unnumbered} ---------------- [The authors thank Žiga Virk for discussions on the material presented in this paper.]{} Polarity and Legendre Transform {#appA} =============================== In this appendix, we give further details how the polarity transform amounts to the Legendre transform for functions of Legendre type. Recall that the polarity maps a point $C = (c, \gamma) \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \times {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ to the function $C^* \colon {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ defined by $C^* (x) = {{\langle c , x \rangle}} - \gamma$, and that it maps $C^$ back to $(C^)^* = C$. Given a second point $S = (s, \sigma) \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n \times {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$, and the corresponding affine function $S^* (x) = {{\langle s , x \rangle}} - \sigma$, the transform preserves the difference between the values: $$\begin{aligned} \sigma - C^* (s) &= \gamma - S^* (c) . \label{eqn:duality}\end{aligned}$$ Indeed, both sides of the equation evaluate to $\gamma + \sigma - {{\langle c , s \rangle}}$. To apply the polarity transform to ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$, consider a point $A = (a, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a))$ and note that the graph of the affine function defined by $B^(x) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) + {{\langle \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) , x-a \rangle}}$ is the hyperplane that touches the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ at $A$. Let $P = (p, {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p))$ be another point on the graph of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ and $Q^(x) = {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) + {{\langle \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) , x-p \rangle}}$ the corresponding affine function. To avoid potential confusion, we note that $B^$ and $A^$ are generally different, and so are $Q^$ and $P^$. Since ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}$ is strictly convex, we have $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({p}\\|{a})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({p}\\|{a})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) - B^(p) \geq 0 , \label{eqn:nonnegative1} \\ {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({a}\\|{p})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({a}\\|{p})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) - Q^(a) \geq 0 , \label{eqn:nonnegative2}\end{aligned}$$ with vanishing lefthand sides iff $a = p$; see Figure \[fig:conjugate\]. Applying the polarity transform, we get two additional point/affine function pairs, namely $B, A^$ and $Q, P^$. We define $b = a^* = \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a)$ and $q = p^* = \nabla {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p)$ in ${{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}^n$ and $\beta, \psi \in {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ such that $B = (b, \beta)$ and $Q = (q, \psi)$; see again Figure \[fig:conjugate\]. Relating the two points with the two lines using , we get , , , by setting $C, S$ to $A, B$, to $P, Q$, to $P, B$, and to $A, Q$, in this sequence: $$\begin{aligned} \beta - A^* (b) &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) - B^* (a), \label{eqn:dual1} \\ \psi - P^* (q) &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) - Q^* (p), \label{eqn:dual2} \\ \beta - P^* (b) &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(p) - B^* (p), \label{eqn:dual3} \\ \psi - A^* (q) &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}(a) - Q^* (a). \label{eqn:dual4}\end{aligned}$$ The two sides in and in vanish by construction of $B^$ and $Q^$. Using and , we see that the terms in and are non-negative, and that they vanish iff $a = p$. Their lefthand sides are the Bregman divergences between $b$ and $q$ under ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$, and their righthand sides can be rewritten using the Duality Lemma: $$\begin{aligned} {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({b}\\|{q})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({b}\\|{q})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^ (b) - P^* (b) , \\ {{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^}{({q}\\|{b})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}^}{({q}\\|{b})}\)}}\fi}} &= {{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^* (q) - A^* (q) .\end{aligned}$$ This provides the crucial inequalities that imply the required properties of ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}^$, as enumerated in Section \[sec2\]. Discrete Morse Theory {#appB} ===================== In this appendix, we present the proof of the GDMF Theorem. Recall that this theorem claims that for ${{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}\colon {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}\to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ of Legendre type, and $X \subseteq {{\ifmmode{\Omega}\else{\mbox{$\Omega$}}\fi}}$ finite and in general position, ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}} \colon \Delta (X) \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ and ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}} \colon {{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({X})}} \to {{\ifmmode{\mathbb{R}}\else{\mbox{$\mathbb{R}$}}\fi}}$ are generalized discrete Morse functions. We consider ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}}$ first. Let $P \subseteq X$ be a $k$-simplex and consider two possibly different dual balls defined for $P$: the smallest including dual ball, ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({p_0};{r_0})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({p_0};{r_0})}\)}}\fi}}$, and the smallest dual circumball, ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({p_1};{r_1})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({p_1};{r_1})}\)}}\fi}}$. The first ball always exists, and by assumption of general position, the second ball exists iff $k \leq n$. We are interested in simplices for which the two balls are the same, which excludes simplices of dimension larger than $n$. They are the lower bounds of the intervals in the generalized discrete gradient [@BaEd16]. Let $P$ be such a simplex, and let $R$ be the set of points $x \in X$ with ${{\ifmmode{{D}_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}{({x}\\|{p_0})}}\else{\mbox{${D}_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}{({x}\\|{p_0})}\)}}\fi}} \leq r_0$. Clearly, $P \subseteq R$, and all simplices $P \subseteq Q \subseteq R$ have ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({p_0};{r_0})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({p_0};{r_0})}\)}}\fi}}$ as the smallest including dual ball. All simplices in $[P, R]$ belong to ${{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r_0})}}$ but none of them belongs to ${{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$ with $r < r_0$. If $r_0 = r_1$ is unique for $P$, then this is the only difference between ${{\ifmmode{\rm \check{C}ech}\else{\mbox{$\rm \check{C}ech$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r_0})}}$ and its immediate predecessors. Else, the difference consists of two or more intervals. By assumption of general position, there are no face relations between the simplices in two different such intervals. It follows that ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm \check{C}ech}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm \check{C}ech}\)}}\fi}}$ is a generalized discrete Morse function. We consider ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}}$ second. The argument is similar, except that the relevant dual balls are different. Besides the smallest dual circumball of $P$, ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({p_1};{r_1})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({p_1};{r_1})}\)}}\fi}}$, we also consider the smallest empty dual circumball, ${{\ifmmode{{B}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}'{({p_2};{r_2})}}\else{\mbox{${B}_{{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}}'{({p_2};{r_2})}\)}}\fi}}$. The latter exists iff $P$ belongs to the Delaunay triangulation of $X$. Again, we are interested in simplices for which the two balls are the same. They are the upper bounds of the intervals in the generalized discrete gradient [@BaEd16]. Let $P$ be such a simplex, and let $R$ be the smallest face of $P$ such that the smallest containing dual ball of $R$ contains $P$. We note that in this case, $R$ has the same smallest empty dual circumball as $P$. Furthermore, $R \subseteq P$, and all simplices $R \subseteq Q \subseteq P$ have the same smallest empty dual circumball. Hence, all simplices in $[R, P]$ belong to ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r_1})}}$, and none of them belongs to ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r})}}$ with $r < r_1$. If $r_1 = r_2$ is unique for $P$, then this is the only difference between ${{\ifmmode{\rm Del}\else{\mbox{$\rm Del$}}\fi}_{{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}}{({X};{r_1})}}$ and its immediate predecessors. Else, the difference consists of two or more intervals, and general position again implies that there are no face relations between the simplices in two different such intervals. It follows that ${{\ifmmode{\varrho_{{{\ifmmode{{F}}\else{\mbox{${F}$}}\fi}}}^{\rm Del}}\else{\mbox{$\varrho_{{{\ifmmode{{F}}\else{\mbox{\({F}$}}\fi}}}^{\rm Del}\)}}\fi}}$ is a generalized discrete Morse function. [15]{} Methods of Information Geometry.* Amer. Math. Soc., Providence, Rhode Island, 2000. Clustering with Bregman divergences. J. Mach. Learn. Res. [6]{} (2005), 1705–1749. The Morse theory of Čech and Delaunay complexes. Trans. Amer. Math. Soc., to appear. Legendre functions and the method of random Bregman projections. J. Convex Analysis [4]{} (1997), 27-67. Using Kullback–Leibler distance for text categorization. In “Proc. 25th European Conf. Inform. Retrieval, 2003”, LNCS [2633]{}, 305–319. Bregman Voronoi diagrams. Discrete Comput. Geom. [44]{} (2010), 281–307. On the imbedding of systems of compacta in simplicial complexes. Fund. Math. [35]{} (1948), 217–234. The relaxation method of finding the common point of convex sets and its applications to the solution of problems in convex programming. USSR Comput. Math. Math. Phys. [7]{} (1967), 200–217. Topology and data. Bull. Amer. Math. Soc. [46]{} (2009), 255–308. Topology of viral evolution. Proc. Natl. Acad. Sci. [110]{} (2013), 18566–18571. Towards persistence-based reconstruction in euclidean spaces. In “Proc. 24th Ann. Sympos. Comput. Geom., 2008”, 232–241. Stability of persistence diagrams. Discrete Comput. Geom. [37]{} (2007), 103–120. Optimality Conditions in Convex Optimization: a Finite-dimensional View. CRC Press, Taylor & Francis Group, Boca Raton, Florida, 2012. Wavelet-based texture retrieval using generalized Gaussian density and Kullback–Leibler distance. IEEE Trans. Image Proc. [11]{} (2002), 146–158. Surface reconstruction by wrapping finite point sets in space. In Discrete and Computational Geometry. The Goodman–Pollack Festschrift, 379–404, eds. B. Aronov, S. Basu, J. Pach and M. Sharir, Springer-Verlag, 2003. Computational Topology. An Introduction. Amer. Math. Soc., Providence, Rhode Island, 2010. Three-dimensional alpha shapes. ACM Trans. Graphics [13]{} (1994), 43–72. Nonnegative matrix factorization with the Itakura–Saito divergence: with application to music analysis. Neural Comput. [21]{} (2009), 793–830. Morse theory for cell complexes. Adv. Math. [134]{} (1998), 90–145. Algebraic Topology. Cambridge Univ. Press, Cambridge, England, 2002. Similarity measures for text document clustering. Proc. 6th New Zealand Computer Science Research Student Conference, 49–56, 2008. An analysis-synthesis telephony based on the maximum likelihood method. In “Proc. 6th Internat. Congress Acoustics, 1968”, Tokyo, Japan, c17–c20. On information and sufficiency. Ann. Math. Stat. [22]{} (1951), 79–86. Sur la forme des espaces topologiques et sur les points fixes des représentations. J. Math. Pures Appl. [24]{} (1945), 95–167. Persistent homology and many-body atomic structure for medium-range order in the glass. Nanotechnology, [26]{} (2015), 304001. On the smallest enclosing information disk. Proc. 18th Canad. Conf. Comput. Geom., 2006. . Numerical Optimization. Springer Science and Business Media, 2006. Persistent topology of syntax. arXiv:1507.05134, 2015. Persistent homology for the evaluation of dimensionality reduction schemes. Computer Graphics Forum [34]{} (2015), 431–440. Convex Analysis. Princeton University Press, Princeton, New Jersey, 1970. The persistent cosmic web and its filamentary structure–I. Theory and implementation. Monthly Notices Royal Astro. Soc. [414]{} (2011), 350–383. Smallest enclosing disks (balls and ellipsoids). In New Results and New Trends in Computer Science, H. A. Maurer (ed.), Springer, LNCS [555]{} (1991), 359–370. Fast construction of Vietoris–Rips complex. Computer & Graphics [34]{} (2010), 263–271. [^1]: This research is partially supported by the [Toposys]{} project FP7-ICT-318493-STREP.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Consider the problem: Alice wishes to send the same key to $n-1$ users (Bob, Carol,. . . , Nathan), while preventing eavesdropper Eve from acquiring information without being detected. The problem has no solution in the classical cryptography but in quantum telecommunication there are some codes to solve the problem. In the paper [@zengetall], Guo-Jyun Zeng, Kuan-Hung Chen, Zhe-Hua Chang, Yu-Shan Yang, and Yao-Hsin Chou from one side and Cabello in [@cabello] from other side, used Hadamard gates, Pauli gates in providing the quantum communication code for two-partity telecommunication with 3 persons and then generalized it to the case of arbitrary number of participants, indicating the position of measurements of participants. We remark that the Hadamard gate with precising the position of measurement is the same as Fourier transform for two qubits and hence use the general Fourier transform for $n$ entangled qubits, in place of Hadamard gates. The result is more natural for arbitrary $n$ qudits.' address: - '${}^1$ Institute of Mathematics, Vietnam National Academy of Science and Technology, Hoang Quoc Viet road, Cau Giay district, 10307 Hanoi, Vietnam' - '${}^2$ Thang Long University, Nghiem Xuan Yem, Hoang Mai district, Hanoi, Vietnam' author: - 'Do Ngoc Diep${}^{1,2}$' title: Multiparty Quantum Telecommunication Using Quantum Fourier Transforms --- Introduction ============ The paper is devoted to the following problem of multiparty quantum telecommunication, see [@cabello] for more details: Alice wishes to send a sequence of random classical bits (a “key”) to Bob, and at the same time preventing that Eve acquires information without being detected. This problem, as known has no solution by classical cryptography, but it can be solved in quantum telecommunication by using the quantum computing. There are some tools to solve the problems: Some peoples use the non-cloning principle, some ones use entanglement particles, some others combine quantum techniques with classical private amplification and compression techniques, or split out the information in several qubits to which eavesdropper Eve has only a sequential access. In the works [@cabello] and [@zengetall] the problem was solved by a procedure that we will remind in the next sections 2. In those code, the Hadamard gates were used to makes measurement of entangled Bell states and/or general entangled GHZ-states some times with indication of position to make measurements and directions[@zengetall]. We remark that in the last the Hadamard gates can be replaced by a more general quantum Fourier transform and the Fourier coefficients can be considered as the results of measurement, which is called quantum Fourier measurements. In section 3 we show how to measure the quantum Fourier transform coefficients. and in section 4 we show how to use it to solve the general problem of multipartite quantum telecommunication. Our main results are Definition \[defn1\] and Theorem \[thm1\]. In section 5 we remind of security estimation and in Section 6 we illustrate the codes with the holographic softwares and finally in Section 7 we make some final conclusion is made. The two-party quantum telecommunication ======================================= The problem of two-party quantum teleportation between 3 persons can be formulated as follows. There three persons, Alice and Bob and Carol, who want to exchange their idea, Alice is some leader and Bob and Carol are participants. In other word, Alice, Bob and Carol have the secret key $S_A$, $S_B$ and $S_C$ respectively. And the final key is $S_A \oplus S_B\oplus S_C$ and that each participant can change the key by their idea, see [@cabello], [@zengetall] for more details. In practice the problem appears in some specialized context, namely [@cabello]: A. Hillery-Bužek-Berthiaume secret sharing using GHZ states “Alice wants to have a secret action taken on her behalf in a distant part. There she has two agents, Bob and Carol, who carry it out for her. Alice knows that one and only one of them is dishonest, but she does not know which one. She cannot simply send a secure message to both of them, because the dishonest one will try to sabotage the action, but she knows that if both carry it out together, the honest one will keep the dishonest one from doing any damage.” B. Multiparty key distribution based on Bell-state entanglement swapping and and Pauli actions Supposed to consider the same situation but produce another protocol for secret sharing using Bell states instead of GHZ states is proposed, but also some Pauli gates $\sigma_z$, $\sigma_x$. This was later develloped in the work [@zengetall]. C. Secret sharing using Bell-state and GHZ entanglement swapping Supposed to consider the same situation but produce a protocol such that once Bob and Carol knows the results of the public measurement, he/she can infer the first bit of the result of Alice’s secret measurement and then use the public measurements from Alice they in cooperating together, can infer the second. The four steps solution of the problem from can be summerized as follows, see [@cabello] for more details. $$\CD \begin{array}{l} \mbox{\bf Step i. Initialization:} \mbox{Alice uses 5 qubits: 1 and 2}\\ \mbox{(unmoving) in Bell state, and 3 (unmoving), A and B }\\ \mbox{(moving for interchange) in GHZ state} \\ \mbox{Carol has 2 qubits: 4(unmoving) and C(moving) in Bell state}\\ \mbox{Bob has 2 qubits: 5(unmoving) and D (moving ) in Bell state}\end{array}\\ @VV\begin{array}{l} \mbox{Alice measures the Bell state in qubits 1 and 2, }\\ \mbox{Alice measures GHZ state in qubits 3, A and B, }\\ \mbox{Carol measures the Bell state in qubits 4 and C}\\ \mbox{Bob measures the Bell state in qubits 5 and D}\\ \end{array}V\\ \begin{array}{l} \mbox{After that the system is for example, namely in the initial state:}\\ \|\psi_i\rangle = \|000\rangle_{3AB} \otimes \|00\rangle_{12} \otimes \|00\rangle_{4C} \otimes \|00\rangle_{5D} \\ \\ \mbox{\bf Step ii. Bell-state measurements} \end{array} \\ @VV\begin{array}{l}\mbox{- Alice sends qubit A to Bob, qubit B to Carol}\\ %\mbox{- Alice performs a \textit{secret} Bell-state measurement}\\ \mbox{ on qubits 2 and 3}\\ \mbox{- Bob performs a Bell-state measurement}\\ \mbox{ on qubits 5 and D }\\ \mbox{- Carol performs a Bell-state measurement}\\ \mbox{ on qubits 4 and C } \end{array} V \\ \begin{array}{l} \mbox{After that it is easy to see that the system is in the state:} \\ \|\psi_{ii}\rangle = \|AP\rangle_{12}\otimes \|BP\rangle_{5D} \otimes \|CP\rangle_{4C} \\ \\ \mbox{\bf Step iii. Secret Bell-state measurements.} \end{array}\\ @VV\begin{array}{l}\mbox{- Alice performs a \textit{secret} Bell-state measurement}\\ \mbox{ on qubits 2 and 3}\\ \mbox{- Bob performs a \textit{secret} Bell-state measurement}\\ \mbox{ on qubits 5 and A }\\ \mbox{- Carol performs a \textit{secret} Bell-state measurement}\\ \mbox{ on qubits 4 and B }\end{array}V \\ \begin{array}{l} \mbox{ After that it is easy to see that the system is in the state:}\\ \|\psi_{iii}\rangle= \|AP\rangle_{1CD} \otimes \|AS\rangle_{23} \otimes \|BS\rangle_{5A}\otimes \|CS\rangle_{4B}\\ \\ \mbox{\bf Step iv. Secret sharing} \end{array}\\ @VV\begin{array}{l} \mbox{- Bob (resp., Carol) sends qubit D (resp., C) out}\\ \mbox{ to Alice}\\ \mbox{ - Alice performs a complete GHZ-state}\\ \mbox{measurement on qubits 1, C, and Dand publishes}\\ \end{array}V \\ \begin{array}{l} \mbox{After that it is easy to see that the system is in the state:}\\ \|\psi_{iv}\rangle = \|AP\rangle_{1CD} \otimes \|AS\rangle_{23} \otimes \|BS\rangle_{5A} \otimes \|CS\rangle_{4B}\\ \end{array} \endCD$$ This four steps are illustrated on Figure 1. ![Steps 1-4: ](figures/2partites-step1.png "fig:") ![Steps 1-4: ](figures/2partites-step3.png "fig:") \[fig:1\] Quantum Fourier Transforms as Quantum Measurements ================================================== The main tools are the entangled Bell state measurements and the entangled GHZ state measurements, which can be described as follows [@zengetall]. \[fig:Bellmeasurement\] ![The entangled measurement (a) The entangled Bell states measurement (b) The entangled GHZ states measurement.](figures/entangledmeasurement_a.png "fig:") \[fig:GHZmeasurement\] ![The entangled measurement (a) The entangled Bell states measurement (b) The entangled GHZ states measurement.](figures/entangledmeasurement_b.png "fig:") We remark that the schemes for entangled Bell state measurments are the same as the Fourier transform on these two qubits and entangled GHZ state measurments are the same as the Fourier transform on three entangled qubits. The next section is devoted to the general measurements of entangled GHZ states, using the quantum Fourier transforms for and arbitrary number $n$ of partites. For a state $\|x\rangle$ we define the Fourier coefficients in a standard basis as the results of Fourier measurement. \[defn1\] For any set of entangled qubit states $$\|\mathbf x\rangle = \frac{1}{\sqrt{2^{n}}}\sum_{y\in F_2^{n}} e^{2\pi i\mathbf x.\mathbf y/2^{n}}\|\mathbf y\rangle,$$ the Fourier coefficients are considered as the results of Fourier measurement. For a fixed orthonormal basis consisting of $2^n$ vectors of $n$ entangled qubits in GHZ states \[basis\] $$\|0\dots 0\rangle_{ij\dots n}=\frac{1}{\sqrt{2}}\left(\|0\rangle_i \otimes \|0\rangle_j\otimes\dots\otimes \|0\rangle_n + \|1\rangle_i \otimes \|1\rangle_j\otimes\dots \otimes\|1\rangle_n\right)$$ $$\|0\dots 1\rangle_{ij\dots n}=\frac{1}{\sqrt{2}}\left(\|0\rangle_i \otimes \|0\rangle_j\otimes\dots\otimes \|0\rangle_n - \|1\rangle_i \otimes \|1\rangle_j\otimes\dots \otimes\|1\rangle_n\right)$$ $$\|11\dots 10\rangle_{ij\dots n}=\frac{1}{\sqrt{2}}\left(\|1\rangle_i \otimes \|1\rangle_j\otimes\dots\otimes \|0\rangle_n + \|0\rangle_i \otimes \|1\rangle_j\otimes\dots \otimes\|1\rangle_n\right)$$ $$\|11\dots 11\rangle_{ij\dots n}=\frac{1}{\sqrt{2}}\left(\|1\rangle_i \otimes \|0\rangle_j\otimes\dots\otimes\|0\rangle_n - \|0\rangle_i \otimes \|1\rangle_j\otimes\dots \otimes\|1\rangle_n\right)$$ the Fourier transform measurements give the coefficients as the values of measurements. Let us first remind the Quantum Fourier transform code picture, Figure 3 from Ekert lectures [@ekertetall]: ![Quantum Fourier transform measurement of multiparty entangled states](figures/40.png) The general case ================ Consider the following problem: Alice wishes to convey the same key to N users (Bob, Carol,. . . , Nathan), while preventing Eve from acquiring information without being detected. This problem, called multiparty key distribution, is a special case of networked cryptographic conferencing. Here I introduce a protocol for using GHZ states for multiparty quantum key distribution that, as far as I know, has not been presented anywhere before. It can be considered as a generalization to many parties of the two-party protocol. \[thm1\] The above multiparty quantum telecommunication problem of secret sharing with quantum key distribution can be solved by a procedure with using the quantum Fourier transform measurements. Proof. The theorem is proved by the following Procedure, the same as in the 3 persons case [@cabello], the system state is changing as follows. $$\CD \|\psi_i\rangle @>>> \|\psi_{ii}\rangle @>>> \|\psi_{iii}\rangle @>>> \|\psi_{iv}\rangle \endCD$$ Let us consider in more details. Step 1. Initialization of 3n qubits. Alice has $n+2$ qubits: $1,2,3, A_1,\dots, A_{n-1}$: qubits 1 and 2 are entangled in Bell state, qubits $3, A_1, \dots, A_{n-1}$ are entangled in GHZ state. $n-1$ persons: Bob, Carol, ...., Nathan, each has 2 entangled qubits $i+3, C_i,i=1\dots,n-1$ namely in null state. Alice produces a Bell state measurement on qubit 1 and 2 and a Fourier measurement on $n$ qubits $3, A_1,\dots, A_{n-1}$. Each of participants makes a Bell state Fourier measurement of entangled $i+3,C_i, i=1\dots n-1$. At the end of this step 1, the system is in the state $$\|\psi_i\rangle = \|0\dots 0\rangle_{3A_1\dots A_{n-1}} \otimes \|00\rangle_{12} \otimes \|00\rangle_{4C_1}\otimes\dots \otimes \|00\rangle_{n+2,C_{n-1}}$$ Step 2. Entangled Bell-state measurements. Alice sends each qubit $A_i$ of her GHZ state out to each $i^{th}$ participant of the other $n-1$ users. The system is in the state $$\|\psi_{ii}\rangle = \|AP\rangle_{3A_1\dots A_{n-1}} \otimes \|BP\rangle_{1C_1} \otimes\|CP\rangle_{2C_2} \otimes\dots \otimes \|NP\rangle_{{n+2},C_{n-1}}$$ Step 3. Secret Bell-state measurement. Next, Alice and each user performs a Bell-state Fourier measurement on the received qubit and one of their qubits. After these measurements the state of the system becomes $$\|\psi_{iii}\rangle = \|AP\rangle_{3A_1\dots A_{n-1}} \otimes \|AS\rangle_{2,3}\otimes \|BS\rangle_{4,A_1} \otimes \dots \otimes \|NS\rangle_{n+2,A_{n-1}},$$ where $\|AP\rangle$ is $n$-qubit GHZ state of the standard basis \[basis\]. Step 4. Secret sharing. The $n-1$ users sends a qubit (the one they have not used) to Alice, and she performs a Fourier measurement to discriminate between the $2^n$ GHZ states, and publicly announces the result $\|AP\rangle_{3C_1\dots C_{n-1}}$. After these measurements the state of the system becomes $$\|\psi_{iv}\rangle = \|AP\rangle_{1C_1\dots C_{n-1}} \otimes \|AS\rangle_{2,3}\otimes \|BS\rangle_{4,A_1} \otimes \dots \otimes \|NS\rangle_{n+2,A_{n-1}},$$ The result AP, and the result of their own secret measurement allow each legitimate user to infer the first bit of Alice’s secret result $AS$. To find out the second bit of AS, all users (except Alice) must cooperate. The proof therefore is achieved. $\Box$ The same is true for qudits in place of qubits. In that situation we do use the phase Fourier coefficients in place of $\pm$. Security ======== It was shown [@cabello] that the protocools guarantees the security in the following sense: The secret that Alice admits has two qubits 2 and 3. by the public entangled Bell-state measurement of qubit 1 and 2, 4 and $C_1$, 5 and $C_2$ etc. and the entangled GHZ state of 1 and $C_1$ and $C_{n-1}$, every participant knows the first bit of the secret of Alice. In order to find out the second bit of Alice’s secret all $n-1$ participand do cooperate together and following the public result of GHZ state measurement of qubits $3, A_1,\dots,A_{n-1}$ every body knows also the second bit of the qubit 3 of Alice. The Alice’s secret is therefore discovered by each participant. A eavesdropper Eve can not do change the situation: Eve need to have the same as each participant - any attempt ot find out one of the secret result of participant will change the Alice’s public GHZ measurement result AP and peoples know about attempting of Eve. Detecting Eve’s presence requires the comparison of fewer bits. The probability of the result of AP to be $n$ times false is $$\frac{1}{2^n}\left( 1+ 2+\dots+2^{n-1}\right)=\frac{2^n-1}{2^n}.$$ In all these cases peoples observe the brochen results of AP and remove the telecommunication, while Eve cannot discover the secret AS. Illustration with holographic softwares ======================================= In this section we review the work of A. Jaffe, Z. Liu, and A. Wozniakowski [@jaffeetal] involving the softwares of sharing problem. The pictures are taken from their work. First let us remind that the measurement can also be produced with the code as shown in Figure 4 ![Measurement and corresponding code](figures/measure1.png "fig:") ![Measurement and corresponding code](figures/measure2.png "fig:") This kind of scheme can then be applied to 1-qubit-teleport problem in Figure 5 ![1-qubit-teleport problem and corresponding code](figures/tele1.png "fig:") ![1-qubit-teleport problem and corresponding code](figures/tele2.png "fig:") The secret sharing problem between Alice, Bob and Carol can be illustrated as the holographic software and code in Figure 6. ![Secret sharing problem between Alice, Bob and Carol illustrated as the holographic software and corresponding code](figures/3-sharing1.png "fig:") ![Secret sharing problem between Alice, Bob and Carol illustrated as the holographic software and corresponding code](figures/3-sharing2.png "fig:") The scheme for Alice-Bob-Carol sharing is easily generalized as BVK code for n-partitie-sharing problem as shown in Figure 7 ![n-partitie-sharing problem and corresponding BVK code](figures/n-sharing1.png "fig:") ![n-partitie-sharing problem and corresponding BVK code](figures/n-sharing2.png "fig:") Conclusion ========== We proposed to use the Fourier transform measurements for entangled GHZ-state of qubits. The result is independent of indicating the positions and directions, as used in [@zengetall]. The result is certainly true also for qudits in place of qubits. The codes can be illustrated by the holographic software and corresponding codes. [xxx]{} , [Multiparty key distribution and secret sharing based on entanglement swapping]{}, arXiv:quant-ph/0009025v1, 2000. , [Basis Concepts In Quantum Computation]{}, Centre for Quantum Computation, University of Oxford, (2000). , [Multiparty Quantum Key Agreement based on Quantum Secret Direct Communication with GHZ states]{}, arXiv:1602.00832v1\[quant-ph\], 2016. , [Holographic Software for Quantum Networks]{}, https://www.researchgate.net/publication/301818865	{ "pile_set_name": "ArXiv" }
--- abstract: 'Using first-principles total-energy pseudopotential calculations, we have studied the properties of chains of potassium and aluminum in nanotubes. For BN tubes, there is little interaction between the metal chains and the tubes, and the conductivity of these tubes is through carriers located at the inner part of the tube. In contrast, for small radius carbon nanotubes, there are two types of interactions: charge-transfer (dominant for alkali atoms) leading to strong ionic cohesion, and hybridization (for multivalent metal atoms) resulting in a smaller cohesion. For Al-atomic chains in carbon tubes, we show that both effects contribute. New electronic properties related to these confined atomic chains of metal are analyzed.' address: ' Department of Physics, University of California at Berkeley, Berkeley, California 94720 and Materials Sciences Division, Lawrence Berkeley Laboratory, Berkeley, California 94720 ' author: - 'Angel Rubio[@add2], Yoshiyuki Miyamoto[@add1], X. Blase, Marvin L. Cohen, and Steven G. Louie' title: 'Theoretical Study of One-dimensional Chains of Metal Atoms in Nanotubes' --- = 10000 Quasi-one-dimensional metals and conducting nanowires are of great interest both from a fundamental point of view and for possible applications. Here we explore theoretical models for examining the structural and transport properties of these systems [@Garcia]. Tubule forms of graphitic carbon [@Iijima] with their expected interesting electronic and mechanical properties [@Hamada; @prl; @Ajayan; @Dujardin; @Guerret; @Heer] are ideal candidates as hosts for one dimensional metal systems. Carbon tubes have been made which are microns in length [@Iijima; @Guerret; @Heer; @Ajayan1; @Iijima1] and have diameters ranging from larger than 100$\AA $ for multi-wall tubes down to less than 10$\AA $ for single-wall tubes [@Iijima1; @Bethune]. Important effects in the conductivity as a function of length and diameter of encapsulated metallic nanowires are expected and confinement might induce new metallic phases. For small diameter tubes, the captured metal atoms inside can form an atomic linear-chain [@prl]. This would provide a new means for producing ideal one-dimensional metallic chains. The metal capillarity and doping of tubes also raise the possibility of changing the electronic structure of the tube itself by either charge-transfer or hybridization. Just as in the case of alkali doped C$_{60}$, we expect curvature effects to alter electron-phonon and superconducting pairing interactions from those of intercalated graphite [@Lorin]. At present, we know of no evidence of alkali intercalation into nanotubes experimentally which is in contrast to the results for graphitic intercalation compounds (GIC’s) [@GIC]. Most of the experiments on metal-atom intercalation have been done on multi-wall tubes with large diameter sizes (more than 100 $\AA$ diameter) [@Ajayan; @Dujardin; @Guerret; @Ajayan1; @Tanaka]. Transition metals inside the tubes were studied [@Guerret], and it was found that the formation of continuous nanowires was connected with the existence of an incomplete electronic shell in the most stable ion configuration of the metal. However, in most cases, the filling material was a crystalline metallic carbide. Lead, bismuth and manganese incorporation has also been reported [@Ajayan; @Ajayan1]. A more detailed study [@Dujardin] has proposed surface tension of the liquid phase of the metals as a key factor in determining whether capillary action (wetting) occurs. The experimental surface tension threshold found for wetting is near 190mN/m. Thus, in this picture, fillings using metals with larger surface tension require external pressure to be experimentally realizable, and it is expected that typical metals will not be drawn into tube cavities by capillary action. On the other hand, a strong ionic cohesion for K inside subnanometer-size carbon tube was theoretically predicted [@prl]. The assumed classical theory of wetting was therefore concluded to be inappropriate for this tube size, and incorporation of other metal-atoms should be possible. Based on the similarities among graphite and hexagonal boron-nitride compounds, we have predicted that BN and B$_x$C$_y$N$_z$ will form stable tubes [@Angel; @bcn]. This was recently proven experimentally by electric arc-discharge synthesis [@bn_exp] as well as by laser-driven gas-phase chemical reaction synthesis [@Willaime]. The electronic properties of this new class of nanotubes are quite different from their carbon counterpart. Namely, the BN nanotubes are stable wide band-gap semiconductors (E$_{\rm g}~\sim$ 5.5 eV) independent of helicity, diameter of the tube, or whether the tube is single wall or multi-wall. Furthermore, the bottom of the conduction band is a nearly free electron (NFE) like-state that derives its character from the weakely bound states of a BN sheet in a band-folding picture [@bn_hex]. Considering that insulators are much less polarizable than metals and semimetals, it is expected that the potential experienced by an electron in the internal vacuum region close to the BN tube surface will be less binding than that for graphite. Therefore, we expect BN to behave like an ideal [non-interacting]{} host for the metal atoms inside. We note here that other composition B-C-N tubes have also been proposed (BC$_2$N and BC$_3$) as stable [@bcn] and observed experimentally [@bcn_exp]. These systems have potential technological applications and have very interesting electronic properties which can be generally explained by rolling the corresponding planar sheets. However, we expect that these B-C-N tubes would not serve as non-interacting hosts since their band-gaps are small. The purpose of the present study is to examine two particular cases of intercalation with metal atoms of different chemical valence: K and Al atoms. We compare and contrast the results for BN tubes with those for carbon nanotubes. The known alkali GIC’s [@GIC] and the intercalation of K-atoms in hexagonal BN [@Doll] supports the possibility of incorporation of this metal into small radius nanotubes. However, there are no reports for Al GIC’s and Al intercalation into hexagonal BN to our knowledge, so no comparison can be made. We show that BN tubes behave as [non-interacting]{} confining hosts for metallic chains made of K (or other alkali atoms) and Al. This is in contrast with metal atoms inside carbon nanotubes where charge-transfer and hybridization effects are obtained. In all cases, we expect that reactions in the gas phase would favor the intercalation of these metallic atoms in both carbon and BN nanotubes. We conclude that the thin tubes react with metals similarly to the process of intercalation of the planar graphitic sheets. Total-energy pseudopotential band-structure calculations are done for the metal and tube systems within the framework of the local density approximation (LDA) using a planewave basis set with a cutoff energy of 36 Ry. The Kleinman-Bylander [ab-initio]{} pseudopotential scheme [@KB] including core-corrections for the exchange-correlation energy [@Steve] is used. The core correction is necessary to reproduce the structural properties of bulk K. The calculations were performed in a supercell geometry with a 5.5$\AA$ distance between the walls of neighboring tubes. This distance is large enough to ensure that tube-tube interactions are negligible. We use two [k]{}-points in the one dimensional irreducible Brillouin zone to get well converged total energies and electronic states. We first perform a structural minimization for the free standing linear chain of K and Al atoms. The obtained bond length of the K chain is 4.08 $\AA$ and the Al chain 2.38 $\AA$, as compared to the calculated values of 4.38 $\AA$ for bulk bcc K and 2.84 $\AA$ for bulk fcc Al. The smaller bond-length can be understood in terms of a reduction of the coordination number going from the bulk system to the linear chain [@Pauling]. In the case of a K chain, a small energy lowering (below 0.5 K) [@prl] was found by dimerization of the K atoms. In contrast, an energy lowering of the Al chain from dimerization was not found in the present calculations. These results however do not rule out a Peierls transition for K and Al chains. Since the energy gain is extremely small, the calculations require a large number of sampling points in momentum space. According to these results, we believe that Peierls distortion would not be important in practical situations. We have estimated the cohesive energy to be 0.47 eV/atom and 1.23 eV/atom for the K and Al chains respectively, as compared to the experimental values of 0.94 eV/atom and 3.34 eV/atom for bulk K and Al respectively [@Handbook]. For the Al chain, the large difference in cohesive energy with respect to the bulk metal can be related to the incomplete formation of bulk $sp$-delocalized bands in the one dimensional chain. Similar effects are expected to happen for the electronic structure when the Al chain is incorporated into a BN nanotube. In order to reduce the computational effort, we have assumed that the linear chain and the nanotube constitute commensurate phases because the change in the binding energy of the linear chain in going from the theoretical bond length to the well-matched tube lattice constant is small [@comment1]. Thus we set the metal chain bond length to be the same as the periodic distance along the tube axis in all the calculations reported here. We note that the K chain is well suited to be incorporated in $(n,0)$ single unit cell tubes or double unit cell $(n,n)$ nanotubes, while the Al chain fits well in the $(n,n)$ single unit cell nanotubes [@comment2]. The index notation for the tubes used here is the same as that given in Ref. [@Hamada]. Furthermore, the metal atoms are assumed to be linearly aligned at the tube center. This is not a severe approximation since for the diameter of the tubes studied, the metallic linear chain is in the only possible arrangement because of the geometrical restrictions for the covalent and metallic radii of the atoms [@Pauling]. In Ref. [@prl] there is a discussion of the important role of having small diameter carbon tubes (diameters close to the GIC’s interlayer distance) for maximizing the heat of formation for incorporating K atoms ($\sim $1 eV/K-atom). The heat of formation is obtained by subtracting the total energy of the doped tube from the sum of the total energies of separated systems of bulk metal and a pure tube. This subnanometer diameter tubes do not obey the classical picture of wetting and capillarity. We here note that non-intercalation of Al atoms is expected for very large diameter tubes because liquid Al has a much larger surface tension [@Handbook] than the threshold for wetting [@Dujardin]. For the intercalation of Al in a C(4,4)-tube, we find that the heat of formation is negative, in contrast to K. However, there is a gain in energy of 0.14 eV/atom comparing the energy of a tube with a linear chain inside to that with the chain separated. We thus expect intercalation of Al in the gas phase could be experimentally realizable under pressure for carbon tube bundles. The main reason for the negative heat of formation is that the coordination number for Al is not large enough to form the $sp$-metallic band of bulk Al. Hence we expect that Al is likely to form larger diameter nanowires (with a tendency to fcc coordination) when the tube diameter is larger but still in the nanometer regime. In Fig. (1) the band structure of Al inside a C(4,4) nanotube is compared with that of an undoped tube. In Fig. (2), a band structure of a free standing Al chain is shown for the comparison. Upon Al incorporation, the Fermi level is shifted upward corresponding to one electron transfer per metal atom as in the case of K-intercalation. On the other hand, some hybridization between the Al-$s$ states with the tube wall is observed. This Al-state is indicated in Fig. (1). The Al p$_{z}$-states and p$_{xy}$-states are above the Fermi-level and also hybridize with the conduction tube states. To gain insight into the physics of intercalation of K atoms into BN tubes, we first performed some calculations for the hexagonal planar phase. Hexagonal BN is an indirect wide-gap semiconductor with a NFE state which lies close to the bottom of the conduction band [@bn_hex]. This NFE state is the one that interacts the most with the K $s$ states. The position of the K-derived level, around 3 eV above the BN valence band top, is consistent with photoluminescence experiments [@Doll]. We remark that the band structure of a single BN-sheet is also a wide band gap insulator but the NFE state now becomes the bottom of the conduction band. According to our calculation, the wetting of the planar BN sheet by a K-monolayer is an exothermic reaction ($\sim$ 0.5 eV/K-atom) with a 2$\times$2 unit cell. In a manner similar to the adsorption of K layers on graphite [@Kmono], multiple-layers are not expected to be adsorbed on the BN sheet. These results support the possible intercalation of K atoms in small radius BN nanotubes. In the three cases studied here: K atoms inside BN(4,4) and BN(8,0), and Al atoms in BN(4,4) the difference in the binding energy of the two systems (energy of the tube with metal outside minus the one with metal inside) is positive but smaller than the value for Al in carbon nanotubes mainly due to the negligible charge-transfer from the chain to the BN-wall. This is an indication of weak interaction between the metal-chain and the BN-tube. Intercalation however could still be possible while growing the nanotube in a vapor phase of the metallic element. The band structures of K and Al chains in BN(4,4) tubes are depicted in Fig. (3). The NFE state interacts with the K chain $s$ states, keeping the wavefunction of the occupied carrier states located in the interior of the tube on the K chain (see a contour map of charge density for the state $\alpha$ at the bottom of Fig. (3)). This is different from K in carbon nanotubes, in which the K chain $s$ states and the carbon NFE state also interact but the corresponding wavefunctions are unoccupied and the K electrons are donated to the graphitic wall. We check that these results are independent of the diameter and helicity of the BN tube by doing the calculation for K in BN(8,0). The same picture is obtained with hybridization among the tube NFE-band and the K-$s$ derived band forming occupied states with wavefunction inside the tube. The band structures of the multiwall BN-nanotubes do not change significantly from those of the single wall tubes, hence we expect the intercalation to be the same as for the single wall tubes. The band structure for an Al chain in a BN(4,4) tube can be understood by the direct addition of the two band structures of the isolated systems (BN tube plus free Al chain). Electronic states in the gap at X are clearly seen as being derived from Al $s$ and $p_{z}$ linear chain states, (see Fig. (2)). This is different from the carbon nanotubes where the Al $p_{z}$ states are located just close to the Fermi level. The Al $s$ state at X is displayed in the bottom part of Fig. (3). Also, plotted is one of the states at $\Gamma$ derived from the interaction of the NFE-state with the Al $p_{xy}$ states ($\beta_1$ and $\beta_2$ states in the figure which have very similar charge density profiles). We note that all of these states have charge density concentrated in the inner part of the tube. Therefore, the conductivity of these doped-tubes will be controlled by carriers in the inside of the tubes whereas, in carbon nanotubes, carriers on the tube wall dominate. When considering the rigid sliding motion of the inner metal chain, we find that the change in energy as a function of position is small and comparable to the one given for K in carbon nanotubes [@prl]. In conclusion, we have shown that boron-nitride nanotubes may be good hosts for the incorporation of metal nanowires. This together with the elastic properties and high thermal conductivity of the BN-matrix suggest that these systems may have technological applications. Similar studies on carbon nanotubes show both charge-transfer and hybridization mechanisms are possible for the formation of the metal/tube system. The conductivity of the intercalated tubes will be dominated by carriers in the inner region of the tube for boron-nitride tubes and on the tube wall for carbon nanotubes. The transition between macroscopic behavior controlled by surface tension and the microscopic level incorporation studied here is still an open question. Noble-metal atoms such as Ag and Au could also be incorporated in these tubes forming continuous nanowires. Quantized conductance versus localization effect as a function of length has been reported for thin Au-wires [@Garcia]. Similar or more striking effects may be observed in a one-dimensional metal chain inside a tube. [Acknowledgements: ]{} This work was supported by National Science Foundation Grant No. DMR-9520554 and by the Director Office of Energy Research, Office of Basic Energy Sciences, Materials Sciences Division of the U.S. Department of Energy under Contact No. DE-AC03-76SF00098. The computations were done on the CRAY-C90 computer at the San Diego Supercomputer Center. X.B. gratefully acknowledges support from the France-Berkeley fund. , Departamento de Física Teórica. Universidad de Valladolid, E-47011 Valladolid, SPAIN , Fundamental Research Laboratories, NEC Corporation, 34 Miyukigaoka, Tsukuba 305, JAPAN J.I. Pascual, J. Méndez, J. Gómez-Herrero, A.M. Baró, N. Garcia, U. Ladman, W.D. Luedtke, E.N. Bogachek and H.P. Cheng, Science [267]{}, 1793 (1995). S. Iijima, Nature [354]{}, 56 (1991). N. Hamada, S. Sawada, and A. Oshiyama, Phys. Rev. Lett. [68]{}, 1579 (1992); R. Saito, M. Fujita, G. Dresselhaus, and M.S. Dresselhaus, Appl. Phys. Lett [60]{}, 2204 (1992); X. Blase, L.X. Benedict, E.L. Shirley and S.G. Louie, Phys. Rev. Lett. [72]{}, 1878 (1994). Y. Miyamoto, A. Rubio, X. Blase, M.L. Cohen and S.G. Louie, Phys. Rev. Lett. [74]{}, 2993 (1995). P.M. Ajayan and S. Iijima, Nature [361]{}, 333 (1993). E. Dujardin, T.W. Ebbessen, H. Hiura, and K. Tanigaki, Science [265]{}, 1850(1994). C. Guerret-Plécourt, Y. Lebouas, A. Loisea and H. Pascard, Nature [372]{}, 761 (1994). W.A. de Heer, W.S. Bacsa, A. Chatelain, T. Gerfin, R. Humphrey-Baker, L. Forro and D. Ugarte, Science [268]{}, 845 (1995). P.M. Ajayan, C. Colliex, J.M. Lambert, P. Bernier, L. Barbedette, M. Tencé and O. Stephan, Phys. Rev. Lett. [72]{}, 1722 (1994). S. Iijima and T. Ichihashi, Nature [363]{}, 603 (1993). D.S. Bethune, C.H. Klang, M.S. de Vrles, G. Gorman, R. Savoy, J. Vazquez, and R. Beyers, Nature [363]{}, 605 (1993). L.X. Benedict, V.H. Crespi, S.G. Louie and M.L. Cohen (preprint) see for example, M.S. Dresselhaus and G. Dresselhaus, Advance in Physics, [30]{}, 139 (1981), and references therein. K. Tanaka, T. Sato, T. Yamabe, K. Okahara, K. Uchida, M. Yumura, H. Niino, S. Ohshima, Y. Kuriki, K. Yase, and F. Ikazaki, Chem. Phys. Lett. [223]{}, 65 (1994). A. Rubio, J.L. Corkill and M.L. Cohen, Phys. Rev. B [49]{}, 5081 (1994); X. Blase, A. Rubio, S.G. Louie and M.L. Cohen, Europhys. Lett. [28]{}, 335 (1994). Y. Miyamoto, A. Rubio, S.G. Louie, and M.L. Cohen, Phys.Rev.B [50]{}, 4976 (1994); [ibid.]{} [50]{} 18360 (1994). N.G. Chopra, J. Luyken, K. Cherrey, V.H. Crespi, M.L. Cohen, S.G. Louie and A. Zettl, Science (in press). L. Boulanger, B. Andriot, M. Cauchetier and F. Willaime, Chem. Phys. Letts. [234]{}, 227 (1995) X. Blase, A. Rubio, S.G. Louie and M.L. Cohen, Phys. Rev. B [51]{}, 6868 (1995); A. Castellani, M. Posternack, A. Baldereschi, H.J.F. Jansen and A.J. Freeman, Phys.Rev.B [32]{}, 6997 (1985). Z. Weng-Sieh, K. Cherrey, N.G. Chopra, X. Blase, Y. Miyamotot, A. Rubio, M.L. Cohen, S.G. Louie, A. Zettl and R. Gronsky, Phys.Rev B [51]{}, 11229 (1995); O. Stephan, P.M. Ajayan, C. Colliex, Ph. Redlich, J.M. Lambert, P. Bernier and P. Lefin, Science [266]{}, 1683 (1994). G.L. Doll, J.S. Speck, G. Dresselhaus, M.S. Dresselhaus, K. Nakamura, and S.I. Tanuma, J. Apply. Phys. [66]{}, 2554 (1989). L. Kleinman and D.M. Bylander, Phys. Rev. Lett. [48]{}, 1425 (1982). S.G. Louie, S. Froyen, and M.L. Cohen, Phys. Rev. B [26 ]{}, 1738 (1982). L. Pauling, [The nature of the chemical bond]{} (Ithaca, New York, Cornell University Press, 1948), see chapter XI. , edited by R.C. Weast. (The Chemical Rubber Co., Cleveland, Ohio, 1972-1973.). The surface tension for liquid Al ranges from 400 to 900 mN/m depending on the temperature and surrounding gas. This change in binding is of 0.011 eV/atom and 0.04 eV/atom for K- and Al-chains, respectively. The lattice constant for $(n,0)$ tubes is 4.29 $\AA$ and 4.32 $\AA$ for carbon and BN, respectively. Meanwhile, the lattice constant for $(n,n)$ tubes is 2.46 $\AA$ for C and 2.48 $\AA$ for BN. Z.Y. Li, K.M. Hock, and P.E. Palmer, Phys. Rev. Lett. [ 67]{}, 1562 (1991).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Appreciating the classical understanding of the elementary particle the “dynamical” Poincaré algebra is developed. It is shown that the “dynamical” Poincaré algebra and the equations of motion of particles with arbitrary spin are gauge invariant and that gauge invariance and relativistic invariance stand on equal footings. A “dynamical” non-minimal interaction is constructed explicitly and the Rarita–Schwinger equation is considered in the framework of this “dynamical” interaction.' --- MZ-TH/09-30\ arXiv:0908.3761\ August 2009 [“Dynamical” interactions and gauge invariance]{} \ $^1$ Loodus- ja Tehnoloogiateaduskond, Füüsika Instituut,\ Tartu Ülikool, Riia 142, 51014 Tartu, Estonia\ $^2$ Institut für Physik der Johannes-Gutenberg-Universität,\ Staudinger Weg 7, 55099 Mainz, Germany PACS numbers: 11.15.-q, 11.30.Cp, 21.10.Hw Introduction ============ Understanding the higher-spin interactions[^1] is a longstanding problem. However, in spite of its 70 years history, the main goal – the construction of a consistent higher-spin theory, even for the electromagnetic interaction, which ought to be the simpliest case – has not been achieved yet. The investigations of higher-spin fields started in the thirties of the last century with papers by Dirac [@Dirac:1936tg], Wigner [@Wigner:1939cj], Fierz and Pauli [@Fierz:1939ix] and followed by the works of Rarita and Schwinger [@Rarita:1941mf], Bargmann and Wigner [@Bargmann:1948ck], and others [@Bhabha:1945zz; @Wild:1947zz; @HarishChandra:1947zz; @Corson:1982wi; @Pursey:1965zz; @Tung:1967zz; @Wightman:1977uc]. The difficulties in higher spin physics revealed themselves when one tried to couple higher-spin fields to an electromagnetic field. In the 1960ies concrete defects of the higher-spin interaction theory were found. Johnson and Sudarshan [@Johnson:1960vt] and Schwinger [@Schwinger:1963zz] demonstrated that in the case of minimal electromagnetic coupling some of the anticommutation relations become indefinite. It appeared that the defects were also present on the classical level. Velo and Zwanziger [@Velo:1969bt] and Shamaly and Capri [@Shamaly:1972zu] showed that in an external electromagnetic field there appeared acausal (superluminal) modes of propagation.[^2] Afterwards other defects – bad high-energy behaviour of the amplitudes, various algebraic problems [etc.]{} – were found. Since the sixties of the last century much work was done to solve the problems, but no result which one can call a breakthrough has been obtained in the framework of ordinary field theory. In case of higher-spin electromagnetic interactions investigations of the last two decades have moved in two directions. One part of community developes the theory on the ground of the minimal electromagnetic coupling, the other part searches for a consistent theory by using non-minimal couplings. The theory of higher spin interactions has never belonged to the “mainstream” theories. The field has been cultivated by groups of enthusiasts. On the other hand, the theory of higher spin interactions is needed for solving many mainstream problems. It is related to the Standard Model (SM) in several ways. By introducing the massive spin-one gauge bosons into the theory one also introduces the higher-spins problems into the Standard Model. Difficulties appear for instance in scattering processes with the charged gauge bosons $W^\pm$ in the initial or final state, or in constructing three-vertex gauge boson self-interactions. A consistent higher-spin interaction theory is also needed in chromodynamics. Quantum Chromodynamics (QCD) does not yet allow to describe low-energy hadronic processes in terms of underlying quark-gluon dynamics. Due to this one has to use a more phenomenological approach in terms of hadronic fields. However, one of the basic problems here is the treatment of hadrons with higher spins [@Pascalutsa:2003vz]. To understand better the problems of modern theories beyond the SM one also needs a better understanding of ordinary higher-spin field theory. String theory for instance is free of many higher-spin problems and due to this it is believed that it can consistently describe quantum gravity. A reason behind this consistent behaviour is that string theories contain an infinite tower of all spin states. But at the same time there exist serious troubles in the physical interpretation of the string theories. The existence of a consistent higher-spin interaction theory would help to understand better the physics behind the string theory. It is believed that if a breakthrough in understanding the basic problems of the ordinary higher-spin field theory would happen, it might become a fashionable topic [@Sorokin:2004ie]. Problems of higher-spin interactions ------------------------------------ The search for a consistent higher-spin interaction theory has been faced with various difficulties. The theory of the relativistic wave equations is based on the representations of the Poincaré group which in the field theory are somewhat specific in their mathematical realization. In addition, the theory of higher-spin fields is altogether rather complicated and due to this the wave functions and Lagrangians proposed have not always been correct. Therefore, it is difficult to understand whether the problems are of technical kind or pertained to the principles. As a matter of fact, the difficulties in higher-spin physics are generic to all field theoretical descriptions of relativistic higher-spin particles. The difficulties are related to the fact that covariant higher-spin field has more components than it is necessary to describe the spin degrees of freedom of the physical particle. To get rid of redundant degrees of freedom one must set up constraints between the field components. Using the language of Lagrangians, one have to construct a free Lagrangian which in addition to the Dirac and Proca type higher-spin equations would yield also constraint equations that reduce the number of degrees of freedom to the physical values. The problem is related to the introduction of interactions. If the interactions are introduced consistently with the free field theory, the number of independent field components remains unchanged. Otherwise the free theory constraints may be violated and unphysical degrees of freedom become involved. In order to put constraints on the field components it is reasonable to use the symmetry framework. To reduce the number of degrees of freedom of the free field to a physical value certain symmetries have to be imposed in formulating the action. Any higher-spin action has to be invariant under a transformation which leaves only the physical value of $2s+1$ degrees of freedom. Because of this, the interacting theory has to obey similar symmetry requirements as the corresponding free theory or, even better, preserves the gauge symmetries of the free theory. The possibility to construct consistent higher-spin theories with gauge invariant couplings was first pointed out by Weinberg and Witten [@Weinberg:1980kq]. However, the realization of this scenario is beset with difficulties. Even though certain progress in understanding of a higher-spin interaction theory have been made [@Sorokin:2004ie; @progress], up to now no general prescription for the construction of a consistent higher-spin field theory for any spin has been found. “Dynamical” representation of the Poincaré algebra -------------------------------------------------- In this paper we used a higher-spin electromagnetic interaction theory developed by ourselves, based on the “dynamical” representation of the Poincaré algebra as a dynamical principle which leads to a non-minimal coupling. The “dynamical” representations are built up by introducing a plane electromagnetic field into the free Poincaré algebra. The representations are constructed from the generators of the free Poincaré algebra and the external field in such a way that the new, field-dependent generators obey the commutation relations of free Poincaré algebra. Introducing the interactions in this way preserves the Poincaré symmetry of the free theory and, hopefully, also the number of degrees of freedom of the free theory. The “dynamical” theory had achieved success in constructing causal spin-$3/2$ equations [@Saar:1999ez] and for justifying the value of gyromagnetic ratio $g=2$ for any spin [@Ots:2001xn]. The paper is organized as follows. In Section 2 we give an introduction to the Lorentz–Poincaré connection. In Section 3 we show how the Poincaré group ${{\cal P}}_{1,3}$ can help to describe physical states. In Section 4 we show that an external field can be introduced consistently by employing a nonsingular transformation. Using this, in Section 5 we explicitly construct the “dynamical” interaction. Finally, in Section 6 we treat the Rarita–Schwinger equation in the framework of a “dynamical” interaction. The Poincaré group ================== Relativistic field theories are based on the invariance under the Poincaré group ${{\cal P}}_{1,3}$ (known also as inhomogeneous Lorentz group ${\cal IL}$ [@Wigner:1939cj; @Bargmann:1954gh; @Fronsdal:1959zz; @Shaw:1964zz; @Joos:1962qq; @Niederer:1974ps; @Ohnuki:1988ai; @Tung:1985na; @Kim:1986wv]) This group is obtained by combining Lorentz transformations $\Lambda$ and space-time translations $a_T$, $$(a,\Lambda)\equiv a_T\Lambda:{\hbox{\it I\kern-3pt E}}_{1,3}\ni x^\mu\to{\Lambda^\mu}_\nu x^\nu+a^\mu.$$ The group’s composition law $(a_1,\Lambda_1)(a_2,\Lambda_2)=(a_1+\Lambda_1a_2,\Lambda_1\Lambda_2)$ generates the semidirect structure of ${{\cal P}}_{1,3}$, $${{\cal P}}_{1,3}={{\cal T}}_{1,3}\odot{{\cal L}}$$ where ${{\cal T}}_{1,3}$ is the abelian group of space-time translations (i.e. the additive group ${\hbox{\it I\kern-3pt R}}^4$) and ${{\cal L}}=\{\Lambda:\det\Lambda=+1,{\Lambda^0}_0\ge 1\}$ is the proper orthochronous Lorentz group acting on the Minkowski space ${\hbox{\it I\kern-3pt E}}_{1,3}$ with metric $$\eta_{\mu\nu}=\mathop{\rm diag}(1,-1,-1,-1).$$ The condition of the metric to be invariant under Lorentz transformations $\Lambda$ takes the form $$\label{eq02} {\Lambda^\mu}_\rho\eta_{\mu\sigma}{\Lambda^\sigma}_\nu=\eta_{\rho\nu}.$$ Transformation of covariant functions ------------------------------------- Under the Lorentz transformation $\Lambda\in{{\cal L}}$ the covariant functions $\psi$ transform according to a representation $\tau(\Lambda)$ of the Lorentz group [@Wigner:1939cj; @Bargmann:1948ck; @Corson:1982wi; @Pursey:1965zz; @Tung:1967zz; @Wightman:1977uc; @Bargmann:1954gh; @Fronsdal:1959zz; @Shaw:1964zz; @Joos:1962qq; @Niederer:1974ps] where the diagram ------------------------------------ --------------------------------------- ------------------- ---------------------------------------------------- $\psi:$ $x\in{\hbox{\it I\kern-3pt E}}_{1,3}$ $\longrightarrow$ $\psi(x)$ $\llap{$\tau(\Lambda)$}\downarrow$ $\downarrow\rlap{$\Lambda$}$ $\downarrow\rlap{$T(\Lambda)$}\quad$ $\tau(\Lambda)\psi:$ $\Lambda x$ $\longrightarrow$ $(\tau(\Lambda)\psi)(\Lambda x)=T(\Lambda)\psi(x)$ ------------------------------------ --------------------------------------- ------------------- ---------------------------------------------------- is commutative, i.e.$$\label{eq01} T(\Lambda)\psi(x)=(\tau(\Lambda)\psi)(\Lambda x)\equiv\psi'(x').$$ The map $T:\Lambda\to T(\Lambda)$ is a finite-dimensional representation of ${{\cal L}}$. If we parametrize the element $\Lambda\in{{\cal L}}$ by $\Lambda(\omega)=\exp\Big(-\frac12\omega_{\mu\nu}e^{\mu\nu})$ where the Lorentz generators are given by $${(e_{\mu\nu})^\rho}_\sigma=-{\eta_\mu}^\rho\eta_{\nu\sigma}+\eta_{\mu\sigma} {\eta_\nu}^\rho$$ and $\omega^{\mu\nu}=-\omega^{\nu\mu}$ are six independent parameters, the parametrization of $T$ reads $$T(\Lambda(\omega))=\exp\Big(-\frac i2\omega_{\mu\nu}s^{\mu\nu}).$$ The Lorentz group ${{\cal L}}$ is non-compact. As a consequence, all unitary representations are infinite dimensional. In order to avoid this, we introduce the concept of $H$-unitarity (see e.g. Ref. [@Niederer:1974ps] and references therein). A finite representation $T$ is called $H$-unitary if there exists a non-singular hermitian matrix $H=H^\dagger$ so that $$\label{eq03} T^\dagger(\Lambda)H=HT^{-1}(\Lambda)\quad\Leftrightarrow\quad s_{\mu\nu}^\dagger H=Hs_{\mu\nu}.$$ Notice that a $H$-unitary metric is always indefinite, so that the inner product $\langle\ ,\ \rangle$ generated by $H$ is sesquilinear sharing the hermiticity condition $\langle\psi,\varphi\rangle=\langle\varphi,\psi\rangle^$. The most famous case of $H$-unitarity is given in the Dirac theory of spin-1/2 particles where $H=\gamma^0$. Transformation of operators --------------------------- The transformation $\tau(\Lambda)$ in Eq. (\[eq01\]) is a covariant transformation for the operator ${{\cal O}}$ [@Giovannini:1977wi; @Janner:1970zz] acting on the $\psi$-space of covariant functions[^3] if the diagramm ------------------------------------ ------------------------------ ------------------- --------------------------------------------------------------------------------------- ${{\cal O}}\psi:$ $x$ $\longrightarrow$ $({{\cal O}}\psi)(x)$ $\llap{$\tau(\Lambda)$}\downarrow$ $\downarrow\rlap{$\Lambda$}$ $\downarrow\ T(\Lambda)$ $\tau(\Lambda)({{\cal O}}\psi):$ $\Lambda x$ $\rightarrow$ $\left(\tau(\Lambda)({{\cal O}}\psi)\right)(\Lambda x)=T(\Lambda)({{\cal O}}\psi)(x)$ ------------------------------------ ------------------------------ ------------------- --------------------------------------------------------------------------------------- is commutative, i.e.$$\label{eq05} (\tau(\Lambda){{\cal O}})(\Lambda x)(\tau(\Lambda)\psi)(\Lambda x) =T(\Lambda){{\cal O}}(x)\psi(x).$$ Using Eq. (\[eq01\]) we obtain $$(\tau(\Lambda){{\cal O}})(\Lambda x)T(\Lambda)\psi(x)=T(\Lambda){{\cal O}}(x)\psi(x).$$ Notice that the covariance of the transformation embodies only the property of equivalence of reference systems. The covariant operator ${{\cal O}}$ is invariant under the transformation (\[eq01\]) if in addition $\tau(\Lambda){{\cal O}}={{\cal O}}$. As a consequence we obtain $$\label{eq06} {{\cal O}}(\Lambda x)T(\Lambda)\psi(x)=T(\Lambda){{\cal O}}(x)\psi(x)$$ or ${{\cal O}}(\Lambda x)T(\Lambda)=T(\Lambda){{\cal O}}(x)$ on the $\psi$-space. The invariance means the symmetry of the physical system and implies the conservation of currents. In particular, the symmetry transformations leave the equations of motion form-invariant. The Lie algebra --------------- While the Lorentz transformation $T(\Lambda)$ changes the wave function $\psi$ itself as well as the argument of this function (cf. Eq. (\[eq01\])), the proper Lorentz transformation $\tau(\Lambda)$ causes a change of the wave function only. On the ground of infinitesimal transformations, this change is performed by the substancial variation. Starting from an arbitrary infinitesimal coordinate transformation $\Lambda(\delta\omega):x^\mu\to x^\mu+\delta\omega^{\mu\nu}x_\nu$, the substancial variation is given by Ref. [@Corson:1982wi] $$\delta_0\psi(x)\equiv\psi'(x)-\psi(x) =-\frac{i}2\delta\omega^{\rho\sigma}M_{\rho\sigma}\psi(x)$$ where $M_{\rho\sigma}=\ell_{\rho\sigma}+s_{\rho\sigma}$, $\ell_{\rho\sigma}=i(x_\rho\partial_\sigma-x_\sigma\partial_\rho)$. The corresponding finite proper Lorentz transformation can be written as $$\tau(\Lambda(\omega))=\exp\left(-\frac{i}2\omega_{\mu\nu}M^{\mu\nu}\right),$$ and the multiplicative structure of the group generates the adjoint action $$\label{adjact} \mbox{Ad}_{\tau(\Lambda)}:M_{\mu\nu}\to\tau^{-1}(\Lambda)M_{\mu\nu} \tau(\Lambda)={\Lambda_\mu}^\rho{\Lambda_\nu}^\sigma M_{\rho\sigma}.$$ Due to Eq. (\[eq03\]) the generators $s_{\rho\sigma}$ fulfill $s_{\rho\sigma}^\dagger H=Hs_{\rho\sigma}$. They depend on the spin of the field but not on the coordinates $x_\mu$. Therefore, we have $[\ell_{\mu\nu},s_{\rho\sigma}]=0$. If a generic element of the translation group is written as $$\exp(+ia_\mu P^\mu),$$ the commutator relations of the Lie algebra are given by $$\begin{aligned} \label{eq08} [M_{\mu\nu},M_{\rho\sigma}]&=&i(\eta_{\mu\sigma}M_{\nu\rho} +\eta_{\nu\rho}M_{\mu\sigma}-\eta_{\mu\rho}M_{\nu\sigma} -\eta_{\nu\sigma}M_{\mu\rho}),\nonumber\\[7pt] [M_{\mu\nu},P_\rho]&=&i(\eta_{\nu\rho}P_\mu-\eta_{\mu\rho}P_\nu), \nonumber\\[7pt] [P_\mu,P_\nu]&=&0.\end{aligned}$$ The Casimir operators of the algebra are $P^2=P_\mu P^\mu$ and $W^2=W_\mu W^\mu$ where $$W^\mu=+\frac12\epsilon^{\mu\nu\rho\sigma}M_{\nu\rho}P_\sigma$$ is the Pauli-Lubanski pseudovector, $[P_\mu,W_\nu]=0$. In coordinate representation we have $P_\mu=i\partial_\mu$, and the finite Poincaré transformation has the form $$\tau(a,\Lambda):\psi(x)\to\left(\tau(a,\Lambda)\psi\right)(x) =T(\Lambda)\psi\left(\Lambda^{-1}(x-a)\right).$$ This relation constitutes the Lorentz–Poincaré connection [@Tung:1985na]. While the representation $T$ generally generates a reducible representation of ${{\cal P}}_{1,3}$, the spectra of the Casimir operators $P^2$ and $W^2$ determine the mass and spin content of the system. Physical states =============== The natural identification of elementary particle systems is the direct geometric transition from space-time to the system under consideration. In fact, the very definition and characterization of distinct species of elementary particles are provided by the set of inequivalent irreducible projective unitary representations of the space-time symmetry group ${{\cal P}}_{1,3}$, the Poincaré group. The physical limitations of this identification are found in the description of interacting systems and in the description of internal quantum numbers in composite systems. Space-time symmetry alone does not account for all the characteristics of today’s elementary particles. According to the conventional understanding of the particle, its physical states of definite mass and spin, labelled by the moment $p^\mu$ and the helicity $\lambda$, arise from the irreducible representation of this symmetry group. The irreducible unitary representations of the Poincaré group are characterized by the eigenvalues of the two Casimir operators $P^2$ and $W^2$ of the Lie algebra $p_{1,3}$, $$\label{eq09} P^2\|m,s\rangle=m^2\|m,s\rangle,\qquad W^2\|m,s\rangle=-m^2s(s+1)\|m,s\rangle.$$ The Pauli-Lubanski pseudovector $W$ has the properties $$\begin{aligned} W^\mu P_\mu&=&0,\nonumber\\[7pt] [W^\mu,P^\nu]&=&0,\nonumber\\[7pt] [W^\mu,M^{\rho\sigma}]&=&i(\eta^{\mu\rho}W^\sigma-\eta^{\mu\sigma}W^\rho), \nonumber\\[7pt] [W^\mu,W^\nu]&=&-i\epsilon^{\mu\nu\rho\sigma}W_\rho P_\sigma,\nonumber\\[3pt] W^2&=&-\frac12M_{\mu\nu}M^{\mu\nu}P^2 +(M_{\mu\rho}P^\rho)(M^{\mu\sigma}P_\sigma).\end{aligned}$$ The independent components of $W^\mu$ form the Lie algebra of the little group of fixed momentum $p^\mu$. For every irreducible unitary representation of the little group one can derive a corresponding irreducible induced representation of the Poincaré group labeled by $(m,s)$, i.e. by the eigenvalues of the Casimir operators in Eq. (\[eq09\]). Notice that the procedure of deriving induced representations [@Niederer:1974ps; @Mackey:1952zz] corresponds very well to the physical idea of first determining the internal degrees of freedom (the helicity) of the system and then all its possible states of motion. As a matter of fact, all physical variables (like position, momentum, etc.), quantum wave functions and fields transform as finite-dimensional representations of the Lorentz group. The reason is that interactions between fundamental particles (as irreducible representations of the Poincaré group) are most conveniently formulated in terms of field operators (i.e., finite-dimensional representations of the Lorentz group) if the general requirements like covariance, causality, etc. are to be incorporated in a consistent way. The relation between these two groups and their representations is given by the Lorentz-Poincaré connection [@Tung:1985na]. As a rule this connection is realized by the relativistic wave equations. If the relativistic wave equation transforms as a finite-dimensional representation of the Lorentz group by Eq. (\[eq01\]), it contains spins exceeding the desired physical spins. In order that the solutions of the field equation correspond to a particle with a definite spin, the equation must act like a projection operator to pick out the desired spin components, i.e. to select the corresponding irreducible representation of the Poincaré group. The wave function ----------------- The wave functions we will consider have the form $$\label{eq10} {{\cal D}}(x,\psi,\partial)\equiv(\beta^\mu\partial_\mu+i\rho)\psi(x)=0$$ where $\psi$ is an $N$-component function, $\beta^\mu$ ($\mu=0,1,2,3$), and $\rho$ are $N\times N$ matrices independent of $x$. Following Bhabha’s conception [@Bhabha:1945zz], it is “… logical to assume that the fundamental equations of the elementary particles must be first-order equations of the form (\[eq10\]) and that all properties of the particles must be derivable from these without the use of any further subsidiary conditions.” The principle of relativity states that a change of the reference frame cannot have implications for the motion of the system. This means that Eq. (\[eq10\]) is invariant under Lorentz transformations. Equivalently, the Lorentz symmetry of the system means the covariance and form-invariance of Eq. (\[eq10\]) under the transformation in Eq. (\[eq01\]), i.e. the transformed wave equation is equivalent to the old one. Therefore, we require that every solution $\psi'(x')$ of the transformed equation $${{\cal D}}'(x',\psi',\partial')=0$$ can be obtained as Lorentz transformation of the solution $\psi(x)$ of Eq. (\[eq10\]) in the original system and that the solutions in the original and transformed systems are in one-to-one correspondence. The explicit form of the covariance follows from Eq. (\[eq05\]), $$\label{eq11} (\tau(\Lambda){{\cal D}})(\Lambda x,\tau(\Lambda)\psi,\Lambda\partial) ={{\cal D}}'(x',\psi',\partial')=T(\Lambda){{\cal D}}(x,\psi,\partial)=0.$$ Due to the linearity in $\psi$ one may write $${{\cal D}}(x,\psi,\partial)\equiv{{\cal D}}(\partial)\psi(x)=0$$ to obtain the explicit Lorentz transformations $$\begin{aligned} \beta^{\prime\mu}&=&{\Lambda^\mu}_\rho T(\Lambda)\beta^\rho T^{-1}(\Lambda), \nonumber\\ \rho'&=&T(\Lambda)\rho T^{-1}(\Lambda).\nonumber\end{aligned}$$ The Lorentz invariance is given by the substitution $${{\cal D}}(\partial)\psi(x)=0\ \buildrel{{\rm Eq.\,(\ref{eq01})}}\over\longrightarrow\ {{\cal D}}(\partial)\psi'(x)=0.$$ or $$T^{-1}(\Lambda)\beta^\mu T(\Lambda)={\Lambda^\mu}_\rho\beta^\rho,\qquad T^{-1}(\Lambda)\rho T(\Lambda)=\rho.$$ The difference of the original and transformed wave equation is given by the wave equation where the wave function $\psi$ is replaced by the substantial variation $\delta_0\psi$, ${{\cal D}}(x,\delta_0\psi,\partial)=0$. As a consequence we obtain $[{{\cal D}},M^{\rho\sigma}]=0$ or $$\label{eq13} [\beta^\mu,s^{\rho\sigma}]=i(\eta^{\mu\rho}\beta^\sigma -\eta^{\mu\sigma}\beta^\rho),\qquad[\rho,s^{\rho\sigma}]=0.$$ An excellent discussion of such matrices $\beta$ can be found in Refs. [@Fierz:1939zz; @Bhabha:1945zz; @Wild:1947zz; @Corson:1982wi; @Gelfand:1963; @Naimark:1964]. The hermiticity of the representation $T$ in Eq. (\[eq03\]) implies the hermiticity of Eq. (\[eq10\]). Including a still unspecified hermitian matrix $H$ the hermiticity condition reads ${{\cal D}}(\partial)^\dagger H\buildrel!\over=({{\cal D}}(\partial)H)^\dagger =-H{{\cal D}}(-\partial)$ or $$\label{eq14} \beta^{\mu\dagger}H=H\beta^\mu,\qquad\rho H=H\rho.$$ Writing $\bar\psi=\psi^\dagger H$, one obains the adjoint equation $$\label{eq15} \bar\psi{{\cal D}}(-{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}\partial)=\bar\psi(-\beta^\mu{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial_\mu}+i\rho) =-(H{{\cal D}}(\partial)\psi)^\dagger=0.$$ Conserved currents ------------------ The physical meaning of hermiticity is the particle-antiparticle symmetry and the conservation laws in elementary particle processes. The technique of Takahashi and Umezawa [@Takahashi:1964] for deriving conservation laws directly from equations is based upon the differential operator $\Gamma_\mu({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftrightarrow}}$}}}\partial)$ with $[\partial_\mu,\Gamma_\nu]=0$ and $$\label{eq16} ({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}{\partial_\mu}+{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial_\mu})\Gamma^\mu({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftrightarrow}}$}}}\partial) ={{\cal D}}({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}\partial)-{{\cal D}}(-{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}\partial).$$ Since the general form of a linear differential equation can be written as $${{\cal D}}(\partial)=\sum_{j=0}^N\beta_{\mu_1\cdots\mu_j}\partial^{\mu_1}\cdots \partial^{\mu_j}$$ with the coefficients $\beta_{\mu_1\cdots\mu_j}$ symmetric in all indices for $j>1$, the operators $\Gamma_\mu$ become $$\label{eq17} \Gamma_\mu({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftrightarrow}}$}}}\partial)=\beta_\mu+\beta_{\mu\nu}({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}{\partial^\nu} -{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial^\nu})+\beta_{\mu\nu\rho}({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}{\partial^\nu} {\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}{\partial^\rho}-{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}{\partial^\nu}{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial^\rho} +{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial^\nu}{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}{\partial^\rho})+\ldots$$ Thanks to identity (\[eq16\]) one can derive a general current continuity equation caused by a continuous infinitesimal transformation $$\label{eq18} x\to x'=x+\delta x,\qquad\psi\to\psi'=\psi+\delta\psi.$$ If we find two functions $F_L[x]$ and $F_R[x]$ of $\psi$ and $\delta\psi$ and their derivatives with $$\label{eq19} F_L{{\cal D}}({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \rightarrow}}$}}}\partial)F_R-F_L{{\cal D}}(-{\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftarrow}}$}}}\partial)F_R=\partial_\mu\Omega^\mu$$ with $\Omega\ne F_L\Gamma_\mu F_R$, the current $$j_\mu=F_L\Gamma_\mu({\rlap{\kern-3pt\raise2pt\hbox{$^{^{\ \leftrightarrow}}$}}}\partial)F_R-\Omega_\mu$$ is conserved, $$\partial_\mu j^\mu=0.$$ Notice that this algebraic technique is very useful in the application to the case of interacting fields and discrete symmetries. Introduction of an external field ================================= The identification of elementary particle systems and irreducible representations of the Poincaré group find its physical limitations in the description of interacting systems and internal quantum numbers of composite systems. Since gauge symmetry is a fundamental concept in Quantum Electrodynamics, all physical quantities and dynamical equations of particles have to be gauge invariant. However, if gauge invariance is realized by minimal coupling, Poincaré invariance is broken at least for the theory of higher-spin fields ($s\ge 1$). The deficits occur both on the classical level (acausality and algebraic inconsistency) as well as on the quantum level (indefiniteness of antimutation relations). A lot of work has been done to solve these problems. If the problem is investigated by group theoretical methods of space-time symmetries of interacting systems, symmetries of interacting systems lead to the general covariance group in case of a charged particle moving in an external electromagnetic field. As a consequence, the group theoretical definition of an elementary particle can be extended to the case where an external field is present. Even though the Poincaré group is not a subgroup of the general covariance group [@Giovannini:1977wi; @Schrader:1972zd], this point of view is of help to solve the problem. A nonsingular transformation ---------------------------- It may be reasonable to introduce an external field directly into the Poincaré algebra which can be applied to classically understand the elementary particle. To do so one has to transform the generators of the Poincaré group to be dependent on the external field in such a way that the new, field-dependent generators obey the commutation relations (\[eq08\]). As it was proposed by Chakrabarti [@Chakrabarti:1968zz] and Beers and Nickle [@Beers:1972xt], the simplest way to build such a field dependent algebra is to introduce the external field by a nonsingular transformation $${{\cal V}}:p_{1,3}\to p_{1,3}^d={{\cal V}}p_{1,3}{{\cal V}}^{-1}=p_{1,3}+[{{\cal V}},p_{1,3}]{{\cal V}}^{-1}.$$ More explicitely, the transformed operators $$\begin{aligned} \label{eq2.1} \Pi^\mu&=&P^\mu+[{{\cal V}},P^\mu]{{\cal V}}^{-1},\nonumber\\[7pt] \xi^\mu&=&x^\mu+[{{\cal V}},x^\mu]{{\cal V}}^{-1},\nonumber\\[7pt] \sigma^{\mu\nu}&=&s^{\mu\nu}+[{{\cal V}},s^{\mu\nu}]{{\cal V}}^{-1},\nonumber\\[7pt] \mu^{\mu\nu}&=&\xi^\mu\Pi^\nu-\xi^\nu\Pi^\mu+\sigma^{\mu\nu}\end{aligned}$$ must satisfy the commutation relations of the Poincaré algebra. The concept of Lorentz covariance raises the requirement that the operator ${{\cal V}}$ has to be of Lorentz type for the generator $s^{\mu\nu}$, i.e.$$\label{eq2.2} {{\cal V}}s^{\mu\nu}{{\cal V}}^{-1}={V^\mu}_\rho{V^\nu}_\sigma s^{\rho\sigma}$$ which is a local extension of Eq. (\[adjact\]). $V=V(x,A)$ is the local Lorentz transformation depending on the external field $A$ and obeying $$\label{eq2.3} V_{\mu\rho}{V^\mu}_\sigma=V_{\rho\mu}{V_\sigma}^\mu=\eta_{\rho\sigma}.$$ If such a local Lorentz transformation exists, the problem is solved. Therefore, in the following we make the attempt to find explicit realizations of the local Lorentz transformation $V_{\mu\nu}$. There is no way to construct the Lorentz transformation $V_{\mu\nu}$ in general. However, as first shown by Taub [@Taub:1969zz], in the case of a plane-wave field we obtain $$\label{eq2.4} V_{\mu\nu}=\eta_{\mu\nu}-\frac{q}{k_P}(k_\mu A_\nu-k_\nu A_\mu) -\frac{q^2}{2k_P^2}A^2k_\mu k_\nu$$ where $q$ is the electric charge of the particle. The plane wave is characterized by its lightlike propagation vector $k_\mu$, $k^2=0$, and it polarization vector $a^\mu$ such that $$\label{eq2.5} a^2=-1,\qquad ka=0.$$ The operator $k_P\equiv k_\mu P^\mu$ commutes with any other and has a special role in the theory. For particles with nonzero mass one has $k_\mu P^\mu\ne 0$. Therefore, for the plane wave the inverse operator $1/k_P$ is well-defined for the plane-wave solution $\psi_P$ of the Klein–Gordon equation. In all other cases, $1/k_P$ is assumed to exist. We write $A_\mu(\xi)=a_\mu f(\xi)$ where the variable $\xi=k_\mu x^\mu$ can be used in place of the proper time. From Eq. (\[eq2.5\]) one obtains the conditions $$\label{eq2.6} \partial_\mu A^\mu=k_\mu\frac{dA^\mu(\xi)}{d\xi}=k_\mu A^{\mu\prime}(\xi)=0, \qquad k_\mu A^\mu=0$$ where we used $A'_\mu(\xi)=dA_\mu(\xi)/d\xi$, while the field $$F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu =k_\mu A'_\nu(\xi)-k_\nu A'_\mu(\xi)=F_{\mu\nu}(\xi)$$ satisfies $$\begin{aligned} \label{eq2.7} \partial_\mu F^{\mu\nu}&=&k_\mu F^{\mu\nu}(\xi)\ =\ 0,\nonumber\\ F_{\mu\nu}{F^\nu}_\rho&=&-k_\mu k_\rho\left(A'(\xi)\right)^2.\end{aligned}$$ It turns out that Eq. (\[eq2.4\]) can be written as $$V_{\mu\nu}=\exp\left(-\frac{q}{k_P}G\right)_{\mu\nu}$$ where $G_{\mu\nu}=k_\mu A_\nu-k_\nu A_\mu$. Note that the exponential series truncates after the second order term. In addition one obtains $$\begin{aligned} \label{eq2.8} V_{\mu\nu}&=&V_{\nu\mu}+\frac{2q}{k_P}G_{\nu\mu}\nonumber\\ V_{\mu\nu}k^\nu&=&V_{\nu\mu}k^\nu\ =\ k_\mu,\nonumber\\ \ [P_\mu,V_{\rho\sigma}]&=&-i\frac{q}{k_P}k_\mu F_{\rho\sigma} -i\frac{q^2}{k_P^2}(AA')k_\mu k_\rho k_\sigma.\end{aligned}$$ From the second equation in (\[eq2.8\]) one concludes that $V_{\mu\nu}$ is an element of the (local) little group ${{\cal L}}g(\xi)$ of the propagation vector $k_\mu$. It is easy and interesting to see that $V_{\mu\nu}$ generates a gauge transformation on $A_\mu$, $$\label{eq2.9} V_{\mu\nu}A^\nu=A_\mu+\partial_\mu\lambda_V(\xi),\qquad \lambda_V(\xi)=-\frac{q}{k_P}\int_{\xi_0}^\xi d\xi'A^2(\xi'),$$ and that the field $F_{\mu\nu}$ is invariant under this gauge transformation, $$\label{eq2.9i} {V^\mu}_\rho{V^\nu}_\sigma F^{\rho\sigma}=F^{\mu\nu}.$$ Therefore, the local Lorentz transformation $V$ is a symmetry. Notice that the local Lorentz transformation (\[eq2.4\]) has been rederived many times [@Brown:1964zz; @Kupersztych:1978dq; @Brown:1984hy] and widely exploited often in the context of its physical implications. In particular, at the classical level the solutions of the Lorentz form equation can be expressed in terms of these local transformations (\[eq2.4\]). Therefore, in the plane-wave case $V_{\mu\nu}$ plays the role of an evolution operator. According to the line of thought presented by the relations (\[eq2.2\]) and (\[eq2.3\]), the realization of (\[eq2.5\]) can be achieved by the singular transformation $${{\cal V}}={{\cal V}}_0{{\cal V}}_s$$ with $$\begin{aligned} \label{eq2.10} {{\cal V}}_0&=&\exp\Bigg\{-\int\frac{d\xi}{2k_P}(2q(AP)-q^2A^2)\Bigg\},\nonumber\\ {{\cal V}}_s&=&\exp\Bigg\{-\frac{iq}{2k_P}G_{\mu\nu}s^{\mu\nu}\Bigg\}.\end{aligned}$$ Collecting the results obtained, the generators of the interacting Poincaré algebra $p_{1,3}$ have the form $$\begin{aligned} \label{eq2.11} \Pi_\mu&=&P_\mu+k_\mu\frac{q}{2k_P}(qA^2-2AP-{F\kern-8pt/\kern2pt}),\nonumber\\ \sigma_{\mu\nu}&=&s_{\mu\nu}-\frac{q}{k_P}\Bigg(\frac{q}{2k_P}A^2 (\eta_{\mu\rho}k_\nu-\eta_{\nu\rho}k_\mu)k_\sigma+\strut\nonumber\\&&\strut +\eta_{\mu\rho}(k_\nu A_\sigma-k_\sigma A_\nu) -\eta_{\nu\rho}(k_\mu A_\sigma-k_\sigma A_\mu)+\strut\nonumber\\&&\strut -\frac{q}{k_P}(k_\mu A_\nu-k_\nu A_\mu)k_\rho A_\sigma\Bigg)s^{\rho\sigma},\\ \xi_\mu&=&x_\mu-\frac{q}{2k_P}\Big[x_\mu,\int d\xi(qA^2-2AP)-{G\kern-8pt/\kern2pt}\Big] \nonumber\end{aligned}$$ where ${F\kern-8pt/\kern2pt}\equiv F_{\mu\nu}s^{\mu\nu}$ and ${G\kern-8pt/\kern2pt}\equiv G_{\mu\nu}s^{\mu\nu}$. The transformed first Casimir operator $\Pi^2$ reads $$\label{eq2.12} \Pi^2=D^2-q{F\kern-8pt/\kern2pt}$$ where $D_\mu=P_\mu-qA_\mu$. The explicit form of the transformed Pauli–Lubanski vector $\Omega_\mu$ is $$\begin{aligned} \label{eq2.13} \Omega_\mu&=&W_\mu-\frac{q}{2k_P}\epsilon_{\mu\nu\rho\sigma} \left\{\eta^{\nu\alpha}\left(\frac{q}{2k_P}A^2k^\rho k^\beta +G^{\rho\beta}\right)-\frac{q}{2k_P}G^{\nu\rho}G^{\alpha\beta}\right\} s_{\alpha\beta}P^\sigma+\strut\nonumber\\&&\strut +\frac{q}{4k_P}\epsilon_{\mu\nu\rho\sigma}k^\sigma\eta^{\nu\alpha} \left(\eta^{\rho\beta}-\frac{2q}{k_P}G^{\rho\beta}\right)s_{\alpha\beta} \left(qA^2-2AP-{F\kern-8pt/\kern2pt}\right)\end{aligned}$$ which yields the transformed second Casimir operator $$\begin{aligned} \label{eq2.14} \Omega^2&=&-\frac12s^2D^2 +s^{\sigma\alpha}s_{\sigma\beta}D_\alpha D^\beta +\frac12qs^2{F\kern-8pt/\kern2pt}+\strut\nonumber\\&&\strut -\frac{q}{2k_P}\left\{k_\alpha s^{\alpha\sigma}{F\kern-8pt/\kern2pt})s_{\sigma\beta} +s_{\sigma\beta}(k_\alpha s^{\alpha\sigma}{F\kern-8pt/\kern2pt})\right\}D^\beta +\frac{q^2}{4k_P}(k_\alpha s^{\alpha\sigma}{F\kern-8pt/\kern2pt}) (k^\beta s_{\beta\sigma}{F\kern-8pt/\kern2pt})+\strut\nonumber\\&&\strut -\frac{iq}{2k_P}(k^\alpha{F\kern-8pt/\kern2pt}s_{\alpha\beta})D^\beta -\frac{iq}{2k_P}(k_\alpha s^{\alpha\sigma})(k^\beta s_{\beta\sigma}{F\kern-8pt/\kern2pt}').\end{aligned}$$ Gauge symmetry of the interacting Poincaré algebra -------------------------------------------------- It is now due time to study the gauge symmetry of the interacting Poincaré algebra $p_{1,3}^d$. As a consequence of the explicit form (\[eq2.10\]) the associated transformation of the evolution operator ${\cal V}(A)$ under the local gauge transformation $$\label{eq2.15} A_\mu(\xi)\rightarrow A_\mu(\xi)+\partial_\mu\lambda(\xi)$$ becomes $$\label{eq2.16} {\cal V}(A)\rightarrow{\cal V}(A+\partial\lambda) =e^{-iq\lambda}{\cal V}(A).$$ On substituting these relations into the interaction algebra one arrives at $$\label{eq2.17} p_{1,3}^d(A)\rightarrow p_{1,3}^d(A+\partial\lambda) =e^{-iq\lambda}p_{1,3}^d(A)e^{iq\lambda}.$$ In particular, one obtains $$\begin{aligned} \Pi_\mu(A+\partial\lambda)&=&\Pi_\mu(A)-k_\mu q\lambda',\nonumber\\[7pt] \Pi^2(A+\partial\lambda)&=&\Pi^2(A)-2qk_P\lambda'.\end{aligned}$$ It has to be mentioned that the evolution operator ${\cal V}(A)$ may be chosen to be $H$-unitary according to the representation $T$ in Eq. (\[eq03\]), i.e.$${\cal V}^\dagger(A)H=H{\cal V}^{-1}(A).$$ The “dynamical” interaction =========================== We have shown that in the case of a particular external electromagnetic field $A$ there exists an evolution operator ${\cal V}(A)$ which transforms the free Poincaré algebra $p_{1,3}$ into the interacting algebra $p_{1,3}^d(A)$, called the “dynamical” representation. By analogy with the free particle case one can realize this representation on the solution space of relativistically invariant equations. Expressing for the “dynamical” representation of the equation the operators explicitly in terms of free-field operators, one obtains “dynamical” interactions. Applying for instance the operator ${\cal V}$ to Eq. (\[eq10\]) one obtains $$\label{eq3.1} {\cal V}(A):(P_\mu\beta^\mu-m)\psi(x)=0\quad\rightarrow\quad (\Pi_\mu\Gamma^\mu-m)\Psi(x,A)=0$$ where $\Gamma^\mu\equiv{\cal V}(A)\beta^\mu{\cal V}^{-1}(A)$ and $$\label{eq3.2} \Psi(x,A)={\cal V}(A)\psi(x).$$ The transformed matrices $\Gamma^\mu$ satisfy the requirement of relativistic invariance with respect to the “dynamical” representation, $$\label{eq3.3} [\Gamma_\mu,\sigma_{\rho\sigma}] =i(\eta_{\mu\rho}\Gamma_\sigma-\eta_{\mu\sigma}\Gamma_\rho)$$ just as in the free-field case (\[eq13\]). On substituting (\[eq2.11\]) into (\[eq3.1\]) and (\[eq3.2\]) one arrives at $$\label{eq3.4} \Gamma_\mu=V_{\mu\rho}\beta^\rho=\beta_\mu-\frac{q}{k_P} \left(\frac{q}{2k_P}A^2k_\mu k_\rho+G_{\mu\rho}\right)\beta^\rho$$ and $$\label{eq3.5} {{\cal D}}^d(A)\Psi\equiv\left(D_\mu\beta^\mu-\frac{q}{2k_P}{/\kern-6pt k}{F\kern-8pt/\kern2pt}-m \right)\Psi=0$$ where ${/\kern-6pt k}\equiv k_\mu\beta^\mu$ and ${F\kern-8pt/\kern2pt}\equiv F_{\mu\nu}s^{\mu\nu}$. It is important to notice that for the “dynamical” interaction the “minimal” replacement $P_\mu\to D_\mu=P_\mu-qA_\mu$ is modified by $$\label{eq3.6} P_\mu\to P_\mu-qA_\mu-\frac{q}{2k_P}k_\mu{F\kern-8pt/\kern2pt}=D_\mu-\frac{q}{2k_P}k_\mu{F\kern-8pt/\kern2pt}.$$ The additional term $-qk_\mu{F\kern-8pt/\kern2pt}/2k_P$ in this non-minimal replacement is nontrivial in the higher spin cases ($s\ge 1$) while in the lower spin cases ($s=0,1/2$) it vanishes identically. In these cases the “dynamical” interaction coincides with the “minimal” coupling. This is investigated in much detail [@Saar:1999ez]. Note also that the substitution (\[eq3.6\]) does not coincide with the formal replacement $P_\mu\to\Pi_\mu$ in Eq. (\[eq2.11\]). Local gauge invariance ---------------------- Having found the “dynamical” interaction as a result of the Lorentz–Poincaré invariance (cf. Eqs. (\[eq2.1\]) and (\[eq2.2\])), it is interesting to ask about the local gauge invariance. The usual way is to use the local gauge invariance to conjecture the field equation on the ground of Lorentz covariance. As a matter of fact, in the “dynamical” model the Lorentz typeness (\[eq2.2\]) yields the gauge covariance, i.e. the diagram $$\label{eq3.7} \begin{tabular}{rlcr} $\Psi:$&$A$&$\longrightarrow$&$\Psi(A)$\\ $\llap{gauge}\downarrow\ $&$\ \downarrow\rlap{gauge}$ &&$\downarrow\rlap{gauge}$\\ $\Psi^g:$&$A+\partial\lambda$&$\rightarrow$ &$\Psi^g(A+\partial\lambda)=e^{-iq\lambda}\Psi(A)$\\ \end{tabular}$$ is commutative (here and in the following we skip the argument $x$ for $\Psi$). By virtue of Eq. (\[eq2.16\]) one has $$\Psi(A+\partial\lambda)=e^{-iq\lambda}\Psi(A)$$ which implies $\Psi^g\equiv\Psi$. We conclude that for the “dynamical” model the relativistic covariance in Eq. (\[eq01\]) and the local gauge covariance are on equal footings. According to the general idea of the invariance (\[eq06\]) it is clear that the “dynamical” interaction in Eq. (\[eq3.5\]) is gauge invariant under the transformation $A\to A+\partial\lambda$, i.e. the diagram $$\label{eq3.8} \begin{tabular}{rlcr} ${{\cal D}}^d\Psi:$&$A$&$\longrightarrow$&${{\cal D}}^d(A)\Psi(A)$\\ $\downarrow\quad$& $\ \downarrow\rlap{gauge}$&&$\downarrow\rlap{gauge}\qquad$\\ ${{\cal D}}^d\Psi:$&$A+\partial\lambda$&$\rightarrow$ &$e^{-iq\lambda}{{\cal D}}^d(A)\Psi(A)$\\ \end{tabular}$$ is commutative, $${{\cal D}}^d(A+\partial\lambda)\Psi(A+\partial\lambda) =e^{-iq\lambda}{{\cal D}}^d(A)\Psi(A).$$ It should be pointed out that the promotion of the local gauge invariance stands for the specified transformation $A\to A+\partial\lambda$ of the electromagnetic field only. The transformation $\Psi\to e^{-iq\lambda}\Psi$ is a consequence of Eq. (\[eq2.16\]). Moreover, due to Eq. (\[eq3.5\]) the gauge invariance does not determine the interaction uniquely and does not demand the minimal substitution only; rather the other principles of symmetry may be reconciled. For the physical quantities $k_\mu$ and $F_{\mu\nu}$ in the model introduced before in Eqs. (\[eq2.8\]) and (\[eq2.9i\]) the external (unquantized) field is acting on the particle without reaction of the particle on the field. The identification of the elementary particle system with the Poincaré group invariants in Eqs. (\[eq2.12\]) and (\[eq2.14\]) lead to the equations $$\begin{aligned} (P^2-m^2)\psi=0&\rightarrow&(\Pi^2-m^2)\Psi =(D^2-qF_{\mu\nu}S^{\mu\nu}-m^2)\Psi=0,\label{eq3.9}\\[7pt] \left(W^2+m^2s(s+1)\right)\psi=0&\rightarrow&\left(\Omega^2+m^2s(s+1)\right) \Psi=0.\label{eq3.10}\end{aligned}$$ These two equations must be satisfied by any field in the presence of the plane-wave field. As a consequence of Eq. (\[eq3.9\]) the gyromagnetic factor is $g=2$ and the Bargmann-Michel-Telegdi equation for the four-polarization vector of the particle takes its simplest form in the proper time frame of the particle [@Ots:2001xn]. The Rarita–Schwinger equation in the framework of a “dynamical” interaction =========================================================================== The spin-3/2 field may be described entirely in terms of the vector-bispinor $\Psi_\mu$ corresponding to the representation of the proper Lorentz group $$\label{eq4.1i} \left({\textstyle\frac12,\frac12}\right)\otimes\left({\textstyle \left(\frac12,0\right)\oplus\left(0,\frac12\right)}\right) ={\textstyle\left(1,\frac12\right)\oplus\left(\frac12,1\right) \oplus\left(\frac12,0\right)\oplus\left(0,\frac12\right)}.$$ The transformation rule according to Eq. (\[eq01\]) is $$\label{eq4.1ii} \left(\tau(\lambda)\psi\right)_\mu(p)=\Lambda_{\mu\nu}T_D(\Lambda) \psi^\nu(\Lambda^{-1}p)$$ where $T_D(\Lambda)$ is the Dirac representation of the Lorentz group. The generators of the representation are $$\begin{aligned} \label{eq4.2} s_{\mu\nu}&=&-ie_{\mu\nu}\otimes{\hbox{\rm 1\kern-3pt l}}_D+{\hbox{\rm 1\kern-3pt l}}_P\otimes s_{D\mu\nu} \ =\nonumber\\ &=&i\left(-\frac12\eta_{\mu\nu}+E_{\mu\nu}\otimes{\hbox{\rm 1\kern-3pt l}}_D -E_{\nu\mu}\otimes{\hbox{\rm 1\kern-3pt l}}_D+\frac12{\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma_\mu\gamma_\nu\right)\end{aligned}$$ where the indices $P$ and $D$ stand for the Proca and Dirac part of the direct product in Eq. (\[eq4.1i\]). Here the 16 matrices $E_{\mu\nu}$ generate the Weyl’s basis of the set of $4\times 4$ matrices, $$(E_{\mu\nu})_{\rho\sigma}=\eta_{\mu\rho}\eta_{\nu\sigma},\qquad E_{\mu\nu}E_{\rho\sigma}=\eta_{\nu\rho}E_{\mu\sigma},$$ and $e_{\mu\nu}=-E_{\mu\nu}+E_{\nu\mu}$ for the Lorentz generators of the vector representation. The $SO_3$ decomposition of the representation (\[eq4.1i\]) is $$\label{eq4.3} 2D^{(3/2)}\oplus 4D^{(1/2)}.$$ Therefore, the representation of the Poincaré group contains spins $3/2$ and $1/2$. The Pauli–Lubanski vector reads $$\label{eq4.4} W_\mu=i\epsilon_{\mu\rho\sigma\nu}\left(E^{\rho\sigma}\otimes{\hbox{\rm 1\kern-3pt l}}_D +\frac14{\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma^\rho\gamma^\sigma\right)P^\nu$$ and its root $$\label{eq4.5} W^2=-\frac{15}4P^2+P^2(E^{\mu\nu}\otimes\gamma_\mu\gamma_\nu) +P_\mu P^\nu(E^{\mu\rho}\otimes\gamma_\rho\gamma_\nu +E^{\rho\mu}\otimes\gamma_\nu\gamma_\rho).$$ Note that $$\begin{aligned} (W^2)^2&=&-\frac94P^2\Bigg\{-\frac{15}4P^2 +P^2(E^{\mu\nu}\otimes\gamma_\mu\gamma_\nu)+\strut\nonumber\\&&\strut +P_\mu P^\nu(E^{\mu\rho}\otimes\gamma_\rho\gamma_\nu) +P_\mu P^\nu(E^{\rho\mu}\otimes\gamma_\nu\gamma_\rho)+\frac58P^2\Bigg\} \ =\nonumber\\ &=&-2s^2P^2\left(W^2+\frac{s^2-1}2P^2\right)\Bigg\|_{s=3/2}\end{aligned}$$ is a pure spin-3/2 object which enables us to construct the Poincaré covariant mass ($m$) and spin ($j$) projectors ($j=3/2$, $1/2$). The free spin-3/2 particle Rarita–Schwinger equation is given as $$\begin{aligned} (P_\nu\gamma^\nu-m)\psi^\mu&=&0,\label{eq4.6i}\\[7pt] \gamma_\mu\psi^\mu&=&0.\label{eq4.6ii}\end{aligned}$$ The other constraints $$\begin{aligned} (P^2-m^2)\psi^\mu&=&0,\label{eq4.7}\\[7pt] P_\mu\psi^\mu&=&0\label{eq4.8}\end{aligned}$$ are a consequence of Eqs. (\[eq4.6i\]) and (\[eq4.6ii\]). It is interesting to note that the static condition (\[eq4.6ii\]) and the dynamic condition (\[eq4.8\]) together eliminate the spin-1/2 state completely, i.e. the equations $$(P^2-m^2)\psi^\mu=0,\qquad\gamma_\mu\psi^\mu=P_\mu\psi^\mu=0$$ with $\psi_\mu$ transforming according to Eq. (\[eq4.1ii\]) gives a theory for spin-3/2 states. Using the explicit form of $W^2$ in Eq. (\[eq4.5\]) it is easy to see that under the constraints (\[eq4.6ii\]) and (\[eq4.8\]) we obtain $$\label{eq4.9} W^2\psi=-\frac{15}4P^2\psi=-s(s+1)P^2\psi\Big\|_{s=3/2}.$$ Therefore, Eqs. (\[eq4.6i\]) and (\[eq4.6ii\]) describe indeed a single particle of mass $m$ and spin $3/2$. The “dynamical” interaction is obtained in the way described in Sec. 3. Taking into account the explicit form (\[eq4.2\]) of the generators $s_{\mu\nu}$, the transformation ${{\cal V}}$ in Eq. (\[eq2.10\]) becomes $$\begin{aligned} \label{eq4.10} {{\cal V}}_{\rm RS}&=&\exp\left(-\frac{iq}{k_P}\int(AP-\frac q2A^2)\right) ({\hbox{\rm 1\kern-3pt l}}_P\otimes{\hbox{\rm 1\kern-3pt l}}_D)\times\strut\nonumber\\&&\strut \times\left\{{\hbox{\rm 1\kern-3pt l}}_P-\frac q{k_P}\left(G_{\rho\sigma}-\frac q{2k_P} (G^2)_{\rho\sigma}\right)E^{\rho\sigma}\right\}\otimes \left\{{\hbox{\rm 1\kern-3pt l}}_D+\frac q{4k_P}G^{\rho\sigma}\gamma_\rho\gamma_\sigma\right\}.\end{aligned}$$ A straightforward calculation yields $$\begin{aligned} \label{eq4.11} P_\mu&\rightarrow&\Pi_\mu=\left(P_\mu+k_\mu\frac q{2k_P}(qA^2-2AP) \right)({\hbox{\rm 1\kern-3pt l}}_P\otimes{\hbox{\rm 1\kern-3pt l}}_D)+\strut\nonumber\\&&\strut -k_\mu\frac{iq}{k_P}F_{\rho\sigma}(E^{\rho\sigma}\otimes{\hbox{\rm 1\kern-3pt l}}_D) -k_\mu\frac{iq}{4k_P}F_{\rho\sigma}({\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma^\rho\gamma^\sigma), \nonumber\\[7pt] s_{\mu\nu}&\rightarrow&\sigma_{\mu\nu}=-i\left(\frac12\eta_{\mu\nu} +\frac q{k_P}G_{\mu\nu}\right)({\hbox{\rm 1\kern-3pt l}}_P\otimes{\hbox{\rm 1\kern-3pt l}}_D) +\strut\nonumber\\&&\strut +i\Bigg\{-\eta_{\mu\rho}\eta_{\nu\sigma} +\frac q{k_P}(\eta_{\mu\rho}G_{\nu\sigma}-\eta_{\nu\rho}G_{\mu\sigma}) +\strut\nonumber\\&&\strut -\frac{q^2}{2k_P^2}\left(\eta_{\mu\rho}(G^2)_{\nu\rho} -\eta_{\nu\rho}(G^2)_{\mu\sigma}+G_{\mu\nu}G_{\rho\sigma}\right)\Bigg\} \left(e^{\rho\sigma}\otimes{\hbox{\rm 1\kern-3pt l}}_D -\frac12{\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma^\rho\gamma^\sigma\right),\nonumber\\[7pt] W_\mu&\rightarrow&\Omega_\mu=-\frac{iq}{2k_P}\epsilon_{\mu\nu\rho\sigma} k^\nu A^\rho P^\sigma{\hbox{\rm 1\kern-3pt l}}_P\otimes{\hbox{\rm 1\kern-3pt l}}_D+\strut\nonumber\\&&\strut -\frac i2\epsilon_{\mu\nu\rho\sigma}\Bigg\{\left({\eta^\nu}_\alpha {\eta^\rho}_\beta -\frac{q}{k_P}{\eta^\nu}_\alpha{G^\rho}_\beta +\frac{q^2}{2k_P^2}G^{\nu\rho}G_{\alpha\beta} -\frac{q^2A^2}{2k_P^2}k^\rho k_\beta{\eta^\nu}_\alpha\right)P^\sigma +\strut\nonumber\\&&\strut\qquad +\frac{q}{2k_P}(qA^2-2AP)k^\sigma{\eta^\nu}_\alpha\left({\eta^\rho}_\beta +\frac{2q}{k_P}A^\rho k_\beta\right)\Bigg\} \Big\{e^{\alpha\beta}\otimes{\hbox{\rm 1\kern-3pt l}}_D -\frac12{\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma^\alpha\gamma^\beta\Big\} +\strut\nonumber\\&&\strut +\frac{q}{4k_P}\epsilon_{\mu\nu\rho\sigma}k^\sigma{\eta^\nu}_\alpha \left({\eta^\rho}_\beta+\frac{2q}{k_P}A^\rho k_\beta\right)F_{\lambda\tau} \times\strut\nonumber\\&&\strut\qquad\times \Big\{e^{\alpha\beta}e^{\lambda\tau}\otimes{\hbox{\rm 1\kern-3pt l}}_D -\frac12e^{\alpha\beta}\otimes\gamma^\lambda\gamma^\tau -\frac12e^{\lambda\tau}\otimes\gamma^\alpha\gamma^\beta +\frac14{\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma^\alpha\gamma^\beta\gamma^\lambda\gamma^\tau \Big\}.\end{aligned}$$ The two Casimir invariants of the “dynamical” Poincaré algebra are $$\label{eq4.12} P^2\ \rightarrow\ \Pi^2=D^2-2iqF^{\rho\sigma} \Big\{(E_{\rho\sigma}\otimes{\hbox{\rm 1\kern-3pt l}}_D) +\frac14({\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma_\sigma)\Big\}$$ and $$\begin{aligned} \label{eq4.13} W^2&\rightarrow&\Omega^2=\left(\frac92 +(e^{\rho\sigma}\otimes\gamma_{\rho\sigma})\right)D^2 +\strut\nonumber\\&&\strut +\frac12\left(-4(E_{\alpha\beta}\otimes{\hbox{\rm 1\kern-3pt l}}_D) +{e_\alpha}^\rho\otimes\gamma_{\rho\beta} +{e_\beta}^\rho\otimes\gamma_{\rho\alpha}\right)D^\alpha D^\beta +\strut\nonumber\\&&\strut -\frac{iq}{2k_P}k^\tau F^{\rho\sigma}\Bigg\{ -\frac32(h_{\tau\beta}\otimes\gamma_{\rho\sigma}) +(h_{\tau\rho}\otimes\gamma_{\sigma\beta})+\strut\nonumber\\&&\strut -\eta_{\sigma\beta}({h_\rho}^\alpha\otimes\gamma_{\alpha\tau}) -\frac i2\epsilon_{\rho\sigma\alpha\beta}({e_\tau}^\alpha\times\gamma^5) \Bigg\}D^\beta+\strut\nonumber\\&&\strut +iqF^{\rho\sigma}\Bigg\{-16(E_{\rho\sigma}\otimes{\hbox{\rm 1\kern-3pt l}}_D) -\frac{29}8({\hbox{\rm 1\kern-3pt l}}_P\otimes\gamma_{\rho\sigma})+\strut\nonumber\\&&\strut -6({e^\alpha}_\sigma\otimes\gamma_{\rho\alpha}) -i\epsilon_{\rho\sigma\alpha\beta}(E^{\alpha\beta}\otimes\gamma_5)\Bigg\} +\strut\nonumber\\&&\strut -\frac q{k_P}k^\alpha k^\beta F^{\prime\rho\sigma}(E_{\alpha\beta} \otimes\gamma_{\rho\sigma})\end{aligned}$$ where we used the abbreviations $\gamma_{\mu\nu}\equiv\gamma_\mu\gamma_\nu$ and $h_{\mu\nu}=E_{\mu\nu}+E_{\nu\mu}$. Applying the operator ${{\cal V}}_{\rm RS}$ to the Rarita–Schwinger equation one obtains $$\begin{aligned} \left\{(D^\mu\gamma_\mu-m)\eta_{\rho\sigma}-\frac{iq}{k_P}(k^\mu\gamma_\mu) F_{\rho\sigma}\right\}\Psi^\sigma&=&0,\label{eq4.14i}\\[7pt] \gamma_\mu\Psi^\mu&=&0\label{eq4.14ii}\end{aligned}$$ where $\Psi(x,A)={{\cal V}}(x,A)\psi(x)$. The first equation is the true equation of motion containing all derivatives $D_\mu\Psi_\sigma$. The static constraint (\[eq4.14ii\]) survives the “dynamical” interaction and eliminates all superfluous spin-1/2 components. As a consequence the other constraints are the Feynman–Gell-Mann equation $$\label{eq4.15} \left\{({\kern1pt/\kern-8pt D}^2-m^2)\eta_{\mu\rho}-2iqF_{\mu\rho}\right\}\Psi^\rho=0$$ and the kinematical constraint $$\label{eq4.16} \left\{D_\mu-\frac{iq}{4k_P}(F^{\rho\sigma}\gamma_\rho\gamma_\sigma)k_\mu \right\}\Psi^\mu=0.$$ Note that as in the free case the “dynamical” interaction is algebraically consistent. Moreover, the second order equation (\[eq4.15\]) describes the causal propagation of waves (assuming the continuity of the first order derivatives of $\Psi$). Conclusions =========== Starting from the Lorentz–Poincaré connection presented in Sec. 2, applied to physical states in Sec. 3, in this paper we constructed a non-minimal interaction by developing a field dependent algebra. For this purpose we introduced a transformation ${\cal V}$ based on a Lorentz type matrix $V$. We showed that the application of this transformation on the gauge field $A_\mu$ generates a gauge transformation. On the other hand, if a local gauge transformation is performed, the transformation ${\cal V}$ generates a phase transformation for the covariant fields. The difference to the standard procedure can be formulated as follows: In order to make quantum mechanics consistent with Maxwell equations, we usually have to impose the local gauge transformation given by the scheme $$\left. {\displaystyle A_\mu(x)\rightarrow A'_\mu(x)=A_\mu(x)+\partial_\mu\lambda(x) \atop\displaystyle\psi(x)\rightarrow\psi'(x)=e^{-iq\lambda}\psi(x)} \right\}\quad\Rightarrow\quad{{\cal D}}'_\mu\psi'=e^{-iq\lambda}{{\cal D}}_\mu\psi.$$ In the “dynamical” model we have a chain of related transformations, $$\begin{aligned} A_\mu(x)\rightarrow A'_\mu(x)&=&A_\mu(x)+\partial_\mu\lambda(x) \quad\Rightarrow\nonumber\\[7pt] \psi(x)\rightarrow\Psi(x,A+\partial\lambda)&=&e^{-iq\lambda}\Psi(x,A) \quad\Rightarrow\nonumber\\[7pt] {{\cal D}}^d(A+\partial\lambda)\Psi(x,A+\partial\lambda) &=&e^{-iq\lambda}{{\cal D}}^d(A)\psi(x,A),\end{aligned}$$ motivated by the Lorentz–Poincaré symmetry, where the “dynamical” extension ${{\cal D}}\rightarrow{{\cal D}}^d$ is a minimal extension of the minimal coupling. Applying to the generalized Dirac equation for arbitrary spin we showed explicitly how the transformation ${\cal V}$ changes the minimal interaction to a non-minimal one. We applied the formalism to the Rarita–Schwinger equation and derived explicit expressions for the field-dependent generators of the algebra. Acknowledgements {#acknowledgements .unnumbered} ---------------- The work is supported by the Estonian target financed projects No. 0182647s04 and No. 0180056s09 and by the Estonian Science Foundation under grant No. 6216. [99]{} M.A. Vasiliev, “Progress in higher spin gauge theories”, talk given at the International Conference on Quantization, Gauge Theory, and Strings: Conference Dedicated to the Memory of Professor Efim Fradkin, Moscow, Russia, 5–10 June 2000, published in [Moscow 2000, Quantization, gauge theory, and strings]{}, vol. 1, p. 452 \[arXiv:hep-th/0104246\] P.A.M. Dirac, “Relativistic wave equations”, Proc. Roy. Soc. Lond. [155A]{} (1936) 447 E.P. Wigner, “On Unitary Representations Of The Inhomogeneous Lorentz Group”, Annals Math. [40]{} (1939) 149 \[Nucl. Phys. Proc. Suppl. [6]{} (1989) 9\] M. Fierz and W. Pauli, Proc. Roy. Soc. Lond. [A173]{} (1939) 211 W. Rarita and J. Schwinger, Phys. Rev. [60]{} (1941) 61 V. Bargmann and E.P. Wigner, Proc. Nat. Acad. Sci. [34]{} (1948) 211 H.J. Bhabha, Rev. Mod. Phys. [17]{} (1945) 200 E. Wild, Proc. Roy. Soc. Lond. [A191]{} (1947) 253 Harish-Chandra, Phys. Rev. [71]{} (1947) 793 E.M. Corson, “Introduction To Tensors, Spinors, And Relativistic Wave Equations”, London and Glasgow, 1953 D.L. Pursey, Annals Phys. [32]{} (1965) 157 W.K. Tung, Phys. Rev. [156]{} (1967) 1385 A.S. Wightman, “Invariant Wave Equations: General Theory And Applications To The External Field Problem”, [SPIRES entry]{} in [Proceedings of the Ettore Majorana International School of Mathematical Physics of Erice, July 1977]{} (New York, N.Y. 1978), pp. 1–101 K. Johnson and E.C.G. Sudarshan, Annals Phys. [13]{} (1961) 126 J. Schwinger, Phys. Rev. [130]{} (1963) 800 G. Velo and D. Zwanziger,\ Phys. Rev. [186]{} (1969) 1337; Phys. Rev. [188]{} (1969) 2218 A. Shamaly and A.Z. Capri, Annals Phys. [74]{} (1972) 503 W. Cox, J. Phys. [A22]{} (1989) 1599 V. Pascalutsa, “Higher spin hadrons as relativistic fields”, invited talk at the Workshop on the Physics of Excited Nucleons (NSTAR 2002), Pittsburgh, Pennsylvania, 9–12 October 2002, published in [Pittsburgh 2002, Physics of excited nucleons]{}, p. 325 \[arXiv:nucl-th/0303005\] S. Weinberg and E. Witten, Phys. Lett. [B96]{} (1980) 59 D. Sorokin, “Introduction to the classical theory of higher spins”,\ a course of lectures given during 2003-2004, AIP Conf. Proc. [767]{} (2005) 172 S. Ferrara, M. Porrati and V.L. Telegdi, Phys. Rev. [D46]{} (1992) 3529;\ V. Pascalutsa and R.G.E. Timmermans, Phys. Rev. [C60]{} (1999) 042201;\ S. Deser, V. Pascalutsa and A. Waldron, Phys. Rev. [D62]{} (2000) 105031;\ M. Napsuciale, S. Rodríguèz, E.G. Delgado-Acosta and M. Kirchbach,\ Phys. Rev. [D77]{} (2008) 014009;\ D.P. Sorokin and M.A. Vasiliev, Nucl. Phys. [B809]{} (2009) 110 R. Saar, I. Ots, R. Tammelo and R.K. Loide, J. Phys. [A32]{} (1999) 2499 I. Ots, R. Saar, R. K. Loide and H. Liivat, Europhys. Lett. [56]{} (2001) 367 V. Bargmann, Annals Math. [59]{} (1954) 1 C. Fronsdal, Phys. Rev. [113]{} (1959) 1367 R. Shaw, Nuovo Cim. [33]{} (1964) 1074 H. Joos, Fortsch. Phys. [10]{} (1962) 65 U.H. Niederer, L. O’Raifeartaigh, Fortsch. Phys. [22]{} (1974) 111 W. J. Hurley, Phys. Rev. [D10]{} (1974) 1185 N. Giovannini, Helv. Phys. Acta [50]{} (1977) 337 A. Janner, Helv. Phys. Acta [43]{} (1970) 296 Y. Ohnuki, “Unitary Representations of the Poincaré Group and Relativistic Wave Equations”, [Singapore, Singapore: World Scientific (1988) 213p]{} W.K. Tung, “Group Theory In Physics”,\ [Singapore, Singapore: World Scientific (1985) 344p]{} Y.S. Kim and M.E. Noz, “Theory and Applications of the Poincaré Group”,\ [Dordrecht, Netherlands: Reidel Publ. Company (1986) 331p]{} G.W. Mackey, Annals Math. [55]{} (1952) 101;\ Annals Math. [58]{} (1953) 193; Acta Math. [99]{} (1958) 265 M. Fierz, Helv. Phys. Acta [12]{} (1939) 3 I.M. Gel’fand, R.A. Minlos and Z.Y. Shapiro, “Representations of Rotation and Lorentz Groups and Their Applications”, Pergamon Press (1963) M.A. Naimark, “Linear Representations of the Lorentz Group”,\ Pergamon Press (1964) Y. Takahashi, H. Umezawa, Nucl. Phys. [51]{} (1964) 193 R. Schrader, Fortsch. Phys. [20]{} (1972) 701 A. Chakrabarti, Nuovo Cim. [A56]{} (1968) 604 B. Beers and H.H. Nickle, J. Math. Phys. [13]{} (1972) 1592 A.H. Taub, Rev. Mod. Phys. [21]{} (1949) 388 L.S. Brown and T.W.B. Kibble, Phys. Rev. [133]{} (1964) A705 J. Kupersztych, Phys. Rev. [D17]{} (1978) 629 R.W. Brown and K.L. Kowalski, Phys. Rev. [D30*]{} (1984) 2602 [^1]: In this paper states with spin one and higher are considered as higher-spin states. This concept is not universally accepted. For a part of the investigators “higher spin” means $s\geq 3/2$. The specialists in supergravity updated the convention of the higher spin to be even $s\geq 5/2$ [@Vasiliev:2001ur]. Nevertheless, at least in the Standard Model the troubles start already from the value $s=1$. Therefore it seems that the convention $s\geq 1$ as the higher-spin region is more justified than the other ones. [^2]: Earnestly, as shown by Cox [@Cox:1989zz] the constraint analysis leading to these acausal pathologies is incomplete. On the contrary, in the complete constraint analysis a new tier of constraints occurs for the critical external field values, reducing the pathology to the field-induced change of the degrees of freedom. Because of these field-dependent constraints the analysis of acausal models is very complicated. [^3]: We have to impose the action on covariant functions because in case of higher spins the relations between operators we obtain are valid only as weak conditions.	{ "pile_set_name": "ArXiv" }
--- address: - 'Osaka University, Osaka, Japan' - 'The University of Tokyo, Tokyo, Japan' - 'Nara Institute of Science and Technology, Nara, Japan' - RIKEN Center for Advanced Intelligence Project author: - Koki Kishimoto - Katsuhiko Hayashi - Genki Akai - Masashi Shimbo bibliography: - 'ref.bib' title: Binarized Canonical Polyadic Decomposition for Knowledge Graph Completion --- Knowledge graph completion ,Tensor factorization ,Model compression Introduction {#intro} ============ Related Work {#sec:related} ============ Notation and Preliminaries {#sec:notation} ========================== Tensor Decomposition for Knowledge Graphs {#sec:tensor} ========================================= Proposed Method {#sec:proposed} =============== Experiments {#sec:exp} =========== Conclusion {#sec:conclusion} ========== Proof of Theorem \[thm:expressiveness\] {#sec:proof} =======================================	{ "pile_set_name": "ArXiv" }
--- abstract: \| The Restricted Isometry Property (RIP) is a fundamental property of a matrix enabling sparse recovery [@CRT06]. Informally, an $m \times n$ matrix satisfies RIP of order $k$ in the $\ell_p$ norm if $\\|Ax\\|_p \approx \\|x\\|_p$ for any vector $x$ that is $k$-sparse, i.e., that has at most $k$ non-zeros. The minimal number of rows $m$ necessary for the property to hold has been extensively investigated, and tight bounds are known. Motivated by signal processing models, a recent work of Baraniuk et al [@BCDH10] has generalized this notion to the case where the support of $x$ must belong to a given [model]{}, i.e., a given family of supports. This more general notion is much less understood, especially for norms other than $\ell_2$. In this paper we present tight bounds for the model-based RIP property in the $\ell_1$ norm. Our bounds hold for the two most frequently investigated models: tree-sparsity and block-sparsity. We also show implications of our results to sparse recovery problems. author: - 'Piotr Indyk[^1]' - 'Ilya Razenshteyn[^2]' bibliography: - '../../../bibtex/ir.bib' title: 'On Model-Based RIP-$1$ Matrices' --- Introduction {#intro} ============ In recent years, a new “linear” approach for obtaining a succinct approximate representation of $n$-dimensional vectors (or signals) has been discovered. For any signal $x$, the representation is equal to $Ax$, where $A$ is an $m \times n$ matrix, or possibly a random variable chosen from some distribution over such matrices. The vector $Ax$ is often referred to as the [measurement vector]{} or [ linear sketch]{} of $x$. Although $m$ is typically much smaller than $n$, the sketch $Ax$ often contains plenty of useful information about the signal $x$. A particularly useful and well-studied problem is that of [stable sparse recovery]{}. We say that a vector $x'$ is $k$-sparse if it has at most $k$ non-zero coordinates. The sparse recovery problem is typically defined as follows: for some norm parameters $p$ and $q$ and an approximation factor $C>0$, given $Ax$, recover an “approximation” vector $x^$ such that $$\label{e:lplq} {\left \lVertx-x^\right \rVert_{p}} \le C \min_{k\text{-sparse } x'} {\left \lVertx-x'\right \rVert_{q}}$$ (this inequality is often referred to as [$\ell_p/\ell_q$ guarantee]{}). Sparse recovery has a tremendous number of applications in areas such as compressive sensing of signals [@CRT06; @D06], genetic data acquisition and analysis and data stream algorithms [@M05; @GI10]. It is known [@CRT06] that there exist matrices $A$ and associated recovery algorithms that produce approximations $x^$ satisfying Equation with $p=q=1$[^3] constant approximation factor $C$, and sketch length $$\label{e:m} m=O(k \log (n/k))$$ This result was proved by showing that there exist matrices $A$ with $m=O(k \log (n/k))$ rows that satisfy the [Restricted Isometry Property (RIP)]{}. Formally, we say that $A$ is a $(k, {\varepsilon})$-RIP-$p$ matrix, if for every $x \in {\mathbb{R}}^n$ with at most $k$ non-zero coordinates we have $$(1 - {\varepsilon}) \\|x\\|_p \leq \\|Ax\\|_p \leq (1 + {\varepsilon}) \\|x\\|_p.$$ The proof of [@CRT06] proceeds by showing that (i) there exist matrices with $m=O(k \log (n/k))$ rows that satisfy $(k, {\varepsilon})$-RIP-$2$ for some constant ${\varepsilon}>0$ and (ii) for such matrices there exist a polynomial time recovery algorithm that given $Ax$ produces $x^$ satisfying Equation \[e:lplq\]. Similar results were obtained for RIP-$1$ matrices [@BGIKS08]. The latter matrices are closely connected to hashing-based streaming algorithms for heavy-hitter problems, see [@GI10] for an overview. It is known that the bound on the number of measurements in Equation is asymptotically optimal for some constant $C$ and $p=q=1$, see [@DIPW10] and [@FPRU10] (building on [@D06; @GG84; @G84; @K77]). The necessity of the “extra” logarithmic factor multiplying $k$ is quite unfortunate: the sketch length determines the “compression rate”, and for large $n$ any logarithmic factor can worsen that rate tenfold. Fortunately, a more careful modeling offers a way to overcome the aforementioned limitation. In particular, after decades of research in signal modeling, signal processing researchers know that not all supports (i.e., sets of non-zero coordinates) are equally common . For example, if a signal is a function of time, large coefficients of the signal tend to occur consecutively. This phenomenon can be exploited by searching for the best $k$-sparse approximation $x^$ whose support belongs to a given “model" family of supports ${\mathcal{M}}_k$ (i.e., $x^$ is ${\mathcal{M}}_k$-[sparse]{}). Formally, we seek $x^$ such that $$\label{e:model} \\|x - x^\\|_p \leq C \cdot \min_{\begin{smallmatrix}\operatorname{supp}(x') \subseteq T \\ T \in {\mathcal{M}}_k\end{smallmatrix}} \\|x - x'\\|_q$$ for some family ${\mathcal{M}}_k$ of $k$-subsets of $[n]$. Clearly, the original $k$-sparse recovery problem corresponds to the case, when ${\mathcal{M}}_k$ is a family of all $k$-subsets of $[n]$. A prototypical example of a sparsity model is [block sparsity]{} [@EM09]. Here the signal is divided into blocks of size $b$, and the non-zero coefficients belong to at most $k/b$ blocks. This model is particularly useful for bursty time signals, where the “activity” occurs during a limited time period, and is therefore contained in a few blocks. Another example is [tree sparsity]{} [@RCB01] which models the structure of wavelet coefficients. Here the non-zero coefficients form a rooted subtree in a full binary tree defined over the coordinates.[^4]. For many such scenarios the size of the family ${\mathcal{M}}_k$ is much smaller than ${n \choose k}$, which in principle makes it possible to recover an approximation from fewer measurements. An elegant and very general model-based sparse recovery scheme was recently provided in a seminal work of Baraniuk et al [@BCDH10]. The scheme has the property that, for any “computationally tractable” family of supports of “small" size, it guarantees a near-optimal sketch length $m=O(k)$, i.e., without any logarithmic factors. This is achieved by showing the existence of matrices $A$ satisfying the [model-based]{} variant of RIP. Formally, we say that $A$ satisfies ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$p$ if $$\label{e:modelrip} (1 - {\varepsilon})\\|x\\|_p \leq \\|Ax\\|_p \leq (1 + {\varepsilon})\\|x\\|_p$$ for any ${\mathcal{M}}_k$-sparse vector $x \in {\mathbb{R}}^n$. In [@BCDH10] it was shown that there exist matrices with $m=O(k)$ rows that satisfy ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$2$ as long as (i) either ${\mathcal{M}}_k$ is the block-sparse model and $b=\Omega(\log n)$ or (ii) ${\mathcal{M}}_k$ is the tree-sparse model. This property can be then used to give an efficient algorithm that, given $Ax$, finds $x^$ satisfying a variant of the guarantee of Equation \[e:model\]. However, the guarantees offered in [@BCDH10], when phrased in the $\ell_1/\ell_1$ framework, results in a super-constant approximation factor $C=\Theta(\sqrt{\log n})$ [@IP11]. The question of whether this bound can be improved has attracted considerable attention in signal processing and streaming communities. In particular, one of the problems[^5] listed in the Bertinoro workshop open problem list [@B11] asks whether there exist matrices $A$ with $m=O(k)$ rows that provide the $\ell_1/\ell_1$ guarantee for the tree-sparse model with some constant approximation factor $C$. #### Our results In this paper we make a substantial progress on this question. In particular: 1. For both block-sparse and tree-sparse models, we show that there exist $m\times n$ matrices $A$ that provide the $\ell_1/\ell_1$ guarantee for some constant approximation factor $C$, such that the number of measurements improves over the bound of Equation \[e:m\] for a wide range of parameters $k$ and $b$. In particular we show that for the block-sparse model we can achieve $m=O(k \cdot (1 + \log_b (n / k)))$. This improves over the $O(k \log(n/k))$ bound of Equation \[e:m\] for any $b = \omega(1)$. In particular, if $b=n^{\Omega(1)}$, we obtain $m=O(k)$. For the tree-sparse model we achieve $m=O(k \log(n/k)/\log \log (n/k))$ as long as $k=\omega(\log n)$. This also improves over the $O(k \log(n/k))$ bound of Equation \[e:m\]. We note, however, that our results are not accompanied by efficient recovery algorithms. Instead, we show the existence of model-based RIP-1 matrices with the given number of rows. This implies that $Ax$ contains enough information to recover the desired approximation $x^$ (see Appendix \[s:imp\] for more details). 2. We complement the aforementioned results by showing that the measurement bounds achievable for a matrix satisfying block-sparse or tree-sparse RIP-1 property cannot be improved (i.e., our upper bounds are tight). This provides strong evidence that the number of measurements required for sparse recovery itself cannot be $O(k)$. Our results show a significant difference between the model-based RIP-$1$ and RIP-$2$ matrices. For the $\ell_2$ norm, the original paper [@BCDH10] shows that the number of measurements is fully determined by the cardinality of the model. Specifically, their proof proceeds by applying the union bound over all elements of ${\mathcal{M}}_k$ on top of the Johnson–Lindenstrauss-type concentration inequality. This leads to a measurement bound of $m = O(k + \log \|{\mathcal{M}}_k\|)$, which is $O(k)$ for the tree-sparse or block-sparse models. In contrast, in case of the $\ell_1$ norm we obtain an upper bound with a very different dependence on $\|{\mathcal{M}}_k\|$: namely, if $\|{\mathcal{M}}_k\| \geq n / k$, then we are able to achieve $$m = O\left(\frac{k \cdot \log \frac{n \log(n / k)}{\log \|{\mathcal{M}}_k\|}}{\log \frac{k \log (n / k)}{\log \|{\mathcal{M}}_k\|}}\right).$$ Moreover, our lower bounds show that this bound is tight for the block- and the tree-sparse cases. #### Our techniques Our lower bounds are obtained by relating RIP-1 matrices to novel combinatorial/geometric structures we call [generalized expanders]{}. Specifically, it is known [@BGIKS08] that any [binary]{} 0-1 matrix $A$ that satisfies $(k, {\varepsilon})$-RIP-$1$ is an adjacency matrix of an unbalanced $(k, {\varepsilon})$-expander (see Section \[s:def\] for the formal definition). The notion of a generalized expander can be viewed as extending the notion of expansion to matrices that are not binary. Formally, we define it as follows. \[def\_gen\_expander\] Let $A$ be an $m \times n$ real matrix. We say that $A$ is a generalized $(k, {\varepsilon})$-expander, if all $A$’s columns have $\ell_1$-norm at most $1 + {\varepsilon}$, and for every $S \subseteq [n]$ with $\|S\| \leq k$ we have $$\sum_{i \in [m]} \max_{j \in S} \|a_{ij}\| \geq \|S\| \cdot (1 - {\varepsilon}).$$ Observe that the notion coincides with the standard notion of expansion for binary 0-1 matrices (after a proper scaling). In this paper we show that any (not necessarily binary) RIP-1 matrix is also a generalized expander. We then use this fact to show that any RIP-1 matrix can be sparsified by replacing most of its entries by 0. This in turn lets us use counting arguments to lower bound the number of rows of such matrix. Our upper bounds are obtained by constructing low-degree expander-like graphs. However, we only require that the expansion holds for the sets from the given model ${\mathcal{M}}_k$. This allows us to reduce the number of the right nodes of the graph, which corresponds to reducing the number of rows in its adjacency matrix. Definitions {#s:def} =========== In this section we provide the definitions we will use throughout the text. \[def\_expander\] Let $G = (U, V, E)$ with $\|U\| = n$, $\|V\| = m$, $E \subseteq U \times V$ be a bipartite graph such that all vertices from $U$ have the same degree $d$. Then we say that $G$ is a $(k, {\varepsilon})$-expander, if for every $S \subseteq U$ with $\|S\| \leq k$ we have $$\|{\left\{v \in V \mid \exists u \in S \; (u, v) \in E\right\}}\| \geq (1 - {\varepsilon}) d \|S\|.$$ Let us call any non-empty subset $${\mathcal{M}}_k \subseteq \Sigma_k = {\left\{A \subseteq [n] \mid \|A\| = k\right\}}$$ a model. In particular, $\Sigma_k$ is a model as well. Suppose that $b, k \in [n]$. Moreover, $b$ divides both $k$ and $n$. Let us partition our universe $[n]$ into $n / b$ disjoint blocks $B_1$, $B_2$, …, $B_{n/b}$ of size $b$. We consider the following block-sparse model: ${\mathcal{B}}_{k, b}$ consists of all unions of $k / b$ blocks. Suppose that $k \in [n]$ and $n = 2^{h+1} - 1$, where $h$ is a non-negative integer. Let us identify the elements of $[n]$ with the vertices of a full binary tree of depth $h$. Then, tree-sparse model ${\mathcal{T}}_k$ consists of all subtrees of size $k$ that contain the root of the full binary tree. Let ${\mathcal{M}}_k \subseteq \Sigma_k$ be any model. We say that a set $S \subseteq [n]$ is ${\mathcal{M}}_k$-sparse, if $S$ lies within a set from ${\mathcal{M}}_k$. Moreover, let us call a vector $x \in {\mathbb{R}}^n$ ${\mathcal{M}}_k$-sparse, if its support is a ${\mathcal{M}}_k$-sparse set. It is straightforward to generalize the notions of RIP-$p$ matrix, expanders and generalized expanders to the case of ${\mathcal{M}}_k$-sparse vectors and sets. Let us call the corresponding objects ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$p$ matrix, ${\varepsilon}$-${\mathcal{M}}_k$-expander and generalized ${\varepsilon}$-${\mathcal{M}}_k$-expander, respectively. Clearly, the initial definitions correspond to the case of $\Sigma_k$-sparse vectors and sets. Our two main objects of interest are ${\mathcal{B}}_{k, b}$- and ${\mathcal{T}}_k$-RIP-$1$ matrices. Sparsification of RIP-$1$ matrices ================================== In this section we show that any $n \times m$ matrix, which is $(k, {\varepsilon})$-RIP-$1$, can be sparsified after removing $(1 - \Omega(1)) n$ columns (Theorem \[sparsification\]). Then we state an obvious generalization of this fact (Theorem \[model\_sparsification\]), which will be useful for proving lower bounds on the number of rows for ${\mathcal{B}}_{k, b}$- and ${\mathcal{T}}_k$-RIP-$1$ matrices. \[sparsification\] Let $A$ be any $m \times n$ matrix, which is $(k, {\varepsilon})$-RIP-$1$. Then there exists an $m \times \Omega(n)$ matrix $B$ which is $(k, O({\varepsilon}))$-RIP-$1$, has at most $O(m / k)$ non-zero entries per column and can be obtained from $A$ by removing some columns and then setting some entries to zero. We prove this theorem via the sequence of lemmas. First we prove that for every matrix $A$ there exists a $\pm 1$-vector $x$ such that $\\|Ax\\|_1$ is small. Let $A$ be any $m \times k$ matrix. Then there exists a vector $x \in {\left\{-1, 1\right\}}^k$ such that $$\label{l1_l2} \\|Ax\\|_1 \leq \sum_{i \in [m]} \left(\sum_{j \in [k]} a_{ij}^2\right)^{1/2}.$$ Let us use a probabilistic argument. Namely, let us sample all coordinates $x_i$ independently and uniformly at random from $\{-1,1\}$. Then $$\begin{gathered} {\mathrm{E}_{}\left[\\|Ax\\|_1\right]} = \sum_{i \in [m]} {\mathrm{E}_{}\left[\left\|\sum_{j \in [k]} a_{ij} x_j\right\|\right]} \leq \sum_{i \in [m]} \left({\mathrm{E}_{}\left[\left(\sum_{j \in [k]} a_{ij} x_j\right)^2\right]}\right)^{1/2} =\\= \sum_{i \in [m]} \left({\mathrm{E}_{}\left[\sum_{j \in [k]} a_{ij}^2 x_j^2\right]}\right)^{1/2} = \sum_{i \in [m]} \left(\sum_{j \in [k]} a_{ij}^2\right)^{1/2}. \end{gathered}$$ Thus, there exists a vector $x \in {\left\{-1, 1\right\}}^k$ that satisfies (\[l1\_l2\]). As a trivial corollary we have the following statement. \[l1\_isometry\_l2\] Let $A$ be any $m \times k$ matrix that preserves (up to $1 \pm {\varepsilon}$) $\ell_1$-norms of all vectors. Then $$\sum_{i \in [m]} \left(\sum_{j \in [k]} a_{ij}^2\right)^{1/2} \geq (1 - {\varepsilon}) k.$$ The next lemma shows that every $(k, {\varepsilon})$-RIP-$1$ matrix is a generalized $(k, O({\varepsilon}))$-expander. This is a generalization of a theorem from [@BGIKS08]. \[RIP\_expander\] Let $A$ be any $m \times n$ matrix, which is $(k, {\varepsilon})$-RIP-$1$. Then, $A$ is a generalized $(k, 3 {\varepsilon})$-expander. For the proof we need the following lemma. \[l1\_l2\_linfty\] For any $y \in {\mathbb{R}}^k$ $$\label{l1_l2_linfty_eq} \\|y\\|_1 - \\|y\\|_{\infty} \leq \left(1 + \frac{1}{\sqrt{2}}\right)(\\|y\\|_1 - \\|y\\|_{2}).$$ Clearly, if $y = 0$, then the desired inequality is trivial. Otherwise, by homogenity we can assume that $\\|y\\|_1 = 1$. If $\\|y\\|_{\infty} = 1$, then $\\|y\\|_2 = 1$, and both sides of (\[l1\_l2\_linfty\_eq\]) are equal to zero. So, we can assume that $\\|y\\|_{\infty} < 1$. Suppose that $\\|y\\|_{\infty} = t$ for some $t \in (0; 1)$. If $1 / n > t \geq 1 / (n + 1)$ (thus, $n = \lceil 1/t - 1 \rceil$) for some positive integer $n$, then, clearly, $\\|y\\|_2 \leq \sqrt{nt^2 + (1 - nt)^2}$. One can check using elementary analysis that for every $t \in (0; 1)$ $$\frac{1 - \\|y\\|_{\infty}}{1 - \\|y\\|_2} \leq \frac{1 - t}{1 - \sqrt{\left\lceil \frac{1}{t} - 1\right\rceil t^2 + \left(1 - \left \lceil \frac{1}{t} - 1\right \rceil t\right)^2 }} \leq 1 + \frac{1}{\sqrt{2}}$$ (equality is attained on $t = 1/2$). This concludes the proof. Let $S \subseteq [n]$ be any subset of size at most $k$. For any $i \in [m]$ let us denote $y_i = (a_{ij})_{j \in S} \in {\mathbb{R}}^S$. We have $$\begin{aligned} \sum_{i \in [m]} \\|y_i\\|_{\infty} & \geq & \left(1 + \frac{1}{\sqrt{2}} \right) \sum_{i \in [m]} \\|y_i\\|_2 - \frac{1}{\sqrt{2}} \cdot \sum_{i \in [m]} \\|y_i\\|_1 \mbox{\ \ \ \ (by Lemma~\ref{l1_l2_linfty})} \\ & \geq & \left(1 + \frac{1}{\sqrt{2}}\right) (1 - {\varepsilon})\|S\| - \frac{1}{\sqrt{2}} \cdot (1 + {\varepsilon}) \|S\| \mbox{\ \ \ \ (by Corollary~\ref{l1_isometry_l2} and RIP-$1$)} \\ & = & (1 - (1 + \sqrt{2}) {\varepsilon})\|S\|. \end{aligned}$$ So, $A$ is a generalized $(k, (1 + \sqrt{2}) {\varepsilon})$-expander. Since $1 + \sqrt{2} < 3$, this concludes the proof. Finally, we prove Theorem \[sparsification\]. By Lemma \[RIP\_expander\] $A$ is a generalized $(k, 3 {\varepsilon})$-expander. Let us partition $[n]$ into $n / k$ disjoint sets of size $k$ arbitrarily: $[n] = S_1 \cup S_2 \cup \ldots \cup S_{n / k}$. Now for every $i \in [m]$ and every $S_t$ let us zero out all the entries $a_{ij}$ for $j \in S_t$ except one with the largest absolute value. Let $A'$ be the resulting matrix. Since $A$ is a generalized $(k, 3 {\varepsilon})$-expander, we know that the (vector) $\ell_1$ norm of the difference $A-A'$ is at most $3 {\varepsilon}n$. Thus, each column of $A-A'$ has the $\ell_1$ norm of at most $3 {\varepsilon}$ on the average. The number of non-zero entries in $A'$is at most $mn / k$, so a column has at most $m / k$ non-zero entries on the average. Thus, by Markov inequality there is a set of $n / 3$ columns such that we have moved each of them by at most $9 {\varepsilon}$ and each of them contains at most $3m / k$ non-zero entries. We define a matrix $B$ that consists of these columns. Since we have modified each of these columns by at most $9 {\varepsilon}$ and $A$ is $(k, {\varepsilon})$-RIP-$1$ we have that $B$ is $(k, 10 {\varepsilon})$-RIP-$1$. The following theorem is a straightforward generalization of Theorem \[sparsification\]. It can be proved via literally the same argument. \[model\_sparsification\] Suppose that a model ${\mathcal{M}}_k \subseteq \Sigma_k$ has the following properties: - for some $l \leq k$ all sets from $\Sigma_l$ are ${\mathcal{M}}_k$-sparse; - there exists a partition of an $\Omega(1)$-fraction of $[n]$ into disjoint subsets of size $\Omega(k)$ such that each of these subsets is ${\mathcal{M}}_k$-sparse. Then if $A$ is an $m \times n$ matrix which is ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$1$ for some sufficiently small ${\varepsilon}>0$, there exists an $m \times \Omega(n)$ matrix $B$ which is $(l, O({\varepsilon}))$-RIP-$1$, has at most $O(m / k)$ non-zero entries per column and can be obtained from $A$ by removing some columns and then setting some entries to zero. Lower bounds for model-based RIP-$1$ matrices ============================================= In this section we prove lower bounds on the number of rows for ${\mathcal{B}}_{k,b}$- and ${\mathcal{T}}_k$-RIP-$1$ matrices. This is done using the following general theorem. \[general\_lower\_bound\] If a model ${\mathcal{M}}_k \subseteq \Sigma_k$ satisfies the statement of Theorem \[model\_sparsification\] and $A$ is an $m \times n$ matrix which is ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$1$ for some sufficiently small ${\varepsilon}>0$, then $$m = \Omega\left(k \cdot \frac{\log(n / k)}{\log(k / l)}\right).$$ The proof is a combination of Theorem \[model\_sparsification\] and a counting argument similar to one used in [@N10]. First, we need the following standard geometric fact. \[volume\_bound\] Let $v_1, v_2, \ldots, v_n \in {\mathbb{R}}^d$ be a set of $d$-dimensional vectors such that - for every $i \in [n]$ we have $\\|v_i\\|_1 \leq 1.1$; - for every $i \ne j \in [n]$ we have $\\|v_i - v_j\\|_1 \geq 0.9$. Then, $n \leq 4^d$. Denote $B(x, r)$ the ball in $\ell_1$-metric with center $x$ and radius $r$. Consider the balls $B_i = B(v_i, 0.45)$. On the one hand, these balls are disjoint, on the other hand, they lie within $B(0, 1.55)$. Thus, if we consider the balls’ volumes we see that $$n \leq \left(\frac{1.55}{0.45}\right)^d < 4^d.$$ The next theorem shows a tradeoff between $m$ and column sparsity for any RIP-$1$ matrix. Its variant was proved in [@N10], but we present here the proof for the sake of completeness. \[sparsity\_m\_tradeoff\] Let $A$ be an $m \times n$ matrix, which is $(k, {\varepsilon})$-RIP-$1$ for some sufficiently small ${\varepsilon}>0$. Moreover, suppose that every column of $A$ has at most $s$ non-zero entries. Then $$s \log \left(\frac{m}{sk}\right) = \Omega\left(\log\left(\frac{n}{k}\right)\right).$$ We need a lemma from [@N10], which is proved by a standard probabilistic argument. There exists a set $X \subseteq {\mathbb{R}}^n$ of $k / 2$-sparse vectors such that - $\log \|X\| = \Omega(k \log(n / k))$; - every vector from $X$ has a unit $\ell_1$-norm; - all pairwise $\ell_1$-distances between the elements of $X$ are at least $1$. Now let us see how $A$ acts on the elements of $X$. Clearly, for every $x \in X$ the vector $Ax$ is $sk$-sparse. By pigeonhole principle we have that for some $S \subseteq [m]$ with $\|S\| \leq sk$ there exists a subset $X' \subseteq X$ with $$\label{php} \|X'\| \geq \frac{\|X\|}{\binom{m}{sk}}$$ such that for every $x \in X'$ the support of $Ax$ lies within $S$. On the other hand, since $A$ is $(k, {\varepsilon})$-RIP-$1$ one can easily see that the set ${\left\{Ax\right\}}_{x \in X'}$ (which lies in the $sk$-dimensional subspace) has the following properties: - every vector from the set has $\ell_1$-norm at most $1 + {\varepsilon}$; - all pairwise distances are at least $1 - {\varepsilon}$. Since this set lies in the $sk$-dimensional subspace by Theorem \[volume\_bound\] its cardinality is bounded by $4^{sk}$ (provided that ${\varepsilon}$ is sufficiently small). Thus, we have by plugging this bound into (\[php\]) $$\frac{2^{\Omega(k \log(n / k))}}{\binom{m}{sk}} \leq 4^{sk}.$$ Now by using a standard estimate $\binom{m}{sk} \leq 2^{O(sk \log(m / sk))}$ we have the desired statement. Now we can finish the proof of Theorem \[general\_lower\_bound\]. By Theorem \[model\_sparsification\] we can get an $m \times \Omega(n)$ matrix $A$ with column sparsity $s = O(m / k)$ and which is $(l, O({\varepsilon}))$-RIP-$1$. Then applying Theorem \[sparsity\_m\_tradeoff\] we have $s \log(m / sl) = \Omega(\log(n / l))$. Since, $s = O(m / k)$ we get the desired bound $$m = \Omega\left(k \cdot \frac{\log(n / k)}{\log(k / l)}\right).$$ Next we apply Theorem \[general\_lower\_bound\] to ${\mathcal{B}}_{k, b}$- and ${\mathcal{T}}_k$-RIP-$1$ matrices. For any $k \geq 2 b$ and sufficiently small ${\varepsilon}>0$ if $A$ is an $m \times n$ matrix which is ${\varepsilon}$-${\mathcal{B}}_{k, b}$-RIP-$1$, then $m = \Omega(k \cdot (1 + \log_b (n / k)))$. Clearly, ${\mathcal{B}}_{k, b}$ satisfies the conditions of Theorem \[model\_sparsification\] for $l = k / b$. Thus, by Theorem \[general\_lower\_bound\] we have $$m = \Omega\left(k \cdot \log_b(n / k)\right).$$ At the same time, the lower bound $\Omega(k)$ is obvious. Combining them together we get the desired result. Let $A$ be an $m \times n$ matrix which is ${\varepsilon}$-${\mathcal{T}}_k$-RIP-$1$. Then, if ${\varepsilon}$ is sufficiently small and $k = \omega(\log n)$, $$m = \Omega\left(k \cdot \frac{\log(n / k)}{\log \log (n / k)}\right).$$ The next Lemma shows that for any $k = \omega(\log n)$ the model ${\mathcal{T}}_k$ satisfies the first condition of Theorem \[model\_sparsification\] with $l = \Omega(k / \log (n / k))$. Let $S \subseteq [n]$ be a subset of the full binary tree. Then there exists a subtree that contains both $S$ and the root with at most $O(\|S\| \log(n / \|S\|))$ vertices. Let $T$ be a subtree that consists of $\log \|S\|$ levels of the full binary tree that are closest to the root. Let $T'$ be a subtree that is a union of $T$ and paths from the root to all the elements of $\|S\|$. It is not hard to see that $\|T' \setminus T\| \leq \|S\| \log (n / \|S\|)$. As a result we get $$\|T'\| \leq \|T\| + \|S\| \log (n / \|S\|) \leq O(\|S\| \log (n / \|S\|)).$$ The second condition of Theorem \[model\_sparsification\] is satisfied as well (here we use that $k = \omega(\log n)$). Thus, applying Theorem \[general\_lower\_bound\] we have $$m = \Omega\left(k \cdot \frac{\log(n / k)}{\log \log(n / k)}\right).$$ Upper bounds for model-based RIP-$1$ matrices {#upper_bounds} ============================================= In this section we complement the lower bounds by upper bounds. We use the following obvious modification of a theorem from [@BGIKS08]. \[expanders\_rip\] If a graph $G = (U, V, E)$ is an ${\varepsilon}$-${\mathcal{M}}_k$-expander for some model ${\mathcal{M}}_k \subseteq \Sigma_k$, then the normalized (by a factor of $d$, where $d$ is the degree of all vertices from $U$) adjacency matrix of $G$ (which size is $\|V\| \times \|U\|$) is an $O({\varepsilon})$-${\mathcal{M}}_k$-RIP-$1$ matrix. Thus, it is sufficient to build ${\mathcal{M}}_k$-expanders with as small $m$ as possible. We use the standard probabilistic argument to show the existence of such graphs. Namely, for every vertex $u \in U$ we sample a subset of $[m]$ of size $d$ ($d$ and $m$ have to be carefully chosen). Then, we connect $u$ and all the vertices from this subset. All sets we sample are uniform (among all $d$-subsets of $[m]$) and independent. We use the following tail inequality, which can be proved using Chernoff bound (and whose slight variant is proved and used in [@BMRV02]). \[concentration\] There exist constants $C > 1$ and $\delta > 0$ such that, whenever $m \geq C dt / {\varepsilon}$, one has for any $T \subseteq U$ with $\|T\| = t$ $${\mathrm{Pr}_{}\left[\|{\left\{v \in V \mid \exists u \in T \; (u, v) \in E\right\}}\| < (1 - {\varepsilon}) dt\right]} \leq \left(\delta \cdot \frac{{\varepsilon}m}{dt}\right)^{-{\varepsilon}dt}$$ For the proof see Appendix (Section \[concentration\_proof\]). For a model ${\mathcal{M}}_k \subseteq \Sigma_k$ denote $\#({\mathcal{M}}_k, t)$ the number of ${\mathcal{M}}_k$-sparse sets of size $t$ (for $t \in [k]$). We use the following simple estimate. $$\forall t \in [k] \quad \#({\mathcal{M}}_k, t) \leq \min{\left\{\|{\mathcal{M}}_k\| \cdot \binom{k}{t}, \binom{n}{t}\right\}}.$$ Now if we combine this Lemma with the standard estimate $\binom{u}{v} \leq (eu / v)^v$, we get the following bound. $$\label{block_estimates} \forall t \in [k] \quad \#({\mathcal{M}}_k, t) \leq \min {\left\{\|{\mathcal{M}}_k\| \cdot \left(\frac{ek}{t}\right)^t, \left(\frac{en}{t}\right)^t\right\}}$$ Now we combine these estimates, Lemma \[concentration\], Theorem \[expanders\_rip\] and the union bound to get upper bounds for $m$ for ${\mathcal{M}}_k$-RIP-$1$ matrices. \[model\_upper\] For any $0 < {\varepsilon}< 1/2$ and any model ${\mathcal{M}}_k$ that is of size at least $n / k$ there exists an ${\varepsilon}$-${\mathcal{M}}_k$-RIP-$1$ matrix with $$m = O\left(\frac{k}{{\varepsilon}^2} \cdot \frac{\log(n / l)}{\log(k / l)}\right),$$ where $$l = \frac{\log \|{\mathcal{M}}_k\|}{\log(n / k)}.$$ The proof is in the appendix. We note that the condition $\|{\mathcal{M}}_k\| \geq n / k$ is needed to make sure that $l \geq 1$. Now we derive corollaries for the block- and tree-sparse cases. \[block\_upper\] For any $0 < {\varepsilon}< 1/2$ and any $1 \leq b \leq k \leq n$ there exists an ${\varepsilon}$-${\mathcal{B}}_{k,b}$-RIP-$1$ matrix with $$m = O\left(\frac{k \cdot (1 + \log_b(n/k))}{{\varepsilon}^2}\right)$$ rows. It is easy to see that $$\|{\mathcal{M}}_k\| = \binom{n/b}{k/b} \geq n/k.$$ Thus, we can apply Theorem \[model\_upper\] with $ l = k/b $ and get the required bound. \[tree\_upper\] For any $0 < {\varepsilon}< 1/2$ and any $k = \omega(\log n)$ there exists an ${\varepsilon}$-${\mathcal{T}}_k$-RIP-$1$ matrix with $$m = O\left(\frac{k}{{\varepsilon}^2}\cdot\frac{\log(n / k)}{\log \log(n/k)}\right)$$ rows. Using a simple estimate on Catalan numbers we can see that $\|{\mathcal{T}}_k\| \leq 4^k$. Since $k = \omega(\log n)$, we are in position to apply Theorem \[model\_upper\] and get the required bound. Acknowledgments =============== This work was supported in part by NSF CCF-1065125 award, by MADALGO (a Center of the Danish National Research Foundation), by Packard Foundation and by Akamai Presidential Fellowship. The second author would like to thank Jelani Nelson for useful discussions. We also thank Ludwig Schmidt who pointed out an error in an earlier version of this paper. RIP-$1$ yields sparse recovery {#s:imp} ============================== In this section we show improved upper bounds on the number of measurements needed to recover good block- or tree-sparse approximations with $\ell_1 / \ell_1$ guarantee and constant approximation factor. This result is folklore, but we include it for completeness. Suppose that ${\mathcal{M}}_k \subseteq \Sigma_k$ is some model. We say that an $m \times n$ matrix $A$ is ${\varepsilon}$-${\mathcal{M}}_k^{(2)}$-RIP-$1$, if for every $x \in {\mathbb{R}}^n$ such that $\operatorname{supp}x \subseteq S_1 \cup S_2$ for some ${\mathcal{M}}_k$-sparse sets $S_1$ and $S_2$ one has $$(1 - {\varepsilon}) \\|x\\|_1 \leq \\|Ax\\|_1 \leq (1 + {\varepsilon}) \\|x\\|_1.$$ Let $A$ be any ${\varepsilon}$-${\mathcal{M}}_k^{(2)}$-RIP-$1$ matrix for a sufficiently small ${\varepsilon}$ such that $\\|A\\|_1 \leq 1 + {\varepsilon}$. Algorithm \[recovery\_algorithm\] (whose running time is exponential in $n$) given $y = Ax$ for some $x \in {\mathbb{R}}^n$ recovers a vector $x^* \in {\mathbb{R}}^n$ such that $$\label{good_approximation} \\|x - x^\\|_1 \leq (3 + O({\varepsilon})) \cdot \min_{\mbox{$x'$ is ${\mathcal{M}}_k$-sparse}} \\|x - x'\\|_1.$$ $y = Ax$ for some $x \in {\mathbb{R}}^n$ a good ${\mathcal{M}}_k$-approximation $x^$ of $x$ $x^* \gets 0$ $\tilde{x} \gets \operatorname{argmin}_{\operatorname{supp}x' \subseteq S} \\|y - Ax'\\|_1$ $x^* \gets \tilde{x}$ Note that the optimization problem within the for-loop can be easily reduced to a linear program. Now let us prove that the resulting vector $x^$ satisfies (\[good\_approximation\]). Denote $x_{{\mathcal{M}}_k}$ an ${\mathcal{M}}_k$-sparse vector that minimizes $\\|x - x_{{\mathcal{M}}_k}\\|_1$. Now we have $$\begin{gathered} \\|x - x^\\|_1 \leq \\|x - x_{{\mathcal{M}}_k}\\|_1 + \\|x_{{\mathcal{M}}_k} - x^\\|_1 \leq \\|x - x_{{\mathcal{M}}_k}\\|_1 + (1 + O({\varepsilon})) \\|A(x_{{\mathcal{M}}_k} - x^)\\|_1 \leq \\ \leq \\|x - x_{{\mathcal{M}}_k}\\|_1 + (1 + O({\varepsilon})) (\\|A(x - x_{{\mathcal{M}}_k})\\|_1 + \\|A(x - x^)\\|_1) \leq \\ \leq (2 + O({\varepsilon})) \\|x - x_{{\mathcal{M}}_k}\\|_1 + (1 + O({\varepsilon})) \\|A(x - x^)\\|_1 \leq \\ \leq (2 + O({\varepsilon})) \\|x - x_{{\mathcal{M}}_k}\\|_1 + (1 + O({\varepsilon})) \\|A(x - x_{{\mathcal{M}}_k})\\|_1 \leq (3 + O({\varepsilon})) \\|x - x_{{\mathcal{M}}_k}\\|_1. \end{gathered}$$ The second inequality is true, since $A$ is ${\varepsilon}$-${\mathcal{M}}_k^{(2)}$-RIP-$1$ and both $x_{{\mathcal{M}}_k}$ and $x^$ are ${\mathcal{M}}_k$-sparse. The fourth inequality is true, because $\\|A\\|_1 \leq 1 + {\varepsilon}$. The fith inequality is due to the construction of the algorithm: clearly, $\\|A(x - x^)\\|_1 \leq \\|A(x - x_{{\mathcal{M}}_k})\\|_1$. It is immediate to see that any ${\varepsilon}$-${\mathcal{B}}_{2k, b}$-RIP-$1$ matrix is ${\varepsilon}$-${\mathcal{B}}_{k, b}^{(2)}$-RIP-$1$. Similarly, any ${\varepsilon}$-${\mathcal{T}}_{2k}$-RIP-$1$ matrix is ${\varepsilon}$-${\mathcal{T}}_k^{(2)}$-RIP-$1$. Moreover, since all the singletons are both block- and tree-sparse, we have that these matrices have $\ell_1$-norm at most $1 + {\varepsilon}$. Thus, plugging Corollaries \[block\_upper\] and \[tree\_upper\] we get the following result. The problem of model-based stable sparse recovery with $\ell_1 / \ell_1$ guarantee and a constant approximation factor can be solved - with $$m = O(k \cdot (1 + \log_b (n / k)))$$ measurements for ${\mathcal{B}}_{k, b}$, - with $$m = O\left(k \cdot \frac{\log(n / k)}{\log \log(n / k)} \right)$$ measurements for ${\mathcal{T}}_k$, provided that $k = \omega(\log n)$. Proof of Lemma \[concentration\] {#concentration_proof} ================================ To prove the lemma we need the following version of Chernoff bound [@MR95]. \[chernoff\] Suppose that $X_1$, $X_2$, …, $X_n$ are independent binary random variables. Denote $\mu = {\mathrm{E}_{}\left[X_1 + X_2 + \ldots + X_n\right]}$. Then for any $\tau > 0$ $${\mathrm{Pr}_{}\left[X_1 + X_2 + \ldots + X_n \geq (1 + \tau) \mu\right]} \leq \left(\frac{e^{\tau}}{(1 + \tau)^{(1 + \tau)}}\right)^{\mu}.$$ We can enumerate all $dt$ outgoing from $T$ edges arbitrarily. Then let us denote $C_u$ the event “$u$-th of these edges collides with $v$-th edge on the right side, where $v < u$”. We would like to upper bound the probability of the event “at least ${\varepsilon}dt$ of events $C_u$ happen”. As shown in [@BMRV02] this probability is at most $${\mathrm{Pr}_{}\left[B\left(dt, \frac{dt}{m}\right) \geq {\varepsilon}dt\right]}.$$ Thus we can apply Theorem \[chernoff\] with $\tau = {\varepsilon}m / dt - 1$ and $\mu = (dt)^2 / m$. Then, the statement of the Lemma can be verified routinely. Proof of Theorem \[model\_upper\] ================================= If we combine Lemma \[concentration\] with the union bound we see that it is sufficient to prove that $$\forall 1 \leq t \leq l \quad \#({\mathcal{M}}_{k}, t) \cdot \left(\delta \cdot \frac{{\varepsilon}m}{dt}\right)^{-{\varepsilon}dt} < \frac{1}{2l},$$ and $$\forall l \leq t \leq k \quad \#({\mathcal{M}}_{k}, t) \cdot \left(\delta \cdot \frac{{\varepsilon}m}{dt}\right)^{-{\varepsilon}dt} < \frac{1}{2k},$$ provided that $m \geq C dk / {\varepsilon}$ (here $C$ and $\delta$ are the constants from the statement of Lemma \[concentration\] and $l$ is the quantity from the statement of Theorem \[model\_upper\]). Now let us plug (\[block\_estimates\]). Namely, for $1 \leq t \leq l$ we use the second estimate, for $l \leq t \leq k$ we use the first estimate. Thus, it is left to choose $d$ such that we have $$\begin{aligned} \label{block_low_inequality} \forall 1 \leq t \leq l &\quad& \left(\frac{en}{t}\right)^t \left(\frac{C'k}{t}\right)^{-{\varepsilon}dt} < \frac{1}{2l}\\ \label{block_high_inequality} \forall l \leq t \leq k &\quad& \|{\mathcal{M}}_k\| \cdot \left(\frac{ek}{t}\right)^t \left(\frac{C'k}{t}\right)^{-{\varepsilon}dt} < \frac{1}{2k}, \end{aligned}$$ for $C'$ being sufficiently large constant. It would allow us to set $m = O(dk / {\varepsilon})$. It is straightforward to check that, whenever $d > 1 / {\varepsilon}$, the left-hand sides of (\[block\_low\_inequality\]) and (\[block\_high\_inequality\]) are log-convex. Thus, to check (\[block\_low\_inequality\]) and (\[block\_high\_inequality\]), it sufficient to check them for $t = 1, l, k$. Let us first handle (\[block\_low\_inequality\]). For it to hold for a given $1 \leq t \leq l$ it is enough to set $$d = O\left(\frac{1}{{\varepsilon}\cdot \log(k / t)} \cdot \left(\frac{\log l}{t} + \log(n/t)\right)\right).$$ For $t = 1$ this expression is $O({\varepsilon}^{-1} \cdot \log_k n)$, for $t = l$ it is $$\label{d_final} O\left(\frac{\log(n / l)}{{\varepsilon}\cdot \log(k / l)}\right).$$ Clearly, the latter bound is always larger than the former one. Now let us handle (\[block\_high\_inequality\]). Similarly, for $l \leq t \leq k$ it is enough to set $$\label{d_second} d = O\left(\frac{1}{{\varepsilon}} \cdot \left(1 + \frac{\log \|{\mathcal{M}}_k\| + \log k}{t \cdot \log(k / t)}\right)\right) = O\left(\frac{1}{{\varepsilon}} \cdot \left(1 + \frac{l \cdot \log(n / k) + \log k}{t \cdot \log(k / t)}\right)\right),$$ where the second step is due to the definition of $l$ ($l = \log \|{\mathcal{M}}_k\| / \log(n / k)$). Now we will prove that for $t = l, k$ the right-hand side of (\[d\_second\]) is at most (\[d\_final\]). For $t = l$ we obviously have $$\frac{l \cdot \log(n / k)}{t \cdot \log(k/t)} = O\left(\frac{\log(n/l)}{\log(k/l)}\right).$$ At the same time $$\frac{\log k}{t \cdot \log(k/t)} = O\left(\frac{\log(n/l)}{\log(k/l)}\right),$$ since $\log k=O(l \cdot \log(n/l))$. Indeed, if $l \geq \log n$, then we are done. Otherwise, $\log(n / l) = \Omega(\log n) = \Omega(\log k)$, and we are again done. Now let us handle $t = k$. Indeed, $$\frac{\log k}{t \cdot \log(k/t)} = O(1)$$ in this case. At the same time, $$\frac{l \cdot \log(n / k)}{t \cdot \log(k/t)} = O\left(\frac{\log(n/l)}{\log(k/l)}\right),$$ since $l \cdot \log(k / l) = O(k)$. Overall, this argument shows that we can take $$d = O\left(\frac{\log(n / l)}{{\varepsilon}\cdot \log(k / l)}\right)$$ and $$m = O\left(\frac{k}{{\varepsilon}^2}\cdot\frac{\log(n / l)}{\log(k / l)}\right).$$ [^1]: MIT CSAIL, `[email protected]` [^2]: MIT CSAIL, `[email protected]` [^3]: In fact, one can prove a somewhat stronger guarantee, referred to as the $\ell_2/\ell_1$ guarantee. [^4]: See Section \[s:def\] for formal definitions of the two models. [^5]: See Question 15: Sparse Recovery for Tree Models. The question was posed by the first author.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Modulo conventional scale factors, the Simon and Simon-Mars tensors are defined for stationary vacuum spacetimes so that their equality follows from the Bianchi identities of the second kind. In the nonvacuum case one can absorb additional source terms into a redefinition of the Simon tensor so that this equality is maintained. Among the electrovacuum class of solutions of the Einstein-Maxwell equations, the expression for the Simon tensor in the Kerr-Newman-Taub-NUT spacetime in terms of the Ernst potential is formally the same as in the vacuum case (modulo a scale factor), and its vanishing guarantees the simultaneous alignment of the principal null directions of the Weyl tensor, the Papapetrou field associated with the timelike Killing vector field, the electromagnetic field of the spacetime and even the Killing-Yano tensor.' address: - ' ${}^{\dag}$Istituto per le Applicazioni del Calcolo M. Picone, C.N.R., I– 00161 Roma, Italy ' - ' ${}^{\ddag}$ International Center for Relativistic Astrophysics, University of Rome, I–00185 Roma, Italy ' - ' ${}^{\ast}$ Department of Physics E.R. Caianiello, University of Salerno, I–84081, Italy ' - ' ${}^{\|\|}$Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth, PO1 2EG, UK' - ' ${}^{\S}$ Department of Mathematical Sciences, Villanova University, Villanova, PA 19085, USA ' - ' ${}^{\P}$Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge, CB3 OHA ' author: - 'Donato Bini${}^{\dag\,\ddag}$, Christian Cherubini${}^{\ast,\|\|}$, Robert T. Jantzen${}^{\S \,\ddag}$, Giovanni Miniutti${}^{\P \,\ddag}$' title: 'The Simon and Simon-Mars Tensors for Stationary Einstein-Maxwell Fields' --- Introduction ============ In the $1+3$ gravitoelectromagnetic approach to general relativity associated with splitting spacetime quantities using a unit timelike vector field $u$, the 4-velocity of an observer congruence, the Bianchi identities for the curvature tensor take a Maxwell-like form for the electric and magnetic parts of the Weyl tensor with sources arising from the Ricci tensor terms, which in turn can be expressed in terms of the energy-momentum tensor of the spacetime through the Einstein field equations [@Ellis73; @Maartens; @Ellis99]. For the stationary case with $u$ directed along the timelike Killing vector field, these equations are greatly simplified and suggested the introduction of the complex Simon tensor [@Simon], which in the vacuum case equals the simpler Simon-Mars tensor [@Mars99; @Mars00; @Mars01; @Fasop99; @Fasop00; @cys1] (modulo conventional scale factors). The Simon-Mars tensor is a natural algebraic combination of the self-dual fields associated with the Papapetrou field of the timelike Killing vector field and the Weyl tensor. Simon introduced his Simon tensor for vacuum stationary spacetimes as a complex generalization of the Cotton-York tensor, to which it reduces in the static case, where its vanishing guarantees that the space sections are conformally flat, as occurs in the Schwarzschild spacetime. Its vanishing also characterizes the Kerr spacetime within certain classes of solutions of the stationary vacuum Einstein equations. Mars has given an equivalent form of the Simon tensor (the Simon-Mars tensor) that has been very useful in understanding the meaning of its vanishing in the Kerr spacetime, where it leads to the alignment of the principal null directions of the Weyl tensor of the spacetime with those of the Papapetrou field associated with the timelike Killing congruence. Here we extend the gravitoelectromagnetic analysis of the vacuum case [@cys1] to electrovac stationary spacetimes which contain a source-free electromagnetic field, where the Simon tensor and the Simon-Mars tensor defined by the same formulas as in vacuum differ by additional source terms. However, these source terms can be absorbed into a redefinition of the Simon tensor so that it continues to be equivalent to the Simon-Mars tensor defined exactly as in the vacuum case, and of course having all the same tensor symmetries. Remarkably for the Kerr-Newman spacetime and even the more general family of Kerr-Newman-Taub-NUT spacetimes, the situation is very similar to the Kerr spacetime. First the vanishing of the Simon-Mars tensor (and hence of the newly defined Simon tensor) can still be explained in terms of the simultaneous alignment of the principal null directions of the Weyl tensor, the Papapetrou field associated with the timelike Killing vector field, the electromagnetic field of the spacetime and even the Killing-Yano tensor. Secondly when written in terms of the Ernst potential, the newly defined Simon tensor including the source terms formally coincides with the vacuum expression (modulo a scale factor). Our extension of the Simon tensor to stationary electrovac spacetimes takes a different approach from the work of Marklund and Perjés on stationary perfect fluid spacetimes [@Perjes]. They are also able to incorporate additional source terms into the definition of the nonvacuum generalization of the Simon tensor in such a way that the symmetry properties of the tensor are respected. However, they focus on the spatial Bianchi identities (for the conformally rescaled spatial metric) rather than the spacetime Bianchi identities. Finally it is worth noting that a prime motivation of our discussion, begun in [@cys1], is to make more accessible the geometry surrounding the Simon tensor, whose existing literature assumes detailed knowledge of stationary exact solution techniques that outsiders simply do not possess. Nonexperts cannot even recognize the equivalence of different formulas for the Simon tensor stated by various experts in these manipulations. Appendix A reconciles these definitions, while the main body recasts the geometry in the natural complex version of the Maxwell-like equations for the Weyl tensor pioneered by Ellis [@Ellis73; @Maartens; @Ellis99]. The $1+3$ Maxwell-like equations for the Weyl tensor with sources ================================================================= The Bianchi identities of the second kind equate the divergence of the Weyl curvature tensor $C_{\alpha\beta\gamma\delta}$ to half the Cotton tensor [@Cotton], which can be thought of as a current in analogy with electromagnetism [@HE] (where $F^{\alpha\beta}{}_{;\beta} = 4\pi J^\alpha$) \[current\] C\^\_[;]{} = -\_[\[]{} ( R\^\_[\]]{} -16 R \^\_[\]]{} ) J\^\_ , following the conventions of Misner, Thorne and Wheeler [@mtw], including the signature $-$+++. The Einstein equations (in Ricci form) \[eq:EEricci\] R\^\_ = (T\^\_- [12]{} T\^\_\^\_) can then be used to replace the Ricci tensor and scalar curvature terms in this current by the energy-momentum tensor. Splitting these equations with respect to a generic timelike congruence $u$ leads to their $1+3$ Maxwell-like form [@Ellis99] given explicitly in our notation in [@cys1]. The Weyl tensor splits into two symmetric tracefree spatial fields, its electric and magnetic parts respectively E \^ \_ = C\^ \_ u\^ u\^ , H \^ \_ = - \^C\^ \_ u\^ u\^= \^ \_ \^ C\^\_ u\^ . If $a$ is the acceleration vector and $\omega$ the vorticity vector of $u$, introducing the complex spatial fields from the gravitoelectromagnetic connection and Weyl curvature fields z=-a-i= g-iH/2 , Z=E-iH , where $g$ and $\vec H$ are the gravitoelectric and gravitomagnetic vector fields, leads to a more efficient representation of those equations. These complex fields are associated with the $O(3,C)$ representation of the Lorentz group exploited in the ‘complex vector / self-dual formalism’ [@Kramer; @Debeve; @Taub] in turn closely related to the Newman-Penrose formalism. Here we specialize to the case of a stationary electrovacuum spacetime and align $u = M^{-1}\xi$ with the associated Killing vector $\xi$, so that the acceleration can be expressed as $a = \nabla \ln M$, where $M=\|-\xi_\alpha \xi^\alpha\|^{1/2}$. Splitting the Weyl divergence equations and expressing them in terms of an adapted frame $e_\top =u, e_a$, $a=1,2,3$, where the spatial frame $e_a$ is orthogonal to $u$, one finds in an index-free notation the identities [@cys1] \[bianchi\] M\^[3]{}= 3 z Z + \^[(G)]{} ,M\^[-1]{}= zZ + i J\^[(G)]{} ,where the complex current fields are defined by \^[(G)]{}\_a=J\^\_[a]{}+iJ\^[\]{} \^\_[a]{} , J\^[(G)]{}\_[ab]{}=-\[J\_[(ab)]{}+i J\^[\]{}\_[(ab)]{}\] . The div and Scurl spatial operators on symmetric spatial 2-tensors $S$ \[(u) S\]\^= (u)\_S\^ , \[(u) S\]\^ = (u)\^[(]{} (u)\_S\^[)]{}\_are defined in terms of the spatial covariant derivative $\nabla(u)$, the spatial volume 3-form $\eta(u)_{\alpha\beta\gamma} =u^\delta \eta_{\delta\alpha\beta\gamma}$ (associated with the spatial duality operation $\dualp{u}$, used for example to define the vorticity vector $\omega(u)^\alpha=\frac12\eta(u)^{\alpha\beta\gamma}\nabla_\beta u_\gamma$ from the corresponding 2-form, see appendix A of [@cys1] for sign conventions) and the spatial cross and dot products associated with $u$, all given explicitly in [@cys1] by spatially projecting spacetime derivatives and quantities orthogonally to $u$. For example, the cross-product of a spatial vector and spatial symmetric tensor is \[X \_u A\]\^ = (u)\^[(]{} X\_ A\^[)]{} \_ . For simplicity we drop the reference to $u$ on these spatially projected operators, resorting to the abbreviation $\vec\nabla=\nabla(u)$ to distinguish the spatial covariant derivative from the spacetime covariant derivative $\nabla$, an unnecessary distinction when acting on stationary scalars. In an adapted frame spatial quantities may be expressed using only Latin indexed components. The spatial cross product of $z$ and $Z$ which appears in the Scurl equation in the projected Bianchi identity (\[bianchi\]) defines the (complex, symmetric and tracefree) Simon-Mars 2-tensor \[simon,cys1\] [SimonMars]{} = zZ . By using that Scurl equation to define the nonvacuum Simon 2-tensor for any stationary spacetime (with any energy-momentum source) as = M\^[-1]{}- i J\^[(G)]{} , this identity then guarantees that the Simon and Simon-Mars tensors also remain equal in the nonvacuum case. Of course once the Ricci tensor in the source term $J^{(G)}$ is replaced using the Einstein equations (\[eq:EEricci\]), this equality between Simon and Simon-Mars is only true if they are each evaluated on a solution of those equations. Apart from a scale factor, our Simon tensor is the spatial dual of the three-index spacetime tensor introduced by Simon in the vacuum case [@Simon], as shown in Appendix A. In the special case in which the source term has the form $J^{(G)} = -i \nabla \ln\sigma \times Z$, then the Scurl identity for any scalar $\sigma$ and any symmetric spatial 2-tensor $A$ =-A + A enables it to be incorporated into the existing Scurl term \[eq:simonsigma\] [Simon]{} = (/M) . This again would re-establish the vacuum property that the Simon tensor is proportional to a Scurl, but with a generalized Simon tensor incorporating the source terms so that it still equals the Simon-Mars tensor. This is the case with the family of Kerr-Newman-Taub-NUT electrovac spacetimes (including the more physically interesting Kerr-Newman spacetime), where the extra factor of $\sigma$ converts this to the same Scurl expression when written in terms of the Ernst potential as in the vacuum formula (modulo a scale factor). The vanishing of the Simon-Mars tensor ($z\times Z=0$) is equivalent to the alignment of $z$ and $Z$ in the sense that $Z$ must be proportional to the tracefree tensor product of $z$ with itself [@cys1] Z=k\[zz\]\^[(TF)]{} , which is an algebraic consequence of the definition of the cross product for symmetric spatial 2-tensors [@cys1]. Thus the vanishing of the generalized Simon tensor corresponds to the alignment of $z$ and $Z$ as in the vacuum case. Maxwell equations and the Papapetrou field ========================================== The stationary Maxwell-like equations (\[bianchi\]) for the Weyl tensor expressed in complex self-dual form are very similar to the stationary Maxwell equations expressed in complex form [@cys1] \[eq:max\] M\^[-2]{}(M\^2 x) = -2 zx+4 ,M\^[-1]{}(M x)=4i J satisfied by the spacetime’s electromagnetic field (using an arrow to distinguish the electric field from the electric part of the Weyl tensor in index-free notation) F=uE + B ,x = E -iB . They only differ in form by the missing $z\times x$ term in the curl equation here, where $\dualp{u}B$ is the spatial dual of the magnetic 1-form field $B$. However, when $z$ and $Z$ are aligned (i.e., $z\times Z=0$) as in the Kerr-Newman-Taub-NUT spacetimes, the correspondence is complete. The sources $\rho$ and $J$ for the spacetime electromagnetic field are zero for the electrovac spacetimes considered here. The curl equation for the electromagnetic field implies that $Mx$ is the gradient of a potential \[potx\] Mx= , x=M\^[-1]{} . Papapetrou [@pap66] observed that a Killing vector field $\xi$ can be interpreted as a Lorentz gauge vector potential for an electromagnetic field later called the Papapetrou field [@Fasop99; @Fasop00] \[eq:papfield\] F\_ =\[d \]\_ = \_\_- \_\_= 2\_\_ satisfying Maxwell’s equations with an effective current source proportional to the Ricci tensor \[eq:papst\] \_[(;)]{}=0 \^\_[;]{}=0 \_[;]{}\^[;]{} = -\_R\^\_ F\_\^[;]{} = 2\_R\^\_ , where this divergence condition is merely the trace of the Ricci identity for a Killing vector field. Rescaling the Papapetrou field $\tilde F$ by the factor $(2M)^{-1}$ [@cys1] and introducing its complex self-dual representation yields the self-dual 2-form $\mathcal{F}$ associated with $z$ \[eq:Papselfdual\] = (2M)\^[-1]{}(F + i F) = uz +i z , which when multiplied by $M$ satisfies the same Maxwell equations but with nonzero sources generated by the electromagnetic field. These equations \[eqpap\] M\^[-3]{}(M\^3z)=-2 zz- x\|x , M\^[-2]{} (M\^2 z) = x \|x . are obtained by splitting the vector Eqs. (\[eq:papst\]) and using the Einstein equations (\[eq:EEricci\]) to replace the Ricci tensor appearing there in terms of the appropriate projections of the electromagnetic energy-momentum tensor of the spacetime T\^\_ = - x\|x , i T\^\_[a]{} = \[x \|x\]\_a , which then leads to the source terms appearing in these Maxwell equations. Inserting the potential representation (\[potx\]) of $Mx$ into the rescaled Papapetrou field curl equation of Eqs. (\[eqpap\]) then yields M\^[-2]{} (M\^2 z) = M\^[-2]{} \|=- M\^[-2]{} (\|) , namely \[eq:MPhi\] (M\^2 z + \|)=0 . This latter quantity in parentheses must therefore be a gradient, defining the Ernst potential $\mathcal{E}$ [@Kramer] (see their Eq. (18.34) with $F=-M^2$, reducing to the integrated Eq. (18.29) in the vacuum case $\Phi=0$) modulo a numerical factor \[eq:ernstpot\] = -2 (M\^2 z + \|) . The remaining divergence equation of Eqs. (\[eqpap\]) is the Ernst equation, which can be expressed in the form (equivalent to Eq. (18.38) of [@Kramer], see Appendix A) \[ernst\] M (M ) = - 2M\^2 z . This result follows from using the definition of $\mathcal{E}$ to re-express $x$ in terms of the gradient of $\Phi$ in the two divergence equations (\[eq:max\]) and (\[eqpap\]) 12 +\|= M\^2 (a+2z)z, = - (a+2z)and using the second of these equations to eliminate $\div \nabla \Phi$ in the first, yielding +a = - 2z , which can then be rewritten as Eq. (\[ernst\]). Simon and the Ernst potential ============================= The Simon-Mars tensor can be evaluated in a form which shows its relation to the Ernst potential expression for the Simon tensor, assuming that the Einstein field equations are satisfied so that the two tensors are in fact equal. Following the notation of [@cys1], the purely spatial components and the mixed time-space components of the Ricci tensor field equations (see [@Kramer], Eqs. 18.15 and 18.16) are respectively $$\begin{aligned} \label{ricci} [{}^{(4)} R_{ab}]^{\rm (TF)} &=& -2M^{-2}[\nabla_{(a}\Phi \nabla_{b)}\bar\Phi]^{\rm (TF)}\ ,\nonumber \\ {}^{(4)}R_{\top b} &=& iM^{-2}[\curl (M^2 z)]_b =-iM^{-2}[\curl (\bar \Phi \nabla \Phi)]_b \ , % &\equiv & \mathcal{R}_{b} %= M^{-2}[\curl (M^2\omega)]_b \nonumber \\ %&=& iM^{-2}[\curl (M^2 z)]_b \nonumber \\ %& =& -iM^{-2}[\curl (\bar \Phi \nabla \Phi)]_b\end{aligned}$$ where $z$ and $\Phi$ are related to the Ernst potential by (\[eq:ernstpot\]). A straightforward computation gives the electric and magnetic part of the Weyl tensor in terms of the gravitoelectromagnetic fields $a$ and $\omega$ and the Ricci tensor of the quotient 3-manifold [@cys1] $$\begin{aligned} E&=& -\frac12 [4 \omega \otimes \omega -\vec\nabla a -a\otimes a -^{(3)}{\rm Ricci}]^{\rm (TF)}\ , \nonumber \\ H&=&-\SYM [\vec\nabla \omega +2 a \otimes \omega]^{\rm (TF)}\ .\end{aligned}$$ Then using the formula \[R3vsR4\] \[\^[(3)]{}[Ricci]{}\]\^[(TF)]{} =\[\^[(4)]{}[Ricci]{} + a + aa +2 \]\^[(TF)]{} , one finds \[Zdef\] Z = E-iH =\[-(z) + z z + 12 \^[(4)]{}[Ricci]{}\]\^[(TF)]{} . The Simon-Mars tensor is then = z(-z+12 \^[(4)]{}[Ricci]{}) since $z\times(z\otimes z)$ is identically zero. By next replacing the first term here using the Scurl identity valid for an arbitrary spatial vector field $X$ (XX) = -X(X) + 32 \[(XX)\]\^[(TF)]{} , and using the Einstein Eqs. (\[ricci\]) to replace the spatial components of the spacetime Ricci tensor, the Simon-Mars tensor becomes $$\begin{aligned} \label{nuovosimon} {\rm SimonMars} &=& \Scurl (z\otimes z) -\frac32 [\SYM (z\otimes \curl \,z)]^{\rm (TF)} \nonumber\\ &&\quad -M^{-2} z\times \SYM (\nabla \Phi \otimes \nabla\bar\Phi )\ .\end{aligned}$$ Finally by substituting $z=M^{-2}(M^2 z)$ into the curl term here expanding the result with the appropriate product rule and incorporating the resulting additional acceleration term into the Scurl term, and then using Eq. (\[eq:ernstpot\]) (M\^2 z) =-(\|) to substitute the other term in this expansion, one finds $$\begin{aligned} \label{nuovosimon1} {\rm SimonMars} &=& M^{-3}\Scurl (M^3 \, z\otimes z) - \frac32 [\SYM (z\otimes (x\times \bar x)]\TF \nonumber\\ && \quad - z\times \SYM (x \otimes \bar x ) \ .\end{aligned}$$ In the vacuum case where $x=0$, this reduces to \[nuovosimon2\] [SimonMars]{} = M\^[-3]{}(M\^3 zz) , which can be rewritten in terms of the Ernst potential by making the substitution $z= -\frac12 M^{-2} \nabla \mathcal{E}$ \[simonmarsvacuum\] [SimonMars]{} = 14 M\^[-3]{} (M\^[-1]{} ) . As shown in appendix A, this is proportional to the form of the Simon tensor given in [@Kramer] and to the original form given by Simon. The Kerr-Newman-Taub-NUT spacetime ================================== In Boyer-Lindquist coordinates $(x^\alpha)=(t,r,\theta,\phi)$, with the abbreviations $(c,s)=(\cos\theta,\sin\theta)$, the (exterior) Kerr-Newman-Taub-NUT (KNTN) spacetime the line element is [@DemNew; @Mil; @NUT; @McGRuf; @DodTur] $$\begin{aligned} \label{metrica} \rmd s^2 &=&-\frac{1}{\Sigma}(\Delta -a^2 s^2)\rmd t^2 +\frac{2}{\Sigma}[\Delta \chi -a(\Sigma +a \chi)s^2]\rmd t \rmd \phi \nonumber \\ &&+\frac{1}{\Sigma}[(\Sigma +a \chi)^2 s^2 -\chi^2 \Delta ]\rmd \phi^2 +\frac{\Sigma}{\Delta}\rmd r^2 +\Sigma d\theta^2\ ,\end{aligned}$$ and the corresponding electromagnetic Faraday tensor can be expressed in terms of the 2-form $$\begin{aligned} \label{max} F&=&\frac{Q}{\Sigma^2}\{ [r^2-(\ell+ac)^2]\rmd r \wedge (\rmd t -\chi\rmd \phi)\nonumber \\ && +2rs (\ell+ac)\rmd \theta \wedge [(r^2+a^2+\ell^2)\rmd \phi -a \rmd t]\}\ ,\end{aligned}$$ which is the exterior derivative of the vector potential A=-(t -) . Here $\Sigma$, $\Delta$, and $\chi$ are defined by = r\^2 +(+ac)\^2, = r\^2-2r-\^2 +a\^2+Q\^2, = a s\^2 -2c . Units are chosen so that $G=c=1$, so the parameters $(\mathcal{M},Q,a,\ell)$ all have the dimension of length. The source has mass $\mathcal{M}$, electric charge $Q$, angular momentum $J=\mathcal{M}a$ (i.e., gravitomagnetic dipole moment) along the $\theta=0$ direction, and gravitomagnetic monopole moment $\mu=-\ell$, where $\ell$ is the NUT parameter. [@Mis; @MisTau; @Bon; @Fei]. The spatial 3-metric $h_{\alpha\beta}$ on the quotient space of the timelike Killing congruence associated with the Killing vector field $\xi=\partial_t$ is $$\rmd l^2=h_{\alpha\beta}\rmd x^\alpha \rmd x^\beta =\frac{\Sigma}{\Delta}\rmd r^2 +\Sigma \rmd \theta^2 +\frac{\Delta s^2 \Sigma}{\Delta-a^2 s^2}\rmd \phi^2\ ,$$ while its convenient conformal rescaling by the square of the lapse function $M^2=-\xi_\alpha \xi^\alpha = (\Delta-a^2 s^2\theta)/\Sigma$ is $$\begin{aligned} \rmd \tilde l{}^2 &=& \tilde h_{\alpha\beta}\rmd x^\alpha \rmd x^\beta = M^2 h_{\alpha\beta}\rmd x^\alpha \rmd x^\beta \nonumber \\ &=& \frac{\Delta-a^2\sin^2\theta}{\Delta}\rmd r^2 +(\Delta-a^2s^2) \rmd \theta^2 +\Delta s^2\rmd \phi^2\ .\end{aligned}$$ Using the latter metric and its covariant derivative $\tilde\nabla$, one easily shows by direct evaluation that both the Ernst potential and $\Phi$ $$\mathcal{E}=1-\frac{2(M+il)}{r+i(\ell+ac)}\ ,\ \Phi= \frac{Q}{r+i(\ell+ac)}$$ satisfy the same Scurlfree property for their gradient ( )=0 that occurs for the Ernst potential in the vacuum case within this family. This is explained by the special form of the Weyl current $$\begin{aligned} \fl J^{(G)} &=& \frac{3\Delta s^2 a Q^2}{[r-i(\ell+ac)][r+i(\ell+ac)]^2\Sigma^{1/2}(\Delta-a^2s^2)^{3/2}} (\rmd r-ias \rmd\theta)\wedge \rmd\phi \nonumber\\\fl &=& -i \nabla \log \sigma \times Z \ ,\end{aligned}$$ where = -34 a factor also which appears in Z=M\^[-2]{}\^[(TF)]{} , and so cancels the factor inside the Scurl in the Simon tensor expression (\[eq:simonsigma\]) when expressed in terms of $\mathcal{E}$ = M\^[-1]{} (\^[-1]{}MZ) = M\^[-1]{}(M\^[-1]{} ) . This latter Scurl expression is the same one originally used to define the Simon tensor in the vacuum case (where $\nabla\sigma \times Z=0$ and so disappears from the formula upon expansion of the first Scurl expression), and as shown in Appendix A is proportional to $\tilde S{\rm curl}\, (\tilde\nabla \mathcal{E} \otimes \tilde\nabla \mathcal{E})$. This is forced to be zero when $z$ and $Z$ are aligned as they are here $$\begin{aligned} \label{eq:zZ} z &=& \frac{\{[r-i(\ell+ac)](\mathcal{M}+i\ell)-Q^2\}}{[r+i(l+ac)]^2(\Delta-a^2s^2)[r-i(l+ac)]} (-\Delta \partial_r+ias \partial_\theta) \ ,\\ Z &=& \frac{3(\Delta-a^2s^2)}{\{[r-i(\ell+ac)](\mathcal{M}+i\ell)-Q^2\}} [z\otimes z]\TF \ ,\end{aligned}$$ In fact all the complex vectors $z$, $\nabla \mathcal{E}$, $\nabla\Phi$, and the one associated with the Killing-Yano tensor are all aligned with (i.e., proportional to) the gradient vector field $\nabla [r+i(\ell+ac)]$. Although we have downplayed the notational distinction between contravariant and covariant tensors, by writing explicitly the index-lowered 1-form z\^=- (r -ias ) , one sees that it is proportional to this gradient (exterior derivative). The Killing-Yano tensor [@Kramer] for this spacetime is naturally defined by using the principal null directions of the Riemann tensor. The principal Newman-Penrose tetrad is $$\begin{aligned} l&=& \frac{r^2+a^2+l^2}{\Delta}\partial_t+\partial_r+\frac{a}{\Delta}\partial_\phi\, ,\nonumber \\ n&=& -\frac{1}{2[r^2+(l+ac)^2]}[(r^2+a^2+l^2)\partial_t+\Delta \partial_r-a\partial_\phi]\, ,\nonumber \\ m&=& \frac{1}{\sqrt{2}[(l+ac)-ir]}[\frac{\chi}{s}\partial_t-i\partial_\theta+\frac{1}{s}\partial_\phi] \ .\end{aligned}$$ The Killing-Yano 2-form is then [@Kramer] (their Eq. (35.64)) $$f=-(l+ac)[l\wedge n]+ir [m\wedge \bar m]\ .$$ By taking its self-dual combination $\mathcal{K}=f+i{}^f$, one finds $$\begin{aligned} \fl \mathcal{K} &=& i[r+i(\ell +ac)] [\rmd t \wedge \rmd r -ias \rmd t \wedge \rmd \theta +\chi \rmd r \wedge \rmd \phi -is(r^2+a^2+\ell^2)\rmd \theta \wedge \rmd \phi]\nonumber \\\fl &=& -4i \frac{\{[r-i(\ell+ac)](\mathcal{M}+i\ell)-Q^2\}}{[r+i(\ell+ac)]^3[r-i(\ell+ac)]} M \mathcal{F} \ ,\end{aligned}$$ namely it is proportional to the self-dual combination (\[eq:Papselfdual\]) of the rescaled Papapetrou 2-form (\[eq:Papselfdual\]) associated with the timelike Killing vector. The (symmetric) Cotton-York tensor [@York] for the quotient-space geometry can be evaluated using either spatial metric since its covariant form is conformally invariant. For the rescaled metric (see [@cys1]), this tensor is y= - [curl]{} ( \^[(3)]{}) and its only nonzero components are y\^[r]{} = y\^ ,y\^ = The condition $\ell^2+\mathcal{M}^2-Q^2=0$ makes this tensor zero, corresponding to the conformal flatness of either spatial metric. The same is true for the static limit $a=0$, where the Cotton-York and Simon tensors coincide. Conclusions =========== The gravitoelectric description of the Simon and Simon-Mars tensors which play a crucial role in characterizing the Kerr spacetime has been generalized to the class of electrovacuum spacetimes where the corresponding Kerr-Newman spacetime exhibits a similar behavior once the Simon tensor is suitably generalized to incorporate the electromagnetic field. Again the alignment of the principal null directions of the Weyl, Papapetrou and now additionally the electromagnetic field is equivalent to the vanishing of the Simon-Mars tensor, which in turn forces the generalized Simon tensor to vanish as well. The Scurl and cross-product notation and their properties also help illuminate some of the manipulations of exact solution theory whose motivation is not so obvious. Acknowledgments {#acknowledgments .unnumbered} =============== The referees of a previous version of this article are kindly acknowledged for useful suggestions and comments. Simon tensor for stationary vacuum spacetimes ============================================= If $h_{ab}$ is the spatial metric, namely the Riemannian metric on the quotient 3-space of a stationary spacetime by its Killing trajectories, then a convenient conformal rescaling $\tilde h_{ab} = M^{2} h_{ab}$ by the absolute value of the norm of the timelike Killing vector field $M^2=-\xi_\alpha \xi^\alpha >0$ (where it remains timelike) leads to simpler formulas for the spacetime Ricci tensor. Like the original definition of Simon, the formula due to Perjes [@Perjes85] quoted as Eq. (18.70) in [@Kramer] relies on this conformal metric to express the Simon tensor in terms of the Ernst potential. We denote this tensor by SimonES (for the book title Exact Solutions) with a tilde (since its indices should be raised and lowered with $\tilde h_{ab}$) to distinguish it from our Simon tensor defined slightly differently using the original spatial metric $$\fl\qquad \label{simonES} \tilde{\rm S}{\rm imonES}^b{}_a =[\Re e\, \mathcal{E}]^{-2} \tilde\eta{}^{bcd} [ \mathcal{E}_c \tilde\nabla_d \mathcal{E}_a -\tilde h_{a c}\tilde h^{f g} \mathcal{E}_{[d} \tilde\nabla_{g]} \mathcal{E}_f ] \ ,$$ where $\mathcal{E}_a=\nabla_a \mathcal{E}=\tilde\nabla_a \mathcal{E}=\tilde\mathcal{E}_a$, $\Re e\, \mathcal{E} = -M^2$, and all the metric and duality operations refer to the conformally rescaled 3-metric. (In this appendix, $\tilde\nabla$ refers to the covariant derivative associated with the metric $\tilde h_{ab}$.) Apart from an extra factor of 2, Krish [@Krish] also uses this definition. One easily sees that this tensor is symmetric, and since the second term is antisymmetric, it must exactly cancel the antisymmetric part of the first term, so only symmetric part of the first term remains and that is just the Scurl operator acting on the gradient of the Simon potential apart from the scalar multiplier, which is automatically tracefree. Thus the Simon tensor is much more efficiently represented by a formula which makes these properties obvious $$\begin{aligned} \fl\quad \tilde {\rm S}{\rm imonES}_{ab} &=&[\Re e\, \mathcal{E}]^{-2} \tilde\eta^{cd}{}_{(a} \tilde\nabla_{\|c\|} \tilde\mathcal{E}_{b)} \tilde\mathcal{E}_d \equiv -[\Re e\, \mathcal{E}]^{-2}[\tilde\nabla \mathcal{E} \tilde\times \tilde\nabla (\tilde\nabla \mathcal{E})]_{ab} \nonumber \\ \fl\quad &=& [\Re e\, \mathcal{E}]^{-2} [\tilde{\rm S}{\rm curl} (\tilde\nabla\mathcal{E} \otimes \tilde\nabla \mathcal{E})]_{ab}\ ,\end{aligned}$$ where the last equality follows from the gradient identity (\[scurlXX\]) in the conformal geometry. Finally one must use the conformal transformation properties of the Scurl operator to re-express this in terms of the original spatial metric. For a symmetric covariant tensor $X_{ab} = \nabla_a \mathcal{E} \nabla_b \mathcal{E}$ which does not transform $\tilde X_{ab}= X_{ab}$, one finds \_[ab]{} = \[(M\^[-1]{} X)\]\_[ab]{} , so \_[ab]{} = \[e \]\^[-2]{} \[[Scurl ]{} (M\^[-1]{} )\]\_[ab]{} = 4 M\^[-1]{} [SimonMars]{}\_[ab]{} , where the last equality follows from Eq. (\[simonmarsvacuum\]). Simon introduced a well-known special transformation $w=(1-\mathcal{E})/(1+\mathcal{E})$ of the Ernst potential which has the property that it only changes the Scurl combination above by an overall factor (w w) = 14 (1+w)\^4 () . As discussed in appendix B, this is actually true for any function of the Ernst potential. The spatial dual of the original definition of Simon C\_[bcd]{} = 4 (1-w\|w)\^[-2]{} (\_[\[d]{} w \_[c\]]{} \_b w - h\_[b\[c]{} U\_[d\]]{}) ,U\_d= h\^[fg]{} \_[\[d]{} w \_[g\]]{} \_f w , can therefore be rewritten in the same way as the above definition $$\begin{aligned} \fl \frac12\tilde\eta_a{}^{cd}\tilde C_{bcd} &=& 2(1-w\bar w)^{-2} \tilde\eta_a{}^{cd} (\tilde\nabla_{[d} w \tilde\nabla_{c]} \tilde\nabla_b w - \tilde h_{bc} \tilde h{}^{fg} \tilde\nabla_{[d} w \tilde\nabla_{g]} \tilde\nabla_f w ) \nonumber\\ \fl &=& 2(1-w\bar w)^{-2} [\tilde{\rm S}{\rm curl}\, (\tilde\nabla w \otimes\tilde\nabla w) ]_{ab} \nonumber\\ \fl &=& 2(1-w\bar w)^{-2}(1+w)^4 [\tilde{\rm S}{\rm curl}\, (\tilde\nabla \mathcal{E} \otimes\tilde\nabla \mathcal{E})]_{ab} \nonumber\\ \fl &=& 2(1-w\bar w)^{-2}(1+w)^4 [\Scurl(M^{-1}\nabla \mathcal{E} \otimes\nabla \mathcal{E})]_{ab} \ .\end{aligned}$$ The Scurl formula for the Simon tensor expressed in terms of the rescaled metric $\tilde h_{ab}$ has an obvious immediate consequence for the interpretation of its vanishing. Because of the Scurl identity (\[scurlXXgrad\]) in the conformally rescaled geometry, the differential condition $ \tilde{\rm S}{\rm curl} (\tilde\nabla\mathcal{E} \otimes \tilde\nabla \mathcal{E}) = 0 $ is equivalent to the algebraic condition () = 0 , which in turn expresses the “alignment" of the two tensors $\tilde \nabla (\tilde \nabla \mathcal{E})$ and $\tilde \nabla \mathcal{E}\otimes \tilde \nabla \mathcal{E}$, i.e. \^[(TF)]{} \^[(TF)]{} . This geometrical meaning has been overlooked perhaps because of the unnecessary complication in previous formulas. For the electrovac case, the conformal transformation formula for a vector divergence X = M\^[-3]{} (MX) helps show the equivalence of Eq. (\[ernst\]) with the original Ernst equation (18.38) of [@Kramer]. Using this formula and substituting for $z$ in terms of $\mathcal{E}$ and $\Phi$ by solving Eq. (\[eq:ernstpot\]) in Eq. (\[ernst\]), one obtains -M\^[-2]{} h\^[ab]{}(\_a+2\|\_a) \_b =0 . Scurl and cross product identities ================================== In this appendix we use the simpler notation $\nabla$ for the spatial covariant derivative and assume all fields are spatial and stationary. A straightforward computation verifies the identity \[scurlXX\] (XX) = 32 \[(X X)\] - \[X(X)\] for any vector field $X$. When the vector field is a gradient $X=\nabla f$, then $\curl X = \curl(\nabla f)=0$ and the covariant derivative $\nabla X = \nabla(\nabla f)=\SYM (\nabla(\nabla f))$ is symmetric, so this reduces to the useful identity (see Eq (3.19) of [@cys1]) \[scurlXXgrad\] (XX) = - X(X) X=f . Notice that for any function $F(f)$ of $f$, then $\nabla F(f) = F^\prime(f) \nabla f$ and one has the property that the rate of change factor passes through the Scurl operation without extra terms, expanding the Scurl by the product rule for a scalar and a tensor $$\begin{aligned} \fl\qquad && \Scurl(\nabla F \otimes \nabla F) =\Scurl(F^{\prime 2} \nabla f \otimes \nabla f) \nonumber\\\fl\qquad &&= (F^{\prime 2})^\prime \nabla f \times (\nabla f \otimes \nabla f) + F^{\prime 2} \Scurl(\nabla f \otimes \nabla f) = F^{\prime 2} \Scurl(\nabla f \otimes \nabla f)\end{aligned}$$ because of the identity $Y\times (Y\otimes Y) = 0$ valid for any vector field $Y$. When performing actual calculations in this approach, one needs to refer to a number of Scurl and cross product identities involving vector fields and symmetric tensors where the symmetric tensors themselve arise from vector fields. For arbitrary spatial vector fields $X$, $Y$ and $Z$ one finds by straightforward evaluation $$\begin{aligned} \fl\quad \Scurl [\SYM (X\otimes Y)]&=& -\frac12 [X\times \SYM (\nabla Y)+Y\times \SYM (\nabla X)]+\nonumber \\ \fl\quad && +\frac34 \SYM [(X\otimes \curl Y+Y\otimes \curl X)]^{\rm (TF)} \ ,\label{scurlXY}\\ \fl\quad \Scurl [\SYM (\nabla X)]&=& - [X\times {}^{(3)}{\rm Ricci} ] +\frac12 \SYM [\nabla (\curl X)]^{\rm (TF)} \ ,\label{scurlnablaX}\\ \fl\quad \SYM [\nabla (X\times Y) ]^{\rm (TF)} &=& X\times \SYM (\nabla Y) - Y\times \SYM (\nabla X) +\nonumber \\ \fl\quad &&+\frac12 \SYM [X\otimes \curl Y-Y \otimes {\rm curl }X ]^{\rm (TF)} \ ,\label{scurlnablaXY}\\ \fl\quad \SYM [X \otimes (Y\times X)]&=&Y\times[X\otimes X] \ ,\label{XtimesYX}\\ \fl\quad X\times \SYM [X\otimes Y]&=& -\frac12Y\times [X\otimes X] \ ,\label{XtimesXoY}\\ \fl\quad X \times \SYM [Y\times Z]&=&\frac12 \SYM [Y\otimes (X\times Z)+Z\otimes (X\times Y)] \ .\end{aligned}$$ References {#references .unnumbered} ========== [00]{} Ellis G F R 1973 in [Cargese Lectures in Physics 6, Lectures at the International Summer School of Physics, Cargese, Corsica 1971]{} ed E Schatzman Gordon and Breach New York 1 Maartens R and Basset B A 1998 [Class. Quantum Grav.]{} [15]{} 705 Ellis G F R and van Elst H 1999 “Cosmological models (Cargèse lectures 1998)" in [Theoretical and Observational Cosmology]{} ed M Lachièze-Rey Kluwer Dordrecht 1 Simon W 1984 [Gen. Relativ. Grav.]{} [16]{} 465 Mars M 1999 [Class. Quantum Grav.]{} [16]{} 2507 Mars M 2000 [Class. Quantum Grav.]{} [17]{} 3353 Mars M 2001 [Class. Quantum Grav.]{} [18]{} 719 Fayos F and Sopuerta C F 1999 [Class. Quantum Grav.]{} [16]{} 2965 Fayos F and Sopuerta C F 2001 [Class. Quantum Grav.]{} [18]{} 353 Bini D, Jantzen R T and Miniutti G 2001 [Class. Quantum Grav.]{} [18]{} 4969 Marklund M and Z. Perjés Z 1997 [J. Mat. Phys.]{} [38]{} 5280 Cotton E 1899 [Ann. Fac. Sci. Toulouse (II)]{} [1]{} 385 Hawking S W and Ellis G F R 1973 [The Large Scale Structure of Space-Time]{} Cambridge University Press Cambridge Misner C W Wheeler J A and Thorne K S 1973 [Gravitation]{} Freeman San Francisco Stephani H Kramer D MacCallum M Hoenselaers C and Herlt E 2003 [Exact Solutions of Einstein’s Field Equations: Second Edition]{} Cambridge University Press Cambridge Debever R 1964 [Le Rayonnement Gravitationnel]{} Universite Libre de Bruxelles Taub A H 1966 “The Riemann-Christoffel Tensor and Tetrad and Self-Dual Formalisms" in [Perspectives in Geometry and Relativity]{} ed B Hoffmann Indiana University Press 360 Papapetrou A 1966 [Ann. Inst. H. Poincare’]{} [A4]{} 83 Demiański M and Newman E T 1966 [Bulletin de l’Académie Polonaise des Sciences]{} [XIV]{} 653 Miller J G 1973 [J. Math. Phys.]{} [14]{} 486 Newman E Tamburino L and Unti T 1963 [J. Math. Phys.]{} [4]{} 915 McGuire P and Ruffini R 1975 [Phys. Rev.]{} [12]{} 3019 Dadhich N and Turakulov Z Ya 2002 [Class. Quantum Grav.]{} [19]{} 2765 Misner C W 1963 [J. Math. Phys.]{} [4]{} 924 Misner C W and Taub A H 1968 [Zh. Eksp. Teor. Fiz.]{} [55]{} 233 \[English translation: [Sov. Phys.- JETP]{} [55]{} 122 1969\] Bonnor W B 1969 [Proc. Camb. Philo. Soc.]{} [66]{} 145 Feinblum D A 1970 [J. Math. Phys.]{} [11]{} 2713 York Jr J W 1972 [Phys. Rev. Lett.]{} [26]{} 1656 Perjés Z 1985 “Improved Characterization of the Kerr Metric" in [Proceedings of the Third Seminar on Quantum Gravity (Moscow, 1984)]{} eds M A Markov V A Berezin and V P Frolov World Scientific Singapore Krish J P 1988 [J. Math. Phys.]{} [29*]{} 446	{ "pile_set_name": "ArXiv" }
INTRODUCTION {#sec:introduction} ============ The subject of exchange-correlation-induced phase transitions has proven to be a rich field of research which has revealed many intriguing phenomena. Normal Fermi-liquid-state instabilities occur when the kinetic energy of the particles in a quantum system is of the same order or smaller than the inter-particle exchange and correlation energies. The instabilities lead to a variety of new electronic states. In Coulombic systems, this situation obtains at low densities and instabilities are more likely in systems of reduced dimensionality or, especially in two dimensions, in an applied magnetic field. Perhaps the best known examples of such new states occur in the in the fractional-quantum-Hall, strong-field regime of two-dimensional electronic systems.[@fqhe] The interest in exchange-correlation-induced phase transitions can be traced in part to the availability of high-quality semiconductor quantum wells, quantum wires, and superlattices. These artificial structures may be fabricated with remarkable precision and quality and possess electron densities that can be varied over a large range through a combination of modulation doping and judicious gating. The one- or two-dimensional nature of the resulting electronic systems, and the low densities realizable in devices of this kind, make them ideal for studies of interaction-driven physics. Another feature of these devices of importance to the current work is the extra degree of freedom available when multiple layers are present, as in multiple quantum wells or superlattices. This new degree of freedom allows transitions to states with order not only in the intra-layer electronic degrees of freedom but also in diagonal or off diagonal inter-layer charge [@MacDonald; @Ruden; @nei; @Varma; @Katayama; @Ying; @Patel; @Conti; @Zheng; @Wen-Zee_double; @ezawa; @iu_double] or spin[@Zheng; @das-tam; @rad-das; @Pinczuk; @Zheng-Rad-Das; @Plaut] observables. In particular, a great deal of attention has recently focussed on the possibility of inter-layer spin ordering in wide single or double quantum wells. This attention is motivated by studies of quantum well structures in which the lowest two subbands are well separated in energy from the higher subbands and the density is sufficiently low so that only these two subbands are occupied. In these structures, an earlier theoretical calculation of the collective spin-density excitations (SDEs) in the absence of a magnetic field showed a complete softening of the inter-subband SDE in a range of densities around that at which the second subband begins to populate.[@das-tam] Subsequent analysis indicated that this softening corresponded to a phase transition to a state in which each well in the double quantum well (or the electron gases on each side of a wide single quantum well) was spin polarized with the polarization vectors antiparallel; i.e., a transition to an antiferromagnetic order in the well spin densities.[@rad-das] Further work predicted that a similar transition to a canted antiferromagnetic phase should occur in the presence of a magnetic field at $\nu = 2$,[@Zheng-Rad-Das] and this transition seems to have been observed experimentally.[@Pinczuk] However, the predicted transition in the zero-field case has not yet been observed.[@Plaut] Several possibilities exist which may explain the absence of the zero-field antiferromagnetic phase in these experiments. The calculations predicting this phase[@rad-das] are based on a mean-field treatment of the interacting system, which is known to overestimate the densities and temperatures at which such symmetry-breaking transitions occur. In addition, electronic scattering by disorder or impurities generally has a detrimental effect on correlation-induced phases. Both of these difficulties are exacerbated by the low dimensionality of the double quantum well system considered. Thus, the zero-field antiferromagnetic phase may exist but may not have been observed due to measurements made at temperatures which are too large in samples of insufficiently high quality. The fact that the $\nu = 2$ transition is observed[@Pinczuk] as predicted by the mean-field calculations[@Zheng-Rad-Das] does not contradict this point of view, since the magnetic field completely quenches the kinetic energy and makes the mean-field theory a controlled approximation to the interacting system. Alternative explanations for the absence of the zero-field phase lie in the structure of the theory itself,[@rad-das] which was derived to explore the qualitative features of the antiferromagnetic phase without considering several confounding effects which may nonetheless be important. In particular, the previous calculation[@rad-das] did not account for interactions between electrons in the same subband, which should be of the same size or stronger than the interactions between electrons in different subbands which were included. This omission can become especially important when the intersubband excitation softens. These interactions could introduce ferromagnetic or charge-ordered phases into the model which are not probed by the current experiments. More seriously, the slight asymmetry present in any realistic double quantum well structure will couple the intra- and inter-subband SDEs, potentially preventing the latter from softening. As this softening was expected to be a hallmark of the antiferromagnetic phase transition,[@das-tam; @rad-das] it seemed reasonable to hypothesize that any asymmetry in the structure might suppress the antiferromagnetic phase entirely. In this paper, we address the question of the robustness of the zero-field antiferromagnetic phase in the presence of intra-subband interactions and asymmetry in the double quantum well structure. First, we argue on general grounds that the antiferromagnetic phase is a direct consequence of the importance of the intra-well exchange interaction at low densities and is stabilized by the inter-well hopping, which leads to an effective superexchange interaction. Thus, this phase should obtain in a suitably constructed heterostructure. Second, we extend the self-consistent Hartree-Fock calculations of earlier work[@rad-das] to include both intra- and inter-subband matrix elements of the model interaction and the effects of an asymmetric double quantum well. These calculations demonstrate that, while the intra-subband interaction does introduce ferromagnetic phases and asymmetry does reduce the region of the phase diagram occupied by the antiferromagnetic phase, the antiferromagnetic phase does not disappear. By examining the collective mode spectrum in the asymmetric structure, we also find that the inter-subband SDE does not soften due to the mode coupling between the intra- and inter-subband SDEs. The antiferromagnetic transition nevertheless occurs as a result of the collapse of the [intra]{}-subband SDE which, through the mode-coupling, has a strongly antiferromagnetic character. The outline of this paper is as follows. In the next Section, we employ a simple model for two weakly coupled two-dimensional electron gases to examine the energetics of the antiferromagnetic transition. Sec. \[sec:formalism\] contains the formalism for the extended self-consistent Hartree-Fock theory used in the remainder of the paper. This formalism includes asymmetry and all matrix elements of the interaction and is used to compute both the ground state and collective mode properties in what follows. In Sec. \[sec:results\], we present the results of our computations for the matrix elements, phase diagram, and collective modes in this model and discuss their implications. Finally, in Sec. \[sec:conclusion\], we summarize our results and conclude. ORIGIN AND STABILITY OF THE ANTIFERROMAGNETIC PHASE {#sec:origin} =================================================== In this Section, we examine a simple model of a double quantum well in order to extract the basic physics underlying the zero-field antiferromagnetic phase. To that end, consider two two-dimensional electron gases separated by a barrier constructed so that the interaction between them is negligible. This would be the case if the two-dimensional layers were widely separated or the barrier were very high and if the dielectric constant of the barrier were large. Suppose the electrons in each layer move freely except for a Hubbard-like point-contact interaction $V({\bf r}) = V_0 \delta({\bf r})$, as used in textbook treatments of itinerant magnetism.[@Doniach] For equal charge density in the two layers, the Hartree-Fock energy $E_{\rm HF}$ of the two-layer system can be derived following Refs. and may be expressed in the form $$\frac{E_{\rm HF}}{A} = \frac{n^2}{8 N_0}\,(1-N_0 V_0) \left( \frac{1 + m_1^2}{2} + \frac{1 + m_2^2}{2} \right) + \frac{1}{4} V_0 n^2, \label{eq:Ehf}$$ where $n$ is the total electronic density in both layers, $N_0$ is the single-spin, two-dimensional density of states, and $$m_i = \frac{n_{i\uparrow}-n_{i\downarrow}} {n_{i\uparrow}+n_{i\downarrow}} \label{eq:m}$$ is the relative spin polarization in layer $i = 1,2$ with partial spin-dependent densities $n_{i\sigma}$. In this equation, the first term represents the contribution to the energy from combined kinetic and exchange effects, and the last term is the Hartree contribution. As is clear from Eq. (\[eq:Ehf\]), when $N_0 V_0 > 1$, it is energetically favorable for both layers to acquire a spontaneous spin polarization $\|m_1\| = \|m_2\| = 1$, while for $N_0 V_0 < 1$, the layer remain unpolarized. This result is simply the Stoner criterion for 2D itinerant magnetism.[@Doniach] This description of spontaneous spin-polarization in two-dimensional electron systems is unrealistic both in the use of a point-contact interaction and in the use of the Hartree-Fock approximation. Our objective in this section is to obtain a qualitative understanding of the influence of weak electronic tunneling between the layers on a double-layer system when isolated single layer systems are close to their ferromagnetic instabilities. We postpone a realistic discussion of the system parameters for which the physics we address in this paper is likely to realized to Sec. \[sec:conclusion\]. For a sufficiently strong intra-layer repulsion, then, the exchange interaction forces both layers to spin polarize, but the relative orientation of the polarizations is unknown. For simplicity, let us restrict our attention to two possibilities for the relative orientation: parallel (ferromagnetic) alignment or antiparallel (antiferromagnetic) alignment. With the spin unpolarized (paramagnetic) phase, the three possible phases for the two-layer system are shown in Fig. \[fig:phases\]. In the absence of any inter-layer coupling, the ferromagnetic and antiferromagnetic phases are degenerate, but this situation changes if we introduce a small amplitude for hopping between the layers. Writing $c_{i{\bf k}\sigma}$ for the annihilation operator of an electron in layer $i=1,2$ with two-dimensional wave vector ${\bf k}$ and spin projection $\sigma$, this inter-layer hopping is governed by the Hamiltonian $$H_{\bot} = -\frac{\Delta_0}{2} \, \sum_{{\bf k}\sigma} \left[ c^{\dag}_{1{\bf k}\sigma} c^{~}_{2{\bf k}\sigma} + {\rm h.c.} \right] \label{eq:Hbot}$$ where $\Delta_0$ is the splitting between the symmetric and antisymmetric single-particle eigenstates. The leading order change in the ground state energy due to $H_{\bot}$ can be calculated using linear response theory, and is proportional to the transverse pseudospin susceptibility of double-layer systems defined in Ref. . For the present model with no inter-layer interactions and contact intra-layer interactions, the time-dependent Hartree-Fock approximation for the susceptibility gives $$\frac{\delta E }{A} = \left\{ \begin{array}{ll} -\frac{1}{2} N_0 \Delta_0^2 & {\rm (paramagnetic)} \\ -\frac{1}{4} N_0 \Delta_0^2 & {\rm (ferromagnetic)} \\ -\frac{1}{2} \Delta_0^2/V_0 & {\rm (antiferromagnetic)} \end{array} \right. \label{eq:Ebot}$$ The fact that the results for paramagnetic and ferromagnetic states are independent of the interaction strength is a special property of the present model related to both the absence of interlayer interactions and the wave vector independence of the exchange self-energy. We see that, in all cases, introducing inter-layer hopping reduces the ground state energy, as one might expect when the confinement of the electrons to the layers is weakened. Comparing the energies of the ferromagnetic and antiferromagnetic phases, we also observe that the degeneracy between these phases is broken by the hopping term. Specifically, the antiferromagnetic phase is found to be more stable than the ferromagnetic phase if $N_0 V_0 < 2$, implying that the inter-layer hopping opens up a region of antiferromagnetic order between the paramagnetic and ferromagnetic phases. The mechanism for this stabilization can be deduced from the form of $\delta E / A$ to be a superexchange interaction; that is, an electron is able to hop from one layer to the other and back at the cost of a Hubbard energy in the intermediate state, yielding an energy savings of $\Delta_0^2 / 2 V_0$. A similar mechanism is blocked by the Pauli exclusion principle in the ferromagnetic phase , since the hopping Hamiltonian preserves spin \[Eq. (\[eq:Hbot\])\]. We argue that this mechanism favoring an antiferromagnetic arrangement of the ordered moments in the two layers will be dominant in most circumstances. The calculations described below indicate that an effective superexchange interaction between the wells should stabilize an antiferromagnetic phase for moderate interaction strengths. This explanation of the zero-field antiferromagnetic phase suggests why the presence of asymmetry and additional interaction matrix elements may not eliminate this phase: both the two-dimensional ferromagnetism within each well and the superexchange interaction between the wells should be fairly insensitive to these perturbations. In the remainder of the paper, we perform a more detailed self-consistent Hartree-Fock calculation to support this statement and to explore the consequences of these perturbations. FORMALISM {#sec:formalism} ========= To accomplish the goal of investigating matrix element and asymmetry effects, we employ an extension of the point-contact model described in Ref. . In the original calculation, the full Coulomb interaction between the electrons in the double quantum well was approximated by a delta function in real space with only the matrix elements of this interaction between the lowest two subbands being kept and the remainder being set to zero.[@rad-das] The use of a delta-function or point-contact interaction is based primarily on a desire to create a simple, solvable model which mimics the qualitative features of the fully interacting system. A quantitative theory would require that the interaction be made realistic, and also that the interactions be treated more accurately than in the Hartree-Fock approximation. In practise this would require quantum Monte-Carlo calculations of some type, which would involve an enormous amount of effort and would not be able to address the excitation spectrum which provides the experimental signature for the state we are proposing. We therefore maintain the point-contact form of the interaction here. However, there is no reason to set matrix elements of this interaction other than those between the lowest subbands to zero, as in previous work. Indeed, we shall see below that these other matrix elements are of the same order as the inter-subband ones. Thus, we shall include all the matrix elements between the lowest two subbands in our calculations. In addition to the issue of intra-subband repulsion, we would also like to study the effects of quantum well asymmetry on the phase diagram and collective modes of this system. This asymmetry arises in real quantum wells through alloy fluctuations across the profile of the well or fluctuations in the well thickness, and we model it by allowing one of the wells to be deeper than the other, as illustrated in Fig. \[fig:dqw\]. Although we have assumed the effective asymmetry to enter through the well depth, the results of our calculations for systems with asymmetric well widths should be qualitatively similar. We therefore consider a three-dimensional electron gas confined along the $z$-direction by a potential $V_{CONF}(z)$ of the type shown in Fig. \[fig:dqw\] which interacts through a [three dimensional]{} point-contact potential $V({\bf R}) = V \delta({\bf R})$. In the absence of the interaction $V({\bf R})$, the electronic eigenstates are given by the solution of the time-independent Schrödinger equation $$\left[ -\frac{\hbar^2}{2m^} \frac{d^2}{dz^2} + V_{CONF} (z) \right] \xi_n (z) = \epsilon_n \xi_n (z), \label{eq:xidef}$$ where $m^$ is the effective mass, which is assumed to be constant throughout the heterostructure. Normalizing these eigenfunctions by $$\int dz \left\| \xi_n(z) \right\|^2 = 1, \label{eq:norm}$$ we write the electron annihilation operator $\psi_{\sigma} ({\bf R})$ as $$\psi_{\sigma} ({\bf R}) = \frac{1}{\sqrt{A}} \sum_{n{\bf k}} e^{i{\bf k \cdot r}} \xi_n (z) c_{n{\bf k}\sigma}, \label{eq:basis}$$ where ${\bf R} = ({\bf r},z) = (x,y,z)$, ${\bf k} = (k_x, k_y)$, $A$ is the transverse area of the sample, and $c_{n{\bf k}\sigma}$ annihilates a quasiparticle in subband $n$, of transverse wave vector ${\bf k}$, and with spin projection $\sigma$ (these conventions will be used throughout this paper). Defining a composite subband and spin index $a = (n_a, \sigma_a)$ with summation over repeated indices implied, the Hamiltonian for this system is $$\begin{aligned} H &=& \sum_{\bf k} \, \epsilon_{a{\bf k}}^{~} \, c^{\dag}_{a {\bf k}} c^{~}_{a {\bf k}} \nonumber \\ &&+ \frac{1}{2 A} \, \sum_{{\bf k} {\bf k'} {\bf q}} \, V_{ad,bc} \, c^{\dag}_{a {\bf k + q}} c^{\dag}_{b {\bf k' - q}} c^{ }_{c {\bf k'}} c^{ }_{d {\bf k}} . \label{eq:H}\end{aligned}$$ Here, the quasiparticle energy $$\epsilon_{a{\bf k}} = \epsilon_n + \frac{\hbar^2 k^2}{2m^} - \mu \label{eq:ea}$$ is measured with respect to the chemical potential $\mu$ and the matrix elements of the interaction are $$\begin{aligned} V_{ab,cd} &=& \delta_{\sigma_a \sigma_b} \delta_{\sigma_c \sigma_d} \, V \, \int dz \,\xi_{n_a}^{} (z) \xi_{n_b}^{~} (z) \, \xi_{n_c}^{} (z) \xi_{n_d}^{~} (z) . \label{eq:V}\end{aligned}$$ We treat this Hamiltonian within self-consistent Hartree-Fock theory allowing for the possibility of phases with broken symmetry in subband and spin indices, but imposing translational invariance within each layer. The electronic Green’s function in the interacting system can therefore be written $$G_{ab} (k_n) = - \int_0^{\beta} d\tau \, e^{i\omega_n \tau} \, \left< T_\tau \left[ c_{a{\bf k}} (\tau) c_{b{\bf k}}^{\dag} (0) \right] \right>, \label{eq:G}$$ where $k_n = ({\bf k},i\omega_n)$, $\beta = 1 / T$ ($\hbar = k_B = 1$ throughout this paper), and the rest of the notation is standard.[@Mahan] This Green’s function is determined self-consistently from the self-energy in the Hartree-Fock approximation, $$\Sigma_{ab} = \left[ V_{ab,dc} - V_{ac,db} \right] \, \frac{T}{A} \sum_{k^{\prime}_m} e^{-i \omega_m 0-} \, G_{cd} (k^{\prime}_m), \label{eq:Sigma}$$ and the Dyson equation, $$\left[ \left( i\omega_n - \epsilon_{a{\bf k}} \right) \delta_{ab} - \Sigma_{ab} \right] \, G_{bc} (k_n) = \delta_{ac}, \label{eq:Dyson}$$ under the constraint of constant sheet density $N_s$, $$N_s = \frac{T}{A} \sum_{k_m} e^{-i \omega_m 0-} \, G_{aa} (k_m) , \label{eq:filling}$$ which determines the chemical potential. We solve these equations in the following way. Given a double quantum well structure defined by the effective mass $m^$, sheet density $N_s$, and the structural parameters shown in Fig. \[fig:dqw\], we compute the eigenfunctions and eigenenergies by solving Eq. (\[eq:xidef\]) with the normalization condition Eq. (\[eq:norm\]). These eigenfunctions are used to compute the matrix elements of the interaction via Eq. (\[eq:V\]) in terms of a single parameter V. The resulting matrix elements and eigenenergies are employed to solve Eqs. (\[eq:G\])-(\[eq:filling\]) self-consistently at $T = 0$ following the procedure outlined in Sec. II.B of Ref. and including [all]{} matrix elements of the interaction between the lowest two subbands. This procedure yields the ground state properties of the system as a function of the sheet density $N_s$, the interaction parameter $V$, and the structural parameters of the double quantum well \[Fig. \[fig:dqw\]\]. To illuminate the properties of the interacting system further, we also compute the generalized density-density response function, which is defined by the analytic continuation of $$\begin{aligned} \Pi^{\mu} ({\bf Q},i\nu_n) &=& - \int \frac{d{\bf R}}{\cal V} \, e^{-i{\bf Q \cdot R}} \, \int_0^{\beta} d\tau \, e^{i\nu_n\tau} \times \nonumber \\ && \left< T_{\tau} \left[ \rho^{\mu}({\bf R},\tau) \, \rho^{\mu}({\bf 0},0) \right] \right> \label{eq:Pi}\end{aligned}$$ to real frequencies. In this expression, ${\cal V}$ is the system volume and the generalized density operator is $$\rho^{\mu}({\bf R}) = \sum_{\sigma\sigma'} \, \psi^{\dag}_{\sigma}({\bf R}) \sigma^{\mu}_{\sigma\sigma'} \psi^{~}_{\sigma'}({\bf R}) \label{eq:dop}$$ with $ \psi_{\sigma}({\bf R})$ given by Eq. (\[eq:basis\]) and $\sigma^{\mu} = (1,\sigma^x,\sigma^y,\sigma^z)$ are the Pauli matrices. This response function is computed from the non-interacting response function in subband and spin space within the conserving approximation described in Sec. V.A of Ref. but with the inclusion of all interaction matrix elements. In addition to the information this response function reveals about the excitations of the interacting system, its imaginary part is proportional to the intensity observed in resonant inelastic light scattering measurements,[@pin-abs; @das:ils] allowing us to make contact with experiment. This is particularly relevant here, because searches for the antiferromagnetic phase in both finite[@Pinczuk] and zero[@Plaut] magnetic field have employed this technique. RESULTS {#sec:results} ======= In this Section, we apply the formalism described in Sec. \[sec:formalism\] to compute the ground state phase diagram and collective modes in a typical GaAs/Al$_x$Ga$_{1-x}$As double quantum well structure which is expected to exhibit the zero-field antiferromagnetic instability. The structure has a well width of 140 Å, a barrier width of 30 Å, and a well depth of 220 meV \[cf. Fig. \[fig:dqw\]\], and an electronic effective mass $m^* = 0.067\,m_e$. For the moment, we leave the asymmetry unspecified. Matrix Elements {#sec:me} --------------- As a first step in obtaining the phase diagram for this structure, we must solve the time-independent Schrödinger equaution \[Eq. (\[eq:xidef\])\] for the lowest two eigenfunctions $\xi_n (z)$ and eigenenergies $\epsilon_n$ at a fixed value of the asymmetry parameter $dU$ \[cf. Fig. \[fig:dqw\]\] and then compute the matrix elements of the interaction through Eq. (\[eq:V\]). Solving Eq. (\[eq:xidef\]) is straightforward and yields a splitting of the non-interacting eigenstates of $\Delta^0_{\rm SAS} = \epsilon_2 - \epsilon_1 = 2.25$ meV for $dU = 0$. As $dU$ is increased, this splitting increases to a maximum of 18.5 meV at $dU = 9.4$ meV; for larger $dU$, the lowest eigenfunction is localized in one well. Since the structures examined experimentally have subband splittings on the order of 1 meV,[@Pinczuk; @Plaut] we restrict our attention to small $dU$ values. The dependence of the matrix elements on the asymmetry parameter is somewhat more interesting and merits a brief discussion. Since our model interaction is a delta function in real space and we have chosen the wave functions to be real, the matrix elements defined by Eq. (\[eq:V\]) are invariant under permutation of the indices. Thus, there are only five independent matrix elements, $V_{11,11}$, $V_{22,22}$, $V_{11,22}$, $V_{11,12}$, and $V_{22,21}$, which are displayed in Fig. \[fig:me\] as a function of the asymmetry parameter $dU$. In a symmetric double quantum well ($dU = 0$), the $V_{11,12}$ and $V_{22,21}$ matrix elements vanish by symmetry, but the remaining inter-subband ($V_{11,22}$) and intra-subband ($V_{11,11}$ and $V_{22,22}$) matrix elements are equal to within 5 %. We remark that for more realistic interaction models $V_{11,22}$, which is roughly[@boebinger] proportional to the difference of intra-layer and inter-layer interactions, is weaker than $V_{11,11}$ and $V_{22,22}$ which are roughly proportional to the sum. The latter matrix elements would be equal if the electrons were localized to the wells; the fact that they are nearly so indicates that the wave function overlap between the wells is small. Additionally, this calculation provides direct evidence that the neglect of the intra-subband repulsion employed in earlier work[@rad-das] is not generally justified for these double quantum well structures. We shall see, however, that their inclusion in the calculation changes the qualitative picture only slightly. As the asymmetry is increased from zero, we discern several features. We see that the intra-subband matrix elements $V_{11,11}$ and $V_{22,22}$ are approximately equal and increase with increasing $dU$ to saturate at a value about twice the $dU = 0$ one. The approximate equality of these diagonal matrix elements follows from the normalization condition, Eq. (\[eq:norm\]), imposed on the two eigenfunctions. The increase in these matrix elements with $dU$, on the other hand, can be attributed to the increasing confinement of the wave functions of the two subbands to opposite wells, similar to what occurs in the presence of an applied electric field.[@Katayama; @Ying] Thus, at large $dU$, the two subband wave functions are almost completely localized in opposite wells, enhancing the magnitude of the diagonal matrix elements. For the same reason, the inter-subband matrix element $V_{11,22}$ decreases with increasing asymmetry: as the wave functions from different subbands are increasingly localized, their overlap, and hence $V_{11,22}$, decreases to zero. Since the zero-field antiferromagnetic transition depends on this matrix element, it is clear that large asymmetry is inimical to this phase. Based on the small values of the observed splitting of the lowest two subbands in the experimental samples,[@Pinczuk; @Plaut] however, we expect the actual samples to be in a regime in which $V_{11,22}$ is still non-negligible. Hence, the antiferromagnetic phase is not immediately excluded. Finally, we note that the mixing terms $V_{11,12}$ and $V_{22,21}$ have opposite signs and increase in magnitude with $dU$ to a maximum around $dU = 1$ meV. The wave function for the lowest ($n = 1$) subband has no nodes and we chose it to be positive. Orthogonality requires that the $n = 2$ wave function have a node and we choose it to be negative in the well where $\|\xi_1(z)\|^2$ is largest; with this convention $V_{11,12}$ is negative and $V_{22,21}$ is positive, as observed in Fig. \[fig:me\]. In addition, since these matrix elements must vanish both in the symmetric ($dU = 0$) limit and when $dU$ is large and the subband wave functions are localized in different wells, the maximum seen in this figure is also expected. These results suggest that the mode coupling between intra- and inter-subband excitations induced by these matrix elements will be maximal around $dU = 1$ meV. Taken together, the behavior of the interaction matrix elements presented in Fig. \[fig:me\] indicate that the zero-field antiferromagnetic phase will probably not be stable against large asymmetry in the quantum wells. However, the current experimental samples have subband splittings more consistent with small asymmetry, and therefore these samples may be of high enough quality to observe this phase, at least in principle. To examine this situation further, we shall compute the phase diagram and collective modes for two choices of $dU$ in the weak asymmetry regime: $dU = 0$ (the symmetric case) and $dU = 0.5$ meV (the asymmetric case). The following subsections discuss the results of these calculations. Phase Diagram {#sec:phases} ------------- As described in the preceding subsection, the structure of the double quantum well yields the eigenenergies and eigenfunctions of the non-interacting system which are then used to compute the interaction matrix elements up to an overall factor $V$ \[Eq. (\[eq:V\])\]. To be consistent with earlier work,[@rad-das] we choose to parameterize the interaction strength by the magnitude of the inter-subband repulsion $V_{11,22} = V_{12}$ rather than by $V$, but all the matrix elements are uniquely determined by either parameter. With the structure and interaction strength fixed, the only other parameter in our model is the sheet density $N_s$. Given these parameters, Eqs. (\[eq:G\])-(\[eq:filling\]) can be solved at $T = 0$ to yield the interacting ground state of the system. The resulting phase diagrams in terms of the dimensionless interaction strength $N_0 V_{12}$ and sheet density $N_s / 2 N_0 \Delta^0_{\rm SAS}$ are presented in Fig. \[fig:phasediagram\] for the double quantum well structure of Fig. \[fig:dqw\]. In these figures, $N_0$ is the single-spin, two-dimensional density of states, and $\Delta^0_{\rm SAS} = \epsilon_2 - \epsilon_1$ is the subband splitting in the non-interacting system. Consider the symmetric ($dU = 0$) case first \[Fig. \[fig:phasediagram\](a)\]. At first glance, this phase diagram is very similar to the one obtained from the neglect of intra-subband repulsion,[@rad-das] even though this repulsion is included in our calculations. At low density, the paramagnetic phase with one spin-degenerate subband occupied (labeled $N_1$ in the figure) is stable, while at higher densities ($N_s / 2N_0\Delta^0_{\rm SAS} > 1$) but weak interaction ($N_0 V_{12} < 1/2$) the paramagnetic phase with two spin-degenerate subbands occupied (labeled $N_2$) is stable. For larger interaction strengths and intermediate densities, we see a broad region in which the antiferromagnetic ($AF$) phase obtains. Thus, the inclusion of intra-subband repulsion does not eliminate the zero-field antiferromagnetic phase from the phase diagram of the model, despite the fact that the intra- and inter-subband repulsions are of the same order \[Fig. \[fig:me\]\]. These matrix elements do have an effect on the phase diagram, however. At higher densities ($N_s/2N_0\Delta^0_{\rm SAS}$ larger than approximately 2) and strong interaction ($N_0V_{12} > 1/2$), two different ferromagnetic phases appear. In these phases, each quantum well is spin polarized, and the polarization vectors are parallel. They differ in the arrangement and filling of the interacting bands, as indicated by the labels in the figure: the phase $FM_i$ corresponds to an interacting band structure with $i$ spin-split subbands occupied. The presence of the ferromagnetic phases is expected when intra-subband repulsion in present; what is surprising is that the ferromagnetic phases do not exclude the presence of the antiferromagnetic phase. As argued in Sec. \[sec:origin\] based on a weak-coupling model, the antiferromagnetic phase is stabilized by a superexchange interaction for intermediate interaction strengths. Similar behavior is seen in Fig. \[fig:phasediagram\](a) based on our strong-coupling computations and presumably originates from the same mechanism. Note that, in the limit of vanishing hopping between the wells, $V_{12}$ may be associated with $V_0/2$ in the model of Sec. \[sec:origin\], implying that the phase boundary $N_0 V_{12} = 1/2$ in Fig. \[fig:phasediagram\](a) is nothing but the Stoner criterion for the related Hubbard model. The fact that the model of Sec. \[sec:origin\] does not have a closer correspondence to Fig. \[fig:phasediagram\](a) suggests that the relative magnitudes of the intra- and inter-subband interaction matrix elements, which are all equal in the model of Sec. \[sec:origin\], are important for determining whether ferro- or antiferromagnetic phases obtain in a specific region of the phase diagram. With the introduction of weak asymmetry ($dU = 0.5$ meV), the qualitative features of the phase diagram do not change, as seen in Fig. \[fig:phasediagram\](b). As before, we find paramagnetic phases at small interaction strengths, the antiferromagnetic phase at larger interaction strengths and intermediate densities, and ferromagnetic phases at larger interaction strengths and higher densities. The identification of these phases precisely matches those in the symmetric case of Fig. \[fig:phasediagram\](a), although the position of the phase boundaries have shifted somewhat. An apparently new feature occurs at low density and large interaction strength, where an $FM_1$ ferromagnetic phase has replaced the paramagnetic $N_1$ phase. However, this phase also occurs in the symmetric structure when $N_0 V_{12}>1$, but is cut-off in Fig. \[fig:phasediagram\](a). It corresponds to the usual ferromagnetic instability in a single, spin-degenerate band which obtains when the interaction is sufficiently strong.[@Doniach] Although weak asymmetry clearly does not destroy the antiferromagnetic phase, it does have observable consequences. The most noticeable effect of the asymmetry on the ground state is that the spin polarizations in the magnetic phases, which can be obtained from the expectation value of the density operator, Eq. (\[eq:dop\]), are no longer of equal magnitude in the wells. This is an obvious consequence of an asymmetric structure which nonetheless does not disturb the identification of ferro- and antiferromagnetic phases, since one can determine whether the spin polarizations are parallel or antiparallel without referring to their magnitudes. Collective Modes {#sec:modes} ---------------- The effects of asymmetry on the antiferromagnetic phase cannot be fully appreciated based on the ground-state properties alone, but must be augmented by an examination of the excitation spectrum of the system. We focus on the spin-density excitations in what follows, since they are the excitations most strongly coupled to the magnetic nature of the ground state and can also be probed experimentally by Raman scattering.[@pin-abs; @das:ils] For these calculations, we compute the spin-spin response function ($\mu = 3$ in Eq. (\[eq:Pi\])) for the appropriate ground state as discussed in Sec. \[sec:formalism\] and identify the collective modes by peaks in the imaginary part of this response function. Since this procedure is used in the Raman scattering measurements, our results have direct implications for experiment and we shall discuss them in this context. For concreteness, we fix the interaction strength and sweep the sheet density through the second-order transition from the $N_1$ to the $AF$ phase in these calculations \[cf. Fig. \[fig:phasediagram\]\]. As an introduction to the general phenomenology of spin-density excitations in double quantum wells, consider the symmetric ($dU = 0$) case first. By appropriately arranging the light scattering geometry, Raman scattering can selectively probe the intersubband spin-density excitations,[@Pinczuk; @Plaut] which, in our approximation, have the form shown in Fig. \[fig:symmode\](a). In addition to a continuum of inter-subband particle-hole excitations, there is a collective spin-density excitation (SDE) with a finite energy at $q = 0$ which disperses with increasing $q$ toward the particle-hole continuum. The magnitude of the $q = 0$ SDE energy is reduced from the subband splitting $\Delta_{\rm SAS}$ by vertex corrections appearing in the response function due the the exchange interaction.[@Ando-RMP; @pinczuk89] In addition, there is an intra-subband SDE which has a linear-in-q dispersion in our model, shown by the dashed line in Fig. \[fig:symmode\](a). If the well is symmetric, this mode will not appear in Raman spectra taken in a scattering geometry meant to observe inter-subband excitations. A symmetric system with identical quantum wells is invariant under inversion about the mid-point between the wells so that all states can be classified by a parity quantum number. Inter-subband excitations, which are odd, and intrasubband excitations, which are even, do not interact and can cross as seen in Fig. \[fig:symmode\](a). As the density increases in our model, the exchange-induced reduction in the $q=0$ inter-subband SDE energy increases until the mode softens entirely, as illustrated in Fig. \[fig:symmode\](b). This complete softening was seen initially in time-dependent, local density approximation calculations of the SDE spectrum in these systems and was the first evidence for the zero-field antiferromagnetic phase.[@das-tam] An analysis of the real-space spin response identified this softening as an antiferromagnetic transition of the well spin polarizations.[@rad-das] As the density is increased past the antiferromagnetic transition, the inter-subband SDE turns into the collective mode associated with amplitude fluctuations of the antiferromagnetic order parameter.[@rad-das] Experimentally, then, one expects to see a complete softening of the inter-subband SDE and the recovery of this amplitude mode as the density is tuned through the transition. The presence of asymmetry in the double quantum well complicates this picture somewhat. The SDE spectrum for our asymmetric ($dU = 0.5$ meV) double quantum well illustrates these complications and is shown in Fig. \[fig:asymmode\](a). Most noticeably, the asymmetry mixes the intra- and inter-subband excitations, so that even in scattering geometries designed to measure only inter-subband response, both intra- and inter-subband excitations will appear.[@note] This effect is seen through both an enlarged particle-hole continuum and the presence of a damped mode in the inter-subband spectrum corresponding to the intra-subband SDE. Furthermore, the asymmetry couples the intra- and inter-subband SDEs themselves, leading to an avoided crossing which may be seen by comparing Fig. \[fig:asymmode\](a) to Fig. \[fig:symmode\](a). This avoided crossing effectively prevents the inter-subband SDE from completely softening on entering the antiferromagnetic phase, as shown in Fig. \[fig:asymmode\](b). In this figure, ones sees that, as the density is tuned toward the antiferromagnetic transition, the $q=0$ inter-subband SDE energy decreases to a finite value. Further increase of the density into the antiferromagnetic phase reverses this trend, and the energy of the amplitude mode of the antiferromagnetic order parameter increases, resulting in a cusp. Despite the fact that the inter-subband SDE does not completely soften, the antiferromagnetic phase appears, demonstrating that this softening is a sufficient but not necessary signature of the zero-field phase transition. We expect some mode to soften at this transition, however, and it turns out that it is the [intra]{}-subband SDE which collapses. Specifically, the mode coupling between inter- and intra-subband SDEs pushes the latter mode down in energy at all $q$, effectively reducing the group velocity of this mode. Approaching the antiferromagnetic transition, the $q \rightarrow 0$ group velocity of the intra-subband SDE decreases until it vanishes at the transition point. The resulting phase has the character of the real-space spin density profile of the intra-subband mode, which direct calculation reveals to be antiferromagnetic. The character of the intra-subband excitation at small $q$ is therefore very similar to the inter-subband excitation in the symmetric well due to the mode coupling between the two excitations. Thus, its softening can lead to an antiferromagnetic transition without inconsistency. CONCLUSION {#sec:conclusion} ========== In this paper, we have examined the origin and robustness of a zero-field antiferromagnetic phase in double quantum wells predicted in previous calculations.[@das-tam; @rad-das] Based on a simple model, we determined that such magnetic phases are a direct consequence of the magnetism expected at low densities in two-dimensional systems where interaction effects dominate. In particular, magnetic phases with either ferro- or antiferromagnetic spin polarizations are possible, and a superexchange interaction between the wells leads to a region at intermediate densities and interaction strengths where the antiferromagnetic phase is preferred. By performing a detailed self-consistent Hartree-Fock calculation on a model of these double quantum well systems, we addressed two features left out of preceding work:[@rad-das] intra-subband repulsion and asymmetry of the heterostructure. Our results support and extend those of Ref. . Specifically, we found that the antiferromagnetic phase is stable in a large region of the model phase diagram despite the presence of intra-subband repulsion that is as strong as the inter-subband repulsion which drives the antiferromagnetic instability. The intra-subband repulsion does, however, introduce ferromagnetic phases, producing a rich phase diagram. Note, however, that no charge ordering phases of the type discussed in Refs. were observed, even though our formalism did allow for that possibility. In addition, both ferro- and antiferromagnetic phases persist in the presence of asymmetry in the quantum well structure; indeed, the phase diagram is qualitatively unaffected by its introduction. Asymmetry does have a strong influence on the collective mode spectrum, though, and induces a mode-coupling between inter- and intra-subband spin-density excitations which prevents the latter from completely softening at the antiferromagnetic phase transition. Instead, the inter-subband spin-density excitation exhibits a cusp at the transition while the $q \rightarrow 0$ group velocity of the [intra]{}-subband excitation vanishes. The coupling between these two modes nonetheless lends an antiferromagnetic character to the intra-subband excitation in the asymmetric system and enables the collapse of its group velocity to yield the antiferromagnetic phase. Taken together, these results strongly indicate that, if an exchange-correlation-induced ferromagnetic transition occurs in a single-layer system at sufficiently low density, then the antiferromagnetic phase in a double-layer system should also occur. However, the issue of whether or not the ferromagnetic phase obtains in a single two-dimensional layer is not settled. Variational Monte Carlo calculations[@cep-ald; @tan-cep] show the presence of a ferromagnetic phase between the paramagnetic and Wigner crystal phases, but Green’s function Monte Carlo computations[@tan-cep] find no intervening ferromagnetism. More recent numerical work based on the Monte Carlo technique[@cep-unp; @RS; @Conti] once again favors the existence of a ferromagnetic transition. Other calculations using density-functional theory in the local-spin-density approximation[@rad-das; @Zheng] also support the existence of a ferromagnetic transition at sufficiently low density. Thus, although a definitive demonstration of ferromagnetism in a single, two-dimensional layer is lacking, a large body of evidence exists which firmly supports this hypothesis. The precise conditions under which these phases would be observable are somewhat harder to elucidate based on the mean-field theory presented in this paper. The principle difficulty is that mean-field calculations will tend to overestimate the temperatures and densities at which exchange-correlation-induced phase transitions occur. The source of this difficulty lies in the neglect of fluctuations in the theory, which play an important role in the low-dimensional structures considered. If one goes to densities and temperatures deep below the critical values of these parameters, we expect that the mean-field theory will give an accurate qualitative picture of the phase.[@MW] In sum, work on the single layer system indicates that its ferromagnetic state is unlikely to occur for electron-gas density parameters smaller than $r_s \approx 10$, a much lower density than would be indicated by the Hartree-Fock approximation, for which the transition to the ferromagnetic state occurs at $r_s \approx 2$. The present work suggests that the two-layer antiferromagnetic phase, as well as two-layer ferromagnetic phases, are likely to be present in double-layer systems when the density per layer approaches the low value at which the single-layer ferromagnetic instability occurs. For the GaAs systems studied experimentally it therefore seems unlikely that the antiferromagnetic state will occur for densities per layer substantially larger than $ \approx 10^{10} {\rm cm}^{-2}$. However, it is exceedingly difficult to estimate the transition density theoretically and one must rely on experiment. Currently, a single experimental publication regarding a search for the antiferromagnetic phase has appeared in the literature,[@Plaut] and the results are equivocal. The authors of this study report inelastic light scattering measurements of the long-wavelength inter-subband collective spin-density excitations as a function of density in a double quantum well which was expected, on the basis of the original theoretical work, to show the antiferromagnetic instability.[@Plaut] Instead of completely softening at a finite density, as predicted in earlier work,[@das-tam; @rad-das] the inter-subband SDE shows no dramatic structure down to the lowest densities measured.[@Plaut] These results could be accounted for in at least two different ways. The most likely explanation is that the electron density of the sample, of order $10^{11}/{\rm cm}^2$,[@Plaut] is above the critical density for the antiferromagnetic transition. Alternatively, the calculations in this paper demonstrate that slight asymmetry in the double quantum well will prevent a complete softening of the inter-subband SDE and yield a cusp as a function of density. Since the energy of the amplitude mode in the antiferromagnetic phase is similar to that of the inter-subband SDE away from the critical density \[cf. Fig. \[fig:asymmode\](b)\], measurements at a closely spaced grid of densities may be required to detect this cusp. In addition, the cusp may be broadened by impurity or fluctuation effects which are beyond our mean-field theory, further increasing the difficulty of detecting the transition. Thus, the current experimental results cannot exclude the existence of the zero-field antiferromagnetic state, and its robustness as demonstrated by the calculations in this paper leaves us optimistic that such a unique exchange-driven phase can occur in nature. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== RJR would like to acknowledge stimulating discussions with H. Ehrenreich regarding the superexchange mechanism in this system. The work of RJR and SDS work was supported by the US-ARO and the US-ONR. The work of AHM was supported by the National Science Foundation under grant DMR-9416906. For a recent review, see [Perspectives in Quantum Hall Effects]{}, edited by S. Das Sarma and A. Pinczuk (Wiley, New York, 1997). A. H. MacDonald, Phys. Rev. B [37]{}, 4792 (1988). P. P. Ruden and Z. Wu, Appl. Phys. Lett. [59]{}, 2165 (1991). L. Świerkowski, D. Neilson, and J. Szymański, Phys. Rev. Lett. [67]{}, 240 (1991); D. Neilson, L. Świerkowski, J. Szymański, and L. Liu, [ibid.]{} [71]{}, 4035 (1993). C. M. Varma, A. I. Larkin, and E. Abrahams, Phys. Rev. B [49]{}, 13999 (1994). Y. Katayama, D. C. Tsui, H. C. Manoharan, and M. Shayegan, Surf. Sci. [305]{}, 405 (1994); Phys. Rev. B [52]{}, 14817 (1995). X. Ying, S. R. Parihar, H. C. Manoharan, and M. Shayegan, Phys. Rev. B [52]{}, 11611 (1995). N. K. Patel, I. S. Millard, E. H. Linfield, P. D. Rose, M. P. Grimshaw, D. A. Ritchie, G. A. C. Jones, and M. Pepper, Phys. Rev. B [53]{}, 15443 (1996). S. Conti and G. Senatore, Europhys. Lett. [36]{}, 695 (1996). L. Zheng, M. W. Ortalano, and S. Das Sarma, Phys. Rev. B [55]{}, 4506 (1997). X.G. Wen and A. Zee, [Phys. Rev. Lett.]{} [69]{}, 1811 (1992); X.G. Wen and A. Zee, [Phys. Rev. B]{} [47]{}, 2265 (1993). Z.F. Ezawa and A. Iwazaki, Int. J. Mod. Phys. B [6]{}, 3205 (1992); [ibid.]{}, [8]{}, 2111 (1994); Phys. Rev. B [47]{}, 7295 (1993); [48]{}, 15189 (1993); Phys. Rev. Lett. [70]{}, 3119 (1993); Z.F. Ezawa, A. Iwazaki, and Y.S. Wu, Mod. Phys. Lett. B [7]{}, 1223 (1993); [7]{}, 1825 (1993); Z.F. Ezawa, Phys. Rev. B, Phys. Rev. B [51]{}, 11152 (1995). Kun Yang, K. Moon, L. Zheng, A.H. MacDonald, S.M. Girvin, D. Yoshioka, and Shou-Cheng Zhang, Phys. Rev. Lett. [72]{}, 732 (1994); K. Moon, H. Mori, Kun Yang, S. M. Girvin, A. H. MacDonald, L. Zheng, D. Yoshioka, and Shou-Cheng Zhang, Phys. Rev. B, [51]{}, 5138 (1995); K. Moon, H. Mori, Kun Yang, S.M. Girvin, A.H. MacDonald, L. Zheng, D. Yoshioka, and Shou-Cheng Zhang, submitted to Phys. Rev. B (1995). S. Das Sarma and P. I. Tamborenea, Phys. Rev. Lett. [73]{}, 1971 (1994). R. J. Radtke, P. I. Tamborenea, and S. Das Sarma, Phys. Rev. B [54]{}, 13832 (1996). V. Pellegrini, A. Pinczuk, B. S. Dennis, A. S. Plaut, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. [78]{}, 310 (1997). L. Zheng, R. J. Radtke, and S. Das Sarma, Phys. Rev. Lett. [78]{}, 2453 (1997). A. S. Plaut, A. Pinczuk, P. I. Tamborenea, B. S. Dennis, L. N. Pfeiffer, and K. W. West, Phys. Rev. B [55]{}, 9282 (1997). See, for example, S. Doniach and E. H. Sondheimer, [Green’s Functions for Solid State Physicists]{} (Addison-Wesley, New York, 1982), Chap. 7. L. Swierkowski and A.H. MacDonald, Phys. Rev. B [55]{}, R16017 (1997). G. D. Mahan, [Many-Particle Physics]{} (Plenum, New York, 1981). A. Pinczuk and G. Abstreiter, in [Light Scattering in Solids V]{}, edited by M. Cardona and G.Guntherodt (Springer-Verlag, Berlin, 1989). S. Das Sarma, in [Light Scattering in Semiconductor Structures and Superlattices]{}, edited by D. J. Lockwood and J. F. Young (Plenum Press, New York, 1991). A.H. MacDonald, P.M. Platzman and G.S. Boebinger, Phys. Rev. Lett. [65]{}, 775 (1990). T. Ando, A. B. Fowler, and F. Stern, Rev. Mod. Phys. [54]{}, 437 (1981). A. Pinczuk, S. Schmitt-Rink, G. Danan, J. P. Valladares, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. [63]{}, 1633 (1989). We shall continue to refer to the SDEs as inter- and intra-subband in an asymmetric double quantum well even though that distinction does not strictly apply. The reader should understand these terms to refer to the dominant character of the collective excitation with respect to the symmetric structure. D. M. Ceperley and B. J. Alder, Phys. Rev. Lett. [45]{}, 566 (1980); J. Perdew and A. Zunger, Phys. Rev. B [23]{}, 5048 (1981). B. Tanatar and D. M. Ceperley, Phys. Rev. B [39]{}, 5005 (1989). D. M. Ceperley, private communication. F. Rapisarda and G. Senatore, Aust. J. Phys. [49]{}, 161 (1996). Technically, the Mermin-Wagner theorem \[N. D. Mermin and H. Wagner, Phys. Rev. Lett. [17]{}, 1133 (1966)\] implies that fluctuations will completely destroy long-range order in two-dimensional systems of infinite extent at non-zero temperature. In practice, however, the finite sample size prohibits arbitrarily long-wavelength fluctuations and yields a region of finite temperature in which the sample is effectively a single, ordered domain. It is in this region where the mean-field theory should yield a good description of the antiferromagnetic phase.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Conformally related metrics and Lagrangians are considered in the context of scalar-tensor gravity cosmology. After the discussion of the problem, we pose a lemma in which we show that the field equations of two conformally related Lagrangians are also conformally related if and only if the corresponding Hamiltonian vanishes. Then we prove that to every non-minimally coupled scalar field, we may associate a unique minimally coupled scalar field in a conformally related space with an appropriate potential. The latter result implies that the field equations of a non-minimally coupled scalar field are the same at the conformal level with the field equations of the minimally coupled scalar field. This fact is relevant in order to select physical variables among conformally equivalent systems. Finally, we find that the above propositions can be extended to a general Riemannian space of $n$-dimensions.' author: - Michael Tsamparlis - Andronikos Paliathanasis - Spyros Basilakos - Salvatore Capozziello title: Conformally related metrics and Lagrangians and their physical interpretation in cosmology --- Introduction ============ The detailed analysis of the current high quality cosmological data (Type Ia supernovae, cosmic microwave background, baryonic acoustic oscillations, etc), converge towards a new emerging “Cosmological Standard Model”. This cosmological model is spatially flat with a cosmic dark sector constituted by cold dark matter and some sort of dark energy, associated with large negative pressure, in order to explain the observed accelerating expansion of the universe (see [@Teg04; @Spergel07; @essence; @Kowal08; @Hic09; @komatsu08; @LJC09; @BasPli10] and references therein). Despite the mounting observational evidences on the existence of the dark energy component in the universe, its nature and fundamental origin remains an intriguing enigma challenging the very foundations of theoretical physics. Indeed, during the last decade there has been an intense theoretical debate among cosmologists regarding the nature of this exotic “dark energy”. The absence of a fundamental physical theory, concerning the mechanism inducing the cosmic acceleration, has opened a window to a plethora of alternative cosmological scenarios. Most are based either on the existence of new fields in nature (dark energy) or in some modification of Einstein’s general relativity, with the present accelerating stage appearing as a sort of geometric effect (see [@Ratra88; @curvature; @mauro; @report; @repsergei; @Oze87; @Weinberg89; @Lambdat; @Bas09c; @Wetterich:1994bg; @Caldwell98; @Brax:1999gp; @KAM; @fein02; @Caldwell; @Bento03; @chime04; @Linder2004; @LSS08; @Brookfield:2005td; @Boehmer:2007qa; @Starobinsky-2007; @Ame10] and references therein). In order to investigate the dynamical properties of a particular dark energy model, we need to specify the covariant Einstein-Hilbert action of the model and find out the corresponding energy-momentum tensor. This methodology provides a mathematically consistent way to incorporate dark energy in cosmology. However, in the literature, there are many Lagrangians [@Ame10] which describe differently the physical features of the scalar field or the modified gravity [@report]. Because of the large amount of dark energy models, it is essential to study them in a unified context in order to discriminate the true physical variables. From our viewpoint, this framework has to be at the level of geometry since the various Lagrangians which describe the nature of dark energy are embedded in the space-time. From a theoretical point of view, an easy way to study dynamics in a unified manner is to look for conformally related Lagrangians. In fact, the idea to use conformally related metrics and Lagrangians as a cosmological tool is not new. In particular, it has been proposed that the existence of conformally equivalent Lagrangians can be used in order to select viable cosmological models [@allemandi]. In general, the presence of scalar fields into the gravitational action can give rise to two classes of theories: minimally and non-minimally coupled Lagrangians. In the first case, the gravitational coupling is the standard Newton constant and the scalar field Lagrangian is simply added to the Ricci scalar. It can consist of $i)$ a kinetic term, $ii)$ a kinetic term and a self-interaction potential, or $iii)$ just an interaction potential. In the first case, the scalar field is nothing else but a cyclic variable and then it is related to a conserved quantity. The second case is the relevant one since, by a variational principle, it is possible to obtain a Klein-Gordon equation where the self-interacting potential $V(\phi)$ leads the dynamics. The third case means that the scalar field has no dynamics. When the coupling is the Newton constant, we are in the so-called [Einstein frame]{}. In fact, as firstly pointed out by Brans and Dicke [@brans], a gravitational theory can be made more “Machian” by relaxing the hypothesis that the coupling is constant. They introduced a scalar field $\phi$ non-minimally coupled to the Ricci scalar $R$ and a kinetic term for such a scalar field into the gravitational action. The result was that the coupling was non-minimal and coordinate dependent. In other words, the gravitational interaction is assumed to change with distance and time. This approach can be generalized considering scalar-tensor theories of gravity where also a self-interaction potential comes into the game or more than one scalar field are taken into account. In general, any gravitational theory not simply linear in the Ricci scalar can be reduced to a scalar-tensor one. In the case of $f(R)$-gravity, it is straightforward to show that an [O’Hanlon representation]{} by scalar fields is possible [@ohanlon]. In this case, a scalar field is non-minimally coupled to the Ricci scalar and a self-interacting potential is present while there is no kinetic term. Scalar field dynamics is guaranteed by the non-minimal coupling and the potential (see [@report] and references therein). In other words, the further gravitational degrees of freedom, coming from the fact that $f(R)\neq R$, can be figured out as a scalar field. As soon as we are considering non-minimal couplings or higher-order terms in the Lagrangian, we are in the so-called [Jordan frame]{}, after Jordan who first introduced this notion [@Jordan38; @Jordan52]. In this article we would like to address the following basic question: In the framework of the scalar field cosmology, is it possible to relate the available Lagrangians in a conformal way? The structure of the paper is as follows. In section II, we discuss the issue of conformal transformations considering its historical development and connection with physical theories. From our point of view, this section is essential in order to fix the problem showing the urgency to discriminate physical variables among the conformally related models. The basic theoretical elements of the conformally related metric are presented in section III, where we also introduce the concept of conformally related Lagrangians and prove a lemma which shows that the field equations for two conformally related Lagrangian are also conformally related if the corresponding Hamiltonian vanishes. In section IV, we discuss the conformal equivalence of Lagrangians for scalar fields in a Riemannian space of dimension $4$ (for an extension see appendix A). In particular we enunciate a theorem which proves that the field equations for a non-minimally coupled scalar field are the same at the conformal level with the field equations of the minimally coupled scalar field. The necessity to preserve Einstein’s equations in the context of Friedmann-Lemaître-Robertson-Walker (FRLW) space-time leads us to apply, in section V, the current general analysis to the scalar field (quintessence or phantom) flat FLRW cosmologies. Finally, we draw our main conclusions and discuss results in section VI. What is the physical frame? =========================== Some considerations are in order at this point. The conformal transformations from the Jordan to the Einstein frame are geometrical maps allowing to set many features of scalar-tensor gravity, $f(R)$ gravity and, in general, any modified theory of gravity. Taking into account both the Jordan and the Einstein conformal frames (infinite conformal frames can be chosen assuming a suitable conformal factor), the question is whether the frames are infinitely many physically equivalent or only mathematically related. In other words, the problem is whether the physical information contained in the theory is “preserved” or not by the conformal transformations. In other words, one has the metric $g_{ij} $ and its conformally related one $\bar{g}_{ij}$: the question is what is the “physical metric”, i.e., the metric from which curvature, geometry, and physical effects have to be calculated and compared with experiments and observations [@mercadante]. More precisely, every Killing and homothetic vector is also a symmetry of the energy-momentum tensor. The latter is also a [metric]{} of spacetime because a “metric” can be considered any second order tensor. Therefore, up to the homothetic group, metric, energy-momentum tensor, and Ricci tensor have the same symmetries. This is not the case for the conformal group and then the requirement that the theory is invariant under conformal transformations is an additional assumption not related to the gauge invariance. Furthermore, we are concerned with Lagrangians which means that we assume the same kinematics on which these Lagrangians set up a dynamics. In general, kinematics has other symmetries with respect to dynamics and these symmetries are not related to those of the metric because no field equation relates kinematic quantities to metric. Kinematic symmetries are constrained only by the Ricci identity which gives the constraint and propagation equations. The issue of “which is the physical frame” has been debated for a long time and it emerges as soon as some authors argue in favor of one frame against the other, and others support the idea that the two frames are physically equivalent. In the latter case, authors claim that the issue is a pseudo-problem. The final result is that there is a good deal of confusion in the literature. Fierz was the first to pose the problem [@GrumillerKummerVassilevich03], but the main argument is due to Dicke, who discussed the conformal transformation for Brans-Dicke theory [@Dicke62]. The point was that physics must be invariant under a rescaling of units and the conformal transformation is a local rescaling: units do not change rigidly over the entire spacetime manifold, but by amounts which are different at different spacetime points. From Dicke’s point of view, the two frames are equivalent provided that mass, length, time, and quantities derived from them scale with appropriate powers of the conformal factor in the Einstein frame [@Dicke62]. From this point of view, it is not difficult to see why many authors consider the issue of which is the physical frame a pseudo-problem. In principle, it is difficult to object to this argument, but there are some difficulties. Even though the above argument is clear in principle, its application to practical situations gives rise to problems. The assumption that the two conformal frames are just different representations of the same theory, similar to different gauges of a gauge theory, has to be checked explicitly by using the field equations of a given system. Physical equivalence is a vague concept because one can consider many different matter (or test) fields in curved spacetime and different types of physics, or different physical aspects of a problem. When checking explicitly the physical equivalence between the two frames, one has to specify which physical field, or physical process is considered and the equations describing it. The equivalence has to be shown explicitly, but there is no proof that holds in any situation (e.g. scalar fields, spinors, cosmology, black holes derived from the same theory). While physical equivalence can be proven for various physical aspects, no proof comprehensive of all physical fields and different physical applications exists. It is important to stress that Dicke seems to mix the concept of dimensional units and the concept of measuring units. For example, spatial distance has the dimensional unit length $L$ but the measuring unit length, as we already know from Special Relativity, has to be defined in a relativistic inertial frame (RIF). After, by some rule, spatial lengths are compared in different RIFs. One of such rules is the Einstein one. It can be defined by means of light signals which are simultaneous in the RIF in which the measurement is taking place. This is the so called [rest length]{} (which, by assumption, coincides with the corresponding Euclidian length measured by a Newtonian inertial observer). This type of measurement has nothing to do with the concept of dimension which, for all RIFs, is the same i.e. the time $T$ (the second, in units, where $c=1$, see e.g. [@mikebook]). Dicke is aware of that and then says [“ Generally there may be more than one feasible way of establishing the equality of units at different spacetime points”]{} [@Dicke62]. It is essential to stress that such an approach is based on conventions which may lead to absurd results in real measurements. Geometrically it is a 1:1 map of two points defining a spacelike interval from the rest space of one observer at a spacetime point to a spacelike plane of another observer at another spacetime point. This definition it is not a coordinate transformation and there is no point or meaning to consider symmetry invariance with this transformation. Dicke says [“It is evident that the equations of motion of matter must be invariant under a general coordinate dependent transformation of units”]{} [@Dicke62]. This is a misunderstanding that can give rise to confusion. The method of comparing lengths at different spacetime points is a kinematic assumption while equations of motion give dynamics. Furthermore, Dicke’s argument is purely classical. In cosmology, black hole physics, and quantum fields in curved space, the equivalence of conformal frames is not clear. At quantum level, this equivalence is not proven due to the lack of a definitive quantum gravity theory: in fact, when the metric $g_{ij}$ is quantized, inequivalent quantum theories can be found [@GrumillerKummerVassilevich03; @AshtekarCorichi03]. One can consider the semiclassical regime in which gravity is classical and matter fields are quantized: again, one would expect the conformal frames to be inequivalent because the conformal transformations can be seen as Legendre transformations [@MagnanoSokolowski94], similar to the Legendre transformations of classical mechanics of point particles which switches from the canonical Lagrangian coordinates $q$ to the variables $\left\{ q,p\right\} $ of the Hamiltonian formalism. Now, it is well known that Hamiltonians that are classically equivalent become inequivalent when quantized, producing different energy spectra and scattering amplitudes [@CalogeroDegasperis04; @GlassScanio77]. However, the conformal equivalence between Jordan and Einstein frame seems to hold to some extent at the semi-classical level [@Flanagan04b]. Again, only a particular kind of physics has been considered and one cannot make statements about all possible physical situations. It is important to point out a very basic argument among particle physicists that relies on the equivalence theorem of Lagrangian field theory. It states that the $S$-matrix is invariant under local (nonlinear) field redefinitions [@Dyson48; @Blasietal99]. Since the conformal transformation is, essentially, a field redefinition, it would seem that quantum physics is invariant under the change of the conformal frame. However, the field theory in which the equivalence theorem is derived applies to gravity only in the perturbative regime in which the fields deviate slightly from the Minkowski space-time. In this regime, tree level quantities can be calculated in any conformal frame with same result, but in the non-perturbative regime field theory and the equivalence theorem do not apply. Unfortunately, the scaling of units in the Einstein frame often produces results that either do not make sense or are incorrect in the Jordan frame, reinforcing the opposite view that the two frames are not equivalent. While Dicke’s explanation is very appealing and several claims supporting the view that the two frames are inequivalent turned out to be incorrect because they simply neglected the scaling of units in the Einstein frame, one should not forget that Dicke’s argument is not inclusive of all areas of physics and it is better to check explicitly that the physics of a certain field does not depend on the conformal representation and not make sweeping statements. Certain points have been raised in the literature which either constitute a problem for Dicke’s view, or, at least, indicate that this viewpoint cannot be applied blindly, including the following ones. For example, massive particles follow time-like geodesics in the Jordan frame, while they deviate from geodesic motion in the Einstein frame due to a force proportional to the gradient of the conformal scalar field [@FaraoniNadeauconfo]. Hence, the Weak Equivalence Principle is satisfied only in the Jordan frame but not in the Einstein frame due to the coupling of the scalar field to the ordinary matter. Since the Equivalence Principle is the foundation of any relativistic theory of gravity, this aspect is important and there are two ways to consider it. One can ask for the two conformal frames are equivalent also with respect to the Equivalence Principle. This means that Equivalence Principle is formulated in a way that depends on the conformal frame representation. Then, a representation-independent formulation must be sought for. However, up to now, no definite result exists in this direction. On the other hand, we can ask for the violation of the Weak Equivalence Principle in the Einstein frame by saying that the “physical equivalence” of the two frames must be precisely defined and this concept cannot be used blindly. In fact the Equivalence Principle holds only in one frame but not in the other. This fact could be used as an argument against the physical equivalence of the frames. However, the fact that the Equivalence Principle holds in a given frame and not in [all frames]{} means that it is not a covariant requirement but a kinematic one. In other words, the Equivalence Principle could not be sufficient to discriminate among conformally related frames. In the scalar-tensor theories of gravity, the energy conditions are easily violated in the Jordan frame, but they are satisfied in the Einstein frame [@MagnanoSokolowski94]. This fact does not eliminate singularities and then the two frames remain equivalent with respect to the singularities and not with respect to the energy conditions. This difficulty arises because part of the matter sector of the theory, in the Einstein frame, comes from the conformal factor; in other words, the conformal transformation mixes matter and geometric degrees of freedom, which is the source of many interpretational problems [@Cap01; @Cap02]. Thus, even if the theory turns out to be independent of the conformal representation, its interpretation is not. There are results in cosmology in which the universe accelerates in one frame but not in the other. From the pragmatic point of view of an astronomer attempting to fit observational data (for example, type Ia supernovae data to a model of the present acceleration of the universe), the two frames certainly do not appear to be “physically equivalent” [@nodicap2; @CapozzielloPrado]. To approach correctly the problem of physical equivalence under conformal transformations, one has to compare physics in different conformal frames at the level of the Lagrangian, of the field equations, and of their solutions [@allemandi]. This comparison may not always be easy but, in certain cases, it is extremely useful to discriminate between frames. It has been adopted, for example, in Ref. [@cno], to compare cosmological models in the Einstein and the Jordan frame. Specifically, it has been shown that solutions of $f(R)$ and scalar-tensor gravity cannot be assumed to be physically equivalent to those in the Einstein frame when matter fields are given by generalized Equations of State. In these and in other situations, one must specify precisely what “physical equivalence” means. In certain situations physical equivalence is demonstrated simply by taking into account the coupling of the scalar field to matter and the varying units in the Einstein frame, but in other cases the physical equivalence is not obvious and it does not seem to hold. At the very least, this equivalence, if it is valid at all, must be defined in precise terms and discussed in ways that are far from obvious. For this reason, it would be too simplistic to dismiss the issue of the conformal frame as a pseudo-problem that has been solved for all physical situations of interest. Conformally related metrics and Lagrangians =========================================== Taking in mind the above discussion, we want to seek for geometrical structures that are conformally invariant. Our aim is to compare cosmological models coming from scalar-tensor gravity in order to select conformal quantities in view of a possible physical meaning. This is a very delicate issue that has to be discussed in details. On one hand, an invariant quantity, (i.e. a quantity that remains the same under conformal transformation) should have a physical meaning. However, in cosmology, this statement does not necessarily hold since the problem of equivalence between the two frames is not well posed and it may happen that one of them has to be taken as the physical one in a particular case and then such invariant quantity would not have a physical meaning. This situation often happen if cosmological solutions fit data and then related quantities are assumed as “physical” [@CapozzielloPrado]. On the other hand, if a scalar field describes an actual particle in a given frame, then its properties (e.g, its mass and couplings to other fields) would change in the conformally-related metric. This fact does not mean that its properties have no physical meaning. In other words, the identification of conformally invariant physical quantities is a very difficult task if it is not based on first principles. With these considerations in mind, let us start with defining some geometrical structures that will be useful in the discussion. A vector field $X^{a}$ is a Conformal Killing Vector [@Yano] (hereafter CKV) of the metric $g_{ij}$ if there exists a function $\psi\left( x^{k}\right) $ so that: $$\mathcal{L}_{X}g_{ij}=2\psi\left( x^{k}\right) g_{ij}\label{CLN.01}$$ where $\mathcal{L}_{X}$ is the Lie derivative. In case that $\psi_{;a}=0$ i.e. $\psi=$constant, the vector $X$ is called homothetic (hereafter HV) while if $\psi=0$ then the vector $X$ is a Killing vector (hereafter KV). In this context, two metrics $g_{ij},\bar{g}_{ij}$ are said to be conformal or conformally related if there exists a function $N^{2}\left( x^{k}\right) $ so that $\bar{g}_{ij}=N^{2}\left( x^{k}\right) g_{ij}$. From the mathematical point of view the CKVs form the so called conformal algebra of the metric. The conformal algebra contains two closed sub-algebras the Homothetic algebra and the Killing algebra. Interestingly the above algebras are related as follows: $$KV_{s}\subseteq HV_{s}\subseteq CKV_{s}\;.\label{CLN.02}$$ The dimension of the conformal algebra of an $n-$ dimensional metric $(n>2)$ of constant curvature equals $\frac{(n+1)(n+2)}{2}$, the dimension of the Killing algebra $\frac{n(n+1)}{2}$ and that of the homothetic algebra $\frac{n(n+1)}{2}+1$. Note that two conformally related metrics have the same conformal algebra [@TsampC], however not the same subalgebras. Indeed if $X$ is a CKV for the metric $g_{ij}$ i.e. $\mathcal{L}_{X}g_{ij}=2\psi\left( x^{k}\right) g_{ij}$ then for the metric $\bar{g}_{ij}$ the vector $X$ is again a CKV with conformal factor $\bar{\psi}\left( x^{k}\right) ,$ that is: $$\mathcal{L}_{X}\bar{g}_{ij}=2\bar{\psi}\left( x^{k}\right) \bar{g}_{ij}\label{CLN.03}$$ where the conformal factors $\psi\left( x^{k}\right) $, $\bar{\psi}\left( x^{k}\right) $ are related as follows: $$\bar{\psi}\left( x^{k}\right) =\psi\left( x^{k}\right) +\mathcal{L}_{X}\left( \ln N\right) .\label{CLN.04}$$ The Ricci scalars of the conformally related metrics $g_{ij},\bar{g}_{ij}$ are related as follows [@HawkingB]:$$\bar{R}=N^{-2}R-2(n-1)N^{-3}\Delta_{2}N-(n-1)(n-4)\Delta_{1}N\label{CLN.04.1}$$ where (note that $\Delta_{2}N$ contains the covariant derivative whereas $\Delta_{1}N$ the partial derivative): $$\begin{aligned} \Delta_{1}N & =g_{ij}N^{,i}N^{,j}\label{CLN.04.2}\\ \Delta_{2}N & =g_{ij}N^{;ij}.\label{CLN.04.3}$$ From the above discussion it becomes clear that all two dimensional spaces are Einstein Spaces $\left( \text{i.e. }R_{ab}=\frac{R}{2}g_{ab}\right) $ and conformally flat. The metric of a two dimensional space can be written in the generic form: $$ds^{2}=N^{2}\left( x,y\right) \left( \varepsilon dx^{2}+dy^{2}\right) ~~,~\varepsilon=\pm1\;.\label{CLN.10}$$ Conformal Lagrangians --------------------- Due to the fact that almost every dynamical system is described by a corresponding Lagrangian, below we study generically, as much as possible, the problem of the conformal Lagrangians and then we apply the current ideas to the scalar field cosmology. To begin with, consider the Lagrangian of a particle moving under the action of a potential $V(x^{k})$ in a Riemannian space with metric $g_{ij}$ $$L=\frac{1}{2}g_{ij}\dot{x}^{i}\dot{x}^{j}-V\left( x^{k}\right) ~,~\dot {x}^{i}=\frac{dx^{i}}{dt}\label{CLN.05}$$ where $t$ is a path parameter. The equations of motion follow from the action $$S=\int dxdtL\left( x^{k},\dot{x}^{k}\right) =\int dxdt\left[ \frac{1}{2}g_{ij}\dot{x}^{i}\dot{x}^{j}-V\left( x^{k}\right) \right] .\label{CLN.06}$$ Changing the variables in Eq.(\[CLN.06\]) from $t$ to $\tau$ via the relation: $$d\tau=N^{2}\left( x^{i}\right) dt \label{tran1A}$$ the action is given by $$S=\int dx\frac{d\tau}{N^{2}\left( x^{k}\right) }\left[ \frac{1}{2}g_{ij}N^{4}\left( x^{k}\right) x^{\prime i}x^{\prime j}-V\left( x^{k}\right) \right] ~~\label{CLN.07}$$ where $x^{\prime i}=\frac{dx^{i}}{d\tau}$. Obviously, the Lagrangian in the new coordinate system $(\tau,x^{i})$ becomes:$$\bar{L}\left( x^{k},x^{\prime k}\right) =\frac{1}{2}N^{2}\left( x^{k}\right) g_{ij}x^{\prime i}x^{\prime j}-\frac{V\left( x^{k}\right) }{N^{2}\left( x^{k}\right) }.\label{CLN.08}$$ Now if we consider a conformal transformation of the metric $\bar{g}_{ij}=N^{2}\left( x^{k}\right) g_{ij}$ and a new potential function $\bar {V}\left( x^{k}\right) =\frac{V\left( x^{k}\right) }{N^{2}\left( x^{k}\right) }$ then the new Lagrangian $\bar{L}\left( x^{k},x^{\prime k}\right) $ takes the following form: $$\bar{L}\left( x^{k},x^{\prime k}\right) =\frac{1}{2}\bar{g}_{ij}x^{\prime i}x^{\prime j}-\bar{V}\left( x^{k}\right) \label{CLN.09}$$ implying that Eq.(\[CLN.09\]) is of the same form as the Lagrangian $L$ in Eq.(\[CLN.05\]). From now on the Lagrangian $L\left( x^{k},\dot {x}^{k}\right) ~$ of Eq.(\[CLN.05\]) and the Lagrangian $\bar{L}\left( x^{k},x^{\prime k}\right) $ of Eq.(\[CLN.09\]) will be called conformal. In this framework, the action remains the same i.e. it is invariant under the change of parameter, the equations of motion in the new variables $(\tau ,x^{i})$ will be the same with the equations of motion for the Lagrangian $L$ in the original coordinates $(t,x^{k})$ . It has been shown [@TsamGE] that the Noether symmetries of a Lagrangian of the form (\[CLN.05\]) follow the homothetic algebra of the metric $g_{ij}.$ The same applies to the Lagrangian $\bar{L}\left( x^{k},x^{\prime k}\right) $ and the metric $\bar{g}_{ij}.$ As we have remarked the conformal algebra of the metrics $g_{ij},\bar{g}_{ij}$ (as a set) is the same however their closed subgroups of HVs and KVs are in general different[^1]. Now, we formulate and prove the following proposition: *Lemma:* The Euler-Lagrange equations for two conformal Lagrangians transform covariantly under the conformal transformation relating the Lagrangians iff the Hamiltonian vanishes. *Proof:* Consider the Lagrangian $L=\frac{1}{2}g_{ij}\dot{x}^{i}\dot{x}^{j}-V\left( x^{k}\right) $ whose Euler-Lagrange equations are: $$\ddot{x}^{i}+\Gamma_{jk}^{i}\dot{x}^{i}\dot{x}^{j}+V^{,i}=0 \label{CLN1.01}$$ where $\Gamma_{jk}^{i}$ are the Christofell symbols. The corresponding Hamiltonian is given by $$E=\frac{1}{2}g_{ij}\dot{x}^{i}\dot{x}^{j}+V\left( x^{k}\right) \;. \label{CLN1.02}$$ For the conformally related Lagrangian $\bar{L}\left( x^{k},x^{\prime k}\right) =\left( \frac{1}{2}N^{2}\left( x^{k}\right) g_{ij}x^{\prime i}x^{\prime j}-\frac{V\left( x^{k}\right) }{N^{2}\left( x^{k}\right) }\right) $ where $N_{,j}\neq0$ the resulting Euler Lagrange equations are$$x^{\prime\prime i}+\hat{\Gamma}_{jk}^{i}x^{\prime j}x^{\prime k}+\frac {1}{N^{4}}V^{,i}-\frac{2V}{N^{5}}N^{,i}=0 \label{CLN1.03}$$ where $$\hat{\Gamma}_{jk}^{i}=\Gamma_{jk}^{i}+(\ln N)_{,k}\delta_{j}^{i}+(\ln N)_{,j}\delta_{k}^{i}-(\ln N)^{,i}g_{jk} \label{CLN1.04}$$ and the corresponding Hamiltonian is$$\bar{E}=\frac{1}{2}N^{2}\left( x^{k}\right) g_{ij}\dot{x}^{i}\dot{x}^{j}+\frac{V\left( x^{k}\right) }{N^{2}\left( x^{k}\right) }. \label{CLN1.05}$$ In order to show that the two equations of motion are conformally related we start from Eq.(\[CLN1.03\]) and apply the conformal transformation $$\begin{aligned} x^{\prime i} & =\frac{dx^{i}}{d\tau}=\frac{dx^{i}}{dt}\frac{dt}{d\tau}=\dot{x}^{i}\frac{1}{N^{2}}\\ x^{^{\prime\prime}i} & =\ddot{x}^{i}\frac{1}{N^{4}}-2\dot{x}^{i}\dot{x}^{j}\left( \ln N\right) _{,j}\frac{1}{N^{4}}.\end{aligned}$$ Replacing in Eq.(\[CLN1.03\]) we find:$$\ddot{x}^{i}\frac{1}{N^{4}}-2\dot{x}^{i}\dot{x}^{j}\left( \ln N\right) _{,j}\frac{1}{N^{4}}+\frac{1}{N^{4}}\hat{\Gamma}_{jk}^{i}\dot{x}^{j}\dot {x}^{k}+\frac{1}{N^{4}}V^{,i}-\frac{2V}{N^{5}}N^{,i}=0\,.$$ Replacing $\hat{\Gamma}_{jk}^{i}$ from Eq.(\[CLN1.04\]) we have $$\begin{aligned} \ddot{x}^{i}-2\dot{x}^{i}\dot{x}^{j}\left( \ln N\right) _{,j}+\Gamma _{jk}^{i}\dot{x}^{j}\dot{x}^{k}+2(\ln N)_{,j}\dot{x}^{j}\dot{x}^{i}\\ -(\ln N)^{,i}g_{jk}\dot{x}^{j}\dot{x}^{k}+V^{,i}-2V(\ln N)^{,i}=0\end{aligned}$$ from which follows$$\ddot{x}^{i}+\Gamma_{jk}^{i}\dot{x}^{j}\dot{x}^{k}+V^{,i}-(\ln N)^{,i}\left( g_{jk}\dot{x}^{j}\dot{x}^{k}+2V\right) =0.$$ Obviously, the above Euler-Lagrange equations coincide with Eqs.(\[CLN1.01\]) if and only if $\left( g_{jk}\dot{x}^{j}\dot{x}^{k}+2V\right) =0$, which implies that the Hamiltonian of Eq.(\[CLN1.02\]) vanishes. The steps are reversible hence the inverse is also true. The physical meaning of such a result is that systems with vanishing energy are conformally invariant at the level of equations of motion. Conformally equivalent Lagrangians in the scalar field Cosmology ================================================================ Let us discuss now the conformal equivalence of Lagrangians for scalar fields in a general Riemannian space of four dimensions (see also [@allemandi]). The field equations in the scalar-tensor cosmology can be derived from two different variational principles. In the first case we consider a scalar field $\phi$ which is minimally coupled to gravity and the equations of motion follow from the action $$S_{M}=\int d\tau dx^{3}\sqrt{-g}\left[ R+\frac{\varepsilon}{2}g_{ij}\phi ^{;i}\phi^{;j}-V\left( \phi\right) \right] \label{CLN.11}$$ where $\varepsilon=\pm 1$ defines quintessence or phantom field cosmology respectively. In the second case we assume a scalar field $\psi$ (different from the minimally coupled scalar field $\phi$) which interacts with the gravitational field (non minimal coupling) and the corresponding action is given by $$S_{NM}=\int d\tau dx^{3}\sqrt{-\bar{g}}\left[ F\left( \psi\right) \bar {R}+\frac{\varepsilon}{2}\bar{g}^{ij}\psi_{;i}\psi_{;j}-\bar{V}\left( \psi\right) \right] \label{CL.12.0}$$ where $F(\psi)$ is the coupling function between the gravitational and the scalar field $\psi$ respectively. Below we pose the following proposition. *Theorem:* The field equations for a non minimally coupled scalar field $\psi$ with Lagrangian $\bar{L}\left( \tau,x^{k},\dot {x}^{k}\right) $ and coupling function $F(\psi)$ in the gravitational field $\bar{g}_{ij}$ are the same with the field equations of the minimally coupled scalar field $\Psi$ for a conformal Lagrangian $L\left( \tau,x^{k},\dot {x}^{k}\right) $ in the conformal metric $g_{ij}$ $=N^{-2}\bar{g}_{ij}$ where the conformal function is $N=\frac{1}{\sqrt{-2F\left( \psi\right) }}$ with $F\left( \psi\right) <0.$ The inverse is also true, that is, to a minimally coupled scalar field it can be associated a unique non minimally coupled scalar field in a conformal metric and with a different potential function. *Proof:* We first start with the action of Eq.(\[CL.12.0\]). Let $g_{ij}$ be the conformally related metric (this is not a coordinate transformation): $$g_{ij}=N^{-2}\bar{g}_{ij}.$$ Then the action provided by Eq.(\[CL.12.0\]) becomes [^2]:$$S_{NM}=\int d\tau dx^{3}N^{4}\sqrt{-g}\left[ F\left( \psi\right) \bar {R}+\frac{\varepsilon}{2}N^{-2}g^{ij}\psi_{;i}\psi_{;j}-\bar{V}\left( \psi\right) \right] .$$ Inserting the Ricci scalar $\bar{R}$ (using $n=4)$ from Eq.(\[CLN.04.1\]) into the latter equation we find: $$\begin{aligned} \label{ICLN} S_{NM} &=&\int d\tau dx^{3}N^{4}\sqrt{-g}\left[ F\left( \psi \right) N^{-2}R-6F\left( \psi \right) N^{-3}\Delta _{2}N +\frac{\varepsilon }{2}N^{-2}\Delta _{1}\psi -\bar{V}\left( \psi \right) \right]\,. $$ Now we can define the conformal function $N$ in terms of the coupling function $F(\psi)$ \[where $F\left( \psi\right) <0)$\]: $$\label{CLN122}N=\frac{1}{\sqrt{-2F\left( \psi\right) }}~.$$ with $$\label{CLN123}N_{;i}=\frac{F_{\psi}\psi_{;i}}{\left( -2F\right) ^{\frac {3}{2}}}\;.$$ Using Eqs.(\[CLN122\]) and (\[CLN123\]) the first term of the integral in Eq.(\[ICLN\]) becomes: $$\int d\tau dx^{3}\sqrt{-g}F\left( \psi\right) N^{2}R=\int d\tau dx^{3}\sqrt{-g}\left( -\frac{R}{2}\right) .$$ On the other hand the second term in Eq.(\[ICLN\]) gives, after integration by parts: $$\begin{aligned} \label{ICLN1} \int d\tau dx^{3}\sqrt{-g}\left[ -6F\left( \psi \right) N\Delta _{2}N\right] &=&\int d\tau dx^{3}\sqrt{-g}\left( -6\frac{F}{\sqrt{-2F}} N_{;ij}g^{ij}\right) \nonumber\\ &=&\int d\tau dx^{3}\sqrt{-g}\left[ -6\frac{F}{\sqrt{-2F}}\frac{1}{\sqrt{-g}}\left( \sqrt{-g}g^{ij}N_{,k}\right) _{,j}\right] \nonumber\\ &=&\int d\tau dx^{3}\left[ -6\frac{F}{\sqrt{-2F}}\left( \sqrt{-g} g^{ij}N_{,k}\right) _{,j}\right] \nonumber\\ &=&\int d\tau dx^{3}\sqrt{-g}\left( 3\frac{F_{\psi }}{\sqrt{-2F}}\psi _{;j}N_{;i}g^{ij}\right) \nonumber \\ &=& \int d\tau dx^{3}\sqrt{-g}\left[ 3\frac{F_{\psi }^{2}}{\left( -2F\right) ^{2}}\psi _{;i}\psi _{;j}g^{ij}\right] \;.\end{aligned}$$ The third term provides: $$\frac{\varepsilon}{2}N^{2}\Delta_{1}\psi=\frac{\varepsilon}{4F}\psi_{;i}\psi_{;j}g^{ij}\,.$$ Finally, collecting all terms and inserting them into Eq.(\[ICLN\]), the action is written as $$\begin{aligned} S_{NM} &=& \int d\tau dx^{3}\sqrt{-g}\left[ -\frac{R}{2}+3\frac{F_{\psi }^{2}}{4F^{2} }\psi _{;i}\psi _{;j}g^{ij}-\frac{\varepsilon }{4F}\psi _{;i}\psi _{;j}g^{ij}-\frac{\bar{V}\left( \psi \right) }{4F^{2}}\right] \nonumber\\ &=& \int d\tau dx^{3}\sqrt{-g}\left[ -\frac{R}{2}+\frac{\varepsilon }{2}\left( \frac{3\varepsilon F_{\psi }^{2}-F}{2F^{2}}\right) \psi _{;i}\psi _{;j}g^{ij}-\frac{\bar{V}\left( \psi \right) }{4F^{2}}\right] \;. \label{CLN.12.5}\end{aligned}$$ Interestingly, introducing the scalar field $\Psi$ with the requirement: $$d\Psi=\sqrt{\left( \frac{3\varepsilon F_{\psi}^{2}-F}{2F^{2}}\right) }d\psi\label{CLN.12.6}$$ the action of Eq.(\[CLN.12.5\]) can be written as follows $$S_{NM}=\int d\tau dx^{3}\sqrt{-g}\left[ -\frac{R}{2}+\frac{\varepsilon}{2}\Psi_{;i}\Psi_{;j}g^{ij}-\frac{\bar{V}\left( \Psi\right) }{4F\left( \Psi\right) ^{2}}\right] \;.\label{CLN.12.7}$$ We conclude that the scalar field $\Psi$ is minimally coupled (modulus a constant) to the gravitational field. In other words, we find that to every non-minimally coupled scalar field, we may associate a unique minimally coupled scalar field in a conformally related space with an appropriate potential. All considerations are reversible, hence the result is reversible. Finally, we would like to remark that the above theorem can be extended to general Riemannian spaces of $n-$dimensions (see appendix A). Conformal Lagrangians in FRLW cosmology ======================================= In this section we consider a spatially flat $\left( K=0\right) $ FRLW spacetime[^3] whose metric is$$ds^{2}=-dt^{2}+a^{2}\left( t\right) \delta_{ij}dx^{i}dx^{j}\label{CLN.13}$$ where $\delta_{ij}$ is the 3-space metric in Cartesian coordinates. The Lagrangian of a scalar field $\phi$ minimally coupled to gravity in this coordinate system $(a,\phi)$ is $$L_{M}=-3a\dot{a}^{2}+\frac{\varepsilon}{2}a^{2}\dot{\phi}^{2}-a^{3}V\left( \phi\right) .\label{CLN.14}$$ On the other hand, the Lagrangian of the non minimally coupled scalar field $\psi$ in the coordinate system $(a,\psi)$ is given by $$L_{NM}=6F\left( \psi\right) a\dot{a}^{2}+6F_{\psi}\left( \psi\right) a^{2}\dot{a}\dot{\psi}+\frac{\varepsilon}{2}a^{3}\dot{\psi}^{2}-a^{3}V\left( \psi\right) \label{CLN.15}$$ where $F(\psi)<0$ is the coupling function. The Hamiltonian of the above Lagrangian is $$E=6F\left( \psi\right) a\dot{a}^{2}+6F_{\psi}\left( \psi\right) a^{2}\dot{a}\dot{\psi}+\frac{\varepsilon}{2}a^{3}\dot{\psi}^{2}+a^{3}V\left( \psi\right) .\label{CLN.15e}$$ We construct a conformal Lagrangian which corresponds to a minimally coupled scalar field. To do that we introduce the following transformation (see [@Cap01; @Cap02]): $$A\left( t\right) =\sqrt{-2F(t)}a(t)\;.\label{CLN.15a}$$ Then the Lagrangian (\[CLN.15\]) takes the form:$$\begin{aligned} L_{NM} & =\frac{1}{\sqrt{-2F}}\left[ -3A\dot{A}^{2}+\frac{\varepsilon}{2}\left( \frac{3\varepsilon F_{\psi}^{2}-F}{2F^{2}}\right) A^{3}~\dot{\psi }^{2}\right] \nonumber\\ & -\frac{A^{3}}{\left( -2F\right) ^{\frac{3}{2}}}V\left( \psi\right) \;.\label{CLN.16}$$ It is interesting to mention here that the cross term $\dot{a}\dot{\psi}$ disappears from Eq.(\[CLN.16\]). Utilizing simultaneously Eq.(\[CLN.12.6\]) and the conformal transformation $$d\tau=\sqrt{-2F\left( \psi\right) }dt\label{CLN.17a}$$ we find, after some algebra, that Eq.(\[CLN.16\]) can be written as $$L_{M}\left( A,A^{\prime},\Psi,\Psi^{\prime}\right) =-3AA^{\prime2}+\frac{\varepsilon}{2}A^{3}\Psi^{\prime2}-A^{3}\bar{V}(\Psi)\label{LLC}$$ where $$\bar{V}(\Psi)=\frac{A^{3}}{\left( -2F\right) ^{\frac{3}{2}}}V\left( \Psi\right) \;.\label{CLN.17}$$ Notice that the prime denotes derivative with respect to the conformal time $\tau$. Evidently, the functional form of the Lagrangian (\[LLC\]) has the general form of Eq.(\[CLN.14\]) proving our assessment. Furthermore, considering in the new coordinates $(\tau,x^{i})$ the metric$$d\bar{s}^{2}=-d\tau^{2}+A^{2}\left( \tau\right) \delta_{ij}dx^{i}dx^{j}\label{CLN.19}$$ we find that the term $3AA^{\prime2}$ equals the Ricci scalar $\bar{R}$ of the conformally flat metric $d\bar{s}^{2}.$ In other words, the Lagrangian (\[LLC\]) can be seen as the Lagrangian of a scalar field $\Psi$ of potential $\bar{V}\left( \Psi\right) $ which is minimally coupled to the gravitational field $\bar{g}_{ij}$ in the space with metric $d\bar{s}^{2}$. Replacing the quantity $A\left( \tau\right) $ and the coordinate $\tau$ from Eq.(\[CLN.15a\]) and Eq.(\[CLN.17a\]) respectively, we obtain: $$d\bar{s}^{2}=\sqrt{-2F}\left[ -dt^{2}+a^{2}\left( t\right) \delta _{ij}dx^{i}dx^{j}\right] =\sqrt{-2F}ds^{2}\label{CLN.20}$$ that is, the metric $d\bar{s}^{2}$ is conformally related to the metric $ds^{2}$ with conformal function $\sqrt{-2F}.$ This means that the non-minimally coupled scalar field in the gravitational field $ds^{2}$ is equivalent to a minimally coupled scalar field - with appropriate potential defined in terms of the coupling function - in the gravitational field $d\bar{s}^{2}.$ For the benefit of the reader, we would like to stress that the above geometrical/dynamical result is reversible in the sense that a minimally coupled scalar field $\phi$ in a metric $ds^{2}$ can be seen as a non-minimally coupled scalar field $\psi$ in the flat FRLW space in which the Eq.(\[CLN.15\]) is equivalent to the minimally coupled scalar field $\Psi=\Psi(\psi)$ in the conformally related metric $d\bar{s}^{2},$ where the conformal function is defined in terms of the coupling function. Equivalently the Lagrangians $L_{M}$ and $L_{NM}$ are conformally related. Finally, we want to stress that the result of the previous lemma is automatically recovered since the Hamiltonian (\[CLN.15e\]) is equal to zero being the $\{0,0\}$ Einstein equation of the system (see also [@CapRev]). Discussion and Conclusions ========================== In this article we have investigated conformally related metrics and Lagrangians in the context of scalar-tensor cosmology. The aim is to select which is the frame where conformally related solutions have an immediate physical meaning. As discussed in section II, no final statement is available for the problem if solutions have to be interpreted either in the Jordan frame or in the Einstein frame since the physical equivalence can be questioned according to several issues (quantum vs classical measurements, energy conditions, choice of physical units, etc.). Due to this situation, it is is too simplistic to consider the problem of conformal frames just a pseudo-problem since we are facing only a mathematical equivalence. Clearly, it has to be addressed at three levels: $i)$ Lagrangians (or in general, effective actions); $ii)$ field equations; $iii)$ solutions. Actually, the last issue means also the choice of a set of observables where the interpretation of solutions is evident. To this goal, seeking for dynamical quantities invariant under conformal transformations is a fundamental issue. However, such quantities have to be related to geometry and possibly to be conserved like Noether symmetries. With this target in mind, we have firstly proved a lemma which shows that the field equations of two conformally related Lagrangians are also conformally related if the corresponding Hamiltonian vanishes. This fact is extremely relevant being the Hamiltonian the energy constraint of a given mechanical system and, in particular, it constitutes a non-holonomic constraint for dynamical systems describing cosmological models. It is the $\{0,0\}$ Einstein equation of the system. Secondly, we have found that to every non-minimally coupled scalar field, we can associate a unique minimally coupled scalar field in a conformally related space with an appropriate potential. The existence of such a connection can be used in order to study the dynamical properties of the various cosmological models, since the field equations of a non-minimally coupled scalar field are the same, at the conformal level, of the field equations of the minimally coupled scalar field. The above propositions can be extended to general Riemannian spaces in $n$-dimensions. It is worth stressing that the above results are in agreement with the so called [Bicknell’s Theorem]{} which states that a general non-linear $f(R)$ Lagrangian is equivalent to a minimally coupled scalar field with a general potential in the Einstein frame. In Ref. [@Bick74], this result is achieved in the case of $R^2$-gravity. In [@schmidt], the result is generalized to any analytic $f(R)$-gravity. We’d like to point out that in a recent paper [@BB], based on the Noether symmetry approach, we have studied the issue of physical solutions in $f(R)$ gravity models and scalar field dark energy models. Starting from these results, it is possible to identify the Noether symmetries, the physical solutions and the corresponding conformal properties of the scalar tensor theories (including $f(T)$ gravity). Such an analysis is in progress. In general, the Noether symmetries play an important role in physics because they can be used to simplify a given system of differential equations as well as to determine the integrability of the system. The latter will provide the necessary platform in order to solve the equations of motion analytically and thus to obtain the evolution of the physical quantities. In cosmology, such a method is extremely relevant in order to compare cosmographic parameters, such as scale factor, Hubble expansion rate, deceleration parameter, density parameters with observations [@CapRev; @BB; @felice; @nesseris]. In this appendix we generalize the theorem of section III to a Riemannian space of dimension $n$. Briefly, we consider the non minimally coupled scalar field $\psi $ whose field equations are obtained from the action: $$\begin{aligned} S_{NM} &=&\int dx^{n}N^{n}\sqrt{-g}\left[ F\left( \psi \right) \bar{R}+\frac{\varepsilon }{2}N^{-2}g^{ij}\psi _{,i}\psi _{,j}-\bar{V}\left( \psi \right) \right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ \begin{array}{c} F\left( \psi \right) N^{n-2}R-2(n-1)F\left( \psi \right) N^{n-3}\Delta _{2}N+ \\ -F\left( \psi \right) N^{n}(n-1)(n-4)\Delta _{1}N+\frac{\varepsilon }{2}N^{n-2}g^{ij}\psi _{,i}\psi _{,j}-N^{n}\bar{V}\left( \psi \right)\end{array}\right] \label{GENA}\end{aligned}$$ where in order to derive the last equality we have used Eq.(\[CLN.04.1\]). Note that we can define the function $N(x^{i})$ in terms of the coupling function $F(\psi)$ by the requirement: $$N^{n-2}=-\frac{1}{2F}\;\;\;\;F=-\frac{N^{2-n}}{2}$$ which also implies that $$N=\frac{1}{\left( -2F\right) ^{\frac{1}{n-2}}}\rightarrow N^{-1}=\left( -2F\right) ^{\frac{1}{n-2}} \label{N11}$$ $$N_{;i}N_{;j}^{-1}=-\frac{F_{\psi}^{2}}{\left( n-2\right) ^{2}F^{2}}\psi _{;i}\psi_{;j}\;. \label{N12}$$ We start now to treat the terms of the action in Eq.(\[GENA\]). In particular, the first term gives: $$\int dx^{n}\sqrt{-g}\left( F\left( \psi\right) N^{n-2}R\right) =\int dx^{n}\sqrt{-g}\left( -\frac{R}{2}\right) \;.\label{ICLN01}$$ If we utilize Eqs.(\[N11\]) and (\[N12\]) then the second (integrating by parts) and the third terms of the general action are $$\begin{aligned} \label{ICLN11} \int dx^{n}\sqrt{-g}\left[ -2(n-1)F\left( \psi \right) N^{n-3}N_{;ij}g^{ij}\right] &=&\int dx^{n}\sqrt{-g}\left[ (n-1)N^{2-n}N^{n-3}N_{;ij}g^{ij}\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ (n-1)N^{-1}\frac{1}{\sqrt{-g}}\left( \sqrt{-g}g^{ij}N_{,k}\right) _{,j}\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ (n-1)N^{-1}N_{;ij}g^{ij}\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ -(n-1)\left( N^{-1}\right) _{;j}N_{;i}g^{ij}\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ \frac{(n-1)}{\left( n-2\right) ^{2}}\frac{F_{\psi }^{2}}{F^{2}}\psi _{;i}\psi _{;j}g^{ij}\right]\;.\end{aligned}$$ $$\begin{aligned} \label{ICLN22} \int dx^{n}\sqrt{-g}\left[ F\left( \psi \right) N^{n}(n-1)(n-4)\Delta _{1}N\right) &=&\int dx^{n}\sqrt{-g}\left( -\frac{N^{2-n}}{2}N^{n}(n-1)(n-4)\Delta _{1}N\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ -\frac{1}{2}N^{2}(n-1)(n-4)\Delta _{1}N\right] \nonumber \\ &=&\int dx^{n}\sqrt{-g}\left[ -\frac{1}{2}\frac{(n-1)(n-4)}{\left( n-2\right) ^{2}}\frac{F_{\psi }^{2}}{\left( -2F\right) ^{\frac{4}{2-n}}F^{2}}\psi _{;i}\psi _{;j}g^{ij}\right] \;.\end{aligned}$$ To this end the final term gives: $$\label{ICLN33}\int dx^{n}\sqrt{-g}\left( \frac{\varepsilon}{2}N^{n-2}g^{ij}\psi_{,i}\psi_{,j}\right) =\int dx^{n}\sqrt{-g}\left( -\frac {\varepsilon}{4}\frac{1}{F}g^{ij}\psi_{;i}\psi_{;j}\right).$$ Now we change the variable $\psi$ to $\Psi$ as follows $$\label{ICLN44} d\Psi =\left[ \frac{2\varepsilon (n-1)}{\left( n-2\right) ^{2}}\frac{F_{\psi }^{2}}{F^{2}}-\varepsilon \frac{(n-1)(n-4)}{\left( n-2\right) ^{2}}\frac{F_{\psi }^{2}}{\left( -2F\right) ^{\frac{4}{2-n}}F^{2}}-\frac{1}{2F}\right] ^{\frac{1}{2}}d\psi \;.$$ Collecting the results of the above terms namely Eqs.(\[ICLN01\]),(\[ICLN11\]),(\[ICLN22\]),(\[ICLN33\]) and (\[ICLN44\]) we find after some non-trivial algebra that the general action of Eq.(\[GENA\]) is written is terms of $\psi$ as follows $$S_{NM}=\int dx^{n}\sqrt{-g}\left[ -\frac{R}{2}+\frac{\varepsilon}{2}\Psi _{;i}\Psi_{;j}g^{ij}-\frac{\bar{V}\left( \Psi\right) }{\left( -2F\right) ^{\frac{n}{n-2}}}\right] .$$ We would like to remind the reader that the new scalar field $\Psi$ is minimally coupled to the gravitational field $g_{ij}$ and that the potential of $\Psi$ is given by $\frac{\bar{V}\left( \Psi\right) }{\left( -2F\right) ^{\frac{n}{n-2}}}$. Notice that for $n=4$ the above expressions boil down to those of section III as they should. The above proof agrees with that provided by Keiser [@Keiser] however in our work we have used a different methodology which is simple and transparent. [99]{} M. Tegmark, et al., Astrophys. J., 606, 702, (2004) D.N. Spergel, et al., Astrophys. J. Suplem., 170, 377, (2007) T.M. Davis et al., Astrophys. J., 666, 716, (2007) M. Kowalski, et al., Astrophys. J., 686, 749, (2008) M. Hicken et al., Astroplys. J., 700, 1097, (2009) E. Komatsu, et al., Astrophys. J. Suplem., 180, 330, (2009); G. Hinshaw, et al., Astrophys. J. Suplem., 180, 225, (2009); E. Komatsu, et al., Astrophys. J. Suplem., 192, 18, (2011) J. A. S. Lima and J. S. Alcaniz, Mon. Not. Roy. Astron. Soc. 317, 893 (2000); J. F. Jesus and J. V. Cunha, Astrophys. J. Lett. 690, L85 (2009) S. Basilakos and M. Plionis, Astrophys. J. Lett, 714, 185 (2010) B. Ratra and P. J. E. Peebles, Phys. Rev D., 37, 3406 (1988). S. Capozziello, Int.J.Mod.Phys. D 11, 483 (2002). S. Capozziello, M. Francaviglia Gen.Rel.Grav. 40, 357 (2008). S. Capozziello, M. De Laurentis, Phys. Rept. 509, 167 (2011). S. Nojiri, S. D. Odintsov , Phys.Rept. 505, 59 (2011). M. Ozer and O. Taha, Nucl. Phys. B 287, 776 (1987). S. Weinberg, Rev. Mod. Phys., 61, 1, (1989) W. Chen and Y-S. Wu, Phys. Rev. D 41, 695 (1990); J. C. Carvalho, J. A. S. Lima and I. Waga, Phys. Rev. D [46]{}, 2404 (1992); J. A. S. Lima and J. M. F. Maia, Phys. Rev D 49, 5597 (1994); J. A. S. Lima, Phys. Rev. D 54, 2571 (1996), \[gr-qc/9605055\]; A. I. Arbab and A. M. M. Abdel-Rahman, Phys. Rev. D 50, 7725 (1994); J. M. Overduin and F. I. Cooperstock, Phys. Rev. D [58]{}, 043506 (1998). S. Basilakos, M. Plionis and S. Solà, Phys. Rev. D. 80, 3511 (2009) C. Wetterich, Astron. Astrophys. 301, 321 (1995) R. R. Caldwell, R. Dave, and P.J. Steinhardt, Phys. Rev. Lett., 80, 1582 (1998). P. Brax, and J. Martin, Phys. Lett. B468, 40 (1999). A. Kamenshchik, U. Moschella, and V. Pasquier, Phys. Lett. B. 511, 265, (2001). A. Feinstein, Phys. Rev. D., 66, 063511 (2002) R. R. Caldwell, Phys. Rev. Lett. B., 545, 23 (2002) M. C. Bento, O. Bertolami, and A.A. Sen, Phys. Rev. D., 70, 083519 (2004) L. P. Chimento, and A. Feinstein, Mod. Phys. Lett. A, 19, 761 (2004) E. V. Linder, Rep. Prog. Phys., 71, 056901 (2008) J. A. S. Lima, F. E. Silva and R. C. Santos, Class. Quant. Grav. 25, 205006 (2008) A. W. Brookfield, C. van de Bruck, D.F. Mota, and D. Tocchini-Valentini, Phys. Rev. Lett. 96, 061301 (2006) C. G. Boehmer, and T. Harko, Eur. Phys. J. C50, 423 (2007) A. A. Starobinsky, JETP Lett. 86, 157 (2007). E. J. Copeland, M. Sami and S. Tsujikawa, Intern. Journal of Modern Physics D, 15, 1753,(2006); R. R. Caldwell and M. Kamionkowski, Ann.Rev.Nucl.Part.Sci., 59, 397, (2009), arXiv:0903.0866; I. Sawicki and W. Hu, Phys. Rev. D., 75, 127502 (2007); L. Amendola and S. Tsujikawa, Dark Energy Theory and Observations, Cambridge University Press, Cambridge UK, (2010); S. Capozziello and V. Faraoni, Beyond Einstein gravity: A Survey of gravitational theories for cosmology and astrophysics, Fundamental Theories of Physics, Vol. 170, Springer, Heidelberg (2010); K. Bamba, S. Capozziello, S. Nojiri, S. D. Odintsov, Astrophys. Space Sci. [342]{}, 155 (2012). G. Allemandi, M. Capone, S. Capozziello, M. Francaviglia, Gen. Rel. Grav. 38, 33 (2006). C.H. Brans, R.H. Dicke, Phys. Rev. [124]{}, 925 (1961). J. OÕHanlon, Phys. Rev. Lett. [29]{} 137 (1972). P. Jordan, [Naturwiss.]{} [26]{}, 417 (1938). P. Jordan, [Schwerkraft und Weltfall, Grundlagen der Theoretische Kosmologie]{}. Braunschweig: Vieweg und Sohns (1952). S. Capozziello, M. De Laurentis, M. Francaviglia, S. Mercadante, Found. Phys. 39, 1161 (2009). D. Grumiller, W. Kummer, D.W. Vassilevich, J. High. Energy Phys. 7, 009 (2003). M. Tsamparlis, [Special Relativity]{}, Springer, Berlin (2010). R. H. Dicke, Phys. Rev. 125, 2163 (1962). A. Ashtekar, A. Corichi, Class. Quantum Grav. 20, 4473 (2003). G. Magnano and L.M. Solowski, Phys. Rev. D 50, 5039 (1994). F. Calogero, A. Degasperis, Am. J. Phys. 72, 1202 (2004). E.N. Glass, J.J.G. Scanio, Am. J. Phys. 45, 344 (1977). E.E. Flanagan, Class. Quantum Grav. 21, 3817 (2004). F.J. Dyson, Phys. Rev. 73, 929 (1948). A. Blasi, N. Maggiore, S.P. Sorella, L.C.Q. Vilar, Phys. Rev. D 59, 121701 (1999). V. Faraoni, S. Nadeau, Phys. Rev. D 75, 023501 (2007). S. Capozziello, S. Nojiri, S.D. Odintsov, A. Troisi, Phys. Lett. B 639, 135 (2006). S. Capozziello, P. Martin-Moruno, C. Rubano, Phys. Lett. B 689, 117 (2010). S. Capozziello, S. Nojiri, S.D. Odintsov, Phys. Lett. B 634, 93 (2006). K. Yano, (1956), “The Theory of Lie derivatives and Its Applications”, North Holland, Amsterdam, The Netherlands M. Tsamparlis, D. Nikolopoulos and P. S. Apostolopoulos, Class. Quantum. Grav. 15 2909 (1998) S. W. Hawking and G. F. R. Ellis, “The Large Scale Structure of Space-Time” Cambridge University Press, Cambridge: New York (1975). S. Capozziello, R. de Ritis, C. Rubano, P. Scudellaro, Riv. Nuovo Cim. 19N4, 1 (1996). S. Capozziello, R. de Ritis and A. A. Marino, Class. Quantum. Grav. 14, 3243 (1997). S. Capozziello, R. de Ritis and A. A. Marino, Class. Quantum. Grav. 14, 3259 (1997). M. Tsamparlis and A. Paliathanasis, Gen. Relativ. Grav. 42, 2957 (2010) S. Basilakos, A. Paliathanasis and M. Tsamparlis, Phys. Rev. D., 83, 103512 (2011); A. Paliathanasis, M. Tsamparlis and S. Basilakos, Phys. Rev. D., 84, 123514 (2011). S. Capozziello and A. De Felice, JCAP 0808, 016 (2008). S. Capozziello, S. Nesseris and L. Perivolaropoulos, JCAP [0712]{}, 009 (2007). G. V. Bicknell, Journal of Phys. A., [7]{} 1061 (1974). H. J. Schmidt, Lectures in Mathematical Cosmology, e-Print: gr-qc/0407095 (2004). D. Keiser, Phys. Rev. D., 81 084044 (2010) [^1]: The Noether symmetries of the conformal Lagrangians (\[CLN.05\])-(\[CLN.09\]) are elements of the common conformal algebra of the metrics $g_{ij},\bar{g}_{ij}$. A clear definition of the Noether symmetries can be founds in [@CapRev; @Cap01; @Cap02; @TsamGE] (for applications to cosmology see [@BB; @felice; @nesseris] and references therein). [^2]: For a $4\times 4$ matrix namely, $A=(a_{ij})$ we have $$\det A=\varepsilon^{ijkl}a_{ij}a_{kl}$$ hence $$\bar{g}=\varepsilon^{ijkl}\bar{g}_{ij}\bar{g}_{kl}=N^{4}g.\label{CLN.12.1}$$ [^3]: Similar results can be achieved for $K\neq0$ [@allemandi].	{ "pile_set_name": "ArXiv" }
--- abstract: 'Given a graph $G$, an incidence matrix $\mathcal{N}(G)$ is defined on the set of distinct isomorphism types of induced subgraphs of $G$. It is proved that Ulam’s conjecture is true if and only if the $\mathcal{N}$-matrix is a complete graph invariant. Several invariants of a graph are then shown to be reconstructible from its $\mathcal{N}$-matrix. The invariants include the characteristic polynomial, the rank polynomial, the number of spanning trees and the number of hamiltonian cycles in a graph. These results are stronger than the original results of Tutte in the sense that actual subgraphs are not used. It is also proved that the characteristic polynomial of a graph with minimum degree 1 can be computed from the characteristic polynomials of all its induced proper subgraphs. The ideas in Kocay’s lemma play a crucial role in most proofs. Kocay’s lemma is used to prove Whitney’s subgraph expansion theorem in a simple manner. The reconstructibility of the characteristic polynomial is then demonstrated as a direct consequence of Whitney’s theorem as formulated here.' author: - \| Bhalchandra D. Thatte\ Allan Wilson Centre for Molecular Ecology and Evolution,\ and Institute of Fundamental Sciences,\ Massey University, Palmerston North, New Zealand\ `[email protected]` date: \| Submitted: June 29, 2004\ Mathematics Subject Classifications: 05C50, 05C60 title: 'Kocay’s lemma, Whitney’s theorem, and some polynomial invariant reconstruction problems' --- Introduction {#sec-intro} ============ Suppose we are given the collection of induced subgraphs of a graph. There is a natural partial order on this collection defined by the induced subgraph relationship between members of the collection. An incidence matrix may be constructed to represent this relationship along with the multiplicities with which members of the collection appear as induced subgraphs of other members. Given such a matrix, is it possible to construct the graph or compute some of its invariants? Such a question is motivated by the treatment of chromatic polynomials in Biggs [@biggs1993]. Biggs demonstrates that it is possible to compute the chromatic polynomial of a graph from its incidence matrix. The idea of Kocay’s lemma in graph reconstruction theory is extremely useful in studying the question for other invariants. In this paper, we present several results on a relationship between Ulam’s reconstruction conjecture and the incidence matrix. Extending the reconstruction results of Tutte and Kocay, we show that many graph invariants can be computed from the incidence matrix. We then consider the problem of computing the characteristic polynomial of a graph from the characteristic polynomials of all induced proper subgraphs. Finally, we present a new short proof of Whitney’s subgraph expansion theorem, and demonstrate the reconstructibility of the characteristic polynomial of a graph using Whitney’s theorem. Notation -------- We consider only finite simple graphs in this paper. Let $G$ be a graph with vertex set $VG$ and edge set $EG$. The number of vertices of $G$ is denoted by $v(G)$ and the number of edges is denoted by $e(G)$. When $VG=\emptyset $, we denote $G$ by $\Phi$, and call the graph a [null graph]{}. When $EG=\emptyset $, we call the graph an [empty graph]{}. When $F$ is a subgraph of $G$, we write $F\subseteq G$, and when $F$ is a proper subgraph of $G$, we write $F\subsetneq G$. The subgraph of $G$ induced by $S \subseteq VG$ is the subgraph whose vertex set is $S$ and whose edge set contains all the edges having both end vertices in $S$. It is denoted by $G_S$. The subgraph of $G$ induced by $VG - S$ is denoted by $G-S$, or simply $G-u$ if $S=\{u\}$. A subgraph of $G$ with vertex set $V\subseteq VG$ and edge set $E\subseteq EG$ is denoted by $G_{(V,E)}$, or just $G_E$ if $V$ consists of the end vertices of edges in $E$. The same notation is used when $E = (e_1,e_2, \ldots ,e_k)$ is a tuple of edges, some of which may be identical. Isomorphism of two graphs $G$ and $H$ is denoted by $G\cong H$. For $i > 0$, a graph isomorphic to a cycle of length $i$ is denoted by $C_i$, and the number of cycles of length $i$ in $G$ is denoted by $\psi_i(G)$, where, as a convention, $C_i \cong K_i$ for $i \in \{1,2\}$. The number of hamiltonian cycles is denoted by a special symbol $ham(G)$ instead of $\psi_{v(G)}(G)$. While counting the number of subgraphs of a graph $G$ that are isomorphic to a graph $F$, it is important to make a distinction between induced subgraphs and edge subgraphs. The number of subgraphs of $G$ that are isomorphic to $F$ is denoted by ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, and the number of induced subgraphs of $G$ that are isomorphic to $F$ is denoted by $\displaystyle\binom{G}{F}$. The two numbers are related by $${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \sum_{H\|VH=VF}\displaystyle\binom{G}{H}{\mbox{$\left[\begin{array}{c}H\\#2\end{array}\right]$}}$$ where the summation is over distinct isomorphism types of graphs $H$. The characteristic polynomial of $G$ is denoted by $P(G;\lambda) = \sum_{i = 0}^{v(G)} c_i(G) \lambda^{v(G)-i}$. The collection $\mathcal{PD}(G) = \{P(G-S;\lambda)\mid S\subsetneq VG\}$ is called the [complete polynomial deck]{} of $G$. Note that a polynomial may appear in the collection more than once. The [rank]{} of a graph $G$, which has $comp(G)$ components, is defined by $v(G) - comp(G)$, and its [co-rank]{} is defined by $e(G)-v(G) + comp(G)$. The rank polynomial of $G$ is defined by $R(G;x,y) = \sum \rho_{rs}x^ry^s$, where $\rho_{rs}$ is the number of subgraphs of $G$ with rank $r$ and co-rank $s$. The set of consecutive integers from $a$ to $b$ is denoted by $[a,b]$; in particular, $N_k = [1,k]$. Ulam’s Conjecture ----------------- The [vertex deck]{} of a graph $G$ is the collection $\mathcal{VD}(G) = \{G-v\mid v \in VG\}$, where the subgraphs in the collection are ‘unlabelled’ (or isomorphism types). Note that the vertex deck is not exactly a set: an isomorphism type may appear more than once in the vertex deck. A Graph $G$ is said to be [reconstructible]{} if its isomorphism class is determined by $\mathcal{VD}(G)$. Ulam [@ulam1960] proposed the following conjecture. Graphs on more than 2 vertices are reconstructible. A property or an invariant of a graph $G$ is said to be reconstructible if it can be calculated from $\mathcal{VD}(G)$. For example, Kelly’s Lemma allows us to count the number of vertex-proper subgraphs of $G$ of any given type. \[lem-kelly\][(Kelly’s Lemma [@kelly1957])]{} If $F$ is a graph such that $v(F) < v(G)$ then $$\label{eq-kelly} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \frac{1}{v(G)-v(F)}\sum_{u\in VG}{\mbox{$\left[\begin{array}{c}G-u\\#2\end{array}\right]$}}$$ therefore, ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ is reconstructible from $\mathcal{VD}(G)$. Also, in Equation (\[eq-kelly\]), ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ and ${\mbox{$\left[\begin{array}{c}G-u\\#2\end{array}\right]$}}$ may be replaced by $\displaystyle\binom{G}{F}$ and $\displaystyle\binom{G-u}{F}$, respectively. Tutte [@tutte1977], [@tutte1984] proved the reconstructibility of the characteristic polynomial and the chromatic polynomial. Tutte’s results were simplified by an elegant counting argument by Kocay [@kocay1981]. This argument is useful to count certain subgraphs that span $VG$. Let $S = \{F_1, F_2, \ldots ,F_k\}$ be a family of graphs. Let $c(S,H)$ be the number of tuples $(X_1, X_2, \ldots ,X_k)$ of subgraphs of $H$ such that $X_i\cong F_i\, \forall \, i$, and $\cup_{i = 1}^{k}X_i = H$. We call it the number of [$S$-covers]{} of $H$. \[lem-kocay\] [(Kocay’s Lemma [@kocay1981])]{} $$\prod_{i = 1}^{k} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \sum_{X} c(S,X){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$$ where the summation is over all isomorphism types of subgraphs of $G$. Also, if $v(F_i)\, <\, v(G)\,\forall\, i$ then $ \sum_{X} c(S,X){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ over all isomorphism types $X$ of spanning subgraphs of $G$ can be reconstructed from the vertex deck of $G$. We refer to [@bondy1991] for a survey of reconstruction problems. The chromatic polynomial and the $\mathcal{N}$-matrix ----------------------------------------------------- Stronger reconstruction results on the chromatic polynomial were implicit in Whitney’s work [@whitney1932], although Ulam’s conjecture had not been posed at the time. Motivation for some of the work presented in this paper comes from Whitney’s work on the chromatic polynomials. The discussion of the chromatic polynomial presented here is based on [@biggs1993]. A graph $G$ is called [quasi-separable]{} if there exists $K\subsetneq VG$ such that $G_K$ is a complete graph and $G-K$ is disconnected. If $\|K\| \leq 1$ then $G$ is said to be separable. (Theorem 12.5 in [@biggs1993]) \[thm-biggs\] The chromatic polynomial of a graph is determined by its proper induced subgraphs that are not quasi-separable. The procedure of computing the chromatic polynomial may be outlined as follows. First a matrix $\mathcal{N}(G) = (N_{ij})$ is constructed. The rows and the columns of $\mathcal{N}(G)$ are indexed by induced subgraphs $\Lambda_1, \Lambda_2, \ldots , \Lambda_I = G$, which are the distinct isomorphism types of non-quasi-separable induced subgraphs of $G$. The list includes $K_1 = \Lambda_1$ and $K_2=\Lambda_2$. The indexing graphs are ordered in such a way that $v(\Lambda_i)$ are in non-decreasing order. The entry $N_{ij}$ is the number of induced subgraphs of $\Lambda_i$ that are isomorphic to $\Lambda_j$. It is a lower triangular matrix with diagonal entries 1. The computation of the chromatic polynomial is performed by a recursive procedure beginning with the first row of the $\mathcal{N}$-matrix, computing at each step certain polynomials in terms of the corresponding polynomials for non-quasi-separable induced subgraphs on fewer vertices. A few observations about the procedure are useful to motivate the work in this paper. The graphs $C_4$ and $K_4$ are the only non-quasi-separable graphs on $4$ vertices. Also, for any $i$, $N_{i1}=v(\Lambda_i)$, and $N_{i2}= e(\Lambda_i)$. Therefore, graphs on 4 or fewer vertices that index the first few rows of the $\mathcal{N}$-matrix can be inferred from the matrix entries. Therefore, we conclude that the computation of the chromatic polynomial can be performed on the matrix entries alone, even if the induced subgraphs indexing the rows and the columns of $\mathcal{N}(G)$ are unspecified. Therefore, we will think of the $\mathcal{N}$-matrix as [unlabelled ]{}, that is, we will assume that the induced subgraphs indexing the rows and the columns are not given. A natural question is what other invariants can be computed from the (unlabelled) $\mathcal{N}$-matrix? Obviously, the characteristic polynomial $P(G;\lambda)$ cannot always be computed from $\mathcal{N}(G)$. For example, the only non-quasi-separable induced subgraphs of any tree $T$ are $K_1$ and $K_2$, so $P(T;\lambda)$ cannot be computed from $\mathcal{N}(T)$. Therefore, we omit the restriction of non-quasi-separability on the induced subgraphs used in the construction of the incidence matrix. We then investigate which invariants of a graph are determined by its $\mathcal{N}$-matrix. The Sections \[sec-n-rec1\] and \[sec-nmatrix\] are devoted to the study of reconstruction from the $\mathcal{N}$-matrix. In Section \[sec-n-rec1\], we formally define the $\mathcal{N}$-matrix, and the related concept of the edge poset of induced subgraphs of a graph. We then prove several basic results on the relationship between the $\mathcal{N}$-matrix, the edge labelled poset and reconstruction. In particular, we show that Ulam’s conjecture is true if and only if the $\mathcal{N}$-matrix itself is a complete graph invariant. We then prove that Ulam’s conjecture is true if and only if the edge labelled poset has no non-trivial automorphisms. We also prove the $\mathcal{N}$-matrix reconstructibility of trees and forests. In Section \[sec-nmatrix\] we compute several invariants of a graph from its $\mathcal{N}$-matrix. We prove that the characteristic polynomial $P(G;\lambda)$ of a graph $G$, its rank polynomial $R(G;x,y)$, the number of spanning trees in $G$, the number of Hamiltonian cycles in $G$ etc., can be computed from $\mathcal{N}(G)$. In the standard proof of the reconstructibility of these invariants, one first counts the disconnected subgraphs of each type, (see [@bondy1991]). In view of Theorem \[thm-equiv-disc\], the proofs in Section \[sec-nmatrix\] are more involved. Theorem \[thm-equiv-disc\] implies that if there are counter examples to Ulam’s conjecture then there are many more counter examples to reconstruction from the $\mathcal{N}$-matrix. Therefore, we hope that the study of $\mathcal{N}$-matrix reconstructibility will highlight new difficulties. Similar generalisations of the reconstruction problem were also suggested by Tutte, (notes on pp. 123-124 in [@tutte1984]). Reconstruction of the characteristic polynomial ----------------------------------------------- The proof of the reconstructibility of the characteristic polynomial of a graph from its $\mathcal{N}$-matrix is also of independent technical interest, since other authors have considered the question of computing $P(G; \lambda)$ given the [polynomial deck]{} $\{P(G-u;\lambda); u \in VG\}$. This question was originally proposed by Gutman and Cvetkovic [@gut-cve1975], and has been studied by others, for example, [@schwenk1979] & [@sciriha2002]. This question remains open. So we consider a weaker question in Section \[sec-poly\]: the question of computing the characteristic polynomial of a graph from its complete polynomial deck. Here we present basic facts about the characteristic polynomial, and outline the idea of Section \[sec-poly\]. A graph is called [elementary]{} if each of its components is $1$-regular or $2$-regular. In other words, each component of an elementary graph is a single edge ($K_2$) or a cycle ($C_r; r >2$). Let $L_i$ be the collection of all unlabelled $i$-vertex elementary graphs. So, $L_0 = \{\Phi \}$, $L_1 = \emptyset$, $L_2 = \{K_2\}$, and so on. \[lem-sachs\] (Proposition 7.3 in [@biggs1993]) Coefficients of the characteristic polynomial of a graph $G$ are given by $$(-1)^i c_i(G) = \sum_{F \in L_i, F\subseteq G} (-1)^{r(F)} 2^{s(F)}$$ where $r(F)$ and $s(F)$ are the rank and the co-rank of $F$, respectively. Thus, $c_0(G) = 1$, $c_1(G) = 0$, and $c_2(G) = e(G)$. (Note 2d in [@biggs1993]) \[lem-derivative\] Let $P'(G;\lambda)$ denote the first derivative of $P(G;\lambda)$ with respect to $\lambda$. Then, $$\label{eq-derivative} P'(G;\lambda) = \sum_{u\in VG}P(G-u;\lambda)$$ From the above two lemmas, it is clear that the problem of reconstructing a characteristic polynomial (either from the vertex deck or the complete polynomial deck) reduces to computing the coefficient $c_{v(G)}(G)$, which is the constant term in $P(G;\lambda)$. This in turn is a problem of counting the elementary spanning subgraphs of $G$ - a problem that can be solved using Kocay’s Lemma in case of reconstruction from the vertex deck. Motivated by Kocay’s Lemma, we ask the following question. Suppose the coefficients $c_{i_1}(G), c_{i_1}(G), \ldots , c_{i_k}(G)$ are known, and $i_1 + i_2 + \ldots + i_k \geq v(G)$. If the coefficients $c_{i_j}; 1 \leq j \leq k$ are multiplied, can we get some information about the spanning subgraphs of $G$? This is especially tempting if $i_1+i_2 + \ldots + i_k = v(G)$, since the product is expected to have some relationship with the disconnected spanning elementary subgraphs of $G$. This idea is explored in Section \[sec-poly\]. In Section \[sec-whitney\], we present a very simple new proof of Whitney’s subgraph expansion theorem, again based on Kocay’s lemma. We then present a more direct argument to compute the characteristic polynomial of a graph from its vertex deck, based on our formulation of Whitney’s theorem. Ulam’s conjecture and the $\mathcal{N}$-matrix {#sec-n-rec1} ============================================== Let $\Lambda(G) = \{\Lambda_i; \,i \in [1,I]\}$ be the set of distinct isomorphism types of nonempty induced subgraphs of $G$. We call this the $\Lambda$-deck of $G$. Let $\mathcal{N}(G)=(N_{ij})$ be an $I$ x $I$ incidence matrix where $N_{ij}$ is the number of induced subgraphs of $\Lambda_i$ that are isomorphic to $\Lambda_j$. Thus $N_{ii}$ is 1 for all $i \in [1,I]$. We call an invariant of a graph [$\mathcal{N}$-matrix reconstructible]{} if it can be computed from the (unlabelled) $\mathcal{N}$-matrix of the graph. As an example, the ladder graph $L_3$ and its collection of distinct induced subgraphs with nonempty edge sets are shown in Figure 1. Below each graph (except $L_3$) is shown its multiplicity as an induced subgraph in $L_3$ and its name. \[fig-l3\] (200,400) (0,100) (200,100) (0,400) (200,400) (100,200) (100,300) (0,100)[(1,0)[200]{}]{} (0,100)[(0,1)[300]{}]{} (0,400)[(1,0)[200]{}]{} (200,100)[(0,1)[300]{}]{} (0,100)[(1,1)[100]{}]{} (200,100)[(-1,1)[100]{}]{} (0,400)[(1,-1)[100]{}]{} (200,400)[(-1,-1)[100]{}]{} (100,200)[(0,1)[100]{}]{} (0,0)[$L_3 = \Lambda_9$]{} (100,400) (100,150) (100,300) (100,150)[(0,1)[150]{}]{} (50,0)[9$\Lambda_1$]{} (200,400) (0,150) (200,150) (100,300) (0,150)[(1,0)[200]{}]{} (50,0)[6$\Lambda_2$]{} (200,400) (0,150) (150,150) (0,300) (0,150)[(1,0)[150]{}]{} (0,150)[(0,1)[150]{}]{} (10,0)[12$\Lambda_3$]{} (200,400) (0,150) (200,150) (100,300) (0,150)[(1,0)[200]{}]{} (0,150)[(2,3)[100]{}]{} (200,150)[(-2,3)[100]{}]{} (50,0)[2$\Lambda_4$]{} (50,400) (50,100) (50,200) (50,300) (50,400) (50,100)[(0,1)[100]{}]{} (50,200)[(0,1)[100]{}]{} (50,300)[(0,1)[100]{}]{} (0,0)[6$\Lambda_5$]{} (200,400) (0,150) (200,150) (100,250) (100,350) (0,150)[(1,0)[200]{}]{} (0,150)[(1,1)[100]{}]{} (200,150)[(-1,1)[100]{}]{} (100,250)[(0,1)[100]{}]{} (50,0)[6$\Lambda_6$]{} (200,400) (0,150) (200,150) (0,350) (200,350) (0,150)[(1,0)[200]{}]{} (0,150)[(0,1)[200]{}]{} (0,350)[(1,0)[200]{}]{} (200,150)[(0,1)[200]{}]{} (50,0)[3$\Lambda_7$]{} (200,400) (0,150) (200,150) (0,350) (200,350) (300,250) (0,150)[(1,0)[200]{}]{} (0,150)[(0,1)[200]{}]{} (0,350)[(1,0)[200]{}]{} (200,150)[(0,1)[200]{}]{} (200,150)[(1,1)[100]{}]{} (200,350)[(1,-1)[100]{}]{} (50,0)[6$\Lambda_8$]{} The rows and the columns of $\mathcal{N}$-matrix of $L_3$ are both indexed by $\Lambda_1$ to $\Lambda_9$. The $\mathcal{N}$-matrix of $L_3$ is shown below. $$\mathcal{N}(L_3) = \begin{pmatrix} 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 2 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ 3 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\ 3 & 2 & 2 & 0 & 1 & 0 & 0 & 0 & 0 \\ 4 & 1 & 2 & 1 & 0 & 1 & 0 & 0 & 0 \\ 4 & 0 & 4 & 0 & 0 & 0 & 1 & 0 & 0 \\ 6 & 3 & 6 & 1 & 2 & 2 & 1 & 1 & 0 \\ 9 & 6 & 12 & 2 & 6 & 6 & 3 & 6 & 1 \\ \end{pmatrix}$$ Let us associate an [edge labelled poset]{} with the graph $G$. Define a partial order $\preceq $ on the set $\Lambda(G)$ as follows: $\Lambda_j \preceq \Lambda_k$ if and only if $\Lambda_j$ is an induced subgraph of $\Lambda_k$. This poset is denoted by $(\Lambda(G), \preceq)$. We make the poset $(\Lambda(G), \preceq)$ an edge labelled poset by assigning a positive integer to every edge of its Hasse diagram, such that if $\Lambda_k$ covers $\Lambda_j$ then the edge label on $\Lambda_j$-$\Lambda_k$ is $\displaystyle\binom{\Lambda_k}{\Lambda_j}$. We say that two edge labelled posets are isomorphic if they are isomorphic as posets, and there is an isomorphism between them that preserves the edge labels. This naturally leads to the notion of the [abstract edge labelled poset]{} of $G$: it is the isomorphism class of the edge labelled poset of $G$. Note that the notion of the abstract edge labelled poset of a graph is not to be confused with the isomorphism class of the Hasse diagram as a graph. An isomorphism from an edge labelled poset to itself is called an automorphism of the edge labelled poset. We denote the abstract edge labelled poset of $G$ by $\mathcal{ELP}(G)$. The Hasse diagram of the abstract edge labelled poset is simply the Hasse diagram of the edge labelled poset of $G$ with labels $\Lambda_i$ removed. The Hasse diagram of $\mathcal{ELP}(L_3)$ is shown in Figure 2. \[fig-hasse\] (600,1300) (300,100) (0,400) (300,400) (600,400) (0,700) (300,700) (600,700) (300,1000) (300,1300) (300,100)[(-1,1)[300]{}]{} (100,200)[2]{} (300,100)[(0,1)[300]{}]{} (470,200)[1]{} (300,100)[(1,1)[300]{}]{} (320,250)[3]{} (0,400)[(0,1)[300]{}]{} (-80,500)[4]{} (0,400)[(1,1)[300]{}]{} (100,550)[2]{} (0,400)[(2,1)[600]{}]{} (120,400)[2]{} (300,400)[(0,1)[300]{}]{} (320,450)[1]{} (600,400)[(0,1)[300]{}]{} (620,500)[2]{} (600,400)[(-1,1)[300]{}]{} (450,465)[1]{} (0,700)[(1,1)[300]{}]{} (50,830)[1]{} (300,700)[(0,1)[300]{}]{} (320,800)[2]{} (600,700)[(-1,1)[300]{}]{} (530,830)[2]{} (300,1000)[(0,1)[300]{}]{} (320,1100)[6]{} \[lem-rank\] There is a rank function on $\rho $ on $\mathcal{ELP}(G)$ such that $\rho(\Lambda_i) = \rho(\Lambda_j) +1$ whenever $\Lambda_i $ covers $\Lambda_j$. Each $\Lambda_i$ in $\Lambda(G)$ is nonempty. Therefore, for each $\Lambda_i$ in $\Lambda(G)$ and for each $k$ such that $2 \leq k \leq v(\Lambda_i)$ there is at least one nonempty induced subgraph $\Lambda_j$ of $\Lambda_i$ such that $v(\Lambda_j) = k$. Moreover, empty induced subgraphs do not belong to $\Lambda(G)$. Therefore, $\rho(\Lambda_i) = v(\Lambda_i)$ meets the requirements of a rank function. Stanley [@stanley1997] defines a rank function such that the $\rho(x) = 0$ for a minimal element $x$. But we have deviated from that convention since $\rho(\Lambda_i) = v(\Lambda_i)$ for each $\Lambda_i \in \Lambda(G)$ is more convenient here. We now demonstrate that $\mathcal{N}(G)$ and $\mathcal{ELP}(G)$ are really equivalent, that is, they can be constructed from each other. \[lem-kelly1\] Let $F$ and $H$ be two graphs, and let $q$ be an integer such that $v(F) \leq q \leq v(H)$. Then $$\label{eq-kelly1} \sum_{X\| v(X) = q}\displaystyle\binom{H}{X}\displaystyle\binom{X}{F} = \displaystyle\binom{v(H)-v(F)}{q-v(F)}\displaystyle\binom{H}{F}$$ where the summation is over distinct isomorphism types $X$. This is similar to Kelly’s Lemma \[lem-kelly\]. Each induced subgraph of $H$ that is isomorphic to $F$ is also an induced subgraph of $\displaystyle\binom{v(H)-v(F)}{q-v(F)}$ induced subgraphs of $H$ that have $q$ vertices. \[lem-elp\_eq\_n\] The structures $\mathcal{N}(G)$ and $\mathcal{ELP}(G)$ can be constructed from each other. We first show how $\mathcal{N}(G)$ is constructed from $\mathcal{ELP}(G)$. The matrix $\mathcal{N}(G)$ is an $I\times I$ matrix where $I$ is the number of points in $\mathcal{ELP}(G)$. Without the loss of generality, suppose that the points of $\mathcal{ELP}(G)$ are labelled from $\Lambda_1$ to $\Lambda_I$ such that if $\rho(\Lambda_i) < \rho(\Lambda_j)$ then $i < j$, where $\rho$ is the rank function defined in Lemma \[lem-rank\]. Correspondingly, the rows and the columns of $\mathcal{N}(G)$ are indexed from $\Lambda_1$ to $\Lambda_I$. The edge labels in $\mathcal{ELP}(G)$ immediately give some of the entries in $\mathcal{N}(G)$: if $\Lambda_i $ covers $\Lambda_j$ then $N_{ij}$ is the label on the edge joining $\Lambda_i $ and $\Lambda_j$. The diagonal entries are 1. Except $N_{11}$, all the other entries in the first row are 0. We construct the remaining entries of $\mathcal{N}(G)$ by induction on the rank. The base case is rank 2. It corresponds to the first row, and is already filled. Let $f(r)$ denote the number of points of $\mathcal{ELP}(G)$ that have rank at most $r$. Suppose now that the first $f(r)$ rows of $\mathcal{N}(G)$ are filled for some $r \geq 2$. Let $\Lambda_i$ be a graph of rank $r+1$, and let $\Lambda_j$ be a graph of rank at most $r$. Then $N_{ij}$ is computed by applying Lemma \[lem-kelly1\] with $q = r$. $$\sum_{\Lambda_k\| \rho(\Lambda_k) = r} \displaystyle\binom{\Lambda_i}{\Lambda_k} \displaystyle\binom{\Lambda_k}{\Lambda_j} = \displaystyle\binom{v(\Lambda_i)-v(\Lambda_j)}{r-v(\Lambda_j)} \displaystyle\binom{\Lambda_i}{\Lambda_j} = (r+1-v(\Lambda_j))N_{ij}$$ On the LHS, $\displaystyle\binom{\Lambda_k}{\Lambda_j}$ are known by induction hypothesis. Since $\Lambda_k$ are the graphs covered by $\Lambda_i$, $\displaystyle\binom{\Lambda_i}{\Lambda_k}$ are the edge labels. Therefore, $N_{ij}$ can be computed. This completes the construction of $\mathcal{N}(G)$ from $\mathcal{ELP}(G)$. To construct $\mathcal{ELP}(G)$ from $\mathcal{N}(G)$, define a partial order $\preceq $ on $\{\Lambda_1, \Lambda_2, \ldots , \Lambda_I\}$ as follows: $\Lambda_j \preceq \Lambda_i$ if $N_{ij} \neq 0$. In this poset, if $\Lambda_i$ covers $\Lambda_j$ then assign an edge label $N_{ij}$ to the edge $\Lambda_j-\Lambda_i$ of the Hasse diagram of the poset. This completes the construction of $\mathcal{ELP}(G)$ from $\mathcal{N}(G)$. \[lem-num-vertices\] Given $\mathcal{N}(G)$, $v(\Lambda_i)$ and $e(\Lambda_i)$ can be counted for each graph in $\Lambda(G)$. There is a unique row in $\mathcal{N}(G)$ that has only one nonzero entry (the diagonal entry 1). This row corresponds to $\Lambda_1 \cong K_2$, and we assume it to be the first row. Now $e(\Lambda_i) = N_{i1}$ for each $\Lambda_i$. By Lemma \[lem-elp\_eq\_n\], $\mathcal{ELP}(G)$ is uniquely constructed. By Lemma \[lem-rank\], the rank function of the poset defined by $\rho(\Lambda_1)=2$ gives $v(\Lambda_i)=\rho(\Lambda_i)$ for each $\Lambda_i$. Now on, without the loss of generality, we will assume that the nonisomorphic induced subgraphs $\Lambda_1, \Lambda_2, \ldots, \Lambda_I$ of a graph $G$ under consideration are ordered so that $v(\Lambda_i)$ are in a non-decreasing order. The first row will correspond to $\Lambda_1 \cong K_2$ and the last row to $\Lambda_I \cong G$. \[lem-kelly3\] The collection $\{\mathcal{N}(G-u)\| u \in VG, e(G-u) > 0\}$ is unambiguously determined by $\mathcal{N}(G)$. Note that this collection is a “multiset”, that is, an $\mathcal{N}$-matrix may appear multiple times in the collection. Let $j \neq I$. The graph $\Lambda_j$ is a vertex deleted subgraph of $G$ if and only if for all $i \neq j \neq I$, $N_{ij}=0$. Now $N(\Lambda_j)$ is obtained by deleting $k$’th row and $k$’th column for each $k$ such that $N_{jk} = 0$. A multiplicity $N_{Ij}$ is assigned to $N(\Lambda_j)$. Equivalently, we can construct $\mathcal{ELP}(G)$ by Lemma \[lem-elp\_eq\_n\], then construct the down set $\mathcal{ELP}(\Lambda_j)$ of each $\Lambda_j$ that is covered by $\Lambda_I = G$, and then construct $\mathcal{N}(\Lambda_j)$, and assign it a multiplicity equal to the edge label on $\Lambda_I-\Lambda_j$. [Remark]{} It is is possible that for distinct $j$ and $k$, the matrices $N(\Lambda_j)$ and $N(\Lambda_k)$ are equal. In this case a multiplicity $N_{Ij}$ is assigned to $N(\Lambda_j)$ and $N_{Ik}$ is assigned to $N(\Lambda_k)$ while constructing the above collection. \[lem-empty\] Let $rK_1$ be the $r$-vertex empty graph. The number of induced subgraphs of $G$ isomorphic to $rK_1$ is determined by $\mathcal{N}(G)$. The required number is $$\displaystyle\binom{G}{rK_1} = \displaystyle\binom{v(G)}{r} - \sum_{j\mid v(\Lambda_j) = r}N_{Ij}$$ where indices $j$ in the summation are determined by Lemma \[lem-num-vertices\]. We are interested in the question of reconstructing a graph $G$ or some of its invariants given $\mathcal{N}(G)$. As indicated earlier, we will assume that the induced subgraphs $\Lambda_i; i \in [1,I]$ are not given. We have the following relationship between Ulam’s conjecture and the $\mathcal{N}$-matrix reconstructibility. \[prop:equiv-un\] Ulam’s conjecture is true if and only if all graphs on three or more vertices are $\mathcal{N}$-matrix reconstructible. Proof of [if]{}: by Lemma \[lem-kelly1\], $\mathcal{N}(G)$ is constructed from $\mathcal{VD}(G)$. Therefore, Ulam’s conjecture is true if all graphs are $\mathcal{N}$-matrix reconstructible. In fact, a graph is reconstructible if it is $\mathcal{N}$-matrix reconstructible. Proof of [only if]{}: this is proved by induction on the number of vertices. Let Ulam’s conjecture be true. Since $N_{i1} = e(\Lambda_i)$ for all $i$, every non-empty three vertex graph is $\mathcal{N}$-matrix reconstructible. Now, let all graphs on at most $n$ vertices, where $n\geq 3$, be $\mathcal{N}$-matrix reconstructible. Let $G$ be a graph on $n+1$ vertices. By Lemma \[lem-kelly3\], the collection $\{\mathcal{N}(G-u); u\in VG, e(G-u) > 0\}$ is unambiguously determined by $\mathcal{N}(G)$. The number of empty graphs in $\mathcal{VD}(G)$ is 0 or 1, and is determined by Lemma \[lem-empty\]. Therefore, by induction hypothesis, $\mathcal{VD}(G)$ is uniquely determined. Now the result follows from the assumption that Ulam’s conjecture is true. Since $\mathcal{N}(G)$ and $\mathcal{ELP}(G)$ are equivalent by Lemma \[lem-elp\_eq\_n\], we rephrase Proposition \[prop:equiv-un\] as follows. \[prop:equiv-elp\] Ulam’s conjecture is true if and only if all graphs on three or more vertices are reconstructible from their abstract edge labelled posets. We would like to point out that reconstructing $G$ from $\mathcal{N}(G)$ or from $\mathcal{ELP}(G)$ is not proved to be equivalent to reconstructing $G$ from $\mathcal{VD}(G)$. This poses a difficulty. For example, proving $\mathcal{N}$-matrix reconstructibility of disconnected graphs is as hard as Ulam’s conjecture, although disconnected graphs are known to be vertex reconstructible. This is proved below. For graphs $X$ and $Y$, we use the notation $X+Y$ to denote a graph that is a disjoint union of two graphs isomorphic to $X$ and $Y$, respectively. Suppose $G$ and $H$ are connected graphs having the same vertex deck. Consider graphs $2G = G+G$ and $2H=H+H$. \[lem-kelly2\] Let $F$ be a graph on fewer than $2v(G)$ vertices. If $F$ has a component isomorphic to $G$ (in which case we write $F=G+X$) then $\displaystyle\binom{2G}{G+X} = \displaystyle\binom{2H}{H+X}$. If $F$ has no component isomorphic to $G$ then $\displaystyle\binom{2G}{F} = \displaystyle\binom{2H}{F}$. When $F=G+X$, $X$ must have fewer than $v(G)-1$ vertices. Since $G$ and $H$ have identical vertex decks, by Kelly’s Lemma \[lem-kelly\], $\displaystyle\binom{G}{X} = \displaystyle\binom{H}{X}$. Therefore, $\displaystyle\binom{2G}{G+X} = \displaystyle\binom{2H}{H+X}$. When $F$ does not have a component isomorphic to $G$, then if $F$ has a component on $v(G)$ vertices then $\displaystyle\binom{2G}{F} = \displaystyle\binom{2H}{F} = 0$. Therefore, assume that all components of $F$ have at most $v(G)-1$ vertices. Any realisation of $F$ as an induced subgraph of $2G$ is a disjoint union of graphs isomorphic to $X$ and $Y$ such that $X$ is an induced subgraph of one component of $2G$ and $Y$ is an induced subgraph of the other component of $2G$. Moreover, $v(X) < v(G)$ and $v(Y) < v(G)$. Now $\displaystyle\binom{2G}{F}=\displaystyle\binom{2H}{F}$ follows from the fact that $G$ and $H$ have identical vertex decks and Kelly’s Lemma. The following corollary is an immediate consequence of the above lemma. \[cor-correspondence\] Define a correspondence $f$ between $\Lambda(2G)$ and $\Lambda(2H)$ as follows. 1. $f(2G) = 2H$ 2. $F\in \Lambda(2G)$ is not $2G$ but has a component isomorphic to $G$. We write $F= G+X$, and set $f(F) = H+X$. 3. $F\in \Lambda(2G)$ has no component isomorphic to $G$. In this case we set $f(F) = F$. The correspondence defined above is a bijection. \[lem-2g2h\] $\mathcal{N}(2G) = \mathcal{N}(2H)$. For the bijection $f$ between non-empty induced subgraphs of $2G$ and $2H$ that was defined in Corollary \[cor-correspondence\], we show that, for any two nonisomorphic induced subgraphs $F_1$ and $F_2$ of $2G$, $$\label{eq-2g2h} \displaystyle\binom{F_2}{F_1} = \displaystyle\binom{f(F_2)}{f(F_1)}$$ In view of Corollary \[cor-correspondence\], it is sufficient to show this when at least one of the graphs $F_1$ and $F_2$ has a component isomorphic to $G$. 1. When $F_2 = 2G$, then Equation (\[eq-2g2h\]) follows from Lemma \[lem-kelly2\] and Corollary \[cor-correspondence\]. 2. When $F_1 = G+X$ and $F_2 = G+Y$, and $v(X) < v(G)$ and $v(Y) < v(G)$, we have $\displaystyle\binom{F_2}{F_1} = \displaystyle\binom{Y}{X} = \displaystyle\binom{H+Y}{H+X} = \displaystyle\binom{f(F_2)}{f(F_1)}$. 3. $F_2 = G+Z$, $v(Z) < v(G)$, but $F_1$ has no component isomorphic to $G$. In this case, any realisation of $F_1$ as an induced subgraph of $F_2$ may be represented (possibly in many ways) as $F_1 = X+Y$ where $X$ is an induced proper subgraph of the component $G$ of $F_2$ and $Y$ is an induced subgraph of $Z$. Moreover, $v(X) < v(G)$ and $v(Y) < v(G)$. Since $\displaystyle\binom{G}{X} = \displaystyle\binom{H}{X}$, we have $\displaystyle\binom{G+Z}{F_1} = \displaystyle\binom{H+Z}{f(F_1)}$. Note that the actual value of $\displaystyle\binom{F_2}{F_1}$ may be written by considering all possible ways of realising $F_1$ as an induced subgraph of $G+Z$. Thus we have shown Equation (\[eq-2g2h\]) for arbitrary non-empty induced subgraphs of $2G$, which implies the result. \[thm-equiv-disc\] Ulam’s conjecture is true if and only if disconnected graphs on three or more vertices are $\mathcal{N}$-matrix reconstructible. The [only if]{} part follows from Proposition \[prop:equiv-un\]. The [if]{} part is proved by contradiction. Suppose $G$ and $H$ are connected nonisomorphic graphs with the same vertex deck, that is, they are a counter example to Ulam’s conjecture. Then $2G$ and $2H$ are nonisomorphic but $\mathcal{N}(2G) = \mathcal{N}(2H)$ by Lemma \[lem-2g2h\]. Therefore, $2G$ and $2H$ are disconnected graphs that are not $\mathcal{N}$-matrix reconstructible. The following result is proved along the lines of Lemma \[lem-2g2h\]. Ulam’s conjecture is true if and only if the edge labelled poset of each graph has only the trivial automorphism. The proof of [only if]{} is done by contradiction. Suppose that $\mathcal{ELP}(G)$ has a nontrivial automorphism $\sigma$. Then there are nonisomorphic induced subgraphs $\Lambda_i$ and $\Lambda_j$ of $G$ such that $\sigma(\Lambda_i) = \Lambda_j$. The downsets (or the edge labelled posets) of $\Lambda_i$ and $\Lambda_j$ must be isomorphic. Therefore, by Proposition \[prop:equiv-elp\], there is a counter example to Ulam’s conjecture. The proof of [if]{} is also done by contradiction. Suppose that Ulam’s conjecture is false, and $G$ and $H$ are connected nonisomorphic graphs having identical vertex decks. We show that $\mathcal{ELP}(G+H)$ has a nontrivial automorphism. Define a bijective map $\sigma : \Lambda(G+H)\rightarrow \Lambda(G+H)$ as follows. 1. The graph $G+H$ is mapped to itself. 2. If $\Lambda_i\in \Lambda(G+H)$ has a component isomorphic to $G$, then denote $\Lambda_i$ by $G+X$, where $X$ is a proper subgraph of the component isomorphic to $H$. In this case, set $\sigma(G+X) = H+X$. 3. If $\Lambda_i$ is $H+X$, where $X$ is a proper subgraph of the component isomorphic to $G$, then set $\sigma(H+X) = G+X$. 4. For all other graphs $\Lambda_i \in \Lambda(G+H)$, $\sigma(\Lambda_i) = \Lambda_i$. We now show that $\sigma$ is an automorphism of $\mathcal{ELP}(G+H)$. That is, we show that $\displaystyle\binom{\Lambda_i}{\Lambda_j} = \displaystyle\binom{\sigma(\Lambda_i)}{\sigma(\Lambda_j)}$ for any two graphs $\Lambda_i $ and $\Lambda_j$ in $\Lambda(G+H)$. We have to consider only the case in which at least one of $\Lambda_i $ and $\Lambda_j$ has a component isomorphic to $G$ or $H$, and $v(\Lambda_j) \leq v(\Lambda_i)$. 1. $\Lambda_j = G+X$ and $\Lambda_i = G+H$. In this case,\ $\displaystyle\binom{G+H}{G+X} = \displaystyle\binom{H}{X} = \displaystyle\binom{G}{X} = \displaystyle\binom{H+G}{H+X} = \displaystyle\binom{\sigma(G+H)}{\sigma(G+X)}$. 2. $\Lambda_j = G+X$ and $\Lambda_i = G+Y$ and $v(Y) < v(G)=v(H)$. In this case,\ $\displaystyle\binom{G+Y}{G+X} = \displaystyle\binom{Y}{X} = \displaystyle\binom{H+Y}{H+X} = \displaystyle\binom{\sigma(G+Y)}{\sigma(G+X)}$. 3. $\Lambda_j = G+X$ and $\Lambda_j = H+Y$ and $Y\ncong G$. In this case,\ $\displaystyle\binom{H+Y}{G+X} = \displaystyle\binom{G+Y}{H+X} = 0$. 4. $\Lambda_j = G+X$ and $\Lambda_i $ has no component isomorphic to $G$ or $H$. In this case,\ $\displaystyle\binom{\Lambda_i}{G+X} = \displaystyle\binom{\Lambda_i}{H+X} = 0$. 5. $\Lambda_j$ has no component isomorphic to $G$ or $H$, and $\Lambda_i = G+H$. This is trivial since $\sigma(\Lambda_j) = \Lambda_j$ and $\sigma(G+H) = G+H$ 6. $\Lambda_j$ has no component isomorphic to $G$ or $H$ and $\Lambda_i = G+X$, where $v(X) < v(G)=v(H)$. In this case, a realisation of $\Lambda_j$ as an induced subgraph of $G+X$ may be written as $\Lambda_j = Y+Z$, where $Y$ is an induced subgraph of $G$ and $Z$ is an induced subgraph of $X$. Since, $\displaystyle\binom{G}{Y}=\displaystyle\binom{H}{Y}$, the number of such realisations is $\displaystyle\binom{G}{Y}\displaystyle\binom{X}{Z} = \displaystyle\binom{H}{Y}\displaystyle\binom{X}{Z}$. By summing over all possible ways of realising $\Lambda_j$ as an induced subgraph of $G+X$, we get $\displaystyle\binom{G+X}{\Lambda_j} = \displaystyle\binom{H+X}{\Lambda_j}$. 7. All the above arguments are valid when $G$ and $H$ are interchanged. Thus we have constructed a non-trivial automorphism, completing the [if]{} part. We conclude this section with a result on trees. \[thm-trees\] Trees and forests are $\mathcal{N}$-matrix reconstructible. The class of simple acyclic graphs is closed under vertex deletion. Therefore, we can use the method in the proof of Proposition \[prop:equiv-un\]. Let $T$ be a tree or a non-empty forest on three or more vertices. We prove by induction on $v(T)$ that $T$ is uniquely reconstructible from $\mathcal{N}(T)$. The base case is $v(T) = 3$. All graphs on 3 vertices are $\mathcal{N}$-matrix reconstructible by Lemma \[lem-num-vertices\]. Suppose each acyclic graphs on at most $k$ can be recognised and reconstructed from its $\mathcal{N}$-matrix. Let $v(T) = k+1$. By Lemma \[lem-kelly3\], the collection $\{\mathcal{N}(T-u)\| u \in VT\}$ is unambiguously determined. Then by induction hypothesis, $T-u$ are determined (along with their multiplicities). The subgraphs in the vertex deck that are not determined by Lemma \[lem-kelly3\] are the ones having no edges. Since Ulam’s conjecture has been proved for trees and disconnected simple graphs in [@kelly1957], $T$ is $\mathcal{N}$-matrix reconstructible. [Remark]{} If Ulam’s conjecture is true for a class of graphs that is closed under vertex deletion, then the class is also $\mathcal{N}$-matrix reconstructible. Tutte-Kocay theory on the $\mathcal{N}$-matrix. {#sec-nmatrix} =============================================== In this section we will compute several invariants of a graph $G$ from its $\mathcal{N}$-matrix. The invariants include the number of spanning trees, the number of spanning unicyclic subgraphs containing a cycle of specified length, the characteristic polynomial and the rank polynomial. [An outline of the proof.]{} First we outline how the above mentioned invariants are calculated from the vertex deck using Kocay’s Lemma. 1. Suppose the graphs $F_1, F_2, \ldots , F_k$ satisfy $\sum_i v(F_i) = v(G)$ and $v(F_i)< v(G) \forall i$. Kocay’s Lemma then gives the number of disconnected spanning subgraphs having components isomorphic to $F_1, F_2, \ldots , F_k$. 2. Kacay’s lemma is then applied to $F_1 = F_2 = \ldots = F_k = K_2$, where $k=v(G)-1$. Since disconnected spanning subgraphs of each type are counted in the first step, we can now count the number of spanning trees. 3. The second step is repeated with $k=v(G)$. Since the number of spanning trees and disconnected spanning subgraphs of each type are known from the first two steps, we can now count the number of hamiltonian cycles. 4. Once the above three steps are completed, many other invariants, such as the characteristic polynomial, rank polynomial, etc. are easily computed. The procedure outlined above cannot be implemented on the $\mathcal{N}$-matrix in a straight forward manner. We do not know all the induced proper subgraphs. But we observe that the above procedure essentially reduces counting certain spanning subgraphs to counting them on vertex proper subgraphs. It turns out that we do not really need the number of vertex proper subgraphs of each type. We only need to know the ‘cycle structure’, that is, $\psi_i(\Lambda_j)$ for each $i \leq v(\Lambda_j)$, for each $j < I$. Next we outline the strategy to construct the cycle structure. Suppose $X, Y, \ldots$ is a list of some graph invariants that are either polynomials or numbers, for example, the number of hamiltonian cycles in a graph or the chromatic polynomial of a graph. We say that an invariant $Z$ can be reduced to invariants $X, Y, \ldots $ (or $Z$ has a reduction on the $\mathcal{N}$-matrix) if for each graph $G$ having a non-empty edge set, 1. $Z(G)$ can be written as $Z(G) = \Theta(X(G_U), Y(G_V), ...)$ where $\Theta(x,y,...)$ is a polynomial in $x,y, \ldots$, and $U, V, \dots $ are proper subsets of $VG$. 2. the coefficient of each term in the polynomial can be computed from $\mathcal{N}(G)$. Proving an identity that gives a reduction of an invariant $Z$ as in the above equation is not in itself sufficient to claim that $Z$ is $\mathcal{N}$-matrix reconstructible. If the invariants $X_1, X_2, \ldots, X_k$ appear on the RHS of the above equation, then it is essential to show that the invariants $X_1, X_2, \ldots, X_k$ themselves can be reduced to $X_1, X_2, \ldots, X_k$. The reconstructibility of $Z$ and $X_1, X_2, \ldots, X_k$ from the $\mathcal{N}$-matrix is then proved by induction on $v(G)$. That is possible because of the requirement that the sets $U,V,\ldots $ are proper subsets of $VG$. It is worth noting that the chromatic polynomial computation given in Biggs [@biggs1993] essentially follows a similar style. In several lemmas that precede the main theorem, we will prove identities of the form $Z(G) = \Theta(X(G_U), Y(G_V), ...)$. It will become clear that in the end all invariants computed here will reduce to the cycle structure of proper subgraphs. \[lem-kelly-cycles\] For $i < v(G)$, the number of cycles of length $i$ in $G$ has a reduction on $\mathcal{N}(G)$ given by $$\psi_i(G) = \frac{1}{v(G)-i}\sum_{u\in VG}\psi_i(G-u) = \frac{1}{v(G)-i}\sum_{j\mid v(\Lambda_j)=v(G)-1}N_{Ij}\psi_i(\Lambda_j)$$ This immediately follows from Kelly’s Lemma \[lem-kelly\] and Lemma \[lem-kelly3\]. \[def-cycle-cover\] Let $X$ be a subset of $VG$. Let $A\equiv (a_i)_{i=1}^k$ be a sequence in $[2,\|X\|]$. A $k$-tuple of cycles in $G_X$, corresponding to the sequence $A$, is a $k$-tuple of cycles in $G_X$ such that the $k$ cycles have lengths $a_1, a_2, \ldots, a_k$, respectively. Additionally, if the cycles in the $k$-tuple span the set $X$, then it is called a [spanning cycle cover]{} of $G_X$, corresponding to the sequence $A$. The number of $k$-tuples of cycles in $G_X$, corresponding to the sequence $A$, is denoted by $p(A\rightarrow G_X)$. The number of spanning cycle covers of $G_X$, corresponding to the sequence $A$, is denoted by $c(A\rightarrow G_X)$. \[lem-p2c\] $$\label{eq-p2c} p(A \rightarrow G_X) = \sum_{Y\subseteq X}c(A\rightarrow G_Y)$$ $$\label{eq-p2c-proof} \begin{split} p(A \rightarrow G_X) & = \prod_{j=1}^{k}\psi_{a_j}(G_X)\\ & = \|\{(F_1,F_2,\ldots,F_{k})\mid (\forall j\in[1,k])(F_j\subseteq G_X,\, F_j\cong C_{a_j})\}\| \\ & = \sum_{Y\subseteq X}c(A\rightarrow G_Y) \end{split}$$ Thus we have grouped together the $k$-tuples of cycles in groups that span each subset of $X$. This is essentially the idea of Kocay’s Lemma \[lem-kocay\]. \[lem-exp-p\] If $A \equiv (a_i)_{i = 1}^k $ is a sequence in $[2,v(G)-1]$ then $p(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$ given by $$p(A \rightarrow G) =\prod_{i=1}^{k}\psi_{a_i}(G) = \prod_{i=1}^{k} \left(\frac{1}{v(G)-a_i} \sum_{j\mid v(\Lambda_j)=v(G)-1}N_{Ij}\psi_{a_i}(\Lambda_j)\right)$$ This follows from the definition of $p(A\rightarrow G)$ and Lemma \[lem-kelly-cycles\]. \[lem-exp-c\] If $A \equiv (a_i)_{i = 1}^k $ is a sequence in $[2,v(G)-1]$ then $c(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$. By Möbius inversion of Equation (\[eq-p2c\]), we write $$c(A\rightarrow G_X) = \sum_{Y\subseteq X}(-1)^{\|X\backslash Y\|} p(A\rightarrow G_Y)$$ which implies $$c(A\rightarrow G) = \sum_{j=1}^I(-1)^{v(G)- v(\Lambda_j)}N_{Ij}p(A \rightarrow \Lambda_j)$$ By Lemma \[lem-exp-p\], the RHS of this equation has a reduction on $\mathcal{N}(G)$. Therefore, $c(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$. The following definition restricts the spanning cycle covers of Definition \[def-cycle-cover\] to connected spanning cycle covers. \[def-connected-spanning-cycle-cover\] Let $X$ be a subset of $VG$. Let $A\equiv (a_i)_{i=1}^k$ be a sequence in $[2,\|X\|]$. A [connected spanning cycle cover]{} of $G_X$, corresponding to the sequence $A$, is a $k$ tuple of cycles in $G_X$ such that the $k$ cycles have lengths $a_1, a_2, \ldots, a_k$, respectively, and together they constitute a connected subgraph spanning the set $X$. More formally, it is a $k$-tuple $(F_1,F_2,\ldots,F_{k})$ such that $(\forall j\in [1,k])(F_j\subseteq G_X, F_j\cong C_{a_j})$, $\cup_{j=1}^k VF_j = X$, and $\cup_{j=1}^kF_j$ is connected. The number of connected spanning cycle covers of $G_X$, corresponding to the sequence $A$, is denoted by $con(A\rightarrow G_X)$. The [disconnected spanning cycle covers]{} are defined similarly, and their number, corresponding to a sequence $A$, is denoted by $discon(A\rightarrow G_X)$. Let $A\equiv (A_i)_{i=1}^l$ be a list of $l$ non-increasing sequences such that $A_i \equiv (a_{ij})_{j=1}^{k_i}$; $a_{ij}\in [2,v(G)]$. Let $B \equiv (b_i)_{i=1}^l$ be a sequence in $[2,v(G)]$. Let $m \leq n$. We now define quantities $Q_m(A,B\rightarrow G)$ and $T_p(A,B\rightarrow G)$ that are based on connected spanning cycle covers as follows. $$\begin{split} Q_m(A,B\rightarrow G) & = \sum_{\substack {S\subseteq V(G)\\ \|S\|=m}} \prod_{i = 1}^{l} \left( \sum_{\substack{X\subseteq S \\ \|X\|=b_i}} con(A_i\rightarrow G_{X}) \right)\\ & = \sum_{\substack{S\subseteq V(G)\\\|S\|=m}} \left( \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \bigcup_{j=1}^{l}X_j\subseteq S\\ \|X_j\|=b_j\forall j}} \left( \prod_{i = 1}^{l}con(A_i\rightarrow G_{X_i}) \right) \right)\\ & = \sum_{p\leq m} T_p(A,B\rightarrow G)\displaystyle\binom{v(G)-p}{m-p} \end{split}$$ where $$\begin{split} T_p(A,B\rightarrow G) & = \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j\subseteq V(G)\\ \|\cup_{j=1}^{l}X_j\| = p \\ \|X_j\|=b_j\forall j} } \left( \prod_{i = 1}^{l}con(A_i\rightarrow G_{X_i}) \right) \end{split}$$ Solving the system of equations for $T_m(A,B\rightarrow G)$, we can write $$\label{eq-q2tp} \begin{split} T_m(A,B\rightarrow G) & = \sum_{p\leq m} (-1)^{m-p} \displaystyle\binom{v(G)-p}{m-p}Q_p(A,B\rightarrow G) \end{split}$$ When $m=v(G)$, this is simply $$\begin{split} T_{v(G)}(A,B\rightarrow G) & = \sum_{p\leq v(G)} (-1)^{v(G)-p} Q_p(A,B\rightarrow G) \end{split}$$ Note that if $m < max_{i,j}(a_{ij})$ for some $i,j$ then $T_m(A,B\rightarrow G)$ and $Q_m(A,B\rightarrow G)$ are both 0. \[lem-exp-q-tp\] If $A_i; i \in [1,l]$ are sequences in $[2,v(G)-1]$, and $B\equiv (b_i)_{i=1}^l$ are sequences in $[2,v(G)-1]$ then $Q_m(A,B\rightarrow G)$ and $T_m(A,B\rightarrow G)$ have reductions on the $\mathcal{N}$-matrix for each $m \leq v(G)$. We write $Q_m(A,B\rightarrow G)$ in terms of $\Lambda_j$ and the entries of $\mathcal{N}(G)$. $$\label{eq-exp-q} \begin{split} Q_m(A,B\rightarrow G) & = \sum_{\substack {S\subseteq V(G)\\ \|S\|=m}} \prod_{i = 1}^{l} \left( \sum_{\substack{X\subseteq S \\ \|X\|=b_i}} con(A_i\rightarrow G_{X}) \right)\\ &= \sum_{p\mid v(\Lambda_p)=m}N_{Ip} \prod_{i = 1}^{l} \left( \sum_{j\mid v(\Lambda_j)=b_i}N_{pj}con(A_i\rightarrow \Lambda_j) \right) \end{split}$$ Since $b_i<v(G)$, Equations (\[eq-exp-q\]) reduce $Q_m(A,B\rightarrow G)$ to the invariants $con(A_i\rightarrow \Lambda_j)$ for $m\leq v(G)$. Therefore, by Equation (\[eq-q2tp\]), $T_m(A,B\rightarrow G)$ are also reduced to the invariants $con(A_i\rightarrow \Lambda_j)$ for each $m\leq v(G)$. Note that if $a_{ik} > v(\Lambda_j)$ for some $k$ then $con(A_i\rightarrow \Lambda_j)=0$. \[lem-con\] If $A\equiv (a_i)_{i=1}^k$ is a sequence in $[2,v(G)-1]$ then $con(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$. The idea of the proof is similar to Kocay’s Lemma. First $c(A\rightarrow G)$ is written as $con(A\rightarrow G) + discon(A\rightarrow G)$. Then $discon(A\rightarrow G)$ is expressed in terms of $con(A_i \rightarrow G_{X_i})$ where $X_i$ are proper subsets of $VG$, and $A_i$ are certain subsequences of $A$. Then $discon(A\rightarrow G)$ are related to $T_{v(G)}(A^P, B\rightarrow G)$ for certain subsequences $A^P$ of $A$ and certain sequences $B$ constructed from appropriate partitions of $VG$. Since the reductions of $c(A\rightarrow G)$ and $T_{v(G)}(A^P, B\rightarrow G)$ have already been obtained, we get a reduction of $con(A\rightarrow G)$. Let $\mathcal{P}_q(N_k)$ be the set of all partitions of $N_k$ in $q$ parts. A partition $P$ in $\mathcal{P}_q(N_k)$, is denoted by $P = \{N_k^1,N_k^2,\ldots ,N_k^q\}$. Consider an arbitrary $k$ tuple $(F_1,F_2,\ldots,F_{k})$ such that $(\forall i\in [1,k])(F_i\subseteq G,F_i\cong C_{a_i})$. It defines a partition $P\in \mathcal{P}_q(N_k)$ so that $h$ and $j$ are in the same part of $P$ if and only if $F_h$ and $F_j$ are subgraphs of the same connected component of $\cup_{i=1}^kF_i$. We denote the contribution to $c(A\rightarrow G)$ from such tuples by $c_P(A\rightarrow G)$, and write $$c(A\rightarrow G) = \sum_{q}\sum_{P\in \mathcal{P}_q}c_P(A\rightarrow G)$$ Let the set of solutions to the equation $\sum_{i=1}^qb_i = v(G)$ be $\mathcal{B}(q)$. We can then write $$c(A\rightarrow G) = \sum_{q} \sum_{P\in \mathcal{P}_q} \sum_{B\in \mathcal{B}(q)} c_{PB}(A\rightarrow G)$$ where one more suffix $B$ in the summand is used to denote those tuples for which the connected component corresponding to part $N_k^i$ has $b_i$ vertices, for $i\in [1,q]$. Now expanding the summand in terms of $con(\ldots)$, we get $$\begin{split} c(A\rightarrow G) = \sum_{q} \sum_{P\in \mathcal{P}_q} \sum_{B\in \mathcal{B}(q)} \sum_{\substack{(X_1,X_2,\ldots,X_q)\mid \\ \cup_{j=1}^{q}X_j= V(G)\\ \|X_j\|=b_j\forall j} } \prod_{i=1}^q con(A_i\rightarrow G_{X_i}) \end{split}$$ where $A_i$ is the subsequence of $A$ with indexing set $N_k^i$. Innermost summation and product are now replaced by $T_{v(G)}(A^P,B\rightarrow G)$, so $$\label{eq-pqb} \begin{split} c(A\rightarrow G) = \sum_{q} \sum_{P\in \mathcal{P}_q} \sum_{B\in \mathcal{B}(q)} T_{v(G)}(A^P,B\rightarrow G) \end{split}$$ where $A^P$ is the collection of subsequences $A_i$ of $A$; $i\in [1,q]$, corresponding to the partition $P$. By Lemma \[lem-exp-c\], the LHS of Equation (\[eq-pqb\]) has a reduction on $\mathcal{N}(G)$. By Lemma \[lem-exp-q-tp\], each term on the RHS, except the term $con(A\rightarrow G)$, which corresponds to $q = 1$, has a reduction on $\mathcal{N}(G)$. This proves that $con(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$. \[cor-trees\] Let $tr(G)$ be the number of spanning trees in $G$, and let $uni(G,r)$ be the number of spanning unicyclic subgraphs of $G$ containing an $r$-cycle, respectively. Then, $tr(G)$ and $uni(G,r); r\in [3,v(G)]$ have reductions on $\mathcal{N}(G)$. Define $A\equiv (a_i)_{i=1}^{v(G)-1}$ such that $a_i=2 \,\forall \, i\in [1,v(G)-1]$. By Lemma \[lem-con\], $con(A\rightarrow G)$ has a reduction on $\mathcal{N}(G)$. But $con(A\rightarrow G) = (v(G)-1)!tr(G)$. Therefore, $tr(G)$ has a reduction $\mathcal{N}(G)$. To reduce $uni(G,r);r\in [3,v(G)-1]$, define $A\equiv(a_j)_{j=1}^{v(G)-r+1}$ such that $a_1 = r$, and $a_j = 2 \,\forall \, j\in [2,v(G)-r+1]$. Again, $con(A\rightarrow G) = (v(G)-r)!uni(G,r)$ has a reduction $\mathcal{N}(G)$ by Lemma \[lem-con\]. Thus $uni(G,r)$ also has a reduction on $\mathcal{N}(G)$. To reduce the number of hamiltonian cycles, let $A\equiv(a_i)_{i=1}^{v(G)}; a_i = 2\,\forall \,i\in [1,v(G)]$. We have, $con(A\rightarrow G) = (v(G)-1)!S(v(G),v(G)-1)tr(G) + \sum_{i=3}^{v(G)-1}v(G)!uni(G,i) + v(G)!ham(G)$, where $S(v(G),v(G)-1)$ is the Sterling number of the second kind computed for $(v(G),v(G)-1)$. Therefore, $ham(G)$, has a reduction $\mathcal{N}(G)$. \[lem-ci\] The coefficients $c_i(G)$ of the characteristic polynomial of $G$ are given by $c_0(G) = 1$ and $$c_i(G) = \frac{1}{v(G)-i}\sum_{j\mid v(\Lambda_j)=v(G)-1}N_{Ij}c_i(\Lambda_j) \,\,\text{for}\,\, 0 < i < v(G).$$ By Lemma \[lem-derivative\] we write $P'(G;\lambda) = \sum_{u\in VG}P(G-u;\lambda)$. Equating identical powers of $\lambda $ on the two sides, we get the result. \[cor-elementary\] If $F$ is an elementary graph on $v(G)$ vertices then ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ has a reduction on $\mathcal{N}(G)$. The case when $F$ is a hamiltonian cycle is handled in Corollary \[cor-trees\]. If $F$ is not a cycle, define a sequence $A\equiv (a_i)_{i=1}^k,\, a_i \in [2,v(G)-1]$ such that $\sum_{i=1}^k a_i = v(G)$. It is uniquely associated with an elementary graph $F\in L_{v(G)}$, so that the components of $F$ are cycles of length $a_i$, or $K_2$ if $a_i=2$. Now $c(A\rightarrow G) = c(A\rightarrow F){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, and $c(A\rightarrow F)$ depends only on the multiplicity of each cycle length in $F$. By Lemma \[lem-exp-c\], $c(A\rightarrow G)$ has a reduction on the $\mathcal{N}$-matrix. Therefore, ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ has a reduction on the $\mathcal{N}$-matrix. The following chart shows how various invariants were reduced to other invariants. For example, it shows that $con(A\rightarrow G); \, 2 \leq a_i \leq v(G)$ can be reduced to computing $ham(G)$, $con(A\rightarrow G); \, 2 \leq a_i < v(G)$ and $p(A\rightarrow G); \, 2 \leq a_i < v(G)$. It is clear from the diagram that computing $con(A\rightarrow G); \, 2 \leq a_i \leq v(G)$ and $ham(G)$ reduces to computing the same invariants for induced proper subgraphs. \[fig-dependency\] This makes the reconstructibility of several invariants obvious. \[thm-everything\] Suppose that $G$ is a simple finite graph, and we are given $\mathcal{N}(G)$. 1. Let $A\equiv (a_i)_{i=1}^k$ be a sequence such that $a_i \in [2,v(G)]$. Then $con(A\rightarrow G)$ and $ham(G)$ are reconstructible from $\mathcal{N}(G)$. 2. If $A_i; i \in [1,l]$ are sequences in $[2,v(G)]$, and $B\equiv (b_i)_{i=1}^l$ is a sequence in $[2,v(G)]$ then $Q_m(A,B\rightarrow G)$ and $T_m(A,B\rightarrow G)$ are reconstructible from $\mathcal{N}(G)$ for each $m \leq v(G)$. 3. the number of spanning trees in $G$, the number of cycles of length $i$, for $3 \leq i \leq v(G)$, the number of unicyclic subgraphs containing a cycle of length $i$, for each $i \in [3,v(G)]$, and the characteristic polynomial $P(G;\lambda)$ are all reconstructible from $\mathcal{N}(G)$. We prove the first item by induction on $v(G)$. The base case is $v(G) = 2$. In this case $G = K_2$. Now suppose that $con(A\rightarrow G)$ and $ham(G)$ are reconstructible from $\mathcal{N}(G)$ when $v(G) < s$ for an arbitrary sequence $A\equiv (a_i)_{i=1}^k$ of integers in $[2,v(G)]$. Now let $v(G) = s$. Lemmas and Corollaries \[lem-kelly-cycles\] to \[cor-trees\] imply that computations of $con(A\rightarrow G)$ and $ham(G)$ reduce to computations on induced proper subgraphs of $G$, thus completing the induction step, and the proof of the first item. Since all other intermediate invariants $T_m(\ldots)$, $Q_m(\ldots)$, the number of spanning trees, the number of unicyclic graphs having a cycle of a specified length, number of cycles of each length, number of elementary spanning graphs of each type, etc. have been reduced to computations of invariants $con(A\rightarrow \Lambda_j); j < I$ and $\psi_i(\Lambda_j); i < v(G), j < I$, the remaining parts of the theorem follow. There is another way of proving the $\mathcal{N}$-matrix reconstructibility of the characteristic polynomial. From $\mathcal{N}(G)$, it is possible to construct $\mathcal{N}(\bar{G})$, and then invoke the result of Hagos [@hagos2000] in the induction step. Hagos proved that the pair $(P(G;\lambda),P(\bar{G};\lambda))$ can be reconstructed from the collection $\{(P(G-u;\lambda),P(\bar{G}-u;\lambda)); u\in VG\}$. We skip the details of this argument. The proof presented here counts many other invariants. It is likely that the deck of pairs of polynomials considered by Hagos contains enough information for counting hamiltonian cycles and spanning trees. Now we count the subgraphs with a given number of components, and a given number of edges in each component, and use it to compute the rank polynomial. Let $\mathcal{G}(p,l,(p_i,q_i)_{i=1}^l)$ be the family of graphs with $p$ vertices and $l$ components, such that the $i$’th component has $p_i$ vertices and $q_i$ edges for $i \in [1,l]$. So, $\sum_i p_i = p$. We also assume that $p_i \geq p_j$ whenever $i < j$. By extending the notation ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ defined earlier, we denote by ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ the number of subgraphs of $G$ that belong to the family $\mathcal{G}(p,l,(p_i,q_i)_{i=1}^l)$. \[lem-kedge\] The number of connected spanning subgraphs of $G$ with $k$ edges, that is,\ ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, is reconstructible from $\mathcal{N}(G)$ for all $k$. When $k < v(G)-1$, ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ is 0. When $k\in [v(G)-1, e(G)]$, we prove the result by induction on $k$. The base case $k = v(G)-1$, which corresponds to the number of spanning trees, was proved in Theorem \[thm-everything\]. Let the claim be true for all $k\in [v(G)-1,q-1]$. To prove the claim for $k=q$, define $A\equiv (a_i)_{i=1}^q$ such that $a_i=2$ for all $i\leq q$. We can write $$\label{eq-kedge-con} con(A\rightarrow G) = \sum_{i=n-1}^q i!S(q,i) {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$$ In Theorem \[thm-everything\], $con(A\rightarrow G)$ was shown to be reconstructible from $\mathcal{N}(G)$. By the induction hypothesis, all terms on the RHS, except ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, are known. Solving Equation (\[eq-kedge-con\]) for ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ completes the induction step and the proof. \[lem-lcompo\] ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, where $\sum_{i=1}^ln_i = v(G)$ and $m_i > 0$ for all $i$, is reconstructible from $\mathcal{N}(G)$. Let $A\equiv (A_i)_{i=1}^l$, where $A_i\equiv (a_{ij})_{j=1}^{m_i}$; $a_{ij} = 2\,\forall i,j$, and $B\equiv (n_i)_{i=1}^l$. By Theorem \[thm-everything\], $T_{v(G)}(A,B\rightarrow G)$ is $\mathcal{N}$-matrix reconstructible. We first express $T_{v(G)}(A,B\rightarrow G)$ in terms of the subgraphs to be counted, and then count the subgraphs by induction. $$\begin{split} T_{v(G)}(A,B\rightarrow G) & = \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j= V(G)\\ \|X_j\|=n_j\forall j} } \prod_{i = 1}^{l}con(A_i\rightarrow G_{X_i}) \\ & = \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j= V(G)\\ \|X_j\|=n_j\forall j} } \prod_{i = 1}^{l}\left( \sum_{q_i = n_i-1}^{m_i} q_i!S(m_i,q_i) {\mbox{$\left[\begin{array}{c}G_{X_i}\\#2\end{array}\right]$}} \right)\\ & = \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j= V(G)\\ \|X_j\|=n_j\forall j} } \sum_{\substack{(q_1,q_2,\ldots,q_l)\mid \\ n_j-1\leq q_j \leq m_j \forall j} } \prod_{i = 1}^{l}\left( q_i!S(m_i,q_i){\mbox{$\left[\begin{array}{c}G_{X_i}\\#2\end{array}\right]$}} \right)\\ & = \sum_{\substack{(q_1,q_2,\ldots,q_l)\mid \\ n_j-1\leq q_j \leq m_j \forall j} } \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j= V(G)\\ \|X_j\|=n_j\forall j} } \prod_{i = 1}^{l} \left( q_i!S(m_i,q_i){\mbox{$\left[\begin{array}{c}G_{X_i}\\#2\end{array}\right]$}} \right)\\ & = \sum_{\substack{(q_1,q_2,\ldots,q_l)\mid \\ n_j-1\leq q_j \leq m_j \forall j} } \left( \prod_{i = 1}^{l} q_i!S(m_i,q_i) \right) \left( \sum_{\substack{(X_1,X_2,\ldots,X_l)\mid \\ \cup_{j=1}^{l}X_j= V(G)\\ \|X_j\|=n_j\forall j} } \prod_{i = 1}^{l} {\mbox{$\left[\begin{array}{c}G_{X_i}\\#2\end{array}\right]$}} \right) \\ \end{split}$$ The sequence $(n_i,q_i)_{i=1}^l$ may be written as $(n_i^\prime,q_i^\prime)^{\mu_i}$; $i = 1$ to $r$, which denotes that the pair $(n_i^\prime,q_i^\prime)$ appears $\mu_i$ times in the sequence $(n_i,q_i)_{i=1}^l$, the pairs $(n_i^\prime,q_i^\prime)$ are all distinct for $i = 1$ to $r$, and that $\sum_{i=1}^r \mu_i = l$. Then each subgraph of $G$ that belongs to the family $\mathcal{G}(v(G),l,(n_i,q_i)_{i=1}^l)$ is counted $\prod_{i=1}^r \mu_i!$ times in the inner summation. Therefore, $$\label{eq-dics-subgraphs} \begin{split} T_{v(G)}(A,B\rightarrow G) & = \sum_{\substack{(q_1,q_2,\ldots,q_l)\mid \\ n_j-1\leq q_j \leq m_j \forall j} } \left(\prod_{i = 1}^{l}q_i!S(m_i,q_i)\right) \left(\prod_{i=1}^r \mu_i!\right) {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \end{split}$$ The LHS of Equation (\[eq-dics-subgraphs\]) is known by Theorem \[thm-everything\]. Now we prove the claim by induction on $\sum_i m_i$. The base case of induction corresponds to the case in which each component in the subgraphs being counted has minimum number of edges, that is, $m_i = n_i -1$ for all $i\leq l$. In this case, there is only one term on the RHS of Equation (\[eq-dics-subgraphs\]), and it contains the unknown ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$, which can be solved for. Suppose the claim is true for $\sum_i m_i < m$. Now let $\sum_i m_i = m$. In this case, as in Lemma \[lem-kedge\], there is only one unknown ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ on the RHS of Equation (\[eq-dics-subgraphs\]). All other terms on the RHS are known by the induction hypothesis. We can compute the unknown term to obtain the desired result. This completes the induction step and the proof. \[thm-rank\] The rank polynomial $R(G;x,y)$ is reconstructible from $\mathcal{N}(G)$. Lemma \[lem-lcompo\] can be applied to all induced subgraphs of $G$. So, we can count the number of subgraphs with $v$ vertices (none of which isolated), $e$ edges and $l$ components for all $v\leq v(G)$, $e\leq e(G)$ and $l\geq 1$. Therefore, $\rho_{rs}$ in the expression for the rank polynomial are known. Computing $P(G;\lambda)$ from $\mathcal{PD}(G)$ {#sec-poly} =============================================== In this section, we consider the problem of computing the characteristic polynomial of a graph from its complete polynomial deck. We prove that elementary spanning subgraphs of each type other than hamiltonian cycles can be counted from the complete polynomial deck of a graph, thus proving that the characteristic polynomial of a non-hamiltonian graph is reconstructible from its complete polynomial deck. Here we apply the idea of Kocay’s Lemma to the coefficients of the characteristic polynomials. Let $A \equiv (a_i)_{i = 1}^k$ be a non-increasing sequence in $[2,v(G)]$. In this section, we define the notation $p(A \rightarrow G_X)$ and $c(A\rightarrow G_X)$ differently. $$\label{eq-pc} \begin{split} p(A \rightarrow G_X) & = \prod_{j=1}^{k}(-1)^{a_j}c_{a_j}(G_X)\\ & = \prod_{j = 1}^{k}\left( \sum_{F\subseteq G_X,\,F\in L_{a_j}}(-1)^{r(F)}2^{s(F)} \right)\\ & = \sum_{\substack{(F_1,F_2,\ldots,F_{k})\mid \\ (\forall j\in [1,k])(F_j\subseteq G_X,\,F_j\in L_{a_j})} } \left((-1)^{\sum_{j=1}^{k} r(F_j)} 2^{\sum_{j=1}^{k} s(F_j)}\right)\\ & = \sum_{Y\subseteq X}c(A\rightarrow G_Y) \end{split}$$ where $$\begin{split} c(A\rightarrow G_Y) & = \sum_{\substack{(F_1,F_2,\ldots,F_{k})\mid \\ (\forall j\in [1,k])(F_j\subseteq G_Y,\,F_j\in L_{a_j})\\ \bigcup_{j=1}^{k}(VF_j)\, =\, Y} } \left((-1)^{\sum_{j=1}^{k} r(F_j)} 2^{\sum_{j=1}^{k} s(F_j)}\right) \end{split}$$ Thus we have grouped together tuples of elementary subgraphs in groups that span each subset of $X$. \[lem-pd2cdec\] If $A$ is a sequence defined over $[2,v(G)-1]$ then $c(A\rightarrow G)$ is reconstructible from the complete polynomial deck of $G$. As in the proof of Lemma \[lem-exp-c\], $p(A \rightarrow G_X)$ can be computed for each induced subgraph $G_X$. By Möbius inversion of Equation (\[eq-pc\]), we write $$\label{eq-pd-mobius} c(A\rightarrow G_X) = \sum_{Y\subseteq X}(-1)^{\|X\backslash Y\|} p(A\rightarrow G_Y)$$ But we cannot compute the RHS of Equation (\[eq-pd-mobius\]) because, in general, for $X\subsetneq VG$, we do not know which polynomials in $\mathcal{PD}(G)$ correspond to the induced subgraphs of $G_X$. But this is not a problem if $X = VG$. We can write $$c(A\rightarrow G) = \sum_{Y\subseteq VG}(-1)^{\|VG\backslash Y\|} p(A\rightarrow G_Y)$$ Now the RHS, and hence $c(A\rightarrow G)$, can be computed. [Remark.]{} Note that we would not be able to compute $p(A\rightarrow G)$ if we defined $A$ in $[2,v(G)]$, because we do not know $c_{v(G)}(G)$. Let $\lambda (m,p) \equiv (x_1, x_2, \ldots , x_p)$ denote a partition of $m$. We assume that $x_1\, \geq \, x_2\, \ldots \, \geq x_p$. We write $\lambda (m) $ when the number of parts $p$ is not relevant. Also, we just write $\lambda $ instead of $\lambda (m,p)$ when $m$ and $p$ are either understood from the context or not relevant. Another partition $\lambda ^\prime (m,p+1) $ can be obtained from $\lambda (m,p)$ by replacing an $x_i$ by $y$ and $z$ such that $y+z = x_i$, and ordering the numbers in a non-increasing order. Any partition that is obtained from $\lambda (m,p)$ by a sequence of such operations is called a [refinement]{} of $\lambda (m,p)$. Also, $\lambda (m,p)$ is a trivial refinement of itself. If $\lambda ^\prime (m,q)$ is a refinement of $\lambda (m,p)$, then we denote it by $\lambda ^\prime (m,q) \preceq \lambda (m,p)$. The relation $\preceq $ between partitions is a partial order. Now consider partitions in which the smallest part $x_p$ is at least 2. Associated with each such partition $\lambda (m,p)$, there is a unique elementary graph $F_{\lambda }\in L_m$, whose $i$’th component is a cycle of length $x_i$, or an edge if $x_i\, =\, 2$. If the sequence $A\equiv (a_i)_{i=1}^k$ is defined such that $\lambda \equiv (a_1,a_2, \ldots, a_k)$ is a non-trivial partition of $v(G)$, then we denote $c(A\rightarrow G)$ by $c(\lambda \rightarrow G)$. \[lem-refinement\] Let $\lambda \equiv (a_1,a_2, \ldots, a_k)$ be a non-trivial partition of $n = v(G)$. Then, $$\label{eq:refinement} \begin{split} c(\lambda \rightarrow G) & = \sum_{\lambda ^\prime \preceq \lambda} c(\lambda \rightarrow F_{\lambda ^\prime}) {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \end{split}$$ From the definition of $c(A\rightarrow G)$, we write $$\begin{split} c(\lambda\rightarrow G) & = \sum_{\substack{(F_1,F_2,\ldots,F_{k})\mid \\ (\forall j\in [1,k])(F_j\subseteq G,\,F_j\in L_{a_j})\\ \bigcup_{j=1}^{k}(VF_j)\, =\, VG} } \left((-1)^{\sum_{j=1}^{k} r(F_j)} 2^{\sum_{j=1}^{k} s(F_j)}\right)\\ & = \sum_{F\subseteq G, F\in L_n} \sum_{\substack{(F_1,F_2,\ldots,F_{k})\mid \\ (\forall j\in [1,k])(F_j\subseteq F,\,F_j\in L_{a_j})\\ \bigcup_{j=1}^{k}F_j\, =\, F} } \left((-1)^{\sum_{j=1}^{k} r(F_j)} 2^{\sum_{j=1}^{k} s(F_j)}\right)\\ & = \sum_{\substack{F\subseteq G,F\in L_n}} c(\lambda \rightarrow F)\\ & = \sum_{\lambda ^\prime \preceq \lambda} c(\lambda \rightarrow F_{\lambda ^\prime}) {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \end{split}$$ The last step may be explained as follows: if $F$ is a disjoint union of elementary graphs $F_j\in L_{a_j}; j \in [1,k]$, where $v(F) = v(G) = n$, then $F$ is isomorphic to an elementary graph $F_{\lambda ^\prime}$ for some refinement ${\lambda ^\prime}$ of $\lambda$. Trivially, if each $F_j$ is the cycle $C_{a_j}$ then $F = F_{\lambda}$. We then group the terms $c(\lambda \rightarrow F)$ by the isomorphism type of $F$. \[lem-non-hamiltonian\] If $F$ is an elementary graph on $n = v(G)$ vertices, other than the cycle, then ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ is reconstructible from $\mathcal{PD}(G)$. Since $F$ is not a cycle, it is isomorphic to $F_{\lambda_0}$ for a unique non-trivial partition $\lambda_0 $ of $v(G)$. From Equation (\[eq:refinement\]) we write $$\label{eq:nonh1} \begin{split} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} &= \frac{1}{c(\lambda_0 \rightarrow F_{\lambda_0 })} \left( c(\lambda_0 \rightarrow G) -\sum_{\lambda \prec \lambda_0} c(\lambda_0 \rightarrow F_{\lambda }) {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \right) \end{split}$$ Now we expand $ {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} $ on the RHS of the above equation by repeated application of the same equation, and obtain the following solution. $$\label{eq:nonh2} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \sum_{\lambda_q \prec \lambda_{q-1} \prec \ldots \prec \lambda_0} \frac{\left(-1\right)^q c(\lambda_q \rightarrow G) \prod_{i=0}^{q-1}c(\lambda_i\rightarrow F_{\lambda_{i+1}})} {\prod_{i=0}^qc(\lambda_i\rightarrow F_{\lambda_i})}$$ where the summation is over all chains $\lambda_q \prec \lambda_{q-1} \prec \ldots \prec \lambda_0$; $q \geq 0$, and an empty product is 1. There are finitely many terms in the above summation since there are finitely many refinements of $\lambda_0$. Since $\lambda_0 $ is a non-trivial partition of $n$, (that is, $x_i < v(G)\,\forall\, i$), by Lemma \[lem-pd2cdec\], $c(\lambda \rightarrow G)$ is reconstructible for each $\lambda \preceq \lambda_0$. Also, for each $\lambda \preceq \lambda_0$, $c(\lambda \rightarrow F_{\lambda })$ is non-zero. (Here we would like to repeat that we have considered only those partitions in which the smallest part is at least 2.) Thus the RHS can be computed. The main theorem in this section now follows from the above lemmas. \[thm-deg1\] If $F$ is an elementary graph on $v(G)$ vertices, other than a cycle, then ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ is reconstructible from $\mathcal{PD}(G)$. Therefore, if there is a vertex of degree 1 in $G$, then the characteristic polynomial of $G$ is reconstructible from its complete polynomial deck. The degree sequence of a graph is reconstructed from its complete polynomial deck as follows. Consider the polynomials of degree $v(G)-1$ in $\mathcal{PD}(G)$. They are the characteristic polynomials of the vertex deleted subgraphs $G-u$ of $G$ for $u\in VG$. Since $c_2(G)$ and $c_2(G-u)$ count the number of edges of $G$ and $G-u$, respectively, we know the degree of $u$ in $G$ for each $u \in VG$. Thus the premise of Theorem \[thm-deg1\] is recognised from $\mathcal{PD}(G)$. The coefficients $c_i(G);i<v(G)$ can be computed using Lemma \[lem-derivative\]. Since there are no hamiltonian cycles in $G$, Lemma \[lem-non-hamiltonian\] implies that the constant term in the characteristic polynomial of $G$ can be calculated. [Remark.]{} Whenever non-hamiltonicity of a graph is recognised from its complete polynomial deck, its characteristic polynomial can be computed as well. Whitney’s Theorem {#sec-whitney} ================= In Section \[sec-intro\], it was stated that the computation of the chromatic polynomial of a graph requires only non-separable induced subgraphs of the graph. Whitney’s proof of this fact was based on his theorem that separable subgraphs can be counted from the counts of non-separable subgraphs. Let $n_t(G)$ denote the number of subgraphs of $G$ of type $t$, where ‘type’ of a graph is determined by the number of blocks of each isomorphism type. He proved, (stated in the terminology of [@biggs1993]), that there is a polynomial $\phi_t$, independent of $G$, such that $$n_t(G) = \phi_t(n_{\sigma}(G), n_{\rho}(G), ...)$$ where $\sigma$, $\rho$, ... are non-separable types with not more edges than $t$. Here we prove result using Kocay’s Lemma. Our presentation explicitly describes the polynomial in Whitney’s theorem. Let $S_0 = \{F_1, F_2, \ldots ,F_k\}$ be a family of non-separable graphs, some of them possibly isomorphic. Thus $S_0$ represents a ‘graph type’. Extending the notation introduced earlier, we write ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ to denote the number of subgraphs of $G$ of type $S_0$. We define a partial order $\preceq $ on graph types as follows. Let $S$ be a graph type. We say that $S\preceq S_0$ if $c(S_0,X)$ is non-zero for some graph $X$ of type $S$. It is easily seen that $c(S_0,X)$ depends only on the type $S$ of $X$, not on a particular choice of $X$. So, we write it as $c(S_0,S)$. For any graph $G$, by Kocay’s Lemma \[lem-kocay\], $$p(S_0) = \prod_{i = 1}^{k} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \sum_{X} c(S_0,X){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$$ Terms on the RHS can be grouped together according to the types of $X$, so we can write $$\begin{split} p(S_0) &= \sum_{S\preceq S_0} c(S_0,S){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \\ &= c(S_0,S_0){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}+\sum_{S\prec S_0} c(S_0,S){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \end{split}$$ Therefore, $${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \frac{1}{c(S_0,S_0)} \left( p(S_0)-\sum_{S\prec S_0}c(S_0,S){\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} \right)$$ Now we repeatedly apply the same equation to ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ on the RHS, as we did in Equation (\[eq:nonh2\]). We thus get the polynomial of Whitney’s theorem. \[thm-whitney1\] $$\label{eq-whitney} {\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = \sum_{S_q \prec S_{q-1} \prec \ldots \prec S_0} \frac{\left(-1\right)^qp(S_q)\prod_{i=0}^{q-1}c(S_i,S_{i+1})} {\prod_{i=0}^qc(S_i,S_i)}$$ where the summation is over all chains $S_q \prec S_{q-1} \prec \ldots \prec S_0$; $q \geq 0$, and an empty product is 1. There are finitely many terms in the summation in Equation (\[eq-whitney\]) because each $S_q$ has fewer blocks in it than $S_{q-1}$. While some other known proofs of this theorem are based on not very different ideas, (for example, see [@biggs1978]), the above explicit formulation of the polynomial seems new. It allows us to argue about the reconstructibility of the characteristic polynomial more directly than in other standard proofs. \[cor-w2p\] The characteristic polynomial of a graph is reconstructible from its vertex deck. Let $G$ be the graph under consideration. Its elementary spanning subgraphs other than the hamiltonian cycles are counted as in the standard proof. Let $n = v(G)$, and let $S_0$ be the type of an $n$-vertex elementary graph $H$ other than $C_n$. Any block in a type $S \prec S_0$ has fewer than $n$ vertices, and $H$ is the only graph of type $S_0$ that has $n$ vertices. So ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}}$ can be counted using Whitney’s Theorem \[thm-whitney1\] and Kelly’s Lemma \[lem-kelly\]. To count hamiltonian cycles, we set $S_0 = \{nK_2\}$, that is, a graph type consisting of $n$ blocks, each one of them a $K_2$. Since no subgraph of $G$ has $n$ blocks isomorphic to $K_2$, ${\mbox{$\left[\begin{array}{c}G\\#2\end{array}\right]$}} = 0$. On the RHS of Equation (\[eq-whitney\]), there is precisely one term that contributes hamiltonian cycles. That is, $S_1 = \{C_n\} \prec S_0$ is the unique chain that contributes $ham(G)$, implying that the terms containing $ham(G)$ cannot cancel out. All other blocks that appear in Equation (\[eq-whitney\]) have fewer than $n$ vertices, so can be counted using Kelly’s Lemma \[lem-kelly\]. Therefore, $ham(G)$ is reconstructible. The reconstructibility of $P(G;\lambda)$ now follows from Lemma \[lem-sachs\]. [Remark.]{} In the standard proof of the reconstructibility of the characteristic polynomial, one applies Kocay’s Lemma directly. As a result one has to proceed step by step, counting spanning trees, then spanning unicyclic subgraphs, etc., as we did in Corollary \[cor-trees\]. These intermediate steps are skipped by the direct application of Whitney’s theorem. Problems and discussion ======================= Expressing $c_n(G)$ or $\psi_n(G)$, where $n = v(G)$, as polynomials in $c_j(G-S)$ or $\psi_j(G-S)$; $S\subsetneq VG$, would be of interest. Alternatively, we would like to construct a generalisation of the characteristic polynomial which can be computed more naturally from the poset of induced subgraphs, and from which the characteristic polynomial can be easily computed. Such a goal is motivated on the one hand by the proofs in Section \[sec-nmatrix\], and, on the other hand, by similar generalisations of the chromatic polynomial, viz, Stanley’s chromatic symmetric function, (see [@stanley1995] & [@stanley1999]), and another recent two variable generalisation of the chromatic polynomial [@dohmen2003]. Both these generalisations are closely related to the lattice of connected partitions of $VG$, (see [@stanley1995] for definitions). A relationship between the poset of induced subgraphs defined in this paper and the lattice of connected partitions of the vertex set defined by Stanley could possibly be established using Kocay’s Lemma. Such a result would be useful in understanding exact relationship between different expansions (and reconstructibility) of several important invariants in a unified way. The reconstruction of the number of hamiltonian cycles is difficult and indirect in the proofs we have presented here, and in the original proof by Tutte as well. A reason for this difficulty is seen in Whitney’s theorem. Observe that for an $n$-vertex graph $G$, the polynomial of Whitney’s theorem contains a term in $ham(G)$ only if the type $S_0$ contains $n$ copies of $K_2$, or a $C_n$, and possibly other types of blocks. As a result, in Corollary \[cor-w2p\] we had to count all possible blocks with at most $n$ edges. But can we count the number of hamiltonian cycles from the $\mathcal{N}$-matrix at least as clearly as in Corollary \[cor-w2p\]? Towards this goal, we would like to understand the relationship between the structure of the edge labelled poset for separable graphs and that for blocks, and prove a generalisation of Whitney’s theorem. The crucial difference between the proofs in Section \[sec-nmatrix\] and Section \[sec-poly\] is in Lemmas \[lem-exp-c\] and \[lem-pd2cdec\]. In Lemma \[lem-pd2cdec\], the use of Möbius inversion was limited to the computation of $c(A\rightarrow G)$ because we did not know the partial order on the induced subgraphs. This suggests that a general ‘expansion’ for the number of hamiltonian cycles would probably involve a summation over chains in $\mathcal{ELP}(G)$. Therefore, counting hamiltonian cycles from $\mathcal{PD}(G)$ and the original problem of Gutman and Chvetković seem difficult. This is probably why many known results on the reconstruction of the characteristic polynomial of a graph from its characteristic polynomial deck assume the graph to contain several pendant vertices. We propose the following generalisation of reconstruction for studying questions similar to the one posed by Gutman and Chvetković. Suppose $f$ is a graph invariant, and we are interested in reconstructing $G$ or partial invariants of $G$ from the deck $\mathcal{D}(G;f) = \{f(G-u); u \in VG\}$. A new collection $\mathcal{D}!(G;f)$ is recursively defined as $\{(f(G-u), \mathcal{D}!(G;f-u)); u\in VG\}$. We then define an equivalence relation $\sim $ on graphs such that $H_1 \sim H_2$ if $(f(H_1), \mathcal{D}!(H_1)) = (f(H_2),\mathcal{D}!(H_2))$. This relation gives an incidence matrix (or an edge labelled poset) on the types of induced subgraphs of $G$, where ‘type’ refers to an equivalence class under the relation $\sim $ defined above. It can be shown that for many invariants $f$, Ulam’s conjecture is true if and only if all graphs $G$ on more than 2 vertices are reconstructible from $\mathcal{D}!(G;f)$. One example of such an invariant is: $f(G) = 1$ if $G$ has a vertex of degree 1, and $f(G) = 0$ otherwise. Another example is $f(G) = P(G;\lambda)$. The proof of this is similar to that of Proposition \[prop:equiv-un\]: the base case follows from the fact that any three vertex graph $G$ is completely determined by $\mathcal{D}!(G;f)$ for the above invariants. The problem of reconstructing $G$ from the deck $\mathcal{D}!(G;f)$ is similar to the generalisation of the reconstruction problem suggested by Tutte, (Notes on pp. 123-124 in [@tutte1984]). We are not really interested in the question of computing $f(G)$ from $\mathcal{D}(G;f)$. Rather we ask the question - what are the incomplete invariants $f$, (that is, the invariants that do not determine a graph completely,) and classes of graphs $G$, for which $\mathcal{D}!(G;f)$ could be constructed from $\mathcal{D}(G;f)$? If we could construct $\mathcal{D}!(G;f)$ from $\mathcal{D}(G;f)$, then we could also prove all the results of Section \[sec-nmatrix\]. We would like to investigate this question when $f(G) = P(G;\lambda)$, and when $f(G) = (P(G;\lambda), P(\bar{G};\lambda))$ - the invariant which was considered by Hagos [@hagos2000]. Acknowledgements {#acknowledgements .unnumbered} ---------------- I take this opportunity to thank Allan Wilson Centre for the support and encouragement. I would also like to thank the referee for several useful suggestions for improving the presentation. [10]{} N. Biggs. On cluster expansions in graph theory and physics. , 29:159–173, 1970. N. Biggs. . Cambridge University Press, 1993. A. Bondy. A graph reconstructor’s manual. , pages 221–252, 1991. K. Dohmen, A. Pönitz, and P. Tittmann. A new two variable generalisation of the chromatic polynomial. , 6:069–090, 2003. I. Gutman and D. M. Cvetković. The reconstruction problem for the characteristic polynomial of graphs. , pages 45–48, 1975. Elias M. Hagos. The characteristic polynomial of a graph is reconstructible from the characteristic polynomials of its vertex-deleted subgraphs and their complements. , 7:\#R12, 2000. P. J. Kelly. A congruence theorem for trees. , 7:961–968, 1957. W. L. Kocay. On reconstructing spanning subgraphs. , 11:301–313, 1981. A. J. Schwenk. On the eigen values of a graph. In L. W. Beineke and R. J. Wilson, editors, [Selected Topics in Graph Theory]{}, pages 307–336. Academic Press, New York, 1979. I. Sciriha. Polynomial reconstruction and terminal vertices. , 356:145–156, 2002. R. P. Stanley. A symmetric function generalisation of the chromatic polynomial of a graph. , 111:166–194, 1995. R. P. Stanley. , volume 1. Cambridge University Press, 1997. R. P. Stanley. , volume 2. Cambridge University Press, 1999. W. T. Tutte. All the king’s horses. In J.A.Bondy and U.S.R.Murthy, editors, [Graph Theory and Related Topics, (Proceedings of the conference held in honour of W. T. Tutte on the occasion of his 60th birthday, Waterloo 1977)]{}, pages 15–33. Academic Press, New York, 1979. W. T. Tutte. , volume 21. Addison-Wesley Publishing Co., Reading, Mass., 1984. S. Ulam. . Wiley (Interscience), New York, 1960. H. Whitney. A logical expansion in mathematics. , 38:572–579, 1932.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We numerically compute axisymmetric Taylor-Couette flow in the presence of axially periodic magnetic fields, with Hartmann numbers up to $Ha^2=10^7$. The geometry of the field singles out special field lines on which Shercliff layers form. These are simple shear layers for insulating boundaries, versus super-rotating or counter-rotating layers for conducting boundaries. Some field configurations have previously studied spherical analogs, but fundamentally new configurations also exist, having no spherical analogs. Finally, we explore the influence of azimuthal fields $B_\phi\sim r^{-1}{\bf\hat e}_\phi$ on these layers, and show that the flow is suppressed for conducting boundaries but enhanced for insulating boundaries.\ $\phantom{x}\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\;$ [Résumé]{}\ Nous modéliserons l’écoulement axisymétrique de Taylor-Couette en présence d’un champ magnétique axialement périodique, avec un nombre de Hartmann jusqu’à $Ha^2=10^7$. La géometrie du champ montre des lignes de champ sous forme de couche de Shercliff. Il y a des couches de cisaillement, lorsque les frontières sont isolantes, tandis que la rotation est excessive ou inversée pour les frontières conductrices. Certaines configurations de champs sont similaires à celles vues sous forme sphérique cependant de nouvelles configurations existent. Enfin, nous découvrirons l’influence de champs azimutaux ($B_\phi\sim r^{-1}{\bf\hat e}_\phi$) sur ces couches et nous montrerons que l’écoulement diminue avec des bords conducteurs alors qu’il s’accentue pour des frontières isolantes. author: - 'Rainer Hollerbach$^1$, Deborah Hulot' title: \| Shercliff layers in strongly magnetic cylindrical\ Taylor-Couette flow --- Introduction ============ Shercliff layers are free shear layers that can occur in the flow of an electrically conducting fluid when a sufficiently strong magnetic field is externally imposed [@Shercliff]. They arise due to the strongly anisotropic nature of the Lorentz force, consisting of a tension along the magnetic field lines. The details of how the spatial structure of the imposed field overlaps with the geometry of the container can then single out special field lines on which Shercliff layers form. For example, suppose we consider spherical Couette flow, the flow induced in a spherical shell where the inner sphere rotates and the outer one is fixed. Consider further two possible choices of magnetic fields to impose, a dipole ${\bf B}_d=2\sigma^{-3}\cos\theta\,{\bf\hat e}_\sigma +\sigma^{-3}\sin\theta\,{\bf\hat e}_\theta$ and a uniform axial field ${\bf B}_a={\bf\hat e}_z=\cos\theta\,{\bf\hat e}_\sigma-\sin\theta\, {\bf\hat e}_\theta$, where $(\sigma,\theta,\phi)$ are standard spherical coordinates, and $(z,r,\phi)$ cylindrical coordinates. For the dipole field, there will be some field lines that link only to the inner sphere, and others that connect the two spheres. Similarly, for the axial field there will be some field lines that link only to the outer sphere, and others that connect the two spheres. The tension in the field lines then ensures that any field lines linked to one boundary only are completely locked to that boundary, with the fluid either co-rotating with the inner sphere, or stationary together with the outer sphere. It is only on field lines that connect to both boundaries that the fluid is faced with conflicting conditions at the two ends of the line, and resolves this conflict by rotating at a rate intermediate between the two end values. The entire domain is therefore naturally divided up into different regions depending on how the field lines connect to the boundaries, with the angular velocity changing abruptly across those field lines separating different regions [@Dormy1; @Starchenko]. Furthermore, it is clear that there is nothing special about either the spherical geometry or these two particular fields. As long as both the container and the imposed field are axisymmetric, the same considerations will apply, and will always result in Shercliff layers forming on these special field lines where the linkage to the boundaries switches from one type to another. The thickness of these layers scales as $Ha^{-1/2}$, where the Hartmann number $Ha$ is a measure of the strength of the imposed field [@Roberts]. Another intriguing result is the influence of the electromagnetic boundary conditions. The conclusion above, that Shercliff layers are simply shear layers on which the angular velocity switches to something intermediate between 0 at the outer boundary and 1 at the inner boundary, is valid only if both boundaries are insulating. If instead the inner sphere is conducting, a dipole field yields a so-called super-rotation, where the fluid within the Shercliff layer rotates faster than the inner sphere [@Dormy1]. Alternatively, if the outer sphere is conducting, an axial field yields a counter-rotation, where the fluid within the Shercliff layer rotates in the opposite direction to the inner sphere [@Hollerbach00]. In both of these cases, the degree of super-rotation or counter-rotation is around 20-30% of the inner sphere’s rotation rate, independent of $Ha$ (for sufficiently large values). Even more unexpected results are obtained if both boundaries are taken to be conducting; in this case the degree of ‘anomalous’ rotation appears to increase indefinitely as $Ha$ is increased in a numerical computation [@Hollerbach00; @Hollerbach01]. Various asymptotic analyses of this problem confirm that the anomalous rotation should be $O(1)$ if only one boundary is conducting, but $O(Ha^{1/2})$ if both boundaries are conducting [@Dormy2; @Mizerski; @Buhler; @Soward]. Motivated by these counter-intuitive results, [@Hollerbach01b] performed a systematic investigation of linear combinations of dipole and axial fields, and showed that it is even possible to obtain both super-rotation and counter-rotation simultaneously. One finds easily enough that combinations of these two basic ingredients, dipole and axial, are sufficient to create all field line topologies that are possible in a spherical shell geometry. The purpose of this paper is to show that other topologies are possible in cylindrical geometry, and to numerically explore what happens in those cases. For example, we will show that it is possible to construct a field having a single field line that is tangent to both the inner and outer cylinders, with the tangency at the outer cylinder then suggesting a super-rotation, but the tangency at the inner cylinder suggesting a counter-rotation. So what does happen in that case? We will further explore what happens when azimuthal fields of the form $r^{-1}{\bf\hat e}_\phi$ are added, which also have no natural analog in spherical geometry. Finally, it is worth noting that there have been several liquid metal experiments related to some of the topics considered here. These include spherical Couette flow in both dipole [@Nataf1; @Brito; @Cabanes] and axial [@Sisan; @Zimmerman] fields, cylindrical Taylor-Couette flow in an axial field [@Spence; @Roach], and even electromagnetically driven flows [@Stelzer1; @Stelzer2]. However, inertia (finite Reynolds number) plays an important role in most of these results, unlike in the ‘pure’ Shercliff layer problem considered here. See also [@Hollerbach07; @Hollerbach09; @Gissinger1; @Gissinger2; @Figueroa; @Nataf2; @Kaplan] for numerical results related to some of these experiments, as well as [@Rudiger] for a general review of magnetohydrodynamic Couette flows. Equations ========= We consider a cylindrical Taylor-Couette geometry with nondimensional radii $r_i=1$ and $r_o=2$. Periodicity is imposed in $z$, with a wavelength $z_0=4$. The precise choice $z_0=4$ is not crucial, with a broad range of $O(1)$ values yielding similar Shercliff layer structures. (Taking $z_0\gg O(1)$ could well lead to different solutions though.) In the inductionless limit, the nondimensional Navier-Stokes and magnetic induction equations are $$\frac{\partial\bf U}{\partial t}=-\nabla p + \nabla^2{\bf U} - Re{\bf U\cdot\nabla U} + Ha^2(\nabla\times{\bf b})\times{\bf B}_0, \eqno(1)$$ $$\nabla^2{\bf b}=-\nabla\times({\bf U\times B}_0),\eqno(2)$$ where $\bf U$ is the fluid flow, ${\bf B}_0$ is the externally imposed magnetic field, and $\bf b$ the induced field. For the axisymmetric solutions that are relevant here, it is convenient to further decompose $\bf U$ and $\bf b$ as $${\bf U}=\nabla\times(\psi\,{\bf\hat e}_\phi) + v\,{\bf\hat e}_\phi, \qquad\qquad {\bf b}=\nabla\times(a\,{\bf\hat e}_\phi) + b\,{\bf\hat e}_\phi.\eqno(3)$$ The two nondimensional parameters are the Hartmann number $Ha={B_0 r_i}/{\sqrt{\mu\rho\nu\eta}}$ measuring the strength $B_0$ of the imposed field, and the Reynolds number $Re={\Omega r_i^2}/{\nu}$ measuring the inner cylinder’s rotation rate $\Omega$. In fact, in this work we are interested in the limit of infinitesimal differential rotation, so we set $Re\to0$ and remove the inertial term $Re{\bf U\cdot\nabla U}$, but again see [@Nataf1; @Brito; @Cabanes; @Sisan; @Zimmerman; @Spence; @Roach; @Stelzer1; @Stelzer2; @Hollerbach07; @Hollerbach09; @Gissinger1; @Gissinger2; @Figueroa; @Nataf2; @Kaplan] for a broad variety of effects that can arise at finite $Re$. The quantities $\mu$, $\rho$, $\nu$, and $\eta$ are the fluid’s permeability, density, viscosity, and magnetic diffusivity, respectively. We next turn to the allowed choices for the imposed field ${\bf B}_0$. The requirements are that it should be axisymmetric, periodic in $z$, and satisfy $\nabla\cdot{\bf B}_0=0$ and $\nabla\times{\bf B}_0={\bf0}$. The condition $\nabla\cdot{\bf B}_0=0$ is of course one of Maxwell’s original equations; the condition $\nabla\times{\bf B}_0={\bf0}$ states that ${\bf B}_0$ is a potential field, and not due to electric currents within the fluid. In addition to the familiar $z$-independent fields ${\bf\hat e}_z$ and $r^{-1}{\bf\hat e}_\phi$, the only other choices that satisfy all four of these conditions are $${\bf B}_I =\cos(\kappa z)I_0(\kappa r)\,{\bf\hat e}_z + \sin(\kappa z)I_1(\kappa r)\,{\bf\hat e}_r,\eqno(4)$$ $${\bf B}_K =\cos(\kappa z)K_0(\kappa r)\,{\bf\hat e}_z - \sin(\kappa z)K_1(\kappa r)\,{\bf\hat e}_r,\eqno(5)$$ where $\kappa=2\pi/z_0$ and $I_0$, $I_1$, $K_0$ and $K_1$ are the modified Bessel functions [@AandS]. ${\bf B}_I$ corresponds to the field that would be generated by an array of Helmholtz coils in the region $r>r_o$, with currents that alternate periodically in $z$; ${\bf B}_K$ is the field that would be generated by squeezing the Helmholtz array into the region $r<r_i$. ![(Color online.) From left to right, the four combinations that are referred to as Fields 1-4 in the text. The thick red lines in each case denote the special field lines across which the linkage to the boundaries changes, and where Shercliff layers are thus expected to arise. The general form is $({\bf\hat e}_z + c_I{\bf B}_I + c_K{\bf B}_K)/c_0$, where $c_I$ and $c_K$ are adjusted to select the desired field topology, and $c_0$ is then chosen to rescale the maximum amplitude to 1. The precise values of $(c_I, c_K)$ are $(0.155,-4.356)$, $(0.124,-3.485)$, $(0.091,-2.562)$, $(0.548,-7.687)$, respectively.[]{data-label="fig1"}](fig1.ps) Various linear combinations of ${\bf\hat e}_z$, ${\bf B}_I$ and ${\bf B}_K$ then correspond to the linear combinations of dipole and axial fields discussed above. One easily finds that it is possible to construct field line topologies that have no analog in the spherical shell geometry. Fig. 1 shows four possible combinations. In Field 1, there are some field lines that thread only the inner cylinder, some that thread only the outer cylinder, and some that connect the two. Based on the spherical results, if both boundaries are conducting, we would expect to find a super-rotating jet on one dividing line, and a counter-rotating jet on the other. In Field 2, the linear combinations have been adjusted slightly, in such a way that the previously separate dividing field lines coincide, and the region of field lines linking both boundaries has been collapsed to this single line that is tangent to both boundaries, but does not penetrate either. So, what happens in this case, when the previously expected super-rotating and counter-rotating jets should now occur on one and the same field line? In Field 3, the linear combinations have been further adjusted, so there is now a central ribbon of field lines that never touch either boundary, but just continue periodically to $z=\pm\infty$. As with Field 2, this scenario also has no spherical analog, so it is again not clear what to expect in this case. Finally, in Field 4 the linear combinations have been chosen to yield an X-type neutral point in the middle of the domain. This topology is in fact achievable in spherical geometry as well [@Hollerbach01b], so is included here primarily for completeness and comparison. We see then that just taking different combinations of ${\bf\hat e}_z$, ${\bf B}_I$ and ${\bf B}_K$ already allows us to construct topologies that have no spherical analogs. To all of these we can further add the azimuthal field $r^{-1}{\bf\hat e}_\phi$, which also has no natural analog in spherical geometry, since it is singular on the $z$-axis, which is part of the domain in spherical geometry but not in cylindrical. Since it is everywhere tangent to the boundaries, this azimuthal field will not alter the fundamental topology of the previously considered fields, but it nevertheless changes the detailed structure of the Shercliff layers that arise on the critical field lines. Indeed, including an azimuthal field component changes the solutions in at least one quite fundamental way: For purely meridional fields, the coupling between the different quantities turns out to be such that in fact $\psi$ and $a$ in Eq. (3) are identically zero (in the $Re\to0$ limit). Adding an azimuthal component to ${\bf B}_0$ introduces new couplings that result in non-zero $\psi$ and $a$. Finally, note also that a [purely]{} azimuthal ${\bf B}_0$ would not yield any interesting dynamics; the solution in that case is simply the original Couette profile $v=(-r+4r^{-1})/3$, and $\psi=a=b=0$. To summarize, the goal of this paper is to explore the Shercliff layers that occur on the critical field lines indicated in Fig. 1, either these fields alone or together with azimuthal fields of the form $r^{-1}{\bf\hat e}_\phi$. This is accomplished by using an axisymmetric, pseudo-spectral code [@Hollerbach08] to numerically solve Eqs. (1)-(3). Very briefly, $\psi$, $v$, $a$ and $b$ are expanded in terms of Chebyshev polynomials in $r$ and Fourier series in $z$. Eq. (1) is time-stepped until a stationary solution emerges; Eq. (2) is directly inverted for $\bf b$ at each time-step of Eq. (1). Resolutions as large as 240 Chebyshev polynomials and 400 Fourier modes were used, and allow Hartmann numbers as large as $Ha^2=10^7$ to be achieved. The associated boundary conditions are no-slip for $\bf U$, and either insulating or perfectly conducting boundaries for $\bf b$, referred to as I and C respectively. Other possible choices could include finitely conducting, or perhaps ferromagnetic, which in other contexts can have a significant influence [@Gissinger3]. For the Shercliff layer problem the asymptotic analyses [@Dormy2; @Mizerski; @Buhler; @Soward] indicate that the most relevant parameter is how the conductance of the exterior regions compares with the conductance of the fluid region; if this ratio is small (large) the results are similar to the insulating (perfectly conducting) case. Our I and C choices are therefore natural limiting cases, and even something at first sight quite different, such as ferromagnetic, is likely to be similar to the I case, since they both have zero conductance of the exterior regions. Results without imposed $B_\phi$ ================================ Fig. 2 shows contours of the angular velocity $\omega=v/r$ for the four choices Fields 1-4 alone, without any additional azimuthal component. In every case, the most prominent features are indeed concentrated on the particular field lines singled out in Fig. 1. The contrast between insulating and conducting boundaries is also clear; conducting boundaries exhibit both super-rotation and counter-rotation, especially for Field 4, whereas insulating boundaries only have very weak counter-rotating regions. ![(Color online.) Contours of the angular velocity $\omega=v/r$, for $Ha^2=10^6$. From left to right are the four choices Fields 1-4. The top row is for insulating boundaries, the bottom row conducting. In each panel only the upper half of the domain is shown; that is, $r$ extends over the full range $[1,2]$, but $z\in[2,4]$. The lower half $z\in[0,2]$ is reflection-symmetric in each case, as seen also in Fig. 1. The white regions indicate values between 0 at the outer boundary and 1 at the inner; the magenta contour lines in these regions have intervals 0.2. The red-shaded regions correspond to super-rotation, where $\omega>1$; the blue-shaded regions correspond to counter-rotation, where $\omega<0$. In both cases the black contour lines in these regions have intervals 1.[]{data-label="fig2"}](fig2.ps) In the insulating case there are also Hartmann layers at the boundaries. These layers are so thin, $O(Ha^{-1})$, that they cannot be seen directly at this scale. Their presence can be inferred though by the magenta contour lines, indicating values between 0.2 and 0.8, that appear to touch the boundaries. The actual imposed boundary conditions of course are $\omega =1$ at $r_i$ and $\omega=0$ at $r_o$, so these contour lines cannot touch the boundaries, and indeed they don’t, but rather remain within the Hartmann layers. These layers were investigated in detail, and always followed the expected $O(Ha^{-1})$ scalings. We therefore concentrate only on the Shercliff layers in the following discussion. To explore the details of the Shercliff layers, we require more precise diagnostics than the two-dimensional contour plots in Fig. 2. Fig. 3 shows one-dimensional cuts along the midplane $z=2$. Such cuts allow much more quantitative information to be extracted, such as how the thicknesses and amplitudes scale with $Ha$. The thicknesses were again always found to be broadly consistent with the expected $O(Ha^{-1/2})$ scalings. Regarding the amplitudes, insulating boundaries are as expected, with hardly any anomalous rotation for any of Fields 1-4. ![(Color online.) The first two rows show $\omega(r)$ at $z=2$, for Fields 1-4 from left to right as indicated, the first row insulating (I) and the second row conducting (C). Within each panel black-red-blue indicate $Ha^2=10^5$, $10^6$, $10^7$, respectively. The third row shows how the amplitude of the super-rotation in the second row scales with $Ha$, and suggests fits of the form $Ha^s$, with $s\approx0.59$, $-0.08$, $-0.43$, $0.56$, respectively.[]{data-label="fig3"}](fig3.ps) Conducting boundaries exhibit precisely the features we were expecting, and which make this problem interesting. Starting with Field 1, we see that the super-rotating jet on the field line tangent to the outer boundary is clearly increasing with increasing $Ha$, apparently scaling as $Ha^{0.59}$. The counter-rotating jet on the field line tangent to the inner boundary at $z=4$ has much the same scaling. Similarly for Field 4, we see the same behaviour even more strongly, for both the super-rotating and counter-rotating jets. Field 4 in particular is not only qualitatively, but even quantitatively very similar to corresponding results in spherical geometry – compare for example with Fig. 3 of [@Hollerbach01b]. In contrast, Field 2 still exhibits a slight super-rotation, but its amplitude seems to be practically independent of $Ha$. Similarly, a cut at $z=4$ has a slight counter-rotation, also with an $Ha$-independent amplitude. We recall that Field 2 is the case where the previously distinct field lines in Field 1 have been made to coincide. Evidently the system adjusts in such a way that weak anomalous rotations remain, but they no longer increase with increasing $Ha$. Finally, for Field 3, having this ribbon of field lines that are not connected to either boundary gives the system so much flexibility in adjusting the shear across the Shercliff layers that the anomalous rotation decreases with increasing $Ha$, apparently scaling as $Ha^{-0.43}$. To understand the origin of the anomalous rotations, we turn to the Lorentz force $Ha^2(\nabla\times{\bf b})\times{\bf B}_0$ in Eq. (1). Fig. 4 shows contours of the streamfunction of the induced current $\bf j=\nabla\times b$, for Fields 1 and 3. For both choices, I and C boundaries yield very similar patterns, consisting of clockwise circulation cells. Focusing attention specifically at the point $(r,z)=(2,2)$, the current in all four cases is therefore in the $-{\bf\hat e}_r$ direction. Since ${\bf B}_0$ at this point is in the ${\bf\hat e}_z$ direction, the Lorentz force will be in the ${\bf\hat e}_\phi$ direction. It is precisely this force which accelerates the fluid from $\omega=0$ at the boundary to $\omega>0$ in the interior. For insulating boundaries this force is just sufficient to achieve $\omega\approx1$ on those field lines linked only to the inner boundary, as seen in Figs. 2 and 3. ![Contours of $b\,r$, which constitutes the streamfunction of the electric current ${\bf j}=\nabla\times(b\,{\bf\hat e}_\phi)$. From left to right Field 1, I and C boundaries, then Field 3, I and C boundaries, and $Ha^2=10^6$ for all four. All circulation cells are clockwise, with recirculation within the Hartmann boundary layers for the I cases, and through the boundaries for the C cases. As in Fig. 2, only the upper half $z\in[2,4]$ is shown; the circulation cells in the lower half are counter-clockwise. Finally, the contour intervals from left to right are $2\cdot10^{-4}$, $2\cdot10^{-3}$, $2\cdot10^{-5}$ and $3\cdot10^{-5}$, respectively, and illustrates how switching the boundaries from I to C has a far greater effect for Field 1 than for Field 3.[]{data-label="fig4"}](fig4.ps) For conducting boundaries the system essentially ‘over-reacts’, and thereby causes the super-rotation in this region. To understand further why the system over-reacts in this way, we need to consider two (closely related) differences between the insulating and conducting results in Fig. 4. Although the patterns are generally similar for both boundary conditions, in the insulating case the current must recirculate through the Hartmann boundary layers (which are again so thin as to be barely visible here), whereas in the conducting case the current can recirculate through the exterior regions. Recirculating the current is therefore much easier in the conducting case, resulting in a stronger current, hence a stronger Lorentz force, hence the over-reaction. As indicated in Fig. 4, for Field 1 the current is an order of magnitude greater for C than for I boundaries, consistent with the increasing super-rotation, whereas for Field 3 it is only moderately greater, consistent with much weaker, and indeed decreasing super-rotation. The various other anomalous rotations, at other locations, and also for Fields 2 and 4, are similarly explained by the orientation of the Lorentz force at the position in question. The results for Fields 1 and 4 are fully consistent with the analogous scalings previously obtained in the spherical problem [@Hollerbach00; @Hollerbach01; @Buhler; @Soward]. The new cases, Fields 2 and 3, would certainly also merit further asymptotic analyses to discover the precise scalings in these cases, and why they differ from the previous results. Results with imposed $B_\phi$ ============================= To all the cases studied so far, we now wish to add azimuthal fields of the form $B_\phi=\beta r^{-1}{\bf\hat e}_\phi$, with amplitudes $\beta>0$. This is again a configuration that has not been considered before, but one that fundamentally alters the nature of the $Re\to0$ ‘pure’ Shercliff layer problem. If ${\bf B}_0$ includes an azimuthal component, then the Lorentz force $Ha^2(\nabla\times{\bf b})\times{\bf B}_0$ in Eq. (1) will include a meridional component, thereby driving a meridional circulation $\nabla\times (\psi\,{\bf\hat e}_\phi)$ that would otherwise be absent. Once $\psi\neq0$, Eq. (2) will similarly induce a field $\nabla\times(a\,{\bf\hat e}_\phi)$. In the process the previous $v$ and $b$ will also be modified. We will focus especially on how the angular velocity is altered, as well as the new flow component $\nabla\times(\psi\,{\bf\hat e}_\phi)$. We gradually increased $\beta$ from 0, and found that $\beta=O(1)$ is already sufficient to noticeably change the previous results. However, the most significant adjustments seem to occur for somewhat larger values, so we fix $\beta=10$ in the following. (That is, the Hartmann number continues to measure the strength of the imposed meridional field, but the imposed azimuthal field is $\sim10$ times stronger.) ![(Color online.) Contours of the angular velocity $\omega=v/r$, for $Ha^2=10^6$, and with $B_\phi=10r^{-1}{\bf\hat e}_\phi$ added to the previous choices Fields 1-4. All eight panels are exactly as in Fig. 2, except that the contour interval is now 0.2 throughout, for both the magenta and the black contour lines.[]{data-label="fig5"}](fig5.ps) ![(Color online.) As in Fig. 3, the first two rows show $\omega(r)$ at $z=2$, for Fields 1-4 and I and C as indicated. Within each panel black-red-blue again corresponds to $Ha^2=10^5$, $10^6$, $10^7$. In the third row, the blue lines ($+$ symbols) and the black lines ($\times$ symbols) show the scalings with $Ha$ of the I and C super-rotations, respectively.[]{data-label="fig6"}](fig6.ps) ![Contours of $\psi\,r$, which constitutes the streamfunction of the meridional circulation $\nabla\times(\psi\,{\bf\hat e}_\phi)$. From left to right Field 1, I and C boundaries, then Field 3, I and C boundaries, and $Ha^2=10^6$ for all four. White indicates negative values, grey positive. The contour intervals are $5\cdot10^{-3}$ for Field 1, and $10^{-3}$ for Field 3.[]{data-label="fig7"}](fig7.ps) Fig. 5 shows the equivalent of Fig. 2. The qualitative features are still similar, but there are also clear differences. Most notably, the very strong anomalous rotations in the conducting case have been substantially reduced. All of the various Shercliff layers also seem to be considerably thicker than before, although an examination of the variation with $Ha$ still suggests a scaling as $O(Ha^{-1/2})$. Fig. 6 again shows cuts at $z=2$. Comparing with Fig. 3, the key differences are: (a) the broadening of the Shercliff layers already noted above, (b) the presence of anomalous rotation in the I case, (c) the strong suppression of anomalous rotation in the C case, and (d) the broadly similar scalings of the anomalous rotations in the I and C cases. We note though that the anomalous rotation scalings in most cases are not as clear as in Fig. 3; for Fields 1-3 one might conjecture scalings roughly as $s\approx0.45$, $0.0$ and $-0.3$, respectively, but for Field 4 one probably should not speculate about a particular exponent at all. Finally, Fig. 7 shows examples of the meridional circulation. As one might expect, it also tends to align with the previously existing Shercliff layers, which continue to dominate the flow, that is, $U_\phi\gg U_z,\,U_r$. For both choices of imposed field the I and C options also yield broadly similar magnitudes of $\psi$. We conclude this section, and this paper, by noting that while the $\beta>0$ case yields Shercliff layers similar in many ways to the previously studied $\beta=0$ case, there are also clear differences, and many of the precise scalings are almost certainly different. An asymptotic analysis of this problem along the lines of the previous analyses [@Dormy2; @Mizerski; @Buhler; @Soward] would be of considerable interest. Acknowledgments =============== DH’s visit to Leeds was supported by an Erasmus+ scholarship and by ‘Region Stages mobilité’ from Haute-Normandie. J.A. Shercliff, The flow of conducting fluids in circular pipes under transverse magnetic fields, J. Fluid Mech. 1 (1956) 644–666. E. Dormy, P. Cardin, D. Jault, MHD flow in a slightly differentially rotating spherical shell, with conducting inner core, in a dipolar magnetic field, Earth Planet. Sci. Lett. 160 (1998) 15–30. S.V. Starchenko, Magnetohydrodynamic flow between insulating shells rotating in strong potential field, Phys. Fluids 10 (1998) 2412–2420. P.H. Roberts, Singularities of Hartmann layers, Proc. Royal Soc. A 300 (1967) 94–107. R. Hollerbach, Magnetohydrodynamic flows in spherical shells, in: C. Egbers, G. Pfister (Eds.), Physics of Rotating Fluids, Springer, 2000. R. Hollerbach, S. Skinner, Instabilities of magnetically induced shear layers and jets, Proc. Royal Soc. A 457 (2001) 785–802. E. Dormy, D. Jault, A.M. Soward, A super-rotating shear layer in magnetohydrodynamic spherical Couette flow, J. Fluid Mech. 452 (2002) 263–291. K.A. Mizerski, K. Bajer, On the effect of mantle conductivity on the super-rotating jets near the liquid core surface, Phys. Earth Planet. Inter. 160 (2007) 245-268. L. Bühler, On the origin of super-rotating layers in magnetohydrodynamic flows, Theoret. Comput. Fluid Dynam. 23 (2009) 491–507. A.M. Soward, E. Dormy, Shear-layers in magnetohydrodynamic spherical Couette flow with conducting walls, J. Fluid Mech. 645 (2010) 145–185. R. Hollerbach, Super- and counter-rotating jets and vortices in strongly magnetic spherical Couette flow, in: P. Chossat, D. Armbruster, J. Oprea (Eds.), Dynamo and Dynamics, a Mathematical Challenge, Springer, 2001. H.-C. Nataf, T. Alboussiere, D. Brito, P. Cardin, N. Gagniere, D. Jault, J.-P. Masson, D. Schmitt, Experimental study of super-rotation in a magnetostrophic spherical Couette flow, Geophys. Astrophys. Fluid Dynam. 100 (2006) 281–298. D. Brito, T. Alboussiere, P. Cardin, N. Gagniere, D. Jault, P. La Rizza, J.-P. Masson, H.-C. Nataf, D. Schmitt, Zonal shear and super-rotation in a magnetized spherical Couette-flow experiment, Phys. Rev. E 83 (2011) 066310. S. Cabanes, N. Schaeffer, H.-C. Nataf, Magnetic induction and diffusion mechanisms in a liquid sodium spherical Couette experiment, Phys. Rev. E 90 (2014) 043018. D.R. Sisan, N. Mujica, W.A. Tillotson, Y.M. Huang, W. Dorland, A.B. Hassam, T.M. Antonsen, D.P. Lathrop, Experimental observation and characterization of the magnetorotational instability, Phys. Rev. Lett. 93 (2004) 114502. D.S. Zimmerman, S.A. Triana, H.-C. Nataf, D.P. Lathrop, A turbulent, high magnetic Reynolds number experimental model of Earth’s core, J. Geophys. Res. Solid Earth 119 (2014) 4538–4557. E.J. Spence, A.H. Roach, E.M. Edlund, P. Sloboda, H. Ji, Free magnetohydrodynamic shear layers in the presence of rotation and magnetic field, Phys. Plasmas 19 (2012) 056502. A.H. Roach, E.J. Spence, C. Gissinger, E.M. Edlund, P. Sloboda, J. Goodman, H.T. Ji, Observation of a free Shercliff layer instability in cylindrical geometry, Phys. Rev. Lett. 108 (2012) 154502. Z. Stelzer, D. Cebron, S. Miralles, S. Vantieghem, J. Noir, P. Scarfe, A. Jackson, Experimental and numerical study of electrically driven magnetohydrodynamic flow in a modified cylindrical annulus. I. Base flow, Phys. Fluids 27 (2015) 077101. Z. Stelzer, S. Miralles, D. Cebron, J. Noir, S. Vantieghem, A. Jackson, Experimental and numerical study of electrically driven magnetohydrodynamic flow in a modified cylindrical annulus. II. Instabilities, Phys. Fluids 27 (2015) 084108. R. Hollerbach, E. Canet, A. Fournier, Spherical Couette flow in a dipolar magnetic field, Eur. J. Mech. B 26 (2007) 729–737. R. Hollerbach, Non-axisymmetric instabilities in magnetic spherical Couette flow, Proc. Royal Soc. A 465 (2009) 2003–2013. C. Gissinger, H. Ji, J. Goodman, Instabilities in magnetized spherical Couette flow, Phys. Rev. E 84 (2011) 026308. C. Gissinger, J. Goodman, H. Ji, The role of boundaries in the magnetorotational instability, Phys. Fluids 24 (2012) 074109. A. Figueroa, N. Schaeffer, H.-C. Nataf, D. Schmitt, Modes and instabilities in magnetized spherical Couette flow, J. Fluid Mech. 716 (2013) 445–469. H.-C. Nataf, Magnetic induction maps in a magnetized spherical Couette flow experiment, Comptes Rendus Phys. 14 (2013) 248–267. E.J. Kaplan, Saturation of nonaxisymmetric instabilities of magnetized spherical Couette flow, Phys. Rev. E 89 (2014) 063016. G. Rüdiger, L.L. Kitchatinov, R. Hollerbach, Magnetic Processes in Astrophysics: Theory, Simulations, Experiments, Wiley-VCH, 2013. M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions, Dover, 1968. R. Hollerbach, Spectral solutions of the MHD equations in cylindrical geometry, Int. J. Pure Appl. Math. 42 (2008) 575–581. C. Gissinger, A. Iskakov, S. Fauve, E. Dormy, Effect of magnetic boundary conditions on the dynamo threshold of von Kármán swirling flows, EPL 82 (2008) 29001.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The method of matched asymptotic expansions is applied to the problem of a collisionless plasma generated by UV illumination localized in a central part of the plasma in the limiting case of small Debye length $\lambda_{D}$. A second-approximation asymptotic solution is found for the double layer positioned at the boundary of the illuminated region and for the un-illuminated plasma for the plane geometry. Numerical calculations for different values of $\lambda_{D}$ are reported and found to confirm the asymptotic results. The net integral space charge of the double layer is asymptotically small, although in the plane geometry it is just sufficient to shield the ambipolar electric field existing in the illuminated region and thus to prevent it from penetrating into the un-illuminated region. The double layer has the same mathematical nature as the intermediate transition layer separating an active plasma and a collisionless sheath, and the underlying physics is also the same. In essence, the two layers represent the same physical object: a transonic layer.' author: - \| M. S. Benilov$^{1}$ and D. M. Thomas$^{2}$\ $^{1}$Departamento de Física, CCCEE, Universidade da Madeira,\ Largo do Município, 9000 Funchal, Portugal\ $^{2}$Blackett Laboratory, Imperial College London, Prince Consort Road,\ London, SW7 2BW, UK title: Asymptotic theory of double layer and shielding of electric field at the edge of illuminated plasma --- Introduction ============ In the first part of this work [@2013i] a collisionless plasma, generated by UV illumination localized in a central part of the plasma, was analyzed. The ions were assumed to be cold and the fluid description was used. Both plane and cylindrical geometries were treated. An approximate analytical solution was found under the approximation of quasi-neutrality and the exact solution was computed numerically for one value of the Debye length $\lambda_{D}$ for each geometry, this value being much smaller than widths of both illuminated and un-illuminated regions. It was found that the ions generated in the illuminated region are accelerated up to approximately the Bohm speed inside the illuminated region. In plane geometry, the ions flow across the un-illuminated region towards the near-wall positive space-charge sheath with a speed which is virtually constant and slightly exceeds the Bohm speed. In cylindrical geometry, the ions continue to be accelerated in the un-illuminated region and enter the near-wall space-charge sheath with a speed significantly exceeding the Bohm speed. In both geometries, a double layer forms where the illuminated and un-illuminated regions meet. A very unusual, if not unique, feature that this simple system reveals in plane geometry is the coexistence of two quasi-neutral plasmas of the same size with the ambipolar electric field being confined in one of them (the illuminated plasma), while the other (the un-illuminated plasma) is to high accuracy electric field-free and uniform. The latter is particularly surprising since in all known models a near-wall space-charge sheath is bordered by a nonuniform quasi-neutral presheath where the ions going to the sheath are accelerated and the voltage drop is of the order of the electron temperature measured in volts. (We set aside cases where the ions are produced on the surface, as in Q-machines or experiments with heated cavities [@Phelps1976], or inside the sheath, as in near-cathode layers of discharges burning in cathode vapour [@2010f].) Moreover, the difference between illuminated and un-illuminated regions in terms of ion momentum is in the presence or absence of ionization friction force, and the fact that the ion fluid is accelerated in the illuminated plasma, where the ionization friction force is present, and is not accelerated in the un-illuminated plasma, where the friction force is absent, is somehow counterintuitive. The above feature is extremely interesting, also from the methodological point of view; note that the classical Bohm sheath solution [@Bohm1949] is sufficient to describe both the sheath and the adjacent (un-illuminated) plasma in such situation. A key to this feature is the double layer, which shields the ambipolar electric field induced in the illuminated region and prevents it from penetrating the un-illuminated region. The quasi-neutral analytical solution [@2013i], while clearly being useful, does not describe the double layer, hence a more sophisticated treatment is needed in order to fully understand this feature and the underlying physics. It is clear that this feature originates in the smallness of $\lambda_{D}$, therefore the relevant procedure is to find an asymptotic solution to the considered problem in the limiting case of small $\lambda_{D}$. The technique of choice to this end is the method of matched asymptotic expansions (e.g., [@van-Dyke1964; @Cole1968; @Nayfeh1973; @Nayfeh1981; @Kevorkian1981; @Nayfeh1985]), which is a standard tool in problems with singular perturbations. Note that this method has been successfully used in the theory of plasma-wall transitions in collisionless plasmas; e.g., reviews [@Riemann1991; @Franklin2003c; @Allen2009; @2009a; @Riemann2009a]. In particular, relevant in the present context are works [@Lam1965; @Franklin1970], where a transition layer separating active plasma and a collisionless sheath was introduced; [@2012b], where a more adequate and simpler mathematical description of this layer was suggested; and [@1999d], where the plasma column in electronegative gases was studied including in the exceptional case where the column comprises an inner plasma, a double layer, an outer plasma, and a near-wall sheath. An approximate analytical solution in the limiting case of small $\lambda_{D}$ found by means of the method of matched asymptotic expansions is reported in this paper, which is thus complementary to [@2013i]. Also reported are results of numerical calculations for different values of $\lambda_{D}$. Questions to be answered include: what is the physics of the double layer; why the ion speed in the un-illuminated plasma deviates from the Bohm speed; how this deviation can be estimated. The outline of the paper is as follows. Equations and boundary conditions are briefly described in Sec. \[Equations\]. An asymptotic solution for plane geometry is given and compared with numerical results in Sec. \[Asymptotic solution\]. An asymptotic solution for cylindrical geometry is outlined in Sec. \[Cylindrical geometry\]. A concluding discussion is given in Sec. \[Discussion\]. Mathematical details are placed in two Appendices in order not to overload the text. Equations and boundary conditions {#Equations} ================================= We refer to [@2013i] for the description of the physical situation. In brief, we consider a plane or cylindrical plasma produced by UV radiation. Governing equations are written in the fluid approximation and are well-known and the same as in [@2013i]; they include the ion conservation equation, the ion momentum equation written without account of collisions, the equilibrium equation for the electrons, and the Poisson equation:$$\frac{1}{x^{\beta}}\frac{d}{dx}\left( x^{\beta}n_{i}v_{i}\right) =G,\label{1}$$$$m_{i}n_{i}v_{i}\frac{dv_{i}}{dx}=en_{i}E-Gm_{i}v_{i}.\label{2a}$$$$\frac{d}{dx}\left( n_{e}kT_{e}\right) +en_{e}E=0,\label{3}$$$$\frac{\varepsilon_{0}}{x^{\beta}}\frac{d\left( x^{\beta}E\right) }{dx}=e\left( n_{i}-n_{e}\right) ,\label{4}$$ where $\beta=0$ and $x$ is the distance from the plane of symmetry for plane geometry, $\beta=1$ and $x$ is the distance from the axis for cylindrical geometry, $G$ is the ionization rate, and other designations are the usual ones. It is convenient for the purposes of this work to replace the ion momentum equation Eq. (\[2a\]) by an equivalent equation which can be derived following [@2000a; @2012b] and reads$$\left( \frac{v_{i}}{c_{s}^{2}}-\frac{1}{v_{i}}\right) \frac{dv_{i}}{dx}=\frac{\varepsilon_{0}}{en_{i}}\left\{ \frac{d}{dx}\left[ \frac {1}{x^{\beta}}\frac{d\left( x^{\beta}E\right) }{dx}\right] +\frac {eE}{kT_{e}x^{\beta}}\frac{d\left( x^{\beta}E\right) }{dx}\right\} -\frac{G}{c_{s}n_{i}}\left( \frac{v_{i}}{c_{s}}+\frac{c_{s}}{v_{i}}\right) +\frac{\beta}{x}.\label{6}$$ Here $c_{s}=\sqrt{kT_{e}/m_{i}}$ is the Bohm speed. Note that for $\beta=0$ this equation coincides with the corresponding equation from [@2012b] except that Eq. (\[6\]) does not account for ion-atom collisions. We consider a situation where the illumination is localized in a central part of the plasma and is uniform there, so $G=G_{0}$ for $0\leq x<\Delta$ and $G=0$ for $\Delta<x\leq L$, where $G_{0}$ is a (positive) constant, $\Delta$ is the halfwidth of irradiated region and $L$ is the halfwidth of the system in the plane case, and $\Delta$ is the radius of the irradiated region and $L$ is the discharge tube radius in the cylindrical case. Note that $n_{i}$, $n_{e}$, $v_{i}$, and electrostatic potential $\varphi$ are, of course, continuous at $x=\Delta$; the derivatives $d\varphi/dx$ and $d^{2}\varphi/dx^{2}$ are also continuous, which follows from Eq. (\[4\]); the derivative $dn_{e}/dx$ is continuous as well, which follows from Eq. (\[3\]); the derivatives $dv_{i}/dx$, $dn_{i}/dx$, and $d^{3}\varphi/dx^{3}$ are discontinuous, which follows from, respectively, Eqs. (\[2a\]), (\[1\]), and (\[4\]). Boundary conditions are also well-known and the same as those in [@2013i]: $$x=0:\quad\quad\frac{dn_{i}}{dx}=0,\quad\quad v_{i}=0,\quad\quad\varphi =0,\quad\quad E=0;$$$$x=L:\quad\quad n_{i}v_{i}=\frac{1}{4}n_{e}\left( \frac{8kT_{e}}{\pi m_{e}}\right) ^{1/2}.\label{8}$$ Eq. (\[1\]) can be integrated directly and the result may be written in terms of the Heaviside step function $H\!\left( x\right) $ ($H\!\left( x\right) =0$ for $x<0$; $H\!\left( x\right) =1$ for $x>0$):$$n_{i}v_{i}=\frac{G_{0}x}{1+\beta}+\frac{G_{0}}{1+\beta}\left( \frac {\Delta^{1+\beta}}{x^{\beta}}-x\right) H\!\left( x-\Delta\right) .$$ Introduce dimensionless variables $$\xi=\frac{x}{\Delta},\quad\quad N_{i}=\frac{n_{i}}{n_{0}},\quad\quad N_{e}=\frac{n_{e}}{n_{0}},\quad\quad V=\frac{v_{i}}{c_{s}},\quad\quad \Phi=\frac{e\varphi}{kT_{e}},\label{5.1}$$ where $n_{0}=G_{0}\Delta/\left( 1+\beta\right) c_{s}$ is a characteristic density of the charged particles. (Note that this normalization differs from the one used in [@2013i] in that it employs normalization factors involving only control parameters.) The governing equations assume the form$$N_{i}V=\xi+\left( \frac{1}{\xi^{\beta}}-\xi\right) \,H\!\left( \xi-1\right) ,\label{5.2}$$$$\begin{aligned} \frac{V^{2}-1}{V}\frac{dV}{d\xi} & =\frac{\varepsilon}{N_{i}}\left\{ -\frac{d}{d\xi}\left[ \frac{1}{\xi^{\beta}}\frac{d}{d\xi}\left( \xi^{\beta }\frac{d\Phi}{d\xi}\right) \right] +\frac{d\Phi}{d\xi}\frac{1}{\xi^{\beta}}\frac{d}{d\xi}\left( \xi^{\beta}\frac{d\Phi}{d\xi}\right) \right\} \label{ion_mom}\\ & -\frac{1+\beta}{\xi}\left( V^{2}+1\right) H\!\left( 1-\xi\right) +\frac{\beta}{\xi},\nonumber\end{aligned}$$$$\frac{dN_{e}}{d\xi}-N_{e}\frac{d\Phi}{d\xi}=0,$$$$\frac{\varepsilon}{\xi^{\beta}}\frac{d}{d\xi}\left( \xi^{\beta}\frac{d\Phi }{d\xi}\right) =N_{e}-N_{i},\label{5.5}$$ where $\lambda_{D}=\left( \varepsilon_{0}kT_{e}/e^{2}n_{0}\right) ^{1/2}$ and $\varepsilon=\left( \lambda_{D}/\Delta\right) ^{2}$. The next section, Sec. \[Asymptotic solution\], is concerned with plane geometry, $\beta=0$. The asymptotic solution for cylindrical geometry, $\beta=1$, is outlined in Sec. \[Cylindrical geometry\]. Asymptotic solution: plane geometry {#Asymptotic solution} =================================== Asymptotic structure of the solution {#The approach} ------------------------------------ Asymptotic zones that need to be considered may be conveniently illustrated by invoking results of numerical calculations. The calculations have been performed in the same way as in [@2013i] and an example is depicted by solid lines in Fig. 1. One can clearly see four regions with apparently different physics: the illuminated plasma, $0\leq\xi<1$; the un-illuminated plasma, $1<\xi<2$; a thin intermediate (double) layer positioned in the vicinity of the point $\xi=1$ and separating the two plasmas; and a thin space-charge sheath positioned at the wall, i.e., at $\xi$ close to $2$. In the course of application of the method of matched asymptotic expansions the same four regions appear as asymptotic zones described by different asymptotic expansions. The illuminated and un-illuminated plasmas are described by straightforward expansions, i.e., the first term of each of these expansions is of the order unity and governed by equations which are obtained from Eqs. (\[5.2\])-(\[5.5\]) by setting $\varepsilon=0$. The latter is, of course, consistent with the plasmas in both the illuminated and un-illuminated regions being quasi-neutral. The double layer is described by the same asymptotic expansion as the one which describes the transition layer separating active plasma and a collisionless sheath [@Franklin1970]. The first term of the expansions of each of the quantities $V$, $N_{i}$, $N_{e}$, and $\Phi$ is constant. Therefore, the double layer appears only in the second approximation and may be ignored in the first one. The near-wall space-charge sheath is to the first approximation the well-known Bohm space-charge sheath. Its asymptotic description in the present problem is similar to the description given in [@Bohm1949] and is skipped for brevity. The first approximation {#The first approximation} ----------------------- The region of illuminated plasma, $0\leq\xi<1$, is described by the straightforward expansion$$V\left( \xi,\varepsilon\right) =V_{1}\left( \xi\right) +\alpha_{1}V_{2}\left( \xi\right) +\dots,\label{10}$$$$N_{i}\left( \xi,\varepsilon\right) =\frac{\xi}{V_{1}\left( \xi\right) }\left[ 1-\alpha_{1}\frac{V_{2}\left( \xi\right) }{V_{1}\left( \xi\right) }\right] +\dots,$$$$N_{e}\left( \xi,\varepsilon\right) =N_{e1}\left( \xi\right) +\alpha _{1}N_{e2}\left( \xi\right) +\dots,$$$$\Phi\left( \xi,\varepsilon\right) =\Phi_{1}\left( \xi\right) +\alpha _{1}\Phi_{2}\left( \xi\right) +\dots,\label{13}$$ where $\alpha_{1}=\alpha_{1}\left( \varepsilon\right) $ is a small parameter which is to be found in the course of analysis as a part of solution. Substituting expansion (\[10\])-(\[13\]) into Eqs. (\[ion\_mom\])-(\[5.5\]), expanding and retaining the leading terms, one obtains equations$$\frac{V_{1}^{2}-1}{V_{1}}\frac{dV_{1}}{d\xi}=-\frac{1}{\xi}\left( V_{1}^{2}+1\right) \,,$$$$\frac{dN_{e_{1}}}{d\xi}-N_{e1}\frac{d\Phi_{1}}{d\xi}=0,$$$$N_{e1}-\frac{\xi}{V_{1}}=0.$$ A solution subject to boundary condition $\Phi_{1}\left( 0\right) =0$ reads$$V_{1}=\frac{C_{1}\xi}{1+\sqrt{1-\left( C_{1}\xi\right) ^{2}}},\;\;\;\;N_{e1}=\frac{1+\sqrt{1-\left( C_{1}\xi\right) ^{2}}}{C_{1}},\;\;\;\Phi_{1}=\ln\frac{1+\sqrt{1-\left( C_{1}\xi\right) ^{2}}}{2},$$ where $C_{1}$ is an integration constant, $0<C_{1}\leq1$. Note that $V_{1}\left( 1\right) $ (i.e., the value of the function $V_{1}\left( \xi\right) $ for $\xi=1$) increases with increasing $C_{1}$ and its maximum value is attained at $C_{1}=1$ and equals $1$. The region of un-illuminated plasma, $1<\xi<S=L/\Delta$, is described by the straightforward expansion $$V\left( \xi,\varepsilon\right) =V_{3}\left( \xi\right) +\alpha_{2}V_{4}\left( \xi\right) +\dots,\label{22}$$$$N_{i}\left( \xi,\varepsilon\right) =\frac{1}{V_{3}\left( \xi\right) }\left[ 1-\alpha_{2}\frac{V_{4}\left( \xi\right) }{V_{3}\left( \xi\right) }\right] +\dots,$$$$N_{e}\left( \xi,\varepsilon\right) =N_{e3}\left( \xi\right) +\alpha _{2}N_{e4}\left( \xi\right) +\dots,$$$$\Phi\left( \xi,\varepsilon\right) =\Phi_{3}\left( \xi\right) +\alpha _{2}\Phi_{4}\left( \xi\right) +\dots,\label{25}$$ where $\alpha_{2}=\alpha_{2}\left( \varepsilon\right) $ is a small parameter which is to be found as a part of solution. The leading terms of this expansion are governed by the equations $$\left( V_{3}-\frac{1}{V_{3}}\right) \frac{dV_{3}}{d\xi}=0,\label{26}$$$$\frac{dN_{e3}}{d\xi}-N_{e3}\frac{d\Phi_{3}}{d\xi}=0,$$$$N_{e3}-\frac{1}{V_{3}\left( \xi\right) }=0.\label{28}$$ Expansion (\[22\])-(\[25\]) is to be matched with a double layer expansion in the vicinity of the point $\xi=1$ and with a sheath expansion in the vicinity of the point $\xi=S$. However, the double layer needs to be taken into account only in the second and higher approximations, so the first terms of the expansions describing the illuminated and un-illuminated plasmas should be patched directly, meaning that $V_{1}\left( 1\right) =V_{3}\left( 1\right) $, $N_{e1}\left( 1\right) =N_{e3}\left( 1\right) $, $\Phi _{1}\left( 1\right) =\Phi_{3}\left( 1\right) $. Solutions to Eqs. (\[26\])-(\[28\]) subject to these boundary conditions are trivial: $V_{3}=V_{1}\left( 1\right) $, $N_{e3}=N_{e1}\left( 1\right) $, $\Phi _{3}=\Phi_{1}\left( 1\right) $. In order that a matching with the sheath expansion be possible, $V_{3}$ must satisfy the Bohm criterion, i.e., it should be $V_{3}\geq1$. Since $V_{1}\left( 1\right) \leq1$ as discussed above, it follows that $V_{1}\left( 1\right) =V_{3}=1$ and $C_{1}=1$. It follows also that $N_{e3}=1$, $\Phi_{3}=-\ln2$. The first-approximation solution is complete now. In [@2013i], the same solution was obtained by patching solutions for the illuminated and un-illuminated plasmas obtained with the use of the condition of quasi-neutrality and the Bohm criterion. This solution is depicted by the dotted lines in Fig. 1. As expected, it provides a reasonable approximation of the ion speed. However, in the double layer this approximation is not smooth, and the approximation of the electric field is not suitable altogether. Poor approximation of derivatives of approximate solutions is a deficiency inherent to patching. In the method of matched asymptotic expansions, this deficiency can be removed by going to the second approximation. The second approximation {#The second approximation} ------------------------ It is natural to seek an asymptotic expansion describing the double layer in the same form as that of the expansion which describes the transition layer separating active plasma and a collisionless sheath [@Franklin1970]:$$V\left( \xi,\varepsilon\right) =1+\varepsilon^{1/5}V_{5}\left( \xi _{5}\right) +\dots,\label{30}$$$$N_{i}\left( \xi,\varepsilon\right) =1-\varepsilon^{1/5}V_{5}\left( \xi _{5}\right) +\dots,\label{31}$$$$N_{e}\left( \xi,\varepsilon\right) =1+\varepsilon^{1/5}N_{e5}\left( \xi _{5}\right) +\dots,\label{32}$$$$\Phi\left( \xi,\varepsilon\right) =-\ln2+\varepsilon^{1/5}\Phi_{5}\left( \xi_{5}\right) +\dots,\label{33}$$ where $\xi_{5}=\left( \xi-1\right) /\varepsilon^{2/5}$. The form of this expansion was derived in [@Franklin1970] with the use of considerations stemming from matching and in [@2012b] by means of estimates of different terms of equation (\[6\]). A convenient illustration of this expansion is provided by Fig. 1: the expansion implies that variations of the quantities $V$, $N_{i}$, $N_{e}$, and $\Phi$ in the double layer are of the order of $\varepsilon^{1/5}$ and the electric field is of the order of $\varepsilon ^{-1/5}$, and this is consistent with Fig. 1. Substituting this expansion into Eqs. (\[ion\_mom\])-(\[5.5\]), expanding and retaining the leading terms, one obtains equations$$2V_{5}\frac{dV_{5}}{d\xi_{5}}=-\frac{d^{3}\Phi_{5}}{d\xi_{5}^{3}}-2H\!\left( -\xi_{5}\right) ,\label{34}$$$$\frac{dN_{e5}}{d\xi_{5}}-\frac{d\Phi_{5}}{d\xi_{5}}=0,\label{35}$$$$N_{e5}+V_{5}=0.\label{36}$$ As seen from Eq. (\[36\]), the plasma in the double layer is quasi-neutral not only to the first approximation, but also to the second one. A consequence is that the derivative $dV_{5}/d\xi_{5}$ is continuous at $\xi_{5}=0$. Coming back to the discussion of Sec. \[Equations\], one can say that the discontinuity of derivative $dv_{i}/dx$ at $x=\Delta$ occurs not in the first approximation in $\varepsilon$ but rather in the second one. The latter is consistent with the second term on the rhs of Eq. (\[2a\]), which is responsible for the discontinuity, being of the order of $\varepsilon^{1/5}$, i.e., asymptotically small, with respect to the first term. Eqs. (\[34\]) and (\[35\]) may be integrated term by term to give$$V_{5}^{2}+\frac{d^{2}\Phi_{5}}{d\xi_{5}^{2}}+2\xi_{5}H\!\left( -\xi _{5}\right) =C_{2}^{2},\label{37}$$$$N_{e5}-\Phi_{5}=C_{3},\label{38}$$ where $C_{2}^{2}$ and $C_{3}$ are integration constants. Note that since every term on the lhs of these equations is continuous at $\xi=0$, the constants $C_{2}^{2}$ and $C_{3}$ do not switch their values between the regions $\xi>0$ and $\xi<0$. The system of Eqs. (\[37\]), (\[38\]), (\[36\]) may be reduced to a single equation, for example, for the unknown $V_{5}$:$$V_{5}^{2}=\frac{d^{2}V_{5}}{d\xi_{5}^{2}}-2\xi_{5}H\!\left( -\xi_{5}\right) +C_{2}^{2}.\label{43}$$ A solution of this equation subject to relevant boundary conditions, which follow from the van Dyke asymptotic matching principle [@van-Dyke1964], is found in Appendix \[Appendix\] and depicted in Fig. 2 by the solid line for $\xi_{5}\leq0$ and dash-dotted line for $\xi_{5}\geq0$. (Here $X=\xi _{5}-0.2254$.) The constant $C_{2}$ is found to be equal to $0.6714$. Also shown in Fig. 2 is the function $V_{5}\left( X\right) $ referring to the transition layer separating active plasma and a collisionless sheath [@Franklin1970]. The two functions coincide in the range $X\leq-0.2254$ and differ for bigger $X$. We will need for subsequent asymptotic matching the two-term asymptotic expansion of the functions $V_{5}\left( \xi_{5}\right) $ and $N_{5}\left( \xi_{5}\right) $ for $\xi\rightarrow-\infty$, the three-term asymptotic expansion of the function $\Phi_{5}\left( \xi_{5}\right) $ for $\xi\rightarrow-\infty$, and the one-term expansion of all these functions for $\xi\rightarrow\infty$. These expansions may be readily found: $$\begin{aligned} \xi_{5}\rightarrow-\infty:\;\;\; & V_{5}=-\sqrt{-2\xi_{5}}-\frac{C_{2}^{2}}{\sqrt{-8\xi_{5}}}+\dots,\;\;\;N_{e5}=\sqrt{-2\xi_{5}}+\frac{C_{2}^{2}}{\sqrt{-8\xi_{5}}}+\dots,\;\;\;\label{42}\\ & \Phi_{5}=\sqrt{-2\xi_{5}}-C_{3}+\frac{C_{2}^{2}}{\sqrt{-8\xi_{5}}}+\dots;\nonumber\end{aligned}$$$$\xi_{5}\rightarrow\infty:\;\;\;V_{5}\rightarrow C_{2},\;\;\;N_{e5}\rightarrow-C_{2},~~~~\Phi_{5}\rightarrow-C_{2}-C_{3}.\label{48}$$ Let us now consider the asymptotic matching of expansions (\[30\])-(\[33\]) and (\[10\])-(\[13\]) accounting for two terms in each expansion. Making use of Eq. (\[42\]), one finds that the matching is possible provided that $\alpha_{1}=\varepsilon^{2/5}$, $C_{3}=0$, and $$V_{2}=-\frac{C_{2}^{2}}{\sqrt{8\left( 1-\xi\right) }}+\dots,\;\;\;\;N_{e2}=\frac{C_{2}^{2}}{\sqrt{8\left( 1-\xi\right) }}+\dots,\;\;\text{\ }\Phi _{2}=\frac{C_{2}^{2}}{\sqrt{8\left( 1-\xi\right) }}+\dots\label{54}$$ for $\xi\rightarrow1-0$. Eq. (\[54\]) represents a set of boundary conditions for differential equations governing functions $V_{2}$, $N_{e2}$, and $\Phi_{2}$, which can be obtained by substituting expansion (\[10\])-(\[13\]) into Eqs. (\[ion\_mom\])-(\[5.5\]), expanding, and retaining second-order terms. However, these functions represent a correction of the order of $\alpha_{1}=\varepsilon^{2/5}$, which is higher than the order to which the solution is known in the double layer ($\varepsilon^{1/5}$). For this reason, we leave finding these functions beyond the scope of this work and only note that the boundary conditions (\[54\]) are compatible with the differential equations. Let us now consider the asymptotic matching of expansions (\[30\])-(\[33\]) and (\[22\])-(\[25\]) accounting for two terms in each expansion. Taking into account Eq. (\[48\]), one finds that the matching is possible provided that $\alpha_{2}=\varepsilon^{1/5}$ and $$V_{4}\left( 1\right) =C_{2},\;\;\;N_{e4}\left( 1\right) =-C_{2},\;\;\;\Phi_{4}\left( 1\right) =-C_{2}.\label{50a}$$ Substituting expansions (\[22\])-(\[25\]) into Eqs. (\[ion\_mom\])-(\[5.5\]), expanding, and retaining second-order terms, one obtains equations$$V_{4}\frac{dV_{4}}{d\xi}=0,\label{50}$$$$\frac{dN_{e4}}{d\xi}-\frac{d\Phi_{4}}{d\xi}=0,\;\;\;N_{e4}+V_{4}=0.\label{52}$$ Solution of Eqs. (\[50\]) and (\[52\]) subject to the boundary conditions Eq. (\[50a\]) is trivial: $V_{4}=C_{2}$, $N_{e4}=-C_{2}$, $\Phi_{4}=-C_{2}$. The above results ensure a description of the illuminated plasma region to the double layer to the un-illuminated region with the error of the order of $\varepsilon^{2/5}$ (which is, presumably, the order of the third terms of the expansions describing the double layer and the un-illuminated region). These results allow one to construct a composite expansion [@van-Dyke1964; @Cole1968; @Nayfeh1973; @Nayfeh1981; @Kevorkian1981; @Nayfeh1985] uniformly valid in all these regions with the error of the order of $\varepsilon^{2/5}$; see Appendix \[Composite expansions\]. Comparison of the above asymptotic solution with results of numerical calculations is shown in Figs. 3-5. While analyzing Figs. 3b-3d, it should be kept in mind that $\Phi_{5}=-V_{5}$, which follows from Eqs. (\[36\]) and (\[38\]). Dotted lines in Figs. 3 and 4 represent the data given by Eqs. (\[61\]), (\[61a\]), and (\[62a\]). Values of the ion speed in the un-illuminated region shown in Fig. 5 have been taken at points where $dV/d\xi$ attains the minimum value except in the case $\varepsilon =0.99\times10^{-5}$, where variation of the ion speed in the un-illuminated region was below the usual floating-point precision. One can see the deviation between the asymptotic and numerical results is reasonably small and decreases with decreasing $\varepsilon$. Cylindrical geometry {#Cylindrical geometry} ==================== One needs to consider the same four asymptotic zones as in the case of plane geometry. A composite first-approximation solution uniformly valid from the illuminated plasma to the double layer to the un-illuminated plasma is governed by Eq. (\[ion\_mom\]) with $\varepsilon=0$ and $\beta=1$:$$\frac{V^{2}-1}{V}\frac{dV}{d\xi}=-\frac{2}{\xi}\left( V^{2}+1\right) H\!\left( 1-\xi\right) +\frac{1}{\xi}.\label{8a}$$ A continuous single-valued solution of this equation may be written in the implicit form:$$\xi=\left\{ \begin{tabular} [c]{lll}$\frac{3^{3/4}V}{\left( 2V^{2}+1\right) ^{3/4}}$ & for & $V\leq1$\\ $\frac{1}{V}\exp\frac{V^{2}-1}{2}$ & for & $V\geq1$\end{tabular} \ \ \right. .\label{12}$$ One can see that $V\left( 1\right) =1$, i.e., the sonic point is positioned at the edge of the illuminated plasma, as in the planar case. The difference is that the solution (\[12\]) is obtained without invoking the Bohm criterion; note that Eq. (\[8a\]) admits continuous solutions with $V\left( 1\right) \neq1$, however these solutions are multi-valued. Another difference is that $V>1$ for $\xi>1$, i.e., the plasma continues to be accelerated in the un-illuminated plasma. Most formulas of Sec. \[The second approximation\] referring to the double layer may be readily generalized in order to become applicable to both plane and cylindrical geometries. In particular, this includes replacing the first term on the rhs of Eq. (\[33\]) with $-\left( 2+\beta\right) \left( 2+2\beta\right) ^{-1}\ln\left( 2+\beta\right) $ and writing Eq. (\[43\]) in the form$$V_{5}^{2}=\frac{d^{2}V_{5}}{d\xi_{5}^{2}}-2\left( 1+\beta\right) \xi _{5}H\!\left( -\xi_{5}\right) +\beta\xi_{5}+C_{2}^{2}.\label{12a}$$ A major difference between the functions $V_{5}\left( \xi_{5}\right) $ for plane and cylindrical geometries is in their asymptotic behavior for $\xi _{5}\rightarrow\infty$: for $\beta=1$ the latter is governed by the term $\beta\xi_{5}$ on the rhs of Eq. (\[12a\]) and reads $V_{5}\approx\sqrt {\xi_{5}}$ , which is similar to the first relation in (\[42\]) rather than (\[48\]). Note that this difference does not lead to significantly different behavior of the normalized electric field and space-charge density, $dV_{5}/d\xi_{5}$ and $dV_{5}^{2}/d\xi_{5}^{2}$: both tend to zero as $\xi _{5}\rightarrow\infty$, although algebraically rather than exponentially as in plane geometry. Expansions (\[10\])-(\[13\]) and (\[22\])-(\[25\]), describing regions of, respectively, illuminated and un-illuminated plasmas, remain applicable in cylindrical geometry except that $\alpha_{2}$ becomes equal to $\varepsilon ^{2/5}$. Concluding discussion {#Discussion} ===================== The four zones with different physics revealed by numerical calculations and shown in Fig. 1, that is, the illuminated plasma, the double layer, the un-illuminated plasma, and the near-wall space-charge sheath appear in a natural way in the course of application of the method of matched asymptotic expansions. The first-order terms of asymptotic expansions of $V$, $N_{i}$, and $\Phi$ describing the illuminated plasma, the double layer, and the un-illuminated plasmas represent the same quasi-neutral solution that was found in [@2013i] by patching solutions for the illuminated and un-illuminated plasmas and (in plane geometry) invoking the Bohm criterion. The second-order terms are of the order of $\varepsilon^{2/5}$ in the illuminated plasma; $\varepsilon^{1/5}$ in the double layer; and $\varepsilon^{1/5}$ in the un-illuminated plasma in plane geometry and $\varepsilon^{2/5}$ in cylindrical geometry. It is interesting to note that these orders explain why the quasi-neutral solution for $V$, $N_{i}$, and $\Phi$ in Fig. 1 of this work and Figs. 2 and 3 of [@2013i] is more accurate in the illuminated plasma than in the double layer and why its accuracy in the un-illuminated plasma is comparable to that in the double layer in plane geometry and in the illuminated plasma in cylindrical geometry. The double layer separating the illuminated and un-illuminated plasmas is quasi-neutral: the separation of charges here is of the order of $\varepsilon^{2/5}$, i.e., asymptotically small, although much bigger than in the illuminated and un-illuminated plasmas. The net integral space charge (per unit cross section) of the double layer is asymptotically small, which can be readily seen: in the first approximation it is proportional to $\left. dV_{5}/d\xi_{5}\right\vert _{-\infty}^{\infty}$ and equals zero by virtue of boundary conditions (\[42\]) and (\[48\]). However, in plane geometry the net integral space charge of the double layer is just sufficient to shield the ambipolar electric field existing in the illuminated region and thus to prevent it from penetrating the un-illuminated region. The effect of the ionization friction force \[the second term on the rhs of Eq. (\[2a\])\] is as well asymptotically small in the double layer, therefore the full energy of an ion is conserved as reflected by the relationship $V_{5}+\Phi_{5}=0$. The double layer has the same mathematical nature as the intermediate transition layer separating an active plasma and a collisionless sheath, which was introduced in [@Lam1965; @Franklin1970] and revisited in, e.g., [@Riemann1997; @Kaganovich2002; @Riemann2005; @Riemann2009a; @2012b]. The underlying physics is also the same and is related to the passage of the ion fluid through the sonic barrier [@2012b]. In essence, the two layers represent the same physical object: the transonic layer, which in one case assumes the form of a double layer separating the illuminated and un-illuminated plasmas and in the other case the form of a (positive-charge) transition layer separating the plasma and a collisionless sheath. In the form of transition layer separating the plasma and a collisionless sheath, the transonic layer is difficult to identify in results of numerical solution of a full problem. In fact, some researchers even believe that the transition layer is merely an artifact produced by the method of matched asymptotic expansions. However, the transonic layer at the edge of illuminated plasma, having the form of a double layer, is easily identifiable; another proof that the transonic layer is distinguished by specific physical processes and is therefore not less real than, e.g., the near-wall space-charge sheath. The normalized ion speed in the un-illuminated plasma in plane geometry given by the asymptotic analysis equals $1+C_{2}\varepsilon^{1/5}$, where $C_{2}=0.6714$. Given that $V_{1}\left( \xi\right) $ equals $1$ when extrapolated to $\xi=1$, a natural interpretation is that the ion fluid is accelerated up to $1$ (the Bohm speed) in the illuminated plasma and from $1$ to $1+C_{2}\varepsilon^{1/5}$ in the double layer. A similar interpretation applies to potential: the potential difference across the illuminated plasma equals $-\ln2$ and the potential difference across the double layer is $-C_{2}\varepsilon^{1/5}$. The double layers in plane and cylindrical geometries are not qualitatively different as far as distributions of the electric field and space-charge densities are concerned. However, distributions of the ion speed are significantly different: the ion fluid in the double layer is continually accelerated in cylindrical geometry. This is a consequence of the appearance in the case $\beta=1$ of an additional term ($\xi_{5}$) on the rhs of Eq. (\[12a\]), which governs distribution of ion speed in the double layer. This difference may look somewhat surprising: the double layer is asymptotically thin, i.e., locally planar, so how can its curvature affect its structure? Note, however, that this is not the only weak effect affecting the double layer: the ionization and separation of charges also play a role; cf. the first and second terms on the rhs of Eq. (\[12a\]). The reason for the latter was discussed in [@2012b] and applies to the curvature effect as well: the balance of forces acting over the ion fluid is delicate in the vicinity of the sonic point since the main effects (inertia force and the electrostatic force) cancel; therefore, the weak effects (ionization, curvature, and separation of charges) also play a role. One should stress in this connection that Eq. (\[12a\]), describing acceleration of the ion fluid in the double layer, appears in the second approximation rather than in the first one. In plane geometry, the ions enter the near-wall space-charge sheath with a speed that equals the Bohm speed in the first approximation in $\varepsilon$. In other words, the Bohm criterion is satisfied with the equality sign, which is a usual situation, and in the first approximation the near-wall sheath is the usual Bohm sheath. In other words, the first term of the asymptotic expansion describing the sheath in the considered problem is exactly the same as in situations involving an active plasma and a collisionless sheath, for example, in the problem treated in [@Franklin1970], or in a situation which would have occurred in the problem considered here if the whole plasma were illuminated. However, the second terms of the asymptotic expansions describing the sheath in this problem and in usual situations should be different, which is due to the different behavior of the second-order term in the asymptotic zone adjacent to the sheath: in this problem, the second-order term of expansion of $V$ in the un-illuminated plasma is constant (and equal to $C_{2}\varepsilon^{1/5}$), while in usual situations the second-order term of expansion of $V$ in the transition layer has a pole at the wall. Thus, the near-wall space-charge sheath in the plane partially illuminated plasma is the usual Bohm sheath to the first approximation in $\varepsilon$ but not to the second approximation. In cylindrical geometry, the ions enter the sheath with a speed exceeding the Bohm speed. In other words, the Bohm criterion is “oversatisfied”; a situation which was known to occur only in a few artificial models [@Riemann1991]. It has been known for many decades that a (quasi-neutral) presheath, in which the ion fluid is accelerated up to the Bohm speed, must involve at least one of the following three mechanisms: ion-atom collisions, ionization, and multidimensional effects; e.g., [@Riemann1991]. As far as ion-atom collisions and ionization are concerned, this result is somehow counterintuitive: in terms of ion momentum both ion-atom collisions and ionization represent a friction force, and why should a friction force be needed for acceleration? An explanation of this paradox is as follows. Of course, the ion fluid is accelerated by the electrostatic force while a friction force has a decelerating effect; cf. Eq. (\[2a\]). On the other hand, the electrostatic force in a collisionless plane subsonic quasi-neutral plasma with frozen ionization would exceed the ion inertia force and the ion momentum balance cannot be ensured; cf. Eq. (\[6\]) with the rhs dropped. In other words, a friction force, while not being the reason of acceleration of a plane subsonic ion flow under conditions of quasi-neutrality, is its necessary attribute. The problem of plane partially illuminated plasma offers a remarkable illustration of this statement: the ion flow is accelerated in the illuminated plasma, where a friction force due to ionization is present; there is no friction force in the un-illuminated plasma - and therefore no acceleration. In cylindrical geometry the rhs of Eq. (\[8a\]) in the illuminated region equals $-\frac{2}{\xi}V^{2}-\frac{1}{\xi}$ and is negative. In the un-illuminated plasma the rhs of Eq. (\[8a\]) equals $1/\xi$ and is positive. In other words, cylindrical geometry provides a retarding effect in the subsonic region and the ion flow is accelerated due to the presence of the ionization friction force. The sonic point occurs where the ionization friction force disappears. Cylindrical geometry provides an accelerating effect in the supersonic region. The possibility of experimental testing was discussed in the first part of this work [@2013i]. In this connection, relevant is the case where the ionization profile is not described by the Heaviside function and decays smoothly. Asymptotic structure of the solution will remain the same in this case provided that the decay is fast (occurs on a length scale much smaller than widths of both illuminated and un-illuminated plasmas). In particular, there will be a quasi-neutral double layer, where the sonic transition occurs. The theory of the double layer developed in this work will remain applicable if length scale of the decay is much smaller than $\varepsilon^{2/5}\Delta$. Acknowledgments The authors appreciate stimulating discussions with R. N. Franklin and J. E. Allen. Work at Universidade da Madeira was supported by FCT - Fundação para a Ciência e a Tecnologia of Portugal through the projects PTDC/FIS-PLA/2708/2012 and PEst-OE/MAT/UI0219/2011. Finding solution for the plane double layer {#Appendix} ============================================ In the region $\xi_{5}>0$, Eq. (\[43\]) does not explicitly depend on the independent variable (is autonomous) and admits an analytical solution, which may be found as follows. Multiplying this equation by $dV_{5}/d\xi_{5} $ and integrating, one finds$$\frac{V_{5}^{3}}{3}=\frac{1}{2}\left( \frac{dV_{5}}{d\xi_{5}}\right) ^{2}+C_{2}^{2}V_{5}+C_{4},\label{44}$$ where $C_{4}$ is an integration constant. Solving this equation for $\left( dV_{5}/d\xi_{5}\right) ^{-1}$, choosing the sign in the square root with the use of the assumption that $dV_{5}/d\xi_{5}>0$ and integrating, one obtains the desired solution in an implicit form $$\int_{C_{5}}^{V_{5}}\frac{dV_{5}}{\left( V_{5}^{3}-3C_{2}^{2}V_{5}-3C_{4}\right) ^{1/2}}=\sqrt{\frac{2}{3}}\xi_{5},\label{45}$$ where $C_{5}$ is a new integration constant. Since the solution should exist for all positive values of $\xi_{5}$, the integral on the lhs of Eq. (\[45\]) should diverge at a certain value of $V_{5}$. The latter value should be finite, since the integral converges as $V_{5}\rightarrow\infty$, and should represents the double root of the cubic polynomial in the parentheses in the denominator. Such a root exists provided that $C_{4}=-2C_{2}^{3}/3$ and equals $C_{2}$. Then the integral may be evaluated analytically and the solution may be transformed to an explicit form:$$V_{5}=3C_{2}\tanh^{2}\left( \sqrt{\frac{C_{2}}{2}}\xi_{5}+\operatorname{arctanh}\sqrt{\frac{C_{5}+2C_{2}}{3C_{2}}}\right) -2C_{2}\text{.}\label{47}$$ The solution involves two integration constants, $C_{2}$ and $C_{5}$, which have the meaning $C_{2}=V_{5}\left( \infty\right) $ and $C_{5}=V_{5}\left( 0\right) $. It should be the case that $C_{2}>0$, $-2C_{2}<C_{5}<C_{2}$. Let us now consider Eq. (\[43\]) in the region $\xi_{5}\leq0$. The boundary condition at $\xi_{5}=0$ is expressed by Eq. (\[44\]) with $C_{4}=-2C_{2}^{3}/3$; we recall that not only $V_{5}\left( \xi_{5}\right) $ but also $dV_{5}\left( \xi_{5}\right) /d\xi_{5}$ are continuous at $\xi_{5}=0$ as discussed in Sec. \[The second approximation\]. Another boundary condition is obtained by matching two terms of the expansion (\[30\]) with the first term of the expansion (\[10\]) and reads$$\xi_{5}\rightarrow-\infty:\;\;V_{5}=-\sqrt{-2\xi_{5}}+\dots.\label{41}$$ Note that the problem (\[43\]), (\[44\]), (\[41\]) coincides with the problem describing the intermediate transition layer separating active plasma and a collisionless sheath [@Franklin1970] except for a different boundary condition at $\xi_{5}=0$. Another substantial difference is in the way in which Eq. (\[43\]) was derived: in [@Franklin1970] it was derived by means of invoking the next (third) term of the asymptotic expansion (\[30\])–(\[33\]), while in this work it was derived in a straightforward way from Eq. (\[ion\_mom\]); see discussion in [@2012b]. Constant $C_{2}$ may be eliminated from Eq. (\[43\]) by means of substitution $\xi_{5}=X+C_{2}^{2}/2$:$$\frac{d^{2}V_{5}}{dX^{2}}=V_{5}^{2}+2X.\label{47a}$$ Note that this equation coincides with the first Painlevé equation (e.g., [@Davis1962; @Polyanin2007]) to the accuracy of transformation $X=3^{1/5}\tilde{X}$, $V_{5}=2\times3^{3/5}Y$. Asymptotic behavior for $X\rightarrow-\infty$ of solutions of Eq. (\[47a\]) satisfying the matching condition (\[41\]) may be found to be$$\begin{aligned} V_{5} & =-\left( -2X\right) ^{1/2}-\frac{1}{8X^{2}}+\dots+\frac{C_{6}}{\left( -X\right) ^{1/8}}\exp\left[ i\frac{2^{11/4}}{5}\left( -X\right) ^{5/4}\right] \left( 1+\dots\right) \nonumber\\ & +\frac{C_{7}}{\left( -X\right) ^{1/8}}\exp\left[ -i\frac{2^{11/4}}{5}\left( -X\right) ^{5/4}\right] \left( 1+\dots\right) ,\label{16}$$ where $C_{6}$ and $C_{7}$ are arbitrary constants. The last two terms on the rhs of this expression are oscillatory and have to be eliminated in order that the asymptotic matching be possible beyond the first order; i.e., one should set $C_{6}=C_{7}=0$. Thus, the proper boundary condition for Eq. (\[47a\]) is given by Eq. (\[16\]) with $C_{6}=C_{7}=0$ and this boundary condition specifies a unique solution of Eq. (\[47a\]). Note that this conclusion coincides with the corresponding conclusion in [@Franklin1970], although Eq. (\[16\]) is not exactly the same. A convenient way of numerically finding the function $V_{5}$ for $\xi_{5}\leq0$ is as follows: Eq. (\[47a\]) is integrated with the initial condition $V_{5}=-\left( -2X\right) ^{1/2}-1/8X^{2}$ from a large negative $X $ in the direction of increasing $X$ until the condition (\[44\]) written in the form$$\frac{V_{5}^{3}}{3}=\frac{1}{2}\left( \frac{dV_{5}}{dX}\right) ^{2}-2XV_{5}-\frac{2}{3}\left( -2X\right) ^{3/2}.\label{49}$$ has been met. The value $X=X_{0}$ at which the latter happens corresponds to $\xi_{5}=0$, so one can find the constants $C_{2}=\sqrt{-2X_{0}}$ and $C_{5}=V_{5}\left( X_{0}\right) $. This calculation results in the function $V_{5}\left( X\right) $ shown in Fig. 1 by the solid line with $X_{0}=-0.2254$, $C_{2}=0.6714$, $C_{5}=-0.9262$. Also shown in Fig. 1 is the dependence $V_{5}\left( X\right) $ in the range $X\geq X_{0}$, described by Eq. (\[47\]); the dash-dotted line. The dashed line represents data obtained when the numerical calculations described in the previous paragraph are not stopped when Eq. (\[49\]) has been satisfied but rather continue until a pole has been encountered; a procedure identical to the one employed in [@Franklin1970]. These data refer to the transition layer separating active plasma and a collisionless sheath. The dotted lines represent asymptotic behavior described by the first two terms on the rhs of Eq. (\[16\]) and by expressions $V_{5}=0.6714$ and $V_{5}=6\left( X-X_{1}\right) ^{-2}$, where $X_{1}$ is the position of the pole evaluated as $3^{1/5}$ times the value $2.384$ determined numerically in [@Franklin1970]. Composite asymptotic expansion {#Composite expansions} ============================== The composite expansion of the ion velocity uniformly valid in the illuminated plasma region to the double layer to the un-illuminated region is obtained by adding the expansion (\[30\]) and the first term of the expansion (\[10\]) and subtracting the common part:$$V\left( \xi,\varepsilon\right) =\frac{\xi}{1+\sqrt{1-\xi^{2}}}+\left[ 1+\varepsilon^{1/5}V_{5}\left( \xi_{5}\right) \right] \ -\left[ 1-\sqrt{2\left( 1-\xi\right) }\right] ,\label{61}$$ where the first and third terms on the rhs should be discarded for $\xi>1$. In the illuminated plasma, the second and third terms on the rhs of Eq. (\[61\]) virtually cancel and the first term is dominating as it should; in the double layer the first and third terms virtually cancel or are discarded and the second term is dominating as it should; in the un-illuminated plasma the first and third terms are discarded and $V=1+C_{2}\varepsilon^{1/5}$, again as it should be. Error of Eq. (\[61\]) is of the order of $\varepsilon^{2/5}$ from the illuminated plasma to the double layer to the un-illuminated plasma. The composite expansion of potential of the same accuracy is obtained in a similar way and reads$$\Phi\left( \xi,\varepsilon\right) =\ln\frac{1+\sqrt{1-\xi^{2}}}{2}+\left[ -\ln2-\varepsilon^{1/5}V_{5}\left( \xi_{5}\right) \right] -\left[ -\ln2+\sqrt{2\left( 1-\xi\right) }\right] .\label{61a}$$ The first-approximation composite expansion of the electric field is obtained by differentiating Eq. (\[61a\]) and reads $$-\frac{d\Phi}{d\xi}=\frac{\xi}{\sqrt{1-\xi^{2}}+1-\xi^{2}}+\varepsilon ^{-1/5}\frac{dV_{5}\left( \xi_{5}\right) }{d\xi_{5}}\ -\frac{1}{\sqrt{2\left( 1-\xi\right) }}.\label{62}$$ Relative error of this formula in the double layer is of the order of $\varepsilon^{1/5}$. A deficiency of Eq. (\[62\]) is a discontinuity at $\xi=1$, which stems from the limit for $\xi\rightarrow1-0$ of the sum of the first and third term on the rhs being non-zero (it equals $-1$). This discontinuity is of the order $\varepsilon^{1/5}$ relative to the main term and disappears in the next approximation. Imagine, for example, that the second term of the expansion of the electric field in the double layer \[i.e., the term associated with the third term of the expansion (\[33\])\] is $-\frac{1}{2}+\frac{1}{2}\tanh\left( \sqrt{\frac{C_{2}}{2}}\xi_{5}\right) $. Then the composite expansion of the electric field would be $$-\frac{d\Phi}{d\xi}=\frac{\xi}{\sqrt{1-\xi^{2}}+1-\xi^{2}}+\left[ \varepsilon^{-1/5}\frac{dV_{5}\left( \xi_{5}\right) }{d\xi_{5}}\ -\frac {1}{2}+\frac{1}{2}\tanh\left( \sqrt{\frac{C_{2}}{2}}\xi_{5}\right) \right] -\left[ \frac{1}{\sqrt{2\left( 1-\xi\right) }}-1\right] .\label{62a}$$ Relative error of this formula in the double layer would be of the order of $\varepsilon^{2/5}$ and the rhs of Eq. (\[62a\]) is continuous. In reality, Eq. (\[62a\]) is of course no more accurate than Eq. (\[62\]): while the second term of the expansion of the electric field in the double layer is absent from Eq. (\[62\]), in (\[62a\]) it is taken into account in a form which can be only qualitatively correct at best. However, Eq. (\[62a\]) gives a continuous distribution of the electric field and is therefore preferable for illustrative purposes. [99]{} R. N. Franklin, J. E. Allen, D. M. Thomas, and M. S. Benilov, Phys. Plasmas 20, 123508 (2013). A. D. R. Phelps and J. E. Allen, Proc. Roy. Soc. A 348, 221 (1976). M. S. Benilov and L. G. Benilova, J. Phys. D: Appl. Phys. 43, 345204 (12pp) (2010). D. Bohm, in The Characteristics of Electrical Discharges in Magnetic Fields, edited by A. Guthrie and R. K. Wakerling (McGraw-Hill, New York, 1949), pp. 77–86. M. van Dyke, Perturbation Methods in Fluid Mechanics (Parabolic Press, Stanford, CA, USA, 1975). J. D. Cole, Perturbation Methods in Applied Mathematics (Blaisdell, Waltham, 1968). A. H. Nayfeh, Perturbation Methods (Wiley, New York, 1973). A. H. Nayfeh, Introduction to Perturbation Techniques (Wiley, New York, 1981). J. Kevorkian and J. D. Cole, Perturbation Methods in Applied Mathematics (Springer-Verlag, New York, 1981). A. H. Nayfeh, Problems in Perturbation (Wiley, New York, 1985). K.-U. Riemann, J. Phys. D: Appl. Phys. 24, 493 (1991). R. N. Franklin, J. Phys. D: Appl. Phys. 36, R309 (2003). J. E. Allen, Plasma Sources Sci. Technol. 18, 014004 (2009). M. S. Benilov, Plasma Sources Sci. Technol. 18, 014005 (14pp) (2009). K.-U. Riemann, Plasma Sources Sci. Technol. 18, 014006 (10pp) (2009). S. H. Lam, Phys. Fluids 8, 73 (1965). R. N. Franklin and J. R. Ockendon, J. Plasma Phys. 4, 371 (1970). N. A. Almeida and M. S. Benilov, Phys. Plasmas 19, 073514 (2012). M. S. Benilov and R. N. Franklin, J. Plasma Phys. 62, 541 (1999). M. S. Benilov, Phys. Plasmas 7, 135 (2000). K.-U. Riemann, Phys. Plasmas 4, 4158 (1997). I. D. Kaganovich, Phys. Plasmas 9, 4788 (2002). K.-U. Riemann, J. Seebacher, D. D. [Tskhakaya Sr]{}, and S. Kuhn, Plasma Phys. Control. Fusion 47, 1949 (2005). H. T. Davis, Introduction to Nonlinear Differential and Integral Equations (Dover, N. Y., 1962). A. D. Polyanin and A. V. Manzhirov, Handbook of Mathematics for Engineers and Scientists (Chapman and Hall/CRC, Boca Raton, 2007). [1.eps]{} \[c\][l]{}Fig. 1. Distributions of the ion speed and electric field. Solid: numerical\ calculations, $S=2$, $\varepsilon=0.89\times10^{-4}$. Dotted: patching. [2.eps]{} \[c\][l]{}Fig. 2. Asymptotic solution for distribution of ion speed in the vicinity of the sonic\ barrier. Plane geometry. Solid + dash-dotted: double layer at the edge of illuminated\ plasma. Solid + dashed: transition layer separating active plasma and a collisionless\ sheath. Dotted: asymptotic behavior for large $\left\vert X\right\vert $. \[c\][cc]{} &\ &\ &\ & \[c\][l]{}Fig. 3. Distributions of normalized parameters in the double layer: the ion speed (a),\ the potential (b), the electric field (c), and the density of space charge (d). Solid:\ numerical calculations for $S=2$ and $\varepsilon=1.26\times10^{-3}$ (line 1) or $0.99\times10^{-5}$ (2).\ Dashed: second-order term of the asymptotic expansion describing the double layer.\ Dotted: composite expansion. $\raisebox{-0cm}{\includegraphics[ trim=0.000000in 0.000000in -0.008770in 0.000000in, height=7.4812cm, width=7.4747cm ]{4.eps}} $ \[c\][l]{}Fig. 4. Distributions of ion speed in the double layer and the un-illuminated plasma.\ Solid: numerical calculations for $S=2$ and $\varepsilon=1.26\times10^{-3}$ (line 1), $0.99\times10^{-5}$ (2),\ $0.89\times10^{-4}$ (3). Dotted: composite expansion. $\raisebox{-0cm}{\includegraphics[ trim=0.000000in 0.000000in -0.008770in 0.000000in, height=7.4812cm, width=7.6045cm ]{5.eps}} $ \[c\][l]{}Fig. 5. Ion speed in the un-illuminated plasma. Line: asymptotic solution.\ Points: numerical calculations.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we consider femtocell CR networks, where femto base stations (FBS) are deployed to greatly improve network coverage and capacity. We investigate the problem of generic data multicast in femtocell networks. We reformulate the resulting MINLP problem into a simpler form, and derive upper and lower performance bounds. Then we consider three typical connection scenarios in the femtocell network, and develop optimal and near-optimal algorithms for the three scenarios. Second, we tackle the problem of streaming scalable videos in femtocell CR networks. A framework is developed to captures the key design issues and trade-offs with a stochastic programming problem formulation. In the case of a single FBS, we develop an optimum-achieving distributed algorithm, which is shown also optimal for the case of multiple non-interfering FBS’s. In the case of interfering FBS’s, we develop a greedy algorithm that can compute near-opitmal solutions, and prove a closed-form lower bound on its performance.' author: - bibliography: - 'cr\_video\_femto.bib' - 'MyWork.bib' title: The Feasibility of Scalable Video Streaming over Femtocell Networks --- Introduction ============ Due to the use of open space as transmission medium, capacity of wireless networks are usually limited by interference. When a mobile user moves away from the base station, a considerably larger transmit power is needed to overcome attenuation, while causing interference to other users and deteriorating network capacity. To this end, femtocells provide an effective solution that brings network infrastructure closer to mobile users. A femtocell is a small (e.g., residential) cellular network, with a [femto base station]{} (FBS) connected to the owner’s broadband wireline network [@Chandrasekhar08; @Kim09; @Guvenc10]. The FBS serves approved users when they are within the coverage. Among the many benefits, femtocells are shown effective on improving network coverage and capacity [@Chandrasekhar08]. Due to reduced distance, transmit power can be greatly reduced, leading to prolonged battery life, improved signal-to-interference-plus-noise ratio (SINR), and better spatial reuse of spectrum. Femtocells have received significant interest from the wireless industry. Although highly promising, many important problems should be addressed to fully harvest their potential, such as interference mitigation, resource allocation, synchronization, and QoS provisioning [@Chandrasekhar08; @Kim09]. It is also critical for the success of this technology to support important applications such as real-time video streaming in femtocell networks. In this paper, we first investigate the problem of data multicast in femtocell networks. It is not atypical that many users may request for the same content, as often observed in wireline networks. By allowing multiple users to share the same downlink multicast transmission, significant spectrum and power savings can be achieved. In particular, we adopt [superposition coding]{} (SC) and [successive interference cancellation]{} (SIC), two well-known PHY techniques, for data multicast in femtocell networks [@Goldsmith06]. With SC, a compound signal is transmitted, consisting of multiple signals (or, layers) from different senders or from the same sender. With SIC, a strong signal can be first decoded, by treating all other signals as noise. Then the decoder will reconstruct the signal from the decoded bits, and subtract the reconstructed signal from the compound signal. The next signal will be decoded from the residual, by treating the remaining signals as noise. And so forth. A special strength of the SC with SIC approach is that it enables simultaneous unicast transmissions (e.g., many-to-one or one-to-many). It has been shown that SC with SIC is more efficient than PHY techniques with orthogonal channels [@Goldsmith06; @Li09]. We adopt SC and SIC for the unique femtocell network environment, and investigate how to enable efficient data multicast from the femtocells to multiple users. We formulate a Mixed Integer Nonlinear Programming (MINLP) problem, which is NP-hard in general. The objective is to minimize the total BS power consumption. Then we reformulate the MINLP problem into a simpler form, and derive upper and lower performance bounds. We also derive a simple heuristic scheme that assigns users to the BS’s with a greedy approach. Finally, we consider three typical connection scenarios in the femtocell network, and develop optimal and near-optimal algorithms for the three scenarios. The proposed algorithms have low computational complexity, and are shown to outperform the heuristic scheme with considerable gains. Then, we investigate the problem of video streaming in femtocell cognitive radio (CR) networks. We consider a femtocell network consisting of a [macro base station]{} (MBS) and multiple FBS’s. The femtocell network is co-located with a primary network with multiple licensed channels. This is a challenging problem due to the stringent QoS requirements of real-time videos and, on the other hand, the new dimensions of network dynamics (i.e., channel availability) and uncertainties (i.e., spectrum sensing and errors) found in CR networks. We adopt Scalable Video Coding (SVC) in our system [@Hu10JSAC; @Hu10TW]. SVC encodes a video into multiple substreams, subsets of which can be decoded to provide different quality levels for the reconstructed video [@Wien07]. Such scalability is very useful for video streaming systems, especially in CR networks, to accommodate heterogeneous channel availabilities and dynamic network conditions. We consider H.264/SVC medium grain scalable (MGS) videos, since MGS can achieve better rate-distortion performance over Fine-Granularity-Scalability (FGS), although it only has Network Abstraction Layer (NAL) unit-based granularity [@Wien07]. The unique femtocell network architecture and the scalable video allow us to develop a framework that captures the key design issues and trade-offs, and to formulate a [stochastic programming]{} problem. It has been shown that the deployment of femtocells has a significant impact on the network performance [@Chandrasekhar08]. In this paper, we examine three deployment scenarios. In the case of a single FBS, we apply [dual decomposition]{} to develop a distributed algorithm that can compute the optimal solution. In the case of multiple non-interfering FBS’s, we show that the same distributed algorithm can be used to compute optimal solutions. In the case of multiple interfering FBS’s, we develop a greedy algorithm that can compute near-optimal solutions, and prove a closed-form lower bound for its performance based on an [interference graph]{} model. The proposed algorithms are evaluated with simulations, and are shown to outperform three alternative schemes with considerable gains. The remainder of this paper is organized as follows. The related work is discussed in Section \[sec:femto\_work\]. We investigate the problem of data multicast over fenmtocell networks in Section \[sec:femto\_mcast\_sic\]. The problem of streaming multiple MGS videos in a femtocell CR network is discussed in Section \[sec:femto\_cr\_video\]. Section \[sec:femto\_conc\] concludes this paper. Background and Related Work {#sec:femto_work} =========================== Femtocells have attracted considerable interest from both industry and academia. Technical and business challenges, requirements and some preliminary solutions to femtocell networks are discussed in [@Chandrasekhar08]. Since FBS’s are distributedly located and are able to spatially reuse the same channel, considerable research efforts were made on interference analysis and mitigation [@Chandrasekhar09; @Lee10]. A distributed utility based SINR adaptation scheme was presented in [@Chandrasekhar09] to alleviate cross-tire interference at the macrocell from co-channel femtocells. Lee, Oh and Lee [@Lee10] proposed a fractional frequency reuse scheme to mitigate inter-femtocell interference. Deploying femtocells by underlaying the macrocell has been proved to significantly improve indoor coverage and system capacity. However, interference mitigation in a two-tier heterogeneous network is a challenging problem. In [@Chu11], the interference from macrocell and femtocells was mitigated by a spatial channel separation scheme with codeword-to-channel mapping. In [@Rangan10], the rate distribution in the macrocell was improved by subband partitioning and modest gains were achieved by interference cancellation. In [@Bharucha09], the interference was controlled by denying the access of femtocell base stations to protect the transmission of nearby macro base station. A novel algorithmic framework was presented in [@Madan10] for dynamic interference management to deliver QoS, fairness and high system efficiency in LTE-A femtocell networks. Requiring no modification of existing macrocells, CR was shown to achieve considerable performance improvement when applied to interference mitigation [@Cheng11]. In [@Kaimaletu11], the orthogonal time-frequency blocks and transmission opportunities were allocated based on a safe/victim classification. SIC has high potential of sending or receiving multiple signals concurrently, which improves the transmission efficiency [@Hu11GC]. In [@Li09], the authors developed MAC and routing protocols that exploit SC and SIC to enable simultaneous unicast transmissions. Sen, et al. investigated the possible throughput gains with SIC from a MAC layer perspective [@Sen10]. Power control for SIC was comprehensively investigated and widely applied to code division multiple access (CDMA) systems [@Jean09; @Park08; @Benvenuto07; @Agrawal05; @Andrews03]. Applying game theory, Jean and Jabbari proposed an uplink power control under SIC in direct sequence-CDMA networks [@Jean09]. In [@Park08], the authors introduced an iterative two-stage SIC detection scheme for a multicode MIMO system and showed the proposed scheme significantly outperformed the equal power allocation scheme. A scheme on joint power control and receiver optimization of CDMA transceivers was presented in [@Benvenuto07]. In [@Agrawal05; @Andrews03], the impact of imperfect channel estimation and imperfect interference cancellation on the capacity of CDMA systems was examined. The problem of video over CR networks has only been studied in a few recent papers [@Shiang08; @Ding09; @Hu12JSAC; @Luo11]. In our prior work, we investigated the problem of scalable video over infrastructure-based CR networks [@Hu10JSAC] and multi-hop CR networks [@Hu10TW]. The preliminary results of video over femtocell CR networks were presented in [@Hu11IDS]. Multicast in Femtocell Networks with Superposition Coding and Successive Interference Cancellation {#sec:femto_mcast_sic} ================================================================================================== In this section, we formulate a Mixed Integer Nonlinear Programming (MINLP) problem of data multicast in femotcell networks, which is NP-hard in general. Then we reformulate the MINLP problem into a simpler form, and derive upper and lower performance bounds. We also derive a simple heuristic scheme that assigns users to the BS’s with a greedy approach. Finally, we consider three typical connection scenarios in the femtocell network, and develop optimal and near-optimal algorithms for the three scenarios. The proposed algorithms have low computational complexity, and are shown to outperform the heuristic scheme with considerable gains. System Model and Problem Statement \[sec:mod3\] ----------------------------------------------- ### System Model Consider a femtocell network with an MBS (indexed $0$) and $M$ FBS’s (indexed from $1$ to $M$) deployed in the area. The $M$ FBS’s are connected to the MBS and the Internet via broadband wireline connections. Furthermore, we assume a spectrum band that is divided into two parts, one is allocated to the MBS with bandwidth $B_0$ and the other is allocated to the $M$ FBS’s. The bandwidth allocated to FBS $m$ is denoted by $B_m$. When there is no overlap between the coverages of two FBS’s, they can spatially reuse the same spectrum. Otherwise, the MBS allocates disjoint spectrum to the FBS’s with overlapping coverages. We assumed the spectrum allocation is known a priori. There are $K$ mobile users in the femtocell network. Each user is equipped with one transceiver that can be tuned to one of the two available channels, i.e., connecting to a nearby FBS or to the MBS. The network is time slotted. We assume block-fading channels, where the channel condition is constant in each time slot [@Goldsmith06]. We focus on a multicast scenario, where the MBS and FBS’s multicast a data file to the $K$ users. The data file is divided into multiple packets with equal length and transmitted in sequence with the same modulation scheme. Once packet $l$ is successfully received and decoded at the user, it requests packet $(l+1)$ in the next time slot. We adopt SC and SIC to transmit these packets [@Goldsmith06], as illustrated in Fig. \[fig:sic\]. In each time slot $t$, the compound signal has $L$ [layers]{} (or, levels), denoted as $D_1(t)$, $\cdots$, $D_L(t)$. Each level $D_i(t)$, $i=1, \cdots, L$, is a packet requested by some of the users in time slot $t$. A user that has successfully decoded $D_i(t)$, for all $i=1$, $\cdots$, $l-1$, is able to subtract these signals from the received compound signal and then decodes $D_l(t)$, while the signals from $D_{l+1}(t)$ to $D_L(t)$ are treated as noise. ### Problem Statement For the SC and SIC scheme to work, the transmit powers for the levels should be carefully determined, such that there is a sufficiently high SNR for the levels to be decodable. It is also important to control the transmit powers of the BS’s to reduce interference and leverage frequency reuse. The annual power bill is a large part of a mobile operator’s costs [@Ulf10]. Minimizing BS power consumption is important to reduce not only the operator’s OPEX, but also the global CO$_2$ emission; an important step towards “green” communications. ![Superposition coding and successive interference cancellation.[]{data-label="fig:sic"}](superposition-coding.eps){width="4.5in"} Therefore, we focus on BS power allocation in this paper. The objective is to minimize the total power of all the BS’s, while guaranteeing a target rate $R_{tar}$ for each user. Recall that the data file is partitioned into equal-length packets. The target rate $R_{tar}$ ensures that a packet can be transmitted within a time slot, for given modulation and channel coding schemes. Define binary indicator $I_m^k$, for all $m$ and $k$, as: $$\label{eq:imk} I_m^k = \left\{ \begin{array}{ll} 1, & \mbox{if user $k$ connects to BS $m$} \\ 0, & \mbox{otherwise.} \end{array} \right.$$ Consider a general time slot $t$ when $L$ data packets, or levels, are requested. We formulate the optimal power allocation problem (termed OPT-Power) as follows. $$\begin{aligned} \mbox{minimize:} && \hspace{-0.2in} \sum_{m=0}^M \sum_{l=1}^L P_l^m \label{eq:ObjFun11} \\ \mbox{subject to:} && \hspace{-0.2in} B_m\log_2(1+\gamma_m^k I_m^k) \ge R_{tar}I_m^k, \mbox{ for all } k \label{eq:cntrate} \\ && \hspace{-0.2in} \sum_{m=0}^M I_m^k=1, \mbox{for all } k \label{eq:cnttransceiver} \\ && \hspace{-0.2in} P_l^m \ge 0, \mbox{ for all } l, m,\end{aligned}$$ where $P_l^m$ is the power of BS $m$ for transmitting the level $l$ packet; $\gamma_m^k$ is the SNR at user $k$ if it connects to BS $m$. Constraint (\[eq:cntrate\]) guarantees the minimum rate at each user. Constraint (\[eq:cnttransceiver\]) is due to the fact that each user is equipped with one transceiver, so it can only connect to one BS. Let $\mathcal{U}_l$ denote the set of users requesting the level $l$ packet. A user $k \in \mathcal{U}_l$ has decoded all the packets up to $D_{l-1}$. It subtracts the decoded signals from the received signal and treats signals $D_{l+1}, \cdots, D_L$ as noise. The SNR at user $k \in \mathcal{U}_l$, for $l=1, \cdots, L-1$, can be written as: $$\begin{aligned} \label{eq:SNR1} %\gamma_m^k=\frac{H_m^kP_l^m}{N_0+H_m^k\sum_{i=l+1}^LP_i^m} \gamma_m^k = H_m^k P_l^m / \left(N_0 + H_m^k \sum_{i=l+1}^L P_i^m \right),\end{aligned}$$ where $H_m^k$ is the random channel gain from BS $m$ to user $k$ and $N_0$ is the noise power. For user $k \in \mathcal{U}_L$ that requests the last packet $D_L$, the SNR is $$\begin{aligned} \label{eq:SNR2} %\gamma_m^k=\frac{H_m^kP_L^m}{N_0} \gamma_m^k = H_m^k P_L^m / N_0.\end{aligned}$$ The optimization variables in Problem OPT-Power consist of the binary variables $I_m^k$’s and the continuous variables $P_l^m$’s. It is an MINLP problem, which is NP-hard in general. In Section \[sec:alg\], we first reformulate the problem to a obtain a simpler form, and then develop effective algorithms for optimal and suboptimal solutions. Reformulation and Power Allocation \[sec:alg\] ---------------------------------------------- In this section, we reformulate Problem OPT-Power to obtain a simpler form, and derive an upper bound and a lower bound for the total BS power. The reformulation also leads to a simple heuristic algorithm. Finally, we introduce power allocation algorithms for three connection scenarios. ### Problem Reformulation Due to the monotonic logarithm functions and the binary indicators $I_m^k$, constraint (\[eq:cntrate\]) can be rewritten as: $$\label{eq:StSNR} \gamma_m^k I_m^k \ge \Gamma_m^k I_m^k, \;\; m=0, 1, \cdots,M,$$ where $\Gamma_m^k = \Gamma_m =: 2^{R_{tar}/B_m} - 1$ is the minimum SNR requirement at user $k$ that connects to BS $m$. To further simplify the problem, define $Q_l^m = \sum_{i=l}^L P_i^m$, with $Q_{L+1}^m=0$. Then power $P_l^m$ is the difference $$\begin{aligned} \label{eq:QrepP} P_l^m=Q_l^m-Q_{l+1}^m.\end{aligned}$$ Problem OPT-Power can be reformulated as: $$\begin{aligned} \mbox{minimize} && \hspace{-0.2in} \sum_{m=0}^M Q_1^m \label{eq:ObjFun2} \\ %\mbox{subject to:} && \hspace{-0.2in} \frac{H_m^k(Q_l^m-Q_{l+1}^m)}{N_0+H_m^kQ_{l+1}^m}I_m^k\ge \Gamma_m I_m^k, \nonumber \\ \mbox{subject to:} && \hspace{-0.2in} H_m^k(Q_l^m \hspace{-0.025in} - \hspace{-0.025in} Q_{l+1}^m) / \hspace{-0.025in} \left( N_0 \hspace{-0.025in} + \hspace{-0.025in} H_m^k Q_{l+1}^m \right) I_m^k \ge \Gamma_m I_m^k, \nonumber \\ && \hspace{0.7in} \mbox{ for all } k \in \mathcal{U}_l, l=1, \cdots, L \label{eq:cntrate2} \\ && \hspace{-0.2in} Q_l^m \ge Q_{l+1}^m, l=1, \cdots, L \\ && \hspace{-0.2in} \sum_{m=0}^M I_m^k=1, \mbox{ for all } k. \end{aligned}$$ For $l\le L$, constraint (\[eq:cntrate2\]) can be rewritten as: $$\begin{aligned} \label{eq:QmlIneq} Q_l^m I_m^k \ge \left[ N_0 \Gamma_m / H_m^k + (1 + \Gamma_m) Q_{l+1}^m \right] I_m^k.\end{aligned}$$ Let $\mathcal{U}_l^m$ be the subset of users connecting to BS $m$ in $\mathcal{U}_l$. Since $Q_l^m \ge Q_{l+1}^m$, (\[eq:QmlIneq\]) can be rewritten as, $$\label{eq:QmlEqu} Q_l^m = \max \left\{ Q_{l+1}^m, \max_{k \in \mathcal{U}_l^m} \left[ N_0 \Gamma_m / H_m^k + (1 + \Gamma_m) Q_{l+1}^m \right] \right\}. % Q_l^m = \max \left\{ Q_{l+1}^m, \max_{k \in \mathcal{U}_l^m} \left[ N_0 \frac{\Gamma_m}{H_m^k} + (1 + \Gamma_m) Q_{l+1}^m \right] \right\}.$$ From (\[eq:QmlEqu\]), we define a function $Q_l^m = F_m(Q_{l+1}^m,\mathcal{U}_l^m)$ as: $$\begin{aligned} \label{eq:FmDef} %&& \hspace{-0.1in} F_m(Q_{l+1}^m,\mathcal{U}_l^m) %\label{eq:FmDef} \\ %&=&\hspace{-0.1in} = \left\{\begin{array}{l l} Q_{l+1}^m,& \mathcal{U}_l^m=\emptyset \\ \max_{k\in\mathcal{U}_l^m} \left\{ \frac{N_0\Gamma_m}{H_m^k}+(1+\Gamma_m)Q_{l+1}^m \right\}, & \hspace{-0.05in} \mathcal{U}_l^m \neq \emptyset. \end{array}\right. %\nonumber\end{aligned}$$ Obviously, $F_m(Q_{l+1}^m,\mathcal{U}_l^m)$ is non-decreasing with respect to $Q_{l+1}^m$. It follows that $$\begin{aligned} \label{eq:ObjFun3} Q_1^m &=& F_m(Q_2^m, \mathcal{U}_1^m) \; = \; F_m(F_m(Q_3^m,\mathcal{U}_2^m),\mathcal{U}_1^m) \nonumber \\ &=& F_m(\cdots (F_m(Q_{L+1}^m, \mathcal{U}_L^m), \mathcal{U}_{L-1}^m), \cdots, \mathcal{U}_1^m) \nonumber \\ &=& F_m(\cdots (F_m(0, \mathcal{U}_L^m), \mathcal{U}_{L-1}^m), \cdots, \mathcal{U}_1^m).\end{aligned}$$ If none of the subsets $\mathcal{U}_l^m$ ($l=1,\cdots,L$) is empty, we can expand the above recursive term using (\[eq:FmDef\]). It follows that $$\label{eq:FoldTerm} Q_1^m = N_0 \Gamma_m \sum_{l=1}^L (1 + \Gamma_m)^{c_l^m} \max_{k \in \mathcal{U}_l^m} \left\{1 / H_m^k \right\},$$ where the exponent $c_l^m$ is defined as $c_1^m=0$ and $c_{l+1}^m=c_l^m+1$. Otherwise, if a subset $\mathcal{U}_l^m = \emptyset$ for some $m$, we have that $Q_l^m=Q_{l+1}^m$, $\max_{k\in\mathcal{U}_l^m} \left\{1/H_m^k\right\}=\max_{k\in \emptyset} \left\{1/H_m^k\right\}=0$, and $c_l^m=c_{l-1}^m$. Eq. (\[eq:FoldTerm\]) still holds true. Finally, the objective function (\[eq:ObjFun2\]) can be rewritten as $$% \mbox{$\sum_{m=0}^M$} Q_1^m = \mbox{$\sum_{m=0}^M$} N_0 \Gamma_m \mbox{$\sum_{l=1}^L$} (1 + \Gamma_m)^{c_l^m} \max_{k \in \mathcal{U}_l^m} \left\{1 / H_m^k \right\}. \sum_{m=0}^M N_0 \Gamma_m \sum_{l=1}^L (1 + \Gamma_m)^{c_l^m} \max_{k \in \mathcal{U}_l^m} \left\{1 / H_m^k \right\}.$$ Since $(1+\Gamma_m)>0$, it can be seen that to minimize the total BS power, we need to keep the $c_l^m$’s as low as possible. ### Performance Bounds \[subsec:bounds\] The reformulation and simplification allow us to derive performance bounds for the total BS power consumption. First, we derive the upper bound for the objective function (\[eq:ObjFun2\]). Define a variable $$% \overline{G}_m = \max_{l\in\{1,\cdots,L\}} \left\{ \max_{k\in\mathcal{U}_l^m} \left\{\Gamma_m/H_m^k\right\} \right\}. \overline{G}_m = \max_{l\in\{1,\cdots,L\}} \max_{k\in\mathcal{U}_l^m} \left\{\Gamma_m/H_m^k\right\},$$ which corresponds to the user with the worst channel condition among all users that connect to BS $m$. It follows that: $$\begin{aligned} \sum_{m=0}^M Q_1^m &\hspace{-0.1in} =& \hspace{-0.1in} N_0 \sum_{m=0}^M \sum_{l=1}^L (1+\Gamma_m)^{c_l^m}\max_{k\in\mathcal{U}_l^m}\left\{\Gamma_m / H_m^k\right\} \nonumber \\ &\hspace{-0.1in} \le& \hspace{-0.1in} N_0 \sum_{m=0}^M \sum_{l=1}^L (1+\Gamma_m)^{c_l^m} \overline{G}_m \nonumber \\ &\hspace{-0.1in} \le& \hspace{-0.1in} N_0 \sum_{m=0}^M \overline{G}_m \sum_{l=1}^L (1+\Gamma_m)^{l-1} \nonumber \\ &\hspace{-0.1in} =& \hspace{-0.1in} N_0 \sum_{m=0}^M \overline{G}_m \left[ (1+\Gamma_m)^L-1 \right] / \Gamma_m. \label{eq:UBound}\end{aligned}$$ In (\[eq:UBound\]), the first inequality is from the definition of $\overline{G}_m$. The second inequality is from the definition of $c_{l+1}^m$. Specifically, $c_1^m=0$; when $\mathcal{U}_l^m \neq \emptyset$, we have $c_{l}^m=c_{l-1}^m+1$; when $\mathcal{U}_l^m = \emptyset$, we have $c_{l}^m=c_{l-1}^m$. It follows that $c_l^m\le l-1$. Therefore, (\[eq:UBound\]) is an upper bound on the objective function (\[eq:ObjFun2\]). Furthermore, by defining $\overline{G} = \max_{m\in\{0,\cdots,M\}} \left\{ \overline{G}_m \right\}$, and $\overline{\Gamma} = \max_{m\in\{0,\cdots,M\}} \left\{ \Gamma_m \right\}$, we can get a looser upper bound from \[eq:UBound\] as $$\sum_{m=0}^M Q_1^m \leq N_0 \overline{G} (M+1)\left[ (1+\overline{\Gamma})^L-1\right]/\overline{\Gamma}.$$ Next, we derive a lower bound for (\[eq:ObjFun2\]). Define $$\left\{ \begin{array}{l} \underline{G}^l = \min_{m\in\{0,\cdots,M\}} \max_{k\in\mathcal{U}_l^m} \left\{ \Gamma_m/ H_m^k \right\} \\ \underline{\Gamma} = \min_{m\in\{0,\cdots,M\}} \left\{ \Gamma_m \right\}. \end{array} \right.$$ We have that $$\begin{aligned} \sum_{m=0}^M Q_1^m &\hspace{-0.1in} =& \hspace{-0.1in} N_0 \sum_{m=0}^M \sum_{l=1}^L (1+\Gamma_m)^{c_l^m} \max_{k\in\mathcal{U}_l^m} \left\{ \Gamma_m / H_m^k \right\} \nonumber \\ &\hspace{-0.1in} \ge& \hspace{-0.1in} N_0 \sum_{m=0}^M \sum_{l=1}^L (1+\Gamma_m)^{c_l^m} \underline{G}^l \nonumber \\ &\hspace{-0.1in} \ge& \hspace{-0.1in} N_0 \sum_{l=1}^L \underline{G}^l \sum_{m=0}^M (1+\underline{\Gamma})^{c_l^m} \nonumber \\ &\hspace{-0.1in} \ge& \hspace{-0.1in} N_0 (M+1) \sum_{l=1}^L \underline{G}^l (1+\underline{\Gamma})^{\frac{\sum_{m=0}^Mc_l^m}{M+1}}\nonumber \\ &\hspace{-0.1in} \ge& \hspace{-0.1in} N_0 (M+1) \sum_{l=1}^L \underline{G}^l(1+\underline{\Gamma})^{\frac{l-1}{M+1}}. \label{eq:LBound}\end{aligned}$$ In (\[eq:LBound\]), the first inequality is from the definition of $\underline{G}^l$. The second inequality is due to the definition of $\underline{\Gamma}$. The third inequality is due to the fact that $(1+\Gamma)^{c_l^m}$ is a convex function. The fourth inequality is because that each level must be transmitted by at least one BS. Thus for each level $l$, there is at least one $c_l^m=c_{l-1}^m+1$ for some $m$. It follows that the sum $\sum_{m=0}^M c_l^m$ should be greater than $l-1$. Therefore, (\[eq:LBound\]) provides a lower bound for (\[eq:ObjFun2\]). Furthermore, by defining $\underline{G} = \min_{l\in\{1,\cdots,L\}} \left\{ \underline{G}^l \right\}$, we can obtain a looser lower bound from (\[eq:LBound\]) as $$\sum_{m=0}^M Q_1^m \geq N_0 \underline{G} (M+1) \frac{(1+\underline{\Gamma})^{\frac{L}{M+1}}-1}{(1+\underline{\Gamma})^{\frac{1}{M+1}}-1}.$$ ### A Simple Heuristic Scheme \[subsec:heuristic\] We first describe a greedy heuristic algorithm that solves OPT-Power with suboptimal solutions. With this heuristic, each user compares the channel gains from the MBS and the FBS’s. It chooses the BS with the best channel condition to connect to, thus the values of the binary variables $I_m^k$ are determined. Once the binary variables are fixed, all the subsets $\mathcal{U}_l^m$’s are determined. Starting with $Q_{L+1}^m=0$, we can apply (\[eq:QmlEqu\]) iteratively to find the $Q_l^m$’s. Finally, the transmit powers $P_l^m$ can be computed using (\[eq:QrepP\]). With this approach, among the users requesting the level $l$ packet, it is more likely that some of them connect to the MBS and the rest connect to some FBS’s, due to the random channel gains in each time slot. In this situation, both MBS and FBS will have to transmit all the requested data packets. Such situation is not optimal for minimizing the total power, as will be discussed in Section \[subsubsec:caseII\]. ### Power Allocation Algorithms \[subsec:proposed\] In the following, we develop three power allocation algorithms for three different connection scenarios with a more structured approach. #### Case I–One Base Station We first consider the simplest connection scenario where all the $K$ users connect to the same BS (i.e., either the MBS or an FBS). Assume all the users connect to BS $m$. Then we have $I_m^k=1$ for all $k$, and all the subsets $\mathcal{U}_l^m$ are non-empty; $I_{m'}^k=0$ for all $k$ and all $m' \neq m$, and all the subsets $\mathcal{U}_l^{m'}$ are empty for $m' \neq m$. From (\[eq:FmDef\]), we can derive the optimal solution as: $$\begin{aligned} \label{eq:OptCase1} Q_l^{m\ast} &=& (1+\Gamma_m) Q_{l+1}^{m\ast} + \max_{k \in \mathcal{U}_l^m} \left\{ N_0\Gamma_m / H_m^k \right\}, \nonumber \\ &=& N_0 \Gamma_m \sum_{i=l}^L (1+\Gamma_m)^{i-l} \max_{k \in \mathcal{U}_l^m} \left\{1/H_m^k\right\}, %\nonumber \\ %&& \hspace{1.3in} \;\; l=1,2, \cdots,L. %\\ %Q_{L+1}^{m\ast} &=& 0,\end{aligned}$$ Recall that $Q_{L+1}^{m\ast} = Q_{L+1}^{m} = 0$, the optimal power allocation for Problem OPT-Power in this case is: $$% P_l^{i\ast}=Q_l^{i\ast}-Q_{l+1}^{i\ast}, i=m, \mbox{ for all } l. P_l^{m'\ast}=\left\{ \begin{array}{ll} Q_l^{m\ast}-Q_{l+1}^{m\ast}, & m'=m, \mbox{ for all } l \\ 0, & m' \neq m, \mbox{ for all } l. \end{array} \right.$$ #### Case II–MBS and One FBS \[subsubsec:caseII\] We next consider the case with one MBS and one FBS (i.e., $M=1$), where each user has two choices: connecting to either the FBS or the MBS. Recall that $\mathcal{U}_l^0$ and $\mathcal{U}_l^1$ are the subset of users who connected to the MBS and the FBS, respectively, and who request the level $l$ packet. Examining (\[eq:FoldTerm\]), we find that the total power of BS $m$ can be significantly reduced if one or more levels are not transmitted, since the exponent $c_l^m$ will not be increased in this case. Furthermore, consider the two choices: (i) not transmitting level $l$, and (ii) not transmitting level $l'>l$ from BS $m$. The first choice will yield larger power savings, since more exponents (i.e., $c_l^m, c_{l+1}^m, \cdots, c_{l'-1}^m$) will assume smaller values. Therefore, we should let these two subsets be empty whenever possible, i.e., either $\mathcal{U}_l^0=\emptyset$ or $\mathcal{U}_l^1=\emptyset$. According to this policy, all the users requesting the level $l$ packet will connect to the same BS. We only need to make the optimal connection decision for each subset of users requesting the same level of packet, rather than for each individual user. Since not transmitting a lower level packet yields more power savings for a BS, we calculate the power from the lowest to the highest level, and decide whether connecting to the MBS or the FBS for users in each level. Define $G_l^0 = \max_{k\in\mathcal{U}_l} \left\{ 1/H_0^k \right\}$ and $G_l^1 = \max_{k\in\mathcal{U}_l} \left\{ 1/H_1^k \right\}$. The algorithm for solving Problem OPT-Power in this case is given in Table \[tab:Case2Algo\]. In Steps $2$–$10$, the decision on whether connecting to the MBS or the FBS is made by comparing the expected increments in the total power. The user subsets $\mathcal{U}_l^0$ and $\mathcal{U}_l^1$ are determined in Steps $4$ and $7$. In Steps $11$–$14$, $Q_l^m$’s and the corresponding $P_l^m$’s are computed in the reverse order, based on the determined subsets $\mathcal{U}_l^0$ and $\mathcal{U}_l^1$. The computational complexity of this algorithm is $\mathcal{O}(L)$. ----- --------------------------------------------------------------------------------------- 1: Initialize all $c_l^0$, $c_l^1$, $Q_{L+1}^0$ and $Q_{L+1}^1$ to zero; 2: FOR $l=1$ TO $L$ 3: $\;\;$ IF ($\Gamma_0(1+\Gamma_0)^{c_l^0}G_l^0 \le \Gamma_1(1+\Gamma_1)^{c_l^1}G_l^0$) 4: $\;\;\;\;$ Set $\mathcal{U}_l^0=\mathcal{U}_l$ and $\mathcal{U}_l^1=\emptyset$; 5: $\;\;\;\;$ $c_l^0=c_l^0+1$; 6: $\;\;$ ELSE 7: $\;\;\;\;$ Set $\mathcal{U}_l^0=\emptyset$ and $\mathcal{U}_l^1=\mathcal{U}_l$; 8: $\;\;\;\;$ $c_l^1=c_l^1+1$; 9: $\;\;$ END IF 10: END FOR 11: FOR $l=L$ TO $1$ 12: $\;\;$ $Q_l^0=F_0(Q_{l+1}^0,\mathcal{U}_l^0)$ and $P_l^0=Q_l^0-Q_{l+1}^0$; 13: $\;\;$ $Q_l^1=F_1(Q_{l+1}^1,\mathcal{U}_l^1)$ and $P_l^1=Q_l^1-Q_{l+1}^1$; 14: END FOR ----- --------------------------------------------------------------------------------------- : Power Allocation Algorithm For Case II \[tab:Case2Algo\] #### Case III–MBS and Multiple FBS’s \[subsubsec:case3\] Finally, we consider the general case with one MBS and multiple FBS’s in the network. Each user is able to connect to the MBS or a nearby FBS. Recall that we define $\mathcal{U}_l$ as the set of users requesting the level $l$ packet, and $\mathcal{U}_l^m$ as the subset of users in $\mathcal{U}_l$ that [connect]{} to BS $m$. These sets have the following properties. $$\label{eq:SetProp1} \left\{ \begin{array}{l} \bigcup_{m=0}^M \mathcal{U}_l^m = \mathcal{U}_l \nonumber \\ \mathcal{U}_l^m \bigcap \mathcal{U}_l^{m'} = \emptyset, \; \mbox{ for all } m' \neq m. \nonumber \end{array} \right.$$ The first property is due to the fact that each user must connect to the MBS or an FBS. The second property is because each user can connect to only one BS. The user subsets connecting to different BS’s do not overlap. Therefore, $\mathcal{U}_l^m$’s is a [partition]{} of $\mathcal{U}_l$ with respect to $m$. In addition, we define $\mathcal{S}_l^m$ as the set of possible users that are [covered]{} by BS $m$ and request the level $l$ packet. These sets have the following properties. $$\label{eq:SetProp2} \left\{ \begin{array}{l} \bigcup_{m=1}^M \mathcal{S}_l^m = \mathcal{S}_l^0=\mathcal{U}_l \nonumber \\ \mathcal{S}_l^m \bigcap \mathcal{S}_l^0 = \mathcal{S}_l^m, \; \mbox{ for all } m \neq 0 \nonumber \\ \mathcal{S}_l^m \bigcap \mathcal{S}_l^{m'} = \emptyset, \; \mbox{ for all } m' \neq m \mbox{ and } m, m'\neq 0. \nonumber \end{array} \right.$$ The first property is because all users in each femtocell are covered by the MBS. The second property indicates that the users covered by FBS $m$ are a subset of the users covered by the MBS. The third property shows that the user subsets in different femtocells do not overlap. We can see that the $\mathcal{S}_l^m$’s, for $m=1,\cdots,M$, are also a partition of $\mathcal{U}_l$. Define $W_m(\mathcal{U})=\max_{k\in\mathcal{U}}\left\{ 1/H_m^k \right\}$, where $\mathcal{U}$ is the set of users and $m=0,\cdots,M$. If the set $\mathcal{U}$ is empty, we define $W_m(\emptyset)=0$. For example, consider Case II where $M=1$. We have $\mathcal{S}_l^0 = \mathcal{S}_l^1 = \mathcal{U}_l$, $W_0(\mathcal{U}_l)=G_l^0$, and $W_1(\mathcal{U}_l)=G_l^1$. The power allocation algorithm for Case III is presented in Table \[tab:Case3Algo\]. The algorithm iteratively picks users from the [eligible]{} subset $\mathcal{S}_l^m$ and assigns them to the [allocated]{} subset $\mathcal{U}_l^m$. In each step $l$, $\Psi$ is the subset of FBS’s that will transmit the level $l$ packet; the complementary set $\overline{\Psi}$ is the subset of FBS’s that will not transmit the level $l$ packet. The expected increment in total power for each partition is computed, and the partition with the smallest expected increment will be chosen. $\Delta_l^m$ is the power of BS $m$ for transmitting the level $l$ data packet. In Steps $6$–$15$, the MBS and FBS combination $\Psi$ is determined for transmitting the level $l$ packet, with the lowest power $\Delta_0$. In Steps $16$–$30$, elements in $\mathcal{S}_l^m$ are assigned to $\mathcal{U}_l^m$ according to $\Psi$. In Steps $31$–$35$, power sums $Q_l^m$ and the corresponding power allocations $P_l^m$ are calculated in the reverse order from the known $\mathcal{U}_l^m$’s. The complexity of the algorithm is $\mathcal{O}(ML)$. ----- ------------------------------------------------------------------------------------------------------------------------ 1: Initialize: $c_l^m=0$ and $Q_{L+1}^m=0$, for all $l$, $m$; 2: FOR $l=1$ TO $L$ 3: $\;\;\;$ FOR $m=0$ TO $M$ 4: $\;\;\;\;\;\;$ $\Delta_l^m=\Gamma_m(1+\Gamma_m)^{c_l^m}W_m(\mathcal{S}_l^m)$; 5: $\;\;\;$ END FOR 6: $\;\;\;$ Set $\Omega=\{1,\cdots,M\}$ and $\Psi=\emptyset$; 7: $\;\;\;$ WHILE ($\Omega \neq \emptyset$) 8: $\;\;\;\;\;\;$ $m'=\arg\min_{m\in\Omega} \Delta_l^m$; 9: $\;\;\;\;\;\;$ Compute $\Delta'=\Gamma_0(1+\Gamma_0)^{c_l^0}W_0(\bigcup_{m\in\overline{\Psi\cup m'}}\mathcal{S}_l^m)$; 10: $\;\;\;\;\;\;$ IF ($(\sum_{m\in\Psi\cup m'}\Delta_l^m + \Delta') < \Delta_0$) 11: $\;\;\;\;\;\;\;\;\;$ Add $m'$ to $\Psi$; 12: $\;\;\;\;\;\;\;\;\;$ $\Delta_0=\sum_{m\in\Psi}\Delta_l^m + \Delta'$; 13: $\;\;\;\;\;\;$ END IF 14: $\;\;\;\;\;\;$ Remove $m'$ from $\Omega$; 15: $\;\;\;$ END WHILE 16: $\;\;\;$ IF ($\Psi = \emptyset$) 17: $\;\;\;\;\;\;$ $\mathcal{U}_l^0=\mathcal{S}_l^0$; 18: $\;\;\;\;\;\;$ $c_l^0=c_l^0+1$; 19: $\;\;\;\;\;\;$ Set $\mathcal{U}_l^m = \emptyset$, for all $m \neq 0$; 20: $\;\;\;$ ELSE 21: $\;\;\;\;\;\;$ $\mathcal{U}_l^0=\bigcup_{m\in\overline{\Psi}}\mathcal{S}_l^m$; 22: $\;\;\;\;\;\;$ IF ($\|\Psi\|<M$) 23: $\;\;\;\;\;\;\;\;\;$ $c_l^0=c_l^0+1$; 24: $\;\;\;\;\;\;$ END IF 25: $\;\;\;\;\;\;$ FOR $m \in \Psi$ 26: $\;\;\;\;\;\;\;\;\;$ $c_l^m=c_l^m+1$; 27: $\;\;\;\;\;\;\;\;\;$ $\mathcal{U}_l^m=\mathcal{S}_l^m$; 28: $\;\;\;\;\;\;$ END FOR 29: $\;\;\;$ END IF 30: END FOR 31: FOR $l=L$ TO $1$ 32: $\;\;\;$ FOR $m=0$ TO $M$ 33: $\;\;\;\;\;\;$ $Q_l^m=F_m(Q_{l+1}^m,\mathcal{U}_l^m)$ and $P_l^m=Q_l^m-Q_{l+1}^m$; 34: $\;\;\;$ END FOR 35: END FOR ----- ------------------------------------------------------------------------------------------------------------------------ : Power Allocation Algorithm For Case III \[tab:Case3Algo\] Performance Evaluation \[sec:sim3\] ----------------------------------- We evaluate the performance of the proposed power allocation algorithms using MATLAB^TM^. Three scenarios corresponding to the three cases in Section \[sec:alg\] are simulated: (i) Case I: a single MBS; (ii) Case II: one MBS and one FBS; and (iii) Case III: one MBS and three FBS’s. Since we do not find any similar schemes in the literature, we made the following comparisons. First, we compare Cases I and II with respect to BS power consumption and interference footprint. In both cases, there are $K=8$ users and $L=4$ levels. In Case I, the MBS bandwidth is $B_0=2$ MHz. In Case II, the MBS and the FBS share the $2$ MHz total bandwidth; the MBS bandwidth is $B_0=1$ MHz and the FBS bandwidth is $B_1=1$ MHz. The target data rate $R_{tar}$ is set to $2$ Mbps. The channel gain from a base station to each user is exponentially distributed in each time slot. The interference footprints in the three dimensional space are plotted in Fig. \[fig:Case1VSCase2\]. The height $B$ of the cylinders indicates the spectrum used by a BS, while the radius $r$ is proportional to the BS transmit power. In Case I when only the MBS is used, the total BS power is $45.71$ dBm and the volume of the cylinder is $\pi r^2 B = 18,841$ MHz m$^2$. In Case II when both the MBS and FBS are used, the total BS power is $34.58$ dBm and the total volume of the two cylinders is $2,378$ MHz m$^2$. Using an additional FBS achieves a $11.13$ dB power saving and the interference footprint is reduced to $12.62$% of that in Case I. This simple comparison clearly demonstrate the advantages of femtocells achieved by bringing BS’s closer to users. \[!t\] ![Case I vs. Case II: interference footprints.[]{data-label="fig:Case1VSCase2"}](PlotCase1VSCase2_br.eps "fig:"){width="4.5in" height="3.0in"} We next consider the more general Case III, using a femtocell network of one MBS and three FBS’s. The MBS bandwidth is $B_0=1$ MHz and each FBS is assigned with bandwidth $B_m=1$ MHz, $m=1, 2, 3$. The target data rate is still $2$ Mbps. In Figs. \[fig:Case3Level\] and \[fig:Case3BW0\], we plot four curves, each obtained with: (i) the heuristic scheme described in Section \[subsec:heuristic\]; (ii) The proposed algorithm presented in Section \[subsubsec:case3\]; (iii) The upper bound; and (iv) the lower bound derived in Section \[subsec:bounds\]. Each point in the figures is the average of $10$ simulation runs. The $95\%$ confidence intervals are plotted as error bars, which are all negligible. In Fig. \[fig:Case3Level\], we examine the impact of the number of packet levels $L$ on the total BS transmit power. We increase $L$ from $2$ to $6$, and plot the total power of base stations. As expected, the more packet levels, the larger the BS power consumption. Both the proposed and heuristic curves lie in between the upper and lower bound curves. When $L$ is increased from $2$ to $6$, the power consumption of the heuristic scheme is increased by $12.22$ dB, while the power consumption of the proposed algorithm is increased by $9.94$ dB. The power savings achieved by the proposed algorithm over the heuristic scheme range from $3.92$ dB to $6.45$ dB. ![Case III: impact of number of levels $L$.[]{data-label="fig:Case3Level"}](PlotCase3Level_largefont.eps){width="4.5in" height="3.0in"} In Fig. \[fig:Case3BW0\], we show the impact of the BS bandwidths. The number of levels is $L=4$. We fix the total bandwidth at $2$ MHz, which is shared by the MBS and FBS’s. We increase the MBS bandwidth from $0.4$ MHz to $1.6$ MHz in steps of $0.2$ MHz, while decrease the bandwidth of FBS’s from $1.6$ MHz to $0.4$ MHz. We find that the total power consumption is increased as $B_0$ gets large. This is due to the fact that as the FBS bandwidth gets smaller, the FBS’s have to spend more power to meet the minimum data rate requirement. The curve produced by the proposed algorithm has a smaller slop than that of the heuristic scheme: the overall increase in the total power of the proposed algorithm is $4.86$ dB, while that of the heuristic scheme is $20.84$ dB. This implies that the proposed scheme is not very sensitive to the bandwidth allocation between the MBS and FBS’s. The proposed algorithm also achieves consider power savings over the heuristic scheme. When $B_0=1.6$ MHz, the total power of the proposed algorithm is $20.75$ dB lower than that of the heuristic scheme. ![Case III: impact of MBS bandwidth $B_0$.[]{data-label="fig:Case3BW0"}](PlotCase3BW0_largefont.eps){width="4.5in" height="3.0in"} Video over CR Femtocell Networks {#sec:femto_cr_video} ================================ In this section, we investigate the problem of video streaming in femtocell cognitive radio (CR) networks and formulate a [stochastic programming]{} problem to examine three deployment scenarios. In the case of a single FBS, we apply [dual decomposition]{} to develop an optimum-achieving distributed algorithm, which is shown also optimal for the case of multiple non-interfering FBS’s. In the case of multiple interfering FBS’s, we develop a greedy algorithm that can compute near-optimal solutions, and prove a closed-form lower bound for its performance based on an [interference graph]{} model. The proposed algorithms are evaluated with simulations, and are shown to outperform three alternative schemes with considerable gains. System Model and Preliminaries \[sec:mod4\] ------------------------------------------- ### Spectrum and Network Model We consider a spectrum consisting of $(M+1)$ channels, including one common, unlicensed channel (indexed as channel $0$) and $M$ licensed channels (indexed as channels $1$ to $M$). The $M$ licensed channels are allocated to a primary network, and the common channel is exclusively used by all CR users. We assume all the channels follow a synchronized time slot structure [@Zhao07a]. The capacity of each licensed channel is $B_1$ Mbps, while the capacity of the common channel is $B_0$ Mbps. The channel states evolve independently, while the occupancy of each licensed channel follows a two-state discrete-time Markov process. The femtocell CR network is illustrated in Fig. \[fig:netmod2\]. There is an MBS and $N$ FBS’s deployed in the area to serve CR users. The $N$ FBS’s are connected to the MBS (and the Internet) via broadband wireline connections. Due to advances in antenna technology, it is possible to equip multiple antennas at the base stations. The MBS has one antenna that is always tuned to the common channel. Each FBS is equipped with multiple antennas (e.g., $M$) and is able to sense multiple licensed channels at the beginning of each time slot. There are $K_i$ CR users in femtocell $i$, $i=1,2,\cdots,N$, and $\sum_{i=1}^N K_i = K$. Each CR user has a software radio transceiver, which can be tuned to any of the $M$+1 channels. A CR user will either connect to a nearby FBS using one or more of the licensed channels or to the MBS via the common channel. ![A femtocell CR network with one MBS and four FBS’s.[]{data-label="fig:netmod2"}](network-mod-jsac-camready.eps){width="4.0in" height="2.7in"} Although the CR users are mobile, we assume constant topology during a time slot. If the topology is changed during a time slot, the video transmission will only be interrupted for the time slot, since the proposed algorithms are executed in every time slot for new channel assignment and schedule. ### Spectrum Sensing and Access The femtocell CR network is within the coverage of the infrastructure-based primary network. Both FBS’s and CR users sense the channels to identify spectrum opportunities in each time slot. Each time slot consists of (i) a [sensing phase]{}, when CR users and FBS’s sense licensed channels, (ii) a [transmission phase]{}, when CR users and FBS’s attempt to access licensed channels, and (iii) an [acknowledgment phase]{}, when acknowledgments (ACK) are returned to the source. Cooperative sensing policy is also adopted here. We also adopt a [hypothesis test]{} to detect channel availability. We assume that each CR user chooses one channel to sense in each time slot, since it only has one transceiver. The sensing results will be shared among CR users and FBS’s via the common channel in the sensing phase. Given $L$ sensing results on channel $m$, the availability of channel $m$, i.e., $P^A_m(\Theta_1^m,\cdots,\Theta_L^m)$, can be computed iteratively as follows. $$\begin{aligned} \label{eq:Iteration2} P^A_m(\Theta_1^m)&=&\left[1+\frac{\eta_m}{1-\eta_m}\times\frac{(\delta_1^m)^{1-\Theta_1^m}(1-\delta_1^m)^{\Theta_1^m}}{(\epsilon_1^m)^{\Theta_1^m}(1-\epsilon_1^m)^{1-\Theta_1^m}}\right]^{-1} \\ %\nonumber \\ P^A_m(\vec{\Theta}_l^m)&=&P^A_m(\Theta_1^m,\Theta_2^m,\cdots,\Theta_l^m) \nonumber \\ &=&\left\{1+\left[\frac{1}{P^A_m(\Theta_1^m,\Theta_2^m,\cdots,\Theta_{l-1}^m)}-1\right]\times \right.\nonumber \\ &&\left. \frac{(\delta_l^m)^{1-\Theta_l^m}(1-\delta_l^m)^{\Theta_l^m}}{(\epsilon_l^m)^{\Theta_l^m}(1-\epsilon_l^m)^{1-\Theta_l^m}}\right\}^{-1},l=2,\cdots,L. %\nonumber\end{aligned}$$ We adopt a probabilistic approach: based on sensing results $\vec{\Theta}_m$, we have $D_m(t)=0$ with probability $P^D_m(\vec{\Theta}_m)$ and $D_m(t)=1$ with probability $1-P^D_m(\vec{\Theta}_m)$. For primary user protection, the collision probability with primary users caused by CR users should be bounded. The probability $P^D_m(\vec{\Theta}_m)$ is determined as follows $$\label{eq:PrDm2} P^D_m(\vec{\Theta}_m)=\min \left\{ \gamma_m / \left[ 1 - P^A_m(\vec{\Theta}_m) \right], 1 \right\}.$$ Let $\mathcal{A}(t) := \{m\|D_m(t)=0\}$ be the set of available channels in time slot $t$. Then $G^t=\sum_{m\in \mathcal{A}(t)} P^A_m(\Theta_1^m)$ is the expected number of available channels. These channels will be accessed in the transmission phase of time slot $t$. ### Channel Model Without loss of generality, we consider independent block fading channels that is widely used in prior work [@Rappaport01]. The channel fading-gain process is piecewise constant on blocks of one time slot, and fading in different time slots are independent. Let $f^{i,j}_X(x)$ denote the [probability density function]{} of the received SINR $X$ from a base station $i$ at CR user $j$. We assume the packet can be successfully decoded if the received SINR exceeds a threshold $H$. The packet loss probability from base station $i$ to CR user $j$ is $$\label{eq:PrFad} P_{i,j}=\Pr\{X \le H\} = \int_0^{H}f_X^{i,j}(x) dx=F_X^{i,j}(H),$$ where $F_X^{i,j}(H)$ is the cumulative density function of $X$. In the case of correlated fading channels, which can be modeled as finite state Markov Process [@Zhang99], the packet loss probability in the next time slot can be estimated from the known state of the previous time slot and the transition probabilities. If the packet is successfully decoded, the CR user returns an ACK to the base station in the ACK phase. We assume ACKs are always successfully delivered. ### Video Performance Measure We assume each active CR user receives a real-time video stream from either the MSB or an FSB. Without loss of generality, we adopt the MGS option of H.264/SVC, for scalability to accommodate the high variability of network bandwidth in CR networks. Due to real-time constraint, each Group of Pictures (GOP) of a video stream must be delivered in the next $T$ time slots. With MGS, enhancement layer NAL units can be discarded from a quality scalable bit stream, and thus packet-based quality scalable coding is provided. Our approach is to encode the video according to the maximum rate the channels can support. During transmission, only part of the MGS video gets transmitted as allowed by the current available channel bandwidth. The video packets are transmitted in decreasing order of their significance in decoding. When a truncated MGS video is received and decoded, the PSNR is computed by substituting the effective rate of the received MGS video into (\[eq:QuaMod\]) given below, thus the original video is not required. Without loss of generality, we assume that the last wireless hop is the bottleneck; video data is available at the MBS and FBS’s when they are scheduled to be transmitted. The quality of reconstructed MGS video can be modeled as [@Wien07]: $$\label{eq:QuaMod} W(R)=\alpha+\beta \times R,$$ where $W(R)$ is the average peak signal-to-noise ratio (PSNR) of the reconstructed video, $R$ is the received data rate, $\alpha$ and $\beta$ are constants depending on the video sequence and codec. We verified (\[eq:QuaMod\]) using an H.264/SVC codec and the [Bus]{}, [Mobile]{}, and [Harbour]{} test sequences. In Fig. \[fig:mgs-rd\], the markers are obtained by truncating the encoded video’s enhancement layer at different positions to obtain different effective rates, while the curves are computed using (\[eq:QuaMod\]). The curves fit well with measurements for the three sequences. It is worth noting that PSNR may not be a good measure of video quality as compared with alternative metrics such as MS-SSIM [@Wang04]. The main reason for choosing PSNR is that there is a closed-form model relating it to network level metrics–video rate. With the closed-form model, we can have a mathematical formulation of the scheduling/resource allocation problem, and derive effective algorithms. Should such closed-form models be available for MS-SSIM, it is possible to incorporate it into the optimization framework as well. ![Rate-distortion curves of three H.264/SVC MGS videos.[]{data-label="fig:mgs-rd"}](MGS-RD.eps){width="4.5in" height="3.0in"} MGS Video over Femtocell CR Networks \[sec:alg2\] ------------------------------------------------- In this section, we address the problem of resource allocation for MGS videos over femtocell CR networks. We first examine the case of a single FBS, and then the more general case of multiple non-interfering or interfering FBS’s. The algorithms for the single and non-interfering FBS cases are distributed ones and optimal. The algorithm for the interfering FBS case is a centralized one that can be executed at the MBS. To simplify notation, we omit the time slot index $t$ for most of the variables in this Section. For example, $x$ represents a variable for time slot $t$, $x^{-}$ represents the variable in time slot $(t-1)$, and $x^{+}$ represents the variable in time slot $(t+1)$. ### Case of Single FBS #### Formulation We first consider the case of a single FBS in the CR network, where the FBS can use all the $G$ available channels to stream videos to $K$ active CR users. Let $w_j$ be the PSNR of CR user $j$ at the beginning of time slot $t$ and $W_j$ the PSNR of CR user $j$ at the end of time slot $t$. In time slot $t$, $w_j$ is already known; $W_j$ is a random variable that depends on channel condition and primary user activity; and $w_j^{+}$ is a [realization]{} of $W_j$. Let $\xi_{0,j}$ and $\xi_{1,j}$ indicate the random packet losses from the MBS and FBS, respectively, to CR user $j$ in time slot $t$. That is, $\xi_{i,j}$ is $1$ with probability $\bar{P}_{i,j}=1-P_{i,j}$ and $0$ with probability $P_{i,j}$. Due to block fading channels, $P_{i,j}$’s do not change within the time slot. Let $\rho_{0,j}$ and $\rho_{1,j}$ be the portions of time slot $t$ when CR user $j$ receives video data from the MBS and FBS, respectively. The average PSNR is computed every $T$ time slots. We first have $W_j(0)=\alpha_j$, when $t=0$. In each time slot $t$, the CR user receives $\xi_{0,j} \rho_{0,j} B_0$ bits through the MBS, and $\xi_{1,j} \rho_{1,j} G B_1$ bits through the FBS (assuming that OFDM is used), which contribute an increase of $\beta (\xi_{0,j} \rho_{0,j} B_0 + \xi_{1,j} \rho_{1,j} G B_1) / T$ to the total PSNR in this $T$ time slot interval, according to (\[eq:QuaMod\]). Therefore we have the following recursive relationship: $W_j = W_j^{-} + \beta (\xi_{0,j} \rho_{0,j} B_0 + \xi_{1,j} \rho_{1,j} G B_1) / T$ = $W_j^{-} + \xi_{0,j} \rho_{0,j} R_{0,j} + \xi_{1,j} \rho_{1,j} G R_{1,j}$, where $R_{0,j}=\beta B_0/T$ and $R_{1,j}=\beta B_1/T$. For proportional fairness, we aim to maximize the sum of the logarithms of the PSNRs of all CR users [@Kelly98]. We formulate a [multistage stochastic programming problem]{} by maximizing the [expectation]{} of the logarithm-sum at time $T$. $$\begin{aligned} \label{eq:MultStage} \mbox{maximize:} && \hspace{-0.2in} \sum_{j=1}^K \mathbb{E}[\log(W_j(T))] \\ \mbox{subject to:} && \hspace{-0.2in} W_j=W_j^{-}+\xi_{0,j} \rho_{0,j} R_{0,j} + \xi_{1,j} \rho_{1,j} G R_{1,j}, %\nonumber \\ %&& \hspace{0.8in} \;\; j=1,\cdots,K, \; t=1,\cdots,T \nonumber \\ && \hspace{-0.2in} \sum_{j=1}^K \rho_{i,j} \leq 1, \;\; i=0,1, \; t=1,\cdots,T \nonumber \\ && \hspace{-0.2in} \rho_{i,j} \ge 0, \; i=0,1, \; j=1,\cdots,K, \; t=1,\cdots,T. \nonumber\end{aligned}$$ $R_{0,j}=\beta_j B_0/T$ and $R_{1,j}=\beta_jB_1/T$ are constants for the $j$-th MGS video. At the beginning of the last time slot $T$, a realization $\bm{\xi}_{[T-1]} = [ \vec{\xi}_1, \vec{\xi}_2, \cdots, \vec{\xi}_{T-1} ]$ is known, where $\vec{\xi}_t = [ \xi_{0,1}^t, \xi_{0,2}^t, \cdots, \xi_{0,K}^t, \xi_{1,1}^t, \cdots, \xi_{1,K}^t ]$, $t = 1, 2, \cdots, T-1$. It can be shown that the multistage stochastic programming problem (\[eq:MultStage\]) can be decomposed into $T$ serial sub-problems, each to be solved in a time slot $t$ [@Hu10TW]. $$\begin{aligned} \label{eq:SingStage} \mbox{maximize:} && \hspace{-0.2in} \sum_{j=1}^K \mathbb{E}\{\log(W_j)\|\bm{\xi}_{[t-1]}\} \\ \mbox{subject to:} && %\nonumber \\ \hspace{-0.2in} W_j=W_j^{-}+\xi_{0,j} \rho_{0,j} R_{0,j} + \xi_{1,j} \rho_{1,j} G R_{1,j}, %\nonumber \\ %\;\; j=1,\cdots,K \nonumber \\ %&& \hspace{1.6in} j=1,\cdots,K \nonumber \\ && \hspace{-0.2in} \sum_{j=1}^K \rho_{i,j} \le 1, \;\; i=0,1 \nonumber \\ && \hspace{-0.2in} \rho_{i,j} \ge 0, \; i=0,1, \; j=1,\cdots,K, \nonumber\end{aligned}$$ where $\mathbb{E}\{\log(W_j)\|\bm{\xi}_{[t-1]}\}$ denotes the [conditional expectation]{} of $\log(W_j)$ given realization $\bm{\xi}_{[t-1]}$. $W_j^{-}$ is known given the realization. When $t=1$, the conditional expectation becomes an unconditional expectation. Since a CR user has only one transceiver, it can operate on either one or more licensed channels (i.e., connecting to the FBS) or the common channel (i.e., connecting to the MBS), but not both simultaneously. Assume CR user $j$ operates on the common channel with probability $p_j$ and one or more licensed channels with probability $q_j$. We then rewrite problem (\[eq:SingStage\]) as $$\begin{aligned} \label{eq:ProbOpt1} \mbox{maximize:} && \hspace{-0.2in} \sum_{j=1}^K \left[ p_j \bar{P}_{0,j} \log(W_j^{-}+\rho_{0,j} R_{0,j}) \; + %q_j \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j} G R_{1,j}) %\right] \\ %\right. \\ %&& \hspace{0.5in} \left. q_j \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j} G R_{1,j}) \right] \nonumber \\ \mbox{subject to:} && \hspace{-0.2in} %\nonumber \\ \sum_{j=1}^K \rho_{i,j} \le 1, \;\; i=0,1 \nonumber \\ && \hspace{-0.2in} p_j + q_j = 1, \;\; j=1,\cdots,K \nonumber \\ && \hspace{-0.2in} \rho_{i,j}, \; p_j, \; q_j \ge 0, \;\; i=0,1, \; j=1,\cdots,K. \nonumber \end{aligned}$$ #### Properties In this section, we analyze the formulated problem (\[eq:ProbOpt1\]) and derive its properties. We have Lemmas 1, 2, and 3 and Theorem 1 and provide the proofs in the following. Problem (\[eq:ProbOpt1\]) is a convex optimization problem. First, it can be shown that the single term $ p_j \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j} R_{0,j}) + q_j \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j} G R_{1,j}) $ is a concave function, because its [Hessian matrix]{} is negative semi-definite. Then, the objective function is concave since the sum of concave functions is also concave. Finally, all the constraints are linear. We conclude that problem (\[eq:ProbOpt1\]) is convex with a unique optimal solution. If $[\rho,p,q]$ is a feasible solution to problem (\[eq:ProbOpt1\]), then $[\rho,q,p]$ is also feasible. Since $[\rho,p,q]$ is feasible, we have $p + q =1$. Switching the two probabilities, we still have $q + p =1$. Therefore, the derived new solution is also feasible. Let the optimal solution be $[\rho^{\ast},p^{\ast},q^{\ast}]$. If $p_j^{\ast} \ge q_j^{\ast}$, then $\bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j}^{\ast} R_{0,j})$ is greater than or equal to $\bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}^{\ast} G R_{1,j})$. And vice versa. Assume $\bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j}^{\ast} R_{0,j})$ is less than $\bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}^{\ast} G R_{1,j})$. Since $p_j^{\ast} \ge q_j^{\ast}$, the sum of the product $p_j^{\ast} \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j}^{\ast} R_{0,j}) + q_j^{\ast} \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}^{\ast} G R_{1,j})$ is smaller than the sum of the product $q_j^{\ast} \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j}^{\ast} R_{0,j}) + p_j^{\ast} \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}^{\ast} G R_{1,j})$. Thus we can obtain an objective value larger than the optimum by switching the values of $p_j^{\ast}$ and $q_j^{\ast}$, which is still feasible according to Lemma 2. This conflicts with the assumption that $[\rho^{\ast},p^{\ast},q^{\ast}]$ is optimal. The reverse statement can be proved similarly. Let the optimal solution be $[\rho^{\ast},p^{\ast},q^{\ast}]$. If $p_j^{\ast} > q_j^{\ast}$, then we have $p_j^{\ast} = 1$ and $q_j^{\ast} = 0$. Otherwise, we have $p_j^{\ast} = 0$ and $q_j^{\ast} = 1$. If $p_j^{\ast} > q_j^{\ast}$, we have $\bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j}^{\ast} R_{0,j}) \geq \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}^{\ast} G R_{1,j})$ according to Lemma 3. Since the objective function is linear with respect to $p_j$ and $q_j$, the optimal value can be achieved by setting $p_j$ to its maximum value 1 and $q_j$ to its minimum value 0. The reverse statement can be proved similarly. According to Theorem 1, a CR user is connected to either the MBS or the FBS for the [entire]{} duration of a time slot in the optimal solution. That is, it does not switch between base stations during a time slot under optimal scheduling. #### Distributed Solution Algorithm To solve problem (\[eq:ProbOpt1\]), we define non-negative [dual variables]{} ${\mathcal \lambda}=[\lambda_0, \lambda_1]$ for the two inequality constraints. The [Lagrangian function]{} is $$\begin{aligned} \label{eq:LagProbOpt1} \hspace{-0.2in} \mathcal{L}(p,\rho,\lambda) \hspace{-0.025in}&=&\hspace{-0.025in} \sum_{j=1}^K \left[ p_j \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j} R_{0,j}) + %\right. \nonumber \\ %&& \left. (1-p_j) \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j} G R_{1,j}) \right] + \nonumber \\ && \lambda_0(1 - \sum_{j=1}^K \rho_{0,j})+ \lambda_1(1 - \sum_{j=1}^K \rho_{1,j}) \nonumber \\ &=& \hspace{-0.025in} \sum_{j=1}^K \mathcal{L}_j(p_j,\rho_{0,j},\rho_{1,j},\lambda_0,\lambda_1) \hspace{-0.025in} + \hspace{-0.025in} \lambda_0 \hspace{-0.025in} + \hspace{-0.025in} \lambda_1, \end{aligned}$$ where $$\begin{aligned} \hspace{0.1in} \mathcal{L}_j(p_j,\rho_{0,j},\rho_{1,j},\lambda_0,\lambda_1) = p_j \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j} R_{0,j}) + \nonumber \\ \hspace{0.2in} (1 - p_j) \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j} G R_{1,j}) - \lambda_0 \rho_{0,j} - \lambda_1 \rho_{1,j}. \nonumber\end{aligned}$$ The corresponding problem can be decomposed into $K$ sub-problems and solved iteratively. In Step $\tau \geq 1$, for given $\lambda_0(\tau)$ and $\lambda_1(\tau)$ values, each CR user $j$ solves the following sub-problem using local information. $$\begin{aligned} \label{eq:ArgSubOpt1} %&& \hspace{-0.3in} [p_j^{\ast}(\tau), \rho_{0,j}^{\ast}(\tau), \rho_{1,j}^{\ast}(\tau)] %\nonumber \\ %&& \hspace{-0.5in} = \stackbin[p_j,\rho_{0,j},\rho_{1,j} \ge 0]{}{\arg\max} \mathcal{L}_j(p_j,\rho_{0,j},\rho_{1,j},\lambda_0(\tau),\lambda_1(\tau)).\end{aligned}$$ There is a unique optimal solution since the objective function in (\[eq:ArgSubOpt1\]) is concave. The CR users then exchange their solutions. The [master dual problem]{}, for given $p(\tau)$ and $\rho(\tau)$, is: $$\begin{aligned} \label{eq:MaterOpt1} %&& \hspace{-0.2in} \min_{\lambda\ge 0} \mathcal{L}(p(\tau),\rho(\tau),\lambda) %\nonumber \\ %&& \hspace{-0.4in} = \sum_{j=1}^K \mathcal{L}_j(p_j(\tau),\rho_{0,j}(\tau),\rho_{1,j}(\tau),\lambda_0,\lambda_1)+\lambda_0+\lambda_1. \end{aligned}$$ Since the Lagrangian function is differentiable, the [gradient iteration]{} approach can be used. $$\label{eq:IterOpt1} \lambda_i(\tau+1) = \left[ \lambda_i(\tau) - s \times \left( 1 - \sum_{j=1}^K \rho_{i,j}^{\ast}(\tau) \right) \right]^+, \; i=0,1,$$ where $s$ is a sufficiently small positive [step size]{} and $[\cdot]^+$ denotes the projection onto the nonnegative axis. The updated $\lambda_i(\tau+1)$ will again be used to solve the sub-problems, and so forth. Since the problem is convex, we have [strong duality]{}; the [duality gap]{} between the primal and dual problems is zero. The dual variables $\lambda(\tau)$ will converge to the optimal values as $\tau$ goes to infinity. Since the optimal solution to (\[eq:ArgSubOpt1\]) is unique, the primal variables $p(\tau)$ and $\rho_{i,j}(\tau)$ will also converge to their optimal values when $\tau$ is sufficiently large. The distributed solution procedure is presented in Table \[tab:Opt1\]. In the table, Steps 3–8 solve the sub-problem in (\[eq:ArgSubOpt1\]); Step 9 updates the dual variables. The threshold $\phi$ is a prescribed small value with $0 \leq \phi \ll 1$. The algorithm terminates when the dual variables are sufficiently close to the optimal values. ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ 1: Set $\tau=0$, $\lambda_0(0)$ and $\lambda_1(0)$ to some nonnegative value; 2: DO % (each CR user $j$ executes Steps 3–8) 3: $\;\;$ $\rho_{0,j}(\tau) \hspace{-0.025in} = \hspace{-0.025in} \left[ \frac{\bar{P}_{0,j}}{\lambda_0(\tau)} \hspace{-0.025in} - \hspace{-0.025in} \frac{W_j^{-}}{R_{0,j}} \right]^+$, $\rho_{1,j}(\tau) \hspace{-0.025in} = \hspace{-0.025in} \left[ \frac{\bar{P}_{1,j}}{\lambda_1(\tau)} \hspace{-0.025in} - \hspace{-0.025in} \frac{W_j^{-}}{R_{1,j} G} \right]^+$; 4: $\;\;$ IF $\left[ \bar{P}_{0,j} \log(W_j^{-}+\rho_{0,j}(\tau) R_{0,j})-\lambda_0(\tau)\rho_{0,j}(\tau) \right] >$ $\;\;$ $\left[ \bar{P}_{1,j} \log(W_j^{-} + \rho_{1,j}(\tau) G R_{1,j}) - \lambda_1(\tau) \rho_{1,j}(\tau) \right]$ 5: $\;\;\;\;\;$ Set $p_j(\tau)=1$ and $\rho_{1,j}(\tau)=0$; 6: $\;\;$ ELSE 7: $\;\;\;\;\;$ Set $p_j(\tau)=0$ and $\rho_{0,j}(\tau)=0$; 8: $\;\;$ END IF 9: $\;\;$ MBS updates $\lambda_i(\tau+1)$ as in (\[eq:IterOpt1\]); 10: $\;\;$ $\tau=\tau+1$; 11: WHILE $\left( \sum_{i=0}^{1}(\lambda_i(\tau+1)-\lambda_i(\tau))^2 > \phi \right)$ ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ : Algorithm for the Case of Single FBS \[tab:Opt1\] ### Case of Multiple Non-interfering FBS’s \[subsec:mulnifbs\] We next consider the case of $N>1$ non-interfering FBS’s. The coverages of the FBS’s do not overlap with each other, as FBS 1 and 2 in Fig. \[fig:netmod2\]. Consequently, each FBS can use all the available licensed channels without interfering other FBS’s. Assume each CR user knows the nearest FBS and is associate with it. Let $\mathcal{U}_i$ denote the set of CR users associated with FBS $i$. The resource allocation problem becomes: $$\begin{aligned} \label{eq:ProbOptDM} \mbox{maximize:} && \hspace{-0.2in} \sum_{j=1}^K p_j \bar{P}_{0,j} \log(W_j^{-} + \rho_{0,j} R_{0,j}) + %\\ %&& \sum_{i=1}^N \sum_{j\in\mathcal{U}_i} q_j \bar{P}_{i,j} \log(W_j^{-} + \rho_{i,j} G R_{i,j}) \\ %\nonumber \\ \mbox{subject to:} && \hspace{-0.2in} \sum_{j=1}^K \rho_{0,j} \leq 1 \nonumber \\ && \hspace{-0.2in} \sum_{j\in\mathcal{U}_i} \rho_{i,j} \le 1, \;\; i=1,\cdots,N \nonumber \\ && \hspace{-0.2in} p_j + q_j = 1, \;\; j=1,\cdots,K \nonumber \\ && \hspace{-0.2in} \rho_{i,j}, \; p_j, \; q_j \ge 0, \;\; i=1,\cdots,N,\; j=1,\cdots,K. \nonumber\end{aligned}$$ Since all the available channels can be allocated to each FBS with spatial reuse, problem (\[eq:ProbOptDM\]) can be solved using the algorithm in Table \[tab:Opt1\] with some modified notation: $\rho_{1,j}(\tau)$ now becomes $\rho_{i,j}(\tau)$ and $\lambda_1(\tau)$ becomes $\lambda_i(\tau)$, $i=1, \cdots, N$. The dual variables are iteratively updated as $$\begin{aligned} && \hspace{-0.4in} \lambda_0(\tau+1)=\left[\lambda_0(\tau)-s \times \left( 1 - \sum_{j=1}^K \rho_{0,j}^{\ast}(\tau) \right) \right]^+ \label{eq:IterOptM0} \\ && \hspace{-0.4in} \lambda_i(\tau+1)=\left[\lambda_i(\tau)-s \times \left( 1 - \sum_{j\in \mathcal{U}_i} \rho_{i,j}^{\ast}(\tau) \right) \right]^+, %\nonumber \\ %&& \hspace{1.8in} \;\; i=1,\cdots,N. \label{eq:IterOptM1}\end{aligned}$$ The modified solution algorithm is presented in Table \[tab:OptDisM\]. As in the case of single FBS, the algorithm is jointly executed by the CR users and MBS, by iteratively updating the dual variables $\lambda_0(\tau)$ and $\lambda_i(\tau)$’s, and the resource allocations $\rho_{0,j}^{\ast}(\tau)$ and $\rho_{i,j}^{\ast}(\tau)$’s. It can be shown that the distributed algorithm can produce the optimal solution for problem (\[eq:ProbOptDM\]). ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- 1: Set $\tau=0$, and $\lambda_0(0)$ and $\lambda_i(0)$ to some nonnegative values, for all $i$; 2: DO % (each CR user $j$ executes Steps 3–8) 3: $\;\;$ $\rho_{0,j}(\tau) \hspace{-0.025in} = \hspace{-0.025in} \left[ \frac{\bar{P}_{0,j}}{\lambda_0(\tau)} \hspace{-0.025in} - \hspace{-0.025in} \frac{W_j^{-}}{R_{0,j}} \right]^+$, $\rho_{i,j}(\tau) \hspace{-0.025in} = \hspace{-0.025in} \left[ \frac{\bar{P}_{i,j}}{\lambda_i(\tau)} \hspace{-0.025in} - \hspace{-0.025in} \frac{W_j^{-}}{R_{i,j} G} \right]^+$ ; 4: $\;\;$ IF $\left[ \bar{P}_{0,j} \log(W_j^{-}+\rho_{0,j}(\tau) R_{0,j})-\lambda_0(\tau)\rho_{0,j}(\tau) \right] >$ $\;\;$ $\left[ \bar{P}_{i,j} \log(W_j^{-} + \rho_{i,j}(\tau) G R_{i,j}) - \lambda_i(\tau) \rho_{i,j}(\tau) \right]$ 5: $\;\;\;\;\;$ Set $p_j(\tau)=1$ and $\rho_{i,j}(\tau)=0$; 6: $\;\;$ ELSE 7: $\;\;\;\;\;$ Set $p_j(\tau)=0$ and $\rho_{0,j}(\tau)=0$; 8: $\;\;$ END IF 9: $\;\;$ MBS updates $\lambda_i(\tau+1)$ as in (\[eq:IterOptM0\]) and (\[eq:IterOptM1\]); 10: $\;\;$ $\tau=\tau+1$; 11: WHILE $\left( \sum_{i=0}^{N}(\lambda_i(\tau+1)-\lambda_i(\tau))^2 > \phi \right)$ ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- : Algorithm for the Case of Multiple Non-Interfering FBS’s \[tab:OptDisM\] ### Case of Multiple Interfering FBS’s #### Formulation Finally, we consider the case of multiple interfering FBS’s. Assume that the coverages of some FBS’s overlap with each other, as FBS 3 and 4 in Fig. \[fig:netmod2\]. They cannot use the same channel simultaneously, but have to compete for the available channels in the transmission phase. Define [channel allocation variables]{} $c_{i,m}$ for time slot $t$ as: $$\label{eq:cimt} c_{i,m} = \left\{ \begin{array}{ll} 1, & \mbox{if channel $m$ is allocated to FBS $i$} \\ 0, & \mbox{otherwise.} \end{array} \right.$$ Given an allocation, the expected number of available channels for FBS $i$ is $G_i \hspace{-0.025in} = \hspace{-0.025in} \sum_{m\in\mathcal{A}(t)} c_{i,m} P_m^A$. We use [interference graph]{} to model the case of overlapping coverages, which is defined below. An [interference graph]{} $G_I=(V_I,E_I)$ is an undirected graph where each vertex represents an FBS and each edge indicates interference between the two end FBS’s. For the example given in Fig. \[fig:netmod2\], we can derive an interference graph as shown in Fig. \[fig:interferencegraph\]. FBS 3 and 4 cannot use the same channel simultaneously, as summarized in the following lemma. ![Interference graph for the femtocell CR network shown in Fig. \[fig:netmod2\].[]{data-label="fig:interferencegraph"}](interference-graph-jsac-camready.eps){width="4.0in"} If channel $m$ is allocated to FBS $i$, the neighboring vertices of FBS $i$ in the interference graph $G_I$, denoted as $\mathcal{R}(i)$, cannot use the same channel $m$ simultaneously. Further define index variables $d_i^k$ as $$\label{eq:dik} d_i^k = \left\{ \begin{array}{ll} 1, & \mbox{if FBS $i$ is an endpoint of link $k \in G_I$} \\ 0, & \mbox{otherwise.} \end{array} \right.$$ The interference constraint can be described as $\sum_{i=1}^N d_i^k c_{i,m} \le 1$, for $m=0,\cdots,M$, and for all link $k \in G_I$. We then have the following problem formulation. $$\begin{aligned} \label{eq:ProbOptM} \mbox{maximize:} && \hspace{-0.25in} \sum_{j=1}^K p_j \bar{P}_{0,j} \log(W_j^{-}+\rho_{0,j} R_{0,j}) + %\\ %&& \sum_{i=1}^N \sum_{j\in\mathcal{U}_i} q_j \bar{P}_{i,j} \log(W_j^{-} + \rho_{i,j} G_i R_{i,j}) \\ %\nonumber \\ \mbox{subject to:} && \hspace{-0.2in} \sum_{j=1}^K \rho_{0,j} \leq 1 \nonumber \\ && \hspace{-0.25in} \sum_{j\in\mathcal{U}_i} \rho_{i,j} \le 1, \;\; i=1,\cdots,N \nonumber \\ && \hspace{-0.25in} p_j + q_j = 1, \;\; j=1,\cdots,K \nonumber \\ && \hspace{-0.25in} G_i = \sum_{m\in\mathcal{A}(t)} c_{i,m} P_m^A, \;\; i=1,\cdots,N \nonumber \\ && \hspace{-0.25in} \sum_{i=1}^N d_i^k c_{i,m} \le 1, m=0,\cdots,M, \mbox{for link } k \in G_I, \nonumber \\ && \hspace{-0.25in} \rho_{i,j}, p_j, q_j, c_{i,m} \ge 0, \;\; %\nonumber \\ %&& i=1,\cdots,N,\; j=1,\cdots,K, \; m=0,\cdots,M. \nonumber\end{aligned}$$ #### Solution Algorithm \[subsubsec:ifbs\] The optimal solution to problem (\[eq:ProbOptM\]) depends on the channel allocation variables $c_{i,m}$. Problem (\[eq:ProbOptM\]) can be solved with the algorithm in Table \[tab:OptDisM\] if the $c_{i,m}$’s are known. Let $Q(\bm{c})$ be the suboptimal objective value for a given channel allocation $\bm{c}$, where $\bm{c}=[\vec{c}_1, \vec{c}_2, \cdots, \vec{c}_N]$ and $\vec{c}_i$ is a vector of elements $c_{i,m}$, for FBS $i$ and channels $m \in \mathcal{A}(t)$. If all the FBS’s are disjointedly distributed with no overlap, each FBS can use all the available channels. We have $c_{i,m}=1$ for all $i$ and $m \in \mathcal{A}(t)$, i.e., it is reduced to the case in Section \[subsec:mulnifbs\]. To solve problem (\[eq:ProbOptM\]), we first apply a [greedy algorithm]{} to allocate the available channels in $\mathcal{A}(t)$ to the FBS’s (i.e., to determine $\bm{c}$). We then apply the algorithm in Table \[tab:OptDisM\] with the computed $\bm{c}$ to obtain a near-optimal solution. Let $\bm{e}_{i,m}$ be a matrix with $1$ at position $\{i,m\}$ and $0$ at all other positions, representing the allocation of channel $m \in \mathcal{A}(t)$ to FBS $i$. The greedy channel allocation algorithm is given in Table \[tab:ChanAloc\], where the FBS-channel pair that can achieve the largest increase in $Q(\cdot)$ is chosen in each iteration. The worst case complexity of the greedy algorithm is $O(N^2 M^2)$. ---- ------------------------------------------------------------------------------------------------------------------------ 1: Initialize $\bm{c}$ to a zero matrix, FBS set $\mathcal{N}=\{1,\cdots,N\}$, and FBS-channel set $\mathcal{C}=\mathcal{N}\times\mathcal{A}(t)$; 2: WHILE ($\mathcal{C}$ is not empty) 3: $\;\;$ Find FBS-channel pair $\{i',m'\}$, such that $\;\;\;\;\;\;$ $\{i',m'\} = \stackbin[\{i,m\} \in \mathcal{C}]{}{\arg\max} \{Q(\bm{c} + \bm{e}_{i,m}) - Q(\bm{c}) \}$; 4: $\;\;$ Set $\bm{c} = \bm{c} + \bm{e}_{i',m'}$; 5: $\;\;$ Remove $\{i',m'\}$ from $\mathcal{C}$; 6: $\;\;$ Remove $\mathcal{R}(i') \times m'$ from $\mathcal{C}$; 7: END WHILE ---- ------------------------------------------------------------------------------------------------------------------------ : Channel Allocation Algorithm for Case of Interfering FBS’s \[tab:ChanAloc\] #### Performance Lower Bound We next present a lower bound for the greedy algorithm. Let $e(l)$ be the $l$-th FBS-channel pair chosen in the greedy algorithm, and $\pi_l$ denote the sequence $\{e(1),e(2),\cdots,e(l)\}$. The increase in object value (\[eq:ProbOptM\]) due to the $l$-th allocated FBS-channel pair is denoted as $$\Delta_l := \Delta(\pi_l, \pi_{l-1}) = Q(\pi_l) - Q(\pi_{l-1}).$$ Since $Q(\pi_0)=Q(\emptyset)=0$, we have $$\begin{aligned} \sum_{l=1}^L \Delta_l &=& Q(\pi_L)-Q(\pi_{L-1})+ \cdots + Q(\pi_1) - Q(\pi_0) \nonumber \\ &=& Q(\pi_L) - Q(\pi_0) = Q(\pi_L). \nonumber\end{aligned}$$ For two FBS-channel pairs $e(l)$ and $e(l')$, we say $e(l)$ [conflicts with]{} $e(l')$ when there is an edge connecting the FBS in $e(l)$ and the FBS in $e(l')$ in the interference graph $G_I$, and the two FBS’s choose the same channel. Let $\Omega$ be the global optimal solution. We define $\omega_l$ as the subset of $\Omega$ that conflicts with allocation $e(l)$ but not with the previous allocations $\{e(1), e(2), \cdots, e(l-1)\}$. \[lemma1:5\] Assume the greedy algorithm in Table \[tab:ChanAloc\] stops in $L$ steps. The global optimal solution $\Omega$ can be partitioned into $L$ non-overlapping subsets $\omega_l$, $l=1,2,\cdots,L$. According to the definition of $\omega_l$, the $L$ subsets of the optimal solution $\Omega$ do not intersect with each other. Assume the statement is false, then the union of these $L$ subsets is not equal to the optimal set $\Omega$. Let the [set difference]{} be $\omega_{L+1} = \Omega \setminus (\cup_{l=1}^L \omega_l)$. By definition, $\omega_{L+1}$ does not conflict with the existing $L$ allocations $\{e(1), \cdots, e(L)\}$, meaning that the greedy algorithm can continue to at least the $(L+1)$-th step. This conflicts with the assumption that the greedy algorithm stops in $L$ steps. It follows that $\Omega = \cup_{l=1}^L \omega_l$. Let $\Delta(\pi_2,\pi_1)=Q(\pi_2)-Q(\pi_1)$ denote the difference between two feasible allocations $\pi_1$ and $\pi_2$. We next derive a lower bound on the performance of the greedy algorithm. We assume two properties for function $\Delta(\pi_2,\pi_1)$ in the following. Consider FBS-channel pair sets $\pi_1$, $\pi_2$, and $\sigma$, satisfying $\pi_1\subseteq\pi_2$ and $\sigma \cap \pi_2 =\emptyset$. We have $\Delta(\pi_2\cup\sigma, \pi_1\cup\sigma) \le \Delta(\pi_2,\pi_1)$. Consider FBS-channel pair sets $\pi$, $\sigma_1$, and $\sigma_2$ satisfying $\sigma_1 \cap \pi = \emptyset$, $\sigma_2 \cap \pi = \emptyset$, and $\sigma_1 \cap \sigma_2 = \emptyset$. We have $\Delta(\sigma_1 \cup \sigma_2 \cup \pi, \pi) \le \Delta(\sigma_1 \cup \pi,\pi) + \Delta(\sigma_2 \cup \pi, \pi)$. In Property 1, we have $\sigma \cap \pi_1 =\emptyset$ since $\pi_1\subseteq\pi_2$ and $\sigma \cap \pi_2 =\emptyset$. This property states that the incremental objective value does not get larger as more channels are allocated and as the objective value gets larger. Property 2 states that the incremental objective value achieved by allocating multiple FBS-channel pair sets does not exceed the sum of the incremental objective values achieved by allocating each individual FBS-channel pair set. These are generally true for many resource allocation problems [@Kelly98]. Since we choose the maximum incremental allocation in each step of the greedy algorithm, we have Lemma \[lemma:step3\] that directly follows Step 3 in Table \[tab:ChanAloc\]. \[lemma:step3\] For any FBS-channel pair $\sigma \in \omega_l$, we have $Q(\pi_{l-1} \cup \sigma) - Q(\pi_{l-1}) = \Delta(\pi_{l-1} \cup \sigma, \pi_{l-1}) \le \Delta_l$. \[lemma:6\] Assume the greedy algorithm stops in $L$ steps, we have $$Q(\Omega)\le Q(\pi_L)+\sum_{l=1}^L \sum_{\sigma\in \omega_{l}} \Delta(\sigma \cup \pi_{l-1}, \pi_{l-1}).$$ The following inequalities hold true according to the properties of the $\Delta(\cdot, \cdot)$ function: $$\begin{aligned} \label{eq:IneqProof} %&&\hspace{-0.2in} Q((\cup_{i=l+1}^L \omega_i)\cup \pi_l) &=& Q((\cup_{i=l+2}^L \omega_i)\cup \pi_l) + %\nonumber\\ %&&\hspace{0.6in} \Delta((\cup_{i=l+1}^L \omega_i)\cup \pi_l,(\cup_{i=l+2}^L \omega_i)\cup \pi_l) \nonumber \\ %&&\hspace{-0.35in} && \le Q((\cup_{i=l+2}^L \omega_i)\cup \pi_l)+\Delta(\omega_{l+1}\cup\pi_l,\pi_l) \nonumber \\ %&&\hspace{-0.35in} &&\le Q((\cup_{i=l+2}^L \omega_i)\cup \pi_{l+1})+\Delta(\omega_{l+1}\cup\pi_l,\pi_l) \nonumber \\ %&&\hspace{-0.35in} && \le Q((\cup_{i=l+2}^L \omega_i)\cup \pi_{l+1}) + \sum_{\sigma\in\omega_{l+1}} \Delta(\sigma \cup \pi_l,\pi_l). \nonumber\end{aligned}$$ We have $\pi_0=\emptyset$ and $\omega_{L+1}=\emptyset$ (see Lemma \[lemma1:5\]). With induction from $l=0$ to $l=L-1$, we have $Q((\cup_{i=1}^L \omega_i) \cup \emptyset) = Q(\Omega)$ and $Q(\Omega) \le Q(\pi_{L}) + \sum_{l=1}^L \sum_{\sigma\in\omega_{l}} \Delta(\sigma\cup\pi_{l-1},\pi_{l-1})$. \[lemma:7\] The maximum size of $\omega_l$ is equal to the degree, in the interference graph $G_I$, of the FBS selected in the $l$-th step of the greedy algorithm, which is denoted as $D(l)$. Once FBS $i$ is allocated with channel $m$, the neighboring FBS’s in $G_I$, $\mathcal{R}(i)$, cannot use the same channel $m$ anymore due to the interference constraint. The maximum number of FBS-channel pairs that conflict with the selected FBS-channel pair $\{i,m\}$, i.e., the maximum size of $\omega_l$, is equal to the degree of FBS $i$ in $G_I$. Then we have Theorem \[th1:2\] that provides a lower bound on the objective value achieved by the greedy algorithm given in Table \[tab:ChanAloc\]. \[th1:2\] The greedy algorithm can achieve an objective value that is at least $\frac{1}{1+D_{max}}$ of the global optimum, where $D_{max}$ is the maximum node degree in the interference graph $G_I$ of the femtocell CR network. According to Lemmas \[lemma:6\] and \[lemma:7\], we have: $$\begin{aligned} \label{eq:OptBound} && \hspace{0.0in} Q(\Omega) \le Q(\pi_L) + \sum_{l=1}^L D(l)\Delta_l = Q(\pi_L)+\bar{D} \sum_{l=1}^L \Delta_l \nonumber \\ && \hspace{0.375in} = (1 + \bar{D}) Q(\pi_L), \end{aligned}$$ where $\bar{D}=\sum_{l=1}^L D(l)\Delta_l / \sum_{l=1}^L \Delta_l$. The second equality is due to the facts that $\sum_{l=1}^L \Delta_l = Q(\pi_L)$. To further simplify the bound, we replace $D(l)$ with the maximum node degree $D_{max}$. We then have $\bar{D} \leq \sum_{l=1}^L D_{max} \Delta_l / \sum_{l=1}^L \Delta_l = D_{max}$ and $$\label{eq:lowerbd} \frac{1}{1 + D_{max}} Q(\Omega) \le Q(\pi_L) \le Q(\Omega),$$ which provides a lower bound on the performance of the greedy algorithm. When there is a single FBS in the CR network, we have $D_{max}=0$ and $Q(\pi_L) = Q(\Omega)$ according to Theorem \[th1:2\]. The proposed algorithm produces the optimal solution. In the case of multiple non-interfering FBS’s, we still have $D_{max}=0$ and can obtain the optimal solution using the proposed algorithm. For the femtocell CR network given in Fig. \[fig:netmod2\] (with interference graph shown in Fig. \[fig:interferencegraph\]), we have $D_{max}=1$ and the low bound is a half of the global optimal. Note that (\[eq:OptBound\]) provides a tighter bound for the optimum than (\[eq:lowerbd\]), but with higher complexity. These are interesting performance bounds since they bound the achievable video quality, an application layer performance measure, rather than lower layer metrics (e.g., bandwidth or time share). Simulation Results \[sec:sim4\] ------------------------------- We evaluate the performance of the proposed algorithms using MATLAB and JSVM 9.13 Video codec. Two scenarios are used in the simulations: a single FBS CR network and a CR network with interfering FBS’s. In every simulation, we compare the proposed algorithms with the following three more straightforward heuristic schemes: - Heuristic 1 based on [equal allocation]{}: each CR user chooses the better channel (i.e., the common channel or a licensed channel) based on the channel conditions; time slots are equally allocated among active CR users; - Heuristic 2 exploiting [multiuser diversity]{}: the MBS and each FBS chooses one active CR user with the best channel condition; the entire time slot is allocated to the selected CR user. - [SCA-MAC]{} proposed in [@Hsu07]: with this scheme, the successful transmission rate is evaluated based on channel packet loss rate and collision probability with primary users; the channel-user pair with the highest transmission probability is selected. We choose SCA-MAC because it adopts similar models and assumptions as in this paper. Once the channels are selected, the same distributed algorithm is used for scheduling video data for all the three schemes. We adopt the Raleigh block fading model and the packet loss probability is between \[0.004, 0.028\]. The frame rate is set to 30 fps and the GoP size is 16. The base layer mode is set to be AVC compatible. The motion search mode is set to Fast Search with search range 32. Each point in the figures presented in this section is the average of 10 simulation runs with different random seeds. We plot 95% confidence intervals in the figures, which are generally negligible. ### Case of Single FBS In the first scenario, there are $M=8$ channels and the channel parameters $P_{01}^m$ and $P_{10}^m$ are set to 0.4 and 0.3, respectively, for all $m$. The maximum allowable collision probability $\gamma_m$ is set to 0.2 for all $m$. There is one FBS and three active CR users. Three Common Intermediate Format (CIF, 352$\times$288) video sequences are streamed to the CR users, i.e., [Bus]{} to CR user 1, [Mobile]{} to CR user 2, and [Harbor]{} to CR user 3. We have $T=10$ as the delivery deadline. Both probabilities of false alarm $\epsilon$ and miss detection $\delta$ are set to 0.3 for all the FBS’s and CR users, unless otherwise specified. First we investigate the convergence of the distributed algorithm. The traces of the two dual variables are plotted in Fig. \[fig:conv\]. To improve the convergence speed, the correlation in adjacent time slots can be exploited. In particular, we set the optimal values for the optimization variables in the previous time slot as the initialization values for the variables in the current time slot. By doing so, the convergence speed can be improved. It can be seen that both dual variables converge to their optimal values after 300 iterations. After convergence, the optimal solution for the primary problem can be obtained. ![Convergence of the two dual variables in the single FBS case.[]{data-label="fig:conv"}](convergence_plot2.eps){width="4.5in" height="3.0in"} Our proposed scheme achieves the best performance among the three algorithms, with up to 4.3 dB improvement over the two heuristic schemes and up to 2.5 dB over SCA-MAC. Such gains are significant with regard to video quality, since a 0.5 dB difference is distinguishable by human eyes. Compared to the two heuristic schemes and SCA-MAC, the video quality of our proposed scheme is well balanced among the three users, indicating better fairness performance. In Fig. \[fig:singleFBSmAll\], we examine the impact of the number of channels $M$ on received video quality. First, we validate the video quality measure used in our formulation by comparing the PSNR value computed using (\[eq:QuaMod\]) with that computed from real decoded video frames. The average PSNR for three received videos are plotted in the figure. It can be seen that the real PSNRs are very close to those predicted by (\[eq:QuaMod\]), with overlapping confidence intervals. This is also consistent with the results shown in Fig. \[fig:mgs-rd\]. Second, as expected, the more licensed channels, the more spectrum opportunities for CR users and the higher PSNR for received videos. SCA-MAC performs better than two heuristics, but is inferior to the proposed scheme. ![Single FBS: received video quality vs. number of channels (computed with (9) and measured by PSNR).[]{data-label="fig:singleFBSmAll"}](SingleFBS_ChanM_Plot_all.eps){width="4.5in" height="3.0in"} We also plot the MS-SSIM of the received videos at the three CR users in Fig. \[fig:singleFBSmSSIM\] [@Wang04]. Similar observations can be made from the MS-SSIM plot. All MS-SSIMs for the four curves are more than 0.97 and very close to 1. The proposed scheme still outperforms the other three schemes. In the remaining figures, we will use model predicted PSNR values, since the model (\[eq:QuaMod\]) is sufficient to predict the real video quality. ![Single FBS: received video quality vs. number of channels (measured by MS-SSIM).[]{data-label="fig:singleFBSmSSIM"}](SingleFBS_ChanM_Plot_SSIM.eps){width="4.5in" height="3.0in"} In Fig. \[fig:singleFBSeta\], we demonstrate the impact of channel utilization $\eta$ on received video quality. The average PSNRs achieved by the four schemes are plotted when $\eta$ is increased from 0.3 to 0.7. Intuitively, a smaller $\eta$ allows more spectrum opportunities for video transmission. This is illustrated in the figure where all the three curves decrease as $\eta$ gets larger. The performance of both heuristics are close and the proposed scheme achieves a gain about 3 dB over the heuristics and 2 dB over SCA-MAC. ![Single FBS: received video quality vs. channel utilization.[]{data-label="fig:singleFBSeta"}](SingleFBS_ChanEta_Plot.eps){width="4.5in" height="3.0in"} We also compare the MGS and FGS videos while keeping other parameters identical. We find that MGS video achieves over 0.5 dB gain in video quality over FGS video. The results are omitted for brevity. ### Case of Interfering FBS’s We next investigate the second scenario with three FBS’s, and each FBS has three active CR users. Each FBS streams three different videos to the corresponding CR users. The coverages of FBS 1 and 2 overlap with each other, and the coverages of FBS 2 and 3 overlap with each other. In Fig. \[fig:multiFBSm\], we examine the impact of the number of channels $M$ on the received video quality. The average PSNRs of all the active CR users are plotted in the figure when we increase $M$ from 12 to 20 with step size 2. As mentioned before, more channels imply more transmission opportunities for video transmission. In this scenario, heuristic 2 (with a multiuser diversity approach) outperforms heuristic 1 (with an equal allocation approach). But its PSNRs are still about 0.3 $\sim$ 0.5 dB lower that those of the proposed algorithm. The proposed scheme has up to 0.4 dB improvement over SCA-MAC. In Fig. \[fig:multiFBSm\], we also plot an upper bound on the optimal objective value, which is obtained as in (\[eq:OptBound\]). It can be seen that the performance of our proposed scheme is close to optimal solution since the gap between the upper bound and our scheme is generally small (about 0.5 dB). ![Interfering FBS’s: received video quality vs. number of channels.[]{data-label="fig:multiFBSm"}](MultiFBS_ChanM_Plot.eps){width="4.5in" height="3.0in"} Next, we examine the impact of sensing errors on the received video quality. In Fig. \[fig:multiFBSepsilon\], we test five pairs of $\{\epsilon, \delta\}$ values: {0.2,0.48}, {0.24,0.38}, {0.3,0.3}, {0.38,0.24}, and {0.48,0.2}. It is interesting to see that the performance of all the four schemes get worse when the probability of one of the two sensing errors gets large. We can trade-off between false alarm and miss detection probabilities to find the optimal operating point for the spectrum sensors. Moreover, the dynamic range of video quality is not big for the range of sensing errors simulated, compared to that in Fig. \[fig:multiFBSm\]. This is because both sensing errors are modeled and treated in the algorithms. Again, our proposed scheme outperforms the two heuristic schemes and SCA-MAC with considerable margins for the entire range. ![Interfering FBS’s: received video quality vs. sensing error probability.[]{data-label="fig:multiFBSepsilon"}](MultiFBS_SenseErr_Plot.eps){width="4.5in" height="3.0in"} We also investigate the impact of the bandwidth of the common channel $B_0$. In this simulation, we fix $B_1$ at 0.3 Mbps and increase $B_0$ from 0.1 Mbps to 0.5 Mbps with step size 0.1 Mbps. The results are presented in Fig. \[fig:multiFBSCtrlBW\]. We notice that the average video quality increases rapidly as the common channel bandwidth is increased from 0.1 Mbps to 0.3 Mbps. Beyond 0.3 Mbps, the increases of the PSNR curves slow down and the curves get flat. This implies that a very large bandwidth for the common channel is not necessary, since the gain for additional bandwidth diminishes as $B_0$ gets large. Again, the proposed scheme outperforms the other three schemes and the gap between our scheme and the upper bound is small. ![Interfering FBS’s: received video quality vs. bandwidth of the common channel.[]{data-label="fig:multiFBSCtrlBW"}](MultiFBS_CtrlBW_Plot.eps){width="4.5in" height="3.0in"} Next, we stop the distributed algorithm after a fixed amount of time, and evaluate the suboptimal solutions. In particular, we vary the duration of time slots, and let the distributed algorithm run for 5% of the time slot duration at the beginning of the time slot. Then the solution obtained this way will be used for the video data transmissions. The results are presented in Fig. \[fig:timedura\]. It can be seen that when the time slot is 5 ms, the algorithm does not converge after 5%$\times$5 = 0.25 ms and the PSNR produced by the distributed algorithm is close to that of Heuristic 1, and lower than those of Heuristic 2 and SCA-MAC. When the time slot is sufficiently large, the algorithm can get closer to the optimal and the proposed algorithm produces better video quality as compared to the two heuristic algorithms and SCA-MAC. Beyond 20 ms, the increase in PSNR is small since all the curves gets flat. Therefore the proposed algorithm could be useful even when there is no time for it to fully converge to the optimal. ![Video quality achieved by the algorithms when they are only executed for 5% of the time slot duration.[]{data-label="fig:timedura"}](MultiFBS_TimeDura_Plot.eps){width="4.5in" height="3.0in"} During the simulations, we find the collision rate with primary users are strictly kept below the prescribed collision tolerance $\gamma$. These results are omitted for brevity. Conclusions {#sec:femto_conc} =========== In this paper, we first investigated data multicast in femtocell networks consisting of an MBS and multiple FBS’s. We adopted SC and SIC for multicast data and investigated how to assign transmit powers for the packet levels. The objective was to minimize the total BS power consumption, while guaranteeing successful decoding of the multicast data at each user. We developed optimal and near-optimal algorithms with low computational complexity, as well as performance bounds. The algorithms were evaluated with simulations and are shown to outperform a heuristic with considerable gains. Next, we investigated the problem of streaming multiple MGS videos in a femtocell CR network. We formulated a multistage stochastic programming problem considering various design factors across multiple layers. We developed a distributed algorithm that can produce optimal solutions in the case of non-interfering FBS’s, and a greedy algorithm for near-optimal solutions in the case of interfering FBS’s with a proved lower bound. The proposed algorithms are evaluated with simulations and are shown to outperform three alternative schemes with considerable gains.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The ultimate goal of quantum error correction is to achieve the fault-tolerance threshold beyond which quantum computers can be made arbitrarily accurate. This requires extraordinary resources and engineering efforts. We show that even without achieving full fault-tolerance, quantum error detection is already useful on the current generation of quantum hardware. We demonstrate this by executing an end-to-end chemical calculation for the hydrogen molecule encoded in the \[\[4, 2, 2\]\] quantum error-detecting code. The encoded calculation with logical qubits significantly improves the accuracy of the molecular ground-state energy.' author: - Miroslav Urbanek - Benjamin Nachman - Wibe Albert de Jong bibliography: - 'main.bib' title: Quantum error detection improves accuracy of chemical calculations on a quantum computer --- Introduction ============ Predicting chemical properties from first principles is a notoriously hard problem. Quantum computing promises efficient methods for such calculations that can reach far beyond the abilities of classical computers [@Reiher:2017]. Large-scale calculations on quantum computers will require an ability to detect and correct errors. However, near-term devices known as noisy intermediate-scale quantum (NISQ) computers [@Preskill:2018] are not expected to be fully fault-tolerant. Despite this limitation, they can still be useful for solving certain problems in physics and chemistry. In particular, the variational quantum eigensolver (VQE) [@Peruzzo:2014; @McClean:2016] is an algorithm designed to work well on NISQ computers. It has been experimentally demonstrated that VQE is able to find the ground state as well as excited states of small quantum systems encountered in quantum chemistry and nuclear physics [@Peruzzo:2014; @OMalley:2016; @Kandala:2017; @Shen:2017; @Colless:2018; @Dumitrescu:2018; @Hempel:2018; @Ganzhorn:2019; @Kokail:2019]. The performance of NISQ algorithms is currently limited by gate errors and device noise. Several novel error mitigation and suppression techniques have been developed to overcome the imperfections of real devices [@Li:2017; @Temme:2017; @McClean:2017; @BonetMonroig:2018; @Endo:2018; @McArdle:2018; @Endo:2019; @Kandala:2019; @McClean:2019; @Otten:2019a; @Otten:2019b; @Sagastizabal:2019]. Quantum error correction (QEC) is a theory developed in the last two decades to address this problem in a systematic way. The ultimate goal of QEC is to achieve the fault-tolerance threshold. Fault-tolerance requires a large number of qubits, long coherence times, and low gate errors [@Nielsen:2010; @Devitt:2013; @Terhal:2015; @Campbell:2017]. However, QEC can still be useful even without achieving the fault-tolerance threshold and even with only a small number of qubits [@Gottesman:2016; @Chao:2018a; @Chao:2018b]. QEC can potentially increase coherence times and reduce error rates in existing devices. There have been efforts to demonstrate that quantum circuits using QEC codes can improve accuracy, or at least break even, in comparison with the original circuits. Previous experiments studied the traditional three-qubit quantum code [@Reed:2012] and also demonstrated necessary improvements in qubit and gate qualities for QEC [@Barends:2014; @Kelly:2015; @Wootton:2018]. There has also been a growing interest in studying the \[\[4, 2, 2\]\] quantum code [@Linke:2017; @Takita:2017; @Roffe:2018; @Vuillot:2018; @Willsch:2018; @Harper:2019]. All these efforts have tested individual steps of QEC protocols separately. QEC has never been demonstrated to provide a tangible benefit in practical applications. In this work we demonstrate that QEC provides an improvement in accuracy in an end-to-end quantum-chemical calculation. We have implemented a two-qubit VQE algorithm for calculating the ground-state energy of the hydrogen dimer in the \[\[4, 2, 2\]\] QEC code [@Vaidman:1996; @Grassl:1997; @Devitt:2013]. Instead of two physical qubits, the calculation uses two logical qubits encoded in four physical qubits. The code facilitates detection of a single bit-flip and phase-flip error in either of the two logical qubits. Our circuit additionally uses two ancillary qubits to perform a syndrome measurement during the initial state preparation and to perform a logical qubit rotation. Using the \[\[4, 2, 2\]\] code goes a step beyond the experiments with the three-qubit quantum code that encodes only a single logical qubit and cannot simultaneously correct a bit-flip and a phase-flip error [@Nielsen:2010; @Devitt:2013]. We first describe the problem being solved, briefly summarize the VQE algorithm, and present a two-qubit circuit implementing the algorithm. We then discuss the \[\[4, 2, 2\]\] code and show how to implement the required logical gates. This leads to a six-qubit encoded circuit. Our numerical simulations predict that the encoded circuit should outperform the physical circuit up to a fairly large error rate. We then discuss the implementation of both the two-qubit and the six-qubit circuit on the IBM Q Experience platform and show that the encoded circuit improves the accuracy of the potential energy curve of the molecule. Quantum algorithm ================= Finding the ground-state energy of the $\mathrm{H}_2$ molecule in the minimal basis is the simplest molecular electronic-structure problem. It is often used as a benchmark to investigate the performance of quantum-chemical algorithms on quantum computers. The respective molecular Hamiltonian can be transformed into a qubit Hamiltonian using the Jordan–Wigner [@Jordan:1928], Bravyi–Kitaev [@Bravyi:2002], or another similar transformation. Here we use an explicit transformation defined in Ref. [@Colless:2018] that maps the subspace of the Hamiltonian corresponding to two electrons with zero total spin to a two-qubit Hamiltonian. The transformed Hamiltonian is given by $$\label{hamiltonian} H = g_1 + g_2 Z_1 + g_3 Z_2 + g_4 Z_1 Z_2 + g_5 X_1 X_2,$$ where $X_i$, $Y_i$, and $Z_i$ denote Pauli operators acting on qubit $i$ and $g_j$ are classically-calculated coefficients that depend on the internuclear separation $R$. We use values of $g_j$ published in Ref. [@Colless:2018]. The VQE algorithm performs particularly well for this problem. It is a hybrid quantum-classical algorithm that uses a quantum computer to create and measure the properties of a parametrized trial wavefunction and a classical computer to optimize the wavefunction parameters. Our trial wavefunction is the unitary coupled-cluster (UCC) ansatz [@Bartlett:1989; @Taube:2006; @OMalley:2016; @Colless:2018; @Hempel:2018] $$\label{ansatz} {\lvert \psi(\theta) \rangle} = e^{-i \theta X_1 X_2 / 2} {\lvert \Phi \rangle},$$ where $\theta$ is a parameter and ${\lvert \Phi \rangle} = {\lvert 00 \rangle}$ is the Hartree–Fock wavefunction. The wavefunction energy is given by $$\label{energy} E(\theta) = g_1 + g_2 {\langle Z_1 \rangle}_\theta + g_3 {\langle Z_2 \rangle}_\theta + g_4 {\langle Z_1 Z_2 \rangle}_\theta + g_5 {\langle X_1 X_2 \rangle}_\theta,$$ where ${\langle O \rangle}_\theta = {\langle \psi(\theta) \rvert} O {\lvert \psi(\theta) \rangle}$. VQE uses a quantum computer to estimate the expectation values included in $E(\theta)$ and a classical optimizer to find the value of $\theta$ that minimizes $E(\theta)$. Since our ansatz depends on a single parameter only, we sample the full domain of $\theta$ and use a peak-finding routine to minimize $E(\theta)$. It is then sufficient to sample the individual expectation values in Eq. only once and use the same data with any set of coefficients $g_j$. A quantum circuit that implements VQE is shown in the top of Fig. \[circuits\]. ![image](circuits) Error-detecting code ==================== Our goal is to compare the performance of a circuit implemented with physical qubits to a circuit implemented with logical qubits of the \[\[4, 2, 2\]\] code. This code maps two logical qubits into a subspace of four physical qubits as $$\label{code} \begin{aligned} \overline{{\lvert 00 \rangle}} & = \frac{1}{\sqrt{2}} \left( {\lvert 0000 \rangle} + {\lvert 1111 \rangle} \right), \\ \overline{{\lvert 01 \rangle}} & = \frac{1}{\sqrt{2}} \left( {\lvert 0011 \rangle} + {\lvert 1100 \rangle} \right), \\ \overline{{\lvert 10 \rangle}} & = \frac{1}{\sqrt{2}} \left( {\lvert 0101 \rangle} + {\lvert 1010 \rangle} \right), \\ \overline{{\lvert 11 \rangle}} & = \frac{1}{\sqrt{2}} \left( {\lvert 0110 \rangle} + {\lvert 1001 \rangle} \right), \\ \end{aligned}$$ where an overline denotes a logical wavefunction. This mapping allows for the detection of one single-qubit error. To implement the circuit we have to construct the required logical gates from the set of available physical gates. Our set of physical gates is limited to arbitrary single-qubit gates and $\mathit{CNOT}$ gates between any pairs of physical qubits. The encoded circuit is shown in the bottom of Fig. \[circuits\]. Its first part is a preparation of the initial logical state $\overline{{\lvert 00 \rangle}}$. The circuit uses an ancilla measurement to detect an error during the preparation [@Gottesman:2016]. The measurement outcome zero corresponds to no error while the outcome one signals an error. Some logical gates can be implemented easily because the corresponding physical gates act transverally, i.e., they can be implemented with only single-qubit physical gates. The \[\[4, 2, 2\]\] code also facilitates a very simple implementation of the logical $\mathit{CNOT}$ gates as $\overline{\mathit{CNOT}}_{12} = \mathit{SWAP}_{12}$ and $\overline{CNOT}_{21} = \mathit{SWAP}_{13}$, where an overline denotes a logical gate and $\mathit{SWAP}_{ij}$ swaps physical qubits $i$ and $j$ [@Harper:2019]. We implement $\mathit{SWAP}_{ij}$ and therefore the $\overline{\mathit{CNOT}}$ gates without performing any physical operation by relabelling the respective qubits. The arbitrary-angle rotation of the first logical qubit $\overline{R_y^1}(\theta)$ cannot be implemented transversally. We apply this gate by entangling the logical qubit with an ancilla and performing a rotation and a measurement on the ancilla. The measurement outcome zero projects the wavefunction onto the rotated logical state. The gate works correctly only if the logical qubit is initially in the $\overline{{\lvert 0 \rangle}}$ state. A general gate would require additional physical gates. Experiment ========== The algorithm can be summarized as follows. We sample the ${\langle Z_1 \rangle}_\theta$, ${\langle Z_2 \rangle}_\theta$, ${\langle Z_1 Z_2 \rangle}_\theta$, and ${\langle X_1 X_2 \rangle}_\theta$ terms for $\theta \in [-\pi, \pi)$ on a quantum computer. The ${\langle Z_1 \rangle}_\theta$, ${\langle Z_2 \rangle}_\theta$, and ${\langle Z_1 Z_2 \rangle}_\theta$ terms can be measured with a single circuit without any basis transformations. However, it is necessary to perform a basis transformation to measure the ${\langle X_1 X_2 \rangle}_\theta$ term. We therefore execute the circuit with $R_t = I$ to measure ${\langle Z_1 \rangle}_\theta$, ${\langle Z_2 \rangle}_\theta$, and ${\langle Z_1 Z_2 \rangle}_\theta$, and with $R_t = H$ to measure ${\langle X_1 X_2 \rangle}_\theta$. We then calculate the ground-state energy for each internuclear separation $R$ by minimizing $E(\theta)$. We ran both the two-qubit logical circuit and the six-qubit encoded circuit on the Tokyo chip on IBM Q Experience. The major errors are readout errors [@Kandala:2017; @Dumitrescu:2018; @YeterAydeniz:2019]. If a qubit is in the ${\lvert 0 \rangle}$ state there is a significant probability of measuring outcome one and vice versa. The readout errors are asymmetric, i.e., the probability of measuring zero when a qubit state is ${\lvert 1 \rangle}$ is higher than the probability of measuring one when the state is ${\lvert 0 \rangle}$. This is mostly due to the readout time being significant in comparison to the $T_1$ coherence time, so the qubit can decay from the ${\lvert 1 \rangle}$ state to the ${\lvert 0 \rangle}$ state during the readout. We employed a readout error correction technique known as unfolding [@Cowan:2002; @Blobel:2013] based on a Bayesian probabilistic model [@Lucy:1974; @Richardson:1972; @DAgostini:1995] to correct the readout errors. We first measured and estimated the probability of each outcome when the qubits were prepared in each computational basis state. We then used this probability matrix to iteratively unfold all measured counts to corresponding true counts. The chip contained 20 qubits arranged in a two-dimensional geometry. There were 72 ways to map our two-qubit physical circuit and 288 ways to map our six-qubit encoded circuit to the chip qubits. We found that the results depended significantly on the chosen qubits and also on the order of the applied gates. The result variability is illustrated in Fig. \[qubits\]. The ${\langle X_1 X_2 \rangle}_\theta$ term is the most sensitive term in Eq. . To find an optimal mapping, we measured ${\langle X_1 X_2 \rangle}_\theta$ for $\theta = -3\pi/4$, $-\pi/2$, $-\pi/4$, $0$, $\pi/4$, $\pi/2$, and $3\pi/4$, applied readout error corrections, and calculated the $\ell_1$ distances between the corrected results and the exact results for each mapping. We used the mappings with the smallest distances to run the final circuits. The compiler reordered commuting gates based on the qubit mapping, so this technique took into account both the qubit mapping and the gate order variability. ![image](qubits) We executed the final calculations for both the two-qubit and the six-qubit circuit using the optimal mappings. The ${\langle Z_1 \rangle}_\theta$, ${\langle Z_2 \rangle}_\theta$, ${\langle Z_1 Z_2 \rangle}_\theta$, and ${\langle X_1 X_2 \rangle}_\theta$ terms were obtained for 257 values of $\theta$ in the $[-\pi, \pi]$ interval. Each measured value was sampled with 8192 shots. For the encoded six-qubit circuit we postselected the outcomes based on their ancilla values. In particular, we measured all six qubits, performed readout error corrections, and discarded outcomes with nonzero ancilla values. The outcomes outside the code space were discarded as well. We renormalized the remaining outcomes and summed counts of constituent basis states in Eq. to calculate the logical state counts. The calculated expectation values of the Hamiltonian terms are shown in Fig. \[terms\]. We then used a peak-finding routine to find $\theta$ that minimized the energy in Eq. for each internuclear separation. The calculated energy potential curves are shown in Fig. \[experiment\]. The results demonstrate that the six-qubit encoded circuit improves the accuracy of the ground-state energy. ![image](terms) ![image](experiment) Discussion ========== In summary, we designed a circuit for a VQE calculation within the \[\[4, 2, 2\]\] quantum error-detecting code space. We executed the circuit on a quantum chip and observed an improvement in the calculated ground-state energy of the $\mathrm{H}_2$ molecule. Our encoded circuit requires more physical qubits and gates than our logical circuit and therefore is more error-prone. However, the gain by using the code was larger than the loss due to the circuit complexity. The results show that quantum error-detection is already useful on NISQ devices even without achieving full fault-tolerance. The presented method can be used in addition to other error mitigation techniques. Our implementation uses two ancillary qubits with postselection on their measured outcomes. In principle, it would be possible to use just one ancilla if we had an ability to perform a qubit reset. Similarly, the postselection in the rotation gate would be unnecessary if we had an ability to apply conditional gates dependent on measurement outcomes. Implementing quantum error detecting and correcting codes tailored to specific applications better utilizes limited resources and extends the possibilities of existing devices. It also informs hardware designers which features will be the most impactful in the next generation of quantum computers. Some of the previous VQE experiments [@OMalley:2016; @Kandala:2017; @Colless:2018; @Hempel:2018; @Kandala:2019] found the ground-state energy of the $\mathrm{H}_2$ molecule with a comparable or better accuracy. They used techniques like higher qubit states measurement [@OMalley:2016], quantum subspace expansion [@Colless:2018], and noise extrapolation [@Kandala:2019] to mitigate errors. We emphasize that our circuits do not use any such techniques. Our QEC method demonstrates that on the same hardware and using the same algorithm, the encoded circuit results in smaller error than the physical circuit. Other error mitigation techniques are complementary to the presented method. We thank Jarrod R. McClean and Mekena Metcalf for helpful discussions. This work is supported by the DOE under contract DE-AC02-05CH11231, by the DOE Office of Advanced Scientific Computing Research (ASCR) through the Quantum Algorithm Team and Quantum Testbed Pathfinder Program, and from Quantum Information Science Enabled Discovery (QuantISED) for High Energy Physics (KA2401032). This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725. Hamiltonian transformation ========================== We use a transformation presented in Ref. [@Colless:2018] to map the electronic-structure space to qubits. The transformed space corresponds to a $\mathrm{H}_2$ molecule with two electrons and zero total spin. In particular, $$\begin{aligned} a_{1\uparrow}^\dagger a_{1\downarrow}^\dagger {\lvert \mathrm{vac} \rangle} & \to {\lvert 00 \rangle}, \\ a_{1\uparrow}^\dagger a_{2\downarrow}^\dagger {\lvert \mathrm{vac} \rangle} & \to {\lvert 01 \rangle}, \\ a_{2\uparrow}^\dagger a_{1\downarrow}^\dagger {\lvert \mathrm{vac} \rangle} & \to {\lvert 10 \rangle}, \\ a_{2\uparrow}^\dagger a_{2\downarrow}^\dagger {\lvert \mathrm{vac} \rangle} & \to {\lvert 11 \rangle}, \\ \end{aligned}$$ where $a_{is}^\dagger$ is an operator that creates an electron with spin $s$ in orbital $i$ and ${\lvert \mathrm{vac} \rangle}$ is the vacuum state. Analytical model ================ We analyze the effect of noise on the calculated ground-state energies using the depolarizing noise model. The noise operation is given by [@Nielsen:2010] $$\epsilon(\rho) = (1-p) \rho + p \frac{I}{2},$$ where $\rho$ is the density matrix and $p$ is the probabilistic error rate. The value of $p = 0$ corresponds to vanishing noise and $p = 1$ corresponds to full noise. We assume that the noise affects only qubits involved in a particular gate application. Separate operations are used for one-qubit gates, $$\epsilon_i(\rho) = (1 - p_1) \rho + \frac{p_1}{4} \sum_{E_i \in \mathcal{P}_i} E_i^\dagger \rho E_i,$$ and for two-qubit gates, $$\epsilon_{i, j} (\rho) = (1 - p_2) \rho + \frac{p_2}{16} \sum_{\substack{E_i \in \mathcal{P}_i \\ E_j \in \mathcal{P}_j}} E_i^\dagger E_j^\dagger \rho E_i E_j,$$ where $\mathcal{P}_i = \{I_i, X_i, Y_i, Z_i\}$ is the set of the unit matrix and the Pauli matrices acting on qubit $i$. The noise operations above are performed on the density matrix after each gate application to a respective set of qubits. We characterize the noise channel with only a single parameter $p$ and use $p_2 = p$ and $p_1 = p/16$ since the single-qubit gates have significantly higher fidelities in hardware. The comparison of the ground-state energy calculated with the noisy logical and encoded circuits is shown in Fig. \[model\]. ![Energy potential curves of the $\mathrm{H}_2$ molecule calculated analytically using the VQE algorithm with the depolarizing noise model. The two-qubit gate error rate is $p = 5\,\%$. The six-qubit encoded circuit performs better for error rates up to about $30\,\%$.[]{data-label="model"}](model) Readout error correction ======================== Correcting measurements of discrete data for readout bias has a long history. For example, in high energy physics experiments, binned differential cross sections are corrected for detector effects in order to compare them with predictions from quantum field theory. In that context, the corrections are called unfolding (sometimes called deconvolution in other fields) and a variety of techniques have been proposed and are in active use [@Cowan:2002; @Blobel:2013]. Quantum readout error correction can be represented as a binned unfolding where each bin corresponds to one of the possible $2^m$ configurations, where $m$ is the number of qubits. We use an iterative Bayesian unfolding technique [@Lucy:1974; @Richardson:1972; @DAgostini:1995]. Given a response matrix $$R_{ij} = \Pr(\text{measure}\ i \,\|\, \text{truth is}\ j),$$ a measured spectrum $m_i = \Pr(\text{measure}\ i)$ and a prior truth spectrum $t_i^0 = \Pr(\text{truth is}\ i)$, the iterative technique proceeds according to an equation $$\label{unfolding} \begin{split} t_i^{n + 1} & = \sum_j \Pr(\text{truth is}\ i \,\|\, \text{measure}\ j) \times m_j \\ & = \sum_j \frac{R_{ji} t_i^n}{\sum_k R_{jk} t_k^n} \times m_j, \end{split}$$ where $n$ is the iteration number. The advantage of Eq. over simple matrix inversion is that the result is a probability (nonnegative and unit measure). We construct $R_{ij}$ by preparing $2^m$ calibration circuits where each qubit computational state is constructed with $X$ gates. The entries of $R_{ij}$ are the fraction of measurements that qubit configuration $i$ is observed in configuration $j$. We use a uniform distribution as the initial spectrum $t_i^0$. The iterative procedure described in Eq. is repeated until convergence. The effect of readout error correction on potential energy curves is shown in Fig. \[raw\]. ![\[raw\] Comparison of potential energy curves obtained with raw measurement outcomes and with outcomes corrected for readout errors for both the two-qubit and six-qubit circuits.](raw) Qubit mappings ============== The availability of qubits and their connections has changed during the data collection. The final data in Fig. \[experiment\] were collected after a connection between qubits three and nine was turned off. Additionally, qubit seven was not available during experiments with the six-qubit circuit. As a result, there were only 70 and 116 possible mappings from the abstract qubits to the physical qubits for the two-qubit and the six-qubit circuits, respectively.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the average method, we derive a closed set of linear equations that describes the spinning up of an harmonically trapped gas by a rotating anisotropy. We find explicit expressions for the time needed to transfer angular momentum as well as the decay time induced by a static residual anisotropy. These different time scales are compared with the measured nucleation time and lifetime of vortices by the ENS group [@VortexENS]. We find a good agreement that may emphasize the role played by the non-condensed component in those experiments.' address: - 'Laboratoire Kastler Brossel, Département de Physique de l’Ecole Normale Supérieure' - '24, Rue Lhomond, F-75231 Paris Cedex 05, France' author: - 'D. Guéry-Odelin' title: Spinning up and down a Boltzmann gas --- Introduction {#introduction .unnumbered} ============ Superfluidity of a Bose-Einstein condensate is naturally investigated by its rotational properties [@Baym; @scissors]. So far, two experimental schemes have been successfully applied to gaseous condensates in order to generate quantized vortices [@VortexJila; @VortexENS]. The first one uses phase engineering by means of laser beams whereas the second one is the analog of the “rotating bucket" experiment, initially suggested by Stringari [@STRINGARI96; @Sandro2]. In the latter, atoms are first confined in a static, axially symmetric Ioffe-Pritchard magnetic trap upon which a non-axisymmetric attractive dipole potential is superimposed by means of a strirring laser beam. In this paper, we address the question of the behaviour of the non-condensed component for this experiment. We investigate the transfer of angular momentum to an ultracold harmonically confined gas by such a time dependent potential. We also study possible mechanisms for its dissipation. First (Sec. I), we recall the Lagragian and the Hamiltonian formalism for a single particle in a rotating frame. Then in Sec. II, we give the expression for the rotating potential that has been used in our study. We briefly expose analytical results for the single particle trajectory in this potential in Sec. III. The rest of the paper deals with the crucial role played by collisions. A classical gas that evolves in this potential thermalizes in the rotating frame, leading to a finite value of its mean angular momentum. We investigate this equilibrium state in Sec. IV. In Sec. V, we derive the analytical expression for the time needed to spin up a classical gas with an approach based on the average method [@AvM]. We evaluate the time needed to transfer angular momentum and thus we deduce the characteristic time for vortex nucleation [via]{} angular momentum transfer from the uncondensed to the condensed component. Finally, in Sec. VI, we consider a related problem: what is the time needed to dissipate a given angular momentum by a static residual anisotropy ? We have in mind the role of the axial asymmetry of a magnetic trap (Ioffe-Pritchard or time-orbiting potential traps) induced by the presence of gravity. We show that this effect may explain the finite lifetime of vortices. A Reminder on the rotating frame ================================ In this section, we recall the hamiltonian for a classical particle in a rotating frame characterized by the fixed rotation vector ${\bf \Omega}$. Without loss of generality, we choose a rotation around the $z$ axis ${\bf \Omega}=\Omega {\bf e}_z$ and the same origin for the rotating frame $\Re'$ as the one of the laboratory frame $\Re$ (see Fig. 1). In the following, quantities with a prime are evaluated in the rotating frame. Coordinates in $\Re'$ are linked to coordinates in $\Re$ just by a rotation of angle $\theta=\Omega t$: $$\left( \begin{array}{c} x' \\ y' \end{array} \right)= \left( \begin{array}{cl} \;\;\;\cos \Omega t & \sin \Omega t \\ -\sin \Omega t & \cos\Omega t \end{array} \right) \left( \begin{array}{c} x \\ y \end{array} \right) \nonumber$$ The lagrangian for the single particle movement in the rotating frame reads [@Landau1]: $$\begin{aligned} {\cal L}'=\frac{1}{2}mv'^2+\frac{1}{2}m({\bf \Omega}\times {\bf r})^2 %\nonumber\\ + m{\bf v}'.({\bf \Omega}\times {\bf r}) -V_{\rm ext}, \label{lagr}\end{aligned}$$ where $V_{\rm ext}$ is the potential energy that describes the role of external forces. The correspondance between the laboratory and rotating frames for momentum, angular momentum and the hamiltonian is given by: $$\begin{aligned} {\bf p}' & = & \frac{\partial {\cal L}'}{\partial {\bf v}'}=m{\bf v}'+m{\bf \Omega} \times {\bf r}={\bf p} \nonumber\\ {\bf L}' & = & {\bf r} \times {\bf p'}={\bf r}\times {\bf p}={\bf L} \nonumber\\ {\cal H}' & = & {\bf p}'.{\bf v}'-{\cal L}'={\cal H}-{\bf \Omega}.{\bf L} \label{srot}\end{aligned}$$ Note that the angular momentum as well as the momentum are the same in both frames, but the link between the momentum and the velocity differs. Using the hamiltonian formalism, one can easily extract the equation of motion in the rotating frame: $$m\frac{d{\bf v}'}{dt}={\bf F}_{\rm ext}+{\bf F}_{\rm cen}+{\bf F}_{\rm cor}$$ where ${\bf F}_{\rm ext}$ refers to the external force derived from the potential $V_{\rm ext}$, ${\bf F}_{\rm cen}$ is the centrifugal force: $${\bf F}_{\rm cen}=-m{\bf \Omega}\times ({\bf \Omega}\times {\bf r})=- \mbox{\boldmath $\nabla$}V_{\rm cen} % -m{\dot {\bf \Omega}}\times {\bf r}$$ with $V_{\rm cen}=-m\Omega^2(x^2+y^2)/2$, and ${\bf F}_{\rm cor}$ is the Coriolis force: $${\bf F}_{\rm cor}=-2m{\bf \Omega}\times {\bf v}'.$$ Note that all formulas of the system (\[srot\]) are still valid even for a time-dependent rotation vector ${\bf \Omega}$. The rotating trap ================= In Bose-Einstein condensation experiments, the magnetic confinement is axially symmetric. In order to spin up the system, one possibility consists of breaking this symmetry by superimposing a small rotating anisotropy as initially suggested by S. Stringari [@STRINGARI96]. This breaking of the rotational invariance of the external potential can be carried out experimentally by adding a rotating stirring dipolar beam to the magnetic field of the trap [@VortexENS]. This combination of light and magnetic trapping induces the following harmonic external potential (see Fig. 1) [@Oxford]: $$V_{\rm ext}=\frac{m\omega_0^2}{2}\bigg((1+\epsilon)x'^2+(1-\epsilon)y'^2+\lambda^2 z^2\bigg) \label{vext}$$ where we have defined the geometric parameter $\lambda=\omega_z/\omega_0$, which is responsible for the shape of the cloud. For instance, if we set $\epsilon=0$, the trap is isotropic for $\lambda=1$, cigar-shaped for $\lambda \ll 1$, and disk-shaped for $\lambda \gg 1$. The potential (\[vext\]) is time-dependent if expressed with laboratory coordinates $(x,y,z)$, but static in the rotating frame, [i.e.]{} in terms of $(x',y',z)$. Single particle trajectory ========================== Let us first investigate the single particle trajectory. Equations in the laboratory frame and for the transverse coordinates are given by: $$\left( \begin{array}{c} \ddot{x} \\ \ddot{y} \end{array} \right)=-\omega_0^2 \left( \begin{array}{c} x \\ y \end{array} \right) -\epsilon\omega_0^2 \left( \begin{array}{cl} \cos 2\Omega t & \;\;\;\sin 2\Omega t \\ \sin 2\Omega t & -\cos 2\Omega t \end{array} \right) \left( \begin{array}{c} x \\ y \end{array} \right) \nonumber$$ In the rotating frame, the corresponding equations are time independent. They can be rewritten by means of the complex quantity: $\xi'=x'+iy'$: $$\ddot{\xi}'+2\Omega i\dot{\xi}'+(\omega_0^2-\Omega^2)\xi'+\epsilon \omega_0^2\bar{\xi}'=0 \label{eqz}$$ where $\bar{\xi}$ denotes the complex conjugate of $\xi$. The second term in Eq. (\[eqz\]) accounts for the Coriolis force, the centrifugal force leads to a reduction of the harmonic strength (third term), and the last term is the contribution of the asymmetry. The stability of the single particle movement is extracted from the dispersion relation of (\[eqz\]). We find a window of instability around the value $\Omega=\omega_0$, [i.e.]{} for $\Omega\in[\omega_0\sqrt{1-\epsilon}; \omega_0\sqrt{1+\epsilon}]$. For $\Omega <\omega_0\sqrt{1-\epsilon}$, the stability is essentially ensured by the harmonic trapping even if reduced by the centrifugal force, whereas for $\Omega >\omega_0\sqrt{1+\epsilon}$ the Coriolis force plays a crucial role in the stabilization of the trajectory. The latter effect is similar to magnetron stabilization in a Penning trap for ions [@magnetron]. Equilibrium state of the gas ============================ In practice, experiments are carried out in the so-called collisionless regime, since on average an atom undergoes less than a collision during a transverse oscillation period. However, collisions are of course essential to explain the dynamics of the gas induced by the potential (\[vext\]). We consider the situation in which a gas at a given temperature $T_0$ is initially at rest in the lab frame, and at $t=0$ the axial asymmetry $\epsilon$ is spinned up at a constant angular velocity ${\bf \Omega}$ as explained above. If $\Omega <\omega_0\sqrt{1-\epsilon}$ one expects that elastic collisions will ensure the thermalization of the gas in the rotating frame. This equilibrium state is defined since a minimum of the effective potential $V_{\rm eff}=V_{\rm ext}+V_{\rm cen}$ always exists in this range of values for $\Omega$. In Fig. 2, we compare the stability of a single particle (upper part of the diagram) with that of the interacting gas. In the rotating frame, one can compute equilibrium quantities by means of the Gibbs distribution $\rho$ which reads [@Landau5]: $$\rho ({\bf r}',{\bf v'}) \propto e^{-{\cal H}'({\bf r}',{\bf v'})/k_BT_0} \label{gd}$$ where ${\cal H}'$ is given by: $${\cal H}'=\frac{mv^{\prime 2}}{2}+V_{\rm ext}(x',y',z)-\frac{m\Omega^2}{2}(x'^2 +y'^2)$$ As regards statistical properties of the gas, the rotation is equivalent to a reduction of the effective transverse frequencies of the trap due to the contribution of the centrifugal force. The Coriolis force plays no role for the equilibrium state. During the thermalization, the mean angular momentum per particle increases from zero to its asymptotic value $\langle L_z\rangle$. This last quantity is easily derived from the Gibbs distribution (\[gd\]): $$\langle L_z\rangle=m\Omega\langle x^2+y^2 \rangle= \frac{2k_BT_0\Omega(\omega_0^2-\Omega^2)}{(\omega_0^2-{\Omega}^2)^2 -\epsilon^2\omega_0^4} \label{eqlz}$$ For typical experimental parameters $\omega_0/2\pi=200$ Hz, $\Omega=\omega_0/2$, $\epsilon=0.05$ and $T_0=1$ $\mu$K, the angular momentum per particle $\langle L_z\rangle /\hbar\simeq 150 $ unlike the superfluid part for which for instance the angular momentum is equal to $\hbar$ per particle in the presence of one vortex. To check that the gas undergoes a full rotation, one can search for a displacement of the critical temperature induced by centrifugal forces [@Sandro2], or perform a time-of-flight measurement. In the latter case, one expects that the ratio between $x$ and $z$ size scales as $({1-\Omega^2/\omega_0^2})^{-1/2}$ for long time expansion. Time needed to reach equilibrium ================================ In this section we derive the expression for the time $t_{\rm up}$ needed to reach the equilibrium state in the rotating coordinate system. In other words, $t_{\rm up}$ corresponds to the time needed to build correlations between $x$, $v_y$, $y$ and $v_x$. To estimate this time, we use the classical Boltzmann equation. Our analysis relies on the use of the average method as explained in [@AvM]. For instance, the equation for $\langle x'v_{y'}-y'v_{x'}\rangle$ involves $\langle x'y'\rangle$ which itself is coupled to $\langle x'v_{y'}+y'v_{x'}\rangle$ and so on. Terms that do not correspond to a conserved quantity in a binary elastic collision lead to a non zero contribution of the collisional integral. For instance, a mean value such as $\langle v_{x'}v_{y'}\rangle$ involves the occurence of quadrupole deformations in the velocity distribution which make the contribution $\langle v_{x'}v_{y'}I_{\rm coll}\rangle\neq 0$. Here, $I_{\rm coll}$ stands for the collisional kernel of the Boltzmann equation [@Huang]. At this stage, we perform a gaussian ansatz for the distribution function. After linearization this quadrupolar contribution results in the so-called relaxation time approximation [@Huang; @Smith]: $$\langle v_{x'}v_{y'}I_{\rm coll}\rangle = -\frac{\langle v_{x'}v_{y'}\rangle}{\tau}$$ This method leads to a closed set of $13\times 13$ linear equations (see Appendix), and provides furthermore an explicit link between the relaxation time $\tau$ and the collisional rate in the sample. It is worth emphasizing that the dynamic transfer of angular momentum to a classical gas by rotating a superimposed axial anisotropy involves a coupling between all quadrupole modes. The average method is fruitful in the sense that it yields a closed set of equations when one deals with only quadratic moments, namely: monopole mode, scissors mode, quadrupole modes, ... Non-inertial forces are linear in position or velocity, and thus give rise only to quadratic moments using the average method. Although unimportant for equilibrium properties, the Coriolis force plays a crucial role for reaching equilibrium. Fig. 3 depicts a typical thermalization of the gas in the rotating frame, leading to the equilibrium value (\[eqlz\]) of the angular momentum, obtained by a numerical integration of the $13\times 13$ set of equations. After a tedious but straightforward expansion of the dispersion relation, one extracts the smallest eigenvalue that drives the relaxation for a given collisional regime. First we focus on the regime in which experiments are performed: collisionless ($\omega_0\tau\simeq 10$) and with $\epsilon^2 \ll \Omega^2/\omega_0^2$. Denoting $t^{\rm CL}_{\rm up}$ as the time needed to spin up the gas in this regime, we find: $$t^{\rm CL}_{\rm up}= \frac{8\tau}{\epsilon^2}\left( \frac{\Omega}{\omega_0}\right)^2$$ The result is independent of the geometrical aspect ratio of the trap $\lambda$ as physically expected. Using the same numerical values as in the previous, we deduce that this time is very long $t^{\rm CL}_{\rm up}\simeq 15$ s. Note that to transfer just $\hbar$ of angular momentum per particle, one needs only 100 ms. This time is on the order of the nucleation time for vortices that has been experimentally observed by the ENS group [@nuclea]. One should nevertheless be carefull since the non-condensed component is actually a Bose gas rather than a classical one that evolves in a non harmonic potential because of the mean field potential due to the condensed component. In the hydrodynamic limit, the characteristic time for spinning up is: $$t^{\rm HD}_{\rm up}=1/(2\epsilon^2\omega_0^2\tau).$$ So far this regime is not accessible for ultracold atom experiments since inelastic collisions prevent the formation of very high density samples. We recover here the special feature of the hydrodynamic regime, [i.e.]{} the time needed to reach equilibrium increases with the collisional rate. In Fig. 4 we have reported the evolution of $t_{\rm up}$ as a function of $\omega_0\tau$ from a numerical integration of the $13\times 13$ system. The smallest value is obtained between the collisionless and hydrodynamic regimes, as is usual for the relaxation of a thermal gas [@AvM]. Time needed to dissipate a given angular momentum ================================================= Consider a gas with a given angular momentum, obtained for instance as explained before. If this gas evolves in an axially symmetric trap, the angular momentum is a conserved quantity. In contrast, if a small asymmetry $\epsilon$ exists between the $x$ and $y$ spring constants, the angular momentum is no longer a conserved quantity and it is thus dissipated. We call $t_{\rm down}$ the typical time for the relaxation of the angular momentum. This problem is very different from the one we faced previously since the rotating frame is not an inertial frame. We thus expect $t_{\rm down}\neq t_{\rm up}$. As in the previous treatment, this problem can be computed with the average method. The corresponding equations are nothing but the scissors mode equations [@scissors], [i.e.]{} a linear set of 4 equations involving $\langle xy\rangle$, $\langle xv_y-yv_x\rangle$, $\langle xv_y+yv_x\rangle$ and $\langle v_xv_y\rangle$ (see Appendix). Searching a solution of this system of the form $\exp(-\lambda t)$, one finds $$\lambda=\frac{1}{4\tau}\bigg(1-\sqrt{1-\epsilon^2/\epsilon_c^2}\bigg), \label{tthdown}$$ where the critical anisotropy is related to $\tau$ and $\omega_0$ by $\epsilon_c=1/(4\omega_0\tau)$. In Fig. (5), we plot different curves for the relaxation of angular momentum depending on the value of $\epsilon$ with respect to $\epsilon_c$. For $\epsilon<\epsilon_c$ (long dashed line), one has a purely damped relaxation; in the limiting case $\epsilon \ll \epsilon_c$, $t_{\rm down}\simeq 1/(2\epsilon^2\omega_0^2\tau)$. On the contrary for $\epsilon> \epsilon_c$ (solid line), one has a damped oscillating behaviour; in the limiting case $\epsilon \gg \epsilon_c$, $t_{\rm down}\simeq 4\tau$, and the oscillating frequency is equal to $\epsilon\omega_0$. Moreover, as we use a linear analysis, this decay time does not depend on the specific initial value of the angular momentum. For the experiment described in [@VortexENS], the axial asymmetry induced by gravity is on the order of 1% and the collisional rate is such that $\omega_0\tau\simeq 10$. The corresponding decay time of the angular momentum is then evaluated from Eq. (\[tthdown\]) to be $\sim$ 500 ms. This value matches surprisingly well the typical observed lifetime of a vortex. Therefore, one possible interpretation may be that the thermal part acts as a reservoir of angular momentum and thus ensures the stabilization of the vortex as long as this thermal part is itself rotating significantly. On the opposite when the angular momentum of the thermal part is zero, the vortices disappear. Conclusion {#conclusion .unnumbered} ========== In this paper we have investigated the dynamic of transfer and dissipation of angular momentum for a classical gas and in all collisional regimes. We derive very different time scales for both processes. The short time deduced to dissipate angular momentum suggests a high sensitivity of the experiment described in [@VortexENS] to residual static anisotropy (see different time scales between Figs. (3) and (5)). In practice, as underlined above one cannot avoid residual fixed anisotropy due for instance to gravity. The competition between a rotating and a static anisotropy may explain why no evidence for a full rotation of the classical gas was experimentally reported in Ref [@VortexENS]. acknoledgements {#acknoledgements .unnumbered} =============== I acknowledge fruitful discussions with V. Bretin, F. Chevy, J. Dalibard, K. Madison, A. Recati, S. Stringari and F. Zambelli. I am grateful to Trento’s BEC group where part of this work was carried out. [Rotating frame with respect to laboratory frame.]{} \[frame\] [Stability diagram for the single particle and for an interacting gas in the potential with a rotating anisotropy.]{} \[stab\] [Mean angular momentum per particle as a function of time.]{} \[lup\] [Time needed to reach equilibrium by rotation of a small anisotropy. A minimum of $t_{\rm up}$ is obtained in-between the collisionless and the hydrodynamic regimes.]{} \[tup\] [Decay of angular momentum due to a small residual anisotropy. We distinguish three different regimes: $\epsilon>\epsilon_c$ damped oscillations (solid line), $\epsilon=\epsilon_c$ (small dashed line) the frontier with the region of purely damped decay ($\epsilon<\epsilon_c$, long dashed line).]{} \[ldown\] \[ldown\] APPENDIX {#A} ======== Hereafter, the 13$\times$13 closed set of equations that describes the dynamic of a classical gas induced by the rotating potential (\[vext\]): $$\begin{aligned} \frac{d\langle x'y'\rangle}{dt}-\langle x'v'_{y'}+y'v'_{x'}\rangle=0 \\ \frac{d\langle x'^2-y'^2\rangle}{dt}-2\langle x'v'_{x'}-y'v'_{y'}\rangle=0 \\ \frac{d\langle x'^2+y'^2+z'^2\rangle}{dt}-2\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle=0 \\ \frac{d\langle x'^2+y'^2-2z'^2\rangle}{dt}-2\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle=0 \\ \frac{d\langle x'v_{y'}-y'v_{x'}\rangle}{dt}-2\omega_0^2\epsilon \langle x'y'\rangle \nonumber\\ +\frac{4\Omega}{3}\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle \nonumber\\ +\frac{2\Omega}{3}\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle=0 \\ \frac{d\langle x'v_{y'}+y'v_{x'}\rangle}{dt}+2(\omega_0^2-\Omega^2)\langle x'y'\rangle \nonumber\\ +2\Omega \langle x'v_{x'}-y'v_{y'}\rangle-2\langle v_{x'}v_{y'}\rangle=0 \\ \frac{d\langle x'v_{x'}-y'v_{y'}\rangle}{dt}+(\omega_0^2-\Omega^2)\langle x'^2-y'^2\rangle \nonumber\\ +\frac{2\omega_0^2\epsilon}{3}\langle x'^2+y'^2+z'^2\rangle-2\Omega \langle x'v_{y'}+y'v_{x'}\rangle \nonumber\\ +\frac{\omega_0^2\epsilon}{3}\langle x'^2+y'^2-2z'^2\rangle-\langle v_{x'}^2-v_{y'}^2\rangle=0 \\ \frac{d\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle}{dt}+\epsilon\omega_0^2 \langle x'^2-y'^2\rangle \nonumber\\ +\frac{2-2\Omega^2+\lambda^2}{3}\omega_0^2\langle x'^2+y'^2+z'^2\rangle \nonumber\\ +\frac{1-\Omega^2-\lambda^2}{3}\omega_0^2\langle x'^2+y'^2-2z'^2\rangle \nonumber\\ -\langle v_{x'}^2+v_{y'}^2+v_{z'}^2\rangle-2\Omega \langle x'v_{y'}-y'v_{x'}\rangle=0 \\ \frac{d\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle}{dt}+\epsilon\omega_0^2 \langle x'^2-y'^2\rangle \nonumber\\ +\frac{2\omega_0^2}{3}(1-\lambda^2-\Omega^2)\langle x'^2+y'^2+z'^2\rangle \nonumber\\ +\frac{\omega_0^2}{3}(1+2\lambda^2-\Omega^2)\langle x'^2+y'^2-2z'^2\rangle \nonumber\\ -\langle v_{x'}^2+v_{y'}^2-2v_{z'}^2\rangle-2\Omega \langle x'v_{y'}-y'v_{x'}\rangle=0 \\ \frac{d\langle v_{x'}v_{y'}\rangle}{dt}+\epsilon\omega_0^2 \langle x'v_{y'}-y'v_{x'}\rangle+2\Omega \langle v_{x'}^2-v_{y'}^2\rangle \nonumber\\ +(\omega_0^2-\Omega^2)\langle x'v_{y'}+y'v_{x'}\rangle=-\frac{\langle v_{x'}v_{y'}\rangle}{\tau} \\ \frac{d\langle v_{x'}^2-v_{y'}^2\rangle}{dt}+2(\omega_0^2-\Omega^2)\langle x'v_{x'}-y'v_{y'}\rangle \nonumber\\ -8\Omega \langle v_{x'}v_{y'}\rangle+\frac{4\epsilon\omega_0^2}{3}\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle+ \nonumber\\ \frac{2\epsilon\omega_0^2}{3}\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle =-\frac{\langle v_{x'}^2-v_{y'}^2\rangle}{\tau} \\ \frac{d\langle v_{x'}^2+v_{y'}^2+v_{z'}^2\rangle}{dt}+2\epsilon\omega_0^2 \langle x'v_{x'}-y'v_{y'}\rangle \nonumber\\ +\frac{2\omega_0^2}{3}(2+\lambda^2-2\Omega^2)\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle+ \nonumber\\ \frac{2\omega_0^2}{3}(1-\lambda^2-\Omega^2)\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle=0 \\ \frac{d\langle v_{x'}^2+v_{y'}^2-2v_{z'}^2\rangle}{dt}+2\epsilon\omega_0^2 \langle x'v_{x'}-y'v_{y'}\rangle \nonumber\\ +\frac{4\omega_0^2}{3}(1-\lambda^2-\Omega^2)\langle x'v_{x'}+y'v_{y'}+z'v_{z'}\rangle \nonumber\\ +\frac{2\omega_0^2}{3}(1+2\lambda^2-\Omega^2)\langle x'v_{x'}+y'v_{y'}-2z'v_{z'}\rangle \nonumber\\ =-\frac{\langle v_{x'}^2+v_{y'}^2-2v_{z'}^2\rangle}{\tau}\end{aligned}$$ One can show from the gaussian ansatz that the relaxation time $\tau$ is the same for all quadrupolar contributions. In this system, linear terms in $\Omega$ account for the Coriolis force whereas quadratic terms in $\Omega$ are due to centrifugal force contributions. All quadrupolar modes are involved in this system. In order to enlight the physics of this system, let us consider some limiting cases. One can check the conservation of energy in the rotating frame $d\langle{\cal H}'\rangle/dt=0$. The stationary state, obtained by setting all time derivatives of moments to zero, is nothing but the equipartition law (see [@Landau5], §44): $(1-\Omega^2+\epsilon)\langle x'^2\rangle=(1-\Omega^2-\epsilon)\langle y'^2\rangle=\lambda^2\langle z^2\rangle =\langle v_{x'}^2\rangle=\langle v_{y'}^2\rangle=\langle v_{z'}^2\rangle$. If $\epsilon=0$ and $\Omega=0$, the trap is axially symmetric and one recovers the conservation of the angular momentum. In addition the system gives rise to 3 independent linear systems on quadrupolar quantities: one for the $m=2,\langle xy\rangle$ mode, another one for $m=2,\langle x^2-y^2\rangle$, and finally the $m=0$ mode that describes the coupling between monopole ($\langle r^2\rangle$) and quadrupole mode ( $m=0,\langle x^2+y^2-2z^2\rangle$) [@AvM]. Finally, one can recover the scissors mode equations for a classical gas by considering the case $\Omega=0$ and $\epsilon\neq 0$ [@scissors]: eqs A1, A5, A6, A10. G. Baym, in [Mathematical Methods in Solid State and Superfluidity Theory]{}, edited by R.C. Clark and E.H. Derrick (Oliver and Boyd, Edimburgh, 1969). D. Guéry-Odelin and S. Stringari, Phys. Rev. Lett. [83]{}, 4452 (1999). M.R. Matthews, B.P. Anderson, P.C. Haljan, C.E. Wieman, and E.A. Cornell, Phys. Rev. Lett. [83]{}, 2498 (1999). K. Madison, F.Chevy, W. Wohlleben and J. Dalibard, Phys. Rev. Lett. [84]{}, 806 (2000). S. Stringari, Phys. Rev. Lett. [76]{}, 1405 (1996). S. Stringari, Phys. Rev. Lett. [82]{}, 4371 (1999). D. Guéry-Odelin, F. Zambelli, J. Dalibard, and S. Stringari, Phys. Rev. A [60]{}, 4851 (1999). L.D. Landau and E.M. Lifshitz, [Mechanics]{}, Vol. 1, §39 (Pergamon Press, Third edition, 1980). Another proposal for such a potential based on the time-orbiting potential is under investigation by the Oxford group, see J.Arlt, O. Maragò, E. Hodby, S.A. Hopkins, G. Hechenblaikner, S. Webster and c.J. Foot, preprint cond-mat/9911201. L.S. Brown and G. Gabrielse, Rev. Mod. Phys., [58]{}, 233 (1986). L.D. Landau and E.M. Lifshitz, [Statistical Physics]{}, Vol. 5, §26 and §34 (Pergamon Press, Third edition, 1980). K. Huang, [Statistical Mechanics]{}, (J. Wiley, New York, 1987), 2nd ed. U. Al Khawaja, C.J. Pethick and H. Smith, preprint cond-mat/9908043. Private communication, J. Dalibard.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Non-ideal MHD effects may play a significant role in determining the dynamics, thermal properties, and observational signatures of radiatively inefficient accretion flows onto black holes. In particular, particle acceleration during magnetic reconnection events may influence black hole spectra and flaring properties. We use representative GRMHD simulations of black hole accretion flows to identify and explore the structures and properties of current sheets as potential sites of magnetic reconnection. In the case of standard and normal (SANE) disks, we find that, in the reconnection sites, the plasma beta ranges from $0.1$ to 1000, the magnetization ranges from $10^{-4}$ to 1, and the guide fields are weak compared to the reconnecting fields. In magnetically arrested (MAD) disks, we find typical values for plasma beta from $10^{-2}$ to $10^3$, magnetizations from $10^{-3}$ to 10, and typically stronger guide fields, with strengths comparable to or greater than the reconnecting fields. These are critical parameters that govern the electron energy distribution resulting from magnetic reconnection and can be used in the context of plasma simulations to provide microphysics inputs to global simulations. We also find that ample magnetic energy is available in the reconnection regions to power the fluence of bright X-ray flares observed from the black hole in the center of the Milky Way.' author: - 'David Ball, Feryal Özel$^{,}$, Dimitrios Psaltis$^{,}$, Chi-Kwan Chan, Lorenzo Sironi' bibliography: - 'david\_bib.bib' title: The Properties of Reconnection Current Sheets in GRMHD Simulations of Radiatively Inefficient Accretion Flows --- Introduction ============ General relativistic magnetohydrodynamic (GRMHD) simulations are often used to study the physics of accretion systems around compact objects and to explain their observed properties. These simulations make a number of simplifications that reduce computation time and facilitate calculations. The most common is the assumption of “ideal” MHD, which enforces that the plasma is infinitely conductive. This assumption leads to a couple of important properties: any electric fields in the fluid frame are shorted out and, as a result, the magnetic fields are frozen into the fluid (see, e.g., @Kulsrud2005). For many systems, the ideal MHD approximation is adequate: astrophysical plasmas are often fully ionized and have extremely low resistivities. However, even in systems where the approximation may globally seem appropriate, there can arise regions that violate the underlying assumptions. One such example is when two fluid elements with opposing magnetic fluxes encounter each-other. Conventional wisdom tells us that this may be a site of reconnection, where magnetic fields can change topology and dissipate their energy into the plasma. Such a configuration will have a rapidly changing magnetic field in space, resulting in a high current density. Ohmic dissipation, which scales as $\eta J^{2}$, where $\eta$ is the resistivity and J is the current density, can lead to significant dissipation where the current is high enough. This shows that dissipative terms can change the energetics of the flow and, given their localization, can lead to time-dependent phenomena that are not captured in global ideal MHD simulations. In the case of magnetic reconnection, even the inclusion of non-ideal terms may not be sufficient to capture the entire behavior of the plasma. For instance, reconnection has been shown to be an efficient source of non-thermal particle acceleration under certain conditions (@drake2013; @sironi2014; @melzani2014; @liguo2015; @guo2015; @werner2016; @sironi2016; see @kagan2015 for review), but the common use of the fluid equations assumes that the particle distribution is a Maxwellian. As a result, typical implementations of MHD do not capture effects related to non-thermal particles, regardless of the inclusion of non-ideal terms. In these cases, we must turn to computational methods of solving the Vlasov equation that do not make assumptions about the particle distribution. The supermassive black hole at the center of the Milky Way, Sgr A\, has an accretion flow that falls into a category broadly referred to as Radiatively Inefficient Accretion Flows (see @yuan2014 for a recent review). These flows are characterized by geometrically thick, optically thin disks, low accretion rates, and low luminosities. In recent years, there have been a number of studies using GRMHD to infer properties of the accretion flow around Sgr A\ (e.g., @narayan2012; @dexter2012; @drappeau2013; @chan2015a; @moscibrodzka2014). Even more recently, studies have begun to evolve the electron entropy equation, accounting for electron heating and anisotropic conduction (@ressler2015, 2016). Additionally, @chael2017 developed a scheme for coevolving a population of non-thermal electrons, including effects of adiabatic compression and expansion as well as radiative cooling. While these simulations successfully match a number of broadband steady-state properties, they show very little X-ray variability, contrary to observations (@eckart2004; @eckart2006; @neilsen2013, 2015). In @ball2016, we showed that a population of non-thermal electrons in highly magnetized regions of a radiatively inefficient accretion flow studied with GRMHD, where they are likely to be accelerated via reconnection, can result in X-ray variability with properties that are roughly consistent with observations. In this paper, we use representative GRMHD simulations to assess whether reconnection regions frequently occur in global simulations. We consider simulations with Standard and Normal Evolution (SANE) and Magnetically Arrested Disk (MAD) initial magnetic field configurations (see @narayan2012 & @sadowski2013). In the SANE case, the magnetic field is initialized with alternating poloidal loops, while the MAD initial field consists of a single poloidal loop, which results in the magnetic field playing a more dominant role in the dynamics of the disk. We devise criteria to locate regions of field reversal and characterize the properties of the plasma in these regions. We focus on the plasma-$\beta$ and magnetization parameter $\sigma$, which have been shown to play an important role in particle acceleration. We also identify field components that are orthogonal to the reversing field, often referred to as guide fields, and quantify their strengths. Our results will guide future particle-in-cell (PIC) studies of low-luminosity accretion flows. Finally, we compute the time-dependent magnetic energy available in reconnection regions to assess whether this is a plausible mechanism to generate the observed X-ray variability of Sgr A\. Characterizing Potential Reconnection Regions in MHD Simulations ================================================================ Magnetic reconnection takes place in regions where there is a reversal of magnetic field over a short characteristic length scale, in which the current density becomes large. In typical simulations of the local dynamics of reconnection, the initial condition is specified in terms of a Harris sheet, which has the magnetic field profile $$\vec{B}=B_{0}\tanh{\frac{x}{L}}\hat{y}.$$ In this geometry, the $y$-component of the magnetic field reverses direction over a characteristic length $L$ in the $x$-direction. This field reversal has a high curl associated with it, leading to a sudden peak in the current density, which scales as $$\vec{J}=\frac{B_{0}}{L}\operatorname{sech}^{2}{\frac{x}{L}}\hat{z}.$$ There are only a small number of parameters that determine the particle heating and acceleration that results from reconnection events. These are the magnetization parameter $$\sigma \equiv \frac{B^{2}}{4 \pi \rho c^2},$$ which is the ratio of magnetic energy density to rest mass energy density, and the plasma-$\beta$ parameter $$\beta \equiv \frac{P_{\rm{gas}}}{P_{\rm{magnetic}}}= \frac{8 \pi n k T}{B^2},$$ which specifies the ratio of gas pressure to magnetic pressure. Here, $\rho$, $n$, and $T$ are the mass density, number density, and temperature of the plasma particles, respectively. Another important quantity to consider for magnetic reconnection is the magnitude and direction, if present, of the so-called guide field. This is the component of the magnetic field in the sheet perpendicular to the reconnecting field. The effect of such a guide field on particle acceleration has been studied in certain regimes (@wangh2016; @dahlin2016; @stanier2016) and, in some cases, can have an effect on the resulting electron energy distribution. Our first goal is to devise an algorithm that will allow us to identify the location and relevant properties of potential reconnection regions, i.e., Harris sheets, in global GRMHD simulations, which we describe in the following. Finding and characterizing current sheets ========================================= As an illustrative example, we use two 60 hr (about $11,000 \; GM/c^{3}$) long GRMHD simulations of a radiatively inefficient accretion flow onto a black hole (@narayan2012; @sadowski2013) that were performed using the HARM code [@gammie2003]. These simulations were employed in a large study of the broadband, time-dependent emission from Sgr A\ (@chan2015b, b) where we coupled HARM to the radiative transfer algorithm GRay [@chan2013] and varied the black hole spin, density normalization, observer inclination, initial magnetic field configuration, as well as the electron thermodynamic prescription. From these investigations, we identified 5 models that best fit the steady-state broadband spectrum as well as the previously observed 1.3 mm image size of Sgr A\* and also characterized their variability properties. In the present study, we use two representative models from @chan2015a: a SANE model with a black hole spin $a=0.7$ and a MAD model with a black hole spin $a=0.9$. In general, the thermal SANE models tend to show short-lived, high amplitude variability in their IR and mm flux, while the thermal MAD models tend to show lightcurves dominated by smooth and long-timescale flux changes (@chan2015b). Our goal is to identify in each snapshot from these simulations potential regions of reconnection. Because of the large shear in the accretion flow, the magnetic fields are primarily toroidal and the alternating components occur primarily in the azimuthal direction. For this reason, we search through the simulation volume for cells that have both a high current relative to the mean value in the snapshot as well as very low values of the azimuthal component of the magnetic field $B_\phi$ in order to pick out the sheets where reconnection may occur. Specifically, we consider 2D slices of the simulation volume at each azimuthal angle $\phi$ at each snapshot and identify the points that ($i$) have current magnitudes $\sqrt{J_\mu J^\mu}$ that are higher than four times the mean current of that snapshot and ($ii$) $\phi$-components of the magnetic field smaller than a fiducial value, characterized by the usual magnetization parameter $\sigma_{\phi}=B_{\phi}^{2}/(4 \pi m n c^{2})$. We use a $\sigma_{\phi}$ threshold of $10^{-6}$. We then apply an algorithm similar to the one described in @zhdankin2013 for identifying and analyzing the statistics of current sheets in shearing box simulations of MHD turbulence. For every point with grid indices ($i,j$) on an azimuthal slice of our domain picked out by the above criteria with current magnitude $J_{\rm{peak}}$, we consider all 4 adjacent points in the grid. If the current at an adjacent point is above $J_{\rm{peak}}/2$, while also satisfying $\sigma_\phi<10^{-6}$, we consider it as part of the same current sheet. We continue this process of scanning every point in the sheet, considering all neighboring points, and applying these criteria to them until no more points are being added to the sheet. In the top panel of Figure \[bphi\_sheet\_example\], we show the result of applying this algorithm on a snapshot of the SANE simulation, with the $B_\phi$ configuration shown in the bottom panel of the same figure. It is evident that the regions between flux tubes of opposing azimuthal magnetic flux are effectively picked out. When we repeat this procedure on all adjacent azimuthal slices, we find that the current sheets show large azimuthal extents throughout the flow, providing ample surface area for neighboring flux tubes to reconnect over. Sampling The Plasma Properties Associated with Current Sheets ------------------------------------------------------------- Once we identify the current sheets in each snapshot, we characterize the plasma parameters of these sheets that are relevant to magnetic reconnection. The location that we want to measure these parameters at is not in the sheet itself but where the magnetic field reaches its asymptotic value some distance away from the sheet in a direction perpendicular to it. This breaks down into two problems: finding the direction perpendicular to the sheet and determining how far to go along this direction until the magnetic field reaches its appropriate asymptotic value. In order to approximate the direction perpendicular to the current sheet at a point ($i,j$) that has been flagged as belonging to the sheet, we first find the local slope of the sheet about this point. To do this, we consider a box around each point ($i,j$) in the sheet, with width S+1, whose corners are at ($i \pm S/2 , j \pm S/2$). We use a value of $S=10$ pixels, which is generally smaller than the radius of curvature of a current sheet. We then calculate the slope from the point in the center of this box ($i,j$) to every other point ($i^{\prime},j^{\prime}$) in the box which is flagged as being part of the current sheet. Taking the inverse tangent of this slope gives the angle with respect to the horizontal of the line that passes through ($i,j$) and ($i^{\prime}$,$j^{\prime}$). We calculate the average of these angles, approximating the angle of the current sheet about point ($i,j$) as $$\theta_{mean}=\frac{1}{N}\sum_{\left(i^{\prime},j^{\prime}\right)_{n=1}}^{\left(i^{\prime},j^{\prime}\right)_{N}} \arctan{\left[\frac{z\left(j^{\prime}_{n}\right)-z\left(j\right)}{r\left(i^{\prime}_{n}\right)-r\left(i\right)}\right]}$$ We then calculate the mean slope: $$m_{\rm{mean}}=\tan{\theta_{\rm{mean}}}$$ and take the direction perpendicular to this slope as $$m_{\perp}=-\frac{1}{m_{\rm{mean}}}$$ We sample the plasma properties at some distance along the normal where the toroidal magnetic field has reached its asymptotic value. We approximate this location by scanning along the normal direction, given by equation (7), until the field profile flattens out. We consider the field sufficiently flat when the fractional change in magnetic field from one computational cell along the normal to the next is less than three percent, averaged over two adjacent cells. We show in Figure \[field\_profile\] the magnetic field and current density profile along the normal of a current sheet picked out by our algorithm. We indeed see a Harris-sheet-like structure, with a magnetic field profile that passes through 0 and asymptotes to a fixed value at a distance $\approx$ 0.2–0.3 $GMc^{-2}$ away from the center; the current density has the expected maximum associated with the steep gradient in magnetic field. To illustrate the variety of current sheets and show their typical length scales and field profiles, we show in Figure \[rand\_b\] a randomly selected sample of current sheets identified in different snapshots and locations. The magnetic field profiles again follow structures reminiscent of Harris sheets, with the magnetic field passing in a linear fashion through 0 and reaching an asymptotic value at a distance that is typically 0.2 to 0.6 $GMc^{-2}$ away from the center of the sheet. ![A random selection of field profiles from the simulation across current sheets, showing that the typical behavior is reminiscent of the idealized Harris sheet structure, passing linearly through 0 and asymptoting to similar values on either side of the sheet. The vertical scale is normalized to the average asymptotic magnetic field for each sheet.[]{data-label="rand_b"}](final_randBplots.pdf){width="50.00000%"} Plasma Properties of Current Sheets =================================== Having established the frequent occurrence and geometry of potential reconnection regions, our second goal is to investigate the properties of current sheets in time-dependent simulations of accretion flows and characterize the parameters relevant to non-thermal particle acceleration to inform further PIC studies. We ultimately wish to determine the role of magnetic reconnection in contributing to the multiwavelength variability of low-luminosity accretion flows. Iterating through timesteps in our simulations, we find the current sheets and, at every point in each sheet, determine the asymptotic values of $\sigma$ and $\beta$ as well as the guide field strength at the center of the sheet for both our SANE and MAD simulations, as described below.\ Properties of SANE Current Sheets --------------------------------- For the SANE simulation, in the regions where reconnection may occur, the magnetization $\sigma$ ranges from $10^{-4}$ to 1, while the plasma-$\beta$ ranges from 0.1 to $10^{3}$, as shown in Figure \[SANE\_2d\_hist\]. \ The anticorrelation evident in Figure \[SANE\_2d\_hist\] (see also Fig. \[MAD\_2d\_hist\]) occurs because the magnetization parameter $\sigma$ scales as $B^{2}/n$, while the plasma-$\beta$ scales as $\left(B^{2}/n\right)^{-1}$. The spread arises because the plasma-$\beta$ also depends on the plasma temperature. The most promising subspace of this region for particle acceleration to be efficient is the high-$\sigma$, low-$\beta$ (bottom-right) regime, where there is maximal magnetic energy to dissipate into the particles and fairly little gas pressure relative to the magnetic pressure, such that the plasma is magnetically dominated. The inferred ranges of $\sigma$ and $\beta$ are interesting for a number of reasons. Studies have only recently begun for this transrelativistic regime [@werner2016] and the physics of particle acceleration and heating in these conditions are not yet fully understood. While the ions in this regime remain non-relativistic (because $\sigma$ is of order 1), the electrons will likely be accelerated (or heated on average) to highly relativistic speeds, since $\sigma_{e} \equiv B^{2}/\left(4 \pi \rho_{e}c^{2}\right) = \sigma m_{i}/m_{e} \approx10^{3}$, which is an estimate of the characteristic electron Lorentz factors expected from reconnection. Finally, in Figure \[guide\_field\] we show a histogram of the relative guide field strengths in the SANE simulation. It is evident that both cases of weak ($B_{r}/B_{\phi} < 0.5$) and of no guide fields are of interest for the purposes of these simulations. Even weak guide fields may play an important and potentially adverse role in determining the outcome of magnetic reconnection and must be explored via PIC simulations in the transrelativistic regime. ![Histogram of guide fields in the SANE simulation, scaled to $B_{\phi}$, the component showing field reversal. A large number of current sheets have no guide fields associated with them and, when present, the guide fields tend to be quite weak[]{data-label="guide_field"}](final_SANE_guide_histograms.pdf){width=".5\textwidth"} Properties of MAD Current Sheets -------------------------------- For the MAD simulation, we find that, in the regions of potential reconnection, $\sigma$ ranges from 10$^{-3}$ to 10, while $\beta$ ranges from 0.03 to 10$^{3}$, as shown in Figure \[MAD\_2d\_hist\]. This is roughly an order of magnitude higher (lower) than the $\sigma$ ($\beta$) values in the SANE simulation, hinting that particle acceleration may be more efficient in these systems. We show the guide fields in the MAD simulation in Figure \[MAD\_guide\_field\]. In stark contrast to the SANE guide fields, which are weak relative to the reconnecting field, the MAD guide fields are stronger and can be comparable to the reconnecting ones. While the typical values of the magnetization $\sigma$ and the plasma-$\beta$ are more favorable in terms of particle acceleration in the MAD simulations, the stronger guide fields may alter the outcome of the reconnection event for the particle distribution. ![Histogram of guide fields in the MAD simulation, scaled to $B_{\phi}$, the component showing field reversal. While many sheets have little to no guide fields present, there are a significant number of current sheets with strong guide fields that will likely impact the efficiency of particle acceleration in these sheets.[]{data-label="MAD_guide_field"}](final_MAD_guide_histograms.pdf){width=".5\textwidth"} Variability of Magnetic Energy Available For Reconnection ========================================================= We finally examine the time variability of energy available to reconnection throughout the accretion flow. One motivation for this is to assess whether reconnection events can contribute substantially to the high energy variability of low luminosity accretion flows, as has been extensively observed in the case of Sgr A\. We integrate the magnetic energy density, $B^{2}/8\pi$, over the reconnecting volume bounded by the surfaces defined by the asymptotic magnetic field location and obtain in this way the total magnetic energy in the reconnection regions throughout the flow. We plot the results of this in Figure \[mag energy\] for both the SANE and MAD simulations. ![image](final_SANE_mag_energy_plot.pdf){width=".5\textwidth"} ![image](final_MAD_mag_energy_plot.pdf){width=".5\textwidth"}\[MAD mag energy\] We see that the turbulent nature of the accretion flow very often leads to the formation of transient current sheets that result in a highly time-varying magnetic energy being available to reconnection. This indicates that magnetic reconnection likely is a significant contributor to the variability of such systems. The SANE model produces persistent variability due to the high levels of turbulence in the disk. The MAD system has less turbulence and, hence, fewer variations, but the higher degree of magnetization means that, when a current sheet does develop, it typically has more magnetic energy associated with it. For this reason, we find that the MAD simulation is characterized by fewer but stronger variations in the magnetic energy available to reconnection. Including the variability properties of non-thermal electrons that are accelerated in these current sheets will likely alter significantly the earlier finding of @chan2015b, who used models that assumed a purely thermal electron distribution. In that early work, the variability of MAD simulations was characterized by smooth long timescale variations in the flux. It is clear, however, from the present analysis of the MAD simulation that there is the potential for having a sudden injection of non-thermal electrons associated with the spikes in Figure \[MAD mag energy\], which can then result in corresponding flares in the lightcurve. Based on previous studies, we expect some fraction of this magnetic energy to go towards accelerating particles. This acceleration efficiency will, in principle, depend on the flow conditions, such as $\sigma$ and $\beta$, and can be found through PIC simulations. In order to go from a picture such as the one shown in Figure \[mag energy\] to the non-thermal particle energy as a function of time, the magnetic energy as a function of time must be combined with the acceleration efficiency as a function of flow parameters, which will likely result in even more dramatic variation of energy on short timescales. To estimate whether the energy available for reconnection is a plausible explanation for flares of these magnitudes, we calculate the total energy from an average X-ray flare from Sgr A\. Observations of X-ray flares from Sgr A\* show typical luminosities from $\sim 10^{34}$ erg s$^{-1}$ to $2 \times 10^{35}$ erg s$^{-1}$ and typical timescales from hundreds of s to 8 ks [@neilsen2013]. With a luminosity of $5 \times 10^{34}$ erg s$^{-1}$ and a duration of 1000 s, about $5 \times 10^{37}$ erg is being released in a typical flare. Considering Figure \[mag energy\], we see for the SANE model that the energy available to reconnection peaks at typical values around $10^{38}$ erg, while typical MAD energies are an order of magnitude higher than this. This shows that there is enough energy available to reconnection in these simulations to plausibly account for the observed energy released during these flares. Moreover, the efficiency $\xi$ that determines the fraction of magnetic energy that goes into particle acceleration must be quite high in the case of SANE models, which have typically lower magnetizations and hence less magnetic energy associated with their current sheets. In @ball2016, we characterized the non-thermal particle distribution using $\eta$, the fraction of non-thermal to thermal energy densities in the fluid and power-law index, p. We found that significant X-ray flares can occur while satisfying the observed quiescent X-ray constraints for values of $\eta = 0.1$ and a conservative power-law index of $p=-3.5$. To connect our present results to these earlier findings, we express $\eta$ in terms of $\beta$ and $\xi$ as $$\eta \equiv \frac{E_{nt}}{E_{th}} = \frac{\xi B^{2}}{nkT8\pi} = \xi \beta^{-1}. \label{beta_constraint}$$ We can rewrite this as a constraint on the plasma-$\beta$ using the $\eta$ found in our previous study to result in significant X-ray variability, i.e., $$0.1 \left(\frac{\eta}{0.1}\right) \beta \leq \xi.$$ This places a constraint on the plasma-$\beta$, given a local $\xi$, which must be found as a function of flow parameters via PIC simulations. Note that $\xi$, by definition, cannot be greater than 1, placing a strict upper limit of $\beta=10$ for regions where there is sufficient magnetic energy to accelerate particle to the energies required to generate the flux excursions demonstrated in @ball2016. More realistically, $\xi$ is likely to be of order 0.1, resulting in an upper limit of $\beta\approx1$. As shown in Figure \[SANE\_2d\_hist\], we find that we indeed identify many current sheets satisfying this condition. Conclusions =========== In this paper, we investigated the detailed structure of current sheets and their plasma properties in GRMHD simulations of radiatively inefficient accretion flows. We found that the regimes of plasma parameters relevant to magnetic reconnection have been relatively unexplored in terms of non-thermal particle acceleration. Specifically, we found that the magnetization $\sigma$ in the vicinity of current sheets in the SANE simulation is of order $10^{-4}$ to 1 , while the plasma-$\beta$ is of order 0.1 to $10^{3}$. Current sheets in the MAD simulation have magnetization $\sigma$ ranging from $10^{-3}$ to 10 and plasma-$\beta$ from 0.03 to $10^{3}$. Additionally we find that, in these regions, there is a relatively small spread in temperature, leading to a tight correlation between the parameters $\sigma$ and $\beta$. We also characterized the guide fields found in current sheets, which can play a role in governing the details of particle acceleration, and found that the ratio of guide field to reconnecting field strength is typically 0–0.5 for SANE simulations, but can be of order unity in MAD simulations. GRMHD simulations need to use subgrid models in order to account for physical effects that cannot be resolved or incorporated in MHD. In order to employ correctly subgrid models of reconnection, we must improve our understanding of particle acceleration and heating in the parameter space we lay out here. In addition to characterizing the plasma properties of current sheets, we also calculated the magnetic energy available to reconnection throughout the simulations. We found that the turbulent nature of the accretion flow leads to current sheets of varying characteristics continuously forming and dissipating in the flow. This leads to a highly variable amount of energy available to reconnect and dissipate into heating and particle acceleration and makes magnetic reconnection a promising candidate for contributing to the X-ray variability of Sgr A\* and other black holes with similar accretion characteristics. Additionally, we found that there is indeed enough energy available to reconnection around current sheets to account for typical flares observed from Sgr A\*. We conclude that if this mechanism is responsible for the X-ray flares, then the acceleration efficiency must be reasonably high for SANE disks and can be lower for the MAD model. We gratefully acknowledge support for this work from Chandra Award No. TM6-17006X and from NASA TCAN award NNX14AB48G. DP acknowledges support from the Radcliffe Institute for Advanced Study at Harvard University. FO gratefully acknowledges a fellowship from the John Simon Guggenheim Memorial Foundation in support of this work. Calculation of Current Density in a Kerr Metric =============================================== In order to find regions in our GRMHD simulation where magnetic reconnection would occur, if it was explicitly included, we need to identify regions of high current density. However, GRMHD simulations typically evolve only 4 quantities, i.e., the magnetic field, density, fluid velocity, and internal energy, and the current density is typically not explicitly computed in the simulation. To understand the structure of the current density throughout the flow, we calculate it from the electromagnetic tensor, $$\label{maxwell tensor} F^{\mu \nu}= \left( \begin{array}{cccc} 0 & -E_{1} & -E_{2} & -E_{3} \\ E_{1} & 0 & B_{3} & -B_{2} \\ E_{2} & -B_{3} & 0 & B_{1}\\ E_{3} & B_{2} & -B_{1} & 0 \end{array} \right)$$ via $$\label{maxwell} J^{\nu}=\nabla_{\mu}F^{\nu \mu}$$ where $\nabla_\mu$ represents the covariant derivative. Breaking up the four-current into its zeroth and i$^{th}$ components (henceforth, Greek indices run from 0-3, while Latin indices go from 1-3), we can rewrite equation (\[maxwell\]) as $$\label{max1} \nabla_{j}F^{ij} - \nabla_{0}F^{0i}=J^{i}$$ $$\label{max2} \nabla_{i}F^{0i}=J^{0},$$ In the comoving frame, $F^{0 i} = 0$, equation (\[max2\]) reads $J^{0} = 0$. Additionally, the displacement current is 0 ($\nabla_{0}F^{0i}=0$). We can then write the three-current as $$\label{ji} J^{i} = \partial_{j}F^{ij} + \Gamma^{i}_{j \lambda}F^{\lambda j} + \Gamma^{j}_{j \lambda}F^{i \lambda}.$$ Considering the 2nd term on the right hand side, $\Gamma^{i}_{j \lambda}F^{\lambda j}$, we see that for every term in the sum, the symmetry of $\Gamma^{i}_{jk}$ about its lower two indices and the anti-symmetry of $F^{\mu \nu}$ implies perfect cancellation. Hence, equation (\[ji\]) reduces to $$\label{ji2} J^{i} = \partial_{j}F^{ij} + \Gamma^{j}_{j \lambda}F^{i \lambda}.$$ We use the Christoffel symbols for the metric of an uncharged Kerr black hole, which in Boyer-Lindquist coordinates is given by $$\label{BL metric} \begin{aligned} ds^{2} = -\frac{\Delta}{\Sigma}\left(dt - a \sin^{2}\theta d\phi \right)^{2} + \frac{\Sigma}{\Delta}dr^{2} + \Sigma d\theta^{2} \\+ \frac{\sin^{2}\theta}{\Sigma}\left(\left(r^{2} + a^{2}\right)d\phi + a dt \right)^{2} \end{aligned}$$ with $$\begin{aligned} \Delta=r^{2} - 2Mr + a^{2} \\ \Sigma = r^{2} + a^{2}\cos^{2}\theta. \end{aligned}$$\ Here, M and a are the mass and angular momentum per unit mass of the black hole. For this metric, the relevant nonzero Christoffel terms in equation (\[ji2\]) are $\Gamma^{1}_{11} , \Gamma^{1}_{1 2} , \Gamma^{3}_{3 1} , \Gamma^{3}_{3 2} , \Gamma^{2}_{2 1} ,\Gamma^{2}_{2 2}.$ With these, we can write out the individual components of $J^{i}$ as $$J^{1} = \partial_{j}F^{1j} + \left(\Gamma^{1}_{12} + \Gamma^{3}_{32} + \Gamma^{2}_{22}\right)F^{12} ,$$ $$J^{2} = \partial_{j}F^{2j} + \left(\Gamma^{1}_{11} + \Gamma^{3}_{31} + \Gamma^{2}_{21}\right)F^{21} ,$$ and $$\begin{aligned} J^{3} = \partial_{j}F^{3j} + \left(\Gamma^{1}_{11} + \Gamma^{2}_{21} + \Gamma^{3}_{31}\right)F^{31} \\+ \left(\Gamma^{1}_{12} + \Gamma^{3}_{32} + \Gamma^{2}_{22}\right)F^{32}. \end{aligned}$$ Using these equations, we solve for the current density in GRMHD simulations at every timestep and identify current sheets where reconnection may take place.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Determining the intended sense of words in text – word sense disambiguation (WSD) – is a long-standing problem in natural language processing. Recently, researchers have shown promising results using word vectors extracted from a neural network language model as features in WSD algorithms. However, a simple average or concatenation of word vectors for each word in a text loses the sequential and syntactic information of the text. In this paper, we study WSD with a sequence learning neural net, LSTM, to better capture the sequential and syntactic patterns of the text. To alleviate the lack of training data in all-words WSD, we employ the same LSTM in a semi-supervised label propagation classifier. We demonstrate state-of-the-art results, especially on verbs.' author: - \| Dayu Yuan Julian Richardson Ryan Doherty Colin Evans Eric Altendorf\ Google, Mountain View CA, USA\ {dayuyuan,jdcr,portalfire,colinhevans,ealtendorf}@google.com bibliography: - 'coling2016.bib' title: 'Semi-supervised Word Sense Disambiguation with Neural Models' --- Introduction ============ Word sense disambiguation (WSD) is a long-standing problem in natural language processing (NLP) with broad applications. Supervised, unsupervised, and knowledge-based approaches have been studied for WSD [@survey]. However, for [all-words]{} WSD, where all words in a corpus need to be annotated with word senses, it has proven extremely challenging to beat the strong baseline, which always assigns the most frequent sense of a word without considering the context [@semeval2007; @survey; @semeval2013; @semeval2015]. Given the good performance of published supervised WSD systems when provided with significant training data on specific words [@IMS], it appears lack of sufficient labeled training data for large vocabularies is the central problem. One way to leverage unlabeled data is to train a neural network language model (NNLM) on the data. Word embeddings extracted from such a NNLM (often Word2Vec [@mikolov2013efficient]) can be incorporated as features into a WSD algorithm. Iacobacci et al. show that this can substantially improve WSD performance and indeed that competitive performance can be attained using word embeddings alone. In this paper, we describe two novel WSD algorithms. The first is based on a Long Short Term Memory (LSTM) [@hochreiter1997long]. Since this model is able to take into account word order when classifying, it performs significantly better than an algorithm based on a continuous bag of words model (Word2vec) [@mikolov2013efficient; @supervised_word2vec], especially on verbs. We then present a semi-supervised algorithm which uses label propagation [@talukdar2009new; @ravi2016] to label unlabeled sentences based on their similarity to labeled ones. This allows us to better estimate the distribution of word senses, obtaining more accurate decision boundaries and higher classification accuracy. The best performance was achieved by using an LSTM language model with label propagation. Our algorithm achieves state-of-art performance on many SemEval all-words tasks. It also outperforms the most-frequent-sense and Word2Vec baselines by $10$% (see Section \[sec:noad-evel\] for details). [Organization: ]{} We review related work in Section \[sec:related-work\]. We introduce our supervised WSD algorithm in Section \[sec:supervised-wsd\], and the semi-supervised WSD algorithm in Section \[sec:semi-supervised-wsd\]. Experimental results are discussed in Section \[sec:experiment\]. We provide further discussion and future work in Section \[sec:discussion\]. Related Work {#sec:related-work} ============ The development of large lexical resources, such as WordNet [@wordnet] and BabelNet [@babelnet], has enabled knowledge-based algorithms which show promising results on all-words prediction tasks [@knowledge-rich; @semeval2013; @semeval2015]. WSD algorithms based on supervised learning are generally believed to perform better than knowledge-based WSD algorithms, but they need large training sets to perform well [@semeval2007; @semeval2007-coarse; @survey; @IMS]. Acquiring large training sets is costly. In this paper, we show that a supervised WSD algorithm can perform well with $\sim 20$ training examples per sense. In the past few years, much progress has been made on using neural networks to learn word embeddings [@mikolov2013efficient; @levy2014neural], to construct language models [@mikolov2011extensions], perform sentiment analysis [@socher2013recursive], machine translation [@sequence_to_sequence] and many other NLP applications. A number of different ways have been studied for using word embeddings in WSD. There are some common elements: - Context embeddings. Given a window of text $w_{n-k}, ..., w_n, ..., w_{n+k}$ surrounding a focus word $w_n$ (whose label is either known in the case of example sentences or to be determined in the case of classification), an embedding for the context is computed as a concatenation or weighted sum of the embeddings of the words $w_i, i \ne n$. Context embeddings of various kinds are used in both [@chen2014unified] and [@supervised_word2vec]. - Sense embeddings. Embeddings are computed for each word sense in the word sense inventory (e.g. WordNet). In [@autoextend], equations are derived relating embeddings for word senses with embeddings for undisambiguated words. The equations are solved to compute the sense embeddings. In [@chen2014unified], sense embeddings are computed first as weighted sums of the embeddings of words in the WordNet gloss for each sense. These are used in an initial bootstrapping WSD phase, and then refined by a neural network which is trained on this bootstrap data. - Embeddings as SVM features. Context embeddings [@supervised_word2vec; @wsd-word-embedding], or features computed by combining context embeddings with sense embeddings [@autoextend], can be used as additional features in a supervised WSD system e.g. the SVM-based [IMS]{} [@IMS]. Indeed Iacobacci et al. found that using embeddings as the only features in [IMS]{} gave competitive WSD performance. - Nearest neighbor classifier. Another way to perform classification is to find the word sense whose sense embedding is closest, as measured by cosine similarity, to the embedding of the classification context. This is used, for example, in the bootstrapping phase of [@chen2014unified]. - Retraining embeddings. A feedforward neural network can be used to jointly perform WSD and adjust embeddings [@chen2014unified; @wsd-word-embedding]. In our work, we start with a baseline classifier which uses $1000$-dimensional embeddings trained on a $100$ billion word news corpus using Word2Vec [@mikolov2013efficient]. The vocabulary consists of the most frequent $1,000,000$ words, without lemmatization or case normalization. Sense embeddings are computed by averaging the context embeddings of sentences which have been labeled with that sense. To classify a word in a context, we assign the word sense whose embedding has maximal cosine similarity with the embedding of the context. This classifier has similar performance to the best classifier in [@supervised_word2vec] when SemCor is used as a source of labeled sentences. The Word2Vec embeddings are trained using a bag of words model, i.e. without considering word order in the training context, and word order is also not considered in the classification context. In Section \[sec:supervised-wsd\] we show that using a more expressive language model which takes account of word order yields significant improvements. Semi-supervised learning has previously been applied successfully to word sense disambiguation. In [@yarowsky1995unsupervised] bootstrapping was used to learn a high precision WSD classifier. A low recall classifier was learned from a small set of labeled examples, and the labeled set then extended with those sentences from an unlabeled corpus which the classifier could label with high confidence. The classifier was then retrained, and this iterative training process continued to convergence. Additional heuristics helped to maintain the stability of the bootstrapping process. The method was evaluated on a small data set. In [@niu2005word], a label propagation algorithm was proposed for word sense disambiguation and compared to bootstrapping and a SVM supervised classifier. Label propagation can achieve better performance because it assigns labels to optimize a [global]{} objective, whereas bootstrapping propagates labels based on [local]{} similarity of examples. In Section \[sec:semi-supervised-wsd\] we describe our use of label propagation to improve on nearest neighbor for classification. Supervised WSD with LSTM ======================== Neural networks with long short-term memory (LSTM) units [@hochreiter1997long] make good language models which take into account word order [@sundermeyer2012lstm]. We train a LSTM language model to predict the held-out word in a sentence. As shown in Figure \[fig:lstm\], we first replace the held-out word with a special symbol \$, and then, after consuming the remaining words in the sentence, project the $h$ dimensional hidden layer to a $p$ dimensional context layer, and finally predict the held out word with softmax. By default, the LSTM model has $2048$ hidden units, $512$ dimensional context layer and $512$ dimensional word embeddings. We also studied other settings, see Section \[sec:exp-noad-lstm\] for details. We train the LSTM on a news corpus of about $100$ billion tokens, with a vocabulary of $1,000,000$ words. Words in the vocabulary are neither lemmatized nor case normalized. Our LSTM model is different from that of Kågebäck and Salomonsson [@bilstm]. We train a LSTM language model, which predicts a held-out word given the surrounding context, with a large amount of unlabeled text as training data. The huge training dataset allows us to train a high-capacity model ($2048$ hidden units, $512$ dimensional embeddings), which achieves high precision without overfitting. In our experiments, this directional LSTM model was faster and easier to train than a bidirectional LSTM, especially given our huge training dataset. Kågebäck and Salomonsson’s LSTM directly predicts the word senses and it is trained with a limited number of word sense-labeled examples. Although regularization and dropout are used to avoid overfitting the training data, the bidirectional LSTM is small with only $74 + 74$ neurons and $100$ dimensional word embeddings [@bilstm]. Because our LSTM is generally applicable to any word, it achieves high performance on [all-words WSD]{} tasks (see Section \[sec:experiment\] for details), which is the focus of this paper. Kågebäck and Salomonsson’s LSTM is only evaluated on [lexical sample WSD]{} tasks of SemEval 2 and 3 [@bilstm]. \[sec:supervised-wsd\] ![LSTM: Replace the focus word $w_3$ with a special symbol \$ and predict $w_3$ at the end of the sentence.[]{data-label="fig:lstm"}](LSTM2.pdf){width="60.00000%"} The behavior of the LSTM can be intuited by its predictions. Table \[tab:LSTM-predictions\] shows the top $10$ words predicted by an LSTM language model for the word ‘stock’ in sentences containing various senses of ‘stock’. [> m[0.5cm]{}m[7cm]{}m[5.5cm]{} > m[0.8 cm]{}]{} id & sentence & top 10 predictions from LSTM & sense\ 1 & Employee compensation is offered in the form of cash and/or [stock]{}. &cash, stock, equity, shares, loans, bonus, benefits, awards, equivalents, deposits &\ (lr)[1-3]{} 2 & The [stock]{} would be redeemed in five years, subject to terms of the company’s debt. & bonds, debt, notes, shares, stock, balance, securities, rest, Notes, debentures &\ 3 & These stores sell excess [stock]{} or factory overruns . &inventory, goods, parts, sales, inventories, capacity, products, oil, items, fuel& sense\#2\ 4 & Our soups are cooked with vegan [stock]{} and seasonal vegetables. & foods, food, vegetables, meats, recipes, cheese, meat, chicken, pasta, milk& sense\#3\ query & In addition, they will receive [stock]{} in the reorganized company, which will be named Ranger Industries Inc. & shares, positions, equity, jobs, awards, representation, stock, investments, roles, funds& ?\ In our initial experiments, we computed similarity between two contexts by the overlap between their bags of predicted words. For example (Table \[tab:LSTM-predictions\]) the top predictions for the query overlap most with the LSTM predictions for ‘sense\#1’ —we predict that ‘sense\#1’ is the correct sense. This bag of predictions, while easily interpretable, is just a discrete approximation to the internal state of the LSTM when predicting the held out word. We therefore directly use the LSTM’s context layer from which the bag of predictions was computed as a representation of the context (see Figure \[fig:lstm\]). Given context vectors extracted from the LSTM, our supervised WSD algorithms classify a word in a context by finding the sense vector which has maximum cosine similarity to the context vector (Figure \[fig:nneighbor\_exp\]). We find the sense vectors by averaging context vectors of all training sentences of the same sense. We observed in a few cases that the context vector is far from the held-out word’s embedding, especially when the input sentence is not informative. For example, the LSTM language model will predict “night" for the input sentence “I fell asleep at \[work\]." when we hold out “work". Currently, we treat the above cases as outliers. We would like explore alternative solutions, e.g., forcing the model to predict words that are close to one sense vector of the held-out word, in further work. As can be seen in SemEval all-words tasks and Tables \[tab:baseline\], this LSTM model has significantly better performance than the Word2Vec models. Semi-supervised WSD {#sec:semi-supervised-wsd} =================== [0.48]{} ![WSD classifiers. Filled nodes represent labeled sentences. Unfilled nodes represent unlabeled sentences.[]{data-label="fig:expander_exp"}](NNeighbor "fig:"){width="50.00000%"} [0.48]{} ![WSD classifiers. Filled nodes represent labeled sentences. Unfilled nodes represent unlabeled sentences.[]{data-label="fig:expander_exp"}](Expander "fig:"){width="50.00000%"} The non-parametric nearest neighbor algorithm described in Section \[sec:supervised-wsd\] has the following drawbacks: - It assumes a spherical shape for each sense cluster, being unable to accurately model the decision boundaries given the limited number of examples. - It has no training data for, and does not model, the sense prior, omitting an extremely powerful potential signal. To overcome these drawbacks we present a semi-supervised method which augments the labeled example sentences with a large number of unlabeled sentences from the web. Sense labels are then propagated from the labeled to the unlabeled sentences. Adding a large number of unlabeled sentences allows the decision boundary between different senses to be better approximated. A [label-propagation graph]{} consists of (a) vertices with a number of labeled seed nodes and (b) undirected weighted edges. Label propagation (LP) [@talukdar2009new] iteratively computes a distribution of labels on the graph’s vertices to minimize a weighted combination of: - The discrepancy between seed labels and their computed labels distributions. - The disagreement between the label distributions of connected vertices. - A regularization term which penalizes distributions which differ from the prior (by default, a uniform distribution). We construct a graph for each lemma with labeled vertices for the labeled example sentences, and unlabeled vertices for sentences containing the lemma, drawn from some additional corpus. Vertices for sufficiently similar sentences (based on criteria discussed below) are connected by an edge whose weight is the cosine similarity between the respective context vectors, using the LSTM language model. To classify an occurrence of the lemma, we create an additional vertex for the new sentence and run LP to propagate the sense labels from the seed vertices to the unlabeled vertices. Figure \[fig:expander\_exp\] (b) illustrates the graph configuration. Spatial proximity represents similarity of the sentences attached to each vertex and the shape of each node represents the word sense. Filled nodes represent seed nodes with known word senses. Unfilled nodes represent sentences with no word sense label, and the ? represents the word we want to classify. With too many edges, sense labels propagate too far, giving low precision. With too few, sense labels do not propagate sufficiently, giving low recall. We found that the graph has about the right density for common senses when we ranked vertex pairs by similarity and connected the pairs above the $95$ percentile. This may still leave rare senses sparsely connected, so we additionally added edges to ensure that every vertex is connected to at least $10$ other vertices. Our experiments (Table \[tab:density\]) show that this setting achieves good performance on WSD, and the performance is stable when the percentile ranges between $85$ to $98$. Since it requires running LP for every classification, the algorithm is slow compared to the nearest neighbor algorithm. Experiments {#sec:experiment} =========== We evaluated the LSTM algorithm with and without label propagation on standard SemEval all-words tasks using WordNet as the inventory. Our proposed algorithms achieve state-of-art performance on many SemEval all-words WSD tasks. In order to assess the effects of training corpus size and language model capacity we also evaluate our algorithms using the New Oxford American Dictionary (NOAD) inventory with SemCor [@miller1993semantic] or MASC [^1]. SemEval Tasks ------------- In this section, we study the performance of our classifiers on Senseval2 [@senseval2], Senseval3 [@senseval3], SemEval-2007 [@senseval7], SemEval-2013 Task 12 [@semeval2013] and SemEval-2015 task 13 [@semeval2015] [^2]. We focus the study on all-words WSD tasks. For a fair comparison with related works, the classifiers are evaluated on all words (both polysemous and monosemous). Following related works, we use SemCor or OMSTI [@taghipour2015one] for training. In our LP classifiers, unlabeled data for each lemma consists either of $1000$ sentences which contain the lemma, randomly sampled from the web, or all OMSTI sentences (without labels) which contain the lemma. --------------------------------------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -- -- -- (lr)[2-3]{} (lr)[4-5]{} (lr)[6-7]{} (lr)[8-9]{} [model]{} [all]{} [n.]{} [all]{} [n.]{} [all]{} [n.]{} [all]{} [n.]{} [n.]{} IMS + Word2Vec (T:SemCor) 0.634 0.742 0.653 0.701 0.578 0.686 IMS + Word2Vec (T:OMSTI) 0.683 0.777 0.682 0.741 0.591 0.715 Taghipour and Ng 0.682 Chen et al. 0.826 [0.853]{} Weissenborn et al. 0.688 0.660 [0.855]{} [0.728]{} Word2Vec (T:SemCor) 0.678 0.737 0.621 0.714 0.585 0.673 0.795 0.814 0.661 LSTM (T:SemCor) 0.736 0.786 0.692 0.723 0.642 [0.723]{} 0.828 0.834 0.670 LSTM (T:OMSTI) 0.724 0.777 0.643 0.680 0.607 0.673 0.811 0.820 0.673 LSTMLP (T:SemCor, U:OMSTI) 0.739 0.797 0.711 0.748 [0.637]{} 0.704 [0.843]{} 0.834 0.679 LSTMLP (T:SemCor, U:1K) 0.738 0.796 [0.718]{} [0.763]{} 0.635 0.717 0.836 0.831 0.695 LSTMLP (T:OMSTI, U:1K) [0.744]{} [0.799]{} 0.710 0.753 0.633 0.717 0.833 0.825 0.681 --------------------------------------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -- -- -- Table \[tab:sem\_eval\] shows the Sem-Eval results. Our proposed algorithms achieve the highest all-words F1 scores except for Sem-Eval 2013. Weissenborn et al. only disambiguates nouns, and it outperforms our algorithms on Sem-Eval 2013 by $4$%, but is ranked behind our algorithms on Senseval-3 and SemEval-7 tasks with an F1 score more than $6$% lower than our algorithms. Unified WSD [@chen2014unified] has the highest F1 score on Nouns (Sem-Eval-7 Coarse), but our algorithms outperform Unified WSD on other part-of-speech tags. #### Settings For a fair comparison of Word2Vec and LSTM, we do not use pre-trained word-embeddings as in [@supervised_word2vec], but instead train the Word2Vec and LSTM models on a $100$ billion word news corpus [^3] with a vocabulary of the most frequent 1,000,000 words. Our self-trained word embeddings have similar performance to the pre-trained embeddings, as shown in Table \[tab:sem\_eval\]. The Word2Vec word vectors are of dimension ${1024}$. The LSTM model has $2048$ hidden units, and inputs are ${512}$-dimensional word vectors. We train the LSTM model by minimizing sampled softmax loss [@sampled_softmax] with Adagrad [@adagrad]. The learning rate is $0.1$. We experimented with other learning rates, and observed no significant performance difference after the training converges. We also downsample frequent terms in the same way as [@mikolov2013efficient]. #### Word2Vec vectors Vs. LSTM To better compare LSTM with word vectors we also build a nearest neighbor classifier using Word2Vec word embeddings and SemCor example sentences, Word2Vec (T:SemCor). It performs similar to IMS + Word2Vec (T:SemCor), a SVM-based classifier studied in [@supervised_word2vec]. Table \[tab:sem\_eval\] shows that the LSTM classifier outperforms the Word2Vec classifier across the board. #### SemCor Vs. OMSTI Contrary to the results observed in [@supervised_word2vec], the LSTM classifier trained with OMSTI performs worse than that trained with SemCor. It seems that the larger size of the OMSTI training data set is more than offset by noise present in its automatically generated labels. While the SVM classifier studied in [@supervised_word2vec] may be able to learn a model which copes with this noise, our naive nearest neighbor classifiers do not have a learned model and deal less well with noisy labels. #### Label propagation We use the implementation of DIST\_EXPANDER [@ravi2016]. We test the label propagation algorithm with SemCor or OMSTI as labeled data sets and OMSTI or $1000$ random sentences from the web per lemma as unlabeled data. The algorithm performs similarly on the different data sets. Table \[tab:sem\_eval\_2015\] shows the results of Sem-Eval 2015. The LSTM LP classifier with an LSTM language model achieves the highest scores on nouns and adverbs as well as overall F1. The LSTM classifier has the highest F1 on verbs. [algorithm]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} ------------------------- --------------- --------------- --------------- --------------- --------------- LIMSI 0.647 0.795 DFKI 0.703 0.577 UNIBA 0.790 BFS Baseline 0.663 0.667 0.551 [0.821]{} 0.825 Word2Vec (T:SemCor) 0.667 0.661 0.555 0.789 0.810 LSTM (T:SemCor) 0.721 0.713 [0.642]{} 0.813 0.845 LSTMLP (T:SemCor, U:1K) [0.726]{} [0.728]{} 0.622 0.813 [0.857]{} : F1 Scores of SemEval-2015 English Dataset. The BFS baseline uses BabelNet first sense. []{data-label="tab:sem_eval_2015"} NOAD Eval {#sec:noad-evel} --------- Many dictionary lemmas and senses have no examples in SemCor or OSTMI, giving rise to losses in all-words WSD when these corpora are used as training data. The above SemEval scores do not distinguish errors caused by missing training data for certain labels or inaccurate classifier. To better study the proposed algorithms, we train the classifiers with the New Oxford American Dictionary (NOAD) [@NOAD], in which there are example sentences for each word sense. ### Word Sense Inventory {#sec:inventory} The NOAD focuses on American English and is based on the Oxford Dictionary of English (ODE) [@ODE]. It distinguishes between coarse ([core]{}) and fine-grained ([sub]{}) word senses in the same manner as ODE. Previous investigations [@Navigli:2006:MCS; @semeval2007-coarse] using the ODE have shown that coarse-grained word senses induced by the ODE inventory address problems with WordNet’s fine-grained inventory, and that the inventory is useful for word sense disambiguation. For our experiments, we use NOAD’s core senses, and we also use lexicographer-curated example sentences from the Semantic English Language Database (SELD)[^4], provided by Oxford University Press. We manually annotated all words of the English language SemCor corpus and MASC corpora with NOAD word senses in order to evaluate performance [^5]. Table \[tab:NOAD\] shows the total number of polysemes (more than one core sense) and average number of senses per polyseme in NOAD/SELD (hereafter, NOAD), SemCor and MASC. Both SemCor and MASC individually cover around $45\%$ of NOAD polysemes and $62\%$ of senses of those polysemes. [ ]{} noun verb adj. adv. -- -------- ------ ------ ------ ------ NOAD 8097 2124 2126 266 SemCor 2833 1362 911 193 MASC 2738 1250 829 181 NOAD 2.46 2.58 2.30 2.47 SemCor 1.44 1.69 1.49 1.84 MASC 1.48 1.66 1.48 2.01 : NOAD polysemous lemma in NOAD, SemCor and MASC[]{data-label="tab:NOAD"} Table \[tab:data\] gives the number of labeled sentences of these datasets. Note that although NOAD has more labeled sentences than SemCor or MASC, the average numbers of sentences per sense of these datasets are similar. This is because NOAD has labeled sentences for each word sense, whereas SemCor (MASC) only covers a subset of lemmas and senses (Table \[tab:NOAD\]). The last column of Table \[tab:data\] shows that each annotated word in SemCor and MASC has an average of more than $4$ NOAD corse senses. Hence, a random guess will have a precision around $1/4$. ------------- --------- --------- --------- ---------- ---------- ------- --------- [dataset]{} all n. v. adj. adv NOAD [580]{} [312]{} [150]{} [95]{} [13]{} 18.43 [3.1]{} SemCor [115]{} [38]{} [57]{} [11.6]{} [8.6]{} 14.27 [4.1]{} MASC [133]{} [50]{} [57]{} [12.7]{} [13.6]{} 17.38 [4.2]{} ------------- --------- --------- --------- ---------- ---------- ------- --------- : Number of examples in each dataset and the average sense count per example.[]{data-label="tab:data"} In the default setting, we use NOAD example sentences as labeled training data and evaluate on SemCor and MASC. We evaluate all polysemous words in the evaluation corpus. ### LSTM classifier {#sec:exp-noad-lstm} We compare our algorithms with two baseline algorithms: - Most frequent sense: Compute the sense frequency (from a labeled corpus) and label word $w$ with $w$’s most frequent sense. - Word2Vec: a nearest-neighbor classifier with Word2Vec word embedding, which has similar performance to cutting-edge algorithms studied in [@supervised_word2vec] on SemEval tasks. ---------------------------------------------------- -------------------- --------------- --------------- ------------ --------------- --------------- --------------- --------------- --------------- --------------- --------------- [eval data]{} (lr)[1-2]{} (lr)[3-7]{} (lr)[8-12]{} [model]{} [train data]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} Frequent Sense SemCor 0.753 0.799 0.713 0.758 0.741 Frequent Sense MASC 0.752 0.751 0.749 0.737 0.789 Word2Vec NOAD 0.709 0.783 0.657 0.736 0.693 0.671 0.790 0.562 0.724 0.638 Word2Vec SemCor 0.692 0.806 0.592 0.754 0.635 Word2Vec NOAD,SemCor 0.678 0.808 0.565 0.753 0.604 Word2Vec MASC 0.698 0.785 0.619 0.766 0.744 Word2Vec NOAD,MASC 0.695 0.801 0.605 0.767 0.719 (lr)[2-12]{} LSTM NOAD 0.786 0.796 0.782 0.781 0.784 0.786 0.805 0.772 0.776 0.786 LSTM SemCor 0.799 0.843 0.767 0.808 0.767 LSTM NOAD,SemCor [0.812]{} [0.846]{} [0.786]{} [0.816]{} [0.798]{} LSTM MASC 0.810 0.825 0.799 0.809 [0.825]{} LSTM NOAD,MASC [0.821]{} [0.834]{} 0.814 [0.818]{} 0.821 ---------------------------------------------------- -------------------- --------------- --------------- ------------ --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------------------------------------------- -------------------- ------------- ------------ ------------ -------------- -------------- --------------------------------------------------------------------------- ------------ ------------ -------------- -------------- [eval data]{} (lr)[1-2]{} (lr)[3-7]{} (lr)[8-12]{} [model]{} [train data]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} (lr)[2-12]{} LSTM NOAD 0.769 0.791 0.759 0.751 0.672 0.780 0.791 0.768 0.780 0.726 LSTM SemCor 0.656 0.663 0.668 0.643 0.581 LSTM NOAD,SemCor 0.796 &0.805 &0.790 &0.794 &0.742\ LSTM & MASC & 0.631 & 0.653 & 0.606 & 0.617 & 0.600 & & & & &\ LSTM & NOAD,MASC & 0.782 &0.803 &0.774 &0.761 &0.688 & & & & &\ ************** ---------------------------------------------------- -------------------- ------------- ------------ ------------ -------------- -------------- --------------------------------------------------------------------------- ------------ ------------ -------------- -------------- Table \[tab:baseline\] compares the F1 scores of the LSTM and baseline algorithms. LSTM outperforms Word2Vec by more than $10\%$ over all words, where most of the gains are from verbs and adverbs. The results suggest that syntactic information, which is well modeled by LSTM but ignored by Word2Vec, is important to distinguishing word senses of verbs and adverbs. #### Change of training data By default, the WSD classifier uses the NOAD example sentences as training data. We build a larger training dataset by adding labeled sentences from SemCor and MASC, and study the change of F1 scores in Table \[tab:baseline\]. Across all part of speech tags and datasets, F1 scores increase after adding more training data. We further test our algorithm by using SemCor (or MASC) as training data (without NOAD examples). The SemCor (or MASC) trained classifier is on a par with the NOAD trained classifier on F1 score. However, the macro F1 score of the former is much lower than the latter, as shown in Table \[tab:baseline-macro\], because of the limited coverage of rare senses and words in SemCor and MASC. #### Change of language model capacity In this experiment, we change the LSTM model capacity by varying the number of hidden units $h$ and the dimensions of the input embeddings $p$ and measuring F1. Figure \[fig:dim\] shows strong positive correlation between F1 and the capacity of the language model. However, larger models are slower to train and use more memory. To balance the accuracy and resource usage, we use the second best LSTM model ($h = 2048$ and $p = 512$) by default. [0.48]{} ![image](semcor.pdf){width="90.00000%"} [0.48]{} ![image](masc.pdf){width="90.00000%"} ### Semi-supervised WSD {#semi-supervised-wsd} We evaluate our semi-supervised WSD classifier in this subsection. We construct the graph as described in Section \[sec:semi-supervised-wsd\] and run LP to propagate sense labels from the seed vertices to the unlabeled vertices. We evaluate the performance of the algorithm by comparing the predicted labels and the gold labels on eval nodes. As can be observed from Table \[tab:semi-supervised\], LP did not yield clear benefits when using the Word2Vec language model. We did see significant improvements, 6.3% increase on SemCor and 7.3% increase on MASC, using LP with the LSTM language model. We hypothesize that this is because LP is sensitive to the quality of the graph distance metric. ---------------------------------------------------- -------------------- ------------- ------------ ------------ -------------- -------------- ------------------------------------------------------------------------------ ------------ ------------ -------------- -------------- [eval data]{} (lr)[1-2]{} (lr)[3-7]{} (lr)[8-12]{} [model]{} [train data]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} [all]{} [n.]{} [v.]{} [adj.]{} [adv.]{} (lr)[2-12]{} Word2Vec LP NOAD 0.642 0.733 0.554 0.705 0.725 0.643 0.752 0.524 0.726 0.664 (lr)[2-12]{} LSTM LP NOAD 0.822 0.859 0.800 0.817 0.816 0.831 0.865 0.806 0.825 0.821 LSTM LP NOAD,SemCor 0.872 &0.897 &0.852 &0.865 &0.868\ LSTM LP & NOAD,MASC & 0.873 &0.883 &0.870 &0.858 &0.874 & & & & &\ **************** ---------------------------------------------------- -------------------- ------------- ------------ ------------ -------------- -------------- ------------------------------------------------------------------------------ ------------ ------------ -------------- -------------- #### Change of seed data: As can be seen in Table \[tab:semi-supervised\], LP substantially improves classifier F1 when the training datasets are SemCor+NOAD or MASC+NOAD. As discussed in Section \[sec:semi-supervised-wsd\], the improvement may come from explicitly modeling the sense prior. We did not see much performance lift by increasing the number of unlabeled sentences per lemma. #### Change of graph density: By default, we construct the LP graph by connecting two nodes if their affinity is above $95$% percentile. We also force each node to connect to at least $10$ neighbors to prevent isolated nodes. Table \[tab:density\] shows the performance of the LP algorithm by changing the percentile threshold. The F1 scores are relatively stable when the percentile ranges between $85$ to $98$, but decrease when the percentile drops to $80$. Also, it takes longer to run the LP algorithm on a denser graph. We pick the $95$ percentile in our default setting to achieve both high F1 scores and short running time. ------------------------------------------ --------------- --------------- --------------- ------------ ------------ ------------ ------------ --------------- --------------- ------------ ------------ ------------ (lr)[2-7]{} (lr)[8-13]{} [pos-tag]{} [98]{} [95]{} [90]{} [85]{} [80]{} [70]{} [98]{} [95]{} [90]{} [85]{} [80]{} [70]{} overall [0.823]{} 0.822 [0.823]{} 0.818 0.813 0.800 0.827 0.831 [0.835]{} 0.830 0.824 0.806 n. 0.848 [0.859]{} 0.852 0.846 0.840 0.828 0.863 0.865 [0.868]{} 0.865 0.861 0.847 v. [0.810]{} 0.800 0.803 0.797 0.792 0.778 0.800 [0.806]{} [0.806*]{} 0.799 0.794 0.769 ------------------------------------------ --------------- --------------- --------------- ------------ ------------ ------------ ------------ --------------- --------------- ------------ ------------ ------------ Conclusions and Future Work {#sec:discussion} =========================== In this paper, we have presented two WSD algorithms which combine (1) LSTM neural network language models trained on a large unlabeled text corpus, with (2) labeled data in the form of example sentences, and, optionally, (3) unlabeled data in the form of additional sentences. Using an LSTM language model gave better performance than one based on Word2Vec embeddings. The best performance was achieved by our semi-supervised WSD algorithm which builds a graph containing labeled example sentences augmented with a large number of unlabeled sentences from the web, and classifies by propagating sense labels through this graph. Several unanswered questions suggest lines of future work. Since our general approach is amenable to incorporating any language model, further developments in NNLMs may permit increased performance. We would also like to better understand the [limitations*]{} of language modeling for this task: we expect there are situations – e.g., in idiomatic phrases – where per-word predictions carry little information. We believe our model should generalize to languages other than English, but have not yet explored this. Character-level LSTMs [@kim2015character] may provide robustness to morphology and diacritics and may prove useful even in English for spelling errors and out of vocabulary words. We would like to see whether our results can be improved by incorporating global (document) context and multiple embeddings for polysemous words [@huang2012improving]. Finally, many applications of WSD systems for nominal resolution require integration with resolution systems for named entities, since surface forms often overlap [@moro2014entity; @babelnet]. This will require inventory alignment work and model reformulation, since we currently use no document-level, topical, or knowledge-base coherence features. We thank our colleagues and the anonymous reviewers for their insightful comments on this paper. [^1]: http://www.anc.org/MASC/About.html/ [^2]: We mapped all senses to WordNet3.0 by using maps in https://wordnet.princeton.edu/wordnet/download/current-version/ and http://web.eecs.umich.edu/ mihalcea/downloads.html [^3]: The training corpus could not be released, but we have plans to open source the well-trained models [^4]: http://oxfordgls.com/our-content/english-language-content/ [^5]: https://research.google.com/research-outreach.html\#/research-outreach/research-datasets	{ "pile_set_name": "ArXiv" }
=1200 \#1 \#1 \#1\#2[[[\#1]{}]{}]{} \#1[\_[\#1]{}]{} \#1\#2[[[\#1]{}]{}]{} \#1\#2 [[\#1]{}]{}[[\#2]{}]{} \#1\#2 \#1\#2[\_[[\#1]{}]{}]{} \#1\#2[\_[[\#1]{}]{}]{} \#1\#2[\_[[\#1]{}]{}]{} =cmr10 scaled5 =cmr10 scaled4 =cmr10 scaled3 =cmr10 scaled2 =cmr10 scaled1 =cmr8 =cmb10 scaled5 =cmb10 scaled4 =cmb10 scaled3 =cmb10 scaled2 =cmb10 scaled1 =cmss10 =cmtt10 \#1 =cmr6 =cmr12 =0 =0 =0 \#1[-\#1 =[A]{}=[B]{} =[C]{} =[D]{} =[E]{} =[F]{} =[G]{} =[H]{} =[I]{} =[J]{} =[K]{} =[L]{} =[M]{} =[N]{} =[O]{} =[P]{} =[Q]{} =[R]{} =[S]{} =[T]{} =[U]{} =[V]{} =[W]{} =[X]{} =[Y]{} =[Z]{}]{} \#1[\#1>0 \#1 ]{} \#1[ 1[ CHAPLABEL\#1]{} CHAPLABEL\#1]{} \#1[?? 16[ \\\Undefined Chapter Reference \#1\\\ ]{} 16[ \\\Undefined Chapter Reference \#1\\\ ]{} 2[Chapter \#1]{}]{} \#1[by 1 1[ EQLABEL\#1]{} EQLABEL\#1 (. )]{} \#1[by 1 1[ EQLABEL\#1]{} EQLABEL\#1 (. )]{} \#1[by1 1[ EQLABEL\#1]{} EQLABEL\#1&(. )]{} \#1[(\\\) 16[ \\\Undefined Equation Reference \#1\\\* ]{} 16[ \\\Undefined Equation Reference \#1\\\ ]{} . = .) 2[Equation \#1]{}]{} \#1[by 1 1[ FIGLABEL\#1]{} FIGLABEL\#1. $^{\hbox{\marginstyle #1}}$]{} \#1[!!!!) 16[ \\\Undefined Equation Reference \#1\\\ ]{} 16[ \\\Undefined Equation Reference \#1\\\ ]{} . = . 2[Figure \#1]{}]{} \#1\#2 . \#2 \#1\#2[[. \#2]{}]{} \#1\#2 [. \#2]{} ) \#1[\#1]{} \#1[\#1]{} \#1.\#2?[\#1]{} \#1.\#2?[\#2]{} \#1 \#1 \#1 \#1 EQLABELassumption EQLABELnp.1 EQLABELnp1 EQLABELrep EQLABELnp2 EQLABELnp3 EQLABELareamin EQLABELnp10 EQLABELnp11 FIGLABELlipdiff FIGLABELdiffae EQLABELdiv FIGLABELmeancur EQLABELdivmc EQLABELproblem EQLABELexcessrn FIGLABELbk EQLABELsamevol EQLABELmceleqk FIGLABELberegular FIGLABELbhregular EQLABELlip EQLABELc1 EQLABELsuminh FIGLABELballinmin EQLABELintersectionisball EQLABELesamemeasure EQLABELintmul FIGLABELdensity FIGLABELforms FIGLABELapproxid EQLABELsumdeg FIGLABELHisK EQLABELequalint EQLABELtoprove EQLABELona FIGLABELhar FIGLABELeconvex FIGLABELenested EQLABELnested EQLABELeunique EQLABELsameint EQLABELequality FIGLABELsmoothconvex FIGLABELnotstrict FIGLABEL2d EQLABELnorm2 FIGLABELrearrangement EQLABELbvinequality =0 =0 \#1/\#2[.1em.5ex-.1em /-.15em.25ex]{} 1.7truein Area Minimizing Sets Subject to a Volume Constraint in a Convex Set 1truein 2= 2 =1 =1 [1. Introduction.]{} In this paper we consider the problem of minimizing area subject to a volume constraint in a given convex set. In precise terms we have the following. Let $ \Omega \subset\R^{n}$ be a bounded convex set. Thus, $ \abs{\Omega }<\infty $ where $ \abs{\Omega }$ denotes Lebesgue measure. For a number $ 0<v<\abs{\Omega }$, let $ E\subset\Omega $ denote a set with $ \abs{E}=v$ such that $$P(E) \leq P(F)$$ for all sets $ F\subset\Omega$ with $ \abs{F}=v$, where $ P(E)$ denotes the perimeter of $ E$. The main question we investigate is whether $ E$ is convex. It should be emphasized that the perimeter of a competitor $ F$ is taken relative to $ \R^{n}$, or what is the same, the perimeter is taken relative to the closure of $ \Omega $ since $ F$ is assumed to be a subset of $ \Omega $. This problem is considerably different from minimizing perimeter relative to the interior of $ \Omega $. This was considered in \[Gr\] where it was shown that a minimizer is regular and intersects $ \partial \Omega $ orthogonally. The question of existence of a solution to our problem is resolved immediately in the context of sets of finite perimeter. Regularity questions have been considered by other authors. Tamanini \[T\] has shown that an area minimizing set $ E$ subject to a volume constraint has the property that $\partial E\cap\Omega $ is real analytic except for a closed set whose Hausdorff dimension does not exceed $ n-8$. Also, under the assumption that $ \partial \Omega \in C^{1}$, it was shown in \[GMT2\] that $ \partial E$ is an $ (n-1)$ manifold of class $ C^{1}$ in some neighborhood of each point in $ \partial E\cap\partial \Omega $. In $\R^2$, and in $R^n,\,n>2$ under an additional condition on $ \Omega$, we are able to obtain regularity results and ultimately establish that a minimizer $ E$ is convex. Assuming only that $ \Omega $ is bounded and convex, the convexity of $ E$ is an open question in $\R^n,\,n>2$. The additional condition we impose on $ \Omega $ if $n>2$ is the following. 3= $$\cases{ \box3&\cr}\eqnlbl{assumption}$$ Also, assuming initially that $ \partial \Omega \in C^{2}$ and strictly convex, we invoke a result of \[BK\] to conclude that $\partial E\in C^{1,1}$ at points near $ \partial \Omega $. We then show, Theorem , that $ E$ is convex. Finally, through an approximation procedure, we show that $ E$ is convex with $ C^{1,1}$ boundary assuming only that $ \Omega $ satisfies a great circle condition. Clearly, there is no uniqueness if $ v$ is too small. However, with $ H_{\Omega }$ denoting the union of all largest balls in $ \Omega $, if $ \abs{H_{\Omega }}\leq v<\abs{\Omega }$, then $ E$ is unique. In addition for such $v$ we show that perimeter minimizers $E$ are nested as a function of $v$. In general for nonconvex $\Omega$ one can expect neither uniqueness or nestedness as indicated by examples in \[GMT1\]. The nestedness of perimeter minimizers allows one to rearrange level sets of functions to create test functions useful in studying minimizers to certain variational problems. For domains $\Omega$ having certain symmetries it is frequently possible to apply symmetrization to gain information on minimizers of functionals such as $$\int_\Omega \|\nabla u\|^p +\int_\Omega F(u) + \int_0^{\|\Omega\|} G(u^,{u^}')$$ over appropriate function classes , where $u^$ is the decreasing rearrangement of $u$. However this greatly restricts the collection of domains which can be considered. In Section 4 for the case $p=1$ we construct a rearrangement which retains various useful properties of symmetrization while allowing a much larger class of domains to be considered, namely those convex domains described above. This rearrangement is useful when one has a boundary condition of the form $u=0$ on $\partial\Omega$ and when it can be established, for instance using truncation, that $u\ge0$ in $\Omega$. Since this rearrangement produces functions of bounded variation it is more accurate to replace $\int \|\nabla u\|$ in the functional above by the BV norm. The results of Section 3 allow one to deduce certain regularity properties for minimizers $u$ from regularity properties of $u^$. In addition they establish the convexity of the sets $\{u>t\}$. Results in \[LS\] show that one can not hope for similar results if $p>1$. =2 [2. Notation and Preliminaries.]{} The Lebesgue measure of a set $E\subset \R^{n}$ will be denoted by $\abs{E}$ and $H^{\alpha}(E)$, $\alpha>0,$ will denote its $\alpha$-dimensional Hausdorff measure. If $\Omega\subset \R^{n}$ is an open set, the class of functions $u\in L^{1}(\Omega)$ whose partial derivatives in the sense of distributions are measures with finite total variation in $\Omega$ is denoted by $BV(\Omega)$ and is called the space of [functions of bounded variation in $\Omega.$]{} The space $BV(\Omega)$ is endowed with the norm $$\norm{u}_{BV(\Omega)}=\norm{u}_{1;\Omega}+\norm{\nabla u}(\Omega)\eqnlbl{np.1}$$ where $\norm{u}_{1;\Omega}$ denotes the $L^{1}$-norm of $u$ on $\Omega$ and where $\norm{\nabla u}$ is the total variation of the vector-valued measure $\nabla u$. A Borel set $E\subset \R^{n}$ is said to have [finite perimeter in $\Omega$]{} provided the characteristic function of $E$, $\upchi_{E}$, is a function of bounded variation in $\Omega$. Thus, the partial derivatives of $\upchi_{E}$ are Radon measures on $\Omega$ and the perimeter of $E$ in $\Omega$ is defined as $$P(E,\Omega)=\norm{\nabla\upchi_{E}}(\Omega).\eqnlbl{np1}$$ A set $E$ is said to be of [locally finite perimeter]{} if $P(E,\Omega)<\infty$ for every bounded open set $\Omega\subset \R^{n}$. The definition implies that sets of finite perimeter are defined only up to sets of measure 0. In other words, each set determines an equivalence class of sets of finite perimeter. In order to avoid this ambiguity, whenever a set $E$ of finite perimeter is considered we shall always employ the measure theoretic closure as the set to represent $E$. Thus, with this convention, we have $$x\in E\;\hbox{if and only if}\;\limsup_{r\to0}\frac{\abs{E\cap B(x,r)}}{\abs{B(x,r)}}>0. \eqnlbl{rep}$$ One of the fundamental results of the theory of sets of finite perimeter is that they possess a measure-theoretic exterior normal which is suitably general to ensure the validity of the Gauss-Green theorem. A unit vector $\nu$ is defined as the exterior normal to $E$ at $x$ provided $$\lim_{r\to0}r^{-n}\abs{B(x,r)\cap\{y:(y-x)\cdot\nu<0,y\notin E\}}=0$$ and $$\lim_{r\to0}r^{-n}\abs{B(x,r)\cap\{y:(y-x)\cdot\nu>0,y\in E\}}=0,\eqnlbl{np2}$$ where $B(x,r)$ denotes the open ball of radius $r$ centered at $x$. The measure-theoretic normal of $E$ at $x$ will be denoted by $\nu(x,E)$ and we define $$\partial^{}E=\{x:\nu(x,E)\;{\rm exists}\}.\eqnlbl{np3}$$ Clearly, $\partial^{}E\subset\partial E$, where $\partial E$ denotes the topological boundary of $E$. A set $E$ of finite perimeter is said to be [area minimizing in an open set]{} $\Omega$ if $$\norm{\nabla \upchi_{E}}(\Omega)\leq\norm{\nabla \upchi_{F}}(\Omega)\eqnlbl{areamin}$$ for every set $F$ with $F\Delta E\subset\subset \Omega$. Here $F\Delta E$ denotes the symmetric difference. The regularity of $\partial E$ will play a crucial role in our development. Suppose $\partial E$ is area minimizing in $ U$ and for convenience of notation, suppose $0\in U\cap\partial E$. For each $r>0$, let $E_{r}=\R^{n}\cap\{x:rx\in E\}$. It is known (cf. \[S,§35\], \[MM,§2.6\]) that for each sequence $\{r_{i}\}\to0$ there exists a subsequence (denoted by the full sequence) such that $\upchi_{ E_{r_{i}}}$ converges in $L^{1}_{\rm loc}(\R^{n})$ to $\upchi_{C}$, where $C$ is a set of locally finite perimeter. In fact, $\partial C$ is area minimizing and is called the tangent cone to $E$ at 0. Although it is not immediate, $C$ is a cone and therefore the union of half-lines issuing from 0. It follows from \[S, §37.6\] that if $\overline{C}$ is contained in $\overline{H}$ where $H$ is any half-space in $\R^{n}$ with $0\in\partial H$, then $\partial E$ is regular at 0. That is, there exists $r>0$ such that $$B(0,r)\cap\partial E\;\,\hbox{is a real analytic hypersurface.}\eqnlbl{np10}$$ Furthermore, $\partial E$ is regular at all points of $\partial ^{}E$ and $$H^{\alpha}((\partial E - \partial ^{}E)\cap U)=0 \quad\hbox{\rm for all}\;\alpha>n-8,\eqnlbl{np11}$$ cf. \[Gi, Theorem 11.8\]. The notion of [excess]{} plays a critical role in the theory of minimal boundaries. It measures how far a set $E$ is from being area minimizing in a ball. Formally, it is defined by $$\psi(x,r)=\norm{\nabla \upchi_{E}}(B(x,r))-\inf\{\norm{\nabla \upchi_{F}}(B(x,r)):F\Delta E\subset\subset B(x,r)\}.$$ Thus, $\psi\equiv0$ when $E$ is area minimizing. If $E$ is an arbitrary set of finite perimeter and $\psi(x,r)\leq Cr^{n-1+2\alpha}$ for some $x\in\partial E$ and all $0<r<R$ with given constants $C, R$ and $0<\alpha<1$, then it follows from a result of Tamanini \[T\] that there is an area minimizing tangent cone to $\partial E$ at $x$. Let $M$ denote a $k$-dimensional $C^{1}$ submanifold of $\R^{n},\>0<k<n,$ and let $f\colon M\to\R$ be an arbitrary function. We will say that $f$ is differentiable at $x_{0}\in M$ if $f$ is the restriction to $M$ of a function $\bar f\colon U\to\R$ where is $U\subset\R^{n}$ is some open set containing $x_{0}$ and where $\bar f$ is differentiable at $x_{0}$. The manifold $M$ near any of its points $x_{0}$ can be represented as the graph of a function defined on some open $n-1$-ball $B'\subset\R^{n-1}$. Thus, there is an open $n$-cylinder $C$ of the form $C=B'\times(a,b)$ such that $C-M$ consists of two nonempty connected, open sets and that each projection of $M\cap C$ onto the top and bottom of $C$ is a homeomorphism. Let points $x\in C$ be denoted by $x=(x',y)$ where $x'\in B'$ and $y\in(a,b)$ and define $\bar f\colon C\to\R$ by $\bar f(x',y)=f(x',y_{M})$ where $(x',y_{M})$ is that unique point on $M\cap C$ that is the projection of $(x',y)$. It is easy to verify that $\bar f\colon C\to\R$ is Lipschitz and therefore, by Rademacher’s theorem, that $\bar f$ is differentiable at (Lebesgue) almost all points of $C$. Let $N$ denote those points at which $\bar f$ is not differentiable. Clearly, if $\bar f$ is differentiable at a point $(x',y_{1})$ then it is differentiable at any other point of the form $(x',y_{2})$. Now define $d\colon C\to\R$ by $d(x',y)=\abs{y-y_{M}}$. Note that $d$ is Lipschitz and that $d^{-1}(t)$ consists of two copies of $M\cap C$, one is a vertical distance of $t$ units above $M\cap C$ and the other is a vertical distance of $t$ units below $M\cap C$. Now employ the co-area formula to obtain $$0=\int_{N}\abs{\nabla d}\;dx=\int_{a}^{b}H^{n-1}[d^{-1}(t)\cap N]\;dt.$$ Thus, for almost every $t\in(a,b),\>\bar f$ is differentiable at $H^{n-1}$ almost all points of $d^{-1}(t)$. Consequently, $\bar f$ is differentiable at the corresponding points of $d^{-1}(0)=M\cap C$; that is, $\bar f$ is differentiable at $H^{n-1}$ almost all points of $M\cap C$, as required.In view of the preceding Lemma, we can define the directional derivative of $f$ relative to $M$ at $H^{n-1}$-almost all $x\in M$ in the usual manner. Given a vector $\tau$ in the tangent space to $M$ at $x$, let $\gamma\colon(-1,1)\to M$ be any $C^{1}$ curve with $\gamma(0)=x$ and $\gamma'(0)=\tau$. Define $$D_{\tau}f(x)=(\bar f\circ\gamma)'(0)$$ where it is understood that $\bar f$ is differentiable at $x$. Observe that this definition is independent of the extension $\bar f$. If we are given a Lipschitz vector field $X\colon M\to\R^{n}$, by using usual methods, it now becomes clear how to define the divergence of $X$ relative to $M$, denoted by $\div_{M}X$. If the closure $\overline{M}$ of $M$ is a $C^{1}$ manifold with boundary $\partial M=\overline{M}-M$ and if $X\colon\R^{n}\to\R^{n}$ is a $C^{1}$ vector field with the property that for each $x\in M,\;X(x)$ is an element of the tangent space to $M$ at $x$, then the classical divergence theorem states $$\int_{M}\div_{M}X\;dH^{n-1}=\int_{\partial M}X\cdot\eta \;dH^{n-2}\eqnlbl{div}$$ where $\eta$ is the outward pointing unit co-normal of $\partial M$. That is, $\abs{\eta}=1$, $\eta$ is normal to $\partial M$, and tangent to $M$. Let $M$ be an oriented $n-1$-dimensional submanifold of $\R^{n}$ of class $C^{1,1}$; that is, $M$ is of class $C^{1}$ and its unit normal $\nu$ is Lipschitz. From Lemma , we have that the components of $\nu$ are differentiable at $H^{n-1}$ almost all points of $M$. Thus, $\div_{M}\nu$ is defined $H^{n-1}$ almost everywhere on $M$. At such points, we define the [mean curvature]{} of $M$ at $x$ as $${\cal H}_{M}(x)=\div_{M}\nu(x)$$ If $X\colon\R^{n}\to\R^{n}$ is a $C^{1}$ vector field, consider its decomposition into its tangent and normal parts relative to $M$, $$X=X^{\top}+X^{\bot}$$ where $$X^{\bot}=(X\cdot\nu)\nu.$$ Then, at $H^{n-1}$ almost all points in $M$, it follows that $$\div_{M}X^{\bot}=(X\cdot\nu)\div_{M}\> \nu.$$ Hence, $$\div_{M}X^{\bot}={\cal H}_{M}X\cdot\nu.$$ On the other hand, from we have $$\int_{M}\div_{M}X^{\top}\;dH^{n-1}=\int_{\partial M}X\cdot\eta\;dH^{n-2}.$$ Since $\div_{M}X=\div_{M}X^{\top}+\div_{M}X^{\bot}$, we obtain $$\int_{M}\div_{M}X\;dH^{n-1}=\int_{M}{\cal H}_{M}X\cdot\nu\;dH^{n-1}+\int_{\partial M}X\cdot\eta\;dH^{n-2}.\eqnlbl{divmc}$$ =3 =0 =0 [3. Main Results]{} In this section we consider the following situation. 2= $$\cases{ \box2&\cr}\eqnlbl{problem}$$ We will first establish boundary regularity and curvature properties for such perimeter minimizers under the assumption that $\Omega$ is [strictly convex and that]{} $\partial \Omega\in C^{2}$. Convexity, nestedness and uniqueness results will then be established under the further assumption that $$n=2 \quad\hbox{or}\quad \Omega \hbox{ satisfies a great circle condition. }$$ The assumption of strict convexity and $C^2$ regularity will then be dispensed with in part through an approximation argument. Associated with is some further notation. We let $H$ denote the convex hull of a minimizer $E$ of , and we denote by $H^{+}$ that part of $H$ that lies “above” the equatorial disk $D_{B_{\Omega }}$ of $B_{\Omega}$ as defined in . Since $P$ divides $H$ into two parts, we arbitrarily call one of them the part that lies “above” $P$. Next, we recall some facts concerning area minimizing sets with a volume constraint. The main result of \[GMT1\] is that if $E$ is area minimizing with a volume constraint, then $$\psi(x,r)\leq Cr^{n}\eqnlbl{excessrn}$$ for each $x\in\partial E$ and for all sufficiently small $r>0$. Consequently, it follows from work of Tamanini \[T\] that an area minimizing set $E$ with a volume constraint possesses an area minimizing tangent cone at each point of $(\partial E)\cap\Omega$. From this it follows that $(\partial E)\cap\Omega$ enjoys the same regularity properties as an area minimizing set; that is, $(\partial E)\cap\Omega$ is real analytic except for a closed singular set $S$ whose Hausdorff dimension does not exceed $n-8$. Furthermore, it was established in \[GMT2, Theorem 3\] that $\partial E$ is an $(n-1)$ manifold of class $C^{1}$ in some neighborhood of each point $x\in\partial E\cap\partial \Omega$. The object of this section is to prove that $E$ is convex and we begin by proving $C^{1,1,}$ regularity of $\partial E$ near $\partial \Omega$. For this we will need the following result of Brézis and Kinderlehrer, \[BK\]. We now apply this result to obtain $C^{1,1}$ regularity of the boundary of a minimizer $E$ of the variational problem near $\partial \Omega$ . Since $\partial E$ is an $(n-1)$ manifold of class $C^{1}$ in some neighborhood of each point $x\in\partial E\cap\partial \Omega$, it follows that near such a point $x$, we may represent both $\partial E$ and $\partial \Omega$ as graphs of functions $u$ and $\beta$, respectively, defined on an open set $U'\in\R^{n-1}$ containing $x'$ where $x=(x',y''),\;y''\in\R$. We will assume $u$ and $\beta$ chosen in such a way that $u\geq\beta,\;u=0$ on $\partial U'$ and $\beta\leq0$ on $\partial U'$. Using the convexity of $\Omega$, this can be accomplished by considering a hyperplane $P_{0}$ passing through $E$ and parallel to the tangent plane to $\partial E$ at $x$. By taking $P_{0}$ sufficiently close to the tangent plane, $U'$ can be defined as $P_{0}\cap E$. Now select $v\in {\bf K}$ and for $0<\varepsilon<1$, define $u_{\varepsilon}$ on $U'$ as $u_{\varepsilon}=u+\varepsilon(v-u)$. We will assume $\varepsilon$ chosen small enough so that the graph of $u_{\varepsilon}$ remains in $\overline{\Omega}$. Note that $u_{\varepsilon}\in {\bf K}$. Select a point $z\in(\partial E)\cap\Omega$ at which $\partial E$ is regular. Thus, $\partial E$ is real analytic near $z$ and its mean curvature is a constant $K$ there. In a neighborhood of $z$, we can represent $\partial E$ as the graph of a function $w$ defined on some open set $V'\subset\R^{n-1}$ containing $z'$ where $z=(z',z'')$. The neighborhoods about $x$ and $z$ where $\partial E$ is represented as a graph are taken to be disjoint. Let $\varphi\in C^{\infty}_{0}(V')$ denote a function with the property that $$\int_{V'}\varphi\;dH^{n-1}=\int_{U'}(v-u)\;dH^{n-1},\eqnlbl{samevol}$$ and define $w_{\varepsilon}=w-\varepsilon\varphi$. The graphs of the functions $u_{\varepsilon}$ and $w_{\varepsilon}$ produce a perturbation of the set $E$, say $E_{\varepsilon}$. Because of , we have that $\abs{E}=\abs{E_{\varepsilon}}$. With $$F(\varepsilon)=\int_{U'}\sqrt{1+\abs{\nabla u_{\varepsilon}}^{2}}+ \int_{V'}\sqrt{1+\abs{\nabla w_{\varepsilon}}^{2}},$$ the minimizing property of $\partial E$ implies that $F(0)\leq F(\varepsilon)$ for all small $\varepsilon$ and therefore that $F'(0)\geq0$. Thus, $$\int_{U'}\frac{\nabla u}{\sqrt{1+\abs{\nabla u}^{2}}}\cdot\nabla (v-u) -\int_{V'}\frac{\nabla w}{\sqrt{1+\abs{\nabla w}^{2}}}\cdot\nabla \varphi\geq0.$$ Since $w$ has constant mean curvature $K$, we obtain $$\int_{V'}\frac{\nabla w}{\sqrt{1+\abs{\nabla w}^{2}}}\cdot\nabla \varphi=-\int_{V'}K\varphi=-K\int_{V'}\varphi=-K\int_{U'}(v-u),$$ and therefore $$\int_{U'}\frac{\nabla u}{\sqrt{1+\abs{\nabla u}^{2}}}\cdot\nabla (v-u) \geq -K\int_{U'}(v-u).\eqnlbl{mceleqk}$$ If $\eta\in C^{\infty}_{0}(U')$ denotes an arbitrary nonnegative test function, then with $v-u=\eta$, states that $u$ is a weak solution of $\mce\leq K$. This combined with the $C^{1,1}$- regularity of $u$ implies that $\mce\le K$ pointwise almost everywhere in a neighborhood of $\partial \Omega$. Since $\mce=K$ in $\partial E\cap(\Omega\setminus S)$ with $H^{n-1}(S)=0$ we have the following result. We now will exploit Theorem to establish both regularity and a mean curvature estimate for the boundary of the convex hull of $E$. Note that the singular set $S$ in $\partial E$ is a closed subset of $\Omega$ and thus separated from $\partial\Omega$, in fact it is contained in the interior of H, for if $x\in\partial E\cap\partial H\cap\Omega$, then the tangent cone to $\partial E$ at $x$ must be a hyperplane because $E\subset H$ and $H$ is convex. Consequently $\partial E$ is regular at $x$. Let $N$ be an open neighborhood of $S$ with compact closure in the interior of $H$. Thus by Theorem and the analyticity of $\partial E$ in $\Omega\setminus S$ we see that $\partial E$ is $C^{1,1}$ at points in $G:=\partial E\setminus N$. Therefore for some $C$ we have $$\|\nu(x)-\nu(z)\|\le C\|x-z\|\qquad\qquad x, z\in G\eqnlbl{lip}$$ where $\nu(x)$ is the outward unit normal to $\partial E$ at $x$. Also since $\partial E$ is $C^1$ at points in $G$ there exists an $\varepsilon$ such that for all $x\in G$ and $z\in\partial E\cap B(x,\varepsilon)$ we have $$\|\nu(x)\cdot(x-z)\|\le\fracnum{1}/{2}\|x-z\|.\eqnlbl{c1}$$ Choose $x\in\partial E\cap\partial H\subset G$ and let $ 0<\alpha <1/2$. Then define $$d=\alpha \min\{\varepsilon, \hbox{ dist}(\partial H, N), {1\over 2C},\hbox{\rm diam}\>E\}.$$ Let $ y=x-d\nu (x)$ and observe that $ y$ is in the interior of $ E$ since $\partial E$ cannot intersect the line segment $\overline{xy}$ at a point $z\not=x$ due to . Let $r=\hbox{ dist}(y,\partial E)$ and note that $0<r\le d$. Now choose any $z\in\partial E$ such that $\|y-z\|=r$. Note that $z\in G$, for otherwise we would have $ z\in N$ and since $ \abs{x-z}\leq\abs{x-y}+\abs{y-z}$, it would follow that $$2d\geq\abs{x-z}\geq\hbox{\rm dist}(\partial H,N)\geq\frac{d}{\alpha }>2d,$$ a contradiction. Then, $\|x-z\|\le\|x-y\|+\|y-z\|\le2 d<\varepsilon$ and both and hold. Thus, since $x=y+d\nu(x)$ and $z=y+r\nu(z)$, we have $\|d-r\|\le\|\nu(x)\cdot(x-z)\|$ and $$\|x-z\|=\|(d-r)\nu(x)+r(\nu(x)-\nu(z))\|\le(\fracnum{1}/{2}+C r)\|x-z\|\le\fracnum{3}/{4}\|x-z\|,$$ (since $r\le d\le \alpha/(2c)\le 1/(4c)$) which implies that $x=z$ and therefore $r=d$. This implies that for every $x\in \partial E\cap\partial H$ there exists a ball $B_x\subset E$ of radius $d$ containing $x$. Given any $p\in \partial H$ we claim that $p$ is a convex combination of points $\{x_i\}$ in $\partial E\cap\partial H$. To see this note that if $C$ is a convex set with $E\subset C$ then $\overline E\subset C$ since if $x\in \overline E$ then either $x\in C$ or $x\in\partial C$; in the later case $x$ lies in a support plane of $C$ so if $x\in \Omega$, regularity theory implies that $x\in E\subset C$, and if $x\in \partial\Omega$ then $x$ is not in the singular set $S$ of $E$ (since $S$ is a compact subset of $\Omega$) so again $x\in E\subset C$. Consequently from the definition of convex hull $H$ of $E$ as the intersection of all convex sets containing $E$, we see that $\overline E\subset H$. Moreover $H$ is the convex hull of $\overline E$ from which we conclude by a well known result that $H$ is closed since $\overline E$ is a compact subset of $\R^n$. Note that the set of finite convex combinations of points from $E$ is convex, contains $E$, and is contained in any convex set which contains $E$ and so equals $H$. Thus if $p\in\partial H$ we have $p\in H$, since $H$ is closed, and consequently $p=\sum_{i=1}^k \lambda_i x_i$ for $x_i\in E$ and $\sum_{i=1}^k \lambda_i=1,\, \lambda_i\ge0, i=1\dots k$. If we take $k$ to be as small as possible then either $k=1$ and $p\in E$ and the claim is trivially true, or $p$ lies in the $k$ dimensional interior of the convex hull $M$ of $\{x_i\}$ in which case no $x_i$ can lie in the interior of $H$ since then the same would be true of $p$. Consequently $x_i\in \partial E\cap\partial H,\,i=1\dots k$, as claimed. Taking the convex hull of $\cup_{i=1}^k B_{x_i}$ we see that there exists a ball $B_p\subset H$ of radius $d$ containing $p$, i.e. $H$ satisfies a uniform interior sphere condition. We claim that this implies $\partial H$ is $C^{1,1}$. To see this, consider the problem of prescribing unit vectors $\nu_1, \nu_2\in\R^n$, and finding a convex set $\tilde H$, satisfying the interior sphere condition noted above, and points $x, y\in\partial\tilde H$ with $\nu(x)=\nu_1, \nu(y)=\nu_2$, such that $\|x-y\|$ is minimized. It is clear that $x, y$ must lie in a two dimensional plane orthogonal to the intersection of two hyperplanes having $\nu_1, \nu_2$ as normals, i.e. one need only consider the two dimensional case where it is easy to see that one must have $B_x=B_y$. Taking the center of this ball to be the origin then $\nu(x)=x/d, \nu(y)=y/d$ and we trivially have $$\|\nu(x)-\nu(y)\|\le{1\over d}\|x-y\|.$$ Since this is the case when $\|x-y\|$ is smallest for fixed $\nu(x), \nu(y)$ we have established that $\nu(x)$ is Lipschitz in general. We now prove that $\mch\leq K\;H^{n-1}$-almost everywhere in $\partial H$. Note that $\mch=\mce$ $H^{n-1}$-almost everywhere on $\partial E\cap\partial H$ by Theorem . Thus we need only consider points $p\in\partial H\setminus\partial E$. In fact since $\partial H$ is $C^{1,1}$ we need only consider $p\in\partial H\setminus\partial E$ at which $\partial H$ is classically twice differentiable. As above, any such $p$ lies in the $k$ dimensional interior of the convex hull $M$ of certain points $p_i\in \partial E,\, i=1,\dots, k$. Note that $k\ne 1$ due to $p\notin\partial E$. Choose a coordinate system such that points in $\R^n$ are represented as $(x,y,z),\,x\in\R^k,\, y\in\R^{n-k-1},\, z\in\R$, with $z=0$ the tangent plane to $\partial H$ at $p$, $p_i=(x_i,0,0),\, i=1,\dots, k$, and $z\ge 0$ in $H$. We will construct an analytic function $g$ whose graph does not lie below $\partial H$, contains $M$, and has mean curvature bounded above by $K+\varepsilon$ (for any $\varepsilon>0)$ in a small neighborhood of $p$. This will lead to the conclusion that $\mch\le K$ at $p$. Let $\partial E$ be represented as $z=f(x,y)$ for $f$ defined in a neighborhood in $\R^{k}\times\R^{n-k-1}$ of $\cup(x_{i},0)$. Thus $$(x_i,y, f(x_i,y))\in \partial E\subset H$$ for small $\|y\|$, and consequently $$\sum_{i=1}^k \lambda_i(x_i,y,f(x_i,y)) \in H \qquad \hbox{ if }\quad \sum_{i=1}^k \lambda_i=1,\, \lambda_i\ge0 \eqnlbl{suminh}$$ for small $\|y\|$. For any given $x$ in $N$, where $N$ is the convex hull of the points $x_i,\,i=1,\dots, k$, let $\lambda=\lambda(x)=(\lambda_1(x),\dots, \lambda_k(x))$ be the unique vector such that $$x=\sum_{i=1}^{k} \lambda_i(x) x_i,\,\quad \sum_{i=1}^k \lambda_i(x)=1,\,\quad \lambda_i(x)\ge0.$$ Thus if we define $$g(x,y)=\sum_{i=1}^k \lambda_i(x) f(x_i,y)$$ we see from for $x\in N$ and small $\|y\|$ that $$(x,y,g(x,y))\in H,$$ and so the surface $z=g(x,y)$ does not lie below $\partial H$ at such $(x,\,y)$. Note that $M\cap\partial \Omega =\nullset$, for otherwise the plane $z=0$, which contains $M$, would be a tangent plane to $\partial \Omega $, thus contradicting the strict convexity of $\partial \Omega$. Also $M$ does not intersect the singular set of $\partial E$ since $M\subset \partial H$. Thus $\partial E$ is analytic at each $p_i$ and therefore both $f(x_i,y)$ and $g(x,y)$ are smooth for small $\|y\|$. Furthermore, $$0\le \Delta_y f(x_i,0) \le \Delta f(x_i,0) \le K$$ since $\nabla f(x_i,0)=0$, $\mce$ equals $ \Delta f$ at points where the gradient is zero, and the second derivatives of $f$ are nonnegative at $(x_i,0)$ due to the fact that $f\ge 0$, $f(x_i,0)=0$ for all $i$. Hence, for any $\varepsilon>0$, $\Delta_y f(x_i,y)\le (K+\varepsilon)$ for small enough $\|y\|$ so $\Delta_y g(x,y)\le (K+\varepsilon)$ as well. However $\Delta_x g=0$ and so $\Delta g \le (K+\varepsilon)$ for small $\|y\|$. Recall that $\partial H$ is trapped between $\{z=0\}$ and the graph of $g$ over a region which contains $p$ in its interior. Since $g(p)=0$ and $\partial H$ is twice differentiable at $p$ we conclude that $\mch(p)\le K$ as required. [ Theorem.]{} [Assume that $\Omega$ is bounded, strictly convex and satisfies a great circle condition. If $E$ is a minimizer of with $\|B_\Omega\|\le\|E\|$ then $$B_\Omega\subset E$$ where $B_\Omega$ is the largest ball in $\Omega$.]{} If $\|E\|=\|B_\Omega\|$ then clearly $E$ must be a ball. Since there is only one largest ball in $\Omega$ due to strict convexity, we have $E=B_\Omega$. Otherwise $\|B_\Omega\|<\|E\|$. In this case translate the upper and lower hemispheres of $B_\Omega$ by a distance $d$ in opposite directions orthogonal to $T_{B_{\Omega }}$ until $H$, the convex hull of the two translated hemispheres, intersects $E$ in a set of measure $\|B_\Omega\|$ i.e. $$\|H\cap E\|=\|B_\Omega\|.\eqnlbl{intersectionisball}$$ This is possible because of the great circle conition and because $\Omega $ is bounded and convex. Now translate the hemispheres back to their original positions while rigidly carrying along the parts of $E$ lying in the exterior of $H$. Let $\tilde E$ be the union of the translated parts of $E$ with $B_\Omega$. Note that $$\|\tilde E\|=\|E\|\quad\hbox{\rm and therefore} \quad P(\tilde E)\geq P(E).\eqnlbl{esamemeasure}$$ Using a standard inequality, cf. \[MM\], we have $$P(E)+P(H)\ge P(E \cap H)+P(E \cup H)$$ where $P(S)$ denotes $P(S,\,\R^n)$. For brevity, write $D=D_{B_{\Omega }}$. Observe that $$P(H)=2dH^{n-2}(\partial D)+P(B_\Omega), \quad P(E \cup H)=P(\tilde E)+2dH^{n-2}(\partial D)$$ and thus $$P(E)+P(B_\Omega)\ge P(E \cap H)+P(\tilde E).$$ In view of it follows that $P(E\cap H)\leq P(B_{\Omega })$. But then the isoperimetric inequality and imply that $E\cap H$ is a ball. However $\Omega$ contains only one largest ball and so we must have $E\cap H=B_\Omega$, i.e. $B_\Omega\subset E$.Suppose $M$ is an oriented $(n-1)$-dimensional $C^{1}$ submanifold of $\R^{n}$ and $f\colon M\to\R^{n-1}$ a $C^{1}$ mapping. Let $Jf(x)$ denote the Jacobian of $f$ at $x$ and note that the sign of the Jacobian depends on the orientation of $M$. We recall the following result, cf. \[Fe, Theorem 3.2.20\]: For any $H^{n-1}$-measurable set $E\subset M$ and any $H^{n-1}$-measurable function $\varphi$, $$\int_{E}\varphi[f(x)]\abs{Jf(x)}\;dH^{n-1}(x)=\int \varphi(y)N(f,E,y)\;dy\eqnlbl{intmul}$$ where $N(f,E,y)$ denotes the number (possibly infinite) of points in $f^{-1}(y)\cap E$. Here equality is understood in the sense that if one side is finite, then so is the other. In our application below, we will know the left side is finite, therefore ensuring that $N(f,E,y)$ is finite for almost all $y$. It follows from that we may as well assume $\partial E$ is area minimizing. In this case the result follows immediately from the fact that $$\frac{H^{n-1}((\partial E)\cap B(x,r))}{r^{n-1}}$$ is nondecreasing in $r$, for $r>0$ sufficiently small, cf. \[Fe, Theorem 3.4.3\].[ Lemma.]{} [For every $\varepsilon>0$ and any open set $V\subset\R^{n}$ containing the singular set $S$ of $\partial E$, there exists an open set $W$ and a Lipschitz function $f$ such that $$\displaylines{ S\subset W\subset\{f=1\}\cr \hbox{\rm spt}\;f\subset V\cr \int_{\partial E}\abs{\nabla f}\;dH^{n-1}\leq \varepsilon.\cr}$$]{} Let $V$ be any open set containing $S$ and let $\delta=1/2(\hbox{\rm dist}\;S,\R^{n}-V)$. Since $H^{n-7}(S)=0$ and $S$ is compact, there is a finite collection of open balls $\{B(x_{i},r_{i})\}_{i=1}^{m}$ such that $2r_{i}<\delta, B(x_{i},r_{i})\cap S\not=\nullset, S\subset\cup_{i=1}^{m}B(x_{i},r_{i})$ and $$\sum_{i=1}^{m}r_{i}^{n-7}< \frac{\varepsilon}{C},$$ $C$ as in Lemma . We will assume that each ball $B(x_{i},r_{i})$ has been chosen so that $r_{i}<1$ and that $2r_{i}$ satisfies Lemma . Let $W$ denote the union of these balls and define $f_{i}$ by $$f_{i}(x)=\cases{ 1&if $\abs{x-x_{i}}\leq r_{i}$\cr 2-\frac{\abs{x-x_{i}}}{r_{i}}&if $r_{i}\leq\abs{x-x_{i}}\leq2r_{i}$\cr 0&if $2r_{i}\leq\abs{x-x_{i}}$.\cr}$$ In view of Lemma , it follows that $$\int_{B(x_{i},r_{i})\cap\partial E}\abs{\nabla f_{i}}\;dH^{n-1}\leq Cr_{i}^{n-2}<Cr_{i}^{n-7}.$$ Now let $f:=\max_{1\leq i\leq m}f_{i}$. Then $f$ is Lipschitz, $W\subset\{f=1\},\;\hbox{\rm spt}\;f\subset V$ and $$\eqalign{ \int_{\partial E}\abs{\nabla f}\;dH^{n-1}&\leq\sum_{i=1}^{m}\int_{B(x_{i},r_{i})\cap\partial E}\abs{\nabla f_{i}}\;dH^{n-1}\cr &<C\sum_{i=1}^{m}r_{i}^{n-7}<\varepsilon.\cr}$$ Clearly, the support of $\partial T$ contains the $n-2$-sphere, but we must rule out the possibility of it containing points of $S$ as well. For this purpose, choose $x\in S$ and let $\varphi$ be any smooth differential form supported in some neighborhood of $x$ that does not meet $(\partial E)^{+}\cap\partial D$. It suffices to show that $T(d\varphi)=0$. Let $\mu$ denote $H^{n-1}$ restricted to $(\partial E)^{+}$. Appealing to Lemma , we can produce a sequence of Lipschitz functions $\{\omega_{i}\}$ such that $$\displaylines{ \omega_{i}\to1\;\mu\;\hbox{\rm a.e.}\cr \abs{\nabla \omega_{i}}\to0\;\mu\;\hbox{\rm a.e.}\cr \omega_{i}\;\hbox{\rm vanishes in a neighborhood of}\;S\cr \int_{(\partial E)^{+}}\abs{\nabla \omega_{i}}\;d\mu\to0.\cr}$$ Thus, we obtain $$\eqalign{ 0=T(d(\varphi\omega_{i})) &=T(d\varphi\wedge\omega_{i})+T(\varphi \wedge d\omega_{i})\cr &=\int_{(\partial E)^{+}}d\varphi\wedge\omega_{i}+\int_{(\partial E)^{+}}\varphi\wedge d\omega_{i}.\cr}$$ The first integral tends to $$\int_{(\partial E)^{+}}d\varphi=T(d\varphi)$$ while the second tends to $0$. Thus, $T(d\varphi)=0$. Let $E$ denote a minimizer of , where $\Omega$ is strictly convex with $C^2$ boundary. Since $\partial E$ is locally an $n-1$-manifold of class $C^{1}$ except for a singular set $S$ whose Hausdorff dimension does not exceed $n-8$, it follows that $\partial E$ can be regarded as an oriented $n-1$ integral current whose boundary is $0$; i.e. an oriented $n-1$ integral cycle. Let $T$ denote the $n-1$ integral current represented by $(\partial E)\cap H^{+}$. Since $\partial E$ is of class $C^{1,1}$ in a neighborhood of each point of $(\partial E)\cap(\partial \Omega)$, it follows that the tangent cone to $\partial E$ at such points is in fact a tangent plane. Consequently, $\partial E$ is analytic near such points and therefore the singular set $S$ of $\partial E$ lies in the interior of $(\partial E)\cap H^{+}$. We know from Lemma that the boundary of $T$ is the $n-2$-sphere determined by $\partial D_{B_{\Omega }}$, the equator of $B_{\Omega}$. Let $p\colon\R^{n}\to T_{B_{\Omega }}$ denote the orthogonal projection and consider the current $R:=p_{\#}(T)$. Note that $\partial R=p_{\#}(\partial T)=\partial D_{B_{\Omega}}$. Furthermore, $D_{B_{\Omega}}$ is the unique current in $T_{B_{\Omega }}$ whose boundary is $\partial D_{B_{\Omega}}$ and therefore, we conclude that $R=D_{B_{\Omega}}$. Let us consider the action of $R$ operating on an $n-1$-form $\varphi$. For this we will let $\alpha(x)$ denote the Grassman $(n-1)$-vector of norm one that is in the tangent plane orthogonal to $\nu(E,x)$, the exterior normal to $E$ at $x$. $\alpha(x)$ is chosen in such a way that $\alpha(x)\wedge \nu(E,x)$ forms the Grassman unit $n$-vector that induces a positive orientation of $\R^{n}$. Also, we let $dp(\alpha(x))$ denote the value of the differential of $p$ operating on $\alpha(x)$. Then, with the help of , we have $$\eqalign{ R(\varphi)&=T(p^{\#}\varphi)\cr &=\int_{(\partial E)\cap H^{+}}p^{\#}\varphi\cdot\alpha\cr &=\int_{(\partial E)\cap H^{+}}\varphi[p(x)]\cdot dp(\alpha(x))\;dH^{n-1}(x)\cr &=\int_{D_{B_{\Omega}}}\varphi(y)[N^{+}(p,\partial E,y) -N^{-}(p,\partial E,y)]\;dy\cr}$$ where $N^{+}(p,\partial E,y)$ denotes the number of points of $p^{-1}(y)\cap\partial E$ at which $Jp$ is positive and similarly, $N^{-}(p,\partial E,y)$ denotes the number of points of $p^{-1}(y)\cap\partial E$ at which $Jp$ is negative. Since $R=D_{B_{\Omega}}$, we conclude that $$N^{+}(p,\partial E,y)-N^{-}(p,\partial E,y)=1\eqnlbl{sumdeg}$$ for almost all $y\in D_{B_{\Omega}}$. First, we recall that $\partial E\cap\overline{\Omega}$ is $C^{1}$ at all of its points except for a singular set $S\subset\partial E\cap\Omega$ whose Hausdorff dimension does not exceed $n-8$. Furthermore, we know that $\partial E\cap\Omega$ is real analytic at all points away from $S$ and that $\partial H$ is $C^{1,1}$. Finally, we know that $E$ contains $B_{\Omega}$. Let $(\partial E)^{+}$ and $(\partial H)^{+}$ denote the parts of $\partial E$ and $\partial H$ respectively that lie above the equatorial plane $P$ of $B_{\Omega}$. Let $p\colon\R^{n}\to P$ denote the orthogonal projection. The mean curvature of $\partial E$ is equal to a constant $K$ at all points of $\partial E\cap(\Omega-S)$. Let $X$ denote the vertical unit vector. We wish to apply with $ (\partial E)^{+}$ replacing $ M$. Referring to the proof of Lemma , we see that this can be done in spite of the singular set $ S\in(\partial E)^{+}$. Thus, applying , we obtain $$\int_{(\partial H)^{+}}{\cal H}_{\partial H}X\cdot\nu_{H}\;dH^{n-1} =\int_{(\partial E)^{+}}{\cal H}_{\partial E}X\cdot\nu_{E}\;dH^{n-1} \eqnlbl{equalint}$$ where $\nu_{H}$ and $\nu_{E}$ denote the unit exterior normals to $H$ and $E$ respectively. Let $$\displaylines{ A=(\partial E)^{+}\cap(\partial H)^{+}\cr B=((\partial H)^{+}-A)\cap\{x:{\cal H}_{\partial H}(x)<K\}\cr C=((\partial H)^{+}-A)\cap\{x:{\cal H}_{\partial H}(x)=K\}.\cr}$$ Since $\mch \le K$ $H^{n-1}$-a.e. in $(\partial H)^{+}\cap\Omega$, it suffices to prove that $$H^{n-1}(B)=0.\eqnlbl{toprove}$$ Observe that both $B$ and $C$ are subsets of $\partial H^+$. Note also that $A,B,$ and $C$ are mutually disjoint subsets of $(\partial H)^{+}$ with $H^{n-1}[(\partial H)^{+}-(A\cup B\cup C)]=0$. Thus, $p(A),p(B)$ and $p(C)$ are mutually disjoint and their union occupies almost all of $D_{B_{\Omega}}$. Clearly, $\nu_{E}$ and $\nu_{H}$ as well as $\mch$ and $\mce$ agree $H^{n-1}$ almost everywhere on $A$. Therefore, $$\int_{A}\mch X\cdot\nu_{H}\;dH^{n-1}=\int_{A}\mce X\cdot\nu_{E}\;dH^{n-1}\;.\eqnlbl{ona}$$ Since $X\cdot\nu_{H}$ is the Jacobian of the mapping $p\colon\partial H^{+}\to D_{B_{\Omega}}$, it follows from that $$\displaylines{ \int_{B}\mch X\cdot\nu_{H}\;dH^{n-1}<K H^{n-1}[p(B)],\cr \int_{C}\mch X\cdot\nu_{H}\;dH^{n-1}=K H^{n-1}[p(C)].\cr}$$ Now let $$\eqalign{ A^{}&=((\partial E)^{+})\cap p^{-1}[p(A)],\cr B^{}&=((\partial E)^{+})\cap p^{-1}[p(B)],\cr C^{}&=((\partial E)^{+})\cap p^{-1}[p(C)].\cr}$$ Next, observe that both $B^{}$ and $C^{}$ are subsets of $\Omega$. To see this, consider $x\in B^{}$. If it were true that $x\in B^{}\cap\partial \Omega $, then $x\in(\partial H)^{+}$ and thus $x\in A$. This is impossible since $p(A)$ and $p(B)$ are disjoint. A similar argument holds for $C^{}$. Referring to and , we obtain $$\eqalign{ &\int_{B^{}}\mce X\cdot\nu_{E}\;dH^{n-1}\cr &\quad=K\int_{B^{}\cap\{x:X\cdot\nu_{E}(x)>0\}} X\cdot\nu_{E}\;dH^{n-1} +K\int_{B^{}\cap\{x:X\cdot\nu_{E}(x)<0\}} X\cdot\nu_{E}\;dH^{n-1}\cr &\quad=K\int_{p(B^{})}N^{+}(p,\partial E,y)-N^{-}(p,\partial E,y)\;dH^{n-1}(y)\cr &\quad=K H^{n-1}[p(B^{})]\cr &\quad=K H^{n-1}[p(B)].\cr}$$ Similarly, $$\int_{C^{}}\mce X\cdot\nu_{E}\;dH^{n-1} =K H^{n-1}[p(C^{})]=K H^{n-1}[p(C)]$$ and $$\int_{A^} K X\cdot\nu_E\;dH^{n-1}=K H^{n-1}(p(A)).$$ Finally, in view of the fact that $A\subset(\partial H)^{+}$ and therefore that $N^{+}(p,A,y)=1$ and $N^{-}(p,A,y)=0$ for $H^{n-1}$-almost all $y\in p(A)$, we obtain $$\int_{A} K X\cdot\nu_E\;dH^{n-1}= K H^{n-1}(p(A)).$$ Now, using the facts that $A^{}-A\subset\Omega$ and $\mce=K$ on $A^{}-A-S$, we obtain $$\eqalign{ \int_{A^}&\mce X\cdot\nu_{E}\;dH^{n-1}\cr &\qquad=\int_{A^}K X\cdot\nu_{E}\;dH^{n-1} +\int_{A^}(\mce-K) X\cdot\nu_{E}\;dH^{n-1}\cr &\qquad=\int_{A^}K X\cdot\nu_{E}\;dH^{n-1} +\int_{A}(\mce-K) X\cdot\nu_{E}\;dH^{n-1} \cr &\qquad=K H^{n-1}(p(A))-K H^{n-1}(p(A))+\int_{A}\mce X\cdot\nu_{E}\;dH^{n-1}\cr &\qquad=\int_{A}\mce X\cdot\nu_{E}\;dH^{n-1}.\cr}$$ Under the assumption $H^{n-1}(B)>0$, we would obtain $$\eqalign{ \int_{(\partial H)^{+}}\mch X\cdot\nu_{H}\;dH^{n-1}& <\int_{A}\mch X\cdot\nu_{H}\;dH^{n-1} +K H^{n-1}[p(B)]+K H^{n-1}[p(C)]\cr &=\int_{A}\mce X\cdot\nu_{E}\;dH^{n-1}+K H^{n-1}[p(B^{})]+K H^{n-1}[p(C^{})]\cr &=\int_{A^{}}\mce X\cdot\nu_{E}\;dH^{n-1}+K H^{n-1}[p(B^{})]+K H^{n-1}[p(C^{})]\cr &=\int_{A^{}}\mce X\cdot\nu_{E}\;dH^{n-1}+\int_{B^{}}\mce X\cdot\nu_{E}\;dH^{n-1}\cr &\qquad+\int_{C^{}}\mce X\cdot\nu_{E}\;dH^{n-1}\cr &=\int_{A^{}\cup B^{}\cup C^{}}\mce X\cdot\nu_{E}\;dH^{n-1}\cr &\leq\int_{(\partial E)^{+}}\mce X\cdot\nu_{E}\;dH^{n-1},\cr}$$ where we have used that $A^{},B^{}$ and $C^{}$ are mutually disjoint. This would contradict , thus establishing . A function $u\in C^{1}(W)$ is called a [weak subsolution [(]{}supersolution[)]{} of the equation of constant $K$ mean curvature]{} if $$Mu(\varphi)=\int_{W}\frac{\nabla u\cdot\nabla \varphi}{\sqrt{1+\abs{\nabla u}^{2}}}-K\varphi\,dx\leq0\quad(\geq 0)$$ whenever $\varphi\in C^{1}_{0}(W),\; \varphi \geq 0$. We note that if $u\in C^{1,1}$ and classically satisfies the equation of constant mean curvature equation almost everywhere, then $u$ is a weak solution. The following result will be stated in the context of $R^{n-1}$ because of its applications in the subsequent development. Define $$\eqalign{ u_{t}&=tu_{1}+(1-t)u_{2}\;\hbox{\rm for}\;t\in[0,1],\cr w&=u_{1}-u_{2},\cr a^{ij}(x)&=\int_{0}^{1}D_{u_{x_{j}}}\left(\frac{D_{i}u_{t}(x)} {\sqrt{1+\abs{\nabla u_{t}}^{2}}}\right)\,dt\cr &=\int_{0}^{1}\frac{1}{\sqrt{1+\abs{\nabla u_{t}}^{2}}}\left(\delta_{ij}- \frac{D_{i}u_{t}(x)D_{j}u_{t}(x)}{(1+\abs{\nabla u_{t}}^{2})}\right)\;dt.\cr}$$ Since both $u_{1}$ and $u_{2}$ are continuously differentiable in $W$, for each open set $V\subset\subset W$ containing $x_{0}$ there exists $M>0$ such that $\abs{\nabla u_{t}(x)}\leq M$ for all $x\in V$ and all $t\in[0,1]$. Hence, $$\eqalign{ a^{ij}(x)\xi_{i}\xi _{j}&\geq\frac{1}{(1+M^{2})^{1/2}}\abs{\xi }^{2},\;\hbox{\rm for all}\;\xi \in R^{n-1},x\in V,\cr \sum_{i,j}a^{ij}(x)^{2}&\leq C,\;\hbox{\rm for all}\;x\in V.\cr }$$ For $\varphi\in C^{1}_{0}(W),\>\varphi\geq0$, we have $$\eqalign{ 0&\leq Mu_{1}(\varphi)-Mu_{2}(\varphi)\cr &=\int_{W}\!\!\int_{0}^{1}\frac{d}{dt}\left(\frac{\nabla u_{t}(x)\cdot\nabla \varphi(x)}{\sqrt{1+\abs{\nabla u_{t}}^{2}}}\right)\,dt\,dx\cr &=\int_{W}a^{ij}(x)D_{j}w(x)D_{i}\varphi(x)\,dx.\cr}$$ Thus, $w$ is a weak supersolution of the equation $$D_{i}(a^{ij}D_{j}w)=0$$ and since $w\geq0$, the weak Harnack inequality \[GT, Theorem 8.18\] yields $$\left( r^{-n}\int_{B(x_{0},2r)}\abs{w(x)}^{p}\,dx\right)^{1/p}\leq C\inf_{B(x_{0},r)}w=0$$ whenever $1\leq n<n/(n-2)$ and $B(x_{0},4r)\subset W$. [ Theorem.]{} [Suppose $\Omega$ is a bounded, strictly convex domain with $C^{2}$ boundary that satisfies a great circle condition. Then any minimizer $E$ of the variational problem is convex.]{} Later we show that neither smoothness of $\partial\Omega$ nor strict convexity are required. In addition, the great circle condition is unnecessary in $\R^{2}$. The same applies to the uniqueness result below. [Proof.]{} It suffices to show that $ H=E$ where $ H$ denotes the convex hull of $ E$. Assume $\partial H\not\subset\partial E$ so there exists $x\in \partial H\setminus\partial E$. Thus, as in the proof of the mean curvature inequality in Theorem , we see that $x$ lies in the convex hull $M$ of distinct points $p_i\in \partial H\cap\partial E$, $i=1,\dots,k,\, k>1$. Futhermore each $p_{i}$ is an element of $\Omega $ due to the fact that they all lie in a single support plane of $H$; hence if one $p_i$ where to lie in $\partial \Omega$ then they all would, thus contradicting strict convexity. Referring to Lemma , we see that $ \partial H$ and $ \partial E$ agree in a neighborhood of the points $p_i$. Since $ M$ is connected, it follows again from Lemma that $ M\subset\partial E\cap\partial H$, which contradicts $ x\not\in\partial E$. Consequently $\partial H\subset\partial E$ and thus $P(H)\le P(E)$. However $E\subset H$ so $\|E\|\le\|H\|$. Assume $\|E\|<\|H\|$. Dilate $H$ to obtain $\tilde H\subset \Omega$ satisfying $\|\tilde H\|=\|E\|$. But then $P(\tilde H)<P(H)\le P(E)$ which contradicts the minimality of $E$. Thus $\|E\|=\|H\|$ so that $E$ and $H$ have the same measure theoretic closure. Hence, due to our convention concerning distinguished representatives for sets of finite perimeter, $E=H$ and $E$ is convex. Note that the assumption of convexity can be relaxed. It is only required that the intersection of $\Omega$ with any vertical line is an interval. In addition $\partial\Omega$ must not contain vertical line segments. To prove we argue by contradiction. If $E$ and $F$ are perimeter minimizers satisfying $\|B_\Omega\|\le\|F\|<\|E\|$ assume $F$ is not a subset of $E$. From Theorems and we see that $E$ and $F$ are convex and contain $B_\Omega$. Since $F$ is not a subset of $E$ one can employ the proof of , with $F$ playing the role of $B_\Omega$, to prove that there is a second perimeter minimizer $E^{}$ which contains $F$ and satisfies $\|E^{}\| = \|E\|$. Let $H$ be the analog of $H$ in the proof of Theorem and let $D^{\circ}$ denote the interior of $D:=D_{B_{\Omega }}$. We will use the properties of perimeter minimizers to show that $\partial H$ and $\partial(H\cup E)$ are analytic and coincide on some open set. By connectedness, this will show they are identical, thus establishing the desired contradiction. Let $O$ be the interior of $\partial H\setminus (H\cup E)^\circ$ relative to $\partial H$, and $\partial O$ represent the boundary of $O$ relative to $\partial H$. Assume there exists a point $$x\in\partial O\cap p^{-1}(D^\circ).$$ Note that $x\in \partial H\cap \partial E\cap \partial (H\cup E)\cap p^{-1}(D^\circ)$. Let $y$ be the point on $\partial F\cap \partial E^{}$ which was translated (as in the definition of $H$) to $x$. Since $\partial O$ has positive $H^{n-2}$ measure ($\partial O\cap p^{-1}(D^\circ)\ne\nullset$) we can assume $y\notin S,\,S$ being the singular set for $E^{}$. Since $x\in\partial E\subset\overline\Omega,\, y$ lies in $\Omega$ and consequently $\partial H$ is analytic in a neighborhood of $x$ since $\partial F$ is analytic in a neighborhood of $y$. Similarly $H\cup E$ inherits analyticity (in a neighborhood of $x$) from $\partial E^{}$ since $x\in \partial(H\cup E)$ and $y\notin S$. However $\partial H\cap O\subset \partial (H\cup E)$ so $\partial H$ and $\partial (H\cup E)$ coincide on open (relative to $\partial H$) subsets of any neighborhood of $x$ so by analyticity $\partial H$ coincides with $\partial (H\cup E)$ in some neighborhood of $x$. But this contradicts $x\in\partial O$ so $\partial O\cap p^{-1}(D^\circ)$ is empty. Note that $(\partial H\setminus E)\cap p^{-1}(D^\circ)$ contains points lying both above and below $D$ since $\Omega$ is strictly convex and $H$ is the hull of the translated halves of $F$ (which contain the hemispheres of the largest ball $B_\Omega$). Thus the same is true of $O\cap p^{-1}(D^\circ)$. Combined with $\partial O\cap p^{-1}(D^\circ)=\nullset$, this implies $\partial H\cap p^{-1}(D^\circ)\cap E^\circ=\nullset$, i.e. $E\subset H$. Of course this is absurd since $\|E\|>\|F\|=\|H\|$. Thus the assumption that $F$ is not contained in $E$ is false i.e. $F\subset E$ as required. Now assume that $\|B_\Omega\|<\|F\| = \|E\|=v$. Choosing a sequence of perimeter minimizers $F_i$ of measure $v_i\uparrow v$, it follows from that $F_i\subset E\cap F$. Consequently $\|E\cap F\|=v$ and so $E=F$. To prove that minimizers are strictly nested in the sense defined above assume that $\|B_\Omega\|\le\|F\|<\|E\|$ and so $F\subset E$. Assume in addition that $G:=(\partial F\cap\partial E\cap\Omega)^\circ$ is not empty. Since $F,\, E$ are analytic in $\Omega$ and nested, it is clear that $\mcf\ge\mce$ at points in $G$. Given that $\mcf,\,\mce$ are constants, say $k_f,\, k_e$, in $\Omega$ and equal almost everywhere on $\partial F\cap\partial E\cap\partial\Omega$, we may derive a contradiction from $k_f\ge k_e$ through the use of . In fact, we obtain $$\int_D \mcf'\;dH^{n-1}=H^{n-2}(\partial D)= \int_D \mce'\;dH^{n-1} \eqnlbl{sameint}$$ where $\mcf'(x):=\mcf(p^{-1}(x)\cap\partial F)$ and $\mce'(x):=\mce(p^{-1}(x)\cap\partial E)$. However with $A:=\partial F\cap\partial \Omega$ and $B:=\partial F\cap\Omega$, we see that $$\eqalign{ \int_D \mce'\;dH^{n-1} &=\int_{p(A)} \mcf'\;dH^{n-1}+\int_{p(B)}\mce'\;dH^{n-1}\cr &\le \int_{p(A)}\mcf'\;dH^{n-1}+\int_{p(B)}\mcf'\;dH^{n-1}\cr &=\int_D\mcf'\;dH^{n-1},\cr} \eqnlbl{equality}$$ and thus we have equality due to . Therefore, $\mce'=k_e=k_f=\mcf'$ on $p(B)$. However, since $\mce'=\mcf'$ almost everywhere on $p(A)$, we obtain $$\mce'=k_e=k_f=\mcf' \qquad \hbox{$H^{n-1}$-almost everywhere on} D.$$ Thus for $x\in \partial F\cap\partial E$, apply Lemma to conclude that $\partial F$ and $\partial E$ coincide in a neighborhood of $x$. Thus $p(\partial F\cap\partial E)$ is both open and closed relative to $D^\circ$, and therefore contains $D^\circ$, a contradiction since $\abs{F}<\abs{E}$. We now dispense with the assumptions of strict convexity and smoothness of $\partial\Omega$. When the assumption of strict convexity is dropped, complications arise because there is no longer a unique largest ball in $\Omega$. Eliminating the smoothness assumption on the boundary forces us to take limits of perimeter minimizers, and to establish convexity of all perimeter minimizers through a uniqueness theorem. One interesting observation is that a perimeter minimizer can be thought of as a smooth approximation of $\Omega$, especially when its measure is close to that of $\Omega$. This is due to the fact that even after we have dispensed with the smoothness assumption on $\partial \Omega$ perimeter minimizers still have $C^{1,1}$ boundaries. For the proof of Theorem below, we need the following lemma. A given open interval containing $c$ in turn contains an open interval $ I=(a',b')$ with $ c\in I$ determined by the constants $ c_{1},c_{2}$ and $ c_{3}$ such that the following three conditions hold: There are points $ x_{1},x_{2}\in I$ with $x_{1}<c<x_{2}$ such that $ f_{1}(x_{1})=f_{2}(x_{2})$. There are polynomials $ p_{i}$ of degree $ 2$ ($i$=1,2) such that $p_{i}(x_{i})=f_{i}(x_{i}) $ and such that the functions $$h_{1}(x):=\cases{ f_{1}(x)& for $ a\leq x\leq x_{1}$\cr p_{1}(x)& for $ x_{1}\leq x\leq c$\cr} \qquad h_{2}(x):= \cases{ f_{2}(x)& for $ x_{2}\leq x\leq b$\cr p_{2}(x)& for $ c\leq x\leq x_{2}$\cr}$$ are $ C^{2}$ and strictly concave on $ I_{i}$. There is a point $c'\in I $ such that $ h_{1}(c')=h_{2}(c')$. Thus, the function $$h:=\cases{ h_{1}& on $ [a,c']$\cr h_{2}& on $ [c',b]$\cr}$$ is strictly concave on $ [a,b]$. We now will mollify $ h$ restricted to $ I$ by using a smooth mollifying kernel $ \varphi $ with the property that $$\varphi _{\varepsilon }\ast p(x)=p(x)$$ whenever $ p$ is a polynomial of degree 2, $\varepsilon >0$, and $ x\in\R$, cf. \[Z, Lemma 3.5.6\]. Thus, for sufficiently small $ \varepsilon >0,\;\varphi _{\varepsilon }\ast h(x)=h(x)$, for $ x\in (a'+\varepsilon ,c'-\varepsilon ) \cup(c'+\varepsilon,b'-\varepsilon )$. Also, $ \varphi _{\varepsilon }\ast h$ is strictly concave since $ h$ is. Thus, our desired function $ g$ is defined by $$g(x) =\cases{ h(x)& for $ a\leq x\leq a'+\varepsilon $\cr \varphi _{\varepsilon }\ast h(x)& for $ a'+\varepsilon < x <c'-\varepsilon $\cr h(x)& for $ c'-\varepsilon \leq x\leq b$.\cr}$$ We define $H_\Omega$ to be the union of all largest balls in $\Omega$. Thus $H_\Omega$ is the convex hull of the two largest balls which are furthest apart. $H_\Omega$ essentially plays the role of $B_{\Omega }$. We first smooth $\Omega$ and then establish the existence of a nested family of convex perimeter minimizers by taking limits. We finish by adapting the uniqueness result of Theorem and the proof of disjointness of boundaries. Let $T_{B_\Omega}$ be the hyperplane which intersects orthogonally the midpoint of the line segment joining the centers of the two largest balls whose hull forms $H_\Omega$. Think of the “vertical” axis as coinciding with this line segment and take the origin of our coordinate system to be the midpoint just mentioned. As defined previously $p$ is orthogonal projection onto $T_{B_\Omega}$. Let $B_\Omega$ be the largest ball in $\Omega$ with equatorial plane in $T_{B_\Omega}$. Let $D_{B_\Omega}=p(\Omega)$ so $D_{B_\Omega}$ is an $(n-1)$-ball. Let $C$ be the interior of the union of a closed right circular cone with base $D_{B_\Omega}$ with its reflection across $T_{B_\Omega}$. Let $B$ be the largest ball in $C$ and note that $C\setminus B$ has three components (four in $\R^2$). Let $C_0$ denote the component (or union of two components in $\R^2$) which intersects $D_{B_\Omega}$ and consider the set $C_1=C\setminus \overline{C_0}$. First we show that $\Omega$ can be approximated arbitrarily closely by strictly convex sets satisfying a great circle condition, then we will approximate the later by sets with $ C^{2}$ boundary of the same type. Note that $\Omega\cap C_1$ is convex and satisfies a great circle condition with $B$ being the largest ball. Also $\partial(\Omega\cap C_1)$ consists of the union of the graphs of functions $f_i,\,i=1,2,\, f_1\ge0,\,f_2\le 0$. Let $\Omega'$ be the set whose boundary is the union of the graphs of $f_1+\varepsilon b,\, f_2-\varepsilon b$ where $\varepsilon>0$ and $b$ is the function whose graph is the upper hemishere of $B$. Note that $\Omega'$ is strictly convex and satisfies a great circle condition. Also, as $\varepsilon\to 0$, $C$ approaches a cylinder, and $\Omega'\to\Omega$ in the Hausdorff sense. We now may assume with out loss of generality that $\Omega$ is strictly convex. Consider $G=\Omega \cap C$. Note that $\partial G$ is the union of graphs of $f_i:\overline D_{B_\Omega}\rightarrow \R,\, i=1,2$ with $f_1\ge0,\, f_2\le 0$. Given $r>0$ let $B_r$ be the ball of radius $r$ concentric to $B_\Omega$, $D_r=D_{B_\Omega}\cap B_r$,and $R$ the radius of $B_\Omega$. Also let $\bar r$ be the distance from $\partial B_\Omega\cap\partial C$ to the vertical axis. Consider $\varepsilon, \, 0<\varepsilon<<R$. For a smooth radially symmetric approximate identity $\eta_\varepsilon$ supported in $B_\varepsilon$ let $f_\varepsilon=f_1\eta_\varepsilon$. Thus $f_\varepsilon$ is defined in $D_{R-\varepsilon}$ and is a surface of revolution in $A_\varepsilon=D_{R-\varepsilon}\setminus D_{\bar r+\varepsilon}$ Now consider $\delta>0$ such that $\bar r<R-\delta$ but $\partial B_{R-\delta}$ does not intersect $\partial C$. Take $\varepsilon$ small enough that the graph of $f_\varepsilon$ does not intersect $\partial B_{R-\delta}$. Let $g_\varepsilon:[\bar r+\varepsilon, R-\varepsilon]\rightarrow \R$ be the function the rotation of whose graph around the vertical axis produces the graph of $f_\varepsilon$ over $A_\varepsilon$. In the $r,\,z$ plane let $C_2$ be a circle of radius $s>>R$ with center on the negative $r$ axis which passes through $(R-\delta,0)$. Let $c:[\bar r+\varepsilon, R-\varepsilon]\rightarrow \R$ be the function whose graph lies in the upper half of $C_2$ and define $h_\varepsilon=\min(g_\varepsilon,\,c)$ on $[\bar r+\varepsilon,\,R-\delta]$. Note $h_\varepsilon$ is a strictly concave function and is smooth except at the point $q$ of intersection of the graphs of $g_\varepsilon$ and $c$ (which exists if $s$ is large enough). Now employ Lemma to alter $h_{\varepsilon }$ in a small neighborhood of $q$ to produce a $C^2$ function which is still strictly concave. Consider the surface obtained by taking the union of the surface of revolution formed by rotating the graph of the smoothed $h_\varepsilon$ with the graph of $f_\varepsilon$ over $D_{\bar r+\varepsilon}$. This is a $C^2$ surface and when combined with a similarly constructed surface for $f_2$ produces the boundary of a strictly convex set $\Omega_\varepsilon$. Note that $\partial \Omega_\varepsilon$ is $C^2$ and that $\Omega_\varepsilon$ satisfies a great circle condition with $B_{R-\delta}$ being the largest ball. Also as $C$ approaches a cylinder and $\delta, \varepsilon\to 0$ we have $\Omega_\varepsilon\to\Omega$ in the Hausdorff sense as required. To make the process of taking limits easier in the following we can dilate the sets $\Omega_\varepsilon$ a small amount so they contain $\Omega$. Thus there exists a sequence of $C^2$ strictly convex sets $\Omega_n$ which contain $\Omega$, satisfy a great circle condition, and which converge to $\Omega$ in the Hausdorff sense. For $v,\,\|B_\Omega\|< v\le\|\Omega\|$ (and $n$ large enough so $\|B_{\Omega_n}\|<v$) let $E_n(v)$ be the unique perimeter minimizer in $\Omega_n$ of measure $v$. It is easy to see that for a dense set of $v_i$’s we can, by repeatedly extracting subsequences and diagonalizing, construct a subsequence of $E_n$ such that for all $i$, $E_n(v_i)$ converges (on the subsequence) to $E(v_i)$, a subset of $\Omega$, in the Hausdorff sense. Nestedness and convexity are clearly inherited. Thus taking intersections of appropriate $E(v_i)$ we extend the definition of $E(v)$ to all $v,\,\|B_\Omega\|< v<\|\Omega\|$. Nestedness allows us to extend convergence to all such $v$. We claim that the sets $E(v)$ are perimeter minimizers relative to $\Omega$. To see this note that given any set $F\subset\Omega$ with $\|F\|=v$ we have $F\subset\Omega_n$ since $\Omega\subset\Omega_n$; consequently by lower semicontinuity of perimeter we have $$P(E(v))\le \liminf P(E_{n}(v)) \le P(F)$$ (with the liminf taken over the subsequence) i.e. $E(v)$ is a perimeter minimizer. For $v,\,0\le v\le\|H_\Omega\|$ we can characterize perimeter minimizers. Assume $E$ is a perimeter minimizer of measure $v$. If $0<v\le\|B_\Omega\|$ then $E$ is clearly a ball. If $\|B_\Omega\|< v\le\|H_\Omega\|$ we claim that $E$ is the convex hull of two largest balls in $\overline\Omega$. In proving this we will also prove for $v \ge \|H_\Omega\|$ that any perimeter minimizer $E$ satisfies $H_\Omega\subset E$. Assume $\|B_\Omega\|< v$. Consider the following extension of the proof of Theorem . As it stands the proof of Theorem implies that $E$ contains a largest (in $\overline\Omega$) ball. In fact one can conclude much more. Let $B_1,\,B_2$ be the closed balls whose convex hull is $H_\Omega$, let $\ell$ be the line through their centers, and consider any set $H$ which is the convex hull of two translates of $B_1$ with centers on $\ell$ such that $\|H\cap E\|=\|B_1\|$ and $H\cap H_\Omega$ contains a translate of $B_1$. A mild variation in the proof of Theorem shows that $H\cap E$ is a translate of $B_1$. We claim that this implies that $E\cap H_\Omega$ is the convex hull of two translates of $B_1$. To see this let $B_3,\,B_4\subset E$ be distinct translates of $B_1$ with $x$ being the midpoint between their centers. Since the hull of $B_3,\,B_4$ has measure larger than $\|B_1\|$ construct $H$ as above using translates of $B_1$ placed symmetrically with respect to $x$. However $H\cap E$ is a translate of $B_1$. Thus there is a translate of $B_1$ contained in $E$ lying strictly between ant two such balls. Therfore the centers of such balls form an interval in $\ell$ Now take $\ell$ to be the vertical axis with $B_1$ lying above $B_2$, let $B_u,\,B_l$ be the the uppermost and lowest translates of $B_1$ in $E$, and $E_u,\,E_l$ the parts of $E$ strictly above and below $B_u,\,B_l$ respectively. Assume $E_u$ is not empty so $\|E_u\|\ne 0$. If $B_u\ne B_1$ construct $H$ as above by translating hemispheres of $B_1$ so that $H$ contains subsets of positive measure from both $E\cap H_\Omega$ and $E_u$. However this is a contradiction since by the above $E\cap H$ is a translate of $B_1$ which cannot possibly intersect $E_u$. Thus $E_u$ not empty implies $B_u=B_1$. Similarly $E_l$ not empty implies $B_l=B_2$. This establishes the claim. Moreover one can conclude that $v\ge\|H_\Omega\|$ implies $H_\Omega\subset E$. To see this note if $v\ge\|H_\Omega\|$ then at least one of $E_u,\,E_l$ is nonempty. If both are nonempty then $H_\Omega\subset E$ as claimed. If only one is nonempty, say $E_u$, then translate $E$ as far down as possible while remaining in $\overline\Omega$ to form a set $E^$ which contains $B_2$ ($E_l$ is empty). Note that $E^$ is also a perimeter minimizer of measure $v$. Thus $E_u^$ nonempty, i.e. $E^=E$ with $H_\Omega\subset E$ as required. Now that we have characterized perimeter minimizers for $v,\,0<v\le \|H_\Omega\|$ we can redefine $E(v)$ so that $E(v)$ is the convex hull of two translates of $B_1$, symmetrically placed in $H_\Omega$, if $\|B_1\|<v\le \|H_\Omega\|$, and $E(v)$ is a symmetrically placed ball if $0<v\le \|B_1\|$. Thus we have a nested collection of convex perimeter minimizers which can be used to establish uniqueness. Given $\bar v$, $\|H_\Omega\|<\bar v$, assume that $E$ is a perimeter minimizer with measure $\bar v$. Recall from above that $H_\Omega\subset E$. Before proceeding we define an auxiliary collection $\{H(v):\, \|H_\Omega\|\le v\le \bar v \}$, $H(v)$ defined analogously to $H$ in Theorem by translating the halves of $E(v)$ the least possible amount such that the resultant hull $H(v)$ satisfies $\|H(v)\cap E\|=v$. Note that the sets $H(v)$ are nested since if $\|H_\Omega\|\le v<w$ and one translates the halves of $E(w)$ the same distance as for $E(v)$ in the definition of $H(v)$, and calls the hull of the translated halves $\tilde H$ then $\|\tilde H\setminus H(v)\|=w-v$ so $\|\tilde H\cap E\|=\|(\tilde H\setminus H(v))\cap E\| +\|H(v)\cap E\|\le (w-v)+v=w$ i.e. $H(v) \subset \tilde H\subset H(w)$ as required. Let $v_0=\sup\{v:E(v)\subset E\}$. If $v_0=\|E\|$ then $E=E(v_0)$, otherwise $v_0<\|E\|$ so $E^\circ\setminus E(v_0)$ is not empty. Let $B$ be a closed ball of positive radius in $E^\circ\setminus E(v_0)$, $v_1=\sup\{v:H(v)\cap B \hbox{ is empty }\}$, and $v_2=\inf\{v:B\subset H(v)\}$. Clearly $v_2=\|H(v_2)\cap E\|\ge \|H(v_1)\cap E\|+\|B\|=v_1+\|B\|$ so choosing $v,\,v_1<v<v_2$ we see that $B$ contains points in $H(v)$ and its complement. Consequently $\partial H(v)$ intersects $B$. One can now proceed as in the proof of Theorem with $H$ replaced by $H(v)$ with the following modifications. In proving that $(\partial H(v)\setminus E)\cap p^{-1}(D_{B_\Omega}^\circ)$ is not empty one uses the fact proved above that $H_\Omega\subset F$ so that $H(v)$ contains a convex hull of “largest balls” which is larger than $H_\Omega$ and thus must intersect the complement of $\Omega$. Finally we see that the conclusion $\partial H(v)\cap p^{-1}(D_{B_\Omega}^\circ)\cap E^\circ=\nullset$ is absurd due to our construction in which $\partial H(v)$ intersects $E^\circ$. Thus the assumption that $v_0<\|E\|$ must be false and consequently $E=E(\|E\|)$ as required. It remains only to prove the disjointness result. The proof is identical to that in Theorem once we have established that minimizers have $C^{1,1}$ boundaries and satisfy the same mean curvature properties as before. Assume $\|H_\Omega\|\le v<\|\Omega\|$. Let $E_n(v)$ be as above and note that $k_n$, the constant mean curvature associated with $\partial E_n(v)$, is bounded uniformly in $n$ since as in the proof of Theorem we have $$H^{n-1}(\partial D_{B_\Omega})=\int_{D_{B_\Omega}} \mcen' \ge k_n H^{n-1}(p(\partial E_n(v)\cap\Omega^\circ))$$ where $H^{n-1}(p(\partial E_n(v)\cap\Omega^\circ))$ is uniformly bounded from zero (on a subsequence) for geometrical reasons since $E_n(v)$ is convex, contains $B_\Omega$, and converges (on a subsequence) to $E(v)$. Consequently $0\le \mcen \le k_n\le M$ almost everywhere and we see that $\partial E_n(v)$ is uniformly $C^{1,1}$ from which we see that $E(v)$ is $C^{1,1}$ as well. Note that tangent planes converge almost everywhere so that locally first derivatives converge almost everywhere and consequently one can take limits in the weak definition of mean curvature to show that if $k_n\rightarrow k$ (on a subsequence) then $\partial E(v)$ has mean curvature $k$ in the interior of $\Omega$, and that $\mce\le k$ as required. [ Theorem.]{} [If $n=2$, and $\Omega$ is as in Theorem , except that the great circle condition is not assumed, then the results of Theorem still hold. Furthermore, if $\|H_\Omega\|\le\|E\|<\|\Omega\|$, then a perimeter minimizer $E$ is the union of all balls in $\overline \Omega$ of curvature equal to the curvature of $\partial E\cap\Omega$, if $\|B_\Omega\|<\|E\|<\|H_\Omega\|$, then $E$ is the union of two largest balls in $\Omega$, if $0<\|E\|\le \|B_\Omega\| $, then $E$ is a ball.]{} Smooth $\partial \Omega$ as before but without requiring the great circle condition. The same regularity properties hold as before for perimeter minimizers $E_n$ in the smoothed domains $\Omega_n$. Note that there is no singular set since $n=2$. Also $\partial E_n\cap \Omega$ consists of circular arcs. Thus if $x\in \Omega$ is a limit point of points $x_n\in\partial E_n$ then it is easy to see geometrically that the curvatures of the circular arcs in $\partial E_n\cap \Omega$ must be uniformly bounded in $n$. Regularity and curvature results for the limiting perimeter minimizer follows as before. We claim that any perimeter minimizer $E$ must be convex. First note that $E$ cannot have an infinite number of components since otherwise $\partial E$ would contain a limit point of points in the boundaries of distinct components of $E$ which would violate the regularity of $\partial E$. In addition each component must be simply connected since otherwise one could add a bounded component of the complement of $E$ to $E$ which would reduce the perimeter of $E$ and increase its measure. Thus a scaling argument as in the proof of Theorem would violate the fact that $E$ is a perimeter minimizer. Also each component must be convex. To see this note that locally $\partial E$ is a graph of a $C^{1,1}$ function $f$. Thus $f'$ is Lipschitz continuous, monotone increasing (if axes are chosen properly) on $f^{-1}(\partial E\cap\partial\Omega)$, and monotone increasing on each component of $f^{-1}(\partial E\cap \Omega)$ from which the claim easily follows. Finally given two components considering the two unique lines which are support lines for both components one sees that one of the components can be translated with out leaving $\Omega$ until it first touches another component. This translation does not change the measure of the overall set and does not increase perimeter so a new perimeter minimizer is created. Due to the regularity of $\partial E$ the point of contact lies in $\Omega$. However this contradicts the fact that the boundary of a perimeter minimizer must be a circular arc locally in $\Omega$. Consequently there must be only one component which we have already shown to be convex so $E$ is convex as claimed. To establish the uniqueness and nestedness properties it is sufficient to characterize perimeter minimizers. In fact we claim that if $E$ is a perimeter minimizer with $\|H_\Omega\|\le\|E\|<\|\Omega\|$ then it is the union of all balls in $\overline\Omega$ of curvature given by the curvature of $\partial E\cap\Omega$. We prove the claim in two parts. We first establish that if a point $x$ lies in $E$ then $x$ lies in a ball contained in $\overline\Omega$ whose boundary has the same curvature as $\partial E\cap\Omega$ . We finish by proving that if $\|H_\Omega\|\le\|E\|<\|\Omega\|$ then $E$ contains all balls with the same curvature as $\partial E\cap\Omega$. Assume $x\in E$ and let $d=\hbox{dist}(x,\,\partial E),\, r={1\over k}$ where $k$ is the curvature of $\partial E\cap\Omega$. If $d\ge r$ then $x$ is clearly in a ball of radius $r$ contained in $\overline\Omega$ as claimed. If $d<r$ then choose a point $y\in\partial E$ closest to $x$. Choose axes so that $y$ is the origin, $x$ lies on the positive horizontal axis, and the vertical axis is tangent to $\partial E$ at $y$. Let $C$ be the upper half of the circle of radius $r$ containing $y$ with center on the positive horizontal axis. Let $(0,\,a)$ be the largest subinterval of $(0,\,2r)$ over which the part of $\partial E$ lying above the horizontal axis is a graph. Let $f:(0,a)\rightarrow \R$ be the function having such a graph. Let $g:[0,\,2r]:\rightarrow\R$ be the function with $C$ as its graph. Integrating the divergence form for curvature over $(\varepsilon,\,t)$ for $t<a,\,\varepsilon>0$ then letting $\varepsilon\rightarrow 0$ one obtains $$1-J(f'(t))=-\int^t_0 [J(f'(s))]'\,ds \le \int^t_0 k\,ds = -\int^t_0 [J(g'(s))]'\,ds =1-J(g'(t))$$ where $J(x)=x/(\sqrt{1+x^2})$ since $f'(0)=g'(0)=\infty$ (recall $\partial E$ is $C^{1,1}$). Since $J(x)$ is monotone increasing this implies that $g'(t)\le f'(t)$ on $(0,\,a)$. However $0=g(0)\le \lim_{s\to 0}f(s)$ so $g(t)\le f(t)$ on $(0,\,a)$. From the estimate on $f'$ and the convexity of $E$ we see that $a=2r$. A similar argument shows that the part of $\partial E$ lying below the horizontal axis in fact lies below the other half of the circle of radius $r$ mentioned above. Thus from the convexity of $E$ we see that this circle lies in $E$ as claimed. Consequently $E$ lies in the union of all balls of radius $r$ which lie in $\overline\Omega$. To prove our second claim let $B$ be a ball of radius $r$ contained in $E$ (such a ball exists by the above argument). Let $D$ be any other ball of radius $r$ contained in $\Omega$ and let $H$ be the convex hull of $B,\, D$. Assume that $D$ is not a subset of $E$ so there exists $x\in D\setminus E$. Thus $\partial E$ separates $x$ from $B$. However $H^\circ$, the interior of $H$, lies in $\Omega$ so $\partial E\cap H^\circ$ is locally a circular arc of radius $r$. The only way a circular arc of radius $r$ can separate $x$ from $B$ is if it is a half circle $C$ tangent at its end points to the line segments in $\partial H$. In such a case $\partial E$ must contain $C$ and the (possibly empty) line segments in $\partial H$ with endpoints in $C$ and $\partial B$. Since $x\notin E$ one can translate $E$ towards $x$ while remaining in $\Omega$ due to the geometrical relationship between $E$ and $H$. The translated set is thus still a perimeter minimizer with end opposite $D$ lying in $\Omega$. Thus the end opposite $D$ is a circular arc and $E$ is the convex hull of two (possibly identical) balls of radius $r$. If $\|E\|\le \|B_\Omega\|$ then clearly $E$ is a ball. If $\|B_\Omega\|<\|E\|\le\|H_\Omega\|$ then $\Omega$ satisfies a great circle condition since the line segments in $\partial H$ must lie in $\partial\Omega$. Thus we can use the characterization of $E$ in Theorem as the convex hull of two largest balls in $\overline\Omega$. If $\|H_\Omega\|<\|E\|$ then $E$ cannot be a ball or a hull of two balls in $\Omega$ as concluded in the last paragraph. Consequently the assumption that $D$ was not a subset of $E$ is false and $D\subset E$. Since $D$ was an arbitrary ball of radius $r$ we see that the union of all such balls lies in $E$. Combining this with our earlier conclusion we see that $E$ is in fact equal to the union of all such balls. The disjointness property for boundaries of perimeter minimizers follows from nestedness of minimizers and the fact that the curvature of the boundary of a perimeter minimizer in $\overline\Omega$ strictly increases as a function of the measure $v$ of the minimizer if $\|H_\Omega\|\le v$, a fact which follows directly from the characterization of minimizers. If $E,\, F$ are minimizers with $E\subset F$, and $\partial E\cap\partial F \cap\Omega$ is not empty then geometrically the curvature of $\partial F\cap\Omega$ can not be larger than the curvature of $\partial E\cap\Omega$. However this contradicts the monotonicity of curvature as a function of measure mentioned above. =4 =0 =0 [4. Eqimeasurable Convex Rearrangement]{} Various standard symmetrizations have the useful property of rearranging functions in an equimeasurable fashion while reducing various norms such as $\\|u\\|_{L^p(\Omega)}+\\|\nabla u\\|_{L^p(\Omega)}$ and $\\|u\\|_{BV(\Omega)}$ (see ). However they alter $\Omega$, the domain of definition of $u$, unless $\Omega$ has appropriate symmetries. This is unfortunate from the point of view of studying minimizers to certain variational problems. Using results of Section 2 we introduce an equimeasurable rearrangement which preserves convex domains, reduces $\\|u\\|_{BV(\Omega)}$, and creates level sets which are boundaries of convex sets, when $u\in BV(\R^n)$ with $u\ge0$ and $u=0$ in $\R\setminus\Omega$. Results of \[LS\] imply that such a rearrangement cannot exist for the norm $\\|u\\|_{L^p(\Omega)}+\\|\nabla u\\|_{L^p(\Omega)}$, $p>1$. Any equimeasurable rearrangement clearly fixes the first term in the BV norm . From the co-area formula we will see that a rearrangement which minimizes the perimeter of sets $\{u>t\}$ will minimize the BV norm over an appropriate class of equimeasurable functions . In minimizing functionals such as $$\\|u\\|_{BV(\Omega)} +\int_\Omega F(u) + \int_0^{\|\Omega\|} G(u^,{u^}')\eqnlbl{norm2}$$ over appropriate function classes , where $u^$ is the decreasing rearrangement of $u$, $u^(v)=\sup\{t\,:\,\|\{u>t\}\|\ge v\}$, it is sometimes straight forward to derive regularity estimates for $u^$. Assuming continuity of $u^$ the results of Theorem imply continuity for minimizers of in $\Omega\setminus H_\Omega$, using the continuity and uniqueness properties of $\tilde u$. Of course to apply Theorem it is necessary that $u=0$ on $\Omega$ is a boundary condition for the variational problem and that one can establish $u\ge 0$ in $\Omega$ for minimizers for instance by using a truncation argument. Behaviour in $H_\Omega$ is also highly constrained by the characterization of level sets up to translation. It is fairly straight forward but more delicate to prove partial regularity results for $\nabla u$ if $\Omega\subset\R^2$ by analyzing interactions between boundaries of perimeter minimizers and $\partial \Omega$. However in higher dimensions this is a difficult open problem. Assume that $\Omega$ is a bounded convex set in $\R^n$. In addition assume that $n=2$, or $\Omega$ satisfies a great circle condition. Thus from Section 2 we have a family of convex nested perimeter minimizers $E(v)$ defined as follows. If $B_\Omega$ is a largest ball in $\Omega$ and $H_\Omega$ is the union of all such balls then if $0<v\le\|B_\Omega\|$ let $E(v)$ be a ball of measure $v$ centered symmetrically in $H_\Omega$, if $\|B_\Omega\|<v\le\|H_\Omega\|$ (in which case $\Omega$ satisfies a great circle condition) then let $E(v)$ be the convex hull of two largest balls symmetrically centered in $H_\Omega$ and of measure $v$, finally if $\|H_\Omega\|<v<\|\Omega\|$ let $E(v)$ be the unique perimeter minimizer of measure $v$ shown to exist in Section 2. Define $$BV^+_0(\Omega)= \{u\in BV(\R^n)\,:\,u\ge0,\, u=0 \hbox{ in } \R^n\setminus\Omega\}$$ and define the convex rearrangement of a function $u\in BV^+_0(\Omega)$ by $$\tilde u(x) = \inf\{s\ge 0\,:\, x\notin E(\|\{u>s\}\|)\}.$$ From the remark after Theorem one sees that it is possible to create a rearrangement even if the convexity assumption is relaxed. However it is unclear that one can in this context establish qualitative information analogous to convexity of $\{\tilde u>t\}$. [Proof.]{} Semicontinuity and continuity results are clear from the definition of $\tilde u$ and the disjointness results on boundaries of perimeter minimizers in $\Omega$. It is also clear that $\tilde u\in BV^+_0(\Omega)$. Due to the convexity and nestedness (which is strict in $\Omega$) of the sets $E(v)$ we see that $$E^\circ(\|\{u>t\}\|)\subset \{\tilde u>t\}\subset E(\|\{u>t\}\|)$$ thus $$\|\{u>t\}\|=\|E(\|\{u>t\}\|)\|= \|\{\tilde u>t\}\|.$$ and $$P(\{\tilde u>t\})=P(E(\|\{u>t\}\|))\le P(\{u>t\}).$$ The result on BV norms then follows from the co-area formula. If one has equality in the BV norm expression then from the co-area formula and the minimization property of the sets $E(v)$ it is clear that $P(\{\tilde u>t\})= P(\{u>t\})$, and consequently $\{\tilde u>t\}$ is a perimeter minimizer for almost all $t$. Let $t_0=\sup\{t\,:\, \|\{\tilde u>t\}\|\ge \|H_\Omega\|\}$ so applying the uniqueness result for perimeter minimizers we see that $\{\tilde u>t\}$ and $\{u>t\}$ have the same measure theoretic closure for almost every $t,\, 0\le t <t_0$. For $t\ge t_0$ we have $\|\{\tilde u>t\}\|<\|H_\Omega\|$ so this is true only up to translation within $H_\Omega$ in which case $\partial^\{\tilde u>t\}$ is a translation of $\partial\{u>t\}$ (recall $\{u>t\}$ is convex) as claimed. This is easily justified for all $t,\,t\ge t_0$ by a limit argument. Returning to the case $0\le t<t_0$ let $E$ be an arbitrary measurable subset of $\Omega$ and $d\mu=\upchi_E\, dx$ where $dx$ represents Lebesgue measure. From Fubini’s theorem we see that $$\int_0^{t_0}\mu(\{u>t\})\, dt=\int\int_0^{t_0}\upchi_{\{u>t\}}\,dt\,d\mu =\int_E \min(u,\,t_0).$$ Using the fact that $\{\tilde u>t\}$ and $\{u>t\}$ have the same measure theoretic closure for almost every $t,\, 0\le t <t_0$ we conclude that $\min(u,\,t_0)=\min(\tilde u,\,t_0)$ almost everywhere. Recalling that $\{u>t\}$ and the set theoretic closure of $\{u>t\}$ are subsets of $H_\Omega$ for $t>t_0$ it is clear that $\tilde u=u$ almost everywhere in $\R\setminus H_\Omega$. References* 50truept Brezis, H., and D. Kinderlehrer, [The smoothness of solutions to nonlinear variational inequalities]{}, Ind. Univ. Math. J., 23(1974), 831 – 844. Federer. H., [Geometric measure theory]{}, Springer Verlag, New York, Heidelberg, 1969. Gonzalez, E., U. Massari and I. Tamanini, [Minimal boundaries enclosing a given volume]{}, Manuscripta Math., 34(1981), 381 – 395. Gonzalez, E., U. Massari and I. Tamanini, [On the regularity of sets minimizing perimeter with a volume constraint]{}, Ind. Univ. Math. J., 32(1983), 25 – 37. Grüter, M., [Boundary regularity for solutions of a partitioning problem]{}, Arch. Rat. Mech. Anal., 97(1987), 261-270. Giusti, E., [Minimal surfaces and functions of bounded variation]{}, Birkhäuser, 1985. Laurence, P., E.W. Stredulinsky, [On Quasiconvex Equimeasurable Rearrangement, a Counterexample and an Example]{}, J. fur Reine Angew. Math., 447 (1994), 63-81. Massari, U. and M. Miranda, [Minimal surfaces of codimension one]{}, Mathematics Studies, North Holland, 91 (1984). Simon, L., [Lectures on geometric measure theory]{}, Proc. Centre Math. Analysis, ANU, 3 (1983). Tamanini, I., [Boundaries of Caccioppoli sets with Hölder-continuous normal vector]{}, J. fur Reine Angew. Math., 334 (1982), 27-39. Ziemer, W.P., [Weakly differentiable functions,]{} Springer-Verlag, GTM Series, 120 1989.	{ "pile_set_name": "ArXiv" }
--- author: - \| Atsushi [Kuriyama]{},$^{1}$ Constança [Providência]{},$^{2}$\ João da [Providência]{},$^{2}$ Yasuhiko [Tsue]{}$^{3}$ and Masatoshi [Yamamura]{}$^{1}$ subtitle: ' In Relation to the su(1,1)- and its Relevant Algebras ' title: \| Note on the Orthogonal Set in\ Six Kinds of Boson Operators --- In recent paper referred to as (I),[@1] we presented a possible form of the orthogonal set for description of many-body systems in six kinds of bosons. The reason why such orthogonal set is necessary was already mentioned in (I) and, then, in this paper, we do not repeat the reason. Its basic idea comes from the paper by the present authors,[@2] in which the Schwinger boson representation of an extended $(M+1)(N+1)$-dimensional algebra containing the $su(M+1)$- and the $su(N,1)$-algebras was formulated. In (I), the case $(M=1, N=2)$ was treated and the $su(2)$-algebra in six kinds of bosons is the object of the investigation. In this note, we treat the case $(M=2, N=1)$, in which the $su(1,1)$-algebra in six kinds of bosons can be formulated. Combining with the deformed boson scheme proposed by the present authors,[@3] both cases may be helpful for analyzing more realistic systems than those treated in Ref. . First, let us recapitulate the $su(1,1)$-algebra in six kinds of bosons in a form slightly different from that developed by the present authors in Ref. including the notations. In the same notations as those in (I), the six kinds of bosons are denoted as $({\hat a}_i , {\hat a}_i^)$ and $({\hat b}_i , {\hat b}_i^)$ ($i=1,2,3$). In terms of these bosons, the following operators are defined : $$\label{1} {\hat T}_{\pm,0}=\sum_{i=1}^3 {\hat T}_{\pm,0}(i) \ ,$$ \[2\] $$\label{2a} {\hat T}_+(i)={\hat b}_i^{\hat a}_i^ \ , \qquad {\hat T}_-(i)={\hat a}_i{\hat b}_i \ , \qquad {\hat T}_0(i)=(1/2)\cdot ({\hat b}_i^{\hat b}_i+{\hat a}_i^{\hat a}_i +1) \ .$$ The set $({\hat T}_{\pm,0})$ obeys the $su(1,1)$-algebra. Associating with the operators ${\hat T}_{\pm,0}$, we introduce the operators defined by $$\label{2b} {\hat T}(i)=(1/2)\cdot({\hat b}_i^{\hat b}_i-{\hat a}_i^{\hat a}_i+1) \ . \qquad (i=1,2,3)$$ Three sets $({\hat T}_{\pm,0}(i) \ ; \ i=1,2,3)$ form the independent $su(1,1)$-algebras in the Schwinger boson representation[@4] and ${\hat T}(i)$ denotes the magnitude of the $i$-th $su(1,1)$-spin. The set $({\hat T}_{\pm,0})$ in this paper plays the same role as that of the set $({\hat S}_{\pm,0})$, which obeys the $su(2)$-algebra, in (I). It should be noted that ${\hat T}(i)$ is not positive-definite. The set $({\hat T}_{\pm,0})$ forms the addition of these three $su(1,1)$-spins. Main aim of this paper is to give the eigenstate of $({\hat {\mib T}}^2 , {\hat T}_0)$ without the limits of the conventional manner, i.e., successive addition of the $su(1,1)$-spins. Here, ${\hat {\mib T}}^2$ denotes the Casimir operator of the $su(1,1)$-algebra : $$\label{3} {\hat {\mib T}}^2={\hat T}_0^2-(1/2)\cdot({\hat T}_+{\hat T}_- +{\hat T}_-{\hat T}_+) ={\hat T}_0({\hat T}_0-1)-{\hat T}_+{\hat T}_- \ .$$ The eigenstate of $({\hat {\mib T}}^2, {\hat T}_0)$ with the eigenvalues $(t(t-1) , t_0)$ satisfies the relation $$\begin{aligned} \label{4} & &{\hat {\mib T}}^2{\| {(\gamma); t,t_0} \rangle}=t(t-1){\| {(\gamma); t,t_0} \rangle} \ , \nonumber\\ & &{\hat T}_0{\| {(\gamma); t,t_0} \rangle}=t_0{\| {(\gamma); t,t_0} \rangle} \ . \end{aligned}$$ Here, $(\gamma)$ denotes a set of four quantum numbers. Solutions of the eigenvalue equation (\[4\]) are classified into two groups : \[5\] $$\begin{aligned} & &t=+3/2 , \ +2,\ 5/2, \cdots, \qquad t_0=t,\ t+1,\ t+2, \cdots \ , \label{5a}\\ & &t=-1/2 , \ -1,\ -3/2, \cdots, \qquad t_0=1-t,\ 2-t,\ 3-t, \cdots \ . \label{5b}\end{aligned}$$ In this paper, we treat the case (\[5a\]). In this case, ${\| {(\gamma); t,t_0} \rangle}$ can be expressed in the form $$\label{6} {\| {(\gamma); t,t_0} \rangle}=({\hat T}_+)^{t_0-t}{\| {(\gamma); t} \rangle} \ .$$ Here, we call the state ${\| {(\gamma); t} \rangle}$ the intrinsic state and it should satisfy the condition $$\label{7} {\hat T}_-{\| {(\gamma); t} \rangle}=0 \ , \qquad {\hat T}_0{\| {(\gamma); t} \rangle}=t{\| {(\gamma); t} \rangle} \ .$$ In this paper, we omit any numerical factor appearing in any state such as normalization constant. In associating with the operators ${\hat T}_{\pm,0}$, we introduce two $su(3)$-algebras in the following form : \[8\] $$\begin{aligned} & &{\hat S}_+(2)={\hat a}_2^{\hat b}_3-{\hat a}_3^{\hat b}_2 \ , \qquad {\hat S}_+(1)={\hat a}_1^{\hat b}_3-{\hat a}_3^{\hat b}_1 \ , \nonumber\\ & &{\hat S}_-(2)={\hat b}_3^{\hat a}_2-{\hat b}_2^{\hat a}_3 \ , \qquad {\hat S}_-(1)={\hat b}_3^{\hat a}_1-{\hat b}_1^{\hat a}_3 \ , \label{8a}\\ & &{\hat S}_2^1={\hat a}_1^{\hat a}_2-{\hat b}_2^{\hat b}_1 \ , \qquad {\hat S}_1^2={\hat a}_2^{\hat a}_1-{\hat b}_1^{\hat b}_2 \ , \label{8b}\\ & &{\hat S}_2^2=(1/2)\cdot({\hat a}_2^{\hat a}_2-{\hat b}_2^{\hat b}_2) +(1/2)\cdot({\hat a}_3^{\hat a}_3-{\hat b}_3^{\hat b}_3) \ , \nonumber\\ & &{\hat S}_1^1=(1/2)\cdot({\hat a}_1^{\hat a}_1-{\hat b}_1^{\hat b}_1) +(1/2)\cdot({\hat a}_3^{\hat a}_3-{\hat b}_3^{\hat b}_3) \ , \label{8c}\end{aligned}$$ \[9\] $$\begin{aligned} & &{\hat R}_+(2)={\hat a}_3^{\hat a}_2-{\hat b}_2^{\hat b}_3 \ , \qquad {\hat R}_+(1)={\hat b}_1^{\hat b}_3-{\hat a}_3^{\hat a}_1 \ , \nonumber\\ & &{\hat R}_-(2)={\hat a}_2^{\hat a}_3-{\hat b}_3^{\hat b}_2 \ , \qquad {\hat R}_-(1)={\hat b}_3^{\hat b}_1-{\hat a}_1^{\hat a}_3 \ , \label{9a}\\ & &{\hat R}_2^1={\hat a}_1^{\hat a}_2-{\hat b}_2^{\hat b}_1 \ , \qquad {\hat R}_1^2={\hat a}_2^{\hat a}_1-{\hat b}_1^{\hat b}_2 \ , \label{9b}\\ & &{\hat R}_2^2=(1/2)\cdot({\hat a}_2^{\hat a}_2-{\hat b}_2^{\hat b}_2) -(1/2)\cdot({\hat a}_3^{\hat a}_3-{\hat b}_3^{\hat b}_3) \ , \nonumber\\ & &{\hat R}_1^1=(1/2)\cdot({\hat a}_1^{\hat a}_1-{\hat b}_1^{\hat b}_1) -(1/2)\cdot({\hat a}_3^{\hat a}_3-{\hat b}_3^{\hat b}_3) \ . \label{9c}\end{aligned}$$ The sets (\[8\]) and (\[9\]) obey the $su(3)$-algebras, respectively, but, they are not mutually commutable. An important relation is shown as follows : $$\label{10} [\ \hbox{\rm any\ of\ $({\hat T}_{\pm,0})$\ , \ any\ of\ the\ $su(3)$-generators\ (\ref{8})\ and\ (\ref{9})}\ ]=0 \ .$$ With the use of the generators (\[8\]) and (\[9\]), we define the following two sets of the operators : $$\begin{aligned} {\hat M}_+&=& +[{\hat S}_+(1) , {\hat S}_-(2) ] =+[{\hat R}_+(2) , {\hat R}_-(1) ] ={\hat S}_2^1={\hat R}_2^1={\hat a}_1^{\hat a}_2-{\hat b}_2^{\hat b}_1 \ , \nonumber\\ {\hat M}_-&=& -[{\hat S}_-(1) , {\hat S}_+(2) ] =-[{\hat R}_-(2) , {\hat R}_+(1) ] ={\hat S}_1^2={\hat R}_1^2={\hat a}_2^{\hat a}_1-{\hat b}_1^{\hat b}_2 \ , \nonumber\\ {\hat M}_0&=& {\hat S}_1^1-{\hat S}_2^2={\hat R}_1^1-{\hat R}_2^2 \nonumber\\ &=&(1/2)\cdot({\hat a}_1^{\hat a}_1-{\hat b}_1^{\hat b}_1) -(1/2)\cdot({\hat a}_2^{\hat a}_2-{\hat b}_2^{\hat b}_2) \ , \label{11}\\ {\hat L}_+&=& +[{\hat S}_+(2) , {\hat R}_-(1) ] ={\hat a}_2^{\hat b}_1-{\hat a}_1^{\hat b}_2 \ , \nonumber\\ {\hat L}_-&=& -[{\hat S}_-(2) , {\hat R}_+(1) ] ={\hat b}_1^{\hat a}_2-{\hat b}_2^{\hat a}_1 \ , \nonumber\\ {\hat L}_0&=& (1/2)\cdot({\hat S}_1^1+{\hat S}_2^2)+(1/2)\cdot( {\hat R}_1^1+{\hat R}_2^2) \nonumber\\ &=&(1/2)\cdot({\hat a}_1^{\hat a}_1-{\hat b}_1^{\hat b}_1) +(1/2)\cdot({\hat a}_2^{\hat a}_2-{\hat b}_2^{\hat b}_2) \ . \label{12}\end{aligned}$$ Two sets $({\hat M}_{\pm,0})$ and $({\hat L}_{\pm,0})$ obey the $su(2)$-algebras independently : \[13\] $$\label{13a} [\ \hbox{\rm any\ of\ $({\hat M}_{\pm,0})$\ , \ any\ of\ $({\hat L}_{\pm,0})$\ }]=0 \ .$$ Further, from the relation (\[10\]) and the definitions (\[11\]) and (\[12\]), we have $$\label{13b} [\ \hbox{\rm any\ of\ $({\hat M}_{\pm,0})$\ and\ $({\hat L}_{\pm,0})$\ , \ any\ of\ $({\hat T}_{\pm,0})$\ }]=0 \ .$$ The Casimir operators ${\hat {\mib M}}^2$ and ${\hat {\mib L}}^2$ are given by $$\begin{aligned} & &{\hat {\mib M}}^2={\hat M}_0^2+(1/2)\cdot ({\hat M}_+{\hat M}_-+{\hat M}_-{\hat M}_+) \ , \label{14}\\ & &{\hat {\mib L}}^2={\hat L}_0^2+(1/2)\cdot ({\hat L}_+{\hat L}_-+{\hat L}_-{\hat L}_+) \ . \label{15}\end{aligned}$$ Definitions of $({\hat M}_{\pm,0})$ and $({\hat L}_{\pm,0})$ shown in the relations (\[11\]) and (\[12\]), respectively, give us $$\label{16} {\hat {\mib L}}^2={\hat {\mib M}}^2 \ .$$ Hereafter, we call the sets $({\hat M}_{\pm,0})$ and $({\hat L}_{\pm,0})$ the $M$- and $L$-spin, respectively. Our next task is to investigate properties of the operators ${\hat S}_+(2)$, ${\hat S}_+(1)$, ${\hat R}_+(1)$ and ${\hat R}_+(2)$ with respect to the $M$- and the $L$-spin. For this aim, the following commutation relations are helpful : $$\begin{aligned} & &[ {\hat M}_+ , {\hat S}_+(2) ]={\hat S}_+(1) \ , \quad [ {\hat M}_- , {\hat S}_+(2) ]=0 \ , \quad [ {\hat M}_0 , {\hat S}_+(2) ]=-1/2\cdot {\hat S}_+(2) \ , \nonumber\\ & &[ {\hat M}_+ , {\hat S}_+(1) ]=0 \ , \quad [ {\hat M}_- , {\hat S}_+(1) ]={\hat S}_+(2) \ , \quad [ {\hat M}_0 , {\hat S}_+(1) ]=+1/2\cdot{\hat S}_+(1) \ , \nonumber\\ & &[ {\hat M}_+ , {\hat R}_+(1) ]={\hat R}_+(2) \ , \quad [ {\hat M}_- , {\hat R}_+(1) ]=0 \ , \quad [ {\hat M}_0 , {\hat R}_+(1) ]=-1/2\cdot {\hat R}_+(1) \ , \nonumber\\ & &[ {\hat M}_+ , {\hat R}_+(2) ]=0 \ , \quad [ {\hat M}_- , {\hat R}_+(2) ]={\hat R}_+(1) \ , \quad [ {\hat M}_0 , {\hat R}_+(2) ]=+1/2\cdot{\hat R}_+(2) , \quad\ \label{17}\\ & &[ {\hat L}_+ , {\hat R}_+(1) ]={\hat S}_+(2) \ , \quad [ {\hat L}_- , {\hat R}_+(1) ]=0 \ , \quad [ {\hat L}_0 , {\hat R}_+(1) ]=-1/2\cdot {\hat R}_+(1) \ , \nonumber\\ & &[ {\hat L}_+ , {\hat S}_+(2) ]=0 \ , \quad [ {\hat L}_- , {\hat S}_+(2) ]={\hat R}_+(1) \ , \quad [ {\hat L}_0 , {\hat S}_+(2) ]=+1/2\cdot{\hat S}_+(2) \ , \nonumber\\ & &[ {\hat L}_+ , {\hat R}_+(2) ]={\hat S}_+(1) \ , \quad [ {\hat L}_- , {\hat R}_+(2) ]=0 \ , \quad [ {\hat L}_0 , {\hat R}_+(2) ]=-1/2\cdot {\hat R}_+(2) \ , \nonumber\\ & &[ {\hat L}_+ , {\hat S}_+(1) ]=0 \ , \quad [ {\hat L}_- , {\hat S}_+(1) ]={\hat R}_+(2) \ , \quad [ {\hat L}_0 , {\hat S}_+(1) ]=+1/2\cdot{\hat S}_+(1) \ . \label{18}\end{aligned}$$ The relation (\[17\]) tells us that the sets $({\hat S}_+(2), {\hat S}_+(1))$ and $({\hat R}_+(1), {\hat R}_+(2))$ are spherical tensors with rank 1/2, respectively, for the $M$-spin. In the same meaning as that of the above, the relation (\[18\]) shows that the sets $({\hat R}_+(1) , {\hat S}_+(2))$ and $({\hat R}_+(2) , {\hat S}_+(1))$ are spherical tensors with rank 1/2, respectively, for the $L$-spin. As is later shown in the relation (\[23\]), this point is useful for constructing the intrinsic state ${\| {(\gamma) ,t} \rangle}$. Further, we note that the operator ${\hat P}_+$ defined in the following is a scalar with respect to both the $M$- and the $L$-spins : $$\begin{aligned} \label{19} {\hat P}_+&=&(1/2)\cdot[{\hat S}_+(2){\hat R}_+(2) -{\hat S}_+(1){\hat R}_+(1)] \nonumber\\ &=&{\hat a}_3^{\hat b}_3[1+(1/2)\cdot({\hat a}_1^{\hat a}_1 +{\hat b}_1^{\hat b}_1)+(1/2)\cdot({\hat a}_2^{\hat a}_2 +{\hat b}_2^{\hat b}_2)] \nonumber\\ & & -({\hat b}_1^{\hat a}_1^+{\hat b}_2^{\hat a}_2^)\cdot(1/2)\cdot {\hat b}_3^{2}-(1/2)\cdot{\hat a}_3^{2}\cdot({\hat a}_1{\hat b}_1 +{\hat a}_2{\hat b}_2) \ .\end{aligned}$$ The operator ${\hat P}_+$ commutes with ${\hat S}_+(2)$, ${\hat S}_+(1)$, ${\hat R}_+(2)$ and ${\hat R}_+(1)$. Finally, we contact with the Casimir operator, which we denote as ${\hat {\mib S}}^2$, for the $su(3)$-algebra defined in the relation (\[8\]). The operator ${\hat {\mib S}}^2$ is given in the following form : $$\begin{aligned} \label{20} {\hat {\mib S}}^2&=& (4/3)\cdot[({\hat S}_1^1)^2-{\hat S}_1^1{\hat S}_2^2+({\hat S}_2^2)^2] +(1/2)\cdot({\hat S}_2^1{\hat S}_1^2+{\hat S}_1^2{\hat S}_2^1) \nonumber\\ & &+(1/2)\cdot[{\hat S}_+(1){\hat S}_-(1)+{\hat S}_-(1){\hat S}_+(1)] +(1/2)\cdot[{\hat S}_+(2){\hat S}_-(2)+{\hat S}_-(2){\hat S}_+(2)] \ . \nonumber\\\end{aligned}$$ The above expression can be rewritten as $$\label{21} {\hat {\mib S}}^2=(1/3)\cdot {\hat S}_0({\hat S}_0-3)+{\hat {\mib M}}^2+ [{\hat S}_+(1){\hat S}_-(1)+{\hat S}_+(2){\hat S}_-(2)] \ .$$ Here, ${\hat S}_0$ is defined in the form $$\begin{aligned} \label{22} {\hat S}_0&=&{\hat S}_1^1+{\hat S}_2^2 \nonumber\\ &=&(1/2)\cdot({\hat a}_1^{\hat a}_1-{\hat b}_1^{\hat b}_1) +(1/2)\cdot({\hat a}_2^{\hat a}_2-{\hat b}_2^{\hat b}_2) +({\hat a}_3^{\hat a}_3-{\hat b}_3^{\hat b}_3) \ . \quad\end{aligned}$$ The above is our framework for the later discussion. From the above framework, we have four mutually commutable hermitian operators, ${\hat {\mib M}}^2$, ${\hat M}_0$, ${\hat {\mib S}}^2$ and ${\hat S}_0$, which are also commuted with ${\hat T}_{\pm,0}$. They are expressed in terms of the $su(3)$-generators defined in the relation (\[8\]). The eigenvalue equations for these operators determine the set of four quantum numbers $(\gamma)$ appearing in the intrinsic state shown in the relation (\[6\]). The idea presented in (I) helps us to determine ${\| {(\gamma);t} \rangle}$ in the following form : \[23\] $$\begin{aligned} {\| {(\gamma);t} \rangle}&=& {\| {k,l,m,m_0;t} \rangle} \nonumber\\ &=&\sum_{k_0l_0}C(k,k_0;l,l_0\|m,m_0) {\hat S}_+(k,k_0) {\hat R}_+(l,l_0)\cdot({\hat b}_3^)^{2t-3} {\| {0} \rangle} \ , \label{23x}\\ {\hat S}_+(k,k_0)&=& \left(\sqrt{(k+k_0)!(k-k_0)!}\right)^{-1} ({\hat S}_+(1))^{k+k_0}({\hat S}_+(2))^{k-k_0} \ , \label{23a}\\ {\hat R}_+(l,l_0)&=& \left(\sqrt{(l+l_0)!(l-l_0)!}\right)^{-1} ({\hat R}_+(2))^{l+l_0}({\hat R}_+(1))^{l-l_0} \ , \label{23b}\end{aligned}$$ In the same meaning as that shown in (I), ${\hat S}_+(k,k_0)$ and ${\hat R}_+(l,l_0)$ are spherical tensors with rank $k$ and $l$, respectively. Of course, $k_0$ and $l_0$ denote the $k_0$-th and the $l_0$-th components, respectively. The quantities $C(k,k_0;l,l_0\|m,m_0)$ is the Clebsch-Gordan(CG)-coefficient. The form (\[23\]) gives us the eigenvalues of ${\hat {\mib M}}^2$ and ${\hat M}_0$, $({\mib M}^2)_{\rm ev}$ and $(M_0)_{\rm ev}$ : $$\begin{aligned} \label{24} & &({\mib M}^2)_{\rm ev}= m(m+1) \ , \nonumber\\ & &(M_0)_{\rm ev}=m_0 \ .\end{aligned}$$ Further, the eigenvalue of ${\hat S}_0$, $(S_0)_{\rm ev}$, is obtained in the form $$\label{25} (S_0)_{\rm ev}=3k-(2t-3-l) \ .$$ For the derivation of the above result, the following relation is useful : $$\begin{aligned} \label{26} & &[{\hat S}_0 , {\hat S}_+(1) ] = 3/2\cdot {\hat S}_+(1) \ , \qquad [{\hat S}_0 , {\hat S}_+(2) ] = 3/2\cdot {\hat S}_+(2) \ , \nonumber\\ & &[{\hat S}_0 , {\hat R}_+(1) ] = 1/2\cdot {\hat R}_+(1) \ , \qquad [{\hat S}_0 , {\hat R}_+(2) ] = 1/2\cdot {\hat R}_+(2) \ . \end{aligned}$$ The eigenvalue of ${\hat {\mib S}}^2$, $({\mib S}^2)_{\rm ev}$, is given as $$\label{27} ({\mib S}^2)_{\rm ev}=(1/3)\cdot(2t-3-l)(2t-l)+l(l+1) \ .$$ The above result is obtained through the relations $$\begin{aligned} & &{\hat {\mib S}}^2{\| {(\gamma);t} \rangle} =\sum_{k_0l_0}C(k,k_0;l,l_0\|m,m_0){\hat S}_+(k,k_0)\cdot {\hat {\mib S}}^2{\| {l,l_0;t} \rangle} \ , \label{28}\\ & &{\| {l,l_0;t} \rangle}={\hat R}_+(l,l_0)({\hat b}_3^)^{2t-3}{\| {0} \rangle} \ , \label{29}\\ & &{\hat S}_-(1){\| {l,l_0;t} \rangle}={\hat S}_-(2){\| {l,l_0;t} \rangle}=0 \ , \label{30}\\ & &{\hat S}_0{\| {l,l_0;t} \rangle}=-(2t-3-l){\| {l,l_0;t} \rangle} \ , \label{31}\\ & &{\hat {\mib M}}^2{\| {l,l_0;t} \rangle}=l(l+1){\| {l,l_0;t} \rangle} \ . \label{32}\end{aligned}$$ As is shown in the relation (\[16\]), ${\hat {\mib L}}^2={\hat {\mib M}}^2$ and, then, the state (\[23\]) is an eigenstate of ${\hat {\mib L}}^2$ with the eigenvalue $m(m+1)$. Further, the state (\[23\]) is an eigenstate of ${\hat L}_0$ with the eigenvalue $(L_0)_{\rm ev}$ : $$\label{33} (L_0)_{\rm ev}=k-l \ .$$ The state (\[23x\]) is of the polynomial form and, therefore, from the reason mentioned in (I), it may be interesting to give the form in terms of the monomial. For this aim, first, we define the following state : $$\label{34} {\| {n,m;t} \rangle\!\rangle}=({\hat P}_+)^n{\| {m,-m;t} \rangle} \ .$$ Here, ${\hat P}_+$ and ${\| {m,-m;t} \rangle}$ are defined in the relations (\[19\]) and (\[29\]), respectively. The state ${\| {n,m;t} \rangle\!\rangle}$ satisfies $$\begin{aligned} \label{35} & &{\hat M}_-{\| {n,m;t} \rangle\!\rangle}={\hat L}_-{\| {n,m;t} \rangle\!\rangle}=0 \ , \nonumber\\ & &{\hat M}_0{\| {n,m;t} \rangle\!\rangle}={\hat L}_0{\| {n,m;t} \rangle\!\rangle}=-m{\| {n,m;t} \rangle\!\rangle} \ . \end{aligned}$$ Next, we introduce the state $$\label{36} {\| {n,k-l,m,m_0;t} \rangle\!\rangle}=({\hat M}_+)^{m_0+m}({\hat L}_+)^{k-l+m} {\| {n,m;t} \rangle\!\rangle} \ .$$ With the help of the relation (\[35\]), we can show that the state (\[36\]) is an eigenstate of ${\hat {\mib M}}^2\ (={\hat {\mib L}}^2)$, ${\hat M}_0$ and ${\hat L}_0$ with the eigenvalues $m(m+1)$, $m_0$ and $k-l$, respectively. These are identical with the eigenvalues shown in the relations (\[24\]) and (\[33\]), respectively. The quantum number $n$ can be determined in the process mentioned below. We can prove the relation $$\label{37} {\hat S}_0{\| {n,k-l,m,m_0;t} \rangle\!\rangle} =[(k-l)+2(m+n)-(2l-3)]{\| {n,k-l,m,m_0;t} \rangle\!\rangle} \ .$$ Then, equating the eigenvalue (\[37\]) with $[3k-(2t-3-l)]$ shown in the relation (\[25\]), we have $$\label{38} n=k+l-m \ .$$ Thus, we are able to rewrite the state (\[23x\]) in terms of the monomial shown in the following form : $$\label{39} {\| {k,l,m,m_0;t} \rangle} =({\hat M}_+)^{m_0+m}({\hat L}_+)^{k-l+m}({\hat P}_+)^{k+l-m} ({\hat R}_+(1))^{2m}({\hat b}_3^)^{2t-3}{\| {0} \rangle} \ .$$ It can be proved that the state (\[39\]) is an eigenstate of ${\hat {\mib S}}^2$, the eigenvalue of which is given in the relation (\[27\]). This state can be rewritten as $$\label{40} {\| {k,l,m,m_0;t} \rangle} =({\hat M}_+)^{m_0+m}({\hat S}_+(2))^{k-l+m}({\hat R}_+(1))^{l-k+m} ({\hat P}_+)^{k+l-m}({\hat b}_3^)^{2t-3}{\| {0} \rangle} \ .$$ The quantities $(k-l+m)$, $(l-k+m)$ and $(k+l-m)$ in the state (\[40\]) should be positive and we have $$m\ge l-k \ , \qquad m\ge k-l \ , \qquad m\le k+l \ , \nonumber$$ that is, $$\label{41} \|k-l\| \le m \le k+l \ .$$ Further, as is proved at the end of this paper, the state (\[39\]) obeys the condition $$\label{42} k+l+m \le 2t-3 \ .$$ Of course, concerning $m$, $m_0$, $t$ and $t_0$, we have the relations $$\begin{aligned} & & -m \le m_0 \le +m \ , \label{43}\\ & &t_0 \ge t \ , \qquad t \ge 3/2 \ . \label{44}\end{aligned}$$ In principle, the normalization constant can be calculated, but, practically, it may be not necessary to give it. Our next problem is to investigate successive addition of three $su(1,1)$-spins. First, we note that there exist six mutually commutable hermitian operators : ${\hat {\mib T}}^2$, ${\hat T}_0$, ${\hat T}(1)$, ${\hat T}(2)$, ${\hat T}(3)$ and ${\hat {\mib T}}_{12}^2$. Here, ${\hat {\mib T}}^2$, ${\hat T}_0$, ${\hat T}(1)$, ${\hat T}(2)$ and ${\hat T}(3)$ are given in the relations (\[3\]), (\[1\]) and (\[2b\]), respectively. The operator ${\hat {\mib T}}_{12}^2$ denotes the Casimir operator of the $su(1,1)$-algebra composed of the two $su(1,1)$-spins specified by the indices 1 and 2 : $$\begin{aligned} \label{45} {\hat {\mib T}}_{12}^2&=& ({\hat T}_0(1)+{\hat T}_0(2))^2 \nonumber\\ & &-(1/2)\!\cdot\![({\hat T}_+(1)\!+\!{\hat T}_+(2)) ({\hat T}_-(1)\!+\!{\hat T}_-(2)) +({\hat T}_-(1)\!+\!{\hat T}_-(2)) ({\hat T}_+(1)\!+\!{\hat T}_+(2))] \ . \nonumber\\\end{aligned}$$ The four operators except for $({\hat {\mib T}}^2 , {\hat T}_0)$ can be expressed as follows : \[46\] $$\begin{aligned} & &{\hat T}(1)=-(1/2)\cdot({\hat L}_0+{\hat M}_0-1)\ , \qquad {\hat T}(2)=-(1/2)\cdot({\hat L}_0-{\hat M}_0-1)\ , \nonumber\\ & &{\hat T}(3)=+(1/2)\cdot({\hat L}_0-{\hat S}_0+1)\ , \label{46a}\\ & &{\hat {\mib T}}_{12}^2={\hat {\mib M}}^2 \ (={\hat {\mib L}}^2) \ . \label{46b}\end{aligned}$$ Here, for ${\hat M}_0$, ${\hat L}_0$ and ${\hat S}_0$, the relations (\[11\]), (\[12\]) and (\[22\]) were used. The relation (\[46b\]) is given by the straightforward calculation. The state (\[39\]) is the eigenstate of ${\hat M}_0$, ${\hat L}_0$, ${\hat S}_0$ and ${\hat {\mib M}}^2$. Therefore, the state (\[39\]) is also the eigenstate of ${\hat T}(1)$, ${\hat T}(2)$, ${\hat T}(3)$ and ${\hat {\mib T}}_{12}^2$, the eigenvalues of which we denote as $t_1$, $t_2$, $t_3$ and $t_{12}(t_{12}-1)$, respectively : \[47\] $$\begin{aligned} & &t_1=-(1/2)\cdot(k-l+m_0-1) \ , \qquad t_2=-(1/2)\cdot(k-l-m_0-1) \ , \nonumber\\ & &t_3=t-1-(k+l) \ , \label{47a}\\ & &t_{12}(t_{12}-1)=m(m+1) \ . \label{47b}\end{aligned}$$ The relation (\[47a\]) is solved inversely : $$\begin{aligned} \label{48} & &k=(1/2)\cdot(t-t_1-t_2-t_3) \ , \qquad k+l=t-t_3-1 \ , \nonumber\\ \hbox{\rm i.e.,}& & \nonumber\\ & &l=(1/2)\cdot(t+t_1+t_2-t_3)-1 \ , \qquad k-l=1-(t_1+t_2) \ , \nonumber\\ & &m_0=t_2-t_1 \ . \end{aligned}$$ The relation (\[47b\]) gives us the following relation : \[49\] $$\begin{aligned} & &m=t_{12}-1 \ , \label{49a}\\ & &m=-t_{12} \ . \label{49b}\end{aligned}$$ Substituting the relation (\[48\]) and (\[49a\]) into the inequality (\[41\])$\sim$(\[43\]), we have the inequality \[50\] $$\begin{aligned} \label{50a} & &t_{12} \ge \|t_1-1/2\|+\|t_2-1/2\|+1 \ , \nonumber\\ & &t \ge \|t_{12}-1\|+\|t_3-1/2\|+3/2 \ .\end{aligned}$$ The relations (\[48\]) and (\[49b\]) also give us the inequality $$\begin{aligned} \label{50b} & &t_{12} \le -(\|t_1-1/2\|+\|t_2-1/2\|) \ , \nonumber\\ & &t \ge \|t_{12}\|+\|t_3-1/2\|+3/2 \ .\end{aligned}$$ The relations (\[50a\]) and (\[50b\]) show the coupling rule for two and three $su(1,1)$-spins. Since $m\ge 0$, we have $\|t_{12}-1\|=t_{12}-1 \ge 0$ and $\|t_{12}\|=-t_{12} \ge 0$. Substitution of the relations (\[48\]) and (\[49\]) into the state (\[40\]) leads us to \[51\] $$\begin{aligned} {\| {t_1,t_2,(t_{12}),t_3;t} \rangle} &=&({\hat M}_+)^{(t_{12}-1)-(t_1-1/2)+(t_2-1/2)} \nonumber\\ & &\times ({\hat S}_+(2))^{(t_{12}-1)-(t_1-1/2)-(t_2-1/2)} ({\hat R}_+(1))^{(t_{12}-1)+(t_1-1/2)+(t_2-1/2)} \nonumber\\ & &\times ({\hat P}_+)^{(t-3/2)-(t_{12}-1)-(t_3-1/2)} ({\hat b}_3^)^{2t-3}{\| {0} \rangle} \ . \quad (t_{12}-1 \ge 0) \label{51a}\\ {\| {t_1,t_2,(t_{12}),t_3;t} \rangle} &=&({\hat M}_+)^{(-t_{12})-(t_1-1/2)+(t_2-1/2)} \nonumber\\ & &\times ({\hat S}_+(2))^{(-t_{12})-(t_1-1/2)-(t_2-1/2)} ({\hat R}_+(1))^{(-t_{12})+(t_1-1/2)+(t_2-1/2)} \nonumber\\ & &\times ({\hat P}_+)^{(t-3/2)-(-t_{12})-(t_3-1/2)} ({\hat b}_3^)^{2t-3}{\| {0} \rangle} \ . \quad (-t_{12} \ge 0) \label{51b}\end{aligned}$$ Here, the state ${\| {t_1,t_2,(t_{12}),t_3;t} \rangle}$ is relabeled from ${\| {k,l,m,m_0;t} \rangle}$. The above is the formalism in terms of the addition of three $su(1,1)$-spins. This paper closes with proving the inequality (\[42\]). For this aim, we take up the following state which is related with the state (\[40\]) : $$\label{52} {\| {s,r,p;T} \rangle\!\rangle}=({\hat S}_+(2))^s({\hat R}_+(1))^r({\hat P}_+)^p ({\hat b}_3^)^T{\| {0} \rangle} \ .$$ The state (\[40\]) can be expressed as $$\label{53} {\| {k,l,m,m_0;t} \rangle}=({\hat M}_+)^{m_0+m}{\| {k-l+m,l-k+m,k+l-m;2t-3} \rangle\!\rangle} \ .$$ The definition of ${\hat P}_+$ shown in the relation (\[19\]) gives us the relation $$\begin{aligned} \label{54} ({\hat P}_+)^p({\hat b}_3^)^T{\| {0} \rangle} &=&\sum_{q=0}^p D(p,q)(T!/(T-p-q)!) \nonumber\\ & &\qquad\qquad\times ({\hat b}_1^{\hat a}_1^+{\hat b}_2^{\hat a}_2^)^q({\hat a}_3^)^{p-q} ({\hat b}_3^)^{T-p-q} {\| {0} \rangle} \ .\end{aligned}$$ Here, $D(p,q)$ obeys the recursion formula $$\label{55} D(p,q)=-(1/2)\cdot D(p-1,q-1)+(q+1)D(p-1,q) -(1/2)\cdot(q+1)(q+2)D(p-1,q+1) \ .$$ Of course, we have $$\label{56} D(0,0)=1 \ , \qquad D(p,q)=0 \quad \hbox{\rm for}\quad q\le -1 \ \ \hbox{\rm and}\ \ q\ge p+1 \ .$$ Some examples are given in the following numbers : $$\begin{aligned} \label{57} & &D(1,0)=1 \ , \qquad D(1,1)=-1/2 \ , \nonumber\\ & &D(2,0)=3/2 \ , \qquad D(2,1)=-3/2 \ , \qquad D(2,2)=1/4 \ , \nonumber\\ & &D(3,0)=3 \ , \quad D(3,1)=-9/2 \ , \quad D(3,2)=3/2 \ , \quad D(3,3)=-1/8 \ . \end{aligned}$$ Further, we rewrite the state (\[52\]) under the relation $$\label{58} {\hat b}_1^{\hat a}_1^+{\hat b}_2^{\hat a}_2^* ={\hat T}_+-{\hat b}_3^{\hat a}_3^ \ , \qquad [ {\hat T}_+ , {\hat b}_3^{\hat a}_3^ ]=0 \ .$$ Then, the binomial theorem gives us $$\label{59} ({\hat b}_1^{\hat a}_1^+{\hat b}_2^{\hat a}_2^)^q =\sum_{v=0}^q (-)^{q-v}(q!/v!(q-v)!)\cdot ({\hat T}_+)^v ({\hat b}_3^{\hat a}_3^)^{q-v} \ .$$ Substituting the relation (\[59\]) into the form (\[54\]), we have $$\begin{aligned} & &({\hat P}_+)^p({\hat b}_3^)^T{\| {0} \rangle} =\sum_{q=0}^p C(p,q)({\hat T}_+)^q({\hat a}_3^)^{p-q} ({\hat b}_3^)^{T-p-q} {\| {0} \rangle} \ , \label{60}\\ & &C(p,q)=\sum_{v=0}^p D(p,v)(T!/(T-p-v)!)\cdot (-)^{v-q} (q!/(v-q)!v!) \ .\end{aligned}$$ Next, with the aid of the form (\[60\]), we rewrite the state (\[52\]). First, the following relation should be noted : $$\label{62} [{\hat S}_+(2) , {\hat T}_+ ]=[{\hat R}_+(1) , {\hat T}_+ ]=0 \ .$$ Then, under the definitions of ${\hat S}_+(2)$ and ${\hat R}_+(1)$ shown in the relations (\[8a\]) and (\[9a\]), respectively, we have $$\begin{aligned} \label{63} {\| {s,r,p;T} \rangle\!\rangle}&=& \sum_{q=0}^p C(p,q)({\hat T}_+)^q({\hat S}_+(2))^s({\hat R}_+(1))^r ({\hat a}_3^)^{p-q}({\hat b}_3^)^{T-p-q}{\| {0} \rangle} \nonumber\\ &=&\sum_{q=0}^p C(p,q)({\hat T}_+)^q({\hat a}_2^{\hat b}_3)^s ({\hat b}_1^{\hat b}_3)^r({\hat a}_3^)^{p-q}({\hat b}_3^)^{T-p-q} {\| {0} \rangle} \nonumber\\ &=&\sum_{q=0}^p C(p,q) ((T-p-q)!/(T-p-q-s-r)!) \nonumber\\ & &\qquad\qquad \times \cdot({\hat T}_+)^q({\hat a}_2^)^s({\hat b}_1^)^r({\hat a}_3^)^{p-q} ({\hat b}_3^)^{(T-p-s-r)-q}{\| {0} \rangle} \ . \end{aligned}$$ For the expression (\[63\]), let us impose the condition $$\label{64} T-p-s-r \le -1 \ .$$ Under the condition (\[64\]), the state ${\| {s,r,p;T} \rangle\!\rangle}$ cannot exist independently of any other condition. Therefore, we can conclude that the condition supporting the existence of the state ${\| {s,r,p;T} \rangle\!\rangle}$ is expressed as $$\label{65} T-p-s-r \ge 0 \ .$$ If $s=k-l+m$, $r=l-k+m$, $p=k+l-m$ and $T=2t-3$, the condition (\[65\]) is rewritten in the form $$\label{66} k+l+m \le 2t-3 \ .$$ The condition (\[66\]) is identical with the inequality (\[42\]). Thus, we can conclude that the orthogonal set in six kinds of bosons can be formulated in relation to the $su(1,1)$-algebra in six kinds of bosons. [99]{} A. Kuriyama, C. Providência, J. da Providência, Y. Tsue and M. Yamamura, Submitted to Prog. Theor. Phys. A. Kuriyama, J. da Providência and M. Yamamura, Prog. Theor. Phys. [103]{} (2000), 285. A. Kuriyama, C. Providência, J. da Providência, Y. Tsue and M. Yamamura, Prog. Theor. Phys. [108]{} (2002), 323. J. Schwinger, in [Quantum Theory of Angular Momentum*]{}, ed. L. Biedenharn and H. Van Dam (Academic Press, New York, 1955), p.229.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Separation Logic with inductive definitions is a well-known approach for deductive verification of programs that manipulate dynamic data structures. Deciding verification conditions in this context is usually based on user-provided lemmas relating the inductive definitions. We propose a novel approach for generating these lemmas automatically which is based on simple syntactic criteria and deterministic strategies for applying them. Our approach focuses on iterative programs, although it can be applied to recursive programs as well, and specifications that describe not only the shape of the data structures, but also their content or their size. Empirically, we find that our approach is powerful enough to deal with sophisticated benchmarks, e.g., iterative procedures for searching, inserting, or deleting elements in sorted lists, binary search tress, red-black trees, and AVL trees, in a very efficient way.' author: - 'Constantin Enea, Mihaela Sighireanu and Zhilin Wu' bibliography: - 'sl.bib' title: 'On Automated Lemma Generation for Separation Logic with Inductive Definitions[^1]' --- Introduction {#sec:intro} ============ Program verification requires reasoning about complex, unbounded size data structures that may carry data ranging over infinite domains. Examples of such structures are multi-linked lists, nested lists, trees, etc. Programs manipulating such structures perform operations that may modify their shape (due to dynamic creation and destructive updates) as well as the data attached to their elements. An important issue is the design of logic-based frameworks that express assertions about program configurations (at given control points), and then to check automatically the validity of these assertions, for all computations. This leads to the challenging problem of finding relevant compromises between expressiveness, automation, and scalability. An established approach for scalability is the use of Separation logic (SL) [@ORY01; @Rey02]. Indeed, its support for local reasoning based on the “frame rule” leads to compact proofs, that can be dealt with in an efficient way. However, finding expressive fragments of SL for writing program assertions, that enable efficient automated validation of the verification conditions, remains a major issue. Typically, SL is used in combination with inductive definitions, which provide a natural description of the data structures manipulated by a program. Moreover, since program proofs themselves are based on induction, using inductive definitions instead of universal quantifiers (like in approaches based on first-order logic) enables scalable automation, especially for recursive programs which traverse the data structure according to their inductive definition, e.g., [@QGS+13]. Nevertheless, automating the validation of the verification conditions generated for [iterative programs]{}, that traverse the data structures using while loops, remains a challenge. The loop invariants use inductive definitions for fragments of data structures, traversed during a partial execution of the loop, and proving the inductiveness of these invariants requires non-trivial lemmas relating (compositions of) such inductive definitions. Most of the existing works require that these lemmas be provided by the user of the verification system, e.g., [@CHL11; @QGS+13; @NC08] or they use translations of SL to first-order logic to avoid this problem. However, the latter approaches work only for rather limited fragments [@PWZ13; @PWZ14]. In general, it is difficult to have lemmas relating complex user-defined inductive predicates that describe not only the shape of the data structures but also their content. To illustrate this difficulty, consider the simple example of a sorted singly linked list. The following inductive definition describes a sorted list segment from the location $E$ to $F$, storing a multiset of values $M$: where ${\mathtt{emp}}$ denotes the empty heap, $E\mapsto\{({\mathtt{next}},X),({\mathtt{data}},v)\}$ states that the pointer field ${\mathtt{next}}$ of $E$ points to $X$ while its field ${\mathtt{data}}$ stores the value $v$, and ${\ast}$ is the separating conjunction. Proving inductive invariants of typical sorting procedures requires such an inductive definition and the following lemma: The data constraints in these lemmas, e.g., $M_1\le M_2$ (stating that every element of $M_1$ is less or equal than all the elements of $M_2$), which become more complex when reasoning for instance about binary search trees, are an important obstacle for trying to synthesize them automatically. Our work is based on a new class of inductive definitions for describing fragments of data structures that (i) supports lemmas [without additional]{} data constraints like $M_1 \le M_2$ and (ii) allows to [automatically synthesize]{} these lemmas using efficiently checkable, almost syntactic, criteria. For instance, we use a different inductive definition for ${\mathit{lseg}}$, which introduces an additional parameter $M'$ that provides a “data port” for appending another sorted list segment, just like $F$ does for the shape of the list segment: The new definition satisfies the following simpler lemma, which avoids the introduction of data constraints: Besides such “composition” lemmas (formally defined in Sec. \[sec:comp\]), we define (in Sec. \[sec:derived\]) other classes of lemmas needed in program proofs and we provide efficient criteria for generating them automatically. Moreover, we propose (in Sec. \[sec:slice\]) a proof strategy using such lemmas, based on simple syntactic matchings of spatial atoms (points-to atoms or predicate atoms like ${\mathit{lseg}}$) and reductions to SMT solvers for dealing with the data constraints. We show experimentally (in Sec. \[sec:exp\]) that this proof strategy is powerful enough to deal with sophisticated benchmarks, e.g., the verification conditions generated from the iterative procedures for searching, inserting, or deleting elements in binary search trees, red-black trees, and AVL trees, in a very efficient way. The appendix contains the proofs of theorems and additional classes of lemmas. Motivating Example {#sec:motivation} ================== Fig. \[fig:bst\_search\] lists an iterative implementation of a search procedure for binary search trees (BSTs). The property that $E$ points to the root of a BST storing a multiset of values $M$ is expressed by the following inductively-defined predicate: [l]{}[0.4]{} The predicate ${\mathit{bst}}(E,M)$ is defined by two rules describing empty (eq. (\[rule:bst-base\])) and non-empty trees (eq. (\[rule:bst-rec\])). The body (right-hand side) of each rule is a conjunction of a pure formula, formed of (dis)equalities between location variables (e.g. $E = {{\sf nil}}$) and data constraints (e.g. $M= \emptyset$), and a spatial formula describing the structure of the heap. The data constraints in eq. (\[rule:bst-rec\]) define $M$ to be the multiset of values stored in the tree, and state the sortedness property of BSTs. The precondition of `search` is ${\mathit{bst}}(\texttt{root},M_0)$, where $M_0$ is a ghost variable denoting the multiset of values stored in the tree, while its postcondition is ${\mathit{bst}}(\texttt{root},M_0) \land (\texttt{key} \in M_0 \rightarrow \mathit{ret} = 1) \land (\texttt{key} \not \in M_0 \rightarrow \mathit{ret} = 0)$, where $\mathit{ret}$ denotes the return value. The while loop traverses the BST in a top-down manner using the pointer variable `t`. This variable decomposes the heap into two domain-disjoint sub-heaps: the tree rooted at `t`, and the truncated tree rooted at `root` which contains a “hole” at `t`. To specify the invariant of this loop, we define another predicate ${\mathit{bsthole}}(E,M_1,F,M_2)$ describing “truncated” BSTs with one hole $F$ as follows: Intuitively, the parameter $M_2$, interpreted as a multiset of values, is used to specify that the structure described by ${\mathit{bsthole}}(E,M_1,F,M_2)$ could be extended with a BST rooted at $F$ and storing the values in $M_2$, to obtain a BST rooted at $E$ and storing the values in $M_1$. Thus, the parameter $M_1$ of ${\mathit{bsthole}}$ is the union of $M_2$ with the multiset of values stored in the truncated BST represented by ${\mathit{bsthole}}(E,M_1,F,M_2)$. Using ${\mathit{bsthole}}$, we obtain a succinct specification of the loop invariant: We illustrate that such inductive definitions are appropriate for automated reasoning, by taking the following branch of the loop: [assume(t != NULL);]{} [assume(t->data > key);]{} [t]{}$'$ [= t->left]{} (as usual, [if]{} statements are transformed into [assume]{} statements and primed variables are introduced in assignments). The postcondition of ${\it Inv}$ w.r.t. this branch, denoted ${\it post}({\it Inv})$, is computed as usual by unfolding the ${\mathit{bst}}$ predicate: The preservation of ${\it Inv}$ by this branch is expressed by the entailment ${\it post}({\it Inv}){\Rightarrow}{\it Inv}'$, where ${\it Inv}'$ is obtained from ${\it Inv}$ by replacing ${\tt t}$ with ${\tt t}'$. Based on the lemmas, this paper also proposes a deterministic proof strategy for proving the validity of entailments of the form $\varphi_1{\Rightarrow}\exists \vec{X}.\varphi_2$, where $\varphi_1,\varphi_2$ are quantifier-free and $\vec{X}$ contains only data variables[^2]. The strategy comprises two steps: (i) enumerating spatial atoms $A$ from $\varphi_2$, and for each of them, carving out a sub-formula $\varphi_A$ of $\varphi_1$ that entails $A$, where it is required that these subformulas do not share spatial atoms (due to the semantics of separation conjunction), and (ii) proving that the data constraints from $\varphi_A$ imply those from $\varphi_2$ (using SMT solvers). The step (i) may generate constraints on the variables in $\varphi_A$ and $\varphi_2$ that are used in step (ii). If the step (ii) succeeds, then the entailment holds. For instance, by applying this strategy to the entailment ${\it post}({\it Inv}){\Rightarrow}{\it Inv}'$ above, we obtain two goals for step (i) which consist in computing two sub-formulas of ${\it post}({\it Inv})$ that entail $\exists M_1'.\ {\mathit{bsthole}}({\tt root},M_0,{\tt t}',M_1')$ and respectively, $\exists M_1''.\ {\mathit{bst}}({\tt t}',M_1'')$. This renaming of existential variables requires adding the equality $M_1=M_1'=M_1''$ to ${\it Inv}'$. The second goal, for $\exists M_1''.\ {\mathit{bst}}({\tt t}',M_1'')$, is solved easily since this atom almost matches the sub-formula ${\mathit{bst}}({\tt t}',M_2)$. This matching generates the constraint $M_1''=M_2$, which provides an instantiation of the existential variable $M_1''$ useful in proving the entailment between the data constraints in step (ii). Computing a sub-formula that entails $\exists M_1'.\ {\mathit{bsthole}}({\tt root},M_0,{\tt t}',M_1')$ requires a non-trivial lemma. Thus, according to the syntactic criteria defined in Sec. \[sec:comp\], the predicate ${\mathit{bsthole}}$ enjoys the following composition lemma: Intuitively, this lemma states that composing two heap structures described by ${\mathit{bsthole}}$ results in a structure that satisfies the same predicate. The particular relation between the arguments of the predicate atoms in the left-hand side is motivated by the fact that the parameters $F$ and $M$ are supposed to represent “ports” for composing ${\mathit{bsthole}}({\tt root},M_0,F, M)$ with some other similar heap structures. This property of $F$ and $M$ is characterized syntactically by the fact that, roughly, $F$ (resp. $M$) occurs only once in the body of each inductive rule of ${\mathit{bsthole}}$, and $F$ (resp. $M$) occurs only in an equality with ${\tt root}$ (resp. $M_0$) in the base rule (we are referring to the rules (\[rule:bsth-base\])–(\[rule:bsth-recl\]) with the parameters of ${\mathit{bsthole}}$ substituted by $({\tt root},M_0,F, M)$). Therefore, the first goal reduces to finding a sub-formula of ${\it post}({\it Inv})$ that implies the premise of (\[eq:bsth-lemma\]) where $M_1'$ remains existentially-quantified. Recursively, we apply the same strategy of enumerating spatial atoms and finding sub-formulas that entail them. However, we are relying on the fact that all the existential variables denoting the root locations of spatial atoms in the premise of the lemma, e.g., $F$ in lemma (\[eq:bsth-lemma\]), occur as arguments in the only spatial atom of the conclusion whose root location is the same as that of the consequent, i.e., ${\mathit{bsthole}}({\tt root},M_0, F, M)$ in lemma (\[eq:bsth-lemma\]). Therefore, the first sub-goal, $\exists F,M.\ {\mathit{bsthole}}({\tt root},M_0, F, M)$ matches the atom ${\mathit{bsthole}}({\tt root},M_0,{\tt t},M_1)$, under the constraint $F={\tt t}\land M=M_1$. This constraint is used in solving the second sub-goal, which now becomes $\exists M_1'.\ {\mathit{bsthole}}({\tt t},M_1,{\tt t}',M_1')$. The second sub-goal is proved by unfolding ${\mathit{bsthole}}$ twice, using first the rule (\[rule:bsth-recl\]) and then the rule (\[rule:bsth-base\]), and by matching the resulting spatial atoms with those in ${\it post}({\it Inv})$ one by one. Assuming that the existential variable $M_1$ from ${\it Inv}'$ is instantiated with $M_2$ from ${\it post}({\it Inv})$ (fact automatically deduced in the first step), the data constraints in ${\it post}({\it Inv})$ entail those in ${\it Inv}'$. This completes the proof of ${\it post}({\it Inv}){\Rightarrow}{\it Inv}'$. Separation Logic with Inductive Definitions {#sec:logic} =========================================== Let ${{\sf LVar}}$ be a set of location variables, interpreted as heap locations, and ${{\sf DVar}}$ a set of data variables, interpreted as data values stored in the heap, (multi)sets of values, etc. In addition, let ${\mathsf{Var}}={{\sf LVar}}\cup {{\sf DVar}}$. The domain of heap locations is denoted by ${{\mathbb L}}$ while the domain of data values stored in the heap is generically denoted by ${{\mathbb D}}$. Let ${\mathcal{F}}$ be a set of pointer fields, interpreted as functions ${{\mathbb L}}{\rightharpoonup}{{\mathbb L}}$, and ${\mathcal{D}}$ a set of data fields, interpreted as functions ${{\mathbb L}}{\rightharpoonup}{{\mathbb D}}$. The syntax of the Separation Logic fragment considered in this paper is defined in Tab. \[tab:syntaxSL\]. Formulas are interpreted over pairs $(s,h)$ formed of a stack $s$ and a heap $h$. The stack $s$ is a function giving values to a finite set of variables (location or data variables) while the heap $h$ is a function mapping a finite set of pairs $(\ell,{\it pf})$, where $\ell$ is a location and ${\it pf}$ is a pointer field, to locations, and a finite set of pairs $(\ell,{\it df})$, where ${\it df}$ is a data field, to values in ${{\mathbb D}}$. In addition, $h$ satisfies the condition that for each $\ell \in {{\mathbb L}}$, if $(\ell,{\it df}) \in {{\sf dom}}(h)$ for some ${\it df} \in {\mathcal{D}}$, then $(\ell,{\it pf}) \in {{\sf dom}}(h)$ for some ${\it pf} \in {\mathcal{F}}$. Let ${{\sf dom}}(h)$ denote the domain of $h$, and ${{\sf ldom}}(h)$ denote the set of $\ell \in {{\mathbb L}}$ such that $(\ell, {\it pf}) \in {{\sf dom}}(h)$ for some ${\it pf} \in {\mathcal{F}}$. $$\begin{array}{l} \begin{array}{ll} X,Y,E \in {{\sf LVar}}\mbox{ location variables} & \rho\ \subseteq\ ({\mathcal{F}}\times {{\sf LVar}})\cup ({\mathcal{D}}\times{{\sf DVar}}) \\[1mm] \vec{F}\in {\mathsf{Var}}^* \mbox{ vector of variables} & P \in {\mathcal{P}}\mbox{ predicates} \\[1mm] x\in{\mathsf{Var}}\mbox{ variable} \hspace{1cm}& \Delta\ \mbox{ formula over data variables} \end{array} \\[0.7cm] \Pi\ ::=\ X = Y \mid X \neq Y \mid \Delta\mid \Pi \land \Pi \hfill \mbox{pure formulas}\\[0.7mm] \Sigma\ ::=\ {\mathtt{emp}}\mid E \mapsto \rho \mid P(E,\vec{F}) \mid \Sigma {\ast}\Sigma \hspace{1cm}\hfill \mbox{spatial formulas} \\[1.2mm] \varphi\ ::=\ \Pi \land \Sigma\mid \varphi\vee \varphi\mid \exists x.\ \varphi \hfill \mbox{formulas} \end{array}$$ Formulas are conjunctions between a pure formula $\Pi$ and a spatial formula $\Sigma$. Pure formulas characterize the stack $s$ using (dis)equalities between location variables, e.g., a stack models $x=y$ iff $s(x)=s(y)$, and constraints $\Delta$ over data variables. We let $\Delta$ unspecified, though we assume that they belong to decidable theories, e.g., linear arithmetic or quantifier-free first order theories over multisets of values. The atom ${\mathtt{emp}}$ of spatial formulas holds iff the domain of the heap is empty. The points-to atom $E \mapsto \{ (f_i,x_i) \}_{i\in\mathcal{I}}$ specifies that the heap contains exactly one location $E$, and for all $i\in \mathcal{I}$, the field $f_i$ of $E$ equals $x_i$, i.e., $h(s(E),f_i)=s(x_i)$. The predicate atom $P(E,\vec{F})$ specifies a heap segment rooted at $E$ and shaped by the predicate $P$; the fragment is parameterized by a set ${\mathcal{P}}$ of inductively defined predicates, formally defined hereafter. Let $P \in {\mathcal{P}}$. An inductive definition of $P$ is a finite set of rules of the form $P(E,\vec{F})::=\exists \vec{Z}. \Pi \land \Sigma$, where $\vec{Z} \in {\mathsf{Var}}^\ast$ is a tuple of variables. A rule $R$ is called a base rule if $\Sigma$ contains no predicate atoms. Otherwise, it is called an inductive rule. A base rule $R$ is called spatial-empty if $\Sigma={\mathtt{emp}}$. Otherwise, it is called a spatial-nonempty base rule. For instance, the predicate ${\mathit{bst}}$ in Sec. \[sec:motivation\] is defined by one spatial-empty base rule and one inductive rule. We consider a class of restricted inductive definitions that are expressive enough to deal with intricate data structures (see Sec. \[sec:exp\]) while also enabling efficient proof strategies for establishing the validity of the verification conditions (see Sec. \[sec:slice\]). For each rule $R: P(E,\vec{F})::=\exists \vec{Z}. \Pi \land \Sigma$ in the definition of a predicate $P(E,\vec{F}) \in {\mathcal{P}}$, we assume that: - If $R$ is inductive, then $\Sigma = \Sigma_1 {\ast}\Sigma_2$ and the following conditions hold: - the root atoms: $\Sigma_1$ contains only points-to atoms and a unique points-to atom starting from $E$, denoted as $E \mapsto \rho$. Also, all the location variables from $\vec{Z}$ occur in $\Sigma_1$. $\Sigma_1$ is called the root of $R$ and denoted by $root(R)$. - connectedness: the Gaifman graph of $\Sigma_1$, denoted by $G_{\Sigma_1}$, is a connected DAG (directed acyclic graph) with the root $E$, that is, every vertex is reachable from $E$, - predicate atoms: $\Sigma_2$ contains only atoms of the form $Q(Z,\vec{Z'})$, and for each such atom, $Z$ is a vertex in $G_{\Sigma_1}$ without outgoing arcs. - If $R$ is a spatial-nonempty base rule, then $\Sigma$ contains exactly one points-to atom $E \mapsto \rho$, for some $\rho$. The classic acyclic list segment definition [@Rey02] satisfies these constraints as well as the first rule below; the second rule below falsifies the “root atoms” constraint: Since we disallow the use of negations on top of the spatial atoms, the semantics of the predicates in ${\mathcal{P}}$ is defined as usual as a least fixed-point. The class of inductive definitions defined above is in general undecidable, since with data fields, inductive definitions can be used to simulate two-counter machines. A variable substitution $\eta$ is a mapping from a finite subset of ${\mathsf{Var}}$ to the set of terms over the respective domains. For instance, if $X \in {{\sf LVar}}$ and $v,v_1 \in {{\sf DVar}}$ be integer variables then the mapping $\eta=\{X \rightarrow {{\sf nil}}, v \rightarrow v_1 + 5\}$ is a variable substitution. We denote by ${\mathtt{free}}(\psi)$ the set of free variables of a formula $\psi$. Composition Lemmas {#sec:comp} ================== As we have seen in the motivating example, the predicate ${\mathit{bsthole}}(E,M_1,F,M_2)$ satisfies the property that composing two heap structures described by this predicate results in a heap structure satisfying the same predicate. We call this property a composition lemma. We define simple and uniform syntactic criteria which, if they are satisfied by a predicate, then the composition lemma holds. The main idea is to divide the parameters of inductively defined predicates into three categories: The source parameters ${\vec{\alpha}}=(E,C)$, the hole parameters ${\vec{\beta}}=(F,H)$, and the static parameters ${\vec{\xi}}\in {\mathsf{Var}}^\ast$, where $E,F \in {{\sf LVar}}$ are called the source and resp., the hole location parameter, and $C,H \in {{\sf DVar}}$ are called the cumulative and resp., the hole data parameter[^3]. Let ${\mathcal{P}}$ be a set of inductively defined predicates and $P \in {\mathcal{P}}$ with the parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$. Then $P$ is said to be syntactically compositional if the inductive definition of $P$ contains exactly one base rule, and at least one inductive rule, and the rules of $P$ are of one of the following forms: - Base rule: $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)::= \alpha_{1} = \beta_{1} \land \alpha_2 = \beta_2 \land {\mathtt{emp}}$. Note that here the points-to atoms are disallowed. - Inductive rule: $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)::= \exists {\vec{Z}}. \ \Pi \land \Sigma$, with (a) $\Sigma \triangleq \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\,)$, (b) $\Sigma_1$ contains only and at least one points-to atoms, (c) $\Sigma_2$ contains only and possibly none predicate atoms, (d) $\vec{\gamma} \subseteq {\vec{Z}}$, and (d) the variables in ${\vec{\beta}}$ do not occur elsewhere in $\Pi \land \Sigma$, i.e., not in $\Pi$, or $\Sigma_1$, or $\Sigma_2$, or $\vec{\gamma}$. Note that the inductive rule also satisfies the constraints “root atom” and “connectedness” introduced in Sec. \[sec:logic\]. In addition, $\Sigma_2$ may contain $P$ atoms. One may easily check that both the predicate ${\mathit{lseg}}(E,M,F,M')$ in eq. (\[rule:lseg-base\])–(\[rule:lseg-rec\]) and the predicate ${\mathit{bsthole}}(E,M_1,F,M_2)$ in eq. (\[rule:bsth-base\])–(\[rule:bsth-recl\]) are syntactically compositional, while the predicate ${\mathit{lseg}}(E,M,F)$ in eq. (\[rule:lseg1-base\])–(\[rule:lseg1-rec\]) is not. A predicate $P \in {\mathcal{P}}$ with the parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$ is said to be semantically compositional if the entailment $\exists {\vec{\beta}}.\ P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\,) \Rightarrow P({\vec{\alpha}}, {\vec{\gamma}},{\vec{\xi}}\,)$ holds. \[thm-one-pred-compos\] Let ${\mathcal{P}}$ be a set of inductively defined predicates. If $P \in {\mathcal{P}}$ is syntactically compositional, then $P$ is semantically compositional. The proof of Thm. \[thm-one-pred-compos\] is done (see [@ESZ15]) by induction on the size of the domain of the heap structures. Suppose $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\,)$, then either $s({\vec{\alpha}})=s({\vec{\beta}})$ or $s({\vec{\alpha}}) \neq s({\vec{\beta}})$. If the former situation occurs, then $(s,h) \models P({\vec{\alpha}}, {\vec{\gamma}},{\vec{\xi}}\,)$ follows immediately. Otherwise, the predicate $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$ is unfolded by using some inductive rule of $P$, and the induction hypothesis can be applied to a sub-heap of smaller size. Then $(s,h) \models P({\vec{\alpha}}, {\vec{\gamma}},{\vec{\xi}}\,)$ can be deduced by utilizing the property that the hole parameters occur only once in each inductive rule of $P$. The syntactically compositional predicates are rather general in the sense that they allow nestings of predicates, branchings (e.g. trees), as well as data and size constraints. Therefore, composition lemmas can be obtained for complex data structures like nested lists, AVL trees, red-black trees, and so on. In addition, although lemmas have been widely used in the literature, we are not aware of any work that uses the composition lemmas as simple and elegant as those introduced above, when data and size constraints are included. Derived Lemmas {#sec:derived} ============== Theorem \[thm-one-pred-compos\] provides a mean to obtain lemmas for one single syntactically compositional predicate. In the following, based on the syntactic compositionality, we demonstrate how to derive additional lemmas describing relationships between different predicates (proofs are detailed in [@ESZ15]). We identify three categories of derived lemmas: “completion” lemmas, “stronger” lemmas, and “static-parameter contraction” lemmas. Based on our experiences in the experiments (cf. Sec. \[sec:exp\]) and the examples from the literature, we believe that the composition lemmas as well as the derived ones are natural, essential, and general enough for the verification of programs manipulating dynamic data structures. For instance, the “composition” lemmas and “completion” lemmas are widely used in our experiments, the “stronger” lemmas are used to check the verification conditions for rebalancing AVL trees and red-black trees. While “static parameter contraction” lemmas are not used in our experiments, they could also be useful, e.g., for the verification of programs manipulating lists with tail pointers. The “completion” lemmas ----------------------- We first consider the “completion” lemmas which describe relationships between incomplete data structures (e.g., binary search trees with one hole) and complete data structures (e.g., binary search trees). For example, the following lemma is valid for the predicates ${\mathit{bsthole}}$ and ${\mathit{bst}}$: Notice that the rules defining ${\mathit{bst}}(E,M)$ can be obtained from those of ${\mathit{bsthole}}(E_1,M_1,F,M_2)$ by applying the variable substitution $\eta=\{F \rightarrow {{\sf nil}}, M_2 \rightarrow \emptyset\}$ (modulo the variable renaming $M_1$ by $M$). This observation is essential to establish the “completion lemma” and it is generalized to arbitrary syntactically compositional predicates as follows. Let $P \in {\mathcal{P}}$ be a syntactically compositional predicate with the parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$, and $P' \in {\mathcal{P}}$ a predicate with the parameters $({\vec{\alpha}},{\vec{\xi}}\,)$. Then $P'$ is a completion of $P$ with respect to a pair of constants ${\vec{c}}=c_1c_2$, if the rules of $P'$ are obtained from the rules of $P$ by applying the variable substitution $\eta=\{\beta_1 \rightarrow c_1, \beta_2 \rightarrow c_2\}$. More precisely, - let $\alpha_1 = \beta_1 \land \alpha_2 = \beta_2 \land {\mathtt{emp}}$ be the base rule of $P$, then $P'$ contains only one base rule, that is, $\alpha_1 = c_1 \land \alpha_2 = c_2 \land {\mathtt{emp}}$, - the set of inductive rules of $P'$ is obtained from those of $P$ as follows: Let $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)::=\exists {\vec{Z}}.\ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\,)$ be an inductive rule of $P$, then $P'({\vec{\alpha}},{\vec{\xi}}\,)::=\exists {\vec{Z}}.\ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P'({\vec{\gamma}},{\vec{\xi}}\,)$ is an inductive rule of $P'$ (Recall that $\vec{\beta}$ does not occur in $\Pi,\Sigma_1,\Sigma_2,{\vec{\gamma}}$). \[thm-completion\] Let $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) \in {\mathcal{P}}$ be a syntactically compositional predicate, and $P'({\vec{\alpha}},{\vec{\xi}}\,) \in {\mathcal{P}}$. If $P'$ is a completion of $P$ with respect to ${\vec{c}}$, then $P'({\vec{\alpha}},{\vec{\xi}}\,) \Leftrightarrow P({\vec{\alpha}},{\vec{c}},{\vec{\xi}}\,)$ and $\exists{\vec{\beta}}.\ P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) {\ast}P'({\vec{\beta}},{\vec{\xi}}\,) \Rightarrow P'({\vec{\alpha}}, {\vec{\xi}}\,)$ hold. The “stronger” lemmas --------------------- We illustrate this class of lemmas on the example of binary search trees. Let ${\it natbsth}(E,M_1, F, M_2)$ be the predicate defined by the same rules as ${\mathit{bsthole}}(E,M_1,F,M_2)$ (i.e., eq. (\[rule:bsth-base\])–(\[rule:bsth-recl\])), except that $M_3 \ge 0$ ($M_3$ is an existential variable) is added to the body of each inductive rule (i.e., eq. (\[rule:bsth-recr\]) and (\[rule:bsth-recl\])). Then we say that ${\it natbsth}$ is stronger than ${\mathit{bsthole}}$, since for each rule $R'$ of ${\it natbsth}$, there is a rule $R$ of ${\mathit{bsthole}}$, such that the body of $R'$ entails the body of $R$. This “stronger” relation guarantees that the following lemmas hold: In general, for two syntactically compositional predicates $P,P' \in {\mathcal{P}}$ with the same set of parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$, $P'$ is said to be stronger than $P$ if for each inductive rule $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)::=\exists {\vec{Z}}. \ \Pi' \land \Sigma_1 {\ast}\Sigma_2 {\ast}P'({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\,)$, there is an inductive rule $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)::=\exists {\vec{Z}}. \ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\,)$ such that $\Pi' {\Rightarrow}\Pi$ holds. The following result is a consequence of Thm. \[thm-one-pred-compos\]. \[thm-strong\] Let $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,),P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) \in {\mathcal{P}}$ be two syntactically compositional predicates. If $P'$ is stronger than $P$, then the entailments $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,){\Rightarrow}P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,)$ and $\exists{\vec{\beta}}.\ P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\,) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\,) {\Rightarrow}P({\vec{\alpha}},{\vec{\gamma}},{\vec{\xi}}\,)$ hold. The “stronger” relation defined above requires that the spatial formulas in the inductive rules of $P$ and $P'$ are the same. This constraint can be relaxed by only requiring that the body of each inductive rule of $P'$ is stronger than a formula obtained by unfolding an inductive rule of $P$ for a bounded number of times. This relaxed constraint allows generating additional lemmas, e.g., the lemmas relating the predicates for list segments of even length and list segments. The “static-parameter contraction” lemmas ----------------------------------------- Let ${\it tailbsth}(E,M_1,F,M_2)$ (resp. ${\it stabsth}(E,M_1,F, M_2,B)$) be the predicate defined by the same rules as ${\mathit{bsthole}}(E,M_1,F,M_2)$, with the modification that the points-to atom in each inductive rule is replaced by $E \mapsto \{({\mathtt{left}},X),({\mathtt{right}},Y),({\tt tail},F),({\mathtt{data}},v)\}$ (resp. $E \mapsto \{({\mathtt{left}},X),({\mathtt{right}},Y),({\tt tail},B),({\mathtt{data}},v)\}$). Intuitively, ${\it tailbsth}$ (resp. ${\it stabsth}$) is obtained from ${\mathit{bsthole}}$ by adding a ${\tt tail}$ pointer to $F$ (resp. $B$). Then ${\it tailbsth}$ is not syntactically compositional since $F$ occurs in the points-to atoms of the inductive rules. On the other hand, ${\it stabsth}$ is syntactically compositional. From the above description, it is easy to observe that the inductive definition of ${\it tailbsth}(E, M_1, F, M_2)$ can be obtained from that of ${\it stabsth}(E, M_1, F, M_2, B)$ by replacing $B$ with $F$. Then the lemma ${\it tailbsth}(E, M_1, F, M_2) \Leftrightarrow {\it stabsth}(E, M_1, F, M_2, F)$ holds. From this, we further deduce the lemma We call the aforementioned replacement of $B$ by $F$ in the inductive definition of ${\it stabsth}$ as the “static-parameter contraction”. This idea can be generalized to arbitrary syntactically compositional predicates as follows. Let $P \in {\mathcal{P}}$ be a syntactically compositional predicate with the parameters $({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}}\,)$, $P' \in {\mathcal{P}}$ be an inductive predicate with the parameters $({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi'}}\,)$, ${\vec{\xi}}=\xi_1\dots \xi_k$, and ${\vec{\xi'}}=\xi'_1 \dots \xi'_l$. Then $P'$ is called a static-parameter contraction of $P$ if the rules of $P'$ are obtained from those of $P$ by a variable substitution $\eta$ s.t. ${{\sf dom}}(\eta)=\vec{\xi}$, for each $i: 1 \le i \le k$, either $\eta(\xi_i)=\xi_i$, or $\eta(\xi_i)=\beta_j$ for some $j=1,2$ satisfying that $\xi_i$ and $\beta_j$ have the same data type, and ${\vec{\xi'}}$ is the tuple obtained from $\eta(\vec{\xi}~)$ by removing the $\beta_j$’s. The substitution $\eta$ is called the contraction function. \[thm-border-contr\] Let $P({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}}\,) \in {\mathcal{P}}$ be a syntactically compositional predicate and $P'({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi'}}) \in {\mathcal{P}}$ be an inductive predicate. If $P'$ is a static-parameter contraction of $P$ with the contraction function $\eta$, then $P'({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi'}}) \Leftrightarrow P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ ))$ and $\exists{\vec{\beta}}.\ P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ )) {\ast}P'({\vec{\beta}}, {\vec{\gamma}}, {\vec{\xi'}}) \Rightarrow P'({\vec{\alpha}}, {\vec{\gamma}}, {\vec{\xi'}})$ hold. The lemmas presented in the last two sections are incomplete in the sense that they may not cover all the lemmas for a given set of inductive predicates. Although various extensions of the lemmas are possible, generating all the possible lemmas can be quite complex in general. Thm. 3 in [@IRV14] shows that generating all the lemmas is at least EXPTIME-hard, even for a fragment restricted to shape properties, without any data or size constraint. A Proof Strategy Based on Lemmas {#sec:slice} ================================ We introduce a proof strategy based on lemmas for proving entailments $\varphi_1{\Rightarrow}\exists\vec{X}.\varphi_2$, where $\varphi_1$, $\varphi_2$ are quantifier-free, and $\vec{X} \in {{\sf DVar}}^\ast$. The proof strategy treats uniformly the inductive rules defining predicates and the lemmas defined in Sec. \[sec:comp\]–\[sec:derived\]. Therefore, we call lemma also an inductive rule. W.l.o.g. we assume that $\varphi_1$ is quantifier-free (the existential variables can be skolemized). In addition, we assume that only data variables are quantified in the right-hand side[^4]. W.l.o.g., we assume that every variable in $\vec{X}$ occurs in at most one spatial atom of $\varphi_2$ (multiple occurrences of the same variable can be removed by introducing fresh variables and new equalities in the pure part). Also, we assume that $\varphi_1$ and $\varphi_2$ are of the form $\Pi\land\Sigma$. In the general case, our proof strategy checks that for every disjunct $\varphi_1'$ of $\varphi_1$, there is a disjunct $\varphi_2'$ of $\varphi_2$ s.t. $\varphi_1'{\Rightarrow}\exists \vec{X}.\varphi_2'$. We present the proof strategy as a set of rules in Fig. \[fig-pf-rule\]. For a variable substitution $\eta$ and a set $\mathcal{X} \subseteq {\mathsf{Var}}$, we denote by $\eta\|_{\mathcal{X}}$ the restriction of $\eta$ to $\mathcal{X}$. In addition, ${{\tt EQ}}(\eta)$ is the conjunction of the equalities $X=t$ for every $X$ and $t$ such that $\eta(X)=t$. Given two formulas $\varphi_1$ and $\varphi_2$, a substitution $\eta$ with ${{\sf dom}}(\eta)=\vec{X}$, the judgement $\varphi_1\models_{\eta} \exists\vec{X}.\varphi_2$ denotes that the entailment $\varphi_1 {\Rightarrow}\eta(\varphi_2)$ is valid. Therefore, $\eta$ provides an instantiation for the quantified variables $\vec{X}$ which witnesses the validity. The rules $\textsc{Match1}$ and $\textsc{Match2}$ consider a particular case of $\models_{\eta}$, denoted using the superscript ${\textit{SUB}}$, where the spatial atoms of $\varphi_2$ are syntactically matched[^5] to the spatial atoms of $\varphi_1$ modulo a variable substitution $\theta$. The substitution of the existential variables is recorded in $\eta$, while the substitution of the free variables generates a set of equalities that must be implied by $\Pi_1\land {{\tt EQ}}(\eta)$. For example, let $\Pi_1 \land \Sigma_1 ::= w=w'\land E\mapsto\{(f,Y),(d_1,v),(d_2,w)\}$, and $\exists \vec{X}.\ \Sigma_2 ::= \exists X,v'.\ E\mapsto\{(f,X),(d_1,v'),(d_2,w')\}$, where $d_1$ and $d_2$ are data fields. If $\theta=\{X \rightarrow Y, v' \rightarrow v, w' \rightarrow w\}$, then $\Sigma_1 = \theta(\Sigma_2)$. The substitution of the free variable $w'$ from the right-hand side is sound since the equality $w=w'$ occurs in the left-hand side. Therefore, $\Pi_1 \land \Sigma_1 \models^{SUB}_{\theta\|_{\{X,v'\}}} \exists X, v'.\ \Sigma_2$ holds. The rule $\textsc{Lemma}$ applies a lemma $L::= \exists \vec{Z}.\ \Pi \land root(L) {\ast}\Sigma {\Rightarrow}A$. It consists in proving that $\varphi_1$ implies the LHS of the lemma where the variables in $\vec{X}$ are existentially quantified, i.e., $\exists\vec{X}\exists \vec{Z}.\ \Pi \land root(L) {\ast}\Sigma$. Notice that $\vec{Z}$ may contain existential location variables. Finding suitable instantiations for these variables relies on the assumption that ${\it root}(L)$ in the LHS of $L$ is either a unique predicate atom or a separating conjunction of points-to atoms rooted at $E$ (the first parameter of $A$) and ${\it root}(L)$ includes all the location variables in $\vec{Z}$. This assumption holds for all the inductive rules defining predicates in our fragment (a consequence of the root and connectedness constraints) and for all the lemmas defined in Sec. \[sec:comp\]–\[sec:derived\]. The proof that $\varphi_1$ implies $\exists\vec{X}\exists \vec{Z}.\ \Pi \land root(L) {\ast}\Sigma$ is split into two sub-goals (i) proving that a sub-formula of $\varphi_1$ implies $\exists\vec{X}\exists \vec{Z}.\ root(L)$ and (ii) proving that a sub-formula of $\varphi_1$ implies $\exists\vec{X}\exists \vec{Z}.\ \Pi\land \Sigma$. The sub-goal (i) relies on syntactic matching using the rule $\textsc{Match1}$, which results in a quantifier instantiation $\eta_1$. The substitution $\eta_1$ is used to instantiate existential variables in $\exists \vec{X} \exists \vec{Z}.\ \Pi \land \Sigma$. Notice that according to the aforementioned assumption, the location variables in $\vec{Z}$ are not free in $\eta_1(\Pi \land \Sigma)$. Let $\eta_2$ be the quantifier instantiation obtained from the second sub-goal. The quantifier instantiation $\eta$ is defined as the extension of $\eta_1 \cup \eta_2$ to the domain $\vec{X} \cup \vec{Z}$ by utilizing the pure constraints $\Pi$ from the lemma[^6]. This extension is necessary since some existentially quantified variables may only occur in $\Pi$, but not in $root(L)$ nor in $\Sigma$, so they are not covered by $\eta_1 \cup \eta_2$. For instance, if $\Pi$ contains a conjunct $M=M_1 \cup M_2$ such that $M_1\in {{\sf dom}}(\eta_1)$, $M_2 \in {{\sf dom}}(\eta_2)$, and $M \not \in {{\sf dom}}(\eta_1 \cup \eta_2)$, then $\eta_1 \cup \eta_2$ is extended to $\eta$ where $\eta(M)=\eta_1(M_1) \cup \eta_2(M_2)$. The rule $\textsc{Slice}$ chooses a spatial atom $A$ in the RHS and generates two sub-goals: (i) one that matches $A$ (using the rules $\textsc{Match2}$ and $\textsc{Lemma}$) with a spatial sub-formula of the LHS ($\Sigma_1$) and (ii) another that checks that the remaining spatial part of the RHS is implied by the remaining part of the LHS. The quantifier instantiations $\eta_1$ and $\eta_2$ obtained from the two sub-goals are used to check that the pure constraints in the RHS are implied by the ones in LHS. Note that in the rule $\textsc{Slice}$, it is possible that $\Sigma_2=\Sigma={\mathtt{emp}}$. The rules in Fig. \[fig-pf-rule\] are applied in the order given in the figure. Note that they focus on disjoint cases w.r.t. the syntax of the RHS. The choice of the atom $A$ in $\textsc{Slice}$ is done arbitrary, since it does not affect the efficiency of proving validity. We apply the above proof strategy to the entailment $\varphi_1{\Rightarrow}\exists M.\ \varphi_2$ where: and ${\mathit{lseg}}$ has been defined in Sec. \[sec:intro\] (eq. (\[rule:lseg-base\])–(\[rule:lseg-rec\])). The entailment is valid because it states that two cells linked by ${\mathtt{next}}$ and storing ordered data values form a sorted list segment. The RHS $\varphi_2$ contains a single spatial atom and a pure part so the rule $\textsc{Slice}$ is applied and it generates the sub-goal $\varphi_1\models_{\eta}\exists M.\ {\mathit{lseg}}(x_1, M, {{\sf nil}},\emptyset)$ for which the syntactic matching (rule <span style="font-variant:small-caps;">Match1</span>) can not be applied. Instead, we apply the rule $\textsc{Lemma}$ using as lemma the inductive rule of ${\mathit{lseg}}$, i.e., eq. (\[rule:lseg-rec\]) (page ). We obtain the RHS $\exists M,X,M_1,v.\ x_1 \mapsto \{({\mathtt{next}},X), ({\mathtt{data}},v)\} {\ast}{\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset) \land M=\{v\}\cup M_1 \land v \le M_1$, where $x_1 \mapsto \{({\mathtt{next}},X), ({\mathtt{data}},v)\}$ is the root. The rule $\textsc{Match1}$ is applied with $\Pi_1 \land \Sigma_1 ::= x_1\neq{{\sf nil}}\land x_2\neq {{\sf nil}}\land v_1 < v_2\land x_1 \mapsto \{({\mathtt{next}},x_2),({\mathtt{data}},v_1)\}$ and it returns the substitution $\eta_1=\{X \rightarrow x_2, v \rightarrow v_1\}$. The second sub-goal is $\Pi_1 \land \Sigma_2 \models_{\eta_2} \exists M,M_1. \psi' $ where $\Pi_1 \land \Sigma_2 ::= x_1\neq{{\sf nil}}\land x_2\neq {{\sf nil}}\land v_1 < v_2 \land x_2 \mapsto \{({\mathtt{next}},{{\sf nil}}), ({\mathtt{data}},v_2)\}$ and $\psi' ::= M=\{v_1\} \cup M_1 \land v_1 \le M_1 \land {\mathit{lseg}}(x_2,M_1,{{\sf nil}},\emptyset)$. For this sub-goal, we apply the rule $\textsc{Slice}$, which generates a sub-goal where the rule $\textsc{Lemma}$ is applied first, using the same lemma, then the rule <span style="font-variant:small-caps;">Slice</span> is applied again, and finally the rule <span style="font-variant:small-caps;">Lemma</span> is applied with a lemma corresponding to the base rule of ${\mathit{lseg}}$, i.e., eq. (\[rule:lseg-base\]) (page ). This generates a quantifier instantiation $\eta_2=\{M \rightarrow \{v_1,v_2\}, M_1 \rightarrow \{v_2\}\}$. Then, $\eta_1 \cup \eta_2$ is extended with the constraints from the pure part of the lemma, i.e., $M=\{v\}\cup M_1 \land v_1 \le M_1$. Since $M \in {{\sf dom}}(\eta_1 \cup \eta_2)$, this extension has no effect. Finally, the rule $\textsc{Slice}$ checks that $\Pi_1 \land {{\tt EQ}}(\eta\|_{\{M\}}) \models \Pi_2$ holds, where ${{\tt EQ}}(\eta\|_{\{M\}}) ::= M = \{v_1,v_2\}$ and $\Pi_2 ::= v_2 \in M$. The last entailment holds, so the proof of validity is done. The following theorem states the correctness of the proof rules. Moreover, since we assume a finite set of lemmas, and every application of a lemma $L$ removes at least one spatial atom from $\varphi_1$ (the atoms matched to ${\it root}(L)$), the termination of the applications of the rule $\textsc{Lemma}$ is guaranteed. Let $\varphi_1$ and $\exists\vec{X}.\varphi_2$ be two formulas such that $\vec{X}$ contains only data variables. If $\varphi_1 \models_{\eta} \exists\vec{X}.\varphi_2$ for some $\eta$, then $\varphi_1{\Rightarrow}\exists\vec{X}.\varphi_2$. Experimental results {#sec:exp} ==================== We have extended the tool [<span style="font-variant:small-caps;">spen</span>]{} [@SPENsite] with the proof strategy proposed in this paper. The entailments are written in an extension of the SMTLIB format used in the competition SL-COMP’14 for separation logic solvers. It provides as output SAT, UNSAT or UNKNOWN, and a diagnosis for all these cases. The solver starts with a normalization step, based on the boolean abstractions described in [@EneaLSV14], which saturates the input formulas with (dis)equalities between location variables implied by the semantics of separating conjunction. The entailments of data constraints are translated into satisfiability problems in the theory of integers with uninterpreted functions, discharged using an SMT solver dealing with this theory. We have experimented the proposed approach on two sets of benchmarks[^7]: RDBI: : verification conditions for proving the correctness of iterative procedures (delete, insert, search) over recursive data structures storing integer data: sorted lists, binary search trees (BST), AVL trees, and red black trees (RBT). SL-COMP’14: : problems in the SL-COMP’14 benchmark, without data constraints, where the inductive definitions are syntactically compositional. \[tab:exp-data\] Tab. \[tab:exp-data\] provides the experiment results[^8] for RDBI. The column gives the number of verification conditions considered for each procedure. The column provides statistics about the lemma applications as follows: \#b and \#r are the number of the applications of the lemmas corresponding to base resp. inductive rules, \#c and \#d are the number of the applications of the composition resp. derived lemmas, and \#p is the number of predicates matched syntactically, without applying lemmas. Column gives the number of entailments between data constraints generated by [<span style="font-variant:small-caps;">spen</span>]{}. Column gives the “system” time spent by [<span style="font-variant:small-caps;">spen</span>]{} on all verification conditions of a function[^9] excepting the time taken to solve the data constraints by the SMT solver, which is given in the column . Tab. \[tab:exp-slcomp\] provides a comparison of our approach (column ) with the decision procedure in [@EneaLSV14] (column ) on the same set of benchmarks from SL-COMP’14. The times of the two decision procedures are almost the same, which demonstrates that our approach, as an extension of that in [@EneaLSV14], is robust. \[tab:exp-slcomp\] --------------------- ---- --------------------------- -------------------------------------------------------- ----------------------------------------------------------- (\#b, \#r, \#p, \#c, \#d) [<span style="font-variant:small-caps;">spen</span>]{} [<span style="font-variant:small-caps;">spen</span>]{}-TA Nested linked lists 16 (17,47,14,8,0) 4.428 4.382 Skip lists 2 levels 4 (11,16,1,1,0) 1.629 1.636 Skip lists 3 levels 10 (16,32,29,17,0) 3.858 3.485 --------------------- ---- --------------------------- -------------------------------------------------------- ----------------------------------------------------------- : Experimental results on benchmark SL-COMP’14 Related work ============ There have been many works on the verification of programs manipulating mutable data structures in general and the use of separation logic, e.g., [@AHJ+13; @AGH+14; @BCO05; @BDP11; @BPZ05; @CDN+12; @CHL11; @CHO+11; @EneaLSV14; @GVA07; @IBI+14; @IRS13; @IRV14; @NC08; @PWZ14; @RBH07; @ZKR08; @ChuJT14]. In the following, we discuss those which are closer to our approach. The prover SLEEK [@CDN+12; @NC08] provides proof strategies for proving entailments of SL formulas. These strategies are also based on lemmas, relating inductive definitions, but differently from our approach, these lemmas are supposed to be given by the user (SLEEK can prove the correctness of the lemmas once they are provided). Our approach is able to discover and synthesize the lemmas systematically, efficiently, and automatically. The natural proof approach DRYAD [@QGS+13; @Pek:2014:NPD:2666356.2594325] can prove automatically the correctness of programs against the specifications given by separation logic formulas with inductive definitions. Nevertheless, the lemmas are still supposed to be provided by the users in DRYAD, while our approach can generate the lemmas automatically. Moreover, DRYAD does not provide an independent solver to decide the entailment of separation logic formulas, which makes difficult to compare the performance of our tool with that of DRYAD. In addition, the inductive definitions used in our paper enable succinct lemmas, far less complex than those used in DRYAD, which include complex constraints on data variables and the magic wand. The method of cyclic proofs introduced by [@BDP11] and extended recently in [@ChuJT14] proves the entailment of two SL formulas by using induction on the paths of proof trees. They are not generating the lemma, but the method is able to (soundly) check intricate lemma given by the user, even ones which are out of the scope of our method, e.g., lemmas concerning the predicate $RList$ which is defined by unfolding the list segments from the end, instead of the beginning. The cyclic proofs method can be seen like a dynamic lemma generation using complex reasoning on proof trees, while our method generates lemma statically by simple checks on the inductive definitions. We think that our lemma generator could be used in the cyclic proof method to cut proof trees. The tool SLIDE [@IRS13; @IRV14] provides decision procedures for fragments of SL based on reductions to the language inclusion problem of tree automata. Their fragments contain no data or size constraints. In addition, the EXPTIME lower bound complexity is an important obstacle for scalability. Our previous work [@EneaLSV14] introduces a decision procedure based on reductions to the membership problem of tree automata which however is not capable of dealing with data constraints. The tool GRASShopper [@PWZ14] is based on translations of SL fragments to first-order logic with reachability predicates, and the use of SMT solvers to deal with the latter. The advantage is the integration with other SMT theories to reason about data. However, this approach considers a limited class of inductive definitions (for linked lists and trees) and is incapable of dealing with the size or multiset constraints, thus unable to reason about AVL or red-black trees. The truncation point approach [@GVA07] provides a method to specify and verify programs based on separation logic with inductive definitions that may specify truncated data structures with multiple holes, but it cannot deal with data constraints. Our approach can also be extended to cover such inductive definitions. Conclusion ========== We proposed a novel approach for automating program proofs based on Separation Logic with inductive definitions. This approach consists of (1) efficiently checkable syntactic criteria for recognizing inductive definitions that satisfy crucial lemmas in such proofs and (2) a novel proof strategy for applying these lemmas. The proof strategy relies on syntactic matching of spatial atoms and on SMT solvers for checking data constraints. We have implemented this approach in our solver [<span style="font-variant:small-caps;">spen</span>]{} and applied it successfully to a representative set of examples, coming from iterative procedures for binary search trees or lists. In the future, we plan to investigate extensions to more general inductive definitions by investigating ideas from [@QGS+13; @ChuJT14] to extend our proof strategy. From a practical point of view, apart from improving the implementation of our proof strategy, we plan to integrate it into the program analysis framework [Celia]{} [@celia]. Proofs in Sec. \[sec:comp\] =========================== [Theorem \[thm-one-pred-compos\]]{}. Suppose that ${\mathcal{P}}$ is a set of inductively defined predicates. If $P \in {\mathcal{P}}$ is syntactically compositional, then $P$ is semantically compositional. Suppose $P$ is syntactically compositional and has parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. It is sufficient to prove the following claim. > For each pair $(s,h)$, if $(s,h) \models P({\vec{\alpha}}_1,{\vec{\alpha}}_2,{\vec{\xi'}}) {\ast}P({\vec{\alpha}}_2,{\vec{\alpha}}_3,{\vec{\xi'}})$, then $(s,h) \models P({\vec{\alpha}}_1,{\vec{\alpha}}_3,{\vec{\xi'}})$. We prove the claim by induction on the size of ${{\sf ldom}}(h)$. Suppose for each $i: 1 \le i \le 3$, ${\vec{\alpha}}_i= E_i\ v_i$, where $E_i$ and $v_i$ are respectively location and data variables. Since $(s,h) \models P({\vec{\alpha}}_1,{\vec{\alpha}}_2,{\vec{\xi'}}) {\ast}P({\vec{\alpha}}_2,{\vec{\alpha}}_3,{\vec{\xi'}})$, there are $h_1,h_2$ such that $h=h_1 {\ast}h_2$, $(s,h_1) \models P({\vec{\alpha}}_1,{\vec{\alpha}}_2,{\vec{\xi'}})$, and $(s,h_2) \models P({\vec{\alpha}}_2,{\vec{\alpha}}_3,{\vec{\xi'}})$. If $(s,h_1) \models \bigwedge \limits_{i=1}^2 \alpha_{1,i} = \alpha_{2,i} \land {\mathtt{emp}}$, then ${{\sf ldom}}(h_1)=\emptyset$, and $h_2=h$. From this, we deduce that $(s,h) \models P({\vec{\alpha}}_1,{\vec{\alpha}}_3,{\vec{\xi'}})$. Otherwise, there are a recursive rule of $P$, say $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )::=\exists {\vec{X}}.\ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$, and an extension of $s$, say $s'$, such that $(s',h_1) \models \Pi' \land \Sigma'_1 {\ast}\Sigma'_2 {\ast}P({\vec{\gamma'}},{\vec{\alpha}}_2,{\vec{\xi'}})$, where $\Pi', \Sigma'_1, \Sigma'_2, \gamma'$ are obtained from $\Pi,\Sigma_1,\Sigma_2,\gamma$ by replacing ${\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}}$ with ${\vec{\alpha_1}}, {\vec{\alpha}}_2,{\vec{\xi'}}$ respectively. From this, we deduce that there are $h_{1,1},h_{1,2},h_{1,3}$ such that $h_1=h_{1,1} {\ast}h_{1,2} {\ast}h_{1,3}$, $(s',h_{1,1}) \models \Sigma'_1$, $(s',h_{1,2}) \models \Sigma'_2$, and $(s',h_{1,3}) \models P({\vec{\gamma'}},{\vec{\alpha}}_2,{\vec{\xi'}})$. Then $(s',h_{1,3} {\ast}h_2) \models P({\vec{\gamma'}},{\vec{\alpha}}_2,{\vec{\xi'}}) {\ast}P({\vec{\alpha}}_2,{\vec{\alpha}}_3, {\vec{\xi'}})$. From the induction hypothesis, we deduce that $(s',h_{1,3} {\ast}h_2) \models P({\vec{\gamma'}},{\vec{\alpha}}_3, {\vec{\xi'}})$. Then $(s',h_{1,1} {\ast}h_{1,2} {\ast}h_{1,3} {\ast}h_2) \models \Pi' \land \Sigma'_1 {\ast}\Sigma'_2 {\ast}P({\vec{\gamma'}},{\vec{\alpha}}_3, {\vec{\xi'}})$. We then deduce that $(s,h) \models \exists {\vec{X}}.\ \Pi' \land \Sigma'_1 {\ast}\Sigma'_2 {\ast}P({\vec{\gamma'}},{\vec{\alpha}}_3, {\vec{\xi'}})$. To prove $(s,h) \models P({\vec{\alpha_1}},{\vec{\alpha}}_3,{\vec{\xi'}})$, it is sufficient to prove that $(s,h) \models \exists {\vec{X}}.\ \Pi'' \land \Sigma''_1 {\ast}\Sigma''_2 {\ast}P({\vec{\gamma''}},{\vec{\alpha}}_3,{\vec{\xi'}})$, where $\Pi'', \Sigma''_1,\Sigma''_2,{\vec{\gamma''}}$ are obtained from $\Pi,\Sigma_1,\Sigma_2,{\vec{\gamma}}$ by replacing ${\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}}$ with ${\vec{\alpha}}_1, {\vec{\alpha}}_3, {\vec{\xi'}}$ respectively. From the fact that no variables from ${\vec{\beta}}$ occur in $\Pi$, $\Sigma_1$, $\Sigma_2$, or ${\vec{\gamma}}$, we know that $\Pi''=\Pi'$, $\Sigma''_1=\Sigma'_1$, $\Sigma''_2 = \Sigma'_2$, and ${\vec{\gamma''}}={\vec{\gamma'}}$. Since $(s,h) \models \exists {\vec{X}}.\ \Pi' \land \Sigma'_1 {\ast}\Sigma'_2 {\ast}P({\vec{\gamma'}},{\vec{\alpha}}_3, {\vec{\xi'}})$, we have already proved that $(s,h) \models \exists {\vec{X}}.\ \Pi'' \land \Sigma''_1 {\ast}\Sigma''_2 {\ast}P({\vec{\gamma''}},{\vec{\alpha}}_3,{\vec{\xi'}})$. The proof is done. Proofs in Sec. \[sec:derived\] ============================== [Theorem \[thm-completion\]]{}. Let $P \in {\mathcal{P}}$ be a syntactically compositional predicate with the parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}})$, and $P' \in {\mathcal{P}}$ with the parameters $({\vec{\alpha}},{\vec{\xi}}\ )$. If $P'$ is a completion of $P$ with respect to ${\vec{c}}$, then $P'({\vec{\alpha}},{\vec{\xi}}\ ) \Leftrightarrow P({\vec{\alpha}},{\vec{c}},{\vec{\xi}}\ )$ and $\exists \vec{\beta}. \ P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P'({\vec{\beta}},{\vec{\xi}}\ ) \Rightarrow P'({\vec{\alpha}}, {\vec{\xi}}\ )$ hold. The fact $P'({\vec{\alpha}},{\vec{\xi}}) \Leftrightarrow P({\vec{\alpha}},{\vec{c}},{\vec{\xi}})$ can be proved easily by an induction on the size of the domain of the heap structures. The argument for $\exists \vec{\beta}.\ P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}) {\ast}P'({\vec{\beta}},{\vec{\xi}}) \Rightarrow P'({\vec{\alpha}}, {\vec{\xi}})$ goes as follows: Suppose $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}) {\ast}P'({\vec{\beta}},{\vec{\xi}})$. Then there are $h_1,h_2$ such that $h=h_1 {\ast}h_2$, $(s,h_1) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}})$, and $(s,h_2) \models P'({\vec{\beta}},{\vec{\xi}})$. From the fact that $P'({\vec{\beta}},{\vec{\xi}}) \Leftrightarrow P({\vec{\beta}},{\vec{c}},{\vec{\xi}})$, we know that $(s,h_2) \models P({\vec{\beta}},{\vec{c}},{\vec{\xi}})$. Therefore, $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}) {\ast}P({\vec{\beta}},{\vec{c}}, {\vec{\xi}})$. From Theorem \[thm-one-pred-compos\], we deduce that $(s,h) \models P({\vec{\alpha}},{\vec{c}}, {\vec{\xi}})$. From the fact $P({\vec{\alpha}},{\vec{c}}, {\vec{\xi}}) \Leftrightarrow P'({\vec{\alpha}}, {\vec{\xi}})$, we conclude that $(s,h) \models P'({\vec{\alpha}}, {\vec{\xi}})$. [Theorem \[thm-strong\]]{}. Let $P,P' \in {\mathcal{P}}$ be two syntactically compositional inductively defined predicates with the same set of parameters $({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. If $P'$ is stronger than $P$, then the entailment $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )\Rightarrow P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$ and $\exists \vec{\beta}. \ P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\ ) \Rightarrow P({\vec{\alpha}},{\vec{\gamma}},{\vec{\xi}}\ )$ hold. We first show that $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )\Rightarrow P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. By induction on the size of ${{\sf ldom}}(h)$, we prove the following fact: For each $(s,h)$, if $(s,h) \models P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$, then $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. Suppose $(s,h) \models P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. If $(s,h) \models \bigwedge \limits_{i=1}^2 \alpha_i = \beta_i \land {\mathtt{emp}}$, since $P'$ and $P$ have the same base rule, we deduce that $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. Otherwise, there are a recursive rule of $P'$, say $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )::= \exists {\vec{X}}.\ \Pi' \land \Sigma_1 {\ast}\Sigma_2 {\ast}P'({\vec{\gamma}},{\vec{\beta}}, {\vec{\xi}}\ )$, and an extension of $s$, say $s'$, such that $(s',h) \models \Pi' \land \Sigma_1 {\ast}\Sigma_2 {\ast}P'({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$. Then there are $h_1,h_2,h_3$ such that $h=h_1 {\ast}h_2 {\ast}h_3$, $(s', h_1) \models \Sigma_1$, $(s', h_2) \models \Sigma_2$, and $(s', h_3) \models P'({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$. From the induction hypothesis, we deduce that $(s', h_3) \models P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$. Moreover, from the assumption, we know that there is a recursive rule of $P$ of the form $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )::= \exists {\vec{X}}.\ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}}, {\vec{\xi}}\ )$, such that $\Pi' \Rightarrow \Pi$ holds. Then it follows that $(s',h_1 {\ast}h_2 {\ast}h_3) \models \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$. We then deduce that $(s,h) \models \exists {\vec{X}}.\ \Pi \land \Sigma_1 {\ast}\Sigma_2 {\ast}P({\vec{\gamma}},{\vec{\beta}},{\vec{\xi}}\ )$. From this, we conclude that $(s,h) \models P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$. We then prove the second claim of the theorem. From the argument above, we know that $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) \Rightarrow P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ )$ holds. Then $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\ ) \Rightarrow P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\ )$ holds. In addition, from Theorem \[thm-one-pred-compos\], we know that $P({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\ ) \Rightarrow P({\vec{\alpha}},{\vec{\gamma}},{\vec{\xi}}\ )$ holds. Therefore,we conclude that $P'({\vec{\alpha}},{\vec{\beta}},{\vec{\xi}}\ ) {\ast}P({\vec{\beta}},{\vec{\gamma}},{\vec{\xi}}\ ) \Rightarrow P({\vec{\alpha}},{\vec{\gamma}},{\vec{\xi}}\ )$. [Theorem \[thm-border-contr\]]{}. Let $P \in {\mathcal{P}}$ be a syntactically compositional predicate with the parameters $({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}}\,)$ and $P' \in {\mathcal{P}}$ be an inductive predicate with the parameters $({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi'}})$. If $P'$ is a static-parameter contraction of $P$ with the contraction function $\eta$, then $P'({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi'}}) \Leftrightarrow P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ ))$ and $\exists{\vec{\beta}}.\ P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ )) {\ast}P'({\vec{\beta}}, {\vec{\gamma}}, {\vec{\xi'}}) \Rightarrow P'({\vec{\alpha}}, {\vec{\gamma}}, {\vec{\xi'}})$ hold. The first claim can be proved by induction on the size of the domain of the heap structures. The argument for the second claim goes as follows: From the fact that $P'({\vec{\beta}}, {\vec{\gamma}}, {\vec{\xi'}}) \Leftrightarrow P({\vec{\beta}}, {\vec{\gamma}}, \eta(\vec{\xi}\ ))$, we deduce that $$\exists{\vec{\beta}}.\ P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ )) {\ast}P'({\vec{\beta}}, {\vec{\gamma}}, {\vec{\xi'}}) \ \ \Rightarrow \ \ P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ )) {\ast}P({\vec{\beta}}, {\vec{\gamma}}, \eta(\vec{\xi}\ )). $$ From Theorem \[thm-one-pred-compos\], we know that $$P({\vec{\alpha}}, {\vec{\beta}}, \eta(\vec{\xi}\ )) {\ast}P({\vec{\beta}}, {\vec{\gamma}}, \eta(\vec{\xi}\ )) \Rightarrow P({\vec{\alpha}}, {\vec{\gamma}}, \eta(\vec{\xi}\ )).$$ Then the second claim follows from the fact $P({\vec{\alpha}}, {\vec{\gamma}}, \eta({\vec{\xi}}\ ) ) \Leftrightarrow P'({\vec{\alpha}}, {\vec{\gamma}}, {\vec{\xi'}})$. Extensions of the lemmas {#app:extensions} ======================== In this section, we discuss how the the basic idea of syntactical compositionality can be extended in various ways. Multiple location and data parameters ------------------------------------- At first, we would like to emphasize that although we restrict our discussions on compositional predicates $P({\vec{\alpha}}, {\vec{\beta}}, {\vec{\xi}})$ to the special case that ${\vec{\alpha}}$ (resp. ${\vec{\beta}}$) contain only two parameters: one location parameter, and one data parameter. But all the results about the lemmas can be generalized smoothly to the situation that ${\vec{\alpha}}$ and ${\vec{\beta}}$ contain multiple location and data parameters. Pseudo-composition lemmas ------------------------- We then consider syntactically pseudo-compositional predicates. We still use the binary search trees to illustrate the idea. Suppose $neqbsthole$ is the predicate defined by the same rules as $bsthole$, with the modification that $E \neq F$ is added to the body of each inductive rule. Then $\it neqbsthole$ is not syntactically compositional anymore and the composition lemma $$\begin{array}{l} \exists E_2,M_2. \ neqbsthole(E_1,M_1,E_2,M_2) {\ast}neqbsthole(E_2,M_2,E_3,M_3) \Rightarrow\\ \hspace{7cm} neqbsthole(E_1,M_1,E_3,M_3) \end{array}$$ does not hold. This is explained as follows: Suppose $h= h_1 {\ast}h_2$ (where $h=h_1 {\ast}h_2$ denotes that $h_1$ and $h_2$ are domain disjoint and $h$ is the union of $h_1$ and $h_2$), $(s,h_1) \models neqbsthole(E_1,M_1,E_2,M_2)$ and $(s,h_2) \models neqbsthole(E_2,M_2,E_3,M_3)$, in addition, both ${{\sf ldom}}(h_1)$ and ${{\sf ldom}}(h_2)$ are nonempty. Then from the inductive definition of $neqbsthole$, we deduce that $s(E_1) \neq s(E_2)$ and $s(E_2) \neq s(E_3)$. On the other hand, $(s,h) \models bsthole1(E_1,M_1,E_3,M_3)$ requires that $s(E_1) \neq s(E_3)$, which cannot be inferred from $s(E_1) \neq s(E_2)$ and $s(E_2) \neq s(E_3)$ in general. Nevertheless, the entailment $$\begin{array}{l} \exists E_2,M_2.\ neqbsthole(E_1,M_1,E_2,M_2) {\ast}neqbsthole(E_2,M_2,E_3,M_3) \ {\ast}\\ \hspace{1.5cm} E_3 \mapsto ((left,X),(right,Y),(data,v)) \Rightarrow \\ \hspace{1.5cm} neqbsthole(E_1,M_1,E_3,M_3) {\ast}E_3 \mapsto ((left,X),(right,Y),(data,v)) \end{array}$$ holds since the information $E_1 \neq E_3$ can be inferred from the fact that $E_3$ is allocated and separated from $E_1$. Therefore, intuitively, in this situation, the composition lemma can be applied under the condition that we already know that $E_1 \neq E_3$. We call this as pseudo-compositionality. Our decision procedure can be generalized to apply the pseudo-composition lemmas when proving the entailment of two formulas. Data structures with parent pointers ------------------------------------ Next, we show how our ideas can be generalized to the data structures with parent pointers, e.g. doubly linked lists or trees with parent pointers. We use binary search trees with parent pointers to illustrate the idea. We can define the predicates $prtbst(E,Pr,M)$ and $prtbsthole(E, Pr_1, M_1, F, Pr_2, M_2)$ to describe respectively binary search trees with parent pointers and binary search trees with parent pointers and one hole. The intuition of $E,F$ are still the source and the hole, while $Pr$ and $Pr_1$ (resp. $Pr_2$) are the parent of $E$ (resp. $F$) (the definition of $prtbst$ is omitted here). $$\begin{array}{l l} prtbsthole(E, Pr_1, M_1,F, Pr_2, M_2) &::= E = F \land {\mathtt{emp}}\land Pr_1 = Pr_2 \land M_1 = M_2 \\[1mm] prtbsthole(E, Pr_1, M_1,F, Pr_2, M_2) &::= \exists X,Y,M_3,M_4,v.\\ & \hspace{-1.5cm} \ E \mapsto \{(left,X), (right,Y),(parent,Pr_1), (data,v)\} \\ & \hspace{-1.5cm} {\ast}\ prtbst(X,E,M_3) {\ast}prtbsthole(Y,E, M_4, F, Pr_2, M_2) \\ & \hspace{-1.5cm} \land \ M_1=\{v\} \cup M_3 \cup M_4 \land M_3 < v < M_4 \\ [1mm] prtbsthole(E, Pr_1, M_1,F, Pr_2, M_2) &::= \exists X,Y,M_3,M_4,v.\\ & \hspace{-1.5cm} E \mapsto \{(left,X), (right,Y),(parent,Pr_1), (data,v)\}\\ & \hspace{-1.5cm} {\ast}\ prtbsthole(X,E,M_3, F, Pr_2, M_2) {\ast}prtbst(Y,E, M_4) \\ & \hspace{-1.5cm} \land \ M_1=\{v\} \cup M_3 \cup M_4 \land M_3 < v < M_4 \end{array}$$ Then the predicate $prtbsthole$ enjoys the composition lemma $$\begin{array} {l} \exists E_2, Pr_2, M_2. \ prtbsthole(E_1, Pr_1, M_1,E_2, Pr_2, M_2) \ {\ast}\\ \hspace{2.3cm} prtbsthole(E_2, Pr_2, M_2,E_3, Pr_3, M_3) \Rightarrow \\ \hspace{6cm} prtbsthole(E_1, Pr_1, M_1,E_3, Pr_3, M_3). \end{array}$$ Points-to atom in base rules ---------------------------- Finally, we discuss the constraint that the base rule of a syntactically compositional predicate has an empty spatial atom. We use the predicates $lsegeven$ and $lsegodd$ to illustrate the idea. $$\begin{array}{ll} lsegeven(E,F)& ::= E=F \land {\mathtt{emp}},\\[1mm] lsegeven(E,F) & ::= \exists X,Y. \ E \mapsto (next,X) {\ast}X \mapsto (next,Y) {\ast}lsegeven(Y,F). \end{array}$$ The definition of $lsegodd(E,F)$ can be obtained from that of $lsegeven(E,F)$ by replacing the base rule with the rule $lsegodd(E,F) ::= E \mapsto (next,F)$. The only difference between the inductive definition of $lsegeven$ and and that of $lsegodd$ is that $lsegeven$ has an empty base rule, while $lsegodd$ does not. From this, we deduce that $$lsegodd(E,F) \Leftrightarrow \exists X.\ E \mapsto \{(next,X)\} {\ast}lsegeven(X,F).$$ This idea can be generalized to arbitrary syntactically compositional predicates. Full example of Sec. \[sec:slice\] {#sec:app-ex} ================================== We provide here the full details of the example considered in Section \[sec:slice\]. Consider the following entailment which states that two cells linked by the ${\mathtt{next}}$ pointer field, and storing ordered data values, form a sorted list segment: where ${\mathit{lseg}}$ has been defined in Sec. \[sec:intro\] (eq. (\[rule:lseg-base\])–(\[rule:lseg-rec\])). For convenience, let $$\begin{aligned} \Pi_1 & ::= x_1\neq{{\sf nil}}\land x_2\neq {{\sf nil}}\land v_1 < v_2, \\ \Sigma_1 & ::= x_1 \mapsto \{({\mathtt{next}},x_2),({\mathtt{data}},v_1)\} {\ast}x_2 \mapsto \{({\mathtt{next}},{{\sf nil}}), ({\mathtt{data}},v_2)\},\\ \Pi_2 & ::= v_2\in M, \\ \Sigma_2 & ::= {\mathit{lseg}}(x_1, M, {{\sf nil}},\emptyset).\end{aligned}$$ [The first application of the rule]{} <span style="font-variant:small-caps;">Slice</span>. Since the right-hand side contains a single spatial atom, the rule $\textsc{Slice}$ generates a sub-goal $\Pi_1 \land \Sigma_1 \models_{\eta} \exists M.\ \Sigma_2$. For the sub-goal, the syntactic matching (rule <span style="font-variant:small-caps;">Match1</span>) cannot be applied. Instead, we apply the rule $\textsc{Lemma}$ using a lemma $L$ that corresponds to the inductive rule of ${\mathit{lseg}}$, i.e., eq. (\[rule:lseg-rec\]) (page ): $$\begin{array}{l} L::=\exists X,M_1,v.\ x_1 \mapsto \{({\mathtt{next}},X), ({\mathtt{data}},v)\} {\ast}{\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset) \land \\ \hspace{4.5cm} M=\{v\}\cup M_1 \land v \le M_1 {\Rightarrow}{\mathit{lseg}}(x_1, M, {{\sf nil}},\emptyset). \end{array}$$ For convenience, let $$\begin{aligned} \Pi & ::= M=\{v\}\cup M_1 \land v \le M_1, \\ \Sigma & ::= x_1 \mapsto \{({\mathtt{next}},X), ({\mathtt{data}},v)\} {\ast}{\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset). \end{aligned}$$ [The first application of the rule]{} <span style="font-variant:small-caps;">Lemma</span>. Since $root(L)::=x_1 \mapsto \{({\mathtt{next}},X), ({\mathtt{data}},v)\}$, the rule <span style="font-variant:small-caps;">Lemma</span> generates a sub-goal $\Pi_1 \land \Sigma'_1 \models^{SUB}_{\eta_1} \exists X,v.\ root(L)$, where $\Sigma'_1::=x_1 \mapsto \{({\mathtt{next}},x_2),({\mathtt{data}},v_1)\}$. Then the rule $\textsc{Match1}$ is applied, resulting in a quantifier instantiation $\eta_1=\{X \rightarrow x_2, v \rightarrow v_1\}$. Note that, since ${{\tt EQ}}(\eta_1\|_{{\mathtt{free}}(\exists X,v. root(L))})::= {\ensuremath{\mathtt{true}}\xspace}$, the entailment $\Pi_1 \land {{\tt EQ}}(\eta_1\|_{\{X,v\}}) \models {{\tt EQ}}(\eta_1\|_{{\mathtt{free}}(\exists X,v. root(L))})$ holds. The variable substitution $\eta_1$ is used to instantiate the existentially quantified variables in the remaining part of the lemma, that is, $\Pi \land {\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset)$, resulting into the formula $$\begin{array}{l c l} \eta_1(\Pi \land {\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset)) & ::= & \\ & & \hspace{-2.5cm} M=\{v_1\} \cup M_1 \land v_1 \le M_1 \land {\mathit{lseg}}(x_2,M_1,{{\sf nil}},\emptyset). \end{array}$$ Then, the rule $\textsc{Lemma}$ generates another sub-goal $\Pi_1 \land \Sigma''_1 \models_{\eta_2} \exists M,M_1.\ \eta_1(\Pi \land {\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset))$, where $\Sigma''_1 ::= x_2 \mapsto \{({\mathtt{next}},{{\sf nil}}), ({\mathtt{data}},v_2)\}$. [The second application of the rule]{} <span style="font-variant:small-caps;">Slice</span>. For the sub-goal $\Pi_1 \land \Sigma''_1 \models_{\eta_2} \exists M,M_1.\ \eta_1(\Pi \land {\mathit{lseg}}(X,M_1,{{\sf nil}},\emptyset))$, the rule $\textsc{Slice}$ is applied again. Since there is a single spatial atom in the RHS, the rule $\textsc{Slice}$ generates a sub-goal $\Pi_1 \land \Sigma''_1 \models_{\eta_3} \exists M_1.\ {\mathit{lseg}}(x_2,M_1,{{\sf nil}},\emptyset)$. [The second application of the rule]{} <span style="font-variant:small-caps;">Lemma</span>. For the sub-goal $\Pi_1 \land \Sigma''_1 \models_{\eta_3} \exists M_1.\ {\mathit{lseg}}(x_2,M_1,{{\sf nil}},\emptyset)$, the rule <span style="font-variant:small-caps;">Lemma</span> is applied again, using the lemma $L'$ (still corresponding to the inductive rule of ${\mathit{lseg}}$), $$\begin{array}{l} L'::=\exists X',M'_1,v'.\ x_2 \mapsto \{({\mathtt{next}},X'), ({\mathtt{data}},v')\} {\ast}{\mathit{lseg}}(X', M'_1,{{\sf nil}},\emptyset) \land \\ \hspace{4.5cm} M_1=\{v'\}\cup M'_1 \land v' \le M'_1 {\Rightarrow}{\mathit{lseg}}(x_2, M_1, {{\sf nil}},\emptyset). \end{array}$$ For convenience, let $$\begin{aligned} \Pi' & ::= M_1=\{v'\}\cup M'_1 \land v' \le M'_1, \\ \Sigma' & ::= x_2 \mapsto \{({\mathtt{next}},X'), ({\mathtt{data}},v')\} {\ast}{\mathit{lseg}}(X', M'_1,{{\sf nil}},\emptyset). \end{aligned}$$ Since $root(L')::= x_2 \mapsto \{({\mathtt{next}},X'), ({\mathtt{data}},v')\}$, the rule <span style="font-variant:small-caps;">Lemma</span> generates a sub-goal $\Pi_1 \land \Sigma''_1 \models^{SUB}_{\eta'_1} \exists X',v'.\ root(L')$. Then the rule $\textsc{Match1}$ is applied, resulting in a quantifier instantiation $\eta'_1=\{X' \rightarrow {{\sf nil}}, v' \rightarrow v_2\}$. Note that, since ${{\tt EQ}}(\eta'_1\|_{{\mathtt{free}}(\exists X',v'. root(L'))})::= {\ensuremath{\mathtt{true}}\xspace}$, the entailment $\Pi_1 \land {{\tt EQ}}(\eta'_1\|_{\{X',v'\}}) \models {{\tt EQ}}(\eta'_1\|_{{\mathtt{free}}(\exists X',v'. root(L'))})$ holds. The variable substitution $\eta'_1$ is used to instantiate the existentially quantified variables in the remaining part of the lemma, that is, $\Pi' \land {\mathit{lseg}}(X',M'_1,{{\sf nil}},\emptyset)$, resulting into the formula $$\begin{array}{l c l} \eta'_1(\Pi' \land {\mathit{lseg}}(X',M'_1,{{\sf nil}},\emptyset)) & ::= & \\ & & \hspace{-1.5cm} M_1=\{v_2\}\cup M'_1 \land v_2 \le M'_1 \land {\mathit{lseg}}({{\sf nil}}, M'_1,{{\sf nil}},\emptyset). \end{array}$$ Then, the rule $\textsc{Lemma}$ generates another sub-goal $\Pi_1 \land {\mathtt{emp}}\models_{\eta'_2} \exists M_1, M'_1.\ \eta'_1(\Pi' \land {\mathit{lseg}}(X',M'_1,{{\sf nil}},\emptyset))$. [The third application of the rule]{} <span style="font-variant:small-caps;">Slice</span>. For the sub-goal $\Pi_1 \land {\mathtt{emp}}\models_{\eta'_2} \exists M_1,M'_1.\ \eta'_1(\Pi' \land {\mathit{lseg}}(X',M'_1,{{\sf nil}},\emptyset))$, since there is only one spatial atom ${\mathit{lseg}}({{\sf nil}}, M'_1,{{\sf nil}},\emptyset)$ in the RHS, the rule <span style="font-variant:small-caps;">Slice</span> generates a subgoal $\Pi_1 \land {\mathtt{emp}}\models_ {\eta'_3} \exists M'_1.\ {\mathit{lseg}}({{\sf nil}}, M'_1,{{\sf nil}},\emptyset)$, for which the rule <span style="font-variant:small-caps;">Lemma</span> is applied, with a lemma $L''$ corresponding to the base rule of ${\mathit{lseg}}$, i.e., eq. (\[rule:lseg-base\]) (page ), $$L'' \hspace{2mm} ::=\hspace{2mm} {{\sf nil}}= {{\sf nil}}\land {\mathtt{emp}}\land M'_1= \emptyset \hspace{2mm} \Rightarrow \hspace{2mm} {\mathit{lseg}}({{\sf nil}}, M'_1,{{\sf nil}},\emptyset).$$ [The third application of the rule]{} <span style="font-variant:small-caps;">Lemma</span>. Since $root(L'')::= {\mathtt{emp}}$, the rule <span style="font-variant:small-caps;">Lemma</span> generates a sub-goal $\Pi_1 \land {\mathtt{emp}}\models^{SUB}_{\eta''_1} root(L'')$. Then the rule $\textsc{Match1}$ is applied, resulting in a quantifier instantiation $\eta''_1=\emptyset$. Note that, since ${{\tt EQ}}(\eta''_1\|_{{\mathtt{free}}(root(L''))})::= {\ensuremath{\mathtt{true}}\xspace}$, the entailment $\Pi_1 \land {{\tt EQ}}(\eta''_1\|_{\emptyset}) \models {{\tt EQ}}(\eta''_1\|_{{\mathtt{free}}(root(L''))})$ holds. The substitution $\eta''_1$ is used to instantiate the existential variables in the remaining part of the lemma, that is, ${{\sf nil}}= {{\sf nil}}\land M'_1= \emptyset$, resulting into the same formula. Then the rule <span style="font-variant:small-caps;">Lemma</span> generates a sub-goal $\Pi_1 \models_{\eta''_2} \exists M'_1.\ {{\sf nil}}= {{\sf nil}}\land M'_1= \emptyset$, which holds clearly with $\eta''_2=\emptyset$. Finally, $\eta''_1 \cup \eta''_2$ is extended with ${{\sf nil}}= {{\sf nil}}\land M'_1= \emptyset$, resulting into $\eta'_3=\{M'_1 \rightarrow \emptyset\}$. [The third application of the rule]{} <span style="font-variant:small-caps;">Slice</span> ([continued]{}). The variable substitution $\eta'_3$ is extended with $M_1=\{v_2\}\cup M'_1 \land v_2 \le M'_1$, resulting into $\eta'_2=\{M_1 \rightarrow \{v_2\} \cup \emptyset,\ M'_1 \rightarrow \emptyset\}$. Then the rule <span style="font-variant:small-caps;">Slice</span> generates a sub-goal $\Pi_1 \land {{\tt EQ}}(\eta'_2) \models M_1=\{v_2\}\cup M'_1 \land v_2 \le M'_1$. Because ${{\tt EQ}}(\eta'_2)=M_1 = \{v_2\} \cup \emptyset \land M'_1 = \emptyset$, we know that ${{\tt EQ}}(\eta'_2) \models M_1=\{v_2\}\cup M'_1 \land v_2 \le M'_1$. Thus the sub-goal holds. [The second application of the rule]{} <span style="font-variant:small-caps;">Lemma</span> ([continued]{}). The variable substitution $\eta'_1 \cup \eta'_2$ should be extended with $\Pi'::=M_1=\{v'\} \cup M'_1 \land v' \le M'_1$. Since $M_1 \in {{\sf dom}}(\eta'_2)$, the extension makes no effect. Then $\eta_3=(\eta'_1 \cup \eta'_2)\|_{\{M_1\}}=\{M_1 \rightarrow \{v_2\} \cup \emptyset\}$. [The second application of the rule]{} <span style="font-variant:small-caps;">Slice</span> ([continued]{}). The variable substitution $\eta_3$ is extended with $M=\{v_1\} \cup M_1 \land v_1 \le M_1$, resulting into $\eta_2=\{M \rightarrow \{v_1\} \cup \{v_2\} \cup \emptyset, M_1 \rightarrow \{v_2\} \cup \emptyset\}$. Then the rule <span style="font-variant:small-caps;">Slice</span> generates a sub-goal $\Pi_1 \land {{\tt EQ}}(\eta_2) \models M=\{v_1\} \cup M_1 \land v_1 \le M_1$. Since $\Pi_1::=x_1\neq{{\sf nil}}\land x_2\neq {{\sf nil}}\land v_1 < v_2$ and ${{\tt EQ}}(\eta_2)::=M = \{v_1\} \cup \{v_2\} \cup \emptyset \land M_1 = \{v_2\} \cup \emptyset$, the sub-goal holds. [The first application of the rule]{} <span style="font-variant:small-caps;">Lemma</span> ([continued]{}). The variable substitution $\eta_1 \cup \eta_2$ should be extended with $\Pi::=M=\{v\} \cup M_1 \land v \le M_1$. Since $M \in {{\sf dom}}(\eta_1 \cup \eta_2)$, this extension makes no effect. Then $\eta$ is obtained from $\eta_1 \cup \eta_2$ by restricting to $\{M\}$. So $\eta=\{M \rightarrow \{v_1\} \cup \{v_2\} \cup \emptyset\}$. [The first application of the rule]{} <span style="font-variant:small-caps;">Slice</span> ([continued]{}). The variable substitution $\eta$ is extended with $\Pi_2::=v_2 \in M$, and still getting $\eta$. Finally, <span style="font-variant:small-caps;">Slice</span> generates the sub-goal $\Pi_1 \land {{\tt EQ}}(\eta) \models \Pi_2$. Since ${{\tt EQ}}(\eta)::=M = \{v_1\} \cup \{v_2\} \cup \emptyset$, the entailment ${{\tt EQ}}(\eta) \models \Pi_2$ holds. Therefore, the sub-goal $\Pi_1 \land {{\tt EQ}}(\eta) \models \Pi_2$ holds as well. [^1]: Zhilin Wu is supported by the NSFC projects (No. 61100062, 61272135, and 61472474), and the visiting researcher program of China Scholarship Council. This work was supported by the ANR project Vecolib (ANR-14-CE28-0018). [^2]: The existential quantifiers in $\varphi_1$ are removed using skolemization. [^3]: For simplicity, we assume that ${\vec{\alpha}}$ and ${\vec{\beta}}$ consist of exactly one location parameter and one data parameter. [^4]: We believe that this restriction is reasonable for the verification conditions appearing in practice and all the benchmarks in our experiments are of this form. [^5]: In this case, the right-hand side contains no pure constraints. [^6]: The extension depends on the pure constraints $\Pi$ and could be quite complex in general. In the experiments of Sec. \[sec:exp\], we use the extension obtained by the propagation of equalities in $\Pi$. [^7]: <http://www.liafa.univ-paris-diderot.fr/spen/benchmarks.html> [^8]: The evaluations used a 2.53GHz Intel processor with 2GB, running Linux on VBox. [^9]: [<span style="font-variant:small-caps;">spen</span>]{} does not implement a batch mode, each entailment is dealt separately, including the generation of lemma. The SMT solver is called on the files generated by [<span style="font-variant:small-caps;">spen</span>]{}.	{ "pile_set_name": "ArXiv" }
=22.5cm 0.3cm -0.5cm -0.3cm [HIP-1999-09/TH]{} [New Magnetically Ordered State in the Electron Plasma of Metals]{}\ $^{)}$\ Division of Theoretical Physics, Physics Department\ P. O. Box 9, FIN-00014 University of Helsinki, Finland\ and\ $^{\dagger) }$\ Helsinki Institute of Physics, P. O. Box 9, FIN-00014 University of Helsinki, Finland\ \ Within the simplest model of metals, namely a gas of electrons with Coulomb interactions, in the presence of a uniform background of positive charge to enforce electric neutrality of the system, we have derived a mechanism, by which the Coulomb interaction between the electrons generates a new kind of magnetism. The ground state of the metal is represented by a magnetically ordered state described by a non-local magnetic field. This non-local magnetic field does not produce spin polarisation of electrons, but induces a special long range correlation between electrons of opposite spin. This mechanism results in a theoretical value for the binding energy per electron, which is more than twice the corresponding value for the unmagnetised state of the metal. The new magnetic order proposed and analysed theoretically here, can in principle be experimentally tested. $^{)}$ e-mail: [email protected]\ $^{\dagger)}$ e-mail: [email protected].\ Permanent address: Department of Theoretical Physics, Comenius University, Mlynska Dolina\ 84215 Bratislava and Department of Physics, Faculty of Natural Sciences at the University of Constantinus Philosopher in Nitra, Slovak Republic. Introduction ============ In considering many body systems, especially condensed matter, using Quantum Field Theory, it is frequently advantageous to use functional integral methods rather than canonical operator methods. A particular benefit of functional integral methods is the possibility to eliminate certain fermionic functional integrals by means of a Grasmannian change of integration variables, and to obtain in stead bosonic integrals involving new bosonic field variables. These bosonic fields can then be associated with existing macroscopic structures or phases appearing in interacting fermionic systems. To explain this in principle, consider a system of interacting fermions described by a given action $S_{f}[\psi^{}, \psi]$, where $\psi^{}(\tau)$ and $\psi(\tau)$ are anticommuting (Grassmann-odd) variables, and $\tau$ is a parameter in the interval $(0, \beta)$, with $\beta$ the inverse temperature. All the termodynamical properties of the system in its equilibrium state are obtained from the partition function $Z$, which is expressed as a functional integral as follows, $$Z = \int D(\psi^{}, \psi) \exp \left \{ - S_{f}[\psi^{}, \psi] \right \}, \label{eq:partfZ}$$ where $D(\psi^{}, \psi)$ is a standard measure over Grassmann variables [@Vasiliev], [@Kleinert], [@Popov]. Making a change of integration variables in (\[eq:partfZ\]) known as the Hubbard-Stratonovich transformation, or more generally a change of variables which introduces Grassmann-even, i.e. bosonic variables $B(\tau)$, one obtains a new functional integral representation [@Kleinert], [@Popov] of the partition function (\[eq:partfZ\]) $$Z = \int D(B, \psi^{}, \psi) \exp \left \{ - S_{bf}[B, \psi^{}, \psi] \right \}. \label{eq:HSpartfZ}$$ The second representation (\[eq:HSpartfZ\]) is exact but not nesessarily useful or interesting, unless one manages to choose the bosonic field variables in some particularly convenient way. This depends naturally on the functional form of the original fermionic action $S_{f}[\psi^{}, \psi]$. There may exist several different actions $S_{bf}[B, \psi^{}, \psi]$, corresponding to the same $S_{f}[\psi^{}, \psi]$. Under certain circumstances one may carry out the integration over the fermionic variables in (\[eq:HSpartfZ\]) explicitly and exactly, obtaining an expression involving a bosonic field only, $$Z = \int D(B) \exp \left \{ - S_{eff}[B] \right \}, \label{eq:Zeff}$$ where $S_{eff}[B]$ is an effective action functional involving only bosonic variables $B$. The functional integral (\[eq:Zeff\]) can in general not be carried out explicitly, but may be very convenient as a a starting point for approximate calculations. We will presently use the machinery presented in general terms so far, to analyse the simplest model of metals, namely an electron gas with Coulomb interaction between electrons, in the presence of a uniform background of positive charge, which is used in order to enforce the electric neutrality of the system. However it is convenient to discuss the approximate evaluation of the functional integral (\[eq:Zeff\]) in general terms before considering detailed calculations. The approximate evaluation of (\[eq:Zeff\]) goes in general as follows. One expands the action $S_{eff}[B]$ as a power series in the fluctuations ${\cal B} = B - \overline{B}$ around its stationary value $S_{eff}[\overline{B}]$. The mean field $\overline{B}$ is a solution to the variational Euler-Lagrange equations, $$\delta \,S_{eff}[B] = 0. \label{eq:EuLa}$$ Writing $$S_{eff}[B] = S_{eff}[\overline{B}] + S_{fl}[{\cal B}] \label{eq:sadl1}$$ one can then write, $$Z = Z_{0} < \exp \left \{ - S_{fl}[{\cal B}] \right \}>_{0}\;, Z_{0} = \exp \left \{-S_{eff}[\overline{B}]\right \}, \label{eq:sadl2}$$ where the symbol $<...>_{0}$ stands for average value with respect to $S_{eff}[\overline{B}]$. In this method, the saddle point contribution is treated exactly, while the effects of the fluctuations around $\overline{B}$, are treated perturbatively, to the desired accuracy. If the replacement of a fermionic integral with a bosonic one is succesfully carried out along the lines indicated above, then the bosonic fields $\overline{B}$ ought to describe macroscopic structures, or phases, of the interacting fermionic system, such as ferromagnetism, superconductivity and so on, for interacting electron systems. The effects of fluctuations $$< \exp \left \{ - S_{fl}[{\cal B}] \right \}>_{0} \label{eq:fluct2}$$ ought to give rise to small perturbative corrections only, to the underlying structure of the system described by the field $\overline{B}$. If one can ignore effects of fluctuations, then the partition function $Z$ of the interacting fermion system is approximated by $Z_{0}$, and one says that the system is described in the mean field approximation represented by the field $\overline{B}$. It was already mentioned above, that in general there may exist several realisations of fields $B$ by means of which one transforms the functional integral (\[eq:partfZ\]) into the equivalent form (\[eq:HSpartfZ\]). Hence, in the same fermionic system, one may in principle obtain several macroscopic structures, or phases, described by a set of mean filds $\overline{B}^{(i)}$, where the index $(i)$ enumerates the different macroscopic structures that can appear in the considered fermionic system, under given external conditions. With each mean field $\overline{B}^{(i)}$ one then associates a corresponding perturbation expansion, given by the relation (\[eq:sadl2\]), for $\overline{B} = \overline{B}^{(i)}$. If there are several mean fields $\overline{B}^{(i)}, i = 1,2,...$ one may ask which of these is actually realised in the system at hand. There is a nearly obvious answer to this; the structure $\overline{B}^{(i)}$ associated with the lowest energy under given external conditions is realised. It may of course happen that the external conditions do not lead to a unique structure $\overline{B}^{(i)}$, in which case we have a coexistence of macroscopic phases. The simplest model of metals ============================ Let us now return to the physical problem of interest, namely the problem of the ground state in the simplest model of metals [@Gell-Mann], which we analyse using the general method outlined above. The problem in question is ancient [@Bloch], [@Fetter-W], and concerns more precisely the question of whether the genuine ground state of the metal is an unmagnetised state or a ferromagnetic state. The model in question consists of a large number $N$ of electrons (an electron gas) in a finite but large volume V, interacting through the ordinary Coulomb interaction, in the presence of a uniform background of positive charge determined so that the system is electrically neutral. The system is described by the following Hamiltonian, $$H = H_{0} + H_{int} - C, \label{eq:Hammod}$$ where $$H_{0} = \int_{V} d^{3}{\bf x}\sum_{\alpha} \left [ - \frac{\hbar^{2}}{2m}\psi^{\dagger}_{\alpha}({\bf x})\nabla^{2}\psi_{\alpha}({\bf x}) \right ] \label{eq:Hzero}$$ describes the free electron gas, and $$H_{int} = \frac{e^{2}}{2}\sum_{\alpha, \alpha'}\int_{V}\int_{V} \frac{ d^{3}{\bf x} d^{3}{\bf x}'} {\mid {\bf x} - {\bf x}' \mid } \left \{\psi^{\dagger}_{\alpha'}({\bf x}')\psi^{\dagger}_{\alpha}({\bf x})\psi_{\alpha}({\bf x})\psi_{\alpha'}({\bf x}') \right \} \label{eq:Hint}$$ describes the Coulomb interaction between the electrons. Here the quantities $\psi^{\dagger}_{\alpha}$ and $\psi_{\alpha}$ are the creation and annihilation operators, respectively, of electrons with spin component $\alpha$. Finally, the constant $C$ in the expression (\[eq:Hammod\]) is the following, $$C = \frac{e^{2}}{2}\eta^{2}\int_{V}\int_{V} \frac{ d^{3}{\bf x} d^{3}{\bf x}'}{\mid {\bf x} - {\bf x}' \mid }, \label{eq:Charge}$$ where $e\eta$ is the density of uniformly distributed positive charge related to the average density of the electron charge as follows, $$e\eta = e\sum_{\alpha} < \psi^{\dagger}_{\alpha}({\bf x})\psi_{\alpha}({\bf x}) >. \label{eq:defeta}$$ Incorporating the constant $C$ as defined in (\[eq:Charge\]) above in the total Hamiltonian (\[eq:Hammod\]), makes the system electrically neutral, as it should be. The ground state energy $E_{0}$ of the unmagnetised state in the model above has been derived in perturbation theory [@Gell-Mann], [@Wigner1], [@Macke], [@Pines], with $H_{0}$ as the unperturbed Hamiltonian , and expressed as a sum of three terms, $$E_{0} = E_{kin.} + E_{exch.} + E_{corr.} \label{eq:GrE1}$$ The expression for the leading term $E_{kin.}$ in (\[eq:GrE1\]) is as follows, $$E_{kin.} = \frac{e^{2}N}{2a_{0}}\,\frac{3}{5} \left (\frac{9\pi}{4} \right )^{\frac{2}{3}} \frac{1}{r_{s}^{2}} \approx \frac{e^{2}N}{2a_{0}}\,\frac{2.21}{r_{s}^{2}}, \label{eq:Ekin}$$ where $N$ is the total number of electrons in the system, $a_{0}$ is the Bohr radius, $r_{s} = r_{0}/a_{0}$ and $r_{0}$ is a parameter related to the volume per electron, i.e. $\frac{4}{3}\pi r_{0}^{3} = V/N$. The term $E_{exch.}$ in (\[eq:GrE1\]) is the exchange energy, i.e. the expectation value of $H_{int.}$ in the ground state of a free electron gas, and has the following form, $$E_{exch.} = - \frac{e^{2} N }{2a_{0}}\,\frac{3}{2\pi} \left (\frac{9\pi}{4} \right )^{\frac{1}{3}} \frac{1}{r_{s}} \approx - \frac{e^{2} N }{2a_{0}}\,\frac{0.916}{r_{s}}. \label{eq:Eexch}$$ The last term in the expression (\[eq:GrE1\]), is the so-called correlation energy, a term introduced by Wigner [@Wigner1]. This term was calculated by Gell-Mann and Brueckner to the accuracy $e^{4}$, by summing an infinite number of so-called ring diagrams, with the result, $$E_{corr.} = \frac{e^{2} N }{2a_{0}}\,\left [\frac{2}{\pi^{2}}(1-\log2)\log r_{s} - 0.096 \right ] \approx \frac{e^{2} N }{2a_{0}}\,\left [0.0622 \log r_{s} - 0.096 \right ]. \label{eq:Ecorr}$$ It was pointed out by F. Bloch already in 1929 [@Bloch] that the ferromagnetic state of the electron plasma in metals could have a lower energy than that of the unmagnetised state. By considering a polarised gas of electrons with $N_{+}$ electrons in the spin-up state, and $N_{-}$ electrons in the spin-down state and defining the polarisation $\xi = (N_{+} - N_{-})/N$, where $N = N_{+} + N_{-}$, one finds the following expression for the ground state energy to the first order in the interaction potential, $$\begin{aligned} \label{eq:Emag} E_{mag} & = & E_{kin.} + E_{exch.}\\ & = & \frac{e^{2}N}{2a_{0}}\, \frac{3}{10} \left (\frac{9\pi}{4}\right )^{\frac{2}{3}} \left [(1+\xi)^{\frac{5}{3}} + (1-\xi)^{\frac{5}{3}}\right ]\frac{1}{r_{s}^{2}}\nonumber \\ & & - \frac{e^{2}N}{2a_{0}}\, \frac{3}{4\pi} \left (\frac{9\pi}{4}\right )^{\frac{1}{3}} \left [(1+\xi)^{\frac{4}{3}} + (1-\xi)^{\frac{4}{3}}\right ]\frac{1}{r_{s}}. \nonumber\end{aligned}$$ One readily infers from the expression (\[eq:Emag\]) that the ferromagnetic state, with $\xi = 1$, has a lower energy than the unmagnetised state with $\xi = 0$, provided that $$r_{s} > \frac{2\pi}{5}\,\left (\frac{9\pi}{4} \right )^{\frac{1}{3}} [2^{\frac{1}{3}} + 1] \approx 5.45. \label{eq:ineq}$$ This problem is also discussed in the textbook by Fetter and Walecka [@Fetter-W]. In the relation (\[eq:Emag\]) the polarisation $\xi$ is a free parameter subject to variation in the interval $[0, 1]$. The expression (\[eq:Emag\]) represents, in fact, only the energy of the electrons in a uniform magnetic field, which gives rise to the polarisation of these electrons. If the electrons become spontaneously polarised, then there should appear a spontaneously self-organised magnetic field ${\bf H}$ in the system, which makes a positive contribution $E_{field}$ to the total energy of the system. Then the total energy of the ferromagnetic state $E_{ferro}$ would be the following, $$E_{ferro} = E_{mag} + E_{field} \label{eq:ferroE}$$ There is the possibility, that the energy $E_{ferro}$ may be lower than the energy of the unmagnetised state given to lowest order by the sum of (\[eq:Ekin\]) and (\[eq:Eexch\]). Thus the ferromagnetic state can in principle be a genuine ground state of the system for certain external conditions. In [@Wigner2] and [@Pines] doubts were expressed concerning the possibility of having a ferromagnetic state as a genuine ground state of the Hamiltonian (\[eq:Hammod\]) in view of the fact, among other things, that the contribution $E_{field}$ was neglected in the considerations based on (\[eq:Emag\]). Independently of arguments for or against the possibility of a ferromagnetic ground state for the system described by the Hamiltonian (\[eq:Hammod\]), the Bloch considerations based on (\[eq:ferroE\]) pose a challenge in electron plasma theory. The challenging task is to derive a mechanism by means of which the Coulomb interaction between electrons in the model described by (\[eq:Hammod\]), can generate magnetism. To the best of our knowledge, this has been an unsolved problem so far. Below we give a solution to this problem, i.e. we show, using the functional methods discussed in general terms in the Introduction, that there is a kind of magnetic field that arises due to the Coulomb interaction, and that the corresponding magnetic state under certain conditions is a genuine ground state of the system under consideration. Magnetic ordering in the electron plasma ======================================== The grand canonical partition function $Z$ corresponding to the Hamiltonian (\[eq:Hammod\]) is given by the functional integral (\[eq:partfZ\]) with the following action $$\begin{aligned} \label{eq:actSf} \lefteqn{S_{f}[\psi^{}, \psi] = } \\ & & \int_{0}^{\beta} d\tau \sum_{\alpha} \int d^{3}{\bf x} \left [ \psi_{\alpha}^{}({\bf x}, \tau) \frac{\partial}{\partial \tau}\psi_{\alpha}({\bf x}, \tau) - \frac{\hbar^{2}}{2m}\psi_{\alpha}^{}({\bf x}, \tau )\nabla^{2} \psi_{\alpha}({\bf x}, \tau ) - \mu \psi_{\alpha}^{}({\bf x}, \tau ) \psi_{\alpha}({\bf x}, \tau ) \right ] \nonumber \\ & & + \frac{e^{2}}{2}\int_{0}^{\beta} d\tau \sum_{\alpha, \alpha'}\int \int \frac{ d^{3}{\bf x} d^{3}{\bf x}'}{\mid {\bf x} - {\bf x}' \mid } \left [\psi^{}_{\alpha'}({\bf x}', \tau )\psi^{}_{\alpha}({\bf x}, \tau )\psi_{\alpha}({\bf x}, \tau ) \psi_{\alpha'}({\bf x}', \tau) \right ] - \beta C, \nonumber\end{aligned}$$ where the electron field variables $\psi_{\alpha}^{}({\bf x}, \tau )$ and $\psi_{\alpha}({\bf x}, \tau)$ are Grassmann variables enumerated by the spin index $\alpha$, the space coordinate ${\bf x}$ and the “time” $\tau$, and $\mu$ is the chemical potential. The electron field variables satisfy the antiperiodic boundary conditions, $$\psi_{\alpha}^{}({\bf x}, \tau + \beta ) = - \psi_{\alpha}^{}({\bf x}, \tau )\,,\; \psi_{\alpha}({\bf x}, \tau + \beta ) = - \psi_{\alpha}({\bf x}, \tau ). \label{eq:antipbc}$$ The unmagnetised electron plasma -------------------------------- We first review one known [@Kleinert], [@Popov] way of introducing bosonic variables in the fermionic system described in terms of the functional integral (\[eq:partfZ\]). This consists of introducing an electric scalar potential $\phi({\bf x}, \tau)$ , in the present case when the action is given by the expression (\[eq:actSf\]) above. As mentioned in the Introduction, there may in general be more than one way of transforming the functional integral (\[eq:partfZ\]) into a functional integral (\[eq:HSpartfZ\]) involving also bosonic (Grassmann even) variables. We consider an alternative to the procedure explained below in the next subsection. The partition function $Z$ involving the electric scalar potential is the following, $$Z = \frac{1}{Z_{0 \phi}}\int D(\phi, \psi^{}, \psi) \exp \left \{ - S_{bf}[\phi, \psi^{}, \psi] \right \}, \label{eq:scalpot}$$ where $$\begin{aligned} \label{eq:scalact} S_{bf}[\phi, \psi^{}, \psi] & = & \frac{1}{8\pi }\int_{0}^{\beta} d\tau \int d^{3} {\bf x}\; \left (\nabla \phi({\bf x}, \tau)\right )^{2} \\ & & + \int_{0}^{\beta} d\tau \int d^{3} {\bf x}\;\sum_{\alpha} \psi_{\alpha}^{}({\bf x}, \tau) \frac{\partial}{\partial \tau}\psi_{\alpha}({\bf x} , \tau) \nonumber\\ & & + \int_{0}^{\beta} d\tau \int d^{3} {\bf x} \sum_{\alpha} \left [ - \frac{\hbar^{2}}{2m}\psi_{\alpha}^{}({\bf x}, \tau )\nabla^{2} \psi_{\alpha}({\bf x}, \tau ) - \mu \psi_{\alpha}^{}({\bf x}, \tau ) \psi_{\alpha}({\bf x}, \tau )\right ] \nonumber \\ & & + ie \int_{0}^{\beta} d\tau \int d^{3} {\bf x}\; \phi({\bf x}, \tau) \left [ \eta - \sum_{\alpha} \psi_{\alpha}^{}({\bf x}, \tau ) \psi_{\alpha}({\bf x}, \tau )\right ]. \nonumber\end{aligned}$$ In Eq. (\[eq:scalact\]) above, the quantity $e\eta$ is a uniform positive charge density and $Z_{0\phi}$ is defined by the following functional integral, $$Z_{0\phi} = \int D(\phi) \exp \left \{- \frac{1}{8\pi }\int_{0}^{\beta} d\tau \int d^{3} {\bf x}\, [\nabla \phi({\bf x}, \tau)]^{2} \right \}. \label{eq:Zophi}$$ The quantiy $\phi({\bf x}, \tau)$ introduced above, is a bosonic, i.e. Grassmann even, integration variable, which satisfies the following periodic boundary conditions, $$\phi({\bf x}, \tau + \beta) = \phi({\bf x}, \tau) \label{eq:bcphi}$$ In the action (\[eq:scalact\]) one can recognize a contribution from the energy density of an electric field ${\bf E}({\bf x}, \tau) = - \nabla \phi({\bf x}, \tau)$, and from a system of electrons interacting through a scalar potential $\phi({\bf x}, \tau)$. Performing the integration over the variable $\phi({\bf x}, \tau)$ in the functional integral (\[eq:scalpot\]) one recovers indeed the functional integral (\[eq:partfZ\]) with the action $S$ given by Eq. (\[eq:actSf\]). However, one may also first integrate over the variables $\psi^{}$ and $\psi$ in Eq. (\[eq:scalpot\]). One then obtains, $$Z = \frac{1}{Z_{0\phi}} \int D[\phi] \exp \left \{ - \int_{0}^{\beta} d\tau \int d^{3} {\bf x}\;\left [\frac{1}{8\pi}[\nabla \phi({\bf x}, \tau)]^{2} + ie \eta \phi({\bf x}, \tau) \right ]\, det\;{\cal F}(\phi) \right \}, \label{eq:Zbos}$$ wherere ${\cal F}(\phi)$ is a functional matrix, the matrix elements of which are as follows, $$<{\bf x}', \tau', \alpha' \mid {\cal F}(\phi) \mid {\bf x}, \tau , \alpha >\; = \left [\frac{\partial}{\partial \tau} - \frac{\hbar^{2}}{2m} \nabla_{\bf x}^{2} - \mu -ie \phi({\bf x}, \tau ) \right ]\, \delta(\tau' - \tau) \,\delta({\bf x}' - {\bf x})\, \delta_{\alpha' \alpha}. \label{eq:Fm-elem}$$ The expression (\[eq:Zbos\]) defines a certain effective action $S_{eff}$. Evaluating the functional integral (\[eq:Zbos\]) approximatively, to the order $e^{4}$, using the general procedure given by Eqns. (\[eq:EuLa\]) and (\[eq:sadl2\]) above, one reproduces the expressions (\[eq:Ekin\]), (\[eq:Eexch\]) and (\[eq:Ecorr\]) in the expression (\[eq:GrE1\]) for the ground state energy $E_{0}$ of the unmagnetised electron plasma. The magnetised electron plasma ------------------------------ We now come to the central issue in this paper, namely that there is another way of recognising bosonic variables in the system described by the action (\[eq:actSf\]), than the one described above, which utilised a scalar field $\phi({\bf x}, \tau )$. The novel transformation involves [bilocal]{} fields. Consider auxiliary commuting (Grassmann even) real variables $A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau )$, enumerated by spin indices $\alpha ,\alpha' = \pm$, by two space coordinates ${\bf x}'$ and ${\bf x}$ (${\bf x}' \neq {\bf x}$) and by a “time-variable” $\tau $. The commuting variables $A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau )$ are taken to be antisymmetric with respect to a simultaneous transposition of space- and spin indices, $$A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) = - A_{\alpha\alpha'}({\bf x}, {\bf x}', \tau ) \label{eq:antisA}$$ Furthermore, these variables are supposed to satisfy periodic boundary conditions, $$A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau + \beta ) = A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) \label{eq:perbcA}$$ The variables $A_{\alpha'\alpha}$ can be regarded as matrix elements of a certain $2\times 2$ matrix, expressed with the aid of the $2\times 2$ unit matrix $\sigma^{0} = 1$ and the usual Pauli matrices $\sigma^{k}, ( k = 1,2,3)$, and a four-component quantity $B = (B_{0},B_{1}, B_{2}, B_{3})$ as follows, $$A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) = \sum_{\ell=0}^{3} \sigma^{\ell}_{\alpha'\alpha} B_{\ell}({\bf x}', {\bf x}, \tau ) \label{eq:Arepr}$$ The symmetry conditions (\[eq:antisA\]) satisfied by the real quantities $A_{\alpha'\alpha}$ imply the following reality and symmetry conditions for the four-component quantity $B$, $$Im \;B_{k}({\bf x}', {\bf x}, \tau ) \equiv 0, \;B_{k}({\bf x}', {\bf x}, \tau ) = - B_{k}({\bf x}, {\bf x}', \tau )\;\;, k = 0,1,3. \label{eq:syprB013}$$ The second component $B_{2}$ is purely imaginary, and symmetric under the interchange of ${\bf x}$ and ${\bf x}'$, $$Re \;B_{2}({\bf x}', {\bf x}, \tau ) \equiv 0, \; B_{2}({\bf x}', {\bf x}, \tau ) = + \, B_{2}({\bf x}, {\bf x}', \tau ) \label{eq:syprB2}$$ The real variables $A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau )$ are used to define an action $S_{0b}$ as follows, $$\begin{aligned} \label{eq:Brepr} S_{0b}[B] & = &\frac{1}{4e^{2}} \int_{0}^{\beta}d\tau \int\int d^{3}{\bf x}d^{3}{\bf x}' \mid {\bf x} - {\bf x}'\mid \sum_{\alpha'\alpha} A^{2}_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) \\ & = & \frac{1}{2e^{2}} \int_{0}^{\beta} d\tau \int\int d^{3}{\bf x}d^{3}{\bf x}'\mid {\bf x} - {\bf x}'\mid \sum_{k=0}^{3} B_{k}^{\dagger}({\bf x}', {\bf x}, \tau ) B_{k}({\bf x}', {\bf x}, \tau ), \nonumber\end{aligned}$$ and the corresponding partition function, $$Z_{0b} = \int D(B) \exp \left \{ - S_{0b}[B] \right \}. \label{eq:Z0b[B]}$$ We wish to evaluate the partition function (\[eq:partfZ\]) with the action $S_{f}$ given by Eq. (\[eq:actSf\]). To this end we multiply Eq. (\[eq:partfZ\]) with the evident identity, $$1 = \frac{1}{Z_{0b}} \int D(B) \exp \left \{ - S_{0b}[B] \right \}. \label{eq:iden}$$ This results in the following expression for the partition function $Z$, $$Z = \frac{1}{Z_{0b}}\int D(B) D(\psi^{}, \psi) \exp \left \{ - S_{0b}[B] - S_{f}[\psi^{}, \psi] \right \} \label{eq:finZ}$$ In the expression (\[eq:finZ\]) we finally make the following change of integration variables, $$\begin{aligned} \label{eq:shift} A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) & \rightarrow & A'_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) = A_{\alpha'\alpha}({\bf x}', {\bf x}, \tau ) + \\ & & + \frac{ie^{2}}{\mid {\bf x}' - {\bf x} \mid } [\psi^{}_{\alpha'}({\bf x}', \tau)\psi_{\alpha}({\bf x}, \tau) - \psi^{}_{\alpha}({\bf x}, \tau)\psi_{\alpha'}({\bf x}', \tau)]. \nonumber\end{aligned}$$ The reason for the shift of integration variable described above, is simply that by evaluating the functional integral (\[eq:finZ\]) in the manner indicated, one cancels the Coulomb interaction term in the expression (\[eq:actSf\]) for $S_{f}[\psi^{}, \psi]$ and is left with the following simple expression, $$Z = \frac{1}{Z_{0b}} \int D(B, \psi^{}, \psi) \exp \left \{- S_{bf}[B, \psi^{}, \psi] \right \} \label{eq:ZpsiB}$$ with an action $S_{bf}[B, \psi^{}, \psi]$ given by the following expression, $$\begin{aligned} \label{eq:Sbfin} S_{bf}[B, \psi^{}, \psi] & = & S_{0b}[B] + \int_{0}^{\beta} \int d^{3}{\bf x} \psi^{}({\bf x}, \tau)\left [ \frac{\partial}{\partial \tau} - \frac{\hbar^{2}}{2m} \nabla^{2} - \mu \right ]\psi({\bf x}, \tau) \\ & & + i\int_{0}^{\beta} \int \int d^{3}{\bf x} d^{3}{\bf x}'\psi^{}({\bf x}', \tau)\;\left (\sum_{k=0}^{3}\sigma^{k} B_{k}({\bf x}', {\bf x}, \tau) \right ) \;\psi({\bf x}, \tau) - \beta C \nonumber\end{aligned}$$ In the expression (\[eq:Sbfin\]) we have used ordinary two-component notation, $$\psi = \left (\begin{array}{c} \psi_{+} \\ \psi_{-} \end{array} \right ) \;,\; \psi^{} = (\psi^{}_{+}, \psi^{}_{-}) \label{eq:spinot}$$ for the Grassmann variables $\psi^{}({\bf x}, \tau)$ and $\psi({\bf x}, \tau)$, respectively. The action (\[eq:Sbfin\]) apparently describes a system of electrons interacting through a bilocal scalar potential field $B_{0}({\bf x}',{\bf x}, \tau)$ and a bilocal magnetic field ${\bf B}({\bf x}',{\bf x}, \tau)$. The special symmetry properties of these field variables under interchange of space coordinates are given in Eqns. (\[eq:syprB013\]) and (\[eq:syprB2\]). The formal transformations (\[eq:finZ\]) and (\[eq:shift\]) leading from the original action (\[eq:actSf\]) to the action (\[eq:Sbfin\]), which describes a system of electrons interacting through a bilocal field $B({\bf x}, {\bf x}', \tau)$, define a mechanism through which magnetic interactions can be generated from the Coulomb interaction. This mechanism is an alternative to the transformation involving only an electric scalar potential $\phi({\bf x}, \tau)$ which resulted in the action $S_{bf}$ given in Eq. (\[eq:scalact\]), by means of which one could reproduce the expressions (\[eq:Ekin\]), (\[eq:Eexch\]) and (\[eq:Ecorr\]) in the formula (\[eq:GrE1\]) for the ground state energy $E_{0}$ of the unmagnetised electron plasma. Clearly, the self-consistency of the mechanism described above will have to be ascertained. We return to this question presently, and continue now with a straightforward analysis based on the action (\[eq:Sbfin\]), which is bilinear in the fields $\psi^{}$ and $\psi$. This circumstance makes it possible to carry out the functional integrations over $\psi^{}$ and $\psi$ in the expression (\[eq:ZpsiB\]) exactly and explicitly, with the result, $$Z = \frac{e^{\beta C}}{Z_{0b}} \int D[B] \exp \left \{ - S_{0b}[B] \right \}\;det \;{\cal F}(B), \label{eq:Zfin2}$$ where ${\cal F}(B)$ is a functional matrix, the matrix elements of which can be written as follows, $$\begin{aligned} \label{eq:functmatr} <{\bf x}', \tau', \alpha' \mid {\cal F}(B) \mid {\bf x}, \tau, \alpha> & = & \left [\frac{\partial}{\partial \tau} - \frac{\hbar^{2}}{2m} \nabla^{2} - \mu \right ]\; \delta_{\alpha'\alpha}\; \delta(\tau' - \tau)\; \delta({\bf x}'- {\bf x}) \\ & & + i\;\sum_{\ell=0}^{3}\sigma^{\ell}_{\alpha'\alpha} B_{\ell}({\bf x}', {\bf x}, \tau) \;\, \delta(\tau' - \tau) \nonumber\end{aligned}$$ The Dirac $\delta$-function $\delta(\tau' - \tau)$ which enters in the relation (\[eq:functmatr\]) above, is here defined in terms of its spectral representation, $$\delta(\tau' - \tau) = \frac{1}{\beta} \sum_{\nu = -\infty}^{+\infty} \exp i\omega_{\nu}(\tau' - \tau),\; \omega_{\nu} = \frac{\pi}{\beta}(2\nu + 1), \label{eq:deltatau}$$ where the summation index $\nu$ runs through all integer values, and the quantities $\omega_{\nu}$ are the so-called Matsubara frequencies. Employing the relation $$det\;{\cal F}(B) = \exp \left \{Tr \log {\cal F}(B) \right \} \label{eq:detxtr}$$ we obtain the final result for the partition function $Z$ in Eq. (\[eq:Zfin2\]), $$Z = \frac{1}{Z_{0b}} \int D(B) \exp \left \{ - S_{eff}[B] \right \}, \label{eq:Zfin3}$$ where the effective action $S_{eff}[B]$ is as follows, $$S_{eff}[B] = S_{0b}[B] - \sum_{\alpha} \int_{0}^{\beta} d\tau \int d^{3}{\bf x}<{\bf x}, \tau, \alpha \mid \log {\cal F}(B) \mid {\bf x}, \tau, \alpha> - \beta C. \label{eq:FinSeff}$$ The equation (\[eq:FinSeff\]) will be the starting point for an approximate evaluation of the partition function $Z$ under consideration. This problem will be considered in the next subsection. The saddle point equations -------------------------- We now look for the extremals of the effective action (\[eq:FinSeff\]), i.e. for solutions to the equations, $$\label{eq:varSeff} \delta \,S_{eff}[B] = 0.$$ It is reasonably simple to evaluate the variations involved in Eq. (\[eq:varSeff\]). Taking in addition into account the reality and symmetry conditions (\[eq:syprB013\]) and (\[eq:syprB2\]) above, one obtains the following four equations from the conditions (\[eq:varSeff\]). For $\ell = 0,1,3$ the equations in question are, $$\begin{aligned} \label{eq:vareqB013} \lefteqn{\frac{1}{e^{2}} \mid {\bf x} - {\bf x}' \mid B_{\ell}({\bf x}, {\bf x}', \tau) = } \\ & & = \frac{i}{2}\sum_{\alpha \alpha'} \sigma^{\ell}_{\alpha \alpha'} \left \{ <{\bf x}', \tau, \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}, \tau, \alpha > - <{\bf x}, \tau, \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}', \tau, \alpha > \right \} \nonumber \end{aligned}$$ For $\ell = 2$ one gets, $$\begin{aligned} \label{eq:vareqB2} \lefteqn{\frac{1}{e^{2}} \mid {\bf x} - {\bf x}' \mid B_{2}({\bf x}, {\bf x}', \tau) = } \\ & & = - \frac{i}{2} \sum_{\alpha \alpha'} \sigma^{2}_{\alpha \alpha'} \left \{ <{\bf x}', \tau, \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}, \tau, \alpha > + <{\bf x}, \tau, \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}', \tau, \alpha > \right \} \nonumber\end{aligned}$$ Using the relation ( \[eq:functmatr\]) one finds readily that the quantity $<{\bf x}', \tau', \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}'', \tau '', \alpha''>$ satisfies the following equation, $$\label{eq:invF(B)} \sum_{\alpha'} \int d^{3}{\bf x}' {\cal O}({\bf x}, \alpha; {\bf x}', \alpha') <{\bf x}', \tau', \alpha' \mid {\cal F}^{-1}(B) \mid {\bf x}'', \tau '', \alpha''> = \delta(\tau' - \tau'') \delta^{3}({\bf x} - {\bf x}'') \delta_{\alpha \alpha''}$$ where the operator ${\cal O}$ is the following, $$\label{eq:operO} {\cal O}({\bf x}, \alpha; {\bf x}', \alpha') = \left \{ \left [- \frac{\partial}{\partial \tau'} - \frac{\hbar^{2}}{2m} \nabla_{{\bf x}}^{2} - \mu \right ] \delta_{\alpha \alpha'} \delta^{3}({\bf x} - {\bf x}') + i \sum_{\ell = 0}^{3} \sigma_{\alpha \alpha'}^{\ell} B_{\ell}({\bf x}, {\bf x}',\tau') \right \}.$$ The equations (\[eq:vareqB013\]) and (\[eq:vareqB2\]) together with the equation (\[eq:invF(B)\]) constitute a complicated set of nonlinear integral equations for the saddle point configuration $B_{\mu}({\bf x}, {\bf x}', \tau),\; \mu = 0,1,2,3$. It remains to show that a nontrivial saddle point configuration exists. We shall not attempt here to discuss the most general solution possible of the set of equations (\[eq:vareqB013\]), (\[eq:vareqB2\]) and (\[eq:invF(B)\]). Rather we confine our attention to the following special $\tau$-independent [ansatz]{}, which we denote by $\overline{B}$, $$\overline{B}_{\ell}({\bf x}, {\bf x}', \tau) = 0, \; \ell = 0,1,3;\; \overline{B}_{2}({\bf x}, {\bf x}', \tau) = -iH(\mid {\bf x} - {\bf x}' \mid ), \label{eq:Hansatz}$$ where $H(\mid{\bf x}\mid)$ is a real-valued function to be determined. Inserting the ansatz (\[eq:Hansatz\]) in the equation (\[eq:invF(B)\]), one finds that this equation, apart from a factor $-1$, is nothing but the equation for the Green’s function for an electron in a magnetic field of variable magnitude $H$ in the direction of the $2$-axis. Under these circumstances it is appropriate to introduce the following notation, $$G_{\alpha' \alpha''}({\bf x}' - {\bf x}'', \tau' -\tau''; H) \equiv - <{\bf x}', \tau', \alpha' \mid {\cal F}^{-1}(\overline{B}) \mid {\bf x}'', \tau '', \alpha''>. \label{eq:defGreen}$$ We consider first the equation (\[eq:vareqB2\]), which becomes an equation for the determination of the function $H(\mid {\bf x} \mid)$ in the ansatz (\[eq:Hansatz\]). In order to analyse this equation, it is convenient to consider the Fourier transform $\tilde{G}_{\alpha \alpha'}$ of the Green’s function $G_{\alpha \alpha'}$ defined by the equation (\[eq:defGreen\]), $$\tilde{G}_{\alpha \alpha'}({\bf k}, \nu; H) \equiv \int_{0}^{\beta} d\tau \int_{V} d^{3}{\bf x}\; G_{\alpha \alpha'}({\bf x}, \tau; H) \exp (-i\omega_{\nu}\tau + i{\bf k}\cdot {\bf x}). \label{eq:FouriG}$$ Using the equation (\[eq:invF(B)\]) in the present case, one readily finds that $$\sum_{\alpha'} \left \{ [i\omega_{\nu} - (\frac{\hbar^{2}{\bf k}^{2}}{2m} - \mu) ]\delta_{\alpha \alpha'} - \sigma^{2}_{\alpha \alpha'} \tilde{H}({\bf k}) \right \} \tilde{G}_{\alpha' \alpha''}({\bf k}, \nu; H) = \delta_{\alpha \alpha''}, \label{eq:EtildG}$$ where $$\tilde{H}({\bf k}) = \int_{V} d^{3}{\bf x} H(\mid {\bf x} \mid) \exp (i{\bf k}\cdot {\bf x}). \label{eq:FouriH}$$ The Eq. (\[eq:EtildG\]) has the following solution, $$\tilde{G}_{\alpha \alpha'}({\bf k}, \nu; H) = \frac{(i\omega_{\nu} - {\cal E}_{0}({\bf k}) + \mu)\,\delta_{\alpha \alpha'} + \sigma^{2}_{\alpha \alpha'}\tilde{H}({\bf k})}{(i\omega_{\nu} - {\cal E}_{+}({\bf k}) + \mu)(i\omega_{\nu} - {\cal E}_{-}({\bf k}) + \mu)}, \label{eq:soltildeG}$$ where $${\cal E}_{0}({\bf k}) = \frac{\hbar^{2} {\bf k}^{2}}{2m}, \;{\cal E}_{\pm}({\bf k}) = \frac{\hbar^{2} {\bf k}^{2}}{2m} \pm \tilde{H}({\bf k}). \label{eq:defcalE}$$ In passing we note the symmetry of the Green’s function in Eq. (\[eq:defGreen\]) under the interchange of ${\bf x}'$ and ${\bf x}''$, which readily follows from the equation (\[eq:soltildeG\]) above. From this symmetry follows that the ansatz (\[eq:Hansatz\]) is [consistent]{} with the remaining equations (\[eq:vareqB013\]). We then continue with our analysis of Eq. (\[eq:vareqB2\]). The equation (\[eq:vareqB2\]) now finally boils down to the following, $$\frac{1}{e^{2}}\mid {\bf x} \mid H(\mid {\bf x} \mid) = - \sum_{\alpha \alpha'} \sigma^{2}_{\alpha \alpha'} G_{\alpha' \alpha}({\bf x}, 0; H), \label{eq:fineqH}$$ where $$G_{\alpha' \alpha}({\bf x}, 0; H) = \frac{1}{V\beta} \sum_{{\bf k}}\,\lim_{N \rightarrow \infty}\sum_{\nu = -N}^{N} \tilde{G}_{\alpha' \alpha}({\bf k}, \nu; H) \exp (-i{\bf k}\cdot{\bf x}). \label{eq:finGeq}$$ Inserting the expression (\[eq:soltildeG\]) in Eq. (\[eq:finGeq\]) and using the well-known series, $$\lim_{N \rightarrow \infty} \sum_{\nu = -N}^{N} \frac{1}{i\pi (2\nu + 1) - \beta E} = \frac{1}{\exp(\beta E) + 1} - \frac{1}{2}, \label{eq:Ferfunct}$$ one obtains from Eq. (\[eq:fineqH\]) the equation for the unknown function $H$ in its final form, $$\mid{\bf x}\mid H(\mid{\bf x}\mid) = \frac{e^{2}}{V}\sum_{{\bf k}} \left \{ \frac{1}{\exp \beta({\cal E}_{-}({\bf k}) - \mu) + 1} - \frac{1}{\exp \beta({\cal E}_{+}({\bf k}) - \mu) + 1} \right \} \exp (-i{\bf k}\cdot{\bf x}). \label{eq:laseqH}$$ The equation (\[eq:laseqH\]) is still a fairly complicated nonlinear integral equation for the determination of the function $H$; it should be remembered that the $H$-dependence on the right hand side of Eq. (\[eq:laseqH\]) occurs through the energy spectra ${\cal E}_{\pm}$ defined in Eq. (\[eq:defcalE\]). In the next subsection we consider a limiting case of Eq. (\[eq:laseqH\]), for which the solution $H$ can be constructed explicitly. Solution of the integral equation at T = 0 ------------------------------------------ In this sub-section we consider a limting case, namely the limit $T \rightarrow 0$, of the integral equation (\[eq:laseqH\]). This limiting case, which corresponds to $\beta \rightarrow \infty$, is physically relevant, and furthermore such that one is able to obtain a fairly complete picture of the solution to Eq. (\[eq:laseqH\]) in the limit in question. Let us first note that Eq. (\[eq:laseqH\]) is equivalent to the following equation, $$\int_{V} d^{3}{\bf x}\; \mid{\bf x}\mid H(\mid{\bf x}\mid) \exp (i{\bf k}\cdot{\bf x}) = e^{2} \left \{\frac{1}{\exp \beta({\cal E}_{-}({\bf k}) - \mu) + 1} - \frac{1}{\exp \beta({\cal E}_{+}({\bf k}) - \mu) + 1} \right \}. \label{eq:FfinH}$$ In the limit $\beta \rightarrow \infty$ we have, $$\lim_{\beta \rightarrow \infty} \frac{1}{\exp\beta(E- \mu) + 1} = \Theta({\cal E}_{F} - E) \label{eq:stepE}$$ where $\Theta$ is the ordinary step function, and we use the notation, $$\lim_{T \rightarrow 0} \mu = {\cal E}_{F} \label{eq:FermE}$$ In the limit $T \rightarrow 0$ the analysis of the equation (\[eq:FfinH\]) boils down to finding the roots of the following equations, $${\cal E}_{\pm}({\bf k}) = {\cal E}_{F} \label{eq:roots}$$ where the quantities ${\cal E}_{\pm}({\bf k})$ have been defined in Eq. (\[eq:defcalE\]). It turns out that the relevant parameter in the analysis of the Eqns. (\[eq:roots\]) is the value of the quantity $\tilde{H}(0)$, which, without essential loss of generality, can be taken to be non-negative, as is easily verified. One still has to consider separately the cases when ${\cal E}_{F} \geq 0$ and ${\cal E}_{F} < 0$, repectively. The necessity to consider also the case with ${\cal E}_{F} < 0$ is a consequence of the fact that a part of the quasiparticle spectrum ${\cal E}_{-}({\bf k})$ is negative. We consider first the case $${\cal E}_{F} \geq 0 \label{eq: FermEpos}$$ There are then two cases to be considered, namely $$0 \leq \tilde{H}(0) < {\cal E}_{F} \label{eq:H0leqmu}$$ and $$\tilde{H}(0) \geq {\cal E}_{F} \label{eq:H0geqmu}$$ We first consider the case when Eq. (\[eq:H0leqmu\]) is in force. Then, assuming reasonable smoothness of the function $\tilde{H}({\bf k})$, both the equations (\[eq:roots\]) can have roots $k = k_{<}$ and $k = k_{>}$, respectively, where $${\cal E}_{+}({k_{<}}) \equiv \frac{\hbar^{2}}{2m} \;k_{<}^{2} + \tilde{H}(k_{<}) = {\cal E}_{F} \label{eq:ksmall}$$ and $${\cal E}_{-}({k_{>}}) \equiv \frac{\hbar^{2}}{2m} \;k_{>}^{2} - \tilde{H}(k_{>}) = {\cal E}_{F} \label{eq:klarge}$$ In the limit $\beta \rightarrow \infty$ one then gets the following equation from Eq. (\[eq:FfinH\]), $$\int_{0}^{\rightarrow \infty} dr\, r^{2} H(r) \sin(kr) = e^{2} \frac{k}{4\pi}\left \{\Theta(k_{>} - k) - \Theta(k_{<} - k) \right \} \label{eq:seHtwor}$$ From Eq. (\[eq:seHtwor\]) follows readily that $$H(r) = \frac{e^{2}}{2\pi^{2} r^{4}}\left \{ [\sin(k_{>}r) - k_{>}r \cos(k_{>}r)] - [\sin(k_{<}r) - k_{<}r \cos(k_{<}r)] \right \}, \label{eq:H(r)etwor}$$ from which follows, $$\tilde{H}(k) = \frac{e^{2}}{\pi} \left \{ k_{>} - k_{<} + \frac{k^{2} - k_{>}^{2}}{2k}\; \log \left \|\frac{k_{>} - k}{k_{>} + k}\right \| - \frac{k^{2} - k_{<}^{2}}{2k}\; \log \left \|\frac{k_{<} - k}{k_{<} + k}\right \| \right \}. \label{eq:etildHktwor}$$ It then remains to determine the actual values of the roots $k_{<}$ and $k_{>}$ from the equations (\[eq:ksmall\]) and (\[eq:klarge\]), respectively. From Eq. (\[eq:etildHktwor\]) one readily gets the following results, $$\tilde{H}(k_{<}) = \frac{e^{2}}{\pi} \left \{ k_{>} - k_{<} - \frac{k_{>}^{2} - k_{<}^{2}}{2k_{<}}\; \log \left \|\frac{k_{>} - k_{<}}{k_{>} + k_{<}}\right \| \right \} \label{eq:etildHksmall}$$ and $$\tilde{H}(k_{>}) = \frac{e^{2}}{\pi} \left \{ k_{>} - k_{<} - \frac{k_{>}^{2} - k_{<}^{2}}{2k_{>}}\; \log \left \|\frac{k_{>} - k_{<}}{k_{>} + k_{<}}\right \| \right \} \label{eq:etildHklarge}$$ Inserting the values for $\tilde{H}(k_{<})$ and $\tilde{H}(k_{>})$ from Eqns. (\[eq:etildHksmall\]) and (\[eq:etildHklarge\]) into the Eqns. (\[eq:ksmall\]) and (\[eq:klarge\]), respectively, one finally obtains a pair of transcendental equations for the determination of the roots $k_{<}$ and $k_{>}$. The analysis of these equations is facilitated by the introduction of a momentum variable $k_{+}$, related to the Fermi energy ${\cal E}_{F}$ as follows, $${\cal E}_{F} \equiv \frac{\hbar^{2}}{2m} k_{+}^{2} \label {eq:defk+}$$ It turns out that the existence of solutions to the transcendental equations (\[eq:ksmall\]) and (\[eq:klarge\]) determining the roots $k_{<}$ and $k_{>}$, depends on the values of the following dimensionless parameter $\alpha_{0}$, $$\alpha_{0} = \frac{1}{\pi a_{0} k_{+}} \label{eq:alfazero}$$ where $a_{0}$ denotes the Bohr radius, $$a_{0} = \frac{\hbar^{2}}{m e^{2}} \label{eq:Bohrrad}$$ It is not difficult to verify that the transcendental equations (\[eq:ksmall\]) and (\[eq:klarge\]) have roots $k_{<}$ and $k_{>}$, with $k_{>} > k_{<} > 0$, if and only if the parameter $\alpha_{0}$ introduced in Eq. (\[eq:alfazero\]) satisfies the following inequality, $$0 < \alpha_{0} < \frac{\sqrt{6}}{12} \approx 0.204. \label{eq:alfa0twor}$$ The inequality (\[eq:alfa0twor\]) yields a rather large lower bound on the Fermi energy ${\cal E}_{F}$. Namely, expressing the scale $k_{+}$ in the expression (\[eq:defk+\]) in terms of the parameter $\alpha_{0}$ defined in Eq. (\[eq:alfazero\]), and using the inequality (\[eq:alfa0twor\]) we obtain, $${\cal E}_{F} = \frac{1}{\pi^{2} \alpha_{0}^{2}}\; \frac{e^{2}}{2a_{0}} > \frac{24}{\pi^{2}}\; \frac{e^{2}}{2a_{0}} \label{eq:lowbFermE}$$ The lower bound (\[eq:lowbFermE\]) is numerically far above the Fermi energies for all existing metals. For completeness we nevertheless continue with the analysis of the case at hand. The pair of transcendental equations (\[eq:ksmall\]) and (\[eq:klarge\]) can only be solved numerically. For this purpose we introduce dimensionless parameters $u$ and $w$ by writing $$k_{<} = uk_{+}\, , k_{>} = wk_{+}. \label{eq:u,w}$$ In Table 1 below, we give values of the parameters $u$ and $w$ introduced above in (\[eq:u,w\]), which have been obtained by solving Eqns. (\[eq:ksmall\]) and (\[eq:klarge\]) numerically, for a set of values of the parameter $\alpha_{0}$ in the interval (\[eq:alfa0twor\]). $\alpha_{0} = 0.1$ $\alpha_{0} = 0.15$ $\alpha_{0} = 0.2$ ----- -------------------- --------------------- -------------------- $u$ 0.98120 0.88503 0.61358 $w$ 1.01794 1.08974 1.19723 : $k_{<} = uk_{+}$, $k_{>} = wk_{+}$ In view of the condition (\[eq:lowbFermE\]), the solution discussed above is presumably irrelevant for realistic physical systems. Hence we turn our attention to the case when the inequality (\[eq:H0geqmu\]) is in force, which is actually a much simpler case than that already considered. We now assume that the inequality (\[eq:H0geqmu\]) is in force. Then, still assuming reasonable smoothness of the function $\tilde{H}({\bf k})$, one finds that Eq. (\[eq:ksmall\]) has no non-negative root at all. Hence the equation (\[eq:FfinH\]) can now be written as follows, in the limit $T \rightarrow \infty$, $$\int_{0}^{\rightarrow \infty} dr\, r^{2} H(r) \sin(kr) = e^{2} \frac{k}{4\pi} \Theta(k_{>} - k) \label{eq:sineH}$$ where $k_{>}$ is determined by the equation (\[eq:klarge\]) alone. From Eq. (\[eq:sineH\]) one readily obtains the following expression, $$H(r) = \frac{e^{2}}{2\pi^{2} r^{4}} \left\{\sin(k_{>}r) - k_{>}r \cos(k_{>}r)\right\}. \label{eq:H(r)}$$ which immediately leads to the following expression for the function $\tilde{H}({\bf k})$, $$\tilde{H}(k) = \frac{e^{2}}{\pi} \left \{ k_{>} + \frac{k^{2} - k_{>}^{2}}{2k}\; \log \left \|\frac{k_{>} - k}{k_{>} + k}\right \| \right \}, \label{eq:tildeHk}$$ We then finally obtain $$\tilde{H}(0) = \lim_{k \rightarrow 0} \tilde{H}(k) = \frac{2 e^{2}}{\pi}\;k_{>} \label{eq:tilde{H}(0)}$$ and $$\tilde{H}(k_{>}) = \frac{ e^{2} }{\pi}\;k_{>}. \label{eq:H(k_{>})}$$ Using the result (\[eq:H(k\_[>]{})\]) in Eq. (\[eq:klarge\]) one obtains a simple quadratic equation for the determination of the root $k_{>}$. It is still expedient to use the momentum scale $k_{+}$ defined by Eq. (\[eq:defk+\]) as well as the dimensionless parameter $\alpha_{0}$ defined by Eq. (\[eq:alfazero\]). Writing $$k_{>} = wk_{+} \label{eq:defw}$$ one finds that Eq. (\[eq:klarge\]) is equivalent to the following equation, $$w^{2} - 2\alpha_{0} w - 1 = 0, \label{eq:w-eq}$$ which gives the final expression for the root $k_{>}$, $$k_{>} = (\alpha_{0} + \sqrt{\alpha_{0}^{2} + 1})\,k_{+} \equiv (1 + \sqrt{1 + \alpha_{0}^{-2}}) \frac{1}{\pi a_{0}} \label{eq:rootk2}$$ Let us recall that the analysis above is done under the assumption that both the inequalities (\[eq: FermEpos\]) and (\[eq:H0geqmu\]) are in force. The latter inequality is now equivalent to the following condition on the dimensionless parameter $\alpha_{0}$, $$\alpha_{0} \geq \frac{\sqrt{6}}{12} \approx 0.204 \label{eq:ineqalf}$$ For future reference we note that the inequality (\[eq:ineqalf\]) implies that $$\frac{2}{\pi a_{0}} \leq k_{>} \leq \frac{6}{\pi a_{0}} \label{eq:ineklarge}$$ It remains to consider the case of negative Fermi energy ${\cal E}_{F}$, $${\cal E}_{F} < 0 \label{eq:FermEneg}$$ In this case, there is only one equation for the determination of the root $k_{>}$, namely Eq. (\[eq:klarge\]). This equation can have a positive root $k_{>}$ under the condition (\[eq:FermEneg\]), if and only if $$\tilde{H}(0) > - {\cal E}_{F} \label{eq:negineqH0}$$ The analysis proceeds along the same lines as in the previous case with only minor modifications. We introduce a momentum scale $k_{-}$, determined by the (negative) Fermi energy, $${\cal E}_{F} = - \frac{\hbar^{2}}{2m} k_{-}^{2} \label{eq:defk-}$$ as well as a dimensionless parameter $\gamma_{0}$ analogous to the parameter $\alpha_{0}$ in the previous case, $$\gamma_{0} = \frac{1}{\pi a_{0} k_{-}} \label{eq:defgamma0}$$ Writing $$k_{>} = vk_{-} \label{eq:defvpar}$$ we now obtain the following quadratic equation from Eq. (\[eq:klarge\]) for the determination of the parameter $v$ $$v^{2} - 2\gamma_{0}v + 1 = 0 \label{eq:v-eq}$$ This gives two roots $k^{\pm}_{>}$, $$k^{-}_{>} = (\gamma_{0} - \sqrt{\gamma_{0}^{2} - 1})\,k_{-} \equiv (1 - \sqrt{1 - \gamma_{0}^{-2}}) \frac{1}{\pi a_{0}} \label{eq:k(-)v2}$$ and $$k^{+}_{>} = (\gamma_{0} + \sqrt{\gamma_{0}^{2} - 1})\,k_{-} \equiv (1 + \sqrt{1 - \gamma_{0}^{-2}}) \frac{1}{\pi a_{0}} \label{eq:k(+)v2}$$ There is the obvious requirement that the roots $k^{\pm}_{>}$ be real; from this follows that we must necessarily have $$\gamma_{0}^{2} \geq 1 \label{eq:cineqga}$$ The equations (\[eq:H(r)\]), (\[eq:tildeHk\]) and (\[eq:tilde[H]{}(0)\]) are now valid as such, but with $k_{>}$ replaced with either the root given in Eq. (\[eq:k(-)v2\]) or in Eq. (\[eq:k(+)v2\]). For negative Fermi energy ${\cal E}_{F} < 0$ there are thus two distinct solutions to the equation (\[eq:FfinH\]) in the limit $T \rightarrow 0$, for each fixed value of the free parameter $\gamma_{0}$ satisfying condition (\[eq:cineqga\]). The inequality (\[eq:negineqH0\]) does not give any additional restrictions on the parameter $\gamma_{0}$ besides the restriction (\[eq:cineqga\]). For future reference we finally note that the roots $k^{\pm}_{>}$ satisfy the following inequalities, as a consequence of the inequality (\[eq:cineqga\]), $$0 \leq k^{-}_{>} < \frac{1}{\pi a_{0}} \label{eq: kmin}$$ and $$\frac{1}{\pi a_{0}} \leq k^{+}_{>} < \frac{2}{\pi a_{0}} \label{eq: kplu}$$ Let us summarise the situation so far. We have shown the existence of solutions $H(r)$ to the equation (\[eq:laseqH\]) in the limit $T \rightarrow 0$. If the Fermi energy ${\cal E}_{F}$ is positive, then the solution $H(r)$ takes the form given in Eq. (\[eq:H(r)etwor\]) if the inequality (\[eq:H0leqmu\]) is in force. This is the case if the dimensionless parameter $\alpha_{0}$ defined by Eq. (\[eq:alfazero\]) satisfies the inequality (\[eq:alfa0twor\]). If the inequality (\[eq:H0geqmu\]) is in force, then the solution takes the form given in Eq. (\[eq:H(r)\]). This is the case if the parameter $\alpha_{0}$ satisfies the inequality (\[eq:ineqalf\]). The latter solution can be obtained from the former in the formal limit $k_{<} \rightarrow 0$. In both cases the solutions depend on a free parameter $\alpha_{0}$ constrained either by the inequality (\[eq:alfa0twor\]) or by the inequality (\[eq:ineqalf\]). If the Fermi energy ${\cal E}_{F}$ is negative, then the solutions $H(r)$ take the form given in Eq. (\[eq:H(r)\]), with either $k^{-}_{>}$ given by Eq. (\[eq:k(-)v2\]) replacing the parameter $k_{>}$, or $k^{+}_{>}$ given by Eq. (\[eq:k(+)v2\]) replacing the parameter $k_{>}$. The demonstration given above, is by no means a replacement of a fullfledged proof of existence of solutions $H(r)$ to the equation (\[eq:laseqH\]) in the general case, i.e. for nonzero finite $T$. However, the demonstration above is a proof of existence of solutions $H(r)$ in the limit $T \rightarrow 0$, by explicit construction. The solutions obtained will be used below to calculate the ground state energy of the system under consideration. Before that, we consider in the next section the effects of fluctuations around the solutions discussed in this subsection. Fluctuations ============ We have found a stationary point $\overline{B}$ of the effective action (\[eq:FinSeff\]), namely the [ansatz]{} (\[eq:Hansatz\]), in which the function $H$ is the solution to the equation (\[eq:laseqH\]). We then have to consider the effect of fluctuations around this stationary point. We define a fluctuation field ${\cal B}$ as follows, $${\cal B}({\bf x}, {\bf x}', \tau) = B({\bf x}, {\bf x}', \tau) - \overline{B}({\bf x}, {\bf x}', \tau) \label{eq:calB}$$ Then one has to expand the effective action (\[eq:FinSeff\]) in powers of the field ${\cal B}$. We will make this expansion up to second order in the field ${\cal B}$. After some calculations one obtains the dollowing result, $$S_{eff}[B] = S_{eff}[\overline{B}] + S_{ob}[{\cal B}] + S_{fl}[{\cal B}] - \beta C \label{eq:finaleff}$$ where $$\label{eq:finCflu} S_{fl}[{\cal B}] = \int_{0}^{\beta} d\tau \int_{0}^{\beta} d\tau ' \int d^{3}{\bf x}\int d^{3}{\bf x}'\int d^{3}{\bf x}''\int d^{3}{\bf x}'''\sum_{\mu, \nu = 0}^{3} \sum_{\alpha \alpha' \alpha'' \alpha'''} R[{\cal B}]$$ The quantity $R[{\cal B}]$ in the expression above is given in terms of the Greens functions $G$ defined in Eq. (\[eq:defGreen\]) as follows, $$R[{\cal B}] = - \frac{1}{2} G_{\alpha\alpha'}({\bf x} - {\bf x}', \tau - \tau'; H) \sigma^{\mu}_{\alpha'\alpha''}{\cal B}_{\mu}({\bf x}', {\bf x}'', \tau') G_{\alpha''\alpha'''}({\bf x}'' - {\bf x}''', \tau' - \tau; H)\sigma^{\nu}_{\alpha'''\alpha}{\cal B}_{\nu}({\bf x}''', {\bf x}, \tau) \label{eq:Rdef}$$ The effective action $S_{eff}[\overline{B}]$ at the stationary point $\overline{B}$ has the following explicit form, $$\begin{aligned} \label{eq:MN37} S_{eff}[\overline{B}] & = & \frac{\beta}{2e^{2}} \int d^{3}{\bf x} d^{3}{\bf x}'\mid {\bf x} - {\bf x}'\mid H^{2}({\bf x} - {\bf x}') \\ & - & \sum_{{\bf k}} \log \left \{[1+ e^{-\beta ({\cal E}_{+}({\bf k}) - \mu)}] [1+ e^{-\beta ({\cal E}_{-}({\bf k}) - \mu)}] \right \}\nonumber\end{aligned}$$ Let us now consider the physical meaning of the stationary point $\overline B$ in some detail. The effective action $S_{eff}[\overline{B}]$ originates from the following functional integral, $$\exp\left \{ - S_{eff}[\overline{B}] \right \} = \exp\left \{ - S_{ob}[\overline{B}] \right \} \int {\cal D}[\psi^{}, \psi] \exp\left \{ - S_{of}[\psi^{}, \psi] \right \} \label{eq:MN38}$$ where $S_{of}[\psi^{}, \psi]$ is the following action, $$\begin{aligned} \label{eq:MN39} S_{of}[\psi^{}, \psi] & = & \int_{0}^{\beta}d\tau \int d^{3} {\bf x} \left \{\psi^{}({\bf x}, \tau) \frac{\partial}{\partial \tau} \psi({\bf x}, \tau) + \psi^{}({\bf x}, \tau)\left [ - \frac{\hbar^{2}}{2m} \nabla^{2} - \mu \right ]\psi({\bf x}, \tau) \right \}\\ & + & \int_{0}^{\beta}d\tau \int d^{3} {\bf x}\int d^{3}{\bf x}' H(\|{\bf x} - {\bf x}'\|)\psi^{}({\bf x}', \tau)\sigma_{2} \psi({\bf x}, \tau) \nonumber\end{aligned}$$ The action (\[eq:MN39\]) describes a system of electrons interacting with a given non-local magnetic field ${\bf H} = (0, H(\|{\bf x}-{\bf x}'\|), 0)$. This magnetic field generates a correlated creation of an electron with spin up (down) at the point ${\bf x}'$, together with a simultaneus annihilation of an alectron of opposite spin down (up) at the point ${\bf x}$. Introducing the following parametrisations for the Grassman variables $\psi$ and $\psi^{}$, respectively, in Eq. (\[eq:MN38\]), $$\psi({\bf x}, \tau) = \frac{1}{\sqrt{V}}\sum_{\bf k} a_{\bf k}(\tau) e^{i{\bf k}\cdot {\bf x}},\;\psi^{}({\bf x}, \tau) = \frac{1}{\sqrt{V}}\sum_{\bf k} a^{}_{\bf k}(\tau) e^{-i{\bf k}\cdot {\bf x}} \label{eq:MN39.5}$$ one can express the action (\[eq:MN39\]) in terms of the variables $a_{\bf k}(\tau)$ and $a^{}_{\bf k}(\tau)$ as follows, $$\begin{aligned} \label{eq:MN40} S_{of}(a^{}, a) & = & \int_{0}^{\beta} d\tau \sum_{\bf k} \left \{ a^{}_{\bf k}(\tau)\frac{\partial}{\partial \tau}a_{\bf k}(\tau) + a^{}_{\bf k}(\tau)[{\cal E}_{\bf k} - \mu] a_{\bf k}(\tau)\right \}\\ & + & \int_{0}^{\beta} d\tau \sum_{\bf k} \tilde{H}({\bf k}) a^{}_{\bf k}(\tau)\sigma_{2}a_{\bf k}(\tau) \nonumber\end{aligned}$$ where $\tilde{H}({\bf k})$ is the Fourier transform of the function $H(\|{\bf x}\|)$ defined as in Eq. (\[eq:FouriH\]). The Grassmann variables $a^{}_{{\bf k}Ê\alpha}(\tau)$ and $a_{{\bf k}Ê\alpha}(\tau)$ are associated with electron states specified by the wave number ${\bf k}$ and spin $\alpha = \pm$. In order to diagonalize the action (\[eq:MN40\]) we introduce new fields $c_{{\bf k}Ê\alpha}$ and $c^{}_{{\bf k}Ê\alpha}$ by the following unitary transformation, $$\begin{aligned} \label{eq:MN41} & & c_{{\bf k} +} = \frac{1}{\sqrt{2}}(a_{{\bf k} +} - i a_{{\bf k} -})\;, \;c_{{\bf k} -} = \frac{1}{\sqrt{2}}(-ia_{{\bf k} +} + a_{{\bf k} -})Ê\\ & & c^{}_{{\bf k} +} = \frac{1}{\sqrt{2}}(a^{}_{{\bf k} +} + i a^{}_{{\bf k} -})\;, \;c^{}_{{\bf k} -} = \frac{1}{\sqrt{2}}(ia^{}_{{\bf k} +} + a^{}_{{\bf k} -}) \nonumber\end{aligned}$$ Then the action (\[eq:MN40\]) can be expressed as follows, $$\label{eq:MN42} S_{of}(c^{},c) = \int_{0}^{\beta}d\tau \sum_{\bf k} \left \{c^{}_{\bf k}\left [\frac{\partial}{\partial \tau} - \mu \right ] c_{\bf k} + {\cal E}_{+}({\bf k})c^{}_{{\bf k}+}c_{{\bf k}+} + {\cal E}_{-}({\bf k})c^{}_{{\bf k}-}c_{{\bf k}-} \right \}$$ where the quantities ${\cal E}_{\pm}({\bf k})$ are the energy spectra (\[eq:defcalE\]) of the quasiparticles described by the fields (\[eq:MN41\]). As evident from the relations (\[eq:MN41\]), the states of these quasiparticles are special superpositions of electron states with spin up and spin down. The probability to find an electron in a spin up or down state is equal to $\frac{1}{2}$ for each quasiparticle spin state. Finally, carrying out the appropriate functional integration in Eq. (\[eq:MN38\]) using the form (\[eq:MN42\]) of the action, one obtains $$\exp\left \{-S_{eff}[\overline B] \right \} = \exp\left \{-S_{ob}[\overline B] \right \} \prod_{\bf k}\left \{[1+ e^{-\beta ({\cal E}_{+}({\bf k}) - \mu)}] [1+ e^{-\beta ({\cal E}_{-}({\bf k}) - \mu)}] \right \} \label{eq:MN43}$$ which correponds exactly to the result given in Eq. (\[eq:MN37\]). It behooves us still to check whether the self-organised magnetic field ${\bf H} = (0, H(\|{\bf x}-{\bf x}'\|), 0)$, which is of order $e^{2}$, survives under fluctuations which are also of order $e^{2}$, or not. This involves an evaluation of the functional integral (\[eq:Zfin3\]), with the effective action given by the second order approximation (\[eq:finaleff\]). One obtains the following result, $$Z = \exp\left \{-S_{eff}[\overline{B}]\right \} Z_{f\ell} \label{eq:MNCC45}$$ where $S_{eff}[\overline{B}]$ is given by Eq. (\[eq:MN37\]) and $$\log Z_{f\ell} = \frac{\beta e^{2}}{4} \int \frac{d^{3}{\bf x}d^{3}{\bf x}'}{\|{\bf x}-{\bf x}'\|}\, [Tr\,G({\bf x}-{\bf x}', 0+; H)]^{2} \label{eq:CCMN45a}$$ In Eq. (\[eq:CCMN45a\]) above, the notation $Tr$ stands for the appropriate spin index sums of the indicated quantities involving the Green’s functions defined by Eq. (\[eq:defGreen\]). The term $\beta C$ in Eq. (\[eq:finaleff\]) was exactly cancelled by the direct Coulomb term which appeared in the average value $<S_{f\ell}({\cal B})>_{0}$. Using the results (\[eq:MNCC45\]) and (\[eq:CCMN45a\]) on gets the grand canonical potential $\Omega$, $$\Omega = -k_{B} T \log Z \label{eq:CCOmega}$$ from which one obtains the energy $E_{mag}$ of the electron system with the self-organized magnetic field ${\bf H}$, $$\begin{aligned} \label{eq:MN46} E_{mag} & = & \frac{1}{2e^{2}} \int d^{3}{\bf x} d^{3}{\bf x}'\mid {\bf x} - {\bf x}'\mid H^{2}({\bf x} - {\bf x}')\\ & + &\sum_{{\bf k}}\left \{{\cal E}_{+}({\bf k})f({\cal E}_{+}({\bf k})) + {\cal E}_{-}({\bf k})f({\cal E}_{-}({\bf k}))\right \} \nonumber \\ & - & \frac{e^{2}}{4}\int \frac{d^{3}{\bf x}d^{3}{\bf x}'}{\|{\bf x}-{\bf x}'\|} \,[Tr\,G({\bf x}-{\bf x}', 0+; H)]^{2} \nonumber\end{aligned}$$ where the function $f({\cal E})$ denotes the Fermi function, $$\label{eq:Fermif} f({\cal E}) = \frac{1}{e^{\beta ({\cal E} - \mu)} + 1}$$ In the expression (\[eq:MN46\]) the first term corresponds to the energy of the self-organised magnetic field ${\bf H}$,and the second term corresponds to the energy of electrons interacting with this self-organised magnetic field. The last term corresponds to the exchange energy and to fluctuations about the self-organised magnetic field (mean field) calculated to the order of $e^{2}$. This term does not dominate over the two first terms in Eq. (\[eq:MN46\]). Hence the fluctuations do not destroy the effects of the self-organised magnetic field. Ground state energy =================== In the case $T = 0$ the function $H(r)$, which determines the self-organised magnetic field is known explicitly, as shown in subsection (3.4.). If the inequality (\[eq:H0leqmu\]) is in force, then the function $H(r)$ is given by Eq. (\[eq:H(r)etwor\]). If the follwing inequality holds, $$\tilde{H}(0) \geq \|{\cal E}_{F}\| \label{eq:geninEF}$$ then the function $H(r)$ is given by Eq. (\[eq:H(r)\]). It should be noted that the latter case is obtained from the former in the formal limit $k_{<} \rightarrow 0$, whence we continue our analysis using the expression (\[eq:H(r)etwor\]). Using this expression, or rather its Fourier transform (\[eq:etildHktwor\]), we can evaluate the energy $E_{mag}$ in Eq. (\[eq:MN46\]) in a fairly explicit form. After straightforward calculations one obtains, $$\begin{aligned} \label{eq:MN47} E_{mag} & = & \frac{V}{2\pi^{2}} \left \{\frac{\hbar^{2}}{10m}(k_{>}^{5} + k_{<}^{5})\right \}\\ & - & \frac{V}{2\pi^{2}} \left \{\frac{e^{2}}{8\pi}\left [3(k_{>}^{4} + k_{<}^{4}) - k_{>}k_{<}(k_{>}^{2} + k_{<}^{2}) + 6(k_{>}^{2} - k_{<}^{2})^{2} \log \frac{k_{>} + k_{<}}{k_{>} - k_{<}}\right ] \right \}\nonumber\end{aligned}$$ Before proceeding further, let us recall that the equations (\[eq:klarge\]) and (\[eq:ksmall\]) have two positive roots $k_{>}$ and $k_{<}$ if and only if the inequality (\[eq:alfa0twor\]) is in force. This inequality in turn implies the lower bound (\[eq:lowbFermE\] ) for the Fermi energy ${\cal E}_{F}$, which numerically is far above the Fermi energies for all known metals. The physically relevant case obtains when the inequality (\[eq:geninEF\]) holds. The corresponding formula for $E_{mag}$ can be obtained by passing to the formal limit $k_{<} \rightarrow 0$ in Eq. (\[eq:MN47\]). Thus, if condition (\[eq:geninEF\]) holds true, then $$E_{0mag} \equiv \lim_{k_{<} \rightarrow 0} E_{mag} = \frac{V}{2\pi^{2}} \left \{\frac{\hbar^{2}}{10m}k_{>}^{5} - \frac{3e^{2}}{8\pi}k_{>}^{4} \right \} \label{eq:E0mag}$$ where the quantity $k_{>}$ is given either by the equation (\[eq:rootk2\]), which implies that $k_{>}$ is in the interval (\[eq:ineklarge\]) or by one or the other of the equations (\[eq:k(-)v2\]) or (\[eq:k(+)v2\]), which restrict the quantity $k_{>}$ to be in one or the other of the intervals (\[eq: kmin\]) or (\[eq: kplu\]). Thus, for realistic metals, we expect the ground state energy to be given by Eq. (\[eq:E0mag\]), which originates from quasiparticles in the spin down states only, with energy spectrum ${\cal E}_{-}({\bf k})$. At this point it is appropriate to recall that we are supposed to calculate the energy per particle. The number of electrons is given as follows, $$\begin{aligned} \label{eq:numberN} N & \equiv & \sum_{\alpha} \int d^{3}{\bf x}<\hat{\psi}_{\alpha}^{\dagger}({\bf x})\hat{\psi}_{\alpha}({\bf x})>\\ & = & \frac{V}{(2\pi)^{3}} \int d^{3} {\bf k} \left [f({\cal E}_{-}({\bf k})) + f({\cal E}_{+}({\bf k}))\right ]\nonumber\end{aligned}$$ where the functions $f({\cal E}_{\pm})$ denote the appropriate Fermi functions (\[eq:Fermif\]). In the limit $T \rightarrow 0$ the integral (\[eq:numberN\]) can immediately be evaluated by use of the formula (\[eq:stepE\]), with the result, $$N = \frac{V}{6\pi^{2}} (k_{>}^{3} + k_{<}^{3}) \label{eq:density}$$ Thus the parameters $k_{>}$ and $k_{<}$ are related to the average particle density as indicated above. In particular, in the case $k_{<} \rightarrow 0$ we obtain the following relation from Eq. (\[eq:density\]), $$\frac{V}{N} = 6\pi^{2}\,k_{>}^{-3} \label{eq:VNk>rel}$$ Using the reult (\[eq:VNk>rel\]) in Eq. (\[eq:E0mag\]) we obtain the energy per particle, $$\frac{E_{0mag}}{N} = \frac{3}{2} e^{2} \left \{ \frac{a_{0}}{5} k_{>}^{2} - \frac{3}{4\pi} k_{>} \right \} \label{eq:finE0k}$$ To obtain the genuine ground state energy, we still have to minimise the expression (\[eq:finE0k\]) with respect to the parameter $k_{>}$, which can vary within certain limits, as discussed above. One readily finds that the expression (\[eq:finE0k\]) has a minumum for the following value of $k_{>}$, $$(k_{>})_{min} = \frac{15}{8} \, \frac{1}{\pi a_{0}} = \frac{1.875}{\pi a_{0}} \label{eq:Ckmin}$$ The corresponding minimum value of the energy per particle is, $$\left (\frac{E_{0mag}}{N} \right )_{min} = - 15 \left (\frac{3}{8\pi} \right )^{2} \, \frac{e^{2}}{2a_{0}} \approx -0.214\, \frac{e^{2}}{2a_{0}} \label{eq:E0magmin}$$ Numerically the result (\[eq:E0magmin\]) corresponds to the following value, $$\left (\frac{E_{0mag}}{N} \right )_{min} \approx - 2.91\; eV \label{eq:numEmin}$$ The minimum (\[eq:E0magmin\]) occurs at the value (\[eq:Ckmin\]) for the parameter $k_{>}$, which is within the limits in Eq. (\[eq: kplu\]). Hence the minimum considered here corresponds to a [negative]{} value for the Fermi energy ${\cal E}_{F}$, $${\cal E}_{F} < 0 \label{eq:negEF}$$ In conclusion we wish to compare the results given above, with the ground state energy $E_{0}$ of the unmagnetised case considered in Sec. 2. The ground state energy in question is obtained by adding the quantities $E_{kin}$ and $E_{exch}$ given by Eq.(\[eq:Ekin\]) and Eq. (\[eq:Eexch\]), respectively, $$\label{eq:MN49} E_{0} = \frac{e^{2} N}{2a_{0}}\left \{\frac{3}{5}\left (\frac{9\pi}{4}\right )^{\frac{2}{3}} \frac{1}{r_{s}^{2}} - \frac{3}{2\pi}\left (\frac{9\pi}{4}\right )^{\frac{1}{3}} \frac{1}{r_{s}} \right \} \approx \frac{e^{2}N}{2a_{0}}\left \{\frac{2.21}{r_{s}^{2}} - \frac{0.916}{r_{s}} \right \}$$ The parameter $r_{s}$ occuring in the equation (\[eq:MN49\]) above, is related to the volume per electron, as detailed in Sec. 2, i.e. $$r_{s} = \frac{r_{0}}{a_{0}} \label{eq:defress}$$ and $$\frac{4}{3} \pi r_{0}^{3} = \frac{V}{N} \label{eq:defr0}$$ Thus, using Eq. (\[eq:VNk>rel\]) and the equations (\[eq:defress\]) and (\[eq:defr0\]) above, we obtain a relation between the parameters $k_{>}$ and $r_{s}$, $$r_{s} k_{>} a_{0} = \left (\frac{9\pi}{2} \right )^{\frac{1}{3}} \label{eq:resska}$$ Eliminating the parameter $k_{>}$ from Eq. (\[eq:finE0k\]) by means of Eq. (\[eq:resska\]) above, one obtains, $$E_{0mag} = \frac{e^{2}N}{2a_{0}} \left \{ \frac{3}{5}\left(\frac{9\pi}{2}\right )^{\frac{2}{3}} \frac{1}{r_{s}^{2}} - \frac{3}{2\pi}\left(\frac{9\pi}{2}\right )^{\frac{1}{3}} \frac{3}{2r_{s}} \right \} \approx \frac{e^{2}N}{2a_{0}} \left \{ \frac{3.508}{r_{s}^{2}} - \frac{1.731}{r_{s}} \right \} \label{eq:MN50}$$ The energy per particle of the magnetically ordered state is smaller than the corresponding energy per particle in the unmagnetised state, provided the parameter $r_{s}$ exceeds a certain limit $r_{sc}$, $$\frac{E_{0mag}}{N} < \frac{E_{0}}{N} \Longleftrightarrow r_{s} > r_{sc} \label{eq:critrs}$$ A comparison of the two expressions (\[eq:MN49\]) and (\[eq:MN50\]) above, yields a numerical estimate for the lower limit $r_{sc}$ in (\[eq:critrs\]), $$r_{sc} \approx 1.592 \label{eq:numrcrits}$$ The condition $r_{s} > 1.592$ is satisfied for all metals, so that the energy per particle in the magnetically ordered state is smaller than the energy per particle in the unmagnetised state in all the relevant cases. Let us finally minimise the energy per particle (\[eq:MN49\]) for the unmagnetised state with respect to $r_{s}$. One obtains, $$(r_{s})_{min} \approx 4.83\;, \; \left (\frac{E_{0}}{N}\right )_{min} \approx - 1.29\; eV \label{eq:MN52.5}$$ The minimum value $- 2.91\; eV$ for the energy per particle in the magnetically ordered state, which was obtained above in Eq (\[eq:E0magmin\]), is much closer to the experimental values for most metals [@Metals] than the value $-1.29\; eV$ obtained above from the energy per particle in the unmagnetised case. The minimum in question occurs at $$(r_{s})_{min} = \frac{8\pi}{15} \left (\frac{9\pi}{2} \right )^{\frac{1}{3}} \approx 4.05 \label{eq:last}$$ as is readily found from Eqns. (\[eq:Ckmin\]) and (\[eq:resska\]). In fact, the binding energies of all metals exceed numerically considerably the value $1.29\; eV$, and are closer to the value $2.91\; eV$ obtained for the magnetised case considered here. Summary and conclusions ======================= Starting form the simplest theory of metals, we have derived a particular mechanism by means of which the Coulomb interaction between electrons can generate a new kind of magnetically ordered state in a metal. The self-organised magnetic field geberated by this mechanism is represented by an order parameter which is a non-local field of a special kind. The self-organised field does not produce spin polarisations of the electrons, but generates particular long range correlations between electrons of opposite spin in different points ${\bf x}$ and ${\bf x}'$. The electron plasma in the model is described in terms of interacting quasiparticles of fermionic type. Each quasiparticle state is shown to be a special superposition of an electron spin up and spin down state, such that the probability of finding an electron with either spin up or spin down in each quasiparticle state equals $\frac{1}{2}$. Hence the spontaneously organised magnetic field does not generate polarisations of the electrons in the metal but only polarisations of the quasiparticles. The ground state of the electron plasma has been shown to be described by quasiparticles with a specific absolute spin-polarisation, but with no spin-polarisations of the electrons, for such values of e.g. the Fermi energy, which are valid for all existing metals. Hence the ground state in the simplest model of metals is not a ferromagnetic state as advocated by Bloch [@Bloch], but rather a state involving the spontaneously generated non-local magnetic field derived in this paper. The ground state properties of this magnetised state give rise to theoretical numerical values for the average distance between electrons and for the binding energy which are in reasonable agreement with experimetal data. It ought to be possible to use our theoretical analysis as a starting point for experimental research in condensed matter physics aiming at observing and studying the magnetically ordered states predicted in this paper. [Acknowledgements]{} A large portion of the work described in this paper was done while one of the authors (M. N.) was a summer visitor at the Helsinki Institute of Physics (HIP) in Helsinki. The hospitality of HIP, in particular T. Ala-Nissil[ä]{}, is gratefully acknowledged. [10]{} A. N. Vasiliev, Functional Methods in Quantium Field Theory and Statistical Physics, Gordon and Breach, Amsterdam, 1998. H. Kleinert, Fortschr. d. Physik, [26]{}, 565 (1978). V. N. Popov, Functional Integrals in Quantium Field Theory and Statistical Physics (in Russian) Atomizdat, Moscow, 1976. English translation: D. Reidel Publishing Company, Dordrecht, 1983. M. Gell-Mann and K. A. Brueckner, Phys. Rev. [106]{}, 364 (1957). F. Bloch, Z. Physik [57]{}, 545 (1929). A. L. Fetter and J. D. Walecka, Quantum Theory of Many Particle Systems, McGraw-Hill, New York, 1971; p. 32. E. P. Wigner, Phys. Rev. [46]{}, 1002 (1934). W. Macke, Z. Naturforschung, [5a]{}, 192 (1950). E. P. Wigner, Trans. Faraday Soc. [34]{}, 678 (1938). D. Pines, Phys. Rev. [92*]{}, 626 (1953). Metals, Reference Book, Edited by C. J. Smithells, Butterworths, London, 1976.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Linden, Massar and Popescu have recently given an optimization argument to show that a single two-qubit Werner state, or any other mixture of the maximally entangled Bell states, cannot be purified by local operations and classical communications. We generalise their result and give a simple explanation. In particular, we show that no purification scheme using local operations and classical communications can produce a pure singlet from any mixed state of two spin-$1/2$ particles. More generally, no such scheme can produce a maximally entangled state of any pair of finite-dimensional systems from a generic mixed state. We also show that the Werner states belong to a large class of states whose fidelity cannot be increased by such a scheme.' address: \| Department of Applied Mathematics and Theoretical Physics, University of Cambridge,\ Silver Street, Cambridge CB3 9EW, U.K. author: - Adrian Kent date: '28 May 1998; revised 23 August 1998' title: Entangled Mixed States and Local Purification --- \#1[\#1\|]{} \#1[\| \#1]{} \#1[\#1 \|]{} \#1[\| \#1]{} \#1\#2[ \#1 \| \#2 ]{} \#1[ \#1 ]{} \#1\#2[ \#1 , \#2 ]{} § [[v]{}]{} [2]{} The relationship between quantum entanglement and locality has been a source of great theoretical interest ever since the discovery of Bell’s theorem, and new subtleties continue to be discovered. Several practical applications of entanglement have also been proposed,[@BBCJPW; @E; @DEJMPS] most of which require that separated parties share a fixed maximally entangled state — conventionally, the Bell singlet state of two spin-$1/2$ particles. Since no preparation method is perfect and no transmission channel is noiseless, in practice we expect the parties to share entangled mixed states rather than pure singlets. This raises a problem which is also of independent theoretical interest: how can singlets be extracted from shared entangled mixed states? One solution, at least in principle, is to use the “entanglement purification” scheme of Bennett et al.[@BBPSSW]. But as this scheme relies on carrying out collective measurements on a large number of shared states, and produces perfectly pure singlets only when the number of shared states is infinite, it is natural to ask whether there is any simpler and more practical method. In particular, it is natural to ask whether there is any way of purifying individual entangled mixed states by a local purification scheme — that is, a scheme which involves only local measurements and unitary operations on the two entangled particles, together with classical communications between the parties. Linden, Massar and Popescu (LMP) have recently addressed this question, showing that the answer is negative for the entangled Werner states[@W] in two dimensions — i.e., for rotationally invariant mixed states of two spin $1/2$ particles whose fidelity to the Bell singlet is greater than $1/2$. More generally, they show that the fidelity of a Werner state — i.e. the proportion of singlets — cannot be increased by local purification schemes, and that no mixture of the maximally entangled Bell states of two spin-$1/2$ particles can be purified by such schemes to a state whose entanglement of formation is greater. As LMP comment, these results may at first sight seem surprising, given that singlets [can]{} be produced from non-maximally entangled pure states by individual measurements.[@BBPS] One might have conjectured that the same should be true of mixed states, at least if they are sufficiently close to being singlets. LMP’s argument shows that this is not true, but does not give any clear intuition as to why pure and mixed states differ in this respect, and leaves open the possibility that generic mixed states of spin-$1/2$ particles, or mixed states of higher-dimensional systems, might be locally purifiable. In this Letter we give a simple argument which generalises LMP’s results and explains why individual mixed states cannot generally be completely purified. We show that no local purification scheme can produce a pure singlet from a single copy of any entangled mixed state of two spin $1/2$ particles. More generally, we show that no local purification scheme can produce a maximally entangled state from a generic mixed state entangling any two finite-dimensional systems. Finally, we show that the Werner and Bell-diagonal states belong to a large class of states whose fidelity to a given state cannot be increased by any local purification scheme. Consider first a general mixed state of two spin-$1/2$ particles, $$\label{m} \rho = \sum_{i=1}^{n} p_i \ket{\psi_i } \bra{\psi_i} \, ,$$ where the $\ket{\psi_i} \in C^2 \otimes C^2$ are distinct, the $p_i$ are positive, and $n \geq 2$ is minimal, i.e. $\rho$ cannot be represented as a mixture of fewer than $n$ pure states. Two parties, Alice and Bob, each have one of the spin-$1/2$ particles, and want if possible to carry out a sequence of local quantum operations and classical communications so as to obtain, with non-zero probability, the singlet $$\label{singlet} \ket{\Psi_-} = {1 \over \sqrt{2}} ( \ket{\uparrow} \ket{\downarrow} - \ket{\downarrow} \ket{\uparrow} ) \, .$$ As the properties of a mixed state are independent of the details of its preparation we may, for clarity, suppose that a third party has actually prepared one of the pure states $\ket{\psi_i}$ for Alice and Bob, but gives them only the statistical information encoded in the decomposition (\[m\]) of $\rho$. In any local purification scheme, Alice and Bob carry out sequences of POV measurements, unitary rotations, and classical communications. They must either conclude from their measurement results that the purification has failed, in which case they jettison the state, or else — if the measurements lie in some specified subset of the possibilities — that it has succeeded. After any finite sequence of operations and measurements, the density matrix takes the form $$\label{ab} \rho' = { {A\otimes B \rho A^\dagger\otimes B^\dagger} \over {\Tr (A\otimes B \rho A^\dagger\otimes B^\dagger) } } \, ,$$ where $A = A_1 ... A_{n_a}$ and $B= B_1 ... B_{n_b}$ are the products of the positive operators and unitary maps corresponding to Alice’s and Bob’s measurements and rotations. Now if $\rho' = \ket{\Psi_-} \bra{\Psi_-}$ then we must have that $$\label{abp} A \otimes B \ket{\psi_i} = a_i \ket{\Psi_-} \, ,$$ for each $i$ from $1$ to $n$ and some constants $a_i$, which need not all necessarily be non-zero. But since $A \otimes B$ maps distinct rays onto the ray $\ket{\Psi_-}$, it cannot have maximal rank, which means that either $A$ or $B$ must have rank one or zero. This means that the $a_i$ must all be zero, contradicting the original hypothesis. The same argument holds true for an infinite sequence of local operations, so that no local purification scheme involving a countable number of steps can produce a singlet with non-zero probability. Suppose now that Alice and Bob share a mixed state $\rho$ of two quantum systems described by spaces $H_A$ and $H_B$ of dimensions $N_A \geq 2$ and $N_B \geq 2$. We can write $$\rho = \sum_{i=1}^{r} p_i \ket{\psi_i} \bra{\psi_i} \, ,$$ where the $\ket{\psi_i}$ are orthogonal, the $p_i$ are positive, and $r$ is the rank of $\rho$. Any sequence of local operations can be implemented by an operator $A \otimes B$ as in (\[ab\]), and if these operations purify $\rho$ to some state $\ket{\Psi}$ then we have $ A \otimes B \ket{\psi_i} = a_i \ket{\Psi}$ for each $i$ from $1$ to $r$ and some constants $a_i$, so that the rank of $ A \otimes B $ is at most $N_A N_B - r + 1$. If $r \geq N_A N_B - 2$, as is true for generic $\rho$, this implies that at least one of $A$ and $B$ has rank one or zero, so that $\ket{\Psi}$ cannot be entangled. Hence no local purification scheme can produce an entangled pure state from a generic mixed state of two separated finite-dimensional systems. Since the set of possible purification operators $A \otimes B$ is compact, any state which cannot be completely purified to a maximally entangled state must have a maximal purifiability — according to any continuous measure — whenever all the purification operators produce a well-defined state. Moreover, there must be some local purification for which this maximum is attained. For example, a generic mixed state $\rho$ of two finite-dimensional systems has maximal rank: $$\rho = \sum_{i=1}^{N_A N_B} p_i \ket{\psi_i} \bra{\psi_i} \, ,$$ where the $\ket{\psi_i}$ are orthonormal. Thus $\Tr ( A \otimes B \rho A^\dagger \otimes B^\dagger )$ is zero only if $A \otimes B$ is the zero operator. Since without loss of generality we can take $\\| A \\| = \\| B \\| = 1$, it follows that there is some $\epsilon ( \rho ) > 0 $ such that $$\label{maxpur} {\rm max}_{A,B; {\ket{\psi}}} {{ \bra{ \psi} A \otimes B \rho A^\dagger \otimes B^\dagger \ket{\psi}} \over { \Tr ( A \otimes B \rho A^\dagger \otimes B^\dagger ) }} = 1 - \epsilon (\rho ) \, ,$$ where the maximum is taken over all local purification operators $A$ and $B$ and all maximally entangled pure states $\ket{\psi}$. It is perhaps worth stressing that this does not contradict the results of Bennett et al.,[@BBPSSW] whose scheme produces a non-zero fractional yield of pure Bell singlets only in the asymptotic limit as the number of purified pairs tends to infinity. Nor does the result apply to all mixed states: it is easy to construct examples of mixed states of lower rank which can be purified to singlets. Examples of lower rank mixed states which can be purified to arbitrary high fidelity, though not fidelity $1$, are also known.[@HHH] Finally, we consider the problem of whether the proportion of some entangled pure state in a mixture can be increased by a local purification scheme. Suppose that $$\label{rhof} \rho_F = F \ket{\Psi} \bra{\Psi} + (1-F) \tilde{\rho}$$ is a density matrix, where $\ket{\Psi}$ is a normalised entangled state, $0 \leq F \leq 1$, and $\tilde{\rho}$ is a density matrix with $\bra{\Psi} \, \tilde{\rho} \, \ket{\Psi} = 0$. Suppose also that there is some $F_0$ with $0 < F_0 < F$ such that $\rho_{F_0}$ is not entangled, i.e. $\rho_{F_0}$ can be written as a convex combination of non-entangled pure states, and such that any state with fidelity greater than $F_0$ to $\ket{\Psi}$ is entangled. Then we can show that no local purification scheme can produce a mixed state $\rho'$ from $\rho_F$ such that the fidelity $F' = \bra{\Psi} \, \rho' \, \ket{\Psi}$ of $\rho'$ to the state $\ket{\Psi}$ is greater than $F$. As before, we can describe the action of any successful local purification scheme by (\[ab\]). If we consider the scheme acting on $\rho_F$, it produces a state of fidelity $$\begin{aligned} \label{ff} \lefteqn{F' (F) =} \qquad \\ & & {{ F \| \bra{\Psi} A \otimes B \ket{\Psi} \|^2 + (1-F) \bra{\Psi} A \otimes B \tilde{\rho} A^\dagger \otimes B^\dagger \ket{\Psi} } \over {\Tr ( A \otimes B ( F \ket{\Psi} \bra{\Psi} + (1-F) \tilde{\rho} ) A^\dagger \otimes B^\dagger ) }} \nonumber \, , \end{aligned}$$ with probability $$\Tr ( A \otimes B ( F \ket{\Psi} \bra{\Psi} + (1-F) \tilde{\rho} ) A^\dagger \otimes B^\dagger ) \, .$$ Note that if this probability is non-zero for any value of $F$ in the range $0 < F < 1$ then it is non-zero throughout the range. We have that $$\begin{aligned} \label{deriv} \lefteqn{{{d^2 } \over {d F^2}} \delta (F) = } \qquad \\ & & {{\beta'} \over { ( \Tr ( A \otimes B ( F \ket{\Psi} \bra{\Psi} + (1-F) \tilde{\rho} ) A^\dagger \otimes B^\dagger ) )^3 } } \nonumber \, , \end{aligned}$$ where $\delta(F) = (F'(F) - F)$ and $\beta'$ is independent of $F$, so that the sign of the second derivative is constant on the range $0 < F < 1$ and $\delta(F)$ is either convex, concave or linear over the range. Now $\delta(F) \geq -F$ and so approaches or exceeds zero as $F \rightarrow 0$ from above; similarly $\delta(F)$ approaches or is less than zero as $F \rightarrow 1$ from below. By hypothesis $\delta ( F_0 ) \leq 0 $, since it is impossible to obtain an entangled state from the unentangled state $\rho_{F_0}$. Hence, if there is an $F$ in the range $F_0 < F < 1$ such that $\delta(F) > 0$, then, given the form of (\[ff\]), $\delta (F)$ must have at least two extrema in the range $0 < F < 1$, which is inconsistent with (\[deriv\]). So there can be no such $F$. Hence no local purification scheme can increase the fidelity of $\rho_F$ for $F \geq F_0$. As a special case, applying the result to a pair of spin-$1/2$ particles, taking $\ket{\Psi}$ to be the Bell singlet state $\ket{\Psi_-}$ and $\tilde{\rho}$ to be a linear combination $a \ket{\Psi_+ }\bra{\Psi_+} + b \ket{\Phi_-}\bra{\Phi_-} + c \ket{\Phi_+}\bra{\Phi_+}$, we obtain LMP’s result that the fidelity of a Werner state, or any Bell-diagonal state, with $F>{1 \over 2}$ cannot be increased by a local purification scheme. In summary, the difficulty with purifying individual mixed states is that any purification scheme must have a well-defined action on each of the pure states in the mixture, and hence on all other mixtures of those states, including any non-entangled mixtures. This imposes strong constraints, which do not arise in purifying pure states, and which make effective purification of mixed states generally inconsistent with locality. We have used only the very simplest constraints here in generalising LMP’s results: more detailed analyses would no doubt give stronger results and bounds. These results remove any remaining hope that noise on a quantum channel can be completely countered by individual measurements on mixed states, whether or not the noise is rotationally symmetric. As LMP’s results already strongly suggest, if teleportation and similar schemes are to become workable, they will require either essentially noisefree channels, or algorithms tolerant to noise, or technology that allows efficient collective measurements. [Acknowledgements]{} I am very grateful to Michal Horodecki, Noah Linden, Serge Massar, Martin Plenio and Sandu Popescu for helpful discussions and to the Royal Society for financial support. [99]{} C. Bennett et al., Phys. Rev. Lett. [70]{} (1993) 1895 A. Ekert, Phys. Rev. Lett. [68]{} (1991) 661 C. Bennett, H. Bernstein, S. Popescu and B. Schumacher, Phys. Rev. A [53]{} (1996) 2046 C. Bennett et al., Phys. Rev. Lett. [76]{} (1996) 722 D. Deutsch et al., Phys. Rev. Lett. [77]{} (1997) 2818; erratum, Phys. Rev. Lett. [80]{} (1998) 2022. R. Werner, Phys. Rev. A [40]{} (1989) 4277 N. Linden, S. Massar and S. Popescu, quant-ph/9805001 M. Horodecki, P. Horodecki and R. Horodecki, quant-ph/9807091	{ "pile_set_name": "ArXiv" }
--- abstract: 'We compute invariant for iterated torus knots $K$ for the standard representation $\pi_1(S^3\setminus K)\to\mathbb{Z}$ given by abelianisation. For algebraic knots, this invariant turns out to be very closely related to an invariant of a plane curve singularity, coming from algebraic geometry.' address: 'Institute of Mathematics, University of Warsaw, ul. Banacha 2, 02-097 Warsaw, Poland' author: - Maciej Borodzik date: 29 June 2009 title: A invariant of iterated torus knots --- [^1] Introduction ============ A von Neumann invariant (also called $L^2$–signature, or $L^2$–eta invariant) of a real closed 3–manifold $M$ is a real number $\rho_\phi(M)$ associated to every representation $\phi:\pi_1(M)\to\Gamma$, where $\Gamma$ is any group satisfying PTFA condition (see [@COT1 Definition 2.1]). As a special case, if $K$ is a knot in a 3–sphere, and we consider representations of the fundamental group of the manifold $S^3_0(K)$ (i.e. a zero surgery on $K$), then we can talk about the [$\rho$–invariants]{} of knots. In particular, the representation $ab:\nobreak\pi_1(S^3\setminus K)\to\mathbb{Z}$, given by abelianization, gives rise to the representation $\tilde{ab}:\pi_1(S^3_0(K))\to\mathbb{Z}$ and the corresponding invariant, ${\rho_{ab}}(K)$, turns out to be the integral over normalised unit circle of the [Tristram–Levine signature]{} of a knot. The invariants for knots have been introduced first in [@ChG]. They were then deeply studied in [@COT1]. In their seminal paper, the authors observed that they are a very subtle obstruction for some knots to be slice. Namely, let us be given a knot $K$ bounding a disk $D$ in the ball $B^4$. Let $Y=\partial(B^4\setminus \nu(D))$, where $\nu$ denotes the tubular neighbourhood. Then $Y$ is canonically isomorphic to $S^3_0(K)$, and, for any representation $\phi:\pi_1(Y)\to\Gamma$ that can be extended to $\tilde{\phi}:\pi_1(B^4\setminus\nu(D))\to\Gamma$, the corresponding invariant must vanish. This allows to construct examples of non-slice knots, undistinguishable from slice knots by previously known methods as the [Tristram–Levine signature]{} or the Casson–Gordon invariants. The difficulty of computability of invariants is the cost of their subtlety. Only in the first nontrivial case of the representation given by $ab$, there is a general method of computing this invariant, namely integrating the [Tristram–Levine signature]{}. In papers [@COT2], [@Ha], and others, these invariants were computed also for some other representations of the knot group. But there, the choice of knots is very specific. In this paper we focus on ${\rho_{ab}}$–invariant and compute it for all iterated torus knots. The computation consists of integrating the [Tristram–Levine signature]{}, which is not a completely trivial task. In fact, we do even more: we compute the Fourier transform of the [Tristram–Levine signature]{} function of iterated torus knot. This transform can be expressed by a surprisingly simple formula. In particular, this method can be used to detect knots, which are connected sums of iterated torus knots and which have identical [Tristram–Levine signature]{}. What we find most interesting and striking about ${\rho_{ab}}$ of algebraic knots, is its relation with deep algebro-geometrical invariants of the plane curve singularity. We state this relation, in terms of a uniform bound (see Proposition \[bound\]) but, honestly speaking, we are far from understanding it. Moreover, this relation is not that clear for algebraic links, as we show on an example. The structure of the paper is the following. In Section 2 we recall, how to compute the [Tristram–Levine signature]{} for iterated torus knots and formulate Theorem \[core\]. Then we deduce some of its corollaries. In Section 3 we prove Theorem \[core\]. In Section 4 we recall definitions of some invariants of plane curve singularities and compare them to ${\rho_{ab}}$ for algebraic knots. We end this section by computing the ${\rho_{ab}}$ for a $(d,d)$ torus link, i.e., the link of singularity $x^d-y^d=0$. We apologise the reader for not giving a definition of the invariant. A precise definition from scratch, including necessary definitions of twisted signature of a $4-$manifold, would make this paper twice as long. Instead we refer to [@COT1 Section 5], or, for more detailed treatment, to a book by Lück [@Lu]. We end this introduction by remarking that the $\rho$ invariants were also studied in the context of mixed Hodge structures of hypersurface singularities. The $\eta$ invariant, defined, for instance, in [@Ne1 Section I], is closely related to the ${\rho_{ab}}$ invariant in the case of plane curve singularities. We refer to [@Ne2; @Ne3] for the detailed study of this invariant. Tristram–Levine signature of torus knots ======================================== We begin this section with some definitions, which we give also to fix the notation used in the article. \[iterdef\] A knot is called an iterated torus knot if it arises from an unknot by finitely many cabling operations. An iterated torus knot is of type $(p_1,q_1,\dots,p_n,q_n)$ if it is a $(p_1,q_1)$ cable of $(p_2,q_2)$ cable of …of $(p_n,q_n)$ cable of an unknot. Fore example, a torus knot $T_{p,q}$ is an iterated torus knot of type $(p,q)$. Let $K$ be a knot, $S$ its Seifert matrix. Let $\zeta\in\mathbb{C}$, $\|\zeta\|=1$. The [Tristram–Levine signature]{}, $\sigma_K(\zeta)$ is the signature of the hermitian form given by $$\label{eq:TLSig} (1-\zeta)S+(1-\bar\zeta)S^T.$$ It is well-known that the form is degenerate (i.e. has non-trivial kernel) if and only if $\zeta$ is a root of the Alexander polynomial $\Delta_K$ of $K$. The function $\zeta\to \sigma_K(\zeta)$ is piecewise constant with possible jumps only at the roots of the Alexander polynomial $\Delta_K(\zeta)$. The value of $\sigma_K$ at such root can a priori be different then left or right limit of $\sigma_K$ at that point. However, there are only finitely many such values and they do not influence the integral. As we do not want to take care of this values, we introduce a very handy notion. We shall say that two piecewise-constant functions from a unit circle (or a unit interval) to real numbers are almost equal if they are equal at all but finitely many points. We would like to compute ${\rho_{ab}}$ for an iterated torus knot. We will use Proposition 5.1 from [@COT2], which we can formulate as follows. For any knot $K\subset S^3$ we have $${\rho_{ab}}(K)=\int_0^1\sigma_K(e^{2\pi ix})dx.$$ Therefore, what we have to do, is to compute the integral of the [Tristram–Levine signature]{} for an iterated torus knot. We begin with recalling results from [@Li], where the function $\sigma_K$ is computed for iterated torus knots. Let $p,q$ be coprime positive integers. Let $x$ be in the interval $[0,1]$. Consider the set $$\Sigma=\Sigma_{p,q}=\left\{\frac{k}{p}+\frac{l}{q}\colon 1\le k<p,\,1\le l<q\right\}\subset [0,2]\cap\mathbb{Q}.$$ The function $s_{p,q}(x)$ is defined as $$s_{p,q}(x)=-2\#\Sigma\cap (x,x+1)+\#\Sigma.$$ If $\zeta=e^{2\pi ix}$ is not a root of the polynomial $(t^{pq}-1)(t-1)/(t^p-1)(t^q-1)$, then the [Tristram–Levine signature]{} of the torus knot $T_{p,q}$ at $\zeta$ is equal to $s_{p,q}(x)$. Therefore, computing the invariant of a torus knot boils down to computing the integral of the function $s_{p,q}(x)$. Before we do this, let us show, how one can compute the Tristram–Levine signatures of an iterated torus knot. We shall need another lemma from [@Li]. \[cable\] Let $K$ be a knot and $K_{p,q}$ be the $(p,q)-$cable on $K$. Then for any $\zeta\in\mathbb{C}$, $\|\zeta\|=1$, we have $$\sigma_{K_{p,q}}(\zeta)=\sigma_{K}(\zeta^q)+\sigma_{T_{p,q}}(\zeta).$$ This allows a recursive computation for an iterated torus knot. Namely, let for $r>1$ $$s_{p,q;r}(x)=s_{p,q}(\lfloor rx\rfloor).$$ \[iterated\] Let $K$ be an iterated torus knot of type $(p_1,q_1,\dots,p_n,q_n)$. Let $x\in[0,1]$ be such that $e^{2\pi ix}$ is not a root of the Alexander polynomial of $K$. Denote by $r_k=q_1\dots q_{k-1}$. Then $$\sigma_{K}(e^{2\pi i x})=\sum_{k=1}^ns_{p_k,q_k,r_k}(x).$$ The core of this section is \[core\] For any $\beta\in\mathbb{C}$ which is not an integer divisible by $r$ we have $$\label{eq:formula1} \int_0^1e^{\pi i\beta x}s_{p,q,r}(x)\,dx=\frac{2e^{\pi i\beta/2}\sin\frac{\pi\beta}{2}}{\pi\beta}n_{p,q;r}(\frac{\pi\beta}{2}),$$ where $$n_{p,q;r}(t)={\cot}\frac{t}{pqr}{\cot}\frac{t}{r}-{\cot}\frac{t}{pr}{\cot}\frac{t}{qr}.$$ In particular, by taking a limit $\beta\to 0$ we get $$\int_0^1s_{p,q,r}=-\frac13(p-\frac1p)(q-\frac1q).$$ The function $n_{p,q;r}(t)$ will be called normalised Fourier transform. We prove Theorem \[core\] in Section \[coreproof\]. Now we pass to corollaries. \[corcore\] The ${\rho_{ab}}$ invariant of an iterated torus knot is equal to $$-\frac13\sum_{k=1}^n(p_k-\frac1{p_k})(q_k-\frac{1}{q_k}).$$ Apart of this corollary, Theorem \[core\] has its interest of its own. In fact, it might help to study possible cobordism relations between iterated torus knot. For example, Litherland showed in [@Li], that the connected sum of knots $T_{2,3}$, $T_{3,5}$ and a $(2,5)$-cable on $T_{2,3}$ has the same [Tristram–Levine signature]{} as a $T_{6,5}$. It might be possible that normalised Fourier transforms of torus knots can help studying similar phenomena. This could be done as follows. Let us be given two finite sets $I$ and $J$ of triples of integers $\{p,q,r\}$. Then the difference $$\label{eq:int} {\Delta_{IJ}(x)}:=\sum_{i\in I}s_{p_i,q_i;r_i}(x)-\sum_{j\in J}s_{p_j,q_j;r_j}(x)$$ is almost equal to zero for $x\in[0,1]$, if and only if the difference $$\label{eq:trans} {\widehat{\Delta}_{IJ}(t)}:=\sum_{i\in I}n_{p_i,q_i;r_i}(t)-\sum_{j\in J}n_{p_j,q_j;r_j}(t)$$ is equal to zero on some open subset in $\mathbb{C}$. The ’only if’ part is trivial. To prove the ’if’ part we observe that ${\widehat{\Delta}_{IJ}(t)}\cdot\frac{t}{e^t\sin t}$ is, up to a multiplicative constant, and up to rescaling of the parameter $t$, the Fourier transform of ${\Delta_{IJ}(x)}$, when we extend ${\Delta_{IJ}(x)}$ by $0$ to the whole real line. On the other hand, vanishing of ${\widehat{\Delta}_{IJ}(t)}$ on some open subset of $\mathbb{C}$ implies that it is everywhere $0$. \[cornew\] The condition that ${\Delta_{IJ}(x)}$ is almost equal to zero is equivalent to the fact, that two following conditions are satisfied at once - $\sum_{i\in I}(p_i-\frac{1}{p_i})(q_i-\frac{1}{q_i})=\sum_{j\in J}(p_j-\frac{1}{p_j})(q_j-\frac{1}{q_j})$. - For any $t_0$ such that $\pi r_kt_0\in\mathbb{Z}$ for some $k\in I\cup J$ the residuum at $t_0$ of ${\widehat{\Delta}_{IJ}(t)}$ is zero. If $T$ is the least common multiplier of $p_kq_kr_k$ for $k\in I\cup J$, then $T\pi$ is the period of ${\widehat{\Delta}_{IJ}(t)}$. It follows that the condition (b) involves only finitely many equations. Vanishing of ${\widehat{\Delta}_{IJ}(t)}$ clearly implies (b). The equality in (a) is equivalent to $3\widehat{\Delta}_{IJ}(0)=0$. We shall prove that (b) implies that ${\widehat{\Delta}_{IJ}(t)}$ is bounded on $\mathbb{C}$. This is done as follows. Observe that, in general, ${\widehat{\Delta}_{IJ}(t)}$ can have poles only at such $t_0$’s, that $\pi r_kt_0\in\mathbb{Z}$, for some $k\in I\cup J$. Moreover, these poles are at most of order $1$: in fact, it is a matter of simple computation, that $n_{p,q;r}$ does not have a pole of order $2$. Therefore, condition (b) implies that the ${\widehat{\Delta}_{IJ}(t)}$ extends holomorphically across points $\frac{n}{\pi r_k}$, where $k\in I\cup J$ and $n\in\mathbb{Z}$. As this function is periodic with real period, for any $\delta>0$ it is bounded on the strip $\|\operatorname{Im}t\|\le\delta$ by some constants, depending of course of $\delta$. A uniform bound on ${\widehat{\Delta}_{IJ}(t)}$ for $\|\operatorname{Im}t\|\ge\delta$ results from the standard estimate $\|{\cot}t\|^2\le 1+\frac{1}{(\operatorname{Im}t)^2}$. Hence, if (b) holds, then the function ${\widehat{\Delta}_{IJ}(t)}$ is a bounded holomorphic function, by Liouville’s theorem it is then constant. The condition (a) implies then that it vanishes at $0$, so it is zero everywhere. Proof of Theorem \[core\] {#coreproof} ========================= To make computations at least a bit more transparent, let us first assume that $r=1$. The function $s_{p,q}$ can be expressed as the sum $$s_{p,q}(x)=2{\sum\limits_{\substack{\alpha<1/2\\\alpha\in\Sigma_{p,q}}}}{\chi_{\left(\alpha,1-\alpha\right)}}(x)-2{\sum\limits_{\substack{\alpha\in(1/2,1)\\\alpha\in\Sigma_{p,q}}}}{\chi_{\left(1-\alpha,\alpha\right)}}(x),$$ where ${\chi_{\left(a,b\right)}}$ is the characteristic function of the interval $(a,b)$. Therefore $$\label{eq:whatwewant} \int_0^1s_{p,q}(x)e^{\pi i\beta x}dx=-\frac{2}{\pi i \beta}{\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha\beta}-e^{\pi i\beta(1-\alpha)}.$$ We have $${\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha\beta}=\sum_{k=1}^{p-1}\sum_{\substack{l=1\\ l<q(1-k/p)}}^{q-1}e^{\pi i\beta(k/p+l/q)}.$$ The internal sum on the right hand side is the sum of geometric series (here we use the assumption that $\beta$ is not an integer) and can be expressed as $$\frac{1}{1-e^{\pi i\beta/q}}(e^{\pi i\beta k/p}-e^{\pi i\beta(k/p+l_k/q)}),$$ where $l_k$ satisfies $$k/p+l_k/q>1>k/p+(l_k-1)/q.$$ So we have $$\label{eq:twosums} {\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha\beta}=\frac{\sum\limits_{k=1}^{p-1}e^{\pi i\beta k/p}-\sum\limits_{k=1}^{p-1}e^{\pi i\beta(k/p+l_k/q)}}{1-e^{\pi i\beta/q}}.$$ The first sum in the denominator is again a geometric series. As to the second one, let us denote $$\gamma_k=k/p+l_k/q.$$ Then $\gamma_k$’s have the following obvious properties - $\gamma_k$’s are all different; - $1+\frac{1}{pq}\le\gamma_k\le 1+\frac{p-1}{pq}$; - each $\gamma_k$ is of the form $1+a_k/pq$ with $a_k$ an integer. By the Dirichlet principle the set $\{\gamma_1,\dots,\gamma_{p-1}\}$ is the same as the set $\{1+1/pq,\dots,1+(p-1)/pq\}$. Therefore, the second sum in the denominator , upon reordering, can be expressed as $$\sum_{m=1}^{p-1}e^{\pi i\beta (1+m/pq)},$$ which again is a geometric series. Putting things all together we get $$\label{eq:universalsum} {\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha\beta}=\frac{1}{1-e^{\pi i\beta/q}}\left(\frac{e^{\pi i\beta/p}-e^{\pi i\beta}}{1-e^{\pi i\beta/p}}- \frac{e^{\pi i\beta(1+1/pq)}-e^{\pi i\beta(1+1/q)}}{1-e^{\pi i\beta/pq}}\right).$$ On the other hand, we have $${\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i(1-\alpha)\beta}=e^{\pi i \beta}{\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha(-\beta)},$$ and the sum on the right hand side is just with $-\beta$ substituted in place of $\beta$. Substituting this into , and applying the formula $e^{\pi i a}-e^{\pi ib}=2ie^{\pi i(a+b)/2}\sin\frac{\pi(a-b)}{2}$ several times, we arrive finally at $$\int_0^1 s_{p,q}(x)e^{\pi i\beta}dx=\frac{2e^{\pi i\beta/2}\sin{\frac{\pi\beta}{2}}}{\pi\beta}({\cot}\frac{\pi\beta}{2pq}{\cot}\frac{\pi\beta}{2}-{\cot}\frac{\pi\beta}{2p}{\cot}\frac{\pi\beta}{2q}).$$ To conclude the proof in the case $r>1$ we observe that $$\begin{aligned} s_{p,q;r}(x)=&2{\sum\limits_{\substack{\alpha<1/2\\\alpha\in\Sigma_{p,q}}}}\sum_{k=0}^{r-1}{\chi_{\left(\frac{\alpha+k}{r},\frac{1-\alpha+k}{r}\right)}}(x)+\\ -&2{\sum\limits_{\substack{\alpha\in(1/2,1)\\\alpha\in\Sigma_{p,q}}}}\sum_{k=0}^{r-1}{\chi_{\left(\frac{1-\alpha+k}{r},\frac{\alpha+k}{r}\right)}}(x)\end{aligned}$$ Thus $$\label{eq:involvedsum} \int_0^1s_{p,q;r}e^{\pi i\beta x}=\frac{-2}{\pi i\beta}{\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}\sum_{k=0}^{r-1}e^{\pi i\beta(\alpha/r+k/r)}-e^{\pi i\beta(1-\alpha/r-k/r)}.$$ Now, for fixed $\alpha$ we have $$\sum_{k=0}^{r-1}e^{\pi i\beta(\alpha/r+k/r)}=e^{\pi i\alpha(\beta/r)}\sum_{k=0}^{r-1}e^{\pi i\beta k/r}=e^{\pi i\alpha(\beta/r)}\frac{1-e^{\pi i\beta}}{1-e^{\pi i\beta/r}}.$$ Therefore, returning to we get $${\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}\sum_{k=0}^{r-1}e^{\pi i\beta(\alpha/r+k/r)}=\frac{1-e^{\pi i\beta}}{1-e^{\pi i\beta/r}}{\sum\limits_{\substack{\alpha<1\\\alpha\in\Sigma_{p,q}}}}e^{\pi i\alpha(\beta/r)}.$$ We can use again, substituting $\beta/r$ in place of $\beta$. Similarly we can deal with a sum of terms $e^{\pi i\beta(1-\alpha/r-k/r)}$. Now straightforward but long computations yield the formula . Relation with algebraic invariants ================================== The setup in this section is the following. Let $(C,0)\subset\mathbb{C}^2$ be germ of a plane curve singularity with one branch. This means that there exists a local parametrisation $C=(x(t),y(t))$, with $x$ and $y$ analytic functions in one variable with $x(0)=y(0)=0$. Let us assume that the Puiseux expansion of $y$ in fractional powers of $x$ written is the multiplicative form (see [@EN page 49]) is $$y=x^{q_1/p_1}(a_1+x^{q_2/p_1p_2}(a_2+\ldots+x^{q_s/p_1p_2p_3\dots p_s}(a_s+\ldots))),$$ with $q_1>p_1$ (otherwise we switch $x$ with $y$), $\gcd(q_i,p_i)=1$ and $p_i,q_i>0$. The pairs $(p_1,q_1),\dots,(p_n,q_n)$ are called characteristic pairs (or Newton pairs) of the singularity. They completely determine the topological type of the singular point. Put $a_1=q_1$ and $a_{k+1}=p_{k+1}p_ka_k+q_{k+1}$. Then the link of the singularity $(C,0)$ is an iterated torus knot. More precisely, it is a $(p_n,a_n)$ cable on $(p_{n-1},a_{n-1})$ cable on $\dots$ on $(p_1,a_1)$ torus knot The ordering of cables in [@EN] is different than in [@Li]. According to Definition \[iterdef\], the link of the singuarity $(C,0)$ above would be an iterated torus knot of type $(p_n,q_n,p_{n-1},q_{n-1},\dots,p_1,q_1)$. The ${\rho_{ab}}$ invariant of an algebraic knot is equal to $$\label{eq:tlint} {\rho_{ab}}=-\frac13\sum_{k=1}^n\left(a_kp_k-\frac{a_k}{p_k}-\frac{p_k}{a_k}+\frac{1}{p_ka_k}\right).$$ It is on purpose that we wrote formula in a different shape that in Corollary \[corcore\]. Let us now resolve the above singularity. This means that we have a map $\pi:(X,E)\to (U,0)$, where $U$ is a neighbourhood of $0$ in $\mathbb{C}^2$, $E$ is the exceptional divisor and $X$ is a complex surface. We require the strict transform $C'$ to be smooth, $C'\cup E$ to have only normal crossings as singularities and the resolution to be minimal, so that we cannot blow-down any exceptional curve without violating one of the two above assumptions. Put $K=K_X$ the canonical divisor on $X$ and let $D=C'+E_{red}$. Here, the subscript ’red’ means that we take a reduced divisor, i.e. coefficients with all components are equal to $1$. Using the notation from this section, we have $$\label{eq:Mbar} (K+D)^2=a_1p_1-{\genfrac{\lceil}{\rceil}{}{1}{a_1}{p_1}}-{\genfrac{\lceil}{\rceil}{}{1}{p_1}{a_1}}+\sum_{k=2}^n\left(a_kp_k-{\genfrac{\lceil}{\rceil}{}{1}{a_k}{p_k}}\right),$$ where $(K+D)^2$ denotes the self-intersection of the divisor $K+D$, and $\lceil x\rceil=\min(n\in\mathbb{Z},\,n\ge x)$. On the one hand $(K+D)^2$ has a very natural meaning. Namely, at least for unibranched singularities, this is the sum of the Milnor number $\mu$ and so called $\bar{M}$ number of singularity. The latter, introduced in [@Or] and studied in [@BZ], can be interpreted as a parametric codimension of a singular point, i.e. the number of locally independent conditions, which are imposed on a curve given in parametric form, by the appearance of the singularity of given topological type. On the other hand there is an apparent similarity of left hand sides of formulae and . To make it even more similar, let us take a Zariski–Fujita [@Fuj] decomposition of the divisor $K+D$. We have then $$K+D=H+N$$ with $H$ nef (its intersection with any algebraic curve in $X$ is non-negative), $N$ effective and $N^2<0$, $H\cdot N'=0$ for any $N'$ supported on $\operatorname{supp}N$. $$\label{eq:H} H^2=a_1p_1-\frac{a_1}{p_1}-\frac{p_1}{a_1}+\sum_{k=2}^n\left(a_kp_k-\frac{a_k}{p_k}\right).$$ In the case of unibranched singularity, the quantity $H^2$ is the sum of Milnor number and so called $M$-number (without a bar) of singular point. Its importance lies in the fact that the sum of $M$-numbers of all singular points of an algebraic curve in $\mathbb{C}P^2$ can be bounded from above by global topological data of the curve, as genus and first Betti number (see [@BZ]). These bounds involve very deep Bogomolov–Miyaoka–Yau inequality from algebraic geometry. Thus the following result seem to be a very mysterious and shows a deep link between knot theory and algebraic geometry. \[bound\] Let ${\rho_{ab}}$ be the integral of the [Tristram–Levine signature]{} of an algebraic knot (see ) and $H^2$ be like in . Then $$0<-3{\rho_{ab}}-H^2<\frac29.$$ It easy to observe that $$\Delta:=-3{\rho_{ab}}-H^2=\frac{1}{a_1p_1}+\sum_{k=2}^n\left(\frac{1}{a_kp_k}-\frac{p_k}{a_k}\right).$$ On the one hand $$\Delta\le\sum_{k=1}^n\frac{1}{a_kp_k}.$$ Recall that $a_{k+1}=a_kp_{k+1}p_k+q_{k+1}$, so $a_{k+1}p_{k+1}>a_kp_kp_{k+1}^2\ge 4a_kp_k$. Hence $$\Delta\le\frac{1}{a_1p_1}\sum_{k=0}^{n-1}\frac{1}{4^k}<\frac{4}{3a_1p_1}.$$ But $a_1p_1\ge 6$, so one inequality is proved. To prove in the second one, let us reorganise terms of $\Delta$ as follows $$\Delta=\sum_{k=1}^{n-1}\left(\frac{1}{a_kp_k}-\frac{p_{k+1}}{a_{k+1}}\right)+\frac{1}{a_np_n}.$$ But $$\frac{1}{a_kp_k}-\frac{p_{k+1}}{a_{k+1}}=\frac{1}{a_kp_k}-\frac{p_{k+1}}{a_kp_kp_{k+1}+q_{k+1}}>\frac{1}{a_kp_k}-\frac{p_{k+1}}{a_kp_kp_{k+1}}=0.$$ We end up the chapter with the simplest example of multibranched singularity, i.e. with a singularity defined locally by $x^d-y^d=0$ with $d\ge 2$. Its link at singularity is the torus link $T_{d,d}$. Let us consider a set $$\Sigma_d=\{\frac{i}{d}+\frac{j}{d},\,1\le i,j\le d-1\}.$$ Here the element $k/d$ appears in $\Sigma_d$ precisely $d-1-\|d-1-k\|$ times, according to possible presentations $k=i+j$, $1\le i,j\le d-1$. Let $s_d(x)$ be the function computing the elements of $\Sigma_d$ in $(x,x+1)$ with a ’$-$’ sign and the others with ’$+$’ sign. Then $s_d$ is almost equal to the [Tristram–Levine signature]{} of link $T_{d,d}$. We have the formula $$s_d=2\sum_{k<d/2}(k-1){\chi_{\left(\frac{k}{d},\frac{d-k}{d}\right)}}-2(k-1)\sum_{k>d/2}{\chi_{\left(\frac{d-k}{d},\frac{k}{d}\right)}}-(d-1).$$ The final term, $-(d-1)$, comes from the $d-1$ elements of the set $\Sigma_d$ of type $d/d$. They belong to any interval $(x,x+1)$. Thus, the integral of $s_d$ is equal to $$\int_0^1s_d=-2\sum_{k=1}^{d-1}(k-1)\frac{2k-d}{d}-(d-1).$$ But an elementary calculus shows that $$\sum_{k=1}^{d-1}(k-1)(2k-d)= \frac{d(d-1)(d-2)}{6}.$$ Hence $$\int_0^1s_d=-\frac13(d-1)(d+1).$$ On the other hand, in order to resolve the singularity of $C$ we need only one blow-up. The exceptional divisor $E$ consists of single rational curve with $E^2=-1$. Then $K=K_X=\alpha E$ and $C'=\beta E$ (as $E$ spans second (co)homology of blown-up space) and $K(K+E)=-2$ by genus formula, so $K=E$ and $C'\cdot E=d$, so $C'=-d\cdot E$. Thus $K+D=K+C'+E=(2-d)E$. Moreover, this divisor is nef, so its Zariski–Fujita decomposition is trivial, $H=(2-d)E$, $N=0$, so in this case $$H^2=(d-2)^2.$$ This shows that, in case of general links, a trivial analogue of Proposition \[bound\] does not hold. The author is very grateful to Tim Cochran and Stefan Friedl for explaining the rudiments of invariants. He wishes also to express his thanks to Andràs Némethi for various discussions on the subject. [999]{} M. Borodzik, Morse theory for plane algebraic curves, preprint, University of Warsaw, Warsaw 2009, available at `http://www.mimuw.edu.pl/~mcboro/pliki/artykuly/signature.pdf` M. Borodzik, H. Żoładek Complex algebraic plane curves via Poincaré–Hopf formula. III. Codimension bounds, J. Math. Kyoto Univ. 48 (2008), no. 3, 529–570. J. Cheeger, M. Gromov, Bounds on the von Neumann dimension of $L\sp 2$-cohomology and the Gauss–Bonnet theorem for open manifolds. J. Diff. Geom. 21(1985), no. 1, 1–34. T. Cochran, K. Orr, P. Teichner, Knot concordance, Whitney towers and $L\sp 2$-signatures, Ann. of Math. 157(2003), no. 2, 433–519. T. Cochran, K. Orr, P. Teichner Structure in the classical knot concordance group, Comment. Math. Helv. 79(2004), no. 1, 105–123. D. Eisenbud, W. Neumann, Three-dimensional link theory and invariants of plane curve singularities, Annals Math. Studies 110, Princeton University Press, Princeton, 1985. T. Fujita, On the topology of non-complete algebraic surfaces, J. Fac. Sci. Univ. Tokyo (Ser. 1A) 29 (1982), 503–566. S. Harvey, Homology cobordism invariants and the Cochran-Orr-Teichner filtration of the link concordance group, Geom. Topol. 12(2008), no. 1, 387–430. R. A. Litherland, Signatures of iterated torus knots in Topology of low-dimensional manifolds (Proc. Second Sussex Conf., Chelwood Gate, 1977), pp. 71–84, Lecture Notes in Math., 722, Springer, Berlin, 1979. W. Lück $L\sp 2$-invariants: theory and applications to geometry and $K$-theory, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics, 44. Springer-Verlag, Berlin, 2002. A. Némethi, The eta-invariant of variation structures. I. Topology and its Applications, 67(1995), 95–111. A. Némethi, Dedekind sums and the signature of $f(x,y)+z^N$. I. Selecta Mathematica, New series, 4(1998), 361–376. A. Némethi, Dedekind sums and the signature of $f(x,y)+z^N$. II. Selecta Mathematica, New Series, 5(1999), 161–179. S. Yu. Orevkov, On rational cuspidal curves. I. Sharp estimate for degree via multiplicities, Math. Ann. 324 (2002), 657–673. S. Yu. Orevkov, M. G. Zaidenberg, On the number of singular points of plane curves, in: Algebraic Geometry, Saithana 1995. [^1]: The author is partially supported by the Foundation for Polish Science	{ "pile_set_name": "ArXiv" }
--- abstract: 'Mars Global Surveyor measurements revealed that the Martian crust is strongly magnetized in the southern hemisphere while the northern hemisphere is virtually void of magnetization. Two possible reasons have been suggested for this dichotomy: A once more or less homogeneously magnetization may have been destroyed in the northern hemisphere by, for example, resurfacing or impacts. The alternative theory we further explore here assumes that the dynamo itself produced a hemispherical field [@Stanley2008; @Amit2011]. We use numerical dynamo simulations to study under which conditions a spatial variation of the heat flux through the core-mantle boundary (CMB) may yield a strongly hemispherical surface field. We assume that the early Martian dynamo was exclusively driven by secular cooling and we mostly concentrate on a cosine CMB heat flux pattern with a minimum at the north pole, possibly caused by the impacts responsible for the northern lowlands. This pattern consistently triggers a convective mode which is dominated by equatorially anti-symmetric and axisymmetric (EAA, @Landeau2011) thermal winds. Convective up- and down-wellings and thus radial magnetic field production then tend to concentrate in the southern hemisphere which is still cooled efficiently while the northern hemisphere remains hot. The dynamo changes from an $\alpha^2$- for a homogeneous CMB heat flux to an $\alpha\Omega$-type in the hemispherical configuration. These dynamos reverse on time scales of about 10 kyrs. This too fast to allow for the more or less unidirectional magnetization of thick crustal layer required to explain the strong magnetization in the southern hemisphere.' author: - \| Wieland Dietrich$^{1,2}$ & Johannes Wicht$^{2}$\ [$ ^2$Max-Planck-Institut für Sonnensystemforschung, ]{}\ [Max-Planck-Strasse 2, 37191 Katlenburg-Lindau, Germany]{}\ [$ ^1$Institute for Geophysics,]{}\ [University of Göttingen, 37073 Göttingen, Germany]{} bibliography: - 'mypaper.bib' title: 'A hemispherical dynamo model : Implications for the Martian crustal magnetization' --- Introduction ============ Starting in 1998 the space probe Mars Global Surveyor (MGS) delivered vector magnetic field data from orbits between $185$ and $400$ km above the planets surface [@Acuna1999]. The measurements reveal a strong but heterogeneous crustal magnetization [@Acuna1999; @Connerney2001]. The more strongly magnetized rocks are mainly localized in the southern hemisphere where the crust is thick and old. The northern hemisphere is covered by a younger and thinner crust which is much weaker magnetized. Two alternative types of scenarios are discussed to explain this dichotomy. One type explores the possibility that an originally more or less homogeneous magnetization was partly destroyed by resurfacing events after the demise of the internal dynamo. Based on the fact that the Hellas and Argyre impact basins are largely void of magnetization, @Acuna2001 conclude that the dynamo stopped operating in the early Noachian before the related impact events happened roughly $3.7-4$ Gyrs ago. Volcanic activity and crustal spreading are two other possibilities to explain the lack of strong magnetization in certain surface areas [@Lillis2008b; @Mohit2004], in particular the northern hemisphere after the dynamo cessation. The alternative scenario explains the dichotomy by an ancient Martian dynamo that inherently produced a hemispherical magnetic field. Numerical dynamo simulations by @Stanley2008 and @Amit2011 show that this may happen when more heat is allowed to escape the core through the southern than through the northern core mantle boundary (CMB). Such north/south asymmetry can for example be caused by larger impacts or low-degree mantle convection [@Roberts2006; @Keller2009; @Yoshida2006]. Due to depth-dependent viscosity and a possible endothermic phase transition [@Harder1996] Martian mantle convection may be ruled in an extreme case by one gigantic plume typically evoked to explain the dominance of the volcanic Tharsis region. However, the single plume convection might have developed after the dynamo ceased. Due to the hotter temperature of the rising material the CMB heat flux can be significantly reduced under such a plume. Though Tharsis is roughly located in the equatorial region it could nevertheless lead to magnetic field with the observed north-south symmetry, as we will show in the following. The possible effects of large impacts on planets and the dynamo in particular are little understood. @Roberts2009 argue that impacts locally heat the underlying mantle and thereby lead to variations in the CMB heat flux. Large impacts may also cause a demise of the dynamo by reducing the CMB heat flux below the value where subcritical dynamo action is still possible [@Roberts2009]. The deposition of heat in the outer parts of the core by impact shock waves could lead to a stably stratified core and thereby also stop dynamo action [@Arkani-Hamed2010] for millions of years until the heat has diffused out of the core. If the iron content of the impactor is large enough it may even trigger a dynamo [@Reese2010]. The thermal state of the ancient Martian core is rather unconstrained [@Breuer2010]. Analysis of Martian meteorites suggests a significant sulphur content and thus a high core melting temperature [@Dreibus1985]. Mars may therefore never have grown a solid inner core, an assumption we also adopt here [@Schubert1990; @Breuer2010]. The ancient Martian dynamo was then exclusively driven by secular cooling and radiogenic heating and has stopped operating when the CMB heat flux became subadiabatic [@Stevenson1983]. Run-away solidification or light element saturation may explain the dynamo cessation in the presence an inner core. The geodynamo, on the other hand, is predominantly driven by the latent heat and light elements emanating from the growing inner core front. Secular cooling and radiogenic heating is typically modeled by homogeneously distributed internal buoyancy sources while the driving associated to inner core growth is modeled by bottom sources [@Kutzner2000]. These latter sources have a higher Carnot efficiency and would likely have kept the Martian dynamo alive if the planet would have formed an inner core. Several authors have explored the influence of a CMB heat flux pattern on dynamos geared to model Earth and report that they can cause hemispherical variations in the secular variation [@Bloxham2000; @Christensen2003; @Amit2006], influence the reversal behavior [@Glatzmaier1999; @Kutzner2004] or lead to inhomogeneous inner core growth [@Aubert2008]. However, the effects where never as drastic as those reported by @Stanley2008 or @Amit2011. In the work of @Amit2011 the reason likely is the increased susceptibility of internally driven dynamos to the thermal CMB boundary condition [@Hori2010]. @Stanley2008 retain bottom driving and employed a particularly strong heat flux variation to enforce a hemispherical field Here, we follow @Amit2011 in exploring the effects of a simple sinusoidal CMB heat flux variation on a dynamo model driven by internal heat sources. The main scope of this paper is to understand the particular dynamo mechanism, to explore its time dependence, and to extrapolate the results to the Martian situation. [Section \[Model\]]{} introduces our model, whereas [section \[hemisphericalaction\]]{} describes the effects of the CMB heat flux anomaly on the convection and the induction process. In [section \[applymars\]]{} we explore the applicability to the ancient Martian dynamo. The paper closes with a discussion in [section \[Discussion\]]{}. Numerical Model {#Model} =============== Using the MagIC code [@Wicht2002; @Christensen2007], we model the Martian core as a viscous, electrically conducting and incompressible fluid contained in a rotating spherical shell with inner core radius $r_{icb}$ and outer radius $r_{cmb}$. Conservation of momentum is described by the dimensionless Navier-Stokes equation for a Boussinesq fluid: $$E \left( \frac{\partial \vec u}{\partial t} + \vec u \cdot \vec \nabla \vec u \right) = -\vec \nabla \Pi + E \nabla^2 \vec u - 2 \hat z \times \vec u + \frac{Ra E}{Pr} \frac {\vec r}{r_{cmb}} T + \frac 1 {Pm} (\vec \nabla \times \vec B ) \times \vec B \, \label{nseq}$$ where $\vec u$ is the velocity field, $\Pi$ the generalized pressure, $\hat z$ the direction of the rotation axis, $T$ the super-adiabatic temperature and $\vec B$ the magnetic field. The conservation of energy is given by $$\label{heat} \frac {\partial T}{\partial t} + \vec u \cdot \vec \nabla T= \frac 1 {Pr} \nabla^2 T + \epsilon \\,$$ where $\epsilon$ is a uniform heat source density. The conservation of magnetic field is given by the induction equation $$\frac {\partial \vec B}{\partial t} = \vec \nabla \times \left( \vec u \times \vec B \right) + \frac 1 {Pm} \nabla^2 \vec B \ . \label{ind_eq}$$ We use the shell thickness $D=r_{cmb}-r_{icb}$ as length scale, the viscous diffusion time $D^2/ \nu$ as time scale and $(\rho \mu \lambda \Omega)^{1/2}$ as the magnetic scale. The mean superadiabatic CMB heat flux density $q_0$ serves to define the temperature scale $q_0 D / c_p\rho\kappa$. Here, $\nu$ is the viscous diffusivity, $\rho$ the constant background density, $\mu$ the magnetic permeability, $\lambda$ the magnetic diffusivity, $\Omega$ the rotation rate, $\kappa$ the thermal diffusivity and $c_p$ the heat capacity. Three dimensionless parameters appear in the above system: the Ekman number $E=\nu / \Omega D^2 $ is a measure for the relative importance of viscous versus Coriolis forces while the flux based Rayleigh number $Ra=\alpha g_0 \|q_0\| D^4 / \rho c_p\kappa^2\nu$ is a measure for the importance of buoyancy. The Prandtl number $P=\nu / \kappa$ and the magnetic Prandtl number $Pm=\nu / \lambda$ are diffusivity ratios. An inner core with $r_{icb}/r_{cmb}=0.35$ is retained for numerical reasons [@Hori2010], but to minimize its influence the heat flux from the inner core is set to zero. The secular cooling and radiogenic driving is modeled by the homogeneous heat sources $\epsilon$ appearing in \[heat\] [@Kutzner2000]. Furthermore we assume an electrically insulating inner core to avoid an additional sink for the magnetic field. We use no-slip, impermeable flow boundary conditions and match $\vec B$ to a potential field at the outer and inner boundary. The results by @Hori2010 and @Aubert2009 suggest that this is a fair approximation to model a dynamo without inner core since an additional reduction of the inner core radius has only a minor impact. The effective heat source $\epsilon$ is chosen to balance the mean heat flux $q_0$ through the outer boundary: $$\begin{aligned} 4 \pi^2 r_{cmb}^2\;q_0 &=-Pr \frac 4 3 \pi (r_{cmb}^3-r_{icb}^3)\;\epsilon \ .\end{aligned}$$ Note that $q_0$ is generally negative. The CMB heat flux pattern is modeled in terms of spherical harmonic contributions with amplitude $q_{lm}$, where $l$ is the degree and $m$ the spherical harmonic order. Here we mostly concentrate on a variation along colatitude ${\vartheta}$ of the form $q_{10} \cos{{\vartheta}}$ with negative $q_{10}$ so that the minimum (maximum) heat flux is located at the north (south) pole. This is the most simple pattern to break the north/south symmetry and has first been used by @Stanley2008 in the context of Mars. We also explore the equatorially symmetric disturbance $q_{11} \sin{{\vartheta}} \sin{\phi}$, which breaks the east/west symmetry, and a superposition of $q_{10}$ and $q_{11}$ to describe a cosine disturbance with arbitrary tilt angle $$\begin{aligned} \alpha = \operatorname{arctan} (\|q_{11}\| / \|q_{10}\|)\;\;. \label{defalpha}\end{aligned}$$ In the following we will characterize the amplitude of any disturbance by its maximum relative variation amplitude in percent $$\begin{aligned} g = 100 \% \max(\|\delta q\|) / \|q_0\|\;\;.\end{aligned}$$ We vary $g$ up to $300\%$, the value used in @Stanley2008. For variations beyond $100\%$ the heat flux becomes subadiabatic in the vicinity of the lowest flux. For severely subadiabatic cases this may pose a problem since dynamo codes typically solve for small disturbances around an adiabatic background state [@Braginsky1995]. The possible implication of this have not been explored so far and we simply assume that the model is still valid. Since the main effects described below do not rely on $g>100\%$ this is not really an issue here. The hemispherical mode triggered by the heat flux variation is dominated by equatorially anti-symmetric and axisymmetric thermal winds [@Landeau2011]. Classical columnar convection found for a homogeneous heat flux, on the other hand, is predominantly equatorial symmetric and non-axisymmetric, at least at lower Rayleigh numbers. We thus use the relative equatorial anti-symmetric and axisymmetric (EAA) kinetic energy to identify the hemispherical mode: $$\begin{aligned} {\mbox{$\mathcal{A}$}}= \frac { \sum_{l_{odd}, m=0} E_{lm} } { \sum_{lm} E_{lm} } \ , \label{defeaa}\end{aligned}$$ where $E_{lm}$ is the rms kinetic energy carried by a flow mode of spherical harmonic degree $l$ and order $m$. For a homogeneous outer boundary heat flux the dynamo is to first order of an $\alpha^2$-type where poloidal and toroidal fields are produced in the individual convective columns [@Olson1999]. As the hemispherical flow mode takes over, the $\Omega$-effect representing the induction of axisymmetric toroidal magnetic field via axisymmetric shearing becomes increasingly important. We measure its relative contribution to toroidal field production by $${\mbox{$\mathcal{O}$}}= \frac { \left[(\bar{\vec{B}}\cdot \vec \nabla )\; \bar{u}_\phi \right]_{\mbox{\scriptsize tor}}^{\mbox{\scriptsize rms}}} { \left[(\vec B\cdot \vec \nabla ) \vec u \right]^{\mbox{\scriptsize rms}}_{\mbox{\scriptsize tor}}} \ . \label{defomega}$$ The lower index tor and upper index rms indicate that rms values of the toroidal field production in the shell are considered. For quantifying to which degree the Martian crustal magnetization and the poloidal magnetic fields in our dynamo simulations are concentrated in one hemisphere we use the hemisphericity measure $${\mbox{$\mathcal{H}$}}(r)= \left\vert \frac { B_r ^{N}(r) - B_r ^{S}(r) } { B_r ^{N}(r) + B_r ^{S}(r) } \right\vert \ , \label{Hdef}$$ where $B_r ^{N}(r)$ and $B_r ^{S}(r)$ are the surface integral over the unsigned radial magnetic flux in the northern and southern hemispheres, respectively. According to this definition both a purely equatorially symmetric and a purely equatorially anti-symmetric field yield $\mathcal{H}=0$. For $\mathcal{H}=1$ the flux is strictly concentrated in one hemisphere which requires a suitable combination of equatorially symmetric and anti-symmetric modes [@Grote2000]. A potential field extrapolation is used to calculate $\mathcal{H}$ for radii above $r_{cmb}$, for example the surface hemisphericity . [Table \[Tab1\]]{} provides an overview of the different parameter combinations explored in this study along with $\mathcal{A}$, $\mathcal{O}$, , , the Elsasser number $\Lambda=B^2 / \mu_0 \lambda \rho \Omega$ and the field strength at the Martian surface in nano Tesla $\bar{B}_{sur}$. Column $14$ lists the respective (if present) dimensionless oscillation frequencies given in units of magnetic diffusion time. We mostly focus on simulations at $E=10^{-4}$ where the relatively moderate numerical resolution still allows to extensively explore the other parameters in the system. A few cases at $E=3\times10^{-5}$ and $E=10^{-5}$ provide a first idea of the Ekman number dependence.The last line in table \[Tab1\] gives estimates for the Rayleigh, Ekman and magnetic Prandtl number of Mars, based on the (rather uncertain) properties of Mars [@Morschhauser2011]. $E$ $Ra$ $Pm$ $g$ $\alpha$ $Rm$ $\Lambda$ $\mathcal{A}$ $\mathcal{O}$ $\bar{B}_{sur}$ $\mathcal{H}_{sur}$ $\mathcal{H}_{cmb}$ freq. ------- ------- ------ ----- ---------- ------- ------- ----------- --------------- --------------- ----------------- --------------------- --------------------- ------- 1e-4 7e6 2 0 0 54.6 0.24 - 2.24e-5 - - - - - 100 0 133.5 122.6 0.1 0.85 0.32 803.2 0.1 0.21 - 5 0 0 117.1 3.95 9.79 1.93e-3 0.21 62510 4e-4 3e-3 - 100 0 326.9 301.5 0.97 0.85 0.66 1469 0.1 0.35 ? 200 0 449.5 417.5 - 0.84 - - - - - 100 90 230.8 - 2.22 5e-3 0.20 7264 - - 18.84 2.1e7 2 0 0 105.9 6.42 4.95 3.64e-3 0.24 64635 1.0e-3 0.03 - 60 0 178.1 149.8 6.24 0.72 0.63 14017 0.12 0.38 - 80 0 228.3 189.7 1.06 0.74 0.53 1684 0.17 0.61 10.69 100 0 247.6 213.6 0.15 0.73 0.58 1001 0.21 0.55 ? 200 0 313.3 272 0.19 0.76 0.77 689 0.79 0.8 56.27 100 90 160.2 - 2.64 6.6e-3 0.18 699 - - ? 4e7 1 100 0 169.3 143.1 0.2 0.73 0.53 922 0.21 0.22 - 200 0 206.5 171.5 - 0.7 - - - - ? 2 0 0 155.6 3.58 6.26 2e-3 0.18 58349 3.0e-3 0.05 - 60 0 283.4 252.4 4.18 0.59 0.65 6154 0.26 0.6 13.96 100 0 338.1 307.2 2.64 0.78 0.76 2219 0.52 0.77 40.83 200 0 409.8 350 1.16 0.74 0.75 1628 0.74 0.75 62.6 100 90 217.3 - 1.47 8.4e-3 0.21 10071 - - 20.77 200 90 226.5 - 5.41 4.1e-3 0.24 10934 - - - 5 100 0 837.5 749.5 6.83 0.81 0.8 3493 0.7 0.65 79.87 8e7 2 0 0 228.7 6.4 7.06 9e-3 0.18 60036 3.3e-3 0.07 - 60 0 400.2 343.6 5.5 0.74 0.65 9268 0.41 0.72 26.97 100 0 457.6 403.2 2.97 0.79 0.73 3240 0.68 0.73 61.3 100 90 297.5 - 3.73 6.4e-3 0.20 16424 - - 24.5 2e8 1 0 0 251.8 54.4 - 0.05 - 0 - - - 60 0 309.9 230.1 2.14 0.63 0.56 6095 0.15 0.56 - 100 0 343.5 276.3 0.34 0.64 0.65 1119 0.42 0.72 24.32 100 90 270.4 - 0.77 4e-3 0.21 7954 - - 17.95 3e-5 1e8 2 0 0 137.1 2.88 7.68 5.7e-3 0.19 80508 1e-3 0.03 - 60 0 210.7 48.2 12.31 0.5 0.53 25567 0.1 0.23 - 100 0 360.4 324.2 5.07 0.81 0.67 5157 0.12 0.46 10.34 100 90 199.5 - 1.4 3.1e-3 0.16 10657 - - ? 3e-5 4e8 2 0 0 316.9 6.62 12.2 5.5e-3 0.26 71085 2.1e-3 0.07 - 60 0 517.1 401.8 29.9 0.64 0.49 30528 0.24 0.51 - 100 0 769.1 682 6.04 0.76 0.70 4584 0.62 0.76 87.6 1e-5 4e8 2 0 0 234.9 586 13.58 0.01 0.18 88763 6e-3 0.09 - 50 0 292.5 146.2 19.07 0.18 0.23 69618 0.03 0.22 - 100 0 441.1 376.4 41.7 0.41 0.26 42194 0.07 0.30 - Mars 3e-15 2e28 1e-6 ? ? 500? ? ? ? ? 5000 0.45 ? ? : Selection of runs performed. Rm - magnetic Reynolds number, $\Lambda$ - Elsasser number of rms field in full core shell, EAA - relative equatorially antisymmetric and axisymmetric kinetic energy, $\omega^\ast$ - relative induction of toroidal field by shearing,$\|B\|_{sur}$ - time averaged field intensity at the Martian surface, $\mathcal{H}_{sur}$ and $\mathcal{H}_{cmb}$ - hemisphericity at the surface and CMB, freq. - rough frequency ($2\pi Pm / \tau_{vis}$) if present. Decaying solutions are marked with ‘-’ in the Elsasser number, stationary dynamos with ‘-’ in the frequency. If not a single frequency could be extracted ‘?’ is used.[]{data-label="Tab1"} Hemispherical Solution {#hemisphericalaction} ====================== We start with discussing the emerging hemispherical dynamo mode promoted by the $l=1,m=0$ heat flux pattern with minimal (maximal) heat flux at the north (south) pole concentrating on cases at $E=10^{-4}$, $Ra=4.0 \times 10^{7}$ and $Pm=2$. The study of @Landeau2011 reports the emergence of the equatorially anti-symmetric and axisymmetric convective mode if the Rayleigh number is sufficiently high. Note, that the authors used a homogeneous heat flux condition at the outer boundary. There the amplitude of the hemispherical convection becomes of equal strength compared to the columnar type in the pure hydrodynamic case and is even more dominant if the magnetic field can act on the flow [@Landeau2011]. Hemispherical Convection ------------------------ ![Zonal average of the temperature (left plots), zonal flow with meridional circulation contours (middle plots) and toroidal field with poloidal field line contours (right plots) for columnar convection dominated and magnetic dipolar reference case (left) and a typical hemispherical dynamo solution with the strong EAA symmetry in the flow (right). See the online-version of the article for the color figure.[]{data-label="zonal"}](fig1.pdf){width="80.00000%"} ![Radial flow (top row) at mid-depth and radial field at CMB (lower row) for the columnar reference case (left) and the hemispherical dynamo (right), indicates the reduction of the magnetic signature at the CMB if the radial motions are limited to the southern polar cusp of high heat flux. Here an Aitoff projection of the spherical CMB is used. See the online-version of the article for the color figure. []{data-label="doublehammer"}](fig2.pdf){width="100.00000%"} Figure \[zonal\] and \[doublehammer\] illustrate the typical hemispherical dynamo configuration emerging at $g=100\%$ and compares this with the typical dipole dominated dynamo found at $g=0\%$. While the southern hemisphere is still cooled efficiently the northern hemisphere remains hot since radial upwellings and the associated convective cooling are predominantly concentrated in the southern hemisphere (figure \[doublehammer\], top row). The flow pattern changes from classical columnar solutions to a thermal wind dominated flow which is a direct consequence of the strong north/south temperature gradient (figure \[zonal\], left bottom). When neglecting inertial, viscous and Lorentz force contributions the azimuthal component of the curl of the Navier-Stokes equation yields: $$2 \frac{\partial \bar{u}_\phi}{\partial z} = \frac {Ra E}{Pr} \frac{1}{r_{cmb}} \frac{\partial \bar{T}}{\partial {\vartheta}}\;\;.$$ This is the thermal wind equation and the respective zonal flows will dominate the solution, indicated by large $\mathcal{A}$ values when the latitudinal temperature gradient is large enough [@Landeau2011]. Since radial flows mainly exist in the southern hemisphere the production of poloidal and thus radial magnetic field is also concentrated there. This results in a very hemispherical magnetic field pattern at the top of the dynamo region (figure \[doublehammer\], bottom row). The figure \[sym\_rel\] demonstrates that the toroidal energy rises quickly with the variation amplitude $g$ while the poloidal energy is much less effected. The growth of the toroidal energy is explained by the increasing thermal wind, which is an equatorial anti-symmetric and axisymmetric (EAA) toroidal flow contribution. At a disturbance amplitude of $g=60\%$ the EAA contribution accounts for already $50\%$ of the total kinetic energy (figure \[sym\_rel\].b ). The maximum EAA contribution of $\mathcal{A}\approx0.8$ is reached at $g=100\%$. When further increasing the variation amplitude, the thermal wind still gains in speed. However, the relative importance of the EAA mode decreases because the strongest latitudinal temperature gradient and thus the thermal wind structure moves further south. This trend is already observed in figure \[zonal\]. The equatorial anti-symmetry of the poloidal kinetic energy rises from $10\%$ for $g=0$ to about $50\%$ for $g=100$% reflecting that upwellings are increasingly concentrated in one (southern) hemisphere. The meridional circulation remains weak (figure \[sym\_rel\].d), and its contribution to the total EAA energy is minor. ![Symmetries and amplitude of total kinetic energy and toroidal/poloidal contributions as a function of $g$. a) total (solid line), toroidal (dashed) and poloidal (dotted) kinetic energy; b) relative amount of axisymmetry (solid), equatorial anti-symmetry (dashed) and the combined symmetries (EAA, dotted) of the full kinetic energy; c) and d) show the same but separated into toroidal and poloidal contributions.[]{data-label="sym_rel"}](fig3.pdf){width="100.00000%"} \[eaadynamics\] Dynamo mechanism ---------------- The upper panel in figure \[tor\_pol\_ome\] demonstrates that the rise in the magnetic Reynolds number $Rm$, that goes along with the increasing toroidal flow amplitude, does not necessarily lead to higher Elsasser numbers. Once more, cases at $E=10^{-4}$, $Ra=4.0 \times 10^{7}$ and $Pm=2$ are depicted here. For small variation amplitudes up to $g=30$% $\Lambda$ still increases due to the additional $\Omega$-effect associated to the growing thermal winds. Figure \[tor\_pol\_ome\] (lower panel) shows that the relative contribution of the $\Omega$-effect to toroidal field production $\mathcal{O}$ grows with $g$. For $g=0$ it is rather weak so that the dynamo can be classified as $\alpha^2$ [@Olson1999]. Around $g=50\%$, $\mathcal{O}$ reaches $50\%$ and the dynamo is thus of an $\alpha^2\Omega$-type. When increasing $g$ further the classical convective columns practically vanish and the associated $\alpha$-effects decrease significantly, leading to both weak poloidal and toroidal fields (figure \[tor\_pol\_ome\], lower panel). For the toroidal field the effect is somewhat compensated by the growing $\Omega$-effect. The hemispherical dynamo clearly is an $\alpha\Omega$-dynamo. At $g=100\%$ the hemispherical mode clearly dominates and the dynamo is of the $\alpha\Omega$-type with $\mathcal{O}\approx0.8$. The Elsasser number has dropped to half its value at $g=0$ while the magnetic Reynolds number has increased by a factor two (figure \[tor\_pol\_ome\], upper panel). The hemispherical dynamo is clearly less effective than the columnar dynamo. ![Upper panel: Flow amplitude in terms of the magnetic Reynolds number (solid line) and magnetic field strength in terms of Elsasser number (dashed)) as function of the CMB heat flux anomaly amplitude $g$, shows the difference between both dynamo regimes in the efficiency of inducing a dynamo. The hemispherical solution, with the $\alpha \Omega$-induction contains large amounts of axisymmetric zonal flows created by the Coriolis force, therefore the kinetic energy is drastically larger than in the columnar regime ($g=0$). The magnetic energy decreases, the more the $g$ increases. The gray shade correspond to the standard deviation due to time variability. Lower panel: Toroidal (dashed) and poloidal (solid) magnetic field in nondimensional units and the relative $\Omega$-effect in terms of ${\mbox{$\mathcal{O}$}}$ (dotted) as a function of $g$ demonstrates the transformation of induction characteristic from an $\alpha^2$-dynamo at $g=0$ (columnar dynamo) towards an $\alpha \Omega$-type from $g=60\%$ (hemispherical solution).[]{data-label="tor_pol_ome"}](fig4.pdf){width="80.00000%"} ![3D visualization of flow and magnetic field generation in a hemispherical dynamo. The meridional cut depicts contours of axisymmetric zonal flows with prograde (retrograde) directions shown in yellow/red (blue). Outward (inward) radial flows are shown as yellow/red (blue) contours of a spherical shell at mid depth $r_{icb}+(r_{cmb}-r_{icb})/2$. Red (blue) isosurfaces depict the 3D structure of convective upwellings (downwellings). Gray fieldlines illustrate the magnetic field configuration. Their thickness is scaled with the local magnetic field energy. See the article online-version for the color figure.[]{data-label="schema"}](fig5.pdf){width="80.00000%"} Figure \[schema\] illustrates the hemispherical dynamo mechanism in a 3D rendering. Magnetic field lines show the magnetic field configuration, their thickness is scaled with the local magnetic energy while red and blue colors intensities indicate the relative inward and outward radial field contribution. Plain gray lines are purely horizontal. Red and blue isosurfaces characterize inward and outward directed radial plume-like motions producing radial field magnetic field. Strong axisymmetric zonal field is produced by a thermal wind related $\Omega$-effect around the equatorial plane. Magnetic Oscillations {#oscillations} --------------------- Figure \[gaussplots\] illustrates the changes in the time behavior of the poloidal magnetic field when the CMB heat flux variation is increased. We concentrate on axisymmetric Gaussian coefficients at the CMB $(r=r_{cmb})$ here. In the reference case $g=0$ (top panel) the axial dipole dominates, varies chaotically in time and never reverses. If $g$ is increased to $50\%$ (second panel) the relative importance of the axial quadrupole component has increased significantly, which indicates the increasing hemisphericity of the magnetic field. To yield a hemispherical magnetic field a similar amplitude in dipolar (equatorial antisymmetric) and quadrupolar (equatorial symmetric) dynamo family contributions is required [@Landeau2011; @Grote2000]. When increasing the variation slightly to $g=60\%$ (third panel) where the hemispherical mode finally dominates, all coefficients assume a comparable amplitude and oscillate in phase around a zero mean with a period of roughly half a magnetic diffusion time. The faster convective flow variations can still be discerned as a smaller amplitude superposition in figure \[gaussplots\]. The oscillation is also present in a kinematic simulation performed for comparison and is thus a purely magnetic phenomenon. Lorentz forces nevertheless cause the flow to vary along with the magnetic field. Since the coefficients vary in phase there are times where the magnetic field and thus the Lorentz forces are particularly weak or particularly strong. Figure \[lorentz\] illustrates the solutions at maximum (top) and minimum (middle) rms field strength. At the minimum the convective columns are still clearly present and the flow is similar to that found in the non-magnetic simulations shown in the lower panel of figure \[lorentz\]. At the maximum the Lorentz forces, in particular those associated with the strong zonal toroidal field, severely suppress the columns. The magnetic field thereby further promotes the dominance of the hemispherical mode [@Landeau2011]. This becomes even more apparent when comparing the relative importance of the EAA mode $\mathcal{A}$ in magnetic and non-magnetic simulations in the top panel of figure \[eaa\_angle\]. In the dynamo run $\mathcal{A}$ is around $35\%$ higher than in the non-magnetic case for mild heat flux variation amplitudes. When further increasing the amplitude of the CMB heat flux pattern, the frequency grows, the time behavior becomes somewhat more complex, and the different harmonics vary increasingly out of phase. In addition, the relative importance of harmonics higher than the dipole increase which indicates a concentration of the field at higher southern latitudes. The impact of the oscillations on the flows decreases since the hemispherical mode now always clearly dominates and the relative variation in the magnetic field amplitude becomes smaller. The appearance of the oscillations may result from the increased importance of the $\Omega$-effect which at $g=60$% starts to dominate toroidal field production (see figure \[tor\_pol\_ome\], lower panel). The $\Omega$-effect could be responsible for the oscillatory behavior of the solar dynamo as has, for example, been demonstrated by @Parker1955 who describes a simple purely magnetic wave phenomenon. @Busse2006 report Parker wave type oscillatory behavior in their numerical dynamo simulations where the stress free mechanical boundary conditions promote strong zonal flows and thus a significant $\Omega$-effect. ![Time evolution of the first five axisymmetric Gauss coefficients at the CMB for the dipole dominated ($g=0$%,$50$%) (first and second panel), the oscillatory ($g=60$%, third panel) and the reversing hemispherical regime ($g=100$% bottom). The colours indicate different spherical harmonic degrees $l$: black $l=1$, dark gray $l=2$, gray $l=3$, light gray $l=4$, faint gray $l=5$. []{data-label="gaussplots"}](fig6_bw.pdf){width="90.00000%"} ![The upper two rows depict the solution at $g=60$% at maximum ($\Lambda=12$) and minimum magnetic field amplitude ($\Lambda=0.06$) respectively. The lower row shows a non-magnetic simulation at identical parameters for comparison. Each row shows from left to right: the radial flow at mid depth in the shell, the z-vorticity in the equatorial plane, the azimuthal magnetic field at the equator and the zonal toroidal field. See the online-version of the article for the color figure.[]{data-label="lorentz"}](fig7.pdf){width="100.00000%"} Arbitrary tilt angle -------------------- To explore to which degree the effects described above still hold when the variation and rotation axis do not coincide we systematically vary the variation pattern tilt angle $\alpha$ (see eq. \[defalpha\]) up to 90 degrees. The lower panel in figure \[eaa\_angle\] shows how $\mathcal{A}$, the relative EAA kinetic energy, varies with $g$ for different tilt angles. Somewhat surprisingly, the hemispherical mode still clearly dominates for tilt angles up to $\alpha=80^\circ$. Only the rather special case of $\alpha=90^\circ$ shows a new behavior, where $\mathcal{A}$ remains negligible. It is thus the general breaking of the north/south symmetry that is essential here. Since it leaves the northern hemisphere hotter than the southern it always leads to the above described dynamo mode. ![upper panel : Effect of the Lorentz force on the relative kinetic energy in the EAA mode for dynamo (black) and non-magnetic (gray) simulations. The time variability is indicated by gray shaded areas in the width of the standard deviation. The magnetic oscillation described in the text lead to the stronger time variability in the dynamo simulations at $g=60$% and $g=100$%. lower panel: The relative equatorially anti-symmetric and axisymmetric energy for different tilting angles follows the onset of EAA convective mode in the case for the axial pattern (black line). For the equatorial orientation (squares) the EAA contribution to total kinetic energy remains Zero. Triangles - $10^\circ$, crosses - $30^\circ$, faint circles- $45^\circ$, dark circles - $60^\circ$, plus symbols - $80^\circ$. See the online-version of the article for the color figure. []{data-label="eaa_angle"}](fig8.pdf){width="80.00000%"} The $90$ degree tilt angle of the $(l=1,m=1)$ pattern forms a special case because the breaking of the north-south symmetry is missing here. Finally, the effects of the east/west symmetry breaking become apparent and supersede the thermal wind related action in the other cases. Figures \[eq\_vphi\] and \[t90\_flow\] illustrate the solution for an equatorial anomaly with $g=200\%$. The resulting east/west temperature difference drives a large scale westward directed flow and a more confined eastward flow in the equatorial region of the outer part of the shell (figure \[eq\_vphi\]). Coriolis forces divert the westward directed flow poleward and inward, and lead to the confinement of the eastward directed flow. Consequently, the westward flow plays the more important role here. The diverted flows feed two distinct downwelling features that form at the latitude of zero heat flux disturbance close to the tangent cylinder. Due to the significant time dependence of the solution these can best be identified in time average flows shown in figure \[t90\_flow\]. Convective columns concentrated in the high heat flux hemisphere but the center of their action is somewhat shifted retrograde, probably due to the action of azimuthal winds. Other authors have shown that this shift, for example, depends on the Ekman number [@Christensen2003]. The remaining columns are small scale and highly time dependent. On time average only one column-like feature remains, identified by a strong downwelling somewhat west to the longitude of highest heat flux. The time averaged flows form two main vorticity structures illustrated in figure \[t90\_flow\]. A long anticyclonic structure associated to the strong equatorial westward flow stretches nearly around the globe and connects the equator with high latitudes inside the tangent cylinder. A smaller cyclonic feature is owed to the eastward equatorial flow. The snapshot and time averaged radial magnetic fields shown in figure \[t90\_flow\] are rather similar which demonstrates that the time dependent small scale convective features are not very efficient in creating larger scale coherent magnetic field. The radial field is strongly concentrated in patches above flow downwelling where the associate inflows concentrate the background field [@Olson1999]. Like in the study for dynamos with homogeneous CMB heat flux by @Aubert2008 the anti-cyclone mainly produces poloidal magnetic field. The cyclone twists the field in the other direction and therefore is responsible for the pair of inverse (outward directed here) field patches located at mid latitudes in the western hemisphere. The exceptional strength of the high latitude normal flux patches suggests that additional field line stretching further intensified the field here. ![Equatorial slice of $u_\phi$ for the homogeneous reference case (left) and the equatorial heat flux anomaly (right). The plus, minus and zero character describe the maximal, minimal and the zero line of the anomaly. See the online-version of the article for the color figure.[]{data-label="eq_vphi"}](fig9.pdf){width="70.00000%"} ![Aitoff projections of spherical surfaces of (from top to bottom) radial field, temperature (both at the CMB), $u_r$, $u_{\vartheta}$ and z-vorticity at $r/r_{cmb} = 0.8$ for a snapshot (left plots) and the time average (right). Parameters: $Ra=4 \times 10^7$, $E=10^{-4}$, $Pm=2$, equatorial ($l=m=1$) perturbation with $g=100\%$ relative amplitude. See the online-version of the article for the color figure.[]{data-label="t90_flow"}](fig10.pdf){width="70.00000%"} Parameter Dependence {#Parameter} -------------------- Focusing again at the axial heat flux anomaly we further study the influence of Rayleigh $Ra$, Ekman $E$ and magnetic Prandtl number $Pm$. In general we find that, independently of the Ekman $E$ and Rayleigh numbers $Ra$, a hemispherical dynamo mode is promoted once $g$ reaches a value of $60\%$. Close to the onset of dynamo action a mild variation can help to maintain dynamo action due to the additional $\Omega$-effect. See the cases at $E=10^{-4}$ and either $Ra=7\times10^6$, $Pm=2$ or $Ra=2\times10^8$, $Pm=1$ in table \[Tab1\]. A strong amplitude of the heat flux anomaly can also suppress dynamo action due to the weakening of convective columns by the Lorentz force. For example, at $E=10^{-4}$, $Ra=4\times10^7$ and $Pm=2$ the dynamo fails once $g$ reaches $200\%$. Figure \[Hboth\] shows how the CMB and surface hemisphericity ( , ) depends on the magnetic Reynolds number based on the equatorially anti-symmetric part of the zonal flow only and therefore useful to quantify the important $\Omega$-effect in the hemispherical dynamo cases. For $E=10^{-4}$ the values first increase linearly with and then saturates around ${\mbox{$\mathcal{H}_{cmb}$}}\approx0.75$ for ${\mbox{${Rm^\star}$}}\ge400$. All cases roughly follow the same curve with the exception of the peculiar $Ra=7\times10^6$, $Pm=2$ and $g=60$% case described above. This means that there is a trade off between $g$, $Ra$ and $Pm$; increasing either parameter leads to larger values. All the solution with hemisphericities ${\mbox{$\mathcal{H}_{cmb}$}}\ > 0.6$ oscillate. The few simulations at smaller Ekman numbers indicate that the degree of hemisphericity decreases with decreasing $E$. This is to be expected since the Taylor Proudman theorem becomes increasingly important [@Landeau2011], inhibiting the ageostrophic hemispherical mode. Larger heat flux variation amplitudes can help to counteract this effect. Since both inertia and Lorentz forces can help to balance the Coriolis force, increasing either $Ra$ or $Pm$ also helps. For $E=3{\mbox{$\times10^{-5}$}}$ an oscillatory case with ${\mbox{$\mathcal{H}_{cmb}$}}=0.76$ is found for the larger Rayleigh number of $Ra=4{\mbox{$\times10^{8}$}}$ and $g=100$%. For $E=10^{-5}$ ${\mbox{$\mathcal{H}_{cmb}$}}$ remains small at $g=100$% and we could not afford to increase $Ra$ here since larger $Ra$ as well as lower $E$ values both promote smaller convective and magnetic length scales and therefore require finer numerical grids. The decrease in length scales has another interesting effect on the radial dependence of hemisphericity. To yield a maximum hemisphericity, equatorially symmetric ($l+m=$ even) and anti-symmetric ($l+m=$ odd) magnetic field contributions must be of comparable strength, i.e. obey a ’whitish’ spectrum (in a suitable normalization) [@Grote2000]. Since, however, the radial dependence of the modes depends on the spherical harmonic degree (they decay like $r^{-(l+2)}$ away from the CMB) the hemisphericity also depends on radius. The spectrum can only be perfectly ’white’ at one radius. The smaller the scale of the magnetic field at $r_{cmb}$ the further this radius lies beyond $r_{cmb}$. This explains why the values shown in the lower panel of figure \[Hboth\] show a much larger scatter than the values. Larger values of $Ra$, $E$, but also $g$ and $Pm$ lead to small magnetic scales and thus larger ratios of ${\mbox{$\mathcal{H}_{sur}$}}$ over ${\mbox{$\mathcal{H}_{cmb}$}}$. ![Top panel: Hemisphericity at CMB versus , the magnetic Reynolds number based on the equatorially anti-symmetric thermal wind. Oscillatory dynamos in gray, stationary in black symbols.bottom panel: Hemisphericity at the (imaginary) Martian surface versus .[]{data-label="Hboth"}](fig11a.pdf "fig:"){width="80.00000%"} ![Top panel: Hemisphericity at CMB versus , the magnetic Reynolds number based on the equatorially anti-symmetric thermal wind. Oscillatory dynamos in gray, stationary in black symbols.bottom panel: Hemisphericity at the (imaginary) Martian surface versus .[]{data-label="Hboth"}](fig11b.pdf "fig:"){width="80.00000%"} Application to Mars {#applymars} =================== Could the hemispherical dynamo models presented above provide an explanation for the crustal magnetization found on Mars? To address this question we rely on the hemisphericity of the crustal magnetization and the magnetic field strength inferred from Martian meteorites. @Amit2011 use MGS data to estimate a hemisphericity between ${\mbox{$\mathcal{H}_{sur}$}}=0.45$ and ${\mbox{$\mathcal{H}_{sur}$}}=0.65$. The magnetization of the Martian meteorite (ALH 84001) suggest a field strength of the ancient dynamo between $5$ and $50\, \textrm{$\mu$T} $ [@Weiss2002]. Because our simulations show that the magnetic field strength also varies significantly with the amplitude of the heat flux pattern, we rescale the dimensionless field strength in our simulations by assuming that the Elsasser number provides a realistic value. Assuming a magnetic diffusivity of $\lambda=1.32\,$m$^2\,$s$^{-1}$ and density of $\rho=7000\,$kg$\,$m$^{-3}$, a rotation rate of $\Omega=7.1\times 10^{-5}\,$s$^{-1}$ and the magnetic vacuum permeability then allows to rescale the Elsasser number to dimensional field strengths. Time is rescaled via the magnetic diffusion time $t_\lambda=D^2 \lambda^{-1}$ with an outer core radius of $1680\,$km. We have included the Martian crustal hemisphericity values in the lower panel of figure \[Hboth\] to show that only oscillatory cases fall in the required range with heat flux variation amplitudes $g\ge 60 $% and ${\mbox{${Rm^\star}$}}\ge 300$. Figure \[H\_rad\_evo\] shows the temporal evolution of and for one of these cases. The variation is surprisingly strong and oscillates at twice the frequency of the individual Gauss coefficients. Since all coefficients roughly oscillate with the same period there are two instances during each period where the hemisphericity is particularly large (around ${\mbox{$\mathcal{H}$}}\approx0.8$) since axial dipole and quadrupole have the same amplitude. Since the mean hemisphericity decreases with radius the variation amplitude is much higher at the planetary surface than at the CMB (figure \[H\_rad\_evo\]). The strong time dependence of oscillatory cases highlights that considerations over which period the magnetization was acquired are extremely important. ![Time evolution of hemisphericity $\mathcal{H}$ at the CMB (solid) and surface (dashed) for $g=100\%$, $Ra=4 \times 10^7$, $E=10^{-4}$ and $Pm=2$.[]{data-label="H_rad_evo"}](fig12.pdf){width="70.00000%"} To translate the dynamo field into a magnetization pattern, @Amit2011 suggest two end-members of how the magnetization was acquired. In the first end-member scenario called ’random’ the crustal magnetization is acquired randomly in time and space and, according to @Amit2011, should reflect the time averaged intensity. In the second end-member called ’continuous’, magnetization is acquired in global thick layers, so that the time intensity of the time average field is considered. However, since the magnetization records the magnetic changes happening during the slow crust formation, the local net magnetization, as seen by an observer, is always proportional to the time averaged local magnetic field possibly slightly dominated by the outermost layers. We therefore think, that the random magnetization scenario does not apply. The strong magnetization found on Mars indicates that a significant portion of the crust is unidirectionally magnetized. @Langlais2004 estimated a magnetization depth of $20 - 40\,$km depending on the magnetization density. Crust formation is a rather slow process that may take millions of years. Typical magnetic time scales can be much shorter. The periods of the reversing strongly hemispherical dynamos discussed above, amount to not more than about ten thousand years. Table \[Tab1\] lists the time averaged rescaled magnetic field intensity at the model Martian surface. For $g\le 60$% the field strengths are similar to that predicted for Mars [@Weiss2002] and fall somewhat below this values for larger $g$-values. In the strongly hemispherical oscillating cases, however, the amplitude of the time average field average to zero on time scales of the crustal magnetization. We therefore conclude that while the hemispherical dynamos can reach hemisphericities similar to that of the Martian crustal magnetization their oscillatory nature makes them incompatible with the rather strong magnetization amplitude. Discussion {#Discussion} ========== We find that an equatorially anti-symmetric convective mode is consistently triggered by a cosine heat flux variation that allows more heat to escape through the southern than through the northern outer boundary of the dynamo region. When the variation is strong enough, convective up- and down-wellings are concentrated at the southern hemisphere and the northern hemisphere remains hot. The associated latitudinal temperature gradients drive strong thermal winds that dominate the flow when, for example, the variation amplitude $g$ exceeds $50\,$% at $E=10^{4}$. Tilting the heat flux pattern axis leaves the solution more or less unchanged with the exception of the $90^\circ$-case where the equatorial symmetry remain unbroken. We conclude that breaking the equatorial symmetry is dynamically preferred over an equatorially oriented heat flux anomaly of the CMB heat flux. Due to the thermal winds, the dynamo type changes from $\alpha^2$ to $\alpha\Omega$ but is generally less efficient. Lorentz forces associated with the toroidal field created via the $\Omega$-effect tend to kill whatever remains of classical columnar convection. This further increases the equatorial anti-symmetry of the solution. Poloidal fields are mainly produced by the southern up- and downwellings which lead to a hemispherical field pattern at the outer boundary. When the hemisphericity approaches values of that found in Martian crustal magnetization, however, all dynamos start to oscillate on (extrapolated) time scales of the order of $10\,$kyr. These oscillations are reminiscent of previously described Parker waves in dynamo simulations [@Busse2006]. As a typical characteristic of Parker waves, the frequency increases with the (square root of the) shear strength, see table \[Tab1\]. The oscillation periods are much shorter than the time over which the deep reaching Martian magnetization must have been acquired [@Langlais2004]. Being a composite of many consecutive layers with alternating polarities the net magnetization would scale with the time averaged field and would therefore likely be much smaller than the predicted strength of the ancient Martian field magnetizing the crust [@Weiss2002]. The maximum hemisphericity for non-oscillatory dynamos amounts to a configuration where the mean northern field amplitude is only $50\%$ weaker than the southern. Additional effects like lava-overflows would then be required to explain the observed hemisphericity. @Amit2011 [@Stanley2008] also studied the effects of the identical sinusoidal boundary heat flux pattern and find very similar hemispherical solutions. @Amit2011 used a very similar setup to ours and also reported oscillations when the dynamo becomes strongly hemispherical. @Stanley2008 do not report the problematic oscillations intensively studied here, which may have to do with differences in the dynamo models. They study stress-free rather than rigid flow boundaries and assume that the growing inner core contributes to drive the dynamo while our model exclusively relies on internal heating. Should a hemispherical dynamo indeed be required to explain the observed magnetization dichotomy, this may indicate that ancient Mars already had an inner core. Alternatively efficient demagnetization mechanisms may have modified an originally more or less homogeneous magnetized crust [@Shahnas2007]. @Landeau2011 observed that similar hemispherical dynamos are found when the Rayleigh number exceeds a critical value. However, albeit the effects are significantly smaller than when triggered via the boundary heat flux. All the cases explored here remain below this critical Rayleigh number. @Landeau2011 also mentioned that the equatorial anti-symmetry, and thus the hemisphericity of the magnetic field, decreases when the Ekman number is decreased. Our simulations at lower Ekman number seem to confirm this trend although a meaningful extrapolation to the Martian value of $E=3\times 10^{-15}$ would require further simulations at lower Ekman numbers. To a certain extent the decrease can be compensated by increasing the heat flux variation amplitude, the Rayleigh number or the magnetic Prandtl number. Our results show that a north-south symmetry breaking induced by lateral CMB heat flux variations can yield surprisingly strong effects. Fierce thermal winds and local southern upwellings take over from classical columnar convection and the dynamo changes from an $\alpha^2$ to an $\alpha \Omega$-type. The dominant $\Omega$-effect seems always linked to Parker-wave-like field oscillations typically discussed for stellar applications. It will be interesting to further explore the aspects independent of the application to Mars. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank Ulrich Christensen for helpful discussions. W. Dietrich acknowledges a PhD fellowship from the Helmholtz Research Alliance ’Planetary Evolution and Life’ and support from the ’International Max Planck Research School on Physical Processes in the Solar System and Beyond’.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A simple version for the extension of the Taylor theorem to the operator functions was found. The expansion was done with respect to a value given by a diagonal matrix for the non-commutative case, and the coefficients are given both by recurrence relations and Cauchy integrals.' author: - \| Ioan Sturzu\ Center for Computational Nanoscience,\ Ball State University, Muncie, IN 47306 title: Taylor expansion for an operator function --- -0.01cm In Quantum Physics, one has to use functions of the operators defined in some Hilbert spaces to describe specific quantities, as evolution operator, state operator. Usually, these are known for some more simple operators, while changing some parameters, the problems become analytically intractable, usually because the involved operators cannot be easy diagonalizable anymore. However, approximate solutions exist when the new operators differ (in some appropriate norm) from the analytically-tractable ones. Here we present two independent mathematical proofs of a useful result, which straightforwardly generalizes the classical result of Taylor expansion from the real analysis to the operator algebra domain. A function of an operator defined on a Hilbert space can be written using a Cauchy-type formula \[1\]: $$\label{funcop} f(A)=\frac 1{2 \pi i}\oint\limits_\gamma \frac{f(z)}{z-A}$$ where $\gamma$ is the closed rectifiable Jordan boundary of an open domain containing the spectrum of the operator $A$. The following lemma applies: \[lemma\]If $\lambda $ is a diagonal matrix, $\left[ \lambda \right] _{ip}=\lambda _i\delta _{ip}$, than: $$\left( {\lambda +\tau }\right) ^p={\lambda }^p(1+\sum_{q=0}^{p-1} \epsilon _q+...\sum_{q=0}^{p-1}\sum_{q_1=q+1}^{p-1}...\sum_{q_{k-1}=q_{k-2}+1}^{p-1}\epsilon _{q_{k-1}}...\epsilon _{q_1}\epsilon _q+...)$$ where: $$\left[ \epsilon _q\right] _{ip}=\left( \frac{{\lambda _p}}{{\lambda _i}} \right) ^q\frac 1{{\lambda _i}}\tau _{ip}$$ Proof : One has: $$\left[ {\lambda }\epsilon _{q_1}...\epsilon _{q_r}{\lambda }\right] _{ip}{ =\lambda }_i{\sum\limits_{m_1,m_2,...m_{r-1}}}\left[ \epsilon _{q_1}\right] _{im_1}\left[ \epsilon _{q_2}\right] _{m_1m_2}...\left[ \epsilon _{q_r}\right] _{m_{r-1}p}{\lambda }_p=$$ $$={\lambda }_i{\sum\limits_{m_1,m_2,...m_{r-1}}}\left[ \epsilon _0\right] _{im_1}\left[ \epsilon _0\right] _{m_1m_2}...\left[ \epsilon _0\right] _{m_{r-1}p}\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^{q_1}\left( \frac{{\lambda _{m_2}}}{{\lambda _{m_1}}}\right) ^{q_2}...\left( \frac{{ \lambda _p}}{{\lambda _{m_r-1}}}\right) ^{q_r}{\lambda }_p=$$ $$={\lambda }_i^2{\sum\limits_{m_1,m_2,...m_{r-1}}}\left[ \epsilon _0\right] _{im_1}\left[ \epsilon _0\right] _{m_1m_2}...\left[ \epsilon _0\right] _{m_{r-1}p}\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^{q_1+1}\left( \frac{{\lambda _{m_2}}}{{\lambda _{m_1}}}\right) ^{q_2+1}...\left( \frac{{\lambda _p}}{{\lambda _{m_r-1}}}\right) ^{q_r+1}{=}$$ $$={\lambda }_i^2{\sum\limits_{m_1,m_2,...m_{r-1}}}\left[ \epsilon _{q_1+1}\right] _{im_1}\left[ \epsilon _{q_2+1}\right] _{m_1m_2}...\left[ \epsilon _{q_r+1}\right] _{m_{r-1}p}=\left[ {\lambda }^2\epsilon _{q_1+1}...\epsilon _{q_r+1}\right] _{ip}$$ The proof is done by mathematical induction: $$\left( {\lambda +\tau }\right) ^1={\lambda }(1+\epsilon _0)$$ $$\left( {\lambda +\tau }\right) ^{p+1}={\lambda }^p(1+\sum_{q=0}^{p-1}\epsilon _q+...\sum_{q=0}^{p-1}\sum_{q_1=q+1}^{p-1}...\sum_{q_{k-1}=q_{k-2}+1}^{p-1}\epsilon _{q_{k-1}}...\epsilon _{q_1}\epsilon _q+...){\lambda }(1+\epsilon _0)=$$ $$={\lambda }^{p+1}(1+\sum_{q=1}^p\epsilon _q+...\sum_{q=1}^p\sum_{q_1=q+1}^p...\sum _{q_{k-1}=q_{k-2}+1}^p\epsilon _{q_{k-1}}...\epsilon _{q_1}\epsilon _q+...)(1+\epsilon _0)=$$ $$={\lambda }^{p+1}(1+\sum_{q=0}^p\epsilon _q+...\sum_{q=0}^p\sum_{q_1=q+1}^p...\sum _{q_{k-1}=q_{k-2}+1}^p\epsilon _{q_{k-1}}...\epsilon _{q_1}\epsilon _q+...)$$ Let $f$ be a real function, consistent with the conditions of the usual Taylor expansion theorem. For the matrices ${\lambda }$[ and ]{}${\tau }$ as in Lemma \[lemma\], one can construct the following Taylor expansion with respect to the value given by the diagonal matrix $\lambda$: $$\left[ {f(\lambda +\tau )}\right] _{ip}=f(\lambda _i)\delta _{ip}+\sum\limits_{n\ge 1}{\sum\limits_{m_1,m_2,...m_{n-1}}{ A_{i,m_1,m_2,...m_{n-1},p}^{(n,\lambda )}}}\tau _{im_1}\tau _{m_1m_2}...\tau _{m_{n-1}p}$$ where: $$A_{ip}^{(1,\lambda )}=\left\{ \begin{array}{l} \frac{{{f(\lambda }_i{)}-{f(\lambda }_p{)}}}{{\lambda _i-\lambda _p}},\text{ if }{\lambda _i\neq \lambda _p} \\ \\ {f}^{\prime }{{(\lambda }_i{),}}\text{ else} \end{array} \right. \label{recc0}$$ $$A_{i,m,p}^{(2,\lambda )}=\left\{ \begin{array}{l} \frac{A_{ip}^{(1,\lambda )}{-}A_{mp}^{(1,\lambda )}}{{\lambda _i-\lambda _m}} ,\text{ if }{\lambda _i\neq \lambda _m} \\ \\ \text{else limit, i.e. } \left\{\begin{array}{l} \frac{{{f(\lambda }_i{)}-{f(\lambda }_p{)}}}{\left( {\lambda _i-\lambda _p} \right) ^2}-\frac{{f}^{\prime }{{(\lambda }_p{)}}}{{\lambda _i-\lambda _p}}, \text{ if }{\lambda _i\neq \lambda _p} \\ \\ \frac 12{f}^{\prime \prime }{{(\lambda }_i{),}}\text{ else} \end{array} \right. \end{array} \right. \label{recc1}$$ $$A_{i,m_1,m_2,...m_{n-1},m_n,p}^{(n+1,\lambda )}= \left\{\begin{array}{l} \frac{{A_{i,m_2,...m_{n-1},p}^{(n,\lambda )}-A_{m_1,m_2,...m_{n-1},p}^{(n,\lambda )}}}{{\lambda _i-\lambda _{m_1}}}, \text{ if }{\lambda _i\neq \lambda _p} \\ \\ \text{else limit} \end{array} \right. \label{recc}$$ Proof : Using (\[funcop\]) one can can write a Taylor expansion with respect to the null operator \[1\]: $$\begin{aligned} {f(\lambda +\tau )} &=&{}\sum_{n\geq 0}\frac 1{n!}{f}^{\prime (n)}{(0)} \left( {\lambda +\tau }\right) ^n=\sum_{n\geq 0}\frac 1{n!}{f}^{\prime (n)}{ (0)\lambda }^n \nonumber \times \\ && \times (1+...\sum_{q=0}^{p-1}\sum_{q_1=q+1}^{p-1}...\sum_{q_{k-1}=q_{k-2}+1}^{p-1}\epsilon _{q_{k-1}}...\epsilon _{q_1}\epsilon _q+...)\end{aligned}$$ the $(i,p)$ matrix element of the $k$ term from the parenthesis can be written as: $${\sum\limits_{m_1,m_2,...m_{k-1}}B_{i,m_2,...m_{k-1},p}^{(k,\lambda ,n)}} \tau _{im_1}\tau _{m_1m_2}...\tau _{m_{n-1}p}$$ where: $$\begin{aligned} {B_{i,m_1,m_2,...m_{k-1},p}^{(k,\lambda ,n)}} &=&{}\sum_{q=0}^{n-1} \sum_{q_1=q+1}^{n-1}...\sum_{q_{k-1}=q_{k-2}+1}^{n-1}\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^{q_{k-1}}\left( \frac{{\lambda _{m_2}}}{{\lambda _{m_1}}}\right) ^{q_{k-2}}... \\ &&...\left( \frac{{\lambda _{m_k-1}}}{{\lambda _{m_k-2}}}\right) ^{q_1}\left( \frac{{\lambda _p}}{{\lambda _{m_k-1}}}\right) ^q\frac 1{{ \lambda _i\lambda _{m_1}...\lambda _{m_k-1}}}\end{aligned}$$ In the corresponding expression for ${B_{i,m_1,m_2,...m_k,p}^{(k+1, \lambda ,n)}}$, if ${\lambda _i\neq \lambda _{m_1}}$ one has: $$\sum_{q_k=q_{k-1}+1}^{n-1}\left( \frac{{\lambda _{m_1}}}{{\lambda _i} }\right) ^{q_{k-1}}=\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^{q_{k-1}+1}\frac{1-\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^{n-1-q_{k-1}}}{1-\frac{{\lambda _{m_1}}}{{\lambda _i}}} \label{sum}$$ so, one obtains after some algebra: $${B_{i,m_1,m_2,...m_k,p}^{(k+1,\lambda ,n)}=}\frac{{B_{i,m_2,...m_k,p}^{(k, \lambda ,n)}-\left( \frac{{\lambda _{m_1}}}{{\lambda _i}}\right) ^nB_{m_1,m_2,...m_k,p}^{(k,\lambda ,n)}}}{{\lambda _i}-{\lambda _{m_1}}} \label{recc1}$$ Identifying the coefficients: $$A_{i,m_1,m_2,...m_{k-1},p}^{(k,\lambda )}=\sum_{n\geq 0}\frac 1{n!}{f} ^{\prime (n)}{(0)\lambda }_i^n{B_{i,m_1,m_2,...m_{k-1},p}^{(k,\lambda ,n)}}$$ and using (\[recc1\]), one obtains the recurrence relation (\[recc\]). If ${\lambda _i=\lambda _{m_1}}$, (\[sum\]) is equal to $(n-1-q_{k-1})$; finally the term $n$ will go to a derivative (which is finite, because $f$ is Taylor), while the contribution of the remainder is also finite. So, for the higher orders, the limiting argument is valid, if for the lower orders this is true. However, for orders $1$ and $2$ this was verified by direct calculation, so the theorem is true, by the argument of mathematical induction. The same result can be obtained starting with a Dyson expansion of the integrand from (\[funcop\]). $$\frac 1{z-\left( {\lambda +\tau }\right) }=\frac 1{z-{\lambda }% }+\sum\limits_{n\ge 1}(\frac 1{z-{\lambda }}{\tau )}^n\frac 1{z-{\lambda }}$$ Using (\[funcop\]) one has: $$\begin{aligned} \left[ {f(\lambda +\tau )}\right] _{ip} &=&f(\lambda _i)\delta _{ip}+\sum\limits_{n\ge 1}{\sum\limits_{m_1,m_2,...m_{n-1}}}% \oint\limits_\gamma \frac{{f(z)dz}}{\left( z-{\lambda }_i\right) \left( z-{% \lambda }_{m_1}\right) ...\left( z-{\lambda }_{m_{n-1}}\right) }\times\\ &&\times \tau _{im_1}\tau _{m_1m_2}...\tau _{m_{n-1}p}\end{aligned}$$ Identifying: $${{A_{i,m_1,m_2,...m_{n-1},p}^{(n,\lambda )}=}}\oint\limits_\gamma \frac{{% f(z)dz}}{\left( z-{\lambda }_i\right) \left( z-{\lambda }_{m_1}\right) ...\left( z-{\lambda }_{m_{n-1}}\right) }$$ after decomposing the path $\gamma$ in closed paths surrounding only one of the values $\lambda$ one can obtain very easy the recurrence relations (\[recc0\]) and (\[recc1\]). References =0.2cm 1. N. Dunford and J. Schwartz, Linear Operators (Interscience Publ., 1957). This work has been supported by The Indiana 21st Century Research and Technology Fund.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent results from the DELPHI collaboration [@DELPHI_NOTE] led us to review the present bounds on the $b''$ quark mass. We use all available experimental data for $m_{b''} > 96$ GeV to constrain the $b''$ quark mass as a function of the Cabibbo-Kobayashi-Maskawa elements in a sequential four generations model. We find that there is still room for a $b''$ with a mass larger than 96 GeV.' author: - 'S. M. Oliveira' - 'R. Santos' title: 'Bounds on the mass of the quark, revisited' --- Introduction ============ It has long been known that a sequential fourth generation within the Standard Model (SM) needs both quarks and leptons. Half a generation would imply that the gauge anomalies associated with triangle diagrams would not cancel. It is also known [@Hagiwara:fs] that SLC, and then LEP have set a bound on the number of light neutrinos ($ m_{\nu} \, < \, M_Z/2$), which is indisputably equal to three. This bound applies to all new fermions that couple to the Z and one has to be extremely open minded to accept a fourth neutrino with a mass larger than around 45 GeV. Thus, there seems to be no strong motivation for the search of a sequential fourth generation (for a review see [@Frampton:1999xi]). So why look for it? Despite the strength of the previous arguments one should try to experimentally exclude the existence of a fourth generation. In fact such evidence does not yet exist. The most recent precision electroweak results [@Novikov:2001md] allow a sequential fourth generation if the quark masses are not too further apart$^1$. The same results also disfavour a degenerate fourth family if both the leptonic and hadronic sector are degenerate. This is in agreement with the conclusions of Erler and Langacker [@Hagiwara:fs]. However, as discussed in ref. [@Frampton:1999xi], there are several reasons to keep investigating this subject starting with the fact that precision results vary with time. In ref. [@Frampton:1999xi] it can be seen that even if one takes a degenerate fourth family of quarks with 150 GeV masses, it is enough to choose a non-degenerate family of leptons with masses of 100 GeV and 200 GeV and a Higgs mass of 180 GeV for the discrepancy with experimental data to fall from roughly three to two standard deviations$^2$. Moreover, it is clear that any new physics will also influence these results. It was shown in refs. [@Frampton:1999xi; @Arhrib:2000ct] that the mass range $\|m_{t'}-m_{b'}\| \leq 60 \, \text{GeV}$, where $t'$ and $b'$ are the fourth generation quarks, is consistent with all available precision electroweak data. This range enable us to say that even if $m_{b'}>m_{t'}$, the decay $b' \rightarrow t' \, W$ is forbidden. The decay $b' \rightarrow t' \, W^$ although allowed, is phase space suppressed [@Sher:1999ae] and consequently extremely small in the mass range under study (from now on we consider $m_{b'} < m_{t'}$). Experimental data allow us to go only up to $m_{b'}$ close to $180$ GeV. Hence, the $b'$ can not decay to a top quark. Furthermore, while some recent studies [@Yanir:2002cq; @Huang:2000xe] have constrained the Cabibbo-Kobayashi-Maskawa (CKM) elements of the fourth generation, they do not influence our results. Nevertheless we will take into account the $2 \sigma$ bound $\|V_{tb}\|^2 + 0.75 \|V_{t'b}\|^2 \leq 1.14$ [@Yanir:2002cq] coming from $Z \rightarrow b \bar{b}$ to constrain the CKM element $V_{cb'}$ as a function of the $b'$ mass. Present experimental bounds on the $b'$ mass above 96 GeV suffer from the drawback of assuming a 100 % branching ratio for a specific decay channel. As stated before the strongest bound on the $b'$ mass comes from LEP [@Decamp:1989fk] and is $m_{b'}\, > \, 46$ GeV. Here all $b'$ decays were considered. There are presently three bounds on the $b'$ mass for $m_{b'}\, > \, 96$ GeV. The first one [@Affolder:1999bs], $m_{b'}\, > \, 199$ GeV, assumes that $Br(b' \rightarrow b \, Z)=100 \%$. We will drop this condition and use instead their plot of $\sigma (p \, \bar{p} \rightarrow b' \bar{b'} + X) \times Br^2(b' \rightarrow b \, Z)$ as a function of the $b'$ mass. The second one [@Abachi:1995ms] $m_{b'}\, > \, 128$ GeV, is based on the data collected in the top quark search. Because the D0 collaboration looked for $t \rightarrow b \, W$, the analysis can be used to set a limit on $\sigma (p \, \bar{p} \rightarrow b' \bar{b'} + X) \times Br^2(b' \rightarrow c \, W)$. By doing so we assume that the $b$ and $c$ quark masses are negligible and that $\sigma (p \, \bar{p} \rightarrow b' \bar{b'}) \approx \sigma (p \, \bar{p} \rightarrow t \bar{t})$. The obtained limit $m_{b'}\, > \, 128$ GeV assumes $Br(b' \rightarrow c \, W) = 100 \%$. The third bound is from CDF [@Abe:1998ee] and is based on the decay $b' \rightarrow b \, Z$ followed by the search for $Z \rightarrow e^+ \, e^-$ with displaced vertices. Their excluded region is inside a rectangle in the lifetime $(c \, \tau)$, $m_{b'}$ plane with $\, 9 \times 10^{-3} \enskip \text{cm}\, < \, c \, \tau \, < \, 12 \enskip \text{cm}$ and $m_{b}+ M_Z \, < \, m_{b'}\, < \, 148$ GeV sides. Hence, the excluded region depends heavily on the $b'$ lifetime. But, contrary to the top quark which has a lifetime of around $10^{-24} \, s$, the lifetime of a sequential $b'$ quark is expected to be extremely large, especially knowing that we are considering a heavy $b'$. In fact, depending on the CKM values and on the $b'$ and $t'$ masses, the decay length can be as large as $10^{-4}$ cm or even $10^{-3}$ cm in extreme cases. Nevertheless, in this model, it is very hard to go beyond that value. It is worth mentioning that even with this huge lifetime, the $b'$ always decays inside the detector and hadronization occurs before it decays. Thus, the limit obtained in [@Abe:1998ee] which on top of what was said assumes $Br(b' \rightarrow b \, Z)=100 \%$ can not be used in our analysis. Hence, we think it is worthwhile to reexamine the limits on the $b'$ mass. We will use the CDF and the D0 data which, together with the new DELPHI data, is all that there is available for $m_{b'} > 96$ GeV. We will draw exclusion plots in the plane ($R_{CKM}$, $m_{b'}$), where $R_{CKM}=\|\frac{V_{cb'}}{V_{tb'} \, V_{tb}}\|$, from 96 GeV to 180 GeV without assuming a definite value for the branching ratios of specific channels. In some regions it is possible to combine all experimental data allowing a larger exclusion area. Notice that the use of the $R_{CKM}$ variable provides a new way to look at the experimental results. This variable enable us to actually use and combine all the available data. Moreover, the new form in which the results are presented will serve as a guide to future experiments since it is possible to know how far one has to go to exclude the regions that are still allowed. To end this section we note that there is, at present, no bound on a sequential $2/3$ charged quark in the PDG but if we assume a 100 % decay to cW the bound is again 128 GeV [@Abachi:1995ms]. The paper is organised as follows. In section II we define the model and discuss the production and the decays of $b'$ quarks. In section III we combine the theoretical and the experimental results to produce exclusion plots in the parameter space. Section IV summarises our results and conclusions. production and decay ===================== There are several ways of extending the SM to accommodate a fourth family of quarks and/or leptons. A review of the different models in the literature is available in [@Frampton:1999xi]. Obviously, the most natural and straightforward way to introduce a fourth family in the SM is just to add a $(t', \, b')$ family with the same quantum numbers and similar couplings to all other known quarks. The same can be done for the lepton sector$^3$. This is called a sequential fourth generation model and is sometimes referred to as SM4. The resulting CKM matrix has a very similar structure to the SM one. It is a $4 \times 4$ unitary matrix and it is assumed to be approximately symmetric. Besides the four new masses, there are 9 additional parameters compared to the SM: 6 mixing angles instead of 3 and 3 complex phases instead of 1. Because we are not concerned with CP-violation we take all CKM values to be real. In the SM4, the CKM elements that are not determined experimentally have more freedom due to the extra parameters introduced. This model has been the subject of wide study in the literature. Production cross sections for lepton and hadron colliders and $b'$ branching fractions were calculated long ago. At LEP, a pair of heavy quarks is produced through the reaction $e^+ \, e^- \rightarrow q \bar{q}$. For consistency with the experimental analysis, the process $e^+ \, e^- \rightarrow b' \, \bar{b'}$ was calculated using <span style="font-variant:small-caps;">Pythia</span> [@Sjostrand:2000wi], with initial state radiation (ISR), final state radiation (FSR) and QCD corrections turned on. We have cross checked the results using a simple program with the formulas of refs. [@Jersak:sp] and [@Bernreuther:1991hy], which also include QCD corrections and ISR. Since the larger contribution to the cross section comes from ISR we have double checked by making use of the formulas presented in [@Kuraev:hb]. The results agree very well with the <span style="font-variant:small-caps;">Pythia</span> results. It should be noticed that near the threshold bound states would surely be formed. Without a detailed analysis of such bound states it is impossible to evaluate whether their contribution to the cross section would be relevant or not. So, if bound states do exist above the threshold, we are assuming that they give a negligible contribution to the cross section. Far away from the threshold the problem ceases to exist and the results we will show for hadron colliders are not affected by this approximation. The equivalent production reaction at the Tevatron is $p \, \bar{p} \rightarrow b' \, \bar{b'} + X$, with the relevant processes being $gg \, (q \, \bar{q}) \rightarrow b' \, \bar{b'}$. Even though this cross section can not be found in the literature it is generally recognised that all massive quark pair production cross section are very similar due to its hadronic nature. The same is true for the subsequent decays into leptons and for the detector efficiency. Thus we can use the exact order $\alpha_s^3$ corrected cross section for the production of top quarks [@Laenen:1993xr]. This approximation is used both by the CDF and the D0 collaborations in their studies on $b'$ production and decay. In [@Abachi:1995ms] it is also assumed that the final states are exactly the same as the top quark ones. Notice that the error in calculating the hadronic cross section is much larger than the corresponding error in the leptonic one. For $m_q = 100$ GeV the error is about 38 % falling to 12 % for $m_q = 200$ GeV. This will be reflected in the exclusion plots. All $b'$ decays were exhaustively studied by Hou and Stuart in [@Hou:1988yu; @Hou:1988sx; @Hou:1989ty; @Hou:1990wz] and by Haeri, Eilam and Soni [@hes]. Hou and Stuart have shown that the $b'$ is peculiar in the sense that 1-loop flavour change decays (FCNC) can dominate over charged current decays (CC). Depending on the values of the CKM matrix elements and as long as the Higgs channel remains closed, there are mainly two processes in competition: $b' \rightarrow b \, Z$ and $b' \rightarrow c \, W$. As soon as the Higgs channel opens the decay $b' \rightarrow b \, H$ can be as large as $b' \rightarrow b \, Z$. Other decays like $b' \rightarrow b \, g$ and $b' \rightarrow b \, \gamma$ and three body decays give smaller contributions but can sometimes be relevant. The three body decays $b' \rightarrow b \, e^+ \, e^-$, $b' \rightarrow b \, \nu \, \bar{\nu}$ and $b' \rightarrow b \, q \, \bar{q}$, including box diagrams were calculated in [@Hou:1989ty]. At that time, the top mass was still unknown and the $t'$ was taken to be much larger than the top mass. Under these conditions and for the range of the $b'$ mass in study, the sum of all three body decays could be as large as $b' \rightarrow b \, g$. It could be even larger for a “small” $t$ mass and a very large $t'$ mass [@Hou:1989ty]. But it turned out that the top mass is $\approx$ 175 GeV and electroweak precision measurements force $m_{t'}$ to be close to $m_{b'}$ for the range of $b'$ mass under consideration. In our case we estimate all three body decays plus the decay $b' \rightarrow b \, \gamma$ to be smaller than $b' \rightarrow b \, g$. Nevertheless, because we want to make a conservative estimate we will take it to be as large as $b' \rightarrow b \, g$. Using the unitarity of the CKM matrix, its approximate symmetry $V_{t'b'} \, V_{t'b} \approx V_{tb} \, V_{tb'}$, and taking $V_{ub'} \, V_{ub} \approx \, 0$ and $V_{cb} \approx \, 10^{-2}$ we can write all branching fractions as a function of three quantities alone: $R_{CKM}$, $m_{t'}$ and $m_{b'}$. Notice that the two last conditions do not play a significant role in the final result. Using a very large value like for instance $V_{ub'} \, V_{ub} \approx \, 10^{-4}$ gives a contribution much less than 1 % to the $b' \rightarrow b \, Z$ decay width. The same is true when we relax the condition $V_{t'b'} \, V_{t'b} \approx V_{tb} \, V_{tb'}$ near to a GIM cancellation region. Relaxing this condition leads to an increase by several orders of magnitude of the values of the NC decay widths but they are always much smaller than the CC decays in that region. One-loop calculations of the NC $b'$ decays were performed using the FeynArts and FeynCalc [@feyn] packages for generating and computing the complete set of diagrams and the LoopTools/FF [@ff] packages for the numerical analysis. We have carried out several checks in the four generations model following [@Arhrib:2000ct; @Hou:1988yu; @Hou:1988sx; @Hou:1989ty; @Hou:1990wz] and in the SM against [@Mele; @Eilam]. We have found full agreement in both cases. The branching ratios depend on three quantities alone and 96 GeV $\lesssim m_{b'}\lesssim$ 180 GeV. So, we just have to decide on what values of $R_{CKM}$ and $m_{t'}$ to use. Since we know that $m_{t'}$ is limited by precision data we will study two extreme cases $m_{t'} \, = \, m_{b'}+ 50$ GeV and the almost degenerate case $m_{t'} \, = \, m_{b'}+ 1$ GeV. In the exclusion plots $R_{CKM}$ is a free parameter and so no assumptions on its variation range were made. However, there is a hint on its most significant values coming from the fact that the competing NC and CC cross at $ 10^{-3} \lesssim R_{CKM} \lesssim 10^{-2}$. We will come back to this point later. In fig. \[fig\_new1\] we present the branching ratios as a function of the $b'$ mass with $R_{CKM}=0.001$ and $m_{t'} - m_{b'}= 50$ GeV. The closer to $m_{b'}=96$ GeV we are the larger $b'\rightarrow \, b \, g$ gets due to phase space suppression of the competing NC $b'\rightarrow \, b \, Z$. In fact, for an almost degenerate fourth family and small values of $R_{CKM}$, $b'\rightarrow \, b \, g$ can be the dominant NC for $m_{b'}=96$ GeV. As soon as one moves away from this value, $b'\rightarrow \, b \, Z$ becomes the dominant NC. If the Higgs channel is closed , for $m_{b'} \geq 97$ GeV, the competition is always between $b'\rightarrow \, c \, W$ and $b'\rightarrow \, b \, Z$. As $m_{b'}$ rises so does the NC except if the GIM mechanism gets in the way. It can be clearly seen in the figure the GIM mechanism acting for $m_{b'} \approx 125$ GeV, that is, $m_{t'}-m_{t}=0$. Then the NC rises again and the CC falls crossing at 140 GeV. When $R_{CKM}$ grows so does $b'\rightarrow \, c \, W$ and the crossing point is shifted to the left. As the mass difference tends to zero the GIM effect is shifted to $m_{b'} \approx m_{t}$. In fig. 2 we show the branching ratios as a function of $R_{CKM}$ with $m_{b'} = 110$ GeV and $m_{t'} - m_{b'}= 1$ GeV. As we already knew, the NC are favoured by small values of $R_{CKM}$ because $R_{CKM}$ is a direct measure of the charged currents. Again, when $m_{b'}$ grows so does $b'\rightarrow \, b \, Z$ and the crossing point is shifted to the left. The same happens when $m_{t'} - m_{b'}$ decreases as explained above. Results and discussion ====================== We are now in a position to draw exclusion plots on the plane $(R_{CKM}, m_{b'})$ with $m_{t'}$ as a parameter. Using the latest experimental data from the DELPHI collaboration and the data from the CDF and D0 collaborations together with the theoretical values of the cross sections and the branching ratios we have drawn the exclusion plots shown in the figures below. The upper regions are excluded by the limits on $Br_{b'\rightarrow \, c \, W}$ and the lower regions by the limits on $Br_{b'\rightarrow \, b \, Z}$. The results based on the DELPHI data, are shown in figs. 3 and 4. The only difference between the two plots is in the value of $m_{t'}$. It can be seen that as $m_{t'}-m_{b'}$ grows, the allowed region gets smaller. This is because $Br_{b' \rightarrow b \, Z}$ decreases with $m_{t'}$ due to a GIM suppression as long as $m_{t'}$ is smaller than $m_{t}$ and $(m_{t'}-m_{t}) \rightarrow 0$. On the contrary, $Br_{b' \rightarrow c \, W}$ does not depend on the $t'$ mass. Hence, as $m_{t'}$ grows, $Br_{b' \rightarrow c \, W}$ becomes dominant and the upper excluded region increases. The reason why there isn’t a lower bound close to 96 GeV in both figures is because of the competing neutral currents. Close to the $Z \, b$ threshold ($\approx$ 96 GeV), $b' \rightarrow b \, g$ dominates over $b' \rightarrow b \, Z$ and the experimental bound on $Br_{b' \rightarrow b \, Z}$ becomes useless. As one moves away from the $Z \, b$ threshold, $b' \rightarrow b \, Z$ becomes the dominant neutral current. $Br_{b' \rightarrow b \, Z}$ falls less sharply with $m_{t'}$ than the other neutral currents and that explains why there is a lower bound for e.g. at $m_{b'} = 100$ GeV in fig. 4 but not in fig. 3. After 102 GeV almost all values are allowed because the experiments are not sensitive to those mass values. In figs. 5 and 6 we show similar plots but using the CDF and the D0 data. The D0 data is responsible for excluding the upper regions because it deals with CC as the CDF excludes the lower regions due to the bounds on NC. The three curves marked upper, central and lower are related with the theoretical error bars in the $b'$ production cross section. Again and for the same reason the excluded region grows with $m_{t'}-m_{b'}$. This means that like the constraints from precision electroweak data, the experimental data also disfavours a fourth family with a large mass difference between the two quarks. In some cases the allowed regions in the CDF/D0 and DELPHI plots overlap and the excluded region grows. For instance, considering $m_{b'} = 100$ GeV and $m_{t'}-m_{b'} = 50$ GeV we get for DELPHI $4.5 \times 10^{-4} < R_{CKM} \, < \, 8.4 \times 10^{-4}$ and for CDF/D0 (lower) $6.7 \times 10^{-4} < R_{CKM} \, < \, 1.1 \times 10^{-3}$. Hence, the resulting excluded region is $6.7 \times 10^{-4} < R_{CKM} \, < \, 8.4 \times 10^{-4}$. With the bound $\|V_{tb}\|^2 + 0.75 \|V_{t'b}\|^2 \leq 1.14$ [@Yanir:2002cq] and assuming $\|V_{tb}\| \approx 1$ it is possible to limit the value of the matrix element $V_{cb'}$. For the same value of the $b'$ mass, $m_{b'} = 100$ GeV we know that $R_{CKM} \, < \, 8.4 \times 10^{-4}$ and so $$V_{cb'} \, < \, 8.4 \times 10^{-4} \sqrt{0.14/0.75} \, \approx \, 3.6 \times 10^{-4}$$ with $m_{t'}= m_{b'} + 50 = 150 \, \text{GeV}$. The bound gets weaker for smaller $m_{t'}$ [@Yanir:2002cq]. Finally we show an exclusion plot with the Higgs channel opened and a Higgs mass of 115 GeV. As we expected, the inclusion of the Higgs makes the excluded region to shrink. By itself, the inclusion of one more channel always diminishes the branching ratios and consequently less values will be excluded. Like $b' \rightarrow b \, Z$, $b' \rightarrow b \, h$ is larger for small $R_{CKM}$ and large $m_{b'}$. Hence in this region of parameter space it competes with $b' \rightarrow b \, Z$ and $b' \rightarrow c \, W$ making the allowed region larger. For a detailed analysis of the so-called cocktail solution see [@Arhrib:2000ct]. Conclusion ========== In this work we have found the allowed $b'$ mass as a function of the CKM elements of a four generations sequential model. Using all available experimental data for $m_{b'}> 96$ GeV we have shown that there is still plenty of room for a $b'$ with a mass larger than 96 GeV. We have also shown that the allowed region decreases as $m_{t'}$ increases. In fact, as the gap between the fourth generation quark masses increases the allowed region shrinks. Notice that this is in full agreement with the tendency of a small mass gap, if not completely degenerated, favoured by the electroweak precision measurements. All plots show that $R_{CKM}$ is for sure smaller than $\approx 10^{-2}$ and it can be as small as $\approx 10^{-4}$. This is not surprising because this region is exactly where we expected it to be. In fact, the CKM values we know so far suggest that $V_{cb'} \approx 10^{-4}-10^{-3}$. If $V_{tb'} \approx 10^{-1}$ then a value of $R_{CKM}$ between $10^{-2}$ and $10^{-4}$ is absolutely natural. Moreover, the limit we have obtained for $V_{cb'}$ in the last section makes it even more natural. We know that the DELPHI analysis [@DELPHI_NOTE] is being improved. In the near future we hope to reduce very much the allowed region in figs. 3 and 4. As far as we know there are no new results from the CDF and the D0 collaborations improving their bounds. For large $m_{t'}-m_{b'}$, and for some values of $m_{b'}$ the CDF/D0 limits almost shrink the allowed region to zero. Hence, a small improvement in the analysis could disallow a large region of the parameter space. As for the future, searches in hadron colliders will have to wait for the RunII of the Tevatron and for the Large Hadron Collider (LHC). The $b' \bar{b'}$ production cross section increases by roughly two orders of magnitude at the LHC compared to the Tevatron. Thus LHC will be a copious source of $b'$ pairs. With high values for cross section and luminosity, if background is suppressed exclusion plots can be drawn for a very wide range of $b'$ masses. However, we have to worry about two problems in future searches. From the theoretical point of view we have to take into account all the possible hierarchies in mass, for instance one could have $m_{t'} < m_{t} < m_{b'}$ or $m_{t} < m_{t'} < m_{b'}$. A careful study, including also the possibility of finding a Higgs has to be done. From the experimental point of view we have to know how the detectors will perform. Nobody knows yet if there is going to be a Next Linear Collider with energies of $\sqrt{s}= 500$ GeV or $\sqrt{s}= 1$ TeV. NLC would allow us to go up $m_{b'}=250$ GeV or $m_{b'}=500$ GeV which is close to the perturbative limit. Depending on the available luminosity, and because a small background is expected, we believe that the excluded region would be very large, probably allowing the exclusion of some values of $m_{b'}$ regardless of the values of the mixing angles. However, if a Higgs boson is found the excluded region will surely be smaller and will depend on the mass and type of Higgs boson found. For a detailed discussion on future searches see [@Frampton:1999xi]. In summary we believe that there is still experimental and theoretical work to be done to find or definitely to exclude a sequential fourth generation of quarks at the electroweak scale. We thank A. Barroso and M. Pimenta for comments and suggestions on the manuscript. We thank A. Onofre for reading of the manuscript. We thank our DELPHI/LIP collaborators and also M. Greco for discussions. This work is supported by Fundação para a Ciência e Tecnologia under contract POCTI/FNU/49523/2002. S.M.O. is supported by FCT under contract SFRH/BD/6455/2001. [1]{} S. Andringa [et al.]{} \[DELPHI Collaboration\], Search for a fourth generation $b'$-quark at LEP-II at $\sqrt{s}$ = 200-209 GeV. Contributed paper to EPS 2003 (Aachen) and LP 2003 (FNAL). K. Hagiwara [et al.]{} \[Particle Data Group Collaboration\], Phys. Rev. D [66]{}, 010001 (2002). P. H. Frampton, P. Q. Hung and M. Sher, Phys. Rept. [330]{}, 263 (2000). V. A. Novikov, L. B. Okun, A. N. Rozanov and M. I. Vysotsky, Phys. Lett. B [529]{}, 111 (2002). H. J. He, N. Polonsky and S. f. Su, Phys. Rev. D [64]{}, 053004 (2001). A. Arhrib and W. S. Hou, Phys. Rev. D [64]{}, 073016 (2001). M. Sher, Phys. Rev. D [61]{}, 057303 (2000). T. Yanir, JHEP [0206]{}, 044 (2002), and references therein. C. S. Huang, W. J. Huo and Y. L. Wu, Phys. Rev. D [64]{}, 016009 (2001), and referencer therein. D. Decamp [et al.]{} \[ALEPH Collaboration\], Phys. Lett. B [236]{}, 511 (1990). P. Abreu [et al.]{} \[DELPHI Collaboration\], Nucl. Phys. B [367]{}, 511 (1991). O. Adriani [et al.]{} \[L3 Collaboration\], Phys. Rept. [236]{}, 1 (1993). M. Z. Akrawy [et al.]{} \[OPAL Collaboration\], Phys. Lett. B [246]{}, 285 (1990). T. Affolder [et al.]{} \[CDF Collaboration\], Phys. Rev. Lett. [84]{}, 835 (2000). S. Abachi [et al.]{} \[D0 Collaboration\], Phys. Rev. D [52]{}, 4877 (1995). C. D. Froggatt, D. J. Smith and H. B. Nielsen, Z. Phys. C [73]{}, 333 (1997). F. Abe [et al.]{} \[CDF Collaboration\], Phys. Rev. D [58]{}, 051102 (1998). T. Sjostrand, P. Eden, C. Friberg, L. Lonnblad, G. Miu, S. Mrenna and E. Norrbin, Comput. Phys. Commun. [135]{}, 238 (2001). J. Jersak, E. Laermann and P. M. Zerwas, Phys. Rev. D [25]{}, 1218 (1982) \[Erratum-ibid. D [36]{}, 310 (1987)\]. W. Bernreuther [et al.]{}, TTP-92-19 [Prepared for Workshops on Future e+ e- Colliders, Hamburg, Germany, Sep 2-3, 1991 and Saariselka, Finland, Sep 9-14, 1991]{}, and references therein. E. A. Kuraev and V. S. Fadin, Sov. J. Nucl. Phys. [41]{}, 466 (1985) \[Yad. Fiz. [41]{}, 733 (1985)\]. E. Laenen, J. Smith and W. L. van Neerven, Phys. Lett. B [321]{}, 254 (1994). W. S. Hou and R. G. Stuart, Phys. Rev. Lett. [62]{}, 617 (1989). W. S. Hou and R. G. Stuart, Nucl. Phys. B [320]{}, 277 (1989). W. S. Hou and R. G. Stuart, Nucl. Phys. B [349]{}, 91 (1991). W. S. Hou and R. G. Stuart, Phys. Rev. D [43]{}, 3669 (1991). B. Haeri, G. Eilam and A. Soni, Phys. Rev. Lett. [62]{}, 719 (1989); Phys. Rev. D [41]{}, 875 (1990). T. Hahn, Comp. Phys. Comm. [140]{}, 418 (2001); J. K" ublbeck, M. B" ohm and A. Denner, [ibid.]{} [60]{}, 165 (1990); R. Mertig, M. B" ohm and A. Denner, [ibid.]{} [64]{}, 345 (1991). G. J. van Oldenborgh and J. A. Vermaseren, Z. Phys. C [46]{}, 425 (1990); G.J. van Oldenborgh, Comp. Phys. Comm. [66]{}, 1 (1991); T. Hahn and M. Perez-Victoria, [ibid.]{} [118]{}, 153 (1999). B. Mele, S. Petrarca and A. Soddu, Phys. Lett. [B435]{}, 401 (1998). G. Eilam, J.L. Hewett and A. Soni, Phys. Rev. D [44]{}, 1473 (1991); Erratum-[ibid.]{} D [59*]{}, 039901 (1999).	{ "pile_set_name": "ArXiv" }
--- author: - 'Nicola Bartolo[^1]' - 'David J. Papoular' - Alessio Recati - Chiara Menotti title: 'A toy model for the dipolar-induced resonance in quasi-one-dimensional systems' --- Introduction {#SecIntro} ============ Thanks to a high degree of experimental control, cold-atom ensembles imposed as ideal candidates to quantum-simulate condensed-matter systems [@BlochNatPhys12]. In this perspective a fundamental role is played by the interparticle potential, usually described in terms of a contact interaction, which can be tuned at will exploiting Feshbach resonances [@ChinRMP10]. Two-body scattering is strongly modified by reducing the dimensionality of the system, leading to confinement-induced resonances (CIR) [@OlshaniiPRL98]. Furthermore, the dipolar interaction, due to its long-range and anisotropic character, is a promising candidate to mimic more general Hamiltonians [@LahayeRPP09]. The investigation of dipolar quantum gases has been boosted by the realization of Bose-Einstein condensates of magnetic dipoles (namely, Cr, Er, and Dy [@DipolarBEC]) and by the recent progresses with heteronuclear molecules (such as RbK and NaK [@Heteronuclear]). A fundamental feature of the low-energy scattering between either magnetic or electric dipoles is the occurrence of dipolar-induced resonances (DIRs) when the dipole strength is varied [@MarinescuPRL98]. ![(Color online) Left: Representation of two polarized dipoles at relative distance $x$, harmonically trapped along a quasi-1D tube, of radial size $l_{\!\perp}=(\hbar/m\omega_{\!\perp})^{1/2}$. The polarization direction and the tube axis form an angle $\theta$. (a) Effective 1D dipolar interaction $V_{\rm d}$ \[Eq.\] and (c) corresponding toy-model potential $V_{\rm toy}$ \[Eq.\]. (b) Numerical results for the even-channel scattering length $a_{\rm d}^e$ associated to $V_{\rm d}$ \[solid (blue)\], as obtained from Eq. for $x\!=\!100l_{\!\perp}$. A single DIR occurs for ${\rho_\theta^}\simeq2.6$, in coincidence with the entrance of a dipolar bound state of energy $E_{\rm d}^{bs}$ \[dashed (green)\]. (d) Even-channel scattering length $a_{\rm toy}^e$ of the toy model. The qualitative behavior is the same of $a_{\rm d}^e$, with a resonance occurring for ${\rho_\theta^}\simeq3.3$, coinciding with the entrance of a bound state of energy $E_{\rm toy}^{bs}$ \[dashed (green)\]. \[FigSystem\]](Fig1){width=".95\textwidth"} In this work we consider the two-body low-energy scattering of polarized dipoles tightly confined along a quasi-one-dimensional (quasi-1D) tube [@BartoloPRA13; @GiannakeasPRL13], represented in Fig. \[FigSystem\](left). As described in Sec. \[SecTube\], a dipolar-induced resonance (DIR) emerges in this system. It is practical to investigate the scattering problem by introducing a toy-model potential that catches the main features of the dipolar interaction (Sec. \[SecToy\]). The toy-model scattering length can be calculated analytically and shows a resonant behavior analogous to that of the DIR. In Sec. \[SecToyGen\], the toy model is finally generalized to investigate the interplay of contact and non-zero-range potentials, giving an intuitive interpretation of the mechanism responsible for the resonance. Dipolar-induced resonance in a quasi-1D tube {#SecTube} ============================================ Generally, the dipolar interaction not only depends on the inter-particle distance, but also on the relative orientation of dipoles in space [@LahayeRPP09]. It is considerably simplified in the case of polarized dipoles, a scheme realizable by aligning the particles with an external field. In this case it takes the form $V_{\rm d}^{\rm 3D}\!=\!\hbar^2 r^* (1-3\cos^2\theta)/(mr^3)$, where we have introduced the dipolar length $r^\!=\!mD^2/\hbar^2$ (being $D$ the dipole strength and $m$ its mass). The system we want to investigate, represented in Fig. \[FigSystem\](left), consists of two polarized dipoles in a quasi-1D tube. The particles are harmonically confined in the radial direction with trapping frequency $\omega_{\!\perp}$, which fixes the length $l_{\!\perp}\!=\!(\hbar/m\omega_{\!\perp})^{1/2}$. Two classical dipoles restricted to 1D motion would simply repel (attract) each other for $\theta_c\!<\!\theta\!<\!\pi/2$ ($0\!<\!\theta\!<\!\theta_c$), where $\theta_c\!\simeq\!54.7^\circ$ is the “magic angle” at which the interaction vanishes. In the quantum case, we need to account for the radial extension of the particles’ wavefunctions. By restricting the analysis to scattering energies $E\!\ll\!\varepsilon_{\!\perp}\!=\!\hbar\omega_{\!\perp}$, we assume the dipoles to lie in the transverse ground state. Properly integrating out the transverse degrees of freedom [@EffectivePot], one can consider the tube as an effective 1D system, in which $V_{\rm d}^{\rm 3D}$ is replaced by the effective relative potential $$\label{DefVDI} V_{\rm d}(x)=\varepsilon_{\!\perp} \rho_\theta^\,\left[ w\left( \frac{x}{l_{\!\perp}} \right)-\frac{2}{3} \delta \left(\frac{x}{l_{\!\perp}} \right) \right],$$ with ${\rho_\theta^}\!=\!r^(1-3\cos^2\theta)/l_{\!\perp}$ and $w(\xi)\!=\!\sqrt{\pi/8}\, (1+\xi^2)\, \exp(\xi^2/2)\, {\rm erfc}(\|\xi\|/\sqrt{2})-\|\xi\|/2$. The potential $V_{\rm d}$ is plotted in Fig. \[FigSystem\](a) for ${\rho_\theta^}\!>\!0$, corresponding to the regime of classical repulsion. Distant particles perceive each other as classical dipoles, so that the potential has the expected $1/x^3$ behavior for $x\!\gg\!l_{\!\perp}$. On the other hand, getting closer, the dipoles’ quantum nature emerges: the interplay of transverse extension and interaction anisotropy results in a non-divergent repulsion plus an attractive contact term. For ${\rho_\theta^}\!<\!0$ the potential is reversed, resulting mainly attractive, and for any small value of $r^$ there exists at least one dipolar bound state. The scattering properties of the 1D potential $V_{\rm d}(x)$ can be investigated by numerically solving the relative-motion zero-energy Schrödinger equation $$\label{1DSch} \bigl[V_{\rm d}(x)-\hbar^2\partial_x^2/m\bigr]\psi_p(x)=0.$$ The parity index $p\!=\!e,o$ distinguishes between even and odd solutions, corresponding, respectively, to bosonic and fermionic particles. The $s$-wave scattering length for each channel is defined by $$\label{Defa} a_{\rm d}^p\!=\!\lim_{x\to\infty}[x-\psi_p(x)/\psi_p'(x)].$$ Due to the long-range character of $V_{\rm d}$, it is not possible to associate a well defined scattering length to it, since Eq. does not converge [@AstrakharchikPRA08]. Anyhow, one can evaluate Eq. for a large, but finite value $x\!=\!x_{\rm max}$. The even-channel scattering length of $V_{\rm d}$ for $x_{\rm max}\!=\!100l_{\!\perp}$ is presented in Fig. \[FigSystem\](b). A DIR occurs at ${\rho_\theta^}\!\simeq\!2.6$, due to the presence of the attractive $\delta$-term which takes over the repulsive wings, allowing for the existence of a dimer state of energy $E_{\rm d}^{bs}$ also in the regime of classical repulsion. Remarkably, the resonance position is unchanged for any $x_{\rm max}\!\gtrsim\!10l_{\!\perp}$. Numerical estimations suggest that this is the only DIR for ${\rho_\theta^}\!>\!0$ in a quasi-1D tube. The contact term is invisible to odd solutions, so that no resonances arise in this case. Toy model for the effective dipolar interaction {#SecToy} =============================================== To investigate the physical properties of $V_{\rm d}$, we propose a versatile toy model for which $a_{\rm toy}^p$ can be evaluated analytically, together with the energy $E_{\rm toy}^{bs}$ of the dimer state appearing at the resonance. Hence, we replace the fast-decaying wings of the real potential with a finite-range step function: $$\label{DefVtoy} V_{\rm toy}(x)=\varepsilon_{\!\perp}{\rho_\theta^}\, \left[ \frac{1}{2}\, \sigma\left(\frac{x}{l_{\!\perp}}\right) - \frac{2}{3}\, \delta\left(\frac{x}{l_{\!\perp}}\right) \right], \quad{\rm with}\quad \sigma\left(\frac{x}{l_{\!\perp}}\right)=\begin{cases} 1 & \|x\|\le l_{\!\perp}, \\ 0 & \|x\|>l_{\!\perp}. \end{cases}$$ The step width $2l_{\!\perp}$ corresponds to the region in which $V_{\rm d}$ deviates from the classical $1/x^3$ behavior, while its height $\varepsilon_{\!\perp}/2$ has been chosen so that areas under the wings $w$ and the step $\sigma$ are the same. The potential $V_{\rm toy}$ is plotted in Fig. \[FigSystem\](c), together with the corresponding even-channel scattering length $a_{\rm toy}^e({\rho_\theta^})$ \[Fig. \[FigSystem\](d)\]. The model is able to reproduce the DIR, with a resonance appearing at ${\rho_\theta^}\!\simeq\!3.3$ for even wave functions, which results close to the value ${\rho_\theta^}\!\simeq\!2.6$ found for the real potential, despite the simplicity of the model. Furthermore, the analytic expression of $a_{\rm toy}^e$ \[reported in Eq. for a more general case\] confirms that only one DIR exists for ${\rho_\theta^}\!>\!0$. Interplay of dipolar and contact interaction {#SecToyGen} ============================================ In addition to the dipolar interaction, we now consider the presence of a contact potential, whose strength is fixed by the 3D scattering length $a_{\rm3D}$. In a quasi-1D tube, this is taken into account by the effective contact potential $V_{\rm c}(x)\!=\!2\varepsilon_{\!\perp}(a_{\rm3D}/l_{\!\perp})\delta(x/l_{\!\perp})$ [@PetrovPRL00], valid as long as $a_{\rm3D}\!\ll\!l_{\!\perp}$. Thus, in a real system, one can change independently the contact and long-range terms of the total interaction $V_{\rm d}+V_{\rm c}$ by tuning $a_{3\rm D}$ via a Feshbach resonance [@ChinRMP10] and changing ${\rho_\theta^}$ with the polarizing field [@LahayeRPP09]. Correspondingly, we can generalize the toy-model potential $$\label{DefVtoyGen} \widetilde{V}_{\rm toy}(x)=\varepsilon_{\!\perp}\, \left[ \beta\, \sigma\left(\frac{x}{l_{\!\perp}}\right) + \alpha\, \delta\left(\frac{x}{l_{\!\perp}}\right) \right],$$ where the parameters $\alpha\!=\!2a_{\rm3D}/l_{\!\perp}-2{\rho_\theta^}/3$ and $\beta\!=\!{\rho_\theta^}/2$ set, respectively, the contact and non-zero-range interaction strengths. By analytically solving the Schrödinger equation for $\widetilde{V}_{\rm toy}$, one gets the scattering lengths: $$\label{DefaToyGen} \widetilde{a}_{\rm toy}^{e}(\alpha,\beta)=1-\frac{1}{\kappa} \frac{\alpha\sinh(\kappa)+2\kappa\cosh(\kappa)}{2\kappa\sinh(\kappa)+\alpha\cosh(\kappa)}, \quad \widetilde{a}_{\rm toy}^{o}(\alpha,\beta)=1-\frac{\tanh(\kappa)}{\kappa},$$ with $\kappa\!=\!\sqrt{\beta}$ for $\beta\!>\!0$ and $\kappa\!=\!i\sqrt{\|\beta\|}$ for $\beta\!<\!0$. ![(Color online) Left: Solid blue (dashed red) lines correspond to solutions of $1/\widetilde{a}_{\rm toy}^e\!=\!0$ ($1/\widetilde{a}_{\rm toy}^o\!=\!0$) \[cf. Eq. \]. The green dot-dashed line marks the condition $E_{\!\alpha}\!=\!E_{\!\beta}$ (cf. text). Gray-shadowed regions ($\beta\!<\!0$) indicate the direction in which the Tonks-Girardeau (TG) and super-Tonks-Girardeau (STG) limits are asymptotically reached. The green-shadowed quadrant ($\alpha\!<\!0$, $\beta\!>\!0$) is the one in which the DIR mechanism occurs. Right: Illustration of the generalized toy-model potential \[Eq. \] at the points (a,b,c,d,e,f) marked on the resonances diagram at left. Green horizontal lines represent the energy $E_\alpha$ of the $\delta$-sustained bound state, shifted upwards by the height of the repulsive energy barrier $E_\beta$. \[FigToyResonances\]](Fig2){width="95.00000%"} In Fig. \[FigToyResonances\](left) we show the position of the resonances of $\widetilde{a}_{\rm toy}^p$ varying $\alpha$ and $\beta$. The contact term is invisible to odd solutions, since $\psi_o(0)\!=\!0$. Hence, the corresponding resonances do not depend on $\alpha$. They exist only for $\beta\!<\!0$ and are simply those of a square well of depth $\|\beta\|$. Even solutions are, instead, strongly affected by the $\delta$-potential. No resonances exist in the purely repulsive quadrant $\alpha,\beta\!>\!0$. For $\alpha\!>\!0$ and $\beta\!<\!0$ \[Fig. \[FigToyResonances\](e)\], when $\alpha\!\to\!\infty$ the system reaches the Tonks-Girardeau limit of impenetrable particles [@GirardeauJMP60]: even wave functions acquire a zero at the origin to avoid a divergent contribution to the energy and, correspondingly, the even resonances tend asymptotically to the odd ones. A similar even-to-odd limit occurs in the region $\alpha,\beta\!<\!0$ \[Fig. \[FigToyResonances\](f)\]. For $\alpha\!\to\!-\infty$ the $\delta$-potential sustains only a single, infinitely deep bound state, so that the other even wavefunctions must acquire a zero at the origin to keep finite their energy. The dipoles become effectively impenetrable, reaching the super-Tonks-Girardeau regime [@SuperTonks]. The DIR occurs if $\alpha\!<\!0$ and $\beta\!>\!0$. It is a (single-channel) shape resonance which results from a competition between the attractive delta term and the repulsive step potential, and it can be understood intuitively as follows. In the absence of the step potential ($\beta=0$), the $\delta$ term would support a bound state with energy $E_\alpha\!=\!-\varepsilon_{\!\perp}\alpha^2/4\!<\!0$. If we now add to the Hamiltonian a step potential whose height $E_\beta\!=\!\beta\varepsilon_{\!\perp}\!>\!0$ is smaller than $\sim\!\|E_\alpha\|$, the discrete level survives and its energy is shifted upwards by $\sim\!E_\beta$. On the other hand, if $E_\beta\!\gtrsim\!\|E_\alpha\|$, the discrete level supported by the delta dissolves into the continuum and disappears. The resonance occurs at the threshold between these two regimes, i.e. for $E_\beta\!\sim\!\|E_\alpha\|$. Our results \[Fig. \[FigToyResonances\](left)\] show that this condition is asymptotically exact (if $\|\alpha\|$ and $\beta$ are both large, the resonance occurs for $E_{\beta}\!=\!\|E_{\alpha}\|$). Conclusions and Perspectives {#SecConc} ============================ We presented a short study on the dipolar interaction in a quasi-1D tube, discussing the emergence of a DIR and the appearance of a dimer state in a regime where classical dipoles would simply repel each other. This feature can be reproduced using a simple toy-model potential for which the low-energy scattering properties are analytically determined. A generalized version of the toy model allows to investigate the interplay of contact and step potential, giving an intuitive description of the DIR as a shape resonance. These results can be extended for the actual quasi-1D dipolar plus contact potential $V_{\rm d}+V_{\rm c}$, which will be likely to present analogous features. Furthermore, interesting effects may arise for $a_{\rm3D}\!\gtrsim\!l_{\!\perp}$, due to the coaction of DIR and CIR. These investigations on the two-body scattering are the building block for the analysis of many-body dipolar systems in optical lattices [@BartoloPRA13] and are useful in determining the stability conditions for systems of attracting dipoles. The effects of the DIR may be probed experimentally by studying the energy levels with spectroscopic techniques or by looking at the two-body losses as a function of the dipolar strength. This work has been supported by ERC through the QGBE grant and by Provincia Autonoma di Trento. [14]{} I. Bloch, J. Dalibard, and S. Nascimbène, Nat. Phys. [8]{}, 267 (2012). C. Chin, R. Grimm, P. S. Julienne, and E. Tiesinga, Rev. Mod. Phys. [82]{}, 1225 (2010). M. Olshanii, Phys. Rev. Lett. [81]{}, 938 (1998). T. Lahaye, C. Menotti, L. Santos, M. Lewenstein, and T. Pfau, Rep. Prog. Phys. [72]{}, 126401 (2009); M. A. Baranov, M. Dalmonte, G. Pupillo and P. Zoller, Chem. Rev. [112]{}, 5012 (2012). A. Griesmaier, J. Werner, S. Hensler, J. Stuhler, and T. Pfau, Phys. Rev. Lett. [94]{}, 160401 (2005); K. Aikawa, A. Frisch, M. Mark, S. Baier, A. Rietzler, R. Grimm, and F. Ferlaino, Phys. Rev. Lett. [108]{}, 210401 (2012); M. Lu, N.Q. Burdick, S.H. Youn, and B.L. Lev, Phys. Rev. Lett. [107]{}, 190401 (2011). K.-K. Ni, S. Ospelkaus, M. G. H. de Miranda, A. Pe’er, B. Neyenhuis, J. J. Zirbel, S. Kotochigova, P. S. Julienne, D. S. Jin, and J. Ye, Science [322]{}, 231 (2008); C.-H. Wu, J. W. Park, P. Ahmadi, S. Will, and M. W. Zwierlein, Phys. Rev. Lett. [109]{}, 085301 (2012). M. Marinescu and L. You, Phys. Rev. Lett. 81, 4596 (1998). N. Bartolo, D.J. Papoular, L. Barbiero, C. Menotti, and A. Recati, Phys. Rev. A 88, 023603 (2013). P. Giannakeas, V. S. Melezhik, and P. Schmelcher, Phys. Rev. Lett. 111, 183201 (2013). S. Sinha and L. Santos, Phys. Rev. Lett. [99]{}, 140406 (2007); F. Deuretzbacher, J. C. Cremon, and S. M. Reimann, Phys. Rev. A [81]{}, 063616 (2010); [87]{}, 039903(E) (2013). G. E. Astrakharchik and Yu. E. Lozovik, Phys. Rev. A [77]{}, 013404 (2008). D.S. Petrov, G.V. Shlyapnikov, and J.T.M. Walraven, Phys. Rev. Lett. [85]{}, 3745 (2000). M. Girardeau, J. Math. Phys. [1]{}, 516–523 (1960). G.E. Astrakharchik, D. Blume, S. Giorgini, and B.E. Granger, Phys. Rev. Lett. [92]{}, 030402 (2004); G.E. Astrakharchik, J. Boronat, J. Casulleras, and S. Giorgini, Phys. Rev. Lett. [95]{}, 190407 (2005); E. Haller, M. Gustavsson, M. J. Mark, J. G. Danzl, R. Hart, G. Pupillo, H.-C. Nägerl, Science [325*]{}, 1224 (2009). [^1]:	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Galactic centre (GC) is a crowded environment: observations have revealed the presence of (molecular, atomic and ionized) gas, of a cusp of late-type stars, and of $\sim{}100$ early-type stars, about half of which lying in one or possibly two discs. In this paper, we study the perturbations exerted on a thin stellar disc (with outer radius $\sim{}0.4$ pc) by a molecular cloud that falls towards the GC and is disrupted by the supermassive black hole (SMBH). The initial conditions for the stellar disc were drawn from the results of previous simulations of molecular cloud infall and disruption in the SMBH potential. We find that most of the gas from the disrupted molecular cloud settles into a dense and irregular disc surrounding the SMBH. If the gas disc and the stellar disc are slightly misaligned ($\sim{}5-20^\circ{}$), the precession of the stellar orbits induced by the gas disc significantly increases the inclinations of the stellar orbits (by a factor of $\sim{}3-5$ in 1.5 Myr) with respect to the normal vector to the disc. Furthermore, the distribution of orbit inclinations becomes significantly broader. These results might be the clue to explain the broad distribution of observed inclinations of the early-type stars with respect to the normal vector of the main disc. We discuss the implications for the possibility that fresh gas was accreted by the GC after the formation of the disc(s) of early-type stars.' title: Perturbations induced by a molecular cloud on the young stellar disc in the Galactic Centre --- methods: numerical – stars: kinematics and dynamics – Galaxy: centre – black hole physics – ISM: clouds Introduction ============ The compact radio source Sgr A$^\ast{}$, located at the very centre of our Galaxy, coincides with a high concentration of mass ($\approx{}4\times{}10^6$ M$_\odot$), almost certainly a supermassive black hole (SMBH, @genzel03; @schodel03; @ghez03, 2005). More than a hundred young massive stars have been observed in the vicinity of Sgr A$^\ast{}$ (@krabbe91; @morris93; @krabbe95; @genzel03). Many of them are O-type and Wolf-Rayet (WR) stars. About half of the early-type stars lie in a thin disc with radius 0.04 pc$\,{}\lesssim{}r\lesssim{}0.5$ pc and average eccentricity $e\sim{}0.3-0.4$ (@bartko09; @lu09; @yelda12; @do13; @lu13). This disc is called clock-wise (CW) disc, because it shows CW motion when projected on the plane of the sky (@genzel03; @paumard06). The CW disc is likely warped, as the orientation of its normal vector changes by several degrees ($\sim{}60^\circ{}$, according to @bartko09) from its inner to its outer edge. A fraction of the remaining early-type stars show counterclockwise motion, which may indicate the presence of a second dissolving disc (@lu06, 2009; @bartko09). The age of the observed early-type stars is $t_{\rm age}\approx{}2.5-6$ Myr (@lu13; see @paumard06 for a previous estimate $t_{\rm age}=6\pm{}2$ Myr). Furthermore, @yusef-zadeh13 found indications of gas outflows, suggesting recent star formation (10$^{4-5}$ yr) within 0.6 pc of SgrA$^\ast$. The $\sim{}20$ stars closest to SgrA$^\ast{}$ ($\lesssim{}0.04$ pc $\sim{}1$ arcsec) are B stars, with an age 20-100 Myr. These, named the S-stars, have very eccentric and randomly oriented orbits (@schodel03; @ghez03, 2005; @eisenhauer05; @gillessen09). The ensemble of the (both young and old) stars in the central few parsecs is often referred to as the nuclear star cluster (NSC) of the Milky Way (MW). The Galactic centre (GC) is a very crowded environment not only for the stellar population, but also for the gas. A molecular torus, the circumnuclear ring (CNR, @genzel85; @gusten87; @yusef-zadeh04; @christopher05; @oka11), is located at $\sim{}2$ pc from SgrA$^\ast$, and is on the verge of forming stars (@yusef-zadeh08). Furthermore, the innermost 3 pc of the GC are extremely rich in ionized gas (e.g., @zhao09, and references therein). Finally, two molecular clouds (the M–0.02–0.07 and the M–0.13–0.08 cloud, @solomon72; @novak00) lie within 20 pc of the GC. The origin of the early-type stars is puzzling, as the strong tidal field exerted by the SMBH is expected to disrupt any molecular cloud before it reaches the distance of the CW disc. On the other hand, a disrupted molecular cloud might spiral towards the SMBH and form a gas disc around it, sufficiently dense to fragment into stars (@levin03; @goodman03; @goodman04; @milosavljevic04; @nayakshincuadra05; @rice05; @alexander08; @collin08; @bonnell08; @mapelli08; @wardle08; @hobbs09; @alig11; @jiang11; @namekata11; @mapelli12, hereafter M12; @lucas13; @alig13). In particular, M12 simulated the infall of a molecular cloud towards the GC. In the M12 simulations, the molecular cloud is disrupted by the SMBH tidal shear, and settles into a dense gas disc, which fragments into self-gravitating clumps. These clumps, or proto-stars, lie in a disc at a distance of 0.1–0.4 pc with moderately eccentric orbits ($e\sim0.2-0.4$). The properties of the stellar disc reproduce quite well the observations of the CW disc, but cannot explain the counterclockwise stars as well as the S-stars. On the other hand, M12 integrates the evolution of the cloud only for $\sim{}0.5$ Myr, while the early-type stars are $\sim{}2.5-6$ Myr old (@lu13). The GC is such a crowded environment that many external forces might have influenced the orbital evolution of the CW disc after its formation. First, the stellar cusp of late-type stars induces precession on the orbits of a stellar disc (e.g. @lockmann09a). This is expected to significantly affect the eccentricity distribution (@madigan09; @gualandris12, hereafter G12). Furthermore, a massive perturber, such as an intermediate-mass black hole (@gualandris09; @gualandris10; @perets10), the CNR (@haas11a; @haas11b), a second stellar disc (@lockmann08; @lockmann09b; @lockmann09a), or a second molecular cloud infalling towards SgrA$^\ast$ is expected to induce a precession on the CW disc, and it may also excite the Kozai resonance (@kozai62; @lidov62). In this paper, we simulate the infall of a second molecular cloud toward the SMBH, and we study the influence that the second cloud has on the orbits of a pre-existing stellar disc. As initial conditions for the pre-existing stellar disc, we adopt the outcomes of run E in M12. N-body simulations ================== For our simulations, we used the N-body/smoothed particle hydrodynamics (SPH) code <span style="font-variant:small-caps;">gasoline</span> (@wadsley04), upgraded with the @read10 optimized SPH (OSPH) modifications, to address the SPH limitations outlined, most recently, by @agertz07. Table 1 shows a summary of the runs that will be presented in this paper. In all the runs, we simulate a disc of 1252 star particles orbiting the SMBH. The SMBH is represented by a sink particle, with initial mass $M_{\rm SMBH}=3.5\times{}10^6$ M$_{\odot{}}$ (@ghez03), sink radius $r_{\rm acc}=5\times{}10^{-3}$ pc and softening radius $\epsilon{}=1\times{}10^{-3}$ pc. The disc of star particles is obtained from run E of M12, with the same procedure as described in G12. In particular, we assume that each self-bound clump formed in run E of M12 becomes a star, without accreting any further gas particles after $t=4.8\times{}10^5$ yr[^1]. Thus, we replace each gas particle with a single star particle having mass equal to the total mass of the gas clump, position and velocity corresponding to those of the centre-of-mass of the clump. The resulting simulated mass of the disc is $4.1\times{}10^3$ M$_{\odot{}}$ (see M12). The best-fitting initial mass function (IMF) is given by a single power law with index $\alpha{}=1.50\pm{}0.06$ (see M12 for details), consistent with Lu et al. (2013, $\alpha{}=1.7\pm{}0.2$). Considering that M12 simulations cannot resolve stellar masses $<1.3$ M$_{\odot{}}$, this corresponds to an expected total mass of the disc $\approx{}4.5\times{}10^3$ M$_{\odot{}}$ (assuming $\alpha{}=1.50$). This value is consistent with most observations (e.g. @paumard06; @bartko09), but is a factor of $\sim{}3$ lower than the most recent results (@lu13). Finally, the stellar disc has average semi-major axis $\langle{}a\rangle{}=0.21\pm{}0.04$ pc, average inclination[^2] $\langle{}{\theta{}}\rangle{}=2.4\pm{}1.5^\circ{}$, and average eccentricity $\langle{}e\rangle{}=0.29\pm{}0.04$. The high density and the clumpiness of the gas disc at $t=4.8\times{}10^5$ yr in run E of M12 make prohibitive to continue the simulation as it is. Thus, in our initial conditions, we instantaneously removed the gas left in the first disc, to prevent our simulations from stalling before the infall of the second cloud. For the issues related to this assumption, see the discussion in Section \[issue\]. In runs A2 and B2, we also add a rigid potential, to account for the stellar cusp surrounding Sgr A$^\ast{}$. The overall density profile of the stellar cusp goes as $\rho{}(r)=2.8\times{}10^6\,{}{\rm M}_{\odot{}}\,{}{\rm pc}^{-3}\,{}(r/\textrm{0.22 pc})^{-\gamma}$, where $\gamma=1.2$ (1.75) for $r<0.22$ pc ($r>0.22$ pc), consistent with the values reported in @schodel07. Runs A2 and B2 will be referred to as runs ‘with cusp’, while runs A1 and B1 will be dubbed runs ‘without cusp’. Run Stellar Cusp$^{\rm a}$ Perturber$^{\rm b}$ $\alpha{}$ ($^\circ{}$)$^{\rm c}$ ----- ------------------------ --------------------- ----------------------------------- A1 No No – A2 Yes No – B1 No Yes 0 B2 Yes Yes 0 : Initial conditions. In runs B1 and B2, we simulate the infall of a second molecular cloud towards the GC, to study the perturbations induced by the cloud on the orbits of the stellar disc. In runs A1 and A2, we do not include the second molecular cloud: the stellar disc evolves under the influence of the SMBH, and, in the case of run A2, also under the influence of the rigid stellar cusp. Runs A1 and A2 were performed just for comparison with the other two runs. The perturber cloud is simulated as in M12. In particular, it is a spherical cloud with a radius of 15 pc and a mass of $1.3\times{}10^5$ M$_{\odot{}}$. The cloud is seeded with supersonic turbulent velocities and marginally self-bound (see @hayfield11). The centre-of-mass of the cloud is initially at 25 pc from the SMBH. The orbit of the cloud was chosen so that the impact parameter with respect to the SMBH is $10^{-2}$ pc and the initial velocity is one tenth of the escape velocity from the SMBH at the initial distance (i.e. the orbit is bound and highly eccentric). The orbit of the cloud lies in the plane of the stellar disc ($\alpha{}=0$, see Table 1). As in run E of M12, we include radiative cooling in all our simulations. The radiative cooling algorithm is the same as that described in @boley09 and in @boley10. @dalessio01 opacities are used, with a 1 $\mu{}$m maximum grain size. The irradiation temperature is $T_{\rm irr}=100$ K everywhere. The mass of the gas particles in each simulation is 0.4 M$_{\odot{}}$ and the softening length $10^{-3}$ pc. Given the low mass resolution, these simulations cannot be used to study gas fragmentation. In a forthcoming paper, we will show new high-resolution runs focused on this aspect, while in the current paper we are interested in the role of the cloud as perturber of the pre-existing stellar disc. The code used for our simulations adopts a kick-drift-kick leapfrog integrator. This scheme is known to be rather inaccurate in the short-term integration of the orbits, while it preserves well energy and angular momentum on the long term evolution (e.g., @zemp07; @dehnen11). This may damp any secular changes in eccentricity (e.g., @zemp07). On the other hand, the adopted code is the best compromise between describing the orbits accurately and following the thermodynamical and dynamical evolution of a living cloud (rather than assuming a rigid potential for the gas, as done in previous work). We checked the limits of our numerical approach by comparing them with the results of G12, who used a Hermite integration scheme (implemented in the direct-summation N-body code $\phi{}$<span style="font-variant:small-caps;">GRAPE</span>, @harfst07). In Fig. \[fig:fig1\], the distribution of semi-major axes and inclinations obtained at $t=1.5$ Myr by G12 are compared with the same distributions derived at $t=1.5$ Myr for our run A2 (whose main ingredients are the same as in G12). No evident differences appear between G12 and our run A2, by looking at Fig. \[fig:fig1\]. The Kolmogorov-Smirnov test indicates a 80 (30) per cent probability that the distribution of semi-major axes (inclinations) is the same in G12 and in run A2. Such probabilities are quite high, considering the intrinsic differences of the two runs. Other checks (e.g. reducing the time stepping, changing the number of neighbours, substituting the SMBH particle with a rigid potential, integrating some analytic orbits) demonstrated that our results are robust and fairly describe the effects of precession on the stellar orbits. Furthermore, stochastic fluctuations (between different realizations of run A2) change the orbital parameter distributions by less than $\sim{}10$ per cent (i.e. do not significantly affect the mean values of the orbital parameters). The simulations presented in this paper ran on the Fermi IBM Blue Gene/Q at CINECA (through CINECA Award N. HP10CL51UF, 2012). Runs B1 and B2 required approximately 100k CPU hours each. Results ======= The disruption of the cloud and the formation of the gas disc {#gasev} ------------------------------------------------------------- In our runs B1 and B2, the second molecular cloud spirals towards the GC, and is quickly disrupted by the tidal force exerted by the SMBH. The disruption of the cloud starts at $t\approx{}10^5$ yr since the beginning of the simulations, i.e. $\approx{}6\times{}10^5$ yr after the infall of the first cloud (simulated in M12). At $t\approx{}3-5\times{}10^5$ yr since the beginning of the simulation, a dense gaseous ring forms, surrounding the SMBH. In the next Myr, the ring is fueled by fresh gas coming from the remnant of the disrupted molecular cloud (see Fig. \[fig:fig2\]). For this reason, the central gas disc looks more like a series of streamers than a coherent thin disc (see Fig. \[fig:fig3\]). We notice that the gravitational potential exerted on the cloud is slightly different in run B1 and B2, since a spherical stellar cusp is included only in run B2. Thus, the orbits of the streamers are slightly different in the two runs. The initial inclination of the orbit of the cloud is nearly preserved during the evolution and the tidal disruption. In runs B1 and B2, the orbit of the second molecular cloud is in the plane of the stellar disc ($\alpha=0$), and the gas disc and the pre-existing stellar disc are nearly coplanar (see Fig. \[fig:fig3\]). On the other hand, the gas disc is not a rigid body, but is instead composed of many concentric annuli and streamers, all of which may have slightly different ($\sim{}5^\circ{}$) mutual inclinations (see Figs. \[fig:fig3\], \[fig:fig4\] and \[fig:fig5\]). Furthermore, the fact that the cloud is seeded with supersonic turbulence implies that clumps and filaments form in the molecular cloud before disruption. This determines local changes in the geometry of the infalling cloud, and is source of important stochastic fluctuations in the structure of the gas disc, between different simulations. In particular, we remind that the initial radius of the cloud is $\sim{}15$ pc, while the radius of the stellar disc is less than one pc: even small density fluctuations inside the cloud can significantly affect the geometry of the encounter between the stellar disc and the cloud. In particular, Figs. \[fig:fig4\] and \[fig:fig5\] show how the gas disc assembles between 0.5 and 1.5 Myr. We notice that the orbits of the streamers can change with time, as long as the disrupted molecular cloud re-fuels the inner disc with fresh gas. This makes the interplay between stellar disc and gas disc(s) very complicated. It is apparent that while the stars define a thin and coherent disc, the gas is distributed in a series of streamers, with different angular momentum directions. Finally, by the end of the simulation, when the cloud is entirely disrupted and the gas discs are completely settled (i.e. the shocks and the cooling induced by the disruption are over and almost no new gas is accreted), the situation is even more complex. The bottom panels of Figs. \[fig:fig4\] and \[fig:fig5\] show the inner 2 pc of run B2 at $t=1.5$ Myr. It is apparent that there are at least two different gas rings: a inner one, with radius $\sim{}0.4$ pc (similar to the stellar disc), and a outer one, with radius $\sim{}1.5$ pc. While the inner gas disc is almost coplanar with the stellar disc, the angular momentum vector of the outer gas disc is inclined by $\sim{}15-20^\circ{}$ with respect to the angular momentum vector of the stellar disc. The outer gas ring may play the same role as the CNR observed in our Galaxy. Figs. \[fig:fig6\] and \[fig:fig7\] show the angular momentum distribution of gas particles moving on bound orbits around the SMBH (with semi-major axis $0.1\le{}a/{\rm pc}\le{}10$) in run B1 and B2, respectively. The top and the bottom panel refer to time $t=0.5$ and 1.5 Myr, respectively. The coordinate system refers to the total angular momentum of the stellar disc at a given time $t$, as defined in Appendix A. We remind that particles belonging to a razor thin disc aligned with the stellar disc have $\cos{\theta{}}=1$ and can assume any possible value of $\phi{}$. Instead, a razor thin disc that is misaligned with respect to the stellar disc defines an infinitely small circle in the $\cos{\theta{}},\,{}\phi{}$ plane. The larger the opening angle of the disc, the larger the circle it defines in the $\cos{\theta{}},\,{}\phi{}$ plane. The inclination between the stellar disc and a second misaligned disc is given by the difference in $\theta{}$. From Figs. \[fig:fig6\] and \[fig:fig7\] it is apparent that the gas discs are neither razor-thin nor regular structures. In particular, the top panels of Figs. \[fig:fig6\] and \[fig:fig7\] show that the situation is rather complicated at $t=0.5$ Myr, i.e. during the first phase of the disc assembly (see also Figs. \[fig:fig4\] and \[fig:fig5\]). At $t=1.5$ Myr the gas discs are more coherent, especially in the case of run B2 (bottom panel of Fig.\[fig:fig7\]), where the angular momentum vectors concentrate at $\cos{\theta{}}=0.964$ and $\phi{}=3$ rad. In both runs B1 and B2, the total angular momentum direction of the main gas disc is similar to the one of the stellar disc, but is slightly offset. The offset at $t=1.5$ Myr is $\sim{}10^\circ$ and $\sim{}15^\circ$ in runs B1 and B2, respectively. These issues are very important for the effects of the gas on the stellar orbits, as it will be discussed in Section \[starev\]. Timescales for precession ------------------------- The two main sources of perturbation of the stellar orbital inclinations in our simulations are: (i) the precession induced by the rigid stellar cusp; (ii) the precession induced by the disrupted molecular cloud. In this section, we give an estimate of the timescales associated with these precessions, for the quantities involved in our simulations. \(i) The stellar cusp, i.e. a spherical potential, induces a precession of the orbits of the disc stars on a timescale (@ivanov05; @lockmannbaumgardt08; @lockmann09a; G12) $$T_{\rm cusp}=\frac{M_{\rm SMBH}}{M_{\rm cusp}}\,{}T_{\rm orb}\,{}f(e),$$ where $M_{\rm SMBH}$ is the mass of the SMBH, $T_{\rm orb}$ is the orbital period of a disc star, $M_{\rm cusp}$ is the mass of the cusp inside the stellar orbit, and $f(e)=\frac{1+\sqrt{1-e^2}}{\sqrt{1-e^2}}$ is a function of the eccentricity $e$ of a disc star. The main effect of this precession is pericentre advance (e.g. @subrhaas12). (ii) The disrupted molecular cloud settles into a dense and irregular gas disc. The newly formed gas disc and the pre-existing stellar disc exert mutual torques on each other, and thus precess about each other. A star orbiting a SMBH of mass $M_{\rm SMBH}$ with semi-major axis $a$, at an inclination $\beta{}$ relative to a gas disc of radius $R_{\rm DISC}$ and mass $M_{\rm DISC}$ precesses on a timescale (@nayakshin05; @lockmann08; @subr09): $$\label{eq:eq2} T_{\rm DISC}\sim{}\frac{T_{\rm K}}{\cos{\beta}}\,{}\frac{\sqrt{1-e^2}}{1+\frac{3}{2}e^2},$$ where $$T_{\rm K}\equiv{}\frac{M_{\rm SMBH}}{M_{\rm DISC}}\,{}\frac{R_{\rm DISC}^3}{a^{3/2}\,{}\sqrt{G\,{}M_{\rm SMBH}}}. $$ $T_{\rm DISC}$ is the timescale for the change of the longitude of the node, i.e. for the precession around the symmetry axis of the gaseous disc (@subrhaas12). $T_{\rm DISC}$ depends on the inclination between the two discs, and in particular $T_{\rm DISC}$ tends to infinity if $\beta{}\approx{}90^\circ{}$: the precession will be very slow if the two discs are perpendicular. Instead, $T_{\rm DISC}$ is minimised, or in other words, the precession occurs at its maximal rate, if the two discs are coplanar. Furthermore, the precession timescale depends on the distance of the star from the SMBH ($T_{\rm DISC}\propto{}a^{-3/2}$). Thus, stars with different $a$ will precess with different speed. If $\beta{}=0$ and the two discs are razor thin, then the precession does not affect the inclinations, as it is just a change of the longitude of the node. Instead, if $\beta{}\gtrsim{}5^{\circ{}}$, this precession significantly influences the inclinations of stellar orbits (@lockmann08; @subr09). In the latter case, the orbits of the outer stars will become inclined with respect to the orbits of the inner stars, producing a warp in the disc, and increasing its thickness. Kozai oscillations are another possible effect induced by the interaction between two discs. They trigger a periodic increase of the eccentricity together with small variations of the inclination, and their characteristic timescale is $\propto{}T_{\rm K}\,{}\sin^{-2}{\beta{}}$ (see @subr09). Kozai oscillations are efficiently damped in presence of a spherical cusp, if $T_{\rm K}>T_{\rm cusp}$ (which is the case of our run B2, see Fig. \[fig:fig8\]). As our integrator is not sufficiently accurate for large eccentricities, we select a cloud orbit with $\alpha{}=0$, so that the final inclination between gas disc and stellar disc remains mild. Thus, Kozai oscillations are negligible[^3]. Finally, relativistic effects are not included in our code. Thus, we cannot account for the relativistic precession of the orbits due to the SMBH. The timescale for relativistic precession is (see @gualandris09): $$T_{\rm RP}=\frac{2\,{}\pi{}\,{}c^2\,{}(1-e^2)a^{5/2}}{3\,{}(G\,{}M_{\rm SMBH})^{3/2}},$$ where $c$ is the speed of light. Fig. \[fig:fig8\] shows a comparison of the relevant timescales for our simulations. It is apparent that the timescale for relativistic precession is $\gtrsim{}3$ orders of magnitude longer than the cusp precession timescale. Thus, relativistic precession is negligible for our simulated stellar disc, and the fact that we do not include relativistic effects in our simulations is not an issue. Fig. \[fig:fig8\] shows that $T_{\rm cusp}$ is almost independent of the semi-major axis, for the potential well of the stellar cusp in the GC (see G12). On the contrary, $T_{\rm K}$, which measures the perturbation by the gas disc, depends on the semi-major axis, and is about one order of magnitude shorter for the outermost simulated stars than for the innermost ones. $T_{\rm K}$ is always longer than $T_{\rm cusp}$. On the other hand, the timescale of disc precession ($T_{\rm DISC}$) depends on the angle $\beta{}$ between the stellar orbit and the gas disc. Furthermore, to derive $T_{\rm K}$ in Fig. \[fig:fig8\], we assumed that the disc is regular, that its mass is $4\times{}10^4$ M$_\odot$, and that its radius is 0.5 pc. Each of these assumptions is a source of uncertainty, as the disc is not regular, the radius is not clearly defined and the mass depends on the radius. Finally, Fig. \[fig:fig8\] is derived in the case of run B2. The other runs have small differences, which depend on the eccentricity distribution. In the next section, we will consider all the runs in detail. The evolution of the stellar disc in the simulations {#starev} ---------------------------------------------------- In this section, we investigate the effect of the perturbations generated by the gas disc on the stellar disc, in our simulations B1 and B2. In agreement with the fact that the precession cannot affect the semi-major axis of the orbit (from energy conservation, e.g. @subr09), the distribution of semi-major axes is not perturbed by the second cloud (bottom row of Fig. \[fig:fig9\]). Instead, the distribution of inclinations is significantly affected by the gas disc, in both run B1 and B2 (top row of Fig. \[fig:fig9\]). In fact, the precession induced by the presence of the gas disc is maximum when $\beta{}$ is small (see equation \[eq:eq2\]), and its effect on inclinations is not suppressed by the presence of the stellar cusp (see e.g. @subr09). Actually, if the gas disc was perfectly razor-thin and coplanar with the stellar disc ($\beta{}=0$), we would not have found significant changes in the inclinations (e.g. @subr09), but just the precession of the ascending node. On the other hand, as we showed in Section \[gasev\], the simulated gas disc is neither razor thin nor perfectly coplanar with the stellar disc. Values of $\beta{}$ as high as $5-20^\circ{}$ occur in some of the streamers, and lead to the increase of the inclinations in a rather complex way. The fact that stellar inclinations in run B2 are generally larger than those in run B1 can be attributed to the different potential well. In fact, in run B2 the cloud is disrupted marginally faster than in run B1, as the potential well of run B2 is the combination of the SMBH and of the stellar cusp. For this reason, the stars in run B2 are exposed to the influence of the gas disc for a longer time than the stars in run B1. Also, stochastic fluctuations leading to slightly different inclinations of the streamers might play an important role. Table 2 summarizes some of the main statistical properties of the simulations. In run B1 (B2), the average inclination passes from 2.4$^\circ{}$ (2.4$^\circ{}$) at $t=0$ to 6.6$^\circ{}$ (12.9$^\circ{}$) at $t=1.5$ Myr since the beginning of the simulation (corresponding to $t=1.98$ Myr since the beginning of run E in M12). The maximum achieved inclination is 17.3$^\circ{}$ and 27.2$^\circ{}$ in run B1 and B2, respectively. We find no statistically significant differences for the average semi-major axes and for the average eccentricities. We also checked whether or not the orbital properties of the simulated stars depend on the stellar mass, and we found no statistically significant difference (at more than 1$\sigma{}$) between stars with mass $>5$ M$_\odot{}$ and stars with mass $<5$ M$_\odot{}$. [lllllll]{} Run & $\langle{}a\rangle{}$ & $\langle{}e\rangle{}$ & $\langle{}\theta{}\rangle{}$ & $a_{\rm max}$ & $e_{\rm max}$ & $\theta{}_{\rm max}$\ & (pc) & & ($^\circ{}$) & (pc) & & ($^\circ{}$)\ A1 & $0.21\pm{}0.05$ & $0.3\pm{}0.1$ & $3.3\pm{}1.7$ & 0.40 & 0.6 & 11.9\ A2 & $0.21\pm{}0.05$ & $0.3\pm{}0.1$ & $3.4\pm{}1.9$ & 0.47 & 0.7 & 14.4\ B1 & $0.21\pm{}0.05$ & $0.3\pm{}0.1$ & $6.6\pm{}3.5$ & 0.37 & 0.6 & 17.3\ B2 & $0.21\pm{}0.05$ & $0.3\pm{}0.1$ & $12.9\pm{}5.0$ & 0.49 & 0.8 & 27.2\ G12 & $0.21\pm{}0.05$ & $0.3\pm{}0.2$ & $3.5\pm{}2.1$ & 0.49 & 0.7 & 13.8\ In Fig. \[fig:fig10\], the inclinations are shown as a function of the semi-major axis for run A2 (without gas disc) and run B2 (with gas disc), at $t=1.5$ Myr. For comparison, we report the initial conditions. In Fig. \[fig:fig10\], the inclinations of the individual stellar orbits were re-binned in four different bins (so that each bin contains at least 25 stars). In the initial conditions, the average inclinations in each bin are well below $7^\circ$. The inner stellar orbits appear more inclined than the outer ones, because the gas disc in run E of M12 formed with this feature (see M12 for details). In run B2, the average inclinations ($>10^\circ$) are significantly higher than those in the initial conditions, and it is apparent that the outer stellar orbits achieve (on average) higher inclinations than the inner ones. This is an effect of the precession induced by the gas disc: as $T_{\rm K}\propto{}a^{-3/2}$, the outer orbits are expected to precess faster than the inner ones. As a further support that this is a genuine effect of precession, induced by the gas disc, the inclinations in run A2 (without gas) remain almost the same as they are in the initial conditions. Furthermore, Fig. \[fig:fig10\] indicates that the half-opening angle of the stellar disc at $t=1.5$ Myr is as large as $\sim{}15^\circ{}$ in run B2, while it does not exceed $\sim{}7^\circ{}$ in run A2. Because of this semi-major axis dependent change in the inclinations, the simulated stellar disc in run B2 (and, similarly, in run B1) may appear tilted and/or warped. Recent observations (@bartko09, 2010; @lu09; @do13; @lu13) show that the opening angle of the CW disc is only $\sim{}10^\circ{}-14^\circ{}$, but about half of the early-type stars in the inner $1-10$ arcsec ($0.04-0.4$ pc) do not belong to the CW disc. Furthermore, the probability of early-type stars being members of the CW disc decreases with increasing projected distance from Sgr A$^\ast{}$ (@bartko09; @lu09). Finally, the CW disc does not seem a flat structure, but rather a significantly warped ($\sim{}64^\circ{}$, @bartko09) and tilted object. Our simulations suggest a reasonable interpretation for such observations: the precession exerted by a slightly misaligned gas disc enhances the inclinations of the outer stellar orbits with respect to the inner stellar orbits. Thus, while the inner disc remains quite coherent, the outer stellar orbits change angular momentum orientation till they may even lose memory of their initial belonging to the same disc. The result is a tilted/warped disc, which is being dismembered in its outer parts. By the time our simulations were stopped ($1.5$ Myr), the warping is still lower than the observed effect. On the other hand, it is reasonable to expect that in few Myr the simulated disc will match the observed features. A similar interpretation was already proposed by Šubr et al. (2009; see also @haas11a, 2011b), which pointed out the possible influence of the CNR on the stellar disc. Our paper supports the results by @subr09, by means of the first self-consistent N-body/SPH simulations. Other studies considered similar effects induced by a second stellar disc (@lockmann09a, 2009b). Alternatively, radiation pressure instability or the Bardeen-Petterson effect were also invoked as possible sources of warping, both driven by the central SMBH (@ulubay13). It is likely that the complex distribution of early-type stars in the GC is due to the concurrence of different proposed mechanisms. One or two gas discs? {#issue} --------------------- One of the most simplistic assumptions of our simulations is that no gas is left in the initial disc, as we instantaneously removed the remaining gas disc in run E of M12. This is necessary to prevent our simulations from stalling before the infall of the second cloud. On the other hand, we do not know whether the feedback from the SMBH and/or from the newly born stars is sufficiently efficient to remove all the gas in $\approx{}0.5$ Myr (e.g. @alexander12). For a typical young star cluster with mass $\sim{}5\times{}10^3$ M$_{\odot{}}$ and radius $\sim{}0.5$ pc, gas expulsion is expected to occur in $\sim{}10^5$ yr, because of the photo-ionizing flux from O stars (@kroupa02). However, the case of the young stellar disc in the GC is more tricky, as the disc is dominated by the gravitational field of the SMBH[^4] and of the stellar cusp. Furthermore, the gas in the disc has a density of $\sim{}10^{7-8}$ cm$^{-3}$, implying that the Strömgren sphere around a O star is only $\sim{}10^{-3}$ pc. Dynamical effects (e.g. @mathews69; @franco90) might increase the radius of the ionized region around early-type stars, but it is more likely that the gas disc can be completely evaporated only by the first supernovae in the stellar disc ($\gtrsim{}3$ Myr). Finally, a possible contribution from the SMBH or other ionizing sources in the GC is even less constrained. In summary, our assumption of instantaneous removal of the gas from the first disc, although necessary to avoid serious computational-time issues, might underestimate the gas evaporation timescale. This does not alter the dynamical evolution dramatically, as the gravitational influence of the SMBH on the stellar disc is larger by more than two orders of magnitude with respect to that of the gas disc. However, the first gas disc, if not completely removed, is expected to contribute to the secular evolution of the stellar disc. This effect is not included in our simulations. On the other hand, our simulations provide important information, even if the actual evaporation timescale is longer. In this case, our simulations can be considered as a robust lower limit to the effect of a second gas disc. In fact, if the first gas disc was not completely evaporated by the time the second gas disc forms, the effects of the latter may combine with to those of the former, and the resulting precession induced on the stellar disc is expected to be even stronger. In the extreme case of almost no gas evaporation in 2 Myr, the results of our simulations mimic the secular influence exerted on the stellar disc by its parent gas disc, rather than the perturbations induced by a second gas disc. This implies that even a long-lived ($>1$ Myr) parent gas disc can significantly affect the inclinations of the stellar orbits. Conclusions {#sec:concl} =========== In this paper, we study the effect of perturbations caused by a molecular cloud spiralling towards the GC on the orbits of a pre-existing stellar disc. It is likely that fresh gas spiralled to the GC in the last few Myrs, as @yusef-zadeh13 observe the traces of very recent star formation ($10^{4-5}$ yr) in the inner parsec. In our N-body/SPH simulations, the pre-existing stellar disc was formed by a first molecular cloud, disrupted by the GC (run E of M12). We simulate a second molecular cloud that falls towards the GC, is disrupted by the SMBH, and settles into a dense gas disc. The gas disc perturbs the orbits of the stellar disc. The details of this perturbation depend on the orbital parameters of the two discs and on the presence of a stellar cusp. The gas disc is neither razor-thin nor regular: it is a collection of different streamers with slightly different inclinations. The inclination of the angular momentum vector of the single streamers with respect to the angular momentum vector of the stellar disc ranges from $\sim{}0$ to $\sim{}20^\circ{}$. Thus, the dynamical interplay between the stellar disc and the gaseous disc(s) is highly complex. Our simulations show that, if the gas disc and the stellar disc are slightly misaligned ($\sim{}5-20^\circ{}$), the precession induced by the gas disc leads to a significant increase of the inclinations of the stellar orbits. The inclinations of the outer stellar orbits increase more than those of the inner ones, as the precession is faster for larger semi-major axes. This has crucial implications for the evolution of a stellar disc in the GC. In fact, the probability of early-type stars being members of the CW disc decreases with increasing projected distance from Sgr A$^\ast{}$ (@bartko09; @lu09). Furthermore, the CW disc does not seem a flat structure, but rather a significantly warped ($\sim{}64^\circ{}$, @bartko09) and tilted object. Our simulations suggest that the outer stellar orbits lose coherence with their original disc because of differential precession. This has two main effects: (i) the stellar disc appears warped/tilted; (ii) the outer parts of the stellar disc are progressively dismembered. This is consistent with previous predictions by @subr09, which proposed that the CNR is the source of differential precession. In our simulations, we assumed that the parent gas disc instantaneously evaporated, and that the naked stellar disc interacts with a second molecular cloud. It is likely that the stellar disc is still embedded in some residual of the original gas disc. Thus, our simulations provide a robust lower limit to the effect of a second gas disc. In fact, if the first gas disc was not completely evaporated, the new disc ‘superimposes’ to the old one, and the resulting precession induced on the stellar disc is expected to be even stronger. Another effect expected to be important is Kozai resonance. In our simulations, Kozai resonance is negligible as the timescale for Kozai oscillations scales as $\propto{}T_{\rm K}\,{}\sin^{-2}{\beta{}}$. Kozai resonance becomes important only if $\beta>>0$, but might be completely suppressed by the stellar cusp (e.g. @subr09). The case with $\beta\sim{}90^\circ{}$ deserves to be investigated with a more suitable integration scheme. Finally, we do not observe fragmentation in the simulations presented in this paper, but the mass resolution of gas particles is a factor of 10 lower than in M12. Thus, the possibility that the second gas disc fragments into self-bound clumps, forming stars, will be addressed in a further study, with higher resolution simulations. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the referee, Ladislav Šubr, for his valuable comments that significantly improved our paper. We also thank H. Perets, E. Ripamonti, L. Mayer, R. Alexander, W. Dehnen, A. King and S. Nayakshin for useful discussions. We thank the authors of <span style="font-variant:small-caps;">gasoline</span> (especially J. Wadsley, T. Quinn and J. Stadel). We acknowledge the CINECA Award N. HP10CL51UF, 2012 for the availability of high performance computing resources and support. MM acknowledges financial support from the Italian Ministry of Education, University and Research (MIUR) through grant FIRB 2012 (‘New perspectives on the violent Universe: unveiling the physics of compact objects with joint observations of gravitational waves and electromagnetic radiation’), and from INAF through grant PRIN-2011-1 (‘Challenging Ultraluminous X-ray sources: chasing their black holes and formation pathways’). [99]{} Agertz O., et al., 2007, MNRAS, 380, 963 Alexander R. D., Armitage P. J., Cuadra J., Begelman M. C., 2008, ApJ, 674, 927 Alexander R. D., Smedley S. L., Nayakshin S., King A. R., 2012, MNRAS, 419, 1970 Alig C., Burkert A., Johansson P. H., Schartmann M., 2011, MNRAS, 412, 469 Alig C., Schartmann M., Burkert A., Dolag K., 2013, ApJ, 771, 119 Bartko H. et al. 2009, ApJ, 697, 1741 Bartko H. et al. 2010, ApJ, 708, 834 Boley A. C., 2009, ApJ, 695, L53 Boley A. C., Hayfield T., Mayer L., Durisen R. H., 2010, Icarus, 207, 509 Bonnell I. A., Rice W. K. M., 2008, Science, 321, 1060 Christopher M. H., Scoville N. Z., Stolovy S. R., Yun M. S., 2005, ApJ, 622, 346 Collin S., Zahn J.-P., 2008, A&A, 477, 419 D’Alessio P., Calvet N., Hartmann L., 2001, ApJ, 553, 321 Dehnen W., Read J. I., 2011, The European Physical Journal Plus, 126, 55 Do T., Lu J. R., Ghez A. M., Morris M. R., Yelda S., Martinez G. D., Wright S. A., Matthews K., 2013, ApJ, 764, 154 Eisenhauer F. et al., 2005, ApJ, 628, 246 Franco J., Tenorio-Tagle G., Bodenheimer P., 1990, ApJ, 349, 126 Genzel R., Crawford M. K., Townes C. H., Watson D. M., 1985, ApJ, 297, 766 Genzel R., et al., 2003, ApJ, 594, 812 Ghez A. M., et al., 2003, ApJ, 586, L127 Ghez A. M., et al., 2005, ApJ, 635, 1087 Gillessen S., Eisenhauer F., Trippe S., Alexander T., Genzel R., Martins F., Ott T., 2009, ApJ, 692, 1075 Goodman J., 2003, MNRAS, 339, 937 Goodman J., Tan J. C., 2004, ApJ, 608, 108 Gualandris A., Merritt D., 2009, ApJ, 705, 361 Gualandris A., Gillessen S., Merritt D., 2010, MNRAS, 409, 1146 Gualandris A., Mapelli M., Perets H. B., 2012, MNRAS, 427, 1793 (G12) Güsten R., Genzel R., Wright M. C. H., Jaffe D. T., Stutzki J., Harris A. I., 1987, ApJ, 318, 124 Haas J., Šubr L., Kroupa P., 2011a, MNRAS, 412, 1905 Haas J., Šubr L.,Vokrouhliký D., 2011b, MNRAS, 416, 1023 Harfst S., Gualandris A., Merritt D., Spurzem R., Portegies Zwart S., Berczik P., 2007, NewAstronomy, 12, 357 Hayfield T., Mayer L., Wadsley J., Boley A. C., 2011, MNRAS, 417, 1839 Hobbs A., Nayakshin S., 2009, MNRAS, 394, 191 Ivanov P. B., Polnarev A. G., Saha P., 2005, MNRAS, 358, 1361 Jiang Y.-F., Goodman J., 2011, ApJ, 730, 45 Kozai Y., 1962, AJ, 67, 591 Krabbe A., Genzel R., Drapatz S., Rotaciuc V., 1991, ApJ, 382, L19 Krabbe A., et al., 1995, ApJ, 447, L95 Kroupa P., Boily C. M., 2002, MNRAS, 336, 1188 Larson R. B., 1981, MNRAS, 194, 809 Levin Y., Beloborodov A. M., 2003, ApJ, 590L, 33 Lidov M. L., 1962, Planetary and Space Science, 9, 719 Löckmann U., Baumgardt H., 2008, MNRAS, 384, 323 Löckmann U., Baumgardt H., Kroupa P., 2008, ApJ, 683, L151 Löckmann U., Baumgardt H., 2009, MNRAS, 394, 1841 Löckmann U., Baumgardt H., Kroupa P., 2009, MNRAS, 398, 429 Lu J. R., Ghez A. M., Hornstein S. D., Morris M., Matthews K., Thompson D. J., Becklin E. E., 2006, JPhCS, 54, 279 Lu J. R., Ghez A. M., Hornstein S. D., Morris M. R., Becklin E. E., Matthews K., 2009, ApJ, 690, 1463 Lu J. R., Do T., Ghez A. M., Morris M. R., Yelda S., Matthews K., 2013, ApJ, 764, 155 Lucas W. E., Bonnell I. A., Davies M. B., Rice K., 2013, MNRAS, 433, 353 Madigan A.-M., Levin Y., Hopman C., 2009, ApJ, 697, L44 Mapelli M., Hayfield T., Mayer L., Wadsley J., 2008, arXiv0805.0185 Mapelli M., Hayfield T., Mayer L., Wadsley J., 2012, ApJ, 749, 168 (M12) Mathews W. G., O’dell C. R., 1969, Annual Review of Astronomy and Astrophysics, 7, 67 Milosavljevic M., Loeb A., 2004, ApJ, 604L, 45 Morris M., 1993, ApJ, 408, 496 Namekata D., Habe A., 2011, ApJ, 731, 57 Nayakshin S., 2005, MNRAS, 359, 545 Nayakshin S., Cuadra J. J., 2005, A&A, 437, 437 Novak G., Dotson J. L., Dowell C. D., Hildebrand R. H., Renbarger T., Schleuning D. A., 2000, ApJ, 529, 241 Oka T., Nagai M., Kamegai K., Tanaka K., 2011, ApJ, 732, 120 Paumard T. et al., 2006, ApJ, 643, 1011 Perets H. B., Gualandris A., 2010, ApJ, 719, 220 Read J. I., Hayfield T., Agertz O.,2010, MNRAS, 405, 1513 Rice W. K. M., Lodato G., Armitage P. J., 2005, MNRAS, 364, L56 Schödel R., Ott T., Genzel R., Eckart A., Mouawad N., Alexander T., 2003, ApJ, 596, 1015 Schödel et al., 2007, A&A, 469, 125 Solomon P. M., Scoville N. Z., Penzias A. A., Wilson R. W., Jefferts K. B., 1972, ApJ, 178, 125 Šubr L., Schovancová J., Kroupa P., 2009, A&A, 496, 695 Šubr L., Haas J., 2012, Journal of Physics: Conference Series, 372, 012018 Ulubay-Siddiki A., Bartko H., Gerhard O., 2013, MNRAS, 428, 1986 Wadsley J. W., Stadel J., Quinn T., 2004, NewAstronomy, 9, 137 Wardle M., Yusef-Zadeh F., 2008, ApJ, 683L, 37 Yelda S., Ghez A. M., Lu J. R., Do T., Meyer L., Morris M. R., 2012, Adaptive Optics Systems III. Proceedings of the SPIE, Volume 8447, article id. 84470A Yusef-Zadeh F., Hewitt J. W., Cotton W., 2004, ApJS, 155, 421 Yusef-Zadeh F., Braatz J., Wardle M., Roberts D., 2008, ApJ, 683, L147 Yusef-Zadeh F., et al., 2013, ApJ, 767, L32 Zemp M., Stadel J., Moore B., Carollo C. M., 2007, MNRAS, 376, 273 Zhao J.-H., Morris M. R., Goss W. M., An T., 2009, ApJ, 699, 186 The coordinate system ===================== The coordinate system is defined in the following way. We define ${\bf J}_{\rm DISC}$ as the total angular momentum of the stellar disc. The normal vector to the total angular momentum of the stellar disc will then be $$\begin{aligned} {\bf n}_{\rm DISC}={\bf J}_{\rm DISC}/J_{\rm DISC}=\nonumber{}\\ =(\cos{\phi{}_{\rm DISC}}\,{}\sin{\theta{}_{\rm DISC}},\,{}\sin{\phi{}_{\rm DISC}}\,{}\sin{\theta{}_{\rm DISC}},\,{}\cos{\theta{}_{\rm DISC}}),\end{aligned}$$ where $J_{\rm DISC}$ is the modulus of ${\bf J}_{\rm DISC}$. We now define the coordinate system so that $\cos{\theta{}_{\rm DISC}}=1$ at any given time $t$. The normal vector to the angular momentum of a single star will then be $$\label{eq:A2} {\bf n}={\bf J}/J=(\cos{\phi{}}\,{}\sin{\theta{}},\,{}\sin{\phi{}}\,{}\sin{\theta{}},\,{}\cos{\theta{}}),$$ where ${\bf J}$ and $J$ are the angular momentum vector of a single star orbit and its modulus, respectively. By construction, $\theta{}=\theta{}_{\rm DISC}=0$ means that the angular momentum of a star is aligned to the total angular momentum of the disc. Thus, we refer to the value of $\theta{}$ as the inclination of the stellar orbit with respect to the disc axis at a given time $t$. The average value of $\theta{}$ ($\langle{}\theta{}\rangle{}$) gives also a measure of the half-opening angle of the disc. Fig. \[fig:figA1\] shows how the stellar orbits populate the plane defined by $\cos{\theta}$ and $\phi{}$ in the initial conditions and in runs A1 and B1. Fig. \[fig:figA2\] is the same as Fig. \[fig:figA1\], but for runs A2 and B2. [^1]: This assumption is quite conservative for the mass of stars, as it is realistic to expect further gas accretion in run E of M12. [^2]: For the definition of the coordinate system, see Appendix A. [^3]: As we discussed in the previous Section, the gas disc is not a rigid body, but a collection of concentric annuli with rather different inclination. On the other hand, we exclude that Kozai oscillations play a role, as we do not see any difference in the eccentricity evolution between run B1 (in which there is no stellar cusp) and run B2 (where Kozai oscillations should be damped by the stellar cusp). [^4]: The radiation pressure exerted by a star on a hydrogen atom overcomes the gravitational pull by the SMBH only when the distance of the hydrogen atom from the star is $d_\ast{}\le{}10^{-3}\,{}(L_\ast/10^{39}\textrm{ erg s}^{-1})^{1/2}\,{}(M_{\rm SMBH}/3.5\times{}10^{6}\textrm{ M}_\odot{})^{-1/2}\,{}\,{}d_{\rm BH}$, where $L_\ast$ is the luminosity of the star and $d_{\rm BH}$ is the distance of the hydrogen atom from the SMBH.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper we derive the Clauser-Horne (CH) inequality for the full electron counting statistics in a mesoscopic multiterminal conductor and we discuss its properties. We first consider the idealized situation in which a flux of entangled electrons is generated by an [entangler]{}. Given a certain average number of incoming entangled electrons, the CH inequality can be evaluated for different numbers of transmitted particles. Strong violations occur when the number of transmitted charges on the two terminals is the same ($Q_1=Q_2$), whereas no violation is found for $Q_1\ne Q_2$. We then consider two actual setups that can be realized experimentally. The first one consists of a three terminal normal beam splitter and the second one of a hybrid superconducting structure. Interestingly, we find that the CH inequality is violated for the three terminal normal device. The maximum violation scales as $1/M$ and $1/M^2$ for the entangler and normal beam splitter, respectively, $2M$ being the average number of injected electrons. As expected, we find full violation of the CH inequality in the case of the superconducting system.' author: - 'Lara Faoro$^{(1)}$, Fabio Taddei$^{(1,3)}$ and Rosario Fazio$^{(3)}$' title: 'Clauser-Horne inequality for electron counting statistics in multiterminal mesoscopic conductors' --- \#1[\|]{} \#1[\|[\#1]{}]{} Introduction ============ Entanglement [@bell87] denotes the nonlocal correlations that exist, even in the absence of direct interaction, between two (spatially separated) parts of a given quantum system. Since the early days of quantum mechanics, understanding the phenomenon of entanglement has been central to the understanding of the foundations of quantum theory. Besides its fundamental importance, a great deal of interest has been brought forth by its role in quantum information [@nielsen00]. Entanglement is believed to be the main ingredient of computational speed-up in quantum information protocols. Most of the work on entanglement has been performed in optical systems with photons [@zeilinger99], cavity QED systems [@rauschenbeutel00] and ion traps [@sackett00]. Only recently attention has been devoted to the manipulation of entangled states in a solid state environment. This interest, originally motivated by the idea to realize a solid state quantum computer [@makhlin01; @awschalom02], has been rapidly growing and by now several works discuss how to generate, manipulate and detect entangled states in solid state systems. It is probably worth to emphasize already at this point that, differently from the situation encountered in quantum optics, in solid state system entanglement is rather common. What is not trivial is its control and detection (especially if the interaction between the different subsystems forming the entangled state is switched off). Despite the large body of knowledge developed in the study of optical systems, new strategies have to be designed to reveal the signatures of non-local correlations in the case of electronic states. For mesoscopic conductors, the prototype scheme was discussed in Ref. [@burkard00]. In this work it has been shown that the presence of spatially separated pairs of entangled electrons, created by some [entangler]{}, can be revealed by using a beam splitter and by measuring the correlations of the current fluctuations in the leads. Provided that the electrons injected are in an entangled state bunching and anti-bunching behavior for the cross-correlations of current fluctuations are found depending on whether the state is a spin singlet or a spin triplet. Not only the noise, but the full counting statistics is sensitive to the presence of entanglement in the incoming beam [@taddei02]. The distribution of transmitted electrons is binomial and symmetric with respect to the average number of transmitted charges. Moreover, this is important for the problem studied in the present work, the joint probability for counting electrons at different leads unambiguously characterizes the state of the incident electrons if one uses spin-sensitive electron counters. In this case the joint probability cannot be expressed as a product of single-terminal probabilities. Given the general setup to detect entanglement an important issue is to understand how to generate it. This has been discussed in several papers. Most of the existing proposals are based on the generation of Bell states by means of electron-electron interaction. This can be achieved through superconducting correlations [@footnote] in hybrid normal - superconducting [@loss00; @loss01; @lesovik01; @samuelsson03] and superconductor - carbon nanotubes systems [@bena01; @bouchiat02], quantum dots in the Coulomb blockade regime [@saraga03] or Kondo-like impurities [@costa01]. Then, by using energy or spin filters, the two electrons forming the Bell state are separated. The entanglement can be created in the spin or in the orbital [@samuelsson03] degrees of freedom. Very recently, as it is also discussed in Section \[ns\], it was shown that in a mesoscopic multi-terminal conductor entanglement can be produced also in absence of electron interaction [@beenakker03]. Besides electrons, it is possible to produce entangled states with Cooper pairs in superconducting nanocircuits [@plastina01] or by coupling a mesoscopic Josephson junctions with superconducting resonators [@buisson00; @marquardt01; @you02; @plastina03]. Since Bell’s work [@bell64], it is known that a classical theory formulated in terms of a hidden variables satisfying reasonable condition of locality, yields predictions which are different from those of quantum mechanics. These predictions were casted into the form of inequalities which any realistic local theory must obey. Bell inequalities have been formulated for mesoscopic multi-terminal conductors in Refs. [@kawabata00; @chtchelkatchev02; @samuelsson03] in terms of electrical noise correlations at different terminals [@footnote2]. A test of quantum mechanics through Bell inequalities in mesoscopic physics is very challenging and most probably it would be rather difficult, if not impossible, to get around all possible loopholes. Although solid state systems are not the natural arena where to test the foundations of quantum mechanics, it is nevertheless very interesting to have access, manipulate and quantify these non-local correlations. In this work we derive a Bell inequality for the full electron counting statistics and discuss its properties. The formulation we follow is based on what is known as the Clauser-Horne (CH) inequality [@clauser74; @mandel95]. We shall show that the joint probabilities for a given number of electrons to pass through a mesoscopic conductor (in a given time) should satisfy, for a classical local theory, an inequality. The paper is organized as follows: in the next Section we motivate our approach to the problem, derive the CH inequality and express the joint probabilities needed in the CH inequality in terms of the scattering properties of the mesoscopic conductor. Section \[results\] is devoted to the discussion of the results. We first consider the idealized situation where an incoming flux of fully entangled electrons is injected into the mesoscopic region. Then we move on to analyze actual setups. Interacting electrons are not necessary to have an entangled state, we show that a three terminal normal device is enough to lead to violation of the CH inequality. For completeness we also consider the case where entanglement is produced by Andreev reflection. In the last Section we present the conclusions and a brief summary of this work. CH inequality for the full counting statistics ============================================== Electron Full Counting Statistics (FCS) refers to the probability that a given number of electrons has traversed, in a time $t$, a mesoscopic conductor. In the long time limit the first and the second moment of the probability distribution are related to the average current and noise, respectively. The reason for which we resort to FCS for analyzing electronic entanglement in a solid state environment resides in the fact that electrons in a conductor are not necessarily sufficiently separated from one another for coincidence counting to make sense, like in optical systems. Furthermore, the measurement of single coincidence events in electronic solid state systems does not seem realizable at present. Zero-frequency noise accounts for long time correlations and we do not expect it to be in general sensitive to coincidence measurements (see however the discussion in Ref. [@samuelsson03] for the limit of small transmission rates). From these premises we suggest that FCS is a natural candidate to formulate a Bell-type inequality for electrons in mesoscopic conductors. In the case where only two entangled electrons are injected, we find a situation similar to that with photons. More generally we discuss the case where a large number of electrons have been injected. In its original version [@bell64], the Bell inequality was derived for dicotomic variables. Here we consider the more general formulation due to Clauser and Horne [@clauser74]. We consider the idealized setup, illustrated in Fig.\[Ent\], which consists of the following parts. On the left we place an entangler that produces $2M$ electrons in a spin entangled state (in Section \[results\] two different situations for the implementation of the entangler are discussed). Two conductors, characterized by some scattering matrix, connect the terminals 3 and 4 of the entangler with the exit leads 1 and 2 so to carry the two particles belonging to each pair into two different spatially separated reservoirs. The electron counting is performed in leads 1 and 2 for electrons with spin aligned along the local spin-quantization axis at angles $\theta_1$ and $\theta_2$. Detection is realized by means of spin-selective counters, [i.e.]{} by counting electrons with the projection of the spin along a given local quantization-axis. In analogy with the optical case we say that the analyzer is not present when the electron counting is spin-insensitive (electrons are counted irrespective of their spin direction). In Section \[derivation\] we present the derivation of the CH inequality for the FCS and in Section \[scatt\] we resume, for completeness, the relation between FCS and the scattering matrix $S$. Derivation of the CH inequality {#derivation} ------------------------------- The basic object for the formulation of the CH inequality is the joint probability $P(Q_1,Q_2)$ for transferring a number of $Q_1$ and $Q_2$ electronic charges into leads $1$ and $2$ over an observation time $t$. We follow closely the derivation given in Ref. [@mandel95]. Our starting point is the following algebraic inequality $$-1 \le xy-xy'+x'y+x'y'-x'-y \le 0 \; \label{Clauser}$$ which holds for any variable $0 \le x,y,x',y' \le 1$. Let us now introduce explicitly a set of hidden variables $\tau$ which take values in a space ${\cal T}$. We assume that the incoming entangled electron states are described by $\tau$ in all the details necessary to determine the probability distributions $P(Q_{\alpha},\tau)$ for transferring a number of $Q_{\alpha}$ electronic charges into lead $\alpha=1,2$. By imposing that the hidden variable theory is local, it follows that the joint probability can be expressed in the following form: $$P(Q_1,Q_2)= \int_{\cal T} {\cal M}(\tau) P(Q_1,\tau) P(Q_2,\tau) d \tau , \label{jpd1}$$ where ${\cal M}(\tau)d\tau$ defines a probability measure on the space ${\cal T}$. The physical meaning of Eq.(\[jpd1\]) is straightforward: it states that the probability distribution on lead ${\alpha}$ does not depend on the probability distribution on the lead ${\beta}$. We now introduce $P^{\theta_1,\theta_2}(Q_1,Q_2)$ as the joint probability for transferring $Q_1$ and $Q_2$ electronic charges when both analyzers are present, while $P^{\theta_1,-}(Q_1,Q_2)$ and $P^{-,\theta_2}(Q_1,Q_2)$ are the corresponding joint probabilities when one of the two analyzers is removed. If the condition $$P^{\theta_\alpha}(Q_\alpha,\tau) \le P(Q_\alpha,\tau) \; \label{noen}$$ (known as [no-enhancement assumption]{} ) is verified, it is possible to identify the variables appearing in Eq.(\[Clauser\]) as follows: $$\begin{aligned} x&=& \frac{P^{\theta_1}(Q_1, \tau)}{P(Q_1,\tau)} ~~~ y=\frac{P^{\theta_2} (Q_2,\tau)}{P(Q_2, \tau)} \; ,\nonumber\\ x'&=& \frac{P^{\theta'_1}(Q_1,\tau)}{P(Q_1,\tau)} ~~~ y'=\frac{P^{\theta'_2} (Q_2, \tau)}{P(Q_2,\tau)} \; ,\label{variabili}\end{aligned}$$ $P^{\theta}(Q_\alpha,\tau)$ being the single terminal probability distribution in the presence of a analyzer. Eq.(\[Clauser\]) can then be rewritten in terms of probabilities by multiplying each side of the equation by $P(Q_1,\tau)P(Q_2,\tau) {\cal M}(\tau)d\tau$ and integrating over the space ${\cal T}$. Finally the following inequality is obtained $$\begin{aligned} {\cal S}_{CH}&=&P^{\theta_1,\theta_2}(Q_1,Q_2)-P^{\theta_1,\theta'_2}(Q_1,Q_2)+ P^{\theta'_1,\theta_2}(Q_1,Q_2)+P^{\theta'_1,\theta'_2}(Q_1,Q_2)\;\nonumber\\ &-&P^{\theta'_1,-}(Q_1,Q_2)-P^{-,\theta_2}(Q_1,Q_2) \le 0 \; . \label{CHin}\end{aligned}$$ Eq.(\[CHin\]) is the CH inequality for the full counting statistics [^1], holding for all values of $Q_1$ and $Q_2$ which satisfy the no-enhancement assumption. We stress that the no-enhancement assumption, upon which Eq.(\[CHin\]) is based, it is not satisfied in general like its optical version. The quantities that we have to compare are probability distributions, so that Eq.(\[noen\]) must be checked over the whole range of $Q$. For a fixed time $t$ and a given mesoscopic system, hence for a given scattering matrix and incident particle state, the no-enhancement assumption is valid only in some range of values of $Q$. In particular, different sets of system parameters correspond to different such ranges. The quantity ${\cal S}_{CH}$ in Eq.(\[CHin\]) depends on $Q_1$ and $Q_2$ so that the possible violation, or the extent of it, also depends on $Q_1$ and $Q_2$. Given a certain average number $M$ of entangled pairs that have being injected in the time $t$, one can look for the maximum violation as a function of the transmitted charges $Q_1$ and $Q_2$. Scattering approach to the full counting statistics {#scatt} --------------------------------------------------- The joint probabilities appearing in Eq.(\[CHin\]) can be determined once the scattering matrix $S$ of the mesoscopic conductor is known. The FCS in electronic systems was first introduced by Levitov [et al.]{} in Ref. [@levitov93; @levitov96] in the context of the scattering theory and later on the Keldysh Green function method [@nazarov99] to FCS was developed in Refs.[@nazarov01] (for a review see Refs. [@FCS]). In this paragraph we briefly describe how the FCS is formulated for a mesoscopic conductor in the scattering approach. Within this framework, the transport properties of a metallic phase-coherent structure attached to $n$ reservoirs are determined by the matrix $S$ of scattering amplitudes [@buttiker92]. Such amplitudes are defined through the scattering states describing particles propagating through the leads. For one dimensional conductors, for example, the scattering state arising from a unitary flux of particles at energy $E$ originating in the $i$-th reservoir reads $$\varphi_i(x)=\frac{e^{ik_i(E)x}+r_i(E) e^{-ik_i(E)x}}{\sqrt{hv_i(E)}} ~,$$ for the $i$-th lead, and $$\varphi_j(x)=\frac{t_{ji}(E) e^{-ik_j(E)x}}{\sqrt{hv_j(E)}} ~,$$ for the $j$-th lead, with $j\ne i$. Here $r_i(E)$ is the reflection amplitude for particles at energy $E$, wave vector $k_i(E)$ and group velocity $v_i(E)$ and $t_{ji}(E)$ is the transmission amplitude from lead $i$ to lead $j$. Note that $\|r_i\|^2$ is the probability for a particle to reflect back into the $i$-th lead and $\|t_{ji}\|^2$ is the probability for the transmission of a particle from lead $i$ to lead $j$. In the second quantization formalism, the field operator $\hat{\psi}_{j\sigma}(x,t)$ for spin $\sigma$ particles in lead $j$ is built from scattering states and it is defined as [@lesovik] $$\hat{\psi}_{j\sigma}(x,t)= \int dE ~ \frac{e^{-\frac{iEt}{\hbar}}} {\sqrt{hv_j(E)}} \left[ \hat{a}_{j\sigma}(E)e^{ik_j x}+\hat{\phi}_{j\sigma}(E)e^{-ik_j x} \right] ~,$$ where $\hat{a}_{j\sigma}(E)$ ($\hat{\phi}_{j\sigma}(E)$) is the destruction operator for incoming (outgoing) particles at energy $E$ with spin $\sigma$ in lead $j$. These operators are linked by the equation $$\left( \begin{array}{c} \hat{\phi}_{1\uparrow}\\ \hat{\phi}_{1\downarrow} \\ \hat{\phi}_{2\uparrow} \\ \vdots \end{array} \right)=~S~ \left( \begin{array}{c} \hat{a}_{1\uparrow}\\ \hat{a}_{1\downarrow} \\ \hat{a}_{2\uparrow} \\ \vdots \end{array} \right) \label{Smat}$$ and obey anti-commutation relations $$\left\{ \hat{a}^{\dagger}_{i\sigma}(E),~\hat{a}_{j\sigma'}(E')\right\} =\delta_{i,j} \delta_{\sigma,\sigma'} \delta(E-E') \,\,\, .$$ In the case of two and three dimensional leads one can separate longitudinal and transverse particle motion. Since the transverse motion is quantized, the wave function relative to the plane perpendicular to the direction of transport is characterized by a set of quantum numbers which identifies the channels of the lead. Such channels are referred to as open when the corresponding longitudinal wave vectors are real, since they correspond to propagating modes. Note that the case of a single open channel corresponds to a one dimensional lead. Let us now turn the attention to the probability distribution for the transfer of charges. Following Ref. [@muzykanskii94], within the scattering approach the characteristic function of the probability distribution for the transfer of particles in a structure attached to $n$ leads at a given energy $E$ can be written as $$\chi_E(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})= \langle \prod_{j=1,n} e^{i\lambda_{j\uparrow} \hat{N}_I^{j\uparrow}}~ e^{\lambda_{j\downarrow} \hat{N}_I^{j\downarrow}} ~ \prod_{j=1,n} e^{-i\lambda_{j\uparrow} \hat{N}_O^{j\uparrow}} e^{{-i\lambda_{j\downarrow} \hat{N}_O^{j\downarrow}}} \rangle ~, \label{chiE}$$ where the brackets $\langle ... \rangle$ stand for the quantum statistical average over the thermal distributions in the leads. Assuming a single channel per lead, $\hat{N}_{I(O)}^{j\sigma}$ is the number operator for incoming (outgoing) particles with spin $\sigma$ in lead $j$ and $\vec{\lambda_{\uparrow}}$, $\vec{\lambda_{\downarrow}}$ are vectors of $n$ real numbers, one for each open channel. In terms of incoming (outgoing) creation operator the number operators can be expressed as follows $$\hat{N}_I^{j\sigma}= \hat{a}_{j\sigma}^{\dagger}\hat{a}_{j\sigma}; \qquad \hat{N}_O^{j\sigma}=\hat{\phi}_{j\sigma}^{\dagger}\hat{\phi}_{j\sigma} \;\; .$$ Eq.(\[chiE\]) can also be recasted in the form [@levitov93]: $$\chi_E(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})= \det (\mathbb{I}-n_E+n_E~ S^\dagger \Lambda^\dagger~ S~ \Lambda) , \label{chiEdet}$$ where $\mathbb{I}$ is the unit matrix, $n_E$ is the diagonal matrix of Fermi distribution functions $f_j(E)$ for particles in the reservoir $j$ and defined as $(n_E)_{j\sigma,j\sigma}=f_j(E)$, whereas $\Lambda$ is a diagonal matrix defined as: $(\Lambda)_{j\sigma,j\sigma}=\exp (i\lambda_{j\sigma})$. For long measurement times $t$ the total characteristic function $\chi$ is the product of contributions from different energies, so that $$\chi(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})=e^{\frac{t}{h}\int dE~\log{\chi_E(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})}} \; . \label{chi}$$ At zero temperature, the statistical average over the Fermi distribution function in Eq.(\[chiE\]) simplifies to the expectation value calculated on the state ${\left\|\psi\right\rangle}$ containing two electrons of both spin species for each channel of a given lead up to the energy corresponding to the chemical potential of such lead. Furthermore, in the limit of a small bias voltage $V$ applied between the reservoirs, the argument of the integral is energy-independent so that Eq.(\[chi\]) can be approximated to $$\chi(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})\simeq \left[ \chi_0(\vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}})\right]^M \label{approx}$$ where only the zero-energy characteristic function appears and $M=eVt/h$ is the average number of injected particles. The joint probability distribution for transferring $Q_{1\sigma}$ spin-$\sigma$ electrons in lead 1, $Q_{2\sigma}$ spin-$\sigma$ electrons in lead 2, etc. is related to the characteristic function by the relation (we assume that no polarizers are present): $$P(Q_{1\uparrow},Q_{1\downarrow},Q_{2\uparrow},\ldots)=\frac{1}{(2\pi)^{2n}} \int_{-\pi}^{+\pi} d\lambda_{1\uparrow} d\lambda_{1\downarrow} d\lambda_{2\uparrow} \ldots ~\chi( \vec{\lambda_{\uparrow}},\vec{\lambda_{\downarrow}} ) ~e^{i\vec{\lambda_{\uparrow}}\cdot\vec{Q_{\uparrow}}} ~e^{i\vec{\lambda_{\downarrow}}\cdot\vec{Q_{\downarrow}}} ~. \label{counting}$$ In the rest of the paper we will consider systems where only two counting terminals are present. In particular, while the counting terminals are kept at the lowest chemical potential, all other terminals are biased at chemical potential $eV$. For later convenience, we write down the most general expression for the characteristic function when spin-$\sigma$ electrons are counted in lead 1 and spin-$\sigma'$ electrons are counted in lead 2: $$\begin{aligned} \chi_E (\lambda_{1\sigma},\lambda_{2\sigma'})=1+ \left( e^{-i\lambda_{1\sigma}} -1 \right) \langle \hat{N}_{\text{O}}^{1\sigma} \rangle + \left( e^{-i\lambda_{2\sigma'}} -1 \right) \langle \hat{N}_{\text{O}}^{2\sigma'} \rangle + \nonumber\\ + \left( e^{-i\lambda_{1\sigma}} -1 \right) \left( e^{-i\lambda_{2\sigma'}} -1 \right) \langle \hat{N}_{\text{O}}^{1\sigma} \hat{N}_{\text{O}}^{2\sigma'} \rangle , \label{chi_s}\end{aligned}$$ in the relevant energy range $0<E<eV$. The parameters $\lambda$ corresponding to all others terminals are set to zero. Using Eqs. (\[approx\]), (\[counting\]) and (\[chi\_s\]), at zero temperature, one can calculate the single terminal probability distribution: $$P(Q_{1\sigma})=\left( \begin{array}{c}M\\Q_{1\sigma} \end{array}\right) \left[ 1- {\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} {\left\|\psi\right\rangle}\right]^{M-Q_{1\sigma}} {\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} {\left\|\psi\right\rangle}^{Q_{1\sigma}} \label{stp}$$ and the joint probability distribution: $$\begin{aligned} \nonumber P (Q_{1\sigma},Q_{2\sigma'})=\sum_{k=\textrm{Max}[M-Q_{1\sigma},M-Q_{2\sigma'}]}^{(M-Q_{1\sigma})+(M-Q_{2\sigma'})} A^{2M-Q_{1\sigma}-Q_{2\sigma'}-k} ~B^{Q_{1\sigma}-M+k} ~C^{k-M+Q_{2\sigma'}}\times \\ \times {\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} \hat{N}_{\text{O}}^{2\sigma'} {\left\|\psi\right\rangle}^{M-k} f(M,Q_{1\sigma},Q_{2\sigma'},k) \label{FCSt}\end{aligned}$$ where $ A=1- {\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\sigma} {\left\|\psi\right\rangle} -{\left\langle\psi\right\|} \hat{N}_{\text{O}}^{2\sigma'} {\left\|\psi\right\rangle} + {\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} \hat{N}_{\text{O}}^{2\sigma'} {\left\|\psi\right\rangle} $, $ B={\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} (1- \hat{N}_{\text{O}}^{2\sigma'}){\left\|\psi\right\rangle} $, $ C={\left\langle\psi\right\|} (1- \hat{N}_{\text{O}}^{1\sigma})\hat{N}_{\text{O}}^{2\sigma'} {\left\|\psi\right\rangle} $ and $ f(M,Q_{1\sigma},Q_{2\sigma'},k) =M!/[(k-M+Q_{2\sigma'})!(2M-k-Q_{1\sigma}-Q_{2\sigma'})!] $. In doing so we have written the expressions for the probability distributions in terms of the expectation values of “outgoing” number operators. For $Q_{1\sigma}=Q_{2\sigma'}=M$, Eq.(\[FCSt\]) reduces to $$P(Q_{1\sigma}=M,Q_{2\sigma'}=M)={\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\sigma} \hat{N}_{\text{O}}^{2\sigma'} {\left\|\psi\right\rangle}^M .$$ When both spin species are counted in one of the terminals the characteristic function is different from the one given in Eq.(\[chi\_s\]). In particular, the characteristic function for counting both spins in terminal 1 reads: $$\begin{aligned} \nonumber \chi_E (\lambda_1,\lambda_{2\sigma'})=1+ \left( e^{-i\lambda_1} -1 \right) \langle \left( \hat{N}_{\text{O}}^{1\uparrow} + \hat{N}_{\text{O}}^{1\downarrow}\right) \rangle + \left( e^{-i\lambda_{2\sigma'}} -1 \right) \langle \hat{N}_{\text{O}}^{2\sigma'} \rangle + \\\nonumber+ \left( e^{-i\lambda_1} -1 \right) \left( e^{-i\lambda_{2\sigma'}} -1 \right) \langle \left( \hat{N}_{\text{O}}^{1\uparrow}+ \hat{N}_{\text{O}}^{1\downarrow}\right) \hat{N}_{\text{O}}^{2\sigma'} \rangle + \left( e^{-i\lambda_1} -1 \right)^2 \langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \rangle + \\ \left( e^{-i\lambda_1} -1 \right)^2 \left( e^{-i\lambda_{2\sigma'}} -1 \right) \langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \hat{N}_{\text{O}}^{2\sigma'} \rangle . \label{chi_s1}\end{aligned}$$ where we have set $\lambda_{1\uparrow}=\lambda_{1\downarrow}\equiv \lambda_1$. The expression for the joint probability distribution is in general complicated, as one can see in Appendix \[PR\] where such expressions for different systems are reported. Results ======= The inequality presented in Eq.(\[CHin\]) can be tested in various multi-terminal mesoscopic conductors. In this Section we present several geometries that can be experimentally realized. In order to get acquainted with the informations that can be retrieved from Eq.(\[CHin\]) we start from an ideal case in which the entangled pair is generated by some [entangler]{} in the same spirit as in the works of Refs. [@burkard00; @taddei02]. In Section \[ns\] we shall demonstrate that a normal beam splitter in the absence of interaction is enough to generate entangled pairs of electrons, therefore constituting a simple realization of an entangler. For comparison we also analyze the role of superconductivity in creating spin singlets. Entangled electrons {#ee} ------------------- In the setup depicted in Fig. \[Ent\] we assume the existence of an entangler that produces electron pairs in the Bell state $${\left\|\psi\right\rangle}=\frac{1}{\sqrt{2}} \left [a^{\dagger}_{3\uparrow}(E) a^{\dagger}_{4\downarrow}(E) \pm a^{\dagger}_{3\downarrow}(E) a^{\dagger}_{4\uparrow} (E)\right ] {\left\|0\right\rangle} , \label{st}$$ of spin triplet (upper sign) or spin singlet (lower sign) in the energy range $0<E<eV$. These electrons propagate through the conductors which connect terminals 3 and 4 with leads 1 and 2, as though terminals 3 and 4 were kept at a potential $eV$ with respect to 1 and 2. Our aim is to test the violation of the CH inequality given in Eq.(\[CHin\]) for such maximally entangled states. When the angles $\theta_1$ and $\theta_2$ are parallel to each other, the scattering matrix of the two conductors, in the absence of spin mixing processes, can be written as: $$S= \left( \begin{array}{cc} \hat{S}_{13} & 0 \\ 0 & \hat{S}_{24} \end{array} \right)\; \label{S-ent}$$ where $$\hat{S}_{13}=\left (\begin{array}{cc} \check{r}_{3} & \check{t}_{31} \\ \check{t}_{13} & \check{r}_{1} \end{array} \right )= \left (\begin{array}{cccc} r_{3 \uparrow} & 0 & t_{31 \uparrow} & 0 \\ 0 & r_{3 \downarrow} & 0 & t_{31 \downarrow} \\ t_{13 \uparrow} & 0 & r_{1 \uparrow} & 0 \\ 0 & t_{13 \downarrow} & 0 & r_{1 \downarrow} \end{array} \right ) .$$ Here $r_{j\sigma}$ ($t_{ij \sigma}$) is the probability amplitude for an incoming particle with spin $\sigma$ from lead $j$ to be reflected (transmitted in lead $i$). For a normal-metallic wire we set $t_{ij \uparrow}=t_{ij \downarrow}=\sqrt{T}$, $t_{ji \uparrow}=t_{ji \downarrow}=-\sqrt{T}$ and $r_{j\uparrow}=r_{j\downarrow}=\sqrt{1-T}$, where $T$ is the transmission probability. The expression for $\hat{S}_{24}$ is written analogously. For simplicity we will assume that $\hat{S}_{13}$ and $\hat{S}_{24}$ are equal. The general scattering matrix relative to non-collinear angles is obtained from $S$ by rotating the spin quantization axis independently in the two conductors (note that this is possible because the two wires are decoupled). The “rotated” S-matrix is obtained [@brataas01] by the transformation $S_{\theta_1,\theta_2}={\cal U} S {\cal U}^{\dagger}$, where ${\cal U}$ is the rotation matrix given by: $${\cal U}=\left( \begin{array}{cccc} U_{\theta_1} & 0 &0 & 0\\ 0&\mathbb{I}&0 & 0\\ 0 & 0 & U_{\theta_2} & 0 \\ 0&0&0& \mathbb{I}\end{array} \right) \label{tra}$$ where $$U_{\theta}=\left( \begin{array}{cc} \cos{\theta\over 2}&\sin{\theta\over 2}\\ -\sin{\theta\over 2}&\cos{\theta\over 2} \end{array} \right) . \label{uteta}$$ The probability distributions are now given by the expressions in Eq. (\[stp\]) and Eq. (\[FCSt\]) where the state ${\left\|\psi\right\rangle}$ is given by Eq. (\[st\]). In the case where both analyzers are present we set $\sigma=\sigma'=\uparrow$. The probability distribution when one of the analyzers is removed also possesses the structure of Eq.(\[FCSt\]) since, in this case, the correlators $\langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \rangle$ and $\langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \hat{N}_{\text{O}}^{2\uparrow} \rangle$ appearing in Eq.(\[chi\_s1\]) vanish. In particular when, for example, the upper analyzer in Fig. \[Ent\] is removed we need to replace $\hat{N}_{\text{O}}^{1\sigma}$ with $\hat{N}_{\text{O}}^{1\uparrow}+\hat{N}_{\text{O}}^{1\uparrow}$ and $\hat{N}_{\text{O}}^{2\sigma'}$ with $\hat{N}_{\text{O}}^{2\uparrow}$. For the other correlators one gets: $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\uparrow}{\left\|\psi\right\rangle}= {\left\langle\psi\right\|}\hat{N}_{\text{O}}^{2\uparrow}{\left\|\psi\right\rangle}=\frac{T}{2} , \label{27}$$ $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\downarrow}{\left\|\psi\right\rangle}=\frac{T}{2} , \label{28}$$ $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\uparrow}\hat{N}_{\text{O}}^{2\uparrow} {\left\|\psi\right\rangle}= \frac{T^2}{2} \sin^2 \left( \frac{\theta_1\pm\theta_2}{2}\right)$$ and $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\downarrow}\hat{N}_{\text{O}}^{2\uparrow} {\left\|\psi\right\rangle}= \frac{T^2}{2} \cos^2 \left( \frac{\theta_1\pm\theta_2}{2}\right) .$$ For the single terminal probability distributions in leads $i=1,2$ we get, in the presence and in the absence of an analyzer, respectively, $$\begin{aligned} P^{\theta_i}(Q_i)&=&\left( \begin{array}{c}M\\Q_i \end{array}\right) \left (\frac{T}{2} \right )^{Q_i} \left (1-\frac{T}{2} \right )^{M-Q_i} \label{prob00}\\ P(Q_i)&=&\left( \begin{array}{c}M\\Q_i \end{array}\right) \left (T \right )^{Q_i} \left (1-T \right )^{M-Q_i} , \label{prob0}\end{aligned}$$ so that the no-enhancement assumption reads: $$\left (1-\frac{T}{2} \right )^{(M-Q_i)} \left (\frac{1}{2} \right )^{Q_i} \le (1-T)^{(M-Q_i)} ~~~~~~~i=1,2\; . \label{noen1}$$ Note that the probabilities in Eqs. (\[prob00\]) and (\[prob0\]) do not depend on the angles $\theta_1$ and $\theta_2$ because the expectation values in Eqs. (\[27\]) and (\[28\]) are invariant under spin rotation. As a consequence, the effect of the analyzer is equivalent to a reduction of the transmission probability $T$ by a factor of 2, resulting in a shift of the maximum of the distribution. From Eq.(\[noen1\]) it follows that, for a given number $M=eVt/h$ of entangled pairs generated by the entangler, the no enhancement assumption can be verified only for certain values of $T$ and of $Q_i$. This makes clear that the CH inequality of Eq.(\[CHin\]) can be tested for violation only for appropriate values of $M$, $T$ and $Q_1$ or $Q_2$. For example, for a given observation time $t$ ([i.e.]{} a given $M$) and a given value of $Q$, CH inequality can be tested only for transmission $T$ less than a maximum value given by the expression $$T_{\text{max}}=\frac{2^{\frac{Q_i}{M-Q_i}}-1}{2^{\frac{Q_i}{M-Q_i}}-\frac{1}{2}} . \label{condition}$$ At the edge of the distribution ($Q_i=M$) the no-enhancement assumption is satisfied for every $T$. The window of allowed $Q_i$ values where the no-enhancement assumption is satisfied gets wider on approaching the tunneling limit. For large $M$, $T_{\text{max}}\simeq 2 (\log2) \frac{Q_i}{M}$. The previous inequality can be also interpreted as a limit for the allowed measuring time given a setup at disposal. Alternatively, given a certain transmission, the no-enhancement assumption is verified for points of the distribution such that: $$\frac{Q_i}{M}\geq \frac{\log\frac{1-T/2}{1-T}}{\log 2+\log\frac{1-T/2}{1-T}} . \label{ultimo}$$ The various probabilities needed to define ${\cal S}_{CH}$ are collected in Appendix \[PR\]. However, it is useful to note here that the joint probabilities with a single analyzer are factorized: $$\begin{aligned} \nonumber P^{\theta_1,-}(Q_1,Q_2)=P^{\theta_1}(Q_1) P(Q_2)\\ P^{-,\theta_2}(Q_1,Q_2)=P(Q_1) P^{\theta_2}(Q_2) ,\end{aligned}$$ while joint probabilities with two analyzers are not factorized. Furthermore, all such probabilities have a common factor, $T^{Q_1+Q_2}/2^M$, which leads to an exponential suppression for large $M$ and $Q_1+Q_2$. We shall address the question of whether this also produces a suppression of ${\cal S}_{CH}$ in case of violation. Let us now analyze the possibility of violation of the CH inequality for different values of $Q_1$ and $Q_2$. First consider the situation where the entangler emits a single entangled pair of electrons in which case $P^{\theta_1,\theta_2}(1,1)={\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{2\uparrow}{\left\|\psi\right\rangle}$, $P^{-,\theta_2}(1,1)={\left\langle\psi\right\|}( \hat{N}_{\text{O}}^{1\uparrow}+\hat{N}_{\text{O}}^{1\downarrow}) \hat{N}_{\text{O}}^{2\uparrow}{\left\|\psi\right\rangle}$ and $P^{\theta_1,-}(1,1)={\left\langle\psi\right\|} \hat{N}_{\text{O}}^{1\uparrow} (\hat{N}_{\text{O}}^{2\uparrow}+ \hat{N}_{\text{O}}^{2\uparrow}){\left\|\psi\right\rangle}$. We find that the CH inequality is maximally violated for the following choice of angles: $\theta_2-\theta_1=\theta'_2-\theta'_1=3\pi /4$. More precisely we obtain: $${\cal S}_{CH}= T^2 \frac{\sqrt{2}-1}{2}$$ which is equal to the result obtain for an entangled pair of photons [@mandel95], where $T$ plays the role of the quantum efficiency of the photon detectors. In the more general case of $Q_1=Q_2=M$, for $M\gg 1$, we have $$\begin{aligned} P^{\theta_1,\theta_2}(M,M)&=&\frac{T^{2M}}{2^M} \left [\sin^2 \left ( \frac{\theta_1 \pm \theta_2}{2} \right ) \right ]^M \; \nonumber \\ P^{\theta_1,-}(M,M)&=& P^{-,\theta_2}(M,M)=\frac{T^{2 M}}{2^M} \;\label{prob}\end{aligned}$$ so that the no-enhancement assumption is always satisfied and the quantity ${\cal S}_{CH}$ can be easily evaluated: $${\cal S}_{CH}=\frac{T^{2M}}{2^M} \left[ \sin^{2M}\frac{\theta_1\pm \theta_2}{2}- \sin^{2M}\frac{\theta_1\pm \theta'_2}{2} + \sin^{2M}\frac{\theta'_1\pm \theta_2}{2} + \sin^{2M}\frac{\theta'_1\pm \theta'_2}{2} -2 \right] .$$ The rotational invariance makes $P^{\theta_1,-}$ and $P^{-,\theta_2}$ independent of angles, and $P^{\theta_1,\theta_2}$ dependent on the angles through $\frac{\theta_1\pm \theta_2}{2}$. This allows us, without loss of generality, to define an angle $\Theta$ such that $2\Theta=\theta_1\pm \theta_2=\theta'_1\pm \theta_2= \theta'_1\pm \theta'_2=(\theta_1\pm \theta'_2)/3$. As a result Eq.(\[CHin\]) takes the form: $${\cal S}_{CH}=3P_{1,2}^{\Theta}(Q_1,Q_2)-P_{1,2}^{3\Theta}(Q_1,Q_2)-P_{1,-}(Q_1,Q_2)-P_{-,2}(Q_1,Q_2) \le 0$$ where $P_{1,2}^{\Theta}=P^{\theta_1,\theta_2}$ and $P_{1,-}=P^{\theta_1,-}$. It is useful to define the reduced quantity $\overline{{\cal S}}_{CH}={\cal S}_{CH}/(T^{2M}/2^M)$ which is plotted in Fig. \[Srid:Q=M\] as a function of $\Theta$ for different values of $M$ (note that since $P^{\theta_1,-}(M,M)=(T^{2M}/2^M)$, $\overline{{\cal S}}_{CH}$ is nothing but ${\cal S}_{CH}/P^{\theta_1,-}(M,M)$). The violation occurs for every value of $M$ in a range of angles around $\Theta=\pi/2$ (note that ${\cal S}_{CH}$ is symmetric with respect to $\pi/2$). The range of angles for which $\overline{{\cal S}}_{CH}$ is positive shrinks with increasing $M$, while the maximum value of $\overline{{\cal S}}_{CH}$ decreases very weakly with $M$ (more precisely, $\overline{{\cal S}}_{CH}^{\text{max}}\propto 1/M$). This means that the effect of the factor $T^{2M}/2^M$ on the value of ${\cal S}_{CH}$ is exponentially strong, making the violation of the CH inequality exponentially difficult to detect for large $M$ and $Q_1=Q_2=M$. The weakening of the violation is mainly due to the suppression of the joint probabilities. As we shall show later, by optimizing all the parameters it is yet possible to eliminate this exponential suppression. Let us now consider the violation of the CH inequality as a function of the transmitted charges. We notice that the CH inequality is not violated for the off-diagonal terms of the distributions (when $Q_1 \neq Q_2$), meaning that one really needs to look at “coincidences”. Therefore we discuss the case $Q_1=Q_2\equiv Q<M$ (remember that the no-enhancement assumption is satisfied only for $T\leq T_{\text{max}}(Q)$). In Fig. \[S:Q<M\] we plot the quantity ${\cal S}_{CH}$ for $M=20$ as a function of $\Theta$ and different values of $Q$. The transmission $T$ is fixed at the highest allowed value by the no-enhancement assumption, which corresponds to the smallest $Q$ considered $T_{\text{max}}(Q=1)=0.06917$. Fig. \[S:Q<M\] shows that the largest positive value of ${\cal S}_{CH}$ and the widest range of angles corresponding to positive ${\cal S}_{CH}$ occur for $Q=1$, [i.e.]{} for a joint probability relative to the detection of a single pair. One should not conclude that, in order to detect the violation of the CH inequality, only very small values of the transmitted charge should be taken. We have in fact considered $T=T_{\text{max}}$ relative to $Q=1$ and the maximum violation, for given $M$ and $Q$, always occurs at $T=T_{\text{max}}$. In order to get the largest violation of the CH inequality at a given $M$ and $Q$ one could, in principle, choose the highest allowed value of $T$ for each value of $Q$ ($T=T_{\text{max}}(Q)$). We show in Fig. \[S1:Q<M\] the corresponding plot, to be compared with Fig. \[S:Q<M\]. For every $Q<M$ the violation occurs in the same range of angles, namely $\pi/4 \le\Theta\le \pi/2$, because of the following properties of the joint probability distributions: $P^{\Theta}_{1,2}(Q_1,Q_2)=P^{3\Theta}_{1,2}(Q_1,Q_2)=P_{1,-}(Q_1,Q_2)$ for $\Theta=\pi/4$. This implies that ${\cal S}_{CH}(\Theta=\pi/4)=0$, and $P^{\Theta}_{1,2}(Q_1,Q_2)\geq P^{3\Theta}_{1,2}(Q_1,Q_2),P^{\Theta}_{1,-}(Q_1,Q_2), P^{\Theta}_{-,2}(Q_1,Q_2)$ for $\pi/4 \le\Theta\le \pi/2$. Furthermore, in this specific case of $M=20$, we find that the maximum values of $S$ occurs at $Q=8$. In Fig. \[Smax:Q<M\] we plot the maximum value of $S$, with respect to $\Theta$ and $T$, as a function of $Q$ for different values of $M$. Several observations are in order. For increasing $M$, the position of the maximum, $Q_{\text{max}}$ is very weakly dependent on $M$. Remarkably, the value of the maximum of the curves does not decreases exponentially, but rather as $1/M^2$. Despite the exponential suppression of the joint probability with $M$, the extent of the maximal violation scales with $M$ much slowly (polynomially). It may be useful to look at the same situation from a different perspective. Given a certain transmission $T$ ([i.e.]{} fixing the transport properties of the conductors) we want to find when the CH inequality is maximally violated. For a given observation time $t$, the no-enhancement assumption Eq. (\[condition\]) imposes a minimum value for $Q$. In Fig. \[Smax:T\] we plot the quantity ${\cal S}_{CH}$, maximized over the angle $\Theta$ and $Q$, as a function of $T$ for different $M$. The curves are piecewise increasing function of $T$, where the discontinuities correspond to an increase of the value of $Q$ by one imposed by the no-enhancement assumption. More precisely, when $T$ is increased above a threshold for which Eq. (\[condition\]) is not satisfied, one needs to increase $Q$ by one unit in order for this condition to be recovered. The result of this is a jump in the values of the probabilities that leads to a discontinuity of ${\cal S}_{CH}$. Fig. \[Smax:T\] allows to choose the best values of $M$ and $Q$ to get the maximum violation. If the entangler is substituted with a source that emits factorized states, the CH inequality given in Eq.(\[CHin\]) is never violated. In this case, in contrast to Eq.(\[st\]), the state emitted by the source reads: $ {\left\|\psi\right\rangle}=a^{\dagger}_{3\uparrow} a^{\dagger}_{4\uparrow} {\left\|0\right\rangle} . $ All the previous calculations can be repeated and we find, as expected, that the characteristic functions factorizes, so that the two terminal joint probability distributions are given by the product of the single terminal probability distributions. Normal beam splitter {#ns} -------------------- We are now ready to analyze realistic structures by replacing the shaded block in Fig. \[Ent\] (which represents the entangler) with a certain system, and discuss the CH inequality along the lines of Section \[ee\]. We first consider a normal beam splitter (shaded block in Fig. \[Stub\]) in which lead 3 is kept at a potential $eV$ and leads 1 and 2 are grounded so that the same bias voltage is established between 3 and 1, and 3 and 2. The two conductors, which connect the beam splitter to the leads 1 and 2, are assumed to be normal-metallic and perfectly transmissive, so that the S-matrix of the system for $\theta_1=\theta_2=0$ is equal to the S-matrix of the beam splitter, which reads [@buttiker84] $$S=\left( \begin{array}{ccc} -(a+b) & \sqrt{\epsilon} & \sqrt{\epsilon} \\ \sqrt{\epsilon} & a & b\\ \sqrt{\epsilon} & b & a \end{array} \right) . \label{S-stub}$$ In this parametrization of a symmetric beam splitter $a=\pm(1+\sqrt{1-2\epsilon})/2$, $b=\mp(1-\sqrt{1-2\epsilon})/2$ and $0<\epsilon <1/2$. For arbitrary angles $\theta_1$ and $\theta_2$, the S-matrix is obtained rotating the quantization axis in the two conductors independently by applying the transformation $S_{\theta_1,\theta_2}={\cal U} S {\cal U}^{\dagger}$, where ${\cal U}$ is the rotation matrix given by: $${\cal U}=\left( \begin{array}{ccc} \mathbb{I} & 0 &0\\0&U_{\theta_1}&0\\ 0&0&U_{\theta_2} \end{array} \right)$$ and $U_{\theta}$ is defined in Eq.(\[uteta\]). This procedure is valid as long as no back scattering is present in the conductors. The probability distributions are given by Eqs. (\[stp\]) and (\[FCSt\]) where the state ${\left\|\psi\right\rangle}$ is now factorisable: $${\left\|\psi\right\rangle}=a^{\dagger}_{1\uparrow}(E) a^{\dagger}_{1\downarrow}(E) {\left\|0\right\rangle} \label{initial}$$ in the energy range $0<E<eV$. Analogously to what was done in Section \[ee\], when both analyzers are present we set $\sigma=\sigma'=\uparrow$. When only one analyzer is present, however, one has to use the correct characteristic function of Eq.(\[chi\_s1\]), since one of the two additional correlators does not vanish. Namely, $\langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \rangle= \epsilon^2$ and $\langle \hat{N}_{\text{O}}^{1\uparrow} \hat{N}_{\text{O}}^{1\downarrow} \hat{N}_{\text{O}}^{2\uparrow} \rangle=0$, when the upper analyzer, for example, in Fig. \[Stub\] is removed. For the other expectation values we get: $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\uparrow}{\left\|\psi\right\rangle}= {\left\langle\psi\right\|}\hat{N}_{\text{O}}^{2\uparrow}{\left\|\psi\right\rangle}=\epsilon , \label{n1s}$$ $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\downarrow}{\left\|\psi\right\rangle}=\epsilon , \label{n2s}$$ $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\uparrow}\hat{N}_{\text{O}}^{2\uparrow} {\left\|\psi\right\rangle}= \epsilon^2 \sin^2 \left( \frac{\theta_1-\theta_2}{2}\right) \label{n3s}$$ and $${\left\langle\psi\right\|}\hat{N}_{\text{O}}^{1\downarrow}\hat{N}_{\text{O}}^{2\uparrow} {\left\|\psi\right\rangle}= \epsilon^2 \cos^2 \left( \frac{\theta_1-\theta_2}{2}\right) , \label{n4s}$$ obtaining the joint probability distributions reported in Appendix \[PR\]. The above number operator expectation values are equal to the case of the entangler when $\epsilon$ is replaced by $T/2$, whereas the cross-terminal correlators are equal in the two cases if $\epsilon$ is replaced with $T/\sqrt{2}$. From this follows that the characteristic functions for the beam splitter possess the same dependence on the angle difference as the corresponding characteristic functions for the entangler (Section \[ee\]) but have a different structure as far as scattering probabilities are concerned. In particular, as expected [@levitov93], the cross-correlations vanish when the two angles are equal. On the contrary, when the angle difference is $\pi$ cross-correlations are maximized. Furthermore, when only one analyzer is present the characteristic function shows no dependence on the angle, but it is not factorisable, in contrast to the case of the entangler. As a result, the single terminal probabilities, given by Eq.(\[stp\]), are equal in the two cases provided that $\epsilon$ is replaced with $T/2$. The joint probabilities for $Q_1=Q_2=M$ are equal in the two cases if $\epsilon$ is replaced with $T/\sqrt{2}$ (however, this replacement is not valid in general for joint probabilities with $Q_1,Q_2\ne M$): $$P^{\theta_1,\theta_2}(M,M)=\left[ \epsilon^2 \sin^2{\left(\theta_1 -\theta_2 \over 2\right)} \right]^M \label{qq1}$$ $$P^{\theta_1,-}(M,M)=\epsilon^{2M} \label{qq2}$$ The no-enhancement assumption is verified when $$\epsilon \leq \frac{1}{2} \frac{2^{Q \over M-Q}-1}{2^{Q \over M-Q}-\frac{1}{2}} ,$$ which equals the condition of Eq. (\[condition\]) once $\epsilon$ is replaced with $T/2$. Let us first consider the case for which $Q_1=Q_2=M$. We obtain an important result: the CH inequality is violated for the same set of angles found for the case of the entangler, although to a lesser extent, since the prefactors in Eqs. (\[qq1\]) and (\[qq2\]) now varies in the range $0\leq \epsilon ^{2M}\leq \frac{1}{4^M}$. In particular, in the simplest case of $M=1$, corresponding to injecting a single pair of electrons, the maximum violation corresponds to ${\cal S}_{CH}=\frac{\sqrt{2}-1}{4}$, which is a half of the value for the entangler. Furthermore, the plot in Fig. \[Srid:Q=M\] is also valid in the present case with $\overline{{\cal S}}_{CH}$ defined as $\overline{{\cal S}}_{CH}={\cal S}_{CH}/\epsilon^{2M}$, [i.e.]{} by replacing $T/\sqrt{2}$ with $\epsilon$. This means that a geometry like that of the beam splitter enables to detect violation of CH inequality without any need to resort to interaction processes to produce entanglement. Also here we consider the case for which $Q_1=Q_2\equiv Q<M$, where interesting differences with respect to the case of the entangler are found. i) We find that the violation of the CH inequality is in general weaker, meaning that the absolute maximum value of ${\cal S}_{CH}$ is smaller than in the ideal case of the entangler. ii) The weakening of the violation with increasing $M$ is determined by the suppression of the probability by the prefactor $(\epsilon^2)^{Q1+Q2}$. Remarkably, the maximum value of ${\cal S}_{\text{max}}$ decreases like $1/M$, therefore even slower than for the ideal case. iii) Violations occur only for values of $Q$ close to 1, even for large values of $M$: to search for violations one has to look at single- or few-pair probabilities and therefore, because of the no-enhancement assumption, to small transmissions $\epsilon$. iv) Interestingly, for $Q=1$ the quantity ${\cal S}_{CH}$ is positive for any angles, although the largest values correspond to $\Theta$ close to $\pi/2$ (see Fig. \[S:angle-stub\]). We do not find any relevant variation, with respect to the discussion in paragraph \[ee\], for probabilities relative to $Q_1\ne Q_2$. It is easy to convince oneself that the final state calculated from the initial one (\[initial\]) using the S-matrix (\[S-stub\]) contains an entangled part. In Ref. [@bose02] this fact was already noticed, but for an incident state composed by a single pair of particles impinging from the two entering arms of a beam splitter. For mesoscopic conductors, entanglement without interaction for electrons injected from a Fermi sea has been discussed by Beenakker [et al]{} [@beenakker03]. Superconducting beam splitter {#ss} ----------------------------- In many proposals superconductivity has been identified as a key ingredient for the creation of entangled pairs of electrons. The idea is to extract the two electrons which compose a Cooper pair (a pair of spin-entangled electrons) from two spatially separated terminals. Here we showed that it is not necessary to have superconducting correlations. Nevertheless, in view of the recent interest in entanglement created by pairing correlations, it is useful to analyze also the case of a superconducting beam splitter [@boerlin02; @samuelsson02] depicted in Fig. \[sup\_stub\], which consists of a superconducting lead (with condensate chemical potential equal to $\mu$) in contact with two normal wires. The wires are then connected to two leads attached to reservoirs kept at zero potential. This is basically what is obtained by replacing the entangler of Fig. \[Ent\] by a superconducting lead with two terminals. The system can be decomposed into two subsystems: on the left-hand-side of Fig. \[sup\_stub\] we place the superconducting slab attached two normal terminals (5 and 6) characterized by a reflection amplitudes matrix $R'_\text{s}$ defined, in terms of the particle operators, by: $$\hat{\phi}_{j\alpha\sigma}(E)=\sum_{k=5,6}\sum_{\beta=e,h}\sum_{\sigma'=\uparrow,\downarrow} \left[ R'_\text{s}(E)\right]_{j\alpha\sigma,k\beta\sigma'}~ \hat{a}_{k\beta\sigma'}(E) .$$ Here $j=5,6$ and the additional indexes $\alpha$ and $\beta$ refer to the particle-hole degree of freedom, in particular $\alpha=e$ for particles and $\alpha=h$ for holes and $[\ldots ]_{j\alpha\sigma,k\beta\sigma'}$ represents the specified element of the matrix. Note that $R'_\text{s}$ is block diagonal in spin indexes so that $$R'_\text{s}=\left( \begin{array}{cc} \mathcal{R}'&0\\0&\mathcal{R}' \end{array}\right)$$ with $$\left( \begin{array}{c} \hat{\phi}_{5e\uparrow}\\ \hat{\phi}_{5h\downarrow}\\ \hat{\phi}_{6e\uparrow}\\ \hat{\phi}_{6h\downarrow} \end{array}\right)=\mathcal{R}' \left( \begin{array}{c} \hat{a}_{5e\uparrow}\\ \hat{a}_{5h\downarrow}\\ \hat{a}_{6e\uparrow}\\ \hat{a}_{6h\downarrow}\\ \end{array}\right),\qquad \qquad \mathcal{R}'=\left( \begin{array}{cccc} \rho_{ee}&\rho_{ph}&\tau_{ee}&\tau_{eh}\\ \rho_{he}&\rho_{hh}&\tau_{he}&\tau_{hh}\\ \tau'_{ee}&\tau'_{eh}&\rho'_{pp}&\rho'_{eh}\\ \tau'_{he}&\tau'_{hh}&\rho'_{hp}&\rho'_{hh} \end{array}\right) ,$$ where $\rho_{ee}$ ($\rho_{hh}$) is the normal reflection amplitude for particles (holes) in terminal 5, $\rho_{eh}$ ($\rho_{he}$) is the Andreev reflection for a hole to evolve into a particle (particle to evolve into a hole) in terminal 5. $\tau_{ee}$ ($\tau_{hh}$) is the normal transmission amplitude for particles (holes) to be transmitted from terminal 5 to terminal 6, $\tau_{eh}$ ($\tau_{he}$) is the Andreev transmission amplitude for holes (particles) in terminal 5 to be transmitted in terminal 6 as particles (holes). Primed amplitudes refer to reflections occurring in lead 6 and transmissions from lead 6 to lead 5. On the right-hand-side of Fig. \[sup\_stub\] we have the subsystem composed of two identical decoupled conductors characterized by the $16\times 16$ scattering matrix $$S_{\text{c}}=\left( \begin{array}{cc} R_{\text{c}} & T'_{\text{c}} \\ T_{\text{c}} & R'_{\text{c}} \end{array}\right) . \label{essep}$$ The four submatrices in Eq.(\[essep\]) are block diagonal in spin space, for example $R_{\text{c}}$ can be written as: $$R_{\text{c}}=\left( \begin{array}{cc} R_{\text{c}}^{\uparrow}&0\\0&R_{\text{c}}^{\downarrow} , \end{array}\right)$$ where $R_{\text{c}}^{\uparrow}$ is a diagonal matrix defined by $$\left( \begin{array}{c} \hat{\phi}_{3e\uparrow}\\ \hat{\phi}_{3h\downarrow}\\ \hat{\phi}_{4e\uparrow}\\ \hat{\phi}_{4h\downarrow} \end{array}\right)=R_{\text{c}}^{\uparrow} \left( \begin{array}{c} \hat{a}_{3e\uparrow}\\ \hat{a}_{3h\downarrow}\\ \hat{a}_{4e\uparrow}\\ \hat{a}_{4h\downarrow}\\ \end{array}\right),\qquad \qquad R_{\text{c}}^{\uparrow}=\left( \begin{array}{cccc} r_{3e\uparrow}&0&0&0\\ 0&r_{3h\downarrow}&0&0\\ 0&0&r_{4e\uparrow}&0\\ 0&0&0&r_{4h\downarrow} \end{array}\right) .$$ $R_{\text{c}}^{\downarrow}$ is defined like $R_{\text{c}}^{\uparrow}$ exchanging $\uparrow$ with $\downarrow$, whereas $T_{\text{c}}^{\sigma}$ is defined similarly to $R_{\text{c}}^{\sigma}$ replacing $r_{3\alpha\sigma}$ with $t_{1\alpha\sigma}$ and $r_{4\alpha\sigma}$ with $t_{2\alpha\sigma}$. The matrices $R_{\text{c}}^{\prime\sigma}$ and $T_{\text{c}}^{\prime\sigma}$ are defined analogously using the amplitudes $r_{1\alpha\sigma}$, $r_{2\alpha\sigma}$, $t'_{1\alpha\sigma}$ and $t'_{2\alpha\sigma}$. The spin quantization axis of the two wires can be rotated independently as in paragraph \[ee\] by applying the transformation $S_{\theta_1,\theta_2}={\cal U} S_{\text{c}} {\cal U}^{\dagger}$, where ${\cal U}$ is defined in Eq. (\[tra\]), obtaining the scattering matrix $$S_{\theta_1,\theta_2}=\left( \begin{array}{cc} \tilde{R}_{\text{c}} & \tilde{T}'_{\text{c}} \\ \tilde{T}_{\text{c}} & \tilde{R}'_{\text{c}} \end{array}\right)~.$$ The overall matrix of reflection amplitudes is calculated by composing the scattering matrices relative to the two subsystems [@datta95]: $$R'_{\text{tot}}=\tilde{R}'_{\text{c}} +\tilde{T}_{\text{c}} \left[ \mathbb{I} - R'_{\text{s}} \tilde{R}_{\text{c}}\right]^{-1} R'_{\text{s}} \tilde{T}'_{\text{c}} .$$ where $R'_{\text{tot}}$ is defined by $$\hat{\phi}_{j\alpha\sigma}(E)=\sum_{k=1,2}\sum_{\beta=e,h}\sum_{\tau=\uparrow,\downarrow} \left[ R'_\text{tot}(E)\right] _{j\alpha\sigma,k\beta\tau}~ \hat{a}_{k\beta\tau}(E) ,$$ with $j$ running from 1 to 2. The characteristic function can now be calculated through Eq. (\[chiEdet\]) taking $R'_\text{tot}(E)$ as scattering matrix. In the present case, where superconductivity is present, the diagonal matrix of Fermi distribution functions is defined as $[n_E]_{j\alpha\sigma,j\alpha\sigma}=f_{j\alpha}(E)$, $f_{j\alpha}(E)=[1+\exp (\frac{E+\alpha\mu}{k_BT})]^{-1}$ and $[\Lambda]_{j\alpha\sigma,j\alpha\sigma}=\exp (i\alpha\lambda_{j\sigma})$, with $j=1,2$. By choosing $\lambda_{1\downarrow}=\lambda_{2\downarrow}=0$ we achieve the goal of counting excitations with spin-up component. The case where one of the analyzers is removed, for example in lead 1, is implemented by setting $\lambda_{1\downarrow}=\lambda_{1\uparrow} =\lambda_1$ and $\theta_1=0$, [i.e.]{} by counting electrons in lead 1 regardless their spin. In the limit of zero temperature and small bias voltage, we only need the scattering amplitudes at the zero energy (Fermi level) so that the overall characteristic function can be approximated like in Eq.(\[approx\]). We parametrize the matrix $S_{\text{c}}$ of the wires as follows: $r_{3e\sigma}=r_{4e\sigma}=\sqrt{1-T}$, $r_{1e\sigma}=r_{2e\sigma}=\sqrt{1-T}$, $t_{1e\sigma}=t_{1e\sigma}=\sqrt{T}$ and $t'_{1e\sigma}=t'_{2e\sigma}=-\sqrt{T}$, where $T$ is the wire transmission probability of the wires. The amplitudes relative to hole degree of freedom are determined from the ones above by making use of the particle-hole symmetry. Although Andreev processes are fundamental for the injection of Cooper pairs, in the case where Andreev transmissions only are non-zero and $T=1$ the joint probabilities factorize in a trivial way $$P^{\theta_1,\theta_2}(Q_1,Q_2)=\delta_{Q_1,2M} \delta_{Q_2,2M} \;\;\;\;\; P^{\theta_1,-}(Q_1,Q_2)=\delta_{Q_1,2M} \delta_{Q_2,4M}\, ,$$ in such a way that the CH inequality is never violated. This apparent contradiction is due to the fact that in this situation the scattering processes occur with unit probability, so that the condition of locality is fulfilled. Non-locality can be achieved by imposing $T< 1$. In the limit $T\ll 1$ we obtain the probabilities $P^{\theta_1,\theta_2}(Q_1,Q_2)$ and $P^{-,\theta_2}(Q_1,Q_2)$ reported, respectively, in Eqs. (A\[probS\]) and (A\[probSs\]) of the Appendix, which reduce to $$P^{\theta_1,\theta_2}(M,M)=\left[ \frac{2T^2A^6}{[A-T(A-1)]^8}\right]^M \left[ \sin^2\left( \frac{\theta_1+\theta_2}{2}\right)\right]^M \label{probS1}$$ and $$P^{-,\theta_2}(M,M)=\left[ \frac{2T^2A^6}{[A-T(A-1)]^8}\right]^M \label{probSs1}$$ for $Q_2=Q_3=M$, with $A=1+\tau_{he}\tau^{\prime\star}_{he}$. Eqs. (\[probS1\]) and (\[probSs1\]) are equal to Eqs. (\[prob\]), relative to the case of an entangler, once $2T^2A^6/[A-T(A-1)]^8$ is replaced with $T^2/2$. From this follows that superconductivity leads to violation of the CH inequality. For $A=2$, [i.e.]{} perfect Andreev transmission, the quantity $2T^2A^6/[A-T(A-1)]^8$ tends to $T^2/2$ in the limit $T\rightarrow 0$ so that the analysis of Section \[ns\] relative to the case $Q_1=Q_2=M$ applies also here. Conclusions =========== In mesoscopic multiterminal conductors it is possible to observe violations of locality in the whole distribution of the transmitted electrons. In this paper we have derived and discussed the CH inequality for the full counting electron statistics. In an idealized situation in which one supposes the existence of an [entangler]{}, we have found that the CH inequality is violated for joint probabilities relative to an equal number of electrons that have passed in different terminals. This is related to the intuition that any violation is lost in absence of coincidence measurements. The extent of the violation is suppressed for increasing $M$ (average number of injected pairs), however such a suppression does not scale exponentially with $M$ like the probability, but instead decreases like $1/M^2$. This means that the detection of violation does not become exponentially difficult with increasing $M$. For fixed transport properties we analyzed the conditions, in terms of $M$ and number of counted electrons, for maximizing the violation of the CH inequality. The violation of the CH inequality could be achieved in an experiment. Indeed we tested the CH inequality for two different realistic systems, namely a normal beam splitter and a superconducting beam splitter. Interestingly we find a violation even for the normal system, even though weaker with respect to the idealized case of the entangler. In this case the violation is again suppressed for increasing observation time, but scales like $1/M$. We analyzed the superconducting case in the limit of small transmissivity and we also find a violation of the CH inequality to the same extent with respect to the case of the entangler. It is important to notice that the analyzers should not affect the scattering properties of the system as in the case of ferromagnetic electrodes. In the latter case, in fact, the probability density of the local hidden variables would also depend on the angles $\theta_1$ and $\theta_2$. We believe that the results derived in this work may be of interest for the understanding of the statistics of electrons in mesoscopic conductors. It is however important to look for experimental tests of our claims. In this respect two possible schemes for measuring the counting statistics have been recently proposed in Ref. [@nazarov01_2]. Since solid state devices are considered promising implementations for quantum computational protocols, this line of research does not seem interesting only from a fundamental point of view, but may be of clear relevance for the actual realization of solid state computers. The authors would like to thank M. Büttiker, P. Samuelsson and E. Sukhorukov for helpful discussions and C.W.J. Beenakker for comments on the manuscript. This work has been supported by the EU (IST-FET-SQUBIT, RTN-Spintronics, RTN-Nanoscale Dynamics). \[PR\] Probability distributions ================================ In this appendix we give the general expressions for the joint probability distributions used in the paper to discuss the CH inequality. Entangler --------- In the case of an entangler we find $$\begin{aligned} P^{\theta_1,-}(Q_1,Q_2)=\frac{T^{(Q_1+Q_2)}}{2^M} {M\choose Q_1}{M\choose Q_2} \left( 2-T\right)^{M-Q_1}\left( 1-T\right)^{M-Q_2}\\ P^{-,\theta_2}(Q_1,Q_2)=\frac{T^{(Q_1+Q_2)}}{2^M} {M\choose Q_1}{M\choose Q_2} \left( 1-T\right)^{M-Q_1}\left( 2-T\right)^{M-Q_2}\end{aligned}$$ and $$\begin{aligned} P^{\theta_1,\theta_2}(Q_1,Q_2) &=& \sum_{k=\text{Max}[Q_1,Q_2]}^{\text{Min}[Q_1+Q_2,M]} {M\choose k} {k\choose 2k-Q_1-Q_2} {2k-Q_1-Q_2\choose k-Q_2}\times \nonumber \\ &\times & \frac{T^{(Q_1+Q_2)}}{2^M} \left[ 2(1-T) +T^2 \sin^2\left( \frac{\theta_1\pm\theta_2}{2}\right)\right]^{M-k} \times \nonumber \\ & \times & \left[ 1-T \sin^2\left( \frac{\theta_1\pm\theta_2}{2}\right)\right]^{2k-Q_1-Q_2} \left[ \sin^2\left( \frac{\theta_1\pm\theta_2}{2}\right)\right]^{Q_1+Q_2-k}\end{aligned}$$ Normal beam splitter {#normal-beam-splitter} -------------------- The joint probability $P^{\theta_1,\theta_2}(Q_1,Q_2)$ used in Section \[ns\] is $$\begin{aligned} P^{\theta_1,\theta_2}(Q_1,Q_2) &=& \sum_{k=\text{Max}[M-Q_1,M-Q_2]}^{\text{Min}[(M-Q_1)+(M-Q_2),M]} {M\choose k} {k\choose M-Q_2} {M-Q_2\choose Q_1-M+k}\times \nonumber \\ &\times & \epsilon^{(Q_1+Q_2)} \left[ 1-2\epsilon +\epsilon^2 \sin^2\left( \frac{\theta_1-\theta_2}{2}\right)\right] ^{2M-Q_1-Q_2-k} \times \nonumber \\ & \times & \left[ 1- \epsilon\sin^2\left( \frac{\theta_1-\theta_2}{2}\right)\right]^{Q_1+Q_2-2M+2k} \left[ \sin^2\left( \frac{\theta_1-\theta_2}{2}\right)\right]^{M-k}\end{aligned}$$ The single-analyzer joint probability $P^{-,\theta_2}(Q_1,Q_2)$ reads: $$\begin{aligned} P^{-,\theta_2}(Q_1,Q_2) &=& \epsilon^{(Q_1+Q_2)} \sum_{k=0}^{Q_1}~~~~~ \sum_{l=\text{Max}[0,(Q_1-k)+(Q_2-k)]}^{\text{Min}[M-k,Q_2]} {M\choose k} {M-k\choose l} {k\choose k+l-Q_2} \times \nonumber \\ &\times & {k+l-Q_2\choose Q_1-k} \left[ 1-3\epsilon +2\epsilon^2 \right]^{M-k-l} \left[ 1- \epsilon \right]^l \left[ 2- 3\epsilon\right]^{2k+l-Q_1-Q_2}\end{aligned}$$ with $0\le Q_1\le 2M$ and $0\le Q_2\le M$ (note that the sum on $l$ has to be performed only when the lower limit is less than or equal to the upper limit). Superconducting beam splitter {#superconducting-beam-splitter} ----------------------------- The joint probability $P^{\theta_1,\theta_2}(Q_1,Q_2)$ used in Section \[ss\] is $$\begin{aligned} \nonumber P^{\theta_1,\theta_2}(Q_1,Q_2) &=& \sum_{k=\text{Max}[Q_1,Q_2]}^{\text{Min}[Q_1+Q_2,M]} {M\choose k} {k\choose 2k-Q_1-Q_2} {2k-Q_1-Q_2\choose k-Q_2}\times \nonumber \\ & \times & \left[ \frac{A^8}{[A-T(A-1)]^8}\right]^M \left(\frac{2T^2}{A^2}\right)^k \times \nonumber \\ & \times & \left[ 1-4T+6T^2+\frac{2T^2}{A^2} \sin^2\left( \frac{\theta_1+\theta_2}{2}\right) \right]^{M-k}\times \nonumber \\ & \times & \left[ \sin^2\left( \frac{\theta_1+\theta_2}{2}\right)\right]^{Q_1+Q_2-k} \left[ \cos^2\left( \frac{\theta_1+\theta_2}{2}\right)\right]^{2k-Q_1-Q_2} \label{probS}\end{aligned}$$ where $A=1+\tau_{hp}\tau^{\prime\star}_{hp}$. The single-analyzer joint probability $P^{-,\theta_2}(Q_1,Q_2)$ reads: $$P^{-,\theta_2}(Q_1,Q_2)={M\choose Q_1} {Q_1\choose Q_2} \left( \frac{A^8}{[A-T(A-1)]^8}\right)^M \left( \frac{2T^2}{A^2}\right)^{Q_1} [1-4T+6T^2]^{M-Q_1} \label{probSs}$$ for $Q_1\ge Q_2$ and $P^{-,\theta_2}(Q_1,Q_2)=0$ for $Q_1< Q_2$. [99]{} J.S. Bell, [Speakable and unspeakable in Quantum Mechanics]{}, Cambridge University Press (1987). M. Nielsen and I. Chuang, [Quantum Computation and Quantum Communication]{}, Cambridge University Press, (2000). A. Zeilinger, Rev. Mod. Phys. [71]{}, S288,(1999). A. Rauschenbeutel, G. Nogues, S. Osnaghi, P. Bertet, M. Brune, J.-M. Raimond, and S. Haroche, Science [288]{}, 2024, (2000). C.A. Sackett, D. Kielpinski, B. E. King, C. Langer, V. Meyer, C.J. Myatt, M. Rowe, Q. A. Turchette, W.M. Itano, D.J. Wineland, and C. Monroe, Nature [404]{}, 256, (2000). Yu. Makhlin, G. Schön, and A. Shnirman, Rev. Mod. Phys. [73]{}, 357 (2001). , D.D. Awschalom, D. Loss, and N. Samarth Eds.: Series on Nanoscience and Technology, Springer-Verlag, Berlin, (2002). G. Burkard, D. Loss and E.V. Sukhorukov, Phys. Rev. B [61]{}, R16303 (2000). F. Taddei and R. Fazio, Phys. Rev. B [65]{}, 075317 (2002). The idea behind these approaches is to use the superconductor as a source of singlet pairs (the Cooper pairs). D. Loss and E.V. Sukhorukov, Phys. Rev. Lett. [84]{}, 1035 (2000). P. Recher, E.V. Sukhorukov and D. Loss, Phys. Rev. B [63]{}, 165314 (2001) G.B. Lesovik, T. Martin, G. Blatter, Eur. Phys. J. B [24]{}, 287 (2001). P. Samuelsson, E.V. Sukhorukov, M. Büttiker, Phys. Rev. Lett. [92]{}, 026805 (2004). C. Bena, S. Vishveshwara, L. Balents, M.P.A. Fisher, Phys. Rev. Lett. [89]{}, 037901 (2002). V. Bouchiat, N. Chtchelkatchev, D. Feinberg, G.B. Lesovik, T. Martin, J. Torres, Nanotechnology 14, [77]{} (2003). D.S. Saraga and D. Loss, Phys. Rev. Lett. [90]{}, 166803 (2003). A.T. Costa and S. Bose, Phys. Rev. Lett. [87]{}, 277901 (2001). C.W.J. Beenakker, C. Emary, M. Kindermann, J.L. van Velsen, Phys. Rev. Lett. [91]{}, 147901 (2003). F. Plastina, R. Fazio and M. Palma, Phys. Rev. B [64]{}, 113306 (2001). O. Buisson, F.W.J. Hekking, in [Macroscopic Quantum Coherence and Quantum Computing]{} edited by D.V. Averin, B. Ruggiero, and P. Silvestrini (Kluwer, New York, 2001), p. 137. F. Marquardt and C. Bruder, Phys. Rev. B [63]{}, 054514 (2001). F. Plastina and G. Falci, Phys. Rev. B [67]{}, 224514 (2003). J.Q. You, J.S. Tsai, and F. Nori, Phys. Rev. Lett. [89]{}, 197902 (2002). J.S. Bell, Physics [1]{}, 195 (1964). S. Kawabata, J. Phys. Soc. Jpn. [70]{}, 1210 (2001) N.M. Chtchelkatchev, G. Blatter, G.B. Lesovik, T. Martin, Phys. Rev. B [66]{}, 161320 (2002). A central issue of these works is the ability to perform coincident measurements which can be achieved either by performing short-time measurement [@kawabata00; @chtchelkatchev02] or by operating in the tunneling limit [@samuelsson03; @beenakker03]. J.F. Clauser and M.A. Horne, Phys. Rev. D [10]{}, 526 (1974). L. Mandel and E. Wolf, [Optical coherence and quantum optics]{}, Cambridge University Press (1995). A. Brataas, Yu. V. Nazarov, and G. E. W. Bauer, Eur. Phys. J. B [22]{}, 99 (2001). L.S. Levitov and G.B. Lesovik, Pis’ma Zh. Eksp. Tepr. Fiz [58]{}, 225 (1993) \[JETP Lett. [58]{}, 225 (1993)\]. L.S. Levitov, H. Lee and G.B. Lesovik, Jour. Math. Phys. [37]{}, 10 (1996). Yu.V. Nazarov, Ann. Phys. (Leipzig) [8]{}, Spec. Issue, SI-193 (1999). W. Belzig, Yu.V. Nazarov, Phys. Rev. Lett. [87]{}, 067006 (2001); W. Belzig, Yu.V. Nazarov, Phys. Rev. Lett. [87]{}, 197006 (2001). L.S. Levitov, in “Quantum Noise in Mesoscopic Physics”, edited by Yu.V. Nazarov (Kluwer, Amsterdam, 2003); M. Kindermann and Yu.V. Nazarov, [ibidem]{}; D.A. Bagrets and Yu.V. Nazarov, [ibidem]{}; W. Belzig, [ibidem]{}. M. Büttiker, Phys. Rev. B [46]{}, 12485 (1992). G. B. Lesovik, Pis’ma Zh. Eksp. Teor. Fiz. [49]{}, 513 (1989) \[JETP Lett. [49]{}, 592 (1989)\]. B.A. Muzykantskii and D.E. Khmelnitskii, Phys. Rev. B [50]{}, 3982 (1994). M. Büttiker, Y. Imry, and M.Ya. Azbel, Phys. Rev. A [30]{}, 1982 (1984). S. Bose and D. Home, Phys. Rev. Lett. [88]{}, 050401 (2002). S. Datta, [Electronic Transport in Mesoscopic Systems]{}, Cambridge University Press (1995). Yu.V. Nazarov and M. Kindermann, Eur. Phys. J. B [35]{}, 413 (2003). J. Börlin, W. Belzig, and C. Bruder, Phys. Rev. Lett. [88]{}, 197001 (2002). P. Samuelsson and M. Büttiker, Phys. Rev. Lett. [89]{}, 046601 (2002). [^1]: The CH inequality given in Eq.(\[CHin\]) is said to be weak in the sense that it holds only when the no-enhancement assumption, Eq.(\[noen\]), is satisfied.	{ "pile_set_name": "ArXiv" }
--- author: - Luis Aparicio title: Esa --- [99]{} S. Weinberg, “The Quantum theory of fields. Vol. 1: Foundations,” Cambridge, UK: Univ. Pr. (1995) 609 p.\ S. Weinberg, “The quantum theory of fields. Vol. 2: Modern applications,” Cambridge, UK: Univ. Pr. (1996) 489 p.\ M. Srednicki, “Quantum field theory,” Cambridge, UK: Univ. Pr. (2007) 641 p S. W. Hawking and G. F. R. Ellis, “The Large scale structure of space-time,” Cambridge University Press, Cambridge, 1973.\ R. M. Wald, “General Relativity,” Chicago, Usa: Univ. Pr. ( 1984) 491p. Q. Ho-Kim, (ed.) and X. -Y. Pham, (ed.), “Elementary particles and their interactions,” Berlin, Germany, Springer-Verlag, 1998. 661p.\ T. P. Cheng and L. F. Li, “Gauge theory of elementary particle physics: Problems and solutions,” Oxford, UK: Clarendon (2000) 306 p. V. Mukhanov, “Physical foundations of cosmology,” Cambridge, UK: Univ. Pr. (2005) 421 p. Y. Fukuda [et al.]{} \[Super-Kamiokande Collaboration\], “Evidence for oscillation of atmospheric neutrinos,” Phys. Rev. Lett. [81]{} (1998) 1562 \[hep-ex/9807003\].\ Q. R. Ahmad [et al.]{} \[SNO Collaboration\], “Measurement of the rate of nu/e + d –> p + p + e- interactions produced by B-8 solar neutrinos at the Sudbury Neutrino Observatory,” Phys. Rev. Lett. [87]{} (2001) 071301 \[nucl-ex/0106015\]. A. J. Conley [et al.]{} \[Supernova Cosmology Project Collaboration\], “Measurement of Omega(m), Omega(lambda) from a blind analysis of Type Ia supernovae with CMAGIC: Using color information to verify the acceleration of the Universe,” Astrophys. J. [644]{} (2006) 1 \[astro-ph/0602411\]. E. Komatsu [et al.]{} \[WMAP Collaboration\], “Seven-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation,” Astrophys. J. Suppl. [192]{} (2011) 18 \[arXiv:1001.4538 \[astro-ph.CO\]\]. F. Derue [et al.]{} \[The EROS Collaboration\], “Observation of microlensing towards the galactic spiral arms. EROS2 2 year survey,” Astron. Astrophys. [351]{} (1999) 87 \[astro-ph/9903209\]. J. Alcaraz [et al.]{} \[AMS Collaboration\], “Search for anti-helium in cosmic rays,” Phys. Lett. B [461]{} (1999) 387 \[hep-ex/0002048\]. A. H. Guth, “The Inflationary Universe: A Possible Solution to the Horizon and Flatness Problems,” Phys. Rev. D [23]{} (1981) 347. A. H. Guth, “Eternal inflation and its implications,” J. Phys. A A [40]{} (2007) 6811 \[hep-th/0702178 \[HEP-TH\]\]. J. Elias-Miro, J. R. Espinosa, G. F. Giudice, G. Isidori, A. Riotto and A. Strumia, “Higgs mass implications on the stability of the electroweak vacuum,” Phys. Lett. B [709]{} (2012) 222 \[arXiv:1112.3022 \[hep-ph\]\]. G. Isidori, V. S. Rychkov, A. Strumia and N. Tetradis, “Gravitational corrections to standard model vacuum decay,” Phys. Rev. D [77]{} (2008) 025034 \[arXiv:0712.0242 \[hep-ph\]\]. S. P. Martin, “A Supersymmetry primer,” In \Kane, G.L. (ed.): Perspectives on supersymmetry II\ 1-153 \[hep-ph/9709356\]. H. Baer and X. Tata, “Weak scale supersymmetry: From superfields to scattering events,” Cambridge, UK: Univ. Pr. (2006) 537 p. J. Polchinski, “String theory. Vol. 1: An introduction to the bosonic string,” Cambridge, UK: Univ. Pr. (1998) 402 p.\ J. Polchinski, “String theory. Vol. 2: Superstring theory and beyond,” Cambridge, UK: Univ. Pr. (1998) 531 p. M. B. Green, J. H. Schwarz and E. Witten, “Superstring Theory. Vol. 1: Introduction,” Cambridge, Uk: Univ. Pr. ( 1987) 469 P. ( Cambridge Monographs On Mathematical Physics).\ M. B. Green, J. H. Schwarz and E. Witten, “Superstring Theory. Vol. 2: Loop Amplitudes, Anomalies And Phenomenology,” Cambridge, Uk: Univ. Pr. ( 1987) 596 P. ( Cambridge Monographs On Mathematical Physics). L. E. Ibanez and A. M. Uranga, “String theory and particle physics: An introduction to string phenomenology,” Cambridge, UK: Univ. Pr. (2012) 673 p. A. Strominger and C. Vafa, “Microscopic origin of the Bekenstein-Hawking entropy,” Phys. Lett. B [379]{} (1996) 99 \[hep-th/9601029\]. J. M. Maldacena, “The Large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys. [2]{} (1998) 231 \[Int. J. Theor. Phys. [38]{} (1999) 1113\] \[hep-th/9711200\]. L. Aparicio, D. G. Cerdeño and L. E. Ibáñez, “Modulus-dominated SUSY-breaking soft terms in F-theory and their test at LHC,” JHEP [0807]{} (2008) 099 \[arXiv:0805.2943 \[hep-ph\]\]. L. Aparicio, D. G. Cerdeno and L. E. Ibanez, “A 119-125 GeV Higgs from a string derived slice of the CMSSM,” arXiv:1202.0822 \[hep-ph\]. L. Aparicio, A. Font, L. E. Ibanez and F. Marchesano, “Flux and Instanton Effects in Local F-theory Models and Hierarchical Fermion Masses,” JHEP [1108]{} (2011) 152 \[arXiv:1104.2609 \[hep-th\]\]. T. Ortin, “Gravity and strings,” Cambridge Unversity, Cambridge University Press, 2004 J. Polchinski, “Dirichlet Branes and Ramond-Ramond charges,” Phys. Rev. Lett. [75]{} (1995) 4724 \[hep-th/9510017\]. J. E. Paton and H. -M. Chan, “Generalized veneziano model with isospin,” Nucl. Phys. B [10]{} (1969) 516. E. Witten, “Bound states of strings and p-branes,” Nucl. Phys. B [460]{} (1996) 335 \[hep-th/9510135\]. J. H. Schwarz, “Gauge Groups For Type I Superstrings,” In \Florence 1982, Proceedings, Lattice Gauge Theory, Supersymmetry and Grand Unification\, 233-245 and Caltech Pasadena - CALT-68-906 (82,REC.APR.) 11p N. Marcus and A. Sagnotti, “Tree Level Constraints on Gauge Groups for Type I Superstrings,” Phys. Lett. B [119]{} (1982) 97. A. M. Uranga, “Intersecting brane worlds,” Class. Quant. Grav. [22]{} (2005) S41. M. Berkooz, M. R. Douglas and R. G. Leigh, “Branes intersecting at angles,” Nucl. Phys. B [480]{} (1996) 265 \[hep-th/9606139\]. H. Arfaei and M. M. Sheikh Jabbari, “Different d-brane interactions,” Phys. Lett. B [394]{} (1997) 288 \[hep-th/9608167\]. V. Balasubramanian and R. G. Leigh, “D-branes, moduli and supersymmetry,” Phys. Rev. D [55]{} (1997) 6415 \[hep-th/9611165\]. M. M. Sheikh Jabbari, “Classification of different branes at angles,” Phys. Lett. B [420]{} (1998) 279 \[hep-th/9710121\]. M. Berkooz and R. G. Leigh, “A D = 4 N=1 orbifold of type I strings,” Nucl. Phys. B [483]{} (1997) 187 \[hep-th/9605049\]. F. Gliozzi, J. Scherk and D. I. Olive, “Supergravity and the Spinor Dual Model,” Phys. Lett. B [65]{} (1976) 282. P. S. Aspinwall, “D-branes on Calabi-Yau manifolds,” hep-th/0403166. G. Aldazabal, S. Franco, L. E. Ibanez, R. Rabadan and A. M. Uranga, “D = 4 chiral string compactifications from intersecting branes,” J. Math. Phys. [42]{} (2001) 3103 \[hep-th/0011073\]. R. Blumenhagen, L. Goerlich, B. Kors and D. Lust, “Noncommutative compactifications of type I strings on tori with magnetic background flux,” JHEP [0010]{} (2000) 006 \[hep-th/0007024\]. R. Blumenhagen, B. Kors, D. Lust and T. Ott, “Intersecting brane worlds on tori and orbifolds,” Fortsch. Phys. [50]{} (2002) 843 \[hep-th/0112015\]. A. M. Uranga, “Local models for intersecting brane worlds,” JHEP [0212]{} (2002) 058 \[hep-th/0208014\]. R. Blumenhagen, V. Braun, B. Kors and D. Lust, “Orientifolds of K3 and Calabi-Yau manifolds with intersecting D-branes,” JHEP [0207]{} (2002) 026 \[hep-th/0206038\]. G. Aldazabal, L. E. Ibanez, F. Quevedo and A. M. Uranga, “D-branes at singularities: A Bottom up approach to the string embedding of the standard model,” JHEP [0008]{} (2000) 002 \[hep-th/0005067\]. C. Bachas, “D-brane dynamics,” Phys. Lett. B [374]{} (1996) 37 \[hep-th/9511043\]. C. Angelantonj, I. Antoniadis, E. Dudas and A. Sagnotti, “Type I strings on magnetized orbifolds and brane transmutation,” Phys. Lett. B [489]{} (2000) 223 \[hep-th/0007090\]. R. Blumenhagen, B. Kors, D. Lust and T. Ott, “The standard model from stable intersecting brane world orbifolds,” Nucl. Phys. B [616]{} (2001) 3 \[hep-th/0107138\]. J. F. G. Cascales and A. M. Uranga, “Chiral 4-D string vacua with D-branes and moduli stabilization,” hep-th/0311250. R. Blumenhagen, V. Braun, T. W. Grimm and T. Weigand, “GUTs in Type IIB Orientifold Compactifications,” Nucl. Phys. B [815]{} (2009) 1 \[arXiv:0811.2936 \[hep-th\]\]. J. H. Schwarz, Nucl. Phys. Proc. Suppl. [49]{} (1996) 183 \[hep-th/9509148\]. C. Vafa, “Evidence for F theory,” Nucl. Phys. B [469]{} (1996) 403 \[hep-th/9602022\]. D. R. Morrison and C. Vafa, “Compactifications of F theory on Calabi-Yau threefolds. 2.,” Nucl. Phys. B [476]{} (1996) 437 \[hep-th/9603161\]. T. Weigand, [“Lectures on F-theory compactifications and model building,”]{} Class. Quant. Grav. [27]{}, 214004 (2010) \[arXiv:1009.3497 \[hep-th\]\]. A. Sen, “An Introduction to nonperturbative string theory,” In \Cambridge 1997, Duality and supersymmetric theories\ 297-413 \[hep-th/9802051\]. J. J. Heckman, [“Particle Physics Implications of F-theory,”]{} arXiv:1001.0577 \[hep-th\]. C. Beasley, J. J. Heckman and C. Vafa, [“GUTs and Exceptional Branes in F-theory - I,”]{} JHEP [0901]{} (2009) 058 \[arXiv:0802.3391 \[hep-th\]\]. C. Beasley, J. J. Heckman and C. Vafa, [“GUTs and Exceptional Branes in F-theory - II: Experimental Predictions,”]{} JHEP [0901]{} (2009) 059 \[arXiv:0806.0102 \[hep-th\]\]. R. Donagi, M. Wijnholt, [“Model Building with F-Theory,”]{} \[arXiv:0802.2969 \[hep-th\]\]. R. Donagi and M. Wijnholt, [“Breaking GUT Groups in F-Theory,”]{} arXiv:0808.2223 \[hep-th\]. B. R. Greene, A. D. Shapere, C. Vafa and S. -T. Yau, “Stringy Cosmic Strings and Noncompact Calabi-Yau Manifolds,” Nucl. Phys. B [337]{} (1990) 1. F. Marchesano, “Progress in D-brane model building,” arXiv:hep-th/0702094;\ R. Blumenhagen, B. Kors, D. Lust and S. Stieberger, “Four-dimensional String Compactifications with D-Branes, Orientifolds and Fluxes,” Phys. Rept. [445]{} (2007) 1 \[arXiv:hep-th/0610327\];\ A. M. Uranga, “Chiral four-dimensional string compactifications with intersecting D-branes,” Class. Quant. Grav. [20]{} (2003) S373 \[arXiv:hep-th/0301032\];\ R. Blumenhagen, M. Cvetic, P. Langacker and G. Shiu, “Toward realistic intersecting D-brane models,” Ann. Rev. Nucl. Part. Sci. [55]{} (2005) 71 \[arXiv:hep-th/0502005\];\ F. G. Marchesano, “Intersecting D-brane models,” \[arXiv:hep-th/0307252\];\ D. Cremades, L. E. Ibáñez and F. Marchesano, “More about the standard model at intersecting branes,” arXiv:hep-ph/0212048. P. G. Camara, L. E. Ibáñez and A. M. Uranga, “Flux-induced SUSY-breaking soft terms on D7-D3 brane systems,” Nucl. Phys. B [708]{} (2005) 268 \[arXiv:hep-th/0408036\]. D. Berenstein, V. Jejjala and R. G. Leigh, “The standard model on a D-brane,” Phys. Rev. Lett. [88]{} (2002) 071602 \[arXiv:hep-ph/0105042\]. H. Verlinde and M. Wijnholt, “Building the standard model on a D3-brane,” JHEP [0701]{} (2007) 106 \[arXiv:hep-th/0508089\];\ D. Malyshev and H. Verlinde, “D-branes at Singularities and String Phenomenology,” Nucl. Phys. Proc. Suppl. [171]{} (2007) 139 \[arXiv:0711.2451 \[hep-th\]\]. D. Cremades, L. E. Ibáñez and F. Marchesano, “Yukawa couplings in intersecting D-brane models,” JHEP [0307]{} (2003) 038 \[arXiv:hep-th/0302105\]. D. Cremades, L. E. Ibáñez and F. Marchesano, “Computing Yukawa couplings from magnetized extra dimensions,” JHEP [0405]{} (2004) 079 \[arXiv:hep-th/0404229\]. M. Cvetic, G. Shiu and A. M. Uranga, “Chiral four-dimensional N = 1 supersymmetric type IIA orientifolds from intersecting D6-branes,” Nucl. Phys. B [615]{} (2001) 3 \[arXiv:hep-th/0107166\];\ M. Cvetic, T. Li and T. Liu, “Standard-like models as type IIB flux vacua,” Phys. Rev. D [71]{} (2005) 106008 \[arXiv:hep-th/0501041\];\ C. M. Chen, V. E. Mayes and D. V. Nanopoulos, “MSSM via Pati-Salam from Intersecting Branes on $T^6/({ Z_2} \times {Z_2'})$,” Phys. Lett. B [648]{} (2007) 301 \[arXiv:hep-th/0612087\]. F. Marchesano and G. Shiu, [“MSSM vacua from flux compactifications,”]{} Phys. Rev. D [71]{}, 011701 (2005) \[arXiv:hep-th/0408059\]. [“Building MSSM flux vacua,”]{} JHEP [0411]{}, 041 (2004) \[arXiv:hep-th/0409132\]. A. Font, L. E. Ibáñez and F. Marchesano, “Coisotropic D8-branes and model-building,” JHEP [0609]{} (2006) 080 \[arXiv:hep-th/0607219\]. A. Font and L. E. Ibáñez, “SUSY-breaking soft terms in a MSSM magnetized D7-brane model,” JHEP [0503]{} (2005) 040 \[arXiv:hep-th/0412150\]. L. E. Ibáñez, C. Muñoz and S. Rigolin, “Aspects of type I string phenomenology,” Nucl. Phys. B [553]{} (1999) 43 \[arXiv:hep-ph/9812397\]. D. Lust, P. Mayr, R. Richter and S. Stieberger, “Scattering of gauge, matter, and moduli fields from intersecting branes,” Nucl. Phys. B [696]{} (2004) 205 \[arXiv:hep-th/0404134\];\ D. Lust, S. Reffert and S. Stieberger, “Flux-induced soft supersymmetry breaking in chiral type IIb orientifolds with D3/D7-branes,” Nucl. Phys. B [706]{} (2005) 3 \[arXiv:hep-th/0406092\] V. Balasubramanian, P. Berglund, J. P. Conlon and F. Quevedo, “Systematics of moduli stabilisation in Calabi-Yau flux compactifications,” JHEP [0503]{}, 007 (2005) \[arXiv:hep-th/0502058\];\ J. P. Conlon, F. Quevedo and K. Suruliz, “Large-volume flux compactifications: Moduli spectrum and D3/D7 soft supersymmetry breaking,” JHEP [0508]{} (2005) 007 \[arXiv:hep-th/0505076\]; J. P. Conlon, S. S. Abdussalam, F. Quevedo and K. Suruliz, “Soft SUSY breaking terms for chiral matter in IIB string compactifications,” JHEP [0701]{} (2007) 032 \[arXiv:hep-th/0610129\];\ J. P. Conlon, C. H. Kom, K. Suruliz, B. C. Allanach and F. Quevedo, “Sparticle Spectra and LHC Signatures for Large Volume String Compactifications,” JHEP [0708]{} (2007) 061 \[arXiv:0704.3403 \[hep-ph\]\]. J. P. Conlon, D. Cremades and F. Quevedo, “Kaehler potentials of chiral matter fields for Calabi-Yau string compactifications,” JHEP [0701]{} (2007) 022 \[arXiv:hep-th/0609180\]. M. Berg, M. Haack and E. Pajer, “Jumping Through Loops: On Soft Terms from Large Volume Compactifications,” JHEP [0709]{} (2007) 031 \[arXiv:0704.0737 \[hep-th\]\]. S. S. AbdusSalam, J. P. Conlon, F. Quevedo and K. Suruliz, “Scanning the Landscape of Flux Compactifications: Vacuum Structure and Soft Supersymmetry Breaking,” arXiv:0709.0221 \[hep-th\]. M. Cicoli, J. P. Conlon and F. Quevedo, “Systematics of String Loop Corrections in Type IIB Calabi-Yau Flux Compactifications,” JHEP [0801]{} (2008) 052 \[arXiv:0708.1873 \[hep-th\]\]; “General Analysis of LARGE Volume Scenarios with String Loop Moduli Stabilisation,” arXiv:0805.1029 \[hep-th\]. R. Blumenhagen, S. Moster and E. Plauschinn, “Moduli Stabilisation versus Chirality for MSSM like Type IIB Orientifolds,” JHEP [0801]{} (2008) 058 \[arXiv:0711.3389 \[hep-th\]\]. P. Candelas, A. Font, S. H. Katz and D. R. Morrison, “Mirror symmetry for two parameter models. 2,” Nucl. Phys. B [429]{} (1994) 626 \[arXiv:hep-th/9403187\];\ F. Denef, M. R. Douglas and B. Florea, “Building a better racetrack,” JHEP [0406]{} (2004) 034 \[arXiv:hep-th/0404257\]. D. Cremades, L. E. Ibanez and F. Marchesano, “SUSY quivers, intersecting branes and the modest hierarchy problem,” JHEP [0207]{} (2002) 009 \[arXiv:hep-th/0201205\]. P. G. Cámara, L. E. Ibáñez and A. M. Uranga, [Flux-induced SUSY-breaking soft terms]{}, Nucl. Phys. B[689]{} (2004) 195, hep-th/0311241;\ M. Graña, T. W. Grimm, H. Jockers and J. Louis, [Soft supersymmetry breaking in Calabi-Yau orientifolds with D-branes and fluxes]{}, Nucl. Phys. B[690]{} (2004) 21, hep-th/0312232; L. E. Ibáñez, “The fluxed MSSM,” Phys. Rev. D [71]{} (2005) 055005 \[arXiv:hep-ph/0408064\]. M. Graña, “MSSM parameters from supergravity backgrounds,” Phys. Rev. D [67]{}, 066006 (2003) \[arXiv:hep-th/0209200\]. G. F. Giudice and A. Masiero, “A Natural Solution to the mu Problem in Supergravity Theories,” Phys. Lett. B [206]{} (1988) 480. A. Brignole, L. E. Ibáñez and C. Muñoz, “Soft supersymmetry-breaking terms from supergravity and superstring models,” in ’Perspectives on Supersymmetry” p. 125, ed.G.L. Kane, World Scientific Co. (1998), arXiv:hep-ph/9707209. F. Gianotti, CERN Public Seminar, Update on the Standard Model Higgs searches in ATLAS, 13th December 2011. The ATLAS Collaboration, ATLAS-CONF-2011-163. G. Tonelli, CERN Public Seminar, Update on the Standard Model Higgs searches in CMS, 13th December 2011. The CMS Collaboration, CMS-PAS-HIG-11-032. A. H. Chamseddine, R. Arnowitt and P. Nath, “Locally Supersymmetric Grand Unification,” Phys. Rev. Lett. [49]{}, 970 (1982);\ L. E. Ibáñez, “Locally Supersymmetric SU(5) Grand Unification,” Phys. Lett. B [118]{}, 73 (1982);\ R. Barbieri, S. Ferrara and C. A. Savoy, “Gauge Models With Spontaneously Broken Local Supersymmetry,” Phys. Lett. B [119]{}, 343 (1982);\ L. J. Hall, J. Lykken and S. Weinberg, “Supergravity As The Messenger Of Supersymmetry Breaking,” Phys. Rev. D [27]{}, 2359 (1983). H. Baer, V. Barger and A. Mustafayev, “Implications of a 125 GeV Higgs scalar for LHC SUSY and neutralino dark matter searches,” arXiv:1112.3017 \[hep-ph\]. L. J. Hall, D. Pinner and J. T. Ruderman, “A Natural SUSY Higgs Near 126 GeV,” arXiv:1112.2703 \[hep-ph\]. A. Arbey, M. Battaglia, A. Djouadi, F. Mahmoudi and J. Quevillon, “Implications of a 125 GeV Higgs for supersymmetric models,” Phys. Lett. B [708]{} (2012) 162 \[arXiv:1112.3028 \[hep-ph\]\]. S. Akula, B. Altunkaynak, D. Feldman, P. Nath and G. Peim, “Higgs Boson Mass Predictions in SUGRA Unification, Recent LHC-7 Results, and Dark Matter,” arXiv:1112.3645 \[hep-ph\]. I. Gogoladze, Q. Shafi and C. S. Un, “Higgs Boson Mass from t-b-$\tau$ Yukawa Unification,” arXiv:1112.2206 \[hep-ph\]. M. Carena, S. Gori, N. R. Shah and C. E. M. Wagner, “A 125 GeV SM-like Higgs in the MSSM and the $\gamma \gamma$ rate,” arXiv:1112.3336 \[hep-ph\]. P. Draper, P. Meade, M. Reece and D. Shih, “Implications of a 125 GeV Higgs for the MSSM and Low-Scale SUSY Breaking,” arXiv:1112.3068 \[hep-ph\]. J. L. Evans, M. Ibe, S. Shirai and T. T. Yanagida, “A 125GeV Higgs Boson and Muon g-2 in More Generic Gauge Mediation,” arXiv:1201.2611 \[hep-ph\]. A. Brignole, L. E. Ibáñez and C. Muñoz, “Towards a theory of soft terms for the supersymmetric Standard Model,” Nucl. Phys. B [422]{} (1994) 125 \[Erratum-ibid. B [436]{} (1995) 747\] \[hep-ph/9308271\]. M. Cvetic, A. Font, L. E. Ibáñez, D. Lust and F. Quevedo, “Target space duality, supersymmetry breaking and the stability of classical string vacua,” Nucl. Phys. B [361]{} (1991) 194. L. E. Ibáñez and D. Lust, “Duality anomaly cancellation, minimal string unification and the effective low-energy Lagrangian of 4-D strings,” Nucl. Phys. B [382]{} (1992) 305 \[arXiv:hep-th/9202046\]. V. S. Kaplunovsky and J. Louis, “Model independent analysis of soft terms in effective supergravity and in string theory,” Phys. Lett. B [306]{} (1993) 269 \[hep-th/9303040\]. B. C. Allanach, A. Brignole and L. E. Ibáñez, “Phenomenology of a fluxed MSSM,” JHEP [0505]{} (2005) 030 \[hep-ph/0502151\];\ K. Choi, A. Falkowski, H. P. Nilles and M. Olechowski, “Soft supersymmetry breaking in KKLT flux compactification,” Nucl. Phys. B [718]{} (2005) 113 \[arXiv:hep-th/0503216\];\ K. Choi and H. P. Nilles, “The gaugino code,” JHEP [0704]{} (2007) 006 \[arXiv:hep-ph/0702146\];\ R. Blumenhagen, J. P. Conlon, S. Krippendorf, S. Moster and F. Quevedo, “SUSY Breaking in Local String/F-Theory Models,” JHEP [0909]{} (2009) 007 \[arXiv:0906.3297 \[hep-th\]\];\ J. J. Heckman, G. L. Kane, J. Shao and C. Vafa, “The Footprint of F-theory at the LHC,” JHEP [0910]{} (2009) 039 \[arXiv:0903.3609 \[hep-ph\]\];\ G. Kane, P. Kumar, R. Lu and B. Zheng, “Higgs Mass Prediction for Realistic String/M Theory Vacua,” arXiv:1112.1059 \[hep-ph\];\ S. P. de Alwis, “Classical and Quantum SUSY Breaking Effects in IIB Local Models,” JHEP [1003]{} (2010) 078 \[arXiv:0912.2950 \[hep-th\]\];\ T. Li, J. A. Maxin, D. V. Nanopoulos and J. W. Walker, “A Higgs Mass Shift to 125 GeV and A Multi-Jet Supersymmetry Signal: Miracle of the Flippons at the $\sqrt{s} = 7$ TeV LHC,” arXiv:1112.3024 \[hep-ph\]. B. S. Acharya, K. Bobkov, G. L. Kane, J. Shao and P. Kumar, “The $G_2$-MSSM - An $M$ Theory motivated model of Particle Physics,” arXiv:0801.0478 \[hep-ph\];\ B. S. Acharya, K. Bobkov, G. L. Kane, P. Kumar and J. Shao, “Explaining the electroweak scale and stabilizing moduli in M theory,” Phys. Rev. D [76]{} (2007) 126010 \[arXiv:hep-th/0701034\];\ B. Acharya, K. Bobkov, G. Kane, P. Kumar and D. Vaman, “An M theory solution to the hierarchy problem,” Phys. Rev. Lett. [97]{} (2006) 191601 \[arXiv:hep-th/0606262\]. S. Kachru, R. Kallosh, A. Linde and S. P. Trivedi, “De Sitter vacua in string theory,” Phys. Rev. D [68]{} (2003) 046005 \[arXiv:hep-th/0301240\]. F. Denef, “Les Houches Lectures on Constructing String Vacua,” arXiv:0803.1194 \[hep-th\]. M. R. Douglas and S. Kachru, “Flux compactification,” Rev. Mod. Phys. [79]{} (2007) 733 \[hep-th/0610102\]. D. Lust, S. Reffert and S. Stieberger, “MSSM with soft SUSY breaking terms from D7-branes with fluxes,” Nucl. Phys. B [727]{} (2005) 264 \[hep-th/0410074\]; E. Cremmer, S. Ferrara, C. Kounnas and D. V. Nanopoulos, “Naturally Vanishing Cosmological Constant In N=1 Supergravity,” Phys. Lett. B [133]{} (1983) 61. S. B. Giddings, S. Kachru and J. Polchinski, [Hierarchies from fluxes in string compactifications]{}, Phys. Rev. D[66]{} (2002) 106006, hep-th/0105097. S. Dimopoulos, S. Raby and F. Wilczek, “Supersymmetry And The Scale Of Unification,” Phys. Rev. D [24]{} (1981) 1681;\ L. E. Ibáñez and G. G. Ross, “Low-Energy Predictions In Supersymmetric Grand Unified Theories,” Phys. Lett. B [105]{} (1981) 439;\ S. Dimopoulos and H. Georgi, “Softly Broken Supersymmetry And SU(5),” Nucl. Phys. B [193]{} (1981) 150. W. M. Yao [et al.]{} \[Particle Data Group\], “Review of particle physics,” J. Phys. G [33]{} (2006) 1. L. E. Ibáñez and G. G. Ross, “$SU(2)_L \times U(1)_L$ Symmetry Breaking As A Radiative Effect Of Supersymmetry Breaking In Guts,” Phys. Lett. B [110]{} (1982) 215. W. Porod, “SPheno, a program for calculating supersymmetric spectra, SUSY particle decays and SUSY particle production at e+ e- colliders,” Comput. Phys. Commun. [153]{} (2003) 275 \[arXiv:hep-ph/0301101\]. E. Barberio [et al.]{} \[Heavy Flavor Averaging Group (HFAG) Collaboration\], “Averages of b-hadron properties at the end of 2006,” arXiv:0704.3575 \[hep-ex\]. M. Misiak and M. Steinhauser, ‘NNLO QCD corrections to the $\bar B \to X_s \gamma$ matrix elements using interpolation in $m_c$’, [Nucl. Phys. B]{} [764]{} (2007) 62 \[arXiv:hep-ph/0609241\]; M. Misiak [et al.]{}, ‘The first estimate of B$(\bar B \to X_s \gamma)$ at ${\cal O} (\alpha_s^2)$’, [Phys. Rev. Lett.]{} [98]{} (2007) 022002 \[arXiv:hep-ph/0609232\]. T. Aaltonen [et al.]{} \[CDF Collaboration\], “Search for $B_s\to\mu^+\mu^-$ and $B_d\to\mu^+\mu^-$ Decays with 2fb$^-1$ of $p\bar p$ Collisions,” Phys. Rev. Lett. [100]{} (2008) 101802 \[arXiv:0712.1708 \[hep-ex\]\]. V. M. Abazov [et al.]{} \[D0 Collaboration\], “Search for $B_s \to \mu^{+} \mu^{-}$ at D0,” Phys. Rev. D [76]{} (2007) 092001 \[arXiv:0707.3997 \[hep-ex\]\]. Muon g-2 Collaboration, G. W. Bennett [et al.]{}, ‘Measurement of the negative muon anomalous magnetic moment to 0.7-ppm’, [Phys. Rev. Lett.]{} [92]{}, 161802 (2004) \[arXiv:hep-ex/0401008\]. T. Kinoshita and W. J. Marciano, in [Quantum Electrodynamics]{}, ed. T. Kinoshita (World Scientific, Singapore, 1990), p. 419; T. Kinoshita, ‘New value of the $\alpha^3$ electron anomalous magnetic moment’, [Phys. Rev. Lett.]{} [75]{} (1995) 4728; A. Czarnecki, B. Krause and W. J. Marciano, ‘Electroweak corrections to the muon anomalous magnetic moment’, [Phys. Rev. Lett.]{} [76]{} (1996) 3267 \[arXiv:hep-ph/9512369\]; S. Eidelman and F. Jegerlehner, ‘Hadronic contributions to $g-2$ of the leptons and to the effective fine structure constant $\alpha (M_Z^2)$’, [Z. Phys. C]{} [67]{} (1995) 585 \[arXiv:hep-ph/9502298\]; K. Adel and F. J. Ynduráin, ‘Improved evaluation of the hadronic vacuum polarization contributions to muon $g-2$ and $\bar\alpha_{\rm QED}(M_Z)$ using high order QCD calculations’, arXiv:hep-ph/9509378; T. Kinoshita, B. Nizic and Y. Okamoto, ‘Hadronic Contributions To The Anomalous Magnetic Moment Of The Muon’, [Phys. Rev. D]{} [31]{} (1985) 2108; J. Bijnens, E. Pallante and J. Prades, ‘Hadronic light by light contributions to the muon $g-2$ in the large N(c) limit’, [Phys. Rev. Lett.]{} [75]{} (1995) 1447 \[Erratum-ibid. [75]{} (1995) 3781\] \[arXiv:hep-ph/9505251\]. M. Davier, S. Eidelman, A. Höcker and Z. Zhang, ‘Updated estimate of the muon magnetic moment using revised results from $e^+ e^-$ annihilation’, [Eur. Phys. J. C]{} [31]{} (2003) 503 \[arXiv:hep-ph/0308213\]; K. Hagiwara, A. D. Martin, D. Nomura and T. Teubner, ‘Predictions for g-2 of the muon and $\alpha_{QED}(M_Z^2)$’, [Phys. Rev. D]{} [69]{}, 093003 (2004) \[arXiv:hep-ph/0312250\]; J. F. de Trocóniz and F. J. Ynduráin, ‘The hadronic contributions to the anomalous magnetic moment of the muon’, [Phys. Rev. D]{} [71]{}, 073008 (2005) \[arXiv:hep-ph/0402285\]. T. Kinoshita and M. Nio, ‘The tenth-order QED contribution to the lepton $g-2$: Evaluation of dominant $\alpha^5$ terms of muon $g-2$,’ [Phys. Rev. D]{} [73]{} (2006) 053007 \[arXiv:hep-ph/0512330\]; K. Hagiwara, A. D. Martin, D. Nomura and T. Teubner, ‘Improved predictions for $g-2$ of the muon and $\alpha_{QED}(M_Z^2)$,’ [Phys. Lett. B]{} [649]{} (2007) 173 \[arXiv:hep-ph/0611102\]. G. Belanger, F. Boudjema, A. Pukhov and A. Semenov, “micrOMEGAs 2.0.7: A program to calculate the relic density of dark matter in a generic model”, Comput. Phys. Commun. [177]{} (2007) 894; “micrOMEGAs: A program for calculating the relic density in the MSSM”, [Comput. Phys. Commun.]{} [149]{} (2002) 103 \[arXiv:hep-ph/0112278\]; “MicrOMEGAs: Version 1.3”, [Comput. Phys. Commun.]{} [174]{} (2006) 577 \[arXiv:hep-ph/0405253\]. J. Dunkley [et al.]{} \[WMAP Collaboration\], “Five-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Likelihoods and Parameters from the WMAP data,” arXiv:0803.0586 \[astro-ph\]. T. T. E. Group [et al.]{} \[CDF Collaboration\], “A Combination of CDF and D0 Results on the Mass of the Top Quark,” arXiv:0803.1683 \[hep-ex\]. J. A. Casas, A. Lleyda and C. Muñoz, ‘Strong constraints on the parameter space of the MSSM from charge and colour breaking minima’, [Nucl. Phys. B]{} [471]{} (1996) 3 \[arXiv:hep-ph/9601357\]. D. G. Cerdeño, E. Gabrielli, M. E. Gómez and C. Muñoz, “Neutralino nucleon cross section and charge and colour breaking constraints,” JHEP [0306]{} (2003) 030 \[arXiv:hep-ph/0304115\]. J. A. Casas, A. Ibarra and C. Muñoz, “Phenomenological viability of string and M-theory scenarios,” Nucl. Phys. B [554]{} (1999) 67 \[arXiv:hep-ph/9810266\]; A. Ibarra, “Charge and color breaking and D-terms in string theory,” JHEP [0201]{}, 003 (2002) \[arXiv:hep-ph/0111085\]. D. G. Cerdeño, T. Kobayashi and C. Muñoz, “Prospects for the direct detection of neutralino dark matter in orbifold scenarios,” JHEP [0801]{} (2008) 009 \[arXiv:0709.0858 \[hep-ph\]\]. K. Griest and D. Seckel, “Three exceptions in the calculation of relic abundances,” Phys. Rev. D [43]{} (1991) 3191. M. Wijnholt, “F-theory and unification,” Fortsch. Phys. [58]{} (2010) 846. W. Porod, “SPheno, a program for calculating supersymmetric spectra, SUSY particle decays and SUSY particle production at e+ e- colliders,” Comput. Phys. Commun. [153]{} (2003) 275 \[arXiv:hep-ph/0301101\]. W. Porod and F. Staub, “SPheno 3.1: Extensions including flavour, CP-phases and models beyond the MSSM,” arXiv:1104.1573 \[hep-ph\]. M. Lancaster \[Tevatron Electroweak Working Group and for the CDF and D0 Collaborations\], “Combination of CDF and D0 results on the mass of the top quark using up to 5.8 fb-1 of data,” arXiv:1107.5255 \[hep-ex\]. G. Belanger, F. Boudjema, P. Brun, A. Pukhov, S. Rosier-Lees, P. Salati and A. Semenov, “Indirect search for dark matter with micrOMEGAs2.4,” Comput. Phys. Commun. [182]{} (2011) 842 \[arXiv:1004.1092 \[hep-ph\]\]. “Search for the rare decay $B^{0}_{s}\to \mu^{+}\mu^{-}$ at the LHC with the CMS and LHCb experimentsCombination of LHC results of the search for $B_s\to\mu^+\mu^-$ decays,” LHCb-CONF-2011-047,CMS-PAS-BPH-11-019. R. Aaij [et al.]{} \[LHCb Collaboration\], “Strong constraints on the rare decays $B_s \to \mu^+ \mu^-$ and $B^0 \to \mu^+ \mu^-$,” arXiv:1203.4493 \[hep-ex\]. S. Chatrchyan [et al.]{} \[CMS Collaboration\], “Search for B(s) and B to dimuon decays in pp collisions at 7 TeV,” Phys. Rev. Lett. [107]{}, 191802 (2011) \[arXiv:1107.5834 \[hep-ex\]\]. R. Aaij [et al.]{} \[LHCb Collaboration\], “Search for the rare decays $B_{s}\to \mu^{+}\mu^{-}$ with 300 pb$^-1$ at LHCb", LHCb-CONF-2011-037. A. J. Buras, M. V. Carlucci, S. Gori and G. Isidori, “Higgs-mediated FCNCs: Natural Flavour Conservation vs. Minimal Flavour Violation,” JHEP [1010]{} (2010) 009 \[arXiv:1005.5310 \[hep-ph\]\]. A. J. Buras, “Minimal flavour violation and beyond: Towards a flavour code for short distance dynamics,” Acta Phys. Polon. B [41]{} (2010) 2487 \[arXiv:1012.1447 \[hep-ph\]\]. H. E. Haber and R. Hempfling, “Can the mass of the lightest Higgs boson of the minimal supersymmetric model be larger than m(Z)?,” Phys. Rev. Lett. [66]{} (1991) 1815;\ J. R. Ellis, G. Ridolfi and F. Zwirner, “Radiative corrections to the masses of supersymmetric Higgs bosons,” Phys. Lett. B [257]{} (1991) 83;\ Y. Okada, M. Yamaguchi and T. Yanagida, “Upper bound of the lightest Higgs boson mass in the minimal supersymmetric standard model,” Prog. Theor. Phys. [85]{} (1991) 1. S. Heinemeyer, W. Hollik and G. Weiglein, “FeynHiggs: A Program for the calculation of the masses of the neutral CP even Higgs bosons in the MSSM,” Comput. Phys. Commun. [124]{} (2000) 76 \[hep-ph/9812320\]. T. Hahn, W. Hollik, S. Heinemeyer and G. Weiglein, “Precision Higgs masses with FeynHiggs 2.2,” eConf C [050318]{} (2005) 0106 \[hep-ph/0507009\]. J. Cao, Z. Heng, D. Li and J. M. Yang, “Current experimental constraints on the lightest Higgs boson mass in the constrained MSSM,” arXiv:1112.4391 \[hep-ph\]. S. Heinemeyer, W. Hollik and G. Weiglein, “Constraints on tan Beta in the MSSM from the upper bound on the mass of the lightest Higgs boson,” JHEP [0006]{} (2000) 009 \[hep-ph/9909540\]. G. Degrassi, S. Heinemeyer, W. Hollik, P. Slavich and G. Weiglein, “Towards high precision predictions for the MSSM Higgs sector,” Eur. Phys. J. C [28]{} (2003) 133 \[hep-ph/0212020\]. B. C. Allanach, A. Djouadi, J. L. Kneur, W. Porod and P. Slavich, “Precise determination of the neutral Higgs boson masses in the MSSM,” JHEP [0409]{} (2004) 044 \[hep-ph/0406166\]. G. W. Bennett [et al.]{} \[ Muon G-2 Collaboration \], “Final Report of the Muon E821 Anomalous Magnetic Moment Measurement at BNL,” Phys. Rev. [D73 ]{} (2006) 072003. \[hep-ex/0602035\]. K. Hagiwara, R. Liao, A. D. Martin, D. Nomura and T. Teubner, “$(g-2)_\mu$ and $\alpha(M_{Z^2})$ re-evaluated using new precise data,” J. Phys. G G [38]{} (2011) 085003 \[arXiv:1105.3149 \[hep-ph\]\]. F. Jegerlehner and A. Nyffeler, “The Muon g-2,” Phys. Rept. [477]{} (2009) 1 \[arXiv:0902.3360 \[hep-ph\]\]. M. Davier, A. Hoecker, B. Malaescu and Z. Zhang, “Reevaluation of the Hadronic Contributions to the Muon g-2 and to alpha(MZ),” Eur. Phys. J. C [71]{} (2011) 1515 \[arXiv:1010.4180 \[hep-ph\]\]. D. Feldman, Z. Liu and P. Nath, “Sparticles at the LHC,” JHEP [0804]{}, 054 (2008) \[arXiv:0802.4085 \[hep-ph\]\]. Z. Ahmed [et al.]{} \[The CDMS-II Collaboration\], “Dark Matter Search Results from the CDMS II Experiment,” Science [327]{} (2010) 1619 \[arXiv:0912.3592 \[astro-ph.CO\]\]. Z. Ahmed [et al.]{} \[CDMS and EDELWEISS Collaborations\], “Combined Limits on WIMPs from the CDMS and EDELWEISS Experiments,” Phys. Rev. D [84]{} (2011) 011102 \[arXiv:1105.3377 \[astro-ph.CO\]\]. E. Aprile [et al.]{} \[XENON100 Collaboration\], “Dark Matter Results from 100 Live Days of XENON100 Data,” Phys. Rev. Lett. [107]{} (2011) 131302 \[arXiv:1104.2549 \[astro-ph.CO\]\]. T. Sjostrand, S. Mrenna and P. Z. Skands, “PYTHIA 6.4 Physics and Manual,” JHEP [0605]{} (2006) 026 \[hep-ph/0603175\]. http://www.physics.ucdavis.edu/ conway/research/software/pgs/pgs.html S. Jadach, Z. Was, R. Decker and J. H. Kuhn, “The tau decay library TAUOLA: Version 2.4,” Comput. Phys. Commun. [76]{} (1993) 361. \[CMS Collaboration\] Search for supersymmetry in all-hadronic events with missing energy, CMS PAS SUS-11-004. G. Aad [et al.]{} \[ATLAS Collaboration\], “Search for squarks and gluinos using final states with jets and missing transverse momentum with the ATLAS detector in sqrt(s) = 7 TeV proton-proton collisions,” arXiv:1109.6572 \[hep-ex\]. C. F. Berger, Z. Bern, L. J. Dixon, F. Febres Cordero, D. Forde, T. Gleisberg, H. Ita and D. A. Kosower [et al.]{}, “Next-to-Leading Order QCD Predictions for W+3-Jet Distributions at Hadron Colliders,” Phys. Rev. D [80]{} (2009) 074036 \[arXiv:0907.1984 \[hep-ph\]\]. K. Melnikov and G. Zanderighi, “W+3 jet production at the LHC as a signal or background,” Phys. Rev. D [81]{} (2010) 074025 \[arXiv:0910.3671 \[hep-ph\]\]. G. Aad [et al.]{} \[ATLAS Collaboration\], “Measurement of the production cross section for W-bosons in association with jets in pp collisions at sqrt(s) = 7 TeV with the ATLAS detector,” Phys. Lett. B [698]{}, 325 (2011) \[arXiv:1012.5382 \[hep-ex\]\]. N. Kidonakis, “Next-to-next-to-leading soft-gluon corrections for the top quark cross section and transverse momentum distribution,” Phys. Rev. D [82]{} (2010) 114030 \[arXiv:1009.4935 \[hep-ph\]\]. J. M. Campbell, R. K. Ellis and C. Williams, “Vector boson pair production at the LHC,” JHEP [1107]{}, 018 (2011) \[arXiv:1105.0020 \[hep-ph\]\]. W. Beenakker, R. Hopker and M. Spira, “PROSPINO: A Program for the production of supersymmetric particles in next-to-leading order QCD,” hep-ph/9611232. R. L. Arnowitt, B. Dutta and Y. Santoso, “Coannihilation effects in supergravity and D-brane models,” Nucl. Phys. B [606]{} (2001) 59 \[hep-ph/0102181\]. R. L. Arnowitt, B. Dutta, T. Kamon, N. Kolev and D. A. Toback, “Detection of SUSY in the stau-neutralino coannihilation region at the LHC,” Phys. Lett. B [639]{} (2006) 46 \[hep-ph/0603128\]. D. Feldman, Z. Liu and P. Nath, “The Landscape of Sparticle Mass Hierarchies and Their Signature Space at the LHC,” Phys. Rev. Lett. [99]{} (2007) 251802 \[Erratum-ibid. [100]{} (2008) 069902\] \[arXiv:0707.1873 \[hep-ph\]\]. T. Jittoh, J. Sato, T. Shimomura and M. Yamanaka, “Long life stau in the minimal supersymmetric standard model,” Phys. Rev. D [73]{} (2006) 055009 \[hep-ph/0512197\]. M. Pospelov, “Particle physics catalysis of thermal Big Bang Nucleosynthesis,” Phys. Rev. Lett. [98]{} (2007) 231301 \[hep-ph/0605215\]. V. Khachatryan [et al.]{} \[CMS Collaboration\], “Search for Heavy Stable Charged Particles in pp collisions at sqrt(s)=7 TeV,” JHEP [1103]{} (2011) 024 \[arXiv:1101.1645 \[hep-ex\]\]. G. Aad [et al.]{} \[ATLAS Collaboration\], “Search for Heavy Long-Lived Charged Particles with the ATLAS detector in pp collisions at sqrt(s) = 7 TeV,” Phys. Lett. B [703]{} (2011) 428 \[arXiv:1106.4495 \[hep-ex\]\]. S. A. Abel and M. D. Goodsell, [“Realistic Yukawa couplings through instantons in intersecting brane worlds,”]{} JHEP [0710]{}, 034 (2007) \[arXiv:hep-th/0612110\]. A. Font and L. E. Ibáñez, [“Yukawa Structure from U(1) Fluxes in F-theory Grand Unification,”]{} JHEP [0902]{}, 016 (2009) \[arXiv:0811.2157 \[hep-th\]\]. J. J. Heckman and C. Vafa, [“Flavor Hierarchy From F-theory,”]{} Nucl. Phys. B [837]{} (2010) 137 \[arXiv:0811.2417 \[hep-th\]\]. H. Hayashi, T. Kawano, R. Tatar and T. Watari, [“Codimension-3 Singularities and Yukawa Couplings in F-theory,”]{} Nucl. Phys. B [823]{}, 47 (2009) \[arXiv:0901.4941 \[hep-th\]\]. A. Font and L. E. Ibáñez, [“Matter wave functions and Yukawa couplings in F-theory Grand Unification,”]{} JHEP [0909]{}, 036 (2009) \[arXiv:0907.4895 \[hep-th\]\]. S. Cecotti, M. C. N. Cheng, J. J. Heckman and C. Vafa, [“Yukawa Couplings in F-theory and Non-Commutative Geometry,”]{} arXiv:0910.0477 \[hep-th\]. J. P. Conlon and E. Palti, [“Aspects of Flavour and Supersymmetry in F-theory GUTs,”]{} JHEP [1001]{}, 029 (2010) \[arXiv:0910.2413 \[hep-th\]\]. H. Hayashi, T. Kawano, Y. Tsuchiya and T. Watari, [“Flavor Structure in F-theory Compactifications,”]{} JHEP [1008]{}, 036 (2010) \[arXiv:0910.2762 \[hep-th\]\]. G. K. Leontaris and G. G. Ross, “Yukawa couplings and fermion mass structure in F-theory GUTs,” JHEP [1102]{} (2011) 108 \[arXiv:1009.6000 \[hep-th\]\]. E. Dudas and E. Palti, “Froggatt-Nielsen models from E(8) in F-theory GUTs,” JHEP [1001]{} (2010) 127 \[arXiv:0912.0853 \[hep-th\]\]. F. Marchesano and L. Martucci, [“Non-perturbative effects on seven-brane Yukawa couplings,”]{} Phys. Rev. Lett. [104]{}, 231601 (2010) \[arXiv:0910.5496 \[hep-th\]\]. P. Koerber and L. Martucci, [“From ten to four and back again: how to generalize the geometry,”]{} JHEP [0708]{}, 059 (2007) \[arXiv:0707.1038 \[hep-th\]\]. D. Baumann, A. Dymarsky, S. Kachru, I. R. Klebanov and L. McAllister, [“D3-brane Potentials from Fluxes in AdS/CFT,”]{} JHEP [1006]{}, 072 (2010) \[arXiv:1001.5028 \[hep-th\]\]. H. Abe, T. Kobayashi and H. Ohki, [“Magnetized orbifold models,”]{} JHEP [0809]{}, 043 (2008) \[arXiv:0806.4748 \[hep-th\]\].\ H. Abe, K. S. Choi, T. Kobayashi and H. Ohki, [“Three generation magnetized orbifold models,”]{} Nucl. Phys. B [814]{}, 265 (2009) \[arXiv:0812.3534 \[hep-th\]\]. J. P. Conlon, A. Maharana and F. Quevedo, [“Wave Functions and Yukawa Couplings in Local String Compactifications,”]{} JHEP [0809]{}, 104 (2008) \[arXiv:0807.0789 \[hep-th\]\]. P. Di Vecchia, A. Liccardo, R. Marotta, F. Pezzella, [“Kähler Metrics and Yukawa Couplings in Magnetized Brane Models,”]{} JHEP [0903 ]{} (2009) 029. \[arXiv:0810.5509 \[hep-th\]\]. P. G. Cámara and F. Marchesano, [“Open string wavefunctions in flux compactifications,”]{} JHEP [0910]{}, 017 (2009) \[arXiv:0906.3033 \[hep-th\]\]. [“Physics from open string wavefunctions,”]{} PoS E [PS-HEP2009]{}, 390 (2009). R. C. Myers, [“Dielectric branes,”]{} JHEP [9912]{}, 022 (1999) \[arXiv:hep-th/9910053\]. H. Fritzsch and Z. Z. Xing, [“Mass and flavor mixing schemes of quarks and leptons,”]{} Prog. Part. Nucl. Phys. [45]{}, 1 (2000) \[arXiv:hep-ph/9912358\]. M. Cvetic, G. Shiu and A. M. Uranga, [“Three-family supersymmetric standard like models from intersecting branes ”]{}, Phys. Rev. Lett. [87]{}, 201801 (2001), hep-th/0107143. [“Chiral type II orientifold constructions as M theory on G(2) holonomy spaces ”]{}, hep-th/0111179. R. Blumenhagen, D. Lust and T. R. Taylor, “Moduli stabilization in chiral type IIB orientifold models with fluxes,” Nucl. Phys. B [663]{}, 319 (2003) \[arXiv:hep-th/0303016\]. J. F. G. Cascales and A. M. Uranga, “Chiral 4d N = 1 string vacua with D-branes and NSNS and RR fluxes,” JHEP [0305]{}, 011 (2003) \[arXiv:hep-th/0303024\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a detailed stellar clustering analysis with the application of the two-point correlation function on distinct young stellar ensembles. Our aim is to understand how stellar systems are assembled at the earliest stages of their formation. Our object of interest is the star-forming region NGC346 in the Small Magellanic Cloud. It is a young stellar system well-revealed from its natal environment, comprising complete samples of pre–main-sequence and upper main-sequence stars, very close to their formation. We apply a comprehensive characterization of the autocorrelation function for both centrally condensed stellar clusters and self-similar stellar distributions through numerical simulations of stellar ensembles. We interpret the observed autocorrelation function of NGC346 on the basis of these simulations. We find that it can be best explained as the combination of two distinct stellar clustering designs, a centrally concentrated, dominant at the central part of the star-forming region, and an extended self-similar distribution of stars across the complete observed field. The cluster component, similar to non-truncated young star clusters, is determined to have a core radius of $\sim$2.5pc and a density profile index of $\sim $2.3. The extended fractal component is found with our simulations to have a fractal dimension of $\sim$2.3, identical to that found for the interstellar medium, in agreement to hierarchy induced by turbulence. This suggests that the stellar clustering at a time very near to birth behaves in a complex manner. It is the combined result of the star formation process regulated by turbulence and the early dynamical evolution induced by the gravitational potential of condensed stellar clusters.' date: To be submitted to MNRAS title: \| The Complex Distribution of Recently Formed Stars.\ Bimodal Stellar Clustering in the Star-Forming Region NGC346. --- \[firstpage\] Magellanic Clouds – stars: pre-main-sequence – stars: statistics – [Hii]{} Regions – ISM: individual objects: LHA115-N66 – open clusters and associations: individual: NGC346. Introduction ============ It is generally accepted that stars form in groups of various sizes and characteristics [@ladalada03], starting with small compact concentrations of protostars embedded in star-forming regions and moving up in [length-scale]{} to large extended loose aggregates of young stars stars. It is, however, suggested that these diverse stellar assembles are not independent from each other, but tightly connected through the star formation process [@elmegreen11]. Small, dense proto-clusters coexist in a symbiotic fashion with larger, less dense subgroups of OB-type stars, which in turn reside in even larger, looser stellar associations, each of these types of objects representing a different [time-scale]{} of star formation within one molecular cloud [@efremovelmegreen98]. This picture of clustered star formation at various length- and time-scales is not always clear in our observations, as e.g., in the massive star formation environments of star-burst clusters. Studies of newly formed stellar systems can identify the conditions that may favor [multiple]{} over [single]{} cluster formation events. Different scenarios for star formation predict different observable properties for the resulting stellar systems. The ‘quiescent’ star formation scenario [e.g. @krumholztan07] predicts large age-spreads among young stars in the same molecular cloud, while according to the ‘competitive accretion’ scenario [e.g. @clark07] star formation is a process of short time-scale, leading to clusters in a variety of forms. The investigation of the clustering of stars at the time of their formation can provide important information on the nature of the star formation process itself [e.g., @schmeja08], and place constraints to the suggested theories. A single compact stellar cluster, embedded in its own H[ii]{} region, would suggest a local monolithic episode of star formation in a dense environment. However, such clusters are rarely found in isolation; they are the densest stellar concentrations of larger stellar structures as seen in dwarf and spiral galaxies [@efremov09; @karampelas09]. In our own Galaxy multiple (or fractured) clusterings of stars in large star-forming regions, are also more commonly observed [@feigelson11; @megeath12]. This clustering behavior favors the existence of multiple processes, such as feedback, controlling star formation, and being active on different scales. Giant Molecular Clouds are hierarchical structures [@elmegreen96; @stutzki98], indicating that scale-free processes determine their global morphology. Turbulence is being widely accepted as the dominant among these processes [@maclowklessen04; @elmegreenscalo04]. The goal of this paper is to establish whether the newly formed stars follow a similar, hierarchical distribution, which may indicate that turbulence also determines the clustering behavior of stars at the time of their formation. This investigation requires a complete census of stars, covering a high dynamic range in masses, distributed over the typical length-scale of giant molecular clouds. Our target of interest is the star-forming complex NGC346, the brightest [Hii]{} region in the Small Magellanic Cloud [LHA115-N66; @henize56]. [Hubble]{} Space Telescope imaging of such regions in the Magellanic Clouds (MCs) provides an unprecedented access to their newly-born stellar populations down to the sub-solar regime over large areas of the sky. Observed by [Hubble]{} (GO Program 10248; PI: A. Nota), this region satisfies, thus, the main observational criteria for our analysis: 1) Observed on size-scales relevant for molecular clouds (50-100pc); 2) High angular resolution ($\sim$0.125); 3) Large number of detected members, covering a significant fraction of the stellar Initial Mass Function, complete to $\sim$0.5 M. Our dataset, obtained with the Advanced Camera for Surveys (ACS), is described in [@gouliermis06]. A rich sample of more than 98,000 stars was detected down to $m_{\rm 555} \simeq 27$ mag with [ -2.truept]{} 50% completeness. It comprises a mixture of stellar generations, with 60% of the stars formed [ -2.truept]{}5Gyr ago [@cignoni11]. The young populations in NGC346 consist mainly of low-mass pre–main-sequence (PMS) stars, identified from their positions on the color-magnitude diagram . It is decidedly troublesome to determine an age for a cluster based on its PMS stars [@jeffries12; @preibisch12]. Nevertheless, an isochronal age of $\sim$3Myr has been established for NGC346 by [@sabbi08] by comparison with PMS evolutionary models. We complete the sample of low-mass PMS stars with the upper–main-sequence (UMS) stars ($m_{\rm 555}-m_{\rm 814} \leq$0.0mag; 12[ -2.truept]{}$m_{\rm 555}$[ -2.truept]{}17mag; ages [ -2.truept]{}10Myr), compiling a total sample of 5,150 stars. The UMS stars correspond to about 7% of the sample. The map of our stellar inventory is shown in Figure\[f:map\]. ![Map of the stars selected for our analysis, i.e., low-mass PMS and more massive upper–main-sequence (UMS) stars, as found in both F555W and F814W filters with [dolphot]{} photometry [@dolphin00] based on ACS/WFC imaging of three pointings on the field of NGC346. Red dots correspond to the UMS stars. The map covers the whole observed area. North is up, and east is to the left. Coordinates are given in seconds of arc in respect to a reference point, and are determined based on the drizzled ACS image in the F814W filter.[]{data-label="f:map"}](map-arcsec-reduced+fields.eps){width="\columnwidth"} In a previous study [@schmeja09 from here on PaperI] we applied a cluster analysis based on the nearest-neighbor density method and we identified ten individual PMS stellar clusters in the region. We established that NGC346 is a multi-clustered environment. This work also provided evidence of hierarchy in the PMS stellar clustering from a graph theory study with the minimum spanning tree and the analysis of the $\cal{Q}$ parameter [@cw04]. While these methods provided a unprecedented insight of the NGC346 clustering, they were not able to quantitatively describe its complexity, or to characterize its self-similar behavior. In addition, the accuracy of the $\cal{Q}$ parameter in interpreting a fractal structure has been challenged on the basis of the effect of projection in elongated star clusters (see @bastian09 and @cw09 for different accessions and viewpoints to the problem). The $\cal{Q}$ parameter for the whole complex of NGC346 was found equal to about 0.8, and thus cannot be used to conclude about the nature of stellar clustering in this region. We revisit the question of stellar clustering in NGC346 with the application of a thorough cluster analysis. We first assess the topology of young stellar clustering in the region with the kernel density estimation technique, and the distribution function of stellar separations. We then decipher the clustering behavior of young stars in NGC346 with the construction of their observed two-point correlation function, i.e., their autocorrelation function and its comparison to those from a series of simulated stellar distributions. We explore, thus, the limitations of this method and provide an accurate interpretation of the observed autocorrelation function. We, thus, carefully characterize the complex clustering behavior of young stars in NGC346. The paper is organized as follows. In Section\[s:tpcf\] we present the stellar surface density map of NGC346, and the distribution function of stellar separations. In Section\[s:acf\] we present the autocorrelation function and discuss the first observable evidence that NGC346 contains (at least) two components with very distinct distributions. We compare the autocorrelation function of NGC346 with simulations of centrally condensed and self-similar stellar distributions in Section\[s:acfint\]. We show that indeed the stellar distribution in the region is the result of two individual stellar components, a central condensed and an extended fractal distribution. In the same section we also constrain the basic parameters of these components using dedicated simulations of mixed stellar distributions. Finally, in Section\[s:sum\] we summarize our results and discuss their implications to our understanding of clustered star formation. Concluding remarks of our study are given in Section\[s:concl\]. Additionally, in Appendix\[s:appendix\] we present our library of simulated autocorrelation functions to be used for the interpretation of that observed in NGC346, and in Appendix\[d3toeta\] we provide an empirical calibration between three-dimensional and two-dimensional fractal dimensions over a wide range of values. The Spatial Distribution of Young Stars {#s:tpcf} ======================================= ![The surface stellar density (significance) map constructed from the sample of young stars in the region of NGC346 with the use of the KDE technique and a 5 Gaussian kernel. Isopleths at different density levels, drawn with different colors, signify the individual stellar clusterings identified in this region. The color-bar indicate different density levels in $\sigma$, the standard deviation of the measured stellar density in the whole region. North is up and East is to the left. Coordinates are given as in Figure\[f:map\].[]{data-label="f:kdemap"}](ngc346_vvumspms_scmap_ckd.eps){width="\columnwidth"} The Kernel Density Estimation Map {#s:kde} --------------------------------- We construct the surface density map using the [kernel density estimation]{} (KDE) technique [@Silverman92]. This technique smooths individual data point locations with Gaussian kernels to form a continuous spatial distribution. The fundamental parameter is the full-width-at-half-maximum (FWHM) of the kernel. For the purposes of our study the selection of this parameter was based on the desired ‘resolution’, i.e., the smallest stellar structure to be identified. Based on the sizes of the detected stellar clusterings in PaperI, we optimize the construction of the surface density map with a FWHM $\sim$5 (equivalent to $\sim$1.5pc)[^1]. The produced KDE density map of NGC346 is shown in Figure\[f:kdemap\]. In this map small stellar clusterings, coinciding with the stellar groupings previously detected using the nearest-neighbor method, appear as over-densities at density levels of [ -2.truept]{}2$\sigma$. However, all these clusterings are part of a large stellar structure at the 1$\sigma$ density level. At the lowest, $\sim1\sigma$ significance level, two large loose stellar over-densities are apparent in the KDE map, which coincide with features that can also be discerned in ionized gas and PAH emission: 1) The central “bar” of the star-forming region, i.e., the bright emission region, extending from southeast to northwest [@rubio00], and 2) the “northern arc-like arm” identified in mid-infrared wavebands to extend from the center of the region to the center-northeast part of the field [@gouliermis08]. The KDE map of the region depicts, similarly to the nearest-neighbor map, that the region of NGC346 includes several separate clusterings, defined by the 2$\sigma$ and 3$\sigma$ isopleths, characterized in previous studies as individual sub-clusters [@sabbi07 PaperI]. However, the KDE map illustrates in a far more meticulous manner another critical characteristic of the stellar clustering in NGC346. There are several compact stellar clumps, which do not appear to be independent, but they emerge as small over-densities within the NGC346 bar and northern arm. The bar itself is revealed as a large stellar structure (aggregate) in the KDE map at 2$\sigma$ significance. Stellar distribution within this aggregate is organized in a segregated fashion with the small compact clusterings surrounding a central massive cluster, which appears circular at $\sim$4$\sigma$ significance and higher (Figure\[f:kdemap\]). Whether there is any hierarchy in the manner the aggregate is assembled, whether the central cluster is indeed centrally concentrated and how all the over-densities are connected to each other within the same large structure, are the questions we intend to explore with our analysis. ![Distribution function for separations between the young stars in our sample for NGC346 (blue line). The separations distribution $p(r)$ for a simulated field of stars with uniform surface-density (i.e., random distribution; green-dashed line), and that for a centrally concentrated cluster with density gradient $\propto r^{-1.7}$ (red line) are plotted for comparison. The construction of both these artificial stellar distributions are discussed in detail in Appendix\[s:appendix\]. The separations distribution for NGC346 shows multiple maxima, and its shape is very close to that for the centrally concentrated cluster (see Section\[s:346ascluster\], for a discussion on this similarity).[]{data-label="f:pdf"}](separations_distributions_plot.eps){width="\columnwidth"} Stellar Separations Distribution Function {#s:pdf} ----------------------------------------- We derive the probability distribution function of stellar separations for the stars in our sample, following the recipe of [@cw04]. The probability function, $p_i(r_j)$, for each star $i$ is calculated as the number of pair separations $N_{ij}$ that fall in the separation bin centered on $r_j$ divided by the total number of separations: $$\begin{aligned} \label{eq:pdf} p_i(r_j) & = & \frac{2\,N_{ij}}{N\,(N-1)\,dr} \,,\end{aligned}$$ where $N$ is the total number of stars. The distribution function $p(r)$ is calculated in every separation bin from the sum of $p_i(r_j)$: $$\begin{aligned} p(r_j) & = & \sum_{i=1}^{N}p_i(r_j).\end{aligned}$$ The probability that the projected separation between two randomly chosen stars is in the interval $(r, r+dr)$ is given by $p(r)dr$. ![image](ngc346_density_profile+model_v2.eps){width="\textwidth"} The separations distribution function for the young stellar population in NGC346 is presented in Figure\[f:pdf\] (blue line). We also show the stellar separations distribution for a uniform (not clustered) stellar field (green-dashed line), and that for a simulated centrally concentrated cluster (red line). Details on the simulations are discussed in Appendix\[s:appendix\]. Perusal of Figure\[f:pdf\] shows that overall the separation distribution compares quite well with that of the centrally concentrated cluster with surface density profile $n \propto r^{-1.7}$. However, this agreement is surprising, since evidently both the observed stellar chart (Figure\[f:map\]) and surface density map (Figure\[f:kdemap\]) exhibit more structures than can be explained by a simple spherical cluster. This similarity is further discussed in terms of the radial stellar density profile in Section\[s:346ascluster\]. The observed probability distribution also displays a number of local maxima, which indicates the existence of some preferred length-scales in NGC346, perhaps resulting from multiple clustering in the region. The clearest local maxima occur at separations of about 50and 100($\sim$14pc and 28pc respectively). The smaller of these scales coincides with the size of a central stellar concentration, as revealed in the KDE map of Figure\[f:map\] at 3$\sigma$ significance. The maximum at the larger length-scale concurs with the average size of the stellar aggregate in the NGC346 bar (at 2$\sigma$ in the KDE map). Describing NGC346 as a single condensed cluster {#s:346ascluster} ----------------------------------------------- Based on the stellar separation distribution function, presented in the previous section, one might be tempted to interpret the entire stellar distribution of NGC346 as originating from a single, centrally concentrated cluster, in spite of the obvious asymmetries in the stellar map. We performed a typical analysis by constructing the radial surface density profile of young stars in the entire observed area, centered on the peak surface density. This profile is shown in blue in Figure\[f:densprof\] (left panel). We then fitted to the observed profile a model cluster density profile with the functional form prescribed by @eff87 [see Appendix\[s:effclusters\]]. The best-fitting model is also shown in Figure\[f:densprof\] (left) in red. Considering that NGC346 is certainly not a single condensed cluster, its radial profile was surprisingly well reproduced by models for clusters with core radii of $r_c \simeq 14.9\arcsec \pm 0.7\arcsec$, and density profile slopes of $\gamma \simeq 1.69 \pm 0.06$. Such a model cluster is used for comparing the stellar separations distributions shown in Figure\[f:pdf\]. This agreement shows that caution is warranted when interpreting radial density profiles, as well as stellar separations distribution functions, in particular in cases where the observations do not allow to recognize prominent asymmetries. It demonstrates the need for a diagnostic tool for stellar distributions, which is equally sensitive to structure afar from the regions with the highest stellar density. To this end we use the two point correlation (autocorrelation) function. The diagnostic power of this function is demonstrated in Figure\[f:densprof\], right panel, which shows that the best-fit cluster distribution fails completely to reproduce the observed autocorrelation function (see Section\[s:onecluster\]), in spite of the excellent fit to the radial profile (Figure\[f:densprof\] left) and stellar separations distribution function (Figure\[f:pdf\]). The Autocorrelation Function of Young Stars in NGC346 {#s:acf} ===================================================== The degree of clustering of stars can be quantified by using the two-point correlation function [@peebles80 Section45]. Applied to stars in the same sample, this function becomes an [autocorrelation function]{} (ACF). In the following, we broadly follow the method introduced by [@peebles80] for cosmological applications and modified by [@gomez93] for characterizing the clustering behavior of TTauri stars in Galactic star-forming regions. Other recent investigations applied this method to observed samples of star clusters in remote galaxies [e.g., @zhang01; @scheepmaker09 in the Antennae and M51 galaxies respectively]. The ACF is defined as: $$\begin{aligned} \label{eq:autocorrelation} 1+\xi (r) & = & \frac{1}{\bar{n}N}\sum_{i=1}^{N} n_{i}(r),\end{aligned}$$ where $n_{i} (r)$ is the number density of stars found in an aperture of radius $r$ centered on, but excluding star $i$. $N$ is the total number of stars and $\bar{n}$ is the average stellar number density. The corresponding uncertainties based on [@peebles80] are given by: $$\begin{aligned} \label{eq:acferr} \delta(r) & = & \sqrt{N} \cdot\left(\frac{1}{2}\sum_{i=1}^{N} n_{\mathrm{p}}(r)\right)^{-1/2}, \end{aligned}$$ where $n_{\mathrm{p}}(r)$ is the number of pairs formed with the central star $i$ of the current aperture, and the factor $1/2$ accounts for not counting every pair twice. In general, $\xi (r)$ is defined such that $\bar{n}[1+\xi(r)] d^2r$ is the probability of finding a neighboring star in a area of radius $r$ from a random star in the sample. This means in effect that $1+\xi (r)$ is a measure for the mean surface density within radius $r$ from a star, divided by the mean surface density of the total sample (i.e. the surface density enhancement within radius $r$ with respect to the global average). Therefore, for a random stellar distribution $1+\xi(r) = 1$, while for a clustered distribution $1+\xi(r) > 1$. For a hierarchical, or fractal, distribution of stars the ACF yields a power-law dependency with radius of the form $1+\xi (r) \propto r^{\eta}$ [@gomez93]. For such a distribution, the total number of stars $N$ within an aperture of radius $r$ increases as $N\propto r^{\eta}\cdot r^{2} = r^{\eta+2}$. The power-law index $\eta$ is related to the two-dimensional fractal dimension $D_{2}$ as $D_{2} = \eta +2$ [@mandelbrot83]. Power-law ACFs have been observed for interstellar gas over a large range of environments, and have been interpreted as indications of hierarchical structuring of the gas. The derived typical (three-dimensional) fractal dimension of $D_{3} \sim2.3$ [e.g. @elmegreen96; @elmegreen01b] apparently comes from the same underlying distribution as found for extragalactic star-forming regions in NGC628 with a two-dimensional fractal dimension of $D_{2}\simeq 1.5$ [@elmegreen06]. However, the conversion $D_{2} = D_{3}-1$ is not generally valid is discussed in [@elmegreen96] and [@elmegreenscalo04]. In Appendix\[d3toeta\] we calibrate this relation over a large range of fractal dimensions. The Effect of Limited Observed Field-of-View {#s:edgeeffect} -------------------------------------------- The ACF, calculated according to Eq.(\[eq:autocorrelation\]), is sensitive to the size of the surveyed area, because apertures around stars close to the edge fall partly outside the survey area, measuring a too low average stellar surface density. This introduces a very steep decrease, dropping well below unity for larger separations due to missing stars outside of the observed field-of-view. This behavior is demonstrated in the ACF of NGC346 shown with a grey line in Figure\[fig:autocorrelation\] (described in Section\[s:acf346\]). The correct way of dealing with this issue is by “masking” the apertures, i.e., by dividing only the part of each aperture that overlaps with the observed field-of-view when calculating average surface densities ($n_{i}$ in Eq.\[eq:autocorrelation\]). With this correction the ACF drops smoothly to the level of the random field for larger separations (blue line in Figure\[fig:autocorrelation\]). From Figure\[fig:autocorrelation\] we assess that the correction for the ACF of NGC346 becomes dominant at separations larger than $\sim$70. We consider this ACF, as well as all ACFs simulated in our analysis, to be reliable for separations up to this limit. ![The ACF of the young stellar population (PMS and UMS stars) in NGC346. The horizontal dashed line corresponds to the ACF of a random field and equals unity. The grey line corresponds to the ACF uncorrected for edge effects, while the blue line depicts the corrected ACF (see discussion in Section\[s:edgeeffect\].[]{data-label="fig:autocorrelation"}](tpcf_young-stars_plot_log.eps){width="\columnwidth"} The Observed ACF of NGC346 {#s:acf346} -------------------------- The ACF for the young stellar population of NGC346, constructed according to Eq.(\[eq:autocorrelation\]), is plotted with respect to projected stellar separations in Figure\[fig:autocorrelation\]. Error bars are determined according to Eq.(\[eq:acferr\]). The clustering of young stars in NGC346 becomes stronger, i.e., larger values of $1+\xi(r)$, at smaller stellar separations. This figure also shows that the stellar clustering in NGC346 changes behavior at different scales. Two distinct parts in the ACF plot can be distinguished at the separation of about 20 ($\sim$5pc). Both parts show an almost linear decrease of the ACF with radial distance, but with significantly different slopes. Each of these power-law dependencies is similar to that expected for a fractal stellar distribution. However, a purely hierarchical stellar distribution exhibits a single-slope increase of $1+\xi$ with smaller separations over all scales unlike the ACF of Figure\[fig:autocorrelation\] that shows a clear break. We verify that the ACF of young stars in NGC346 is well-described by a broken power-law, and we determine its slopes, $\eta$, by fitting such a power-law function. We establish the indexes of the two corresponding linear parts by applying a Levenberg–Marquard nonlinear least square minimization technique [@levenberg44; @marquardt63], as implemented in IDL by [@markwardt09]. The two power-law slopes in the ACF, as well as the position of the break point along the abscissa are the free parameters in our fit. The break in the slope occurs at separations of 20.85 ($\sim$5.8pc). The inner part (r [$\stackrel {<}{_{\sim}}$]{}21) has a power law index $\eta_{\rm in} \simeq -0.18$, which corresponds to a 2D fractal dimension of $D_{2} \simeq 1.8$. For separations 21[$\stackrel {<}{_{\sim}}$]{} r [$\stackrel {<}{_{\sim}}$]{}70 (6pc [$\stackrel {<}{_{\sim}}$]{} r [$\stackrel {<}{_{\sim}}$]{}20pc) the ACF has a power-law index of $\eta_{\rm out} \simeq -0.58$ ($D_{2} \simeq 1.4$). As we discussed earlier, separations beyond the limit of $\sim$70 cannot be considered in our analysis, since at these separations the ACF correction for the finite observed field is dominant. The fractal dimension found for the small stellar separations in NGC346 is quite close to the geometrical 2D dimension, suggesting a smooth stellar distribution at these scales. It is interesting to note that at larger scales the derived smaller fractal dimension suggests a more clumpy distribution of stars. This fractal dimension of $D_{2} \simeq 1.4$ agrees very well with that found by several authors for the interstellar gas in the Milky Way and other galaxies [e.g., @falgarone91; @westpfahl99; @kim03], as well as with that found for stellar clusters in external galaxies [e.g., @zhang01; @elmegreen06; @scheepmaker09], and for young stars in several dwarf galaxies [@odekon06]. These investigations argue that the derived fractal dimensions originate from the turbulent motions in the ISM. Projected fractal dimensions with values between $D_{2} = 1.3$ and 1.5 are considered to be consistent with analogous measures of the 3D fractal dimension of $D_{3} \sim 2.4$ derived by both fractional Brownian motion simulations [see, e.g., @stutzki98] and from laboratory or numerical turbulent flows [@mandelbrot83; @sreenivasan91; @federrath09]. It should be noted that this argument is valid only when the simple conversion $D_3=D_2+1$ holds. Based on our simulations of self-similar stellar distributions (see Appendix\[s:fractaldistr\]), we show that [the conversion $D_3=D_2+1$ is not generally valid]{} (see Appendix\[d3toeta\]). We also establish an empirical relation between the original $D_3$ used for constructing the distributions and the $D_2$ values derived with the use of the ACF. Our findings on the ACF of NGC346 clearly suggest that stellar clustering at scales larger than $\sim$21 is hierarchically driven by turbulence and thus related to the structure of the ISM. Our simulations of fractal distributions, discussed in Appendix\[s:fractaldistr\], show that indeed hierarchical stellar clustering produces a linear dependency of ACF in log-log, but this dependency extends with the same index $\eta$ across the whole considered length-scales range. Cases where this has been observed are reported by [@gomez93], [@larson95], and [@simon97] in star-forming regions of the Milky Way. The fact that the ACF of Figure\[fig:autocorrelation\] behaves like a broken power-law implies that the stellar distribution in NGC346 is neither a purely hierarchical, nor a pure centrally condensed distribution. In fact, the clear change of slope in the observed ACF provides a strong indication that there are multiple components present, whose distribution is quantitatively and significantly different. The first, high density, component – e.g., a centrally concentrated cluster – affects preferentially the most populated regions of the observed field, while the second, more extended component has significant members throughout the region. This is consistent with the observation that the bar of NGC346 is a loose stellar congregate, with a definitive compact stellar concentration at its center (Figure\[f:kdemap\]). We test the two-components hypothesis by a simple experiment. We repeat our calculation of the ACF while excluding the central compact part of the observed field. Specifically, we mask the stellar sample at significance levels varying between 3 and 8$\sigma$ from the KDE map of NGC346 (Figure\[f:kdemap\]). We find that for values between 5 and 7$\sigma$ the resulting ACF is very close to a single power law, with an exponent of $\sim -0.2$. The derived ACF for masking value of 6$\sigma$ is plotted in red in Figure\[f:cclusmask\]. For higher values of $\sigma$, too many stars in the central compact part are retained and the corresponding ACF still exhibits the characteristic break. This experiment provides more confidence that indeed NGC346 is the conjugated result of two different stellar distributions. The following section is dedicated to obtaining quantitative constraints on the nature of these separate components through numerical simulations. ![The ACF of the young stellar population in NGC346 through its complete extend (in blue) compared to the ACF of the same population with the central compact region masked. The ACF derived after masking the stellar sample at 6$\sigma$ KDE-significance is plotted in red. This ACF exhibits a single slope in contrast to the broken power-law observed for the whole sample. This supports the interpretation that the stellar distribution in NGC346 is governed by two components, a compact and an extended one. The vertical dashed line indicates the separation of 70 beyond which the effect of finite field-of-view becomes dominant.[]{data-label="f:cclusmask"}](ccluster_masking_new4.eps){width="\columnwidth"} Numerical Simulations {#s:acfint} ===================== We derive the nature of the components that dominate the stellar clustering by performing a thorough set of numerical simulations. These simulations are essential for the correct interpretation of observed ACFs, and quantify the uncertainties of the derived clustering parameters. In Appendix\[s:appendix\] we present a library of ACFs on the basis of simulations of centrally condensed and self-similar stellar distributions. These simulations aid the interpretation of the ACF assuming a [single]{} type of clustering (compact or fractal). Here we discuss simulations directly applicable to NGC346, which shows a more complicated clustering behavior with the co-existence of [multiple]{} stellar components. All considered simulations contain the same number of stars as observed in NGC346 in comparable field-of-view. This implies that an applicable simulation should reproduce not only the shape of the observed ACF but also the absolute values. Simulations of a Single Centrally Condensed Cluster {#s:onecluster} --------------------------------------------------- In Section\[s:tpcf\] we found that both the stellar separations distribution function (Figure\[f:pdf\]) and the stellar surface density profile of young stars in NGC346 (Figure\[f:densprof\]) imply that the young stars are mostly distributed in a centrally condensed fashion. Therefore, the simplest possible model to describe the ACF of NGC346 would consist of a single, centrally concentrated cluster. In Appendix\[s:ccclusters\], among the considered types of centrally condensed cluster models, that proposed by @eff87 [from hereon the EFF model] is the most appropriate, since it refers to non-tidally truncated clusters like those found embedded in star-forming regions. We simulated an EFF cluster whose density profile fits best that of NGC346 (shown in Section\[s:346ascluster\]) and constructed its ACF. We compare the ACF of NGC346 to that of this cluster in Figure\[f:densprof\] (right panel). In this figure, it can be seen that while the density profiles fit very well, the corresponding ACFs are strikingly different from each other. At smaller separations the trend of $1+\xi(r)$ cannot be reproduced by the ACF of the EFF cluster, because it is flat. The decrease of $1+\xi$ with separation for larger scales with index $-0.4$ does not reproduce that found for NGC346 ($-0.6$; Section\[s:acf346\]) either. This result highlights the fact that while the surface density profile of a stellar concentration may give the impression of a centrally concentrated cluster, its ACF may clearly indicate that this is not the case. As a consequence the use of surface density profiles for characterizing uncertain stellar distributions (e.g., open clusters and loose stellar groups) should be made with caution, as it may not represent reality. We explored the possibility that other EFF clusters may reproduce the observed ACF, but we found that [all]{} simulated EFF profiles fail to reproduce the main characteristics of the observed ACF. They all produce constant $(1+\xi)$ values at scales smaller than their core-radii, not agreeing with the observed ACF of NGC346 at these scales. In particular, those clusters that had stellar densities at shorter separations compatible to that of NGC346 contained unrealistically large numbers of stars within their core radii. The ACF of the “most successful” set of our simulations of a single centrally condensed cluster is plotted in Figure\[f:wfitacf\] (orange line), where it is shown that the single-cluster scenario cannot reproduce successfully the observed ACF. From this analysis we conclude that while the stellar density map of NGC346 indicates the existence of a condensed stellar cluster, the ACF indicates that not all young stars in NGC346 can belong to the central cluster and thus another, more extended, component should be also considered. This result is additionally supported by the decomposed ACF of NGC346 (Figure\[f:cclusmask\]), constructed by masking the central condensed cluster In the following sections we explore the constraints we can obtain on the nature of this extended component. We identify the type of this component by applying simulations of the combination of two individual stellar distributions. The total number of stars and the field-of-view in these simulations are again identical to the observed ones in NGC346. The basic input parameters in our experiments are (1) the fraction of stars with respect to the total considered number that belong to each component, (2) the core radii and $\gamma$ indexes of each considered condensed cluster, and (3) the fractal dimensions of the simulated self-similar distributions. ![The ‘most successful’ ACFs of stellar distributions, constructed according to the scenarios discussed in Sections\[s:onecluster\]-\[s:oneclusterandfield\]. We show the results for a single cluster (orange line), a double cluster (green line) and a cluster in a homogeneous field (blue line). They all fail to explain the observed ACF (shown with the black line). We demonstrate the spread in these simulations by also showing, in dashed lines, the results after changing the density profile index of the considered clusters by $\gamma = \pm 0.2$. Even the ‘best’ simulations of those that do not include a self-similar stellar component fail to reproduce both the shape and absolute values of the ACF of NGC346.[]{data-label="f:wfitacf"}](tpcf_young-stars_worse-models_v3.eps){width="\columnwidth"} Simulations of two Centrally Condensed Clusters {#s:twocluster} ----------------------------------------------- We study the possibility of reproducing the main characteristics of the observed ACF, assuming two centrally condensed stellar components. This is motivated by the fact that in the stellar surface density (KDE) map of Figure\[f:kdemap\] a small number ($<$4) of isolated density peaks can be seen, depending on the chosen significance level. We produced mock stellar catalogs with the addition of two EFF clusters, with a total number of observed stars equal to that of NGC346 (5,150 stars). We produced a grid of combinations in the stellar distribution by varying the following parameters: 1) Projected separation between the clusters, which varied from 10to 25, 2) the fraction of stars belonging to the first cluster (0.5 to 0.7), 3) the EFF profile indexes $\gamma$ (between 1.6 and 2.2), and 4) the core radii, $r_c$, which was set between 6and 12. The best correspondences between models and observations were found for low values of $\gamma$ ($\sim$1.8) and core radii ($\sim$8-10). Distributing the stars in roughly equal amounts over two clusters allows for somewhat smaller individual clusters, which alleviates to some extent the problem of having too many stars within the core radius – and thus absolute ACF values higher than the observed. This is demonstrated in Figure\[f:wfitacf\], where the most well representative ACF from our simulations of two centrally condensed clusters is shown (in green). If the separation between the clusters is tuned to be roughly consistent with the observed break of about 21, the smaller clusters combined can still yield a weak break at this separation in the ACF, but cannot reproduce the power-law trend of the observed ACF at scales below 15. In general, all of the considered combinations of two EFF clusters completely failed to reproduce the behavior of the ACF of NGC346 in the complete range of separations. Simulations of a Single Cluster plus a Homogeneous Field {#s:oneclusterandfield} -------------------------------------------------------- One of the striking features of the distribution of the young population in NGC346 is that there are stars found distributed throughout the observed region (see also PaperI), which may be an indication of a truly dispersed population, i.e., a (young) background field population. We simulate thus clusters that are located within a homogeneous field. In these simulations we vary the following parameters: 1) The fraction of stars that belong to the EFF cluster ($f_{\rm cl}$: 0.4 to 0.7), 2) the core radius (3-15) and 3) the slope $\gamma$ (1.5-3.0) of the cluster profile. The best agreement with the observed ACF is found for $f_{\rm cl} \simeq 0.5$, core radii between 8and 10and quite high $\gamma$ values from 2.2 to 2.5. For these parameters the peak stellar surface density (i.e., the ACF values at the smallest separation bin) and the ACF at separations larger than 21 are well reproduced. However, all simulated distributions fail to reproduce the ACF behavior at separations smaller than 15, in identical manner as the two previous scenarios (see Figure\[f:wfitacf\]). The origin of this general mismatch with the observations is the fact that the stellar surface density of both the underlying EFF distributions and the random field does not increase at continuously smaller spatial scales. ![The observed ACF of NGC346 (solid black line) with the area filled by the ACFs of the best-representative simulated mixed distributions that assume a central compact stellar component and an extended fractal one (light grey). The useful separations range of [ -2.truept]{}70, where the ACF indexes were determined, is indicated by the vertical dashed line. These models succeed in reproducing the observed ACF at the shortest and longest stellar separations. They have, though, the tendency to over-predict the ACF values at separation of $\sim$10. The ACF of the best-matching model is plotted with a short dashed line. These simulations provide constraints for the basic parameters of the contributing stellar distributions, with the central cluster having a core radius of $\sim$9($\sim$2.5pc) and extending with a slope $\gamma \sim$2.27, and the self-similar distribution having a fractal dimension of $D_{3} \sim$2.3.[]{data-label="f:bfitacf"}](tpcf_best-models3_mono.eps){width="\columnwidth"} Simulations of a purely Self-Similar Distribution {#s:purefractal} ------------------------------------------------- In Appendix\[s:fractaldistr\] we present a detailed account of the behavior of the ACF for self-similar, i.e. fractal, stellar clusterings, based on artificial stellar distributions. It is shown that hierarchical stellar distributions are characterized by a monotonous (i.e. single-slope) linear dependency of the ACF on stellar separations in log-log, which extends across the whole considered length-scales range. The shape of the observed ACF for NGC346 (Figures\[fig:autocorrelation\] and \[f:cclusmask\]) is proven to behave like a broken power-law with different slopes for different separations ranges (Section\[s:acf346\]). As a consequence, with our simulations of pure fractal stellar distributions we are able to reproduce each one of the two linear parts of the observed ACF, but not its complete shape across the whole separations range. As we discuss in Section\[s:acf346\], and show in the following section, a second stellar component must be considered in order to successfully reproduce the broken power-law behavior of this ACF. Simulations of a Single Cluster plus a Self-Similar Distribution {#s:oneclusterandfractal} ---------------------------------------------------------------- The decomposed ACF of Figure\[f:cclusmask\], constructed by masking the central condensed stellar concentration of the region, strongly suggests the existence of an underlie self-similar distribution of young stars in NGC346, and it provided the first evidence that the system may be the combined result of two such diverse stellar distributions. Indeed, we find that the most successful simulations in replicating the observed behavior of ACF for NGC346 are those which contain both a centrally concentrated dense cluster and a large-scale hierarchical (fractal) distribution. We constructed a sample of distributions by varying the following parameters. (1) The fraction of stars that belong to the cluster $f_{\rm cl}$ (between 0.3 and 0.7), the core radius of the EFF cluster (7-12), the $\gamma$ index of its profile (1.8-3.0) and the three-dimensional fractal dimension $D_3$ of the fractal distribution (from 2 to 2.6). There is only a small subset of these models that is able to reproduce satisfactorily that observed ACF, i.e. the peak value of $1+\xi$, the power-law slope for separations between 1 and 21, and the steeper slope for larger separations (21to 70). The simulations that represent best the observed ACF have 2,300 to 2,500 stars in the cluster, which has a core radius of 4 to 9 and $\gamma$ between 2.25 and 2.35. The best-matching extended component has a high fractal dimension (2.2 $<$ $D_3$ $<$ 2.4). Smaller clusters systematically over-predict the ACF peak value (at the shortest-scale bin). Clusters with core radii [ -2.truept]{}10fail to simultaneously reproduce the AVF behavior at both small and large separations. They produce too much correlation for large separations, which can be remedied by increasing the fraction of stars in the cluster but then the power-law behavior at shorter spacings is lost. With these dedicated simulations we succeeded in reproducing both the shallow linear behavior of NGC346 ACF at short separations, as well as its steeper drop at large separations. We reproduced the ACF index for every part of its broken power-law, as well as the separation limit, where the break occurs. Twelve of our combined simulations reproduce the ACF of NGC346, providing constraints to the basic parameters of the assumed stellar distributions. In all these simulations the core radius of the central cluster is found to be practically unchanged and equal to $\sim$8-9. The fraction of stars in the EFF cluster varies around 0.4, and the slope $\gamma$ of the assumed cluster profile has values $\gamma\simeq$2.20 to 2.35. Finally, the input three-dimensional fractal dimension of the best-fitting simulations varies at values between $\sim$2.20 and 2.36. The observed ACF of NGC346 and those of the twelve most successful simulations are shown in Figure\[f:bfitacf\]. It should be noted that the modeled ACFs are very sensitive to the chosen parameters. We established the best-matching models by iteratively refining the input parameters. In Figure\[f:kdemodel\] the KDE surface density map of the most successful distribution (short dashed line in Figure\[f:bfitacf\]) with $f_{\rm cl} =$0.4, $r_{\rm c} =$9, $\gamma = 2.3$, and $D_{3}=$2.32, is shown to be compared to the KDE map of NGC346. This density map shows secondary peaks, resembling concentrations seen in NGC346. Nevertheless, we could not reproduce the amplitude of the second most important stellar clump of NGC346, which may mean that there is a second (much smaller) compact cluster in the region. ![Surface stellar density (significance) map constructed from the synthetic stellar distribution that represents best the stellar clustering in NGC346. The map is constructed with the use of the KDE technique and a 5 Gaussian kernel, in comparison to the NGC346 map of Figure\[f:kdemap\]. The color-bar indicates different density levels in $\sigma$, the standard deviation of the measured stellar density in the regions. The artificial field-of-view is identical to the observed field-of-view. Coordinates are given as in Figure\[f:kdemap\] for NGC346.[]{data-label="f:kdemodel"}](best-model_scmap.eps){width="\columnwidth"} Summary and Discussion {#s:sum} ====================== In this study we present a thorough analysis of the clustering behavior of young stars in the star-forming region NGC346. In particular, we construct the KDE density map and the separations distribution function, and we determine the ACF of the rich sample of low-mass PMS stars (supplemented by the young massive stars) in the region. We show that the ACF is a robust method to characterize the clustering of stars, providing that its behavior is accurately understood through dedicated simulations. We characterize the stellar clustering in NGC346, and interpret its ACF, through the construction and study of the ACFs of simulated centrally condensed and fractal stellar distributions. We previously found indications that NGC346 includes multiple compact stellar over-densities, and that the distribution of PMS stars in the region is hierarchical (PaperI). With our analysis here we demonstrate that the young stellar design of NGC346 is far more complicated. The comparison of the observed ACF with those of simulated stellar distributions shows clear evidence that NGC346 includes a hierarchical stellar component, extended across the whole field-of-view, and a compact stellar cluster, which appears more dominant in the central part of the observed region. The observed ACF of young stars in NGC346 is successfully reproduced if we assume that $\sim$40% of the stellar population belongs to a condensed cluster with core radius of $r_{\rm c} \sim$2.5pc and a density profile index $\gamma \sim$2.3, “embedded” in a self-similar stellar distribution with a fractal dimension of $D_{3}\sim$2.3. NGC346 is, thus, a very interesting example of composed stellar distributions, on length-scales typical for molecular clouds, providing an unprecedented insight of the topology of star formation. The fact that we still observe self-similar stellar clustering even after $\sim$3Myr is also interesting, because this time-scale provides a constraint on how long it takes for the stars to lose their primordial spatial distribution. The derived three-dimensional fractal dimension of $D_3 \simeq 2.3$, for the self-similar stellar component, fits very well to the value derived from numerical experiments of supersonic isothermal turbulence [@federrath09], and with measurements inferred from observations of the interstellar gas [@elmegreen01b]. This agreement implies that the self-similar stellar distribution in NGC346 is possibly inherited by the turbulent interstellar gas of the natal molecular cloud. In our study for the determination of $D_3$ we use the ACF, [@federrath09] argue that the $\Delta$-variance [@stutzki98; @ossenkopf08] is the most reliable method, and [@elmegreen01b] use the size distribution function of stellar aggregates. Although different, all three studies agree in the derived value of $D_3 \sim$2.3. On the other hand, recent numerical simulations by [@girichidis12] and [@dale13] find a higher degree of hierarchy for stars formed from turbulent molecular clouds with smaller values of $D_3 \sim$1.6. These studies, which apply the $\cal{Q}$-parameter [@cw04], find similar fractal dimensions with that inferred from the same method for the Taurus molecular complex ($D_3 \sim$1.5), in line with the fractal dimensions of the ISM in this region [see, e.g., @alfaro11]. While the differences in the derived fractal dimensions may reflect discrepancies in the methods used, the may as well demonstrate the fact that stars have a different spatial distribution to the gas from which they formed, or that the gas is not similarly structured everywhere. Observations of nearby Galactic star-forming regions show that while the gas filaments have a smooth, radially decreasing density profile [e.g., @arzoumanian11], the distribution of recently-formed stars ($\tau \sim$1Myr) has a range of different morphologies, some being fractal and others centrally concentrated [@cw04]. Whether or not every stellar distribution forms with substructure has yet to be determined. N-body simulations suggest that even a moderate amount of dynamical interactions will partly erase substructure, forbidding a star-forming region from retaining a strong signature of the primordial ISM distribution [e.g., @scally02; @gw04]. Considering that subsequent substructure is difficult to be formed, and dynamics may have aided the formation of a condensed cluster in NGC346, the observed substructure may be an upper limit of the primordial value. The existence of the central compact cluster naturally complicates somewhat the star formation picture in the region. Assuming that the clustered stellar population is coeval with the extended hierarchical stellar component, one has to consider two scenarios for its formation: Either (1) this cluster was formed as a distinct prominent compact concentration amid a distributed stellar population, or (2) it was originally formed as part of the extended stellar component, and soon became centrally condensed due to rapid dynamical evolution. The first scenario implies the co-existence of two different modes of star formation; one that produces a centrally concentrated cluster plus another that is responsible for the distributed PMS stars. The formation of both clustered and distributed populations of young stars in a single molecular cloud is numerically described by [@bonnell11], where both modes are determined by the local gravitational binding of the cloud. The bimodal star formation scenario is further supported by the large number of ‘unclustered’ PMS stars (see also PaperI). The second scenario assumes that the centrally condensed cluster in NGC346 is the merging product of distinct compact newly-born sub-clusters within the natal cloud. This scenario is supported by hierarchical fragmentation of the turbulent molecular cloud, which forms stars in many small sub-clusters [@klessenburkert]. These sub-clusters will interact and merge to form a dominant stellar cluster through closer and more frequent dynamical interactions [@bonnell03]. The existence of compact PMS sub-clusters along the whole extend of NGC346 makes this scenario favorable. Recent N-body simulations of the dynamical evolution of star-forming regions by [@parker13] also support this scenario. In these simulations, initially sub-structured super-virial agglomerations will evolve to form a multi-clustered region, similar to NGC346, within 5Myr. The simulations by [@parker13] predict that star-forming regions will dynamically evolve to form bound clusters or unbound associations, depending on their initial conditions, i.e., their virial status and fractal dimension. Based on the present status of the central cluster, we cannot know which of the above scenarios explains its origins. Specifically, it is unclear if the cluster is still under formation (through merging or not), and if it is undergoing core collapse that evaporates its low-mass stars due to gas expulsion [@baumgardtkroupa07], or through rapid dynamical interactions [@allison10]. Considering that all simulations discussed here deal with length-scales and stellar numbers smaller than that of NGC346, modeling of richer stellar samples in larger areas is surely necessary for constraining the initial conditions that led to the formation of NGC346. Most simulations predict central segregation of the most massive stars, but they measure it with different methods and explain it with different mechanisms. Mass segregation may be primordial because massive stars are born in situ at the central region, or due to rapid equipartition, and so it is the product of dynamical evolution. Mass segregation is observed in NGC346 through the comparison of its isochronal age with the time-scale for energy equipartition, i.e., for mass segregation [@spitzer87 p. 74, see also @kroupa04]. [@sabbi08] found that the latter is one order of a magnitude larger, impliying that the observed mass segregation is likely due to initial conditions, rather than dynamical evolution. The present dynamical status of whole region of NGC346 can be naively[^2] determined in terms of its relaxation and crossing time-scales. The relaxation time, $T_{\rm relax} \approx \left( 0.1\,N/\ln{N} \right)\,T_{\rm cr}$, is the time for significant energy redistribution to occur in a cluster [@bt87 p. 37], where $T_{\rm cr} \approx 2r_{\rm h}/\sigma$, is the crossing time of a typical star through the cluster which has a characteristic radius $r_{\rm h}$ and a velocity dispersion $\sigma \approx GM/r_{\rm h}$ [see also @kroupa08]. $G = 0.0045\,{\rm pc}^3/\left({\rm M}{\solar}\,{\rm Myr}^2\right)$ is the gravitational constant, $N$ the total number of stars, and $M$ that total stellar mass of the cluster. Using a radius $r_{\rm h} \simeq$9pc and a total mass $M \simeq 3.9\,10^{5}$M, [@sabbi08] derives $T_{\rm relax} \simeq$570Myr, far larger than the age of NGC346. This result, supported by the apparent substructure in the region suggests that NGC346 is not dynamically relaxed, in agreement with the assumption that mass segregation in NGC346 is primordial. [@gpz11] use the ratio of stellar age over the crossing time, $\Pi = {\rm Age}/T_{\rm cr}$, to distinguish bound from unbound stellar systems, with $\Pi < 1$ for unbound, i.e., expanding, objects. Using their own definition of the crossing time in terms of empirical cluster parameters (their Eq. 1), and the values for age, $r_{\rm h}$ and $M$ from Sabbi et al., they derive $T_{\rm cr} = 6.4$Myr, and thus $\Pi \approx 0.5$, classifying NGC346 as unbound agglomerate. However, assuming a velocity dispersion of $\sigma = 10$kms$^{-1}$, typical for star-forming complexes of the size of NGC346, Sabbi et al. derives $T_{\rm cr} = 1.8$Myr, which classifies NGC346 as a bound system. The difference in the results between the above studies is important, because they influence our understanding of the origin of the region and, thus, the nature of its bimodal stellar clustering. If NGC346 is a gravitationally bound object, but not relaxed yet, as the findings of [@sabbi08] suggest, then probably the condensed stellar component is the [undergoing result of dynamical interactions]{} between smaller clusters. On the other hand if NGC346 expands as the result of [@gpz11] implies, then most probably the central cluster [was originally formed condensed]{}. Under these circumstances a clarification about the origin of the central cluster and of the bimodal clustering behavior of the young stars in NGC346 can only be achieved with the accurate determination of stellar dynamics, i.e., kinematics, of the region. By reversing the above analysis we make a prediction for the velocity dispersion that the system would have if it was dynamically stable or unstable. The condition for stability for NGC346 ($\Pi = 1$) requires $T_{\rm cr} = 3$Myr, which implies a velocity dispersion of $\sigma \simeq$6kms$^{-1}$ for $r_{\rm h} = 9$pc. (This value is in excellent agreement with the line-of-sight velocity dispersion of stars within a projected distance of 5pc from the centre of the young massive cluster R136 in the Tarantula nebula; see @brunet12.) It should be noted that the half-number radius, i.e., the characteristic radius, in our young stellar sample is larger than that determined by Sabbietal., and equal to $r_{\rm h} \simeq 16.5$pc. Assuming this radius, the velocity dispersion at the stability limit raises to $\sigma \simeq$11kms$^{-1}$, still within typical values [e.g., @sana13]. The predicted values of 6-11kms$^{-1}$ for the velocity dispersion of NGC346 provide a limit for characterizing the stability of the system as a whole. If the velocity dispersion is larger that this limit, NGC346 is bound and collapsing (since $\Pi > 1$), while for values lower than this estimate, the system is unbound, and thus possibly under dissolution. In any case, different stellar components in the region are expected to demonstrate different velocity dispersions, depending on their own dynamical status. In the above discussion we assume a single age of $\sim$3Myr for the entire region, based on its PMS stars [@sabbi08]. However, single-epoch photometry of PMS stars is significantly affected by their physical characteristics (unresolved binarity, accretion, circumstellar extinction, variability, etc) that dislocate these stars from their theoretical positions on the color-magnitude diagram. This effect causes a color-luminosity spread of the PMS stars, which translates to an artificial age-spread, making the measurement of stellar ages (and masses) quite uncertain [see review by @gouliermis12]. This issue, for [Hubble]{} imaging of PMS stars in the Magellanic Clouds, can only be accurately addressed through probabilistic determination of their physical parameters [e.g., @dario10]. Naturally, a more accurate determination of PMS ages in NGC346 and verification of an age-spread, which may even be positional dependent, would have noticeable implications to our understanding of star formation in the region. Concluding Remarks {#s:concl} ================== Concluding remarks, derived from our analysis, are summarized to the following: Large star-forming complexes are excellent stellar ‘ecosystems’ for the investigation of clustered star formation, because the represent the typical 100-pc scale of giant molecular clouds. The combined application of various cluster analysis tools is necessary for thoroughly characterizing young stellar clustering. In particular the autocorrelation function emerges as a robust method, because, supported by the appropriate simulations, it is capable to distinguish different clustering styles. We determine that the clustering of young stellar populations in the star-forming complex NGC346 has two distinct components; An extended self-similar, i.e., hierarchical, stellar distribution, and at least a centrally condensed young stellar cluster. Dedicated simulations of combined stellar distributions show that the condensed stellar component of fits to a spherical cluster, including 40% of the stars, with a core radius between 8and 9($\sim$2.5pc) and a power-law surface density profile with slope between 2.20 and 2.35. The remaining population is fractally distributed across the whole extent of the region with a fractal dimension $D_3$ between 2.20 and 2.36. Our findings suggest that the present clustering behavior of young stars in NGC346 is the product of both star formation (turbulent-induced hierarchy), and early dynamical evolution (merging towards or dissolution of a centrally condensed cluster). Both processes seem to be still active after a time-scale of $\sim$3Myr, i.e., the assumed evolutionary age of the region. The origin of this bimodal clustering behavior is not clearly understood. Considering that it influences the future evolution of the region, we assess that if the complete system is under disruption then possibly the central cluster was formed condensed. If NGC346 contracts under its own gravity then this would imply that the central cluster may be the product of this contraction. We determine that the velocity dispersion limit for stability of the region as a whole is $\sigma {\ \raise -2.truept\hbox{\rlap{\hbox{$\sim$}}\raise5.truept\hbox{$>$}\ }}$6-11kms$^{-1}$. However, based on the heterogeneous clustering of the region, the [local]{} velocity dispersion of individual sub-structures should demonstrate variations. Considering the excellent coincidence of the fractal dimension derived for the stellar clustering with that measured and theoretically predicted for the turbulent interstellar medium, a natural step forward for this study is to understand how the stellar assembling process relates to the structuring behavior of the natal interstellar medium. This topic will be addressed in a subsequent study, where we compare the interstellar gas structure to that presented here for the young stars in NGC346. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Richard Parker and Stefan Schmeja for their critical views and useful discussions that helped improving our interpretation. D.A.G. kindly acknowledges financial support by the German Research Foundation through grant GO1659/3-1. S.H. and R.S.K. acknowledge support from the Collaborative Research Center “The Milky Way System” (SFB881), particularly subproject B5, of the German Research Foundation. Based on observations made with the NASA/ESA [Hubble Space Telescope]{}, obtained from the data archive at the Space Telescope Science Institute (STScI). STScI is operated by the Association of Universities for Research in Astronomy, Inc. under NASA contract NAS 5-26555. Library of Simulated Autocorrelation Functions {#s:appendix} ============================================== The Requirement for Simulated ACFs {#s:acflim} ---------------------------------- A cluster analysis method, developed by [@larson95] as a modification of the standard two-point angular correlation function [@gomez93], is the correlation of the mean surface density of stellar companions (MSDC) against pair separation (in logarithmic scales). In general, the MSDC behaves in analogy to the ACF[^3], with a power-law index $\eta$ being identically associated to the fractal dimension $D_2$. Both methods, however, are subject to observational limitations in their interpretation. A break in the power-law of the MSDC of Galactic star-forming regions is suggested to be strongly influenced by the overall stellar surface density [@simon97; @bate98], rather than representing a characteristic scale, as was previously proposed. Moreover, a single power-law index may not even be related to the fractal dimension of the clustering, merely reflecting the large-scale density gradient in a centrally concentrated cluster [@bate98; @klessenkroupa]. These limitations, in addition to edge effects introduced by the unavoidably limited observed fields [@cw04 see also Section\[s:edgeeffect\]], require the simulation of distinctive stellar clusterings for the correct characterization of the ACF, and its use as a cluster analysis diagnostic. In this Appendix we build a library of typical ACFs, based on simulations of centrally concentrated clusters and fractal stellar distributions. We describe these simulations and their ACFs in Appendices\[s:ccclusters\] and \[s:fractaldistr\], where we also verify the behavior of the ACF for a random stellar distribution, i.e., for a sample of non-clustered stars (Appendices\[s:king\] and\[s:effclusters\]). These ‘field’ ACFs are consistent with a value of $\simeq$1, and remain unchanged with the separation range. The spatial coverage and stellar numbers of all considered artificial distributions are scaled to be identical to the field-of-view and stellar sample covered by our [Hubble]{} ACS observations of NGC346. The Autocorrelation Function of Centrally Concentrated Stellar Distributions {#s:ccclusters} ---------------------------------------------------------------------------- We examine the autocorrelation of stars spatially related to each other within a spherically symmetric cluster. We compose artificial centrally-condensed stellar distributions that follow typical stellar surface density profiles, and we construct the corresponding ACFs in order to comprehend their behavior in comparison to the observed ACF of NGC346. All simulations are performed for the same number of stars with that in our observed sample (5,150) confined within a field-of-view identical to that covered by the three ACS/WFC pointings mosaic. We consider two types of radial stellar density profiles for the synthetic clusters: Profiles following the empirical law by [@king62], and those represented by the model of [@eff87]. The latter, representing clusters that are not tidally truncated, are best-suited for extended clusters surrounded by well populated stellar fields, as is usually the case in large star forming regions like NGC346. Therefore we base our analysis on these clusters, the simulations of which are presented in Appendix\[s:effclusters\]. Below, we also present the King star cluster profiles for reasons of completion in our study, and for a reference for future studies on the ACF of star clusters. ### King-profile Clusters {#s:king} We further investigate the ACF of more realistic centrally concentrated stellar clusters in dynamical equipartition. Artificial spherical clusters were simulated to follow stellar surface density profiles defined by King’s semi-empirical model [e.g., @king62]. The functional form of these profiles is: $$\begin{aligned} \label{eq:king} f(r) & \propto & \Bigl(\frac{1}{[1+(r/r_{\rm c})^2]^\frac{1}{2}}-\frac{1} {[1+(r_{\rm t}/r_{\rm c})^2]^\frac{1}{2}}\Bigr)^{2}, \end{aligned}$$ where $f$ is the stellar surface density, and $r$ is radial distance. The clusters are constructed with various values in their concentration parameter $c$, defined as $c=\log{r_{\rm t}/r_{\rm c}}$, where $r_{\rm t}$ and $r_{\rm c}$ are the tidal and core radius of the cluster. ![ACFs for a sample of four centrally concentrated clusters that follow King-type stellar density profiles with different concentration parameters, $c$, as indicated in the plot. Vertical color dotted lines correspond to the $r_{\rm h}$ of the clusters, and the black dashed line corresponds to the $r_{\rm t}$ of the clusters, which is chosen to be identical and equal to 55pc. The ACF of a random stellar distribution is plotted for reference in red.[]{data-label="fig:acfmodel"}](tpcf_king-clusters_plot_v2.eps){width="\columnwidth"} The ACFs for a sample of such clusters are shown in Figure\[fig:acfmodel\] along with that for the random field (plotted in red). Clusters are constructed with concentration parameters equal to 0.5 (plotted in orange), 1.0 (light green), 1.5 (green) and 3.0 (blue) respectively. The latter represents the observed extreme in the parameters space of known Galactic globular clusters [Pal12, $c=2.98$; @harris96 2010 Edition[^4]]. From these plots it is seen that while indeed for all clusters the autocorrelation function is $1+\xi(r) > 1$, it does not again remain constant through the whole range of stellar separations $r$, but it drops at larger separations towards the outskirts of the clusters converging to the random field value of unity. The ACFs reach this value at the radii where essentially the clusters meet the surrounding field, comparable to their tidal radii. For reasons of simplicity, in the examples shown in Figure\[fig:acfmodel\] all clusters are constructed to have the same $r_{\rm t} = 55$pc ($\simeq$196). ![ACFs for a sample of five centrally concentrated clusters that follow EFF-type stellar density profiles. All clusters in the plot have the same number of stars, $\sim5,000$, and a fixed outer profile slope, $\gamma=3$, typical for young clusters in the Magellanic Clouds, while their core radii, $r_{\rm c}$, vary. They are selected to have typical values of 2, 4, 6, 8, and 10, indicated with vertical dashed lines in magenta, blue, green, light-green and orange respectively. The ACF of a simulated random (unclustered) field of the same number of stars is shown in red for reference.[]{data-label="f:effacs1"}](tpcf_eff-clusters_plot_v1.eps){width="\columnwidth"} ![ACFs for a sample of five centrally concentrated clusters that follow EFF-type stellar density profiles. All clusters have the same number of stars, $\sim5,000$, and a typical fixed core radius, $r_{\rm c}=8\arcsec$, indicated in the plot with the vertical grey dashed line. Their outer density profile indexes, $\gamma$, are selected to have values typical for young Magellanic Clouds clusters, of 1, 2, 3, 4, and 5. The case of $\gamma=1$ is an unobserved extreme used here for demonstrating the flatness of its ACF.[]{data-label="f:effacs2"}](tpcf_eff-clusters_plot_v2.eps){width="\columnwidth"} There are two derivatives from these simulations: (1) The maximum values of $1+\xi$ depend on the degree of concentration, i.e. the concentration parameter, of each cluster; More centrally concentrated clusters have higher ACF values, and they are thus ‘more clustered’. This trend, which is more obvious at small separations, becomes less important for clusters with $c {\ \raise -2.truept\hbox{\rlap{\hbox{$\sim$}}\raise5.truept\hbox{$>$}\ }}1.5$. (2) The shape of the ACF drops at larger separations (up to the field value) also depends on the concentration parameter. It naturally depends also on the limiting radius of the cluster, i.e., its tidal radius; Clusters with larger $r_{\rm t}$ and smaller $c$ have smoother drop in their ACFs. In the examples shown in Figure\[fig:acfmodel\] all clusters have the same $r_{\rm t} $, and thus the steepness of the drop of their ACF depends only on their concentration parameter. ![image](fractal-distr-plots_qudr-hor.eps){width="97.50000%"} ### EFF-Profile Clusters {#s:effclusters} The outskirts of young stellar clusters, located in star-forming regions of the Magellanic Clouds, demonstrate extended outer envelopes, which cannot be represented by King’s semi-empirical model [e.g., @king62], designed for tidally truncated globular clusters. [@eff87] developed an empirical model more suitable to describe the stellar surface density profile of such clusters (from hereon the EFF model). We base our simulations of centrally-concentrated clusters on this model. The surface stellar density of the cluster according to the EFF model is described as: $$\begin{aligned} \label{eq:eff} f(r) & = & f_{0} {\left[ 1 + {\left( \frac{r}{\alpha} \right)}^{2} \right]}^{-\gamma/2}+f_{\rm{field}},\end{aligned}$$ where $f_{0}$ is the central stellar surface density, $\alpha$ is a measure of the core radius and $\gamma$ is the power-law slope which describes the decrease of surface density of the cluster at large radii; $f(r)\propto r^{-\gamma/2}$ for $r\gg a$. The uniform background density is given by $f_{\rm field}$. For comparison, a King profile (Eq.\[eq:king\]) for $r_{\rm t} \gg r_{\rm c}$ would be described as: $$\begin{aligned} \label{eq:kingeff} f(r) & = & f_{0} {\left[ 1 + {\left( \frac{r}{r_{\rm c}} \right)}^{2} \right]}^{-1}+f_{\rm{field}}.\end{aligned}$$ The basic parameters of the EFF model, $\alpha$, $\gamma$ and $f_{\rm field}$, are measured from the fitting of observed profiles with the function of Eq.(\[eq:eff\]). In order to determine the ACF of clusters following EFF profiles, and compare it to that for NGC346, we simulate clusters that contain the same number of stars as the observed field of NGC346 placed in a comparable field-of-view. Apart from the total number of stars of the cluster, we construct our EFF centrally concentrated clusters by providing as input parameters its core radius, $r_{\rm c}$, and the outer surface density index, $\gamma$. For reasons of simplicity we ignore the constant contribution of the field, assuming $f_{\rm field} = 0$. According to the EFF model, the relation of $\alpha$ to the core radius is given from Eq.(\[eq:eff\]) assuming no contribution from the field: $$\begin{aligned} r_{\rm c} & = & \alpha(2^{2/\gamma}-1)^{1/2} \label{eq-rc}. \end{aligned}$$ Our simulated clusters have $r_{\rm c}$ and $\gamma$ values comparable to typical values for young clusters in the Magellanic Clouds [@mg03lmc; @mg03smc]. In Figure\[f:effacs1\] we show the ACFs of five EFF clusters with the same typical value of $\gamma = 3$ and different core radii. Figure\[f:effacs2\] demonstrates the constructed ACFs for such clusters keeping the core radius constant at the typical value of $r_{\rm c} = 8\arcsec$ while varying the profile index $\gamma$ within the range of observed values. Both Figures\[f:effacs1\] and\[f:effacs2\] demonstrate the dependance of the shape of the ACF of centrally condensed clusters on the assumed structural parameters of the clusters. In Figure\[f:effacs1\] can be seen that, for clusters with the same number of stars, larger $r_{\rm c}$ produced more shallow ACFs in particular at smaller separations. The reason for this behavior is that stars would be confined in a narrow area (with short separations) if the core were small. A larger core for the cluster forces them to be distributed over a wider area with higher separations among them, flattening the ACF at small separations. These clusters will appear in the ACF plot ‘less clustered’ than those which are more centrally concentrated, i.e., in smaller cores. On the other hand, as shown in Figure\[f:effacs2\], if the core remains unchanged, the surface density profile slopes change accordingly the decrease of the ACF at large separations. Steeper density profiles lead to ACFs, which are steeper at large separations. The slope of the ACF at smaller separations remains unchanged, while its absolute values are smaller (less clustered) for more shallow stellar surface density profiles at the outskirts of the clusters. Both figures with the ACFs of EFF-profile clusters show a change in their separations dependence, i.e., a ‘breaking’ of the ACF slope, at different scales. In Figure\[f:effacs1\], the use of clusters with the same $\gamma$ demonstrates that the radial distance where the separation dependence of ACF changes, i.e., where the power-law ‘breaks’, depends on how well concentrated the clusters are, i.e., on their $r_{\rm c}$. Therefore, the separation, where the break occurs may be related to a radial scale, which is characteristic for its stability[^5]. The Autocorrelation Function of Self-Similar Distributions {#s:fractaldistr} ---------------------------------------------------------- ![ACFs of six different self-similar stellar distributions, having fractal dimensions of $D_{3} =$1,1.6,2,2.3,2.6and 2.8. A sample of one distribution out of several constructed for each $D_{3}$ is shown here. These plots show that the ACFs of fractal stellar concentrations should monotonically increase for smaller stellar separations, unlike what we observe in NGC346.[]{data-label="fig:acfsfractal"}](tpcf_fractals_plot_v1.eps){width="\columnwidth"} In this section we consider the ACF of self-similar, i.e., fractal stellar distributions. We construct three-dimensional artificial fractal distributions with the application of a reverse box-counting algorithm by first defining a cube of side-length 1. Next, this cube is divided into ${\cal N}_{\rm div}^3$ equal sub-cubes, of side-length $1/{\cal N}_{\rm div}$, which is the first generation of ‘children’. ${\cal N}_{\rm ran}$ of these sub-cubes are then randomly selected to become parents themselves and further divided into ${\cal N}_{\rm div}^3$ child sub-cubes, and the process is repeated recursively, terminating at the desired level of recursion. At this stage each of the smallest (final generation) sub-cubes has a star placed in it. Finally, we normalize the three-dimensional coordinates of these stars to a total side-length of the order of our observed field-of-view[^6]. The counting-box fractal dimension depends on both ${\cal N}_{\rm div}$ and ${\cal N}_{\rm ran}$, and is defined as $$\begin{aligned} \label{eq:cboxfract} D \equiv \frac{\ln{({\cal N}_{\rm ran})}}{\ln{({\cal N}_{\rm div})}}. \end{aligned}$$ Normally we use ${\cal N}_{\rm div} = 2$, in which case there are 8 sub-cubes (children) in every generation. The probability that a child cube will further be divided and become a parent is ${\cal N}_{\rm div}^{(D-3)}$. For lower $D$, this probability is lower and the distribution becomes more ‘porous’. This technique was originally implemented by various authors [see, e.g., @bate98; @cw04]. The produced fractal distributions contain specific stellar numbers by culling randomly the stars that populate the final samples. The number of iterations, i.e., of produced generations, depends on the total number of stars to populate the distribution, $N_{\star}$ and ${\cal N}_{\rm ran}$ as $N_{\rm gen} = \log{(N_{\star})}/\log{({\cal N}_{\rm ran})}$. Normally we require 10,000 stars to populate our fractal distributions, and thus we produce about 13 generations in each simulation. To avoid an obviously regular structure, we add a bit noise to the positions of the stars, by repositioning every star by a random infinitely small fraction of the sub-cube size. The examples of three such fractal distributions are shown in Figure\[f:frctsmpl\]. The ACFs of indicative self-similar distributions of approximately 5,000 stars with fractal dimensions $D_3$ between 1.0 and 2.8, are shown in Figure\[fig:acfsfractal\]. These ACFs show a monotonous decrease with stellar separation, and no indication of a change of slope at a specific scale, unlike what we find in NGC346. We examined the effect of different viewing angles of the distributions to their two-dimensional projections and the subsequent effect to the produced ACF, and we found that – at least for these stellar numbers of $\sim$5,000 – the derived ACFs show no difference at all, with the corresponding indexes remaining essentially unaffected. It should be noted that the pure 3D fractals generated by the box counting method, after projected on the 2D plane, produce a “wiggle” in their ACF, which nevertheless does not affect their monotonous trend. The scale where this wiggle appears seems to depend on the fractal dimension, occurring for low fractal dimensions at larger scales than for higher fractal dimensions. In the case of the best representative simulations ($D_3$ $\sim$ 2.3) this wiggle occurs at separations [ -2.truept]{}2. Relation between the Fractal Dimensions $D_3$ and $D_2$ through the ACF Index $\eta$ {#d3toeta} ==================================================================================== The application of the ACF provides a measurement of the two-dimensional fractal dimension $D_2$ of the considered ensemble through its index $\eta$ (see, e.g., Section\[s:acf\]). However, we construct our simulated fractal distributions in a volume, providing as basic input parameter the three-dimensional fractal dimension $D_3$, and there is no direct relation between $D_3$ and $D_2$. A simple conversion $D_3=D_2+1$ is usually cited, which, however, applies only if the perimeter-area dimension of a projected 3D structure is the same as the perimeter-area dimension of a slice [@elmegreenscalo04]. We utilize our simulations of self-similar stellar distributions to provide a more general conversion between $D_3$ and $\eta$, and consequently between $D_3$ and $D_2$. This empirical conversion can be very useful for comparing results from methods that measure $D_2$ to those from methods that provide measurements for $D_3$ [see, e.g., @federrath09 for a discussion on the available methods], with no need for assuming the simple relation $D_3=D_2+1$. Since in our box-counting simulations we have to use integer numbers for ${\cal N}_{\rm ran}$ and ${\cal N}_{\rm div}$, there is a limited number of input values for $D_3$ that can be applied. We performed simulations assuming two numbers for ${\cal N}_{\rm div}$ (2 and 3) and established a set of 8 self-similar distributions with different $D_3$, which span the complete realistic range of values between $D_3 =$0.5 and [ -2.truept]{}3. Then we constructed their ACFs and determined the corresponding indexes $\eta$, as well as the corresponding two-dimensional fractal dimensions $D_2$ from the relation $D_2 = 2+ \eta$. The derived empirical calibration is shown in Figure\[f:d3eatd2\]. From this figure one can see that $D_2$ almost equals $D_3$ only for $D_3$[ -2.truept]{}1.6; this one-to-one relation is represented by the left-hand dashed grey line. For larger values of $D_3$, the derived $D_2$ converges toward its maximum possible value of 2 while $D_3$ approaches its own maximum of 3, having another linear dependence to each other, demonstrated by the right-hand dashed grey line. A simple functional form that can thus represent this relation is: $$\begin{aligned} D_{2} \simeq \left\{ \begin{array}{r@{\,\,}l} D_{3} & ,\,{\rm for}\, D_3 \le 1.6\\ & \\ 1.1 + 0.3\,D_{3} & ,\,{\rm for}\, D_3 \ge 1.6\\ \end{array}\right. \label{eq:funcform}\end{aligned}$$ It is worth noting that for typical $D_3$ values derived for turbulent-induced hierarchy, measured in interstellar gas ($D_3 \sim 2.3$), the corresponding $D_2$ value according to our conversion is $D_2 \sim 1.8$. The errors shown in Figure\[f:d3eatd2\] are the standard deviations of the derived values for $\eta$ after performing few realizations in the construction of the synthetic fractal distributions for each $D_3$ value. ![Calibration relation between the three-dimensional fractal dimension $D_3$, the ACF index $\eta$, and the corresponding two-dimensional fractal dimension $D_2$, derived from our simulated self-similar stellar distributions.[]{data-label="f:d3eatd2"}](fractaldimension_eta_plus_final-model.eps){width="\columnwidth"} Alfaro, E. J., & S[á]{}nchez, N. 2011, Computational Star Formation, 270, 81 Allison, R. J., Goodwin, S. P., Parker, R. J., Portegies Zwart, S. F., & de Grijs, R. 2010, , 407, 1098 Arzoumanian, D., Andr[é]{}, P., Didelon, P., et al. 2011, , 529, L6 Bastian, N., Gieles, M., Ercolano, B., & Gutermuth, R. 2009, , 392, 868 , M. R., [Clarke]{}, C. J., & [McCaughrean]{}, M. J. 1998, , 297, 1163 Baumgardt, H., & Kroupa, P. 2007, , 380, 1589 Binney, J., & Tremaine, S. 1987, Galactic Dynamics, Princeton University Press, Princeton, NJ Bonnell, I. A., Smith, R. J., Clark, P. C., & Bate, M. R. 2011, , 410, 2339 Bonnell, I. A., Bate, M. R., & Vine, S. G. 2003, , 343, 413 Cartwright, A., & Whitworth, A. P. 2004, , 348, 589 Cartwright, A., & Whitworth, A. P. 2009, , 392, 341 Cignoni, M., Tosi, M., Sabbi, E., Nota, A., & Gallagher, J. S. 2011, , 141, 31 Clark, P. C., Klessen, R. S., & Bonnell, I. A. 2007, , 379, 57 Da Rio, N., Gouliermis, D. A., & Gennaro, M. 2010, , 723, 166 Dale, J. E., Ercolano, B., & Bonnell, I. A. 2013, , 430, 234 Dolphin, A. E. 2000, , 112, 1383 Efremov, Y. N., & Elmegreen, B. G. 1998, , 299, 588 Efremov, Y. N. 2009, Astronomy Letters, 35, 507 Elmegreen, B. G. 2011, EAS Publications Series, 51, 31 Elmegreen, B. G. & Falgarone, E. 1996, ApJ, 471, 816 , B. G. & [Elmegreen]{}, D. M. 2001, , 121, 1507 Elmegreen, B. G., & Scalo, J. 2004, , 42, 211 , B. G., [Elmegreen]{}, D. M., [Chandar]{}, R., [Whitmore]{}, B., & [Regan]{}, M. 2006, , 644, 879 Elson, R. A. W., Fall, S. M., & Freeman, K. C. 1987, , 323, 54 Falgarone, E., Phillips, T. G., & Walker, C. K. 1991, , 378, 186 Federrath, C., Klessen, R. S., & Schmidt, W. 2009, , 692, 364 Feigelson, E. D., Getman, K. V., Townsley, L. K., et al. 2011, , 194, 9 Gieles, M., & Portegies Zwart, S. F. 2011, , 410, L6 Girichidis, P., Federrath, C., Allison, R., Banerjee, R., & Klessen, R. S. 2012, , 420, 3264 , M., [Hartmann]{}, L., [Kenyon]{}, S. J., & [Hewett]{}, R. 1993, , 105, 1927 Goodwin, S. P., & Whitworth, A. P. 2004, , 413, 929 Gouliermis, D. A. 2012, , 169, 1 Gouliermis, D. A., Dolphin, A. E., Brandner, W., & Henning, T. 2006, , 166, 549 Gouliermis, D. A., Chu, Y.-H., Henning, T., et al. 2008, , 688, 1050 Harris, W. E. 1996, , 112, 1487 H[é]{}nault-Brunet, V., Evans, C. J., Sana, H., et al. 2012, , 546, A73 Henize, K. G. 1956, , 2, 315 Jeffries, R. D. 2012, Star Clusters in the Era of Large Surveys, Astrophysics and Space Science Proceedings, 163 Karampelas, A., Dapergolas, A., Kontizas, E., et al. 2009, , 497, 703 Kim, S., Staveley-Smith, L., Dopita, M. A., et al. 2003, , 148, 473 King, I. 1962, , 67, 471 Klessen, R. S., & Burkert, A. 2000, , 128, 287 Klessen, R. S., & Kroupa, P. 2001, , 372, 105 Kroupa, P. 2004, , 48, 47 Kroupa, P. 2008, The Cambridge N-Body Lectures, 760, 181 Krumholz, M. R., & Tan, J. C. 2007, , 654, 304 Lada, C. J., & Lada, E. A. 2003, , 41, 57 , R. B. 1995, , 272, 213 Levenberg, K. 1944, Q. Appl. Math., 2, 164 Mac Low, M.-M., & Klessen, R. S. 2004, Reviews of Modern Physics, 76, 125 Mackey, A. D., & Gilmore, G. F. 2003a, , 338, 85 Mackey, A. D., & Gilmore, G. F. 2003b, , 338, 120 Mandelbrot, B. B. 1983, The fractal geometry of nature (Freeman, San Francisco) Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251 Marquardt, D. 1963, SIAM J. Appl. Math., 11, 431 Megeath, S. T., Gutermuth, R., Muzerolle, J., et al. 2012, , 144, 192 Odekon, M. C. 2006, , 132, 1834 Ossenkopf, V., Krips, M., & Stutzki, J. 2008, , 485, 917 Parker, R. J., Wright, N. J., Goodwin, S. P., & Meyer, M. R. 2013, Accepted (arXiv:1311.3639) , P. J. E. 1980, [The large-scale structure of the universe]{} (Research supported by the National Science Foundation. Princeton, N.J., Princeton University Press, 1980. 435 p.) Preibisch, T. 2012, Research in Astronomy and Astrophysics, 12, 1 Rubio, M., Contursi, A., Lequeux, J., et al. 2000, , 359, 1139 Sabbi, E., Sirianni, M., Nota, A., et al. 2007, , 133, 44 (Erratum: 2007, , 133, 2430) Sabbi, E., Sirianni, M., Nota, A., et al. 2008, , 135, 173 Sana, H., de Koter, A., de Mink, S. E., et al. 2013, , 550, A107 Scally, A., & Clarke, C. 2002, , 334, 156 Schmeja, S., Kumar, M. S. N., & Ferreira, B. 2008, , 389, 1209 Schmeja, S., Gouliermis, D. A., & Klessen, R. S. 2009, , 694, 367 Spitzer, L. 1987, Dynamical Evolution of Globular Clusters, Princeton University Press, Princeton, NJ Sreenivasan, K. R. 1991, Annual Review of Fluid Mechanics, 23, 539 Silverman, B. W. 1992, [Density Estimation]{}, London:Chapman & Hall Simon, M. 1997, , 482, L81 Stutzki J., Bensch F., Heithausen A., Ossenkopf V., & Zielinsky M. 1998, 336, 697 Scheepmaker, R. A., Lamers, H. J. G. L. M., Anders, P., & Larsen, S. S. 2009, , 494, 81 Westpfahl, D. J., Coleman, P. H., Alexander, J., & Tongue, T. 1999, , 117, 868 Zhang, Q., Fall, S. M., & Whitmore, B. C. 2001, ApJ, 561, 727 [^1]: In PaperI we applied an unsupervised cluster detection, based on the 20th nearest-neighbor density map of the region. [^2]: Due to its asymmetrical stellar clustering, the determination of a crossing time for the whole region may not make sense. It is, though, useful in understanding its dynamical status, especially in comparison to other objects. [^3]: An advantage of the ACF over the MSDC is that the former is normalized by the average density in the survey area. This allows the clear distinction between strongly and weakly clustered samples (from the absolute ACF values), and the direct comparison of the shape of ACF for different ensembles. [^4]: [http://physwww.physics.mcmaster.ca/$\sim$harris/mwgc.dat](http://physwww.physics.mcmaster.ca/~harris/mwgc.dat) [^5]: We explored this dependance for clusters following a King profile in Appendix\[s:king\]. [^6]: We assume an average side-length of 6656 pixels, corresponding to $\simeq$5.55, or $\simeq$94pc.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We report an experimental test of complementarity using clock-triggered single-photon pulses emitted by an individual N-V color center in a diamond nanocrystal. The single photons are sent into a Mach-Zehnder interferometer with an output beamsplitter of adjustable reflection coefficient $R$. In addition, the choice of introducing or removing this beamsplitter is random and relativistically space-like separated from the entering of the photon inside the interferometer, as required for the Wheeler’s delayed-choice regime. Each set value of $R$ allows us to observe interference with visibility $V$ and to obtain incomplete which-path information characterized by the distinguishability $D$. The measured values of $V$ and $D$ are found to obey the complementarity relation $V^{2}+D^{2}\leq 1$.' author: - 'Vincent Jacques$^1$, E Wu$^{1,2}$, Frédéric Grosshans$^1$, François Treussart$^1$, Philippe Grangier$^3$, Alain Aspect$^3$ and Jean-François Roch$^1$' title: 'Delayed-choice test of complementarity with single photons' --- As emphasized by Bohr [@Bohr], complementarity lies at the heart of quantum mechanics. A celebrated example is the illustration of wave-particle duality by considering single particles in a two-path interferometer [@Feynman], where one chooses either to observe interference fringes, obviously associated to a wave-like behavior, or to know which path of the interferometer has been followed, according to a particle-like behavior [@footnote; @no; @entanglement]. Although interference has been observed at the individual particle level with electrons [@Tonomura], neutrons [@Neutrons], atoms [@Carnal; @Keith], molecules [@Arndt], only a few experiments with massive particles have explicitly checked the mutual exclusiveness of which-path information (WPI) and interference [@Pfau; @Pritchard; @Buks; @RempeNature; @Haroche].\ In the case of photons, it has been pinpointed that meaningful two-path interference experiments demand a single-photon source [@Grangier] for which full and unambiguous WPI can be obtained, complementary to the observation of interference [@Grangier; @Kurtsiefer; @Jacques]. In order to rule out a too naive view of complementarity, which would assume that the particle could “know” when entering the apparatus which experimental configuration has been set (record of interference or determination of WPI) and would then adjust its behavior accordingly [@Greenstein], Wheeler proposed the “delayed choice" scheme where the choice between the two complementary measurements is made long after the particle entered the interferometer [@Wheeler]. Realizations of that gedanken experiment [@Zajonc; @Martiennsen; @Baudon; @DCECachan] have confirmed that the chosen observable can be determined with perfect accuracy even if the choice, made by a quantum random number generator, is space-like separated from the entering of the particle into the interferometer [@DCECachan].\ In 1978, Wooters and Zurek [@Zurek] considered an intermediate situation in which interaction with the interferometer considered as a quantum device allows one to gain an imperfect – but significant – knowledge of WPI, without destroying the interference pattern, which remains observable with a good – although reduced – visibility. In 1988, Greenberger and Yasin noticed that in an unbalanced interferometer as used in some neutron interferometry experiments, one has partial WPI while keeping interference with limited visibility [@GreenbergerYasin]. The complementary quantities WPI and interference visibility could then be partially determined simultaneously.\ Consistent theoretical analysis of both schemes, independently published by Jaeger et al. [@Jaeger_PRA1995] and by Englert [@Englert] leads to the inequality [@inequality] $$\label{ComplIneq} V^{2}+D^{2}\leq 1$$ which puts an upper bound to the maximum values of simultaneously determined interference visibility $V$ and path distinguishability $D$, a parameter that quantifies the available WPI on the quantum system.\ The all-or-nothing cases $(V=1,D=0)$ or $(V=0,D=1)$ [@Tonomura; @Neutrons; @Carnal; @Keith; @Arndt; @Grangier; @Kurtsiefer; @Jacques] obviously fulfill inequality . Intermediate situations, corresponding to partial WPI and reduced visibility, have been investigated using atoms [@Rempe], nuclear spins [@Peng] and faint laser light [@attenuated; @light; @pulses]. However none of them has been realized in the delayed-choice scheme. We report here an experimental test of the complementarity inequality in intermediate regimes using true single-photon pulses, and in the delayed-choice operation mode.\ Following Englert [@Englert], we point out that the distinguishablility $D$ constrained by inequality actually corresponds to two different notions. The a priori distinguishability, also called “predictability", refers to a WPI obtained by using an unbalanced interferometer with different particle flux along the two paths. Only the case where path distinguishability is introduced a posteriori, i.e. after the entering of the particles into the interferometer, offers the opportunity of a delayed choice test of complementarity. This a posteriori distinguishability can be introduced either by creating entanglement between the particle and a which-path marker [@Haroche; @ScullyNature] or by using an interferometer with an unbalanced output beamsplitter [@Rempe]. We have chosen the latter case by implementing the scheme depicted on Fig. 1, where a single-photon pulse is sent into a Mach-Zehnder interferometer with a variable output beamsplitter (VBS) of adjustable reflection coefficient $R$. When $R$ is not $0.5$, one can have some WPI by observing which detector ($\rm P_{1}$ or $\rm P_{2}$) is fired. The choice of introducing or removing this beamsplitter is random and relativistically space-like separated from the entering of the photon inside the interferometer, as required for the delayed-choice regime.\ ![Theroretical sketch of the delayed-choice complementarity-test experiment. A single-photon pulse is sent into a Mach-Zehnder interferometer, composed of a 50/50 input beamsplitter (BS) and a variable output beamsplitter (VBS). The reflection coefficient is randomly set either to the null value or to an adjustable value $R$, after the photon has entered the interferometer. The single-photon photodetectors $\rm P_{1}$ and $\rm P_{2}$ allow to record both the interference and the WPI. []{data-label="FigSetup"}](DCEvarth.ps){width="8cm"} The experiment starts from a clock-triggered single-photon source, based on the photoluminescence of a single N-V color center in a diamond nanocrystal [@Kun]. The linearly polarized single-photon pulses are then directed to a polarization Mach-Zehnder interferometer described in Ref. [@DCECachan]. The input polarization beamsplitter $\rm BS$ splits the light pulse into two spatially separated components of equal amplitudes, associated with orthogonal S and P polarizations. The two beams then propagate in free space for 48 m, which corresponds to a time of flight of 160 ns.\ The variable output beamsplitter VBS is the association of a polarization beamsplitter (PBS) which spatially overlaps the two beams, an electro-optical modulator (EOM) which acts as an adjustable waveplate, and a Wollaston prism (WP) with its polarization eigenstates corresponding to the S and P polarized channels of the interferometer (Fig. \[FigVBS\]). Given the relative orientation $\beta$ of the EOM, the VBS reflection coefficient $R$ depends on the voltage $V_{\rm EOM}$ applied to the EOM, according to the relation: $$R=\sin^{2} 2\beta \times \sin^{2}\left(\frac{\pi}{2} \frac{ V_{\rm EOM}}{ V_{\pi}}\right) \label{defR}$$ where $V_{\pi}$ is the half-wave voltage of the EOM. The parameters $\beta$ and $V_{\pi}$ have been independently measured for our experimental conditions, and found equal to $\beta=24\pm 1°$ and $V_{\pi}=217\pm 1\ \rm V$ at the wavelengh $\lambda=670 \ \rm nm$ which is the emission peak of the negatively charged N-V color center [@Kun]. This allows $R$ to vary between $0$ and $0.5$ when $V_{\rm EOM}$ is varied between $0$ and $170$ V.\ When $R=0$, the VBS is equivalent to a perfectly transparent (or absent) beamsplitter. Then, each “click” of one of the two photodetectors ($\rm P_{1}$ or $\rm P_{2}$) placed on the output ports of the interferometer is associated to a specific path. It then gives access to the full WPI ($D=1$) and no interference effect will be observed ($V=0$). When $R\neq0$, paths $1$ and $2$ are partially recombined by the VBS. The WPI is then partially washed out, up to be totally erased when $R=0.5$. On the other hand, interference can be observed when the dephasing $\Phi$ between paths 1 and 2 is varied. The experiment will consist in checking the relation between $D$ and $V$ for a given value of $R$, controlled by the EOM voltage $V_{\rm EOM}$. ![Variable output beamspltter (VBS) implementation. The optical axis of the polarization beamsplitter (PBS) and the polarization eigenstates of the Wollaston prism (WP) are aligned, and make an angle $\beta$ with the optical axis of the EOM. The voltage $V_{\rm EOM}$ applied to the EOM is randomly chosen accordingly to the output of a Quantum Random Number Generator (QRNG), located at the output of the interferometer and synchronized on the 4.2 MHz clock that triggers the single-photon emission. []{data-label="FigVBS"}](DCE-VBS.ps){width="8cm"} In order to perform the experimental test of complementarity in the delayed-choice regime, the choosen configuration of the interferometer, defined by $R$, has to be causally isolated from the entering of the photon into the interferometer. This condition is ensured by a relativistically space-like separated random choice. For each measurement, the value of the reflection coefficient of VBS is randomly chosen between $0$ and a given value of $R$, using a quantum random number generator (QRNG) located at the output of the interferometer (Fig. 2) [@DCECachan]. The random numbers are generated from the amplified shotnoise of a white light beam which is an intrinsic quantum random process. At the experiment clock-frequency, i.e. every $\tau_{\rm rep}=238$ ns, fast comparison of the amplified shotnoise to the zero level generates a binary random number $0$ or $1$ which changes the VBS reflectivity between $0$ and $R$, by applying or not the corresponding voltage to the EOM (see Eq. ). In the laboratory framework, the random choise is realized simultaneously with the entering of the photon into the interferometer, ensuring the required space-like separation [@DCECachan].\ As meaningfull illustration of complementarity requires the use of single particles, the quantum behavior of the light field is first tested using the two output detectors feeding single and coincidence counters with no voltage applied to the EOM. In this situation of an absent output beamsplitter, we measure the correlation parameter $\alpha$ [@Grangier; @Jacques], which is equivalent to the second order correlation function at zero delay $g^{(2)}(0)$. For an ideal single-photon source, quantum optics predicts a perfect anticorrelation $\alpha=0$, in agreement with the particle-like image that the photon cannot be detected simultaneously in the two paths of the interferometer. With our source [@Kun], we find $\alpha = 0.15 \pm 0.01$, a value much smaller than one, showing that we are indeed close to the pure single-photon regime [@alpha]. ![Interference visibility $V$ measured in the delayed choice regime for different value of $V_{\rm EOM}$ applied randomly to the EOM. (a), (b), (c) correspond respectively to $V_{\rm EOM}\approx 150 \ $V ($R=0.43$ and $V=93\pm 2\%$), $V_{\rm EOM}\approx 40$ V ($R=0.05$ and $V=42\pm 2\%$) and $V_{\rm EOM}=0$ ($R=0$ and $V=0$). Each point is recorded with $1.9$ s acquisition time. Detectors dark counts, of about $60 \ \rm s^{-1}$ for each, have been substracted to the data.[]{data-label="Figresult"}](fig3_bisPRL.ps){width="8cm"} The delayed-choice test of complementarity with single-photon pulses is performed with the EOM randomly switched for each photon sent in the interferometer, corresponding to a random choice between two values $0$ and $R$ of the VBS reflectivity. The phase-shift $\Phi$ between the two arms of the interferometer is varied by tilting the polarization beamsplitter PBS of VBS with a piezoelectric actuator (see Fig.\[FigVBS\]). For each photon, we record the chosen configuration of the interferometer, the detection events, and the actuator position. All raw data are saved in real time and are processed only after a run is completed. The events corresponding to each configuration of the interferometer are finally sorted. For a given value $R$, the wave-like information of the light field is obtained by measuring the visibility of the interference, predicted to be $$V=2\sqrt{R(1-R)} \ . \label{defV}$$ ![\[Figresult\] Delayed-choice test of complementarity with single-photon pulses. (a)-Wave-like information $V^{2}$ and which-path information $D^{2}$ as a function of the EOM voltage corresponding to a given value $R$ of the VBS reflectivity. The solid lines are the theoretical expectations, with $\beta=24 °$ and $V_{\pi}=217$, using Eqs , and . (b)- $V^{2}+D^{2}$ as a function of the EOM voltage.](fig4_PRL.ps){width="8cm"} The results, depicted in Fig. 3, show a reduction of $V$ when the value of $R$ randomly applied decreases.\ To test inequality , a value of the distinguishability $D$ is then required, to qualitatively qualify the amount of WPI which can be extracted for each value of $R$. We introduce the quantity $D_{1}$ and $D_{2}$, respectively associated to the WPI on path 1 and path 2: $$\begin{aligned} D_{1}=&\left\|p({\rm P}_{1},{\rm path} \ 1)-p(\rm P_{2},\rm path \ 1)\right\|\\ D_{2}=&\left\|p({\rm P}_{1},{\rm path} \ 2)-p(\rm P_{2},\rm path \ 2)\right\| \label{defI1}\end{aligned}$$ where $p(\rm P_{i},\rm path \ j)$ is the probability that the particle follows path j and is detected on detector $\rm P_{i}$. For a single particle arriving on the output beamsplitter, one obtains $$D_{1}=D_{2}=\frac{1}{2}-R \ .$$ The distinguishability parameter $D$ is finally defined as [@Englert] $$D=D_{1}+D_{2}=1-2 R \ . \label{Predic}$$ In order to test this relation, we estimate the values of $D_{1}$ and $D_{2}$ by blocking one path of the interferometer and measuring the number of detections $N_{1}$ and $N_{2}$ on detectors $\rm P_{1}$ and $\rm P_{2}$, which are statisticallly related to $D_{1}$ and $D_{2}$ according to [@Rempe; @DCECachan] : $$\begin{aligned} D_{1}=&\left.\left\|\frac{N_{1}-N_{2}}{N_{1}+N_{2}} \right\| \right]_{\text{path 2 blocked}}\\ D_{2}=&\left.\left\|\frac{N_{1}-N_{2}}{N_{1}+N_{2}} \right\| \right]_{\text{path 1 blocked}}. \end{aligned}$$ These measurements are also performed in the delayed-choice regime, using the procedure described above. We finally obtain independent measurement of $D$ and $V$ for different values of the reflection coefficient $R$, randomly applied to the interferometer. The final results, depicted on Fig. 4, leads to $V^{2}+D^{2}=0.97 \pm 0.03$, close to the ideal balance between $V$ and $D$ constrained by inequality , eventhough each quantity varies from $0$ to $1$.\ The effects observed in this delayed-choice experiment are in perfect agreement with quantum mechanics predictions. No change is observed between a so called “normal-choice” experiment and the “delayed-choice” version. It demonstrates that the complementarity principle cannot be interpreted in a naive way, assuming that the photon at the input of the interferometer could adjust its nature according to the experimental setup installed. As Bohr pointed out , “it obviously can make no difference as regards observable effects obtainable by a definite experimental arrangement, whether our plans of constructing or handling the instrument are fixed beforehand or whether we prefer to postpone the completion of our planning until a later moment when the particle is already on its way from one instrument to another”. Such intriguing property of quantum mechanics forces one to renounce to some common-sense representations of the physical reality.\ We warmly thank A. Clouqueur and A. Villing for the realization of the electronics, and J.-P. Madrange for all mechanical realization of the interferometer. This work is supported by Institut Universitaire de France. N. Bohr, Naturwissenschaften 16, 245 (1928). R. P. Feynman, R. B. Leighton, and M. L. Sands, Lectures on Physics (Addison Wesley, 1963). We stick to the original discussions on complementarity for single particles and do not consider schemes with pairs of entangled particles, where a measurements on one particle allows one to obtain which-path information on the other particle. A. Tonomura, J. Endo, T. Matsuda, T. Kawasaki, and H. Ezawa, Am. J. Phys. 57, 117 (1989). J. Summhammer, G. Badurek, H. Rauch, U. Kischko, and A. Zeilinger, Phys. Rev. A 27, 2523 (1983). O. Carnal and J. Mlynek, Phys. Rev. Lett. 66, 2689 (1991). D. S. Keith et al., Phys. Rev. Lett. 66, 2693 (1991). M. Arndt et al., Nature 401, 680 (1999). T. Pfau, S. Spalter, Ch. Kurtsiefer, C. R. Ekstrom, and J. Mlynek, Phys. Rev. Lett. 73, 1223 (1994). M. S. Chapman et al., Phys. Rev. Lett. 75, 3783 (1995). E. Buks, R. Schuster, M. Heiblum, D. Mahalu and V. Umansky, Nature 391, 871-874 (1998). S. Durr, T. Nonn, and G. Rempe, Nature 395, 33 (1998). P. Bertet et al., Nature 411, 166 (2001). P. Grangier, G. Roger, and A. Aspect, Europhys. Lett. 1, 173 (1986). C. Braig, P. Zarda, C. Kurtsiefer, and H. Weinfurter, Appl. Phys. B 76, 113 (2003). V. Jacques et al., Eur. Phys. J. D 35, 561 (2005). G. Greenstein and A. G. Zajonc, The Quantum Challenge, 2nd edition (Jones and Bartlett Publishers, 2005). J. A. Wheeler, pp. 182-213 in Quantum Theory and Measurement, J. A. Wheeler and W. H. Zurek eds (Princeton University Press, 1984). T. Hellmut, H. Walther, A. G. Zajonc, and W. Schleich, Phys. Rev. A 35, 2532 (1987). J. Baldzuhn, E. Mohler, and W. Martienssen, Z. Phys. B 77, 347 (1989). B. J. Lawson Daku et al., Phys. Rev. A 54, 5042 (1996). V. Jacques et al., Science [315]{}, 966 (2007). W. K. Wootters and W. H. Zurek, Phys. Rev. D 19, 473 (1979). D. M. Greenberger and A. Yasin, Physics Letters A [128]{}, 391 (1988). G. Jaeger, A. Shimony and L. Vaidman, Phys. Rev. A 51, 54-67 (1995). B.-G. Englert, Phys. Rev. Lett. [77]{}, 2154 (1996). A similar inequality is found in P. Grangier, Thèse d’état (1986), Institut d’Optique et Université Paris 11; available online at http://tel.ccsd.cnrs.fr/tel-00009436. S. Durr, T. Nonn, and G. Rempe, Phys. Rev. Lett. 81, 5705 (1998). X. Peng et al., J. Phys. A 36, 2555 (2003). P. Mittelstaedt, A. Prieur, and R. Schieder, Found. Phys. 17, 891 (1987). P. D. Schwindt, P. G. Kwiat, and B.-G. Englert, Phys. Rev. A 60, 4285 (1999). As stressed in [@Rempe], these experiments can actually be fully interpreted in the framework of classical electrodynamics since they do not correspond to the case of true single-photon pulses (see Ref. [@Grangier]). M. O. Scully, B. G. Englert, and H. Walther, Nature 351, 111 (1991). . A. Beveratos et al., Eur. Phys. J. D 18, 191 (2002). The non-ideal value of the $\alpha$ parameter is due to residual background photoluminescence of the diamond sample and to its two-phonon Raman scattering line, which both produce uncorrelated photons associated to Poissonian statistics. N. Bohr, in Albert Einstein: Philosopher Scientist, 2nd ed., edited by P. A. Schilpp (Library of Living Philosophers, Evanston, IL, 1951), p.230.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We reexamine the simplified dark matter (DM) models with fermionic DM particle and spin-0 mediator. The DM-nucleon scattering cross sections of these models are low-momentum suppressed at tree-level, but receive sizable loop-induced spin-independent contribution. We perform one-loop calculation for scalar-type and twist-2 DM-quark operators and complete two-loop calculation for scalar-type DM-gluon operator. By analyzing the loop-level contribution from new operators, we find that future direct detection experiments can be sensitive to a fraction of parameter space. The indirect detection and collider search also provide complementary constraints on these models.' author: - Tong Li - Peiwen Wu bibliography: - 'refs.bib' title: Simplified dark matter models with loop effects in direct detection and the constraints from indirect detection and collider search --- Introduction {#sec:Intro} ============ Although the existence of Dark Matter (DM) has been established by substantial cosmological and astronomical observations, the microscopic nature of DM particles is still unknown. An appealing candidate of DM is the Weakly Interacting Massive Particle (WIMP) arising from various extensions of the Standard Model (SM). The experimental searches for WIMP consist of four main categories, i.e. the direct detection (DD) of possible scattering between DM and SM target materials, the indirect detection (ID) looking for signals of DM annihilation/decay products from the sky, the collider searches for signals from DM productions at high energy accelerators, and the gravitational and/or cosmological effects originating from the DM in the early and/or the current Universe. Among the aforementioned four categories, the DD experiments have achieved significantly improved sensitivity in the past two decades, but yield null results up to now and very stringent bounds on the WIMP-nucleon scattering cross section. A natural explanation of the absence of a confirmed DM signal is that the scattering rate is highly suppressed by the small typical value of transfer momentum of the process and/or the relative velocity between DM and nucleon. A simple but compelling scenario resulting in the suppressed rate at tree-level is that fermionic DM particles $\chi$ scatter off the target nucleon $N$ through a pseudo-scalar mediator in the $t$-channel scattering process [@Boehm:2014hva; @Ipek:2014gua]. The corresponding tree-level DM-nucleon contact interaction can reduce down to a non-relativistic contact operator $$\begin{aligned} \bar{\chi}i\gamma_5\chi N i\gamma_5 N \to (\mathbf{s}_\chi\cdot \mathbf{q})(\mathbf{s}_N\cdot \mathbf{q}),\end{aligned}$$ in the non-relativistic limit. Here, $\mathbf{s}_\chi$ ($\mathbf{s}_N$) is the DM (target nucleon) spin, and the scattering exchange momentum $\mathbf{q}$ is only of order 10 MeV. As a result, this scenario leads to a momentum-suppressed spin-dependent (SD) scattering cross section and thus an undetectable signal rate. The suppression of tree-level scattering rate makes it appealing to further scrutinize the high-order effect from one-loop induced processes (see early discussions in e.g. [@Drees:1993bu; @Freytsis:2010ne]) and the possibly detectable signal at the upgrades of DD experiments. Integrating out the one-loop diagrams can induce distinct scalar-type operators giving non-momentum-suppressed spin-independent (SI) scattering cross section. The enhancement of loop-level SI cross section by the squared total nucleon number in a nucleus competes with the loop suppression and may dominate the WIMP-nuclei cross section over the suppressed tree-level scattering. This one-loop effect in direct DM detection has been investigated in both simplified frameworks and UV complete models [@Haisch:2013uaa; @Ipek:2014gua; @Arcadi:2017wqi; @Sanderson:2018lmj; @Li:2018qip; @Han:2018gej; @Abe:2018bpo; @Ghorbani:2018pjh; @Mohan:2019zrk]. Recent progresses on the pseudo-scalar mediator scenario go beyond the one-loop processes for scalar-type DM-quark operator $m_q\bar{\chi}\chi \bar{q}q$ and include the dedicated contributions from two-loop scattering diagrams for scalar-type DM-gluon operator ${\alpha_s\over \pi}\bar{\chi}\chi GG$ after integrating out both heavy quarks and the mediator. It has been shown in Refs. [@Abe:2018emu; @Ertas:2019dew] that the full two-loop calculation deviates sizably from the result obtained by the conventional relation between the scalar-type current of heavy quarks and that of the gluon. This discrepancy is caused by the failure of the quark momentum expansion for heavy quarks and an ignored two-loop diagram for gluon emission when one utilized the relation for DM-gluon scattering. In this work we revisit the loop effect in the DD of simplified DM models including but not limited to a pseudo-scalar mediator. We consider the hypotheses with one spin-0 mediator only coupled to the SM quarks and fermionic DM particles, giving momentum-suppressed WIMP-nuclei scattering cross sections at tree-level. The latest approaches of dedicated loop calculations are utilized in the high-order contributions to the cross sections, together with the estimate of the Renormalization Group Equations (RGE) running effects. We also take into account the constraints from other DM detection categories, e.g. the DM relic abundance, the ID constraint in terms of gamma-ray emission, as well as the current status of collider search. These synergistic discussions are regarded as a more complete improvement of the recently appeared works [@Abe:2018emu; @Ertas:2019dew]. The paper is outlined as follows. In Sec. \[sec:Model\] we describe the simplified dark matter models. Then we give the effective DM-nucleon interactions at tree-level and the corresponding DM-nucleus scattering cross sections. In Sec. \[sec:Loop\], we present the effective operators and Wilson coefficients at loop-level for the DM-nucleon cross section. The numerical results are given in Sec. \[sec:Result\]. Our conclusions are drawn in Sec. \[sec:Con\]. Some technical details for loop calculations are collected in the Appendix. Simplified dark matter hypothesis {#sec:Model} ================================= In this work, we focus on the simplified DM frameworks which consist of Majorana fermion DM $\chi$ and a spin-0 mediator $a$ coupled to $\chi$ and the SM quarks with strength $g_\chi$ and $g_q$ respectively. We consider each of the following three scenarios at a time $$\begin{aligned} &&\mathcal{L}_{\rm D2} = -{g_{\chi}\over 2} a \bar{\chi} i\gamma_5 \chi -g_q {m_q\over v_0} a \bar{q} q, \\ &&\mathcal{L}_{\rm D3} = -{g_{\chi}\over 2} a \bar{\chi} \chi -g_q {m_q\over v_0} a \bar{q} i\gamma_5 q, \\ &&\mathcal{L}_{\rm D4} = -{g_{\chi}\over 2} a \bar{\chi} i\gamma_5 \chi -g_q {m_q\over v_0} a \bar{q} i\gamma_5 q.\end{aligned}$$ Here the $a\bar{q}q$ coupling is also scaled by the SM-like Yukawa coupling with $v_0 = 246$ GeV being the SM Higgs vacuum expectation value. The model D4 with $a$ being a pure pseudo-scalar is designated as the pseudo-scalar mediator DM model in most of literatures. Models D2 and D3 are induced by more specific UV complete models with CP violation [@LopezHonorez:2012kv; @Beniwal:2015sdl; @Baek:2017vzd; @Athron:2018hpc; @Abe:2019wku; @Ertas:2019dew] and correspond to the cases with specific CP angles in Ref. [@Ertas:2019dew]. Based on the DM interactions with quarks and gluons at tree-level, the DM-nucleon contact interactions are described by the effective Lagrangians as follows $$\begin{aligned} &&\mathcal{L}^{eff}_{\rm D2} = {C_N^{\rm tree}({\rm D2})\over 2m_{a}^2} \bar{\chi} i\gamma_5 \chi \bar{N} N, \ \mathcal{L}^{eff}_{\rm D3} = {C_N^{\rm tree}({\rm D_3})\over 2m_a^2} \bar{\chi} \chi \bar{N} i\gamma_5 N, \ \mathcal{L}^{eff}_{\rm D4} = {C_N^{\rm tree}({\rm D_4})\over 2m_a^2} \bar{\chi} i\gamma_5 \chi \bar{N} i\gamma_5 N, \nonumber \\\end{aligned}$$ where the tree-level coefficients are defined as $$\begin{aligned} &&C_N^{\rm tree}({\rm D2}) = \sum_{q=u,d,s}{m_N\over m_q} C_q f_q^{N} + \sum_{q=c,b,t}{m_N\over m_q} C_q {2\over 27} f_G^{N}, \\ &&C_N^{\rm tree}({\rm D4}) = \sum_{q=u,d,s}{m_N\over m_q} \left(C_q-C\right) \Delta_q^{N}, \ \ \ C = \bar{m}\sum_{q=u,\cdots,t}{C_q\over m_q}, \ \ \ \bar{m}^{-1}=\sum_{q=u,d,s}m_q^{-1}.\end{aligned}$$ The coefficient $C_N^{\rm tree}({\rm D_3})$ is equal to $C_N^{\rm tree}({\rm D_4})$ as models D3 and D4 share the same quark bilinear form $\bar{q}\gamma_5 q$. Here $\Delta_q^{N}$, $f_q^{N}$ and $f_G^{N}$ are quark/gluon-nucleon form factors as numerically used in micrOMEGAs [@Belanger:2013oya]. The quark-level constant is defined as $C_q = g_\chi g_q{m_q\over v_0}$. Consequently, their differential DM-nucleus scattering cross sections read as $$\begin{aligned} &&{d\sigma_{\rm SI}({\rm D2})\over dE_R} = {1\over 32\pi} {m_T\over m_\chi^2 m_N^2 v^2} {4m_N^2q^2\over m_a^4} \sum_{N,N'=p,n}C_N^{\rm tree}({\rm D_2}) C_{N'}^{\rm tree}({\rm D_2}) F_{M}^{(N,N')}(q^2) , \label{D2xsectree} \\ &&{d\sigma_{\rm SD}({\rm D3})\over dE_R} = {1\over 32\pi} {m_T\over m_\chi^2 m_N^2 v^2} {4m_\chi^2 q^2\over m_a^4} \sum_{N,N'=p,n}C_N^{\rm tree}({\rm D_3}) C_{N'}^{\rm tree}({\rm D_3}) F_{\Sigma''}^{(N,N')}(q^2) , \label{D3xsectree}\\ &&{d\sigma_{\rm SD}({\rm D4})\over dE_R} = {1\over 32\pi} {m_T\over m_\chi^2 m_N^2 v^2} {q^4\over m_a^4} \sum_{N,N'=p,n}C_N^{\rm tree}({\rm D_4}) C_{N'}^{\rm tree}({\rm D_4}) F_{\Sigma''}^{(N,N')}(q^2) , \label{D4xsectree}\end{aligned}$$ where $m_T$ is the nucleus mass, $v$ is the DM speed in Earth’s frame, $E_R$ is the nuclear recoil energy and $F_{\Sigma''}^{(N,N')}(q^2), F_{M}^{(N,N')}(q^2)$ are the form factors defined in Ref. [@Fitzpatrick:2012ix]. The tree-level WIMP-nucleus scattering cross sections of the above simplified models are all dependent on the transfer momentum $q=\sqrt{2E_R m_T}$. As seen above, they are suppressed by $m_N^2q^2/ m_a^4$, $m_\chi^2 q^2/ m_a^4$ and $q^4/m_a^4$ for models D2, D3 and D4, respectively. Loop effect in direct detection {#sec:Loop} =============================== In this section, we derive the loop-level effect in direct DM detection of the above simplified models, followed by the estimate of the scale effects in terms of renormalization group evolutions. Loop effect from scalar-type quark/gluon operators -------------------------------------------------- The general Lagrangian for the non-momentum-suppressed DM-nucleon SI cross section is given by $$\begin{aligned} \mathcal{L}_{eff} &=& {1\over 2}\sum_{q=u,d,s}C'_q m_q \bar{\chi}\chi \bar{q}q + {1\over 2}C_G\left(-{9\alpha_s\over 8\pi}\bar{\chi}\chi G^a_{\mu\nu}G^{a\mu\nu}\right)\nonumber \\ &+& {1\over 2}\sum_{q=u,d,s,c,b}\left[C_q^{(1)}\bar{\chi}i\partial^\mu \gamma^\nu \chi \mathcal{O}^q_{\mu\nu}+C_q^{(2)}\bar{\chi}i\partial^\mu i\partial^\nu \chi \mathcal{O}^q_{\mu\nu}\right], \label{Lagrangian}\end{aligned}$$ where $\mathcal{O}^q_{\mu\nu}={i\over 2}\bar{q}\left(\partial_\mu\gamma_\nu+\partial_\nu\gamma_\mu-{1\over 2}g_{\mu\nu}\cancel{\partial}\right)q$ is the twist-2 operator. For the models we consider, the Wilson coefficients in Eq. (\[Lagrangian\]) are all zero at tree-level but can be generated at loop-level, denoted by $C'_q=C_q^{\rm box}$, $C_G=C_G^{\rm box}$, $C_q^{(1)}=C_q^{\rm (1)box}$, $C_q^{(2)}=C_q^{\rm (2)box}$. The coefficients for scalar-type DM-quark operator and the twist-2 operator, i.e. $C_q^{\rm box}$, $C_q^{\rm (1)box}$ and $C_q^{\rm (2)box}$, are generated by the box diagrams in the top panels of Fig. \[loops\]. The two-loop diagrams in Fig. \[loops\] with only heavy quark $Q$ in the quark loop contribute to the scalar-type DM-gluon operator and the coefficient $C_G^{\rm box}$. ![Loop diagrams for the DM-quark currents (top) and DM-gluon current (bottom). []{data-label="loops"}](plots/loops.jpg){width="14cm"} Following the non-relativistic limit used in Ref. [@Abe:2018emu], we expand the small momentum of valence quarks in the amplitude of the DM-quark scattering box diagrams. The coefficients $C_q^{\rm box}$, $C_q^{\rm (1)box}$ and $C_q^{\rm (2)box}$ are then obtained by reading out the DM-quark effective operators. For the DM-gluon coefficient $C_G^{\rm box}$, one needs to calculate the amplitude of two-loop diagrams and find the effective operator $\bar{\chi}\chi G^a_{\mu\nu}G^{a\mu\nu}$. The complete two-loop calculations ensure the validity of the obtained $C_G^{\rm box}$ for any values of mediator mass $m_a$. For model D4, the above Wilson coefficients are equivalent to those for pseudo-scalar mediator model as derived in Ref. [@Abe:2018emu] $$\begin{aligned} C_q^{\rm box}({\rm D4})&=&{-m_\chi\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^2} [6X_{001}(m_\chi^2,m_\chi^2,0,m_a^2)+m_\chi^2X_{111}(m_\chi^2,m_\chi^2,0,m_a^2)\nonumber \\ &-&6X_{001}(m_\chi^2,m_\chi^2,m_a^2,0)-m_\chi^2X_{111}(m_\chi^2,m_\chi^2,m_a^2,0)],\\ C_q^{\rm (1)box}({\rm D4})&=&{-8\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^2} [X_{001}(m_\chi^2,m_\chi^2,0,m_a^2)-X_{001}(m_\chi^2,m_\chi^2,m_a^2,0)],\\ C_q^{\rm (2)box}({\rm D4})&=&{-4m_\chi\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^2} [X_{111}(m_\chi^2,m_\chi^2,0,m_a^2)-X_{111}(m_\chi^2,m_\chi^2,m_a^2,0)],\\ C_G^{\rm box}({\rm D4})&=& \sum_{Q=c,b,t}{-m_\chi\over 432\pi^2}\left({m_Q\over v_0}\right)^2 g_\chi^2 g_q^2 {\partial F(m_a^2)\over \partial m_a^2},\end{aligned}$$ where the loop functions $X_{001}, X_{111}, F$ are given in Ref. [@Abe:2018emu] and the references therein. Following the same procedure, for models D2 and D3, we obtain the corresponding Wilson coefficients which are related to those in model D4 $$\begin{aligned} C_q^{\rm box}({\rm D2})&=&C_q^{\rm box}({\rm D4})\nonumber \\ &+&{4m_\chi\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^2}[C_2(m_\chi^2,m_a^2,m_\chi^2)+{1\over m_a^2}B_1(m_\chi^2,0,m_\chi^2)-{1\over m_a^2}B_1(m_\chi^2,m_a^2,m_\chi^2)],\nonumber \\ \\ C_q^{\rm box}({\rm D3})&=&C_q^{\rm box}({\rm D4})+{-8m_\chi\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^4}[X_{00}(m_\chi^2,m_a^2,m_\chi^2)+{m_\chi^2\over 4}X_{11}(m_\chi^2,m_a^2,m_\chi^2)],\\ C_q^{\rm (1)box}({\rm D2})&=&C_q^{\rm (1)box}({\rm D3})=C_q^{\rm (1)box}({\rm D4}), \ \ \ C_q^{\rm (2)box}({\rm D2})=C_q^{\rm (2)box}({\rm D4}),\\ C_q^{\rm (2)box}({\rm D3})&=&C_q^{\rm (2)box}({\rm D4})+{-8m_\chi\over (4\pi)^2}\left({m_q\over v_0}\right)^2 {g_\chi^2 g_q^2\over m_a^4}X_{11}(m_\chi^2,m_a^2,m_\chi^2),\\ C_G^{\rm box}({\rm D2})&=&\sum_{Q=c,b,t}{-m_\chi\over 432\pi^2}\left({m_Q\over v_0}\right)^2 g_\chi^2 g_q^2 {\partial F'(m_a^2)\over \partial m_a^2},\\ C_G^{\rm box}({\rm D3})&=&\sum_{Q=c,b,t}{-m_\chi\over 432\pi^2}\left({m_Q\over v_0}\right)^2 g_\chi^2 g_q^2 {\partial F''(m_a^2)\over \partial m_a^2}.\end{aligned}$$ The new loop functions here are collected in Appendix. In our numerical calculation, we use Package-X [@Patel:2015tea] to compute the above loop functions. Based on the above effective operators for SI DM-nucleon scattering and the corresponding Wilson coefficients, we define the DM-nucleon constant at loop-level $$\begin{aligned} C_N^{\rm loop}&=&m_N\left[\sum_{q=u,d,s}C_q' f_{q}^N + C_G f_{G}^N + {3\over 4}\sum_{q=u,d,s,c,b}\left(m_\chi C_q^{(1)}+m_\chi^2C_q^{(2)}\right)\left(q^N(2)+\bar{q}^N(2)\right)\right], \label{WCDMnucleon}\end{aligned}$$ where the second moments of the parton distribution functions for quarks $q^N(2)$ and anti-quarks $\bar{q}^N(2)$ are taken from the CTEQ PDFs [@Pumplin:2002vw]. The SI cross section of the DM interaction with nucleon is thus given by $$\begin{aligned} \sigma_{\rm SI}={1\over \pi}\left({m_\chi m_N\over m_\chi + m_N}\right)^2 \|C_N^{\rm loop}\|^2. \label{sigmaloop}\end{aligned}$$ In terms of the form factor function $F_{M}^{(N,N')}(q^2)$, the differential SI cross section of the DM interaction with nucleus with mass $m_T$ is $$\begin{aligned} {d\sigma_{\rm SI}\over dE_R} = {1\over 2\pi} {m_T\over v^2} \sum_{N,N'=p,n}C_N^{\rm loop} C_{N'}^{\rm loop} F_{M}^{(N,N')}(q^2). \label{xsecloop}\end{aligned}$$ Loop effect from RGE running ---------------------------- Another manifestation of loop effect is the mixing of operators according to the RGE. This RGE effect can be important if one considers DM phenomenology at vastly different energy scales. For instance, the DM annihilation typically happens at the electroweak scale and the DM particles are possibly produced near TeV scale at colliders. The energy scale for the DM-nucleon scattering in DD experiments is of the order of the hadron scale $\mu_{had}$. One usually starts with a gauge-invariant renormalizable DM model defined near or above the electroweak (EW) scale $\mu_{EW}\simeq m_Z$, but studies the non-relativistic DM-nucleon scattering rate with characteristic scale $\mu_{had}\simeq 1$ GeV. A series of effective field theories (EFTs) should be properly constructed by integrating out particles heavier than the current EFT scale $\mu_{EFT}$ and reasonably matched when passing thresholds of particles lighter than $\mu_{EFT}$ where they are integrated out in a similar way. Between the thresholds, the evolutions and mixings of the EFT operators should be performed according to RGE. The scale of the first EFT constructed in the whole analysis determines the procedures of RGE and threshold matching. The above procedures have been well elaborated in e.g. [@Hisano:2010ct; @Hisano:2015bma; @Hisano:2015rsa; @Hill:2014yka; @Hill:2014yxa] and implemented in packages such as DirectDM [@Bishara:2017nnn], runDM [@DEramo:2016gos], Wilson [@Aebischer:2018bkb] for specific or generic models. An important difference should be emphasized between the scalar-type and twist-2 operators in Eq. (\[Lagrangian\]). The scalar-type form factors $f_q^{N}$ for light quarks $q=u,d,s$ are attributed to the non-perturbative QCD effects with energy scale around or below 1 GeV, and are obtained from the lattice QCD simulations. Thus, the scalar-type Wilson coefficients of light quarks $u,d,s$ and gluon must take values around 1 GeV. Depending on the scale of the first EFT constructed in the whole analysis, e.g. at $\mu=m_Z$, this implies the procedures of RGE and threshold matching when calculating the scalar-type operator contributions. At the scale of about 1 GeV, the heavy quarks $Q=c,b,t$ have been integrated out into the scalar-type gluon operator using the full two-loop calculations, as emphasized in Refs. [@Abe:2018emu; @Ertas:2019dew]. In turn, the twist-2 form factors $q^N(2), \bar{q}^N(2)$ in Eq. (\[Lagrangian\]) can be calculated perturbatively using parton PDFs at various scales [@Hisano:2010ct; @Hisano:2015bma; @Hisano:2015rsa; @Hill:2014yka; @Hill:2014yxa; @Drees:1993bu], e.g. 1 GeV or $m_Z$. One can choose a convenient scale to calculate the twist-2 contributions, with the proper active field contents (e.g. 5 flavor quarks $u,d,s,c,b$ at $\mu=m_Z$) and the Wilson coefficients and form factors evaluated at that scale. Note that we ignored the negligible contributions from twist-2 gluon operator, since its Wilson coefficient is suppressed by an additional $\alpha_s/\pi$ due to the operator definition [@Hisano:2015bma; @Hisano:2015rsa]. We note that in Refs. [@Abe:2018emu; @Ertas:2019dew], the values of $q^N(2), \bar{q}^N(2)$ are evaluated at $\mu=m_Z$ in the calculations. Since the same (similar) box diagrams and model parameters are used to obtain the scalar-type and twist-2 Wilson coefficients for quarks (gluons), a more consistent implementation should involve the RGE running effects for the scale-type operators from $m_Z$ to 1 GeV as discussed above. Since the coupling between the mediator and the SM quarks is chosen to mimic the SM Yukawa structure, the DM-nucleon constant from scalar-type DM-gluon interaction $C_G f_{G}^N$ in Eq. (\[WCDMnucleon\]) is dominated by the top quark loop. The constant $C_G f_{G}^N$ also dominates over the scalar-type and twist-2 DM-quark interactions. To have a conservative estimate of the scale effects on the scalar type DM-gluon operator $-{9\alpha_s\over 8\pi}\bar{\chi}\chi G^a_{\mu\nu}G^{a\mu\nu}$, we utilize the package DirectDM to perform its RGE running from $m_Z$ to 1 GeV. We find that the scale effects give a negative correction of $1\%\sim 2\%$, and thus do not affect our main conclusions in this work. Other dark matter constraints ============================= In this section, we consider DM constraints on the above simplified models from other categories mentioned in the Introduction, including the relic abundance, indirect detection and collider search. Assuming that DM particles have frozen out in the early Universe, as standard thermal relics, they acquire their present abundance through annihilation processes. The pairs of DM particle $\chi$ in the simplified models can either annihilate into SM quark or gluon pairs via $s$-channel processes $\chi \chi\to a\to q\bar{q}, gg$ or annihilate into two mediators $\chi \chi\to a a$ when kinematically allowed [@Arina:2014yna; @Abdallah:2015ter; @Boveia:2016mrp; @Balazs:2017hxh; @Li:2017nac]. The amplitudes of the two annihilation channels are governed by $g_\chi g_q$ and $g_\chi^2$, respectively. We assume all kinematically accessible final states of the DM annihilation and use micrOMEGAs 5.0 [@Belanger:2018mqt] to calculate the relic abundance. Note that the WIMP candidate here may account only a fraction of the total DM of the Universe, referred as multi-component DM scenario [@Zurek:2008qg; @Profumo:2009tb]. In this scenario, the DM energy density measured by PLANCK [@Ade:2015xua] is imposed as an upper limit on the WIMP relic abundance. Dwarf galaxies are the search targets for DM annihilation into gamma rays. The Fermi Large Area Telescope (LAT) has detected no excess of gamma ray emission from the dwarf spheroidal satellite galaxies (dSphs) of the Milky Way. Thus, an upper limit on the DM annihilation cross section can be placed from a combined analysis of multiple Milky Way dSphs [@Ackermann:2015zua; @Fermi-LAT:2016uux]. For individual dwarf galaxy target, Fermi-LAT provided tabulated values of delta-log-likelihood as a function of the energy flux bin-by-bin. The gamma ray energy flux from DM annihilation for the $j$th energy bin and the $k$th dwarf is given by $$\begin{aligned} \Phi^E_{j,k}(m_{\chi},\langle \sigma v\rangle,J_k)=\frac{\langle \sigma v\rangle}{16\pi m_{\chi}^2}J_k\int^{E^{\rm max}_j}_{E^{\rm min}_j}E\frac{dN_\gamma}{dE}dE,\end{aligned}$$ where $J_k$ is the J factor for the $k$th dwarf. The energy flux only depends on $m_{\chi}$, $\langle \sigma v\rangle$ and $J_k$, and is thus calculable for the DM annihilation process from the above simplified models. We use the PPPC4DMID package [@Cirelli:2010xx] to obtain the spectrum of photons $dN_\gamma/dE$. The likelihood for $k$th dwarf is $$\begin{aligned} &&\mathcal{L}_k(m_{\chi},\langle \sigma v\rangle,J_k)=\mathcal{L}_J(J_k\|\bar{J}_k,\sigma_k)\prod_j \mathcal{L}_{j,k}(\Phi^E_{j,k}(m_{\chi},\langle \sigma v\rangle,J_k)),\end{aligned}$$ where $\mathcal{L}_{j,k}$ is the tabulated likelihood provided by Fermi-LAT for each dwarf and energy flux. The uncertainty of the J factors is taken into account by profiling over $J_k$ in the likelihood below [@Ackermann:2015zua] $$\begin{aligned} &&\mathcal{L}_J(J_k\|\bar{J}_k,\sigma_k)={1\over \ln(10)J_k\sqrt{2\pi}\sigma_k}\times e^{-(\log_{10}(J_k)-\log_{10}(\bar{J}_k))^2/2\sigma_k^2},\end{aligned}$$ with the measured J factor $\bar{J}_k$ and error $\sigma_k$. A joint likelihood for all dwarfs can then be performed as $$\begin{aligned} \mathcal{L}(m_{\chi},\langle \sigma v\rangle,\mathbb{J})=\prod_k \mathcal{L}_k(m_{\chi},\langle \sigma v\rangle,J_k),\end{aligned}$$ where $\mathbb{J}$ is the set of J factors $J_k$. In our numerical implementation, we adopt the corresponding values of $\mathcal{L}_{j,k}$ and $\bar{J}_k, \sigma_k$ for 19 dwarf galaxies considered in Ref. [@Fermi-LAT:2016uux]. According to the maximum likelihood analysis adopted by Fermi-LAT, the delta-log-likelihood is given by $$\begin{aligned} -2\Delta \ln \mathcal{L}(m_{\chi},\langle \sigma v\rangle)=-2\ln\left({\mathcal{L}(m_{\chi},\langle \sigma v\rangle,\widehat{\widehat{\mathbb{J}}})\over \mathcal{L}(m_{\chi},\widehat{\langle \sigma v\rangle},\widehat{\mathbb{J}})}\right),\end{aligned}$$ where $\widehat{\langle \sigma v\rangle}$ and $\widehat{\mathbb{J}}$ maximize the likelihood at any given $m_{\chi}$, and $\widehat{\widehat{\mathbb{J}}}$ maximizes the likelihood for given $m_{\chi}$ and $\langle \sigma v\rangle$. The 95% C.L. upper limit on the annihilation cross section for a given $m_{\chi}$ is determined by demanding $-2\Delta \ln\mathcal{L}(m_{\chi},\langle \sigma v\rangle)\leq 2.71$. We perform the likelihood analysis and obtain the upper limit using Minuit [@James:1975dr]. Once the annihilation cross section calculated by a certain set of model parameters is larger than the limit, we claim the corresponding parameter values are excluded by Fermi-LAT dSphs. The Large Hadron Collider (LHC) performed the search for DM in association with energetic jet [@Aaboud:2017phn; @Sirunyan:2017jix] or the third generation quarks [@Aaboud:2017rzf; @Aaboud:2017aeu; @Sirunyan:2017xgm] for simplified DM models with spin-0 mediator at $\sqrt{s}=13$ TeV collisions. The most severe limits are from final states with $t\bar{t}$ and missing transverse momentum [@Aaboud:2019yqu]. For model D4 with pseudo-scalar mediator, assuming unitary couplings $g_\chi=g_q=1$, the range of mediator mass between 15 and 25 GeV is excluded [@Aaboud:2017aeu]. This limit is valid for all DM masses as long as $m_a>2m_\chi$ and closely related models D2 and D3 should have very similar collider constraint. Results {#sec:Result} ======= By combining the theoretical calculations of DM-nucleus scattering cross sections with a certain velocity distribution for DM particles, we can calculate WIMP signal rates for DD experiments. The differential event rate with respect to the recoil energy is given by $$\begin{aligned} {dN\over dE_R}={\rho_\chi\over m_\chi}\int d^3v vf(\vec{v}){d\sigma(v,E_R)\over dE_R},\end{aligned}$$ where $\rho_\chi$ is the local DM density which is fixed to be $0.3 \ {\rm GeV}/{\rm cm}^3$. We take a cut-off Maxwell-Boltzmann velocity distribution for $f(\vec{v})$ with the escape velocity as $v_{esc}=544 \ {\rm km}/{\rm s}$. Together with the differential scattering cross sections obtained above, the differential event rate can be evaluated and in practice we employ DMFormFactor [@Fitzpatrick:2012ix; @Anand:2013yka] for the numerical calculation on xenon nucleus $^{129}{\rm Xe}$. The recoil energy spectra of models D2, D3 and D4 are shown in Figs. \[dNdE\_D2\], \[dNdE\_D3\] and \[dNdE\_D4\], respectively. Model D2 leads to SI scattering cross section with a strong enhancement for large nuclei. Thus, although there is momentum suppression at tree-level, the spectrum at tree-level for model D2 still dominates over the loop-level contribution. As being suppressed by $q^4$ at tree-level, in turn, the loop-level spectrum is much greater than that at tree-level for model D4. Although models D3 and D4 both lead to SD cross section at tree-level, this discrepancy is smaller in model D3 as its tree-level scattering cross section is suppressed by $q^2$ only. The loop-level spectrum of model D3 dominates over the tree-level one only in the range of low recoil energy. ![ Recoil energy spectra of model D2 with $m_a=10$ GeV, $g_\chi g_q=1$ and $m_\chi = 40$ GeV (left) or $m_\chi = 400$ GeV (right). The tree-level and loop-level spectra are denoted by red and black curves, respectively. []{data-label="dNdE_D2"}](plots/dNdE_ma10_mchi40_D2.jpg "fig:"){width="7cm"} ![ Recoil energy spectra of model D2 with $m_a=10$ GeV, $g_\chi g_q=1$ and $m_\chi = 40$ GeV (left) or $m_\chi = 400$ GeV (right). The tree-level and loop-level spectra are denoted by red and black curves, respectively. []{data-label="dNdE_D2"}](plots/dNdE_ma10_mchi400_D2.jpg "fig:"){width="7cm"} ![ Recoil energy spectra of model D3, as labeled in Fig. \[dNdE\_D2\]. []{data-label="dNdE_D3"}](plots/dNdE_ma10_mchi40_D3.jpg "fig:"){width="7cm"} ![ Recoil energy spectra of model D3, as labeled in Fig. \[dNdE\_D2\]. []{data-label="dNdE_D3"}](plots/dNdE_ma10_mchi400_D3.jpg "fig:"){width="7cm"} ![Recoil energy spectra of model D4, as labeled in Fig. \[dNdE\_D2\]. []{data-label="dNdE_D4"}](plots/dNdE_ma10_mchi40_D4.jpg "fig:"){width="7cm"} ![Recoil energy spectra of model D4, as labeled in Fig. \[dNdE\_D2\]. []{data-label="dNdE_D4"}](plots/dNdE_ma10_mchi400_D4.jpg "fig:"){width="7cm"} As the recoil spectrum of the SI scattering induced by loop diagrams is dominant in model D4, the prediction of SI DM-nucleon cross section in Eq. (\[sigmaloop\]) can be compared directly to the limits set by DD experiments to yield a bound on $g_\chi g_q$. As shown in Fig. \[D4-g\], for $g_\chi g_q \leq 1$ in model D4, the SI scattering cross sections in the red region are below the XENON1T exclusion limit [@Aprile:2017iyp; @Aprile:2018dbl] but above the neutrino floor. The green region gives cross sections below the neutrino floor. Future DD experiments can thus be more sensitive to the region of $g_\chi g_q > 0.4$ and $m_\chi < 200$ GeV in the case of $m_a=10$ GeV. The cross section of DM annihilation into SM quarks for model D4 is proportional to $m_q^2/v_0^2$, thus the $t\bar{t}$ channel dominates if kinematically allowed. The annihilation to mediator pairs is governed by $m_a/m_\chi$ and plays a crucial role in small $m_\chi$ range. The Fermi-LAT dSphs exclude a majority of parameter region for $m_{\chi}\lesssim 100$ GeV and $m_{\chi}\gtrsim m_t$, as shown in light blue region in Fig. \[D4-g\]. The region around $m_{\chi}\simeq 100$ GeV evades the ID constraint due to the fact that the $\chi\chi\to a\to t\bar{t}$ channel is not kinematically allowed and the annihilation to mediator pairs is suppressed [@Balazs:2017hxh; @Li:2017nac]. Above the black contours in Fig. \[D4-g\], the DM relic abundance satisfies $\Omega h^2\leq 0.12$. ![The region of $g_\chi g_q$ vs. $m_\chi$ below the Xenon1T exclusion limit and above the neutrino floor (red) and the region below the neutrino floor (green) for model D4, with $m_a=10$ GeV (left) and $m_a=30$ GeV (right). Assuming $g_\chi = 1$, the light blue values are excluded by Fermi-LAT dSphs and the DM relic abundance satisfies $\Omega h^2\leq 0.12$ above the black curves. []{data-label="D4-g"}](plots/D4_g_ma10new.jpg "fig:"){width="8cm"} ![The region of $g_\chi g_q$ vs. $m_\chi$ below the Xenon1T exclusion limit and above the neutrino floor (red) and the region below the neutrino floor (green) for model D4, with $m_a=10$ GeV (left) and $m_a=30$ GeV (right). Assuming $g_\chi = 1$, the light blue values are excluded by Fermi-LAT dSphs and the DM relic abundance satisfies $\Omega h^2\leq 0.12$ above the black curves. []{data-label="D4-g"}](plots/D4_g_ma30new.jpg "fig:"){width="8cm"} Conclusion {#sec:Con} ========== We reexamined the loop-level correction to the WIMP-nucleon scattering cross section in the framework of simplified DM models. A spin-0 mediator is assumed to couple with fermionic DM particle and the SM quarks in each model. The cross sections of these models are low-momentum suppressed at tree-level, but receive sizable loop-induced SI contribution. Following the recent progress on the loop-level correction, we perform one-loop calculation for scalar-type and twist-2 DM-quark operators and complete two-loop calculation for scalar-type DM-gluon operator. By including the loop-level SI cross section, we find that future DD experiments can be sensitive to a fraction of parameter space which gives no detectable signal with only tree-level contribution. The sensitivity of DD experiments to these models is also complementary to the constraints from ID and collider search. T. L. would like to thank Tomohiro Abe, Motoko Fujiwara and Junji Hisano for helpful discussion. P. W. would like to thank Chengcheng Han for the beneficial argument. The work of T. L. is supported by “the Fundamental Research Funds for the Central Universities”, Nankai University (Grant Number 63196013). Loop functions ============== The used loop functions for models D2 and D3 are $$\begin{aligned} \int {d^D \ell \over (2\pi)^D} {\ell_\mu \ell_\nu \over [(\ell+p)^2-m_\chi^2](\ell^2-m_a^2)^2\ell^4} = {i\over (4\pi)^2 m_a^4} \left[ g_{\mu\nu} X_{00}(p^2,m_a^2,m_\chi^2)+p_\mu p_\nu X_{11}(p^2,m_a^2,m_\chi^2)\right] \nonumber \\\end{aligned}$$ and $$\begin{aligned} &&F'(m_a^2)=\int^1_0 dx [3Y_1(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)-m_Q^2{-x^2-3x\over x^2(1-x)^2}Y_2(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)\nonumber \\ &&+4m_Q^4{x^2(1-2x)\over x^3(1-x)^3}Y_3(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)],\\ &&F''(m_a^2)=\int^1_0 dx [3Y_1(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)-m_Q^2{9x-5x^2\over x^2(1-x)^2}Y_2(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)\nonumber \\ &&-2m_Q^4{2x^2\over x^3(1-x)^3}Y_3(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)]\nonumber \\ &&+2\int^1_0 dx [3Z_1(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)-m_Q^2{9x-5x^2\over x^2(1-x)^2}Z_2(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)\nonumber \\ &&-2m_Q^4{2x^2\over x^3(1-x)^3}Z_3(m_\chi^2,m_\chi^2,m_a^2,m_Q^2)],\end{aligned}$$ with $$\begin{aligned} &&\int {d^D\ell\over (2\pi)^D}{1\over [(\ell+p)^2-m_\chi^2](\ell^2-m_a^2)[\ell^2-{m_Q^2\over x(1-x)}]}={i\over (4\pi)^2}Z_1(p^2,m_\chi^2,m_a^2,m_Q^2),\\ &&\int {d^D\ell\over (2\pi)^D}{1\over [(\ell+p)^2-m_\chi^2](\ell^2-m_a^2)[\ell^2-{m_Q^2\over x(1-x)}]^2}={i\over (4\pi)^2}Z_2(p^2,m_\chi^2,m_a^2,m_Q^2),\\ &&\int {d^D\ell\over (2\pi)^D}{1\over [(\ell+p)^2-m_\chi^2](\ell^2-m_a^2)[\ell^2-{m_Q^2\over x(1-x)}]^3}={i\over (4\pi)^2}Z_3(p^2,m_\chi^2,m_a^2,m_Q^2).\end{aligned}$$ Here the $Y_i (i=1,2,3)$ functions are shown in Ref. [@Abe:2018emu] and the references therein.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss how much we can probe the effective number of neutrino species $N_\nu$ with cosmic microwave background alone. Using the data of WMAP, ACBAR, CBI and BOOMERANG experiments, we obtain a constraint on the effective number of neutrino species as $0.96< N_\nu <7.94$ at 95%C.L. for a power-law $\Lambda$CDM flat universe model. The limit is improved to be $1.39 < N_\nu < 6.38$ at 95%C.L. if we assume that the baryon density, $N_\nu$ and the helium abundance are related by the big bang nucleosynthesis theory. We also provide a forecast for the PLANCK experiment using a Markov chain Monte Carlo approach. In addition to constraining $N_\nu$, we investigate how the big bang nucleosynthesis relation affects the estimation for these parameters and other cosmological parameters.' --- July 2008\ .5in [Probing the Effective Number of Neutrino Species with Cosmic Microwave Background ]{} .45in [Kazuhide Ichikawa$^{1,2}$, Toyokazu Sekiguchi$^1$, and Tomo Takahashi$^3$ ]{} .45in [$^1$ Institute for Cosmic Ray Research, University of Tokyo, Kashiwa 277-8582, Japan\ $^2$ Department of Physics and Astronomy, University College London, Gower Street, London, WC1E 6BT, U.K.\ $^3$ Department of Physics, Saga University, Saga 840-8502, Japan ]{} .4in Introduction ============ Cosmology is now becoming a precision science, and cosmological observations can give us a lot of information for our understanding of the universe. Moreover, the interplay between cosmology and particle physics in various contexts has also been discussed vigorously. One of such examples is the effective number of neutrino species $N_\nu$. Although collider experiments such as LEP have measured the number of light active neutrino types to be $2.92 \pm 0.06$ [@Yao:2006px], it is important to cross-check this value because cosmological measurements may lead to different value. This could be due to an extra radiation component which is predicted by some models of particle physics such as sterile neutrinos (see Ref. [@Dolgov:2002wy] and references therein), or due to incomplete thermalization of neutrinos in the low-scale reheating universe in which the reheating temperature $T_{\rm reh}$ can be as low as $T_{\rm reh} \sim \mathcal{O}(1)$MeV and $N_\nu$ is predicted to be less than three [@Kawasaki:1999na; @Kawasaki:2000en; @Hannestad:2004px; @Ichikawa:2005vw]. If such a non-standard ingredient exists, it can affect big bang nucleosynthesis (BBN), cosmic microwave background (CMB), large scale structure (LSS) and so on; thus precise cosmological observations can probe these scenarios through the effective number of neutrino species. Constraints on $N_\nu$ have been investigated in the literature using the information of CMB and LSS, sometimes with priors on the Hubble constant, cosmic age and Helium abundance [@Kneller:2001cd; @Hannestad:2001hn; @Bowen:2001in; @Crotty:2003th; @Pierpaoli:2003kw; @Hannestad:2003xv; @Barger:2003zg; @Crotty:2004gm; @Hannestad:2005jj; @Spergel:2006hy; @Seljak:2006bg; @Hannestad:2006mi; @Cirelli:2006kt; @Ichikawa:2006vm; @Mangano:2006ur; @Friedland:2007vv; @Hamann:2007pi; @Ichikawa:2007fa; @deBernardis:2007bu; @Popa:2008nz; @Komatsu:2008hk; @Dunkley:2008ie; @Simha:2008zj; @Popa:2008tb]. Although CMB in general can constrain various quantities severely, since the effects of $N_\nu$ on CMB are degenerate with some cosmological parameters, the studies so far have combined CMB data with some other observations such as LSS to obtain a sensible constraint on $N_\nu$. However, when one uses the data from LSS, constraints can become different depending on how one treats non-linear correction/bias on small scales for the matter power spectrum [@Hamann:2007pi]. Furthermore, different LSS data seem to give different constraints on $N_\nu$ [@Spergel:2006hy; @Seljak:2006bg; @Ichikawa:2006vm; @Mangano:2006ur; @Hamann:2007pi]. Regarding the prior on the Hubble constant $H_0$, as is summarized in Ref. [@Ichikawa:2006vm], it can yield some constraints on $N_\nu$ when combined with CMB data (without LSS data) [@Hannestad:2001hn; @Bowen:2001in; @Crotty:2003th; @Hannestad:2003xv; @Barger:2003zg], but they depend on the $H_0$ prior adopted. One may consider that we can use the usually assumed prior on the Hubble constant based on the result by Freedman et al. $H_0 = 72 \pm 8$ [@Freedman:2000cf], but another group reported a somewhat lower value as $H_0 = 62.3 \pm 5.2 $ [@Sandage:2006cv]. If the lower value for $H_0$ is adopted as the prior, a resulting constraint on $N_\nu$ would be different. Having these considerations in mind, it is desirable to investigate a constraint on $N_\nu$ without these kind of uncertainties. In this paper, we study a constraint on $N_\nu$ from CMB experiments alone. By making the analysis of CMB data alone, we can avoid such subtleties as the galaxy-bias/non-linear corrections and the value for the prior on the Hubble constant. However, as is mentioned above, the effects of $N_\nu$ are strongly degenerate in CMB with other cosmological parameters such as energy density of matter, the Hubble constant, and the scalar spectral index, and, in fact, we could not obtain a meaningful bound only with WMAP3 [@Ichikawa:2006vm; @Hamann:2007pi]. Recent WMAP5 alone analysis gives a better constraint but it still cannot give an upper bound [@Dunkley:2008ie; @Komatsu:2008hk]. As we will discuss later, the degeneracy is significant up to about the 2nd/3rd peak of the CMB power spectrum where the observation of WMAP has precisely measured. To break this degeneracy to some extent, it would be helpful to have the information at higher multipoles where signals unique to relativistic neutrinos are expected to appear [@Bashinsky:2003tk]. Recently, the data from ACBAR which probes CMB at higher multipoles than those of WMAP has been updated [@Reichardt:2008ay]. By using this data in addition to other small scale observations such as BOOMERANG and CBI, we can obtain a relatively severe constraint on $N_\nu$ which is comparable to that have been obtained previously with LSS data. The organization of this paper is as follows. In the next section, we start with the discussion how $N_\nu$ affects the CMB power spectrum, which helps to understand our results for the constraint on $N_\nu$. In Section \[sec:current\], we study the current constraint on $N_\nu$ using observations of CMB alone. We use the data from WMAP5, the recent ACBAR, BOOMERANG and CBI. Furthermore, we forecast the constraint from the future Planck experiment. In the final section, we summarize our results and discuss its implications for some models of particle physics/the early universe. Effects of $N_\nu$ on CMB {#sec:effects} ========================= The effective number of neutrino species $N_\nu$ represents the energy density stored in relativistic components as $$\label{eq:N_nu} \rho_{\rm rad} = \rho_\gamma + \rho_\nu + \rho_x = \left[ 1 + \frac{7}{8} \left(\frac{4}{11}\right)^{4/3} N_\nu \right] \rho_\gamma$$ where $\rho_\gamma$, $\rho_\nu$ and $\rho_x$ are energy densities of photons, three species of massless active neutrinos and some possible extra radiation components, respectively. In this paper, we assume that neutrinos are massless and have no chemical potential. For the case with the standard three neutrino flavors without an extra relativistic component, the effective number of neutrino is $N_\nu = 3.046$ where some corrections from the incomplete decoupling due to a slight interaction of neutrinos with electrons/positrons and finite temperature QED effect to the electromagnetic plasma are taken into account [@Mangano:2005cc]. Any deviation of $N_\nu$ from this value implies that there exists an extra relativistic component and/or some non-standard thermal history takes place such as the low reheating temperature scenario. To illustrate the effects of $N_\nu$ on CMB, we plot CMB power spectra for several values of $N_\nu$ in Fig. \[fig:cl\]. Other cosmological parameters are assumed as the mean values of WMAP5 alone analysis for a power-law flat $\Lambda$CDM model. As seen from the figure, as we increase the value of $N_\nu$, the height of the 1st peak is strongly enhanced and the positions of acoustic peaks are shifted to higher multipoles. Also, the amplitude on small scales (higher multipoles) is suppressed. Below, we discuss where these changes are coming from. One of the main effects of $N_\nu$ comes from the change of the epoch of the radiation-matter equality. By increasing (decreasing) the value of $N_\nu$, the radiation-matter equality occurs later (earlier). Thus the increase (decrease) of $N_\nu$ gives an almost the same effect of the decrease (increase) of energy density of matter. One of noticeable features is that the height of the first acoustic peak is enhanced by increasing the value of $N_\nu$. This is due to the early integrated Sachs-Wolfe (ISW) effect in which fluctuations of the corresponding scale, having crossed the sound horizon in the radiation-dominated epoch are boosted by the decay of the gravitational potential. Thus a larger amount of relativistic species drives the first peak higher. Another effect is the shift of the position of acoustic peaks due to the change of the radiation-matter equality through the change of $N_\nu$. The position of acoustic peaks is well captured by the so-called acoustic scale $\theta_A$ which is inversely proportional to the peak position and written as $$\label{eq:theta_A} \theta_A = \frac{r_s (z_{\rm rec})}{r_\theta (z_{\rm rec})}$$ where $r_\theta (z_{\rm rec})$ and $r_s (z_{\rm rec})$ are the comoving angular diameter distance to the last scattering surface and the sound horizon at the recombination epoch $z_{\rm rec}$, respectively. Although $r_\theta (z_{\rm rec})$ almost remains the same for different values of $N_\nu$, $r_s (z_{\rm rec})$ becomes smaller when $N_\nu$ is increased. Thus the positions of acoustic peaks are shifted to higher multipoles (smaller scales) by increasing the value of $N_\nu$. Furthermore, since the position of the $n$-th peak can be roughly written as $l_n \sim n \pi / \theta_A$, separations of the peaks become also greater for larger $N_\nu$. Another important effect is free-streaming of ultrarelativistic neutrinos [@Bashinsky:2003tk]. The perturbation of ultrarelativistic neutrino propagates with the speed of light, which is faster than the sound speed of acoustic oscillations of photon-baryon fluid. The coupled photon-baryon component behaves to oscillate like a compressional fluid; on the other hand, ultrarelativistic neutrinos free-stream to erase their fluctuations. These two components are coupled via gravity; thus the fluctuations of photons can also be affected by the free-streaming of neutrinos, which results in the damping of the amplitude and the shift of the acoustic oscillations. The effects are significant on small scales where fluctuations of a given scale enter the horizon while the energy density of ultrarelativistic neutrinos takes a significant portion of that of the universe. Although the effects of the standard cosmological parameters on the heights and positions of acoustic peaks are well known, here we discuss them in some phenomenological way including the effects of $N_\nu$. For this purpose, we calculated the response of the heights and positions of the acoustic peaks to the change of the cosmological parameters up to the 5th peak around the fiducial values. As a fiducial parameter set, we take those of the mean value from WMAP5-alone analysis for a power-law flat $\Lambda$CDM model. The shifts of the positions of acoustic peaks are $$\begin{aligned} \Delta l_1 &=& 15.58\frac{\Delta \omega_b}{\omega_b} -26.99\frac{\Delta \omega_m}{\omega_m} +36.01\frac{\Delta n_s}{n_s} +0.94\frac{\Delta Y_p}{Y_p} -44.58\frac{\Delta h}{h} +15.53\frac{\Delta N_\nu}{N_\nu} , \label{eq:l1} \\ \Delta l_2 &=& 62.57\frac{\Delta \omega_b}{\omega_b} -74.90\frac{\Delta \omega_m}{\omega_m} +14.69\frac{\Delta n_s}{n_s} +2.81\frac{\Delta Y_p}{Y_p} -108.60\frac{\Delta h}{h} +47.73\frac{\Delta N_\nu}{N_\nu} , \label{eq:l2} \\ \Delta l_3 &=& 74.23\frac{\Delta \omega_b}{\omega_b} -143.47\frac{\Delta \omega_m}{\omega_m} +9.88\frac{\Delta n_s}{n_s} +4.64\frac{\Delta Y_p}{Y_p} -152.39\frac{\Delta h}{h} +81.82\frac{\Delta N_\nu}{N_\nu} , \label{eq:l3} \\ \Delta l_4 &=& 110.84\frac{\Delta \omega_b}{\omega_b} -181.89\frac{\Delta \omega_m}{\omega_m} +7.29\frac{\Delta n_s}{n_s} +6.77\frac{\Delta Y_p}{Y_p} -220.21\frac{\Delta h}{h} +112.76\frac{\Delta N_\nu}{N_\nu} , \label{eq:l4} \\ \Delta l_5 &=& 136.88\frac{\Delta \omega_b}{\omega_b} -237.18\frac{\Delta \omega_m}{\omega_m} +6.20\frac{\Delta n_s}{n_s} +7.79\frac{\Delta Y_p}{Y_p} -276.02\frac{\Delta h}{h} +145.34\frac{\Delta N_\nu}{N_\nu} , \label{eq:l5} \end{aligned}$$ where $l_i$ means the position of the $i$-th peak. $\omega_b$, $\omega_m$, $n_s$, $Y_p$ and $h$ are energy densities of baryon and matter, the scalar spectral index, the primordial helium abundance and the normalized Hubble constant. $n_s$ is defined at the wave number $k = 0.05$Mpc$^{-1}$. The positive derivatives of the peak positions with respect to $N_\nu$ demonstrate the decrease in $r_s (z_{\rm rec})$ due to the increase in $N_\nu$. The responses of the heights of acoustic peaks to the change of various cosmological parameters are $$\begin{aligned} \frac{\Delta \mathcal{C}_{l_1}}{\mathcal{C}_{l_1}} &=& 0.429\frac{\Delta \omega_b}{\omega_b} -0.632\frac{\Delta \omega_m}{\omega_m} -0.947\frac{\Delta n_s}{n_s} -0.0065\frac{\Delta Y_p}{Y_p} +0.141\frac{\Delta N_\nu}{N_\nu}, \label{eq:Cl1} \\ \frac{\Delta \mathcal{C}_{l_2}}{\mathcal{C}_{l_2}} &=& -0.211\frac{\Delta \omega_b}{\omega_b} -0.579\frac{\Delta \omega_m}{\omega_m} -0.034\frac{\Delta n_s}{n_s} -0.035\frac{\Delta Y_p}{Y_p} +0.083\frac{\Delta N_\nu}{N_\nu} , \label{eq:Cl2} \\ \frac{\Delta \mathcal{C}_{l_3}}{\mathcal{C}_{l_3}} &=& 0.026\frac{\Delta \omega_b}{\omega_b} -0.136\frac{\Delta \omega_m}{\omega_m} +0.276\frac{\Delta n_s}{n_s} -0.071\frac{\Delta Y_p}{Y_p} -0.080\frac{\Delta N_\nu}{N_\nu} , \label{eq:Cl3} \\ \frac{\Delta \mathcal{C}_{l_4}}{\mathcal{C}_{l_4}} &=& -0.044\frac{\Delta \omega_b}{\omega_b} -0.229\frac{\Delta \omega_m}{\omega_m} +0.587\frac{\Delta n_s}{n_s} -0.125\frac{\Delta Y_p}{Y_p} -0.108\frac{\Delta N_\nu}{N_\nu} , \label{eq:Cl4} \\ \frac{\Delta \mathcal{C}_{l_5}}{\mathcal{C}_{l_5}} &=& 0.149\frac{\Delta \omega_b}{\omega_b} -0.006\frac{\Delta \omega_m}{\omega_m} +0.776\frac{\Delta n_s}{n_s} -0.172\frac{\Delta Y_p}{Y_p} -0.216\frac{\Delta N_\nu}{N_\nu}. \label{eq:Cl5_ratio} \end{aligned}$$ where ${\cal C}_l = l(l+1)C^{TT}_l/2\pi$. Since $h$ gives only a negligible change to the height of the peaks, we omit it. By increasing the value of $N_\nu$, the height of the 1st peak is enhanced due to the early ISW effect, as previously mentioned. For the third and the higher peaks, the heights are damped more by increasing $N_\nu$, which is inferred from the negative coefficients. This is due to the effect of free streaming of neutrinos [@Bashinsky:2003tk]. In addition, for later convenience, we also show the derivatives of the peak heights relative to the first peak height following Refs. [@Hu:2000ti]. Here, $H_i \equiv {\cal C}_i/{\cal C}_1$ for $i=2$–5. They are useful quantities when we interpret degeneracies since the dependence on the overall amplitude is cancelled out. $$\begin{aligned} \Delta H_2 &=& -0.291\frac{\Delta \omega_b}{\omega_b} +0.023\frac{\Delta \omega_m}{\omega_m} +0.396\frac{\Delta n_s}{n_s} -0.013 \frac{\Delta Y_p}{Y_p} -0.026\frac{\Delta N_\nu}{N_\nu} , \label{eq:H2} \\ \Delta H_3 &=& -0.177\frac{\Delta \omega_b}{\omega_b} +0.206\frac{\Delta \omega_m}{\omega_m} +0.514\frac{\Delta n_s}{n_s} -0.028 \frac{\Delta Y_p}{Y_p} -0.098\frac{\Delta N_\nu}{N_\nu} , \label{eq:H3} \\ \Delta H_4 &=& -0.102\frac{\Delta \omega_b}{\omega_b} +0.082\frac{\Delta \omega_m}{\omega_m} +0.317\frac{\Delta n_s}{n_s} -0.025 \frac{\Delta Y_p}{Y_p} -0.054\frac{\Delta N_\nu}{N_\nu} , \label{eq:H4} \\ \Delta H_5 &=& -0.040\frac{\Delta \omega_b}{\omega_b} +0.084\frac{\Delta \omega_m}{\omega_m} +0.236\frac{\Delta n_s}{n_s} -0.023 \frac{\Delta Y_p}{Y_p} -0.052\frac{\Delta N_\nu}{N_\nu}. \label{eq:H5} \end{aligned}$$ Constraint on $N_\nu$ from observations of CMB {#sec:current} ============================================== In this section, we present our result for the constraint on $N _\nu$ from CMB alone. First, we give some details of our analysis. We use the CMB data of WMAP5 [@Komatsu:2008hk; @Dunkley:2008ie; @Hinshaw:2008kr; @Hill:2008hx; @Nolta:2008ih], BOOMERANG [@Jones:2005yb; @Piacentini:2005yq; @Montroy:2005yx], CBI [@Sievers:2005gj] and ACBAR [@Reichardt:2008ay]. We performed a Markov chain Monte Calro (MCMC) analysis to obtain constraints on cosmological parameters using [cosmomc]{} code [@Lewis:2002ah] with some modifications which are described in the following. We explore a 9 dimensional parameter space which consists of $\omega_b$, $\omega_c$, $\tau$, $\theta_s$, $n_s$, $A_s$, $A_{SZ}$, $Y_p$ and $N_\nu$. Here, $\omega_c$ is the energy density of dark matter, $\tau$ is the optical depth of reionization, $\theta_s$ is the acoustic peak scale [@Kosowsky:2002zt], $A_s$ is the amplitude of primordial fluctuations and $A_{SZ}$ is the amplitude of thermal Sunyaev-Zel’dovich (SZ) effect which is normalized to the $C_l^{SZ}$ template from Ref. [@Komatsu:2002wc]. ----------------------- ------------------------ ----------------- ----------------- \[0pt\]\[0pt\][Mean]{} 68%$\uparrow$ 95%$\uparrow$ 68%$\downarrow$ 95%$\downarrow$ WMAP5 \[0pt\]\[0pt\][5.65]{} $7.88$ $9.96$ ($Y_p=0.24$: fixed) $3.02$ $1.92$ CMB all \[0pt\]\[0pt\][4.24]{} $5.47$ $7.94$ ($Y_p$: free) $2.03$ $0.96$ CMB all \[0pt\]\[0pt\][3.71]{} $4.80$ $6.38$ ($Y_p$: BBN relation) $2.27$ $1.39$ CMB all \[0pt\]\[0pt\][3.89]{} $4.89$ $6.84$ ($Y_p=0.24$: fixed) $2.19$ $1.28$ ----------------------- ------------------------ ----------------- ----------------- : The mean values, 68% and 95% limits of $N_\nu$ for several current CMB data sets and assumptions of $Y_p$. []{data-label="table:N_nu_current"} -- -- -- -- -- -- -- -- As far as CMB is concerned, these cosmological parameters can be considered to be independent. However, when we take into account the BBN theory, $Y_p$ is determined once $\omega_b$ and $N_\nu$ are given. In this case, we should relate these parameters to each other and sample an 8 dimensional parameter space. We shall refer to this relation among $Y_p$, $\omega_b$ and $N_\nu$ as the BBN relation. For this purpose, we calculate $Y_p$ as a function of $\omega_b$ and $N_\nu$ using the Kawano BBN code [@Kawanocode] with some update in the nuclear reaction network part based on Ref. [@NACRE]. Such relation is considered in the CMB analyses in Refs. [@Huey:2003ef; @Ichikawa:2006dt; @Hamann:2007sb; @Ichikawa:2007js; @Popa:2008nz]. In passing, we would like to make a comment on the fitting formula for the BBN calculation presented in Ref. [@Serpico:2004gx] which has been used in the authors’ previous work [@Ichikawa:2006dt]. We do not adopt the formula here since, as we will see later, our MCMC chains for the constraints from the present data sets sometimes go to the region beyond the range over which their fitting formula is valid, $0 \le N_\nu \le 6$. For the Planck forecast, where the chains are contained in that region, we obtain the same results with the Kawano code and the fitting formula of Ref. [@Serpico:2004gx]. Furthermore, we also consider the case with fixing the helium abundance to $Y_p=0.24$ since, in most analyses, the primordial helium abundance is fixed to this value. Finally, it should be noted that this BBN relation is not necessarily realized in some cases. We can think of more exotic scenarios in which the BBN theory cannot relate those parameters. For exmaple, $\omega_b$ and $Y_p$ may vary between BBN and CMB epochs [@Ichikawa:2004pb] or an increase in $N_\nu$ may take place [@Ichikawa:2007jv]. Now, we present our results in order. In Fig. \[fig:1Dcurrent\], the posterior 1D distributions for $N_\nu$ are shown for the analysis with WMAP5 alone and WMAP5+ACBAR+BOOMERANG+CBI (CMB all). The former is shown by a red solid line. For the latter case, the results for different assumptions on $Y_p$ are depicted: $Y_p$ being fixed as $Y_p=0.24$ (green dashed line), $Y_p$ determined from the BBN relation (blue dotted line) and $Y_p$ being treated as a free parameter (magenta dot-dashed line). Corresponding constraints on $N_\nu$ are summarized in Table \[table:N\_nu\_current\]. We also show 2D contours of 68% and 95% C.L. in the planes of $N_\nu$ v.s. several other cosmological parameters in Fig. \[fig:2Dcurrent\]. Table \[table:current\_all\] summarizes the derived constraints on these parameters. As seen from Fig. \[fig:1Dcurrent\], the likelihood for $N_\nu$ from WMAP alone has irregular shape, far from Gaussian. It has the maximum around $N_\nu \sim 5$, declines slowly as $N_\nu$ increases and drop abruptly at $N_\nu \sim 9$. The abrupt cut can be traced to the prior on the cosmic age $t_0$ which is implicitly assumed as $ 10\, {\rm Gyr} < t_0 < 20\, {\rm Gyr}$ in the analysis. In particular, the lower limit $t_0 > 10\,{\rm Gyr}$ makes the cut (see Fig \[fig:2Dcurrent\]). We can regard this prior to be very conservative on the observational ground since it is far looser than the astrophysical lower bound of the cosmic age e.g. $t_0 > 11.2\,{\rm Gyr}$ (95% C.L.) from the age estimates of globular clusters [@Krauss:2003em]. Moreover, we should include such prior from the practical reason. As can be seen by the relatively slow decline of the likelihood for $5 \lesssim N_\nu \lesssim 9$ or the elongated contours in Fig \[fig:2Dcurrent\], the degeneracy is so severe that we cannot produce MCMC chains which are well converged within a reasonable time. Although we can formally calculate a constraint using this data as shown in Table \[table:N\_nu\_current\], since the likelihood is so irregular, we would conclude that it is not meaningful to constrain $N_\nu$ from WMAP5 alone. However, it may be instructive to understand how the degeneracies arise in the WMAP-alone analysis. As clearly shown in Fig. \[fig:2Dcurrent\], $N_\nu$ most notably degenerates with $\omega_c$ and $H_0$ (or $\theta_s$). There are also some degeneracies with $n_s$ and $A_s$ but not as severe as $\omega_c$ or $H_0$. These degeneracies are understood as follows. First, to produce the same amount of the early ISW effect, $\omega_c$ has to be increased as $N_\nu$ increases. It roughly scales as $(\omega_b+\omega_c) \propto N_\nu$ to make the matter-radiation equality same. At the same time, under the flatness assumption, $\Omega_m$ has to be preserved in order to have the same distance to the last scattering surface. Then, since $\Omega_m = (\omega_b+\omega_c)/h^2$, $h$ has to increase for larger $N_\nu$. The slight enhancement in $n_s$ and $A_s$ can be attributed to their effects to cancel the suppression around the diffusion damping scales due to the increase in $N_\nu$. A more quantitative argument based on the derivatives presented in the previous section may be useful. The degeneracy as regards the same matter-radiation equality is given by setting $\Delta {\cal C}_{l_1} = 0$ to be $\Delta \omega_m/\omega_m \sim 0.2\, \Delta N_\nu/N_\nu$. This is equivalent to $\Delta \omega_c/\omega_c \sim 0.3\, \Delta N_\nu/N_\nu$, which roughly gives the slope in the $N_\nu$–$\omega_c$ plane in Fig \[fig:2Dcurrent\]. Using this relation with $\Delta l_1=0$ shows the $N_\nu$–$h$ degeneracy. From Eq. , we obtain $\Delta h/h \sim 0.2\, \Delta N_\nu/N_\nu$, which appear in the $N_\nu$–$h$ contour in Fig \[fig:2Dcurrent\]. The $N_\nu$–$n_s$ degeneracy is given by further requiring $\Delta H_2 = 0$. Plugging $\Delta \omega_m/\omega_m$ and $\Delta h/h$ in Eq. yields $\Delta n_s/n_s \sim 0.05\,\Delta N_\nu/N_\nu$. This 5% increase in the best fit value of $n_s$ for $\Delta N_\nu = 3$ is consistent with the $N_\nu$–$n_s$ contour in Fig \[fig:2Dcurrent\]. Although WMAP has measured the CMB power spectrum very precisely, since it is just up to around the 2nd peak, the effects of $N_\nu$ are absorbed in the changes of $\omega_c$, $h$, $n_s$ and $A_s$ and we cannot constrain $N_\nu$. For a visual illustration of the degeneracy, in Fig. \[fig:cl\_deg\], we show CMB power spectra for the case with $N_\nu=1$ (blue dotted line), $3$ (red solid line) and $5$ (green dashed line) with other cosmological parameters being chosen such that they give the degenerate spectra up to the 2nd/3rd peak. We can see that these curves coincide up to the 2nd peak but begin to separate around the 3rd or the higher peaks. On small scales, the change caused by $N_\nu$ cannot be fully canceled just by tuning other parameters. In particular, $n_s$ affects the spectrum in the whole scales; thus, even if the spectra is almost degenerate up to the 2nd peak by tuning the value of $n_s$, it cannot cancel the damping on small scales. ------------------- -------------------------------- --------------------------------- --------------------------------- --------------------------------- parameters CMB all CMB all CMB all WMAP5 ($Y_p$: free) ($Y_p$: BBN relation) ($Y_p=0.24$) ($Y_p=0.24$) $\omega_b$ $0.0229^{+0.00056}_{-0.00054}$ $0.02291^{+0.00052}_{-0.00059}$ $0.02291^{+0.00058}_{-0.00053}$ $0.02275^{+0.00060}_{-0.00062}$ $\omega_c$ $0.132^{+0.018}_{-0.037}$ $0.124^{+0.017}_{-0.027}$ $0.127^{+0.017}_{-0.029}$ $0.153^{+0.036}_{-0.047}$ $\theta_s$ $1.0390^{+0.0071}_{-0.0088}$ $1.0413^{+0.0039}_{-0.0051}$ $1.0402^{+0.0044}_{-0.0060}$ $1.0334^{+0.0042}_{-0.0084}$ $\tau$ $0.088^{+0.016}_{-0.019}$ $0.088^{+0.016}_{-0.017}$ $0.088^{+0.015}_{-0.019}$ $0.088^{+0.016}_{-0.019}$ $n_s$ $0.977^{+0.026}_{-0.022}$ $0.975^{+0.024}_{-0.023}$ $0.975^{+0.024}_{-0.021}$ $0.989^{+0.030}_{-0.019}$ $\ln(10^{10}A_s)$ $3.104^{+0.0067}_{-0.0061}$ $3.097^{+0.066}_{-0.061}$ $3.098^{+0.065}_{-0.059}$ $3.128^{+0.080}_{-0.054}$ $Y_p$ $0.220^{+0.105}_{-0.085}$ $0.256^{+0.015}_{-0.016}$ — — $N_\nu$ $4.24^{+1.23}_{-2.21}$ $3.72^{+1.07}_{-1.45}$ $3.89^{+1.00}_{-1.70}$ $5.65^{+2.63}_{-2.23}$ $A_\mathrm{SZ}$ $1.04^{+0.96}_{-0.34}$ $1.07^{+0.93}_{-0.33}$ $1.05^{+0.95}_{-0.34}$ $1.00^{+0.85}_{-0.50}$ $\Omega_m$ $0.265^{+0.026}_{-0.032}$ $0.266^{+0.025}_{-0.030}$ $0.265^{+0.025}_{-0.030}$ $0.260^{+0.028}_{-0.033}$ Age\[Gyr\] $12.9^{+1.3}_{-1.2}$ $13.2^{+1.0}_{-1.0}$ $13.1^{+1.0}_{-1.1}$ $12.1^{+7.7}_{-1.9}$ $H_0$ $76.5^{+6.4}_{-9.7}$ $74.5^{+5.5}_{-6.7}$ $75.2^{+5.6}_{-7.3}$ $82.0^{+9,9}_{-8.9}$ ------------------- -------------------------------- --------------------------------- --------------------------------- --------------------------------- : Mean values and 68% C.L. errors from current observations of CMB for the cases with WMAP5 alone and all data combined.[]{data-label="table:current_all"} When we include the data at higher multipoles such as ACBAR, BOOMERANG and CBI, the likelihood functions now have a well-behaved peak close to Gaussian and we can obtain meaningful constraints. The bound on $N_\nu$ is $0.96\le N_\nu \le 7.94$ at 95 % C.L. when $Y_p$ is treated as a free parameter. At higher multipoles, the free streaming of neutrinos damps the spectrum, which cannot be compensated by above-mentioned parameters. Hence the degeneracy can be removed to some extent. That is the reason why we can have a severer constraint on $N_\nu$ by including the data on small scales. In fact, $Y_p$ also suppresses the amplitude on small scales via diffusion damping (see Eqs. and ); thus the constraint on $N_\nu$ slightly changes for different treatments of $Y_p$ but the differences are very small as seen in Fig. \[fig:1Dcurrent\] and Table \[table:N\_nu\_current\]. Since current CMB observations on small scales are not so precise, it does not make much difference how we treat $Y_p$. Imposing the BBN relation tightens the constraint to $1.39 \le N_\nu \le 6.38$ at 95 % C.L., but it is not so different from the $Y_p$-free case. Also, the limit does not differ much even if we set $Y_p = 0.24$. Similarly, the estimates for the other cosmological parameters are not affected by the assumption on $Y_p$ as shown in Table \[table:current\_all\]. ----------------------- -------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- \[0pt\]\[0pt\][Mean]{} 68%$\uparrow$ 95%$\uparrow$ 68%$\uparrow$ 95%$\uparrow$ 68%$\uparrow$ 95%$\uparrow$ 68%$\downarrow$ 95%$\downarrow$ 68%$\downarrow$ 95%$\downarrow$ 68%$\downarrow$ 95%$\downarrow$ CMB all \[0pt\]\[0pt\][$4.24$]{} $5.47$ $7.94$ $5.51$ $8.19$ (3.046) (3.046) ($Y_p$: free) $2.03$ $0.96$ $(3.046)$ $(3.046)$ $2.05$ $1.17$ CMB all \[0pt\]\[0pt\][$3.71$]{} $4.80$ $6.38$ $4.70$ $6.35$ ($Y_p$: BBN relation) $2.27$ $1.39$ $(3.046)$ $(3.046)$ CMB all \[0pt\]\[0pt\][$3.89$]{} $4.89$ $6.84$ $4.87$ $6.88$ (3.046) (3.046) ($Y_p=0.24$: fixed) $2.19$ $1.28$ $(3.046)$ $(3.046)$ $2.12$ $1.27$ Planck \[0pt\]\[0pt\][$3.11$]{} $3.44$ $3.83$ $3.45$ $3.87$ (3.046) (3.046) ($Y_p$: free) $2.72$ $2.41$ $(3.046)$ $(3.046)$ $2.72$ $2.43$ Planck \[0pt\]\[0pt\][$3.06$]{} $3.26$ $3.44$ $3.25$ $3.44$ ($Y_p$: BBN relation) $2.87$ $2.68$ $(3.046)$ $(3.046)$ Planck \[0pt\]\[0pt\][$3.19$]{} $3.43$ $3.67$ $3.37$ $3.63$ (3.046) (3.046) ($Y_p=0.24$: fixed) $2.95$ $2.72$ $(3.046)$ $(3.046)$ $2.87$ $2.67$ ----------------------- -------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- : The mean values and 68% and 95% limits of $N_\nu$ for current and future CMB data. []{data-label="table:N_nu"} Up to now, we have assumed no prior on $N_\nu$. However, if we consider an extra radiation component such as sterile neutrinos and so on, the effective number of neutrino species just increases. In this case, $N_\nu$ cannot be less than the standard value of 3.046. Thus it may be appropriate to study adopting the prior $N_\nu > 3.046$ to constrain a scenario with such an extra radiation component. We denote it as $\Delta N^{\rm ext}_\nu \equiv N_\nu - 3.046$. With this prior, we obtain an upper bound on an extra radiation component as $N_\nu <8.19$ (or $\Delta N^{\rm ext}_\nu < 5.14$) at 95 % C.L. when $Y_p$ is taken as a free parameter and improve to be $N_\nu < 6.35$ (or $\Delta N^{\rm ext}_\nu < 3.30$) at 95 % C.L. when the BBN relation is assumed. Notice that these limits are weaker than those with no priors on $N_\nu$, which are $N_\nu < 7.94$ ($\Delta N^{\rm ext}_\nu < 4.89$) and $N_\nu < 6.38$ ($\Delta N^{\rm ext}_\nu < 3.33$), respectively. This somewhat peculiar fact stems from the shape of the likelihood shown in Fig. \[fig:1Dcurrent\], which is not symmetric with respect to $N_\nu = 3.046$. Since the differences due to the $N_\nu$ prior are not negligible, caution is needed when we use these constraints regarding the prior on $N_\nu$. Even when we limit ourselves to the case with three active neutrino species, a deviation from the standard value of $N_\nu = 3.046$ is possible. In a scenario with low (MeV scale) reheating temperature, $N_\nu$ can be less than 3.046. In this case, $N_\nu$ only takes the value less than the standard one. Thus it may be interesting to investigate a constraint on $N_\nu$ assuming $N_\nu < 3.046$. As regards the treatment of $Y_p$, we do not consider the case with adopting the BBN relation here because, in a scenario with MeV reheating temperature, $Y_p$ should be calculated taking into account the non-thermal neutrino distribution functions and oscillation effects [@Kawasaki:1999na; @Kawasaki:2000en; @Hannestad:2004px; @Ichikawa:2005vw]. These effects drive $Y_p$ to increase as $N_\nu$ decreases contrary to the usual case where $N_\nu$ just represents a measure of the expansion rate. (This is why we are not showing constraints for the prior $N_\nu < 3.046$ with the BBN relation in Table \[table:N\_nu\].) Since taking into account this effect is beyond the scope of this paper, we show the constraint for the case with $Y_p$ being varied freely, which can be considered as the conservative one for the prior $N_\nu < 3.046$. We obtained constraints $N_\nu > 1.27$ and $N_\nu > 1.17$ at 95% C.L. for the cases with $Y_p$ being fixed as $Y_p=0.24$ and $Y_p$ being assumed as a free parameter, respectively. For a scenario with MeV scale reheating temperature, these limits are translated into the lower bound on the reheating temperature as $T_{\rm reh}>2.0$MeV [@Ichikawa:2005vw]. Finally, we investigate a future constraint on cosmological parameters paying particular attention to $N_\nu$ and its effects on constraints on other parameters. We use the data expected from the future Planck experiment and make a MCMC analysis following the method in Ref. [@Perotto:2006rj]. As for the specification of Planck, we adopt the following parameters for the instrument. For the frequency channels of $\nu = 100$, 143 and 217GHz, the width of the beam and the sensitivities per pixel for temperature and polarization are adopted as $(\theta_{\rm FWHW} {\rm [arcmin]}, \sigma_T [\mu {\rm K}], \sigma_P [\mu {\rm K}] ) =(9.5, 6.8, 10.9), (7.1, 6.0, 11.4)$ and $(5.0, 13.1, 26.7)$, respectively. Other frequency channels are assumed to be used to remove foregrounds. We make use of the data up to $l =2500$ in order that our results will not be affected by the SZ effect and the marginalization over $A_{SZ}$ is not performed. These setups for the Planck forecast are similar to the recent works performed in Refs. [@Hamann:2007sb; @Popa:2008nz; @Popa:2008tb], but explored parameter spaces are different. We make a simple extension by adding $N_\nu$ and $Y_p$ to the standard 6 dimensional parameter space, but theirs include neutrino masses and/or lepton asymmetry. When one would like to check the constraint on an extra radiation component in a simple scenario, one can refer our results here. However, when some other particular setups are considered such as a scenario with large lepton asymmetry and massive neutrinos, the above mentioned references should be consulted. Our results are summarized in Tables \[table:N\_nu\] and \[table:forecast\]. As seen from the Table \[table:N\_nu\], the constraint is most stringent when the BBN relation is adopted, and in this case, we obtained a future constraint as $2.68 \le N_\nu \le3.44$ at 95 % C.L. Another point which should be noted is that fixing of $Y_p = 0.24$ can bias the determination of some other cosmological parameters such as $\omega_b$ and $n_s$, which was already pointed out in Refs. [@Hamann:2007sb; @Ichikawa:2007js]. However, when we vary the value of $N_\nu$, the effect of fixing of $Y_p=0.24$ is partly cancelled by the change in $N_\nu$. In fact, this in turn results in biases of other cosmological parameters such as $\omega_c$ and $ \theta_s$ which are strongly correlated with $N_\nu$. Therefore, $Y_p$ should be carefully treated in investigating cosmological constraints with future CMB data. ------------------- --------------------------------- --------------------------------- --------------------------------- Planck Planck Planck parameters ($Y_p$: free) ($Y_p$: BBN relation) ($Y_p=0.24$) $\omega_b$ $0.02275^{+0.00025}_{-0.00028}$ $0.02275^{+0.00026}_{-0.00027}$ $0.02273^{+0.00027}_{-0.00026}$ $\omega_c$ $0.1108^{+0.0046}_{-0.0056}$ $0.1101^{+0.0028}_{-0.0028}$ $0.1120^{+0.0033}_{-0.0036}$ $\theta_s$ $1.0404^{+0.0014}_{-0.0014}$ $1.04060^{+0.00044}_{-0.00049}$ $1.04000^{+0.00055}_{-0.00062}$ $\tau$ $0.0881^{+0.0050}_{-0.0064}$ $0.0881^{+0.0053}_{-0.0063}$ $0.0880^{+0.0056}_{-0.0059}$ $n_s$ $0.964^{+0.009}_{-0.010}$ $0.964^{+0.010}_{-0.010}$ $0.963^{+0.010}_{-0.009}$ $\ln(10^{10}A_s)$ $3.066^{+0.016}_{-0.016}$ $3.065^{+0.014}_{-0.015}$ $3.068^{+0.015}_{-0.015}$ $Y_p$ $0.246^{+0.020}_{-0.018}$ $0.2488^{+0.0027}_{-0.0027}$ — $N_\nu$ $3.11^{+0.33}_{-0.39}$ $3.06^{+0.20}_{-0.19}$ $3.19^{+0.24}_{-0.24}$ $\Omega_m$ $0.256^{+0.010}_{-0.010}$ $0.256^{+0.009}_{-0.010}$ $0.255^{+0.009}_{-0.010}$ Age\[Gyr\] $13.63^{+0.34}_{-0.31}$ $13.67^{+0.20}_{-0.21}$ $13.56^{+0.22}_{-0.25}$ $H_0$ $72.3^{+2.2}_{-2.4}$ $72.0^{+1.7}_{-1.6}$ $72.7^{+1.8}_{-1.9}$ ------------------- --------------------------------- --------------------------------- --------------------------------- : Forecasts on mean values and 68% errors of $N_\nu$ and other cosmological parameters. []{data-label="table:forecast"} We would like in the end to comment on how our discussion so far can be affected by theoretical uncertainties in the recombination process [@Seager:1999bc; @Lewis:2006ym; @Wong:2007ym; @Chluba:2005uz; @Chluba:2006bc; @Chluba:2007yp; @Switzer:2007sn; @Hirata:2007sp; @Switzer:2007sq]. Since the change of $Y_p$ can influence the recombination process, its uncertainties might affect the cosmological parameter determination in some way. Thus it may be worth mentioning here on the effects. For this purpose, we proceed with the same analysis as have been done in Ref. [@Ichikawa:2007js] but varying $N_\nu$ here. Two parameters $F_H$ and $b_{He}$, which represent the uncertainties in the recombination modeling, are included among other free parameters. (See Ref. [@Ichikawa:2007js] and references therein for more details). We impose top-hat priors, $0<F_H<2$ and $0<b_{He}<1.5$, which are very conservative ones, to take into account the uncertainties in the recombination theory. We made the analyses for the two cases where $Y_p$ is given from the BBN relation and $Y_p$ is treated as a free parameter. In both cases, we found that the constraints on other cosmological parameters including $N_\nu$ are scarcely affected even by very conservative prior on $F_H$ and $b_{He}$. The mean values are unchanged and errors increase only very slightly (no more than 10% for any parameters other than $F_H$ and $b_{He}$). Therefore we can say that the uncertainties parametrized with $F_H$ and $b_{He}$ do not change much our results of the Planck forecast discussed above. However, we would need more understanding of uncertainties in the recombination theory for the precise determination of cosmological parameters in future CMB surveys. Summary {#sec:summary} ======= We discussed the issue of probing the effective number of neutrino species $N_\nu$ from CMB data alone. Although a constraint on $N_\nu$ has been investigated by many authors, in most analyses, some combinations of data set such as CMB+LSS, CMB+$H_0$, CMB+LSS+$H_0$ have been used to constrain $N_\nu$. This is partly because $N_\nu$ has severe degeneracies in WMAP with some other cosmological parameters such as $\omega_m$ and $h$; thus $N_\nu$ can be more constrained by combining some data sets. However, when we combine data from LSS, some subtleties can arise: a constraint from LSS data depends on how we treat non-linear corrections/bias. Furthermore, different galaxy data lead to slightly different constraints on $N_\nu$. In addition, as for the Hubble prior, the prior usually adopted is $H_0 = 72 \pm 8$ which is from the result of Freedman et al [@Freedman:2000cf]. However, another group has reported somewhat different value as $H_0 = 62.3 \pm 5.2$ [@Sandage:2006cv]. Since different priors on the Hubble constant can give different results, in this respect, the constraint obtained by assuming some prior on $H_0$ should be regarded taking into account the above uncertainty. Taking these issues into consideration, it may be interesting to study a constraint on $N_\nu$ removing such subtleties, which can be done by using CMB data alone. In this paper, first we discussed the effects of $N_\nu$ on CMB and the issues of degeneracies with some other cosmological parameters. Phenomenological descriptions of its effects on the heights and the positions of acoustic peaks were also given. Then, in section \[sec:current\], a constraint on $N_\nu$ was studied by using CMB data alone. We made use of the data from WMAP5, ACBAR, BOOMERANG and CBI. As discussed there, although WMAP measurement is very accurate, its precision is limited up to the 2nd peak/3rd peak. We have explicitly shown that the information up to the 2nd/3rd peak is not enough to constrain $N_\nu$ severely. This was demonstrated by making the analysis with WMAP data alone, in which a sensible constraint cannot be obtained. However, if we include the data on small scales, the degeneracies of $N_\nu$ with some other cosmological parameters can be removed to some extent; then a stronger constraint can be obtained. In fact, on small scales, the amplitude is suppressed due to the free streaming effect by increasing $N_\nu$, which is similar to the effects of $Y_p$ through the diffusion damping. Thus we have studied the constraint on $N_\nu$ assuming different priors on $Y_p$: adopting the BBN relation to derive $Y_p$ for given $N_\nu$ and $\omega_b$, assuming $Y_p$ as a free parameter and usual fixing of $Y_p=0.24$. Depending on the prior, the constraint slightly changes. We obtained the 95 % limits as $0.96 \le N_\nu \le 7.94$ for the case with $Y_p$ being free, $1.28 \le N_\nu \le 6.84$ for $Y_p$ being fixed as $Y_p=0.24$ and $1.39 \le N_\nu \le 6.38$ when the BBN relation being adopted. It should be noted that these constraints are comparable to that obtained using CMB+LSS in previous works. One of the main purposes of constraining the effective number of neutrino species using cosmological data is to check the standard value of $N_\nu$ independently from particle physics experiments. Thus we primarily focus on the analysis with no prior on $N_\nu$. However, from the viewpoint of constraining extra radiation which may be motivated from some particle physics models, a constraint obtained by assuming the prior $N_\nu > 3.046$ may be interesting since an extra radiation always increases the value of $N_\nu$. In this respect, we also made an analysis adopting this prior and obtained the constraint on the effective number of neutrino as $N_\nu \le 6.35$ and $N_\nu \le 8.14$ for the cases where the BBN relation is adopted and $Y_p$ is treated as a free parameter. On the other hand, in a scenario with low-reheating temperature, the effective number of neutrino species can be reduced. In this case, another prior may be motivated to be assumed for a simple scenario of low-reheating temperature with three relativistic neutrino species. In this regard, we have also studied the case with the prior $N_\nu < 3.046$ and obtained the constraints as $N_\nu > 1.17$ for the cases with $Y_p$ being assumed as a free parameter. This can be translated into the lower bound on the reheating temperature as $T_{\rm reh} > 2.0$ MeV. We have also discussed a future constraint on $N_\nu$ using the expected data from Planck experiment. It was shown that the attainable constraint on $N_\nu$ from Planck is $2.68 \le N_\nu \le3.44$ at 95% C.L. when the BBN relation is adopted for $Y_p$, which is most stringent compared to the other cases. Since Planck experiment can probe CMB down to smaller scales than WMAP, Planck alone can give a stringent constraint on $N_\nu$. The interplay between particle physics and cosmology is now becoming more important in the era of precision cosmology. One of such examples is the number of neutrino species, which was investigated in this paper. In light of upcoming more precise observations of cosmology, research of this kind will bring us fruitful insight for particle physics and cosmology. [Acknowledgments:]{} This work is supported in part by the Japan Society for the Promotion of Science (K.I. and T.S.), the Sumitomo Foundation (T.T.), and the Grant-in-Aid for Scientific Research from the Ministry of Education, Science, Sports, and Culture of Japan, No.18840010 (K.I.) and No.19740145 (T.T.). [100]{} W. M. Yao [et al.]{} \[Particle Data Group\], J. Phys. G [33]{} (2006) 1. A. D. Dolgov, Phys. Rept. [370]{}, 333 (2002) \[arXiv:hep-ph/0202122\]. M. Kawasaki, K. Kohri and N. Sugiyama, Phys. Rev. Lett. [82]{}, 4168 (1999) \[arXiv:astro-ph/9811437\]. M. Kawasaki, K. Kohri and N. Sugiyama, Phys. Rev. D [62]{}, 023506 (2000) \[arXiv:astro-ph/0002127\]. S. Hannestad, Phys. Rev. D [70]{}, 043506 (2004) \[arXiv:astro-ph/0403291\]. K. Ichikawa, M. Kawasaki and F. Takahashi, Phys. Rev. D [72]{}, 043522 (2005) \[arXiv:astro-ph/0505395\]. J. P. Kneller, R. J. Scherrer, G. Steigman and T. P. Walker, Phys. Rev. D [64]{}, 123506 (2001) \[arXiv:astro-ph/0101386\]. S. Hannestad, Phys. Rev. D [64]{}, 083002 (2001) \[arXiv:astro-ph/0105220\]. R. Bowen, S. H. Hansen, A. Melchiorri, J. Silk and R. Trotta, Mon. Not. Roy. Astron. Soc. [334]{}, 760 (2002) \[arXiv:astro-ph/0110636\]. P. Crotty, J. Lesgourgues and S. Pastor, Phys. Rev. D [67]{}, 123005 (2003) \[arXiv:astro-ph/0302337\]. E. Pierpaoli, Mon. Not. Roy. Astron. Soc. [342]{}, L63 (2003) \[arXiv:astro-ph/0302465\]. S. Hannestad, JCAP [0305]{}, 004 (2003) \[arXiv:astro-ph/0303076\]. V. Barger, J. P. Kneller, H. S. Lee, D. Marfatia and G. Steigman, Phys. Lett. B [566]{}, 8 (2003) \[arXiv:hep-ph/0305075\]. P. Crotty, J. Lesgourgues and S. Pastor, Phys. Rev. D [69]{}, 123007 (2004) \[arXiv:hep-ph/0402049\]. S. Hannestad, JCAP [0601]{}, 001 (2006) \[arXiv:astro-ph/0510582\]. D. N. Spergel [et al.]{} \[WMAP Collaboration\], Astrophys. J. Suppl. [170]{}, 377 (2007) \[arXiv:astro-ph/0603449\]. U. Seljak, A. Slosar and P. McDonald, JCAP [0610]{}, 014 (2006) \[arXiv:astro-ph/0604335\]. S. Hannestad and G. G. Raffelt, JCAP [0611]{}, 016 (2006) \[arXiv:astro-ph/0607101\]. M. Cirelli and A. Strumia, JCAP [0612]{}, 013 (2006) \[arXiv:astro-ph/0607086\]. K. Ichikawa, M. Kawasaki and F. Takahashi, JCAP [0705]{}, 007 (2007) \[arXiv:astro-ph/0611784\]. G. Mangano, A. Melchiorri, O. Mena, G. Miele and A. Slosar, JCAP [0703]{}, 006 (2007) \[arXiv:astro-ph/0612150\]. A. Friedland, K. M. Zurek and S. Bashinsky, arXiv:0704.3271 \[astro-ph\]. J. Hamann, S. Hannestad, G. G. Raffelt and Y. Y. Y. Wong, JCAP [0708]{}, 021 (2007) \[arXiv:0705.0440 \[astro-ph\]\]. K. Ichikawa, arXiv:0706.3465 \[astro-ph\]. F. de Bernardis, A. Melchiorri, L. Verde and R. Jimenez, JCAP [0803]{}, 020 (2008) \[arXiv:0707.4170 \[astro-ph\]\]. L. A. Popa and A. Vasile, arXiv:0801.3928 \[astro-ph\]. E. Komatsu [et al.]{} \[WMAP Collaboration\], arXiv:0803.0547 \[astro-ph\]. J. Dunkley [et al.]{} \[WMAP Collaboration\], arXiv:0803.0586 \[astro-ph\]. V. Simha and G. Steigman, JCAP [0806]{}, 016 (2008) \[arXiv:0803.3465 \[astro-ph\]\]. L. A. Popa and A. Vasile, JCAP [0806]{}, 028 (2008) \[arXiv:0804.2971 \[astro-ph\]\]. W. L. Freedman [et al.]{} \[HST Collaboration\], Astrophys. J. [553]{}, 47 (2001) \[arXiv:astro-ph/0012376\]. A. Sandage, G. A. Tammann, A. Saha, B. Reindl, F. D. Macchetto and N. Panagia, Astrophys. J. [653]{}, 843 (2006) \[arXiv:astro-ph/0603647\]. S. Bashinsky and U. Seljak, Phys. Rev. D [69]{}, 083002 (2004) \[arXiv:astro-ph/0310198\]. C. L. Reichardt [et al.]{}, arXiv:0801.1491 \[astro-ph\]. G. Mangano, G. Miele, S. Pastor, T. Pinto, O. Pisanti and P. D. Serpico, Nucl. Phys. B [729]{}, 221 (2005) \[arXiv:hep-ph/0506164\]. W. Hu, M. Fukugita, M. Zaldarriaga and M. Tegmark, Astrophys. J. [549]{}, 669 (2001) \[arXiv:astro-ph/0006436\]. G. Hinshaw [et al.]{} \[WMAP Collaboration\], arXiv:0803.0732 \[astro-ph\]. R. S. Hill [et al.]{} \[WMAP Collaboration\], arXiv:0803.0570 \[astro-ph\]. M. R. Nolta [et al.]{} \[WMAP Collaboration\], arXiv:0803.0593 \[astro-ph\]. W. C. Jones [et al.]{}, Astrophys. J. [647]{}, 823 (2006) \[arXiv:astro-ph/0507494\]. F. Piacentini [et al.]{}, Astrophys. J. [647]{}, 833 (2006) \[arXiv:astro-ph/0507507\]. T. E. Montroy [et al.]{}, Astrophys. J. [647]{}, 813 (2006) \[arXiv:astro-ph/0507514\]. J. L. Sievers [et al.]{}, arXiv:astro-ph/0509203. A. Lewis and S. Bridle, Phys. Rev. D [66]{}, 103511 (2002) \[arXiv:astro-ph/0205436\]. A. Kosowsky, M. Milosavljevic and R. Jimenez, Phys. Rev. D [66]{}, 063007 (2002) \[arXiv:astro-ph/0206014\]. E. Komatsu and U. Seljak, Mon. Not. Roy. Astron. Soc. [336]{}, 1256 (2002) \[arXiv:astro-ph/0205468\]. L. H. Kawano, FERMILAB-Pub-92/04-A (1992). C. Angulo [et al.]{}, Nucl. Phys. A. [656]{}, 3 (1999). G. Huey, R. H. Cyburt and B. D. Wandelt, Phys. Rev. D [69]{}, 103503 (2004) \[arXiv:astro-ph/0307080\]. K. Ichikawa and T. Takahashi, Phys. Rev. D [73]{}, 063528 (2006) \[arXiv:astro-ph/0601099\]. J. Hamann, J. Lesgourgues and G. Mangano, JCAP [0803]{}, 004 (2008) \[arXiv:0712.2826 \[astro-ph\]\]. K. Ichikawa, T. Sekiguchi and T. Takahashi, Phys. Rev. D [78]{}, 043509 (2008) \[arXiv:0712.4327 \[astro-ph\]\]. P. D. Serpico, S. Esposito, F. Iocco, G. Mangano, G. Miele and O. Pisanti, JCAP [0412]{}, 010 (2004) \[arXiv:astro-ph/0408076\]. K. Ichikawa, M. Kawasaki and F. Takahashi, Phys. Lett. B [597]{}, 1 (2004) \[arXiv:astro-ph/0402522\]. K. Ichikawa, M. Kawasaki, K. Nakayama, M. Senami and F. Takahashi, JCAP [0705]{}, 008 (2007) \[arXiv:hep-ph/0703034\]. L. M. Krauss and B. Chaboyer, Science [299]{}, 65 (2003). L. Perotto, J. Lesgourgues, S. Hannestad, H. Tu and Y. Y. Y. Wong, JCAP [0610]{}, 013 (2006) \[arXiv:astro-ph/0606227\]. S. Seager, D. D. Sasselov and D. Scott, Astrophys. J. Lett. [523]{}, L1 (1999) arXiv:astro-ph/9909275. J. Chluba and R. A. Sunyaev, Astron. Astrophys. [446]{}, 39 (2006) \[arXiv:astro-ph/0508144\]. A. Lewis, J. Weller and R. Battye, Mon. Not. Roy. Astron. Soc. [373]{}, 561 (2006) \[arXiv:astro-ph/0606552\]. J. Chluba, J. A. Rubino-Martin and R. A. Sunyaev, Mon. Not. Roy. Astron. Soc. [374]{}, 1310 (2007) \[arXiv:astro-ph/0608242\]. J. Chluba and R. A. Sunyaev, arXiv:astro-ph/0702531. E. R. Switzer and C. M. Hirata, Phys. Rev. D [77]{}, 083006 (2008) \[arXiv:astro-ph/0702143\]. C. M. Hirata and E. R. Switzer, Phys. Rev. D [77]{}, 083007 (2008) \[arXiv:astro-ph/0702144\]. E. R. Switzer and C. M. Hirata, Phys. Rev. D [77]{}, 083008 (2008) \[arXiv:astro-ph/0702145\]. W. Y. Wong, A. Moss and D. Scott, arXiv:0711.1357 \[astro-ph\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We construct a continuous 1-parameter family of smooth complete Ricci-flat metrics of cohomogeneity one on vector bundles over $\mathbb{CP}^2$, $\mathbb{HP}^2$ and $\mathbb{OP}^2$ with respective principal orbits $G/K$ the Wallach spaces $SU(3)/T^2$, $Sp(3)/(Sp(1)Sp(1)Sp(1))$ and $F_4/{\mathrm{Spin}}(8)$. Almost all the Ricci-flat metrics constructed have generic holonomy. The only exception is the complete $G_2$ metric discovered in [@bryant1989construction][@gibbons1990Einstein]. It lies in the interior of the 1-parameter family on $\bigwedge_-^2\mathbb{CP}^2$. All the Ricci-flat metrics constructed have asymptotically conical limits given by the metric cone over a suitable multiple of the normal Einstein metric on $G/K$.' author: - Hanci Chi bibliography: - 'C1EMoW.bib' title: 'Invariant Ricci-flat Metrics of Cohomogeneity One with Wallach Spaces as Principal Orbits' --- Introduction ============ Background and Main Result -------------------------- A Riemannian manifold $(M,g)$ is Ricci-flat if its Ricci curvature vanishes: $$\label{general Ricci-flat} {\mathrm{Ric}}(g)=0.$$ A Ricci-flat manifold is the Euclidean analogy of a vacuum solution of the Einstein field equations. In this article, we study complete noncompact Ricci-flat manifolds of cohomogeneity one. A Riemannian manifold $(M,g)$ is of cohomogeneity one if a Lie Group $G$ acts isometrically on $M$ such that the principal orbit $G/K$ is of codimension one. The Ricci-flat condition is then reduced to a system of ODEs. Many examples of cohomogeneity one Ricci-flat metrics have special holonomy. These include the first example of an inhomogeneous Einstein metric, which is also a Kähler metric. It was constructed in [@calabi1975construction] on a non-compact open set of $\mathbb{C}^n$. A complete Calabi–Yau metric was constructed on $T^\mathbb{S}^2$ independently in [@calabi1979metriques][@eguchi1979self]. The construction was generalized to $T^\mathbb{CP}^n$ in [@calabi1979metriques] and those Ricci-flat metrics are hyper-Kähler. Cohomogeneity one Kähler–Einstein metrics were constructed on complex line bundles over a product of compact Kähler–Einstein manifolds in [@bergery1982nouvelles][@KEcoho1]. Complete metrics with $G_2$ or ${\mathrm{Spin}}(7)$ holonomy can be found in [@bryant1989construction][@gibbons1990Einstein] [@cvetivc2002cohomogeneity][@cvetivc2004new][@foscolo_infinitely_2018]. Ricci-flat metrics with generic holonomy, for example, were constructed on various vector bundles in [@bergery1982nouvelles][@bohm_inhomogeneous_1998][@wang1998Einstein][@chen2011examples]. It is further shown in [@buzano2015family][@buzano_non-kahler_2015] that for infinitely many dimensions, there exist examples which are homeomorphic but not diffeomorphic. The case where the isotropy representation of the principal orbit contains exactly two inequivalent irreducible summands was studied in [@bohm_inhomogeneous_1998][@wink2017cohomogeneity]. In this article, we consider examples with three inequivalent summands. Specifically, let $(G,H,K)$ be one of $$\label{Cases} \begin{split} &\text{I. } (SU(3),S(U(2)U(1)),S(U(1)U(1)U(1))),\\ &\text{II. } (Sp(3),Sp(2)Sp(1),Sp(1)Sp(1)Sp(1)),\\ &\text{III. }(F_4,{\mathrm{Spin}}(9),{\mathrm{Spin}}(8)). \end{split}$$ For these triples, we construct Ricci-flat metrics on the corresponding cohomogeneity one vector bundles $M$ with unit sphere bundle $H/K\hookrightarrow G/K\rightarrow G/H.$ The singular orbits $G/H$’s are respectively $\mathbb{CP}^2$, $\mathbb{HP}^2$ and $\mathbb{OP}^2$. The principal orbits $G/K$’s are Wallach spaces. They appeared explicitly in Wallach classification of even dimensional homogeneous manifolds with positive sectional curvature [@wallach1972compact]. Throughout this paper, the letters $j,k,l$ will denote three distinct numbers in $\{1,2,3\}$ whenever more than one of them appear in a formula together. Let $d=\dim(H/K)$ and $n=\dim(G/K)$. As will be shown in Section \[Smoothness\], each $M$ is in fact an irreducible (sub)bundle of $\bigwedge^{d}_-T^(G/H)$. In all three cases, we can rescale the normal metric on $G/K$ to a metric $Q$, whose restriction on $H/K$ is the standard metric with constant sectional curvature 1. Take $Q$ as the background metric for $G/K$. As will be shown in Section \[Cohomogeneity One Ricci-flat Equation\], the isotropy representation $\mathfrak{g}/\mathfrak{k}$ has $\mathbb{Z}_3$-symmetry among its three inequivalent irreducible summands. By Schur’s lemma, any $G$-invariant metric on $G/K$ has the form $$\label{invariant metric on g/k} g_{G/K}=f_1^2\left.Q\right\|_{\mathfrak{p}_{1}}\oplus f_2^2\left.Q\right\|_{\mathfrak{p}_{2}}\oplus f_3^2\left.Q\right\|_{\mathfrak{p}_{3}}$$ for some $f_j>0$. Correspondingly, the Ricci endomorphism $r$ of $G/K$, defined by $g_{G/K}(r(\cdot),\cdot)={\mathrm{Ric}}(\cdot,\cdot)$, has the form $$\label{ricci on principaltotal}r=r_1\left.Q\right\|_{\mathfrak{p}_{1}}\oplus r_2\left.Q\right\|_{\mathfrak{p}_{2}}\oplus r_3\left.Q\right\|_{\mathfrak{p}_{3}},$$ where $$\label{ricci on principal} \begin{split} &r_j=\dfrac{a}{f_j^2}+b\left(\dfrac{f_j^2}{f_k^2f_l^2}-\dfrac{f_k^2}{f_j^2f_l^2}-\dfrac{f_l^2}{f_j^2f_k^2}\right) \end{split}$$ for some constants $a$ and $b$. Their values were computed in [@nikonorov2014classification]. We have Case $d$ $n$ $a$ $b$ ------ ----- ----- --------------- --------------- I 2 6 $\frac{3}{2}$ $\frac{1}{4}$ II 4 12 $4$ $\frac{1}{2}$ III 8 24 $9$ $1$ : []{data-label="constant"} . \[scalr of spher basic ineq\] A basic observation on $a$ and $b$ is $a-2b=d-1.$ This is not surprising since $Q$ is the sectional curvature 1 metric on $\mathbb{S}^{d}$. Another observation is $ a-6b\geq 0,$ where the equality is achieved in Case I. These observations are frequently used in this article, especially in Section \[Completensese\] and Section \[Entrance Zone\]. Note that all three possible $\frac{f_j^2}{f_k^2f_l^2}$’s appear in . An important motivation for our choices of principal orbits to consider is to study the complications that arise from the simultaneous presence of the terms $\frac{f_1^2}{f_2^2f_3^2}$,$\frac{f_2^2}{f_1^2f_3^2}$ and $\frac{f_3^2}{f_1^2f_2^2}$. If two of $f_j$’s are identical, say $f_2\equiv f_3$, the Ricci endomorphism takes a simpler form, with $r_1=\frac{a-2b}{f_1^2}+b\frac{f_1^2}{f_2^4}$ and $r_2\equiv r_3=\frac{a}{f_2^2}-b\frac{f_1^2}{f_2^4}$. The Ricci-flat ODE system for this special case then reduces to the one for $\mathfrak{g}/\mathfrak{k}$ with two inequivalent irreducible summands considered in [@bohm_inhomogeneous_1998][@wink2017cohomogeneity]. It is noteworthy that the functional $\widehat{\mathcal{G}}$ introduced in [@bohm_inhomogeneous_1998] does not have any positive real root for Case I. Nevertheless, the two summands case can be viewed as the subsystem of the ODE system studied in this article. The invariant compact set constructed in Section \[Completensese\] can be used to prove the existence of complete Ricci-flat metric for this special case. With the condition $f_2\equiv f_3$ relaxed, we prove the following theorem. \[main 1\] There exists a continuous 1-parameter family of non-homothetic complete smooth invariant Ricci-flat metrics on each $M$. Ricci-flat metrics constructed in Case II and Case III all have generic holonomy. In Case I, the 1-parameter family of smooth Ricci-flat metrics contains in its interior the complete smooth $G_2$ metric that was first constructed in [@bryant1989construction][@gibbons1990Einstein]. The other metrics in that family all have generic holonomy. Therefore, for $M$ in Case I, the moduli space $\mathcal{M}_{G_2}$ of $G_2$ metric is not* isolated in $\mathcal{M}_0$ the moduli space of Ricci-flat metric in the $C^0$ sense. Such a phenomenon cannot occur on a simply connected spin closed manifold, for example, by Theorem 3.1 in [@wang1991preserving]. \[AC\] Let $(N,g_N)$ and $(M,g_M)$ be Riemannian manifolds of respective dimension $n$ and $n+1$. Let $t$ be the geodesic distance from some point on $M$. Then $M$ has one asymptotically conical (AC) end if there exists a compact subset $\check{M}\subset M$ such that $M\backslash \check{M}$ is diffeomorphic to $(1,\infty)\times N$ with $g_M=dt^2+t^2g_N+o(1)$ as $t\to \infty.$ With further analysis on the asymptotic behavior of Ricci-flat metrics in Theorem \[main 1\], we are able to prove the following: \[main 2\] Each Ricci-flat metric in Theorem \[main 1\] has an AC end with limit the metric cone over a suitable multiple of the normal Einstein metric on $G/K$. In Case I, the normal Einstein metric on the principal orbit $SU(3)/T^2$ admits a (strict) nearly Kähler structure. Hence the metric cone over $G/K$ is the singular $G_2$ metric which was first constructed in [@bryant1987metrics]. The other two principal orbits, however, do not admit (strict) nearly Kähler structure [@davila_homogeneous_2012]. Organization ------------ This paper is structured as followings. In Section \[Cohomogeneity One Ricci-flat Equation\] and \[Smoothness\], we discuss some details of the geometry of the cohomogeneity one manifolds $M$. Based on the work in [@eschenburg2000initial], we reduce to a system of ODEs with a conservation law . A $G$-invariant Ricci-flat metric around $G/H$ is hence represented by an integral curve defined on $[0,\epsilon)$. We derive the condition for smooth extension to $G/H$ using Lemma 1.1 in [@eschenburg2000initial]. If in addition, the integral curve is defined on $[0,\infty)$, the corresponding Ricci-flat metric is complete. In Section \[Local Existence\], we apply the coordinate change introduced in [@dancer2008non][@dancer2008some]. The ODE system is transformed to a polynomial one. Invariant Einstein metrics on $G/H$ and $G/K$ are transformed to critical points of the new system. We carry out linearizations at these critical points and prove the local existence of invariant Ricci-flat metrics around $G/H$. An integral curve defined on $[0,\epsilon)$ is transformed to a new one that is defined on $(-\infty,\epsilon')$ for some $\epsilon'\in\mathbb{R}$. Each integral curve represents a Ricci-flat metric on $M$ up to homothety. It is determined by a parameter $s_1$ that controls the principal curvature of $G/H$ at $t=0$. To show the completeness of the metric is equivalent to proving that the new integral curve is defined on $\mathbb{R}$. The proof of completeness of the metric is divided into two sections. In Section \[Completensese\], we construct a compact invariant set whose boundary contains critical points that represent the invariant metric on $G/H$ and the normal Einstein metric on $G/K$. The construction is almost the same for all three cases with a little difference in Case I. Section \[Entrance Zone\] proves that as long as $s_1$ is close enough to zero, integral curves of Ricci-flat metrics enter the compact invariant set constructed in Section \[Completensese\] in finite time, hence proving the completeness. In Section \[aymptoticlimit\], we analyze the asymptotic behavior of all the Ricci-flat metrics constructed in Section \[Entrance Zone\]. There also exist solutions to the polynomial system that represent singular Ricci-flat metrics. They are discussed in Section \[a digression\]. Results in this article are summarized by a plot at the end. With similar techniques introduced in Section \[Completensesetotal\], we can also show that there exists a 2-parameter family of Poincaré–Einstein metrics on each $M$. More details will appear in another upcoming article. Acknowledgement. The author is grateful to his PhD supervisor, Prof. McKenzie Wang for his guidance and encouragement. Local Solution Near Singular Orbit {#Local EXISTENCEEE} ================================== Cohomogeneity One Ricci-flat Equation {#Cohomogeneity One Ricci-flat Equation} ------------------------------------- In this section, we derive the system of ODEs whose solutions give Ricci-flat metrics of cohomogeneity one on $M$. Since $M$ is of cohomogeneity one, there is a $G$-diffeomorphism between $M\backslash({G/H})$ and $(0,\infty)\times G/K$. We construct a Ricci-flat metric $g$ on $M$ by setting $(0,\infty)$ as a geodesic and assigning a $G$-invariant metric $g_{G/K}$ to each hypersurface $\{t\}\times G/K$, i.e., define $$\label{Ricci-flat metric on M} g=dt^2+g_{G/K}(t)$$ on $M$. By [@eschenburg2000initial], if $g_{G/K}(t)$ satisfies $$\label{old Ricci-flat1} \dot{g}_{G/K}=2g_{G/K}(L\cdot,\cdot),$$ $$\label{old Ricci-flat2} \dot{L}=-{\mathrm{tr}}(L)L+r,$$ $$\label{old Ricci-flat3} {\mathrm{tr}}(\dot{L})=-{\mathrm{tr}}(L^2),$$ $$\label{old Ricci-flat4} d({\mathrm{tr}}(L))+\delta^{\nabla} L=0,$$ on $(0,\epsilon)$, where $\delta^{\nabla}\colon \Omega^1(G/K,T(G/K))\rightarrow T^(G/K)$ is the divergence operator composed with the musical isomorphism, then $g$ is a Ricci-flat metric on $(0,\epsilon)\times G/K$. Note that provides a formula for computing $L(t)$ the shape operator of hypersurface $\{t\}\times G/K$ for each $t\in (0,\epsilon)$. By [@back1986local] and [@eschenburg2000initial], Equation automatically holds for a $C^3$ metric satisfying and if there exists a singular orbit of dimension smaller than $\dim(G/K)$. Canceling the term ${\mathrm{tr}}{(\dot{L})}$ using and yields the conservation law $$\label{old conservation} R-({\mathrm{tr}}{(L)})^2+{\mathrm{tr}}{(L^2)}=0.$$ We shall focus on deriving specific formulas for , and on $M$. It requires a closer look at isotropy representations of $G/K$ and $G/H$. We fix notations first. Each irreducible complex representation is characterized by inner products between the dominant weight and simple roots on nodes of the corresponding Dynkin diagram. We use $[a]$ for class $A_1=B_1=C_1$; $[a,b]$ for $C_2=B_2$ with the shorter root on the right end; $[a,b,c,d]$ for $B_4$ with the shorter root on the right end. Furthermore, let $\mathfrak{t}$ be the Lie algebra of $S(U(1)U(1)U(1))$. Choose $Q$-orthogonal decomposition $\mathfrak{t}=\mathfrak{t}_1\oplus\mathfrak{t}_2$, where $$\mathfrak{t}_1={\mathrm{span}}_\mathbb{R}\left\{ \begin{bmatrix} i&&\\ &-i&\\ &&0 \end{bmatrix} \right\},\quad \mathfrak{t}_2={\mathrm{span}}_\mathbb{R}\left\{ \begin{bmatrix} i&&\\ &i&\\ &&-2i \end{bmatrix} \right\}.$$ Let $\theta_j^a$ denote the complexified irreducible representation of circle generated by $\mathfrak{t}_j$ with weight $a$. We use $\Lambda_8$ and $\Delta_8^\pm$ to respectively denote the complexified standard representation and spin representations of ${\mathrm{Spin}}(8)$. We use $\mathbb{I}$ to denote the trivial representation. The formula of $g_{G/K}$ is given by . With $(G,H,K)$ listed in , we have the following $Q$-orthogonal decomposition for $\mathfrak{g}$: $$\label{decomposition} \begin{split} \mathfrak{g}&=\mathfrak{h}\oplus \mathfrak{q} \quad \text{as a representation of $\left.Ad(G)\right\|_{H}$}\\ &=(\mathfrak{k}\oplus \mathfrak{p}_{1})\oplus (\mathfrak{p}_{2}\oplus\mathfrak{p}_{3})\quad \text{as a representation of $\left.Ad(G)\right\|_{K}$}. \end{split}$$ Irreducible $K$-modules $\mathfrak{p}_{j}$’s are all of dimension $d$ but they are inequivalent to each other. Specifically, we have Table \[Kmodule\]. Case $\mathfrak{p}_{1}\otimes \mathbb{C}$ $\mathfrak{p}_{2}\otimes \mathbb{C}$ $\mathfrak{p}_{3}\otimes \mathbb{C}$ ------ -------------------------------------- -------------------------------------- -------------------------------------- I $\theta_1^2\otimes\mathbb{I}$ $\theta_1^1\otimes\theta_2^3$ $\theta_1^{-1}\otimes\theta_2^3$ II $[1]\otimes[1]\otimes\mathbb{I}$ $[1]\otimes\mathbb{I}\otimes[1]$ $\mathbb{I}\otimes[1]\otimes[1]$ III $\Lambda_8$ $\Delta_8^+$ $\Delta_8^-$ : []{data-label="Kmodule"} By Schur’s lemma, a $G$-invariant metric on $G/K$ has the form of . \[metric\] The formula of Ricci endomorphism on $(G/K,g_{G/K})$ is given by and with constants $a$ and $b$ listed in Table \[constant\]. Since the Ricci endomorphism is also $G$-invariant, it has the form of . To compute its formula, use (7.39) in [@besse2007einstein] to derive the scalar curvature on $G/K$ and then apply variation. For each case, since $[\mathfrak{p}_{j},\mathfrak{p}_{j}]\subset \mathfrak{k}$ and $[\mathfrak{p}_j,\mathfrak{p}_k]\subset \mathfrak{p}_l$, each $r_j$ in has the form of . Take $M$ as an associated vector bundle to principal $H$-bundle $G\rightarrow G/H$ of cohomogeneity one. As the orbit space is of dimension one, the action of $H$ on the unit sphere of $\mathbb{R}^{d+1}$ must be transitive. Then the group $K$ is taken as an isotropy group of a fixed nonzero element in $\mathbb{R}^{d+1}$, say $v_0=(1,0,\dots,0)$. It is clear that $H/K=\mathbb{S}^d$. Hence $G/K$ is indeed a unit sphere bundle over $G/H$. In this setting, $g_{G/K}(t)$ is an $S^2(\mathfrak{p}_{1}\oplus\mathfrak{p}_{2}\oplus\mathfrak{p}_{3})^{K}$-valued function with each $f_j$ in as a positive function. Correspondingly, the Ricci endomorphism $r$ in is also an $S^2(\mathfrak{p}_{1}\oplus\mathfrak{p}_{2}\oplus\mathfrak{p}_{3})^{K}$-valued function. \[ricci\] For $(G,H,K)$ listed in , Ricci-flat conditions and respectively become $$\label{Ricci-flat0} L=\frac{\dot{f}_1}{f_1}\left.Q\right\|_{\mathfrak{p}_{1}}\oplus \frac{\dot{f}_2}{f_2}\left.Q\right\|_{\mathfrak{p}_{2}}\oplus \frac{\dot{f}_3}{f_3}\left.Q\right\|_{\mathfrak{p}_{3}},$$ $$\label{Ricci-flat1} \dfrac{\ddot{f_j}}{f_j}-\left(\dfrac{\dot{f_j}}{f_j}\right)^2=-\left(d\frac{\dot{f_1}}{f_1}+d\frac{\dot{f_2}}{f_2}+d\frac{\dot{f_3}}{f_3}\right)\dfrac{\dot{f_j}}{f_j}+r_j,\quad j=1,2,3$$ and $$\label{old conservation law} \begin{split} &-d\sum_{j=1}^3\left(\dfrac{\dot{f_j}}{f_j}\right)^2=-\left(\sum_{j=1}^3d\dfrac{\dot{f_j}}{f_j}\right)^2+R. \end{split}$$ The proof is complete by computation results in Proposition \[metric\] and Proposition \[ricci\]. In summary, constructing a smooth complete cohomogeneity one Ricci-flat metric on $M$ is essentially equivalent to solving $g_{G/K}(t)$ that satisfies , and . The fundamental theorem of ODE guarantees the existence of solution on neighborhood around $\{t_0\}\times G/K$ for any $t_0\in (0,\infty)$. In order to have a smooth complete Ricci-flat metric on $M$, we need to show that 1\. (Smooth extension) the solution exists on a tubular neighborhood around $G/H$ and extends smoothly to the singular orbit; 2\. (Completeness) the solution exists on $(0,\infty)\times G/K$. We discuss the smooth extension in Section \[Smoothness\] and Section \[Local Existence\]. The proof for completeness is in Section \[Completensesetotal\]. Smoothness Extension {#Smoothness} -------------------- It is not difficult to guarantee the smoothness of $g_{G/K}(t)$ at $t=0$ as a $S^2(\mathfrak{p}_{1}\oplus\mathfrak{p}_{2}\oplus\mathfrak{p}_{3})$-valued function. However, the smooth function does not guarantee the smooth extension of $g=dt^2+g_{G/K}(t)$ as a metric on $G/H$ as $t\to 0$. By Lemma 1.1 in [@eschenburg2000initial], the question boils down to studying the slice representation $\chi=\mathbb{R}^{d+1}$ of M and the isotropy representation $\mathfrak{q}$ of $G/H$. We rephrase the lemma below. [@eschenburg2000initial] \[EW\] Let $g(t): [0,\infty)\rightarrow S^2(\chi\oplus \mathfrak{q})^K$ be a smooth curve with Taylor expansion at $t=0$ as $\sum_{l=0}^\infty g_lt^l$. Let $W_l={\mathrm{Hom}}(S^l(\chi),S^2(\chi\oplus \mathfrak{q}))^H$ be the space of $H$-equivariant homogeneous polynomials of degree $l$. Let $\iota\colon W_l\rightarrow S^2(\chi\oplus \mathfrak{q})$ denote the evaluation map at $v_0=(1,0,\dots,0)$. Then the map $g(t)$ has a smooth extension to $G/H$ as a symmetric tensor if and only if $g_l\in\iota(W_l)$ for all $l$. To compute $W_l$, we need to identify $\chi$ and $\mathfrak{q}$ first. Since $H$ acts transitively on $H/K$, the slice representation $\chi=\mathbb{R}^{d+1}$ of $M$ is irreducible and hence can be identified. Recall that $\mathfrak{q}$ is an irreducible $H$-module in decomposition . Hence we have Table \[Slicerepn\]. Case $\chi\otimes\mathbb{C}$ as an $H$-module $\chi \otimes\mathbb{C}$ as a $K$-module $\mathfrak{q}\otimes\mathbb{C}$ as an $H$-module ------ ------------------------------------------ ---------------------------------------------------- ------------------------------------------------------------ I $[2]\otimes \mathbb{I}$ $\mathbb{R}\oplus(\theta_1^2\otimes\mathbb{I})$ $([1]\otimes \theta_2^3)\oplus ([1]\otimes \theta_2^{-3})$ II $[1,0]\otimes \mathbb{I}$ $\mathbb{R}\oplus([1]\otimes[1]\otimes\mathbb{I})$ $[0,1]\otimes[1]$ III $[1,0,0,0]$ $\mathbb{R}\oplus \Lambda_8$ $[0,0,0,1]$ : []{data-label="Slicerepn"} . \[spin(d+1)\] Recall the background metric $Q$ on $G/K$ is chosen that $\left.Q\right\|_{\mathfrak{p}_1}$ is the standard metric on $\mathbb{S}^d$. Therefore, the Euclidean inner product $\langle\cdot,\cdot\rangle$ on $\chi$ can be written in “polar coordinate” as $dt^2+t^2\left.Q\right\|_{\mathfrak{p}_1}$. As shown in the first column of Table \[Slicerepn\], the action of $H$ is essentially the standard representation of ${\mathrm{Spin}}(d+1)$ on $\chi$ and it preserves $\langle\cdot,\cdot\rangle$. In the following discussion, we take $\langle\cdot,\cdot\rangle\oplus\left.Q\right\|_{\mathfrak{p}_2}\oplus\left.Q\right\|_{\mathfrak{p}_3}$ as the background metric of $T_pM=\chi\oplus T_p(G/H)$ for $p=[H]\in G/H$. Compare the second column of Table \[Slicerepn\] to the first column of Table \[Kmodule\]. It is clear that $\chi=\mathbb{R}\oplus \mathfrak{p}_{1}$ as a $K$-module. Since $\chi$ and $\mathfrak{q}$ are inequivalent $H$-modules, we have $$\label{decomposition2} S^2(\chi\oplus \mathfrak{q})^{K}=S^2(\chi)^{K}\oplus S^2(\mathfrak{q})^{K}.$$ Hence we have decomposition $W_l=W_l^+\oplus W_l^-$ where $W_l^+$ and $W_l^-$ are respectively valued in $S^2(\chi)$ and $S^2(\mathfrak{q})$. We are ready to compute each $W_l^\pm$. \[EACHW\] For each $M$, we have $$W_l^+\cong\left\{\begin{array}{ll} \mathbb{R}& l=0\\ 0& l\equiv1\mod 2\\ \mathbb{R}^2& l\equiv 0\mod 2,\quad l\geq 2\\ \end{array}\right.,\quad W_l^-\cong\mathbb{R}$$ From Table \[Slicerepn\], we can derive the decomposition of complexified symmetric products $S^l(\chi)\otimes \mathbb{C}$ and $S^l(\mathfrak{q})\otimes \mathbb{C}$ as $H$-modules, as shown in Table \[Taylor\] below. The proof is complete. Case $S^{2m-1}(\chi)\otimes\mathbb{C}$ $S^{2m}(\chi)\otimes\mathbb{C}$ $S^2(\mathfrak{q})\otimes\mathbb{C}$ ------ ------------------------------------------------------------ ---------------------------------------------------------- ------------------------------------------------------------------------------------------------------------ I $\bigoplus^m_{i=1}\left([4i-2]\otimes \mathbb{I}\right)$ $\bigoplus^m_{i=0}\left([4i]\otimes \mathbb{I}\right)$ $([2]\otimes \theta_2^{6} )\oplus([2]\otimes \theta_2^{-6} )\oplus([2]\otimes \mathbb{I})\oplus\mathbb{I}$ II $\bigoplus^m_{i=1}\left([2i-1,0]\otimes \mathbb{I}\right)$ $\bigoplus^m_{i=0}\left([2i,0]\otimes \mathbb{I}\right)$ $([0,2]\otimes[2])\oplus([1,0]\otimes\mathbb{I})\oplus\mathbb{I}$ III $\bigoplus^m_{i=1}[2i-1,0,0,0]$ $\bigoplus^m_{i=0}[2i,0,0,0]$ $[0,0,0,2]\oplus[1,0,0,0]\oplus\mathbb{I}$ : []{data-label="Taylor"} In order to apply Lemma \[EW\], we need to find generators of each $W_l^\pm$ in Proposition \[EACHW\]. Note that $W_l^{\pm}$ can be viewed as subspaces of $W_{l+2}^{\pm}$ by multiplying each element with $\sum_{i=0}^dx_i^2$. Hence we only need to find generators of $W_0^\pm$ , $W_2^+$ and $W_1^-$. It is clear that $W_0^+$ is spanned by $I_{d+1}\in S^2(\chi)$ and $W_0^-$ is spanned by $I_{2d}\in S^2(\mathfrak{q})$. It is also clear that $W_2^+$ is generated by the identity map and $(\sum_{i=0}^d x_i^2)I_{d+1}$. Note that the identity map in the form of a homogeneous polynomial is a symmetric matrix $\Pi$ with $\Pi_{ij}=x_ix_j$ for $i,j\in\{0,1,\dots,d\}$. The computation for $W_1^-$ is a bit more complicated. We follow Chapter 14 in [@harvey_spinors_1990] and consider $\chi=\mathbb{R}\oplus \mathbb{F}$ with $\mathbb{F}$ as one of $\mathbb{C},\mathbb{H}$ and $\mathbb{O}$ for Case I, II and III, respectively. \[w1-\] $W_1^-$ is generated by the $\mathbb{R}$-linear map $$\begin{aligned} \Phi\colon \chi&\rightarrow S^2(\mathfrak{q})\\ (x_0,\mathbf{x})&\mapsto \begin{bmatrix} x_0 I_{d}& \mathsf{L}_\mathbf{x}\\ \mathsf{L}_{\bar{\mathbf{x}}} &-x_0 I_{d}, \end{bmatrix}\end{aligned}$$ where $\mathsf{L}_\mathbf{x}$ is the real matrix representation of left multiplication of $\mathbf{x}\in\mathbb{F}$, as shown in Table \[Only for j1\] below. -------------------------------------------------------------------------------------------------------------- Case I II III ------------------------- ------------------ ---------------------- ------------------------------------------ $\mathsf{L}_\mathbf{x}$ $\begin{bmatrix} $\begin{bmatrix} $\begin{bmatrix} x_1&-x_2\\ x_1&-x_2&-x_3&-x_4\\ x_1&-x_2&-x_3&-x_4&-x_5&-x_6&-x_7&-x_8\\ x_2&x_1 x_2&x_1&-x_4&x_3\\ x_2&x_1 &-x_4&x_3 &-x_6&x_5 &x_8 &-x_7\\ \end{bmatrix} x_3&x_4&x_1&-x_2\\ x_3&x_4 &x_1 &-x_2&-x_7&-x_8&x_5 &x_6\\ $ x_4&-x_3&x_2&x_1 x_4&-x_3&x_2 &x_1 &-x_8&x_7 &-x_6&x_5\\ \end{bmatrix}$ x_5&x_6 &x_7 &x_8 &x_1 &-x_2&-x_3&-x_4\\ x_6&-x_5&x_8 &-x_7&x_2 &x_1 &x_4 &-x_3\\ x_7&-x_8&-x_5&x_6 &x_3 &-x_4&x_1 &x_2\\ x_8&x_7 &-x_6&-x_5&x_4 &x_3 &-x_2&x_1 \end{bmatrix}$ -------------------------------------------------------------------------------------------------------------- : []{data-label="Only for j1"} Consider $i\Phi(\chi)$ a subspace of $\mathbb{C}\otimes_\mathbb{R} S^2(\mathfrak{q})$. Since $(i\Phi(x_0,\mathbf{x}))^2=-(x_0^2+\\|\mathbf{x}\\|^2)I_{d+1}$, it is clear that the matrix multiplication of $i\Phi(\chi)$ generates a Clifford algebra and hence ${\mathrm{Spin}}(d+1)$. Specifically, the group is generated by elements $\Xi(y_0,\mathbf{y}):=\Phi(-1,\mathbf{0})\Phi(y_0,\mathbf{y})$ with $y_0^2+\\|\mathbf{y}\\|^2=1$. Since each $\mathbb{F}$ is an alternative algebra that satisfies Moufang identity, computations show $$Ad(\Xi(y_0,\mathbf{y}))(\Phi(x_0,\mathbf{x}))=\Xi(y_0,\mathbf{y}) (\Phi(x_0,\mathbf{x})) \Xi(y_0,\mathbf{y})^{-1}=\Phi(z_0,\mathbf{z}),$$ where $z_0=(y_0^2-\\|\mathbf{y}\\|)x_0+2y_0\langle\mathbf{y},\mathbf{x}\rangle$ and $\mathbf{z}=y_0^2\mathbf{x}-2x_0y_0\mathbf{y}-(\mathbf{y}\bar{\mathbf{x}})\mathbf{y}$. Hence $\Phi(\chi)$ is an $Ad_{{\mathrm{Spin}}(d+1)}$-invariant subspace in $S^2(\mathfrak{q})$. Moreover, since $$(Ad(\Xi(y_0,\mathbf{y}))(\Phi(x_0,\mathbf{x})))^2=(\Phi(x_0,\mathbf{x})))^2=(x_0^2+\\|\mathbf{x}\\|^2)I_{d+1},$$ The adjoint action on $\Phi(\chi)$ induces the standard representation $\Lambda_{d+1}$ on $\mathbb{R}^{d+1}$. Therefore, $$\begin{aligned} \Phi\colon (\chi,\Lambda_{d+1})&\rightarrow \left(\Phi(\chi),Ad_{{\mathrm{Spin}}(d+1)}\right)\end{aligned}$$ is $H$-equivariant and generates $W_1^-$. With the generators known, we are ready to prove the following proposition. \[Smooooooth\] The necessary and sufficient conditions for a metric $g=dt^2+g_{G/K}(t)$ on $M$ to extend to a smooth metric in a tubular neighborhood of the singular orbit $G/H$ are $$\label{old initial} \lim_{t\to 0}(f_1,f_2,f_3,\dot{f_1},\dot{f_2},\dot{f_3}) =(0,h_0,h_0,1,-h_1,h_1)$$ for some $h_0>0$ and $h_1\in\mathbb{R}$. The metric $g$ in LHS of can be identified with a map $$\label{metric map!} g(t)\colon [0,\epsilon) \rightarrow S^2(\chi)^{K}\oplus S^2(\mathfrak{q})^{K}$$ with Taylor expansion $$\label{taylorforg} g(t)=\sum_{l=0}^\infty g_lt^l.$$ Write $g(t)=D(t)\oplus J(t)$, where $D(t)\colon [0,\infty)\rightarrow S^2(\chi)$ and $J(t)\colon [0,\infty)\rightarrow S^2(\mathfrak{q})$. The Taylor expansion can be rewritten as $$\label{specific taylor} \begin{split} &D(t)=D_0+D_1t+D_2t^2+\dots\\ &J(t)=J_0+J_1t+J_2t^2+\dots \end{split}$$ Since $W_2^+/W_0^+\cong \mathbb{R}$, in principle there is a free variable for the second derivative of a smooth $D(t)$. However, with the geometric setting that $t$ is a unit speed geodesic, the choice of $D_2$ is in fact determined by $D_0$. Hence we take $D_0=I_{d+1}$ and $D_2$ must be a multiple of $\left((\sum_{i=0}^d x_i^2)I_{d+1}-\Pi\right)(v_0)= \begin{bmatrix} 0&\\ &I_d \end{bmatrix} $ with the multiplier determined by the choice of $D_0$. Since $H/K$ is and irreducible sphere, it is expected that there is no indeterminacy from $D(t)$. By Lemma \[EW\], the smooth condition for $D(t)$ with respect to background metric $\langle\cdot,\cdot\rangle$ is $ D(t)=I_{d+1}+O(t^2). $ This is consistent with Lemma 9.114 in [@besse2007einstein]. As $g$ degenerates to an invariant metric on $G/H$ and the isotropy representation of $G/H$ is irreducible, $J_0$ is a positive multiple of $I_{2d}$. The evaluation of $\Phi$ at $v_0$ in Proposition \[w1-\] is $\begin{bmatrix} I_d&\\ &-I_d \end{bmatrix}. $ Hence by Lemma \[EW\], the smoothness condition for $J(t)$ is $$J(t)=\begin{bmatrix} f_2^2(t)I_{d}&\\ &f_3^2(t)I_{d} \end{bmatrix}=c_0I_{2d}+c_1\begin{bmatrix} I_d&\\ &-I_d \end{bmatrix}t+O(t^2)$$ for some $c_0>0$ and $c_1\in\mathbb{R}$. Recall \[spin(d+1)\], note that $\langle\cdot,\cdot\rangle=dt^2+t^2\left.Q\right\|_{\mathfrak{p}_1}$. Switch the background metric to $dt^2+Q$, we conclude that the smoothness condition for $g$ is $$\begin{split} &f_1^2(t)=t^2+O(t^4)\\ &f_2^2(t)=c_0+c_1t+O(t^2)\\ &f_3^2(t)=c_0-c_1t+O(t^2) \end{split}$$ Then the proof is complete. \[already\] The Ricci-flat ODE system and is invariant under the homothetic change $\kappa^2(dt^2+g_{G/K})$ with $ds=\kappa dt$. The smooth initial condition \[old initial\] is transformed to $(0, \kappa h_0, \kappa h_0, 1, h_1, -h_1).$ Hence if we abuse the notation. Multiplying $h_0$ by $\kappa>0$ while having $\dot{f}_j(0)$ unchanged give the smooth initial condition for metrics in the same homothetic family. Therefore, in the original coordinate, $h_1$ is the free variable that gives non-homothetic metrics. As shown in (2.27), only $h_1$ matters in producing different curves in the polynomial system. Combine the analysis in Proposition \[Smooooooth\] with the main result in [@eschenburg2000initial], we conclude that there exists a 1-parameter family of Ricci-flat metric on a neighborhood around $G/H$ in $M$. We derive the same result in Section \[Local Existence\] using a new coordinate. Note that we always have $\lim\limits_{t\to 0}\frac{\dot{f}_3}{f_3}+\frac{\dot{f}_2}{f_2}=0$, i.e., the mean curvature of $G/H$ vanishes at $t=0$. This is consistent with Corollary 1.1 in [@hsianglawson]. The last two components of shows that the smooth extension does not require $G/H$ to be totally geodesic. If $h_1$ in vanishes, then we recover cases in [@bohm_inhomogeneous_1998][@wink2017cohomogeneity] with $f_2\equiv f_3$. \[symmetry\] It is worth pointing out that Equations and are symmetric among $f_1$, $f_2$ and $f_3$. Therefore, initial condition has two other counterparts where $f_2$ or $f_3$ collapses initially depending how $H$ is embedded in $G$. Without loss of generality, we will consider initial condition in this article. We end this section by identifying each vector bundle $M$ as a (sub)bundle of ASD $d$-form of lowest rank. Table \[MidForm\] lists out $H$-decomposition of $\bigwedge\nolimits^{d}\mathfrak{q}\otimes\mathbb{C}$ and dimension of each irreducible summand. The subspace $\bigwedge\nolimits^{d}_-\mathfrak{q}\otimes\mathbb{C}$ consists of summands in brace brackets. Decomposition below is mostly computed via software $\mathsf{LiE}$, with reference in [@brown1972riemannian][@salamon1989riemannian][@lopez2009canonical]. [ l\| l l]{} Case & $H$ & $H$-decomposition of $\bigwedge\nolimits^{d}\mathfrak{q}\otimes\mathbb{C}$ and Dimension of each Summand\ I & $S(U(2)U(1))$ &$ \begin{array}{r@{}l@{}} \\ \bigwedge\nolimits^2 \left(([1]\otimes \theta_2^3)\oplus([1]\otimes \theta_2^3)\right) & {}=(\mathbb{I}\otimes\theta_2^6)\oplus(\mathbb{I}\otimes\theta_2^{-6})\oplus\mathbb{I}\oplus\{[2]\otimes\mathbb{I}\}\\ \\ \mathbf{6} & {}=\mathbf{1}+\mathbf{1}+\mathbf{1} +\{\mathbf{3}\}\end{array} $\ \ II & $Sp(2)Sp(1)$ & $\begin{array} {r@{}l@{}} \\ \bigwedge\nolimits^4 [01]\otimes[1] & {}=[01]\otimes[2]\oplus[02]\otimes\mathbb{I}\oplus\mathbb{I}\otimes[4]\oplus\mathbb{I}\oplus\{[02]\otimes[2]\oplus[01]\otimes\mathbb{I}\}\\ \\ \mathbf{70} & {}= \mathbf{15}+\mathbf{14}+\mathbf{5}+\mathbf{1}+\{\mathbf{30}+\mathbf{5}\} \end{array}$\ \ III & ${\mathrm{Spin}}(9)$ & $\begin{array} {r@{}l@{}} \\ \bigwedge\nolimits^8 [0001]&{}=[2010]\oplus[0020]\oplus[1002]\oplus[0200]\oplus[4000]\oplus[0010]\oplus[2000]\oplus\mathbb{I} \\ & \quad \oplus\{[2002]\oplus[0110]\oplus[1010]\oplus[3000]\oplus[0002]\oplus[1000]\}\\ \\ \mathbf{12870} & {}= \mathbf{2457}+\mathbf{1980}+\mathbf{924}+\mathbf{495}+\mathbf{450}+\mathbf{84}+\mathbf{44}+\mathbf{1} \\ & \quad +\{\mathbf{3900}+\mathbf{1650}+\mathbf{594}+\mathbf{156}+\mathbf{126}+\mathbf{9}\}\end{array}$\ \ For Case I, it is known that the trivial representation generates the invariant Kähler form on $\mathbb{CP}^2$. The bundle that we study in this paper is the associated bundle with respect to representation $[2]\otimes\mathbb{I}$, which is the bundle of ASD 2-form $\bigwedge^2_-T^\mathbb{CP}^2$ that admits a complete smooth $G_2$ metric [@bryant1989construction][@gibbons1990Einstein]. For Case II, the trivial representation generates a canonical 4-form for Quaternionic Kähler manifolds, as described in [@salamon1989riemannian]. Explicitly, given a Quaternionic Kähler manifold with a triple of complex structures $(I,J,K)$ and corresponding symplectic forms $(\omega_I,\omega_J,\omega_K)$, the canonical 4-form is defined as $ \Omega=\omega_I\wedge\omega_I+\omega_J\wedge\omega_J+\omega_K\wedge\omega_K. $ By Table \[Slicerepn\], $M$ is an associate bundle with respect to representation $[01]\otimes\mathbb{I}$ in $\bigwedge_-^4\mathfrak{q}_2\otimes\mathbb{C}$. Therefore, $M$ is indeed an irreducible subbundle of $\bigwedge_-^4T^\mathbb{HP}^2$. For Case III, the trivial representation generates the canonical 8-form, whose existence is proved in [@brown1972riemannian]. Explicit formula for the canonical 8-form can be found in [@lopez2009canonical]. The 9-dimensional representation $[1000]$ is the (twisted) adjoint representation of ${\mathrm{Spin}}(9)$ on $\mathbb{R}^9$. Similar to Case II, the bundle that we consider in this paper is an irreducible subbundle of $\bigwedge^8_-T^\mathbb{OP}^2$. In conclusion, the name “(sub)bundle of ASD $d$-form of lowest rank” for $M$ is justified. Coordinate Change and Linearization {#Local Existence} ----------------------------------- We apply the coordinate change introduced in [@dancer2008non][@dancer2008some] to the Ricci-flat system in this section. The original ODE system is transformed to a polynomial one. As described in Remark \[important criti\], some critical points of the new system carry geometric data. Linearizations at these critical points provide guidance on how integral curves potentially behave, which help us to construct a compact invariant set in Section \[Completensesetotal\] to prove the completeness. As predicted by the result in the previous section (Remark \[already\]), analysis on the new system shows that there exists a 1-parameter family of integral curves with each represents a homothetic class of Ricci-flat metrics on a neighborhood around $G/H$. Consider $$\label{coordinate change} d\eta={\mathrm{tr}}(L) dt.$$ Define $$\label{def of variable} X_j:=\dfrac{\frac{\dot{f_j}}{f_j}}{{\mathrm{tr}}(L)},\quad Z_j:=\dfrac{\frac{f_j}{f_kf_l}}{{\mathrm{tr}}(L)}.$$ And define $$\mathcal{R}_j:=\frac{r_j}{({\mathrm{tr}}(L))^2}=aZ_kZ_l+b\left(Z_j^2-Z_k^2-Z_l^2\right),\quad \mathcal{G}:=\sum_{j=1}^3 d X_j^2, \quad \mathcal{H}:=\sum_{j=1}^3dX_j.$$ Use $'$ to denote derivative with respect to $\eta$. In the new coordinates given by and , the system is transformed to $$\label{new Ricci-flat} \begin{split} \begin{bmatrix} X_1\\ X_2\\ X_3\\ Z_1\\ Z_2\\ Z_3 \end{bmatrix}'=V(X_1,X_2,X_3,Z_1,Z_2,Z_3):=\begin{bmatrix} X_1(\mathcal{G}-1)+\mathcal{R}_1\\ X_2(\mathcal{G}-1)+\mathcal{R}_2\\ X_3(\mathcal{G}-1)+\mathcal{R}_3\\ Z_1\left(\mathcal{G}-\frac{\mathcal{H}}{d}+2X_1\right)\\ Z_2\left(\mathcal{G}-\frac{\mathcal{H}}{d}+2X_2\right)\\ Z_3\left(\mathcal{G}-\frac{\mathcal{H}}{d}+2X_3\right) \end{bmatrix}, \end{split}$$ and the conservation law becomes $$\label{conservation law} C\colon \mathcal{G}-1+d\sum_j \mathcal{R}_j=0.$$ As $\left(\frac{1}{{\mathrm{tr}}(L)}\right)'=\frac{\mathcal{G}}{{\mathrm{tr}}(L)}$, the original variables can be recovered by $$\label{recovery} t=\int_{\eta_0}^{\eta} \exp\left(\int_{\tilde{\eta}_0}^{\tilde{\eta}} \mathcal{G} d\tilde{\tilde{\eta}}+\tilde{t}_0\right) d\tilde{\eta}+t_0,\quad f_j=\dfrac{\exp\left(\int_{\eta_0}^\eta \mathcal{G} d\tilde{\eta}+t_0\right)}{\sqrt{Z_kZ_l}}.$$ \[themeaningofz\] The new variables $X_j$’s record the relative size of each principal curvature of $G/K$. Variables $Z_j$’s carry the data of relative size of each $f_j$’s. Note that $\frac{Z_j}{Z_k}=\frac{f_j^2}{f_k^2}$. In the original coordinates, a smooth solution to is an integral curve with variable $t\in [0,\epsilon)$. Since by , $\lim\limits_{t\to 0}\eta=\lim\limits_{t\to 0}\ln\left(f_1^{d}f_2^{d}f_3^{d}\right)+\hat{\eta}=-\infty$, the original solution is transformed to an integral curve with variable $\eta\in (-\infty,\epsilon')$ for some $\epsilon'\in\mathbb{R}$. Note that the graph of the integral curve does not change when homothetic change is applied to the original variable. Hence each integral curve to the new system represent a solution in the original coordinate up to homothety. \[symmetry 2\] It is clear that the symmetry mentioned in Remark \[symmetry\] remains among pairs $(X_j,Z_j)$’s in the new system with . In addition, by the observation on $Z_j$’s derivative. It is clear that they do not change sign along the integral curve. Without loss of generality, we focus on the region where these three variables are positive. This observation provides basic estimates needed in our construction of compact invariant set (the set $P$ introduced in ). \[smooth conservation\] It is clear that $\mathcal{H}\equiv 1$ by the definition variable $X_j$. In fact, since $\mathcal{H}'=(\mathcal{H}-1)(\mathcal{G}-1)$ on $C$, the set $C\cap \{\mathcal{H}\equiv 1\}$ is flow-invariant. Furthermore, $C\cap \{\mathcal{H}\equiv 1\}$ is diffeomorphic to a level set $$dX_1^2+dX_2^2+d\left(\frac{1}{d}-X_1-X_2\right)^2-1+d\sum_j \mathcal{R}_j=0$$ in $\mathbb{R}^5$. Therefore, $C\cap \{\mathcal{H}\equiv 1\}$ is a $4$-dimensional smooth manifold by the inverse function theorem. System can be restricted to a $4$-dimensional subsystem on $C\cap \{\mathcal{H}\equiv 1\}$. \[critical points\] The complete list of critical points of system in $C\cap \{\mathcal{H}\equiv 1\}$ is the following: I. the set $\left\{(x_1,x_2,x_3,0,0,0)\mid \sum_j^3x_j^2=\frac{1}{d},\quad \sum_j^3x_j=\frac{1}{d}\right\}$; II. $\left(-\frac{1}{d},\frac{1}{d},\frac{1}{d},\pm\frac{1}{d}\sqrt{\frac{3-d}{b}},0,0\right)$ and its counterparts with pairs $(X_j,Z_j)$’s permuted. This critical point occurs only for Case I; III. $\left(\frac{1}{d},0,0,0,\pm\frac{1}{d},\pm\frac{1}{d}\right)$ and its counterparts with pairs $(X_j,Z_j)$’s permuted; IV. $\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\pm\frac{2b}{d-1}\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}},\pm\frac{2b}{d-1}\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}},\pm\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}}\right)$ and its counterparts with pairs $(X_j,Z_j)$’s permuted; V. $\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\pm\frac{1}{n}\sqrt{\frac{n-1}{a-b}},\pm\frac{1}{n}\sqrt{\frac{n-1}{a-b}},\pm\frac{1}{n}\sqrt{\frac{n-1}{a-b}}\right)$. The proof is processed by direct computations. By Remark \[symmetry 2\], we focus on critical points with non-negative $Z_j$’s. \[important criti\] Some critical points in Proposition \[critical points\] have further geometric significance. - $p_0:=\left(\frac{1}{d},0,0,0,\frac{1}{d},\frac{1}{d}\right)$\ This critical point is the initial condition under the new coordinate -, i.e., becomes $\lim\limits_{\eta\to-\infty}(X_1,X_2,X_3,Z_1,Z_2,Z_3)=p_0$. Hence we study integral curves emanating from $p_0$. In order to prove the completeness, we construct a compact invariant set in Section \[Completensesetotal\] that contains $p_0$ in its boundary and traps the integral curve initially. By Remark \[symmetry\] and Remark \[symmetry 2\], its two other counterparts $p_0'=\left(0,\frac{1}{d},0,\frac{1}{d},0,\frac{1}{d}\right)$ and $ p_0''=\left(0,0,\frac{1}{d},\frac{1}{d},\frac{1}{d},0\right)$ also have the similar geometric meaning depending on how $H$ is embedded in $G$. - $p_1:=\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\frac{1}{n}\sqrt{\frac{n-1}{a-b}},\frac{1}{n}\sqrt{\frac{n-1}{a-b}},\frac{1}{n}\sqrt{\frac{n-1}{a-b}}\right)$\ This critical point is symmetric among all $(X_j,Z_j)$’s. Note that $\frac{f_j^2}{f_k^2}(p_1)=\frac{Z_j}{Z_k}(p_1)=1$, all $f_j$’s are equal at this point. We prove in Section \[aymptoticlimit\] that $p_1$ represents an AC end for the complete Ricci-flat metric represented by the integral curve emanating from $p_0$. The conical limit is a metric cone over a suitable multiple of the normal Einstein metric on $G/K$. - $p_2:=\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\frac{2b}{d-1}\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}},\frac{2b}{d-1}\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}},\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}}\right) $\ Since $r_j(p_2)$ are all equal, this point represent an invariant Einstein metric on $G/K$ other than the one represented by $p_1$. In the following text, we call the metric the “alternative Einstein metric”. For Case I, it is a Kähler–Einstein metric. It has two other counterparts with permuted $Z_j$’s. Although we do not find any integral curve with its limit as $p_2$, we show in Section \[a digression\] that there exists an integral curve emanating from $p_2$ and tends to $p_1$, representing a singular Ricci-flat metric with a conical singularity and an AC end. The linearization $\mathcal{L}$ of vector field $V$ in is $$\label{linearization} \begin{bmatrix} \mathcal{G}-1+2dX_1^2& 2dX_1X_2& 2dX_1X_3& 2bZ_1& aZ_3-2bZ_2& aZ_2-2bZ_3\\ 2dX_1X_2&\mathcal{G}-1+2dX_2^2& 2dX_2X_3& aZ_3-2bZ_1& 2bZ_2& aZ_1-2bZ_3\\ 2dX_1X_3& 2dX_2X_3& \mathcal{G}-1+2dX_3^2& aZ_2-2bZ_1& aZ_1-2bZ_2& 2bZ_3\\ (2dX_1+1)Z_1& (2dX_2-1)Z_1& (2dX_3-1)Z_1& \mathcal{G}-\frac{\mathcal{H}}{d}+2X_1& 0& 0\\ (2dX_1-1)Z_2& (2dX_2+1)Z_2& (2dX_3-1)Z_2& 0& \mathcal{G}-\frac{\mathcal{H}}{d}+2X_2& 0\\ (2dX_1-1)Z_3& (2dX_2-1)Z_3& (2dX_3+1)Z_3& 0& 0& \mathcal{G}-\frac{\mathcal{H}}{d}+2X_3 \end{bmatrix}$$ With we can compute the dimension of the unstable subspace at $p_0$. As we are considering system on $C\cap \{\mathcal{H}\equiv 1\}$, we require each unstable eigenvector to be tangent to $C\cap \{\mathcal{H}\equiv 1\}$. The normal vector field to the hypersurfaces $C$ and $\{\mathcal{H}\equiv 1\}$ are respectively $$\label{normal vector} N_C=\begin{bmatrix} 2dX_1\\ 2dX_2\\ 2dX_3\\ adZ_2+adZ_3-2bdZ_1\\ adZ_1+adZ_3-2bdZ_2\\ adZ_2+adZ_1-2bdZ_3 \end{bmatrix}, \quad N_{\{\mathcal{H}\equiv 1\}}=\begin{bmatrix} 1\\ 1\\ 1\\ 0\\ 0\\ 0 \end{bmatrix}.$$ \[unstable dimension\] The unstable subspace of system at $p_0$, restricted on $C\cap \{\mathcal{H}\equiv 1\}$, is of dimension 2. Hence the linearization at $p_0$ is $$\label{linerization matrix <} \mathcal{L}(p_0)=\begin{bmatrix} \frac{3}{d}-1&0&0&0&\frac{a-2b}{d}&\frac{a-2b}{d}\\ 0&\frac{1}{d}-1&0&\frac{a}{d}&\frac{2b}{d}&-\frac{2b}{d}\\ 0&0&\frac{1}{d}-1&\frac{a}{d}&-\frac{2b}{d}&\frac{2b}{d}\\ 0&0&0&\frac{2}{d}&0&0\\ \frac{1}{d}&\frac{1}{d}&-\frac{1}{d}&0&0&0\\ \frac{1}{d}&-\frac{1}{d}&\frac{1}{d}&0&0&0 \end{bmatrix}.$$ Eigenvalues and corresponding eigenvectors of are $$\lambda_1=\frac{1}{d},\quad \lambda_2=\lambda_3=\frac{2}{d},\quad\lambda_4=\lambda_5=\frac{1}{d}-1,\quad \lambda_6=-1.$$ $$\label{list of eigen} v_1=\begin{bmatrix} 0\\ -1\\ 1\\ 0\\ -2\\ 2 \end{bmatrix},\quad v_2=\begin{bmatrix} 2\\ 0\\ 0\\ 0\\ 1\\ 1 \end{bmatrix},\quad v_3=\begin{bmatrix} 0\\ \frac{a}{d+1}\\ \frac{a}{d+1}\\ 1\\ 0\\ 0 \end{bmatrix},\quad v_4=\begin{bmatrix} 1-d\\ 0\\ 0\\ 0\\ 1\\ 1 \end{bmatrix},\quad v_5=\begin{bmatrix} 0\\ 1\\ 1\\ 0\\ 0\\ 0 \end{bmatrix},\quad v_6=\begin{bmatrix} 0\\ 4b\\ -4b\\ 0\\ -1\\ 1 \end{bmatrix}.$$ With Remark \[scalr of spher basic ineq\], Remark \[smooth conservation\] and , it is clear that $$T_{p_0}(C\cap \{\mathcal{H}\equiv 1\})={\mathrm{span}}\{v_1,(d+1)v_3-av_2,2v_4+(d-1)v_5,v_6\}.$$ By , an unstable subspace at $p_0$ is spanned by $v_1$ and $(d+1)v_3-av_2$. Solutions of the linearized equations at $p_0$ have the form $$\label{linearized solution<} \begin{split} \begin{bmatrix} X_1\\ X_2\\ X_3\\ Z_1\\ Z_2\\ Z_3 \end{bmatrix}&=p_0+s_0e^{\frac{2\eta}{d}}((d+1)v_4-av_3)+s_1 e^{\frac{\eta}{d}}v_1=\begin{bmatrix} \frac{1}{d}\\ 0\\ 0\\ 0\\ \frac{1}{d}\\ \frac{1}{d} \end{bmatrix}+s_0e^{\frac{2\eta}{d}}\begin{bmatrix} -2a\\ a\\ a\\ d+1\\ -a\\ -a \end{bmatrix} +s_1 e^{\frac{\eta}{d}}\begin{bmatrix} 0\\ -1\\ 1\\ 0\\ -2\\ 2 \end{bmatrix}, \end{split}$$ for some $s_0>0$ and $s_1\in\mathbb{R}$. Recall Remark \[symmetry 2\]. In order to let $Z_1$ be positive initially, the assumption $s_0>0$ is necessary. It is clear that there is a 1 to 1 correspondence between the germ of linearized solution around $p_0$ and $[s_0:s_1^2]$ in $\mathbb{RP}^2$. We fix $s_0>0$ in the following text. By Hartman–Grobman theorem, there is a 1 to 1 correspondence between each and local solution to . Hence for a fixed $s_0>0$, there is no ambiguity to use $\gamma_{s_1}$ to denote an integral curve to system on with $$\gamma_{s_1}\sim p_0+s_0e^{\frac{2\eta}{d}}((d+1)v_4-av_3)+s_1 e^{\frac{\eta}{d}}v_1$$ near $p_0$. Analysis above shows that there exists a 1-parameters family of short-time existing integral curves of system on . Since each curve corresponds to a homothetic class of Ricci-flat metrics defined on a neighborhood around singular orbit $G/H$, there exists a 1-parameters family of non-homothetic Ricci-flat metrics defined on a neighborhood around $G/H$. Recall Remark \[already\], the result is consistent with the main theorem in [@eschenburg2000initial]. \[y a subcase\] By the unstable version of Theorem 4.5 in [@coddington1955theory], from we know that $$\label{s1ish1} \frac{2h_1}{\sqrt{d}}=\lim\limits_{t\to 0}\frac{\left(\frac{\dot{f_3}}{f_3}-\frac{\dot{f_2}}{f_2}\right)\sqrt{f_2f_3}}{\sqrt{{\mathrm{tr}}(L)f_1}}=\lim_{\eta\to \infty}\frac{X_3-X_2}{\sqrt{Z_1}}=\dfrac{2s_1}{\sqrt{(d+1)s_0}}.$$ Hence the parameter $s_1$ vanishes if and only if $h_1$ does. The solution with $s_1=0$ corresponds to the subsystem of where $(X_2,Z_2)\equiv (X_3,Z_3)$ is imposed, which corresponds to the subsystem of the original system where $f_2\equiv f_3$ is imposed. The reduced system is essentially the same as the one for the case where the isotropy representation has two inequivalent irreducible summands. For Case I, $\gamma_{0}$ represents the smooth complete $G_2$ metric in [@bryant1989construction][@gibbons1990Einstein]. For Case II and Case III, Ricci-flat metrics with $s_1=0$ are proved to be complete in [@bohm_inhomogeneous_1998][@wink2017cohomogeneity]. Our construction does not assume the vanishing of $s_1$. By the symmetry of the ODE system, we mainly focus on the situation where $s_1\geq 0$ without loss of generality. Suppose an integral curve $\gamma_{s_1}$ is defined on $\mathbb{R}$, then by Lemma 5.1 in [@buzano2015family], functions $f_j(t)$’s are defined on $[0,\infty)$. Therefore, Theorem \[main 1\] is proved once $\gamma_{s_1}$ is shown to be defined on $\mathbb{R}$. Completeness {#Completensesetotal} ============ With smooth extension of metrics represented by $\gamma_{s_1}$ proved, the next step is to show that $\gamma_{s_1}$ is defined on $\mathbb{R}$ so that the Ricci-flat metric it represents is complete. Our construction is divided into two parts. The first part is to find an appropriate compact invariant set $\hat{S}_3$ with $p_0$ sitting on its boundary. Although $p_0$ is in the boundary of $\hat{S}_3$, integral curves are not trapped in the set initially unless $s_1=0$. In the second step, we construct another compact set that serves as an entrance zone. It traps $\gamma_{s_1}$ initially as long as $s_1$ is close enough to zero. Moreover, integral curves trapped in this set cannot escape through some part of its boundary and they are forced to enter $•\hat{S}_3$. Hence such a $\gamma_{s_1}$ must be defined on $\mathbb{R}$. Compact Invariant Set {#Completensese} --------------------- We describe the first step in this section. There is a subtle difference between the compact invariant set for Case I and ones for Case II and III. We first construct the set for Case II and III since it is simpler. Let $\rho=\sqrt{\frac{a+2b}{2}}$. It is clear that $\rho\geq 1$ and equality holds exactly in Case I. Define $$\label{set S3's friends} \begin{split} &P=\left\{Z_1,Z_2,Z_3\geq 0\right\}\\ &\tilde{S_3}=\bigcap_{j=1}^2 \left\{Z_3-Z_j\geq 0,\quad X_3-X_j+\rho(Z_3-Z_j)\geq 0,\quad X_3\geq 0\right\}. \end{split}$$ And define $$\label{Set S3} S_3=C\cap \{\mathcal{H}\equiv 1\}\cap P\cap \tilde{S_3}.$$ Before doing further analysis on $S_3$, we give some explanations as to why it is constructed in this way. Note that the positivity of $Z_j$’s are immediate by Remark \[symmetry 2\]. The first inequality in $\tilde{S_3}$ is to require $Z_3$ to be the largest variable among $Z_j$’s. Equivalently, it requires $f_3$ to be the largest among $f_j$’s in the original coordinate. This condition is indicated by the subscript of $\tilde{S}_3$ and $S_3$. A direct consequence of this assumption is that we can assume $X_3\geq 0$ along $\gamma_{s_1}$ as shown in . It is easy to check that $p_0\in S_3$ hence the set is nonempty. Each inequality in defines a closed subset in $\mathbb{R}^7$ whose boundary is defined by the equality. Therefore, a point $x\in \partial S_3$ if there exists at least one defining inequality in reaches equality at $x$. For Case II and III, functions $$\label{where p_0 locate} X_3, Z_1, Z_3-Z_2, X_3-X_2+\rho(Z_3-Z_2)$$ among those in vanish at $p_0$. The point is hence in $\partial S_3$. Substitute to functions in . It is clear that $\gamma_{s_1}$ is trapped in $S_3$ initially if $s_1\geq 0$. By Remark \[y a subcase\], we know that $\gamma_0$ is trapped in $\partial S_3$ with $(X_2,Z_2)\equiv (X_3,Z_3)$. \[sharp\] In the set $S_3\cap\{2bZ_3-a(Z_1+Z_2)\leq 0\}$, we have estimate $$\label{sharp Z_1+Z_2} Z_1+Z_2\leq 2\sqrt{\frac{n-1}{n^2(a-b)}}.$$ By the conservation law , it follows that $$\label{1} 0\geq \frac{1}{n}-1+da(Z_2Z_3+Z_1Z_3+Z_1Z_2)-db(Z_1^2+Z_2^2+Z_3^2).$$ Note that the RHS of is symmetric between $Z_1$ and $Z_2$. It is convenient to find the maximum of $Z_1+Z_2$ on $S_3\cap \{Z_2\geq Z_1\}$ first. By the symmetry between $Z_1$ and $Z_2$ in , such a maximum is the maximum of $Z_1+Z_2$ in $S_3$. With the assumption $Z_2\geq Z_1$, we write $Z_1=\nu Z_2$ for some $\nu\in [0,1]$. Fix such a $\nu$. Then becomes $$\label{2} \begin{split} 0&\geq \frac{1}{n}-1+da(Z_2Z_3+\nu Z_2Z_3+\nu Z_2^2)-db(\nu^2 Z_2^2+Z_2^2+Z_3^2)\\ &= \frac{1}{n}-1+d(-bZ_3^2+a(1+\nu)Z_2Z_3+(a\nu -b(1+\nu^2))Z_2^2). \end{split}$$ Define $\mathcal{F}(Z_3)=-bZ_3^2+a(1+\nu)Z_2Z_3+(a\nu-b(1+\nu^2))Z_2^2.$ Consider the set $S_3\cap\{2bZ_3-a(Z_1+Z_2)\leq 0\}\cap \{Z_1=\nu Z_2\}$, we have $$Z_2\leq Z_3\leq \frac{a}{2b}(1+\nu)Z_2.$$ Hence for any fixed $\nu$ and $Z_2$, the minimum of $\mathcal{F}$ in $S_3\cap\{2bZ_3-a(Z_1+Z_2)\leq 0\}\cap \{Z_1=\nu Z_2\}$ is reached at $Z_3=Z_2$. Therefore, computation continues as $$\label{3} \begin{split} 0&\geq \frac{1}{n}-1+d\left(-b+a(1+\nu)+(\nu a-b(1+\nu^2))\right)Z_2^2\\&=\frac{1}{n}-1+d\left(-b\nu^2+2a\nu +a-2b\right)Z_2^2. \end{split}$$ The coefficient of $Z_2^2$ in can be easily checked to be positive. It follows that $$\begin{split} \label{4} (Z_1+Z_2)^2=(1+\nu)^2Z_2^2&\leq \left(1-\frac{1}{n}\right)\dfrac{(1+\nu)^2}{d(-b\nu^2+2a\nu +a-2b)}\\ &=\left(1-\frac{1}{n}\right)\dfrac{1}{d\left(-b+2(a+b)\frac{1}{1+\nu}-(a+3b)\frac{1}{(1+\nu)^2}\right)}. \end{split}$$ Consider function $h\left(\frac{1}{1+\nu}\right)=-(a+3b)\frac{1}{(1+\nu)^2}+2(a+b)\frac{1}{1+\nu}-b$. Since by Remark \[scalr of spher basic ineq\], we have $ \frac{1}{2}\leq \frac{a+b}{a+3b}\leq 1, $ the minimum of $h$ is either $h\left(\frac{1}{2}\right)$ or $h(1)$. Computation shows $h\left(\frac{1}{2}\right)<h(1)$. We conclude that $ (Z_1+Z_2)^2\leq \left(1-\frac{1}{n}\right)\frac{1}{d}\frac{1}{h\left(\frac{1}{2}\right)}=4\frac{n-1}{n^2(a-b)}. $ Hence the proof is complete. Note that the equality in is reached by $p_1$. \[S\_3\] For Case II and III, integral curves $\gamma_{s_1}$ to system on $C_0\cap \{\mathcal{H}\equiv 1\}$ emanating from $p_0$ with $s_1\geq 0$ do not escape $S_3$. Two perspectives can be taken in the following computations that frequently appear through out this article. First is to view algebraic expressions in as functions along $\gamma_{s_1}$ and they all vanish at $p_0$. Integral curves emanating from $p_0$ being trapped in $S_3$ initially is equivalent to these defining functions being positive near $p_0$. To show that $\gamma_{s_1}$ does not escape $S_3$ is to show the non-negativity of these functions along the integral curves. Suppose one of these functions vanishes at some point along the integral curves for the first time. We want to show that its derivative at that point is non-negative. The second perspective is to consider $\partial S_3$ as a union of subsets of a collection of linear and quadratic varieties. Require the restriction of the vector field $V$ in on each of these subsets to point inward $S_3$. If such a requirement is met, then it is impossible for the integral curves to escape if they are initially in $S_3$. Both perspectives lead to the same computation of inner product between $V$ and the gradient of each defining function in . Then require the inner product to be non-negative if the gradient points inward $S_3$. It might not be true that the inner product is non-negative on each variety globally. But all we need is the non-negativity on its subsets that $\partial S_3$ consists of. By definition of $S_3$, we automatically have $$\label{ricci3} \begin{split} \mathcal{R}_3&=aZ_1Z_2+b(Z_3^2-Z_1^2-Z_2^2)=\left\{\begin{array}{cc} Z_2(aZ_1-bZ_2)+b(Z_3^2-Z_1^2)\geq 0& \text{if $Z_1\geq Z_2$}\\ Z_1(aZ_2-bZ_1)+b(Z_3^2-Z_2^2)\geq 0& \text{if $Z_2\geq Z_1$} \end{array}\right.. \end{split}$$ On $X_3=0$, we have $ \left. \langle\nabla(X_3),V\rangle \right\|_{X_3=0}=\mathcal{R}_3\geq 0 $ by . Hence $X_3$ is non-negative along every $\gamma_{s_1}$ that is trapped in $S_3$ initially. Next we need to show that the integral curves cannot escape from the part of $\partial \tilde{S_3}$ that is in $\partial S_3$. For distinct $j,k\in\{1,2\}$, it follows that $$\begin{split} \left. \langle\nabla(Z_3-Z_j),V\rangle \right\|_{Z_3-Z_j=0}&=Z_3\left(\mathcal{G}-\frac{1}{d}+2X_3\right)-Z_j\left(\mathcal{G}-\frac{1}{d}+2X_j\right)\\ &=2Z_3(X_3-X_j) \quad \text{since $Z_3-Z_j=0$}\\ &\geq 2\rho Z_3(Z_j-Z_3) \quad \text{by definition of $S_3$}\\ &=0 \quad \text{since $Z_3-Z_j=0$}. \end{split}$$ Although it is not clear if $X_3-X_j\geq 0$ along $\gamma_{s_1}$, we impose a weaker condition, which is the second inequality in $\tilde{S_3}$. What it means is to allow $Z_3-Z_j$ to decrease, yet the rate of its decreasing cannot be too steep so that $Z_3-Z_j$ increases before it could decrease to zero. Fortunately, the weaker condition does hold along the integral curves. $$\label{X3-X_1+k(Z_3-Z_1)pre} \begin{split} &\left. \langle\nabla(X_3-X_j+\rho(Z_3-Z_j)),V\rangle \right\|_{X_3-X_j+\rho(Z_3-Z_j)=0}\\ &=(X_3-X_j+\rho(Z_3-Z_j))\left(\mathcal{G}-1\right)+\mathcal{R}_3-\mathcal{R}_j+\rho Z_3\left(1-\frac{1}{d}+2X_3\right)-\rho Z_j\left(1-\frac{1}{d}+2X_j\right)\\ &= (Z_3-Z_j)\left(2b(Z_3+Z_j)-aZ_k+\rho\left(1-\frac{1}{d}\right)+2\rho X_3-2\rho^2Z_j\right) \quad \text{since $X_j=X_3+\rho(Z_3-Z_j)$}\\ &\geq (Z_3-Z_j)\left(2bZ_3-a(Z_j+Z_k)+\rho\left(1-\frac{1}{d}\right)\right)\quad \text{since $X_3\geq 0$ in $S_3$}\\ &= (Z_3-Z_j)\left(2bZ_3-a(Z_1+Z_2)+\rho\left(1-\frac{1}{d}\right)\right) \end{split}.$$ If $2bZ_3-a(Z_1+Z_2)\geq 0$, then the last line of computation above is obviously non-negative. If $2bZ_3-a(Z_1+Z_2)\leq 0$, then continues as $$\label{X3-X_1+k(Z_3-Z_1)} \geq (Z_3-Z_j)\left((b-a)(Z_1+Z_2)+\rho\left(1-\frac{1}{d}\right)\right)$$ since $Z_3\geq \frac{Z_1+Z_2}{2}$ in $S_3$. Apply Proposition \[sharp\], we know that is non-negative if $$\label{i=1 fail} \dfrac{\rho(d-1)}{d(a-b)}\geq 2\sqrt{\frac{n-1}{n^2(a-b)}}.$$ Straightforward computations show that $$ Case $\rho$ $\dfrac{\rho(d-1)}{d(a-b)}$ $2\sqrt{\frac{n-1}{n^2(a-b)}}$ ------ ----------------------- --------------------------------------- -------------------------------------- I 1 $\frac{2}{5}$ $\frac{2}{3}$ II $\sqrt{\frac{5}{2}}$ $\frac{3\sqrt{10}}{28}\approx 0.339$ $\frac{\sqrt{154}}{42}\approx 0.295$ III $\sqrt{\frac{11}{2}}$ $\frac{7\sqrt{22}}{128}\approx 0.257$ $\frac{\sqrt{46}}{48}\approx 0.141$ . $$ Inequality holds only for Case II and III. Hence for Case II and III, integral curves $\gamma_{s_1}$ emanating from $p_0$ does not escape $S_3$ if $s_1\geq 0$. Although estimate is sharp in $S_3$, inequality has room to be improved as we dropped a non-negative term $2\rho X_3$ in the computation. It turns out can be proved to be non-negative for Case I with an additional inequality, as demonstrated in Proposition \[T\_3\]. We move on to Case I. Recall that the construction in Proposition \[S\_3\] is not successful just because inequality does not hold in this case. To fix this issue, an additional inequality is needed. Define $$\label{G2condition} F_j:=X_k+X_l-Z_j.$$ Computations show $$\begin{split} &\langle \nabla F_j,V\rangle=F_j\left(\mathcal{G}-1\right)+\dfrac{3Z_j}{2}\left(\dfrac{1}{3}F_j-F_k-F_l\right). \end{split}$$ \[G2\] The condition $F_1\equiv F_2\equiv F_3\equiv 0$ is in fact the $G_2$ condition on cohomogeneity one manifold with principal orbit $SU(3)/T^2$. Hence $\cap_{j=1}^3\{F_j\equiv 0\}$ is flow-invariant and it contains the integral curve $\gamma_0$ that represents the complete smooth $G_2$ metric on $M$, which is firstly discovered in [@bryant1989construction][@gibbons1990Einstein]. In the following text, we still use $\tilde{S_3}$ and $S_3$ to denote invariant sets constructed. If necessary, we use the phrase such as “$S_3$ for Case I” to refer to the case in particular. Define $$\label{set T3's friends} \begin{split} &\tilde{S_3}=\bigcap_{j=1}^2 \left\{Z_3-Z_j\geq 0,\quad F_j-F_3\geq 0, \quad X_3\geq 0\right\}\cap \{3F_1+3F_2-F_3\geq 0\}. \end{split}$$ And define $$\label{Set T3} S_3=C\cap \{\mathcal{H}\equiv 1\}\cap P\cap \tilde{S_3}.$$ Note that $F_j-F_3\geq 0$ is simply the second defining inequality in the $\tilde{S_3}$ in with $\rho =1$. It is easy to check that $p_0\in S_3$ hence $S_3$ is nonempty. Since functions $X_3, Z_1, Z_3-Z_2, F_j-F_3$ and $3F_1+3F_2-F_3$ vanish at $p_0$ among those in , the point is in $\partial S_3$. With the same argument as the one for Case II and III, we know that $\gamma_{s_1}$ is trapped in $S_3$ initially if $s_1\geq 0$. \[T\_3\] Integral curves $\gamma_{s_1}$ to system on $C\cap \{\mathcal{H}\equiv 1\}$ emanating from $p_0$ with $s_1\geq 0$ do not escape $S_3$. The idea of proving Proposition \[T\_3\] is the same as the one of Proposition \[S\_3\]. Besides, almost all computations for Proposition \[S\_3\] still hold except the one for $F_j-F_3\geq 0$ since is not true for Case I. With the additional inequality, it follows that $$\begin{split} &\left.\langle\nabla (F_j-F_3),V\rangle\right\|_{F_j-F_3=0}\\ &=(F_j-F_3)\left(\mathcal{G}-1\right)+\dfrac{3Z_j}{2}\left(\dfrac{1}{3}F_j-F_k-F_3\right)-\dfrac{3Z_3}{2}\left(\dfrac{1}{3}F_3-F_j-F_k\right)\\ &=\dfrac{3Z_j}{2}\left(\dfrac{1}{3}F_j-F_k-F_3\right)-\dfrac{3Z_3}{2}\left(\dfrac{1}{3}F_3-F_j-F_k\right)\quad \text{since $F_j=F_3$}\\ &=\dfrac{3Z_j}{2}\left(\dfrac{1}{3}F_j-F_k-F_j\right)-\dfrac{3Z_3}{2}\left(\dfrac{1}{3}F_j-F_j-F_k\right) \quad \text{since $F_j=F_3$}\\ &=F_k\dfrac{3}{2}(Z_3-Z_j)+F_j(Z_3-Z_j)\\ &=\dfrac{1}{2}(Z_3-Z_j)(3F_j+3F_k-F_3) \quad \text{since $F_j=F_3$}\\ &\geq 0. \end{split}$$ Notice that we do not drop any non-negative term in the computation above like we do in . The estimate for $\left.\langle\nabla (F_j-F_3),V\rangle\right\|_{F_j-F_3=0}$ hence becomes sharper. Finally, we need to show that the additional inequality holds along the integral curves. Indeed, since $$\begin{split} &\left.\langle\nabla (3F_1+3F_2-F_3),V\rangle\right\|_{3F_1+3F_2-F_3=0}\\ &=(3F_1+3F_2-F_3)\left(\mathcal{G}-1\right)\\ &\quad +\dfrac{3Z_1}{2}\left(F_1-3F_2-3F_3\right)+\dfrac{3Z_2}{2}\left(F_2-3F_1-3F_3\right)-\dfrac{3Z_3}{2}\left(\dfrac{1}{3}F_3-F_1-F_2\right)\\ &=\dfrac{3Z_1}{2}\left(F_1-3F_2-3F_3\right)+\dfrac{3Z_2}{2}\left(F_2-3F_1-3F_3\right)\quad \text{since $3F_1+3F_2-F_3=0$}\\ &=\dfrac{3Z_1}{2}\left(4F_1-4F_3\right)+\dfrac{3Z_2}{2}\left(4F_2-4F_3\right)\quad \text{since $3F_1+3F_2-F_3=0$}\\ &\geq 0\quad \text{definition of $S_3$ for $i=1$} \end{split},$$ $3F_1+3F_2-F_3$ remains non-negative along the integral curves. Therefore, integral curves $\gamma_{s_1}$ do not escape $S_3$ in Case I if $s_1\geq 0$. One may want to integrate the additional inequality in $S_3$ for Case I to the other two cases so that all cases can be discussed by a single construction. Specifically, one can define $$F_j:=X_k+X_l-\rho Z_j.$$ Then the additional inequality analogous to $3F_1+3F_2-F_3\geq 0$ for Case II and III is $ aF_1+aF_2-2bF_3\geq 0. $ But $$\begin{split} &\left.\langle\nabla (aF_1+aF_2-2bF_3),V\rangle\right\|_{aF_1+aF_2-2bF_3=0}\\ &=\frac{aZ_1}{k}(a+2b)(F_1-F_3)+\frac{aZ_2}{k}(a+2b)(F_2-F_3)+\dfrac{\zeta}{k}(aZ_1+aZ_2-2bZ_3) \end{split},$$ where $\zeta=\frac{(3-d)a-(2+2d)b}{2d}\leq 0$. It only vanishes in Case I. Hence whether $aF_1+aF_2-2bF_3$ is non-negative along the integral curves in $S_3$ is not clear. The analogous $F_j$ defined for Case II and Case III may not have too much meaning after all because there is no special holonomy for odd dimension other than $7$. We are ready to construct the compact invariant set mentioned at the beginning of this section. Define $$\hat{S}_3=S_3\cap \{Z_1+Z_2-Z_3\geq 0\}\cap\{Z_1(X_1-X_3)+Z_2(X_2-X_3)\geq 0\}$$ for all three cases. We have the following lemma. \[never escape\] $\hat{S}_3$ is a compact invariant set. Because $Z_1+Z_2-Z_3\geq 0$ in $\hat{S}_3$, we can apply Proposition \[sharp\] so that $Z_1+Z_2$ is bounded above. Then all $Z_j$’s are bounded in $\hat{S}_3$. By conservation law , we immediately conclude that all variables are bounded. The compactness of $\hat{S}_3$ is hence proved. To check that $\hat{S_3}$ is flow invariant, consider the hyperplane $Z_1+Z_2-Z_3=0$. It follows that $$\begin{split} \left. \langle\nabla(Z_1+Z_2-Z_3),V\rangle \right\|_{Z_1+Z_2-Z_3=0}&=(Z_1+Z_2-Z_3)\left(\mathcal{G}-\frac{1}{d}\right)+2Z_1X_1+2Z_2X_2-2Z_3X_3\\ &=2Z_1(X_1-X_3)+2Z_2(X_2-X_3) \quad \text{since $Z_1+Z_2-Z_3=0$}\\ &\geq 0 \quad \text{definition of $\hat{S}_3$} \end{split}.$$ On hypersurface $Z_1(X_1-X_3)+Z_2(X_2-X_3)=0$, we have $$\label{messy!} \begin{split} &\left. \langle\nabla(Z_1(X_1-X_3)+Z_2(X_2-X_3)),V\rangle \right\|_{Z_1(X_1-X_3)+Z_2(X_2-X_3)=0}\\ &=\left. \left\langle\nabla \left(Z_3\left(\frac{Z_1}{Z_3}(X_1-X_3)+\frac{Z_2}{Z_3}(X_2-X_3)\right)\right),V\right\rangle \right\|_{Z_1(X_1-X_3)+Z_2(X_2-X_3)=0} \\ &= Z_3\left(\mathcal{G}-\frac{1}{d}+2X_3\right)\left(\frac{Z_1}{Z_3}(X_1-X_3)+\frac{Z_2}{Z_3}(X_2-X_3)\right)+Z_3\left(2\dfrac{Z_1}{Z_3}(X_1-X_3)^2+2\dfrac{Z_2}{Z_3}(X_2-X_3)^2\right)\\ &\quad+Z_1\left((X_1-X_3)\left(\mathcal{G}-1\right)+\mathcal{R}_1-\mathcal{R}_3\right)+Z_2\left((X_2-X_3)\left(\mathcal{G}-1\right)+\mathcal{R}_2-\mathcal{R}_3\right)\\ &=2Z_1(X_1-X_3)^2+2Z_2(X_2-X_3)^2+Z_1(\mathcal{R}_1-\mathcal{R}_3)+Z_2(\mathcal{R}_2-\mathcal{R}_3)\\ &\quad \text{ since $Z_1(X_1-X_3)+Z_2(X_2-X_3)=0$}\\ &\geq Z_1(\mathcal{R}_1-\mathcal{R}_3)+Z_2(\mathcal{R}_2-\mathcal{R}_3)\\ &=Z_1(Z_3-Z_1)(aZ_2-2b(Z_3+Z_1))+Z_2(Z_3-Z_2)(aZ_1-2b(Z_3+Z_2)) \end{split}$$ For distinct $j,k\in\{1,2\}$, take $A_j=Z_j(Z_3-Z_j)$ and $B_j=aZ_k-2b(Z_3+Z_j)$. Apply identity $$A_1B_1+A_2B_2=\frac{1}{2}((A_1+A_2)(B_1+B_2)+(A_1-A_2)(B_1-B_2)).$$ Then the computation continues as $$\label{usefulfor p3} \begin{split} &= \frac{1}{2}(Z_1(Z_3-Z_1)+Z_2(Z_3-Z_2))((a-2b)(Z_1+Z_2)-4bZ_3)+\frac{1}{2}(Z_1-Z_2)^2(Z_1+Z_2-Z_3)(a+2b)\\ &\geq \frac{1}{2}(Z_1(Z_3-Z_1)+Z_2(Z_3-Z_2))(a-6b)Z_3 \quad \text{since $Z_1+Z_2\geq Z_3$}\\ &\geq 0 \quad \text{Remark \ref{scalr of spher basic ineq}} \end{split}.$$ Therefore, $\hat{S}_3$ is flow-invariant. \[S2\] By the symmetry between $(X_2,Z_2)$ and $(X_3,Z_3)$, constructions of $S_3$ and $\hat{S}_3$ above can be carried over to defining $S_2$ and $\hat{S}_2$. With the same arguments, it can be shown that $\gamma_{s_1}$ does not escape $S_2$ whenever $s_1\leq 0$ and $\hat{S}_2$ is a compact invariant set. \[Mcontainy\] It is clear that $p_0\in\partial \hat{S}_3$. One can check that $\gamma_{0}$ is trapped in $\hat{S}_3$ initially. Hence the long time existence for $\gamma_0$ is proved. By Remark \[y a subcase\], it is trapped in $\hat{S}_3\cap\{X_2\equiv X_3, Z_2\equiv Z_3\}$. Hence $\hat{S}_3$ can be used to prove the long time existence for the special case where $(X_2,Z_2)\equiv (X_3,Z_3)$ is imposed. In fact, the compact invariant set for cohomogeneity one manifolds of two summands can be constructed by a little modification on $\hat{S}_3\cap \{X_2\equiv X_3, Z_2\equiv Z_3\}$, reproducing the same result in [@bohm_inhomogeneous_1998][@wink2017cohomogeneity]. For Case I in particular, $\gamma_0$ represents the complete $G_2$ metric discovered in in [@bryant1989construction][@gibbons1990Einstein]. \[positivericci\] Not only $\mathcal{R}_3$ is non-negative in $\hat{S}_3$. This is in fact the case for all $\mathcal{R}_j$’s. For distinct $j,k\in\{1,2\}$, we have $$\begin{split} \mathcal{R}_j=aZ_3Z_k+b(Z_j^2-Z_k^2-Z_3^2)&\geq aZ_3Z_k+b(Z_j^2+Z_k^2-(Z_j+Z_k)^2)\quad \text{by definition of $\hat{S}_3$}\\ &=aZ_3Z_k-2bZ_jZ_k\\ &\geq (a-2b)Z_jZ_k \quad \text{by definition of $\hat{S}_3$}\\ &\geq 0 \end{split}.$$ Therefore, one geometric feature of complete Ricci-flat metrics represented by $\gamma_0$ is that hypersurface has positive Ricci tensor for all $t\in(0,\infty)$. As discussed in Remark \[posiscaler\], Ricci-flat metrics represented by $\gamma_{s_1}$ with $s_1\neq 0$ does not hold such a property. Although $\gamma_{s_1}$ is trapped in $S_3$ if $s_1\geq 0$, functions $Z_1+Z_2-Z_3$ and $Z_1(X_1-X_3)+Z_2(X_2-X_3)$ are negative initially if $s_1> 0$. Hence $\gamma_{s_1}$ is not trapped in $\hat{S}_3$ initially if $s_1> 0$. To include the case where $s_1> 0$, we need to enlarge $\hat{S}_3$ a little bit so that it initially traps all $\gamma_{s_1}$ with $s_1$ close enough to zero. That leads us to the second step of our construction. Entrance Zone {#Entrance Zone} ------------- In this section, we assume $s_1>0$ and work with the set $S_3$. We construct an entrance zone that forces $\gamma_{s_1}$ to enter $\hat{S}_3$ eventually. Our goal is to show that for all small enough $s_1>0$, $\gamma_{s_1}$ will enter $\hat{S}_3$ in a compact set. As shown in computation , it is more convenient to compute with variables $ \omega_1=\frac{Z_1}{Z_3}$ and $\omega_2=\frac{Z_2}{Z_3}, $ whose respective derivatives are $ \omega_1'=2\omega_1(X_1-X_3)$ and $\omega_2'=2\omega_2(X_2-X_3). $ By the definition of $S_3$, we have $Z_3\geq Z_1,Z_2$. Therefore $\omega_1,\omega_2\in[0,1]$. For another point of view, we can also consider the problem on $\omega_1\omega_2$-plane as shown Figure \[Fig1\]. Whatever $\gamma_{s_1}$ looks like, we can always project its $Z_1$ and $Z_2$ coordinate to $\omega_1 \omega_2$-plane. And we want to prove the projection is bounded away from $(0,0)$ and hopefully going through the line $$l_0\colon \omega_1+\omega_2-1=0,$$ which is the projection of hyperplane $Z_1+Z_2-Z_3=0.$ Note that any homogeneous variety in $Z_j$’s of degree $D$ can be projected to an algebraic curve on $\omega_1\omega_2$-plane by dividing by $Z_3^D$. Before the construction, we establish the following basic fact. ![Projection to $\omega_1\omega_2$-plane[]{data-label="Fig1"}](Projection1.jpg){width="2.5in"} \[at least to Z\_1=Z\_2\] $\frac{Z_1}{Z_2}$ is strictly increasing along $\gamma_{s_1}$ as long as $Z_2> Z_1$. Initially we have $(X_1-X_2)(p_0)=\frac{1}{d}$. If $Z_2>Z_1$, we have $$\begin{split} \left.\left(X_1-X_2\right)'\right\|_{X_1-X_2=0}&=(X_1-X_2)(\mathcal{G}-1)+\mathcal{R}_1-\mathcal{R}_2\\ &=(Z_2-Z_1)(aZ_3-2bZ_1-2bZ_2)\\ &\geq (Z_2-Z_1)(a-4b)Z_2 \quad \text{since $Z_3\geq Z_2>Z_1$}\\ &> 0 \quad \text{Remark \ref{scalr of spher basic ineq}} \end{split}.$$ Hence $X_1-X_2>0$ along $\gamma_{s_1}$ when $Z_2>Z_1$. But then $$\begin{split} \left(\frac{Z_1}{Z_2}\right)'&=2\frac{Z_1}{Z_2}(X_1-X_2)> 0 \end{split}$$ when $Z_2>Z_1$. Therefore $\frac{Z_1}{Z_2}$ is strictly increasing along $\gamma_{s_1}$ as long as $Z_2>Z_1$. Substitute solution of linearized equation to $\mathcal{R}_1-\mathcal{R}_3$ and $\mathcal{R}_3-\mathcal{R}_2$. It is clear that they are positive initially. Hence at the beginning, the integral curve is trapped in $$\label{U_0} U_0=S_3\cap \{Z_1+Z_2-Z_3\leq 0, \mathcal{R}_1-\mathcal{R}_3\geq 0, \mathcal{R}_3-\mathcal{R}_2\geq 0\},$$ whose projection on $\omega_1\omega_2$-plane for all three cases is illustrated in Figure \[Fig 2\]. [.5]{} ![Projection of $U_0$ (enclosed by bold line segments) on $\omega_1\omega_2$-plane for all three cases[]{data-label="Fig 2"}](Projection2.jpg "fig:"){width="1\linewidth"} [.5]{} ![Projection of $U_0$ (enclosed by bold line segments) on $\omega_1\omega_2$-plane for all three cases[]{data-label="Fig 2"}](Projection2-1.jpg "fig:"){width="1\linewidth"} [.5]{} ![Projection of $U_0$ (enclosed by bold line segments) on $\omega_1\omega_2$-plane for all three cases[]{data-label="Fig 2"}](Projection2-2.jpg "fig:"){width="1\linewidth"} By Proposition \[at least to Z\_1=Z\_2\], we know that in principal, the projection of $\gamma_{s_1}$ on $\omega_1\omega_2$-plane can get arbitrarily closed to $\omega_1-\omega_2=0$, represented the dashed lines in Figure \[Fig 2\]. Therefore, an integral curve that is initially trapped in $U_0$ has to escape. The question is whether it will escape $U_0$ through $Z_1+Z_2-Z_3=0$, represented by the red line segment. It turns out that a subset of $U_0$ can be constructed in a way that it contains a part of $Z_1+Z_2-Z_3=0$ and $\gamma_{s_1}$ has to escape that subset through $Z_1+Z_2-Z_3=0$. Specifically, the construction is based on the following three ideas. 1\. Since $Z_1+Z_2-Z_3\leq 0$ initially along $\gamma_{s_1}$, the main task is to bound $Z_3$ from above. For computation conveniences, we prefer to bound $Z_3$ from above by some homogeneous algebraic varieties in $Z_j$’s. In other words, defining inequalities of the entrance zone should include $Z_1+Z_2-Z_3\leq 0$ and $B(Z_1,Z_2,Z_3)\geq 0$ for some homogeneous polynomial $B$ in $Z_j$’s. 2\. In order to show that $\gamma_{s_1}$ does not escape through $B=0$, we need to show that $\left.\langle \nabla(B),V\rangle\right\|_{B=0}$ is non-negative along $\gamma_{s_1}$ in the entrance zone. This idea is discussed in the proof of Proposition \[Set S3\]. It might be difficult to determine the sign of $\left.\langle \nabla(B),V\rangle\right\|_{B=0}$ even we are allowed to mod out $B=0$ in the computation result. But notice that $B':=\left.\langle \nabla(P),V\rangle\right\|_{B=0}=0$ vanishes at $p_0$, and inequality $B'\geq 0$ can potentially be added to the definition of the entrance zone. 3\. If we want to impose $B'\geq 0$, the trade-off is to show that $\left.\langle \nabla(B'),V\rangle\right\|_{B'=0}\geq 0$ along $\gamma_{s_1}$ in the entrance zone. The homogeneous polynomial $B$ that we find consists of two parameters. They allow us to tune the entrance zone to satisfy some technical inequalities. Once these inequalities are satisfied, we can show that $\left.\langle \nabla(B'),V\rangle\right\|_{B'=0}\geq 0$ in the entrance zone and $\gamma_{s_1}$ is forced to escape through $Z_1+Z_2-Z_3=0$. We first proceed the construction by having those technical inequalities in part 3 ready. In this process, the first parameter for $B$ is introduced and how they interact with these technical inequalities are explained. Then we reveal the definition for $B$ and its last parameter. \[X\_3+X\_2>0\] In $S_3$, $X_2+X_3> 0$ along $\gamma_{s_1}$ always. It is clear that $X_2+X_3$ is positive initially along the curves. Since $$\begin{split} \left. \langle\nabla(X_2+X_3),V\rangle \right\|_{X_2+X_3=0}&=(X_2+X_3)(\mathcal{G}-1)+\mathcal{R}_2+\mathcal{R}_3\\ &=\mathcal{R}_2+\mathcal{R}_3\quad \text{since $X_2+X_3=0$}\\ &\geq Z_1(a(Z_2+Z_3)-2bZ_1)\\ &> 0 \quad \text{since $Z_3\geq Z_1$ and $a-2b=d-1>0$} \end{split},$$ $X_2+X_3$ stays positive along $\gamma_{s_1}$. \[compare X\_3 with X\_2\] For any fixed $\delta\geq 0$, $X_3- (1+\delta)X_2>0$ initially along $\gamma_{s_1}$ and stay positive in the region where $\mathcal{R}_3-(1+\delta) \mathcal{R}_2\geq 0$. Substitute solution of linearized equation to $X_3-(1+\delta)X_2$. We have $$(2+\delta)s_1e^{\frac{\eta}{d}}-a\delta s_0e^\frac{2\eta}{d} \sim (2+\delta)s_1e^{\frac{\eta}{d}}>0$$ near $p_0$. Since $$\label{comp for X_3and X_2} \begin{split} \left. \langle\nabla(X_3-(1+\delta X_2)),V\rangle \right\|_{X_3-(1+\delta) X_2=0} &=(X_3-(1+\delta) X_2)\left(\mathcal{G}-1\right)+\mathcal{R}_3-(1+\delta)\mathcal{R}_2\\ &=\mathcal{R}_3-(1+\delta)\mathcal{R}_2\quad \text{since $X_3-(1+\delta) X_2=0$}, \end{split}$$ the proof is complete. Define $$\label{Udelta} U_\delta=U_0\cap \{\mathcal{R}_3-(1+\delta)\mathcal{R}_2\geq 0\}.$$ It is easy to check that $U_\delta$ is a subset of $U_0$ and $\gamma_{s_1}$ is initially trapped in $U_{\delta}$ if $s_1>0$. Therefore, $X_3-(1+\delta)X_2>0$ when $\gamma_{s_1}$ is in $U_\delta$ by Proposition \[compare X\_3 with X\_2\]. The fixed value of $\delta$ needs to be picked in a certain range for the following two technical reasons. Firstly, we want inequality $X_3-(1+\delta)X_2> 0$ to hold at least until $\gamma_{s_1}$ enters $\hat{S}_3$. Hence by proposition \[compare X\_3 with X\_2\], we need to pick $\delta$ that make $\mathcal{R}_3-(1+\delta)\mathcal{R}_2\geq 0$ contains a subset of $Z_1+Z_2-Z_3=0$. Secondly, because $U_0\subset S_3\cap \{Z_2-Z_1>0\}$ and the behavior of $\gamma_{s_1}$ is better known in $U_0$, we want $\gamma_{s_1}$ passes though the part of $Z_1+Z_2-Z_3=0$ that $Z_2-Z_1\geq 0$ is satisfied. In summary, we have the following proposition. \[R3-R2\] If $\delta\in \left(\frac{6b-a}{2(d-1)},\frac{4b}{d-1}\right)$, then $\{\mathcal{R}_3-(1+\delta)\mathcal{R}_2\geq 0\}$ contains a subset of $\{Z_1+Z_2-Z_3=0\}\cap \{Z_2-Z_1>0\}$ in $U_0$. If $\delta\in\left(\frac{6b-a}{2(d-1)},\frac{4b}{d-1}\right)$, then we have $\frac{4b-(d-1)\delta}{(d-1)(1+\delta)}\in (0,1)$. Suppose $\frac{4b-(d-1)\delta}{(d-1)(1+\delta)}\geq \frac{Z_1}{Z_2}$, then $$\label{R3R2} \begin{split} &\left.(\mathcal{R}_3-(1+\delta)\mathcal{R}_2)\right\|_{Z_1+Z_2-Z_3=0}\\ &=(Z_3-Z_2)(2b(Z_3+Z_2)-aZ_1)-\delta(aZ_1Z_3+b(Z_2^2-Z_1^2-Z_3^2))\quad \text{since }Z_1+Z_2-Z_3=0\\ &=Z_1(2b(2Z_2+Z_1)-aZ_1)-\delta(aZ_1Z_2+aZ_1^2-2bZ_1^2-2bZ_1Z_2)\quad \text{since }Z_1+Z_2-Z_3=0\\ &=Z_1(-(d-1)(1+\delta)Z_1+(4b-(d-1)\delta)Z_2)\quad \text{Remark \ref{scalr of spher basic ineq}}\\ &\geq 0 \end{split}.$$ The proof is complete. Perhaps a better way to illustrate Proposition \[R3-R2\] is to consider the projection on the $\omega_1\omega_2$-plane. For $\mathcal{R}_3-(1+\delta)\mathcal{R}_2=0$, we obtain an algebraic curve $$l_1\colon (1-\omega_2)(2b(1+\omega_2)-a\omega_1)-\delta(a\omega_1+b(\omega_2^2-\omega_1^2-1))=0.$$ ![$\delta=0.7$ for Case I[]{data-label="Fig 3"}](Projection3.jpg){width="2.6in"} Straightforward computation shows that $l_1$ intersect with $\omega_1+\omega_2=1$ at points $\left(0,1\right)$ and $\left(\frac{4b-\delta(d-1)}{a+2b},\frac{(d-1)(1+\delta)}{a+2b}\right)$. If $\delta \in \left(\frac{6b-a}{d-1},\frac{4b}{d-1}\right)$, then the second intersection point $\left(\frac{4b-\delta(d-1)}{a+2b},\frac{(d-1)(1+\delta)}{a+2b}\right)$ is in the region where $\omega_2-\omega_1> 0$. Hence $U_\delta$, denoted by the darker area in Figure \[Fig 3\], can include a segment of $l_0$ in $U_0$, represented by the bold segment, that is away from $\omega_1-\omega_2=0$. Note that Case I is the only case where the admissible $\delta$ must be positive. The entrance zone we construct is a subset of $U_\delta$. We impose $\delta\in \left(0,\frac{4b}{d-1}\right).$ As shown in the following technical proposition, $\delta>0$ is needed for the sake of conveniences. The first parameter in the definition of $B$ is also introduced. \[from X to scalar curvature!\] In $U_\delta$, we can find a $p$ large enough such that $$\label{from X to scalar curvature} ((X_1-X_2)+(p-1)(X_3-X_2))(X_1-X_2+(p+1)(X_3-X_2))\geq \frac{1-\mathcal{G}}{d(d-1)}$$ along $\gamma_{s_1}$ in $U_\delta$. Since $X_1=\frac{1}{d}-X_2-X_3$, we can write inequality with respect to $\tilde{X}=X_3+X_2$ and $\tilde{Y}=X_3-X_2$. Straight forward computation shows that inequality is equivalent to $$\label{huge inequality in X} \begin{split} &\left(\left(p-\frac{1}{2}\right)\left(p+\frac{3}{2}\right)+\frac{1}{2(d-1)}\right)\tilde{Y}^2-\left(3p+\frac{3}{2}\right)\tilde{X}\tilde{Y}\\ &+\left(\frac{9}{4}+\frac{3}{2(d-1)}\right)\tilde{X}^2+\frac{2p+1}{d}\tilde{Y}-\frac{3d-1}{d(d-1)}\tilde{X}\geq 0. \end{split}$$ Note that $\tilde{X}$ and $\tilde{Y}$ are positive along $\gamma_{s_1}$ in $U_\delta$ by Proposition \[X\_3+X\_2>0\] and \[compare X\_3 with X\_2\]. Moreover, in $U_\delta$, we have $X_3- (1+\delta)X_2> 0$ along $\gamma_{s_1}$ by Proposition \[compare X\_3 with X\_2\]. Rewrite this condition in terms of $\tilde{X}$ and $\tilde{Y}$ so we have $ (2+\delta)\tilde{Y}-\delta\tilde{X}> 0 $ along $\gamma_{s_1}$ in $U_\delta$. Hence the LHS of is larger than $$\begin{split} &\left(\left(\left(p-\frac{1}{2}\right)\left(p+\frac{3}{2}\right)+\frac{1}{2(d-1)}\right)\frac{\delta}{2+\delta}-\left(3p+\frac{3}{2}\right)\right)\tilde{X}\tilde{Y}\\ &+\left(\frac{9}{4}+\frac{3}{2(d-1)}\right)\tilde{X}^2+\left(\frac{2p+1}{d}\frac{\delta}{2+\delta}-\frac{3d-1}{d(d-1)}\right)\tilde{X} \end{split}.$$ Since $\delta\in \left(0,\frac{4b}{d-1}\right)$ is fixed, we can choose $p$ large enough so that $$\begin{split} \label{lower bound for p} &\left(\left(p-\frac{1}{2}\right)\left(p+\frac{3}{2}\right)+\frac{1}{2(d-1)}\right)\frac{\delta}{2+\delta}\geq 3p+\frac{3}{2}\\ &\frac{2p+1}{d}\frac{\delta}{2+\delta}\geq \frac{3d-1}{d(d-1)} \end{split}$$ are satisfied. Then inequality is satisfied. Now we are ready to reveal the definition for $B$ and its last parameter. Define $$B_{p,k}(Z_1,Z_2,Z_3):=kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2.$$ For a fixed $\delta\in \left(0,\frac{4b}{d-1}\right)$, choose a $p$ that satisfies inequalities . Then define $$\label{yes} \begin{split} U_{(\delta,p,k)}&=S_3\cap \{Z_1+Z_2-Z_3\leq 0\}\cap \{B_{p,k}\geq 0\}\\ &\quad \cap\{(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)\geq 0\}, \end{split}$$ More requirements on the choice of $p$ and $k$ are added later. Before that, we prove the following. \[initial trap\] For any fixed $k>0$, $\gamma_{s_1}$ is initially trapped in $U_{(\delta,p,k)}$ as long as $s_1\in \left(0,\sqrt{\frac{ks_0(d+1)}{16d}}\right)$. With discussion in Section \[Completensese\], we know that $\gamma_{s_1}$ is initially in $S_3$ if $s_1>0$. Since all the other inequalities in reach equality at $p_0$, we need to substitute solution of linearized equation in each one of them. For $Z_1+Z_2-Z_3$, we have $$(d+1)s_0e^{\frac{2\eta}{d}}-4s_1e^{\frac{\eta}{d}}\sim-4s_1e^{\frac{\eta}{d}}<0$$ if $s_1<0$. Substitute solution of linearized equation to $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2$, we have $$\begin{split} &k(d+1)s_0e^{\frac{2\eta}{d}}\left(\frac{1}{d}-as_0e^{\frac{2\eta}{d}}+2s_1e^{\frac{\eta}{d}}\right)^{p+1}-\left(\frac{1}{d}-as_0e^{\frac{2\eta}{d}}-2s_1e^{\frac{\eta}{d}}\right)^p 16 s_1^2e^\frac{2\eta}{d}\\ &\sim \left(\frac{1}{d}\right)^p\left(\frac{ks_0(d+1)}{d}-16s_1^2\right) e^{\frac{2\eta}{d}}. \end{split}$$ Hence $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2>0$ initially along the projection of $\gamma_{s_1}$ when $s_1^2<\frac{ks_0(d+1)}{16d}.$ Finally, for $(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)$, we have $$\label{linearised deriv} 4s_1e^{\frac{\eta}{d}}\left(\frac{1}{d}-3as_0e^{\frac{2\eta}{d}}-s_1e^{\frac{\eta}{d}}\right)+\left(4ps_1e^{\frac{\eta}{d}}-2\left(\frac{1}{d}-as_0e^{\frac{2\eta}{d}}-2s_1e^{\frac{\eta}{d}}\right)\right)2s_1e^{\frac{\eta}{d}}\sim (4+8p)s_1^2e^{\frac{2\eta}{d}}>0.$$ Hence $\gamma_{s_1}$ is indeed trapped in $U_{(\delta,p,k)}$ initially when $s_1\in \left(0,\sqrt{\frac{ks_0(d+1)}{16d}}\right)$. We now specify our choice for $p$ and $k$. Projected to the $\omega_1\omega_2$-plane, the first two inequalities in is equivalent to $$\frac{\omega_2^p(1-\omega_2)^2}{k}\leq \omega_1\leq 1-\omega_2.$$ Write $l_0$ as a function $\mathcal{C}_0(\omega_2)=1-\omega_2$. Define $l_2\colon \mathcal{C}_2(\omega_2)=\frac{\omega_2^p(1-\omega_2)^2}{k}$. It is clear that $\mathcal{C}_0-\mathcal{C}_2=0$ at $\omega_2=1$. Our goal is to choose $p$ and $k$ so that $\mathcal{C}_0-\mathcal{C}_2$ vanishes again at some $\omega_<1$. Then we define $\hat{U}_{(\delta,p,k)}$ to be the compact subset of $U_{(\delta,p,k)}$ where $\omega_2\in [\omega_,1]$ and $\mathcal{C}_0>\mathcal{C}_2$ for $\omega_2\in (\omega_,1)$. Moreover, because we want to utilize Proposition \[from X to scalar curvature!\], parameters $p$ and $k$ are chosen to guarantee that $\omega_$ is not too small so that $\hat{U}_{(\delta,p,k)}\subset U_{\delta}.$ Specifically, we have the following proposition. \[verytechnicalthatitsiboring\] Let $p\geq 2$ be a fixed number large enough that it satisfies inequalities and $$\label{more lower bound for p} \frac{p}{p+1}\geq \frac{(d-1)(1+\delta)}{a+2b}.$$ Let $k>0$ be a number small enough so that $$\label{small k} k<\left(\frac{p}{p+1}\right)^p\frac{1}{p+1}.$$ Then there exists some $\omega_\in \left(\frac{p}{p+1},1\right)$ such that $$\label{hatyes} \hat{U}_{(\delta,p,k)}:=U_{(\delta,p,k)}\cap \{Z_2-\omega_Z_3\geq 0\}$$ is a compact subset of $U_\delta$. Although the proposition is true as long as $p>0,$ the technical condition $p\geq 2$ is imposed for computations in Lemma \[upper X\_3\] and . We first claim that $p$ exists. Because $\delta$ is a fixed number in $(0,\frac{4b}{d-1})$, we have $\frac{(d-1)(1+\delta)}{a+2b}<\frac{d-1+4b}{a+2b}=1.$ Hence we can choose $p$ large enough on top of inequalities to satisfies inequalities . Consider the function $$\label{function C} \mathcal{C}=\mathcal{C}_0-\mathcal{C}_2=1-\omega_2-\frac{\omega_2^p(1-\omega_2)^2}{k}=\frac{1-\omega_2}{k}\left(k-\omega_2^p(1-\omega_2)\right).$$ It is clear that $\mathcal{C}$ vanishes at $\omega_2=1$ and $\mathcal{C}>0$ near that point. Let $\tilde{\mathcal{C}}=k-\omega_2^p(1-\omega_2)$. Since $ \frac{d\tilde{\mathcal{C}}}{d\omega_2}=\omega_2^{p-1}(\omega_2-p(1-\omega_2)), $ we have $$ $\omega_2$ $0$ $\left(0,\frac{p}{p+1}\right)$ $\frac{p}{p+1}$ $\left(\frac{p}{p+1},1\right)$ $1$ ------------------------------------------ ----- -------------------------------- ----------------- -------------------------------- ----- $\frac{d\tilde{\mathcal{C}}}{d\omega_2}$ $0$ $<0$ $0$ $>0$ 1 $\tilde{\mathcal{C}}$ $k$ $k$ . $$ Therefore, for an arbitrary $p$, inequality is satisfied if and only if $\tilde{\mathcal{C}}\left(\frac{p}{p+1}\right)<0$. Then there exists some $\omega_\in \left(\frac{p}{p+1},1\right)$ such that $\tilde{\mathcal{C}}(\omega_)=0$ and $\tilde{\mathcal{C}}(\omega_2)> 0$ in $\left(\omega_,1\right)$. Since $\omega_2\leq 1$, that means for such an $\omega_$, we must have $\mathcal{C}(\omega_)=0$ and $\mathcal{C}(\omega_2)> 0$ in $\left(\omega_,1\right)$. The $\omega_2$-coordinate of the intersection point between $l_0$ and $l_1$ is $\frac{(d-1)(1+\delta)}{a+2b}$. By and Remark \[scalr of spher basic ineq\], the root $\omega_$ discussed above satisfies $$\label{thankyouomega} \omega_>\frac{p}{p+1}\geq \frac{(d-1)(1+\delta)}{a+2b}>\frac{a-2b}{a+2b}$$ We are ready to prove that $\hat{U}_{(\delta,p,k)}\subset U_\delta$. In other words, with our choice of $p$ and $k$ above, inequalities in the definition of $U_{(\delta,p,k)}$ and of $\hat{U}_{(\delta,p,k)}$ imply all inequalities in definition of $U_0$ and of $U_\delta$. Firstly, we need to show $\hat{U}_{(\delta,p,k)}\subset U_0$. With $-Z_1\geq Z_2-Z_3$ and $Z_2\geq \omega_Z_3$ satisfied in $S_3$, we have $$\begin{split} \label{living in U0 1} \mathcal{R}_1-\mathcal{R}_3 & =(Z_3-Z_1)(aZ_2-2bZ_1-2bZ_3)\\ &\geq (Z_3-Z_1)((a+2b)Z_3\omega_-4bZ_3)\\ &\geq (Z_3-Z_1)((a+2b)Z_3\omega_-(a-2b)Z_3)\quad \text{Remark \ref{scalr of spher basic ineq}}\\ &\geq 0 \quad \text{by \eqref{thankyouomega} and definition of $S_3$} \end{split}$$ and $$\begin{split} \label{living in U0 2} \mathcal{R}_3-\mathcal{R}_2 & =(Z_3-Z_2)(2bZ_3+2bZ_2-aZ_1)\\ &\geq (Z_3-Z_2)((a+2b)\omega_Z_3-(a-2b)Z_3) \\&\geq 0 \quad \text{by \eqref{thankyouomega} and definition of $S_3$} \end{split}.$$ Hence $\hat{U}_{(\delta,p,k)}\subset U_0$. In $\hat{U}_{(\delta,p,k)}$, we have $$\begin{split} \mathcal{R}_3-(1+\delta)\mathcal{R}_2&=(Z_3-Z_2)(2b(Z_3+Z_2)-aZ_1)-\delta(aZ_1Z_3+b(Z_2^2-Z_1^2-Z_3^2))\\ &=2b(Z_3^2-Z_2^2)-aZ_1(Z_3-Z_2)-\delta a Z_1Z_3-\delta b Z_2^2+\delta b Z_1^2+\delta b Z_3^2\\ &=(2+\delta )bZ_3^2-(1+\delta)aZ_1Z_3+\delta bZ_1^2-(\delta b+2b) Z_2^2+aZ_1Z_2. \end{split}$$ Treat the result of the computation above as a function of $Z_3$. It is a parabola centered at $\frac{(1+\delta)a}{(2+\delta)2b}Z_1.$ By , it is clear that $\frac{1}{1-\omega_}>\frac{a+2b}{4b}$. Since $\delta\in\left(0,\frac{4b}{a-2b}\right)$, it is straightforward to deduce that $\frac{a+2b}{4b}>\frac{(1+\delta)a}{(2+\delta)2b}$. From $Z_2\geq \omega_Z_3\geq \omega_(Z_1+Z_2)$ we also deduce $$\label{Z_1Z_2} Z_2\geq \frac{\omega_}{1-\omega}Z_1.$$ Therefore, we know that $ Z_1+Z_2\geq \frac{1}{1-\omega_}Z_1\geq \frac{a+2b}{4b}Z_1\geq \frac{(1+\delta)a}{(2+\delta)2b}Z_1 $ in $\hat{U}_{(\delta,p,k)}$. Hence $$\begin{split} \mathcal{R}_3-(1+\delta)\mathcal{R}_2&\geq \left.(\mathcal{R}_3-(1+\delta)\mathcal{R}_2)\right\|_{Z_3=Z_1+Z_2}\\ &\geq 0 \quad\text{by \eqref{R3R2}} \end{split}.$$ Finally, we need to show that $\hat{U}_{(\delta,p,k)}$ is compact. Since $Z_2-\omega_Z_3\geq 0$, we automatically have $ Z_1+Z_2\geq \omega_Z_3 $ in $\hat{U}_{(\delta,p,k)}$. By , we can deduce $\omega_>\frac{a-2b}{a+2b}>\frac{2b}{a}$ in $\hat{U}_{(\delta,p,k)}$, where the last inequality is from Remark \[scalr of spher basic ineq\]. Hence $a(Z_1+Z_2)-2b Z_3\geq 0$ in $\hat{U}_{(\delta,p,k)}$. Proposition \[sharp\] can be applied and all $Z_j$’s are bounded above. By the conservation law , we know that all variables are bounded. Hence $\hat{U}_{(\delta,p,k)}$ is compact. The proof is complete. \[Fig 4\] ![$\delta=0.7$, $p=12$, $k=\frac{1}{13+1}\left(\frac{12}{13}\right)^{12}$ for Case I](Projection4.jpg){width="3in"} We are ready show that $\hat{U}_{(\delta,p,k)}$ is the entrance zone. \[only through l0\] For $s_1\in\left(0,\sqrt{\frac{k(d+1)s_0}{16d}}\right)$ and suitable choice of $\delta,p$ and $k$ as described above, the integral curve $\gamma_{s_1}$ escapes $\hat{U}_{(\delta,p,k)}$ through $Z_1+Z_2-Z_3=0$. Suppose $\gamma_{s_1}$ does not escape through $Z_1+Z_2-Z_3=0$, then it can only escape through either $kZ_1Z_3^{p+1}-Z_2^{p}(Z_3-Z_2)^2=0$ or $(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)=0$. We prove that these situations are impossible. Since $$\label{The bump} \begin{split} &\left. \langle\nabla(kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2),V\rangle \right\|_{kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0}\\ &=\left. \langle\nabla(Z_3^{p+2}(k\omega_1-\omega_2^p(1-\omega_2)^2)),V\rangle \right\|_{kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0}\\ &=(p+2)Z_3^{p+2}\left(\mathcal{G}-\frac{1}{d}+2X_3\right)(k\omega_1-\omega_2^p(1-\omega_2)^2)\\ &\quad +Z_3^{p+2}(2k\omega_1(X_1-X_3)-2p\omega_2^p(X_2-X_3)(1-\omega_2)^2+4\omega_2^p(1-\omega_2)\omega_2(X_2-X_3))\\ &=Z_3^{p+2}(2k\omega_1(X_1-X_3)-2p\omega_2^p(X_2-X_3)(1-\omega_2)^2+4\omega_2^p(1-\omega_2)\omega_2(X_2-X_3))\\ &\quad \text{ since $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0$}\\ &=Z_3^{p+2}(2\omega_2^p(1-\omega_2)^2(X_1-X_3)-2p\omega_2^p(X_2-X_3)(1-\omega_2)^2+4\omega_2^p(1-\omega_2)\omega_2(X_2-X_3))\\ &\quad \text{ since $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0$}\\ &=2Z_2^p(Z_3-Z_2)((Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2))\\ &\geq 0 \quad \text{ definition of $\hat{U}_{(\delta,p,k)}$} \end{split},$$ it is impossible for $\gamma_{s_1}$ to escape $\hat{U}_{(\delta,p,k)}$ through $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0$. For the other defining inequality, we have $$\label{Bump Derivative} \begin{split} &\left. \langle\nabla((Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)),V\rangle \right\|_{(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)=0}\\ &=\left. \langle\nabla(Z_3((1-\omega_2)(X_1-X_3)+(p(1-\omega_2)-2\omega_2)(X_3-X_2))),V\rangle \right\|_{(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2)=0}\\ &=Z_3\left(\mathcal{G}-\frac{1}{d}+2X_3\right)((1-\omega_2)(X_1-X_3)+(p(1-\omega_2)-2\omega_2)(X_3-X_2)))\\ &\quad +Z_3((1-\omega_2)(X_1-X_3)+(p(1-\omega_2)-2\omega_2)(X_3-X_2)))(\mathcal{G}-1)\\ &\quad +Z_3((1-\omega_2)(\mathcal{R}_1-\mathcal{R}_3)+(p(1-\omega_2)-2\omega_2)(\mathcal{R}_3-\mathcal{R}_2))\\ &\quad +Z_3(2\omega_2(X_3-X_2)(X_1-X_3)+2(p+2)\omega_2(X_3-X_2)^2)\\ &=Z_3((1-\omega_2)(\mathcal{R}_1-\mathcal{R}_3)+(p(1-\omega_2)-2\omega_2)(\mathcal{R}_3-\mathcal{R}_2))\\ &\quad +Z_3(2\omega_2(X_3-X_2)(X_1-X_3)+2(p+2)\omega_2(X_3-X_2)^2)\\ &\quad \text{ since $(Z_3-Z_2)(X_1-X_3)+(p(Z_3-Z_2)-2Z_2)(X_3-X_2))=0$}\\ &=(Z_3-Z_2)(\mathcal{R}_1-\mathcal{R}_3)+(p(Z_3-Z_2)-2Z_2)(\mathcal{R}_3-\mathcal{R}_2)\\ &\quad +2Z_2(X_3-X_2)(X_1-X_3)+2(p+2)Z_2(X_3-X_2)^2\\ &=(Z_3-Z_2)(\mathcal{R}_1-\mathcal{R}_3+p(\mathcal{R}_3-\mathcal{R}_2))-2Z_2(Z_3-Z_2)(2bZ_3+2bZ_2-aZ_1)\\ &\quad +2Z_2(X_3-X_2)((X_1-X_3)+(p+2)(X_3-X_2))\\ &=(Z_3-Z_2)(\mathcal{R}_1-\mathcal{R}_3+p(\mathcal{R}_3-\mathcal{R}_2)+2Z_2(aZ_1-2bZ_2-2bZ_3))\\ &\quad +(Z_3-Z_2)((X_1-X_2)+(p-1)(X_3-X_2))(X_1-X_2+(p+1)(X_3-X_2))\\ &\quad \text{ since $2Z_2(X_3-X_2)=(Z_3-Z_2)(X_1-X_3)+p(Z_3-Z_2)(X_3-X_2)$} \end{split}.$$ Because $\hat{U}_{(\delta,p,k)}\subset U_\delta$, we can apply Proposition \[from X to scalar curvature!\] to the last line of and continue the computation as $$\begin{split} &\geq (Z_3-Z_2)\left(\mathcal{R}_1-\mathcal{R}_3+p(\mathcal{R}_3-\mathcal{R}_2)+2Z_2(aZ_1-2bZ_2-2bZ_3)+\frac{1-\mathcal{G}}{d(d-1)}\right) \\ &=(Z_3-Z_2)\left(\mathcal{R}_1-\mathcal{R}_3+p(\mathcal{R}_3-\mathcal{R}_2)+2Z_2(aZ_1-2bZ_2-2bZ_3)+\frac{1}{d-1}(\mathcal{R}_1+\mathcal{R}_2+\mathcal{R}_3)\right) \quad \text{by \eqref{conservation law}}\\ &=\left(2-\frac{1}{d-1}\right)bZ_1^2+\left(aZ_3((p+1)\omega_2-p)+\frac{a}{d}(Z_2+Z_3)\right)Z_1\\ &\quad +Z_3^2\left(-\left(2bp+4b+\frac{b}{d-1}\right)\omega_2^2+\left( \frac{a}{d-1}+a-4b\right)\omega_2+\left(2pb-2b-\frac{b}{d-1}\right)\right). \end{split}$$ The first term of the computation result above is obviously positive. The second term is positive because $\omega_2\geq \omega_>\frac{p}{p+1}$ in $\hat{U}_{(\delta,p,k)}$. The positivity of the last term depends on the one of parabola $$\pi(\omega_2)=-\left(2bp+4b+\frac{b}{d-1}\right)\omega_2^2+\left( \frac{a}{d-1}+a-4b\right)\omega_2+\left(2pb-2b-\frac{b}{d-1}\right).$$ Since we impose $p\geq 2$, it is clear that $\pi(0)$ is positive. As the coefficient of the first term is negative, we know that $\pi$ has two roots with different signs. It is easy to verify that $\pi(1)=0$. Then we conclude that $\pi$ is non-negative for all $\omega_2\in [0,1]$. Therefore, the computation of is non-negative and only vanishes when $Z_1=0$ and $Z_2=Z_3$. Notice that there is no need to check the possibility that $\gamma_{s_1}$ may escape through $Z_2-\omega_Z_3=0$. Because when the equality of $Z_2-\omega_Z_3\geq 0$ is reached at some point $\gamma_{s_1}(\eta_)$, it implies that the function $\mathcal{C}$ in vanishes at that point. Specifically, we have $ 1-\omega_2=\frac{\omega_2^p(1-\omega_2)^2}{k} $ at that point. But then $$Z_1Z_3^{p+1}\leq Z_3^{p+1}(Z_3-Z_2)=\frac{Z_2^p(Z_3-Z_2)^2}{k}\leq Z_1Z_3^{p+1},$$ which implies $kZ_1Z_3^{p+1}-Z_2^p(Z_3-Z_2)^2=0$ at that point and this case is included in the computation at the beginning of the proof. \[critical points in Uhat\] The only critical points in $\hat{U}_{(\delta,p,k)}$ are $p_0$ and those of Type I. By proposition \[critical points\], it is clear that $p_0$ and critical points of Type I are in $\hat{U}_{(\delta,p,k)}$. We first eliminate critical points with negative $Z_j$ entry. Since $\hat{U}_{(\delta,p,k)}\subset S_3$, we can eliminate critical points with $Z_3$ smaller than the other two $Z_j$’s. Because $X_3\geq 0$ in $S_3$, there is no critical points of Type II. Since $Z_2\geq pZ_1\geq Z_1$ in $\hat{U}_{(\delta,p,k)}$ by and , there is no critical points other than $p_0$ and those of Type I in $\hat{U}_{(\delta,p,k)}$. \[Z1Z2Z3\] The function $Z_1Z_2Z_3$ stays positive and increases along $\gamma_{s_1}$. Since $\mathcal{H}\equiv 1$, it is clear that $\mathcal{G}\geq \frac{1}{n}$. Hence $$\label{increase} (Z_1Z_2Z_3)'=Z_1Z_2Z_3\left(3\mathcal{G}-\frac{1}{d}\right)\geq 0.$$ Since $Z_1Z_2Z_3$ is initially positive along $\gamma_{s_1}$, the proof is complete. We are ready to prove the completeness of Ricci-flat metrics represented by $\gamma_{s_1}$ with $s_1$ close enough to zero. \[long existence\] There exists a $k>0$ such that an unstable integral curve $\gamma_{s_1}$ to on $C\cap \{\mathcal{H}\equiv 1\}$ emanating from $p_0$ is defined on $\mathbb{R}$ if $s_1\in \left(-\sqrt{\frac{k(d+1)s_0}{16d}},\sqrt{\frac{k(d+1)s_0}{16d}}\right)$. If $s_1>0$, the curve $\gamma_{s_1}$ is initially trapped in $\hat{U}_{(\delta,p,k)}$ as long as $s_1\in\left(0,\sqrt{\frac{k(d+1)s_0}{16d}}\right)$. The function $Z_1+Z_2-Z_3$ vanishes at $p_0$ and it is negative along $\gamma_{s_1}$ in $\hat{U}_{(\delta,p,k)}$. By Lemma \[only through l0\], the function $Z_1+Z_2-Z_3$ must vanish at $\gamma_{s_1}(\eta_)$ for some $\eta_\in\mathbb{R}$. Then we must have $(Z_1+Z_2-Z_3)'(\gamma_{s_1}(\eta_))\geq 0$. But $$\begin{split} (Z_1+Z_2-Z_3)'(\gamma_{s_1}(\eta_))&=\langle\nabla(Z_1+Z_2-Z_3),V\rangle(\gamma_{s_1}(\eta_))\\ &=\left(\left.\langle\nabla(Z_1+Z_2-Z_3),V\rangle\right\|_{Z_1+Z_2-Z_3=0}\right)(\gamma_{s_1}(\eta_))\\ &=\left(Z_1(X_1-X_3)+Z_2(X_2-X_3)\right)(\gamma_{s_1}(\eta_)). \end{split}$$ Hence $\gamma_{s_1}(\eta_)$ is in $\partial \hat{S}_3.$ By Proposition \[verytechnicalthatitsiboring\], we know that $\hat{U}_{(\delta,p,k)}$ is in $U_0$, where $\mathcal{R}_1-\mathcal{R}_3\geq 0$ and $\mathcal{R}_3-\mathcal{R}_2\geq 0$ hold. Then with the similar argument in Proposition \[compare X\_3 with X\_2\], we know that $X_1>X_3>X_2$ along $\gamma_{s_1}$ in $\hat{U}_{(\delta,p,k)}$. Hence the intersection point $\gamma_{s_1}(\eta_)$ is not $p_0$. By Proposition \[Z1Z2Z3\], we know that $\gamma_{s_1}(\eta_)$ cannot be a critical point of Type I. By Proposition \[critical points in Uhat\], we know that $\gamma_{s_1}(\eta_)$ is not a critical point. Then by Lemma \[never escape\], $\gamma_{s_1}$ continue to flows inward $\hat{S}_3$ from $\gamma_{s_1}(\eta_)$ and never escape. Therefore, such a $\gamma_{s_1}$ is defined on $\mathbb{R}$. By symmetry, similar result can be obtained for $s_1\in\left(-\sqrt{\frac{k(d+1)s_0}{16d}},0\right)$. If $s_1=0$, then we are back to the special case by Remark \[Mcontainy\]. By the discussion at the end of Section \[Local Existence\], Lemma \[long existence\] proves the first half of Theorem \[main 1\]. \[posiscaler\] For $\gamma_{s_1}$ with $s_1\in \left(0, \sqrt{\frac{k(d+1)s_0}{16d}}\right)$, it can be shown that $\mathcal{R}_2$ is negative initially by substituting . Hence the Ricci-flat metrics represented does not have the property introduced in Remark \[positivericci\]. By straightforward computation, however, it processes a weaker condition that the scalar curvature of each hypersurface remain positive. Asymptotic Limit {#aymptoticlimit} ================ In this section, we study the asymptotic behavior of complete Ricci-flat metrics constructed above. Each integral curve $\gamma_{s_1}$ mentioned below satisfies the condition in Lemma \[long existence\], i.e., each $\gamma_{s_1}$ is trapped in $\hat{U}_{(\delta,p,k)}$ initially and then enter $\hat{S}_3$ in finite time. \[omega\_1+omega\_2isnever1\] Let $\gamma_{s_1}$ be a long time existing integral curve that intersects with $\hat{S}_3$ at a non-critical point $\gamma_{s_1}(\eta_)$. Then function $\omega_1+\omega_2>1$ along $\gamma_{s_1}(\eta)$ for $\eta\in(\eta_,\infty)$. Note that $(\omega_1+\omega_2)(\gamma_{s_1}(\eta_))=1$. By Lemma \[never escape\], we know that $\gamma_{s_1}(\eta)\in \hat{S}_3$ for $\eta\geq \eta_$. We have $$\begin{split} (\omega_1+\omega_2)'(\gamma_{s_1}(\eta_))&=(2\omega_1(X_1-X_3)+2\omega_2(X_2-X_3))(\gamma_{s_1}(\eta_))\\ &\geq 0\quad\text{by definition of $\hat{S}_3$}. \end{split}$$ Suppose $(\omega_1+\omega_2)'(\gamma_{s_1}(\eta_))=0$. Recall in the proofs of Lemma \[long existence\], we know that $X_1>X_3>X_2$ at $\gamma_{s_1}(\eta_)$. By and , we have $$\begin{split} (\omega_1+\omega_2)''(\gamma_{s_1}(\eta_))&\geq \left(4\omega_1(X_1-X_3)^2+4\omega_2(X_2-X_3)^2\right)(\gamma_{s_1}(\eta_))>0 \end{split}.$$ Suppose there exists $\eta_{1}\in(\eta_,\infty)$ that $(\omega_1+\omega_1)(\gamma_{s_1}(\eta_{1}))=1$. We know from the computation above that there exists $\eta_{2}\in(\eta_{},\eta_1)$ such that $(\omega_1+\omega_2)(\gamma_{s_1}(\eta_{2}))>1$. By mean value theorem, there exists $\eta_3\in[\eta_2,\eta_1]$ such that $(\omega_1+\omega_2)'(\gamma_{s_1}(\eta_3))=(2\omega_1(X_1-X_3)+2\omega_2(X_2-X_3))(\gamma_{s_1}(\eta_3))<0,$ a contradiction to the definition of $\hat{S}_3$. \[upper X\_3\] The variable $X_3$ is smaller than $\frac{1}{n}$ along integral curves $\gamma_{s_1}$. Since $\mathcal{H}\equiv 1$, $X_3\leq \frac{1}{n}$ is equivalent to $X_1+X_2-2X_3\geq 0$. The function $X_1+X_2-2X_3$ is positive at $p_0$. Suppose the function vanishes along $\gamma_{s_1}$ at some point in $\hat{U}_{(\delta,p,k)}$, then we have $$\label{X_3deri} \begin{split} \left.(X_1+X_2-2X_3)'\right\|_{X_1+X_2-2X_3=0}&=(X_1+X_2-2X_3)(\mathcal{G}-1)+\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3\\ &=\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3 \quad \text{since $X_1+X_2-2X_3=0$}\\ &=a(Z_2Z_3+Z_1Z_3-2Z_1Z_2)-2b(2Z_3^2-Z_1^2-Z_2^2). \end{split}$$ Consider the computation result above as a function $$\mathcal{J}(Z_3)=-4bZ_3^2+a(Z_1+Z_2)Z_3+2bZ_1^2+2bZ_2^2-2aZ_1Z_2.$$ Since $Z_1+Z_2\leq Z_3\leq \frac{Z_2}{\omega_}$ in $\hat{U}_{(\delta,p,k)}$, the positivity of $\mathcal{J}$ is implied by those of $\mathcal{J}(Z_1+Z_2)$ and $\mathcal{J}\left(\frac{Z_2}{\omega_}\right)$. With the choice $p\geq 2$, inequality implies $\omega_>\frac{p}{p+1}\geq \frac{2}{3}\geq \frac{4b}{a}$. Hence it is sufficient to prove a stronger condition: the positivity of $\mathcal{J}(Z_1+Z_2)$ and $\mathcal{J}\left(\frac{a}{4b}Z_2\right)$. We have $$\label{Joneend1} \begin{split} \mathcal{J}(Z_1+Z_2)&=(a-2b)(Z_1^2+Z_2^2)-8bZ_1Z_2\\ &\geq 4b(Z_1-Z_2)^2\quad\text{Remark \ref{scalr of spher basic ineq}}\\ &\geq 0 \end{split}.$$ And we have $$\label{Joneend2} \begin{split} \mathcal{J}\left(\frac{a}{4b}Z_2\right)&=\left(\frac{a^2}{4b}-2a\right)Z_1Z_2+2b(Z_1^2+Z_2^2)\\ &\geq \left(\frac{a^2}{4b}-2a\right)Z_1Z_2+4bZ_1Z_2\\ &\geq 0 \end{split}.$$ All $Z_j$’s are positive along $\gamma_{s_1}$. Hence by and , computation can vanish only if $Z_1=Z_2=\frac{Z_3}{2}$. But with $p\geq 2$ imposed, $Z_2\geq \omega_Z_3\geq \frac{2}{3}Z_3\geq\frac{Z_3}{2}$ in $\hat{U}_{(\delta,p,k)}$. Hence $\mathcal{J}$ can only vanish at the origin of $Z$-space, which is impossible for $\gamma_{s_1}$ to reach by . Therefore, $X_1+X_2-2X_3$ never vanishes along $\gamma_{s_1}$ at least till $\gamma_{s_1}$ intersect with $\partial \hat{S}_3$ at some $\gamma_{s_1}(\eta_)$. $\gamma_{s_1}$ is in $\hat{S}_3$ for $\eta\in[\eta_,\infty)$. The function $X_1+X_2-2X_3$ is positive at $\gamma_{s_1}(\eta_)$. Suppose the function vanishes at some point along $\gamma_{s_1}$ in $\hat{S}_3$, then $$\label{X3der} \begin{split} \left.(X_1+X_2-2X_3)'\right\|_{X_1+X_2-2X_3=0}&=\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3\\ &=a(Z_2Z_3+Z_1Z_3-2Z_1Z_2)-2b(2Z_3^2-Z_1^2-Z_2^2)\\ &\geq a(Z_2Z_3+Z_1Z_3-2Z_1Z_2)-\dfrac{a}{3}(2Z_3^2-Z_1^2-Z_2^2) \quad \text{Remark \ref{scalr of spher basic ineq}}\\ &\geq \dfrac{a}{3}(Z_1+Z_2-Z_3)(2Z_3-Z_1-Z_2)\\ &\geq 0 \quad\text{definition of $\hat{S}_3$} \end{split}.$$ By Proposition \[Z1Z2Z3\], there is no need to consider the case where each $Z_j$ vanishes. For Case I-III, suppose computation above vanishes at some point on $\gamma_{s_1}$. Then one possibility is that $Z_1=Z_2=Z_3$ at that point. But then $X_3-X_j+\rho(Z_3-Z_j)=X_3-X_j=\frac{1}{n}-X_j\geq 0$ at that point by the definition of $\hat{S}_3$. Then we must have $X_j=\frac{1}{n}$ for each $j$. Hence the point must be the critical point $p_1$, a contradiction. For Case I in particular, there is an extra possibility where $Z_1=Z_2=\frac{Z_3}{2}$ at that point. It is ruled out by Lemma \[omega\_1+omega\_2isnever1\]. Hence $X_3<\frac{1}{n}$ along $\gamma_{s_1}$ all the way. We can now describe the asymptotic limit of $\gamma_{s_1}$. \[ricci flat limit\] The integral curve $\gamma_{s_1}$ converges to $p_1$. Since $\gamma_{s_1}$ does not hit any critical point in $\hat{U}_{(\delta,p,k)}$ by Lemma \[long existence\], we can focus on the behavior of the integral curve in the set $\hat{S}_3$. By Proposition \[Z1Z2Z3\], we know that $Z_1Z_2Z_3$ converges to some positive number along $\gamma_{s_1}$. There exists a sequence $\{\eta_m\}$ such that $\lim\limits_{m\to \infty}\eta_m=\infty$ and $\lim\limits_{m\to \infty}\mathcal{G}=\frac{1}{n}$. Hence $\lim\limits_{m\to \infty}X_j(\eta_m)=\frac{1}{n}$ for each $j$. But then $$\begin{split} 0&=\lim_{m\to\infty} (X_1+X_2-2X_3)'(\eta_m)\\ &=\lim_{m\to\infty}\left((X_1+X_2-2X_3)(\mathcal{G}-1)+\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3\right)(\eta_m)\\ &=\lim_{m\to\infty}\left(\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3\right)(\eta_m)\\ &\geq 0 \quad\text{by \eqref{X3der}} \end{split}.$$ Therefore, either $\lim\limits_{m\to \infty} Z_3(\omega_1+\omega_2-1)(\eta_m)=0$ or $\lim\limits_{m\to\infty}(2Z_3-Z_1-Z_2)(\eta_m)=0$. It is clear that $\lim\limits_{m\to\infty}Z_3(\eta_m)\neq 0$ as $Z_3\geq Z_1,Z_2$ in $S_3$ and $Z_1Z_2Z_3$ increases along $\gamma_{s_1}$. By Lemma \[omega\_1+omega\_2isnever1\], we know that $\lim\limits_{m\to\infty}(\omega_1+\omega_2-1)(\eta_m)\neq 0$. Hence $\lim\limits_{m\to\infty}(2Z_3-Z_1-Z_2)(\eta_m)=0$. Since $Z_3\geq Z_1,Z_2$ in $S_3$, we conclude that $\lim\limits_{m\to \infty}(Z_3-Z_1)(\eta_m)=\lim\limits_{m\to \infty}(Z_3-Z_2)(\eta_m)=0$. With , we conclude that $\lim\limits_{m\to\infty} \gamma_{s_1}(\eta_m)=p_1$. Hence $p_1$ is in the $\omega$-limit set of $\gamma_{s_1}$. Consider $p_1=\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\alpha,\alpha,\alpha\right)$, where $\alpha=\frac{1}{n}\sqrt{\frac{n-1}{a-b}}$. By , the linearization at $p_1$ is $$\label{linearizationat p1} \mathcal{L}(p_1)=\begin{bmatrix} \frac{5}{3n}-1&\frac{2}{3n}&\frac{2}{3n}&2b\alpha&(a-2b)\alpha&(a-2b)\alpha\\ \frac{2}{3n}&\frac{5}{3n}-1&\frac{2}{3n}&(a-2b)\alpha&2b\alpha &(a-2b)\alpha\\ \frac{2}{3n}&\frac{2}{3n}&\frac{5}{3n}-1&(a-2b)\alpha &(a-2b)\alpha &2b\alpha\\ \frac{5}{3}\alpha&-\frac{1}{3}\alpha&-\frac{1}{3}\alpha&0&0&0\\ -\frac{1}{3}\alpha&\frac{5}{3}\alpha&-\frac{1}{3}\alpha&0&0&0\\ -\frac{1}{3}\alpha&-\frac{1}{3}\alpha&\frac{5}{3}\alpha&0&0&0\\ \end{bmatrix}.$$ Its eigenvalues and corresponding eigenvectors are $$\lambda_1=\frac{1}{n}-1,\quad \lambda_2=\lambda_3=\beta_1,\quad\lambda_4=\lambda_5=\beta_2,\quad \lambda_6=\frac{2}{n}.$$ $$v_1=\begin{bmatrix} n-1\\ n-1\\ n-1\\ -n\alpha\\ -n\alpha\\ -n\alpha \end{bmatrix}, v_2=\begin{bmatrix} -\frac{\beta_1}{2\alpha}\\ \frac{\beta_1}{2\alpha}\\ 0\\ -1\\ 1\\ 0 \end{bmatrix},v_3=\begin{bmatrix} -\frac{\beta_1}{2\alpha}\\ 0\\ \frac{\beta_1}{2\alpha}\\ -1\\ 0\\ 1 \end{bmatrix}, v_4=\begin{bmatrix} -\frac{\beta_2}{2\alpha}\\ \frac{\beta_2}{2\alpha}\\ 0\\ -1\\ 1\\ 0 \end{bmatrix}, v_5=\begin{bmatrix} -\frac{\beta_2}{2\alpha}\\ 0\\ \frac{\beta_2}{2\alpha}\\ -1\\ 0\\ 1 \end{bmatrix}, v_6=\begin{bmatrix} 2\\ 2\\ 2\\ n\alpha\\ n\alpha\\ n\alpha \end{bmatrix},$$ where $$\beta_1=-\dfrac{n-1+\sqrt{(n-1)^2-8n^2\alpha^2(a-4b)}}{2n}<0,\quad \beta_2=-\dfrac{n-1-\sqrt{(n-1)^2-8n^2\alpha^2(a-4b)}}{2n}<0.$$ Evaluate at $p_1$, it is clear that $ T_{p_1}(C\cap \{\mathcal{H}\equiv 1\})={\mathrm{span}}\{v_2,v_3,v_4,v_5\}.$ Critical point $p_1$ is a sink. Hence $\lim\limits_{\eta\to\infty}\gamma_{s_1}=p_1$. \[ACCCCCmetric\] Ricci-flat metrics represented by $\gamma_{s_1}$ are AC. For each $j$, we have $$\label{ACrate} \begin{split} \lim_{t\to \infty} \dot{f_j}=\lim_{\eta\to \infty}\frac{X_j}{\sqrt{Z_kZ_l}}=\sqrt{\dfrac{a-b}{n-1}}. \end{split}$$ Therefore by Definition \[AC\], the Ricci-flat metric represented by $\gamma_{s_1}$ has conical asymptotic limit $dt^2+t^2\dfrac{a-b}{n-1} Q.$ Lemma \[ricci flat limit\] and Lemma \[ACCCCCmetric\] imply Theorem \[main 2\]. Singular Ricci-flat Metrics {#a digression} =========================== This section is dedicated to singular Ricci-flat metrics. Note that critical points $p_1$ and $p_2$ can be viewed as integral curves defined on $\mathbb{R}$. They correspond to singular Ricci-flat metrics $g=dt+t^2\dfrac{a-b}{n-1} Q$. This is consistent with the fact that the Euclidean metric cone over a proper scaled homogeneous Einstein manifold is Ricci-flat. For Case I in particular, the normal Einstein metric on $G/K$ is strict nearly Kähler. Hence the metric cone represented by $p_1$ is the singular $G_2$ metric discovered in [@bryant1987metrics]. Note that functions $F_j$’s in do note vanish at $p_2$. Therefore, the Euclidean metric cone over the Kähler–Einstein metric has generic holonomy. There are also singular Ricci-flat metrics represented by nontrivial integral curves. Recall Remark \[G2\], The cohomogeneity one $G_2$ condition is given by $F_j\equiv 0$ for each $j$. Eliminate $X_j$’s in the conservation law $C$ shows that $$\begin{split} \blacktriangle= &C\cap \{\mathcal{H}\equiv 1\} \cap P\cap \{F_1\equiv F_2\equiv F_3\equiv 0\}\\ &=\{Z_1+Z_2+Z_3-1\equiv 0\}\cap P\cap \{F_1\equiv F_2\equiv F_3\equiv 0\} \end{split}$$ is an invariant 2-dimensional plane with boundary. Its projection in $Z$-space is plotted in Figure \[Fig G2\]. Black squares are critical poitns of Type II. Linearization at these points shows that they are sources. Furthermore, for any $\xi\in\mathbb{R}$, $\blacktriangle\cap \{Z_3(Z_1-Z_2)-\xi Z_2(Z_1-Z_3)\equiv 0\}$ is a pair of integral curves that connects three critical points. If $\xi\neq 0,1$, then these two integral curves connect $p_1$ with two distinct critical points of Type II. These integral curves represent singular cohomogeneity one $G_2$ metrics on $(0,\infty)\times G/K$ that do not have smooth extension to $G/H$[@cvetivc2002cohomogeneity][@cleyton_cohomogeneity-one_2002]. They all share the same AC limit as the metric cone over $G/K$ equipped with the normal Einstein metric. When $\xi=0,1$, then one of the integral curve connects a critical point of Type II with $p_1$ and the other one connects a critical point of Type III with $p_1$. In particular, if $\xi=1$, then we recover $\gamma_0$ that represents the $G_2$ metric, connecting $p_0$ and $p_1$. ![Integral curves on $\blacktriangle$ with $0<\xi\leq 1$[]{data-label="Fig G2"}](SingularG2.png){width="6.3in"} There are singular metrics with generic holonomy. We construct a new compact invariant set whose boundary includes $p_1$ and $p_2$. Consider $$\check{S}_3=S_3\cap \{X_1\equiv X_2, Z_1\equiv Z_2\}\cap\{X_1+X_2-2X_3\geq 0\}\cap \left\{(d-1)^2Z_1Z_2-4b^2Z_3^2\geq0\right\}.$$ $\check{S}_3$ is a compact invariant set. It is easy to show that $\{X_1\equiv X_2, Z_1\equiv Z_2\}$ is flow invariant. In fact, even if we define $\check{S}_3$ without $\{X_1\equiv X_2, Z_1\equiv Z_2\}$, the set is still compact and invariant. However, considering the subsystem does make the computation easier. In $\check{S}_3$, we have $$\begin{split} 4b^2Z_3^2&\leq (d-1)^2Z_1Z_2<a^2Z_1Z_2\leq a^2(Z_1+Z_2)^2. \end{split}$$ Hence we can apply Proposition \[sharp\] and conclude that inequality holds in $\check{S}_3$. As $Z_3$ is bounded above by $\frac{d-1}{2b}\sqrt{Z_1Z_2}$ in $\check{S}_3$, the compactness follows by . To show that $\check{S}_3$ is invariant, consider $$\begin{split} \left. \langle\nabla(X_1+X_2-2X_3),V\rangle \right\|_{X_1+X_2-2X_3=0}&=(X_1+X_2-2X_3)(\mathcal{G}-1)+\mathcal{R}_1+\mathcal{R}_2-2\mathcal{R}_3\\ &=2\mathcal{R}_2-2\mathcal{R}_3\quad \text{ since $Z_1\equiv Z_2$ in $\check{S}_3$ and $X_1+X_2-2X_3=0$}\\ &=2(Z_3-Z_2)((d-1)\sqrt{Z_1Z_2}-2bZ_3)\quad \text{ since $Z_1\equiv Z_2$ in $\check{S}_3$}\\ &\geq 0 \quad \text{by definition of $\check{S}_3$} \end{split}.$$ Moreover, we have $$\begin{split} &\left. \langle\nabla((d-1)^2Z_1Z_2-4b^2Z_3^2),V\rangle \right\|_{(d-1)^2Z_1Z_2-4b^2Z_3^2=0}\\ &=\left.\nabla\left(Z_3^2\left((d-1)^2\dfrac{Z_1Z_2}{Z_3^2}-4b^2\right)\right)\right\|_{(d-1)^2Z_1Z_2-4b^2Z_3^2=0}\\ &= \left((d-1)^2Z_1Z_2-4b^2Z_3\right)\left(\mathcal{G}-\frac{1}{d}+2X_3\right)+2(d-1)^2Z_1Z_2\left(X_1+X_2-2X_3\right)\\ &=2(d-1)^2Z_1Z_2\left(X_1+X_2-2X_3\right)\quad \text{ since $(d-1)^2Z_1Z_2-4b^2Z_3^2=0$}\\ &\geq 0 \quad \text{ by definition of $\check{S}_3$} \end{split}.$$ Hence $\check{S}_3$ is a compact invariant set. \[singularcurve\] There exists an integral curve $\Gamma$ defined on $\mathbb{R}$ emanating from $p_2$ in $\check{S}_3$. Consider $p_2=\left(\frac{1}{n},\frac{1}{n},\frac{1}{n},\frac{2}{n}\sqrt{\frac{(n-1)b}{(d-1)(a+2b)}},\frac{2}{n}\sqrt{\frac{(n-1)b}{(d-1)(a+2b)}},\frac{1}{n}\sqrt{\frac{(n-1)(d-1)}{b(a+2b)}}\right)$. For simplicity, denote $Z_=\frac{2}{n}\sqrt{\frac{(n-1)b}{(d-1)(a+2b)}}$. The linearization at $p_2$ is $$\mathcal{L}(p_2)=\begin{bmatrix} \frac{5}{9d}-1&\frac{2}{9d}&\frac{2}{9d}&2bZ_&\left(\frac{a(d-1)}{2b}-2b\right)Z_&2bZ_\\ \frac{2}{9d}&\frac{5}{9d}-1&\frac{2}{9d}&\left(\frac{a(d-1)}{2b}-2b\right)Z_&2bZ_&2bZ_\\ \frac{2}{9d}&\frac{2}{9d}&\frac{5}{9d}-1&(d-1)Z_&(d-1)Z_&(d-1)Z_\\ \frac{5}{3}Z_&-\frac{1}{3}Z_&-\frac{1}{3}Z_&0&0&0\\ -\frac{1}{3}Z_&\frac{5}{3}Z_&-\frac{1}{3}Z_&0&0&0\\ -\frac{d-1}{6b}Z_&-\frac{d-1}{6b}Z_&-\frac{5(d-1)}{6b}Z_&0&0&0\\ \end{bmatrix}$$ Straightforward computation shows that for all cases, $L(p_2)$ is a hyperbolic critical point that has only one unstable eigenvalues with the corresponding eigenvector as $$\check{\lambda}=\dfrac{1}{2n}\left(\sqrt{(n-1)^2+96n(d-1)(a-4b)Z_^2}-(n-1)\right),\quad \check{v}=\begin{bmatrix} b\check{\lambda}\\ b\check{\lambda}\\ -2b\check{\lambda}\\ 2bZ_\\ 2bZ_\\ -2(d-1)Z_ \end{bmatrix}$$ Evaluate at $p_2$, it is clear that $\check{v}$ are tangent to $C\cap \{\mathcal{H}\equiv 1\}$. Fix $\check{s}_0>0$, there exists a unique trajectory $\Gamma$ emanating from $p_2$ with $ \Gamma\sim p_2+\check{s}_0e^{\check{\lambda}\eta}\check{v}. $ It is easy to check that $p_2\in\partial\check{S}_3$ with only $X_1+X_2-2X_3$ and $(d-1)^2Z_1Z_2-4b^2Z_3^2$ vanished at $p_2$. By straightforward computation, we know that $\Gamma$ is trapped in $\check{S}_3$ initially. The integral curve is hence defined on $\mathbb{R}$. Functions $f_j$’s that correspond to solutions $\Gamma$ are defined on $[0,\infty)$. \[limitGamma\] The integral curve $\Gamma$ converges to $p_1$. Since $\check{S}_3$ is a compact invariant set with $X_3\leq \frac{1}{n}$. Arguments in Lemma \[ricci flat limit\] and \[ACCCCCmetric\] carry over. Hence for $\Gamma$, we have $\lim\limits_{\eta\to\infty}\Gamma=p_1$. For each $j$, we have $$\label{ACpointy} \begin{split} &\lim_{t\to 0} \dot{f_j}=\lim_{\eta\to -\infty}\frac{X_j}{\sqrt{Z_kZ_3}}=\frac{\check{\lambda}}{2Z_} \quad j,k\in\{1,2\}\\ &\lim_{t\to 0} \dot{f_3}=\lim_{\eta\to -\infty}\frac{X_3}{\sqrt{Z_1Z_2}}=\frac{b\check{\lambda}}{(d-1)Z_} \end{split}.$$ Hence $f_1=f_2\sim \frac{\check{\lambda}}{2Z_}t$ and $f_3\sim \frac{b\check{\lambda}}{(d-1)Z_}t$ as $t\to 0$. Since $\lim\limits_{t\to \infty} \dot{f_1}\neq \lim\limits_{t\to \infty} \dot{f_3}$, $\Gamma$ represents a singular metric whose end at $t\to 0$ is a conical singularity as a metric cone over the alternative Einstein metrics. Lemma \[singularcurve\] and Lemma \[limitGamma\] then prove the following theorem. \[pointy1\] Up to homothety, there exists a unique singular Ricci-flat metric on $(0,\infty)\times G/K$ that at the end with $t\to 0$, it admits conical singularity as the metric cone over $G/K$ with alternative Einstein metric. It has an AC limit at the end with $t\to \infty$ as the metric cone over $G/K$ with normal Einstein metric. Results of this article can be summarized by the plot in the following page. It shows the projection of integral curves to on the $Z$-space for Case I. It is computed by MATLAB using the 4th order Runge–Kutta method. \[Lambda\] ------------------------------------------------------------------------------------------------- Integral Curves Metric Type --------------------------- --------------------------------------------------------------------- $\gamma_{0}$ $\begin{array}{l} \text{Smooth metric with vanished principal curvatures on } G/H;\\ \text{Nonsmooth metric with non-vanishing mean curvatures on } G/H \end{array}$ $\gamma_{s_1}, s_1\neq 0$ $\begin{array}{l} \text{Smooth metrics with non-zero principal curvatures on } G/H \\ \text{Nonsmooth metrics with non-vanishing mean curvatures on } G/H \end{array}$ $\Gamma$ $\begin{array}{l} \text{Conical Singularity as alternative Einstein metric on $G/K$}; \end{array}$ ------------------------------------------------------------------------------------------------- . ![image](Ricciflat.jpg){width="6.3in"} \[Fig 1\]	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this manuscript, we automate the procedure of grading of diabetic retinopathy and macular edema from fundus images using an ensemble of convolutional neural networks. The availability of limited amount of labeled data to perform supervised learning was circumvented by using transfer learning approach. The models in the ensemble were pre-trained on a large dataset comprising natural images and were later fine-tuned with the limited data for the task of choice. For an image, the ensemble of classifiers generate multiple predictions, and a max-voting based approach was utilized to attain the final grade of the anomaly in the image. For the task of grading DR, on the test data (n=56), the ensemble achieved an accuracy of 83.9%, while for the task for grading macular edema the network achieved an accuracy of 95.45% (n=44).' address: '[email protected], [email protected], [email protected], [email protected]' bibliography: - 'refs.bib' title: Ensemble of Convolutional Neural Networks for Automatic Grading of Diabetic Retinopathy and Macular Edema --- Convolutional Neural Networks, Ensemble, Transfer learning, Diabetic Retinopathy. Introduction {#sec:intro} ============ Diabetic Retinopathy (DR), one of the leading causes of blindness in humans, is a consequence of rupture of blood vessels in the eye and thereby leading to the discharge of blood and fluid to the surrounding tissues [@carrera2017automated]. Structures such as microaneurysms, hemorrhages and hard exudates are closely associated with DR and the presence of each of aforementioned anomaly determines the grade of DR in the patient. Diabetic Macular Edema (DME) is a condition that could occur at any stage of DR and is characterized by the appearance of exudates close to the macula or retinal thickening & thus affects the central vision of the patient [@al2016diabetic]. The treatment administered to subjects with DR or DME is dependent on the grade of each anomaly and thus classifying the degree of severity of DR and DME is of utmost importance. For a variety of classification and pattern recognition based tasks, Convolutional Neural Networks (CNNs) have outperformed the traditional machine learning approaches [@alexnet]. The superior performance of the CNNs comes at the cost of requiring millions of high quality labeled data for training the network. The presence of huge amount of labeled data in the domain of medical image analysis is extremely rare. In circumstances as stated above, a transfer learning based approach is utilized, wherein the model is first trained on a large dataset of natural images or digits and then fine-tuned for the task of choice using the limited dataset. The performance of CNNs are dependent on the architecture of the model and the connectivity pattern explored in the model. However, there exists no single architecture or connectivity which guarantees best or ideal performance. For the reason stated above, it is a common practice in deep learning to create an ensemble of classifiers with different architecture and connectivity patterns. This approach also helps in reducing the variance in the predictions made by the network. This manuscript explains our approach for the automatic grading of DR and DME from fundus images. For the task of DR grading, we make use of an ensemble of classifiers to differentiate between Normal and various variants of Non-proliferative diabetic retinopathy (NPDR) such as mild, moderate & severe. Additionally, an expert ensemble of classifiers were used to differentiate between Severe NPDR and proliferative diabetic retinopathy (PDR). For the task of grading DME, we make use of a one versus rest (OVR) based technique so as to mitigate the data imbalance problem in the provided dataset. Materials and Methods ===================== Data ---- The dataset [@IRID] was made available as part of the Segmentation and Grading challenge held in conjunction with the 2018 IEEE International Symposium on Biomedical Imaging. For addressing the issue of limited number of images from the “mild" NPDR, we made use of a publicly available mild NPDR database [@Additional_Data]. Pre-processing of Images ------------------------ The images were resized to dimension of 224 $\times$ 224 using bilinear interpolation and the intensity was normalized between the 0 and 1 using Eq. \[eq:norm\]. $$\label{eq:norm} I_{norm} = \frac{I- min(I)}{max(I)- min(I)}$$ Further the intensity scaled eye images were normalized to have zero mean and unit variance, Eq. \[eq:znorm\]. The images were normalized by the same statistics (mean $\mu$ and standard deviation $\sigma$) as the ones used for pre-training the models on the Imagenet database. $$\label{eq:znorm} I_{z-norm} = \frac{I_{norm}- \mu}{\sigma}$$ Convolution Neural Networks --------------------------- We make use of different variants of Residual networks (ResNet) [@he2016identity] and densely connected networks [@huang2017densely]. For the task of DR grading, 5 variants of ResNets namely ResNet-18, 34, 50, 101 & 152 and 3 variants of DenseNets, i.e., DenseNet-121, 169 & 201. For the task of DME grading, we use 2 variants of Resnets (Resnet-34 & 50) and 3 variants of Densenets (Densenet-161, 169 & 201). All the networks were pre-trained on the Imagenet dataset [@deng2009imagenet] and the networks used were made available by PyTorch [@paszke2017automatic]. Grading of Diabetic Retinopathy ------------------------------- ### Training {#training .unnumbered} For the task of classifying the severity of DR in an image, we make use of 2 models; namely the primary and expert classifiers. Both the classifiers are composed of ensemble of CNNs. For training the primary classifier, the classes Severe NPDR and PDR were unified to form a new class “S-(N)-PDR". The primary classifier classifies a fundus image as one of the 4 classes namely Normal, Mild NPDR, Moderate NPDR or S-(N)-PDR. Each model in the ensemble was trained and validated using 70% and 20% of the entire training data (n=502). The models were initialized with the pre-trained weights and the parameters of networks were optimized by reducing the cross entropy loss with ADAM [@adam] as the optimizer. The learning rate was initialized to $10^{-3}$ and the learning rate was reduced by a factor 10% every instance when the validation loss failed to drop. Each network was trained for 30 epochs and the model parameters that yielded the lowest validation loss were used for inference. The expert classifier helps in demarcating fundus images as either one with Severe NPDR or PDR. The models in the expert classifiers were trained and validated exclusively on the images with the aforementioned classes. The training regime and all hyper-parameters such as learning rate, optimizer, learning rate decay, etc. were similar to the ones used in training the primary classifier. ### Model Pruning {#model-pruning .unnumbered} Both the primary and expert classifiers are composed of 8 CNN models which varying either in terms of the depth of the network or the connectivity pattern. From the list of 8 models in the primary classifier, the model which yielded the highest number of true positives on the validation data was retained as the “benchmark model". Furthermore, models in the ensemble those produce at least 95 percentile of number of true positives generated by the “benchmark model" were also retained. A similar approach was used to prune the models from the expert classifier ensemble also. ### Testing {#testing .unnumbered} During inference, each fundus image is first resized to have a dimension of 256 $\times$ 256. From the resized images a total of 10 cropped images with a dimension of 224 $\times$ 224 were extracted. 5 cropped images were created by first cropping the image along the four corners and one along the centre of the image. Further, the image was flipped along the horizontal axis and the 5 cropped images as mentioned earlier were created to attain a total of 10 cropped images. The images are first passed through the primary classifier and optionally through the expert classifier. For a model in the pruned ensemble of primary classier, a total of 10 different variants of the test image are fed to the network. The network predicts each of the 10 variants as one of 4 classes namely “Normal", “Mild", “Moderate" or “S-(N)-PDR". The final class assigned to the test image by the model was decided by taking a maximum voting scheme. The above procedure is done for all the models in the pruned-ensemble of primary classifiers. Thus for a given test image, the primary classifier returns as many predictions as the number of models in the pruned-ensemble. On the array of predictions made, a maximum voting scheme is further applied so as to classify the test image as one of the 4 classes. If the image is classified as “S-(N)-PDR" by the primary classifier, then the 10 cropped images are presented to pruned ensemble of the expert classifier. Each model in the ensemble classifies each cropped image as “Severe NPDR" or “PDR" and a maximum voting scheme is utilized so that the model assigns a class to the image. Further, on the array of predictions made by different models in the expert classifier, a second max voting scheme is used and thereby achieve the final class assigned to the image by the expert classifier. The overall test regime is illustrated in Fig. \[ \_pipeline\]. Grading of Diabetic Macular Edema --------------------------------- ### Training {#training-1 .unnumbered} The task of DME grading is to classify each fundus image into Grade 0 (No apparent hard exudate(s)), Grade 1 (Presence of hard exudate(s) outside the radius of one disc diameter from the macula center) or Grade 2 (Presence of hard exudate(s) within the radius of one disc diameter from the macula center) depending on the presence/absence and location of hard exudate. For achieving this, we make use of 2 models based on one versus rest approach. Model 1, an ensemble of pre-trained DenseNet-161, DenseNet-169 and DenseNet-201 was trained to classify an image as class “no apparent exudates" (Grade 0) or class “presence of exudates" (Grade 1 or Grade 2). On the other hand, Model 2, an ensemble of DenseNet-161, Resnet-34 and Resnet-50 was trained to classify an image as “Grade 2" DME or not. The models were trained and validated on 70% and 20% respectively of the entire training data. The parameters of the network were initialized with the pre-trained Imagenet data weights and the parameters were learnt by minimizing cross entropy loss with ADAM as the optimizer. The learning rate of both the networks were initialized to value of $10^{-4}$ and the learning rate was annealed step-wise with step size of 10 and the multiplicative factor of learning rate decay value of 0.9. ### Testing {#testing-1 .unnumbered} ![Testing pipeline for Automated Diabetic Macular Edema grading.[]{data-label="dme_test"}](dme_test_14){width="50.00000%"} During inference, the images were fed to both Model 1 and Model 2 to check the presence of exudates in fundus image. The predictions of each model are then sent to a decision maker which gives out the final predictions following the rules as given in Table (\[Decision Maker\]). The testing pipeline used for grading DME is given in Fig. \[dme\_test\]. [c\|c\|c\|c\|]{} & -------------- Model 1 (Presence of No Exudates) -------------- : DME Decision Maker[]{data-label="Decision Maker"} & ------------------- Model 2 (Presence of Grade 2 Exudates) ------------------- : DME Decision Maker[]{data-label="Decision Maker"} & ------------ Final Prediction ------------ : DME Decision Maker[]{data-label="Decision Maker"} \ & True & False & Class 0\ & False & True & Class 2\ & False & False & Class 1\ & True & True & Class 2\ Results and Discussion ====================== Grading of DR ------------- For grading of DR, the proposed scheme was tested on a held out test data (n=56). It was observed that the technique achieved an accuracy of 84%. The confusion matrix achieved on the held out test data is shown in Table (\[my-label\]). When compared to the best model in the ensemble, ensembling predictions from multiple models in the ensemble produced a boost in accuracy by 1%. Similarly we observed that rather than using all models in the ensemble, pruning or using a subset of models in the ensemble had a positive impact in the overall performance. The model pruning explored in this work helped in achieving an improvement of 1.78% when compared to using all models in the ensemble. Providing 10 different variants of the image to the pruned ensemble during inference was observed to produce a classification accuracy of 83.6%, while removing the 10 crop step during inference led to a dip in classification accuracy of 6.82%. The performance of the classifier upon using Ten Crops, pruning, etc., is given in Table (2). The performance of the automated DR grading pipeline on the entire training data (n=502) is given in Table (\[train-label\]). For classifying images as PDR in the dataset, we observe that the expert model achieves an accuracy below the expected performance, however, the expert model helps in improving the accuracy by 14% as the ensemble of classifiers trained to demarcate all 5 classes yielded an accuracy of only 65%. \[effect1\] Method Accuracy(%) ------------------------ ------------------------- Without model Ensemble 70-74(individual model) With model Ensemble 75 Without model Pruning 75 With model Pruning 76.78 Without Ten Crops 76.78 With Ten Crops 85.7 Our method 85.7 : Effects of each step on results. \[effect\] Grading of DME -------------- On the held out test data (n=44), the proposed automated DME grading networks achieve an accuracy of 95.45 %. The confusion matrix of the proposed scheme on the test data is given in Table (\[dme-test-confusion\]). On the entire training data, the proposed scheme achieves an accuracy of 96.85%. [\|l\|l\|c\|c\|c\|c]{} &&\ & Grade 0 & Grade 1 & Grade 2\ & Grade 0 & 18 & 1 & 0\ & Grade 1 & 0 & 5 & 0\ & Grade 2 & 0 & 1 & 19\ [\|l\|l\|c\|c\|c\|c]{} &&\ & Grade 0 & Grade 1 & Grade 2\ & Grade 0 & 172 & 5 & 0\ & Grade 1 & 2 & 38 & 1\ & Grade 2 & 2 & 3 & 190\ Conclusion {#sec:pagestyle} ========== In this manuscript, we make use of ensemble of pre-trained classifiers for automating grading of DR and DME from fundus images. For the task of grading of DR, we observed that: - Ensemble of classifiers produces better performance when compared to using a single model. - Selective pruning of the ensemble aids in achieving higher accuracy when compared to using all the models in the ensemble. - Using an expert model to differentiate severe NPDR and PDR aids in improving the performance considerably. - Providing 10 variants of the test image during inference aids in attaining a classification accuracy of 83.6%. - On the entire training data (n =502, including additional data), the model achieved an accuracy of 89.4% For the task of grading of DME: - The task was accomplished by training 2 ensemble of classifiers in a one vs rest fashion. - Similar to DR grading, ensemble of classifiers produced better performance than a single model. - On the entire training data, the scheme attained an accuracy of 96.85%.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We propose a semiclassical approach to calculate multiparticle cross sections in scalar theories, which have been strongly argued to have the exponential form $\exp(\lambda^{-1}F(\lambda n,\epsilon))$ in the regime $\lambda\to0$, $\lambda n$, $\epsilon=$ fixed, where $\lambda$ is the scalar coupling, $n$ is the number of produced particles, and $\epsilon$ is the kinetic energy per final particle. The formalism is based on singular solutions to the field equation, which satisfy certain boundary and extremizing conditions. At low multiplicities and small kinetic energies per final particle we reproduce in the framework of this formalism the main perturbative results. We also obtain a lower bound on the tree–level cross section in the ultra–relativistic regime.' author: - \| D. T. Son\ [Institute for Nuclear Research of the Russian Academy of Sciences]{}\ [60th October Anniversary Prospect 7a, Moscow 117312 Russia]{}\ and\ [Department of Physics and Astronomy, Rutgers University]{}\ [Piscataway, New Jersey 08855–0849 USA]{} date: May 1995 title: Semiclassical approach for multiparticle production in scalar theories --- 6.25 in 8.5in -27pt .5in [RU–95–31]{} Introduction ============ Recently, the problem of multiparticle production in weakly coupled scalar field theories has received close attention. This problem has been initiated by the qualitative observation [@Cornwall; @Goldberg] that in the ${\lambda\over4}\phi^4$ theory at tree level the probability of processes producing large number of bosons exhibit factorial dependence on the multiplicity of the final state. This dependence originates from the large number of tree graphs contributing to the process of multiparticle production: at multiplicity $n\gg1$ the number of graphs is of order $n!$. At $n\sim1/\lambda$ this factor is sufficient to compensate the suppression due to the smallness of the coupling constant, and the tree–level multiparticle cross sections become large. Much efforts have been made to understand how this behavior is changed by loop corrections, though a conclusive result on this issue is still lacking right now. So far, quantitative calculations have been performed mostly at, or near, the multiparticle threshold, where some perturbative techniques have been developed and extensively explored. The tree amplitude of transition from one initial virtual particle to $n$ real bosons at rest (the $1\to n$ process) can be computed either by summing Feynman diagrams or by using some appropriate classical solution [@Vol; @AKP; @Brown] and the result reads $$A^{tree}_n(0)=n!\left({\lambda\over8}\right)^{n-1\over2} \label{Atree}$$ (the boson mass in this formula and further is set to 1). The same methods have been applied for calculating the amplitude beyond the threshold or beyond the tree level [@LRST]. In the first case it has been found that when the final particles are non–relativistic, the tree amplitude is an exponent of the total kinetic energy of final particles, $$A^{tree}_n(\epsilon)=A^{tree}_n(0) \e^{-{5\over6}n\epsilon} \label{intro_1}$$ where $A^{tree}_n(0)$ is given by eq.(\[Atree\]) and $\epsilon$ is the kinetic energy per particle in the final state. The exponential fall of the tree amplitude beyond threshold in eq.(\[intro\_1\]) is not sufficient, however, to make the cross section small. The second result concerns loop corrections at exact threshold and reads that the leading–$n$ contributions from each loop level (namely, the $\lambda n^2$ contribution from the first loop, $\lambda^2n^4$ from the second and $\lambda^kn^{2k}$ from the $k$–th) sum up to an exponent, so at not very large $n$, (when subleading on $n$ contributions can be neglected, presumably at $\lambda n\ll1$) the $1\to n$ amplitude at threshold has the form, $$A_n(0)=A^{tree}_n(0)\e^{B\lambda n^2} \label{intro_2}$$ where $B$ is a constant that depends on the number of spatial dimensions, $$B=\int\!{d{\bf k}\over(2\pi)^d}\,{9\over8\omega_{\bf k} (\omega_{\bf k}^2-1)(\omega_{\bf k}^2-4)},\qquad \omega_{\bf k}=\sqrt{{\bf k}^2+1}$$ In particular, in (3+1) dimensions ($d=3$), the numeric value of $B$ is $$B=-{1\over64\pi^2}(\ln(7+4\sqrt{3})-i\pi)$$ The physically interesting quantity is however not the amplitude, but the cross section, or transition rate. For the $1\to n$ process near threshold this quantity is easy to evaluate at $n\ll1/\lambda$, having on hand the two results above. In fact, if $\epsilon$ is small enough for the amplitude to be constant in the whole phase volume, the cross section is equal to $$\sigma(E,n)=\|A_n(\epsilon)\|^2V_n$$ where $A_n(\epsilon)=A^{tree}_n\exp(-{5\over6}n\epsilon+B\lambda n^2)$, and $V_n$ is the bosonic phase volume. An important point to note is that at large $n$, $V_n$ has the exponential form except from the factor of $1/n!$, $$V_n\propto{1\over n!}\exp\left({dn\over2}\left( \ln{\epsilon\over\pi d}+1\right)+{d-2\over4}n\epsilon+O(n\epsilon^2) \right) \label{Vn}$$ and now it is easy to verify that the cross section is exponential, $$\sigma(E,n)\propto\exp\left(n\ln{\lambda n\over16}-n+ {dn\over2}\left(\ln{\epsilon\over\pi d}+1\right)+ \left({d-2\over4}-{5\over3}\right)n\epsilon+2\mbox{Re}B\lambda^2n \right) \label{page3}$$ Though eq.(\[page3\]) is valid only at small $\epsilon$ and $\lambda n$, the form of $\sigma(E,n)$ strongly supports the hypothesis that in the most interesting regime $\lambda n\sim1$, $\epsilon\sim1$ the cross section is also exponential [@LRST]. $$\sigma(E,n)\propto\exp\left({1\over\lambda}F(\lambda n,\epsilon) \right),\qquad\epsilon={E-n\over n} \label{intro_exp}$$ Moreover, there are indications [@LST] that the exponent $F(\lambda n,\epsilon)$ is independent of the few–particle initial state (i.e. the cross section of $2\to n$, $3\to n$, etc. processes coincide, with exponential accuracy, with that of $1\to n$). The function $F(\lambda n,\epsilon)$ is unknown, but some terms of its expansion at small $\lambda n$ and $\epsilon$ can be found from (\[page3\]), $$F(\lambda n,\epsilon)=\lambda n\ln{\lambda n\over16}-\lambda n +{d\over2}\left(\ln{\epsilon\over\pi d}+1\right)\lambda n+ \left({d-2\over4}-{5\over3}\right)\lambda n\epsilon+ 2\mbox{Re}B\lambda^2n^2+$$ $$+O(\lambda^3n^3)+O(\lambda^2n^2\epsilon)+ O(\lambda n\epsilon^2) \label{Fpert}$$ The situation that has emerged here shows a complete analogy to the instanton–like processes at high energies. As in the latter case, the exponential form of the multiparticle cross sections is a strong argument in favor to the semiclassical calculability of the exponent $F$, but says nothing about the nature of possible calculation schemes. In this paper we propose a semiclassical method to calculate the multiparticle cross section. By using the coherent state formalism, we reduce the calculation to the problem of solving classical field equation with certain boundary conditions in the asymptotic regions $t\to\pm\infty$. The technique, as well as the boundary value problem are very similar to the those in the case of instanton transitions. The most essential difference from the latter case is that the boundary value problem for multiparticle production possesses only singular solutions. In particular, the field configuration defining the cross section is the one that is singular at one point $t={\bf x}=0$ and regular elsewhere in the Minkowskian space–time. Note that some other approaches utilizing singular solutions have been proposed recently for both multiparticle processes in scalar theories and instanton–like transitions [@VolLandau; @Khlebnikov; @CornTikt; @DiakPetr], most being inspired by the Landau procedure for calculating the semiclassical matrix elements [@Landau]. We emphasize, however, that in the framework of our formalism the singular field configuration is determined in a unique way by the boundary conditions and the structure of its singularity in the Minkowskian space–time. The field configuration with the required properties can also be found in a different setting. Namely, if one makes analytical continuation to the complex times and looks for solutions to the boundary value problem which is singular on some surface (in the simplest case the surface lies in the Euclidean space–time) and extremizes the transition rate over all possible forms of this surface , one obtains the same field configuration as one would find by solving the boundary value problem. Actually, this formalism sometimes appears to be simpler and will be applied for making quantitative calculations in this paper, which include reproducing eq.(\[page3\]) at small $\lambda n$ and $\epsilon$ and finding a lower bound on the tree–level cross section in the ultra–relativistic limit of the final state. This paper is organized as follows. In Sect.2 we derive the classical problem for the multiparticle cross sections. Sect.3 is devoted to the tree amplitude at threshold. A simplified, entirely Euclidean version of the classical problem is presented and the perturbative result for the energy dependence of the tree amplitude near threshold is reproduced. In the opposite, ultra–relativistic, limit, we obtain a lower bound on the tree cross section. In Sect. 4 we consider the amplitude (with loops) at exact threshold, where the procedure for calculating the amplitude is derived from the general formalism. We reproduce the exponentiation factor coming from leading–$n$ loop in the limit $\lambda n\ll1$. Finally, Sect. 5 contains concluding remarks. General formalism ================= The boundary value problem -------------------------- We consider the scalar field theory without symmetry breaking in $(d+1)$–dimensional space–time, $$S=\int\!d^{d+1}x\,\left({1\over2}(\partial_{\mu}\phi)^2- {1\over2}\phi^2-{\lambda\over4}\phi^4\right)$$ The quantity in interest is the total transition rate from an initial few–particle state to all possible final states having given energy $E$ and multiplicity $n$. One writes, $$\sigma(E,n)=\sum_f \|\langle f\|P_EP_n\hat{S}\hat{A}\|0\rangle\|^2 \label{sigma}$$ where $\hat{A}$ is the operator that creates the initial state from the vacuum, $\hat{S}$ is the $S$–matrix, $P_E$ and $P_n$ are the projection operators to states with energy $E$ and number of particles $n$, and the sum runs over all final states $\|f\rangle$. Different choices of the operator $\hat{A}$ corresponds to different initial states: for example, $\hat{A}=\phi$ corresponds to the $1\to n$ process (for the operator $\hat{A}$ describing the $2\to n$ process see [@LST]). We recall perturbative calculations in [@LST] indicating that $F$ does not depend on the particular choice of $\hat{A}$, providing the latter is independent of $\lambda$ parametrically. Making use of this fact, we will calculate (\[sigma\]) for the operator $\hat{A}$ most convenient for our purpose. Namely, we choose $\hat{A}$ in the following exponential form, $$\hat{A}=\e^{j\phi(0)}$$ where $j$ is some arbitrary number. Following the technique of [@RST], we derive the classical boundary value problem for $\sigma(E,n)$. Using the coherent state formalism [@KRT], one rewrites eq.(\[sigma\]) in the following integral form, $$\sigma(E,n)=\int\!db_{\bf k}^\,db_{\bf k}\,d\xi\,d\eta\, {\cal D}\phi\,{\cal D}\phi'\,\exp\left( -\int\!d{\bf k}\,b_{\bf k}^b_{\bf k}\e^{i\omega_k\xi+i\eta} +iE\xi+in\eta+\right.$$ $$+\left.B_i(0,\phi_i)+B_f(b_{\bf k}^,\phi_f)+B_i^(0,\phi'_i)+ B_f^(b_{\bf k},\phi'_f)+iS[\phi]-iS[\phi'] +j\phi(0)+j\phi'(0)\right) \label{int_repr}$$ In eq.(\[int\_repr\]), $B$’s stay for the boundary terms, $$B_i(0,\phi_i)=-{1\over2}\int\!d{\bf k}\,\omega_{\bf k}\phi_i({\bf k}) \phi_i(-{\bf k})$$ $$B_f(b^_{\bf k},\phi_f)=-{1\over2}\int\!d{\bf k}\, b_{\bf k}^b_{\bf -k}^\e^{2i\omega_kT_f}+ \int\!d{\bf k}\,\sqrt{2\omega_{\bf k}}b_{\bf k}^\phi_f({\bf k}) \e^{i\omega_kT_f}- {1\over2}\int\!d{\bf k}\,\omega_{\bf k}\phi_f({\bf k})\phi_f({\bf -k})$$ where $T_f$ is some final time moment (the limit $T_f\to+\infty$ is assumed), while $\phi_i({\bf k})$ and $\phi_f({\bf k})$ are the Fourier transformations of the field in the initial and final asymptotic regions. It is easy to see that if one takes $j$ to be of order $1/\sqrt{\lambda}$, then the integral (\[int\_repr\]) has the semiclassical nature in the limit $\lambda\to0$, $\lambda E$, $\lambda n=$ fixed, and one can expect that it is saturated by a saddle point, where $\phi$, $\phi'$, $b_{\bf k}$ and $b^_{\bf k}$ are of order $1/\sqrt{\lambda}$, and $\xi$ and $\eta$ are of order 1. In this case, the cross section certainly has the exponential form (\[intro\_exp\]). However, the assumption $j\sim1/\sqrt{\lambda}$ contradicts the requirement that $j$ does not depend on $\lambda$, which forces us to take $j$ parametrically smaller, $j\sim1$. This difficulty is obviously the consequence of the non–semiclassical nature of the initial state that contains few energetic particles, in contrast with the final state consisting of many soft ones. A method to overcome this difficulty has been suggested for the instanton transitions [@RT; @T] and can be applied for our problem in an analogous way. In this method one evaluates the integral (\[int\_repr\]) for $j=\alpha/\sqrt{\lambda}$, where $\alpha$ is some constant, in saddle–point approximation, and find the cross section in the exponential form $\sigma\sim\e^W$, where $W$ is of order $1/\lambda$ and depends on $\alpha$. After that one takes the limit $\alpha\to0$. The claim is that in this limit one reproduces the value of $W$ at $j\sim1$. Apparently, the weak point in this way of reasoning is the limit $\alpha\to0$: it is neither obvious that this limit is smooth, nor that it reproduces the amplitude with one initial particle. Moreover, the exponent of the cross section of the $1\to n$ process contains contribution from [loops]{} (the term $2\mbox{Re}B\lambda n^2$ in eq.(\[Fpert\])) that seems to be of quantum, rather than classical, nature. Nevertheless, it turns out that this nontrivial exponentiated loops are reproduced by the semiclassical calculations (see Sect.4 below), which is a sound argument in favor to the hypothesis about the $\alpha\to0$ limit. Therefore, we assume that $j\sim1/\sqrt{\lambda}$ and look for the saddle point of the exponent of the integrand in (\[int\_repr\]). The saddle–point equations can be divided into two groups. The first group contains the equations for $\phi$, $${\delta S\over\delta\phi}=ij\delta(x) \label{phi_fe}$$ $$i\dot{\phi}_i({\bf k})+\omega_{\bf k}\phi_i({\bf k})=0 \label{phi_in}$$ $$i\dot{\phi}_f({\bf k})-\omega_{\bf k}\phi_f({\bf k})+ \sqrt{2\omega_{\bf k}}b^_{\bf -k}\e^{i\omega_kT_f}=0 \label{phi_fin1}$$ $$-b_{\bf k}\e^{i\omega_ k\xi+i\eta}- b_{\bf -k}^\e^{2i\omega_kT_f}+ \sqrt{2\omega_{\bf k}}\phi_f({\bf k})\e^{i\omega_kT_f}=0 \label{phi_fin2}$$ and similar equations for $\phi'$. The equations from the second group relate the energy $E$ and the number of final particles $n$ to other parameters, $$E=\int\!d{\bf k}\,\omega_{\bf k}b^_{\bf k}b_{\bf k} \e^{i\omega_k\xi+i\eta} \label{Ebb}$$ $$n=\int\!d{\bf k}\,b^_{\bf k}b_{\bf k}\e^{i\omega_k\xi+i\eta} \label{nbb}$$ First, let us consider the equations for $\phi$. Eq.(\[phi\_fe\]) is simply the field equation with a $\delta$–like source, while eqs.(\[phi\_in\]), (\[phi\_fin1\]), (\[phi\_fin2\]) can be considered as boundary conditions in initial and final asymptotics $t\to\mp\infty$. It is convenient to rewrite these boundary conditions in a more transparent form. For this end we note that since $\phi$ is a superposition of plane waves in the limit $t\to-\infty$ and $t\to+\infty$, eq.(\[phi\_in\]) can be satisfied only if the initial asymptotics of $\phi$ is purely Feynman, $$\phi_i({\bf k})=a^_{\bf -k}\e^{i\omega_kt},\qquad t\to-\infty \label{phi_in_as}$$ where $a_{\bf k}$ are arbitrary Fourier components, while eq.(\[phi\_fin2\]) implies the following asymptotics of $\phi$ in the final asymptotic region, $$\phi_f({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( b_{\bf k}\e^{i\omega_k\xi+i\eta-i\omega_kt}+ b^_{\bf -k}\e^{i\omega_kt}\right),\qquad t\to+\infty \label{phi_fin_as}$$ It is easy to see that eq.(\[phi\_fin\_as\]) satisfies the condition (\[phi\_fin1\]) automatically. Let us turn to the equations from the second group. The physical meaning of eqs.(\[Ebb\\]) and (\[nbb\\]) is simple: they read that $E$ and $n$ are the energy and the number of particles of the field $\phi$ in its final asymptotics (\[phi\_fin\_as\]). Since $\phi$ satisfies the sourceless field equation at all values of $t$ but $t=0$ (where the source is not vanishing), the energy of the field conserves in the two regions $t>0$ and $t<0$ separately, but may have discontinuity at $t=0$. The energy in the region $t>0$ is equal to that in the limit $t\to+\infty$ and therefore is $E$. To find the energy at $t<0$ we make use of the $\phi$’s initial asymptotics, and since the latter contains only Feynman components, it vanishes. So, we find that the energy has a finite jump at $t=0$ which, naturally, is associated with the $\delta$–functional source located at $t={\bf x}=0$. To simplify further discussions, let us make two conjectures that, as we will see in what follows, do not lead to contradiction. The first is that the saddle–point values of $b_{\bf k}$ and $b^_{\bf k}$ are complex conjugated to each other. The physical meaning of this assumption is that the sum over final states in eq.(\[sigma\]) is saturated by a single coherent state. The second conjecture is that the saddle point values of $\xi$ and $\eta$ are purely imaginary, $$\xi=-iT,\qquad \eta=i\theta$$ where $T$ and $\theta$ are real (for further convenience we choose different sign conventions for $T$ and $\theta$). With the two conjectures formulated above, the field configuration describing the multiparticle process at given energy $E$ and multiplicity $n$ is the solution to the field equation with source $${\delta S\over\delta\phi}=ij\delta^{d+1}(x) \label{14}$$ with the two boundary conditions, $$\phi({\bf k})=a_{\bf k}\e^{i\omega_kt},\qquad t\to-\infty \label{14*}$$ $$\phi({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( b_{\bf k}\e^{\omega_kT-\theta-i\omega_kt}+ b_{\bf -k}^\e^{i\omega_kt}\right),\qquad t\to+\infty \label{phi_bc}$$ It is easy to notice that the initial boundary conditions (\[14\\\]) can be reformulated in Euclidean language. In fact, making the Wick rotation to the Euclidean time $\tau=-it$, eq.(\[14\\\]) reads that $\phi({\bf k})$, as a function of $\tau$, contains only the decaying component in the asymptotics $\tau\to+\infty$, $$\phi({\bf k})=a_{\bf k}\e^{-\omega_k\tau},\qquad\tau\to+\infty \label{in_as}$$ In contrast, the the final asymptotics (\[phi\_bc\]) contains both frequencies and cannot be rewritten in Euclidean language. Moreover the field in the final asymptotics is, in general, complex. The boundary value problem for $\phi'$ can be derived in analogous way, $${\delta S\over\delta\phi}=-ij\delta^{d+1}(x)$$ $$\phi'({\bf k})=a_{\bf k}\e^{-i\omega_kt},\qquad t\to-\infty$$ $$\phi'({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( b_{\bf k}\e^{-i\omega_kt}+ b_{\bf -k}^\e^{\omega_kT-\theta+i\omega_kt}\right), \qquad t\to+\infty \label{phi'_bc}$$ Notice that if $\phi$ is a solution to the boundary value problem (\[14\\], \[14\\\] \[phi\_bc\]), its complex conjugate $\phi^$ satisfies eqs.(\[phi’\_bc\]). This fact simplifies our calculations, since we need to solve only one boundary value problem instead of two. In further discussions we will assume that $\phi'=\phi^$. The saddle point of the integral (\[int\_repr\]) determines the cross section, which has the exponential form, $$\sigma(E,n)\sim\e^{W(E,n)}$$ where $$W(E,n)={1\over\lambda}F(\lambda n,\epsilon) =ET-n\theta-2\mbox{Im}S[\phi] \label{W}$$ The relation between $E$, $n$ and $T$, $\theta$, can be found either from eqs.(\[Ebb\\]) and (\[nbb\\]), or, equivalently, from $$2{\partial\mbox{Im}S\over\partial T}=E,\qquad 2{\partial\mbox{Im}S\over\partial\theta}=n. \label{dS/dT}$$ which can be easily understood if one recalls that $\xi=-iT$ and $\eta=i\theta$ are the saddle point of the integrand in eq.(\[int\_repr\]). From (\[W\]) one sees that $W(E,n)$ is the Legendre transformation of $2\mbox{Im}S(T,\theta)$, and therefore one obtains the following important relations, $${\partial W\over\partial E} = T,\qquad {\partial W\over\partial n} = -\theta \label{dW/dn}$$ Having derived the boundary value problem for calculating the transition rate at finite $j$, let us discuss the limit $j\to0$. It can be shown that in this limit the field configuration becomes singular at $x=0$. In fact, according to eq.(\[14\\]), $\phi$ has discontinuity at $t=0$, $$\delta\dot{\phi}({\bf x})= \dot{\phi}({\bf x})\|_{t=+0}-\dot{\phi}({\bf x})\|_{t=-0}= ij\delta^d({\bf x})$$ This discontinuity leads a jump of the energy at $t=0$, since the latter contains the term ${1\over2}\int\!d{\bf x}\,\dot{\phi}^2$. On the other hand, the discontinuity of the energy $E$ is supposed to be finite while that of $\dot{\phi}$ is proportional to $j$ and tends to 0. When $j$ is small one has $$E=\int\!d{\bf x}\,\dot{\phi}(0,{\bf x})\delta\dot{\phi}(0,{\bf x}) =ij\dot{\phi}(0)$$ One sees that when $E$ is fixed and $j\to0$, $\dot{\phi}(0)$ goes to infinity, which means that the field configuration becomes singular at $x=0$ in the limit of vanishing source. This is not surprising, since in the limit $j\to0$ eq.(\[phi\_fe\]) becomes the sourceless field equation, whose regular solutions conserve the energy and therefore do not obey the boundary conditions. So, to evaluate the transition rate one should find the solution to the field equation which obeys the boundary conditions and has singularity at $t={\bf x}=0$, but remains regular elsewhere in the Minkowskian space–time. For doing calculations, however, we will use another formulation of the boundary value problem. Extremization procedure ----------------------- Let us discuss the structure of singularities of our solution. Recall that in Minkowskian space–time, $\phi$ is regular everywhere except a point–like singularity at $x=0$. However, if one extrapolates $\phi$ into Euclidean times, it may occur that $\phi$ develops more singularities beside that at $x=0$. Let us consider a simple possibility that $\phi$ is singular on some $d$–dimensional surface $A$ in the Euclidean space–time, which we will parametrize either as $x_{\mu}=x_{\mu}(s_i)$, where $s_i$ are $d$ coordinates on the surface, or $\tau=\tau_0({\bf x})$. In the region near $A$, $\phi$ is inverse to the distance to the surface, $$\phi\sim\sqrt{2\over\lambda}{1\over l(x)} \label{leadsing}$$ where $l(x)$ is the distance form the point $x$ to $A$. In fact eq.(\[leadsing\]) is the only possible form of the leading singularity of $\phi$ in the region near the singularity surface $A$. In fact, one can even develop a perturbation theory on $l$ and find the correction to eq.(\[leadsing\]), which is of order $l(x)$. However, one will see soon that ambiguity begins at the order of $O(l^3)$ (on other words, the terms higher than $O(l^3)$ is not defined uniquely by the form of $A$), which reflect the fact that the solution is not defined uniquely by the surface where it is singular. If two solutions are singular on the same surface $A$, the difference between them goes to 0 when one approaches the surface as $l^3$. Note that if $A$ touches the plane $\tau=0$ only at one point $x=0$, then in Minkowskian space–time $\phi$ has the required structure, i.e. is singular only at $x=0$. Let us take an arbitrary surface $A$ satisfying the latter requirement and determine a field configuration $\phi$, which consists of two parts $\phi_1$ and $\phi_2$, as follows. Both $\phi_1$ and $\phi_2$ are supposed to be solutions to the field equation and singular on $A$, but each of them obeys one boundary condition from eqs.(\[14\\\]) and (\[phi\_bc\]). Namely, we require that $\phi_1$ decreases in the Euclidean asymptotics, $$\phi_1\to0,\qquad\tau\to+\infty$$ while $\phi_2$ obeys the boundary condition in the Minkowskian limit, $$\phi_2({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( b_{\bf k}\e^{\omega_kT-\theta-i\omega_kt}+ b_{\bf -k}^\e^{i\omega_kt}\right),\qquad t\to+\infty$$ One may imagine the field $\phi$ is defined on a particular contour on the complex time plane, which at each value of ${\bf x}$ goes along the Euclidean time axis from $i\infty$ to some $i\tau_0({\bf x})$ lying on the singularity surface and then goes back to 0 and then along the Minkowskian time axis to $\infty$ (fig.\[contour\]). The field $\phi_1$ is defined on the first part of the contour, $(i\infty,i\tau_0)$, while $\phi_2$ on the two final parts, $(i\tau_0,0)$ and $(0,\infty)$. Note that there is a region on the Euclidean time axis where both $\phi_1$ and $\phi_2$ are defined, namely $(i\tau_0,0)$. Despite the fact that $\phi$ obeys the boundary conditions of the boundary value problem, it is not the solution to the latter, since $\phi_1$ and $\phi_2$ need not to be equal at $t=0$. In other words, for a generic surface $A$, $\phi$ contains discontinuities on the whole plane $t=0$, instead of being singular at one point $t={\bf x}=0$. On the other hand, if one manages to choose the surface $A$ in such a way that $\phi_1(0,{\bf x})=\phi_2(0,{\bf x})$ for any ${\bf x}\neq0$, $\phi$ would be the solution to the boundary value problem. Let us define the Euclidean action of $\phi$ as the sum of the action of $\phi_1$ and $\phi_2$, $S[\phi_1]$ and $S[\phi_2]$, each calculated along the corresponding part of the contour, $$S_E[\phi]=-\int\limits_{+\infty}^{\tau_0({\bf x})}\!d\tau d{\bf x} \left({1\over2}(\partial_\mu\phi_1)^2+V(\phi_1)\right)- \int\limits_{\tau_0({\bf x})}^0\!d\tau d{\bf x} \left({1\over2}(\partial_\mu\phi_2)^2+V(\phi_2)\right)- i\int\limits_0^{\infty}\!dt d{\bf x}\,L(\phi_2)$$ Since $\phi_{1,2}$ are singular at $\tau=\tau_0({\bf x})$, both $S[\phi_1]$ and $S[\phi_2]$ are infinite, but since the integration contour for $\phi_1$ and $\phi_2$ goes at different directions near the singularities, one can hope that their sum $S[\phi]$ is nevertheless finite. Now we will show that the “correct” singularity surface $A$ determined by the condition that $\phi_1=\phi_2$ at $t=0$ can be found by extremizing the real part of the Euclidean action $S_E[\phi]$ with respect to all possible form of the surface $A$, with the requirement that the point $\tau={\bf x}=0$ lies on the latter. First let us regularize $\phi$ to avoid dealing with infinities in intermediate calculations. For this end we replace the condition that $\phi$ is singular on $A$ by the condition that $\phi=\phi_0$ on the same surface, where $\phi_0$ is some large, but finite number, which will eventually tends to infinity. So, we set $\phi_1=\phi_2=\phi_0$ on $A$. Since $\phi_1$ and $\phi_2$ are, in general, different in the region near the surface, we expect that the derivatives of $\phi_{1,2}$ are different on $A$. Let us denote $$\partial_n(\phi_1-\phi_2)=j(s_i) \label{page11}$$ where $\partial_n$ is the derivative along the direction normal to $A$. The configuration $\phi$, thus, can be regarded as the solution to the field equation with a source that is distributed over the surface $A$, $${\partial S_E\over\partial\phi}=j(x)=\int\!ds_i\, j(s_i)\delta(x-x_\mu(s))$$ (in this formula we assume that the appropriate metric factor has been included in $ds_i$). Now, as an intermediate step, we show that once $S_E[\phi]$ has been extremized with respect to $A$, the source $j(x)$ is proportional to $\delta(x)$ (in other words, the source is located at the point $x=0$ rather than distributed over $A$). Let us take an arbitrary singularity surface $A$ and deform it slightly to $A'$. This can be represented as shifting each point $x_\mu$ on $A$ by a small vector $\delta x_\mu=n_\mu\delta x$, where $n_\mu$ is the unit vector perpendicular to $A$, so that $x_\mu+\delta n_\mu$ lies on $A'$. To ensure that $x=0$ is always a singular point, we require that $\delta x_\mu\|_{x=0}=0$ (fig.\[shift\]). The new surface $A'$ thus corresponds to new configurations $\phi_{1,2}'=\phi_{1,2}+\delta\phi_{1,2}$. We will evaluate the variation of $S[\phi_1]$ and $S[\phi_2]$ separately. The variation of $S[\phi_1]$ is due to two factors: the first is the variation of the field, $\phi_1\to\phi_1+\delta\phi_1$ and the second is the change of the integration region. The first contribution can be reduced to a boundary integral, since $\phi_1$ is a solution to the field equation, $$-\!\!\int\limits_{\infty}^{\tau_0({\bf x})}\!\!\!d\tau d{\bf x} \delta\left({1\over2}(\partial_\mu\phi_1)^2+V(\phi_1)\right) =-\int\limits_A\!dx\, [\partial_\mu\phi_1\partial_\mu\delta\phi_1-V'(\phi_1)\delta\phi_1]= -\int\limits_{A}\!ds\,(\partial_\mu\phi_1\cdot n_\mu)\delta\phi_1$$ If $\delta x_\mu$ is small, the second contribution that is associated with the change of the integration region can be also reduced to a boundary integral, where each point on the surface $A$ is integrated with the weight $\delta x$. Therefore, the full variation of $S[\phi_1]$ is $$\delta S_E[\phi_1]=\int\limits_A\!ds\, \left[(\partial_\mu\phi\cdot n_\mu)\delta\phi -\left({1\over2}(\partial_\mu\phi_1)^2+V(\phi_1)\right)\delta x(s)\right]$$ Let us make use of the fact that $\phi_1=\phi_0$ on $A$, and $\phi_1'=\phi_0$ on $A'$. We have $$\delta\phi(x_\mu)=\phi'(x_\mu)-\phi(x_\mu)= \phi'(x_\mu)-\phi'(x_\mu+\delta x_\mu)= -(\partial_\mu\phi\cdot n_\mu)\delta x(s)$$ Therefore, $$\delta S[\phi_1]=\int\limits_A\!ds\,\left[(\partial_\mu\phi_1)^2- \left({1\over2}(\partial_\mu\phi_1)^2+V(\phi_1)\right) \right]\delta x(s) \label{deltaS1}$$ $$=\int\limits_A\!ds\,\left[{1\over2}(\partial_n\phi_1)^2-V(\phi_1)\right]$$ where we have made use of the fact that on $A$ the derivatives of $\phi$ along directions tangent to $A$ vanish (since $\phi$ is constant on $A$). Let us turn to the variation of $S[\phi_2]$. The computation is completely analogous to the case of $S[\phi_1]$, with the exception that now there is a boundary term at $t=+\infty$. So we have, $$\delta S_E[\phi_2]=-\int\!ds\, \left[{1\over2}(\partial_n\phi_2)^2-V(\phi_2)\right] -i\int\!d{\bf x}\,\partial_0\phi_2\delta\phi_2\|_{t=+\infty} \label{deltaS2}$$ However one can show that the boundary term at $t=+\infty$ is purely imaginary. In fact, since $\phi_2$ and $\phi_2'$ obey the boundary condition (\[phi\_bc\]) with the same $T$ and $\theta$, we have at $t\to\infty$, $$\delta\phi_2({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( \delta b_{\bf k}\e^{\omega_kT-\theta-i\omega_kt}+ \delta b_{\bf -k}^\e^{i\omega_kt}\right),\qquad t\to+\infty$$ Substituting this, as well as the asymptotics for $\phi_2$, to the boundary term at $t=+\infty$, we see that the latter term in the r.h.s. of eq.(\[deltaS2\]) is in fact purely imaginary $$i\int\!d{\bf x}\,\partial_0\phi_2\delta\phi_2= \int\!{d{\bf k}\over(2\pi)^d}\, \left(b_{\bf k}\delta b_{\bf k}^-b^_{\bf k}\delta b_{\bf k}\right) \e^{\omega_kT-\theta}$$ and since we are interested only in the real part of the Euclidean action, this term can be dropped it in further calculations. Taking the sum of (\[deltaS1\]) and (\[deltaS2\]), one finds, $$\delta\mbox{Re} S_E[\phi]={1\over2}\int\limits_A\!ds\,\left( (\partial_n\phi_1)^2-(\partial_n\phi_2)^2\right)\delta x(s)$$ Now when one takes $\phi_0$ to be large, both $\partial_n\phi_{1,2}$ are large but the difference between them is small. Making use of eq.(\[page11\]) one writes $$\delta\mbox{Re} S[\phi]=\int\!ds\,(\partial_n\phi)\delta x(s)j(s)$$ Now we see that the requirement that $\delta\mbox{Re}S[\phi]=0$ for all variations $A$ obeying $\delta x\|_{x=0}=0$ can be satisfied only if $j(s)$ is proportional to the delta function, $j(x)=j_0\delta(x)$. So, when we extremize the real part of the action $S_E[\phi]$, varying the surface $A$, the source $j$, which is at first distributed over $A$, gather to a localized delta–functional source $j(x)=j_0\delta(x)$ Suppose that we have performed this extremization procedure. Since now $\phi_1$ and $\phi_2$ are equal on $A$ and there normal derivatives are also equal (except from the point $x=0$), these fields coincide in the region where they are both determined, namely, at $\tau_0({\bf x})<\tau<0$. In particular, $\phi_1=\phi_2$ everywhere on the plane $\tau=0$ but $\tau={\bf x}=0$. So far we have been dealing with the regularized field configurations. Let us now take the limit $\phi_0\to\infty$. The coincidence of $\phi_1$ and $\phi_2$ at $t=0$ remains in this limit, however now the point $t={\bf x}=0$ becomes singular. The strength of the delta–functional source, $j_0$, goes to 0 in order to keep the jump of energy finite. So, $\phi$ obeys the sourceless field equation in Minkowskian space–time, and is singular only at $x=0$, thus it is the solution to the boundary value problem. To summarize, we have shown that the solution to the boundary value problem can be found by extremizing the real part of the Euclidean action $S_E[\phi]$ (or the imaginary part of the Minkowskian action) over all singularity surfaces $A$ containing the point $t={\bf x}=0$. Let us now apply this formalism to find the cross section in various limiting cases. Tree–level cross sections ========================= General consideration --------------------- Consider the exponent for the cross section, $F(\lambda n,\epsilon)$, at small $\lambda n$. Keeping in mind the formula (\[Fpert\]), one expect that at $\epsilon\sim1$, $\lambda n\ll1$, the function $F$ has the following form, $$F(\lambda n,\epsilon)=\lambda n\ln{\lambda n}-\lambda n+ \lambda nf(\epsilon)+O(\lambda^2n^2) \label{Ftree}$$ where $f(\epsilon)$ is some function of $\epsilon$. In what follows we will neglect the terms $O(\lambda^2n^2)$ and higher. Since these terms come from loops, it is equivalent to considering the tree level. Another way to see this is to compute the cross section corresponding to (\[Ftree\]), $$\sigma(E,n)\sim\exp\left({1\over\lambda}F(n\lambda,\epsilon)\right) \sim n!\lambda^n\e^{nf(\epsilon)} \label{simple}$$ We see that the cross section depends on the coupling constant $\lambda$ as $\lambda^n$, which is natural for tree diagrams whose number of vertices is $n/2$. Even at tree level, the cross section at arbitrary $\epsilon$ has not been calculated (for lower bound see ref. [@Vol_bound]). In the non–relativistic limit $\epsilon\ll1$ the perturbative result (\[Fpert\]) yields the following formula for $f(\epsilon)$, $$f(\epsilon)=-\ln16+{d\over2}\ln{\epsilon\over\pi d}+{d\over2} +\left({d-2\over4}-{5\over3}\right)\epsilon+O(\epsilon^2) \label{fepsilon}$$ In this section we make no attempt to compute $f(\epsilon)$ at arbitrary value of $\epsilon$. Our main goal is to reproduce eq.(\[fepsilon\]) from the formalism of Sect.2. We will also try to estimate the tree cross section from below in the limit $\epsilon\to\infty$. Before considering the small–$\epsilon$ limit, let us first point out that the calculation procedure can be considerably simplified if one restricts himself to the tree level. We will work in the “extremization” formalism, not in the framework of the original Minkowskian boundary value problem, so we take an arbitrary surface $A$ and calculate $S_E[\phi]$ (we will deal only with the Euclidean action, so for simplicity further we will drop the index $E$). First note that from eqs.(\[dW/dn\]) and (\[Ftree\]) one finds $$\theta=-{1\over\lambda}{\partial F\over\partial n} =-\ln{\lambda n\over16}-f(\epsilon)$$ So, in the limit $\lambda n\to0$, $\theta\gg1$ independent of the form of the function $f(\epsilon)$. We see that in the final asymptotics the Feynman part of $\phi$, $b_{\bf k}\e^{-i\omega_kt+\omega_kT-\theta}$ is much smaller than the anti–Feynman part, $b^_{\bf k}\e^{i\omega_kt}$. Recall that the initial asymptotics of $\phi_1$ is the same as that of $\phi$ eq.(\[in\_as\]). By construction, $\phi_1$ is defined on the first part of the contour of fig.\[contour\], $(i\infty,0)$, however one can always analytically continue $\phi_1$ to other parts. Let us investigate the difference between $\phi_1$ and $\phi_2$ on the parts 2 and 3. Denote $$\tilde{\phi}=\phi_2-\phi_1$$ so $\tilde{\phi}$ is regular (in fact, equal to 0) on the surface $A$ and obeys the boundary condition $$\tilde{\phi}({\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( b_{\bf k}\e^{-i\omega_kt+\omega_kT-\theta}+ (b^_{\bf -k}-a^_{\bf -k})\e^{i\omega_kt}\right)$$ in the final asymptotics. Let us show that the anti–Feynman part of $\tilde{\phi}$, $b^_{\bf -k}-a^_{\bf -k}$ is small and proportional to $\e^{-\theta}$. Supposing that this is true, then the whole $\tilde{\phi}$ is a small perturbation on $\phi_1$ and thus obeys the linearized equation $$(\partial_{\mu}^2+1+3\lambda\phi_1^2)\tilde{\phi}=0 \label{lin_eq}$$ with two boundary conditions $$\tilde{\phi}\|_{A}=0$$ $$\tilde{\phi}(k)={1\over\sqrt{2\omega_k}}\,a_{\bf k} \e^{-i\omega_kt+\omega_kT-\theta}+~ \mbox{any anti--Feynman part},\qquad t\to+\infty \label{asympt2}$$ (we have made use of the assumption that $b_{\bf k}\approx a_{\bf k}$). Since the equation is linear and the final boundary conditions contains a factor of $\e^{-\theta}$, the solution $\tilde{\phi}$ is also proportional to $\e^{-\theta}$, which is consistent with our starting assumption. Now making use of the smallness of $\tilde{\phi}$, the action, up to the contributions of order $\e^{-\theta}$, is equal to $$S[\phi]=S_1[\phi_1]+S_{\{2,3\}}[\phi_1]-\int_{\{3\}}\!dx\,\left[ (\partial_\mu\phi_1)(\partial_\mu\tilde{\phi})- V'(\phi_1)\tilde{\phi}\right] \label{S123}$$ It is easy to see that the action of $\phi_1$ on the whole contour (the first two term in the r.h.s. of eq.(\[S123\])) vanishes (one can see this, for example, by making the Wick rotation of the third part of the contour). The last term in eq.(\[S123\]) is reduces to boundary integrals. The boundary term on $A$ is equal to 0 since near $A$ $\tilde{\phi}(x)$ tends to 0 as $l^3$, where $l$ is the distance from $x$ to $A$, while $\partial\phi_1\sim l^{-2}$. The boundary term at $t=+\infty$ is $$S[\phi]=i\int\!d{\bf x}\,\tilde{\phi}\partial_t\phi_1\|_{t=+\infty} ={1\over2}\int\!d{\bf k}\,a^_{\bf k}\, a_{\bf k}\e^{\omega_kT-\theta} \label{Stree}$$ So, to compute the action, one need not really solve eq.(\[lin\_eq\]): only the knowledge of $\phi_1$ (more precisely, its Euclidean asymptotics) is required. Note that the Euclidean action is real. Once the action is found, one should extremize (\[Stree\]) over all solutions that are singular at $x=0$. In fact, one can see that one should extremize only $\int\!d{\bf k}\,a^_{\bf k}a_{\bf k}\e^{\omega_kT}$, since $\e^{-\theta}$ is just an overall factor in eq.(\[Stree\]). Moreover, since the action is bounded from below by 0, the extremum of the action is most likely the true minimum. We also expect that the action contains a classical factor $1/\lambda$, so, the extremized action has the following form $$2S(T,\theta)={1\over\lambda}\e^{g(T)-\theta} \label{g_def}$$ where $g(T)$ is some function. Eqs.(\[dS/dT\]) then read $$E=2{\partial S\over\partial T}={1\over\lambda}g'(T)\e^{g(T)-\theta} \label{Etree}$$ $$n=-2{\partial S\over\partial\theta}={1\over\lambda}\e^{g(T)-\theta} \label{ntree}$$ These equations should be solved with respect to $T$ and $\theta$. Dividing (\[Etree\]) to (\[ntree\]), one obtains $$1+\epsilon=g'(T) \label{Teps}$$ We see that the parameter $T$ depends on $\epsilon$ but not on $\theta$. Regarding (\[Teps\]) as an equation on $T$, we denote its solution as $$T=T(\epsilon)$$ and from eq.(\[ntree\]) one finds $\theta$ as a function of $\epsilon$ and $n$, $$\theta=g(T(\epsilon))-\ln(\lambda n)$$ Let us substitutes the solution to the exponent of the cross section. One obtains $${1\over\lambda}W=ET-n\theta-2S(T,\theta)= n(1+\epsilon)T(\epsilon)-n(g(T(\epsilon))-\ln\lambda n)-n$$ Comparing the last equation with (\[Ftree\]), one finds $$f(\epsilon)=(1+\epsilon)T(\epsilon)-g(T(\epsilon)) \label{fefin}$$ Therefore, the problem of finding the tree cross section at any value of $E$ and $n$ can be formulated entirely in the Euclidean space–time. One looks for all solutions $\phi_1(\tau,{\bf x})$ to the Euclidean field equations which are singular at $\tau={\bf x}=0$ and decay as $\tau\to+\infty$, and calculates for each solution the corresponding Fourier components $a_{\bf k}$ from its asymptotics at $\tau\to\infty$. Then one should maximize the integral $\int\!d{\bf k}\,a^_{\bf k}a_{\bf k}\e^{\omega_kT}$ and determines the function $g(T)$ (eq.(\[g\_def\])). The required $f(\epsilon)$ can be calculated using eq.(\[fefin\]), where $T(\epsilon)$ is the solution to eq.(\[Teps\]) Unfortunately, we are unable to carry out this program for arbitrary values of $\epsilon$ due to the non–linearity of the field equation. At small $\epsilon$ (which corresponds to non–relativistic final states) we will see that it reproduces the result found by perturbative calculations. Non–relativistic regime ----------------------- ### Leading order At small $\epsilon$, the typical momentum of final particles is $k_0=\epsilon^{1/2}$ of much smaller than the mass, so one can expect that $\phi_1$ is a slowly varying function of ${\bf x}$. The ${\bf x}$–independent solution to the field equation which is singular at $\tau=0$ and decays as $\tau\to\infty$ is known explicitly, $$\phi_1(\tau)=\sqrt{2\over\lambda}\,{1\over\sinh\tau}\, , \label{homo}$$ Let us restrict ourselves to the leading order. The field configuration in the first approximation can be obtained from (\[homo\]) by a simple modification, $$\phi_1(\tau,{\bf x})=\sqrt{2\over\lambda}\, {1\over\sinh(\tau-\tau_0({\bf x}))} \label{leading}$$ where $\tau_0({\bf x})$ is a slowly varying function of ${\bf x}$. One can check that eq.(\[leading\]) satisfies the field equation to the accuracy of $O((\partial_{\bf x}\tau_0)^2)$. Note that $\tau=\tau_0({\bf x})$ is the surface of singularities of $\phi_1$. The point $\tau={\bf x}=0$ should lie on the latter, so we require that $\tau_0(0)=0$. First, from eq.(\[leading\]) one can relate $a_{\bf k}$ to $\tau_0(x)$. At $\tau\to\infty$, the asymptotics of $\phi_1$ in eq.(\[leading\]) is $$\phi(\tau,{\bf x})=\sqrt{8\over\lambda}\, \e^{\tau_0({\bf x})}\e^{-\tau} \label{equal}$$ On the other hand, $\phi_1$ can be expanded into plane wave in this region, $$\phi_1=\int\!{d{\bf k}\over(2\pi)^{d/2}\sqrt{2\omega_k}}\,a_{\bf -k} \e^{i{\bf kx}-\omega_k\tau} \label{equal'}$$ Recalling that typical momentum ${\bf k}$ is small, as the first approximation one can replace $\omega_{\bf k}$ by 1 and the r.h.s. of eq.(\[equal’\]) behaves like $\e^{-\tau}$, as that of eq.(\[equal\]). Comparing eqs.(\[equal\]) and (\[equal’\]), one obtains, $$\int\!{d{\bf k}\over(2\pi)^{d/2}}\,a_{\bf k}\e^{i{\bf kx}}= \sqrt{16\over\lambda}\,\e^{\tau_0({\bf x})} \label{btau}$$ Taking in this equation ${\bf x}=0$, and recalling the requirement $\tau({\bf x})=0$, one finds the following constraint on $a_{\bf k}$, $$\int\!{d{\bf k}\over(2\pi)^{d/2}}\,a_{\bf k}= \sqrt{16\over\lambda} \label{constraint}$$ To find $W$ one should extremize the r.h.s of eq.(\[Stree\]) with the constraint (\[constraint\]). We follow the standard technique and introduce the term $$C\left(\int\!{d^d{\bf k}\over(2\pi)^{d/2}}\,a_{\bf k}-(2\pi)^{d/2} \sqrt{16\over\lambda}\right)$$ where $C$ is the Lagrange multiplier to the r.h.s. of eq.(\[Stree\]) and, varying with respect to $a^_{\bf k}$, we obtain $a_{\bf k}$, $$a_{\bf k}=C\e^{-\omega_kT} \label{a}$$ It is easy to notice that (\[a\]) in fact the minimizes of the action. The constant $C$ can be determined from eq.(\[constraint\]), $$C=\sqrt{16\over\lambda}T^{d/2}e^T$$ Let us now calculate the function $g(T)$. According to eq.(\[g\_def\]), $$g(T)=\ln\lambda\left(\int\!d{\bf k}\, a^_{\bf k}a_{\bf k}\e^{\omega_kT}\right) =T+{d\over2}\ln(2\pi T)+\ln16$$ so eq.(\[Teps\]) becomes $$1+{d\over2T}=1+\epsilon$$ whose solution is obviously $$T(\epsilon)={d\over2\epsilon} \label{page17}$$ Now substituting $T(\epsilon)$ and $g(T)$ into eq.(\[fefin\]) one finds finally $$f(\epsilon)={d\over2}-{d\over2}\ln{\pi d\over\epsilon}-\ln16$$ which is nothing but the leading terms in eq.(\[fepsilon\]). Before considering the $O(\epsilon)$ correction to $f(\epsilon)$, let us discuss the form of the surface of singularities $\tau=\tau_0({\bf x})$. From eqs.(\[btau\]), (\[a\]) one finds $$\tau_0({\bf x})={{\bf x}^2\over T}={2\epsilon\over d}{\bf x}^2 \label{surface}$$ One sees that the surface of singularity has the form of a paraboloid, whose curvature is proportional to the typical momentum of the final particles $k_0$. At small $\epsilon$ the curvature radius is much larger than the inverse boson mass, i.e. 1. One should keep in mind, however, that eq.(\[surface\]) is valid only for not very large values of ${\bf x}$ (not much larger than $k^{-1}$. At larger ${\bf x}$ the actual behavior of the singularity surface is unknown. ### Next–to–leading order To find the first correction to the solution, eq.(\[leading\]), let us substitute the ansatz $$\phi_1=\sqrt{2\over\lambda}{1\over\sinh(\tau-\tau_0({\bf x}))} +\tilde{\phi}_1(\tau,{\bf x})$$ to the field equation. We expect that $\tilde{\phi_1}$ is suppressed, compared with the leading order, by a factor of $k_0^2$, where $k_0$ is the typical momentum of the final particle, so we drop terms containing $\phi_1^2$ and $\phi_1^3$. We expect also that each derivative with respect to ${\bf x}$ adds an additional factor of $k_0$. Denoting $y=\tau-\tau_0$, the equation for $\tilde{\phi_1}$ is linear, $$\hat{O}\tilde{\phi}\equiv\left[\partial_\tau^2+1+ {6\over\sinh^2y}\right]\tilde{\phi_1} =\sqrt{2\over\lambda} -\left(\partial_{\bf x}^2\tau_0\cdot{\cosh y\over\sinh^2y}+ (\partial_{\bf x}\tau_0)^2\left({1\over\sinh y}+{2\over\sinh^3y} \right)\right) \label{Ophi1}$$ We have ignored, for example, $\partial_{\bf x}\tilde{\phi_1}$, since it is of order $O(k_0^4)$. The simplest way to solve eq.(\[Ophi1\]) is to make use of the following simple relations, $$\hat{O}\cosh y=-6{\cosh y\over\sinh^2y},\qquad \hat{O}\sinh y=-{6\over\sinh y}$$ $$\hat{O}{1\over\sinh y}=-{4\over\sinh^3y}$$ to see that one solution to (\[Ophi1\]) is $$\sqrt{2\over\lambda}\left( {\partial_{\bf x}^2\tau_0\over6}\cosh y+{(\partial_{\bf x}\tau_0)^2\over6} \sinh y+{(\partial_{\bf x}\tau_0)^2\over2\sinh y}\right) \label{nonhom}$$ However, one can also add to (\[nonhom\]) an arbitrary solution to the homogeneous equation $\hat{O}\phi=0$. The general solution to the homogeneous equation depends on two arbitrary functions of spatial coordinates, $F_{1,2}({\bf x})$ and has the form, $$\phi=F_1({\bf x})f_1(\tau,{\bf x})+F_2({\bf x})f_2(\tau,{\bf x})$$ where $$f_1(\tau,{\bf x})={\cosh y\over\sinh^2y}$$ $$f_2(\tau,{\bf x})=y{\cosh y\over\sinh^2y}-{\sinh y\over3}- {1\over\sinh y} \label{f2}$$ Note that $f_1$ has a double pole at $y=\tau-\tau({\bf x})=0$, while $f_2$ is regular on the singularity surface. As we expect that the field has a pole of the first order at $y=0$, so we should rule out $f_1$ and set $F_1({\bf x})=0$ [^1]. The function $F_2({\bf x})$ can be chosen from the requirement that $\phi_1$ decreases in the Euclidean asymptotics $\tau\to+\infty$. In fact, in this limit (\[nonhom\]) becomes $$\sqrt{2\over\lambda}\cdot{1\over12} \left(\partial_{\bf x}^2\tau_0+(\partial_{\bf x}\tau_0)^2 \right)\e^{\tau-\tau_0}$$ while the asymptotics of $f_2$ is also growing $$f_2(\tau,{\bf x})\sim-{1\over6}\e^{\tau-\tau_0}$$ but they cancel each other when we choose $F_2({\bf x})={1\over2}\sqrt{2\over\lambda} [\partial^2_{\bf x}\tau_0+(\partial_{\bf x}\tau_0)^2]$. Therefore, the solution to the field equation with singularities at $\tau=\tau_0({\bf x})$ and decays at $\tau\to+\infty$, up to the order of $k_0^2$, is $$\phi_1=\sqrt{2\over\lambda}\left[{1\over\sinh y}+ {\partial^2_{\bf x}\tau_0\over6}\cosh y+ {(\partial_{\bf x}\tau_0)^2\over6}\sinh y+ {(\partial_{\bf x}\tau_0)^2\over2\sinh y}+\right.$$ $$+\left.{\partial_{\bf x}^2\tau_0+(\partial_{\bf x}\tau_0)^2\over2} \left(y{\cosh y\over\sinh^2y}- {1\over3}\sinh y-{1\over\sinh y}\right)\right] \label{2order}$$ To find $a_{\bf k}$, in complete analogy with the leading order, one calculates the asymptotics of $\phi_1$ from eq.(\[2order\]), $$\phi_1(\tau,{\bf x})=\sqrt{2\over\lambda}\left[\left(2-{5\over6} \partial_{\bf x}^2\tau_0\right)\e^{-y}+ \left(\partial_{\bf x}^2\tau_0+(\partial_{\bf x}\tau_0)^2\right) y\e^{-y}\right] \label{2asympt}$$ Let us note that, in contrast with the leading order, there is a term proportional to $y\e^{-y}$ in the asymptotics of $\phi_1$. This structure also emerges when one evaluate the r.h.s. of eq.(\[equal’\]) to the next–to–leading order, replacing $\e^{\omega_k\tau}$ by $\e^{-\tau}(1-{{\bf k}^2\over2}\tau)$. Introducing the function $$a({\bf x})=\int\!{d{\bf k}\over(2\pi)^{d/2}\sqrt{2\omega_{\bf k}}} \,a_{\bf k}\e^{i{\bf kx}}$$ eq.(\[equal’\]) reads $$\phi_1(\tau,{\bf x})=\int\!{d{\bf k}\over(2\pi)^{d/2} \sqrt{2\omega_{\bf k}}}\left(1-{{\bf k}^2\over2}\right)a_{\bf k} \e^{-\tau+i{\bf kx}}=\left(a({\bf x})+ {\tau\over2}\partial_{\bf x}^2a({\bf x})\right)\e^{-\tau} \label{2asympt'}$$ Comparing eqs.(\[2asympt\]) and (\[2asympt’\]), one finds the relation between $a({\bf x})$ and $\tau(x)$, the next–to–leading–order version of eq.(\[btau\]), $$a({\bf x})=\sqrt{2\over\lambda}\left(2-{5\over6}\partial_{\bf x}^2 \tau_0-\tau_0(\partial_{\bf x}^2\tau_0+(\partial_{\bf x}\tau_0)^2) \right)\e^{\tau_0({\bf x})} \label{ax}$$ Let us find out the constraint on $a({\bf x})$ comes from the requirement that $\tau_0(0)=0$. Taking ${\bf x}=0$ and noticing that $\partial_{\bf x}\tau_0\|_{{\bf x}=0}=0$ since the singularity surface is tangent to the plane $\tau=0$ at ${\bf x}=0$, it is easy to show that this constraint is $$a({\bf x})+{5\over12}\partial_{\bf x}^2a({\bf x})\|_{{\bf x}=0}= \sqrt{8\over\lambda}$$ In momentum representation this relation reads, $$\int\!{d{\bf k}\over(2\pi)^{d/2}\sqrt{2\omega_{\bf k}}} \left(1-{5\over12}{\bf k}^2\right)a_{\bf k}=\sqrt{8\over\lambda} \label{amom}$$ The difference of the constraint from that of the leading order is there is a small correction ${5\over12}{\bf k}^2$. The calculation is now straightforward and similar to that of the leading order. Introducing the Lagrange multiplier and taking variation with respect to $a_{\bf k}$, one finds $$a_{\bf k}={C\over\sqrt{2\omega_{\bf k}}}\left(1-{5\over12}{\bf k}^2 \right)$$ where $C$ is a constant that can be found from (\[amom\]). The final result of the calculations is $$f(\epsilon)= {d\over2}\left(\ln{\epsilon\over\pi d}+1\right)+\ln16 +\left({d-2\over4}-{5\over3}\right)\cdot\epsilon$$ which, as expected, coincides with the perturbative result for the tree cross section near threshold. One can, in principle proceed further in this direction and calculate more terms in the expansion of $f$ on the small parameter $\epsilon$ by the same technique. In this way one could find $O(\epsilon^2)$ and higher corrections to eq.(\[fepsilon\]), which have not been found by standard perturbative techniques. However, let us stop here and turn to the opposite limit $\epsilon\gg1$. Ultra–relativistic limit in (3+1) dimensions. Lower bound on tree cross section ------------------------------------------------------------------------------- In the ultra–relativistic limit $\epsilon\to\infty$, the boson mass can be neglected. We will consider the most interesting case $d=3$, where $\lambda$ is dimensionless. In this case the theory becomes scale invariant and the exponent of the cross section should be independent of the energy (the dependence of the pre–exponent on the energy can be determined solely by dimensional analysis). In other words, we expect that the function $F(\lambda n,\epsilon)$ becomes independent of $\epsilon$ at large $\epsilon$, $$\lim_{\epsilon\to\infty}F(\lambda n,\epsilon)=F(\lambda n)$$ At tree level, this statement means that $f(\epsilon)$ has a limit when $\epsilon\to\infty$. From the first identity in eq.(\[dW/dn\]) one sees that $T$ should tends to 0 faster than $1/\epsilon$ (otherwise $W=\int\!dE\,T(E)$ diverges in the ultraviolet). Now let us take the limit $\epsilon\to\infty$ in eq.(\[fefin\]). One finds, $$f(\infty)=-g(0)$$ Recall that $\lambda^{-1}\e^{g(0)}$ is the minimal value of $\int\!d{\bf k}\,a_{\bf k}^a_{\bf k}$ over all solutions of the massless theory that are singular at $\tau={\bf x}=0$. We are unable to find $g(0)$. Instead, we will try to bound $g(0)$ from above. We will take a solution that is known analytically, calculate for it $a_{\bf k}$ and plug the result into the definition of $g(0)$. If the extremum over $a_{\bf k}$ is the true minimum (recall our discussion on the natrue of the extremum in subsection 3.1), this procedure would gives us an upper bound on $g(0)$. We take the following trial singular solution to the massless field equation [@Khlebnikov], $$\phi(\tau,{\bf x})=\sqrt{8\over\lambda}\,{\rho\over {\bf x}^2+(\tau-\rho)^2-\rho^2}$$ The Fourier components of this configuration is $$a_{\bf k}=\sqrt{8\over\lambda}\sqrt{\pi\over\omega_{\bf k}}\rho \e^{-\omega_k\rho}$$ which implies the following bound on $g(0)$, $$g(0)<\ln(8\pi^2)$$ The corresponding lower bound on the cross section is $$\sigma_n\geq n!\left({\lambda\over 8\pi^2}\right)^n$$ Note that this lower bound grows factorially as $n\to\infty$, which reflects the fact that at tree level not only amplitudes, but also cross sections becomes large at $n\sim\lambda^{-1}$. Loop corrections at threshold ============================= General consideration --------------------- In this Section we consider another limiting case. Namely, we take arbitrary $\lambda n$ but small $\epsilon\ll1$, and will be interested in the exponent of the cross section to the leading order of $\epsilon$. In this limit, we expect from eq.(\[Fpert\]) that the result has the following form, $$F(\lambda n,\epsilon)=\lambda n\ln{\lambda n\over16}-\lambda n+ {d\over2}\left(\ln{\epsilon\over\pi d}+1\right)\lambda n +g(\lambda n) \label{gdef}$$ where the first term in the expansion of $g(\lambda n)$ at small $\lambda n$ should be $2\mbox{Re}B\lambda^2n^2$. The meaning of the function $g(\lambda n)$ becomes clear when one recalls that in the limit $\epsilon\to0$ the cross section is the product of the square of the threshold amplitude and the phase volume. Dividing the cross section corresponding to eq.(\[gdef\]) to the phase volume at small $\epsilon$, one finds the absolute value of the threshold amplitude, $$\|A_n\|\propto n!\left({\lambda\over8}\right)^{n/2} \e^{{1\over\lambda}g(\lambda n)}$$ Comparing with eq.(\[Atree\]), one finds that $g$ is nothing but the contributions of the loops to the amplitude at threshold. Note that the reproduction of of even the leading term in $g$, $2\mbox{Re}B\lambda^2n^2$, is a nontrivial argument in favor to the validity of the whole semiclassical ideology, since it shows that the exponentiated part of loop contributions is semiclassically calculable. To find the form of the singularity surface, let us begin by recalling the result of Sect.3 that $\epsilon\to0$ corresponds to $T\to\infty$ (eq.(\[page17\])), which means that the surface of singularities has a very large curvature radius and in the limit $\epsilon\to0$ can be considered as a plane. However, the discussions in Sect.3 is based on the assumption that the limit $\lambda n\to0$ is taken. When one drops this assumption, one could expect that the presence of the source at $x=0$ deforms the surface of singularities near its location. This change should be local and the curvature of the singularity surface should tend to 0 at large ${\bf x}$. At finite $\lambda n$, one expects that the surface of singularities $\tau=\tau_0({\bf x})$ has the form similar to that sketched in fig.\[fig1\]. The requirement of zero curvature at large distances from ${\bf x}=0$ can be satisfied if $\tau_0({\bf x})$ tends to some constant $\tau_\infty$ as ${\bf x}\to\infty$. If the singularity surface is just the plane $\tau=\tau_\infty$, the solution would be ${\bf x}$–independent and equal to $$\phi=\sqrt{2\over\lambda}{1\over\sinh(\tau-\tau_\infty)}$$ In the case when the singularity surface has the form shown in fig.\[fig1\], the general form of $\phi$ is $$\phi=\sqrt{2\over\lambda}\,{1\over\sinh(\tau-\tau_{\infty})}+ \tilde{\phi}(\tau,{\bf x}) \label{tildephi}$$ where $\tilde{\phi}(\tau,{\bf x})$ vanishes at ${\bf x}\to\infty$ and decays into plane waves in the asymptotics regions, since it comes from the local (about ${\bf x}=0$) deviation of the singularity surface from the plane $\tau=\tau_\infty$. Let us consider the asymptotics of $\phi$ at $t\to+\infty$. Strictly speaking, the ${\bf x}$–independent part of $\phi$ remains non–linear at any $t$ and does not decays into plane waves. This is an artifact of the limit $T\to\infty$: at any finite $T$ the field $\phi$ is a linear superposition of waves at sufficiently large $t$. To make $\phi$ linear, let us take the final part of the contour to form at a small angle with the Minkowskian time axis. Namely, we take $t=(1+i\delta)\alpha$, where $\delta$ is small and $\alpha$ is a real parameter going to infinity. This leads to the replacement $${1\over\sinh(\tau-\tau_\infty)}\to2\e^{(\tau-\tau_\infty)}$$ in the asymptotic region. This part of $\phi$ is homogeneous on ${\bf x}$, so its Fourier transformation is a delta–function in the momentum space. In contrast, the Fourier components of $\tilde{\phi}$ in the asymptotics $t\to+\infty$ are supposed to be well behaved functions of ${\bf k}$. Denoting $f_{\bf k}$ and $g_{\bf k}$ to be the Feynman and anti–Feynman components of $\tilde{\phi}$ at the final asymptotics, $$\tilde{\phi}(t,{\bf k})={1\over\sqrt{2\omega_{\bf k}}}\left( f_{\bf k}\e^{-i\omega_kt}+g_{\bf -k}\e^{i\omega_kt}\right)$$ one finds that the asymptotics of $\phi$ is $$\phi(t,{\bf k})={1\over\sqrt{2\omega_{\bf k}}} \left[f_{\bf k}\e^{-i\omega_kt}+\left( (2\pi)^{d/2}\sqrt{16\over\lambda}\delta({\bf k})\e^{\tau_\infty} +g_{\bf -k}\e^{i\omega_kt}\right)\right] \label{fg}$$ Comparing eqs.(\[phi\_bc\]) and (\[fg\]), one finds, $$b_{\bf k}\e^{\omega_kT-\theta}=f_{\bf k} \label{bf}$$ $$b_{\bf k}^=(2\pi)^{d/2}\sqrt{16\over\lambda}\delta({\bf k}) \e^{\tau_\infty}+g_{\bf k} \label{bg}$$ From eq.(\[bf\]) one obtains $b_{\bf k}=f_{\bf k}\e^{-\omega_kT+\theta}$. Since $T$ is large, $b_{\bf k}$ looks like a delta–function in the ${\bf k}$–space: $b_{\bf k}$ is non–zero only in a small region near ${\bf k}=0$, namely, $k\sim T^{-1/2}$. As $f_{\bf k}$ is a smooth function of ${\bf k}$, one can replace $f_{\bf k}$ by its value at ${\bf k}=0$ and write $$b_{\bf k}=f_0\e^{-\omega_kT+\theta}$$ Consider now eq.(\[bg\]). Since $b^_{\bf k}$ has the form of the delta–function, $g_{\bf k}$ vanishes. This important result means that $\tilde{\phi}(t,{\bf x})$ is purely Feynman in the final asymptotics. Eq.(\[bg\]) now reads $$f_0^\e^{-\omega_kT+\theta}=(2\pi)^{d/2}\sqrt{16\over\lambda} \delta({\bf k})\e^{\tau_\infty} \label{f0T}$$ This equation should be understood as a symbolic representation of a relation between $f_0$, $T$, $\theta$ and $\tau_\infty$. To find the latter let us take the integral of both sides of eq.(\[f0T\]) over $d{\bf k}$. We obtain $$f_0^=\sqrt{16\over\lambda}\,T^{d/2}\e^{T-\theta+\tau_{\infty}}$$ The multiplicity and the (small) kinetic energy of the final state can be also expressed in terms of $T$, $\theta$ and $\tau_\infty$, $$n=\int\!d{\bf k}\,b^_{\bf k}b_{\bf k}\e^{\omega_kT-\theta} ={16\over\lambda}(2\pi T)^{d/2}\e^{T-\theta+2\tau_{\infty}} \label{n}$$ $$n\epsilon=\int\!d{\bf k}\,{{\bf k}^2\over2}b^_{\bf k}b_{\bf k} \e^{\omega_kT-\theta}= {16\over\lambda}(2\pi T)^{d/2}\e^{T-\theta+2\tau_{\infty}} \cdot{d\over2T} \label{E}$$ Solving eqs.(\[n\]) and (\[E\]) with respect to $T$ and $\theta$, one finds, $$T={d\over2\epsilon}$$ $$\theta=-\ln{\lambda n\over16}+{d\over2}\left(\ln{\pi d\over\epsilon} +1\right)+2\tau_{\infty}$$ Note that $T$ and $\theta$ depend on the surface of singularities through $\tau_\infty$ but are insensitive of the behavior of the function $\tau_0({\bf x})$ at finite ${\bf x}$. In contrast, the action depends on the the precise form of this surface. In Sect.3 we have seen that a part of the action (that is linear on $\tilde{\phi}$) can be reduced to the boundary term $\tilde{\phi}\partial_t\phi\|_{t=\infty}$ and is equal to ${in\over2}$. It is convenient to separate this contribution from the action and introduce the notation $S'=\mbox{Im}S-{n\over2}$. The exponent of the cross section is now equal to $$W=ET-n\theta-2\mbox{Im}S= {dn\over2}\left(ln{\pi d\over\epsilon}+1\right)+ n\ln{\lambda n\over16}-n-2S'[\tau_0({\bf x})]-2n\tau_{\infty} \label{WS'}$$ where $S'$ a functional depending on $\tau_0({\bf x})$. Notice that the first three terms in eq.(\[WS’\]) coincide with those in eq.(\[gdef\]), we conclude that the exponent for the loop corrections to the amplitude at threshold can be obtained by maximizing the expression $${1\over\lambda}g(\lambda n)=-n\tau_\infty-S'[\tau_0({\bf x})] \label{gS'}$$ with respect to all possible surfaces of singularities $\tau_0({\bf x})$. The problem has a simple geometric interpretation: it is equivalent to finding the equilibrium configuration of a surface under the force $n$ acting to the point ${\bf x}=0$, when the energy of the surface depends on its form through the functional $S'[\tau_0({\bf x})]$. Let us summarize our results. To find the amplitude at threshold, one should - Find for any function $\tau=\tau_0({\bf x})$ (which goes to a constant value $\tau_\infty$ as ${\bf x}\to\infty$) the solution to the field equation which is singular on the surface $\tau=\tau_0({\bf x})$ and decays at $\tau\to+\infty$ and at $t=+\infty$ has the form (\[tildephi\]) where $\tilde{\phi}$ is purely Feynman. - calculate the action of the field, and find $S'=\mbox{Im}S-{n\over2}$ - Extremize the r.h.s. of eq.(\[gS’\]) over all surfaces $\tau=\tau_0({\bf x})$. Unfortunately, $S'$ seems to be a very complicated functional of $\tau_0({\bf x})$. One region when it can be calculated is the regime $\tau_0\ll1$, where one can develop perturbation theory on $\tau_0$. Let us see that this is the regime of low multiplicities, $n\ll1/\lambda$. Low multiplicities, $\lambda n\ll1$ ----------------------------------- When $\tau_0({\bf x})$ is small the field configuration can be found by solving the field equation separately at $\tau\ll1$ and $\tau\gg\tau_{\infty}$ and matching in the intermediate region $\tau_0\ll\tau\ll1$. For convenience we will deform the contour as shown in fig.\[fig2\], where $\tau_c$ is some value in the intermediate region, $\tau_{\infty}\ll\tau_c\ll1$. So, the contour in the $t$ plane consists of four parts: (I) $(i\infty,i\tau_c)$, (II) $(i\tau_c,0)$, (III) $(0,i\tau_c)$, (IV) $(i\tau_c,i\tau_c+\infty)$. The energy on parts (I) and (II) is 0, while on parts (III) and (IV) of the contour it is equal to $E$. First consider parts (II) and (III). In these part $\tau\ll1$, and the field is given by the following formula, $$\phi(\tau,{\bf x})=\sqrt{2\over\lambda}\,{1\over\tau-\tau_0({\bf x})}+\mbox{small corrections} \label{smalltau}$$ Eq.(\[smalltau\]) is valid on both parts (II) and (III). While the leading singular terms coincide, there is difference between $\phi$ on (II) and (III). Denote $$\delta\phi(\tau,{\bf x})=\phi_{III}(\tau,{\bf x})- \phi_{II}(\tau,{\bf x})$$ Since we expect that this difference is much smaller than the leading term in eq.(\[smalltau\]), $\delta\phi$ satisfies the linearized equation $$(\partial_{\tau}^2+\partial_{\bf x}^2+1+3\lambda\phi_0^2) \delta\phi=0$$ Furthermore, at $\tau\ll1$, the mass and the spatial derivatives can be neglected, so $\delta\phi$ satisfies the equation $(\partial_{\tau}^2+3\lambda\phi_0^2)\delta\phi=0$, the general solution to which that vanishes on the singularity surface is $$\delta\phi(\tau,{\bf x})=\sqrt{2\over\lambda}\, W({\bf x})(\tau-\tau_0({\bf x}))^3 \label{deltaphi}$$ where the function $W({\bf x})$ is yet to be determined. To the linear order on $\delta\phi$, the action on the parts (II) and (III) can be reduced to the boundary term at $\tau=\tau_c$, $$S'_{II+III}=-\int\!d{\bf x}\, \partial_\tau\phi\cdot\delta\phi\|_{\tau=\tau_c}= {2\over\lambda}\left((\tau_c-\tau_\infty) \int\!d{\bf x}\,W({\bf x})- \int\!d{\bf x}\,W({\bf x})c({\bf x})\right) \label{Ssmalltau}$$ where $c({\bf x})=\tau_0({\bf x})-\tau_\infty$. The next correction to $S'_{II+III}$ is suppressed by a factor of $\tau_\infty^2$ compared to the leading result (\[Ssmalltau\]) and will be neglected. Consider now the parts (I) and (IV). On these parts of the contours, the solutions to the field equation can be represented in the form of the perturbative series on the background of the homogeneous field, $$\phi=\phi_{(0)}+\phi_{(1)}+\phi_{(2)}+\cdots \label{largetau}$$ where $\phi_{(0)}$ is the ${\bf x}$–independent solution to the field equation, $$\phi_0=\sqrt{2\over\lambda}\,{1\over\sinh(\tau-\tau_\infty)} \label{phi0}$$ $\phi_{(1)}$ is the linear wave on the background (\[phi0\]) satisfying the equation $$\partial_\mu^2\phi_{(1)}-V'(\phi_{(0)})\phi_{(1)}=0 \label{phi(1)}$$ everywhere except the singularities and having the following explicit form, $$\phi_{(1)}=\sqrt{2\over\lambda} \int\!{d{\bf k}\over(2\pi)^d}\,{1\over3}c_{\bf k}f^k_2(\tau) \e^{i{\bf kx}},\qquad\mbox{on part (I)} \label{phi11}$$ $$\phi_{(1)}=\sqrt{2\over\lambda} \int\!{d{\bf k}\over(2\pi)^d}\,{1\over3}c_{\bf k}f^k_1(\tau) \e^{i{\bf kx}},\qquad\mbox{on part (IV)} \label{phi14}$$ where $c_{\bf k}$ are some function of ${\bf k}$, and $f_{1,2}^k(\tau)$ are mode functions on the background field (\[phi0\]) [@Volprop; @AKPprop], $$f^k_1(\tau)=\e^{\omega_k(\tau-\tau_\infty)}\left(\omega_{\bf k}^2- {3\omega_{\bf k}\over\tanh(\tau-\tau_\infty)}+2+ {3\over\sinh^2(\tau-\tau_\infty)}\right)$$ $$f^k_2(\tau)=f^k_1(2\tau_\infty-\tau)$$ and $\phi_{(2)}$, etc. are higher corrections. The small parameter governing the expansion (\[largetau\]) is in fact $\tau_\infty$. The relation between $c_{\bf k}$ and the form of the surface of singularities can be found from the matching condition between eqs.(\[largetau\]) and (\[smalltau\]) at intermediate values of $\tau$. One finds, $$c_{\bf k}=\int\!d{\bf x}\,c({\bf x})\e^{i{\bf kx}}$$ One can also relate $c_{\bf k}$ with the function $W({\bf x})$. In fact, at $\tau=\tau_c\ll1$ one finds from eqs.(\[phi11\]) and (\[phi14\]) $$\delta\phi_{(1)}=\sqrt{2\over\lambda}{(\tau_c-\tau_\infty)^3\over45} \int\!{d{\bf k}\over(2\pi)^d}\,W_{\bf k}c_{\bf k}\e^{i{\bf kx}} \label{deltaphi(1)}$$ where $$W_{\bf k}=2\omega_{\bf k}(\omega^2_{\bf k}-1)(\omega^2_{\bf k}-4). \label{Wk}$$ On the other hand, since $\tau_c\gg\tau_0$ eq.(\[deltaphi\]) can be expanded on the parameter $\tau_0/\tau_c$ as follows, $$\delta\phi=\sqrt{2\over\lambda}W({\bf x})(\tau_c-\tau_\infty)^3 -3W({\bf x})c({\bf x})(\tau_c-\tau_\infty)^2+\cdots \label{deltaphi(1)'}$$ The r.h.s. of eq.(\[deltaphi(1)\]) should coincide with the first term of the r.h.s. of (\[deltaphi(1)’\]), so we obtain the following relation, $$W({\bf x})={1\over45} \int\!{d{\bf k}\over(2\pi)^d}\,W_{\bf k}c_{\bf k}\e^{i{\bf kx}}$$ Let us now evaluate the action on the part (I). Up to the second order of the small parameter the action has the form $$S_I'=S_I[\phi_0]+\int\limits_{\tau_c}^\infty\!\left( \partial_\mu\phi_{(0)}\partial_\mu\phi_{(1)}+ \partial_\mu\phi_{(0)}\partial_\mu\phi_{(2)}+ {1\over2}(\partial_\mu\phi_{(1)})^2+ V'(\phi_{(0)})(\phi_{(1)}+\phi_{(2)})+\right.$$ $$\left.{1\over2}V''(\phi_{0})(\phi_{(1)})^2\right) \label{SI'}$$ Making use of eq.(\[phi(1)\]) and the fact that $\phi_{(0)}$ satisfies the field equation, the integral in eq.(\[SI’\]) can be taken in part and the result is $$S_I'=S_I[\phi_0]-\int\!d{\bf x}\left( (\phi_{(1)}+\phi_{(2)})\partial_\tau\phi_{(0)}+ {1\over2}\phi_{(1)}\partial_\tau\phi_{(1)} \right)\|_{\tau=\tau_c}$$ The action on the part (IV) can be calculated in an analogous way. Since $S_I[\phi_0]+S_{IV}[\phi_0]=0$, the sum of the action on parts (I) and (IV) is given by the following boundary terms, $$S'_{I+IV}=\int\!{d{\bf x}}\, \left(\partial_\tau\phi_{(0)}\cdot\delta\phi_{(1)}+ \partial_\tau\phi_{(0)}\cdot\delta\phi_{(2)}+ {1\over2}\partial_\tau\phi_{(1)}\cdot\delta\phi_{(1)}+ {1\over2}\partial_\tau\delta\phi_{(1)}\cdot\phi_{(1)}\right)_{\tau=\tau_c} \label{S14}$$ The analytical expressions for $\phi_2$ are rather complicated. Fortunately, the calculation of the action requires only $\delta\phi_2$ at $\tau=\tau_c$, which is equal to the second term in the r.h.s. of eq.(\[deltaphi(1)’\]), $$\delta\phi_2(\tau_c,{\bf x})=-3W({\bf x})c({\bf x})(\tau_c-\tau_\infty)^2$$ Substitute this to eq.(\[S14\]) one obtains $$S'_{I+IV}={2\over\lambda}\left(-(\tau_c-\tau_\infty) \int\!d{\bf x}\,W({\bf x})+ {7\over2}\int\!d{\bf x}\,W({\bf x})c({\bf x})\right) \label{Slargetau}$$ The full action can be obtained by taking the sum of (\[Slargetau\]) and (\[Ssmalltau\]). The dependence on $\tau_c$ disappears, as one can anticipate, and the action is quadratic on $c_{\bf k}$, $$S'={5\over\lambda}\int\!d{\bf x}\,W({\bf x})c({\bf x})= {1\over9\lambda}\int\!{d{\bf k}\over(2\pi)^d}\, W_{\bf k}\|c_{\bf k}\|^2 \label{56}$$ Now to extremize (\[gS’\]) we note that $$\tau_\infty=-c(0)=\int\!{d{\bf k}\over(2\pi)^d}\,c_{\bf k} \label{}$$ so $${1\over\lambda}g=n\int\!{d{\bf k}\over(2\pi)^d}\,c_{\bf k}- {1\over9\lambda}\int\!{d{\bf k}\over(2\pi)^d}\,W_{\bf k} \|c_{\bf k}\|^2 \label{*}$$ Differentiating $g$ with respect to $c_{\bf k}$, one obtains $$c_{\bf k}={9\lambda\over2}{n\over W_{\bf k}} \label{'}$$ and $${1\over\lambda}g={9\lambda\over4}n^2\int\!{d{\bf k}\over(2\pi)^d}\, {1\over W_{\bf k}}=B\lambda n^2$$ which coincides with the perturbative result for exponentiated leading–$n$ loops. Let us find the condition for the approximation we use here to be valid. The small parameter is $\tau_\infty$, and from eqs.(\[\\\]) and (\[\’\]) we see that $\tau_\infty\sim\lambda n$. So, our calculations are reliable when $n$ is small, $n\ll1/\lambda$. Another remark should be made on the nature of the extremum. Since $W_{\bf k}$ does not have definite sign (it is negative at ${\bf k}^2<3$ and positive at ${\bf k}^2>3$, see eq.(\[Wk\])), the extremum of the r.h.s. of (\[\\\\]) is neither maximum nor minimum, but rather a saddle point (in contrast with the maximum for tree–level cross section considered in sect.3). In the theory with broken symmetry $W_{\bf k}$ is a positively defined function [@Smith], and we have in this case the true minimum. Conclusion ========== We have seen that the problem of calculation multiparticle cross sections can be reduced to a certain problem of the classical theory. The field configuration describing these processes is a singular solution to the field equation with certain boundary conditions that optimizes the transition rate. Though most results obtained in this paper can also be found by making use of various perturbative methods, our discussions show that they can be derived from a single approach. We have pointed out an important fact that the exponentiated loop corrections can be calculated semiclassically. We also obtain a new result, namely, the lower bound on tree cross section in the ultra–relativistic regime. Hopefully, the formalism developed in this paper can be used in further analytical or numerical calculations of the multiparticle cross section The author would like to thank V.A. Rubakov and P.G. Tinyakov for numerous stimulating discussions and valuable comments. The author is grateful to L. McLerran and M.B. Voloshin for discussions of the results, and thanks Rutgers University and Theoretical Physics Institute, University of Minnesota, for hospitality. This work is supported, in part, by Russian Foundation for Fundamental Research, grant \# 93-02-3812 and INTAS grant \# INTAS-93-1630. [99]{} J.M. Cornwall, Phys. Lett. [B243]{}, 271 (1990). H. Goldberg, Phys. Lett. [B246]{}, 445 (1990). M. B. Voloshin, Nucl. Phys. [B383]{}, 233 (1992). E. N. Argyres, R. H. P. Kleiss and C. G. Papadopoulos, Nucl. Phys. [B391]{}, 42 (1993). L. S. Brown, Phys. Rev. [D46]{}, 4125 (1992). M.V. Libanov, V.A. Rubakov, D.T. Son and S.V. Troitsky, Phys. Rev. [D50]{}, 7553 (1994) M.V. Libanov, D.T. Son and S.V. Troitsky, [Exponentiation of multiparticle amplitudes in scalar theories. II. Universality of the exponent]{}, preprint RU–95–19, 1994, hep–ph/9503412. M.B. Voloshin, Phys. Rev. [D43]{}, 1726 (1991). S.Yu.D Khlebnikov, Phys. Lett. [B282]{}, 459 (1992). J.M. Cornwall and G. Tiktopoulos, Phys. Rev. [D47]{}, 1629 (1993). D. Diakonov and V. Petrov, Phys. Rev. [D50]{}, 266 (1994). L.D. Landau, Phys. Zs. Sowiet. [1]{}, 88 (1932).\ L.D. Landau and E.M. Lifshits, Quantum mechanics, Non-Relativistic Theory, Third edition, Pergamon Press, 1977; section 52. V.A. Rubakov, D.T. Son and P.G. Tinyakov, Phys. Lett. [B287]{}, 342 (1992). S.Yu. Khlebnikov, V.A. Rubakov, P.G. Tinyakov, Nucl. Phys. [B350]{}, 441 (1991). V.A. Rubakov and P.G. Tinyakov, Phys. Lett. [B279]{}, 165 (1992). P. G. Tinyakov, Phys. Lett. [B284]{}, 410 (1992). S.Yu. Khlebnikov, Nucl. Phys. [B436]{}, 428 (1995). M.B. Voloshin, Phys. Lett. [B293]{}, 389 (1992). M.B. Voloshin, Phys. Rev. [D47]{}, 357 (1993). E.N. Argyres, R.H.P. Kleiss, C.G. Papadopoulos, Phys. Lett. [B308]{}, 292 (1993). B.H. Smith, Phys. Rev. [D47]{}, 3518 (1993). (4,3) (0,1)[(1,0)[4]{}]{} (0,0)[(0,1)[3]{}]{} (0,0)[(0,1)[2]{}]{} (0,0)[(1,0)[.15]{}]{} (.15,0)[(0,1)[1]{}]{} (.15,1)[(1,0)[2]{}]{} (0.1,2.9)[$\tau$]{} (3.9,0.7)[$t$]{} (-.3,0.9)[0]{} (-.3,-0.1)[$\tau_0$]{} (6,2)(0,1) (3,1)[(0,1)[2.5]{}]{} (3.1,3.4)[$\tau$]{} (5.9,2.7)[$x$]{} (0,3)[(1,0)[6]{}]{} (0.55,1)(3,5)(5.45,1) (0,1)(3,5)(6,1) (5,1.67)[(3,2)[0.27]{}]{} (4.7,1.3)[$\delta x_\mu$]{} (0.3,1)[$A$]{} (0,1.3)[$A'$]{} (6,3) (3,0)[(0,1)[3]{}]{} (0,1)[(1,0)[6]{}]{} (0,0)[(1,0)[1]{}]{} (5,0)[(1,0)[1]{}]{} (3,0)[(1,0)[.1]{}]{} (1,0)(1.5,0)(2,0.5) (2,0.5)(2.5,1)(3,1) (3,1)(3.5,1)(4,0.5) (4,0.5)(4.5,0)(5,0) (3.1,2.9)[$\tau$]{} (5.9,0.7)[$t$]{} (3.2,0)[$\tau_\infty$]{} (4,3) (0,0)[(1,0)[4]{}]{} (0,0)[(0,1)[3]{}]{} (0,1)[(-1,0)[.1]{}]{} (0,0)[(0,1)[2]{}]{} (0,0)[(1,0)[.15]{}]{} (.15,0)[(0,1)[1]{}]{} (.15,1)[(1,0)[2]{}]{} (0.1,2.9)[$\tau$]{} (3.9,-0.3)[$t$]{} (-.3,0.9)[$\tau_c$]{} (-.3,-0.1)[$\tau_0$]{} (-.3,1.4)[I]{} (-.3,0.4)[II]{} (0.25,0.4)[III]{} (1,1.1)[IV]{} [^1]: Actually, the term proportional to $f_1$ arises when one shifts the singularity surface $\tau_0\to\tau_0+\delta\tau_0$ and try to expand the function $\sinh^{-1}(\tau-\tau_0-\delta\tau_0)$ on $\delta\tau_0$. Since we want the singularity surface to be $\tau=\tau_0({\bf x})$, this term should be excluded.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We describe a simple algorithm that calculates the distributions of electrons and positrons under multiphoton beamstrahlung at a future linear collider. The evolution equation as given by Chen is solved by a Monte Carlo algorithm. Explicit multiple beamstrahlung photons are generated. We present first results from an implementation of beamstrahlung into the Monte Carlo event generator [WOPPER]{}, that calculates the QED radiative corrections to the process $e^+e^- \to 4$ fermions through resonating W pairs. --- -0.25in .3cm IKDA 96/6\ hep-ph/9602397\ February 1996 [A Monte Carlo algorithm for Multiphoton\ Beamstrahlung in Monte Carlo event generators[^1][^2] ]{} [Harald Anlauf]{}\ [Technische Hochschule Darmstadt, 64289 Darmstadt, Germany\ and\ Universität Siegen, 57076 Siegen, Germany ]{} Introduction ============ At future linear $e^+e^-$ colliders with center of mass energies of 0.5 TeV and above, beamstrahlung [@overview; @Chen88], the synchrotron radiation from the colliding $e^+e^-$ beams, will become of significant physical interest. On one hand, beamstrahlung may carry away a substantial fraction of primary beam energy and lead to a degradation of the effective center of mass energy for $e^+e^-$ collisions, on the other hand, the lower energy $e^+e^-$ and $\gamma$’s contribute to background processes. In general, the full treatment of beam-beam interactions and beamstrahlung is a complicated many-body problem, and dedicated codes for the calculation of beam-beam effects exist [@ABEL] or are being developed [@CAIN]. However, from the high energy physics perspective the availability of a simplified analytical or numerical treatment is desired for the study of interesting processes, (e.g., $e^+e^- \to W^+W^-$), at the level of Monte Carlo event generators. As long as the average number of beamstrahlung photons per beam particle is much less than unity, the radiation spectrum is given by the Sokolov-Ternov spectrum [@ST86]. In the case of the proposed linear colliders, the average number of radiated photons per beam particle at design luminosity is typically ${\cal O}(1)$. Therefore, the effects of multiple radiation of beamstrahlung photons have to be taken into account. Multiphoton beamstrahlung has already been discussed in [@BCK91; @Chen92] (and references quoted therein). In this note we will follow the approach by Chen et al., however, without using any approximation to the radiation spectrum. We shall derive a recursive solution of the evolution equation, which is well suited for a numerical implementation. The Monte Carlo algorithm for beamstrahlung presented below is considered complementary to a full treatment of all beam-beam effects as performed in [@ABEL]. The main purpose of this Monte Carlo approach is to provide a simple package that can be incorporated into or interfaced with Monte Carlo event generators for physics at future $e^+e^-$ colliders. In section 2 we will review the electron energy spectrum under multiphoton beamstrahlung. Section 3 outlines our Monte Carlo approach, while in section 4 we present preliminary results. Further details of the calculation will be presented elsewhere [@Anl96]. Electron Energy Spectrum ======================== In this section we recapitulate the derivation of the electron energy spectrum due to multiphoton beamstrahlung in the spirit of Chen et al. [@BCK91; @Chen92]. We use $\hbar=c=1$. Let $\psi(x,t)$ be the distribution function of the electron with energy fraction $x = E/E_0$ at time $t$, normalized such that $$\label{eq:norm} \int_0^1 dx \; \psi(x,t) = 1 \; .$$ We shall assume that the emission of the photons takes place in an infinitesimally short time interval. Therefore the interference between successive emissions may be neglected, and one can describe the time evolution of the electron distribution by the rate equation $$\label{eq:evolution} \frac{\partial\psi}{\partial t} = - \nu(x) \psi(x,t) + \int_x^1 dx' \; F(x,x') \psi(x',t) \; ,$$ where the first and second term on the r.h.s. correspond to the sink and source for the evolution of $\psi$, respectively. $F$ is the spectral function of radiation, and $$\nu(x) = \int_0^x dx'' \; F(x'',x)$$ represents the average number of photons radiated per unit time by an electron with energy fraction $x$. Electrons and positrons from pair production in photon-photon processes are not taken into account. The spectral function of radiation is conventionally characterized by the beamstrahlung parameter $\Upsilon$, which is defined as $$\label{eq:Upsilon} \Upsilon = \frac{E_0}{m_e} \frac{B}{B_c} \; ,$$ where $B$ is the effective field strength in the beam, and $B_c = m_e^2/e$ is the Schwinger critical field. The $e^+e^-$ beams generally have Gaussian charge distributions, and thus the local field strength varies inside the beam volume. However, it has been shown [@Chen88] through integration over the impact parameter and the longitudinal variations, that the overall beamstrahlung effect can be simply described as if all particles experience a uniform mean field $B_{mean}$ during an effective collision time $\tau = l_{eff}/2 = \sqrt{3} \sigma_z$. In what follows, we will assume the obtained mean beamstrahlung parameter $\Upsilon_{mean}$ and effective beam length $l_{eff}$ but drop the subscripts. The spectral function of synchrotron radiation was derived by Sokolov and Ternov [@ST86] and reads $$\begin{aligned} F(x,x') & = & \frac{\nu_{cl}}{\Upsilon} \frac{2}{5\pi} \frac{1}{x'^2} \left\{ \int_\eta^\infty du \; K_{5/3}(u) + \frac{(\xi\eta)^2}{1+\xi\eta} K_{2/3}(u) \right\} \theta(x'-x) \; ,\end{aligned}$$ where $\xi = (3\Upsilon/2) x'$, $\eta=(2/3\Upsilon) (1/x-1/x')$, and $K_\nu$ denotes the modified Bessel function of order $\nu$. We define $$\nu(x) = \int_0^x dx' \; F(x',x) \equiv \nu_{cl} \cdot U_0(x\Upsilon) \; ,$$ where $\nu_{cl}$ is the number of photons per unit time in the classical limit, $$\nu_{cl} = \frac{5}{2\sqrt{3}} \frac{\alpha m_e^2}{E_0} \Upsilon \; ,$$ which is independent of the particle energy for a given field strength due to (\[eq:Upsilon\]). The function $U_0(y)$ behaves as $$U_0(y) = \left\{ \begin{array}{ll} 1 & y \to 0 \\ const. \times y^{-1/3} \quad & y \to \infty \end{array} \right. \; ,$$ exhibiting the asymptotic approach to the classical limit and the suppression of radiation in the deep quantum regime, respectively. Finally, the differential $e^+e^-$ luminosity can be expressed as the convolution of the effective electron energy distributions of the colliding beams [@Chen92], $$\label{eq:diff-lumi} \frac{ d{\cal L}(s) }{ds} = {\cal L}_0 \int dx_1 dx_2 \; \delta(s-x_1 x_2 s_0) \psi(x_1) \psi(x_2) \; ,$$ where ${\cal L}_0$ is the nominal luminosity of the collider, including the enhancement factor due to beam disruption, and the effective energy distribution $\psi(x)$ is obtained by averaging over the longitudinal position within the beam, $$\label{eq:eff-energy} \psi(x) = \frac{2}{l} \int_0^{l/2} \psi(x,t) \; .$$ Multiphoton Beamstrahlung by Monte Carlo ======================================== The evolution equation (\[eq:evolution\]) may be solved by a Monte Carlo algorithm. Here we will outline only the basic idea, deferring the full presentation and discussion to a forthcoming publication [@Anl96]. To determine the electron distribution $\psi(x,t)$, let us decompose it into a suitable set of time independent functions $\phi_n(x)$ using the ansatz $$\label{eq:ansatz} \psi(x,t) = \sum_{n=0}^\infty \phi_n(x) \cdot \frac{(\nu t)^n}{n!} e^{-\nu t} \; ,$$ where $\nu$ is a parameter that will be determined below. The normalization condition (\[eq:norm\]) translates into $$\label{eq:norm-phi} \int_0^1 dx \; \phi_n(x) = 1 \; .$$ Plugging the ansatz (\[eq:ansatz\]) into the evolution equation (\[eq:evolution\]) we obtain a recursion relation for the $\phi_n$, $$\label{eq:recursion} \phi_n(x) = \int_0^1 dx' \left[ \left(1 - \frac{\nu(x)}{\nu}\right) \delta(x-x') + \frac{F(x,x')}{\nu} \right] \phi_{n-1}(x') \quad \mbox{ for } n \geq 1 \; ,$$ with initial condition $$\phi_0(x) = \psi(x,0) \; .$$ If we neglect the intrinsic (Gaussian) energy spread of the accelerator, the initial condition of the electron distribution is simply given by $$\label{eq:initial-cond} \psi(x,0) = \delta(1-x) \; ,$$ and thus $$\phi_0(x) = \delta(1-x) \; .$$ The functions $\phi_n(x), n > 0$, can be determined, e.g., by numerical integration. Here we suggest a different way to exploit the recursion relations. First note that the kernel given by the square brackets in (\[eq:recursion\]) is positive for $x' > x$. Furthermore, since $\nu(0) \geq \nu(x) \geq \nu(1)$, the kernel is positive if we choose $\nu = \nu(0) = \nu_{cl}$. The functions $\phi_n$ will then be positive, and together with the normalization condition (\[eq:norm-phi\]) they will allow for a probabilistic interpretation. Using our ansatz (\[eq:ansatz\]), the effective electron energy distribution (\[eq:eff-energy\]) reads $$\label{eq:psi} \psi(x) = \sum_n \left[ 1 - \frac{\gamma(1+n; N_{cl})}{n!} \right] \phi_n(x) \; ,$$ where $\gamma(1+n; N_{cl})$ is the incomplete gamma function, and $N_{cl} = \nu_{cl} \cdot l/2$ is the average number of photons per beam particle radiated during the entire collision, in the classical limit. Due to the positivity of the term in square brackets in (\[eq:psi\]) and the functions $\phi_n$, the electron energy distribution can be determined straightforwardly by a multichannel Monte Carlo method [@Anl96]. Results {#sec:results} ======= Of the abundance of available parameter sets for future linear $e^+e^-$ colliders, we have taken a hopefully representative selection from [@Weise] that comprises many of the proposed linear collider designs. Table \[tab:NLC-params\] lists some of the parameters that are relevant for our numerical results. $\Upsilon_{mean}$ was determined similarly to [@Chen92]. Parameter SBLC TESLA NLC JLC-S JLC-X CLIC VLEPP ----------------------------------------------------- ------- ------- ------- ------- ------- ------- ------- Luminosity \[$10^{33} {\rm cm}^{-2} {\rm s}^{-1}$\] 5.3 6.1 6.9 4.6 5.1 ? 9.7 Particles/bunch $[10^{10}]$ 1.1 3.63 0.65 1.44 0.63 0.8 20 $\sigma_x$ \[nm\] 335 845 320 260 260 250 2000 $\sigma_y$ \[nm\] 15.1 18.9 3.2 3.04 3.04 7.5 3.9 $\sigma_z$ \[$\mu$m\] 300 700 100 120 90 200 750 $\Upsilon_{mean}$ 0.049 0.029 0.090 0.213 0.120 0.072 0.064 $N_{cl} = \nu_{cl} \cdot l/2$ 1.42 1.93 0.87 2.47 1.04 1.4 4.6 ${\cal L}(s>0.99\,s_0)/{\cal L}_0$ 0.58 0.55 0.68 0.32 0.62 0.56 0.18 We have calculated the differential luminosity of the $e^+e^-$ beams under beamstrahlung, eq. (\[eq:diff-lumi\]), using our Monte Carlo algorithm based on (\[eq:recursion\]). The result of the simulation for the normalized luminosity is shown in figure \[fig:luminosity\]. For the sake of brevity we have selected only a subset of four designs which essentially cover the interesting parameter range: the TESLA design with rather large bunch length but small effective beamstrahlung parameter and a correspondingly rather soft beamstrahlung spectrum, the JLC-S design with shorter bunches but large $\Upsilon$ and large mean energy loss due to a hard beamstrahlung spectrum, NLC with short bunches and moderate $\Upsilon$, having a small number of radiated photons $N_{cl}$, and VLEPP with a large multiplicity of beamstrahlung photons. We have also calculated the relative luminosity for collisions with $s > 0.99\, s_0$, the result being displayed in the last line of table \[tab:NLC-params\]. In high energy $e^+e^-$ processes, besides beamstrahlung one also has to take into account the radiative corrections to the process under consideration. The single most important universal contribution of the radiative corrections are the leading logarithms $\alpha^n \log^n(s/m_e^2)$, which are due to QED initial state radiation (ISR). These large logarithms can be resummed to all orders in the structure function approach, with the electron structure function [@StrFun]: $$D(z; \mu^2) = \frac{\beta}{2} (1-z)^{\frac{\beta}{2}-1} \left(1+\frac{3}{8}\beta \right) - \frac{1}{4}\beta(1+z) + {\cal O}(\beta^2) \; ,$$ where $$\beta = \frac{2\alpha}{\pi} \left( \log\frac{\mu^2}{m_e^2} - 1 \right) \; ,$$ $\mu^2$ is the factorization scale of order $s$, and $z$ is the electron energy fraction after initial state radiation. As an application, we have interfaced our Monte Carlo routines for beamstrahlung with the Monte Carlo event generator [WOPPER]{} [@WOPPER]. [WOPPER]{} simulates the process $e^+e^- \to (W^+W^-) \to 4f$ including QED radiative corrections in the abovementioned leading logarithmic approximation with explicit photons. In figure \[fig:WWinvmass\] we show the normalized invariant mass distribution $M^2_{WW}$ of the final state (assuming a 100% reconstruction efficiency for simplicity), $$\frac{1}{\sigma_0} \frac{d\sigma}{dM^2_{WW}} = \int dx_1 dx_2 dz_1 dz_2 \; \psi(x_1) \psi(x_2) \; D(z_1;\mu^2) D(z_2;\mu^2) \cdot \frac{1}{\sigma_0} \sigma_0(x_1 z_1 x_2 z_2 s) \; ,$$ using the same collider designs as in figure \[fig:luminosity\]. The thin continuous line represents the corresponding distribution from ISR alone. For events with a significantly reduced invariant mass the corrections due to beamstrahlung are numerically much larger than due to pure ISR, and the shape of the distribution is essentially dominated by the beamstrahlung spectrum. This result clearly shows the importance of beamstrahlung and the need to include beamstrahlung into Monte Carlo event generators for physics at future $e^+e^-$ colliders. [19]{} For an overview see e.g. D.V. Schroeder, Ph.D. thesis, SLAC-371, UC-414, (unpublished), and references quoted therein. P. Chen, in [Frontiers of Particle Beams]{}, Lectures Notes in Physics 296, Springer Verlag, Berlin 1988. K. Yokoya, [Nucl. Instr. Meth.]{} [A251]{} (1986) 1; SLAC-AAS-27, 1987 (unpublished). P. Chen, G. Horton-Smith, T. Ohgaki, A.W. Weidemann, and K. Yokoya, [Nucl. Instr. Meth.]{} [A355]{} (1995) 107. A.A. Sokolov and I.M. Ternov, in [Radiation from Relativistic Electrons]{}, AIP Translation Series, AIP, New York, 1986. T. Barklow, P. Chen, and W. Kozanecki, in [Proceedings of the Workshop $e^+e^-$ Collisions at 500 GeV: The Physics Potential – Munich, Annecy, Hamburg, February 4 to September 3, 1991]{}, DESY 92-123. P. Chen, [Phys. Rev.]{} [D46]{} (1992) 1186. H. Anlauf, in preparation. H. Weise, private communication. E.A. Kuraev and V.S. Fadin, [Yad. Fiz. 41*]{} (1985) 733;\ G. Altarelli and G. Martinelli, in [Physics at LEP]{}, CERN 86-02, Geneva 1986;\ W. Beenakker, F.A. Berends and W.L. van Neerven, [Proceedings of the Ringberg workshop on Electroweak Radiative Corrections]{}, April 1989. H. Anlauf, J. Biebel, H.D. Dahmen, A. Himmler, P. Manakos, T. Mannel, and W. Schönau, [Comp. Phys. Comm.]{} [79]{} (1994) 487;\ H. Anlauf, H.D. Dahmen, P. Manakos, and T. Ohl, [WOPPER]{} [Version 1.5]{}, work in progress. [^1]: Contribution to the Proceedings of the Workshop on [Physics with $e^+e^-$ Linear Colliders,*]{} Annecy, Gran Sasso, Hamburg, February 4 to September 1, 1995. [^2]: Supported by Bundesministerium für Bildung, Wissenschaft, Forschung und Technologie, Germany.	{ "pile_set_name": "ArXiv" }
--- abstract: 'I classify projective modules over idempotent semirings that are free on a monoid. The analysis extends to the case of the semiring of convex, piecewise-affine functions on a polyhedron, for which projective modules correspond to convex families of weight polyhedra for the general linear group.' author: - 'Andrew W. Macpherson' bibliography: - 'proj.bib' title: Projective modules over polyhedral semirings --- Introduction ============ This paper begins the geometric study of module theory over a class of idempotent semirings that are of basic importance in skeletal and tropical geometry: those generated freely by a monoid of monomials. The projective modules over these semirings can be described in terms of a simpler category of partially ordered modules over the underlying monoid. In the related case of the semiring of convex, piecewise-affine functions on a polyhedron, this latter category can itself be realised in terms of convex geometric data. This last intepretation of the classification suggests explicit descent criteria for modules over such ‘polyhedral’ semirings. The only existing work of which I am aware in this direction is [@crumpy], which addresses finite projective modules over the real semifield $\R_\vee=\R\sqcup\{-\infty\}$ (therein denoted $\mathbb T$). As far as I know, the classification scheme given in the present paper is new even in that case. Results {#results .unnumbered} ------- The moral of the paper will be that over semirings generated freely by monomials, projective modules are defined by monomial inequalities. As a warm-up, we have the following result for the case of the simplest possible semiring, the Boolean semifield $\B=\{-\infty,0\}$: Let $\mu$ be a finite $\B$-module. The following are equivalent: 1. $\mu$ is projective; 2. $\mu$ is dualisable; 3. $\mu$ is flat; 4. $\mu$ is free on a poset; 5. $\mu$ has unique irredundant primitive decompositions. A careful extension of the same analysis helps us to understand projectivity of modules over the free or fractional ideal semiring $\B[A;A^+]$ associated to a pair of $\F_1$-algebras (monoids with zero) $(A;A^+)$. We carry out this analysis in §\[MON\]. In the case that $A$ is a domain, one obtains a complete classification in terms of partially ordered $(A;A^+)$-modules. I do not reproduce the classification here, but refer the reader to corollary \[MON\_DOMAIN\_FIRST\]. The geometric part §\[POLY\] of the paper concerns modules over the semifield $H_\vee=H\sqcup\{-\infty\}$ associated to a totally ordered group $H\subseteq\R$. Let $H\subseteq\R$ be an additive subgroup. The category of finite projective $H_\vee$-modules is anti-equivalent to a category of extended $H$-integral general linear weight polyhedra and convex, piecewise-affine maps whose linear parts are fundamental weights.[^1] By duality for finite projectives these categories are also equivalent - I have just found it more natural to phrase the result in the form of a duality. Taking $H=0$ in this result recovers a more geometric formulation of the (only non-trivial) equivalence i)$\Leftrightarrow$iv) of theorem \[B\_PRIM\_THM\]: finite projective $\B$-modules are dual to certain convex cones in the Coxeter complex of $\GL_n$. We cannot immediately extend our classification scheme to the more geometric setting of the semiring of convex, piecewise-affine functions on a polyhedron, since the latter is not actually free on the group of affine functions. Rather, it is the normalisation of the free semiring [@norm]. Making this replacement helps us to get a geometric classification: Let $\Delta$ be an $H$-rational polytope. The category of finite projective ${\mathrm{CPA}}(\Delta,\Z)$-modules is anti-equivalent to a category of convex families of $\GL$ weight polyhedra over $\Delta$. For a more precise description of the latter category and the duality, see §\[POLY\_FAM\]. Acknowledgement {#acknowledgement .unnumbered} --------------- I’d like to thank Jeff Giansiracusa, a conversation with whom gave me the idea for this paper. Preliminaries {#ALG} ============= Points {#ALG_PTS} ------ It will be convenient to switch between the categories of pointed and unpointed sets. We do this using the strongly monoidal functor $$-\tens\F_1:(\mathbf{Set},\times)\rightarrow (\mathbf{Set}_,\tens_{\F_1}=\wedge)$$ that adjoins a disjoint base point. This exhibits the category $\mathbf{Set}_$ of pointed sets as the universal way to attach a zero object to $\mathbf{Set}$. This functor extends to the pointed versions of all essentially algebraic theories in $\mathbf{Set}$. In particular, a monoid $Q$ can be replaced by a monoid with zero $\F_1[Q]$, and a partially ordered set by a pointed partially ordered set (the point is the minimum). These constructions have the same universal property: $\mathbf{POSet}_$ (resp. ${\mathrm{Mod}}_{\F_1[Q]}$) is the universal pointed extension of $\mathbf{POSet}$ (resp. ${\mathrm{Mod}}_Q$). We will adhere to the convention of writing monoids without zero (mainly appearing in §\[POLY\]) and idempotent semirings additively, and pure $\F_1$-algebras (§\[MON\]) multiplicatively. If $(Q;Q^+)$ is a pair of unpointed monoids, we will write $\F_1[Q]$ as a shorthand for the associated $\F_1$-algebra pair. When $H\subseteq\R$ is an additive subgroup, $\F_1[H]$ denotes the $\F_1$-algebra pair that in [@rig2] was (more suggestively) labelled $\F_1(\!(t^{-H})\!)$. Projectives ----------- Let $A$ be a commutative monoid (with or without zero) or semiring. (In fact, the following definition is standard for any commutative algebraic monad in the sense of [@Durov].) An $A$-module is projective* if it satisfies the equivalent conditions 1. $\Hom(P,-)$ is exact; 2. any surjection $F\twoheadrightarrow P$ has a section; 3. $P$ is a retract of a free module. The situation when $A$ is a monoid - with or without zero - is very simple. A free $A$-module splits uniquely as a sum of cyclic factors $$M\cong \bigoplus_i A_i$$ with each $A_i\simeq A$ (where the coproduct $\oplus$ is disjoint union in the unpointed case and wedge sum when $0\in A$). The factor of a non-zero element $x\in M$ is the index of the cyclic submodule to which it belongs. This splitting, and in particular, its index set, is natural in $M$. In other words, every matrix over $A$ is a product of a permutation and a diagonal matrix. It follows: \[ALG\_CYCLIC\]Let $A$ be a commutative monoid. Every projective $A$-module splits uniquely as a coproduct of cyclic submodules, each isomorphic to the image of an idempotent in $A$. Let us call a monoid (resp. monoid with zero $A$) a domain if it is cancellative (resp. $A\setminus 0$ is a cancellative submonoid). If $A$ is a domain, then every projective $A$-module is free. Flatness and duality over semirings ----------------------------------- We now restrict attention to the case that $A$ is a semiring, whence ${\mathrm{Mod}}_A$ is semiadditive (i.e. finite coproducts are products). Let $A$ be a (not necessarily idempotent) semiring, $M$ an $A$-module. $M$ is said to be: 1. dualisable if there exists a module $M^\vee$ such that $-\tens_AM^\vee$ is adjoint to $-\tens_AM$; 2. flat if $-\tens_AM$ is exact. The linear dual of a module is the object $$M^\vee:=\Hom_A(M,A);$$ a module is said to be reflexive if $M\tilde\rightarrow (M^\vee)^\vee$. This is strictly weaker than being dualisable, for while we always have the evaluation map $M\tens_A M^\vee\rightarrow A$, there may be no ‘identity matrix’ $A\rightarrow M^\vee\tens_AM$. Flatness, at least with this definition, does not make much sense when $A$ is a monoid (more generally, when ${\mathrm{Mod}}_A$ is not semiadditive), because the class of flat modules is not closed under finite coproducts. In particular, free modules on more than one generator are never flat. Some authors [@bougie] have studied variants of the notion of flatness adapted to $A$-modules (or ‘$A$-acts’) in the unpointed case. The relations between these properties are, much as in the case of commutative rings, as follows: - Any coproduct, filtered colimit, or retract of a flat module is flat. In particular, projective modules are flat. - Any finite coproduct or retract of a dualisable module is dualisable. In particular, finitely generated projective modules are dualisable. - Dualisable modules are flat, since in this case $-\tens_A M^\vee$ is left adjoint to $-\tens_AM$. - Conversely, finitely presented flat modules are projective (corollary \[ALG\_FLAT\]). The following fact is no doubt well-known - indeed, the proof for the case of rings [@stacksproject [00HK](http://stacks.math.columbia.edu/tag/00HK)] carries through with only minor modifications. \[ALG\_EQN\]Let $M$ be a flat module, $v:A^n\rightarrow M$ a homomorphism from a finite free module. Suppose that $v$ satisfies a relation $f\sim g$: $$A \underset{g}{\stackrel{f}{\rightrightarrows}} A^n \stackrel{v}{\rightarrow} M.$$ Then $v$ factors through a finite free module in which $f\sim g$. Dualising the relation, we obtain an equaliser sequence $$K\rightarrow A^n\rightrightarrows A,$$ and hence by flatness of $M$, an equaliser $$K\tens M \rightarrow M^n\rightrightarrows M.$$ It follows that $v\in M^n$ actually lies in $K\tens M$. By writing $v$ as a sum of decomposable elements, one obtains morphisms $A^m\rightarrow K\subseteq A^n$ and $A^m\rightarrow M$ that induce the composite $$A^m \stackrel{\Sigma}{\rightarrow} A \stackrel{v}{\rightarrow} K\tens M$$ on the tensor product. The transpose of $A^m\rightarrow A^n$ is the desired factorisation. Every homomorphism from a finitely presented module into a flat module factors through a finite free module. \[ALG\_FLAT\]Every finitely presented flat module is projective. \[ALG\_LAZARD\]A module is flat if and only if it is a filtered colimit of finite free modules. $\B$-modules {#B} ============ For the rest of the document, all semirings will be additively idempotent. I continue with the convention of [@mac] in denoting idempotent semirings and their modules by lowercase Greek letters, and their operations by $\vee$ (‘max’) and $+$. Correspondingly, the closed monoidal structure on the category ${\mathrm{Mod}}_\alpha$ of modules over an idempotent semiring $\alpha$ is denoted $\oplus_\alpha$ (i.e. this does not denote the coproduct of modules). As in the introduction, we will denote by $\B$ the Boolean semifield, the initial object in the category of idempotent semirings. It will be illuminating to understand $\B$ as a monad via its free functor $$\B:\mathbf{Set}\rightarrow\mathbf{Set}$$ that takes a set $S$ to the set $\B S$ of its finite subsets, with operations $$1_\mathbf{Set} \stackrel{\{-\}}{\longrightarrow}\B \stackrel{\bigcup}{\longleftarrow}\B\B$$ given by singleton and union, respectively. This monad is the unique algebraic (i.e. commuting with filtered colimits) extension of the monad of power set and union on the category of finite sets. A finite, free $\B$-module is nothing more than the power set of a finite set. The same statements remain valid, mutatis mutandi, with $\mathbf{Set}$ replaced by the category $\mathbf{Set}_$ of pointed sets. The free $\B$-module functor factorises $$\mathbf{Set}\rightarrow\mathbf{Set}_\rightarrow{\mathrm{Mod}}_\B.$$ We will use this in §\[POLY\] to apply the results of §\[MON\], couched in the setting of monoids with zero, in the unpointed regime. More concretely, a $\B$-module $(\mu,\vee)$ is nothing more than a join semilattice, that is, a partially ordered set with finite joins, and a $\B$-linear morphism is a right exact monotone map. In [@mac], these were called ‘spans’. The following are equivalent for a $\B$-module $\mu$: 1. $\mu$ is finite as an object of ${\mathrm{Mod}}_\B$; 2. $\mu$ is compact as an object of ${\mathrm{Mod}}_\B$; 3. $\mu$ is a finite set. Free module on a poset ---------------------- Let us denote by $\mathbf{POSet}$ the category of partially ordered sets and monotone maps. Since every $\B$-module is in particular a poset, we have a faithful, conservative functor $${\mathrm{Mod}}_\B\rightarrow\mathbf{POSet}.$$ In fact, this functor is monadic; its left adjoint takes a poset $(S,\leq)$ to the poset $\B(S,\leq)$ of finitely generated lower subsets. Since any union of lower subsets is lower, union makes this poset a $\B$-submodule of the power set $\sh P(S)$. It is called the free $\B$-module on $(S,\leq)$. By general principles, $\B:\mathbf{POSet}\rightarrow{\mathrm{Mod}}_\B$ is strongly monoidal. It also respects Hom sets in the following way: if $S_1,S_2$ are posets, then the assignment $$(X,Y):Z\mapsto \left\{\begin{array}{ll} Y & \text{if }X\leq Z \\ -\infty & \text{otherwise}\end{array} \right.$$ defines a monotone map $S_1^\mathrm{op}\times S_2 \rightarrow \Hom(\B S_1,\B S_2)$; if $S_1$ is finite, then this map extends to an isomorphism $$\B(S_1^\mathrm{op}\times S_2) \cong \Hom_\B(\B S_1,\B S_2)$$ so that maps $\B S_1\rightarrow\B S_2$ are finite monotone correspondences from $S_1$ to $S_2$. The same logic holds over the category $\mathbf{POSet}_$ of pointed* posets, that is, posets equipped with a distinguished minimum and monotone maps that preserve this minimum. #### Flatness and projectivity The free module on a poset comes equipped with a natural set of generators $$\B S\twoheadrightarrow \B(S,\leq).$$ To understand projectivity, we will need to know when this map admits a section. A natural candidate for a splitting is the (right ind-adjoint) inclusion $\B(S,\leq)\hookrightarrow\sh P(S)$. As remarked above, it is automatically a $\B$-module homomorphism. This map factors through $\B S$ - thus defining an honest adjoint - if and only if $(S,\leq)$ satisfies the condition - any principal lower set $S_{\leq X}$ is finite.lower finite This has the flavour of a ‘finite presentation’ condition for posets: it is equivalent that the quotient $\B S\rightarrow\B (S,\leq)$ be defined by finitely many relations. It is satisfied, in particular, whenever $S$ is finite. It turns out that this is the only way to split this epimorphism: \[B\_POSET\_SECTION\]Let $p:(S_1,\leq)\rightarrow(S_2,\leq)$ be a bijective monotone map, $\sigma:\B(S_2,\leq)\rightarrow\B(S_1,\leq)$ a monotone section of its $\B$-linear extension. Then $\sigma$ is right adjoint to $\B p$ and $\B$-linear. Let $\sigma:\B(S_2,\leq)\rightarrow \B(S_1,\leq)$ be any section, $S_0\in\B(S_2,\leq)$. Then $p\sigma S_0\subseteq S_0$ generates $S_0$ as a lower set. In particular, for any $X\in S_2$, $X\in p\sigma X$, i.e. $p^{-1}X\in\sigma X$. If $\sigma$ is monotone, then this shows that $\sigma p$ is increasing on $\B(S_1,\leq)$. Thus $Y\leq\sigma X$ if and only if $pY\leq X$, that is, $\sigma$ is right adjoint to $p$. The proof of the lemma \[B\_POSET\_SECTION\] on monotone sections depends on the fact that a lower set has a unique minimal set of generators. We will use variations of the latter fact, and the lemma, repeatedly in the sequel. \[B\_POSET\_PROJ\]Let $(S,\leq)$ be a poset. The following are equivalent: 1. $\B(S,\leq)$ is projective; 2. $\B(S,\leq)$ is lower finite; 3. $(S,\leq)$ is lower finite. From Lazard’s theorem \[ALG\_LAZARD\] it also follows: \[B\_PRIM\_FLAT\]The free $\B$-module on any poset is flat. The converse to corollary \[B\_PRIM\_FLAT\] is false: while every flat module is a filtered colimit of modules free on a poset, there is no requirement that the transition maps preserve these posets. For example, the set $\mu$ of compact open subsets of $\Z_2$ has an expression as a colimit $$\mu=\colim_k\B(\Z/2^k\Z) \simeq \colim_k\B^{2^k}$$ with transition maps given by inverse image along $\Z/2^k\Z\twoheadrightarrow\Z/2^{k-1}\Z$, and is therefore flat by Lazard’s theorem \[ALG\_LAZARD\]; however, it has no primitive elements (cf. §\[B\_PRIM\] below) and so cannot be free on a poset. Primitives {#B_PRIM} ---------- To achieve our goal of classifying projective modules, we must still characterise which $\B$-modules appear through this construction. \[B\_PRIM\_LEMMA\]Let $X$ be an element of a $\B$-module $\mu$. The following are equivalent: 1. $X=\bigvee_{i\in I}X_i$, with $I$ a finite set, implies $X=X_i$ for some $i\in I$; 2. if $S\subseteq\mu $ generates - that is, if $\B S\rightarrow\mu$ is surjective - then $X\in S$. \[B\_PRIM\_DEF\]An element $X$ of a $\B$-module $\mu$ satisfying the equivalent conditions of lemma \[B\_PRIM\_LEMMA\] is said to be $\vee$-primitive, or simply primitive if no confusion can arise. (Note that $-\infty$ is never primitive.) The set of primitive elements of $\mu$ is denoted $\mathrm{Prim}\mu$. It is usually not functorial in either direction. We have tautological maps $$\B({\mathrm{Prim}}\mu)\twoheadrightarrow\B({\mathrm{Prim}}\mu,\leq)\rightarrow\mu$$ where ${\mathrm{Prim}}\mu\subset \mu$ carries the induced partial order. If these modules surject onto $\mu$, we say that $\mu$ has primitive decompositions. A primitive decomposition of an element $X\in\mu$ is a lift to $\B({\mathrm{Prim}}\mu)$. Such a decomposition is said to be irredundant if it is minimal in its fibre of $\B({\mathrm{Prim}}\mu)\twoheadrightarrow\B({\mathrm{Prim}}\mu,\leq)$. The set of possible primitive decompositions (resp. irredundant decompositions) of elements of $\mu$ is precisely $\B({\mathrm{Prim}}\mu)$ (resp. $\B({\mathrm{Prim}}\mu,\leq)$). In equations, a primitive decomposition of $X\in\mu$ is an expression $X=\bigvee_{i\in I}X_i$ with $X_i$ primitive, and it is irredundant if there are no order relations among different $X_i$. Beware that an irredundant decomposition in $\mu$ is not necessarily minimal: one may perfectly well have a relation $$X_1\vee X_2\vee X_3=X_1\vee X_2$$ between primitives in $\mu$, but unless $X_3\leq X_2$ or $X_3\leq X_1$, both decompositions will be irredundant. Of course, in light of theorem \[B\_PRIM\_THM\], this cannot happen for projective modules. #### Frees A module is free if and only if $\B({\mathrm{Prim}}\mu)\rightarrow\mu$ is an isomorphism; that is, if it has unique primitive decompositions. \[B\_PRIM\_FREE\]Let $\mu$ be a $\B$-module. The following are equivalent: 1. $\mu$ is free; 2. $\mu$ has unique primitive decompositions. #### Projectives More generally, a module is free on a poset if and only if it has unique irredundant primitive decompositions. \[B\_PRIM\_THM\]Let $\mu$ be a $\B$-module. The following are equivalent: 1. $\mu$ is projective; 2. $\mu$ is free on a lower finite poset; 3. $\mu$ has unique irredundant primitive decompositions, and the set of primitives is lower finite. To complete the proof of this theorem, we must show that projective modules have unique irredundant primitive decompositions. \[B\_PRIM\_SUB\]Any submodule of a $\B$-module having primitive decompositions itself has primitive decompositions. \[B\_PRIM\_SECTION\]Let $S\subseteq\mu$ be a subset, $\sigma$ any section of $$\B(S,\leq)\rightarrow\mu.$$ If $X\in\mu$ is primitive, then $\sigma X=S_{\leq X}$. Let $\sigma$ be a section, and let $X\in\mu$. Since $\bigvee_{Y\in\sigma X}Y=X$, $\sigma X$ must certainly be contained in $S_{\leq X}$. If $X$ is primitive, then in fact $X\in\sigma X$ and so $\sigma X=S_{\leq X}$. A projective module $\mu$ is a submodule of a free module, and hence by lemma \[B\_PRIM\_SUB\], $\B({\mathrm{Prim}}\mu,\leq)\rightarrow\mu$ is surjective. By projectivity, it admits a $\B$-linear section. Applying the lemma \[B\_PRIM\_SECTION\] on sections over primitives to $S={\mathrm{Prim}}\mu$ shows that it is an isomorphism. Modules over free semirings {#MON} =========================== Here we generalise and discuss modules over semirings that are free on a pair $(A;A^+)$ of monoids with $A$ an $A^+$-algebra. Our convention in this section will be that monoids are multiplicative with zero; the category of such pairs is denoted $\mathrm{Pair}_{\F_1}$. Following the remarks of §\[ALG\_PTS\], the results translate straightforwardly into the unpointed regime. Usually, $A$ will be a localisation of $A^+$; correspondingly, we will typically consider idempotent semirings $\alpha$ that are a localisation of their semiring of integers $\alpha^\circ:=\alpha_{\leq 0}$. However, these assumptions are not actually necessary for the conclusions of §§\[MON\_FREE\]-\[MON\_PRIM\]. We will later need to assume that $A^+$ is integrally closed in $A$ (§\[MON\_DOMAIN\]) and that $\alpha$ is normal in the sense of [@norm] (§\[POLY\_FAM\]). The structure of this section is as follows. The first two subsections and §\[MON\_PRIM\] are parallel to the structure of §\[B\]. In §\[MON\_POMOD\] we derive formal criteria for the free module on a partially ordered projective $A$-module to be projective, and in §\[MON\_PRIM\] we show that conversely, every projective module over the free semiring $\B[A]$ on $A$ comes from a partially ordered $A$-module. The technical heart of the paper lies in §\[MON\_QUIVER\], where we discuss a method of presenting partial orders via quivers. This will help us to unpack the meanings of the conditions appearing in §\[MON\_POMOD\]. In particular, we define the notion of non-degeneracy of partially ordered free modules, which has crucial finiteness implications. When $A$ is a domain, we obtain a complete classification in §\[MON\_DOMAIN\]. The remaining poings are, first, that the partially ordered $A$-module behind a projective $\B[A]$-module is itself projective, and second, that non-degeneracy is a necessary condition for lower finiteness. Free semirings and modules {#MON_FREE} -------------------------- Let $(A;A^+)$ be an $\F_1$-algebra pair, and define, as in [@mac], the fractional ideal semiring $$\B[A;A^+]:=\{\text{finite }A^+\text{-submodules of }A\}$$ with semiring of integers $$\B[A;A^+]^\circ=\B[A^+;A^+].$$ We also write simply $\B[A]$ when $A^+$ is considered implicit. There is natural monoid homomorphism $$\log:A\rightarrow\B[A;A^+], \quad \log(A^+)\leq 0$$ the ‘universal valuation’ of $(A;A^+)$. This valuation is injective if and only if $A^+$ is sharp (i.e. has no invertible elements other than $1$). In general, the image of $A^+$ in the monoid semiring is its universal sharp quotient. For the purposes of studying $\B[A]$-modules, then, we can and will always assume that $A^+$ is sharp. Note that in contrast to the situation for commutative rings, the fractional ideal functor has a monadic right adjoint forgetful functor $${\mathrm{Alg}}_\B \rightarrow \mathrm{Pair}_{\F_1},\quad \alpha \mapsto (\alpha;\alpha^\circ)$$ which forgets the $\vee$ operation. In this language, $\log$ is the unit of the adjunction. More generally, if $M$ is an $A$-module, one can form the disc (or fractional submodule) set $$\B(M;A^+):=\{\text{finite }A^+\text{-submodules of }M\},$$ more briefly, $\B(M)$, which carries a natural structure of a $\B[A]$-module. This fits into an adjunction $$\B(-;A^+):{\mathrm{Mod}}_A\leftrightarrows{\mathrm{Mod}}_{\B[A]}$$ with the evident forgetful functor. It follows that $\B(-;A^+)$ preserves colimits, and hence the classes of free, projective, and (by the Lazard theorem \[ALG\_LAZARD\]) flat modules. Partially ordered modules {#MON_POMOD} ------------------------- If $A\in\mathrm{Pair}_{\F_1}$, then any $A$-module carries a natural pointed $A^+$-divisibility partial order $$fv\leq_{A^+}v \quad \forall f\in A^+$$ which is non-degenerate if and only if $A^+$ is sharp. An $A$-linear map is automatically monotone with respect to this order. The divisibility order on a smash product is the smash product of the orders on the factors. It therefore makes $(A,\leq_{A^+})$ into an algebra object in the closed monoidal category $(\mathbf{POSet}_,\tens_{\F_1})$ of pointed posets (cf. §\[ALG\_PTS\]), and $(M,\leq_{A^+})$ into an $(A,\leq_{A^+})$-module. This defines fully faithful functors $$\mathrm{Pairs}_{\F_1}\hookrightarrow\mathrm{Alg}(\mathbf{POSet}_,\tens_{\F_1})$$ and, for each $(A;A^+)$, $${\mathrm{Mod}}_A\hookrightarrow{\mathrm{Mod}}_{(A,\leq_{A^+})}(\mathbf{POSet}_).$$ In the spirit of [@G2], the free $\B[A]$-module on an $A$-module $M$ can be presented via this partial order $$\B(M;A^+)\cong\B(M;\leq_{A^+})=\B(M;\mathbf{Set}_)/(fv\leq v\|\forall f\in A^+)$$ which shows that, in particular, $\B(M)$ has unique irredundant primitive decompositions. By corollary \[B\_PRIM\_FLAT\], it is moreover flat as a $\B$-module. A partially ordered module, or po-module, over a $\F_1$-algebra pair $(A;A^+)$ is an $(A,\leq_{A^+})$-module object in $(\mathbf{POSet}_,\tens_{\F_1})$. That is, it is an $A$-module together with a partial order such that 1. $fv\leq gv$ for all $v\in M$, $g\in A$ and $f\in A^+g$; 2. $v\leq w\Rightarrow fv\leq fw$ for all $v,w\in M$ and $f\in A$. The category of partially ordered $A$-modules is abbreviated $\mathbf{PO}{\mathrm{Mod}}_A:={\mathrm{Mod}}_{(A,\leq_{A^+})}(\mathbf{POSet}_)$. It is closed monoidal and compactly generated. By analogy with the case of $\B$-modules, we can define a ‘free’ $\B[A]$-module on any partially ordered $A$-module: $$\B(M,\leq):=\{\text{finitely generated lower $A^+$-submodules of $M$}\};$$ here a lower submodule is ‘finitely generated’ if it is the lower hull of a finite $A^+$-submodule; and hence a monadic adjunction $$\mathbf{PO}{\mathrm{Mod}}_A\leftrightarrows{\mathrm{Mod}}_{\B[A]}$$ extending the one defined in §\[MON\_FREE\]. Intuitively, $\B[A]$-modules that are free over a po-$A$-module are those ‘defined by monomial relations’. \[MON\_POMOD\_THM\]The free module on a partially ordered projective $A$-module $(M,\leq)$ is a projective $\B[A]$-module if and only if the following conditions are satisfied: 1. the lower hull of a finite disc in $M$ is finite; lower finite 2. if $D\subseteq M$ is a lower disc, then so is $fD$ for $f\in A$. lower saturated Indeed, we have already observed that $\B(-;A^+)$ preserves projectivity, and so assuming $M$ projective, by the monotone section lemma \[B\_POSET\_SECTION\] we must check that the canonical generator $$\B(M;A^+)\twoheadrightarrow \B(M,\leq)$$ has a right adjoint. Condition i) is enough to show that the right adjoint exists as a map of posets, and ii) is the condition that it be $A$-linear. We will expand upon the meanings of the other two conditions in §\[MON\_QUIVER\]. A $\B[A]$-module is flat if it is free on a lower cancellative po-$A$-module that is flat as an $A$-module. Partial orders from quivers {#MON_QUIVER} --------------------------- In the case of projective $A$-modules - or more generally, direct sums of cyclic modules - we can give a fairly explicit method for defining partial orders. We will use this method as an auxiliary tool to obtain a good classification theorem \[MON\_DOMAIN\_FIRST\]. Let $M\cong \bigoplus_{i\in I} A_i$ be such a module, and let $\sh Q$ be a (possibly infinite) quiver with vertex set $I$. Suppose that we are given the structure of a representation of $\sh Q$ on $M$ - that is, for each edge in $\sh Q$ from $i$ to $j$, a map $A_i\rightarrow A_j$. If we choose generators $x_i\in A_i$ for the cyclic factors of $M$ - since $A^+$ is sharp, this is the same as choosing a cyclic $A^+$-structure for each factor - then such can be represented by attaching an element of $A\twoheadrightarrow\Hom_A(A_i,A_j)$ to each edge of $\sh Q$. With a choice of generators, we can exchange a quiver representation for a presentation $$\begin{aligned} \label{q1} \{(ex_i,x_i)\|e\in\mathrm{Edge}(\sh Q)\}\subset M^2 \end{aligned}$$ of a partial order on $M$. In particular, this presentation is finite if and only if $\sh Q$ has finitely many edges. Conversely, any presentation $R\subset M^2$ of this form - with the right-hand term always the chosen generator of its cyclic factor - can be obtained from a quiver representation, for which $R$ is the set of edges via the natural projection $M\setminus0\rightarrow I$. More invariantly, the action of the path category $\mathrm{Path}(\sh Q)$ of $\sh Q$ defines an $A$-invariant pre-order $$\begin{aligned} \label{q2}x\leq^{\sh Q}\!\! y \quad\Leftrightarrow\quad \exists\gamma\in\mathrm{Path}(\sh Q):x=\gamma y \end{aligned}$$ on $M$. It is the transitive and $A$-invariant closure of the relation defined by the set in (\[q1\]). The union of this pre-order with the $A^+$-divisibility order is an $A$-module pre-order $$\begin{aligned} \label{q3} x\leq^\sh{Q}_{A^+}\!\! y \quad\Leftrightarrow\quad \exists\gamma\in\mathrm{Path}_{A^+}(\sh Q):x=\gamma y \end{aligned}$$ where $\mathrm{Path}_{A^+}$ denotes the $A^+$-linear extension of the path category. The latter is degenerate if there is a cycle $\gamma:i\rightarrow i$ in $\mathrm{Path}(\sh Q)$ such that $1\leq_{A^+}\gamma\in\End(A_i)$; otherwise, it defines the structure of a po-$A$-module on $M$. We call it the partial order presented by $\sh Q\curvearrowright M$. #### Lower saturation Partial orders on $M$ can be equivalently described by producing, for each $x\in M$, the lower hull $M_{\leq x}$. For lower saturated partial orders (cf. proposition \[MON\_POMOD\_THM\], ii) it is enough to define the lower hull for $x_i$ the generators of the cyclic factors of $M$, since by definition in that case, $$M_{\leq fx_i}=fM_{\leq x_i}$$ gives the formula for general elements $fx_i$ of $M$. From such a module we can produce a quiver whose edges $i\rightarrow j$ are a generating set for $M_{\leq x_i}\cap A_j$ as an $A^+$-module. This quiver presents the partial order via (\[q1\]). Conversely, the partial order presented by a quiver action $\sh Q\curvearrowright M$ has lower sets $$\begin{aligned} \label{lower} M_{\leq x}=\mathrm{Path}_{A^+}(\sh Q)x = A^+\mathrm{Path}(\sh Q)x \end{aligned}$$ whose definition manifestly commutes with the action of $A$. In other words, quiver partial orders are lower saturated. \[MON\_DOMAIN\_SECOND\]Let $(M,\leq)$ be a partially ordered direct sum of cyclic $A$-modules. The following are equivalent: 1. $(M,\leq)$ is lower saturated; 2. $\leq$ can be presented by a quiver. #### Lower finiteness When $(M,\leq)$ is lower finite, in particular each $M_{\leq x_i}$ is finitely generated, and so by sharpness of $A^+$ has a unique set of primitive generators. Let us call the quiver $\sh Q$ with these generators as its set of edges the canonical quiver. (If $A^+_{\leq i}\cap A_i^+=A_i^+$, then this algorithm yields a trivial loop at $i$, which we may exclude.) It has finitely edges departing from each vertex. In particular: Let $M$ be finite and $\leq$ lower finite. Then $(M,\leq)$ is finitely presented. #### Non-degeneracy Unfortunately, being presented by a quiver with finitely many edges departing each vertex is not sufficient to guarantee lower finiteness; by the formula \[lower\], it is the action of the path algebra at $i$ that we need to worry about, and the latter is infinite whenever there is a cycle at $i$. We need a way to disregard such cycles. Let $M$ be a direct sum of cyclic $A$-modules. A partial ordering on $M$ is said to be non-degenerate if the induced partial order on each cyclic factor is the $A^+$-divisibility order $\leq_{A^+}$. Non-degeneracy of a partial order entails that the element of $A$ attached to any path in a presenting quiver $\sh Q$ with the same start and end point $i$ actually lies in $A^+\subseteq \End(A_i)$, and so $$M_{\leq x_i}\cap A_i=A^+x_i$$ for any $x_i\in A_i$. Such cycles can therefore be disregarded in the presentation (\[q3\]), and we may restrict attention to the set $\sh Q^\emptyset\subseteq\mathrm{Path}(\sh Q)$ of acyclic paths. The latter is finite as soon as $\sh Q$ is. \[MON\_DOMAIN\_NONDEGEN\]Let $(M,\leq)$ be lower saturated and finitely presented. If $\leq$ is non-degenerate, then it is lower finite. All that remains to obtain an equivalence $$\text{lower finite} \qquad \Leftrightarrow \qquad \text{finitely presented } + \text{ non-degenerate}$$ is to show that lower finite orders are non-degenerate. We conclude this for domains in §\[MON\_DOMAIN\]. Primitives {#MON_PRIM} ---------- Free modules on a po-module have the following distinguishing features: - they have unique irredundant primitive decompositions (cf. §\[B\_PRIM\]); - the set of primitive elements, together with $-\infty$, is an $A$-submodule. If, in general, these criteria are satisfied, we can by passing to primitive elements reverse the procedure and obtain a po-$A$-module from a $\B[A]$-module. This is possible for finite projective $\B[A]$-modules, by a partial converse to proposition \[MON\_POMOD\_THM\]: \[MON\_PRIM\_THM\]Every projective $\B[A]$-module is free on a partially ordered $A$-module. We have seen that free $\B[A]$-modules are, in particular, free $\B$-modules on a poset. By this and lemma \[B\_PRIM\_SUB\], any projective $\B[A]$-module has primitive decompositions. Now let $\mu$ be projective, $A{\mathrm{Prim}}\mu\subseteq \mu$ the $A$-submodule generated by the primitive elements. Then $A{\mathrm{Prim}}\mu$ generates $\mu$ as a $\B$-module, that is, $$\B(A{\mathrm{Prim}}\mu,\leq)\twoheadrightarrow\mu$$ is surjective. By projectivity of $\mu$, it has a $\B[A]$-linear section. By applying the lemma \[B\_PRIM\_SECTION\] on sections over primitives to $S=A{\mathrm{Prim}}\mu$, the section surjects onto the set of elements of $\B(A{\mathrm{Prim}}\mu,\leq)$ of the form $(A{\mathrm{Prim}}\mu)_{\leq X}$ with $X\in{\mathrm{Prim}}\mu$. Since these generate it as a $\B[A]$-module, it is an isomorphism. (In particular, $A{\mathrm{Prim}}\mu={\mathrm{Prim}}\mu$ is an $A$-submodule of $\mu$.) Domains {#MON_DOMAIN} ------- It remains to determine whether it is necessary that $M$ be projective, that is, that only projective $A$-modules can yield projective $\B[A]$-modules. We have a complete result in the case that $A$ is an $\F_1$-domain. A monoid pair $(A;A^+)$ is called an $\F_1$-domain if $A^+\setminus 0$ is cancellative, $A$ is a localisation of $A^+$, and $A^+$ is integrally closed (i.e. saturated) in $A$. I remind the reader that $A^+$ is always assumed to be sharp. Since $A^+$ is integrally closed in $A$, the multiplicative torsion of $A$ must belong to $A^+$ and therefore be trivial. \[MON\_POSET\_PROJ\]Let $M$ be a partially ordered $A$-module. 1. If $\B(M;\leq)$ is finitely generated over $\B[A]$, then $M$ is finitely generated over $A$. 2. Suppose $A$ is a domain. If $\B(M;\leq)$ is projective over $\B[A]$, then $M$ is free over $A$. i) If $I\subseteq\B(M,\leq)$ is a generating set, then so too is the set $I^\prime\subset M$ of irredundant primitive factors of elements of $I$. Thus $AI^\prime\subseteq M$ generates $\B(M,\leq)$ as a $\B$-module, and therefore contains $M$. In particular, if $I$ is finite, then $I^\prime$ is finite and $M$ is finitely generated. ii) Let $p:\bigoplus_iA_i\twoheadrightarrow M$ be a surjection from a free module, where each $A_i\simeq A$ is free cyclic. By eliminating indices $i$ such that $pA_i$ is strictly contained in another $pA_j$, and identifying indices with the same image, we may assume that $p$ is an irredundant generator, i.e. that no factor can be removed without destroying surjectivity of $p$. By projectivity of $\B(M;\leq)$, there exists a section $\sigma$ to the induced projection $$\B(\bigoplus_iA_i;A^+)\twoheadrightarrow\B(M;\leq).$$ Using the natural identification $\bigoplus_i\B(A_i)\cong\B(\bigoplus_iA_i)$, we will study $\sigma$ via its components $$\sigma_i:\B(M;\leq)\rightarrow\B(\oplus_iA_i;A^+)\stackrel{\mathrm{pr}_i}{\rightarrow}\B A_i.$$ \[MON\_DOMAIN\_UNIQUE\_GENS\]If $x\in A_i$, then $x\in\sigma_ipv$. By $A$-linearity, it will be enough to show this for $x$ a generator of $A_i$. Irredundancy implies that $px$ is not in the image of $A_j$ for any $j\neq i$. Thus $\sigma_ipx$ contains a lift $fx$ of $px$, and $f$ fixes $px$. Applying $\sigma_i$ to the relation $fpx=px$ shows that $f$ acts by an automorphism of $\sigma_ipx$. Since $A^+$ is sharp, this automorphism must permute the finite set of primitive generators. In particular, the orbit of $fx$ is finite. Thus $f$ is torsion, therefore $1$. \[MON\_POSET\_LEMMA\]$M$ is a direct sum of the cyclic modules $pA_i$. Since $A_i$ and $A_j$ are free cyclic, there are identifications $A_i\simeq A\simeq A_j$. We will show that if $v_i,v_j$, are the images of $1$ in $M$ under $p$ with respect to these two identifications, then any relation of the form $$f_iv_i=f_jv_j$$ in $M$ will force $i=j$. By lemma \[MON\_DOMAIN\_UNIQUE\_GENS\], applying $\sigma_j$ (resp. $\sigma_i$) to the relation shows that there exist $e_i\in\sigma_j v_i$ (resp. $e_j\in\sigma_iv_j$) such that $$f_ie_i=f_j, \quad f_i=f_je_j$$ in $A$, and so in particular $e_ie_jf_i=f_i$. By cancellativity of $A\setminus 0$, $e_ie_j=1$. Applying $p$ to the relation $\sigma_jv_i\leq\sigma v_i$ gives $e_iv_j\leq v_i$, and vice versa. We deduce $$v_i=e_ie_jv_i\leq e_iv_j\leq v_i$$ in $M$, that is, $v_i=e_iv_j$. By irredundancy of $I$, $i=j$. It remains to show that the cyclic factors of $M$ are free. This follows from a ‘high-brow’ argument based on the fact that $\B$ is strongly monoidal and commutes with equalisers. Since $\B(M)$ is by hypothesis projective, the tensor sum $\B(M)\oplus_{\B[A]}-$ is left exact. By taking primitives, this implies that $M\tens_A-$ commutes with equalisers. It follows from the argument for the equational criterion for flatness \[ALG\_EQN\] that any relation $$A_i\underset{f_j}{\stackrel{f_i}{\rightrightarrows}} A_i\stackrel{p}{\rightarrow} M$$ in a cyclic summand of $M$ must be trivial. \[MON\_DOMAIN\_FIRST\]Let $A$ be an $\F_1$-domain. A $\B[A]$-module $\mu$ is finite projective if and only if: 1. it has unique irredundant primitive decompositions; 2. ${\mathrm{Prim}}\mu$ is a free $A$-submodule of $\mu$, whose induced order is 3. non-degenerate, and 4. can be presented by a finite quiver. Alternatively, the last condition iv) can be replaced by the two conditions 1. $M$ is a lower submodule of $M\tens_AK_A$, and 2. $(M,\leq)$ is finitely presented, that is, compact as an object of $\mathbf{PO}{\mathrm{Mod}}_A$. By lemma \[MON\_PRIM\_THM\], a projective module $\mu$ is free on $({\mathrm{Prim}}\mu,\leq)$, and by proposition \[MON\_POSET\_PROJ\] ${\mathrm{Prim}}\mu$ is itself finite projective. By proposition \[MON\_POMOD\_THM\], the order on ${\mathrm{Prim}}\mu$ is lower finite and lower saturated; loc. cit. also handles the converse. The remaining meat of the theorem is therefore that lower finiteness and lower saturation equate to conditions iii) and iv), and that iv) is equivalent to iva)$+$ivb). More precisely, lower saturation is equivalent to iva), and lower finiteness is equivalent to non-degeneracy together with finite presentation. ($\mathrm{ls}\Leftrightarrow\mathrm{iva})$) This follows from unravelling explicitly the definition of lower saturation: for all $x,y\in M$ and $f\in A$ such that $x\leq fy$, there exists a symbol $[x/f]\in M_{\leq y}$ such that $[x/f]f=x$; $$x\leq fy \quad \Rightarrow \quad \exists x/f \quad x/f\leq y.$$ When $M\cong \bigoplus_{i\in I}A_i$ is a direct sum of cyclic modules, it follows that every relation can be deduced from one of the form $\tilde x\leq y$ with $y$ a generator of its cyclic factor. When $f$ is invertible on $M$, the symbol $[x/f]$ automatically exists and is unique: $[x/f]=xf^{-1}$. More generally, when $f:M\rightarrow M$ is injective, then $[x/f]$ is unique if it exists, and the condition only asks that $$M_{\leq y}=M\cap M[f^{-1}]_{\leq y}.$$ When $A$ is a domain and $M$ is free, this works for all $f\neq 0$, hence the result. ($\mathrm{lf}\Leftrightarrow\mathrm{ndg}+\mathrm{fp}$) After lemmas \[MON\_DOMAIN\_SECOND\] and \[MON\_DOMAIN\_NONDEGEN\], it remains to show that lower finiteness implies non-degeneracy. Any degenerate relation $fx\leq x$ in $M$ entails an infinite sequence of relations $f^nx\leq x$. Since $M$ is free, we may localise our study to a cyclic factor and assume $M=A$, and by lower saturation, that $x=1$. The condition that the set $\{f^n\}_{n\in\N}$ be contained in a finite $A^+$-submodule is precisely that $f$ be integral over $A^+$. Since we have assumed that $A^+$ is integrally closed in $A$, non-degeneracy is a necessary condition for lower finiteness. \[MON\_DUALITY\]There is in general no good theory of duality for modules over $\F_1$-algebras - since $A^{\oplus n}\not\simeq A^n=\Hom(A^{\oplus n},A)$ as soon as $n>1$, free modules are typically not reflexive. However, in light of corollary \[MON\_DOMAIN\_FIRST\], at least for certain po-modules over a domain $A$ we can ‘borrow’ the theory of duality for finite projective $\B[A]$-modules, and define a restricted dual $$M^:= {\mathrm{Prim}}(\Hom(\B M,\B A)).$$ The result implies that if $M$ is a lower finite and lower saturated partially ordered, finite, free $A$-module, then so is $M^$, and that this operation is an involution (though still not actually a duality with respect to $\tens_A$) on such modules. We end this section with some counterexamples. A more involved argument in lemma \[MON\_POSET\_LEMMA\] shows that the result holds for $M$ with the divisibility order whenever $A$ is idempotent-free. The same line of thought yields a counterexample with idempotents to part ii) of proposition \[MON\_POSET\_PROJ\]. Let $$A=A^+:= \F_1[f_i,e_i]_{i=1,2} / (e_i^2e_j=e_i, e_if_i=f_j)_{i\neq j}.$$ The idempotent $\xi=e_1e_2$ fixes all elements except $0$ and $1$, and multiplication by $f_i$ is injective on its image - so induces an isomorphism between the fixed set of $\xi$ and the principal ideal $(f_1)=(f_2)$. Now define a module $M$ over $A$ with two generators and the relations $$(f_1,0)=(0,f_2),\quad (e_2,0)=(0,\xi),\quad (\xi,0)=(0,e_1).$$ Both of the generating cyclic submodules are free, but $M$ itself is not, so by proposition \[ALG\_CYCLIC\] it cannot be projective. Note also: - the fixed set of $\xi$ on $M$ is isomorphic to the fixed set of $\xi$ of $A$ itself; in particular, it is projective cyclic; - the base change of $A$ to $\Z$ splits as a product $\Z\times\Z[e_1^{\pm 1}]$, whereupon $M$ becomes the union of a trivial line bundle over the $\G_m$ part and a rank two free module over the point - in particular, it is projective. I claim that $\B(M;A)$ is a projective $\B[A]$-module. Indeed, a section to the given map $\B[A]^2\twoheadrightarrow\B(M)$ is given $$\sigma(1,0):=(1,0)\vee(0,e_1)$$ and vice versa. I leave it to the reader to check that this map is indeed a section and obeys the relations of $M$. When working with monoids one somewhat inevitably encounters pathological examples like this one, as these kinds of relations can arise after a simple sequence of relative tensor products between finite free monoids $\simeq \N^m$. Note that since finite $\B[A]$-modules no longer have finite underlying sets, they need not have primitive decompositions. For example, take $\Z_\vee^2/(e_1+\lambda\vee e_2=e_1)$, where $\lambda\leq 0$. The only primitives in this module are the translates of $e_2$; in particular, it is not generated by primitives. It follows that it cannot be embedded in a free module, and is in particular non-reflexive. Let $\Gamma\subseteq\Gamma^\prime$ be an extension of totally ordered Abelian groups. Then $\Gamma^\prime_\vee$ is a flat $\Gamma_\vee$-module. We will prove directly the equational criterion (prop. \[ALG\_EQN\]). Let $v:\Gamma_\vee^n\rightarrow\Gamma^\prime_\vee$ be a $\Gamma$-linear homomorphism, $\Gamma_\vee\rightrightarrows\Gamma^n_\vee$ a relation. These data are determined by a set of (up to) $n$ elements $\{v_i\}_{i=1}^n\subseteq\Gamma^\prime$ and a relation $$\max\{v_i+f_i\}=\max\{v_i+g_i\}_{i=1}^n\in \Gamma^\prime$$ with $(f_i),(g_i)\in\Gamma^n$. Let $i_f$, resp. $i_g$, be the index for which the maximum on the left (resp. right) is realised. If $i_f=i_g$, then $f_{i_f}=g_{i_g}$ and the projection $\Gamma^n\rightarrow\Gamma$ on the $i_f$th factor factors $v$. Otherwise, we may identify the $i_f$th and $i_g$th factor via the primitive of the given relation, and project out the other co-ordinates. On the other hand, if the rank of $\Gamma^\prime$ is strictly greater than that of $\Gamma$, then submodules generated by $n>1$ elements are in general infinitely presented, and so cannot be projective. For instance, this applies to the $\Z_\vee$-submodule of $\Z^2_\mathrm{lex}$ generated by $(0,0)$ and $(0,1)$. Polyhedra {#POLY} ========= In this section, we will work with additively notated monoids without zero. The associated fractional submodule functor $\B$ factors through adjoining zero. Cf. §\[ALG\_PTS\]. Weight polyhedra {#POLY_WEIGHT} ---------------- Let $H$ be a totally ordered group - for example, an additive subgroup of $\R$. We will study finite modules over the semifield $$H_\vee = \B[H;H^\circ] \cong H\sqcup\{-\infty\}$$ via partially ordered $H$-modules (or ‘$H$-acts’). This setting enjoys two substantial simplifications over the general case: - since $H$ is a group, lower saturation is automatic (cf. part iva of corollary \[MON\_DOMAIN\_FIRST\]); - since $H$ is totally ordered, non-degeneracy is automatic; and so all we have to worry about is finite presentation. When $M\simeq H^{\sqcup n}$ is free, the dual $H$-module $${\mathbb V}(M):=\Hom_H(M,H)\simeq H^n$$ is an $H$-affine space - that is, a torsor for $\Hom_\Z(\Lambda,H)$ with $\Lambda$ a free $\Z$-module of rank $n$ - whose co-ordinate functions, up to translation, are the elements of $M$. When $M$ is free as an $H$-module but carries a possibly non-trivial partial order, it embeds in ${\mathbb V}(M)$ as the subset $${\mathbb V}(M,\leq):=\Hom_{\mathbf{PO}{\mathrm{Mod}}_H}(M,H)$$ cut out by relations defining the partial order of $M$. When $(M,\leq)$ is finitely presented, it is a polyhedral subset defined by inequalities between co-ordinates. We formalise the sets arising in this way as follows: \[POLY\_WEIGHT\_DEF\]Let $\Phi\subset V$ be a root system in a Euclidean space with a weight lattice $V_\Z$. An $H$-integral weight polyhedron $(\Delta,N,V,\Phi)$ for $\Phi$ is a finite, non-empty intersection of half-spaces orthogonal to the roots, and having co-ordinates in $H$, in an affine space for $\Hom_\Z(V_\Z, H)$. A general linear $H$-integral weight polyhedron is a weight polyhedron with $\Phi$ the root system of the general linear group $\GL_n$ of a finite set $n$.[^2] The character lattice $V_\Z$ of $\GL_n$ has a canonical unordered basis $V_\Z\cong\Z^n$. Let us call the elements of this basis the fundamental weights. The roots are the differences of pairs of distinct fundamental weights. Any $\GL_n$ weight polyhedron $\Delta$ is invariant under the diagonal action of $H$; quotienting by this action one obtains an $\SL_n$ weight polyhedron $\P\Delta$. An affine map $(\Delta_1,N_1,V_1,\Phi_1)\rightarrow(\Delta_2,N_2,V_2,\Phi_2)$ of general linear weight polyhedra is an $H$-equivariant map $\psi:N_1\rightarrow N_2$ such that $\psi(\Delta_1)\subseteq\Delta_2$ and $d\psi:V_2\rightarrow V_1$ preserves the set of fundamental weights (i.e. is the linear extension of a map $I_2\rightarrow I_1$). An affine map of $\GL_n$ weight polyhedra preserves the centre $\G_m\subseteq\GL_n$ and descends under $\P$ to an affine map of $\SL_n$ weight polyhedra preserving the root system $\Phi\sqcup\{0\}$. The category of general linear weight polyhedra is denoted $\mathbf{Poly}_H^\GL$. By diagonal invariance of $\GL$ weight polytopes, the mapping sets $$\Aff_H^\GL(\Delta_1,\Delta_2) = \Hom_{{\mathbf{Poly}}_H^\GL}(\Delta_1,\Delta_2)$$ are also $H$-modules. By replacing $\Delta$ with its affine span, a $\GL_n$ weight polyhedron may always be assumed $n$-dimensional. Note that this definition is more relaxed than the usual notion of weight polytope, since there is no requirement of Weyl group symmetry. When the Weyl group does act, a weight polytope in the sense of definition \[POLY\_WEIGHT\_DEF\] is polar to the usual weight polytope. \[POLY\_WEIGHT\_EG\]If $\Delta_1$ is one-dimensional (i.e. a $\GL_1$ weight polyhedron), then $\Delta_1\simeq H$ as an $H$-act and the differential of any map $\Delta_1\rightarrow\Delta_2$ is fixed to be the map $\Z^{I_2}\rightarrow\Z$ sending all basis vectors to $1$. To specify such a map, it is therefore enough to say where a single point $0\in\Delta_1$ goes. In particular, $\Aff_H^\GL(H,\Delta_2)=\Delta_2$. Conversely, a map $\Delta_2\rightarrow\Delta_1$ is necessarily a co-ordinate projection. In particular, the set $\Aff_H^\GL(\Delta_2,H)=:\Aff_H^\GL(\Delta_2)$ of such is a finite $H$-module. #### Co-ordinate functions on a weight polyhedron Let $N$ be a torsor under $\Hom_\Z(V_\Z,H)$. Then in particular $N$ is a weight polyhedron, and $$\Aff_H^\GL(N):=\Hom_{{\mathbf{Poly}}_H^\GL}(N,H)\simeq H^{\sqcup n}$$ is the set of $H$-affine functions whose differentials are fundamental weights. It is a finite free $H$-module indexed by the set of fundamental weights, functorial for $H$-affine maps. Now let $\Delta\subseteq N(H)$ be a weight polyhedron, and suppose $\Delta$ is not contained in any proper affine subspace of $N$, i.e. $\dim\Delta=\dim N$. Then $\Aff_H^\GL(N)$ acquires a partial ordering based on the values of functions on $\Delta$. Let us call this partially ordered module $\Aff_H^\GL(\Delta)$. It is functorial for maps of weight polyhedra. This defines a contravariant functor $$\Aff_H^\GL:{\mathbf{Poly}}_H^\GL\rightarrow\mathbf{PO}{\mathrm{Mod}}^\mathrm{free}_H$$ into the category of partially ordered free $H$-modules. As an a posteriori consequence of theorem \[POLY\_WEIGHT\_THM\], we will see that in fact $\Aff_H^\GL(\Delta)$ is always finitely presented. #### Dual to a partially ordered $H$-module Dualising an object $(M,\leq)$ of $\mathbf{PO}{\mathrm{Mod}}_H$ with respect to the unit object yields an embedding ${\mathbb V}(M,\leq)\subseteq\Hom_H(M,H)$ into an $H$-affine space; we also saw that if $(M,\leq)$ is finitely presented, ${\mathbb V}(M,\leq)$ is a general linear weight polyhedron. A module homomorphism $M_1\rightarrow M_2$ yields a pullback map $\Hom_H(M_2,H)\rightarrow\Hom_H(M_1,H)$ of the ambient affine spaces whose differential is a permutation. Thus ${\mathbb V}$ defines a contravariant functor $$\mathbb V:\mathbf{PO}{\mathrm{Mod}}_H^\mathrm{pf/free}\rightarrow {\mathbf{Poly}}_H^\GL$$ on the category of finite free $H$-modules with finitely presented partial order. (Note that since $H$ is totally ordered, po-$H$-modules are automatically non-degenerate.) Evaluation gives us maps $$1\rightarrow \Aff_H^\GL{\mathbb V}, \quad 1 \rightarrow {\mathbb V}\Aff_H^\GL$$ that put ${\mathbb V}$ and $\Aff_H^\GL$ into adjunction. \[POLY\_WEIGHT\_THM\]The adjoint functors ${\mathbb V},\Aff_H^\GL$ restrict to an anti-equivalence $${\mathbf{Poly}}_H^\GL \cong \mathbf{PO}{\mathrm{Mod}}_H^\mathrm{pf/free}$$ between the category of $\GL$ weight polyhedra and the category of finitely presented and non-degenerate partially ordered free $H$-modules. Let $\Delta\subseteq N$ be a weight polyhedron. A monotone homomorphism $\Aff_H^\GL(\Delta)\rightarrow H$ is a point $p$ of $N$ such that for every pair of affine functions $F,G$ on $V$, $F\|_\Delta \leq G\|_\Delta$ implies $F(p)\leq G(p)$. Since $\Delta$ is cut out from $N$ by such inequalities, $p\in\Delta$. The more subtle part of the proof is to show that when $(M,\leq)$ is finitely presented, conversely, the elements of ${\mathbb V}(M)$ ‘separate points’ of $M$. More precisely: \[POLY\_WEIGHT\_MAINLEMMA\]Let $M\in\mathbf{PO}{\mathrm{Mod}}^\mathrm{free}_H$ be finitely presented. For any $F\not\leq G\in M$ there exists a monotone homomorphism $\phi_{F,G}:M\rightarrow H_\vee$ such that $\phi_{F,G}G=0$ and $\phi_{F,G} F>0$. By non-degeneracy, a homomorphism $M\rightarrow H$ is monotone if and only if it obeys the relations of the acyclic path set $\sh Q^\emptyset$ of a presenting quiver $\sh Q$. In other words, if $\phi_GG=0$, then we must produce for each cyclic factor $H_i$ an identification $\phi_i:H_i\tilde\rightarrow H$ such that $$\begin{aligned} \label{arooga} \phi_i\gamma G\leq 0\leq\phi_i\tilde\gamma G \end{aligned}$$ for all $\gamma$ in $\sh Q^\emptyset$ from $dG$ to $i$, resp. $\tilde\gamma$ from $i$ to $dG$. Variables whose differentials are not connected to $dG$ in $\sh Q$ may be set arbitrarily. We distinguish three cases for $F$: \[$dG\rightarrow dF$\] Since $H_\vee$ is totally ordered, we can define our identifications by $$\phi_i\left(\max_{\gamma:dG\rightarrow i}\gamma G\right):=0.$$ Since, by non-degeneracy, $\max_{\gamma:dG\rightarrow i}\gamma < \tilde\gamma G$ for any $\tilde\gamma:i\rightarrow dG$, the other relations $0<\phi_{G,i}\tilde\gamma G$ are also satisfied. \[$dG\not\rightarrow dF$, $dF\rightarrow dG$\] There exists an extension $\tilde\phi_i$ of $\phi_G$ to the factor of $F$. If $\tilde\phi_i(F)\leq 0$, replace it with $\phi_i=\tilde\phi_i-\tilde\phi_i(F)+\epsilon$ where $\epsilon>0$. \[$dF\not\leftrightarrow dG$\] We may set $\phi_iF>0$ arbitrarily. It follows that if $F\leq G$ on ${\mathbb V}(M)\subseteq H^n$, then in fact $F\leq G$ in $M$. In other words, the unit $$M\rightarrow \Aff_H^\GL({\mathbb V}(M))$$ is an isomorphism of partially ordered modules. In particular, the po-$H$-module associated to a weight polyhedron is finitely presented. Combining this with the classification theorem \[MON\_DOMAIN\_FIRST\], we obtain a faithful and conservative contravariant embedding $${\mathbf{Poly}}_H^\GL\hookrightarrow{\mathrm{Mod}}_{H_\vee}$$ that surjects onto the set of non-zero finite projectives. #### Extended weight polyhedron By duality of finite projective $H_\vee$-modules, the category of weight polyhedra can also be embedded covariantly into ${\mathrm{Mod}}_{H_\vee}$. This inclusion can be realised geometrically by adding strata at infinity directly to a weight polyhedron $\Delta$. A one-dimensional $H$-affine space $N$ has a natural extension to an invertible $H_\vee$-module $\bar N:=N\sqcup\{-\infty\}$. The same logic allows us to ‘compactify’ a finite product of lines to a free $H_\vee$-module. In terms of functions, it admits the description $$\bar N=\bar{\mathbb V}(\Aff_H^\GL(N)):=\Hom_H(\Aff_H^\GL(N),H_\vee);$$ in particular, $\bar N$ is a free $H_\vee$-module. If now $\Delta\subseteq N$ is a $\GL$ weight polyhedron, then we can define the ‘closure’ $\bar\Delta$ of $\Delta/\bar N$ by extending the inequalities defining $\Delta$ to $\bar N$: $$\bar\Delta=\bar{\mathbb V}(\Aff_H^\GL(\Delta),\leq) := \Hom_{\mathbf{PO}{\mathrm{Mod}}_H}(\Aff_H^\GL(\Delta),H_\vee).$$ If $H=\R$, then this is the same as the topological closure of $\Delta$ in $\bar N$ with respect to the order topology on $\R$. The resulting object is an extended general linear weight polyhedron. If $\Delta$ is laterally compact - meaning that the associated $\SL_n$ polyhedron $\P(\Delta)$ is bounded - then the extension only affixes the single point $-\infty$. More generally, each stratum of $\bar\Delta$ is a $\GL$ weight polyhedron in the stratum of $\bar N$ that contains it. In general, extended weight polyhedra are dual to finitely presented po-modules over the pointed version $\F_1[H]$ of $H$. For this reason, we adopt the convention that $\{-\infty\}$ is also an extended weight polyhedron. #### Mapping sets The covariant embedding of ${\mathbf{Poly}}_H^\GL$ in ${\mathrm{Mod}}_{H_\vee}$ allows us to give a more geometric description of the $H_\vee$-module of homomorphisms between projective modules. We define an affine map $\bar\Delta_1\rightarrow\bar\Delta_2$ between extended weight polyhedra to be an affine map - in the sense of the category ${\mathbf{Poly}}_H^\GL$ - from $\Delta_1$ into a stratum of $\bar\Delta_2$, that is, $$\Hom_H^\GL(\bar\Delta_1,\bar\Delta_2):=\coprod_{\Delta_2/\rho}\Hom_H^\GL(\Delta_1,\Delta_2/\rho)$$ where $\rho$ ranges over faces of the recession cone of $N_-\cap\Delta_2$, $N_-$ a negative orthant in $N_2$. Such maps are exactly those arising from homomorphisms of the extended dual $\F_1[H]$-modules $\Hom_H^\GL(\Delta_i,H)\tens_H\F_1[H]$. The inclusion $$\Hom_H\left(\Aff_H^\GL(\Delta_2),\hspace{3pt} \Aff_H^\GL(\Delta_1)\tens_H\F_1[H]\right)\subseteq \Hom_{H_\vee}(\B\Aff_H^\GL(\Delta_2),\hspace{3pt}\B\Aff_H^\GL(\Delta_1))$$ is generating as a $\B$-module; applying theorem \[POLY\_WEIGHT\_THM\] on the left and duality for projective $\B[A]$-modules on the right, the same is true of $$\Hom_H^\GL(\bar\Delta_1,\bar\Delta_2)\subseteq \Hom_{H_\vee}(\bar\Delta_1,\bar\Delta_2).$$ It follows that $\Hom_{H_\vee}(\bar\Delta_1,\bar\Delta_2)$ is the set of convex piecewise affine maps $\Delta_1\rightarrow\bar\Delta_2$ whose linear parts preserve the fundamental weights. \[POLY\_WEIGHT\_COR\]The duality of theorem \[POLY\_WEIGHT\_THM\] extends $\B$-linearly to a duality $${\mathrm{Mod}}_{H_\vee}^\mathrm{f/proj}\cong {\mathbf{Poly}}_{H,\vee}^\GL$$ between the category of finite projective $H_\vee$-modules and the category of extended $\GL$ weight polyhedra and convex, piecewise-affine maps whose linear parts are fundamental weights. Bundles of weight polyhedra {#POLY_FAM} --------------------------- Let $H\subseteq\R$ be an additive subgroup, and let $\Delta\subseteq N$ be an $H$-rational, strongly convex polyhedron in an $H$-affine space $N$. More precisely, $\Delta\subseteq N(\Q H)$ is a subset defined over the divisible hull $\Q H$ of $H$ cut out by finitely many affine equations over $H$ and having at least one vertex. By multiplying out denominators, the condition of $H$-rationality is the same as rationality over $\Q H$.[^3] To $\Delta$ we will associate the monoid pair $$\Aff^+(\Delta,H) \subseteq\Aff(\Delta,H)$$ of $H$-affine functions $\Aff(\Delta,H)$ on $\Delta$ with integral slopes and values in $H$, with integers the (saturated, sharp) submonoid $\Aff^+(\Delta,H)$ of functions bounded above by zero. We do not add strata to $\Delta$ at infinity - but see §\[POLY\_EXTRA\]. Since $\Aff(\Delta,H)=\Aff(N,H)$ is a group, we retain the benifit that we enjoyed in §\[POLY\_WEIGHT\] that po-modules are automatically lower saturated. However, non-degeneracy is no longer automatic. We will be interested in the ‘polyhedral’ semiring ${\mathrm{CPA}}(\Delta,H_\vee)$ of convex piecewise-affine functions on $\Delta$ with integer slopes and values in $H_\vee$. This is generated as a $\B$-module by the pair $\Aff(\Delta,H)$, and is closely related to the free semiring. Let $\Delta\subset N$ be a strongly convex $H$-rational polyhedron. A convex family of weight polyhedra $E$ over $\Delta$ is a convex polyhedron $\Delta_E$ inside an $H$-affine space $N_E/N$ such that for each point $p\in\Delta(H^\prime)$ defined over an ordered extension $H^\prime$ of $H$, $\pi^{-1}(p)$ is an $H^\prime$-rational $\GL$ weight polyhedron. It is enough to check for finite extensions of $H^\prime$. The dimension of the fibre at any interior point is called the rank of the family. A morphism of convex families of weight polyhedra is an $H$-affine map over $N$ that restricts to a morphism in ${\mathbf{Poly}}_{H^\prime}^\GL$ on the fibre over any $H^\prime$-point. Note that $\pi$ is necessarily surjective, and the rank does not depend on the point. #### Dual to a partially ordered $\Aff$-module Let us abbreviate $A=(\Aff(\Delta,H),\Aff^+(\Delta,H))$. To a free $A$-module $M$, we associate for every finite extension $H^\prime$ of $H$ a set of $H^\prime$-points $${\mathbb V}(M)(H^\prime):=\coprod_{p\in\Delta(H^\prime)}\Hom_A(M,H^\prime_p)$$ of pairs $(p,\phi)$ consisting of an $H$-algebra map $p:A\rightarrow H^\prime$ and an $A$-module homomorphism $\phi:M\rightarrow H^\prime$, where $H^\prime$ is considered an $A$-module via $p$. For any fixed $H^\prime$, ${\mathbb V}(M)(H^\prime)$ is a trivial $H^\prime$-affine space bundle over $\Delta(H^\prime)$. If $M$ carries a partial order, we define $${\mathbb V}(M,\leq)(H^\prime):=\coprod_{p\in\Delta(H^\prime)} \Hom_{\mathbf{PO}{\mathrm{Mod}}_A}(M,H_p^\prime)\subseteq {\mathbb V}(M)(H^\prime)$$ to be the subset of pairs $(p,\phi)$ for which $\phi$ is monotone. #### Co-ordinate functions on families of weight polyhedra Let $\Delta\times H$ denote the trivial affine line bundle over $\Delta$ - more precisely, the affine bundle whose $H^\prime$-points are $\Delta(H^\prime)\times H^\prime$. If $E$ is a convex family of weight polyhedra, then the set $\Aff_H^\GL(E)$ of affine functions $E\rightarrow\Delta\times H$ whose relative differentials are fundamental weights has a natural structure of a partially ordered, finite, free module over $\Aff(\Delta,H)$. One obtains a contravariant adjunction $$\Aff^\GL_\Delta:{\mathbf{Poly}}_\Delta^\GL \leftrightarrows \mathbf{PO}{\mathrm{Mod}}_{\Aff(\Delta,H)}^\mathrm{free}:{\mathbb V}$$ that restricts to an equivalence between the category of affine bundles over $\Delta$ on the left and of free (unordered) $A$-modules on the right. However, the analogue of theorem \[POLY\_WEIGHT\_THM\] fails. More precisely, lemma \[POLY\_WEIGHT\_MAINLEMMA\] is false in this setting: \[POLY\_FAM\_EG\]Let $\Delta=[a,b]$ be an interval. Then $A^+$ is generated as a monoid by $-1$ and the functions $X-b,a-X$. The free semiring on $A$ is not equivalent to the semiring of convex piecewise-affine functions: $0\not\leq X\vee -X$ (provided $a<0<b$). In other words, $\B[A;A^+]$ is abnormal (unless $a=b$). We may thus define a rank two counterexample presented by the quiver $$\xymatrix{ \bullet \ar@/^/[r]^{X}\ar@/_/[r]_{-X} & \bullet }$$ which fails to see the relation $(-\infty,0)\leq(0,-\infty)$ which must hold on any map into a totally ordered group. This can be viewed as a problem on the $\B[A]$-module side and can be solved by considering instead its ‘normalisation’ as in [@norm]. In geometric terms, this is achieved by replacing $\mathbf{PO}{\mathrm{Mod}}_A$ with the associated stack for the rigid analytic topology $\Spec(A;A^+)$. #### Normal projectives Here we carry out a sketch of the programme above, restricting attention to the category ${\mathrm{Mod}}_{\B[A]}^\nu$ of normal $\B[A]$-modules. This is the same as the category of modules over ${}^\nu\B[A]$, the normalisation of $\B[A]$, which in the case $A=\Aff_\Z(\Delta,H)$ is just the semiring ${\mathrm{CPA}}_\Z(\Delta,H)$ of convex, piecewise-affine functions on $\Delta$. This is the only part of the paper relying on [@norm]. Let $A$ be an $\F_1$-domain. Every projective ${}^\nu\B[A]$-module is the normalisation of a $\B[A]$-module free on a lower finite, lower saturated partially ordered free $A$-module. The proof follows exactly the lines of §\[MON\]. ${\mathrm{Prim}}\mu$ is an $A$-module: The primitives of ${}^\nu\B[A]$ are again $A$. It follows from lemma \[B\_PRIM\_SUB\] that projective modules over ${}^\nu\B[A]$ also have primitive decompositions. We may therefore repeat the argument of §\[MON\_PRIM\] to conclude that $$\B({\mathrm{Prim}}\mu,\leq)\oplus_{\B[A]}{}^\nu\B[A]\rightarrow \mu$$ is an isomorphism for any projective $\mu\in{\mathrm{Mod}}^\nu_{\B[A]}$. Indeed, by projectivity of $\mu$ there is at least a $\B$-linear section $\sigma$; composing this with a set-theoretic section of $\B[A]\rightarrow{}^\nu\B[A]$ gives a section of $\B({\mathrm{Prim}}\mu,\leq)\rightarrow\mu$. By the lemma \[B\_PRIM\_SECTION\] on sections over primitives, this necessarily commutes with the two embeddings of ${\mathrm{Prim}}\mu$. It follows that $\sigma$ is an isomorphism. This part of the proof does not require $A$ to be a domain. The remainder of the argument depends crucially on the fact that a finitely generated $A^+$-module has a unique minimal set of generators. With a little work, the same holds true in the integrally closed regime as well: \[POLY\_FAM\_CRUX\]Let $A$ be an $\F_1$-field, $M$ a free $A$-module. Let $K$ be the integral closure of a finitely generated $A^+$-submodule of $M$. There is a unique minimal set of generators of $K$ as an integrally closed fractional submodule. We write in additive notation. By intersecting $K$ with the cyclic factors of $M$ and choosing a trivialisation, we may reduce to the case $M=A$. In this case, integral closure of $K$ coincides with convexity [@norm lemma 5.4]. Let $(X_i)_{i\in I}$ and $(Y_j)_{j\in J}$ be two minimal sets of generators of $K$ as a convex $A^+$-submodule. Fixing an index $0\in I$, this give us equations $$\begin{aligned} \nonumber nX_0=\sum_{j\in J}n_jY_j,&\quad n=\sum_{j\in J}n_j \\ \nonumber m_jY_j=\sum_{i\in I}m_{ji}X_i,&\quad m_j=\sum_{i\in I}m_{ji} \end{aligned}$$ which we will use to show that $X_0\in\{Y_j\}_{j\in J}$. Indeed, $$\prod_{j\in J_0}m_jnX_0=\sum_{j\in J_0}n_j\prod_{\ell\neq j}m_\ell\sum_{i\in I}m_{j i}X_i = \sum_{i\in I}\left(\sum_{j\in J_0}n_j\prod_{\ell\neq j}m_\ell m_{j i}\right)X_i$$ so either $X_0$ can be eliminated from its generating set by cancelling it from the right, or $$\sum_{j\in J}n_jm_{j 0}\prod_{\ell\neq j}m_\ell =1;\qquad \sum_{j\in J}n_jm_{j i}\prod_{\ell\neq j}m_\ell =0,\quad i\neq0.$$ Since all $m_\ell$ are non-zero, there’s exactly one index $j\in J$ for which $n_j$ and $m_{j0}$ are both non-vanishing, and $m_{ji}=0$ for all $i\neq 0$. Thus $m_j=m_{j0}=1$ and so $Y_j=X_0$. $({\mathrm{Prim}}\mu,\leq)$ is lower finite and lower saturated. Since $\mu$ is projective, there is a ${}^\nu\B[A]$-linear section to $${}^\nu\B({\mathrm{Prim}}\mu;A^+)\rightarrow {}^\nu\B({\mathrm{Prim}}\mu,\leq).$$ The proof of the monotone section lemma \[B\_POSET\_SECTION\] applies, in light of lemma \[POLY\_FAM\_CRUX\], and shows that the section is automatically a right adjoint. ${\mathrm{Prim}}\mu$ is free. We need clarify only two points: - to prove that the submodule spanned by each generator of $M$ was free, we used the fact that $\B$ is strongly monoidal and commutes with equalisers. These properties also hold for ${}^\nu\B = {}^\nu\B[A]\oplus_{\B[A]}(-)$; the first is clear, and the second follows from the fact that the normalisation of a semiring is always flat. - The crucial lemma \[MON\_DOMAIN\_UNIQUE\_GENS\] that $M$ is a direct sum of cyclics uses the uniqueness of primitive generators for fractional submodules of a free cyclic module, and so passes with another application of lemma \[POLY\_FAM\_CRUX\]. That said, the proof of proposition \[MON\_POSET\_PROJ\] runs verbatim. Every projective ${}^\nu\B[A]$-module is the normalisation of a projective $\B[A]$-module. Not every partially order on a free $A$-module $M$ is induced from its inclusion into the normalisation of $\B(M,\leq)$, as example \[POLY\_FAM\_EG\] shows. However, since $\Hom(\alpha,H_\vee)\tilde\rightarrow\Hom({}^\nu\alpha,H_\vee)$ for any semiring $\alpha$ and totally ordered group $H$ [@norm Prop. 1.2.ii)], the geometric dual ${\mathbb V}$ factors through this replacement. The missing element of the proof is to show that conversely, $M\rightarrow \Aff_\Delta^\GL({\mathbb V}(M,\leq))$ is an order isomorphism for $M$ the module of primitives in a normal projective $\B[A]$-module. This follows from a corrected version of lemma \[POLY\_WEIGHT\_MAINLEMMA\]: \[POLY\_FAM\_MAINLEMMA\]Let $M\in\mathbf{PO}{\mathrm{Mod}}^\mathrm{free}_A$ be finitely presented, and suppose the partial order on $M$ is that induced by its inclusion into $\B(M,\leq)$. For any $F\not\leq G\in M$ there exists a $\phi_{F,G}:M\rightarrow H$ such that $\phi_{F,G}G=0$ and $\phi_{F,G} F >0$. By the same argument as in lemma \[POLY\_WEIGHT\_MAINLEMMA\], it will be enough to find an identification $A_{dF}\tilde\rightarrow A$ and algebra homomorphism $A\rightarrow H$ satisfying the inequalities \[arooga\] and $\phi_{F,G}>0$. By non-degeneracy, we may assume $dF\neq dG$. We have $F\not\leq\bigvee_{\gamma:dG\rightarrow dF}\gamma G$ in ${}^\nu\B(A_{dF})$. Since ${}^\nu\B(A_{dF})$ admits a representation as semiring of convex piecewise-affine functions ([@norm] ex. 5.6), there is some point $q\in\Delta(\Q H)$ at which $\bigvee_\gamma\gamma G(q)<F(q)$. If $\gamma_0$ is the path achieving the maximum in this formula, then identifying $A_{dF}$ with $A$ by setting $\gamma_0G=0$ gives us the desired functional. Finally, following the same line of argumentaion preceding corollary \[POLY\_WEIGHT\_COR\], we extend our definitions to a category of families of extended $\GL$ weight polyhedra $\bar\Delta_E\subseteq\bar N_E\twoheadrightarrow\Delta$ and piecewise-affine maps, thereby deriving an analogue of loc. cit.: \[POLY\_FAM\_COR\]The adjunction between $\Aff_\Delta^\GL$ and ${\mathbb V}$ extends $\B$-linearly to a duality $${\mathbf{Poly}}_{\Delta,\vee}^\GL \cong {\mathrm{Mod}}_{{\mathrm{CPA}}(\Delta)}^{\mathrm{fproj}}$$ between the category of finite projective $\mathrm{CPA}(\Delta)$-modules and the category of convex families of extended $\GL$ weight polyhedra over $\Delta$ with convex piecewise-affine maps whose vertical linear parts are fundamental weights. Extensions {#POLY_EXTRA} ---------- The results of §\[POLY\_FAM\] were couched in the setting of $\F_1$-fields, but using corollary \[MON\_DOMAIN\_FIRST\] it can easily be extended to the case of domains: we simply need to identify the lower free $A$-structures on partially ordered $K_A$-modules. We obtain a supply of $\F_1$-domains by partially compactifying our polyhedra at infinity. This means that we replace $\Aff(\Delta,H)$ with the submonoid $\Aff(\bar\Delta,H)$ of functions bounded above on $\Delta$. We may also define various partial compactifications of $\Delta$ dual to intermediate saturated submonoids $\Aff(\bar\Delta)\subseteq A\subseteq\Aff(\Delta)$. A free $A$-structure on a module $M$ over $K_A$ is the same data as a reduction of the structure group of ${\mathbb V}(M)$ from $\Aff(\Delta,H)$ to the group of bounded affine functions. For instance, one may fix a trivialisation ${\mathbb V}(M)\simeq \Delta\times H^n$ (which even determines an $A^+$-structure). The $A$-structure $M_A$ is then the set of functions on ${\mathbb V}(M)$ bounded above by some co-ordinate function of the trivialisation. Such a reduction defines a lower $A$-structure on $M$ if and only if every function bounded above on ${\mathbb V}(M,\leq)$ is bounded above on all of ${\mathbb V}(M)$. This is the case if and only if there exists a constant subset $$K\times\Delta\subseteq{\mathbb V}(M,\leq)\subseteq{\mathbb V}(M)$$ whose fibre $K\subseteq H^n$ has non-empty interior. More generally, if $\bar\Delta$ is not compact, such subsets must exist on rays approaching the boundary. \[POLY\_EXTRA\_EG\]Let $\bar\Delta=[-\infty,0]$ with co-ordinate $X$, and let $E$ be the rank two family of polyhedra cut out by the equation $Y_1\leq Y_2+X$. Its restriction to $X=-\infty$ is free of rank one. The dual module $\Aff^\GL_\Delta(E)$ is not lower finite because of the function $Y_1-X$, and so this does not correspond to a projective ${\mathrm{CPA}}_\Z(\bar\Delta,H)$-module. However, its base change to $\Delta:=(-\infty,0]$ is projective; it is the module presented by the quiver $$\xymatrix{\bullet & \bullet \ar[l]_{-X} }$$ It is trivialised by the change of basis $Y_2\leftrightarrow Y_2+X$; in fact, it is the pullback of a projective $\B$-module. This provides us with a trivial, and in particular projective, extension over $-\infty$. If we drop the non-emptiness requirement in the definition of weight polyhedra - and hence that $E\rightarrow\Delta$ be surjective - we obtain a class of modules that can be obtained by ‘pushing forward’ projective modules from sub-polyhedra. More interestingly, by pushing forward projective modules from boundary strata one can construct modules in polyhedra whose rank jumps up at infinity. The construction of example \[POLY\_EXTRA\_EG\] gives also families whose rank decreases at the boundary. Finally, it will be important in the future to understand modules that correspond to families of ‘valuated matroids’ - but that is another story for another day. [^1]: This result is closely related to the more specific theorem 1.5 of [@crumpy], which states that every projective $\R_\vee$-submodule is isomorphic to a submodule of $\R_\vee^n$ closed under co-ordinate-wise minimum, as well as maximum. By [@Xtracrumpy Thm. B], such sets are automatically convex polyhedra, and it is an elementary matter to observe what kinds of supporting half-spaces are allowed. Moreover, [@crumpy] even provides criteria to determine when a given submodule of $\R_\vee^n$ is projective. [^2]: That is, the general linear group of a vector space with a direct sum decomposition into lines indexed by $n$. There is then a canonical Cartan subgroup which is naturally identified with $\G_m^n$. [^3]: Beware that this category of polytopes differs from the category ${\mathbf{Poly}}_H^\GL$ discussed in §\[POLY\_WEIGHT\] in several key ways. First, by $H$-invariance $\GL$ weight polyhedra are never strongly convex. Second, the co-ordinates of the supporting half-spaces for a weight polyhedron are required to lie in $H$, rather than merely in $\Q H$. Finally, while it is always natural to add strata at (minus) infinity to a weight polyhedron, in this section for simplicity we will consider our base polyhedra to be ‘punctured’ at infinity.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Deep models have achieved impressive performance for face hallucination tasks. However, we observe that directly feeding the hallucinated facial images into recognition models can even degrade the recognition performance despite the much better visualization quality. In this paper, we address this problem by jointly learning a deep model for two tasks, i.e. face hallucination and recognition. In particular, we design an end-to-end deep convolution network with hallucination sub-network cascaded by recognition sub-network. The recognition sub-network are responsible for producing discriminative feature representations using the hallucinated images as inputs generated by hallucination sub-network. During training, we feed LR facial images into the network and optimize the parameters by minimizing two loss items, i.e. 1) face hallucination loss measured by the pixel wise difference between the ground truth HR images and network-generated images; and 2) verification loss which is measured by the classification error and intra-class distance. We extensively evaluate our method on LFW and YTF datasets. The experimental results show that our method can achieve recognition accuracy $97.95\%$ on 4x down-sampled LFW testing set, outperforming the accuracy $96.35\%$ of conventional face recognition model. And on the more challenging YTF dataset, we achieve recognition accuracy $90.65\%$, a margin over the recognition accuracy $89.45\%$ obtained by conventional face recognition model on the 4x down-sampled version.' author: - \| Junyu Wu\ Sun Yat-sen University\ [[email protected]]{} - \| Shengyong Ding\ Sun Yat-sen University\ [[email protected]]{} - \| Wei Xu\ Sun Yat-sen University\ [[email protected]]{} - \| Hongyang Chao\ Sun Yat-sen University\ [[email protected]]{} bibliography: - 'egbib.bib' title: Deep Joint Face Hallucination and Recognition --- Introduction ============ Face hallucination and recognition are critical components for a lot of applications, e.g. law enforcement and video surveillance. Face hallucination aims at producing HR (high-resolution) facial images from LR (low-resolution) images [@DBLP:journals/ijcv/LiuSF07]. Face recognition targets at verifying whether two facial images are from the same identity by designing discriminative features and similarities [@turk1991eigenfaces]. Empirical studies [@Lui2009A] in face recognition proved that a minimum face resolution between $32 \times 32$ and $64 \times 64$ is required for stand-alone recognition algorithms. [@DBLP:journals/pami/TorralbaFF08] reported a significant performance drop when the image resolution is decreased below $32 \times 32$ pixels. It is natural to expect that hallucinated face images can improve the recognition performance for LR facial images. Unfortunately, we find that this expectation does not hold in a lot of cases. As an example, Figure \[fig:srexample\] shows typical LR versions of LFW [@LFWTech] and its hallucinated counterparts generated by SRCNN [@srcnn]. We can clearly see that hallucinated versions have much better details and sharpness. However, feeding the hallucinated versions to a state-of-the-art recognition model can even degrade the recognition performance compared with the LR versions (from $96.35\%$ to $96.30\%$). ![Images generated by SRCNN.[]{data-label="fig:srexample"}](figures/srexample.png){width="40.00000%"} [@DBLP:conf/pricai/ZhangGLC16] reported similar conclusion: SR algorithms may perform poorly on recognition task since SR algorithms focus more on visual enhancement rather than classification accuracy. Considering the SR model and recognition model are trained separately, this phenomenon is not hard to be explained as each model has no signals or feedbacks from the other one during the training. Thus we propose a novel method to jointly optimize these two models under a unified convolutional neural network. Our Joint Model is based on an end-to-end CNN which can be seen as composed of two sub-networks, i.e. hallucination sub-network followed by recognition sub-network. During testing, one LR image is fed to the end-to-end network so that the hallucination sub-network produces a hallucinated facial image (as intermediate feature maps). Then this hallucinated image is fed to the recognition sub-network to generate a representation vector for recognition. In order to jointly solve these two tasks, LR face images are provided with its HR versions as well as their identities in the training stage. With these enriched training samples, we introduce two loss items to solve the parameters, i.e. hallucination loss and recognition loss. The hallucination loss is defined as the squared difference between the generated image and ground truth HR. The recognition loss follows the recently published literature [@centerloss] which is defined as the weighted sum of classification error and intra-class distance (the distance between each sample and its center in the feature space). Intuitively, classification error is to separate different classes as far as possible while the intra-class distance is to shrink the samples of one class. To the best of our knowledge, there are few works studying the joint learning of hallucination and recognition for face images. The most similar work to ours is proposed by Z Wang et al [@studyvlr]. In this work, the authors first train a SR network. Then two fully-connected layers are stacked on this pretrained SR network to learn a classification model. During the learning of this classification model, the super resolution loss is not applied anymore, i.e. SR module only acts as pretraining rather than joint supervision. In contrast to this work, we focus on face domain and extensively study the joint effect of SR and recognition using state-of-the-art network architectures to rigorously evaluate the improvements brought by the Joint Model. We extensively evaluate our method on public dataset, i.e. LFW and the YTF. We obtain a set of models for thorough comparison to demonstrate the effect of the Joint Model. Our experimental results show that the result of Joint Model outperforms the independently trained models by a margin of $0.63\%$ on LFW. In summary, our contributions are mainly two folded: - A joint end-to-end model which simultaneously solve hallucination task and recognition task. - Extensive performance reports of hallucination and recognition performance on facial dataset. Related Work ============ The related work to our method can be roughly divided into 3 groups as follows. Face Recognition ---------------- The shallow models, e.g. Eigen face [@turk1991eigenfaces], Fisher Face [@belhumeur1997eigenfaces], and Gabor based LDA [@liu2002gabor], and LBP based LDA [@li2007illumination] usually rely on handcrafted features and are evaluated on early datasets in controlled environments. Recently, a set of deep face models have been proposed and greatly advanced the progress [@taigman2014deepface; @sun2014deep; @yi2014learning; @DBLP:FaceNet]. DeepID [@sun2014deep] uses a set of small networks with each network observing a patch of the face region for recognition. FaceNet [@DBLP:FaceNet] is another deep face model proposed recently, which are trained by relative distance constraints with one large network. Using a huge dataset, FaceNet achieves 99.6% recognition rate on LFW. [@DBLP:conf/eccv/WenZL016] proposed a loss function (called center loss) to minimize the intra-class distances of the deep features, and achieved 99.2% recognition rate on LFW using web-collected training data. Super Resolution and Face Hallucination --------------------------------------- A category of state-of-art SR approaches [@DBLP:journals/ijcv/FreemanPC00; @DBLP:conf/cvpr/ChangYX04; @DBLP:conf/cvpr/YangWHM08] learn a mapping between LR / HR patches. There have been some studies of using deep learning techniques for SR [@srcnn] [@vdsr]. SRCNN [@srcnn] is a representative state-of-art method for deep learning based SR approach, which directly models HR images with 3 layers: patch extraction / representation, non-linear mapping, and reconstruction. [@vdsr] proposed a Very Deep Super-Resolution convolutional network, modeling high frequency information with a 20 weighted $3 \times 3$ layers network. Conventional hallucination methods [@DBLP:conf/fgr/BakerK00; @DBLP:journals/tsmc/WangT05] are often designed for controlled settings and cannot handle varying conditions. Deep models are also applied to face hallucination tasks [@fhwild; @DBLP:conf/eccv/ZhuLLT16]. [@fhwild] proposed a Bi-channel Convolutional Neural Network, which extracts robust face representations from raw input by using deep convolutional network, then adaptively integrates 2 channels of information to predict the HR image. Low Resolution Face Recognition ------------------------------- Low-resolution face recognition (LR FR) aims to recognize faces from small size or poor quality images with varying pose, illumination, expression, etc. [@verylrfrproblem] reported a degradation of the recognition performance when face regions became smaller than $16 \times 16$. [@studyvlr] proposed a Partially Coupled Super-Resolution Networks (PCSRN), as the pre-training part of recognition model. Joint Model =========== We use one end-to-end network to jointly solve face hallucination and recognition. Figure \[fig:DARI\] illustrates the overall principle. This network consists of two parts, i.e. face hallucination layers and recognition layers, which will be abbreviated as SRNET and FRNET respectively for convenience. In testing stage, the hallucination layers produce a high resolution facial image ${I^{h}}$ for a low resolution facial image ${I^{l}}$. The recognition layers then generates face representations ${x}$ using ${I^{h}}$ as input which serves face recognition task. As these two parts are cascaded, these two steps will be executed by one forward propagation, i.e. in end-to-end fashion. ![image](figures/CNN_2.pdf){width="80.00000%"} An intuitive approach to implement this end-to-end network is cascade one well trained SRNET and one FRNET. However, as we aforementioned, such direct cascading will even degrade the overall recognition performance despite the output of the SRNET has better visualization and PSNR since the well trained FRNET has never seen samples generated by SRNET. In order to address this problem, we propose to jointly optimize these two networks so that each network can benefit from the other one. Figure \[fig:DARI\] illustrates the overall principle. Given a set of low resolution facial images $I_i^l$ with their high resolution versions $\tilde I_i^h$ and the labels of identities $c_i$, the end-to-end model produces predicted high resolution facial images $I_i^h$ by SRNET and feature vectors $x_i$ by FRNET. This end-to-end network is jointly optimized so that $I_i^h$ are as close as possible to $\tilde I_i^h$ and $x_i$ should be able to separate different identities in the feature space. These two constraints can be further formulated as two loss items $L_h$ and $L_r$ in the overall objective function $L$ as follows where $W_h$ denotes the parameter set of SRNET and $W_r$ denotes the parameter set of FRNET with $\alpha$ and $\beta$ controls the weight of these two items: $$L=\alpha L_h(I_i^l;W_h)+\beta L_r(I_i^l;W_h,W_r)$$ Note $L_r$ depends on both parameter set $W_r$ and $W_h$ as FRNET uses the outputs of SRNET as inputs. For the loss item $L_h$, we use the pixel wise difference between $I_i^h$ and $\tilde I_i^h$ as below: $$L_h=\Sigma_{i=1}^{m}\|\|I_i^h-\tilde I_i^h\|\|^2$$ And for the recognition loss, we want to obtain representations that can discriminate different identities in the feature space under some similarity measure. We follow the recently published method, named center loss to model this constraint. In particular, this loss includes two items, i.e. classification error $L_c$ and the center loss $L_d$ which is defined as the mean intra-class distance between the samples and their centers. We use $W_j$ to denote the $j$th column of the softmax weight matrix $W$ and $b_j$ for the bias terms, then $L_c$ can be defined as below where $n$ is the number of training samples: $$L_c=-\Sigma_{i=1}^{n} \log \frac{e^{W_{c_i} x_i+b_{c_i}}}{\Sigma_{j=1}^{n}e^{W_{j} x_{j}+b_{j}}}$$ By using $m_{c_i}$ to represent the center of class $c_i$ in the feature space, $L_d$ is then defined as: $$L_d=\Sigma_{i=1}^{n}\|\|x_i-m_{c_i}\|\|^2$$ In order to balance the softmax loss $L_c$ and center loss $L_d$, we can introduce weight parameters $\alpha$, $\beta_1$ and $\beta_2$ and define the overall loss function as: $$L= \alpha L_h+\beta_1 L_c+\beta_2 L_d$$ In the next section, we will give a method to solve this model in the end-to-end fashion. Optimization ------------ In this section, we show how to jointly solve our end-to-end model. As the softmax and center-loss are introduced, we use $W_s$ and $M=[m_1, m_2,..., m_k]$ to denote the softmax parameter set and center vectors of $k$ classes and give a parameterized version of the loss function to show the dependency of different items on the parameter set. $$\begin{aligned} L(I_i^l;W_h,W_r,W_s,M) = {} & \alpha L_h(I_i^l;W_h) \\ & + \beta_1 L_c(I_i^l;W_h,W_r,W_s) \\ & +\beta_2 L_d(I_i^l;W_h,W_r,M) \end{aligned}$$ Due to the non-convexity of the loss function, we apply gradient descent algorithm to find the local minimum, i.e. calculating the gradient $\nabla W=[\nabla W_h \nabla W_r \nabla W_s]$ and update $W$ by this gradient with a learning rate iteratively. Note the update of $M$ is replaced by an approximate mechanism as adopted in literature [@centerloss] rather than the gradient method. Graident with respect to $W_r$ ($\nabla W_r$): This gradient is relatively simple and can be obtained by running the standard back propagation algorithm after we calculate $\frac{\partial L}{\partial x_i}$ as the following chain rule holds: $$\frac{\partial L}{\partial W_r}=\Sigma_{i=1}^{m}\frac{\partial L}{\partial x_i}\frac{\partial x_i}{\partial W_r}$$ Actually, $\frac{\partial L}{\partial x_i}$ involves two terms according to the definition as below: $$\frac{\partial L}{\partial x_i}=\beta_1 \frac{\partial L_c}{\partial x_i} + \beta_2 \frac{\partial L_d}{\partial x_i}$$ The first term is rather simple according to the definition of $L_c$. However, the second term is a little bit complicated as $L_d$ depends on class center $m_j$ which further depends on $x_i$. In order to simplify the optimization algorithm, we follow the approach in literature [@centerloss], i.e. fixing the center $m_j$ during calculation of $\frac{\partial L_d}{\partial x_i}$. This simplification gives us: $$\frac{\partial L_d}{\partial x_i}=x_i-m_{c_i}$$ Gradient with respect to $W_h$ ($\nabla W_h$): For the parameter $W_h$, we give the chain rule as in equation \[equ:SRNETChain\] considering the hallucination loss $L_h$ is added to the intermediate feature map $I_i^h$: $$\label{equ:SRNETChain} \frac{\partial L}{\partial W_h}=\Sigma_{i=1}^{n}\frac{\partial L}{\partial I_i^h}\frac{\partial I_i^h}{\partial W_h}$$ This shows we can run back propagation to get the gradient with respect to parameter set $W_h$ after we correct set the partial derivative of the loss function with respect to $I_i^h$. And by expanding $L$, we get: $$\label{equ:SRNETChainExpand} \frac{\partial L}{\partial I_i^h}=\alpha \frac{\partial L_h}{\partial I_i^h} +( \beta_1 \frac{\partial L_c}{\partial I_i^h}+\beta_2 \frac{\partial L_d}{\partial I_i^h})$$ According to the definition, $\frac{\partial L_h}{\partial I_i^h}$ is quite simple as follows: $$\frac{\partial L_h}{\partial I_i^h}=2(I_i^h-\tilde I_i^h)$$ And the remaining part, i.e.$( \beta_1 \frac{\partial L_c}{\partial I_i^h}+\beta_2 \frac{\partial L_d}{\partial I_i^h})$ can not be analytically expressed as $L_c$ and $L_d$ is not directly defined on $I_i^h$, however, it is just the result of back propagation of recognition layers. Gradient with respect to $W_s:$ This can be directly calculated according to the definition of softmax loss $L_c$. Center update $M$: In deriving the gradient with respect to output feature $x_i$, we assume the center $m_j$ is fixed. However, during the training, $x_i$ will be inevitably changed, which requires to update $m_j$ accordingly. We strictly follow the mechanisms adopted in the literature [@centerloss] by updating the center $m_j$ with a learning rate $\gamma$ as it has been proven very effective: $$m_j=m_j-\gamma \Delta m_j$$ where $\Delta m_j$ is defined as: $$\Delta m_j=\frac{\Sigma_{i=1}^{n}\delta(c_i=j) (m_j-x_i)}{1+\Sigma_{i=1}^{n}\delta(c_i=j)}$$ With these gradient, we can easily run gradient descent algorithm iteratively to find the local minimum. We summarize the optimization algorithm in Algorithm \[alg:JointOptimization\]: \ Training samples $\mathcal{I}=\{<I_l^i,\tilde I_h^i,c_i>\}$; \ Model parameter set $W=[W_h W_r]$ t=t+1; calculate the partial derivative $\frac{\partial L}{\partial W_s}$; update the parameter set $W_s$ by $W_s^{t+1}=W_s^t-\theta \frac{\partial L}{\partial W_s}$; calculate the partial derivative $\frac{\partial L}{\partial x_i}$; execute back propagation from top layer to the bottom layer of FRNET to obtain $\frac{\partial L}{\partial W_r}$; calculate the partial derivative $\frac{\partial L_h}{\partial I_h^i}$; add the $\frac{\partial L_h}{\partial I_h^i}$ to the derivative $\frac{\partial L}{\partial I_h^i}$ obtained in step 6; execute back propagation from the top layer to the bottom layer of SRNET to obtain $\frac{\partial L}{\partial W_h}$; update the parameter W by $W^{t+1}=W^t-\theta \nabla$ W; calculate $\Delta m_j$; update the center $m_j$ by $m_j=m_j-\gamma \Delta m_j$; Experiments =========== In this section, we give the experimental results of our model. We first describe the experimental setting including the data preparation, network architecture and evaluation protocol. Then we give the performance of our models under different settings. Also, we compare performance of our SRNET with other state-of-art methods. ------------- ------------------ ----- ----- Convolution $9 \times 9, 64$ $0$ $1$ Convolution $1 \times 1, 32$ $0$ $1$ Convolution $1 \times 1, 3$ $0$ $1$ ------------- ------------------ ----- ----- --------------- ---------------------------------- ---------------------------------- ---------------------------------- Convolution $3 \times 3, 64$ $0$ $1$ Max Pooling $2 \times 2$ $0$ $2$ Residual $\begin{bmatrix} $\begin{bmatrix} $\begin{bmatrix} 3 \times 3, 64 \\ 1 \\ 1 0 \\ 0 3 \times 3, 64 \end{bmatrix} \times 1 $ \end{bmatrix} \times 1 $ \end{bmatrix} \times 1 $ Convolution $3 \times 3, 128$ $0$ $1$ Max Pooling $2 \times 2$ $0$ $2$ Residual $\begin{bmatrix} $\begin{bmatrix} $\begin{bmatrix} 3 \times 3, 128 \\ 1 \\ 1 0 \\ 0 3 \times 3, 128 \end{bmatrix} \times 2 $ \end{bmatrix} \times 2 $ \end{bmatrix} \times 2 $ Convolution $3 \times 3, 256$ $0$ $1$ Max Pooling $2 \times 2$ $0$ $2$ Residual $\begin{bmatrix} $\begin{bmatrix} $\begin{bmatrix} 3 \times 3, 256 \\ 1 \\ 1 0 \\ 0 3 \times 3, 256 \end{bmatrix} \times 5 $ \end{bmatrix} \times 5 $ \end{bmatrix} \times 5 $ Convolution $3 \times 3, 512$ $0$ $1$ Max Pooling $2 \times 2$ $0$ $2$ Residual $\begin{bmatrix} $\begin{bmatrix} $\begin{bmatrix} 3 \times 3, 512 \\ 1 \\ 1 0 \\ 0 3 \times 3, 512 \end{bmatrix} \times 3 $ \end{bmatrix} \times 3 $ \end{bmatrix} \times 3 $ Inner product $512$ - - --------------- ---------------------------------- ---------------------------------- ---------------------------------- ----------------------------------------- ------- ----------- ----------- -- DeepFace [@DBLP:conf/cvpr/TaigmanYRW14] 4M $97.35\%$ $91.40\%$ DeepID-2+ [@DBLP:conf/cvpr/SunWT15] - $98.70\%$ $-$ FaceNet [@DBLP:conf/acl/FlekovaG16] 200M $99.63\%$ $95.10\%$ Center loss [@centerloss] 0.7M $99.28\%$ $94.90\%$ FRNET 0.49M $98.63\%$ $91.30\%$ ----------------------------------------- ------- ----------- ----------- -- --- -------------- -------------- ----------- ----------- ----------- ----------- 1 HR HR $98.63\%$ $94.73\%$ $91.30\%$ $65.70\%$ 2 HR LR $96.35\%$ $74.66\%$ $89.45\%$ $46.10\%$ 3 HR Hallucinated $96.30\%$ $72.44\%$ $89.36\%$ $43.29\%$ 4 LR LR $97.22\%$ $82.40\%$ $90.45\%$ $61.20\%$ 5 Hallucinated Hallucinated $97.61\%$ $83.03\%$ $88.20\%$ $39.90\%$ 6 LR LR $97.95\%$ $88.73\%$ $90.65\%$ $58.50\%$ --- -------------- -------------- ----------- ----------- ----------- ----------- Experimental Setting -------------------- Data Preparation We use 3 datasets in our experiments: CASIA-WebFace [@yi2014learning], LFW [@LFWTech], and YTF [@ytf]. LR-CASIA, LR-LFW and LR-YTF are down-sampled versions of CASIA-WebFaces, LFW and YTF by a factor of $4$. All the face images are aligned with 5 landmarks (two eyes, noise and mouth corners) detected with algorithm [@mtcnn] for similarity transformation. The faces are cropped to $124 \times 108$ RGB images. Each pixel in RGB images is normalized by subtracting 127.5 then dividing by 128. The only data augmentation we used is horizontal flipping. Network Architecture This network consists of two parts: SRNET to hallucinate LR inputs and SRNET to extract deep discriminative features from input images. Details of SRNET and FRNET are given in Table \[tab:srnet\] and Table \[tab:frnet\]. The notation follows [@resnet]’s convention. Evaluation Protocol We report our results on 3 metrics: 1) Verification accuracy on LR-LFW and LR-YTF, 2) True positive rate at low false positive rate $0.1\%$ (TP for short), and 3) Average PSNR gains on LR-LFW. Implemenation Details We implement the SRNET and FRNET using the Caffe [@jia2014caffe] library with our modifications. We extract the deep features by concatenating the output of the first fully-connected layer of the FRNET for each image and its horizontal flip. Verification task is done on the score computed by the cosine distance of two features after PCA. For fair comparisons, we train the networks with batch size $128$. We choose a learning rate $0.00001$ for SRNET and a learning rate $0.1$ for FRNET, and divide the learning rates by $10$ after $16000$ and $24000$ iterations. The training procedure is finished after $28000$ epochs, in no more than $7$ hours on a single TITAN X GPU. Recognition Perfomance and Comparison ------------------------------------- One important goal of our model is to achieve better recognition performance for low resolution facial images. Thus we conduct the experiment using low resolution images for testing and compare with the methods that also use LR images as input. Setting 1: HR-training and HR-testing In order to show the drop caused by low resolution images, we first give the recognition performance trained and tested by normal images, i.e. trained on CASIA-WebFaces and tested on LFW and YTF. For LFW testing set, the verification accuracy is $98.63\%$, and TP is $94.73\%$. For YTF testing set, the verification accuracy is $91.30\%$, and TP is $65.70\%$. We call the network trained by HR dataset as FRNET-HR. Also, we give a comparison of FRNET with other state-of-the-art models in Table \[tab:verif\_performance\]. Setting 2: HR-training and LR-testing The simplest way to run face recognition for low resolution image is to directly feeding the up-scaled image into a network that is trained by the normal dataset (in our experiment, trained by HR-CASIA). On LR-LFW testing set, we achieve accuracy $96.35\%$ and TP $74.66\%$. On LR-YTF, we achieve accuracy $89.45\%$ and TP $46.10\%$. We can see a large drop ($2.28\%$) on LFW compared with the number of using HR as inputs. Setting 3: HR-training and Hallucinated-testing As SRNET produces hallucinated HR versions, we can also use the hallucinated images generated by the SRNET for testing. Thus we first train the SRNET using CASIA-Webfaces. By using the hallucinated versions of LR-LFW, we achieve verification accuracy $96.30\%$ on LFW, from which we can clearly see the hallucinated inputs even degrade the recognition performance compared with directly feeding the LR images to the network ($96.35\%$). Setting 4: LR-training and LR-testing Another direct means to support LR testing is to train the network with LR-CASIA. We call this trained network FRNET-LR. On LR-LFW testing set, we achieve accuracy $97.22\%$ and TP $82.40\%$. On LR-YTF, we achieve accuracy $90.45\%$ and TF $61.20\%$. FRNET-LR performs slightly better than FRNET-HR on LR versions of testing sets. Setting 5: Hallucinated-training and Hallucinated-testing In order to directly benefit from the output of SRNET, we can train the FRNET by using the outputs of SRNET to improve the recognition performance. More precisely, we first train our SRNET and generate hallucinated version of LR-CASIA with SRNET, which are further used to train FRNET. In testing stage, we get the hallucinated versions of LR-LFW and LR-YTF and use the hallucinated versions for testing. Not surprisingly, we get accuracy $97.61\%$ on LR-LFW and $88.20\%$ on LR-YTF respectively. Surprising, It shows a improvement over previous settings on LR-LFW, and poses a negative impact to performance on LR-YTF. We believe that the performance degradation on LR-LFW is caused by video compression artifacts which prevent the SRNET from working properly, and more discriminative features can be learned from hallucinated face images to help recognition task. Setting 6: Joint End-to-end Training and Testing In this setting, we give the recognition performance of our Joint Model. We train the network by taking LR-CASIA images as inputs and CASIA-WebFaces images as ground-truths. The weight of $\alpha$, $\beta_1$ and $\beta_2$ are set $0.01$, $1$ and $0.008$ respectively. We get accuracy $97.95\%$, TP $88.73\%$ on LR-LFW, and accuracy $90.95\%$, TP $59.40\%$ on LR-YTF, which shows a improvement over setting 5. Results of setting 5 and setting 6 support our hypothesis that not only FRNET can learn better from hallucinated images containing more discriminative features, but also SRNET can learn how to produce images more helpful to face recognition task. We give accuracies and TPs under all 6 settings in Table \[tab:settings\]. SR Performance and Comparison ----------------------------- Our Joint Model serves not only for face recognition purpose, but also generates visually pleasing hallucinated images. We trained a SRCNN from scratch as [@srcnn], and compare it with our models. Also, we find the Joint Model has slightly out-performanced stand-alone SRNET and SRCNN (trained on CASIA-WebFaces) by a $0.01$ dB. Conclusion ========== In this paper, we have proposed a Joint Multi-tasking Model for LR face recognition and face SR. By joining the SR network to our face recognition, the power of extracting deep feature from LR is greatly enhanced. Experiments on several LR version of face benchmarks have convincingly demonstrated the effectiveness of the proposed approach.	{ "pile_set_name": "ArXiv" }
--- abstract: \| There has been interest during the last decade in properties of the sequence $\gcd(a^n-1,b^n-1), \ \ n=1,2,3,....$ where $a,b$ are fixed (multiplicatively independent) elements in one of ${\mathbb{Z}}, {\mathbb{C}}[T]$, or ${\mathbb{F}}_q[T]$ . In the case of ${\mathbb{Z}}$, Bugeaud, Corvaja and Zannier have obtained an upper bound $\exp (\epsilon n)$ for any given $\epsilon >0$ and all large $n$, and demonstrate its sharpness by extracting from a paper of Adleman, Pomerance, and Rumely a lower bound $\exp(\exp(c\frac{\log n}{\log\log n}))$ for infinitely many $n$, where $c$ is an absolute constant. This paper generalizes these results to $\gcd(\Phi_N(a^n), \Phi_N(b^n))$ for any positive integer $N$, where $\Phi_N(x)$ is the $N$th cyclotomic polynomial, the preceding being the case $N=1$. The upper bound follows easily from the original, but not the lower bound, which is the focus of this paper. The lower bound has been proved in the first author’s Ph.D. thesis for the case $N=2$, i.e. for $\gcd(a^n+1,b^n+1)$. In this paper we prove the lower bound for arbitrary $N$ under GRH (the generalized Riemann Hypothesis). The analogue of the lower bound for $\gcd(a^n-1,b^n-1)$ over ${\mathbb{F}}_q[T]$ was proved by Silverman; we prove a corresponding generalization unconditionally. address: - \| Department of Mathematics\ Technion — Israel Institute of Technology\ Haifa, 32000\ Israel - \| Department of Mathematics\ Technion — Israel Institute of Technology\ Haifa, 32000\ Israel author: - Joseph Cohen - Jack Sonn title: 'On $GCD(\Phi_N(a^n),\Phi_N(b^n))$' --- Introduction {#sec:intro} ============ In recent years there has been interest [@BCZ],[@AR],[@si] in sequences of the form $$\gcd(a^n-1,b^n-1), \ \ n=1,2,3,....$$ where $a,b$ are fixed elements in one of ${\mathbb{Z}}, {\mathbb{C}}[T]$, or ${\mathbb{F}}_q[T]$. Motivated by recurrence sequences and the Hadamard quotient theorem, Bugeaud, Corvaja and Zannier [@BCZ], bounded the cancellation in the sequence $\frac{b^n-1}{a^n-1}$ by proving the following upper bound result: 0.5em \[thm:upper bound\][@BCZ] Let $a,b$ be multiplicatively independent positive integers, $\epsilon>0$. Then $\log\gcd(a^n-1,b^n-1)< \epsilon n$ for all sufficiently large $n$. Moreover, it is conjectured in [@AR] that if the additional (necessary) condition $\gcd(a-1,b-1)=1$ holds, then $\gcd(a^n-1,b^n-1)=1$ for infinitely many $n$. Returning to [@BCZ], in order to show that Theorem \[thm:upper bound\] is close to best possible, it is remarked in [@BCZ] that one can derive from a paper of Adleman, Pomerance, and Rumely [@APR] a lower bound result: \[thm:lower bound\] [@BCZ] For any two positive integers $a,b$, there exist infinitely many positive integers $n$ for which $\log\gcd (a^n-1,b^n-1)>\exp(c\frac{\log n}{\log\log n})$, where $c$ is an absolute constant. The result in [@APR] from which this is derived is an improvement of a result of Prachar [@pr]: \[prachar\] [@pr] Let $\delta(n)$ denote the number of divisors of $n$ of the form $p-1$, with $p$ prime. Then there exist infinitely many $n$ such that $\delta(n)> \exp(c \log n/(\log \log n)^2).$ The improvement in [@APR] (with a similar proof) removes the exponent 2 (and the $p-1$ are squarefree): \[APR\] [@APR] Let $\delta(n)$ denote the number of divisors of $n$ of the form $p-1$, with $p$ prime and $p-1$ squarefree. Then there exist infinitely many $n$ such that $\delta(n)> \exp(c \log n/\log \log n).$ It is interesting to note that in [@pr], Prachar was motivated by a paper of Nöbauer [@no] which dealt with the group of invertible polynomial functions on ${\mathbb{Z}}/n{\mathbb{Z}}$ and particularly the subgroup of functions of the form $x^k$, whereas in [@APR], Adleman, Pomerance and Rumely were motivated by the computation of a lower bound on the running time of a primality testing algorithm. 0.5em In his Ph.D. thesis [@coh], the first author tests the robustness of these results and asks what happens to Theorem \[thm:lower bound\] if $\gcd(a^n-1,b^n-1)$ is replaced by $\gcd(a^n+1,b^n+1)$ or by $\gcd(a^n+1,b^n-1)$, and proceeds to prove the analogous results for these sequences, using [@APR]: \[thm:yossi\] [@coh] For any two positive nonsquare integers $a,b$, there exist infinitely many positive integers $n$ for which $\log\gcd (a^n+1,b^n+1)>\exp(c\frac{\log n}{\log\log n})$, where $c$ is a constant depending on $a$ and $b$. The same result holds for $\gcd (a^n+1,b^n-1).$ (The corresponding analogues of Theorem \[thm:upper bound\] follow immediately from $x^n\pm1\|x^{2n}-1$.) If one observes that the polynomials $x-1$ and $x+1$ are the first and second cyclotomic polynomials $\Phi_N(x)$, $N=1,2$, we ask if Theorems \[thm:upper bound\] and \[thm:lower bound\] also hold for $\gcd(\Phi_N(a^n), \Phi_N(b^n))$ for any positive integer $N$, or even for $\gcd(\Phi_M(a^n), \Phi_N(b^n))$ for suitable positive integers $M,N$. For Theorem \[thm:upper bound\], this is immediate from $\Phi_N(x)\|x^N-1$. In this paper we deal with this generalization for Theorems \[thm:lower bound\] (and \[thm:yossi\]). 0.5em It should be remarked that Corvaja and Zannier have made far-reaching generalizations of Theorem \[thm:upper bound\] in [@CZ], in other directions. 0.5em In Section 2 we prove the above generalization for $\gcd(\Phi_N(a^n), \Phi_N(b^n))$ for any positive integer $N$ under the Generalized Riemann Hypothesis (GRH). The explanation for this is that the generalization of Prachar’s argument in this situation leads to an application of the effective Chebotarev density theorem to a tower of Galois extensions $L_d/{\mathbb{Q}}$, where the exceptional zeros of the corresponding zeta functions of the $L_d$ are required to be bounded away from $1$ as $d$ goes to infinity [^1]. Since we do not know if the exceptional zeros in our tower are bounded away from $1$, we apply the stronger GRH version of the effective Chebotarev density theorem in which there are no exceptional zeros. An additional attempt to avoid GRH using the Bombieri-Vinogradov theorem has so far not been successful. 0.5em Silverman [@si] has proved an analogue of Theorem \[thm:lower bound\] for the global function fields ${\mathbb{F}}_q(T)$: \[thm:si\] Let ${\mathbb{F}}_q$ be a finite field and let $a(T),b(T) \in {\mathbb{F}}_q(T)$ be nonconstant monic polynomials. Fix any power $q^k$ of $q$ and any congruence class $n_0 +q^k{\mathbb{Z}}\in {\mathbb{Z}}/q^k{\mathbb{Z}}$. Then there is a positive constant $c=c(a,b,q^k)>0$ such that $$\deg (\gcd (a(T)^n-1,b(T)^n-1)) \geq cn$$ for infinitely many $n \equiv n_0 \ (mod \ q^k)$. In Section 3 we apply the method of Section 2 to prove (unconditionally) the corresponding cyclotomic generalization of Silverman’s theorem. 0.5em Acknowledgment. We are grateful to Zeev Rudnick, Ram Murty and Jeff Lagarias for helpful discussions at various stages of the preparation of this paper. We also thank Joe Silverman for helpful comments on the initial draft. 1em The case $a,b\in {\mathbb{Z}}$ {#sec:rational case} =============================== 1em \[thm:rational case\] Let $N$ be a positive integer, $N=\ell_1^{s_1}\cdots \ell_r^{s_r}, \ell_1<\ell_2<\cdots \ell_r$, the factorization of $N$ into primes. Let $a,b$ be positive integers, relatively prime to $N$, which are not $\ell_i$th powers in ${\mathbb{Q}}$ for $i=1,...,r$. Then there exist infinitely many positive integers $n$ such that $$\log\gcd (\Phi_N(a^n),\Phi_N(b^n))>\exp (\frac{c\log n}{\log \log n}),$$ where $c$ is a positive constant depending only on $a,b,N$. 1em Suppose $p$ is a prime congruent to $1$ mod $N$ such that neither $a$ nor $b$ is a $\ell_i$th power mod $p$ for $i=1,...,r$. Suppose also that $n$ is a positive integer prime to $N$ and divisible by $\frac{p-1}{N}$. Then $p\mid{\text {gcd}}(\Phi_N(a^n),\Phi_N(b^n))$. Indeed, $(a^n)^N\equiv 1$ (mod $p$). The orders of $a$ and of $a^n$ mod $p$ are equal and divide $N$. If $a$ has order less than $N$, then there is a prime $\ell\|N$ such that $a^{N/\ell}\equiv 1$ mod $p$, so $a^{(p-1)/\ell}\equiv 1$ mod $p$, whence $a$ is an $\ell$th power mod $p$, contrary to hypothesis. The idea of the proof of the theorem, a generalization of the proof in Prachar’s paper, is to use the pigeonhole principle to produce, for large $x$, an $n \leq x^2$ with more than $\exp(c\frac{\log x}{\log \log x})$ divisors of the form $\frac{p-1}{N}$, $p$ prime, $c$ an absolute constant. The result then follows. Fix $0<\delta<1$. Let $x$ be a positive real number and let $K=K_{\delta}(x)$ be the product of all the primes $p\leq \delta \log x,$ $p \nmid N$. Let $A$ be the set of pairs $(m,p)$, $m$ a positive integer, $p$ a prime, $m \leq x$, $p \leq x$, $\gcd(m,N)=1$, $p\equiv 1$ (mod $N$), $p \not\equiv 1$ (mod $N\ell_i$), $i=1,...,r$, neither $a$ nor $b$ is an $\ell_i$th power mod $p$, $i=1,...,r$, and $K\|m\frac{p-1}{N}$. Now for each $d\|K$, let $A_d$ be the subset of $A$ consisting of pairs $(m,p)\in A$ such that $(m,K)=K/d$ and $d\|\frac{p-1}{N}$. Let $N_0:=\ell_1\cdots \ell_r.$ We first bound $\|A_d\|$ from below by bounding the following subset of $A_d$ of the form $A_d'\times A_d''$, where $$A_d'=\{m\leq x:(m,N_0K)=K/d\}$$ and $$A_d''=\{p \leq x:p\equiv 1 \mod N,\ p \not\equiv 1 \mod N\ell_i,\ i=1,...,r, \ d\mid\frac{p-1}{N},$$ $$\text{ and neither } a \text{ nor } b \text{ is an } \ell_i\text{th power mod } p , \text{ } i=1,...,r \}.$$ 0.5em To bound $\|A_d' \times A_d''\|$ from below, it suffices to bound each of $\|A_d'\|,\|A_d''\|$ from below and take the product of the two lower bounds. 0.5em First, writing $d'=K/d$, $$\|A_d'\|=\|\{m \leq x:d'\|m,(m/d',N_0K/d')=1\}\|=\|\{m/d' \leq x/d':(m/d',N_0K/d')=1\}\|$$ $$\geq\phi(N_0K/d')[\frac{x/d'}{N_0K/d'}]=\phi(N_0d)[x/N_0K]$$ where $\phi$ denotes Euler’s $\phi$-function and $[-]$ the integer part. To bound $\|A_d''\|$ from below we use the effective form of Chebotarev’s density theorem due to Lagarias and Odlyzko [@LO] as formulated by Serre [@Se] under the generalized Riemann Hypothesis (GRH). The condition $d\mid \frac{p-1}{N}$ is equivalent to $p\equiv 1$ (mod $Nd$), which is equivalent to $p$ splits completely in ${\mathbb{Q}}(\mu_{Nd})$, where $\mu_n$ denotes the group of $n$th roots of unity. The condition $a$ is an $\ell$th power mod $p$ ($\ell$ prime) is equivalent to the condition $x^\ell-a$ has a root mod $p$, which for $p\equiv 1$ modulo $\ell$ is equivalent to the condition $x^\ell-a$ splits into linear factors mod $p$, which is equivalent to the condition $p$ splits completely in (the Galois extension) ${\mathbb{Q}}(\mu_\ell,\root{\ell}\of{a})$ of ${\mathbb{Q}}$, which for $p\equiv 1$ (mod $Nd$) and $\ell\mid N$ is equivalent to $p$ splits completely in ${\mathbb{Q}}(\mu_{Nd},\root{\ell}\of{a})$. Consider the Galois extension $F_d={\mathbb{Q}}(\mu_{NdN_0},\root{N_0}\of{a},\root{N_0}\of{b)}$ of ${\mathbb{Q}}$, with Galois group $G_d=G(F_d/{\mathbb{Q}})$, and the subset $$C_d=G(F_d/{\mathbb{Q}}(\mu_{Nd}))\setminus\{[\bigcup_i G(F_d/{\mathbb{Q}}(\mu_{Nd},\root{\ell_i}\of{a}))]\bigcup [\bigcup_i G(F_d/{\mathbb{Q}}(\mu_{Nd},\root{\ell_i}\of{b}))]\bigcup [\bigcup_i G(F_d/{\mathbb{Q}}(\mu_{Nd\ell_i}))]\}$$ of $G_d$. ($C_d$ is the complement in the first group you see in the display, of the union of all the other (sub)groups you see in the display.) It is easily verified that $C_d$ is $G_d$-invariant under conjugation, i.e. a union of conjugacy classes in $G_d$. 0.5em It follows from the definition of $C_d$ that $$A_d''=\{p \leq x:p \text { unramified in } F_d, (p,F_d/{\mathbb{Q}})\subseteq C_d\}$$ where $(p,F_d/{\mathbb{Q}})$ denotes the Artin symbol. Set $$\pi_{C_d}(x):= \|A_d''\|=\|\{p \leq x: p \text { unramified in } F_d, (p,F_d/{\mathbb{Q}})\subseteq C_d\}\|.$$ By the effective Chebotarev density theorem cited above, under GRH for the Dedekind zeta function of $F_d$, $$R_d(x):=\|\pi_{C_d}(x)-\frac{\|C_d\|}{\|G_d\|}Li(x)\|\leq c_1\frac{\|C_d\|}{\|G_d\|}x^{1/2}(\log D_{F_d}+n_{F_d}\log x)$$ where $c_1$ is an absolute constant, $D_{F_d}$ is the discriminant of $F_d$, $n_{F_d}=[F_d:{\mathbb{Q}}]$ is the degree of $F_d$ over ${\mathbb{Q}}$, and $Li(x)$ is the logarithmic integral $\int_2^x \frac{dt}{\log t}$. 0.5em We have $\|G_d\|=\phi({Nd})\prod_i \ell_i^{e_i}$, where $e_i=2$ or $3$ according to whether or not $a,b$ are multiplicatively dependent mod $\ell_i$th powers in ${\mathbb{Q}}$. 0.5em Claim: $\|C_d\|=\prod_i (\ell_i-1)^{e_i}$. 0.5em Proof: First we look at the case $r=1$ ($N$ is a power of $\ell_1$) and write $\ell=\ell_1$. We need an elementary observation. Let $H$ be the direct product of three cyclic groups of order $\ell$: $H=U\times V \times W$ with $U,V,W$ cyclic of order $\ell$. Then $(u,v,w)\in H\setminus ((U\times V)\cup (V\times W )\cup (U \times W)) \iff u\neq 1, v\neq 1, w\neq 1.$ Hence $\|H\setminus ((U\times V)\cup (V\times W )\cup (U \times W))\|=(\ell-1)^3$. 0.5em Now write $G(F_d/{\mathbb{Q}}(\mu_{Nd}))\cong H_1 \times\cdots \times H_r$, where $H_i= U_i \times V_i\times W_i$ if $a,b$ are multiplicatively independent mod $\ell_i$th powers in ${\mathbb{Q}}$, and $H_i= U_i \times W_i$ if not. The subgroups whose union we are looking at (in the definition of $C_d$) can be identified with the subgroups of the form $H_1 \times \cdots H_{i-1} \times U_i \times V_i \times H_{i+1} \times \cdots \times H_r$ or $H_1 \times \cdots H_{i-1} \times V_i \times W_i \times H_{i+1} \times \cdots \times H_r$ or $H_1 \times \cdots H_{i-1}\times U_i \times W_i \times H_{i+1} \times \cdots \times H_r$ for those $i$ for which $a,b$ are multiplicatively independent mod $\ell_i$th powers in ${\mathbb{Q}}$, and the subgroups $H_1 \times \cdots H_{i-1} \times U_i \times H_{i+1} \times \cdots \times H_r$ or $H_1 \times \cdots H_{i-1} \times W_i \times H_{i+1} \times \cdots \times H_r$ for those $i$ for which $a,b$ are multiplicatively dependent mod $\ell_i$th powers in ${\mathbb{Q}}$. An element $(h_1,...,h_r)$ is in the union of these $\iff$ some $h_i \in (U_i\times V_i)\cup (V_i\times W_i )\cup (U_i \times W_i)$ for $i$ of the first kind, or $h_i\in U_i\cup W_i$ for some $i$ of the second kind. Hence $(h_1,...,h_r)$ lies in the complement (in $H$) of the union $\iff h_i=(u_i,v_i,w_i)$ with $u_i,v_i,w_i\neq 1$ for all $i$ of the first kind and $h_i=(u_i,w_i)$ with $u_i,w_i\neq 1$ for all $i$ of the second kind. It follows that the complement has order $\prod_i ( \ell_i-1)^{e_i}$, proving the claim. 0.5em We conclude that $$\frac{\|C_d\|}{\|G_d\|}=\frac{\prod_i (\ell_i-1)^{e_i}}{\phi({Nd})\prod_i \ell_i^{e_i}}.$$** 0.5em By [@Se], Prop. 5, p. 128, $$\log D_{F_d} \leq (n_{F_d}-1) \sum_{p\|Nabd}\log p + n_{F_d}\log n_{F_d}\|\{p:p\|Nabd\}\|.$$ Now $$n_{F_d}=\phi({Nd})\prod_i \ell_i^{e_i}\leq\phi(Nd)N^3=\phi(N)N^3\phi(d),$$ so $$\log D_{F_d} \leq (\phi(N)N^3\phi(d)-1)\log^2(Nabd)+ \phi(N)N^3\phi(d)\log(\phi(N)N^3\phi(d))\log(Nabd)$$ $$\leq \phi(N)N^3\phi(d)\log^2(Nabd)+ \phi(N)N^3\phi(d)\log(\phi(N)N^3\phi(d))\log(Nabd)$$ $$\leq 2(\phi(N)N^3ab)^3\phi(d)\log^2d$$ $$=f(N,a,b)\phi(d)\log^2d.$$ It now follows that $$\|A_d\|\geq\|A_d'\times A_d''\|=\|A_d'\|\|A_d''\|=\|A_d'\|\pi_{C_d}(x)$$ $$\geq \phi(N_0d)[\frac{x}{N_0K}](\frac{\|C_d\|}{\|G_d\|}Li(x)-c_1\frac{\|C_d\|}{\|G_d\|}x^{1/2}(\log D_{F_d}+n_{F_d}\log x))$$ $$\geq\phi(N_0d)[\frac{x}{N_0K}]\frac{\|C_d\|}{\|G_d\|}(Li(x)-c_1x^{1/2}(\log D_{F_d}+n_{F_d}\log x))$$ where $c_1$ is an absolute constant. We now bound $$Li(x)-c_1x^{1/2}(\log D_{F_d}+n_{F_d}\log x)$$ from below. First, $$\log D_{F_d}+n_{F_d}\log x \leq f(N,a,b)\phi(d)\log^2d + \phi(N)N^3\phi(d)\log x$$ $$\leq g(N,a,b)\phi(d) \log^2x\leq g(N,a,b)x^{\delta}\log x \leq g(N,a,b)x^{\delta+\epsilon}$$ (using $\phi(d)<d<K<x^{\delta}$ and $\log x < x^{\epsilon}$ for any given $\epsilon$ and sufficiently large $x$). From this, $$Li(x)-c_1x^{1/2}(\log D_{F_d}+n_{F_d}\log x) \geq \frac{x}{2\log x}-c_1x^{\frac{1}{2}+\delta +\epsilon}g(N,a,b)\geq \frac{x}{4\log x}$$ (for sufficiently large $x$, using $Li(x) \sim \frac{x}{\log x}$). We then have $$\|A_d'\|\pi_{C_d}(x)\geq \phi(N_0d)[\frac{x}{N_0K}]\frac{x}{4\log x}\frac{\|C_d\|}{\|G_d\|}$$ $$\geq \frac{1}{2}\phi(N_0d)\frac{x}{N_0K}\frac{x}{4\log x}\frac{\phi(N_0)^2}{\phi(N)N_0^2}\cdot\frac{1}{\phi(d)}$$ $$\geq \frac{1}{8}\frac{\phi(N_0)^3 x^2}{\phi(N)N_0^3 K \log x}=\frac{h(N)}{K}\frac{x^2}{\log x}.$$ It then follows that $$\|A\|=\sum_{d\|K}\|A_d\|\geq \frac{h(N)}{K}\frac{x^2}{\log x}\sum_{d\|K}1=\frac{h(N)}{K}\frac{x^2}{\log x}2^{\omega(K)}$$ $$\geq \frac{h(N)}{K}\frac{x^2}{\log x}2^{\frac{\frac{1}{4}\delta\log x}{\log \log x}}$$ where $\omega(K)$ denotes the number of primes dividing $K$. For the last inequality we use [@HW], 22.2, p. 341, and 22.10, p. 355: $$\omega(K)\sim \frac{\log K}{\log\log K}\Rightarrow \omega(K)\geq \frac{\log K}{2\log\log K}\geq \frac{1}{4}\frac{\delta\log x}{\log\log x}.$$ Now the number of positive integers $n\leq x^2$ such that $K\|n$ is at most $\frac{x^2}{K}$. Furthermore, for every pair $(m,p)\in A$, $m\frac{p-1}{N}$ is such an $n$. Therefore there exists an $n\leq x^2$ such that $K\|n$ with at least $$\frac{\|A\|}{x^2/K}>\frac{h(N)}{\log x}2^{\frac{\frac{1}{4}\delta\log x}{\log \log x}}=h(N)\exp(c_2\delta\frac{\log x}{\log\log x}-\log\log x)>\exp(c_3\frac{\log x}{\log\log x})$$ representations of the form $m\frac{p-1}{N}$, for $x$ sufficiently large, where $c_2, c_3$ are absolute constants. It follows that $GCD(\Phi_N(a^n),\Phi_N(b^n))$ is a product of at least $\exp(c_3\frac{\log x}{\log\log x})$ primes, hence is itself at least $\exp\exp(c_4\frac{\log x}{\log\log x})$. As $n\leq x^2$ and $\frac{\log x}{\log\log x}$ is an increasing function (for $x>e^e$), the last expression is $\geq \exp\exp(c_5\frac{\log n}{\log\log n})$. 1em The proof of Theorem 2.1 can be generalized to yield the following 0.5em \[contingent on GRH\] \[thm:genrationalcase\] Let $M,N$ be positive integers. Let $D=\gcd(M,N)$ and assume $\gcd(M/D,D)=\gcd(N/D,D)=1$. Let $L=lcm(M,N)=\ell_1^{s_1}\cdots \ell_r^{s_r}, \ell_1<\ell_2<\cdots \ell_r$, the factorization of $L$ into primes. Let $a,b$ be positive integers, relatively prime to $L$, which are not $\ell_i$th powers in ${\mathbb{Q}}$ for $i=1,...,r$. Then there exist infinitely many positive integers $n$ such that $$\gcd (\Phi_M(a^n),\Phi_N(b^n))>(\exp (\exp (\frac{c\log n}{\log \log n}))),$$ where $c$ is a positive constant depending only on $a,b,N$. 0.5em The proof is similar to the proof of Theorem \[thm:rational case\]; we omit the details. Also here, the case $M=1$, $N=2$ was proved unconditionally in [@coh]. 0.5em 2em The case $a=a(T),b=b(T)\in{\mathbb{F}}_q(T)$ {#sec:function field case} ============================================= 1em In this section we will generalize Silverman’s Theorem \[thm:si\] [@si]: Let ${\mathbb{F}}_q$ be a finite field and let $a(T),b(T) \in {\mathbb{F}}_q(T)$ be nonconstant monic polynomials. Fix any power $q^k$ of $q$ and any congruence class $n_0 +q^k{\mathbb{Z}}\in {\mathbb{Z}}/q^k{\mathbb{Z}}$. Then there is a positive constant $c=c(a,b,q^k)>0$ such that $$\deg (\gcd (a(T)^n-1,b(T)^n-1)) \geq cn$$ for infinitely many $n \equiv n_0 \ (mod \ q^k)$. 1em The generalization will be as in the preceding section, replacing $a(T)^n-1$ with $\Phi_m(a(T)^n)$ for an arbitrary fixed positive integer $m$. The proof will be similar in parts to the proof of Theorem 1, but there will be some changes in notation. 1em \[function field case\] Let ${\mathbb{F}}_q$ be a finite field, and let $m$ be a positive integer prime to $q$, $m=\ell_1^{e_1}\cdots \ell_s^{e_s}, \ell_1<\ell_2<\cdots \ell_s$, the factorization of $m$ into primes. Let $a(T),b(T) \in {\mathbb{F}}_q(T)$ be nonconstant monic polynomials which are not $\ell_i$th powers in ${\mathbb{F}}_q(T)$ for $i=1,...,s$. Fix a power $q^k$ of $q$, and any congruence class $n_0 +q^k{\mathbb{Z}}\in {\mathbb{Z}}/q^k{\mathbb{Z}}$. Then there is a positive constant $c=c(m,a,b,q^k)>0$ such that $$\deg (\gcd (\Phi_m(a(T)^n),\Phi_m(b(T)^n)) \geq cn$$ for infinitely many n $\equiv n_0 \ (mod \ q^k)$. 1em Proof. Assume first that $(n_0,q)=1$. Choose the smallest positive integer $r$ such that $(r,m)=1$ and $rmn_0 \equiv -1 \ (mod \ q^k).$ Let $Q=q^t$, where $t\geq k$ and $q^t\equiv 1$ mod $mr$ (e.g. $t=k\phi(mr)$. Let $n=\frac{Q^N-1}{mr}$, where $N$ is a positive integer. Let $\pi=\pi(T)$ be a monic irreducible polynomial of degree $N$ in ${\mathbb{F}}_Q[T]$ not dividing $a(T)b(T)$ (this holds e.g. if $\deg(\pi)>\deg(a(T)b(T))$). Then, writing $a=a(T),b=b(T)$, $\pi\|\Phi_m(a^n)$ if and only if $a^n$ is a primitive $m$th root of unity mod $\pi$, i.e. $a^{nm}\equiv 1$ mod $\pi$ and $a^{mn/\ell}\not\equiv 1$ mod $\pi$ for every $\ell\|m$. Substituting $n=\frac{Q^N-1}{mr}$, this holds $\Leftrightarrow$ $$a^{\frac{Q^N-1}{r}}\equiv 1 \ (mod \ \pi)$$ and $$a^{\frac{Q^N-1}{r\ell}}\not\equiv 1 \ (mod \ \pi)$$ for all $\ell\|m$. The first condition holds $\Leftrightarrow$ there exists $A\in {\mathbb{F}}_Q[T]$ such that $a\equiv A^r \ (mod \ \pi)$. For such an $A$, the second condition is equivalent to $$A^{\frac{Q^N-1}{\ell}}\not\equiv 1 \ (mod \ \pi)$$ which is equivalent to saying that $A$ is not an $\ell$th power mod $\pi$, and since $(r,\ell)=1$, this is equivalent to saying that $a$ is not an $\ell$th power mod $\pi$. It follows that the two conditions hold together $\Leftrightarrow$ $a$ is an $r$th power mod $\pi$ and $a$ is not an $\ell$th power mod $\pi$ for all $\ell$ dividing $m$. We conclude that $\pi\|\Phi_m(a^n)$ if and only if $a$ is an $r$th power mod $\pi$ and $a$ is not an $\ell$th power mod $\pi$ for all $\ell$ dividing $m$. Similarly, $\pi\|\Phi_m(b^n)$ if and only if $b$ is an $r$th power mod $\pi$ and $b$ is not an $\ell$th power mod $\pi$ for all $\ell$ dividing $m$. 0.5em To count the number of $\pi$ dividing $\gcd (\Phi_m(a^n),\Phi_m(b^n))$, we will use an effective version of Chebotarev’s density theorem for global function fields [@FJ], p. 62, Prop. 5.16. For this purpose, let $$F:={\mathbb{F}}_{Q^N}(T)(\root{r}\of{a},\root{r}\of{b})$$ and let $$E:={\mathbb{F}}_{Q^N}(T)(\root{\ell_1}\of{a},\root{\ell_1}\of{b},...,\root{\ell_s}\of{a},\root{\ell_s}\of{b}).$$ 0.5em Since $\deg\pi=N$, $\pi$ splits completely in ${\mathbb{F}}_{Q^N}(T)$. Therefore $a$ and $b$ are $r$th powers mod $\pi$ if and only if $\pi$ splits completely in $F$. Furthermore, $a$ and $b$ are not $\ell$th powers mod $\pi$ for all $\ell$ dividing $m$ if and only if $\pi$ does not split completely in ${\mathbb{F}}_{Q^N}(T)(\root{\ell}\of{a})$ nor in ${\mathbb{F}}_{Q^N}(T)(\root{\ell}\of{b})$ for all $\ell$ dividing $m$. Accordingly, proceeding as in Section 2, consider the Galois extension $EF/{\mathbb{F}}_Q(T)$ with Galois group $G_N$, and let $$C_N=G(EF/{\mathbb{F}}_Q(T))\setminus \{[\bigcup_i G(EF/F(\root{\ell_i}\of{a}))]\bigcup [\bigcup_i G(EF/F(\root{\ell_i}\of{b}))]\}.$$ Then $\pi$ splits completely in $F$ and $\pi$ does not split completely in ${\mathbb{F}}_{Q^N}(T)(\root{\ell}\of{a})$ nor in ${\mathbb{F}}_{Q^N}(T)(\root{\ell}\of{b})$ for all $\ell$ dividing $m$, if and only if $(\pi, EF/{\mathbb{F}}_Q(T))\subseteq C_N$. 0.5em Now the same counting argument as in the preceding section gives $\|G_N\|=Nr^2\prod_i\ell_i^{e_i}$ and $\|C_N\|=\prod_i(\ell_i-1)^{e_i}$. Applying [@FJ], p. 62, Prop. 5.16 [^2] (and observing that a conjugacy class can be replaced by any union of conjugacy classes in that theorem), we get $$\|\{\pi\in {\mathbb{F}}_Q[T]:\pi \ \text{monic irreducible of degree} \ N, \ (\pi, EF/{\mathbb{F}}_Q(T))\subseteq C_N\}\|$$ $$=\frac{\|C_N\|}{\|G_N\|}Q^N+O(Q^{N/2})=\frac{\prod_i(\ell_i-1)^{e_i}}{Nr^2\prod_i\ell_i^{e_i}}Q^N+O(Q^{N/2}).$$ It follows that $$\deg (\gcd (\Phi_m(a(T)^n),\Phi_m(b(T)^n)) \geq N(\frac{\prod_i(\ell_i-1)^{e_i}}{Nr^2\prod_i\ell_i^{e_i}}Q^N+O(Q^{N/2}))$$ $$= \frac{\prod_i(\ell_i-1)^{e_i}}{r^2\prod_i\ell_i^{e_i}}Q^N+O(NQ^{N/2})\geq cn$$ for some constant not depending on $N$, and $n=Q^N-1$. This proves Theorem 2 when $(n_0,q)=1$. The case $(n_0,q)\neq 1$ follows from the case $(n_0,q)=1$ as in [@si]. 0.5em As in the previous section, the proof of Theorem 3.1 can be generalized to yield the following \[thm:genfunctionfieldcase\] Let ${\mathbb{F}}_q$ be a finite field, $u,v$ be positive integers, $d=\gcd(u,v)$, and assume $\gcd(u/d,d)=\gcd(v/d,d)=1$. Let $a=a(T)$, resp. $b=b(T)\in {\mathbb{F}}_q[T]$ be monic nonconstant polynomials which are not $\ell$th powers in ${\mathbb{F}}_q[T]$ for all $\ell\|u$, resp. $\ell\|v$. Fix a power $q^k$ of $q$, and any congruence class $n_0 +q^k{\mathbb{Z}}\in {\mathbb{Z}}/q^k{\mathbb{Z}}$. Then there is a positive constant $c=c(u,v,a,b,q^k)>0$ such that $$\deg (\gcd (\Phi_u(a(T)^n),\Phi_v(b(T)^n)) \geq cn$$ for infinitely many $n \equiv n_0 \ (mod \ q^k)$. 0.5em The details are omitted. 2em [10]{} L. M. Adleman, C. Pomerance and R.S. Rumely, On distinguishing prime numbers from composite numbers, 117: 173–206, 1983 N. Ailon and Z. Rudnick, Torsion points on curves and common divisors of $a^k-1,b^k-1$, 113:31–38, 2004 Y. Bugeaud, P. Corvaja and U. Zannier, An upper bound for the G.C.D. of $a^n - 1$ and $b^n -1$, 243:79–84, 2003 J. Cohen, Primitive roots in algebraic number fields, , 2004 P. Corvaja and U. Zannier, A lower bound for the height of a rational function at S-unit points, 144;3:203–-224, 2005 M. Fried and M. Jarden, , Springer-Verlag , New York-Heidelberg, 1986 G.H. Hardy and E.M. Wright, , Oxford Univ. Press, Oxford, 1979 J. Lagarias and A.M. Odlyzko, Effective versions of the Chebotarev density theorem, , Academic Press, London, 409-464, 1977 W. Nöbauer, Über eine Gruppe der Zahlentheorie, , 58:181-192 (1954) K. Prachar, Über die Anzahl der Teiler einer natürlichen Zahl, welche die Form $p-1$ haben, , 59:91–-97 (1955) J.-P. Serre, Quelques applications du theoreme de densite de Chebotarev, , 54:123-201, 1982 J. Silverman, Common divisors of $a^n-1$ and $b^n-1$ over function fields, , 10:37–43 , 2004 [^1]: The 2-part of $[L_d:{\mathbb{Q}}]$ is unbounded as $d\rightarrow \infty$ so the results of Stark and of Odlyzko and Skinner do not seem to apply. [^2]: This is an effective Chebotarev density theorem for global function fields, implied by the Riemann Hypothesis for curves over finite fields, which is a theorem.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove surjectivity of the comparison map from continuous bounded cohomology to continuous cohomology for Hermitian Lie groups with finite center. For general semisimple Lie groups with finite center, the same argument shows that the image of the comparison map contains all the even generators. Our proof uses a Hirzebruch type proportionality principle in combination with Gromov’s results on boundedness of primary characteristic classes and classical results of Cartan and Borel on the cohomology of compact homogeneous spaces.' address: - 'Departement Mathematik, ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland' - 'Departement Mathematik, ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland' author: - Tobias Hartnick - Andreas Ott title: Surjectivity of the comparison map in bounded cohomology for Hermitian Lie groups --- [^1] Introduction ============ Let $G$ be a connected semisimple Lie group without compact factors. $G$ is called Hermitian if the symmetric space associated with $G$ admits a $G$-invariant complex structure. We denote by $H^\bullet_{cb}(G; {\mathbb{R}})$ the continuous bounded cohomology ring of $G$ and by $H^\bullet_{c}(G; {\mathbb{R}})$ the continuous cohomology ring of $G$, both with trivial real coefficients (see [@BC; @BoWa] or Sections \[SubsecExistenceUniversal\] and \[SubsecGromov\] for the definitions). These two rings are related by a natural comparison map $H^\bullet_{cb}(G; {\mathbb{R}}) \to H^\bullet_{c}(G; {\mathbb{R}})$ [@BC Def.9.2.1]. The purpose of this article is to prove the following result. \[MainThm\] Let $G$ be a connected semisimple Lie group without compact factors and with finite center. If $G$ is Hermitian, then the comparison map $H^\bullet_{cb}(G;{\mathbb{R}}) \to H^\bullet_{c}(G;{\mathbb{R}})$ is surjective. Our methods still apply if the condition that $G$ be Hermitian is dropped, albeit with a weaker conclusion. In this case one has to distinguish between even and odd generators of $H^\bullet_{c}(G; {\mathbb{R}})$ (see Section \[SecOverview\] for the definitions). \[MainThm2\] Let $G$ be a semisimple Lie group with finite center and without compact factors. Then the subring of $H^\bullet_c(G; {\mathbb{R}})$ generated by the even generators consists of bounded classes. We will see in Section \[SecOverview\] that Theorem \[MainThm2\] implies Theorem \[MainThm\]. Before we turn to the proofs of the theorems, let us explain their context. Continuous bounded cohomology of locally compact groups was introduced by Burger and Monod [@BuMo] as a tool to compute the bounded cohomology groups in the sense of Gromov [@Gromov] of compact locally symmetric spaces of the non-compact type, and has found applications beyond that purpose in recent years [@MonodSurvey; @Surface]. Burger and Monod obtained a complete understanding of continuous bounded cohomology of connected Lie groups in degree $2$. More precisely, they showed the following. - If $G$ is a connected Lie group with radical $R(G)$, then $$H^2_{cb}(G; {\mathbb{R}}) \cong H^2_{cb}(G/R(G); {\mathbb{R}}).$$ This reduces the computation of $H^2_{cb}(G; {\mathbb{R}})$ to the case of semisimple Lie groups with finite center. Similarly, one can eliminate compact factors. - If $G$ is a connected semisimple Lie group without compact factors and with finite center, then the degree $2$ comparison map $H^2_{cb}(G; {\mathbb{R}}) \to H^2_{c}(G; {\mathbb{R}})$ is an isomorphism. Since $H^2_{c}(G; {\mathbb{R}})$ is well-known, this allows one to compute $H^2_{cb}(G; {\mathbb{R}})$ for arbitrary connected Lie groups. One would like to prove a similar result for higher degree bounded cohomology groups. While the reduction step (i) is based on amenability methods and hence works in arbitrary degree, the key ingredient in the proof of step (ii) is double ergodicity, which does not have any analog in higher degrees. A higher degree generalization of step (ii) would therefore require methods different from those used by Burger and Monod. Nevertheless, it is commonly believed that such a higher degree generalization exists [@Guido Conjecture 16.1], [@MonodSurvey]. \[MainConj\] If $G$ is a connected semisimple Lie group without compact factors and with finite center, then the comparison map $H^\bullet_{cb}(G; {\mathbb{R}}) \to H^\bullet_{c}(G; {\mathbb{R}})$ is an isomorphism. While the injectivity part of this conjecture is rather new (apparently first suggested in [@MonodSurvey]), the surjectivity part goes back to a $30$ years old question of Dupont [@Dupont Remark3]. In fact, Dupont even suggests an explicit candidate for a bounded cocycle in a given cohomology class, namely the unique cocycle in the image of the van Est map. This stronger form of the conjecture was verified by Dupont himself in degree $2$ [@Dupont], but is still open in higher degree. Theorem \[MainThm\] may be regarded as a positive answer to the cohomological version of Dupont’s conjecture for Hermitian Lie groups. In Section \[SecOverview\] we give an outline of the proofs of Theorem \[MainThm\] and Theorem \[MainThm2\]. The organization of this article will be described at the end of that section. [Acknowledgments.]{} We are grateful to Marc Burger for pointing out reference [@Borel] and, in particular, Proposition \[CorBorel\] to us. While working on the proof of Theorem \[MainThm\] we learned from Michelle Bucher-Karlsson that a similar theorem should be true for certain even-degree classes in arbitrary semisimple Lie groups. We are indepted to her for this suggestion, which led us to discover Theorem \[MainThm2\]. The second author would like to thank the Department of Mathematics at Rutgers University for their hospitality and excellent working conditions. Outline of the proof {#SecOverview} ==================== In this section we outline the proofs of Theorem \[MainThm\] and Theorem \[MainThm2\]. Let $G$ be an arbitrary semisimple Lie group with finite center and without compact factors. The main technical result of this article is Proposition \[MainConvenient\] below, which exhibits an identification of two a priorily different maps between the singular cohomology $H^\bullet(BG;{\mathbb{R}})$ of the classifying space $BG$ of $G$ and the continuous group cohomology $H^\bullet_c(G;{\mathbb{R}})$ of $G$ (see Section \[SubsecExistenceUniversal\] for the definition of continuous group cohomology). The first of these maps is characterized by a universal property and will therefore be referred to as the universal map $${\sigma_G\colonH^\bullet(BG;{\mathbb{R}})\ltoH^\bullet_c(G;{\mathbb{R}})}.$$ In order to define $\sigma_G$, we first recall that for any discrete group $\Gamma$ there is a natural isomorphism ${\sigma_\Gamma\colonH^\bullet(B\Gamma;{\mathbb{R}})\rightarrowH^\bullet(\Gamma;{\mathbb{R}})}$ (see Section \[SubsecExistenceUniversal\]). Then $\sigma_G$ is defined to be the unique extension of the family $\{\sigma_\Gamma\}$ to a natural transformation, that is, the unique map such that for every discrete group $\Gamma$ and every representation ${\rho\colon\Gamma\rightarrowG}$, the diagram $$\begin{xy}\xymatrix{ H^\bullet(BG;{\mathbb{R}}) \ar[d]_{(B\rho)^} \ar[r]^{\sigma_G} & H^\bullet_c(G;{\mathbb{R}}) \ar[d]^{\rho^} \\ H^\bullet(B\Gamma;{\mathbb{R}}) \ar[r]^{\sigma_\Gamma} & H^\bullet(\Gamma;{\mathbb{R}}) \\ }\end{xy}$$ commutes. We will prove in Section \[SubsecExistenceUniversal\] that the universal map ${\sigma}_{G}$ actually exists. It is related to the question of boundedness of classes in the continuous cohomology of $G$ by the next proposition, which is an immediate consequence of Gromov’s result on boundedness of primary characteristic classes [@Gromov] and will be proved in Section \[SubsecGromov\]. \[GromovConvenient\] The image of the universal map $\sigma_G: H^\bullet(BG; {\mathbb{R}}) \to H^\bullet_c(G; {\mathbb{R}})$ is contained in the image of the comparison map $H^\bullet_{cb}(G;{\mathbb{R}}) \to H^\bullet_c(G;{\mathbb{R}})$. We shall compare the universal map ${\sigma}_{G}$ to the geometric map $${T_G\colonH^\bullet(BG;{\mathbb{R}})\ltoH^\bullet_c(G;{\mathbb{R}})}$$ that is defined in terms of explicit geometric data in the following way. Recall that $H^\bullet_c(G; {\mathbb{R}})$ gets identified via the van Est isomorphism with the singular cohomology of the compact dual symmetric space ${{\mathcal{X}}}_u$ of $G$ (see Section \[SubsecVanEst\]). Here ${{\mathcal{X}}}_u = G_u/K$, where $K$ is a maximal compact subgroup of $G$, and $G_u$ is the compact dual group of $G$ (see Section \[SubsecAuxiliary\]). Then the geometric map is defined by $$\begin{aligned} \label{T_Gexplicit} T_G\!: H^\bullet(BG;{\mathbb{R}}) \cong H^\bullet(BK;{\mathbb{R}}) {\xrightarrow}{f_{G_u}^} H^\bullet({{\mathcal{X}}}_u; {\mathbb{R}}) \cong H^\bullet_c(G;{\mathbb{R}})\end{aligned}$$ in terms of the classifying map ${f_{G_u}\colon{{\mathcal{X}}}_u\rightarrowBK}$ of the canonical $K$-bundle $G_{u} \to {{\mathcal{X}}}_{u}$. The next proposition provides the desired identification between the universal and the geometric map. \[MainConvenient\] Let $G$ be an arbitrary semisimple Lie group with finite center and without compact factors. Then the universal map $\sigma_G: H^\bullet(BG; {\mathbb{R}}) \to H^\bullet_c(G; {\mathbb{R}})$ and the geometric map $T_G: H^\bullet(BG; {\mathbb{R}}) \to H^\bullet_c(G; {\mathbb{R}})$ agree up to sign. We will prove Proposition \[MainConvenient\] in Section \[SectionExplicitDescription\]. We now come to the proof of Theorem \[MainThm2\]. Let us denote by ${f_{G_u}\colon{{\mathcal{X}}}_u\rightarrowBK}$ the classifying map of the canonical $K$-bundle ${p_{G_u}\colonG_u\rightarrow{{\mathcal{X}}}_u}$ over the compact dual symmetric space of $G$. By [@Cartan2 Sec.10], there is an isomorphism of algebras $$H^\bullet ({{\mathcal{X}}}_u; {\mathbb{R}})\cong f_{G_u}^ H^\bullet(BK;{\mathbb{R}}) \otimes p_{G_u}^* H^\bullet ({{\mathcal{X}}}_u; {\mathbb{R}})$$ which intertwines $p_{G_u}^$ with the projection onto the second factor. We obtain from this identification generators for the cohomology ring $H^\bullet_c(G; {\mathbb{R}}) \cong H^\bullet({{\mathcal{X}}}_u; {\mathbb{R}})$ in the following way. The singular cohomology ring $H^\bullet_{sing}(K;{\mathbb{R}})$ is a Hopf algebra and is therefore generated by odd degree primitive elements [@Hopf41]. These primitive elements give rise to even degree generators of $H^\bullet(BK;{\mathbb{R}})$ via trangression in the universal $K$-bundle. Taking only those generators of $H^\bullet(BK;{\mathbb{R}})$ that are mapped non-trivially under $f_{G_u}^$ we then obtain generators of the first factor of $H^\bullet ({{\mathcal{X}}}_u; {\mathbb{R}})$. These generators are unique up to real multiples and we shall refer to them as even generators of $H^\bullet ({{\mathcal{X}}}_u;{\mathbb{R}})$. The second factor $p_{G_u}^* H^\bullet ({{\mathcal{X}}}_u;{\mathbb{R}}) \subset H^\bullet_{sing}(G_u;{\mathbb{R}})$ is generated by certain primitive elements lying in the image of the trangression map of the universal $G_u$-bundle. We will call these primitive elements the odd generators of $H^\bullet ({{\mathcal{X}}}_u;{\mathbb{R}})$. (We refer the reader to [@Cartan2] and [@Borel] for details.) Now, by definition, $H^\bullet_c(G;{\mathbb{R}})$ is generated by the even and odd generators, and the image of the map $f_{G_u}^$ contains all even generators. Hence we see from (\[T\_Gexplicit\]) that all even generators of $H^\bullet ({{\mathcal{X}}}_u;{\mathbb{R}})$ are contained in the image of the geometric map $T_{G}$ and hence, by Proposition \[MainConvenient\], also in the image of the universal map ${\sigma}_{G}$. They are thus bounded by Proposition \[GromovConvenient\]. This proves Theorem \[MainThm2\]. We close this subsection explaining how Theorem \[MainThm\] follows from Theorem \[MainThm2\]. By [@Cartan2 Sec.10], the number of odd generators of $H^\bullet ({{\mathcal{X}}}_u;{\mathbb{R}})$ is given by ${\rm rk}(G_u) - {\rm rk}(K)$. Whence $f_{G_u}^$ is onto if and only if ${\rm rk}(G_u) = {\rm rk}(K)$. The next proposition shows that this condition is always satisfied if $G$ is Hermitian. \[CorBorel\] Let $G$ be a Hermitian semsimimple Lie group without compact factors and with finite center, let $K$ be its maximal compact subgroup, and let $G_u$ be its compact dual. Then ${\rm rk}(G_u) = {\rm rk}(K)$. In particular, $H^\bullet_c(G;{\mathbb{R}})$ is generated by its even generators. This follows form the structure theory of Hermitian Lie groups [@Knapp Chapter VII.9]: Namely, if $G$ is Hermitian, then its Lie algebra ${\mathfrak}g$ admits a compact Cartan subalgebra ${\mathfrak{h}}\subset {\mathfrak{k}}$. By definition this means that ${\mathfrak{h}}\otimes {\mathbb{C}}\subset {\mathfrak{g}}\otimes {\mathbb{C}}$ is a Cartan subalgebra. But since ${\mathfrak{g}}\otimes {\mathbb{C}}= {\mathfrak{g}}_u \otimes {\mathbb{C}}$ this implies that ${\mathfrak{h}}\subset {\mathfrak{g}}_u$ is a Cartan subalgebra as well. Hence ${\rm rk}_{\mathbb{R}}(K) = \dim {\mathfrak{h}}= {\rm rk}_{\mathbb{R}}(G_u)$. Theorem \[MainThm\] is now an immediate consequence of Theorem \[MainThm2\] and Proposition \[CorBorel\]. Note that, for general semisimple Lie groups $G$, the subring of $H^\bullet_{c}(G; {\mathbb{R}})$ generated by the even generators may be quite small. For example, for $G = SL(n, {\mathbb{R}})$ it is generated by the Euler class and does not contain any stable classes. The remainder of this article is organized as follows. In Section \[SecUniversal\], we show existence of the universal map $\sigma_G$ and prove Proposition \[GromovConvenient\]. Section \[SecHirzebruch\] then provides a generalization of Hirzebruch’s proportionality principle [@HirzAutom], which will play a central role in the proof of Proposition \[MainConvenient\] in Section \[SectionExplicitDescription\]. Throughout this article we will frequently make use of an auxiliary lattice $\Gamma$ in $G$. The role of this lattice will be discussed in the next section. The role of the auxiliary lattice {#SubsecAuxiliary} ================================= Let $G$ be a semisimple Lie group with finite center and without compact factors. Throughout this article we shall use the following notation. We fix a maximal compact subgroup $K$ of $G$. Writing ${\mathfrak{g}}$ and ${\mathfrak{k}}$ for the Lie algebras of $G$ and $K$, and denoting by ${\mathfrak{p}}$ the Killing orthogonal complement of ${\mathfrak{k}}$ in ${\mathfrak{g}}$, we have a Cartan decomposition ${\mathfrak{g}}= {\mathfrak{k}}\oplus {\mathfrak{p}}$. This yields an identification of the tangent space $T_{o}{{\mathcal{X}}}$ of the symmetric space ${{\mathcal{X}}}= G/K$ of $G$ at the basepoint $o = eK$ with ${\mathfrak{p}}$. By definition the compact dual group $G_u$ of $G$ is the analytic subgroup of the universal complexification $G_{\mathbb{C}}$ of $G$ with Lie algebra ${\mathfrak{g}}_u := {\mathfrak{k}}\oplus i\,{\mathfrak{p}}$. Its homogeneous space ${{\mathcal{X}}}_u {\mathrel{\mathop:}=}G_u/K$ is called the compact dual symmetric space of ${{\mathcal{X}}}$. Its tangent space $T_{o_{u}}{{\mathcal{X}}}_{u}$ at the basepoint $o_{u} = eK$ then gets identified with $i\,{\mathfrak{p}}$. The assignment ${{\mathcal{X}}}\mapsto {{\mathcal{X}}}_u$ gives rise to a duality between Riemannian symmetric spaces (without Euclidean factors) of the non-compact and compact type [@Helgason]. However, this duality is not directly reflected in cohomology, since ${{\mathcal{X}}}$ is contractible and so $H^\bullet ({{\mathcal{X}}};{\mathbb{R}})$ is trivial. To overcome this problem, we fix a cocompact lattice $\Gamma$ in $G$ and consider the locally symmetric space $M := \Gamma\backslash {{\mathcal{X}}}$ instead of ${{\mathcal{X}}}$. As we shall see in Section \[SecHirzebruch\], the cohomology rings of $M$ and ${{\mathcal{X}}}_u$ do reflect the duality between ${{\mathcal{X}}}$ and ${{\mathcal{X}}}_u$. This will be made precise in the generalized Hirzebruch principle of Proposition \[PropositionDualityOfCharacteristicClasses\]. It will be convenient to choose $\Gamma$ torsion-free so that $M$ becomes a manifold. Note that the group cohomology of $\Gamma$ is canonically isomorphic to the singular cohomology of the manifold $M$. The next lemma shows that, by our assumptions, this cohomology ring contains the continuous cohomology ring of $G$. \[LemmaTransfer\] Let $\Gamma$ be a torsion-free, cocompact lattice in $G$ and let us denote by the inclusion. Then $${\iota_\Gamma^\colonH^\bullet(G;{\mathbb{R}})\ltoH^\bullet(\Gamma;{\mathbb{R}}) \cong H^\bullet(M;{\mathbb{R}})}$$ is injective. A left inverse of $\iota_\Gamma^$ is given by the transfer map [@BC], which on the level of cochains is given by $$T^n: C(G^{n+1})^{\Gamma} \to C(G^{n+1})^{G}, \quad f \mapsto \bar f,$$ where $\bar f$ is given in terms of the $G$-invariant probability measure $\mu$ on $G/\Gamma$ by $$\bar f(g_0, \dots, g_n) = \int_{G/\Gamma} f(\dot gg_0, \dots, \dot gg_n)d\mu(\dot g).$$ The lemma will later allow us to carry out computations in cohomology using concrete harmonic differential forms on $M$ rather than abstract cohomology classes. Moreover, it gives rise to the following characterization of the universal map $\sigma_G$. \[NaturalCharacterized\] Let $G$ be a semisimple Lie group without compact factors and with finite center and let $\Gamma$ be a torsion-free, cocompact lattice in $G$. Denote by the inclusion. Let be a homomorphism. Then $\sigma$ agrees with the universal map $\sigma_G$ if and only if the diagram $$\begin{xy}\xymatrix{ H^\bullet(BG;{\mathbb{R}}) \ar[d]_{(B\iota_\Gamma)^} \ar[r]^\sigma & H^\bullet_c(G;{\mathbb{R}}) \ar[d]^{\iota_\Gamma^} \\ H^\bullet(B\Gamma;{\mathbb{R}}) \ar[r]^{\sigma_\Gamma} & H^\bullet(\Gamma;{\mathbb{R}}) \\ }\end{xy}$$ commutes. The existence of the diagram follows from the universal property of $\sigma_G$. Conversely, the diagram determines $\sigma$ uniquely since the right down arrow is injective by Lemma \[LemmaTransfer\]. We now fix a cocompact, torsion-free lattice $\Gamma$ in $G$ once and for all. All subsequent arguments will be independent of this choice. The image of the universal map {#SecUniversal} ============================== The aim of this section is to give an explicit construction of the universal map , whose existence we postulated in Section \[SecOverview\], and to deduce Proposition \[GromovConvenient\] from Gromov’s results on boundedness of primary characteristic classes [@Gromov]. Existence of the universal map {#SubsecExistenceUniversal} ------------------------------ For an arbitrary Hausdorff locally compact topological group $G$, the continuous cohomology $H^\bullet_c(G; {\mathbb{R}})$ of $G$ with real coefficients is defined as the cohomology of the complex $(C^\bullet_c(G;{\mathbb{R}}), d)$, where $$C^n_c(G; {\mathbb{R}}) := C(G^{n+1}, {\mathbb{R}})^G, \quad df(g_0, \dots, g_n) = \sum_{i=0}^n (-1)^i f(g_0, \dots, \widehat{g_i}, \dots, g_n).$$ Here $C(\cdot, {\mathbb{R}})$ stands for real-valued continuous functions and $(-)^G$ denotes the functor of $G$-invariants. Dropping the continuity requirement on the cochains we obtain the usual group cohomology $H^\bullet(G; {\mathbb{R}})$. Thus, if $G^\delta$ denotes $G$ equipped with the discrete topology, then by definition $$H^\bullet(G; {\mathbb{R}}) = H^\bullet_c(G^\delta; {\mathbb{R}}).$$ An important difference between continuous cohomology and ordinary group cohomology is that the latter has a direct geometric interpretation. Namely, if $BG^{\delta}$ denotes any classifying space for $G^\delta$, then $H^\bullet(G; {\mathbb{R}}) \cong H^\bullet(BG^{\delta}; {\mathbb{R}})$. However, it is in general not true that $H^\bullet_c(G; {\mathbb{R}}) \cong H^\bullet(BG; {\mathbb{R}})$. There are various ways to understand the difference between the groups $H^\bullet_c(G; {\mathbb{R}})$ and $H^\bullet(BG;{\mathbb{R}})$. The classical point of view is to consider $H^\bullet_c(G; {\mathbb{R}})$ as relative cohomology groups on a space with two topologies; this was pioneered by Bott [@Bott] and developed further by M. A. Mostow [@Mostow]. We will take a different point of view here, which is inspired from groupoid cohomology [@Tu]. Namely, given a Hausdorff topological group $G$ we define a topological simplicial object $G_\bullet$ by $G_n := G^{n}$, $n \geq 0$ with the usual face and degeneracy maps as described in [@DupontLNM p.76] (where $G_\bullet$ is denoted $NG(\bullet)$). In order to avoid technicalities we will assume that $G$ is either a Lie group or a (not necessarily countable) discrete group. Then the fat geometric realization $B_G := \\|G_\bullet\\|$ of $G_\bullet$ is a model for the classifying space of $G$, and this model is functorial. In fact, it coincides with the Milnor model [@Milnor]. On each of the spaces $G_n$ we can consider the sheaf $\underline{{\mathbb{R}}}$ of locally constant real valued functions and the sheaf $C^0$ of continuous real valued functions. Since these sheaves are compatible with the face and degeneracy maps, we obtain sheaves $\underline{{\mathbb{R}}}$ and $C^0$ over the simplicial space $G_\bullet$. (For the notion of a sheaf over a simplicial space see [@Tu Sec.3].) We can thus form the corresponding sheaf cohomology groups. Now the sheaf $\underline{{\mathbb{R}}}$ admits a flabby resolution $$\underline{{\mathbb{R}}} \to C^0 {\xrightarrow}{d} C^1 \to \dots$$ by the sheaves $C^q$ of singular real $q$-cochains (i.e. for $U_n \subset G_n$ the group $C^q(U_n)$ consists of singular real $q$-cochains in $U_n$.) Hence $H^\bullet(G_\bullet; \underline{{\mathbb{R}}})$ is the cohomology of the total complex associated to the double complex $\{C^q(G_n)\}$. Then [@DupontLNM Prop.5.15] applies and we obtain $$H^\bullet(G_\bullet; \underline{{\mathbb{R}}}) \cong H^\bullet(\\|G_\bullet\\|; \underline{{\mathbb{R}}}) \cong H^\bullet(B_G; {\mathbb{R}}).$$ On the other hand, the sheaves $C^0$ on $G_n$ are flabby, hence acyclic, and thus the double complex computing $H^\bullet(G_\bullet; C^0)$ collapses to the inhomogeneous bar resolution for $ H_c^\bullet(G; {\mathbb{R}})$. This implies $$H^\bullet(G_\bullet; C^0) \cong H_c^\bullet(G; {\mathbb{R}}).$$ In particular, the inclusion $i: \underline{{\mathbb{R}}} \to C^0$ of sheaves induces a map $$\begin{aligned} \label{DefSigma}\sigma_G: H^\bullet(B_G; {\mathbb{R}}) \cong H^\bullet(G_\bullet; \underline{{\mathbb{R}}}) \xrightarrow{i^} H^\bullet(G_\bullet; C^0) \cong H_c^\bullet(G; {\mathbb{R}}).\end{aligned}$$ We can think of this map as follows. The complex which computes $H_c^\bullet(G; {\mathbb{R}}) \cong H^\bullet(G_\bullet; C^0)$ is the $0$-th row of the double complex $\{C^q(G_n)\}$ which computes $H^\bullet(B_G; {\mathbb{R}})$. The map $\sigma_G$ is then induced by the projection maps $$\bigoplus_{p+q = n} C^q(G_p) \to C^0(G_n).$$ Let us spell out the naturality property of $\sigma_G$ explicitly. and $G \mapsto H_c^\bullet(G; {\mathbb{R}})$ define two functors on the category formed by Lie groups and discrete groups with the obvious morphisms, and $\sigma_G$ is a natural transformation between these functors. Moreover, if $\Gamma$ is discrete, then every continuous function is locally constant, whence the sheaves $\underline{{\mathbb{R}}}$ and $C^0$ on $\Gamma_\bullet$ coincide, so that $\sigma_\Gamma$ is an isomorphism in this case. We have thus shown that $\sigma_G$ has the universal property used to define the universal map in Section \[SecOverview\]. By Corollary \[NaturalCharacterized\] this property characterizes $\sigma_G$ uniquely, so that we may indeed refer to $\sigma_G$ as the* universal map from $H^\bullet(B_G; {\mathbb{R}})$ to $H_c^\bullet(G; {\mathbb{R}})$. Gromov’s boundedness theorem {#SubsecGromov} ---------------------------- We claimed in Section \[SecOverview\] that Proposition \[GromovConvenient\] is a direct consequence of Gromov’s results on boundedness of primary characteristic classes [@Gromov]. We shall now make this precise. We begin by recalling the necessary background on bounded cohomology. For an arbitrary locally compact Hausdorff topological group $G$, the continuous bounded cohomology ring $H_{cb}^\bullet(G;{\mathbb{R}})$ of $G$ is defined as be the cohomology of the complex $(C^\bullet_{cb}(G;{\mathbb{R}}), d)$ of continuous bounded functions, where $$C^n_{cb}(G, {\mathbb{R}}) := C_b(G^{n+1}, {\mathbb{R}})^G, \quad df(g_0, \dots, g_n) = \sum_{i=0}^n (-1)^i f(g_0, \dots, \widehat{g_i}, \dots, g_n).$$ If $\Gamma$ is a discrete group, we drop the subscript $c$ from notation, writing $H^\bullet(\Gamma;{\mathbb{R}}) $ and $H_{b}^\bullet(\Gamma;{\mathbb{R}})$. In this case, the groups $H_b^\bullet(\Gamma; {\mathbb{R}})$ are precisely Gromov’s bounded cohomology groups from [@Gromov]. The generalization to topological groups is due to Burger and Monod [@BuMo]. Its basic properties are summarized in [@BC], from which we recall the following facts. Firstly, the inclusion of complexes $$(C^\bullet_{cb}(G;{\mathbb{R}}), d) \hookrightarrow (C^\bullet_{c}(G;{\mathbb{R}}), d)$$ induces a comparison map* $${c^\bullet_G\colonH_{cb}^\bullet(G;{\mathbb{R}})\ltoH_c^\bullet(G;{\mathbb{R}})}$$ which is natural in $G$. Secondly, if $H \subset G$ is a subgroup of finite covolume then there exists a bounded transfer map $${T_b^\bullet\colonH_{cb}^\bullet(H;{\mathbb{R}})\ltoH_{cb}^\bullet(G;{\mathbb{R}})}$$ which provides a left-inverse to the restriction map $H_{cb}^\bullet(G;{\mathbb{R}}) \to H_{cb}^\bullet(H;{\mathbb{R}})$, and on the level of cochains is given by the same formula as the usual transfer map defined in the proof of Lemma \[LemmaTransfer\] above. We now return to the case where $G$ is a semisimple Lie group without compact factors and with finite center, and denote by $G^\delta$ the discrete group underlying $G$. We denote by $B_$ a functorial model of the classifying space (say, the Milnor model) and consider the map $B_\iota_{G^\delta}: B_G^\delta \to B_G$ induced by the continuous map $\iota_{G^\delta}: G^\delta \to G$. The elements in the image of the map $$H^\bullet(B_G) \xrightarrow{(B_\iota_{G^\delta})^} H^\bullet(B_G^\delta) \cong H^\bullet(G^\delta; {\mathbb{R}})$$ are called primary characteristic classes. The following result of Gromov ensures that primary characteristic classes are bounded. \[Gromov\] Every primary characteristic class lies in the image of the comparison map $c_{G^\delta}^\bullet: H_b^\bullet(G^\delta;{\mathbb{R}}) \to H^\bullet(G^\delta;{\mathbb{R}})$. For an alternative approach to Theorem \[Gromov\] we refer the reader to [@Michelle], where a stronger version of the theorem is proved. We are now ready to prove Proposition \[GromovConvenient\]. We shall be using the (torsion-free, cocompact) auxiliary lattice $\Gamma$ in $G$ introdcued in Section \[SubsecAuxiliary\]. The inclusion $\iota_\Gamma: \Gamma \hookrightarrow G$ factors as $\iota_\Gamma = \iota_\delta \circ \iota_\Gamma^\delta$, where $\iota_\Gamma^\delta: \Gamma \to G^\delta$ and $\iota_\delta: G^\delta \to G$. Then the universal property of the universal map $\sigma_G$ yields the following commutative diagram: $$\begin{aligned} \begin{xy} \xymatrix{ & H_{c}^{\bullet}(G; {\mathbb{R}}) \ar[d]_{\iota_{\delta}^} & \ar_{\sigma_G}[l] \ar[d]^{(B_\iota_{\delta})^} H^{\bullet}(B_G; {\mathbb{R}}) \\ H^{\bullet}_b(G^\delta; {\mathbb{R}}) \ar[r]^{c_{G^\delta}^\bullet} \ar[d]_{(\iota_\Gamma^\delta)^} & H^{\bullet}(G^\delta; {\mathbb{R}}) \ar[d]_{(\iota_\Gamma^\delta)^}& \ar[l]_{\sigma_{G^\delta}} H^{\bullet}(B_G^\delta; {\mathbb{R}}) \\ H^{\bullet}_b(\Gamma; {\mathbb{R}}) \ar[r]^{c_G^\bullet} \ar[d]_{T_b^{\bullet}} & H^{\bullet}(\Gamma; {\mathbb{R}}) \ar[d]_{T^\bullet}& \\ H_{cb}^{\bullet}(G; {\mathbb{R}}) \ar[r]^{c_G^\bullet} & H_{c}^{\bullet}(G; {\mathbb{R}})& \\ } \end{xy}\end{aligned}$$ Thus if $\alpha$ lies in the image of the universal map ${\sigma}_{G}$, then $\iota_{\delta}^\alpha$ is a primary characteristic class. By Theorem \[Gromov\] we thus find $\beta \in H^{\bullet}_b(G^\delta; {\mathbb{R}})$ with $\iota_{\delta}^\alpha = c^{\bullet}_{G^\delta}(\beta)$. Then commutativity of the diagram yields $$\alpha = T^\bullet(\iota_{\Gamma}^\alpha) = T^\bullet \circ \bigl( \iota_\Gamma^\delta \bigr)^* \bigl( \iota_{\delta}^\alpha \bigr) = T^\bullet \circ \bigl( \iota_\Gamma^\delta \bigr)^ \bigl( c^{\bullet}_{G^\delta}(\beta) \bigr) = c_G^{\bullet} \bigl( T_b^{\bullet}\circ\iota_{\delta}^(\beta) \bigr),$$ that is, ${\alpha}$ lies in the image of the comparison map $c^{\bullet}_{G}$. Duality and cohomology of symmetric spaces {#SecHirzebruch} ========================================== In this section we recall the van Est isomorphisms, which relates the continuous cohomology of a semisimple Lie group $G$ to the singular cohomology of its compact dual symmetric space ${{\mathcal{X}}}_u$. The latter is then related to the cohomology of the locally symmetric space $M$ of the fixed auxiliary lattice $\Gamma$ in $G$ by means of a generalized Hirzebruch proportionality principle. We emphasize that while the classical Hirzebruch proportionality principle [@HirzAutom] and its generalizations [@KamberTondeur] relate characteristic numbers of bundles over ${{\mathcal{X}}}_u$ to characteristic numbers of bundles over $M$, our generalization provides a relation for characteristic classes of arbitrary degree. The van Est isomorphism {#SubsecVanEst} ----------------------- \[SubSectionDualityOfSymmetricSpaces\] Throughout this section, $G$ denotes a semisimple Lie group without compact factors and with finite center. Moreover, we denote by $\iota_\Gamma: \Gamma \hookrightarrow G$ the inclusion of the fixed auxiliary lattice $\Gamma$, and by $M := \Gamma\backslash{{\mathcal{X}}}$ the associated locally symmetric space. We will again be considering the groups $H_c^n(G; {\mathbb{R}})$, but from a point of view different from the one of the last section. Namely, we will take the point of view of [@BoWa], thinking of $H_c^n(G; {\mathbb{R}})$ as derived functors rather than as sheaf cohomology groups of a simplicial manifold. Accordingly, we can use any $s$-injective resolution $$\begin{aligned} \label{GRes} 0 \to {\mathbb{R}}\to A^0 {\xrightarrow}{d_0} A_1 \to \dots\end{aligned}$$ in order to compute $H_c^\bullet(G; {\mathbb{R}})$ as the cohomology of the complex $$\begin{aligned} \label{InvComplex} (A^0)^G {\xrightarrow}{d_0} (A_1)^G \to \dots\end{aligned}$$ of $G$-invariants (see [@BoWa p.261]). In fact, there exists a continuous chain map, unique up to $G$-chain homotopy, from the complex into the augmented homogeneous bar resolution, which identifies the cohomology of with $H_c^\bullet(G; {\mathbb{R}})$. The complex is also $\Gamma$-injective, whence there exists an isomorphism between the cohomology of the complex $$\begin{aligned} \label{InvComplexGamma}(A^0)^{\Gamma}{\xrightarrow}{d_0} (A_1)^{\Gamma}\to \dots\end{aligned}$$ and $H^\bullet(\Gamma; {\mathbb{R}})$. Via these isomorphisms the map $\iota_\Gamma^: H_c^\bullet(G; {\mathbb{R}}) \to H^\bullet(\Gamma; {\mathbb{R}})$ is intertwined with the inclusion map $(A_n)^G \hookrightarrow (A_n)^{\Gamma}$. Now, for the symmetric space ${{\mathcal{X}}}$ of $G$ $$0 \to {\mathbb{R}}\to \Omega^0({{\mathcal{X}}}) {\xrightarrow}{\operatorname{d}} \Omega^1({{\mathcal{X}}}) \to \dots$$ is an $s$-injective resolution [@BoWa Prop. 5.4]. Using the fact that $G$-invariant forms on ${{\mathcal{X}}}$ are harmonic, we thus obtain an isomorphism $${\iota^\bullet_{vE}\colon\Omega^\bullet({{\mathcal{X}}})^G \cong H^\bullet(\Omega^\bullet({{\mathcal{X}}})^G, \operatorname{d})\ltoH_c^\bullet(G; {\mathbb{R}})},$$ called the van Est isomorphism (see [@DupontTopology] for an explicit description of this map on the level of cochains). A more algebraic model of the space $\Omega^\bullet({{\mathcal{X}}})^G$ can be obtained as follows. Since the action of $G$ on ${{\mathcal{X}}}$ is transitive, any $G$-invariant differential form $\omega \in \Omega^\bullet({{\mathcal{X}}})^G$ is uniquely determined by its value $\omega_o$ at the base point. We thus have an isomorphism $$\Omega^\bullet({{\mathcal{X}}})^G \cong \left(\bigwedge{}^\bullet {\mathfrak{p}}^\right)^K, \quad \omega \mapsto \omega_o.$$ The right hand side actually is the $({\mathfrak{g}}, K)$-cohomology $H^\bullet({\mathfrak{g}}, K; {\mathbb{R}})$ with trivial coefficients. This $({\mathfrak{g}}, K)$-cohomology has the following interpretation in terms of the compact dual symmetric space of $G$: Identify the singular cohomology groups of ${{\mathcal{X}}}_u$ with the de Rham cohomology groups $H^{\bullet}\bigl( \Omega^{\bullet}({{\mathcal{X}}}_{u}), \operatorname{d}^{\bullet} \bigr)$. Now any differential form on ${{\mathcal{X}}}_{u}$ can be made $G_u$-invariant by integration, and this integration does not affect the cohomology classes of closed forms. Moreover, $G_u$-invariant forms are automatically harmonic. Thus Hodge theory provides an isomorphism $$\begin{aligned} \label{MapIsosHodge} H^{\bullet}\bigl( \Omega^{\bullet}({{\mathcal{X}}}_{u}), \operatorname{d}^{\bullet} \bigr) \cong H^{\bullet}\bigl( \Omega^{\bullet}({{\mathcal{X}}}_{u})^{G_{u}}, \operatorname{d}^{\bullet} \bigr) \cong \Omega^{\bullet}({{\mathcal{X}}}_u)^{G_u},\end{aligned}$$ where $\Omega^{\bullet}({{\mathcal{X}}}_{u})^{G_{u}}$ denotes the subcomplex of $G_{u}$-invariant forms. Furthermore, restricting $G_{u}$-invariant forms on ${{\mathcal{X}}}_{u}$ to the basepoint $o_{u}$ we obtain an isomorphism $$\begin{aligned} \label{XuAlgebraic} \Omega^{\bullet}({{\mathcal{X}}}_{u})^{G_u} \cong \left( \bigwedge{}^{\bullet}\,T_{o_{u}}^{\ast}{{\mathcal{X}}}_{u} \right)^{K} \cong \left( \bigwedge{}^\bullet (i\,{\mathfrak{p}})^{\ast} \right)^{K}, {\hspace{5mm}}{\omega}\mapsto {\omega}_{o_{u}}\end{aligned}$$ which identifies $G_{u}$-invariant forms on ${{\mathcal{X}}}_{u}$ with $K$-invariant anti-symmetric multilinear forms on $i\,{\mathfrak{p}}$. The linear map ${{\iota}\colon{\mathfrak{p}}\rightarrowi\,{\mathfrak{p}}}$, $X \mapsto iX$ then induces an isomorphism $$\begin{aligned} \label{MapIsoiota} {{\iota}^{\ast}\colon\left( \bigwedge{}^{\bullet} (i\,{\mathfrak{p}})^{\ast} \right)^{K}\xrightarrow{\quad\cong\quad}\left( \bigwedge{}^{\bullet} {\mathfrak{p}}^{\ast} \right)^{K}}\end{aligned}$$ which acts on n-linear forms by $$({\iota}^{\ast}\,{\alpha})(X_{1},\dots,X_{n}) = {\alpha}(iX_{1},\dots,iX_{n}), {\hspace{5mm}}X_{1},\ldots,X_{n} \in {\mathfrak{p}}.$$ We have thus obtained an isomorphism $$H^\bullet({{\mathcal{X}}}_u; {\mathbb{R}}) \cong \left( \bigwedge{}^{\bullet} {\mathfrak{p}}^{\ast} \right)^{K}$$ realizing the $({\mathfrak}g, K)$-cohomology groups in question as singular cohomology groups of ${{\mathcal{X}}}_u$. In particular we obtain isomorphisms $$\begin{aligned} \label{CompactModel} {\Phi_G\colonH^\bullet({{\mathcal{X}}}_u;{\mathbb{R}}){\longrightarrow}\Omega^\bullet({{\mathcal{X}}})^G}\end{aligned}$$ and $$\begin{aligned} \label{CompactModel2} {\Psi_G := \iota_{vE}^\bullet \circ \Phi_G\colonH^\bullet({{\mathcal{X}}}_u;{\mathbb{R}})\ltoH_c^\bullet(G;{\mathbb{R}})}.\end{aligned}$$ Here the isomorphism $\Phi_G$ is completely explicit on cochains. We will now relate these isomorphisms to the auxiliary lattice $\Gamma$ and the associated locally symmetric space $M = \Gamma\backslash {{\mathcal{X}}}$. Recall at this point that $\Gamma$ is assumed to be cocompact and torsion free. We denote by ${\pi\colon{{\mathcal{X}}}\rightarrowM}$ the canonical projection and observe that the image of the inclusion ${\Omega}^{\bullet}({{\mathcal{X}}})^{G} \hookrightarrow {\Omega}^{\bullet}_{\rm{cl}}({{\mathcal{X}}})^{{\Gamma}} \cong \Omega^\bullet(M)$ consist of closed forms. In particular, we obtain a push-forward map $$\begin{aligned} \label{MapFromXToM} {\pi_{!}\colon{\Omega}^{\bullet}({{\mathcal{X}}})^{G}\ltoH^{\bullet}(M;{\mathbb{R}})}.\end{aligned}$$ Since the inclusion ${{\iota}_{\Gamma}\colon\Gamma\rightarrowG}$ induces the restriction map , we have thus proved the following proposition. \[SmoothImplementation\] If $i_M: H^\bullet_{dR}(M; {\mathbb{R}}) \to H^\bullet(\Gamma; {\mathbb{R}})$ denotes the canonical isomorphism, then the diagram $$\begin{xy}\xymatrix{ \Omega({{\mathcal{X}}})^G \ar[d]_{\iota_{vE}} \ar[r]^{\pi_!}& H^\bullet_{dR}(M; {\mathbb{R}}) \ar[d]^{i_M}\\ H^\bullet_c(G; {\mathbb{R}}) \ar[r]^{\iota_\Gamma^}& H^\bullet(\Gamma; {\mathbb{R}}) }\end{xy}$$ commutes. The generalized Hirzebruch proportionality principle will be formulated in terms of the homomorphism ${\Phi_\Gamma\colonH^\bullet({{\mathcal{X}}}_u;{\mathbb{R}})\rightarrowH^\bullet(M; {\mathbb{R}})}$ defined by $$\begin{aligned} \label{MapBetweenCohomologyOfSymmetricSpaces} {\Phi_{{\Gamma}} := \pi_! \circ \Phi_G\colonH^{\bullet}({{\mathcal{X}}}_{u};{\mathbb{R}})\ltoH^{\bullet}(M;{\mathbb{R}})}.\end{aligned}$$ It will thus be convenient to have the following reformulation of Proposition \[SmoothImplementation\]. \[LatticeInclusion\] The diagram $$\begin{xy}\xymatrix{ H^\bullet({{\mathcal{X}}}_u;{\mathbb{R}}) \ar[r]^{\Phi_\Gamma} \ar[d]_{\Psi_G} & H^\bullet_{dR}(M;{\mathbb{R}}) \ar[d]^{i_M}\\ H_{c}^\bullet(G;{\mathbb{R}}) \ar[r]^{\iota_\Gamma^} & H^\bullet(\Gamma;{\mathbb{R}}).} \end{xy}$$ commutes. Characteristic classes of principal bundles {#SubSectionCharacteristicClasses} ------------------------------------------- In this subsection we collect some basic facts from Chern-Weil theory, that is, the theory of characteristic classes of principal bundles over smooth manifolds. We follow the exposition in [@DupontLNM]. Let $BK$ denote a classifying space for the compact Lie group $K$. The elements of $H^{\bullet}(BK;{\mathbb{R}})$ are characteristic classes of principal $K$-bundles. More specifically, given a smooth manifold $X$ and a principal $K$-bundle $P \to X$ with classifying map ${f_{P}\colonX\rightarrowBK}$, any characteristic class $c \in H^{\bullet}(BK;{\mathbb{R}})$ gives rise to a corresponding characteristic class $$c(P) {\mathrel{\mathop:}=}f_{P}^{\ast}\,c \in H^{\bullet}(X;{\mathbb{R}})$$ of the bundle $P$. For any connection 1-form $A \in {\Omega}^{1}(P,{\mathfrak{k}})$ on $P$, its curvature 2-form $F_{A} \in {\Omega}^{2}(P,{\mathfrak{k}})$ is given by $$F_{A} = \operatorname{d}\!A + \frac{1}{2} [A \wedge A].$$ Here $[A \wedge A]$ denotes the 2-form defined by $$[A \wedge A](v,w) = [A(v),A(w)] - [A(w),A(v)] = 2\,[A(v),A(w)]$$ for tangent vectors $v,w$ on $P$. The curvature form $F_{A}$ is horizontal and thus descends to a 2-form $$F_{A} \in {\Omega}^{2}\bigl( X,P({\mathfrak{k}}) \bigr)$$ on $X$ with values in the adjoint bundle $P({\mathfrak{k}}) := P \operatorname{\times}_{K} {\mathfrak{k}}$. Recall that $K$ acts on the space $S^{k}({\mathfrak{k}}^{\ast})$ of symmetric $k$-multilinear functions on ${\mathfrak{k}}$ via the diagonal adjoint action. We denote by $I^{k}({\mathfrak{k}}^{\ast}) \subset S^{k}({\mathfrak{k}}^{\ast})$ the subset of those functions that are invariant under this action. Given a $K$-invariant symmetric function $f \in I^{k}({\mathfrak{k}}^{\ast})$ and a connection 1-form $A \in {\Omega}^{1}(P,{\mathfrak{k}})$, we obtain a well-defined closed 2-form $$f(F_{A},\cdots,F_{A}) \in {\Omega}^{2}(X)$$ on $X$ which then defines a class in $H^{2}_{\rm{dR}}(X)$. We now have the following lemma (see Theorem 8.1 in [@DupontLNM]): \[LemmaChernWeil\] Let $K$ be a compact Lie group. 1. There are no characteristic classes of principal $K$-bundles in odd degree, that is, $$H^{2k+1}(BK; {\mathbb{R}}) = \{0\}$$ for all $k \geq 0$. 2. For every characteristic class $c \in H^{2k}(BK;{\mathbb{R}})$ there exists a unique $f \in I^{k}({\mathfrak{k}}^{\ast})$ such that the following holds. For every principal $K$-bundle $P \to X$ over a compact smooth manifold $X$ and every connection 1-form $A \in \Omega^1(P,{\mathfrak{k}})$ with curvature form $F_{A} \in \Omega^{2}(X;P({\mathfrak{k}}))$, the corresponding characteristic class $c(P) \in H^{2}(X;{\mathbb{R}})$ of the bundle $P$ gets identified with the class $$[f(F_{A},\cdots,F_{A})] \in H^{2}_{\rm{dR}}(X)$$ under the de Rham isomorphism. Duality of characteristic classes {#SubSectionDualityOfCharacteristicClasses} --------------------------------- The symmetric spaces ${{\mathcal{X}}}$ and ${{\mathcal{X}}}_{u}$ come along with canonical principal $K$-bundles $$G {\longrightarrow}{{\mathcal{X}}}=G/K, {\hspace{5mm}}G_{u} {\longrightarrow}{{\mathcal{X}}}_{u}=G_{u}/K.$$ The former bundle further descends to a principal $K$-bundle $${\Gamma}\backslash G {\longrightarrow}M={\Gamma}\backslash {{\mathcal{X}}}$$ over the locally symmetric space $M$. As we have seen in Section \[SubSectionCharacteristicClasses\] above, any characteristic class $c \in H^{2k}(BK;{\mathbb{R}})$ yields corresponding characteristic classes $$c({\Gamma}\backslash G) \in H^{\bullet}(M;{\mathbb{R}}) {\hspace{5mm}}\text{and} {\hspace{5mm}}c(G_{u}) \in H^{\bullet}({{\mathcal{X}}}_{u};{\mathbb{R}})$$ of the canonical $K$-bundles ${\Gamma}\backslash G \to M$ and $G_{u} \to {{\mathcal{X}}}_{u}$, respectively. The following proposition gives an explicit relation between these classes. \[PropositionDualityOfCharacteristicClasses\] Let $k \ge 0$, and let $c \in H^{2k}(BK;{\mathbb{R}})$ be a characteristic class. Then the corresponding characteristic classes of the canonical $K$-bundles $G_{u} \to {{\mathcal{X}}}_{u}$ and ${\Gamma}\backslash G \to M$ are related by $$\Phi_{{\Gamma}}\bigl( c(G_{u}) \bigr) = (-1)^{k} \cdot c({\Gamma}\backslash G),$$ where ${\Phi_{{\Gamma}}\colonH^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}})\rightarrowH^{2k}(M;{\mathbb{R}})}$ is the homomorphism (\[MapBetweenCohomologyOfSymmetricSpaces\]). We will prove the claimed relation by unraveling the definition of the homomorphism $\Phi_{{\Gamma}}$ given in Section \[SubSectionDualityOfSymmetricSpaces\] above, using the results from Chern-Weil theory discussed in the previous subsection. First of all, we introduce appropriate connection 1-forms on the bundles ${\Gamma}\backslash G \to M$ and $G_u \to {{\mathcal{X}}}_u$. Let us denote by ${\theta}_{G} \in {\Omega}^{1}(G;{\mathfrak{g}})$ and ${\theta}_{G_{u}} \in {\Omega}^{1}(G_{u};{\mathfrak{g}}_{u})$ the left Maurer-Cartan forms on $G$ and $G_{u}$ respectively, and denote by $${\pi_{{\mathfrak{k}}}\colon{\mathfrak{g}}= {\mathfrak{k}}\oplus {\mathfrak{p}}{\longrightarrow}{\mathfrak{k}}}, {\hspace{5mm}}{\pi_{{\mathfrak{k}}}^{u}\colon{\mathfrak{g}}_{u} = {\mathfrak{k}}\oplus i\,{\mathfrak{p}}{\longrightarrow}{\mathfrak{k}}}$$ the canonical projections. Then $$\tilde{A} {\mathrel{\mathop:}=}\pi_{{\mathfrak{k}}} \circ {\theta}_{G} \in {\Omega}^{1}(G;{\mathfrak{k}})$$ defines a connection 1-form on the bundle $G \to {{\mathcal{X}}}$. This form is invariant under the action of the auxiliary lattice ${\Gamma}$ in $G$ and hence descends to a connection 1-form $A \in {\Omega}^{1}({\Gamma}\backslash G;{\mathfrak{k}})$ on the bundle ${\Gamma}\backslash G \to M$. Likewise, $$A_{u} {\mathrel{\mathop:}=}\pi_{{\mathfrak{k}}}^{u} \circ {\theta}_{G_{u}} \in {\Omega}^{1}(G_{u};{\mathfrak{k}})$$ defines a connection 1-form on the bundle $G_{u} \to {{\mathcal{X}}}_{u}$.\ Now we are ready for the actual proof of the proposition: Let $c \in H^{2k}(BK; {\mathbb{R}})$ be a characteristic class and denote by $f \in I^{k}({\mathfrak{k}}^{\ast})$ the corresponding $k$-multilinear invariant function on ${\mathfrak{k}}$ as in Lemma \[LemmaChernWeil\]. By Lemma \[LemmaChernWeil\], the characteristic class $$c(G_{u}) \in H^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}})$$ is represented as a class in de Rham cohomology by the closed form $$f(F_{A_{u}},\dots,F_{A_{u}}) \in \Omega^{2k}({{\mathcal{X}}}_{u}).$$ Since the Maurer-Cartan form ${\theta}_{G_{u}} \in {\Omega}^{1}(G_{u};{\mathfrak{k}})$ is $G_{u}$-inavariant, it follows that the connection 1-form $A_{u} \in{\Omega}^{1}(G_{u};{\mathfrak{k}})$ and hence also the form $f(F_{A_{u}},\dots,F_{A_{u}}) \in{\Omega}^{2k}({{\mathcal{X}}}_{u})$ on ${{\mathcal{X}}}_{u}$ is $G_{u}$-invariant. Thus the class $[f(F_{A_{u}},\dots,F_{A_{u}})]$ is mapped under (\[MapIsosHodge\]) to the invariant form $$f(F_{A_{u}},\dots,F_{A_{u}}) \in{\Omega}^{2k}({{\mathcal{X}}}_{u})^{G_{u}}.$$ Restricting this form to the basepoint $o_{u}$ we see that it gets mapped further under (\[XuAlgebraic\]) to the $K$-invariant $2k$-linear form $$\bigl( f(F_{A_{u}},\dots,F_{A_{u}}) \bigr)_{o_{u}} \in \left( \bigwedge{}^{2k} (i\,{\mathfrak{p}})^{\ast} \right)^{K}$$ on $i\,{\mathfrak{p}}$. In order to figure out the image of this form under the isomorphism (\[MapIsoiota\]) we use the following fact. $${\iota}^{\ast} \Bigl( \bigl( f(F_{A_{u}},\dots,F_{A_{u}}) \bigr)_{o_{u}} \Bigr) = (-1)^{k} \cdot \bigl( f(F_{\tilde{A}},\dots,F_{\tilde{A}}) \bigr)_{o}.$$ We will prove the claim by a direct calculation. First, we note that it follows from the definitions in a straightforward way that, for $iX_{1},\ldots,iX_{2k}\in i\,{\mathfrak{p}}\cong T_{o_{u}}{{\mathcal{X}}}_{u}$, we have $$\begin{gathered} (f(F_{A_{u}},\dots,F_{A_{u}}) \bigr)_{o_{u}}(iX_{1},\ldots,iX_{2k}) \\ = \frac{1}{(2k)!} \sum_{\sigma \in \mathfrak S_{2k}} (-1)^\sigma f((F_{A_{u}})_{e}(iX_{\sigma(1)},iX_{\sigma(2)}),\dots,(F_{A_{u}})_{e}(iX_{\sigma(2k-1)},iX_{\sigma(2k)})), \end{gathered}$$ where $e\in G_{u}$ is the unit element. On the left hand side of this identity we regard $F_{A_{u}}$ as a form on ${{\mathcal{X}}}_{u}$ whereas on the right hand side we regard it as a form on $G_{u}$. Likewise, for $X_{1},\ldots,X_{2k}\in {\mathfrak{p}}\cong T_{o}{{\mathcal{X}}}$ we have $$\begin{gathered} (f(F_{\tilde{A}},\dots,F_{\tilde{A}}))_{o}(X_{1},\ldots,X_{2k})\\ = \frac{1}{(2k)!} \sum_{\sigma \in \mathfrak S_{2k}} (-1)^\sigma f((F_{\tilde{A}})_{e}(X_{\sigma(1)},X_{\sigma(2)}),\dots, (F_{\tilde{A}})_{e}(X_{\sigma(2k-1)},X_{\sigma(2k)})),\end{gathered}$$ where $e\in G$ is the unit element of $G$. Again, on the left hand side of this identity $F_{\tilde{A}}$ is regarded as a form on ${{\mathcal{X}}}$ whereas on the right hand side it is considered as a form on $G$. Since $f$ is $k$-linear we see from this and the definition of the isomorphism (\[MapIsoiota\]) that it will be enough to establish the relation $$\bigl( F_{A_{u}} \bigr)_{e}(iX,iY) = - \bigl( F_{\tilde{A}} \bigr)_{e}(X,Y)$$ for $X,Y\in{\mathfrak{p}}$. To this end, we recall that the Maurer-Cartan form on $G_{u}$ satisfies the identities $$({\theta}_{G_{u}})_{o_{u}}(iX) = iX, \, X \in {\mathfrak{p}}{\hspace{5mm}}\text{and} {\hspace{5mm}}\operatorname{d}\!{\theta}_{G_{u}} + \frac{1}{2} \bigl[ {\theta}_{G_{u}} \wedge {\theta}_{G_{u}} \bigr] = 0.$$ Then we obtain $$\begin{split} \bigl( F_{A_{u}} \bigr)_{e}(iX, iY) &= \left( \bigl( dA_{G_{u}} \bigr)_{e} + \frac{1}{2} \bigl[ A_{G_{u}} \wedge A_{G_{u}} \bigr]_{e} \right) (iX,iY) \\ &= \left( \pi_{{\mathfrak{k}}}^{u} \circ \bigl( \operatorname{d}\!{\theta}_{G_{u}} \bigr)_{e} + \frac{1}{2} \bigl[ \pi_{{\mathfrak{k}}}^{u} \circ \operatorname{d}\!{\theta}_{G_{u}} \wedge \pi_{{\mathfrak{k}}}^{u} \circ \operatorname{d}\!{\theta}_{G_{u}} \bigr]_{e} \right)(iX,iY) \\ &= \frac{1}{2} \left( -\pi_{{\mathfrak{k}}}^{u} \circ \bigl[ {\theta}_{G_{u}} \wedge {\theta}_{G_{u}} \bigr]_{e} + \bigl[ \pi_{{\mathfrak{k}}}^{u} \circ \operatorname{d}\!{\theta}_{G_{u}} \wedge \pi_{{\mathfrak{k}}}^{u} \circ \operatorname{d}\!{\theta}_{G_{u}} \bigr]_{e} \right)(iX,iY) \\ &= - \frac{1}{2} \bigl[ {\theta}_{G_{u}} \wedge {\theta}_{G_{u}} \bigr]_{e} (iX,iY) \\ &= - \frac{1}{2} \Bigl( \bigl[ ({\theta}_{G_{u}})_{e}(iX),({\theta}_{G_{u}})_{e}(iY) \bigr] - \bigl[ ({\theta}_{G_{u}})_{e}(iY),({\theta}_{G_{u}})_{e}(iX) \bigr] \Bigr) \\ &= - [iX,iY] \\ &= [X,Y]. \end{split}$$ A similar computation shows that $$\bigl( F_{\tilde{A}} \bigr)_{e}(X,Y) = -[X,Y].$$ This proves the claim. Continuing with the proof of the proposition, we note that since the Maurer-Cartan form ${\theta}_{G} \in {\Omega}^{1}(G;{\mathfrak{k}})$ is $G$-invariant, it follows that the connection 1-form $\tilde{A} \in{\Omega}^{1}(G;{\mathfrak{k}})$ and hence also the form $f(F_{\tilde{A}},\dots,F_{\tilde{A}}) \in{\Omega}^{2k}({{\mathcal{X}}})$ on ${{\mathcal{X}}}$ is $G$-invariant. Hence the $2k$-linear form $\bigl( f(F_{\tilde{A}},\dots,F_{\tilde{A}}) \bigr)_{o}$ on ${\mathfrak{p}}$ is mapped to the form $$f(F_{\tilde{A}},\dots,F_{\tilde{A}}) \in{\Omega}^{2k}({{\mathcal{X}}})^{G}.$$ This form is in particular ${\Gamma}$-invariant, so it follows from Lemma \[LemmaChernWeil\] that it gets mapped under (\[MapFromXToM\]) to the characteristic class $c({\Gamma}\backslash G)$. The proposition is proved. A proportionality principle {#SubSectionAProportionalityPrinciple} --------------------------- In order to clarify the relation between our results and the generalized Hirzebruch proportionality principle obtained in [@KamberTondeur Sec.4.14], we derive a corollary of Proposition \[PropositionDualityOfCharacteristicClasses\] concerning characteristic numbers. For any compact oriented manifold $X$ of dimension $m$, we denote by $${{\langle}\,\cdot\,,[X]{\rangle}\colonH^{m}(X;{\mathbb{R}}){\longrightarrow}{\mathbb{R}}}$$ the pairing of classes of top degree in cohomology with the fundamental class $[X] \in H_{m}(X;{\mathbb{R}})$ in the singular homology of $X$. Then we have the following proportionality principle. Write $m {\mathrel{\mathop:}=}\dim(M) = \dim({{\mathcal{X}}}_{u})$, and fix an orientation of ${{\mathcal{X}}}_{u}$ and $M$. Then there exists a real number $a(\Gamma) \neq 0$ such that for any collection $c_{1},\dots,c_{r} \in H^{\bullet}(BK;{\mathbb{R}})$ of characteristic classes satisfying $$\deg(c_{1}) + \dots + \deg(c_{r}) = m,$$ we have $$\big{\langle}c_{1}(G_{u}) \wedge \dots \wedge c_{r}(G_{u}),[{{\mathcal{X}}}_{u}] \big{\rangle}= a(\Gamma) \cdot \big{\langle}c_{1}({\Gamma}\backslash G) \wedge \dots \wedge c_{r}({\Gamma}\backslash G),[M] \big{\rangle}.$$ By the assumption and Lemma \[LemmaChernWeil\](i) we may without loss of generality assume $m$ to be even. Since $H^{m}({{\mathcal{X}}}_{u};{\mathbb{R}}) \cong {\mathbb{R}}$ there exists a real number $a^{\prime}$ such that the linear functionals $${\big{\langle}\,\cdot\,,[{{\mathcal{X}}}_{u}]\big{\rangle},\big{\langle}\Phi_{{\Gamma}}(\,\cdot\,),[M]\big{\rangle}\colonH^{m}({{\mathcal{X}}}_{u};{\mathbb{R}}){\longrightarrow}{\mathbb{R}}}$$ are related by $$\big{\langle}\,\cdot\,,[{{\mathcal{X}}}_{u}]\big{\rangle}= a^{\prime} \cdot \big{\langle}\Phi_{{\Gamma}}(\,\cdot\,),[M]\big{\rangle}.$$ Then Proposition \[PropositionDualityOfCharacteristicClasses\] yields $$\begin{split} \big{\langle}c_{1}(G_{u}) \wedge \dots \wedge c_{r}(G_{u}),[{{\mathcal{X}}}_{u}] \big{\rangle}&= a^{\prime} \cdot \big{\langle}\Phi_{{\Gamma}}\bigl( c_{1}(G_{u}) \wedge \dots \wedge c_{r}(G_{u}) \bigr),[M] \bigr) \big{\rangle}\\ &= (-1)^{m/2} \cdot a^{\prime} \cdot \big{\langle}c_{1}({\Gamma}\backslash G) \wedge \dots \wedge c_{r}({\Gamma}\backslash G),[M] \big{\rangle}. \end{split}$$ This shows in particular that $a^{\prime} \neq 0$. Now define $a(\Gamma) {\mathrel{\mathop:}=}(-1)^{m/2} \cdot a^{\prime}$. Geometric vs. universal map {#SectionExplicitDescription} =========================== The goal of this section is to identify the universal map ${\sigma}_{G}$ and the geometric map $T_{G}$ up to sign. This will prove Proposition \[MainConvenient\], thereby completing the proof of Theorem \[MainThm2\]. Reformulation of the statement {#SubSectionTransferMap} ------------------------------ \[SubSectionResult\] Throughout this section, $G$ denotes a semisimple Lie group without compact factors and with finite center. We keep the notation introduced in the last section. In particular, given a subgroup $H < G$ we denote by $\iota_H: H \hookrightarrow G$ the corresponding inclusion map. Since $K$ is a maximal compact subgroup in $G$, the inclusion $\iota_K: K \hookrightarrow G$ is a homotopy equivalence and thus induces an isomorphism $${(B_{\ast}\iota_{K})^{\ast}\colonH^{\bullet}(B_{\ast}G;{\mathbb{R}})\ltoH^{\bullet}(B_{\ast}K;{\mathbb{R}})}.$$ Together with the isomorphism ${\Psi_{G}\colonH_{c}^{\bullet}(G;{\mathbb{R}})\ltoH^{\bullet}({{\mathcal{X}}}_{u};{\mathbb{R}})}$ from in Section\[SubsecVanEst\], this can be used to intertwine the classifiying map ${f_{G_{u}}\colon{{\mathcal{X}}}_{u}\rightarrowB_{\ast}K}$ of the canonical $K$-bundle $G_{u} \to {{\mathcal{X}}}_{u}$ with the geometric map $$\begin{aligned} \label{MapTransferMap} {T_{G}\colonH^{\bullet}(B_{\ast}G;{\mathbb{R}})\ltoH_{c}^{\bullet}(G;{\mathbb{R}})}.\end{aligned}$$ We would like to compare $T_G$ to the universal map $\sigma_G$. In odd degrees we have $H^{2k+1}(B_{\ast}K;{\mathbb{R}}) = \{0\}$ by Lemma \[LemmaChernWeil\], whence $T_G = \sigma_G$ for trivial reasons. In even degree we claim that $$\begin{aligned} \label{PreciseSigns} \sigma_{G} = (-1)^{k} \cdot T_{G}: H^{2k}(B_{\ast}G;{\mathbb{R}}) \to H_{c}^{2k}(G;{\mathbb{R}}).\end{aligned}$$ This refines the statement of Proposition \[MainConvenient\]. We see from Corollary \[NaturalCharacterized\] that in order to prove (\[PreciseSigns\]) it actually suffices to show the following proposition. \[PropositionExplicitDescriptionOfNaturalTransformation\] The diagram $$\begin{xy}\xymatrix{ H^{2k}(B_G;{\mathbb{R}}) \ar[d]_{(B_\iota_\Gamma)^} \ar[rr]^{(-1)^k \cdot T_G} && H^{2k}_c(G;{\mathbb{R}}) \ar[d]^{\iota_\Gamma^} \\ H^{2k}(B_\Gamma;{\mathbb{R}}) \ar[rr]^{\sigma_\Gamma} && H^{2k}(\Gamma;{\mathbb{R}}) \\ }\end{xy}$$ commutes. The proof of this proposition will occupy the rest of this article. Proof of Proposition \[PropositionExplicitDescriptionOfNaturalTransformation\] {#SubSectionProofOfProposition} ------------------------------------------------------------------------------ As before, we denote by $M := \Gamma \backslash G/K$ locally symmetric space associated to the fixed auxiliary lattice $\Gamma$ in $G$. We denote by ${f_{G_{u}}\colon{{\mathcal{X}}}_{u}\rightarrowB_{\ast}K}$ the classifying map of the $K$-bundle $G_{u} \to {{\mathcal{X}}}_{u}$, and by the classifying map of the $K$-bundle ${\Gamma}\backslash G \to M$. For every characteristic class $c \in H^{2k}(B_{\ast}K;{\mathbb{R}})$ we obtain characteristic classes $$c(G_{u}) = f^{\ast}_{G_{u}}c \in H^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}}) {\hspace{5mm}}\text{and} {\hspace{5mm}}c({\Gamma}\backslash G) = f^{\ast}_{{\Gamma}\backslash G} c \in H^{2k}(M;{\mathbb{R}})$$ of the bundles $G_{u} \to {{\mathcal{X}}}_{u}$ and ${\Gamma}\backslash G \to M$, respectively. Now Proposition \[PropositionDualityOfCharacteristicClasses\] implies $$f^{\ast}_{{\Gamma}\backslash G} c = c({\Gamma}\backslash G) = (-1)^{k} \cdot \Phi_{{\Gamma}} \bigl( c(G_{u}) \bigr) = \Phi_{{\Gamma}} \circ \bigl( (-1)^{k} \cdot f^{\ast}_{G_{u}} \bigr) c,$$ thus the triangle $$\begin{aligned} \label{DiagramCommutativeTriangle} \begin{xy} \xymatrix{ & H^{2k}(B_{\ast}K;{\mathbb{R}}) \ar[ld]_{(-1)^{k} \cdot f_{G_{u}}^{\ast} \quad} \ar[rd]^{f_{{\Gamma}\backslash G}^{\ast}} & \\ H^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}}) \ar[rr]^{\Phi_{{\Gamma}}} && H^{2k}(M;{\mathbb{R}}) } \end{xy} $$ commutes. Combining this with Corollary \[LatticeInclusion\] we obtain a commutative diagram $$\begin{aligned} \label{DiagramBigCommutativeSquare} \begin{xy} \xymatrix{ H^{2k}(B_{\ast}K;{\mathbb{R}}) \ar[rr]^{(-1)^{k} \cdot f_{G_{u}}^{\ast}} \ar[d]_{f_{{\Gamma}\backslash G}^{\ast}} && \ar[r]^{\Psi_G} H^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}}) \ar[lld]_{\Phi_{{\Gamma}}} & H^{2k}(G;{\mathbb{R}}) \ar[d]^{{\iota}_{{\Gamma}}^{\ast}} \\ H^{2k}(M;{\mathbb{R}}) \ar[rrr]^{i_M} &&& H^{2k}({\Gamma};{\mathbb{R}}) } \end{xy} $$ where ${{\iota}_{{\Gamma}}\colon{\Gamma}\hookrightarrowG}$ denotes the inclusion. (In order to keep the diagram readable we do not include the de Rham isomorphisms.) The last step in our argument requires the following lemma. We denote by ${f_{{{\mathcal{X}}}}\colonM\rightarrowB_{\ast}{\Gamma}}$ the classifying map of the ${\Gamma}$-bundle ${{\mathcal{X}}}\to M$. The diagram $$\begin{aligned} \label{DiagramCommutativeSquare} \begin{xy} \xymatrix{ H^{2k}(B_{\ast}G;{\mathbb{R}}) \ar[d]_{(B_{\ast}{\iota}_{{\Gamma}})^{\ast}} \ar[rr]^{(B_{\ast}\iota_{K})^{\ast}} && H^{2k}(B_{\ast}K;{\mathbb{R}}) \ar[d]^{f_{{\Gamma}\backslash G}^{\ast}} \\ H^{2k}(B_{\ast}{\Gamma};{\mathbb{R}}) \ar[rr]^{f_{{{\mathcal{X}}}}^{\ast}} && H^{2k}(M;{\mathbb{R}}) } \end{xy} $$ commutes. To show that diagram (\[DiagramCommutativeSquare\]) is commutative it suffices to prove that the square $$\begin{xy} \xymatrix{ B_{\ast}K \ar[rr]^{B_{\ast}\iota_{K}} && B_{\ast}G \\ M \ar[u]^{f_{{\Gamma}\backslash G}} \ar[rr]^{f_{{{\mathcal{X}}}}} && B_{\ast}{\Gamma}\ar[u]_{B_{\ast}{\iota}_{{\Gamma}}} } \end{xy}$$ is commutative up to homotopy. So we have to show that there is a homotopy $$B_{\ast}{\iota}_{K} \circ f_{{\Gamma}\backslash G} \,\simeq\, B_{\ast}{\iota}_{{\Gamma}} \circ f_{{{\mathcal{X}}}}$$ between maps from $M$ to $B_{\ast}G$. Since homotopy classes of maps from $M$ to $B_{\ast}G$ are in one-to-one correspondence with isomorphism classes of $G$-bundles over $M$ this is in turn equivalent to the existence of an isomorphism of $G$-bundles $$\begin{aligned} \label{MapPullBackBundlesIsomorphic} (B_{\ast}{\iota}_{K} \circ f_{{\Gamma}\backslash G})^{\ast}EG \,\cong\, (B_{\ast}{\iota}_{{\Gamma}} \circ f_{{{\mathcal{X}}}})^{\ast}EG\end{aligned}$$ over $M$. Now we see from the pullback diagrams $$\begin{xy} \xymatrix{ ({\Gamma}\backslash G) \operatorname{\times}_{K} G \ar[r] \ar[d] & E_{\ast}K \operatorname{\times}_{K} G \ar[r]\ar[d] & E_{\ast}G \ar[d] \\ M \ar[r]^{f_{{\Gamma}\backslash G}} & B_{\ast}K \ar[r]^{B_{\ast}{\iota}_{K}} & B_{\ast}G } \end{xy}$$ and $$\begin{xy} \xymatrix{ {\Gamma}\backslash ({{\mathcal{X}}}\times G) \ar[r] \ar[d] & E_{\ast}{\Gamma}\operatorname{\times}_{{\Gamma}} G \ar[r] \ar[d] & E_{\ast}G \ar[d] \\ M \ar[r]^{f_{{{\mathcal{X}}}}} & B_{\ast}{\Gamma}\ar[r]^{B_{\ast}{\iota}_{{\Gamma}}} & B_{\ast}G } \end{xy}$$ that $$(B_{\ast}{\iota}_{K} \circ f_{{\Gamma}\backslash G})^{\ast}EG \,\cong\, ({\Gamma}\backslash G) \operatorname{\times}_{K} G {\hspace{5mm}}\text{and} {\hspace{5mm}}(B_{\ast}{\iota}_{{\Gamma}} \circ f_{{{\mathcal{X}}}})^{\ast}EG \,\cong\, {\Gamma}\backslash ({{\mathcal{X}}}\operatorname{\times}G)$$ as $G$-bundles over $M$. Here, in the second diagram the quotient ${\Gamma}\backslash ({{\mathcal{X}}}\times G)$ is taken with respect to the diagonal action induced by the standard left action of ${\Gamma}$ on ${{\mathcal{X}}}=G/K$ and $G$. Thus (\[MapPullBackBundlesIsomorphic\]) is a consequence of the following fact: The bundles $({\Gamma}\backslash G) \operatorname{\times}_{K} G$ and ${\Gamma}\backslash ({{\mathcal{X}}}\times G)$ are isomorphic as $G$-bundles over $M$. To prove the claim we write down the isomorphism explicitly. Let us use the notation $${\pi_{1}\colonP_{1} {\mathrel{\mathop:}=}({\Gamma}\backslash G) \operatorname{\times}_{K} G\ltoM}, {\hspace{5mm}}[{\Gamma}g_{1},g_{2}] \mapsto {\Gamma}g_{1}K$$ and $${\pi_{2}\colonP_{2} {\mathrel{\mathop:}=}{\Gamma}\backslash ((G/K) \times G)\ltoM}, {\hspace{5mm}}[g_{1}K,g_{2}] \mapsto {\Gamma}g_{1}K$$ for the two bundles. Then the map $${{\varphi}\colonG \operatorname{\times}G\ltoG \operatorname{\times}G}, {\hspace{5mm}}(g_{1},g_{2}) \mapsto (g_{1}, g_{1}g_{2}).$$ descends to a map ${\overline{{\varphi}}\colonP_{1}\rightarrowP_{2}}$. Similarly, the inverse $${{\varphi}^{-1}\colonG \operatorname{\times}G\ltoG \operatorname{\times}G}, {\hspace{5mm}}(g_{1},g_{2}) \mapsto (g_{1},g_{1}^{-1}g_{2})$$ descends to a map $P_{2} \to P_{1}$. This proves the claim and finishes the proof of the lemma. Attaching diagram (\[DiagramCommutativeSquare\]) to diagram (\[DiagramBigCommutativeSquare\]) from the left, we obtain a commutative diagram $$\begin{xy} \xymatrix{ H^{2k}(B_{\ast}G;{\mathbb{R}}) \ar[r]^{(B_{\ast}{\iota}_{K})^{\ast}} \ar[d]_{(B_{\ast}{\iota}_{{\Gamma}})^{\ast}} & H^{2k}(B_{\ast}K;{\mathbb{R}}) \ar[rr]^{(-1)^{k} \cdot f_{G_{u}}^{\ast}} && \ar[r]^{\Psi_G} H^{2k}({{\mathcal{X}}}_{u};{\mathbb{R}}) & H^{2k}(G;{\mathbb{R}}) \ar[d]^{{\iota}_{{\Gamma}}^{\ast}} \\ H^{2k}(B_{\ast}{\Gamma};{\mathbb{R}}) \ar[rrrr]_{\sigma_{\Gamma}} &&&& H^{2k}({\Gamma};{\mathbb{R}}) } \end{xy}$$ By definition of the geometric map $T_{G}$, the upper row coincides with $(-1)^k \cdot T_G$. This finishes the proof of Proposition \[PropositionExplicitDescriptionOfNaturalTransformation\]. [10]{} Guido’s book of conjectures. , 54(1-2):3–189, 2008. A gift to Guido Mislin on the occasion of his retirement from ETHZ, June 2006, Collected by Indira Chatterji. A. Borel. Sur la cohomologie des espaces fibrés principaux et des espaces homogènes de groupes de [L]{}ie compacts. , 57:115–207, 1953. A. Borel. , volume 1954 of [Lectures given at the University of Chicago]{}. Springer-Verlag, Berlin, 1967. A. Borel and N. Wallach. , volume 67 of [Mathematical Surveys and Monographs]{}. American Mathematical Society, Providence, RI, second edition, 2000. R. Bott. Some remarks on continuous cohomology. In [Manifolds—Tokyo 1973 (Proc. Internat. Conf., Tokyo, 1973)]{}, pages 161–170. Univ. Tokyo Press, Tokyo, 1975. M. Bucher-Karlsson. Finiteness properties of characteristic classes of flat bundles. , 53(1-2):33–66, 2007. M. Burger, A. Iozzi, and A. Wienhard. Surface group representations with maximal [T]{}oledo invariant. , to appear. M. Burger and N. Monod. Bounded cohomology of lattices in higher rank [L]{}ie groups. , 1(2):199–235, 1999. H. Cartan. La transgression dans un groupe de [L]{}ie et dans un espace fibré principal. In [Colloque de topologie (espaces fibrés), [B]{}ruxelles, 1950]{}, pages 57–71. Georges Thone, Liège, 1951. J. L. Dupont. . Lecture Notes in Mathematics, Vol. 640. Springer-Verlag, Berlin, 1978. J. L. Dupont. Bounds for characteristic numbers of flat bundles. In [Algebraic topology, Aarhus 1978 (Proc. Sympos., Univ. Aarhus, Aarhus, 1978)]{}, volume 763 of [Lecture Notes in Math.]{}, pages 109–119. Springer, Berlin, 1979. J. L. Dupont. Simplicial de [R]{}ham cohomology and characteristic classes of flat bundles. , 15(3):233–245, 1976. M. Gromov. Volume and bounded cohomology. , (56), 1982. F. Hirzebruch. Automorphe [F]{}ormen und der [S]{}atz von [R]{}iemann-[R]{}och. In [Symposium internacional de topología algebraica International symposium on algebraic topology]{}, pages 129–144. Universidad Nacional Autónoma de México and UNESCO, Mexico City, 1958. H. Hopf. Über die [T]{}opologie der [G]{}ruppen-[M]{}annigfaltigkeiten und ihre [V]{}erallgemeinerungen. , 42:22–52, 1941. S. Helgason. Differential geometry, [L]{}ie groups, and symmetric spaces. , 34, Corrected reprint of the 1978 original, American Mathematical Society, Providence, RI, 2001. F. Kamber and P. Tondeur. , 1968. A. W. Knapp. , 2002. J. Milnor. Construction of universal bundles. [II]{}. , 63, 1956. N. Monod. , volume 1758 of [Lecture Notes in Mathematics]{}. Springer-Verlag, Berlin, 2001. N. Monod. An invitation to bounded cohomology. In [International [C]{}ongress of [M]{}athematicians. [V]{}ol. [II]{}]{}, pages 1183–1211. Eur. Math. Soc., Zürich, 2006. M. A. Mostow. Continuous cohomology of spaces with two topologies. , 7(175), 1976. J.-L. Tu. Groupoid cohomology and extensions. , 358(11), 2006. [^1]: The first author was partially supported by SNF grant PP002-102765. The second author was supported by ETH Research Grant TH-01 06-1.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the implications of a light tetraquark on the chiral phase transition at nonzero temperature $T$: The behavior of the chiral and four-quark condensates and the meson masses are studied in the scenario in which the resonance $f_{0}(600)$ is described as a predominantly tetraquark state. It is shown that the critical temperature is lowered and the transition softened. Interesting mixing effects between tetraquark, and quarkonium configurations take place.' address: '$^a$Institute for Theoretical Physics, Johann Wolfgang Goethe University, Max-von-Laue-Str. 1, D–60438 Frankfurt am Main, Germany $^b$Frankfurt Institute for Advanced Studies, Johann Wolfgang Goethe University, Ruth-Moufang-Str. 1, D–60438 Frankfurt am Main, Germany' author: - \| Achim Heinz$^{a}$, Stefan Strüber$^{a}$, Francesco Giacosa$^{a} $ and Dirk H. Rischke$^{a,b}$ title: 'Light tetraquark state at nonzero temperature[^1]' --- Introduction ============ In the last decades theoretical and experimental work on light scalar mesons with masses below $\sim 1.8\text{ GeV }$[@pdg] initiated an intense debate about their nature. Quarkonia, tetraquark and mesonic molecular assignments, together with the inclusion of a scalar glueball state around $1.5\text{ GeV}$ as suggested by lattice simulations, have been proposed and investigated in a variety of combinations and mixing patterns [@amslerrev]. Nowadays evidence toward a full nonet of scalars below $1\text{ GeV}$ is mounting: $f_{0}(600)$, $f_{0}(980)$, $a_{0}(980)$, and $K_{0}^{\ast }(800)$. An elegant way to explain such resonances is the tetraquark assignment proposed by Jaffe long ago [@jaffe]. The reversed mass ordering is naturally explained in this way and also decays can be successfully reproduced [@maiani]. Within this context the lightest scalar resonance $% f_{0}(600)$ is interpreted as a predominantly tetraquark state $1/2[u,d][% \bar{u},\bar{d}],$ where the commutator indicates an antisymmetric flavour configuration of the diquark. The lightest quark-antiquark state, i.e., the chiral partner of the pion with flavor wavefunction $\bar{n}n=\sqrt{1/2}(\bar{u}u+\bar{d}d)$, is then predominantly identified with the broad resonance $f_{0}(1370)$. The fact that scalar quarkonia are $p$-wave states supports this choice. According to this picture quarkonia states, together with the scalar glueball, lie above $1$ GeV, see Ref. [@refs] and refs. therein. It is natural to ask how the scenario outlined here affects the physics at nonzero temperature $T$. It is in fact different from the usual assumptions made in hadronic models at $T>0,$ where the chiral partner of the pion has a mass of about $600$ MeV. Moreover, besides the chiral condensate, new quantities emerge: a tetraquark condensate and the mixing of tetraquark and quarkonium states in the vacuum and at nonzero $T.$ Remarkably, the mixing angle increases for increasing $T$ and the behavior of the chiral condensate is affected by the presence of the tetraquark field. Details can be found in Ref. [@tqft], on which these proceedings are based. The Model {#2} ========= We work with a simple chiral model with the following fields: the pion triplet $\vec{\pi}$, the bare quarkonium field $\varphi \equiv \bar{n}n$, and bare tetraquark field $\chi \equiv \frac{1}{2}[u,d][\bar{u},\bar{d}]$. The chiral potential was derived in Ref. [@tqmix]: $$V=\frac{\lambda }{4}\left( \varphi ^{2}+\vec{\pi}^{2}-F^{2}\right) ^{2}-\varepsilon \varphi +\frac{1}{2}M_{\chi }^{2}\chi ^{2}-g\chi (\varphi ^{2}+\vec{\pi}^{2})\text{ }, \label{pot}$$where, besides the usual Mexican hat, the parameter $g$ describes the interaction strength between the quark-antiquark fields and the tetraquark field $\chi .$ In the limit $g\rightarrow 0$ the field $\chi $ decouples, and a simple linear sigma model for $\varphi $ and $\vec{\pi}$ emerges. The minimum of the potential (\[pot\]) is, to order $O(\varepsilon )$: $$\varphi _{0}\simeq \frac{F}{\sqrt{1-2g^{2}/(\lambda M_{\chi }^{2})}}+\frac{% \varepsilon }{2\lambda F^{2}}\text{ },~~\chi _{0}=\frac{g}{M_{\chi }^{2}}% \varphi _{0}^{2}\text{ }, \label{vev}$$ and $\vec{\pi}_{0}=0$. The condensate $\varphi _{0}$ is identified with the pion decay constant $f_{\pi }=92.4\text{ MeV}$. Note that the tetraquark condensate $\chi _{0}$ is proportional to $\varphi _{0}^{2}$: it is induced by spontaneous symmetry breaking in the quarkonium sector. Shifting the fields by their vacuum expectation values (vev’s) $\varphi \rightarrow \varphi +\varphi _{0}$ and $\chi \rightarrow \chi +\chi _{0}$, and expanding around the minimum, we obtain up to second order in the fields $$V=\frac{1}{2}(\chi ,\varphi )\left( \begin{array}{cc} M_{\chi }^{2} & -2g\varphi _{0} \\ -2g\varphi _{0} & M_{\varphi }^{2}% \end{array}% \right) \left( \begin{array}{c} \chi \\ \varphi \end{array}% \right) +\frac{1}{2}M_{\pi }^{2}\vec{\pi}^{2}+\ldots \,, \label{vexp}$$where $$M_{\varphi }^{2}=\varphi _{0}^{2}\left( 3\lambda -\frac{2g^{2}}{M_{\chi }^{2}% }\right) -\lambda F^{2},~~M_{\pi }^{2}=\frac{\epsilon }{\varphi _{0}} \text{ }.$$Since the mass matrix has off-diagonal terms, the fields $\varphi $ and $\chi $ are not mass eigenstates. The mass eigenstates $H$ and $S$, identified with the resonances $f_{0}(600)$ and $f_{0}(1370)$, respectively, are obtained from an $SO(2)$ rotation of the fields $\varphi $ and $\chi $, $$\left( \begin{array}{c} H \\ S% \end{array}% \right) =\left( \begin{array}{cc} \cos \theta _{0} & \sin \theta _{0} \\ -\sin \theta _{0} & \cos \theta _{0}% \end{array}% \right) \left( \begin{array}{c} \chi \\ \varphi \end{array}% \right) \,,\text{ }\theta _{0}=\frac{1}{2}\arctan \frac{4g\varphi _{0}}{% M_{\varphi }^{2}-M_{\chi }^{2}} \text{ }. \label{diag2}$$The tree-level masses of $H$ and $S$ are $$\begin{aligned} M_{H}^{2}& =M_{\chi }^{2}\cos ^{2}\theta _{0}+M_{\varphi }^{2}\sin ^{2}\theta _{0}-2g\varphi _{0}\sin (2\theta _{0}), \\ M_{S}^{2}& =M_{\varphi }^{2}\cos ^{2}\theta _{0}+M_{\chi }^{2}\sin ^{2}\theta _{0}+2g\varphi _{0}\sin (2\theta _{0}).\end{aligned}$$For the reasons discussed in the Introduction, the bare tetraquark is chosen to be lighter than the bare quarkonium, thus: $M_{S}>M_{\varphi }>M_{\chi }>M_{H}$. The state $H\equiv f_{0}(600)$ is the predominantly tetraquark state, and the state $S\equiv f_{0}(1370)$ is the predominantly quarkonium state. Results and discussions {#3} ======================= In order to investigate the nonzero $T$ behavior, we employ the CJT formalism in the Hartree-Fock approximation [@cjt]; for specification of the method in the case of mixing we refer to Ref. [@juergen]. The CJT-formalism leads to temperature-dependent masses $M_{S}(T)$, $% M_{H}(T)$, $M_{\pi }(T)$, and a temperature-dependent mixing angle $\theta (T) $. Moreover, both scalar-isoscalar fields have a $T$-dependent vev, for the quarkonium $\varphi _{0}\rightarrow \varphi (T)$ and for the tetraquark $% \chi _{0}\rightarrow \chi (T)$. For both fields the zero-temperature limits are $% \varphi (0)=\varphi _{0}$ and $\chi (0)=\chi _{0}$. When the tetraquark decouples (limit $g\rightarrow 0$), $S$ is a pure quarkonium and $H$ is a pure tetraquark state. The transition is crossover for $M_{S}\leq 0.95\text{ GeV}$ and first order above this value. This is a well-established result, see e.g. Ref. [@stefan]. The fact that a heavy chiral partner (i.e., mass larger $1\text{ GeV}$) of the pion leads to a first order phase transition disagrees with lattice QCD calculations [karschfodor]{}. The inclusion of the tetraquark state changes this conclusion as shown in Fig. 1: In Fig. 1.a $M_{H}=0.4\text{ GeV}$ is fixed and the parameters $M_{S} $ and $g$ are varied. In Fig. 1.b the behavior of the quark condensate for fixed $M_{S}=1.0\text{ GeV}$ and $M_{H}=0.4\text{ GeV}$ is shown for different values of the parameter $g$. One observes that for increasing values of the coupling $g$ the critical temperature $T_{c}$ decreases: while $% T_{c}=250\text{ MeV}$ for $g\rightarrow 0$, a value $T_{c}\simeq 200\text{ MeV}$ is obtained for $g=2.0\text{ GeV}$. Also the order of the phase transition is affected: when increasing the parameter $g$, the first-order transition is softened and, if the coupling is large enough, becomes a crossover. ![ Panel (a): Order of the phase transition as a function of the parameters of the model. $M_{H}=0.4$ GeV and $M_{S}$ and $g$ are varied. The forbidden area violates the constraint $\left\vert M_{S}^{2}-M_{H}^{2}\right\vert \geq 4g\protect\varphi _{0}$ [tqft]{}. On the border line between the first-order and the crossover transitions a second-order phase transition is realized. Panel (b): the chiral condensate is shown for $M_{H}=0.4$ GeV and $M_{S}$=1.0 GeV for different values of $g$ (step of $0.5\text{ GeV)}$. The dots in panel (a) correspond to the curves in panel (b).[]{data-label="22"}](parameterraumsteps.eps "fig:") ![ Panel (a): Order of the phase transition as a function of the parameters of the model. $M_{H}=0.4$ GeV and $M_{S}$ and $g$ are varied. The forbidden area violates the constraint $\left\vert M_{S}^{2}-M_{H}^{2}\right\vert \geq 4g\protect\varphi _{0}$ [tqft]{}. On the border line between the first-order and the crossover transitions a second-order phase transition is realized. Panel (b): the chiral condensate is shown for $M_{H}=0.4$ GeV and $M_{S}$=1.0 GeV for different values of $g$ (step of $0.5\text{ GeV)}$. The dots in panel (a) correspond to the curves in panel (b).[]{data-label="22"}](phi.eps "fig:") We now turn to the explicit evaluation of masses, condensates, and the mixing angle at nonzero $T$. The masses are chosen to be in the range quoted by Refs. [@pdg; @pelaez]: $M_{S}=1.2\text{ GeV}$ and $M_{H}=0.4\text{ GeV}$. The coupling strength is set to $g=3.4\text{ GeV}$ in order to obtain a crossover phase transition. Together with the pion mass $M_{\pi }=139\text{ MeV}$ and the pion decay constant $\varphi _{0}=f_{\pi }=92.4\text{ MeV}$ the parameters are determined as: $\lambda =52.85$, $M_{\chi }=0.96\text{ GeV} $, and $F=64.2\text{ MeV}$. ![Condensates (panel a), masses (panel b), and mixing angle (panel c) as function of $T$. (From Ref. [@tqft].)](condensate.eps "fig:")![Condensates (panel a), masses (panel b), and mixing angle (panel c) as function of $T$. (From Ref. [@tqft].)](mass.eps "fig:")![Condensates (panel a), masses (panel b), and mixing angle (panel c) as function of $T$. (From Ref. [@tqft].)](angle.eps "fig:") The behavior of the two condensates is shown in Fig. 2.a. At $T_{c}=180\text{ MeV}$ the quark condensate $\varphi (T)$ drops and approaches zero, thus restoring chiral symmetry. Below $T_{c}$ the tetraquark condensate $\chi (T)$ follows the quark condensate, but above $T_{c}$ the condensate starts to increase. (This result could be different if additional terms $\sim \chi ^{4} $ in Eq. (\[pot\]) were included.) By increasing $T$ the function $M_{S}(T)$ first drops softly, but at a certain temperature $T_{s}\simeq 160\text{ MeV}$ a step-like decrease occurs, while the function $M_{H}(T)$ undergoes a step-like increase. The solid line in Fig. 2.b describes the state $S$ according to the following criterion: $S$ is the state containing the largest amount of the bare quarkonium state $\varphi $. For $T<T_{s}$ it corresponds to the heavier state, for $T>T_{s}$ to the lighter one. A similar analysis holds for the dashed line referring to $H$ as the state with the largest bare tetraquark amount. The mixing angle $\theta (T)$ is shown in Fig. 3.c. At $T_{s}$ the mixing becomes maximal and the angle jumps suddenly from $\pi /4$ to $-\pi /4$, $% \lim_{T\rightarrow T_{s}}=\mp \pi /4$. At $T_{s}$ the two physical states $H$ and $S$ have the same amount (50%) of quarkonium and tetraquark. Conclusions {#4} =========== We have shown that the interpretation of $f_{0}(600)$ as a predominantly tetraquark state sizably affects the thermodynamical properties of the chiral phase transition: the behavior of the quark condensate is softened rendering the order of the phase transition cross-over for a sufficiently large tetraquark-quarkonium interaction, and the value of the critical temperature $% T_{c}$ is reduced, in agreement with recent Lattice simulations [karschfodor]{}. In future studies one should include a complete treatment of the other scalar-isoscalar states $f_{0}(980)$, $f_{0}(1500)$, and $f_{0}(1710)$ which appear in an $N_{f}=3$ context (together with the inclusion of the scalar glueball). Also, (axial-)vector mesons shall be considered [@kr]. Nevertheless, the emergence of mixing of tetraquark and quarkonium states is general, and it is expected to play a relevant role at nonzero temperature also in this generalized context. [99]{} W. M. Yao et al. \[Particle Data Group\], J. Phys. G 33 (2006) 1. C. Amsler and N. A. Tornqvist, Phys. Rept. 389, 61 (2004); F. E. Close and N. A. Tornqvist, J. Phys. G 28, R249 (2002); E. Klempt and A. Zaitsev, Phys. Rept. 454 (2007) 1; F. Giacosa, Phys. Rev. D 80 (2009) 074028. R. L. Jaffe, Phys. Rev. D 15 (1977) 267. R. L. Jaffe, Phys. Rev. D 15 (1977) 281. L. Maiani, F. Piccinini, A. D. Polosa and V. Riquer, Phys. Rev. Lett. 93 (2004) 212002; F. Giacosa, Phys. Rev. D 74 (2006) 014028; G. ’t Hooft, G. Isidori, L. Maiani, A. D. Polosa and V. Riquer, Phys. Lett. B 662 (2008) 424; F. Giacosa and G. Pagliara, Nucl. Phys. A 833 (2010) 138; A. H. Fariborz, R. Jora and J. Schechter, Phys. Rev. D [79]{} (2009) 074014; A. H. Fariborz, R. Jora and J. Schechter, Phys. Rev. D [72]{} (2005) 034001. C. Amsler and F. E. Close, Phys. Lett. B 353, 385 (1995); W. J. Lee and D. Weingarten, Phys. Rev. D 61, 014015 (1999); F. E. Close and A. Kirk, Eur. Phys. J. C 21, 531 (2001); F. Giacosa, T. Gutsche, V. E. Lyubovitskij and A. Faessler, Phys. Rev. D 72, 094006 (2005); F. Giacosa, T. Gutsche, V. E. Lyubovitskij and A. Faessler, Phys. Rev. D 72 (2005) 094006. A. Heinz, S. Struber, F. Giacosa and D. H. Rischke, Phys. Rev. D 79 (2009) 037502 F. Giacosa, Phys. Rev. D 75 (2007) 054007. J. M. Cornwall, R. Jackiw and E. Tomboulis, Phys. Rev. D 10 (1974) 2428. J. T. Lenaghan, D. H. Rischke and J. Schaffner-Bielich, Phys. Rev. D 62 (2000) 085008. S. Strüber and D. H. Rischke, Phys. Rev. D 77 (2008) 085004. F. Karsch, arXiv:hep-ph/0701210; Z. Fodor and S. D. Katz, JHEP 0404 (2004) 050. J. R. Pelaez, Phys. Rev. Lett. 92 (2004) 102001; I. Caprini, G. Colangelo and H. Leutwyler, Phys. Rev. Lett. 96 (2006) 132001. S. Gasiorowicz and D. A. Geffen, Rev. Mod. Phys. 41 (1969) 531; P. Ko and S. Rudaz, Phys. Rev. D 50 (1994) 6877; D. Parganlija, F. Giacosa and D. H. Rischke, arXiv:1003.4934 \[hep-ph\]. [^1]: Presented by A. Heinz at the Excited QCD Workshop, 31.1.-6.2.2010, in Tatranska Lomnica (Slovakia)	{ "pile_set_name": "ArXiv" }
--- abstract: \| We study factorization of processes involving two fragmentation functions in the case of very small transverse momenta. We consider two-hadron production in inclusive $e^+ e^-$ annihilation and demonstrate a new simple and illustrative method of factorization for such processes including leading order $\alpha_S$ corrections. PACS numbers: 12.38.-t, 13.85.Ni; 13.88.+e author: - 'I.V. Anikin, O.V. Teryaev' title: 'Factorization and transverse momentum for two-hadron production in inclusive $e^+\,e^-$ annihilation' --- I. Introduction =============== The electron-positron annihilation process is one of the basic hard scattering processes where the composite structure of hadrons is investigated. Especially, the two hadron production in inclusive $e^+\,e^-$ annihilation (THPIA), [i.e.]{} the process in which two hadrons are detected in the final state, attracts a lot of attention due to the possibility for both the experimental and theoretical studies of several fragmentation functions, in particular the chiral-odd Collins functions [@Col93]-[@Efr-Col]. Since available experimental data [@Efremov1998; @SMC; @Belle] correspond to substantially different values of $Q^2$, taking into account the evolution of fragmentation (in particular, Collins) functions becomes an important theoretical task. In turn, the studies of evolution are intimately correlated to the investigation of QCD factorization. However, it is well-known that the factorization theorems do not hold for an arbitrary kinematic regime. Indeed, if a transverse momentum of the produced hadron is of the same order of magnitude as the large photon virtuality, $p_T\sim Q$, the corresponding hadronic tensor can be factorized and expressed in terms of the integrated fragmentation functions (see, for instance, a detailed analysis of $e^+ e^-$ annihilation in [@Collins81]). Even if $p_T$ is much smaller than the large photon virtuality, but on the other hand, is much bigger than the characteristic hadronic size, $\Lambda_{QCD}$, factorization of the $e^+ e^-$ annihilation cross section can still be proven with methods similar to those used for semi-inclusive DIS or Drell-Yan processes [@Ellis1978; @Collins84; @Ji04]. However, the regime for which the transverse momentum of produced hadron is of the order of $\Lambda_{QCD}$ faces a conceptual problem how to identify the hard subprocess. Therefore, factorization becomes vague and requires a special care [@Ji04]. Factorization in THPIA, starting from the seminal paper [@Altar], was only studied in the collinear approximation (see also [@deFlor]). Besides, the procedure of integration over the transverse virtual photon momentum (see [@DBoer]) was restricted to Born approximation. Since the transverse momentum is of crucial importance for the Collins function, we perform a detailed analysis of its role in THPIA. We use ideas of [@Rad77; @ER81] where it was shown that integration over the transverse momentum of a produced $\mu^+\mu^-$ pair or a hadron in the Drell-Yan process provides an effective propagator of a highly virtual photon and generates the hard subprocess structure. This approach was recently generalized [@Ter04] to include the case of the (weighted) transverse momentum average of semi-inclusive deep-inelastic scattering (SIDIS) with transverse momentum dependent fragmentation functions. In this paper, we develop these ideas further and present a new method of factorization which may be applied for any two-current process. In particular, we demonstrate the application of the method in the case of $e^+ e^-$ annihilation when two hadrons, belonging to different back-to-back jets, are produced. To demonstrate the method we are proposing, we consider, at the first stage, the simplest case of the spin-independent $k_\perp$-integrated fragmentation functions. We analyze contributions up to leading order $\alpha_S$ corrections to the hard part to obtain the evolutions of the corresponding fragmentation functions. Also, we briefly outline a way to extend our approach for the study of the Collins effects. II. Kinematics ============== We consider the process $e^+(l_1)+ e^-(l_2)\to H(P_1)+H(P_2)+ X(P_X)$ where positron and electron carrying momenta $l_1$ and $l_2$, respectively, annihilate into a time-like photon with momentum $q=l_1+l_2$ for which $q^2=Q^2$ is large. This time-like photon creates then two outgoing hadrons with momenta $P_i\, (i=1,\,2)$ (these hadrons belong to two distinct jets) and an undetected bunch of hadrons with total momentum $P_X$. For such kind of a process, it is convenient to introduce two invariants $$\begin{aligned} \label{zvar} z_i=\frac{2P_i\cdot q}{Q^2}\end{aligned}$$ which are analogous to the Bjorken variable. Note that the variables $z_1$ and $z_2$ are the energy fractions of the detected hadrons in the $e^+\,e^-$ center of mass system. They obey the following constraint due to the energy conservation: $$\begin{aligned} \label{cmsrel} \frac{z_1+z_2}{2} < 1.\end{aligned}$$ Moreover, due to the momentum conservation there is a stronger constraint on $z_1$ and $z_2$: $$\begin{aligned} \label{zreg} z_1 < 1, \quad z_2 < 1,\end{aligned}$$ Within this region, these variables can vary independently. To perform the Sudakov decomposition, we choose two dimensionless light-cone basis vectors: $$\begin{aligned} \label{lcb} n^_{\mu}=(1/\sqrt{2},\,{\bf 0}_T,\,1/\sqrt{2}), \quad n_{\mu}=(1/\sqrt{2},\,{\bf 0}_T,\,-1/\sqrt{2}), \quad n^\cdot n =1.\end{aligned}$$ In this paper, we choose the kinematics such that the photon and one of the hadron have purely longitudinal momenta while the other hadron has both longitudinal and perpendicular momenta: $$\begin{aligned} \label{Pframe} &&P_{2\, \mu}=\frac{z_2 Q}{\sqrt{2}} n^_{\mu} + \frac{M_2^2}{z_2 Q \sqrt{2}} n_{\mu}, \quad q_{\mu}=\frac{Q}{\sqrt{2}} n^_{\mu} + \frac{Q}{\sqrt{2}} n_{\mu}, \nonumber\\ &&P_{1\, \mu}=\frac{z_1 Q}{\sqrt{2}} n_{\mu} + \frac{M_1^2+\vec{{\bf P}}_{1}^{\perp \,2}}{ z_1 Q \sqrt{2}} n^_{\mu} + P_{1\, \mu}^{\perp}.\end{aligned}$$ The leptonic momenta are $$\begin{aligned} \label{lepmom} &&l_{1\, \mu}=\frac{Q[1-\cos\theta_2]}{2\sqrt{2}}n^_{\mu} + \frac{Q[1+\cos\theta_2]}{2\sqrt{2}} n_{\mu} + l_{1\, \mu}^{\perp}, \quad l_{1\, \mu}^{\perp}=\left( \frac{Q}{2}\,\sin\theta_2,\, 0\right); \nonumber\\ &&l_{2\, \mu}=\frac{Q[1+\cos\theta_2]}{2\sqrt{2}}n^_{\mu} + \frac{Q[1-\cos\theta_2]}{2\sqrt{2}} n_{\mu} + l_{2\, \mu}^{\perp}, \quad l_{2\, \mu}^{\perp}=\left( -\frac{Q}{2}\,\sin\theta_2,\, 0\right),\end{aligned}$$ where $\theta_2$ is the angle between $\vec{P}_2$ and $\vec{l}_1$. This frame is called the $``\perp"$-frame (or the perpendicular frame) [@DBoer; @Lev93]. Below, we will omit terms of order $M^2/Q^2$. Such a kinematics has advantage for analysis of the experimental situation, where the momentum of one of the produced hadrons is measured. In order to ensure that the two hadrons are in different jets, we introduce two different variables [@Altar]: $$\begin{aligned} \label{z2u} {\cal Z}=\frac{2P_1\cdot q}{Q^2}\equiv z_1, \quad {\cal U}=\frac{P_1\cdot P_2}{P_1\cdot q}.\end{aligned}$$ However, using eq. (\[Pframe\]), the difference between ${\cal U}$ and $z_2$ is of the order of $1/Q^2$: $$\begin{aligned} {\cal U}=z_2\left[ 1 + \vec{{\bf P}}_{1}^{\perp \,2}/(z^2_1\,Q^2)\right]^{-1} \ ,\end{aligned}$$ and can be neglected in the leading order approximation which we are considering in this paper. The perpendicular projection tensor is defined as usual $$\begin{aligned} \label{gP} g_{\mu\nu}^{\perp}=g_{\mu\nu}-\hat q_\mu \hat q_\nu+\hat T_\mu \hat T_\nu,\end{aligned}$$ where the two normalized vectors $\hat q$ and $\hat T$ are constructed as $$\begin{aligned} \label{norvec} &&\hat T_{\mu}=\frac{T_{\mu}}{T}, \quad T_{\mu}=P_{2\, \mu}-\frac{P_2\cdot q}{Q^2} q_{\mu}, \nonumber\\ &&\hat q_{\mu}=\frac{q_{\mu}}{Q}.\end{aligned}$$ For simplicity, we consider the unpolarized case. The differential cross section of the corresponding $e^+ e^-$ annihilation is given by $$\begin{aligned} \label{xsec1} d\sigma(e^+e^-)=\frac{1}{2\,Q^2} \frac{d^3\, \vec{P}_1}{(2\pi)^3\, 2E_1}\, \frac{d^3\, \vec{P}_2}{(2\pi)^3 \,2E_2} \sum_X \int \frac{d^3\, \vec{P}_X}{(2\pi)^3 2E_X} \, (2\pi)^4 \, \delta^{(4)}(q-P_1-P_2-P_X)\, \left\| {\cal M} (e^+e^-)\right\|^2.\end{aligned}$$ In terms of leptonic and hadronic tensors, we have $$\begin{aligned} \label{xsec2} d\sigma(e^+e^-)=\frac{\alpha^2}{4\,Q^6}\, \frac{d^3\, \vec{P}_1}{E_1}\,\frac{d^3\, \vec{P}_2}{E_2}\, {\cal L}^{\mu\nu} {\cal W}_{\mu\nu},\end{aligned}$$ where the hadronic tensor ${\cal W}_{\mu\nu}$ is defined as $$\begin{aligned} \label{hadten} {\cal W}_{\mu\nu}=\sum_X \int \frac{d^3\, \vec{P}_X}{(2\pi)^3 2E_X} \, \delta^{(4)}(q-P_1-P_2-P_X)\, \langle 0\| J_{\mu}(0)\|P_1,P_2,P_X\rangle \langle P_1,P_2,P_X\| J_{\nu}(0)\|0\rangle.\end{aligned}$$ We can rewrite the part of the phase space corresponding to the detected hadron with momentum $P_1$ as $$\begin{aligned} \label{inphsp} \frac{d^3\, \vec{P}_1}{(2\pi)^3\, 2E_1}= \frac{d\,z_1}{(2\pi)^3\, 2 z_1} \, d^2\,\vec{{\bf P}}_{1\, \perp}.\end{aligned}$$ Because the leptonic tensor ${\cal L}_{\mu\nu}$ is independent of the hadronic momentum $P_1$, it is useful to introduce the averaged hadronic tensor, $\overline{\cal W}_{\mu\nu}^{(\perp)}$, $$\begin{aligned} \label{averhadtenPer} \overline{\cal W}_{\mu\nu}^{(\perp)}&=&\int d^2\, \vec{{\bf P}}_{1\, \perp} \, {\cal W}_{\mu\nu}^{(\perp)}.\end{aligned}$$ Finally, using (\[inphsp\]) and (\[averhadtenPer\]), the cross section (\[xsec2\]) can be presented as $$\begin{aligned} \label{xsecAV} E_2\, \frac{d\sigma(e^+e^-)}{d^3\, \vec{P}_2}= \frac{\alpha^2}{4\, Q^6}\, \frac{d\, z_1}{z_1} \,{\cal L}^{\mu\nu} \overline{{\cal W}}_{\mu\nu}^{(\perp)}.\end{aligned}$$ Note that this averaging procedure produces the hard subprocess in the Born approximation. III. Factorization procedure: Born diagram ========================================== Let us now discuss factorization of hadronic tensors corresponding to the hadron production in $e^+\, e^-$-annihilation at low $p_T$. In this case, as pointed out before, a serious conceptual problem is known to be associated with the difficulty to identify (or, in other words, to separate out) the hard subprocess. In [@ER81], it was shown that in the case of Drell-Yan process a suitable integration over the transverse momentum provides effectively the separation of the hard subprocess. We extend this approach to the factorization of the $e^+\, e^-$-annihilation hadronic tensor including leading order $\alpha_S$ corrections. Furthermore, the reproduction of the well-known DGLAP evolution kernel for both the quark and anti-quark fragmentation functions can be considered as a proof of longitudinal factorization. First, we consider a simple Born diagram depicted on Fig. \[F1\] (a). The corresponding hadronic tensor reads: $$\begin{aligned} \label{hadten2} {\cal W}_{\mu\nu}=\int d^4 k_1\, d^4 k_2\, \delta^{(4)}\left( k_1+k_2-q \right) \, {\rm tr} \left[\gamma_{\nu} \,\Theta(k_2) \,\gamma_{\mu} \,\bar\Theta(k_1) \right]+ (1\leftrightarrow 2),\end{aligned}$$ where the four-dimension $\delta$-function, representing the momentum conservation at the quark-photon vertex, will be treated as the “hard" part. The non-perturbative quark and anti-quark correlation functions $\Theta(k_2)$ and $\bar\Theta(k_1)$ are given by $$\begin{aligned} \label{FF1} &&\Theta_{\underline{\alpha\, \beta}}(k_2)= \int \frac{d^4\, \xi}{(2\pi)^4} \,e^{ik_2\cdot\xi} \,\langle 0\| \psi_{\underline{\alpha}}(\xi)\|P_2, P_{X_2}\rangle \langle P_2, P_{X_2}\| \bar\psi_{\underline{\beta}}(0)\|0 \rangle, \nonumber\\ &&\bar\Theta_{\underline{\alpha\, \beta}}(k_1)= - \int \frac{d^4\, \eta}{(2\pi)^4} \,e^{-ik_1\cdot\eta}\, \langle 0\| \bar\psi_{\underline{\beta}}(0)\|P_1, P_{X_1}\rangle \langle P_1, P_{X_1}\| \psi_{\underline{\alpha}}(\eta)\|0 \rangle,\end{aligned}$$ where the underlined indices are the Dirac indices. The summation and integration over the intermediate undetected states are implied. Note that owing to the delta function in (\[hadten2\]) the hadronic tensor cannot be expressed through the factorized $k_\perp$-integrated fragmentation functions. To express the corresponding correlation functions through the fragmentation functions, we suggest a factorization scheme inspired by the Ellis-Furmanski-Petronzio (EFP) approach [@EFP]. We first consider the formal identical transformation: $$\begin{aligned} \label{replace} &&d^4 k_1 \to d^4 k_1 \, dz^{\prime}_1\, \delta(P_1^-/k_1^- - z^{\prime}_1)= d^4 k_1 \, \frac{dz^{\prime}_1}{(z^{\prime}_1)^2}\, \delta(k_1\cdot \tilde n^-1/z^{\prime}_1), \nonumber\\ &&d^4 k_2 \to d^4 k_2 \, dz^{\prime}_2\, \delta( P_2^+/k_2^+ - z^{\prime}_2)= d^4 k_2 \, \frac{dz^{\prime}_2}{(z^{\prime}_2)^2}\, \delta(k_2\cdot \tilde n-1/z^{\prime}_2),\end{aligned}$$ where the two vectors $$\begin{aligned} \tilde n^_\mu=\frac{n^_\mu}{P_1\cdot n^}, \quad \tilde n_\mu=\frac{n_\mu}{P_2\cdot n}\end{aligned}$$ have been introduced. Variables $z^{\prime}_i$ may be interpreted as partonic fractions of the corresponding momenta of produced hadrons. As mentioned above, we treat the four-dimensional $\delta$-function as the hard part of the corresponding tensor. This assumption will be justified below. Our analysis is limited by the study of the leading twist contributions. It means that we keep only the first terms of the expansion: $$\begin{aligned} \label{deltafun} &&\delta^{(4)}\left( k_1+k_2-q \right)=\delta^{(4)}\left( \frac{P_{1\, \mu}}{z^{\prime}_1}+ \frac{P_{2\, \mu}}{z^{\prime}_2}-q \right)+ O(k_\perp)\simeq \nonumber\\ && \delta\left( \frac{P_2^+}{z^{\prime}_2}- q^+ \right)\, \delta\left( \frac{P_1^-}{z^{\prime}_1}- q^- \right) \, \delta^{(2)}\left( \frac{\vec{{\bf P}}_1^\perp}{z^{\prime}_1}\right) \ .\end{aligned}$$ The hadronic tensor is now rewritten as $$\begin{aligned} \label{hadtenPer01} {\cal W}_{\mu\nu}^{(\perp)}= \int \frac{dz^{\prime}_1}{(z^{\prime}_1)^2}\, \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \delta\left( \frac{P_2^+}{z^{\prime}_2}- q^+ \right)\, \delta\left( \frac{P_1^-}{z^{\prime}_1}- q^- \right) \, \delta^{(2)} \left( \frac{\vec{{\bf P}}_1^\perp}{z^{\prime}_1}\right)\, {\rm tr}\left[ \gamma_\nu \, \Theta(z^{\prime}_2)\, \gamma_\mu \, \bar\Theta(z^{\prime}_1) \right],\end{aligned}$$ where $$\begin{aligned} \label{Thetaz2} &&\Theta(z^{\prime}_2)\stackrel{def}{=} \int d^4 k_2 \,\delta(k_2\cdot \tilde n - 1/z^{\prime}_2) \,\Theta(k_2), \\ \label{Thetaz1} &&\bar\Theta(z^{\prime}_1)\stackrel{def}{=} \int d^4 k_1 \,\delta(k_1\cdot \tilde n^-1/z^{\prime}_1) \,\bar\Theta(k_1)\end{aligned}$$ Since we study spin-independent fragmentation functions, we first have to project the correlation functions (\[Thetaz2\]) and (\[Thetaz1\]) onto the corresponding Lorentz vector structures. Starting from the quark correlator function, we write $$\begin{aligned} \label{qcorrfun} \Theta(z^{\prime}_2)\Longrightarrow \frac{1}{4} {\rm tr}\left[ \gamma_\alpha \,\Theta(z^{\prime}_2)\right]\, \gamma_\alpha\end{aligned}$$ Using (\[FF1\]) and the integral representation of the $\delta$-function in (\[Thetaz2\]), the vector projection of the quark correlation function (see (\[qcorrfun\])) can be represented in terms of the spin-independent fragmentation function as $$\begin{aligned} \label{D2fun} &&\frac{1}{4} \int \frac{d\lambda_2}{2\pi} \,e^{i\lambda_2 /z^{\prime}_2} \, \int d\, \xi^+ \, d\, \xi^- \,d^2\, \vec{\xi}_\perp\, \delta(\lambda \tilde n^- - \xi^-) \, \delta(\xi^+)\, \delta^{(2)}(\vec{\xi}_\perp) \nonumber\\ &&{\rm tr} [\gamma_{\alpha} \, \langle 0\| \psi(\xi^+, \xi^-, \vec{\xi}_\perp)\|P_2, P_{X_2}\rangle \langle P_2, P_{X_2}\| \bar\psi(0)\|0 \rangle ] \, \gamma_{\alpha} =\frac{D(z^{\prime}_2)}{z^{\prime}_2}\, \hat P_2.\end{aligned}$$ Here, the fragmentation function $D(z^{\prime}_2)$ can also be written as $$\begin{aligned} \label{D2fun2} D(z^{\prime}_2)= \frac{z^{\prime}_2}{4(2\pi)}\int d\xi^- \,e^{i P^+_2 \xi^-/z^{\prime}_2} \, \, {\rm tr} [\gamma^+ \, \langle 0\| \psi(0,\xi^-,\vec{{\bf 0}}_\perp)\|P_2, P_{X_2}\rangle \langle P_2, P_{X_2}\| \bar\psi(0)\|0 \rangle ] \ ,\end{aligned}$$ provided that the minus co-ordinate component $\xi^-$ is equal to $\lambda_2 \tilde n^-$. Note that this fragmentation function obeys the momentum conservation sum rules [@Collins81; @Coll-Sop81; @Lev93]: $$\begin{aligned} \label{sr} \sum_{h,\, s} \int dz\,z \,D(z) = 1.\end{aligned}$$ In a similar manner, we project the anti-quark correlation function: $$\begin{aligned} \label{D1fun} &&\frac{1}{4}{\rm tr}\left[ \gamma_\alpha \, \bar\Theta(z^\prime_1)\right] \gamma_\alpha= \frac{1}{4} \int \frac{d\lambda_1}{2\pi} e^{-i\lambda_1 /z^{\prime}_1} \int d\eta^+\, d\eta^-\,d^2 \vec{\eta}_\perp \, \delta(\lambda \tilde n^{\, +} - \eta^+)\, \delta(\eta^-)\, \delta^{(2)}(\vec{\eta}_\perp) \nonumber\\ &&{\rm tr} [\gamma_{\alpha} \,\langle 0\| \bar\psi(0)\|P_1, P_{X_1}\rangle \langle P_1, P_{X_1}\| \psi(\eta^+,\eta^-,\vec{\eta}_\perp)\|0 \rangle ] \, \gamma_{\alpha} =\frac{\bar D(z^{\prime}_1)}{z^{\prime}_1}\, \hat P_1.\end{aligned}$$ Taking into account Eqs. (\[qcorrfun\])-(\[D1fun\]), we bring the hadronic tensor in the following form: $$\begin{aligned} \label{hadtenPer2} {\cal W}_{\mu\nu}^{(\perp)}&=& \int \frac{dz^{\prime}_1}{(z^{\prime}_1)^3}\, \bar D(z^{\prime}_1) \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^3} \, D(z^{\prime}_2) \nonumber\\ && \delta(P_1^-/z^{\prime}_1 - q^-) \delta (P_2^+/z^{\prime}_2-q^+) \delta^{(2)}(\vec{{\bf P}}_1^\perp/z^{\prime}_1) \, {\rm tr} \left[\gamma_{\nu} \,\hat P_2 \,\gamma_{\mu} \,\hat P_1 \right].\end{aligned}$$ Now, inserting $P_2^+$ and $P_1^-$ defined via the kinematical variables $z_2$ and $z_1$ (see (\[Pframe\])) and calculating the trace in (\[hadtenPer2\]), we get: $$\begin{aligned} \label{hadtenPer3} {\cal W}_{\mu\nu}^{(\perp)}&=& -4 g_{\mu\nu}^\perp \, \delta^{(2)}(\vec{{\bf P}}_1^\perp) \, \biggl[ z_1 \int \,dz^{\prime}_1 \, \bar D(z^{\prime}_1) \delta(z_1-z^{\prime}_1)\biggr] \biggl[ z_2 \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2}\, D(z^{\prime}_2) \delta(z_2-z^{\prime}_2)\biggr], \nonumber\\ &=&-4 g_{\mu\nu}^\perp\, \delta^{(2)}(\vec{{\bf P}}_1^\perp) \, \left[z_1 \bar D(z_1)\right] \, \left[ \frac{D(z_2)}{z_2} \right].\end{aligned}$$ From (\[hadtenPer3\]), one can see that, though the hadronic tensor ${\cal W}_{\mu\nu}^{(\perp)}$ has a formally factorized form, it does not have much sense because of the two-dimensional $\delta$-function. To eliminate it, we have to integrate over the perpendicular momentum or, in other words, to go over to the averaged hadronic tensor (\[averhadtenPer\]), $$\begin{aligned} \label{averhadtenPer3} \overline{{\cal W}}_{\mu\nu}^{(\perp)}= -4 g_{\mu\nu}^\perp \, \int \,d^2\, \vec{{\bf P}}_{1\,\perp} \, \delta^{(2)}(\vec{{\bf P}}_1^\perp) \, \left[z_1 \bar D(z_1)\right] \, \left[ \frac{D(z_2)}{z_2} \right]= -4 g_{\mu\nu}^\perp \, \left[z_1 \bar D(z_1)\right] \, \left[ \frac{D(z_2)}{z_2} \right].\end{aligned}$$ On the other hand, as we will discuss below, integration over $d^2\, \vec{{\bf P}}_{1\,\perp}$ will generate the effective diagram depicted in Fig. \[F1\] (b). Indeed, after integration, one can represent the one-dimensional $\delta$-function $\delta(z_2-z^{\prime}_2)$ as the imaginary part of the effective propagator: $$\begin{aligned} \delta(z_2-z^{\prime}_2)\Longrightarrow \Im{\rm m}\frac{1}{z_2-z^{\prime}_2} \Longrightarrow \Im{\rm m}\frac{1}{[q-P_2/z^{\prime}_2]^2}.\end{aligned}$$ The latter will reduce the diagram in Fig. \[F1\] (a) to the diagram plotted in Fig. \[F1\] (b). Notice that the appearance of this hard effective propagator justifies [a posteriori]{} the suggested generalization of the EFP factorization scheme. Namely, the four-dimensional $\delta$-function should be treated as the hard part, and the entering parton momenta should be replaced, at the leading twist level, by their longitudinal parts. Let us now discuss the above-mentioned procedure from the viewpoint of the approach of averaging given by $$\begin{aligned} \label{reldd} d^2\,\vec{{\bf P}}_{1\, \perp}= 2 z_1\, \int\int d^4\, P_1 \, \delta(P_1^2)\, \delta \left(\frac{2P_1\cdot q}{Q^2}-z_1\right).\end{aligned}$$ Indeed, integrating over $d^4\, P_1$ with the four-dimensional $\delta$-functions (see (\[hadtenPer2\])), one can observe that the one-dimensional $\delta$-function which is responsible for the mass-shell condition goes to the $\delta$-function that can be understood as the imaginary part of some effective “propagator" with large photon virtuality, [i.e.]{} $$\begin{aligned} \delta(P_1^2) \Longrightarrow \delta([q-P_2/z^{\prime}_2]^2)\sim \Im{\rm m}\frac{1}{[q-P_2/z^{\prime}_2]^2}.\end{aligned}$$ Diagrammatically, it corresponds to the case when the Born diagram transforms to the diagram plotted in Fig. \[F1\] (b). The dashed line implies the effective propagator or the factorization link. So, the averaged hadronic tensor can now be written in terms of the factorization link, $$\begin{aligned} \label{averhadtenPer4} \overline{\cal W}_{\mu\nu}^{(\perp)}=\int \,d^2\, \vec{{\bf P}}_{1\,\perp} \, {\cal W}_{\mu\nu}^{(\perp)}= \frac{4g_{\mu\nu}^\perp}{\pi} \,\left[ z_1 \bar D(z_1)\right] \biggl[ z_2 \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2}\, D(z^{\prime}_2) \Im{\rm m}\frac{1}{z_2-z^{\prime}_2}\biggr].\end{aligned}$$ The spuriously asymmetric form of (\[averhadtenPer3\]) or (\[averhadtenPer4\]) with respect to the inter changing of $z_1$ and $z_2$ emerges because of the integration over the momentum $P_1$ of one of the produced hadrons. Restoring the flavour dependence omitted above and by inserting (\[averhadtenPer3\]) into (\[xsecAV\]), we get: $$\begin{aligned} \label{xsecAV2} \frac{d\sigma(e^+e^-)}{d\,z_1\,d\,z_2\, d\cos\theta_2 }= \frac{3\,\pi\,\alpha^2}{2\, Q^2}\, (1+\cos^2\theta_2)\, \sum\limits_{a,\bar a} e^2_a \,\bar D^a(z_1) \, D^a(z_2),\end{aligned}$$ where the only remaining asymmetry is reflected in the polar angle of the detected hadron. After integration over $\theta_2$, the result becomes completely symmetric: $$\begin{aligned} \label{xsecAV3} \int d\cos\theta_2\, \frac{d\sigma(e^+e^-)}{d\,z_1\,d\,z_2\, d\cos\theta_2}= \frac{4\pi\,\alpha^2}{Q^4}\, \sum\limits_{a,\bar a} e^2_a \,\bar D^a(z_1) \, D^a(z_2).\end{aligned}$$ Concluding this section, we would like to stress that the factorization of the Born diagram can be implemented in a similar way as in the case of Drell-Yan process [@DBoer]. In this case, we transform the two-dimensional integration $d^2\,\vec{{\bf P}}_{1\, \perp}$ in the phase space (see (\[inphsp\])) to the two-dimensional integration over the photon transverse momentum $d^2\,\vec{{\bf q}}_{T}$ using the well-known relation: $d^2\,\vec{{\bf P}}_{1\, \perp}=z^2_1 d^2\,\vec{{\bf q}}_{T}$ [@DBoer]. However, the leptonic tensor ${\cal L}^{\mu\nu}$ depends not only on $Q^2$ but also on $\cos\theta_2$. Therefore, factorization with the transformation to the integration over $d^2\,\vec{{\bf q}}_{T}$ does not allow for calculating the angular dependence of the corresponding cross section. IV. Factorization procedure: leading order $\alpha_S$ corrections ================================================================= Now we proceed with the analysis of leading $\alpha_S$ corrections which are associated with the diagrams that involve gluon emission. It is known that large logarithms appearing in this case can be absorbed into the corresponding evolved fragmentation functions. We will pay a special attention to the terms with mass singularities which are extracted from the diagrams with the emission of real gluons. The corresponding Feynman diagrams with $\alpha_S$ corrections are depicted in Fig. \[F3\]. The domain of integration over the loop momentum in each diagram is in the region where the considered parton is collinear to either $P_1$ or $P_2$ momentum directions. For the sake of definiteness, we assume that the hadron with momentum $P_2$ belongs to the quark jet, while the hadron with the momentum $P_1$ – to the anti-quark jet. The opposite situation is trivially obtained by the interchange of the labels. IVa. Evolution of the quark fragmentation function -------------------------------------------------- To begin with, let us consider evolution of the quark fragmentation function $D(z)$. At the Born level, the quark which decays into the detected hadron with the momentum $P_2$ and a bunch of undetected hadrons can emit a real gluon before and after interaction with the virtual photon. The diagram depicted in Fig. \[F3\] (a) corresponds to the interaction of quark with the virtual photon before the emission of the real gluon. Choosing Feynman gauge for the gluon fields, we write the hadronic tensor corresponding to this diagram in the form resembling the Born diagram tensor: $$\begin{aligned} \label{NLOa} {\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(a)})&=& g^2 C_F \int d^4 k_1\, d^4 p\, \delta^{(4)}\left( k_1+p-q \right) \, {\rm tr} \left[\gamma_{\nu} \,\bar\Theta(k_1) \,\gamma_{\mu} \, \Omega(p) \right].\end{aligned}$$ However, instead of the quark correlator $\Theta(k_2)$, we have here the modified correlator $\Omega(p)$ defined as $$\begin{aligned} \label{Omega} \Omega(p)=\frac{\hat p}{p^2} \, \gamma_\alpha \int \frac{d^4 k_2}{(2\pi)^3}\, \delta([p-k_2]^2)\, \Theta(k_2) \gamma_\alpha \, \frac{\hat p}{p^2}.\end{aligned}$$ In principle, factorization of (\[Omega\]) can be implemented in a similar way. However, in difference with the Born diagram, the modified tensor (\[Omega\]) in (\[NLOa\]) is not completely soft because of the $p$-dependence. To factorize (\[NLOa\]), let us first express the parton momenta in terms of the corresponding fractions of the hadron momenta by means of the integral representation of unity. The definitions of the fractions for parton momenta $k_1$ and $k_2$ are the same as in (\[replace\]) whereas for the loop momentum $p$ we write $$\begin{aligned} \label{replace2} &&d^4 p \to d^4 p \, \frac{d\,y^{\prime}}{(y^{\prime})^2}\, \delta\left(p\cdot \tilde n-1/y^{\prime}\right).\end{aligned}$$ The pure soft part of (\[Omega\]) is associated with the quark correlator function $\Theta(k_2)$. At the same time, the quark propagators and the $\delta$-function emerging from the imaginary part of the gluon propagator have to be associated with the hard part of (\[Omega\]). Therefore, after expanding the $\delta$-function in (\[Omega\]) about the momentum $k_2$ around the direction defined by the hadron momentum $P_2$, we get for the correlator: $$\begin{aligned} \label{Omega2} \Omega(p)=\frac{\hat p}{p^2} \, \gamma_\alpha \int \frac{d z^{\prime}_2}{(z^{\prime}_2)^2}\, \delta\left(p^2-\frac{2p\cdot P_2}{z^{\prime}_2} \right) \, \Theta(z^{\prime}_2) \gamma_\alpha \, \frac{\hat p}{p^2}.\end{aligned}$$ From this tensor, which is the same as in the Born diagram, we single out the spin-independent quark fragmentation function (see eqs. (\[qcorrfun\]) and (\[D2fun\])). Inserting the tensor (\[Omega2\]) into the hadronic tensor (\[NLOa\]), we get: $$\begin{aligned} \label{NLOa2} &&{\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(a)})= \frac{g^2 C_F}{4(2\pi)^3} \int \frac{dz^{\prime}_1}{(z^{\prime}_1)^2}\, \int \frac{d\,y^{\prime}}{(y^{\prime})^2}\, \delta\left(\frac{P_2^+}{y^{\prime}}-q^+ \right) \delta \left(\frac{P_1^-}{z^{\prime}_1}-q^- \right) \delta^{(2)} \left(\frac{\vec{{\bf P}}_{1\, \perp}}{z^{\prime}_1} \right) \, \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \nonumber\\ &&\int dp^+\, \int d^2\,\vec{{\bf p}}_{\perp}\, \delta\left( \frac{1}{y^{\prime}}-\frac{p^+}{P_2^+} \right) \, \, \frac{p^+ - P_2^+/z^{\prime}_2}{p^4_\perp (P_2^+/z^{\prime}_2)^2} \, \, {\rm tr} \left[ \gamma_{\nu} \, \bar\Theta(z^{\prime}_1) \,\gamma_{\mu} \, \hat p\, \gamma_\alpha\, \Theta(z^{\prime}_2) \, \gamma_\alpha\, \hat p \right] \biggl\|_{p^- \sim 1/Q},\end{aligned}$$ where we decomposed again the four-dimensional $\delta$-function around the corresponding hadron directions, $$\begin{aligned} \label{4dimdelta} \delta^{(4)}\left( k_1+p-q \right) \Longrightarrow \delta\left(\frac{P_2^+}{y^{\prime}}-q^+ \right) \delta \left(\frac{P_1^-}{z^{\prime}_1}-q^- \right) \delta^{(2)} \left(\frac{\vec{{\bf P}}_{1\, \perp}}{z^{\prime}_1} \right),\end{aligned}$$ and calculated the integral over $dp^-$ with the $\delta$-function coming from the imaginary part of the gluon propagator which fixes the minus component of the loop momentum: $$\begin{aligned} \label{minp} \delta([p-k_2]^2)\Longrightarrow \frac{1}{2[p^+-P_2^+/z^{\prime}_2]} \delta\left( p^- + \frac{p^2_\perp}{2[p^+-P_2^+/z^{\prime}_2]}\right).\end{aligned}$$ We remind that $p^+$ and $P^+_2$ are large vectors. Therefore, as one can see from (\[minp\]), the minus component is thus suppressed as $1/Q$ and can be discarded in (\[NLOa2\]). In (\[NLOa2\]), we still have integration over the plus and perpendicular components of the loop momentum. Just like in Fig. \[F3\] (a), the quark with momentum $p$ emits a real gluon and transforms into the quark with momentum $k_2$. This means that $$\begin{aligned} \label{loopdec} p^+= \xi \, k_2^+, \quad dp^+=\frac{P_2^+}{z^{\prime}_2}\, d\xi\end{aligned}$$ Using (\[loopdec\]) and calculating the corresponding integral and the trace in (\[NLOa\]), we get the following contribution of the diagram \[F3\] (a) to the hadronic tensor: $$\begin{aligned} \label{NLOaFac} &&{\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(a)})= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi^2}\, C_F \, \delta^{(2)}\left( \vec{{\bf P}}_{1\, \perp} \right) \left[ z_1 \bar D(z_1)\right] \nonumber\\ &&\left[ \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \, D(z^{\prime}_2)\, \int d\xi \, \delta\left( \xi - \frac{z^{\prime}_2}{z_2}\right) \left(1-\xi \right) \int \, \frac{d^2\, \vec{{\bf p}}_\perp}{\vec{{\bf p}}_\perp^2}\right].\end{aligned}$$ As in the case of the Born diagram, one can show that the $\delta$-function $\delta(\xi-z^\prime_2/z_2)$ in (\[NLOaFac\]) should be associated with the imaginary part of the hard effective propagator, see \[F4\] (a). Note that similar arguments are valid for other diagrams needed for the study of the quark fragmentation function evolution. Integration over the two-component loop momentum $\vec{{\bf p}}_T$ should be implemented with the lower limits defined by the infrared cut-off $\lambda^2$ and the upper limit of $Q^2$. Therefore, the factor ${\rm ln}(Q^2/\lambda^2)$ which appears after this integration reflects the collinear singularity. Turning back to the averaged hadronic tensor, we thus derive $$\begin{aligned} \label{NLOaFac2} \overline{\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(a)})= \int \,d^2\, \vec{{\bf P}}_{1\,\perp} \, {\cal W}_{\mu\nu}^{(\perp)}= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right)\, \left[ z_1 \bar D(z_1)\right] \left[ \int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \, D(z^{\prime}_2) \left(1-\frac{z^{\prime}_2}{z_2} \right) \right].\end{aligned}$$ The tensor (\[NLOaFac2\]) has a completely factorized form and corresponds to \[F4\] (a) which plays a role of a ladder diagram. The diagram \[F3\] (b) does not contribute to the terms containing the mass singularity. Indeed, the hadronic tensor corresponding to such a diagram is given by $$\begin{aligned} \label{NLOb} {\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(b)})&=& g^2 C_F \int d^4 k_1\, d^4 p\, \delta^{(4)}\left( k_1+p-q \right) \, {\rm tr} \left[\bar\Theta(k_1) \, \Omega_{\mu\nu}(k_1,p) \right],\end{aligned}$$ where $$\begin{aligned} \label{Omegamunu} \Omega_{\mu\nu}(k_1, p)=\gamma_{\alpha} \int \frac{d^4 k_2}{(2\pi)^3}\, \delta([p-k_2]^2)\, \frac{\hat p + \hat k_1 - \hat k_2}{(p + k_1 - k_2)^2}\, \gamma_{\mu} \Theta(k_2)\, \gamma_{\nu}\, \frac{\hat p + \hat k_1 - \hat k_2}{(p + k_1 - k_2)^2}\, \gamma_\alpha .\end{aligned}$$ In the same manner as before, one can see that, after integration over $d\,p^-$ with the $\delta$-function originating from the gluon propagator, the denominator of (\[NLOb\]) does not contain the term with $\vec{{\bf p}}_\perp^2$ that produces a mass singularity. As a result, this contribution can be discarded. It is necessary to note that the diagram [\[F3\]]{} (b) effectively corresponds to the diagram Fig. \[F4\] (b) with the self-energy insertion into the quark propagator. Two diagrams presented on \[F3\] (c) and (d) contribute to the hadronic tensor in a following way: $$\begin{aligned} \label{NLOcd} {\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(c+d)})=2 g^2 C_F \int d^4 k_1\, d^4 p\, \delta^{(4)}\left( k_1+p-q \right) \, {\rm tr} \left[ \gamma_{\nu}\, \bar\Theta(k_1) \, \Omega_{\mu}(k_1,p) \right],\end{aligned}$$ where $$\begin{aligned} \label{Omeganu} \Omega_{\mu}(k_1, p)=\gamma_{\alpha} \int \frac{d^4 k_2}{(2\pi)^3}\, \delta([p-k_2]^2)\, \frac{\hat p + \hat k_1 - \hat k_2}{(p + k_1 - k_2)^2}\, \gamma_{\mu} \Theta(k_2)\, \gamma_{\alpha}\, \frac{\hat p}{p^2}.\end{aligned}$$ In (\[NLOcd\]), the denominator contains the necessary power of $\vec{{\bf p}}_\perp^2$ and we have to keep only the zeroth order of $\vec{{\bf p}}_\perp^2$ in the trace. Then, following the scheme outlined earlier for other diagrams, we derive the following expressions for the hadronic tensor $$\begin{aligned} \label{NLOcdFac} {\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(c)+(d)})= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi^2}\, C_F \, \delta^{(2)}\left( \vec{{\bf P}}_{1\, \perp} \right) \left[z_1 \bar D(z_1) \right] \, \left[\int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \, D(z^{\prime}_2)\, \frac{2\, z^{\prime}_2/z_2}{1-z^{\prime}_2/z_2}\int \, \frac{d^2\, \vec{{\bf p}}_\perp}{\vec{{\bf p}}_\perp^2} \right];\end{aligned}$$ and for the averaged hadronic tensor: $$\begin{aligned} \label{NLOcdFac2} \overline{\cal W}_{\mu\nu}^{(\perp),\, q}({\rm Fig.\ref{F3}(c)+(d)})= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \, \left[z_1 \bar D(z_1) \right] \, \left[\int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \, D(z^{\prime}_2)\, \frac{2\, z^{\prime}_2/z_2}{1-z^{\prime}_2/z_2} \right].\end{aligned}$$ Thus, we have all the ingredients for the derivation of the factorized hadronic tensor including the $\alpha_S$ corrections and mass singularities. Summing of eqs (\[NLOaFac\]) and (\[NLOcdFac\]) together with the contributions related to the virtual gluon emission gives $$\begin{aligned} \label{DEE} \overline{\cal W}_{\mu\nu}^{(\perp),\, q}= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \left[z_1 \bar D(z_1) \right] \, \left[\int \frac{dz^{\prime}_2}{(z^{\prime}_2)^2} \, D(z^{\prime}_2)\, \left( \frac{1+(z^{\prime}_2/z_2)^2}{1-z^{\prime}_2/z_2} \right)_+\, \right].\end{aligned}$$ The factorization scale $\mu_F$ can be introduced by the standard decomposition: ${\rm ln}(Q^2/\lambda^2)={\rm ln}(Q^2/\mu^2_F) + {\rm ln}(\mu_F^2/\lambda^2)$ where the first term should be combined with the hard part of the corresponding hadronic tensor whereas the second one with the soft part. If we choose $\mu_F^2=Q^2$, the sum of (\[DEE\]) with the contribution of the Born diagram (\[averhadtenPer3\]) leads to the following substitution for the quark fragmentation function: $$\begin{aligned} \label{DQ2} D(z_2) \Longrightarrow D(z_2)+ \frac{\alpha_S}{2\pi}\, C_F\, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \, \int\limits_{z_2}^1 \frac{dy_2}{y_2} \, D(z_2/y_2)\, \left(\frac{1+y_2^2}{1-y_2}\right)_+.\end{aligned}$$ As a result, the quark fragmentation function acquires $Q^2$-dependence and satisfies the DGLAP evolution equation: $$\begin{aligned} \label{DGLAPq} \frac{d \, D(z_2)}{d\, {\rm ln} \,Q^2}= \int\limits_{z_2}^1 \frac{dy_2}{y_2} \, D(z_2/y_2)\, V_{qq}(y_2), \quad V_{qq}(y)=\frac{\alpha_S}{2\pi}\, C_F\, \left(\frac{1+y^2}{1-y}\right)_+ .\end{aligned}$$ IVb. Evolution of the anti-quark fragmentation function ------------------------------------------------------- We will now consider the anti-quark fragmentation functions. The anti-quark sector in the fragmentation can be studied in similarly to the quark fragmentation function case. However, there are some minor differences in comparison to the quark fragmentation function evolution. First of all, the diagram \[F3\] (b) will now play a role of the ladder diagram. Let us write down the hadronic tensor corresponding to this diagram: $$\begin{aligned} \label{antiNLOb} {\cal W}_{\mu\nu}^{(\perp),\, \bar q}({\rm Fig.\ref{F3}(b)})&=& g^2 C_F \int d^4 k_2\, d^4 m\, \delta^{(4)}\left( k_2+m-q \right) \, {\rm tr} \left[\gamma_{\mu} \, \Theta(k_2) \,\gamma_{\nu} \, \bar\Omega(m) \right],\end{aligned}$$ where $$\begin{aligned} \label{antiOmega} \bar\Omega(m)=\frac{\hat m}{m^2} \, \gamma_\alpha \int \frac{d^4 k_1}{(2\pi)^3}\, \delta([m-k_1]^2)\, \bar\Theta(k_1) \gamma_\alpha \, \frac{\hat m}{m^2}.\end{aligned}$$ As in the case of the quark sector, we first introduce the definition of the corresponding parton fractions (see eqs. (\[replace\]) and (\[replace2\])). Then, the corresponding $\delta$-functions have to be decomposed in the directions defined by the hadron momenta. We obtain $$\begin{aligned} \label{del1} \delta^{(4)}\left( k_2+m-q \right) \Longrightarrow \delta\left(m^+ + \frac{P_2^+}{z^{\prime}_2} -q^+\right) \, \delta\left(\frac{P_1^-}{y^{\prime}} -q^-\right) \, \delta^{(2)}\left( \frac{\vec{{\bf P}}_1^{\perp}}{y^{\prime}}\right)\end{aligned}$$ for the four-dimensional $\delta$-function which is responsible for the momentum conservation at the local vertex: $$\begin{aligned} \label{del2} \delta([m-k_1]^2) \Longrightarrow \frac{1}{2[m^- - P^-_1/z^{\prime}_1]} \delta\left( m^+ + \frac{m^2_{\perp}}{2[m^- - P^-_1/z^{\prime}_1]} \right)\end{aligned}$$ for the $\delta$-function in the tensor (\[antiOmega\]). Integration over $dm^+$ with $\delta$-function (\[del2\]) fixes the plus component of the loop momentum to be a small variable. As result, $m^+$ can be neglected in the corresponding expressions. Omitting the details, since all stages of calculations are exactly the same as for the quark sector. Below, we write the expression for the averaged hadronic tensor corresponding to Fig. \[F3\] (b): $$\begin{aligned} \label{antiqFF4av} \overline{\cal W}_{\mu\nu}^{(\perp),\, \bar q}({\rm Fig.\ref{F3}(b)})&=& -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \left[\frac{D(z_2)}{z_2} \right] \left[z_1^2\int \frac{dz^{\prime}_1}{(z^{\prime}_1)^2}\, \bar D(z^{\prime}_1) \left( 1-\frac{z_1^{\prime}}{z_1} \right) \right].\end{aligned}$$ The next non-zero contribution comes from the diagrams presented on Fig. \[F3\] (c) and (d). These diagrams give us the following expression: $$\begin{aligned} \label{antiNLOcd} {\cal W}_{\mu\nu}^{(\perp),\, \bar q}({\rm Fig.\ref{F3}(c+d)})&=& 2 g^2 C_F \int d^4 k_2\, d^4 m\, \delta^{(4)}\left( k_2+m-q \right) \, {\rm tr} \left[\Theta(k_2) \,\gamma_{\nu} \, \bar\Omega_{\mu}(k_2, m) \right],\end{aligned}$$ where $$\begin{aligned} \label{antiOmegacd} \bar\Omega_{\mu}(k_2, m)=\frac{\hat m}{m^2} \, \gamma_\alpha \int \frac{d^4 k_1}{(2\pi)^3}\, \delta([m-k_1]^2)\, \bar\Theta(k_1) \gamma_{\mu}\, \frac{\hat m - \hat k_1 + \hat k_2}{( m - k_1 + k_2)^2} \,\gamma_\alpha .\end{aligned}$$ Calculating the tensor (\[antiNLOcd\]) in a way similar to the derivation of (\[NLOcd\]), we obtain the averaged hadronic tensor: $$\begin{aligned} \label{antiqFFcdav} \overline{\cal W}_{\mu\nu}^{(\perp),\, \bar q}({\rm Fig.\ref{F3}(c)+(d)})= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \left[\frac{D(z_2)}{z_2} \right] \left[z_1^2\int \frac{dz^{\prime}_1}{(z^{\prime}_1)^2}\, \bar D(z^{\prime}_1) \frac{2 z_1^{\prime}/z_1}{1-z_1^{\prime}/z_1} \right].\end{aligned}$$ Diagram \[F3\] (a) in this case does not contribute to the evolution of the anti-quark fragmentation function due to the same arguments as we presented for the quark sector discussing the diagram \[F3\] (b). Thus, combining all the diagrams and adding the contributions from the virtual gluon emissions, we get: $$\begin{aligned} \label{DbarEE} \overline{\cal W}_{\mu\nu}^{(\perp),\, \bar q}= -2\, g_{\mu\nu}^\perp\, \frac{\alpha_S}{\pi}\, C_F \, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \left[z_1^2 \int \frac{dz^{\prime}_1}{(z^{\prime}_1)^2} \, \bar D(z^{\prime}_1)\, \left(\frac{1+(z^{\prime}_1/z_1)^2}{1-z^{\prime}_1/z_1}\right)_+ \right] \, \left[\frac{D(z_2)}{z_2} \right].\end{aligned}$$ Again, as for the quark case, the summation of the hadronic tensor (\[DbarEE\]) with the Born hadronic tensor modifies the anti-quark fragmentation function: $$\begin{aligned} \label{DbarQ2} \bar D(z_1) \Longrightarrow \bar D(z_1)+ \frac{\alpha_S}{2\pi}\, C_F\, {\rm ln}\left(\frac{Q^2}{\lambda^2}\right) \, \int\limits_{z_1}^1 \frac{dy_1}{y_1} \, \bar D(z_1/y_1)\, \left(\frac{1+y_1^2}{1-y_1}\right)_+.\end{aligned}$$ As a result of it, the anti-quark fragmentation function becomes $Q^2$-dependent and obeys the DGLAP evolution equation (see, (\[DGLAPq\])). Besides, the introduction of the hard effective propagators as in the quark case leads to the factorized Feynman diagrams which are completely analogous to the diagrams on Fig. \[F4\]. V. $k_T$-dependent functions: Born approximation ================================================ In the preceding sections, we focused on the case of the spin-independent integrated fragmentation functions which in a way was a test of our approach. Since we plan to extend our approach to $k_T$-unintegrated functions (in particular, Collins function) we discuss the Born approximation of $e^+ e^-$-annihilation, involving both the Collins fragmentation function and the spin-dependent fragmentation function. Namely, we consider the production of transverse polarized quark-antiquark pair associated with $\sigma_{\mu\nu}$-structures in the corresponding matrix elements, see below. More exactly, the nonperturbative blob (see, for example, Fig. \[F1\]), related to the detected baryon, whose transverse polarization is correlated to the transverse polarization of the quark can be described by the spin-dependent fragmentation function. At the same time, in the other blob, the transition of the transverse polarized antiquark with the intrinsic transverse momentum into the unpolarized hadron is described by the Collins fragmentation function. In latter case, the mentioned transition is related to the azimuthal asymmetry distribution of hadrons. Note that the symmetric case, with two similar fragmentation functions, has been well-studied from both the theoretical and experimental points of view [@Efremov1998; @Kumano; @Bell; @Boer:2008fr]. The asymmetric situation requires the simultaneous measurement of the meson azimuthal asymmetry in one jet and baryon (say, $\Lambda$) polarization in another one, Such measurements may be performed at BELLE [^1]. We use the co-ordinate (or the impact parameter) representation where the explicit definition of the transverse momentum is not necessary. In this case, the corresponding hadronic tensor takes the form: $$\begin{aligned} \label{HTxspace} \Delta {\cal W}_{\mu\nu}=\int \frac{d^4 \xi}{(2 \pi)^4} e^{i(q\cdot\xi)}\, {\rm tr}\left[ \gamma_{\mu} \hat\Theta_1(\xi) \gamma_\nu \hat{\bar\Theta}_2(\xi)\right].\end{aligned}$$ Here, the spin-dependent fragmentation function has been defined in the co-ordinate space as $$\begin{aligned} \label{sp} \hat\Theta_1(\xi) \Rightarrow \sigma_{\alpha \beta}\gamma_5 P_2^\alpha S^\beta \int\limits_0^1 \frac{dz_2^\prime}{(z_2^{\prime})^2} e^{-i (P_2\cdot \xi)/{z_2^\prime} } H_{1\, T}(z_2^\prime),\end{aligned}$$ where $S$ denotes the hadron transverse polarion, and the Collins function’s analog in the coordinate space is given [@Ter04] by $$\begin{aligned} \label{coll} \hat{\bar\Theta}_2(\xi) \Rightarrow && i M \sigma_{\alpha \beta} P_1^\alpha \xi^\beta \int\limits_0^1 \frac{dz_1^\prime}{(z_1^{\prime})^2} e^{-i (P_1\cdot \xi)/{z_1^\prime} } \bar H^{\perp}_1(z_1^\prime), \quad \bar H^{\perp}_1(z_1^\prime) = \int d k_T^2 \, \frac {k_T^2 }{M^2}\, \bar H^{\perp}_1(z_1^\prime, k_T^2 ),\end{aligned}$$ where $\xi$ is the (transverse) position in co-ordinate space; while $M$, introduced due to the dimensional analysis, implies a parameter of the order of jet mass being the only dimensionful parameter in the soft part. The position in the coordinate space $\xi$ (see, (\[coll\])) yields the derivative over $P_1$ in momentum space. Therefore, inserting (\[sp\]) and (\[coll\]) in (\[HTxspace\]), we get: $$\begin{aligned} \label{HTxspace2} \Delta {\cal W}_{\mu\nu}=\int\frac{dz_1^\prime}{z_1^{\prime}} \int\frac{dz_2^\prime}{(z_2^{\prime})^2} \biggl\{ \frac{\partial}{\partial P_1^-} \delta(q^- - \frac{P^-_1}{z_1^{\prime}})\biggr\} \delta(q^+ - \frac{P^+_2}{z_2^{\prime}}) \delta^{(2)}(\frac{\vec{{\bf P}}_1^\perp}{z_1^{\prime}}) \bar H^{\perp}_1(z_1^\prime) \, H_{1\, T}(z_2^\prime) T_{\mu\nu}(P_1,P_2,S_T),\end{aligned}$$ where $$\begin{aligned} T_{\mu\nu}(P_1,P_2,S_T)={\rm tr}\left[ \gamma_{\mu} (i M \sigma_{+-} P_1^-) \gamma_{\nu} (\sigma_{-T}\gamma_5 P_2^+ S_T)\right].\end{aligned}$$ Using the kinematics defined above, we derive $$\begin{aligned} \label{HTxspace3} \Delta {\cal W}_{\mu\nu}= T_{\mu\nu} \delta^{(2)}(\vec{{\bf P}}_1^\perp) \biggl[ \int dz_1^\prime \left(z_1^\prime \bar H^{\perp}_1(z_1^\prime)\right) \delta^{(1)}(z_1^\prime -z_1)\biggr] \biggl[\int\frac{dz_2^\prime}{z_2^{\prime}}H_{1\, T}(z_2^\prime) \delta(z_2^\prime-z_2)\biggr].\end{aligned}$$ Then, calculating the averaged hadronic tensor, one has the following: $$\begin{aligned} \label{averHTxspace} \overline{\Delta\cal W}_{\mu\nu}= \int \,d^2\, \vec{{\bf P}}_{1\,\perp} \, \delta^{(2)}(\vec{{\bf P}}_1^\perp) \Delta {\cal W}_{\mu\nu}= T_{\mu\nu} \biggl[ z_1 \bar H^{\perp}_1(z_1) \biggr]^{\prime} \biggl[\frac{H_{1\, T}(z_2)}{z_2}\biggr].\end{aligned}$$ Note that the analogous asymmetric combination corresponding to SIDIS process related to ours by crossing was also considered in [@Ter04]. In that case, the incoming quark in SIDIS is described by the transversity distribution function: $$\begin{aligned} \hat h(\eta) = \sigma_{\mu \nu} \gamma_5 P_1^\mu S^\nu \int \limits_0^1 dx e^{i x (P_1\cdot\eta) } h (x), \label{tr}\end{aligned}$$ where $S_\mu$ is the target polarization. Using (\[coll\]) and (\[tr\]), the corresponding hadronic tensor takes the form: $$\begin{aligned} \label{spidis} \Delta {\cal W}^{\mu \nu} = \int \frac{d^4 \xi}{(2 \pi)^4} e^{-i(q\cdot\xi)} \, {\rm tr} [\hat h(\xi) \gamma^\mu \hat H(\xi) \gamma^\nu ].\end{aligned}$$ To study the $k_T$ (or $\vec{{\bf P}}_{2\,\perp}$) distributions in SIDIS and the related asymmetries, we should consider the weighted hadronic tensor which projects out the corresponding moment of the Collins function: $$\begin{aligned} \label{spidisn} \Delta_n \overline{{\cal W}}^{\,\mu \nu} = \int d^4 P_2 \delta( P_2^2) (P_2\cdot n_\perp) \delta \left(\frac{P_1\cdot P_2 }{P_1\cdot q} - z\right) \Delta {\cal W}^{\mu \nu},\end{aligned}$$ where $n_\perp$ is the unit transverse $4-$vector $(n_\perp\cdot P_1=n_\perp\cdot q=0,\, n_\perp^2=-1)$ which defines the transverse direction. Using (\[coll\]), (\[tr\]) and (\[spidisn\]), one can see that the derivative $\partial^\alpha\equiv \partial/\partial P_2^\alpha$ in $$\begin{aligned} \label{spiddn} &&\Delta_n \overline{{\cal W}}^{\,\mu \nu}= i M \int d^4 P_2 \,(P_2\cdot n_\perp)\, \delta(P_2^2) \,\delta \left(\frac{P_1\cdot P_2}{P_1\cdot q} - z\right) \nonumber\\ &&\int dx \, dz^{\prime} \partial^\alpha \delta(x P_1+ q -P_2/z^{\prime}) h(x) (z^{\prime} I (z^{\prime})) {\rm tr} [\gamma_5 \hat P_1 \hat S \gamma^\mu [\hat P_2 \gamma_\alpha] \gamma^\nu]\end{aligned}$$ should act only on $(P_2\cdot n_\perp)$, so that the result [@Ter04] is equal to the standard expression for the contribution of Collins function, except that the role of the direction of intrinsic transverse momentum is played by the auxiliary transverse vector $n_\perp$: $$\begin{aligned} {\rm tr} [\hat p_1 \hat S \gamma_5 \gamma^\mu \hat p_3 \hat n_\perp \gamma^\nu] \Longrightarrow {\rm tr} [\hat p_1 \hat S \gamma_5 \gamma^\mu \hat p_3 \hat k_T \gamma^\nu].\end{aligned}$$ This substitution does not change the azimuthal dependence, as the weighted integration corresponds to azimuthal average: $$\begin{aligned} \langle d \sigma(\phi_h) \cos(\phi_h - \phi_n)\rangle = \cos \phi_n \langle d \sigma(\phi_h) \cos(\phi_h) \rangle + \sin \phi_n \langle d \sigma(\phi_h) \sin(\phi_h)\rangle.\end{aligned}$$ As a result the azimuthal dependence of the cross-section is transferred to the dependence on the angle $\phi_n$, and $I(z)$ corresponds to the [moment]{} of the Collins function: $$I(z) \sim \int d k_T^2 \frac {k_T^2 }{M^2} H_1(z,k_T^2 ).$$ Thus, to describe the Collins effect we suggest the $k_T$-dependent fragmentation function to be written in the co-ordinate space, where no specification of intrinsic $k_T$ is required. Therefore, the calculation of radiation corrections and evolution may be performed in the same way as in this paper. Moreover, this approach may be applied [@Ter04] for the description of higher weighted moments in $k_T$ and $p_T$. Say, choosing the term of order $\xi^2$ in the expansion of unpolarized fragmentation function in the coordinate space and taking the $p_T$ moment of the hadronic tensor weighted with $p_T^2$ corresponds to the account of the width of the $k_T$-dependent unpolarized fragmentation function. The exponential shape of the latter, in turn, corresponds to partial resummation [@Ter04] of the infinite series of higher twists, analogous to that leading to the appearance of non-local vacuum condensates, when vacuum rather than hadronic averages are considered. The expansion in coordinate space accompanied by taking the respective weighted moments of cross-sections provides a complementary definition of observables, accounting for the shape of $k_T$-dependent fragmentation (and distribution) functions. VI. Conclusions =============== We described a method which allows us to prove a factorization of the process with two fragmentation functions. We would like to point out that the presented method can be applied for any two-current process. The difficulties of factorization for such kind of processes emerge for the case when the kinematical transversalities inside the hadron are rather small. It leads to the problem in the definition of what is the hard subprocess for the process. Following the idea of the paper [@ER81], we showed that the corresponding $\delta$-functions in the hadronic tensors should be treated as the hard parts. It is based on the observation that these $\delta$-functions can be associated with the imaginary parts of the effective propagators related to the well-defined hard subprocess. As a result, we finally have the completely factorized expression for the hadronic tensor with the evolved fragmentation functions. In this paper, the proposed method has been tested in the simplest case when the differential cross section of $e^+ e^-$ annihilation is related to the spin-independent integrated fragmentation functions. We also extended our approach to the study of the spin-dependent structures and $k_T$-dependent fragmentation functions (Collins function and transversity fragmentation function). This will allow us to perform the leading order QCD fits of relevant experimental data [@Kumano]. Acknowledgments =============== We would like to thank D. Boer, A.V. Efremov, A. Bacchetta, B. Pire, A.V. Radyushkin and S. Taneja for stimulating discussions. I.V.A. would like to express the gratitude for very warm hospitality at Ecole Polytechnique. This investigation was partially supported by the Heisenberg-Landau Programme (Grant 2008), the Deutsche Forschungsgemeinschaft under contract 436RUS113/881/0, the EU-A7 Project Transversity, and the RFBR (Grants 06-02-16215, 08-02-00896, and 07-02-91557),Russian Federation Ministry of Science and Education (Grant MIREA 2.2.2.2.6546), and the EcoNet program. [99]{} Collins J.C. // Nucl.Phys.B. 1993. V.396. P.161; Efremov A.V., Smirnova O.G. and Tkachev L.G.// Nucl.Phys.Proc.Suppl. 1999. V.74 P.49; Bravar A.// Nucl. Phys.Proc.Suppl. 1999. V.79. P.520; Abe K. [et al.*]{} \[Belle Collaboration\] hep-ex/0507063; Efremov A.V., Goeke K. and Schweitzer P.// Phys.Rev.D 2006. V. 73 P.094025; Collins J.C. and Soper D.E.// Nucl.Phys.B 1981. V.193 P.381; Ellis R.K., Georgi H., Machacek M., Politzer H.D. and Ross G.G.// Nucl.Phys.B 1979. V.152 P.285; Collins J.C., Soper D.E. and Sterman G.// Nucl.Phys.B 1985. V.250 P.199; Ji X.D., Ma J.P. and Yuan F.// Phys.Rev.D 2005. V.71 P.034005; Altarelli G., Ellis R.K., Martinelli G. and Pi S.Y.// Nucl.Phys.B 1979. V.160 P.301; de Florian D. and Vanni D.// Phys.Lett.B 2004. V.578 P.139; Boer D., Jakob R. and Mulders P.J.// Nucl.Phys.B 1997. V.504 P.345; Radyushkin A.V.// Phys.Lett.B 1977. V.69 P.245; Efremov A.V. and Radyushkin A.V.// Theor.Math.Phys. 1981. V.44 P.774; Teryaev O.V.// Phys.Part.Nucl. 2004. V.35 P.S24; Levelt J. and Mulders P.J.// Phys.Rev.D 1994 V.49 P.96; Ellis R.K., Furmanski W. and Petronzio R.// Nucl.Phys.B 1983. V.212 P.29; Collins J.C. and Soper D.E.// Nucl.Phys.B 1982. V.194 P.445; M. Hirai, S. Kumano, M. Oka and K. Sudoh // Phys.Rev.D 2008. V.77 P.017504; Belle Collaboration (R. Seidl et al.) // BELLE-2008-14, KEK-2008-8, May 2008. 24pp. arXiv:0805.2975 \[hep-ex\] D. Boer // arXiv:0804.2408 \[hep-ph\]. $$\includegraphics[width=12cm]{Born-a-Fac.ps}$$ $$\includegraphics[width=12cm]{NLO-F3.ps}$$ $$\includegraphics[width=12cm]{NLO-Fac-F4.ps}$$ [^1]: M. Grosse Perdekamp, private communication.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We outline the proof of a conjecture of Kontsevich on the isomorphism between the group of polynomial symplectomorphisms in $2n$ variables and the group of automorphisms of the $n$-th Weyl algebra over complex numbers. Our proof uses lifting of polynomial symplectomorphisms to Weyl algebra automorphisms by means of approximation by tame symplectomorphisms and gauging of the lifted morphism. Approximation by tame symplectomorphisms is the symplectic version of the well-known theorem of D. Anick and is a result of our prior work.' author: - 'Alexei Kanel-Belov[^1]' - 'Andrey Elishev[^2]' - 'Jie-Tai Yu[^3]' title: 'Automorphisms of Weyl Algebra and a Conjecture of Kontsevich' --- Introduction ============ The objective of this short note is to explain a method of attack on a conjecture formulated by one of us together with Kontsevich in [@BKK1], which is henceforth referred to as the Kontsevich conjecture. The conjecture, in its most straightforward form, states that the automorphism group of the $n$-th Weyl algebra over an algebraically closed field $\mathbb{K}$ of characteristic zero is isomorphic to the group of polynomial symplectomorphisms – that is, polynomial automorphisms preserving the symplectic structure – of the affine space $\mathbb{A}^{2n}_{\mathbb{K}}$. The statement can be reformulated as $\operatorname{Aut}W_n(\mathbb{K})\simeq \operatorname{Aut}P_n(\mathbb{K})$ – where the $n$-th Weyl algebra $W_{n}(\mathbb{K})$ over $\mathbb{K}$ is by definition the quotient of the free associative algebra $$\mathbb{K}\langle a_1,\ldots,a_n,b_1,\ldots,b_n\rangle$$ by the two-sided ideal generated by elements $$b_ia_j-a_jb_i-\delta_{ij},\;\;a_ia_j-a_ja_i,\;\;b_ib_j-b_jb_i,$$ with $1\leq i,j\leq n$, while the algebra $P_n(\mathbb{K})$ is the commutative polynomial algebra $\mathbb{K}[x_1,\ldots,x_{2n}]$ carrying an additional structure of the Poisson algebra via the standard Poisson bracket – that is, a bilinear map $$\lbrace \;,\;\rbrace:\mathbb{K}[x_1,\ldots,x_{2n}]\otimes \mathbb{K}[x_1,\ldots,x_{2n}]\rightarrow \mathbb{K}[x_1,\ldots,x_{2n}]$$ that turns $\mathbb{K}[x_1,\ldots,x_{2n}]$ into a Lie algebra and acts as a derivation with respect to polynomial multiplication (therefore, automorphisms of such algebras are required to preserve the additional structure). The standard Poisson bracket is defined as $$\lbrace x_i,x_j\rbrace = \delta_{i,n+j}-\delta_{i+n,j},$$ with $\delta_{ij}$ meaning the Kronecker delta. Several generalizations of Kontsevich conjecture are known; the most obvious one is obtained by replacing the algebraically closed ground field $\mathbb{K}$ with the field $\mathbb{Q}$ of rational numbers. Other generalizations are somewhat more elaborate and are discussed at length in [@BKK1]. The setting in which Conjecture 1 naturally arises is that of deformation quantization of polynomial algebra. The commutative Poisson algebra $P_n$ serves as the classical counterpart to the algebra $W_n$ of polynomial differential operators. It is therefore sensible to ask whether the quantization preserves the automorphism group. One then, in order to answer this question, tries to construct either a direct homomorphism $$\operatorname{Aut}W_n(\mathbb{K})\rightarrow \operatorname{Aut}P_n(\mathbb{K})$$ or an inverse $$\operatorname{Aut}P_n(\mathbb{K})\rightarrow \operatorname{Aut}W_n(\mathbb{K}).$$ Known ways of accomplishing that goal are somewhat involved. We briefly comment on the relatively accessible case of (direct) group homomorphism $$\operatorname{Aut}W_n(\mathbb{K})\rightarrow \operatorname{Aut}P_n(\mathbb{K})$$ with $\mathbb{K}$ being algebraically closed [^4]. The idea is to realize the Weyl algebra $W_n$ as a subalgebra in an algebra whose center is large enough to contain the polynomial algebra $\mathbb{C}[x_1,\ldots,x_{2n}]$, and then, starting with an automorphism $\varphi\in\operatorname{Aut}W_n(\mathbb{C})$, restrict it to the polynomial algebra. At this point, reduction to positive characteristic starts playing an important role. Namely, one represents the ground field as a reduced direct product of algebraically closed fields of characteristic $p$, $$\mathbb{C}\simeq \prod_{p}\; '\; \mathbb{F}_p$$ where $p$ runs over all prime numbers, and reduction is taken with respect to a free ultrafilter $\mathcal{U}$ on the index set of natural numbers. In other words, $\prod'_{p} \mathbb{F}_p$ is the quotient of the direct product $\prod_{p} \mathbb{F}_p$ by the maximal [^5] ideal generated by $\mathcal{U}$ as in [@FMS]. Such a procedure is sometimes referred to in literature as reduction modulo infinite prime (the infinite prime being the sequence of prime numbers that indexes the direct product – in this case the standard prime number sequence; one could as well take for such a sequence any sequence of primes which is not equivalent under the chosen ultrafilter to a stationary sequence). The point of this construction consists in the fact that in positive characteristic the Weyl algebra $W_n$ has a huge center isomorphic to the polynomial algebra $\mathbb{F}_p[x_1^p,\ldots,x_n^p,d_1^p,\ldots,d_n^p]$ (with $x_i,\; d_j$ being the generators), while nothing of the sort is the case of characteristic zero. Therefore, the reduced product of Weyl algebras will just be the larger algebra one is looking for – one whose center contains a copy of the polynomial algebra. A significant detail is given by the fact that the Weyl algebra commutator naturally induces a Poisson structure on the polynomial subalgebra of the larger center, thus making the resulting automorphism symplectic. Thus one constructs a homomorphism $\operatorname{Aut}W_n(\mathbb{C})\rightarrow \operatorname{Aut}P_n(\mathbb{C})$ which is a candidate for the simplest version of the Kontsevich conjecture. For the sake of brevity we have left out the details and refer the interested reader to the works [@BKK1; @BKK2] and [@K-BE2]. This homomorphism is injective, and induces an isomorphism of subgroups of tame automorphisms (the definition of tame automorphism is given below). An identical procedure produces a monoid homomorphism between the sets of endomorphisms of $W_n$ and $P_n$; this fact has been used to establish a stable equivalence between the Dixmier conjecture [@Dix] (any endomorphism of $W_n$ is invertible – open for all $n$ as of time of writing) and the Jacobian conjecture, cf. [@BKK2; @Tsu1; @Tsu2]. The direct homomorphism can be made explicitly independent of the prime number sequence by means of a non-standard (inverse) Frobenius twist of coefficients. It is, however, insufficient to guarantee its independence of the choice of infinite prime and the ultrafilter completely, for integer combinations of coefficients (coming from applying Weyl algebra commutation relations) could differ for different such choices. The question of independence of infinite prime is therefore non-trivial. In our prior work [@K-BE2] we have provided a proof of independence, which however relies on the homomorphism in question to be one-to-one. The present paper focuses on the reverse approach. Starting with a polynomial symplectomorphism, we construct an automorphism of the power series completion of the Weyl algebra and then argue that the power series that are images of the generators of $W_n$ must be polynomials. The procedure is referred to as the lifting throughout the text. Central to it is the fact that the subgroup of tame symplectomorphisms ${\operatorname{TAut}}P_n$ is dense (with respect to power series topology) in $\operatorname{Aut}P_n$ – a fact we have obtained recently [@KGE]. Approximation of polynomial automorphisms by tame automorphisms was developed by Anick [@An] and has already become a classical result in algebraic geometry. Our work serves, in a way, as a development of Anick’s results to the symplectic case. Tame symplectomorphisms, topology, and approximation ==================================================== This section reviews the background and results on approximation by tame automorphisms necessary in our context. Most of the theory, as well as detailed proofs, can be found in the classical work of Anick [@An]. Tame symplectomorphism approximation is the main result of our recent work with S. Grigoriev and W. Zhang [@KGE]. Let $A_N=\mathbb{K}[x_1,\ldots, x_{N}]$ be the commutative polynomial algebra over a field $\mathbb{K}$, and let $\varphi$ be an algebra endomorphism. Any such endomorphism can be identified with the ordered set $$(\varphi(x_1),\;\varphi(x_2),\;\ldots,\; \varphi(x_N))$$ of images of generators of the algebra. The polynomials $\varphi(x_i)$ may be represented as sums of their respective homogeneous components; this may be expressed formally as $$\varphi = \varphi_0+\varphi_1+\cdots,$$ where $\varphi_k$ is a string of length $N$ whose entries are homogeneous polynomials of total degree $k$.[^6] The height $\operatorname{ht}(\varphi)$ of an endomorphism $\varphi$ is defined as $$\operatorname{ht}(\varphi)=\inf\lbrace k\;\|\;\varphi_k\neq 0\rbrace,\;\;\operatorname{ht}(0)=\infty.$$ This is not to be confused with degree of endomorphism, which is defined as $$\operatorname{deg}(\varphi)=\sup\lbrace k\;\|\;\varphi_k\neq 0\rbrace.$$ The height $\operatorname{ht}(f)$ of a polynomial $f$ is defined quite similarly to be the minimal number $k$ such that the homogeneous component $f_k$ is non-zero. Evidently, for an endomorphism $\varphi=(\varphi(x_1),\;\ldots,\;\varphi(x_N))$ one has $$\operatorname{ht}(\varphi)=\inf\lbrace \operatorname{ht}(\varphi(x_i))\;\|\;1\leq i\leq N\rbrace.$$ The function $$d(\varphi,\psi)=\exp(-\operatorname{ht}(\varphi-\psi))$$ is a metric on the set $\operatorname{End}\mathbb{K}[x_1,\ldots,x_N]$; the corresponding topology will be referred to as the power series topology. We turn to automorphisms and define the tame subgroup. We say that an automorphism $\varphi \in \operatorname{Aut}A_N$ is elementary if it is given either by a linear change of generators $$(x_1,\;\ldots,\;x_N)\mapsto (x_1,\;\ldots,\;x_N)A,\;\;A\in\operatorname{GL}(N,\mathbb{K})$$ or by a transvection – a change of variables of the form $$(x_1,\;\ldots,\;x_N)\mapsto (x_1,\;\ldots,\;x_k+f(x_1,\ldots,x_{k-1},x_{k+1},\ldots,x_N),\;\ldots,\;x_N)$$ (that is, all generators are kept fixed with the exception of $x_k$, to which a polynomial free of $x_k$ is added). The subgroup ${\operatorname{TAut}}A_N$ of tame automorphisms is the subgroup of $\operatorname{Aut}A_N$ generated by elementary automorphisms defined as above. Whenever $N=2n$ is even, the polynomial algebra can be made into $P_n$ by partitioning the set of its generators into two even subsets, $\lbrace x_1,\ldots, x_n\rbrace$ and $\lbrace p_1,\ldots, p_n\rbrace$, and introducing the corresponding Poisson bracket. Under the identification $$\mathbb{K}[x_1,\ldots,x_{n},p_1,\ldots,p_n]\simeq \mathcal{O}(\mathbb{A}^{2n}_{\mathbb{K}})$$ the generators $x_i$, $p_j$ become the Darboux coordinate functions for the standard symplectic form $\omega = \sum_{i} dx_i\wedge dp_i$. The group $\operatorname{Aut}P_n$ is then the subgroup of $\operatorname{Aut}\mathbb{K}[x_1,\ldots,x_{n},p_1,\ldots,p_n]$ of automorphisms which preserve the symplectic (or Poisson) structure. Its intersection with ${\operatorname{TAut}}\mathbb{K}[x_1,\ldots,x_{n},p_1,\ldots,p_n]$ is the subgroup ${\operatorname{TAut}}P_n$ of tame symplectomorphisms. Automorphisms that are not tame are called wild. Wild automorphisms exist – the most well-known example being due to Nagata [@Shes2; @Shes3]: $$\varphi_N:\mathbb{K}[x,y,z]\rightarrow \mathbb{K}[x,y,z],$$ $$\varphi_N:(x,y,z)\mapsto(x+(x^2-yz)x,\;y+2(x^2-yz)x+(x^2-yz)^2z,\;z).$$ In dimension two, all automorphisms are tame – a fact that allows for an explicit description of $\operatorname{Aut}P_1$ (cf. [@Jung; @VdK]) and $\operatorname{Aut}W_1$ and, consequently, positive resolution of Kontsevich conjecture in this case. The latter is due to Makar-Limanov [@ML1; @ML2]. It turns out [@ML3] that the tameness of the planar case is not specific to the commutative polynomial algebra, but rather is a property of a broader class of objects. It is not known whether in even dimensions all symplectomorphisms are tame; that fact, if it were to be established, would pave the way for a quick resolution of Kontsevich conjecture, for the direct homomorphism $$\operatorname{Aut}W_n(\mathbb{C})\rightarrow \operatorname{Aut}P_n(\mathbb{C})$$ restricts to an isomorphism of tame subgroups. We proceed by formulating basic results on approximation by tame automorphisms. An elementary linear automorphism is a symplectomorphism if and only if its matrix $A$ is symplectic, $A\in \operatorname{Sp}(2n, \mathbb{K})$. A transvection defined by a polynomial $f$ is a symplectomorphism if and only if $f$ is free of all generators of the type that has the generator to which $f$ is added. That is, if $f$ is added to $x_k$, then $f$ must be a function of $p_1,\ldots, p_n$ only for it to be a symplectomorphism, and vice versa. Straightforward. We now formulate the basic results of approximation by tame automorphisms. Anick’s theorem states that the subgroup ${\operatorname{TAut}}A_N$ is dense in $\operatorname{Aut}A_N$ in power series topology. The unit Jacobian requirement is not essential (indeed, any automorphism must have a constant Jacobian – an easy exercise; forcing an automorphism to have unit Jacobian amounts then to a rescaling), yet convenient. One may, without loss of generality, develop approximation for automorphisms in the neighborhood of the identity automorphism – that is, automorphisms which are identity modulo terms of certain height. In this framework, the unit Jacobian requirement becomes redundant. Let $\varphi=(\varphi(x_1),\;\ldots,\;\varphi(x_N))$ be an automorphism of the polynomial algebra $A_N=\mathbb{K}[x_1,\ldots,x_N]$ over a field $\mathbb{K}$ of characteristic zero, such that its Jacobian $$\operatorname{J}(\varphi)=\operatorname{det}\left[\frac{\partial \varphi(x_i)}{\partial x_j}\right]$$ is equal to $1$. Then there exists a sequence $\lbrace \psi_k\rbrace\subset {\operatorname{TAut}}\mathbb{K}[x_1,\ldots,x_N]$ of tame automorphisms which converges to $\varphi$ in power series topology. The symplectic version of Anick’s theorem is a recent development [@KGE]. Let $\sigma=(\sigma(x_1),\;\ldots,\;\sigma(x_n),\;\sigma(p_1),\;\ldots,\;\sigma(p_n))$ be a symplectomorphism of $\mathbb{K}[x_1,\ldots,x_n,p_1,\ldots,p_n]$ with unit Jacobian. Then there exists a sequence $\lbrace \tau_k\rbrace\subset {\operatorname{TAut}}P_n(\mathbb{K})$ of tame symplectomorphisms which converges to $\sigma$ in power series topology. The reader is encouraged to browse the proof of this statement in [@KGE] in order to gain a somewhat broader understanding of the context of Kontsevich conjecture and associated situations. In order to utilize the approximation theory, we need to be able to make sense of the lifted limit of a tame sequence $\lbrace \sigma_k\rbrace$. Just as the automorphisms $\psi_k$ lifted from $\sigma_k$ are defined by means of formal power series in $\hbar$ in the framework of deformation quantization (see below), so will be the lifted limit $\Psi$. However, while the tame automorphisms $\psi_k$ will have entries polynomial in $\hbar$ (which is an immediate consequence of the quantization formula) and also the coefficients at each $\hbar^n$ will be polynomial in the generators of the Poisson algebra $P_n$, it will generally not be the case for arbitrary $\sigma$. In order for the lifted limit to be well defined, one therefore needs statements on convergence of the appropriate power series: the power series in $\hbar$ with respect to the $\hbar$-adic topology as well as the power series which determine the coefficients (in the $\mathfrak{m}$-adic topology obtained from the scheme structure on $\operatorname{Aut}P_n$). The needed statements translate into the following theorems [@KGE]: Let $\sigma$ be a symplectomorphism and let $\mathcal{O}_{\sigma}$ be the local ring of $\operatorname{Aut}P_n(\mathbb{C})$ with its maximal ideal $\mathfrak{m}$. Then there exists a sequence of tame symplectomorphisms $\lbrace \sigma_k\rbrace$ which converges to $\sigma$ in power series topology, such that the coordinates of $\sigma_k$ converge to coordinates of $\sigma$ in $\mathfrak{m}$-adic topology. Suppose given a symplectomorphism $\sigma$ and $\lbrace \sigma_k\rbrace$ is a tame sequence converging to it. Then the $\hbar$-series which define the lifted tame automorphisms $\psi_k$ converge to the power series that define the limit $\Psi$ in the $\hbar$-adic topology. Lifting of polynomial symplectomorphisms ======================================== Given an arbitrary symplectomorphism $\sigma \in\operatorname{Aut}P_n(\mathbb{C})$ and a sequence of tame symplectomorphisms $\lbrace \sigma_k\rbrace$ converging to it, we can construct a sequence $\lbrace \psi_k\rbrace$ of Weyl algebra automorphisms in the following way. Let $\psi_k$ be the pre-image of $\sigma_k$ under the direct homomorphism $\operatorname{Aut}W_n\rightarrow \operatorname{Aut}P_n$ described in the first section. It is an isomorphism of the tame subgroups, therefore the assignment is well defined and unique. Alternatively, we could start with the Poisson algebra $P_n=\mathbb{C}[x_1,\ldots,x_n,p_1,\ldots, p_n]$ and perform the deformation quantization [@Kon; @Kell] according to a formal parameter $\hbar$ and the associative star product $\star$. It is straightforward to deform elementary symplectomorphisms, and the procedure yields, under appropriate identifications, the same result as the one involving the direct homomorphism. In either case, we refer to thus described procedure as the lifting of polynomial symplectomorphisms. It seems slightly more convenient to work with the deformed family $W_n(\hbar)$ of Weyl algebras depending upon $\hbar$ rather than with a single Weyl algebra. The most important thing to bear in mind, however, is the fact that a given symplectomorphism $\sigma$ will specify (by imposition of associativity and Weyl algebra commutation relations on images $\sigma(x_i)$, $\sigma(p_j)$) a new star product $\star_{\sigma}$, which differs from the original one by a gauge transformation and defines a new family of associative algebras $W_n(\hbar, \sigma)$. The main theorem then admits a reformulation in the following way: There is an algebra embedding $$W_n(\hbar, \sigma)\hookrightarrow W_n(\hbar).$$ This is equivalent to the Kontsevich conjecture and can therefore be perceived as the principal subject of this study. It is worth mentioning that Myung and Oh [@MyOh] have recently conducted a study of deformations of Poisson algebras with the purpose similar to that of the present paper. It can be inferred from their results that the larger algebras $S_n(\hbar)$ and $S_n(\hbar, \sigma)$ of formal power series (while $W_n(\hbar)$ consists of expressions polynomial in $\hbar$) are isomorphic to each other. The statement of the Kontsevich conjecture (provided by the Main theorem above) is stronger. Whenever $\sigma$ is tame, the statement of the theorem is straightforward. Any tame symplectomorphism $\sigma$ lifts to an object of the following form $$(\Psi_1(x_1,\ldots,x_n,p_1,\ldots,p_n,\hbar),\ldots,\Psi_{2n}(x_1,\ldots,x_n,p_1,\ldots,p_n,\hbar)).$$ Here $\Psi_l(x_1,\ldots,x_n,p_1,\ldots,p_n,\hbar)$ are power series in $\hbar$ whose coefficients are polynomials in (commuting) variables $x_i$ and $p_j$. They are obtained by applying, say, the Kontsevich quantization formula [@Kon] (rather, a special case corresponding to the $2n$-dimensional affine space) to the polynomials $\sigma(x_1),\;\ldots,\;\sigma(p_n)$. It is clear from the fact that the coefficients at $\hbar^n$ are the images under certain bidifferential operators that the power series $\Psi_l$ are really polynomials in $\hbar$, whose degree (in $\hbar$) depends on the total degree of $\sigma$. The case of general $\sigma$ cannot be processed in this way. Indeed, if $\lbrace\sigma_k\rbrace$ is a sequence of tame symplectomorphisms converging to $\sigma$, one can take the lifted automorphisms $\psi_k$ and define the lifted limit $\psi$. The lifting procedure based on the initial star product, however, when applied to the sequence $\sigma_k$, will in the limit return an object defined by power series in $\hbar$ (rather than polynomials). Moreover, the coefficients at $\hbar^n$ will also in general be power series in commuting variables $x_i$ and $p_j$ (although that particular problem can be dealt with, cf. lemma in the next section). At this point (in accordance with the remark at the end of the previous chapter), it may be somewhat comforting to note that these coefficients in $x_i$ and $p_j$ will always be given by power series with sufficiently good behavior. In fact, they will be convergent with respect to the $\mathfrak{m}$-adic topology of the local ring $\mathcal{O}_{\sigma}$ of the scheme $\operatorname{Aut}P_n$ at the point $\sigma$. More precisely, we have the following theorem [@KGE]: Let $\sigma$ be a symplectomorphism and let $\mathcal{O}_{\sigma}$ be the local ring of $\operatorname{Aut}P_n(\mathbb{C})$ with its maximal ideal $\mathfrak{m}$. Then there exists a sequence of tame symplectomorphisms $\lbrace \sigma_k\rbrace$ which converges to $\sigma$ in power series topology, such that the coordinates of $\sigma_k$ converge to coordinates of $\sigma$ in $\mathfrak{m}$-adic topology. This result says that the power series which constitute the coefficients at $\hbar^n$ of the lifted limit are well defined (thus making the lifted limit well defined also). To prove the main theorem, one has to gauge the lifting in a certain way in order to cut off the infinite series and, consequently, obtain an embedding of algebras. In the next section we argue that such gauging always exists. Lifted limit as a Weyl algebra automorphism =========================================== Let $\sigma$ be an arbitrary polynomial symplectomorphism as before, and let $\psi$ be the lifted limit of a tame symplectomorphism sequence $\lbrace\sigma_k\rbrace$. In this section we actively use the embedding of the Weyl algebra over $\mathbb{C}$ into the reduced direct product of Weyl algebras over algebraically closed fields of positive characteristic $p$, which runs over all prime numbers. To avoid the conflict of notation, we denote the generators of $P_n$ and $W_n$ by letters $x$ and $y$ rather than (more classical) $x$ and $p$. We first observe the following The power series $\Psi_l(x_1,\ldots,p_n,\hbar)$ which make up the lifted limit $\psi$ correspond to rational functions in $\hbar$. Namely, for each positive characteristic $p$ in the ultraproduct decomposition, the central elements $\Psi_l^p$ are rational functions in $\hbar$. Indeed, for a given positive characteristic $p$ marking a component in the ultraproduct, the correspoding Weyl algebra is Azumaya. This in particular means that it is isomorphic to the algebra of $p\times p$ matrices over its center. The two matrix algebras – the initial Weyl algebra and the one which results from $\sigma$ – are then isomorphic to each other, which is equivalent to the fact the the centers $C$ and $C_{\sigma}$ of the corresponding algebras are isomorphic[^7]. Therefore, $\Psi_l^p$ are rational functions. We turn now to the proof of the main theorem. Our objective is to show that our lifting can be appropriately modified so that the resulting object will be given by polynomials. Working for each $p$ in the ultraproduct decomposition, we look for a gauge of the lifting that will leave the $\hbar$-independent part of the center, given by $\mathbb{F}_p[x_1^p,\ldots,y_n^p]$, unperturbed. We have the following The lifting can be gauged in such a way that stabilizes the $\hbar$-independent center $\mathbb{F}_p[x_1^p,\ldots,y_n^p]$. Another important lemma is the following. There is a transformation of the lifting that results in the lifted automorphism $\Psi$ being defined by power series in $\hbar$ which have coefficients polynomial in $x_i$ and $y_j$. The proof of this statement will be addressed in our further work. We next observe that the gauging must obviously be performed by means of an Azumaya algebra automorphism. In terms of matrix algebra, that means that (for fixed $p$) the desired transform must be a matrix conjugation. For $l=1,\ldots, 2n$, we therefore consider the expressions of the form $$(1+\hbar Q_l)\Psi_l(1+\hbar Q_l)^{-1},$$ where $Q_l$ is a rational function in $\hbar$. Note that $Q_l$ is, generally, not an element of the Weyl algebra over $\mathbb{C}$ (in the standard sense), but rather a class (modulo ultrafilter $\mathcal{U}$ that realizes the ultraproduct decomposition) of elements $Q_{l,p}$ which are rational in $\hbar$. Any such $Q_l$ leads to a gauging. We use the following result. The polynomials $(\Psi_l)_1(x_1,\ldots, y_n)$, $l=1,\ldots, 2n$ which give the coefficients of $\hbar$-power series $\Psi_l$ at $\hbar$ are of total degree less than $\operatorname{deg}\sigma$. This is not immediately obvious in light of remarks on the modified star product $\star_{\sigma}$ (although it is apparent for lifted tame symplectomorphisms). This statement provides the base case for the processing of the higher-order terms. The gauging acts upon the term at $\hbar^2$. Indeed, the leading term of the difference between the gauged and ungauged expression is given by the commutator (in the initial star product) $$[\Psi_l, Q_l]\hbar,$$ which is of order $\hbar^2$ by definition of $\Psi_l$ and $Q_l$. For $\sigma$ that are close enough to the identity symplectomorphism, this translates into correction terms of the form $$\frac{\partial Q_l}{\partial x_i}\;\;\text{and}\;\;\frac{\partial Q_l}{\partial y_j}.$$ This can always be fulfilled. Now, in order to tweak the higher-degree (in $\hbar$) terms, one applies consecutive conjugations according to the method described above. For the terms of degree higher than $\operatorname{deg}\sigma$, the existence of compensating terms amounts to the vanishing of the appropriate differential form. We illustrate the algorithm by applying it to the base case $n=1$, so that the coefficients that need to be processed are power series in two generators $x$ and $y$. Firstly, we can always find a conjugation such that the resulting lifted limit will send $x$ to itself. Once this is done, the power series $$1+\hbar Q$$ can only depend on $x$ ($Q$ corresponds to $y$). Conjugating the image of $y$ under $\Psi$ by $1+\hbar Q$, we can dispose of the terms that do not contain $y$. Indeed, the leading term of conjugation with $y$ is given by differentiation, so the amending term is constructed by taking a primitive of the appropriate polynomial. On the other hand, because of the fact that the commutator $[y,x]=\hbar$ produces a power of $\hbar$ and shifts the terms one notch, no extra terms in the fixed term $\hbar^k$ (with which we are currently working) can appear. This completes the proof. The algorithm can be easily modified for the case of arbitrary $n$. Acknowledgments =============== The work is supported by the Russian Science Foundation grant No. 17-11-01377. [0]{} , [Automorphisms of Weyl algebras]{}, [ Letters in mathematical physics]{} [74]{} (2005), 181-199. , [The Jacobian Conjecture is stably equivalent to the Dixmier Conjecture]{}, arXiv: math/0512171v2, 2005. , [Independence of the B-KK Isomorphism of Infinite Prime]{}, arXiv: 1512.06533. , [Limits of tame automorphisms of $\mathbb{K}[x_1,\ldots,x_N]$]{}, [ J. Algebra]{}, [82]{}, 459-468 (1983). , [Lifting of Polynomial Symplectomorphisms and Deformation Quantization]{}, arXiv:1707.06450, to appear in Commun. in Algebra. , [Deformation quantization of Poisson manifolds, I]{}, Lett.Math.Phys. 66: 157-216, 2003, arXiv: q-alg/9709040. , [Sur les algebres de Weyl]{}, Bull. Soc. Math. France [96]{} (1968), 209-242. , [ Mem. Fac. Sci, Kochi Univ. Ser. A Math.]{} [24]{} (2003), 43-59. , [Endomorphisms of Weyl algebra and $p$-curvatures]{}, [ Osaka Journal of Mathematics]{}, vol. [42]{} no. [2]{} (2005). , [Reduced Direct Products]{}, Fund. Math. No. 51 (1962), 195-228. , [Uber ganze birationale Transformationen der Eben]{}, [ J. Reine angew. Math.]{}, [184]{} (1942), 161-174. , [On polynomial rings in two variables]{}, [ Nieuw Arch. Wisk.]{}, [1]{} (1953), 33-41. , [On automorphisms of Weyl algebra]{}, [ Bull. S. M. F.]{}, tome [112]{} (1984), 359-363. , [Automorphisms of a free algebra with two generators]{}, [Funkts. Anal. Prilozh.]{}, Vol. [4]{}, Issue 3 (1970), 107-108. , [Automorphisms and derivations of free Poisson algebras in two variables]{}, [ J. Algebra]{} [322]{}, No. [9]{} (2009), 3318-3330. , [The Nagata automorphism is wild]{}, Proc. Nat. Acad. Sci. USA, 100 (2003), No. 22, 12561 – 12563. , [Tame and wild automorphisms of rings of polynomials in three variables]{}, J. Amer. Math. Soc., 17 (2004), 197 – 227. , [Notes for an Introduction to Kontsevich’s quantization theorem]{}, <https://webusers.imj-prg.fr/~bernhard.keller/publ/emalca.pdf>. , [Automorphism groups of Weyl algebras]{}, arXiv: 1710.00432. [^1]: [email protected] [^2]: [email protected] [^3]: [email protected] [^4]: One makes a straightforward observation that the Kontsevich conjecture, along with objects accompanying it, are statements which can be formulated by means of first-order logic; therefore, in the case of algebraically closed ground field $\mathbb{K}$ of characteristic zero, one may well work with complex numbers, $\mathbb{K}=\mathbb{C}$, in accordance with the Lefschetz principle. [^5]: As well as minimal – recall that the product of fields is always von Neumann regular. [^6]: We set $\operatorname{deg}x_i=1$. [^7]: The fact that Morita equivalence of commutative rings implies their isomorphism is well known.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Single white dwarfs with photospheric metal absorption and no dusty disks are believed to be actively accreting metals from a circumstellar disk of gas caused by the disruption and sputtering of asteroids perturbed by post-main sequence planetary systems. We report, for the first time, the detection of circumstellar Ca II gas in absorption around the metal-polluted white dwarf WD 1124-293 at $>$7 $R_{\rm WD}$ and $<$32000 AU, with a probable distance of $\sim$54 R$_{\rm WD}$. This detection is based on several epochs of high-resolution optical spectroscopy around the Ca II H and K lines ($\lambda$=3968Å, 3933Å) with the MIKE spectrograph on the Magellan/Clay Telescope. We confirm the circumstellar nature of the gas by observing stars with small angular separations to WD 1124-293 and larger distances from Earth–none of the reference stars show absorption consistent with the presence of appreciable local interstellar medium at the distance of the white dwarf. By combining our observations over four years with previous spectra of WD 1124-293, we have measured the equivalent width of the two photospheric Ca lines over a period of 11 years. We see $<$ 5-7% epoch-to-epoch variation in equivalent widths over this time period, and no evidence for long term trends in the strength of the Ca II lines. Since it is likely that the circumstellar gas arises from a disk edge-on to our line of sight, we also place limits to short period transiting substellar and planetary companions with R $>$ R$_{\rm \oplus}$ using the WASP survey. The presence of gas in orbit around WD 1124-293 implies that most metal rich WDs could harbor relic planetary systems. Since roughly 25-30% of WDs show metal line absorption, the dynamical process for delivering small bodies close to a WD must be robust.' author: - 'J. H. Debes,M. Kilic, F. Faedi, E. L. Shkolnik,M. Lopez-Morales, A. J. Weinberger, C. Slesnick, R. G. West' bibliography: - 'wd\_chap.bib' - 'Faedi.bib' title: 'Detection of Weak Circumstellar Gas around the DAZ White Dwarf WD 1124-293: Evidence for the Accretion of Multiple Asteroids' --- Introduction ============ In the absence of externally accreted material, radiative levitation, or convective dredge-up, white dwarfs (WDs) should have pure hydrogen or pure helium photospheres. All other elements heavier than helium should settle out of view on timescales much shorter than the cooling age of WDs. Hydrogen WDs (type DA) have photospheric settling times that can range from days to thousands of years, while helium WDs (type DB) have settling times as long as 1 Myr. Due to their high gravity and thin photospheres, WDs are some of the most sensitive probes to low levels of external accretion. A typical 0.6[M$_\odot$]{} hydrogen WD with an effective temperature of 10000 K can show observable Ca II lines from the accretion of $\sim$10$^{6}$ g/s of solar abundance material, corresponding to the accretion of a small asteroid’s worth of material per year [@paquette86; @koester05; @koester09]. WDs are sensitive probes of the accretion from the surrounding interstellar medium (ISM) [@dupuis92; @dupuis93; @dupuis93b], the winds from companions [@sion84; @zuckerman03; @debes06], and the accretion of asteroidal material from orbiting dust disks [e.g. @zuckerman07; @klein10; @vennes10; @dufour10; @debes11a; @debes11b]. All three of these phenomena can explain the presence of metal lines in WDs, but a significant fraction of metal polluted WDs (type DAZ for hydrogen atmosphere WDs and type DBZ or DZ for helium atmosphere WDs) are apparently single, do not possess any dust disks, and do not have sufficient surrounding ISM to explain their accretion [@aanestad93; @kilic07]. ISM accretion was the first logical explanation for metal polluted WDs [@wesemael78; @wesemael82; @dupuis92; @dupuis93; @dupuis93b]. However, studies of polluted WDs show no strong correlation between velocity and metal abundance as would be expected for Bondi-Hoyle accretion of the ISM [@aanestad93; @zuckerman03; @zuckerman10; @kilic07]. Nor is there any correlation between metal abundance and constraints on the local ISM [@kilic07]. Most observed DAZs and DZs reside well within the low density Local Bubble, and are not physically located near known dense molecular clouds. The lack of appreciable hydrogen in DZs also cannot easily be explained by the accretion of solar abundance material. The presence of polluted DAZs that are members of WD+red dwarf binaries raises the possibility that previously unseen companions could be the cause of many polluted WDs through the accretion of a companion wind [@zuckerman03]. While it is physically plausible that a close ($a <$1 AU) companion can pollute a WD with its wind [@debes06], most WDs that have been observed with [Spitzer]{} are constrained to have no close stellar or brown dwarf companions [@mulally08; @debes07; @farihi09]. The most massive objects that could orbit these WDs would have masses of 5-15 M$_{\rm Jup}$. Dusty WDs all show pollution due to the direct accretion of metal rich grains onto the WD surface. However, only $\sim$30% of metal polluted WDs show infrared excesses due to dusty disks [@kilic08; @farihi10]. Some disks may be close to edge-on to our line of sight and hidden, but inclination effects alone cannot explain the relative number of dusty and non-dusty WDs. Tidally disrupted planetesimals perturbed by planets are believed to be the origin of dusty WDs [@debes02; @jura03; @bonsor11]. Similarly, apparently single WDs with metal pollution may be caused by hitherto unseen disks that are either dusty, optically thin, and undetectable in the infrared, or entirely gaseous. Gaseous disks with some mass in gas of $M_{\rm disk}$ that arise from the sputtering of debris from tidally disrupted planetesimals have been postulated to explain the relative number of dusty and non-dusty WDs [@jura08]. In this scenario, only planetesimals larger than $\sim$2 M$_{\rm disk}$ create an optically thick dusty disk, while smaller objects are completely ablated by the relative velocity between the debris and the existing gaseous disk, further populating the disk with gaseous material. Circumstellar gas emission lines have been observed toward several hotter WDs that also possess dust disks [@gaensicke06; @gaensicke07; @gaensicke08; @melis10], demonstrating that significant gas can be generated from either the sublimation, collision, or photoionization of dust grains. Similarly, a handful of hotter WDs that have been observed in the UV show circumstellar gas absorption [@holberg98; @bannister03; @lallement11; @dickinson12]. These hotter WDs may be evaporating circumstellar dust disks or local ISM, but to date none of these systems has been shown to possess a dust disk–their origins are less clear. As WDs become hotter, radiative levitation of metals becomes a more important process and can explain the presence of metals in a WD atmosphere for effective temperatures larger than $\sim$25000 K [@chayer10]. In this work we present the detection at 8-$\sigma$ of weak circumstellar Ca II absorption in orbit around the cooler DAZ WD 1124-293. Based on observations of other sight lines close to WD 1124-293 and the relative velocity of the circumstellar gas, we confine the spatial extent of the absorption to 7 $R_{\rm WD} < r <$ 32000 AU from WD 1124-293. We argue that it must arise from a circumstellar gaseous disk in orbit around the WD at a probable radius of $\sim$54 $R_{\rm WD}$, well within the WD’s tidal disruption radius. In §\[sec:obs\] we describe the MIKE spectroscopic observations of WD 1124-293 and present the results of those observations in §\[sec:res\]. We also search WASP data for planetary transits around WD 1124-293 in §\[s:trans\]. We search for variability in the accretion of material on the WD in §\[sec:var\] and present our conclusions in §\[sec:conc\]. Observations {#sec:obs} ============ We have taken optical spectra of WD 1124-293 over 16 separate days spanning 4 years with the MIKE spectrograph [@bernstein03] on the Clay Telescope at Las Campanas Observatory between 2007 and 2011 (See Table \[tab:obs\]) as part of a large DAZ monitoring project [@debmuk10]. WD 1124-293 was first discovered in the Edinburgh-Cape Survey and listed in @mccook99. Based on spectral models of WDs, WD 1124-293 has M=0.66 [M$_\odot$]{} and T$_{\rm eff}$=9400 K [@koester06], with a spectroscopically determined distance of 33.6 pc [@pauli06]. Photospheric metal lines of Ca II were detected with moderate resolution spectroscopy from Keck HIRES (R$\sim$30000) and VLT UVES (R$\sim$18500) [@zuckerman03; @koester05], which showed moderately strong Ca II H and K lines ($\lambda$=3968.469,3933.664) with a heliocentrically corrected redshift of 29.5 km s$^{-1}$ (29$\pm$1 km s$^{-1}$) for Keck (VLT) in December 1998 and April 1999 (April and May 2000). The calcium cannot be explained by accretion of the local ISM based on measurements of the local ISM density [@kilic07]. The equivalent width (EW) of the Ca II line over the four separate epochs with VLT and Keck averaged $\approx$106 mÅ with no significant variation over this period, implying a steady state accretion rate of 4$\times10^{15}$ g/yr [@koester06], assuming 1% metal rich material relative to solar abundance. In January and June of 2010 we observed four bright F and A stars located near WD 1124-293 on the sky in order to probe the surrounding ISM at the distance of WD 1124-293. For our MIKE spectra, we used a 07$\times$5 slit which corresponded to a resolution of $\sim$40,000 at the Ca II K line. Th-Ar comparison lamp spectra were taken near in time to each spectrum of WD 1124-293. In addition, we obtained the previously reduced spectra of this WD from the VLT and Keck [@koester05; @zuckerman03]. The VLT spectra were kindly provided by D. Koester, while the Keck data was publicly available with the journal article it was published in. Our MIKE data were flatfielded and bias-subtracted using the MIKE reduction pipeline written by D. Kelson, with methodology described in @kelson00 and @kelson03. We initially extracted the spectra using the full pipeline, but discovered that the pipeline was subtracting the background at a level of precision insufficient for high quality equivalent width measurements. To rectify this, we performed our own extraction, background subtraction, and wavelength solution for 8-10Å windows centered on the Ca H and K lines for all of our MIKE spectra. The spectra were also corrected for heliocentric velocity shifts between observations. Since the Ca II K line falls on two orders, we combined each order by fitting the continuum with a polynomial and averaging the orders for each night to increase signal-to-noise (S/N) to the values reported in Table \[tab:obs\]. All the fitted spectra from MIKE were then averaged with a S/N weighting to produce a final spectrum of the region around each line. In order to check for variability in our spectra between 2007-2011 as well as with the previous observations, we calculated the EW of the two calcium lines from each night’s spectrum. To calculate the EW, we chose a window around the line equivalent to $\pm$3 times the standard deviation of the line as determined by a Gaussian fit, which also returned a measure of the radial velocity for the WD on each night. We chose several different polynomial fits to the continuum and added any systematic uncertainty in the continuum fit to our uncertainties in the EW calculation. This approach has the advantage of calculating a more accurate EW despite differences in resolution between different epochs, since the total integrated line flux is conserved from observation to observation. Table \[tab:meas\] details the observed photospheric line radial velocities and equivalent widths. From the individual measures of the K line with MIKE, we obtain a median radial velocity for WD 1124-293 of 29.0$\pm$0.5 km s$^{-1}$. Results {#sec:res} ======= The final, S/N weighted spectrum of WD 1124-293 at the Ca H and K lines is shown in Figure \[fig:f1\]. Just to the blue of the main photospheric lines we detect weak lines indicative of another Ca II component, with line depths of 3.4- and 8.5-$\sigma$ below the continuum for the H and K line respectively. As seen in Table \[tab:meas\], the lines at both wavelengths are separated by $\approx$30 km s$^{-1}$ to the blue of the photospheric Ca II. Weak lines in the spectrum could be due to additional photospheric species, circumstellar Ca II gas, intervening ISM Ca II gas, or an artifact from the spectrograph or reduction process. In this section we argue that the lines arise from circumstellar Ca II gas in orbit around the WD. Ruling out artifacts and other atomic species --------------------------------------------- There are three separate conditions that rule out the possibility that the lines could arise from the data reduction process or the MIKE spectrograph. First, we have verified that the line near the photospheric Ca II K line was detected significantly on a S/N combined spectrum of each of the two echelle orders that the line fell on. The position of the line on the two orders is separated by $\approx$1100 pixels, ruling out possible artifacts from the instrument. For both orders, the line is detected at the proper position and at 4- and 6-$\sigma$ below the continuum. Secondly, the line is detected significantly in the final combined spectra when using the full Kelson MIKE pipeline, which used different methodology to perform order extraction and background subtraction. Finally, when we combine the data into subsets, we recover the line as well. Figure \[fig:f2\] shows three separate epochs: March 2008, May 2008, and a combination of our June 2010 and June 2011 data. In particular, the heliocentric velocity correction in March 2008 and our June 2010+2011 data differed by 23 km s$^{-1}$, ruling out any terrestrial origin for the line, since the epoch-to-epoch centers match to within 0.01Å. These lines could be weak photospheric features. We searched atomic line lists, as well as metal lines other than calcium reported for other polluted WDs [e.g. @klein10], and found no atomic lines that could plausibly be matched to the observed spectrum of WD 1124-293. Furthermore, the line depths and relative equivalent widths are consistent, within the uncertainties, to the expected line strength ratio between the Ca II H and K lines. Based on these arguments, the weak features must arise from intervening ISM or circumstellar Ca II gas. Ruling out Local Interstellar Absorption ---------------------------------------- While ISM absorption could explain the presence of the weak features we observe, we rule out an ISM contribution. The Sun resides in a low-density ISM structure known as the Local Bubble. WD 1124-293 is at a spectroscopically determined distance of 33.6 pc, well inside the Local Bubble [@pauli06]. The local interstellar medium is relatively devoid of cold neutral gas, until the accumulation of dense material at $\sim$80 pc [@welsh10]. However, warm, partially ionized clouds have been detected through weak ISM absorption features in the spectra of stars within 50 pc of the Sun [see @redfield08; @welsh10 for a review of the properties of the local interstellar medium]. If the weak Ca II absorption feature in WD 1124-293 is due to ISM accretion, other stars in close proximity on the sky should also show a similar absorption feature. In order to test this possibility, we observed several bright stars at angular separations $<1.6\arcdeg$ from WD 1124-293 to search for absorption features at the same velocity. This technique, as well as searching for line variability, has been used in the past to confirm circumstellar absorption around HD 32297, a young star with an edge-on debris disk [@redfield07]. We observed four stars with constraints on their distance from Hipparcos and with the same setup as WD 1124-293: HIP 55864 (F5V, r=0.27$\arcdeg$) HIP 55901 (A0V, r=0.71$\arcdeg$), HIP 55731 (A9V, r=1.10$\arcdeg$), and HIP 55968 (A3V, r=1.58$\arcdeg$) (See Table \[tab:obs\] ). Another star, HIP 56280A (F8V, r=1.19$\arcdeg$), had a spectrum that was kindly provided to us by S. Desidera. Figure \[fig:f3\] shows how each ISM standard is arranged on the sky relative to WD 1124-293. The distances for each star are given in the latest reduction of Hipparcos data [@leeuwan07]. HIP 56280A and HIP 55864 have secure distances (parallax measurements with errors of $<$20%), while HIP 55731 and HIP 55901 have marginally detected parallaxes ($\sim$3-$\sigma$). HIP 55968’s parallax is consistent with zero. For the marginal detections, we take as the stellar distance the 3-$\sigma$ upper limit to the parallax, which is 6.62 mas (151 pc) and 5.67 mas (176 pc) for HIP 55731 and HIP 55901 respectively. HIP 55968’s optical and NIR magnitudes and spectral type are consistent with this star lying between HIP 55731 and HIP 55901, so we take the lower limit to its distance to be 151 pc. Each star had broad Ca II K photospheric absorption which we removed by applying a polynomial fit to the photospheric line, an example of which is shown in Figure \[fig:f4\]. For HIP 56280A, its narrow line core was fitted with two polynomials joined at the line core center of 3933.704Å. The resulting continuum fitted spectra were inspected for narrow absorption components. HIP 56280A (to lower significance) and HIP 55864 show no appreciable absorption features comparable in strength to the WD 1124-293 line, while HIP 55731 shows a weak (presumably ISM) component far from the radial velocity or our observed feature. HIP 55901 and HIP 55968, both constrained by Hipparcos to be more distant than HIP 55864, each showed absorption from at least three ISM components. We attempted to simultaneously fit these three components for the two stars and present our best fits in Figure \[fig:f5\], which shows WD 1124-293’s Ca II K line region compared with our ISM standards. Table \[tab:ism\] shows the resulting velocity components for these fits, consistent with at least three distinct complexes of ISM gas that slowly vary on size scales of $\sim$1-2 pc, one of which is at a radial velocity within a few km s$^{-1}$ of the circumstellar feature. Despite the similar radial velocity of the two features to what is observed in WD 1124-293, the non-detection of any similar lines for HIP 56280, HIP 55864, and HIP 55731 constrains the location of these ISM components to be at a distance $>$151 pc. The non-detection of the faint candidate circumstellar component for HIP 55864 constrains the extent of the Ca II gas near WD 1124-293 to $<0.27\arcdeg$ at d=33.6 pc, or roughly 32000 AU (0.16 pc). @welsh10 find a slowly increasing equivalent width with increasing distance, but depending on whether local gas clouds are encountered or not, they find both low and high values of volume density within 30 pc. Out of the 1857 stars analyzed by @welsh10, 50 are within 40 pc of the Sun. Only one of these 50 stars, HD 159561 ($l=36\arcdeg, b=+23\arcdeg$), shows an ISM absorption feature with a equivalent width stronger than what is observed for the weak feature in WD1124-293 ($l=282\arcdeg, b=+30\arcdeg$). The ISM complexes observed for our more distant stars are consistent with a relatively dense cloud of material that lies $>$90 pc along the Galactic sightline to these stars [@welsh10], and thus cannot account for the weak Ca II feature we measure for WD 1124-293. Hence, it is unlikely that this feature is due to an intervening warm ISM cloudlet. The only scenario that the data allows is if an extremely narrow warm ISM column near WD 1124-293 (width $\sim$0.06 pc and height $\sim$1.3 pc) with a velocity gradient intersected the sightlines to WD 1124-293, HIP 55901, and HIP 55968, but did not intersect the sightline of HIP 55864. Such a pathological structure is not supported by studies of the nearby ISM. @redfield01 studied the small scale (0.05-1.2 pc) structure of the local interstellar cloud and find that the column densities do not vary by more than a factor of two for $\le0.6$ pc scales. The implied Ca II column densities by the much stronger ISM absorption observed towards HIP 55901 and HIP 55968 are an order of magnitude larger than for WD 1124-293, and would have been easily detected in the ISM standards that showed no component at the velocity of our candidate circumstellar feature. Hence, the observations of our ISM standards limits the amount of interstellar Ca II between WD 1124-293 and us at the velocity of the Ca II K feature to EW $<$ 6 mÅ. The source of metals in high surface gravity WDs has been a puzzle for decades [see @dupuis93]. The search for a correlation between the local ISM clouds and the locations of metal-rich WDs was inconclusive [@zuckerman03]. @kilic07 used the observed ISM column densities toward stars in close proximity to the known DAZ WDs to demonstrate that there is no correlation between the accretion density required to supply metals observed in DAZs with the densities observed in their interstellar environment, including WD 1124-293. Therefore, the DAZ WD population as a whole argues against ISM accretion being the dominant contributor to the metals in WDs. In conclusion, it is more likely that the weak Ca II feature is due to circumstellar gas in orbit around WD1124-293 because (1) no similar absorption features are detected in neighboring sightlines as close as 0.16 pc and the local interstellar medium is relatively homogeneous at such scales [@redfield01], (2) the strength of the absorption feature is uncommon for stars within 40-50 pc of the Sun, (3) the absorption feature matches the stellar radial velocity, and (4) there is no correlation between the metal abundances of known DAZ stars and the ISM densities [@kilic07]. Limits to the inner extent of circumstellar gas ----------------------------------------------- If the line arises from gas in orbit around the WD, the relative velocity between the photospheric line and the circumstellar feature can be used to determine how deep the gas lies in the gravitational well of the WD. The photospheric Ca II line is offset from the true systemic velocity of the WD due to the gravitational redshift of WD 1124-293. Using the latest values of WD 1124-293’s $\log$ g=8.096 and its T$_{\rm eff}$=9420$\pm$150 K (D. Koester, private communication), we have calculated masses from cooling models of WDs [@bergeron95; @holberg06; @tremblay11; @bergeron11] which use the mass-radius relation of @fontaine01 to determine the gravitational redshift of WD 1124-293. The uncertainty in the gravity of the WD dominates the uncertainty in the expected gravitational redshift, but is also not well known. To estimate this possible uncertainty we compare our value to that given in @koester01 and @koester09, which vary between 8.04 to 8.099. We therefore estimate the systematic uncertainty in the WD gravity to be $\approx$0.06. Using this value and uncertainty and calculating a gravitational redshift ($v_{\rm grav})$, we obtain $v_{\rm grav}$=34.9$\pm3.7$ km s$^{-1}$, implying a systemic velocity $\gamma$=-5.9$\pm3.7$ km s$^{-1}$ Previous calculations of WD 1124-293’s $\gamma$ (based on its inferred mass and radius from synthetic spectral modeling) find -2.7$\pm$4.3 km s$^{-1}$, consistent with our calculated $\gamma$ [@pauli03; @pauli06]. The primary difference between our result and previous calculations of $\gamma$ come from a higher WD gravity (The earlier values of $\log~g$ being 0.05 dex lower), highlighting the dependence of this calculation on a detailed knowledge of a WD’s properties. The measured relative velocity between the photospheric line and the circumstellar line ($v_{\rm gas}$) from our S/N weighted spectrum is -29.9$\pm$0.8 km s$^{-1}$. The orbital radius of the gas is then $R_g$=$v_{\rm grav}/(v_{\rm gas}+v_{\rm grav})$ R$_{\rm WD}$, implying a minimum radius to the gas of 7$^{+11}_{-3}~R_{\rm WD}$. Given the uncertainties, the faint line could correspond to the systemic velocity and reside further from the WD, thus directly probing the gravitational redshift of the WD. This minimum radius then should be treated as a lower limit to the gas radius. For optically thin silicate dust, the sublimation radius is given by [@jura08; @kimura02; @debesbook11]: $$R_{\rm sub}=3.7\times10^{10} \left(\frac{L_{\rm WD}}{0.001~L_\odot}\right)^{1/2}\left(\frac{1000~K}{T_{\rm sub}}\right)^2~\rm cm.$$ From the cooling models used to determine WD 1124-293’s gravitational redshift, we also determine a luminosity of almost exactly 0.001 $L_\odot$, leading to a sublimation radius of $\approx$44 $R_{\rm WD}$, which is close to the inferred radius of the gas from kinematic arguments in §\[sec:disc\]. Properties and Origin of the Circumstellar Gas {#sec:disc} ---------------------------------------------- Further characterization of the Ca II lines permits additional constraints on the origin and properties of the circumstellar gas around WD 1124-293. In particular, the different constraints point to a localized disk of gas a few tens of WD radii from WD 1124-293 and well within its tidal disruption radius. In order to determine the column density of the Ca II gas, we minimized a $\chi^2$ metric for an optically thin gas column by varying the column density ($\log~N$) and FWHM of the line, using the proper oscillator strengths and assuming a Gaussian line profile. From the minimization we find a best fitting line model with (98% confidence) $\log N$=11$^{+0.1}_{-0.2}~cm^{-2}$ and FWHM=0.2$\pm$0.1 Å with a $\chi_\nu^2$=1.08. The best fit line, compared to the observed circumstellar line is shown in Figure \[fig:f6\]. The model overpredicts the expected line strength of the Ca II H line, but given the low S/N of the line it is hard to determine if this difference is significant. Since this is the ionized state of calcium, it is impossible to tell how much total gas resides in the system without a) other gas species or b) a measurement of Ca I to determine the temperature and electron density of the gas. Further characterization of the gas will be the focus of future work, but is beyond the scope of this paper. We can also constrain the location of the gas by the width of the weak Ca II line–an upper limit to the orbital velocity of the gas can be determined from the line FWHM, and thus a characteristic radius for the disk. We assume that the gas is in a circular orbit and that the line FWHM corresponds to the range of orbital radial velocities within the circumstellar gaseous disk that intersects the angular disk of the WD on the sky: $$r_{\rm disk} \approx \left(GM_{\rm WD} \right)^{1/3}\left[\frac{2 R_{\rm WD}}{FWHM(\rm cm~s^{-1})}\right]^{2/3}$$ From our gravitational redshift calculations, $R_{\rm WD}\approx$8.4$\times 10^8$ cm, and the FWHM of the line is 16 km s$^{-1}$, giving a radius for the disk of 4.5$\times10^{10}$ cm or $\approx$54 $R_{\rm WD}$. This is consistent with the gas being close to the WD and within the tidal disruption radius of WD 1124-293, as well as being from the sublimation of optically thin dust. The constraints on the location of the gas and the presence of Ca II combined with the lack of any appreciable ISM all point to a similar origin for WD 1124-293 as that posited for WDs with near-IR excesses due to dusty disks. These observations fit in with the scenario where dusty excesses are caused by massive asteroids or comets that are dynamically perturbed by a planetary system [@debes02; @bonsor11; @bonsor12; @debes12] and then tidally disrupted [@jura03]. However, a size distribution of planetesimals will be disrupted over time, and for smaller bodies, sputtering renders most tidally disrupted dust into a gaseous phase quite quickly, forming circumstellar gas disks instead of circumstellar dust disks [@jura08; @farihi12]. Circumstellar gas absorption roughly at the same radius as that observed for dusty disks around WDs is also suggestive of a tidally disrupted asteroidal origin for our observations of WD 1124-293. Limits to Transiting Companions to WD 1124-293 {#s:trans} ============================================== Given the possible edge-on inclination of this system due to the presence of circumstellar gas, the probability of any detectable planetary companion transiting WD 1124-293 might be higher than a randomly inclined system. To that end we have gathered all data of WD 1124-293 from the Wide Angle Search for Planets (WASP) transit database. From previous observations, WD 1124-293 does not possess close-in brown dwarf or stellar companions. In their comprehensive [Spitzer]{} search for IR excesses due to companions around a sample of WDs, @farihi08 placed a definitive upper limit of 12 M$_{Jup}$ to the presence of any unresolved companions in orbit around WD 1124-293. The WASP North and South telescopes are located in La Palma (ING - Canary Islands) and Sutherland (SAAO - South Africa), respectively. Each telescope consists of 8 Canon 200 mm $f/1.8$ focal lenses coupled to e2v 2048$\times$2048 pixel CCDs, yielding a field-of-view of 7.8$\times$7.8 square degrees with a pixel scale of 13.7 [@pollacco06]. The WASP observing strategy is such that each field is observed with a typical cadence of the order of 8–10 minutes and typical exposures of 1 minute (30 sec exposure plus over heads). WASP provides good quality photometry with a precision exceeding 1% per observation in the approximate magnitude range $9 \le {\rm V} \le 12$. WD 1124-293 was observed by the WASP-South telescope in the 2007 and 2008 seasons covering the interval 2007 January 04 to 2007 June 05, and 2008 January 05 to 2008 May 28, respectively. The 10618 pipeline-processed photometric measurements were de-trended using the Tamuz algorithm [@Tamuz05] to account for linearly-correlated systematic noise in the data. We show the WASP light-curve in Figure \[fig:f7\] (bottom-panel). We used the modified implementation of the Box-Least Square (BLS) algorithm described in @Faedi2011 to search for transits and eclipses of sub-stellar and planetary companions in close orbits around the white dwarf WD1124-293. We searched a parameter space defined by orbital periods ranging from 2 hours to 15 days, and companion sizes ranging from Moon-size to Jupiter-size. No transiting planet has been identified. Figure \[fig:f6\] (top-panel), shows the BLS power-spectrum for the star WD1124-293. The dashed-line indicates the detection threshold as defined in @Faedi2011 for 10% noise levels. We have phase-folded the WASP light-curve to investigate the possible presence of an astrophysical signal at the periods detected above the threshold. However, none showed real variation. Finally, we have used the modified Lomb-Scargle periodogram [@Scargle82] to search for generic variability. We find that a real power is assigned to a peak at P=27.42 days which is most probably related to the Moon cycle. We used these null results together with the results of our simulations to estimate an upper-limit to the frequency of close companions to WD 1124-293. To assess the chances of detecting eclipses and transits of substellar and planetary companions to WD1124-293, we performed an extensive set of Monte Carlo simulations. The approach we adopted was to create realistic synthetic light-curves containing eclipse and transit signatures of the expected depth ($\delta_{\rm tr}$) and duration for a range of companion sizes and orbital periods, then to attempt to detect these signatures using the modified BLS mentioned above. By noting the rate at which the BLS search recovered the transit ($f_{\rm det}$) at the correct period (or an integer multiple or fraction) we were able to estimate the detection limits of such systems in an automated manner. Details of the synthetic transit signal generated, the corresponding transit probability, depth and duration are described in detail in @Faedi2011. Here we present the results of simulations covering the orbital period-planet radius parameter space defined by seven trial periods spaced approximately logarithmically ($P=0.08$, 0.22, 0.87, 1.56, 3.57, 8.30 and 14.72 days), and five planet radii $R_{\rm p}=10.0$, 4.0, 1.0, 0.55, and 0.27R$_{\oplus}$. Figure \[fig:f8\] shows the detection rate of the simulated transit signals injected in the WASP light curve of WD1124-293. We regard as a match any trial in which the most significant detected period is within 1% of being an integer fraction or multiple from $1/5\times$ to $5\times$ the injected transit signal. From our result on the simulations conducted using the WASP light curve of WD1124-293 (Figure \[fig:f8\], see also results on simulations conducted on WASP data detailed in @Faedi2011), we would have been able to detect the transit signal due to a companion of radius 1R$_{\oplus}$ and period between 2 and 16h with a confidence of 75%. For objects of radius similar to that of the Earth but slightly bigger (e.g. up to 3R$_{\oplus}$, or so-called SuperEarths) we would have been able to detect the eclipse signal for periods between 2h and 12h with a confidence of 95% and to about 1d with 75% confidence. Eclipsing gas giant planets and substellar objects would have been detectable at orbital periods between 2h and 1.7d with 95% confidence, and with orbital periods of 2d with 75% confidence. On the other end, no stringent constraint can be placed on objects with radius $<1 {\rm R}_{\oplus}$ even at short periods. The same is true even for large bodies such as Jupiter with orbital periods longer 2 days. From Figures 1 and 2 of @agol11 and using the white dwarf temperature of 9420 K and mass of 0.66 M$_{\odot}$ we find that for WD1124-293 the white dwarf habitable zone (WDHZ) extends for a region comprised between 0.02 and 0.045 AU for which, in our simulations, we can only put weak constraints for Neptune/Uranus-size ($\sim3-4{\rm R_{\oplus}}$) and larger objects. However, eventually WD 1124-293 will cool further and when it cools to below 7000-9000 K, planets with orbital separations of 5$\times10^{-3}$-0.02 AU (P$\sim$0.16-1.3 days) will be in a “continuous habitable zone” which lasts for $>$3 Gyr [@agol11]. Our observations thus rule out a significant fraction of $>$ Earth-sized planets in this system that might reside in the continuous habitable zone of WD 1123-293. We investigated the various factors which affect the efficiency of detection of these transit signals. When generating each synthetic light-curve we can easily assess [a priori]{} whether it will fail the tests requiring a minimum number of individual transits ($>5$) and in-transit data points ($>5$). We find that our requirements alone render transiting companions essentially undetectable at our longest trial periods (8.30 and 14.72 days); the transits are too short in duration and too infrequent to be adequately sampled. For companions around 1[R$_{\oplus}$]{} and larger however there is a good chance of detection out to periods of $\approx$2 days. Limits to a variable Accretion Rate to WD 1124-293 {#sec:var} ================================================== Given the stochastic origins to dust around WDs, it might be possible to observe variability in WD accretion rates. Ca variability was claimed to be present in the dusty DAZ G 29-38 [@vonhippel07b], but disputed with other observations [@debes08]. Discrepant abundances of Si between optical and UV observations of some dusty WDs may suggest possible variable accretion [@gaensicke12]. The multiple spectra of WD 1124-293 allow a careful look at the variability in the accretion rate of material onto the surface of the WD through EW measurements of its photospheric Ca II lines. WD 1124-293’s metal settling timescale for calcium is 194 yr, much longer than the total time high quality spectra of the Ca line have been available, but short enough that small changes due to settling may be observed. Similarly, if there were any short duration accretion events that briefly increased the accretion rate, we may have observed a sudden increase in the Ca EW. Figure \[fig:f8\] shows the EW as a function of Julian date (JD). Two measurements of the Ca II K line, from May 2009 and July 2010, differ significantly from the other measurements. Similarly, the Ca II H line measurements from 1998 and March 2007 also deviate significantly. The cause of these discrepancies most likely due to lower S/N and poor seeing, but given the good agreement between the other EW measurements and the lack of similar variability at the same time from the complementary Ca II line, we neglect these measurements in our analysis. To determine whether WD 1124-293’s Ca II line strength was varying, we calculated a $\chi^2$ value for a constant value EW, determined from the median value of the observed EWs. With this measure, the Ca EW has a $\chi^2$ value of 23 (14) for the Ca II K (H) line, which corresponds to a probability that the EW is constant to 0.03 (0.3). Based on our full survey of other DAZs for variability [@debmuk10], we require the probability of constant EW to be less than 10$^{-3}$ to be considered significantly variable. The standard deviation of our MIKE measurements, neglecting the anomalous measurements, is 4.5% for the K line and 7% for the H line. The roughly constant flow of material over $\sim$11 years for this system places a firm lower limit on the amount of metal-rich material recently deposited, 4.6$\times10^{16}$ g, once again assuming metal rich material at 1% the solar abundance values, and a total amount of material over one settling time of 9$\times10^{18}$ g assuming a settling time of 194 yr [@koester05]. These masses correspond to asteroids with sizes of 1.5 and 9 km respectively, assuming a bulk density of 3 g/cm$^{3}$. These radii are comparable to the sizes of asteroids that should survive post-main sequence evolution at an orbital radius of a few AU [@jura08]. Conclusions {#sec:conc} =========== We have obtained moderate resolution optical spectroscopy of the DAZ WD 1124-293 over several epochs and found a weak circumstellar absorption feature of Ca II consistent with a tenuous gaseous disk supplied by small rocky bodies. Steady state accretion of material onto the WD surface appears to be occurring despite the relatively longer inferred settling timescale for metals, consistent with a recent disruption of an asteroidal body that was relatively small. WD 1124-293 has also been searched for transiting companions given that any planetary system is nearly edge-on to our line of sight, though none are found for short periods. Our results do not rule out the possibility for planetary companions at wider separations. Emission from gaseous disks have been observed around metal enriched WDs with disks that have been actively accreting dust [@gaensicke06; @gaensicke07; @gaensicke08]. The origin of that gas could arise from the viscous heating of material [@werner09] or from the photoionization of gas [@melis10]. However, these systems may be unusual in the fact that they may represent an early stage in the creation of a dusty disk or represent a massive example of a dusty disk in orbit around a WD. The accretion of many small planetesimals should proceed with little observable dust, and a primarily gaseous circumstellar disk [@jura08]. These properties match WD 1124-293, representing the first example of a single metal enriched WD with a direct link to the tidal disruption of planetesimals and thus to a relic planetary system. The large fraction of single WDs that show metal pollution implies a large fraction of post-main sequence objects may house remnant planetary systems with a robust mechanism for delivering small objects in WD-grazing orbits. A large spectroscopic survey of apparently single, dust-free WDs with high S/N may reveal more systems similar to WD 1124-293. We wish to thank Detlev Koester for unceasing aid in discussing the fundamental parameters of WD 1124-293, and Silvano Desidera for kindly providing valuable spectra of HIP 56280. We also wish to thank Jean-Rene Roy for discussions on the kinematic broadening of spectral absorption lines, which provided a key impetus for constraining the location of WD 1124-293’s gaseous disk. Finally, we want to thank the anonymous referee, whose attention to detail greatly increased the quality of this paper. This work made significant use of the SIMBAD and VIZIER databases, operated at CDS, Strasbourg, France. [ccccc]{} 1998-12-12 (51159.629)& WD 1124-293 & Keck/HiRES & 37 & R=30,000\ 1999-4-19 (51288.328) & WD 1124-293 & Keck/HiRES & 26 & R=30,000\ 2000-4-23 (51657.663) & WD 1124-293 & VLT/UVES & 36 & R=18500\ 2000-5-19 (51681.588) & WD 1124-293 & VLT/UVES & 14 & R=18500\ 2007-3-31 (54189.682) & WD 1124-293 & Clay/MIKE & 24 &\ 2008-3-21 (54546.605) & WD 1124-293 & Clay/MIKE & 50 &\ 2008-3-22 (54547.656) & WD 1124-293 & Clay/MIKE & 50 &\ 2008-5-13 (54599.536) & WD 1124-293 & Clay/MIKE & 60 &\ 2008-5-14 (54600.551) & WD 1124-293 & Clay/MIKE & 49 &\ 2008-5-15 (54601.477) & WD 1124-293 & Clay/MIKE & 52 &\ 2008-5-16 (54602.502) & WD 1124-293 & Clay/MIKE & 56 &\ 2009-4-16 (54937.686) & WD 1124-293 & Clay/MIKE & 39 &\ 2009-5-18 (54969.548)& WD 1124-293 & Clay/MIKE & 38 &\ 2010-1-1 (55197.727) & HIP 55731 & Clay/MIKE & 36 & ISM Standard\ 2010-1-1 (55197.733) & HIP 55864 & Clay/MIKE & 40 & ISM Standard\ 2010-1-1 (55197.738) & HIP 55901 & Clay/MIKE & 100 & ISM Standard\ 2010-1-1 (55197.743) & HIP 55968 & Clay/MIKE & 47 & ISM Standard\ 2010-6-17 (55365.467) & HIP 55731 & Clay/MIKE & 78 & ISM Standard\ 2010-6-17 (55365.472) & HIP 55901 & Clay/MIKE & 162 & ISM Standard\ 2010-6-17 (55365.479) & HIP 55864 & Clay/MIKE & 105 & ISM Standard\ 2010-6-17 (55365.486) & WD 1124-293 & Clay/MIKE & 73 &\ 2010-6-17 (55365.575) & HIP 55968 & Clay/MIKE & 56 & ISM Standard\ 2010-7-07 (55384.500) & WD 1124-293 & Clay/MIKE & 33 &\ 2011-6-09 (55385.493) & WD 1124-293 & Clay/MIKE & 70 &\ [cccccc]{} Photospheric & 51159.629 & 104$\pm$ 3 & 39$\pm$ 4 & 30.3 & 31.3\ & 51288.328 & 98$\pm$ 5 & 57$\pm$ 6 & 28.7 & 27.5\ & 51657.663 & 111$\pm$ 6 & 58$\pm$12 & 29.1 & 28.0\ & 51681.588 & 109$\pm$17 & ... & 29.8 & ...\ & 54189.682 & 106$\pm$ 4 & 34$\pm$ 4 & 29.7 & 34.5\ & 54546.605 & 116$\pm$ 2 & 52$\pm$ 3 & 29.7 & 29.8\ & 54547.658 & 113$\pm$ 2 & 54$\pm$ 3 & 29.0 & 29.4\ & 54599.536 & 113$\pm$ 2 & 53$\pm$ 2 & 28.3 & 28.3\ & 54600.551 & 114$\pm$ 3 & 49$\pm$ 3 & 28.6 & 28.8\ & 54601.477 & 113$\pm$ 2 & 55$\pm$ 2 & 29.0 & 30.1\ & 54602.502 & 118$\pm$ 2 & 55$\pm$ 2 & 29.2 & 29.4\ & 54937.686 & 113$\pm$ 3 & 53$\pm$ 3 & 28.8 & 27.7\ & 54969.548 & 127$\pm$ 3 & 43$\pm$ 3 & 29.1 & 27.8\ & 55364.486 & 115$\pm$ 2 & 52$\pm$ 1 & 28.7 & 29.5\ & 55385.493 & 95$\pm$ 3 & 59$\pm$ 4 & 28.4 & 29.7\ & 55721.598 & 111$\pm$ 2 & 54$\pm$ 2 & 29.7 & 29.4\ Circumstellar & March 2008 & 11$\pm$2 & ... & -0.5$\pm$0.9 & ...\ & May 2008 & 12$\pm$1 & ... & -2$\pm$1 & ...\ & June 2010+June 2011 & 10$\pm$1 & ... & -1$\pm$1 & ...\ & Total & 11$\pm$1 & 2$\pm$0.5 & -0.9$\pm$0.5 & -0.4$\pm$0.5\ [ccccccccc]{} HIP 56280 & ... & $<$10 & ... & $<$10 & ... & $<$10 & ... & $<$10\ HIP 55864 & ... & $<$ 7 & ... & $<$ 7 & ... & $<$ 7 & ... & $<$ 7\ HIP 55731 & ... & $<$ 6 & ... & $<$ 6 & 8.7 & 22$\pm$2 &... & $<$ 6\ HIP 55901 & -10.3 & 47$\pm$1 & -1.7 & 34$\pm$1 & 9.5 & 17$\pm$1 & ... & $<$2\ HIP 55968 & -6 .3& 72$\pm$2 & 2.0 & 43$\pm$2 & 11.8 & 56$\pm$2 & 21.2 & 19$\pm$2\ [ccccc]{} \[tab:WASP\] Jupiter &10.0 & 100 & 0.08 & 100\ & & & 0.22 & 100\ & & & 0.87 & 100\ & & & 1.56 & 91\ & & & 3.57 & 8\ & & & 8.30 & 0\ \ Uranus & 4.0 & 100 & 0.08 & 100\ & & & 0.22 & 100\ & & & 0.87 & 99\ & & & 1.56 & 82\ & & & 3.57 & 1\ & & & 8.30 & 0\ \ Earth & 1.0 & 49& 0.08 & $-$\ & & & 0.22 & 99\ & & & 0.87 & 61\ & & & 1.56 & 1\ & & & 3.57 & 0\ \ \	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this advance, we study the system of strongly interacting spinor bosons in a square lattice subject to any of the linear combinations of the Rashba and Dresselhaus spin-orbit coupling (SOC). The SOC leads to novel and rich quantum and topological phenomena in a lattice system. They include masses generated from “order from quantum disorder ” mechanism, collinear spin-bond correlated magnetic phase, quantum commensurate ( C ) and In-commensurate ( IC ) non-coplanar Skyrmion crystal phases, stripe co-planar ( spiral ) C and IC phases, quantum Lifshitz C-IC transitions, metastable states, hysteresis, topological rational and irrational winding numbers, incomplete and complete devil staircases, Cantor sets, multi-fractals, etc. Various perspectives and possible intimate connections with topological spin liquids due to geometric frustrations or quantum spin glass due to quenched disorders are discussed. Implications on current or near future cold atom systems and on 4d or 5d strongly correlated materials with SOC are discussed.' author: - 'Fadi Sun$^{1,2,3}$ and Jinwu Ye $^{1,2,3}$' title: ' In-commensurate Skyrmion crystals, devil staircases and multi-fractals due to spin orbit coupling in a lattice system ' --- [INTRODUCTION. ]{} It was well known that the strong correlations among bosons or fermions lead to many quantum, topological phases and phase transitions in materials [@scaling; @sachdev; @aue]. Its combinations with geometric frustrations may lead to new phases of matter such as coplanar spiral phases, especially topological quantum spin liquids [@sachdev; @aue; @bert; @spinliquid]. Its combinations with quenched disorders also lead to new states of matter such as the quantum spin glass [@glass1; @glass2; @glass3; @syk0] which is closely related to black hole physics through $ AdS/CFT $ correspondences [@mess3; @kittalk; @syk1; @syk2; @syk3]. On the other forefront, Rashba or Dresselhaus spin-orbit coupling (SOC) is ubiquitous in various 2d or layered non-centrosymmetric insulators, semi-conductor systems, metals and superconductors [@rashba; @ahe; @socsemi; @ahe2; @she; @niu; @aherev; @sherev]. There are also recent remarkable experimental advances in generating any linear combinations of the 2d Rashba and Dresselhaus SOC for both fermions and spinor bosons in both continuum and optical lattices [@expk40; @expk40zeeman; @2dsocbec; @clock; @clock1; @clock2; @SDRb; @ben]. New many body phenomena due to the interplay among strong interactions, the SOC and lattice geometries are being investigated in the current cold atom experiments. It becomes urgent, topical and important to investigate what would be the new quantum or topological phenomena due to the the interplay between the strong correlations and the ubiquitous Rashba or Dresselhaus SOC on various lattices. In this work, we address this outstanding problem and discover that the interplay leads to many novel quantum or topological phenomena summarized in the global quantum phase diagram Fig.1. They include masses generated from “order from quantum disorder ” phenomena, the collinear spin-bond correlated Y-x phase and its novel excitation spectrum, quantum commensurate ( C ) and In-commensurate ( IC ) non-coplanar Skyrmion crystal ( SkX ) phases, stripe co-planar ( spiral ) C and IC phases, quantum Lifshitz C-IC transitions, metastable states, hysteresis, topological fractional and irrational winding numbers, incomplete and complete devil staircases, Cantor sets, multi-fractals, etc. We establish the Fig.1 by the combinations of the approaches from the extremely anisotropic limit [@rh] $ (\alpha=\pi/2, \beta ) $, the isotropic Rashba limit $ 0 <\alpha=\beta < \pi/2 $ and near the Abelian line $ 0 < \alpha < \pi/2, \beta=0 $. Our results demonstrate that the interplay among strong correlations, Rashba SOC and lattice geometries opens a new avenue to explore whole new classes of quantum or topological phenomena which may have wide implications in cold atoms and various materials with SOC to be discussed near to the end of the paper. The tight-binding Hamiltonian of ( pseudo)-spin $ 1/2 $ bosons ( or fermions ) at integer ( or half ) fillings hopping in a two-dimensional square optical lattice subject to any combination of Rashba and Dresselhaus SOC is: $$\mathcal{H}_{b/f}= -t\sum_{\langle ij\rangle}(b_{i\sigma}^\dagger U_{ij}^{\sigma\sigma'} b_{j\sigma'}+h.c.) + \frac{U}{2} \sum_i ( n_{i}-N )^2 \label{hubbardint}$$ where $ t $ is the hopping amplitude along the nearest neighbors $\langle ij\rangle$, the non-Abelian gauge fields $U_{iㄛi+\hat{x} } =e^{i \alpha \sigma_x}$, $U_{iㄛi+\hat{y} }=e^{i \beta \sigma_y}$ are put on the two links in a square lattice. $ \alpha=\pm \beta $ stands for the Rashba ( Dresselhaus ) case. $ \alpha \neq \beta $ corresponds to any linear combination of the two. $U>0$ is the Hubbard onsite interaction. In the strong coupling limit $ U/t \gg 1 $, to the order $O(t^2/U)$, we obtain the effective spin $ s=N/2 $ Rotated Heisenberg model: $$\mathcal{H}_{RH} = -J\sum_i [\mathbf{S}_i R(\hat{x},2\alpha)\mathbf{S}_{i+\hat{x}} +\mathbf{S}_i R(\hat{y},2\beta)\mathbf{S}_{i+\hat{y}}] \label{rhgeneral}$$ with $J= \pm 4t^2/U > 0 $ for bosons/fermions, the $R(\hat{x},2\alpha)$, $R(\hat{y},2\beta)$ are the two SO(3) rotation matrices around the $ X $ and $ Y $ spin axis by angle $2\alpha$, $2\beta$ putting on the two bonds along $\hat{x} $, $\hat{y} $ respectively. In this paper, for simplicity, we first focus on spinor bosons. The fermions will be discussed in a separate publication. The RFHM Eq.\[rhgeneral\] at a generic $ (\alpha, \beta ) $ has the translational, the time reversal $ {\cal T} $, the three spin-orbital coupled $ Z_2 $ symmetries $ {\cal P}_x, {\cal P}_y, {\cal P}_z $ symmetries [@rh]. Along the extremely anisotropic limit $\alpha=\pi/2, 0<\beta<\pi/2$, there is a hidden spin-orbital coupled $ U(1) $ symmetry generated by $ U_1(\phi)=e^{ i \phi \sum_{i} (-1)^x S^{z}_i } $ and also the Mirror symmetry $ {\cal M} $ which consists of the local rotation $\tilde{\mathbf{S}}_{i} =R(\hat{x},\pi ) R(\hat{z},\pi n_2) \mathbf{S}_{i}$ followed by a Time reversal transformation $ {\cal T } $, $ (\beta, h ) \rightarrow ( \pi/2 - \beta, h ) $. However, any deviation from the extremely anisotropic line $ \alpha \neq \pi/2 $ spoils the $ U(1) $ and Mirror symmetry. Along the isotropic Rashba limit $ \alpha=\beta $, the $ {\cal P}_z $ symmetry is enlarged to the spin-orbital coupled $ C_4 \times C_4 $ symmetry around the $ z $ axis. Of course, along the bottom Abelian line $ 0 < \alpha < \pi/2, \beta=0 $, it has the $ \tilde{SU}(2) $ symmetry in the $ \tilde{SU}(2) $ basis $ \tilde{\mathbf{S}}_n= R(\hat{x}, 2 \alpha n )\mathbf{S}_n $. Because $ \beta < \alpha $ lower-half is related to the $ \beta > \alpha $ upper half in Fig.\[phasedia\] by the $ C_4 \times C_4 $ transformation, so in the following, we mainly focus on the lower half. We will approach the global phase diagram Fig.1 from all the three lines: the solvable line $ \alpha=\pi/2, 0<\beta<\pi/2$, the Abelian line $ 0 < \alpha < \pi/2, \beta=0 $ and the diagonal line $ 0 < \alpha=\beta < \pi/2 $. ![ Phase diagram of the strongly interacting spinor bosons at a generic SOC $ ( \alpha, \beta) $ in a square lattice. Along the diagonal line $ \alpha_{in} < \alpha < \pi/2 $, there is a gap opening generated by the order from disorder mechanism. There is a quantum Lifshitz transition at $ \alpha= \alpha_{in} $ with the dynamic exponent $ z=1 $, from the collinear Y-x ( or X-y ) phase to the non-coplanar IC-SkX/Y-x ( or IC-SkX/X-y ) phase, then a second one from the IC-SkX/Y-x to the commensurate $ 3 \times 3 $ SkX crystal phase at $ \alpha= \alpha_{33} $ which is a bi-critical point. The segment $ \alpha_{33} < \alpha < \alpha_{in} $ is the first order transition line between the IC-SkX/Y-x and IC-SkX/X-y. All the other segments are first order transition lines between $ N \times 1 $ and $ 1 \times N $ co-planar spiral phase after $ N \geq 4 $. M is the multicritical point located at $ ( \alpha_M, \beta_M ) $ where the $ ( 0, \pm 2\pi/3) $ counter line of the Y-x phase hits the corner of the $ 3 \times 3 $ SkX crystal. There is a second order transition from the Y-x phase to the non-coplanar IC-SkX/Y-x phase driven by the condensations of C-IC with $ \pi-\pi/3 < k^{0}_y < \pi- q_{ic} $. The Y-x state becomes metastable between the first order transition line and the second order transition ( dashed ) line due to the condensations of the $ C-C_0 $ and C-IC with $ 0 < k^{0}_y < 2\pi/3 $. The in-complete devil’s staircases at a small $ \beta < \alpha=\pi (1/2,1/3,1/4,....)=\pi/N $ stand for commensurate stripe co-planar spiral phases. The IC means the stripe co-planar in-commensurate phases. Immersed inside the IC phase are some small devil staircases ( such as at $ 3\pi/7, 2\pi/5, 2\pi/7...... $ ) displaying fractal structures. The relevant numbers are $ \alpha^{0}_{in} \sim 0.3611 \pi, \alpha_{ic} \sim 0.3526 \pi, \alpha_{33} \sim 0.3402 \pi, (\alpha_M,\beta_M) \sim(0.33952\pi,0.31284\pi) $ and $ q_{ic} \sim 0.18 \pi $. []{data-label="phasedia"}](contours2.eps){width="50.00000%"} [RESULTS ]{} [C and IC Magnons in the Y-x state and their condensations. ]{} The firmly established results and physical insights [@rh] achieved on the extremely anisotropic line $ ( \alpha=\pi/2, \alpha < \beta ) $ pave the way to study the physics at generic $ ( \alpha, \beta ) $ in Fig.\[phasedia\]. Especially, we will follow how the three kinds of magnons response and evolve when moving away from the line. Making a globe rotation $R_x(\pi/2)$ to align spin along the Z-axis and then introducing Holstein-Primakoff bosons $ a $ and $ b $ for the two sublattice, we can expand the Hamiltonian in the powers of $1/\sqrt{S}$, $$\begin{aligned} H=E_0+2JS\Big[H_2 +\Big(\frac{1}{\sqrt{S}}\Big)H_3 +\Big(\frac{1}{\sqrt{S}}\Big)^2H_4+\cdots\Big] \label{swcubic}\end{aligned}$$ where the symbol $H_n$ denotes the $n$-th polynomial of the boson operators, $E_0=-2NJS^2\sin^2\alpha$ is the classical ground state energy of the Y-x state. Performing a unitary transformation, then a Bogoliubov transformation on $ H_2$, one can diagonize $ H_2 $ as: $$\begin{aligned} H_2= E_2 +2\sum_k(\omega_k^+\alpha_k^\dagger\alpha_k +\omega_k^-\beta_k^\dagger \beta_k)\end{aligned}$$ where $ E_2=\sum_k(\omega_k^++\omega_k^--2\sin^2\alpha) $ is the quantum correction to the ground state energy at the LSW order, $\omega_k^{\pm}=\sqrt{(\lambda_k^{\pm})^2-\chi_k^2}$, $\lambda_k^{\pm}=\sin^2\alpha-\frac{1}{2}\cos2\beta\cos k_y \pm\frac{1}{2}\sqrt{\sin^4\alpha\cos^2k_x+\sin^22\beta\sin^2 k_y}, \chi_k=\frac{1}{2}\cos^2\alpha\cos k_x $. Obviously, $\omega_k^{\pm}=\omega_{-k}^{\pm}$ which is dictated by the symmetries of the Hamiltonian and the Y-x state. Note that to the LSW order, the dispersion still has the Mirror symmetry under the $ \beta \rightarrow \pi/2-\beta $. However the mirror symmetry will be spoiled by the higher order terms starting at $ H_3 $. As shown in [@rh], at $ \alpha=\pi/2 $, the Y-x state is the exact ground state, $ \chi_k=0 $, there is no need for the extra Bogoliubov transformation, the spin wave dispersion reduces to $ \omega_k^{\pm}=\lambda_k^{\pm} $. As shown in [@rhtran], any transverse field $ h_x $ or $ h_z $ transfers the Y-x state into a co-planar canted state. In sharp contrast, here, under $ \pi/2 - \alpha \neq 0 $, the Y-x state remains the classical state, but not the exact eigenstate anymore due to the quantum fluctuations introduced by $ \alpha \neq \pi/2 $. From $ \omega_k^{-} $, one can identify the minimum $ (0,k_y^0)$ of spin-wave dispersion corresponding to the magnons C-$C_0$, C-IC, C-$C_{\pi} $ respectively ( See SM ). Near $ (0,k_y^0)$, their dispersions take the relativistic form: $$\omega_{-}(q) = \sqrt{ \Delta^2 + v^{2}_x q^{2}_x + v^{2}_y q^{2}_y } \label{relagap}$$ The gap and the two velocities are given in the SM. The Staggered magnetization and specific heat of the Y-x phase at $ T \ll \Delta $ are: $$\begin{aligned} M(T) & \sim & M(T=0)- \frac{T\Delta}{2\pi v_xv_y}\sqrt{1+\frac{\cos^4\alpha}{4\Delta^2}}e^{-\Delta/T} \nonumber \\ C(T) & \sim & \frac{1}{2\pi v_xv_y}\frac{\Delta^3}{T}e^{-\Delta/T} \label{mc}\end{aligned}$$ where $ M(T=0)= S-\frac{1}{N} \sum_k(\frac{\lambda_k^+}{2\omega_k^+} +\frac{\lambda_k^-}{2\omega_k^-}-1) $ is the $ T=0 $ magnetization. At $ \alpha=\pi/2 $, replacing $v_x$ by $\sqrt{\Delta/m_x}$ and $v_y$ by $\sqrt{\Delta/m_y}$, Eqn.\[mc\] gives back to those along the solvable line in [@rh]. Solving $\Delta=0$ leads to the 3 segments of their condensation boundary: $$\begin{aligned} \alpha= \begin{cases} &\pi/2-\beta, \\ & \arcsin\left[ \frac{\sqrt{6}\sin2\beta}{\sqrt{9\sin^22\beta-1}} \right], \\ & \beta, \end{cases} \label{segments}\end{aligned}$$ for $ 0\leq\beta\leq\pi/2-\arccos(1/\sqrt{6}), \pi/2-\arccos(1/\sqrt{6})\leq\beta\leq\arccos(1/\sqrt{6}) $ and $ \arccos(1/\sqrt{6})\leq\beta\leq\pi/2 $ respectively. At the LSW order, it still has the mirror symmetry under $ \beta \rightarrow \pi/2-\beta $. The C-C$_\pi$ magnons condense along the diagonal line $ \arccos(1/\sqrt{6})\leq \beta \leq \pi/2 $ with the gapless relativistic dispersion: $$\begin{aligned} \omega_-(q)=\sqrt{v_x^2 q_x^2+v_y^2 q_y^2} \label{gapless}\end{aligned}$$ where $ v_x=\cos(\alpha)/2,~ v_y=\cos(\alpha)\sqrt{1-6\cos^2(\alpha)}/2 $. Obviously, both velocities vanish at the Abelian point $ \alpha=\pi/2, \beta=\pi/2 $ dictated by the enlarged $ SU(2) $ symmetry. Moving away from the diagonal line $ \alpha=\beta $, $ v_x$ keeps increasing, but $ v_y$ increases first, reaches a maximum, then vanishes at the boundary between C-$ C_{\pi} $ and C-IC magnons $ \alpha^{0}_{ic}= arccos(1/\sqrt{6} ) \sim 0.36614 \pi $. When pushing to higher orders, $ \omega_-(q)=\sqrt{v_x^2 q_x^2+v_y^2 q_y^2 + u^2 q_y^4+ \cdots } $, we find it is a putative $ ( z_x=1, z_y=2 ) $ quantum Lifshitz transition from the Y-x state to an incommensurate state ( Fig.4a). However, as to be shown below, the gapless mode along the diagonal line and the mirror symmetry under $ \beta \rightarrow \pi/2 - \beta $ are just spurious facts of the LSW approximation. However, the quantum Lifshitz transition remains, but with a different dynamic exponent than $ ( z_x=1, z_y=2 ) $. ![ Some most robust Collinear, spiral, vortex and non-coplanar states in Fig.\[phasedia\]. Top layer: the $ 2 \times 1 $ ( Y-x ) state $ S^{y}= (-1)^x $ is the exact quantum ground state [@rh] at $ \alpha=\pi/2 $. When $ \beta $ is small, the $ 3 \times 1 $ spiral state is close to be a FM state in the rotated basis $ \tilde{\mathbf{S}}_n = R(\hat{x}, 2 \alpha n )\mathbf{S}_n $. All the red arrows in $ 120^{\circ} $ structure ( connected by the dashed line ) will be transformed to a FM state in the rotated basis, the blue arrows are actual spiral spin orientations which only deviate slightly from the red arrows. The deviation angles increases as $ \beta $ increases in the $ 3 \times 1 $ staircase. The spiral states at $ 4 \times 1 $, $ 5 \times 1 $ and other devil’s staircases ( not shown ) can be similarly constructed. The inset show the spin axis for the collinear and spiral states. The degeneracy is $ 2 N $ for odd $ N $ and $ N $ for even $ N $. There is also a small magnetization for $ N $ odd, but exactly zero for $ N $ even. Bottom layer: the classically degenerate ( 2 in , 2 out ) $ 2 \times 2 $ vortex state along the diagonal line $ \alpha=\beta $ is simply a FM state in the XY plane in the rotated basis $ \tilde{\mathbf{S}}_n= R(\hat{x},\pi n_1) R(\hat{y},\pi n_2)\mathbf{S}_n $. The $ 3 \times 3 $ skyrmion crystal ( non-coplanar ) state with non-vanishing skyrmion density $ \vec{S}_i \cdot \vec{S}_j \times \vec{S}_{k} \neq 0 $ happens near $ \alpha=\beta=\pi/3 $ which is the most frustrated regime in the Wilson loop [@rh]. Its detailed spin configuration is given in the Fig.S2 in the SM. The inset show the spin axis for the vortex and SkX states. []{data-label="allphases"}](ncross1nature.eps){width="8.2cm"} [Order from disorder along the diagonal line near the $ \alpha=\beta=\pi/2 $ Abelian point. ]{} It is important to understand what is the true quantum ground state along the diagonal line near the Abelian point $ \alpha=\beta=\pi/2 $. At the classical level, the $ 2 \times 1 $ Y-x stripy state $ S^y=(-1)^x $ is degenerate with the $ 1 \times 2 $ X-y stripy state $ S^x=(-1)^y $. In fact, we find there is a family of states called $ 2 \times 2 $ vortex states in Fig.\[allphases\]: $ \mathbf{S}_i= ( (-1)^{i_y}\cos\phi,(-1)^{i_x}\sin \phi,0 ) $ which are degenerate at the classical level. In general, this family breaks the $ C_4 \times C_4 $ symmetry except at $ \phi= \pm \pi/4, \pm 3\pi/4 $. When $ \phi=0, \pi/2 $, it recovers to the X-y and Y-x state respectively. Quantum fluctuations ( “order from disorder” mechanism ) are needed to find the unique quantum ground state upto the $ C_4 \times C_4 $ symmetry in this regime. To perform a LSW calculation, one need to introduce a 4 sublattice structure $ A, B, C, D $ shown in Fig.\[allphases\]. After making suitable rotations to align the spin quantization axis along the Z axis, we introduce 4 HP bosons $ a, b, c, d $ to perform a systematic $ 1/ S $ expansion shown in Eqn.\[swcubic\] where $E_0=-2NJS^2(1-\cos2\alpha\sin^2\phi-\cos2\beta\cos^2\phi)$ is the classical ground state energy, $ H_2 $ can be diagonized by a unitary transformation, then followed by a Bogoliubov transformation as: $$\begin{aligned} H_2= E_2 + 2\sum_{n,k} \omega_n (k)\alpha_{n,k}^\dagger \alpha_{n,k}\end{aligned}$$ where $ n=1,2,3, 4 $ is the sum over the 4 branches of spin wave spectrum in the Reduced BZ $ -\pi/2 < k_x, k_y < \pi/2 $ and $ E_2 $ is the $ 1/S $ quantum correction to the ground-state energy: $$\begin{aligned} E_2 =\sum_{k,n}[\omega_n(k) -(1-\cos2\alpha\sin^2\phi-\cos2\beta\cos^2\phi)/2] \label{quantumphi}\end{aligned}$$ Obviously, near the Abelian point $ \alpha=\beta=\pi/2 $, if $ \alpha > \beta $, it picks the Y-x state with $ \phi=\pi/2 $. If $ \alpha < \beta $, it picks the X-y state with $ \phi= 0 $. Setting $ \alpha=\beta $, the $ E_0 =-2NJS^2(1-\cos2\alpha) $ becomes $ \phi $ independent, indicating the classical degenerate family of states characterized by the angle $ \phi $ along the whole diagonal line $ \alpha=\beta $. Fortunately, the quantum correction $ E_2(\phi)=\sum_{k,n}[\omega_n(k,\phi)-\sin^2\alpha] $ does depend on $ \phi $. As shown in Fig.\[orderdis\]a, we find that $E_2(\phi)$ reach its minimum at $\phi=0$ ( X-y state ) or $\phi=\pi/2$ ( Y-x state) which is related to each other by the $ C_4 \times C_4 $ symmetry. Expanding $E_2(\phi)$ ( in unit of $ 2 J S $ ) around one of its minima $\phi=0$, $E_2(\phi)=E_2^0+ \frac{1}{2} B(\alpha) \phi^2+ \cdots $, one can identify the coefficient $ B(\alpha )$ as plotted in the Fig.\[orderdis\]b. ![ The order from disorder and the gap opening on the spurious gapless mode along the diagonal line in Fig.\[phasedia\]. (a) The quantum correction to the ground-state energy from the LSW. $\phi=0$ corresponds to X-y state and $\phi=\pi/2$ corresponds to Y-x state. So the quantum fluctuations pick up Y-x or X-y as the ground state which is related to each other by the $ C_4 \times C_4 $ symmetry. (b) The classical coefficient $ A( \alpha)/J $ and the quantum one $ B(\alpha )/J $. Both vanish at the Abelian point $ \alpha=\beta =\pi/2 $ as $ \sim ( \pi/2- \alpha)^2 $ and are monotonically increasing function when moving away from the Abelian point. The Dashed line is located at $ \alpha^{0}_{in} \sim 0.3661 \pi $ where the Y-x state becomes unstable at the LSW order. After incorporating the gap opening, the $ \alpha^{0}_{in} $ is shifted to a smaller value $ \alpha_{in} \sim 0.3526 \pi $. []{data-label="orderdis"}](EAB.eps){width="50.00000%"} The quantum order from disorder selection of the Y-x or X-y state along the diagonal line shows that there is a direct first order transition from the Y-x state to the X-y state along the diagonal line in Fig.1. So along the diagonal line, there is any mixture of the Y-x and X-y state. Similar first order transition between vacancy induced supersolid ( SS-v ) and interstitial induced supersolid ( SS-i ) and any mixtures of the two along the particle-hole symmetric line at the half filling in a triangular lattice were discussed in [@dual1; @dual2; @dual3]. [The magnon gap generated by the order from disorder mechanism. ]{} The gapless nature of the spin wave spectrum Eqn.\[gapless\] is just a spurious fact of the LSW approximation. It will be gapped out by the higher order terms in the $1/S $ expansion Eqn.\[swcubic\]. As shown in the next section, the quantum Lifshitz transition remains, but with a different dynamic exponent $ z=1 $ than that $ ( z_x=1, z_y=2 ) $ got within the LSW. It turns out that the leading order corrections to the gap at the minimum $ (\pi,0) $ of the C-$ C_{\pi} $ magnons can be achieved by the spin coherent state path integral formulation [@aue; @swgap]. A general uniform state at $ \vec{q}=0 $ can be taken as a FM state with the polar angle $ (\theta, \phi) $ in the $ \tilde{\tilde{SU}}(2) $ basis with $ \vec{\tilde{\tilde{S}}}_{i} = R(\hat{x},\pi n_1) R(\hat{y},\pi n_2) \vec{S}_{i} $ at the $ \alpha=\beta=\pi/2 $ Abelian point. After transforming back to the original basis by using $ \tilde{\tilde{S}}_1=R_z(\pi)S_1,~ \tilde{\tilde{S}}_2=R_y(\pi)S_2, \tilde{\tilde{S}}_3=R_x(\pi)S_3, \tilde{\tilde{S}}_4=S_4 $, it leads to a $ 2 \times 2 $ state characterized by the two angles $ \theta $ and $ \phi $. Along the diagonal line, its classical energy becomes $$\begin{aligned} H_0=J[-2 \sin^2 \alpha - 2\cos^2 \alpha \sin^2 \theta ] \label{classtheta}\end{aligned}$$ which is, as expected, $ \phi $ in-dependent. But one can see any deviation from the Abelian point picks up the XY plane with $ \theta=\pi/2 $. So it reduces to the $ 2\times 2 $ vortex state in Fig.\[allphases\] used in the “ order from disorder ” analysis in the last section. Expanding around the minimum $ H_0 =J[-2\sin^2\alpha+ 2\cos^2\alpha (\theta-\frac{\pi}{2})^2+\cdots] $ gives the stiffness $ A =2J\cos^2\alpha$ shown in Fig.\[orderdis\]b. Using the spin coherent state analysis, we can write down the quantum spin action at $ \vec{q}=0 $: $${\cal L}( \vec{q}=0 )= i S \cos \theta \partial_{\tau} \phi + \frac{1}{2} S^2 A ( \theta-\pi/2)^2 + \frac{1}{2} S B \phi^2 \label{action00}$$ where we put back the spin $ S $, the first term is the spin Berry phase term, $ A \sim (\pi/2-\alpha)^2 $ and $ B \sim (\pi/2-\alpha)^2 $ are from the classical analysis in Eqn.\[classtheta\] and the quantum order from disorder analysis to LSW order in Eqn.\[quantumphi\] respectively. Eqn.\[action00\] leads to the quantized Hamiltonian $ H(\vec{q}=0)= (a^{\dagger}_0 a_0 + 1/2 ) \hbar \Delta_B $ with the gap $ \Delta_B= \sqrt{S A B} \propto \sqrt{S} $. In fact, there are also corrections from the cubic $ H_3 $ and quartic $ H_4 $ terms in Eqn.\[swcubic\], but they only contribute to order of $ 1 $ which is subleading to the $ \sqrt{S} $ order in the $ 1/S $ expansion [@sw1; @japan]. As shown in Fig.\[orderdis\]b, because both $ A $ and $ B $ are monotonically increasing along the diagonal line, so the gap also increase. Plugging their values at $ \alpha=\alpha^{0}_{in} = \arccos(1/\sqrt{6}) $, When taking $A/J=1/3, B/J \approx0.008 $ and $S=1/2$, we find the maximum gap near the quantum Lifshitz transition $ \Delta_B/J \sim 0.036 $. [Quantum Lifshitz transition from the Y-x ( X-y ) to IC-SkX/Y-x ( IC-SkX/X-y ) state. ]{} As shown in the last section, there is a gap $ \Delta_B $ opening at $ \vec{q}=0 $ along the diagonal line, so the quantum Lifshitz transition point will shift to a smaller value of $ \alpha $. Incorporating the gap $ \Delta_B $ into the spin-wave dispersion $\omega_k$ in Eqn.8 at the LSW order leads to $\Omega_q=\sqrt{\Delta^2_B+\omega_q^2}$ ( See the Methods ). Because the spectrum along $ q_x $ is non-critical, so one can just put $ q_x=0 $: $$\begin{aligned} \omega_{-}(q_x=0, q_y)=\sqrt{ \Delta^2_B+v_y^2 q_y^2 + u^{2} q_y^4 + \cdots } \label{gap}\end{aligned}$$ where $ v^{2}_y= a ( \alpha^{0}_{in} - \alpha ) $ changes sign at $ \alpha = \alpha^{0}_{in} \sim 0.3611 \pi $ ( Fig.4a ). From the gap vanishing condition [@loff] at the IC wave-vectors $ q_{ic}= \pm (\Delta_B/u)^{1/2} $, one can see the quantum Lifshitz transition is shifted to $ \alpha_{ic} = \alpha^{0}_{in}- 2 u \Delta_B/a $. Plugging in the values of $ \Delta_B $ and $ u $, we find $ q_{ic} \sim 0.18 \pi $ ( Fig.4d ) and the shift is so small that $ \alpha_{ic} \sim 0.3526 \pi $ remains larger than $\alpha_{33} \sim 0.3402 \pi $ ( to be defined in the next section ) as shown in Fig.\[phasedia\]. So there must be an In-commensurate phase intervening between the Y-x state and the $ 3 \times 3 $ state when $ \alpha_{33} < \alpha < \alpha_{ic} $ in Fig.1. The transition from the Y-x to IC state is a quantum Lifshitz transition with the dynamic exponent $ z=1 $ ( Fig. S1d) instead of the one with $ ( z_x=1, z_y=2 ) $ at the LSW order in Fig.4a. The IC phase has the 4 orbital order wave-vectors $ ( 0, \pm ( \pi-q^{0}_y ) ) $ and $ ( \pi, \pm ( \pi-q^{0}_y ) ) $ with $ q^{0}_{y} \geq q_{ic} $. The spin structure of this IC phase remains to be determined. It should be a non-coplanar IC Skyrmion crystal phase which we name as IC-SkX/Y-x phase. Similarly, starting form the X-y phase, one can reach the IC-SkX/X-y phase with the 4 orbital order wave-vectors $ ( \pm ( \pi-q^{0}_x ), 0 ) $ and $ ( \pm ( \pi-q^{0}_x ), \pi ) $. The Y-x state has the C-$ C_{\pi} $ magnons when $ \alpha^{0}_{in} < \alpha < \pi/2 $, the C-IC magnons at the two minima $ ( 0, \pm k^{0}_y) $ with $ \pi- q_{ic} < k^{0}_y < \pi $ when $ \alpha_{in} < \alpha < \alpha^{0}_{in} $ as shown in Fig.4c. So along the diagonal line $ \alpha_{ic} < \alpha < \pi/2 $ ( $ \alpha_{33} < \alpha < \alpha_{ic} $ ), there must be co-existence of the Y-x and X-y ( IC-SkX/Y-x and IC-SkX/X-y ) phases with any ratios ( Fig.\[phasedia\] and its inset ). This physical picture will be substantiated further from the anisotropic line $ ( \alpha=\pi/2, \beta ) $ approached from the right. [Non-coplanar $ 3 \times 3 $ Skyrmion Crystal phase and Co-planar spiral phases along the diagonal line $ \alpha=\beta $. ]{} [1. $ 3 \times 3 $ non-coplanar Skyrmion Crystal phase (SkX). ]{} Near $ \alpha=\beta=\pi/3 $, it is natural to take a $3\times3$ ansatz: $S_{(i_x,i_y)}=S_{(i_x+3m,i_y+3n)}$ with $m,n\in \mathbb{Z}$. We estimate its classical ground-state energy by minimizing $ E_{3\times3}(\{\phi_i,\theta_i\}_{0\leq i\leq 9})$ over its 18 variables. Along the diagonal line ($\alpha=\beta$), as long as $\alpha$ is not too small, the minimization of $E_{3\times3}$ always leads to a $ C_4 \times C_4 $ symmetric $ 3 \times 3 $ SkX state ( See SM) shown in Fig.\[allphases\]. This is in sharp contrast to the case near $ \alpha=\beta=\pi/2 $ where the classical analysis only leads to the degenerate family of $ 2 \times 2 $ vortex states shown in Fig.\[allphases\]. A quantum “ order from disorder ” analysis is needed to show the $ 2 \times 2 $ vortex state phase separates into any mixtures of the Y-x state and X-y state along the diagnose line. Comparing the classical ground energy of the $ 3 \times 3 $ SkX with that of the Y-x state $E_{Y-x}=-2J\sin^2\alpha$ leads to a putative first order transition between the two states at $\alpha_{33} \approx0.340188\pi $ which is smaller than $ \alpha_{ic} \sim 0.3526 \pi $. So a putative direct first order transition between the Y-x state and the $ 3 \times 3 $ SkX splits into 2 second order quantum Lifshitz transitions with $ z=1 $ with the IC-SkX intervening between them. In fact, $ \alpha_{33} $ also shifts to a smaller value due to the intervening of the IC-SkX/Y-x phase, but for simplicity, we still use the same symbol. The point $ \alpha= \alpha_{33} $ in Fig.\[phasedia\] is a bi-critical point. Similarly, by LSW, we can determine the excitations spectra above the $ 3 \times 3 $ SkX. We expect the transition from the IC-SKY/Y-x state to the $ 3 \times 3 $ SkX is also a quantum Lifshitz transition with $ z=1 $. [2. The Y-x state to the IC-SkX/Y-x transition through the condensations of C-IC magnons ]{} As shown in the inset of Fig.\[phasedia\], approaching from the right in the Y-x phase, the crossing point between the $ ( 0, \pm 2\pi/3) $ counter line of the Y-x phase and the C-IC condensation boundary just hits the corner of the $ 3 \times 3 $ SkX crystal at the multicritical M point located at $ ( \alpha_M, \beta_M ) $, so there is a Y-x to the IC-SkX/Y-x phase due to the condensations of C-IC magnons with the ordering wavevectors $ \pi-\pi/3 < k^{0}_{y} < \pi- 0.18 \pi $. It leads to the non-coplanar IC-SkX/Y-x phase with the ordering wavevecrtors $ (0, \pm k^{0}_{y}) $ and $ ( \pi, \pm k^{0}_{y}) $. This confirms the picture achieved along the diagonal line in the previous section. Putting $ \alpha=\beta= \alpha_{33} $ into the formula for the constant contour at $ (0, k^{0}_{y} ) $ listed in the SM, one gets $ k^{0}_{y} \sim \pi-0.24 \pi $. So one can see that along the diagonal line $ \alpha_{33} < \alpha < \alpha_{ic} $, the ordering wavevector of the IC-SkX/Y-x is $ 0.18 \pi < q^{0}_{y} < 0.24 \pi $. While the transition from the $ 3 \times 3 $ SkX to the IC-SkX/Y-x on the right is through the condensations of C-IC magnons with $ 0.24 \pi < q^{0}_{y} < \pi/3 $. The non-coplanar IC-SkX/X-y phase with the ordering wavevecrtors $ ( \pm k^{0}_{x}, 0 ) $ and $ ( \pm k^{0}_{x}, \pi ) $. is related to IC-SkX/Y-x by the $ C_4 \times C_4 $ rotation. [3. Co-planar spiral states near $ \alpha=\beta=\pi/N $ and first order transition line. ]{} We find even at the classical level, there is a first order transition from the $ 4 \times 1 $ state to the $ 1 \times 4 $ state along the diagonal line ( Fig.S2). While one need resorts “ order from quantum disorder” mechanism to select out Y-x and X-y state as the quantum ground state near $ \alpha=\pi/2 $. This may be due to the fact that only near $ \alpha=\pi/2 $, the Y-x and X-y are collinear states and orthogonal to each other, while all the other commensurate states near $ \alpha=\pi/N, N > 2 $ are non-collinear ( but co-planar in the YZ plane ) spiral phases and not orthogonal to each other. It turns out the $ 3 \times 3 $ SkX is the only commensurate non-coplanar state along the diagonal line ( see SM). All the other phases separate into $ N \times 1 $ and $ 1 \times N $ co-planar spiral phase in the YZ plane. There are no stable C phases at $ \alpha=\frac{\pi}{N} n $ with $ n > 1 $. However, as shown below, this kind of co-planar phases at $ n > 1 $ can be stable near the Abelian line $ ( 0 < \alpha < \pi/2, \beta=0 ) $. Taking $ N \rightarrow \infty $ limit, one may approach the $ \alpha=\beta=0 $ Abelian point. It suggests some IC phase near the point. To test this prediction, we first identify a $ U(1) $ family of degenerate classical state which is a FM state within XY plane. Then by performing a LSW on this degenerate manifold, the linear term indeed vanishes, but the spin wave spectrum always become negative. This fact indicates the FM is always unstable, the true ground state should be some IC phases corresponding to $ N \rightarrow \infty $ limit in the FK model. [Co-planar spiral phases at a small $ 0< \beta < \alpha <\pi/2 $ near the Abelian line, pre-empty of the magnon condensation transitions and the meta-stable Y-x phase. ]{} [1. Mapping to the one dimensional classical Frenkel-Kontorova (FK) Model at $ S=\infty $. ]{} Now we try to understand the global phase diagram Fig.1 near the whole Abelian line at the bottom $ 0 < \alpha < \pi/2, \beta =0 $. We will establish the classical phase diagram by mapping its lower half $ \beta < \alpha=\pi/N $ to the Frenkel-Kontorowa (FK) model with $ N \times 1 $ ( stripe ) ansatz. We consider a $ N \times 1 $ spin-orbital structure commensurate with a lattice with $ N \times M $ lattice sites. We will reach the incommensurate limit by taking $ N \rightarrow \infty $ limit. Within a general $ N \times 1 $ ansatz, applying the local spin rotation $ \tilde{\mathbf{S}}_n= R(\hat{x}, 2 \alpha n )\mathbf{S}_n $ in Eqn.\[rhgeneral\] to get rid of the $ R $ matrix along the $ x $ bonds, writing the spin as a classical unit vector in the rotated basis $ \tilde{\mathbf{S}}_n=(\cos\tilde{\eta}_n,\sin\tilde{\xi}_n\sin\tilde{\eta}_n,\cos\tilde{\xi}_n\sin\tilde{\eta}_n ) $, we find that any $ \beta > 0 $ picks up $\tilde{\eta}_n=\pi/2$ ( namely, a coplanar state in $ \tilde{Y}\tilde{Z} $ plane ) and the trial energy per site is $ E_{tri}( N \times 1 )=-\frac{J}{N}\sum_{n=1}^N [\cos(\tilde{\xi}_n-\tilde{\xi}_{n+1})-\sin^2\beta\cos(2\tilde{\xi}_n+ 4 \alpha n )+\cos^2\beta] $ which can be transformed back to the original frame using $ \xi_n=\tilde{\xi}_n+2n\alpha $ ( so the spins remain in a coplanar state in the original $ YZ $ plane shown in Fig.\[allphases\] ). $$\begin{aligned} E_{FK}& = & -\frac{J}{N}\sum_{n=1}^N[ \cos(\xi_{n+1}-\xi_{n}-2\alpha)- \sin^2\beta\cos2 \xi_n \nonumber \\ & + & \cos^2\beta] \label{fk}\end{aligned}$$ One can see that at a small $ \beta $, by using $ \cos(\xi_{n+1}-\xi_{n}-2\alpha) \sim 1- \frac{1}{2}(\xi_{n+1}-\xi_{n}-2\alpha)^2 $, $ E_{tri}( N \times 1 ) $ maps to the 1d Frenkel-Kontorova (FK) Model discussed in [@tom] at a finite size $ N $ with the periodic boundary condition or 2d Pokrovsky-Talapov (PT) which was used to discuss C-IC transition in 2d Bilayer quantum Hall systems [@blqh]. Some insights can be achieved from the FK model at a small $ \beta $. The kinetic term favors $ \xi_{n+1}=\xi_{n}+2\alpha $, while the potential term favors $ \xi_n = \pm \pi/2 $. When $ \alpha=\pi/2 $, this leads to the Y-x state as the exact ground state. However, when $ \alpha=\pi/N, N=3,4,5,\cdots $, frustrations comes in. At a small $ \beta $, the kinetic term dominates over the potential term, so $ \xi_{n+1} \sim \xi_{n}+2\alpha $ still holds approximately as shown for the $ 3 \times 1, 4 \times 1, 5 \times 1 $ spiral state in Fig.2. There are qualitative even-odd differences: The net magnetization in the $ N \times 1 $ unit cell is small, but non-vanishing for odd $ N $, exactly vanishes for even $ N $. Even $ N $ always has a larger stable regime than its previous odd $ N-1 $. There is a always cyclic degeneracy $ N $ for both even and odd $ N $. The $ {\cal T } $ gives a different state for $ N $ odd, but not for even $ N $. So the degeneracy is $ 2 N $ for odd $ N $, just $ N $ for even $ N $. One can check other symmetries operations $ {\cal P}_x, {\cal P}_y, {\cal P}_z $ do not generate new states. [2. The pre-empty of the magnon condensations by first order transitions in the Y-x phase ]{} For any parameter $ \beta < \alpha=\pi/N $, Eq. gives the best estimation of the ground-state energy as $\min\limits_{N\in[1,\infty)}E_{N\times1}$ which can be compared to that of the Y-x state $E_{Y-x}=-2J\sin^2\alpha $. If one finds $\min E_{N\times1}<E_{Y-x}$ for some $ N $, then it means Y-x becomes unstable against some stripe spiral IC phase. Note that even $\min E_{N\times1}$ may not give real ground-state energy, but it does give a upper bound for the ground-state energy of the stripe spiral IC phase whose precise nature is difficult to determine using the $ N \times 1 $ ansatz in a finite size calculation. The first order transition line from the Y-x to some IC phases is drawn in Fig.\[phasedia\]. It matches very precisely the line achieved from the previous works [@classdm1; @classdm2] using classical Monte-Carlo simulations. It also hits the $ \pm 2\pi/3 $ contour line inside the Y-x phase at one corner of the $ 3 \times 3 $ SkX phase which is a multi-critical ( M ) point at $ (\alpha_M,\beta_M)\approx(0.33952\pi,0.31284\pi) $ of several commensurate and In-commensurate phases in Fig.1. So all the C-$C_0$ regime and the C-IC regime with $ 0 < k^{0}_y < 2\pi/3 $ in the Y-x phase are pre-emptied by some spiral IC phases through the first-order transition line. So the Y-x state becomes only a meta-stable state ( just a local minimum in energy landscapes ) between the first-order transition line and the putative 2nd-order condensation boundary of the C-$C_0$ and the C-IC magnons. So hysteresis behaviors are expected in this regime. There is no mirror symmetry anymore beyond the LSW. For example, the C-$C_\pi$ regime at $ \alpha= \beta $ entertains the gap opening process due to the quantum order from disorder phenomena. However, its mirror image which is the C-$C_0$ regime at $ \alpha=\pi/2 - \beta $ in Eqn.\[segments\] sustains no such phenomena. Of course, the condensation boundary at $ 0 < \beta < \beta_M \sim 0.31284\pi $ suffers a small shift due to the higher order terms in Eqn.\[swcubic\]. But it is completely pre-emptied by the first order transition anyway. [3. Co-planar spiral states near $ \alpha=\frac{\pi}{N} n $. ]{} In contrast to the Y-x state near $ \alpha=\pi/2 $, the commensurate phases near $ \alpha=\pi/3,\pi/4,\pi/5,\cdots $ are stripe co-planar ( in the YZ plane ) spiral phases shown in Fig.\[allphases\] instead of a collinear phase. As shown in the last section, we also find stable co-planar spiral phase at $ \alpha=\pi/N, N> 2 $ along the diagonal line. So these phases found near the Abelian line $ \beta \ll 1 $ will extend all the way to the diagonal line $ \alpha=\beta $. However, the co-planar phases for $ n=2,3,\cdots, [N/2] $ will not survive to the diagonal line. It is straightforward to extend our calculations near $ \alpha=\pi/2 $ to near $ \alpha= \frac{\pi}{N} n $ to determine the excitation spectra in these phases which are expected to take the same form as the $ 2 \times 1 $ Y-x state derived in Eq.\[relagap\]: $ E_{-}( \boldmath{q} )= \sqrt{ \Delta^2 + v^2_x q^2_x + v^2_y q^2_y } $. As $ N $ gets bigger, $ \Delta $ and $ v_y $ gets smaller, so the stability regimes ( or the widths of the devil staircases ) in Fig.\[phasedia\] gets smaller. At an IC phase, $ \Delta \rightarrow 0 $ and $ v_y \rightarrow 0 $, it becomes a gapless state which is responsible for its zero width in Fig.\[phasedia\]. We expect it has the anisotropic dispersion $ E( \boldmath{q} )= \sqrt{ v^2_x q^2_x + u^2 q^4_y } $ which may be called an anisotropic phason mode. Interesting, if this form is indeed correct, it takes a similar form as that of the lattice phonon mode in the LSDW+CDW which is one of the Itinerant magnetic phases in a three-dimensional repulsively interacting Fermi gas with Weyl type of spin-orbit coupling [@socsdw]. It is also interesting to determine the classical first order transition boundaries between these robust C phases at $ \alpha=\frac{\pi}{N} $ with some IC phases achieved from the FK model Eq.\[fk\] and then to check if they will pre-empty the second order phase transitions due to the condensations of these magnons. If so, then these spiral phase become meta-stable between the first order transitions and their corresponding second order condensation boundaries just like the collinear Y-x state shown as dashed lines in Fig.1. [Rational and irrational topological winding numbers, In-complete and complete devil’s staircases. ]{} From all the co-planar spiral phases in Fig.\[allphases\], one can define the topological winding number $ W= (\xi_N-\xi_0)/N $. For the C-phase at $ \alpha=\pi/N $, $ W/2\pi=2 \alpha/2\pi=1/N $ is a rational winding number which is dependent of the intermediate values of $ \xi_{n}, n=0,1,.....,N-1 $. For the other C phases at $ \alpha=\frac{\pi}{N} n $ with $ n > 1 $, the winding number is found to be $ W/2\pi=2 \alpha/2\pi=\frac{n}{N} $. The quantum fluctuations will certainly reduce the magnitude of spin at a given site $ \vec{M}_i= \langle \vec{S}_i \rangle $. However, as long as the $ \vec{M}_i \neq 0 $, one can still define the fractional winding number $ W/2\pi $ in terms of its phase. So due to its topological features, the definition of the winding number $ W/2\pi $ also hold in the quantum case. For an In-commensurate phase, one can still define $ W= Lim_{N\rightarrow \infty } (\xi_N-\xi_0)/N $ which becomes an irrational number. Each C phase occupies a step with the length $ \Delta_c $, the total C length $ L_{c}= \sum _{ \{C \} } \Delta_c $, its ratio over the total length $ L_0 $ gives the measure of all the C phases $ L_c/L_0 $. For an in-complete devil staircases, $ L_c/L_0 < 1 $, the rest $ 1-L_c/L_0 >0 $ goes to the measure of the IC phases. For a complete devil staircases, $ L_c/L_0 =1 $, while the IC phases intervening the C phases become a set of measure zero. For a harmless devil staircases, there is a direct first order transition between the two C phases with no intervening IC phases. [1. In-complete devil’s staircases near the Abelian line. ]{} Near the Abelian line $ \beta \ll 1 $, Eq.\[fk\] can be mapped to the FK model in the weak locking regime. As shown in Fig.\[phasedia\], in addition to the $ N \times 1 $ C phase, the C phases at $ \alpha=\frac{\pi}{N} n $ with $ n > 1 $ also contribute to $ L_c= \sum _{ \{C \} } \Delta_c $. The total length $ L_0=1/2- \beta/\pi $. The C measure $ L_c/L_0 < 1 $, the IC measure $ 1-L_c/L_0 > 0 $. We expect that there are two following limiting cases in Fig.\[phasedia\]: As $ \beta \rightarrow 0 $ approaches the Abelian line, $ L_c/L_0 \rightarrow 0, 1- L_c/L_0 \rightarrow 1^{-} $, so the IC phases takes almost all the measures. As $ \beta $ approaches the diagonal line, $ L_c/L_0 \rightarrow 1^{-}, 1- L_c/L_0 \rightarrow 0^{+} $, the C phases takes almost all the measures. At the transition point $ \alpha=\beta $, it just becomes the complete devil staircases where the IC phases take the Cantor set with a fractal dimension. [2. Complete devil’s staircases along the diagonal line. ]{} Near the diagonal line $ \alpha=\beta $, the mapping to the FK model may not be precise anymore. Along the diagonal line, as shown in Fig.\[phasedia\], only the $ N \times 1 $ C phase occupies a step with the length $ \Delta_{ N \times 1 } $, the total C length $ L_{C}= \sum^{\infty}_{N=4} \Delta_{ N \times 1 } $ and the total length $ L_0= \alpha_{4 \times 1} \sim 0.295 \pi $. As shown in the appendix, the $ N \times 1, N \geq 4 $ ( or $ 1 \times N $ ) forms a complete devil staircases instead of a harmless one. In fact, as speculated above, they just lie at the transition point from the in-complete devil staircases below the diagonal line to the complete ones along the diagonal line, so the IC phases form a Cantor set with a non-integer fractal dimension. [Implications on cold atom experiments and materials with SOC. ]{} The orbital ordering wavevectors in all the phases in Fig.1 have been listed in the SM. As said below Eq.\[rhgeneral\], there is no spin-orbital coupled $ U(1) $ symmetry away from the line $ (\alpha=\pi/2, \beta) $, so one need to calculate the $ 3 \times 3 $ tensor $ \langle S_{i}(\vec{k},\omega) S_{j}(-\vec{k},-\omega) \rangle $. Following [@rh], one can work out the thermodynamic quantities such as magnetization, uniform and staggered susceptibilities, specific heat at the low temperatures in all the phases in Fig.\[phasedia\]. Similarly, one can work out various kinds of spin correlation functions at the low and high temperatures. In the cold atom contexts, all these physical quantities can be detected by atom or light Bragg spectroscopies [@lightatom1; @lightatom2], specific heat measurements [@heat1; @heat2] and the [In-Situ ]{} measurements [@dosexp]. In materials, they can be easily measured by magnetic resonant X-ray diffraction or neutron scattering techniques [@kitaevlattice; @kitaevlattice1; @kitaevlattice2]. There have been some remarkable experimental advances to generate various kinds of 2D SOC for cold atoms in both continuum and optical lattices. A 2d Rashba SOC was implemented by Raman scheme in the fermion $ ^{40} K $ gas [@expk40; @expk40zeeman]. Using an optical Raman lattice scheme, Wu [et al]{} [@2dsocbec] realized the tunable quantum Anomalous Hall (QAH) SOC of spinor bosons $ ^{87} Rb $ in a square lattice. More recently, the fermionic optical lattice clock [@clock] scheme was successfully implemented to generate a strong SOC for $ ^{87} Sr $ clock [@clock1], $ ^{173} Yb $ clock [@clock2] and also $ ^{87} Rb $ [@SDRb], where the heating and atom loss from spontaneous emissions are eliminated, the exceptionally long lifetime $ \sim 100 s $ of the excited clock state have been achieved. In parallel, by using the most magnetic fermionic element dysprosium to eliminate the heating due to the spontaneous emission, the authors in [@ben] created a long-lived SOC gas of quantum degenerate atoms. The long lifetime of this weakly interacting SOC degenerate Fermi gas will also facilitate the experimental study of quantum many-body phenomena manifest at longer time scales. These ground-breaking experiments set-up a very promising platform to observe novel many-body phenomena shown in Fig.\[phasedia\] due to interplay between SOC and interaction in optical lattices. Naively, due to its microscopic bosonic nature, the RFHM Eq.\[rhgeneral\] may not be useful to describe the magnetism in so called Kitaev materials such as Iridates or Osmates [@spinliquid]. However, as shown in [@rh], the RFHM can be expanded as Heisenberg- ferromagnetic Kitaev [@kit]- Dzyaloshinskii-Moriya (DM) [@dm1] form with a dominant FM Kitave term which is indeed the case in these materials. Of course, the FM Kitave sign in these materials originates from the Hunds rules instead of the spinor bosonic nature of the underlying microscopic models. In the so called 5d Kitaev materials such $ A_2 Ir O_3 $ with $ A=Na_2, Li_2 $ or more recent 4d materials $ \alpha-Ru Cl_3 $, so far, only Zig-Zag phase or an IC-SkX phase were observed experimentally [@kitaevlattice; @kitaevlattice1; @kitaevlattice2], no spin liquids have been found. For example, an IC-SkX phase with the ordering wavevector $ \vec{q}=(0,0,q), q= \pi + \delta, \delta \sim 0.14 \pi $ lying along the orthorhombic $ \vec{a} $ axis was also detected on 3d hyperhoneycomb iridates $\alpha,\beta,\gamma$-Li$_2$IrO$_3$ by resonant magnetic X-ray diffractions [@kitaevlattice; @kitaevlattice1; @kitaevlattice2]. As shown in [@rh], the RFHM can be written as the Heisenberg- ferromagnetic Kitaev-DM form. One can estimate their separate numerical values near $ \alpha= \alpha^{0}_{in}=\arccos \frac{1}{\sqrt{6}} $ in the IC-SkX with the ordering wavevectors $ ( 0, \pm ( \pi-q^{0}_y ) ) $ and $ ( \pi, \pm ( \pi-q^{0}_y ) ) $ with $ 0.18 \pi < q^{0}_y < \pi/3 $ in the inset of Fig.\[phasedia\]: the Heisenberg interaction $ J^{x}_{H}= J^{y}_{H} \sim \cos 2 \alpha=-2/3 $, so it is an AFM coupling, the Ferromagnetic Kitaev interaction $ J^{x}_{K}= J^{y}_{K}= 2 \sin^{2} \alpha \sim 5/3 $, the DM term $ J^{x}_{D}= J^{y}_{D}= 2 \sin 2\alpha \sim \sqrt{5}/3 $. So the model becomes a dominant FM Kitaev term plus a small AFM Heisenberg term and a small DM term. So the RFHM could be an alternative minimal model to the Heisenberg-Kitaev-Ising $(J,K,I)$ model used in [@kitaevlattice1; @kitaevlattice2] or Heisenberg-Kitaev-Exchange ($J,K,\Gamma$) model used in [@kim; @kim1] to fit the experimental data phenomenologically. One common thing among all the three models is a dominant FM Kitaev term plus a small AFM Heisenberg term. It was known that there is no such IC-SkX phase in the Heisenberg-Kitaev model with only $ ( J, K ) $ term. Various IC-SkX phase have also been observed in some helical magnets with a strong Dzyaloshinskii-Moriya (DM) interaction [@dm1]. Indeed, a 2D skyrmion lattice has been observed between $ h_{c1}=50$ mT and $ h_{c2}=70$ mT in some chiral magnets MnSi or a thin film of Fe$_{0.5}$Co$_{0.5}$Si [@sky4]. [DISCUSSIONS ]{} The order from disorder phenomena at the Rashba limit $ \alpha=\beta $ in the range $ \alpha_{33} \sim 0.34 \pi < \alpha < 0.5 \pi $ is due to the Rashba SOC which is a completely different mechanism than that due to the geometric frustrations [@sachdev; @aue; @bert; @spinliquid]. The collinear spin-orbital coupled magnetic phase, the $ 3 \times 3 $ SkX and the IC-SkX, stripe co-planar C and IC phases in Fig.1 can be contrasted to collinear and co-planar phases in geometrically frustrated magnets [@scaling; @sachdev; @aue; @bert]. It was also speculated strong geometric fluctuations may lead to possible quantum spin liquids [@spinliquid]. In fact, the SOC could be a new mechanism leading to new classes of quantum spin liquids which may have a good chance to be sandwiched by two commensurate magnetic phases. For example, the non-coplanar IC-SkX and co-planar IC phases in Fig.\[phasedia\] are particularly vulnerable to some small parameter changes. So it would be important to study if a spin liquid phase can creep in a honeycomb lattice with three SOC parameters $ ( \alpha, \beta, \gamma ) $ putting on its three bonds. For example, it is tempting to see if the two kinds of IC phases will be fractionized into chiral spin liquids or spin liquids under the third SOC parameter $ \gamma $ in a honeycomb lattice. Indeed, there remains unknown deep connections among the in-commensurability, the hierarchy structures of fractals, spin liquids and their fractionalized excitations. The strong correlations and quenched disorders lead to a new class of state of matter: quantum spin glass [@glass1; @glass2; @glass3; @syk0]. The multiple local ( meta-stable ) states, hysteresis and multi-fractals in Fig.\[phasedia\] may resemble the complex multiple local minimum landscapes in quantum spin glass. However, the former is SOC induced, the latter is due to quenched disorders. So the SOC may induce some similar complex glassy phenomena as the disorders do. Indeed, the non-coplanar IC-SkX phase and the coplanar IC phase in Fig.\[phasedia\] completely breaks the translational symmetry along the $ y $ axis and the $ x $ axis respectively. The possible connections between the quantum fluctuations in the multi-fractals in Fig.1 and the quantum chaos in the quantum spin glass deserve to be explored further [@glass1; @glass2; @glass3; @syk0; @mess3; @kittalk; @syk1; @syk2]. In this paper, we only discussed the spinor bosons. Starting from the results achieved in both weak and strong coupling along the anisotropic line $ (\alpha=\pi/2, \beta) $ in the fermionic case [@rafhm], we will map out the global quantum phase diagram for spin 1/2 fermion case in the generic $ (\alpha, \beta ) $ SOC parameter space in both weak and strong coupling limits. We may also explore the quantum or topological transitions between the two limits. [METHODS ]{} In the main text, by the spin-coherent state path integral, we obtained the gap opening in the Y-x state along the diagonal $\tilde{\Delta}_B=J\sqrt{SAB}=4JS\sqrt{AB/(16S)}=4JS\Delta_B$. Note that in Eqn.3, the unit of spin wave dispersion at the LSW order is in the unit of $ 4JS $. Incorporating the gap $ \Delta_B $ into the spin-wave dispersion $\omega_k$ in Eqn.8 at the LSW order leads to $\Omega_q=\sqrt{\Delta^2_B+\omega_q^2}$ whose evolution is shown in Fig.\[cictransition\] ( because $ B \ll A $ as shown in Fig.3b, so we ignore its small contribution to $ \omega_q $ ). Plugging in the value of the gap $ \Delta_B \sim 0.036 $, we estimate the new quantum critical point is shifted to $\alpha=\alpha_{in}=0.3526\pi$ with the onset orbital order $ q_{ic} \sim 0.18 \pi $ as shown in Fig.\[cictransition\]d. This is a quantum Lifshitz C-IC transition from the Y-x state to the IC-SkX/Y-x with the dynamic exponent $ z=1 $. Because the IC-SkX/Y-x phase has the 4 orbital ordering wave-vectors $ ( 0, \pm ( \pi-q^{0}_y ) ) $ and $ ( \pi, \pm ( \pi-q^{0}_y ) ) $, the most general spin structures for the two transverse components can be written as: $$\begin{aligned} S^x(x,y)= [A + B (-1)^{x} ] e^{i (\pi-q^{0}_y ) y }+ h.c \nonumber \\ S^z(x,y)= [C + D (-1)^{x} ] e^{i (\pi-q^{0}_y ) y }+ h.c \label{skxansatz}\end{aligned}$$ where $ q_{ic} \sim 0.18\pi < q^{0}_y < 0.24 \pi $. The longitudinal component $ S^y(x,y) $ can also be written as: $$S^y(x,y)= [E + F (-1)^{x} ] e^{i (\pi-q^{0}_y ) y }+ h.c \label{skxansatzY}$$ The 6 complex numbers $ A, B, C, D $ and $ E, F $ should satisfy the classical constraint $ \sum^{3}_{\alpha=1} S^{2}_{\alpha}=S^2 $. at any IC momentum $ q^{0}_y $. These parameters can be determined by the minimization of the classical energy of the IC-SkX Eqn.\[skxansatz\] in the range $ \alpha_{33} < \alpha < \alpha_{in} $ along the diagonal line. As shown in the main text, there is also a transition from the Y-x state to the IC-SkX/Y-x from the Y-x state on the right due to the condensations of the C-IC magnons with $ \pi- \pi/3 < k^{0}_y < \pi -0.18 \pi $ ( or equivalently $ 0.18\pi < q^{0}_y < \pi/3 $ ). Eq.\[skxansatz\] can also be written as $ S_{\alpha}= Re [ ( \psi_{1 \alpha} + \psi_{2 \alpha}(-1)^{x} ) e^{i (\pi-q^{0}_y ) y } ] $ where $ \alpha=x,z $ stands for the two transverse components. It is interesting to construct a GL action in terms of the order parameters $ \psi_{a \alpha}, a=1,2, \alpha=x,z $ to describe the quantum Lifshitz transition. Note that although the IC-SkX/Y-x phase in Eq.\[skxansatz\] breaks the crystal translational symmetry along the $ x $ axis only to two sites per unit cell, it completely breaks the crystal translational symmetry along the $ y $ axis. It is infinitely degenerate, but discrete and countable. So its excitation spectrum should still have a gap. Because the crystal momentum $ k_y $ is not a good quantum number anymore, so there maybe dis-commensurations or domain walls along the $ y $ axis. It remains interesting to determine the distributions of these dis-commensurations and their repulsive interactions in the IC-SKX phases. ![ The quantum Lifshitz C-IC transition from the Y-x state to the IC-SkX/Y-x state along the diagonal line $ \alpha=\beta $. The momentum is expanded near $ \vec{k} = (0, \pi ) + \vec{q} $. (a) The transition happens at $\alpha=\alpha_{\rm in}^0=0.3661\pi$ at the LSW order with the dynamic exponent $ ( z_x=1, z_y=2 ) $. (b) Order from disorder mechanism generates a gap $ \Delta_B $ to the spin wave spectrum at $\alpha=\alpha_{\rm in}^0=0.3661\pi$. (c) As $ \alpha $ decreases further, the Y-x state supports the C-IC magnons at $ ( 0, k^{0}_y ) $. (d) The C-IC transition due to the condensations of the C-IC magnons happens at $\alpha=\alpha_{\rm in}=0.3552\pi$ with the onset in-commensurate order $ q_{ic}=\pm (\Delta_B/u)^{1/2} \sim 0.18 \pi $ and the dynamic exponent $ (z_x=1, z_y=1 ) $ as shown in the inset. []{data-label="cictransition"}](addDelta){width="50.00000%"} [SUPPLEMENTARY MATERIALS ]{} Supplementary materials for this article is available at http://advances.sciencemag.org/ [99]{} A. V. Chubukov, S. Sachdev, and J. Ye, Theory of two-dimensional quantum Heisenberg antiferromagnets with a nearly critical ground state, Phys. Rev. B 49, 11919(1994). S. Sachdev, Quantum Phase transitions, (2nd edition, Cambridge University Press, 2011). A. Auerbach, Interacting electrons and quantum magnetism, (Springer Science & Business Media, 1994). B.I. Halperin and W.M. Saslow, Phys. Rev. B 16, 2154 (1977). L. Savary and L. Balents, Quantum Spin liquids, arXiv:1601.03742 (2016). J. Ye, S. Sachdev and N. Read, A solvable spin glass of quantum rotors, Phys. Rev. Lett. 70, 4011 (1993) S. Sachdev and J. Ye, Gapless spin-fluid ground state in a random quantum Heisenberg magnet, Phys. Rev. Lett. 70, 3339 (1993) N. Read, S. Sachdev and J. Ye, Landau theory of quantum spin glasses of rotors and Ising spins, Phys.Rev.B,52, 384 (1995 Yi-Zhuang You, Andreas W. W. Ludwig, Cenke Xu, Sachdev-Ye-Kitaev Model and Thermalization on the Boundary of Many-Body Localized Fermionic Symmetry Protected Topological States, arXiv:1602.06964 Subir Sachdev, Holographic Metals and the Fractionalized Fermi Liquid, Phys. Rev. Lett. 105, 151602 每 Published 4 October 2010. A. Kitaev, A simple model of quantum holography," KITP strings seminar and Entanglement 2015 program (Feb. 12, April 7, and May 27, 2015) . http://online.kitp.ucsb.edu/online/entangled15/. Joseph Polchinski, Vladimir Rosenhaus, The Spectrum in the Sachdev-Ye-Kitaev Model, arXiv:1601.06768. Antal Jevicki, Kenta Suzuki, Junggi Yoon, Bi-Local Holography in the SYK Model, arXiv:1603.06246. Juan Maldacena, Douglas Stanford, Comments on the Sachdev-Ye-Kitaev model, arXiv:1604.07818. Y. A. Bychkov and E.I. Rashba, Oscillatory effects and the magnetic susceptibility of carriers in inversion layers, J. Phys. C 17, 6039 (1984). Jinwu Ye, Yong Baek Kim, A. J. Millis, B. I. Shraiman, P. Majumdar, and Z. Tesanovic Berry phase theory of the Anomalous Hall Effect: Application to Colossal Magnetoresistance Manganites, Phys. Rev. Lett. 83, 3737 (1999) Lev P. Gor’kov and Emmanuel I. Rashba, Superconducting 2D System with Lifted Spin Degeneracy: Mixed Singlet-Triplet State, Phys. Rev. Lett. 87, 037004 (2001). T. Jungwirth, Qian Niu and A. H. MacDonald, Anomalous Hall Effect in Ferromagnetic Semiconductors, Phys. Rev. Lett. 88, 207208 (2004). Jairo Sinova, Dimitrie Culcer, Q. Niu, N. A. Sinitsyn, T. Jungwirth, and A. H. MacDonald, Universal intrinsic spin Hall effect, Phys. Rev. Lett. 92, 126603 (2004). Wang Yao and Qian Niu, Berry Phase Effect on the Exciton Transport and on the Exciton Bose-Einstein Condensate, Phys. Rev. Lett. 101, 106401 (2008). Naoto Nagaosa, Jairo Sinova, Shigeki Onoda, A. H. MacDonald, and N. P. Ong, Anomalous Hall effect, Rev. Mod. Phys. 82, 1539 (2010) - Published 13 May 2010. Jairo Sinova, Sergio O. Valenzuela, J. Wunderlich, C.?H. Back, and T. Jungwirth, Spin Hall effects, Rev. Mod. Phys. 87, 1213 (2015) - Published 27 October 2015 Lianghui Huang, [et.al]{}, Experimental realization of a two-dimensional synthetic spin-orbit coupling in ultracold Fermi gases, arXiv:1506.02861. To appear in Nature Physics. Zengming Meng, [et.al]{}, Experimental observation of topological band gap opening in ultracold Fermi gases with two-dimensional spin-orbit coupling, arXiv:1511.08492. Zhan Wu, [et.al]{}, Realization of Two-Dimensional Spin-orbit Coupling for Bose-Einstein Condensates, arXiv:1511.08170 Michael L. Wall, [et.al]{}, Synthetic Spin-Orbit Coupling in an Optical Lattice Clock, Phys. Rev. Lett. 116, 035301 (2016). L. F. Livi, G. Cappellini, M. Diem, L. Franchi, C. Clivati, M. Frittelli, F. Levi, D. Calonico, J. Catani, M. Inguscio, L. Fallani, Synthetic dimensions and spin-orbit coupling with an optical clock transition, Phys. Rev. Lett. 117, 220401 每 Published 23 November 2016. Editors’ Suggestion. S. Kolkowitz, S.L. Bromley, T. Bothwell, M.L. Wall, G.E. Marti, A.P. Koller, X. Zhang, A.M. Rey, J. Ye, Spin-orbit coupled fermions in an optical lattice clock, arXiv:1608.03854. Fangzhao Alex An, Eric J. Meier, Bryce Gadway, Direct observation of chiral currents and magnetic reflection in atomic flux lattices, arXiv:1609.09467. Nathaniel Q. Burdick, Yijun Tang, and Benjamin L. Lev, Long-Lived Spin-Orbit-Coupled Degenerate Dipolar Fermi Gas, Phys. Rev. X 6, 031022 每 Published 17 August 2016. J. Dalibard, F. Gerbier, G. Juzeli\|[u]{}ūnas, and P. Öhberg, Colloquium: Artificial gauge potentials for neutral atoms, Rev. Mod. Phys. 83, 1523 (2011). Fadi Sun, Jinwu Ye, Wu-Ming Liu, Quantum magnetism of spinor bosons in optical lattices with synthetic non-Abelian gauge fields at zero and finite temperatures, Phys. Rev. A 92, 043609 (2015). Zi Cai, Xiangfa Zhou, and Congjun Wu, Magnetic phases of bosons with synthetic spin-orbit coupling in optical lattices, Phys. Rev. A 85, 061605(R) 每 Published 20 June 2012 J. Radić, A. Di Ciolo, K. Sun, and V. Galitski, Exotic Quantum Spin Models in Spin-Orbit-Coupled Mott Insulators, Phys. Rev. Lett. 109, 085303 (2012). W. S. Cole, S. Zhang, A. Paramekanti, and N. Trivedi, Bose-Hubbard Models with Synthetic Spin-Orbit Coupling: Mott Insulators, Spin Textures, and Superfluidity, Phys. Rev. Lett. 109, 085302 (2012). A. Kitaev, Anyons in an exactly solved model and beyond. In the context of the present paper, the difference between honeycomb and a square lattice is not essential, Ann. Phys. 321, 2 (2006). I. Dzyaloshinskii, J. Phys. Chem. Solids 4, 241 (1958); T. Moriya, Phys. Rev. 120, 91 (1960). Fadi Sun, Jinwu Ye, Wu-Ming Liu, Classification of magnons in Rotated Ferromagnetic Heisenberg model and their competing responses in transverse fields, Phys. Rev. B 94, 024409 ( 2016 ). Shang-Shun Zhang, Jinwu Ye, Wu-Ming Liu, Itinerant magnetic phases and Quantum Lifshitz transitions in repulsively interacting spin-orbit coupled Fermi gas, Phys. Rev. B 94, 115121 (2016). Ye, J. Duality, magnetic space group and their applications to quantum phases and phase transitions on bipartite lattices in several experimental systems. Nucl. Phys. B 805, 418 (2008). Chen, Y. & Ye, J., Characterizing boson orders in lattices by vortex degree of freedoms. Philos. Mag. 92, 4484-4491 (2012). Ye, J. & Chen, Y. Quantum phases, Supersolids and quantum phase transitions of interacting bosons in frustrated lattices. Nucl. Phys. B 869, 242 (2013). Murthy, G., Arovas, D. & Auerbach, A. Superfluids and supersolids on frustrated two-dimensional lattices. Phys. Rev. B 55, 3104 (1997). R. T. Scalettar, G. G. Batrouni, A. P. Kampf, and G. T. Zimanyi, Simultaneous diagonal and off-diagonal order in the Bose-Hubbard Hamiltonian, Phys. Rev. B 51, 8467 (1995). Jun-ichi Igarashi, 1/S expansion for thermodynamic quantities in a two-dimensional Heisenberg antiferromagnet at zero temperature, Phys. Rev. B 46, 10763每10771 (1992); Jun-ichi Igarashi and Tatsuya Nagao, 1決S-expansion study of spin waves in a two-dimensional Heisenberg antiferromagnet, Phys. Rev. B 72, 014403 (2005). For a classical bosonic Lifshitz type of transitions, see: Longhua Jiang and Jinwu Ye, Lattice structures of Larkin-Ovchinnikov-Fulde - Ferrell (LOFF) state, Phys. Rev. B 76, 184104 (2007). For scaling functions with the anisotropic dynamic exponents $ ( z_x=2, z_y=1 ) $ where $ q_x $ is the colliding direction across a fermionic Lifshitz type of transitions, see F. Sun, X.-L. Yu, J. Ye, H. Fan, and W.-M. Liu, Topological Quantum Phase Transition in Synthetic Non-Abelian Gauge Potential: Gauge Invariance and Experimental Detections, Sci. Rep. 3, 2119 (2013). Fadi Sun, Jinwu Ye, Wu-Ming Liu, Hubbard model with Rashba or Dresselhaus spin-orbit coupling and Rotated Anti-ferromagnetic Heisenberg Model, arXiv:1601.01642 J. Ye, J. M. Zhang, W. M. Liu, K. Zhang, Y. Li, and W. Zhang Light-scattering detection of quantum phases of ultracold atoms in optical lattices, Phys. Rev. A 83, 051604 (2011). J. Ye, K. Y. Zhang, Y. Li, Y. Chen, and W. P. Zhang, Optical Bragg, atom Bragg and cavity QED detections of quantum phases and excitation spectra of ultracold atoms in bipartite and frustrated optical lattices, Ann. Phys. 328, 103 (2013). J. Kinast, A. Turlapov, J. E. Thomas, Q. Chen, J. Stajic, and K. Levin, Heat Capacity of a Strongly Interacting Fermi Gas, Science 307, 1296 (2005). M. J. H. Ku, A. T. Sommer, L. W. Cheuk, and M. W. Zwierlein, Revealing the Superfluid Lambda Transition in the Universal Thermodynamics of a Unitary Fermi Gas, Science 335, 563 (2012). N.Gemelke, X. Zhang, C. L. Huang, and C. Chin, In situ observation of incompressible Mott-insulating domains in ultracold atomic gases, Nature (London) 460, 995 (2009). X. Z. Yu, Y. Onose, N. Kanazawa, J. H. Park, J. H. Han, Y. Matsui, N. Nagaosa and Y. Tokura, Real-space observation of a two-dimensional skyrmion crystal, Nature 465,901每904(17 June 2010). A. Biffin, [et.al]{}, Noncoplanar and Counterrotating Incommensurate Magnetic Order Stabilized by Kitaev Interactions in 污Li 2 IrO 3, Phys. Rev. Lett. 113, 197201 A. Biffin, [et. al ]{}, Unconventional magnetic order on the hyperhoneycomb Kitaev lattice in 汕?Li2IrO3: Full solution via magnetic resonant x-ray diffraction, Phys. Rev. B 90, 205116 ( 2014 ) Itamar Kimchi, Radu Coldea, and Ashvin Vishwanath, Unified theory of spiral magnetism in the harmonic-honeycomb iridates 汐,汕 and 污Li 2 IrO 3, Phys. Rev. B 91, 245134 ( 2015 ). Jeffrey G. Rau, Eric Kin-Ho Lee, and Hae-Young Kee, Generic Spin Model for the Honeycomb Iridates beyond the Kitaev Limit, Phys. Rev. Lett. 112, 077204 (2014) Eric Kin-Ho Lee and Yong Baek Kim, Theory of magnetic phase diagrams in hyperhoneycomb and harmonic-honeycomb iridates, Phys. Rev. B 91, 064407 ( 2015 ). P. M. Chaikin and T. C. Lubensky principles of condensed matter physics( Cambridge university press,1995.) For areview on a bilayer quantum Hall systems, see S. M. Girvin and A. H. Macdonald, in Persepctives in Quantum Hall effects, edited by S. Das Sarma and Aron Pinczuk ( Wiley, new York, 1997 ). [Acknowledgements ]{} We thank W. M. Liu for encouragements and AFOSR FA9550-16-1-0412 for supports. The work at KITP was supported by NSF PHY11-25915. [Author Contributions ]{} F. S and J. Y did the calculations. F. S wrote the notes leading to the manuscript. J.Y. designed this work and wrote the manuscript based on the notes. Both reviewed the manuscript. [Competing Interests:]{} The authors declare no competing financial interests. [Correspondence:]{} Correspondence and requests for materials should be addressed to Jinwu Ye at [email protected]	{ "pile_set_name": "ArXiv" }
--- author: - 'L. A. Cobus' - 'B. A. van Tiggelen' - 'A. Derode' - 'J. H. Page[^1]' bibliography: - 'DynamicCBS\_DiffuseBvTtheory\_2016.bib' title: 'Dynamic coherent backscattering of ultrasound in three-dimensional strongly-scattering media' --- Introduction ============ Coherent backscattering (CBS) has for several decades been used to measure transport parameters of disordered media in the diffuse regime. The CBS effect is caused by interference between multiply-scattered waves travelling reciprocal paths inside a disordered medium [@Sheng2006]. Experimentally, this phenomenon may be observed as an enhancement (of around 2) in intensity at exact backscattering. Away from exact backscattering the CBS intensity profile decreases, forming a ‘cone’ shape which contains valuable information about scattering parameters of the medium [@VanAlbada1985; @Wolf1985; @Kuga1985; @Akkermans1988]. As has been observed experimentally for various types of diffuse waves and scattering media [@VanAlbada1985; @Wolf1985; @Kuga1985; @Akkermans1988; @Bayer1993; @Tourin1997; @Jonckheere2000; @Wolf1988], the width of the static (single frequency or time-integrated dynamic) CBS profile is directly related to the transport mean free path, $\l^$. The dynamic (time-dependent) CBS profile provides opportunities to measure additional quantities. In the diffuse regime, the dynamic CBS profile can directly yield a measurement of the Boltzmann diffusion coefficient $D_B$ without the influence of absorption. Most measurements of $D_B$ using dynamic CBS have been performed for acoustic waves in 2D media [@Tourin1997; @Mamou2005; @Aubry2007; @Bayer1993]. However, the first acoustic study of CBS by Bayer et al. in 1993 [@Bayer1993] also investigated 3D samples, in which dynamic CBS from a very thick (effectively semi-infinite) gravel medium was observed. These data were interpreted using theory taken directly from electromagnetics, which includes assumptions which may not be justified for acoustics. Here, we present a microscopic derivation of dynamic CBS for acoustic waves in 3D, in which we also take into account the conditions encountered in experiments with ultrasonic transducer arrays. The theory takes into account the diffuse near field, which would be measured if a detector were placed at a distance from the sample that is comparable to, or less than, the width of the diffuse halo at the surface. In our experiments, the transducer array is placed much farther from the sample surface than this distance, so that the theory is specifically evaluated in the diffuse far field limit that applies to our data. We compare our theory with experimental measurements of time-dependent acoustic CBS from a 3D medium, and show that our approach yields an accurate measurement of the Boltzmann diffusion coefficient. Experiment ========== ![(a) Sample L1 has a thickness of $L=25 \pm 2$ mm and a cross-section of 230$\times$250 mm$^2$. (b) The bead structure of sample L1.[]{data-label="figa"}](Fig1a_rev3.eps){width="0.8\columnwidth"} ![(a) Sample L1 has a thickness of $L=25 \pm 2$ mm and a cross-section of 230$\times$250 mm$^2$. (b) The bead structure of sample L1.[]{data-label="figa"}](Fig1b_rev4.eps){width="0.8\columnwidth"} Backscattered ultrasound was measured from a slab-shaped mesoglass sample composed of aluminum beads brazed together to form a disordered elastic network (Fig. \[figa\]). The bead volume fraction in the sample was $\sim 55$%, and the mean bead diameter was 3.93 mm with a polydispersity of about 20%, which helps to randomize bead positions. The sample has a cross-section of 230$\times$250 mm$^2$ much larger than its thickness $L= 25 \pm 2$ mm, which helps to minimize contributions from the edges of the sample when performing backscattering experiments. Other details of the sample characteristics have been described in Refs. [@Aubry2014; @Cobus2016; @Cobus2016phd]. An ultrasonic array with central frequency $f_c = 1.6$ MHz was used to measure backscattered field from the sample. The experiment was done in a large plexiglass water tank, with sample and array immersed in water, parallel to each other and separated by a distance of $a=182$ mm. Before the experiment was performed, the sample was waterproofed, and for the duration of the experiment the pores between the beads were held under vacuum so that the propagation of ultrasonic waves inside the sample occurs only in the elastic network. Note that in this set-up, compressional and shear waves propagate in the sample, but at the transducers all excitations have been mode-converted back to compressional waves. As a result, the CBS is essentially a scalar phenomenon, despite the vector nature of the waves in the sample. The experimental acquisition process is sketched in Fig. \[fig0\]: a single element emits a short pulse, and then all (64) elements record the time-dependent backscattered field. By repeating this process, emitting with each element in turn, the time-dependent ’response matrix’ was acquired [@Aubry2014; @Tourin1997]. Configurational averaging was performed by translating the array parallel to the sample surface and acquiring the response matrices for 302 different positions. Prior to each experiment, careful checks were carried out to minimize and hopefully eliminate any spurious reflections that could have contributed to the backscattered field from the sample, including the small possible contribution from signals that had travelled through the sample, reflected off either a sample support or a tank wall, and then travelled back through the sample en route to the detector. ![First step of the acquisition sequence for an ultrasonic array of $N$ elements.[]{data-label="fig0"}](fig0.eps){width="0.9\columnwidth"} To study a particular frequency range, the data were filtered using a Gaussian envelope of standard deviation 0.025 MHz, centered in this case around $f=1.65$ MHz. This frequency has been shown to exhibit conventional diffusive behaviour of ultrasound [@Cobus2016phd] (as opposed to subdiffusive or localized behaviour, which has been studied for this sample at other frequencies [@Aubry2014; @Cobus2016]). However, the scattering is still very strong at 1.65 MHz, as is evidenced by the significant contribution to the total backscattered intensity from recurrent scattering processes, which reduce the CBS enhancement below 2 [@Aubry2014]. On average, over all times investigated here (between $\sim 20 - 220$ $\mu$s), recurrent scattering constitutes as much as $37\%$ of the total backscattered intensity, and for the latest times (between $170 - 220$ $\mu$s), the observed recurrent scattering contribution is still more than $15\%$ [@Aubry2014; @Cobus2016phd]. This contribution complicates the analysis of CBS, since it adds to the flat, angle-independent background intensity level (the intensity contribution given by Equation \[fourDI\] in the next section) [@Wiersma1995; @vanTiggelen1995; @Aubry2014]. The recurrent scattering contribution was removed from the total backscattered intensity using the approach developed by Aubry et al. [@Aubry2014]. The result of experiments and data-processing is a large set of configurationally-averaged, time-dependent backscattered intensity profiles $I(\rho,t)$, where $\rho$ is the distance between source and receiver elements of the ultrasonic array, and $t$ is time. To eliminate the effect of absorption, $I(\rho,t)$ was normalized by $I(0,t)$, since at time $t$ the effect of absorption is the same for both numerator and denominator of this ratio, and therefore should cancel [@Hu2008; @Cobus2016]. Theory ====== Here we outline our theory for diffuse, strong scattering of acoustic waves in 3D samples. The geometry of the system is shown in Fig. \[fig1\]. Source S is positioned at $(0,0,-a)$, and detector D at $(0,-\rho,-a+w)$, where $a$ is the perpendicular distance between S and the sample surface, and $w$ describes any additional distance between S and D, in the $z$ direction. In an experiment with an ultrasonic array, $w$ is minimized as much as possible by aligning the array as parallel as possible to the sample surface, but it is still useful to be able to account for any residual misalignment in the theoretical analysis. The first scattering event occurs in the ‘skin layer’ of the sample, at a distance equal on average to the scattering mean free path $\ell_s$. The theory presented here assumes that the scattering length is small, i.e. $a>>\ell_s$, and neglects any phase shift which might occur in the skin layer. Additionally, we assume that the problem is symmetric in $\phi$, and that $k_0a>>1$, where $k_0$ is the wave vector in the water in front of the sample. ![Schematic for the experimental geometry and coordinate systems. The source/detector plane (an ultrasonic array) is approximately parallel to sample surface, although source S and detector D may be separated in the $z$ direction by a small distance $w$ (exaggerated in the diagram for clarity). Inside the sample, scattering is described by the coordinate system shown on the lower right. The origin of this $R,\phi$ system, $(0,0,0)$, is on the ray normal to the emitting transducer array element at S. Vector $\vec{R}$ is in the $xy$ plane, which is parallel to the input/output sample surface.[]{data-label="fig1"}](fig3_sm2.eps){width="0.8\columnwidth"} We start with the general expression for intensity from a scattering medium: \|\_[§]{}\|\^2&& G(\_1) G\^\(\_3)(\_1,\_2,\_3,\_4) G(\_2)G\^\(\_4)d\_1 d\_2 d\_3 d\_4, \[eq1\] where propagation between source and sample is given by ensemble averaged Green’s functions $G(\vec{r}_1)$ and $G(\vec{r}_3)$, and propagation between sample and detector is given by $G(\vec{r}_2)$ and $G(\vec{r}_4)$. In the far-field approximation $\|\vec{R}-\vec{r}_1\|\rightarrow\infty$, and in the diffusive regime where $k\ell_s>>1$ ($k=2\pi f/v_p$ is the wavevector in the medium, $v_p$ is the velocity of the longitudinal waves), the ensemble averaged Green’s function between source and sample in 3D may be approximated as G(S)=- \[GS\], with $\mu_{1,3}=\cos\gamma=a/\sqrt{a^2+R_{1,3}^2}$. To more accurately express the intensity drop due to propagation from the last point in the sample at $(\vec{R}_{2,4},z)$, to detector D at $(\rho,-a+w)$, the ensemble averaged Green’s function between sample and detector is written as G([(\_[2,4]{},z)]{}D)=-e\^[-]{}\ . \[GD\] All scattering inside the medium is described by vertex $\Gamma(\vec{r}_1,\vec{r}_2,\vec{r}_3,\vec{r}_4)$. For the incoherent contribution to backscattered intensity, $\Gamma=F_{3D}(\vec{r}_1,\vec{r}_2)\delta_{1,3}\delta_{2,4}$, where in the ladder approximation (the diffuse regime), the 3D ladder kernel $F_{3D}$ is a solution to the diffusion equation. For a slab-shaped medium with partial reflection of energy at the boundaries, i.e. $F_{3D}=0$ at $z=z_0$, $z=L+z_0$, the 3D ladder kernel is &F\_[3D]{}&(R, z\_1, z\_2, t) = C \_[m=1]{}\^e\^[-m\^2\^2D\_B t/B\^2]{}\ && ,\ & & where $D_B = v_e l^/3$ is the Boltzmann diffusion coefficient in 3D (c.f. Ref. [@Schriemer1997]) and $R=\left\|\vec{R_1}-\vec{R_2}\right\|$. The extrapolation length in 3D is $z_0=(2/3)(1+R_\textrm{refl})/(1-R_\textrm{refl})\ell^$, effective sample thickness is defined as $B\equiv L+2z_0$, and $C$ is a constant. Through the boundary conditions, the transport mean free path $\ell^$ has now been introduced, and is allowed to differ from the scattering mean free path $\ell_s$ (anisotropic scattering). We now evaluate equation \[eq1\] for our particular experimental geometry. The integrations over $z$ are the same as in the theory for 2D [@Tourin1999] with some differences in the denominators, and once evaluated, give for the incoherent background intensity &&\^2(t) = \_0\^dR\_1 R\_1\_0\^dR\_2 R\_2 \_0\^[2]{}d\_1\_0\^[2]{}d\_2\ & & , \[fourDI\] where the $L_m$ terms are [@Tourin1999] L\_m(\_[1]{}, \_2)= \[Lm\] with $a_\textrm{1,2} = 1/l_s\mu_\textrm{1,2}$, $b_m = \pi m/B$, A\_[1,2]{}(m) & = & (b\_m z\_0)a\_[1,2]{}\ B\_[1,2]{}(m) & = & (b\_m z\_0)b\_m. To calculate the backscattered intensity due to the CBS effect (the [coherent]{} contribution), the scattering vertex is $\Gamma=F_{3D}(\vec{r}_1,\vec{r}_2)\delta_{2,3}\delta_{1,4}$. This means that contributions from interferences between pairs of reciprocal paths through the sample are taken into account. The expression for the coherent contribution to backscattered intensity thus includes an extra phase factor compared to the incoherent case: &&\^2(t) = \_0\^dR\_1 R\_1\_0\^dR\_2 R\_2 \_0\^[2]{}d\_1\_0\^[2]{}d\_2\ & &\ & & , \[fourDC\] for $w<<\sqrt{a^2+R_{1,2}^2}$ and $\rho<<\sqrt{a^2+R_{1,2}^2}$, and where $s_{1,2}=\sin{\gamma}=\sqrt{1-{\mu_{1,2}}^2}$. To partially evaluate the integral, we apply the diffuse far-field assumption $a^2>>4D_B t$. This corresponds physically to a diffuse halo on the sample that is much smaller than the distance between the sample and transducer. For strong scattering, it is important to assess whether or not the experimental data obey this assumption. In our experiment, the longest times collected are around 220 $\mu$s, and for a diffusion coefficient $D_B$ of order $0.7$ mm$^2/\mu$s (see the next section), the assumption $a^2=(182$ mm$)^2\approx 33000$ mm$^2>>4D_B t\approx4(0.7$ mm$^2/\mu$s$)($220 $\mu$s$)\approx 600$ mm$^2$ holds. In our experiment, the array misalignment $w$ was too small to be measurable, and thus is set to zero from now on. Because of the exponential factor in the integral, the dominant contributions in the diffuse far field approximation come from the points $\vec{R_1}$ and $\vec{R_2}= \vec{R_1} + \Delta \vec{R}$ separated by small distances relative to $a$, so that it is convenient to change coordinates from $\vec{R_1}$ and $\vec{R_2}$ to $\vec{R}$ and $\Delta\vec{R}$. Then, Eq. 7 can be simplified by expanding the argument of the cosine factor to first order in $\Delta\vec{R}$ and integrating over $\Delta\vec{R}$. The coherent backscattered intensity then simplifies to &\^2&(t) \_0\^dR R \_0\^[2]{}d\ & & , \[twoDCnear\] where $\mu=a/\sqrt{a^2+R^2}$ [@wzero]. In the diffuse coda, $t > \tau_{D} = B^2/\pi^2D_B$ ($\tau_D$ is the diffusion time), only the $m=1$ term in Eq. \[Lm\] survives, simplifying the calculation. This $m=1$ term can be further simplified if the optically thick slab approximation, $B>>\ell^$, applies. In our slab, the smallest possible value of $B=L+2z_0=L+2(2/3)l^(1+R_\textrm{refl})/(1-R_\textrm{refl})$ is $B=25 \textrm{ mm}+2(2/3)(4 \textrm{ mm})(1+0.65)/(1-0.65)\approx50$ mm. Thus, the approximation of $B>>\ell^$ is obeyed since $B=50\textrm{ mm}>>4\textrm{ mm}$. However, in our experimental situation we measure up to 220 $\mu$s, so most times considered are smaller than the smallest possible $\tau_{D} \approx (50 \textrm{ mm})^2/\pi^2(0.7 \textrm{ mm}^2/\mu\textrm{s}) \approx$ 360 $\mu\s$. Thus, the diffuse coda approximation does not hold, and all of the terms in the $L_m$ series of Eq. \[Lm\] are included in our calculations. As shown by Eq. \[twoDCnear\], the shape of the CBS dynamic cone is determined by the dimensionless parameter $k_0\rho(D_B t/a^2)$. At each time $t$ the CBS intensity profile has an almost Gaussian shape, with a width $(k\rho)_\textrm{FWHM}$ that depends on time as (k\_0)\_= a\^2/D\_B t, \[k0rho\] where $\Gamma$ is a dimensionless constant. This coherent contribution adds to the uniform incoherent background given by Eq. \[fourDI\]. At $\rho=0$, the coherent and incoherent contributions to total backscattered intensity should be equal. For the sake of completeness, we note that the static CBS intensity profile may be found (in the absence of absorption) simply by integrating Eqs. \[fourDI\] and \[twoDCnear\] over all time. Assuming $B>>\ell^$, the ratio of coherent to incoherent intensity can be expressed analytically as & & \_0\^1 d\_0\^[2]{} , \[twoDstat\] where Q = k\_0. As has been found for other systems, the width of the static profile depends inversely on $\ell^$ [@vanderMark1988; @Akkermans1988; @Tourin1997].\ For technical reasons, experimental results for the static CBS intensity profile are not included in this work. This is mainly due to the fact that at early times the signal is dominated by large specular reflections which could not be completely eliminated with the recurrent scattering filter [@Aubry2014]. In addition, we do not have data at sufficiently late times to accurately calculate the static cone; due to the strong scattering nature of the sample, the dynamic backscattered ‘cones’ do not narrow very quickly, but the range of accessible times in the measurements is limited to times before the arrival of the next echo between sample surface and transducer (after a time interval of $2a/v_\textrm{water}$ $\approx 240$ $ \mu$s). These considerations demonstrate the advantages of our dynamic CBS measurements, which do not require data for all times but can yield accurate results as long as the [range]{} of times experimentally available is sufficient to demonstrate the dynamics of the CBS profiles. Extensions of theory to account for experimental conditions =========================================================== Several modifications to the above theory were made for a more accurate comparison between theory and experiment. The theory so far assumes point sources and detectors, whereas experimentally the source and detector have a finite rectangular shape, with width $W=0.25$ mm (in the $x$ direction of Fig. \[fig1\]) much less than height $H=12$ mm ($y$ direction of Fig. \[fig1\]). This means that the directivity (directional dependence) of each element should be taken into account, especially in the $x$ direction where the spreading of waves due to diffraction from the narrow elements is not insignificant. Here we estimate the directivity along $x$ using the ideal profile for a rectangular transducer of width $W$ in the far field: () = (0)() The same expression can be used in the $y$ direction, so that for an array in 3D, the total angular sensitivity is (\_x,\_y)=\^2()\^2() The correction is incorporated in Eq. \[twoDCnear\] as: &&\^2(t) = \_0\^dR R \_0\^[2]{}d (\_[x,S]{},\_y)(\_[x,D]{},\_y)\ & &\ & & , \[twoDCnear\_dircorr\] where $\theta_{x,S}$, $\theta_{x,D}$ and $\theta_y$ are found from $$\begin{array}{ccccc} \tan(\theta_{x,S}) & = & \frac{R\cos\phi}{a}\nn\\ \tan(\theta_{x,D}) & = & \frac{R\cos\phi - \xi}{a}\nn\\ \tan(\theta_y) & = & \frac{R\sin\phi}{a}. \nn \end{array}$$ An additional correction was performed to account for the height of the array elements, since the height of $H=12$ mm means that signal is being collected over a significantly greater area than was supposed by our theory. The effect is not so large that interference cancellation of ultrasonic field at the array surface is important, so an integration over detected intensity is sufficient to account for the influence of element height. The correction consists of performing an explicit (numerical) integration of the intensity distribution $\|\psi(\rho,t)\|^2$ over all possible source points ($y_1$) and receiver points ($y_2$), by calculating an effective $\rho$ for each pair of points: \|(,t)\|\^2\_ &=& \_[-H/2]{}\^[H/2]{}dy\_1\_[-H/2]{}\^[H/2]{}dy\_2 \|(,t)\|\^2. The same procedure (with the replacement of $H$ by $W$) is performed to account for the finite width of the array elements. Overall, the geometrical corrections presented in this section do not change the global trend of Eq. \[k0rho\], but do change the multiplicative factor $\Gamma$. Fitting and Results =================== ![image](fig2_rev1.eps){width="\textwidth"} Representative experimental CBS profiles are shown in Fig. \[fig2\] (symbols). In principle, after the removal of the recurrent scattering contribution, the incoherent background intensity level should be at 0.5. Our experimental data deviate from this value slightly at some times, with the deviations being especially small at late times. This may be caused by an inaccuracy in the recurrent scattering filter, especially at early times where the initial specular reflection is very large and difficult to remove entirely [@Aubry2014]. Additionally, our theory shows that the enhancement factor may be slightly changed due to the finite size of the array elements. To measure the diffusion coefficient $D_B$, the experimental CBS profiles $I_{\textrm{exp}}(\rho,t_\textrm{exp})$ were fit with the predictions of the diffusion theory outlined in the previous section. The theory calculations require several parameters as input, including scattering mean free path $\ell_s$ and reflection coefficient $R_\textrm{refl}$. From measurements of the coherent ballistic pulse in transmission [@Page1997], we can determine $\ell_s \simeq 1.1$ mm. We also measure the longitudinal phase velocity $v_L\simeq2.8$ mm$/\mu$s [@Cobus2016phd] inside the sample, which is required to calculate $R_\textrm{refl}$. This calculation, based on methods developed by [@Zhu1991; @Page1995; @Ryzhik1996; @Turner1995], assumes that after a few scattering events, there is equipartition of energy between all polarizations of waves inside the sample, allowing equipartition to be taken into account when determining $R_\textrm{refl}$. In this calculation, the phase velocity of the dominant shear (transverse) waves inside the sample, $v_T$, is estimated to be $v_T\sim v_L/2\approx 1.4$ mm/$\mu$s. Because shear waves dominate inside the sample, but longitudinal waves are detected outside the sample, the average reflection coefficient is large, $R_\textrm{refl} \approx 0.75$. Theoretical backscattering profiles $I_{\textrm{theory}}(\rho,D_Bt)$ were calculated as a function of parameter $D_Bt$ (diffusion coefficient multiplied by time) using Equations \[fourDI\], \[twoDCnear\], and \[twoDCnear\_dircorr\]. Then, all experimental CBS profiles were compared to all theoretical profiles, i.e. for each time $t_\textrm{exp}$, $I_{\textrm{exp}}(\rho,t_\textrm{exp})$ was fit with each theoretical CBS profile $I_{\textrm{theory}}(\rho,D_Bt)$. In this way, a best-fit value of $D_B t$ is determined for each $t_\textrm{exp}$. Figure \[fig2\] shows representative experimental CBS profiles for four different times $t_\textrm{exp}$, along with the best theoretical fits. Figure \[fig3\] shows the $D_B t$ values resulting from each best fit versus $t_\textrm{exp}$. The slope of a linear fit to these points gives a direct estimate of $D_B$. The data are well-described by the linear fit, confirming the prediction that $\Delta\rho^{-2}\propto D_B t$. ![Results of the method used to determine the diffusion coefficient. Shown here are the $D_B t_\textrm{theory}$ values obtained from the best fit of the theory to the rho-dependent experimental CBS profiles at each time (symbols). The slope of a weighted linear fit to the data (dashed line) gives the overall value of $D_B$. Notable parameters used to calculate the theory in this plot are $\ell_s = 1.1$ mm, $\ell^ = 4$ mm, and $R_{\textrm{refl}}=0.75$ mm. From this analysis of the data, we measure a value of $D_B=0.74\pm 0.03$ mm$^2/\mu$s. Error bars on each data point are determined from the goodness of fit of $I_{\textrm{theory}}(\rho,D_Bt)$ to $I_{\textrm{exp}}(\rho,t_\textrm{exp})$ (see text).[]{data-label="fig3"}](DBtvst_new.eps){width="0.8\columnwidth"} Since there is some uncertainty in the values of input parameters $R_{\textrm{refl}}$ and $\ell^$ (c.f. [@Cobus2016; @Cobus2016phd]), fitting was performed over the range of physically reasonable values for these parameters: $\ell^=4-8$ mm and $R_{\textrm{refl}}=0.65-0.85$. It was found that over these ranges, the best-fit value of $D_B$ only varies by less than $0.01$ mm$^2/\mu$s, which is less than the fitting uncertainty in the determination of $D_B$ for any pair of these parameters ($\pm0.02$). This shows that the measurement of $D_B$ by dynamic CBS is largely insensitive to the precise values of $\ell^$ and $R_{\textrm{refl}}$, which only enter into the theory via the boundary conditions. Thus, our uncertainty in these parameters is unimportant for determining $D_B$ reliably, and the fitting of the CBS data with the theory presented here gives an accurate and precise measurement of $D_B=0.74\pm 0.03$ mm$^2/\mu$s. This value agrees with results from the measurement on the same sample of the transverse spread of the transmitted intensity, which gives $D_B = 0.71 \pm 0.02$ mm$^2/\mu$s [@Cobus2016phd]. It is worth noting that the transmitted transverse width $=\sqrt{4Dt}$ is known to be independent of absorption and boundary conditions [@Page1995; @Hu2008], so that the excellent agreement between these reflection and transmission methods further confirms the accuracy of our present analysis of the dynamic CBS profiles. The value of $D_B$ seems small if we compare the measured value with a very rough calculation using estimates of the equipartitioned velocity ($1.6$ mm$^2/\mu$s, which is close to the shear wave velocity) and transport mean free path (4 mm); this would give $D_B\sim 2$. Given the plausible range of $\ell^$ values, which are supported by independent transmission experiments, we infer that the energy velocity $v_E$ itself must be very small; we find from our analysis that $v_E$ ($=3 D_B/\ell^$) is between $0.2$ and $ 0.6$ mm/$\mu$s. Such values are much smaller (around $2.5-7$ times smaller) than either the shear or equipartitioned velocities of elastic waves in the sample (and $5-15$ times smaller than the longitudinal velocity), directly indicating how very slow the transport of energy by diffuse waves is in this strongly scattering sample. This result is in striking contrast with the surprising large values of $v_E$ (approximately $3 - 5$ times larger* than the velocity of longitudinal waves) previously deduced from the analysis of data for a similar sample in the localization regime [@Hu2008]. This suggests that the large values of transport velocities observed previously are associated with Anderson localization. Our results for $D_B$ demonstrate that careful consideration of sample and experimental details is necessary to obtain an accurate measurement of $D_B$. Without taking into account details about the geometry of the experiment, the theoretical prediction for the diffusion coefficient is given simply by $\Delta\rho^{-2} =k^2 D_B t$ (see the previous section, and [@Bayer1993; @Tourin1997]). If this relation were used to measure $D_B$ from our experimental data, we would obtain $D_B = 0.5$ mm$^2\mu$s, which differs by 30% from the actual value. Conclusions =========== In this work, we have studied the dynamic coherent backscattering of ultrasound from a 3D medium in the diffusion regime. We used an ultrasonic transducer array to measure the backscattered ultrasound from a slab-shaped ‘mesoglass’, enabling us to perform a very substantial amount of configurational averaging and to use a sophisticated filtering technique to remove, for times $>20$ $\mu$s, other contributions from specular reflections, single scattering and recurrent scattering. As a result, accurate measurements of the dynamic CBS intensity profiles were obtained as a function of time and space/angle. We have described a microsopic derivation of the theory of CBS for acoustic waves in 3D, which provides an excellent description of the experimental data. The fitting of our CBS data with theory enables an absorption-free measurement of the diffusion coefficient of ultrasound in our sample, which was determined to be $D_B=0.74\pm 0.03$ mm$^2/\mu$s and which agrees with results from separate transmission experiments. Roger Maynard was an inspiration in our community. We are all deeply grateful to Roger for so many pleasant moments and so many enlightening discussions, as well as for his enthusiastic support and encouragement over many years. In the context of this paper, J.H.P. would also like to thank Roger for introducing him to the first observations of coherent backscattering of acoustic waves in 1993, thereby initiating J.H.P.’s interest in studying this effect with ultrasound and motivating the current work. We would like to thank Matthew Hasselfield for his contributions to the code used to calculate the theoretical predictions for the CBS profiles, and for the guidance provided by his previous analysis of preliminary data on different samples. We would also like to thank Victor Mamou for carrying out similar experiments and first analyses on 2D synthetic samples. J.H.P. and L.A.C. acknowledge the support of NSERC (Discovery Grant RGPIN/9037-2001, Canada Government Scholarship, and Michael Smith Foreign Study Supplement), the Canada Foundation for Innovation and the Manitoba Research and Innovation Fund (CFI/MRIF, LOF Project 23523). A.D. benefited from funding by LABEXWIFI (Laboratory of Excellence ANR-10-LABX-24), within the French Program Investments for the Future under Reference No. ANR-10-IDEX-0001-02 PSL\*. B.v.T. and J.H.P. are also grateful for support from the PICS program of the CNRS (project Ultra-ALT) and the Agence Nationale de la Recherche (grant ANR-14-CE26-0032 LOVE). [^1]:	{ "pile_set_name": "ArXiv" }
--- abstract: 'We investigate modules for which vanishing of Tor-modules implies finiteness of homological dimensions (e.g., projective dimension and G-dimension). In particular, we answer a question of O. Celikbas and Sather-Wagstaff about ascent properties of such modules over residually algebraic flat local ring homomorphisms. To accomplish this, we consider ascent and descent properties over local ring homomorphisms of finite flat dimension, and for flat extensions of finite dimensional differential graded algebras.' address: 'School of Mathematical and Statistical Sciences, Clemson University, O-110 Martin Hall, Box 340975, Clemson, S.C. 29634 USA' author: - 'Sean K. Sather-Wagstaff' title: Ascent Properties for Test Modules --- Introduction {#sec190908a} ============ Throughout, $(R,{{\mathfrak{m}}},k){\xrightarrow}{\varphi}(S,{{\mathfrak{n}}},l)$ is a local homomorphism between (commutative noetherian) local rings. A classical result in homological commutative algebra says that for each finitely generated $R$-module $N$, if ${\operatorname{Tor}^{R}_{i}(k,N)}=0$ for all $i\gg 0$, then ${\operatorname{pd}}_R(N)<\infty$. In the language of [@celikbas:tgp], this says that the residue field $k$ is a “${\operatorname{pd}}$-test $R$-module”: a finitely generated $R$-module $M$ is a ${\operatorname{pd}}$-test module over $R$ provided that for each finitely generated $R$-module $N$, if ${\operatorname{Tor}^{R}_{i}(M,N)}=0$ for all $i\gg 0$, then ${\operatorname{pd}}_R(N)<\infty$. A key point of our work with O. Celikbas [@celikbas:tgp] is to prove that if $M$ is ${\operatorname{pd}}$-test over $R$, then the ${{\mathfrak{m}}}$-adic completion ${\widehat{M}}$ is ${\operatorname{pd}}$-test over ${\widehat{R}}$. (The converse is relatively straightforward to prove.) This is accomplished by considering the more general notion of ${\operatorname{pd}}$-test complexes. This level of generality allows us to overcome certain technical difficulties involved in considering this ascent problem. However, it did not answer the following. \[q191102a\] If ${\varphi}$ is flat with regular closed fibre $S/{{\mathfrak{m}}}S$ and $M$ is a ${\operatorname{pd}}$-test $R$-module, must ${S\otimes_{R}M}$ be ${\operatorname{pd}}$-test over $S$? It is straightforward to show that the assumption on the closed fibre in this question is necessary: if $R=k$ and $S=k[X]/\langle X^2\rangle$, then $k$ is ${\operatorname{pd}}$-test over $k$, but $S={S\otimes_{k}k}$ is not ${\operatorname{pd}}$-test over $S$ since $S$ is not regular. In the current paper, we answer Question \[q191102a\] in the case where the induced residue field extension $k\to l$ is algebraic; see Theorem \[thm140915a\]. Moreover, the proof shows that the general case reduces to the special case of a purely transcendental field extension with finite transcendence degree. Theorem \[thm140915a\] (module case). Assume that ${\varphi}$ is flat with regular closed fibre, and let $M$ be a finitely generated $R$-module. Assume the induced field extension $k \to l$ is algebraic. Then $M$ is a ${\operatorname{pd}}$-test module over $R$ if and only if ${S\otimes_{R}M}$ is a ${\operatorname{pd}}$-test module over $S$. Our proof of Theorem \[thm140915a\] uses two more significant expansions of our original context. First, we consider the more general setting of local ring homomorphisms of finite flat dimension. This perspective was pioneered by Avramov and Foxby in a sequence of articles [@avramov:glh; @avramov:lgh; @avramov:glp; @avramov:cmporh; @avramov:dalp; @avramov:solh; @avramov:bsolrhoffd] where they used this notion to answer open questions about flat local ring homomorphisms. Section \[sec150429a\] consists of foundational material about test modules and homomorphisms of finite flat dimension. Second, we expand our perspective from modules and complexes over local rings to the setting of differential graded (DG) modules over commutative DG algebras. This point of view was developed by Avramov and his collaborators; see, e.g., [@avramov:ifr; @avramov:cslrec3; @avramov:glh; @avramov:lgh; @avramov:dgha; @avramov:dalp; @avramov:bsolrhoffd; @avramov:tlg; @avramov:holh; @avramov:htecdga; @avramov:phcnr; @avramov:psmlrsec]. It provides a construction whereby if one can solve a homological commutative algebra problem for finite dimensional algebras over a field, then one can sometimes solve the general problem by passing to an associated finite dimensional DG algebra over a field where one can solve a related problem. We have used this approach in several projects [@altmann:csc; @beck:sgidgca; @christensen:dvke; @nasseh:ldgm; @nasseh:egdgm; @nasseh:gart]. For the current setting, we develop some technology around ${\operatorname{pd}}$-test DG modules over finite dimensional DG algebras in Section \[sec190908c\] and then use it to prove Theorem \[thm140915a\]. To be clear, we consider these generalizations because we do not know how to prove Theorem \[thm140915a\] without them. In addition, our techniques allow us to answer similar ascent questions for objects that test for finiteness of G-dimensions. This is the subject of Section \[sec190913a\]. In addition, we include some necessary background material in Section \[sec140831\]. Derived Categories and Semidualizing Complexes {#sec140831} ============================================== Throughout this paper we work in the derived category ${{\mathcal{D}}}(R)$ whose objects are the chain complexes of $R$-modules, i.e., the $R$-complexes $$X=\qquad\cdots{\xrightarrow}{\partial^X_{i+1}} X_i{\xrightarrow}{\partial^X_{i}} X_{i-1}{\xrightarrow}{\partial^X_{i-1}}\cdots.$$ References for this include [@christensen:dcmca; @hartshorne:rad; @verdier:cd; @verdier:1]. We denote by ${\mathbf{R}\!\operatorname{Hom}_{R}(X,Y)}$ and ${X\otimes^{\mathbf{L}}_{R}Y}$ the derived Hom-complex and derived tensor product of two $R$-complexes $X$ and $Y$. Isomorphisms in ${{\mathcal{D}}}(R)$ are identified by the symbol $\simeq$. A complex $X\in{{\mathcal{D}}}(R)$ has finite projective dimension, signified ${\operatorname{pd}}_R(X)<\infty$, if it is isomorphic to a bounded complex of projective $R$-modules. Finiteness of flat dimension and injective dimension are signified and defined similarly. The ring homomorphism $R{\xrightarrow}{\varphi}S$ has finite flat dimension provided that ${\operatorname{fd}}_R(S)<\infty$. The subcategory of ${{\mathcal{D}}}(R)$ consisting of homologically bounded $R$-complexes (i.e., complexes $X$ such that ${\operatorname{H}}_i(X)=0$ for $\|i\|\gg 0$) is denoted ${{{\mathcal{D}}}_{\text{b}}}(R)$. The subcategory of ${{\mathcal{D}}}(R)$ consisting of homologically finite $R$-complexes (i.e., complexes $X$ such that ${\operatorname{H}}(X):=\oplus_{i\in{\mathbb{Z}}}{\operatorname{H}}_i(X)$ is finitely generated) is denoted ${{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. A complex $C\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$ is semidualzing if the natural homothety morphism $R\to{\mathbf{R}\!\operatorname{Hom}_{R}(C,C)}$ in ${{\mathcal{D}}}(R)$ is an isomorphism. Consequently, an $R$-module is semidualizing if and only if ${\operatorname{Hom}_{R}(C,C)}\cong R$ and ${\operatorname{Ext}_{R}^{i}(C,C)}=0$ for $i{\geqslant}1$. In particular, $R$ is a semidualizing $R$-module. A dualizing $R$-complex is a semidualizing $R$-complex of finite injective dimension. Dualizing complexes were introduced in [@hartshorne:rad]. The more general semidualizing complexes were defined in [@christensen:scatac], based on special cases in [@avramov:rhafgd; @foxby:gmarm; @golod:gdagpi; @vasconcelos:dtmc]. \[ch150423c1\] If $R$ is a homomorphic image of a local Gorenstein ring $Q$, then $R$ has a dualizing complex, by [@hartshorne:rad §V.10]. (The converse holds by [@kawasaki:mns Corollary 6.2].) In particular, the Cohen Structure Theorem shows that the completion ${\widehat{R}}$ has a dualizing complex. When $R$ has a dualizing complex $D$, and $C$ is a semidualizing $R$-complex, the dual ${\mathbf{R}\!\operatorname{Hom}_{R}(C,D)}$ is also semidualizing over $R$, by [@christensen:scatac (2.12) Corollary]. \[ch150423c2\] Assume that ${\varphi}$ has finite flat dimension, and let $C$ be a semidualizing $R$-complex. Then the $S$-complex ${S\otimes^{\mathbf{L}}_{R}C}$ is semidualizing, by [@christensen:scatac (5.10) Theorem]. If ${\varphi}$ is flat and the closed fibre $S/{{\mathfrak{m}}}S$ is Gorenstein and $R$ has a dualizing complex $D^R$, then $D^S:={S\otimes^{\mathbf{L}}_{R}D^R}$ is dualizing for $S$ by [@avramov:lgh (5.1) Theorem]. In order to describe another class of ring homomorphisms where dualizing complexes base change to dualizing complexes, we need the following. A Cohen factorization of ${\varphi}$ is a decomposition of ${\varphi}$ as a composition of local ring homomorphisms $R{\xrightarrow}{\dot{\varphi}}R'{\xrightarrow}{{\varphi}'}S$ such that $\dot{\varphi}$ is flat with regular closed fibre, $R'$ is complete, and ${\varphi}'$ is surjective. Such a decomposition exists if and only if $S$ is complete by [@avramov:solh (1.1) Theorem]. In particular, the semicompletion of ${\varphi}$, which is the composition $\grave{\varphi}$ of ${\varphi}$ with the natural map from $S$ to its completion ${\widehat{S}}$, has a Cohen factorization. One idea with Cohen factorizations is that it allows one to transfer questions and properties for $S$ as an $R$-module to related questions and properties for ${\widehat{S}}$ has an $R'$-module. This is valuable because ${\widehat{S}}$ is finitely generated over $R'$ while $S$ often is not finitely generated over $R$. An example of this transfer is found in [@avramov:solh (3.2) Lemma] which implies that ${\operatorname{fd}}({\varphi})<\infty$ if and only if ${\operatorname{pd}}_{R'}({\widehat{S}})<\infty$. Recall that an ideal $I$ of $R$ is Gorenstein if $g={\operatorname{pd}}_R(R/I)<\infty$ satisfies ${\operatorname{Ext}_{R}^{i}(R/I,R)}=0$ for all $i\neq g$ and ${\operatorname{Ext}_{R}^{g}(R/I,R)}\cong R/I$. Our local ring homomorphism ${\varphi}$ is Gorenstein provided that in some (equivalently, every) Cohen factorization $R{\xrightarrow}{\dot{\varphi}}R'{\xrightarrow}{{\varphi}'}{\widehat{S}}$ of $\grave{\varphi}$ the ideal ${{\operatorname{Ker}}}({\varphi}')$ of $R'$ is Gorenstein; see [@avramov:glh; @avramov:lgh]. \[para190908b\] If ${\varphi}$ is flat, then ${\varphi}$ is Gorenstein if and only if its closed fibre is Gorenstein. In general, if ${\varphi}$ is Gorenstein and $R$ has a dualizing complex $D^R$, then $D^S:={S\otimes^{\mathbf{L}}_{R}D^R}$ is dualizing for $S$ by [@avramov:lgh (5.1) Theorem]. As the name suggests, semidualizing complexes are built for duality. Let $C$ be a semidualizing $R$-complex and $X\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. We say that $X$ is derived $C$-reflexive and write ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ when ${\mathbf{R}\!\operatorname{Hom}_{R}(X,C)}\in{{{\mathcal{D}}}_{\text{b}}}(R)$ and the natural morphism $X\to{\mathbf{R}\!\operatorname{Hom}_{R}({\mathbf{R}\!\operatorname{Hom}_{R}(X,C)},C)}$ in ${{\mathcal{D}}}(R)$ is an isomorphism. In the case $C=R$, we write ${\mathrm{G}\text{-}\!\dim}_R(X)<\infty$ instead of ${\mathrm{G}_{R}\text{-}\!\dim}_R(X)<\infty$. In the special case where $C$ is dualizing, this is from [@hartshorne:rad], while the case $C=R$ comes from [@auslander:adgeteac; @auslander:smt; @yassemi:gd]. The general situation is in [@christensen:scatac]. \[ch150423a1\] A semidualizing complex $C$ is dualizing if and only if every $R$-complex in ${{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$ is derived $C$-reflexive, by [@christensen:scatac (8.4) Proposition]. In particular, $R$ is Gorenstein if and only if every $R$-complex $X\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$ has ${\mathrm{G}\text{-}\!\dim}_R(X)<\infty$. \[ch150423a3\] Assume that ${\operatorname{fd}}({\varphi})<\infty$. If $X\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, then ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_S({S\otimes^{\mathbf{L}}_{R}X})<\infty$, by [@christensen:scatac (5.10) Theorem]; Fact \[ch150423c2\] shows that this is reasonable. Auslander and Bass classes, defined next, arrived in special cases in [@avramov:rhafgd; @foxby:gmarm], again with the general case described in [@christensen:scatac]. Let $C$ be a semidualizing $R$-complex. The Auslander class ${{\mathcal{A}}_C}(R)$ consists of the $R$-complexes $X\in{{{\mathcal{D}}}_{\text{b}}}(R)$ such that ${C\otimes^{\mathbf{L}}_{R}X}\in{{{\mathcal{D}}}_{\text{b}}}(R)$ and the natural morphism $\gamma^C_X\colon X\to{\mathbf{R}\!\operatorname{Hom}_{R}(C,{C\otimes^{\mathbf{L}}_{R}X})}$ in ${{\mathcal{D}}}(R)$ is an isomorphism. The Bass class ${{\mathcal{B}}_C}(R)$ consists of all the $R$-complexes $X\in{{{\mathcal{D}}}_{\text{b}}}(R)$ such that ${\mathbf{R}\!\operatorname{Hom}_{R}(C,X)}\in{{{\mathcal{D}}}_{\text{b}}}(R)$ and such that the natural evaluation morphism $\xi^C_X\colon{C\otimes^{\mathbf{L}}_{R}{\mathbf{R}\!\operatorname{Hom}_{R}(C,X)}}\to X$ in ${{\mathcal{D}}}(R)$ is an isomorphism. \[ch150423b1\] When $R$ has a dualizing complex $D$, given an $R$-complex $X\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, one has ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if $X\in{{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{R}(C,D)}}(R)$, by [@christensen:scatac (4.7) Theorem]; this uses Facts \[ch150423c1\] and \[ch150423a1\], which imply that ${\mathbf{R}\!\operatorname{Hom}_{R}(C,D)}$ is semidualizing and $C\simeq{\mathbf{R}\!\operatorname{Hom}_{R}({\mathbf{R}\!\operatorname{Hom}_{R}(C,D)},D)}$. \[ch150423b2\] Assume that ${\operatorname{fd}}({\varphi})<\infty$, and let $X\in{{{\mathcal{D}}}_{\text{b}}}(S)$. Then $X\in{{\mathcal{A}}_C}(R)$ if and only if $X\in{{\mathcal{A}}}_{{S\otimes^{\mathbf{L}}_{R}C}}(S)$, by [@christensen:scatac (5.3.a) Proposition and (5.10) Theorem]. The proof of the following lemma is similar to that of [@iyengar:golh 7.3 Corollary], but it is different enough that we include a proof. \[lem150421a\] Assume that ${\varphi}$ is Gorenstein, and let $X\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(S)$ be such that each homology module ${\operatorname{H}}_i(X)$ is finitely generated over $R$. 1. \[lem150421a1\] One has ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_S(X)<\infty$. 2. \[lem150421a2\] One has ${\mathrm{G}\text{-}\!\dim}_R(X)<\infty$ if and only if ${\mathrm{G}\text{-}\!\dim}_S(X)<\infty$. Case 1: $R$ has a dualizing compelx $D^R$. From Fact \[para190908b\], we have $D^S\simeq{S\otimes^{\mathbf{L}}_{R}D^R}$ is dualizing for $S$. And Facts \[ch150423b1\]–\[ch150423b2\] say that ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if $X\in{{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{R}(C,D^R)}}(R)$ if and only if $X\in{{\mathcal{A}}}_{{S\otimes^{\mathbf{L}}_{R}{\mathbf{R}\!\operatorname{Hom}_{R}(C,D^R)}}}(S)$. Standard isomorphisms give $${S\otimes^{\mathbf{L}}_{R}{\mathbf{R}\!\operatorname{Hom}_{R}(C,D^R)}}\simeq {\mathbf{R}\!\operatorname{Hom}_{S}({S\otimes^{\mathbf{L}}_{R}C},{S\otimes^{\mathbf{L}}_{R}D^R})}\simeq {\mathbf{R}\!\operatorname{Hom}_{S}({S\otimes^{\mathbf{L}}_{R}C},D^S)}$$ so ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if $X\in{{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{S}({S\otimes^{\mathbf{L}}_{R}C},D^S)}}(S)$; Facts \[ch150423a1\] and \[ch150423b1\] show that these are equivalent to ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_S(X)<\infty$, as desired. Case 2: we have ${{\mathfrak{m}}}{\operatorname{H}}(X)=0$. In this case, let ${\widehat{R}}$ and ${\widetilde}S$ denote the ${{\mathfrak{m}}}$-adic completions of $R$ and $S$, respectively, and let ${\widetilde}{\varphi}\colon{\widehat{R}}\to{\widetilde}S$ be the ${{\mathfrak{m}}}$-adic completion of ${\varphi}$, which is Gorenstein. By Fact \[ch150423a3\] we have ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if ${\mathrm{G}_{{{\widehat{R}}\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_{{\widehat{R}}}({{\widehat{R}}\otimes^{\mathbf{L}}_{R}X})<\infty$. As in the proof of [@iyengar:golh 7.1 Theorem] we have an ${\widehat{R}}$-isomorphism ${{\widehat{R}}\otimes^{\mathbf{L}}_{R}X}\simeq {{\widetilde}S\otimes^{\mathbf{L}}_{S}X}$, so ${\mathrm{G}_{{{\widehat{R}}\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_{{\widehat{R}}}({{\widehat{R}}\otimes^{\mathbf{L}}_{R}X})<\infty$ if and only if ${\mathrm{G}_{{{\widetilde}S\otimes^{\mathbf{L}}_{{\widehat{R}}}({{\widehat{R}}\otimes^{\mathbf{L}}_{R}C}})}\text{-}\!\dim}_{{\widetilde}S}({{\widetilde}S\otimes^{\mathbf{L}}_{S}X})<\infty$ by Case 1, i.e., if and only if ${\mathrm{G}_{{{\widetilde}S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_{{\widetilde}S}({{\widetilde}S\otimes^{\mathbf{L}}_{S}X})<\infty$. Again by Fact \[ch150423a3\] we have ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_S(X)<\infty$ if and only if ${\mathrm{G}_{{{\widetilde}S\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}C}})}\text{-}\!\dim}_{{\widetilde}S}({{\widetilde}S\otimes^{\mathbf{L}}_{S}X})<\infty$ if and only if ${\mathrm{G}_{{{\widetilde}S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_{{\widetilde}S}({{\widetilde}S\otimes^{\mathbf{L}}_{S}X})<\infty$, as desired. Case 3: the general case. Let $K^R$ be the Koszul complex over $R$ on a finite generating set for ${{\mathfrak{m}}}$. Then ${{\mathrm{G}_{C}\text{-}\!\dim}}_R(X)<\infty$ if and only if ${{\mathrm{G}_{C}\text{-}\!\dim}}_R({K^R\otimes^{\mathbf{L}}_{R}X})<\infty$ by [@frankild:rrhffd Theorem 4.4], that is, if and only if ${{\mathrm{G}_{C}\text{-}\!\dim}}_R({({K^R\otimes^{\mathbf{L}}_{R}S})\otimes^{\mathbf{L}}_{S}X})<\infty$. Case 2 says that this is equivalent to ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_R({({K^R\otimes^{\mathbf{L}}_{R}S})\otimes^{\mathbf{L}}_{S}X})<\infty$, which is equivalent to ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}_S(X)<\infty$ by another application of [@frankild:rrhffd Theorem 4.4]. This is the special case $C=R$ of part . Test Complexes and Ring Homomorphisms of Finite Flat Dimension {#sec150429a} ============================================================== In this section, let $C$ be a semidualizing $R$-complex. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, and let ${\operatorname{H-dim}}$ denote either ${\operatorname{pd}}$ or ${{\mathrm{G}_{C}\text{-}\!\dim}}$. Then $M$ is an ${\operatorname{H-dim}}$-test complex over $R$ if the following condition holds for all $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$: If ${\operatorname{Tor}^{R}_{i}(M,N)}=0$ for all $i\gg 0$, i.e., if ${M\otimes^{\mathbf{L}}_{R}N}\in{{{\mathcal{D}}}_{\text{b}}}(R)$, then ${\operatorname{H-dim}}_R(N)<\infty$. See [@celikbas:tgp Section 3] for examples and basic properties of these objects. The following two results are proved like [@celikbas:tgp Theorems 3.2 and 3.4], using Lemma \[lem150421a\] and Fact \[ch150423a3\]. \[prop140825a\] Assume that ${\operatorname{fd}}({\varphi})<\infty$, and let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. 1. \[prop140825a2\] If ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $S$, then $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$. 2. \[prop140825a2’\] If ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $S$, then $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $R$. 3. \[prop140825a1\] If ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $S$, then $M$ is ${\operatorname{pd}}$-test over $R$. \[thm140825ax\] Assume that ${\varphi}$ is Gorenstein, and let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Assume that the induced residue field extension $k\to l$ is finite. 1. \[thm140825ax1\] $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $S$. 2. \[thm140825ax2\] $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $S$. \[disc190914a\] Assume that ${\varphi}$ is flat with regular closed fibre. Let ${\mathbf{y}}=y_1,\ldots,y_n\in{{\mathfrak{n}}}$ be a sequence that forms a regular system of parameters for the regular local ring $S/{{\mathfrak{m}}}S$. Then ${\mathbf{y}}$ is $S$-regular and that the composition $\tau{\varphi}\colon R\to {\overline}S=S/({\mathbf{y}})$ of ${\varphi}$ with the quotient map $\tau\colon S\to {\overline}S$ is flat by, e.g., [@matsumura:crt Corollary to Theorem 22.5]. By construction, the closed fibre of $\tau{\varphi}$ is $S/({{\mathfrak{m}}},{\mathbf{y}})=l$. Here is one of our main results. One point is that one can prove better results about flat local maps by widening the context to finite flat dimension. Note that the case where ${{\mathfrak{m}}}S={{\mathfrak{n}}}$ is covered by [@celikbas:tgp Theorem 3.5]. As we note in [@celikbas:tgp Example 3.6], if $S/{{\mathfrak{m}}}S$ is not regular, then the ascent portion of this result fails. \[thm140825a\] Assume that ${\varphi}$ is flat with regular closed fibre, and let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Assume the induced field extension $k \to l$ is finite. Then $M$ is a ${\operatorname{pd}}$-test complex over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is a ${\operatorname{pd}}$-test complex over $S$. One implication is covered by Proposition \[prop140825a\]. For the reverse implication, assume that $M$ is a ${\operatorname{pd}}$-test complex over $R$. Follow the notation from Remark \[disc190914a\]. From [@celikbas:tgp Theorem 3.5], it follows that ${{\overline}S\otimes^{\mathbf{L}}_{R}M}={{\overline}S\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}M})}$ is ${\operatorname{pd}}$-test over ${\overline}S$, so Proposition \[prop140825a\] implies that ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $S$. PD-Test Results {#sec190908c} =============== As we discuss in the introduction, the point of this section is to prove Theorem \[thm140915a\], and the proof relies heavily on a version of pd-test objects for finite dimensional DG algebras. One may consult any of the following for background on DG algebras and derived categories of DG modules [@avramov:ifr; @avramov:dgha; @avramov:htecdga; @beck:sgidgca; @felix:rht]. \[defn190913a\] We say that a (positively graded commutative) DG algebra $A$ is weakly local if ${\operatorname{H}}_0(A)$ is local and noetherian and ${\operatorname{H}}(A)$ is finitely generated as a module over ${\operatorname{H}}_0(A)$. In particular, ${\operatorname{H}}(A)$ is bounded. In this situation, let ${{\mathfrak{m}}}_A$ be the augmentation ideal of $A$ corresponding to the maximal ideal ${{\mathfrak{m}}}_{{\operatorname{H}}_0(A)}\subset{\operatorname{H}}_0(A)$, and set $k=A/{{\mathfrak{m}}}_A$. We sometimes summarize this by writing that $(A,{{\mathfrak{m}}}_A,k)$ is a weakly local DG algebra. \[disc190913b\] If $k$ is a field and $A$ is a finite dimensional DG $k$-algebra such that $A_0=k$ and ${\operatorname{H}}_0(A)\neq 0$, then $A$ is weakly local with ${{\mathfrak{m}}}_A=A_+$ and $\partial^A_1=0$ and $A/{{\mathfrak{m}}}_A\cong k\cong{\operatorname{H}}_0(A)$. \[fact190913a\] Let $(A,{{\mathfrak{m}}}_A,k)$ be a weakly local DG algebra. For a given $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$, the following conditions are equivalent. 1. \[fact190913a1\] $N$ has a semi-free resolution $F\simeq N$ over $A$ with a finite semi-basis. 2. \[fact190913a2\] For all $L\in{{{\mathcal{D}}}_{\text{b}}}(A)$, one has ${N\otimes^{\mathbf{L}}_{A}L}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. 3. \[fact190913a3\] One has ${N\otimes^{\mathbf{L}}_{A}k}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. For the implication $\implies$, replace $A$ with a bounded truncation and use $F$ to compute ${N\otimes^{\mathbf{L}}_{A}L}$. The implication $\implies$ is in [@avramov:htecdga Proposition B.9]. Note that in this situation, the resolution $F$ will be homologically bounded but not necessarily bounded, unless $A$ is bounded. Furthermore, for the implication $\implies$, it is crucial that $A$ be homologically bounded. Here is the DG version of ${\operatorname{pd}}$-test objects we use to prove Theorem \[thm140915a\]. \[defn190913b\] Let $A$ be a weakly local DG algebra. We say that $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ is perfect over $A$ if $M$ satisfies the equivalent conditions of Fact \[fact190913a\]. Then $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ is ${\operatorname{pd}}$-test over $A$ if the following condition holds for all $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$: If ${M\otimes^{\mathbf{L}}_{A}N}\in{{{\mathcal{D}}}_{\text{b}}}(A)$, then $N$ is perfect. \[disc190913a\] If $(A,{{\mathfrak{m}}}_A,k)$ is a weakly local DG algebra, then Fact \[fact190913a\] implies that $k$ is ${\operatorname{pd}}$-test over $A$. Our proof of Theorem \[thm140915a\] follows almost directly from the following result via a construction of Avramov. \[thm190913a\] Let $A$ be a finite-dimensional DG $k$-algebra with $A_0=k$ and ${\operatorname{H}}_0(A)\neq 0$. Let $k\to l$ be a field extension, set $B={l\otimes_{k}A}$, and consider the natural morphism of DG algebras $A\to B$. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ be given. 1. \[thm190913a1\] If ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $B$, then $M$ is ${\operatorname{pd}}$-test over $A$. 2. \[thm190913a2\] The converse of part holds if the extension $k\to l$ is algebraic. We argue as in the proof of [@celikbas:tgp Theorem 3.2]. Assume that ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $B$. To show that $M$ is ${\operatorname{pd}}$-test over $A$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ be such that ${M\otimes^{\mathbf{L}}_{A}N}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. As $N$ is homologically finite over $A$, which is finite dimensional over $k$, we have $${B\otimes^{\mathbf{L}}_{A}N}\simeq{({l\otimes_{k}A})\otimes^{\mathbf{L}}_{A}N}\simeq{l\otimes_{k}N}\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(l)$$ so ${B\otimes^{\mathbf{L}}_{A}N}\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$, and similarly ${B\otimes^{\mathbf{L}}_{A}M},{B\otimes^{\mathbf{L}}_{A}({M\otimes^{\mathbf{L}}_{A}N})}\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$. Moreover, we have the following isomorphisms in ${{\mathcal{D}}}(B)$: $${({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}({B\otimes^{\mathbf{L}}_{A}N})} \simeq{({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A}N} \simeq{B\otimes^{\mathbf{L}}_{A}({M\otimes^{\mathbf{L}}_{A}N})}.$$ As ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $B$, the DG $B$-module ${B\otimes^{\mathbf{L}}_{A}N}$ is perfect. It is straightforward to show (arguing as above or using a minimal semi-free resolution of $N$ over $A$, with Fact \[fact190913a\]) that $N$ is perfect over $A$, as desired. Assume now that $M$ is ${\operatorname{pd}}$-test over $A$. Case 1: $k\to l$ is finite. To show that ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $B$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$ be such that ${({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(B)$. Since $k\to l$ is finite, we have $B\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$, so $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$. Moreover, ${M\otimes^{\mathbf{L}}_{A}N}\simeq{({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. As $M$ is ${\operatorname{pd}}$-test over $A$, the DG module $N$ is perfect over $A$. To conclude that $N$ is perfect over $B$, it suffices to show that ${N\otimes^{\mathbf{L}}_{B}l}\in{{{\mathcal{D}}}_{\text{b}}}(B)$, equivalently, that ${N\otimes^{\mathbf{L}}_{B}l}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. By assumption, we have ${N\otimes^{\mathbf{L}}_{A}k}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. By construction, we have $${B\otimes_{A}k}\cong{({l\otimes_{k}A})\otimes_{A}k}\cong{l\otimes_{k}k}\simeq l$$ so $${N\otimes^{\mathbf{L}}_{B}l}\simeq{N\otimes^{\mathbf{L}}_{B}({B\otimes_{A}k})}\simeq{N\otimes^{\mathbf{L}}_{A}k}\in{{{\mathcal{D}}}_{\text{b}}}(A)$$ as desired. Case 2: $k\to l$ is algebraic. To show that ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $B$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$ be such that ${({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(B)$. Truncate a degreewise finite $B$-semifree resolution of $N$ if necessary to assume without loss of generality that $N$ is finite dimensional over $l$. It follows that the differential and scalar multiplication on $N$ are represented by matrices consisting of finitely many elements of $l$. Let $k'$ be the intermediate field extension $k\to k'\to l$ generated over $k$ by this finite set of elements. Since $l$ is algebraic over $k$, the extension $k\to k'$ is finite. Set $A'={k'\otimes_{k}A}$. Since $k'$ contains the elements representing the differential and scalar multiplication on $N$, there is a DG $A'$-module $L$ that is bounded and degreewise finite over $k'$ such that $N\cong {B\otimes_{A'}L}$. Using these constructions, we compute: $$\begin{aligned} {({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N} &\simeq{({B\otimes^{\mathbf{L}}_{A'}({A'\otimes^{\mathbf{L}}_{A}M})})\otimes^{\mathbf{L}}_{B}({B\otimes_{A'}L})}\\ &\simeq{B\otimes^{\mathbf{L}}_{A'}({({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L})}\\ &\simeq{({l\otimes^{\mathbf{L}}_{k'}A'})\otimes^{\mathbf{L}}_{A'}({({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L})}\\ &\simeq{l\otimes^{\mathbf{L}}_{k'}({({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L})}\end{aligned}$$ Thus, we have ${l\otimes^{\mathbf{L}}_{k'}({({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L})}\simeq{({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(B)$. Since $k'\to l$ is faithfully flat, it follows that ${({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L}$ is homologically bounded as well. Case 1 implies that ${A'\otimes^{\mathbf{L}}_{A}M}$ is ${\operatorname{pd}}$-test over $A'$, so the homological boundedness of ${({A'\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{A'}L}$ implies that $L$ is perfect over $A'$. It follows readily that $N\simeq{B\otimes^{\mathbf{L}}_{A'}L}$ is perfect over $B$, as desired. The next result uses the exterior DG algebra structure on the Koszul complex. \[lem190915a\] Let $K=K^R({\mathbf{x}})$ be the Koszul complex on ${\mathbf{x}}=x_1,\ldots,x_n\in{{\mathfrak{m}}}$, and let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. 1. \[lem190915a1\] If ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $K$, then $M$ is ${\operatorname{pd}}$-test over $R$. 2. \[lem190915a2\] The converse of part holds when ${\mathbf{x}}$ is part of a minimal generating sequence for ${{\mathfrak{m}}}$. Argue as in the descent result [@celikbas:tgp Theorem 3.2]. We argue as in the proof of Theorem \[thm190913a\], with a twist. Assume that $M$ is a ${\operatorname{pd}}$-test complex over $R$. To show that ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $K$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(K)$ be such that ${({K\otimes^{\mathbf{L}}_{R}M})\otimes^{\mathbf{L}}_{K}N}\in{{{\mathcal{D}}}_{\text{b}}}(K)$. Since $K\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, we have $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Moreover, ${M\otimes^{\mathbf{L}}_{R}N}\simeq{({K\otimes^{\mathbf{L}}_{R}M})\otimes^{\mathbf{L}}_{K}N}\in{{{\mathcal{D}}}_{\text{b}}}(R)$. As $M$ is ${\operatorname{pd}}$-test over $R$, the DG module $N$ is perfect over $R$, so ${k\otimes^{\mathbf{L}}_{R}N}\in{{{\mathcal{D}}}_{\text{b}}}(R)$. To conclude that $N$ is perfect over $K$, we compute: $$\begin{aligned} {k\otimes^{\mathbf{L}}_{R}N} &\simeq{({k\otimes^{\mathbf{L}}_{R}K})\otimes^{\mathbf{L}}_{K}N} \simeq{{\operatorname{H}}({k\otimes^{\mathbf{L}}_{R}K})\otimes^{\mathbf{L}}_{K}N} \simeq{{\operatorname{H}}({k\otimes^{\mathbf{L}}_{R}K})\otimes^{\mathbf{L}}_{k}({k\otimes^{\mathbf{L}}_{K}N})}\end{aligned}$$ The second isomorphism here is from [@MR2592508 Theorem 9.1]; this is where we use the fact that ${\mathbf{x}}$ is part of a minimal generating sequence for ${{\mathfrak{m}}}$. Since ${k\otimes^{\mathbf{L}}_{R}N}$ is homologically bounded, the Künneth formula implies that ${k\otimes^{\mathbf{L}}_{K}N}\in{{{\mathcal{D}}}_{\text{b}}}(K)$, so $N$ is perfect. Here is our main result about ${\operatorname{pd}}$-text complexes. \[thm140915a\] Assume that ${\varphi}$ is flat with regular closed fibre, and let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Assume the induced field extension $k \to l$ is algebraic. Then $M$ is a ${\operatorname{pd}}$-test complex over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is a ${\operatorname{pd}}$-test complex over $S$. As with Theorem \[thm140825a\], one implication is covered by Proposition \[prop140825a\]. For the reverse implication, assume that $M$ is a ${\operatorname{pd}}$-test complex over $R$. Case 1: $R$ and $S$ are complete and ${{\mathfrak{m}}}S={{\mathfrak{n}}}$. Let $\tau\colon Q\to R$ be a minimal Cohen presentation of $R$, so the map is a surjective ring homomorphism where $Q$ is a complete regular local ring with the same embedding dimension as $R$. Apply [@avramov:solh (1.6) Theorem] to find a commutative diagram of local homomorphisms of complete local rings $$\xymatrix{Q\ar[r]^-\alpha\ar[d]_-\tau&Q'\ar[d]^-{\tau'}\\ R\ar[r]^-{\varphi}&S}$$ such that $\alpha$ is flat with regular closed fibre, $\tau'$ is surjective, and the induced map ${R\otimes_{Q}Q'}\to S$ is an isomorphism. Moreover, the proof of loc. cit. shows that ${\operatorname{edim}}(Q'/{{\mathfrak{m}}}_Q Q')={\operatorname{edim}}(S/{{\mathfrak{m}}}S)=0$, so the closed fibre of $\alpha$ is $Q'/{{\mathfrak{m}}}_Q Q'\cong l$. Let $F{\xrightarrow}\simeq R$ be a bounded degreewise finite semi-free DG algebra resolution of $R$ over $Q$. The isomorphism ${R\otimes_{Q}Q'}\cong S$ from the previous paragraph implies that $F'={F\otimes_{Q}Q'}{\xrightarrow}\simeq S$ is a bounded degreewise finite semi-free DG algebra resolution of $S$ over $Q'$ since $\alpha$ is flat. Let ${\mathbf{x}}$ be a minimal generating sequence for ${{\mathfrak{m}}}$. Our assumptions imply that ${\varphi}({\mathbf{x}})$ is a minimal generating sequence for ${{\mathfrak{n}}}$. Set $K^R=K^R({\mathbf{x}})$ and $K^S=K^S({\varphi}({\mathbf{x}}))$ and consider the following standard commutative diagram of morphisms of DG algebras where $K^Q$ is the Koszul complex over $Q$ on a lift ${\mathbf{y}}$ of the sequence ${\mathbf{x}}$, and $K^{Q'}=K^{Q'}(\alpha({\mathbf{y}}))\cong{Q'\otimes_{Q}K^Q}$. $$\begin{split}\label{eq191118a}\xymatrix{ R\ar[r]\ar[d]_-{\varphi}&K^R\ar[d]&{R\otimes_{Q}K^Q}\ar[d]\ar[l]_-\cong&{F\otimes_{Q}K^Q}\ar[d]\ar[l]_-\simeq\ar[r]^-\simeq&{F\otimes_{Q}k}\ar[d] \\ S\ar[r]&K^S&{S\otimes_{Q'}K^{Q'}}\ar[l]_-\cong &{F'\otimes_{Q'}K^{Q'}}\ar[l]_-\simeq\ar[r]^-\simeq &{F'\otimes_{Q'}l} }\end{split}$$ Using the properties catalogued above, it is straightforward to show that each square in this diagram is a pushout square. Since $M$ is ${\operatorname{pd}}$-test over $R$, Lemma \[lem190915a\] implies that ${K^R\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $K^R$. Since the other arrows in the top row of the above diagram are equivalences, it is straightforward to show that ${K^R\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over ${F\otimes_{Q}K^Q}$ by restriction of scalars, and ${F\otimes^{\mathbf{L}}_{\protect{{({F\otimes_{Q}k})\otimes_{Q}K^Q}}}({K^R\otimes^{\mathbf{L}}_{R}M})}$ is ${\operatorname{pd}}$-test over ${F\otimes_{Q}k}$ by base-change. From our assumptions, it is straightforward to show that ${F'\otimes_{Q'}l}\cong{l\otimes_{k}({F\otimes_{Q}k})}$. Thus, Theorem \[thm190913a\] implies that the following is ${\operatorname{pd}}$-test over ${F'\otimes_{Q'}l}$. $$\begin{gathered} {F\otimes^{\mathbf{L}}_{\protect{{({F'\otimes_{Q'}l})\otimes_{Q}k}}}({F\otimes^{\mathbf{L}}_{\protect{{({F\otimes_{Q}k})\otimes_{Q}K^Q}}}({K^R\otimes^{\mathbf{L}}_{R}M})})}\\ \simeq {F'\otimes^{\mathbf{L}}_{\protect{{({F'\otimes_{Q'}l})\otimes_{Q'}K^{Q'}}}}({F\otimes^{\mathbf{L}}_{\protect{{({F'\otimes_{Q'}K^{Q'}})\otimes_{Q}K^Q}}}({K^R\otimes^{\mathbf{L}}_{R}M})})}\end{gathered}$$ From the bottom-right equivalence in the diagram , we conclude that the DG module ${F\otimes^{\mathbf{L}}_{\protect{{({F'\otimes_{Q'}K^{Q'}})\otimes_{Q}K^Q}}}({K^R\otimes^{\mathbf{L}}_{R}M})}$ is ${\operatorname{pd}}$-test over ${F'\otimes_{Q'}K^{Q'}}$ by restriction of scalars. From the other equivalences in the bottom row of the diagram, we see that the following DG module is ${\operatorname{pd}}$-test over $K^S$; $$\begin{aligned} {F\otimes^{\mathbf{L}}_{\protect{{({F\otimes_{Q}k})\otimes_{Q}K^Q}}}({K^R\otimes^{\mathbf{L}}_{R}M})} &\simeq{K^S\otimes^{\mathbf{L}}_{K^R}({K^R\otimes^{\mathbf{L}}_{R}M})} \simeq{K^S\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}M})}\end{aligned}$$ the first isomorphism comes from the pushout property of the diagram. Lemma \[lem190915a\] implies that ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $S$. This concludes the proof in Case 1. Case 2: ${{\mathfrak{m}}}S={{\mathfrak{n}}}$. In this case, consider the following natural diagram of flat local ring homomorphisms. $$\xymatrix{ R\ar[r]\ar[d]&S\ar[d]\\ {\widehat{R}}\ar[r]&{\widehat{S}}}$$ Then ${{\widehat{R}}\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over ${\widehat{R}}$ by Theorem \[thm140825a\]. Case 1 then implies that the following complex is ${\operatorname{pd}}$-test over ${\widehat{S}}$. $$\begin{aligned} {{\widehat{S}}\otimes^{\mathbf{L}}_{{\widehat{R}}}({{\widehat{R}}\otimes^{\mathbf{L}}_{R}M})} &\simeq{{\widehat{S}}\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}M})}\end{aligned}$$ Another application of Theorem \[thm140825a\] implies that ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\operatorname{pd}}$-test over $S$. Case 3: the general case follows from Case 2 as in the proof of Theorem \[thm140825a\]. \[disc190915a\] It would be nice to answer [@celikbas:tgp Question 3.7] in general. This is equivalent to proving that the conclusion of Theorem \[thm140915a\] holds without the assumption that $k\to l$ is algebraic. Using a transcendence basis, because of Theorem \[thm140915a\] and its proof, this reduces to proving that the converse to Theorem \[thm190913a\] holds in the case where $k\to l$ is purely transcendental of finite transcendence degree. Unfortunately, we do not know how to prove this even when $k\to l$ is purely transcendental of transcendence degree 1. G-Dim-Test Results {#sec190913a} ================== Throughout this section, $(A,{{\mathfrak{m}}}_A,k)$ is a weakly local DG algebra. Here, we analyze ascent properties for ${{\mathrm{G}_{C}\text{-}\!\dim}}$- and ${\mathrm{G}\text{-}\!\dim}$-test modules using the techniques of the preceding section. Our main result here is Theorem \[thm190915b\]. We say that $C\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ is a semidualizing DG $A$-module if the natural homothety morphism $\chi^A_C\colon A\to{\mathbf{R}\!\operatorname{Hom}_{A}(C,C)}$ is an isomorphism in ${{\mathcal{D}}}(A)$. For instance, since $A$ is homologically bounded, $A$ itself is a semidualizing DG $A$-module. Assume that $C$ is a semidualizing DG $A$-module. We say that $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ is derived $C$-reflexive over $A$ if ${\mathbf{R}\!\operatorname{Hom}_{A}(N,C)}\in{{{\mathcal{D}}}_{\text{b}}}(A)$ and the natural biduality morphism $\delta^N_C\colon N\to{\mathbf{R}\!\operatorname{Hom}_{A}({\mathbf{R}\!\operatorname{Hom}_{A}(N,C)},C)}$ is an isomorphism in ${{\mathcal{D}}}(A)$. Then $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $A$ if the following condition holds for all $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$: If ${M\otimes^{\mathbf{L}}_{A}N}\in{{{\mathcal{D}}}_{\text{b}}}(A)$, then $N$ is derived $C$-reflexive. In the case $C=A$, we write ${\mathrm{G}\text{-}\!\dim}$-test instead of ${\mathrm{G}_{A}\text{-}\!\dim}$-test. Fact \[fact190913a\] implies that $k$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $A$ for each semidualizing DG $A$-module $C$ because every perfect DG $A$-module is derived $C$-reflexive. In particular, $k$ is ${\mathrm{G}\text{-}\!\dim}$-test. We proceed as in the preceding section with some preparatory lemmas. \[lem190929a\] Assume that $A$ is a finite-dimensional DG $k$-algebra with $A_0=k$ and ${\operatorname{H}}_0(A)\neq 0$. Let $k\to l$ be a field extension, set $B={l\otimes_{k}A}$, and consider the natural morphism of DG algebras $A\to B$. Let $C\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ be given. Then ${B\otimes^{\mathbf{L}}_{A}C}$ is semidualizing over $B$ if and only if $C$ is semidualizing over $A$. Note that since $l$ is free over $k$, we have $B\simeq{l\otimes^{\mathbf{L}}_{k}A}$. Furthermore, the $A$-algebra $B={l\otimes_{k}A}$ is free with semi-basis concentrated in degree 0. These facts justify three of the unspecified isomorphisms in the following diagram, while the fourth one is tensor-cancellation. $$\xymatrix@R=6mm{ {l\otimes^{\mathbf{L}}_{k}A}\ar[r]^-{{l\otimes^{\mathbf{L}}_{k}\chi^A_C}}\ar[d]_-\simeq&{l\otimes^{\mathbf{L}}_{k}{\mathbf{R}\!\operatorname{Hom}_{A}(C,C)}}\ar[d]^-\simeq\\ B\ar[d]_-{\chi^B_{{B\otimes^{\mathbf{L}}_{A}C}}}&{({l\otimes^{\mathbf{L}}_{k}A})\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}(C,C)}}\ar[d]^-\simeq\\ {\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}C},{B\otimes^{\mathbf{L}}_{A}C})}&{B\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}(C,C)}}\ar[l]_-\simeq }$$ It follows that $\chi^B_{{B\otimes^{\mathbf{L}}_{A}C}}$ is an isomorphism if and only if ${l\otimes^{\mathbf{L}}_{k}\chi^A_C}$ is an isomorphism; since $l$ is free and non-zero over $k$, the morphism ${l\otimes^{\mathbf{L}}_{k}\chi^A_C}$ is an isomorphism is and only if $\chi^A_C$ is one. Since ${B\otimes^{\mathbf{L}}_{A}C}\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$, the desired conclusion now follows. \[lem190929b\] Let $A$ be a finite-dimensional DG $k$-algebra with $A_0=k$ and ${\operatorname{H}}_0(A)\neq 0$. Let $k\to l$ be a field extension, set $B={l\otimes_{k}A}$, and consider the natural morphism of DG algebras $A\to B$. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ be given, and let $C$ be a semidualizing DG $A$-module. Then ${B\otimes^{\mathbf{L}}_{A}M}$ is derived $({B\otimes^{\mathbf{L}}_{A}C})$-reflexive over $B$ if and only if $M$ is derived $C$-reflexive over $A$. As in the proof of Lemma \[lem190929a\], in ${{\mathcal{D}}}(A)$ we have ${\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}M},{B\otimes^{\mathbf{L}}_{A}C})}\simeq{l\otimes^{\mathbf{L}}_{k}{\mathbf{R}\!\operatorname{Hom}_{A}(M,C)}}$ and therefore ${\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}M},{B\otimes^{\mathbf{L}}_{A}C})}\in{{{\mathcal{D}}}_{\text{b}}}(B)$ if and only if ${\mathbf{R}\!\operatorname{Hom}_{A}(M,C)}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. Assume therefore that ${\mathbf{R}\!\operatorname{Hom}_{A}(M,C)}\in{{{\mathcal{D}}}_{\text{b}}}(A)$, and use the next diagram $$\xymatrix@R=6mm{ {l\otimes^{\mathbf{L}}_{k}M}\ar[r]^-{{l\otimes^{\mathbf{L}}_{k}\delta^M_C}}\ar[dd]_-\simeq&{l\otimes^{\mathbf{L}}_{k}{\mathbf{R}\!\operatorname{Hom}_{A}({\mathbf{R}\!\operatorname{Hom}_{A}(M,C)},C)}}\ar[d]^-\simeq\\ &{({l\otimes^{\mathbf{L}}_{k}A})\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}({\mathbf{R}\!\operatorname{Hom}_{A}(M,C)},C)}}\ar[d]^-\simeq\\ {({l\otimes^{\mathbf{L}}_{k}A})\otimes^{\mathbf{L}}_{A}M}\ar[dd]_-\simeq&{B\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}({\mathbf{R}\!\operatorname{Hom}_{A}(M,C)},C)}}\ar[d]^-\simeq\\ &{\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}(M,C)}},{B\otimes^{\mathbf{L}}_{A}C})} \\ {B\otimes^{\mathbf{L}}_{A}M}\ar[r]^-{\delta^{{B\otimes^{\mathbf{L}}_{A}M}}_{{B\otimes^{\mathbf{L}}_{A}C}}}&{\mathbf{R}\!\operatorname{Hom}_{B}({\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}M},{B\otimes^{\mathbf{L}}_{A}C})},{B\otimes^{\mathbf{L}}_{A}C})}\ar[u]_-\simeq }$$ as in the proof of Lemma \[lem190929a\] to complete the argument. Here are the versions of Theorem \[thm190913a\] and Lemma \[lem190915a\] for this context. \[thm190913a’\] Let $A$ be a finite-dimensional DG $k$-algebra with $A_0=k$ and ${\operatorname{H}}_0(A)\neq 0$. Let $k\to l$ be a field extension, set $B={l\otimes_{k}A}$, and consider the natural morphism of DG algebras $A\to B$. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$ be given, and let $C$ be a semidualizing DG $A$-module. 1. \[thm190913a’1\] If ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\mathrm{G}_{{B\otimes^{\mathbf{L}}_{A}C}}\text{-}\!\dim}$-test over $B$, then $M$ is ${\mathrm{G}_{C}\text{-}\!\dim}$-test over $A$. 2. \[thm190913a’2\] The converse of part holds if the extension $k\to l$ is algebraic. 3. \[thm190913a’11\] If ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $B$, then $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $A$. 4. \[thm190913a’22\] The converse of part holds if the extension $k\to l$ is algebraic. Argue as in the proof of Theorem \[thm190913a\], using Lemma \[lem190929b\]. Assume now that $M$ is a ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test complex over $A$. Case 1: $k\to l$ is finite. To show that ${B\otimes^{\mathbf{L}}_{A}M}$ is ${\mathrm{G}_{{B\otimes^{\mathbf{L}}_{A}C}}\text{-}\!\dim}$-test over $B$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(B)$ be such that ${({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(B)$. Since $k\to l$ is finite, we have $B\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$, so $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(A)$. Moreover, ${M\otimes^{\mathbf{L}}_{A}N}\simeq{({B\otimes^{\mathbf{L}}_{A}M})\otimes^{\mathbf{L}}_{B}N}\in{{{\mathcal{D}}}_{\text{b}}}(A)$. As $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $A$, the DG module $N$ is derived $C$-reflexive over $A$. To conclude that $N$ is derived $({B\otimes^{\mathbf{L}}_{A}C})$-reflexive over $B$, we use the Auslander class. Since $A$ is finite dimensional over $k$, it has a dualizing DG module $D^A$. From the definition of $B$ as ${l\otimes_{k}A}$, it is straightforward to show that $D^B={B\otimes^{\mathbf{L}}_{A}D^A}$ is a dualizing DG $B$-module. As in [@christensen:scatac (4.7) Theorem] or [@avramov:rhafgd (4.1.7)] one shows that $N$ is derived reflexive over $A$ if and only if $N$ is in the Auslander class ${{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{A}(C,D^A)}}(A)$, and similarly over $B$. Moreover, arguing as in [@christensen:scatac (5.3) Proposition], one sees that $N$ is in ${{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{A}(C,D^A)}}(A)$ if and only if it is in ${{\mathcal{A}}}_{{B\otimes^{\mathbf{L}}_{A}{\mathbf{R}\!\operatorname{Hom}_{A}(C,D^A)}}}(B)={{\mathcal{A}}}_{{\mathbf{R}\!\operatorname{Hom}_{B}({B\otimes^{\mathbf{L}}_{A}C},D^B)}}(B)$, that is, $N$ is derived $C$-reflexive over $A$ if and only if it is derived $({B\otimes^{\mathbf{L}}_{A}C})$-reflexive over $B$. This concludes the proof in Case 1. Case 2: $k\to l$ is algebraic. Argue as in Case 2 of the proof of Theorem \[thm190913a\]. – These are the special case $C=A$ of parts –. As in Section \[sec190908c\], we use the exterior DG algebra structure on the Koszul complex to prove our ascent result for ${\mathrm{G}\text{-}\!\dim}$-test complexes. \[fact190929a\] Let $K=K^R({\mathbf{x}})$ be the Koszul complex on ${\mathbf{x}}=x_1,\ldots,x_n\in{{\mathfrak{m}}}$, and let $C\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Then ${K\otimes^{\mathbf{L}}_{R}C}$ is semidualizing over $K$ if and only if $C$ is semidualizing over $R$ by [@christensen:dvke Lemma A.3(a)]. \[lem190915a’\] Let $K=K^R({\mathbf{x}})$ be the Koszul complex on ${\mathbf{x}}=x_1,\ldots,x_n\in{{\mathfrak{m}}}$. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$ be given, and let $C$ be a semidualizing $R$-complex. 1. \[lem190915a’1\] ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{K\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $K$ if and only if $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$. 2. \[lem190915a’2\] ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $K$ if and only if $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $R$. Again, we focus on part . For one implication, argue as in the descent result [@celikbas:tgp Theorem 3.2]. For the converse, assume that $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$. Case 1: $R$ has a dualizing complex $D^R$. To show that ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{K\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $K$, let $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(K)$ be such that ${({K\otimes^{\mathbf{L}}_{R}M})\otimes^{\mathbf{L}}_{K}N}\in{{{\mathcal{D}}}_{\text{b}}}(K)$. Since $K\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, we have $N\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$. Moreover, ${M\otimes^{\mathbf{L}}_{R}N}\simeq{({K\otimes^{\mathbf{L}}_{R}M})\otimes^{\mathbf{L}}_{K}N}\in{{{\mathcal{D}}}_{\text{b}}}(R)$. As $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$, the DG module $N$ is derived $C$-reflexive over $R$, Now argue as in the proof of Theorem \[thm190913a’\] using $D^R$ and $D^K={K\otimes^{\mathbf{L}}_{R}D^R}$ to prove that $N$ is derived ${K\otimes^{\mathbf{L}}_{R}C}$-reflexive over $K$. Case 2: $({\mathbf{x}})R$ is ${{\mathfrak{m}}}$-primary. Proposition \[thm140825ax\] implies that ${{\widehat{R}}\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{{\widehat{R}}\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test for ${\widehat{R}}$. Set ${\widehat{K}}={{\widehat{R}}\otimes^{\mathbf{L}}_{R}K}$. Recall that ${\widehat{R}}$ has a dualizing complex, so Case 1 implies that the DG ${\widehat{K}}$-module $${{\widehat{K}}\otimes^{\mathbf{L}}_{{\widehat{R}}}({{\widehat{R}}\otimes^{\mathbf{L}}_{R}M})}\simeq{{\widehat{K}}\otimes^{\mathbf{L}}_{R}M}\simeq{K\otimes^{\mathbf{L}}_{R}M}$$ is ${\mathrm{G}_{{{\widehat{K}}\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test for ${\widehat{K}}$; the second isomorphism here uses the assumption on $({\mathbf{x}})R$, which implies that the natural morphism $K\to{\widehat{K}}$ is a quasiisomorphism of DG algebras. This quasiisomorphism furthermore implies that this DG module is ${\mathrm{G}_{{K\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test for $K$, by restriction of scalars. Case 3: the general case. Let ${\mathbf{y}}=y_1,\ldots,y_m\in{{\mathfrak{m}}}$ be a generating sequence for ${{\mathfrak{m}}}$. By Case 2, the DG $K^R({\mathbf{x}},{\mathbf{y}})$-module ${K^R({\mathbf{x}},{\mathbf{y}})\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{K^R({\mathbf{x}},{\mathbf{y}})\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $K^R({\mathbf{x}},{\mathbf{y}})$. Argue as in the descent implication, from the DG algebra morphism $K\to K^R({\mathbf{x}},{\mathbf{y}})$ to conclude that ${K\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{K\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$ test over $K$, as desired. \[thm140915a’\] Assume that ${\varphi}$ is flat with regular closed fibre such that the induced field extension $k \to l$ is algebraic. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$ be given, and let $C$ be a semidualizing $R$-complex. 1. \[thm140915a’1\] $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}M}}\text{-}\!\dim}$-test over $S$. 2. \[thm140915a’2\] $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $S$. Argue as in the proof of Theorem \[thm140915a\]. We conclude with the main result of this section. \[thm190915b\] Assume that ${\varphi}$ is Gorenstein such that the induced field extension $k \to l$ is algebraic. Let $M\in{{{\mathcal{D}}}^{\text{f}}_{\text{b}}}(R)$, and let $C$ be a semidualizing $R$-complex. 1. \[thm190915b1\] $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $S$. 2. \[thm190915b2\] $M$ is ${\mathrm{G}\text{-}\!\dim}$-test over $R$ if and only if ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}\text{-}\!\dim}$-test over $S$. Again we focus on part . One implication is from Proposition \[prop140825a\]. For the converse, assume that $M$ is ${{\mathrm{G}_{C}\text{-}\!\dim}}$-test over $R$. Let $R\to R'\to {\widehat{S}}$ be a Cohen factorization of the semicompletion $\grave{\varphi}\colon R\to{\widehat{S}}$, This provides isomorphisms $$\begin{gathered} {{\widehat{S}}\otimes^{\mathbf{L}}_{R'}({R'\otimes^{\mathbf{L}}_{R}C})}\simeq{{\widehat{S}}\otimes^{\mathbf{L}}_{R}C}\simeq{{\widehat{S}}\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}C})}\\ {{\widehat{S}}\otimes^{\mathbf{L}}_{R'}({R'\otimes^{\mathbf{L}}_{R}M})}\simeq{{\widehat{S}}\otimes^{\mathbf{L}}_{R}M}\simeq{{\widehat{S}}\otimes^{\mathbf{L}}_{S}({S\otimes^{\mathbf{L}}_{R}M})}.\end{gathered}$$ Theorem \[thm140915a’\] implies that ${R'\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{R'\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $R'$. Because of the displayed isomorphisms, Proposition \[thm140825ax\] implies that ${{\widehat{S}}\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{{\widehat{S}}\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over ${\widehat{S}}$. Thus, ${S\otimes^{\mathbf{L}}_{R}M}$ is ${\mathrm{G}_{{S\otimes^{\mathbf{L}}_{R}C}}\text{-}\!\dim}$-test over $S$ by Proposition \[prop140825a\]. Acknowledgments {#acknowledgments .unnumbered} =============== I am grateful to Olgur Celikbas, Craig Huneke, Srikanth Iyengar, Mohsen Gheibi, and Justin Lyle for useful conversations about this work. Much of this work was completed around the conference Morgantown Algebra Days (MAD) 2019. I appreciate the fact that the organizers of MAD 2019 created a productive workshop environment. Lastly, I am thankful for all the hard work Roger and Sylvia Wiegand have done over their careers to create/discover inspiring mathematics and to build a welcoming and inclusive environment in the commutative algebra community. They are two of the kindest people I know. [10]{} H. Altmann and S. Sather-Wagstaff, Chains of semidualizing complexes, in preparation. M. Auslander, Anneaux de [G]{}orenstein, et torsion en algèbre commutative, Séminaire d’Algèbre Commutative dirigé par Pierre Samuel, vol. 1966/67, Secrétariat mathématique, Paris, 1967. [MR ]{}[37 \#1435]{} M. Auslander and M. Bridger, Stable module theory, Memoirs of the American Mathematical Society, No. 94, American Mathematical Society, Providence, R.I., 1969. [MR ]{}[42 \#4580]{} L. L. Avramov, Infinite free resolutions, Six lectures on commutative algebra (Bellaterra, 1996), Progr. Math., vol. 166, Birkhäuser, Basel, 1998, pp. 1–118. [MR ]{}[99m:13022]{} L. L. Avramov, A cohomological study of local rings of embedding codepth 3, J. Pure Appl. Algebra 216 (2012), no. 11, 2489–2506. [MR ]{}[2927181]{} L. L. Avramov, R.-O. Buchweitz, S. B. Iyengar, and C. Miller, Homology of perfect complexes, Adv. Math. 223 (2010), no. 5, 1731–1781. [MR ]{}[2592508]{} L. L. Avramov and H.-B. Foxby, Gorenstein local homomorphisms, Bull. Amer. Math. Soc. (N.S.) 23 (1990), no. 1, 145–150. [MR ]{}[1020605 (90k:13009)]{} [to3em]{}, Locally [G]{}orenstein homomorphisms, Amer. J. Math. 114 (1992), no. 5, 1007–1047. [MR ]{}[1183530 (93i:13019)]{} [to3em]{}, Grothendieck’s localization problem, Commutative algebra: syzygies, multiplicities, and birational algebra (South Hadley, MA, 1992), Contemp. Math., vol. 159, Amer. Math. Soc., Providence, RI, 1994, pp. 1–13. [MR ]{}[1266174 (94m:13011)]{} [to3em]{}, Ring homomorphisms and finite [G]{}orenstein dimension, Proc. London Math. Soc. (3) 75 (1997), no. 2, 241–270. [MR ]{}[98d:13014]{} [to3em]{}, Cohen-[M]{}acaulay properties of ring homomorphisms, Adv. Math. 133 (1998), no. 1, 54–95. [MR ]{}[1492786 (99c:13043)]{} L. L. Avramov, H.-B. Foxby, and S. Halperin, Differential graded homological algebra, in preparation. [to3em]{}, Descent and ascent of local properties along homomorphisms of finite flat dimension, J. Pure Appl. Algebra 38 (1985), no. 2-3, 167–185. [MR ]{}[814175 (87b:13026)]{} L. L. Avramov, H.-B. Foxby, and B. Herzog, Structure of local homomorphisms, J. Algebra 164 (1994), 124–145. [MR ]{}[95f:13029]{} L. L. Avramov, H.-B.Foxby, and J. Lescot, Bass series of local ring homomorphisms of finite flat dimension, Trans. Amer. Math. Soc. 335 (1993), no. 2, 497–523. [MR ]{}[93d:13026]{} L. L. Avramov and S. Halperin, Through the looking glass: a dictionary between rational homotopy theory and local algebra, Algebra, algebraic topology and their interactions ([S]{}tockholm, 1983), Lecture Notes in Math., vol. 1183, Springer, Berlin, 1986, pp. 1–27. [MR ]{}[846435 (87k:55015)]{} L. L. Avramov, S. Iyengar, and C. Miller, Homology over local homomorphisms, Amer. J. Math. 128 (2006), no. 1, 23–90. [MR ]{}[2197067]{} L. L. Avramov, S. B. Iyengar, S. Nasseh, and S. Sather-Wagstaff, Persistence of homology over commutative noetherian rings, in preparation. [to3em]{}, Homology over trivial extensions of commutative [DG]{} algebras, Comm. Algebra 47 (2019), no. 6, 2341–2356, see also `arxiv:1508.00748`. [MR ]{}[3957101]{} L. L. Avramov, A. R. Kustin, and M. Miller, Poincaré series of modules over local rings of small embedding codepth or small linking number, J. Algebra 118 (1988), no. 1, 162–204. [MR ]{}[961334 (89k:13013)]{} K. Beck and S. Sather-Wagstaff, A somewhat gentle introduction to differential graded commutative algebra, Connections Between Algebra, Combinatorics, and Geometry, Proceedings in Mathematics and Statistics, vol. 76, Springer, New York, Heidelberg, Dordrecht, London, 2014, pp. 3–99. [MR ]{}[3213517]{} O. Celikbas and S. Sather-Wagstaff, Testing for the [G]{}orenstein property, Collect. Math. 67 (2016), no. 3, 555–568. [MR ]{}[3536062]{} L. W. Christensen, Semi-dualizing complexes and their [A]{}uslander categories, Trans. Amer. Math. Soc. 353 (2001), no. 5, 1839–1883. [MR ]{}[2002a:13017]{} L. W. Christensen, H.-B. Foxby, and H. Holm, Derived category methods in commutative algebra, in preparation. L. W. Christensen and S. Sather-Wagstaff, Descent via [K]{}oszul extensions, J. Algebra 322 (2009), no. 9, 3026–3046. [MR ]{}[2567408]{} Y. F[é]{}lix, S. Halperin, and J.-C. Thomas, Rational homotopy theory, Graduate Texts in Mathematics, vol. 205, Springer-Verlag, New York, 2001. [MR ]{}[1802847]{} H.-B. Foxby, Gorenstein modules and related modules, Math. Scand. 31 (1972), 267–284 (1973). [MR ]{}[48 \#6094]{} A. Frankild and S. Sather-Wagstaff, Reflexivity and ring homomorphisms of finite flat dimension, Comm. Algebra 35 (2007), no. 2, 461–500. [MR ]{}[2294611]{} E. S. Golod, [$G$]{}-dimension and generalized perfect ideals, Trudy Mat. Inst. Steklov. 165 (1984), 62–66, Algebraic geometry and its applications. [MR ]{}[85m:13011]{} R. Hartshorne, Residues and duality, Lecture Notes in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. [MR ]{}[36 \#5145]{} S. Iyengar and S. Sather-Wagstaff, G-dimension over local homomorphisms. [A]{}pplications to the [F]{}robenius endomorphism, Illinois J. Math. 48 (2004), no. 1, 241–272. [MR ]{}[2048224 (2005c:13016)]{} T. Kawasaki, On [M]{}acaulayfication of [N]{}oetherian schemes, Trans. Amer. Math. Soc. 352 (2000), no. 6, 2517–2552. [MR ]{}[1707481 (2000j:14077)]{} H. Matsumura, Commutative ring theory, second ed., Studies in Advanced Mathematics, vol. 8, University Press, Cambridge, 1989. [MR ]{}[90i:13001]{} S. Nasseh and S. Sather-Wagstaff, Liftings and quasi-liftings of [DG]{} modules, J. Algebra 373 (2013), 162–182. [MR ]{}[2995021]{} [to3em]{}, Extension groups for [DG]{} modules, Comm. Algebra 45 (2017), no. 10, 4466–4476. [MR ]{}[3640821]{} [to3em]{}, Geometric aspects of representation theory for [DG]{} algebras: answering a question of [V]{}asconcelos, J. Lond. Math. Soc. (2) 96 (2017), no. 1, 271–292. [MR ]{}[3687949]{} W. V. Vasconcelos, Divisor theory in module categories, North-Holland Publishing Co., Amsterdam, 1974, North-Holland Mathematics Studies, No. 14, Notas de Matemática No. 53. \[Notes on Mathematics, No. 53\]. [MR ]{}[0498530 (58 \#16637)]{} J.-L. Verdier, Catégories dérivées, SGA 4$\frac{1}{2}$, Springer-Verlag, Berlin, 1977, Lecture Notes in Mathematics, Vol. 569, pp. 262–311. [MR ]{}[57 \#3132]{} [to3em]{}, Des catégories dérivées des catégories abéliennes, Astérisque (1996), no. 239, xii+253 pp. (1997), With a preface by Luc Illusie, Edited and with a note by Georges Maltsiniotis. [MR ]{}[98c:18007]{} S. Yassemi, G-dimension, Math. Scand. 77 (1995), no. 2, 161–174. [MR ]{}[97d:13017]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the distribution of angular positions and angular separations of unbound hypervelocity stars (HVSs). HVSs are spatially anisotropic at the 3-$\sigma$ level. The spatial anisotropy is significant in Galactic longitude, not in latitude, and the inclusion of lower velocity, possibly bound HVSs reduces the significance of the anisotropy. We discuss how the observed distribution of HVSs may be linked to their origin. In the future, measuring the distribution of HVSs in the southern sky will provide additional constraints on the spatial anisotropy and the origin of HVSs.' author: - 'Warren R. Brown, Margaret J. Geller, Scott J. Kenyon' - 'Benjamin C. Bromley' title: THE ANISOTROPIC SPATIAL DISTRIBUTION OF HYPERVELOCITY STARS --- INTRODUCTION ============ Unbound HVSs were predicted by @hills88 as the natural consequence of the massive black hole (MBH) in the Galactic center. Following the discovery of the first HVS [@brown05], observers have reported the discovery of at least 16 unbound HVSs and evidence for a similar number of bound HVSs ejected by the same mechanism [@hirsch05; @edelmann05; @brown06; @brown06b; @brown07a; @brown07b; @brown08c]. Follow-up observations of 4 HVSs establish they are main sequence B stars [@fuentes06; @bonanos08; @przybilla08; @przybilla08b; @lopezmorales08] like the S-stars orbiting Sgr A$^$ today [@ghez03; @eisenhauer05; @martins08]. Although not all unbound stars are necessarily HVSs – the massive B star HD 271791 is the first example of an unbound “hyper-runaway” ejected from the outer disk [@heber08; @przybilla08c] – runaway ejection velocities are limited to $\sim$300 km s$^{-1}$ for 3 M$_{\sun}$ stars [@leonard88; @leonard90; @leonard91; @leonard93; @portegies00; @davies02; @gualandris05]. Thus the 14 unbound 2.5-4 M$_{\sun}$ stars found in the @brown07b [@brown08c] targeted surveys are almost certainly HVSs ejected from the Galactic center. Remarkably, 8 of the 14 HVSs in the @brown07b [@brown08c] targeted surveys are located in just two constellations, Leo and Sextans, even though the surveys cover $1/5^{\rm th}$ of the sky. This spatial anisotropy is almost certainly linked to the origin of the HVSs. In §2 we show that the observed distribution of HVSs on the sky is anisotropic at the 3-$\sigma$ level. In §3 we discuss plausible explanations for the observed anisotropy of HVSs. OBSERVED ANISOTROPY =================== Sample ------ We consider the 14 HVSs from the combined surveys of @brown07b [@brown08c]. Our surveys use the MMT telescope to measure radial velocities for stars with the colors of 2.5-4 M$_{\sun}$ stars. Heliocentric velocities are converted to Galactocentric velocities assuming that the local rotation speed is 220 km s$^{-1}$ and that the Sun moves at (U,V,W)=(10, 5.2, 7.2) km s$^{-1}$ relative to the local standard of rest [@dehnen98]. The original HVS survey [@brown07b] is 100% complete for stars with $17<g'_0<19.5$ over 7300 deg$^2$ covered by the Sloan Digital Sky Survey Data Release 6. The new HVS survey [@brown08c] is 59% complete for stars with $19.5<g'_0<20.5$ over the same region of sky. Figure \[fig:polar\] plots the spatial distribution of stars with observed velocities. The combined HVS survey contains 693 stars and 14 HVSs in the 7300 deg$^2$ Sloan region covering the north Galactic cap. HVS2 [@hirsch05] is also located in this region (see Fig. \[fig:polar\]), however it falls outside our color / magnitude criteria. Thus we exclude HVS2 from this analysis. Significant Anisotropy ---------------------- Figure \[fig:gll\] plots the cumulative Galactic longitude and latitude distributions of the HVSs and the other survey stars. Kolmogorov-Smirnov (K-S) tests find 0.007 and 0.11 likelihoods that the HVSs are drawn from the same longitude and latitude distributions, respectively, as the survey stars. Thus the distribution of HVS longitudes appears anisotropic at the 3-$\sigma$ level. As a second test, we explore the anisotropy in terms of the distribution of angular separations, $\theta$, of the HVSs compared to the survey stars. Because the new HVS survey is not yet complete, we calculate $\theta$’s for all unique pairs of stars in the new and original surveys separately. The original survey includes HVS4-HVS10; a K-S test finds a 0.031 likelihood that those HVSs are drawn from the same distribution of $\theta$ as the original survey stars. The new survey includes HVS1 and HVS11-HVS16; a K-S test finds a $7\times10^{-9}$ likelihood that those HVSs are drawn from the same distribution of $\theta$ as the new survey stars. Figure \[fig:angcor\] plots the $\theta$’s of both surveys concatenated together. The likelihood of the combined set of HVSs is $7\times10^{-8}$; thus the distribution of HVS angular separations differs from the distribution of survey star angular separations at the 5-$\sigma$ level. As a third test, we measure the clustering of HVSs using the two-point angular correlation function $w(\theta)$. We use a Monte Carlo estimator [@landy93] and compare the observed HVSs against 10$^5$ sets randomly drawn from the survey region. The lower panel of Figure \[fig:angcor\] plots the resulting $w(\theta)$ in 15$^{\circ}$ bins. Errorbars are determined by Poisson statistics. HVSs are clustered at small angular separations $\theta<45^{\circ}$ and missing at large angular separations $\theta>60^{\circ}$ with $\sim$3.5-$\sigma$ significance. Velocity Dependence ------------------- We now consider the spatial anisotropy of lower velocity stars that may also be HVSs. @brown08c identify 4 “possible HVSs,” stars that are bound in the @kenyon08 potential model but unbound in the @xue08 potential model. Adding the 4 possible HVSs to the above analysis reduces the significance of the anisotropy to the 2-$\sigma$ level. There are also 8 possibly “bound HVSs,” stars with $v_{rf}>+275$ km s$^{-1}$ that are significant outliers from the overall velocity distribution [@brown08c]. Adding the bound HVSs to the above analysis yields an insignificant anisotropy. Thus lower velocity stars have a more isotropic distribution, a trend noted previously in @brown07a. HVS Pairs --------- There are 3 pairs of unbound HVSs with angular separations less than 3.5$^{\circ}$ (see Fig. \[fig:polar\]): HVS7 & HVS15 near ($l,b$)=(265$^{\circ}$, 55$^{\circ}$), HVS12 & HVS13 and HVS12 & HVS14 near ($l,b$)= (245$^{\circ}$, 52$^{\circ}$). Any physical association between the individual HVSs, however, appears unlikely. HVS7 & HVS15 are separated by $2.5^{\circ}$ but have velocities and distances that imply a $\sim$70 Myr difference in travel time from the Galactic Center [see Fig. 3 of @brown08c]. HVS12 has a 429 km s$^{-1}$ minimum rest-frame velocity very similar to that of HVS13 and HVS14, but it has half the distance of the other two HVSs. Thus none of the HVS pairs share a common ejection event. ORIGIN OF THE SPATIAL ANISOTROPY ================================ We observe that the spatial anisotropy of unbound HVSs is statistically robust, that lower velocity HVSs are systematically more isotropic, and that apparent close pairs of HVSs are physically unrelated. Possible explanations for the observations include: [Selection Effect.]{} Previously, we argued that the HVS anisotropy may be a selection effect of our magnitude-limited survey and the Sun’s off-center location in the Galaxy [@brown07a]. However, this selection effect can account only for an extra $\sim$10% HVSs in the anti-center hemisphere, not all of the HVSs in the anti-center hemisphere. Moreover, the observed HVSs cluster around $l=240^{\circ}$, not $l=180^{\circ}$. [Runaways.]{} Runaway stars like HD 271791 may contaminate the population of HVSs. However, we expect runaways ejected from the disk to have an isotropic distribution in Galactic longitude, as demonstrated by the @martin06 Hipparcos-selected sample of runaway B stars. Moreover, because runaways are systematically ejected at low velocities [e.g., @portegies00], the fastest runaways are those ejected in the direction of Galactic rotation and thus preferentially found at low Galactic latitudes. Thus the expected distribution of runaway longitudes and latitudes are contrary to the observed distribution of HVSs. [Large Scale Structure.]{} The distribution of Local Group dwarf galaxies is anisotropic, possibly due to a tidal origin [e.g., @metz08]. A tidal debris origin appears supported by the clumping of HVS travel times around 100-200 Myr, however the travel times are simply a product of the HVS’s $\sim$500 km s$^{-1}$ velocities and our magnitude-limited survey depth of 50-100 kpc. HVS travel times are in fact problematic for a tidal debris origin because the times are a significant fraction of the stars’ main sequence lifetimes, and multiple (gas-rich) tidal disruption events would be required to explain the full $2\times10^8$ yr span of HVS travel times. No dwarf galaxy in the Local Group travels with radial velocities comparable to the unbound HVSs; known dwarf galaxy remnants like the Sgr stream [@ibata94] are bound. We thus consider tidal debris an unlikely explanation for the observed set of HVSs [however, see @abadi08]. [Binary Black Hole.]{} While an equal-mass binary MBH is ruled out in the Galactic Center [@reid04], theorists speculate that the massive star clusters in the Galactic Center form intermediate mass black holes (IMBHs) in their cores. If such IMBHs exist, dynamical friction causes them to in-spiral into the central MBH, preferentially ejecting HVSs from their orbital planes. Thus the expected signature of a IMBH in-spiral is a ring of HVSs around the sky [@gualandris05; @levin06; @sesana06]. @baumgardt06 argue, however, that stellar interactions perturb the orbital plane of an in-spiraling IMBH; the resulting HVS distribution in this scenario may in fact be isotropic. Moreover, a single IMBH in-spiral event happens on timescales 10-100$\times$ shorter than the observed span of HVS travel times; multiple IMBH in-spiral events are required to explain the observed HVSs. [Galactic Center Structure.]{} The Galactic center contains many well-defined structures. As illustrated in @paumard06, the molecular gas circum-nuclear disk and the ionized northern arm are roughly aligned with the plane of the Milky Way. The gaseous minispiral is perpendicular to the plane of the Milky Way. Notably, the stellar disk 0.1 pc from the MBH is roughly perpendicular to the gaseous components [@lu08]. The stellar disk contains massive stars [@tanner06; @paumard06], possibly formed in-situ from a gas accretion disk [@genzel03; @levin03]. Dynamical interactions between a pair of stellar disks may scatter stars in towards the MBH, explaining both the S-stars and the HVSs [@lockmann08b; @perets08c]. Clearly, the Galactic center contains non-isotropic distributions of stars and gas which may provide a natural source for the observed anisotropy of HVSs ejected from the Galactic center. However, it is unclear if the observed structures can persist long enough to explain the anisotropic distribution of HVSs. CONCLUSION ========== Unbound HVSs are spatially anisotropic at the 3-$\sigma$ level. The anisotropy is most significant in Galactic longitude, and not in latitude. Lower velocity HVSs are systematically more isotropic, and apparent close pairs of HVSs are physically unrelated. The observed distribution of HVSs is linked to the origin of the HVSs. @abadi08 propose a tidal debris explanation, although this appears difficult to reconcile with all the observations. We investigate other physical models for the anisotropy in a separate paper. In the future, measuring the distribution of bound and unbound HVSs over the southern sky will allow us to better constrain the anisotropy and the origin of HVSs. This work is based on observations obtained at the MMT Observatory, a joint facility of the Smithsonian Institution and the University of Arizona. This research makes use of NASA’s Astrophysics Data System Bibliographic Services. We thank the anonymous referee and Oleg Gnedin for helpful comments. This work was supported by the Smithsonian Institution. [Facilities:*]{} [MMT (Blue Channel Spectrograph)]{} [45]{} natexlab\#1[\#1]{} , M. G., [Navarro]{}, J. F., & [Steinmetz]{}, M. 2008, , submitted , H., [Gualandris]{}, A., & [Portegies Zwart]{}, S. 2006, , 372, 174 , A. Z., [L[ó]{}pez-Morales]{}, M., [Hunter]{}, I., & [Ryans]{}, R. S. I. 2008, , 675, L77 , W. R., [Geller]{}, M. J., & [Kenyon]{}, S. J. 2008, , accepted , W. R., [Geller]{}, M. J., [Kenyon]{}, S. J., & [Kurtz]{}, M. J. 2005, , 622, L33 —. 2006, , 640, L35 —. 2006, , 647, 303 , W. R., [Geller]{}, M. J., [Kenyon]{}, S. J., [Kurtz]{}, M. J., & [Bromley]{}, B. C. 2007, , 660, 311 —. 2007, , 671, 1708 , M. B., [King]{}, A., & [Ritter]{}, H. 2002, , 333, 463 , W. & [Binney]{}, J. J. 1998, , 298, 387 , H., [Napiwotzki]{}, R., [Heber]{}, U., [Christlieb]{}, N., & [Reimers]{}, D. 2005, , 634, L181 , F. [et al.]{} 2005, , 628, 246 , C. I., [Stanek]{}, K. Z., [Gaudi]{}, B. S., [McLeod]{}, B. A., [Bogdanov]{}, S., [Hartman]{}, J. D., [Hickox]{}, R. C., & [Holman]{}, M. J. 2006, , 636, L37 , R. [et al.]{} 2003, , 594, 812 , A. M. [et al.]{} 2003, , 586, L127 , A., [Portegies Zwart]{}, S. P., & [Sipior]{}, M. S. 2005, , 363, 223 , U., [Edelmann]{}, H., [Napiwotzki]{}, R., [Altmann]{}, M., & [Scholz]{}, R.-D. 2008, , 483, L21 , J. G. 1988, , 331, 687 , H. A., [Heber]{}, U., [O’Toole]{}, S. J., & [Bresolin]{}, F. 2005, , 444, L61 , R. A., [Gilmore]{}, G., & [Irwin]{}, M. J. 1994, , 370, 194 , S. J., [Bromley]{}, B. C., [Geller]{}, M. J., & [Brown]{}, W. R. 2008, , 680, 312 , S. D. & [Szalay]{}, A. S. 1993, , 412, 64 , P. J. T. 1991, , 101, 562 , P. J. T. 1993, in ASP Conf. Ser. 45, Luminous High-Latitude Stars, ed. D. Sasselov, 360 , P. J. T. & [Duncan]{}, M. J. 1988, , 96, 222 —. 1990, , 99, 608 , Y. 2006, , 653, 1203 , Y. & [Beloborodov]{}, A. M. 2003, , 590, L33 , U., [Baumgardt]{}, H., & [Kroupa]{}, P. 2008, , 683, L151 , M. & [Bonanos]{}, A. Z. 2008, , 685, L47 , J. R., [Ghez]{}, A. M., [Hornstein]{}, S. D., [Morris]{}, M. R., [Becklin]{}, E. E., & [Matthews]{}, K. 2008, , accepted , J. C. 2006, , 131, 3047 , F., [Gillessen]{}, S., [Eisenhauer]{}, F., [Genzel]{}, R., [Ott]{}, T., & [Trippe]{}, S. 2008, , 672, L119 , M., [Kroupa]{}, P., & [Libeskind]{}, N. I. 2008, , 680, 287 , T. [et al.]{} 2006, , 643, 1011 , H. B., [Gualandris]{}, A., [Merritt]{}, D., & [Alexander]{}, T. 2008, astro-ph/0807.2340 , S. F. 2000, , 544, 437 , N., [Nieva]{}, M. F., [Heber]{}, U., & [Butler]{}, K. 2008, , 684, L103 , N., [Nieva]{}, M. F., [Heber]{}, U., [Firnstein]{}, M., [Butler]{}, K., [Napiwotzki]{}, R., & [Edelmann]{}, H. 2008, , 480, L37 , N., [Nieva]{}, M. F., [Tillich]{}, A., [Heber]{}, U., [Butler]{}, K., & [Brown]{}, W. R. 2008, , 488, L51 , M. J. & [Brunthaler]{}, A. 2004, , 616, 872 , A., [Haardt]{}, F., & [Madau]{}, P. 2006, , 651, 392 , A. [et al.]{} 2006, , 641, 891 , X. [et al.]{} 2008, , 684, 1143	{ "pile_set_name": "ArXiv" }
--- author: - 'Yiqian Shi$^{}$ and Bin Xu$^{\dagger}$' title: 'Gradient estimate of a Dirichlet eigenfunction on a compact manifold with boundary' --- [^1] Introduction ============ Let $(N,\ g)$ be an $n$-dimensional compact smooth Riemannian manifold with smooth boundary $\p N$ and $\Delta_N$ the positive Dirichlet Laplacian on $N$. Let $L^2(N)$ be the space of square integrable functions on $N$ with respect to the Riemannian density $dv(N)=\sqrt{{\bf g}(x)}\,dx:= \sqrt{\det\ (g_{ij})}\ dx$. Let $e_1(x),\,e_2(x),\,\cdots$ be a complete orthonormal basis in $L^2(N)$ for the Dirichlet eigenfunctions of $\D_N$ such that $0<\l_1^2< \l_2^2\leq \l_3^2\leq\,\cdots$ for the corresponding eigenvalues, where $e_j(x)$ ($j=1,2,\dots$) are real valued smooth function on $N$ and $\l_j$ are positive numbers. Also, let ${\bf e}_j$ denote the projection of $L^2(N)$ onto the 1-dimensional space ${\bf C}e_j$. Thus , an $L^2$ function $f$ can be written as $f=\sum_{j=0}^\infty {\bf e}_j(f)$, where the partial sum converges in the $L^2$ norm. Let $\l$ be a positive real number $\geq 1$. We define the spectral function and the unit band spectral projection operator $\chi_\l$ as follows: $$e(x,y,\l):=\sum_{\l_j\leq \l}\,e_j(x)\,e_j(y),$$ $${\chi}_\l f:=\sum_{\l_j\in (\l,\,\l+1]}{\bf e}_j(f)\ .$$ Grieser [@Gr] and Sogge [@S3] proved the $L^\infty$ estimate of $\chi_\l$, $$\label{equ:sup} \|\|{ \chi}_\l f\|\|_\infty\leq C\l^{(n-1)/2} \|\|f\|\|_2\,$$ where $\|\|f\|\|_r$ ($1\leq r\leq \infty$) means the $L^r$ norm of the function $f$ on $N$. In the whole of this paper $C$ denotes a positive constant which depends only on $N$ and may take different values at different places, if there is no otherwise stated. The idea of Grieser and Sogge is to use the standard wave kernel method outside a boundary layer of width $C \lambda^{-1}$ and a maximum principle argument inside that layer. By using the maximum principle argument and the estimate (\[equ:sup\]), Xu [@X2] proved the gradient estimate of $\chi_\lambda$ $$\label{equ:2infgrad} \|\|\nabla\, { \chi}_\l f\|\|_\infty\leq C\l^{(n+1)/2} \|\|f\|\|_2\ .$$ Here $\nabla$ is the Levi-Civita connection on $N$. In particular, $\nabla f=\sum_j\, g^{ij}\partial f/\partial x_j$ is the gradient vector field of a $C^1$ function $f$, the square of whose length equals $\sum_{i,j} g^{ij}(\p f/\p x_i)(\p f/\p x_j)$. One of his motivation is to prove the H[" o]{}rmander multiplier theorem on compact manifolds with boundary. Seeger and Sogge [@SS] firstly proved that theorem by using the parametrix of the wave kernel on manifolds without boundary. All the results mentioned in the introduction have their analog on compact manifolds without boundary. See the details in the introduction of Shi-Xu [@SX] and the references therein. In general, the method used in manifolds without boundary is not valid for the problems on manifolds with boundary. In particular, on manifolds with boundary the H[" o]{}rmander multiplier theorem cannot be obtained by the standard pseudo-differential operator calculus as done on manifolds without boundary, since the square root of the Dirichlet Laplaican is not a pseudo-differential operator any more and one cannot obtain the $L^\infty$ bounds for $\chi_\l$ and $\nabla \chi_\l$ only by using the Hadamard parametrix of the wave kernel. In the paper, by rescaling $\chi_\l f$ at the scale of $\l^{-1}$ both outside and inside the boundary layer of width $C\l^{-1}$, we obtain by elliptic apriori estimates a slightly stronger estimate than (\[equ:2infgrad\]) as follows:\ Putting $f(\cdot)=\sum_{\l_j\in (\l,\,\l+1]}\, e_j(x)e_j(\cdot)$ in (\[equ:sup\]), we obtain the uniform estimate of eigenfunctions for all $x\in N$, $$\sum_{\l_j\in (\l,\,\l+1]}\, \|e_j(x)\|^2\leq C\l^{n-1}.$$ Actually it can be proved that those two estimates are equivalent. By using the Cauchy-Scahwarz inequality, we obtain the gradient estimate (\[equ:2infgrad\]) from Theorem 1 together with the above inequality. Similarly, that estimate is equivalent to the uniform estimate for all $x\in N$, $$\label{equ:gradpointwise} \sum_{\l_j\in (\l,\,\l+1]}\, \|\nabla e_j(x)\|^2\leq C\l^{n-1}.$$ By the finite propagation speed of the wave equation, the asymptotic formula of derivatives of the spectral function $e(x,y,\l)$ in Theorem 1 [@XuB] which is proved by the standard wave kernel method, also holds for each interior point $x$ of $N$. Much more general asymptotic formulae are given in Theorems 1.8.5 and 1.8.7 of Safarov-Vassiliev [@SV]. In particular, we have the following asymptotic formula that as $\l\to\infty$ $$\sum_{\l_j\leq \l}\|\nabla\,e_j(x)\|^2=\frac{n\,\l^{n+2}}{2\,(4\pi)^{n/2}\,\Gamma(2+\frac{n}{2})}+{\rm O}_x(\l^{n+1}),$$ where the constant in the reminder term ${\rm O}_x(\l^{n+1})$ depends on the distance of $x$ to the boundary of $N$. Hence, the exponents of $\l$ in estimates (\[equ:2infgrad\]) and (\[equ:gradpointwise\]) are sharp at $x$ as $\l\to\infty$. For each point $z$ on the boundary of $N$, Ozawa [@Oz] used the heat kernel method to show the asymptotic formula that as $\l\to\infty$, $$\sum_{\l_j\leq \l}\left\|\frac{\p e_j}{\p \nu}(z)\right\|^2=\frac{\l^{n+2}}{(4\pi)^{n/2}\,\Gamma(2+\frac{n}{2})}+{\rm o}(\l^{n+2}),$$ where $\nu$ is the unit outer normal vector field on the boundary of $N$. Hence, (\[equ:2infgrad\]) and (\[equ:gradpointwise\]) are also sharp on the boundary.\ Putting $f=e_\l$ in the estimate (\[equ:grad2\]), we obtain the corollary.\ Actually, by the basic geometric property of nodal sets of an eigenfunction, we can find a complete picture for the $L^\infty$ norm of $\nabla e_\l$ in the following:\ \ The authors [@SX] proved the analog of Theorems 1 and 2 on compact manifolds without boundary. The proof in this paper is more complicated than there because we need to do analysis at points near the boundary. We believe that there also hold the analog for $k$-covariant derivatives $\nabla^k\,\chi_\l f$ and $\nabla^k \,e_\l$ on $N$. We plan to discuss this question in a future paper.\ Let $\psi_j$ be the normal derivative of $e_j$ at the boundary $\p N$ of $N$. The lower bound estimate $$\\|\psi_j\\|_{L^\infty(\p N)}\geq C\\|e_j\\|_{L^\infty(N)}$$ does not hold in general. Using Examples 3-5 in Hassell-Tao [@HT] and doing a little bit more computations, we can see that the above estimate does not hold on the flat cylinder, the hemisphere and the spherical cylinder. We hope to find a sufficient condition for the lower bound estimate in a future work.\ We conclude the introduction by explaining the organization of this paper. In Section 2, we show the lower bound of the gradient $\nabla\, e_\l$ by the basic geometrical property of the nodal set of eigenfunctions. In Section 3 we use the rescaling method and the H[" o]{}lder estimate about elliptic PDEs to show (\[equ:grad2\]) and the upper bound part of (\[equ:grad\]). The point is to do the rescaling both outside and inside the boundary layer of width $C\,\l^{-1}$. Lower bound of $\nabla e_\l$ ============================ The nodal set of an eigenfunction $e_\l$ of $\D_N$ is the zero set $$Z_{e_\l}:=\{x\in N: e_\l(x)=0\}.$$ A connected component of the open set $N\backslash Z_{e_\l}$ is called a nodal domain of the eigenfunction $e_\l$. We have the same definition for manifolds without boundary.\ (Br[" u]{}ning [@Br]) [Let $M$ be a compact Riemannian manifold without boundary. Let $\l\geq 1$ and $e_\l$ be an eigenfunction of the positive Laplacican $\D_M$: $\D_M\,e_\l=\l^2 e_\l$. Then there exists a constant $C$ only depending on $M$ such that each geodesic ball of radius $C/\l$ in $M$ must intersect the nodal set $Z_{e_\l}$ of $e_\l$.]{} A proof written in English is given by Zelditch in pp. 579-580 of [@Z].\ We need a manifold-with-boundary version of Lemma 1 as follows:\ \ We here adapt the proof of Zelditch [@Z] with a slight modification.\ [Step 1]{}We show the following fact: There exists a constant $C$ such that for each interior point $p$ of $N$ and each positive number $r>0$ satisfying that the distance $d(p,\,\p N)<r$ from $p$ to $\p N$ is less than $r$, the smallest Dirichlet eigenvalue $\l_1^2\bigl(B(p,\,r)\bigr)$ of the geodesic ball $B(x,\,r)$ is bounded from above by $C/r^2$. Since $d(p,\,\p N)>r$, we may assume that there exists a geodesic normal coordinate chart $(x_1,\dots, x_n)$ on the ball $B(p,\,r)$. Let $g$ be the Riemannian metric of $N$ on $B(p,\,r)$ with coefficients $g_{ij}=g(\p/\p x_i,\,\p/\p x_j)$ and $g_0$ be the Euclidean metric $dx_1^2+\cdots+dx_n^2$ on $B(x,\,r)$. Take $0<c_1<1$ depending only on $N$ so that the Euclidean ball $B(p,c_1r\,;\,g_0)$ is contained in the metric ball $B(p,r\,;\,g)$ of $N$. Since $c_1<1$, by the definition of Rayleigh quotient, $$\l_1^2\bigl(B(p,r\,;\,g)\bigr)\leq \l_1^2\bigl(B(p,c_1r\,;\,g)\bigr).$$ Since $N$ is compact, by comparing Rayleigh quotients, there exists $c_2>0$ depending only on $N$ such that $$\l_1^2\bigl(B(p,c_1r\,;\,g)\bigr)\leq c_2 \l_1^2\bigl(B(p,c_1r\,;\,g_0)\bigr).$$ On the other hand, by change of variables, we have $$\l_1^2\bigl(B(p,c_1r\,;\,g_0)\bigr)=\frac{\l_1^2\bigl(B(p,1\,;\,g_0)\bigr)}{(c_1r)^2}.$$ Combining the above three inequalities and setting $C=c_2c_1^{-2}\l_1^2\bigl(B(p,1\,;\,g_0)\bigr)$, we complete the proof.\ [Step 2]{} Take a geodesic ball $B(p,\,r)$ in $N$ such that $d(p,\,\p N)>r$. Suppose that it is disjoint from the nodal set $Z_{e_\l}$. Then it is completely contained in a nodal domain $D_j$ of $e_\l$. But $\l^2=\l_1^2(D_j)\leq \l_1^2\bigl(B(p,\,r)\bigr)\leq C/r^2$. Hence, $r\leq \sqrt{C}{\l}$. Taking $D=2\sqrt{C}$, we complete the proof.\ Take a point $x$ in $N$ such that $\|e_\l(x)\|=\\|e_\l\\|_\infty$. By the Dirichlet boundary condition, the distance $d$ from $x$ to $\p N$ is positive. [Case 1]{} Assume $d>D/\l$. Then there exists point $y$ in the geodesic ball $B(x,\,D/\l)$ with center $x$ and radius $D/\l$ such that $e_\l(y)=0$. We may assume $\l$ so large that there exists a geodesic normal chart $(r, \theta)\in [0,\,D/\l]\times {\Bbb S}^{n-1}(1)$ in the ball $B(x,\,D/\l)$. By the mean value theorem, there exists a point $z$ on the geodesic segment connecting $x$ and $y$ such that $$\left\|\frac{\p e_\l}{\p r}(z)\right\|\geq \frac{\l}{D}\|e_\l(x)\|=\frac{\l}{D}\\|e_\l\\|_\infty.$$ [Case 2]{} Assume $d\leq D/\l$. We may assume $\l$ so large that there exists a unique geodesic $\gamma:[0,\,d]\to N$ of arc length parameter connecting $x$ and $\p N$, $$\gamma(0)=x,\quad \gamma(d)\in \p N.$$ Since $e_\l(\gamma(d))=0$, by the mean value theorem, there exists $t_0$ in $(0,\,d)$ such that $$\left\|\frac{d e_\l\bigl(\gamma(t)\bigr)}{dt}(t_0)\right\|=\frac{\|e_\l(x)\|}{d}\geq \frac{\l}{D}\,\\|e_\l\\|_\infty.$$ Estimate for $\nabla \chi_\l f$ =============================== Outside the boundary layer -------------------------- Recall the principle: [On a small scale comparable to the wavelength $1/\l$, the eigenfunction $e_\l$ behaves like a harmonic function.]{} It was developed in H. Donnelly and C. Fefferman [@DF1] [@DF2] and N. S. Nadirashvili [@Na] and was used extensively there. Recently Mangoubi [@Ma1] applied this principle to studying the geometry of nodal domains of eigenfunctions. In this section, for a square integrable function $f$ on $N$ we give a modification of this principle, which can be applied to the Poisson equation $$\D_N\,\chi_\l f=\sum_{\l_j\in (\l,\,\l+1]}\, \l_j^2 {\bf e}_j(f)\quad {\rm in}\quad {\rm Int}(N)$$ with the Dirichlet boundary condition $\chi_\l f=0$ on $\p N$. In particular, in this subsection, we do the analysis outside the boundary layer $L_{1/\l}=\{z\in N:\,d(z,\,\p N)\leq 1/\l\}$ of width $1/\l$. Take an arbitrary point $p$ with $d(p,\,\p N)\geq 1/\l$. We may assume that $1/\l$ is sufficiently small such that there exists a geodesic normal coordinate chart $(x_1,\dots, x_n)$ on the geodesic ball $B(p,\,2/\l)$ in $N$. In this chart, we may identify the ball $B(p,\,2/\l)$ with the $n$-dimensional Euclidean ball ${\Bbb B}(2/\l)$ centered at the origin $0$, and think of the function $\chi_\l f$ in $B(p,\,2/\l)$ as a function in ${\Bbb B}(2/\l)$. Our aim in this subsection is to show the inequality $$\label{equ:outside2} \|(\nabla \chi_\l f)(p)\|\leq C\left(\l\\|\chi_\l f\\|_{L^\infty\bigl({\Bbb B}(2/\l)\bigr)}+\l^{-1}\\|\D_N\,\chi_\l f\\|_{L^\infty\bigl({\Bbb B}(2/\l)\bigr)}\right).$$ For simplicity of notions, we rewrite $u=\chi_\l f$ and $v=\D_N\chi_\l f$ in what follows. The Poisson equation satisfied by $u$ in ${\Bbb B}(1/\l)$ can be written as $$-\frac{1}{\sqrt{g}}\sum_{i,j}\,\p_{x_i}\left(g^{ij}\sqrt{g}\p_{x_j} u\right)=v.$$ Consider the rescaled functions $u_\l(y)=u(y/\l)$ and $v_\l(y)=v(y/\l)$ in the ball ${\Bbb B}(2)$. The above estimate we want to prove is equivalent to its rescaled version $$\label{equ:scale} \|(\nabla u_\l)(0)\|\leq C\left(\\|u_\l\\|_{L^\infty\bigl({\Bbb B}(2)\bigr)}+\l^{-2}\\|v_\l\\|_{L^\infty\bigl({\Bbb B}(2)\bigr)}\right).$$ On the other hand, the rescaled version of the Poisson equation has the expression, $$\label{equ:PoissonScaled} \sum_{i,j}\,\p_{y_i}\left(g^{ij}_\l\sqrt{g_\l}\p_{y_j} u_\l\right)=-\l^{-2}\,\sqrt{g_\l}\,v_\l,$$ where $g_{ij,\,\l}(y)=g_{ij}(y/\l)$, $g^{ij}_\l(y)=g^{ij}(y/\l)$ and $\sqrt{g_r}(y)=(\sqrt{g})(y/\l)$. For each $0<\alpha<1$, there exists $K>0$ such that the $C^\alpha$ norm of the coefficients $g^{ij}_\l\sqrt{g_\l}$, $\sqrt{g_\l}$ in $\BB(2)$ are bounded uniformly from above by $K$, and the smallest eigenvalue of the $n\times n$ matrix $(g^{ij}_\l\sqrt{g_\l})_{ij}$ in $\BB(2)$ bounds from below by $1/K$, for all $\l\geq 1$. By Theorem 8.32 in page 210 of Gilbarg-Trudinger [@GT], there exists constant $C=C(n, \, \alpha,\, K)$ such that $$\\|u_\l\\|_{C^{1,\,\alpha}\bigl({\Bbb B}(1)\bigr)}\leq C\left( \\|u_\l\\|_{L^\infty\bigl(\BB(2)\bigr)}+\l^{-2}\,\\|v_\l\\|_{L^\infty\bigl(\BB(2)\bigr)}\right),$$ This is stronger than the estimate (\[equ:scale\]). Therefore, we complete the proof of Theorem 1.1 outside the boundary layer $L_\l$. Inside the boundary layer ------------------------- Using the notions in subsection 3.1, We are going to prove the following estimate: $$\label{equ:inside} \\|\nabla u\\|_{L^\infty(L_{1/\l})}\leq C\left(\l \\|u\\|_\infty+\l^{-1}\\|v\\|_\infty\right),$$ with which combining (\[equ:outside2\]) completes the proof of Theorem 1.1. We may assume that $\l$ is sufficiently large so that there exists a geodesic normal coordinate chart $(z',\,z_n)$ on the boundary layer $L_{3/\l}=\{p\in N:\,d(p,\,\p N)\leq 3/\l\}$ with respect to the boundary $\p N$. Hence, for each point $(z',\,z_n)\in L_{3/\l}$, we have $0\leq z_n\leq 3/\l$ and $$d\bigl((z',\,z_n),\,\p N\bigr)=z_n.$$ For each point $q\in \p N$ and $r>0$, denote by $B_+(q,\,r)$ the set of points of $N$ with distance less than $r$ to $q$. Denote by $\BB_+(r)$ the upper half Euclidean ball $$\{x=(x_1,\dots,x_n)\in {\bf R}^n:\|x\|<r,\,x_n\geq 0\}$$ centered at the origin and with radius $r$. Then, for each $q\in \p N$, there exists a geodesic normal chart on $B_+(q,\,3/\l)$ such that the exponential map $\exp_q$ at $q$ gives a diffeomorphism from $\BB_+(3/\l)$ onto $B_+(q,\,3/\l)$. Since $\{B_+(q,\,2/\l):q\in \p N\}$ forms an open cover of $L_{1/\l}$, the question can be reduced to showing the analog of (\[equ:inside\]) on $B_+(q,\,2/\l)$ for each $q$. We only need to prove its equivalent rescaled version, $$\label{equ:insidescaled} \\|\nabla u_\l\\|_{L^\infty\bigl(\BB_+(2)\bigr)}\leq C\left( \\|u_\l\\|_{L^\infty\bigl(\BB_+(3)\bigr)}+\l^{-2}\\|v_\l\\|_{L^\infty\bigl(\BB_+(3)\bigr)}\right),$$ where $u_\l$ and $v_\l$ are the the rescaling function of $u$ an $v$, respectively. Observe that $u_\l$ and $v_\l$ satisfy the Poisson equation $(\ref{equ:PoissonScaled})$ in the upper half Euclidean ball $\BB(3)$ and the Dirichlet boundary condition, $$u_\l=0\quad {\rm on\ the\ portion}\quad \{x\in \BB(3):x_n=0\}$$ of the boundary of $\BB_+(3)$. For each $0<\alpha<1$, there exists $K>0$ such that the $C^\alpha$ norm of the coefficients $g^{ij}_\l\sqrt{g_\l}$, $\sqrt{g_\l}$ in $\BB_+(3)$ are bounded uniformly from above by $K$, and the smallest eigenvalue of the $n\times n$ matrix $(g^{ij}_\l\sqrt{g_\l})_{ij}$ in $\BB_+(3)$ bounds from below by $1/K$, for all $\l\geq 1$. By Theorem 8.36 in page 212 of Gilbarg-Trudinger [@GT], there exists constant $C=C(n, \, \alpha,\, K)$ such that $$\\|u_\l\\|_{C^{1,\,\alpha}\bigl({\Bbb B}_+(2)\bigr)}\leq C\left( \\|u_\l\\|_{L^\infty\bigl(\BB_+(3)\bigr)}+\l^{-2}\,\\|v_\l\\|_{L^\infty\bigl(\BB_+(3)\bigr)}\right).$$ This is a stronger estimate than (\[equ:insidescaled\]).\ Yiqian Shi is supported in part by the National Natural Science Foundation of China (No. 10671096 and No. 10971104) and Bin Xu by the National Natural Science Foundation of China (No. 10601053 and No. 10871184). [99]{} [^1]: $^*$ Department of Mathematics, University of Science and Technology of China, Hefei 230026 China.\ $^\dagger$ E-mail of the correspondent author: [email protected]	{ "pile_set_name": "ArXiv" }
=1200 = \#1[\#1]{} \#1[ \#1]{} \#1 [=1=0by1]{}=1 0.8plus 0.8 minus 0.8$\S$ [. \#1]{} 100000.6plus 0.8minus 0.6 =1 \#1 ** 8truept minus 8truept =0 $\S\S\;$$\bf\sectionlabel.\the\subsecnumber$ \#1 by1 100006truept minus 6truept \#1 [=1by1]{}=1 0.8plus 0.8 minus 0.8[Appendix . \#1]{} 0.8plus 0.8 minus 0.8 \#1 2.4 pt [.]{} \#1 2.4 pt[by1]{} \#1[[\[\#1\]]{}]{} \#1[\^[\#1]{}\_p]{} Determinants of Laplacians, the Ray-Singer Torsion on Lens Spaces and the Riemann zeta function 1.25 by 1.25 Charles Nash and D. J. O’ Connor Department of Mathematical PhysicsSchool of Theoretical Physics St. Patrick’s CollegeDublin Institute for Advanced Studies Maynooth10 Burlington Road IrelandDublin 4 Ireland 3[[Abstract*]{}: We obtain explicit expressions for the determinants of the Laplacians on zero and one forms for an infinite class of three dimensional lens spaces $L(p,q)$. These expressions can be combined to obtain the Ray-Singer torsion of these lens spaces. As a consequence we obtain an infinite class of formulae for the Riemann zeta function $\zeta(3)$. The value of these determinants (and the torsion) grows as the size of the fundamental group of the lens space increases and this is also computed. The triviality of the torsion for just the three lens spaces $L(6,1)$, $L(10,3)$ and $L(12,5)$ is also noted.]{} Topological phenomena are now known to play an important part in many quantum field theories. This is especially true of gauge theories. There are also [topological quantum field theories]{} in which the excitations are purely topological and the classical phase spaces of these theories usually reduce to a finite dimensional space: these spaces can be zero dimensional discrete sets, or whole moduli spaces. The semi-classical, or stationary phase, approximation to the functional integral of such a theory is then a weighted sum, or integral, over the [finite dimensional]{} phase space. In addition, for some of these theories, this approximation is [exact]{} providing thereby a reduction of the functional integral to a finite dimensional integral. If a topological quantum field theory contains a gauge field the reduced functional integral mentioned above often consists of sums or integrals over [flat]{} connections; the non-triviality of such connections is determined purely by their holonomy, and, if $A$ is a flat connection over a manifold $M$, its holonomy is an element of the fundamental group $\pi_1(M)$. This means that an ideal laboratory within which to study such theories is provided by taking the manifold $M$ to have a non-trivial fundamental group but to be otherwise topologically rather simple. An ideal way to do this is to take $M$ to be the quotient of a sphere $S^n$ by a finite cyclic group $G$. This quotient $S^n/G$, described in more detail below, is what is called a lens space, written as $L(p,q)$. In this paper we take $M$ to be a lens space on which is placed a topological field theory whose classical phase space consists of flat connections. Our approach is to take the model given by the field theory and analyse it in detail on a whole infinite class of lens spaces. We work in three dimensions and realise $M$ as the quotient of the manifold $S^3$ by the action of a discrete group $Z_{p}$. The resulting partition function on this manifold is a combinatorial invariant of the manifold known as the Ray-Singer torsion of the manifold. However the field theory gives this partition function as the ratio of a set of determinants. A standard technique in field theory has been to define these functional determinants through the analytic continuation of the zeta functions of the associated operators. In this work we investigate the individual determinants that arise and obtain highly explicit expressions for them. Our expressions have an intriguing structure of their own For example, on the lens space $L(2,1)$, we find that $$\eqalign{\ln\Det d^d_{0}&=-{3\zeta(3)\over2\pi^2}+\ln2\cr \ln\Det d^d_1&=-{3\zeta(3)\over\pi^2}-2\ln2\cr}\no$$ Similar, though more complicated, expressions occur for each of the lens spaces $L(p,1)$ for $p=3,4,\dots$. This in turn leads to non-trivial formulae for $\zeta(3)$: to give two examples we find that $$\eqalign{\zeta(3)&={2\pi^2\over7}\ln(2) -{8\over7}\int_0^{\pi/2}dz\,z^2\cot(z)\cr \zeta(3)&={2\pi^2\over13}\ln 3 -{9\over 13}\int_{0}^{\pi\over3}dz\,z(z+{\pi\over3})\cot(z) -{9\over 13}\int_{0}^{2\pi\over3}dz\,z(z-{\pi\over3})\cot(z) \cr}\no$$ these being the formulae that come from $L(2,1)$ and $L(3,1)$ respectively. The structure of the paper is as follows. In Section 2 we describe the topological field theory under consideration. In section 3 we define the Ray-Singer torsion and describe the lens spaces with which we work; we also carry out the non-trivial task of obtaining the eigenvalues and degeneracies for the Laplacians on $p$-forms acting on these spaces. Section 4 deals with the lens space $L(2,1)$ ($SO(3)$) and is a construction of the analytic continuation of the appropriate $p$-form zeta functions followed by a calculation of their associated determinants. Sections 5 and 6 describe the analogous calculation and expressions for the infinite classes of manifolds corresponding to $L(p,1)$ for $p$ odd and even respectively. Finally in section 7 we present our conclusions, some comments on the torsion of $L(p,q)$ for general $q$, and some graphical data for the resulting determinants and torsion. The torsion studied in this paper has its origins in the 1930’s, cf. Franz \[1\], where it was combinatorially defined and used to distinguish various lens spaces from one another. Given a manifold $M$ and a representation of its fundamental group $\pi_1(M)$ in a flat bundle $E$, this Reidemeister-Franz torsion is a real number which is defined as a particular product of ratio’s of volume elements $V^i$ constructed from the cohomology groups $H^i(M;E)$. Since volume elements are essentially determinants then, for any alternative definition of a determinant, an alternative definition of the torsion can be given. Now if one uses de Rham cohomology to compute $H^i(M;E)$ then these determinants become determinants of Laplacians $\Delta^E_p$ on $p$-forms with coefficients in $E$. But zeta functions for elliptic operators can be used to give finite values to such infinite dimensional determinants and so an analytic definition of the torsion results and this is the analytic torsion of Ray and Singer \[2,3,4\] given in the 1970’s; furthermore this torsion was proved by them to be independent of the Riemannian metric used to define the Laplacian’s $\Delta^E_p$. This analytic torsion coincided, for the case of lens spaces, with the combinatorially defined Reidemeister-Franz torsion. Finally Cheeger and Müller \[5,6\] independently proved that the analytic Ray-Singer torsion coincides with the combinatorial Reidemeister-Franz torsion in all cases. Infinite dimensional determinants also occur naturally in quantum field theories when computing correlation functions and partition functions. In 1978 Schwarz \[7\] showed how to construct a quantum field theory on a manifold $M$ whose partition function is a power of the Ray-Singer torsion on $M$. Schwarz’s construction uses an Abelian gauge theory but in three dimensions a non-Abelian gauge theory—the $SU(2)$ Chern-Simons theory—can be constructed and has deep and important properties established by Witten in 1988: Its partition function is the Witten invariant for the three manifold $M$ and the correlation functions of Wilson loops give the Jones polynomial invariant for the link determined by the Wilson loops—cf. \[8,9\]. Finally the weak coupling limit of the partition function is a power of the Ray-Singer torsion. To define the Ray-Singer torsion, or simply torsion, we take a closed compact Riemannian manifold $M$ over which we have a flat bundle $E$. Let $M$ have a non-trivial fundamental group $\pi_1(M)$ which is represented on $E$—this latter property arises very naturally in the physical gauge theory context where it corresponds simply to the space of flat connections all of whose content resides in their holonomy—In any case the torsion is then the real number $T(M,E)$ where $$\ln T(M,E)=\sum_0^n(-1)^qq\ln \Det \Delta^E_q,\quad n=\dim M\no$$ The metric independence of the torsion requires that we assume, in the above definition, that the cohomology ring $H^(M;E)$ is trivial; this means that the Laplacians $\Delta^E_q$ have empty kernels and so are strictly positive definite. Given this fact one may use zeta functions to define $ \Det \Delta^E_q$ in the standard way. Recall that if $P$ is a positive elliptic differential or pseudo-differential operator with spectrum $\{\mu_n\}$ and degeneracies $\{\Gamma_n\}$ then its associated zeta function $\zeta_P(s)$ is a meromorphic function of $s$, regular at $s=0$, which is given by $$\zeta_P(s)=\sum_{\mu_n}{\Gamma_n\over\mu_n^s}\no$$ and its determinant $\Det P$ is defined by $$\ln\Det P=-\left.{d\zeta_P(s)\over ds}\right\vert_{s=0}\no$$ Using this we have $$\ln T(M,E)= -\sum_0^n(-1)^qq\left.{d\zeta_{\Delta_q^{E}}(s)\over ds}\right\vert_{s=0}\no$$ Quantum field theories of the type alluded to above are usually referred to as topological quantum field theories or simply topological field theories. It turns out that more than one topological field theory can be used to give the torsion, for an excellent review of this question cf. Birmingham et al. \[10\]. For example one can take the action $$S[\omega]=i\int_{M}\omega_{n}d\omega_{n},\qquad \dim M=2n+1\no$$ where $\omega_n$ is an $n$-form. The partition function is then $$Z[M]=\int {\cal D}\omega\mu[\omega]\exp[S[\omega]] \no$$ $S[\omega] $ has a gauge invariance whereby $S[\omega]=S[\omega+d\lambda]$ and therefore to define the partition function it is necessary to integrate over only inequivalent field configurations. The measure ${\cal D}\omega\mu[\omega]$ thus contains functional delta functions which constrain the integration and play the role of gauge fixing, together with their associated determinants. This measure can be constructed using, for example, the Batalin-Vilkovisky BRST construction \[11,12\]. We shall be concerned here with the special situation of three dimensions and with the case where the three manifold $M$ is a lens space. The topological field theory of interest to us in this paper is given by the action $$S[\omega]=i\int_{M}\omega_1 d\omega_1 \no$$ where $\omega_1$ is now a $1$-form. To construct the integration measure we will follow the Batalin-Vilkovisky BRST construction \[11,12\]. The essential element of this construction is what is termed a gauge Fermion” whose BRST variation gives the gauge fixing and ghost portion of the BRST invariant action. Integrating out these fields yields the contribution $\mu[\omega]$ to the measure. The gauge Fermion is constructed by choosing a gauge fixing for the field $\omega_1$ (which we take to be $\dstar \omega=0$), and multiplying the condition by an anti-ghost $c_{\bar{0}}$, which is a $3$-form denoted by its conjugated Poincaré dual label, this indicates its anti-ghost nature also. Thus the gauge Fermion is given by $$\Psi = c_{\specialbar {0}} \dstar \omega_{1}\no$$ The associated BRST variations of these fields are $$\eqalign{\delta\omega_1 &= -d c_{0},\qquad \delta c_{0}=0\cr \delta c_{\specialbar {0}}&=i\omega_{\specialbar {0}} \qquad\delta \omega_{\specialbar {0}}=0\cr }\no$$ With these definitions it is easy to check that $\delta^2=0$. The BRST gauge fixed action is then $${\cal L}= i\omega_1d\omega_1+\delta\Psi$$ which expands to $${\cal L} = i\omega_1 d \omega_1 - c_{\specialbar {0}}\dstar dc_{0} + i\omega_{\specialbar {0}}\dstar \omega_{1} \no$$ If we integrate out all fields as they appear the resulting partition function is $$Z = (\Det L_{-})^{-{1\over2}}\Det \dstar d_{0} \no$$ where the operator $L_{-}$ is obtained by integrating out the $\omega_1$ fields and is a linear operator acting on odd forms. The partition function is therefore $$Z={\Det\Delta_{0}\over {\Det\Delta_{1}}^{1\over 4}{\Det\Delta_{3}}^{1\over 4}}$$ Using Poincaré duality the logarithm of this partition function is then given by $$\ln Z={1\over 4}\left(3\ln\Det \Delta_{0} -\ln\Det \Delta_{1}\right)$$ and we see it is proportional to the logarithm of the Ray-Singer torsion. Our task in what follows is to evaluate the individual components of this expression both for their usefulness in their own right and to verify that the combined result agrees with the Ray-Singer torsion. We do this in the restricted setting where $M$ belongs to a class of three dimensional lens spaces. In the next section we specify the lens spaces that we work with and obtain the eigenvalues and their degeneracies of the Laplacians on these spaces. We now want to turn to field theories defined on lens spaces—for general background on lens spaces cf. \[3,4\] and references therein—briefly, a lens space can be constructed as follows: Take an odd dimensional sphere $S^{2n+1}$, considered as a subset of ${\bf C}^n$, on which a finite cyclic group of rotations $G$, say, acts. The quotient $S^{2n+1}/G$ of the sphere under this action is a lens space. More precisely, suppose that $G$ is of order $p$, $(z_1,\ldots,z_n)\in {\bf C}^n$ and the group action takes the form $$\eqalign{{}&(z_1,\ldots,z_n)\longmapsto (\exp(2\pi i q_1/p)z_1,\ldots, \exp(2\pi i q_n/p)z_n)\cr {}&\hbox{with }q_1,\dots,q_n\quad\hbox{integers relatively prime to }p\cr} \no$$ then the quotient $S^{2n+1}/G$ is a lens space often denoted by $L(p;q_1,\dots,q_n)$. A formula for the torsion of these spaces was first worked out by Ray \[2\]. To our knowledge however there is no computation of the individual determinants of Laplacians on these spaces in the literature. Since these are of independent field theoretic significance and from these the torsion is constructed it is instructive to examine these separately and construct the torsion from them. This we will proceed to do in the next sections focusing on the situation that obtains when $n=2$ and $G$ is the group ${\bf Z}_p\equiv {\bf Z}/p{\bf Z}$. For simplicity we shall denote the resulting lens space $S^3/{\bf Z}_p=L(p;1,1)$ by $L(p)$, we shall also denote the lens space $L(p,1,q)$ by $L(p,q)$; in passing we note that when $p=2$ we have $L(2)={\bf R}P^3\simeq SO(3)$. The group action above defines a representation $V$, say, of $\pi_1(L(p))$ and also determines a flat bundle $F=(V\times S^3)/Z_p$, over $L(p)$. It is the determinants of Laplacians and the resulting torsion of this $F$ over $L(p)$ with which we are concerned here. Using zeta functions the torsion of these lens spaces is therefore given by $$\ln T(L(p),F)=-\sum_0^3(-1)^qq\left.{d\zeta_{\Delta_q^{F}}(s)\over ds}\right\vert_{s=0}\no$$ As an aid to the calculation of $\ln T(L(p),F)$ it is useful to introduce the notation $$\eqalign{\tau(p,s)&=-\sum_0^3(-1)^qq\zeta_{\Delta_q^{F}}(s)\cr T(p)&=T(L(p),F)\cr}$$ The relationship between the two functions being clearly $$\ln T(p)=\left.{d \tau(p,s) \over ds}\right\vert_{s=0}\no$$ For $\tau(p,s)$ itself we now have $$\eqalign{\tau(p,s)&=\zeta_{\Delta_1^{F}}(s)-2\zeta_{\Delta_2^{F}}(s)+ 3\zeta_{\Delta_3^{F}}(s)\cr {}&=3\zeta_{\Delta_0^{F}}(s)-\zeta_{\Delta_1^{F}}(s),\quad \hbox{using Poincar\'e duality}\cr}\no$$ Combining the standard decomposition $\Delta_p=(\dstar d+d\dstar)_p$, with the fact that $\ker\dstar\cap\ker d=\emptyset$, we further obtain the formula $$\tau(p,s)=2\zeta_{{\dstar d}_0}(s)-\zeta_{{\dstar d}_1}(s)\no$$ We now simplify our notation by labelling $$\tau_{+}(p,s)=2\zeta_{{\dstar d}_0}(s),\qquad \tau_{-}(p,s)=\zeta_{{\dstar d}_1}(s)$$ For the individual zeta functions we denote the eigenvalues and their degeneracies by $\lambda_n(q,p)$ and $\Gamma_n(q,p)$ respectively giving the expressions $$\tau_{+}(p,s)=2\sum_n{\Gamma_n(0,p)\over \lambda_n^s(0,p)},\qquad \tau_{-}(p,s)=\sum_n{\Gamma_n(1,p)\over \lambda_n^s(1,p)}\no$$ It remains to compute these eigenvalues and degeneracies. The former are actually independent of $p$ and are fairly easily calculated by the technique of starting with harmonic forms in ${\bf R}^{2n}$ and then restricting successively to $S^{2n+1}$ and $L(p)$. In any case they are given by $$\eqalign{\lambda_n(0,p)&=n(n+2),\;n=1,2,\dots\cr \lambda_n(1,p)&=(n+1)^2,\;n=1,2,\dots\cr}\no$$ To calculate the degeneracies is more difficult; we make use of the fact that $S^3$ is a group manifold and proceed as follows: Consider the Laplacians $\dstar d_q$ on $S^3$, and $\dstar d_q^{F}$ on $L(p)$ also, if $\lambda$ is an eigenvalue, denote the corresponding eigenspaces by $\Lambda_q(\lambda)$ and $\Lambda^{F}_q(\lambda)$ respectively. Let $$v(z)\in \Lambda_q(\lambda),\;\hbox{with }z\in S^3\subset {\bf C}^2, \qquad \hbox{and }g\in {\bf Z}_p,\; \hbox{where }g\equiv\exp[2\pi i j/p],\;0\le j\le (p-1)\no$$ The element $g$ acts on $v(z)$ to give $g\cdot v(z)$ where $$\eqalign{g\cdot v(z)&=v(gz)\cr gz&=(\exp[2\pi i j/p]z_1,\exp[2\pi i j/p]z_2)\cr}\no$$ The above definitions allow us to define the projection $P(\lambda)$ on $\Lambda_q(\lambda)$ by $$P(\lambda)v={1\over p}\sum_{g\in {\bf Z}_p}\exp[-2\pi i j/p]g\cdot v\no$$ Evidently $$[P(\lambda),\dstar d_q]=0\no$$ and so $P(\lambda)$ projects the space $\Lambda_q(\lambda)$ onto the space $\Lambda_q^{F}(\lambda)$. Finally this means that we obtain a formula for the degeneracy $\Gamma_n(q,p)$, namely $$\eqalign{\Gamma_n(q,p)&= \tr\left(\left.P\right\vert_{\Lambda_q^{F}(\lambda)}\right)\cr {}&={1\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\,\tr\left(\left.g\right\vert_{\Lambda_q^{F}(\lambda)}\right)\cr }\no$$ To actually apply this formula we now add in the fact that $S^3$ is the group manifold for $SU(2)$. The Peter–Weyl theorem tells us, in this case where all representations are self-conjugate, that $$L^2(S^3)=L^2(SU(2))={\textstyle \bigoplus\limits_{\mu}}\,\,c_{\mu}D_{\mu}= {\textstyle \bigoplus\limits_{\mu}}\,\,D_\mu\otimes D_\mu\no$$ where $c_{\mu}$ measures the multiplicity of the representation $\mu$ which must therefore be $\dim D_\mu$. But Hodge theory gives us the alternative decomposition $$L^2(S^3)={\textstyle \bigoplus\limits_{\lambda}}\,\,\Lambda_0(\lambda)\no$$ In addition the Casimir operator for $SU(2)$ is a multiple of the Laplacian and, if the representation label $\mu$ is taken to be the usual half-integer $j$, then we know that this Casimir has eigenvalues $j(j+1)$, and also that $\dim D_j=2j+1$. These facts identify the Laplacian $\Delta_0=\dstar d_0$ as four times the Casimir and identify $\Lambda_0(\lambda)$ as $\dim D_j$ copes of $D_j$. Thus if we set $n=2j$, so that $n$ is always integral, then we have the degeneracy formula $$\Gamma_n(0,p)={(n+1)\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\,\chi^{n/2}(2\pi j/p)\no$$ where $\chi^j(\theta)$ denotes the $SU(2)$ character, on $D_j$, for rotation through the angle $\theta$; i.e. $$\chi^j(\theta)={\sin((2j+1)\theta)\over\sin(\theta)}\no$$ Hence our explicit degeneracy formula for $0$-forms on $L(p)$ is $$\Gamma_n(0,p)={(n+1)\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\,{\sin(2\pi(n+1)j/p)\over\sin(2\pi j/p)}\no$$ We now have to find the analogous formula for the $1$-forms. The formula that results is $$\Gamma_n(1,p)={1\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\left\{n\chi^{(n+1)/2}(2\pi j/p)+(n+2)\chi^{(n-1)/2}(2\pi j/p)\right\}\no$$ or, more explicitly, $$\Gamma_n(1,p)={1\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\left\{n{\sin(2\pi(n+2)j/p)\over\sin(2\pi j/p)}+ (n+2){\sin(2\pi nj/p)\over\sin(2\pi j/p)}\right\}\no$$ To simplify the notation we introduce the $p$-averaged character’ $\charav{j}$ which we define by $$\charav{j}={1\over p}\sum_{j=0}^{(p-1)}\exp[-2\pi i j/p]\,\chi^j(2 pi j/p)\no$$ Finally this gives us a concrete expression for $\tau(p,s)$, i.e. $$\eqalign{\tau(p,s)&= \sum_n\left\{{2(n+1)\charav{n/2}\over \{n(n+2)\}^{2s}} -{n\charav{(n+1)/2}+(n+2)\charav{(n-1)/2}\over (n+1)^{2s}}\right\}\cr {}&=\tau_+(p,s)-\tau_-(p,s)\cr \hbox{where}\quad&\cr \tau_+(p,s)&=\sum_n{2(n+1)\charav{n/2}\over \{n(n+2)\}^{s}}, \quad \tau_-(p,s)=\sum_n{n\charav{(n+1)/2}+(n+2)\charav{(n-1)/2}\over (n+1)^{2s}}\cr}\no$$ To make further progress towards a computation of the determinants and torsion we need to be able to evaluate these $p$-averaged characters. This is a somewhat non-trivial combinatorial task but this task is eased if we use for $\chi^j(\theta)$, the alternative expression $$\chi^j(\theta)=\sum_{m=-j}^j\exp[2 i m\theta]\no$$ It is also necessary to divide $n$ up into its conjugacy classes mod $p$ by writing $$n=pk-j,\;k\in{\bf Z},\, j=0,1,\ldots,(p-1)\no$$ We eventually discover that $$\charav{(pk-j)/2}= \cases{ \cases{k&for $j=0,2,\ldots,(p-1)$\cr k&for $j=1$\cr (k-1)&for $j=3,5,\ldots,(p-2)$\cr} &if $p$ is odd\cr \strut&\null\cr \cases{0&for $j=0,2,\dots,(p-2)$ \cr 2k&for $j=1$\cr (2k-1)&for $j=3,5,\ldots,(p-1)$\cr} &if $p$ is even\cr} \no$$ We now lack only one ingredient among those necessary for a calculation of the determinants and the resulting torsion: this is the construction of the analytic continuation of the series for $\tau(p,s)$. We shall construct this in the next section. The technique we shall use will be more easily followed if we first use it in a more simple case. Thus, to begin with, we set $p=2$ and then construct the continuation. The series to be continued are $$\tau_{+}(p,s)=\sum_{n=1}^{\infty} {2(n+1)\charav{n/2}\over\{n(n+2)\}^{s}}$$ $$\tau_{-}(p,s)=\sum_{n=1}^{\infty} {n\charav{(n+1)/2}+(n+2)\charav{(n-1)/2}\over(n+1)^{2s}}$$ and their difference which leads to the torsion $$\tau(p,s)=\sum_n\left\{{2(n+1)\charav{n/2}\over \{n(n+2)\}^{s}} -{n\charav{(n+1)/2}+(n+2)\charav{(n-1)/2}\over(n+1)^{2s}}\right\} \no$$ These already converges for $\hbox{Re}\, s>3/2$; however a calculation of the determinants and the torsion requires us to work at $s=0$, hence we see the need for, and the extent of, the analytic continuation. Our interest in this section for illustrative purposes is in the case $p=2$ where we have $$\eqalign{\tau(2,s)&=\tau_+(2,s)-\tau_-(2,s)\cr {}&=\sum_n\left\{{2(n+1)\left\langle\chi^{n/2} \right\rangle_2\over \left\{n(n+2)\right\}^s}- {n\left\langle\chi^{(n+1)/2}\right\rangle_2+ (n+2)\left\langle\chi^{(n-1)/2}\right\rangle_2 \over(n+1)^{2s}}\right\}\cr}\no$$ But using we find that $$\eqalign{\left\langle\chi^{(n+1)/2}\right\rangle_2&= \left\langle\chi^{(2k-j+1)/2}\right\rangle_2,\quad (n=2k-j)\cr \strut&{}\cr {}&=\cases{0,& $j=1$\cr 2k+2,& $j=0$\cr} \quad\equiv\cases{0,& $n$ odd\cr 2k+2,& $n$ even\cr}\cr}\no$$ Similarly $$\eqalign{\left\langle\chi^{(n-1)/2}\right\rangle_2&= \left\langle\chi^{(2k-j)/2}\right\rangle_2,\quad (n=2k-j)\cr \strut&{}\cr {}&=\cases{2k,& $j=1$\cr 0,& $j=0$\cr} \quad\equiv\cases{(n+1),& $n$ odd\cr 0,& $n$ even\cr}\cr}\no$$ Thus $\tau(2,s)$ becomes $$\eqalign{\tau(2,s)&=\tau_+(2,s)-\tau_-(2,s)\cr {}&=\sum_{n\;{\rm odd}}{ 2(n+1)^2\over\left\{n(n+2)\right\}^s} -\sum_{n\;{\rm even}}{2n(n+2)\over(n+1)^{2s}}\cr}\no$$ Setting $n=(2m-1)$ in $\tau_+(2,s)$ and $n=2m$ in $\tau_-(2,s)$ we have $$\tau_{+}(2,s)=\sum_{m=1}^\infty{8 m^2\over(4m^2-1)^s}, \qquad\tau_{-}(2,s)=\sum_{m=0}^\infty{4m(2m+2)\over (2m+1)^{2s}}$$ and $$\tau(2,s)=\sum_{m=1}^\infty{8 m^2\over(4m^2-1)^s} -\sum_{m=0}^\infty{2\over (2m+1)^{(2s-2)}} +\sum_{m=0}^\infty{2\over (2m+1)^{2s}}\no$$ Now if we use the fact that $$\sum_{n=1,3,5,\dots}{1\over n^s}=(1-2^{-s})\zeta(s)\no$$ where $\zeta(s)$ is the usual Riemann zeta function then we get $$\tau_{-}(2,s)=2(1-2^{-(2s-2)})\zeta(2s-2)-2(1-2^{-2s})\zeta(2s)\no$$ and denoting $$A_{2}(m,0,s)={{(2m)}^2\over{\left\{{(2m)}^2-1\right\}}^s}, \qquad {\rm and}\qquad A_{2}(0,s)=\sum_{m=1}^{\infty}A_{2}(m,0,s)$$ (This notation is used to agree with the general case to be discussed in the next section. See also Appendix A.) Thus we have $$\tau_{+}(2,s)=2A_{2}(0,s)$$ and these combine to give $$\tau(2,s)=2A_{2}(0,s)-2(1-2^{-(2s-2)})\zeta(2s-2)+2(1-2^{-2s})\zeta(2s)\no$$ Since the terms involving the Riemann zeta function already have a well defined continuation it remains to continue $A_{2}(0,s)$. Now $$\eqalign{A_{2}(m,0,s)&={4m^2\over(4m^2-1)^s}={4m^2\over (4m^2)^s} \left(1-{1\over4m^2}\right)^{-s}\cr {}&={1\over(4m^2)^{(s- 1)}}\left\{1+{s\over4m^2}+\cdots\right\}\cr {}&={1\over(4m^2)^{(s-1)}}+{s\over(4m^2)^s}+R(m,s), \qquad(\hbox{def. of }R(m,s))\cr}\no$$ So that the remainder term $R(m,s)$ is given by $$\eqalign{R(m,s)&=A_{2}(0,m,s)-{1\over(4m^2)^{(s-1)}} -{s\over(4m^2)^s}\cr &={4m^2\over(4m^2-1)^s}-{1\over(4m^2)^{(s-1)}}- {s\over(4m^2)^s}\cr}\no$$ The definition of the remainder term is chosen to ensure that $$\left\vert R(m,s)\right\vert\le{(\ln m)^\alpha\over m^2}\no$$ and this has the vital consequence that the operations $d/ds$ (at $s=0$) and $\sum_m$ [commute ]{} when applied to $R(m,s)$. Defining $$R(s)=\sum_{m=0}^\infty R(m,s)\no$$ allows us to tidy our expressions up somewhat. Collecting our regulated expressions we therefore have $$\eqalign{\tau_{+}(2,s)&={8\over{4}^{s}}\zeta(2s-2) +{2s\over 4^s}\zeta(2s)+2 R(s)\qquad {\rm and}\cr \tau_{-}(2,s)&=2(1-2^{-(2s-2)})\zeta(2s-2)-2(1-2^{-2s})\zeta(2s)\cr}\no$$ In fact the expression for $\tau(2,s)$ can be further tidied up to give $$\tau(2,s)=2\left\{{8\over4^{(s)}}-1\right\}\zeta(2s-2) +2\left\{1+{(s-1)\over4^s}\right\}\zeta(2s)+2R(s)\no$$ The series for $R(s)$ is [guaranteed]{} to be convergent and the analytic continuation is now complete. Evaluating our expressions at $s=0$ we find $$\eqalign{\tau_{+}(2,0)&=8\zeta(-2) +2R(0) \qquad \tau_{-}(2,0)=-6\zeta(-2)\cr &{\rm and}\qquad \tau(2,0)=14\zeta(-2)\cr }\no$$ Observe that with our continuation $R(0)$ is automatically zero. Thus noting also that $\zeta(-2)=0$, we conclude $$\tau_{+}(2,0)=0\, , \qquad \tau_{-}(2,0)=0\, , \qquad and \qquad\tau(2,0)=0\no$$ That $\tau_{\pm}(p,0)=0$ is quite generally true for arbitrary $p$; we shall see this in the next section and this agrees with general considerations for generalised zeta functions of second order operators on compact odd dimensional manifolds. We can now take the final step which is to differentiate and obtain $\tp(2,0)$ and $\tm(2,0)$, which we denote by $\tp(2)$, $\tm(2)$ respectively, and hence the torsion $T(2)$. The resulting expressions are $$\tp(2)= 16\zeta '(-2)+2\zeta(0)+2R^\prime(0)$$ $$\tm(2)= -12\zeta '(-2)-2\ln4\zeta(0)$$ and for the torsion $$\ln T(2)={d\tau(2,0)\over ds}=28\zeta^\prime(-2) +2(1+\ln4)\zeta(0)+2R^\prime(0)\no$$ But $$\zeta(0)=-1/2,\qquad\hbox{ and }\qquad \zeta^\prime(-2)=- {\zeta(3)\over4\pi^2}\quad\hbox{from the functional relation}\no$$ and by our remark above concerning the motive for our choice of definition for $R(m,s)$ we have $$\eqalign{R^\prime(0)&= {d\over ds}\sum_m\left. R(m,s)\right\vert_{s=0}\cr \Rightarrow R^\prime(0)&=\sum_m\left.{d R(m,s)\over ds}\right\vert_{s=0}\cr &=\sum_m\left[4m^2\left\{\ln(4m^2)- \ln(4m^2-1)\right\}-1\right]=- \sum_m\left[4m^2\ln(1-1/4m^2)+1\right]\cr}\no$$ Hence $$\eqalign{ \tp(2)&=-{4\over\pi^2}\zeta(3)-1 -2\sum_m\left[4m^2\ln(1-1/4m^2)+1\right]\cr \tm(2)&={3\over\pi^2}\zeta(3)+2\ln2\qquad {\rm and}\cr \ln T(2)&=-{7\over\pi^2}\zeta(3)-1-2\ln(2)- 2\sum_m\left[4m^2\ln(1-1/4m^2)+1\right]}\no$$ However the series for $R^\prime(0)$ can be expressed as a trigonometric integral; in fact, as a special case of more general results which will be derived below, we have $$\sum_{m=1}^\infty\left[4m^2\ln(1-1/4m^2)+1\right]=-{1\over2}+ {4\over\pi^2}\int_0^{\pi/2}dz\,z^2\cot(z)\no$$ which means that $$\eqalign{\tp(2)&=-{4\over \pi^2}\zeta(3) -{8\over\pi^2}\int_{0}^{\pi\over2}dz z^2\cot(z)\cr \tm(2)&={3\over\pi^2}\zeta(3)+2\ln2\qquad {\rm and}\cr \ln T(2)&=-{7\over\pi^2}\zeta(3)-2\ln(2)- {8\over\pi^2}\int_0^{\pi/2}dz\,z^2\cot(z)}\no$$ The formula above for $T(2)$ can be pushed even further: By using it with Ray’s expression \[2\] for the torsion we can deduce that $$\eqalign{\ln T(p)&=-{4\over p}\sum_{j=1}^{(p-1)}\sum_{k=1}^p\cos({2jk\pi\over p})\ln(2\sin({2k\pi\over p}))\exp[{2k\pi i\over p}]\cr &=-4\ln\left[2 \sin({\pi\over p})\right]\cr}\no$$ which, for $p=2$, becomes simply $$\ln T(2)=-4\ln(2)\no$$ Hence we straightaway have the identity $$-4\ln(2)=-{7\over\pi^2}\zeta(3)-2\ln(2)- {8\over\pi^2}\int_0^{\pi/2}dz\,z^2\cot(z)\no$$ Or $$\zeta(3)={2\pi^2\over7}\ln(2)-{8\over7}\int_0^{\pi/2}dz\,z^2\cot(z)\no$$ In other words our computation of the torsion has given us a formula for the zeta function at its first odd argument. Equivalently we can use this relation to eliminate the integral and obtain quite simple expressions for the logarithms of the determinants of the Laplacians on zero and one forms respectively. We conclude this section by quoting these results Noting first of all that the expressions for $\tp(2)$ and $\tm(2)$ in their simplest form now become $$\tp(2)= {3\over \pi^2}\zeta(3)-2\ln 2\qquad {\rm and } \qquad\tm(2)={3\over \pi^2}\zeta(3)+2\ln 2\no$$ and from the definitions of these objects we have at once that $$\eqalign{ \ln \Det d^{}d_{0} &= -{3\over 2\pi^2}\zeta(3) +\ln2\cr \ln \Det d^{}d_{1} &= -{3\over \pi^2}\zeta(3)-2 \ln2\cr }\no$$ It is interesting to note the role that the Riemann zeta function $\zeta(3)$ plays in these expressions. Since these are expressions for volume elements on the discrete moduli spaces associated with the Laplacians, we expect that there are deeper things to be learned from a further study of such expressions. In the next section we tackle the continuation for arbitrary $p$. The analytic continuation for a general value of $p$ naturally divides into two cases: $p$ odd and $p$ even; in fact we shall see below that the case for $p$ even further divides into two subcases which correspond to $p=0,2\,{\rm mod\,}4$. Due to the size of the expressions it is now much more convenient to continue $\tau_+(p,s)$ and $\tau_-(p,s)$ separately and then combine them into expressions for the torsion. We deal first with $\tau_+(p,s)$ and begin with $p$ odd. Let us recall that $\tau_+(p,s)$ is given by $$\tau_+(p,s)=\sum_{n=1}^\infty {2(n+1)\left\langle\chi^{n/2} \right\rangle_p\over \left\{n(n+2)\right\}^s} \no$$ Reference to the general character formula shows that we must resolve $n$ into its conjugacy classes mod $p$ by writing $$n=pk-j,\;\,j=0,\ldots,(p-1)\no$$ and that we must distinguish the two parities of $j$. To implement these requirements we set $$p=2r+1,\qquad\hbox{and parametrise } \quad\cases{j& odd by $j=2l+1$, $l=0,1,\ldots,(r-1)$\cr j& even by $j=2l$, $l=0,1,\ldots,r$\cr}\no$$ This gives $$\eqalign{\tau_+(p,s)&=2\sum_{j=0}^{p-1}\sum_{k=1}^\infty {(pk-j+1)\left\langle\chi^{(pk-j)/2}\right\rangle_p \over\left\{(pk-j)(pk-j+2)\right\}^s}\cr &=2\sum_{k=1}^\infty \left[\sum_{l=0}^{r-1}{(pk-2l) \left\langle\chi^{(pk-2l-1)/2}\right\rangle_p \over\left\{(pk-2l-1)(pk-2l+1)\right\}^s} +\sum_{l=0}^r{(pk-2l+1) \left\langle\chi^{(pk-2l)/2}\right\rangle_p \over\left\{(pk-2l)(pk-2l+2)\right\}^s}\right]\cr}\no$$ Then when we use for $p$ odd we get $$\eqalign{\tau_+(p,s) =\sum_{k=1}^\infty\left[{2pk^2\over\left\{(pk-1)(pk+1)\right\}^s}\right. & +\sum_{l=1}^{r-1}{2(k-1)(pk-2l)\over\left\{(pk-2l-1)(pk- 2l+1)\right\}^s}\cr &+\sum_{l=0}^r\left.{2k(pk-2l+1)\over\left\{(pk-2l)(pk- 2l+2)\right\}^s}\right]\cr}\no$$ To aid in marshalling the combinatorics of $\tau_+(p,s)$ we define $H_{p}(k,s,x)$ by $$H_{p}(k,s,x)={pk(pk+x) \over\left\{(pk+x-1)(pk+x+1)\right\}^s}\no$$ The point being that each of the three summands in is of the form $H_{p}(k,s,\lambda)$ for appropriate $\lambda$. To see this we introduce precisely $p$ constants of the type $\lambda$ defined by $$\cases{\lambda_0=0&\cr \lambda_l=-2l+1,\;\,l=0,\ldots,r&\cr \bar\lambda_l=-2l+p,\;\,l=1,\ldots,(r-1)&\cr}\no$$ With this notation it can be checked that $\tau_+(p,s)$ is given by $$\tau_+(p,s)={2\over p}\sum_{k=1}^\infty\left[H_{p}(k,s,\lambda_0) +\sum_{l=1}^{(r-1)}H_{p}(k,s,\bar\lambda_l) +\sum_{l=0}^{r}H_{p}(k,s,\lambda_l)\right]\no$$ Also if we denote the entire set of $\lambda$’s by $\{\lambda\}$ i.e. $$\{\lambda\}\equiv\{\lambda_0,\lambda_l,\bar\lambda_l\}=\{-(p-2), \ldots,-5,-3,-1,0,1,3,5\ldots(p-2)\}\no$$ then we have the even more concise expression $$\tau_+(p,s)={2\over p}\sum_{\{\lambda\}}\sum_{k=1}^\infty H_{p}(k,\lambda,s)\no$$ The functions $H_{p}(\lambda,s)$ obtained by summing over $k$ then form a set of $p$ functions whose derivative at $s=0$ can be viewed as living on the appropriate space of sections for the Laplacian acting on $0$ forms on the lens space $L(p)$; taking the trace over these functions viewed as forming a matrix then gives the analytic continuation of the determinant of the Laplacian. Next observing $$H_{p}(k,x,s) ={{(pk+x)}^2\over\{{(pk+x)}^2-1\}^s} -x{(pk+x)\over\{{(pk+x)}^2-1\}^s} \no$$ we see that there are therefore two additional functions of interest here i.e. $$A_{p}(k,x,s)={{(pk+x)}^2\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ and $$B_{p}(k,x,s)={(pk+x)\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ and we have $$H_{p}(k,x,s)=A_{p}(k,x,s)-xB_{p}(k,x,s)\no$$ and we can equally write $$\tau_{+}(p,s) ={2\over p}\sum_{\{\lambda\}}\left[A_{p}(\lambda,s) -\lambda B_{p}(\lambda,s)\right]\no$$ Let us further note that in the set $\{\lambda\}$ the non zero elements come in pairs of the form $\{\lambda,-\lambda\}$. Thus we can further write $$\tau_{+}(p,s)={2\over p}A_{p}(0,s)+{2\over p}\sum_{l=1}^{r-1} \left[A_{p}^{+}(2l+1,s)-(2l+1)B_{p}^{-}(2l+1,s)\right]$$ where the $\mp$ superscripts refer to the symmetric or anti-symmetric combination with respect to the first argument: i.e. $A_{p}^{+}(x,s)=A_{p}(x,s)+A_{p}(-x,s)$ and $B_{p}^{-}(x,s)=B_{p}(x,s)-B_{p}(-x,s)$. We relegate the details of the computation of these functions and their analytic continuation to appendices A and B—the calculations are generalisations of those performed for $p=2$. Quoting here from appendices A and B we have that the relevant functions and their derivatives at $s=0$ are given by $$\eqalign{A_{p}(x,0)&=p^2\zeta(-2,1+{x\over p}) = -{x(x+p)(2x+p)\over 6p}\cr B_{p}(x,0)&=p\zeta(-1,1+{x\over p})+{1\over 2p}=-{p\over 12}-{x\over p}+{x^2+1\over 2 p}\cr H_{p}(x,0)&=-{px\over 12}-{x\over 2p}+{x^3\over 6p}\cr }\no$$ Which immediately implies that $H_{p}^{+}(x,0)=0$ and we have our first result that for $p$ odd $$\tau_{+}(p,0)=0$$ The significance of this is that the analytic continuation of the scaling dimension of the determinant is zero. Next we note again from appendices A and B that $$\eqalign{A_{p}^{+}(x) =-{p^2\over\pi^2}\zeta(3) &-{p^2\over\pi^2}\int_{0}^{(x+1)\pi\over p} dz\, z(z-{2\pi x\over p})\cot(z) -{p^2\over\pi^2}\int_{0}^{(x-1)\pi\over p} dz\, z(z-{2\pi x\over p})\cot(z)\cr &-x^2\ln\left[{2\sin\left({(x+1)\pi/ p}\right)\over x+1}\right] -x^2\ln\left[{2\sin\left({(x-1)\pi/ p}\right)\over x-1}\right]\cr }\no$$ and $$\eqalign{B_{p}^{-}(x) &= {p\over\pi}\int_{0}^{(x+1)\pi\over p}dz\,z\cot(z) +{p\over\pi}\int_{0}^{(x-1)\pi\over p}dz\, z\cot(z)\cr &\qquad-x\ln\left[{2\sin\left({(x+1)\pi/ p}\right)\over x+1}\right] -x\ln\left[{2\sin\left({(x-1)\pi/ p}\right)\over x-1}\right]\cr }\no$$ Now combining our expressions , and summing over $\{\lambda\}$ we obtain for $\tp(p)$ the expression $$\eqalign{\tp(p) &=-{2\over p}\left[{p^3\over2\pi^2}\zeta(3) +{p^2\over\pi^2}\int_{0}^{\pi\over p}dz\, z^2\cot(z) +{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p} dz\, z(z-{(p-2)\pi\over p})\cot(z)\right.\cr &\qquad\left.+2\sum_{l=1}^{(p-3)\over2} {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz\, z(z-{2l\pi\over p})\cot(z) \right]\cr }\no$$ for $p$ odd. But the expression $\tp(p)$ above is $2\ln\Det\dstar d_0 $: i.e. twice the logarithm of the Laplacian on 0-forms for the lens space $L(p)$. More precisely our analytic continuation has shown us that $$\eqalign{\ln\Det\dstar d_0 &={1\over p}\left[{p^3\over2\pi^2}\zeta(3) +{p^2\over\pi^2}\int_{0}^{\pi\over p}dz z^2\cot(z) +{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p}dz z(z-{(p-2)\pi\over p})\cot(z)\right.\cr &\qquad\left.+2\sum_{l=1}^{(p-3)\over2} {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz z(z-{2l\pi\over p})\cot(z) \right]\cr }\no$$ Let us now turn to $\tau_{-}(p,s)$ for $p$ odd. $$\tau_{-}(p,s)=\sum_{n=1}^{\infty} {n\left<\chi^{(n+1)\over2}\right>_{p} +(n+2)\left<\chi^{(n-1)\over2}\right>_{p}\over {(n+1)}^{2s}}\no$$ which on decomposing $n$ over the conjugacy classes mod $p$ $n=pk-j,\quad j=0,\dots,(p-1)$ as for $\tau_{+}(p,s)$, distinguishing the two parities as in yields $$\eqalign{\tau_{-}(p,s) &=\sum_{k=1}^{\infty}{kp{\left\langle\chi^{kp+1\over2}\right\rangle}_{p} +(kp+2){\left\langle\chi^{kp-1\over2}\right\rangle}_{p}\over{(kp+1)}^{2s}} +\sum_{k=1}^{\infty}{(kp-2){\left\langle\chi^{kp-1\over2}\right\rangle}_{p} +kp{\left\langle\chi^{kp-3\over2}\right\rangle}_{p}\over{(kp-1)}^{2s}}\cr &\qquad+\sum_{k=1}^{\infty}\sum_{l=2}^{(p-3)\over2} {(kp-2l){\left\langle\chi^{kp-2l+1\over2}\right\rangle}_{p} %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR +(kp-2l+2){\left\langle\chi^{kp-2l-1\over2}\right\rangle}_{p}\over{(kp-2l+1)}^{2s}}\cr %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR &\qquad+\sum_{k=1}^{\infty}{(kp-p+1){\left\langle\chi^{kp-p\over2}\right\rangle}_{p} %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR +(kp-p+3){\left\langle\chi^{kp-p-2\over2}\right\rangle}_{p}\over{(kp-p+2)}^{2s}}\cr &\qquad+\sum_{k=1}^{\infty}\sum_{l=0}^{r-1} {(kp-2l-1){\left\langle\chi^{kp-2l\over2}\right\rangle}_{p} %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR +(kp-2l+1){\left\langle\chi^{kp-2l-2\over2}\right\rangle}_{p}\over{(kp-2l)}^{2s}}\cr }\no$$ Using our degeneracy formula we have with a little re-arrangement $$\eqalign{\tau_{-}(p,s) &=\sum_{k=1}^{\infty}{2(kp+1)k+kp\over{(kp+1)}^{2s}} +\sum_{k=1}^{\infty}{2(kp-1)k-kp\over{(kp-1)}^{2s}}\cr &\qquad+\sum_{l=-{(p-3)\over2}}^{(p-3)\over2} \sum_{k=1}^{\infty}{2(kp+2l)k\over{(kp+2l)}^{2s}}\cr }\no$$ We now observe that this is of the form $$\eqalign{\tau_{-}(p,s) &={2\over p}\sum_{\{\nu\}}\sum_{k=1}^{\infty} {kp(kp+\nu)\over{(kp+\nu)}^{2s}}\cr &+\sum_{k=1}^{\infty}{kp\over{(kp+1)}^{2s}} -\sum_{k=1}^{\infty}{kp\over{(kp-1)}^{2s}}\cr }\no$$ where we have denoted by $\{\nu\}$ the set $$\{\nu\}\equiv\{\nu_0,\nu_l,\bar\nu_l\}=\{-(p-3), \ldots,-6,-4,-2,-1,0,1,2,4,6\ldots(p-3)\}\no$$ The sums occuring in are naturally expressible in terms of Hurwitz zeta functions and this provides the natural analytic continuation of this expression. We also note that $$\sum_{\{\nu\}}\sum_{k=1}^{\infty}{1\over {(kp+\nu)}^{(2s-2)}} =\zeta(2s-2)-1-\sum_{l=1}^{(p-3)\over2}{(2l)}^{(2-2s)}\no$$ On utilising these relations we find that $$\eqalign{\tau_{-}(p,s) &={2\over p} \left[\zeta(2s-2)-1-\sum_{l=1}^{(p-3)\over2}{(2l)}^{2-2s}\right]\cr &\qquad (p-2){p}^{-2s} \left[\zeta(2s-1,1+{1\over p})-\zeta(2s-1,1-{1\over p})\right]\cr &\qquad-2{p}^{-2s}\sum_{l=-{(p-3)\over2}}^{(p-3)\over2} (2l)\zeta(2s-1,1+{2l\over p})\cr &\qquad-{p}^{-2s} \left[\zeta(2s,1+{1\over p})+\zeta(2s,1-{1\over p})\right] }\no$$ If we analytically continue the RHS of this expression to $s=0$, and observe that $\zeta(-1,1+x)-\zeta(-1,1-x)=-x$, then we obtain $$\tau_{-}(p,0)=-{2\over p} \left[1+\sum_{l=1}^{(p-3)\over2}{(2l)}^2\right] +(p-2)(-{1\over p}) -2\sum_{l=1}^{(p-3)\over2}(2l)(-{2l\over p})\no$$ which immediately gives $$\tau_{-}(p,0)=0\no$$ Thus again the scaling dimension of the associated determinant is zero. Passing to $\tm(p)$ by taking the derivative at $s=0$ gives $$\eqalign{\tm(p) &={4\over p}\left[\zeta '(-2)-\ln{p} +\sum_{l=1}^{(p-3)\over2}{(2l)}^2\ln[{2l\over p}]\right]\cr &\qquad+2(p-2)\left[\zeta '(-1,1+{1\over p}) -\zeta '(-1,1-{1\over p})\right]\cr &\qquad-4\sum_{l=1}^{(p-3)\over 2}(2l) \left[\zeta '(-1,1+{2l\over p})-\zeta '(-1,1-{2l\over p})\right]\cr &\qquad-2 \left[\zeta '(0,1+{1\over p})+\zeta '(0,1-{1\over p})\right]\cr }\no$$ A useful identity for Hurwitz zeta functions derived in appendix C is $$\zeta '(-1,1+x) -\zeta '(-1,1-x)= -x\ln\left[{2\sin(\pi x)\over x}\right]+{1\over \pi}\int_{0}^{\pi x}dz z\cot(z)$$ Using this and the expression $\zeta '(-2)=-{\zeta(3)/ 4{\pi}^2}$ we conclude that $$\eqalign{\tm(p) &= -{1\over p{\pi}^2}\zeta(3) +{4\over p}\ln\left[2\sin({\pi\over p})\right] +{4\over p}\sum_{l=1}^{(p-3)\over 2} {(2l)}^2\ln\left[2\sin({2\pi l\over p})\right]\cr &\qquad+{2(p-2)\over \pi}\int_{0}^{\pi\over p}dz z\cot(z) -{4\over\pi}\sum_{l=1}^{(p-3)\over 2} (2l)\int_{0}^{2\pi l\over p}dz z\cot(z)\cr }\no$$ But $-\tm(p)$ is the analytic continuation which gives $\ln \Det d^d_{1}$. Hence we find that $$\eqalign{\ln \Det d^d_{1} &= {1\over p{\pi}^2}\zeta(3) +{4\over p}\ln\left[2\sin({\pi\over p})\right] +{4\over p}\sum_{l=1}^{(p-3)\over 2} {(2l)}^2\ln\left[2\sin({2\pi l\over p})\right]\cr &\qquad+{2(p-2)\over \pi}\int_{0}^{\pi\over p}dz z\cot(z) -{4\over\pi}\sum_{l=1}^{(p-3)\over 2} (2l)\int_{0}^{2\pi l\over p}dz z\cot(z)\cr }\no$$ We are therefore now in a position to put these together and obtain an expression for the torsion. The first observation is that since $\tau_{+}(p,0)$ and $\tau_{-}(p,0)$ are both zero we have $$\tau(p,0)=0\no$$ This vanishing of $\tau(p,0)$ is related to the metric independence of the torsion something which has been established quite generally by Ray and Singer \[3\]. The torsion $T(p)$ itself is given by the difference of and . Combining these two expressions we find, upon a little simplification, that $T(p)$ is determined by the equation $$\eqalign{&\ln T(p) =-{2\over p}\left[{(p^3-1)\over2\pi^2}\zeta(3) +2\ln\left[2\sin({\pi\over p})\right] +2\sum_{l=1}^{(p-3)\over 2} 4l^2\ln\left[2\sin({2l\pi\over p})\right] \right.\cr &\qquad+{p^2\over\pi^2}\int_{0}^{\pi\over p}dz\, z(z+{(p-2)\pi\over p})\cot(z) +{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p}dz\, z(z-{(p-2)\pi\over p})\cot(z)\cr &\qquad\left.+2\sum_{l=1}^{(p-3)\over 2} {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz\, z(z-{4l\pi\over p})\cot(z) \right]\cr }\no$$ Now the expression for the torsion from Ray’s calculation gave the alternative expression i.e. $$\ln T(p) = -4\ln\left[2 \sin({\pi\over p})\right]\no$$ We have verified that these two expressions and agree numerically, yet it is not transparent by inspection that this should be so; also using C.41 of appendix C we can reduce to Ray’s expression. One may conclude that, by following two alternate derivations, we have arrived at what is a sequence of non-trivial identities. As we saw in the case of $p=2$ utilising these identities one can obtain non-trivial formulae for $\zeta(3)$ and also can be used to further simplify the expressions for the individual determinants. The resulting expressions for $\zeta(3)$ are $$\eqalign{\zeta(3)&={2\pi^2\over (p^3-1)}\left[ 2(p-1)\ln\left[2\sin({\pi\over p})\right] -2\sum_{l=1}^{(p-3)\over 2} 4l^2\ln\left[2\sin({2l\pi\over p})\right] \right.\cr &\qquad-{p^2\over\pi^2}\int_{0}^{\pi\over p}dz\, z(z+{(p-2)\pi\over p})\cot(z) -{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p}dz\, z(z-{(p-2)\pi\over p})\cot(z)\cr &\qquad\left.-2\sum_{l=1}^{(p-3)\over 2} {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz\, z(z-{4l\pi\over p})\cot(z) \right]\cr }\no$$ For the sake of illustration let us quote the implications of these formulae for the simplest odd case: $p=3$. On utilising all of the information at our disposal we find that $$\eqalign{\tp(3)&=-{\zeta(3)\over 3\pi^2}-{4\over3}\ln3 +{2\over\pi}\int_{0}^{\pi\over3}dz\,z \cot(z)\cr \tm(3)&=-{\zeta(3)\over 3\pi^2}+{2\over3}\ln3 +{2\over\pi}\int_{0}^{\pi\over3}dz\,z \cot(z)\qquad {\rm and}\cr \ln T(3)&=-2\ln3\cr }$$ The relation between the expressions we have obtained which we expressed in terms of a formula for $\zeta(3)$ (the analog of for $p=2$) becomes $$\zeta(3)={2\pi^2\over13}\ln 3 -{9\over 13}\int_{0}^{\pi\over3}dz\,z(z+{\pi\over3})\cot(z) -{9\over 13}\int_{0}^{2\pi\over3}dz\,z(z-{\pi\over3})\cot(z)$$ We will now turn to the case of even $p$. When $p$ is even we follow a slightly different route to that used in the previous section but we arrive at expressions of a similar general form for the respective determinants and their corresponding torsion. Let us first obtain series expressions for $\tau_{+}(p,s)$ for $p$ even, and observe that the same functions as those encountered for $p$ odd enter these also. Recalling our expression for $\tau_{+}(p,s)$ $$\tau_{+}(p,s)= \sum_{n=1}^{\infty} {2(n+1)\left<\chi^{n\over2}\right>_{p}\over{\left\{{(n+1)}^2-1\right\}}^s} \no$$ We again resolve $n$ into its conjugacy classes mod $p$ by writing $$n=pk-j,\;\,j=0,\ldots,(p-1)\no$$ and distinguish the two parities of $j$ by setting $$p=2r,\qquad\hbox{and parametrise } \quad\cases{j& odd by $j=2l+1$, $l=0,1,\ldots,(r-1)$\cr j& even by $j=2l$, $l=0,1,\ldots,(r-1)$\cr}\no$$ Reference to the general character formula shows that only $j$ odd contributes and we obtain $$\tau_{+}(p,s)=\sum_{k=1}^{\infty}{4p k\over{\left\{{(pk)}^2-1\right\}}^s} +\sum_{l=0}^{(p-2)\over2} \sum_{k=1}^{\infty} {2(pk-2l)(2k-1)\over{\left\{{(pk-2l)}^2-1\right\}}^s} \no$$ which gives $$\tau_{+}(p,s)={4\over p}\sum_{k=1}^{\infty}\left[\sum_{l=0}^{(p-2)\over2} {{(pk-2l)}^2\over{\left\{{(pk-2l)}^2-1\right\}}^s} +\sum_{l=1}^{(p-2)\over2} (2l-{p\over2}){(pk-2l)\over{\left\{{(pk-2l)}^2-1\right\}}^s}\right] \no$$ We recognise the expressions arising as the functions from the preceding analysis in the case of $p$ odd and which are analysed in appendices A and B. We can therefore write as $$\tau_{+}(p,s)={4\over p}\sum_{l=0}^{r-1} \left[A_{p}(-2l,s)+(2l-r)B_{p}(-2l,s)\right] \no$$ Some further rearrangement will allow us to write these again in terms of the symmetric and anti-symmetric parts of $A_{p}(x,s)$ and $B_{p}(x,s)$ respectively. Note first of all, however that the term involving $A_{p}$ is a sum over all even conjugacy classes i.e. $$\eqalign{\sum_{l=1}^{r-1}A(-2l,s) &=\sum_{l=1}^{r-1}\sum_{k=1}^{\infty} {{(2rk-2l)}^2\over{\left\{{(2rk-2l)}^2-1\right\}}^{s}}\cr &=\sum_{m=1}^{\infty}{{(2m)}^2\over{\left\{{(2m)}^2-1\right\}}^{s}}\cr }\no$$ and is therefore ${\tau_{+}(2,s)/2}$; our expression for $\tau_{+}(p,s)$ can hence be written as $$\tau_{+}(p,s)={2\over p}\left[\tau_{+}(2,s) +2\sum_{l=1}^{r-1}(2l-r)B_{p}(-2l,s)\right] \no$$ Now when $l$ ranges from $1$ to $r-1$, $(2l-r)$ ranges over the set $$-(r-2),(r-4),\dots,(r-4),(r-2)\no$$ This allows us to divide the range up into a sum from $1$ up to the integer part of ${(r-1)/ 2}$ which we denote by $\left[{(r-1)/2}\right]$ Thus $$\tau_{+}(p,s)={2\over p}\left[\tau_{+}(2,s) +2\sum_{l=1}^{\left[{(r-1)\over2}\right]}(r-2l) \left[B_{p}(-p+2l,s)-B_{p}(-2l,s)\right]\right] \no$$ Observing that $$B_{p}(-p+x,s)={x\over{\left\{{x}^2-1\right\}}^s}+B_{p}(x,s)$$ This gives $$\tau_{+}(p,s)={2\over p}\left[\tau_{+}(2,s) +2\sum_{l=1}^{\left[{(p-2)\over4}\right]}({p\over2}-2l) \left[B_{p}^{-}(2l,s) +{2l\over{\left\{{x}^2-1\right\}}^s}\right]\right] \no$$ Noting that $\tau_{+}(2,0)=0$ and that $B_{p}^{-}(x)=-x$ we see again immediately that $$\tau_{+}(p,0)=0\no$$ Again as we expect the scaling dimension of the associated determinant is zero. Proceeding now to the expression for the determinant itself, we find the resulting expression from for the derivative at $s=0$ is $$\tau_{p}'(p,0)={2\over p}\left[\tp(2) +\sum_{l=1}^{\left[{(p-2)\over4}\right]}(p-4l) \left[B_{p}^{-}(2l) -2l\ln[{2l}^2-1]\right]\right] \no$$ Substituting for $B_{p}^{-}(2l)$ from gives $$\eqalign{\tp(p)&={2\over p}\left[\tp(2) +{p\over\pi}\sum_{l=1}^{\left[{(p-2)\over4}\right]}(p-4l) \left[\int_{0}^{(2l+1)\pi\over p}z\cot(z) +\int_{0}^{(2l-1)\pi\over p}dz z\cot(z)\right.\right.\cr &\qquad\left.\left. -{2l\pi\over p}\ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right] -{2l\pi\over p}\ln\left[2\sin\left({(2l-1)\pi\over p}\right)\right] \right]\right]\cr }\no$$ where from $\tp(2)={3\over\pi^2}\zeta(3)-2\ln2$ . As we see the case of even $p$ divides naturally into two classes $p=2\, {\rm mod\,}4$ and $p=0\, {\rm mod\,}4$. Making this division we can further simplify things to obtain $$\eqalign{\tp(p) &={2\over r}\left[{3\over 2\pi^2}\zeta(3) -2\sum_{l=1}^{(r-3)\over 2}\left[(2l+1)(r-2l-1)-1\right]\ln\left[2\sin\left({(2l+1)\pi\over 2r}\right)\right]\right.\cr &\left. +{2r(r-2)\over\pi}\int_{0}^{\pi\over 2r}dz\,z\cot(z) +{4r\over\pi}\sum_{l=1}^{(r-2)\over2}(r-2l-1) \int_{0}^{(2l+1)\pi\over 2r}dz\,z\cot(z)\right]\; p=2\mod4\cr }$$ and $$\eqalign{\tp(p) &={2\over r}\left[{3\over 2\pi^2}\zeta(3)-\ln2 -2\sum_{l=1}^{(r-2)\over 2}\left[(2l+1)(r-2l-1)-1\right]\ln\left[2\sin\left({(2l+1)\pi\over 2r}\right)\right]\right.\cr &\left. +{2r(r-2)\over\pi}\int_{0}^{\pi\over 2r}dz\,z\cot(z) +{4r\over\pi}\sum_{l=1}^{(r-2)\over2}(r-2l) \int_{0}^{(2l+1)\pi\over 2r}dz\,z\cot(z)\right]\;p=0\mod4\cr }$$ We note that these expressions agree with the result for $p=2$ and, for $p=4$, we note in passing that inspection of the series shows that $\tau_{\pm}'(4,0)={\tau_{\pm}'(2,0)/2}$; this turns out to be also a property of (the logarithm of) the torsion itself, i.e. $T(p)$ satisifies $\ln T(4)=(\ln T(2))/2$. We therefore have from an expression for the appropriate logarithmic determinant on the lens space $L(p)$ namely $$\eqalign{\ln\Det \dstar d_0 &= -{3\over p\pi^2}\zeta(3)+{2\over p}\ln2\cr &\quad-{2\over p} \sum_{l=1}^{\left[{(p-2)\over4}\right]}({p\over2}-2l) \left[{p\over\pi}\int_{0}^{(2l+1)\pi\over p}z\cot(z) +{p\over\pi}\int_{0}^{(2l-1)\over p}dz z\cot(z)\right.\cr &\quad\left.-2l\ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right] -2l\ln\left[2\sin\left({(2l-1)\pi\over p}\right)\right] \right]\cr }\quad\; p=2r\no$$ Let us now turn to the evaluation of $\tau_{-}(p,s)$. $$\tau_{-}(p,s)=\sum_{n=1}^{\infty} {n\left\langle\chi^{n+1\over2}\right\rangle_p +(n+2)\left\langle\chi^{n-1\over2}\right\rangle_p \over {(n+1)}^{2s}} \no$$ Decomposing the sum over $n$ into the different conjugacy classes and using our general character formula we have $$\eqalign{\tau_{-}(p,s)=&\sum_{k=1}^{\infty}{pk(pk+1) +(pk+2)2k\over{(pk+1)}^{2s}}\cr &\qquad+\sum_{k=1}^{\infty}{(pk-2)2k+pk(2k-1)\over{(pk-1)}^{2s}}\cr &\qquad+\sum_{l=2}^{r-1}{2(pk-2l+1)(2k-1)\over{(pk-2l+1)}^{2s}}\cr }\no$$ After some rearrangement we arrive at $$\eqalign{\tau_{-}(p,s)=&\sum_{k=1}^{\infty}\left\{ {4\over p}\sum_{l=0}^{r-1}{1\over{(pk-2l+1)}^{2s-2}}\right.\cr &\qquad+(1-{4\over p}) \left[{1\over{(pk+1)}^{2s-1}}-{1\over{(pk-1)}^{2s-1}}\right]\cr &\qquad-\left[{1\over{(pk+1)}^{2s}}+{1\over{(pk-1)}^{2s}}\right]\cr &\qquad+\left.{4\over p} \sum_{l=2}^{r-1}(2l-1-r){1\over{(pk-2l+1)}^{2s-1}}\right\}\cr }\no$$ Since the first term involves a sum over all odd conjugacy classes we have $$\eqalign{\sum_{k=1}^{\infty}\sum_{l=0}^{r-1}{1\over{(pk-2l+1)}^{2s-2}} &=\sum_{m=1}^{\infty}{1\over{(2m+1)}^{2s-2}}\cr &=(1-{1\over {2}^{2s-2}})\zeta(2s-2)-1\cr }\no$$ Thus $$\eqalign{\tau_{-}(p,s) =&{4\over p}\left((1-{1\over{2}^{2s-2}})\zeta(2s-2)-1\right)\cr &\qquad+(p-4){1\over {p}^{2s}} [\zeta(2s-1,1+{1\over p})-\zeta(2s-1,1-{1\over p})]\cr &\qquad-{1\over {p}^{2s}} [\zeta(2s,1+{1\over p})+\zeta(2s,1-{1\over p})]\cr &\qquad-{2\over{p}^{2s}} \sum_{l=2}^{r-1}\left(p-(2l-1)\right)\zeta(2s-1,1-{2l-1\over p})\cr }\no$$ Noting that $p-(2l-1)$ ranges from $-(p-6)$ to $(p-6)$ in steps of $2$ when $l$ ranges from $2$ to $r-1$ the final sum in is therefore of the form $$\eqalign{\sum_{l=2}^{r-1} \left(p-(2l-1)\right)\zeta(2s-1,1-{2l-1\over p}) &=\sum_{l=2}^{[{r-1\over 2}]}\left(p-2(2l-1)\right)\left[\zeta(2s-1,1-{2l-1\over p})\right.\cr &\left.-\zeta(2s-1,{2l-1\over p})\right]\cr }\no$$ Substituting back we obtain $$\eqalign{\tau_{-}(p,s) &={4\over p}\left((1-{1\over{2}^{2s-2}})\zeta(2s-2)-1\right)\cr &\qquad+(p-4){1\over {p}^{2s}} [\zeta(2s-1,1+{1\over p})-\zeta(2s-1,1-{1\over p})]\cr &\qquad-{1\over {p}^{2s}} [\zeta(2s,1+{1\over p})+\zeta(2s,1-{1\over p})]\cr &\qquad+{2\over{p}^{2s}} \sum_{l=2}^{[{r-1\over 2}]} \left(p-(2l-1)\right) \left[\zeta(2s-1,{2l-1\over p})-\zeta(2s-1,1-{2l-1\over p})\right]\cr }\no$$ Our first observation is that using $\zeta(-2)=0$,$\zeta(-1,1+a)-\zeta(-1,1-a)=-a$,and $\zeta(0,1+a)+\zeta(0,1-a)=-1$.$$\tau_{-}(p,0)=0\no$$ Now differentiation of with respect to $s$ and evaluating the expression at $s=0$, and using some of our relations from Appendix C gives $$\eqalign{\tm(p) &=-{24\over p}\zeta '(-2) +2(p-4)[\zeta '(-1,{1\over p})-\zeta '(-1,1-{1\over p})] -2[\zeta '(0,{1\over p})+\zeta '(0,1-{1\over p})]\cr &\qquad+4\sum_{l=1}^{[{r-3\over 2}]} \left(p-(2l-1)\right) \left[\zeta '(-1,{2l-1\over p})-\zeta '(-1,1-{2l-1\over p})\right]\cr }\no$$ Further use or our the relations derived in Appendix C allows us to express the result in terms of integrals over trigonometric functions as in the preceding sections to yield $$\eqalign{\tm(p) &=4\ln(2\sin{\pi\over p})+{6\over p{\pi}^2}\zeta(3) +2(p-4) [{1\over\pi}\int_{0}^{\pi\over p}dz z \cot(z)-{1\over p}\ln(2\sin{\pi\over p})]\cr &\quad+4\sum_{l=1}^{[{r-3\over 2}]} \left(p-(2l+1)\right) \left[{1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz z \cot(z)-{(2l+1)\over p}\ln(2\sin{\pi(2l+1)\over p})\right]\cr }\no$$ This again decomposes into the two cases $p=0,2\mod4$ and these yield the expressions $$\eqalign{\tm(p) &=4\ln(2\sin{\pi\over p})+{6\over p{\pi}^2}\zeta(3) +2(p-4) [{1\over\pi}\int_{0}^{\pi\over p}dz z \cot(z)-{1\over p}\ln(2\sin{\pi\over p})]\cr &+4\sum_{l=1}^{t-1} \left(p-(2l+1)\right) \left[{1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz z \cot(z)-{(2l+1)\over p}\ln(2\sin{\pi(2l+1)\over p})\right]\cr & \hskip0.8\hsize p=2\mod4\cr}\no$$ and $$\eqalign{\tm(p) &=4\ln(2\sin{\pi\over p})+{6\over p{\pi}^2}\zeta(3) +2(p-4) [{1\over\pi}\int_{0}^{\pi\over p}dz z \cot(z)-{1\over p}\ln(2\sin{\pi\over p})]\cr &+4\sum_{l=1}^{t-1} \left(p-(2l+1)\right) \left[{1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz z \cot(z)-{(2l+1)\over p}\ln(2\sin{\pi(2l+1)\over p})\right]\cr & \hskip0.8\hsize p=0\mod4\cr}\no$$ We can now combine our results for $\tau_{+}(p,0)$ and $\tau_{-}(p,0)$ to obtain expressions for the torsion in the present case where $p$ is even. Note again that since $\tau_{\pm}(p,0)=0$ we have $$\tau(p,0)=0\no$$ for the case of $p$ even, and again this ensures that the torsion is metric independent. Combining the expressions and for $\tp(p)$ and $\tm(p)$ we obtain two expressions for $\ln T(p)$: one for each conjugacy class; these are $$\eqalign{\ln T(p) &=-4\ln(2\sin{\pi\over p})+[{2(p-4)\over p}\ln(2\sin{\pi\over p})]\cr &\qquad+{4\over p} \sum_{l=1}^{(p-6)/4}\left[{(2l+1)}^2-2\right] \ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right]\cr &\qquad -{8\over\pi}\sum_{l=1}^{(p-6)/4}(2l+1) \int_{0}^{(2l+1)\pi\over p}dz\,z\cot(z)\cr }\qquad p=2\mod4\no$$ and $$\eqalign{\ln T(p) &=-4\ln(2\sin{\pi\over p}) -{4\over p}\ln2+2(p-4)[{1\over p}\ln(2\sin{\pi\over p})]\cr &\qquad+{4\over p}\sum_{l=1}^{(p-4)/4}\left[(2l+1)(2l+1)-2\right] \ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right]\cr &\qquad-{4\over\pi}\sum_{l=1}^{(p-4)/4}(2l-1) \int_{0}^{(2l+1)\pi\over p}dz\,z\cot(z)\cr }\qquad p=0\mod4\no$$ We note that for the above two formulae the torsion is already given by the first term on their RHS’s and so we obtain somewhat non-trivial integration formulae for the integrals therein. On using the relations derived at the end of appendix B we can also reduce the above expressions to Ray’s expression for the torsion. We have therefore now obtained a complete list of the determinants of Laplacians for 0 and 1 forms on the lens spaces $L(p)$ for all integer $p\ge 2$, as well as the torsion $T(p)$ for all $p\ge2$. In the preceeding sections we analysed by direct computation the the determinants of Laplacians on 0 and 1-forms on the lens spaces $L(p)$, defined via the derivatives of their associated zeta functions. In this concluding section we collect our results and present in graphical form the behaviour of the sequence of determinants we have analysed and their related torsion. For 0-forms we found that $$\eqalign{\ln\Det d^d_0=&{1\over 2p\pi^2}\zeta(3) +\ln\left[2\sin({\pi\over p})\right] +{(p-2)\over\pi}\int_{0}^{\pi/p}\ln\left[2\sin(z)\right]\cr &\qquad-{2\over \pi}\sum_{l=1}^{(p-3)/2}(2l)\int_{0}^{2l\pi/p} \ln\left[2\sin(z)\right]\cr }\qquad p=3,5,\ldots\no$$ in the case of $p$ odd and $$\eqalign{\ln\Det d^d_{0} &=-{3\over p{\pi}^2}\zeta(3) +{(p-4)\over\pi}\int_{0}^{\pi\over p}dz\ln\left[2\sin(z)\right]\cr &+2\sum_{l=1}^{t-1} \left(p-(2l+1)\right) {1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz\, \ln\left[2\sin(z)\right]\cr }\qquad p=2\mod4\no$$ and $$\eqalign{\ln\Det d^d_{0} &=-{6\over p{\pi}^2}\zeta(3)\cr &\qquad+ {(p-4)\over\pi}\int_{0}^{\pi\over p}dz\,\ln\left[2\sin(z)\right]\cr &\qquad+2\sum_{l=1}^{t-1} \left(p-(2l+1)\right){1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz\, \ln\left[2\sin(z)\right]\cr }\qquad p=0\mod4\no$$ These are plotted in [figure 1]{}. While for 1-forms we found $$\eqalign{\ln\Det d^d_1=&{1\over p\pi^2}\zeta(3) -2\ln\left[2\sin({\pi\over p})\right] +{2(p-2)\over\pi}\int_{0}^{\pi/p}\ln\left[2\sin(z)\right]\cr &\qquad-{4\over \pi}\sum_{l=1}^{(p-3)/2}(2l)\int_{0}^{2l\pi/p} \ln\left[2\sin(z)\right]\cr }\qquad p=3,5,\ldots\no$$ while when $p$ is even we found that $$\eqalign{\ln\Det d^d_{1} &=-4\ln(2\sin{\pi\over p})-{6\over p{\pi}^2}\zeta(3) +2(p-4) {1\over\pi}\int_{0}^{\pi\over p}dz\,\ln\left[2\sin(z)\right]\cr &+4\sum_{l=1}^{t-1} \left(p-(2l+1)\right) {1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz\, \ln\left[2\sin(z)\right]\cr }\qquad p=2\mod4\no$$ and $$\eqalign{\ln\Det d^d_{1} &=-4\ln(2\sin{\pi\over p})-{6\over p{\pi}^2}\zeta(3)\cr &\qquad+2(p-4) {1\over\pi}\int_{0}^{\pi\over p}dz\, \ln\left[2\sin(z)\right]\cr &\qquad+4\sum_{l=1}^{t-1} \left(p-(2l+1)\right) {1\over\pi}\int_{0}^{{\pi(2l+1)\over p}}dz\, \ln\left[2\sin(z)\right]\cr }\qquad p=0\mod4\no$$ and these results are displayed in [figure 2]{}. The [difference]{} $$\eqalign{\ln T(p)&=\ln\Det d^d_1-2\ln\Det d^d_0\cr &=-4\ln\left[2\sin({\pi\over p})\right]\cr }\no$$ which gives the torsion itself, is plotted in [figure 3]{}. A perusal of figure 3 shows that $\ln T(p)$ is negative for small $p$ and large and positive for large $p$. This raises the question as to whether $\ln T(p)$ crosses the $p$ axis at an integer value or not. If so this corresponds to a trivial value for the torsion. In fact this clearly does happen for the value $p=6$ i.e. we have $$\ln T(6)=0\no$$ We show the more detailed behaviour of the torsion for small $p$ in [figure 4]{}. Further interesting results are that that if we work with $L(p,q)$ rather than $L(p)\equiv L(p,1)$ then the torsion, now denoted by $T(p,q)$, is trivial for [only]{} two other three dimensional lens spaces, namely $L(10,3)$ and $L(12,5)$: we find that $$\eqalign{\ln T(p,q)&=-2 \ln\left[4\sin\left({\pi\over p}\right) \sin\left({\pi q^\over p}\right)\right]\cr \hbox{where }q^& \hbox{ satisfies }q q^=1\,\mod p\cr}\no$$ It is then possible to prove that, for $p>12$, $\ln T(p,q)$ is [strictly positive]{}; while for $p\le12$ a check of the finite number of cases yields triviality in just the three cases given above. We conjecture that this may be understandable using some form of supersymmetry. These formulae have yet to be elucidated further. The precise meaning of our formulae such as $$\eqalign{\zeta(3)&={2\pi^2\over7}\ln(2) -{8\over7}\int_0^{\pi/2}dz\,z^2\cot(z)\cr \zeta(3)&={2\pi^2\over13}\ln 3 -{9\over 13}\int_{0}^{\pi\over3}dz\,z(z+{\pi\over3})\cot(z) -{9\over 13}\int_{0}^{2\pi\over3}dz\,z(z-{\pi\over3})\cot(z) \cr}\no$$ and the more general $$\eqalign{\zeta(3)&={2\pi^2\over (p^3-1)}\left[ 2(p-1)\ln\left[2\sin({\pi\over p})\right] -2\sum_{l=1}^{(p-3)\over 2} 4l^2\ln\left[2\sin({2l\pi\over p})\right] \right.\cr &\qquad-{p^2\over\pi^2}\int_{0}^{\pi\over p}dz\, z(z+{(p-2)\pi\over p})\cot(z) -{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p}dz\, z(z-{(p-2)\pi\over p})\cot(z)\cr &\qquad\left.-2\sum_{l=1}^{(p-3)\over 2} {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz\, z(z-{4l\pi\over p})\cot(z) \right],\qquad p=3,5,\ldots\cr }\no$$ is, as yet, unclear. There may be some number theoretic matters underlying them as seems to be the case in other work on lens spaces, cf. \[15\]. A thought provoking fact is that $\zeta(3)$ occurs in a recent paper of Witten \[16\] where, after multiplication by a known constant, it gives the [volume]{} of the symplectic space of flat connections over a [non-orientable]{} Riemann surface. The corresponding calculation for [orientable]{} surfaces (where the volume element is a rational cohomology class) allows a cohomological rederivation of the irrationality of $\zeta(2),\, \zeta(4),\dots$. This paper also involves the torsion but in two dimensions rather than three. The proof that $\zeta(3)$ is irrational was only obtained in 1978 cf. \[17\] and the rationality of $\zeta(5),\,\zeta(7),\ldots$ is at present open. However there are now other proofs \[18\], one of which uses the characters of conformal quantum field theory. Finally, our technique, applied in five dimensions instead of three, would yield formulae for $\zeta(5)$ but their nature has not yet been explored. =1.5 FIGURES: The Determinants* 6truein -1.75truein [Figure 1:]{} $2\ln \dstar d_0$ versus $p$. 6truein -1.75truein [Figure 2:]{} $\ln \dstar d_1$ versus $p$. =.666 =1.5 FIGURES: The Torsion 6truein -1.75truein [Figure 3:]{} $\ln T(p)$ versus $p$. 6truein -1.75truein [Figure 4:]{} $\ln T(p)$ versus $p$ for small p. =.666 In this appendix we analyze the function $$A_{p}(x,s) =\sum_{k=1}^{\infty}{{(pk+x)}^2\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ We are interested in particular in the value of this function and its derivative with respect to $s$ at $s=0$. For this purpose we denote $$A_{p}(k,x,s)={{(pk+x)}^2\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ which has the expansion $$\eqalign{A_{p}(k,x,s) &={1\over {(pk)}^{2s-2}}\left[1+2(1-s){x\over pk} +\left(s+(s-1)(2s-1)x^2\right){1\over {(pk)}^2}\right.\cr &\qquad\left. -2sx\left(s+{(s-1)(2s-1)\over3}x^2\right){1\over{(pk)}^3}+\dots\right] \cr }\no$$ Summing over $k$ leads to $$\eqalign{A_{p}(x,s) &={1\over{p}^{2s-2}}\zeta(2s-2)+{2x\over{p}^{2s-1}}(1-s)\zeta(2s-1) +{1\over{p}^{2s}}\left(s+(1-s)(1-2s)x^2\right)\zeta(2s)\cr &\qquad-{1\over{p}^{2s+1}}\left(s+{(1-s)(1-2s)\over3}x^2\right)x2s\zeta(2s+1) +\hat A_{p}(x,s)\cr }\no$$ where $$\eqalign{\hat A_{p}(x,s) &=\sum_{k=1}^{\infty}\left\{ {{(pk+x)}^2\over{\left\{{(pk+x)}^2-1\right\}}^s} -{1\over {(pk)}^{2s-2}}\left[1+2(1-s){x\over pk}\right.\right.\cr &\left.\left.+\left(s+(s-1)(2s-1)x^2\right){1\over {(pk)}^2} -2sx\left(s+{(s-1)(2s-1)\over3}x^2\right){1\over{(pk)}^3} \right]\right\}\cr }\no$$ Now the function $\hat A_{p}(x,s)$ is such that the processes of summation and differentiation with respect to $s$ at $s=0$ commute. Also it is such that $\hat A_{p}(x,0)=0$, which yields $$A_{p}(x,0)=-{x(x+p)(2x+p)\over 6p}\no$$ Next evaluating the derivative at $s=0$ we have $$\eqalign{A_{p}'(x,0) &=-A_{p}(x,0)\ln{p}^2+2{p}^{2}\zeta '(-2)+\zeta(0) +2px\left[2\zeta '(-1)-\zeta(-1)\right]\cr &\qquad+\left[2\zeta '(0)-3\zeta(0)\right]x^2 -{x\over p}+{1\over{3p}}\left[3-2\gamma\right]x^3 +\hat A_{p}'(x,0)\cr }\no$$ which on using $\zeta(0)=-{1\over 2}$, $\zeta(-1)=-{1\over12}$, $2s\zeta(2s+1)=1+2\gamma s+\dots$ and $\zeta'(0)=-{1\over2}\ln{2\pi}$ we have $$\eqalign{A_{p}'(x,0)&=-A_{p}(x,0)\ln{p}^2+2p^2\zeta '(-2)-{1\over2} +\left[4p\zeta'(-1)+{p\over6}-{1\over p}\right]x\cr &\qquad+\left[{3\over2}-\ln{[2\pi]}\right]x^2 +{1\over 3p}\left[3-2\gamma\right]x^3+\hat A_{p}'(x,0)\cr }\no$$ Since the processes of differentiation with respect to $s$ and performing the sum over $k$ commute for $\hat A_{p}(x,s)$ we analyze this function by first taking the derivative and then performing the sum. $$\eqalign{\hat A_{p}'(k,x,0) &=-{(pk +x)}^2\left[\ln\left[1+{(x+1)\over pk}\right] +\ln\left[1+{(x-1)\over pk}\right]\right]\cr &+{(pk)}^2\left[{2x\over pk}-{(1-3x^2)\over {(pk)^2}}+{2x^3\over{3(pk)}^3}\right]\cr }\no$$ We note that $$\eqalign{\ln{(1+{(x+1)\over pk})}+\ln{(1+{(x-1)\over pk})} &=\sum_{m=1}^{\infty} {{(-1)}^{(m+1)}\,\left[{(x+1)}^{m} +{(x-1)}^{m}\right]\over m\,\,{(pk)}^{m}}\cr }\no$$ Hence combining this with $\hat A_{p}'(k,x,0)$ gives $$\eqalign{\hat A_{p}'(k,x,0) &=\sum_{m=4}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{(pk)}^{(m-2)}}\cr &\qquad+2x\sum_{m=3}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{(pk)}^{(m-1)}}\cr &+x^2\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{(pk)}^{m}}\cr }\no$$ which gives $$\eqalign{\hat A_{p}'(x,0) &=\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{(m+2)}+{(x-1)}^{(m+2)}\right] \over (m+2)\,\,{p}^{m}}\zeta(m)\cr &\qquad-2x\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{(m+1)}+{(x-1)}^{(m+1)}\right] \over (m+1)\,\,{p}^{m}}\zeta(m)\cr &+x^2\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{p}^{m}}\zeta(m)\cr }\no$$ We now observe the identity $$\sum_{m=2}^{\infty}{(-1)}^{m}{a}^{m-1}\zeta(m,\alpha) =\psi(\alpha+a)-\psi(\alpha) \no$$ which yields $$\eqalign{\sum_{m=2}^{\infty}{{(-1)}^{m}{a}^{m+\nu}\over m+\nu}\zeta(a,\alpha) &=\int_{0}^{a}dy\,y^{\nu}\left[\psi(y+\alpha)-\psi(\alpha)\right]\cr }\no$$ for $\nu\geq 0$. Now using this identity we have $$\eqalign{\hat A_{p}'(x,0)&=p^2\int_{0}^{(x+1)/ p}dy\, {(y-{x\over p})}^2 \left[\psi(1+y)-\psi(1)\right]\cr &\qquad+p^2\int_{0}^{(x-1)/ p}dy\, {(y-{x\over p})}^2\left[\psi(1+y)-\psi(1)\right]\cr }\no$$ Finally using $\psi(1)=-\gamma$ we arrive at at our expression for the desired analytic continuation of the derivative at zero, $$\eqalign{A_{p}'(x,0)=&-A_{p}(x,0)\ln{p}^2 -{p^2\over 2\pi^2}\zeta(3)-{1\over2} +{p x\over6} \left[24\zeta '(-1)+1-{6\over p^2}\right] +x^2({3\over 2}- \ln[2\pi])\cr &\qquad+{x^3\over p} +p^2\int_{0}^{(x+1)\over p}{(y-{x\over p})}^2\psi(1+y) +p^2\int_{0}^{(x-1)\over p}{(y-{x\over p})}^2\psi(1+y)\cr}\no$$ We are interested in particular in the symmetric sum $$A_{p}^{+}(x)=A_{p}'(x,0)+A_{p}'(-x,0)\no$$ which from on using $$\Gamma(1+x)\Gamma(1-x)={\pi x\over\sin(\pi x)}\no$$ we find to be $$\eqalign{A_{p}^{+}(x) =-{p^2\over\pi^2}\zeta(3) &-{p^2\over\pi^2}\int_{0}^{(x+1)\pi\over p} dz\, z(z-{2\pi x\over p})\cot(z) -{p^2\over\pi^2}\int_{0}^{(x-1)\pi\over p} dz\, z(z-{2\pi x\over p})\cot(z)\cr &-x^2\ln\left[{2\sin\left({(x+1)\pi\over p}\right)\over x+1}\right] -x^2\ln\left[{2\sin\left({(x-1)\pi\over p}\right)\over x-1}\right]\cr }\no$$ Alternatively one can see this directly from the expansion of $\hat A_{p}^{+}$ in the series representation and the identities $$\sum_{l=1}^{\infty}{1\over l+1}{1\over a^{2l}}\zeta(2l) ={1\over2}-{a^2\over{\pi}^2}\int_{0}^{\pi\over a}dz z^2\cot(z) \no$$ $$\sum_{l=1}^{\infty}{1\over 2l+1}{1\over{a}^{2l}}\zeta(2l) ={1\over2}-{a\over 2\pi}\int_{0}^{\pi\over a}dz z\cot(z) \no$$ $$\sum_{l=1}^{\infty}{1\over l}{1\over{a}^{2l}}\zeta(2l) =-\ln\left[{\sin({\pi\over a})\over{\pi\over a}}\right] \no$$ A useful expression is obtained by summing over all conjugacy classes to obtain an expression independent of $p$. Hence note that $$\eqalign{\sum_{j=0}^{p-1}\sum_{k=1}^{\infty}A_{p}(k,-j,s) &=\sum_{\{\nu_{p}\}}\sum_{k=1}^{\infty} {{(pk-j)}^2\over{\{{(pk-j)}^2-1\}}^s}\cr &=\sum_{n=1}^{\infty}{{(n+1)}^2\over{\{{(n+1)}^2-1\}}^s}\cr }\no$$ is independent of $p$ since we have merely decomposed it as a sum over conjugacy classes mod $p$. We label this sum $A(s)$ and its derivative at $s=0$ simply $A$. Now denoting $$\eqalign{A_{p} &=A_{p}'(0,0) +\sum_{l=0}^{(p-3)\over2}A_{p}^{+}(2l+1)\cr}$$ As a consequence of the above we have that $$A=A_{p}-\sum_{l=1}^{(p-3)\over2}{(2l+1)}^2\ln[{(2l+1)}^2-1]\no$$ is independent of $p$. This invariant can be written as $$\eqalign{ A&=-{p^3\over2\pi^2}\zeta(3) -\sum_{l=1}^{(p-3)\over2}{(2l+1)}^2\ln[{(2l+1)}^2-1] -{p^2\over\pi^2}\int_{0}^{\pi\over p}dz z^2 \cot(z)\cr &-{p^2\over\pi^2}\sum_{l=0}^{(p-3)\over2} \left[\int_{0}^{(2l+2)\pi\over p}dz z(z-{2\pi(2l+1)\over p})\cot(z) +\int_{0}^{2l\pi\over p}dz z(z-{2\pi(2l+1)\over p})\cot(z)\right]\cr &-\sum_{l=0}^{(p-3)\over2}{(2l+1)}^2 \left[\ln\left[{2\sin\left({(2l+2)\pi\over p}\right)\over 2l+2}\right] +\ln\left[{2\sin\left({2l\pi\over p}\right)\over 2l}\right]\right]\cr }\no$$ Which on simplifying gives $$\eqalign{A &=-{p^3\over 2\pi^2}\zeta(3) -{(p-2)}^2\ln[2\sin({\pi\over p})]-\ln[{\pi\over p}] -{p^2\over\pi^2}\int_{0}^{\pi\over p}dz z^2\cot(z)\cr &\qquad-{p^2\over\pi^2}\int_{0}^{(p-1)\pi\over p}dz z(z-{2(p-2)\pi\over p})\cot(z)\cr &\qquad-2\sum_{l=1}^{(p-3)\over2}\left[ {p^2\over\pi^2}\int_{0}^{2l\pi\over p}dz z(z-{4l\pi\over p})\cot(z) +(4l^2+1)\ln[2\sin({2l\pi\over p})]\right]\cr }\no$$ This constant has the value $A=-1.20563$ One can use this together with the expression for $B_{p}^{-}(x)$ obtained in the next appendix to give an alternative form of the expression for $\tp(p)$. Explicitly $$\eqalign{\tp(p)=&{2\over p}A_{p}'(0,0) +{2\over p}\sum_{l=0}^{(p-3)/2} \left[A_{p}^{+}(2l+1)-(2l+1)B_{p}^{-}(2l+1)\right]\cr =&{2\over p}\left[A -\sum_{l=1}^{(p-3)/p}(2l+1)\left[B_{p}(2l+1)-(2l+1) \ln\left[{(2l+1)}^2-1\right]\right]\right]}$$ which works out to be $$\eqalign{\tp(p) &={2\over p}\left[A +\ln [{\pi\over p}] +{(p-2)}^2\ln[2\sin({\pi\over p})]-{(p-2)p\over \pi}\int_{0}^{(p-1)\pi\over p}dz z\cot(z) \right.\cr &\qquad\left.-2{p\over \pi}\sum_{l=1}^{(p-3)\over2}2l\int_{0}^{2l\pi\over p}dz z \cot(z) +\sum_{l=1}^{(p-3)\over2}(4l^2+1) \ln\left[2\sin({2l\pi\over p})\right]\right]\cr }\no$$ In this appendix we analyze the function $$B_{p}(x,s) =\sum_{k=1}^{\infty}{(pk+x)\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ We are again interested in the value of this function and its derivative with respect to $s$ at $s=0$. For this we denote $$B_{p}(k,x,s)={(pk+x)\over{\left\{{(pk+x)}^2-1\right\}}^s}\no$$ The continuation now requires us to extract the appropriate large $k$ behaviour from $B_{p}(x,s)$. Proceeding to do this we find $$\eqalign{B_{p}(k,x,s)=& {(pk)}^{1-2s}-(2s-1)x{(pk)}^{-2s}+s(1+(2s-1)x^2){(pk)}^{-2s-1} +\hat B_{p}(k,x,s)\cr }\no$$ Hence summing over $k$ yields $$B_{p}(x,s)={1\over {p}^{2s}}\left[p\zeta(2s-1) -x(2s-1)\zeta(2s)+{s\left(1+(2s-1)x^2\right)\over p}\zeta(2s+1)\right]+\hat B_{p}(s,x)\no$$ where $$\eqalign{\hat B_{p}(x,s)&=\sum_{k=1}^{\infty}\left[ {{(pk+x)}\over{\left\{{(pk+x)}^2-1\right\}}^s} -{1\over {(pk)}^{2s-1}}\left\{1-x{(2s-1)\over pk}\right.\right.\cr &\left.\left.+{s\left(1+(2s-1)x^2\right)\over{(pk)}^2} -{x\over3}{s(2s+1)\left(3+(2s-1)x^2\right)\over {(pk)}^3} \right\}\right]\cr }\no$$ $\hat B_{p}(x,s)$ has the property that the process of taking the limit differentiating with respect to $s$ and taking $s\rightarrow 0$ commutes with the sum over $k$. The second property is that $\hat B_{p}(x,0)=0$. Hence this expression has the analytic continuation to $s=0$ $$\eqalign{B_{p}(x,0)&=p\zeta(-1,1+{x\over p})+{1\over 2p}\cr &=-{p\over 12}-{x\over 2}+{(1-x^2)\over 2p}\cr }\no$$ Differentiating with respect to $s$ and evaluating at $s=0$ yields $$\eqalign{B_{p}'(x,0)&= -B_{p}(x,0)\ln p^2 +2p\zeta '(-1)+2x\left[\zeta '(0)-\zeta(0)\right] +{x^2\over p}+{\gamma(1-x^2)\over p} + \hat B_{p}'(x,0)\cr }\no$$ We choose to do the evaluate $\hat B_{p}'(x,0)$ by doing the derivative first and the sum last. Therefore $$\eqalign{\hat B_{p}'(k,x,0)&=\sum_{m=2}^{\infty} {{(-1)}^{(m+1)}\left[{(1+x)}^{(m+1)}+{(x-1)}^{(m+1)}\right] \over (m+1)\,\,{pk}^{m}}\cr &\qquad+x\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{pk}^{m}}\cr }\no$$ and $$\eqalign{\hat B_{p}'(x,0)&=\sum_{m=2}^{\infty} {{(-1)}^{(m+1)}\left[{(1+x)}^{(m+1)}+{(x-1)}^{(m+1)}\right] \over (m+1)\,\,{p}^{m}}\zeta(m)\cr &\qquad+x\sum_{m=2}^{\infty} {{(-1)}^{m}\left[{(1+x)}^{m}+{(x-1)}^{m}\right] \over m\,\,{p}^{m}}\zeta(m)\cr }\no$$ Again using the summation as in appendix A we have $$\eqalign{\hat B_{p}'(x,0)&=-p\,\int_{0}^{(x+1)/ p}dy\, (y-{x\over p})\left(\psi(1+y)-\psi(1)\right)\cr &\qquad-p\,\int_{0}^{(x-1)/ p}dy\, (y-{x\over p})\left(\psi(1+y)-\psi(1)\right)\cr} \no$$ Thus we arrive at $$\eqalign{B_{p}'(x,0)&=-B_{p}(x,0)\ln p^2 +2p\zeta '(-1)+\left[1-\ln[2\pi]\right]x +{x^2\over p}\cr &-p\,\int_{0}^{(x+1)/ p}dy\, (y-{x\over p})\psi(1+y) -p\,\int_{0}^{(x-1)/ p}dy\, (y-{x\over p})\psi(1+y)\cr }\no$$ For the function $B_{p}$ it is the anti-symmetric part that contributes to the quantities of interest thus defining $$\eqalign{B_{p}^{-}(x)&=B_{p}'(x,0)-B_{p}'(-x,0)\cr &=x\ln{p}^2+\left[1-\ln[2\pi]\right]2x+\hat B_{p}^{-}(x)\cr }\no$$ we have $$\eqalign{B_{p}^{-}(x) &= {p\over\pi}\int_{0}^{(x+1)\pi/p}dz\,z\cot(z) +{p\over\pi}\int_{0}^{(x-1)\pi/ p}dz\, z\cot(z)\cr &\qquad-x\ln\left[{2\sin\left({(x+1)\pi\over p}\right)\over x+1}\right] -x\ln\left[{2\sin\left({(x-1)\pi\over p}\right)\over x-1}\right]\cr }\no$$ One can combine the results above with those in Appendix A to obtain an expression for $H_{p}'(x,0)$, where we have defined $$H_{p}(x,s)=A_{p}(x,s)-xB_{p}(x,s)\no$$ Explicitly we have $$\eqalign{H_{p}'(x,0)&=-H_{p}(x,0)\ln{p}^2 -{p^2\over 2\pi^2}\zeta(3)-{1\over2} + {p x\over 6}\left(12\zeta '(-1) + 1 - {6\over p^2}\right)\cr &\qquad+ {x^2\over2}+p^2\int_{0}^{(x+1)\over p} y(y-{x\over p})\psi(1+y) +p^2\int_{0}^{(x-1)\over p}dy y (y-{x\over p})\psi(1+y)\cr }\no$$ where $$H_{p}(x,0)=-{px\over12}-{x\over 2p}+{x^3\over 6p}\no$$ By analogy with $A_{p}^{+}(x)$ we can define $$H_{p}^{+}(x)=A_{p}^{+}(x)-xB_{p}^{-}(x)\no$$ which is given by $$H_{p}^{+}(x)=-{p^2\over\pi^2}\zeta(3) -{p^2\over\pi^2}\int_{0}^{(x+1)\pi/p}dz\,z(z-{\pi x\over p})\cot(z) -{p^2\over\pi^2}\int_{0}^{(x-1)\pi/ p}dz\,z(z-{\pi x\over p})\cot(z)\no$$ The task of this appendix is to obtain an expressions for $\zeta '(0,1+a)$, $\zeta '(-1,1+a)$ and $\zeta '(-2,1+a)$. We begin with $\zeta '(0,1+a)$. Note first of all that $$\zeta(s,1+a)=\sum_{n=1}^{\infty}{1\over{(n+a)}^{s}}\no$$ has the series expansion $$\zeta(s,1+a)=\zeta(s)+s\zeta(s+1)a+\hat\zeta(s,1+a)\no$$ where $$\eqalign{\hat\zeta(s,1+a)&=\sum_{n=1}^{\infty}\left[{1\over{(n+a)}^s} -{1\over n^s}+{sa\over n^{s+1}}\right]\cr &=\sum_{k=2}^{\infty}{{(-1)}^{k}\Gamma(s+k)\over\Gamma(s)k!} \zeta(s+k)a^k\cr }\no$$ Which gives the analytic continuation for $\zeta(s,1+a)$ to $s=0$. We therefore deduce $$\zeta(0,1+a)=\zeta(0)-a\no$$ and $$\zeta '(0,1+a)=-{1\over 2}\ln[2\pi]-\gamma a+\hat\zeta '(0,1+a)$$ where we used $\zeta '(0)=-{1\over2}\ln[2\pi]$ and $s\zeta(s+1)=1+\gamma s+\dots$. Differentiation with respect to $s$ at $s=0$ and summation over $n$ commute for $\hat\zeta(s,1+a)$ we therefore have $$\eqalign{\hat\zeta '(0,1+a)&=-\sum_{n=1}^{\infty}\left[ \ln[1+{a\over n}]-{a\over n}\right]\cr &=\sum_{k=2}^{\infty}{{(-1)}^k a^k\over k}\zeta(k)\cr }\no$$ Thus using we have $$\eqalign{\hat\zeta '(0,1+a) &=\int_{0}^{a}dy\,\left[\psi(1+y)-\psi(1)\right]\cr }\no$$ We therefore obtain that $$\zeta '(0,1+a)=-{1\over2}\ln[2\pi]+\ln\Gamma(1+a)\no$$ The symmetric part of this function on using can be expressed as is expressed as $$\zeta '(0,1+a)+\zeta '(0,1-a)=-\ln\left[{2\sin(\pi a)\over a}\right]\no$$ Finally by noting that $$\zeta(s)=p^{-s}\sum_{j=0}^{p-1}\zeta(s,1-{j\over p})$$ is a decomposition over conjugacy classes of $\zeta(s)$ Differentiating and evaluating at $s=0$ we obtain $$\zeta '(0)=-\sum_{j=0}^{p-1}\zeta(0,1-{j\over p})\ln p +\sum_{j=0}^{p-1}\zeta '(0,1-{j\over p})$$ Now the first term is just a decomposition of $\zeta(0)$ over conjugacy classes therefore gives the identity $$\sum_{j=0}^{p-1}\zeta '(0,1-{j\over p})=\zeta '(0)+\zeta(0)\ln p$$ which for $p$ odd gives $$\sum_{l=1}^{(p-1)/2} \ln\left[2\sin({2l\pi\over p})\right]={1\over2}\ln p$$ or $$\sum_{l=0}^{(p-1)/2} \ln\left[2\sin\left({2l+1)\pi\over p}\right)\right]={1\over2}\ln p$$ Now for $p$ even takes the form $$\sum_{l=0}^{r-1}\left[\zeta '(0,1-{l\over r}) +\zeta '(0,1-{2l+1\over 2r})\right]=\zeta '(0)+\zeta(0)\ln[2r]$$ The first term in is of the form of the original expresson and therefore the resulting expression is $$\sum_{l=0}^{r-1}\zeta '(0,1-{2l+1\over 2r})=-{1\over2}\ln2$$ For $p=4t+2$ we divide the $l$-summation into the two ranges, $0$ to $t-1$, and, $t$ to $2t-1$; rearanging the latter sum and using then makes then gives us $$-\sum_{l=0}^{t-1}\ln\left[2\sin\left({2l+1)\pi\over p}\right)\right]+\zeta '(0,{1\over2})=-{1\over2}\ln2$$ Now noting that $\zeta '(0,{1\over2})=-{1\over2}\ln2$ we have the identity. $$\sum_{l=1}^{(p-6)/4}\ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right] =0$$ We find for $p=4t$, by a similar procedure, that $$\sum_{l=0}^{(p-4)/2}\ln\left[2\sin\left({(2l+1)\pi\over p}\right)\right] ={1\over 2}\ln2$$ These are the first set of identities we obtain by decomposing over conjugacy classes. Next we turn to $\zeta '(-1)$, thus note that $$\zeta(s-1,1+a)=\sum_{n=1}^{\infty}{1\over{(n+a)}^{(s-1)}}\no$$ has the series expansion in terms of $a$ $$\zeta(s-1,1+a) =\sum_{k=0}^{\infty}{{(-)}^{k}\Gamma(s+k-1)\,a^k \over\Gamma(s-1)k!}\zeta(k+s-1)\no$$ which on extracting the divergent part gives $$\zeta(s-1,1+a)=\zeta(s-1)-(s-1)a\zeta(s) +{s(s-1)a^2\over2}\zeta(s+1)+\hat\zeta(s-1,1+a)\no$$ where $$\hat\zeta(s-1,1+a) =\sum_{n=1}^{\infty}\left[{1\over{(n+a)^{s-1}}}-{1\over n^{s-1}} +{(s-1)a\over n^s}-{s(s-1)a^2\over n^{s+1}}\right]\no$$ This is now expressed in terms of our prescription for analytic continuation. Hence we have $$\zeta(-1,1+a)=-{1\over 12}-{a(1-a)\over 2}\no$$ where $\zeta(-1)=-{1/2}$, $\zeta(0)=-{1/2}$ and $s\zeta(s+1)=1+\gamma s+\dots$ have been used. Differentiation with respect to $s$ and setting it to zero yields $$\zeta '(-1,1+a)=\zeta '(-1) +\left[1-\ln[2\pi]\right]{a\over2}+{(\gamma-1)a^2\over 2} +\hat\zeta '(-1,1+a)\no$$ Since differentiation with respect to $s$ at $s=0$ and summation over $n$ commute we have $$\eqalign{\hat\zeta '(-1,1+a) &=-(n+a)\ln(1+{a\over n})+a+{a^2\over 2n}\cr &=-\sum_{k=2}^{\infty}{(-1)^{k}a^{k+1}\over k+1}\zeta(k) +a\sum_{k=2}^{\infty}{{(-)}^{k}a^k\over k}\zeta(k)\cr }\no$$ Now using the identity $$\sum_{k=2}^{\infty}{{(-)}^{k}a^{k+\nu}\over k+\nu} \zeta(k)=\int_{0}^{a}dy\,y^{\nu}\left(\psi(1+y)-\psi(1)\right) \no$$ valid for $a<1$ to remain away from the poles of $\psi(y)$ which occur at $y=-n$ for $n=0,1,2,\dots$ we have $$\hat\zeta '(-1,1+a) =-\int_{0}^{a}dy\,(y-a)\left[\psi(1+y)-\psi(1)\right] \no$$ and therefore $$\zeta '(-1,1+a)=\zeta '(-1)+\left[1-\ln(2\pi)\right]{a\over2} -{a^2\over 2} -\int_{0}^{a}dy\,(y-a)\psi(1+y)\no$$ Finally we conclude that the anti-symmetrised sum is given by $$\zeta '(-1,1+a)-\zeta '(-1,1-a)= %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR \left[1-\ln[2\pi]\right]a-\int_{0}^{a}dy\,(y-a)\left[\psi(1+y)-\psi(1-y)\right]\no$$ Now using $$\psi(1+y)-\psi(1-y) =-{d\over dy}\ln\left[{\sin(\pi y)\over\pi y}\right]\no$$ we obtain $$\zeta '(-1,1+a)-\zeta '(-1,1-a) = -a\ln\left[{2\sin(\pi a)\over a}\right]+{1\over\pi} \int_{0}^{\pi a}dz\,z\cot(z)\no$$ An alternative way of expressing this which is useful in the main text and simplifies some of the expressions is $$\zeta '(-1,a)-\zeta '(-1,1-a) =-{1\over\pi}\int_{0}^{\pi a}\ln\left[2\sin(z)\right]\no$$ Let us now examine the consequences of decomposing over conjugacy classes. Proceeding as for $\zeta '(0)$, on decomposing over conjugacy classes we obtain the identity $$\zeta '(-1)={1\over 12}\ln p+p\sum_{j=0}^{p-1}\zeta '(-1,1-{j\over p})$$ For $p=2r+1$ this yields the identity $$\eqalign{\zeta '(-1)&={1\over p^2-1}\left[p\sum_{l=1}^{(p-1)/2}\left[ {2l\over p}\ln\left[{2l\over p}\right]+{\left({2l\over p}\right)}^2\right.\right.\cr &\qquad\left.\left. +\int_{0}^{2l/p}dy\,\left[\psi(1+y)+\psi(1-y)\right]\right]-{1\over 12}\ln p\right]\cr }$$ i.e. $$\eqalign{\zeta '(-1)&={1\over 6}-{1\over 4}\ln p\cr &\qquad+{1\over p^2-1}\left[\sum_{l=1}^{(p-1)/2}\left[ 2l\ln[2l]+p\int_{0}^{2l/p}dy\,\left[\psi(1+y)+\psi(1-y)\right]\right] -{1\over 12}\ln p\right]\cr }$$ This can be used numerically to verify that $\zeta '(-1)=0.16791$. Now noting that $$\int_{0}^{\pi(1-a)}dz\,\ln\left[2\sin(z)\right]=-\int_{0}^{\pi a}dz\, \ln\left[2\sin(z)\right]$$ We find $$\sum_{j=1}^{p-1}\int_{0}^{j\pi/p}dz\,\ln\left[2\sin(z)\right]=0$$ For $p$ even we find that by observing the two decompositions $$\zeta '(-1,1+\alpha)=-\zeta '(-1,1+\alpha)\ln p +p\sum_{l=0}^{p-1}\zeta '(-1,1-{l-\alpha\over p})$$ for $\alpha$ not necessarily integer and $$\zeta '(-1,1+\alpha)=-\zeta(-1,1+\alpha)\ln p +p\sum_{l=1}^{p}\zeta '(-1,{l+\alpha\over p})$$ valid for arbitrary $p$ can be used to simplify the expression for $$\sum_{l=1}^{(p-2)/2}\int_{0}^{(2l+1)\pi/p}dz\,\ln\left[2\sin(z)\right] ={2\ln2\over p}-{1\over 2}\ln2-{4\over p}\ln[p/2]$$ These are used in simplifying the expression for the torsion in the case of $p$ even. We now turn to $\zeta '(-2,1+a)$ which we obtain from $$\zeta(s-2,1+a)=\sum_{n=1}^{\infty}{1\over {(n+a)}^{s-2}}\no$$ We again extract the divergent parts from the sum to obtain $$\eqalign{\zeta(s-2,1+a)&=\zeta(s-2)-(s-2)a\zeta(s-1) +{(s-1)(s-2)a^2\over 2}\zeta(s)\cr &\qquad-{s(s-1)(s-2)\zeta(s+1)a^3\over 6}+\hat\zeta(s-2,1+a)\cr}\no$$ where $$\eqalign{\hat\zeta(s-2,1+a)&=\sum_{n=1}^{\infty}\left[ {1\over{(n+a)}^{s-2}}-{1\over n^{s-2}}+{(s-2)a\over{n}^{s-1}}\right.\cr &\qquad\left.-{(s-1)(s-2)a^2\over 2{n}^{s}} +{s(s-1)(s-2)a^3\over6 n^{s+1}}\right]\cr}\no$$ Thus $$\zeta(-2,1+a)=-{a\over 6}-{a^2\over 2}-{a^3\over 3}\no$$ and differentiating with respect to $s$ and evaluating at $s=0$ we find $$\eqalign{\zeta '(-2,1+a)&=\zeta '(-2) +\left[2\zeta '(-1)+{1\over12}\right]a +\left[{3\over 4}-{1\over2}\ln[2\pi]\right]a^2 \cr &-\left[{\gamma\over 3}-{1\over2}\right]a^3+\hat\zeta '(-1,1+a)\cr }\no$$ We evaluate $\hat\zeta '(-2,1+a)$ by differentiating with respect to $s$ first and then performing the sum over $n$ obtaining $$\eqalign{\hat\zeta '(-2,1+a) &=-\sum_{n=1}^{\infty}\left[ {(n+a)}^2\ln[1+{a\over n}] -n\,a-{3a^2\over2}-{a^3\over 3n}\right]\cr &=\sum_{k=2}^{\infty}{{(-1)}^{k}a^{k+2}\over k+2}\zeta(k) -2a\sum_{k=2}^{\infty}{{(-1)}^{k}a^{k+1}\over k+1}\zeta(k) +a^2\sum_{k=2}^{\infty}{{(-1)}^ka^k\over k}\zeta(k)\cr }\no$$ Thus using the summation formula we have $$\hat\zeta '(-2,1+a) =\int_{0}^{a}{(y-a)}^2\left[\psi(1+y)-\psi(1)\right] \no$$ We therefore obtain $$\eqalign{\zeta '(-2,1+a) &=\zeta '(-2) +{1\over12}\left[24\zeta '(-1)+1\right]a +{1\over4}\left[3-2\ln[2\pi]\right]a^2\cr &\qquad+{1\over2}a^3 +\int_{0}^{a}dy\,{(y-a)}^2\psi(1+y)\cr } \no$$ We note that our expression for $\zeta '(-2,1+a)$ implies $$\eqalign{\zeta '(-2,1+a)+\zeta '(-2,1-a) &=2\zeta '(-2) +{1\over2}\left[3-2\ln[2\pi]\right]a^2\cr &\qquad+\int_{0}^{a}dy {(y-a)}^2\left[\psi(1+y)-\psi(1-y)\right]\cr }\no$$ Differentiating the functional relation $$\xi(s)=\xi(1-s),\quad\hbox{ where }\xi(s)=\Gamma(s/2)\pi^{-s/2}\zeta(s)\no$$ with respect to $s$ we find for $s=-2$ $$\zeta '(-2)=-{1\over 4\pi^2}\zeta(3)\no$$ Finally using and we have $$\eqalign{\zeta '(-2,1+a)+\zeta '(-2,1-a) &=-{1\over 2\pi^2}\zeta(3)-a^2\ln\left[{2\sin(\pi a)\over a}\right]\cr &\qquad-{1\over\pi^2}\int_{0}^{\pi a}dz\,z (z-2\pi a)\cot(z)\cr }\no$$ We can obtain an identity by decomposing over conjugacy classes mod $p$, by noting that $$\zeta(s-2)=p^{2-s}\sum_{j=0}^{p-1}\zeta(s-2,1-{j\over p})\no$$ implies $$\zeta '(-2)=p^2\sum_{j=0}^{p-1}\zeta(-2,1-{j\over p})\ln p +p^2\sum_{j=0}^{p-1}\zeta '(-2,1-{j\over p}) \no$$ The first sum is a decomposition of $\zeta(-2)$ and therefore zero. If we decompose $j$ into odd and even elements, by setting $j=2l+1$ and $j=2l$ respectively, we find for $p$ odd i.e. $p=2r+1$ $$\zeta '(-2)=p^2\zeta '(-2)+p^2\sum_{l=1}^{r} \left[\zeta '(-2,1-{2l\over p})+\zeta '(-2,1-{2l+1\over p})\right] \no$$ which can be rewritten as either $$\zeta '(-2)=p^2\zeta '(-2) +p^2\sum_{l=1}^{r}\left[\zeta '(-2,1-{2l\over p}) +\zeta '(-2,{2l\over p})\right] \no$$ or $$\zeta '(-2)=p^2\zeta '(-2)+p^2\sum_{l=0}^{r-1} \left[\zeta '(-2,{2l+1\over p})+\zeta '(-2,1-{2l+1\over p})\right] \no$$ Therefore we have that $$\zeta '(-2)=p^3\zeta '(-2)-\sum_{l=1}^{r}\left[4l^2\ln\left[2\sin({2l\pi\over p})\right] +{p^2\over\pi^2}\int_{0}^{2l\pi/p}dz\, z(z-{4l\pi\over p})\cot(z)\right]\no$$ which on using gives $${(1-p^3)\over 4\pi^2}\zeta(3)= \sum_{l=1}^{r}\left[4l^2\ln\left[2\sin({2l\pi\over p})\right] +{p^2\over\pi^2}\int_{0}^{2l\pi/p}dz\, z(z-{4l\pi\over p})\cot(z)\right]\no$$ or $$\eqalign{\zeta '(-2)&=p^3\zeta '(-2)-\sum_{l=0}^{r-1}\left[ {(2l+1)}^2\ln\left[2\sin({(2l+1)\pi\over p})\right]\right.\cr &\qquad\left.+{p^2\over \pi^2}\int_{0}^{(2l+1)\pi/p}dz\, z(z-{2(2l+1)\pi\over p})\cot(z)\right]\cr} \no$$ which gives $$\eqalign{{(1-p^3)\over4\pi^2}\zeta(3)&=\sum_{l=0}^{r-1} \left[{(2l+1)}^2\ln\left[2\sin({(2l+1)\pi\over p})\right]\right.\cr &\qquad\left.+{p^2\over \pi^2}\int_{0}^{(2l+1)\pi/p}dz\, z(z-{2(2l+1)\pi\over p})\cot(z)\right]\cr} \no$$ We can equally establish that $${(1-p^3)\over 4\pi^2}\zeta(3)= \sum_{j=1}^{r}\left[j^2\ln\left[2\sin({j\pi\over p})\right] +{p^2\over\pi^2}\int_{0}^{j\pi/p}dz\, z(z-{2j\pi\over p})\cot(z)\right]\no$$ Finally we tabulate here for future reference some useful identities regarding the Hurwitz and Riemann zeta functions $$\zeta(s,a)={a}^{-s}+\zeta(s,1+a)\no$$ $$\eqalign{ \zeta(-2)&=0\qquad\zeta(-1,1+a)-\zeta(-1,1-a)=-a\cr &\zeta(0,a)+\zeta(0,1-a)=0\qquad\zeta(0,1+a)+\zeta(0,1-a)=-1\cr \zeta '(0,1+a)&=\zeta '(0,a)+\ln a\qquad \zeta '(-1,1+a)=\zeta '(-1,a)+a\ln a\cr \zeta '(0,1+a)+\zeta '(0,1-a) &=-\ln\left[{2\sin(\pi a)\over a}\right]\cr}\no$$ $$\eqalign{ \zeta '(-1,1+a)-\zeta '(-1,1-a) &=-a\ln\left[{2\sin(\pi a)\over a}\right] +{1\over\pi}\int_{0}^{\pi a}dz\,z \cot(z)\cr \zeta '(-1,a)-\zeta '(-1,1-a) &=-a\ln(2\sin\pi a)+{1\over\pi}\int_{0}^{\pi a}dz\, z \cot(z)\cr }\no$$ and $$\eqalign{\zeta '(-2,1+a)+\zeta '(-2,1-a) &=-{1\over 2\pi^2}\zeta(3)-a^2\ln\left[{2\sin(\pi a)\over a}\right]\cr &\qquad-{1\over\pi^2}\int_{0}^{\pi a}dz\,z (z-2\pi a)\cot(z)\cr }\no$$ $$\eqalign{\zeta '(-2,a)+\zeta '(-2,1-a) &=-{1\over 2\pi^2}\zeta(3)-a^2\ln\left[2\sin(\pi a)\right] -{1\over\pi^2}\int_{0}^{\pi a}dz\,z (z-2\pi a)\cot(z)\cr }\no$$ References 0.5	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'library.bib' --- [Wurtzite InP Microdisks: from Epitaxy to Room-Temperature Lasing]{} Philipp Staudinger^+^, Svenja Mauthe^+^, Noelia Vico Triviño, Steffen Reidt, Kirsten E. Moselund, and Heinz Schmid\* IBM Research Zurich, Säumerstrasse 4, 8803 Rüschlikon, Switzerland\ ^+^ These authors contributed equally to this work.\ \ Correspondence to [email protected]* [r]{}[0.35]{} ![image](Figures/FigTOC.pdf){width="35.00000%"} Metastable wurtzite crystal phases of semiconductors comprise enormous potential for high-performance electro-optical devices, owed to their extended tunable direct band gap range. However, synthesizing these materials in good quality and beyond nanowire size constraints has remained elusive. In this work, the epitaxy of wurtzite InP microdisks and related geometries on insulator for optical applications is explored. This is achieved by an elaborate combination of selective area growth of fins and a zipper-induced epitaxial lateral overgrowth, which enables co-integration of diversely shaped crystals at precise position. The grown material possesses high phase purity and excellent optical quality characterized by STEM and -PL. Optically pumped lasing at room temperature is achieved in microdisks with a lasing threshold of 365 J/cm^2^, thus demonstrating promise for a wide range of photonic applications. [2]{} ![image](Figures/Fig1_concept.pdf){width="\textwidth"} ![image](Figures/Fig2_SEM.pdf){width="\textwidth"} ![image](Figures/Fig3_TEM.pdf){width="\textwidth"} ![image](Figures/Fig4_OpticalChar.pdf){width="\textwidth"} III-V materials have attracted great interest throughout the past decades, due to their unique optical and electronic properties. The direct and tunable bandgap allows for scaled electro-optical devices with efficient emission and absorption properties. Novel concepts like on-chip optical communication [@Miller:17], complex optical networks [@Gaio2019], topological photonics [@Zhao2018], or metasurfaces [@Liu2016] have been proposed and rely on the unique features of scaled III-V devices such as microdisk lasers [@Wirths2018] or photonic crystals [@Akabori2003]. Using a top down approach based on wafer bonding, lithography and etching, advanced III-V nanostructures on insulator can readily be formed for photonic applications [@Crosnier2017; @Morthier2015; @DeGroote2016; @Roelkens2007]. These approaches however, come at the cost of fabrication complexity and surface damage created during the etching process which negatively impacts device performance. An alternative approach is the direct monolithic growth of semiconductor micro- and nanostructures [@Mayer2016; @Wang2015; @Kunert2016; @Lourdudoss2012; @Han2019]. Advancements in epitaxy techniques recently enabled the synthesis of high quality single crystalline material in various shapes, like nanowires [@Tomioka2011], nano- and microfins [@Seidl2019], microdisks [@Mauthe2019], or microrings [@Mayer2019]. In contrast to semiconductor geometries obtained from etching, these structures are enclosed by as-grown crystal facets with superior quality. Moreover, materials that are not available in bulk form, like metastable wurtzite (WZ) III-Vs, can be grown and investigated [@Gao2014b; @Li2015; @Kitauchi2010]. Synthesizing these materials in their thermodynamically less stable phase is challenging, but highly interesting in terms of optical properties. For instance, the entire composition spectrum of WZ InAlGaP shows a direct band gap transition, in contrast to their natural zinc-blende (ZB) counterparts [@De2010; @Assali2013a; @Gagliano2018]. Hence, metastable III-V materials are seen as a potential solution for the long-standing problem of realizing efficient green LEDs and laser diodes, commonly referred to as “green gap” [@AufDerMaur2016a]. Using conformal epitaxy or zipper-induced epitaxial lateral overgrowth (ELO) techniques, we demonstrated the growth of extended and pure WZ InP layers with layer dimensions ranging from <1 m^2^ to >100 m^2^ [@Staudinger2018a; @Staudinger2020]. In addition, the growth of various WZ microstructures, such as membrane-, prism-, and ring-like shapes by selective area epitaxy was recently shown [@Wang2019]. In this work, we present a versatile optical platform by demonstrating the growth of various InP micro- and nanostructures on an InP substrate with 300 nm oxide to provide optical isolation from the substrate. The structures are grown in a single growth run and thorough optical characterization is performed. The platform allows for exact positioning of epitaxially grown hexagons and related micro- and nano-sized shapes enclosed by a subset of the low-energy m- and a-plane facets, such as triangles, rhombi, fins, stars, or wires. Complex arrangements, e.g. arrays of hexagons with diameters ranging from 500 nm to 5 m and precise spacing can be achieved. Since the epitaxially grown crystals are only connected to the substrate via thin fins and otherwise separated by an oxide layer, optical isolation is achieved. Devices are fabricated using a three-step growth sequence: (1) inducing the transition of crystal phase to WZ, (2) anisotropic growth of WZ-fins and (3) zipper-induced ELO. The main steps of the formation process for a cavity of a microdisk laser are sketched in Figure \[fig1\](a)-(d). A standard InP(111)A substrate is covered by 300 nm PECVD SiO~2~ after which narrow (\~50 nm) lines are patterned by e-beam lithography and dry-etching techniques (see for details on the fabrication). The diameter of the microdisk will be determined by the length of the lines which are connected centrally and follow the three equivalent $\langle$110$\rangle$-directions. Metal-organic vapor phase epitaxy (MOVPE) allows for the selective nucleation of InP in the trenches and switching to WZ phase, followed by a highly selective and vertical growth along the \[111\]A direction. After the fins extend out of the oxide layer, the zipper-points at the center induce an ELO process until the stable {1-100} (m-plane) or metastable {11-20} (a-plane) WZ facets are formed [@Wang2019]. Figure \[fig1\](e) depicts a scanning electron microscope (SEM) image of a typical InP microdisk obtained during a growth time of 5 min, and after removing the oxide masking layer to reveal the underlying pedestal fin structure. The resulting microdisk has noticeable lower sidewall surface roughness compared to the underlying fin structures as well as to InP structures obtained by reactive ion etching processes. The growth technique offers interesting options for designing a specific photonic structure, but also combinations of structures that can be grown in a single step. This includes hexagons of different sizes which can resemble either (nano-)wires or (micro-)disks. Crystal heights are mainly determined by deposition time, but also by diffusion mechanisms, which results in higher growth rates for hexagons with smaller diameter as illustrated in Figure \[fig2\](a). Moreover, the resulting features are accurately defined by lithography and the respective resolution limit, including gap sizes as demonstrated in Figure \[fig2\](b). Periodic arrays consisting of multiple structures, like hexagons, can be formed. Additionally, a wide range of different geometries and polygons can be grown together with no additional fabrication steps after epitaxy. This principle is exemplified in Figure \[fig2\](c) and (d). Angled lines expand to rhomboids [@Wang2019] that are enclosed by low energy facets and form structures that could serve as waveguides. Similarly, an elongated hexagon can be obtained by extending one of the seed lines. In principle, any convex shape can be envisioned which can be formed by m- and a-type facets, including for example triangles and rectangles. The well controlled and deliberate expansion of the exemplified structures beyond the width of the underlying fin together with the thick oxide mask layer are key for optical isolation and photonic functionality. To obtain high quality optical devices from InP nanostructures, excellent crystalline quality needs to be achieved. Hence, we investigate the crystal structure and phase purity of the grown structures. Figure \[fig3\](a) depicts a cross-sectional scanning transmission electron microscope (STEM) image of a typical microdisk with 1.5 m diameter imaged along a $\langle$110$\rangle$ zone-axis. Substrate and microdisk are divided by an oxide layer (dark), which is interrupted by three vertical bright stripes. These correspond to the line openings (compare Figure \[fig1\](a)), as they are observed in a lamella which is slightly off-centered with respect to the hexagon-axis (inset of Figure \[fig3\](a)). In order to investigate the crystal phase and defects we perform HR-STEM characterization in Figure \[fig3\](b)-(g). We observe a transition from ZB to WZ phase after the crystal has extended approximately 5 nm into the oxide template. This is attributed to the initial growth stage at lower temperature, combined with the surface roughness of the substrate. At closer inspection the abruptness of this phase transition is revealed. The high-angle annular dark field (HAADF) image in Figure \[fig3\](c) suggests that only a single bilayer (stacking fault) divides a pure ZB (ABCABC…) from a pure WZ (ABAB…) stacking sequence. Following this transition, fault-free WZ phase is observed (exemplified in two additional HR-STEM images) up to the flat top facet in the central area. Close to the side-facets, a ring-like structure terminates the top-surface, which is approximately 30 nm high. In stark contrast to the phase purity of the microdisk below, this ring shows ZB phase with high amount of stacking faults as indicated in Figure \[fig3\](g). We assign this to the continued growth at reduced temperatures during the first seconds of the cool-down (see ). In order to assess optical properties, the devices are characterized using a -photoluminescence (PL) setup with a ps-pulsed supercontinuum laser at 750 nm wavelength. A 100× objective is used to illuminate the devices from the top and collect their optical response. Figure \[fig4\](a) depicts the photoluminescence response of a 2 m WZ microdisk for increasing excitation fluences (20 – 295 J/cm^2^). At low illumination, the spontaneous emission of the microdisk is visible with a peak position of 880 nm corresponding to the bandgap energy of WZ InP and hence, confirming the crystal phase [@Staudinger2018a]. Upon increasing laser power, a resonant mode starts forming at 875 nm. This can be observed in Figure \[fig4\](b), where the integrated linear light in light out (LL) curve of the measured cavity mode is plotted. A clear kink marks the onset of lasing. By performing a linear fit on the right-hand side of the graph, the threshold of the resonant cavity mode can be determined to 365 J/cm^2^. In logarithmic representation, a clear S-shape is revealed which is characteristic for a laser. Alongside with the strong increase of the emission in the resonant cavity mode, we observe a linewidth narrowing, another indication for lasing (see inset in Figure \[fig4\](b)). The increase of the linewidth at high pump powers is due to a blueshift of the resonant wavelength arising from refractive index changes and is widely observed in similar systems [@Bennett1990; @Wang2015]. Additional insight can be obtained by analyzing far field images of the optical response taken with a standard camera. As shown in Figure \[fig4\](c) the spontaneous emission is visible at low excitation fluences. With increasing pump energy, the emission of the devices gets stronger and fringes start forming. At high excitation fluences, a clear far-field radiation pattern with interference fringes is visible because of the extended first-order coherence of the laser emission. Room-temperature lasing operation in as-grown microdisks confirms the sufficient optical isolation provided by the 300 nm thick oxide layer introduced during fabrication, which was chosen for this first proof-of-concept. To further quantify the material quality and the lasing performance, we perform time-correlated single photon counting (TCSPC) measurements. The sample is again illuminated with a 750 nm ps-pulsed supercontinuum laser using a 100× objective. Figure \[fig4\](d) depicts the results of the TCSPC measurements. At room temperature and low excitation fluence, the ZB substrate and WZ microdisk exhibit carrier lifetimes of 1.57 ns and 239 ps, respectively. This is in agreement with results obtained in a previous work [@Staudinger2018a], and is attributed to the higher oscillator strength of the optical transition at the $\Gamma$-point in WZ phase [@PhysRevB.73.235308; @Wilhelm2012]. Under strong excitation power, the measured carrier lifetime of the WZ microdisk reduces strongly resulting in a value below the resolution limit of the measurement setup (<50 ps). This strong reduction of carrier lifetime is attributed to the stimulated emission in the resonant mode observed under high excitation powers, which further confirms lasing operation of the hexagonal microdisk with laser pulse durations below <50 ps. This constitutes, to the best of our knowledge, the first demonstration of a WZ InP microdisk laser. In combination with the possibility to controllably grow a range of different geometries side-by-side, our platform could have many promising future photonic applications. In conclusion, we developed a methodology based on selective area epitaxy specifically tailored for the fabrication of photonic devices. High resolution lithographic patterning and thick oxide masks enable precise positioning and optical isolation, while zipper-ELO leads to large devices with defect-free and smooth crystal surfaces. Besides hexagonal microdisks we show a range of additional shapes to exemplify the capabilities of our platform. The unique fabrication approach allowed for the demonstration of optically driven lasing from hexagonal WZ InP cavities with a threshold of 365 J/cm^2^. Finally, the concepts introduced are generic and could therefore also find applications in other commercially important material systems such as InAlGaP and III-nitrides. Experimental Section {#Experimental .unnumbered} -------------------- Substrate fabrication: An InP(111)A wafer was covered with nominally 300 nm plasma-enhanced chemical vapor deposited (PECVD) SiO~2~ (288 nm measured during STEM). This thickness provides a compromise between sufficient optical isolation of the optical mode in the III-V material and fabrication simplicity. Line openings with widths of approximately 50-100 nm and varying lengths were patterned along the three equivalent $\langle$110$\rangle$-directions by e-beam lithography (EBL) and reactive ion etching (RIE) in a CHF~3~/O~2~ plasma. Prior to growth the substrate was cleaned in acetone, isopropanol and O~2~ plasma. MOVPE growth: InP growth was carried out in a cold-wall showerhead MOCVD reactor using H~2~ carrier gas, tertiarybutylphosphine (TBP) and trimethylindium (TMIn) at a total pressure of 8000 Pa. Partial pressures for TBP and TMIn were 8.2 Pa and 82 mPa, respectively, resulting in a nominal V/III ratio of 100. Before growth, the reactor was heated to 630 °C in a TBP atmosphere. Subsequently, deposition was initiated by introducing TMIn into the chamber. The substrate temperature was ramped up to 640 °C within the first 60 s, the final growth temperature. This procedure was developed in order to prevent substrate desorption prior the nucleation process. Deposition time was between 5 min and 30 min. Rapid cooling was started for the last 50 s of the growth, after which TMIn supply was terminated and cooling proceeded in TBP atmosphere. Optical characterization: Optical characterization is performed using a -photoluminescence (PL) setup with a ps-pulsed supercontinuum laser at 750 nm (78 MHz repetition rate). The sample is illuminated from the top using a 100× objective (NA 0.6) which also collects the optical response of the sample. All measurements are performed at 300 K and under ambient conditions. The collected optical response is analyzed using a monochromator and an InGaAs CCD camera. Time-correlated single photon counting (TCSPC) measurements are performed in the same setup, using the 750 nm ps-pulsed supercontinuum laser and the 100× objective. A Si single photon detector is used to measure the optical response. Since the signal of the sample is very strong, a high-density filter was used to reduce the power of optical light incident on the detector. Using a TCSPC measurement module (PicoHarp 300), the lifetime of the sample can be determined. The resolution of the setup is 50 ps due to the temporal pulse shape of the supercontinuum laser. Acknowledgements {#acknowledgements .unnumbered} ---------------- The authors gratefully acknowledge Preksha Tiwari, Marilyne Sousa, Anna Fontcuberta i Morral and Heike Riel for fruitful technical discussions, as well as, the BRNC staff for technical support. The work presented here has received funding from the European Union H2020 program SiLAS (Grant Agreement No. 735008) and the ERC Starting Grant project PLASMIC (Grant Agreement No. 678567). References {#references .unnumbered} ----------	{ "pile_set_name": "ArXiv" }
--- author: - \| [Supreeth Shastri]{}\ [Computer Science]{}\ [University of Texas at Austin]{} - \| [Melissa Wasserman]{}\ [School of Law]{}\ [University of Texas at Austin]{} - \| [Vijay Chidambaram]{}\ [Computer Science]{}\ [University of Texas at Austin]{} bibliography: - 'paper.bib' title: \| The Seven Sins of\ Personal-Data Processing Systems Under GDPR ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'Centaurus A (Cen A) is one of the most famous galaxies hosting an active galactic nucleus (AGN), where the interaction between AGN activities and surrounding interstellar and intergalactic media has been investigated. Recent studies reported detections of the H$\alpha$ emission from clouds in the galactic halo toward the northeast and southwest of the nucleus of Cen A, suggesting that AGN jets may have triggered star formation there. We performed near-infrared line mapping of Cen A with the IRSF 1.4-m telescope, using the narrow-band filter tuned for Pa$\beta$, from which we find that the Pa$\beta$ emission is not detected significantly from either northeast or southwest regions. The upper limit of the Pa$\beta$/H$\alpha$ ratio in the northeast region is compatible with that expected for a typical H region, in line with the scenario that AGNs have triggered star formation there. On the other hand, the upper limit of Pa$\beta$/H$\alpha$ in the southwest region is significantly lower than that expected for a typical H region. A possibility to explain the low Pa$\beta$/H$\alpha$ ratio in the southwest region is the scattering of H$\alpha$ and Pa$\beta$ photons from the center of Cen A by dust grains in the halo clouds. From the upper limit of Pa$\beta$/H$\alpha$ in the southwest region, we obtain constraints on the dust size distribution, which is found to be compatible with those seen in the interstellar medium of our Galaxy.' author: - 'Risako <span style="font-variant:small-caps;">Katayama</span>' - 'Hidehiro <span style="font-variant:small-caps;">Kaneda</span>' - 'Takuma <span style="font-variant:small-caps;">Kokusho</span>' - 'Kumiko <span style="font-variant:small-caps;">Morihana</span>' - 'Toyoaki <span style="font-variant:small-caps;">Suzuki</span>' - 'Shinki <span style="font-variant:small-caps;">Oyabu</span>' - 'Mitsuyoshi <span style="font-variant:small-caps;">Yamagishi</span>' - 'Takuro <span style="font-variant:small-caps;">Tsuchikawa</span>' title: ' Hydrogen recombination near-infrared line mapping of Centaurus A with IRSF/SIRIUS ' --- Introduction ============ Active galactic nuclei (AGNs) are known to play an important role in the evolution of host galaxies through injecting kinetic energy into the surrounding interstellar and intergalactic media (ISM and IGM) via outflows or jets ([@HB], and references therein); however it is yet to be understood how this AGN feedback mechanism affects the host galaxy. It has often been considered that AGN blows out cold gas from the host galaxy or prevents gas from cooling to suppress star formation (i.e., negative feedback; e.g., [@Fabian]; [@Harrison]). On the contrary, it is also suggested that the AGN activity can compress gas and trigger star formation (i.e., positive feedback; e.g., [@Silk]; [@Zinn]). Hence, the effect of AGN feedback is likely to be not straightforward ([@ZB]; [@Shin]). The observational evidence of AGN-triggered star formation has been reported for nearby galaxies harboring AGNs, such as Minkowski’s Object and 3C 285. Minkowski’s Object is a star-forming peculiar object, located at the end of the radio jet from the nearby galaxy NGC 541, which has long been suggested that its star formation is triggered by the radio jet ([@Brodie]; [@vanB]; [@Salome]). At a distance of nearly $70\rm\ kpc$ from the galaxy center of 3C 285, @vanBD discovered a small H$\alpha$-emitting object near the eastern radio jet (3C 285/09.6), suggesting jet-induced star formation through compressing dense surrounding material. Centaurus A (Cen A or NGC 5128), at a distance of 3.8 Mpc ([@Harris]), is the radio galaxy nearest to us, hosting an AGN which emits powerful radio and X-ray jets on the scales ranging from 1.35 kpc up to 250 kpc ([@Israel]). Thus, Cen A is an important target to understand interaction between AGN jets and the surrounding ISM and IGM. At a distance of approximately 8 kpc from the nucleus, optically bright filaments and possible young stars therein are observed in regions close to the northeastern radio jet, which could be the result of jet-ISM interaction ([@Blanco]; [@Osmer]; [@GP]; [@Morganti]; [@Rejkuba]; [@Crockett]). Additionally, recent studies report detection of hydrogen recombination line emissions in the galactic halo around the jets, indicating jet-induced star formation. In the northeast of Cen A, @Santoro find an H$\alpha$-emitting region at about 15 kpc away from the nucleus, suggesting that the AGN may have indeed triggered star formation there. On the opposite side of that region with respect to the nucleus, @Keel also find an H$\alpha$-emitting region in a halo cloud at about 12 kpc away from the nucleus along the axis of the southwestern jet lobe, again suggesting a positive AGN feedback taking place in this region. On the other hand, there is a possibility of the scattering of H$\alpha$ photons from a host galaxy by dust grains in halo clouds. At about $4\rm\ kpc$ far from the galaxy disk of M 82, strongly polarized H$\alpha$ emission is observed, and its polarization pattern suggests that the H$\alpha$ emission originates from the dust scattering of H$\alpha$ photons coming from the central near-InfraRed (IR) nucleus ([@Scarrott]; [@Yoshida]). Therefore, in order to evaluate the effect of the AGN feedback accurately, it is important to investigate whether or not the hydrogen recombination line is truly of the origin of star formation therein. In the previous studies, only a small portion of Cen A was studied. In order to investigate the interaction between AGN feedback and the intergalactic clouds in Cen A, however, it is required to observe wider areas of Cen A along the axis of the jet with the hydrogen recombination lines tracing star-forming regions. Thus in the present study, we perform a wider area mapping of Cen A along the axis of the jet using near-IR narrow-band filters. We obtain the physical parameters of the hydrogen recombination lines, combining our data with the data of the previous studies, to verify the possibility of the dust scattering in the halo clouds of Cen A. Observations and data analysis ============================== Near-IR imaging toward Cen A was carried out in March, May and June 2019 with the SIRIUS (Simultaneous InfraRed Imager of Unbiased Survey; [@Nagashima], [@Nagayama]) camera on the InfraRed Survey Facility (IRSF) 1.4-m telescope at the South African Astronomical Observatory. We observed Cen A using the narrow-band filter tuned for Pa$\beta$ 1.282$\rm\ \mu m$ with the dithering of 10 frames and exposure time of 60 seconds per frame. The effective bandwidth of the filter for Pa$\beta$ is 0.029$\rm\ \mu m$. In addition, to subtract the Pa$\beta$ continuum, we also observed Cen A using the notch filter which has two peaks of transmittance at about 1.268 and 1.297$\rm\ \mu m$. The camera has a field of view of $\timeform{7'.7}\times \timeform{7'.7}$ with a pixel scale of $\timeform{0''.45}$. We observed five fields in total, and summarized the detail of the observation of each field in Table \[sum\_ob\]. For the continuum filter, we observed for twice as long total exposure times as those listed in Table \[sum\_ob\]. \[tab:first\] The image data were processed with the standard data reduction by using the pipeline software of pyIRSF[^1], which includes dark subtraction, flat-fielding, sky subtraction and dithered-image-combining. We performed astrometric and photometric calibrations with the 2MASS Point Source Catalog (PSC; [@Skrutskie]), assuming that the magnitudes of point sources are the same between the $J$-band and the narrow-band images of Pa$\beta$ and the continuum. For photometry, we selected stars with the $J$-band fluxes higher than 12.5 mag which are not saturated, and with errors smaller than 0.05 mag from the 2MASS PSC. Then, determining sky regions in each field and subtracting the sky emission from the images, we adjusted the sky level of each observed field to be zero. Finally, we subtracted the Pa$\beta$-continuum image from the narrow-band image to obtain the Pa$\beta$ intensity map of Cen A. Results ======= Figure \[32map\] shows the Pa$\beta$ intensity map of Cen A obtained with the narrow-band filters, from which we confirm the presence of the Pa$\beta$ emission extended along the galaxy structure, but not clearly associated with the jet lobes. Figure \[NE-SW\] shows the close-up images of the northeast and southwest regions where the H$\alpha$ emission is detected in the previous studies. In the northeast region, @Santoro spectroscopically estimated the H$\alpha$ fluxes of Regions A and B to be $2.3\times10^{-15}$ and $3.8\times10^{-15}\rm\ erg\ s^{-1}\ cm^{-2}$, respectively. Similarly, in the southwest region, @Keel measured the H$\alpha$ flux densities from 4 H$\alpha$-bright local blobs of Main, Main $N$, Diffuse, and North with the integral-field spectroscopy, to be $9.3\times10^{-15}$, $1.3\times10^{-15}$, $1.1\times10^{-15}$ and $1.0\times10^{-15}\rm\ erg\ s^{-1}\ cm^{-2}$, respectively. We performed aperture photometry with a radius of $\timeform{1.''5}$ for Regions A, B, and 4 H$\alpha$-bright local blobs using the present Pa$\beta$ map. As a result, we find that the Pa$\beta$ lines are not detected significantly from any of these regions (Table \[photometry flux\]); the 3$\sigma$ upper limits of the Pa$\beta$ fluxes measured within the northeast and southwest regions are $7.2\times10^{-16}$ and $6.2\times10^{-16}\rm\ erg\ s^{-1}\ cm^{-2}$, respectively, where we use the sum of the photometry results for both regions. Assuming the Case B recombination with the electron density $n_e=10^2\rm\ cm^{-3}$ and electron temperature $T_e=10^4\rm\ K$ for a typical H region, we estimate that the Pa$\beta$ flux from the H$\alpha$ flux of the previous study in the northeast region would be $3.4\times10^{-16}\rm\ erg\ s^{-1}\ cm^{-2}$. Thus the 3$\sigma$ upper limit of the observed flux ($<7.2\times10^{-16}\rm\ erg\ s^{-1}\ cm^{-2}$) is compatible with the Pa$\beta$ flux which is expected from the Case B recombination, indicating that the jet-induced star formation may have indeed taken place. On the other hand, in the southwest region, we similarly estimate the Pa$\beta$ flux from the H$\alpha$ fluxes to be 7.2$\times10^{-16}\rm\ erg\ s^{-1}\ cm^{-2}$, assuming the Case B recombination, which is not compatible with the observed 3$\sigma$ upper limit in the southwest region ($<6.2\times10^{-16}\rm\ erg\ s^{-1}\ cm^{-2}$). Thus, the H$\alpha$ fluxes from this region may not be attributed to the emission of ionized gas in a star-forming region. ![(a) Pa$\beta$+continuum, (b) continuum, and (c) continuum-subtracted Pa$\beta$ maps of Centaurus A. The maps have been smoothed with a Gaussian kernel of $\timeform{2''.7}$ in sigma, and color scales are given in units of $\rm erg\ s^{-1}\ cm^{-2}\ arcsec^{-2}$. In panel (c), contours show the Herschel 500$\rm\ \mu m$ map from @Auld, where the contour levels correspond to $0.10, 0.27, 0.72, 1.9, 5.2, 14, 37\rm\ mJy\ beam^{-1}$. Red squares whose sizes are $30''\times45''$ and $60''\times90''$ indicate the northeast and the southwest regions where the H$\alpha$ emission is detected by @Santoro and @Keel, respectively. []{data-label="32map"}](map_color.eps){width="16cm"} ![Comparison of the H$\alpha$ and the Pa$\beta$ line intensity maps of the northeast and the southwest regions of Centaurus A, which correspond to the red squares denoted in Figure \[32map\], overlaid with the same contours of the Herschel 500$\rm\ \mu m$ map as in Figure \[32map\]. The Pa$\beta$ maps have been smoothed with a Gaussian kernel of $\timeform{1''.13}$ in sigma. The H$\alpha$ maps of the northeast and southwest regions are both taken from @Keel[]{data-label="NE-SW"}](NE_SW.eps){width="16cm"} \[tab:second\] Discussion ========== We obtain the observed upper limits of the Pa$\beta$/H$\alpha$ ratios of $1.2\times10^{-1}$ and $4.9\times10^{-2}$ in the northeast and southwest regions, respectively, which are compared with those predicted by models. In particular, we obtain the strongest constraint on the upper limit of the Pa$\beta$/H$\alpha$ ratio of 3.5$\times 10^{-2}$ for Main in the southwest region which is the brightest H$\alpha$ blob, and discuss its implications below. First, we calculate the Pa$\beta$/H$\alpha$ ratios expected for the Case A and B recombinations for a typical H region, as shown in Table \[CaseAB\], which are not consistent with the observed upper limit of the Pa$\beta$/H$\alpha$ ratio. Even if we assume the Case B recombination for wide ranges of $n_e$ and $T_e$ of $10^2-10^6\rm\ cm^{-3}$ and $5000-20000\rm\ K$, respectively, the expected Pa$\beta$/H$\alpha$ ratios exceed $5.3\times10^{-2}$ ([@OF]). Thus, we find that neither Case A nor B satisfies the observed upper limit of the Pa$\beta$/H$\alpha$ ratio in the southwest region. Then we consider the effect of dust extinction; however it would further increase the Pa$\beta$/H$\alpha$ ratio to be observed, because the Pa$\beta$ emission is less attenuated by dust compared to the H$\alpha$ emission. It is possible that the jet-driven shock excites the optical emission line. For example, @Sutherland suggest that the inner emission-line filaments in Cen A at approximately 8 kpc away from the nucleus of Cen A can be produced by the interaction between the jet and the surrounding ISM. Actually we estimate the Pa$\beta$/H$\alpha$ ratio assuming shock excitation with $n_e$ and $T_e$ of $10^2-10^4$ and $500\rm\ K$, respectively, to find that the resultant Pa$\beta$/H$\alpha$ ratio is larger than $2.1\times 10^{-2}$ ([@OF]), which is compatible with the observed 3$\sigma$ upper limit of the Pa$\beta$/H$\alpha$ ratio. In this shock excitation model, however, a strong radio jet needs to be present in the vicinity of the emission-line filaments. In the northeast region of Cen A, @Santoro15 claim that the jet-cloud interaction is unlikely to be important because the jet is too diffuse in this region. Thus, the jet-driven shock excitation may not be taking place in the southwest region as well, where the galactic cloud is 12$\rm\ kpc$ away from the nucleus and the jet is as diffuse as in the northeast region. Yet another possibility to explain the observed upper limit of the Pa$\beta$/H$\alpha$ ratio in the southwest region is the dust scattering of the H$\alpha$ and Pa$\beta$ photons emitted from the center of the host galaxy (Table \[CaseAB\]). In this case, the scattering efficiency of Pa$\beta$ photons is expected to be significantly lower than that of H$\alpha$ photons. Indeed, in the southwest region, @Auld identified a dust cloud in the 500-$\rm \mu m$ emission, who excluded the possibility that the 500-$\rm \mu m$ emission is due to synchrotron radiation. A scattering scenario would also be useful to explain the low dust temperature and the lack of UV emission observed by @Auld for the dust cloud in the southwest region. On the other hand, @Keel find that the optical spectrum in the southwest region is consistent with photoionization in typical H regions. Since scattering preserves local line ratios, it would suggest scattering not so much of AGN radiation as that from the inner starburst disk. The dust-scattered H$\alpha$ emission from the host galaxy is actually observed in the galactic halo of M 82 ([@Yoshida]). Assuming that it is also the case in Cen A, we measured the Pa$\beta$ flux of the host galaxy with an aperture radius of $\timeform{2'.5}$. The resultant Pa$\beta$ flux is $(8.0\pm 0.7)\times 10^{-13}\rm\ erg\ s^{-1}\ cm^{-2}$ received by us and $(8.0\pm 0.7)\times 10^{-8}\rm\ erg\ s^{-1}\ cm^{-2}$ by dust in the southwestern cloud with the distance of 12 kpc from the galaxy to the cloud. Using these results, we calculate the H$\alpha$ flux of the host galaxy for the Case B recombination, and then we estimate the Pa$\beta$ and H$\alpha$ fluxes scattered by dust in the southwestern cloud. First, in order to get a rough picture, we tentatively use silicate and graphite dust of a constant size of $0.35\rm\ \mu m$ in radius, for which the scattering coefficients at wavelengths of the Pa$\beta$ and H$\alpha$ lines, $Q_{\rm Pa\beta}$ and $Q_{\rm H\alpha}$, are taken from @LD. As a result, we find that the estimated Pa$\beta$/H$\alpha$ ratio, both scattered by the silicate and graphite dust of $0.35\rm\ \mu m$ in size in the southwestern cloud, would be $(3.0\pm 0.3)\times10^{-2}$ and $(5.2\pm 0.4)\times10^{-2}$, respectively (Table \[CaseAB\]). Hence the silicate dust scattering is compatible with the 3$\sigma$ upper limit of the observed Pa$\beta$/H$\alpha$ ratio in the southwest region, whereas the graphite dust scattering is not. \[tab:third\] low density limit, $T_e=10^4\rm\ K$([@OF])\ $n_e=10^2\rm\ cm^{-3}, T_e=10^4\rm\ K$([@OF])\ Scattering efficiency from @LD. A single dust size of $0.35\rm\ \mu m$ is assumed. The scattered Pa$\beta$/H$\alpha$ photons originate from the Case B recombination lines from the host galaxy.\ Then we assume a more realistic dust size distribution such as that in @MRN with $Q_{\rm Pa\beta}$ and $Q_{\rm H\alpha}$ in @LD, the result of which is shown in Table \[MRN scattering\]. We adopt the ranges of silicate and graphite dust sizes of $0.025-0.25\rm\ \mu m$ and $0.005-1\rm\ \mu m$, respectively (MRN77). Table \[MRN scattering\] also shows the results for power-law indices different from that of MRN77. As compared to Table \[CaseAB\], the Pa$\beta$/H$\alpha$ ratios in Table \[MRN scattering\] are relatively small due to the presence of dust smaller than 0.35 $\rm \mu m$ in size, and therefore the observed result favors the presence of such small-size dust. \[tab:forth\] Consequently, the size distribution following the MRN77 power-law function can reproduce the Pa$\beta$/H$\alpha$ ratio compatible with the observed upper limit. As shown in Table \[MRN scattering\], as the power-law index is steeper, the compatibility with our observational result is more secure. Actually, in the galactic halo, @HL show a relative high abundance in small grains with a radius of less than 0.03$\rm\ \mu m$ with their numerical simulation, which may be caused by dust shattering during the galactic outflow processes. In order to give a stronger constraint on the dust size distribution in this region, we need further observations of Pa$\beta$ with higher sensitivity. Conclusion ========== We have conducted near-IR line mapping of Cen A with IRSF, using the narrow-band filter tuned for the Pa$\beta$ line to study the interaction between the AGN activity and the surrounding IGM in the northeast and the southwest regions of the halo of Cen A. In the previous studies, the H$\alpha$ emission is detected in these regions ([@Santoro]; [@Keel]), suggesting that AGN-triggered star formation may have taken place therein. Our observations, however, do not detect the Pa$\beta$ emission, and in the southwest region, the upper limit of the Pa$\beta$/H$\alpha$ ratio is significantly lower than that expected for a typical H region. One possibility to explain the observed ratio is the scattering of H$\alpha$ and Pa$\beta$ photons from the center of the host galaxy by dust in the galactic halo. If this is the case, the H$\alpha$ fluxes detected in these regions are not relevant to the AGN-induced star formation, but just reflecting the presence of dusty clouds where the H$\alpha$ photons originating from the galactic center of Cen A are scattered by the dust grains therein. According to the scattering scenario, the observed 3$\sigma$ upper limit of the Pa$\beta$/H$\alpha$ ratio is consistent with the size distribution of the interstellar dust in our Galaxy. Finally our result suggests the importance of observing Pa$\beta$ emission in addition to H$\alpha$ to identify the origin of the H$\alpha$ emission in the halo of a galaxy. We are grateful to the referee for giving us useful comments. The IRSF project is a collaboration between Nagoya University and the SAAO supported by the Grants-in-Aid for Scientific Research on Priority Areas (A) (Nos. 10147207 and 10147214) and Optical & Near-Infrared Astronomy Inter- University Cooperation Program, from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan and the National Research Foundation (NRF) of South Africa. K.M. is financially supported by MEXT/JSPS KAKENHI Grand number 17K18019. Auld, R., et al. 2012, , 420, 1882 Blanco, V. M., Graham, J. A., Lasker, B. M., & Osmer, P. S. 1975, , 198, L63 Brodie, J. P., Bowyer, S., & McCarthy, P. 1985, , 293, L59 Crockett, R. M., et al. 2012, , 421, 1603 Draine, B. T. & Lee, H. M. 1984, , 285, 89 Fabian, A. C. 2012, Annual Review of Astronomy and Astrophysics, 50, 455 Graham, J. A. & Price, R. M. 1981, , 247, 813 Harris, G. L. H., Rejkuba, M., & Harris, W. E. 2010, PASA, 27, 457 Harrison, C. M. 2017,Astronomy 1, 0165 Heckman, T. M. & Best, P. N. 2014, , 52, 589 Hirashita, H., & Lin, C. Y. 2018, Planet. Space Sci, 183, 104504 Israel, F. P. 1998, , 8, 237 Keel, W. C., Banfield, J. K., Medling, A. M., & Neff, S. G., 2019, , 157, 7 Laor, A. & Draine, B. T. 1993, , 402, 441 Mathis, J. S., Rumpl, W., & Nordsieck, K. H., 1977, , 217, 425 Morganti, R., et al. 1991, , 249, 91 Nagashima, C., Nagayama, T., & Nakajima, Y., 1999, Star Formation 1999, ed. T. Nakamoto (Nagano: Nobeyama Radio Obs.), 397 Nagayama, T., Nagashima, C., & Nakajima, Y., 2003, Proc. SPIE, 4841, 459 Osmer, P. S. 1978, , 226, L79 Osterbrock, D. E., & Ferland, G. J., ed. 2006, Astrophysics of Gaseous Nebulae and Active Galactic Nuclei 2nd edition (Sausalito, California: University Science Books) , 72-299 Rejkuba, M., Minniti D., Courbin, F., & Silva, D R. 2002, , 564, 688 Salomé, Q., Salomé, P., & Combes, F. 2015, , 574, 9 Santoro, F., Oonk, J. B. R., Morganti, R., Oosterloo, T. A., & Tadhunter, C. 2015, , 574, 6 Santoro, F., Oonk, J. B. R., Morganti, R., Oosterloo, T. A., & Tadhunter, C. 2016, , 590, 7 Scarrott, S. M., Eaton, N., & Axon, D. J. 1991, , 252, 12 Shin, J., Woo, J., Chung, A., Baek, J., Cho, K., Kang, D., Bae, H. 2019, , 881, 13 Silk, J. 2013, , 772, 10 Skrutskie, M. F., Cutri, R. M., & Stiening, R. et al. 2006, , 131, 1163 Sutherland, R. S., Bicknell, G. V., & Dopita, M. A. 1993, , 414, 510 Weingartner, J. C., & Draine, B. T., 2001, , 548, 296 Yoshida, M., Kawabata, K. S., Ohyama, Y., Itoh, R., & Hattori, T. 2019, , 71, 87 van Breugel, W. J. M. & Dey, A. 1993, , 414, 563 van Breugel, W. J. M., Filippenko, A. V., Heckman, T., & Miley, G. 1985, , 293, 83 Zinn, P. C., Middleberg, E., Norris, R. P., & Dettmar, R. J., 2013, , 774, 10 Zubovas, K., & Bourne, M. A. 2017, , 468, 4956 [^1]: https://sourceforge.net/projects/irsfsoftware/	{ "pile_set_name": "ArXiv" }
--- abstract: 'Multipath propagation is a common phenomenon in wireless communication. Knowledge of propagation path parameters such as complex channel gain, propagation delay or angle-of-arrival provides valuable information on the user position and facilitates channel response estimation. A major challenge in channel parameter estimation lies in its multidimensional nature, which leads to large-scale estimation problems which are difficult to solve. Current approaches of sparse recovery for multidimensional parameter estimation aim at simultaneously estimating all channel parameters by solving one large-scale estimation problem. In contrast to that we propose a sparse recovery method which relies on decomposing the multidimensional problem into successive one-dimensional parameter estimation problems, which are much easier to solve and less sensitive to off-grid effects, while providing proper parameter pairing. Our proposed decomposition relies on convex optimization in terms of nuclear norm minimization and we present an efficient implementation in terms of the recently developed STELA algorithm.' author: - '[^1]' bibliography: - 'refferences.bib' nocite: - '[@mdRare]' - '[@mdRare]' title: \| Multidimensional Sparse Recovery for\ MIMO Channel Parameter Estimation --- =0.11cm =0.11cm =0.09cm =0.09cm MIMO Channel Parameters, Multidimensional Parameter Estimation, Sparse Recovery, Nuclear Norm, STELA Introduction ============ Mobile communication generally underlies multipath propagation, due to reflection, refraction or scattering on surrounding objects. In a static environment, a propagation path can be characterized by a complex gain coefficient, a propagation delay and an angle-of-arrival (AoA) at the receiver. Knowledge of the channel parameters can be exploited in different ways, e.g., to estimate the user locations by exploiting the characteristics of the dominant multipath components [@4558782], or to estimate the channel response using parametric channel models [@sparseMultipath]. The channel parameter estimation problem can be formulated as a multidimensional (MD) estimation problem [@1547647]-[@sparseUnderwater]. In the simplest form the estimation could be performed by means of an MD Fourier transform, leading, however, to poor resolution in the case of multiple propagation paths. Better performance can be achieved by subspace-based methods. In [@1547647] a variation of the MUSIC algorithm for joint MD channel parameter estimation has been presented, which is applicable to arbitrary spatio-temporal sampling. However, since the MD-MUSIC is a spectrum-search based estimator, high estimation accuracy requires a fine MD parameter grid which, in turn, leads to large and complex estimation problems, i.e. in terms of the computational complexity associated with MD-peak search. Search-free methods, such as the MD-RARE [@mdRare] and MD-ESPRIT [@mdEsprit], rely on specific structures in the spatio-temporal sampling, e.g., uniform or centrosymmetric sampling. For proper performance, subspace-based methods further rely on a good estimate of the signal covariance matrix which either depends on a large number of signal snapshots or techniques like forward-backward averaging and smoothing, where the latter techniques require centrosymmetric and uniform spatio-temporal sampling, respectively. A drawback of smoothing techniques is the fact that the effective sampling aperture is reduced which degrades the resolution capability. In recent years, sparse recovery techniques came into the focus of parameter estimation research and, similarly to subspace-based methods, have been shown to provide the superresolution property [@doi:10.1137/0523074]. However, in contrast to subspace-based methods, sparse recovery techniques do not rely on the estimation of a signal covariance matrix and thus show good performance even in the case of low number of signal snapshots or irregular sampling. Two major categories of sparse recovery methods are greedy algorithms, e.g., orthogonal matching pursuit, and convex relaxation techniques, e.g., LASSO or basis pursuit, where the latter category usually shows better estimation performance at the cost of higher computational complexity. For the application of parameter estimation, a major drawback of sparse recovery methods is the requirement for a fine discretization of the parameter space to a grid in order to achieve a high resolution and avoid basis mismatch or off-grid effects [@BasisMismatch]. The fine parameter grid leads to large dictionary matrices and in turn to large-scale estimation problems. In case of MD parameter estimation, this aspect becomes even more critical, since the dictionary size increases exponentially with the number of parameter dimensions. A sparse recovery method for MD channel parameter estimation was presented in [@sparseUnderwater]. The presented method relies on joint discretization of all associated parameter spaces, resulting in a large dictionary matrix. Due to its large size, the corresponding estimation problem can hardly be solved by convex relaxation methods, hence the authors propose a greedy method referred to as Space-Alternating Orthogonal Matching Pursuit (SA-OMP) which mainly addresses the computational complexity involved with MD-peak search and grid refinement. As we will show by numerical experiments, SA-OMP suffers from a large bias in estimation performance, due to its greedy nature. In [@sparseMultipath] a similar approach of discretizing the MD parameter space was presented. The objective in [@sparseMultipath] is, however, not the estimation of the channel parameters, but the estimation of the channel response by exploiting the channel’s sparse representation in the parameter space. In this paper we propose a novel convex problem formulation which decomposes the MD channel parameter estimation problem into successive one-dimensional estimation problems based on convex relaxation in form of nuclear norm minimization. Due to space limitations we restrict our presentation to the two-dimensional parameter estimation problem, but the extension to estimation problems of higher dimensions is straightforward. In contrast to the MD parameter grid for simultaneous parameter estimation, as presented in [@sparseUnderwater], our method requires discretization of one parameter space at a time only. From the estimation results found in the investigated parameter space, the parameters in the remaining dimensions are estimated successively. The successive estimation approach leads to smaller subproblems as compared to the simultaneous estimation of all channel parameters, reduces the sensitivity to off-grid effects caused by inaccuracies in joint discretization of all parameter spaces, and simplifies the peak search. For efficient computation of our novel problem formulation we furthermore present implementation in form of the recently proposed STELA method, which has been shown to have superior convergence speed as compared to, e.g., gradient methods and, admits parallel implementation [@stela]. Signal Model ============ Consider a single-antenna terminal transmitting a reference signal which propagates over a multipath channel to a receiving linear antenna array, as illustrated in Fig. \[fig:model\]. The receiving linear antenna array consists of $M$ omnidirectional antennas with positions given by $r_m \in \mathbb{R}$, for $m=1,\ldots,M$, relative to the first antenna, i.e., $r_1=0$. The reference signal is known to both, the transmitter and receiver. We assume that Orthogonal Frequency Division Multiplexing (OFDM) symbols of length $N$ are transmitted, preceded by a cyclic prefix of duration $T_{\text{cp}}$, sampled at an interval $T_s$, and further assume that the narrowband assumption holds. The OFDM reference signal is denoted as ${\ensuremath{\boldsymbol{x}}}(l) = [x_1(l), \ldots, x_N(l)]^{\text{T}}$ in the frequency domain, where $x_n(l)$ denotes the data symbol, e.g. quadrature phase-shift keying (QPSK) with $\|x_n(l)\|=1$, on the $n$th subcarrier in the $l$th OFDM symbol. The multipath propagation is modeled by $P$ propagation paths. The scattering objects as well as the transmitter are assumed to be located in the farfield region of the antenna array. Each propagation path is characterized by a complex channel gain coefficient $h_{p}^{{\circ}}$, a propagation delay $\tau^{{\circ}}_{p}$ and an AoA $\theta^{{\circ}}_{p}$, with $p=1,\ldots,P$. The maximum propagation delay is assumed to be smaller than the cyclic prefix duration: $T_{\text{cp}} > \max_{p} \tau^{{\circ}}_{p}$. Under the given assumptions, the signal received by the antenna array in time instant $l$ is modeled by the $M \times N$ receive signal matrix $$\begin{aligned} {\ensuremath{\boldsymbol{Y}}}(l) = {\ensuremath{\boldsymbol{A}}}^{{\circ}} {\ensuremath{\boldsymbol{H}}}^{{\circ}}(l) {\ensuremath{\boldsymbol{B}}}^{{\circ}{\text{T}}} {\ensuremath{\boldsymbol{X}}}(l) + {\ensuremath{\boldsymbol{W}}}(l) , \label{eq:fdSignalSmv}\end{aligned}$$ where $[{\ensuremath{\boldsymbol{Y}}}(l)]_{m,n}$ denotes the signal received by antenna $m$ on subcarrier $n$, for $m=1,\ldots,M$ and $n=1,\ldots,N$ (cmp. [@sparseMultipath]-[@sparseUnderwater]). In , the $M \times P$ array steering matrix is given as $$\begin{aligned} {\ensuremath{\boldsymbol{A}}}^{{\circ}} = [ {\ensuremath{\boldsymbol{a}}}(\theta^{{\circ}}_{1}), \ldots, {\ensuremath{\boldsymbol{a}}}(\theta^{{\circ}}_{P}) ],\end{aligned}$$ where ${\ensuremath{\boldsymbol{a}}}(\theta) \!=\! [ 1, {\ensuremath{\text{e}^{-{\ensuremath{\text{j}}}r_2 \xi(\theta) }}}, \scalebox{0.7}[1]{\ldots}, {\ensuremath{\text{e}^{-{\ensuremath{\text{j}}}r_M \xi(\theta) }}} ]^{\text{T}}/ \sqrt{M}$ denotes the normalized array steering vector for AoA $\theta$, with electric angle $\xi(\theta)=\frac{2\pi}{\lambda} \cos \theta$ and signal carrier wavelength $\lambda$. The $N \times P$ frequency response matrix is defined as $$\begin{aligned} {\ensuremath{\boldsymbol{B}}}^{{\circ}} = [ {\ensuremath{\boldsymbol{b}}}(\tau^{{\circ}}_{1}), \ldots, {\ensuremath{\boldsymbol{b}}}(\tau^{{\circ}}_{P}) ]\end{aligned}$$ with ${\ensuremath{\boldsymbol{b}}}(\tau) = [ 1, {\ensuremath{\text{e}^{-{\ensuremath{\text{j}}}\frac{2 \pi}{N T_s} \tau }}}, \scalebox{0.7}[1]{\ldots}, {\ensuremath{\text{e}^{-{\ensuremath{\text{j}}}\frac{2 \pi (N-1)}{N T_s} \tau }}} ]^{\text{T}}/ \sqrt{N}$ denoting the normalized frequency response vector corresponding to a path delay of $\tau$. The $P \times P$ diagonal channel gain matrix $$\begin{aligned} {\ensuremath{\boldsymbol{H}}}^{{\circ}}(l) = {\ensuremath{\text{diag}\left(h_{1}^{{\circ}}(l), \cdots, h_{P}^{{\circ}}(l)\right)}} \end{aligned}$$ contains the complex channel gain coefficients on its main diagonal and the $N \times N$ diagonal reference signal matrix $$\begin{aligned} {\ensuremath{\boldsymbol{X}}}(l) = {\ensuremath{\text{diag}\left(x_1(l), \ldots, x_N(l)\right)}}\end{aligned}$$ contains the elements of the OFDM reference signal on its main diagonal. The elements of the $M \times N$ additive noise matrix ${\ensuremath{\boldsymbol{W}}}(l)$ represent spatially and temporally white Gaussian noise with variance $\sigma_w^2$. ![Multipath signal model for single antenna transmitter and $M=4$ element antenna array and propagation over $P=3$ paths[]{data-label="fig:model"}](channelModelStandalone) We collect $L$ snapshots, as defined in , where we assume that the propagation delays $\{\tau_p^{{\circ}}\}_{p=1}^P$ and AoAs $\{\theta_p^{{\circ}}\}_{p=1}^P$ remain constant, while the channel gain coefficients $\{h_p^{{\circ}}(l)\}_{p=1}^P$ may vary between snapshots. Upon defining $$\begin{aligned} \tilde{{\ensuremath{\boldsymbol{Y}}}} \!\!=\,& [{\ensuremath{\text{vec}({\ensuremath{\boldsymbol{Y}}}(1) {\ensuremath{\boldsymbol{X}}}^{-1}(1))}}, \ldots, {\ensuremath{\text{vec}({\ensuremath{\boldsymbol{Y}}}(L){\ensuremath{\boldsymbol{X}}}^{-1}(L))}}], \label{eq:mmv1}\\ \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} \!\!=\,& [ {\ensuremath{\text{vecd}({\ensuremath{\boldsymbol{H}}}^{{\circ}}(1))}}, \ldots, {\ensuremath{\text{vecd}({\ensuremath{\boldsymbol{H}}}^{{\circ}}(L))}} ], \\ \tilde{{\ensuremath{\boldsymbol{W}}}} \!\!=\,& [{\ensuremath{\text{vec}({\ensuremath{\boldsymbol{W}}}(1){\ensuremath{\boldsymbol{X}}}^{-1}(1))}}, \ldots, {\ensuremath{\text{vec}({\ensuremath{\boldsymbol{W}}}(L){\ensuremath{\boldsymbol{X}}}^{-1}(L))}}], \label{eq:mmv3}\end{aligned}$$ where ${\ensuremath{\text{vec}({\ensuremath{\boldsymbol{Y}}})}}$ denotes vectorization of matrix ${\ensuremath{\boldsymbol{Y}}}$ and ${\ensuremath{\text{vecd}({\ensuremath{\boldsymbol{H}}})}}$ denotes vectorization of the elements on the main diagonal of ${\ensuremath{\boldsymbol{H}}}$, the multiple snapshot signal model is given by $$\begin{aligned} \tilde{{\ensuremath{\boldsymbol{Y}}}} = ( {\ensuremath{\boldsymbol{B}}}^{{\circ}} {\ast}{\ensuremath{\boldsymbol{A}}}^{{\circ}} ) \, \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \label{eq:fdSignalMmv}\end{aligned}$$ where ${\ast}$ denotes the Khatri-Rao product, i.e., the columnwise Kronecker product. A Novel MD Sparse Recovery Method ================================= Following the ideas in [@mdRare], we start by rewriting the signal model in as $$\begin{aligned} \tilde{{\ensuremath{\boldsymbol{Y}}}} =& ( {\ensuremath{\boldsymbol{B}}}^{{\circ}} {\ast}{\ensuremath{\boldsymbol{A}}}^{{\circ}} ) \, \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \nonumber \\ =& ( {\ensuremath{\boldsymbol{B}}}^{{\circ}} \otimes {\ensuremath{\boldsymbol{I}}}_M ) ({\ensuremath{\boldsymbol{I}}}_P {\ast}{\ensuremath{\boldsymbol{A}}}^{{\circ}}) \, \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \nonumber \\ =& ( {\ensuremath{\boldsymbol{B}}}^{{\circ}} \otimes {\ensuremath{\boldsymbol{I}}}_M ) \, {\ensuremath{\boldsymbol{G}}}_A^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \label{eq:nucModel}\end{aligned}$$ with $\otimes$ denoting the Kronecker product and ${\ensuremath{\boldsymbol{I}}}_M$ denoting the $M \times M$ identity matrix. The extended channel gain matrix $$\begin{aligned} {\ensuremath{\boldsymbol{G}}}_A^{{\circ}} = ({\ensuremath{\boldsymbol{I}}}_P {\ast}{\ensuremath{\boldsymbol{A}}}^{{\circ}}) \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}}\end{aligned}$$ of size $MP \times L$ is composed of $P$ submatrices, i.e., ${\ensuremath{\boldsymbol{G}}}_A^{{\circ}} = [{\ensuremath{\boldsymbol{G}}}_{A,1}^{{\circ}{\text{T}}}, \ldots, {\ensuremath{\boldsymbol{G}}}_{A,P}^{{\circ}{\text{T}}}]^{\text{T}}$, where each $M \times L$ submatrix $$\begin{aligned} {\ensuremath{\boldsymbol{G}}}_{A,p}^{{\circ}} = {\ensuremath{\boldsymbol{a}}}(\theta_p^{{\circ}}) \tilde{{\ensuremath{\boldsymbol{h}}}}\vphantom{{\ensuremath{\boldsymbol{h}}}}_p^{{\circ}{\text{T}}} \label{eq:doubleSparse1}\end{aligned}$$ is of rank-one, with $\tilde{{\ensuremath{\boldsymbol{h}}}}\vphantom{{\ensuremath{\boldsymbol{h}}}}_p^{{\circ}}$ representing the $p$th row of $\tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}}$. For a sparse representation of we uniformly discretize the delay parameter space in $Q$ points $\{ \tau_q \}_{q=1}^Q$, with $\max_q \tau_q < T_{\text{cp}}$, and form the $MN \times MQ$ dictionary matrix $$\begin{aligned} {\ensuremath{\boldsymbol{C}}}_B = [{\ensuremath{\boldsymbol{C}}}_{B,1}, \ldots, {\ensuremath{\boldsymbol{C}}}_{B,Q}] = {\ensuremath{\boldsymbol{B}}} \otimes {\ensuremath{\boldsymbol{I}}}_M,\end{aligned}$$ with ${\ensuremath{\boldsymbol{B}}} = \left[ {\ensuremath{\boldsymbol{b}}}(\tau_1), \ldots, {\ensuremath{\boldsymbol{b}}}(\tau_Q) \right]$ representing the $N \times Q$ frequency response dictionary matrix, such that each submatrix ${\ensuremath{\boldsymbol{C}}}_{B,q} = {\ensuremath{\boldsymbol{b}}}(\tau_q) \otimes {\ensuremath{\boldsymbol{I}}}_M$, for $q=1,\ldots,Q$, contains the elements of the frequency response for path delay $\tau_q$. In correspondence with equations - and under the assumption that the true propagation delays lie on the parameter grid, i.e., $\{ \tau_p^{{\circ}} \}_{p=1}^P~\in~\{ \tau_q \}_{q=1}^Q$, we formulate the block sparse matrix $$\begin{aligned} {\ensuremath{\boldsymbol{G}}}_A = [{\ensuremath{\boldsymbol{G}}}_{A,1}^{{\text{T}}}, \ldots, {\ensuremath{\boldsymbol{G}}}_{A,Q}^{{\text{T}}}]^{\text{T}}\end{aligned}$$ of size $MQ \times L$, with $$\begin{aligned} {\ensuremath{\boldsymbol{G}}}_{A,q} = \begin{cases} {\ensuremath{\boldsymbol{G}}}_{A,p}^{{\circ}} \quad & \text{ if } \tau_q = \tau_p^{{\circ}} \\ {\ensuremath{\boldsymbol{0}}} & \text{ else}, \end{cases} \label{eq:doubleSparse2}\end{aligned}$$ such that we arrive at the sparse representation $$\begin{aligned} \tilde{{\ensuremath{\boldsymbol{Y}}}} = \, {\ensuremath{\boldsymbol{C}}}_B {\ensuremath{\boldsymbol{G}}}_A + \tilde{{\ensuremath{\boldsymbol{W}}}} . \label{eq:nucModel2}\end{aligned}$$ As can be seen from and , ${\ensuremath{\boldsymbol{G}}}_{A}$ exhibits two levels of sparsity, block sparsity and rank sparsity. As proposed in [@Steffens] the twofold sparse structure can be exploited in the following convex optimization problem $$\begin{aligned} \hat{{\ensuremath{\boldsymbol{G}}}}_A = \arg \min_{{\ensuremath{\boldsymbol{G}}}_A} \; \frac{1}{2} \big\\| \tilde{{\ensuremath{\boldsymbol{Y}}}} - {\ensuremath{\boldsymbol{C}}}_B {\ensuremath{\boldsymbol{G}}}_{A} \big\\|_{\text{F}}^2 + \mu \sum_{q=1}^Q \big\\| {\ensuremath{\boldsymbol{G}}}_{A,q} \big\\|_{} \label{eq:optProblem}\end{aligned}$$ where $\mu > 0$ is a regularization parameter, determining the sparsity of the solution $\hat{{\ensuremath{\boldsymbol{G}}}}_A$, and the nuclear norm of the $q$th submatrix ${\ensuremath{\boldsymbol{G}}}_{A,q}$ can be computed as the sum of its singular values $\sigma_{q,i}$, with $i=1,\ldots,r$, and $r=\min (M,L)$, i.e., $$\begin{aligned} \left\\| {\ensuremath{\boldsymbol{G}}}_{A,q} \right\\|_{} = {\textstyle \sum_{i=1}^{r}} \sigma_{q,i} .\end{aligned}$$ While the minimization of the nuclear norm terms leads to rank sparsity in the form of low-rank submatrices ${\ensuremath{\boldsymbol{G}}}_{A,q}$, the nonlinear coupling in form of the nuclear norm terms leads to a block sparse structure, i.e., the elements in submatrices ${\ensuremath{\boldsymbol{G}}}_{A,q}$ are either jointly zero or jointly nonzero. The formulation in can be considered as a one-dimensional parameter estimation problem, i.e. a block sparse representation in terms of the path delays. To recover all the channel coefficients from the matrix estimate $\hat{{\ensuremath{\boldsymbol{G}}}}_A$ in , let us first define the support set as $$\begin{aligned} \mathcal{I} = \big\{ q \big\| \hat{{\ensuremath{\boldsymbol{G}}}}_{A,q} \neq {\ensuremath{\boldsymbol{0}}} \big\}, \end{aligned}$$ i.e., the indices of the nonzero submatrices $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}$. Given the support set $\mathcal{I}$ of cardinality $\hat{P} = \|\mathcal{I}\|$, the estimated channel delays are found from the parameter grid $\{ \tau_q \}_{q=1}^Q$ as $$\begin{aligned} \{ \hat{\tau}_p \}_{p=1}^{\hat{P}} = \left\{ \tau_q \big\| q \in \mathcal{I} \right\} .\end{aligned}$$ Furthermore, let the singular value decomposition of the $q$th nonzero submatrix be given as $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q} = {\ensuremath{\boldsymbol{U}}}_q {\ensuremath{\boldsymbol{\Sigma}}}_q {\ensuremath{\boldsymbol{V}}}_q^{\text{H}}$, then the corresponding AoAs $\hat{\theta}_q$ and complex channel gain coefficients $\hat{\tilde{{\ensuremath{\boldsymbol{h}}}}}_q$ can be estimated from the principal singular vectors: in the case of a rank-one submatrix $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}$, the AoA estimation can be performed by $$\begin{aligned} \hat{\theta}_q = \textstyle \arg \max_{\theta} \; \big\| {\ensuremath{\boldsymbol{a}}}^{\text{H}}(\theta) {\ensuremath{\boldsymbol{u}}}_{q,1} \big\| \qquad \text{for } q \in \mathcal{I} \label{eq:estAoA}\end{aligned}$$ where ${\ensuremath{\boldsymbol{u}}}_{q,1}$ is the principal left singular vector of $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}$. In the case of higher rank submatrices, $r_q = \text{rank} (\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}) > 1$, there are $r_q$ propagation paths with path delays $\hat{\tau}_q = \ldots = \hat{\tau}_{q+r_q-1}$. The corresponding AoAs $\hat{\theta}_q, \ldots, \hat{\theta}_{q+r_q-1}$ are found from the $r_q$ array steering vectors ${\ensuremath{\boldsymbol{a}}}(\hat{\theta}_q), \ldots, {\ensuremath{\boldsymbol{a}}}(\hat{\theta}_{q+r_q-1})$ that best span the column subspace ${\ensuremath{\boldsymbol{U}}}_q$ of $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}$ and can be estimated by standard methods, e.g., MUSIC or sparse recovery. Similarly, for a rank-one submatrix $\hat{{\ensuremath{\boldsymbol{G}}}}_{A,q}$, the complex channel gain coefficients in $\hat{\tilde{{\ensuremath{\boldsymbol{h}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}}_{q} = [\hat{h}_q(1), \ldots, \hat{h}_q(L)]^{\text{T}}$ can be recovered from the principal singular values and corresponding principal right singular vectors as $$\begin{aligned} \hat{\tilde{{\ensuremath{\boldsymbol{h}}}}\vphantom{{\ensuremath{\boldsymbol{h}}}}}_{q} = \sigma_{q,1} {\ensuremath{\boldsymbol{v}}}^{}_{q,1} / \sqrt{ML} \qquad \text{for } q \in \mathcal{I}, \label{eq:channelGainEst}\end{aligned}$$ such that all channel parameters of interest are paired with its corresponding propagation path. Note that alternatively the reformulation in can be performed with respect to the array steering matrix ${\ensuremath{\boldsymbol{A}}}^{{\circ}}$, i.e., $$\begin{aligned} \tilde{{\ensuremath{\boldsymbol{Y}}}} =& ( {\ensuremath{\boldsymbol{B}}}^{{\circ}} {\ast}{\ensuremath{\boldsymbol{A}}}^{{\circ}} ) \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \nonumber \\ =& ( {\ensuremath{\boldsymbol{I}}}_N \otimes {\ensuremath{\boldsymbol{A}}}^{{\circ}} ) {\ensuremath{\boldsymbol{J}}} {\ensuremath{\boldsymbol{J}}}^{\text{T}}({\ensuremath{\boldsymbol{B}}}^{{\circ}} {\ast}{\ensuremath{\boldsymbol{I}}}_P ) \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}} \nonumber \\ =& {\ensuremath{\boldsymbol{C}}}_A^{{\circ}} {\ensuremath{\boldsymbol{G}}}_B^{{\circ}} + \tilde{{\ensuremath{\boldsymbol{W}}}}\end{aligned}$$ where ${\ensuremath{\boldsymbol{C}}}_A^{{\circ}} = ({\ensuremath{\boldsymbol{I}}}_N \otimes {\ensuremath{\boldsymbol{A}}}^{{\circ}}) {\ensuremath{\boldsymbol{J}}}$ and ${\ensuremath{\boldsymbol{G}}}_B^{{\circ}} = ({\ensuremath{\boldsymbol{I}}}_P {\ast}{\ensuremath{\boldsymbol{B}}}^{{\circ}}) \tilde{{\ensuremath{\boldsymbol{H}}}}\vphantom{{\ensuremath{\boldsymbol{H}}}}^{{\circ}}$, with proper permutation matrix ${\ensuremath{\boldsymbol{J}}}$, such that ${\ensuremath{\boldsymbol{J}}}^{\text{T}}({\ensuremath{\boldsymbol{B}}}^{{\circ}}~{\ast}~{\ensuremath{\boldsymbol{I}}}_P ) = {\ensuremath{\boldsymbol{I}}}_P~{\ast}~{\ensuremath{\boldsymbol{B}}}^{{\circ}}$ and ${\ensuremath{\boldsymbol{J}}} {\ensuremath{\boldsymbol{J}}}^{\text{T}}= {\ensuremath{\boldsymbol{I}}}_{NP}$. Analogous to the steps - we can form a related estimation problem to recover the AoAs $\{ \hat{\theta}_p \}_{p=1}^{\hat{P}}$ in the first step and the delays $\{ \hat{\tau}_p \}_{p=1}^{\hat{P}}$ and complex channel gain coefficients $\{ \hat{\tilde{{\ensuremath{\boldsymbol{h}}}}}_p \}_{p=1}^{\hat{P}}$ in the second step. From our experiments we found that it is advantageous to perform the sparse representation with respect to the estimation parameter which provides more measurements, i.e., number of antennas and subcarriers. This brings benefits in estimation performance and reduces computational complexity, since the computation of the nuclear norm terms in is performed for smaller submatrices. Furthermore, we remark that similar to the signal models in [@1547647]-[@sparseUnderwater], the model in can easily be extended to incorporate more parameter dimensions, such as two-dimensional angle-of-arrival, angle-of-departure or Doppler shift. The major advantage, that only discretization of one parameter space per successive estimation step is required for the estimation problem in remains unchanged by this extension. Algorithmic Implementation ========================== The nuclear norm minimization problem in can be solved in various ways, e.g., by semidefintie programming [@Boyd:RankMinimization] or the block coordinate descent method [@Steffens]. One major challenge for the evaluation of the estimation problem is the problem size, which cannot be solved with ease by the previously mentioned solvers. We propose to use the Soft-Thresholding with Exact Line search Algorithm (STELA) [@stela] since it is most suitable for solving large-scale sparse optimization problems. For proper selection of the regularization parameter $\mu$ we further propose the computation of the solution path, i.e., the set of solutions to problem for a sequence of regularization parameters $\{\mu_\kappa\}_{\kappa=1}^{\kappa_\text{max}}$, by means of the STELA method. STELA ----- Here, we will shortly describe the major steps of STELA and refer to [@stela] for details. To simplify the notation we use ${\ensuremath{\boldsymbol{G}}} = {\ensuremath{\boldsymbol{G}}}_{A,B}$ and ${\ensuremath{\boldsymbol{C}}} = {\ensuremath{\boldsymbol{C}}}_{B,A}$. The key idea of STELA is to solve a sequence of approximate problems instead of the original problem in . In iteration $t$, we approximate the problem in by $Q$ independent subproblems of the following form $$\begin{aligned} {\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q =& \arg \min_{{\ensuremath{\boldsymbol{\Gamma}}}_q} \frac{1}{2} \big\\| \tilde{{\ensuremath{\boldsymbol{Y}}}} - {\ensuremath{\boldsymbol{C}}}_{-q} {\ensuremath{\boldsymbol{G}}}^{(t)}_{-q} - {\ensuremath{\boldsymbol{C}}}_{q} {\ensuremath{\boldsymbol{\Gamma}}}_{q} \big\\|_{\text{F}}^2 \nonumber \\ & \phantom{\arg \min_{{\ensuremath{\boldsymbol{\Gamma}}}_q} } \; + \mu \textstyle{\sum_{q=1}^Q} ( \big\\| {\ensuremath{\boldsymbol{G}}}^{(t)}_{-q} \big\\|_{} + \big\\| {\ensuremath{\boldsymbol{\Gamma}}}_{q} \big\\|_{} ) , \label{eq:coordOptProb}\end{aligned}$$ for $q=1,\ldots,Q$, with ${\ensuremath{\boldsymbol{C}}}_{\!-q} = [ {\ensuremath{\boldsymbol{C}}}_{1}, \scalebox{0.9}[1]{\ldots}, {\ensuremath{\boldsymbol{C}}}_{q-1}, {\ensuremath{\boldsymbol{C}}}_{q+1}, \scalebox{0.9}[1]{\ldots}, {\ensuremath{\boldsymbol{C}}}_{Q} ]$ and ${\ensuremath{\boldsymbol{G}}}^{(t)}_{-q} = [ {\ensuremath{\boldsymbol{G}}}^{(t){\text{T}}}_{1}, \scalebox{0.9}{\ldots}, {\ensuremath{\boldsymbol{G}}}^{(t){\text{T}}}_{q-1}, {\ensuremath{\boldsymbol{G}}}^{(t){\text{T}}}_{q+1}, \scalebox{0.9}[1]{\ldots}, {\ensuremath{\boldsymbol{G}}}^{(t){\text{T}}}_{Q} ]^{\text{T}}$ denoting the dictionary matrix and an approximation of the optimal extended channel gain matrix in iteration $t$, with the $q$th submatrix removed, respectively. Note that in contrast to first or second order approximations of the objective function in , the approximation in relies on the concept of best-response, i.e., optimization for one submatrix while fixing the others. The matrix ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q$ in denotes the $q$th submatrix of the best-response matrix ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)} \!= \! [{\ensuremath{\boldsymbol{\Gamma}}}^{(t){\text{T}}}_{1}\!, \scalebox{0.7}[1]{\ldots}, {\ensuremath{\boldsymbol{\Gamma}}}^{(t){\text{T}}}_Q]^{\text{T}}$ in iteration $t$. For the case of unitary submatrices with ${\ensuremath{\boldsymbol{C}}}_{q}^{\text{H}}{\ensuremath{\boldsymbol{C}}}_{q} = {\ensuremath{\boldsymbol{I}}}$, as given in this work, it was shown in [@Steffens; @Candes:SoftThresholding] that the problem has the closed form solution $$\begin{aligned} {\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q = \mathcal{S}_{\mu} ( \bar{{\ensuremath{\boldsymbol{\Gamma}}}}\vphantom{{\ensuremath{\boldsymbol{\Gamma}}}}^{(t)}_q ) \label{eq:coordOptProb2}\end{aligned}$$ where $$\begin{aligned} \bar{{\ensuremath{\boldsymbol{\Gamma}}}}\vphantom{{\ensuremath{\boldsymbol{\Gamma}}}}^{(t)}_q = {\ensuremath{\boldsymbol{C}}}_q^{\text{H}}( \tilde{{\ensuremath{\boldsymbol{Y}}}} - {\ensuremath{\boldsymbol{C}}}_{-q} {\ensuremath{\boldsymbol{G}}}^{(t)}_{-q} )\end{aligned}$$ is the Least-Squares estimate of ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q$, with compact singular value decomposition $\bar{{\ensuremath{\boldsymbol{\Gamma}}}}\vphantom{{\ensuremath{\boldsymbol{\Gamma}}}}^{(t)}_q = \bar{{\ensuremath{\boldsymbol{U}}}}^{(t)}_q \bar{{\ensuremath{\boldsymbol{\Omega}}}}_q^{(t)} \bar{{\ensuremath{\boldsymbol{V}}}}_q^{(t) {\text{H}}}$, and $$\begin{aligned} \mathcal{S}_{\mu} ( {\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q ) = \bar{{\ensuremath{\boldsymbol{U}}}}^{(t)}_q \big( \bar{{\ensuremath{\boldsymbol{\Omega}}}}_q^{(t)} - \mu {\ensuremath{\boldsymbol{I}}} \big)_{+} \bar{{\ensuremath{\boldsymbol{V}}}}_q^{(t) {\text{H}}} \label{eq:minFun6}\end{aligned}$$ denotes the singular value thresholding operator [@Candes:SoftThresholding], with $\left[ ( {\ensuremath{\boldsymbol{X}}} )_{+} \right]_{ij} = \max (\left[ {\ensuremath{\boldsymbol{X}}} \right]_{ij}, 0)$. Returning to the original problem in , it is shown in [@stela] that ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)} - {\ensuremath{\boldsymbol{G}}}^{(t)}$ is a descent direction of the objective function in . Therefore we perform a variable update according to $$\begin{aligned} {\ensuremath{\boldsymbol{G}}}^{(t+1)} = {\ensuremath{\boldsymbol{G}}}^{(t)} + \gamma^{(t)} ( {\ensuremath{\boldsymbol{\Gamma}}}^{(t)} - {\ensuremath{\boldsymbol{G}}}^{(t)} ), \label{stelaUpdate}\end{aligned}$$ with stepsize $\gamma^{(t)}$. Convergence of the sequence $\{{\ensuremath{\boldsymbol{G}}}^{(t)}\}_{t=1}^{\infty}$ to a stationary point strongly depends on proper selection of the stepsize parameter $\gamma^{(t)}$, e.g., by successive or exact line search methods. Following the ideas in [@stela], exact line search is performed according to $$\begin{aligned} \gamma^{(t)} \!=& \operatorname{arg\,min}_{0 \leq \gamma \leq 1} \frac{1}{2} \big\\| \tilde{{\ensuremath{\boldsymbol{Y}}}} - {\ensuremath{\boldsymbol{C}}} ( {\ensuremath{\boldsymbol{G}}}^{(t)} + \gamma ({\ensuremath{\boldsymbol{\Gamma}}}^{(t)} - {\ensuremath{\boldsymbol{G}}}^{(t)} )) \big\\|_{\text{F}}^2 \nonumber \\ &\phantom{\arg \min_{0 \leq \gamma \leq 1}} + \gamma \mu \textstyle{\sum_{q=1}^Q} ( \\| {\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q \\|_{} - \\| {\ensuremath{\boldsymbol{G}}}^{(t)}_q \\|_{} ) \nonumber \\ =& \Big[ {\ensuremath{\text{Re}\big}}\{{\ensuremath{\text{Tr}\big}}( ( {\ensuremath{\boldsymbol{C}}} {\ensuremath{\boldsymbol{G}}}^{(t)} \!- \tilde{{\ensuremath{\boldsymbol{Y}}}} )^{\text{H}}{\ensuremath{\boldsymbol{C}}} ( {\ensuremath{\boldsymbol{\Gamma}}}^{(t)} \!- {\ensuremath{\boldsymbol{G}}}^{(t)} ) \big)\big\} \nonumber \\ & + \mu \textstyle{\sum_{q=1}^Q} ( \\| {\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q \\|_{} \!- \\| {\ensuremath{\boldsymbol{G}}}^{(t)}_q \\|_{} ) / \big\\| {\ensuremath{\boldsymbol{C}}} ( {\ensuremath{\boldsymbol{\Gamma}}}^{(t)} \!\!- {\ensuremath{\boldsymbol{G}}}^{(t)} ) \big\\|_{\text{F}}^2 \Big]_0^1 . \label{eq:stepSize}\end{aligned}$$ In contrast to standard exact line search approaches, which generally have to be evaluated numerically, for problem the stepsize parameter proposed in can be computed in closed-form which significantly reduces the computational complexity. Similarly, as seen from eq. -, all submatrices ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)}_q$, for $q=1,\ldots,Q$, admit closed form expressions and can be computed independently and, thus, in parallel. It can be verified that the problem in and the approximate subproblems in fulfill all assumptions specified by Theorem 2 in [@stela], such that the sequence $\{{\ensuremath{\boldsymbol{G}}}^{(t)}\}_{t=1}^{\infty}$ converges to the global optimum $\hat{{\ensuremath{\boldsymbol{G}}}}$ of problem . Furthermore, due to the exact line search, parallel updates and closed form expressions, the convergence speed of STELA is generally much faster than that of block coordinate descent methods or gradient-based methods, as we will show in the numerical results later. In summary, one iteration of the STELA method consists of computing the submatrices of the best-response matrix ${\ensuremath{\boldsymbol{\Gamma}}}^{(t)}$ according to , the stepsize $\gamma{\raisebox{4pt}[\ht\strutbox]{$\scriptstyle (t)$}}$ according to and updating the approximate solution ${\ensuremath{\boldsymbol{G}}}^{(t)}$ according to . Initialization at $t=0$ can be performed by exploiting a-priori information or in the simplest form by ${\ensuremath{\boldsymbol{G}}}^{(0)}={\ensuremath{\boldsymbol{0}}}$. Regularization Parameter Selection {#sec:regPar} ---------------------------------- A key parameter determining the estimation performance of the minimization problem in is the regularization parameter $\mu$. While direct estimation of a proper regularization parameter $\mu$ is still an open problem, in this work we follow the approach of computing the solution path to problem as proposed in [@Steffens]. The solution path is defined as the set of solutions $\{\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_\kappa)\}_{\kappa=1}^{\kappa_\text{max}}$ to for a regularization parameter sequence $\mu_1 > \mu_2 > \ldots > \mu_{\kappa_\text{max}} >0$. Similar to the LASSO problem [@Osborne99onthe], the starting point of the solution path can be computed in closed form as $$\begin{aligned} \mu_1 = \textstyle \max_q \big\\| {\ensuremath{\boldsymbol{C}}}_q^{\text{H}}\tilde{{\ensuremath{\boldsymbol{Y}}}} \big\\|_2\end{aligned}$$ which is the smallest regularization parameter giving a zero solution $\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_1)={\ensuremath{\boldsymbol{0}}}$ for problem . A small change in $\mu_\kappa$ results in a small change in the estimate $\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_{\kappa})$. This makes the STELA method particularly useful for the pathwise approach since the estimates $\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_{\kappa-1})$ can be used as an initial value ${\ensuremath{\boldsymbol{G}}}^{(0)}(\mu_{\kappa})$ in the STELA iterations, leading to rapid convergence of the STELA method. Given the solution path $\{\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_\kappa)\}_{\kappa=1}^{\kappa_\text{max}}$, we select the smallest regularization parameter $\mu_{\kappa}$ which generates a number of $P$ peaks in the spatial spectrum of the solution $\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_\kappa)$. The corresponding estimate $\hat{{\ensuremath{\boldsymbol{G}}}}(\mu_{\kappa})$ represents $P$ dominant propagation paths with the best Least-Squares data fit in . Numerical Results ================= We consider a uniform linear array of $M=4$ antennas. The OFDM reference signal consists of QPSK-symbols on $N=16$ subcarriers in the frequency domain and is sampled uniformly in time at the receiver. We assume that $L=3$ snapshots are available for parameter estimation. The complex channel gain coefficients are modeled as constant magnitude and uniform random phase. Let $(\|h_{u,p}\|, {\tau}_{p}/T_s, \theta_{p}/\text{deg.})$ denote the parameter triple defining the $p$th propagation path. We model the channel parameters as $(1.12, 0.10, 64.98)$, $(0.85, 1.23, 46.54)$, $(0.71, 1.97, 94.71)$, $(0.52, 3.57, 121.17)$ and $(0.41, 5.02, 105.32)$ such that there are $P=5$ propagation paths in total. Note that MD-ESPRIT is not directly applicable to this scenario, since it cannot resolve $P>\min(M,N)$ signals. The signal-to-noise ratio is defined as $\text{SNR} = 1/\sigma_w^2$. In a first experiment we compare the convergence speed of STELA with that of the block coordinate descent (BCD) method [@Steffens], by comparing the error $\\|\hat{{\ensuremath{\boldsymbol{G}}}}_A - {\ensuremath{\boldsymbol{G}}}^{(t)}_A\\|_{\text{F}}$, where $\hat{{\ensuremath{\boldsymbol{G}}}}_A$ denotes the solution to and ${\ensuremath{\boldsymbol{G}}}^{(t)}_A$ denotes the approximate solution in iteration $t$. The BCD performs sequential submatrix updates, such that we count one submatrix update as an iteration, while for STELA we consider one parallel update of all submatrices as one iteration. For the simulation we assume an SNR of 5dB and select the regularization parameter as $\mu = \mu_1/4$ and a uniform grid of $Q=160$ points. As can be seen in Fig. \[fig:convergence\], STELA clearly outperforms BCD in terms of convergence speed by an order of 2. For statistical evaluation we compare the proposed method to the MD MUSIC estimator [@1547647], the MD root-RARE estimator [@mdRare], the space-alternating orthogonal matching pursuit (SA-OMP) [@sparseUnderwater] and the lower Cramér-Rao bound (CRB) [@Zoltowski:Crb]. In order to achieve a sample covariance matrix of sufficiently high rank for the MUSIC and root-RARE estimator we employ smoothing and forward-backward averaging in the frequency domain. We emphasize that these data pre-processing techniques require linear sampling, whereas our proposed method can be applied to arbitrary sampling schemes. For the proposed method we discretized the delay parameter space in $Q=160$ points and compute the solution path of by means of STELA to find a proper estimate according to section \[sec:regPar\]. Note that with the proposed approach of regularization parameter selection, all evaluated methods exploit the same a-priori information in form of the number of propagation paths. To achieve higher resolution we perform a final estimation with a refined grid. The AoAs are estimated from the principal singular vectors according to equation . We perform 100 Monte Carlo runs and compute the root-mean-square errors (RMSEs) for the estimated propagation delays $\{\hat{\tau}_p\}_{p=1}^{\hat{P}}$ and AoAs $\{\hat{\theta}_p\}_{p=1}^{\hat{P}}$. As can be seen from Fig. \[fig:rmseDelay\] and \[fig:rmseDoa\], the SA-OMP scheme has the best thresholding performance among all evaluated methods. This is due to the joint estimation of delays and AoAs. However, for high SNR the RMSE reaches an estimation bias, which can be explained by the greedy nature of the algorithm [@sparseUnderwater]. Similar to the SA-OMP method, the MUSIC method jointly estimates delays and AoAs. While the asymptotic behavior for high SNR approaches the CRB, the thresholding performance is slightly decreased as compared to SA-OMP, which is due to the low number of snapshots. Our proposed sparse estimation method has thresholding performance similar to the MUSIC estimator but has a much smaller complexity, due to the successive nature of the algorithm. Furthermore it outperforms the Root-RARE estimator in thresholding performance, which performs a similar successive parameter estimation. As can be seen from Fig. \[fig:rmseDelay\] and \[fig:rmseDoa\] direct estimation of path delays and indirect estimation of the AoAs show equally good performance. Concluding Remarks ================== We have presented a new method for MIMO channel parameter estimation based on nuclear norm minimization. In contrast to existing spectrum-based methods which perform large-scale and expensive multidimensional estimation, our proposed approach successively performs one-dimensional parameter estimation. The successive one-dimensional estimation requires discretization of only one parameter space at a time, which significantly reduces the computational complexity and the sensitivity to off-grid effects, while providing proper parameter pairing. Due to space limitations we have restricted the discussion to the two-dimensional parameter estimation problem, but the proposed approach can easily be extended to incorporate, e.g., two-dimensional angle-of-arrival, angle-of-departure or Doppler shift estimation (cmp. [@mdRare]). We conclude that the aforementioned benefits of our proposed method become even more significant in higher-dimensional parameter estimation problems. Furthermore, we have presented implementation based on the recently presented STELA method, which admits for fast convergence and simple implementation. [^1]: The authors acknowledge the financial support of the Seventh Framework Programme for Research of the European Commission under grant number ADEL-619647 and the EXPRESS project within the DFG priority program CoSIP (DFG-SPP 1798).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Many real-world functions are defined over both categorical and category-specific continuous variables and thus cannot be optimized by traditional Bayesian optimization (BO) methods. To optimize such functions, we propose a new method that formulates the problem as a multi-armed bandit problem, wherein each category corresponds to an arm with its reward distribution centered around the optimum of the objective function in continuous variables. Our goal is to identify the best arm and the maximizer of the corresponding continuous function simultaneously. Our algorithm uses a Thompson sampling scheme that helps connecting both multi-arm bandit and BO in a unified framework. We extend our method to batch BO to allow parallel optimization when multiple resources are available. We theoretically analyze our method for convergence and prove sub-linear regret bounds. We perform a variety of experiments: optimization of several benchmark functions, hyper-parameter tuning of a neural network, and automatic selection of the best machine learning model along with its optimal hyper-parameters (a.k.a automated machine learning). Comparisons with other methods demonstrate the effectiveness of our proposed method.' author: - \| Dang Nguyen, Sunil Gupta, Santu Rana, Alistair Shilton, Svetha Venkatesh\ Applied Artificial Intelligence Institute (A^2^I^2^), Deakin University, Geelong, Australia\ {d.nguyen, sunil.gupta, santu.rana, alistair.shilton, svetha.venkatesh}@deakin.edu.au bibliography: - 'aaai2020.bib' title: 'Bayesian Optimization for Categorical and Category-Specific Continuous Inputs' --- Introduction\[sec:Introduction\] ================================ Related Background\[sec:Related-Work\] ====================================== The Proposed Method\[sec:Framework\] ==================================== Discussion\[sec:Discussion\] ============================ Experiments\[sec:Experiments\] ============================== Conclusion\[sec:Conclusion\] ============================ Acknowledgment {#acknowledgment .unnumbered} ============== This research was partially funded by the Australian Government through the Australian Research Council (ARC). Prof Venkatesh is the recipient of an ARC Australian Laureate Fellowship (FL170100006). Theoretical Analysis ==================== Under Assumption 1 (refer to the main paper), when using Thompson sampling (TS) for BO of a function $f_{c}(x)$ where $x\in\mathcal{X}_{c}$, the Bayesian regret after $T_{c}$ iterations is given as [@russo2014learning; @kandasamy2018parallelised] $$\begin{aligned} \text{BayesRegret}(T_{c}) & =\mathbb{E}\sum_{t_{c}=1}^{T_{c}}[f_{c}(x_{c}^{})-f_{c}(x_{t_{c}}^{c})]\nonumber \\ & \leq\mathcal{O}\left(\sqrt{T_{c}\text{log}T_{c}\gamma_{T_{c}}}\right)\label{eq:BayesRegret_definition}\end{aligned}$$ where $\gamma_{T_{c}}$ is the maximum information gain about $f_{c}(x)$ due to any $T_{c}$ function observations and the expectation is with respect to the distribution over all possible functions $f_{c}$ in our hypothesis space and any randomness in the algorithm, particularly the random sampling of TS. Let us define Bayesian simple regret* as $$\text{BayesSimpleRegret}(T_{c})\triangleq\mathbb{E}[f_{c}(x_{c}^{})-\max_{t_{c}\leq T_{c}}f_{c}(x_{t_{c}}^{c})]$$ Since $\max_{t_{c}\leq T_{c}}f_{c}(x_{t_{c}}^{c})\geq\frac{1}{T_{c}}\sum_{t_{c}=1}^{T_{c}}f_{c}(x_{t_{c}}^{c})$, from Eq. (\[eq:BayesRegret\_definition\]), we have $$\begin{aligned} \text{BayesSimpleRegret}(T_{c}) & =\mathbb{E}[f_{c}(x_{c}^{})-\max_{t_{c}\leq T_{c}}f_{c}(x_{t_{c}}^{c})]\nonumber \\ & \leq\mathcal{O}\left(\sqrt{\frac{\text{log}T_{c}\gamma_{T_{c}}}{T_{c}}}\right)\label{eq:BayesSimpleRegret_Definition}\end{aligned}$$ Proof of Lemma 1 (Upper Bound on $R_{T}^{\textrm{MAB}}$) -------------------------------------------------------- To have the proof, we use the connection between the TS and the UCB algorithms as established by [@russo2014learning]. We can write the regret $R_{T}^{\text{MAB}}$ as $$\begin{aligned} R_{T}^{\text{MAB}} & =\mathbb{E}\sum_{t=1}^{T}\left(f_{c^{}}(x^{})-f_{c_{t}}(x_{c}^{})\right)\nonumber \\ & =\mathbb{E}\sum_{t=1}^{T}(f_{c^{}}^{}-U_{t}(c^{}))+\mathbb{E}\sum_{t=1}^{T}(U_{t}(c^{})-f_{c_{t}}^{})\nonumber \\ & =\underset{\text{Term-1}}{\underbrace{\mathbb{E}\sum_{t=1}^{T}(f_{c^{}}^{}-U_{t}(c^{}))}}+\underset{\text{Term-2}}{\underbrace{\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{})}}\label{eq:two-terms}\end{aligned}$$ In the above expression, we have defined $f_{c^{}}^{}\triangleq f_{c^{}}(x^{})$ and $f_{c_{t}}^{}\triangleq f_{c_{t}}(x_{c}^{})$. Also, we have introduced an upper confidence bound function $U_{t}(c)$. The last equality follows as conditioned on $D_{t}$, the optimal arm $c^{}$ and the arm $c_{t}$ selected by TS are identically distributed and $U_{t}$ is deterministic [@russo2014learning]. We provide a detailed explanation. Note that here we are analyzing Bayesian regret (instead of the usual regret), which is defined over a set $\mathcal{F}$ of problem instances along with a probability distribution over $\mathcal{F}$. Our function $f(c,x)$ is just one of these problem instances and can be considered a random function from $\mathcal{F}$. Note that $c^{}$ being the optimal arm for $f(c,x)$ is then also a random variable. Since we do not know the function $f(c,x)$, but only have $t$ observations from it, therefore, given $D_{t}$ $c^{}$ is distributed as $p_{t}(c)=P(c=c^{}\|D_{t})$. Since Thompson sampling precisely uses this posterior distribution to propose $c_{t}$ at iteration $t$, both $c_{t}$ and $c^{}$are identically distributed conditioned on $D_{t}$. This argument is fundamental in the Bayesian regret proof of Thompson sampling and was first used by Russo and Van Roy [@russo2014learning]. Several papers since then have used this argument e.g. see [@Bubeck_Liu_2013prior], where step-1 in the proof of Theorem 1 exactly uses this argument. The uncertainty in $f_{c}^{}$ for $c$-th arm depends on the uncertainty of the Gaussian process (GP) posterior, which reduces with increasing evaluations of $f_{c}(x)$. To write an appropriate upper confidence bound for $f_{c}^{}$, we can use the TS regret analysis of BO. In Eq. (\[eq:BayesSimpleRegret\_Definition\]), plugging $\gamma_{t_{c}}\sim\mathcal{O}(t_{c}^{\alpha})$ (Assumption 2), we have $\mathbb{E}[f_{c}^{}-\max_{\{t'\|c_{t}=c\}}f(x_{t'})]\leq\mathcal{O}(\sqrt{\frac{\log t_{c}}{t_{c}^{1-\alpha}}})$. Given this inequality, we can define an upper confidence bound $U_{t}(c)$ on $f_{c}^{}$ as follows $f_{c}^{}\leq U_{t}(c)=\mathbb{E}[\max_{\{t'\|c_{t}=c\}}f(x_{t'})]+a\sqrt{\frac{\log t_{c}}{t_{c}^{1-\alpha}}})$, where $t_{c}$ is the number of times the arm $c$ has been selected in the first $t$ iterations and $a$ is an appropriate positive constant. We also define a lower confidence bound $L_{t}\left(c\right)=\mathbb{E}[\max_{\{t'\|c_{t}=c\}}f(x_{t'})]$, which holds trivially i.e. $L_{t}(c)=\mathbb{E}[\max_{\{t'\|c_{t}=c\}}f(x_{t'})]\leq f_{c}^{}$. The Term-1 of Eq. (\[eq:two-terms\]) is always negative as $f_{c^{}}^{}\leq U_{t}(c^{})$ by the definition of $U_{t}(c^{})$ and therefore can be ignored when considering an upper bound on $R_{T}^{\text{MAB}}$. We will next derive an upper bound on the Term-2 of Eq. (\[eq:two-terms\]). For this, consider $$\begin{aligned} \mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{}) & \leq\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-L_{t}(c_{t}))\\ & =a\sum_{c=1}^{C}\sum_{t_{c}=1}^{T_{c}}\sqrt{\frac{\log t_{c}}{t_{c}^{1-\alpha}}})\\ & \leq a\sqrt{\log T}\sum_{c=1}^{C}\sum_{t_{c}=1}^{T_{c}}\frac{1}{\sqrt{t_{c}^{1-\alpha}}}\end{aligned}$$ We write $\sum_{t_{c}=1}^{T_{c}}\frac{1}{\sqrt{t_{c}^{1-\alpha}}}=\sum_{t_{c}=1}^{T_{c}}\frac{t_{c}^{\alpha/2}}{\sqrt{t_{c}}}\leq T_{c}^{\alpha/2}\sum_{t_{c}=1}^{T_{c}}\frac{1}{\sqrt{t_{c}}}$ then using the identity $\sum_{t_{c}=1}^{T_{c}}\frac{1}{\sqrt{t_{c}}}\leq\int_{0}^{T_{c}}\frac{1}{\sqrt{r}}dr=2\sqrt{T_{c}}$ we obtain $$\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{})\leq2a\sqrt{\log T}\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}$$ By using the Cauchy-Schwarz inequality, we can bound $\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}\leq\sqrt{C\sum_{c=1}^{C}T_{c}^{\alpha+1}}$. Next we set $q_{c}=T_{c}^{\alpha+1}$ and $\eta=1/(\alpha+1)$ in the identity $\left(\sum_{c=1}^{C}q_{c}\right)^{\eta}\leq\sum_{c=1}^{C}q_{c}^{\eta}$, where $0<\eta\leq1$, to get $\sum_{c=1}^{C}T_{c}^{\alpha+1}\leq\left(\sum_{c=1}^{C}T_{c}\right)^{\alpha+1}$. Therefore, we can write $\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{})\leq2a\sqrt{\log T}\sqrt{C\left(\sum_{c=1}^{C}T_{c}\right)^{\alpha+1}}=2a\sqrt{CT^{\alpha+1}\log T}$ and we get $R_{T}^{\text{MAB}}\leq2a\sqrt{CT^{\alpha+1}\log T}$. Proof of Lemma 2 (Upper Bound on $R_{T}^{\textrm{BO}}$) ------------------------------------------------------- By definition, we have $R_{T}^{\text{BO}}=\mathbb{E}\sum_{t=1}^{T}[f_{c_{t}}(x_{c_{t}}^{})-f_{c_{t}}(x_{t})]=\mathbb{E}\sum_{c=1}^{C}\sum_{t_{c}=1}^{T_{c}}[f_{c}(x_{c}^{})-f_{c}(x_{t_{c}})]$. From Eq. (\[eq:BayesRegret\_definition\]), we have $\mathbb{E}\sum_{t_{c}=1}^{T_{c}}[f_{c}(x_{c}^{})-f_{c}(x_{t_{c}})]\leq\mathcal{O}\left(\sqrt{T_{c}\text{log}T_{c}\gamma_{T_{c}}}\right)\leq b\sqrt{T_{c}^{\alpha+1}\log T_{c}}$ under Assumption 2. Therefore, we have $R_{T}^{\text{BO}}\leq b\mathbb{E}\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}\log T_{c}}\leq b\sqrt{\log T}\mathbb{E}\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}$. As in the proof of Lemma 1, we can bound as $\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}\leq\sqrt{C\sum_{c=1}^{C}T_{c}^{\alpha+1}}\leq\sqrt{CT^{\alpha+1}}$ and therefore have $R_{T}^{\text{BO}}\leq b\sqrt{CT^{\alpha+1}\log T}$. Proof of Theorem 4 (Batch Setting) ---------------------------------- The Bayesian regret for the batch scheme can be written similar to the sequential case as $$\text{BayesRegret}(T)=R_{T}^{\text{MAB}}+R_{T}^{\text{BO}},$$ where $R_{T}^{\text{MAB}}\triangleq\mathbb{E}\sum_{t=1}^{T}[f_{c^{}}(x^{})-f_{c_{t}}(x_{c_{t}}^{})]$ and $R_{T}^{\text{BO}}\triangleq\mathbb{E}\sum_{t=1}^{T}[f_{c_{t}}(x_{c_{t}}^{})-f_{c_{t}}(x_{t})]$. Similar to the sequential case, for batch setting too, we have $$\begin{aligned} R_{T}^{\text{MAB}} & =\underset{\text{Term-1}}{\underbrace{\mathbb{E}\sum_{t=1}^{T}(f_{c^{}}^{}-U_{t}(c^{}))}}+\underset{\text{Term-2}}{\underbrace{\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{})}}\end{aligned}$$ Similar to Eq. (\[eq:BayesRegret\_definition\]) and (\[eq:BayesSimpleRegret\_Definition\]) of the sequential setting, we have, for batch setting, under Assumptions 1 and 2 of the main paper, TS Bayesian regret* and Bayesian simple regret (for BO of $f_{c}(x)$) bounded by $\mathcal{O}(\psi_{B}\sqrt{T_{c}^{\alpha+1}\log T_{c}})$ and $\mathcal{O}(\psi_{B}\sqrt{\frac{\log T_{c}}{T_{c}^{1-\alpha}}})$ respectively [@kandasamy2018parallelised]. Using the Bayesian simple regret we can write $\mathbb{E}[f_{c}^{}-\max_{\{t'\|c_{t}=c\}}f(x_{t'})]\leq\mathcal{O}(\psi_{B}\sqrt{\frac{\log t_{c}}{t_{c}^{1-\alpha}}})$. Given this inequality, we can define an upper confidence bound on $f_{c}^{}$ as $f_{c}^{}\leq U_{t}(c)=\mathbb{E}[\max_{\{t'\|c_{t}=c\}}f(x_{t'})]+a\psi_{B}\frac{\log t_{c}}{\sqrt{t_{c}^{1-\alpha}}}$, where $t_{c}$ is the number of times the arm $c$ has been selected in the first $t$ iterations and $a$ is an appropriate positive constant. As before we define a lower confidence bound $L_{t}\left(c\right)=\mathbb{E}[\max_{\{t'\|c_{t}=c\}}f(x_{t'})]$, which holds trivially i.e.* $L_{t}(c)\leq f_{c}^{}$. The remaining proof follows the same set of arguments as in the sequential case (see the proof of Lemma 1) and we obtain $\mathbb{E}\sum_{t=1}^{T}(U_{t}(c_{t})-f_{c_{t}}^{})\leq2a\psi_{B}\sqrt{CT^{\alpha+1}\log T}$ and since $\mathbb{E}\sum_{t=1}^{T}(f_{c^{}}^{}-U_{t}(c^{}))$ is always negative, we have, for the batch setting, $R_{T}^{\text{MAB}}\leq2a\psi_{B}\sqrt{CT^{\alpha+1}\log T}$. The analysis of finding the upper bound on $R_{T}^{\text{BO}}$ mimics the proof steps in Lemma 1 and for our batch setting, it is upper bounded as $R_{T}^{\text{BO}}\leq\mathbb{E}\sum_{c=1}^{C}b\psi_{B}\sqrt{T_{c}^{\alpha+1}\log T_{c}}\leq b\psi_{B}\sqrt{\log T}\mathbb{E}\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}$, where $b$ being an appropriate constant. As in the proof of Lemma 1, we can prove $\sum_{c=1}^{C}\sqrt{T_{c}^{\alpha+1}}\leq\sqrt{CT^{\alpha+1}}$ and therefore, $R_{T}^{\text{BO}}\leq b\psi_{B}\sqrt{CT^{\alpha+1}\log T}$. Finally combining $R_{T}^{\text{MAB}}$ and $R_{T}^{\text{BO}}$, we have $\text{BayesRegret}(T)$ for our batch setting* bounded by $2a\psi_{B}\sqrt{CT^{\alpha+1}\log T}+b\psi_{B}\sqrt{CT^{\alpha+1}\log T}$, which can be summarized as a sub-linear term in $T$ as $\mathcal{O}(\psi_{B}\sqrt{CT^{\alpha+1}\log T})$. Additional Experimental Results =============================== Experiments with Other Synthetic Functions ------------------------------------------ ### 2d function (1 categorical variable + 1 continuous variable) We compare our Bandit-BO with other methods on a 2d function as follows $$f([c,x])=\exp(-(z_{1}-2)^{2})+\exp(\frac{-(z_{1}-6)^{2}}{10})+\frac{1}{z_{2}^{2}+1}+\frac{c}{2},$$ where $z_{1}=x-0.05\times c$, $z_{2}=x+0.05\times c$, and $x\in[-2,10]$. Figures \[fig:Optimization-results-2d-function-b\] and \[fig:Optimization-results-2d-function-c\] show the optimization results for the function with different batch sizes and different numbers of categories respectively. From the results, we can see that Bandit-BO is the best method followed by SMAC and TPE. ![image](1d_C6_compare_batch_1)![image](1d_C6_compare_batch_5)![image](1d_C6_compare_batch_10) ![image](1d_C25_compare_batch_5_new)![image](1d_C50_compare_batch_5_new)![image](1d_C100_compare_batch_5_new) ![image](4d_C6_compare_batch_1)![image](4d_C6_compare_batch_5)![image](4d_C6_compare_batch_10) ![image](4d_C25_compare_batch_5_new)![image](4d_C50_compare_batch_5_new)![image](4d_C100_compare_batch_5_new) ### 5d function (1 categorical variable + 4 continuous variables) We also compare our method Bandit-BO with others methods on a 5d function created by combining a set of modified Alpine-4d functions as follows $$f([c,x])={\displaystyle \prod_{i=1}^{4}\sqrt{z_{i}}}\sin(z_{i})+(2\times c),$$ where $z_{i}=x_{i}+2\times c$ and $x\in[1,10]^{4}$. Figures \[fig:Optimization-results-5d-function-b\] and \[fig:Optimization-results-5d-function-c\] show the optimization results for the function with different batch sizes and different numbers of categories respectively. Our Bandit-BO is the best method while SMAC is the second-best method. TPE and Merchan-Lobato are generally better than One-hot-Encoding. Experiments with Automated Machine Learning ------------------------------------------- ### Machine learning models and their hyper-parameters Table \[tab:Machine-learning-models\] summarizes 14 machine learning models (classifiers) along with their hyper-parameters used in our experiment automated machine learning (see the main paper). We name the models and their hyper-parameters following the notation in the Python Machine Learning library scikit-learn[^1]. The value range for each hyper-parameter is adopted from the automated machine learning package Auto-sklearn [@Feurer2019]. Model Hyper-parameter Value range ------------------ ---------------------- ------------------------- n\_estimators $[50,100]$ learning\_rate $\log[0.01,2]$ learning\_rate $\log[0.01,1]$ subsample $[0.01,1]$ max\_features $[0.1,1]$ Decision Tree max\_depth $[0,2]$ Extra Trees max\_features $[0,1]$ n\_estimators $[10,50]$ max\_features $[0,1]$ Bernoulli NB alpha $\log[10^{-2},100]$ Multinomial NB alpha $\log[10^{-2},100]$ LDA shrinkage $[0,1]$ QDA reg\_param $[0,1]$ C $\log[2^{-5},2^{15}]$ C $\log[2^{-5},2^{15}]$ gamma $\log[2^{-15},2^{3}]$ C $\log[10^{-5},10]$ alpha $\log[10^{-7},10^{-1}]$ l1\_ratio $\log[10^{-9},1]$ eta0 $\log[10^{-7},10^{-1}]$ hidden\_layer\_sizes $\log[128,256]$ alpha $\log[10^{-7},10^{-1}]$ learning\_rate $\log[10^{-4},10^{-1}]$ ### Classification results on all 30 benchmark datasets Table \[tab:Overall-accuracy\] shows the classification accuracy of each method on 30 benchmark datasets. Compared with other methods, our method Bandit-BO shows an improved classification accuracy on most of the datasets. [\|l\|l\|r\|r\|r\|>p[0.1]{}\|>p[0.1]{}\|>p[0.1]{}\|>p[0.1]{}\|]{} Dataset & Format & $\|D\|$ & $\|F\|$ & $\|L\|$ & Bandit-BO & Hyperopt-sklearn & Auto-sklearn & TPOT[\ ]{} wine & tabular & 178 & 13 & 3 & 98.33 (0.00) & 97.78 (0.01) & 97.50 (0.01) & 96.67 (0.01)[\ ]{} breast\_cancer & tabular & 569 & 30 & 2 & 97.02 (0.01) & 95.44 (0.00) & 96.40 (0.00) & 96.84 (0.00)[\ ]{} analcatdata\_authorship & text & 841 & 70 & 4 & 99.76 (0.00) & 99.47 (0.00) & 99.41 (0.00) & 99.53 (0.00)[\ ]{} diabetes & tabular & 768 & 8 & 2 & 77.40 (0.01) & 73.70 (0.02) & 76.95 (0.01) & 77.01 (0.01)[\ ]{} electricity & tabular & 45,312 & 8 & 2 & 92.29 (0.00) & 92.21 (0.00) & 90.89 (0.00) & 90.94 (0.00)[\ ]{} wall\_robot\_navigation & trajectory & 5,456 & 24 & 4 & 99.73 (0.00) & 99.73 (0.00) & 99.43 (0.00) & 99.46 (0.00)[\ ]{} vehicle & tabular & 846 & 18 & 4 & 81.71 (0.01) & 78.71 (0.01) & 80.24 (0.01) & 78.12 (0.01)[\ ]{} cardiotocography & tabular & 2,126 & 35 & 10 & 100.0 (0.00) & 100.0 (0.00) & 99.98 (0.00) & 100.0 (0.00)[\ ]{} artificial\_characters & text & 10,218 & 7 & 10 & 90.47 (0.01) & 90.94 (0.00) & 82.49 (0.00) & 87.75 (0.01)[\ ]{} monks1 & tabular & 556 & 6 & 2 & 100.0 (0.00) & 99.82 (0.00) & 99.73 (0.00) & 100.0 (0.00)[\ ]{} monks2 & tabular & 601 & 6 & 2 & 98.26 (0.01) & 97.69 (0.01) & 97.36 (0.01) & 99.92 (0.00)[\ ]{} steel\_plates\_fault & tabular & 1,941 & 33 & 2 & 100.0 (0.00) & 100.0 (0.00) & 100.0 (0.00) & 100.0 (0.00)[\ ]{} phoneme & tabular & 5,404 & 5 & 2 & 90.23 (0.00) & 90.21 (0.00) & 89.25 (0.00) & 89.58 (0.00)[\ ]{} waveform & tabular & 5,000 & 40 & 3 & 86.45 (0.00) & 86.42 (0.00) & 86.19 (0.00) & 86.28 (0.00)[\ ]{} balance\_scale & tabular & 625 & 4 & 3 & 98.48 (0.01) & 97.20 (0.01) & 89.04 (0.01) & 92.32 (0.01)[\ ]{} digits & image & 1,797 & 64 & 10 & 98.25 (0.00) & 98.67 (0.00) & 98.08 (0.00) & 97.86 (0.00)[\ ]{} iris & tabular & 150 & 4 & 3 & 94.33 (0.01) & 92.00 (0.01) & 95.33 (0.01) & 94.67 (0.01)[\ ]{} blood\_transfusion & tabular & 748 & 4 & 2 & 77.87 (0.01) & 76.07 (0.01) & 77.67 (0.01) & 78.07 (0.01)[\ ]{} qsar\_biodeg & tabular & 1,055 & 41 & 2 & 85.12 (0.01) & 84.79 (0.01) & 86.54 (0.01) & 85.97 (0.00)[\ ]{} letter & image & 20,000 & 16 & 26 & 97.25 (0.00) & 97.00 (0.01) & 95.72 (0.00) & 95.72 (0.00)[\ ]{} australian & tabular & 690 & 14 & 2 & 83.77 (0.01) & 83.84 (0.01) & 84.13 (0.01) & 84.28 (0.01)[\ ]{} olivetti & image & 400 & 4,096 & 40 & 99.25 (0.00) & 96.25 (0.01) & 94.00 (0.01) & 97.22 (0.01)[\ ]{} spambase & text & 4,601 & 57 & 2 & 95.10 (0.00) & 95.16 (0.00) & 95.44 (0.00) & 94.80 (0.00)[\ ]{} hill\_valley & graph & 1,212 & 100 & 2 & 75.19 (0.02) & 70.82 (0.04) & 90.29 (0.01) & 82.96 (0.03)[\ ]{} eeg\_eye\_state & temporal & 14,980 & 14 & 2 & 92.47 (0.00) & 92.34 (0.00) & 95.10 (0.00) & 88.70 (0.02)[\ ]{} churn & tabular & 5,000 & 20 & 2 & 95.76 (0.00) & 95.50 (0.00) & 95.80 (0.00) & 95.65 (0.00)[\ ]{} kc1 & tabular & 2,109 & 21 & 2 & 86.00 (0.01) & 85.66 (0.01) & 85.07 (0.01) & 86.00 (0.01)[\ ]{} kc2 & tabular & 522 & 21 & 2 & 83.33 (0.01) & 83.05 (0.01) & 82.95 (0.01) & 83.43 (0.01)[\ ]{} segment & image & 2,310 & 19 & 7 & 94.00 (0.00) & 94.20 (0.00) & 94.00 (0.00) & 93.98 (0.00)[\ ]{} gas\_drift & tabular & 13,910 & 128 & 6 & 99.60 (0.00) & 99.58 (0.00) & 99.52 (0.00) & 99.44 (0.00)[\ ]{} Average & & & & & 92.25 (0.00) & 91.48 (0.01) & 91.82 (0.00) & 91.77 (0.01)[\ ]{} [^1]: <https://scikit-learn.org/>	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Hajnal–Szemerédi theorem states that for any integer $r \ge 1$ and any multiple $n$ of $r$, if $G$ is a graph on $n$ vertices and $\delta(G) \ge (1 - 1/r)n$, then $G$ can be partitioned into $n/r$ vertex-disjoint copies of the complete graph on $r$ vertices. We prove a very general analogue of this result for directed graphs: for any integer $r \ge 4$ and any sufficiently large multiple $n$ of $r$, if $G$ is a directed graph on $n$ vertices and every vertex is incident to at least $2(1 - 1/r)n - 1$ directed edges, then $G$ can be partitioned into $n/r$ vertex-disjoint subgraphs of size $r$ each of which contain every tournament on $r$ vertices. A related Turán-type result is also proven.' author: - 'Andrzej Czygrinow, Louis DeBiasio, Theodore Molla and Andrew Treglown' title: Tiling directed graphs with tournaments --- [^1] \[firstpage\] Introduction ============ Tilings in graphs ----------------- Given two (di)graphs $H$ and $G$, an $H$-tiling in $G$ is a collection of vertex-disjoint copies of $H$ in $G$. An $H$-tiling is called perfect if it covers all the vertices of $G$. Perfect $H$-tilings are also referred to as $H$-factors or perfect $H$-packings. If $H$ is connected and of order at least three, the problem of deciding whether a graph $G$ contains a perfect $H$-tiling is NP-complete [@hell]. In light of this, it is natural to ask for simple sufficient conditions which force a graph to contain a perfect $H$-tiling. A cornerstone result in extremal graph theory is the following theorem of Hajnal and Szemerédi [@hs]. \[hs\] Every graph $G$ whose order $n$ is divisible by $r$ and whose minimum degree satisfies $\delta (G) \geq (1-1/r)n$ contains a perfect $K_r$-tiling. Notice that the minimum degree condition in Theorem \[hs\] is tight. Earlier, Corrádi and Hajnal [@corradi] proved Theorem \[hs\] in the case when $r=3$. More recently, Kierstead and Kostochka [@short] gave a short proof of the Hajnal–Szemerédi theorem. Over the last three decades there has been much work on generalising the Hajnal–Szemerédi theorem. One highlight in this direction is a result of Kühn and Osthus [@kuhn; @kuhn2] that characterises, up to an additive constant, the minimum degree which ensures that a graph $G$ contains a perfect $H$-tiling for an arbitrary graph $H$. Other notable results include an Ore-type analogue of the Hajnal–Szemerédi theorem of Kierstead and Kostochka [@kier] and an $r$-partite version of the Hajnal–Szemerédi theorem of Keevash and Mycroft [@my]. See [@survey] for a survey including many of the results on graph tiling. There has also been interest in tiling problems for directed graphs and hypergraphs. A recent survey of Zhao [@zsurvey] gives an extensive overview of the latter problem. In this paper we prove a directed analogue of the Hajnal–Szemerédi theorem. Tilings in directed graphs {#sec1} -------------------------- Throughout this paper, the digraphs we consider do not have loops and we allow for at most one edge in each direction between any pair of vertices. An oriented graph is a digraph without $2$-cycles. For digraphs there is more than one natural notion of degree: The minimum semidegree $\delta ^0 (G)$ of a digraph $G$ is the minimum of its minimum outdegree $\delta ^+ (G)$ and its minimum indegree $\delta ^- (G)$. The minimum degree $\delta (G)$ of $G$ is the minimum number of edges incident to a vertex in $G$. For oriented graphs, there has been some progress on obtaining degree conditions that force tilings. Denote by $T_r$ the transitive tournament of $r$ vertices and by $C_3$ the cyclic triangle. Yuster [@yuster] observed that an oriented graph $G$ on $n\in 3\mathbb N$ vertices and with $\delta (G) \geq 5n/6$ contains a perfect $T_3$-tiling (and also gave a minimum degree condition which forces a perfect $T_r$-tiling for $r >3$). More recently, Balogh, Lo and Molla [@blm] determined the minimum semidegree threshold which ensures a perfect $T_3$-tiling in an oriented graph, thereby resolving a conjecture from [@problem]. Keevash and Sudakov [@keevs] showed that every oriented graph $G$ on $n$ vertices with $\delta ^0 (G) \geq (1/2-o(1))n$ contains a $C_3$-tiling covering all but at most $3$ vertices. (There are examples that show even $\delta ^0 (G) \geq (n-3)/2$ does not guarantee a perfect $C_3$-tiling.) Denote by $\mathcal T_r$ the set of all tournaments on $r$ vertices. Let $T \in \mathcal T_r$. For digraphs, the minimum semidegree threshold that forces a perfect $T$-tiling was characterised in [@treg]. \[tregthm\][@treg] Given an integer $r \geq 3$, there exists an $n_0 \in \mathbb N$ such that the following holds. Suppose $T \in \mathcal T_r$ and $G$ is a digraph on $n \geq n_0$ vertices where $r$ divides $n$. If $$\delta ^0 (G)\geq (1-1/r)n$$ then $G$ contains a perfect $T$-tiling. Notice that the minimum semidegree condition in Theorem \[tregthm\] is tight. Note also that Theorem \[tregthm\] implies the Hajnal–Szemerédi theorem for large graphs. An earlier result of Czygrinow, Kierstead and Molla [@ckm] gives an asymptotic version of Theorem \[tregthm\] for perfect $C_3$-tilings. It is natural to ask whether Theorem \[tregthm\] can be strengthened by replacing the minimum semidegree condition with a minimum degree condition. In particular, can one replace the minimum semidegree condition in Theorem \[tregthm\] with $\delta (G) \geq 2(1-1/r)n-1$? However, when $T=C_3$ the answer is no. Indeed, an example of Wang [@wang] shows that $\delta (G) \geq ({3n-5})/{2}$ does not ensure a perfect $C_3$-tiling. On the other hand, he showed that minimum degree $\delta (G) \geq {(3n-3)}/{2}$ does force a perfect $C_3$-tiling in a digraph $G$. This led to the following question being raised in [@treg]. \[ques1\] Let $n , r \in \mathbb N$ such that $r$ divides $n$. Let $T \in \mathcal T_r \setminus \{C_3\}$. Does every digraph $G$ on $n$ vertices with $\delta (G) \geq 2(1-1/r)n-1$ contain a perfect $T$-tiling? Czygrinow, DeBiasio, Kierstead and Molla [@cdkm] answered Question \[ques1\] in the affirmative for perfect $T_r$-tilings and also in the case when The main result of this paper gives an exact solution to a stronger version of Question \[ques1\] for all $r \geq 4$. \[mainthm\] Given an integer $r \geq 4$, there exists an $n_0 \in \mathbb N$ such that the following holds. Suppose $G$ is a digraph on $n \geq n_0$ vertices where $r$ divides $n$. If $$\delta(G)\geq 2(1-1/r)n-1,$$ then $G$ contains $n/r$ vertex-disjoint subdigraphs each of which contains every tournament on $r$ vertices. The following theorem from [@ckm], in some sense, answers the analogous question when $r=3$. \[mainthmr3\][@ckm] Suppose $G$ is a digraph on $n$ vertices where $3$ divides $n$. If $\delta(G) \ge 4n/3 -1$, then there exist $n/3$ vertex-disjoint subdigraphs each of size $3$ such that each subdigraph contains a $T_3$ and all but at most one contains $C_3$ as well. This is best possible in the following two senses: (i) there exist digraphs $G$ for which $\delta(G) = 4n/3 - 2$ and which do not contain a triangle factor of any kind and; (ii) by Wang’s example in [@wang], there exist digraphs such that $\delta(G) \ge 4n/3 - 1$ that do not have a perfect $C_3$-tiling. However, there is perhaps more to say about the case when $r=3$, as the following conjecture, which originally appeared in [@molla13phd], suggests. \[conj:2strong\][@molla13phd] Suppose $G$ is a digraph on $n$ vertices where $3$ divides $n$. If $\delta(G) \ge 4n/3 -1$ and $G$ is strongly $2$-connected, then there exist $n/3$ vertex-disjoint subdigraphs such that each of these subdigraphs contain both $T_3$ and $C_3$. It should be noted that in [@ckm] Conjecture \[conj:2strong\] was proven when $\delta(G) \ge (3n - 3)/2$. Note that when $\delta(G) \ge (3n - 3)/2$, $G$ is strongly $2$-connected. Tilings in multigraphs ---------------------- Instead of proving Theorem \[mainthm\] directly, we will prove a more general result concerning tilings in multigraphs. A similar approach was taken in [@ckm] and [@cdkm]. Suppose that $M$ is a multigraph. The minimum degree $\delta (M)$ of $M$ is the minimum number of edges incident to a vertex in $M$. For $x,y \in V(M)$ we write $\mu(xy)$ to denote the number of edges between $x$ and $y$ in $M$. We say that a loopless multigraph $M$ is standard if $\mu (xy) \leq 2$ for all $x,y \in V(M)$. Given vertices $x,y$ in a standard multigraph $M$ we say that $xy$ is a light edge if $\mu(xy)=1$ and a heavy edge if $\mu (xy)=2$. Given a digraph $G$, the underlying multigraph $M$ of $G$ is the standard multigraph obtained from $G$ by ignoring the orientations of edges. Given a standard multigraph $M$, an orientation of the edges is legal if the resulting graph $G$ is a digraph (i.e. there is at most one edge in each direction between any pair of vertices in $G$). A standard multigraph $M$ on $r$ vertices is universal if, given any legal orientation $G$ of $M$, we have that $T \subseteq G$ for every $T \in \mathcal T_r$. For example, let $M$ be a standard multigraph on $r$ vertices where $\mu(xy)=1,2$ for all distinct $x, y \in V(M)$ and the collection of light edges in $M$ forms a matching. Then $M$ is universal. On the other hand if $M$ is a standard multigraph on $n$ vertices that contains a cycle on light edges then $M$ is not universal. (There is a legal orientation of $M$ without a copy of $T_r$.) Write $\mathcal U_r$ for the set of all universal standard multigraphs on $r$ vertices. Given a collection of (multi)graphs $\mathcal X$, an $\mathcal X$-tiling in a (multi)graph $M$ is a collection of vertex-disjoint copies of elements of $\mathcal X$ in $M$. An $\mathcal X$-tiling is called perfect if it covers all the vertices of $M$. We refer to the elements of an $\mathcal X$-tiling as tiles. The next result (originally conjectured in [@cdkm]) ensures a standard multigraph of high minimum degree contains a perfect $\mathcal U_r$-tiling. \[mainthm2\] Given an integer $r \geq 4$, there exists an $n_0 \in \mathbb N$ such that the following holds. Suppose $M$ is a standard multigraph on $n \geq n_0$ vertices where $r$ divides $n$. If $$\delta (M)\geq 2(1-1/r)n-1$$ then $M$ contains a perfect $\mathcal U_r$-tiling. Notice that Theorem \[mainthm2\] implies Theorem \[mainthm\]. ![From left to right, the tightness examples $M_1, M_2, M_3, M_4$[]{data-label="fig:extremalcases"}](spbarrier1.pdf "fig:") ![From left to right, the tightness examples $M_1, M_2, M_3, M_4$[]{data-label="fig:extremalcases"}](spbarrier2.pdf "fig:") ![From left to right, the tightness examples $M_1, M_2, M_3, M_4$[]{data-label="fig:extremalcases"}](divbarrier1.pdf "fig:") ![From left to right, the tightness examples $M_1, M_2, M_3, M_4$[]{data-label="fig:extremalcases"}](divbarrier2.pdf "fig:") There are four different examples which show that the minimum degree condition in Theorem \[mainthm2\] is tight. Let $M_1$ and $M_2$ be standard multigraphs on $n$ vertices which contain all possible edges except that $M_1$ contains an independent set $U$ of size $n/r + 1$ and $M_2$ contains a set $U$ of size $2n/r + 1$ such that between any two distinct vertices $u,v \in U$ $\mu(uv) = 1$. For both $i=1,2$, $\delta(M_i) = 2(1 - 1/r)n - 2$, but $M_i$ does not contain a perfect $\mathcal U_r$-tiling. In the case of $M_1$, this is because every element in $\mathcal U_r$ has at most $1$ vertex in $U$ and in the case of $M_2$, it is because every element in $\mathcal U_r$ has at most $2$ vertices in $U$. Suppose $n/r$ is odd. We define the standard multigraph $M_3$ on $n$ vertices as follows: Take two disjoint sets $X, Y$ of size $n/r$. Inside the sets $X, Y$ place all heavy edges, and between $X$ and $Y$ place no edges. From $X\cup Y$ to the remaining vertices, place all heavy edges. Now partition the remaining $(1-2/r)n$ vertices into sets of size $n/r$ or $2n/r$. Between all such sets, place all heavy edges. Inside the sets of size $2n/r$ place all light edges and inside the sets of size $n/r$ place no edges. We have $\delta(M_3)=2(1-1/r)n-2$. If $M_3$ contained a perfect ${\mathcal{U}}_r$-tiling, each copy of ${\mathcal{U}}_r$ would intersect the sets of size $n/r$ from $V(M_3)\setminus (X\cup Y)$ in exactly one vertex and the sets of size $2n/r$ in exactly two vertices, and furthermore every copy of ${\mathcal{U}}_r$ has exactly $2$ vertices from $X$ or exactly $2$ vertices from $Y$. However, since $\|X\|$ and $\|Y\|$ are odd, $M_3$ does not contain a perfect ${\mathcal{U}}_r$-tiling. Suppose $2n/r\equiv 2\bmod{4}$. We form the standard multigraph $M_4$ on $n$ vertices similarly: Take two disjoint sets $X, Y$ of size $2n/r$. Inside the sets $X, Y$ place all heavy edges and between $X$ and $Y$, place all light edges. From $X\cup Y$ to the remaining vertices, place all heavy edges. Now partition the remaining $(1-4/r)n$ vertices into sets of size $n/r$ or $2n/r$. Between all such sets, place all heavy edges. Inside the sets of size $2n/r$ place all light edges and inside the sets of size $n/r$ place no edges. We have $\delta(M_4)=2(1-1/r)n-2$. Consider the standard multigraph $S_r$ with $r$ vertices in which there are all possible edges except that there is a vertex with precisely three light neighbours. If $r=4,5$ it is easy to check that $S_r$ is not universal (however, it is universal for $r>5$). With this in mind, suppose that $r=4,5$. Then if $M_4$ contained a perfect ${\mathcal{U}}_r$-tiling, each copy of ${\mathcal{U}}_r$ would intersect the sets of size $n/r$ from $V(M_4)\setminus (X\cup Y)$ in exactly one vertex and the sets of size $2n/r$ in exactly two vertices, and furthermore every copy of ${\mathcal{U}}_r$ has exactly $4$ vertices from $X$ or exactly $4$ vertices from $Y$. However, since $\|X\|$ and $\|Y\|$ are not divisible by $4$, $M_4$ does not contain a perfect ${\mathcal{U}}_r$-tiling. Note that if $r>5$ then $M_4$ does actually contain a perfect ${\mathcal{U}}_r$-tiling. Throughout the paper, instead of dealing with the set $\mathcal U_r$ itself, we will mainly work with three subsets of $\mathcal U_r$: $\bar{\K}_r$, $\hat{\K}_r$ and $\K'_r$. The elements of each of these three subsets are obtained from the complete standard multigraph on $r$ vertices by removing the edges from a set of vertex-disjoint (light) paths $P_1, \dotsc, P_t$ where $\|P_1\| \ge \|P_2\| \ge \dotsc \ge \|P_t\|$. The elements of $\bar{\K}_r$ are formed by removing the edges of $P_1, \dotsc, P_t$ from the complete standard multigraph on $r$ vertices where we stipulate that $\|P_i\| \le 2$ for all $i \in [t]$, i.e. the elements of $\bar{\K}_r$ are formed by removing a (light) matching from the complete standard multigraph on $r$ vertices. The elements of $\hat{\K}_r$ are formed in the same way, but we stipulate that $\|P_1\| \le 3$ and $\|P_i\| \le 2$ for all $i \ge 2$. To form elements of $\K'_r$, we stipulate that either $\|P_1\| \le 4$ and $\|P_i\| \le 2$ for all $i \ge 2$, or $\|P_1\|, \|P_2\| \le 3$ and $\|P_i\| \le 2$ for all $i \ge 3$. In the proof of Theorem \[mainthm2\] we (implicitly) produce a perfect $\mathcal U_r$-tiling where most of the tiles are elements from $\bar{\K}_r$ (in fact, we actually we produce a perfect $\K'_r$-tiling). The following example, however, demonstrates that we need the minimum degree to be greater than $2(1 - 1/r)n$ to guarantee a perfect $\bar{\K}_r$-tiling. For any $k \in \mathbb N$ and $r \ge 3$, let $G$ be a standard multigraph containing $n := 2rk$ vertices constructed in the following way. Let $\{U_1, \dotsc, U_{r-1}\}$ be a partition of $V(G)$ such that $\|U_1\| = 3k+1$, $\|U_2\| = 3k - 1$ and $\|U_i\| = 2k$ for $i \in \{3, \dotsc, r-1\}$. Place all possible light edges between $U_1$ and $U_2$, and for all other pairs of distinct sets $U_i$ and $U_j$, place all possible heavy edges. Also, inside inside both $G[U_1]$ and $G[U_2]$ place all possible heavy edges. The minimum degree of $G$ is $2(2rk - 2k) = 2(1 - 1/r)n$, which is witnessed by any $v \in U_3 \cup \dotsm \cup U_{r-1}$. Now suppose that $G$ contains $2k$ vertex-disjoint elements from $\bar{\K}_r$. Each one of these elements must have exactly $1$ vertex in each of $U_3, \dotsc, U_{r-1}$ and exactly $3$ vertices in $G[U_1 \cup U_2]$; however, $G[U_1 \cup U_2]$ does not contain a perfect $\bar{\K}_3$-tiling. Note that when $r=3$, this corresponds to Wang’s example in [@wang]. As mentioned above, in the proof of Theorem \[mainthm2\] we actually produce a perfect $\K'_r$-tiling. Notice that for any $r\geq 4$, $M_4$ does not contain a perfect $\K'_r$-tiling. So this multigraph is an extremal example for the perfect $\K'_r$-tiling problem for every $r\geq 4$. We now prove that $\bar{\K}_r\subseteq \hat{\K}_r\subseteq \K'_r\subseteq {\mathcal{U}}_r$ for $r\geq 4$. While this follows from a very strong result of Havet and Thomassé [@ht00], which states that every tournament $T$ on $n$ vertices contains every oriented path $P$ on $n$ vertices except when $P$ is the anti-directed path and $n \in \{3, 5, 7\}$, we prove it directly as we do not need the full strength of their result. For $r\geq 2$, $\bar{\K}_r\subseteq {\mathcal{U}}_r$ and for $r \ge 4$, $\bar{\K}_r\subseteq \hat{\K}_r\subseteq \K'_r\subseteq {\mathcal{U}}_r$. Let $r\geq 2$, $T$ be a tournament on $r$ vertices, and let $\vec{K}$ be a legal orientation of $K\in \bar{\K}_r$, where $\vec{K}$ has exactly $t\leq {\left\lfloorr/2\right\rfloor}$ light edges. Form a bijection from $V(T)$ to $V(\vec{K})$ by choosing $t$ independent edges in $T$ and mapping their endpoints to the light edges of $\vec{K}$ with the correct orientation. Then complete the bijection by mapping all other vertices of $T$ to $V(\vec{K})$ arbitrarily. Since all other edges of $\vec{K}$ are double edges, we have $T\subseteq \vec{K}$. Let $r\geq 4$, $T$ be a tournament on $r$ vertices, and let $\vec{K}$ be a legal orientation of $K\in \K'_r$, where the light edges of $\vec{K}$ form $t$ vertex-disjoint paths $P_1,\dots, P_t$ with $\|P_1\|\geq\dots\geq \|P_t\|$ where either $\|P_1\| \le 4$ and $\|P_i\| \le 2$ for all $i \ge 2$, or $\|P_1\|, \|P_2\| \le 3$ and $\|P_i\| \le 2$ for all $i \ge 3$. The statement follows from the following two facts which are straightforward to verify: (1) Every tournament on $4$ vertices contains every orientation of a path on $4$ vertices; (2) Every tournament on $6$ vertices contains two vertex-disjoint transitive triangles. We use this to first find an isomorphic copy of $P_1$ and $P_2$ (if applicable) in $T$, then we complete the embedding as in the first paragraph. Overview of the proof of Theorem \[mainthm2\] ============================================= As with many proofs in the area, the proof of Theorem \[mainthm2\] divides into extremal and non-extremal cases. Roughly speaking, in the extremal case we consider those standard multigraphs that are ‘close’ to the extremal examples $M_1$, $M_2$, $M_3$ and $M_4$ that were introduced after the statement of Theorem \[mainthm2\]. We deal with these extremal cases in one unified approach in Section \[extremal\]. Suppose that $G$ is as in Theorem \[mainthm2\]. Further, suppose that there is a ‘small’ set $M \subseteq V(G)$ with the property that both $G[M]$ and $G[M\cup Q]$ contain perfect $\mathcal U_r$-tilings for any ‘very small’ set $Q \subseteq V(G)$ where $\|Q\|\in r \mathbb N$. Then notice that, to find a perfect $\mathcal U_r$-tiling in $G$, it suffices to find an ‘almost’ perfect $\mathcal U_r$-tiling in $G':=G\setminus M$. Indeed, suppose that $G'$ contains a $\mathcal U_r$-tiling $\mathcal M_1$ covering all but a very small set of vertices $Q$. Then by definition of $M$, $G[M\cup Q]$ contains a perfect $\mathcal U_r$-tiling $\mathcal M_2$. Thus, $\mathcal M_1 \cup \mathcal M_2$ is a perfect $\mathcal U_r$-tiling in $G$, as desired. Roughly speaking, we refer to such a set $M$ as an ‘absorbing set’ (see Section \[secabs\] for the precise definition of such a set). The ‘absorbing method’ was first used in [@rrs2] and has subsequently been applied to numerous embedding problems in extremal graph theory. In general, a multigraph $G$ as in Theorem \[mainthm2\] may not contain an absorbing set. Indeed, consider the multigraph $G$ with disjoint vertex classes $V_1,\dotsc,V_r$ each of size $n/r$ in which there are all possible heavy edges except that each $V_i$ is an independent set. Then if $Q$ is any set of $r$ vertices in $V_1$, there is no set $M \subseteq V(G)$ such that both $G[M]$ and $G[M\cup Q]$ contain perfect $\mathcal U_r$-tilings. Note that this multigraph is close to the extremal example $M_1$. It turns out that when $G$ is non-extremal, we can always find an absorbing set $M$; we construct this set in Section \[secabs\]. Thus, to complete the proof in the non-extremal case we must find an ‘almost’ perfect $\mathcal U_r$-tiling in $G\setminus M$. Actually in Section \[sec6\] we prove a result that ensures any multigraph $G$ as in Theorem \[mainthm2\] contains a $\mathcal {\bar K}_r$-tiling covering almost all the vertices of $G$, see Corollary \[almostthm\_cor\]. However, this does not quite guarantee a large enough $\mathcal {U}_r$-tiling in $G\setminus M$. Indeed, the leftover set $Q$ obtained by applying Corollary \[almostthm\_cor\] to $G\setminus M$ will be slightly larger than the absorbing set $M$, and thus $M$ will not be able to absorb $Q$. To overcome this we again have to use the property that our multigraph $G$ is non-extremal. Using the $\mathcal {\bar K}_r$-tiling obtained from Corollary \[almostthm\_cor\] we build a significantly bigger $\mathcal {K}'_r$-tiling so that now the leftover set is very small compared to $M$. This is another delicate part of the proof and is dealt with in Section \[sec7\]. In Section \[sec6\] we will apply a Turán-type result for standard multigraphs; this is introduced in Section \[secturan\] (see Theorem \[dituran\]). We also introduce a multigraph regularity lemma in Section \[secreg\] and make use of this in Sections \[sec6\] and \[sec7\]. Notation ======== For the rest of the paper, when we write multigraph, we mean standard multigraph. Let $G$ be a multigraph. We write $e(G)$ for the total number of edges in $G$ and $e_2 (G)$ for the number of heavy edges in $G$. Given a subset $X \subseteq V(G)$, we write $G[X]$ for the submultigraph of $G$ induced by $X$. We write $G \setminus X$ for the submultigraph of $G$ induced by $V (G) \setminus X$ and define $\overline{X}:=V(G)\setminus X$. In a multigraph $G$, for $i=1,2$ let $N^i_G(v):=\{u: uv\in E(G) \text{ and } \mu(uv)=i\}$ and $d^i_G(v):=\|N^i_G(v)\|$. Let $N_G(v):=N^1_G(v)\cup N^2_G(v)$. We define the degree $d_G(v)$ of $v$ to be the sum of the multiplicities of the edges incident with $v$, i.e. $d_G(v):=d^1_G(v)+2d^2_G(v)$. Note that $d_G(v) = \|N_G(v)\| + \|N^2_G(v)\|$. Given a set $X \subseteq V(G)$ (or subgraph $X$ of $G$) we write $d_G(v,X)$ for the total number of edges in $G$ incident to $v$ whose other endpoint lies in $X$ (or $V(X)$). We define $d^2 _G(v,X)$ similarly. Given disjoint $X,Y \subseteq V(G)$ (or subgraphs $X,Y$ of $G$) we write $e_G(X,Y)$ for the total number of edges in $G$ with one endpoint in $X$ (or $V(X)$) and the other in $Y$ (or $V(Y)$); We write $e_2 (X,Y)$ for the total number of heavy edges in $G$ with one endpoint in $X$ (or $V(X)$) and the other in $Y$ (or $V(Y)$), and let $E_2(X,Y)$ denote the set of all such edges. In all the aforementioned notation we omit the subscript $G$ if the multigraph is clear from the context. When ${\mathcal{U}}$ is a family of multigraphs (digraphs) and $G$ is a multigraph (digraph) we write ${\mathcal{U}}\subseteq G$ to mean that some $U \in {\mathcal{U}}$ is a subgraph of $G$. If $U \in {\mathcal{U}}$ we say that $U$ is a copy of ${\mathcal{U}}$. If ${\mathcal{U}}$ is a family of multigraphs and $G$ is a digraph we write ${\mathcal{U}}\subseteq G$ to mean that there is a legal orientation $\vec{U}$ of some $U \in {\mathcal{U}}$ such that $\vec{U}$ is a subdigraph of $G$. Given a graph $G$ we let $G(t)$ denote the graph obtain from $G$ by replacing each vertex $x \in V(G)$ with a set $V_x$ of $t$ vertices so that, for all $x,y \in V(G)$: - if $x \neq y$ then $V_x \cap V_y = \emptyset$; - If $xy \in E(G)$ then there are all possible edges in $G(t)$ between $V_x$ and $V_y$; - If $xy \not \in E(G)$ then there are no edges in $G(t)$ between $V_x$ and $V_y$. Similarly, given a multigraph $G$ we let $G(t)$ denote the multigraph obtain from $G$ by replacing each vertex $x \in V(G)$ with a set $V_x$ of $t$ vertices so that, for all $x,y \in V(G)$: - if $x \neq y$ then $V_x \cap V_y = \emptyset$; - If $\mu(xy)=2$ in $G$ then there are all possible heavy edges in $G(t)$ between $V_x$ and $V_y$; - If $\mu(xy)=1$ in $G$ then there are all possible light edges in $G(t)$ between $V_x$ and $V_y$. - If $\mu(xy)=0$ in $G$ then there are no edges in $G(t)$ between $V_x$ and $V_y$. Given a set $X$ we write, for example $X+v$, $X-v$ and $X+v-w$ for $X \cup \{v\}$, $X\setminus \{v\}$ and $(X\setminus \{w\})\cup \{v\}$ respectively. Similarly given multigraphs $T$, $G$ where $T\subseteq G$ and $v,w \in V(G)$, we write, for example, $T+v$, $T-v$, $T+v-w$ for the multigraphs $G[V(T)+v]$, $G[V(T)-v]$ and $G[V(T)+v-w]$ respectively. We define, for example, $T-X$, $T+X$, $T-X+Y$ similarly where $X,Y \subseteq V(G)$. Throughout the paper, we write $0<\alpha \ll \beta \ll \gamma$ to mean that we can choose the constants $\alpha, \beta, \gamma$ from right to left. More precisely, there are increasing functions $f$ and $g$ such that, given $\gamma$, whenever we choose $\beta \leq f(\gamma)$ and $\alpha \leq g(\beta)$, all calculations needed in our proof are valid. Hierarchies of other lengths are defined in the obvious way. Turán-type results for universal multigraphs and digraphs {#secturan} ========================================================= In this section we determine the density threshold that ensures a standard multigraph contains a universal graph, and therefore determine the threshold that forces a digraph to contain any tournament of a given size. Let $t_{r-1}(n)$ be the maximum number of edges in an $(r-1)$-partite graph on $n$ vertices and let $T_{r-1}(n)$ be the $(r-1)$-partite graph that realises this bound. Note that when $n \ge r - 1$ $$\label{difference} t_{r-1}(n)-t_{r-1}(n-(r-1))=\frac{(r-1)(r-2)}{2}+(r-2)(n-(r-1)), $$ and for any $n$ $$t_{r-1}(n) \le \left(1-\frac{1}{r-1}\right) \frac{n^2}{2},$$ with equality when $r-1$ divides $n$. \[obs1\] Let $D_{r-1}(n)$ be the digraph obtained by replacing every edge of $T_{r-1}(n)$ with two oppositely oriented directed edges and let $M_{r-1}(n)$ be the underlying multigraph of $D_{r-1}(n)$. Then $D_{r-1}(n)$ contains no tournament on $r$ vertices and $M_{r-1}(n)$ contains no graph on $r$ vertices whose underlying graph is complete. Brown and Harary [@bh] proved that if a digraph $D$ on $n$ vertices contains more than $2t_{r-1}(n)$ edges, then $D$ contains every tournament on $r$ vertices. The following theorem strengthens their result by showing that $D$ contains subdigraph of $D$ on $r$ vertices which itself contains every tournament on $r$ vertices; in fact, we prove an even more general result about multigraphs. \[dituran\] Let $r\geq 2$ and let $G$ be a multigraph on $n$ vertices. If $e(G)>2t_{r-1}(n)$, then ${\bar{\mathcal K}}_{r} \subseteq G$. The proof proceeds by double induction on $r$ and $n$. Clearly the result holds for $r=2$. Let $r\geq 3$ and let $G$ be a standard multigraph on $n$ vertices such that ${\bar{\mathcal K}}_{r}\not\subseteq G$. If $n\leq r-1$, then $e(G)\leq 2t_{r-1}(n)$; so suppose $n\geq r$. Either $e(G)\leq 2t_{r-2}(n)\leq 2t_{r-1}(n)$, or by induction, there exists a copy of ${\bar{\mathcal K}}_{r-1}$ in $G$; let $H$ be a copy of ${\bar{\mathcal K}}_{r-1}$ with the maximum number of edges. If there exists $v\in V(G)\setminus V(H)$ such that $d(v, H)\geq 2(r-1)-1$, then we can either add $v$ to $H$ to make a copy of ${\bar{\mathcal K}}_{r}$ or we can swap $v$ with a vertex in $H$ to increase the number of edges in $H$; either way, a contradiction. So for all $v\in V(G)\setminus V(H)$ we have $d(v, H)\leq 2(r-1)-2=2(r-2)$. Thus by and induction on $n$, $$\begin{aligned} e(G)&=e(H)+e(G-H, H)+e(G-H)\\ &\leq (r-1)(r-2)+2(r-2)(n-(r-1))+2t_{r-1}(n-(r-1))\\ &=2\left(\frac{(r-1)(r-2)}{2}+(r-2)(n-(r-1))+t_{r-1}(n-(r-1))\right)\\ &=2t_{r-1}(n).\end{aligned}$$ \[cc2\] Let $G$ be a multigraph on $n$ vertices. If $\delta(G)>2(1-\frac{1}{r-1})n$ or $e(G)>(1-\frac{1}{r-1})n^2$, then ${\bar{\mathcal K}}_{r}\subseteq G$. First note that Theorem \[dituran\] and Corollary \[cc2\] immediately imply the analogous digraph versions. Observation \[obs1\] shows that the density conditions in Theorem \[dituran\] and Corollary \[cc2\] to force a copy of ${\bar{\mathcal K}}_r$ are best-possible; however, one may wonder if the same density conditions could force a multigraph $K$ whose complement contains a matching on at most ${\left\lfloorr/2\right\rfloor}-1$ light edges. The following observation shows that this is not the case. Let $K$ be a multigraph on $r$ vertices such that the complement of $K$ is a matching with at most ${\left\lfloorr/2\right\rfloor}-1$ light edges. If $r$ is even, let $n \in r \mathbb N$ and if $r$ is odd, let $n\in (r+1)\mathbb{N}$. For sufficiently large $n$, there exists a multigraph $G$ on $n$ vertices with (significantly) more than $2t_{r-1}(n)$ edges for which $K\not\subseteq G$. First suppose $r$ is even and $n=rk$. Let $G$ be an $r/2$-partite multigraph with all parts of size $2k=2n/r$. Inside each part put all possible light edges and between the parts put all possible heavy edges. We have $$e(G)=n^2-n-\frac{r}{2}\binom{2n/r}{2}=\left( 1-\frac{1}{r}-\frac{1}{2n}\right )n^2,$$ which is much larger than $( 1-\frac{1}{r-1} )n^2\geq 2t_{r-1}(n)$. Now suppose $r$ is odd and $n=(r+1)k$. Let $G$ be an $(r+1)/2$-partite multigraph with all parts of size $2k=2n/(r+1)$. Inside each part put all possible light edges and between the parts put all possible heavy edges. We have $$e(G)=n^2-n-\frac{r+1}{2}\binom{2n/(r+1)}{2}=\left (1-\frac{1}{r+1}-\frac{1}{2n}\right )n^2,$$ which is much larger than $( 1-\frac{1}{r-1} )n^2\geq 2t_{r-1}(n)$. Note that in each case $G$ contains no copy of $K$. Finally, we address the issue of the structure of ${\bar{\mathcal K}}_r$-free multigraphs with $2t_{r-1}(n)$ edges. Let $\D_2^(n)$ be the family of digraphs obtained by partitioning $n$ as $n=n_1+\dots+n_k$ such that $n_1,\dots, n_k$ are positive integers and at most one of the $n_i$s is odd, and taking $k$ disjoint copies, $D_1,\dots, D_k$ of $D_2(n_1), \dots, D_2(n_k)$, then adding all edges directed from $D_i$ to $D_j$ for all $1\leq i<j\leq k$. In particular, note that $D_2(n)\in \D_2^(n)$. Brown and Harary [@bh] proved that if $T\in \mathcal T_r$ and $D$ is a $T$-free digraph on $n$ vertices with $2t_{r-1}(n)$ edges, then $D\cong D_{r-1}(n)$ unless $T=C_3$ in which case $D\in \D_2^(n)$. The following observation shows that in our case, there is a whole family of tightness examples. Let $\M_2^(n)$ be the family of multigraphs underlying the digraphs in $\D_2^(n)$. Given $r\geq 4$, let $n\in r\mathbb{N}$ and let $\M^_{r-1}(n)$ be the family of multigraphs on $n$ vertices which can be obtained from $M_{r-1}(n)$ by the following process. Take disjoint pairs of colour classes and replace each such pair with a copy of $M\in \M^_2(2n/r)$, leaving all other edges between the sets as they were. Then every $M\in \M^_{r-1}(n)$ does not contain $\mathcal {\bar K}_r$. Let $M\in \M^_{r-1}(n)$, let $X_1,\dots, X_s$ be the colour classes from $M$ which were not modified, and let $Y_1, \dots, Y_t$ be the sets from $M$ which appeared as a result of merging two of the original colour classes. We have $r-1=s+2t$ and thus any copy of ${\bar{\mathcal K}}_r$ must contain at least $2$ vertices from some $X_i$, which is clearly not possible, or at least $3$ vertices from some $Y_j$, which would imply that $Y_j$ contains a copy of ${\bar{\mathcal K}}_3$, which is not the case. Thus ${\bar{\mathcal K}}_r \notin M$. It would be interesting to determine whether every ${\bar{\mathcal K}}_r$-free multigraph on $2t_{r-1}(n)$ edges is a member of $\M^_{r-1}(n)$, and more generally, whether every ${\bar{\mathcal K}}_r$-free multigraph on $2t_{r-1}(n)-o( n^2)$ edges is sufficiently “close” (in edit-distance) to some member of $\M^_{r-1}(n)$. A Regularity Lemma for Standard Multigraphs {#secreg} =========================================== In the proof of Theorems \[mainthm2\] and \[almostthm\_stability\] we will apply a version of Szemerédi’s regularity lemma [@reglem] for multigraphs. Before we state it we need some more definitions. The density* of a bipartite graph $G=(A,B)$ with vertex classes $A$ and $B$ is defined to be $$d_G (A,B):=\frac{e_G(A,B)}{\|A\|\|B\|}.$$ We will write $d(A,B)$ if this is unambiguous. Given any $\varepsilon >0$ we say that $G$ is [$\varepsilon$-regular]{} if for all $X\subseteq A$ and $Y \subseteq B$ with $\|X\|>\varepsilon \|A\|$ and $\|Y\|> \varepsilon \|B\|$ we have that $\|d(X,Y)-d(A,B)\|<\varepsilon$. Given disjoint vertex sets $A$ and $B$ in a graph $G$, we write $(A,B)$ for the induced bipartite subgraph of $G$ whose vertex classes are $A$ and $B$. If $G$ is a multigraph and $A, B \subseteq V(G)$ are disjoint, then we write $(A,B)^i _G$ for the bipartite graph with vertex classes $A$ and $B$ where $a \in A$ and $b \in B$ are adjacent in $(A,B)^i _{G}$ precisely if $\mu (ab)=i$ in $G$. The next well-known observation (see [@ko] for example) states that a large subgraph of a regular pair is also regular. \[slice\] Let $0< \eps < \alpha$ and $\eps ':= \max \{ \eps /\alpha , 2\eps \}$. Let $(A,B)$ be an $\eps$-regular pair of density $d$. Suppose $A' \subseteq A$ and $B' \subseteq B$ where $\|A'\|\geq \alpha \|A\|$ and $\|B'\|\geq \alpha \|B\|$. Then $(A',B')$ is an $\eps '$-regular pair with density $d'$ where $\|d'-d\|<\eps$. The following result will be applied in the proof of Theorem \[almostthm\]. It is (for example) a special case of Corollary 2.3 in [@alony]. \[red\] Let $\eps, d>0$ and $m,r,t \in \mathbb N$ such that $0<1/m \ll \eps \ll d \ll 1/r$ and $t\leq r$. Let $H$ be a graph obtained from $K_r$ by replacing every vertex of $K_r$ with $m$ vertices and replacing each edge of $K_r$ with an $\eps ^2$-regular pair of density at least $d$. Then $H$ contains a $K _t$-tiling covering all but at most $\eps m r$ vertices. We apply the following version of the regularity lemma, which is an immediate corollary of a $2$-coloured regularity lemma from [@blssw] (Theorem 2.4). This result in turn is easy to derive from the many-colour regularity lemma presented in [@ks] (Theorem 1.18). \[2colordegreeform\] For any ${\varepsilon}>0$ and $M'\in \mathbb N$, there exists $M=M({\varepsilon}, M')$ such that the following holds. Let $G$ be a standard multigraph on $n$ vertices and let $0\leq d\leq 1$. Then there exists a partition $\{V_0, V_1,\dots, V_k\}$ of $V(G)$ with $M' \le k \le M$ and a spanning subgraph $G'$ of $G$ with the following properties: 1. $\|V_0\| \le {\varepsilon}n$; 2. all clusters $V_i$, $i\in [k]$, are of the same size $\frac{1-{\varepsilon}}{M}n\leq \frac{n-\|V_0\|}{k}= \|V_1\|\le \frac{n}{M'}$; 3. $d_{G'}(v)>d_G(v)-(4d+2{\varepsilon})n$ for all $v\in V(G')$; 4. $e(G'[V_i]) = 0$ for all $i \in [k]$; 5. for all $1 \le i < j \le k$ and $c\in [2]$, the pair $(V_i,V_j)^c _{G'}$ is ${\varepsilon}$-regular with density either 0 or at least $d$. We call $V_1, \dots, V_k$ clusters, $V_0$ the exceptional set and $G'$ the pure multigraph. Given a multigraph $G$, and parameters ${\varepsilon}, d, M'$, we define the reduced multigraph $\Gamma$ as follows: Let $\{V_0,V_1,\dots, V_k\}$ be the partition and $G'$ be the subgraph of $G$ obtained from an application of Lemma \[2colordegreeform\] with parameters ${\varepsilon}, d, M'$. We let $V(\Gamma) = \{V_1,\dots, V_k\}$ and (i) if $(V_i,V_j)^2 _{G'}$ has density at least $d$ we place a heavy edge between $V_i$ and $V_j$ in $\Gamma$; (ii) if $(V_i,V_j)^2 _{G'}$ has density $0$ and $(V_i,V_j)^1 _{G'}$ has density at least $d$ we place a light edge between $V_i$ and $V_j$ in $\Gamma$; (iii) otherwise $V_i$ and $V_j$ are not adjacent in $\Gamma$. The next result implies that the minimum degree of a multigraph is almost inherited by its reduced multigraph. \[inherit\] Let ${\varepsilon}>0$, $d\in [0,1]$, $M',n \in \mathbb N$ and let $G$ be a multigraph on $n$ vertices. Let $G'$ be the pure multigraph and $\Gamma$ be the reduced multigraph obtained by applying Lemma \[2colordegreeform\] to $G$ with parameters ${\varepsilon}$, $d$ and $M'$. Then $\delta(\Gamma)\geq (\delta(G)/n-(8d+6{\varepsilon}))\|\Gamma\|$. Note that for all $V_i\in V(\Gamma)$, $$d_{\Gamma}(V_i)=d^1_{\Gamma}(V_i)+2d^2_{\Gamma}(V_i)=\|N_{\Gamma}(V_i)\|+d^2_{\Gamma}(V_i).$$ Let $v\in V_i$. Notice that $N_{G'}(v)$ intersects at least $(\|N_{G'}(v)\|-\|V_0\|)/\|V_1\|$ clusters and thus by Lemma \[2colordegreeform\], $$\|N_{\Gamma}(V_i)\|\geq (\|N_{G'}(v)\|-\|V_0\|)/\|V_1\|\geq (d^1_G(v)+d^2_G(v)-(4d+3{\varepsilon})n)/\|V_1\|.$$ Also note that $N^2_{G'}(v)$ intersects at least $(\|N^2_{G'}(v)\|-\|V_0\|)/\|V_1\|$ clusters and thus by Lemma \[2colordegreeform\], $$d^2_{\Gamma}(V_i)\geq (\|N^2_{G'}(v)\|-\|V_0\|)/\|V_1\|\geq (d^2_G(v)-(4d+3{\varepsilon})n)/\|V_1\|.$$ Altogether this gives $$\begin{aligned} d_{\Gamma}(V_i)=\|N_{\Gamma}(V_i)\|+d^2_{\Gamma}(V_i)&\geq (d^1_G(v)+d^2_G(v)-(4d+3{\varepsilon})n)/\|V_1\|+(d^2_G(v)-(4d+3{\varepsilon})n)/\|V_1\|\\ &=(d^1_G(v)+2d^2_G(v)-(8d+6{\varepsilon})n)/\|V_1\|\\ &\geq (\delta(G)/n-(8d+6{\varepsilon}))\|\Gamma\|.\end{aligned}$$ Therefore $\delta(\Gamma)\geq (\delta(G)/n-(8d+6{\varepsilon}))\|\Gamma\|$, as claimed. Almost tiling multigraphs with $\bar{\mathcal K}_r$ {#sec6} =================================================== In order to prove Theorem \[mainthm2\], we will apply (a corollary of) the following result. Roughly speaking, it states that every standard multigraph with minimum degree slightly greater than that in Theorem \[mainthm2\] contains an almost perfect $\bar{\mathcal K}_r$-tiling. \[almostthm\] Let $n,r \in \mathbb N$ where $r \geq 2$ and $\eta >0$ such that $0<1/n \ll \eta \ll 1/r$. Suppose that $G$ is a standard multigraph on $n$ vertices such that $$\delta (G) \geq 2(1-1/r+\eta )n.$$ Then $G$ contains a ${\bar{\mathcal K}}_r$-tiling covering all but at most $\eta n$ vertices. The next result is the key tool in the proof of Theorem \[almostthm\]. \[expand\] Let $\eta ,\gamma >0$ and $n,r \geq 2$ be integers such that $0 < 1/n \ll \gamma \ll \eta \ll 1/r$. Let $G$ be a standard multigraph on $n$ vertices so that $$\begin{aligned} \label{min} \delta (G) \geq 2(1-1/r+\eta)n.\end{aligned}$$ Further, suppose that the largest ${\bar{\mathcal K}}_r$-tiling in $G$ covers precisely $n' \leq (1- \eta )n$ vertices. Then there exists a $({\bar{\mathcal K}}_r \cup \bar{\mathcal K}_{r+1})$-tiling in $G$ that covers at least $n'+ \gamma n$ vertices. Certainly Theorem \[dituran\] and (\[min\]) imply that $n' \geq \eta n$. Let $\mathcal M$ denote a ${\bar{\mathcal K}}_r$-tiling in $G$ containing precisely $n'$ vertices so that the total number of edges in $\mathcal M$ is maximised. Set $n'' :=n-n'$ and $G'':=G\setminus V(\mathcal M)$. \[claimy\] There are at least $\gamma n$ vertices $x \in V(G'')$ such that $d _G (x, V(\mathcal M))\geq 2(1-1/r)n' + 2\gamma n$. Suppose for a contradiction the claim is false. Then by (\[min\]), at least $n'' -\gamma n$ vertices $y \in V(G'')$ are such that $d_{G''} (y) \geq 2(1-1/r+\eta )n -2(1-1/r)n'- 2\gamma n \geq 2(1-1/r+\eta /2)n''$. Thus by Theorem \[dituran\], $G''$ contains a copy of some $U \in {\bar{\mathcal K}}_r$. But then together with $\mathcal M$, this forms a $\bar{\mathcal K}_r$-tiling on $n'+r$ vertices in $G$, a contradiction to the maximality of $\mathcal M$. This proves the claim. Given any $x \in V(G'')$ such that $d_G (x, V(\mathcal M)) \geq 2(1-1/r)n'+ 2\gamma n$, there are at least $\gamma n$ elements $U$ in $\mathcal M$ so that $d_G(x, U)\geq 2r-1$. If $d_G(x, U)=2r$, then since $U \in {\bar{\mathcal K}}_r$, $V(U) \cup \{x\}$ spans a copy of an element of $\bar{\mathcal K}_{r+1}$ in $G$. Otherwise there is precisely one vertex $y \in V(U)$ such that $xy$ is a light edge. Suppose that for some $z \in V(U)\setminus \{y\}$, $zy$ is a light edge in $G$. Then since there are all possible edges between $x$ and $V(U)\setminus \{y\}$, $(V(U)\setminus \{y\})\cup \{x\}$ spans a copy of some $U'\in {\bar{\mathcal K}}_r$ such that $\|E(U')\|=\|E(U)\|+1$ . This is a contradiction to the choice of $\mathcal M$. Thus for all $z\in V(U)\setminus \{y\}$, $yz$ is a heavy edge. This implies that $V(U)\cup \{x\}$ spans a copy of some $U' \in \bar{\mathcal K}_{r+1}$. Claim \[claimy\] implies there are at least $\gamma n$ vertices $x \in V(G'')$ such that $d _G (x, V(\mathcal M))\geq 2(1-1/r)n' +2\gamma n$. So for at least $\gamma n$ such vertices $x$, we can pair them off with distinct elements $U$ of $\mathcal M$ so that $V(U)\cup \{x\}$ spans a copy of an element of $\bar{\mathcal K}_{r+1}$. This therefore implies that there exists a $({\bar{\mathcal K}}_r \cup \bar{\mathcal K}_{r+1})$-tiling in $G$ that covers at least $n'+ \gamma n$ vertices, as desired. The next simple observation will be used in the proof of Theorem \[almostthm\] to convert a $(\bar{\K}_r \cup \bar{\mathcal K}_{r+1})$-tiling in the reduced multigraph $\Gamma$ of $G$ into a $\bar{\K}_r$-tiling in the blow-up $\Gamma(r)$ of $\Gamma$. \[fact2\] Suppose that $r ,t\in \mathbb N$ such that $r$ divides $t$. If $U \in (\bar{\K}_r \cup \bar{\mathcal K}_{r+1})$ then $U(t)$ contains a perfect $\bar{\K}_r$-tiling. We are now ready to prove Theorem \[almostthm\]. We will repeatedly apply Lemma \[expand\] and Fact \[fact2\] to obtain an almost perfect $\bar{\K}_r$-tiling in a blow-up of the reduced multigraph of $G$. Applying Lemma \[red\] will then yield an almost perfect $\bar{\K}_r$-tiling in $G$. Arguments of a similar nature were applied in [@ko; @hlad; @triangle]. [Proof of Theorem \[almostthm\].]{} Define additional constants $\eps, d, \gamma$ and $M' \in \mathbb N$ so that $0<1/n \ll 1/M' \ll \eps \ll d \ll \gamma \ll \eta \ll 1/r$. Set $z:= \lceil 1/\gamma \rceil $. Apply Lemma \[2colordegreeform\] with parameters $\eps, d$ and $M'$ to $G$ to obtain clusters $V_1, \dots , V_k$, an exceptional set $V_0$ and a pure multigraph $G'$. Set $m:=\|V_1\|=\dots =\|V_k\|$. Let $\Gamma$ be the reduced multigraph of $G$ with parameters $\eps, d$ and $M'$. Lemma \[inherit\] implies that $$\begin{aligned} \label{ds} \delta (\Gamma) \geq 2(1-1/r+\eta /2)k.\end{aligned}$$ \[blowclaim\] $\Gamma':=\Gamma(r^z)$ contains a $\bar{\K}_r$-tiling covering at least $(1-\eta /2)kr^z=(1-\eta /2)\|\Gamma'\|$ vertices. If $\Gamma$ contains a $\bar{\K}_r$-tiling covering at least $(1-\eta /2)k$ vertices then Fact \[fact2\] implies that Claim \[blowclaim\] holds. So suppose that the largest $\bar{\K}_r$-tiling in $\Gamma$ covers precisely $\ell \leq (1- \eta /2)k$ vertices. Then by Lemma \[expand\], $\Gamma$ contains a $(\bar{\K}_r \cup \bar{\K}_{r+1})$-tiling that covers at least $\ell+ \gamma k$ vertices. Thus, by Fact \[fact2\], $\Gamma(r)$ contains a $\bar{\K}_r$-tiling covering at least $(\ell+\gamma k)r$ vertices. (So at least a $\gamma$-proportion of the vertices in $\Gamma(r)$ are covered.) Further, by definition of $\Gamma (r)$ and (\[ds\\]), $$\begin{aligned} \delta (\Gamma(r)) \geq 2(1-1/r+\eta /2)kr.\end{aligned}$$ If $\Gamma(r)$ contains a $\bar{\K}_r$-tiling covering at least $(1-\eta /2)kr$ vertices then again Fact \[fact2\] implies that the claim holds. So suppose that the largest $\bar{\K}_r$-tiling in $\Gamma(r)$ covers precisely $\ell' \leq (1- \eta /2)kr$ vertices. Recall that $\ell' \geq (\ell+\gamma k)r$. By Lemma \[expand\], $\Gamma(r)$ contains a $(\bar{\K}_r \cup \bar{\K}_{r+1})$-tiling that covers at least $\ell' +\gamma kr \geq (\ell+2\gamma k)r$ vertices. Thus, by Fact \[fact2\], $\Gamma(r^2)$ contains a $\bar{\K}_r$-tiling covering at least $(\ell+2\gamma k)r^2$ vertices. (So at least a $2\gamma$-proportion of the vertices in $\Gamma(r^2)$ are covered.) Repeating this argument at most $z$ times we see that the claim holds. For each $1 \leq i \leq k$, partition $V_i$ into classes $V^* _i, V_{i,1}, \dots , V_{i, r^z}$ where $m':=\|V_{i,j}\|= \lfloor m/r^z \rfloor \geq m/(2r^z)$ for all $1 \leq j \leq r^z$. Since $mk \geq (1-\eps)n$ by Lemma \[2colordegreeform\], $$\begin{aligned} \label{m'} m'\|\Gamma '\| = \big \lfloor {m}/{r^z} \big \rfloor kr^z \geq mk-kr^z \geq (1-2\eps)n.\end{aligned}$$ Let $c\in [2]$. Lemma \[slice\] implies that if $(V_{i_1}, V_{i_2})^{c}_{G'}$ is $\eps$-regular with density at least $d$ then $(V_{i_1, j_1}, V_{i_2,j_2})^{c}_{G'}$ is $2\eps r^z$-regular with density at least $d-\eps \geq d/2$ (for all $1\leq j_1,j_2 \leq r^z$). In particular, we can label the vertex set of $\Gamma '$ so that $V(\Gamma ')=\{V_{i,j} : 1 \leq i \leq k , \ 1 \leq j \leq r^z \}$ where, for $c \in [2]$, $\mu (V_{i_1, j_1} V_{i_2,j_2})=c$ in $\Gamma'$ implies that $(V_{i_1, j_1}, V_{i_2,j_2})^c _{G'}$ is $2\eps r^z$-regular with density at least $d/2$. By Claim \[blowclaim\], $\Gamma '$ has a $\bar{\K}_r$-tiling $\mathcal M$ that contains at least $(1-\eta /2)\|\Gamma'\|$ vertices. Consider any element $U$ in $\mathcal M$ and let $V(U)=\{ V_{i_1, j_1}, V_{i_2,j_2}, \dots , V_{i_r, j_r} \}$. Set $V'$ to be the union of $ V_{i_1, j_1}, V_{i_2,j_2}, \dots , V_{i_r, j_r}$. Note that $0<1/m' \ll 2 \eps r^z \ll d/2 \ll \gamma \ll 1/r$. Thus, Lemma \[red\] implies that $G'[V']$ contains a $\bar{\K}_r$-tiling covering all but at most $\sqrt{2 \eps r ^z} m' r \leq \gamma m' r$ vertices. (Here we are using that a heavy edge in $\Gamma '$ corresponds to a $2\eps r^z$-regular pair in $G'$ consisting only of heavy edges, and a light edge in $\Gamma '$ corresponds to a $2\eps r^z$-regular pair in $G'$ consisting only of light edges.) By considering each element in $\mathcal M$ we conclude that $G'\subseteq G$ contains a $\bar{\K}_r$-tiling covering at least $$(1-\gamma )m'r \times (1- \eta /2)\|\Gamma'\|/r \stackrel{(\ref{m'})}{\geq} (1-\gamma)(1-\eta /2) (1-2\eps )n \geq (1-\eta )n$$ vertices, as desired. The following result is a simple consequence of Theorem \[almostthm\]. \[almostthm\_cor\] Let $r\geq 2$ and $0 < 1/n \ll \eta \ll 1/r$. Suppose that $G$ is a multigraph on $n$ vertices such that $$\delta(G) \geq 2(1-1/r-\eta)n.$$ Then $G$ contains a $\bar{\K}_r$-tiling covering all but at most $4r^2\eta n$ vertices. Add $n': = {\left\lceil2 \eta n/(1/r - \eta)\right\rceil}$ vertices to $G$ which send out heavy edges to all other vertices (including each other). Call the resulting multigraph $G^$. Since $$\delta(G^)=\delta(G) + 2 n' = \delta(G) + 2(1 - 1/r + \eta)n' + 2(1/r - \eta)n' \ge 2(1 - 1/r + \eta)(n + n'),$$ we can apply Theorem \[almostthm\] to $G^$ to obtain a $\bar{\K}_r$-tiling in $G^$ covering all but at most $\eta(n+n')$ vertices. Removing all those tiles that contain vertices from $V(G^)\setminus V(G)$, we obtain a $\bar{\K}_r$-tiling in $G$ that covers all but at most $\eta(n + n') + (r-1)n' \le 4r^2 \eta n$ of the vertices of $G$, as desired. Almost perfect tilings in the non-extremal case {#sec7} =============================================== Suppose that, in the proof of Theorem \[mainthm2\] we have found a small absorbing set $M$. Ideally, we would next like to apply Corollary \[almostthm\_cor\] to conclude that $G\setminus V(M)$ contains an almost perfect $\mathcal U_r$-tiling $\mathcal M$, and then use $M$ to cover the remaining vertices, thereby obtaining a perfect $\mathcal U_r$-tiling in $G$. However, to achieve this we would require that the set of vertices uncovered by $\mathcal M$ is much smaller* than the size of the absorbing set $M$. Corollary \[almostthm\_cor\] does not guarantee this though. Indeed, this is because the size of the set of uncovered vertices in Corollary \[almostthm\_cor\] is large compared to the parameter $\eta$. Worst still, it is easy to see that the conclusion of Corollary \[almostthm\_cor\] is false if we replace $4r^2 \eta n$ with a term significantly smaller than $\eta n$. Therefore, instead we will show that the conclusion of Corollary \[almostthm\_cor\] can be strengthened in the desired way if our multigraph $G$ is far from extremal. (This strengthening will be at the cost of no longer guaranteeing an almost perfect $\bar{\K}_r$-tiling, but rather an almost perfect ${\K}'_r$-tiling.) This will ensure that we can then use the above approach in the non-extremal case (we then have to deal with the extremal case separately). To precisely describe the multigraphs that are far from extremal, we use the following definition. \[def:extremal\] Given $\gamma >0$ and $r \in \mathbb N$, we say that a multigraph $G$ on $n$ vertices is $(1/r, \gamma)$-extremal if 1. there exists $S\subseteq V(G)$ such that $\|\|S\|-n/r\|< \gamma n$ and $e(G[S])<\gamma n^2$, or 2. there exists $S\subseteq V(G)$ such that $\|\|S\|-2n/r\|<\gamma n$ and $e_2(G[S])<\gamma n^2$. The aim of this section is to prove the following result. \[almostthm\_stability\] Let $n,r \in \mathbb N$ where $r \geq 2$ and $\alpha, \eta, \gamma >0$ such that $0<1/n \ll \alpha \ll \eta \ll \gamma \ll 1/r$ and let $G$ be a multigraph on $n$ vertices. If $G$ is not $(1/r, \gamma)$-extremal and $$\delta (G) \geq 2(1-1/r-\eta )n,$$ then $G$ contains a $\K_r'$-tiling covering all but at most $\alpha n$ vertices. The proof of Theorem \[almostthm\_stability\] makes use of Corollary \[almostthm\_cor\]. The next result will be used to convert an almost perfect tiling of a multigraph with universal graphs into a perfect tiling. \[lem\_oneleftover\] Let $n,r \in \mathbb N$ where $r \geq 2$ and $\tau , \gamma '>0$ such that $0<1/n \ll \tau \ll \gamma' \ll 1/r$ and let $G$ be a multigraph on $n$ vertices. If $G$ is not $(1/r, \gamma')$-extremal, $\delta (G) \geq 2(1-1/r-\tau)n,$ and there exists a $\bar{\K}_r$-tiling covering all but one vertex, then $G$ contains a perfect $(\K_r'\cup \K_{r+1}')$-tiling in which all but at most three of the tiles are copies of $\bar{\K}_r$. Let $\T=\{T_1, \dots, T_m\}$ be the $\bar{\K}_r$-tiling in $G$ and let $v^$ be the leftover vertex. If there exists $T\in \T$ such that $d(v^, T)\geq 2r-1$, then we obtain a perfect $(\K_r'\cup \K_{r+1}')$-tiling in which all but one of the tiles are copies of $\bar{\K}_r$. So suppose that this is not the case. Then for all but at most $$\label{eq:2r-2} (2r - 2)\|\T\| - \delta(G) = 2(1 - 1/r)(n-1) - \delta(G) \le 3 \tau n$$ of the $T\in \T$, $d(v^, T)=2r-2$. If there exists $T\in \T$ such that $d(v^, T)= 2r-2$, then we could move $v^$ into $T$ to create a copy of $\K_{r+1}'$ unless: 1. there exists $u\in V(T)$ such that $\mu(v^u)=0$ in $G$ or; 2. $u_1, u_2\in V(T)$ are light neighbours of $v^$ in $G$ and either 1. \[2a\] $u_1u_2$ is a light edge or; 2. \[2b\] $u_1$ and $u_2$ are incident with distinct light edges in $T$. So we may suppose that one of ($\alpha$) and ($\beta$) holds whenever $d(v^, T)= 2r-2$. Let $$B^_i:=\{u\in T\in \T: d(v^, T)=2r-2 \text{ and } \mu(v^u)=i\}.$$ Set $B^:=B^_0\cup B^_1$. Let $\T^$ to be the set of tiles in $\T$ which contain a vertex from $B^$. Call an edge of $G$ useful if it either has both endpoints in $B_0^$, or it has one endpoint in $B_0^$ and the other in $B_1^$, or it is a heavy edge with both endpoints in $B_1^$. Given distinct $T,T' \in \mathcal T^$, we say that the ordered pair $(T,T')$ is bad* if 1. there exists $b\in B^\cap T$ such that $d(b, T') \neq 2r-2$ or failing this; 2. there exists some $b'\in B^ \cap T'$ such that $d(b', T)=2r-2$, and some $x \in T$ such that $b'x$ is not a heavy edge and $d(x, T') \neq 2r-2 $. If neither $(T,T')$ nor $(T',T)$ are bad, then we say that $\{T, T'\}$ is good. First we will show that non-extremality guarantees that there is a useful edge $bb'$ between a good pair $\{T,T'\}$. Then we will show how to use such a configuration to get the desired tiling which uses $v^$. Define an auxiliary digraph ${\mathcal{D}}$ with vertex set $\T^$ where there is an edge from $T$ to $T'$ if $(T,T')$ is bad. Note that if there exists a $b \in B^$ where $b\in T\in \T^$ and a $T'\in \T \setminus \{T\}$ such that $d(b, T')\geq 2r-1$, then we obtain our desired perfect $(\K_r'\cup \K_{r+1}')$-tiling by moving $b$ to $T'$ to create a $\hat{\K}_{r+1}$ and then moving $v^$ to $T$ to create a $\hat{\K}_r$ ($b$ was a problem vertex for $v^$; moving it out of $T$ means that we can move $v^$ in). So we may assume that such a $b$ does not exist. Therefore, for all $b$ in $B^$, by a computation similar to , we have that $d(b, T')=2r-2$ for all but at most $3 \tau n$ of the $T'\in \T$. In particular, for a fixed $T \in \mathcal T^$ there are at most $6 \tau n$ $T' \in \mathcal T^$ such that $(T,T')$ satisfies (i) in the definition of bad. Given distinct $T,T' \in \mathcal T^$, suppose there exists $b'\in B^ \cap T'$ such that $d(b', T)=2r-2$, and some $x \in T$ such that $b'x$ is not a heavy edge and $d(x, T')\geq 2r-1$. By the previous paragraph, $x \not \in B^$. So $T\cup T' \cup \{v^\}$ spans two disjoint copies of $\hat{K}_r $ and $K'_{r+1}$. (The vertex set of the former tile is $T-x+b'$, the latter $T'-b'+x+v^$.) In particular, we obtain our desired perfect $(\K_r'\cup \K_{r+1}')$-tiling. So we may assume that such an $x$ does not exist. Given distinct $T,T' \in \mathcal T^$, suppose there exists $b'\in B^* \cap T'$ such that $d(b', T)=2r-2$, and some $x \in T$ such that $b'x$ is not a heavy edge. Further, suppose there exists $T''\in \T \setminus \{T,T'\}$ such that $d(x, T'')\geq 2r-1$. We can move $x$ to $T''$ to create a $\hat{\K}_{r+1}$, move $b'$ to $T$ to create a $\hat{\K}_r$ and move $v^$ to $T'$ to create a $\hat{\K}_r$. In particular, we obtain our desired perfect $(\K_r'\cup \K_{r+1}')$-tiling. Thus, we may assume that this is not the case. Fix $T \in \mathcal T^$. Suppose there are at least $3r \tau n$ $T'\in \mathcal T^$ such that $(T,T')$ satisfies (ii) in the definition of bad. Then there exists some vertex $w\in T$ that plays the role of $x$ in (ii) for at least $3\tau n$ such $T'$. But then the previous two paragraphs imply that $$d_G (w) \leq 3\tau n (2r-3)+(\|\mathcal T\|-3\tau n)(2r-2)+2 < \delta (G),$$ a contradiction. Altogether this implies that ${\mathcal{D}}$ has maximum out-degree at most $6r \tau n$ and so $e({\mathcal{D}})\leq 6r\tau n^2$. We will now show that there are more than $6 r\tau n^2$ (unordered) pairs $\{T,T'\}$ where $T,T' \in \mathcal T^$ and so that there is a useful edge in $G$ with one endpoint in $T$ and the other in $T'$. Then for at least one such $\{T,T'\}$ we have that neither $(T,T')$ nor $(T',T)$ is a bad pair. By the non-extremality of $G$, if $\|B^_0\| \ge (1 - \gamma')\frac{n}{r}$ we have at least $\gamma' n^2\gg 6 r\tau n^2$ useful edges in $G[B^_0]$. If $\|B^\|\geq (1-\gamma')\frac{2n}{r}$, then we have at least $\gamma' n^2 \gg 6 r\tau n^2$ heavy edges in $G[B^]$, all of which are useful (and at most $n$ of these edges go between vertices in the same tile from $\mathcal T^$). Note that there are at most $4$ useful edges between any $T$ and $T'$ from $\mathcal T^$. So we can assume that both $\|B^_0\| < (1 - \gamma')\frac{n}{r}$ and $\|B^\| < (1 - \gamma')\frac{2n}{r}$. With the fact that $\|\T\| =(n-1)/r$, implies that, $$\label{eq:size_of_B^} \|B^\| + \|B^_0\| =2\|B^_0\|+\|B^_1\|=2\|\mathcal T^\| \ge 2(\|\T\| - 3\tau n) \ge \left(1 - \frac{\gamma'}{4}\right)\frac{2n}{r}.$$ So $\|B^\| < (1 - \gamma')\frac{2n}{r}$ implies that $\|B^_0\| > \gamma' n/r$, and $\|B^_0\| < (1 - \gamma')\frac{n}{r}$ implies that $\|B^\| > (1 + \gamma'/2)\frac{n}{r}$. Therefore, each of the at least $\gamma' n/r$ vertices in $B^_0$ is incident to at least $\|B^\| + \delta(G)/2 - n \ge (\gamma' n)/(4r)$ useful edges. In total we have at least $(1/2) \times(\gamma' n)/(4r) \times (\gamma' n)/r \gg 6 r\tau n^2$ useful edges in $G$ which ensures we find our desired pair $\{T,T'\}$. Now that we have a useful edge between a good pair, the next two claims show that this is sufficient to give us the desired perfect $(\K_r'\cup \K_{r+1}')$-tiling. \[2r-2switch\] Let $T, T'\in \T^$ be distinct and let $X\subseteq T$ and $X'\subseteq T'$ such that for all $x\in X$, $d(x, T')=2r-2$ and for all $x'\in X'$, $d(x', T)=2r-2$. 1. If there exist $x\in X$ and $x'\in X'$ such that $\mu(xx')=0$, then $T-x+x'$ and $T'-x'+x$ are both copies of $\bar{\K}_r$. 2. If there exist $x\in X$ and $x'\in X'$ such that $\mu(xx')=1$, then $T-x+x'$ and $T'-x'+x$ are both copies of $\hat{\K}_r$. 3. If the bipartite graph of light edges induced by $X,X'$ is $2$-regular, then $T-X+X'$ and $T'-X'+X$ are both copies of $\bar{\K}_r$. The claim follows immediately if $\mu(xx')=0$. If $\mu(xx')=1$, then each vertex $x,x'$ has one other light neighbour, each of which would create a $\hat{\K}_r$ after the switch. In the last case, all of the light neighbours of each vertex $x,x'$ are being moved to the other side. \[caseanalysis\] Suppose $T, T'\in \T^$ are distinct and there is a useful edge $bb'$ where $b \in T$, $b' \in T'$ such that $d(b,T')=2r-2$ and $d(b',T)=2r-2$. Further suppose that for all $w\in T$, if $wb'$ is not heavy, then $d(w, T')=2r-2$ and for all $w'\in T'$ if $w'b$ is not heavy, then $d(w', T)=2r-2$. Then there is a $(\K_r'\cup \K_{r+1}')$-tiling in $G$ covering precisely the vertices in $ V(T) \cup V(T')\cup \{v^\}$. To prove the claim, we split the argument into three cases.\ Case 1:* There exists $w \in T\setminus \{b\}$ or $w' \in T'\setminus \{b'\}$ such that either $\mu(bw') = 0$ or $\mu(b'w) = 0$. Without loss of generality, suppose $\mu(b'w) = 0$. Switch $b'$ and $w$. By Claim \[2r-2switch\](i), $T-w+b'$ and $T'-b'+w$ are $\bar{\K}_r$s. Since $bb'$ is a useful edge, $v^$ sends at least $2r-2$ edges to $T'-b'+w$. If $v^$ sends at least $2r-1$ edges to $T'-b'+w$ then $T'-b'+w+v^$ is a copy of $\hat{\K}_{r+1}$. If $v^$ sends precisely $2r-2$ edges to $T'-b'+w$ then $w$ must be a light neighbour of $v^$ in $G$. Further, as $\mu(b'w) = 0$ and $d(w, T')=2r-2$, we have that $w$ sends all possible edges to $T'-b'+w$, i.e. $d(w, T'-b'+w)=2r-2$. In particular, $(\alpha)$ and $(\beta)$ do not hold (where $T'-b'+w$ is playing the role of $T$). Thus, $T'-b'+w+v^$ is a copy of ${\K}'_{r+1}$. Case 2: $b\in B_0^$ or $b'\in B_0^$.\ Without loss of generality, suppose $b\in B_0^$. Since we are not in the first case, $b$ has two light neighbours in $T'$. In particular, there exists $x' \in V(T') \setminus B_0^$ that is a light neighbour of $b$. Switch $b$ and $x'$. By Claim \[2r-2switch\](ii), $T-b+x'$ and $T'-x'+b$ are copies of $\hat{\K}_r$. In particular, $T-b+x'$ has the property that if $T-b+x'$ contains a light path on $3$ vertices, then $x'$ is an endpoint of this path. Furthermore since $x'\notin B_0^$ and $\mu(v^b)=0$, $v^$ sends at least $2r-1$ edges to $T-b+x'$. Moreover, if $d(v^,T-b+x')=2r-1$, then $\mu(v^x')=1$ and thus $T-b+x'+v^$ is a copy of $\K_{r+1}'$. ![Case 3: Note that the light edges in $T$ and $T'$ (not shown) form a matching.[]{data-label="fig:case3"}](case3i.pdf "fig:") ![Case 3: Note that the light edges in $T$ and $T'$ (not shown) form a matching.[]{data-label="fig:case3"}](case3ii.pdf "fig:") ![Case 3: Note that the light edges in $T$ and $T'$ (not shown) form a matching.[]{data-label="fig:case3"}](case3iii.pdf "fig:") Case 3: $b,b'\in B_1^$.\ Let $B_1^\cap T=\{b_1, b_2\}$ and $B_1^\cap T'=\{b_1', b_2'\}$ with $b_1 = b$ and $b_1' = b'$. By the definition of a useful edge, $\mu(b_1b_1') = 2$. Since we are not in Case 1, there exists $x'\in V(T')\setminus B_1^$ that is a light neighbour of $b_1$ and there exists $x \in V(T) \setminus B_1^$ that is a light neighbour of $b_1'$ (see Figure \[fig:case3\](i)). Since $T' - x' + b_1 \in \hat{\K}_r$, we may assume that $T - b_1 + x' + v^ \notin \K'_{r+1}$ – this implies that the other light neighbour of $x'$ in $T$ must be $b_2$ and that $b_2$ must have a light neighbour in $T - b_1$. Similarly, we can assume $T' - b_1' + x + v^* \notin \K'_{r+1}$, so $xb_2'$ is a light edge and $b_2'$ has a (unique) light neighbour in $T' - b_1'$. Therefore, both $b_1b_2$ and $b_1'b_2'$ are heavy edges (see Figure \[fig:case3\](ii)). Suppose $b_1b_2'$ is not a light edge, so $b_1$ has a light neighbour $x_2' \in V(T - x') \setminus B_1^$. As in the previous case, since $T' - x_2' + b_1 \in \hat{\K}_r$ we may assume that $T - b_1 + x_2' + v^ \notin \K'_{r+1}$. So it must be that $x_2'b_2$ is a light edge. Now $b_1x'b_2x_2'$ forms a $4$-cycle of light edges and by Claim \[2r-2switch\](iii), we can switch $b_1,b_2$ for $x',x_2'$ and then add $v^$ to $T - b_1 - b_2 + x' + x_2'$ to obtain disjoint copies of $\bar{\K}_r$ and $\bar{\K}_{r+1}$. Likewise, we would be done if $b_1'b_2$ is not a light edge. So suppose both $b_1b'_2$ and $b'_1b_2$ are light edges (see Figure \[fig:case3\](iii)). Then $b_1 b'_2 x b_1'b_2x'$ forms a $6$-cycle of light edges; we simultaneously switch $b_1, b_2, x$ for $b_1',b_2',x'$ to obtain two disjoint copies of $\bar{\K}_r$. Recall $b_1b_2$ is a heavy edge so they both have distinct light neighbours in $T$. Hence, at most one of $b_1$ and $b_2$ is a light neighbour of $x$. Therefore we can add $v^$ to $T' - b'_1 - b'_2 - x' + b_1 + b_2 + x$ to form an element of $\K_{r+1}'$. This completes the proof of the claim and thus the lemma. We now combine Corollary \[almostthm\_cor\] and Lemma \[lem\_oneleftover\] to obtain the following result. \[almostthm\_reducedstability\] Let $n,r \in \mathbb N$ where $r \geq 2$ and $\eta, \gamma >0$ such that $0<1/n \ll \eta \ll \gamma \ll 1/r$ and let $G$ be a multigraph on $n$ vertices. If $G$ is not $(1/r,\gamma)$-extremal and $$\delta (G) \geq 2(1-1/r-\eta )n,$$ then $G$ contains a perfect $(\K_r'\cup \K_{r+1}')$-tiling. Choose $\tau, \gamma '$ so that $\eta \ll \tau\ll \gamma' \ll \gamma\ll 1/r$. By Corollary \[almostthm\_cor\] there exists a $\bar{\K}_r$-tiling $\mathcal T$ covering all but at most $4 r^2 \eta n$ vertices. Set $U :=V(G) \setminus V(\mathcal T)$. To construct a perfect $(K_r'\cup K_{r+1}')$-tiling in $G$, we perform the following iterative procedure. For each vertex $v^$ uncovered by $\T$ we apply Lemma \[lem\_oneleftover\] once. In each iteration, we modify at most three elements of $\T$. Each time we apply Lemma \[lem\_oneleftover\], the multigraph under consideration is the subgraph of $G$ induced by $V(\T')\cup \{v^\}$ where $\T' \subseteq \T$ is the set of tiles in $\T$ that have not been modified in any of the previous steps. Suppose we have performed this procedure for every vertex in some $U' \subseteq U$. Let $G'$ be the subgraph under consideration and note that $\|G'\| \ge n - 3r\|U'\| - (\|U\| - 1)$, so $$\delta(G') \ge 2\left(1-1/r-\eta\right)n - (n - \|G'\|) \geq 2\left(1-1/r-\tau\right)\|G'\|.$$ Furthermore, if $S \subseteq V(G')$ and $\|\|S\| - \|G'\|/r\| \le \gamma' \|G'\|$ or $\|\|S\| - 2\|G'\|/r\| \le \gamma' \|G'\|$, then $\|\|S\| - n/r\| \le \gamma n$ or $\|\|S\| - 2n/r\| \le \gamma n$, respectively. Therefore, $G'$ is not $(1/r, \gamma')$-extremal, because $\gamma n^2 \ge \gamma' \|G'\|^2$. Hence, we may apply Lemma \[lem\_oneleftover\] a total of $\|U\|$ times to complete the proof. We now apply the regularity lemma together with Proposition \[almostthm\_reducedstability\] to prove Theorem \[almostthm\_stability\]. [Proof of Theorem \[almostthm\_stability\].]{} Define additional constants $\eps, d$ and $M' \in \mathbb N$ so that $1/n \ll 1/M' \ll \eps \ll d \ll \alpha$. Apply Lemma \[2colordegreeform\] with parameters $\eps, d$ and $M'$ to $G$ to obtain clusters $V_1, \dots , V_k$, an exceptional set $V_0$ and a pure multigraph $G'$. Set $m:=\|V_1\|=\dots =\|V_k\|$. Let $\Gamma$ be the reduced multigraph of $G$ with parameters $\eps, d$ and $M'$. Lemma \[inherit\] implies that $$\begin{aligned} \delta (\Gamma) \geq 2(1-1/r-2\eta )k.\end{aligned}$$ Suppose that there exists $S \subseteq V(\Gamma)$ such that $\|\|S\|-k/r\|<\gamma k/4$ and $e(\Gamma [S])<\gamma k^2 /4$. Let $S =\{V_{i_1} , \dots , V_{i_t}\}$ and $S':=V_{i_1} \cup \dots \cup V_{i_t}$. Then $\|\|S'\|-km/r\|<\gamma km/4$ and so by Lemma \[2colordegreeform\](ii), $\|\|S'\|-n/r\|<\gamma n$. Moreover, by Lemma \[2colordegreeform\](iv) and the definition of the reduced multigraph $\Gamma$, $e(G'[S'])< (\gamma k^2/4)\cdot2m^2 \leq \gamma n^2/2$. Thus, by Lemma \[2colordegreeform\](iii), $e(G[S'])< \gamma n^2$, a contradiction as $G$ is not $(1/r,\gamma)$-extremal. A similar argument shows that there is no set $S \subseteq V(\Gamma)$ such that $\|\|S\|-2k/r\|<\gamma k/4$ and $e_2(\Gamma [S])<\gamma k^2 /4$. Thus, $\Gamma$ is not $(1/r,\gamma/4)$-extremal. Therefore, by Proposition \[almostthm\_reducedstability\], $\Gamma$ contains a perfect $(\K_r'\cup \K_{r+1}')$-tiling $\mathcal T$. Note that every induced subgraph of a copy of $\mathcal K'_{r+1}$ of size $r$ is itself a copy of $\mathcal K'_{r}$. Since $\eps \ll d \ll \alpha$, by repeatedly applying Lemma \[red\] for each of the tiles in $\mathcal T$ we obtain a $\K_r'$-tiling in $G$ covering all but at most $\alpha n$ vertices, as required. The Absorbing Lemma {#secabs} =================== Let $G$ be a multigraph and $\mathcal H$ be a collection of multigraphs. We call a set $M \subseteq V(G)$ an $\mathcal H$-absorbing set for $W \subseteq V(G)$ if both $G[M]$ and $G[M\cup W]$ contain perfect $\mathcal H$-tilings. Suppose that a non-extremal multigraph $G$ as in Theorem \[mainthm2\] contains a small set $M \subseteq V(G)$ that is a $\mathcal K'_r$-absorbing set for any very small set $W \subseteq V(G)$. Theorem \[almostthm\_stability\] ensures that $G\setminus M$ contains an almost perfect $\mathcal K'_r$-tiling; let $W$ denote the set of uncovered vertices in $G\setminus M$. Then $G[M\cup W]$ contains a perfect $\mathcal K'_r$-tiling, and thus $G$ contains a perfect $\mathcal K'_r$-tiling, as required. We will show that such an absorbing set $M$ exists if $G$ is non-extremal and if additionally, in the case when $r=4$, $G$ is not ‘splittable’. We use the following definition to make this precise. \[def:splittable\] Given $\gamma >0$ and $r =2,4$, a multigraph $G$ on $n$ vertices is $(1/r,\gamma)$-splittable if there exist disjoint sets $U_1,U_2 \subseteq V(G)$ such that - $\|U_1\|, \|U_2\| \ge (1/2 - \gamma)n$ and; - if $r=2$ then $e(U_1, U_2) \le \gamma n^2$; if $r=4$ then $e_2(U_1, U_2) \le \gamma n^2$. The next result gives a condition which forces a multigraph to contain an absorbing set. \[lo\] Let $h,t \in \mathbb N$ and let $\gamma >0$. Suppose that $\mathcal H$ is a collection of multigraphs, each on $h$ vertices. Then there exists an $n_0 \in \mathbb N$ such that the following holds. Suppose that $G$ is a multigraph on $n \geq n_0$ vertices so that, for any $x,y \in V(G)$, there are at least $\gamma n^{th-1}$ $(th-1)$-sets $X \subseteq V(G)$ such that both $G[X \cup \{x\}]$ and $G[X \cup \{y\}]$ contain perfect $\mathcal H$-tilings. Then $V(G)$ contains a set $M$ so that - $\|M\|\leq (\gamma/2)^h n/4$; - $M$ is an $\mathcal H$-absorbing set for any $W \subseteq V(G) \setminus M$ such that $\|W\| \in h \mathbb N$ and $\|W\|\leq (\gamma /2)^{2h} hn/32 $. Lo and Markström [@lo] proved Lemma \[lo\] for hypergraphs, however, the proof of this result for multigraphs is identical. The next result together with Lemma \[lo\] implies our multigraph $G$ contains an absorbing set. \[main-absorb\] Let $r \ge 3$, $0 < 1/n \ll \eta , \phi \ll \gamma \ll 1/r$, and let $G$ be a multigraph on $n$ vertices. If $\delta(G) \ge 2(1 - 1/r - \eta)n$ and $G$ is not $(1/r, \gamma)$-extremal and either $r \neq 4$ or $G$ is not $(1/r, \gamma)$-splittable, then for all distinct $x_1,x_2 \in V(G)$ there exist at least $(\phi n)^{r-1}$ $(r-1)$-sets $Y\subseteq V(G)$ such that $G[Y \cup \{ x_1\}]$ and $G[Y \cup \{ x_2\}]$ both contain $\K_r'$. We will need the following lemma in the proof of Lemma \[main-absorb\]. \[prop:heavy\_triangle\_count\] Let $0 < 1/n \ll \eta \ll \lambda \ll \gamma \ll 1$, and let $G$ be a multigraph on $n$ vertices with $\delta(G) \ge (3/2 - 2\eta)n$ which is not $(1/4, \gamma)$-extremal and not $(1/4, \gamma)$-splittable. For any disjoint sets $U_1, U_2 \subseteq V(G)$ such that $\|U_1\|, \|U_2\| \ge (1/2 - \gamma/5)n $, there exists a collection $\T$ of copies of $\bar{\K}_3$ in $G[U_1 \cup U_2]$ such that $\|\T\| \ge \lambda n^3$ and for every $T \in \mathcal{T}$, both $V(T) \cap U_1$ and $V(T) \cap U_2$ are non-empty, and if $T$ contains a light edge $u_1u_2$, then $u_1\in U_1$ and $u_2\in U_2$. For $T\in \bar{\K}_3$ such that $T\subseteq G[U_1 \cup U_2]$, we say that $T$ is nice if both $V(T) \cap U_1$ and $V(T) \cap U_2$ are non-empty, and either $T$ has no light edges or $T$ has exactly one light edge $u_1u_2$ with $u_1\in U_1$ and $u_2\in U_2$. Since $G$ is not $(1/4, \gamma)$-splittable, there exist at least $\gamma n^2$ heavy edges in $G$ with one endpoint in $U_1$ and one endpoint in $U_2$. For each such edge $u_1u_2$, we will either find (i) at least $\gamma n/2$ vertices $u$ such that $u_1u_2u$ is a nice $\bar{\K}_3$, or (ii) at least $\gamma n^2$ edges $e$ such that $u_ie$ is a nice $\bar{\K}_3$ for some $i \in [2]$. A simple calculation then implies that we obtain our desired collection of nice $\bar{\K}_3$. Let $u_1u_2$ be an edge in $E_2(U_1, U_2)$ such that $u_i \in U_i$ for $i \in [2]$. Pick $i$ so that $d^2(u_i) \ge d^2(u_{3-i})$ and note that $$(3/2 - 2\eta)n \le \|N^2(u_{3-i})\| + \|N(u_{3-i})\| \le \|N^2(u_i)\| + \|N(u_{3-i})\|,$$ so $\|N^2(u_i) \cap N(u_{3-i})\| \ge n/2 - 2 \eta n$. Since for every $u \in N^2(u_i) \cap N(u_{3-i}) \cap U_i$, $u_1u_2u$ is a nice $\bar{\K}_3$, if $\|N^2(u_i) \cap N(u_{3-i}) \cap U_i\| \ge \gamma n/2$ we are done. Otherwise, $$\begin{aligned} \|N^2(u_i) \cap U_{3-i}\| &\ge \|N^2(u_i) \cap N(u_{3-i}) \cap (U_1 \cup U_2)\| - \|N^2(u_i) \cap N(u_{3-i}) \cap U_i\| \\ &\ge \left[n/2 - 2 \eta n + 2(1/2 - \gamma/5)n - n \right] - \gamma n/2 \ge (1/2 - \gamma)n. \end{aligned}$$ Therefore, since $G$ is not $(1/4, \gamma)$-extremal, there are at least $\gamma n^2$ heavy edges in $G[N^2(u_i) \cap U_{3-i}]$ and for each such edge $e$, $u_ie$ is a nice $\bar{\K}_3$. We are now ready to prove Lemma \[main-absorb\]. [Proof of Lemma \[main-absorb\].]{} Define $\lambda$ so that $\eta,\phi\ll \lambda \ll \gamma$. Fix distinct vertices $x_1$, $x_2\in V(G)$ and let $X := \{x_1, x_2\}$. For any $U \subseteq V(G)$ (with $0 \le \|U\|\leq r$) and integer $i \ge 0$, let $$S_i(U) := \{v \in V(G)\setminus U : d(v, U) \ge 2\|U\| - i\}.$$ Note that when $U = \emptyset$, we trivially have $S_i(U) = V(G)$. By the degree condition, $$(2r- 2)\|U\|n/r - 2\|U\| \eta n-\|U\|(\|U\|-1) \le e_G(U, V(G)\setminus U) \le \|S_0(U)\| + \|S_1(U)\| + (2\|U\| - 2)n.$$ Therefore, $$\label{eq:U_0_plus_U_1} \|S_0(U)\| + \|S_1(U)\| \ge (2r - 2\|U\|)n/r - 3\|U\|\eta n.$$ So since $\|S_1(U)\| \le n$ and $S_0(U) \subseteq S_1(U)$, $$\begin{aligned} \label{eq:U_0} \|S_0(U)\| &\ge (r - 2\|U\|)n/r - 3\|U\| \eta n\text{ and } \\ \label{eq:U_1} \|S_1(U)\| &\ge (r - \|U\|)n/r - 3\|U\|\eta n/2. \end{aligned}$$ Call an $(r-1)$-set $Y\subseteq V(G)$ good if both $G[Y + x_1]$ and $G[Y + x_2]$ contain $\K'_r$. For $t\geq 0$ and $0\leq l\leq {\left\lfloort/2\right\rfloor}$, we say that a $t$-set $Y \subseteq S_0(X)$ is $l$-acceptable if $G[Y]$ has exactly $l$ light edges and either: - $t = 0$, - $t > 0$ and $G[Y] \in \bar{\K}_t$, or - $t = r - 3$ and $G[Y] \in \hat{\K}_t$. If $Y$ is $l$-acceptable for some $0\leq l\leq {\left\lfloort/2\right\rfloor}$, then we say that $Y$ is acceptable; note that if $Y$ is an acceptable $t$-set, then both $G[Y + x_1]$ and $G[Y + x_2]$ contain $\hat{\K}_{t+1}$. For any acceptable $t$-set $Y$, let $S^_1(Y)$ be the set of vertices $v \in S_1(Y) \cap S_0(X)$ such that if $y$ is the unique light neighbour of $v$ in $Y$, then $y$ is incident to a light edge in $Y$. Note that if $Y$ is an $l$-acceptable $t$-set, $v \in S^_1(Y)$ and $y$ is unique light neighbour of $v$ in $Y$, then $Y' := Y - y + v$, is either $(l-1)$-acceptable or $(l-2)$-acceptable depending on whether $y$ is incident to one or two light edges in $Y$. Recall that $y$ can only be incident to two light edges in $Y$ if $t = r- 3$ and $y$ is the middle vertex of a path on three vertices in $Y$ that consists of light edges. For $0\leq t\leq r-3$ and $0\leq l\leq {\left\lfloort/2\right\rfloor}$, say that an $l$-acceptable $t$-set $Y$ is $\lambda$-extendible if at least one of the following four conditions holds: 1. $t = 0$, $r$ is even and $\|S_0(X)\| \ge \lambda n$; 2. $t = 0$, $r=4$, and there are at least $(\lambda n)^3$ $3$-sets $Z$ such that $Z$ is a good $3$-set; 3. $(r-1) - t$ is even, and there are at least $(\lambda n)^2$ $2$-sets $Z$ such that $Y \cup Z$ is either an acceptable $(t+2)$-set or a good $(r-1)$-set; or 4. $\|S_1^(Y)\| \ge \lambda n$. If we assume that (iv) never holds, then Claim \[claim:extendible\] below will imply that there exists at least $(\phi n)^{r-1}$ good $(r-1)$-sets. Indeed, in this case - if $r$ is odd, then we can build good $(r-1)$-sets two vertices at a time by repeatedly using (iii); - if $r$ is even and (i) holds, then we can construct good $(r-1)$-sets by first selecting any of the $\lambda n$ vertices in $S_0(X)$ and then finish the construction by repeatedly applying (iii); and - if $r$ is even and (i) does not hold, then (ii) must hold which immediately implies that there are $(\lambda n)^{r-1}$ good $(r-1)$-sets. \[claim:extendible\] Let $0\leq t\leq r-3$ and $0\leq l\leq {\left\lfloort/2\right\rfloor}$, and $Y$ be an $l$-acceptable $t$-set. If $t = 0$ or $(r-1) - t$ is even, then $Y$ is $\lambda$-extendible. To prove the claim, we may assume that (iv) does not hold throughout, i.e. $$\label{eq:not_v} \|S_1^(Y)\| < \lambda n.$$ Note that, by , $$\label{eq:size_of_S_0_X} \|S_0(X)\| \ge (r-4)n/r - 6\eta n.$$ First assume that $(r-1) - t$ is even and let $$\begin{aligned} U := S_0(X) \cap S_0(Y) = S_0(X \cup Y). \end{aligned}$$ Also define $$\begin{aligned} U' &:= \left(S_0(X) \cap S_1(Y)\right) \setminus S_1^(Y) && \text{when $t \le r - 5$, or} \\ U' &:= S_1(X \cup Y) \setminus S_1^(Y) && \text{when $t = r - 3$.} \end{aligned}$$ When $t \le r - 5$, , and imply that $$\label{eq:U_plus_Uprime} \begin{split} \|U\| + \|U'\| &\ge (\|S_0(Y)\| + \|S_0(X)\| - n) + (\|S_1(Y)\| - \|S^_1(Y)\| + \|S_0(X)\| - n) \\ &\ge 2(r - t - 4)n/r - 2 \lambda n. \end{split}$$ Assume that $t \le r - 7$. If $\|U\| \ge \gamma n/2$, then first pick any $z \in U$, and then pick any vertex $z' \in N^2(z) \cap U'$ and note that $Y + z + z'$ is an acceptable $(t+2)$-set. By , the minimum degree condition and the fact that $U \subseteq U'$, we have that there are at least $$3n/r - \lambda n - (2/r + 2\eta) n \ge n/r - 2\lambda n$$ choices for $z'$. Note that there are at least $(\gamma n/2)\times (n/r-2\lambda n)\times (1/2) >(\lambda n)^2$ choices for $\{z,z'\}$, so (iii) in the definition of $\lambda$-extendible holds, as required. If $\|U\| < \gamma n/2$, then pick any $z \in U' \setminus U$ and let $y \in Y$ be the unique light neighbour of $z$ in $Y$. Recall that since $z \notin S^_1(Y)$, $y$ has no light neighbours in $Y$. By , $\|U' \setminus U\| \ge 6n/r - 3 \lambda n$ and for any of the at least $2n/r - 2 \gamma n$ vertices $z' \in N^2(z) \cap N^2(y) \cap U'$, $Y + z + z'$ is an acceptable $(t+2)$-set. Note that there are at least $(6n/r - 3\lambda n)\times (2n/r - 2 \gamma n)\times (1/2) >(\lambda n)^2$ choices for $\{z,z'\}$, so (iii) in the definition of $\lambda$-extendible holds, as required. When $t = r - 5$, implies that $$\|U\| + \|U'\| \ge 2n/r - 2\lambda n,$$ and when $t = r - 3$, and give that $$\|U\| + \|U'\| \ge \|S_0(X\cup Y)\| + \|S_1(X\cup Y)\| - \|S_1^(Y)\| \ge 2n/r - 3(r-1)\eta n - \lambda n.$$ Therefore, when $t \in \{r-3, r-5\}$, $$\label{eq:U_plus_Uprime_t_large} \|U\| + \|U'\| \ge 2n/r - \gamma n / 3.$$ We will either find at least $\lambda n^2$ light edges $zz'$ where $z \in U$ and $z' \in U'$, or at least $\lambda n^2$ heavy edges in $G[U']$. Note that, in either of these two cases, when $\|Y\| = r-5$, $G[Y + z + z' + x_i] \in \hat{\K}_{r-2}$ for $i \in [2]$, and when $\|Y\| = r-3$, $G[Y + z + z' + x_i] \in {\K}'_{r}$ for $i \in [2]$, so this will prove the claim. Suppose that we cannot find at least $\lambda n^2$ such edges. By non-extremality, this implies that $\|U\| < (1/r - \gamma)n$ and $\|U'\| < (2/r - \gamma)n$. Hence, by , $\|U\| \ge \frac{2\gamma n}{3}$ and $\|U'\| \ge (1/r + \frac{2}{3}\gamma)n$. Therefore, by the degree condition, any vertex in $z \in U$, is adjacent to at least ${\gamma n}/{2}$ vertices $z' \in U'$, a contradiction. We will now show that if $r$ is even and $t = 0$ and (i) in the definition of $\lambda$-extendible does not hold, then (ii) must hold. So assume $t = 0$, $r$ is even and $\|S_0(X)\| < \lambda n$. This, with , implies that $r = 4$ so $G$ is not $(1/r, \gamma)$-splittable. By the degree condition, $\|S_0(X)\| < \lambda n$ implies that $$2n/r - 2\eta n \le \|N^2(x_i)\| \le 2n/r + 2\lambda n \text{ and } \|N(x_i) \|\ge (1 - 3\lambda)n \text{ for $i \in [2]$}.$$ Therefore, if we let $$U_i := (N^2(x_i) \cap N(x_{3-i})) \setminus N^2(x_{3-i}) \text{ for $i \in [2]$},$$ then $U_1$ and $U_2$ are disjoint, and $$\|U_i\| \ge (\|N^2(x_i)\| + \|N(x_{3-i})\| - n) - \|S_0(X)\| \ge (2/r - 5\lambda)n \text{ for $i \in [2]$}.$$ By Lemma \[prop:heavy\_triangle\_count\], there are at least $\lambda n^3$ nice copies of $\bar{\K}_3$ in $G[U_1\cup U_2]$ (i.e., copies of $\bar{\K}_3$ that intersect both $U_1$ and $U_2$ and that have at most one light edge and such a light edge has one endpoint in $U_1$ and the other in $U_2$). The vertex set of any such nice $\bar{\K}_3$ is a good $(r-1)$-set, as desired. This completes the proof of the claim. Assume for a contradiction that there are at most $(\phi n)^{r-1}$ good $(r-1)$-sets. Let $0 \le t \le r-3$ be the maximum $t$ for which we have $(\phi n)^t$ acceptable $t$-sets and either $t = 0$ or $(r-1) - t$ is even. Such a $t$ exists, since $\emptyset$ is an acceptable $0$-set. We can also assume that $t > 0$, since if $Y = \emptyset$, then $S^_1(Y) = \emptyset$ and (iv) cannot hold, so Claim \[claim:extendible\] implies that one of (i), (ii) or (iii) must hold, which violates the maximality of $t$ or the assumption that there are at most $(\phi n)^{r-1}$ good $(r-1)$-sets. Let $l$ be minimal such that if $\mathcal{Y}$ is the set of $l$-acceptable $t$-sets, then $\|\mathcal{Y}\| \ge \lambda(\lambda/16)^{{\left\lfloort/2\right\rfloor}-l}(\phi n)^t$. There exists such an $l$, because $\lambda \ll 1/r$. By Claim \[claim:extendible\], each set $Y\in \mathcal{Y}$ is $\lambda$-extendible in one of two ways, so in particular there is some subset of $\mathcal{Y}$ of order at least $\|\mathcal{Y}\|/2$ for which all elements are extendible in the same way, either (iii), or (iv). If the elements in this subset are all $\lambda$-extendible by (iii), then, because $\phi \ll \lambda \ll 1/r$, we have at least $$\frac{(\lambda n)^{2} \cdot \|\mathcal{Y}\|/2}{\binom{t+2}{2}} \ge \frac{\lambda^{2}/2 \cdot \lambda(\lambda/16)^{{\left\lfloort/2\right\rfloor}}}{\binom{t+2}{2}} \cdot \phi^{-2} \cdot (\phi n)^{t+2} > (\phi n)^{t+2}$$ acceptable $(t+2)$-sets or, if $t=r-3$, more than $(\phi n)^{r-1}$ good $(r-1)$-sets. This contradicts the maximality of $t$ or the assumption that there are at most $(\phi n)^{r-1}$ good $(r-1)$-sets. (We divide by $\binom{t+2}{2}$ in the above calculation to account for the fact there are $\binom{t+2}{2}$ different ways a $(t+2)$-set can be constructed by adding $2$ vertices to a $t$-set.) Now assume that there are at least $\|\mathcal{Y}\|/2$ $l$-acceptable $t$-sets that are extendible by (iv). Let $\mathcal{Y}'$ be the collection of $(l-1)$-acceptable and $(l-2)$-acceptable $t$-sets and set $$\mathcal{Z} := \{ Y + z : Y \in \mathcal{Y} \text{ and } z \in S_1^(Y) \}.$$ Our aim is to find a lower bound on $\|\mathcal{Y}'\|$ which contradicts the minimality of $l$. Let $Z \in \mathcal{Z}$ and let $Y \in \mathcal{Y}$ and $z \in S_1^(Y)$ such that $Y + z = Z$. Note that the choice of $Y$ is not necessarily unique. Let $z'$ be the unique light neighbour of $z$ in $Y$ and note that $Y' = Y - z' + z$ is in $\mathcal{Y}'$. Therefore, for every $Z \in \mathcal{Z}$ there exists $Y' \in \mathcal{Y}'$ and $z' \in V(G)$, such that $Z = Y' + z'$ which implies that $\|\mathcal{Y'}\| \cdot n \ge \|\mathcal{Z}\|$. Any $Z \in \mathcal{Z}$ is constructed by adding to some $Y \in \hat{\K}_t$ a vertex $z \in S_1^(Y)$. So in any $Z \in \mathcal{Z}$ there are most four vertices $z' \in Z$ such that there exists $z'' \in Z$ such that $z'z''$ is a light edge and $z''$ has at least two light neighbours in $Z$. Therefore, for every $Z \in\mathcal{Z}$, there are at most four different pairs $(Y, z)$ such that $Y \in \mathcal{Y}$, $z \in S_1^(Y)$ and $Z = Y + z$. This implies that $$\|\mathcal{Z}\| \ge \frac{\|\mathcal{Y}\|/2 \cdot \lambda n}{4} \ge (\lambda/8) \cdot \lambda (\lambda/16)^{{\left\lfloort/2\right\rfloor}-l}(\phi n)^t \cdot n,$$ so $\|\mathcal{Y}'\| \ge (\lambda/8) \cdot \lambda (\lambda/16)^{{\left\lfloort/2\right\rfloor}-l}(\phi n)^t$. Therefore, there are at least $$\lambda(\lambda/16)^{{\left\lfloort/2\right\rfloor}-(l-1)}(\phi n)^t$$ $(l-1)$-acceptable $t$-sets or $(l-2)$-acceptable $t$-sets, a contradiction to the minimality of $l$. The stability result ==================== We now combine Lemmas \[lo\] and \[main-absorb\] together with Theorem \[almostthm\_stability\] to prove the following result which ensures that Theorem \[mainthm2\] holds in the case when $G$ is non-extremal and additionally if $r=4$, non-splittable. \[mainthm-stability\] Let $n,r \in \mathbb N$ where $r $ divides $n$ and define $\eta, \gamma>0$ such that $0<1/n \ll \eta \ll \gamma \ll 1/r$. Let $G$ be a multigraph on $n$ vertices. If $\delta(G) \ge 2(1 - 1/r - \eta)n$ and $G$ is not $(1/r, \gamma)$-extremal and either $r \notin \{2, 4\}$ or $G$ is not $(1/r, \gamma)$-splittable, then $G$ contains a perfect $\K_r'$-tiling. First assume that $r \ge 3$. Define $\alpha, \phi, \eta', \gamma' >0$ so that $0<1/n \ll \alpha \ll \phi \ll \eta \ll \eta' \ll \gamma' \ll \gamma$. Let $G$ be as in the statement of the theorem. By Lemma \[main-absorb\], given any $x_1,x_2 \in V(G)$, there exist at least $(\phi n)^{r-1}$ $(r-1)$-sets $Y \subseteq V(G)$ such that both $G[Y \cup \{x_1\}]$ and $G[Y\cup \{x_2\}]$ contain elements of $\K'_r$. Thus, by Lemma \[lo\], $V(G)$ contains a set $M$ so that - $\|M\| \leq ((\phi)^{r-1}/2)^r n/4$; - $M$ is a $\K'_r$-absorbing set for any $W \subseteq V(G) \setminus M$ such that $\|W\| \in r \mathbb N$ and $\|W\| \leq ((\phi)^{r-1}/2)^{2r}rn/32$. Let $G':=G\setminus V(M)$ and $n':=\|G'\|$. So as $\phi \ll \eta \ll \eta'$, $$\delta(G') \geq 2\left(1 - 1/r - \eta' \right)n'.$$ Further, as $\phi \ll \gamma' \ll \gamma$, $G'$ is not $(1/r, \gamma')$-extremal. Thus, by Theorem \[almostthm\_stability\], $G'$ contains a $\K'_r$-tiling covering all but at most $\alpha n' < ((\phi)^{r-1}/2)^{2r} rn/ 32$ vertices. Let $W$ denote the set of these uncovered vertices. Then by definition of $M$, $G[M\cup W]$ contains a perfect $\K'_r$-tiling. Altogether this implies that $G$ contains a perfect $\K'_r$-tiling, as desired. Now assume that $r \le 2$. If $r = 1$ the theorem is trivial. For $r = 2$, we show that there exists a perfect matching in $H$, the graph underlying the multigraph $G$, i.e. $H$ is the graph on $V(G)$ in which there is an edge between $x$ and $y$ if and only if there is either a light or heavy edge between $x$ and $y$ in $G$. We have that $n$ is even, $\eta \ll \gamma \ll 1/2$, $\delta(H) \ge (1/2 - \eta)n$, $G$ is not $(1/2, \gamma)$-extremal and $G$ is not $(1/2, \gamma)$-splittable. Let $M$ be a maximum matching in $H$ and suppose $M$ is not perfect. Note that the vertices unsaturated by $M$ form an independent set. Let $w_1, w_2$ be two vertices unsaturated by $M$ and for $i\in [2]$, define $S_i:=\{v: uv\in M \text{ and } u\in N(w_i)\}$. Note that $\|S_1\|, \|S_2\|\geq (1/2 - \eta)n$ and there are no edges in $H$ with one endpoint in $S_1$ and the other in $S_2$ as this would give us an $M$-augmenting path, contradicting the maximality of $M$. Since $G$ is not $(1/2, \gamma)$-splittable, it cannot be the case that $S_1$ and $S_2$ are disjoint, so let $v\in S_1\cap S_2$. We have $N(v)\cap (S_1\cup S_2)=\emptyset$ and thus $\|S_1\cup S_2\|\leq (1/2+\eta)n$ which implies $\|S_1\cap S_2\|\geq (1/2-3\eta)n\geq (1/2-\gamma)n$, contradicting the fact that $G$ is not $(1/2, \gamma)$-extremal. The extremal case {#extremal} ================= In this section we prove the following theorem. \[thm:main\_extremal\] For any $r \in {\mathbb{N}}$, there exists $n_0\in \mathbb N$ such that the following holds. If $G$ is a multigraph on $n \ge n_0$ vertices, $n$ is divisible by $r$ and $$\label{eq:precise} \delta(G) \ge 2(1 - 1/r)n - 1,$$ then $G$ contains a perfect $\mathcal{K}'_r$-tiling. Note that Theorem \[thm:main\_extremal\] immediately implies Theorem \[mainthm2\] (and thus Theorem \[mainthm\]). Our results from the previous sections will ensure Theorem \[thm:main\_extremal\] holds in the ‘non-extremal’ cases. Therefore most of the work in this section concerns the extremal cases. Throughout this section we consider a standard multigraph $G$ on $n$ vertices that satisfies the hypothesis of Theorem \[thm:main\_extremal\]. In particular, we may assume $1/n\ll 1/r$. We denote the vertex set of $G$ by $V$. Preliminary claims ------------------ We will use the following well-known and simple lemma in this section. A proof is included for completeness. \[lem:matching\] For any graph $H$ there is a matching of order at least $\min\{{\left\lfloor\|H\|/2\right\rfloor}, \delta(H)\}$. Let $M$ be a maximal matching in $H$ and suppose that $\|M\| < \min\left\{{\left\lfloor\|H\|/2\right\rfloor}, \delta(H)\right\}$. Let $U$ be the vertices incident to an edge in $M$ and let $W := \overline{U}$. Note that $\|W\| \ge 2$ and that, by the maximality of $\|M\|$, $W$ is an independent set. Therefore, there exist distinct $x,y \in W$ where $d(x, U) + d(y, U) \ge 2 \delta(H) > 2\|M\|$. Hence, there exists $e \in M$ such that $d(x, e) + d(y, e) \ge 3$, and this implies that there exists a matching of order $\|M\| + 1$ in $H$. The next claim gives us a minimum degree condition for $G[U]$ where $U$ is any set of size close to $sn/r$ for some $s\in \mathbb N$. \[clm:deg\_into\_subsets\] Suppose $0<1/n \ll c \ll 1/r$ and $s \in \mathbb N$ where $1 \le s \le r$. Let $v \in V$ and $U \subseteq V$. If $$sn/r - cn \le \|U\| \le sn/r + cn,$$ then $$d(v, U) \ge 2(1 - 1/s - rc/s)\|U\|.$$ When $s = 1$ the statement is trivially true, so assume $s \ge 2$. Therefore, $$(1/s + rc/s)\|U\| \ge (1/s + rc/s)(s/r - c)n = (1/r + (s-1)c/s - rc^2/s)n > n/r + 1.$$ Hence, by , $$d(v, U) \ge \delta(G) - 2\|\overline{U}\| \ge 2(\|U\| - (n/r + 1)) > 2(1 - 1/s - rc/s)\|U\|. \qedhere$$ Let $c$ be a constant such that $0 < c < 1$. We call a set $U \subseteq V$, - a $(1, c)$-independent set if $\|U\| = n/r$ and $e(G[U]) < c n^2$, or - a $(2, c)$-independent set if $\|U\| = 2n/r$ and $e_2(G[U]) < c n^2$. If $U,U' \subseteq V$ are vertex-disjoint, we say that the pair $\{U,U'\}$ is - a $(1, c)$-disconnected pair if $\|U\| = \|U'\| = n/r$ and $e(U, U') < c n^2$, or - a $(2, c)$-disconnected pair if $\|U\| = \|U'\| = 2n/r$ and $e_2(U, U') < c n^2$. If $1 \le s \le r$ and $U \subseteq V$, we call $U$ an $(s, c)$-tolerant set when $\|U\| = sn/r$ and - for $t \in \{1, 2\}$, $U$ does not contain a $(t, c)$-independent set, and - if $s \in \{2,4\}$, then $U$ does not contain an $(s/2, c)$-disconnected pair. The preceding definitions are closely related to the notion of being $(1/r, \gamma)$-extremal (Definition \[def:extremal\]) or $(1/r, \gamma)$-splittable (Definition \[def:splittable\]) which the following simple claim makes explicit. \[clm:tolerant-implies-not-extremal\] Let $1 \le s \le r$ where $s \in \mathbb N$, $0<1/n \ll c' \ll \gamma \ll c \ll 1/r$, $U \subseteq V$ such that $\|U\| = sn/r$ and $U' \subseteq U$ such that $\|U \triangle U'\| \le c' n$. If $U \subseteq V$ is $(s,c)$-tolerant, then $G[U']$ is not $(1/s,\gamma)$-extremal and, when $s \in \{2, 4\}$, $G[U']$ is not $(1/s, \gamma)$-splittable. Suppose that $G[U']$ is $(1/s, \gamma)$-extremal. So there exists $W \subseteq U'$ such that either $e(G[W]) < \gamma \|U'\|^2$, and $\|W\| \ge (1/s - \gamma) \|U'\| \ge (1/r - 2\gamma)n$, or $e_2(G[W]) < \gamma \|U'\|^2$, and $\|W\| \ge (2/s - \gamma) \|U'\| \ge (2/r - 2\gamma)n$. Because $c' \ll \gamma \ll c \ll 1/r$ and $\|U \triangle U'\| \le c' n$, it is easy to see we can add vertices from $U$ to $W \cap U$ or delete vertices from $W \cap U$ to create either a $(1, c)$-independent or $(2, c)$-independent set in $U$. This implies that $U$ is not $(s, c)$-tolerant. A similar argument implies that when $s \in \{2, 4\}$, if $G[U']$ is $(1/s, \gamma)$-splittable, then $U$ is not $(s, c)$-tolerant. Claim \[clm:tolerant\_sets\] below is meant to capture all of the necessary facts about $(s, c)$-tolerant sets in a form that will be convenient. In some sense, Claim \[clm:tolerant\_sets\] is just a restatement of the main theorems from the previous sections. With Claim \[clm:deg\_into\_subsets\], (i) and (ii) follow from Corollary \[almostthm\_cor\], and (iii) and (iv) follow from Proposition \[almostthm\_reducedstability\] and Theorem \[mainthm-stability\], respectively. Note that, after the proof of this claim, we do not appeal to Corollary \[almostthm\_cor\], Proposition \[almostthm\_reducedstability\] and Theorem \[mainthm-stability\] again. \[clm:tolerant\_sets\] Let $1 \le s \le r$ where $s \in \mathbb N$ and suppose $1/n \ll \eta \ll \phi \ll \gamma \ll 1/r$, $U \subseteq V$ such that $U$ is $(s, \gamma)$-tolerant and $U' \subseteq V$ such that $\|U \triangle U'\| \le \eta n$. 1. \[clm:Kbar\_in\_Uprime\] If $W \subseteq U'$ such that $\|W\| \ge (s-1)n/r + \phi n$ then $G[W]$ contains a copy of $\bar{\K}_{s}$. 2. \[clm:v\_Khat\_s\_plus\_1\] If $v \in V$ such that $d(v, U') \ge 2(s-1)n/r + \phi n$, then $G[U']$ contains a copy $T$ of $\bar{K}_s$ to which $v$ sends at least $2s-1$ edges, so $G[T + v]$ contains a copy of $\hat{\K}_{s+1}$. 3. \[clm:large\_cliques\] There exist at least $\phi n$ vertex-disjoint copies of $\K'_{s+1}$ in $G[U']$. 4. \[clm:factor\] If $\|U'\|$ is divisible by $s$, then there exists a perfect $\K'_s$-tiling in $G[U']$. Note that, by Claim \[clm:deg\_into\_subsets\], $$\label{eq:min_deg_H} \delta(G[U']) \ge 2(1 - 1/s - r\eta/s)\|U'\|.$$ We first prove (i) and (ii), so let $W$ and $v$ be as in the statement of the claim. When $s \ge 2$, Corollary \[almostthm\_cor\] and imply that there exists a $\bar{K}_s$-tiling $\mathcal{T}$ of $G[U']$ on all but at most $4 s r \eta \|U'\|$ vertices, and when $s = 1$, this is trivially true. Let $Z := V(\T)$ be the vertices in $U'$ that are covered by $\mathcal{T}$. Note that $$(s-1)\|\T\| \le (s-1)\|U'\|/s < (s-1)n/r + \phi n/3.$$ Since $$\|W \cap Z\| \ge \|W\| - \|U' \setminus Z\| \ge \|W\| - 4sr \eta \|U'\| > (s-1)n/r + \phi n/3 > (s-1)\|\mathcal{T}\|,$$ there exists $T \in \mathcal{T}$ such that $V(T) \subseteq W$, and this proves (i). Because $$d(v, U') \ge 2((s - 1)n/r + \phi n/3) + \phi n/3 > 2(s - 1)\|\mathcal{T}\| + 2\|U' \setminus Z\|$$ there exists $T \in \mathcal{T}$ such that $d(v, V(T)) \ge 2s - 1$, so $v$ has at most one light neighbour in $T$. This proves (ii). We will now prove (iii) and (iv). To this end, let $U'' \subseteq U'$ such that $$\|U''\| \ge \|U'\| - (s+1) {\left\lceil \phi n \right\rceil}.$$ Note that $$\|U'' \triangle U\| \le \|U'' \triangle U'\| + \|U' \triangle U\| \le 4(s + 1)\phi n/3 + \eta n \le 3 s \phi n$$ so, by Claim \[clm:deg\_into\_subsets\], $$\delta(G[U'']) \ge 2(1 - 1/s - 3 r \phi)\|U''\|.$$ For $\gamma'$ such that $1/n \ll 3r \phi \ll \gamma' \ll \gamma \ll 1/s$, by Claim \[clm:tolerant-implies-not-extremal\] we have that $G[U'']$ is not $(1/s, \gamma')$-extremal and when $s \in \{2, 4\}$ is not $(1/s, \gamma')$-splittable. If $s=1$, (iv) is vacuously true. So suppose $s \geq 2$. Then Proposition \[almostthm\_reducedstability\] implies that $G[U'']$ has a perfect $(\K'_s \cup \K'_{s+1})$-tiling and, when $s$ divides $\|U''\|$, Theorem \[mainthm-stability\] implies that $G[U'']$ has a perfect $\K'_s$-tiling. Taking $U'' = U'$ then gives (iv). Furthermore, we can greedily select ${\left\lceil \phi n \right\rceil}$ copies of elements from $\K'_{s+1}$ from $G[U']$, since any subset of $U'$ that has order greater than $\|U'\| - (s+1){\left\lceil \phi n \right\rceil}$ contains a copy of $\K'_{s+1}$: When $s \ge 2$, this is true because a perfect $(\K'_s\cup \K'_{s+1})$-tiling in a multigraph of order not divisible by $s$ implies the existence of an element from $\K'_{s+1}$, and, when $s = 1$, this is true because $G[U']$ not being $(1, \gamma')$-extremal implies that $G[U']$ contains at least $\gamma' \|U'\|^2$ edges. This proves (iii). Initial partitioning and sorting -------------------------------- Suppose that $$0 < 1/n \ll \gamma_0 \ll \gamma_1 \ll \dotsm \ll \gamma_{r+2} \ll 1/r$$ and, in addition, for every $i \in [r+2]$, we have $\beta_{i}$ and $\psi_{i}$ such that $$\gamma_{i-1} \ll \beta_{i} \ll \psi_{i} \ll \gamma_{i}.$$ We start by trying to find, for either $s_1=1$ or $s_1 = 2$, an $(s_1, \gamma_1)$-independent set which we will call $A_1$. We then try to find, for $s_2 = 1$ or $s_2 = 2$, an $(s_2, \gamma_2)$-independent set $A_2$ disjoint from $A_1$. We continue in this manner for as long as possible, so in the end we have (a possibly empty, in which case $p=0$) collection of disjoint sets $A_1, \dotsc, A_p$ and integers $s_1, \dotsc, s_p$, such that $A_i$ is $(s_i, \gamma_i)$-independent for each $i \in [p]$. Let $U := \overline{\bigcup_{i \in [p]} A_i}$ and set $s := r - \left(s_1 + \dotsm + s_p\right)$. If $s = 0$, then $U = \emptyset$ and we set $q:=0$ to indicate this case. If $U$ is $(s, \gamma_{p+1})$-tolerant, then set $A_{p+1} := U$, $s_{p+1} := s$ and $q := 1$. Otherwise, $U$ is not $(s, \gamma_{p+1})$-tolerant and because the initial process terminated, $U$ contains neither a $(1, \gamma_{p+1})$-independent set nor a $(2, \gamma_{p+1})$-independent set. By the definition of a tolerant set, it must therefore be that $s \in \{2, 4\}$ and that $U$ has a partition $\{A_{p+1}, A_{p+2}\}$ that is $(s/2, \gamma_{p+1})$-disconnected. We set $s_{p+1}: = s_{p+2} := s/2$ and $q := 2$, to indicate this case. If $q = 1$, set $\tau := \gamma_{p}$, $\beta := \beta_{p+1}$, $\psi := \psi_{p+1}$ and $\gamma := \gamma_{p+1}$. Otherwise, set $\tau := \gamma_{p+1}$, $\beta := \beta_{p+2}$, $\psi := \psi_{p+2}$ and $\gamma := \gamma_{p+2}$. Therefore, $$\label{eq:constants} 0 < 1/n \ll \tau \ll \beta \ll \psi \ll \gamma \ll 1,$$ and we have proved the following claim. \[clm:first\_partition\] There exists a partition $\{A_1, \dotsc, A_{p+q}\}$ of $V$ where $p + q \le r$ and $q \in \{0, 1, 2\}$, and non-negative integers $s, s_1, \dotsc, s_{p+q}$ such that the following holds: 1. $\|A_i\| = s_i n/r$ for $i \in [p+q]$; 2. for every $i \in [p]$, $s_i \in \{1, 2\}$ and $A_i$ is an $(s_i, \tau)$-independent set; 3. if $q = 0$, then $s = 0$; 4. \[clm:properties\_of\_A\_p+1\] if $q = 1$, then $s = s_{p+1}$ and $A_{p+1}$ is $(s, \gamma)$-tolerant; and 5. if $q = 2$, then $s \in \{2, 4\}$ and $\{A_{p+1}, A_{p+2}\}$ is an $(s/2, \tau)$-disconnected pair. Note the relationship between Claim \[clm:first\_partition\] and the examples shown in Figure \[fig:extremalcases\]. In order to discuss the case when $q=2$ and the case when $q\neq2$ in a consistent way, we define a permutation $\sigma$ of $[p+q]$ in the following way. If $q \neq 2$ we let $\sigma$ be the identity permutation, and if $q = 2$ we let $\sigma$ be the transposition of $p+1$ and $p+2$. Note that when $p+1 \le i \le p+q$, and $q \ge 1$, $\|A_i \cup A_{\sigma(i)}\| = sn/r$. When $q \neq 1$, we let $\Lambda := [p+q]$, and when $q = 1$, we let $\Lambda := [p]$. We say that a vertex $v$ is $(i,c)$-typical if $i \in \Lambda$ and $$\text{$s_i = 1$ and $d(v, A_{\sigma(i)}) \le cn$, or $s_i = 2$ and $d^2(v, A_{\sigma(i)}) \le cn$;}$$ or, if $q = 1$, $i = p+1$ and $d^2(v, \overline{A_{p+1}}) \ge \|\overline{A_{p+1}}\| - c n$. \[clm:typical\] For any $i \in [p+q]$, if $v$ is $(i,c)$-typical, then $d^2(v, \overline{A_{\sigma(i)}}) \ge \|\overline{A_{\sigma(i)}}\| - cn - 1$, and, furthermore, if $i \in \Lambda$ and $s_i = 2$, then $\|N(v)\| \ge (1-c)n - 1$. Consider any $i \in [p+q]$ and suppose $v$ is $(i,c)$-typical. If $q=1$ and $i = p+1$, then, recalling that $i = \sigma(i) = p+1$, we have $d^2(v, \overline{A_{\sigma(i)}}) \ge \|\overline{A_{\sigma(i)}}\| - c n$ by definition. Otherwise, using , if $s_i = 1$, we have that $$d^2(v, \overline{A_{\sigma(i)}}) \ge \delta(G) - \|\overline{A_{\sigma(i)}}\| - d(v, A_{\sigma(i)}) \ge \|\overline{A_{\sigma(i)}}\| - cn - 1$$ and if $s_i = 2$, we have that $$d^2(v, \overline{A_{\sigma(i)}}) \ge \delta(G) - (n-1) - d^2(v, A_{\sigma(i)}) \ge \|\overline{A_{\sigma(i)}}\| - cn.$$ When $s_i = 2$, we also have that $$\|N(v)\| \ge \delta(G) - d^2(v) \ge \delta(G) - \|\overline{A_{\sigma(i)}}\| - cn \ge (1 - c)n - 1. \qedhere$$ \[clm:nontypical\_bound\] For any $0 <c \le 1$ and $i \in [p+q]$, there are at most $3r \tau n/c$ vertices in $A_i$ that are not $(i, c)$-typical. Let $i \in [p+q]$ and let $t$ be the number of vertices in $A_i$ that are not $(i,c)$-typical. If $i \in \Lambda$, then we have that $$\begin{aligned} {3} t c n &\le e(A_i, A_{\sigma(i)}) \le 2\tau n^2 &&\text{ when $s_i = 1$, and,} \\ t c n &\le e_2(A_i, A_{\sigma(i)}) \le 2\tau n^2 &&\text{ when $s_i = 2$,} \end{aligned}$$ so $t \le 2 \tau n/c$. Here we define $e(A_i,A_i):=2e(A_i)$ and $e_2(A_i,A_i):=2e_2(A_i)$. If $q = 1$ and $i = p+1$, then our assumption gives us $$\|A_{p+1}\|\|\overline{A_{p+1}}\| - t c n \ge e_2(A_{p+1}, \overline{A_{p+1}}),$$ so if we can show that $$\label{eq:A_p_plus_edge_one_lower_bound} e_2(A_{p+1}, \overline{A_{p+1}}) \ge \|A_{p+1}\|\|\overline{A_{p+1}}\| - 3p\tau n^2,$$ this will imply $t \le 3p \tau n/c$, which will prove the claim. To show , let $j \in [p]$, and recall that $j = \sigma(j)$ and $\|A_j\| \in \{n/r, 2n/r\}$. Clearly, $$e_2(A_j, \overline{A_j}) \ge \|A_j\|\delta(G) - e(A_j, A_j) - \|A_j\|\|\overline{A_j}\|.$$ Therefore, when $\|A_{j}\| = n/r$, $\delta(G) \ge 2 \|\overline{A_j}\| - 1$, and $e(A_j, A_j) \le 2\tau n^2$, so $$e_2(A_j, \overline{A_j}) \ge \|A_j\|\|\overline{A_j}\| - \|A_j\| - 2 \tau n^2 \ge \|A_j\|\|\overline{A_j}\| - 3 \tau n^2.$$ When $\|A_j\| = 2n/r$, $\delta(G) \ge 2\|\overline{A_j}\| + \|A_j\| - 1$, and $e(A_j, A_j) \le \|A_j\|^2 + 2\tau n^2$, so $$e_2(A_j, \overline{A_j}) \ge \|A_j\|\|\overline{A_j}\| + \|A_j\|(\|A_j\| - 1) - (\|A_j\|^2 + 2 \tau n^2) \ge \|A_j\|\|\overline{A_j}\| - 3 \tau n^2.$$ Therefore, $$e_2(A_{p+1}, \overline{A_{p+1}}) = \sum_{j = 1}^{p} e_2(A_{p+1}, A_j) \ge \sum_{j=1}^{p} \left(\|A_{p+1}\|\|A_j\| - 3 \tau n^2\right) = \|A_{p+1}\|\|\overline{A_{p+1}}\| - 3p \tau n^2.$$ Let $\mathcal{U} = (U_1, \dotsc, U_{p+q})$ be an ordered collection of $p+q$ pairwise disjoint subsets of $V$. We say that an $r$-set $T \subseteq \bigcup_{i=1}^{p+q} U_i$ is $\mathcal{U}$-balanced, if - $G[T]$ contains a copy of $\K'_r$, - $\|V(T) \cap U_i\| = s_i$ for all $i \in [p]$, and - $\|V(T) \cap \left(U_{p+1} \cup U_{\sigma(p+1)}\right)\| = s$ when $q > 0$. A set $T$ is called $\mathcal{U}$-well-balanced, if $T$ is $\mathcal{U}$-balanced, and when $q = 2$, $T$ intersects exactly one of the two sets $U_{p+1}$ and $U_{p+2}$, i.e. for some $i \in \{p+1, p+2\}$, $\|V(T) \cap U_i\| = s$ and $\|V(T) \cap U_{\sigma(i)}\| = 0$. Note that when $q \neq 2$ every $\mathcal{U}$-balanced set is a $\mathcal{U}$-well-balanced set. We say that a vertex is excellent for $i$ if it is $(i, \beta^2)$-typical and we say that a vertex is good for $i$ if it is $(i, \psi^2)$-typical. We make the following definitions: $$B_i := \{v \in A_i : v \text{ is excellent for $i$} \} \text{ and } B := \bigcup_{ i \in [p+q] } B_i,$$ and we let $$C_i := B_i \cup \{v \in \overline{B} : v \text{ is good for $i$} \} \text{ and } C := \bigcup_{ i \in [p+q] } C_i.$$ We let $\mathcal{B}$ be the ordered collection $(B_1, \dotsc, B_{p+q})$ and $\mathcal{C}$ be the ordered collection $(C_1, \dotsc, C_{p+q})$. Note that, by Claim \[clm:typical\], every vertex is good for at most one index $i \in [p+q]$, so the sets $C_1, \dotsc, C_{p+q}$ are pairwise disjoint. Note that, for every $i \in [p+q]$, each vertex $v \in \overline{C}$ is not good for $i$, so it has a large number of edges into $B_{\sigma(i)}$. Since each vertex $v\in B_i$ is adjacent to almost everything in $\overline{B_\sigma(i)}$, we can argue below that there exists a set $T$ that is $\mathcal{B}$-well-balanced and such that $G[T + v]$ contains a copy of $\K'_{r+1}$. As we will see, this will allow us to easily distribute the vertices in $\overline{C}$ to sets in $\mathcal{C}$. Claims \[clm:typical\] and \[clm:nontypical\_bound\] and the fact that $\tau \ll \beta \ll \psi \ll 1/r$ imply the following claim. \[clm:second\_partition\] The following holds: 1. \[clm:size\_B\_i\_C\_i\] $\|\overline{C}\| \le \|\overline{B}\| \le \beta^2 n$, and, in particular, for every $i \in [p+q]$, $$\text{$\|A_i \triangle C_i\|, \|A_i \triangle B_i\| \le \beta^2 n$ and $\|C_i \setminus B_i\| \le \beta^2 n$;}$$ 2. \[clm:excellent\] for every $i \in [p+q]$,\ $v \in B_i \implies d^2(v, \overline{B_{\sigma(i)}}) \ge \|\overline{B_{\sigma(i)}}\| - \beta n$, and\ $v \in C_i \implies d^2(v, \overline{B_{\sigma(i)}}) \ge \|\overline{B_{\sigma(i)}}\| - \psi n$; 3. \[clm:s\_i\_equals\_2\] for every $i \in \Lambda$, if $s_i = 2$, then $$\text{ $v \in B_i \implies \|N(v)\| \ge (1 - \beta) n$ and $v \in C_i \implies \|N(v)\| \ge (1 - \psi) n$. }$$ Looking ahead, we will construct an ordered collection $\mathcal{D} = (D_1, \dotsc, D_{p+q})$ such that $\{D_1, \dotsc, D_{p+q}\}$ is a partition of $V$ and such that there exists a perfect $\mathcal K_{r}'$-tiling in $G$ such that every element in the tiling is $\mathcal{D}$-well-balanced. This trivially implies that $\mathcal{D}$ must have the following properties: 1. \[proper\_std\] for every $i \in [p]$ and for $i = p+1$ when $q = 1$, $\|D_i\| = s_i \cdot n/r$, 2. \[proper\_q\_equals\_2\] when $q = 2$, both $\|D_{p+1}\|$ and $\|D_{p+2}\|$ are divisible by $s$ and $\|D_{p+1} \cup D_{p+2}\| = s n/r$. For any multigraph $G' \subseteq G$, call $\mathcal{D} = (D'_1, \dotsc, D'_{p+q})$ a proper ordered collection of $G'$ if the sets in $\mathcal{D}$ form a partition of $V(G')$ and it meets conditions (\[proper\_std\]) and (\[proper\_q\_equals\_2\]) (with $n$ replaced by $\|G'\|$). Let $c_i := \|C_i\| - s_in/r$ for every $i \in [p+q]$, so $\|\overline{C}\| + c_1 + \dotsm + c_{p+q} = 0$. In order to make the rest of the section easier to understand, we now give a brief, informal description of the remainder proof for the case when $q \neq 2$. It is very similar to the approach taken by Koml[ó]{}s, S[á]{}rk[ö]{}zy and Szemer[é]{}di in their proof of the Alon–Yuster conjecture [@komlos2001proof]. We ignore the case when $q=2$ in this description to avoid technical details. Our main goal is to balance the sizes of the sets $C_1, \dotsc, C_{p+q}$. We begin by considering the sets that are too large, i.e. for every $i \in [p+q]$ such that $c_i > 0$, we move exactly $c_i$ vertices out of $C_i$ to form the set $D_i$. We want to ensure that the vertices which are moved out of $C_i$ can eventually be covered by a copy of $\K'_r$ that has $s_i$ other vertices in $C_i$. Therefore, we argue that we can find a $\K'_{s_i + 1}$-tiling in $G[C_i]$ of size $c_i$. From each element of this tiling we arbitrarily select a vertex $v$ to remove from $C_i$ when forming $D_i$, and, temporarily, place $v$ into a set we call $F'$. Then we extend each element of this tiling to form a copy of $\K'_{r+1}$ that has exactly $s_j$ vertices in $C_j$ for each $j \in [p+q] \setminus i$, and use the label $T_v$ for this copy of $\K'_{r+1}$. So $T_v - v$ is $\mathcal{C}$-well-balanced . We let $\T = \{T_v : v \in F'\}$. Note that, after this process has completed, we have that, for every $c_i \ge 0$, $D_i = C_i \setminus F'$ and $\|D_i\| = s_i n/r$. Next, we prepare to distribute the vertices that were not assigned to some set $C_i$. To do this, for every such $v \in \overline{C}$, we find a $\B$-well-balanced $r$-set $T'$ such that $G[T' + v]$ contains an element of $\K'_{r+1}$. We then label $T' + v$ as $T_v$, and add $T_v$ to $\T$. Throughout, we ensure that the elements in $\T$ are vertex-disjoint. We let $F = F' \cup \overline{C}$. Note that by Claim \[clm:second\_partition\](\[clm:size\_B\_i\_C\_i\]), $$\|F\| = \sum_{i \in [p+q];~c_i < 0} -c_i \le \|\overline{B}\| \le \beta^2 n.$$ So we can distribute the small number of vertices in $F$ arbitrarily to every $D_i$ such that $\|D_i\| < s_i n/r$ until $\|D_i\| = s_in/r$ for every $i \in [p+q]$. Suppose $v \in F$ has been assigned to $D_i$; so $T_v$ has exactly $s_i$ vertices in $C_i$ and, with $v$, has $s_i + 1$ vertices in $D_i$. We can then arbitrarily remove one element from $T_v \cap C_i$ to create a $\mathcal{D}$-well-balanced set. After this has been done for every $v \in F$, we have that $\T$ is a $\K'_r$-tiling in which every element corresponds to a $\mathcal{D}$-well-balanced $r$-set. We let $G' = G - V(\T)$ and $D'_i = D_i \setminus V(\T)$ for every $i \in [p+q]$. So $(D'_1, \dotsc, D'_{p+q})$ is a proper ordered collection of $G'$ and $D'_i \subseteq C_i$ for every $i \in [p+q]$. Claim \[clm:proper\_partitions\] below will then complete the proof. When $q = 2$, our approach is similar. The main difference is that we have to be somewhat careful to ensure that $\|D_{p+1}\|$ and $\|D_{p+2}\|$ are both divisible by $s$. The details for all cases are in Section \[sec:finishing\]. We now continue the proof of Theorem \[thm:main\_extremal\] by proving Claim \[clm:proper\_partitions\] which will be used at the very end of the proof to construct the vast majority of elements of our tiling of $G$. \[clm:proper\_partitions\] Let $G'$ be an induced subgraph of $G$ such that $\|G'\| \ge \|G\| - \beta n$, and let $\mathcal{D}' = (D'_1, \dotsc, D'_{p+q})$ be a proper ordered collection of $V(G')$. If $D'_i \subseteq C_i$ for every $i \in [p+q]$, then there exists a perfect $\K'_r$-tiling in $G'$. For $i \in [p]$, we let $\mathcal{T}_i$ be a perfect $\bar{\K}_{s_i}$-tiling of $G[D'_i]$. When $s_i=1$ such a tiling trivially exists, and, when $s_i=2$, it exists by Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]). When $q \ge 1$, we let $\mathcal{T}_{p+1}$ be a perfect $\K'_s$-tiling of $G[D'_{p+1} \cup D'_{\sigma(p+1)}]$. By Claim \[clm:second\_partition\](\[clm:excellent\]), this is easy to find when $q = 2$ by, say, applying the Hajnal–Szemerédi theorem to the graph induced by the heavy edges of $D'_{p+1}$ and then to the graph induced by the heavy edges of $D'_{p+2}$. (Actually the use of the Hajnal–Szemerédi theorem here is overkill; it is very easy to directly argue the desired perfect tiling in $G[D'_{p+1} \cup D'_{\sigma(p+1)}]$ exists.) When $q=1$, it is implied by Claim \[clm:tolerant\_sets\](\[clm:factor\]), because $\beta \ll \gamma$ and $$\|A_{p+1} \triangle D'_{p+1}\| \le \|A_{p+1} \triangle C_{p+1}\| + (\|G\| - \|G'\|) \le (\beta^2 + \beta)n .$$ Let $t = p$ when $q=0$, or let $t = p+1$, when $q \ge 1$, and let $H$ be a $t$-partite graph with vertex classes $\T_1, \dotsc, \T_t$ such that, for every distinct $i,i' \in [t]$, $T \in \mathcal{T}_i$ and $T' \in \mathcal{T}_{i'}$, $T$ is adjacent to $T'$ when every vertex in $T$ is heavily adjacent to every vertex in $T'$. Note that $H$ is balanced with each vertex class of size $m = \|G'\|/r$. So we are done if there is a perfect $K_t$-tiling in $H$. By Claim \[clm:second\_partition\](\[clm:excellent\]), when $i,i'$ are distinct and $T \in \mathcal{T}_i$, the number of neighbours of $T$ in $\mathcal{T}_{i'}$ is at least $$\label{eq:deg_in_H} m - \left\| \bigcup_{u \in T} \left( D'_{i'} \setminus N^2(u)\right) \right\| \ge m - r \psi n \ge (1 - 2r^2 \psi)m.$$ For some $1 \le t' < t$, assume we have a perfect $K_{t'}$-tiling $\mathcal{K}$ of $H[\mathcal{T}_1 \cup \dotsm \cup \mathcal{T}_{t'}]$. We extend $\mathcal{K}$ to a perfect $K_{t' + 1}$-tiling of $H[\mathcal{T}_1 \cup \dotsm \cup \mathcal{T}_{t'+1}]$ by finding a perfect matching in the bipartite graph with vertex classes $\mathcal{K}$ and $\mathcal{T}_{t' + 1}$ in which $K \in \mathcal{K}$ is adjacent to $T \in \mathcal{T}_{t'+1}$ when $K + T$ is a copy of $K_{t'+1}$ in $H$. By , this bipartite graph has minimum degree at least $(1 - 2r^2 t' \psi)m \ge m/2$, so Hall’s Theorem implies that it has a perfect matching. Since this holds for every $t' < t$, there exists a perfect $K_t$-tiling in $H$. In Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_std\]) below we establish that, for some $k \in [p+q]$, we can extend a copy of $\K'_{s_k + 1}$ (resp. $\K'_{s_k}$) that is contained in some $C_k$ where $k \le p$ or extend a copy of $\K'_{s+1}$ (resp. $\K'_{s}$) contained in $C_{p+1} \cup C_{\sigma(p+1)}$ to a copy of $\K'_{r+1}$ (resp. $\K'_{r}$) with the correct number of vertices in every $C_j$ for $j \in [p+q] - i$. This is used when we remove elements from the sets $C_k$ that are too large as described in the overview above. When $q = 2$, we use Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_std\]) with $t = 0$ when $\|C_{p+1} \cup C_{p+2}\| \ge sn/r$ to move vertices between $C_{p+1}$ to $C_{p+2}$ to make the order of both sets divisible by $s$. The second part of the lemma, (\[clm:extend\_partial\_clique\_special\]), which is only used when $q = 2$, is similar but we start with a copy of $\K'_{s_k + 1}$ in some $C_k$ with $k \le p$ and extend it only into a copy of $\K'_r$. Furthermore, for any $0 \le \ell \le s - 1$, this $\K'_r$ will have exactly $\ell$ vertices in $C_{p+1}$ and $s - \ell - 1$ vertices in $C_{p+2}$. This is used because sometimes when we must move a vertex from some $C_k$ with order greater than $s_k n/r$, to $C_{p+1}$ because the order of $C_{p+1} \cup C_{p+2}$ is slightly less than $sn/r$, we may also have to move some vertices in $C_{p+2}$ to $C_{p+1}$ to ensure that both sets are divisible by $s$. \[clm:extend\_partial\_clique\] Let $W \subseteq V$ such that $\|W\| \le \beta n$, let $k \in [p+q]$ and $t \in \{0,1\}$. When $k \le p$, let $S \subseteq C_k$ such that $\|S\| = s_k + t$ and when $k \ge p + 1$, let $S \subseteq C_k \cup C_{\sigma(k)}$ such that $\|S\| = s + t$. Suppose that $G[S]$ contains an element of $\K'_{\|S\|}$. 1. \[clm:extend\_partial\_clique\_std\] There exists a set $T$ that avoids $W$ such that $T \cup S$ is $\C$-balanced when $t = 0$, or, when $t = 1$, $T \cup (S - v)$ is $\C$-balanced for any $v \in S$. Furthermore, when $q = 2$, the set $T$ does not intersect $C_{p+2}$. 2. \[clm:extend\_partial\_clique\_special\] When $q = 2$, $t = 1$, and $k \notin \{p+1, p+2\}$, for any $j \in \{p+1, p+2\}$ and $1 \le \ell \le s - 1$, there exists a set $T$ that avoids $W$ such that $G[T \cup S]$ contains an element of $\K'_r$, $\|T \cap C_j\| = \ell$, $\|T \cap C_{\sigma(j)}\| = s-1 -\ell$ and $\|T \cap C_i\| = s_i$ for each $i \in [p] - k$. We will construct $T$ iteratively, and throughout, we let $$C'_i := \left(C_i \setminus W \right) \cap \left(\bigcap_{u \in T \cup S} N^2(u) \right).$$ Assuming $T \subseteq C$, Claim \[clm:second\_partition\](\[clm:excellent\]) implies that for all $i$ such that $C_{\sigma(i)} \cap \left(S \cup T \right) = \emptyset$, $$\label{eq:size_of_Cprime} \|C'_i\| \ge \|C_i\| - \|W\| - \|T \cup S\|\psi n \ge \|C_i\| - 2r \psi n.$$ We start the construction by adding vertices from $C'_{p+1} \cup C'_{\sigma(p+1)}$ to $T$. If $q = 0$, or $q \ge 1$ and $k \ge p+1$, we do not add any vertices from $C'_{p+1} \cup C'_{\sigma(p+1)}$ to $T$. Otherwise, we know $\|C'_{p+1} \cup C'_{\sigma(p+1)}\|$ is large by since $(C_{p+1} \cup C_{\sigma(p+1)}) \cap S = \emptyset$. If $q = 1$, then, since $\beta \ll \psi \ll \gamma$, Claim \[clm:tolerant\_sets\](\[clm:Kbar\_in\_Uprime\]) and Claim \[clm:second\_partition\](\[clm:size\_B\_i\_C\_i\]) imply that we can let $T \subseteq C'_{p+1}$ such that $\|T\| = s$ and $G[T]$ contains an element of $\bar{\K}_s$. To prove (\[clm:extend\_partial\_clique\_std\]) when $q = 2$, note that and Claim \[clm:second\_partition\](\[clm:excellent\]), imply that we can let $T$ induce a clique of size $s$ on heavy edges in $C'_{p+1}$. To prove (\[clm:extend\_partial\_clique\_special\]), we start by using and Claim \[clm:second\_partition\](\[clm:excellent\]) to find vertices $u_1,\dotsc,u_{\ell} \in C'_j$ that form a clique on heavy edges and add these vertices to $T$. If $s - 1 - \ell = 0$, we are done, so assume that this is not the case, which implies that $s = 4$. Therefore, Claim \[clm:second\_partition\](\[clm:excellent\]) and Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]), imply that we can find $s-1-\ell$ vertices in $N(u_1) \cap \dotsm \cap N(u_\ell) \cap C'_{\sigma(j)}$ that form a clique on heavy edges. We then add these vertices to $T$. Note that in this case when $s=4$ either $\ell=1$ and $s-1-\ell=2$ or $\ell =2$ and $s-1-\ell=1$. So in this case currently $G[T]$ contains all possible edges except that it may have a light path on $3$ vertices. Now, for every $i \in \{1, \dotsc, p\} - k$, in turn we use , to add either one vertex, when $s_i = 1$, or two adjacent vertices, when $s_i = 2$, from $C'_i$ to $T$. When $s_i = 2$ we can easily find two adjacent vertices in $C'_i$ using Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]). Further, note that by definition of $C'_i$, in this step any selected vertices in $C'_i$ send heavy edges to all previously selected elements of $S \cup T$. This ensures that there are all possible edges in $G[S \cup T]$ except for perhaps a collection of vertex-disjoint light edges, and at most one path on $3$ vertices (in the case when $s=4$). That is, $G[S\cup T]$ is a copy of $\mathcal K' _{\|S\cup T\|}$. It is now easy to see that the claim holds. The main purpose of Claim \[clm:extend\_overline\_C\](\[clm:extend\_overline\_C\_std\]) is to help distribute the vertices $v \in \overline{C}$. We construct a $\B$-well balanced set $T$ such that $G[T + v]$ contains an element of $\K'_{r+1}$; thus we can then move $v$ to any set $D_i$ and remove any element in $B_i \cap T$ from $T$ leaving a $\D$-well-balanced set. When $q=2$, we may need to move some $v \in \overline{C}$ to one of $D_{p+j}$ for $j \in \{1, 2\}$ and, at the same time, move some vertices from $D_{p+3-j}$ to $D_{p+j}$ to ensure that $D_{p+1}$ and $D_{p+2}$ are both divisible by $s$. This is the reason for Claim \[clm:extend\_overline\_C\](\[clm:extend\_overline\_C\_special\]). Note that the following claim is essentially the reason we define both the sets $B_{1}, \dotsc, B_{p+q}$ and the sets $C_{1}, \dotsc, C_{p+q}$. \[clm:extend\_overline\_C\] Let $W \subseteq V$ such that $\|W\| \le \beta n$ and $v \in \overline{C}$. 1. \[clm:extend\_overline\_C\_std\] There exists a $\B$-well-balanced set $T$ that avoids $W$ such that $G[T + v] \in \K'_{r+1}$. 2. \[clm:extend\_overline\_C\_special\] When $q = 2$, for $j \in \{p+1, p+2\}$ and $1 \le \ell \le s - 1$, there exists a set $T$ that avoids $W$ such that $G[T + v]$ contains an element of $\K'_r$, $\|T \cap B_j\| = \ell$, $\|T \cap B_{\sigma(j)}\| = s - 1 - \ell$ and $\|T \cap B_i\| = s_i$ for each $i \in [p]$. For any $i \in [p+q]$, if $i \in \Lambda$, since $v$ it not $(i, \psi^2)$-typical, Claim \[clm:second\_partition\](\[clm:size\_B\_i\_C\_i\]) implies that $$\text{when $s_i = 1$, $d(v, B_i) \ge \psi^2 n/2$, and, when $s_i = 2$, $d^2(v, B_i) \ge \psi^2 n/2$.}$$ Similarly, when $q = 1$, since $v$ is not $(p+1, \psi^2)$-typical, $d^2(v, \overline{A_{p+1}}) \le \|\overline{A_{p+1}}\| - \psi^2 n$ and $$\label{eq:bad_vertices} \begin{split} d(v, B_{p+1}) &\ge \delta(G) - d(v, \overline{B_{p+1}}) \\ &\ge \delta(G) - \left(\|\overline{A_{p+1}}\| + d^2(v, \overline{A_{p+1}}) + 2\|\overline{B_{p+1}} \setminus \overline{A_{p+1}}\|\right) \\ &\ge 2(1 - 1/s)sn/r + \psi^2 n/2. \end{split}$$ For $i \in [p+q]$, let $B'_{i} := B_i \setminus W$. We have that, by Claim \[clm:second\_partition\](\[clm:size\_B\_i\_C\_i\]), $$\label{eq:A_i_symdiff_B_i_prime} \|A_i \triangle B'_i\| \le \|A_i \triangle B_i\| + \|W\| \le 2\beta n.$$ Let $$I := \{i \in \Lambda : s_i = 1 \text{ and } d^2(v, B_i) \le \psi^2 n/2 \}$$ and note that there are at least $\sum_{i \in I} \left(\|B_i\| - \psi^2 n/2\right) \ge \|I\|(1/r - \psi^2)n$ vertices in $\bigcup_{i \in I} B_i$ that are not heavy neighbours of $v$. By , when $r > 2$, $d^2(v) \ge \delta(G) - (n-1) \ge r \psi n/2$, so there exists $j \in [p + q] \setminus I$, and $$d^2(v, B_{j}) \ge \delta(G) - (n - 1) - \left(\|\overline{B_{j}}\| - \|I\|(1/r - \psi^2)n \right) \ge \|B_j\| - (2/r - \|I\|/r + \|I\|\psi^2)n.$$ Hence, in all cases, $\|I\| \le 2$, and, for any $j \in [p+q] \setminus I$, $$\label{eq:I_equals_2} d^2(v, B_{j}) \ge \|B_{j}\| - 2 \psi^2 n \qquad \text{ if $\|I\| = 2$.}$$ We prove (\[clm:extend\_overline\_C\_std\]) and (\[clm:extend\_overline\_C\_special\]) simultaneously. Let $t \in \{0,1\}$, so that $s - 1 + t$ is the number of vertices of $T$ that will intersect $B_{p+1} \cup B_{\sigma(p+1)}$, i.e. when we are proving (\[clm:extend\_overline\_C\_std\]) we have $t = 1$ and when we are proving (\[clm:extend\_overline\_C\_special\]) we have $t = 0$. We now give a brief overview of our plan for constructing the set $T$. Clearly we must construct $T$ so that every pair of vertices in $T + v$ are adjacent. We will also have that - the only light edges in $G[T + v]$ that are not incident to $v$, are in the subgraphs $G[T \cap B'_i]$ where $s_i = 2$, or are in $G[T \cap \left(B'_{p+1} \cup B'_{\sigma(p+1)}\right)]$ when $q \ge 1$; - $v$ is heavily adjacent to every vertex $u$ in $T \cap B'_i$ when $i \in \Lambda \setminus I$; - if $q \ge 1$, $G[T \cap \left(B'_{p+1} \cup B'_{\sigma(p+1)}\right) + v]$ will contain an element of $\K'_{s - 1 + t}$. If $v$ has at most one light neighbour in $T$, this is enough to give us that $G[T + v]$ contains a an element of $\K'_{r+t}$, which would prove the claim. However, we can only ensure that $v$ has at most two light neighbours in $T$. To prove the claim, we will then also meet one of the following conditions when $v$ has two light neighbours in $T$: - the two light neighbours of $v$ are in different sets $B'_i$ such that $i \in I$ and $i \le p$ and if $q \ge 1$, $G[T \cap (B'_{p+1} \cup B'_{\sigma(p+1)})]$ contains an element of $\overline{\K}_{s-1+t}$ except possibly when $q=2$, $s=4$, $t = 0$ and $\ell \in \{1, 2\}$, and in this case it contains an element of $\hat{\K}_{s-1+t}$; or - $q = 2$, $s=2$, $t=1$, $I = \{p+1, p+2\}$ and the two light neighbours of $v$ are in some $B'_i$ where $i \in I$ and these two light neighbours are heavily adjacent; or - $q = 2$, $s=2$, $t=0$, $\|I\| = 2$, $j \in I \cap \{p+1, p+2\}$, $v$ has one light neighbour in $B'_j$ and the other light neighbour is in some $B'_i$ where $i \in I$ and $i \le p$; or - $q=1$, $\|I\| = 1$, one light neighbour of $v$ is in the set $B'_i$ such that $i \in I$ and the other light neighbour of $v$ is in $B'_{p+1}$ and $G[T \cap B'_{p+1} + v]$ contains an element of $\hat{\K}_{s+t}$. We build the set $T$ iteratively. We start by adding vertices from $B'_{p+1} \cup B'_{\sigma(p+1)}$ to $T$, so when $q = 0$ we do not add anything to $T$. Recall that $\beta \ll \psi \ll \gamma$. If $q = 1$ and $\|I\| \le 1$, then , and Claim \[clm:tolerant\_sets\](\[clm:v\_Khat\_s\_plus\_1\]), imply that we can choose $T \subseteq B_{p+1}'$, such that $\|T\| = s$, $G[T + v]$ contains an element of $\hat{\K}_{s+1}$ and in which $v$ has one light neighbour. If $q = 1$ and $\|I\| = 2$, then by , , and Claim \[clm:tolerant\_sets\](\[clm:Kbar\_in\_Uprime\]) we can let $T \subseteq B_{p+1}'$ be such that $T \subseteq N^2(v)$ and $G[T]$ contains an element of $\bar{\K}_s$. Now assume $q = 2$. To prove (\[clm:extend\_overline\_C\_std\]), when $I = \{p+1, p+2\}$, we pick $j \in \{p+1, p+2\}$ arbitrarily, otherwise we let $j \in \{p+1, p+2\} \setminus I$. Recall that $j \in I$ implies that $s = 2$. Let $Z := N(v) \cap B'_j$ when $j \in I$ and let $Z := N^2(v) \cap B'_j$ when $j \notin I$. Note that, $\|Z\| \ge \psi^2 n/2 - \|W\| \ge \psi^2 n / 3$, in either case. We now use Claim \[clm:second\_partition\](\[clm:excellent\]) to find an $s$-set $T \subseteq Z$ such that $T$ induces a clique on heavy edges in $G$. To prove (\[clm:extend\_overline\_C\_special\]), we assume $j \in \{p+1, p+2\}$ is given. Note that if $j \in I$, then we must have that $s = 2$ and $\ell = 1$. Again, we let $Z: = N(v) \cap B'_j$ when $j \in I$ and let $Z := N^2(v) \cap B'_j$ when $j \notin I$, so $\|Z\| \ge \psi^2 n / 3$. Now we use Claim \[clm:second\_partition\](\[clm:excellent\]) to find vertices $u_1,\dotsc,u_{\ell} \in Z$ such that they induce a clique on heavy edges in $G$, and we add these vertices to $T$. If $s - 1 - \ell = 0$, we are done, so assume that this is not the case, which implies that $s = 4$. Therefore, Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]) and Claim \[clm:second\_partition\](\[clm:excellent\]), imply that we can find $s-1-\ell$ vertices in $N(u_1) \cap \dotsm \cap N(u_\ell) \cap N^2(v) \cap B'_{\sigma(j)}$ that induce a clique on heavy edges in $G$. We then add these vertices to $T$. If $\ell = \{1, 2\}$, then $G[T]$ contains an element of $\hat{K}_3$ and when $\ell = 3$ it is a clique on heavy edges. Note that in all cases the way we have constructed $T$ so far ensures that ($\alpha _3$) holds. Now, in turn for each $i$ with $1\leq i \leq p$, we will add $s_i$ vertices from $B'_i$ to $T$. At each step, when $i \in I$, we let $$Z := B'_i \cap N(v) \cap \left(\bigcap_{u \in T} N^2(u)\right), \text{ and otherwise let } Z := B'_i \cap N^2(v) \cap \left(\bigcap_{u \in T} N^2(u)\right).$$ (So the definition of $Z$ gets updated at each step, as we add more elements to $T$.) Note that, with Claim \[clm:second\_partition\](\[clm:excellent\]), $\|Z\| \ge \psi^2 n/2 - \|W\| - r \beta n \ge \psi^2 n/3$ in both cases. When $s_i = 1$, we we add one vertex from $Z$ to $T$, and when $s_i = 2$, we can add two adjacent vertices in $Z$ to $T$, since Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]) implies that there exists an edge in $G[Z]$. This completes the construction of $T$. Note that ($\alpha_1$) and ($\alpha_2$) immediately hold. Further, one of ($\beta_1$)–($\beta _4$) holds in each case. It is easy to see that in any case we obtain a set $T$ as desired. Finishing the proof {#sec:finishing} ------------------- We now finish the proof by constructing $\mathcal{D}=(D_1, \dots, D_{p+q})$ a proper ordered collection of $G$ and a collection $\T$ of vertex-disjoint $\mathcal{D}$-well-balanced sets as described above. We build the collection $\T$ iteratively, and, at times, it may include $(r+1)$-sets, as well as $r$-sets. Let $c_i := \|C_i\| - s_i n/r$ for every $i \in [p+q]$. For each $i \in [p]$ such that $c_i \ge 0$, we will find a $\hat{K}_{s_i+1}$-tiling $\mathcal{S}_i$ of $G[C_i]$ containing exactly $c_i$ elements. When $s_i = 1$, each vertex in $C_i$ has at least $${\left\lceil(\delta(G) - 2\|\overline{C_i}\|)/2\right\rceil} \ge \|C_i\| - n/r = c_i$$ neighbours in $C_i$. Therefore, by Lemma \[lem:matching\], we can let $\mathcal{S}_i$ be a matching of size $c_i$ in $G[C_i]$. Similarly, for $i \in [p]$ such that $s_i = 2$ and $\|C_i\| \ge 2n/r$, there exists a matching $M_i$ containing at least $$\delta(G) - (n - 1) - \|\overline{C_i}\| \ge \|C_i\| - 2n/r=c_i$$ heavy edges in $G[C_i]$, and, by Claim \[clm:second\_partition\](\[clm:s\_i\_equals\_2\]), we can pair each edge in $e \in M_i$ to a distinct vertex $v_e$ such that $v_e$ is a neighbour of both endpoints of $e$. Therefore, we have a collection $\mathcal{S}_i$ of $c_i$ vertex-disjoint elements of $\hat{\K}_3$ in $G[C_i]$. We let $\mathcal{S}$ be the union of the sets $\mathcal{S}_i$ constructed so far. If $q = 1$ and $\|C_{p+1}\| > sn/r$, we can use Claim \[clm:tolerant\_sets\](\[clm:large\_cliques\]) to find a $\K'_{s+1}$-tiling of size $c_{p+1} = \|C_{p+1}\| - sn/r$ in $G[C_{p+1}]$, we call this tiling $\mathcal{S}_{p+1}$ and we add it to $\mathcal{S}$. Now suppose $q = 2$ and $\|C_{p+1} \cup C_{p+2}\| \ge sn/r$; we find a tiling consisting of exactly $ \|C_{p+1} \cup C_{p+2}\|-sn/r$ copies of $\K'_{s+1}$ in $C_{p+1}$. Note that this is trivial to do, by Claim \[clm:second\_partition\](\[clm:excellent\]). Indeed, we can easily find $\|C_{p+1} \cup C_{p+2}\| - sn/r$ vertex disjoint $(s+1)$-sets in $C_{p+1}$, each forming a clique in the heavy edges of $G[C_{p+1}]$. We call this set $\mathcal{S}_{p+1}$ and we add the sets in $\mathcal{S}_{p+1}$, to $\mathcal{S}$. Note that there is slack in the argument here: given a single fixed set $X \subseteq V$ where $\|X\|\leq \gamma n$ we can additionally ensure no tile in $\mathcal{S}_{p+1}$ intersects $X$. We will use this property shortly. We have now completely defined $\mathcal S$. Our next goal is to construct the sets $D_1, \dots, D_{p+q}$ such that $\mathcal{D}=(D_1, \dots, D_{p+q})$ is a proper ordered collection of $G$. We will first define the $D_i$ for $i$ such that $c_i\geq 0$. Once we have done this we will then define the remaining $D_i$. Suppose that $q = 2$ and $\|C_{p+1} \cup C_{p+2}\| \ge sn/r$. Remove $c_{p+1}$ vertices from $C_{p+1}$ and call the resulting set $D_{p+1}$; we do this in such a way that these removed vertices consist of precisely one vertex from each tile in $\mathcal S_{p+1}$. Place these $c_{p+1}$ vertices into a set $F'$. Let $D_{p+2}:=C_{p+2}$. This ensures that $\|D_{p+1} \cup D_{p+2}\| = sn/r$. However, we also need to ensure that both $\|D_{p+1}\|$ and $\|D_{p+2}\|$ are divisible by $s$. Therefore, we will find a set $X \subseteq C_{p+1}$ and a set $Y \subseteq C_{p+2}$ such that, $\|X \cup Y\| = s$, $G[X \cup Y]$ contains an element of $\K'_s$, and $$\|Y\| \equiv \|C_{p+2}\| \pmod s.$$ We will then move the vertices in $Y$ from $D_{p+2}$ to $D_{p+1}$. We will need the exact minimum degree condition to construct these sets $X,Y$. So actually formally what we do is first construct $X$ and $Y$ then the collection $\mathcal{S}_{p+1}$ as before such that $\mathcal{S}_{p+1}$ is disjoint from the set $X$. Also, for consistency, we will construct the sets $X$ and $Y$ even when there is no divisibility issue, i.e. when $\|C_{p+2}\|$ is divisible by $s$. In this case, we let $Y = \emptyset$ and $X$ be a clique on $s$ vertices in the heavy edges of $G[C_{p+1}]$, which can be found easily using Claim \[clm:second\_partition\](\[clm:excellent\]). Therefore, it only remains to show how we construct $X$ and $Y$ when $\|C_{p+2}\|$ is not divisible by $s$. If $s = 2$, then, for some $i \in \{p + 1, p + 2\}$, $\|C_i\| \ge n/r$, so $\delta(G) - 2(\|\overline{C_i}\| - 1) \ge 1$, which implies that every vertex in $\overline{C_i}$ has a neighbour in $C_i$, and, in particular, there exists an edge $xy$ such that $x \in C_{p+1}$ and $y \in C_{p+2}$, and we let $X := \{x\}$ and $Y := \{y\}$. If $s = 4$, then for some $i \in \{p+1, p+2\}$, $\|C_{i}\| \ge 2n/r$, and $\delta(G) - (n-1) - (\|\overline{C_i}\| - 1) \ge 1$, which implies that every vertex in $\overline{C_i}$ has a heavy neighbour in $C_i$ and we can let $x_1y_1$ be a heavy edge such that $x_1 \in C_{p+1}$ and $y_1 \in C_{p+2}$. Let $j$ be such that $1 \le j \le 3$ and $j \equiv \|C_{p+2}\| \pmod 4$. By Claim \[clm:second\_partition\](\[clm:excellent\])–(iii), we can find vertices $y_2, \dotsc, y_j$ in $N(x_1) \cap (C_{p+2} - y_1)$, such that $Y = \{y_1, y_2, \dotsc, y_j\}$ induces a clique on heavy edges, and then find vertices $x_2, \dotsc, x_{4 - j}$ in $N(y_1) \cap \dotsm \cap N(y_j) \cap (C_{p+1} - x_1)$ such that $X = \{x_1, \dotsc, x_{4-j}\}$ induces a clique on heavy edges. Note that $G[X \cup Y]$ contains an element of $\K'_4$. We have described how to define $D_{p+1}$ and $D_{p+2}$ in the case when $q=2$ and $c_{p+1}\geq 0$. We now describe in general how to construct $D_i$ when $c_i \geq 0$. Using Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_std\]), we can find, for each set in $S \in \mathcal{S}$, a set $T'$ such that $G[T' \cup S]$ contains an element of $\K'_{r+1}$ and when we arbitrarily select a vertex $v \in S$, the set $T' \cup (S - v)$ is $\mathcal{C}$-balanced. We let $F'$ be the set of these arbitrarily selected vertices. Recall that when $q = 2$, the set $S$ does not intersect $C_{p+2}$, so $T' \cup (S - v)$ is actually $\mathcal{C}$-well-balanced. We label $T' \cup S$ as $T_v$ and $S$ as $S_v$. By Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_std\]), we can assume that, for every $v \in F'$, the sets $T_v$ were constructed so as to be vertex-disjoint and, when $q = 2$, disjoint from $X \cup Y$. For every $v \in F'$, we add $T_v$ to $\T$. When $q = 2$, using Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_std\]), we find a set $T'$, disjoint from all of the previously constructed sets, such that $G[T' \cup X \cup Y]$ contains an element of $\K'_r$ and $\|T' \cap C_i\| = s_i$ for each $i \in [p]$, and we add $T' \cup X \cup Y$ to $\mathcal{T}$. We now let $$D_i := C_i \setminus F' \text{ for all $i \in [p+q]$},$$ and note that for every $i \in [p]$, if $c_i \ge 0$, then $\|D_i\| = s_i n/r$. Furthermore, when $q \ge 1$ and $\|C_{p+1} \cup C_{\sigma(p+1)}\| \ge s n/r$, we have that $\|D_{p+1} \cup D_{\sigma(p+1)}\| = s n/r$, and, when $q = 2$, we also have that both $\|D_{p+1}\|$ and $\|D_{\sigma(p+1)}\|$ are divisible by $s$. For each vertex $v \in \overline{C}$, we use Claim \[clm:extend\_overline\_C\](\[clm:extend\_overline\_C\_std\]) to find a vertex set $T_v$ that is $\mathcal{B}$-well-balanced and such that $G[T_v + v]$ contains a $\K'_{r+1}$. We add $T_v + v$ to $\T$ and ensure that these sets are disjoint from the sets already in $\T$. Let $F := F' \cup \overline{C}$ and recall that, for every $i \in [p]$ such that $\|C_i\| \le s_in/r$, we currently have that $D_i = C_i$. At this point, every vertex in $V$ is either in one of the sets $\{D_1, \dotsc, D_{p+q}\}$ or is in $F$. We now move vertices from $F$ to sets in $\mathcal{D}$ that are “too small” until we have the desired proper ordered collection. When we do this we also make small changes to the collection $\T$ so that every $T \in \T$ will be a $\mathcal{D}$-well-balanced $r$-set. In detail, for every $v \in F$ and $i \in [p+q]$, when we say we assign $v$ to $D_i$ we mean that we add $v$ to the set $D_i$ and update $T_v$ by removing one $u \in C_i$ from $T_v$ and add $v$ to $T_v$. This is only well-defined when there is initially some $u \in C_i$ in $T_v$. Note that in this case the updated version of $T_i$ is $\mathcal{D}$-well-balanced. If initially $T_v$ contains no element from $C_i$ then $q = 2$, $i \in \{p+1, p+2\}$, and $T_v$ intersects $C_{\sigma(i)}$ instead of $C_i$. Furthermore, since we are moving a vertex to either $C_{p+1}$ or $C_{p+2}$ from $F$, it must be that $\|C_{p+1} \cup C_{p+2}\| < sn/r$, and so if $v \in F'$ then $S_v \subseteq C_i$ for some $i \in [p]$ such that $c_i > 0$. In this case when we assign $v$ to $D_i$ we instead complete the following process: We first remove $T_v$ from $\T$. Then, if $v \in \overline{C}$, we use Claim \[clm:extend\_overline\_C\](\[clm:extend\_overline\_C\_special\]), to find a set $T' \subseteq B \setminus V(\T)$ such that $G[T '+ v]$ contains an element of $\K'_r$, $\|T' \cap B_j\| = s_j$ for every $j \in [p]$, $\|T' \cap B_i\| = s - 1$, $\|T' \cap B_{\sigma(i)}\| = 0$. Then let $T_v := T' + v$ and add it $\T$. Similarly, when $v \in F'$, we use Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_special\]) to find a set $T' \subseteq C \setminus V(\T)$ such that $G[T' \cup S_v]$ contains an element of $\K'_r$, $\|T' \cap C_k\| = 0$ where $k \in [p]$ such that $S_v \subseteq C_k$, $\|T' \cap C_j\| = s_j$ for every $j \in [p] - k$, $\|T' \cap C_i\| = s - 1$, $\|V(T') \cap C_{\sigma(i)}\| = 0$, and let $T_v := T' \cup S_v$ and add it $\T$. We then add $v$ to $D_i$. Note that now $T_v$ is $\D$-well-balanced. For any $i \in [p+q]$ such that $c_i < 0$, we then arbitrarily assign exactly $-c_i$ of the remaining vertices in $F$ to $D_i$, except when $q = 2$ and $i \in \{p+1, p+2\}$. When $q = 2$ and $\|C_{p+1} \cup C_{p+2}\| < sn/r$, we again have to be careful to ensure that, in the end, both $\|D_{p+1}\|$ and $\|D_{p+2}\|$ are divisible by $s$. Therefore, assume that, when $q = 2$, we assign vertices from $F$ to $D_{p+1}$ and $D_{p+2}$ before we assign vertices in $F$ to any $D_i$ for $i \le p$. Also, note that, because $\|D_{p+1} \cup D_{p+2}\| < sn/r$, $\|F\| \ge 1$. To help us organise the assignment of vertices, we let $j \in \{p+1, p+2\}$, so that if we let $1 \le k_j, k_{\sigma(j)}\le s$ be such that $$k_j \equiv \|D_j\| \text{ and } k_{\sigma(j)} \equiv \|D_{\sigma(j)}\| \pmod s,$$ then $k_{\sigma(j)} \le k_j$. If $\|F\| \ge s - k_j$, we assign $s - k_j$ vertices in $F$ to $D_{j}$ and then assign vertices from $F$ to $D_{\sigma(j)}$ until $\|D_{p+1} \cup D_{p+2}\| = sn/r$. Otherwise, we can assume $1 \le \|F\| < s - k_j$, which implies $s = 4$, and $\|F\| \le 2$. Note that if $\|F\| = 1$, then $\|F \cup D_{p+1} \cup D_{p+2}\| = 4n/r$, so $k_j + k_{\sigma(j)} = 3$. Therefore, exactly one of the two following conditions must hold: $$(i) \ \|F\| = 2 \text{ and } k_j = k_{\sigma(j)} = 1; \text{ or } (ii) \ \|F\| = 1, k_j = 2 \text{ and } k_{\sigma(j)} = 1.$$ In either case, we arbitrarily pick $v \in F$, add $v$ to $D_{\sigma(j)}$ and delete $T_v$ from $\T$. We then use Claim \[clm:extend\_overline\_C\](\[clm:extend\_overline\_C\_special\]) (if $v \in \overline{C}$) or Claim \[clm:extend\_partial\_clique\](\[clm:extend\_partial\_clique\_special\]) (if $v \in F'$), to construct a $\D$-balanced set $T_v$ containing $v$ that has one vertex in $C_{\sigma(j)}$ and two vertices in $C_{j}$. We then move both of the vertices in $T_v \cap D_{j}$ from $D_{j}$ to $D_{\sigma(j)}$. In both cases, we now have that $\|D_{\sigma(j)}\|$ is divisible by $4$, and $T_v$ is $\mathcal{D}$-well-balanced. We then assign the possibly one remaining vertex in $F$ to $D_{j}$, so $\|D_{j}\|$ is divisible by $4$ as well. For every $i \in [p+q]$, we let $D_i' = D_i \setminus V(\T)$, and note that $D'_i \subseteq C_i$. Therefore, if $G' = G - V(\T)$, then $\mathcal{D}' = (D'_1, \dotsc, D'_{p+q})$ is a proper ordered collection of $G'$, so, by Claim \[clm:proper\_partitions\] there exists a perfect $\K'_r$-tiling $\T'$ of $G'$, and $\T \cup \T'$ is a perfect $\K'_r$-tiling of $G$. Acknowledgements {#acknowledgements .unnumbered} ================ This research was partially carried out whilst the last three authors were visiting the Institute for Mathematics and its Applications at the University of Minnesota. The authors would like to thank the institute for the nice working environment. The last two authors were also supported by the BRIDGE strategic alliance between the University of Birmingham and the University of Illinois at Urbana-Champaign, as part of the ‘Building Bridges in Mathematics’ BRIDGE Seed Fund project. [10]{} N. Alon and A. Shapira, Testing subgraphs in directed graphs, J. Comput. System Sci. [69]{} (2004), 354–382. N. Alon and R. Yuster, Almost $H$-factors in dense graphs, Graphs Combin. [8]{} (1992), 95–102. F.S. Benevides, T. [Ł]{}uczak, A. Scott, J. Skokan and M. White, Monochromatic cycles in $2$-coloured graphs, Combin. Probab. Comput. [21]{} (2012), 57–87. J. Balogh, A. Lo and T. Molla, Transitive triangle tilings in oriented graphs, submitted. W.G. Brown and F. Harary, Extremal digraphs, Coll. Math. Soc. J. Bolyai [4]{} (1969), 135–198. K. Corrádi and A. Hajnal, On the maximal number of independent circuits in a graph, Acta Math. Acad. Sci. Hungar. [14]{} (1964), 423–439. A. Czygrinow, L. DeBiasio, H.A. Kierstead and T. Molla, An extension of the Hajnal–Szemerédi theorem to directed graphs, Combin. Probab. Comput. [24]{} (2015), 754–773. A. Czygrinow, H.A. Kierstead and T. Molla, On directed versions of the Corrádi–Hajnal Corollary, European J. Combin. [42]{} (2014), 1–14. C. Grosu and J. Hladký, The extremal function for partial bipartite tilings, European J. Combin. [33]{} (2012), 807–815. A. Hajnal and E. Szemerédi, Proof of a conjecture of Erdős, Combinatorial Theory and its Applications vol. II [4]{} (1970), 601–623. F. Havet, S. Thomassé, Oriented hamiltonian paths in tournaments: A proof of Rosenfeld’s conjecture, J. Combin. Theory B 78, (2000), 243–273. P. Hell and D.G. Kirkpatrick, On the complexity of general graph factor problems, SIAM J. Computing [12]{} (1983), 601–609. S. Janson, T. [Ł]{}uczak and A. Ruciński, [Random Graphs]{}, Wiley, 2000. P. Keevash and R. Mycroft, A multipartite Hajnal–Szemerédi theorem, J. Combin. Theory B [114]{} (2015), 187–236. P. Keevash and B. Sudakov, Triangle packings and 1-factors in oriented graphs, J. Combin. Theory B [99]{} (2009), 709–727. H.A. Kierstead and A.V. Kostochka, An Ore-type Theorem on Equitable Coloring, J. Combin. Theory B [[98]{}]{} (2008), 226–234. H.A. Kierstead and A.V. Kostochka, A short proof of the Hajnal–Szemerédi Theorem on Equitable Coloring, Combin. Probab. Comput. [[17]{}]{} (2008), 265–270. J. Komlós, Tiling Turán Theorems, Combinatorica [20]{} (2000), 203–218. J. Koml[ó]{}s, G. S[á]{}rk[ö]{}zy, and E. Szemer[é]{}di, [Proof of the Alon–Yuster conjecture]{}, Disc. Math. 235 (2001), 255–269. J. Koml[ó]{}s and M. Simonovits, Szemer[é]{}di’s [R]{}egularity [L]{}emma and its applications in graph theory. Bolyai Society Studies 2, Combinatorics, [P]{}aul [E]{}rdős is [E]{}ighty 2** (1996), 295–352. D. Kühn and D. Osthus, Critical chromatic number and the complexity of perfect packings in graphs, 17th ACM-SIAM Symposium on Discrete Algorithms (SODA 2006), 851–859. D. Kühn and D. Osthus, The minimum degree threshold for perfect graph packings, Combinatorica [29]{} (2009), 65–107. D. Kühn and D. Osthus, Embedding large subgraphs into dense graphs, in Surveys in Combinatorics (S. Huczynska, J.D. Mitchell and C.M. Roney-Dougal eds.), London Math. Soc. Lecture Notes 365, 137–167, Cambridge University Press, 2009. A. Lo and K. Markström, $F$-factors in hypergraphs via absorption, Graphs Combin. [31]{} (2015), 679–712. T. Molla, Tiling directed graphs with cycles and tournaments, Ph.D. thesis, Arizona State University, Tempe, Arizona, 2013. V. Rödl, A. Ruciński and E. Szemerédi, A Dirac-type theorem for 3-uniform hypergraphs, Combin. Probab. Comput. [15]{} (2006), 229–251. E. Szemerédi, Regular partitions of graphs, Problémes Combinatoires et Théorie des Graphes Colloques Internationaux CNRS [260]{} (1978), 399–401. A. Treglown, A note on some oriented graph embedding problems, J. Graph Theory [69]{} (2012), 330–336. A. Treglown, On directed versions of the Hajnal–Szemerédi theorem, Combin. Probab. Comput. [24]{} (2015), 873–928. A. Treglown, A degree sequence Hajnal-Szemerédi theorem, J. Combin. Theory B [118]{} (2016), 13–43. H. Wang, Independent directed triangles in a directed graph, Graphs Combin. [16]{} (2000), 453–462. H. Wang and D. Zhang, Disjoint directed quadrilaterals in a directed graph, J. Graph Theory [50]{} (2005), 91–104. R. Yuster, Tiling transitive tournaments and their blow-ups, Order [20]{} (2003), 121–133. Y. Zhao, Recent advances on Dirac-type problems for hypergraphs, Recent Trends in Combinatorics, the IMA Volumes in Mathematics and its Applications 159. Springer, New York, 2015. Vii 706. Andrzej Czygrinow School of Mathematics and Statistics Arizona State University Tempe, AZ 85281 US Louis DeBiasio Department of Mathematics Miami University Oxford, OH 45056 US Theodore Molla Department of Mathematics University of Illinois at Urbana-Champaign Urbana, IL 61801 US [ Andrew Treglown School of Mathematics University of Birmingham Edgbaston Birmingham B15 2TT UK ]{} [^1]: The second author is supported by Simons Foundation Grant \# 283194. The third author is supported by NSF Grant DMS-1500121. The fourth author is supported by EPSRC grant EP/M016641/1.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Die übliche Interpretation der physikalischen Zustände in Quantenfeldtheorien vom Typ der Quantenelektrodynamik bereitet Schwierigkeiten aufgrund des Auftretens masseloser Teilchen, wie dem Photon. In Stoßprozessen geladener Teilchen werden unendlich viele dieser masselosen Teilchen erzeugt, die zu einer enormen Vielfalt von Zuständen führen zwischen denen theoretisch keine Interferenzen möglich sind (sie gehören zu verschiedenen Superauswahlsektoren); experimentell lassen sich diese Zustände jedoch nicht unterscheiden. Dieser scheinbare Widerspruch löst sich auf wenn man berücksichtigt, dass reale Experimente stets in zukunftsgerichteten Gebieten der Raumzeit (Lichtkegeln) durchgeführt werden. Versäumte Messungen in der Vergangenheit lassen sich aufgrund der Richtung des Zeitpfeils nicht nachholen. Die Theorie kann sich daher auf die Beschreibung und Interpretation solcher Lichtkegeldaten beschränken. In diesem Artikel wird erläutert, wie dieser Gedanke im allgemeinen Rahmen der Quantenfeldtheorie mathematisiert wird und zu einem konsistenten physikalischen Bild führt. Zustände gleicher Ladung aber inkohärenter masseloser Anteile lassen sich bei Einschränkung auf die Observablen in einem gegebenen Lichtkegel zu Ladungsklassen zusammenfassen; jede dieser Klassen vereinigt eine Vielzahl von Superauswahlsektoren. Es zeigt sich, dass zu jeder Ladungsklasse eine konjugierte Klasse existiert, die Zustände der entgegengesetzten Ladung (Antimaterie) beschreibt. Alle Zustände in einer Klasse und der entsprechenden konjugierten Klasse genügen der gleichen (Bose oder Fermi) Statistik. Obwohl auf Lichtkegeln nur die Halbgruppe zukunftsgerichteter Zeittranslationen wirkt, lässt sich sowohl die relativistische Kovarianz als auch die energetische Stabilität (Positivität der Energie) der Zustände jeder Ladungsklasse etablieren. Der Formalismus ist somit geeignet, die Eigenschaften der Zustände in Quantenfeldtheorien mit langreichweitigen Kräften in physikalisch sinnvoller Weise zu beschreiben.' author: - \| Detlev Buchholz\ Institut für Theoretische Physik und Courant Forschungszentrum\ “Strukturen höherer Ordnung in der Mathematik”, Universität Göttingen,\ 37077 Göttingen, Deutschland title: \| Masselose Teilchen und Zeitpfeil\ in der relativistischen Quantenfeldtheorie --- Einleitung ========== Das Verständnis der Struktur des physikalischen Zustandsraumes in Quantenfeldtheorien mit langreichweitigen Kräften, prominentestes Beispiel ist die Quantenelektrodynamik, ist ein sehr altes Problem. Seine unterschiedlichen Aspekte wurden in zahlreichen Arbeiten untersucht, siehe zum Beispiel die entsprechenden Kapitel in den Monographien [@Ha; @St; @Str] und dort angegebene Referenzen. Allen diesen Untersuchungen ist gemeinsam, dass die physikalische Raumzeit als unendlich ausgedehnter Minkowskiraum modelliert wird, in dem die quantisierte Materie und die von masselosen Teilchen erzeugte Strahlung beschrieben wird (die Gravitation bleibt außer Acht). Geht man von diesem idealisierten Bild aus, so ergibt sich als Konsequenz, dass man beliebig langwellige Strahlung bzw. masselose Teilchen beliebig kleiner Energie behandeln muss, obwohl diese in realistischen Experimenten nicht nachgewiesen werden können. Dies führt sowohl bei der Konstruktion der Theorie als auch bei ihrer Interpretation zu Schwierigkeiten, die zuweilen als Infrarotkatastrophe bezeichnet werden. Man hat zwar gelernt, wie man diese Schwierigkeiten im Rahmen der Theorie bei einigen Fragestellungen umgehen kann, etwa bei der Berechnung von “inklusiven” Wirkungsquerschnitten von Stoßprozessen. Dort summiert man über unendlich viele unbeobachtbare [niederenergetische]{} masselose Teilchen, die bei der Wechselwirkung geladener Teilchen unvermeidlich erzeugt werden. Doch ist diese Methode, die aufgrund der Überidealisierung der experimentellen Situation eingehandelten Schwierigkeiten durch ad hoc Rezepte wieder zu beseitigen, konzeptionell nicht befriedigend. Zudem bleiben grundlegende Fragen, etwa nach der Sektorstruktur des physikalischen Zustandsraumes, nach den Wurzeln der Teilchenstatistik oder der Existenz von Antimaterie bei diesem Zugang unbeantwortet. Bei Abwesenheit von langreichweitigen Kräften konnten diese Fragen dagegen vor langem umfassend beantwortet werden [@Ha]. Ein ganz neuer Zugang zu diesem Problemkreis, bei dem die raumzeitlichen Einschränkungen bei realen Experimenten von vornherein im Rahmen der Theorie berücksichtigt werden, wurde in einer kürzlich erschienenen Arbeit vorgeschlagen [@BuRo]. Die gewonnen Einsichten sollen hier erläutert werden ohne auf technische Details einzugehen. Vorausgesetzt werden beim Leser lediglich Kenntnisse von Grundbegriffen der Quantentheorie und der speziellen Relativitätstheorie. Grundlagen ========== Es sei zunächst daran erinnert, dass die Ensembles eines physikalischen Systems durch normierte Vektoren (Zustandsvektoren) $\Phi$ eines Hilbertraumes $\cH$ beschrieben werden, die Messgrößen (Observablen) wirken auf diesem Raum durch lineare hermitesche Operatoren $A = A^$. Operatoren lassen sich addieren, per Komposition multiplizieren und mit komplexen Zahlen skalieren, sie erzeugen durch diese Operationen eine Algebra $\cA$. Die Theorie macht Vorhersagen über Mittelwerte der Messgrößen in vorgegebenen Ensembles, deren Schwankungsquadrate . Sie werden beschrieben durch Erwartungswertfunktionale $\varphi$ auf der Observablenalgebra $\cA$. Diese Funktionale erhält man aus den Hilbertraumzuständen $\Phi$ gemäß der fundamentalen Formel $$\varphi(A) = \langle \Phi, A \Phi \rangle \, , \quad A \in \cA \, ,$$ wobei, wie üblich, die eckige Klammer das Skalarprodukt zwischen Hilbertraumvektoren beschreibt. Es ist hier wichtig, dass man umgekehrt aus jedem solchen (linearen und positiven) Erwartungswertfunktional – wir sprechen im Folgenden kurz von Zuständen – einen Hilbertraum und eine konkrete Darstellung der Observablen auf diesem Raum rekonstruieren kann, so dass obige Relation gilt (sogenannte GNS–Konstruktion [@Ha]). Dies ist insbesondere dann von Bedeutung, wenn man Ensembles behandeln möchte, die nicht von vornherein durch Vektoren im vorgegebenen Hilbertraum beschrieben werden können (zum Beispiel ladungstragende Zustände wenn man von neutralen Zuständen ausgeht). Sie lassen sich durch geeignete Grenzwerte der zur Verfügung stehenden Funktionale darstellen. Man kann so den gegebenen Hilbertraum “verlassen” und zu einer konsistenten Beschreibung der gewünschten Ensembles unter Beibehaltung der Observablen kommen. Genauer gesagt, man erhält aus den Limesfunktionalen mittels der GNS–Konstruktion eine neue Darstellung der gegebenen Observablenalgebra auf einem Hilbertraum. In ihm gelten nach wie vor die ursprünglichen algebraischen Relationen zwischen den Observablen (man denke an Vertauschungsrelationen, Feldgleichungen ). In der relativistischen Quantenfeldtheorie benutzt man zweckmäßigerweise das Heisenbergbild. Das heißt, die Symmetrietransformationen im zugrundeliegenden Minkowskiraum $\cM$, wie die Verschiebung des Zeitpunktes und Ortes von Messungen oder Drehungen der Messgeräte und der Übergang von einem Inertialsystem zu einem anderen wirken auf die entsprechende Observablen in $\cA$. Für jede derartige Poincaré–Transformation $\lambda$ wird diese Wirkung beschrieben durch eine lineare, multiplikative und symmetrische Abbildung $\alpha_\lambda$ von $\cA$ auf sich (Automorphismus). Man hat ebenfalls Informationen über die raumzeitlichen Lokalisierungseigenschaften der Observablen. Messungen in einem gegebenen Raumzeitgebiet $\cR \subset \cM$ werden beschrieben durch Observablen in entsprechenden Unteralgebren $\cA(\cR) \subset \cA$. Da diese Messungen auch in jedem größeren Gebiet ausgeführt werden können, gilt aus Konsistenzgründen $\cA(\cR_1) \subset \cA(\cR_2)$ falls $\cR_1 \subset \cR_2$. Weiterhin ist die Wirkung der Symmetrietransformationen $\lambda$ in $\cM$ konsistent mit der Zuordnung von Observablen zu Gebieten, d.h. Observablen in $\cA(\cR)$ werden durch die Wirkung von $\alpha_\lambda$ in die Algebra $\cA(\lambda \cR)$ des transformierten Gebietes $\lambda \cR$ überführt, in Formeln $\alpha_\lambda \, \cA(\cR) = \cA(\lambda \cR)$. Das Einsteinsche Kausalitätsprinzip, also die Aussage, dass sich physikalische Effekte nicht mit Überlichtgeschwindigkeit ausbreiten (dies gilt auch für die Quanteneffekte von Messungen), lässt sich in diesem Rahmen ebenfalls in sehr einfacher Weise formulieren. Dazu sei daran erinnert, dass Messungen, die sich gegenseitig nicht beeinflussen, durch Observablen beschrieben werden, deren Produkt kommutativ ist. Gemäß Kausalitätsprinzip müssen daher Paare von Observablen in raumartig getrennten (nicht durch kausale Weltlinien verbindbaren) Gebieten $\cR_1, \cR_2$ miteinander kommutieren, d.h. $A_1 A_2 = A_2 A_1$ für alle $A_1 \in \cA(\cR_1)$. $A_2 \in \cA(\cR_2)$; man bezeichnet dies als Lokalitätsbedingung. Jede physikalische Theorie, die den Grundprinzipien der [Quantentheorie]{} und der speziellen Relativitätstheorie genügt, erfüllt diese Bedingungen. Es ist bemerkenswert, dass der so abgesteckte allgemeine Rahmen vollständig ausreicht, um die hier interessierenden konzeptionellen Fragen zu diskutieren. Geht man von der Vorstellung aus, dass der zugrunde gelegte Hilbertraum $\cH$ die Zustände sämtlicher elementarer Systeme einer Theorie beschreibt (z.B. alle Zustände endlicher Energie) so stellt sich die Frage nach dessen Sektorstruktur. Dabei trägt man der Tatsache Rechnung, dass in der Quantenfeldtheorie das Superpositionsprinzip nur eingeschränkt gilt: Zustandsvektoren unterschiedlicher Gesamtladung lassen sich zwar auf dem Papier addieren, ihre relative Phase spielt jedoch experimentell keine Rolle, sie sind nicht kohärenzfähig. Die Familien kohärenzfähiger Zuständen in $\cH$ nennt man (Superauswahl) Sektoren. Sie kann man dadurch charakterisieren, dass man von jedem gegebenen Zustand in einem Sektor zu jedem anderen Zustand im gleichen Sektor durch die Quanteneffekte von Messungen gelangen kann, die Observablen operieren transitiv auf den Sektoren. Verschiedene Sektoren lassen sich durch klassische Observablen unterscheiden, die man als Grenzwerte von geeigneten Folgen von Observablen in $\cA$ erhält. Sie vertauschen mit allen Elementen von $\cA$ und haben in jedem Sektor einen scharfen (nicht statistisch fluktuierenden) Wert, die entsprechende Superauswahlladung des Sektors. Prominente Beispiele solcher Superauswahlladungen sind die elektrische Ladung und die sogenannte Univalenz, mit der man bosonische und fermionische Sektoren unterscheiden kann. Es zeigt sich jedoch, dass es in Quantenfeldtheorien mit langreichweitigen Kräften, die durch masselose Teilchen vermittelt werden, eine ungeheure Vielzahl von weiteren Sektoren gibt. Sie unterscheiden sich durch facettenreiche Infrarotwolken unendlich vieler masseloser Teilchen, die bei Stoßprozessen geladener Teilchen erzeugt werden. Die theoretische Beschreibung und Klassifikation aller dieser Infrarotsektoren ist mathematisch ein hoffnungsloses Problem. Zum Glück spielt sie experimentell keine Rolle. Denn es ist praktisch unmöglich unendliche Konfigurationen masseloser Teilchen mit beliebig kleiner Energie scharf zu unterscheiden. Bei Anwendungen der Theorie trägt man dieser Tatsache meist dadurch Rechnung, dass man ad hoc* spezielle Sektoren auswählt und in Erwartungswerten über alle unbeobachtbaren masselosen Teilchen summiert. Dadurch kann man zwar eine Reihe von Schwierigkeiten umgehen, die Methode ist jedoch aus zwei Gründen konzeptionell unbefriedigend. Zum einen erfordert die Unterscheidung zwischen hoch– und niederenergetischen Teilchen die Wahl eines Inertialsystems, man bricht damit die relativistische Invarianz der Theorie. Zum anderen zerstören die Quanteneffekte der lokalisierten Observablen diese Unterscheidung wegen der Unschärferelation. Jedes Ensemble niederenergetischer Teilchen enthält nach einer lokalen Messung auch Anteile beliebig hoher Energie. Diese Tatsache erschwert die Diskussion von Konsequenzen des Einsteinschen Kausalitätsprinzips bei diesem Zugang. Die kürzliche Lösung dieses Problems beruht auf der Einsicht, dass der Zeitpfeil bereits bei der Interpretation der mikroskopischen Theorie berücksichtigt werden muss. Um Missverständnissen vorzubeugen: der Zeitpfeil soll hier nicht erklärt werden, er wird lediglich als empirische in den theoretischen Rahmen eingebaut. Es ist ein Faktum, dass es unmöglich ist, versäumte Messungen in der Vergangenheit heute oder in der Zukunft nachzuholen. Man kann also nicht ausschließen, dass gewisse Daten, die man in der Vergangenheit hätte bestimmen können, unwiederbringlich verloren sind. Anders gesagt, die Theorie kann sich darauf beschränken, die Ergebnisse von Messungen in Raumzeitgebieten zu beschreiben und zu erklären, die in realistischen Experimenten (zumindest prinzipiell) zugänglich sind. Dies sind vorwärts gerichtete Lichtkegel $V$ mit Spitze in einem beliebig gewählten Raumzeitpunkt $a$ in der Vergangenheit; der Rand des Kegels wird von den von der Spitze ausgehenden Lichtstrahlen gebildet. Wo die Spitze des Kegels liegt spielt praktisch keine Rolle. Man kann z.B. den Geburtstag und Geburtsort von Aristoteles wählen, der das Wort Physik ($\phi \upsilon \sigma \iota \kappa \eta$) eingeführt hat; wir alle wissen davon und befinden uns also innerhalb des entsprechenden Kegels. Eine andere Wahl wäre z.B. der Tag und Ort, an dem die Finanzierung eines geplanten Experiments genehmigt wurde. Wichtig ist nur, dass in der Vergangenheit des gewählten Punktes $a$ keine experimentellen Daten berücksichtigt werden müssen bzw. können. Natürlich erstrecken sich reale Experimente nur über einen bestimmten Zeitraum. Doch könnten im Prinzip zukünftige Generationen ein Experiment bis in alle Ewigkeit weiterführen. Lichtkegel sind also die maximalen Raumzeitgebiete in denen Experimente möglich sind. Im Rahmen der Theorie werden die Observablen in einem gegebenen Lichtkegel $V$ durch Elemente der entsprechenden Unteralgebra $\cA(V) \subset \cA$ beschrieben. Geht man davon aus, dass nur diese Observablen in einem Experiment zur Verfügung stehen, so kann man bei gegebenem Zustand $\varphi$ eines Systems nur die entsprechenden Erwartungswerte $\varphi(A)$, $A \in \cA(V)$ bestimmen. Man erhält also nur partielle Informationen über den globalen Zustand $\varphi$ und bezeichnet dessen Einschränkung $\varphi \rest \cA(V)$ daher als partiellen Zustand. Es stellt sich dann die Frage, welche Information man aus solchen partiellen Zuständen extrahieren kann. Bei genauerer Analyse zeigt sich, dass in Theorien, in denen es ausschließlich massive Teilchen gibt (also in einer hypothetischen Welt), die partiellen Zustände bereits die vollständige Information über den globalen Zustand enthalten, egal wie man den Lichtkegel $V$ wählt, d.h. wann und wo man mit den Messungen begonnen hat. Dies kann man auch heuristisch verstehen wenn man bedenkt, dass die Weltlinien massiver Teilchen, die sich ja mit weniger als Lichtgeschwindigkeit bewegen, irgendwann einmal in den Lichtkegel eintreten müssen. Aus den Daten in $V$ lassen sich daher die globalen Daten mit Hilfe der Theorie rekonstruieren. Sobald masselose Teilchen in der Theorie auftreten (“Es werde Licht”) ändert sich die Situation jedoch drastisch. Werden in der Vergangenheit des gegebenen Lichtkegels $V$ durch Wechselwirkungsprozesse masselose Teilchen erzeugt, die danach ungehindert durch die Raumzeit propagieren, so können diese $V$ nicht erreichen und dort nachgewiesen werden, denn sie bewegen sich stets mit Lichtgeschwindigkeit (Huygensches Prinzip). Aus den Zuständen $\varphi \rest \cA(V)$ lassen sich diese auslaufenden masselosen Anteile daher nicht rekonstruieren. Man erhält somit in Experimenten nur eingeschränkte Informationen über die Beiträge masseloser Teilchen in den Zuständen. Wie wir noch sehen werden, ist diese Tatsache dafür verantwortlich, dass man die unterschiedlichen Infrarotwolken in den partiellen Zuständen nicht scharf unterscheiden kann. Die Zustände in verschiedenen Infrarotsektoren stimmen bei Restriktion auf Lichtkegel überein und man kann auf diese Weise Klassen von physikalisch äquivalenten (weil ununterscheidbaren) Zuständen bilden. Von Bedeutung ist dabei, dass die masselosen Teilchen in den Zuständen nicht mehr in einen energetisch weichen und einen harten Anteil aufgeteilt werden (mit den erwähnten Problemen). Stattdessen wird der masselose Teilcheninhalt getrennt in einen marginalen (nicht notwendig niederenergetischen) Teil, der in $V$ keine Beiträge liefert und einen essentiellen Teil, der Messungen in $V$ zugänglich ist. Diese Aufspaltung ist mit der relativistischen Invarianz der Theorie und dem Einsteinschen Kausalitätsprinzip verträglich und bildet den Ausgangspunkt der Untersuchungen in [@BuRo], die wir im Folgenden erläuteren wollen. Das Vakuum ========== Sowohl bei der Konstruktion von Theorien als auch bei deren Interpretation spielt der Zustand des Vakuums eine ausgezeichnete Rolle. Dies liegt daran, dass man eine ganze Reihe von Eigenschaften dieses speziellen Zustandes nennen kann, ohne die Theorie konstruiert zu haben: Der Vakuumzustand ist für alle inertialen Beobachter gleich, das heißt Messergebnisse in diesem Zustand ändern sich nicht, wenn man die Positionen der Messapparate durch Poincaré–Transformationen ändert. Ferner ist das Vakuum energielos, d.h. für alle inertialen Beobachter ist es der Zustand niedrigster Energie (Grundzustand). Da wir nur Messungen in Lichtkegeln betrachten wollen, müssen wir bei der mathematischen Umsetzung dieser Charakterisierung beachten, dass die von den Poincaré–Transformationen $\lambda$ erzeugte Poincaré–Gruppe $\cP$ keine Symmetriegruppe der Kegel ist; denn kein Lichtkegel $V$ ist invariant unter beliebigen derartigen Transformationen. Die Untergruppe $\cL$ der Drehungen und Geschwindigkeitstransformationen bei festgehaltener Spitze des Kegels (Lorentz–Gruppe) lässt $V$ zwar invariant, doch ändert jede raumzeitliche Translation die Lage von $V$ im Minkowski–Raum. Es gibt jedoch Translationen, bei denen man $V$ zumindest nicht verlässt. Dies sind zukunftsgerichtete zeitartige (oder auch lichtartige) Verschiebungen. Inertiale Beobachter in $V$ müssen wenig tun, um derartige Verschiebungen ihrer Messgeräte zu bewerkstelligen: sie müssen nur warten, der Zeitpfeil tut das Übrige. Die Familie der Poincaré–Transformationen, die $V$ auf bzw. in sich abbilden, erzeugt eine Semigruppe $\cS \subset \cP$ (“semi” deshalb, weil die Elemente von $\cS$ sich zwar komponieren lassen, es gibt jedoch kein Inverses, da sich der Zeitpfeil nicht umkehren lässt). Nach diesen Vorbemerkungen können wir nun obige formlose Beschreibung des Vakuums in unserem Formalismus präzisieren. Ein partieller Zustand $\omega \rest \cA(V)$ wird als Vakuum in $V$ gedeutet wenn $$\omega(\alpha_\lambda(A)) = \omega(A) \quad \mbox{für alle} \quad A \in \cA(V), \ \lambda \in \cS \, .$$ Man ergänzt diese Relation noch durch einige physikalisch sinnvolle Bedingungen an die Natur der Korrelationen zwischen Paaren von Observablen in diesem Zustand, siehe [@BuRo]. Mit Hilfe der weiter oben erwähnten GNS–Konstruktion kann man dann einen Hilbertraum $\cH$ konstruieren, auf dem die Observablen in $\cA(V)$ operieren, sowie einen Zustandsvektor $\Omega \in \cH$ angeben, so dass $\omega(A) = \langle \Omega, A \Omega \rangle$ für alle $A \in \cA(V)$. Die Vektoren in $\cH$ beschreiben die durch Quanteneffekte von Messungen erzeugten lokalen Anregungen des Vakuums. Der interessante Punkt ist nun, dass die oben erwähnten charakteristischen Eigenschaften des partiellen Vakuumzustandes bereits ausreichen, um auf $\cH$ eine Darstellung der vollen Poincaré Gruppe $\cP$ durch unitäre Operatoren $U(\lambda)$, $\lambda \in \cP$ zu konstruieren [@BuRo]. Für ihre Einschränkung auf $\cS$ gelten die Relationen $$U(\lambda) A U(\lambda)^{-1} = \alpha_\lambda(A) \, , \quad U(\lambda) \Omega = \Omega \, , \quad \lambda \in \cS \, .$$ Man kann also auch im Fall der auf Lichtkegel eingeschränkten Messungen alle Symmetrietransformationen durch unitäre Operatoren beschreiben, die sich aus den entsprechenden in Weise rekonstruieren lassen. Darüber hinaus haben die Generatoren der resultierenden Raumzeittranslationen, die wie üblich als Energie–Impulsoperatoren gedeutet werden, die erwarteten spektralen Eigenschaften (Positivität der Energie für alle inertialen Beobachter). Sie lassen sich jedoch nicht in der üblichen Weise als Observablen deuten, da in ihnen die Fluktuationen des Energieinhalts in $V$ infolge von in der Vergangenheit abgestrahlten masselosen Teilchen nur pauschal berücksichtigt werden. Die Situation ist ähnlich wie in der Quantenstatistischen Mechanik, wo die Unkenntnis der Mikrozustände thermischer Systeme in die Liouvilleoperatoren statistisch eingeht. Ein charakteristisches Energie–Impulsspektrum, wie man es zum Beispiel für die lokalen Anregungen des Vakuums in der erwartet, ist in Figur \[spektrum\] dargestellt. Eingezeichnet sind die Zustände, die zu den jeweiligen Energie–Impulswerten beitragen: Die Spitze des Kegels entspricht dem Vakuum, der Rand den Energie–Impulswerten eines einzelnen Photons. Die Spektralwerte in der Nähe des Randes gehören zu Streuzuständen von Photonen (Delbrück–Streuung). Photonen genügend hoher Energie können bei Stoßprozessen Elektron–Positron Paare erzeugen, . Ladungsklassen ============== Auch wenn das Vakuum für die Konstruktion und Interpretation der Theorie von Bedeutung ist, so sind doch vor allem seine lokalen Anregungen und die daraus resultierenden ladungstragenden Zustände von physikalischem Interesse. Bevor wir uns der Konstruktion und Analyse dieser Zustände zuwenden, müssen wir zunächst besser verstehen, weshalb die bei der Interpretation der Theorie im Minkowskiraum auftretenden Infrarotprobleme bei unserem Zugang verschwinden. Gemäß der Maxwellschen Elektrodynamik erzeugen beschleunigte Punktladungen Strahlung. Im Rahmen der Quantenfeldtheorie kann man diese Strahlung durch kohärente Zustände von Photonen beschreiben. Es stellt sich bei genauerer Analyse heraus, dass für Punktladungen mit verschiedenem einlaufenden und auslaufenden Impuls $p_{\mbox{\tiny \it ein}} \neq p_{\mbox{\tiny \it aus}}$ diese Zustände mit Sicherheit unendlich viele niederenergetische Photonen enthalten (sie können nicht durch Vektorzustände im Fockraum der Photonen beschrieben werden). Man spricht daher, wie bereits mehrfach erwähnt, von Infrarotwolken. Es zeigt sich ferner, dass diese Infrarotwolken von den präzisen Werten der asymptotischen Impulse abhängen. Es gibt daher eine ungeheure Vielfalt von solchen global unterscheidbaren (im Sinne von Superauswahlsektoren) Infrarotwolken. Doch gibt es einen interessanten Spezialfall: Stimmen die beiden asymptotischen Impulse der Punktladung exakt überein, so erhält man kohärente Zustände im Fockraum der Photonen, d.h. mit endlicher Teilchenzahl. Diese Tatsache bleibt meist unbeachtet, da es ja extrem unwahrscheinlich ist, dass in realen Prozessen ein solches Ereignis auftritt. Doch spielt gerade dieser Spezialfall bei unseren Überlegungen eine wichtige Rolle. Stellt man sich nämlich die Frage, ob sich die Photonenwolken der beiden in Fig. \[stroeme\] angedeuteten Prozesse durch Messungen in einem Lichtkegel $V$ scharf lassen, so verdeutlicht Fig. \[fock\] dass dies nicht so sein sollte. Denn durch Messung an den Photonen in $V$ wird sich nicht mit Sicherheit feststellen lassen, welche einlaufenden Impulse zwei Punktladungen mit dem gleichen auslaufenden Impuls in ferner Vergangenheit einmal hatten; durch Wechselwirkungsprozesse in der und dabei abgestrahlte marginale Photonen wird diese in den Superauswahlsektoren enthaltene globale Information völlig verwischt. Die Messergebnisse werden daher konsistent sein mit denen, die man in Zuständen mit übereinstimmenden einlaufenden und auslaufenden Impulsen erhalten würde. Anders gesagt, globale Infrarotwolken sollten sich bei Messungen in $V$ nicht von Zuständen im Fockraum der Photonen unterscheiden lassen. Tatsächlich lässt sich dies im Rahmen der Quantenfeldtheorie beweisen [@Bu]. Während die Infrarotsektoren sich also bei Messungen in Lichtkegeln nicht unterscheiden lassen, kann der Gesamtwert der von massiven Teilchen in einem Zustand getragenen Ladungen in Lichtkegeln ermittelt werden. Dies veranschaulicht Fig. \[ladung\], in der die Weltlinien von massiven geladenen Teilchen eingezeichnet sind, die sich natürlich alle mit weniger als Lichtgeschwindigkeit bewegen. Massive Teilchen entgegengesetzter Ladung können in Stoßprozessen vernichtet werden (wobei z.B. hochenergetische Photonen entstehen) oder aber durch solche Prozesse in Paaren entgegengesetzter Ladung erzeugt werden. Diese Ereignisse ändern die eines Zustands nicht. Man kann daher aus der Analyse des massiven Teilcheninhalts in die von diesen Teilchen getragene Gesamtladung bestimmen. Diese wichtige Einsicht lässt sich im Rahmen der Quantenfeldtheorie präzise formulieren und begründen [@Bu]. In Analogie zu den Superauswahlsektoren im Minkowskiraum $\cM$ kann man daher im Rahmen der Theorie die Zustände mit gleicher Gesamtladung und auf gegebenem Lichtkegel $V$ ununterscheidbaren Infrarotwolken in Ladungsklassen einteilen. Dabei besteht jede Ladungsklasse aus Zuständen, deren Einschränkungen auf $\cA(V)$ durch die Quanteneffekte physikalischer Operationen in $V$ ineinander überführt werden können. Wir erinnern daran, dass Operationen in der Quantentheorie durch die Wirkung von unitären Operatoren auf Zustandsvektoren beschrieben werden: Stört man die Dynamik des Systems für einige Zeit durch einen in $V$ lokalisierten Operator, so kann man den Effekt dieser Störung im Wechselwirkungsbild berechnen. Nach Abschaltung der Störung erhält man einen unitären Operator $W \in \cA(V)$, der die Änderung der Zustände aufgrund dieser Operation beschreibt. Dies führt zu folgender Definition, die wir hier in etwas vereinfachter Form wiedergeben und auch so benutzen werden. Die präzise Formulierung findet sich in [@BuRo]. *Ladungsklassen:* Zwei reine (ungemischte) Zustände $\varphi_1$, $\varphi_2$ auf der Observablenalgebra $\cA$ gehören zur gleichen Ladungsklasse, wenn es zu jedem gegebenem Lichtkegel $V$ einen entsprechenden unitären Operator $W \in \cA(V)$ gibt, so dass $$\varphi_2 \rest \cA(V) = \varphi_1 {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\mbox{Ad} \, W \rest \cA(V) \, ,$$ d.h. die partiellen Zustände werden durch die adjungierte Wirkung $\mbox{Ad} \, W = W \cdot W^{-1}$ von $W$ auf $\cA(V)$ ineinander überführt. Das Symbol ${\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}$ bezeichnet die Komposition von Abbildungen. Ladungen und Morphismen ======================= Geht man, so wie bei der Konstruktion von Quantenfeldtheorien üblich, vom Vakuumzustand aus, so kann man mit Hilfe der vorangegangenen Definition alle Zustände in der Ladungsklasse des Vakuums charakterisieren, also alle neutralen Zustände, die keine globale Ladung tragen. Es stellt sich dann die Frage, wie man im Rahmen der Theorie aus diesen neutralen Zuständen geladene Zustände erhält. Auch hier ergibt sich die Antwort aus physikalische Überlegungen. Wir wählen im Folgenden einen festen Lichtkegel $V$. Stellt man sich vor, dass inertiale Beobachter, die in der Spitze des Kegels mit unterschiedlichen Geschwindigkeiten in verschiedene Richtungen gestartet sind, alle zur gleichen Eigenzeit den Raumzeitpunkt, den sie erreicht haben, markieren, so erhält man einen Hyperboloid in $V$, den wir als Zeitschale bezeichnen, siehe Fig.\[zeitschale\]. Die asymptotisch lichtartigen Punkte auf einer solchen Zeitschale, die man nur mit Lichtgeschwindigkeit erreichen könnte, liegen raumartig zu allen kompakten Gebieten im Inneren von $V$. Durch lokale Operationen mit genügend großem Energieübertrag kann man nun auf einer gegebenen Zeitschale aus dem Vakuum Paare von Teilchen mit entgegengesetzter Ladung erzeugen (Vakuumpolarisation), daneben entstehen Wolken von masselosen Teilchen. Um aus solchen global neutralen Paaren einen für Beobachter in $V$ effektiv ladungstragenden zu präparieren, muss die unerwünschte entgegengesetzte Ladung auf Lichtgeschwindigkeit gebracht werden. Sie befindet sich dann am asymptotisch lichtartigen Rand der Zeitschale und verschwindet dadurch aus dem kausalen Einflussgebiet von Beobachtern im Inneren von $V$. Dieser Limes würde in der Praxis unendlich viel Energie erfordern, die von der kompensierenden Ladung weggetragen wird. Man betrachtet hier also, wie häufig in der Theoretischen Physik, eine Idealisierung, da sie es gestattet, die Eigenschaften der Ladungen in Reinkultur zu analysieren. Wie in Fig. \[ladungen\] angedeutet, kann die Erzeugung von Paaren in kompakten Gebieten erfolgen, die Erzeugung der geladenen Limeszustände erfordert dagegen unendlich ausgedehnte Gebiete, die sich zur Begrenzung der von den masselosen Teilchen abgestrahlten Energie asymptotisch weiten müssen. Sie haben daher die Form hyperbolischer Kegel $\rm K$. Es ist hier wichtig, dass die Richtung und der Öffnungswinkel dieser Kegel bei der Präparation von Zuständen mit gegebener Ladung frei gewählt werden kann. Wir bezeichnen die kausalen Abhängigkeitsgebiete (Abschlüsse) der Kegel $\rm K$ mit kalligrafischen Buchstaben $\cK$ und nennen sie kurz Hyperkegel. Nach diesen Vorbemerkungen können wir nun beschreiben, wie man in der Theorie, ausgehend vom Vakuumzustand $\omega$, geladene Zustände erhält. Dazu gibt man sich zunächst einen Hyperkegel $\cK \subset V$ vor, in dem die Präparation erfolgen soll. Dann wählt man Folgen von in $\cK$ lokalisierten unitären Operatoren, $\{W_n \in \cA(\cK)\}_{n \in \NN}$, die die Operation der Paarerzeugung beschreiben, wobei für wachsendes $n$ die kompensierende Ladung immer weiter in asymptotisch lichtartiger Richtung verschoben wird. Bei geeigneter Wahl der Operatorfolge konvergieren die resultierenden partiellen Zuständen in der Ladungsklasse des Vakuums, $\{ \omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\mbox{Ad} \, W_n \rest \cA(V) \}_{n \in \NN}$, gegen einen partiellen Zustand $\varphi \rest \cA(V)$ aus der gewünschten Ladungsklasse. Die kompensierende Ladung ist in $\varphi \rest \cA(V)$ nicht mehr nachweisbar, sie befindet sich im raumartigen Komplement der in $V$ zur Verfügung stehenden Observablen. Ausgehend von dieser physikalisch transparenten aber umständlichen Methode kann man zu einer bequemeren, aber etwas abstrakteren Beschreibung der Ladungserzeugung übergehen. Es zeigt sich nämlich, dass auch die Folgen der von den unitären Operatoren induzierten Abbildungen $\{ \mbox{Ad} \, W_n \}_{n \in \NN}$ der Observablenalgebra $\cA(V)$ auf sich selbst konvergieren, wobei wir hier nicht auf die in [@BuRo] diskutierten mathematischen Details eingehen können. Für die Grenzwerte führen wir die Notation ein $$\sigma_\cK = \lim_{n \rightarrow \infty} \, \mbox{Ad} \, W_n \, .$$ Der Index $\cK$ erinnert daran, dass $\sigma_\cK$ Grenzwert von Operationen im Hyperkegel $\cK$ ist. Die so definierten Abbildungen $\sigma_\cK$ sind, wie ihre Approximationen, verträglich mit der algebraischen Struktur von $\cA(V)$: sie sind linear, multiplikativ und symmetrisch; der Einfachheit halber nehmen wir ferner an, dass ihr Bildbereich ebenfalls in $\cA(V)$ liegt. Wir bezeichnen diese Abbildungen kurz als (in $\cK$ lokalisierte) Morphismen. Die in der Monographie [@Ha] erläuterten tiefen Resultate über die Sektorstruktur von Quantenfeldtheorien massiver Teilchen basieren auf der Analyse solcher Morphismen. Auch im vorliegenden Kontext von Theorien mit masselosen Teilchen erweisen sie sich als nützliches analytisches Instrument. Denn in ihnen sind in mathematisch präziser Weise alle physikalischen Informationen enthalten, die wir in den vorangegangenen Abschnitten zusammengetragen haben. Wir fassen sie in der folgenden Liste zusammen, in der einige technische Details etwas vereinfacht dargestellt sind. (Die präzise Formulierung findet sich in [@BuRo].) *Lokalisierte Morphismen:* Zu gegebener Ladungsklasse und beliebigem Hyperkegel $\cK \subset V$ gibt es einen auf $\cA(V)$ definierten lokalisierten Morphismus $\sigma_\cK$ mit den folgenden Eigenschaften. - $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_\cK \rest \cA(V)$ stimmt mit einem partiellen Zustand aus der gegebenen Ladungsklasse überein. - Für jedes Raumzeitgebiet $\cR \subset V$, das $\cK$ enthält, gilt $\sigma_\cK(\cA(\cR)) \subseteq \cA(\cR)$. In den Fällen wo man Gleichheit hat, bezeichnet man die Ladungsklasse und die entsprechenden Morphismen als “einfach”. - Für jedes Raumzeitgebiet $\cR \subset V$, das raumartig von $\cK$ getrennt ist, gilt $\sigma_\cK \rest \cA(\cR) = \iota$, worin $\iota$ die identische (triviale) Abbildung bezeichnet. - Zu jedem Paar von Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$, die mit der Ladungsklasse affiliiert sind, existieren unitäre Operatoren $W_{2 1} \in \cA(V)$, die die Morphismen verknüpfen, . Man nennt die Operatoren $W_{2 1}$ Verknüpfungsoperatoren. Punkt (a) dieser Liste wurde bereits ausführlich erläutert. Die Punkte (b) und (c) folgen aus den Eigenschaften der approximierenden Operatoren $W_n$. Im Fall von (b) gilt $W_n \in \cA(\cK) \subset \cA(\cR)$ und somit $\mbox{Ad} \, W_n (\cA(\cR)) \subseteq \cA(\cR)$ und im Fall (c) gilt aufgrund der Vertauschbarkeit von in $\cR$ und $\cK$ lokalisierten Operatoren (Lokalitätsprinzip) $\mbox{Ad} \, W_n \rest \cA(\cR) = \iota$. Diese Eigenschaften übertragen sich auf den Limes $\sigma_\cK$. In Punkt (d) kommt die Tatsache zum Ausdruck, dass die Superauswahlsektoren der bei der Erzeugung von Ladungen unvermeidlich entstehenden Infrarotwolken bei Messungen in $V$ nicht unterschieden werden können. Die Zustände $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} \rest \cA(V)$ und $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2} \rest \cA(V)$ gehören daher zur gleichen Ladungsklasse und können durch die adjungierte Wirkung unitärer Operatoren in $\cA(V)$ ineinander transformiert werden. Diese Operatoren verknüpfen auch die lokalisierten Morphismen. Wir werden uns im Folgenden auf die Diskussion der in (b) charakterisierten einfachen Ladungsklassen bzw. Morphismen beschränken. Zu ihnen gehört die uns hier besonders interessierende elektrische Ladung und die Univalenz. Ladungskonjugation und Statistik ================================ Mit Hilfe der lokalisierten Morphismen lässt sich die Struktur der Ladungsklassen in der Quantenfeldtheorie in systematischer Weise analysieren und klassifizieren. Dazu muss man voraussetzen, dass die Theorie alle in Hyperkegeln $\cK$ prinzipiell möglichen Messungen und Operationen beschreibt, d.h. dass sich die Algebren $\cA(\cK)$ nicht durch Hinzufügen weiterer Operatoren vergrößern lassen, ohne mit dem Lokalitätsprinzip in Konflikt zu geraten. Sie sind also in diesem Sinne maximal. Diese physikalisch motivierte Forderung an die Theorie lässt sich mathematisch in Form einer Dualitätsbedingung ausdrücken [@BuRo]. Wie bereits angedeutet, werden wir uns im Folgenden auf die Diskussion der einfachen Ladungsklassen beschränken, ohne dies jedes Mal zu erwähnen. Der erste und entscheidende Schritt bei der Analyse ist der Beweis, dass sich Ladungsklassen komponieren lassen. Anschaulich gesprochen entspricht diese Komposition der Addition des Ladungsinhaltes von Zuständen. Genauer gesagt, man kann zeigen [@BuRo], dass für gegebene Ladungsklassen $L_1$, $L_2$ von Zuständen und Lichtkegel $V$ die jeweils zu diesen Klassen assoziierten, in Hyperkegeln lokalisierten Morphismen $\sigma_{\cK_1}$, $\tau_{\cK_2}$ komponiert werden können, $\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Die resultierenden partiellen Zustände $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} \rest \cA(V)$ gehören unabhängig von der Wahl von $\cK_1$, $\cK_2$ alle zu einer bestimmten Ladungsklasse $L_3$, die sich ebenfalls durch lokalisierte Morphismen $\rho_{\cK_3}$ beschreiben lässt. Insbesondere gibt es also unitäre Operatoren $W_{3 2 1} \in \cA(V)$, die diese Morphismen verknüpfen, $\rho_{\cK_3} = \mbox{Ad} \, W_{3 2 1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Dies Resultat zeigt, dass bei der Komposition von Ladungen keine neuartigen Typen von Infrarotwolken entstehen, deren Superauswahlsektoren sich in $V$ unterscheiden lassen. Bei unserer heuristischen Diskussion der Ladungserzeugung waren wir von der Existenz von entgegengesetzten Ladungen ausgegangen, die sich gegenseitig neutralisieren. Es zeigt sich, dass auch diese empirische Tatsache bereits in der Struktur der lokalisierten Morphismen kodiert ist [@BuRo]: Zu jeder Ladungsklasse $L$ von Zuständen und gegebenem Lichtkegel $V$ mit dazu assoziierten in Hyperkegeln lokalisierten Morphismen $\sigma_\cK$ existiert eine konjugierte Ladungsklasse $\overline{L}$ mit entsprechend lokalisierten Morphismen $\overline{\sigma}_\cK$ für die gilt $\sigma_\cK {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\overline{\sigma}_\cK = \overline{\sigma}_\cK {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_\cK = \iota$. Das heißt, alle Effekte der Ladungen in $L$ lassen sich durch Komposition mit bestimmten Konfigurationen der konjugierten Ladungen aus $\overline{L}$ vollständig neutralisieren und umgekehrt. Die mit $\overline{L}$ affiliierten Zustände werden daher als zu $L$ ladungskonjugierte Antimaterie interpretiert. Eine weitere fundamentale Eigenschaft von Ladungen, nämlich die Tatsache, dass sie bei Vertauschung in Zuständen (je nach Ladungstyp) entweder der Bose– oder der Fermistatistik genügen, lässt sich ebenfalls aus den Eigenschaften der lokalisierten Morphismen herleiten [@BuRo]. Dazu betrachtet man bei gegebener Ladungsklasse $L$ und Lichtkegel $V$ zwei assoziierte, beliebig lokalisierte Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$ und vergleicht die resultierenden komponierten Morphismen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2}$ bzw. $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Wie oben erläutert, gehören diese komponierten Morphismen zur gleichen Ladungsklasse (mit verdoppelter Ladung) und sind durch unitäre Verknüpfungsoperatoren miteinander verbunden. Tatsächlich gibt es eine kanonische Wahl für diesen Verknüpfungsoperator, der in einem feldtheorischen Rahmen als gruppentheoretischer Kommutator von ladungstragenden (und daher nicht observablen) unitären Operatoren gedeutet werde könnte. Um diesen Sachverhalt zu beleuchten, nehmen wir für einen Moment an, dass es ladungstragende, in $\cK_j$ lokalisierte unitäre Operatoren $V_j$ gibt so dass $\sigma_{\cK_j} = \mbox{Ad} \, V_j$, $j = 1,2$. Der Operator $W_{2 1} = V_2 V_1^{-1}$ wäre dann ein Verknüpfungsoperator zwischen $\sigma_{\cK_1}$ und $\sigma_{\cK_2}$ im Sinne der oben angegebenen Definition und der gruppentheoretische Kommutator $V_2 V_1 V_2^{-1} V_1^{-1}$ wäre ein Verknüpfungsoperator zwischen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2} = \mbox{Ad} \, V_1 V_2$ und $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} = \mbox{Ad} \, V_2 V_1$. Der interessante Punkt ist nun, dass sich dieser gruppentheoretische Kommutator auch ohne Kenntnis von ladungstragenden Operatoren berechnen lässt gemäß $$V_2 V_1 V_2^{-1} V_1^{-1} = (V_2 V_1^{-1}) V_1 ( V_1 V_2^{-1}) V_1^{-1} = W_{2 1} \, \mbox{Ad} \, V_1 (W_{2 1}^{-1}) = W_{2 1} \sigma_{\cK_1}(W_{2 1}^{-1}) \, .$$ Auf der rechten Seite dieser Gleichung steht eine Größe, die in unserem nur auf Observablen aufgebauten Formalismus bestimmt werden kann. Diese Überlegungen legen es nahe, für die gegebenen Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$ und entsprechendem Verknüpfungsoperator $W_{2 1} \in \cA(V)$ den Operator $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2}) = W_{2 1} \, \sigma_{\cK_1}(W_{2 1}^{-1})$ einzuführen. Man stellt fest, dass $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2})$ nicht von der Wahl von $W_{2 1}$ abhängt sowie die komponierten Morphismen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2}$ und $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$ verknüpft. In diesem Sinne ist der Operator kanonisch. Der physikalisch bedeutsame Punkt ist nun, dass für Morphismen, die in raumartig getrennten Hyperkegeln $\cK_1, \cK_2$ lokalisiert sind, siehe Fig. \[statistik\], der Operator $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2})$ eine besonders einfache Form hat. Man kann nämlich zeigen [@BuRo], dass es zu jeder Ladungsklasse $L$ eine Zahl $\varepsilon_L \in \{\pm 1 \}$ (genannt Statistikparameter) gibt, so dass $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2}) = \varepsilon_L 1$ für alle mit der Klasse assoziierten, in raumartig getrennten Hyperkegeln lokalisierten Morphismen $\sigma_{\cK_1}, \sigma_{\cK_2}$. Die physikalische Bedeutung dieses Ergebnisses wird klar, wenn man es durch die oben betrachteten ladungstragenden Operatoren $V_j$, $j=1,2$ reformuliert: Falls $\varepsilon_L = 1$ müssen diese Operatoren bei raumartigen Abständen miteinander kommutieren (bosonischer Fall), falls $\varepsilon_L = -1$ müssen sie antikommutieren (fermionischer Fall). Jede Ladungsklasse $L$ genügt also entweder der Bose oder der Fermi Statistik. Andere Möglichkeiten gibt es für die hier betrachteten einfachen Ladungen nicht. Man kann ferner zeigen [@BuRo], dass die Statistik der konjugierten Klasse $\overline{L}$ stets mit der von $L$ übereinstimmt, $\varepsilon_{\overline{L}} = \varepsilon_L^{ }$. Es sei betont, dass diese Ergebnisse ganz wesentlich von der Tatsache abhängen, dass die physikalische Raumzeit vierdimensional ist. In hypothetischen Welten niedrigerer Dimension würde die gleiche Analyse ergeben, dass Ladungsklassen auch einer anderen (z.B. anyonischen) Statistik genügen können. Man gelangt somit, wie im massiven Fall [@Ha], auch in Theorien mit masselosen Teilchen zu einem vertieften Verständnis der Wurzeln der Teilchenstatistik. Kovarianz und Spektrum ====================== Ein weiterer physikalisch wichtiger Punkt ist die Frage, wie man im vorliegend Rahmen die relativistische Invarianz der Ladungsklassen und die energetischen Eigenschaften der entsprechenden Zustände beschreibt. Wie bereits erwähnt, wirkt auf einen gegebenen Lichtkegel $V$ in natürlicher Weise nur die Semigruppe $\cS \subset \cP$ der Lorentztransformationen und zukunftsgerichteten Translationen, die $V$ auf bzw. in sich abbilden. Wir müssen daher zunächst diskutieren, wie man in physikalisch sinnvoller Weise den Transport von lokalisierten Morphismen $\sigma$ . (Da wir im folgenden einen festen Morphismus betrachten spielt dessen Lokalisationsgebiet keine Rolle.) Zur Lösung dieses Problems gehen wir von einigen physikalisch motivierten Annahmen aus. Für eine gegebene Transformation $\lambda \in \cS$ sollte es im Prinzip möglich sein, die durch $\sigma$ beschriebene Ladungserzeugung in $V$ in exakt der gleichen Weise auf dem transformierten Lichtkegel zu bewerkstelligen (Wiederholbarkeit des Experiments). Es sollte also Morphismen ${}^\lambda \sigma$ auf $\cA(V)$ geben für die gilt ${}^\lambda \sigma {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\alpha_\lambda = \alpha_\lambda {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma$, wobei daran erinnert sei, dass . Der Ladungsinhalt ändert sich beim Transport nicht, die Morphismen ${}^\lambda \sigma$ gehören somit alle zur gleichen Ladungsklasse und es gibt unitäre Verknüpfungsoperatoren $W_\lambda \in \cA(V)$, so dass $\mbox{Ad} \, W_\lambda {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}{}^\lambda \sigma = \sigma$, $\lambda \in \cS$. Die relativistische Invarianz der Theorie kommt dann dadurch zum Ausdruck, dass die durch $W_\lambda$ beschriebenen physikalischen Operationen auf allen transportierten Lichtkegeln $\mu V$, $\mu \in \cS$ in analoger Weise operieren, d.h. $\alpha_\mu(W_\lambda)$ ist auch Verknüpfungsoperator zwischen den transportierten Morphismen ${}^{\mu \lambda} \sigma$ und ${}^{\mu} \sigma$ für $\lambda, \mu \in \cS$. (Hier benutzt man die Tatsache dass $\cS$ eine Semigruppe ist, dass also das dem sukzessiven Transport entsprechende Produkt erklärt ist.) Wir fassen diese physikalisch erwarteten Eigenschaften in folgender Definition zusammen. *Kovariante Morphismen:* Ein Morphismus $\sigma$ auf $\cA(V)$ heißt kovariant, wenn es eine dazu assoziierte Familie von transportierten Morphismen ${}^\lambda \sigma$, $\lambda \in \cS$ gibt mit den oben angegebenen Eigenschaften. Es ist bemerkenswert, dass diese zwar physikalisch gut begründete aber implizite Charakterisierung der kovarianten Morphismen ausreicht, um die gewohnte Formulierung der relativistischen Kovarianz zu etablieren [@BuRo]. Ausgehend von der durch das Vakuum fixierten unitären Darstellung $U$ der vollen Poincaré Gruppe $\cP$ kann man nämlich zeigen, dass die gemäß $U_\sigma(\lambda) = U(\lambda) W_\lambda$, $\lambda \in \cS$ definierten unitären Operatoren die Transformation der Observablen in der Ladungsklasse von $\sigma$ implementieren, $$\mbox{Ad} \, U_\sigma(\lambda) {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma = \sigma {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\alpha_\lambda \, , \quad \lambda \in \cS \, .$$ Darüber hinaus existiert stets eine Wahl der Verknüpfungsoperatoren $W_\lambda$, so dass sich die Operatoren $U_\sigma(\lambda)$, $\lambda \in \cS$ eindeutig zu einer unitären Darstellung von $\cP$ (bzw. der Überlagerungsgruppe von $\cP$) fortsetzen lassen. Beobachter in $V$ können somit in eindeutiger Weise über den Energieinhalt und Drehimpuls (Spin) von Zuständen in der gegebenen Ladungsklasse sprechen. Diese ergeben sich aus den Generatoren der Darstellung $U_\sigma$, die jedoch nicht direkt als Observablen werden können. Observablen im üblichen Sinne sind die unitären (und damit im Sinne der Spektraltheorie normalen) Verknüpfungsoperatoren $W_\lambda = U(\lambda)^{-1} U_\sigma(\lambda) \in \cA(V)$, $\lambda \in \cS$ aus denen man z.B. Informationen über den Energieunterschied zwischen Zuständen in der Ladungsklasse des Vakuums und der geladenen Klasse erhält. Mit der getroffenen Wahl des Vakuums als Referenzzustand und der sich daraus eindeutig ergebenden Darstellung $U$ von $\cP$ kann man diese Informationen dann mithilfe der Darstellungen $U_\sigma$ in gewohnter Weise beschreiben. Ein weiteres Resultat, das die physikalische Konsistenz unseres theoretischen Zugangs untermauert, betrifft die Form des Energie–Impuls Spektrums in den Ladungsklassen. Durch eine detaillierte Analyse des sich bei der Komposition von konjugierten Ladungsklassen ergebenden Spektrums lässt sich zeigen [@BuRo], dass die Generatoren der raumzeitlichen Verschiebungen $U_\sigma \rest \RR^4$ in allen Ladungsklassen die relativistische Spektrumsbedingung erfüllen (Positivität der Energie in allen Inertialsystemen). Diese energetische Stabilität der Ladungsklassen ist also eine Konsequenz der Struktur der Observablen und des Vakuums und muss nicht extra postuliert werden. Fazit ===== Die vorangegangene Diskussion hat verdeutlicht, dass die notorischen Infrarotprobleme bei der Interpretation von Quantenfeldtheorien mit langreichweitigen Kräften auf einer Überidealisierung beruhen: Nimmt man im Rahmen der Theorie wie gewöhnlich an, dass der masselose Teilcheninhalt der Zustände durch Messungen im gesamten Minkowskiraum bestimmt werden kann, so ist man gezwungen, eine Vielzahl unendlicher Infrarotwolken masseloser Teilchen in Betracht zu ziehen. Berücksichtigt man jedoch die Existenz des Zeitpfeils, so ist offensichtlich, dass sich Messungen bestenfalls in zukunftsgerichteten Lichtkegeln durchführen lassen. Versäumte Messungen in der Vergangenheit bedeuten im Allgemeinen einen unwiederbringlichen Verlust an Informationen, und die Theorie kann sich daher auf die Beschreibung und Interpretation von Lichtkegeldaten beschränken. Wie wir gesehen haben, führt diese Beschränkung zu einer sowohl mit den Symmetrien des Minkowskiraums als auch dessen kausaler Struktur verträglichen Infrarotregularisierung. Das aus der Einsteinschen Kausalitätsforderung ableitbare Huygenschen Prinzips bewirkt die geometrische Aufteilung des globalen masselosen Teilcheninhalts in einen in Lichtkegeln experimentell zugänglichen essentiellen und einen in der Vergangenheit abgestrahlten und daher für alle Zeiten verlorenen marginalen Anteil. Angesichts dieser Fakten wird im Rahmen der Theorie verständlich, weshalb sich globale Infrarotwolken nicht scharf unterscheiden lassen. Auch die Frage, welche physikalischen Aspekte einer Theorie in Lichtkegeln bestimmt werden können, konnte beantwortet werden. Dies sind zum einen die Eigenschaften massiver Teilchen, insbesondere deren Ladungen und Statistik, sowie die Existenz und Eigenschaften von Antimaterie. Darüber hinaus lassen sich aus den auf Lichtkegel eingeschränkten partiellen Zuständen auch deren energetische und rotatorischen Eigenschaften bestimmen. Materie kann also auch in diesem neuen theoretischen Zugang detailliert beschrieben werden. Der in Lichtkegeln nachweisbare masselose Teilcheninhalt lässt sich durch Zustände mit endlicher Teilchenzahl (also im Fockraum) beschreiben. Es sei betont, dass dies auch masselose Teilchen einschließt, die den Lichtkegel aus fernen Galaxien erreichen. Diese Teilchen wurden zwar vor Äonen erzeugt, doch kommen sie nicht aus der Vergangenheit des Beobachters, sondern aus zum Beobachter zuvor raumartig getrennt liegenden Gebieten. Stellt man sich nun vor, dass durch diese Teilchen Informationen über das Ergebnis von Experimenten unserer Kollegen übermittelt werden, so kann man sich die Frage stellen, ob diese Informationen im Widerspruch zu den in hiesigen Lichtkegeln gewonnenen physikalischen Einsichten stehen könnten. Die Antwort ist beruhigend: Auch die extragalaktischen Kollegen können ihre Experimente bestenfalls in Lichtkegeln durchführen und man kann zeigen [@BuRo], dass die Vorhersagen der Theorie nicht von der Wahl eines Lichtkegels abhängen. Somit ist die innere Konsistenz unserer zwar ungewohnten aber dafür problemfreien Interpretation der relativistischen Quantenfeldtheorie gesichert. Danksagung {#danksagung .unnumbered} ========== Mein Dank gilt dem Evangelischen Studienwerk, besonders Herrn Prof. Dr. Eberhard Müller, für die langjährige fruchtbare Zusammenarbeit und stete Förderung des Promotionsschwerpunktes “Wechselwirkung”. [99]{} D. Buchholz, The physical state space of quantum electrodynamics.\ Commun. Math. Phys. [85]{} (1982) 49 D. Buchholz und J. E. Roberts, New light on infrared problems: Sectors, statistics, symmetries and spectrum.\ e–print arXiv:1304.2794 R. Haag, Local Quantum Physics: Fields: Particles, Algebras.\ Berlin, Springer (1992) O. Steinmann, Perturbative Quantum Electrodynamics and Axiomatic Field Theory.\ Berlin Springer (2000) F. Strocchi, An Introduction to Non-Perturbative Foundations of Quantum Field Theory.\ Oxford University Press (2013)	{ "pile_set_name": "ArXiv" }
--- author: - Gerhard Hensler title: Morphological Mutations of Dwarf Galaxies --- 2[/pc\^2]{} Introduction {#intro} ============ By the continuous growth of telescope size and advanced detector sensitivity the panchromatic view of galaxies is enabling us since the HST time to trace the evolution of massive galaxies observationally back to high redshifts. As examples the existence of intact gas-rich galaxy disks around redshift 2 has provided us a new insight into the gas accumulation and causes for the high star-formation rates (SFRs). Although dwarf galaxies (DGs) also exist already at that early epoch, but because of their faintness, those observations are not as feasible for them so that our wisdom of DG formation and evolution depends on assumptions from numerical simulations and from their comparison with stellar population studies of DGs in the local universe. Nevertheless, due to the improved observational facilities also for the DGs, details of their properties have affected our picture of their formation and evolution. The first impression from decades ago, that DGs possess simple structures and evolve morphologically clearly separated, has changed totally in the sense that a classical morphological division of them is meaningless in the view of the variety of DG types: there are e.g. dwarf irregular galaxies (dIrrs) with exceedingly strong star formation (SF), called starburst DGs (SBDGs), and also short but intense epochs of SF in the past, dwarf elliptical galaxies (dEs) with recent SF or central gas content, and last but not least, dwarf spheroidals (dSphs) at the faint end of dEs as satellite galaxies down to about -5$^m$. ”Normal” DGs have a brightness range between M$_V \ge -18^m$ to -10$^m$. For this brightness reason, dSphs are only detectible within the Local Group by refined search algorithms from surveys as e.g. SDSS. Also in the Virgo Cluster an archival work of detailed dE properties is expensive in observing time. In their studies of Virgo cluster DGs already Sandage & Biggeli [@sb84] found that dEs dominate the cluster galaxy population by far, in contrast to their number fraction in the field where dIrrs are the most common DGs. This fact cannot be interpreted from the different local origins of DGs but because matter accumulates to clusters also dIrrs fall in from the cosmic web continuously whereby they have to change their morphology. Not only because of such morphological mutation but also due to the occurrence of enhanced SF in dIrrs, [@sb84] emphasized already the necessity of various links between the DG types by morphological transitions. From the $\Lambda$CDM cosmology the baryonic matter should settle within Dark Matter (DM) halos, which originally preferred to form low-mass subhalos and hierarchically accumulate to massive galaxies. If the baryonic matter would follow this bottom-up structure formation, the subhalos should also assemble their gas at first and by this also evolve with SF to become the oldest galactic objects in the universe. That this picture seems to be too naive is simply understandable by three major physical principles: 1\. The gas assembly timescale should behave as the free-fall timescale $\tff$, namely, dependent on the gas density as $\rg^{-1/2}$, because gas is accreted through gravitation. Whether this accretion leads to the same enhancement of SF in DGs as observed and theoretically expected [@kho09] is still a matter of debate (see also sect. 2). Because the virial temperature in the DG gravitational potential does not accomplish values above 10$^5$ K, on the one hand, cold accretion [@dek09] is not necessarily required as for massive halos. 2\. The SF timescale $\tSF$ is defined as M$_g/\Psi$ with $\Psi$ as the SFR that, on the other hand, in the self-regulated SF mode depends on $\rg^2$ [@koe95]. Let me already emphasize here, that the theoretical treatment and modelling of SF self-regulation has to allow for the stellar energy of radiation and winds by massive stars already released in SF regions during their lives, i.e. prior to the explosion of supernovae typeII (SNeII). This necessity becomes clear when one continues a high SFR conditioned by gas infall and unaffected for a few million years until the first SNeII emerge. Since lower galaxy masses lead to less dense gas, SF is stretched over time for DGs. And 3. As SF couples to stellar energy release, and since the counteracting cooling process depends on $\rg^2$, the gas expands due to pressure support and reduces the SFR so that the effect of SF self-regulates non-linearly. Another important effect that seems to affect the whole network of galaxy formation and evolution is ionizing radiation from the first cosmological objects (supermassive stars, black holes, galaxies). Due to the re-ionization of the gas in the universe, its thermodynamical state is changed so that its accretion onto low-mass objects was reduced [@dij04] and gas already caught in minihalos evaporated again [@bar99]. Since massive objects remained almost unaffected by the re-ionization phase, while DGs should have experienced delayed SF [@noe07], this evolutionary dichotomy is observed as downsizing [@cow96]. Nevertheless, the assumption that all DGs were affected in the re-ionization era and in the same way would request overlapping Stroemgren bubbles in an almost uniformly ionized Universe. This, however, must be questioned and is contrasted by the existence and amplification of cosmological density structures [@par10]. Another possible paths of DG origin is the formation of SF density enhancements in the tidal tails of merger galaxies [@duc07]. Since they should be free of DM and it is not yet well understood whether their SF acts in self-regulation the survival probability of these galaxy-like entities needs to be explored [@rec07b]. Dwarf irregular galaxies and their extreme evolutionary stages {#dIrrs} =============================================================== dIrrs are characterized by large gas fractions, ongoing SF, and low metallicities. That dIrrs contain the same or slightly higher gas fractions than giant spiral galaxies and mostly suffer the same SF efficiency, but appear with lower metallicity $Z$ than spirals, cannot be explained by simple evolutionary models. When gas is consumed by astration but replenished partly by metal-enhanced stellar mass loss, the general analytical derivation relates the element enrichment $Z(t)$ with the logarithm of decreasing remaining gas fraction $\mu = M_{\rm g}(t)/M_{\rm g}(0)$ as $Z(t) = y\, [-ln(\mu)]$, where $y$ as the slope is determined by the stellar yield (see e.g. textbooks like [@pag10] or reviews as e.g. by [@pra08a; @hen10]). As demonstrated by [@gar02] and [@zee01], however, the effective yields of gas-rich galaxies decrease to smaller galaxy masses. This means that their element abundances, particularly O measured in regions, are smaller than those released by a stellar population and confined in a “closed box”. Two processes can reduce the metal abundances in the presence of old stellar populations: loss of metal-enriched gas by galactic outflows or infall of metal-poor (even pristine) intergalactic gas (IGM). It is widely believed, that a fundamental role in the chemical evolution of dIrrs is played by galactic winds, because freshly produced metals in energetic events are carried out from a shallow potential well of DGs through a wind (which will be therefore metal-enhanced). Some SBDGs are in fact characterized by galactic winds [@mar95] or by large expanding supernova type II (SNeII)-driven X-ray plumes (e.g. [@hen98; @mar02]). Studies have raised doubts to whether the expanding $\Ha$ loops, arcs, and shells mostly engulfing X-ray plumes, really imply gas expulsion from the galaxies because their velocities are mostly close to escape, but adiabatic expansion against external gas tends to hamper this. As an extreme, [@bab92] speculated that galactic winds are able to empty DGs from its fuel for subsequent SF and, by this, transform a gas-rich dIrr to a fading gas-poor system. In order to manifest this scenario and to study mass and abundance losses through galactic winds numerous numerical models are performed under various, but mostly uncertain conditions and with several simplifications (e.g. [@mlf99; @str04]). The frequently cited set of models by MacLow & Ferrara [@mlf99] (rotationally supported, isothermal disks of dIrrs with fixed structural relations for four different gas masses between M$_{\rm g} = 10^6 - 10^9 \Msun$ and three different SNII luminosities in the center corresponding to SN rates of one per $3\times 10^4$ yrs to 3 Myrs) is mostly misinterpreted: The hot gas is extremely collimated from the center along the polar axis, but cannot sweep-up sufficient surrounding ISM to produce significant galactic mass loss. On the other hand, the loss of freshly released elements from massive stars is extremely high. Moreover, these models lack of realistic physical conditions, as e.g. the existence of an external pressure, self-consistent SFRs, a multi-phase inhomogeneous ISM, and further more. -- -- -- -- Also more detailed numerical simulations [@db99; @rec06a], show that galactic winds are not very effective in removing gas from a galaxy. Although galactic winds develop vertically, while the horizontal transport along the disk is very limited, their efficiency depends very sensitively on the galaxy structure and ISM properties, as e.g. on the disk shape [@rec09]. Fig. \[fig:wind\] reveals clearly that the more eccentric the disk is, the more pronounced does the superbubble expand. On the one hand, the hot SN gas has to act against the galactic ISM, exciting turbulence and mixing between the metal-rich hot gas with the surrounding . Not taken into account in present-day models is the porosity of the ISM, consisting of clouds and diffuse less dense gas. In particular, the presence of clouds can hamper the development of galactic winds through their evaporation. This so-called mass loading reduces the wind momentum and internal energy. Since the metallicity in those clouds are presumably lower than the hot SNII gas, also the abundances in the outflow are diminished as e.g. observed in the galactic X-ray outflow of NGC 1569 [@mar02] for which a mass-loading factor of 10 is derived to reduce the metallicity to 1-2 times solar. In recent simulations [@sca10] demonstrate that turbulent mixing can effectively drive a galactic wind. Although they stated that their models lead to a complex, chaotic distribution of bubbles, loops and filaments as observed in NGC 1569, other observational facts have not been compared. Detailed numerical simulations of the chemical evolution of these SBDG by [@rec06b] could simultaneously reproduce both, the oxygen abundance in the warm gas as well as the metallicity in the hot outflow. Surprisingly, [@rec07a] demonstrated that the leakage of metals from a SBDG is not prevented by the presence of clouds because the clouds pierce holes into the wind shells. This leads to a final metallicity a few tenths of dex lower than in models without clouds. Consequently, the crucial question must be answered which physical processes trigger such enormous SFRs as observed in SBDGs and consume all the gas content within much less than the Hubble time. One possibility which has been favoured until almost two decades ago was that at least some of these objects are forming stars nowadays for the very first time. Today it is evident that the most all metal-poor ones (like I Zw 18) contain stars at least 1 Gyr old [@mom05], and most SBDGs have several Gyrs old stellar populations. This means that SF in the past should have proceeded in dIrrs, albeit at a low intensity and long lasting, what can at best explain their chemical characteristics, like for instance the low \[$\alpha$/Fe\] ratio [@lm04]. The \[$\alpha$/Fe\] vs. \[Fe/H\] behaviour is representative of the different production phases, $\alpha$-elements from the short-living massive stars and 2/3 of iron from type Ia SNe of longer-living binary systems. If the SF duration in a galaxy is very short, type Ia SNe do not have sufficient time to enhance the ISM with Fe and most of the stars will be overabundant in \[$\alpha$/Fe\]. [lr]{} & \[fig:KS\] In most SBDGs large reservoirs enveloping the luminous galactic body have been detected (NGC 1569 [@sti02], NGC 1705 [@meu98], NGC 4449 [@hun98], NGC 5253 [@kob08], I Zw 18 [@zee98c], II Zw 40 [@zee98b]) with clearly disturbed gas kinematics and disjunct from the galactic body. Nevertheless, in not more than two objects, NGC 1569 [@mue05] and NGC 5253 [@kob08] gas infall is proven, while for the other cases the gas kinematics obtrudes that the gas reservoir feeds the engulfed DGs. In another object, He 2-10, the direct collision with an intergalactic gas cloud [@kob95] is obviously triggering a huge SB. Reasonably, for their measurable surface density the SFRs of most of these objects exceed those of ”normal” gas disks (Fig.\[fig:KS\]). Yet it is not clear, what happens to dIrrs if they move into a region with increasing external pressure as e.g. by means of a denser IGM and of ram pressure when they fall into galaxy clusters. In sect.\[dEs\] we will discuss the effect of ram pressure on the structure of the ISM for which numerical models for spiral galaxies (e.g. [@roe05]) as well as for gas-rich DGs (e.g. [@mor00]) exist, but only hints from observations (as e.g. for the Magellanic stream). The effect on the SFR due to compression of the ISM is observed, but not yet fully understood and convincingly studied by models. [@cor06] e.g. observed a coherent enhancement of SF in group galaxies falling into a cluster. ![ The abundance ratio N/O as a function of oxygen abundance observed in spiral and irregular galaxies (shaded area, after [@zee98a]) overlayed with evolutionary loops due to infall of primordial intergalactic gas clouds. These have different mass fractions $M_{cl}/M_{SF}$ with respect to the mass involved in the SF region The crosses represent evolutionary timesteps of models, the arrows depict the direction of the evolutionary paths. The dashed straight line represents a simple model relation for purely secondary nitrogen production. [For discussion see [@koe05].]{} []{data-label="fig:NO"}](NO_OH_loop.pdf){width="10cm"} Although the mass-metallicity relation also holds for dIrrs and even steepens its slope [@tre04], what can be interpreted by galactic mass loss and the corresponding lower effective yield [@gar02], the abundance ratios are unusual. As mentioned above, O/Fe reaches already solar values for subsolar oxygen or iron abundances. While this can be explained by a long SF timescale, another characteristic signature is that the ratio stays at about -1.6 to -1.5 with O abundances below 1/10 solar and with an increasing scatter with O (see Fig.\[fig:NO\]). Their regime of N/O–O/H values overlaps with those of regions in the outermost disk parts of spirals at about = 8.0 ... 8.5 [@zee98a]. In the 90th several authors have tried to model these observations by SF variations with gas loss through galactic winds under the assumptions that these dIrrs and blue compact DGs (BCDs) are young and experience their first epochs of SF (for a detailed review see [@hens99]). Stellar population studies contradict this youth hypothesis, so that another process must be invoked. Since these objects are embedded into envelops and are suggested to suffer gas infall as manifested e.g. in NGC 1569 (see above, [@sti02; @mue05]), the influence of metal-poor gas infall into an old galaxy with continuous SF on particular abundance patterns were exploited by Koeppen & Hensler [@koe05]. With the reasonable assumption that the fraction of infalling-to-existing gas mass increases with decreasing galaxy mass, their results could match not only the observational regime of BCDs in the \[\]-space but also explain the shark-fin shape of observational data distribution [@koe05]. Fig.\[fig:NO\] demonstrates how self-enriched galaxies which have reached the secondary nitrogen-track already within 2-3 Gyrs of their evolution are thrown back in O abundance by gas infall while N/O stays the same. After a time delay depending on the mass fraction of infalling gas to that existing already within the SF site, along a loop-like evolutionary paths in the \[\]-diagram the ISM abundances reach again the starting point. In summary, one can state that old dIrrs [mutate]{} temporarily to youngly appearing examples with respect to their gas abundances. Dwarf elliptical galaxies {#dEs} ========================= dEs are frequently denoted as examples of “stellar fossile” systems in which the bulk of their SF occurred in the past. They are preferentially located in morphologically evolved environments [@tre02], i.e. in regions with high galaxy densities and dominate the morphological types of galaxies in clusters, as e.g. Virgo, Coma, Fornax, and Perseus. Furthermore, dEs cluster strongly around luminous elliptical/S0 galaxies [@tul08]. The evolution of this galaxy type should be mainly caused by gas and tidal effects on SF and structure and indicates that it is strongly affected by environment. Already [@bot85] found that cluster dEs are usually almost free of interstellar gas and contain few young stars. In trying to understand the dE population, structural regularities and correlations must be studied, as it is known since the 80th, between optical surface brightness and luminosity [@kor85; @fer94] and between luminosity and stellar velocity dispersion which also correlates with metallicity (e.g. [@pet93]). Boselli et al. [@bos08b] proposed to understand these ”Kormendy” relations by processes having transformed dIrrs after their cluster infall, but accuse the still existing lack of numerical simulations of such transformation. Furthermore, dEs often have flattened profiles but are mostly kinematically supported by their stellar velocity dispersion rather than by rotation. The combination of low gas-mass fractions and moderate-to-low stellar metallicities in dE (about 0.1 of solar or less) is a key feature of this class. Their lower stellar abundances [@haa97] suggest that extensive gas loss occurred during their evolution and SF ceased due to a lack of raw material rather than exhaustion of the gas supply through SF. Galactic winds are therefore a hallmark of modern models for dE evolution, starting from the basic consideration by [@lar74] and continued with the study by [@ds86]. They are commonly assumed to have cleaned out DGs soon after their formation. As mentioned in sect. \[dIrrs\], however, gas expulsion by means of galactic winds is inefficient from our understanding of the multi-phase ISM and requires even in low-mass systems a DM-to-baryonic matter ratio [@mlf99] much smaller than assigned to DGs in the classical formation picture (e.g. [@mat98]). There are two competing scenarios for the origin of cluster dEs. On the one hand, those low-mass galaxies are believed to constitute the building blocks in $\Lambda$CDM cosmology and should therefore have evolved congruently with the mass accumulation to more massive entities, galaxies and galaxy clusters. For those, orbiting in a cluster the stellar component must be heated continuously by harrasment of more massive cluster galaxies and thus be pressure supported. On the other hand, a variety of observations are available which also support discrepant scenarios of dEs evolution. Recent studies of Virgo cluster dEs [@con03] and also those of the Fornax cluster (see e.g. [@mic04]) have unveiled that a small but significant fraction of them contains gas, has experienced recent SF, and can be argued from internal kinematics and cluster distribution data to represent an infalling class of different types of gas-rich galaxies in or after the state of morphological transformation. Further findings of a significant fraction of rotationally supported dEs in the Virgo cluster [@zee04] and also disk features as e.g. spiral arms and bars [@lis06] support the possibility of morphological transformation from gas-rich progenitor DGs to dEs thru gas exhaustion. Boselli et al. [@bos08a] have comprehensively discussed the different processes of dE origin. A separation into dE subclasses with respect to their origin should also be visible in an intermediate-age stellar population, blue centers, and flatter figure shape. Indeed, dEs in the Virgo cluster can be divided into different subclasses [@lis07] which differ significantly in their morphology and clustering properties, however, do not show any central clustering, but are distributed more like the late-type galaxies. These types of dEs show different disk signatures, such as bars and spiral structures, are not spheroidal, but rather thick disk-like galaxies. Similar shapes were also found for the brighter, non-nucleated dEs. There is only a small fraction of nucleated dEs whithout any disk features or cores, which keep the image of spheroidal objects consisting of old stars. [lr]{} & \[fig:VCdE\] A figure analysis of Virgo dEs correlates with the averaged orbit velocity in the sense that flatter dEs show on average a larger orbital velocity (700 km/s) than those originating within the cluster (300 km/s) [@lis09] (see Fig. 4). This kinematical dichotomy is expected because galaxies formed in virial equilibrium within the cluster retain their initial kinetic energy while the cluster mass grew. Galaxies falling into the present cluster potential must therefore possess larger velocities. To obtain information about both evolutionary stages, the young infalling vs. the late cluster members, [@got09] studied SDSS data. The basic model is that dIrrs which are formed outside the Virgo Cluster and becoming stripped on their infall, by this being transformed into dEs, should reveal properties recognizably different from dEs which have already aged in the cluster, as e.g. colors, effecitve radius, radial stellar distribution, and abundances. One result by [@got09] is that for the two dE populations, with and without cores, distinguished by their Sersic parameter, there is only a slight indication that non-nucleated dEs are more concentrated towards the inner cluster regions, whereas the fraction of nuclated dE is randomly distributed, while [@lis07] found it to increase with distance. An analysis of the relation between the central surface brightness and the Sersic parameter shows the expected tendency to higher values for brighter galaxies. Furthermore, there were no further relations found of the Sersic parameter, the effective radius, or the distance from M 87. Deeper insights are provided by spectra. Koleva et al. [@kol09] found most dEs in the Fornax cluster to be roundish and to contain significant metallicity gradients already in the old stellar population. They argue that this is due to a lack of sufficient mixing. In contrast, rotationally supported dEs have flat metallicity distributions. Compared with simulations this can be attributed to galactic winds, but the picture of metallicity and gradients is not yet clear. While [@spo09] show a tight positive correlation between the total metallicity \[Z/H\] and the mass, [@kol09] do not find any trend involving \[Fe/H\] for Fornax-cluster and nearby-group dEs. Moreover, from the deconvolution of the SF history of their sample dEs with respect to the central 1 arcmin and within the effective radius [@kol09] draw the conclusions that for a few objects SF episodes occurred in the very center even within the last 1 Gyr. From a systematic study of the central Fornax-cluster dEs’ dynamics [@rij10] conclude that these objects stem from an infall population of late-type DGs and has been transformed to dEs by ram-pressure stripping (RPS). Toloba et al. [@tolo09] derive for Coma cluster dEs to be weaker in carbon than dEs in low-density environments, while similar in nitrogen. Actually, they [@tol11] also find that pressure supported Virgo dEs show higher dynamical mass-to-light ratios than rotationally supported dEs of similar luminosity and further that dEs in the outer parts of the cluster are mostly rotationally supported with disky shapes. Rotationally supported dEs even follow the Tully-Fisher relation. One fundamental and most spectacular result [@tol11] is, however, that dEs are not DM-dominated galaxies, at least up to the half-light radius. Correlations of both signatures, SF history and metallicity gradients, for cluster-member dEs vs. infall dEs should be derived for more clusters, but observations are unfortunately very time-expensive if possible at all. In addition to classical dEs, ultra-compact DGs (UCDs) have been detected and classified as a new type of cluster dEs that differ by their intrinsic structure and brightness (see e.g. [@phi01; @gre09]). The origin of these peculiar DGs is mysterious and not yet understood but requests transformation if they are surviving nuclei of tidally stripped nucleated DGs [@goe08]. Morphological transitions ========================= As the transformation picture from late-type dwarfs into dEs is still not completely understood and only qualitative, in sect.\[dEs\] we tried to shed light on the expected witnesses of differences in system parameters which allow to distinguish between two different populations of dEs. That almost all DGs can be associated with morphological types and that only a few exceptions show morphological transformations, implies that the act of mutation seems to happen rapidly and thus to be observable with only low probability. During the approach to galaxy clusters, ram pressure should act on dIrrs already at group conditions or in the outskirts of clusters [@mar03; @roe05]. When this process pushes the gas out of dIrrs, wouldn’t one expect to observe many head-tail structures of stellar body vs. stripped-off ISM? Several candidates exist which, however, are characterized as BCDs with decentered bright star-forming knots (see e.g. [@noe00]). The recently detected best candidate in the rapid phase of RPS in the Virgo cluster is VCC1217/IC4318 [@fum11]: A main almost old stellar body of $3 \times 10^8 \Msun$ leaves behind a bundle of $\Ha$ and near-UV emitting knots. Nevertheless, the gas distribution and SF progression within some blobs are not fully understandable. It is still, in general, unclear to what extent the ram pressure can trigger SF by compression of the ISM. Other BCDs are observed although their gas is already exhausted [@del08]. Peebles et al. [@pee08] find a number of dIrrs with excessively high metallicity, what they interpret as the last stage of gas consumption before they reach the dE state. On the other hand, rejuvenation of a fraction of cluster dEs seems also to occur which are found to harbour central warm gas [@mic04]. Whether this fact is indicative of a possible re-transformation from gas-free to gaseous DGs by gas accretion is a matter of debate, since it seems impossible within cluster environments, but plausible in less dense regions where gas infall enhances SF or even triggers starbursts (see sect.\[dIrrs\]) and should not only affect dIrrs. Dwarf spheroidal galaxies {#dSphs} ========================= The possible cycle of morphological mutations, i.e. from gas-rich objects to gas-poor systems by means of gas expulsion and back to a significant gas content by gas accretion, can be explored in the local environment, namely, in the galaxies around the Milky Way, their satellites. Except the Magellanic Clouds most of these can be characterized as gas-free spheroidal systems which manifest the faint end of dEs. Since these dSph are gathered around massive galaxies like our MWG and M31 orbiting them as satellites, go down to the lightest and most metal-poor end of galaxies, they have attracted increasing attention over the last years with the advent of more advanced observing facilities. Understanding their formation and evolution is of substantial relevance for our astrophysical picture of cosmological structure formation and of galaxy evolution. Four main questions are addressed: 1. How and when did they form? They all harbour a very old stellar population [@tol09] and, therefore, seem to have been unaffected by the re-ionzation era [@gre04]. 2. Is their existence as satellite system typical for all massive galaxies? Their origin and DM content is still questioned by some authors [@kro10] because of the large discrepancy of the number of objects really observed vs. expected from $\Lambda$CDM cosmology and because of their orbit concentration to the so-called disk-of-satellites, also found around M 31. This invokes the preference of their tidal-tail origin [@met09]. The observed large velocity dispersions, which are otherwise applied as representative to derive the M/L ratio is then caused by tidal effects. 3. How is their evolution determined by the vicinity of the massive mature galaxy? Not only the tidal field must have a disruptive effect, but also a gaseous halo of the central galaxy will interact with the ISM of the dSphs [@may07]. That the relation of the gas fraction bound to the dSphs is increasing with distance from the MWG [@har01], points into that direction. 4. Vice versa the question arises, how do the satellites influence the structure and evolution of the mature galaxy, here the MWG. The first three questions also concern the morphological transition from gas-rich satellites to dSphs. Nevertheless, dSphs follow a mass-metallicity relation [@dek03; @gre03] and continue the total brightness vs. central surface-brightness relation from normal dEs to the faint end [@gre03]. As the first models, Hensler et al. [@hen04] performed simulations [@hen03] of spherical low-mass galaxies in order to study galaxy survival, SF epochs and rates, gas loss, and (final) metallicity. They demonstrate that due to the SF self-regulation only short but vehement initial SF epochs occur and lead to mass-dependent gas loss. Nonetheless, the DGs remain gravitationally bound with the further issue that more cool gas survives than it is observed, but it forms a halo around the visual body. Although the stellar energetic feedback is the driving mechanism to expel the gas, its effect is not as dramatic as obtained in semi-analytic models [@sal08] and the amount of unbound mass is considerably lower. To get lost, this gas has to be stripped off additionally [@gre03] what probably happens by means of ram pressure of the galactic halo gas [@may07] or by tidal stripping [@rea06]. Otherwise it can return to the DG and produce subsequent events, from a second SF epoch to SF oscillations. The external gas reservoirs around some dSphs [@bou06], in particular also the that is kinematically coupled with the Scl dSph, might witness this effect. The fascinating wealth of data and their precision on stellar ages and kinematics, on their chemical abundances, abundance gradients, and tidal tails of dSphs (for most recent reviews see e.g. [@koc09] and [@tol09]) have triggered numerous numerical models. Although they are advanced since [@hen04] to 3D hydrodynamics (see e.g. [@mar08] and [@rev09]), they still lack of a self-consistent treatment of both, internal processes, as e.g. SF self-regulation (see sect.\[intro\]), and the environmental influences as e.g. tidal effects, external gas pressure, gas inflow, etc. In a recent paper [@rev09], e.g. a large set of DG models is constructed with the method of smooth-particle hydronamics (SPH), but considers all of them in isolation. In most of their models sufficient gas mass is retained and can fuel further SF epochs, if it would not be stripped of by ram pressure or tidal forces, as the authors mention. Those models that fit the presently best studied dSphs Fnx, Car, Scl, and Sex, are than chosen as test cases for further exploration. Although their results do not deviate too much from the further observational data, in addition to the already mentioned neglections, three further caveats exist: 1) If models are selected according to any agreement with one or two observed structural parameters, it is not surprising if also other values would not deviate significantly. 2) The numerical mass resolution of the SPH particles is too low to allow quantitative issues of galactic winds, heating and cooling, etc. 3) Because of the single gas-phase description released metals are too rapidly mixed with the cool gas and the metal-enrichment happens too efficiently. Despite these facts, with appropriate initial conditions always models in agreement to observations can be found. ![ Cubes of 200 kpc length around the Milky Way (at their center). [upper panel:]{} Initial conditions of the Milky Way’s satellite system: Distribution of Dark Matter (DM) subhalos within a sphere of 40 kpc radius around the Milky Way at redshift z=4.56. [lower panel:]{} Snapshot of the satellites’ dynamical evolution 1 Gyr after the numerical onset, i.e. at redshift z=2.76. The DM subhalos are filled with baryonic gas of mass fraction of 17%, form stars, and lose mass of all constituents due to tidal interactions among the satellite system. For discussion see text. []{data-label="fig:sat"}](z4_56.pdf "fig:"){width="12cm"} ![ Cubes of 200 kpc length around the Milky Way (at their center). [upper panel:]{} Initial conditions of the Milky Way’s satellite system: Distribution of Dark Matter (DM) subhalos within a sphere of 40 kpc radius around the Milky Way at redshift z=4.56. [lower panel:]{} Snapshot of the satellites’ dynamical evolution 1 Gyr after the numerical onset, i.e. at redshift z=2.76. The DM subhalos are filled with baryonic gas of mass fraction of 17%, form stars, and lose mass of all constituents due to tidal interactions among the satellite system. For discussion see text. []{data-label="fig:sat"}](z2_76.pdf "fig:"){width="12cm"} Although the advancement to a two-phase ISM treatment in SPH is not trivial and implies various numerical problems, but is not impossible [@ber03; @har06; @sca06], such treatment would be absolutely necessary in order to approach reality and to achieve reliable results. In addition, the interaction processes must be applied [@ber03]. From the $\Lambda$CDM structure-formation paradigm and from numerical simulations with different computational tools, subhalos are expected to assemble around massive halos and to accumulate their masses. If they have already experienced SF, their stars should be merged into a spheroid and be identifiable by their kinematics and chemical abundances. Although these low-mass subhalos, their baryonic content as dSphs, and the accretion scenario [@joh08], therefore, serve as the key to pinpoint this cosmological paradigm, observational detections of stellar streams within the Milky Way (MW) halo are rare [@sea08]. Furthermore, the stellar abundances in present-day dSphs deviate mostly from the halo, in particular $\alpha$/Fe with Fe/H, which characterizes the SF timescale ([@tol09] and e.g. for the Car dSph [@koc08]). Yet evolutionary models of dSph with respect to SF, chemistry, and gas expulsion and their comparison with the Milky Way halo are still too simplistic. While their accretion epoch occurred continuously over the Hubble time some models [@pra08b] only considered it as a short early event; their gas is not only removed by tidal stripping and RPS [@may07] but also re-accreted on their orbits around the MW [@bou06]; in general, consideration as isolated systems lacks reality [@rev09]. To model instead of isolated subhalos the evolution of the system of dSphs in the gravitational field of the MWG for which the accretion by the host galaxy is probable over the Hubble time, the Via Lactea II [@die08] simulation was used. Since an acceptible computational time limits the number of gas particles to two million as also for the DM and in order to reach a high mass resolution of 10$^3 \Msun$ per SPH particle, only 250 subhaloes as DM progenitors of dSphs in the mass range of $10^6 < M_{sat}/\Msun < 6\cdot10^8$ from $z=4.56$ could be followed. Unfortunately, this fact limits the radius of consideration to within a radius of 40 kpc around the MW’s center of mass. In order to study the construction of the MWG halo by accretion of subhalos including baryonic matter, both gas and stars, as a first step, the evolution of the dSph system is followed for the first Gyr, i.e. until redshift $z=2.76$ [@ph11] (see Fig.\[fig:sat\]). For the simulations an advanced version of the single-gas chemo-dynamical SPH/N-body code is applied, treating the production and chemical evolution of 11 elements. Starting with a $10^4$ K warm gas of 17% of the subhalo masses in virial equilibrium and under the assumption that re-ionization is improbable to have affected the Local Group dSphs [@gre04], cooling allows the gas particles to achieve SF conditions in all satellites, but its efficiency directly depends on the mass of a satellite and its dynamical history (merging with other satellites or disruption by the MW gravitational potential). The stellar feedback by SNeII releases sufficient energy to expel hot gas from the main bodies of less massive dSphs, facilitated by tidal interactions. This gas accumulates in the MW halo while massive dSphs merge and continue SF. For the first $10^8$ yr of the simulation there is a considerable variance of stellar oxygen abundance in the whole system $(-5. \leq [O/H] \leq -0.5)$ reflecting the very inhomogeneous production and distribution of enriched gas. After $10^8$ yrs the merging of satellites’ ISM promotes the mixing of heavy elements. Finally, almost completely recycling of the gas erases the abundance inhomogeneities so that O in stars converges to $-1. \leq [O/H] \leq 0.$ with a small dispersion. Detailed analyses of the SF history, gas exchange, stellar abundance evolution of dSphs and the MW halo in the early universe are presented in a comprehensive paper [@ph11] and will be discussed with their implications for our cosmological picture. The author is grateful to Alessandro Boselli, Joachim Koeppen, Thorsten Lisker, Polis Papaderos, Mykola Petrov, Simone Recchi, and further more for their contributions and continuous discussions to this field and to the referee for valuable comments.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A scattering problem (or more precisely, a transmission-reflection problem) of linearized excitations in the presence of a dark soliton is considered in a one-dimensional nonlinear Schrödinger system with a general nonlinearity: $ \mathrm{i}\partial_t \phi = -\partial_x^2 \phi + F(\|\phi\|^2)\phi $. If the system is interpreted as a Bose-Einstein condensate, the linearized excitation is a Bogoliubov phonon, and the linearized equation is the Bogoliubov equation. We exactly prove that the perfect transmission of the zero-energy phonon is suppressed at a critical state determined by Barashenkov’s stability criterion \[Phys. Rev. Lett. 77, (1996) 1193.\], and near the critical state, the energy-dependence of the reflection coefficient shows a saddle-node type scaling law. The analytical results are well supported by numerical calculation for cubic-quintic nonlinearity. Our result gives an exact example of scaling laws of saddle-node bifurcation in time-reversible Hamiltonian systems. As a by-product of the proof, we also give all exact zero-energy solutions of the Bogoliubov equation and their finite energy extension.' address: 'Department of Basic Science, The University of Tokyo, Tokyo 153-8902, Japan' author: - 'Daisuke A. Takahashi' bibliography: - 'genNLS.bib' title: 'Soliton-phonon scattering problem in 1D nonlinear Schrödinger systems with general nonlinearity' --- Nonlinear Schrödinger equation ,Bose-Einstein condensate ,Bogoliubov equation ,saddle-node bifurcation ,universal scaling laws ,cubic-quintic nonlinear Schrödinger equation Introduction ============ In this paper, we solve a scattering problem of linearized excitations in the presence of a dark soliton in one-dimensional(1D) nonlinear Schrödinger (NLS) equation with a general nonlinearity: $$\begin{aligned} \mathrm{i}\partial_t\phi = -\partial_x^2\phi+F(\|\phi\|^2)\phi, \label{eq:intro001} \end{aligned}$$ and discuss the physical and mathematical significance of our results. For a schematic picture, see Fig. \[introfigure\]. The precise mathematical definition of the problem will be given in Sec. \[sec:fundamental\]. If we regard the system as a Bose-Einstein condensate(BEC), the linearized excitation is a Bogoliubov phonon, so the problem can be also called a soliton-phonon scattering problem, as this paper entitled.\ The NLS equation (\[eq:intro001\]) has a great number of applications in nonlinear optics, superconductors, magnetism, BECs, and so on. Particularly, much attention has been focused on the experimental realizations of BECs in laser-cooled ultracold atoms for more than a decade, because of high-controllability of the system parameters. By using elongated laser beams, low-dimensional systems are realized, and a dark soliton can be created via the phase imprinting method[@BurgerPhaseImprinting]. The Bogoliubov theory is also well confirmed[@Andrews; @Stamper; @Steinhauer].\ It is known that 1D NLS with a cubic nonlinearity is completely integrable[@ZakharovShabat; @ZakharovShabat2]. Because of integrability, the linearized equation is also solved exactly[@ChenChenHuang; @Kovrizhin], and the phonon excitations are shown to be completely reflectionless against a soliton for any excitation energy. Thus, the problem is trivial in this case. However, when the nonlinear term is generalized, the phonon has a finite reflection coefficient in general. It is worthy to note that the soliton decay dynamics in the laser-trapped quasi-1D BEC has been well explained by the quintic term, which appears as a second-order perturbation of the trapping effect[@Muryshev; @SinhaChernyKovrizhinBrand; @KhaykovichMalomed], and yields the frictional force between thermal excitation clouds and solitons[@Muryshev; @SinhaChernyKovrizhinBrand]. Thus, knowing the scattering properties between solitons and linearized excitations is essential to understand and control the transport of solitons, that is, the transport of stable wave packets. We also mention that the theory of nonpolynomial NLS equation is formulated to describe the confinement effect[@SalasnichParolaReatto; @MateoDelgado; @MateoDelgadoAnnPhys; @Salasnich]. The quintic NLS also appears in an effective mean-field description of the Tonks-Girardeau gas[@Kolomeisky].\ The NLS equation with an integrability breaking factor is also interesting from the viewpoint of an infinite-dimensional dynamical system and the bifurcation theory. When the potential barrier is added in the cubic NLS equation, there exist stable and unstable stationary supercurrent-flowing solutions[@BaratoffBlackburnSchwartz; @Hakim], if the condensate velocity does not exceed a certain critical value. Near the critical point, which separates the stable branch and the unstable branch, it is known that many physical quantities obey saddle-node type scaling, such as an emission period of dark solitons[@Hakim; @PhamBrachet], an eigenvalue of a growing mode for the unstable solution[@PhamBrachet], and a transmission coefficient of linearized excitations[@Kovrizhin; @Kagan; @DanshitaYokoshiKurihara]. It is quite nontrivial that the time-reversible Hamiltonian system exhibits the scaling behaviors of saddle-node type, since this bifurcation is normally understood to emerge in time-irreversible phenomena. However, it is not easy to prove these properties analytically or exactly, because of the infinite dimensional character of the system.\ On the other hand, as another way to break the integrability, one can consider the generalization of the nonlinearity, that is what we will consider in the present paper. When the nonlinear term includes a competing interaction, the dark soliton is no longer always stable. One typical example of an unstable dark soliton is a “bubble” in a cubic-quintic NLS (CQNLS) system[@BarashenkovMakhankov; @BarashenkovPanova]. (See also [@JYang].) The most general criterion for the stability of the dark soliton has been shown by Barashenkov[@Barashenkov], and the critical velocity of the soliton is determined by $ \partial P/\partial v=0 $, where $ v $ is a velocity of the soliton and $ P $ is a renormalized momentum. The existence of the critical velocity lower than the sound velocity and the separation of stable and unstable regions are similar to the phenomena of superflows against a potential barrier. Therefore, we can expect some scaling behavior near the critical state. Furthermore, in the present case, the preserved translational symmetry of the fundamental equation makes it possible to access the problem analytically.\ In this paper, we solve the scattering problem of linearized phonon excitations, and exactly show the following: (i) At the critical state determined by Barashenkov’s criterion[@Barashenkov], the transparency of the zero-energy phonon is suppressed, and only partial transmission occurs. (ii) Near the critical state, the energy-dependence of the reflection coefficient of low-energy phonons shows saddle-node scaling behavior, regarding the renormalized momentum as a parameter of a normal form of saddle-node bifurcation. The obtained analytical results are well confirmed by comparison with the numerical results of CQNLS equation. Our result gives an exact example of scaling laws of saddle-node bifurcation in time-reversible Hamiltonian systems. The proof is based on the exact low-energy expansion of the solution of the linearized equation. Since the exact zero-energy solutions given in this paper are quite general, we believe that our method will also be useful to derive other low-energy physical properties.\ The organization of the paper is as follows. In Sec. \[sec:fundamental\], we introduce fundamental equations and see the fundamental properties. The definition of the transmission-reflection problem is also given. In Sec. \[sec:mainresult\], we give a main result and verify it by numerical study of CQNLS equation. Sections \[sec:proof1\] and \[sec:proof2\] are devoted to the proof of main results. Discussions, future perspectives, and conclusions are given in Sec. \[sec:summary\]. Some mathematically technical formulae are treated in Appendices. ![\[introfigure\]A schematic picture of the problem that we consider in this paper. $ p $ represents a half of the velocity of the dark soliton. The problem is always considered in the comoving frame of the soliton. $ \mathrm{e}^{\mathrm{i}k_1x} $ is an incident wave of a linearized excitation, $ t\mathrm{e}^{\mathrm{i}k_1x} $ is a transmitted wave, and $ r\mathrm{e}^{\mathrm{i}k_2x} $ is a reflected wave. For more detailed definitions of each quantity, see Sec. \[sec:fundamental\].](introfiguretex2img.eps) Fundamental Equations and Definition of the Problem {#sec:fundamental} =================================================== Fundamental equations --------------------- We begin with the NLS equation with a general nonlinearity $$\begin{aligned} \mathrm{i}\partial_t\phi=-\partial_x^2\phi+F(\|\phi\|^2)\phi. \label{eq:nls} \end{aligned}$$ Here, $ F(\rho) $ is a real-valued function such that $ F(0)=0 $. The energy functional (Hamiltonian) which yields this equation is $$\begin{aligned} H = \int\!\mathrm{d}x \left(\|\partial_x\phi\|^2+U(\|\phi\|^2)\right), \end{aligned}$$ where $$\begin{aligned} U(\rho) = \int_0^\rho\mathrm{d}\rho' F(\rho'). \end{aligned}$$ Letting $ \phi=\phi+\delta\phi $ in Eq. (\[eq:nls\]), and discarding higher order terms of $ \delta\phi $, one obtains the following linearized equation: $$\begin{aligned} \mathrm{i}\partial_t\delta\phi=\left[-\partial_x^2+ F(\|\phi\|^2)+\|\phi\|^2F'(\|\phi\|^2) \right]\delta\phi+\phi^2F'(\|\phi\|^2)\delta\phi^. \end{aligned}$$ Writing $ \delta\phi=u $ and $ -\delta\phi^=v $, one obtains $$\begin{aligned} \mathrm{i}\partial_t\begin{pmatrix}u\\v\end{pmatrix}=\mathcal{L}\begin{pmatrix}u\\v\end{pmatrix}, \label{eq:tdbogo} \end{aligned}$$ where $ \mathcal{L} $ is a $ 2\times2 $ matrix operator whose components are $$\begin{aligned} \mathcal{L}_{11}^{}&=-\mathcal{L}_{22}^{}=-\partial_x^2+F(\|\phi\|^2)+\|\phi\|^2F'(\|\phi\|^2), \\ \mathcal{L}_{12}^{}&=-\mathcal{L}_{21}^=-\phi^2F'(\|\phi\|^2). \end{aligned}$$ We use the notation $ (u,v) $ since it is commonly used by condensed matter physicists. If we interpret the system as BEC, this equation is the Bogoliubov equation which describes the Bogoliubov phonon (or Bogoliubov quasiparticle) [@Bogoliubov]. (As a review or a textbook, see, e.g., [@FetterWalecka; @DalfovoGiorginiPitaevskiiStringari; @PethickSmith].) For this reason, henceforth, we call $ \phi $ the condensate wavefunction or the order parameter, and $ (u,v) $ the Bogoliubov (quasiparticle) wavefunction, though the NLS equation itself has more applications in various fields.\ Henceforth we mainly consider the stationary (i.e., time-independent) problem. The stationary NLS equation with chemical potential $ \mu $ is obtained by setting $ \phi(x,t)=\phi(x)\mathrm{e}^{-\mathrm{i}\mu t} $: $$\begin{aligned} (-\mu-\partial_x^2+F(\|\phi\|^2))\phi=0. \label{eq:nls2} \end{aligned}$$ As will be seen, the value of $ \mu $ is fixed by the asymptotic form of $ \phi $ . The stationary Bogoliubov equation with eigenenergy $ \epsilon $ is obtained by setting $ u(x,t)=u(x)\mathrm{e}^{-\mathrm{i}(\epsilon+\mu)t} $ and $ v(x,t)=v(x)\mathrm{e}^{-\mathrm{i}(\epsilon-\mu)t} $: $$\begin{aligned} \epsilon\begin{pmatrix}u\\v\end{pmatrix}=\mathcal{L}_\mu\begin{pmatrix}u\\v\end{pmatrix} \label{eq:bogos} \end{aligned}$$ with $$\begin{aligned} \mathcal{L}_\mu:= \mathcal{L}+\begin{pmatrix}-\mu &0 \\ 0& \mu \end{pmatrix}. \label{eq:bogos2} \end{aligned}$$ Bogoliubov phonons in a uniform condensate {#subsec:uniform} ------------------------------------------ Let us derive the dispersion relation (the energy-momentum relation) of Bogoliubov phonons when the condensate is flowing uniformly: $ \phi(x) = \sqrt{\rho_\infty}\mathrm{e}^{\mathrm{i}(px+\varphi)} $. In order for this $ \phi(x) $ to be the solution of Eq. (\[eq:nls2\]), the chemical potential must be $$\begin{aligned} \mu=p^2+F(\rho_\infty). \label{eq:cp} \end{aligned}$$ The four solutions of Bogoliubov equation (\[eq:bogos\]) are given by $$\begin{aligned} w_i(x,\varphi):= \begin{pmatrix} \bar{u}_i\, \mathrm{e}^{\mathrm{i}(px+\varphi)} \\ \bar{v}_i\, \mathrm{e}^{-\mathrm{i}(px+\varphi)} \end{pmatrix} \mathrm{e}^{\mathrm{i}k_i x}, \label{eq:bogouniform} \end{aligned}$$ where $ i=1,\,2,\,3, \text{ and }4 $, and the wavenumber $ k_i $s are the roots of the following quartic equation: $$\begin{aligned} (\epsilon-2kp)^2=k^2(k^2+2\rho_\infty F'(\rho_\infty)). \label{eq:uniformdisp} \end{aligned}$$ Equation (\[eq:uniformdisp\]) gives the dispersion relation, and from this dispersion one can see that a half of the Landau’s critical velocity (or a half of the sound wave velocity) is given by $$\begin{aligned} p_{\text{L}}=\sqrt{\frac{\rho_\infty F'(\rho_\infty)}{2}}. \label{eq:defofpl} \end{aligned}$$ (Note that the sound wave velocity is not $ p_{\text{L}} $ but $ 2p_{\text{L}} $; see the next subsection.) Since $ p_{\text{L}} $ must be real, in order for the uniform condensate to be stable, $ F'(\rho_\infty)>0 $ must hold. The coefficients $ \bar{u}_i $ and $ \bar{v}_i $ can be, e.g., chosen as follows: $$\begin{aligned} \bar{u}_{i}&=\sqrt{1+\frac{\epsilon-2p k_i^{}+k_i^2}{2p_{\text{L}}^2}}, \label{eq:ucoeff}\\ \bar{v}_{i}&=\sqrt{1-\frac{\epsilon-2p k_i^{}-k_i^2}{2p_{\text{L}}^2}}. \label{eq:vcoeff} \end{aligned}$$ When $ \epsilon>0 $ and $ -p_{\text{L}}<p<p_{\text{L}} $, the quartic equation (\[eq:uniformdisp\]) has one real positive root, one real negative root, and two complex roots conjugate to each other. We call a real positive (negative) root $ k_1 \ (k_2) $, and a complex root with positive (negative) imaginary part $ k_3 \ (k_4) $. The low-energy expansions of them are given by $$\begin{aligned} k_1 &= \frac{\epsilon}{2(p+p_{\text{L}})}+O(\epsilon^3), \label{eq:kexpand1} \\ k_2 &= \frac{\epsilon}{2(p-p_{\text{L}})}+O(\epsilon^3), \label{eq:kexpand2} \\ k_3 &= 2\mathrm{i}\sqrt{p_{\text{L}}^2-p^2}+\frac{p\epsilon}{2(p_{\text{L}}^2-p^2)}+O(\epsilon^2), \label{eq:kexpand3} \\ k_4 &= -2\mathrm{i}\sqrt{p_{\text{L}}^2-p^2}+\frac{p\epsilon}{2(p_{\text{L}}^2-p^2)}+O(\epsilon^2). \label{eq:kexpand4} \end{aligned}$$ $ w_1 $ and $ w_2 $ are plane wave solutions propagating in the positive and negative directions, respectively. $ w_3 $ and $ w_4 $ are exponentially divergent unphysical solutions. Dark soliton solution in comoving frame --------------------------------------- Let us consider the dark soliton solution of stationary NLS Eq. (\[eq:nls2\]) in the comoving frame of the soliton. In this coordinate, the soliton is static but the surrounding condensate is flowing. Let us seek the solution with the asymptotic form $$\begin{aligned} \phi(x\rightarrow \pm\infty)=\sqrt{\rho_\infty}\mathrm{e}^{\mathrm{i}(px\pm\frac{\delta}{2})}. \label{eq:nlsdsasym} \end{aligned}$$ It should be noted that the velocity of the soliton is not $ -p $ but $ -2p $, because the Galilean covariance of NLS equation leads to the following property: $$\begin{aligned} \begin{split} &\text{$ \phi(x,t) $ is a solution.} \\ \leftrightarrow \quad & \text{$\tilde{\phi}(x,t,\alpha)=\phi(x+2\alpha t,t)\mathrm{e}^{-\mathrm{i}\alpha x}\mathrm{e}^{-\mathrm{i}\alpha^2t}$ is a solution.} \end{split}\label{eq:galilei} \end{aligned}$$ So, if one has the solution in the form $ \phi(x,t)=\mathrm{e}^{-\mathrm{i}\mu t}\mathrm{e}^{\mathrm{i}px}f(x) $, the corresponding soliton-moving solution is given by $ \tilde{\phi}(x,t,p)=\mathrm{e}^{-\mathrm{i}(\mu-p^2)t}f(x+2pt) $. However, for brevity, we sometimes call $ p $ “velocity”, ignoring the difference of twice factor.\ From the conservation laws of mass and momentum, one can immediately find two integration constants: $$\begin{aligned} j&=\frac{\phi^\phi_x-\phi\phi_x^}{2\mathrm{i}}, \\ j_m &= \|\phi_x\|^2+\mu\|\phi\|^2-U(\|\phi\|^2). \label{eq:genjm} \end{aligned}$$ Here $ j $ is a mass current density and $ j_m $ is a momentum current density. Let us write the density and the phase of the condensate as $ \phi=\sqrt{\rho}\mathrm{e}^{\mathrm{i}S} $. Taking account of the asymptotic form (\[eq:nlsdsasym\]), the chemical potential $ \mu $ becomes the same as (\[eq:cp\]), and the above constants are determined as $$\begin{aligned} j&=\rho_\infty p, \label{eq:dsj} \\ j_m&=2\rho_\infty p^2+\rho_\infty F(\rho_\infty)-U(\rho_\infty). \label{eq:dsjm} \end{aligned}$$ The conservation laws are then rewritten as $$\begin{aligned} S_x &= \frac{j}{\rho} =\frac{\rho_\infty p}{\rho} \ \leftrightarrow \ S = p\int_0^x\frac{\rho_\infty\mathrm{d}x}{\rho}, \label{eq:phasecond} \\ \frac{(\rho_x)^2}{4} &= -p^2(\rho_\infty-\rho)^2+\rho\left[ U(\rho)-U(\rho_\infty)-(\rho-\rho_\infty)U'(\rho_\infty) \right]. \label{eq:momconsrv} \end{aligned}$$ Thus, one can at least obtain the formal solution $$\begin{aligned} \pm 2(x-x_0) = \int\!\!\frac{\mathrm{d}\rho}{\sqrt{\text{R.H.S. of Eq. (\ref{eq:momconsrv})}}}, \end{aligned}$$ even though it is not easy in general to carry out this integration and obtain the solution in closed form “$ \rho(x)=\dots $”. Henceforth we do not need this formal solution, but we assume the existence of a dark soliton solution which has no singularity and satisfies the asymptotic condition (\[eq:nlsdsasym\]).\ From Eq. (\[eq:phasecond\]), The phase shift $ \delta $ in Eq. (\[eq:nlsdsasym\]) can be written down explicitly: $$\begin{aligned} \delta = p\int_{-\infty}^\infty\!\!\mathrm{d}x\left( \frac{\rho_\infty}{\rho}-1 \right). \end{aligned}$$ We also introduce the symbol for the particle number of the dark soliton for later convenience: $$\begin{aligned} N := \int_{-\infty}^\infty\!\!\mathrm{d}x\left( \rho-\rho_\infty \right)<0. \end{aligned}$$ Barashenkov’s criterion ----------------------- The stability of the soliton is described by the following renormalized momentum[@Barashenkov]: $$\begin{aligned} P = \int_{-\infty}^\infty\mathrm{d}x\left( \frac{\tilde{\phi}^\tilde{\phi}_x-\tilde{\phi}\tilde{\phi}^_x}{2\mathrm{i}} \right)\left( 1-\frac{\rho_\infty}{\|\tilde{\phi}\|^2} \right) \end{aligned}$$ Here $ \tilde{\phi}(x,t) $ is the dark soliton solution in the frame where the surrounding condensate is at rest and the soliton is moving. Writing the soliton velocity $ v $, the stability criterion for the dark solitons is expressed by $ \partial P/\partial v<0 $.\ We can rewrite the above integral by the density profile $ \rho(x) $: $$\begin{aligned} \begin{split} P &= -p\int_{-\infty}^\infty\mathrm{d}x\left( \rho_\infty-\rho \right)\left( \frac{\rho_\infty}{\rho}-1 \right) \\ &= -p N-\rho_\infty \delta. \end{split} \end{aligned}$$ Here remember that the soliton velocity is given by $ v=-2p $, as stated in the preceding subsection. The stability condition is rewritten as $ \partial (-P)/\partial p <0 $. Definition of the scattering problem {#sec:defofsp} ------------------------------------ In this subsection, we define the transmission and reflection problem of Bogoliubov phonons shown in Fig. \[introfigure\]. Since the linearized equation does not satisfy simple particle number conservation, we must define transmission and reflection coefficients via the conservation of excitation energy. The conservation of excitation energy corresponds to the constancy of the following Wronskian:[@Kagan; @DanshitaYokoshiKurihara]: $$\begin{aligned} W = u^\partial_xu-u\partial_xu^+v^\partial_xv-v\partial_xv^. \end{aligned}$$ Let us assume that the asymptotic form of the condensate wavefunction $ \phi(x) $ is given by Eq. (\[eq:nlsdsasym\]). In this situation, sufficiently far from the origin, the Bogoliubov equations have the plane wave (and exponentially decaying/diverging) solutions given by Eq. (\[eq:bogouniform\]) with $ \varphi=\pm\frac{\delta}{2} $.\ The solution of the scattering problem is defined by the one that has the following asymptotic form[@Kovrizhin; @Kagan; @DanshitaYokoshiKurihara]: $$\begin{aligned} \begin{pmatrix}u \\ v \end{pmatrix} \rightarrow \begin{cases} w_1\left(x,-\frac{\delta}{2}\right)+r\, w_2\left(x,-\frac{\delta}{2}\right) & (x\rightarrow-\infty) \\ t\, w_1\left(x,+\frac{\delta}{2}\right) & (x\rightarrow+\infty). \end{cases} \label{eq:uvasymptotic} \end{aligned}$$ Here the exponentially decaying waves, which are $ w_4(x,-\frac{\delta}{2}) $ in $ x\rightarrow -\infty $ and $ w_3(x,+\frac{\delta}{2}) $ in $ x\rightarrow+\infty $, can be also included. However, they are irrelevant in the definition of transmission and reflection coefficients. The calculation of $ W $ shows that $$\begin{aligned} \frac{W(+\infty)}{2\mathrm{i}}=&\|t\|^2\left[ (k_{1}+p)\|\bar{u}_{1}\|^2+(k_{1}-p)\|\bar{v}_{1}\|^2 \right]\!, \\ \begin{split} \frac{W(-\infty)}{2\mathrm{i}}=&(k_{1}+p)\|\bar{u}_{1}\|^2+(k_{1}-p)\|\bar{v}_{1}\|^2 \\ &\ +\|r\|^2\left[ (k_{2}+p)\|\bar{u}_{2}\|^2+(k_{2}-p)\|\bar{v}_{2}\|^2 \right]. \end{split} \end{aligned}$$ Since $ W(+\infty)=W(-\infty) $, the transmission coefficient $ T $ and the reflection coefficient $ R $ are naturally defined as $$\begin{aligned} T&=\|t\|^2, \label{eq:tc} \\ R&=\frac{(-k_{2}-p)\|\bar{u}_{2}\|^2+(-k_{2}+p)\|\bar{v}_{2}\|^2}{(k_{1}+p)\|\bar{u}_{1}\|^2+(k_{1}-p)\|\bar{v}_{1}\|^2}\|r\|^2. \label{eq:defofR} \end{aligned}$$ By this definition, $ T+R=1 $ always holds. If one chooses the normalization $ \bar{u}_i $ and $ \bar{v}_i $ as Eq. (\[eq:ucoeff\]) and (\[eq:vcoeff\]), one can show $$\begin{aligned} \frac{(-k_{2}-p)\|\bar{u}_{2}\|^2+(-k_{2}+p)\|\bar{v}_{2}\|^2}{(k_{1}+p)\|\bar{u}_{1}\|^2+(k_{1}-p)\|\bar{v}_{1}\|^2} = 1-\frac{pp_{\text{L}}\epsilon^2}{2(p^2-p_{\text{L}}^2)^3}+O(\epsilon^4). \end{aligned}$$ So, if one is only interested in the leading order, one can approximate $ R \simeq \|r\|^2 $. Summary of Main Result and Numerical Verification {#sec:mainresult} ================================================= In this section we present the main results of this paper and verify them by numerical study of CQNLS equation. The proof will be given in Secs. \[sec:proof1\] and \[sec:proof2\]. Main result {#subsec:mainresult} ----------- In the scattering problem of linearized excitations defined in Subsec. \[sec:defofsp\], the amplitude of the reflected component $ r $ in Eq. (\[eq:uvasymptotic\]) is given by the following Padé approximant-like form: $$\begin{aligned} r = \frac{- \mathrm{i}\epsilon(d+d_1P_p)+O(\epsilon^2)}{a P_p - \mathrm{i}\epsilon(b+b_1P_p)+O(\epsilon^2)}. \label{eq:mainr} \end{aligned}$$ Here $ X_p:= \partial X/\partial p $ and $$\begin{aligned} a &= 4p_{\text{L}}\rho_\infty, \\ b &= (N+pN_p)^2+(p_{\text{L}}N_p)^2, \\ d &= (N+pN_p)^2-(p_{\text{L}}N_p)^2, \\ b_1 &= N-\frac{p_{\text{L}}^2+p^2}{p_{\text{L}}^2-p^2}\widetilde{N}, \\ d_1 &= N+\widetilde{N} \end{aligned}$$ with $$\begin{aligned} \widetilde{N}:=p \frac{\partial N}{\partial p}-\rho_\infty\frac{\partial N}{\partial \rho_\infty}. \end{aligned}$$ From (\[eq:mainr\]), the energy dependence of the reflection coefficient $ R $ (\[eq:defofR\]) becomes $$\begin{aligned} R = \begin{cases} \displaystyle \left( \frac{d+d_1P_p}{aP_p} \right)^2\epsilon^2+O(\epsilon^4) & (P_p\ne0) \\ \displaystyle \left( \frac{d}{b} \right)^2+O(\epsilon^2) & (P_p=0). \end{cases} \label{eq:mainrR} \end{aligned}$$ When $ P_p \ne0$, the zero-energy phonon transmits perfectly: $ \lim_{\epsilon\rightarrow0}R=0 $. On the other hand, when the soliton velocity reaches a critical value, i.e. $ P_p=0 $, this perfect transmission disappears. We note that the rational form expression (\[eq:mainr\]) makes it possible to unify the description of low-energy behaviors in both critical and non-critical cases. If we use a simple Taylor series, the singular behavior at the critical velocity state cannot be expressed.\ The above (\[eq:mainrR\]) is one good result valid for any soliton velocity, even if it is far from the critical state. However, when the velocity comes close to the critical value, i.e., $ P_p $ comes close to zero, we can derive a more powerful scaling law as below. Let us assume that coefficients of $ \epsilon^n \ (n\ge2) $ in (\[eq:mainr\]) are all finite in the limit $ P_p\rightarrow0 $. and take the limit $ \epsilon\rightarrow 0 $ and $ P_p\rightarrow0 $ with a constraint $ \epsilon/P_p=\text{fix} $. We then obtain $$\begin{aligned} r \rightarrow \frac{-\mathrm{i}d(\epsilon/P_p)}{a-\mathrm{i}b(\epsilon/P_p)}, \end{aligned}$$ and in the same limit, the universal* form of reflection coefficient $$\begin{aligned} \lim_{\substack{\epsilon\rightarrow0,\, P_p\rightarrow0,\\ \epsilon/P_p:\text{fix}}} R = \frac{d^2(\epsilon/P_p)^2}{a^2+b^2 (\epsilon/P_p)^2} \label{eq:scaledR} \end{aligned}$$ follows. Here, the values of $ a, b, $ and $ d $ in the critical state must be substituted when we use Eq. (\[eq:scaledR\]). We remark that Eq. (\[eq:scaledR\]) contains only $ p $-derivatives, whereas the expression before taking the scaling limit contains $ \rho_\infty $-derivatives in addition to $ p $-derivatives.\ Let $ p_c $ be a critical velocity of the dark soliton, i.e., $ P'(p_c)=0 $. The expansion of $ P $ near $ p=p_c $ gives $$\begin{aligned} P(p) &\simeq P(p_c)+\frac{1}{2}P''(p_c)(p-p_c)^2+\dotsb, \\ \rightarrow \|P_p\| &= \|P'(p)\| \simeq \|2P''(p_c)(P(p)-P(p_c))\|^{1/2}. \end{aligned}$$ Therefore we obtain $$\begin{aligned} \frac{\epsilon}{\|P_p\|} \simeq \frac{\epsilon}{\|2P''(p_c)(P(p)-P(p_c))\|^{1/2}}. \label{eq:energyscale} \end{aligned}$$ This is an expected scaling behavior from the normal form of saddle-node bifurcation[@GuckenheimerHolmes; @PhamBrachet], if we regard the renormalized momentum $ P $ as a parameter of normal form. Comparison with Numerical Results in CQNLS System {#subsec:cqnls} ------------------------------------------------- In this subsection, we numerically verify the analytical results of the preceding subsection in the CQNLS system. We first derive the expressions for the dark soliton solution and the renormalized momentum in Subsec. \[subsubsec:darksolitoncqnls\], and solve the scattering problems of linearized excitations for (i) the purely cubic case in Subsec. \[sec:purecubic\], (ii) the purely quintic case in Subsec. \[sec:purequintic\], and (iii) the case where a non-trivial critical velocity exists in Subsec. \[sec:criticalcase\]. ### Dark soliton solution and renormalized momentum {#subsubsec:darksolitoncqnls} In the CQNLS system, the nonlinear term is defined by $$\begin{aligned} U(\rho) &= a_1\rho^2+a_2\rho^3, \\ F(\rho)&=U'(\rho) = 2a_1\rho+3a_2\rho^2, \end{aligned}$$ and the NLS equation (\[eq:nls\]) has the cubic-quintic nonlinearity: $$\begin{aligned} \mathrm{i}\partial_t\phi=-\partial_x^2\phi+2a_1\|\phi\|^2\phi+3a_2\|\phi\|^4\phi. \end{aligned}$$ The stationary linearized equation, i.e., the stationary Bogoliubov equation (\[eq:bogos\]) is given by $$\begin{aligned} \!\!\!\!\!\!&\begin{pmatrix}-\partial_x^2-\mu+4a_1\|\phi\|^2+9a_2\|\phi\|^4 & -2a_1\phi^2-6a_2\|\phi\|^2\phi^2 \\ 2a_1\phi^{2}+6a_2\|\phi\|^2\phi^{2} & \!\! \partial_x^2+\mu-4a_1\|\phi\|^2-9a_2\|\phi\|^4 \end{pmatrix}\begin{pmatrix}u \\ v \end{pmatrix} \nonumber \\ &=\epsilon\begin{pmatrix}u\\v\end{pmatrix}. \label{eq:bogocqnls} \end{aligned}$$ It is known that a bubble and unstable dark solitons appear when $ a_1<0 $ and $ a_2>0 $[@BarashenkovMakhankov; @BarashenkovPanova]. This case is considered in Subsec. \[sec:criticalcase\]. As shown in [@BarashenkovPanova], when the soliton velocity is smaller than the critical value, a small perturbation induces “nucleation dynamics”, and the soliton cannot preserve its shape any more. So, this instability is not convective but absolute.\ The Landau velocity (\[eq:defofpl\]) is given by $$\begin{aligned} p_{\text{L}} = \sqrt{\rho_\infty(a_1+3a_2\rho_\infty)}. \end{aligned}$$ The necessary condition $ a_1+3a_2\rho_\infty>0 $ follows for a uniform condensate to be stable. The dark soliton solution is given by $$\begin{aligned} \phi(x,p,\rho_\infty) =\mathrm{e}^{\mathrm{i}px}\frac{\kappa\rho_0+\mathrm{i}p(\rho_\infty-\rho_0)\tanh\kappa x}{\sqrt{\rho_0(\kappa^2-a_2(\rho_\infty-\rho_0)^2\tanh^2\kappa x)}} \label{eq:dscqnls} \end{aligned}$$ with $$\begin{aligned} \kappa &=\sqrt{p_{\text{L}}^2-p^2}, \\ \rho_0 &=\rho(x=0)= \frac{-(2a_2\rho_\infty+a_1)+\sqrt{(2a_2\rho_\infty+a_1)^2+4a_2p^2}}{2a_2}. \label{eq:rho0cqnls} \end{aligned}$$ See \[app:cqnls\] for a detailed derivation. Since $ \kappa $ and $ \rho_0 $ are the functions of $ (p,\rho_\infty) $, the dark soliton solution has two parameters $ (p,\rho_\infty) $. From (\[eq:rho0cqnls\]), $$\begin{aligned} \lim_{p\rightarrow0}\rho_0=\begin{cases} 0 & (2a_2\rho_\infty+a_1>0) \\ \frac{1}{2a_2}\|2a_2\rho_\infty+a_1\| & (2a_2\rho_\infty+a_1<0), \end{cases} \end{aligned}$$ so the bubble (= a non-topological dark soliton) appears when $ \rho_\infty<-a_1/(2a_2) $. A particle number of soliton $ N $ and a phase difference $ \delta $ are calculated as $$\begin{aligned} N &= -\frac{2}{\sqrt{a_2}}\tanh^{-1}\left[ \frac{\sqrt{a_2}(\rho_\infty-\rho_0)}{\kappa} \right],\label{eq:cqnlsN} \\ \delta &= 2 \tan^{-1}\left[ \frac{p(\rho_\infty-\rho_0)}{\rho_0\kappa} \right]. \label{eq:cqnlsdelta} \end{aligned}$$ From them we can calculate the renormalized momentum $ -P=pN+\rho_\infty \delta $. An example is shown in Fig. \[fig:renP\]. The case where the unstable region exists is analyzed in Subsec. \[sec:criticalcase\] in detail. ![\[fig:renP\](Color online) Plot of renormalized momentum $ -P/\rho_\infty=\delta+pN/\rho_\infty $ in CQNLS. Here we set $ (a_1,a_2)=(-1,1) $. The values of $ \rho_\infty $ of each curve are set, from top to bottom, $ \rho_\infty=0.55,\, 0.52,\, 0.502,\, 0.5,\, 0.498,\, 0.48,\, \text{ and } 0.45 $, respectively. The dark solitons are stable in the regions of the solid lines, while unstable in the regions of the dashed lines. The critical points are marked by black dots. The unstable soliton appears when $ \rho_\infty<-a_1/(2a_2)=0.5 $. ](renP102.eps) ### Purely cubic case {#sec:purecubic} As a first example, let us consider the case $ a_2=0 $, i.e., the nonlinearity is purely cubic. As mentioned in the Introduction, the NLS equation is integrable in this case and the Bogoliubov phonons are reflectionless for any energy. Let us see that our analytical result Eq. (\[eq:mainrR\]) is consistent with these known facts.\ Without loss of generality, we can set $ a_1=1 $. The dark soliton solution (\[eq:dscqnls\]) is then reduced to $$\begin{aligned} \phi = \mathrm{e}^{\mathrm{i}px}\left( p+\mathrm{i}\kappa\tanh \kappa x \right), \quad \kappa = \sqrt{\rho_\infty-p^2}. \end{aligned}$$ The exact solution of the linearized equation, i.e., the Bogoliubov equation is given by the squared Jost solution[@ChenChenHuang; @Kovrizhin]: $$\begin{aligned} u &= \mathrm{e}^{\mathrm{i}(k_j+p)x}\left( \mathrm{i}\kappa\tanh \kappa x+\frac{k_j}{2}+\frac{\epsilon}{2k_j} \right)^2, \\ v &= \mathrm{e}^{\mathrm{i}(k_j-p)x}\left( \mathrm{i}\kappa\tanh \kappa x+\frac{k_j}{2}-\frac{\epsilon}{2k_j} \right)^2, \end{aligned}$$ where $ k_j $s are given by the roots of the dispersion relation Eq. (\[eq:uniformdisp\]) with $ F'(\rho_\infty)=2 $.\ From the above explicit expression, it is obvious that the phonons are reflectionless. Therefore, the coefficient of $ \epsilon^2 $ in Eq. (\[eq:mainrR\]) must vanish. Let us check it. For the cubic case, it follows that $$\begin{aligned} N=-2\kappa,\quad \delta = 2 \tan^{-1}\frac{\kappa}{p}, \end{aligned}$$ by taking the limit $ a_2\rightarrow0 $ of Eqs. (\[eq:cqnlsN\]) and (\[eq:cqnlsdelta\]). With the use of them, we can show $$\begin{gathered} P_p = -N-pN_p-\rho_\infty \delta_{p}=4\kappa, \\ d=4(\kappa^2-3p^2),\quad d_1=\frac{-\kappa^2+3p^2}{\kappa}. \end{gathered}$$ Thus we obtain $ d+d_1P_p=0 $, as expected. We also note that the soliton is always stable since $ -P_p<0 $ for all velocities.\ It is also possible to discuss the reflection properties when the quintic term is small by expanding Eqs (\[eq:cqnlsN\]) and (\[eq:cqnlsdelta\]) with respect to $ a_2 $, but the expression is not so simple. In this case, one can derive an approximate formula valid not only for small energy but for arbitrary energy by the method given in Refs. [@Muryshev; @SinhaChernyKovrizhinBrand]. ### Purely quintic case {#sec:purequintic} Next, we treat the purely (self-defocusing) quintic case. As already mentioned, the quintic NLS equation is known to describe the dynamics of the Tonks-Girardeau gas[@Kolomeisky].\ Without loss of generality, we can set $ a_1=0 $ and $ a_2=1 $. Though we can also set $ \rho_\infty=1 $, we keep it for a moment because we need the differentiation of $ \rho_\infty $ to calculate the reflection coefficient (\[eq:mainrR\]). Eqs. (\[eq:cqnlsN\]) and (\[eq:cqnlsdelta\]) are reduced to $$\begin{aligned} N &= -\tanh^{-1}\left[ \frac{\sqrt{3(1-y^2)}}{2} \right], \\ \delta &= 2 \tan^{-1}\left[ \frac{1-3y^2+\sqrt{1+3y^2}}{3y\sqrt{1-y^2}} \right] \\ \text{with}\quad y &:= \frac{p}{p_{\text{L}}} = \frac{1}{\sqrt{3}}\frac{p}{\rho_\infty}. \end{aligned}$$ From them one can plot the renormalized momentum $ -P/\rho_\infty=\sqrt{3}yN+\delta $ and can show that $ -P_p<0 $ always holds, i.e., the soliton is always stable. One can also obtain the coefficient of $ \epsilon^2 $ in Eq. (\[eq:mainrR\]) as follows: $$\begin{aligned} -\frac{d+d_1P_p}{aP_p} &= \frac{1}{4\sqrt{3}\rho_\infty^2}\frac{\gamma\tanh^{-1}\gamma}{\gamma+2(1-\gamma^2)\tanh^{-1}\gamma}, \label{eq:quinticcoeff} \\ \gamma&:=\frac{\sqrt{3(1-y^2)}}{2}. \end{aligned}$$ ![\[fig:quinticRC\](Color online) Energy-dependence of reflection coefficient $ R $ of linearized excitations for various soliton velocities in the purely quintic system. Here we set $ (a_1,a_2)=(0,1) $ and $ \rho_\infty=1 $. The Landau velocity is given by $ p_{\text{L}}=\sqrt{3}\rho_\infty=\sqrt{3} $. Parabolic curves represent the theoretical approximate expression (\[eq:mainrR\]) with (\[eq:quinticcoeff\]).](quinticRC002.eps) The energy-dependence of the reflection coefficient $ R $ of linearized excitations is obtained by solving the Bogoliubov equation (\[eq:bogocqnls\]) numerically, and the results are shown in Fig. \[fig:quinticRC\]. We can verify that the expression (\[eq:mainrR\]) with (\[eq:quinticcoeff\]) is valid for low-energy region. From this figure we can also see that the soliton with zero velocity is the strongest scatterer. It is intuitively clear since the shape of the soliton becomes shallower and wider if the velocity of the soliton increases. However, this intuitive understanding is not always correct, as the integrable cubic case in Subsec. \[sec:purecubic\] and the instability-induced anomaly in Subsec. \[sec:criticalcase\] illustrate. ### The case with $ a_1<0 $ and $ a_2>0 $ {#sec:criticalcase} Finally, we consider the case with $ a_1<0 $ and $ a_2>0 $, which is most interesting from the viewpoint of critical phenomena, since the soliton can become unstable and the reflection coefficient can show the singular and scaling behavior.\ If both $ a_1 $ and $ a_2 $ are nonzero, we can set $ \|a_1\|=\|a_2\|=1 $ without loss of generality by the following scale transformation: $$\begin{gathered} \bar{x}=\frac{x}{\xi},\ \bar{t}=\frac{t}{\xi^2},\ \bar{\phi}(\bar{x},\bar{t})=\frac{1}{\eta}\phi(x,t), \\ \xi=\frac{\sqrt{\|a_2\|}}{\|a_1\|},\ \eta=\sqrt{\frac{\|a_1\|}{\|a_2\|}}. \end{gathered}$$ So we performed numerical calculations by setting $ (a_1,a_2)=(-1,1)$. Note that $ \rho_\infty $ cannot be normalized to be unity if we choose $ \xi $ and $ \eta $ as the above. Another choice of $ \eta $ is possible to normalize $ \rho_\infty=1 $, but in this case either $ a_1 $ or $ a_2 $ cannot be normalized.\ Using the dark soliton solution (\[eq:dscqnls\]), we numerically solved the stationary Bogoliubov equation (\[eq:bogocqnls\]), and constructed the solution with the asymptotic form (\[eq:uvasymptotic\]). Figure \[fig:RC\] shows the reflection coefficient with $ \rho_\infty=0.45 $ for various soliton velocities. We can observe that the zero-energy phonon transmits perfectly, unless the soliton velocity is equal to the critical one. When the soliton velocity comes close to the critical value, the slope of the reflection coefficient becomes very steep, and at the critical state, the perfect transmission eventually vanishes. The approximate expression (\[eq:mainrR\]) is good for sufficiently low excitation energy. Figure \[fig:scaledR\] shows the scaling behavior of reflection coefficient $ R $. Here, based on Eq. (\[eq:energyscale\]), the horizontal axis is chosen to be the scaled energy $ \tilde{\epsilon}=\epsilon/\sqrt{2P''(p_c)(P(p)-P(p_c))} $. We can see that if the soliton velocity is close to the critical one, numerically calculated points are well fitted to the universal curve (\[eq:scaledR\]). Thus, the theoretical results are well confirmed in this example. ![\[fig:RC\](Color online) Energy-dependence of reflection coefficient $ R $ of linearized excitations for various soliton velocities. Here we set $ (a_1,a_2)=(-1,1) $ and $ \rho_\infty=0.45 $. The critical velocity of the dark soliton is given by $ p_c = (0.206597\dots)\times p_{\text{L}}$. (See the lowest curve of Fig. \[fig:renP\].) A reflection coefficient of zero-energy phonon at the critical state is given by $ (d/b)^2\simeq 0.5718 $. Parabolic curves represent the theoretical approximate expression (\[eq:mainrR\]).](RC103.eps) ![\[fig:scaledR\](Color online) Scaling behavior of reflection coefficient $ R $. Here we set $ (a_1,a_2)=(-1,1) $ and $ \rho_\infty=0.45 $. $ \tilde{\epsilon}:=\epsilon/\sqrt{2P''(p_c)(P(p)-P(p_c))} $ is a scaled excitation energy. “Theory” represents the universal form of reflection coefficient (\[eq:scaledR\]).](scaledR002.eps) Proof – Step 1: Exact Zero-Energy Solutions {#sec:proof1} =========================================== In this and the next section, we prove the main result. This section is particularly devoted to the construction of exact zero-energy solutions. As an important tool, parameter derivatives are introduced. Parameter derivative {#subsec:paradera} -------------------- As seen in the asymptotic form (\[eq:nlsdsasym\]) or in the example of the CQNLS system in Subsec. \[subsec:cqnls\], the dark soliton solution has two parameters, i.e., $ (p,\rho_\infty) $. So we can consider two kinds of parameter derivatives:[^1] $ \partial_p\phi $ and $ \partial_{\rho_\infty}\phi $. We can use arbitrary coordinates to “label” the two-dimensional parameter space $ (\alpha,\beta)=(\alpha(p,\rho_\infty), \beta(p,\rho_\infty)) $, unless the Jacobian of coordinate transformation is singular. Obviously, the final result must not depend on the choice of coordinates. In order to make the story general, we always use these general parameter derivatives, and henceforth, we write the parameter derivative simply by the subscript, i.e., $ \phi_\alpha := \partial_\alpha\phi $ and $ \phi_\beta:=\partial_\beta\phi $. We also introduce the following symbol: $$\begin{aligned} [A,B]_{\alpha\beta}:=A_\alpha B_\beta-A_\beta B_\alpha. \end{aligned}$$ Note that the ratio $$\begin{aligned} \frac{[A,B]_{\alpha\beta}}{[C,D]_{\alpha\beta}} \label{eq:invratio} \end{aligned}$$ has a coordinate-free meaning, in other words, it is invariant under coordinate transformations of parameter space. We often construct coordinate-free solutions in such a ratio form.\ An immediate application of parameter derivatives is that one can obtain a particular solution of zero-energy Bogoliubov equation (\[eq:bogos\]). By differentiation of the stationary NLS eq. (\[eq:nls2\]), one obtains $$\begin{aligned} \mathcal{L}_\mu \begin{pmatrix} \phi_\alpha \\ -\phi^_\alpha \end{pmatrix} = \mu_\alpha \begin{pmatrix} \phi \\ \phi^ \end{pmatrix}. \label{eq:1storder} \end{aligned}$$ The same expression follows by replacement $ \alpha\rightarrow\beta $. Thus, taking the difference of double parameter derivatives, one can obtain the following zero-energy solution: $$\begin{aligned} \mathcal{L}_\mu \begin{pmatrix} [\mu,\phi]_{\alpha\beta} \\ -[\mu,\phi^]_{\alpha\beta} \end{pmatrix} = 0. \end{aligned}$$ It must be emphasized that this solution exists even when a localized potential barrier is added, in other words, when the fundamental equation loses a translational symmetry. What we only need is two kinds of parameter derivatives. So, this solution is not a symmetry-originated zero-mode. (For a symmetry consideration, see \[app:symmetry\].)\ Some technical (but crucially important) identities are derived in \[app:idnty\]. Equation (\[eq:1storder\]) will be used again in the process of energy expansions. Density fluctuation and phase fluctuation {#subsec:fg} ----------------------------------------- Here we introduce notations for the linearized density fluctuations and phase fluctuations, and rewrite the Bogoliubov equation with respect to these variables. They are convenient for both calculations and physical interpretations. Through the symbols $ (u,v)=(\delta\phi,-\delta\phi^) $, the density and phase fluctuations are expressed as $$\begin{aligned} \delta \rho &= \delta(\phi\phi^)=\delta\phi\phi^+\phi\delta\phi^=u\phi^-v\phi \\ \delta S &= \delta\left( \frac{1}{2\mathrm{i}}\log\frac{\phi}{\phi^} \right)=\frac{1}{2\mathrm{i}}\left( \frac{\delta\phi}{\phi}-\frac{\delta\phi^}{\phi^} \right)=\frac{1}{2\mathrm{i}}\left( \frac{u}{\phi}+\frac{v}{\phi^} \right) \end{aligned}$$ Therefore, if one defines $ f $ and $ g $ as $$\begin{aligned} \begin{cases} \displaystyle f=\frac{1}{2\mathrm{i}}\left( \frac{u}{\phi}+\frac{v}{\phi^} \right) \\ g=u\phi^-v\phi \end{cases} \quad\leftrightarrow\quad \begin{cases}\displaystyle u=\phi\left( \mathrm{i}f+\frac{g}{2\rho} \right) \\ \displaystyle v=\phi^\left( \mathrm{i}f-\frac{g}{2\rho} \right), \end{cases} \end{aligned}$$ then $ f $ has the meaning of the phase fluctuation, and $ g $ the density fluctuation. The stationary Bogoliubov equation (\[eq:bogos\]) are rewritten as follows: $$\begin{aligned} \left( \rho f'+\frac{jg}{\rho} \right)'&=\frac{\mathrm{i}\epsilon}{2}g, \label{eq:bogofg1}\\ \left( g'-\frac{\rho'}{\rho}g \right)'-2\rho F'(\rho)g-4jf'&=-2\mathrm{i}\epsilon\rho f \label{eq:bogofg2}. \end{aligned}$$ For the equation with $ \epsilon=0 $, we obtain the following integration constant: $$\begin{aligned} \rho f'+\frac{jg}{\rho}=\text{const.} \label{eq:zeroBogoconst} \end{aligned}$$ Exact zero-energy solutions {#subsec:zerosol} --------------------------- Let us derive all four linearly independent solutions for the zero-energy Bogoliubov equation (\[eq:bogos\]), or equivalently, Eqs. (\[eq:bogofg1\]) and (\[eq:bogofg2\]) with $ \epsilon=0 $. From global phase symmetry and translational symmetry, we can immediately find two zero-modes: $ (u,v)=(\mathrm{i}\phi,\mathrm{i}\phi^) $ and $ (\phi^{}_x,-\phi^_x) $. In addition to them, we already have the third solution in Subsec. \[subsec:paradera\]. It is well known that if we have $ n-1 $ linearly independent solutions for an $ n $-th order linear differential equation, the last one can be obtained by reduction of order. So, the fourth solution also can be found. See \[app:roo\] for a detailed calculation.\ Thus, all four zero-energy solutions are given by $$\begin{aligned} \begin{pmatrix}f_1 \\ g_1 \end{pmatrix}&=\begin{pmatrix}1 \\ 0 \end{pmatrix}, \\ \begin{pmatrix}f_2 \\ g_2 \end{pmatrix}&=\frac{1}{[\mu,j]_{\alpha\beta}}\begin{pmatrix}[\mu,S]_{\alpha\beta} \\ [\mu,\rho]_{\alpha\beta} \end{pmatrix},\\ \begin{pmatrix}f_3 \\ g_3 \end{pmatrix}&=\begin{pmatrix}S_x-p \\ \rho_x \end{pmatrix}=\begin{pmatrix}p(\rho_\infty-\rho)/\rho \\ \rho_x \end{pmatrix}, \label{eq:f3g3} \\ \begin{pmatrix}f_4 \\ g_4 \end{pmatrix}&= -\frac{\rho g_2}{\rho_\infty-\rho}\begin{pmatrix}0 \\ 1\end{pmatrix}+\left[\rho_\infty \int_0^x\!\!\frac{ g_2\mathrm{d}x}{(\rho_\infty-\rho)^2} \right]\begin{pmatrix} f_3 \\ g_3 \end{pmatrix}. \label{eq:zerosol4} \end{aligned}$$ For reference, we also write down the same solutions by $ (u,v) $ notation: $$\begin{aligned} \begin{pmatrix}u_1 \\ v_1 \end{pmatrix}&=\begin{pmatrix}\mathrm{i}\phi \\ \mathrm{i}\phi^ \end{pmatrix}, \\ \begin{pmatrix}u_2 \\ v_2 \end{pmatrix}&=\frac{1}{[\mu,j]_{\alpha\beta}}\begin{pmatrix} [\mu,\phi]_{\alpha\beta} \\ -[\mu,\phi^]_{\alpha\beta} \end{pmatrix},\\ \begin{pmatrix}u_3 \\ v_3 \end{pmatrix}&=\begin{pmatrix}\phi_x-\mathrm{i}p\phi \\ -\phi_x^-\mathrm{i}p\phi^* \end{pmatrix}, \label{eq:u3v3}\\ \begin{pmatrix}u_4 \\ v_4 \end{pmatrix}&= \frac{g_2}{2(\rho_\infty-\rho)}\begin{pmatrix}-\phi \\ \phi^\end{pmatrix}+\left[\rho_\infty \int_0^x\!\!\frac{ g_2\mathrm{d}x}{(\rho_\infty-\rho)^2} \right]\begin{pmatrix} u_3 \\ v_3 \end{pmatrix} \end{aligned}$$ Here $ g_2 = [\mu,\rho]_{\alpha\beta}/[\mu,j]_{\alpha\beta} = \mathrm{i}(u_1v_2-u_2v_1) $.\ $ (f_4,g_4) $ is chosen so that the integration constant (\[eq:zeroBogoconst\]) vanishes. So, only $ (f_2,g_2) $ contributes to this constant: $$\begin{aligned} \rho f_2'+\frac{jg_2}{\rho}&=1, \\ \rho f_i'+\frac{jg_i}{\rho}&=0, \quad (i=\text{1, 3, and 4.}). \end{aligned}$$ If we assume that the density $ \rho(x) $ is an even function, the parities and asymptotic behaviors of the four solutions are summarized as follows: $$\begin{aligned} (f_1,g_1)\, &: \ (\text{even},\text{odd}), \quad \text{bounded.} \label{eq:zerosolparity1} \\ (f_2,g_2)\, &: \ (\text{odd},\text{even}), \quad \text{linearly divergent.} \\ (f_3,g_3)\, &: \ (\text{even},\text{odd}), \quad \text{exponentially decreasing.} \\ (f_4,g_4)\, &: \ (\text{odd},\text{even}), \quad \text{exponentially increasing.} \label{eq:zerosolparity4} \end{aligned}$$ Keeping in mind the above (\[eq:zerosolparity1\])–(\[eq:zerosolparity4\]) will help to understand and imagine the construction of the solution in the next section. More detailed forms of asymptotes are given in \[sec:asymofzero\].\ Deriving all solutions including divergent ones might seem to be only of academic interest, but they will become necessary for the exact energy expansion, which will be performed in the next section.\ We note that four solutions are also obtained without using the parameter derivatives as shown in \[app:zerosol\]. However, those expressions are almost useless for solving a scattering problem. Proof – Step 2: Low-Energy Expansion {#sec:proof2} ==================================== In this section, with the use of zero-energy solutions derived in the preceding section, we construct finite-energy solutions up to second order by an exact energy expansion method. Using them, we solve the scattering problem of linearized excitations, and prove the main result. Exact low-energy expansion -------------------------- We construct the finite energy solution by energy expansion: $$\begin{aligned} \begin{pmatrix}u \\ v \end{pmatrix} = \sum_{n=0}^\infty \epsilon^n \begin{pmatrix} u^{(n)} \\ v^{(n)} \end{pmatrix}, \label{eq:expansion} \end{aligned}$$ or equivalently, $ (f,g) = \sum_{n=0}^\infty \epsilon^n (f^{(n)},g^{(n)}) $. The expansion only using asymptotic forms was used in Ref. [@KatoNishiwakiFujita], and the completely exact expansion was first used in Ref. [@TakahashiKato] to discuss tunneling behaviors of Bogoliubov phonons in the presence of potential walls. In order to prove the singular behavior at the critical point, the exact expansion is essentially necessary [@TakahashiKato; @TakahashiKatoConf].\ The substitution of (\[eq:expansion\]) into Bogoliubov equation (\[eq:bogos\]) yields $$\begin{aligned} \mathcal{L}_\mu\begin{pmatrix}u^{(n)} \\ v^{(n)} \end{pmatrix} = \begin{pmatrix}u^{(n-1)} \\ v^{(n-1)} \end{pmatrix}, \quad (n=1,2,\dots), \label{eq:expansion2} \end{aligned}$$ or in $ (f,g) $ expression, $$\begin{aligned} \left( \rho \left(f^{(n)}\right)'+\frac{jg^{(n)}}{\rho} \right)'&=\frac{\mathrm{i}}{2}g^{(n-1)}, \label{eq:recurf} \\ \left( \left(g^{(n)}\right)'-\frac{\rho'}{\rho}g^{(n)} \right)'-2\rho F'(\rho)g^{(n)}-4j\left(f^{(n)}\right)' &=-2\mathrm{i}\rho f^{(n-1)}. \label{eq:recurg} \end{aligned}$$ It is equivalent to the zero-energy Bogoliubov equation with an inhomogeneous term $ (u^{(n-1)},v^{(n-1)}) $. Since we already know all four solutions for the homogeneous equation, we can solve it by variation of parameters. Since we want the solution of the form (\[eq:uvasymptotic\]), in the process of energy expansions, we must cancel out the exponentially divergent terms, and must take up the power-law divergent terms, because $ \mathrm{e}^{\mathrm{i}k_ix}=1+\mathrm{i}k_ix-k_i^2x^2/2+\dotsb $ and $ k_1,k_2 \propto \epsilon $ for low-energy. (See eqs. (\[eq:kexpand1\]) and (\[eq:kexpand2\]).) First order solutions --------------------- The first order solutions are, in fact, directly calculated without the use of variation of parameters. When we choose $ (u^{(0)},v^{(0)})=(u_1,v_1)=(\mathrm{i}\phi,\mathrm{i}\phi^) $, Eq. (\[eq:1storder\]) gives the particular solution for $ (u^{(1)},v^{(1)}) $, that is, $$\begin{aligned} \begin{pmatrix}\tilde{u}_{\text{A}}^{(\text{finite }\epsilon)} \\ \tilde{v}_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} \mathrm{i}\phi \\ \mathrm{i}\phi^* \end{pmatrix} + \frac{\mathrm{i}\epsilon}{\mu_\alpha}\begin{pmatrix} \phi_\alpha \\ -\phi_\alpha^* \end{pmatrix}+O(\epsilon^2) \intertext{and} \begin{pmatrix}\tilde{u}_{\text{B}}^{(\text{finite }\epsilon)} \\ \tilde{v}_{\text{B}}^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} \mathrm{i}\phi \\ \mathrm{i}\phi^* \end{pmatrix} + \frac{\mathrm{i}\epsilon}{\mu_\beta}\begin{pmatrix} \phi_\beta \\ -\phi_\beta^* \end{pmatrix}+O(\epsilon^2) \end{aligned}$$ give two kinds of first order solutions of the finite energy solution. From Galilean symmetry of the NLS equation, we can also obtain the first order solution when we set $ (u^{(0)},v^{(0)})=(\phi_x,-\phi^_x) $. (See \[app:symmetry\]) It is given by $$\begin{aligned} \begin{pmatrix}\tilde{u}_3^{(\text{finite }\epsilon)} \\ \tilde{v}_3^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix} + \frac{\mathrm{i}\epsilon x}{2}\begin{pmatrix} \mathrm{i}\phi \\ \mathrm{i}\phi^* \end{pmatrix}+O(\epsilon^2). \label{eq:1stu3v3} \end{aligned}$$ We also write down them in $ (f,g) $ notation: $$\begin{aligned} \begin{pmatrix}\tilde{f}_{\text{A}}^{(\text{finite }\epsilon)} \\ \tilde{g}_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} 1 \\ 0 \end{pmatrix} + \frac{\mathrm{i}\epsilon}{\mu_\alpha}\begin{pmatrix} S_\alpha \\ \rho_\alpha \end{pmatrix}+O(\epsilon^2), \label{eq:firsta}\\ \begin{pmatrix}\tilde{f}_{\text{B}}^{(\text{finite }\epsilon)} \\ \tilde{g}_{\text{B}}^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} 1 \\ 0 \end{pmatrix} + \frac{\mathrm{i}\epsilon}{\mu_\beta}\begin{pmatrix} S_\beta \\ \rho_\beta \end{pmatrix}+O(\epsilon^2), \label{eq:firstb}\\ \begin{pmatrix}\tilde{f}_3^{(\text{finite }\epsilon)} \\ \tilde{g}_3^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix} S_x \\ \rho_x \end{pmatrix} + \frac{\mathrm{i}\epsilon x}{2}\begin{pmatrix} 1 \\ 0 \end{pmatrix}+O(\epsilon^2). \label{eq:first3} \end{aligned}$$ Second order calculation – Identification of bounded solutions {#subsec:boundedsols} -------------------------------------------------------------- Let us calculate the second order. Before beginning, we explain the outline of the calculation. It is important to notice that only two of four solutions are the bounded solutions, i.e., propagating waves for finite positive energy $ \epsilon $, as seen in Subsec. \[subsec:uniform\]. This means, in the energy expansion, that the two of four solutions are power-law divergent and the remaining two must be exponentially divergent. On the other hand, however, we have three linearly divergent solutions (\[eq:firsta\])–(\[eq:first3\]) in the previous subsection. So, one of the three must be a dummy. In fact, the solutions calculated from the variation of parameters shows that all three solutions diverge exponentially; The parity of the second-order exponential divergent term becomes $ (f^{(2)},g^{(2)})=(\text{even},\text{odd}) $, so it cannot be canceled out by adding the homogeneous solution $ (f_4,g_4) $. In order to kill out this exponential divergence, we must choose the special linear combinations of the three. Through this process, we have two power-law divergent solutions, as expected.\ Now, let us construct the solution of Eqs. (\[eq:recurf\]) and (\[eq:recurg\]) with $ n=2 $ by the above-mentioned method. In accordance with the method of variation of parameters, let the particular solution be $ (f^{(2)},g^{(2)})=\sum_{i=1,2,3,4}c_i(f_i,g_i) $, and let $ c_i $s have the $ x $-dependence. After a little calculation, the equations for $ c_i $s are summarized as follows: $$\begin{aligned} 2\mathrm{i}\begin{pmatrix}c_1' \\ c_2' \end{pmatrix} &= \begin{pmatrix} f_2 & -g_2 \\ -f_1 & g_1 \end{pmatrix}\begin{pmatrix} g^{(1)} \\ f^{(1)} \end{pmatrix}, \label{eq:vopc12}\\ 2\mathrm{i}p\begin{pmatrix}c_3' \\ c_4' \end{pmatrix} &= \begin{pmatrix} f_4 & -g_4 \\ -f_3 & g_3 \end{pmatrix}\begin{pmatrix} g^{(1)} \\ f^{(1)} \end{pmatrix}. \label{eq:vopc34} \end{aligned}$$ Let us assume that $ f^{(1)} $ is a linearly divergent odd function and $ g^{(1)} $ is a bounded even function. From the asymptotic behaviors (\[eq:zerosolparity1\])–(\[eq:zerosolparity4\]) and \[sec:asymofzero\], we can see that $ c_1(f_1,g_1),\,c_2(f_2,g_2) $, and $ c_3(f_3,g_3) $ are always power-law divergent functions. $ c_4 $ is an asymptotically constant function, so the product $ c_4(f_4,g_4) $ diverges exponentially in general. Exceptionally, if $ c_4 $ decays exponentially, $ c_4(f_4,g_4) $ becomes a power-law divergent function. Therefore, in order for the particular solution to be a power-law divergent function, $ \lim_{\|x\|\rightarrow\infty}c_4=0 $, or equivalently, $$\begin{aligned} \int_{-\infty}^{\infty}c_4'\mathrm{d}x=0 \quad\leftrightarrow\quad \int_{-\infty}^{\infty} \left(f_3 g^{(1)}-g_3 f^{(1)}\right)\mathrm{d}x=0 \label{eq:nondiverge} \end{aligned}$$ must hold.\ Let us construct the solution that satisfies the above condition (\[eq:nondiverge\]). Consider the following linear combination: $$\begin{aligned} \begin{split} \begin{pmatrix}f_{\text{A}}^{(\text{finite }\epsilon)} \\ g_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix} &:= \begin{pmatrix}\tilde{f}_{\text{A}}^{(\text{finite }\epsilon)} \\ \tilde{g}_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix}+2\xi_{\text{A}}\begin{pmatrix}\tilde{f}_3^{(\text{finite }\epsilon)} \\ \tilde{g}_3^{(\text{finite }\epsilon)} \end{pmatrix} \\ &=\begin{pmatrix} 1+2\xi_{\text{A}} S_x \\ 2\xi_{\text{A}} \rho_x \end{pmatrix} + \frac{\mathrm{i}\epsilon}{\mu_\alpha}\begin{pmatrix} S_\alpha+\mu_\alpha\xi_{\text{A}}x \\ \rho_\alpha \end{pmatrix}+O(\epsilon^2). \end{split} \end{aligned}$$ $ \xi_{\text{A}} $ which satisfies Eq. (\[eq:nondiverge\]) is determined as $$\begin{aligned} &\int_{-\infty}^\infty\left[ f_3\rho_\alpha-g_3(S_\alpha+\mu_\alpha\xi_{\text{A}}x) \right]\mathrm{d}x=0 \\ \leftrightarrow \ & \xi_{\text{A}} = \frac{\rho_\infty\delta_\alpha+p N_\alpha}{N \mu_\alpha}. \end{aligned}$$ The same calculation can be performed with respect to the other parameter $ \beta $, so $ (f_{\text{B}}^{(\text{finite }\epsilon)},g_{\text{B}}^{(\text{finite }\epsilon)}) $ and $ \xi_{\text{B}} $ are defined in the same manner.\ Thus, we have two non-exponentially-divergent solutions $ (f_{\text{A}}^{(\text{finite }\epsilon)},g_{\text{A}}^{(\text{finite }\epsilon)}) $ and $ (f_{\text{B}}^{(\text{finite }\epsilon)},g_{\text{B}}^{(\text{finite }\epsilon)}) $. Furthermore, we can get coordinate-free solutions by making the linear combinations of these two. They are defined as follows: $$\begin{aligned} \begin{pmatrix} f_1^{(\text{finite }\epsilon)} \\ g_1^{(\text{finite }\epsilon)} \end{pmatrix} &:= \frac{1}{\xi_{\text{A}}-\xi_{\text{B}}}\left[ \xi_{\text{A}}\begin{pmatrix}f_{\text{B}}^{(\text{finite }\epsilon)} \\ g_{\text{B}}^{(\text{finite }\epsilon)} \end{pmatrix}-\xi_{\text{B}}\begin{pmatrix}f_{\text{A}}^{(\text{finite }\epsilon)} \\ g_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix} \right], \\ \begin{pmatrix} f_3^{(\text{finite }\epsilon)} \\ g_3^{(\text{finite }\epsilon)} \end{pmatrix} &:= \frac{1}{2(\xi_{\text{A}}-\xi_{\text{B}})}\left[ \begin{pmatrix}f_{\text{A}}^{(\text{finite }\epsilon)} \\ g_{\text{A}}^{(\text{finite }\epsilon)} \end{pmatrix}-\begin{pmatrix}f_{\text{B}}^{(\text{finite }\epsilon)} \\ g_{\text{B}}^{(\text{finite }\epsilon)} \end{pmatrix} \right]-p\begin{pmatrix} f_1^{(\text{finite }\epsilon)} \\ g_1^{(\text{finite }\epsilon)} \end{pmatrix}. \end{aligned}$$ If we write $$\begin{aligned} \begin{pmatrix} f_i^{(\text{finite }\epsilon)} \\ g_i^{(\text{finite }\epsilon)} \end{pmatrix} = \sum_{n=0}^\infty \epsilon^n \begin{pmatrix} f_i^{(n)} \\ g_i^{(n)} \end{pmatrix} \quad (i=\text{1 or 3}), \end{aligned}$$ then the zeroth and first order terms are given by $$\begin{aligned} \begin{pmatrix} f_1^{(0)} \\ g_1^{(0)} \end{pmatrix} &= \begin{pmatrix} f_1 \\ g_1 \end{pmatrix} = \begin{pmatrix} 1 \\ 0 \end{pmatrix}, \\ \begin{pmatrix} f_1^{(1)} \\ g_1^{(1)} \end{pmatrix} &= \frac{\mathrm{i}}{\rho_\infty[\delta,\mu]_{\alpha\beta}+p[N,\mu]_{\alpha\beta}}\begin{pmatrix} \rho_\infty[\delta,S]_{\alpha\beta}+p[N,S]_{\alpha\beta} \\ \rho_\infty[\delta,\rho]_{\alpha\beta}+p[N,\rho]_{\alpha\beta} \end{pmatrix}, \label{eq:f11g11}\\ \begin{pmatrix} f_3^{(0)} \\ g_3^{(0)} \end{pmatrix} &= \begin{pmatrix} f_3 \\ g_3 \end{pmatrix} = \begin{pmatrix} S_x-p \\ \rho_x \end{pmatrix}, \\ \begin{pmatrix} f_3^{(1)} \\ g_3^{(1)} \end{pmatrix} &= \frac{\mathrm{i}x}{2}\begin{pmatrix}1 \\ 0 \end{pmatrix}-\frac{\mathrm{i}N[\mu,j]_{\alpha\beta}}{2(\rho_\infty[\delta,\mu]_{\alpha\beta}+p[N,\mu]_{\alpha\beta})}\begin{pmatrix} f_2 \\ g_2 \end{pmatrix}-p\begin{pmatrix} f_1^{(1)} \\ g_1^{(1)} \end{pmatrix}\label{eq:f31g31}, \end{aligned}$$ and the second order terms are, from Eqs. (\[eq:vopc12\]) and (\[eq:vopc34\]), given by $$\begin{aligned} \begin{split} \begin{pmatrix} f_i^{(2)} \\ g_i^{(2)} \end{pmatrix} &= \left[ \frac{1}{2\mathrm{i}}\int_0^x\left(f_2g_i^{(1)}-g_2f_i^{(1)}\right)\mathrm{d}x \right]\begin{pmatrix}f_1 \\ g_1 \end{pmatrix}\\ &\quad+\left[ \frac{-1}{2\mathrm{i}}\int_0^xg_i^{(1)}\mathrm{d}x \right]\begin{pmatrix}f_2 \\ g_2 \end{pmatrix} \\ &\qquad+\left[ \frac{1}{2\mathrm{i}p}\int_0^x\left(f_4g_i^{(1)}-g_4f_i^{(1)}\right)\mathrm{d}x \right]\begin{pmatrix}f_3 \\ g_3 \end{pmatrix} \\ &\qquad\quad+\left[ \frac{-1}{2\mathrm{i}p}\int_0^x\left(f_3g_i^{(1)}-g_3f_i^{(1)}\right)\mathrm{d}x \right]\begin{pmatrix}f_4 \\ g_4 \end{pmatrix} \end{split}\label{eq:2ndbound} \end{aligned}$$ with $ i=\text{1 or 3} $.\ Every bounded solution for finite positive energy $ \epsilon $ must be constructed as a linear combination of $ (f_1^{(\mathrm{finite \ }\epsilon)}, g_1^{(\mathrm{finite \ }\epsilon)}) $ and $ (f_3^{(\mathrm{finite \ }\epsilon)}, g_3^{(\mathrm{finite \ }\epsilon)}) $, so must the solution of the scattering problem (\[eq:uvasymptotic\]). Our remaining work is to calculate their asymptotic behavior. Second order calculation – Asymptotics {#subsec:2ndasym} -------------------------------------- The evaluation of asymptotic form of the second order term (\[eq:2ndbound\]) is tedious but straightforward. All calculations can be carried out by using the expressions in \[sec:asymofzero\]. For brevity, we introduce the following symbols for the asymptotic forms of $ f^{(1)}_i $ and $ \int_0^x g^{(1)}_i\mathrm{d}x $ $ (i=1\text{ or }3.)$: $$\begin{aligned} f^{(1)}_i &\rightarrow \mathrm{i}(l_{i1}x+l_{i0}\operatorname{sgn}x), \label{eq:f1asym} \\ \int_0^xg^{(1)}_i\mathrm{d}x &\rightarrow \mathrm{i}(m_{i1}x+m_{i0}\operatorname{sgn}x). \label{eq:g1asym} \end{aligned}$$ Explicit expressions for $ l_{ij}s $ and $ m_{ij} $s are given in \[app:asymptote\]. The asymptotic form up to second order is given by $$\begin{aligned} \begin{split} &\begin{pmatrix} f_i^{(\text{finite }\epsilon)} \\ g_i^{(\text{finite }\epsilon)} \end{pmatrix} \rightarrow \begin{pmatrix} \delta_{i1} \\ 0 \end{pmatrix} + \mathrm{i}\epsilon \begin{pmatrix} l_{i1}x+l_{i0}\operatorname{sgn}x \\ m_{i1} \end{pmatrix} \\ &\quad-\frac{\epsilon^2}{2\kappa^2}\begin{pmatrix}\frac{1}{\rho_\infty}\Bigl[ (p_{\text{L}}^2m_{i1}-jl_{i1})\frac{x^2}{2}+(p_{\text{L}}^2m_{i0}-jl_{i0})\|x\|\Bigr]\!+\!c_{i0}\! \\ (\rho_\infty l_{i1}-pm_{i1})x+(\rho_\infty l_{i0}-pm_{i0})\operatorname{sgn}x \end{pmatrix}+O(\epsilon^3). \end{split} \end{aligned}$$ Here $ c_{i0} $ is a certain constant whose form is not important here.\ Let $ (u^{(\text{finite }\epsilon)}_i,v^{(\text{finite }\epsilon)}_i) $ be the counterpart of $ (f^{(\text{finite }\epsilon)}_i,g^{(\text{finite }\epsilon)}_i) $ in $ (u,v) $ notation. If we write them in plane wave form $$\begin{aligned} \begin{pmatrix} u^{(\text{finite }\epsilon)}_1 \\ v^{(\text{finite }\epsilon)}_1 \end{pmatrix} &\rightarrow \frac{\mathrm{i}\sqrt{\rho_\infty}}{2pp_{\text{L}}}\left[ C_1^{\pm}w_1\left( x,\pm\tfrac{\delta}{2} \right)+C_2^{\pm}w_2\left( x,\pm\tfrac{\delta}{2} \right) \right] \quad (x\rightarrow\pm\infty) \\ \intertext{and} \begin{pmatrix} u^{(\text{finite }\epsilon)}_3 \\ v^{(\text{finite }\epsilon)}_3 \end{pmatrix} &\rightarrow \frac{\mathrm{i}\sqrt{\rho_\infty}}{2pp_{\text{L}}}\left[ D_1^{\pm}w_1\left( x,\pm\tfrac{\delta}{2} \right)+D_2^{\pm}w_2\left( x,\pm\tfrac{\delta}{2} \right) \right] \quad (x\rightarrow\pm\infty), \end{aligned}$$ then the energy dependence of coefficients can be obtained as follows: $$\begin{aligned} C_1^\pm &= p(p+p_{\text{L}}+2\kappa^2l_{11})\pm\mathrm{i}\epsilon p_{\text{L}}(p-p_{\text{L}})l_{10}+O(\epsilon^2), \label{eq:C1} \\ C_2^\pm &= p(-p+p_{\text{L}}-2\kappa^2l_{11})\pm\mathrm{i}\epsilon p_{\text{L}}(p+p_{\text{L}})l_{10}+O(\epsilon^2), \\ D_1^\pm &= 2p\kappa^2 l_{31}\pm\mathrm{i}\epsilon p_{\text{L}}\left( \frac{p_{\text{L}}N}{4\rho_\infty}+(p-p_{\text{L}})l_{30} \right)+O(\epsilon^2), \\ D_2^\pm &= -2p\kappa^2 l_{31}\pm\mathrm{i}\epsilon p_{\text{L}}\left( -\frac{p_{\text{L}}N}{4\rho_\infty}+(p+p_{\text{L}})l_{30} \right)+O(\epsilon^2). \label{eq:D2} \end{aligned}$$ The striking feature is that the zeroth order of $ D_1^\pm $ and $ D_2^\pm $ vanishes when the soliton velocity becomes the critical value, see Eq. (\[eq:l31\]). This is an immediate cause of the singular behavior of the reflection coefficient.\ A solution of the scattering problem (\[eq:uvasymptotic\]) is constructed as follows: $$\begin{aligned} \begin{pmatrix}u \\ v \end{pmatrix}=D_2^+\begin{pmatrix} u^{(\text{finite }\epsilon)}_1 \\ v^{(\text{finite }\epsilon)}_1 \end{pmatrix}-C_2^+\begin{pmatrix} u^{(\text{finite }\epsilon)}_3 \\ v^{(\text{finite }\epsilon)}_3 \end{pmatrix}, \end{aligned}$$ and coefficients $ t $ and $ r $ are given by $$\begin{aligned} t&=\frac{C_2^+D_1^+-D_2^+C_1^+}{C_2^+D_1^--D_2^+C_1^-}, \\ r&=\frac{C_2^+D_2^--D_2^+C_2^-}{C_2^+D_1^--D_2^+C_1^-}. \end{aligned}$$ Finally, moving to the particular coordinate system $ (\alpha,\beta)=(p,\rho_\infty) $, and using the expressions given in \[app:asymptote\], we get the main result $$\begin{aligned} -\frac{2\rho_\infty^2\delta_p}{p^2p_{\text{L}}^2}(C_2^+D_1^--D_2^+C_1^-) &= a P_p-\mathrm{i}\epsilon\left( b+b_1P_p \right)+O(\epsilon^2) \label{eq:final1}\\ \intertext{and} -\frac{2\rho_\infty^2\delta_p}{p^2p_{\text{L}}^2}(C_2^+D_2^--D_2^+C_2^-) &= -\mathrm{i}\epsilon\left( d+d_1P_p \right)+O(\epsilon^2) \label{eq:final2} \end{aligned}$$ with $$\begin{aligned} a &= 4p_{\text{L}}\rho_\infty, \\ b &= (N+pN_p)^2+(p_{\text{L}}N_p)^2, \\ b_1 &= N-\left( 1+\frac{2p^2}{\kappa^2} \right)[N,j]_{p\rho_\infty}, \\ d &= (N+pN_p)^2-(p_{\text{L}}N_p)^2, \\ d_1 &= N+[N,j]_{p\rho_\infty}. \label{eq:final7} \end{aligned}$$ It gives the result in Subsec. \[subsec:mainresult\]. Discussions and Concluding Remarks {#sec:summary} ================================== In this last section, we give a few discussions and future perspectives. Local density fluctuation at the critical point ----------------------------------------------- In the system with a potential wall, the emergence of a zero-energy local density fluctuation was a key of the destabilization of superflow[@TakahashiKato; @TakahashiKatoConf]. In the present case of solitons, an amplification of the zero-energy local density fluctuation also occurs at the critical point, but its mathematical structure slightly differs. Let us see it in detail.\ Before beginning, one should recall that $ f $ and $ g $ have the meaning of phase and density fluctuations. (See subsection \[subsec:fg\] again.) So, the solution $ (u,v)\propto(\phi,\phi^) \,\leftrightarrow\, (f,g)\propto(1,0) $ has no density fluctuation. If another solution, e.g., a parameter-derivative solution, is added, a non-zero density fluctuation arises.\ For a superflow state against the potential barrier[@BaratoffBlackburnSchwartz; @Hakim; @PhamBrachet], it is known[@TakahashiKato; @TakahashiKatoConf] that $$\begin{aligned} \lim_{\epsilon\rightarrow0}\begin{pmatrix} u \\ v \end{pmatrix} &= \begin{pmatrix} \phi \\ \phi^ \end{pmatrix} & \text{(for non-critical states)}. \\ \lim_{\epsilon\rightarrow0}\begin{pmatrix} u \\ v \end{pmatrix} &= \begin{pmatrix} \phi \\ \phi^* \end{pmatrix}+c \dfrac{\partial }{\partial \varphi}\begin{pmatrix} \phi \\ -\phi^* \end{pmatrix} & \text{(for a critical state)}. \end{aligned}$$ Here $ \varphi $ is a Josephson phase difference and $ c $ is a certain constant. Thus, the density fluctuation represents the anomaly of the critical point.\ In the case of solitons, however, because of spontaneous translational symmetry breaking, the density fluctuating zero-mode $ (f_3,g_3) $ (or equivalently, $ (u_3,v_3) $ ) always exists, and this mode indeed contributes to the solution of the scattering problem: $$\begin{aligned} \lim_{\epsilon\rightarrow0}\begin{pmatrix} u \\ v \end{pmatrix} = c_1 \begin{pmatrix} u_1 \\ v_1 \end{pmatrix} +c_3 \begin{pmatrix} u_3 \\ v_3 \end{pmatrix} \quad \text{(for non critical states).} \end{aligned}$$ Here $ c_1=\lim_{\epsilon\rightarrow0}(-C_2^+) $ and $ c_3=\lim_{\epsilon\rightarrow0}D_2^+ $ are constants. (See Subsec. \[subsec:2ndasym\] for more detailed expressions.) Therefore, the local density fluctuation always exists regardless of whether the soliton is stable or unstable. However, when the soliton velocity comes closer to the critical one, the ratio $ c_3/c_1 $ becomes larger, and it becomes infinite at the critical point. That is to say, at the critical velocity state, $$\begin{aligned} \lim_{\epsilon\rightarrow0}\begin{pmatrix} u \\ v \end{pmatrix} = \begin{pmatrix} u_3 \\ v_3 \end{pmatrix} \quad \text{(for a critical state)} \end{aligned}$$ holds. Thus, we can say that the amplification of the local density fluctuation also plays a key role in the case of the destabilization of solitons. Conclusions and Future perspectives ----------------------------------- In this paper, we have solved the scattering problem of linearized excitations (Bogoliubov phonons) against a dark soliton in a generalized NLS system. We have exactly shown that the perfect transmission of a zero-energy phonon vanishes when the soliton velocity reaches the critical value, and near the critical velocity state, the reflection coefficient obeys a saddle-node type universal scaling law. Our result has a fundamental importance because it provides an exact example of saddle-node scaling in infinite dimensional time-reversible Hamiltonian systems. Through the proof, we have also obtained the exact zero-energy solutions and their finite energy generalizations. In the derivation of them, the use of two kinds of parameter derivatives has played an important role. This method will also be useful to elucidate the low-energy physics of other systems, such as higher dimensional systems or multi-component systems.\ Even though we have shown the example of scaling laws, the derivation of a normal form of saddle-node bifurcation remains unsolved. The similar problem for the supercurrent-flowing system in the presence of an obstacle[@Hakim; @PhamBrachet] also exists. Compared to the problem of solitons treated in this paper, the system with an obstacle seems to be a little more difficult, because all four zero-energy solutions for linearized equation are not yet obtained except for the critical velocity state[@TakahashiKato]. These issues are left as future works.\ In this paper we have treated the soliton-phonon scattering problem. One interesting generalization is the multi-soliton scattering process. It was shown that the existence of the two solitons makes the reflection coefficient of phonons non-trivial even for the integrable cubic case[@KolbyshovaSadreev]. The soliton collision problem in the non-integrable generalized KdV equation was recently investigated[@MartelMerleInventMath; @MartelMerleAnnMath]. The similar problem of dark solitons in non-integrable NLS systems is also important to understand the generic solitary characters of dark solitons.\ An emergence of singularity of scattering properties at a critical state which separates the stable and unstable branches is expected to be a universal character in more general systems, because the emergent or amplified zero-modes can affect the transmission properties of low-energy modes. Since the scattering problem of linearized excitations is easier and more analytically tractable than the existence-proof of an unstable mode or construction of a Lyapunov function, it will be useful to “conjecture” the criterion for stability of solitons, even though it is not a direct proof of the stability itself. For example, to the best knowledge of the present author, the stability criterion of dark solitons in multi-component NLS systems is an untouched problem. In ultracold atoms, the binary mixture of Rb atoms is realized[@HoShenoy; @Myatt] and it is known that the corresponding coupled NLS equation becomes integrable only for the Manakov case. Also, the Bose condensates with spin degree of freedom are created by optical trap[@StamperKurn], and Wadati group members have studied the spin-1 solitons at the integrable point[@TsuchidaWadati; @IedaMiyakawaWadati; @UchiyamaIedaWadati]. When the coupling constant deviates from the integrable one, it is no longer ensured that a dark soliton is always stable. So, it may be an interesting problem to study the stability of solitons through the study of scattering problems of linearized excitations.\ #### Acknowledgment The author is grateful to Y. Kato and M. Kunimi for helpful comments. This work was supported by a Grant-in-Aid for JSPS Fellows (No. 22-10058). Identities for parameter derivatives {#app:idnty} ==================================== Here we derive a few identities for parameter derivatives. The NLS equation (\[eq:nls2\]) expressed in terms of the density $ \rho(x) $ is given by $$\begin{aligned} \rho_{xx}=2\rho\left( \frac{\rho_x^2}{4\rho^2}+\frac{j^2}{\rho^2}-\mu+F(\rho) \right). \end{aligned}$$ From the parameter derivative of Eq. (\[eq:momconsrv\]) and the above equation, one obtains $$\begin{aligned} \rho_\alpha\rho_{xx}-\rho_{\alpha x}\rho_x = 2(\rho_\infty-\rho)(2p j_\alpha-\mu_\alpha \rho). \label{eq:syspar} \end{aligned}$$ Here recall that $ j=\rho_\infty p $ and $ \mu = p^2+F(\rho_\infty) $. The same expression follows by replacement $ \alpha\rightarrow\beta $. From them, an important identity $$\begin{aligned} g_3g_2'-g_2g_3' = -4p(\rho_\infty-\rho) \label{eq:wsimplify} \end{aligned}$$ follows, where $ g_2 $ and $ g_3 $ are defined in Subsec. \[subsec:zerosol\].\ Dividing both sides of (\[eq:syspar\]) by $ \rho $, and integrating them from $ -\infty $ to $ +\infty $, we obtain $$\begin{aligned} 2j_\alpha\delta+\mu_\alpha N=-\frac{\partial }{\partial \alpha}\int_{-\infty}^\infty\frac{(\rho_x)^2\mathrm{d}x}{2\rho}. \end{aligned}$$ Since the same expression for $ \beta $ holds, we obtain the second important identity $$\begin{aligned} \begin{split} & (2j_\alpha\delta+\mu_\alpha N)_\beta=(2j_\beta\delta+\mu_\beta N)_\alpha \\ \leftrightarrow\quad & 2[j,\delta]_{\alpha\beta}+[\mu,N]_{\alpha\beta}=0. \label{eq:idntydev} \end{split} \end{aligned}$$ Derivation of the fourth zero-energy solution {#app:roo} ============================================= Let us derive the expression for the fourth zero-energy solution $ (f_4,g_4) $. Eliminating $ f $ from Eqs. (\[eq:bogofg1\]) and (\[eq:bogofg2\]), one obtains the third order differential equation for $ g $: $$\begin{aligned} g'''+A g'+ B=0. \end{aligned}$$ Here $ A $ and $ B $ are some functions, but their forms are not important here. Knowing two solutions $ g_2 $ and $ g_3 $, the last one is obtained by reduction of order: $$\begin{aligned} \tilde{g}_4 &:= g_2 \int_0^x\frac{g_3\mathrm{d}x}{w^2}-g_3\int_0^x\frac{g_2\mathrm{d}x}{w^2} \intertext{with} w &= g^{}_3g_2'-g^{}_2g_3'. \end{aligned}$$ Since $ w $ is given by (\[eq:wsimplify\]), the above expression is simplified as $$\begin{aligned} 16p^2\tilde{g}_4 = \left[ \frac{1}{\rho_\infty-\rho}-\frac{1}{\rho_\infty-\rho_0} \right]g_2- g_3\int_0^x\frac{g_2\mathrm{d}x}{(\rho_\infty-\rho)^2} \end{aligned}$$ Here $ \rho_0:=\rho(0) $. The solution which cancels out the constant (\[eq:zeroBogoconst\]) is given by the following linear combination with $ g_2 $: $$\begin{aligned} g_4:=-\frac{\rho_0}{\rho_\infty-\rho_0}g_2-16p^2\rho_\infty \tilde{g}_4, \end{aligned}$$ and $ f_4 $ is calculated from $ f_4' = -jg_4/\rho^2 $. Asymptotics of zero-energy solutions {#sec:asymofzero} ==================================== Let us write the asymptotic form of $ \rho(x) $ as $$\begin{aligned} \rho(x) = \rho_\infty - \rho_\infty^{(1)}\mathrm{e}^{-2\kappa\|x\|}+\dotsb, \end{aligned}$$ then a half of the healing length $ \kappa $ is given by $$\begin{aligned} \kappa=\sqrt{p_{\text{L}}^2-p^2}. \end{aligned}$$ From (\[eq:nlsdsasym\]) or (\[eq:phasecond\]), an asymptotic form of the phase $ S(x) $ is given by $$\begin{aligned} S(x) \rightarrow px+(\operatorname{sgn}x)\frac{\delta}{2}. \end{aligned}$$ Using the invariant nature of the ratio (\[eq:invratio\]), one can show $$\begin{aligned} \frac{[\mu,p]_{\alpha\beta}}{[\mu,j]_{\alpha\beta}} &= \frac{[\mu,p]_{p\rho_\infty}}{[\mu,j]_{p\rho_\infty}} = \frac{p_{\text{L}}^2}{\rho_\infty \kappa^2}, \\ \frac{[\mu,\rho_\infty]_{\alpha\beta}}{[\mu,j]_{\alpha\beta}} &= \frac{[\mu,\rho_\infty]_{p\rho_\infty}}{[\mu,j]_{p\rho_\infty}} = -\frac{p}{\kappa^2}. \end{aligned}$$ With the use of the above relations, the asymptotic forms of zero energy solutions are evaluated as follows: $$\begin{aligned} f_2 & \rightarrow \frac{p_{\text{L}}^2}{\rho_\infty \kappa^2}x+(\operatorname{sgn}x)\frac{1}{2}\frac{[\mu,\delta]_{\alpha\beta}}{[\mu,j]_{\alpha\beta}}, \\ g_2 & \rightarrow -\frac{p}{\kappa^2}, \\ f_3 & \rightarrow \frac{p\rho_\infty^{(1)}}{\rho_\infty}\mathrm{e}^{-2\kappa\|x\|}, \\ g_3 & \rightarrow (\operatorname{sgn}x) 2\kappa \rho_\infty^{(1)}\mathrm{e}^{-2\kappa\|x\|}, \\ f_4 & \rightarrow -(\operatorname{sgn}x)\frac{p^2}{4\rho_\infty^{(1)}\kappa^3}\mathrm{e}^{2\kappa\|x\|}, \\ g_4 & \rightarrow \frac{j}{2\rho_\infty^{(1)}\kappa^2}\mathrm{e}^{2\kappa\|x\|}. \end{aligned}$$ Symmetry consideration on zero-modes {#app:symmetry} ==================================== In this appendix we consider how a symmetry plays a role in finding a solution of a linearized equation. Particularly, we emphasize that the information we can obtain from a Galilean symmetry is not about the zero energy solution but about the first order correction of a finite energy solution.\ Let $ \phi(x,t) $ be a solution of time-dependent NLS equation (\[eq:nls\]) and $ \tilde{\phi}(x,t,\alpha) $ be a family of solutions with continuous parameter $ \alpha $ such that $ \tilde{\phi}(x,t,0)=\phi(x,t) $. Here we assume that $ \alpha $ is free from any system parameter. (The parameter derivative introduced in Subsec. \[subsec:paradera\] is a more general concept, because it can depend on the system parameters which appear, e.g., in a boundary condition (\[eq:nlsdsasym\]).) Differentiation of NLS equation (\[eq:nls\]) with respect to $ \alpha $ immediately yields $$\begin{aligned} (\mathrm{i}\partial_t-\mathcal{L})\begin{pmatrix} \phi_\alpha \\ -\phi^_\alpha \end{pmatrix} = 0 \end{aligned}$$ with a definition $ \phi_\alpha:= [\partial_\alpha\tilde{\phi}]_{\alpha=0} $. Thus, $ \phi_\alpha $ is a solution of a time-dependent* linearized equation (\[eq:tdbogo\]) in the presence of the condensate wavefunction $ \phi(x,t) $. Particularly, if one sets $ \tilde{\phi}(x,t,\alpha)=\phi(x,t)\mathrm{e}^{\mathrm{i}\alpha} $, $ \tilde{\phi}(x,t,\alpha)=\phi(x+\alpha,t) $, and Eq. (\[eq:galilei\]), which represent a global phase symmetry, a translational symmetry, and a Galilean symmetry, respectively, one obtains the following particular solutions: $$\begin{aligned} (\mathrm{i}\partial_t-\mathcal{L})\begin{pmatrix} \mathrm{i}\phi \\ \mathrm{i}\phi^* \end{pmatrix} &= 0, \\ (\mathrm{i}\partial_t-\mathcal{L})\begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix} &= 0, \\ (\mathrm{i}\partial_t-\mathcal{L})\begin{pmatrix} 2t\phi_x-\mathrm{i}x\phi \\ -2t\phi^_x-\mathrm{i}x\phi^* \end{pmatrix} &= 0. \end{aligned}$$ It is a result for a time-dependent equation. In order to interpret these results to a stationary problem, let us set $ \phi(x,t)=\phi(x)\mathrm{e}^{-\mathrm{i}\mu t} $. From the assumption stated above, $ \mu $ does not depend on $ \alpha $. (On the other hand, $ \mu $ can depend on the parameter in Subsec. \[subsec:paradera\].) The above equations are rewritten as $$\begin{aligned} \mathcal{L}_\mu\begin{pmatrix} \mathrm{i}\phi \\ \mathrm{i}\phi^* \end{pmatrix} &= 0, \label{eq:appsymzero1} \\ \mathcal{L}_\mu\begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix} &= 0, \label{eq:appsymzero2} \\ 2\mathrm{i}\begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix}-2t\mathcal{L}_\mu\begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix}+\mathcal{L}_\mu\begin{pmatrix} \mathrm{i}x\phi \\ \mathrm{i}x\phi^ \end{pmatrix} &=0. \label{eq:appsymzero3} \end{aligned}$$ Here $ \mathcal{L}_\mu $ defined by Eqs. (\[eq:bogos\]) and (\[eq:bogos2\]) is a differential operator of the stationary Bogoliubov equation. (\[eq:appsymzero1\]) and (\[eq:appsymzero2\]) immediately give zero-energy solutions, whereas we need a consideration on Eq. (\[eq:appsymzero3\]). Since this equation must hold for any time $ t $, both coefficients of $ t^0 $ and $ t^1 $ must vanish. The equation for the $ t^1 $-coefficient reproduces (\[eq:appsymzero2\]). From the $ t^0 $-coefficient, one obtains $$\begin{aligned} -\frac{1}{2}\mathcal{L}_\mu\begin{pmatrix} x\phi \\ x\phi^* \end{pmatrix} = \begin{pmatrix} \phi_x \\ -\phi^_x \end{pmatrix}. \end{aligned}$$ It represents Eq. (\[eq:expansion2\]) with $ n=1 $ and $ (u^{(0)},v^{(0)})=(\phi_x,-\phi^_x) $. Therefore, it gives the first order solution of (\[eq:1stu3v3\]). Thus, a Galilean symmetry gives us a piece of information about first order solutions, and gives no new information about zero-energy solutions. A linearly divergent zero-energy solution, which exists even when the system does not have a translational or a Galilean symmetry, can be obtained by two kinds of parameter-derivatives stated in Subsec. \[subsec:paradera\].\ We further note that Eq. (\[eq:1stu3v3\]) itself does not describe a physically meaningful non-divergent solution. As we show in Subsec. \[subsec:boundedsols\], we must construct a linear combination of solutions so that the second order term is free from exponential divergence. A correct first order solution with non-divergent character is given by (\[eq:f31g31\]). Formulae for calculation of asymptotes {#app:asymptote} ====================================== Let us derive some formulae for $ l_{ij} $s and $ m_{ij} $s defined by Eqs. (\[eq:f1asym\]) and (\[eq:g1asym\]). Using the relations $$\begin{aligned} [X,j]_{\alpha\beta} &= \rho_\infty[X,p]_{\alpha\beta}+p[X,\rho_\infty]_{\alpha\beta}, \label{eq:xj}\\ [X,\mu]_{\alpha\beta} &= 2p[X,p]_{\alpha\beta}+F'(\rho_\infty)[X,\rho_\infty]_{\alpha\beta}, \label{eq:xmu} \end{aligned}$$ one can show $$\begin{aligned} \rho_\infty[X,\mu]_{\alpha\beta}-2p[X,j]_{\alpha\beta}&=2\kappa^2[X,\rho_\infty]_{\alpha\beta}. \label{eq:xjxmu} \end{aligned}$$ Here remember that $ p_{\text{L}} $ and $ F'(\rho_\infty) $ are related to each other by (\[eq:defofpl\]), and $ \kappa^2=p_{\text{L}}^2-p^2 $. Using it and (\[eq:idntydev\]), one obtains $$\begin{aligned} \rho_\infty[\delta,\mu]_{\alpha\beta}+p[N,\mu]_{\alpha\beta}&=2\kappa^2[\delta,\rho_\infty]_{\alpha\beta}. \label{eq:deltamu} \end{aligned}$$ From (\[eq:f11g11\]), (\[eq:f31g31\]), and (\[eq:deltamu\]), we can obtain $$\begin{aligned} l_{11} &= \frac{\rho_\infty[\delta,p]_{\alpha\beta}+p[N,p]_{\alpha\beta}}{2\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}, \\ m_{11} &= \frac{\rho_\infty[\delta,\rho_\infty]_{\alpha\beta}+p[N,\rho_\infty]_{\alpha\beta}}{2\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}, \\ l_{10} &= \frac{p[N,\delta]_{\alpha\beta}}{4\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}, \\ m_{10} &= \frac{\rho_\infty[\delta,N]_{\alpha\beta}}{4\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}, \\ l_{31} &= \frac{1}{2}-\frac{N[\mu,p]_{\alpha\beta}}{4\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}-pl_{11}, \\ m_{31} &= -\frac{N[\mu,\rho_\infty]_{\alpha\beta}}{4\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}-pm_{11}, \\ l_{30} &= -\frac{N[\mu,\delta]_{\alpha\beta}}{8\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}-pl_{10}, \\ m_{30} &= -\frac{N[\mu,N]_{\alpha\beta}}{8\kappa^2[\delta,\rho_\infty]_{\alpha\beta}}-pm_{10}. \end{aligned}$$ Further, using them and (\[eq:xj\]), (\[eq:xmu\]), and (\[eq:xjxmu\]), $$\begin{aligned} 2p l_{11}+F'(\rho_\infty)m_{11} &= 1, \label{eq:l11m11} \\ \rho_\infty l_{11}+pm_{11}&=-\frac{1}{2}\frac{[N,\rho_\infty]_{\alpha\beta}}{[\delta,\rho_\infty]_{\alpha\beta}}, \\ 2p l_{10}+F'(\rho_\infty)m_{10} &= -\frac{1}{2}\frac{[N,\delta]_{\alpha\beta}}{[\delta,\rho_\infty]_{\alpha\beta}}, \\ \rho_\infty l_{10}+pm_{10}&=0, \label{eq:l10m10} \\ 2p l_{31}+F'(\rho_\infty)m_{31} &= 0, \label{eq:l31m31}\\ \rho_\infty l_{31}+pm_{31}&=-\frac{1}{2}\frac{[P,\rho_\infty]_{\alpha\beta}}{[\delta,\rho_\infty]_{\alpha\beta}}, \\ 2p l_{30}+F'(\rho_\infty)m_{30} &=\frac{1}{2}\frac{[pN,\delta]_{\alpha\beta}}{[\delta,\rho_\infty]_{\alpha\beta}}, \\ \rho_\infty l_{30}+pm_{30}&=\frac{N}{4}. \label{eq:l30m30} \end{aligned}$$ We can eliminate $ m_{ij} $s by Eqs. (\[eq:l11m11\]), (\[eq:l10m10\]), (\[eq:l31m31\]), and (\[eq:l30m30\]), and the expressions (\[eq:C1\])–(\[eq:D2\]) are derived in that way.\ Now we consider the particular coordinate $ (\alpha,\beta)=(p,\rho_\infty) $. $ l_{ij} $s are then given by $$\begin{aligned} l_{11}&=-\frac{j\delta_p+p_{\text{L}}^2N_p}{2\rho_\infty\kappa^2\delta_p}, \label{eq:l11} \\ l_{10}&=\frac{p[N,\delta]_{p\rho_\infty}}{4\kappa^2\delta_p}, \\ l_{31}&=-\frac{p_{\text{L}}^2 P_p}{2\rho_\infty\kappa^2\delta_p}, \label{eq:l31}\\ l_{30}&=-\frac{N[\mu,\delta]_{p\rho_\infty}}{8\kappa^2\delta_p}-pl_{10}. \end{aligned}$$ From $ P = -pN-\rho_\infty\delta $ and (\[eq:idntydev\]), we obtain $$\begin{aligned} \delta_p &= -\frac{N+pN_p+P_p}{\rho_\infty}, \\ \delta_{\rho_\infty} &= -\frac{pN-\kappa^2N_p+pP_p+jN_{\rho_\infty}}{\rho_\infty^2}. \label{eq:deltarhoinfty} \end{aligned}$$ Derivatives of $ \delta $ are eliminated by using them. With the use of Eqs. (\[eq:l11\])–(\[eq:deltarhoinfty\]), we can obtain the final result (\[eq:final1\])–(\[eq:final7\]). Zero-energy solutions without using parameter derivatives {#app:zerosol} ========================================================= We can write down the general zero-energy solution without using parameter derivatives: $$\begin{aligned} \begin{pmatrix}u\\ v \end{pmatrix}&=\sum_{i=1,2,3,4} c_i \begin{pmatrix}u_i^{} \\ -u_i^* \end{pmatrix}, \\ u_1 &= \mathrm{i}\phi, \\ u_2 &= \phi_x, \\ u_3 &= \mathrm{i}\phi\int\!\frac{\mathrm{d}x}{\rho}+4\mathrm{i}j^2\phi\int\!\frac{\mathrm{d}x}{\rho\rho_x^2}-4j\phi_x\int\!\frac{\mathrm{d}x}{\rho_x^2}, \\ u_4 &= \mathrm{i}j\phi\int\!\frac{\mathrm{d}x}{\rho_x^2}-\phi_x\int\!\frac{\rho\mathrm{d}x}{\rho_x^2}. \end{aligned}$$ These expressions, however, have a fatal flaw; Namely, they have artificial singularities at the origin (or more precisely, the points where $ \rho'(x)=0 $.) It means that these expressions do not give global solutions which continuously connect the solutions from $ x=-\infty $ to $ x=+\infty $, instead, only give local solutions which satisfy the differential equation at each point. For this reason they are not so useful in scattering problems, which are equivalent to the determination of global behavior of given solutions. On the other hand, the parameter-derivative solutions given in Subsec. \[subsec:zerosol\] have no singularities if the density $ \rho(x) $ does not cross the value $ \rho_\infty $. Stationary solutions of CQNLS equation {#app:cqnls} ====================================== In the CQNLS system, the equation for momentum conservation (\[eq:genjm\]) becomes $$\begin{aligned} \frac{(\rho_x)^2}{4}=-j^2+j_m\rho-\mu\rho^2+a_1\rho^3+a_2\rho^4. \label{eq:cqdiff} \end{aligned}$$ Here we want general stationary solutions, so we do not concentrate on the solution with the asymptotic form (\[eq:nlsdsasym\]), and therefore $ \mu, j, $ and $ j_m $ need not be given by (\[eq:cp\]), (\[eq:dsj\]), and (\[eq:dsjm\]), respectively. If the right hand side of (\[eq:cqdiff\]) is factored as $ a_2\prod_{i=1}^4(\rho-\rho_i) $, the solution of this differential equation is given by the following cross-ratio form: $$\begin{aligned} \operatorname{sn}^2\left( \kappa x \| m \right)&=\frac{(\rho(x)-\rho_1)(\rho_3-\rho_2)}{(\rho(x)-\rho_2)(\rho_3-\rho_1)} \\ \intertext{with} \kappa&=\sqrt{a_2(\rho_4-\rho_2)(\rho_3-\rho_1)}, \label{eq:kappaapp} \\ m&=\frac{(\rho_4-\rho_1)(\rho_3-\rho_2)}{(\rho_4-\rho_2)(\rho_3-\rho_1)}, \end{aligned}$$ or equivalently, $$\begin{aligned} \rho(x)=\frac{\rho_2(\rho_3-\rho_1)-\rho_1(\rho_3-\rho_2)\operatorname{sn}^2\left( \kappa x \| m \right)}{(\rho_3-\rho_1)-(\rho_3-\rho_2)\operatorname{sn}^2\left( \kappa x \| m \right)}. \label{eq:rhoapp} \end{aligned}$$ Here we use Mathematica’s definition for Jacobi elliptic functions.[^2]\ For the case of the dark soliton solution, $ \rho_3=\rho_4=\rho_\infty $ holds, so $ m=1 $ follows and the sn function reduces to the tanh function. $ \rho_1 $ and $ \rho_0:=\rho_2 $ are given by the roots of $$\begin{aligned} a_2\rho^2+(2a_2\rho_\infty+a_1)\rho-p^2=0. \label{eq:quadracqnls} \end{aligned}$$ $ \rho_0 $ (\[eq:rho0cqnls\]) is the root of Eq. (\[eq:quadracqnls\]) with a plus sign. $ \rho_1 $ can be eliminated in two ways; From Eq. (\[eq:kappaapp\]) or from the fact that $ \rho_0 $ and $ \rho_1 $ solve Eq. (\[eq:quadracqnls\]), $$\begin{aligned} \rho_\infty-\rho_1=\frac{\kappa^2}{a_2(\rho_\infty-\rho_0)}, \qquad \rho_0\rho_1=-\frac{p^2}{a_2}. \label{eq:eliminaterho1} \end{aligned}$$ The expression (\[eq:rhoapp\]) is then rewritten as $$\begin{aligned} \rho(x)=\rho_\infty-\frac{\kappa^2(\rho_\infty-\rho_0)\operatorname{sech}^2\kappa x}{\kappa^2-a_2(\rho_\infty-\rho_0)^2\tanh^2\kappa x}. \end{aligned}$$ The following expressions are also useful for calculation of the phase shift $ \delta $ and the particle number of soliton $ N $: $$\begin{aligned} S_x=\frac{j}{\rho}&= p+\frac{(\rho_0\kappa)[ p(\rho_\infty-\rho_0)\tanh\kappa x]'}{(\rho_0\kappa)^2+[ p(\rho_\infty-\rho_0)\tanh\kappa x]^2}, \label{eq:Sxinapp} \\ \rho-\rho_\infty &= -\frac{1}{\sqrt{a_2}}\frac{\kappa [\!\sqrt{a_2}(\rho_\infty-\rho_0)\tanh\kappa x]'}{\kappa^2-[\!\sqrt{a_2}(\rho_\infty-\rho_0)\tanh\kappa x ]^2}, \end{aligned}$$ where Eq. (\[eq:Sxinapp\]) is obtained by eliminating both $ a_2 $ and $ \rho_1 $ with the use of (\[eq:eliminaterho1\]). Here we can use the formulae $$\begin{aligned} \frac{ay'}{a^2+y^2} &= \Bigl(\tan^{-1}\frac{y}{a}\Bigr)', \\ \frac{ay'}{a^2-y^2} &= \Bigl(\tanh^{-1}\frac{y}{a}\Bigr)'. \end{aligned}$$ Thus the content in Subsec. \[subsec:cqnls\] is reproduced.\ [^1]: Here we do not consider the derivative with respect to the parameters included in the definition of the nonlinear term, e.g., $ a_1 $ and $ a_2 $ in the CQNLS equation, because these parameter-derivatives do not yield a solution of the linearized equation. [^2]: The Wolfram Functions Site, http://functions.wolfram.com/	{ "pile_set_name": "ArXiv" }
--- abstract: 'Shashlyk electromagnetic calorimeter modules with an energy resolution of about $3\%/\sqrt{E\ \mathrm{(GeV)}}$ for $50\mathrm{-}1000\ \mathrm{MeV}$ photons has been developed, and a prototype tested. Details of these improved modules, including mechanical construction, selection of wave shifting fibers and photo-detectors, and development of a new scintillator with improved optical and mechanical properties are described. How the modules will perform in a large calorimeter was determined from prototype measurements. The experimentally determined characteristics of the calorimeter prototype show energy resolution of $\sigma_E/E=(1.96\pm0.1)\%\oplus(2.74\pm0.05)\%/\sqrt{E},$ time resolution of $\sigma_T = (72\pm4)/\sqrt{E}\oplus(14\pm2)/E\ \mathrm{(ps)}$, where photon energy $E$ is given in $\mathrm{GeV}$ units and $\oplus$ means a quadratic summation. A punch-through inefficiency of photon detection was measured to be $\epsilon \approx 5\times 10^{-5}\;\;(\Theta _{\mathrm{beam}} >5\;\mathrm{mrad})$.' address: - 'Institute for Nuclear Research of Russian Academy of Sciences, Moscow 117312, Russia' - 'Institute for High Energy Physics, Protvino 142284, Russia' - 'Physics Department, Yale University, New Haven, CT 06511, USA' author: - 'G.S. Atoian' - 'G.I. Britvich' - 'S.K. Chernichenko' - 'S. Dhawan' - 'V.V. Issakov' - 'O.V. Karavichev' - 'T.L. Karavicheva' - 'V.N. Marin' - 'A.A. Poblaguev' - 'I.V. Shein' - 'A.P. Soldatov' - 'M.E. Zeller' date: 27 September 2007 title: An Improved Shashlyk Calorimeter --- , , , , , , , , , , , and Shashlyk calorimeter ,scintillator ,WLS fiber ,APD ,WFD ,Monte-Carlo simulation 29.40.Vj ,07.05.Tp Introduction. ============= The Shashlyk technique for electromagnetic colorimetry [@NIM_A320_144] has been in use for several years. In designing a large calorimeter for the KOPIO experiment [@KOPIO] we have developed an improved Shashlyk module for such a device. This paper describes the design and construction of the module, as well as the unit’s performance in prototype tests. It is further development of the work described in Ref [@NIM_A531_467]. The requirements of the KOPIO experiment led to the following specifications: - Energy range: $50\mathrm{-}1000$ $\mathrm{MeV}$. - Energy resolution: $\approx 3.0\%/\sqrt{E(\mathrm{GeV})}$. - Time resolution: $\approx 100\ \mathrm{ps}\,/\sqrt{E\ \mathrm{(GeV)}}$. - Effective noise per channel: $\leq 0.5\ \mathrm{MeV}$ - Photon detection inefficiency: $\le10^{-4}$. - Granularity: $\sim10\ \mathrm{cm}$. - Size of the Calorimeter: $4.4\times4.4$ m$^2$. - Operation in a magnetic field of up to 500 Gauss. The Shashlyk approach, in which modules are constructed of lead-scintillator sandwiches read out by means of Wavelength-Shifting (WLS) fibers passing through holes in the scintillator and lead is appropriate for such specifications. The performance level can be achieved economically with well understood and reliable techniques. We describe a module with significantly improved performance over previous manifestations for which the technique is well proved by us, [e.g.]{} experiment E865 at Brookhaven [@NIM_A320_144; @NIM_A479_349], and has been adopted by others, [e.g.]{} the PHENIX RHIC detector [@NIM_A499_521], the HERA-B detector at DESY [@NIM_A461_332], and the LHCb detector at CERN [@NIM_A462_233]. Approach to improving the Shashlyk module ========================================= At the outset of improving the module a detailed simulation model was developed [@NIM_A531_467]. This model is based on a GEANT3 [@GEANT3] description of electromagnetic shower and a special optical simulation of the light collection in the scintillator plates. Input for the model was experimental and test beam data. The model output correlated well with data. Analyzing the performance of the prototype Shashlyk module with energy resolution of about $4\%/\sqrt{E(\mathrm{GeV})}$ [^1] we found, that energy resolution of about $3.0\%/\sqrt{E(\mathrm{GeV})}$ could be reached after some optimization of the module’s mechanical and optical properties (see details in Ref. [@NIM_A531_467]). We revisited the mechanical and optical construction as described below, and optimized the selection of WLS fibers and the photo-detector. A new scintillator with improved optical and mechanical properties was specially developed at the IHEP scintillator facility (Protvino, Russia) [@IHEP]. The corresponding improvements to the module design were implemented in the new KOPIO Calorimeter prototype modules, which were equipped with an Avalanche Photo Diode (APD) and Wave-Form Digitizer (WFD) readout. Mechanical construction of the improved Shashlyk module. -------------------------------------------------------- ![The Shashlyk module design.[]{data-label="module_design"}](fig01){width="75.00000%"} ![The Shashlyk modules at different stages of assembly.[]{data-label="module_assembly"}](fig02){width="65.00000%"} The design of the new prototype module is shown in Figures \[module\_design\], \[module\_assembly\]. The module is a sandwich of alternating perforated stamped lead and injection-molded polystyrene-based scintillator plates. The cross sectional size of a module is $110\times110$ mm$^2$. There are 300 layers, each layer consisting of a 0.275-mm lead plate and 1.5-mm scintillator plate. The lateral size of a lead or scintillator plate is $109.7\times109.7$ mm$^2$. Each plate has 144 holes equidistantly arranged in a $12\times 12$ matrix, in which the spacing between the holes is 9.3 mm. The diameter of the holes is 1.3 mm, both in the lead and the scintillator plates. Inserted into the holes are 72 WLS fibers, with each fiber looped at the front of the module so that both ends of a fiber are viewed by a photo-detector. Such a loop (radius $\sim$2.5 cm) may be approximated by a mirror with a reflection coefficient of about 95% [@KURARAY]. The fiber ends are collected in one bunch, squeezed, cut and polished, and connected to a photo-detector though a small air gap. The complete stack of all plates is held in compression by four 1-mm stainless steel wires. The module is wrapped with 150-$\mu$m TYVEK paper which has light reflection efficiency of about 80%. The mechanical parameters of the module are summarized in Table \[module\]. -------------------------------------- -------------------------------------------------------------------- Transverse size $110\times110~\mathrm{mm}^2$ Scintillator thickness $1.5~\mathrm{mm}$ Spacing between scintillator tiles $0.350~\mathrm{mm}$ Lead absorber thickness $0.275~\mathrm{mm}$ Number of the layers 300 WLS fibers per module $72\times1.5~\mathrm{m} = 108~\mathrm{m}$ Fiber spacing $9.3\ \mathrm{mm}$ Hole diameter (lead/scint.) 1.3 mm Diameter of WLS fiber (Y11-200MS) 1.0 mm Fiber bunch diameter 14.0 mm External wrapping (TYVEK paper) 150 $\mu$m Effective radiation length, $X_0$ 34.9 mm Effective Moliere radius, $R_M$ 59.8 mm Effective density 2.75 g/cm$^3$ Active depth 555 mm (15.9 $X_0$) Total depth (without photo-detector) 650 mm Total weight 21 kg -------------------------------------- -------------------------------------------------------------------- : Parameters of the improved Shashlyk module.[]{data-label="module"} The module is assembled in a special assembling berth that allows one to subject the assembled modules to cyclic longitudinal loadings up to 800 kg. This procedure prevents subsequent longitudinal shrinkage of the assembled modules and provides long-term stability for the length of the module to an accuracy of $\simeq 1\ \mathrm{mm}$. Improvements of the Module Geometry ----------------------------------- The mechanical construction of the module was revisited to minimize the insensitive area, to increase the effective radiation density, and to improve the sampling ratio and transverse light collection uniformity. An important innovation in the mechanical design of the module is the “LEGO”-type locks for the scintillator tiles shown in Fig. \[module\_design\]. These locks, four per tile, maintain the position of the scintillators and the 350-$\mu$m gaps, providing sufficient room for the 275-$\mu$m lead tiles without optical contacts between lead and scintillator. This mechanical structure enables removal of the 600 paper tiles that were in earlier modules, and allows reduction of the diameter of the fiber holes to 1.3 mm and removal of the compressing steel tapes at the sides of the module. Compared to the earlier Shashlyk module, the holes/cracks and other insensitive areas were reduced from 2.5% to 1.6%, and the module’s mechanical properties such as dimensional tolerances and stiffness were significantly improved. By removing the paper tiles, the effective radiation length could be reduced from 4.0 cm to 3.5 cm. ![The dependence of the light collection efficiency in the scintillator tile on the $x$-coordinate of the point-like light source. Solid lines are the simulations for the specified reflection efficiencies $\varepsilon_R$ of the wrapping material.[]{data-label="tile_uniform"}](fig03){width="65.00000%"} The sampling, i.e., the relation between thicknesses of lead and scintillator tiles, dominates in the energy resolution of the Shashlyk module. However, one has only limited ability to improve the “pure" sampling contribution to the energy resolution of the Shashlyk module. Decreasing the thickness of the lead will increase the length of the module, while the proportional decreasing both the lead and scintillator thicknesses will reduce the light collection efficiency. Nevertheless, by removing the paper between the lead and scintillator tiles, both the sampling could be improved and the length of the module could be shortened. Compared to the design of the previous module [@NIM_A531_467], the sampling ratio was improved by a factor of 1.25. ![ The effective attenuation of the light in the fibers of Shashlyk module. Experimental data (marks) are fit by the exponential dependence $\exp{(-x/\lambda )}$ (solid lines), where $x$ is the distance to the photo-detector and $\lambda $ is the effective attenuation length. []{data-label="wls_attenuation"}](fig04){width="65.00000%"} The dominant source of non-uniformity of the light collection is the absorption of light at the edges of a scintillator tile. The reflection efficiency on the edges of the scintillator tile is crucial. In the new module, the WLS fiber density was effectively increased by reducing the size of the tiles to 10.97 cm. This allows the outer fibers to be closer to the edge of the scintillator tile than than they were with the the “uniform” size of 11.16 cm for 12 fibers with 0.93-cm spacing. In addition, the module was wrapped with TYVEK paper (reflection efficiency about 80%). As a result, the light collection efficiency at the edges of the scintillator tile is only 2.3% smaller than in the center of the tile for the point-like light source. In the case of a 250-MeV photon shower, the difference is only 0.5%, which is negligible compared to the energy resolution of about 6% for such photons. The experimental results for the light collection uniformity for TYVEK and Xerox copier papers are presented in Fig. \[tile\_uniform\]. The measurements were made with a scintillator tile exposed to collimated $^{90}$Sr electrons. For comparison, the simulated light collection efficiencies are shown. One can see that there is consistency between the optical simulations and measurements. Further improvement could be achieved if Millipore paper with reflection efficiency of 90% were used [@refl_eff]. WLS Fibers ---------- A main concern about the WLS fibers for the Shashlyk readout is the light attenuation length in a fiber. Longitudinal fluctuations of electromagnetic showers are about 3-4 cm (one radiation length). The effective attenuation length in fibers, including the effect of the fiber loop and the contribution of the short-distance component of light, must be greater than 3–4 m to have this contribution to the energy resolution be much smaller than the sampling contribution. We have measured (see Fig. \[wls\_attenuation\]) the light attenuation in few multi-clad WLS fibers using a $1\times1\ \mathrm{cm}^2$ wide muon beam penetrating the module transversely. The effective attenuation length of 3.6 m in KURARAY Y11(200)MS fiber satisfies our requirements. In comparison with other fibers, this commercial fiber also provides the best reemission efficiency of blue scintillation light, and has excellent mechanical properties, high tensile and bending strength, and high uniformity in cross-sectional dimensions. For example, its light reemission efficiency is a factor of 1.5 larger than that for any Bicron WLS fiber, and the diameter for any round fiber is more uniform than others in that it varies by no more than $2.0\%$. Scintillator ------------ An important contribution for the improvement of the photo-statistics over earlier designs of Shashlyk modules is the use of new scintillator tiles with an increased light collection efficiency. An optimization of the light yield of the scintillator tiles for the KOPIO Shashlyk modules has been developed and carried out at the IHEP scintillator facility (Protvino, Russia) [@IHEP]. In the previous Shashlyk calorimeters, scintillator based on PSM115 polystyrene was used. The new modules employ BASF143E-based scintillator. ![The parametric dependence of the light collection efficiency on the effective attenuation length due to the reflection efficiency. The solid line is the result of a calculation based on the model of Ref. [@NIM_A531_467]. For each scintillator, the calculated reflection efficiency is displayed.[]{data-label="tile_reflection"}](fig05){width="65.00000%"} Though there is no actual increase in the amount of light produced by a charge particle, the light collection efficiency in the new scintillator tile has a gain by a factor of 1.6. Because the index of refraction for the polystyrene-based scintillator is 1.59, only light from total internal reflection on the large side of the scintillator tile can be captured by a WLS fiber. The total internal reflection efficiency can potentially be as large as 100%. However, this value is not reachable for realistic surfaces. Many reflections usually occur before light is captured and re-emitted by a WLS fiber. Both the effective attenuation length and the light collection efficiency in a scintillator tile depends on the light reflection efficiency. As shown in Fig. \[tile\_reflection\], the parametric dependence of the light collection efficiency on the attenuation length, calculated from our model of module optics [@NIM_A531_467], is in a good agreement with the experimental results. ![Comparison of the light-yield spread for scintillator tiles used in earlier (PSM115) and new (BASF143E) Shashlyk modules.[]{data-label="tile_light_spread"}](fig06){width="65.00000%"} The new plastic base of the scintillator material and the new production technology of tiles yields a 97% reflection efficiency from the scintillator surface, and strongly improves the reproduction quality (see Fig. \[tile\_light\_spread\]) of the tiles. The latter is crucial for good longitudinal light collection uniformity. Due to the fluctuation of the depth of the electromagnetic shower in the calorimeter module, good reproduction quality of the tiles is a mandatory condition for an appropriate performance of the $3\%/\sqrt{E\ \mathrm{(GeV)}}$ module. ![The closed circles and solid line are the measured emission spectra of the BASF143E-based scintillator. The dashed line is the absorption spectrum of KURARAY Y11(200)MS WLS fibers.[]{data-label="abs_emit_2"}](fig07){width="65.00000%"} Fluors for the new scintillator composition, BASF143E polystyrene + 1.5% PTP + 0.04% POPOP were selected [@scint1; @scint2] to match well with the absorption spectrum of the Kuraray Y11(200)MS WLS fiber (Fig. \[abs\_emit\_2\]). With a BASF143E-based scintillator and KURARAY fibers, the effective light yield in the KOPIO Shashlyk module (at the entrance to the photo-detector) becomes $N_{\gamma }\approx\ 60$ photons per 1 MeV of the incident photon energy. The radiation hardness of the scintillator may be a limiting factor for using the Shashlyk calorimeter in the modern high statistics experiment. It was reported [@scint1] that the radiation stability of BASF143E-based scintillator is a dose level 120 krad, and the recovery time is about 80 days. This limitation would not be a problem for the KOPIO Calorimeter for which this Shashlyk module was designed. Photo-detector -------------- The Avalanche Photo Diode, a detector with high quantum efficiency, provides another important improvement of the photo-statistics contribution to the energy resolution of the Shashlyk module. Our previous experimental studies of fine-sampling Shashlyk modules [@NIM_A531_467] have shown that the performance of Shashlyk calorimeter modules with photomultiplier (PMT) readouts agrees with the simulations and nearly satisfies the requirements formulated in the introduction. However, these test measurements were made in an optimal environment for PMT readout: no magnetic fields, short measuring runs, and continuous and stable electron beams. These conditions do not properly represent those that would be encountered in the KOPIO experimental environment. Our main concern is related to the leakage of the magnetic field from the downstream magnet of up to 500 Gauss. Another serious problem is the cycling of real beam every several seconds, with possible significant variations in intensity. This will produce short- and long-term variations of the PMT gain, which will destroy the response stability of the detector performance. In addition, the quantum efficiency of PMTs, even for the best “standard” green-sensitive tubes, e.g. PMT 9903B of Electron Tube Inc., is relatively low, about 20% at 500 nm (the region of WLS-fiber response). ----------------------------------------------- --------------- -------------------- -------------- -------------- Manufacturer API RMD EG&G Hamamatsu Model number 630-70-74-510 S1315 C30703E S8664-1010N Active area, $\mathrm{mm}^2$ 200 169 100 100 $\oslash16$ $13\times13$ $10\times10$ $10\times10$ Quantum Eff., % $ 94 $ $ 65 $ $ 75 $ $ 80 $ Capacitance, pF $ 130 $ $ 110 $ $ 80 $ $ 270 $ Series resistance, $\Omega$ $ 15 $ $ 50 ( 280)$ $ 10 $ $< 5 $ Signal rise time, ns $ 6 $ $ 9 $ $ 5 $ $ 4.4$ Signal fall time, ns $ 16 $ $ 24 ( 80)$ $ 15 $ $ 25$ Gain $\leq 600 $ $\leq 1000$ $\leq 200 $ $\leq 300 $ Excess noise factor $ 2.2 $ $ 2.05$ $ 3.2 $ $ 2.5 $ $I_S$, nA $ 50 $ $ 310$ $ 90 $ $ 3 $ $I_B$, nA $ 0.6 $ $ 3.8$ $ 0.4 $ $ 0.3 $ $\frac{dM}{M\,dT}$, ${}^\circ\mathrm{C}^{-1}$ $ -0.03$ $ -0.03$ $ -0.04$ $ -0.045$ $\frac{dM}{M\,dV}$, $\mathrm{V}^{-1}$ $ 0.014$ $ 0.016$ $ 0.03$ $ 0.03$ ----------------------------------------------- --------------- -------------------- -------------- -------------- : Some experimental and catalog parameters of large-area APDs. Quantum efficiency is given for green light, $\lambda=500\ \mathrm{nm}$. Signal rise and fall time are measured without amplifier. Excess noise factor, which approximately depends on the gain $M$ as $2+kM$ is given for $M=100$. $I_S$ and $I_B$ are components of the dark current $I_d=I_S+I_BM$. The gain dependences on temperature, $T$, and high voltage, $V$, are given for $M\sim100$. For RMD APD, we have observed a significantly different results for values of series resistance (and consequently for signal fall time) which is indicated in the Table. []{data-label="Tab1"} Consideration was thus given to alternative photo-detectors. Recent progress in the development of new APDs with large active areas (up to 200 mm$^2$), low capacitance, low dark current, high gain (up to 1,000), and high quantum efficiency (up to 90%) allows us to consider these photo-detectors as primary candidates for the improved Shashlyk Calorimeter. We have studied the large area ($\geq100\ \mathrm{mm}^2$) APDs of [(i) Advanced Photonix Inc.]{} (API), [(ii) RMD Instruments Inc.]{} (RMD), [(iii) Perkin-Elmer Inc.]{} (EG&G), and [(iv) Hamamatsu Photonix K.K]{}. Some characteristics of these APDs are summarized in Table \[Tab1\]. Here, we briefly report the results of our study. - Size of the active area.\ The fibers from the Shashlyk module, described in this paper, are collected in a 14 mm diameter bunch. This size is well matched to the API APD, which has a diameter of 16 mm. Optical light guides must be be used with other APDs that were considered, to match the size of the active area. For $10\times10\ \mathrm{mm}^2$ active area, this will result in a light loss of at least 10-15%. - Response uniformity of the active area.\ This feature is important for the Shashlyk module because each fiber delivers light to only a small part of the total sensitive area. The response of the selected APD varies by less than 3% over the active area. Usually, PMTs exhibit poor spatial response uniformity ($\ge20\%$). ![ The quantum efficiency of considered APDs: $1-$ API ($\mathrm{SiO}_2$ window), $1a-$ API (epoxy window), $2-$ RMD (epoxy window), $3-$ EG&G ($\mathrm{SIN}_x$ window), $4-$ Hamamatsu (epoxy window). For comparison, quantum efficiency of the 9903B PMT is also shown. The histogram is an emission spectrum of the Kuraray Y11200MS WLS fiber. []{data-label="apd_qe"}](fig08){width="65.00000%"} ![The APD excess noise factor [vs.]{} the APD gain.[]{data-label="apd_f-factor"}](fig09){width="65.00000%"} - Quantum efficiency of the sensitive area.\ Photo-statistics, directly dependent on the quantum efficiency ($Q$) of the photo-detector, is an important contribution to the energy resolution of a calorimeter. For green light emitted by Kuraray fibers, the APDs have much higher quantum efficiencies than PMTs (see Fig. \[apd\_qe\]). For example, the quantum efficiency of the API APD is 94%, a factor 5 higher than quantum efficiency of the 9903B PMT. It should be noted that that the photo-statistics contribution depends also on the fluctuations of the photo-detector gain, $(\sigma_E/E)^2_\mathrm{ph.stat.}= F/QN_\gamma$, where F is the so-called excess noise factor and $N_\gamma$ is the number of primary photons at the entrance of the photo-detector. For an ideal photo-detector, $F=1$, for a high-linearity PMT it is usually between 1.3 and 1.6. For an APD, excess noise factor dependences on gain $M$ may be well approximated as $F \simeq 2 + kM$ if $M\geq20$ [@APD-th]. The experimentally measured behavior of $F$ versus $M$ for the APDs under consideration is shown in Fig. \[apd\_f-factor\]. Taking into account both quantum efficiency $Q$ and excess noise factor $F$, we conclude that API APD has best “photo-statistics quality” ( $Q/F\simeq43\%\ (M_\mathrm{APD}\simeq100)$) of about a factor 3.5 better than 9903B PMT. - Electronic noise of the photo-detector/amplifier chain.\ In the case of a Charge Sensitive Amplifier (CSA) with a simple $CR$-$RC$ shaper, the contribution of the electronic noise from the photo-detector/amplifier chain to the energy resolution can be estimated as: $$\sigma_\mathrm{noise} = \frac{1}{n_\gamma Q} \sqrt{\frac{I_S/M^2 + I_BF}{q}\Delta t + \frac{\sigma^2_{\mathrm{amp}}(C,\Delta t)}{M^2}} \label{eq:2}$$ where $n_\gamma$ is number of photons at the entrance of photo-detector per unit of the deposited energy ($n_\gamma\approx60\ \mathrm{MeV}^{-1}$ for the improved Shashlyk module), $q=1.6\cdot10^{-19}\ \mathrm{C}$ is electron charge, and $\Delta t$ is the measurement gate width. The photo-detector noise can be related to the $I_S$ (surface leakage current) and $I_B$ (bulk current), the components of the dark current $I_d=I_S+I_BM$. The amplifier noise, $\sigma_\mathrm{amp}$, defined here as Equivalent Noise Charge (ENC) depends on the gate width $\Delta t$ and the photo-detector capacitance $C$. ![The APD dark current.[]{data-label="apd_dc"}](fig10){width="65.00000%"} For a PMT readout, the photo-detector noise is negligible due to a small ($\sim0.1\ \mathrm{nA}$) PMT dark current and big ($\geq10^5$) gain. The experimentally measured behavior of the APD dark current versus gain $M$ for our APDs is shown in Fig. \[apd\_dc\]. The contribution of $I_S$ to the noise becomes negligible if $M\geq100$. For the API APD, the photo-detector contribution to the noise can be estimated as $0.5\ \mathrm{MeV}$ for $\Delta t\approx 100\ \mathrm{ns}$, $M \geq 100$, and environmental temperature $T=25\ {}^\circ\mathrm{C}$. The value of the amplifier noise $\sigma_{\mathrm{amp}}$ is usually measured experimentally. We have used a fast low noise CSA, that has been designed for KOPIO Calorimeter to optimize the signal-to-noise ratio and double-pulse resolution. Its charge-sensitive part was designed as a cascade amplifier with two parallel-connected low capacitance JFET transistors. We considered the Russian KP341A and Japanese SK2394/YJ5 transistors for this amplifier. Both transistors are characterized by a high transconductance ($g_m>20\ \mathrm{mA/V}$) for moderate noise current and low input capacitance ($C_\mathrm{JFET}<10\ \mathrm{pF}$). For a $100\ \mathrm{ns}$ gate, the ENC dependence on the photo-detector capacity $C\ \mathrm{(pF)}$ was measured to be $270+18.8C$ for the KP341A and $500+10.9C$ for the SK2394/YJ5. We used an amplifier with two SK2394/YJ5 transistors in the test beam measurements. For the API APD ($C=130\ \mathrm{pF}$) the amplifier noise may be estimated as $0.3\ \mathrm{MeV}$ resulting in a total photo-detector/amplifier chain noise of $0.6\ \mathrm{MeV}$ for $\Delta t=100\ \mathrm{ns}$, $M=100$, and $T=25\ {}^{\circ}C$. It should be underlined, that large photo-statistics, $n_\gamma$, is a crucial factor in achieving required effective noise suppression. ![ Effective noise of the API APD/amplifier chain [vs.]{} the APD gain for various environmental temperatures. []{data-label="apd_noise"}](fig11){width="65.00000%"} - Gain and noise dependence on the temperature.\ The APD performance is very sensitive to the environmental temperature. Because $\frac{1}{M}\frac{dM}{dT}\simeq -4\%$, temperature and gain monitoring is important in using APDs in the experiment. For the API APD, the noise dependence on temperature is shown in Fig. \[apd\_noise\]. A thermostable cooling system and appropriate increase of the APD gain may allow one to lower the effective noise down to the level of 0.2-0.3 MeV. - Gain dependence on High Voltage.\ According to our experimental study, $\frac{1}{M}\frac{dM}{dV}\approx (2-3)\%$ depending on the environment temperature and APD gain. The dependence is similar to those of PMTs. - Gain dependence on the rate.\ Contrary to PMTs, there is almost no dependence of APD gain on rate. In our test measurements with API APDs, we did not find any gain variations for the rates up to few $\mathrm{MHz}$. However, APDs are used with preamplifiers. For slow preamplifiers, there may be a rate dependence due to the pile-up. ![ PMT-9903B (upper) and API APD (lower) responses to short, 3 ns, light pulses. The intensity of the light pulse corresponds to the 700 MeV photon signal.[]{data-label="pmt-apd_response"}](fig12a "fig:"){width="50.00000%"} ![ PMT-9903B (upper) and API APD (lower) responses to short, 3 ns, light pulses. The intensity of the light pulse corresponds to the 700 MeV photon signal.[]{data-label="pmt-apd_response"}](fig12b "fig:"){width="50.00000%"} - Time response of the API APD.\ The time response of the API APD (without amplifier) to a short-duration (3 ns) light pulse with an intensity corresponding to a 700-MeV photon, is shown in Fig. \[pmt-apd\_response\]. The APD rise time of 5.9 ns is comparable with 3.7 ns for the 9903B PMT response to the same light pulse. Due to higher APD capacitance, the fall time of the APD pulse, 18.1 ns, is longer than that for the PMT response of 7.2 ns. The difference between a PMT and an APD time response becomes less important if we take into account the Shashlyk module signal (light pulse) duration. (see Fig. \[wfd\_signal\]). However, both the photo-detector capacitance $C$ and serial resistance $R_s$ are important parameters for the time response of the APD. - Operation in magnetic field.\ As opposed to PMTs, APDs can operate in a magnetic field of up to 80 kG [@APD_magnet]. From the above consideration of the available APDs, we conclude that the Advanced Photonix Inc. APD 630-70-74-510 is the optimal choice for our goals. Having the largest active area which matches well with the fiber bunch size, this commercial APD has best photo-statistics quality, $Q/F$, and sufficiently short time response when compared to other large area APDS . The relatively high gain and low dependence of excess noise factor on gain, allows us to optimize detector performance, e.g., to reduce the effective photo-detector/amplifier noise to the 0.2-0.3 MeV level for reasonable values of the environmental temperature. These characteristics of the API APD together with the improved light yield in the new Shashlyk module reduces the photo-statistics contribution to the energy resolution of the calorimeter to a negligible level of $0.7\%/\sqrt{E\ \mathrm{(GeV)}}$. The electronic noise contribution is also negligible. Experimental Study of the Calorimeter Prototypes ================================================ Test beam --------- Test measurements of the prototype of a Shashlyk calorimeter with energy resolution of $3\%/\sqrt{E\ \mathrm{(GeV)}}$ have been made with the photon beam from the Laser Electron Gamma Source (LEGS) facility [@LEGS]. LEGS is located at the National Synchrotron Light Source (NSLS) at Brookhaven National Laboratory. A continuous photon beam with an energy range of 150 to 400 MeV is produced by laser photons Compton backscattered by 2.58-GeV storage ring electrons. The energy of each backscattered photon was known by detecting the Compton scattered electron. The tagged photon beam had an average intensity of $\sim3\times10^5$ Hz, horizontal and vertical sizes of $\sim$1.5 cm at the detector position, and an angular spread of $\sim$2 mrad. The photon energy was tagged with an accuracy of $\delta E_\gamma/E_\gamma\approx1.5\%$. The timing accuracy of the tagging spectrometer was $\simeq$1 ns. Several monochromatic photon energy lines in the range of 220 MeV to 370 MeV were triggered in our measurements. Calorimeter Prototypes ---------------------- Two arrays of $3\times 3$ prototype modules with the same sampling of 275 $\mu\mathrm{m}$ of lead and 1.5 mm of scintillator were tested as shown in Fig. \[Beam-setup\]. The first array (prototype 1) contained 9 module with paper between lead and scintillator (earlier design) with the 30-mm-diameter, green-sensitive PMT-9903B of [Electron Tube Inc.]{} ![View of an APD unit.[]{data-label="apd_unit"}](fig14){width="65.00000%"} The second array (prototype 2) contained a nonet of new design modules equipped with the 16 mm diameter API APD. The APD detector housing, an instrumented unit including the APD itself, the APD amplifier, and the APD HV bias, is shown in Fig. \[apd\_unit\]. Because the new Calorimeter photo-detector (APD) draws a typical photo-current of less than 0.1 mA, an economical way to build the Calorimeter HV system is to develop an active, compact individual HV-unit mounted directly into each APD housing. This kind of HV system eliminates expensive and bulky HV cables and connectors, lowers the power consumption per individual power supply, and reduces the electrical HV hazard associated with traditional HV supplies. In the test measurements, two types of APD HV supplies were tested: traditional HV supply: a NIM standard [BERTAN]{} 377P, and a HV “built-in" unit (see Fig. \[apd\_unit\]) with a new commercial, compact, regulated and programmable LV-HV converter (C20), produced by EMCO. The stability of both bias systems was better than 0.1 V/hour, which provided an APD gain stability better than 0.3%. Cooling ------- The photo-detectors (APD) of the Prototype 2 were placed in a thermo-isolated cooling unit. During the test period of 20 hours, the APD temperature was kept at $18\ ^\circ\mathrm{C}$. Variation did not exceed $0.2\ ^\circ\mathrm{C}$. Such cooling system allowed us to neglect the temperature dependent effects in data analysis. Readout ------- One of the goal of the photon beam measurements was to test the readout prototype for the KOPIO experiment. The readout electronics had to be capable of measuring energies with a digitization uncertainty of 0.5 MeV in a dynamic range of 2–1000 MeV, and the time of arrival with respect to the beam micro-bunch clock with an uncertainty of less than a few hundred ps. Operation of this readout electronics was required to be completely pipelined with no dead time. It also had to be capable of forming and discriminating the total energy signals and the total number of hits in the Calorimeter. To meet these performance requirements for energy and timing resolution with PMTs or APDs, in an economical way, we have investigated using Wave Form Digitizers (WFDs). Our WFD was based on an8-bit 140-MHz WFD which has been developed by Yale University [@WFD]. Its measurements were compared with those of conventional 12-bit LeCroy QDCs. Calibration and Monitoring -------------------------- The energy resolution of $\sim$3.0%/$\sqrt{E}$ requires that a pre-calibration and monitoring system be developed. We have used vertical cosmic-ray muons for pre-calibration of the calorimeter prototypes, as we did in Experiment E865 [@NIM_A479_349]. The accuracy of this pre-calibration was estimated to be about 4%. To monitor prototype performance during our measurements and to correct the variation of the photo-detector gains with an accuracy of 0.4%, we have used a prototype of the KOPIO Calorimeter monitoring system. This system employs high-brightness Light Emitting Diodes (LEDs NSPB-310A) to inject blue light into the scintillator via a clear fiber (see Fig. \[module\_design\]). Stabilization of the LED light was obtained by means of an optical feedback provided by a PIN photo-diode as shown in References [@PIN1] and [@PIN2]. This system satisfied a number of our specific requirements: - short light pulse duration, less than 15 ns; - variable light pulse intensity ($5,000\mathrm{-}20,000$ photons/channel); - variable pulse repetition rate (up to $1MHz$); - high long-term and short-term temperature stability, better than 0.1%; - small variation in the flash amplitude, less than 0.2%. The monitoring system can also serve as a pre-calibration of the module readout chain at the 10% level. For final calibration of the prototypes we used the photon beam. Both calorimeter prototypes were mounted on platforms that could be moved horizontally and vertically with respect to the beam line, so that each prototype module could be calibrated (with an accuracy $\leq 1\%$) by using the 250-MeV photon beam that passed through the central region of the module at normal incidence. Experimental Results ==================== Calorimeter prototype response ------------------------------ The pedestal response, the sum of electronic noise and pile-up for individual modules, was measured during the test runs by using a special gate signal that was shifted from the photon timing by up to 1 $\mu$s. The contribution of this effective noise to the energy resolution was $(0.5\pm 0.1)$ MeV for the case of the “APD+QDC” readout. The total equivalent noise for the “APD+WFD” readout was $(1.0\pm 0.2)$ MeV. This latter value is twice as large as that for the first one due to a digitization uncertainty of the 8-bit WFD. The lowest equivalent noise, $(0.2\pm 0.1)$ MeV, was obtained with the PMT-9903B tubes. The typical response of the nonet of the “APD+WFD"-instrumented modules to a 340 MeV photon hitting the nonet at the center of the central module, is shown in Fig. \[wfd\_signal\]. The measured signal distribution in the WFD was fit by a function $Af(t-t_0)+P$, where $f(\tau)$ is an experimentally determined pulse shape function. The signal amplitude $A$, signal time $t_0$, and pedestal $P$ were free parameters in the fit. Energy Resolution ----------------- ![ Comparison of the experimental (marks) and simulated (solid lines) energy spectra in the Shashlyk prototype with APD+WFD readout. []{data-label="shashlyk_spectra"}](fig15){width="65.00000%"} The energy resolution was measured for both prototypes 1 and 2. The photon beam was directed to the center of the module nonets. Energy spectra of 250- and 370-MeV photons in the Shashlyk nonet with the APD+WFD readout are shown in Fig. \[shashlyk\_spectra\]. Note the good agreement between Monte-Carlo and experimental distributions, that gives us a confidence that our simulation model, which includes effects of the beam energy spread, energy loss upstream calorimeter, photon shower evolution, light collection in scintillator tiles and light transmission in WLS fibers, the response of the photo-detector, and noise of the entire electronic chain, properly reproduces the actual calorimeter response. ![Energy resolution of a prototype of the Shashlyk Calorimeter.[]{data-label="e_res"}](fig16){width="65.00000%"} The energy-resolution results for various readouts are displayed in Fig. \[e\_res\]. The best result was achieved for the “APD+WFD" readout. A fit to these experimental data gives $$\sigma_E/E=(1.96\pm0.1)\%\oplus(2.74\pm0.05)\%/\sqrt{E\ \mathrm{(GeV)}}\:,$$ where $\oplus$ means a quadratic summation. The relatively large constant term of 2% may be explained by the short, 15.9-X$_0$ radiation length of the module. The constant term may be improved by increasing length of the module. However, this term is not essential for the photon energy range of $50\mathrm{-}1000$ MeV. Time Resolution --------------- To estimate the time resolution, we measured the time difference for the signals produced by the same shower in two neighbor modules. By using this measurement technique, the 340 MeV photon beam was directed between central module (\#5) and left module (\#4), see Fig. \[Beam-setup\]. ![ Experimental evaluation of the time resolution of a Shashlyk module (see text for details). $E_4$ is energy measured in module \#4. Solid line is an expectation of the time difference R.M.S. $\sigma_{(T_4-T_5)}(E_4)=\sqrt{\sigma^2_T(E_4)+\sigma^2_{T}(0.33-E_4)}$, where $\sigma_T(E)=0.072/\sqrt{E}\oplus0.014/E$ (shown by dashed line) is a fit-estimated time resolution in each module, \#4 and \#5. []{data-label="t_res-2"}](fig17){width="65.00000%"} Only events with the full photon energy ($E_4 + E_5$ = $330\pm 20$ MeV) deposited in these two modules were selected for analysis. The dependence of the time difference variance on the energy deposited in module \#4 is shown in Fig. \[t\_res-2\]. The time difference resolution drops significantly if one of the two deposited energies is close to zero or, alternatively, if one of the energies is close to 330 MeV and the other is necessarily close to zero. Assuming that both modules have the same time resolution, we have obtained from Fig. \[t\_res-2\] $$\sigma_T = \frac{ (72\pm4)\ \mathrm{ps} }{ \sqrt{E\ \mathrm{(GeV)}}} \oplus \frac{ (14\pm2)\ \mathrm{ps} }{ E\ \mathrm{(GeV)}}$$ The signals we use to evaluate the time resolution are strongly correlated because they are produced by the same electromagnetic shower. This may result in a wrong value of the time resolution. For example, the contribution of the longitudinal fluctuation of the shower is suppressed in such measurements. To ensure that our test-beam result does not underestimate the actual time resolution, we carried out a Monte Carlo simulation of the time resolution. To simulate the time response and time resolution of a Shashlyk module, a Monte Carlo model was upgraded for the evolution of the light signal in a module. Ionization produced by a charged particle in a scintillator tile survives several transformations before the corresponding light signal appears at the entrance of a photo-detector. Ionization produced by an electromagnetic shower occur at different space points and different times, resulting in a time spread of the photo-electron emission in the photo-detector. In addition, the actual emission delays depend randomly on the decay times in the scintillator and the WLS fibers, light collection time in the scintillator fiber, and the propagation of light in the WLS fiber. The following effects have been taken into account in the model: - the space-time distribution of hits in the Shashlyk module scintillator; - the decay time in scintillator; - the light collection time in the scintillator tile; - the decay time in WLS fibers; - the effective velocity of light propagation in WLS fibers; - the light attenuation length in WLS fibers; - bending loss in the WLS fiber loop; and - the response function of the detector chain, including photo-detector (APD), preamplifier, cables and WFD. ![ A 340 MeV photon signal in WFD. Closed circles are experimental data. Solid line is the simulation, including readout chain response function. Dashed line is the light pulse shape at the entrance of the APD. []{data-label="wfd_signal"}](fig18){width="65.00000%"} The simulated response of the Shashlyk module for the 340 MeV photons is shown in Fig. \[wfd\_signal\]. The shape of the simulated signal in a Wave Form Digitizer matches well an experimentally measured signal. In the Monte-Carlo simulation of our time resolution measurements, we have obtained $\sigma_{T4-T5}= 280\ \mathrm{ps}$ if $E_4\approx E_5$ which is in a good agreement with the experimental value of 285 ps (see Fig. \[t\_res-2\]). ![ Results of a Monte Carlo simulations of the time resolution of the Shashlyk module for photons, electrons, and Minimum Ionizing Particles (MIP). The solid line corresponds to a $0.072/\sqrt{E}\oplus0.014/E\ \mathrm{(ns)}$, an approximation of the time resolution from our test-beam measurements. The simulated dependence of the mean measured time on type of particle and its energy is shown on the left. []{data-label="time_mc"}](fig19){width="80.00000%"} The Monte-Carlo simulations of the mean measured time and time resolution dependence on type of particle and its energy is shown in Fig. \[time\_mc\]. To understand the behavior of mean time values we should note that by fitting the signal shape in a WFD we actually measure the center of gravity of the signal time distribution. The main contributions to the relative signal delay are electron/photon tracking time before the hit and time of light propagation in the fiber. The average of time delays in both ends of the fiber does not depend on the actual position of the hit if we can neglect the light attenuation in the fiber. In such an approximation, the relative delay of the measured hit is solely defined by the time at which this hit occurred. In other words, the deeper shower penetrates in a module the larger time delay is measured. Since photons interact more deeply in the module than electrons, the measured time for photons is expected to be about 100 ps larger that for electrons of the same energy. It is interesting to note that one can suppress the dependence of the measured time on the type of particle and its energy if the loss of light in the loop were equal to $2v/(c+v)\approx0.7$, where $c$ is the speed of light in vacuum and $v$ is the effective velocity of signal propagation in fiber. However, such a degradation of the loop would reduce the photo-statistics by about 30%, and also the contribution to the shower longitudinal fluctuations will be increased. According to the Monte-Carlo simulation, the Shashlyk module time resolution obtained by our method, based on the measurement of time difference in the neighbor modules, is consistent with the actual time resolution of photon/electron detection. Photon Detection Inefficiency ----------------------------- ![The dependence of the photon detection inefficiency on the incident angle.[]{data-label="ineff"}](fig20){width="65.00000%"} To estimate the photon detection inefficiency due to fiber holes, the prototype 1 was located behind the prototype 2 array. The absence of a signal in the front array, while the total photon energy was deposited in the second, was interpreted as a penetration of the photon through the first prototype without interaction, [e.g.]{} through a fiber hole. Photons of 250 MeV were directed onto the face of the first array. The measured dependence of the photon detection inefficiency on the incident angle of the beam is compared with the simulation in Fig. \[ineff\]. The relatively coarse step in the measured angles does not allow us to accurately compare the experimental results with the simulation. However, our measurements clearly indicate that the effect of the WLS fiber holes is negligible if the incident angle is $\ge 5$ mrad. For such angles, our experimental results agree well with the probability of photon interactions in about 8.25 cm of lead and 45 cm of scintillator. It has to be pointed out that the measurement of photon detection inefficiency in this way is insensitive to the photon “disappearance” in the Calorimeter, e.g. due to photo-nuclear reactions. Long-term APD Stability ----------------------- ![ Illustration of the long-term stability and gain dependence on the rate for the APD and PMT readout. The mean values of the LED signals were measured during 20 hours run. The 250 MeV, 300 kHz photon beam was switched from Prototype 2 with APD readout to Prototype 1 with PMT readout after 10 hours of the run. Vertical lines indicate the measured R.M.S. of the LED signal. []{data-label="beam_intensity"}](fig21){width="65.00000%"} To test the long-term stability of the photo-detector gain, the APD and PMT gains were monitored using LED signals over a 20 hour period. Experimental data resulting from this test are shown in Fig. \[beam\_intensity\]. One can see a rate effect for the PMT-9903B after changing the photon beam intensity from 0 kHz to 300 kHz. No APD gain dependence on the photon beam intensity was observed after changing the photon beam rate over this range. The variation of the APD gain did not exceed 1%. Summary ======= Modules for a Shashlyk Calorimeter with energy resolution $$\sigma_E/E=(1.96\pm0.1)\%\oplus(2.74\pm0.05)\%/\sqrt{E\ \mathrm{(GeV)}}\,,$$ time resolution of $$\sigma_T = (72\pm4\ \mathrm{ps})/\sqrt{E\ \mathrm{(GeV)}} \oplus (14\pm2\ \mathrm{ps})/( E\ \mathrm{(GeV)})\,,$$ and an inefficiency in photon detection due to the nature of the modules of $$\epsilon \approx 5\times 10^{-5}\;\;(\Theta _{\mathrm{beam}} >5\;\mathrm{mrad})$$ have been constructed and experimentally tested. The characteristics experimentally determined for the Calorimeter prototype well meet the design goals of the KOPIO experiment. To optimize the Calorimeter readout design, a Monte Carlo simulation model of the light transmission in scintillator tiles and WLS fibers, the response of the photo-detector, and noise of the entire electronic chain. This model accurately describes the experimental data. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported in part by the US Department of Energy, the National Science Foundations of the USA and Russia. We thank the directorate of Institute for Nuclear Research (Moscow) and Institute for High Energy Physics (Protvino) for their support of this work. We are gratefully acknowledge all participants of the KOPIO Collaboration for numerous discussions of the Shashlyk Calorimeter performance. We are grateful to A. Sandorfi for the opportunity to use LEGS equipment in our test beam measurements. We are further indebted to E.N. Guschin, Yu.V. Musienko, and P. Rehak for many valuable discussions and suggestions. [99]{} G.S. Atoyan, [et al.]{}, Nucl. Instr. and Meth, A [320]{}, 144 (1992). I.-H. Chiang et al., [AGS Experimental Proposal 926]{} (1996). G.S. Atoian, [et. al.]{}, Nucl. Instr. and Meth. A [531]{}, 467 (2004). R. Appel [et. al.]{}, Nucl. Instr. and Meth. A [479]{}, 349 (2002). L. Aphecetche, [et al.]{} $[$ The PHENIX Collaboration $]$ Nucl. Instr. and Meth. A [499]{}, 521 (2003). G. Avoni [et al.]{} $[$The HERA-B ECAL Collaboration$]$, Nucl. Instr. and Meth. A [461]{}, 332 (2001). F. Muheim $[$For the LHCb Collaboration$]$, Nucl. Instr. and Meth. A [462]{}, 233 (2001). GEANT, Detector description and simulation tool, CERN Program Library, Long Writeup W5013 (1994). http://www1.ihep.su/ihep/ihepsc/index.html . Kuraray Co., LTD, [Scintillation materials. Plastic fibers.]{} S.P. Stoll, PHENIX Note \#245 (1996). I.G. Britvich [et al.]{} Instr. Exp. Tech. [44]{}, 472 (2001). I.G. Britvich [et al.]{} Instr. Exp. Tech. [45]{}, 644 (2002). J. Marler [et al.]{}, Nucl. Instr. and Meth. [449]{}, 311 (2000). R.J. McIntyre, IEEE Trans. Electron Devices [ED-13]{}, 164 (1966). D. H. Dowell et al., Prog. Rep. BNL 37623 (1985), p. 29. H. Huang and K. Kurita, AIP Conf. Proc. [868]{}, 3 (2006). W.L. Reiter [et. al.]{}, Nucl. Instr. and Meth. [173]{}, 275 (1980). A. Fyodorov [et. al.]{}, Nucl. Instr. and Meth. [A413]{},352-356 (1998). [^1]: This module was described in details in Ref. [@NIM_A531_467]. In this paper we will refer to it as to “earlier” Shashlyk module.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We apply the method of topological quantization to obtain the bosonic string topological spectrum propagating on a flat background. We define the classical configuration of the system, and construct the corresponding principal fiber bundle (pfb) that uniquely represents it. The topological spectrum is defined through the characteristic class of the pfb. We find explicit expressions for the topological spectrum for particular configurations of the bosonic strings on a Minkowski background and show that they lead to a discretization of the total energy of the system.' author: - Gustavo Arciniega - Francisco Nettel - Leonardo Patiño - Hernando Quevedo bibliography: - 'referencias.bib' title: Topological discretization of bosonic strings --- =5000 Introduction {#sec:int} ============ The main motivation to develop the method of topological quantization is to find an alternative to the ideas that prevail about the quantization of the gravitational fields. Nevertheless, as the method evolved we found ourselves exploring other classical theories with well established quantum counterparts, such as nonrelativistic quantum mechanics (finite number of degrees of freedom) and the bosonic string theory. So far, the canonical quantization, the most successful method to describe the discrete features of nature, still has to work out some answers for the theory of General Relativity. There are some unsolved challenges in the quantization of gravity, among them we find the problem of time, the lack of understanting of the ultimate meaning of the quantization of spacetime, the causality issues due to a fluctuant metric, the reconstruction problem and the appearance of nonrenormalizable divergences [@carlip; @isham2]. At the present time, the main candidates for a quantum theory of gravity, i.e., string theory and loop quantum gravity, both use canonical quantization as it stands. We propose, with the method of topological quantization, to extract the discrete nature of physical systems without making assumptions or putting by hand any rule external to the geometric/topological structure that we use to represent the system under study. Dirac’s idea [@dirac1931] about the discretization of the relation between the charge of a magnetic monopole and a moving electron in the field generated by the former was the starting point to propose topological quantization as an alternative way of understanding the discrete nature of physical systems. In particular, the method was utilized to analyze the case of gravitational fields [@patquev2; @patquev] and developed further for mechanical systems [@netque2; @netquemo; @netque]. The concept of topological quantization and the fundamental idea beneath these calculations has been broadly used in different contexts related to charge quantization in Yang-Mills theories [@deguchi], instantons and monopoles configurations [@schwarz; @zhong], topological models of electromagnetism [@ranada], current quantization of nanostructures [@bulgadaev] and in the theory of superconductors [@choi; @leone]. Its relation to the cohomology theory has been also analyzed [@alvarez]. Examples of topological quantization can also be found in text books where its geometric formulation is applied to physical systems described by a hermitian line bundle [@frankel]. We will generalize this approach to include the case of an arbitrary physical system in the sense that we will provide a strict mathematical definition of classical configurations. Furthermore, the complete picture of topological quantization should also include the definition of states and its dynamical evolution in terms of geometric/topological structures, which, currently, is under research. In the beginning of this program we already established the geometric representation of the physical systems and from it we defined the topological spectrum. In the next section we briefly review some general aspects of the bosonic string theory. In section \[sec:tq\] we give some elements of topological quantization and state the existence and unicity of the principal fiber bundle (pfb) that represents the physical system, followed by the general definition of topological spectrum. Section \[sec:gen\] addresses the construction of the particular pfb and the definition of the topological spectrum for a bosonic string in a general background spacetime. In section \[sec:flat\], we turn our attention to the case of the bosonic string on a Minkowski background and its pfb. The analysis of the topological spectrum for some particular configurations is carried out. Finally, in section \[sec:disc\] we discuss our results and consider their implications over the embedding energy of the string. General aspects of the bosonic string {#sec:bs} ===================================== The action integral for the free bosonic string moving in a general spacetime is given by the Nambu-Goto (N-G) action [@pol; @joh], which is proportional to the area of the worldsheet that describes the propagation of the string over a fixed background. To review this, consider a two-dimensional manifold $\mathcal{M}$ parametrized by $x^a$, $a=1,2$, and a $D$-dimensional manifold $\mathcal{N}$ with coordinates $X^\mu$, $\mu=0,\ldots, D-1$ and a metric tensor ${\boldsymbol}{G}$. Let $X:\mathcal{M} \to \mathcal{N}$ be a smooth map from $\mathcal{M}$, the worldsheet, to the spacetime $\mathcal{N}$. The induced metric on the embedded worldsheet is given by the pullback of ${\boldsymbol}{G}$ through the $X$ mapping, ${\boldsymbol}{g} = X^{\boldsymbol}{G}$, whose components are, $$\label{indmet} g_{ab} = \frac{\partial X^\mu}{\partial x^a} \frac{\partial X^\nu}{\partial x^b} G_{\mu\nu}.$$ Then, the N-G action is written out in terms of the induced metric as, $$\label{NG} S_{NG} = -T\int d^2x \sqrt{\|g\|},$$ where $T$ is the tension of the string and $g \equiv \det(g_{ab})$. The N-G action has two symmetries, the invariance under diffeomorphism on the worldsheet $x'^a = x'^a(x)$ and the invariance under diffeomorphisms on the spacetime $X'^\mu = X'^\mu(X)$. It is usual to start from an action, classically equivalent to (\[NG\]), in which an auxiliary metric field ${\boldsymbol}{\gamma}$ is introduced on the worldsheet, $$\label{Polyakov} S_P = -\frac{1}{4\pi\alpha'} \int d^2x \sqrt{\|\gamma\|} \,\gamma^{ab} g_{ab},$$ where $\alpha'$ is related to the string tension by $T= \frac{1}{2\pi\alpha'}$. This is known as the Polyakov action [@pol; @joh] and from a mathematical point of view is a harmonic map (or nonlinear sigma model) [@misner]. If we vary the Polyakov action with respect to the field ${\boldsymbol}{\gamma}$ we obtain the two-dimensional energy-momentum tensor $T_{ab} = \frac{4\pi}{\sqrt{\gamma}}\frac{\delta S_{P}}{\delta \gamma^{ab}}$ for the worldsheet, $$\label{gammag} T_{ab} = g_{ab} - \frac{1}{2}\gamma^{cd}g_{cd}\gamma_{ab} = 0,$$ which can be understood as a set of constraints that, among other things, suffice to prove the equivalence of (\[NG\]) and (\[Polyakov\]). Varying with respect to $X^\mu$ the equations of motion that determine the dynamics of the string propagating in the spacetime follow, $$\label{eqmot} \frac{1}{\sqrt{\|\gamma\|}}\partial_a \left(\sqrt{\|\gamma\|}\, \gamma^{ab}\partial_b X^\mu \right) + \Gamma^{\mu}_{\;\alpha\beta}\, \gamma^{ab} \partial_a X^\alpha \partial_b X^\beta = 0,$$ with $\partial_a \equiv \frac{\partial}{\partial x^a}$. When the background metric is $G_{\mu\nu} = \eta_{\mu\nu}$ the equations become, $$\label{eqmotflat} \partial_a \left(\sqrt{-\gamma}\,\gamma^{ab}\partial_b X^\mu \right) = 0.$$ We are interested in exact solutions to (\[eqmotflat\]) as they will be necessary to find the induced metric which has a fundamental role in the explicit calculation of the topological spectrum. The Polyakov action possesses, besides the two symmetries of the N-G action, a third invariance under the Weyl transformation, a local rescaling of the metric tensor ${\boldsymbol}{\gamma}' = e^{\omega(x)}\boldsymbol{\gamma}$. In section \[sec:flat\] we will analyze the general solution of (\[eqmotflat\]) in order to find the topological spectrum for some configurations. Fundamentals of topological quantization {#sec:tq} ======================================== A complete description of a physical system must include observables, states and its dynamical evolution. We further know that sometimes the observables have a discrete behavior. It is the aim of topological quantization to provide these three elements for any physical system from a geometric/topological outset and to find out if there is a discrete pattern in such description. Nowadays, we have established the first part of the method, which refers to the definition of the topological spectrum for some observables, meanwhile the definition of states and their dynamics remains as work in progress. We present here some basic elements for the definition of the topological spectrum. We define the classical configuration as a unique pair $(\mathcal{M}, \omega)$ composed by a Riemannian manifold $\mathcal{M}$ and a connection $\omega$ that represents the physical system. Uniqueness, in this case, means that two isomorphic manifolds with the same connection are identical classical configurations. As an example consider a gauge theory over a Minkowski spacetime $M_\eta$; this Riemannian manifold together with the connection one-form $A$, which takes values in the Lie algebra of a gauge group $G$, form the classical configuration. Furthermore, with the classical configuration we can build the pfb $\mathcal{P}$, using the Riemannian manifold $\mathcal{M}$ as the base space and the symmetry group of the theory $G$ as the structure group identical to the standard fiber. Given a local section $s_i$ which bears a local trivialization $(U_i, \phi_i)$ where $U_i \subset \mathcal{M}$ and $\phi_i : U_i \times G \to \pi^{-1}(U_i)$ [@nak], it is possible to introduce a connection $\tilde{\omega}$ on $\mathcal{P}$ through the pullback $s_i^\tilde{\omega} = \omega_i$ where $\omega_i$ is the connection $\omega$ on the open set $U_i$ in the base space $\mathcal{M}$. It can be shown that using these elements and the reconstruction theorem [@naber; @nak] a unique principal fiber bundle exists which represents the physical system for the considered classical configuration. This has been done in the context of gravitational fields [@patquev] and for mechanical systems [@netquemo]. We shall show a similar result for the case in turn. Once we have constructed the principal fiber bundle $\mathcal{P}$ from the classical configuration $(\mathcal{M}, \omega)$ the topological invariant properties of $\mathcal{P}$ can be used to characterize the physical system. This can be done employing the characteristic class of the pfb $C(\mathcal{P})$, that integrated over a cycle of $\mathcal{M}$ constitutes also an invariant of the bundle. The characteristic class $C(\mathcal{P})$, properly normalized leads to, $$\int C(\mathcal{P}) = n,$$ where $n$ is an integer called the characteristic number [@damas]. For the cases we analyze, the symmetry group of the theory may be reduced to an orthogonal group $SO(k)$ by introducing an orthonormal frame on the base manifold. Then, the characteristic class for such bundles is the Pontrjagin class $p(\mathcal{P})$, or the Euler class $e(\mathcal{P})$ in case of $k$ being an even integer. These characteristic classes can be spelled out in terms of the curvature two-form $R$ of the base space by means of the polynomials invariant under the action of the structure group $SO(k)$ [@nashsen], $$\det \left(It - \frac{R}{2\pi}\right) = \sum_{j=0}^{k} p_{k-j}(R)t^j.$$ The Euler class $e(\mathcal{P})$, only defined for even $k$, is expressed in terms of the curvature two-form $R$ of a (pseudo-)Riemannian connection on the base space as [@damas; @nashsen], $$\label{eulergen} e(\mathcal{P}) = \frac{(-1)^m}{2^{2m}\pi^m m!}\epsilon_{i_1 i_2 \cdots i_{2m}} R^{i_1}_{\;i_2} \wedge R^{i_3}_{\;i_4} \wedge \cdots \wedge R^{i_{2m-1}}_{\;i_{2m}},$$ where $2m=k$. It is clear that being in terms of the curvature form, the characteristic classes depend on some parameters $\lambda_i$, $i=1,\ldots,s$ which bear physical information of the system; thus, once we integrate the characteristic class, we end up with a discrete relation for $\lambda_i$, $$\int C(\mathcal{P}) = f(\lambda_1, \ldots, \lambda_s) = n,$$ where $n \in \mathbb{Z}$. This relationship is what we define as the topological spectrum and constitutes a discretization for some of the parameters determining the properties of the physical system of interest. In the next section we will explore in detail these definitions and the existence and uniqueness of the pfb for the bosonic string system. Bosonic string on a general background {#sec:gen} ====================================== In this section we construct explicitly the principal fiber bundle for the bosonic string on a general background. It is natural in this case to consider the worldsheet $\mathcal{M}$ embedded in the spacetime $\mathcal{N}$ as the base space provided with the induced metric ${\boldsymbol}{g} = X^{\boldsymbol}{G}$. Hence, the classical configuration is $(\mathcal{M}_{{\boldsymbol}{g}}, \omega)$, where $\omega$ is the Levi-Civita connection on $\mathcal{M}$ compatible with ${\boldsymbol}{g}$. We take the invariance under diffeomorphisms on the worldsheet as the structure group (isomorphic to the standard fiber), since this is the fundamental symmetry of the two dimensional action integral. The group of diffeomorphisms on $\mathcal{M}$ can be reduced to the orthogonal group by introducing a semiorthonormal frame. Indeed, given $\{e_i\}$ with $i = 1, 2$, an orientable orthonormal frame on $\mathcal{M}$ such that ${\boldsymbol}{g}(e_i, e_j) = \eta_{ij}$, two distinct bases are related by an orthogonal transformation, $e'_i = e_j(\Lambda^{-1})^{j}_{\;\;i}$, where $\Lambda \in SO(1,1)$. There is a one-form basis $\{\theta^i\}$ dual to the orthonormal frame from which it is possible to express the induced metric tensor as ${\boldsymbol}{g} = \eta_{ij}\; \theta^i \!\otimes \theta^j$; thus, the reduction of the symmetry group to $SO(1,1)$ is accomplished. Therefore, the principal fiber bundle $\mathcal{P}$ can be constructed from the classical configuration $(\mathcal{M}_{{\boldsymbol}{g}},\omega')$, with $\omega'$ the spin connection taking values in the Lie algebra $so(1,1)$, and $SO(1,1)$ as the structure group. This is summarized in the following result: Theorem:* A bosonic string propagating in a general background $(\mathcal{N}, {\boldsymbol}{G})$ described by the Nambu-Goto action can be represented by a unique principal fiber bundle $\mathcal{P}$, with the semi-Riemannian manifold $(\mathcal{M}, {\boldsymbol}{g})$ as the base space, $SO(1,1)$ as the structure group (identical to the standard fiber) and with a ${\boldsymbol}{g}-$compatible connection ${\boldsymbol}{\omega}$ which takes values in the Lie algebra $so(1,1)$. The proof of this theorem is completely analogous to the one that appears in previous works [@patquev; @netquemo] and we refer the reader to them for the details. It should be sufficient to mention that it rests on the reconstruction theorem for fiber bundles [@naber]. The Euler characteristic class for the principal fiber bundle $\mathcal{P}$ with a two-dimensional base space and $SO(1,1)$ as the structure group reduces from (\[eulergen\]) to $$\label{Eulerform} e(\mathcal{P}) = -\frac{1}{2\pi} R^{1}_{\;2}.$$ In the conformal gauge, using coordinates $\{\tau,\sigma\}$ in Eq.(\[indmet\]), the worldsheet metric turns out to be conformal to the two-dimensional Minkowski metric, ${\boldsymbol}{g} = g_{\sigma\sigma}{\boldsymbol}{\eta}$. In this gauge the Euler characteristic class takes the following explicit form, $$\label{Eulerformalt} e(\mathcal{P}) = -\frac{1}{4\pi} \left[ \partial_\tau \left(\frac{1}{g_{\sigma\sigma}}\partial_\tau g_{\sigma\sigma} \right) - \partial_\sigma \left(\frac{1}{g_{\sigma\sigma}}\partial_\sigma g_{\sigma\sigma} \right) \right] d\tau \wedge d\sigma.$$ Consequently, the determination of the topological spectrum reduces to the computation of the conformal factor $g_{\sigma\sigma}$ and the integral $\int C(\mathcal{P}) = n \in \mathbb{Z}$, [regardless of the background metric]{}. This shows for this particular case that the formalism of topological quantization is background independent. We will use this property in the following sections to determine specific topological spectra on diverse backgrounds. Bosonic string on a Minkowski background {#sec:flat} ======================================== The worldsheet that minimizes the action of a bosonic string propagating in a flat background, $G_{\mu\nu} = \eta_{\mu\nu}$ is described by the set of embedding functions $\{X^\mu \}$ satisfying the equations of motion $$\label{wave} \left(-\partial_{\tau}^{2} + \partial_{\sigma}^{2}\right)X^\mu(\tau,\sigma) = 0,$$ with general solution $$\label{fyg} X^\mu(\tau, \sigma) = F^\mu(\tau + \sigma) + G^\mu(\tau - \sigma).$$ We have chosen the conformal gauge to write these and the forthcoming expressions. The set (\[gammag\]) of constraint equations takes the form $$\begin{aligned} \label{const1} \left( \partial_\tau X^\mu \partial_\tau X^\nu + \partial_\sigma X^\mu \partial_\sigma X^\nu \right)\eta_{\mu\nu} &= 0, \nonumber \\ \partial_\tau X^\mu \partial_\sigma X^\nu \eta_{\mu\nu} &= 0,\end{aligned}$$ from which the conformal factor of the induced metric can be computed. Let us see how in the case of a Minkowski background choosing the light cone gauge leaves no residual gauge freedom. Consider a whole class of gauges given by [@zwiebach], $$\begin{aligned} \label{clasenormas} \hat{n}\cdot X(\tau,\sigma) &= \beta \alpha' (\hat{n}\cdot p)\tau, \nonumber \\ (\hat{n}\cdot p)\sigma &= \frac{2\pi}{\beta}\int_0^{\sigma} d\tilde{\sigma}\,\, \hat{n}\cdot \!P^\tau(\tau,\tilde{\sigma}), \end{aligned}$$ where $\hat{n}$ is a unitary vector which fixes the relation between the parameters of the worldsheet with the spacetime coordinates, and $\hat{n}\cdot X = \hat{n}^\mu X^\nu \eta_{\mu\nu}$. The constant $\beta$ determines whether we are dealing with an open ($\beta=2$) or closed ($\beta=1$) string; $P^\tau$ is the momentum density along the string, and $p$ the four momentum. Using light cone coordinates for the background space, $$\begin{aligned} X^+ &= \frac{X^0 + X^1}{\sqrt{2}}, \nonumber \\ X^- &= \frac{X^0 - X^1}{\sqrt{2}}, \nonumber \\ X^I &= X^I, \quad \text{con} \quad I = 2,\ldots, D-1,\end{aligned}$$ the line element for the Minkowski spacetime takes the following form $$\label{dsGnulas} ds_G^2 = -2dX^+dX^- + dX^IdX^J\delta_{IJ}.$$ The light cone gauge is fixed choosing the unitary vector $\hat{n}$ as, $$n^\mu = \left(-\frac{1}{\sqrt{2}},\frac{1}{\sqrt{2}},0,\ldots,0\right).$$ Then, the equations (\[clasenormas\]) that determine this specific gauge read, $$\begin{aligned} \label{sconodeluz} X^+(\tau,\sigma) &= \beta\alpha' p^+ \tau, \nonumber \\ p^+ \sigma &= \frac{2\pi}{\beta} \int_0^{\sigma} d\tilde{\sigma} P^\tau{}^+(\tau,\sigma),\end{aligned}$$ where $\hat{n} \cdot P^\tau$ is constant along the string and consequently $p^+$ too, and we notice that the gauge is completely fixed. From this we also see that the parameter $\sigma$ takes values in the interval $[0, 2\pi]$ for a closed string (periodic boundary conditions). From the constraints equations (\[const1\]) in this gauge, $$\begin{aligned} \label{constconoluz} \partial_\tau X^- &= \frac{1}{2\alpha'p^+}\left(\partial_\tau X^I\partial_\tau X^J + \partial_\sigma X^I\partial_\sigma X^J\right)\delta_{IJ}, \nonumber \\ \partial_\sigma X^- &= \frac{1}{\alpha'p^+}\partial_\tau X^I \partial_\sigma X^J \delta_{IJ},\end{aligned}$$ we observe that the component $X^-$ can be found once the transverse sector $X^I(\tau,\sigma)$, $I= 2, \ldots, D-1$, is solved; therefore, it does not represent a dynamical degree of freedom. To obtain the topological spectrum integrating the Euler form (\[Eulerformalt\]), we must first find the conformal factor of the induced metric $g_{\sigma\sigma}$, which in view of the constraints (\[constconoluz\]) reduces to $$\label{gssconoluz} g_{\sigma\sigma} = \partial_\sigma X^I \partial_\sigma X^J \delta_{IJ}.$$ It is clear now that the conformal factor only depends on the dynamics of the string, that is, the transverse sector $X^I$ for the solution to the equations of motion. Topological spectrum for the closed bosonic string {#sub:closedst} -------------------------------------------------- In this section we will obtain the topological spectrum for some particular configurations (solutions) of the closed bosonic string. In this case periodic boundary conditions must be imposed [@joh], $$\begin{aligned} X^\mu(\tau,\sigma_1) &= X^\mu(\tau, \sigma_2) \\ \partial_\sigma X^\mu(\tau,\sigma_1) &= \partial_\sigma X^\mu(\tau,\sigma_2) \\\gamma_{ab}(\tau,\sigma_1) &= \gamma_{ab}(\tau,\sigma_2),\end{aligned}$$ where $\sigma_1 = 0$ and $\sigma_2 = 2\pi$. The solutions are described through two sets of oscillation modes, which are usually interpreted as left moving $\{\tilde{\alpha}^{\mu}_{k}\}$ and right moving $\{\alpha^{\mu}_{k}\}$ waves along the string [@pol]. In the conformal gauge the solutions may be expressed as, $$\begin{gathered} \label{cssol} X^\mu(\tau,\sigma) = x^{\mu}_{0} + \sqrt{2\alpha'} \alpha^{\mu}_{0}\tau + \sqrt{\frac{\alpha'}{2}}\sum_{k = 1}^{\infty} \frac{1}{\sqrt{\omega_k}} \left( \alpha^{\mu}_{k} e^{-i\omega_k(\tau-\sigma)} \right. \\ \left. + \alpha^{\mu}_{k}{}^* e^{i\omega_k(\tau-\sigma)} + \tilde{\alpha}^{\mu}_{k} e^{-i\omega_k(\tau+\sigma)} + \tilde{\alpha}^\mu_k{}^* e^{i\omega_k(\tau+\sigma)} \right),\end{gathered}$$ where, here and throughtout this section, $\mu = 0,\ldots, D-1$ and $\omega_k = k$. The periodicity in $\sigma$ has been considered, leading to the condition that the zero modes are equal, $\tilde{\alpha}^{\mu}_{0} = \alpha^{\mu}_{0}$. The constraints can also be expressed as two independent sets of equations in terms of the modes of oscillation $$\label{csconst} L_k = \frac{1}{2} \sum_{p \in \mathbb{Z}} \alpha^{\mu}_{p - k} \alpha^{\nu}_{p} \eta_{\mu\nu} = 0 \qquad \tilde{L}_k = \frac{1}{2} \sum_{p \in \mathbb{Z}} \tilde{\alpha}^{\mu}_{p - k} \tilde{\alpha}^{\nu}_{p} \eta_{\mu\nu} = 0.$$ In the light cone gauge the dynamical fields $X^I(\tau,\sigma)$ take the same form as above (\[cssol\]), just considering the transverse index $I$ instead of the spatiotemporal $\mu$. Only these transverse fields enter in the expression for the conformal factor $g_{\sigma\sigma}$. We introduce the polar notation for the modes coefficients, $\alpha_k^I = r_k^I e^{-i\gamma_k^I}$ and $\tilde{\alpha}_k^I = \tilde{r}_k^I e^{-i\tilde{\gamma}_k^I}$, such that the solutions for the transverse fields are written as $$\begin{gathered} X^I(\tau,\sigma) = x_0^I + \sqrt{2\alpha'}\alpha_0^I \tau \\ + \sqrt{2\alpha'} \sum_{k=1}^\infty \frac{1}{\sqrt{\omega_k}} \left[ r_k^I \cos \omega_k(\tau -\sigma + \gamma_k^I) + \tilde{r}_k^I \cos \omega_k(\tau +\sigma + \tilde{\gamma}_k^I) \right].\end{gathered}$$ Then, the metric function $g_{\sigma\sigma}$ which determines the Euler characteristic class (\[Eulerformalt\]) in this gauge is given in general by a infinite sum of oscillation modes, $$\begin{gathered} g_{\sigma\sigma}(\tau,\sigma) = 2\alpha'\sum_{k,l=1}^{\infty} \sqrt{\omega_k \omega_l} \bigg[r_k^I \sin \omega_k(\tau -\sigma + \gamma_k^I) - \tilde{r}_k^I \sin \omega_k(\tau +\sigma + \tilde{\gamma}_k^I)\bigg] \\ \times \bigg[r_l^J \sin \omega_l(\tau -\sigma + \gamma_l^J) - \tilde{r}_l^J \sin \omega_l(\tau +\sigma + \tilde{\gamma}_l^J)\bigg]\delta_{IJ}.\end{gathered}$$ It then follows that the integration of the corresponding topological invariant involves the manipulation of infinite series with the consequent technical difficulties. Hence, we take into account particular configurations with only a few nonvanishing modes of oscillation that allow us to reach concrete expressions for their spectra. Topological spectrum of particular configurations {#sub:partconf} ------------------------------------------------- To investigate how the interaction of different modes of oscillation affects the geometric properties of the underlying pfb, let us consider the case of a right mode $\alpha_k^{J_1} \neq 0$ in the direction $J_1$, and a left mode in a different direction $J_2$, $\tilde{\alpha}_l^{J_2} \neq 0$. The transverse fields that involve these modes of oscillation are $$\begin{aligned} \label{solcasomin} X^{J_1} &= x_0^{J_1} + \sqrt{2\alpha'}\alpha_0^{J_1}\tau + \sqrt{2\alpha'}\,\frac{r_k^{J_1}}{\sqrt{\omega_k}} \cos \,\omega_k(\tau -\sigma + \gamma_k^{J_1}), \nonumber \\ X^{J_2} &= x_0^{J_2} + \sqrt{2\alpha'}\alpha_0^{J_2}\tau + \sqrt{2\alpha'}\,\frac{\tilde{r}_l^{J_2}}{\sqrt{\omega_l}} \cos \,\omega_l(\tau +\sigma + \gamma_l^{J_2}),\end{aligned}$$ where we have expressed the coefficients in the polar notation. In all the other transverse directions $J \neq J_1, J_2$, the fields describe only the motion of the center of mass, $X^J(\tau,\sigma) = x_0^J + \sqrt{2\alpha'} \alpha_0^J \tau$. The conformal factor for the induced metric is, $$g_{\sigma\sigma} = 2\alpha' \left[\omega_k (r_k^{J_1})^2 \sin^2 \omega_k(\tau -\sigma +\gamma_k^{J_1}) + \omega_l (\tilde{r}_l^{J_2})^2 \sin^2 \omega_l(\tau +\sigma +\gamma_l^{J_2}) \right],$$ and the Euler characteristic class, $$\label{eulercasomin} e(\mathcal{P}) = \frac{\omega_k^2 \omega_l^2 (r_k^{J_1} \tilde{r}_l^{J_2})^2 \sin 2\omega_k(\tau -\sigma +\gamma_k^{J_1}) \sin 2\omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2})}{\pi \left[ (r_k^{J_1})^2 \sin^2 \omega_k(\tau -\sigma +\gamma_k^{J_1}) + (\tilde{r}_2^{J_2})^2 \sin^2 \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2}) \right]^2} d\tau \wedge d\sigma.$$ In order to integrate the Euler form (\[eulercasomin\]) we must specify the limits in the domain of integration. For the parameter $\sigma$ the interval is $[0,2\pi]$ and is fixed, while for $\tau$ we notice that the above expression is periodic in this parameter and we may choose a complete cycle. To perform the integral it is convenient to use null-like coordinates patches that cover the entire domain of integration, for the details of the calculation we refer the reader to the appendix \[app:uno\]. In this case, it turns out that the integral of the Euler class vanishes identically, meaning that no discrete relation between the parameters $r_k^{J_1}$ and $\tilde{r}_l^{J_2}$ is established. This is so due to the lack of interaction between the modes as they point in perpendicular directions of the background spacetime. Next we calculate the topological spectrum for the string with two nonvanishing modes of oscillation in the same transverse direction, that is, a right $k$-mode $\alpha_k^J$ and a left $l$-mode $\tilde{\alpha}_l^J$. The transverse field in the relevant direction $I=J$ is $$\begin{gathered} \label{soltwomodes} \quad X^J(\tau,\sigma) = x_0^J + \sqrt{2\alpha'}\alpha_0^J \tau \\ + \sqrt{2\alpha'}\left[\frac{r_k}{\sqrt{\omega_k}} \cos \omega_k(\tau -\sigma +\gamma_k) + \frac{\tilde{r}_l}{\sqrt{\omega_l}} \cos \omega_l(\tau +\sigma + \tilde{\gamma}_l) \right],\end{gathered}$$ where we have used again the polar notation, $\alpha_k^J = r_k e^{-i\gamma_k}$ y $\tilde{\alpha}_l^J = \tilde{r}_l e^{-i\tilde{\gamma}_l}$. In all the remaining directions, $I \neq J$, the solutions describe the motion of the center of mass and only depend on $\tau$. The conformal factor is given by $$g_{\sigma\sigma}(\tau,\sigma) = 2\alpha \left[\sqrt{\omega_k}\, r_k \sin \omega_k(\tau -\sigma + \gamma_r) - \sqrt{\omega_l}\, \tilde{r}_l \sin \omega_l(\tau +\sigma + \tilde{\gamma}_l) \right]^2,$$ and the Euler form as $$\label{eulercasonomintau} e(\tau,\sigma) = -\frac{2(\omega_k\omega_l)^{\frac{3}{2}}\,r_k\tilde{r}_l \, \cos \omega_k(\tau -\sigma +\gamma_k) \cos \omega_l(\tau +\sigma +\tilde{\gamma}_l)}{\pi \left[\sqrt{\omega_k}\, r_k \sin \omega_k(\tau -\sigma +\gamma_k) - \sqrt{\omega_l}\,\tilde{r}_l \sin \omega_l(\tau +\sigma +\tilde{\gamma}_l) \right]^2} \,\, d\tau \wedge d\sigma.$$ To obtain the topological spectrum we must integrate this expression for $\sigma \in [0,2\pi]$ and a period in $\tau$. We use the coordinate transformation (\[coordcasomin\]) and cover the region of integration as explained in Appendix \[app:uno\]. For the regions $I$ and $IV$ the Euler characteristic class takes the following form $$e(x,y) = -\frac{1}{\pi} \frac{r_k \tilde{r}_l\,\sqrt{\omega_k\omega_l}}{\left(\sqrt{\omega_k}\,r_k x - \sqrt{\omega_l}\,\tilde{r}_l y \right)^2} \,\, dx \wedge dy,$$ and for the type $II$ and $III$ we have $$e(x,y) = -\frac{1}{\pi} \frac{r_k \tilde{r}_l\,\sqrt{\omega_k\omega_l}}{\left(\sqrt{\omega_k}\, r_k x + \sqrt{\omega_l}\, \tilde{r}_l y \right)^2} \,\, dx \wedge dy.$$ The outcome of the integration yields a discrete relation between the amplitudes of the oscillation modes $r_k$ and $\tilde{r}_l$, $$\label{casonominspec1} \frac{4}{\pi}\, \omega_k\omega_l\, \ln \left[\frac{\left(\sqrt{\omega_l}\,\tilde{r}_l + \sqrt{\omega_k}\,r_k\right)^2}{\left(\sqrt{\omega_l}\,\tilde{r}_l - \sqrt{\omega_k}\, r_k\right)^2} \right] = n,$$ where $n$ is an integer. This is the topological spectrum for the case of two nonvanishing modes of oscillation (right and left) in the same direction of the flat background space. We show in figure \[casonominespe\] the allowed values for $r_k$ and $\tilde{r}_l$ according to the relation (\[casonominspec1\]). ![Illustration of the topological spectrum in the case of two nonvanishing modes of oscillation (right and left) that point in the same spacetime direction. The lines on the surface show the simultaneous values of $r_k$ and $\tilde{r}_l$ that are permitted by the discrete relation for $k=\omega_k=1$ and $l=\omega_l=1$.[]{data-label="casonominespe"}](casonominespe.jpg) Now we can add another nonvanishing mode of oscillation to the ones we had in the previous case. Then, there are two modes in the direction $I=J_1$, $\alpha_k^{J_1}$, $\tilde{\alpha}_l^{J_1}$ right and left respectively, and a third right $k$-mode in a independent direction $I=J_2$, $\alpha_k^{J_2}$ (the case of including a left mode instead can be treated in a similar fashion). The relevant transverse fields are, $$\begin{gathered} \quad X^{J_1}(\tau,\sigma) = x_0^{J_1} + \sqrt{2\alpha'}\alpha_0^{J_1} \\ + \sqrt{2\alpha'} \left[\frac{r_k^{J_1}}{\sqrt{\omega_k}} \cos \omega_k(\tau -\sigma + \gamma_k^{J_1}) + \frac{\tilde{r}_l^{J_1}}{\sqrt{\omega_l}} \cos \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_1}) \right],\end{gathered}$$ $$X^{J_2}(\tau,\sigma) = x_0^{J_2} + \sqrt{2\alpha'}\alpha_0^{J_2} + \sqrt{2\alpha'} \frac{r_k^{J_2}}{\sqrt{\omega_k}} \cos \omega_k(\tau -\sigma +\gamma_k^{J_2}).$$ Integrating as the two former cases we obtain the topological spectrum which generalizes the relation (\[casonominspec1\]) $$\label{ts2} \frac{4}{\pi}\,\omega_k\omega_l\, \ln \left[ \frac{\omega_k\left(r_k^{J_2}\right)^2 + \left(\sqrt{\omega_k}r_k^{J_1} + \sqrt{\omega_l}\tilde{r}_l^{J_1}\right)^2}{\omega_k\left(r_k^{J_2}\right)^2 + \left(\sqrt{\omega_k}r_k^{J_1} - \sqrt{\omega_l}\tilde{r}_l^{J_1}\right)^2} \right] = n.$$ If a left $l$-mode in the $I=J_2$ direction is included to the modes of the preceding case we obtain the following discrete relation, $$\label{ts3} \frac{4}{\pi}\, \omega_k\omega_l\, \ln \left[ \frac{\left(\sqrt{\omega_k}r_k^{J_2} + \sqrt{\omega_l}\tilde{r}_l^{J_2}\right)^2 + \left(\sqrt{\omega_k}r_k^{J_1} + \sqrt{\omega_l}\tilde{r}_l^{J_1}\right)^2}{\left(\sqrt{\omega_k}r_k^{J_2} - \sqrt{\omega_l}\tilde{r}_l^{J_2}\right)^2 + \left(\sqrt{\omega_k}r_k^{J_1} - \sqrt{\omega_l}\tilde{r}_l^{J_1}\right)^2} \right] = n,$$ giving the guideline to generalize the topological spectrum to the case in which there are two or more nonvanishing modes (right and left) in each spacetime direction. Discretization of the energy ---------------------------- Let us now find out how the restrictions imposed by the topological spectrum reflect on a physical quantity such as the Hamiltonian function. We shall do this for the particular configuration described by the solutions (\[soltwomodes\]) that lead to the relation (\[casonominspec1\]). The Hamiltonian density in the light cone gauge is [@joh] $$\mathcal{H} = \frac{1}{4\pi\alpha'} \left[ \partial_\tau X^K \partial_\tau X^L + \partial_\sigma X^K \partial_\sigma X^L \right] \delta_{KL},$$ so that the Hamiltonian function $H = \int_0^{2\pi} \mathcal{H} d\sigma$ for this particular configuration is $$\label{hamil} H = H_0 + \omega_k r_k^2 + \omega_l \tilde{r}_l^2, \quad H_0 = \sum_K \left(\alpha_0^K\right)^2.$$ On the other hand, from the topological spectrum (\[casonominspec1\]) we can derive an expression for the term $\omega_k r_k^2 + \omega_l \tilde{r}_l^2$ which, when replaced in the above Hamiltonian, yields $$\label{hamil1} H = H_0 - 2 \sqrt{\omega_k\omega_l}\, r_k \tilde{r}_l \left( \frac { 1 + e^{n/\omega_{kl}}}{1 - e^{n/\omega_{kl}} }\right), \quad \omega_{kl} = \frac{4}{\pi} \omega_k\omega_l \ ,$$ or, equivalently for $\sqrt{\omega_k} r_k > \sqrt{\omega_l}\tilde r _l$, $$\label{hamil2} H = H_0 + \omega_k r_k^2 \left[1+\left(\frac{1-e^{n/2\omega_{kl}}}{1+e^{n/2\omega_{kl}}}\right)^2\right] \ ,$$ and for $\sqrt{\omega_k} r_k < \sqrt{\omega_l}\tilde r _l$, $$\label{hamil3} H = H_0 + \omega_k r_k^2 \left[1+\left(\frac{1+e^{n/2\omega_{kl}}}{1-e^{n/2\omega_{kl}}}\right)^2\right] \ .$$ We conclude that the topological quantization leads to a discrete Hamiltonian function. In fact, for any given bosonic string configuration, which corresponds to fixed values of the frequencies and amplitudes, the Hamiltonian can take only those values that are allowed by the discrete relationship (\[hamil2\], \[hamil3\]) which depends explicitly on the integer $n$. This is the main result of our analysis. It is interesting to notice that the spectrum of the Hamiltonian is not equidistant, an effect that can be interpreted as a result of the interaction of different modes of oscillation. For large values of $n$ the value of the Hamiltonian tends to a constant value $H_\infty = H_0 - 2\sqrt{\omega_k\omega_l}\, r_k \tilde{r}_l =H_0+2\omega_k r_k^2 $. This behavior is illustrated in figure \[tophamilfigmesh\] for both cases and in figure \[tophamilfig3d1\] for $\sqrt{\omega_k}r_k > \sqrt{\omega_l}\tilde{r}_l$. ![Graphic of the Hamiltonian function with $k=\omega_k=1$, $l=\omega_l=1$, $H_0=1$ and $r_k = 1$, showing a discrete behavior. The cases $\sqrt{\omega_l}\tilde{r}_l > \sqrt{\omega_k}r_k$ and $\sqrt{\omega_l}\tilde{r}_l < \sqrt{\omega_k}r_k$ are indicated by squares and triangles, respectively.[]{data-label="tophamilfigmesh"}](tophamilfigmesh.jpg) ![Graphic of the Hamiltonian function showing a discrete behaviour for the case $\sqrt{\omega_k}r_k > \sqrt{\omega_l}\tilde{r}_l$ with $k=\omega_k=1$, $l=\omega_l=1$, $H_0=1$ and $r_k = 1$, showing a discrete behavior.[]{data-label="tophamilfig3d1"}](tophamilfig3d1b.jpg) The complexity of the calculation rises as we increase the number of distinct modes of oscillation in a given direction. Nevertheless, based on the expressions given above for the spectra of different modes and the simple final form of the Hamiltonian, one can expect similar results for other more complicated configurations. Discussion {#sec:disc} ========== We consistently developed the method to obtain the topological spectrum for a bosonic string moving in a general background. We computed the spectra for some particular configurations in the case of a Minkowski background space. The results are discrete relations between the amplitude of the modes of oscillation that describe the dynamics of the string. These relations account for an allowed set of embeddings of the worldsheet in the background spacetime. That is, the solutions $X^\mu=X^\mu(\tau,\sigma)$ which describe the embedding of the string in the spacetime are parametrized by a set of numbers constituting the modes of oscillation; then, in principle, one might be tempted to believe that any combination of these numbers may be realized. The topological spectrum sets restrictions to the values that these numbers can take, in particular the amplitudes of the modes of oscillation, thus limiting the cases of valid embeddings, which in our perspective seems very interesting. As a consequence, the Hamiltonian corresponding to the energy of the worldsheet becomes a discrete quantity that corresponds to each allowed embedding. Due to the complexity of the computations, the above discretization was performed only for a limited number of oscillations. Nevertheless, the symmetry of the expressions for the solution and the topological spectrum allows us to conjecture the behavior of the discreteness in general. In fact, the general Hamiltonian for a closed string can be shown to be $$H = H_0 + \sum_I \sum_k \omega_k (r_k^I)^2 + \sum_J \sum_l \omega_l (\tilde{r}_l^J)^2 \ . \label{gham}$$ Then, we can infer the general spectrum $$\frac{4}{\pi}\, \ \prod_{kl} \omega_k\omega_l\, \ln \left[ \frac{\sum_I\left(\sum_k \sqrt{\omega_k}r_k^{I} + \sum_l\sqrt{\omega_l}\tilde{r}_l^{I}\right)^2 } {\sum_I \left(\sum_k \sqrt{\omega_k}r_k^{I} - \sum_l\sqrt{\omega_l}\tilde{r}_l^{I}\right)^2 } \right] = n \ , \label{gts}$$ which reduces to the spectra (\[casonominspec1\]), (\[ts2\]), and (\[ts3\]) in the corresponding limiting cases. Moreover, notice that if we consider the simple case of one single oscillation in only one direction, or one single oscillation in different directions, the expression inside the logarithm reduces to one, so that $n=0$ and no discretization appears. It then follows that oscillations in different transverse directions do not interact with each other. As soon as we consider a configuration with at least two different modes of oscillation in the same direction, the topological spectrum becomes nontrivial, leading to discrete relationships between different modes. The general spectrum (\[gts\]) could be used to rewrite the general Hamiltonian (\[gham\]) in such a way that the discreteness of the energy becomes plausible, as in the particular Hamiltonian (\[hamil1\]). The final expression of the Hamiltonian, however, will depend on the relation between different modes of oscillation as, for example, given in Eq.(\[hamil2\]). One important result of the investigation of the topological spectrum of bosonic strings is that it does not depend on the background spacetime, in the sense that the expression for the spectrum depends only on the conformal factor of the induced metric which, in turn, can easily be derived, independently of the specific form of the background metric. This opens the possibility of investigating discretization conditions for bosonic strings moving on curved backgrounds in the same manner as described in the present work. This issue is currently under investigation. It would be interesting to compare the discretization conditions which follow from topological quantization with those that appear in the context of canonical quantization. However, this comparison is not yet possible. In fact, as mentioned before, two important elements of the quantization procedure are still lacking in the approach presented here, namely, the concepts of quantum states and quantum evolution. Acknowledgments {#acknowledgments .unnumbered} =============== This work was partially support by DGAPA-UNAM No. IN106110 and No. IN108309. F. N. acknowledges support from DGAPA-UNAM (postdoctoral fellowship). Details on the integration of the Euler form {#app:uno} ============================================ In order to integrate the Euler form (\[eulercasomin\]) we must specify the limits in the domain of integration. For the parameter $\sigma$ the interval is $[0,2\pi]$ and is fixed, while for $\tau$ we notice that the expression for the Euler form is periodic in this parameter and we may choose a complete cycle. To perform the integral it is convenient to use null-like coordinates patches connected to the conformal coordinates by the following transformations, $$\begin{aligned} \label{coordcasomin} \eta = x_I &= \sin \omega_k(\tau -\sigma +\gamma_k^{J_1}) \quad \text{and} \quad \xi = y_I = \sin \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2}), \nonumber \\ x_{II} &= \sin \omega_k(\tau -\sigma +\gamma_k^{J_1}) \quad \text{and} \quad y_{II} = -\sin \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2}), \nonumber \\ x_{III} &= -\sin \omega_k(\tau -\sigma +\gamma_k^{J_1}) \quad \text{and} \quad y_{III} = \sin \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2}), \nonumber \\ x_{IV} &= -\sin \omega_k(\tau -\sigma +\gamma_k^{J_1}) \quad \text{and} \quad y_{IV} = -\sin \omega_l(\tau +\sigma +\tilde{\gamma}_l^{J_2}), \end{aligned}$$ where four types of regions are used to cover the whole domain of integration as seen in figure \[fig2\]. ![Domain of integration for the case of a right mode of oscillation with $k=\omega_k=1$ in the $J_1$ direction and left mode $l=\omega_l=2$ in the $J_2$ direction. Distinct regions are shown which correspond to the change of coordinates $I$ to $IV$.[]{data-label="fig2"}](regioncasomin.jpg) The Euler form has the following aspect in this gauge, $$\label{eulercasominxy} e(\mathcal{P}) = \pm \frac{2}{\pi}\frac{(r_k^{J_1}\tilde{r}_l^{J_2})^2\omega_k\omega_l x y}{\left[\omega_k\,(r_k^{J_1})^2 x^2 + \omega_l\,(\tilde{r}_l^{J_2})^2 y^2 \right]^2} dx \wedge dy,$$ with the positive sign for regions $I$ and $IV$ and the negative one for $II$ and $III$. The parameters take values in the intervals $x \in [-1,1]$ and $y \in [-1,1]$. To cover the entire region of integration it is necessary to consider $2kl$ regions of the type $I$ and $IV$, and the same number of type $II$ and $III$. In this case, for any type of region the integral of the Euler class vanishes, $$\label{tscasominxy} \int_{-1}^1 \int_{-1}^1 \, dxdy \, e(x,y) = 0.$$ This procedure to perform the integration is employed for the other particular configurations considered.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the systematic decomposition of the dimension nine neutrinoless double beta decay operator, focusing on mechanisms with potentially small contributions to neutrino mass, while being accessible at the LHC. We first provide a ($d=9$ tree-level) complete list of diagrams for neutrinoless double beta decay. From this list one can easily recover all previously discussed contributions to the neutrinoless double beta decay process, such as the celebrated mass mechanism or “exotics”, such as contributions from left-right symmetric models, R-parity violating supersymmetry and leptoquarks. More interestingly, however, we identify a number of new possibilities which have not been discussed in the literature previously. Contact to earlier works based on a general Lorentz-invariant parametrisation of the neutrinoless double beta decay rate is made, which allows, in principle, to derive limits on all possible contributions. We furthermore discuss possible signals at the LHC for mediators leading to the short-range part of the amplitude with one specific example. The study of such contributions would gain particular importance if there were a tension between different measurements of neutrino mass such as coming from neutrinoless double beta decay and cosmology or single beta decay.' --- IFIC/12-079, MPP-2012-134, STUPP-12-212 [Systematic decomposition of the\ neutrinoless double beta decay operator]{} 0.2cm [Florian Bonnet$^{a,}$[^1], Martin Hirsch$^{b,}$[^2], Toshihiko Ota$^{c,d,}$[^3], and Walter Winter$^{a,}$[^4]]{} [$^{a}$ Institut für Theoretische Physik und Astrophysik, Universität Würzburg,\ 97074 Würzburg, Germany ]{}\ [$^{b}$ AHEP Group, Instituto de Física Corpuscular – C.S.I.C./Universitat de Val[è]{}ncia\ Edificio de Institutos de Paterna, Apartado 22085, 46071 Val[è]{}ncia, Spain ]{}\ [$^{c}$ Max-Planck-Institut für Physik (Werner-Heisenberg-Institut),\ Föhringer Ring 6, 80805 München, Germany ]{}\ [$^{d}$ Department of Physics, Saitama University,\ Shimo-Okubo 255, 338-8570 Saitama-Sakura, Japan ]{} 0.3truecm Introduction {#sect:Intro} ============ Neutrinoless double beta ([$0\nu\beta\beta$ ]{}) decay is mostly known as a sensitive probe for Majorana neutrino masses [@Avignone:2007fu; @GomezCadenas:2011it; @Rodejohann:2012xd; @Barabash:1209.4241]. However, the mass mechanism is only one out of many possible contributions to the [$0\nu\beta\beta$ ]{}decay amplitude [@Rodejohann:2011mu; @Deppisch:2012nb]. The aim of the current paper is to provide a (tree-level) complete list of all possible contributions to the neutrinoless double beta decay dimension nine ($d=9$) operator: $$\mathcal{O} \propto \bar u \bar u \, d d \, \bar e \bar e \, \label{equ:effop}$$ From this list one can easily recover all known contributions to [$0\nu\beta\beta$ ]{}decay. More interestingly, however, we will identify a number of new possibilities to generate [$0\nu\beta\beta$ ]{}decay not discussed in the literature previously. (360,140) (10,30)(70,30) (70,30)(130,30) (70,30)(70,90)[2]{}[5]{} (70,90)(130,110) (130,70)(70,90) (230,70)(290,90) (290,90)(230,110) (290,90)(290,30)[2]{}[5]{} (290,30)(230,30) (350,30)(290,30) (130,10)(230,130) (40,25)\[t\][$\nu$]{} (100,28)\[t\][$e^-$]{} (65,60)\[r\][$W^-$]{} (100,72)\[t\][$d$]{} (100,108)\[b\][$u$]{} (320,25)\[t\][$\nu$]{} (260,28)\[t\][$e^-$]{} (295,60)\[l\][$W^-$]{} (260,72)\[t\][$d$]{} (260,108)\[b\][$u$]{} (180,58)\[b\][$0\nu\beta\beta$]{} The Black Box theorem [@Schechter:1981bd; @Nieves:1984sn; @Takasugi:1984xr] states that since observation of [$0\nu\beta\beta$ ]{}indicates that lepton number is not conserved, it proves that neutrinos must be Majorana particles.[^5] Graphically the theorem can be depicted as shown in Fig. \[Fig:BlackBox\]: If [$0\nu\beta\beta$ ]{}decay is observed, Majorana neutrino masses are generated at least at the 4-loop order, which is a model-independent statement. This does not mean that the [$0\nu\beta\beta$ ]{}decay contribution from Fig. \[Fig:BlackBox\] is the leading contribution to neutrino mass; in fact, in specific models, neutrino mass is often generated at a lower loop order. A recent calculation [@Duerr:2011zd] indeed confirms that the neutrino mass generated through this 4-loop diagram with the [$0\nu\beta\beta$ ]{}operator of the size of the current limits is roughly $m_{\nu} \simeq {\cal O}(10^{-24})$ eV. Obviously, this number is too small to explain the neutrino masses observed in oscillation experiments, and also many orders of magnitude smaller than the current sensitivity of [$0\nu\beta\beta$ ]{}decay via the mass mechanism. The correct interpretation of the black box theorem thus is: If [$0\nu\beta\beta$ ]{}decay is observed, neutrinos are Majorana particles, whether the contribution from neutrino mass dominates [$0\nu\beta\beta$ ]{}decay or not. Up to now, [$0\nu\beta\beta$ ]{}decay has not been observed. The best half-life limits on [$0\nu\beta\beta$ ]{}decay come from experiments on two isotopes: $^{76}$Ge and $^{136}$Xe. The Heidelberg-Moscow collaboration gives $T^{0\nu\beta\beta}_{1/2}(^{76}{\rm Ge}) \ge 1.9 \cdot 10^{25}$ yr [@KlapdorKleingrothaus:2000sn],[^6] while the recent results from EXO-200 and KamLAND-ZEN quote $T^{0\nu\beta\beta}_{1/2}(^{136}{\rm Xe}) \ge 1.6 \cdot 10^{25}$ yr [@Auger:2012ar] and $T^{0\nu\beta\beta}_{1/2}(^{136}{\rm Xe}) \ge 1.9 \cdot 10^{25}$ yr [@Collaboration:2012zm], both at the 90 % CL. There is, however, reasonable hope that the half-lifes in excess of $10^{26}$ yr will be probed within the next few years, since a number of next generation [$0\nu\beta\beta$ ]{}experiments are under construction or already taking data. For recent reviews and a list of experimental references, see for example [@Barabash:1209.4241; @GomezCadenas:2011it]. Moreover, proposals for ton-scale next-to-next generation [$0\nu\beta\beta$ ]{}experiments claim that even sensitivities in excess $T^{0\nu\beta\beta}_{1/2} \sim 10^{27}$ yr can be reached for $^{136}$Xe [@KamLANDZen:2012aa; @MacLallen:2012aa] and $^{76}$Ge [@Abt:2004yk; @Guiseppe:2011me]. For a brief summary of the long-term prospects for [$0\nu\beta\beta$ ]{}experiments, see for example [@Barabash:2011fs]. The interpretation of these half-life limits in terms of particle physics parameters requires assumptions, such as which contribution dominates the [$0\nu\beta\beta$ ]{}decay amplitude. If neutrinos have Majorana masses, [$0\nu\beta\beta$ ]{}decay can be mediated by Majorana neutrino propagation, depending on the magnitude of the effective neutrino mass given by ${\langle m_{\nu}\rangle}=\sum_j U_{ej}^2 m_j$. This mechanism is hence forth referred to as the [the mass mechanism]{}. The mass mechanism has attracted most of the attention within the community. The reason for this “bias” is rather straightforward: Neutrinos exist and oscillation experiments [@Fukuda:1998mi; @Ahmad:2002jz; @Eguchi:2002dm; @Abe:2011sj; @Adamson:2011qu; @Abe:2011fz; @An:2012eh; @Ahn:2012nd] have shown that (at least two) neutrinos have non-zero masses. Thus, if neutrinos are indeed Majorana particles, the mass mechanism is guaranteed to give a contribution to the [$0\nu\beta\beta$ ]{}decay amplitude — in this sense the mass mechanism is the minimal possibility to generate [$0\nu\beta\beta$ ]{}decay. With the assumption of the mass mechanism being the dominant contribution, the current limits for the half-life of [$0\nu\beta\beta$ ]{}decay [@KlapdorKleingrothaus:2000sn; @Auger:2012ar] correspond to the limits ${\langle m_{\nu}\rangle}\lsim (0.2-0.35)$ eV \[${\langle m_{\nu}\rangle}\lsim (0.17-0.30)$ eV\] for $^{76}$Ge \[$^{136}$Xe\] using the latest QRPA matrix elements of [@Faessler:2012ku] to ${\langle m_{\nu}\rangle}\lsim 0.53$ eV \[${\langle m_{\nu}\rangle}\lsim 0.34$ eV\] with the matrix elements calculated within the shell model [@Menendez:2008jp; @Menendez:2011zza]. Future limits of order $T^{0\nu\beta\beta}_{1/2} \sim 10^{27}$ yr will then probe ${\langle m_{\nu}\rangle}\lsim (0.02-0.06)$ eV that is of the order of the mass scale suggested by atmospheric neutrino oscillations, $\sqrt{\Delta m^2_{\rm atm}} \simeq 0.05$ eV [@Tortola:2012te]. If the next generation of [$0\nu\beta\beta$ ]{}decay experiments detects a signal, one might expect that future cosmological data [@Lesgourgues:2006nd; @Hannestad:2010kz; @Wong:2011ip] also provide indications for non-zero neutrino masses. Contradicting results from [$0\nu\beta\beta$ ]{}(observation) and cosmology (limit) then might point to a non-standard explanation for [$0\nu\beta\beta$ ]{}decay, see [@Bergstrom:2011dt] for a detailed discussion of the interplay of different data. Contributions to the [$0\nu\beta\beta$ ]{}decay rate can be divided into a short-range [@Pas:2000vn] and a long-range [@Pas:1999fc] part. The long-range contributions, which the neutrino mass mechanism belongs to, can lead in some cases to very stringent limits on the new physics scale $\Lambda \gsim$ $\lambda^{\mathrm{eff}}_{\rm LNV}\times$($10^2 - 10^3$) TeV, where $\lambda^{\mathrm{eff}}_{\rm LNV}$ is some effective Lepton Number Violating (LNV) coupling that depends on the model under consideration. Therefore, for some of the “exotic” mechanisms discussed in the literature, falling into this category, the half-life limits from [$0\nu\beta\beta$ ]{}decay themselves yield the most stringent bounds. On the other hand, in the case of the short-range contribution, i.e., the [$0\nu\beta\beta$ ]{}amplitude is mediated only by heavy mediators with masses at a high energy scale $\Lambda$, the effective Lagrangian describing [$0\nu\beta\beta$ ]{}decay is simply proportional to $1/\Lambda^5$. In that case, the next generation [$0\nu\beta\beta$ ]{}decay experiments are sensitive to new physics at scales $\Lambda \gsim$ (few) TeV. Data from the LHC will probe physics at the TeV scale, i.e., of a similar scale as the sensitivity of [$0\nu\beta\beta$ ]{}decay experiments in case of short-range contributions. In fact, first limits on particular models have already been published. Just to mention one particular example, in Left-Right (LR) symmetric models, [$0\nu\beta\beta$ ]{}decay can be generated by $W_R-N-W_R$ exchange, where $N$ is a heavy Majorana neutrino and $W_{R}$ is the charged gauge boson of the right-handed $SU(2)$ [@Riazuddin:1981hz].[^7] Using the nuclear matrix elements of [@Hirsch:1996qw], the limit on the half-life [@Auger:2012ar] corresponds to $\langle m_N\rangle = m_{W_R} \gsim 1.3$ TeV (assuming that the gauge coupling of the right-handed and left-handed $SU(2)$s are equal), while the recent analysis by the ATLAS [@ATLAS:2012ak] and the CMS collaborations [@CMS:PAS-EXO-12-017] give $m_{W_R} \ge (2.3-2.5)$ TeV for $m_N \lsim 1.3$ TeV. [^8] Because of this complementarity of [$0\nu\beta\beta$ ]{}decay and LHC, we will pay particular attention to the short-range part of the [$0\nu\beta\beta$ ]{}decay amplitude. However, our decomposition of the $d=9$ operator, which will be shown in Sec. \[Sec:decom\], is general, and the corresponding limits for the long-range part can be easily derived as well, using the recipes described below and in the appendix. Our paper is not the first work attempting a systematic analysis of the $d=9$ [$0\nu\beta\beta$ ]{}decay operator, relevant earlier papers include [@Pas:1999fc; @Pas:2000vn; @Babu:2001ex; @deGouvea:2007xp; @delAguila:2012nu]. The authors of [@Pas:1999fc; @Pas:2000vn] worked out a general Lorentz-invariant parametrisation for the [$0\nu\beta\beta$ ]{}decay rate. This approach is motivated from the nuclear physics point of view of [$0\nu\beta\beta$ ]{}decay. At low energies, adequate for [$0\nu\beta\beta$ ]{}decay studies, any of the [$0\nu\beta\beta$ ]{}diagrams in which heavy mediation fields are inserted among the six fermions $\bar{u}\bar{u} dd \bar{e} \bar{e}$ will be reduced to a finite set of combinations of the hadronic and the leptonic currents corresponding to a basic set of nuclear matrix elements. This approach leaves the LNV parameters in the [$0\nu\beta\beta$ ]{}decay rate unspecified, and our current work can be understood as providing a (tree-level) complete list of all possible ultraviolet completions (“models”) for [$0\nu\beta\beta$ ]{}decay. As expected, at low energies any information on the particle models is reduced then to one (or a combination of more than one) of the coefficients $\epsilon_i$ of the effective [$0\nu\beta\beta$ ]{}currents presented in [@Pas:1999fc; @Pas:2000vn]. Our approach also has some overlap with [@Babu:2001ex; @deGouvea:2007xp]. These authors write down all effective LNV operators from $d=5$ (the famous Weinberg operator [@Weinberg:1979sa]) to $d=11$. The main motivation of those papers is to identify all possible Majorana neutrino mass models via the effective LNV operators [@Babu:2001ex]. This effective operator treatment allows to estimate the scale $\Lambda$ at which new physics appears, if these operators give neutrino masses or a [$0\nu\beta\beta$ ]{}decay amplitude of the order of the current experimental sensitivity [@deGouvea:2007xp]. However, our current work is complementary to these papers, in that we list all possible [decompositions]{} of a particular LNV operator, that is the $d=9$ [$0\nu\beta\beta$ ]{}operator. We also go one step further than these works in the estimation of the bounds on the LNV operator, by making contact with the nuclear matrix element calculation of [@Pas:1999fc; @Pas:2000vn], instead of simply relying on dimensional arguments.[^9] Finally, there is also the recent paper [@delAguila:2012nu], where the authors study [$0\nu\beta\beta$ ]{}decay from an effective Lagrangian point of view. Operators of $d>9$ are considered, in which the Standard Model (SM) Higgs doublets are additionally inserted to the $d=9$ operator Eq. . Note, however, that [@delAguila:2012nu] considers only the case when new physics is confined to the leptonic part of the [$0\nu\beta\beta$ ]{}decay amplitude. The rest of this paper is organised as follows. In the next section, as a preparation for our approach, we will first recapitulate the characteristics of the long-range and short-range contributions to [$0\nu\beta\beta$ ]{}decay. Then, in Sec. \[Sec:decom\] we will focus on the decomposition of effective operators to find all possible models generating the $d=9$ [$0\nu\beta\beta$ ]{} operator. Through an example, we will study the crucial role that the LHC can play in discriminating such models in Sec. \[sec:example\]. The relations between the list of particle models and the general decay rate of [@Pas:1999fc; @Pas:2000vn] are given in tabular form in the appendix for the short-range part for the case of scalar exchange. The corresponding relations for the other parts of the decay rate can be easily derived from the recipes spelled out below and in the appendix. Model-independent parametrisation of the $\boldsymbol{0\nu\beta\beta}$ decay rate ================================================================================= A general Lorentz-invariant parametrisation of new physics contributions to [$0\nu\beta\beta$ ]{}has been developed in [@Pas:1999fc; @Pas:2000vn]. This formalism allows to derive limits on any LNV new physics contributing to [$0\nu\beta\beta$ ]{}decay without recalculation of nuclear matrix elements. In order to make contact with this formalism, we recapitulate the main results and definitions of [@Pas:1999fc; @Pas:2000vn] in this section. The total amplitude of [$0\nu\beta\beta$ ]{}is most conveniently divided into two parts: Long-range and short-range contributions, see Fig. \[Fig:LongShort\]. Long-range contributions ------------------------ (400,110) (10,10)(40,10) (40,10)(70,10) (40,10)(40,30)[2]{}[5]{} (40,30)(70,30) (40,30)(40,50) (40,50)(70,50) (40,50)(40,70)[2]{}[5]{} (10,70)(40,70) (40,70)(70,70) (5,10)\[r\][$d$]{} (5,70)\[r\][$d$]{} (75,70)\[l\][$\overline{u}$]{} (75,10)\[l\][$\overline{u}$]{} (75,50)\[l\][$e^-$]{} (75,30)\[l\][$e^-$]{} (35,40)\[r\][$\nu$]{} (35,60)\[r\][$W^-$]{} (35,20)\[r\][$W^-$]{} (40,5)\[t\][(a)]{} (110,10)(140,10) (140,10)(170,10) (140,10)(140,30)[2]{}[5]{} (140,30)(170,30) (140,30)(140,50) (110,60)(140,60) (140,60)(170,70) (140,60)(170,50) (140,60)[10]{}[0.5]{} (105,10)\[r\][$d$]{} (105,60)\[r\][$d$]{} (175,70)\[l\][$\overline{u}$]{} (175,10)\[l\][$\overline{u}$]{} (175,50)\[l\][$e^-$]{} (175,30)\[l\][$e^-$]{} (135,40)\[r\][$\nu$]{} (135,20)\[r\][$W^-$]{} (140,5)\[t\][(b)]{} (210,20)(240,20) (240,20)(270,30) (240,20)(270,10) (240,30)(240,50) (210,60)(240,60) (240,60)(270,70) (240,60)(270,50) (240,60)[10]{}[0.5]{} (240,20)[10]{}[0.5]{} (205,20)\[r\][$d$]{} (205,60)\[r\][$d$]{} (275,70)\[l\][$\overline{u}$]{} (275,10)\[l\][$\overline{u}$]{} (275,50)\[l\][$e^-$]{} (275,30)\[l\][$e^-$]{} (235,40)\[r\][$\nu$]{} (240,5)\[t\][(c)]{} (310,30)(340,30) (310,50)(340,50) (350,45)(380,70) (350,45)(380,45) (350,35)(380,35) (350,35)(380,10) (345,40)[15]{}[0.5]{} (305,30)\[r\][$d$]{} (305,50)\[r\][$d$]{} (385,70)\[l\][$\overline{u}$]{} (385,10)\[l\][$\overline{u}$]{} (385,50)\[l\][$e^-$]{} (385,30)\[l\][$e^-$]{} (345,5)\[t\][(d)]{} Consider first the long-range part. Here, we can sub-divide the amplitudes into parts (a)-(c) as shown in the figure. In case (a), a massive Majorana neutrino is exchanged between two SM charged current vertices, while cases (b) and (c) contain one and two (unspecified) non-standard interactions respectively, indicated by the black blobs. At low energy, we can write the relevant part of the effective Lagrangian with the leptonic ($j$) and hadronic ($J$) charged currents as $$\begin{aligned} \nonumber \mathcal{L}^{\text{4-Fermi}} &=& \mathcal{L^{\rm SM}} + \mathcal{L^{\rm LNV}} \\ & =& \frac{G_F}{\sqrt{2}} \left[ j^{\mu}_{V-A} J_{V-A,\mu} + \hspace{-0.3cm} \sum_{\begin{minipage}{1.5cm} {\tiny $\alpha,\beta \neq V-A$ } \end{minipage}} \hspace{-0.3cm} \epsilon_{\alpha}^{\beta} \hspace{0.1cm} j_{\beta}J_{\alpha}\right]\,.\label{eq:defLR}\end{aligned}$$ Here, we follow the notations of $j$ and $J$ adopted in [@Pas:1999fc], which are[^10] $$\begin{gathered} \label{eq:CurrLR} J^{\mu}_{V\pm A} = (J_{R/L})^{\mu} \equiv \overline{u}\gamma^{\mu}(1\pm\gamma_5)d\,, \qquad j_{V\pm A}^{\mu} \equiv \overline{e}\gamma^{\mu}(1\pm\gamma_5)\nu\,, \\ \nonumber J_{S\pm P} = J_{R/L} \equiv \overline{u}(1\pm\gamma_5)d\,, \qquad j_{S\pm P}\equiv\overline{e}(1\pm\gamma_5)\nu\,, \\ \nonumber J^{\mu\nu}_{T_{R/L}} = (J_{R/L})^{\mu \nu} \equiv \overline{u}\gamma^{\mu\nu}(1\pm\gamma_5)d\,, \qquad j_{T_{R/L}}^{\mu\nu} \equiv\overline{e}\gamma^{\mu\nu}(1\pm\gamma_5)\nu\ ,\end{gathered}$$ where the Lorenz tensor matrix $\gamma^{\mu \nu}$ is defined as $\gamma^{\mu \nu} = \frac{\rm i}{2} [\gamma^{\mu}, \gamma^{\nu}]$. Recall that $P_{L/R} =\frac{1}{2}(1 \mp \gamma_5)$ and we will use the short-hand notation $L$ and $R$ for left-handed and right-handed fermions, respectively. The first term of Eq. is the SM charged current interaction, and the second term contains the new physics contributions, which do not take the Lorenz structure of the standard four-Fermi interaction $(V-A)(V-A)$. The coefficients $\epsilon_{\alpha}^{\beta}$ for the exotic four-Fermi interactions are normalised to the SM charged current strength $G_F/\sqrt{2}$. If these dimensionless coefficients take numbers smaller than one, diagram (c) in Fig. \[Fig:LongShort\] is of order $\epsilon^2$ and becomes immediately sub-dominant in the [$0\nu\beta\beta$ ]{}amplitudes. The neutrino propagator in diagrams (a)-(c) contains two terms: $m_{\nu} + {q\!\!\!/}$, the mass and the momentum terms. Since the charged current for leptons in the SM is purely left-handed, it picks out the $m_{\nu}$ part, i.e., the amplitude of the (standard) mass mechanism of [$0\nu\beta\beta$ ]{}is proportional to $m_{\nu}$. Clearly, then not all $\epsilon_{\alpha}^{\beta}$ can be constrained from [$0\nu\beta\beta$ ]{}decay, due to the absence of a lower bound on $m_{\nu}$. On the other hand, if the new physics in diagram (b) generates a right-chiral lepton interaction $(j_{V+A}, j_{S+P}, j_{T_{R}})$, the ${q\!\!\!/}$-term in the neutrino propagator will enter the amplitude. The size of the 3-momentum $\|{\vec q}\|$ can be estimated from the typical inter-nucleon distance of two neutrons in the nucleus to be of the order of 100 MeV. Therefore, the amplitudes with ${q\!\!\!/}$-terms are highly enhanced in comparison with those with the $m_{\nu}$-term.[^11] For this reason, the coefficients $\epsilon_{\alpha}^{\beta}$ with right-chiral leptonic interactions, are heavily constrained by [$0\nu\beta\beta$ ]{}decay. The hadronic and leptonic currents are best defined as currents of definite chirality (as defined at Eq. ) to take care of this fact. In Table \[Tab:LimitsLong\] we give the updated bounds for all $\epsilon_{\alpha}^{\beta}$, which are taken from [@Deppisch:2012nb]. Note that the index $\beta$ for the leptonic current in the table takes neutrino interactions with the chirality $R$ in the exotic four-Fermi interaction, while the hadronic currents can be of either $L$- or $R$-type. Note also that, while the $\epsilon_{\alpha}^{\beta}$ are defined as dimensionless coefficients, they scale like $\epsilon_{\alpha}^{\beta} \propto (\lambda^{\mathrm{eff}}_{\rm LNV}/\Lambda)^2$. Isotope $\|\epsilon^{V+A}_{V-A}\|$ $\|\epsilon^{V+A}_{V+A}\|$ $\|\epsilon^{S+P}_{S-P}\|$ $\|\epsilon^{S+P}_{S+P}\|$ $\|\epsilon^{TR}_{TL}\|$ $\|\epsilon^{TR}_{TR}\|$ ------------ -------------------------- -------------------------- -------------------------- -------------------------- ------------------------ ------------------------ $^{136}$Xe $2.8 \cdot 10^{-9}$ $5.6 \cdot 10^{-7}$ $6.8 \cdot 10^{-9}$ $6.8 \cdot 10^{-9}$ $4.8 \cdot 10^{-10}$ $8.1\cdot 10^{-10}$ : Limits on effective long-range interactions from $T^{\text{{$0\nu\beta\beta$ }}}_{1/2}({}^{136}\text{Xe}) \gsim 1.6 \cdot 10^{25}$ ys [@Auger:2012ar] which corresponds to approximately $\langle m_{\nu} \rangle \lsim 0.35 $ eV in the mass mechanism. These limits are taken from [@Deppisch:2012nb] and are derived assuming only one $\epsilon$ is different from zero at a time.[]{data-label="Tab:LimitsLong"} Short-range contributions ------------------------- The short-range contributions encompass all processes where no light neutrinos are exchanged, and can be understood as one $d=9$ effective vertex diagram as shown in diagram (d) in Fig. \[Fig:LongShort\]. In this case, one can use the basis of low energy hadronic currents $J$ as defined in Eq. (\[eq:CurrLR\]), while for the currents $j$ of two electrons, one defines $$\begin{aligned} \label{eq:CurrSR} j_{L/R} &\equiv&\overline{e}(1 \mp \gamma_5)e^c\,,\\ \nonumber (j_{L/R})^{\mu} &\equiv&\overline{e}\gamma^{\mu}(1 \mp \gamma_5)e^c , \end{aligned}$$ to express the effective Lagrangian for short-range [$0\nu\beta\beta$ ]{} as [@Pas:2000vn] $$\begin{aligned} \mathcal{L}^{\text{eff}} =\frac{G_F^2}{2}m_{P}^{-1} \left[ \epsilon_1 JJj+\epsilon_2 J^{\mu\nu}J_{\mu\nu}j +\epsilon_3 J^{\mu}J_{\mu}j+\epsilon_4 J^{\mu}J_{\mu\nu}j^{\nu} +\epsilon_5J^{\mu}Jj_{\mu} \right]\, , \label{eps_short}\end{aligned}$$ where $m_{P}$ is the mass of proton. Here we omitted the indices for clarity. However, if chirality changes play a role in the value of the [$0\nu\beta\beta$ ]{}decay rate, one needs to maintain the chirality indices and define $\epsilon_{i}=\epsilon_{i}^{xyz}$, with $x,y,z \in \{L,R\}$; cf., App. \[app:decay\] for details. Since $(j_{R})^{\mu}=-(j_{L})^{\mu}$, we define $(j)^{\mu}=(j_{R})^{\mu}=-(j_{L})^{\mu}$. Current limits on short-range type $\epsilon_i$ are summarised in Tab. \[Tab:LimitsShort\]. Note that, different from the $\epsilon_{\alpha}^{\beta}$ of the long-range part, here $\epsilon_i$ scale as $\epsilon_i \propto (\lambda^{\rm eff}_{\rm LNV})^4/\Lambda^5$. For $\lambda^{\rm eff}_{\rm LNV} \simeq g_L$ the limits given in Tab. \[Tab:LimitsShort\] then correspond to $\Lambda \gsim (1-3)$ TeV. Isotope $\|\epsilon_1\|$ $\|\epsilon_2\|$ $\|\epsilon_3^{LLz (RRz)} \|$ $\|\epsilon_3^{LRz(RLz)}\|$ $\|\epsilon_4\|$ $\|\epsilon_5\|$ ------------ --------------------- --------------------- ----------------------------- --------------------------- --------------------- --------------------- $^{136}$Xe $2.6 \cdot 10^{-7}$ $1.4 \cdot 10^{-9}$ $1.1 \cdot 10^{-8}$ $1.7 \cdot 10^{-8}$ $1.2 \cdot 10^{-8}$ $1.2 \cdot 10^{-7}$ : Limits on effective short-range interactions. These limits are taken from [@Deppisch:2012nb] and are derived assuming only one $\epsilon$ is different from zero at a time.[]{data-label="Tab:LimitsShort"} General decomposition of the $\boldsymbol{d=9}$ $\boldsymbol{0\nu\beta\beta}$ decay operator {#Sec:decom} ============================================================================================ (420,110) (10,10)(50,50) (50,50)(10,90) (50,50)(90,50)[3]{}[4]{} (90,50)(90,90) (90,50)(130,50) (130,50)(130,90) (130,50)(170,50)[3]{}[4]{} (170,50)(210,10) (170,50)(210,90) (50,50)[3]{} (90,50)[3]{} (130,50)[3]{} (170,50)[3]{} (250,10)(290,50) (250,90)(290,50) (290,50)(370,50)[3]{}[8]{} (330,50)(330,90)[3]{}[4]{} (330,90)(290,130) (330,90)(370,130) (370,50)(410,10) (370,50)(410,90) (290,50)[3]{} (330,50)[3]{} (330,90)[3]{} (370,50)[3]{} (295,45)\[t\][$\text v_1$]{} (330,45)\[t\][$\text v_2$]{} (340,85)\[t\][$\text v_3$]{} (370,45)\[t\][$\text v_4$]{} (110,5)\[t\][[Topology I]{}]{} (330,5)\[t\][[Topology II]{}]{} (55,45)\[t\][$\text v_1$]{} (90,45)\[t\][$\text v_2$]{} (130,45)\[t\][$\text v_3$]{} (170,45)\[t\][$\text v_4$]{} The $d=9$ effective operator generating [$0\nu\beta\beta$ ]{}decay at the quark level can be written as in Eq. . In this section, we decompose this operator, following the techniques developed in [@Gavela:2008ra; @Bonnet:2009ej; @Bonnet:2011yx; @Bonnet:2012kz], in terms of the SM quantum numbers of the mediators. Note that the results obtained in this section are valid for both short- and long-range contributions, while the quantitative impact of the contribution to [$0\nu\beta\beta$ ]{} depends on the Lorentz nature. Details are left for the appendix, where we tabulate all possibilities for the short-range part mediated by fermions and scalars. At tree-level, there are only two possible topologies for this operator, which are shown in Fig. \[Fig:0nbbTopologies\]. For topology I (T-I) the internal particles between vertices $\text{v}_1-\text{v}_2$ and $\text{v}_3-\text{v}_4$ can be either scalars (S) or vectors (V), while the particle between $\text{v}_2-\text{v}_3$ must be a fermion (F). The topology thus contains three classes of diagrams: VFV, SFS and SFV. The inner particles for topology II (T-II), on the other hand, must all be either scalars or vectors, all possible combination in principle can occur (SSS, VVV, VVS and SSV). One has multiple choices for assigning the fermions to the outer legs. Once a particular assignment is chosen, the electric charge and the colour of the internal particles are fixed (the latter up to a two-fold ambiguity), but not the $U(1)_{Y}$ hypercharge. The assignments of hypercharge are fixed as well, once the chiralities of the six outer fermions are determined. We will come back to this point later in the discussion, and more details are given in the appendix. We will now discuss the general decompositions of T-I and T-II in turn. Decomposition of topology I --------------------------- In Tab. \[Tab:TopoI\] we list the general decompositions for the [$0\nu\beta\beta$ ]{}decay operator for T-I. The chiralities of the outer fermions are left unspecified here for a more compact presentation. In total there are 18 (times 2 for the choice of colour) possibilities for realizing T-I. Because of the $SU(3)$ multiplication rules, ${\bf \bar{3}} \otimes {\bf \bar{3}} = {\bf 3}_a \oplus {\bf \bar{6}}_s$ and ${\bf 3} \otimes {\bf \bar{3}} = {\bf 1} \oplus {\bf 8}$, there are always two possible colour assignments for the internal particles. The table is valid for all three possible classes of diagrams (VFV, SFS and SFV). In some of the cases listed in Tab. \[Tab:TopoI\], only VFV or SFV exchange is possible, because in the decomposition ${\bf \bar{3}} \otimes {\bf \bar{3}} = {\bf 3}_a \oplus {\bf \bar{6}}_s$ the coupling of a scalar to two identical quarks (two ${\bf \bar{3}}$) vanishes for the ${\bf 3}_a$. These affects all the cases in T-I-3, T-I-4-ii and T-I-5-ii, see table for nomenclature. Each of the models listed in this table leads to an effective operator with a different Lorentz structure, which needs to be projected onto the basis shown in Eqs. (\[eq:defLR\]) and (\[eps\_short\]), once the chiralities are fixed. The projection can be done by Fierz transformations and the transformations of the SM gauge group indices. This allows to identify immediately if a model gives important contributions to [$0\nu\beta\beta$ ]{}decay. More details and results of this procedure are shown in the appendix. From this table we can identify all (T-I) contributions discussed in the literature, once the nature of the internal bosons and the chirality of the outer fermions are chosen. For example, the mass mechanism corresponds to T-I-1-i, with the bosons being vectors ($W^{\pm}$) and all the outer fermions being left-handed. In this case, the quantum numbers of the internal fermion $\psi$ are equal to those of a (light) neutrino, and the mass term is picked out from the propagator. Other examples can be identified as easily. To list a few more, the afore-mentioned $W_R-N-W_R$ exchange diagram is also contained in T-I-1-i with vectors ($W_{R}^{\pm}$), all outer fermions now being right-handed. Since $N$ must be heavy, however, this is now a short-range contribution. Concrete models can lead to the occurrence of more than one of the operators listed, an example is provided by (trilinear) R-parity violating (RPV) supersymmetry (SUSY). The six short-range diagrams [@Hirsch:1995zi; @Hirsch:1995ek] for the RPV SUSY mechanism of [$0\nu\beta\beta$ ]{}decay [@Mohapatra:1986su], correspond to class SFS and are identified as T-I-1-i (${\tilde e}-\tilde{\chi}^{0}-{\tilde e}$ diagram), T-I-2-i-b (${\tilde e}-\tilde{\chi}^{0}-{\tilde d}$), T-I-2-ii-b (${\tilde e}-\tilde{\chi}^{0}-{\tilde u}$), T-I-2-iii-a (${\tilde u}-\tilde{\chi}^{0}/\tilde{g}-{\tilde d}$), T-I-4-i (${\tilde u}-\tilde{\chi}^{0}/\tilde{g}-{\tilde u}$) and T-I-5-i (${\tilde d}-\tilde{\chi}^{0}/\tilde{g}-{\tilde d}$). In those diagrams, the neutralino $\tilde{\chi}^{0}$ corresponds to $\psi(0,{\bf 1})$, while the gluino $\tilde{g}$ corresponds to $\psi(0,{\bf 8})$. Finally, the leptoquark (LQ) mechanism of [@Hirsch:1996qy; @Hirsch:1996ye] is a long-range contribution of the class SFV with T-I-2-i-b and T-I-2-ii-b. The internal fermion $\psi(0,{\bf 1})$ is again identified with a light neutrino. A few more comments on Tab. \[Tab:TopoI\] might be in order. There are a total of six possibilities in which the intermediate fermion transforms $\psi(0,{\bf 1})$ under the gauge symmetries $(U(1)_{\rm em}, SU(3)_{c})$. Only they can lead to long-range contributions, all the other models in the list are necessarily of the short-range type. Among those six, only the cases marked (a) or (b) in the column “Long Range?” can lead to interesting constraints, since the remaining three cases marked as (c) are suppressed with $\epsilon^2$. Note, however, that all of these cases can also be of short-range type. There are 12 (times two) cases listed, which require that the internal fermion has a fractional electric charge. As far as we know, none of them have been discussed in the literature before. All of these new “models” not only require fractionally charged fermions, but also exotic bosons. The latter can be doubly-charged bileptons [@Cuypers:1996ia], diquarks, or leptoquarks [@Buchmuller:1986zs]. The [$0\nu\beta\beta$ ]{}decay process violates lepton number $L$. We can easily identify the different possibilities for LNV from Tab. \[Tab:TopoI\] (cf., Fig. \[Fig:0nbbTopologies\], left panel). If the internal fermion is neutral, it can have a Majorana mass and a mass insertion leads then to $\Delta(L)=2$. This is the case, for example, in the mass mechanism and in the $W_R-N-W_R$ diagram of LR-symmetric models. The other possibility is to have LNV vertices. For the cases with a doubly charged bilepton $S'(+2, {\bf 1})$, for example, one can have one $\Delta(L)=2$ vertex. And, finally, it is possible to have models with two $\Delta(L)=1$ vertices. An example is trilinear RPV SUSY, e.g., from the superpotential $\mathcal{W} = \lambda' {\widehat L}{\widehat Q}{\widehat D}^c$. Most of the new models we find are of the short-range type and thus should be testable at the LHC. We will discuss one particular example in greater detail in Sec. \[sec:example\]. Decomposition of topology II ---------------------------- ---- --------------------------------------------- ----------------------------- ----------------------------- ----------------------------- ----------------------------------------------------- \# Decomposition $S$ or $V_{\rho}$ $S'$ or $V'_{\rho}$ $S'' $ or $V''_{\rho}$ Models/Refs./Comments 1 $(\bar{u} d) (\bar{u} d) (\bar{e} \bar{e})$ $ (+1,{\bf 1})$ $ (+1,{\bf 1})$ $ (-2, {\bf 1})$ Addl. triplet scalar [@Mohapatra:1981pm] LR-symmetric models [@Rizzo:1982kn; @Hirsch:1996qw] $ (+1,{\bf 8})$ $ (+1,{\bf 8})$ $ (-2, {\bf 1})$ 2 $(\bar{u} d) (\bar{u} \bar{e}) (\bar{e} d)$ $(+1,{\bf 1})$ $(-1/3,{\bf 3})$ $(-2/3,\overline{\bf 3})$ $(+1,{\bf 8})$ $(-1/3,{\bf 3})$ $(-2/3,\overline{\bf 3})$ 3 $(\bar{u} \bar{u}) (dd) (\bar{e} \bar{e})$ $(+4/3,\overline{\bf 3})$ $(+2/3,{\bf 3})$ $(-2,{\bf 1})$ only with $V_{\rho}$ and $V'_{\rho}$ $(+4/3,{\bf 6})$ $(+2/3,\overline{\bf 6})$ $(-2,{\bf 1})$ 4 $(\bar{u} \bar{u}) (\bar{e} d) (\bar{e} d)$ $ (+4/3, \overline{\bf 3})$ $ (-2/3, \overline{\bf 3})$ $ (-2/3, \overline{\bf 3})$ only with $V_{\rho}$ $ (+4/3, {\bf 6})$ $ (-2/3, \overline{\bf 3})$ $ (-2/3, \overline{\bf 3})$ 5 $(\bar{u} \bar{e}) (\bar{u} \bar{e}) (d d)$ $ (-1/3,{\bf 3})$ $ (-1/3,{\bf 3})$ $ (+2/3, {\bf 3})$ only with $V''_{\rho}$ $ (-1/3,{\bf 3})$ $ (-1/3,{\bf 3})$ $ (+2/3, \overline{\bf 6})$ [@Gu:2011ak; @Kohda:2012sr] ---- --------------------------------------------- ----------------------------- ----------------------------- ----------------------------- ----------------------------------------------------- : \[Tab:TopoII\] Decomposition of topology II. As in Tab \[Tab:TopoI\], we do only give electric and colour charges of the internal bosons here. The mediators can be either scalars ($S$, $S'$, $S''$) or vectors ($V_\rho$, $V_\rho'$, $V_\rho''$). All listed possibilities give short-range contributions. There are a total of five (again times two due to colour) possibilities to assign the outer fermions to T-II. These are listed together with the electric and colour charges of the possible mediators in Tab. \[Tab:TopoII\]. As before, this table is valid for both V and S intermediate states and we do not specify the chiralities of the outer fermions here. In the appendix, we give a table for the operator decompositions with fixed chiralities. Much fewer models with T-II have been studied in the literature than for T-I. In fact, we have found only the cases T-II-1 and T-II-5 (with a scalar ${\bf\bar 6}$) have been considered previously. The best known is the case T-II-1 with bosons transforming as $V(+1,{\bf 1})$. This diagram can be produced by adding an $SU(2)_L$ triplet scalar to the SM particle content [@Mohapatra:1981pm] (case SSS).[^12] The same diagram can occur in LR-symmetric models [@Rizzo:1982kn] (as VVS). In this case, the outer fermions are again all right-handed, and this possibility has been considered in [@Hirsch:1996qw]. Note that in all these models, there is always an additional contribution from T-I, which is not necessarily present for all T-II models. In the case of the LR-symmetric model, T-II can be comparable with T-1, but can never completely dominate the contributions to [$0\nu\beta\beta$ ]{}decay. Finally there is the recent paper [@Gu:2011ak], in which the author constructs a model with the particle content corresponding to the scalars in T-II-5 (with ${\bf\bar 6}$ representation under $SU(3)_{c}$). In this model neutrinos are pseudo-Dirac, such that T-II-5 gives the dominant contribution to [$0\nu\beta\beta$ ]{}decay. This is the only example in the literature, which we are aware of that a T-II contribution dominates the [$0\nu\beta\beta$ ]{}decay amplitude.[^13] Note, however, that our Tab. \[Tab:TopoII\] allows to construct a number of additional models with this property. An example for short-range $\boldsymbol{0\nu\beta\beta}$ decay, and its test at the LHC {#sec:example} ======================================================================================= As mentioned above, one expects that exotic models with LNV, which lead to short-range contributions for [$0\nu\beta\beta$ ]{}decay, yield testable phenomenology at the LHC. The classical example is the LR-symmetric model, and there is also a recent paper that has made a study for trilinear RPV SUSY and [$0\nu\beta\beta$ ]{}decay [@Allanach:2009iv]. In this section we will discuss basic LHC phenomenology of one particular example in our Tab. \[Tab:TopoI\], based on the decomposition T-I-4-ii-a. We have chosen this particular case basically for two reasons: (i) it belongs to the class of models, which have not been studied in the literature before, and (ii) it leads to richer phenomenology at the LHC than either the LR-symmetric model or RPV SUSY. By using the different signals which we will discuss in the following, one could distinguish this model from the other possibilities, such as the LR-symmetric model and RPV SUSY. For the decomposition T-I-4-ii-a one needs to introduce three new particles to the SM particle content, which are (a) a diquark, which is either a vector $V_{{\rm DQ}}^{4/3}$ or a scalar $S_{\rm DQ}^{4/3}$, (b) an exotic coloured (${\bf 3}$) vector-like fermion $\Psi^{5/3}$ with electric charge ${5/3}$, and (c) a leptoquark, which is again either a vector $V_{{\rm LQ}}^{2/3}$ or a scalar $S_{\rm LQ}^{2/3}$. We will concentrate on the case where both the diquark and the leptoquark are scalars. The vector case is qualitatively similar from the point of view of LHC phenomenology.[^14] The [$0\nu\beta\beta$ ]{}decay is generated through the diagram shown in Fig. \[Fig:0nbbExample\]. (250,130) (10,10)(60,60) (10,120)(60,60) (60,60)(100,60)[5]{} (100,60)(150,60) (150,60)(190,60)[5]{} (190,60)(240,120) (190,60)(240,10) (100,60)(100,120) (150,60)(150,120) (5,10)\[r\][$\overline{u_{R}}$]{} (5,120)\[r\][$\overline{u_{R}}$]{} (100,125)\[b\][$d_{L}$]{} (150,125)\[b\][$\overline{e_{R}}$]{} (245,10)\[l\][$d_{R}$]{} (245,120)\[l\][$\overline{e_{L}}$]{} (80,55)\[t\][$S_{\rm DQ}^{4/3}$]{} (125,55)\[t\][$\Psi^{5/3}$]{} (170,55)\[t\][$S^{2/3}_{\rm LQ}$]{} So far, the chiralities of the outer fermions are not determined. When taking into account the fact that the effective operator should be a component of a SM gauge invariant operator, the number of the choices is limited. There are three possibilities of the SM gauge invariant operators that can contain the T-I-4-ii-a operator,[^15] which are $(\overline{Q} \overline{Q}) (d_{R}) (\overline{L}) (\overline{L} d_{R})$, $(\overline{u_{R}} \overline{u_{R}}) (Q) (\overline{e_{R}}) (\overline{L} d_{R})$, and $(\overline{u_{R}} \overline{u_{R}}) (d_{R}) (\overline{L}) (\overline{e_{R}} Q)$. We note in passing that they correspond to the effective operators (\#11, \#20, and also \#20 respectively) shown in [@Babu:2001ex]. Here, we take the second one for our example, and consequently the chiralities of the outer fermions are fixed as $(\overline{u_{R}} \overline{u_{R}}) (d_{L}) (\overline{e_{R}}) (\overline{e_{L}} d_{R})$. Note that in this example, neutrino mass can be generated at the 3-loop order, with one effective vertex given by the [$0\nu\beta\beta$ ]{} operator. This contribution should be negligible by similar arguments as in Ref. [@Duerr:2011zd]. Diquarks at the LHC have been studied recently in [@Han:2010rf]. Among them, the relevant one for our [$0\nu\beta\beta$ ]{}example is a scalar colour-sextet ([6]{}) diquark $S_{\rm DQ}^{+4/3} ({\bf 6}, {\bf 1})_{+4/3}$ which is an $SU(2)_{L}$ singlet $(\bf 1)$ and has +4/3 $U(1)_{Y}$-hypercharge.[^16] It interacts with two right-handed up-quarks as $$\label{eq:DQ} \mathcal{L}_{\rm{DQ}} = \left[ \lambda_{\rm DQ}^{\alpha \beta} (\overline{u_{\alpha R}})^{I a} (T_{\bf \bar{6}})_{IJ}^{X} ({u_{\beta R}}^{c})^{J}_{a} (S_{\rm DQ}^{+4/3})_{X} + h.c. \right] - m_{\rm DQ}^2 (S_{\rm DQ}^{-4/3})^{X} (S_{\rm DQ}^{+4/3})_{X}.$$ where the indices $\alpha$ and $\beta$ indicate the generations of the up-type quarks, $I$ and $J$ label the fundamental representations (${\bf 3}$ and ${\bf \bar{3}}$) of $SU(3)_{c}$ ($I,J=1,2,3$), and $a$ is the index for the 2-component left-handed (=conjugate of right-handed) spinor. Here, the matrices $(T_{\bf \bar{6}})_{IJ}^{X}$ ($X=$1-6) provides $SU(3)_c$ Clebsch-Gordan coefficients, which are symmetric under the exchange of $I$ and $J$. The concrete form of $T_{\bf \bar{6}}$ is given in App. \[app:fcol\]. If the diquark is chosen as a colour ${\bf \bar{3}}$ (i.e., $\epsilon_{IJK} (\overline{u_{\alpha R}})^{I} ({u_{\beta R}}^{c})^{J} (S_{\rm DQ}^{+4/3})^{K}$) instead of a [6]{}, the coefficient $\lambda_{\rm DQ}^{\alpha\beta}$ must be antisymmetric in the indices, due to the transformation properties of the ${\bf 3}$ that is made from an antisymmetric combination in colour indices of two ${\bf \bar{3}}$ (two $\bar{u}$). For [$0\nu\beta\beta$ ]{}decay, only the choice $\alpha = \beta = u$ is relevant. Thus, the scalar diquark $S_{\rm DQ}$ must be a colour ${\bf 6}$ representation.[^17] A diquark with couplings as in Eq. (\[eq:DQ\]) will be copiously produced at the LHC. In [@Han:2010rf], the authors evaluated the production cross sections $\sigma$, which are as large as $\sigma/(\lambda_{\rm DQ}^2 \rm{BR}_{jj}) = 400$ $(1)$ pb for $m_{\rm DQ}= 1$ (3) TeV. Here, $\rm{BR}_{jj}$ is the branching ratio for the diquark decaying to two jets. Due to the s-channel resonance in the cross section, $\sigma$ scales approximately as $\sigma \propto \lambda_{\rm DQ}^2$ (in the narrow width approximation). Recently, the CMS [@Chatrchyan:2011ns] and ATLAS [@Aad:2012xz; @Aad:2012yz] collaborations have searched for resonances in the dijet mass spectrum and upper limits on $\sigma \times \rm{BR}_{jj}\times {\cal A}$ have been derived as a function of the invariant dijet mass. Here, ${\cal A}$ is the acceptance, which is estimated to be ${\cal A} \simeq 0.6$ for isotropic decays [@Chatrchyan:2011ns]. The experimental upper limits range from $\sigma\times \rm{BR}_{jj}\times {\cal A} \simeq 1$ ($0.01$) pb for $m_{\rm DQ} = 1$ (3) TeV. These limits get stronger for larger values of $m_{\rm DQ}$, because of the larger QCD background for smaller invariant masses. These limits, together with the theoretical calculation of the cross sections [@Han:2010rf], imply upper limits on $\lambda_{\rm DQ}^{uu}$ of the order of roughly $\lambda_{\rm DQ}^{uu} \lsim 0.2$ over the whole mass range explored ($m_{\rm DQ} \sim (1-4)$ TeV). Note that a scalar diquark has been proposed [@Dorsner:2009mq; @Dorsner:2011ai; @Kosnik:2011jr] as a possible explanation for the unexpectedly larger $t{\bar t}$ asymmetry observed at the Tevatron. However, these papers consider only a scalar ${\bf 3}_{a}$, which will not contribute to [$0\nu\beta\beta$ ]{}decay, as explained above. A ${\bf {\bar 6}}_{s}$ would probably be able to give a similar enhancement, but a recent paper by the ATLAS collaboration claims that most of the parameter space of [@Dorsner:2011ai] is now ruled out by LHC data [@Aad:2012em]. We will therefore not enter into a detailed discussion of this possibility. The mediator $\Psi^{5/3}$ is a heavy vector-like coloured (${\bf 3}$) fermion, aka Vector-like Quark (VLQ). The LHC phenomenology of such states has been recently studied by a number of authors [@Cacciapaglia:2010vn; @Cacciapaglia:2011fx; @Okada:2012gy; @Cacciapaglia:2012dd]. From the SM gauge invariance, this exotic fermion should be a component field of an $SU(2)_{L}$ doublet $\Psi=(\Psi^{5/3},\Psi^{2/3})^{T}$ with hypercharge $7/6$. Current limits from pair production have been summarised recently in [@Okada:2012gy]. For the $\Psi^{5/3}$ the ATLAS search for pair-produced heavy quarks decaying to $WqWq$ gives $m_{\Psi^{5/3}} \gsim 350$ GeV [@Okada:2012gy]. Note that vector-like quarks have received a lot of attention recently [@Azatov:2012rj; @Bonne:2012im; @Batell:2012ca; @Bertuzzo:2012bt; @McKeen:2012av; @Bae:2012ir; @Davoudiasl:2012ig; @Batell:2012mj; @An:2012vp; @Martin:2012dg; @Wang:2012gm; @Iwamoto:2012hh; @Endo:2011xq] as a possibility to explain the larger than expected event rate in $h\to \gamma\gamma$ observed by the ATLAS [@Aad:2012gk] and CMS [@CMS:2012gu] collaborations. The other scalar mediator $S^{2/3}_{\rm LQ}$, which interacts with $d_{R}$ and $L$, can be identified as so-called a first generation Leptoquark (LQ) which interacts only with the first generation fermions and it comes from the $SU(2)_{L}$ doublet with hypercharge, $1/6$, $S_{\rm LQ} = (S_{\rm LQ}^{2/3}, S_{\rm LQ}^{-1/3})^{T}$ [@Buchmuller:1986zs]. The Lagrangian relevant for generating the [$0\nu\beta\beta$ ]{}decay diagram of Fig. \[Fig:0nbbExample\] contains $$\mathcal{L}_{\rm LQ} = \left[ \lambda_{\rm LQ} (\overline{L})^{i}_{\dot{a}} (d_{R})_{I}^{\dot{a}} ({\rm i} \tau^{2})_{ij} (S_{\rm LQ}^{})^{I j} + \rm{h.c.} \right] - m_{\rm LQ}^{2} (S_{\rm LQ}^{\dagger})^{I i} (S_{\rm LQ})_{I i}\,, \label{eq:LQ-int}$$ where $({\rm i}\tau^{2})$ is an antisymmetric tensor for the $SU(2)_{L}$ indices. At the LHC, the first generation LQs are studied through pair production via the strong interaction. The produced LQs can then decay into $eq$ or $\nu q$ pairs through the interaction shown in Eq. . This allows to derive absolute bounds on the LQ mass (nearly independent of $\lambda_{\rm LQ}$). The current bounds [@Aad:2011ch] from ATLAS are $m_{\rm LQ}>660\,\rm{GeV}$. The CMS searches for the LQ give the limits which range from 830 GeV to 640 GeV for first generation LQs with Br$_{eq}=1$ to Br$_{eq}=0.5$. The HERA experiment [@Abramowicz:2012tg] has also searched for LQs, but via single LQ production. This leads to limits in the parameter plane $\lambda_{\rm LQ}$-$m_{\rm LQ}$. However, practically all the HERA-excluded combinations are now superseded by the recent LHC limits. For the diagram generating [$0\nu\beta\beta$ ]{}shown in Fig. \[Fig:0nbbExample\], two more interaction terms among the different mediators are needed: $$\begin{aligned} \label{eq:exaLNV} \mathcal{L}_{\Psi} =& \lambda_{{\rm DQ}\Psi}^{\alpha} (\overline{{Q_{\alpha}}^{c}})_{I i}^{a} (T_{\bf 6})^{IJ}_{X} ({\rm i} \tau^{2})^{ij} (\Psi_{L})_{J j a} (S_{\rm DQ}^{-4/3})^{X} + \lambda_{{\rm LQ}\Psi} (\overline{\Psi_{R}})^{I i a} ({e_{R}}^{c})_{a} (S_{\rm LQ})_{Ii} + h.c. $$ Note that Eq. (\[eq:exaLNV\]), together with the previously specified pieces of Lagrangians, necessarily violates lepton (but not baryon) number, as is necessary for the generation of a finite [$0\nu\beta\beta$ ]{}decay amplitude. No constraints on $\lambda_{{\rm DQ}\Psi}$ and $\lambda_{{\rm LQ}\Psi}$ exist in the literature up to now. After integrating out all the heavy fields, the LNV $d=9$ effective Lagrangian for [$0\nu\beta\beta$ ]{}decay is given as $$\label{eq:Lageff} {\cal L}^{\rm eff} = \frac{ \lambda_{\rm DQ} \lambda_{{\rm DQ}\Psi} \lambda_{{\rm LQ}\Psi} \lambda_{\rm LQ} } { m_{\rm DQ}^{2} m_{\Psi} m_{\rm LQ}^{2} } \left[ (\overline{u_{R}})^{I' a} (T_{\bf \bar{6}})_{I'J'}^{X} ({u_{R}}^{c})^{J'}_{a} \right] \left[ (\overline{{d_{L}}^{c}})_{I}^{b} (T_{\bf 6})^{IJ}_{X} ({e_{R}}^{c})_{b} \right] \left[ (\overline{e_{L}})_{\dot{c}} (d_{R})_{J}^{\dot{c}} \right] + h.c.$$ As demonstrated in App. \[app:fcol\], we arrive at a linear combination of the basis operators — the effective current description of [@Pas:2000vn] — after Fierz and the colour-index transformation, which is, $$\label{eq:Fierzexa} \mathcal{L}^{\text{eff}} = \frac{ \lambda_{\rm DQ} \lambda_{{\rm DQ}\Psi} \lambda_{{\rm LQ}\Psi} \lambda_{\rm LQ} } { m_{\rm DQ}^{2} m_{\Psi} m_{\rm LQ}^{2} } \frac{1}{32} \left[ {\rm i} (\mathcal{O}_{4})_{LR} - (\mathcal{O}_{5})_{LR} \right] \equiv \mathcal{C}_4 \, (\mathcal{O}_{4})_{LR} - \mathcal{C}_5 \, (\mathcal{O}_{5})_{LR} \,$$ with correspondingly defined coefficients $\mathcal{C}_i$ ($\propto \Lambda^{-5}$) and $\|\mathcal{C}_4\| = \|\mathcal{C}_5\|$ in this particular model. The basis operators are defined with the chirality indices $L$ and $R$ as described in App. \[app:decay\], cf., Eqs. (\[equ:o4\]) and (\[equ:o5\]). Note that the transition from Eq. (\[eq:Lageff\]) to the basis in Eq. (\[eq:Fierzexa\]) can be directly read off from our tables in App. \[app:topI\]. This example corresponds to the second line of T1-4-ii-2 in Tab. \[Tab:Decom-345\], where Eq. (\[eq:Fierzexa\]) can be read off from the last column. Therefore, this example serves to illustrate how to use the tables in our appendix. The general formula to calculate the half-life time is shown in [@Pas:2000vn], cf., Eq. (\[eq:Tinv\]) in App. \[app:decay\], and is given with the normalised (mass dimensionless) coefficients $\epsilon_{i} \equiv 2m_{P} \mathcal{C}_{i}/G_{F}^{2}$. The relevant part is $$\begin{aligned} (T_{1/2}^{\text{{$0\nu\beta\beta$ }}})^{-1} = G_{2} \left\| \sum_{i=4}^{5} \epsilon_{i} \mathcal{M}_{i} \right\|^{2} \label{eq:T-half-in-example}\end{aligned}$$ in this example. Here $G_{2}$ \[yr$^{-1}$\] is a phase space factor, and $\mathcal{M}_{i}$ are the nuclear matrix element parts of the total amplitude, which are normalised to be mass dimensionless. The half-life Eq. is dominated by the $\epsilon_{4}$ contribution, because of $\mathcal{M}_{4} \gg \mathcal{M}_{5}$.[^18] Using the experimental bounds to $\epsilon_{4}$ listed in Tab. \[Tab:LimitsShort\], we obtain the bounds for the masses of the heavy particles as a function of the couplings involved: $$\|\mathcal{C}_4 \| = \frac{ \left\| \lambda_{\rm DQ} \lambda_{{\rm DQ}\Psi} \lambda_{{\rm LQ}\Psi} \lambda_{\rm LQ} \right\| } { m_{\rm DQ}^{2} m_{\Psi} m_{\rm LQ}^{2} } \frac{1}{32} = \frac{G_F^2}{2 m_{P}} \epsilon_{4} <\frac{G_F^2}{2 m_{P}} 1.2 \cdot 10^{-8} \,$$ which, assuming that all masses are of order $\Lambda$, leads to $$\Lambda \gtrsim 2.0 \lambda_{\text{eff}}^{\frac{4}{5}} \, \text{ TeV} ,$$ where $\lambda_{\text{eff}}\equiv(\lambda_{\rm DQ}\lambda_{{\rm DQ} \Psi} \lambda_{{\rm LQ} \Psi}\lambda_{\rm LQ})^{(1/4)}$. While the individual masses involved in these expressions have been constrained from the searches at colliders, which were discussed above, the most direct test of this model — a signal directly related to the diagram shown in Fig. \[Fig:0nbbExample\] — can be done at the LHC in the following way. In the case where $m_{\rm LQ} < m_{\Psi} < m_{\rm DQ}$, once a diquark $S^{4/3}_{\rm DQ}$ is produced, it will decay to the VLQ $\Psi^{5/3}$ with a branching ratio of roughly (neglecting kinematical factors) ${\rm Br}\sim \frac{\Gamma (S_{\rm DQ} \rightarrow \Psi Q)}{ \Gamma (S_{\rm DQ} \rightarrow \Psi Q) + \Gamma (S_{\rm DQ} \rightarrow u_{R} u_{R})} \sim \frac{\lambda_{\rm DQ\Psi}^{2}} {\lambda_{\rm DQ\Psi}^2 + \lambda_{\rm DQ}^2}$, i.e., ${\rm Br} \sim 1/2$ for $\lambda_{{\rm DQ} \Psi}=\lambda_{\rm DQ}$, and the VLQ $\Psi$ will then further decay to the LQ $S^{2/3}_{\rm LQ}$ plus a lepton. This decay channel will usually dominate over the 3-body decay $\Psi^{5/3} \to 3 j$ via an off-shell diquark. The total signal for this decay chain is then $e^+e^+jj$.[^19] This signal is the same as the process searched for by the ATLAS [@ATLAS:2012ak] and the CMS collaborations [@CMS:PAS-EXO-12-017] in the context of the LR-symmetric model, and thus the result of the search can already be used to derive limits on the parameter space of the mediator fields, DQ, VLQ, and LQ. The upper limit on this channel reported in [@CMS:PAS-EXO-12-017] is around 2.5 fb with an assumption of $m_{W_{R}}=3$ TeV in the LR symmetric model. This bound corresponds then to roughly $\lambda_{{\rm DQ}\Psi} = \lambda_{\rm DQ} \lsim 0.07$ for $m_{\rm DQ}=3$ TeV (with an assumption on the mass difference $m_{\rm DQ}-m_{\Psi} \gsim 100$ GeV due to the experimental cuts). While this limit provides already some interesting constraints on this example “model” shown in Fig. \[Fig:0nbbExample\], a more detailed analysis is required, before it is ruled out as the dominant mechanism of [$0\nu\beta\beta$ ]{}decay process. We expect, of course, that much more stringent limits will be provided by the forth-coming LHC run with $\sqrt{s}=14$ TeV. Summary and conclusions ======================= We have systematically decomposed the [$0\nu\beta\beta$ ]{}operator with mass dimension nine ($d=9$), resulting in a tree-level complete list of possible contributions to [$0\nu\beta\beta$ ]{}decay. Our main results are summarised in Tables \[Tab:TopoI\] and \[Tab:TopoII\]. Our list encompasses all previously discussed contributions to [$0\nu\beta\beta$ ]{}decay and, more interestingly, demonstrates that actually most cases have not been discussed yet. The new options typically require not only fractionally charged fermions, but also exotic bosons. The latter can be doubly-charged bileptons, diquarks, or leptoquarks. For topology II (cf., right panel of Fig. \[Fig:0nbbTopologies\]), we have also found a possibility with integer charges and scalars (or vectors) only. In fact, almost all of the topology II possibilities have not been discussed in the literature before as leading contribution to [$0\nu\beta\beta$ ]{}decay. The $d=9$ [$0\nu\beta\beta$ ]{}decay operator is genuinely suppressed by $1/\Lambda^5$ if the lightest mediator is heavier than a few GeV. On the other hand, the [$0\nu\beta\beta$ ]{}decay operator mediated by a light neutrino with right-chiral interactions (long-range contribution) is weaker suppressed, and therefore, the new physics scale $\Lambda$ could be quite high, unreachable in current collider experiments. [$0\nu\beta\beta$ ]{}decay itself does very likely give the strongest constraint on such mediators. In the short-range case, $\Lambda$ points towards the TeV scale, if a signal is detected at the next generation [$0\nu\beta\beta$ ]{}decay experiments, which means that the mediators leading to [$0\nu\beta\beta$ ]{}decay may be constrained at the LHC. We have therefore focused on the short-range case for a more detailed analysis. In order to translate the bound on the half-life $T_{1/2}^{\text{{$0\nu\beta\beta$ }}}$ into a bound for the masses and couplings of a particular model, the Lorentz structure of the effective [$0\nu\beta\beta$ ]{}decay operator is important, since each model will be sensitive to a distinct combination of Nuclear Matrix Elements (NME). For the short-range case mediated by scalars and a fermion, we have therefore expanded all models in terms of the chiralities of their effective low-energy operators. The results are given in tabular form, and can be used to directly translate new bounds or re-computed NMEs into a mass and coupling limit for a specific model. From our lists and the recipes discussed in the appendix, also the corresponding vector cases can be derived in a straightforward manner. We have also worked out one example which can be tested at the LHC in greater detail. This example requires a diquark, an exotic colour-triplet vector-like fermion with electric charge $5/3$, and a leptoquark. While the individual mediators can be produced and tested at the LHC, not all couplings needed for [$0\nu\beta\beta$ ]{}decay are directly accessible in all kinematically possible configurations. However, in case the exotic fermion is lighter than the diquark, the [$0\nu\beta\beta$ ]{}decay diagram can be directly tested by processes with two like-sign leptons and two jets in the final state. While our example only serves as a prototype, we expect that a more systematic study of all short-range [$0\nu\beta\beta$ ]{}decay contributions for the LHC is feasible. In conclusion, a discovery of [$0\nu\beta\beta$ ]{} guarantees physics beyond the Standard Model. Whether this new physics is due to the $d=5$ Weinberg operator which implies heavy mediators, such as heavy right-handed neutrinos in the famous (type I) seesaw mechanism, or some other mechanism, is an open question. There are many different possibilities to mediate [$0\nu\beta\beta$ ]{}decay without Majorana neutrinos, even at tree level. We have discussed that many of these options have in common that the mediators should be found at the LHC, or that LHC will provide very stringent constraints. Finally, the most interesting case may be that [$0\nu\beta\beta$ ]{}decay is discovered in conflict with neutrino mass bounds from tritium endpoint experiments [@Osipowicz:2001sq] or cosmology [@Lesgourgues:2006nd; @Hannestad:2010kz; @Wong:2011ip] which would point towards one of our exotic mechanisms. Acknowledgements {#acknowledgements .unnumbered} ================ M.H. acknowledges support from the Spanish MICINN grants FPA2011-22975, MULTIDARK CSD2009-00064 and by the Generalitat Valenciana grant Prometeo/2009/091 and the EU Network grant UNILHC PITN-GA-2009-237920. F. B. and W. W. acknowledge support from DFG grants WI 2639/3-1 and WI 2639/4-1. T.O acknowledges support from Grant-in-Aid for Scientific Research No. 24340044 by Japan Society for the Promotion of Science. This work has been also supported by the EU FP7 project “In$\nu$isibles” (Marie Curie Actions, PITN-GA-2011-289442). From effective Lagrangians to the decay rate ============================================ Our general decomposition of the [$0\nu\beta\beta$ ]{}decay operator is given in terms of the quark (and the lepton) currents. However, [$0\nu\beta\beta$ ]{}is a low-energy process in which we must treat hadronic currents — neutrons are converted into protons in a nucleus. Therefore, the derivation of the decay rate for any model leading to an effective Lagrangian of the form Eqs. and , involves a number of steps. However, since this derivation has been studied several times in the literature, we summarise only the relevant definitions necessary for making contact with the general Lorentz-invariant description of [@Pas:1999fc; @Pas:2000vn] in this appendix. Decay rate {#app:decay} ---------- Although we have already given the effective Lagrangian for short-range contributions in Eq. , here we re-define them with the chiralities: $$\begin{aligned} \mathcal{L}^{\text{eff}} =& \frac{G_{F}^{2}}{2} m_{P}^{-1} \left[ \sum_{i=1}^{3} \epsilon_{i}^{\{XY\}Z} (\mathcal{O}_{i})_{\{XY\}Z} + \sum_{i=4}^{5} \epsilon_{i}^{XY} (\mathcal{O}_{i})_{XY} \right], \label{eq:short-w-chirality}\end{aligned}$$ where the effective operators are described as $$\begin{aligned} (\mathcal{O}_{1})_{\{XY\}Z} \equiv& J_{X} J_{Y} j_{Z}, \\ (\mathcal{O}_{2})_{\{XY\}Z} \equiv& (J_{X})^{\mu \nu} (J_{Y})_{\mu \nu} j_{Z}, \\ (\mathcal{O}_{3})_{\{XY\}Z} \equiv& (J_{X})^{\mu} (J_{Y})_{\mu} j_{Z}, \\ (\mathcal{O}_{4})_{XY} \equiv& (J_{X})^{\mu \nu} (J_{Y})_{\mu} (j)_{\nu}, \label{equ:o4} \\ (\mathcal{O}_{5})_{XY} \equiv& J_{X} (J_{Y})^{\mu} (j)_{\mu}. \label{equ:o5}\end{aligned}$$ The following formula directly relates the inverse half-life with the effective Lagrangian Eq. : [^20] $$\begin{aligned} \label{eq:Tinv} \left(T^{0\nu\beta\beta}_{1/2}\right)^{-1} =& G_{1} \left\| \sum_{i=1}^{3} \epsilon_i \mathcal{M}_i \right\|^2 + G_{2} \left\| \sum_{i=4}^{5} \epsilon_i \mathcal{M}_i \right\|^2 + G_{3} {\rm Re} \left[ \left( \sum_{i=1}^{3} \epsilon_i \mathcal{M}_i \right) \left( \sum_{i=4}^{5} \epsilon_i \mathcal{M}_i \right)^{} \right],\end{aligned}$$ Eq. (\[eq:Tinv\]) contains the product of three distinct factors, $G_{i}$, $\mathcal{M}_{i}$ and $\epsilon_{i}$. Here, $G_{i\in\{1, 2, 3\}}$ are the leptonic phase space integrals, which can be calculated accurately, e.g., [@Doi:1985dx]. The nuclear Matrix Elements (NME) ${\cal M}_{i\in\{1\text{-}5\}}$ are different for different short-range contributions $\epsilon_{i\in\{1\text{-}5\}}$, detailed definitions can be found in [@Pas:2000vn], numerical values are given in [@Pas:2000vn; @Deppisch:2012nb]. The contribution from the mass mechanism can be expressed as $$\begin{aligned} \label{eq:Tinv2} \left(T^{0\nu\beta\beta}_{1/2}\right)^{-1} =& G_{1} \left\| \frac{\langle m_{\nu} \rangle}{m_{e}} \left[ \mathcal{M}_{\rm GT} - \frac{g_{V}^{2}}{g_{A}^{2}} \mathcal{M}_{\rm F} \right] \right\|^{2},\end{aligned}$$ where $\mathcal{M}_{\rm F}$ and $\mathcal{M}_{\rm GT}$ are the standard Fermi and Gamow-Teller transition matrix elements, $g_{V}$ and $g_{A}$ are the vector and axial-vector couplings of the hadron current, and $m_{e}$ is the mass of electron.[^21] Given the numerical values of $G_{i}$ and $\mathcal{M}_{i}$, once the combination of $\epsilon_i$’s for a given model have been identified (see the tables below), the derivation of the limits from $T_{1/2}^{\text{{$0\nu\beta\beta$ }}}$ becomes straightforward. Fierz and colour-index transformations {#app:fcol} -------------------------------------- The general effective Lagrangians given in Eqs. (\[eq:defLR\]) and (\[eps\_short\]) are described with [the standard form $J$ of the quark current]{}, which (i) is a singlet under the colour $SU(3)_{c}$ and (ii) takes the bi-linear form $({\bar u} \Gamma d)$ in terms of the Lorentz structure, where $\Gamma\in\{1\pm\gamma^{5}, \gamma^{\mu} (1 \pm \gamma^{5}), \gamma^{\mu \nu} (1 \pm \gamma^{5})\}$. It is suitable for the calculation of the hadron transition amplitudes of the type $\langle P(p)\|({\bar u} \Gamma d) \|N(p')\rangle$, to describe the conversion of neutrons to protons. On the other hand, the decomposed effective operators which are listed in Tabs. \[Tab:TopoI\] and \[Tab:TopoII\] do not take this standard form (except for T-I-1 and T-II-1). In order to bring them to the standard form shown as the effective Lagrangian Eq. , we need to transform the Lorentz and the colour indices in the effective operators. First, we discuss the treatment of colour indices. In the operator decomposition, we introduce the antisymmetric tensors $\epsilon_{IJK}$ and $\epsilon^{IJK}$ ($I,J,K =1\text{-}3$), the symmetric matrices $(T_{\bf 6})_{X}^{IJ}$ and $(T_{\bf \bar{6}})^{X}_{IJ}$ ($X=1\text{-}6$) under the exchange of $I$ and $J$, and the Gell-Mann matrices ${(\lambda^{A})_{I}}^{J}$ ($A=1\text{-}8$), where the lower $I$-index is for ${\bf 3}$ representation ($d_{I}$), while the upper one for ${\bf \bar{3}}$ ($\bar{u}^{I}$). Here, the matrices $T_{\bf 6}$ and $T_{\bf \bar{6}}$ are explicitly defined as $$\begin{gathered} (T_{\bf 6})_{1}^{IJ} = (T_{{\bf \bar{6}}})^{1}_{IJ} = \begin{pmatrix} 1 && \\ &0&\\ &&0 \end{pmatrix}, \qquad (T_{\bf 6})_{2}^{IJ} = (T_{{\bf \bar{6}}})^{2}_{IJ} = \begin{pmatrix} 0 &1/\sqrt{2}& \\ 1/\sqrt{2}&0&\\ &&0 \end{pmatrix}, \nonumber \\ (T_{\bf 6})_{3}^{IJ} = (T_{{\bf \bar{6}}})^{3}_{IJ} = \begin{pmatrix} 0 && \\ &1&\\ &&0 \end{pmatrix}, \qquad (T_{\bf 6})_{4}^{IJ} = (T_{{\bf \bar{6}}})^{4}_{IJ} = \begin{pmatrix} 0 && 1/\sqrt{2} \\ &0&\\ 1/\sqrt{2} &&0 \end{pmatrix}, \nonumber \\ (T_{\bf 6})_{5}^{IJ} = (T_{{\bf \bar{6}}})^{5}_{IJ} = \begin{pmatrix} 0 && \\ &0& 1/\sqrt{2}\\ &1/\sqrt{2}&0 \end{pmatrix}, \qquad (T_{\bf 6})_{6}^{IJ} = (T_{{\bf \bar{6}}})^{6}_{IJ} = \begin{pmatrix} 0 && \\ &0&\\ &&1 \end{pmatrix}.\end{gathered}$$ The transformation rules relevant to our work are summarised as $$\begin{aligned} \epsilon^{IJK} \epsilon_{KI'J'} =& \delta_{I'}^{I} \delta_{J'}^{J} - \delta_{I'}^{J} \delta_{J'}^{I}, \\ (T_{\bf 6})_{X}^{IJ} (T_{\bf \bar{6}})^{X}_{I'J'} =& \frac{1}{2} \left[ \delta_{I'}^{I} \delta_{J'}^{J} + \delta_{I'}^{J} \delta_{J'}^{I} \right], \label{eq:T6T6-dd} \\ {(\lambda^{A})_{I'}}^{I} {(\lambda^{A})_{J'}}^{J} =& - \frac{2}{3} \delta_{I'}^{I} \delta_{J'}^{J} + 2 \delta_{I'}^{J} \delta_{J'}^{I}.\end{aligned}$$ Next, we transform the spinor indices. General formulas for Fierz transformations can be found in the literature in many references, see for example [@Dreiner:2008tw] for 2-component spinor representations. In some decompositions, we must Fierz-transform all the six fermions, to which the well-known transformation rules for four fermions are not applicable. However, after appropriate successive transformations with the formulae shown in [@Dreiner:2008tw], one can reach the standard form. Here, we demonstrate the procedure of the operator projection with an operator as an example, which is we saw in Sec. \[sec:example\], $(\overline{u_{R}} \overline{u_{R}}) (d_{L})(\overline{e_{R}}) (\overline{e_{L}} d_{R})$. Writing down the operator again with all the indices explicitly, we have $$\begin{aligned} \mathcal{O}_{\text{example}} \equiv& \left[ (\overline{u_{R}})^{I' a} (T_{\bf \bar{6}})^{X}_{I'J'} ({u_{R}}^{c})^{J'}_{a} \right] \left[ (\overline{{d_{L}}^{c}})_{I}^{b} (T_{\bf 6})_{X}^{IJ} ({e_{R}}^{c})_{b} \right] \left[ (\overline{e_{L}})_{\dot{c}} (d_{R})_{J}^{\dot{c}} \right].\end{aligned}$$ Applying Eq. , we obtain the following colour-singlet $\bar{u}d$ combinations, that however do not form standard quark currents yet: $$\begin{aligned} \mathcal{O}_{\text{example}} =& \frac{1}{2} \left[ \delta_{I'}^{I} \delta_{J'}^{J} + \delta_{I'}^{J} \delta_{J'}^{I} \right] (\overline{u_{R}})^{I' a} ({u_{R}}^{c})^{J'}_{a} (\overline{{d_{L}}^{c}})_{I}^{b} ({e_{R}}^{c})_{b} (\overline{e_{L}})_{\dot{c}} (d_{R})_{J}^{\dot{c}} \nonumber \\ =& (\overline{u_{R}})^{I a} (\overline{{d_{L}}^{c}})_{I}^{b} ({u_{R}}^{c})^{J}_{a} (d_{R})_{J}^{\dot{c}} (\overline{e_{L}})_{\dot{c}} ({e_{R}}^{c})_{b}. \label{eq:Operator-Projection-example-Nr1}\end{aligned}$$ In order to obtain the standard form, we transform also the spinor indices. The formulae that are necessary for this transformation will be shown later. $$\begin{aligned} &\text{(RHS) of Eq.~\eqref{eq:Operator-Projection-example-Nr1}} \nonumber \\ =& \delta_{b}^{d} \delta_{\dot{f}}^{\dot{c}} (\overline{u_{R}})^{I a} (\overline{{d_{L}}^{c}})_{I}^{b} ({u_{R}}^{c})^{J}_{a} (d_{R})_{J}^{\dot{f}} (\overline{e_{L}})_{\dot{c}} ({e_{R}}^{c})_{d} \nonumber \\ =& - \frac{1}{32} \left[ J_{L} (J_{R})^{\mu} (j)_{\mu} + \frac{1}{{\rm i}} (J_{L})^{\mu \nu} (J_{R})_{\mu} (j)_{\nu} \right]. \label{eq:Operator-Projection-example-Nr2}\end{aligned}$$ Here, Fierz transformations are carried out in the 2-component representation (as in [@Dreiner:2008tw]), which is related to the 4-component representation in the following manner. We take so-called chiral representation for a 4-component spinor, i.e., the Lorentz vector matrices $\sigma^{\mu}\equiv (1, \sigma^{a})$ and $\overline{\sigma}^{\mu} \equiv (1, -\sigma^{a})$ for 2-component spinors are introduced, which are given as the components of the $\gamma^{\mu}$ matrices as $$\begin{aligned} \gamma^{\mu} = \begin{pmatrix} & (\sigma^{\mu})_{a\dot{b}} \\ (\overline{\sigma}^{\mu})^{\dot{a} b} & \end{pmatrix}.\end{aligned}$$ The Lorentz tensor matrices $\sigma^{\mu \nu}$ for 2-component spinors, which appear in Eq. , are defined with $\sigma^{\mu}$ and $\overline{\sigma}^{\mu}$ as[^22] $$\begin{aligned} {(\sigma^{\mu \nu})_{a}}^{b} \equiv& \frac{1}{4} \left[ (\sigma^{\mu})_{a\dot{a}} (\overline{\sigma}^{\nu})^{\dot{a} b} - (\sigma^{\nu})_{a\dot{a}} (\overline{\sigma}^{\mu})^{\dot{a} b} \right], \\ {(\overline{\sigma}^{\mu \nu})^{\dot{a}}}_{\dot{b}} \equiv& \frac{1}{4} \left[ (\overline{\sigma}^{\mu})^{\dot{a} a} (\sigma^{\nu})_{a \dot{b}} - (\overline{\sigma}^{\nu})^{\dot{a} a} (\sigma^{\mu})_{a \dot{b}} \right],\end{aligned}$$ which are related to the matrices $\gamma^{\mu \nu}$ for 4-component spinors as $$\begin{aligned} \gamma^{\mu \nu} = 2 {\rm i} \begin{pmatrix} {(\sigma^{\mu \nu})_{a}}^{b} & \\ & {(\overline{\sigma}^{\mu \nu})^{\dot{a}}}_{\dot{b}} \end{pmatrix}.\end{aligned}$$ The relations between the currents in the 2-component representation and those (defined at Eq. ) in the 4-component representation are explicitly described as $$\begin{aligned} (\overline{u_{R}})^{I a} (d_{L})_{I a} =& \frac{1}{2} J_{L}, \\ (\overline{u_{R}})^{I a} (\sigma^{\mu})_{a \dot{b}} (d_{R})_{I}^{\dot{b}} =& \frac{1}{2} (J_{R})^{\mu}, \\ (\overline{u_{R}})^{I a} {(\sigma^{\mu \nu})_{a}}^{b} (d_{L})_{I b} =& \frac{1}{4 {\rm i}} (J_{L})^{\mu \nu}, \\ (\overline{e_{L}})_{\dot{a}} (\overline{\sigma}^{\mu})^{\dot{a} b} ({e_{R}}^{c})_{b} =& \frac{1}{2} (j_{L})^{\mu}.\end{aligned}$$ In the steps of the transformation shown in Eq. , we applied the following formulae of Fierz-transformation, $$\begin{aligned} \delta_{a}^{d} \delta_{\dot{b}}^{\dot{c}} =& \frac{1}{2} (\sigma^{\mu})_{a \dot{b}} (\overline{\sigma}^{\mu})^{\dot{c} d}, \\ \epsilon_{ac} \epsilon^{bd} =& - \frac{1}{2} \left[ \delta_{a}^{b} \delta_{c}^{d} + {(\sigma^{\rho \sigma})_{a}}^{b} {(\sigma_{\rho \sigma})_{c}}^{d} \right],\end{aligned}$$ and the nature of the sigma matrices, $$\begin{aligned} {(\sigma^{\mu \nu})_{a}}^{b} (\sigma^{\rho})_{b \dot{b}} =& \frac{1}{2} \left[ (\sigma^{\mu})_{a \dot{b}} g^{\nu \rho} - (\sigma^{\nu})_{a \dot{b}} g^{\mu \rho} + {\rm i} \epsilon^{\mu \nu \rho \sigma} (\sigma_{\sigma})_{a \dot{b}} \right], \\ {\rm i} \epsilon^{\mu \nu \rho \sigma} {(\sigma_{\rho \sigma})_{a}}^{b} =& -2 {(\sigma^{\mu \nu})_{a}}^{b}.\end{aligned}$$ Full decompositions for the short-range scalar-mediated topology I {#app:topI} ------------------------------------------------------------------ In case of topology I the intermediate states contain a fermion, leading to a different treatment in case of long-range and short-range contributions. For the long-range part of the amplitude, in order to derive interesting constraints, one has to pick out the neutrino momentum $q_{\nu}$ from the propagator, because of $q_{\nu} \gg m_{\nu}$. On the other hand, in the short-range case, the relevant amplitude must take the mass part $m_{\psi}$ in the propagator of the fermion mediator $\psi$, because of $q_{\psi} \ll m_{\psi}$. Thus, not all possibilities to assign chiralities of the six outer fermions lead to interesting models. When the decomposition is symbolically written as $(a b)(c)(d)(e f)$ where each parenthesis corresponds to each vertex $\text{v}_{i}$ in the left panel of Fig. \[Fig:0nbbTopologies\], i.e., $(ab)$ corresponds to the outer fermions on the vertex $\text{v}_{1}$, and $(c)$ does to $\text{v}_{2}$, and so on, in order to pick out $m_{\psi}$ from the propagator, only the combinations $(\overline{{c_L}^{c}})(d_L)$ and $(\overline{{c_R}^{c}})(d_R)$ need to be considered. Note, that $\overline{a_R}\equiv \overline{a}P_L$. Since we concentrate on the SFS case, in which both the mediators between the vertices $\text{v}_{1}$ and $\text{v}_{2}$ and that between $\text{v}_{3}$ and $\text{v}_{4}$ are scalars, the chirality structures of the outer fermions on $\text{v}_{1}$ and $\text{v}_{4}$ are also restricted to be $(ab) \in \{(\overline{{a_{L}}^{c}}b_{L}),(\overline{{a_{R}}^{c}}b_{R})\}$ and $(ef) \in \{ (\overline{{e_{L}}^{c}}f_{L}), (\overline{{e_{R}}^{c}}f_{R})\}$. At this stage, there are eight choices for the combinations of chiralities of the six outer fermions. However, some of them are not generated from the $d=9$ SM gauge invariant operators which are listed in [@Babu:2001ex; @deGouvea:2007xp]. We exclude the chirality choices which require an additional Higgs doublet(s) (i.e., those originated from the operators of $d>9$), and list all the possibilities inspired from the $d=9$ SM gauge invariant operators. The results are summarised in Tabs. \[Tab:Decom-1-i\]-\[Tab:Decom-345\]. The id-numbers indicated in the column “\#” correspond to those in Tab. \[Tab:TopoI\]. One can find the chirality choices explicitly in the column “Operators”. The id-numbers in the column “BL” tells the correspondence to the $d=9$ SM gauge invariant operators listed by Babu and Leung [@Babu:2001ex]. The SM charges of all possible mediators are fully identified in the column “Mediators”.[^23] The basis operators, cf. Eq. (\[eps\_short\]), resulting after Fierz and the colour-index transformations are given in the column “Basis op.”. Some of the decompositions appear necessarily with different ones at the same time. These associated decompositions are listed in the column “Appears with”. The use and application of our tables is illustrated with the example in Sec. \[sec:example\], see discussion after Eq. (\[eq:Lageff\]). Once a specific model is chosen, the basis operators with the corresponding coefficients can be directly read off from the table. These basis operators are directly related with the NMEs, i.e., the table lists which NMEs specific models are testing. As a consequence, the lifetime bounds can be translated into constraints on masses and couplings for a specific model — as it is illustrated in our example. -------- ----------------------------------------------- ------ ---------------------------- -------------------------------------- ---------------------------- ------------------------------------------------------------------------------------------------ \# Operators BL $S$ $\psi$ $S'$ Basis op. 1-i $(\overline{u_{L}} d_{R}) (\overline{e_{L}}) \#11 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 1})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{RR \} R}$ (\overline{e_{L}}) (\overline{u_{L}} d_{R})$ $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 3})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ s.a.a $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 1})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ $ - \frac{5}{24} (\mathcal{O}_{1})_{\{RR\}R}$ $ - \frac{1}{32} (\mathcal{O}_{2})_{\{RR\}R}$ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 3})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ s.a.a $(\overline{u_{L}} d_{R}) (\overline{e_{L}}) \#14 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 1})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LR \} R}$ (\overline{e_{L}}) (\overline{u_{R}} d_{L})$ $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 3})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ s.a.a $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 1})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ $ -\frac{1}{12} (\mathcal{O}_{1})_{\{LR\}R} $ $ - \frac{1}{8} (\mathcal{O}_{3})_{\{LR\}R} $ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 3})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ s.a.a $(\overline{u_{R}} d_{L}) (\overline{e_{L}}) \#12 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 1})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LL \} R}$ (\overline{e_{L}}) (\overline{u_{R}} d_{L})$ $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 1}, {\bf 3})_{0}$ $({\bf 1},{\bf 2})_{-1/2}$ s.a.a $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 1})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ $ - \frac{5}{24} (\mathcal{O}_{1})_{\{LL\}R} $ $ - \frac{1}{32} (\mathcal{O}_{2})_{\{LL\} R} $ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 8}, {\bf 3})_{0}$ $({\bf 8},{\bf 2})_{-1/2}$ s.a.a 1-ii-a $(\overline{u_{L}} d_{R}) (\overline{u_{L}}) \#11 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 3})_{+2/3}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{RR \} R}$ (d_{R}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 3})_{+2/3}$ $({\bf 1},{\bf 3})_{+1}$ $-\frac{5}{24} (\mathcal{O}_{1})_{\{RR\} R} $ $-\frac{1}{32} (\mathcal{O}_{2})_{\{RR\} R} $ $(\overline{u_{L}} d_{R}) (\overline{u_{R}}) \#14 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 2})_{+7/6}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LR \} R}$ (d_{L}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 2})_{+7/6}$ $({\bf 1},{\bf 3})_{+1}$ $ - \frac{1}{12} (\mathcal{O}_{1})_{\{LR\}R}$ $ - \frac{1}{8} (\mathcal{O}_{3})_{\{LR\} R}$ $(\overline{u_{R}} d_{L}) (\overline{u_{L}}) \#14 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 3})_{+2/3}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LR \} R}$ (d_{R}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 3})_{+2/3}$ $({\bf 1},{\bf 3})_{+1}$ $-\frac{1}{12} (\mathcal{O}_{1})_{\{LR\}R}$ $- \frac{1}{8} (\mathcal{O}_{3})_{\{LR\}R}$ $(\overline{u_{R}} d_{L}) (\overline{u_{R}}) \#12 $({\bf 1},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 2})_{+7/6}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LL \} R}$ (d_{L}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $({\bf 3}, {\bf 2})_{+7/6}$ $({\bf 1},{\bf 3})_{+1}$ $-\frac{5}{24} (\mathcal{O}_{1})_{\{LL\}R} $ $- \frac{1}{32} (\mathcal{O}_{2})_{\{LL\}R}$ 1-ii-b $(\overline{u_{L}} d_{R}) \#14 $({\bf 1},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LR \} R}$ (d_{L}) (\overline{u_{R}}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ $({\bf 1},{\bf 3})_{+1}$ $ - \frac{1}{12} (\mathcal{O}_{1})_{\{LR\}R}$ $ - \frac{1}{8} (\mathcal{O}_{3})_{\{LR\}R}$ $(\overline{u_{L}} d_{R}) \#11 $({\bf 1},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 2})_{+5/6}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{RR \} R}$ (d_{R}) (\overline{u_{L}}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 2})_{+5/6}$ $({\bf 1},{\bf 3})_{+1}$ $ - \frac{5}{24} (\mathcal{O}_{1})_{\{RR\}R} $ $ - \frac{1}{32} (\mathcal{O}_{2})_{\{RR\}R} $ $(\overline{u_{R}} d_{L}) \#12 $({\bf 1},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LL \} R}$ (d_{L}) (\overline{u_{R}}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ $({\bf 1},{\bf 3})_{+1}$ $- \frac{5}{24} (\mathcal{O}_{1})_{\{LL\}R} $ $- \frac{1}{32} (\mathcal{O}_{2})_{\{LL\}R} $ $(\overline{u_{R}} d_{L}) \#14 $({\bf 1},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 2})_{+5/6}$ $({\bf 1},{\bf 3})_{+1}$ $\frac{1}{8} (\mathcal{O}_{1})_{\{LR \} R}$ (d_{R}) (\overline{u_{L}}) (\overline{e_{L}} \overline{e_{L}})$ $({\bf 8},{\bf 2})_{+1/2}$ $(\overline{\bf 3}, {\bf 2})_{+5/6}$ $({\bf 1},{\bf 3})_{+1}$ $-\frac{1}{12} (\mathcal{O}_{1})_{\{LR\}R}$ $- \frac{1}{8} (\mathcal{O}_{3})_{\{LR\}R}$ -------- ----------------------------------------------- ------ ---------------------------- -------------------------------------- ---------------------------- ------------------------------------------------------------------------------------------------ : The results of decomposition and projection of the operators categorised to \#1. We also show the ID-number of the lepton number violating operators listed by Babu and Leung (BL) [@Babu:2001ex] (see also Ref. [@deGouvea:2007xp]). For T-I-i-1, there are only three independent choices of chiralities. Here, we assume that the effective operators are originated from the SM gauge ($SU(3)_{c} \times SU(2)_{L} \times U(1)_{Y}$) invariant $d=9$ operators. Note that the scalar mediators $S$ and $S'$ of \#1-i-(2) take interactions with different combinations of quarks, although their SM charges are the same. The abbreviation “s.a.a” in “Basis op.” column means “same as above”. The hypercharge $Y$ is defined as $Y \equiv Q_{\rm em} - I_3$. []{data-label="Tab:Decom-1-i"} --------- --------------------------------------- ------ -------------------------------------- -------------------------------------- -------------------------------------- ------------------------------------------------------------------------------------------------- ----------- Appears \# Operators BL $S$ $\psi$ $S'$ Basis op. with 2-iii-a $(d_{R} \overline{e_{L}}) \#11 $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 1})_{0}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $ \frac{1}{32} (\mathcal{O}_{1})_{\{ RR \} R}$ $+ \frac{1}{128} (\mathcal{O}_{2})_{\{ RR \}R} $ 4-i & 5-i (\overline{u_{L}}) (d_{R}) (\overline{u_{L}} \overline{e_{L}})$ $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 3})_{0}$ $(\overline{\bf 3},{\bf 3})_{+1/3}$ s.a.a 4-i & 5-i $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 1})_{0}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $- \frac{7}{48} (\mathcal{O}_{1})_{\{RR\}R} $ $ - \frac{1}{192} (\mathcal{O}_{2})_{\{RR\}R}$ 4-i & 5-i $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 3})_{0}$ $(\overline{\bf 3},{\bf 3})_{+1/3}$ s.a.a 4-i & 5-i $(d_{R} \overline{e_{L}}) \#19 $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 1})_{0}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $- \frac{1}{32 {\rm i}} (\mathcal{O}_{4})_{RR}$ $- \frac{1}{32} (\mathcal{O}_{5})_{RR} $ 4-i & 5-i (\overline{u_{L}}) (d_{R}) (\overline{u_{R}} \overline{e_{R}})$ $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 1})_{0}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $ + \frac{1}{48{\rm i}} (\mathcal{O}_{4})_{RR}$ $ +\frac{7}{48} (\mathcal{O}_{5})_{RR} $ 4-i & 5-i $(d_{R} \overline{e_{L}}) \#14 $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $\frac{1}{32} (\mathcal{O}_{3})_{\{ LR \} R} $ (\overline{u_{R}}) (d_{L}) (\overline{u_{L}} \overline{e_{L}})$ $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 3})_{+1/3}$ s.a.a $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $ - \frac{1}{8} (\mathcal{O}_{1})_{\{LR\}R} $ $ - \frac{1}{48}(\mathcal{O}_{3})_{\{LR\}R} $ $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 3})_{+1/3}$ s.a.a $(d_{R} \overline{e_{L}}) \#20 $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 1}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $ \frac{1}{32 {\rm i}} (\mathcal{O}_{4})_{LR} $ $ - \frac{1}{32} (\mathcal{O}_{5})_{LR} $ (\overline{u_{R}}) (d_{L}) (\overline{u_{R}} \overline{e_{R}})$ $(\overline{\bf 3},{\bf 2})_{-1/6}$ $({\bf 8}, {\bf 2})_{+1/2}$ $(\overline{\bf 3},{\bf 1})_{+1/3}$ $ - \frac{1}{48{\rm i}} (\mathcal{O}_{4})_{LR} $ $ + \frac{7}{48} (\mathcal{O}_{5})_{LR} $ 2-iii-b $(d_{R} \overline{e_{L}}) \#14 $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ \frac{1}{16} (d_{L}) (\overline{u_{R}}) (\mathcal{O}_{1})_{\{LR\}R} (\overline{u_{L}} \overline{e_{L}})$ $ $ + \frac{1}{32} (\mathcal{O}_{3})_{\{LR\}R} $ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 3})_{-1/3}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ s.a.a $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ - \frac{1}{32} (\mathcal{O}_{1})_{\{LR\}R} $ $+ \frac{1}{64} (\mathcal{O}_{3})_{\{LR\}R} $ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 3})_{-1/3}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ s.a.a $(d_{R} \overline{e_{L}}) \#20 $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ \frac{1}{32{\rm i}} (d_{L}) (\overline{u_{R}}) (\mathcal{O}_{4})_{LR} (\overline{u_{R}} \overline{e_{R}})$ $ $ - \frac{3}{32} (\mathcal{O}_{5})_{LR} $ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ \frac{1}{64{\rm i}} (\mathcal{O}_{4})_{LR} $ $ + \frac{1}{64} (\mathcal{O}_{5})_{LR} $ $(d_{R} \overline{e_{L}}) \#11 $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ \frac{3}{32} (d_{R}) (\overline{u_{L}}) (\mathcal{O}_{1})_{\{RR\}R}$ $+ (\overline{u_{L}} \overline{e_{L}})$ \frac{1}{128} (\mathcal{O}_{2})_{\{RR\}R} $ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ s.a.a $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ -\frac{1}{64} (\mathcal{O}_{1})_{\{RR\}R}$ $+ \frac{1}{256} (\mathcal{O}_{2})_{\{RR\}R}$ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 3})_{+1/3}$ s.a.a $(d_{R} \overline{e_{L}}) \#19 $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ (d_{R}) (\overline{u_{L}}) - (\overline{u_{R}} \overline{e_{R}})$ \frac{1}{32{\rm i}} (\mathcal{O}_{4})_{RR}$ $- \frac{3}{32} (\mathcal{O}_{5})_{RR} $ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 2})_{+1/6}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $ - \frac{1}{64{\rm i}} (\mathcal{O}_{4})_{RR}$ $+ \frac{1}{64} (\mathcal{O}_{5})_{RR} $ --------- --------------------------------------- ------ -------------------------------------- -------------------------------------- -------------------------------------- ------------------------------------------------------------------------------------------------- ----------- : Decomposition \#2-iii. In case of decomposition \#2-iii-a, we will have not only \#2-iii-a, but also \#4-i and \#5-i.[]{data-label="Tab:Decom-2-iii"} Topology II ----------- ![\[fig:split\]The four different insertions for topology-II: SSS, VVV, SSV and VVS. Note, that both VVV and SSV have necessarily derivative couplings. For a discussion see text.](Topo_II_SSS.eps "fig:") ![\[fig:split\]The four different insertions for topology-II: SSS, VVV, SSV and VVS. Note, that both VVV and SSV have necessarily derivative couplings. For a discussion see text.](Topo_II_VVV.eps "fig:")\ ![\[fig:split\]The four different insertions for topology-II: SSS, VVV, SSV and VVS. Note, that both VVV and SSV have necessarily derivative couplings. For a discussion see text.](Topo_II_SSV.eps "fig:") ![\[fig:split\]The four different insertions for topology-II: SSS, VVV, SSV and VVS. Note, that both VVV and SSV have necessarily derivative couplings. For a discussion see text.](Topo_II_VVS.eps "fig:") We can decompose the [$0\nu\beta\beta$ ]{}operator with T-II in practically the same way as T-I. However, there are also some differences, which we will briefly discuss. As mentioned above, and shown in Fig. \[fig:split\], there are four possible combinations of scalars and vectors to complete the T-II decomposition. As shown in the figure, both VVV and SSV necessarily involve derivative couplings. They lead to an effective $d=10$ operator, $$\begin{aligned} \mathcal{O}_{d=10} \propto \frac{1}{\Lambda^{6}} \partial_{\rho} (\bar{u} \bar{u} dd \bar{e} \bar{e})^{\rho} = \frac{q}{\Lambda} \frac{1}{\Lambda^{5}} (\bar{u} \bar{u} dd \bar{e} \bar{e}),\end{aligned}$$ where $q$ is a typical momentum of the [$0\nu\beta\beta$ ]{}process. Therefore, they are suppressed by a factor of $q/\Lambda$ in comparison with decompositions of the $d=9$ operators without derivatives and can be safely neglected. Diagrams of the type SVV come from vectors being gauge bosons, i.e., from the covariant derivative of the scalar field $\mathcal{S}$: $$\begin{aligned} \mathcal{L}_{\rm gauge} = (D_{\mu} \mathcal{S})^{\dagger}(D^\mu \mathcal{S}) \supset g^2 \mathcal{S}^{\dagger} \mathcal{S} V_{\mu}V^{\mu} \supset g^2 \langle \mathcal{S} \rangle S V_{\mu}V^{\mu},\end{aligned}$$ if the scalar $\mathcal{S}$ can take a vacuum expectation value $\langle \mathcal{S} \rangle$. Here, $S$ is a fluctuation around the vev $\mathcal{S} = \langle \mathcal{S} \rangle + S$, which would be a scalar mediator. If this vev breaks $SU(2)_L$, this leads to a suppression order $v/\Lambda$, but, as can be seen from the example of LR symmetry, an SM singlet vev can produce a coupling whose order is of $\Lambda$, such that the total amplitude for T-II is again proportional to $\Lambda^{-5}$. Similarly, for diagrams of the type SSS, the coupling has a dimension of mass, leading potentially to a $\Lambda^{-5}$ total factor for the diagram. Since new vectors require an extension of the gauge group, we consider the SSS case to be the more easily motivated choice. In our detailed decomposition of T-II, presented in Tab. \[Tab:Top2-3S\], we therefore concentrate on the case of SSS. The results for SVV can be derived easily from the recipes discussed above. Errata for version 1 ==================== We have discovered a few misprints in tables \[Tab:Decom-2-iii\] and \[Tab:Decom-345\] of the first version of the paper. Some quantum numbers for the intermediate scalar or fermion where incorrectly given. The correct quantum numbers are given in table \[Tab:Misprints\], where the corrections are emphasized in red color. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ \# Operators BL $S$ $\psi$ $S'$ --------- -------------------------------------- ------ ------------------------------------------------ ---------------------------------------------- ---------------------------------------------------------- 2-iii-b $(d_{R} \overline{e_{L}}) \#20 $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $({\bf 3}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{{\color[rgb]{1,0,0}}+1/3}$ (d_{L}) (\overline{u_{R}}) (\overline{u_{R}} \overline{e_{R}})$ $(\overline{\bf 3}, {\bf 2})_{-1/6}$ $(\overline{\bf 6}, {\bf 1})_{-1/3}$ $(\overline{\bf 3}, {\bf 1})_{{\color[rgb]{1,0,0}}+1/3}$ 5-i $(\overline{u_{L}} \overline{e_{L}}) \#11 $({\bf 3}, {\bf{\color[rgb]{1,0,0}}3})_{-1/3}$ $({\bf 1}, {\bf 3})_{0}$ $(\overline{\bf 3}, {\bf{\color[rgb]{1,0,0}}3})_{+1/3}$ (d_{R}) (d_{R}) (\overline{u_{L}} \overline{e_{L}}) $ $({\bf 3}, {\bf{\color[rgb]{1,0,0}}3})_{-1/3}$ $({\bf 8}, {\bf 3})_{0}$ $(\overline{\bf 3}, {\bf{\color[rgb]{1,0,0}}3})_{+1/3}$ $(\overline{u_{R}} \overline{e_{R}}) \#19 $({\bf 3}, {\bf 1})_{-1/3}$ $({\bf 1}, {\bf {\color[rgb]{1,0,0}}1})_{0}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ (d_{R}) (d_{R}) (\overline{u_{L}} \overline{e_{L}})$ $({\bf 3}, {\bf 1})_{-1/3}$ $({\bf 8}, {\bf {\color[rgb]{1,0,0}}1})_{0}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ $(\overline{u_{R}} \overline{e_{R}}) — $({\bf 3}, {\bf 1})_{-1/3}$ $({\bf 1}, {\bf {\color[rgb]{1,0,0}}1})_{0}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ (d_{R}) (d_{R}) (\overline{u_{R}} \overline{e_{R}})$ $({\bf 3}, {\bf 1})_{-1/3}$ $({\bf 8}, {\bf {\color[rgb]{1,0,0}}1})_{0}$ $(\overline{\bf 3}, {\bf 1})_{+1/3}$ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ : Misprints in tables 8 and 9. The corrected quantum numbers are emphasized in red colour.[]{data-label="Tab:Misprints"} In addition, in table \[Tab:Decom-345\] two possible decompositions were not listed by accident. The missing decompositions correspond to two cases in decomposition 4-i, BL\#11, and are given in table \[Tab:Missing\]. All the corrections listed above are included in tables of the current version. [100]{} I. Avignone, Frank T., S. R. Elliott, and J. Engel, Rev. Mod. Phys. [80]{}, 481 (2008), arXiv:0708.1033. J. Gomez-Cadenas, J. Martin-Albo, M. Mezzetto, F. Monrabal, and M. Sorel, Riv. Nuovo Cim. [35]{}, 29 (2012), arXiv:1109.5515. W. Rodejohann, arXiv:1206.2560. A. Barabash, arXiv:1209.4241. W. Rodejohann, Int.J.Mod.Phys. [E20]{}, 1833 (2011), arXiv:1106.1334. F. F. Deppisch, M. Hirsch, and H. Pas, arXiv:1208.0727. J. Schechter and J. Valle, Phys. Rev. [D25]{}, 2951 (1982). J. F. Nieves, Phys. Lett. [B147]{}, 375 (1984). E. Takasugi, Phys. Lett. [B149]{}, 372 (1984). M. Hirsch, S. Kovalenko, and I. Schmidt, Phys. Lett. [B642]{}, 106 (2006), arXiv:hep-ph/0608207. M. Duerr, M. Lindner, and A. Merle, JHEP [1106]{}, 091 (2011), arXiv:1105.0901. H. Klapdor-Kleingrothaus [et al.]{}, Eur. Phys. J. [A12]{}, 147 (2001), arXiv:hep-ph/0103062. H. Klapdor-Kleingrothaus and I. Krivosheina, Mod. Phys. Lett. [A21]{}, 1547 (2006). EXO Collaboration, M. Auger [et al.]{}, Phys. Rev. Lett. [109]{}, 032505 (2012), arXiv:1205.5608. K.-Z. collaboration, arXiv:1211.3863. KamLAND-Zen Collaboration, A. Gando [et al.]{}, Phys. Rev. [C85]{}, 045504 (2012), arXiv:1201.4664. EXO-200 Collaboration, R. MacLallen, Recontres de Moriond, [http://moriond.in2p3.fr/]{} (2012). GERDA Collaboration, I. Abt [et al.]{}, arXiv:hep-ex/0404039. Majorana Collaboration, C. Aalseth [et al.]{}, Nucl. Phys. Proc. Suppl. [217]{}, 44 (2011), arXiv:1101.0119. A. Barabash, AIP Conf. Proc. [1417]{}, 5 (2011), arXiv:1109.6423. Super-Kamiokande Collaboration, Y. Fukuda [et al.]{}, Phys. Rev. Lett. [81]{}, 1562 (1998), arXiv:hep-ex/9807003. SNO Collaboration, Q. Ahmad [et al.]{}, Phys. Rev. Lett. [89]{}, 011301 (2002), arXiv:nucl-ex/0204008. KamLAND Collaboration, K. Eguchi [et al.]{}, Phys. Rev. Lett. [90]{}, 021802 (2003), arXiv:hep-ex/0212021. T2K Collaboration, K. Abe [et al.]{}, Phys. Rev. Lett. [107]{}, 041801 (2011), arXiv:1106.2822. MINOS Collaboration, P. Adamson [et al.]{}, Phys. Rev. Lett. [107]{}, 181802 (2011), arXiv:1108.0015. DOUBLE-CHOOZ Collaboration, Y. Abe [et al.]{}, Phys. Rev. Lett. [108]{}, 131801 (2012), arXiv:1112.6353. DAYA-BAY Collaboration, F. An [et al.]{}, Phys. Rev. Lett. [108]{}, 171803 (2012), arXiv:1203.1669. RENO collaboration, J. Ahn [et al.]{}, Phys. Rev. Lett. [108]{}, 191802 (2012), arXiv:1204.0626. A. Faessler, V. Rodin, and F. Simkovic, arXiv:1206.0464. J. Menendez, A. Poves, E. Caurier, and F. Nowacki, Nucl. Phys. [A818]{}, 139 (2009), arXiv:0801.3760. J. Menendez, A. Poves, E. Caurier, and F. Nowacki, J. Phys. Conf. Ser. [312]{}, 072005 (2011). D. Forero, M. Tortola, and J. Valle, arXiv:1205.4018. J. Lesgourgues and S. Pastor, Phys. Rept. [429]{}, 307 (2006), arXiv:astro-ph/0603494. S. Hannestad, Prog. Part. Nucl. Phys. [65]{}, 185 (2010), arXiv:1007.0658. Y. Y. Wong, Ann. Rev. Nucl. Part. Sci. [61]{}, 69 (2011), arXiv:1111.1436. J. Bergstrom, A. Merle, and T. Ohlsson, JHEP [1105]{}, 122 (2011), arXiv:1103.3015. H. Pas, M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys. Lett. [B498]{}, 35 (2001), arXiv:hep-ph/0008182. H. Pas, M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys. Lett. [B453]{}, 194 (1999). Riazuddin, R. Marshak, and R. N. Mohapatra, Phys. Rev. [D24]{}, 1310 (1981). T. G. Rizzo, Phys. Lett. [B116]{}, 23 (1982). W.-Y. Keung and G. Senjanovic, Phys. Rev. Lett. [50]{}, 1427 (1983). M. Hirsch, H. Klapdor-Kleingrothaus, and O. Panella, Phys. Lett. [B374]{}, 7 (1996), arXiv:hep-ph/9602306. ATLAS Collaboration, G. Aad [et al.]{}, Eur. Phys. J. [C72]{}, 2056 (2012), arXiv:1203.5420. CMS Collaboration, PAS EXO-12-017. K. Babu and C. N. Leung, Nucl. Phys. [B619]{}, 667 (2001), arXiv:hep-ph/0106054. A. de Gouvea and J. Jenkins, Phys. Rev. [D77]{}, 013008 (2008), arXiv:0708.1344. F. del Aguila, A. Aparici, S. Bhattacharya, A. Santamaria, and J. Wudka, JHEP [1206]{}, 146 (2012), arXiv:1204.5986. S. Weinberg, Phys. Rev. Lett. [43]{}, 1566 (1979). H. Primakoff and P. S. Rosen, Ann. Rev. Nucl. Part. Sci. [31]{}, 145 (1981). M. Doi, T. Kotani, and E. Takasugi, Prog. Theor. Phys. Suppl. [83]{}, 1 (1985). J. H. Missimer, R. Mohapatra, and N. C. Mukhopadhyay, Phys. Rev. [D50]{}, 2067 (1994). E. Takasugi, Nucl. Instrum. Meth. [A503]{}, 252 (2003). S. Kanemura, Y. Kuno, and T. Ota, arXiv:1205.5681. M. Gavela, D. Hernandez, T. Ota, and W. Winter, Phys. Rev. [D79]{}, 013007 (2009), arXiv:0809.3451. F. Bonnet, D. Hernandez, T. Ota, and W. Winter, JHEP [0910]{}, 076 (2009), arXiv:0907.3143. F. Bonnet, M. Gavela, T. Ota, and W. Winter, Phys. Rev. [D85]{}, 035016 (2012), arXiv:1105.5140. F. Bonnet, M. Hirsch, T. Ota, and W. Winter, JHEP [1207]{}, 153 (2012), arXiv:1204.5862. R. Mohapatra, Phys. Rev. [D34]{}, 3457 (1986). M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys. Rev. Lett. [75]{}, 17 (1995). M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys. Rev. [D53]{}, 1329 (1996), arXiv:hep-ph/9502385. S. Goswami and W. Rodejohann, Phys. Rev. [D73]{}, 113003 (2006), arXiv:hep-ph/0512234. M. Blennow, E. Fernandez-Martinez, J. Lopez-Pavon, and J. Menendez, JHEP [1007]{}, 096 (2010), arXiv:1005.3240. A. Ibarra, E. Molinaro, and S. Petcov, JHEP [1009]{}, 108 (2010), arXiv:1007.2378. S. Choubey, M. Duerr, M. Mitra, and W. Rodejohann, JHEP [1205]{}, 017 (2012), arXiv:1201.3031. M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys. Lett. [B378]{}, 17 (1996), arXiv:hep-ph/9602305. M. Hirsch, H. Klapdor-Kleingrothaus, and S. Kovalenko, Phys.Rev. [D54]{}, 4207 (1996), arXiv:hep-ph/9603213. F. Cuypers and S. Davidson, Eur. Phys. J. [C2]{}, 503 (1998), arXiv:hep-ph/9609487. W. Buchmuller, R. Ruckl, and D. Wyler, Phys. Lett. [B191]{}, 442 (1987). R. N. Mohapatra and J. Vergados, Phys. Rev. Lett. [47]{}, 1713 (1981). P.-H. Gu, Phys. Rev. [D85]{}, 093016 (2012), arXiv:1101.5106. M. Kohda, H. Sugiyama, and K. Tsumura, arXiv:1210.5622. L. Wolfenstein, Phys. Rev. [D26]{}, 2507 (1982). W. Haxton, S. P. Rosen, and G. Stephenson, Phys. Rev. [D26]{}, 1805 (1982). B. Allanach, C. Kom, and H. Pas, Phys. Rev. Lett. [103]{}, 091801 (2009), arXiv:0902.4697. T. Han, I. Lewis, and Z. Liu, JHEP [1012]{}, 085 (2010), arXiv:1010.4309. CMS Collaboration, S. Chatrchyan [et al.]{}, Phys. Lett. [B704]{}, 123 (2011), arXiv:1107.4771. ATLAS Collaboration, G. Aad [et al.]{}, arXiv:1210.1718. ATLAS Collaboration, G. Aad [et al.]{}, ATLAS CONF-2012-088. I. Dorsner, S. Fajfer, J. F. Kamenik, and N. Kosnik, Phys. Rev. [D81]{}, 055009 (2010), arXiv:0912.0972. I. Dorsner, J. Drobnak, S. Fajfer, J. F. Kamenik, and N. Kosnik, JHEP [1111]{}, 002 (2011), arXiv:1107.5393. N. Kosnik, I. Dorsner, J. Drobnak, S. Fajfer, and J. F. Kamenik, arXiv:1111.0477. ATLAS Collaboration, G. Aad [et al.]{}, arXiv:1209.6593. G. Cacciapaglia, A. Deandrea, D. Harada, and Y. Okada, JHEP [1011]{}, 159 (2010), arXiv:1007.2933. G. Cacciapaglia [et al.]{}, JHEP [1203]{}, 070 (2012), arXiv:1108.6329. Y. Okada and L. Panizzi, arXiv:1207.5607. G. Cacciapaglia, A. Deandrea, S. Perries, V. Sordini, and L. Panizzi, arXiv:1211.4034. A. Azatov [et al.]{}, arXiv:1204.0455. N. Bonne and G. Moreau, arXiv:1206.3360. B. Batell, S. Gori, and L.-T. Wang, arXiv:1209.6382. E. Bertuzzo, P. A. Machado, and R. Z. Funchal, arXiv:1209.6359. D. McKeen, M. Pospelov, and A. Ritz, arXiv:1208.4597. K. J. Bae, T. H. Jung, and H. D. Kim, arXiv:1208.3748. H. Davoudiasl, H.-S. Lee, and W. J. Marciano, arXiv:1208.2973. B. Batell, D. McKeen, and M. Pospelov, arXiv:1207.6252. H. An, T. Liu, and L.-T. Wang, arXiv:1207.2473. S. P. Martin and J. D. Wells, Phys. Rev. [D86]{}, 035017 (2012), arXiv:1206.2956. L. Wang and X.-F. Han, arXiv:1206.1673. S. Iwamoto, AIP Conf. Proc. [1467]{}, 57 (2012), arXiv:1206.0161. M. Endo, K. Hamaguchi, S. Iwamoto, and N. Yokozaki, Phys. Rev. [D85]{}, 095012 (2012), arXiv:1112.5653. ATLAS Collaboration, G. Aad [et al.]{}, Phys. Lett. [B716]{}, 1 (2012), arXiv:1207.7214. CMS Collaboration, S. Chatrchyan [et al.]{}, Phys. Lett. [B716]{}, 30 (2012), arXiv:1207.7235. ATLAS Collaboration, G. Aad [et al.]{}, Phys. Lett. [B709]{}, 158 (2012), arXiv:1112.4828. ZEUS Collaboration, H. Abramowicz [et al.]{}, arXiv:1205.5179. KATRIN Collaboration, [http://www.katrin.kit.edu]{}, A. Osipowicz [et al.]{}, (2001), arXiv:hep-ex/0109033. H. K. Dreiner, H. E. Haber, and S. P. Martin, Phys. Rept. [494]{}, 1 (2010), arXiv:0812.1594. [^1]: [email protected] [^2]: [email protected] [^3]: [email protected] [^4]: [email protected] [^5]: For a recent version of the black box theorem including lepton flavour violation, see [@Hirsch:2006yk]. [^6]: There is also a claim [@KlapdorKleingrothaus:2006ff] for the observation of [$0\nu\beta\beta$ ]{}decay in $^{76}$Ge, which is, however, not supported by any other experiment. [^7]: The idea to use accelerator data to test [$0\nu\beta\beta$ ]{}decay contributions in left-right symmetric models has a long history. “Inverse” neutrinoless double beta decay, i.e., $e^-e^- \to W_RW_R$ has been first discussed in [@Rizzo:1982kn]. $W_R$ production at hadron colliders, followed by the decay $W_R\to \mu^+\mu^+ jj$ was first studied in [@Keung:1983uu]. [^8]: Here, $(\langle m_N\rangle)^{-1}\equiv\sum_j V_{ej}^2/m_{N_j}$ is a sum over all heavy neutrinos coupling to the electron, while in the limit from the LHC experiments it is assumed that only one heavy neutrino (the lightest) has a mass below the mass of $W_R$. [^9]: We also found a $d=9$ LNV operator $\overline{u_{R}} \overline{u_{R}} d_{R} d_{R} \overline{e_{R}} \overline{e_{R}}$ that was not listed in the earlier papers [@Babu:2001ex; @deGouvea:2007xp]. [^10]: Note that the difference in normalisation of Eq. (\[eq:CurrLR\]) and the normal convention for $L$/$R$ in particle physics leads to various powers of two, see appendix, when relating models with the $\epsilon_{\alpha}^{\beta}$ of Eq. (\[eq:defLR\]). [^11]: This momentum-enhancement mechanism has also been discussed in the context of the other LNV processes, such as $\mu^{-} N \rightarrow e^{+} N$ [@Primakoff:1981sx; @Doi:1985dx] and $\mu^{-} N \rightarrow \mu^{+} N$ [@Missimer:1994xd; @Takasugi:2003ah]. A possibility of a direct experimental test of the four-Fermi LNV interaction with a right-chiral lepton current is examined in [@Kanemura:2012br]. [^12]: However, it was shown in [@Schechter:1981bd; @Wolfenstein:1982bf; @Haxton:1982ff] that this contribution is always sub-dominant. [^13]: A similar diagram can be found in Fig. 2 of [@Kohda:2012sr]. The model therein was constructed to describe neutrino mass at two-loop order. Note that neutrino mass may be obtained at a lower than 4-loop order (as postulated by the black box diagram), in a specific model. That, however, necessarily requires that the effective operator can be promoted to an $SU(2)$ invariant operator with one or two lepton doublets, which cannot be done for every model. The $SU(2)$ invariant effective operator with one lepton doublet and a right-handed electron may induce neutrino mass at the 3-loop level (as our example in Sec. \[sec:example\]), and the operator with two lepton doublets may induce neutrino mass at the 2-loop level. If one assumes that the SM gauge invariant effective operator is constrained by the current limit of $0\nu\beta\beta$ decay, the 2- and 3-loop-induced neutrino masses via the associated operators $SU(2)_{L}$ should be sub-dominant. [^14]: Of course, some numerical factors are different in the vector cases from the scalar case. [^15]: The chiralities of the outer fermions on the vertices $\text{v}_{2}$ and $\text{v}_{3}$ are chosen so that the mass term in the propagator of the fermion mediator $\Psi^{5/3}$ is picked out. See appendix for more details. [^16]: Here we use the notation $(SU(3)_{c}, SU(2)_{L})_{U(1)_{Y}}$ from the appendix. [^17]: This is strictly true only for unmixed (valence) quarks. Note also that for vector diquarks in principle both, the ${\bf 6}_{s}$ and the ${\bf \bar{3}}_{a}$ can contribute. [^18]: In fact, given our tables in the appendix, one does not need to rely on the assumption of only one dominating NME since the coefficients are explicitely given, and one can directly translate the half-life time into the model constraints for a specific model. In this specific example, we use this dominance for the sake of simplicity. [^19]: If $m_{\Psi}< m_{\rm LQ}$ both the $3j$ and the $eej$ final states of the $\Psi$ decay will occur. [^20]: The formulae for the long-range contributions are given in [@Pas:1999fc]. Note that, when one decomposes the amplitudes into a long-range and a short range parts, one assumes implicitly that there is no new physics with a mass scale similar to the nuclear Fermi scale, i.e., ${\cal O}(100)$ MeV. [^21]: The NME for the mass mechanism have been calculated in a number of papers in the literature. Unfortunately, however, some of the NME of the general decay rate have so far been calculated only in [@Pas:1999fc; @Pas:2000vn; @Deppisch:2012nb]. [^22]: Note that the definition of 2-component tensor matrices here is different from [@Dreiner:2008tw] by an imaginary unit [i]{}. [^23]: Here, the charges are fixed, following the charge flow $\text{v}_{1} \leftarrow \text{v}_{2} \leftarrow \text{v}_{3} \leftarrow \text{v}_{4}$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We introduce complex intersection bodies and show that their properties and applications are similar to those of their real counterparts. In particular, we generalize Busemann’s theorem to the complex case by proving that complex intersection bodies of symmetric complex convex bodies are also convex. Other results include stability in the complex Busemann-Petty problem for arbitrary measures and the corresponding hyperplane inequality for measures of complex intersection bodies.' address: - \| Department of Mathematics\ University of Missouri\ Columbia\ Missouri 65211 - \| Department of Mathematics\ Texas A $\&$ M University\ College Station\ TX 77843- 3368\ - \| Department of Mathematics\ University of Crete\ Heraklio\ Crete author: - 'A. Koldobsky' - 'G. Paouris' - 'M. Zymonopoulou' title: Complex intersection bodies --- \[section\] \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Corollary]{} Ł[[L]{}]{} [[F]{}]{}[[F]{}]{} \|[c]{} v ł Introduction ============ The concept of an intersection body was introduced by Lutwak [@L], as part of his dual Brunn-Minkowski theory. In particular, these bodies played an important role in the solution of the Busemann-Petty problem. Many results on intersection bodies have appeared in recent years (see [@G; @K1; @KY] and references there), but almost all of them apply to the real case. The goal of this paper is to extend the concept of an intersection body to the complex case. Let $K$ and $L$ be origin symmetric star bodies in $\R^n.$ Following [@L], we say that $K$ is the [intersection body of]{} $L$ if the radius of $K$ in every direction is equal to the volume of the central hyperplane section of $L$ perpendicular to this direction, i.e. for every $\xi\in S^{n-1},$ $$\label{intbodyofstar} \\|\xi\\|_K^{-1} = \|L\cap \xi^\bot\|,$$ where $\\|x\\|_K=min\{a\ge 0:\ x\in aK\}$, $\xi^\bot=\{x\in \R^n:\ (x,\xi)=0\},$ and $\|\cdot\|$ stands for volume. By a theorem of Busemann [@Bu1] the intersection body of an origin symmetric convex body is also convex. However, intersection bodies of convex bodies form just a small part of the class of intersection bodies. In particular, by results of Hensley [@H] and Borell [@Bor], intersection bodies of symmetric convex bodies are isomorphic to an ellipsoid, i.e. $d_{BM}(I(K), B_{2}^{n})\le c$ where $d_{BM}$ is the Banach-Mazur distance and $c>0$ is a universal constant. The right-hand side of (\[intbodyofstar\]) can be written using the polar formula for volume: $$\\|\xi\\|_K^{-1}= \frac{1}{n-1} \int_{S^{n-1}\cap \xi^\bot} \\|\theta\\|_L^{-n+1} d\theta= \frac{1}{n-1} {\cal{R}}(\\|\cdot\\|_L^{-n+1})(\xi),$$ where the operator ${\cal{R}}: C(S^{n-1})\to C(S^{n-1})$ is the spherical Radon transform defined by $${\cal{R}}f(\xi) = \int_{S^{n-1}\cap \xi^\bot} f(x) dx.$$ This means that a star body $K$ is the intersection body of a star body if and only if the function $\\|\cdot\\|_K^{-1}$ is the spherical Radon transform of a continuous positive function on $S^{n-1}.$ A more general class of bodies was introduced in [@GLW]. A star body $K$ in $\R^n$ is called an [intersection body]{} if there exists a finite Borel measure $\mu$ on the sphere $S^{n-1}$ so that $\\|\cdot\\|_K^{-1}= R\mu$ as functionals on $C(S^{n-1}),$ i.e. for every continuous function $f$ on $S^{n-1},$ $$\label{defintbody} \int_{S^{n-1}} \\|x\\|_K^{-1} f(x)\ dx = \left( {\cal{R}}\mu, f \right)= \int_{S^{n-1}} Rf(x)\ d\mu(x).$$ We introduce complex intersection bodies along the same lines. In Section \[CIBofstarbodies\] we define complex intersection bodies of complex star bodies, and in Section \[bus\] we study complex intersection bodies of convex bodies. While the complex version of Busemann’s theorem requires a serious effort, the extension of the Hensley-Borell theorem to the complex case follows from a result of Ball [@Ball1]. In Section \[radonandfourier\] we prove that the complex spherical Radon transform and the Fourier transform of distributions coincide (up to a constant) on a class of $(-2n+2)$-homogeneous functions on $\R^{2n}$ with symmetries determined by the complex structure. A similar result in the real case was crucial for the study of real intersection bodies. We use this result in Section \[CIB\], where we define complex intersection bodies and prove a Fourier characterization of intersection bodies: an origin symmetric complex star body $K$ in $\R^{2n}$ is a complex intersection body if and only if the function $\\|\cdot\\|_K^{-2}$ represents a positive definite distribution. We use this characterization in Section \[Char of CIB\] to show that the class of complex intersection bodies coincides with the class of real 2-intersection bodies in $\R^{2n}$ and, at the same time, with the class of generalized 2-intersection bodies, provided that bodies from the real classes possess symmetries determined by the complex structure of $\R^{2n}.$ The latter allows to extend to the complex case a result of Goodey and Weil [@GW] by showing that all symmetric complex intersection bodies can be obtained as limits in the radial metric of complex radial sums of ellipsoids. Finally, Section \[stab-hyper\] deals with stability in the complex Busemann-Petty problem for arbitrary measures and related hyperplane inequalities. Complex intersection bodies of star bodies {#CIBofstarbodies} ========================================== The theory of real convex bodies goes back to ancient times and continues to be a very active field now. The situation with complex convex bodies is different, as no systematic studies of these bodies have been carried out, and results appear only occasionally; see for example [@KKZ; @KZ; @AB; @Ru1; @Zy; @Zy1]. origin symmetric convex bodies in $\C^n$ are the unit balls of norms on $\C^n.$ We denote by $\\|\cdot\\|_K$ the norm corresponding to the body $K:$ $$K=\{z\in \C^n:\ \\|z\\|_K\le 1\}.$$ In order to define volume, we identify $\C^n$ with $\R^{2n}$ using the standard mapping $$\xi = (\xi_1,...,\xi_n)=(\xi_{11}+i\xi_{12},...,\xi_{n1}+i\xi_{n2}) \mapsto (\xi_{11},\xi_{12},...,\xi_{n1},\xi_{n2}).$$ Since norms on $\C^n$ satisfy the equality $$\\|\lambda z\\| = \|\lambda\|\\|z\\|,\quad \forall z\in \C^n,\ \forall\lambda \in \C,$$ origin symmetric complex convex bodies correspond to those origin symmetric convex bodies $K$ in $\R^{2n}$ that are invariant with respect to any coordinate-wise two-dimensional rotation, namely for each $\theta\in [0,2\pi]$ and each $\xi= (\xi_{11},\xi_{12},...,\xi_{n1},\xi_{n2})\in \R^{2n}$ $$\label{rotation} \\|\xi\\|_K = \\|R_\theta(\xi_{11},\xi_{12}),...,R_\theta(\xi_{n1},\xi_{n2})\\|_K,$$ where $R_\theta$ stands for the counterclockwise rotation of $\R^2$ by the angle $\theta$ with respect to the origin. We shall say that $K$ is a [complex convex body in $\R^{2n}$]{} if $K$ is a convex body and satisfies equations (\[rotation\]). A compact set $K$ in $\R^n$ is called a star body if the origin is an interior point of $K$, every straight line passing through the origin crosses the boundary of $K$ at exactly two points and the Minkowski functional of $K$ defined by $$\\|x\\|_K= \min\{a\ge 0:\ x\in aK\},\qquad \forall x\in \R^n$$ is a continuous function on $\R^n.$ The radial function of $K$ is given by $$\rho_K(x) = \max\{a>0:\ ax\in K\}.$$ If $x\in S^{n-1},$ then $\rho_K(x)$ is the radius of $K$ in the direction of $x.$ Note that for any unit vector $\xi$, $\rho_K(\xi)=\\|\xi\\|_K^{-1}.$ The radial metric in the class of star bodies is defined by $$\rho(K,L)=\max_{\xi\in S^{n-1}} \|\rho_K(\xi)-\rho_L(\xi)\|.$$ If the Minkowski functional of a star body $K$ in $\R^{2n}$ is $R_\theta$-invariant (i.e. satisfies equations (\[rotation\])), we say that $K$ is a [complex star body]{} in $\R^{2n}.$ For $\xi\in \C^n, \|\xi\|=1,$ denote by $$H_\xi = \{ z\in \C^n:\ (z,\xi)=\sum_{k=1}^n z_k\overline{\xi_k} =0\}$$ the complex hyperplane through the origin, perpendicular to $\xi.$ Under the standard mapping from $\C^n$ to $\R^{2n}$ the hyperplane $H_\xi$ turns into a $(2n-2)$-dimensional subspace of $\R^{2n}$ orthogonal to the vectors $$\xi=(\xi_{11},\xi_{12},...,\xi_{n1},\xi_{n2})\quad {\rm and } \quad \xi^\perp =(-\xi_{12},\xi_{11},...,-\xi_{n2},\xi_{n1}).$$ The orthogonal two-dimensional subspace $H_\xi^\bot$ has orthonormal basis $\left\{\xi,\xi^\bot\right\}.$ A star (convex) body $K$ in $\R^{2n}$ is a complex star (convex) body if and only if, for every $\xi\in S^{2n-1},$ the section $K\cap H_\xi^\bot$ is a two-dimensional Euclidean circle with radius $\rho_K(\xi)= \\|\xi\\|_K^{-1}.$ We introduce complex intersection bodies of complex star bodies using a definition under which these bodies play the same role in complex convexity, as their real counterparts in the real case. We use the notation $\|K\|$ for the volume of $K;$ the dimension where we consider volume is clear in every particular case. \[intof\] Let $K, L$ be origin symmetric complex star bodies in $\R^{2n}.$ We say that $K$ is the complex intersection body of $L$ and write $K=I_c(L)$ if for every $\xi \in \R^{2n}$ $$\label{def-intbody} \|K\cap H_\xi^{\perp}\|= \|L\cap H_\xi\|.$$ Since $K\cap H_\xi^{\perp}$ is the two-dimensional Euclidean circle with radius $\\|\xi\\|_K^{-1}$, (\[def-intbody\]) can be written as $$\label{intbody} \pi \\|\xi\\|_{I_c(L)}^{-2} = \|L\cap H_\xi\|.$$ All the bodies $K$ that appear as complex intersection bodies of different complex star bodies form [the class of complex intersection bodies of star bodies]{}. In Section \[CIB\], we will introduce a more general class of [complex intersection bodies]{}. The Radon and Fourier transforms of $R_\theta$-invariant functions {#radonandfourier} ================================================================== Denote by $C_c(S^{2n-1})$ the space of $R_\theta$-invariant continuous functions, i.e. continuous real-valued functions $f$ on the unit sphere $S^{2n-1}$ in $\R^{2n}$ satisfying $f(\xi)=f(R_\theta(\xi))$ for all $\xi\in S^{2n-1}$ and all $\theta\in [0,2\pi].$ The complex spherical Radon transform is an operator ${\cal{R}}_c: C_c(S^{2n-1})\to C_c(S^{2n-1})$ defined by $${\cal{R}}_cf(\xi) = \int_{S^{2n-1}\cap H_\xi} f(x) dx.$$ Writing volume in polar coordinates, we get that for every complex star body $L$ in $\R^{2n}$ and every $\xi\in S^{2n-1},$ $$\label{v-r} \|L\cap H_\xi\| = \frac1{2n-2} \int_{S^{2n-1}\cap H_\xi} \\|x\\|_L^{-2n+2} dx = \frac1{2n-2} {\cal{R}}_c\left(\\|\cdot\\|_L^{-2n+2}\right)(\xi),$$ so the condition (\[intbody\]) reads as $$\label{int-radon} \\|\xi\\|_{I_c(L)}^{-2} = \frac1{2\pi(n-1)} {\cal{R}}_c\left(\\|\cdot\\|_L^{-2n+2}\right)(\xi).$$ This means that a complex star body $K$ is a complex intersection body of a star body if and only if the function $\\|\cdot\\|_K^{-2}$ is the complex spherical Radon transform of a continuous positive $R_\theta$-invariant function on $S^{2n-1}.$ We use this observation in Section \[CIB\], where we introduce a more general class of complex intersection bodies (not depending on the underlying star body), like it was done in the real case in [@GLW]. But before that we need several facts connecting the Radon transform to the Fourier transform in the complex setting. We use the techniques of the Fourier transform of distributions; see [@GS] for details. As usual, we denote by ${\cal{S}}(\R^n)$ the Schwartz space of rapidly decreasing infinitely differentiable functions (test functions) in $\R^n,$ and ${\cal{S}}^{'}(\R^n)$ is the space of distributions over ${\cal{S}}(\R^n).$ Suppose that $f$ is a locally integrable complex-valued function on $\R^n$ with [power growth at infinity]{}, i.e. there exists a number $ \beta>0$ so that $$\lim_{\|x\|_2\to \infty} \frac{f(x)} {\|x\|_2^\beta}=0,$$ where $\|\cdot\|_2$ stands for the Euclidean norm on $\R^n.$ Then $f$ represents a distribution acting by integration: for every $\phi\in {\mathcal S},$ $$\langle f, \phi \rangle = \int_{\R^n} f(x) \phi(x)\ dx.$$ The Fourier transform of a distribution $f$ is defined by $\langle\hat{f}, \phi\rangle= \langle f, \hat{\phi} \rangle$ for every test function $\phi.$ If $\phi$ is an even test function, then $(\hat{\phi})^\wedge = (2\pi)^n \phi$, so the Fourier transform is self-invertible (up to a constant) for even distributions. A distribution $f$ is called even homogeneous of degree $p\in \R$ if $$\langle f(x), \phi(x/\alpha) \rangle = \|\alpha\|^{n+p} \langle f,\phi \rangle$$ for every test function $\phi$ and every $\alpha\in \R,\ \alpha\neq 0.$ The Fourier transform of an even homogeneous distribution of degree $p$ is an even homogeneous distribution of degree $-n-p.$ We say that a distribution is [positive definite]{} if its Fourier transform is a positive distribution in the sense that $\langle \hat{f},\phi \rangle \ge 0$ for every non-negative test function $\phi.$ Schwartz’s generalization of Bochner’s theorem (see, for example, [@GV p.152]) states that a distribution is positive definite if and only if it is the Fourier transform of a tempered measure on $\R^n$. Recall that a (non-negative, not necessarily finite) measure $\mu$ is called tempered if $$\int_{\R^n} (1+\|x\|_2)^{-\beta}\ d\mu(x)< \infty$$ for some $\beta >0.$ Our definition of a star body $K$ assumes that the origin is an interior point of $K.$ If $0<p<n,$ then $\\|\cdot\\|_K^{-p}$ is a locally integrable function on $\R^n$ and represents an even homogeneous of degree $-p$ distribution. If $\\|\cdot\\|_K^{-p}$ represents a positive definite distribution for some $p\in (0,n),$ then its Fourier transform is a tempered measure which is at the same time a homogeneous distribution of degree $-n+p.$ One can express such a measure in polar coordinates, as follows. \[posdef\] ([@K1 Corollary 2.26]) Let $K$ be an origin symmetric convex body in $\R^n$ and $p\in (0,n).$ The function $\\|\cdot\\|_K^{-p}$ represents a positive definite distribution on $\R^n$ if and only if there exists a finite Borel measure $\mu$ on $S^{n-1}$ so that for every even test function $\phi,$ $$\int_{\R^n} \\|x\\|_K^{-p} \phi(x)\ dx = \int_{S^{n-1}} \left( \int_0^\infty t^{p-1} \hat\phi(t\xi) dt \right) d\mu(\xi).$$ For any even continuous function $f$ on the sphere $S^{n-1}$ and any non-zero number $p\in \R,$ we denote by $f\cdot r^{p}$ the extension of $f$ to an even homogeneous function of degree $p$ on $\R^n$ defined as follows. If $x\in \R^n,$ then $x=r\theta,$ where $r=\|x\|_2$ and $\theta= x/\|x\|_2.$ We put $$f\cdot r^p(x) = f\left(\theta\right) r^p.$$ It was proved in [@K1 Lemma 3.16] that, for any $p\in (-n,0)$ and infinitely smooth function $f$ on $S^{n-1},$ the Fourier transform of $f\cdot r^{-p}$ is equal to another infinitely smooth function $h$ on $S^{n-1}$ extended to an even homogeneous of degree $-n+p$ function $h\cdot r^{-n+p}$ on the whole of $\R^n.$ The following Parseval formula on the sphere was proved in [@K1 Corollary 3.22]. Let $f,g$ be even infinitely smooth functions on $S^{n-1},$ and $p\in (0,n).$ Then $$\label{parseval} \int_{S^{n-1}} (f\cdot r^{-p})^\wedge(\theta) (g\cdot r^{-n+p})^\wedge(\theta) = (2\pi)^n \int_{S^{n-1}} f(\theta)g(\theta) \ d\theta.$$ We need a simple observation that will, however, provide the basis for applications of the Fourier transform to complex bodies. \[const\] Suppose that $f\in C_c(S^{2n-1})$ is an even infinitely smooth function. Then for every $0<p<2n$ and $\xi\in S^{2n-1}$ the Fourier transform of the distribution $f\cdot r^{-p}$ is a constant function on $S^{2n-1}\cap H_\xi^\bot.$ By [@K1 Lemma 3.16], the Fourier transform of $f\cdot r^{-p}$ is a continuous function outside of the origin in $\R^{2n}.$ The function $f$ is invariant with respect to all $R_\theta$, so by the connection between the Fourier transform of distributions and linear transformations, the Fourier transform of $f\cdot r^{-p}$ is also invariant with respect to all $R_\theta.$ Recall that the two-dimensional space $H_\xi^\bot$ is spanned by vectors $\xi$ and $\xi^\bot$ (see the Introduction). Every vector in $S^{2n-1}\cap H_\xi^\bot$ is the image of $\xi$ under one of the coordinate-wise rotations $R_\theta$, so the Fourier transform of $f\cdot r^{-p}$ is a constant function on $S^{2n-1}\cap H_\xi^\bot.$ The following connection between the Fourier and Radon transforms is well-known; see for example [@K1 Lemma 3.24]. \[f-r\] Let $1\le k <n,$ and let $\phi\in {\cal{S}}(\R^n)$ be an even test function. Then for any $(n-k)$-dimensional subspace $H$ of $\R^n$ $$\int_H \phi(x) dx = \frac1{(2\pi)^k} \int_{H^\bot} \hat{\phi}(x) dx.$$ We also use the spherical version of Proposition \[f-r\]; see [@K1 Lemma 3.25]. \[perp\] Let $\phi$ be an even infinitely smooth function on $S^{n-1},$ let $0<k<n,$ and let $H$ be an arbitrary $(n-k)$-dimensional subspace of $\R^n.$ Then $$\int_{S^{n-1} \cap H} \phi(\theta)d\theta = \frac1{(2\pi)^k} \int_{S^{n-1}\cap H^\bot} \left(\phi\cdot r^{-n+k}\right)^\wedge(\theta) d\theta.$$ Let us translate the latter fact to the complex situation. \[radon-fourier\] Let $\phi\in C_c(S^{2n-1})$ be an even infinitely smooth function. Then for every $\xi\in S^{2n-1}$ $${\cal{R}}_c\phi(\xi)= \frac1{2\pi} \left(\phi\cdot r^{-2n+2}\right)^\wedge (\xi).$$ By Proposition \[perp\], $${\cal{R}}_c\phi(\xi) = \int_{S^{2n-1} \cap H_\xi} \phi(\theta)d\theta= \frac1{(2\pi)^2} \int_{S^{2n-1}\cap H_\xi^\bot} \left(\phi\cdot r^{-2n+2}\right)^\wedge (\theta) d\theta,$$ and, by Lemma \[const\], the function under the integral in the right-hand side is constant on $S^{2n-1}\cap H_\xi^\bot.$ The value of this constant is the function value at $\xi \in S^{2n-1}\cap H_\xi^\bot.$ Also, recall that $S^{2n-1}\cap H_\xi^\bot$ is the two-dimensional Euclidean unit circle, so $$\int_{S^{2n-1}\cap H_\xi^\bot} \left(\phi\cdot r^{-2n+2}\right)^\wedge (\theta) d\theta = 2\pi \left(\phi\cdot r^{-2n+2}\right)^\wedge (\xi).$$ \[self-dual\]The complex spherical Radon transform is self-dual, i.e. for any even functions $f,g\in C_c(S^{2n-1}),$ $$\int_{S^{2n-1}} {\cal{R}}_cf(\xi) g(\xi) d\xi = \int_{S^{2n-1}} f(\theta) {\cal{R}}_c g(\theta) d\theta.$$ By approximation, it is enough to consider the case where $f,g$ are infinitely smooth. For some infinitely smooth even function $h\in C_c(S^{2n-1}),$ we have $g\cdot r^{-2n+2}= \left(h\cdot r^{-2}\right)^\wedge,$ then $ \left(g\cdot r^{-2n+2}\right)^\wedge= (2\pi)^{2n} h\cdot r^{-2}.$ By Lemma \[radon-fourier\] and the spherical Parseval formula (\[parseval\]), $$\int_{S^{2n-1}} {\cal{R}}_cf(\xi) g(\xi) d\xi = \frac1{2\pi} \int_{S^{2n-1}} \left(f\cdot r^{-2n+2}\right)^\wedge(\xi) (g\cdot r^{-2n+2}) (\xi) d\xi$$ $$= \frac{(2\pi)^{2n}}{2\pi} \int_{S^{2n-1}} \left(f\cdot r^{-2n+2}\right)^\wedge(\xi) (h\cdot r^{-2})^\wedge (\xi) d\xi$$ $$= \frac1{2\pi} \int_{S^{2n-1}} f(\theta) \left(g\cdot r^{-2n+2}\right)^\wedge(\theta) d\theta = \int_{S^{2n-1}} f(\theta) {\cal{R}}_c g(\theta) d\theta.$$ We now prove Lemma \[radon-fourier\] without smoothness assumption. This result is a complex version of [@K1 Lemma 3.7]. We say that a distribution $f$ on $\R^{2n}$ is $R_\theta$-invariant if $\langle f, \phi(R_\theta(\cdot))\rangle = \langle f, \phi \rangle$ for every test function $\phi \in {\cal{S}}(\R^{2n})$ and every $\theta\in [0,2\pi].$ If $f$ and $g$ are $R_\theta$-invariant distributions, and $\langle f, \phi \rangle=\langle g, \phi \rangle$ for any test function $\phi$ that is invariant with respect to all $R_\theta,$ then $f=g.$ This follows from the observation that the value of an $R_\theta$-invariant distribution on a test function $\phi$ does not change if $\phi$ is replaced by the function $\frac1{2\pi}\int_0^{2\pi} \phi(R_\theta(\cdot))d\theta.$ \[r-s\] Let $f\in C_c(S^{2n-1})$ be an even function. Then the Fourier transform of $f\cdot r^{-2n+2}$ is a continuous function on the sphere extended to a homogeneous function of degree -2 on the whole $\R^{2n}.$ Moreover, on the sphere this function is equal (up to a constant) to the complex spherical Radon transform of $f$: for any $\xi\in S^{2n-1},$ $${\cal{R}}_cf(\xi) = \frac1{2\pi} \left(f\cdot r^{-2n+2}\right)^\wedge(\xi).$$ Let $\phi \in {\cal{S}}(\R^{2n})$ be any even $R_\theta$-invariant test function. Then $\hat{\phi}$ is also an even $R_\theta$-invariant test function. By Lemma \[const\], for any $\xi\in S^{2n-1},$ $$\label{eq1} \int_{H_\xi^\bot} \hat{\phi}(x) dx = \int_{S^{2n-1}\cap H_\xi^\bot} \left(\int_0^\infty r\hat{\phi}(r\theta)dr \right) d\theta = 2\pi \int_0^\infty r\hat{\phi}(r\xi)dr.$$ By Proposition \[f-r\] and Lemma \[self-dual\], $$\langle \left(f\cdot r^{-2n+2}\right)^\wedge, \phi \rangle = \int_{\R^{2n}} \|x\|_2^{-2n+2} f(x/\|x\|_2) \hat{\phi}(x) dx$$ $$= \int_{S^{2n-1}} f(\xi) \left(\int_0^\infty r\hat{\phi}(r\xi) dr \right) d\xi = \frac1{2\pi} \int_{S^{2n-1}} f(\xi) \left(\int_{H_\xi^\bot} \hat{\phi}(x) dx\right) d\xi$$ $$= 2\pi \int_{S^{2n-1}} f(\xi) \left(\int_{H_\xi} \phi(x) dx \right) d\xi$$ $$= 2\pi \int_{S^{2n-1}} f(\xi) \left(\int_{S^{2n-1}\cap H_\xi} \left(\int_0^\infty r^{2n-3} \phi(r\theta) dr \right) d\theta \right) d\xi$$ $$= 2\pi \int_{S^{2n-1}} f(\xi)\ {\cal{R}}_c\left(\int_0^\infty r^{2n-3}\phi(r\cdot)dr\right)(\xi) d\xi$$ $$= 2\pi \int_{S^{2n-1}} \left(\int_0^\infty r^{2n-3}\phi(r\theta)dr\right) {\cal{R}}_cf(\theta) d\theta$$ $$= 2\pi \int_{\R^{2n}} \|x\|_2^{-2} {\cal{R}}_cf(x/\|x\|_2) \phi(x) dx.$$ We get that for every even $R_\theta$-invariant test function $\phi,$ $$\langle \left(f\cdot r^{-2n+2}\right)^\wedge, \phi \rangle = 2\pi \langle \|x\|_2^{-2} {\cal{R}}_cf(x/\|x\|_2) , \phi \rangle,$$ so even $R_\theta$-invariant distributions $ \left(f\cdot r^{-2n+2}\right)^\wedge$ and $2\pi \|x\|_2^{-2} {\cal{R}}_cf(x/\|x\|_2)$ are equal. Lemma \[r-s\] implies the following Fourier transform formula for the volume of sections of star bodies. Note that the real version of this formula was proved in [@K2], and that the complex formula below was proved in [@KKZ] for infinitely smooth bodies by a different method; here we remove the smoothness condition. \[volume-ft\] Let $K$ be an origin symmetric complex star body in $\R^{2n}, n\ge 2.$ For every $\xi\in S^{2n-1},$ we have $$\|K\cap H_\xi\| = \frac{1}{4\pi(n-1)} \left(\\|x\\|_K^{-2n+2}\right)^\wedge(\xi).$$ By (\[v-r\]) and Lemma \[r-s\] applied to the function $f(\theta)= \\|\theta\\|_K^{-2n+2},$ $$\|K\cap H_\xi\| = \frac1{2n-2} {\cal{R}}_c\left(\\|\cdot\\|_K^{-2n+2}\right)(\xi)= \frac{1}{4\pi(n-1)} \left(\\|x\\|_K^{-2n+2}\right)^\wedge(\xi).$$ We use Theorem \[volume-ft\] to prove the complex version of the Minkowski-Funk theorem saying that an origin symmetric star body is uniquely determined by volume of its central hyperplane sections; see [@K1 Corollary 3.9]. If $K,L$ are origin symmetric complex star bodies in $\R^{2n},$ and their intersection bodies $I_c(K)$ and $I_c(L)$ coincide, then $K=L.$ The equality of intersection bodies means that, for every $\xi\in S^{2n-1},$ $\|K\cap H_\xi\|= \|L\cap H_\xi\|.$ By Theorem \[volume-ft\], homogeneous of degree -2 continuous on $\R^{2n}\setminus \{0\}$ functions $\left(\\|\cdot\\|_K^{-2n+2}\right)^\wedge$ and $\left(\\|\cdot\\|_L^{-2n+2}\right)^\wedge$ coincide on the sphere $S^{2n-1}$, so they are also equal as distributions on the whole of $\R^{2n}.$ The result follows from the uniqueness theorem for the Fourier transform of distributions. Complex intersection bodies {#CIB} =========================== We are going to define the class of complex intersection bodies by extending the equality (\[int-radon\]) to measures, as it was done in the real case in [@GLW]. We say that a finite Borel measure $\mu$ on $S^{2n-1}$ is $R_\theta$-invariant if for any continuous function $f$ on $S^{2n-1}$ and any $\theta\in [0,2\pi]$, $$\int_{S^{2n-1}} f(x) d\mu(x) = \int_{S^{2n-1}} f(R_\theta x) d\mu(x).$$ The complex spherical Radon transform of an $R_\theta$-invariant measure $\mu$ is defined as a functional ${\cal{R}}_c\mu$ on the space $C_c(S^{2n-1})$ acting by $$\left({\cal{R}}_c\mu, f \right) = \int_{S^{2n-1}} {\cal{R}}_cf(x) d\mu(x).$$ \[int\] An origin symmetric complex star body $K$ in $\R^{2n}$ is called a [complex intersection body]{} if there exists a finite Borel $R_\theta$-invariant measure $\mu$ on $S^{2n-1}$ so that $\\|\cdot\\|_K^{-2}$ and ${\cal{R}}_c\mu$ are equal as functionals on $C_c(S^{2n-1}),$ i.e. for any $f\in C_c(S^{2n-1}),$ $$\int_{S^{2n-1}} \\|x\|\|_K^{-2} f (x)\ dx = \int_{S^{2n-1}} {\cal{R}}_c f(\theta) d\mu(\theta).$$ Clearly, ${\cal{R}}_c\mu$ is a finite Borel $R_\theta$-invariant measure on $S^{2n-1}.$ Also, an easy consequence of self-duality of the complex spherical Radon transform (see Lemma \[self-dual\]) is that if $\mu$ has density $f$ on $S^{2n-1}$, then the measure ${\cal{R}}_c\mu$ has density ${\cal{R}}_cf.$ The latter, in conjunction with (\[int-radon\]), immediately implies that every complex intersection body of a star body is a complex intersection body in the sense of Definition \[int\]. Many results on real intersection bodies depend on the following Fourier characterization (see [@K5 Theorem 1]): an origin symmetric star body $K$ in $\R^n$ is an intersection body if and only if the function $\\|\cdot\\|_K^{-1}$ represents a positive definite distribution. Complex intersection bodies admit a similar characterization. To see the connection with the Fourier transform, combine the definition of the intersection body of a star body (\[def-intbody\]) with the result of Theorem \[volume-ft\]: for every $\xi\in S^{2n-1},$ $$\\|\xi\\|_{I_c(L)}^{-2} = \frac1{\pi}\|L\cap H_\xi\| = \frac{1}{4\pi^2(n-1)} \left(\\|x\\|_L^{-2n+2}\right)^\wedge(\xi).$$ Both sides of the latter equality are even homogeneous functions of degree -2, so these functions are equal as distributions on the whole of $\R^{2n}.$ Since the Fourier transform of even distributions is self-invertible (up to a constant), we get $$\left(\\|\cdot\\|_{I_c(L)}^{-2}\right)^\wedge = \frac{(2\pi)^{2n}}{4\pi^2(n-1)} \\|\cdot\\|_L^{-2n+2} >0,$$ so the distribution $\\|\cdot\\|_{I_c(L)}$ is positive definite. Moreover, if the Fourier transform of $\\|\cdot\\|_K^{-2}$ is an even strictly positive $R_\theta$-invariant function on the sphere, then one can use the latter equality to construct a complex star body $L$ such that $K=I_c(L).$ This connection holds for arbitrary complex intersection bodies, as shown in the following theorem. \[posdef\] An origin symmetric complex star body $K$ in $\R^{2n}$ is a complex intersection body if and only if the function $\\|x\\|_K^{-2}$ represents a positive definite distribution on $\R^{2n}.$ Suppose that $K$ is a complex intersection body with the corresponding measure $\mu.$ To prove that $\\|\cdot\\|_K^{-2}$ is a positive definite distribution, it is enough to show that $\langle (\\|\cdot\\|_K^{-2})^\wedge, \phi\rangle \ge 0$ for every even $R_\theta$-invariant non-negative test function $\phi.$ By Definition \[int\] and Proposition \[f-r\], $$\langle (\\|\cdot\\|_K^{-2})^\wedge, \phi\rangle = \int_{\R^{2n}} \\|x\\|_K^{-2}\hat{\phi}(x) dx$$ $$= \int_{S^{2n-1}} \\|\theta\\|_K^{-2} \left(\int_0^\infty r^{2n-3} \hat{\phi}(r\theta) dr \right) d\theta$$ $$= \int_{S^{2n-1}} {\cal{R}}_c\left(\int_0^\infty r^{2n-3} \hat{\phi}(r\cdot) dr\right)(\xi) d\mu(\xi)$$ $$= \int_{S^{2n-1}} \left( \int_{H_\xi} \hat{\phi}(x) dx \right) d\mu(\xi)$$ $$= (2\pi)^{2n-2} \int_{S^{2n-1}} \left( \int_{H_\xi^\bot} \phi(x) dx \right) d\mu(\xi) \ge 0.$$ Now suppose that $\\|\cdot\\|_K^{-2}$ is a positive definite distribution. By Proposition \[posdef\], there exists a finite Borel measure $\mu$ on $S^{2n-1}$ such that for any even test function $\phi$ $$\label{eq2} \int_{\R^{2n}} \\|x\\|_K^{-2} \phi(x)\ dx = \int_{S^{2n-1}} \left(\int_0^\infty t \hat\phi(t\xi) dt \right) d\mu(\xi).$$ Recall that $K$ is $R_\theta$-invariant, so we can assume that $\mu$ is $R_\theta$-invariant and the latter equality holds only for even $R_\theta$-invariant test functions $\phi.$ For each such test function, we have by (\[eq1\]) that $$\int_0^\infty t \hat{\phi}(t\xi) dt = \frac1{2\pi}\int_{H_\xi^\bot} \hat{\phi}(x) dx.$$ Using this and Proposition \[f-r\] and then writing the interior integral in polar coordinates, we get that the right-hand side of (\[eq2\]) is equal to $$\frac1{2\pi} \int_{S^{2n-1}} \left( \int_{H_\xi^\bot} \hat{\phi}(x) dx \right) d\mu(\xi)= 2\pi \int_{S^{2n-1}} \left( \int_{H_\xi} \phi(x) dx \right) d\mu(\xi)$$ $$= 2\pi \int_{S^{2n-1}} {\cal{R}}_c\left(\int_0^\infty r^{2n-3} \phi(r\cdot) dr \right)(\xi)\ d\mu(\xi).$$ Writing the left-hand side of (\[eq2\]) in polar coordinates we get $$\int_{S^{2n-1}} \\|\theta\\|_K^{-2} \left(\int_0^\infty r^{2n-3} \phi(r\theta) dr \right) d\theta$$ $$\label{eq3} =2\pi \int_{S^{2n-1}} {\cal{R}}_c\left(\int_0^\infty r^{2n-3} \phi(r\cdot) dr \right)(\xi)\ d\mu(\xi)$$ for any even $R_\theta$-invariant test function $\phi.$ Now let $\phi(x)=u(r)v(\theta)$ for every $x\in \R^{2n},$ where $x=r\theta,\ r\in [0,\infty),\ \theta\in S^{2n-1},$ $u\in {\cal{S}}(\R)$ is a non-negative test function on $\R,$ and $v$ is an arbitrary infinitely differentiable $R_\theta$-invariant function on $S^{2n-1}.$ Then $$\int_0^\infty r^{2n-3} \phi(r\theta) dr = v(\theta) \int_0^\infty r^{2n-3} u(r) dr,$$ so the equality (\[eq3\]) turns into $$\int_{S^{2n-1}} \\|\theta\\|_K^{-2} v(\theta) d\theta =2\pi \int_{S^{2n-1}} {\cal{R}}_c v(\xi)\ d\mu(\xi).$$ Since infinitely smooth functions are dense in $C_c(S^{2n-1}),$ the latter equality also holds for any function $v\in C_c(S^{2n-1}),$ which means that $K$ is a complex intersection body. Characterizations of complex intersection bodies {#Char of CIB} ================================================ Complex intersection bodies are related to two generalizations of the concept of a real intersection body. These relations allow to apply to the complex case many results established originally in the real case. The concept of a $k$-intersection body was introduced in [@K3; @K4]. For an integer $k,\ 1\le k <n$ and star bodies $D,L$ in $\R^n,$ we say that $D$ is the $k$-intersection body of $L$ if for every $(n-k)$-dimensional subspace $H$ of $\R^n,$ $$\|D\cap H^\bot\|= \|L\cap H\|.$$ The class of $k$-intersection bodies was defined in [@K4] (see also [@K1 Section 4.2]) as follows. Let $1\le k < n.$ We say that an origin symmetric star body $D$ in $\R^n$ is a $k$-intersection body if there exists a finite Borel measure $\mu$ on $S^{n-1}$ so that for every even test function $\phi$ in $\R^n,$ $$\label{defkintgeneral} \int_{\R^n} \\|x\\|_D^{-k} \phi(x)\ dx = \int_{S^{n-1}} \left( \int_0^\infty t^{k-1} \hat\phi(t\xi)\ dt\right) d\mu(\xi).$$ The class of $k$-intersection bodies is related to a certain generalization of the Busemann-Petty problem in the same way as intersection bodies are related to the original problem (see [@K1] for details; this generalization offers a condition that allows to compare volumes of two bodies in arbitrary dimensions). An equivalent and probably more geometric way to define $k$-intersection bodies would be to say that these bodies are limits in the radial metric of $k$-intersection bodies of star bodies (see [@EMilman1] or [@Ru] for a proof of equivalence of this property to the original definition from [@K4]). It was shown in [@K4] that an origin symmetric star body $K$ in $\R^n$ is a $k$-intersection body if and only if the function $\\|\cdot\\|_K^{-k}$ represents a positive definite distribution. By Theorem \[posdef\], \[Char2\] An origin symmetric complex star body $K$ in $\R^{2n}$ is a complex intersection body if and only if it is a 2-intersection body in $\R^{2n}$ satisfying (\[rotation\]). It was proved in [@KKZ Theorem 3] that every origin symmetric complex convex body $K$ in $\R^{2n}$ is a $(2n-4)$-, $(2n-3)$- and $(2n-2)$-intersection body in $\R^{2n}.$ It follows that \[n<4\]Every origin symmetric complex convex body in $\R^6$ and $\R^4$ is a complex intersection body. This is no longer true in $\R^{2n},\ n\ge 4$ as shown in [@KKZ Theorem 4]. The unit balls of complex $\ell_q$-balls with $q>2$ are not $k$-intersection bodies for any $1\le k < 2n-4.$ Zhang in [@Zh3] introduced another generalization of intersection bodies. For $1\leq k\leq n-1$, the ($n-k$)-dimensional spherical Radon transform is an operator ${\cal{R}}_{n-k}:C(S^{n-1})\mapsto C(G(n,n-k))$ defined by $${\cal{R}}_{n-k}(f)(H)=\int_{S^{n-1}\cap H}f(x)dx,\quad H\in G(n,n-k).$$ Here $G(n,n-k)$ is the Grassmanian of $(n-k)$-dimensional subspaces of $\R^n.$ Denote the image of the operator ${\cal{R}}_{n-k}$ by X: $${\cal{R}}_{n-k}\left(C(S^{n-1})\right)=X\subset C(G(n,n-k)).$$ Let $M^+(X)$ be the space of linear positive continuous functionals on $X$, i.e. for every $\nu\in M^+(X)$ and non-negative function $f\in X$, we have $\nu(f)\geq0$. (Zhang [@Zh3]) An origin symmetric star body $K$ in $\R^n$ is called a [generalized $k$-intersection body]{} if there exists a functional $\nu\in M^+(X)$ so that for every $f\in C(S^{n-1})$, $$\int_{S^{n-1}}\\|x\\|_K^{-k} f(x)d x= \nu({\cal{R}}_{n-k}{f}).$$ It is easy to see that every complex intersection body in $\R^{2n}$ is a generalized 2-intersection body in $\R^{2n}.$ If $K$ is a complex intersection body, then there exists an even $R_\theta$-invariant measure $\mu$ on $S^{2n-1}$ such that for every $f\in C(S^{2n-1}),$ $$\int_{S^{2n-1}} \\|x\\|_K^{-2} f(x) dx = \int_{S^{2n-1}} {\cal{R}}_cf(\xi) d\mu(\xi)$$ $$=\int_{G(2n,2n-2)} {\cal{R}}_{2n-2} f(H) d\nu(H),$$ where $\nu$ is a measure on $G(2n,2n-2)$ which is the image of $\mu$ under the mapping $\xi\mapsto H_\xi$ from $S^{2n-1}$ to $G(2n,2n-2).$ The measure $\nu$ can be considered as a positive continuous functional on $X$ acting by $$\nu({\cal{R}}_{n-k}f)= \int_{G(2n,2n-2)} {\cal{R}}_{n-k}f(H) d\nu(H),$$ which implies that $K$ is a generalized 2-intersection body in $\R^{2n}.$ On the other hand, it was shown in [@K4] (see also [@K1 Theorem 4.23]) that every generalized $k$-intersection body in $\R^n$ is a $k$-intersection body. So we have shown the following An origin symmetric complex star body $K$ in $\R^{2n}$ is a complex intersection body if and only if it is a generalized $2$-intersection body in $\R^{2n}$ satisfying (\[rotation\]). Let us point out that the latter result is surprising. Combined with Corollary [\[Char2\]]{} implies that under the invariance assumption (\[rotation\]) the class of $2$-intersection bodies is exactly the class of generalized $2$-intersection bodies. Without the invariance assumption (\[rotation\]) this is no longer true as it follows from an example of E. Milman [[@EMilman2]]{}. Goodey and Weil in [[@GW]]{} proved that any intersection body is the limit (in the radial metric topology) of finite radial sums of ellipsoids. This result has been extended by Grinberg and Zhang [@GZ] (see another proof in [@EMilman1]) to the case of generalized $k$-intersection bodies where the radial sum is replaced by the $k$-radial sum. Now we are going to prove a complex version of the result of Goodey and Weil. We do it by adjusting to the complex case the proofs from [@EMilman1] and [@KY Theorem 3.10]. We define the complex radial sum $K_{1}+^{c} K_{2}$ of two complex star bodies $K_{1}, K_{2}$ as the complex star body that has radial function $$\rho_{K_{1}+^{c} K_{2}}^{2} = \rho_{K_{1}}^{2} + \rho_{K_{2}}^{2} .$$ The latter can be written as $$\\|\cdot\\|_{K_{1}+^{c} K_{2}}^{-2} = \\|\cdot\\|_{K_1}^{-2} + \\|\cdot\\|_{K_1}^{-2}.$$ \[Thm:approximation\] Let $K$ be an origin symmetric complex star body in $\R^{2n}$. Then $K$ is a complex intersection body if and only if $ \\|\cdot\\|_K^{-2}$ is the limit (in the metric of the space $C_c(S^{2n-1})$) of finite sums $\\|\cdot\\|_{E_1}^{-2}+\cdots+ \\|\cdot\\|_{E_m}^{-2} $, where $E_1$,...,$E_m$ are complex ellipsoids in $\R^{2n}$ (i.e. those ellipsoids in $\R^{2n}$ that are complex convex bodies). In other words, an origin symmetric complex star body is an intersection body if and only if it is the limit (in the radial metric) of complex radial sums of complex ellipsoids. To prove this result we need a few lemmas. For fixed $\xi\in S^{2n-1}$, $a>0$, $b>0$, let $E_{a,b}(\xi)$ be an ellipsoid in $\R^{2n}$ with the norm $$\\|x\\|_{E_{a,b}(\xi)}=\left(\frac{( x, \xi )^2+ (x, \xi^\bot)^2}{a^2}+ \frac{\|x\|_2-( x, \xi )^2- (x,\xi^\bot)^2}{b^2}\right)^{\frac12}, \ x\in \R^{2n}.$$ Clearly, $E_{a,b}(\xi)$ is a complex ellipsoid. In fact, $(x, \xi )^2+ (x, \xi^\bot)^2 = \|(x,\xi)_c\|^2$ is the modulus squared of the complex scalar product of $x$ and $\xi$ considered as vectors from $\C^n.$ The latter does not change when $x$ is multiplied by any complex number of modulus 1, which means that the norm of $E_{a,b}(\xi)$ is invariant with respect to all rotations $R_\theta.$ Using the formula for the Fourier transform of powers of the Euclidean norm in $\R^{2n}$ (see [@GS p.192]), we get $$\left(\|x\|_2^{-2}\right)^\wedge(\theta)=C(n) \|\theta\|_2^{ -2n+2},$$ where $C(n) = 2^{2n-3}\pi^n \Gamma(n-1).$ By the connection between linear transformations and the Fourier transform, $$\label{linear transform} \left( \\|Tx\\|^{-1}\right)^\wedge(y)=\|\det T\|^{-1}\left( \\|x\\|^{-1} \right)^\wedge((T^{})^{-1} y)$$ one can easily compute the following: \[Lem:FTellipsoid\] For all $\theta\in S^{2n-1}$, $$\left( \\|x\\|^{-2}_{E_{a,b}(\xi)}\right)_x^\wedge(\theta) = \frac{C(n)}{a^{2n-4}} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2}.$$ By (\[linear transform\]) with $T$ being the composition of a rotation and a diagonal operator, $$\begin{aligned} \left( \\|x\\|^{-2}_{E_{a,b} (\xi)}\right)_x^\wedge(\theta) = C(n) {a^2b^{2n-2}} \\|\theta\\|_{E_{1/a,1/b}(\xi)}^{-2n+2}=\frac{C(n)}{a^{2n-4}} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2}.\end{aligned}$$ \[Step1\] Let $K$ be an origin symmetric complex star body, then the function $ \\|\xi\\|_K^{-2}$ can be approximated in the space $C_c(S^{2n-1})$ by functions of the form $$\label{eqn:delta-seq} f_{a,b}(\xi)=\frac{C(n)}{a^{2n-4}} \int_{S^{n-1}} \\|\theta\\|_K^{-2} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2} d\theta,$$ as $a\to 0$ and $b$ is chosen appropriately. Using Parseval’s formula (\[parseval\]) and the previous Lemma we get $$\frac{C(n)}{ a^{2n-4} } \int_{S^{2n-1}} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2} d\theta$$ $$= \frac{ 1}{ a^{2n-4} } \int_{S^{2n-1}} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2} \left(\|x\|_2^{-2}\right)^\wedge(\theta) d\theta$$ $$= \frac{ 1}{ a^{2n-4} } \int_{S^{2n-1}} \left(\\|x\\|_{E_{b,a}(\xi)}^{-2n+2}\right)^\wedge(\theta) \|\theta\|_2^{-2}d\theta = \frac{ 1}{C(n) } \int_{S^{n-1}} \\|x\\|_{E_{a,b}(\xi)}^{-2} dx$$ $$= \frac1{C(n)}\int_{S^{2n-1}} \left(\frac{( x, \xi )^2+ (x, \xi^\bot)^2}{a^2}+ \frac{1-( x, \xi )^2- (x,\xi^\bot)^2}{b^2}\right)^{-1} dx.$$ For every fixed $a,$ the latter integral goes to infinity when $b\to \infty,$ and it goes to zero when $b\to 0,$ so for every $a>0$ there exists $b=b(a)$ such that $$\frac{ C(n)}{ a^{2n-4} } \int_{S^{2n-1}} \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta =1.$$ Note that, by rotation invariance, the value of $b(a)$ does not depend on the choice of $\xi.$ Now for every $\xi\in S^{2n-1}$ we have $$\left\|\\|\xi\\|_K^{-2} - \frac{ C(n)}{ a^{2n-4} } \int_{S^{2n-1}} \\|\theta\\|_K^{-2} \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta\right\|$$ $$\le \frac{ C(n)}{ a^{2n-4} }\int_{S^{2n-1}}\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\| \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta$$ $$= \frac{ C(n)}{ a^{2n-4} } \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2\ge\delta}\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\| \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta$$ $$+ \frac{ C(n)}{ a^{2n-4} } \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2<\delta}\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\| \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta =I_1+I_2.$$ for every $\delta\in (0,1).$ Since $K$ is a complex star body, the norm of $K$ is constant on vectors of the form $u\xi+v\xi^\bot$ with $u^2+v^2=1.$ Vectors of this form are the only solutions on the sphere of the equation $(\theta,\xi)^2+(\theta,\xi^\bot)^2=1.$ Using this and the uniform continuity of $ \\|x\\|_K^{-2}$ on the sphere $S^{2n-1}$, for any given $\epsilon>0$ we can find $\delta\in(0,1)$, close to $1$, so that $(\theta,\xi)^2+(\theta,\xi^\bot)^2\ge\delta$ implies $\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\|<\epsilon/2$. Therefore $$I_1= \frac{ C(n)}{ a^{2n-4} } \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2\ge\delta}\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\| \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta$$ $$\le \frac{\epsilon}{2} \left[\frac{ C(n)}{ a^{2n-4} } \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2\ge\delta} \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta \right]\le \frac{\epsilon}{2}.$$ Now fix $\delta$ chosen above and estimate the integral $I_2$ as follows $$I_2=\frac{ C(n)}{ a^{2n-4} } \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2<\delta}\Big\| \\|\xi\\|_K^{-2} - \\|\theta\\|_K^{-2} \Big\| \\|\theta\\|_{E_{b(a),a}(\xi)}^{-2n+2} d\theta$$ $$\le \frac{ C_1(n,K)}{a^{2n-4}} \int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2<\delta} \\|\theta\\|_{E_{b(a),a}(x)}^{-2n+2} d\theta =\frac{ C_1(n,K)}{a^{2n-4}}\times$$$$\int_{(\theta,\xi)^2+(\theta,\xi^\bot)^2<\delta} \left(\frac{( \theta, \xi )^2+ (\theta, \xi^\bot)^2}{(b(a))^2}+ \frac{1-( \theta, \xi )^2- (\theta,\xi^\bot)^2}{a^2}\right)^{-n+1} d\theta$$ $$\le a^2 (1-\delta)^{-n+1}C_1(n,K) \|S^{2n-1}\| .$$ where $$C_1(n,K)={2 C(n) \max_{x\in S^{2n-1}} \\|x\\|_K^{-2} } .$$ Now we can choose $a$ so small that $I_2\le \epsilon/2$. \[Step2\] If $\mu$ is a finite measure on $S^{2n-1}$ and $a,b>0$, then the function $$f(\xi)=\int_{S^{n-1}} \\|\theta \\|^{-2}_{E_{a,b}(\xi)} d\mu(\theta)$$ can be approximated in $C_c(S^{2n-1})$ by the sums of the form $$\sum_{i=1}^m \\|\xi\\|_{E_i}^{-2},$$ where $E_1$,...,$E_m$ are complex ellipsoids. Let $\sigma>0$ be a small number and choose a finite covering of the sphere by spherical $\sigma$-balls $B_\sigma (\eta_i)=\{\eta\in S^{n-1}: \|\eta-\eta_i\|<\sigma\}$, $\eta_i\in S^{2n-1}$, $i=1,\dots, m=m(\delta)$. Define $$\widetilde{B}_\sigma(\xi_1)=B_\sigma (\xi_1)$$ and $$\widetilde{B}_\sigma(\xi_i)=B_\sigma (\xi_i)\setminus\bigcup_{j=1}^{i-1}B_\sigma(\xi_j), \quad\mbox{ for } i=2,...,m.$$ Let $1/p_i=\mu(\widetilde{B}_\sigma(\xi_i))$. Clearly, $1/p_1+\cdots+1/p_m=\mu(S^{2n-1})$. Let $\rho(E_{a,b}(\xi),x)$ be the value of the radial function of the ellipsoid $E_{a,b}(\xi)$ at the point $x$, that is $$\rho(E_{a,b}(\xi),x)=\\|x\\|_{E_{a,b}(\xi)}^{-1}.$$ Note that $\rho(E_{a,b}(\xi),x)=\rho(E_{a,b}(x),\xi),$ because both depend only on the modulus of the complex scalar product of $x$ and $\xi$, therefore $$\|\rho^2(E_{a,b}(\xi),x)- \rho^2(E_{a,b}(\theta),x)\|\le C_{a,b} \|\xi - \theta\|,$$ with a constant $C_{a,b}$ that depends only on $a$ and $b$. Then, $$\left\|\int_{S^{n-1}} \rho^2(E_{a,b}(\xi),x) d\mu(\xi)-\sum_{i=1}^m\frac{1}{p_i} \rho^2(E_{a,b}(\xi_i),x)\right\|$$ $$=\left\|\sum_{i=1}^m\left(\int_{\widetilde{B}_\sigma(\xi_i)} \rho^2(E_{a,b}(\xi),x) d\mu(\xi)-\int_{\widetilde{B}_\sigma(\xi_i)} \rho^2(E_{a,b}(\xi_i),x) d\mu(\xi)\right)\right\|$$ $$\le \sum_{i=1}^m \int_{\widetilde{B}_\sigma(\xi_i)} \left\| { \rho^2(E_{a,b}(\xi),x)}- {\rho^2(E_{a,b}(\xi_i),x)} \right\|d\mu(\xi)$$ $$\le \sum_{i=1}^m \int_{\widetilde{B}_\sigma(\xi_i)} C_{a,b} \|\xi-\xi_i\| d\mu(\xi) \le C_{a,b}\ \sigma \mu(S^{2n-1}).$$ Since $\sigma$ is arbitrarily small, the result follows after we define ellipsoids $E_i$ by $$\\|x\\|_{E_i}^{-2} = \frac{1}{p_i} \rho^2(E_{a,b}(\xi_i),x). \qed$$ [Proof of Theorem \[Thm:approximation\].]{} The “if" part immediately follows from Theorem \[posdef\], since for any ellipsoid $E$ the distribution $\\|\cdot\\|_E^{-2}$ is positive definite, as the linear perturbation of the the same function for the Euclidean ball. To prove the converse, suppose that $K$ is a complex intersection body and $\mu$ is the measure on $S^{2n-1}$ corresponding to $K$ by the definition of complex intersection body. By Lemma \[Step1\], $ \\|\xi\\|_K^{-2}$ can be uniformly approximated by the integrals of the form $$\label{eq8} \frac{C(n)}{ a^{2n-4} } \int_{S^{2n-1}} \\|\theta\\|_K^{-2} \\|\theta\\|_{E_{b,a}(\xi)}^{-2n+2} d\theta,$$ as $a\to 0$. By Lemma \[Lem:FTellipsoid\] and Lemma \[r-s\], $$\frac{C(n)}{ a^{2n-4} }{\cal{R}}_c(\\|\cdot\\|_{E_{b,a}(\xi)}^{-2n+2}) = (2\pi)^{2n-1}\\|\cdot\\|_{E_{a,b}(\xi)}^{-2},$$ and by the definition of complex intersection body, (\[eq8\]) is equal to $$(2\pi)^{2n-1} \int_{S^{2n-1}} \\|\theta \\|_{E_{a,b}(\xi)}^{-2} d\mu(\theta).$$ Now, by Lemma \[Step2\], $\int_{S^{2n-1}} \\|\theta \\|_{E_{a,b}(\xi)}^{-2} d\mu(\theta)$ can be uniformly approximated by sums of the form $\sum_{i=1}^m \\|\xi\\|_{E_i}^{-1},$ where $E_i$ are complex ellipsoids. Stability in the Busemann-Petty problem and hyperplane inequalities {#stab-hyper} =================================================================== Intersection bodies played an important role in the solution of the Busemann-Petty problem posed in [@BP] in 1956. Suppose that $K$ and $L$ are origin symmetric convex bodies in $\R^n$ so that, for every $\xi\in S^{n-1},$ $$\|K\cap \xi^\bot\|\le \|L\cap \xi^\bot\|.$$ Does it follow that $\|K\|\le \|L\|?$ The problem was completely solved at the end of 1990’s, and the answer is affirmative if $n\le 4$ and negative if $n\ge 5.$ The solution appeared as the result of a sequence of papers [@LR], [@Ba], [@Gi], [@Bo4], [@L], [@Pa], [@G1], [@G2], [@Z1], [@Z2], [@K5], [@K9], [@Z4], [@GKS] (see [@G Chapter 8] or [@K1 Chapter 5] for details). One of the main ingredients of the solution was a connection between intersection bodies and the Busemann-Petty problem established by Lutwak [@L]: if $K$ is an intersection body then the answer to the Busemann-Petty problem is affirmative for any star body $L.$ On the other hand, if $L$ is a symmetric convex body that is not an intersection body then one can construct $K$ giving together with $L$ a counterexample. The complex Busemann-Petty problem can be formulated as follows. Suppose that $K,L$ are origin symmetric complex convex bodies in $\R^{2n}$ and, for every $\xi \in S^{2n-1},$ we have $\|K\cap H_\xi\|\le \|L\cap H_\xi\|.$ Does it follow that $\|K\|\le \|L\|?$ As proved in [@KKZ], the answer is affirmative if $n\le 3,$ and it is negative if $n\ge 4.$ The proof is based on a connection with intersection bodies, similar to Lutwak’s connection in the real case (see [@KKZ Theorem 2]): \(i) If $K$ is a complex intersection body in $\R^{2n}$ and $L$ is any origin symmetric complex star body in $\R^{2n},$ then the answer to the question of the complex Busemann-Petty problem is affirmative; \(ii) if there exists an origin symmetric complex convex body in $\R^{2n}$ that is not a complex intersection body , then one can construct a counterexample to the complex Busemann-Petty problem. These connections were formulated in [@KKZ] in terms of positive definite distributions, so one has to use Theorem \[posdef\] to get the statements in terms of convex intersection bodies. Zvavitch [@Zv] found a generalization of the Busemann-Petty problem to arbitrary measures, namely, one can replace volume by any measure $\gamma$ with even continuous density in $\R^n.$ In particular, if $n\le 4,$ then for any origin symmetric convex bodies $K$ and $L$ in $\R^n$ the inequalities $$\gamma(K\cap \xi^\bot) \le \gamma(L\cap \xi^\bot), \qquad \forall \xi\in S^{n-1}$$ imply $$\gamma(K)\le \gamma(L).$$ Zvavitch also proved that this is generally not true if $n\ge 5,$ namely, for any $\gamma$ with strictly positive even continuous density there exist $K$ and $L$ providing a counterexample. In [@Zy] the result of Zvavitch was extended to complex convex bodies. In this section we are going to prove stability in the affirmative part of the result from [@Zy]; see Theorem \[stab-complex\] below. Note that stability in the original Busemann-Petty problem was established in [@K7], and for the complex Busemann-Petty problem it was done in [@K6]. Stability in Zvavitch’s result was proved in [@K8], and in [@KM] the result of [@K8] was extended to sections of lower dimensions in place of hyperplane sections. Let $f$ be an even continuous non-negative function on $\R^{2n},$ and denote by $\gamma$ the measure on $\R^{2n}$ with density $f$ so that for every closed bounded set $B\subset \R^n$ $$\gamma(B)=\int_B f(x)\ dx.$$ Since we apply $\gamma$ only to complex star bodies, we can assume without loss of generality that the measure $\gamma$ and the function $f$ are $R_\theta$-invariant. We need a polar formula for the measure of a complex star body $K:$ $$\label{meas-polar} \gamma(K) = \int_K f(x)\ dx = \int_{S^{n-1}} \left(\int_0^{\\|\theta\\|_K^{-1}} r^{n-1}f(r\theta)\ dr \right) d\theta.$$ For every $\xi\in S^{2n-1},$ $$\gamma(K\cap H_\xi) = \int_{K\cap H_\xi} f(x) dx$$ $$= \int_{S^{2n-1}\cap H_\xi} \left(\int_0^{\\|\theta\\|_K^{-1}}r^{2n-3} f(r\theta) dr \right)d\theta$$ $$\label{meas-sect-polar} = {\cal{R}}_c \left(\int_0^{\\|\cdot\\|_K^{-1}} r^{2n-3} f(r\cdot)\ dr \right)(\xi),$$ We need the following elementary lemma, which was also used by Zvavitch in [@Zv]. \[zvavitch\] Let $a,b \in [0,\infty),\ n\in \N, n\ge 2$, and let $g$ be a non-negative integrable function on $[0,\max\{a,b\}].$ Then $$\int_0^a r^{2n-1} g(r)\ dt-a^2\int_0^ar^{2n-3} g(r)\ dr$$ $$\le \int_0^b r^{2n-1} g(r)\ dr - a^2\int_0^br^{2n-3} g(r)\ dr.$$ Denote by $$d_n= \frac{\|B_2^{2n}\|^{\frac{n-1}n}}{\|B_2^{2n-2}\|},$$ where $B_2^n$ stands for the unit Euclidean ball in $\R^n,$ and $$\|B_2^n\|=\frac{\pi^{n/2}}{\Gamma(1+\frac n2)}.$$ Note that $d_n<1$ for every $n\in \N;$ this easily follows from the log-convexity of the $\Gamma$-function. It is well-known (see for example [@K1 p.32]) that the surface area of the sphere $S^{n-1}$ in $\R^n$ is equal to $$\label{ball} \|S^{n-1}\| = n \|B_2^n\|.$$ The following stability result extends [@K8 Theorem 2] to the complex case. The proof is similar to that of [@KM Theorem 2]. \[stab-complex\] Let $K$ and $L$ be origin symmetric complex star bodies in $\R^{2n},$ let $\e>0$ and let $\gamma$ be a measure on $\R^{2n}$with even continuous non-negative density $f.$ Suppose that $K$ is a complex intersection body, and that for every $\xi\in S^{2n-1}$ $$\label{stab-ineq} \gamma(K\cap H_\xi) \le \gamma(L\cap H_\xi) + \e.$$ Then $$\gamma(K)\le \gamma(L) + \frac n{n-1} d_n \e \|K\|^{\frac 1n}.$$ By (\[meas-sect-polar\]), the condition (\[stab-ineq\]) can be written as $${\cal{R}}_c \left(\int_0^{\\|\cdot\\|_K^{-1}} r^{2n-3} f(r\cdot)\ dr \right)(\xi)$$ $$\le {\cal{R}}_c \left(\int_0^{\\|\cdot\\|_L^{-1}} r^{2n-3} f(r\cdot)\ dr \right)(\xi) + \e, \qquad \forall \xi\in S^{2n-1}.$$ Integrating the latter inequality with respect to the measure $\mu$ on $S^{2n-1},$ corresponding to the body $K$ by Definition \[int\], and then using the equality of Definition \[int\], we get $$\label{eq6} \int_{S^{n-1}} \\|\theta\\|_K^{-2} \left(\int_0^{\\|\theta\\|_K^{-1}} r^{2n-3} f(r\theta)\ dr\right) d\theta$$ $$\le \int_{S^{2n-1}} \\|\theta\\|_L^{-2} \left(\int_0^{\\|\theta\\|_L^{-1}} r^{2n-3} f(r\theta)\ dr\right) d\theta + \e \int_{S^{2n-1}} d\mu(\xi).$$ Applying Lemma \[zvavitch\] with $a=\\|\theta\\|_K^{-1}$, $b=\\|\theta\\|_L^{-1}$ and $g(r)=f(r\theta)$ and then integrating over the sphere, we get $$\int_{S^{2n-1}} \left(\int_0^{\\|\theta\\|_K^{-1}} r^{2n-1} f(r\theta)\ dr \right) d\theta$$ $$- \int_{S^{2n-1}} \\|\theta\\|_K^{-k} \left(\int_0^{\\|\theta\\|_K^{-1}} r^{2n-3} f(r\theta)\ dr \right) d\theta$$ $$\le \int_{S^{2n-1}} \left(\int_0^{\\|\theta\\|_L^{-1}} r^{2n-1} f(r\theta)\ dr \right) d\theta$$ $$\label{eq7} - \int_{S^{2n-1}} \\|\theta\\|_K^{-k} \left(\int_0^{\\|\theta\\|_L^{-1}} r^{2n-3} f(r\theta)\ dr \right) d\theta.$$ Adding (\[eq6\]) and (\[eq7\]) and using (\[meas-polar\]) we get $$\gamma(K)\leq\gamma(L)+\e \int_{S^{2n-1}} d\mu(\xi).$$ Since ${\cal{R}}_c 1 = \|S^{2n-3}\| 1,$ where $1(\xi)\equiv 1,$ we again apply Definition \[int\], Hölder’s inequality, the polar formula for volume and (\[ball\]): $$\int_{S^{2n-1}} d\mu(\xi) = \frac 1{\|S^{2n-3}\|} \int_{S^{2n-1}} {\cal{R}}_c1(\xi) d\mu(\xi)$$ $$= \frac 1{\|S^{2n-3}\|} \int_{S^{2n-1}} \\|x\\|_K^{-2} dx \le \frac 1{\|S^{2n-3}\|} \left(\int_{S^{2n-1}} \\|x\\|_K^{-2n}dx \right)^{\frac 1n} \|S^{2n-1}\|^{\frac{n-1}n}$$ $$= \frac{(2n)^{\frac 1n}\|S^{2n-1}\|^{\frac{n-1}n}}{\|S^{2n-3}\|} \|K\|^{\frac 1n} = \frac n{n-1} d_n \|K\|^{\frac1n}.$$ Interchanging $K$ and $L$ in Theorem \[stab-complex\], we get a complex version of [@K8 Corollary 1]. If $K$ and $L$ are complex intersection bodies in $\R^{2n},$ then $$\|\gamma(K) - \gamma(L)\|$$ $$\le \frac n{n-1} d_n \max_{\xi\in S^{2n-1}} \|\gamma(K\cap H_\xi)-\gamma(L\cap H_\xi)\| \max \left\{\|K\|^{\frac1n}, \|L\|^{\frac 1n} \right\}.$$ Putting $L=\emptyset$ in the latter inequality, we extend to the complex case the hyperplane inequality for real intersection bodies from [@K8 Theorem 1]. \[hyper-ineq\] If $K$ is a complex intersection body in $\R^{2n}$, and $\gamma$ is an arbitrary measure on $\R^{2n}$ with even continuous density, then $$\gamma(K) \le \frac n{n-1} d_n \max_{\xi\in S^{2n-1}} \gamma(K\cap H_\xi)\ \|K\|^{\frac1n}.$$ By Corollary \[n<4\], this inequality holds for any origin symmetric complex convex body $K$ in $\R^4$ or $\R^6.$ The constant in Theorem \[hyper-ineq\] is optimal, as can be easily seen from the same example as in [@K8]. Let $K=B_2^n$ and, for every $j\in N,$ let $f_j$ be a non-negative continuous function on $[0,1]$ supported in $(1-\frac 1j,1)$ and such that $\int_0^1 f_j(t) dt =1.$ Let $\gamma_j$ be the measure on $\R^{2n}$ with density $f_j(\|x\|_2),$ where $\|x\|_2$ is the Euclidean norm in $\R^{2n}.$ Then a simple computation shows that $$\lim_{j\to \infty} \frac{\gamma_j(B_2^{2n})}{\max_{\xi\in S^{2n-1}} \gamma_j(B_2^{2n}\cap H_\xi)\ \|B_2^{2n}\|^{1/n}} = \frac n{n-1}d_n.$$ Note that in the case of volume (when the density $f\equiv 1$) the inequality of Theorem \[hyper-ineq\] follows from [@K6 Corollary 1] and the constant is just $d_n$ without the term $n/(n-1).$ One has to follow the proof of Theorem 1 from [@K6] to restore the constant $d_n$ which is estimated by 1 everywhere in [@K6]. The result of Theorem \[hyper-ineq\] is related to the famous hyperplane problem asking whether there exists an absolute constant $C$ so that for any origin symmetric convex body $K$ in $\R^n$ $$\label{hyper} \vol_n(K)^{\frac {n-1}n} \le C \max_{\xi \in S^{n-1}} \vol_{n-1}(K\cap \xi^\bot),$$ where $\xi^\bot$ is the central hyperplane in $\R^n$ perpendicular to $\xi.$ The problem is still open, with the best-to-date estimate $C\sim n^{1/4}$ established by Klartag [@Kl], who slightly improved the previous estimate of Bourgain [@Bo3]. Complex intersection bodies of convex complex bodies {#bus} ==================================================== In this section we extend two classical results about intersection bodies of convex bodies to the complex setting. The well-known result of Busemann [@Bu1] is that the intersection body of a symmetric convex body is also symmetric convex. We prove a complex version of this result. \[CBusemann\] Let $K$ be an origin symmetric convex body in $\C^n$ and $ I_{c}(K)$ the complex intersection body of $K.$ Then $ I_{c}(K)$ is also an origin symmetric convex body in $\C^{n}$. Before we prove Theorem \[CBusemann\] we need some preparations. We write $O_{2}$ for the set or all rotations in $\R^{2}$ as described in and $H_{0}$ for the (real) $2$-dimensional subspace of $\R^{2n}$, spanned by the standard unit vectors $e_{1}, e_{2}$. If $V_{H_{0}}$ is any orthogonal transformation in $H_{0}$ ($V_{H_{0}} \in O_{2}$), we define $U_{V_{H_0}}\in O_{2n}$ to be the orthogonal transformation in $\R^{2n}$ that has (as a matrix) $n$ copies of $V_{H_{0}}$ on its diagonal. Then implies that $ U_{V_{H_0}} K = K$. We will use this property in the following form: $$\label{characterization} {\bf 1}_{K} ( U_{V_{H_0}}x)= {\bf 1}_{K} (x) , \ \forall x\in \R^{2n} , \ \forall \ V_{H_0} \in O_2.$$ Assume that $n\geq 3$. Let $u_{1}, u_{2} \in \C^n, \ \|u_1\|_{2}=\|u_2\|_{2}=1,$ with $H_{u_i}^{\perp}= {\rm span} \{u_i, \ u_i^{\perp}\},$ as introduced earlier, and $\theta_i\in S_{H_{u_i}^{\perp}}, \ i=1, 2.$ We define $u_{3}:= \frac{ u_{1}+u_{2}}{ \|u_{1}+ u_{2}\|_{2}} ,$ with $H^{\perp}_{u_3}$ and $\theta_{3}:= \frac{ \theta_{1}+ \theta_{2}}{ \|\theta_{1}+ \theta_{2}\|_{2}}\in S_{H^{\perp}_{u_3}}$ such that $\|\theta_{1}+ \theta_{2}\|_{2}=\|u_{1}+ u_{2}\|_{2}$. We can assume that $H_{u_{1}}^{\perp}\cap H_{u_{2}}^{\perp}=\{0\}$. Now, let $r_{1}, r_{2}>0$. We define $r_{3}, t$ (as functions of $r_{1}, r_{2}$) such that $$\label{r,t} t:= \frac{r_{1}}{ r_{1}+ r_{2}} , \ r_{3} := \frac{1}{ \frac{1}{r_{1}} + \frac{1}{r_{2}} }\| u_{1}+ u_{2}\|_{2} .$$ If $S:=(H_{u_1}^{\perp} + H_{u_2}^{\perp})^{\perp},$ we write $ E_{i}:= {\rm span}\{ H_{u_{i}}^{\perp}, S\}, \ i=1,2,3$. We define the functions $g_{i}:H_{u_i}^{\perp} \rightarrow \R$, $h_{i}:[0,\infty)\rightarrow [0,\infty), \ i=1,2,3$ to be $$\label{g,h} g_{i}(x):= \int_{S+x} {\bf 1}_{K}(y) d y = \|K \cap ( S+ x)\| , \ h_{i}(r):= g_{i}(r\theta_{i} ).$$ In the following we exploit the fact that $K$ satisfies . \[l2.1\] For $i=1,2,3$, we have that $$\label{eqKEi}\| K\cap E_{i}\| = 2\pi \int_{0}^{\infty} r h_{i}(r) d r .$$ First we will show that the functions $g_{i}$ are rotation invariant. Let $S_{0}:= {\rm span}\{ e_{5}, \cdots, e_{2n}\}$ and $\theta_{i}^{(1)}, \theta_{i}^{(2)} \in S_{H^{\perp}_{U_i}} \ i=1,2,3.$ There exists $ U_{i}\in O_{2n}$ such that $ U_{i}H^{\perp}_{u_i} = H_{0}$ and $ U_{i}S= S_{0},$ and $\phi_{i}^{(1)}, \phi_{i}^{(2)} \in S_{H_0},$ such that $U^{\top}_i (\phi_{i}^{(1)})=\theta_{i}^{(1)}$ and $U^{\top}_i (\phi_{i}^{(2)})=\theta_{i}^{(2)}, \ i=1,2,3.$ Moreover, there exists $V\in O_2$ such that $V(\phi_i^{(2)})=\phi_i^{(1)}.$ Let $V_0\in O_{2n}$ be the diagonal operator with $V$ on its diagonal entries. Then, it is clear that $V_0(\phi_{i}^{(2)})=\phi_{i}^{(1)}$ and $V_0S_0=S_0.$ Then, by , we have that for every $r>0,$ $$g_{i}(r\theta_{i}^{(1)}) = \int_{S+r\theta_{i}^{(1)}} {\bf 1}_{K}(y) d y = \int_{U_{i}^{\top} S_{0} + r U_{i}^{\top} (\phi_{i}^{(1)}) } {\bf 1}_{K} (y ) d y$$ $$= \int_{U_{i}^{\top} (S_{0} + r \phi_{i}^{(1)} )} {\bf 1}_{K} (y ) d y = \int_{S_{0} + r\phi_{i}^{(1)} } {\bf 1}_{K} ( U_{i} y ) d y$$ $$= \int_{V_{0} S_{0} + rV_{0}(\phi_{i}^{(2)}) } {\bf 1}_{K} ( U_{i} y ) d y = \int_{V_{0} (S_{0} + r\phi_{i}^{(2)}) } {\bf 1}_{K} ( U_{i} y ) d y$$ $$= \int_{ S_{0} + r \phi_{i}^{(2)} } {\bf 1}_{K} ( V^{\top}_{0} U_{i} y ) d y = \int_{ S_{0} + r\phi_{i}^{(2)} } {\bf 1}_{K} ( U_{i} y ) d y$$ $$= \int_{ U_{i}^{\top} S_{0} + r U_{i}^{\top}( \phi_{i}^{(2)}) } {\bf 1}_{K} ( y ) d y = \int_{ S + r \theta_{i}^{(2)} } {\bf 1}_{K} ( y ) d y = g_{i}(r\theta_{i}^{(2)}) .$$ So, by Fubini’s theorem we have that for $i=1,2,3$, $$\|K\cap E_{i}\|= \int_{H_{u_{i}}^{\perp}} \int_{S+ x} {\bf 1}_{K}(y) d y dx = \int_{H_{u_{i}}^{\perp}} g_{i}(x) d x$$ $$= \int_{S_{H_{u_{i}}^{\perp}}} \int_{0}^{\infty} r g_{i}(r\theta) d r d\theta = 2\pi \int_{0}^{\infty} r g_{i}(r\theta_{i}) d r = 2\pi \int_{0}^{\infty} r h_{i}(r ) d r.$$ This finishes the proof. $\hfill\Box $ The convexity of $K$ is exploited in the following [\[l2.2\]]{} With the above notation we have that $$\label{2.6} h_{3}(r_{3}) \geq h_{1}^{(1-t)}(r_{1})h_{2}^{t}(r_{2}) .$$ Note that $ h_{i}(r):= \| K \cap ( S + r\theta_{i})\|$, for $i=1,2,3$. Also observe that the sets $ K\cap (S + r_{i} \theta_{i})$ all lie in the same hyperplane, and by convexity, $$(1-t) (K\cap (S+ r_{1}\theta_{1}) ) + t (K\cap (S+ r_{2}\theta_{2}) )\subseteq (K\cap (S+ r_{3}\theta_{3}) ) .$$ The result follows from the Brunn-Minkowski inequality. $\hfill\Box $ Finally we need the following result of K. Ball [@Ball1]. It can be seen as an extension of the inequality of Busemann (see also [@BFranz]). In [@Ball1] this proposition has been proved but has not been stated in this form. In this form it can be found in [@Kl]. [\[pr2.3\]]{} Let $r_{1}, r_{2}>0$. Define $t, r_{3}$ as follows: $$\label{2.7} t:= \frac{r_{1}}{ r_{1}+ r_{2}} , \ r_{3} := \frac{2}{ \frac{1}{r_{1}} + \frac{1}{r_{2}} } .$$ Assume that $h_{1}, h_{2}, h_{3}:[0,\infty)\rightarrow [0,\infty)$ such that $$\label{2.8} h_{3}(r_{3}) \geq h_{1}^{(1-t)}(r_{1})h_{2}^{t}(r_{2}) , \ \forall r_{1}, r_{2}>0 .$$ Let $p\geq 1$ and denote $$A:= \left(\int_{0}^{\infty} r^{p-1} h_{1}(r) d r\right)^{\frac{1}{p}} ,$$ $$B:=\left( \int_{0}^{\infty} r^{p-1} h_{2}(r) d r\right)^{\frac{1}{p}} ,$$ $$C:= \left(\int_{0}^{\infty} r^{p-1} h_{3}(r) d r\right)^{\frac{1}{p}} .$$ Then, $$\label{2.10} C\geq \frac{2}{ \frac{1}{A}+ \frac{1}{B}} .$$ We rewrite the previous proposition in a form that fits our setting: [\[co2.4\]]{} Let $r_{1}, r_{2}>0$ and let $\alpha>0$. Define $t, r_{3}$ as follows: $$\label{2.11} t:= \frac{r_{1}}{ r_{1}+ r_{2}} , \ r_{3} := \frac{\alpha}{ \frac{1}{r_{1}} + \frac{1}{r_{2}} } .$$ Assume that $h_{1}, h_{2}, h_{3}:[0,\infty)\rightarrow [0,\infty)$ such that $$\label{2.12} h_{3}(r_{3}) \geq h_{1}^{(1-t)}(r_{1})h_{2}^{t}(r_{2}) , \ \forall r_{1}, r_{2}>0 .$$ Let $p\geq 1$ and denote $$A:= \left(\int_{0}^{\infty} r^{p-1} h_{1}(r) d r\right)^{\frac{1}{p}} ,$$ $$B:= \left(\int_{0}^{\infty} r^{p-1} h_{2}(r) d r\right)^{\frac{1}{p}} ,$$ $$C:= \left(\int_{0}^{\infty} r^{p-1} h_{3}(r) d r\right)^{\frac{1}{p}} .$$ Then, $$\label{2.14} C\geq \frac{\alpha}{ \frac{1}{A}+ \frac{1}{B}} .$$ Let $H_{3}:[0,\infty)\rightarrow [0,\infty)$ be such that $ H_{3}( r) = h_{3}(\frac{\alpha}{2} r),$ and let $ C^{\prime}:= \int_{0}^{\infty} r^{p-1} H_{3}(r) d r$. Let $r_{3}^{\prime}:= \frac{2}{\alpha} r_{3}$. Then $r_{1}, r_{2}, r_{3}^{\prime}$ satisfy ([\[2.7\]]{}) and $$H_{3}(r_{3}^{\prime}) = h_{3}(r_{3}) \geq h_{1}^{(1-t)}(r_{1})h_{2}^{t}(r_{2}) ,$$ so $h_{1}, h_{2}, H_{3}$ satisfy also ([\[2.8\]]{}). So by Proposition \[pr2.3\] we have that $$\frac{2}{ \frac{1}{A}+ \frac{1}{B}} \leq C^{\prime} = \left(\int_{0}^{\infty} r^{p-1} h_{3}\left(\frac{\alpha}{2}r\right) d r\right)^{\frac{1}{p}} = \frac{2}{\alpha} \left( \int_{0}^{\infty} s^{p-1} h_{3}(s) d s\right)^{\frac{1}{p}}$$ $${\rm or} \ \ \ \ C\geq \frac{\alpha}{ \frac{1}{A}+ \frac{1}{B}} .$$ This completes the proof. $\hfill\Box $ [\[pr2.5\]]{} In the notation introduced above, if $\alpha := \| \theta_{1} + \theta_{2}\|_{2} $ then $$\label{2.15} \frac{\alpha}{\|K\cap E_{3}\|^{\frac{1}{2}}} \leq \frac{1}{\|K\cap E_{1}\|^{\frac{1}{2}}} + \frac{1}{\|K\cap E_{2}\|^{\frac{1}{2}}} .$$ By Lemma \[l2.2\] we have that $h_{1}, h_{2}, h_{3}, r_{1}, r_{2}, r_{3},t$ satisfy (\[2.11\]) and (\[2.12\]). By Corollary \[co2.4\], applied to $p=2,$ we have that for $A,B,C$ as in Corollary \[co2.4\], $$\label{2.16} \frac{\alpha}{C}\leq \frac{1}{A}+ \frac{1}{B} .$$ Note that by Lemma [\[l2.1\]]{}, $$A^{2}:= \int_{0}^{\infty} r h_{1}(r) d r = \frac{\|K\cap E_{1}\|}{2\pi}, \ B^{2}:= \int_{0}^{\infty} r h_{2}(r) d r= \frac{\|K\cap E_{2}\|}{2\pi} \ {\rm and}$$ $$\label{2.17} C^{2}:= \int_{0}^{\infty} r^{p-1} h_{3}(r) d r= \frac{\|K\cap E_{3}\|}{2\pi} .$$ By (\[2.16\]) and (\[2.17\]), we complete the proof. $\hfill\Box $ [\[co2.6\]]{} Let $K$ be a symmetric complex convex body in $\R^n$. Let $H$ be an $(n-2)$-dimensional subspace of $\C^{n}$. Let $u\in H^{\perp}$ complex unit vector and let $H_{u}:= {\rm span}\{ H, u\}$ and $ r(u):= \| K \cap H_{u}\|^{\frac{1}{2}}$. Then $r:H^{\perp} \cap S^{2n-1}\rightarrow (0,\infty)$ is the boundary of a complex convex body in $H^{\perp}$. In order to show that the curve $r$ is the boundary of a convex body in $\C^{n}$ it is enough to show that $r^{-1}$ is the restriction of a norm to $H^{\perp}$. So if $u_{1}, u_{2}$ are two non-parallel unit vectors in $H^{\perp},$ let $u_{3}:= \frac{u_{1}+u_{2}}{ \|u_{1}+ u_{2}\|_{2}}=: \frac{1}{\alpha} (u_{1}+ u_{2})$. It is enough to show that $$\frac{\alpha}{r(u_{3})} \leq \frac{1}{r(u_{1})} +\frac{1}{r(u_{2})} .$$ In the notation of this section we have that $ H_{u_{i}}=E_{i}$ and $ r(u_{i}) = \|K \cap E_{i}\|^{\frac{1}{2}}$. The result follows from Proposition [\[pr2.5\]]{}. $\hfill\Box $ [Proof of Theorem [\[CBusemann\]]{}:]{} In the case where $n=2$ the body $I_{c}(K)$ is simply a rotation of $K$, so the result is obvious. Let $n\geq 3$. Then Corollary \[co2.6\] implies that $I_{c}(K)\cap H^{\perp}$ is convex for every $(n-2)$-dimensional subspace $H$ of $\C^{n}$. This implies that $ I_{c}(K)$ is convex. The symmetry of $I_{c}(K)$ is obvious from the definition. Finally, it is not difficult to see that $I_{c}(K)$ satisfies (\[rotation\]). This implies that $I_{c}(K)$ is a complex convex body. $\hfill\Box $ In the case where $K\subset \R^n$ is convex, by results of Hensley [@H] and Borell [@Bor], one has that the intersection body of $K$, $I(K)$, is isomorphic to an ellipsoid; i.e. $ d_{BM}(I(K), B_{2}^{n})\leq c$ where $d_{BM}$ stands for the Banach-Mazur distance and $c>0$ is a universal constant. Recall that the Banach-Mazur distance of two symmetric convex bodies $K_{1}, K_2$ in $\R^n$ as $$d_{BM}(K_{1}, K_{2}) := \inf_{T\in GL_{n}} \inf\{ a>0 : K_{1}\subseteq T K_{2} \subseteq a K_{1}\} .$$ One can show that the same result holds also in the complex case by using a result of K. Ball [[@Ball1]]{}. However, we can immediately deduce the “complex Hensley" theorem by using a more general fact (where the result of K. Ball has been used) proved in [@KPZ]: Let $K$ be an origin symmetric convex body in $\R^n$ and assume that the $2$-intersection body, $ I_{2}(K)$, exists and it is convex. Then $$d_{BM} (I_{2}(K), B_{2}^{n})\leq c ,$$ where $c>0$ is an absolute constant. Combining the above result with Proposition [\[Char2\]]{} and Theorem [\[CBusemann\]]{}, we immediately get the following Let $K$ be an origin symmetric convex body in $\C^{n}$. Then $$d_{BM}( I_{c}(K), B_{2}^{n}(\C)) \leq c,$$ where $c>0$ is an absolute constant and $ B_{2}^{n}(\C):= B_{2}^{2n}$ is the Euclidean ball in $\C^{n}$. [Acknowledgments:]{} The first named author wishes to thank the US National Science Foundation for support through grant DMS -1001234. The second named author wishes to thank the A. Sloan Foundation and US National Science Foundation for support through grant DMS-0906051 . Part of this work was carried out when the third named author was visiting the Mathematics Department of Texas A$\&$M University, which she thanks for hospitality. [100]{} , [Projection bodies in complex vector spaces]{}, Adv. Math. 227 (2011) 830–846. , [Logarithmically concave functions and sections of convex sets in $\R^n$]{}. Studia Math. 88, no. 1, (1988) 69–84. , [Some remarks on the geometry of convex sets]{}, Geometric aspects of functional analysis (1986/87), Lecture Notes in Math. 1317 (Springer-Verlag, Berlin-Heidelberg-New York 1988) 224–231. , [Complements of Lyapunov’s inequality]{}, Math. Ann. 205 (1973) 323–331. , [Eine Verallgemeinerung des Busemannschen Satzes vom Brunn-Minkowskischen]{}, Typ. Math. Ann. 144 (1961) 183–198. , [On the distribution of polynomials on high-dimensional convex sets]{}, Geometric aspects of functional analysis, Israel seminar (1989?90), Lecture Notes in Math.1469 (Springer, Berlin, 1991) 127–137. , [On the Busemann-Petty problem for perturbations of the ball]{}, Geom. Funct. Anal. 1 (1991) 1–13. , [A theorem on convex bodies of the Brunn-Minkowski type]{}, Proc. Nat. Acad. Sci. U.S.A. 35 (1949) 27–31. , Math. Scand. 4 (1956) 88–94. , [Geometric Tomography]{}, Second edition, Cambridge University Press, New York, 2006. , [Intersection bodies and the Busemann-Petty problem]{}, Trans. Amer. Math. Soc. 342 (1994), 435–445. , [A positive answer to the Busemann-Petty problem in three dimensions]{}, Ann. of Math. (2) 140 (1994), 435–447. , [An analytic solution to the Busemann-Petty problem on sections of convex bodies]{}, Ann. of Math. (2) 149 (1999), 691–703. , [Generalized functions, vol. 1. Properties and operations]{}, Academic Press, New York, 1964. , [Generalized functions, vol. 4. Applications of harmonic analysis]{}, Academic Press, New York, 1964. , [A note on a problem of H. Busemann and C. M. Petty concerning sections of symmetric convex bodies]{}, Mathematika 37 (1990) 239–244. , [Convolutions, transforms and convex bodies]{}, Proc. London Math. Soc. (3) 78 (1999) 77–115. , [Functional analytic characterization of classes of convex bodies]{}, Math. Z. 222 (1996) 363–381. , [Intersection bodies and ellipsoids]{}, Mathematika 42 (1995) 295–304. , [Slicing convex bodies, bounds of slice area in terms of the body’s covariance]{}, [Proc. Amer. Math. Soc.]{} 79 (1980) 619–625. , [On convex perturbations with a bounded isotropic constant]{}, Geom. and Funct. Anal. (GAFA) 16 (2006) 1274–1290. , [Fourier analysis in convex geometry]{}, Mathematical Surveys and Monographs, vol 116 AMS 2005. , [An application of the Fourier transform to sections of star bodies]{}, Israel J. Math 106 (1998), 157-164. , [A generalization of the Busemann-Petty problem on sections of convex bodies]{}, Israel J. Math. 110 (1999) 75–91. , [A functional analytic approach to intersection bodies]{}, Geom. Funct. Anal. 10 (2000) 1507–1526. , [Intersection bodies, positive definite distributions and the Busemann-Petty problem]{}, Amer. J. Math. 120 (1998) 827–840. , [Stability in the Busemann-Petty and Shephard problems]{}, Adv. Math. 228 (2011) 2145–2161. , [Stability of volume comparison for complex convex bodies]{}, Arch. Math. (Basel) 97 (2011) 91–98. , [A hyperplane inequality for measures of convex bodies in $\R^n,\ n\le 4$]{}, Discrete Comput. Geom. In press, doi:10.1007/s00454-011-9362-8. , [Intersection bodies in ${\R^4}$]{}, Adv. Math. 136 (1998) 1–14. , [The complex Busemann-Petty problem on sections of convex bodies]{}, Adv. Math. 218 (2008) 352–367. , [Stability and slicing inequalities for intersection bodies]{}, Preprint, 2011, arxiv:1108.2631v1. , [Isomorphic properties of intersection bodies]{}. J. Funct. Anal. 261 (2011) 2697–2716. , [The interface between convex geometry and harmonic analysis]{}, CBMS Regional Conference Series in Mathematics (108, AMS, Providence, RI, 2008). , [Extremal sections of complex $l_p$-balls, $0<p\le2$]{}, Studia Math. 159 (2003) 185–194. , [The existence of a centrally symmetric convex body with central sections that are unexpectedly small]{}, Mathematika 22 (1975) 164–175. , [Intersection bodies and the dual mixed volumes]{}, Adv. Math. 71 (1988) 232–261. , [Generalized Intersection Bodies]{}, Journal of Functional Analysis 240 (2006) (2) 530–567. , [Generalized Intersection Bodies are not equivalent]{}, Advances in Mathematics 217 (2008) (6) 2822–2840. , [On the Busemann-Petty problem about convex, centrally symmetric bodies in $\R^n$]{}, Mathematika 39 (1992) 258–266. , [Intersection bodies and generalized cosine transforms]{}, Adv. Math. 218 (2008) 696–727. , [Comparison of volumes of convex bodies in real, complex, and quaternionic spaces]{}, Adv. Math. 225 (2010) no. 3 1461–1498. , [Convex Bodies: The Brunn-Minkowski Theory]{}, Encyclopedia of Mathematics and its Applications 44 (Cambridge University Press, Cambridge, 1993). , [Centered bodies and dual mixed volumes]{}, Trans. Amer. Math. Soc. 345 (1994), 777–801. , [Intersection bodies and Busemann-Petty inequalities in $\R^4$]{}, Ann. of Math. (2) 140 (1994) 331–346. , [Sections of convex bodies]{}, Amer. J. Math. 118 (1996) 319–340. , [A positive answer to the Busemann-Petty problem in four dimensions]{}, Ann. of Math. (2) 149 (1999) 535–543. , [The Busemann-Petty problem for arbitrary measures]{}. Math. Ann. 331 (2005) 267–887. , [The complex Busemann-Petty problem for arbitrary measures]{}, Arch. Math. (Basel) 91 (2008) 436–449. , [The modified complex Busemann-Petty problem on sections of convex bodies*]{}, Positivity 13 (2009) 717–733.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We review the recent progress in understanding the [ properties of]{} spin-split superconductors under non-equilibrium conditions. Recent experiments and theories demonstrate a rich variety of transport phenomena occurring in devices based on such materials that suggest direct applications in thermoelectricity, low-dissipative spintronics, radiation detection and sensing. We discuss different experimental situations and present [a theoretical]{} framework based on quantum kinetic equations. Within this framework we provide an accurate description of the non-equilibrium distribution of charge, spin and energy, which are the relevant non-equilibrium modes, in different hybrid structures. We also review experiments on spin-split superconductors and show how transport measurements reveal the properties of the non-equilibrium modes and their mutual coupling. We discuss in detail [spin injection and diffusion and very large thermoelectric effects in spin-split superconductors.]{}' author: - 'F. Sebastian Bergeret' - Mikhail Silaev - Pauli Virtanen - 'Tero T. Heikkilä' bibliography: - 'superferro.bib' title: 'Colloquium: Nonequilibrium effects in superconductors with a spin-splitting field' --- Introduction\[sec-introduction\] ================================ Ferromagnetism and spin-singlet superconductivity are antagonist orders and hardly coexist in bulk materials. However, hybrid nanostructures allow for the possibility of combining the two phenomena via mutual proximity effects. The combination leads to the [emergence]{} of novel features not present in either system alone. We can make a distinction among those characteristics affecting the spectral properties of the materials, showing up when the probed systems are in equilibrium, and those related to nonequilibrium phenomena. The emphasis of our text is in the latter phenomena, especially related to steady-state currents or voltages applied across the structures. Both superconductors and ferromagnets are examples of electron systems with spontaneously broken symmetries, and thereby characterized by order parameters. The order parameter for a conventional spin singlet superconductor is the amplitude of (Cooper) pairing between electrons in states with opposite spins and momenta [@bardeen1957theory]. The presence of this complex pairing amplitude $F$ leads to two characteristic features of conventional superconductivity [@TinkhamBook; @deGennes:566105]: An equilibrium supercurrent that is proportional to the gradient of the phase of $F$ and that can be excited without voltage, and to the quasiparticle spectrum exhibiting an energy gap proportional to the absolute value of $F$. The resulting density of states (DOS, Eq. for $h_{\rm eff}=0$) is strongly energy dependent and results into a non-linear nonequilibrium response of superconductors. The main defining feature of ferromagnets is the broken spin-rotation symmetry into the direction of magnetization, and the associated exchange energy $h$ that splits the spin up and down spectra. This also leads to a strong spin dependence (spin polarization) of the observables related to ferromagnets. There are two mechanisms that prevent most of the ferromagnetic materials from becoming superconducting. One of them is the orbital effect due to the intrinsic magnetic field in ferromagnets. When this field exceeds a certain critical value, superconductivity is suppressed [@Ginzburg:1957]. The second mechanism is the paramagnetic effect @Chandrasekhar:1962 [@PhysRevLett.9.266; @saint1969type]. This is due to the intrinsic exchange field of the ferromagnet that shows up as a splitting of the energy levels of spin-up and spin-down electrons and hence prevents the formation of Cooper pairs. We focus here on the regime where this spin-splitting field is present, but not yet too large to kill superconductivity. In superconductors the spin-splitting field can be generated either due to the Zeeman effect in magnetic field or as a result of the exchange interaction between the electrons forming Cooper pairs and those which determine the magnetic order. Such fields can lead to drastic modifications of the ground state of a spin-singlet superconductor. The best-known example is the formation of the spatially inhomogeneous superconducting state predicted by @fulde_ferrell.1964 and @larkin_ovchinnikov.1964 and dubbed as FFLO. Although extensively studied in the literature, the FFLO phase only takes place in a narrow parameter window and therefore its experimental realization is challenging. ![image](./NewSemiconductorPicture10){width="1.8\columnwidth"} Other more robust phenomena related to the spin-splitting fields in superconductors have their origin in the quasiparticle spectrum modification. In the central panel of Fig. \[fig:SemiconductorPicture\] we show the resulting spin-split density of states. This was first explored experimentally by @PhysRevLett.25.1270 [@meservey1975tunneling] through the spin-valve effect in the superconductor/ferromagnet (Al/Ni) tunnel junctions (Fig. \[fig:SemiconductorPicture\]a). In these experiments the magnetic field was applied in the plane of a thin superconducting film, such that the paramagnetic effect dominates. The spin-split DOS was utilized to determine the spin polarization of an adjacent ferromagnet [@PhysRevLett.26.192; @tedrow1973spin; @PhysRevB.16.4907; @PhysRevB.22.1331; @meservey1994spin]. The basis of this spin-valve effect is the spin-resolved tunneling into the superconductor with spin splitting, shown in Fig. \[fig:SemiconductorPicture\]a. This schematic picture illustrates how by properly tuning the voltage across the junction, the electronic transport is dominated by only one of the spin species. That results in peculiar asymmetric differential conductance curves $dI/dV (V)\neq dI/dV (-V)$ observed in experiments and revealing the spin polarization. This idea has been used more recently to probe the spatially resolved spin polarization of different magnetic materials by means of scanning tunneling microscopy with spin-split superconducting tips [@Eltschka2014; @Eltschka2015]. Similar effects can also arise in thin superconducting films by the magnetic proximity effect from an adjacent ferromagnetic material [@PhysRevLett.56.1746]. In such a case, the spin splitting of the density of states can be observed for small magnetic fields or even at zero field, as discussed in Sec. \[sec-superwithh\]. The combination of spin-splitting fields with strong spin-orbit interaction in superconducting nanowires has also raised considerable interest as a platform for realizing topological phases and Majorana fermions, with possible applications in topological quantum computation [@aasen2016milestones]. Although these effects are beyond the focus of this review, the physics discussed below may help in understanding transport properties of the devices studied in that context. Due to the different nature of their broken symmetry, combining superconductors (S) and ferromagnets (FM) in hybrid structures leads to a multitude of effects where magnetism affects superconductivity and vice versa. Some of these effects show up already in equilibrium properties, especially studied in the context of proximity effects in superconducting/metallic ferromagnet hybrids and reviewed for example by @RevModPhys.77.935 and @Bergeret2005. The latter usually focus on the unusual behavior of Cooper pairs leaking from a superconductor into a metallic ferromagnet generating, for example, oscillating pair wave functions analogous to the FFLO state [@Buzdin:1982; @demler1997superconducting] and long-range spin triplet correlations [@bergeret2001long] induced by the coupling between the intrinsic exchange field of the ferromagnet and the leaked superconducting condensate [@bergeret2001long]. These effects manifest themselves in measurable equilibrium effects, such as the density of states and critical temperature oscillations in S/FM bilayers [@jiang1996superconducting; @kontos2001inhomogeneous], triplet spin valve effects in the critical temperature of FM/S/FM structures [@PhysRevX.5.021019], and unusual Josephson effects in SC/FM/SC junctions [@ryazanov2001coupling; @Singh2016]. Inversely, a magnetic proximity effect can arise when the triplet pairs, created in the FM region, leak back into the superconductor in a FM/S metallic bilayer, generating a non-vanishing magnetic moment in the SC within a coherence length $\xi_s$ from the SM/FM interface [@bergeret2004induced]. In contrast to these equilibrium proximity effects, here we focus on nonequilibrium properties of a superconducting material with a built-in spin-splitting field. The interest in studying such systems has been intensified recently due to the technological advances which allow for a controllable generation of spin splitting in thin superconducting films either by applying an external in-plane magnetic field [@Hubler2012a; @Quay2013] or by an adjacent ferromagnetic insulator. Structures with insulating FMs avoid the proximity effect suppressing superconductivity. Such nonequilibrium properties are studied by applying currents or voltages across the structures. The focus of our Colloquium is on steady-state nonequilibrium effects with time independent driving fields, but we also mention works studying alternating current (ac) responses. Often the nonequilibrium effects can survive to much higher distances than $\xi_s$, as their decay scales are determined via the various inelastic and spin-flip scattering lengths. Moreover, they can be studied at a weak tunneling contact to ferromagnets, making the analysis in some cases more straightforward than in proximity experiments. Nonequilibrium properties are related to the deviation of the electron distribution function from its equilibrium form, which leads to a nonequilibrium distribution (imbalance) of charge, energy or spin degrees of freedom. We refer to these different types of deviations from equilibrium as nonequilibrium modes.[^1] Specifically, we explore the coupling between these modes in superconductors with a spin-splitting field, and discuss unusually strong thermoelectric response and long-range spin signals. The above mentioned ability to characterize the spin polarized Fermi surface of metallic magnets with the help of spin-split superconductors has a direct connection with spintronics, and in particular with the search for spin valves with larger efficiencies than in the structures exhibiting large magnetoresistance [@PhysRevLett.61.2472; @PhysRevB.39.4828; @moodera1995large]. Indeed, a superconductor with a spin-splitting field has also an intrinsic energy dependent spin polarization around the Fermi level. This allows for studying different spintronic effects in a setting of a controllable non-linearity arising from the superconducting gap. Some of these effects are schematically shown in Fig. \[fig:SemiconductorPicture\]. This review explains those phenomena in detail. In normal metals and superconductors a spin accumulation, or spin imbalance, can be created by injection of a charge current from a ferromagnetic electrode [@PhysRevLett.55.1790; @van1987boundary; @jedema:345; @Johnson1994; @gu2002direct; @PhysRevB.71.144513; @Takahashi2003; @poli2008spin]. This state is characterized by the excess population in one of the spin subbands, determined by the balance between spin injection and relaxation or spin diffusion rates. In normal metals the nonequilibrium spin imbalance decays due to spin-flip scattering at typical distances of several hundreds of nanometers. In the superconducting state, at low temperatures $k_B T\ll \Delta$ the injection of any amount of carriers just above the energy gap shifts the chemical potential of quasiparticles rather strongly due to the large amount of quasiparticles at the gap edge \[Fig. \[fig:SemiconductorPicture\](c)\]. This leads to a strong spin signal in SF junctions [@Takahashi2003; @poli2008spin]. The spin relaxation length in normal metals depends only weakly on the temperature $T$. In the superconducting state, however, the scattering length is drastically modified with $T$. According to the first theory and experiments on spin injection in superconductors, the spin relaxation length was found to be reduced compared to the normal state [@morten2004spin; @poli2008spin]. However, subsequent experiments showed, contrary to expectations, an increase of the spin decay length [@Hubler2012a; @Quay2013]. It is now understood that these findings can only be explained by taking into account the spin-splitting field inside the superconductor [@silaev2015long; @krishtop2015-nst; @bobkova2016injection; @bobkova2015]. Due to this field, as shown in Sec. \[sec:noneq\_quasiclas\], it is necessary to take into account four types of nonequilibrium modes describing spin, charge, energy, and spin-energy imbalances. These modes provide the natural generalization of the charge and energy imbalances introduced by @schmid1975. In Sec. \[spininjection\] we show how the spin-splitting field couples pairwise these modes: charge to spin energy and spin to energy. Such a coupling leads to striking effects. For example, the coupling between the spin and energy modes leads to the long-range spin-accumulation observed in the experiments by @Hubler2012a [@Quay2013]. As we show in Sec. \[spininjection\] this long-range effect is related to the fact that the energy mode can only relax via inelastic processes which at low temperatures are rare. The coupling between different modes shows up also in tunnel contacts with spin-split superconductors. Because the spin-splitting field shifts the spin-resolved DOS away from the chemical potential of the superconductor, the system exhibits a strong spin-dependent electron-hole asymmetry. The spin-averaged density of states is still electron-hole symmetric, and therefore does not violate fundamental symmetries of the (quasiclassical) superconducting state. This spin-resolved electron-hole asymmetry leads to a large [spin Seebeck effect]{} shown schematically in Fig. \[fig:SemiconductorPicture\]b and discussed in Sec. \[subs:spinseebeck\]. A temperature difference across a tunneling interface between a normal metal and a spin-split superconductor drives a pure spin current between the electrodes, without transport of charge. If one of the electrodes is small so that the spin injection rate is large or comparable to the rate for spin relaxation, a spin accumulation forms in this electrode. However, it was noticed in several recent works [@Ozaeta2014a; @Machon2014; @Machon2013; @kalenkov12] that in certain situations the relevant observables are not spin-averaged, resulting in an effective electron-hole asymmetry showing up also in the charge current. The spin components are weighted differently in a setup consisting of the spin-filter junction connected to the spin-split superconductor [@Ozaeta2014a], shown schematically in Fig. \[fig:SemiconductorPicture\]d. As a result of this effective electron-hole symmetry breaking, the system exhibits a very large thermoelectric effect. This is discussed in Sec. \[thermoel\]. The main body of the review is organized as follows. In Sec. \[sec-superwithh\] we describe spin-split superconductors and give an overview of the quasiclassical theory that can be used for describing both their equilibrium and nonequilibrium properties. In Sec. \[sec:modes\] we describe the nonequilibrium modes in superconducting systems driven out of equilibrium in terms of the quasiclassical formalism. Section \[spininjection\] focuses on the spin injection and diffusion in superconducting systems, and reviews experiments performed to detect spin and charge imbalance in superconductors with and without spin-splitting. In Sec. \[thermoel\] we describe the giant thermoelectric response of a system exhibiting spin-polarized tunneling into a superconductor with a spin-splitting field. Finally, we present our conclusions and an outlook on possible future developments in the field in Sec. \[sec:outlook\]. A longer version of this review, along with comprehensive technical detail, can be found at [@bergeret2017nonequilibrium]. Superconductor with an exchange field \[sec-superwithh\] {#sec:SCwithh} ======================================================== The main focus of this colloquium is on superconductors with a spin-split density of states (DOS). As discussed in the introduction such a splitting can originate either by an external magnetic field [@PhysRevLett.25.1270] or by the exchange field induced by an adjacent ferromagnetic insulator [@PhysRevLett.56.1746]. The split DOS was observed in spectroscopy experiments [@PhysRevLett.25.1270; @PhysRevLett.26.192; @PhysRevB.16.4907; @PhysRevB.22.1331; @Hao:1990; @PhysRevLett.106.247001]. Formally, the normalized DOS of a spin-split Bardeen-Cooper-Schrieffer (BCS) superconductor is expressed as the sum of the DOS of each spin species, $N=N_\uparrow+N_\downarrow$, $$N=\frac{1}{2}{\rm Re}\frac{ \varepsilon+h_{\rm eff}}{\sqrt{\left(\varepsilon+h_{\rm eff}\right)^{2}-\Delta^{2}}}+\frac{1}{2}{\rm Re}\frac{\varepsilon-h_{\rm eff}}{\sqrt{\left(\varepsilon-h_{\rm eff}\right)^{2}-\Delta^{2}}}\;, \label{eq:split_Dos}$$ where $\pm h_{\rm eff}$ is the effective spin-splitting field. Equation (\[eq:split\_Dos\]) is a simplified description because it does not take into account the effect of magnetic impurities or spin-orbit coupling (SOC) [@meservey1994spin] discussed below. Often inelastic processes are described by $\varepsilon \mapsto \varepsilon + i \Gamma$, where $\Gamma$ is the @dynes84 parameter. In the case when the exchange field is induced by an adjacent ferromagnetic insulator (FI) there is no need of applying an external magnetic field [@PhysRevLett.56.1746; @Hao:1990; @Wolf2014; @PhysRevLett.106.247001; @Moodera_review; @spin_filter_Blamire]. Microscopically, the spin splitting originates from the exchange interaction between conduction electrons and the magnetic moments of the FI localized at the S/FI interface [@PhysRevB.38.8823; @khusainov1996indirect; @Kushainov_review]. The ferromagnetic ordering in the FI is due to a direct exchange coupling between the localized magnetic moments. In usual FIs the direct coupling is strong enough that one can assume that the magnetic configuration of the FI is [only]{} weakly affected by the superconducting state [@buzdin1988ferromagnetic; @Bergeret2000]. [The modification of the DOS is non-local]{} and survives over distances away from the FI/S interface of the order of the coherence length $\xi_{s}$ [@PhysRevB.38.8823; @bergeret2004induced]. [If the thickness $d$ of the S film is much smaller than $\xi_{s}$, the spin splitting can be assumed as homogeneous across the film. Thus the density of states can be approximated by Eq. (\[eq:split\_Dos\]) with an effective exchange field ${\bf h}_{\rm eff}\approx J_{\rm ex}\langle{\bf S_r}\rangle a/d$ [@deGennes1966coupling; @khusainov1996indirect; @PhysRevB.38.8823], where $a$ is the characteristic distance between the localized spins, $J_{\rm ex}$ is the exchange coupling between conduction electrons and localized moments, and $\langle{\bf S_r}\rangle$ is the average of the latter]{}. In Fig. \[fig:exp\_eus\_al\] we show an example of the measured differential conductance of an EuS/Al/Al$_2$O$_3$/Al junction. The Al layer adjacent to the EuS has a spin-split density of states that shows up as the splitting peaks (bright stripes in the figure) in $dI/dV$. Even at zero applied magnetic field the splitting is nonzero. The magnetization reversal of EuS at $H_c\approx -18.5$ mT manifests as a discontinuity of the conductance peaks [@strambini2017revealing]. As a first approach the DOS inferred from Fig. \[fig:exp\_eus\_al\] can be well described by the expression (\[eq:split\_Dos\]). The advantage of using a FI instead of an external magnetic field is that one avoids the depairing effects and all complications caused by the need to apply magnetic fields in superconducting devices. Moreover, because the electrons of the superconductor cannot propagate into the FI, superconducting properties are only modified by the induced spin-splitting field at the S/FI interface, and not by the leakage of Cooper pairs into the FI [as would happen in the case of metallic ferromagnets]{}. In addition, FIs can also be used as spin-filter barriers [@Moodera_review], in some cases with a very high spin-filtering efficiency, and therefore they play a crucial role for different applications as discussed below. ![Color plot of measured differential conductance, $dI/dV$ of a EuS/Al/Al$_2$O$_3$/Al junction as a function of the applied voltage and external magnetic field. $H_{co}$ denotes the coercive field of the EuS layer when the magnetization switches. Figure adapted from the work by @strambini2017revealing \[fig:exp\_eus\_al\] ](./FigEUS_AL2){width="\columnwidth"} In Table \[table:FIS\] we show a list of FI/S combinations and the reported induced exchange splittings and spin-filter efficiencies (barrier spin polarizations). Material Combination Barrier polarization Exchange Splitting (applied field) ---------------------- ------------------------ ------------------------------------ //AlO$_3$/Al $^1$ no spin-filter barrier 1 T (0.1 T)-1.73 T(0.4 T) Au/ / $^2$ 0.8 1.6 T (0 T) / /Al $^3$ 0.6-0.85 1.9-2.6 T (0T) Ag/ / $^4$ $>$ 0.97 none at zero field //AlO$_3$/Ag $^4$ no spin-filter barrier 4 T (0.6 T) / / $^5$ 0.75 / /TiN $^6$ 0.97 1.4 T (0T) ![Calculated density of states of a thin superconducting film at $T\rightarrow0$. We only show the DOS for one of the spin species, $N_\uparrow$. Shown by dashed red lines is the DOS in the absence of relaxation $\tau_{sn}=1/(\tau_{sf}^{-1}+\tau_{so}^{-1}) =\infty$ and zero exchange field $h=0$ which corresponds to a gap $\Delta_0$. Other curves are plotted for $h=0.4\Delta_0$ and different spin relaxation rates. (a) Spin-flip relaxation $\beta=(\tau_{\rm so}-\tau_{\rm sf})/(\tau_{\rm so}+\tau_{\rm sf})=1$, curves from top to bottom correspond to an increasing $(\tau_{sn}\Delta_0)^{-1} $, varying equidistantly from $0$ by $0.04$ steps. (b) Spin-orbit relaxation $\beta=-1$, curves from top to bottom correspond to an increasing $(\tau_{sn}\Delta_0)^{-1} $, varying equidistantly from $0$ by steps of $3.4$. For clarity the curves are shifted along the vertical axis. \[fig:Calculated-density-of\]](./figDOS1){width="\columnwidth"} [The paramagnetic effect, that leads to the spin-splitting, is modified by spin relaxation and orbital depairing. In their absence]{} the superconductivity survives the spin-splitting field up to the Chandrasekhar-Clogston limit [@PhysRevLett.9.266; @Chandrasekhar:1962] $h=\Delta_0/\sqrt{2}$[, where $\Delta_0$ is the order parameter at zero-field and zero-temperature.]{} At this field the system experiences a first-order phase transition into the normal state when the order parameter changes abruptly from $\Delta_0$ to zero. This picture changes qualitatively due the presence of magnetic impurities and spin-orbit scattering. Even at $T=0$ and $h=0$ the spin-flip processes induced by magnetic impurities result in the pair breaking effect closing the energy gap [@Abrikosov1961] at $\tau_{sf}\Delta_0=3/4$ and suppresses superconductivity completely at a certain critical value of the spin-flip time $\tau_{sf}$. For values of $\tau_{sf}$ larger than the critical one the phase transition switches from the first to the second order at [@bruno1973-mfs] $\tau_{sf}\Delta_0=0.461$ and the gapless state appears at a certain value of $h(\tau_{sf})$ (see Fig. \[fig:Calculated-density-of\]a) . Contrary to the spin-flip processes, the spin-orbit scattering alone does not have any effect on the superconducting state. However, in combination with $h\neq 0$ it tends to smear out the spin-splitted DOS singularities provided the spin-orbit relaxation time, $\tau_{so}$, is not very short (see Fig. \[fig:Calculated-density-of\]b). At short relaxation times [ $\tau_{so} \ll \Gamma/\Delta^2$, where $\Gamma$ is the depairing parameter [@dynes84] ]{} the effect of spin splitting is eliminated and the usual BCS density of states is recovered (see Fig. \[fig:Calculated-density-of\]b). Therefore in this case the critical spin-splitting field is strongly increased above the Chandrasekhar-Clogston limit[@bruno1973-mfs]. Besides broadening of the DOS singularities, the spin-orbit and spin-flip relaxation processes have an important effect on the paramagnetic spin susceptibility of the superconductor as it becomes non-vanishing even in the zero-temperature limit [@yosida1958paramagnetic; @abrikosov1961problem; @bruno1973-mfs]. The static spin susceptibility characterizes the paramagnetic response of the superconductor to an external magnetic field. In a usual normal metal the Zeeman field produces the same magnetization as a spin-dependent chemical potential shift $\delta \mu$ of the same magnitude when the distribution functions in different spin subbands are given by $f_{\uparrow}(E) = f_0(E+\delta\mu)$ and $f_{\downarrow}(E) = f_0(E-\delta\mu)$. This is different in superconductors where the paramagnetic susceptibility is determined by both the spin-polarized quasiparticles and the emergent spin-triplet superconducting correlations [@Abrikosov1962; @abrikosov1961problem]. On the other hand, the non-equilibrium spin modes as systematically described in Sec. \[sec:modes\] are determined only by the quasiparticle contribution. In the next sections we review the transport properties of diffusive hybrid structures with spin-split superconductors by contrasting existing theories and experiments. For this sake, in the next section we briefly introduce the quasiclassical Green’s function formalism for superconductors in the presence of spin-dependent fields and spin-polarised interfaces. [It is in our opinion]{} the most suitable formalism for the description of diffusive hybrid structures. Brief overview of the quasiclassical theory of diffusive superconductors\[sec:Usadel\] -------------------------------------------------------------------------------------- Quasiclassical Keldysh Green’s function technique is a useful and well-established way to describe transport and nonequilibrium properties of good metals, where the relevant physical length scales affecting different observables are long compared to the Fermi wave length $\lambda_F$, and where in particular disorder plays a major role. Several reviews explain this technique for various applications [@belzig1999quasiclassical; @Bergeret2005]. Here we just outline the main features relevant for spin-split superconductors. Briefly, the Keldysh Green’s functions (GFs), $\check G(\bf{r},\bf{r'};t,t')$ are two-point correlation functions which depend on two coordinates and two times. Here the “check” $\check G$ denotes GFs that live in a structure formed by the direct product of Keldysh, spin and Nambu spaces. The equation of motion for $\check G$ can be [written as]{} a kinetic-like equation for the Wigner transformed GF, $\check G(\bf{R},\bf{p})$, where $\bf{R}$ and $\bf{p}$ are the center of mass coordinate and $\bf{p}$ the momentum after Fourier transformation with respect to the relative component. A significant simplification can be done in the case of metals by noticing that the Green’s functions are peaked at the Fermi level. This allows for an integration of the equations over the quasiparticle energy, [related to]{} the magnitude of $\bf{p}$. This procedure leads to the quasiclassical GFs, $\check g (\bf{R},\bf{n})$, which only depend on the direction of the momentum at the Fermi level and on two times in the case of non-stationary problems, or only on a single energy $\varepsilon$ in the stationary case. These functions obey the @Eilenberger1968 equation. One of the advantages of using the quasiclassical GFs is that in the normal state, the spectral part is trivial, [i.e.]{}, the retarded and advanced GFs are energy independent. All transport information of the normal metal is encoded in the quasiclassical Wigner distribution function $\hat f (\bf{R},\bf{n})$ and quasiclassical equation for it resembles the classical Boltzmann equation [@Langenberg1986]. In contrast, the superconducting case distinguishes itself by a non-trivial spectrum, and therefore requires taking into account the full Keldysh structure of the GFs, [i.e.]{} $${\check g}=\left(\begin{array}{cc} {\hat g^{R}} & {\hat g^{K}}\\ 0 & {\hat g^{A}} \end{array}\right)\;. \label{eq:gmatrix}$$ This GF satisfies the normalization condition[@Eilenberger1968] $$\check{g}^{2}=\check{1}\;.\label{normalization}$$ In the diffusive limit the elastic mean free path $l$ due to scattering at non-magnetic impurities is much smaller than any other length involved in the problem except $\lambda_F$. Within this limit the Eilenberger equation can be reduced to a diffusive-like equation, in the same way as the Boltzmann equation is simplified in the diffusive limit. This quasiclassical diffusion equation for superconductors is the @usadel.1970 equation (we set $\hbar=k_B=1$) $$D\nabla\cdot(\check{g}\nabla\check{g})+[i \varepsilon\tau_{3}-i{\bm{h}}\cdot{\bm{\sigma}}\tau_{3}-\check{\Delta}- \check\Sigma,\check{g}]=0.\label{eq:Usadel}$$ Here $D$ is the diffusion coefficient, $\check{g}({\bf r},\varepsilon)$ is the isotropic (momentum independent) quasiclassical GF, $\bm{h}$ the spin-splitting field either generated by an external field or by the magnetic proximity effect in a FI/S junction, and $\check{\Delta}=\Delta e^{i\varphi \tau_3}\tau_1$ depends on the superconducting order parameter $\Delta$ that has to be determined self-consistently. Here $\tau_i$ and $\sigma_i$ are Pauli spin matrices in Nambu and spin space, respectively. The self-energy $\check{\Sigma}$ in Eq. (\[eq:Usadel\]) describes different scattering processes, such as elastic spin-flip or spin-orbit scattering, $\check \Sigma_{\rm el}$ and inelastic electron-phonon and electron-electron scattering, $\check \Sigma_{\rm in}$. Equation is central in the description of diffusive superconducting structures. Whereas the spectral properties can be obtained by solving the retarded [(R)]{} and advanced [(A)]{} components of this equation, nonequilibrium properties are described by the kinetic equation obtained by taking the Keldysh (K) component of Eq. (\[eq:Usadel\]). This can be compactly written as $$\nabla_k j_{kb}^a = H^{ab}+R^{ab}+I_{\rm coll}^{ab}, \label{eq:kineticeq}$$ where we introduce the spectral current tensor $j_{kb}^{a}$, $$j_{kb}^{a}=\frac{1}{8}{\rm Tr}\tau_{b}\sigma_{a}(\check{g}\nabla_{k}\check{g})^{K}\;. \label{eq:general_current}$$ The different current density components (charge, spin, energy, spin-energy) can be obtained from Eq (\[eq:general\_current\]). For example, the charge current density reads $$J_{k} = \frac{\sigma_{N}}{2e}\int_{-\infty}^\infty d\varepsilon\,j_{k3}^0, \label{eq:charge_current}$$ Here $\sigma_N=e^2 \nu_F D$ and $\nu_F$ are the normal-state conductivity and density of states at the Fermi level respectively. In Eq. (\[eq:kineticeq\]) the term $H^{ab}={\rm Tr} \tau_b \sigma_a [-i{\mathbf h} \cdot \bm{\sigma} \tau_3,\hat g^K]/8$ describes the Hanle precession of spin caused by the exchange field, and $R^{ab}={\rm Tr} \tau_b \sigma_a [\hat\Delta,\hat g^K]/8$ the conversion between quasiparticles and the superconducting condensate. Finally $I_{\rm coll}^{ab}={\rm Tr} \tau_b \sigma_a [\check \Sigma,\check g]^K/8$ in Eq. (\[eq:kineticeq\]) is the collision integral describing the different scattering process with self-energy $\check \Sigma$. We discuss next different scattering processes. Elastic self-energy terms. We consider elastic contributions to $\check{\Sigma}_{el}$ due to scattering at impurities with spin-orbit coupling (relaxation time $\tau_{so}$) and the spin flips at magnetic impurities ($\tau_{sf}$) [@maki1966effect]. [Within the Born approximation, they read $\check{\Sigma}_{so}={\bm{\sigma}}\cdot\check{g}{\bm{\sigma}}/(8\tau_{so})$, $\check{\Sigma}_{sf}={\bm{\sigma}}\cdot\tau_{3}\check{g}\tau_{3}{\bm{\sigma}}/(8\tau_{sf})$.]{} In the normal state they contribute to the energy-independent total spin-relaxation time $\tau_{\rm sn}^{-1}=\tau_{\rm so}^{-1}+\tau_{\rm sf}^{-1}$. In contrast, in the superconducting case the spin-relaxation time and length acquire energy dependence, which is different for the spin-orbit and spin-flip scattering [@maki1966effect; @morten2004spin; @morten2005spin]. Therefore it is convenient to describe the relative strength of these two scattering mechanisms in terms of the parameter $\beta=(\tau_{\rm so}-\tau_{\rm sf})/(\tau_{\rm so}+\tau_{\rm sf})$. In diffusive superconducting thin films one can also describe the depairing effect of an in-plane magnetic field with a self-energy term $\check{\Sigma}_{orb}=\tau_{3}\check{g}\tau_{3}/\tau_{orb}$ characterized by the orbital depairing time $\tau_{orb}$ [@deGennes:566105; @anthore2003]. This term also contributes to charge imbalance relaxation [@schmid1975; @nielsen1982pair]. The parameters $\tau_{\rm sn}^{-1}$ and $\beta$ are material specific. For example, in Al films, the reported values from a set of spin injection experiments are $\tau_{\rm sn} \approx 100$ ps [@jedema:713; @poli2008spin] and $\beta \approx 0.5$ indicating the dominance of spin-flip relaxation over spin-orbit scattering, whereas the reported value of $\tau_{sn}$ in Nb is only 0.2 ps, and is strongly dominated by spin-orbit scattering [@wakamura2014spin]. They affect both the spectrum of a bulk superconductor (see Fig. \[fig:Calculated-density-of\]) and the spin relaxation as described in Sec. \[sec:nl\_detection\]. Inelastic self-energies. The relevant inelastic processes entering the self-energy in Eq. , are the particle–phonon and particle–particle collisions. These processes do not conserve the energies of colliding quasiparticles, but conserve the total spin. The coupling between quasiparticles and phonons limits some of the effects discussed in the following sections. Due to the energy dependence of the phonon density of states, this coupling decreases rapidly towards low temperatures, and eventually phonons decouple from electrons, and the main heat relaxation occurs via other processes such as quasiparticle diffusion. Superconductivity modifies the electron-phonon heat conduction [@eliashberg1972-iec; @kopnin2001-ton; @kaplan1976], as also the electronic spectrum is energy dependent, and is affected by the spin splitting [@grimaldi1997; @virtanen2016stimulated]. Particle-particle collisions in superconductors and superfluids are discussed by @eliashberg1972-iec [@serene1983-qat; @kopnin2001-ton], although mainly within contact interaction models disregarding screening effects [@narozhny1999; @feigelman2000; @kamenev2009]. The collision integrals can have spin structure also in the normal state [@Dimitrova2008; @chtchelkatchev2008-ers]. The far-from-equilibrium results discussed in Sec. \[spininjection\] disregard the particle-particle collisions, as the simpler theory already describes effects not very far from the measured ones. On the other hand, Sec. \[thermoel\] mostly concentrates on the quasiequilibrium limit, where also spin accumulation is lost due to a strong spin relaxation. Hybrid interfaces. In subsequent sections we apply the kinetic equation, Eq. (\[eq:kineticeq\]), in different situations. For the description of transport in hybrid structures, we need in addition a description of interfaces between different materials in the form of boundary conditions. Such interfaces usually are described by sharp changes of the potential and material parameters over atomic distances, and thus cannot be included [directly]{} in the quasiclassical equations which describe properties over distances much larger than $\lambda_F$. The description of hybrid interfaces requires then the derivation of suitable boundary conditions, first done in the quasiclassical approach by @zaitsev1984quasiclassical. Boundary conditions for the Usadel equation trace back to the work of @Kupriyanov1988. These boundary conditions are applicable for non-magnetic N-N, S-S and S-N interfaces with low transmissivity [@PhysRevB.55.6015]. Later [@nazarov.1999]{} generalized these boundary conditions for an arbitrary interface transparency. @PhysRevB.38.8823 derived the boundary condition for an interface between a superconductor and a ferromagnetic insulator and introduced the concept of the spin-mixing angle, which describes the spin-dependent phase shifts acquired by the electrons after being scattered at the FI/S interface. Later @PhysRevB.70.134510 [@PhysRevB.80.184511] extended these boundary conditions to magnetic metallic structures, such as F-S or S-F-S systems, though with low polarization. Boundary conditions for large polarization and low transmission have been presented by @Bergeret2012 [@Machon2013]. General boundary conditions for arbitrary spin polarization and transmission have been extensively discussed by @eschrig2015general. Here we mainly deal with low transmissive barriers between a mesoscopic superconductor and normal and magnetic leads and use the description presented by @Bergeret2012. [In this description, the component of the spectral current density multiplied by $\sigma_N$ perpendicular to the interface is continuous across it, and given by $$\begin{aligned} \sigma_N j_{\perp,b}^a = -\frac{1}{8 e R_{\square}} {\rm Tr}\tau_{b}\sigma_{a}\left[\hat{\Gamma}\check{g}_{2}\hat\Gamma^{\dagger},\check{g}\right]^{K} \label{eq:currents_at_interface} \,,\end{aligned}$$ where $R_{\square}$ is the spin-averaged barrier resistance per unit area, and the spin-dependent transmission is characterized by the tunneling matrix $\hat{\Gamma}=t\tau_{3}+u\sigma_{3}$, assuming polarization in the $z$-direction. The normalized transparencies satisfy $t^{2}+u^{2}=1$ and are determined from the interface polarization $\|P\|\le1$ via $2ut=P$. The Green’s function $\check{g}_{2}$ in the r.h.s of Eq. (\[eq:currents\_at\_interface\]) corresponds to the opposite side of the junction.]{} Nonequilibrium modes in spin-split superconductors {#sec:modes} ================================================== The out-of-equilibrium state in superconducting systems is characterized by the presence of nonequilibrium modes associated with the different electronic degrees of freedom. For example, injection of an electric current from a normal electrode into a superconductor generates a charge imbalance mode [@PhysRevLett.28.1363; @PhysRevB.6.1747; @PhysRevLett.28.1366; @yagi2006charge; @hubler2010charge] that diffuses into the S region. This nonequilibrium mode reflects an imbalance of the quasiparticle population between the electron-like and hole-like spectrum branches. The charge imbalance measurements made in the 1970s were to our knowledge the first to study such nonequilibrium modes in non-local multiterminal settings. This technique was later adapted to spintronics, to study the nonequilibrium spin accumulation induced by spin-polarized electrodes [@PhysRevLett.55.1790]. ![image](./dFUN-newnew) Schematically, nonequilibrium modes can be represented in terms of the electron/hole branches in the spectrum of the superconductor [@TinkhamBook], as illustrated in Fig. \[fig:DistributionFunction\]. For example the charge mode can be understood as the imbalance between the electron and hole branches (Fig. \[fig:DistributionFunction\]b). In the absence of spin-dependent fields there is one more nonequilibrium mode: the energy imbalance mode (Fig. \[fig:DistributionFunction\]e) . It describes the excess energy stemming from an equal change in the quasiparticle populations of the electron-like and hole-like branches. This energy mode affects charge transport properties indirectly via the self-consistency equation for $\Delta$. This mechanism explains, for example, the enhancement of the superconducting transition temperature in the presence of a microwave field [@Eliashberg_enhacement; @Klapwijk_enhancement]. In this section we generalize the description in terms of nonequilibrium modes to account for superconductors with spin-split density of states. The spin splitting [(energy difference $2h=2\mu_B B$ between the black and red dispersion curves in Fig. \[fig:DistributionFunction\] for spin up/down quasiparticles)]{} gives rise to four distinct quasiparticle branches, electron/hole and spin up/down. These four nonequilibrium modes and their coupling are at the basis of the main effects discussed in this review. Description of nonequilibrium modes in superconductors with spin splitting {#SubSec:Modes} -------------------------------------------------------------------------- At this point [we]{} combine the pictorial description of the nonequilibrium modes (Fig. \[fig:DistributionFunction\]) with the quasiclassical formalism introduced in Sec. \[sec:Usadel\] and [in particular,]{} the Usadel equation. For a description of non-equilibrium properties we need to consider the Keldysh component $\hat g^K$ of the quasiclassical GF \[Eq. (\[eq:gmatrix\])\]. For clarity we first consider a unique spin polarization direction parallel to the $z$-axis. From the normalization condition, Eq. (\[normalization\]), $\hat g^K$ can be expressed in terms of the retarded and advanced components and the generalized matrix distribution function $\hat f$ [@Langenberg1986] $$\hat{g}^{K}=\hat{g}^{R}\hat{f}-\hat{f}\hat{g}^{A}\;.\label{eq:K_param}$$ In the case of only one spin polarization axis, the 4$\times$4 matrix distribution function $\hat{f}$ can be written as the sum of four components[^2] $$\hat{f}=f_{L}\hat{1}+f_{T}\tau_{3}+(f_{T3}\sigma_{3}+f_{L3}\sigma_{3}\tau_{3})\;.\label{eq:distribution_function0}$$ For historical reasons we use the labeling introduced by @schmid1975, generalized for the spin-dependent case. The $L$-labeled functions describe longitudinal modes, the (spin) energy degrees of freedom, and are antisymmetric in energy with respect to the Fermi level, $\varepsilon=0$. The $T$-labeled functions describe transverse modes and are symmetric in energy. In equilibrium, the distribution function is proportional to the unit matrix in Nambu and spin space, and given by $$\hat{f}_{eq}(\varepsilon)=(1-2n_F)\hat 1=\tanh(\varepsilon/2T)\hat 1\; .\label{eq:feq}$$ We can now turn to the pictorial description of Fig. \[fig:DistributionFunction\] and associate each component of $\check f$ in Eq. (\[eq:distribution\_function0\]) with a nonequilibrium mode as discussed next. As shown in Figs. \[fig:DistributionFunction\](b)–(e), two of these modes have electron-hole branch imbalance, $f_T$ and $f_{L3}$, while $f_{T3}$ and $f_L$ are particle-hole symmetric. The filled circles in Fig. \[fig:DistributionFunction\] represent the occupied states. As a reference, panel (a) corresponds to the equilibrium distribution function $\hat f = f_L^0 \hat 1=\tanh(\varepsilon/2T) \hat 1$. In order to excite the nonequilibrium modes, $f_T$, $f_{T3}$ and $f_{L3}$, one only needs to move the populated states (filled circles) between the different spectral branches in an elastic process, i.e., between equal-energy states (marked by horizontal dashed arrows). These modes can also relax back to equilibrium due to elastic scattering processes. The relaxation mechanisms depend on intrinsic material properties, degree and type of disorder, and also on the superconducting spectrum, and are discussed in more detail below. The last nonequilibrium mode, the deviation of $f_L$ from $f_L^0$, is characterized by a change in the total quasiparticle number and energy content, corresponding to an increase or decrease of the effective temperature. It can be excited by increasing the number of occupied states to higher energies, and its relaxation requires inelastic processes. In the absence of spin splitting, the charge imbalance is determined by $f_T$, and the energy imbalance by $f_L$. The spin splitting changes the system properties, mixing the coupling between spin-dependent modes and physical observables \[see Eqs. (\[Eq:ChPot0\]) and (\[Eq:ChPotZ\]) below\]. Qualitatively, the outcome can be seen by counting the number of occupied states on the different branches in Fig. \[fig:DistributionFunction\]. For example, the charge imbalance $\mu$ is determined by the difference between the number of occupied states in the electron and hole branches. Both $f_T$ and $f_{L3}$ components contribute to it, as seen in Figs. \[fig:DistributionFunction\]b and d. On the other hand, a nonzero spin accumulation $\mu_z$ can be induced by exciting the modes $f_{T3}$ or $f_L$ \[Figs. \[fig:DistributionFunction\](c),(e)\]. These two contributions to the total spin accumulation have important differences: The mode $f_{T3}$ contributes to spin imbalance also in the absence of spin splitting. Spin imbalance in this mode can be induced for example by a spin-polarized injection from a ferromagnetic electrode, in both the normal and the superconducting state. The relaxation of the spin accumulation created in this way is determined by elastic scattering processes. The second mechanism of inducing spin accumulation is by exciting the longitudinal mode $f_L$, in the presence of spin splitting \[Fig. \[fig:DistributionFunction\](e)\]. Since energy-conserving transitions do not result in the relaxation of the $f_L$ mode, this component of the spin imbalance is not suppressed by elastic scattering. In other words, its relaxation can be only provided by inelastic processes, e.g., electron-phonon and electron-electron scattering. This result, obtained here on a phenomenological level, is crucial in understanding the long-range spin signal observed in superconductors, for example by @Hubler2012a and discussed in the next sections. [Accumulations]{} in terms of the non-equilibrium modes \[sec:noneq\_quasiclas\] -------------------------------------------------------------------------------- [Quantitatively, we define the charge and spin accumulations]{} based on the Keldysh component of the GF, Eq. , $$\begin{aligned} \mu({\bf r},t) & = & -\int_{-\infty}^{\infty} \frac{d\varepsilon}{16} \operatorname{Tr}\hat g^{K}(\varepsilon,{\bf r},t)\label{eq:def_mu}\\ \mu_{sa}({\bf r},t) & = & \int_{-\infty}^{\infty} \frac{d\varepsilon}{16} \operatorname{Tr}\tau_{3}\sigma_{a}[ \hat g_{\rm eq}^{K}(\epsilon,\mathbf{r},t)- \hat g^{K}(\varepsilon,\mathbf{r},t)]\; , \label{eq:def_mus}\end{aligned}$$ whereas the local energy and spin-energy accumulations are given by $$\begin{aligned} \label{Eq:HeatDens} q({\bf r},t) & = & \int_{-\infty}^{\infty} \frac{d\varepsilon}{16} \varepsilon \operatorname{Tr}\tau_3 [\hat g_{\rm eq}^{K}(\varepsilon,{\bf r},t)- \hat g^{K}(\varepsilon,\mathbf{r},t)]\label{eq:def_q} \\ \label{Eq:SpinEnergyDens} q_{sa}({\bf r},t) & = & \int_{-\infty}^{\infty} \frac{d\varepsilon}{16} \varepsilon \operatorname{Tr}\sigma_{a}[ \hat g_{\rm eq}^{K}(\varepsilon,\mathbf{r},t)- \hat g^{K}(\varepsilon,\mathbf{r},t)]. \label{eq:def_qs}\;\end{aligned}$$ Above, $a=1,2,3$ denotes the polarization direction of the nonequilibrium spins and energy is counted with respect to the potential $\mu_S$ of the superconducting condensate (see below). In terms of the distribution functions, the charge and spin accumulations read [(we assume magnetization in $z$-direction)]{} $$\begin{aligned} \label{Eq:ChPot0} & \mu = -\frac{1}{2}\int_{-\infty}^{\infty} d\varepsilon ( N_+ f_T+ N_- f_{L3}) \\\label{Eq:ChPotZ} & \mu_z = - \frac{1}{2}\int_{-\infty}^{\infty} d\varepsilon [ N_+ f_{T3}+ N_-(f_{L}-f_{\rm eq})],\end{aligned}$$ where $N_+ = N_\uparrow + N_\downarrow$ is the total density of states (DOS), $N_- = N_\uparrow - N_\downarrow $ is the DOS difference between the spin subbands, and $f_{\rm eq}(\varepsilon) = \tanh(\varepsilon/2T)$ is the equilibrium distribution function. [Similarly for (\[Eq:HeatDens\],\[Eq:SpinEnergyDens\]) we get $$\begin{aligned} \label{Eq:HeatDensQuasi} & q = \frac{1}{2}\int_{-\infty}^{\infty} d\varepsilon \varepsilon [ N_+ (f_L-f_{eq}) + N_-f_{T3}] \\ \label{Eq:SpinEnergyDensQuasi} & q_{sa} = \frac{1}{2}\int_{-\infty}^{\infty} d\varepsilon \varepsilon [ N_- (f_{L}-f_{\rm eq}) + N_+f_{T3}].\end{aligned}$$ ]{} [ All these quantities, Eqs. (\[eq:def\_mu\]-\[eq:def\_qs\]) are directly related to experimental observables. ]{} The charge imbalance $\mu$ characterizes the potential of the quasiparticles in the superconductor [@artemenko1979-efc]. In nonequilibrium situations, $\mu$ can differ from the condensate potential $\mu_S$. In the problems discussed in this Colloquium, $\Delta$ can be chosen time-independent and $\mu_S=0$. The charge density depends on $\mu$ via $\rho=-\nu_Fe^2\phi-e\nu_F\mu$ where $\phi$ is the electrostatic scalar potential [@kopnin2001-ton]. In metals, local charge neutrality is maintained on length scales large compared to the Thomas–Fermi screening length, so that $-e\phi=\mu$ and charge imbalance is associated with static electric fields. In the quasiclassical formulation used here, electrochemical potential differences appear explicitly in energy shifts in the boundary conditions for the distribution functions [@belzig1999quasiclassical]. The Fermi distribution at potential $V$ corresponds to $$\begin{aligned} f_{\mathrm{eq},L(T)}(E)=\frac{1}{2}[\tanh\bigl(\frac{E+eV}{2T}\bigr) +(-) \tanh\bigl(\frac{E-eV}{2T}\bigr)] \,.\end{aligned}$$ For superconductor at equilibrium, $V=0$ in this description. However, $V=\phi\ne0$ can describe voltage-biased normal ($\Delta=0$) reservoirs. Spin accumulation is a standard observable in spintronics [@PhysRevLett.55.1790; @jedema:713]. The local energy accumulation is typically measured via electron thermometry [@giazotto2006]. The spin-energy accumulation was measured recently in normal-state nanopillar spin valves [@dejene2013]. To our knowledge this quantity has not been directly studied experimentally in superconducting systems. In the normal state the spectrum is trivial, $g^{R(A)}=\pm \tau_3\sigma_0$. Thus, according to Eq. (\[eq:K\_param\]), the Keldysh component is simply proportional to the distribution function. In other words, the different [modes decouple]{} in Eqs. (\[eq:def\_mu\]-\[eq:def\_qs\]). Moreover, in the normal state it is unnecessary to separate between transverse and longitudinal modes, and rather consider the spin-dependent full distribution function $f_j(E)=[1-f_{Lj}(E)-f_{Tj}(E)]/2$. Solutions of the kinetic equation in the normal state are discussed for example by @Brataas:2006en. In the superconducting case the situation is more complex. First, the spectrum is strongly energy dependent around the Fermi level and the spectral GFs have a non-trivial structure in spin space. Components proportional to the unit matrix in spin space describes the BCS singlet GFs, whereas terms proportional to the Pauli matrices $\sigma_{j}$, $j=1,2,3$, describe the triplet state [@bergeret2001long; @Bergeret2005]. Second, due to this energy dependence and non-trivial spin structure, the spectral functions enter (\[eq:def\_mu\]-\[eq:def\_qs\]) and lead to a coupling between the different non-equilibrium modes that in turns couple all electronic degrees of freedom, as discussed next. Spin injection and diffusion in superconductors\[spininjection\] ================================================================ Non-equilibrium modes can be experimentally studied by means of non-local transport measurements. In this section we review experiments on charge and spin injection in superconductors, and [apply the kinetic equation approach described in the]{} previous sections to describe different experimental situations. Detection of spin and charge imbalance: Non-local transport measurements\[sec:nl\_detection\] --------------------------------------------------------------------------------------------- ![Scanning electron microscopy image of the lateral structure used by @Hubler2012a. From @Hubler2012a. []{data-label="fig:experiemnt_beckmann"}](Beckmannspin_inj){width="\columnwidth"} Studies of the nonequilibrium modes started with the pioneering experiment of @PhysRevLett.28.1363, who realized a way of detecting the charge imbalance in a superconductor. The main idea of this experiment is to inject a current from a normal metal (injector) into a superconductor. This current creates a charge imbalance that [corresponds to]{} a shift of the chemical potential of the quasiparticles with respect to the one of the condensate. This shift of the chemical potential can be detected by a second electrode (detector) that probes the voltage between the superconductor and the detector. More recent experiments used the same non-local measurement to explore the charge, energy and spin modes in mesoscopic superconducting lateral structures [@beckmann2004evidence; @hubler2010charge; @wolf2014charge; @PhysRevB.87.024517; @Quay2013; @poli2008spin]. A scanning electron microscopy image of such a lateral structure is shown in Fig. \[fig:experiemnt\_beckmann\]. A detailed overview of the experiments on charge and energy imbalance can be found in the recent topical reviews by @beckmann2016spin [@quay2017up]. Whereas the charge and energy modes were known for a long time, it was first in the 1990s that theorists predicted that electronic charge and spin degrees of freedom can be separated in a superconductor [@Kivelson1990; @PhysRevB.52.3632]. First experiments on F-S-F layered structures, however, did not show any evidence of such a spin-charge separation [@Johnson1994; @gu2002direct] and the different relaxation times for spin and charge accumulation in superconductors remained an open question. First clear insight into the separation of the spin and charge modes was obtained in experiments using lateral nanostructures with ferromagnetic injectors and detectors [@beckmann2004evidence; @PhysRevB.71.144513; @cadden2007charge; @Hubler2012a; @Kolenda2016; @PhysRevB.87.024517; @Wolf2014; @poli2008spin; @Quay2013; @yang2010extremely]. First theoretical works on spin injection into mesoscopic superconductors [@morten2004spin; @morten2005spin] showed that the spin-relaxation length in the superconducting state strongly depends on the energy of the injected quasiparticles and on the spin relaxation mechanism. In particular, for a dominating spin-orbit scattering, superconductivity suppresses the spin relaxation [rate $\tau_s^{-1}$]{}, which can be qualitatively understood as the decrease in the cross section of the quasiparticle momentum scattering at the energies near the gap edge $\varepsilon\sim \Delta$. The suppression of $\tau_{s}^{-1}$ is however compensated by the decrease in the quasiparticle group velocity $v_g\sim{}v_F\sqrt{1-\|\Delta\|^2/\varepsilon^2}$ so that the spin relaxation length $\lambda_{so} \sim v_g \tau_{s} $ remains almost unchanged in the superconducting state. On the contrary, if the spin-flip mechanism dominates, the spin relaxation is not related to the momentum scattering because the interaction with magnetic impurities does not depend on the propagation direction and the quasiparticle spin does not depend on energy. This results in an increase of $\tau_{s}^{-1}$ which is equivalent to a decrease of the spin-relaxation length in the superconducting state. Although these works provided an explanation to some experiments, two important features observed in subsequent works could not be explained in terms of that theory: First, the spin accumulation [was]{} detected at distances from the injector much larger than the spin-relaxation length measured in the normal state [@Quay2013; @Hubler2012a; @PhysRevB.87.024517]. Second, an unexpected spin accumulation was observed even if the current was injected from a non-magnetic electrode [@PhysRevB.87.024517]. [ In order to explain these two observations one needs to take into account the spin splitting in the superconductor. ]{} Non-local conductance measurements in spin-split superconductors ---------------------------------------------------------------- Specifically, one of the setups studied by @Hubler2012a, was a lateral non-local spin valve [(see Fig. \[fig:experiemnt\_beckmann\])]{} where the experimentalists determined the non-local differential conductance $$g_{nl}= \frac{d I_{\rm det}}{d V_{\rm inj}} \; .\label{eq:gnl0}$$ Typical experimental curves are shown in Fig. \[Fig:Cond-nl\]a, adapted from @Hubler2012a and Fig. \[Fig:Cond-nl\]b shows the results calculated from the kinetic equations. If the detector is a ferromagnet with magnetization collinear with the spin accumulation in the wire, the current at the detector [for $V_{\rm det}=0$ is obtained from Eqs. (\[eq:charge\_current\],\[eq:currents\_at\_interface\]),]{} $$\label{Eq:ZeroCurrentYGen} I_{\rm det}= (\mu + P_{\rm det} \mu_z)/R_{\rm det},$$ where $R_{\rm det}=R_\square/A$ is the detector interface resistance in the normal state, $A$ is the cross-sectional area of the detector, $\mu$ is the charge imbalance and $\mu_z$ the spin imbalance defined in Eqs. (\[eq:def\_mu\],\[eq:def\_mus\]). According to the explicit expressions (\[Eq:ChPot0\],\[Eq:ChPotZ\]), the full description of the non-local current requires all four non-equilibrium modes. Particularly interesting is the contribution from the second term in the r.h.s. of Eq. (\[Eq:ChPotZ\]). It is nonzero when the spin splitting described by $N_-$ is nonzero and it provides a qualitative explanation of the experiments by @Quay2013 [@Hubler2012a; @PhysRevB.87.024517]: The spin imbalance $\mu_z$, being related to the energy nonequilibrium mode $f_L$, once excited can only relax via inelastic processes, especially mediated by the electron-phonon interaction. At low temperatures the corresponding decay length can be much larger than the spin decay length in normal metals. This explains the long-range non-local signal observed in the experiments. The observed long-range spin accumulation can thus be understood to result from the spin accumulation generated by the effective heating of the superconducting wire caused by the injection of nonequilibrium quasiparticles with energies larger than the superconducting gap [@silaev2015spin; @silaev2015long; @virtanen2016stimulated; @bobkova2016injection; @bobkova2015; @krishtop2015-nst]. Such a heating can originate, for example, by an injected current even from the non-ferromagnetic electrode. The heating is not sensitive to the sign of the bias voltage at the injector and hence the generated spin imbalance must be an even function of the voltage, $\mu_z(V_{\rm inj})=\mu_z(-V_{\rm inj})$. This leads to an antisymmetric shape of the non-local spin signal in $g_{nl}$ with respect to $V_{\rm inj}$, [in agreement with the experimental observation [@wolf2014charge].]{} All these features occur only if the superconductor has a spin-split density of states induced either by an external magnetic field or by the proximity to a ferromagnetic insulator. ![\[Fig:Cond-nl\] (a) Nonlocal conductance measured as a function of the injecting voltage, $g_{nl}(V_{\rm inj})$ adopted from @Hubler2012a. (b) The same quantity calculated using the kinetic theory for $\alpha_{orb}=1.33$, $\beta=0.5$, $(\tau_{sn} T_{c0})^{-1} =0.2$, $T=0.05 T_{c0}$, effective inelastic relaxation length $L=20\lambda_{sn}$, $L_{\rm det}=5\lambda_{sn}$. Black solid and red dash-dotted curves correspond to the injection from non-ferromagnetic ($P_{inj}=0$) and ferromagnetic ($P_{inj}=0.5$) electrodes, respectively at the spin-splitting $h=0.2\Delta_0$. Blue dashed line corresponds to $h=0$. The conductance is normalized to $g_0 = R_\xi/(R_{inj}R_{det})$, where $R_\xi = \xi/(A_s \sigma_N)$ is the normal-state resistance of the wire with length $\xi$ and cross section $A_s$. ](./figCond){width="\columnwidth"} [ A quantitative description of these effects can be provided by solving the kinetic equations for superconductors with spin-split subbands [@silaev2015long]. In this case the diffusion couples non-equilibrium modes pairwise. In particular, the kinetic equations take the form]{} $$\label{eq:KineticEquation} \nabla\cdot\begin{pmatrix} j_e\\ j_s\\ j_c\\ j_{se} \end{pmatrix}= \begin{pmatrix} 0& 0& 0& 0\\ 0& 0& 0& S_{T3}\\ 0& 0& R_T& R_{L3}\\ 0& 0& R_{L3} & R_T+S_{L3} \end{pmatrix} \begin{pmatrix} f_L\\ f_{T3}\\ f_T\\ f_{L3} \end{pmatrix},$$ where the spectral energy $j_e$, spin $j_s$, charge $j_c$ and spin energy $j_{se}$ currents derived from the general Eq.(\[eq:general\_current\]) are $$\label{eq:Current1} \begin{pmatrix} j_e\\ j_s\\ j_c\\ j_{se} \end{pmatrix}= \begin{pmatrix} D_L & D_{T3} & 0 & 0\\ D_{T3} & D_L & 0 & 0\\ 0 & 0 & D_T & D_{L3}\\ 0 & 0 & D_{L3} & D_T \end{pmatrix} \begin{pmatrix} \nabla f_L\\\nabla f_{T3}\\\nabla f_T\\\nabla f_{L3} \end{pmatrix}.$$ Here $D_{L/T/T3/L3}$ are kinetic coefficients related to the spectral GFs [@silaev2015long], $S_{T3/L3}$ are parts of collision integrals describing spin relaxation, and $R_{T/L3}$ the coupling between the quasiparticles and the superconducting condensate. On the one hand, the charge is coupled to the spin-energy mode \[lower right block of Eq. \]. The relaxation of both of these modes, right hand side of Eq. , is nonvanishing for all energies, below and above the gap due to the magnetic pair breaking effects [@schmid1975; @nielsen1982pair]. On the other hand, the spin-splitting field couples the spin and energy modes, $f_{L}$ and $f_{T3}$ respectively \[upper left block of Eq. \]. As explained above, the energy mode can only decay via inelastic scattering which at low temperature can be disregarded compared to the spin relaxation. Solutions of Eqs. (\[eq:KineticEquation\],\[eq:Current1\]) along with Eqs. (\[Eq:ChPot0\],\[Eq:ZeroCurrentYGen\]) reproduce the main features of the measured non-local conductance presented in Fig. \[Fig:Cond-nl\]a. Depending on the magnitudes of the spin-splitting field $h$ and the injector polarization $P_{\rm inj}$, we can identify three distinct parameter regimes affecting the symmetry of $g_{\rm nl}$. (i) When $h=P_{\rm inj}=0$ (blue dashed curve in Fig. \[Fig:Cond-nl\]b), the only contribution to the detector current comes from charge imbalance and $g_{\rm nl}$ is a symmetric function of the injection voltage. [In the absence of spin splitting and depairing effects, $R_T=0$ for $\varepsilon > \Delta$,]{} and hence charge imbalance decays only via inelastic scattering neglected here. [This explains the monotonic increase of $g_{nl}$ in Fig. \[Fig:Cond-nl\]b at large voltages]{}. (ii) For $P_{\rm inj}=0$ but in the presence of an applied field leading to $h\neq 0$ (black solid curve), charge relaxation is strongly enhanced due to the orbital depairing. The main long-range contribution comes from $\mu_z$ produced by the heating effect described above. The resulting $g_{\rm nl}$ is an antisymmetric function of $V_{\rm inj}$. (iii) When both $h\neq 0$ and $P_{\rm inj}\neq 0$ (red dash-dotted curve), an additional [symmetric long-range]{} contribution in $g_{\rm nl}$ results due to a thermoelectric effect at the injector. Note that in the case $h=0$, $P_{\rm inj} \neq 0$, there would be another symmetric contribution to $g_{\rm nl}$ due to the regular spin injection also present in the normal state. However, this is a short-range mode (decays via spin relaxation), and therefore does not show up beyond the spin relaxation length. [ In the experiments by @Quay2013 [@Hubler2012a; @wolf2014charge] the spin-splitting field was caused by an external magnetic field. Therefore one needs to take into account the orbital depairing effect of the magnetic field in addition to the Zeeman effect. The relative strength of the orbital depairing and the spin-splitting field is described by the dimensionless parameter $\alpha_{\rm orb}= (h\tau_{orb})^{-1}$. In Fig. \[Fig:Cond-nl\] we choose the value $\alpha_{\rm orb}=1.33$, which should correspond to the experiment by @Hubler2012a.]{} In the presence of a supercurrent, all coefficients of the matrix in Eq. are nonzero [@aikebaier2017supercurrent]. As a result, for example the spin and charge modes are directly coupled by diffusion. Spin Hanle effect {#sec:spinhanle} ----------------- In the previous sections we assume that all magnetizations and the applied field are collinear. If one lifts this assumption, the applied field leads to a precession of the injected spin around the field direction. This is the spin Hanle effect that [in the normal state]{} has been extensively studied in the literature and observed in several experiments [@PhysRevLett.55.1790; @jedema:345; @jedema2003spin; @Villamor2015; @yang2008giant]. The Hanle precession can be measured via the non-local conductance in a setup such as the one shown in Fig. \[fig:experiemnt\_beckmann\]. The non-local measured signal oscillates and decays as a function of the amplitude of the applied field. Formally the Hanle effect is described by the first term on the r.h.s of Eq. . [Indeed, one can derive the Bloch-Torrey transport equation [@Torrey1956] for the magnetic moment $\bm{m}(\varepsilon,x)={\rm Tr} (\tau_{3}\bm{\sigma}g^{K})/8$ from Eq. [@silaev2015spin]. It reads $$\label{LL-1} \frac{\partial \bm m}{\partial t} + \nabla\cdot \bm j_{s} = \gamma{\bm m}\times {\bm h}_s - {\bm m}/\tau_S .$$ Here $\gamma=-2$ is the electron gyromagnetic ratio and $\bm j_{s}$ is the spin current density tensor. In the normal state the spin relaxation $\tau_S$ and Zeeman field ${\bm h}_s$ [are]{} energy independent. This explains why the [nonlocal resistance vs. field curve]{} does not depend either on temperature or on the type of spin relaxation (magnetic or spin-orbit impurities). In contrast, [they are predicted to be strongly energy dependent in the superconducting state, and the precession and decay of the nonlocal signal disappear at $T \rightarrow 0$, whereas the shape of the curves at intermediate temperatures depends on the type of spin relaxation [@silaev2015spin].]{} Experimental evidence of the Hanle effect in the superconducting state has not been reported [so far]{}. ]{} Spin imbalance by ac excitation {#sec:acdynamics} ------------------------------- The quasiparticle $f_{T,j}$ mode — [or equivalently, the quasiparticle magnetic moment $\bm{m}(\varepsilon,x)$ above]{} — can be excited by an external ac magnetic field, which via the Zeeman coupling generates a conduction electron spin resonance [@aoi1970-tes; @maki1973-tes; @vier1983-oce; @yafet1984-ces; @nemes2000-ces]. This was recently studied experimentally in spin-split thin Al films by @quay2015-qsr. As the $f_{T,j}$ mode can relax rapidly via elastic spin-flip scattering, the linewidth seen in such experiments is generally [$\tau_S^{-1}\simeq\tau_{sn}^{-1}$]{} instead of the time scale of the long-ranged non-local spin signal. Spin-flip scattering also provides a channel via which electromagnetic fields can generate spin imbalance through the orbital coupling [@bentum1986-fia; @virtanen2016stimulated]. For high enough driving amplitude, the imbalance modifies the self-consistent $\Delta(T)$ relation, which develops additional features in the spin-split case [@eliashberg1970-fss; @virtanen2016stimulated]. Effects related to spin-splitting and relaxation can moreover be probed with tunnel junctions at low frequencies [@quay2016-fdm] or via photoassisted tunneling [@marchegiani2016-soq]. Thermoelectric effects\[thermoel\] ================================== Thermoelectric effects relate temperature differences to charge currents, and electrical potentials to heat currents. Thermoelectric effects are typically described via the linear response relation between charge and heat currents $I$, $\dot Q$ and bias voltage and temperature difference $V$ and $\Delta T$ across a contact: [^3] $$\label{eq:thermoel} \left(\begin{array}{c} I \\ \dot Q \end{array}\right) = \left(\begin{array}{cc} G & {\alpha}\\ {\alpha}& G_{\rm th} T \end{array}\right) \left(\begin{array}{c} V \\ -\Delta T/T \end{array}\right).$$ Here $G$ is the conductance and $G_{\rm th}$ the heat conductance of the contact. ${\alpha}$ is the thermoelectric coefficient. With a non-zero ${\alpha}$, electrical energy may be converted to heat or cooling, or reciprocally a temperature difference may be converted to electrical power. The efficiency of this conversion is typically described by the thermoelectric figure of merit, $$ZT = \frac{{\alpha}^2}{G_{\rm th} G T - {\alpha}^2}=\frac{S^2 GT}{\tilde G_{\rm th}},$$ where $S={\alpha}/(GT)$ is the thermopower (Seebeck coefficient) and $\tilde G_{\rm th}=G_{\rm th}-{\alpha}^2/(GT)$ is the thermal conductance at a vanishing current. In particular, the maximum efficiency of a thermoelectric heat engine is [@snyder2003] ${\rm max} \eta = \eta_{\rm Carnot} \frac{\sqrt{1+ZT}-1}{\sqrt{1+ZT}+1}$ with $\eta_{\rm Carnot}=\Delta T/T$. Maximum efficiencies of the device are obtained when $ZT\rightarrow \infty$. At or above room temperature, the record-high figures of merit are obtained in certain strongly doped semiconductor structures [@zhao2016; @kim2015]. A typical record value for those cases is $ZT\gtrsim 1\dots 2$. The traditional view of thermoelectric effects in superconductors is that if they exist, they must be very weak. In bulk superconductors, this is partially because any thermoelectrically generated quasiparticle current is easily screened by a supercurrent [@meissner27thermoel]. Alternatively, one could then measure this supercurrent via an additional constraint to the phase of the superconducting order parameter in bimetallic multiply connected structures [@Ginsburg:1944vv]. However, even this thermally created phase gradient tends to be weak, owing to the near-complete electron-hole symmetry in superconductors. @galperin74thermoel showed that $$\alpha = \alpha_N G(\Delta/T),\quad G(x)=\frac{3}{2\pi^2} \int_x^\infty \frac{y^2 dy}{\cosh^2(y/2)}, \label{eq:superalpha}$$ where the latter form comes from the reduction of the quasiparticle density in the superconducting state, and $\alpha_N$ is the value of the thermoelectric coefficient in the normal state. The precise value of ${\alpha}_N$ depends on the exact electronic spectrum. For example, for a simple quadratic dispersion ${\alpha}_N = \frac{\pi^2 G_T k_B^2 T}{6e E_F}$, where $E_F$ is the Fermi energy. At temperatures $T \ll \Delta/k_B$, ${\alpha}$ is thus expected to be a product of two small coefficients, ${\alpha}_N \propto k_B T/E_F$, and $G(\Delta/T)$. This is very small and not easy to measure quantitatively. However, superconductors do contain some ingredients for strong thermoelectric effects, because the latter typically require strongly energy dependent density of states of the charge carriers. This is provided by the BCS density of states. Hence, if one can break the electron-hole symmetry of the transport process via some mechanism, superconductors can become very strong thermoelectrics. This is precisely what happens in spin-split superconductors, as an exchange field breaks the symmetry in each spin sector, but so that the overall spin-summed energy spectrum remains electron-hole symmetric. Transport through a spin filter to a spin-split superconductor then can provide large thermoelectric effects because the two spins are weighed differently [@Ozaeta2014a; @Machon2013; @Machon2014]. We discuss these effects in this section. Charge and heat currents at a spin-polarized interface to a spin-split superconductor {#sec_thermoelectric_sp} ------------------------------------------------------------------------------------- Consider a tunnel contact [from a non-superconducting reservoir R to a superconductor S in a spin-splitting field.]{} Let us assume that the tunnel contact is magnetic, so that the conductance through it is spin-polarized. Denoting the spin-dependent conductances in the normal state by $G_\uparrow,G_\downarrow$ we can parameterize them by the total conductance $G_T=G_\uparrow+G_\downarrow$ and the spin polarization $ P=(G_\uparrow-G_\downarrow)/G_T$. The total tunneling quasiparticle charge and heat currents are now expressed as a sum over spin-dependent contributions, but otherwise of the standard form [@giaver1961; @giazotto2006]. Denoting the spin-dependent reduced density of states via $N_+=N_\uparrow+N_\downarrow$ and $N_-=N_\uparrow-N_\downarrow$ the spin-averaged tunnel currents [ can be obtained from the Keldysh component of Eq. after taking the corresponding traces:]{} $$\begin{aligned} I&=\frac{G_T}{2e}\int_{-\infty}^\infty d\varepsilon \left(N_++P N_-\right) (f_R-f_S)\label{eq:Ith} \\ \dot Q_i&=\frac{G_T}{2e}\int_{-\infty}^\infty d\varepsilon (\varepsilon-\mu_i) (N_+ + P N_-) (f_R-f_S).\label{th_qi}\end{aligned}$$ Here $f_{R/S}=n_F(E-\mu_{R/S};T_{R/S})$, $n_F(E;T)=\{\exp[E/(k_B T)]+1\}^{-1}$ are the (Fermi) functions of the reservoirs biased at potentials $\mu_{R/S}$ and temperatures $T_{R/S}$. The reduced density of states in the superconductor for spin $\sigma$ is $N_\sigma(\varepsilon)$. The heat current $\dot Q_\sigma^i$ is calculated separately for $i=R,S$, using the potential $\mu_{R/S}$, because the two heat currents differ by the Joule power $I(\mu_R-\mu_S)/e$. In the analysis below, we disregard the spin relaxation effects on the density of states, because this assumption allows for some analytically treatable limits and because it is a fair approximation in the case of often used Al samples. [The heat current from R is a non-monotonous function of voltage]{} even in the absence of spin polarization or temperature difference. In particular, for voltage $V=(\mu_R-\mu_S) \approx \Delta/e$, it is positive, i.e., reservoir $R$ cools down [@nahum1996; @leivo1996; @pekola2004]. This heat current is quadratic in the voltage, and therefore it does not result from the usual Peltier effect \[Eq. for $\dot Q$\] where the cooling power is linear in voltage. Interestingly, in the presence of spin polarization $P$ and with a non-zero spin-splitting field $h$ in the superconductor, the cooling power is nonzero even in the linear response regime, [i.e.]{} low voltages [@Ozaeta2014a]. As an example we show in Fig. \[fig:coolingpower\] the cooling power from reservoir R as a function of voltage for various values of $h$, assuming the ideal case of unit spin polarization $P=1$. Contrary to the spin-independent case, the N-FI-S element can also be used to refrigerate the superconductor. Electron refrigeration using magnetic elements have been studied by @Rouco2017. ![\[fig:coolingpower\] Cooling power from reservoir R vs. voltage for different values of the exchange field $h$, assuming a unit polarization $P=1$ at the temperature $k_B T=0.3\Delta$ close to that yielding optimal cooling for $P=h=0$. The exchange fields are given in units of $\Delta$. Changing the sign of $P$ or $h$ inverts the voltage dependence with respect to $V=0$.](coolingpower-new){width="\columnwidth"} Linear response: heat engine based on a superconductor/ferromagnet structure {#subs:linearresponseheatengine} ---------------------------------------------------------------------------- As can be seen in Fig. \[fig:coolingpower\], the simultaneous presence of the non-vanishing spin polarization $P$ and a spin-splitting field $h$ lead to a heat current that has a linear component in the voltage $V$. This component is nothing else than the Peltier effect. In the limit $k_B T \ll \Delta-h$ the linear-response coefficients evaluate to [@Ozaeta2014a] $$\begin{aligned} G&\approx G_T\sqrt{2\pi\tilde\Delta} \cosh(\tilde h)e^{-\tilde\Delta} \,,\label{eq:NFISconductance} \\ G_{\rm th}&\approx \frac{k_B G_T\Delta}{e^2}\sqrt{\frac{\pi}{2\tilde\Delta}}e^{-\tilde\Delta}\left[e^{\tilde h}(\tilde\Delta-\tilde h)^2+e^{-\tilde h}(\tilde\Delta+\tilde h)^2 \right] \,, \\ {\alpha}&\approx \frac{G_T P }{e}\sqrt{2\pi\tilde\Delta}e^{-\tilde\Delta}\left[\Delta\sinh(\tilde h)-h\cosh(\tilde h)\right] \,,\label{eq:alpha} \end{aligned}$$ where $\tilde \Delta=\Delta/(k_B T)$ and $\tilde h=h/(k_B T)$. These yield the thermopower $$S=\frac{{\alpha}}{GT} \approx \frac{P \Delta}{e T}[\tanh(\tilde{h})-h/\Delta].$$ At low temperatures the thermopower is maximized for $h=k_B T {\rm arcosh}[\Delta/(k_B T)]$, where it is $$S_{\rm max} \approx \frac{k_B}{e} P \left[\frac{\Delta}{k_B T}-{\rm arcosh}\left(\sqrt{\frac{\Delta}{k_B T}}\right)\right]. \label{eq:thermopower}$$ It can hence become much larger than the “natural scale” $k_B/e$, and even diverge towards low temperatures. However, such a divergence comes together with the vanishing of the conductance, Eq. , and therefore is in practice either cut off by circuit effects, where the impedance to the voltmeter becomes lower than the contact impedance, due to spin relaxation neglected above, or alternatively by additional contributions beyond the BCS model. The latter ones are described in more detail by @Ozaeta2014a. Nevertheless, with proper circuit design one should be able to measure a thermopower much exceeding $k_B/e$ in this setup. ![\[fig:beckmannexp\] a) Schematic setup for measuring the thermoelectrically induced current, used by @Kolenda2016. S, F, and N stand for a superconductor, ferromagnet and a normal metal, whereas FI is a ferromagnetic insulator. b) Setup used for a direct measurement of the Seebeck effect. c) Heat engine realized in a lateral setup with “n-doped” and “p-doped” junctions using a FNF trilayer with antiparallel magnetization directions. To disregard spin accumulation, the island has to be large compared to the spin relaxation length. d) Heat engine with a spin-split superconducting island. The ferromagnets can also be replaced by a normal metal if the interfaces to the superconductor contain a ferromagnetic insulator. In (c) and (d), the heating power $P_{\rm heat}$ is partially converted to “useful” work $P_{\rm work}$ dissipated on the load.](thermoeldevicesketches){width="\columnwidth"} The above theoretical predictions in the linear response regime were confirmed experimentally by @Kolenda2016 [@Kolenda2017]. In particular, they prepared a sample containing a crossing of three types of metals, a normal-metallic Cu, ferromagnetic Fe, and superconducting Al. The measured configuration is sketched in Fig. \[fig:beckmannexp\]a. The electrons in the ferromagnetic wire were heated with the heater current $I_{\rm heat}$, producing a temperature difference between the ferromagnet and the superconductor. The contact between the ferromagnet and the normal metal is ohmic and therefore the temperature difference between them is negligibly small. Then the thermoelectric current was measured as a function of the magnetic field ${\bm{B}}$ applied parallel to the ferromagnetic wire. The agreement between the experimental results and the above described tunneling theory was excellent (see Fig. \[fig:beckmannthermoel\]). The temperature difference between the ferromagnet and the superconductor was a fitting parameter, whereas the polarization $P$ was fitted from the conductance spectrum. In the experiment it was fitted to the value $P=0.08$, a modest value attributed to the thin oxide barrier between the Fe and the Al layers. In principle larger values of $P$ can be obtained by increasing the thickness of the oxide barrier [@munzenberg2004superconductor], but this of course would reduce the amplitude of the thermoelectric current. ![\[fig:beckmannthermoel\] Thermoelectric current as a function of the applied magnetic field, measured in [@Kolenda2016]. The circles show the measurement values, the solid lines show a comparison to Eq. . The three solid lines correspond to slightly different temperature differences; for further details, see [@Kolenda2016]. From @Kolenda2016. ](Beckmannfig4a-new){width="\columnwidth"} In the experiment, the thermoelectric current was measured rather than the voltage. In that case the impedance of the sample dominated that of the measurement lines. This is why the measurement yielded the exponentially low thermoelectric current, which nevertheless was [sizeable]{}. The measurement configuration in Fig. \[fig:beckmannexp\]b would have directly measured the generated voltage drop (i.e., Seebeck effect) instead of the current. This voltage results from the ratio of two exponentially small functions, the thermoelectric coefficient ${\alpha}$ and the conductance $G$, and itself is not small. Such a measurement would then tell about spurious effects, for example due to spin relaxation, or due to the presence of fluctuations or states inside the gap. These effects would limit the diverging Seebeck coefficient at low temperatures [@Ozaeta2014a]. Better still, replacing the normal metal with another superconductor with an inverse spin-splitting field, would have resulted to twice as large signal (corresponding to a series of p- and n-doped thermoelectric elements), but would not be possible to create as such with a magnetic field. The solution would be furthermore to replace the ferromagnetic wire by an FNF heterostructure \[Fig. \[fig:beckmannexp\]c, where the ferromagnets have antiparallel magnetizations, for example due to different coercive fields, and the normal metal N would serve as a spacer between them\]. To reach high figures of merit, the ferromagnetic metals should also be replaced by ferromagnetic insulators, which can reach very high values of spin polarization (see Table \[table:FIS\]), with $P$ exceeding 0.9. The island setup in Figs. \[fig:beckmannexp\](c) and (d) also realizes a thermally isolated structure, in contrast to those in (a) and (b). This allows realizing a heat engine, where the voltage measurement is replaced by the “device” to be powered with the engine, with resistance that should be matched to the thermoelectric element. If only the electrons of the ferromagnetic island are heated, the main spurious heat conduction mechanism is due to electron-phonon coupling. In that case it is advantageous to use the structure (d), because the electron-phonon heat conductance is weaker in a superconductor [@kaplan1976; @heikkilaup2017] than in a normal metal [@wellstood1994]. For example, Fig. \[fig:ZTsuper2\] shows a prediction for the resulting temperature dependence of the thermoelectric figure of merit ZT in structure (d), including this spurious heat conduction. In an optimized structure, very large ZT could thus be expected. In the picture, $g=5 k_B^5 \sqrt{2\pi} e^2 \Sigma \Omega\Delta^3/(2 G_T)$ is a dimensionless quantity characterizing the relative strength of electron-phonon coupling (characterized by $\Sigma$ [@giazotto2006]) to the tunnel coupling of the thermoelectric element in an island with volume $\Omega$. For example, for $\Omega=0.005$ $\mu$m$^3$, $\Sigma=10^{9}$ W $\mu$m$^{-3}$K$^{-5}$ and $1/G_T=30$ k$\Omega$, $g=1000$. ![\[fig:ZTsuper2\] Figure of merit in a N-FI-S-FI-N heat engine as a function of temperature for polarizations $P$ of the junction. The figure has been calculated with $h=0.5\Delta$ and $g=1000$, without calculating $\Delta$ self-consistently. The solid lines correspond to $\Gamma=10^{-6}\Delta$ and the dashed lines to $\Gamma=10^{-4}\Delta$. The figure of merit at low temperatures reaches very close to $P^2/(1-P^2)$ unless $P$ is very close to unity, but the exact temperature scale where this happens depends on the value of polarization. At the lowest temperatures $ZT$ is limited by another spurious heat conduction process, due to nonzero density of states inside the gap, described here by the Dynes $\Gamma$ parameter. ](ZTfigure){width="\columnwidth"} Note that it is really the presence of the spurious electron-phonon heat conduction that limits the highest available values of $ZT$. Often such spurious mechanisms are disregarded from the theoretical analysis, for example in the case of quantum dots [@hwang16]. [Even if the true figure of merit of the type of heat engine discussed above can be made high, these systems cannot obviously be used to replace room-temperature thermoelectric devices to be applied for example in energy harvesting. However, there are other applications where the large figure of merit may turn out to be essential. For example, this type of thermoelectric heat engine can be used for thermal radiation sensing at low temperatures [@giazotto2006; @heikkilaup2017]. Another possible use of the thermoelectric effects would be in non-invasive low-temperature thermometry [@giazotto2015b], where the temperature (difference) profiles could be read from the thermopower, without having to apply currents. In a scanning mode this would hence be a low-temperature version of the method used by @menges16.]{} Note that the above discussion disregards the effect of spin-orbit or spin-flip scattering on the superconducting state. It limits $ZT$ especially in heavy-metal superconductors. The associated effects were considered by @bergeret2017nonequilibrium [@rezaei2017spin]. Spin Seebeck effect \[subs:spinseebeck\] ---------------------------------------- Besides the large thermoelectric effect, the contact between spin-split superconductors with other conducting materials can exhibit a large (longitudinal) spin Seebeck effect, where a temperature difference drives spin currents to/from the spin-split superconductor [@Ozaeta2014a]. In this case the charge, heat, spin and spin heat currents are described by the full [[@onsager31; @jacquod12; @Machon2013]]{} Onsager linear-response matrix $$\label{Eq:Onsager} \left( \begin{array}{ccc} I \\ \dot Q \\ I_s \\ \dot Q_s \end{array} \right) = \left( \begin{array}{cccc} G & \alpha & PG & \tilde \alpha \\ \alpha & G_{\rm th}T & \tilde \alpha & PG_{\rm th}T \\ PG & \tilde \alpha & G & \alpha \\ \tilde \alpha & PG_{\rm th}T & \alpha & G_{\rm th}T \\ \end{array} \right) \left(\begin{array}{ccc} V \\ - \Delta T/T \\ V_s/2 \\ - \Delta T_s/2T \end{array} \right),$$ where for $k_B T \ll \Delta-h$ the coefficients $G$, $G_{\rm th}$ and $\alpha$ are given in Eqs. (\[eq:NFISconductance\]-\[eq:alpha\]), and $\tilde \alpha=\alpha/P$. Here $V_s$ and $\Delta T_s$ refer to spin-dependent biases [@bergeret2017nonequilibrium]. The spin currents induced in the case of two spin-split superconductors, and the additional effects of Josephson coupling, magnetization texture and spin-orbit effects are discussed by @linder2016 [@bathen2017]. When either of the two materials realizes an island, the spin current can be converted into a spin accumulation $\mu_z$ that is determined from the balance between thermally induced spin currents and spin relaxation within the island. The above discussion on heat engines assumes a structure size much longer than the spin-relaxation length, and hence disregards this spin accumulation. The effect of the thermally induced spin accumulation on the superconducting gap was considered by @bobkova2017, who predicted the associated changes in the critical temperature. This spin Seebeck effect should be contrasted to the analogous phenomenon discussed in non-superconducting materials [@Uchida2014]. There, [a major contribution to the spin Seebeck signal is due to the thermally induced spin pumping [@hoffman2013]]{}. Thermophase in a S(FI)S contact {#sec:thermophase} ------------------------------- The large thermoelectric effect described [above]{} allows for a large thermally induced phase gradient. This was theoretically investigated by @giazotto2015. The total current in this case consists of the sum of a thermoelectric current $I_{\rm th}$ and the supercurrent, $$I=I_{\rm th} + I_c \sin(\varphi),$$ where $I_{\rm th}$ is obtained from and $I_c$ is the critical current for the junction with a phase difference $\varphi$ of the order parameters across the contact. The critical current is proportional to $\sqrt{1-P^2}$ [@Bergeret2012] and depends on the spin-splitting field in S [@bergeret2014] In an electrically open configuration, the two currents must cancel, and instead a [thermophase]{} $\varphi^{\rm th}$ develops across the junction. This is obtained from $$\sin(\varphi^{\rm th}) = -\frac{I_{\rm th}}{I_c}. \label{eq:thermophase}$$ The thermophase can be detected using a bimetallic loop with two contacts, characterized by critical currents $I_{c1,2}$ and thermophases $\varphi_{1,2}^{\rm th}$. For non-zero exchange field and spin polarization $P$, the resulting thermophases can be much larger than in ordinary bulk superconductors. Hence the temperature dependence of the inductances play a more minor role [than in the case of superconductors without spin splitting [@vanharlingen80thermoel; @Shelly16]]{}. For junctions with non-equal thermophases and for negligible loop inductance (in practice, $2e L I_{c1,2} /\hbar \ll 1$) in the absence of an external flux the circulating current is $$I_{\rm circ} = \frac{I_{c1} I_{c2}}{I_{c1}+I_{c2}} \left[\sin(\varphi^{\rm th}_1)-\sin(\varphi^{\rm th}_2)\right].$$ In the case of symmetric junctions both thermophases are the same and the circulating current in the absence of an external flux vanishes. However, as discussed by @giazotto2015, the thermoelectric current affects the response of the circulating current to the external flux, allowing for their measurement also in that case. Equation requires that both sides of the equation have an absolute value of at most unity, i.e., $\|I_{\rm th}\| < I_c$. For a very large thermoelectric current, its cancellation with a supercurrent is no longer possible, and instead a voltage across the contact forms. In this case the direct current response of the junction is more similar to the case discussed above in the linear response limit for a N-FI-S junction. This regime was investigated in detail by @linder2016. Moreover, the nonvanishing dc voltage across the superconducting junction leads to Josephson oscillations at the frequency $2eV/h$, where $h$ is the Planck constant. Hence, the device can be used as a temperature (difference) to frequency converter as discussed in more detail by @giazotto2015b. Summary and Outlook {#sec:outlook} =================== This review focuses on transport and thermal properties of superconducting hybrid structures with a spin-split density of states. Such a splitting can be achieved either by an external magnetic field, or, more interestingly, by placing a ferromagnetic insulator (FI) adjacent to a superconducting layer (S) (Sec. \[sec-superwithh\]). We discuss several experimental situations with the help of a theoretical framework (see Sec. \[sec:Usadel\] and \[sec:noneq\_quasiclas\]) based on the quasiclassical formalism, with which one can account for both thermodynamical and nonequilibrium properties of such hybrid structures. In order to account for effects beyond quasiclassics, as for example strong spin polarization, we combine the quasiclassical equations with effective boundary conditions. Out-of equilibrium superconductivity by itself leads to a decoupling between the charge and energy degrees of freedom of the electronic transport. In this review we show that the combination between superconductivity and magnetism requires on one hand a description of additional nonequilibrium modes, spin and spin energy, and on the other hand to couples them all. This leads to novel and intriguing phenomena discussed in this review with direct impact in latest research activities and proposed future technologies based on superconductors and spin dependent-fields [@eschrig2011spin; @eschrig2015spin; @linder2015superconducting]. By using the theoretical formalism presented in this review one can predict and explain phenomena such as the spin injection and relaxation (Sec. \[spininjection\]) in superconductors with an intrinsic exchange field along with their consequences in the transport properties. We also discuss a number of striking thermoelectric effects in superconductors with a spin-splitting field (Sec. \[thermoel\]). The best scenario for the phenomena and applications discussed here, and in particular for the thermoelectric effects, are FI-S systems where the spin splitting can be achieved without the need of an applied magnetic field. Hence it becomes important to look for ideal FI-S material combinations. So far europium chalcogenides (EuO, EuS and EuSe) together with Aluminum films have shown large splittings and hence these are the best combination. In addition, thin films of EuO or EuS can be used as almost perfect spin filters (see Table \[table:FIS\]) and hence they are good candidates for realizing the near-optimal heat engines proposed in Sec. \[thermoel\]. One of the main challenges from this perspective is to find FI-S combinations with large superconducting critical temperature and simultaneously a large spin splitting. Superconductors like Nb or Pb on the one hand increase $T_C$ with respect to Al-based structures, but on the other the spin-orbit coupling may spoil the sharp splitting as discussed in Sec. \[sec:SCwithh\]. Recent experiments on GdN-NbN suggest large splittings [@PhysRevB.92.180510] but further research in this direction is needed. In Sec. \[sec:acdynamics\] we [briefly]{} discuss the dynamics of spin-split superconductors in rf fields. Historically, magnetic resonance effects in superconductors are well studied, but fewer experiments have probed spin-split thin films. Besides the effects discussed in this review, several theoretical studies made striking predictions in mesoscopic systems with spin-split superconductors, such as the creation of highly polarized spin currents [@absolute_spin_valve; @PhysRevB.77.132501; @giazotto2013quantum], large supercurrents in FI-S-I-S-FI junctions [@bergeret2001enhancement], junctions with switchable current-phase relations [@strambini2015mesoscopic], and an almost ideal heat valve based on S-FI elements [@Giazotto:2013ei]. Although many of the transport phenomena in spin-split superconductors are now well-understood, we foresee a number of exciting avenues for future research. One further perspective of the present work is the extension of the Keldysh quasiclassical formalism used in this review to include magneto-electric effects associated with the spin-orbit coupling (SOC). For a linear in momentum SOC the generalization of this can be done by introducing an effective SU(2) gauge potential. The quasiclassical equations in this case have been derived by @Bergeret:2013il [@bergeret2014spin; @bergeret2016manifestation]. Effects such as the spin-Hall and spin-galvanic effect in superconductors have been studied in the equilibrium case [@konschelle2015theory]. Extending these results to a nonequilibrium situation, and also to time-dependent fields, would be an interesting further development and would allow for a detailed study of the well-controlled non-linearities associated to these effects in superconductors. First steps in this direction have been taken in [@espedal2017]. Recent discoveries of skyrmionic states in chiral magnets [@Nagaosa2012] have attracted a lot of attention due to the effects resulting from the interplay of magnetism and SOC [@Soumyanarayanan2016] which can induce chiral Dzyaloshinskii-Moriya interactions between magnetic moments. Currently it is very interesting to study these effects in the presence of the additional component — superconductivity, when the exchange interaction is mediated by the Cooper pairs [@DEGENNES196610]. One can expect that in such systems superconductivity can induce a non-trivial magnetic ordering and dynamics. These effects can show up in various systems including ferromagnet/superconductor bilayers, surface magnetic adatoms and bulk magnetic impurities inducing the localized Yu-Shiba-Rusinov states modified by the SOC [@PhysRevLett.115.116602]. Superconducting structures with strong spin-orbit coupling and exchange fields are also of high interest in view of engineering a platform for realization of topological phases and Majorana bound states [@qi2011-tia; @beenakker2013-smf; @alicea2012; @hasan2010-cti]. Understanding and controlling the behavior and relaxation of nonequilibrium quasiparticles in these systems is also of importance, not least because of their influence on the prospects of solid-state topological quantum computation [@nayak2008]. This review focuses exclusively on the nonequilibrium properties of superconductors in proximity to magnets. We expect the inclusion of the magnetization dynamics and its coupling to the electronic degrees of freedom via the reciprocal effects of spin transfer torque and spin pumping [@tserkovnyak2005] in the far-from equilibrium regime to lead to completely new type of physics, as the two types of order parameters affect each other. The coupling of supercurrent on magnetization dynamics and texture has been studied during the past decade [@waintal2002; @houzet2008; @richard2012-aci], but the work where both systems are out of equilibrium has been mainly concentrated on Josephson junctions [@holmqvist2011; @hikino2011; @mai2011; @holmqvist2014; @kulagina2014] and much less attention has been paid to quasiparticle effects [@skadsem2011; @linder2012; @trif2013-dme]. [Besides the rich physics offered by spin-split superconductors, they have been long used as tools to characterize equilibrium properties of magnets, especially their spin polarization. In this review (see end of Sec. \[subs:linearresponseheatengine\]) we outline two further possibilities related to their large thermoelectric response: accurate radiation sensing and non-invasive scanning thermometry. We believe there are also many other avenues to be uncovered, opened by the possibility for realizing a controlled combination of magnetism and superconductivity.]{} We thank Faluke Aikebaier, Marco Aprili, Detlef Beckmann, Wolfgang Belzig, Irina Bobkova, Alexander Bobkov, Matthias Eschrig, Yuri Galperin, Francesco Giazotto, Vitaly Golovach, Kalle Kansanen, Alexander Mel’nikov, Jagadeesh Moodera, Risto Ojajärvi, Asier Ozaeta, Charis Quay, Jason Robinson, Mikel Rouco, and Elia Strambini for useful discussions. This work was supported by the Academy of Finland Center of Excellence (Project No. 284594), Research Fellow (Project No. 297439) and Key Funding (Project No. 305256) programs, the European Research Council (Grant No. 240362-Heattronics), the Spanish Ministerio de Economía y Competitividad (MINECO) (Projects No. FIS2014-55987-P and FIS2017-82804-P), the European Research Council under the European Union’s Seventh Framework Program (FP7/2007- 2013)/ERC Grant agreement No. 615187-COMANCHE. [^1]: The term “mode” here refers to the changes of the electron distribution function with respect to its equilibrium form. It should be distinguished from collective modes such as the @carlson1973superconducting or the amplitude mode [@higgs64] that affect the response of superconductors at temperatures close to the critical temperature or at high frequencies. [^2]: Here we assume a unique spin polarization direction. In the most general case the distribution function has all spin components $ \hat{f}=f_{L}\hat{1}+f_{T}\tau_{3}+\sum_{j}(f_{Tj}\sigma_{j}+f_{Lj}\sigma_{j}\tau_{3})\label{eq:distribution_function}$. [^3]: [In the case of thermoelectric effects, it is customary to talk about heat currents $\dot Q$ instead of energy currents $\dot U$, and we adapt this convention here. These are related by [@ashcroftmermin] $\dot Q=\dot U-\mu I/e$, where $\mu$ is a reference energy compared to the Fermi level. At linear response we can set $\mu=0$ in which case $\dot Q=\dot U$. On the other hand, when considering heat balance at non-vanishing voltages as in Sec. \[sec\_thermoelectric\_sp\], the two are not the same and rather the heat current $\dot Q$ should be used.]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Computations in general relativity have revealed an interesting phase diagram for the black hole – black string phase transition, with three different black objects present for a range of mass values. We can add charges to this system by ‘boosting’ plus dualities; this makes only kinematic changes in the gravity computation but has the virtue of bringing the system into the near-extremal domain where a microscopic model can be conjectured. When the compactification radius is very large or very small then we get the microscopic models of 4+1 dimensional near-extremal holes and 3+1 dimensional near-extremal holes respectively (the latter is a uniform black string in 4+1 dimensions). We propose a simple model that interpolates between these limits and reproduces most of the features of the phase diagram. These results should help us understand how ‘fractionation’ of branes works in general situations.' --- [A Microscopic Model for the Black hole – Black string\ Phase Transition]{}\ $^{a,}$[^1], [Stefano Giusto]{}$^{b,}$[^2] [and Samir D. Mathur]{}$^{a,}$[^3]\ $^{a}$Department of Physics,\ The Ohio State University,\ Columbus, Ohio, USA 43210\ $^{b}$Department of Physics,\ University of Toronto,\ Toronto, Ontario, Canada M5S 1A7\ Introduction {#intr} ============ Semiclassical physics gives us the entropy and Hawking radiation rates of black holes. In string theory we can understand extremal and near-extremal holes in terms of branes, and thereby reproduce the entropy and radiation from a microscopic description. In this paper we will consider another property of black holes that can be described in classical general relativity: the transition between black holes and black strings when a transverse circle is made large or small. We suggest a simple microscopic picture of this transition, which will reproduce many of the broad features of the ‘phase diagram’ of the system. The ‘phase diagram’ ------------------- In Fig.\[fig:BH-BS\_PT\](a) we depict a small black hole in a spacetime with a compact transverse direction which we will call $z$; the length of this transverse circle is $L$. In Fig.\[fig:BH-BS\_PT\](b) we depict a hole with a larger mass; the horizon now feels a significant distortion from the compactification. In Fig.\[fig:BH-BS\_PT\](c) we have increased the mass still further; the black hole horizon has now turned into a ‘black string’ horizon. ![(a) A small black hole in a space with a compact circle of length $L$ (b) The horizon distorts when the mass is increased so that its radius becomes comparable to $L$ (c) At still larger masses we get a black string which wraps uniformly around the compact circle.[]{data-label="fig:BH-BS_PT"}](BH-BS_Phase_Transition.eps){width="5in"} (a)(b)(c) The black string has a nonzero ‘tension’ ${\cal T}$ along the compact circle. The hole in Fig.\[fig:BH-BS\_PT\](b) will also have some tension in this direction. When we take the hole to be very small, as in Fig.\[fig:BH-BS\_PT\](a), the hole does not notice the compactification, and this tension becomes small too, vanishing for infinitesimal holes. The relevant parameter here is $r_0/L$, where $r_0$ is the radius of the horizon. When this parameter is small the effects of the compactification go away, and the tension becomes ignorable in the determination of the geometry of the hole. Through a large number of studies, some analytical and some numerical, an interesting ‘phase diagram’ has emerged for this system [@GL]-[@sahakian]. In Fig.\[fig:nVSmu\_gravity\](a) we reproduce this diagram for the case of 5-D; i.e. 4 noncompact space dimensions and the compact circle of length $L$. In Fig.\[fig:nVSmu\_gravity\](b) we give the diagram for 4-D; i.e., 3 noncompact space directions and the compact circle. (Our interest will be in the 4-D case, but we give the diagram for the 5-D case as well since it is more complete – numerical computations are easier in this case since the gravitational fields fall off faster at infinity.) ![The phase diagram for (a) 4 noncompact space directions and (b) 3 noncompact space directions. The vertical axis measures the dimensionless relative tension and and the horizontal axis gives a dimensionless mass parameter. The solid line denotes the uniform black string, the dashed line denotes the black hole, and the dot-dashed line denotes the non-uniform black string.[]{data-label="fig:nVSmu_gravity"}](6DnVSmu_gravity.eps "fig:"){width="3in"} ![The phase diagram for (a) 4 noncompact space directions and (b) 3 noncompact space directions. The vertical axis measures the dimensionless relative tension and and the horizontal axis gives a dimensionless mass parameter. The solid line denotes the uniform black string, the dashed line denotes the black hole, and the dot-dashed line denotes the non-uniform black string.[]{data-label="fig:nVSmu_gravity"}](5DnVSmu_gravity.eps "fig:"){width="3in"}\ (a)(b) On the horizontal axis we have the mass, scaled by some constant and termed $\mu$. On the vertical axis we have a dimensionless tension $n={{\cal T} L\over M}$. For small $\mu$ we have just the ‘black hole’ phase, with $n\r 0$ for $\mu\r 0$. At very large $\mu$ we have the ‘black string’ phase, with $n$ independent of $\mu$. But the most striking feature of these phase diagrams is that we do not see just a black hole and a black string; for a range of $\mu$ there is a third branch, called the ‘non-uniform black string’. In each of these phases we can compute the Bekenstein entropy from the area of the horizon. In Fig.\[fig:sVSmu\_gravity\](a),(b) we reproduce the entropy graphs. The horizontal axis again gives the mass, and on the vertical axis we plot ${s\over s_\mathrm{us}}$, where $s_\mathrm{us}$ is the entropy of the uniform string. Thus the uniform string branch just gives a horizontal line at height unity. The black hole branch is seen to have higher entropy at low mass, and lower entropy at high mass. The non-uniform string is seen to have an entropy that is always lower than the other two branches. ![The entropies for (a) 4 noncompact space directions and (b) 3 noncompact space directions. The vertical axis gives the ratio of the entropy to the entropy of the uniform black string, and the horizontal axis gives a dimensional mass. The three kinds of lines label the three phases in the same way as in Fig.(\[fig:nVSmu\_gravity\])[]{data-label="fig:sVSmu_gravity"}](6DsVSmu_gravity.eps "fig:"){width="3in"} ![The entropies for (a) 4 noncompact space directions and (b) 3 noncompact space directions. The vertical axis gives the ratio of the entropy to the entropy of the uniform black string, and the horizontal axis gives a dimensional mass. The three kinds of lines label the three phases in the same way as in Fig.(\[fig:nVSmu\_gravity\])[]{data-label="fig:sVSmu_gravity"}](5DsVSmu_gravity.eps "fig:"){width="3in"}\ (a)(b) The microscopic picture ----------------------- Our goal is to see if this phase diagram can be understood in terms of a microscopic description of black holes. The microscopic picture works best for extremal and near-extremal objects. This is because supersymmetry prevents strong dependences on coupling and thus allows computations in terms of weakly interacting excitations. So our first goal is to map the phase diagram of neutral objects to the phase diagram for near-extremal objects. A detailed study of this map was made in [@harmarkoberscharged; @hkor], and we will use many of their results. We will work in type IIB string theory. Consider the compactification M\_[9,1]{}M\_[4,1]{}T\^4S\^1 The compact circle of length $L$ will be chosen from the spatial directions in $M_{4,1}$, we will call it $\t S^1$. We will add large D1 and D5 charges to the neutral object. The addition of charges is done by ‘boosting’ the system in the compact direction $S^1$; this gives a momentum charge P in that direction. Dualities convert this to a D1 charge along $S^1$. We boost again along $S^1$, and dualize to get a D5 charge on $T^4\times S^1$. First let $L\r\infty$, so we have a black object in 4+1 noncompact spacetime dimensions. This is the near extremal D1-D5 black hole studied in [@cm; @ms]. In the microscopic description, the D1-D5 bound state produces an ‘effective string’ winding along $S^1$ with total winding number $n_1n_5$. The non-extremal energy of the system is carried by a dilute gas of $P, \bar P$ excitations along the $S^1$ direction, running along this effective string. The entropy of these excitations is given by S=2(+) \[en1\] With no net $P$ charge, we have $n_p=\bar n_p$, and the value of $n_p$ is given by the non-extremal energy. So we have a microscopic picture of ‘small black holes’, i.e. black holes where $r_0/L$ is small. Now let us imagine that $L$ is small. Then the effective compactification is M\_[9,1]{}M\_[3,1]{}T\^4S\^1S\^1 \[com2\] With this compactification we can make an extremal black hole with 4 kinds of charges: D1, D5 as above, P along $S^1$, and KK monopoles which have $\t S^1$ as their nontrivially fibered circle. What we have however is two nonzero charges (D1, D5) plus nonextremal energy. Note that the 4 kinds of charges can be permuted in all possible ways by dualities. Using this fact, we can write the entropy of this near-extremal system from the expression found in [@malda], getting S=2(+)(+) \[en2\] If the net P, KK charges are zero then $n_p=\bar n_p$, $n_k=\bar n_k$, and the values of $n_p, n_k$ are determined by extremising $S$ over all possibilities for given total nonextremal energy. This entropy describes a black object that is uniformly smeared in all the compact directions of (\[com2\]), so it is a ‘uniform black string’ in the direction $\t S^1$. One can write an expression similar to (\[en1\]),(\[en2\]) for the neutral hole as well [@hms; @hlm]. But we have chosen to add charges to our system and look at the near-extremal systems (\[en1\]),(\[en2\]) because for these cases we also have a microscopic derivation of low energy Hawking radiation [@dm1; @ms; @km]. Thus not only is the count of states correct, their microscopic description must also be correct because they lead to the expected dynamics. Let us return to our problem of interest. For large $L$, we have the ‘black hole phase’, described by (\[en1\]), and for small $L$ we have the ‘uniform black string’ phase, described by (\[en2\]). What happens when we reduce $L$ from large values to small values? We seek a microscopic picture of the intermediate $L$ regime, which will help us understand the three phases that emerge in the gravity computation; in particular we wish to understand the ‘non-uniform black string’ in the microscopic picture. Outline of our computations --------------------------- We proceed in the following steps: \(A) A basic property of gravity solutions with a horizon is the Smarr relation. This is a scaling relation that relates the tension ${\cal T}$ of the solution to the mass, charge and entropy. Since our microscopic model must describe a black object with a horizon, the Smarr relation must be satisfied by the model. We thus recall the Smarr relation and sketch its derivation. \(B) The gravity solution has a nontrivial metric in $M_{4,1}$. The compact directions $T^4\times S^1$ can be ‘added on trivially’ to the neutral solution in $M_{4,1}$; i.e. the metric is flat and constant in these directions. From Einstein’s equations one then finds that the tension in these ‘trivial’ directions must be [nonzero]{} in general. Thus from a microscopic perspective, these directions are not so ‘trivial’: one must wrap branes around these directions to obtain the required tension. The tension in these ‘trivial’ directions is also important in getting the correct changes of physical parameters under the ‘boosting’ which adds charge. We recall the equations that relate old and new quantities upon boosting; these were derived in [@harmarkoberscharged], but we also supply a physical derivation valid at weak gravitational coupling since this makes some of the physics more transparent. \(C) We next make a microscopic model that describes the transition from the black hole to the black string. The entropy (\[en1\]) arises from momentum modes $n_p, \bar n_p$ that are fractionated in units of $1/n_1n_5$ by the presence of the D1, D5 branes [@dmfrac; @maldasuss]. The entropy (\[en2\]) comes from two kinds of excitations (momentum and KK monopoles) which are again fractionated by the D1, D5 brane charges [@malda]. We make a simple model where a part $E_1$ of the nonextremal energy $E$ goes to creating entropy of the type (\[en1\]) and the rest $E-E_1$ goes to creating entropy of type (\[en2\]). In this process we assume that a fraction $(1-x)$ of the ‘fractionating charge’ $n_1n_5$ goes to fractionating the excitations used in (\[en1\]), and the fraction $x$ goes to fractionating the excitations for (\[en2\]). We then extremise over $E_1, x$ to find the preferred state of the microscopic system. We find that below some energy $E$ there is just one phase, the ‘black hole’ phase, where all entropy is of type (\[en1\]). But for $E_1$ greater than a certain critical value, there are three extrema of the entropy. One is a ‘black hole’ with entropy entirely of type (\[en1\]), one is a ‘uniform black string’ which has entropy entirely of type (\[en2\]), and a third exremum where the entropy is partly of type (\[en1\]) and partly of type (\[en2\]). This last extremum is a saddle point of the entropy, and we identify it with the non-uniform black string. The tension and entropy graphs for this very simple microscopic model are given in section \[microsection\], in Fig.\[fig:4DnVSmu\_micro\] and Fig.\[fig:4DsVSmu\_micro\] respectively. \(D) While this simple model reproduces the three phases seen in the gravity computations, we notice one simple qualitative difference between the microscopic and gravity computations: the microscopic model has a vanishing tension in the black hole phase, while the gravity description gives a tension that rises from zero as the mass is increased from zero. We argue that this tension should be found in the microscopic picture if we include the ‘twist operator interaction’ in the microscopic CFT, something we have not done in our leading order computation. An indirect approach to computing the tension of the black hole branch was taken in [@costaperry; @hkor], and we extend this approach here. One computes the tension from the gravity description, and uses this to read off the changes in excitations in the microscopic description. In the microscopic description we can imagine integrating out the effects of the KK monopole pairs (which are only weakly excited at large $L$) to obtain an ‘effective interaction’ between the $P, \bar P$ excitations of the black hole. We first use the leading order gravity result to find this interaction to leading order in the microscopic picture. Then we use analyticity in the microscopic variables to predict the tension to next order on the gravity side, and observe agreement with the known gravity computation at that order. We thus obtain some understanding of tension in the black hole phase. We note some agreements and also some disagreements between the gravity phase diagram and the microscopic diagram. We hope to return in future work to exploring further the interactions in the microscopic picture in order to obtain a more accurate phase diagram. Some basic relations ==================== The Smarr relation ------------------ We recall the Smarr relation [@harmarkobersII], and sketch a derivation; derivations along these lines can be found for example in [@kolsorpirI]. The full theory of quantum gravity has an inbuilt length scale – the planck length $l_p$. But $l_p$ involves $\hbar$, and the classical theory does not have this length scale. The action $\int \sqrt{-g}~ R$ does not have any natural length scale, so if we have one solution to the vacuum equations, we can obtain another by uniformly scaling up all lengths by the same factor. Let the length of the compact direction $z$ change as L L+dL \[zero\] Consider a black hole in a spacetime with one compact direction $z$ of length $L$. Let the number of noncompact spacetime dimensions be $d$; thus including the compact circle $z$ the total spacetime dimension is $d+1$. In this dimension the mass $M$ scales as $L^{d-2}$. (The mass appears in the metric through the combination $G_N^{(d+1)}M$, which has units $L^{d-2}$; $G_N^{(d+1)}$ is held fixed, and thus $M$ scales as $L^{d-2}$.) Thus we will get M M(1+[dLL]{})\^[d-2]{}, dM=M(d-2)[dLL]{} \[one\] The entropy is proportional to the surface area of the horizon. The horizon is a $d-1$ dimensional surface, so it will scale as S S(1+[dLL]{})\^[d-1]{}, dS=S(d-1)[dLL]{} \[two\] The charge $Q$ appears in the metric the same way that the mass does, so we have Q Q(1+[dLL]{})\^[d-2]{}, dQ=Q(d-2)[dLL]{} \[three\] Now assume that the solution satisfies the first law of thermodynamics (this will be true for all black objects) dE=TdS+[T]{} dL+ dQ where ${\cal T}$ is the ‘tension’ in the $z$ direction and $\phi$ the potential at the horizon. This law will be satisfied in particular for the changes (\[zero\])-(\[three\]), since these changes lead from one valid solution to another. We thus get E(d-2)=TS(d-1)+[T]{} L+Q(d-2) or TS=[(d-2)(E-Q)-[T]{} Ld -1]{} \[smarrnewp\] This is the Smarr relation. Behavior under boosting {#boostsec} ----------------------- Consider the metric of a black object, by which we mean any solution of gravity with a single connected horizon. We assume that the solution is neutral, i.e. it carries no charge. The metric is of the form ds\^2=-Udt\^2+ds\_B\^2 \[original\] where $U$ is independent of $t$ and the “base metric” $ds_B^2$ involves only the spatial coordinates. The horizon is at the location $U=0$. We can add one or more ‘trivial directions’ to this metric; we single out a special one called $y$ for later use: ds\^2=-Udt\^2+ds\_B\^2+dy\^2+dz\_adz\_a \[neutral\] We will think of the added directions as compact circles. In particular let $0\le y<L_y$. We can start with the solution (\[neutral\]) and add charges to it by boosting plus dualities. Thus writing t=t’+y’, y=t’+ y’\[boost\] we get ds\^2=H\^[-1]{}\[-Udt’\^2+H\^2 (dy’-A\_[t’]{} dt’)\^2\]+ds\_B\^2+dz\_adz\_a \[charged\] where H=1+(1-U) \^2, A\_[t’]{}=(H\^[-1]{}-1) This process adds ‘momentum charge $P_y$’. We can change this to a different kind of charge by S,T dualities, and add further charges by boosting. We will write down a 2-charge solution in the Appendix, but for now we will look at the simple case having just $P_y$ charge. We can extract the following properties of the boosted solution; the derivations can be seen from the 2-charge case discussed in the Appendix. Let the neutral solution (\[neutral\]) have mass $M$, entropy $S$, and temperature $T$. (The temperature is extracted from the surface gravity at the horizon in the usual way.) The part $ds_B^2$ describes a $(d-1)$ dimensional noncompact space and a compact direction $z$, and we will let ${\cal T}$ denote the tension in the $z$ direction. What are the properties of the charged solution (\[charged\])? Note that $y$ was a compact direction, and we do not have the boost symmetry (\[boost\]) for compact $y$. But since the classical solution is homogeneous in $y$, we can lift the solution to the covering space of $y$, apply the boost, and then ‘recompactify’ the new coordinate $y'$. Since the boosted solution is again homogeneous in $y'$, we can pick any coordinate length to compactify $y'$. Let us choose $0\le y'<L_y$, thus keeping the length of the compact circle the same as in the neutral solution. Then we find the following [@harmarkoberscharged]: \(a) The mass $M'$ of the charged solution (\[charged\]) is given by M’=M(1+[d-2-nd-1]{} \^2) \[massgr\] where the relative tension $n$ is defined as n \[relative\] \(b) The tension ${\cal T}'$ of the charged solution is the same as that of the neutral solution ’=[T]{} \[tensiongr\] \(c) The charge of the solution is Q=M [d-2-nd-1]{} \[chargegr\] \(d) The entropy $S'$ of the charged solution is S’=S \[entropygr\] \(e) The temperature $T'$ of the charged solution is T’=[Tc]{} Note that T’S’=TS \(f) The electric potential at the horizon is = Understanding the ‘boost relations’ ----------------------------------- The above relations (a)-(f) between the neutral and charged solutions are straightforward to derive in the gravity description of black objects. But our goal is to understand black objects from a microscopic point of view. From this ‘matter’ point of view these relations appear somewhat nontrivial. They tell us the thermodynamic properties of a charged system once we know the properties of the neutral system. But abstractly speaking, there is no general relation between charged and neutral systems. So we must be looking at very special systems, and our microscopic model must be made to reproduce the properties of such special systems. To understand these properties better, we now rederive the relations of the above section from a microscopic perspective. The matter stress tensor ------------------------ Consider the neutral system. Let the total spacetime dimension be $D$. The Einstein equations are R\_[AB]{}-[12]{} g\_[AB]{} R=8G T\_[AB]{} We can rewrite these as R\_[AB]{}=8GT\_[AB]{} where T\_[AB]{}=T\_[AB]{}-[1D-2]{} g\_[AB]{} T We will now assume that the gravitational field is weak. This will not be the case in general for the systems we finally consider, but we are only trying to get a physical feeling for the relations (a)-(f) which are already known to be correct, and this weak-field assumption will allow us to extract some basic ideas in a simple way. Writing $g_{AB}=\eta_{AB}+h_{AB}$ we get (in the gauge $h_{AB,}{}^B=0$) -[12]{}h\_[AB]{}-[12]{} h\_[,AB]{}=8GT\_[AB]{} \[final\] Recall that we had started with a metric of the form (\[original\]) and added ‘trivial’ directions to obtain the metric (\[neutral\]). It seems from the gravity point of view that these extra directions play no role in the physics of the black object. But from a microscopic viewpoint we see a different story. Let the weak gravity system (\[final\]) describe this situation where we have trivially added directions $y, z_a$, and the metric is nontrivial in the noncompact directions $x_\mu$ and the compact direction $z$. Consider one of the ‘trivial’ directions, say $y$. First, we have [homogeneity]{} in $y$, so fields do not depend on $y$. Thus in (\[final\]) we will get $h_{,yy}=0$. Second, we have [uniformity]{} of the $y$ circle, which means that the size of this circle does not change from place to place. This gives $h_{yy}=0$. Eq. (\[final\]) then gives $\t T_{yy}=0$, which is T\_[yy]{}-[1D-2]{} T=T\_[yy]{}-[1D-2]{}\[-T\_[00]{}+T\_[zz]{}+T\_[yy]{}+T\_[z\_az\_a]{}\]=0 \[basic\] Note that $T\ne 0$ in general, so $T_{yy}\ne 0$ in general. This means that our microscopic model [cannot]{} consist of branes that extend only in the directions $x_i, z$ while ignoring the trivial directions $y, z_a$. Branes have a tension only along their worldvolumes, and such a model would give $T_{yy}=0$. Instead, we must have at least one kind of brane wrapping each of the ‘trivial’ directions. As an example consider the Schwarzschild solution in 4+1 noncompact dimensions. We can add the other 5 directions of string theory as ‘trivial directions’ in the gravity solution. But a microscopic model for this system can be written in terms of branes and antibranes of type D1, D5, P [@hms], and these objects are seen to wrap around the 5 ‘trivial’ directions. The mass of the neutral system is M= T\_[00]{} The tension ${\cal T}$ along the $z$ direction is =-[ T\_[zz]{}L]{} First note that under the boost (\[boost\]) we have $T'_{zz}=T_{zz}$, so the tension remains unchanged ’=[T]{} which is the relation (\[tensiongr\]). Next, note that M’= T’\_[00]{}= (T\_[00]{}\^2+T\_[yy]{} \^2) \[mass\] In (\[basic\]) the ‘trivial’ directions $y, z_a$ are all on the same footing, so $T_{yy}=T_{z_az_a}$. We then find T\_[yy]{}=-[1d-1]{}(T\_[00]{}-T\_[zz]{}) \[tyy\] Substituting this in (\[mass\]) we find M’=M \^2-[1d-1]{}(M+[T]{} L)\^2=M+M [d-2 -nd-1]{} \^2\[massnew\] where we have used the definition $n={\cal T} L/M$. Thus we reproduce (\[massgr\]). The charge $P_y$ is Q= T’\_[0y]{}= (T\_[00]{} +T\_[yy]{} ) Using (\[tyy\]) we find Q=[(d-2)M -[T]{} Ld-1]{} =M [d-2 -nd-1]{} \[chargenew\] which agrees with (\[chargegr\]). ### Entropy Now consider the entropy $S$. The neutral system was homogeneous in the direction $y$; we used this fact in lifting to the noncompact covering space before boosting. The classical entropy is linear in the coordinate length of the $y$ circle, since the area of the horizon is proportional to this length. Thus the entropy of the microscopic model must be extensive in the length $L_y$. Under boosting along $y$ we would expect that the length of the $y$ direction would contract by a factor $\gamma={1\over \sqrt{1-v^2}}=\c$; thus it would fit in a length $L'_y=L_y/\c$. We could have chosen any length of recompactification after boosting, but our choice was to keep $L_y$ unchanged. This means that we have considered a ‘larger amount of matter’ in the charged case – larger by a factor $\c$. This accounts for the relation S’=S \[entropynew\] ### The relation $\phi=\tanh\alpha$ The potential of the charged system is defined as =([M’Q]{})\_[S’]{} Setting $dS'=0$ in (\[entropynew\]) we find S d+[dSdM]{}dM=0 \[alpha\] Then from (\[massnew\]) we find (using (\[alpha\]) to eliminate $d\alpha$) dM’=dMFrom (\[chargenew\]) we find dQ=dMAs it stands, the ration ${dM'\over dQ}$ does not simplify. But now we use the Smarr relation for the neutral system TS=M [d-2-nd-1]{} =1 \[smarrnew\] Then we find =[dM’dQ]{}= ### The relation $T'S'=TS$ We have T’=([d M’dS’]{})\_Q Setting $dQ=0$ in (\[chargenew\]) we find d=-[dMM]{}[d-2-n-M[dndM]{}d-2-n]{} [c\^2+\^2]{} Using this we find dM’=dMdS’=[dMM]{}cAs it stands the expression T’S’=[dM’dS’]{} S’ does not simplify. But if we use the Smarr relation (\[smarrnew\]) we find T’S’=TS The near extremal system ------------------------ Let us now note some results that will be useful in our study of the near extremal system that we will use. We will for the most part use D1,D5 charges, but will also have occasion to consider the system with three nonzero charges – D1,D5,P. ### Two charges {#2chargesection} Let us take the solution (\[neutral\]) and add D1-D5 charges. We find the solution (we give the derivation in the Appendix) ds\^2\_E &=& H\_1\^[1/4]{}H\_5\^[-1/4]{}\[H\_1\^[-1]{}(- U dt\^2+ dy\^2) + H\_5 ds\^2\_B + dz\_a dz\_a\]\ G\^[(3)]{}&=&-U\^[-1/2]{} \_r H\_5 \_5 \_B dr-\_r H\_1\^[-1]{}\_1 drdtdy\ e\^[2]{}&=&[H\_1H\_5]{},H\_1 = 1+(1-U)\^2\_1,H\_5 = 1+(1-U)\^2\_5 \[chargedE\] (Here $ds_E^2$ is the Einstein metric.) In this case the analog of (\[massnew\]) is M’=M\[mprime\] Since we will be working close to extremality, let us compute the energy above extremality EM’-M\_[ex]{}=M’-Q\_1-Q\_5 \[mex\] The charges $Q_1, Q_5$ are given by the analog of (\[chargenew\]) Q\_1=M[2-n3]{}\_1\_1, Q\_5=M[2-n3]{}\_5\_5 \[q1q5charges\] In the near extremal limit we have $\alpha_1, \alpha_5\gg1$, so that \_i\_i\^2\_i+[12]{}, i=1,5 Using this in (\[mex\]) we get EM [1+n3]{} \[m2e\] For the near extremal case it is convenient to define the relative tension in a manner similar to the definition (\[relative\]) but with the mass replaced by the mass above extremality [@harmarkoberscharged] r \[defr\] From (\[tensiongr\]) and (\[m2e\]) we find r=[L[T]{}E]{}=[L[T]{}M]{}[ME]{}=[3n1+n]{} \[n2r\] We can similarly find the transformation law for the entropy. With two charges the analogue of (\[entropygr\]) is S’=S \_1 \_5 \[entropy2ch\] For later use let us define a ‘rescaled entropy’ as \[rescaledsgrav\] From (\[entropy2ch\]) and (\[q1q5charges\]) we find =[SM]{} [32-n]{}[\_1\_5]{} In the limit of large charges $\alpha_1,\alpha_5 \gg 1$ we get S=[SM]{} [32-n]{} \[s2shat\] ### Three charges We will also have occasion to turn on 3 charges, D1, D5, P. We will have the D1, D5 charges large as above, but the P charge will be arbitrary. The relation (\[mprime\]) now becomes M’=M\[mprimep\] and we have Q\_p=M[2-n3]{}\_p\_p \[qpcharge\] We define the energy above extremality $E$ again as the energy above the mass of the large charges D1, D5, and find EM’-Q\_1-Q\_5=M [1+n3]{}+M [2-n3]{} \^2\_p \[enew\] For later use, it will be helpful to write this relation in another way. For the neutral system (i.e. before any charges were added) we have the Smarr relation TS=M [2-n3]{} This implies M=[32]{} TS+[nM2]{}=[32]{} TS+[[T]{}L2]{} Substituting this in (\[enew\]) we find E=[TS2]{}2\_p +[[T]{}L2]{} \[erelation\] Phase Transition: Microscopic Model =================================== \[microsection\] We are looking at type IIB string theory with the compactification $M_{9,1}\r M_{4,1}\times T^4\times S^1\times \t S^1$. On these 6 compact directions we can wrap 4 mutually BPS charges: \(i) D5 branes wrapped on $T^4\times S^1$. \(ii) D1 branes wrapped on $S^1$. \(iii) Momentum modes P that along along $S^1$. \(iv) KK-monopoles that have $\t S^1$ as their nontrivial fiber, and that extend uniformly along $T^4\times S^1$. The four charges above can be permuted in all possible ways among themselves by S,T dualities. For this reason we will often list the number of these charges as $n_1, n_2, n_3, n_4$, instead of writing $n_5, n_1, n_p, n_k$. Work on black hole microscopics has given an understanding of the entropy of black objects for various compactifications. Let us recall these results. First let the length $L$ of $\t S^1$ be infinite, so that we have only the compactification $T^4\times S^1$. Then the mass of the KK charge becomes infinite, and it drops out of the computations and the entropy is an expression in the three remaining charges $n_1, n_2, n_3$. The entropy for all such cases can be derived from the following abstract expression [@hms] S=2(+)(+)(+) \[threec\] There are 6 variables on the right hand side. We have to vary these variables to achieve a maximum of $S$, subject to the constraints that we are given the net values of the 3 types of charges n\_i=n\_i-\|n\_i, i=1,2,3 \[threed\] and the total energy M=\_[i=1]{}\^3 m\_i (n\_i+\|n\_i) \[threee\] where $m_i$ is the mass of each quantum of type $i$. Roughly speaking, extremising $S$ tells us that we create pairs of those charges that are lighter, and that have [smaller]{} net charge values; i.e., we create pairs of the charges that we [don’t]{} have. In particular, if some charge $\hat n_i\r \infty$, then the nonextremal energy will not create pairs for this charge. Let us list those cases that will be relevant to us in what follows: \(a) Let there be 2 large charges $n_1, n_2\r \infty$, $\hat n_3=0$, and a small amount of nonextremal energy $E$. Then we find n\_3=\|n\_3=[E2 m\_3]{} and S=2(+)=4 \[en1p\] \(b) Let there be one large charge $n_1\r \infty$, $\hat n_2=\hat n_3=0$, and a small amount of nonextremal energy $E$. Then we have to extremise S=2(+)(+)=8 \[en23\] subject to E=2m\_2n\_2+2 m\_3n\_3 We find n\_2=\|n\_2=[E4 m\_2]{}, n\_3=\|n\_3=[E4 m\_3]{} and S=2 \[en24\] Now let $\t S^1$ be finite, so that all 4 charges are relevant. The analog of (\[threec\])-(\[threee\]) are [@hlm] S=2(+)(+)(+)(+) \[threecq\] n\_i=n\_i-\|n\_i, i=1,2,3,4 \[threedq\] M=\_[i=1]{}\^4 m\_i (n\_i+\|n\_i) \[threeeq\] The analogues of cases (a), (b) above are (a’) We have $\hat n_1, \hat n_2, \hat n_3\to \infty$, $\hat n_4=0$. We get n\_4=\|n\_4=[E2 m\_4]{} and S=4 \[formulap\] (b’) We have $\hat n_1, \hat n_2\r\infty$, $\hat n_3=\hat n_4=0$. We get n\_3=\|n\_3=[E4 m\_3]{}, n\_4=\|n\_4=[E4 m\_4]{} \[formula\] and S=2 [E]{} \[en2p\] Consider any of these expressions, say (\[en1p\]). We can regard the entropy as arising from ‘fractional’ $n_p, \bar n_p$ pairs. We have $n_1n_5n_p$ units of fractional $P$ excitations, and $n_1n_5\bar n_p$ units of fractional $\bar P$ excitations. Our interest is in understanding more about how fractionation works. We will therefore examine different possible fractionations below, and see how well they fit the physics of the black hole – black string transition. Modeling the phase transition {#microsectionmodel} ----------------------------- Let us now come to the main computations of this paper. We have seen that in the limit of large $L$ and small $L$ the entropy of the system is given by (\[en1\]) and (\[en2\]) respectively. We now wish to make a simple model for the transition between these two possibilities. We have two large charges D1, D5 (with integer values $n_1, n_5$ respectively); we have added these to the system to make the system ‘near-extremal’. On the gravity side the addition of charges induced a simple map between neutral and charged systems, so no generality was lost by looking at this near-extremal case. Let the energy above extremality be $E$. We let a part $E_1$ of this energy go to exciting $P, \bar P$ pairs. If these were the only excitations present then they would describe a ‘black hole’ in 4+1 noncompact dimensions, and give an entropy of type (\[en1\]), which we have rewritten in the form (\[en1p\]) above. The rest of the energy $E-E_1$ will go to entropy of the form (\[en2\]); we have rewritten this entropy in the form (\[en2p\]) above. If all the energy was in this latter form we would get a uniform black string. The important question now is what happens to the charges $n_1, n_5$. There are three possibilities: \(a) A fraction $(1-x)$ of the charges $n_1$ go to the entropy of type (\[en1p\]), and the rest $x$ go to the entropy (\[en2p\]); similarly a fraction $(1-x)$ of the charges $n_5$ contribute to (\[en1p\]), and the rest $x$ contribute to (\[en2p\]). (We take the fractions $x$ equal for the two charges by the symmetry between these charges, but we can consider different fractions too.) Thus we have two disjoint systems, and the ‘fractionation’ available for the two types of entropy are $x^2n_1n_5$ and $(1-x)^2 n_1n_5$ respectively. \(b) A fraction $(1-x)$ of the product $n_1n_2$ goes to the entropy (\[en1p\]) and a fraction $x$ to the entropy (\[en2p\]). Thus the D1 and D5 branes make a bound state with $n_1n_5$ effective degrees of freedom, and it is these effective degrees of freedom that get partitioned into the two subsystems. It is useful to define N=n\_1n\_5 and the partitioning gives $n_1'n_5'=N-\delta N$ to the entropy (\[en1p\]) and $n_1''n_5''=\delta N$ to the entropy (\[en2p\]). \(c) All the charges $n_1$ and all the charges $n_5$ contribute to each kind of entropy. Thus the full product $n_1n_5$ appears in each of the expressions of type (\[en1p\]) and (\[en2p\]). It turns out that possibility (b) gives a phase diagram that resembles the gravity computation, and we will adopt this choice below. We will discuss what happens with possibilities (a),(c) at the end of the section. ![(a) The entropy is plotted on the vertical axis, as a function of $x, \epsilon_1$, for $\epsilon=0.5$, a value below the point where the three phase structure appears; there is only one maximum at the boundary of the parameter space (b) The entropy graph for $\epsilon=2.4$, a value at which there are three phases; there are two maxima at the boundary of the parameter space and a saddle point in the middle. []{data-label="fig:Entropy_Plot"}](5D_entropy_plot_p5.eps "fig:"){width="2in"} ![(a) The entropy is plotted on the vertical axis, as a function of $x, \epsilon_1$, for $\epsilon=0.5$, a value below the point where the three phase structure appears; there is only one maximum at the boundary of the parameter space (b) The entropy graph for $\epsilon=2.4$, a value at which there are three phases; there are two maxima at the boundary of the parameter space and a saddle point in the middle. []{data-label="fig:Entropy_Plot"}](5D_entropy_plot_2p4.eps "fig:"){width="2in"} (a)(b) Thus adopting possibility (b) we write the entropy as S&=&2+2\ &=&[2]{}\[major\] If we introduce the dimensionless quantities =[Em\_k]{}, \_1=[E\_1m\_k]{},x=[NN]{} \[microscales\] we can rewrite the entropy as S=[2]{}\[\_1,x;\] \[rescaleds\] where \[\_1,x;\]=+(-\_1) \[sxx\] where for a given $\epsilon$ we should extremize in $\epsilon_1, x$. Let us now study the properties of the function $\hat{s}$ for different $\epsilon$. The extrema of $\epsilon$ occur at points where =0,=0 There is only one such point: \_1 = -1,x=[12-1]{} This point lies inside the allowed region $\epsilon_1\in[0,\epsilon]$, $x\in[0,1]$ for 1 The matrix of second derivatives at the extremal point is =-[1-1]{} Its determinant is =-[(2-1)\^28 (-1)]{} which is negative for $\epsilon>1$. Thus the extremal point is a saddle point, which is a maximum along some direction and a minimum along another direction. The value of $\hat{s}$ at the extremum point is \_c= Since the extremum we have found is a saddle, the actual maxima of the entropy must occur at the boundary of the allowed parameter space. We plot $\hat s$ against $x, \epsilon_1$ in fig.(\[fig:Entropy\_Plot\]). When $\epsilon<1$, as in fig.(\[fig:Entropy\_Plot\]) (a), there is one maximum at the boundary \_1=,x=0 \_a= \[eqfirst\] This corresponds to having all the energy and entropy in the mode of type (\[en1p\]) (we call this mode (a)). When $\epsilon>1$ we have the situation in fig.(\[fig:Entropy\_Plot\]) (b), where there are two maxima at the two ends. One as in (\[eqfirst\]), and the other at \_1=0,x=1 \_b=This latter extremum corresponds to having all the energy and entropy in a mode of type (\[en2p\]) (we call this mode (b)). Both maxima $\hat{s}_a$ and $\hat{s}_b$ are always greater than $\hat{s}_c$. To summarize, we have the following situation: \(1) For $\epsilon<1$, there is only one maximum corresponding to the ‘black hole phase’, phase (a). \(2) For $\epsilon>1$ we have three phases, that we interpret as folllows: \(a) black hole phase: $\epsilon_1=\epsilon$, $x=0$, $\hat{s}_a= \sqrt{2\epsilon}$; \(b) uniform black string phase: $\epsilon_1=0$, $x=1$, $\hat{s}_b=\epsilon$; \(c) non-uniform black string phase: $\epsilon_1=\epsilon-1$, $x=1/(2\epsilon-1)$, $\hat{s}_c=\sqrt{2\epsilon-1}$. Phase (c) is always unstable. Computing the tension --------------------- The energy above extremality comes from $P,\bar P$ and $KK,\overline{KK}$ pairs E=m\_p(n\_p+\|n\_p)+m\_k(n\_k+\|n\_k)=2m\_p n\_p+2 m\_k n\_k Note that $m_k$ depends on $L$ but $m_p$ does not. Thus the tension comes only from the excitations of type (b), where both P and KK pairs are excited. Since $m_k$ depends quadratically on $L$ we have =2 [m\_kL]{} \[forpp\] Thus we get =[EL]{}=[E\_bL]{}=(n\_k+\|n\_k) [m\_kL]{}=4 n\_k [m\_kL]{} Using (\[formula\]) we get[^4] =4 [E-E\_b4 m\_k]{} [m\_kL]{}=[E-E\_bL]{} The rescaled tension, defined in (\[defr\]), is r = [ LE]{}=[E-E\_1E]{}=[-\_1]{} In the uniform black string phase $\epsilon_1=0$ and thus the rescaled tension is r\_b= 1 In the black hole phase $\epsilon-\epsilon_1=0$ and the tension vanishes r\_a = 0 In the non-uniform string phase $\epsilon-\epsilon_1=1$ and our microscopic prediction for the rescaled tension is r\_c=[1]{} \[rc\] Note that the value $r_a=1$ we have obtained for the uniform string tension is the one expected from gravity (a fact that had already been noted in [@hms]). This can be seen as follows. The rescaled tension of the neutral system in the uniform string phase is obtained by setting $c_z=0$ in (\[neutralmass\]), and it is given by n\_=[12]{} Substituting this value in (\[n2r\]) gives the rescaled tension for the charged uniform string: r\_=1 We thus see that $r_a=r_\mathrm{us}=1$. Comparison between gravity and the microscopic model ---------------------------------------------------- In this subsection we compare various results from gravity with the predictions following from our simple microscopic model. It will be convenient to introduce rescaled energies as M, E Using the fact that =[R\_z\^2 R\_y Vg\^2’\^4]{}=m\_k we see that the quantity $\epsilon$ defined here coincides with the one introduced in (\[microscales\]). ### Mapping from the charged system to the neutral one ![The phase diagram predicted by the leading order microscopic model. The solid line is the uniform black string. The black hole branch is a horizonal (dashed) line that overlaps with the x-axis – the tension is zero. The dot-dashed line is the non-uniform string branch.[]{data-label="fig:4DnVSmu_micro"}](5DnVSmu_micro.eps){width="3in"} ![The entropy graphs for the three phases, with the same coding for the lines as in fig.(\[fig:4DnVSmu\_micro\]).[]{data-label="fig:4DsVSmu_micro"}](5DsVSmu_micro.eps){width="3in"} We would now like to compare the graphs for tension and entropy derived from model (b) with the corresponding graphs obtained from gravity (shown in Figs. \[fig:nVSmu\_gravity\], \[fig:sVSmu\_gravity\]). The graphs obtained from the gravity computation describe the neutral system, while the results from microscopics are for the charged system. We will thus map the variables in the microscopic computation to the corresponding neutral system, and then plot the results. The relations required to effect the needed map were given in section \[2chargesection\]. Eq.(\[n2r\]) relates the relative tension $r$ of the charged case to the relative tension $n$ of the neutral case. Eq.(\[m2e\]) reads as $\epsilon=\mu{1+n\over 3}$ in scaled variables, and relates the energy above extremality of the charged case to the mass in the neutral case. The entropy of the charged case is $S'$, but it is more convenient to look at $\hat S$ defined in (\[rescaledsgrav\]), which is related to the quantities of the neutral case by (\[s2shat\]). It turns out that $\hat S$ exactly equals the quantity $\hat s$ that we have used in our microscopic analysis. Checking this needs the exact constants relating the $Q_i$ to the corresponding $n_i$, but even without doing this we note that the ratio between $\hat S$ and $\hat s$ is a constant, and this constant will cancel when we plot ratios of entropies. Carrying out this map to the neutral system, we find the following for the three phases: \(a) black hole phase: n=0,=[4\^2]{} ([23]{})\^[3/2]{} \(b) uniform string phase: n=[12]{},=1 \(c) non-uniform string phase: n=[1]{},=[4\^2]{}([2-13]{})\^[3/2]{} ### Tension and entropy graphs With the above map to the neutral system we plot the tension and entropy graphs obtained from the microscopic model in figs.(\[fig:4DnVSmu\_micro\]),(\[fig:4DsVSmu\_micro\]) respectively. We should compare these graphs to the graphs of fig.(\[fig:nVSmu\_gravity\]) (b) and fig.(\[fig:sVSmu\_gravity\]) (b) respectively. The first feature we observe is that the microscopic model exhibits three phases, and the tension and entropies of these phases are in the correct order when compared to the gravity graphs; i.e., the tension of the non-uniform string branch is lower than the tension of the uniform string and higher than the tension of the black hole, and the entropy of the non-uniform string is always lower than the entropy of the other two phases. We will now compare the microscopic results to the gravity results in more detail. We note that the microscopic model gives vanishing tension for the black hole phase while the gravity computations do not; we will discuss this in the next section. ### The Gregory-Laflamme point Let the Gregory Laflamme (GL) point in the phase diagram be the point where the non-uniform string branch starts off from the uniform string branch. In the microscopic model this point is at \_[GL]{}=1, r\_[GL]{}=1 Mapping this to the neutral system we get \_[GL]{}=2 \_[GL]{} = 2 \[mugl2\] ### Transition energy On the gravity side, the energy at which the GL transition happens has been determined numerically in [@gubser; @wiseman; @sorkin]. The value of this energy depends on the number of non-compact dimensions, and for $d=4$ non-compact dimensions it is given by \_[GL]{}M\_[GL]{}3.52 Since the microscopic model gives $\mu_{GL}=2$, we find =1.76 Thus we do not get perfect agreement on the GL point in our leading order microscopic model. But we will see below that once we scale energies so that they are all compared to the GL point, we get good agreement for the tension graph of the non-uniform string. ### Non-uniform string tension near GL In [@harmarkobersIII], from an analysis of the $d=5$ numerical solution of [@wiseman], it was found that the following relation describes with good accuracy the non-uniform string phase near to the GL point: -1=x(d)([MM\_[GL]{}]{}-1) Let us assume that an analogous relation is valid for a generic dimension $d$, with a dimension-dependent costant $x(d)$. One can combine this relation with the Smarr identity[^5] TS=[d-2-nd-1]{}M,T\_[GL]{} S\_[GL]{} = [d-3d-2]{} M\_[GL]{} \[smarrGL\] to obtain a relation for the rescaled tension n = [(d-1)(d-3)d-2]{}(x(d)-1)[M\_[GL]{}M]{}+ [(d-2)\^2-x(d)(d-1)(d-3)d-2]{} \[nneargl\] The $M$-independent term on the r.h.s. of the above equation vanishes if x(d)=\|x(d)=[(d-2)\^2(d-1)(d-3)]{} \[xd\] For $d=5$, the numerically derived value for the constant $x$ is x(5)=1.12 We note that this number is very close to \|x(5)=[98]{}=1.125 Thus for $d=5$ the M-independent term in (\[nneargl\]) vanishes with good accuracy. We do not know the numerical value for $x(4)$. It seems however reasonable to conjecture that the M-independent term vanishes also for $d=4$. This happens if x(4)=\|x(4)=[43]{}1.33 We can support this conjecture by expanding (\[nneargl\]) in powers of $M-M_{GL}$: the first term in this expansion is known, for $d=4$, from the work of [@gubser]: n=[12]{}-(-\_[GL]{})+O((-\_[GL]{})\^2),=.14 \[nexp\] Expanding (\[nneargl\]) we find, for $d=4$ n=[32]{}(x(4)-1)[M\_[GL]{}M]{}+2-[32]{}x(4)=[12]{}-[32]{}(x(4)-1)[-\_[GL]{}\_[GL]{}]{}+ O((-\_[GL]{})\^2) \[n1\] Comparison of (\[nexp\]) and (\[n1\]) gives =[32]{}[x(4)-1\_[GL]{}]{}x(4)=[23]{}\_[GL]{}+11.33 This confirms, with good accuracy, that $x(4)$ is such that the M-independent term in (\[nneargl\]) vanishes. If we use $x(4)=\bar x(4)$ in (\[nneargl\]), we then have n = [12]{} [M\_[GL]{}M]{}=[12]{} [\_[GL]{}]{} \[n2\] In the microscopic model we have found that for the non-uniform string $n={1\over \mu}$. Noting that $\mu_{GL}=2$ in this model (eq.(\[mugl2\])), we can write for the microscopic model n=[12]{}[\_[GL]{}]{} which is in perfect agreement with (\[n2\]). ### The Smarr relation On the gravity side all black objects satisfy the Smarr relation. In the microscopic model, the black hole and uniform black string phases correspond to the 4+1 dimensional black hole and the 3+1 dimensional black hole respectively, and the entropies $S(E, Q_i)$ in each case are known to agree with the gravity computation. Thus the Smarr relation is satisfied for these two phases. For the non-uniform string, the microscopic model gives S\_(E,N)=S\_(E\_1,N\_1)+S\_(E\_2,N\_2) where $E=E_1+E_2$, $N=n_1 n_5 = N_1 + N_2$. It is easy to check that if a system is made of two non-interacting subsystems, i.e. E=E\_1+E\_2, Q\_i=Q\_[i,1]{}+Q\_[i,2]{}, S=S\_1+S\_2 \[mix\] and if each subsystem satisfies the Smarr relation, then the total system will satisfy the Smarr relation. Thus it is consistent to identify the saddle point of the microscopic model with the non-uniform string black string. Even though satisfying the Smarr relation appears trivial in this leading order microscopic model, it will be nontrivial if we look at higher order corrections where we need to add interactions that link the two different modes of excitation. We will then not have (\[mix\]), and the interactions will have to be chosen such that the Smarr relation continues to be satisfied. Note that a generic field theory model will not satisfy the required Smarr relation. The dimension $d$ of the spacetime appears in (\[smarrnewp\]), and this dimension is in general different from the dimension of the dual field theory. Thus the Smarr relation appears to be nontrivial from the viewpoint of the microscopic model. Other models of fractionation ----------------------------- At the start of section \[microsectionmodel\] we had noted three different possible models of fractionation. We analyzed model (b) above in detail; here we note what happens if we look at models (a) and (c). ### Model (a) ![The tension graph for model (a). Note that the non-uniform string branch starts off at $\mu=0$, which is not in accord with the gravity diagrams (fig.\[fig:nVSmu\_gravity\]). []{data-label="fig:4DsVSmu_micro_alternate_a"}](5DnVSmu_micro_alt_a.eps){width="3in"} The analog of (\[sxx\]) is s=(1-x)+x(-\_1) Extremising in $x, \epsilon_1$ gives the point x=[1]{}, \_1=[12]{}(-1)\^2 \[sol22\] The determinant of the matrix of second derivatives is =-(1+[1]{})\^2 Since this is negative, we have a saddle point, and the actual maxima are again at the endpoints of the domain of parameters. The black hole branch is at $(x=0, \epsilon_1=\epsilon)$, and the uniform black string branch is at $(x=1, \epsilon_1=0)$. But note that the ‘non-uniform black string branch’, given by (\[sol22\]), is present for [all]{} values of $\epsilon$. The tension versus mass graph is given in Fig.\[fig:4DsVSmu\_micro\_alternate\_a\]. We see that the point where the non-uniform black string branch meets the uniform black string branch is at $\epsilon=0$, rather than at some positive value of $\epsilon$. Thus the phase diagram has a qualitative difference from the gravity diagram. ### Model (c) In this case the analog of (\[sxx\]) is s=+(-\_1) Setting ${\p \hat s\over \p \epsilon_1}=0$ gives \_1=[12]{} and ${\p^2\hat s\over \p \epsilon_1^2}=-1< 0$. Thus this point is a maximum when it can be attained. Thus for $\epsilon<{1\over 2}$ the maximum is at $\epsilon_1=\epsilon$, while for $\epsilon>{1\over 2}$ the maximum is at $\epsilon_1={1\over 2}$. The phase diagram is sketched in Fig.\[fig:4DsVSmu\_micro\_alternate\_c\]. We see that there is only one phase for all $\epsilon$, so we have a qualitative difference from the gravity picture. ![The phase diagram for model (c). There is only one phase for each $\mu$, which does not agree with the gravity diagrams (fig.\[fig:nVSmu\_gravity\]).[]{data-label="fig:4DsVSmu_micro_alternate_c"}](5DnVSmu_micro_alt_c.eps){width="3in"} A physical picture of model (b) ------------------------------- Having seen above that models (a),(c) give phase diagrams that are less in accord with gravity than model (b), let us look at why a model like (b) is suggested by what we know about microscopics. Consider first the case where we have the compactification $M_{4,1}\times T^4\times S^1$, and let there be just one large charge. We let this charge be $n_5$ units of D5; the excitations can then be (fractional) pairs of D1 and P charges. First suppose that only one kind of charge pairs is excited. We can let this be $P\bar P$. But this situation is very similar to exciting $P\bar P$ pairs on a bound state of D1 branes; the D5 wrapped on $T^4$ acts like an effective string in 6-D and the $P\bar P$ excitations run on this effective string. The picture of this bound state is simple: the strands of the effective string join together to make one ‘long string’, and the $P\bar P$ excitations are transverse vibrations of this long string. The vibrations cause the strands of the long string to separate from each other and spread over some transverse region. Now consider the excitation mode where two kinds of charge pairs are excited. This situation was studied in [@malda], and it was found that the entropy was exactly reproduced by assuming that the energy of excitation is carried by a ‘fractional tension D1 brane’ living in the D5 worldvolume; the tension of the D1 is fractionated by the factor $1/n_5$. The string excitations represent both $P\bar P$ pairs and $D1\overline {D1}$ pairs. The entropy of such a string is given by (\[en24\]). It seems plausible that to get the required fractionation of the tension the D5 branes must be all ‘stuck together’, and the D1 should be ‘dissolved’ in this D5 bound state to give rise to the effective string with fractional tension. Thus we see that in the mode where one kind of charge pairs are excited the D5 branes [separate]{} from each other in carrying the excitation, while in the mode where we have two kinds of excitations we need the D5 branes to be [together]{}. If we are in the process of transition from one mode to the other, then we will have some excitation in each of these modes. But this suggests that we will need to partition the D5 branes into two sets: a fraction $(1-x)$ which will contribute to the first mode and a fraction $x$ which will contribute to the second mode. Now return to our actual problem, where the compactification is $M_{3,1}\times T^4\times S^1\times \t S^1$ and instead of the D5 charge we have D1-D5 charge. When we have the mode with only one kind of charge pair the entropy is given by (\[en1p\]), where the effective string has winding $n_1n_5$ instead of $n_5$. When we have two kinds of charge pairs, we just note that by dualities we can map this to near extremal D5-KK, which was also studied in [@malda]. It was found that we again get a fractional tension effective string living in the worldvolume made by the D5 and KK branes; all one has to do is replace $n_5$ with $n_5n_k$, which in our duality frame corresponds to replacing $n_5$ everywhere with $n_1n_5$. Thus extending our intuition from the D5 charge case to the D1-D5 charge case we expect that in the phase transition we will need to partition the product $n_1n_5$ into two parts: a fraction $(1-x)$ for the first mode and a fraction $x$ for the second mode. This is just model (b). Microscopic model at large momentum ----------------------------------- So far we have added only two large charges, D1, D5, to bring the system near to extremality. But we can also add a large P charge by another boost. In this subsection we consider how the black hole – black string phase transition will look in the near extremal 3-charge system. The transition happens when the horizon radius parameter $r_0$ becomes comparable to the compactification length $L$. We have now chosen all the charges D1,D5,P large, so the charge radii for these charges are all much larger than $L$. We can thus assume that we are looking at a situation with three nonzero charges in 3+1 noncompact dimensions, with a small amount of nonextremality. Since the 4 possible charges for this compactification can be freely permuted by dualities, we can equally well look at a system where the nonzero charges are D1,D5,KK, and the nonextremality then creates pairs of fractional $P\bar P$ excitations. This system has been studied before. If the numbers of the three nonzero charges are $n_1, n_2, n_3$, then we get an effective string with total winding number $n_1n_2n_3$. The kind of entropy that we will encounter here was studied in [@emission].[^6] If the system is extremal, then we get entropy from partitioning the effective string in all possible ways into component strings, getting an entropy $S=2\pi\sqrt{n_1n_2n_3}$. If there is some nonextremal energy then pairs of $P\bar P$ excitations can be created on this effective string in the following way. A fraction $f$ of the effective string winding creates entropy as before by breaking up into component strings of different length; this gives an entropy $S=2\pi\sqrt{n_1n_2n_3f}$. The remainder of the effective string joins up into one long component string (thus having no entropy of its own) but this enables the $P\bar P$ excitations to be fractionated in units of $1/(n_1n_2n_3(1-f))$. This gives an entropy $4\pi\sqrt{n_1n_2n_3(1-f)}\sqrt{E\over 2 m_4}$, by (\[formulap\]). We then extremize over $f$ to find out how the effective string will partition itself into its two kinds of roles. Let us apply this 3-charge model to our present problem, calling the result ‘model A’. The entropy is thus S=2+ 4=2 (+) where $\t N=n_1n_2n_3$, and we have again defined $\epsilon=E/m_4=E/m_k$ (we have reverted to our duality frame where the nonzero charges are D1, D5, P and the excitations are pairs of KK). The entropy $S$ has only one extremum, as a function of $f$, given by -=0f=[12+1]{} It is easy to check that this extremum is a maximum. Thus we find a unique maximum for every value of the energy: According to model A the complicated phase structure we have seen emerging in section \[microsectionmodel\] disappears, at large P, and we are left with a unique stable phase for every $E$. The value of the entropy at the maximum is S\_[max]{}=2=2 \[smax\] We note that this entropy agrees with the extremal 3-charge black hole entropy [@sv] S\_=2 for small energies ($\epsilon\ll 1$), and with the near-extremal 3-charge uniform black-string entropy (eq.(\[formulap\])) S\_=2 for large energies above extremality ($\epsilon\gg 1$). We can also compute the value of the rescaled tension $r$ predicted by this model. The rescaled tension is defined as r=[LE]{}=[LE]{}([EL]{})\_[S, Q\_i]{} where $E$ is the energy above extremality (so that the total energy is given by $M'=Q_{1}+Q_{5}+Q_p+E$). Consider the entropy given by (\[smax\]). Since $m_k\sim L^2$, we find that $S$ is kept constant if the variations of $E,L$ satisfy E= 2 E [LL]{} Thus r= [LE]{}[2 EL]{}=2 \[forpq\] This agrees with the result from the gravity computation when we have the extremal three large charges limit. In [@hkor] it was shown that in this limit the $r$ vs. $\epsilon$ graphs for the three phases in the gravity computation all degenerate[^7] to the line $r=2$. Thus, as far as the tension is concerned, model A gives a prediction which is consistent with gravity. But this analysis may also look a little puzzling, for the following reason. In the gravity computation the three phases collapse to one horizontal line in the $r-\epsilon$ diagram, but there should still be three overlapping phases for each $\epsilon$ in the range where the non-uniform string exists. On the other hand the analysis of model A gives a unique maximum for the entropy for all $\epsilon$, so there seems to be only one phase. To emphasize the puzzle still further consider the model (\[sxx\]) that we have analyzed in the case of two large charges D1, D5, and where we did get three phases for a range of $\epsilon$. In this model we can let $n_p\ne \bar n_p$, thus getting a nonzero P charge. All the results derived above for this model can be extended to the nonzero $P$ case by changing $n_p, \bar n_p$ in the following way to account for the P charge: n\_pn\_p e\^,\|n\_p \|n\_p e\^[-]{} \[pmap\] The limit of large P is given by taking $\sigma\gg 1$. Let us denote the microscopic model obtained in this way as model B. From the point of view of model B, it is clear that there is a one to one map between the phase diagram we have found at $P=0$ and the one at $P\gg 1$, given by (\[pmap\]). In particular model B still predicts three phases (which all degenerate to the line $r=2$ as $\sigma\r\infty$). We thus need to ask: If we go slightly off extremality, how does the single line in the tension graph of model A split into three closely spaced lines? We expect that we will see this split if we take into account interactions in the CFT which describes the near extremal 3-charge system. This is a general feature of adding charges to a system; we simplify the system by going closer to extremality, but we may need to take into account interactions between the degrees of freedom of the system to recover all the physics. We will discuss this issue in more detail in the next section when we talk about the need for interactions to understand the tension in the black hole phase. Tension in the black hole phase =============================== Our microscopic model gives a phase diagram that has several features of the phase diagram obtained by numerically solving the equations of general relativity. But one important feature that is missing is the tension of the black hole branch: the microscopic model gives ${\cal T}=0$ while the numerical results show a tension rising from zero as the mass is increased. This tension was recently related to microscopics in [@hkor]. In this section we discuss why this tension did not show up in our calculation of the last section, and extract some properties of the tension from microscopic arguments. The source of tension in the black hole phase --------------------------------------------- ### The leading order gravity calulation We first recall the origin of this tension in the gravity picture. A similar calculation can be found for instance in [@gorkol2]. ![(a) A small black hole in a space with a compact transverse circle (b) Lifting to the covering space replaces the compactification by an infinite set of images of the hole. Gravitational attraction between the images is given by graviton exchange, which is a one loop of open strings in the dual channel, with all modes of the open string contributing.[]{data-label="fig:SqueezedBH"}](SqueezedBH.eps){width="3in"} (a)(b) Fig.\[fig:SqueezedBH\](a) shows a black hole with radius much smaller than the compactification length. In Fig.\[fig:SqueezedBH\](b) we lift to the covering space, getting an infinite set of images of the black hole. Each copy of the hole feels the attraction of all other copies, and this creates a negative potential energy which is the effect of the compactification. At leading order we can just use the Newtonian potential to compute this energy. In 4+1 dimensions the gravitational potential energy between two bodies of mass $M_1,M_2$ at distance $r$ is V=-[4 G\^[(5)]{}\_N3]{}[M\_1M\_2r\^2]{} The energy between a hole and all its images is V\_[array]{}=-[43]{}[G\_N\^[(5)]{}M\^2L\^2]{}2 \_[n=1]{}\^[1n\^2]{}=-[49]{}[G\_N\^[(5)]{}M\^2L\^2]{} where we have used that \_[n=1]{}\^=[\^26]{} The part of this energy that we must attribute to the compactification length $L$ is V\_[int]{}=[V\_[array]{}2]{}= -[29]{}[G\_[N]{}\^[(5)]{} M\^2L\^2]{} \[vint\] The tension is given by ${\cal T}=({\p E\over \p L})_S$. Since the black hole horizon is not distorted to leading order, the entropy is independent of $L$, and we can just compute the $L$ derivative of $V_{int}$ to obtain L[T]{} \[tension\] where the approximation symbol denotes the fact that we have just computed the interaction energy at leading order in the size of the black hole. The above calculation was for neutral holes, but we can obtain $V_{int}$ for charged holes by relating them to neutral holes via the relations of section \[boostsec\]. We need large D1,D5 charges, and finite P charge. Let $E_0$ be the energy above extremality in the absence of the interaction $V_{int}$. With no interaction the tension $n$ vanishes, so that at leading order one has E\_0 2\_p \[e0\] The tension is invariant under the addition of charges (eq. (\[tensiongr\])), so we just get (\[tension\]) with $M$ obtained through (\[e0\]) L[T]{}([E\_0\_p]{})\^2 \[tensionone\] The momentum charge carried by the hole is Q\_pm\_p P 2\_p \[p0\] (here $P$ is the integer charge). We can write for the charged hole =([EL]{})\_[Q,S]{}=[V\_[int]{}L]{} Note that the entropy changes in a known way under addition of charge by boosting ($S\r S \cosh\alpha$). From this relation and (\[p0\]) we see that (to the order we are working at) we keep $Q_p,S$ fixed if we keep $M,\alpha_p$ fixed. Thus the tension is related to $V_{int}$ only through variations of $L$, and we find for the charged case a relation analogous to (\[vint\]) V\_[int]{}-2([E\_02\_p]{})\^2 \[vintcharged\] ### The effect in the dual field theory Let us now ask how the energy (\[vintcharged\]) would be obtained in a field theory calculation. We have seen that this energy is obtained from the long range gravitational attraction between the hole and its images. The exchanged graviton can be seen in a dual channel as a 1-loop open string diagram, but note that since $L\gg \sqrt{\alpha'}$ all modes of the open string contribute in the loop integral, not just the lowest energy modes. ![(a) A string stretched between two well separated branes has a closely spaced tower of excitations (b) When a large brane charge is added to the system we only get the ground states of open strings, but the interactions between these encode the original tower of excitations.[]{data-label="fig:IntrBranes"}](IntrBranes.eps){width="4in"} (a)(b) Now let us ask how this interaction would be seen in a dual field theory description. First look at a simpler system: an open string stretched between two D3-branes, with the length of this string being $L\gg\sqrt{\alpha'}$; this is pictured in Fig.\[fig:IntrBranes\](a). The low energy dynamics of this open string has a tower of closely spaced vibration modes. Now consider such a string, still stretched between two D3 branes, in a space with large D3 brane charge – i.e., in $AdS_5$. In the gravity picture, the string still has a tower of oscillation modes. But in the dual CFT the open string is a gauge boson, Higgsed because of the separation between the D3 branes at its ends. At weak coupling this gauge boson has no tower of excitation modes. But at larger coupling, we can sum a set of ladder diagrams and see that the tower of oscillation modes is reproduced [@klebanovmalda]. Thus we see that the tower of modes in the gravity picture of fig.\[fig:IntrBranes\](a) is reproduced in the CFT by degrees of freedom that arise from only the [ground]{} states of open strings; the price we pay is that we have to consider all orders of interaction between these open strings. Let us draw lessons from this example for our own system. In the gravity picture fig.\[fig:SqueezedBH\](b) we have all modes of the open string stretched between images of the black hole. The dual CFT in our case is complicated, but can be assumed to arise in some way analogous to the field theory of open strings on branes. To see the vibration modes of the open string in the CFT we thus expect that we will need to consider all orders of interaction between the basic degrees of freedom of the CFT. But in our microscopic computation of the phase transition we considered only the free CFT. This should explain why we did not see the tension in the black hole phase. Interactions in the CFT are difficult to study. In the D1-D5 CFT they involve the insertion of a ‘twist’ operator in the orbifold CFT [@lmorb]. With the presence of the KK-monopole degrees of freedom, we will get other interactions as well. If we do not study these interactions, how can we proceed to study the phase transition? In [@costaperry] a method was developed to get an indirect picture of the CFT effects, and this method was used further in [@hkor]. We will recall this approach below, and study it a further to extract some properties of the effective interactions in the CFT. Microscopics ------------ First consider the black hole in the limit $L\r\infty$. Then we have just a black hole carrying D1,D5 charges in 4+1 dimensions, and this is described by the 1+1 dimensional CFT. The energy above extremality and the P charge are given by E\_0 = m\_p (n\_p+[\|n]{}\_p),P = n\_p-[\|n]{}\_p \[ep\] Using (\[e0\]) and (\[p0\]) we find n\_p=[M e\^[2\_p]{}6 m\_p]{},\_p = [M e\^[-2\_p]{}6 m\_p]{} \[npnpbar\] In [@costaperry; @hkor] a ‘mixed’ approach was taken. Suppose that for $L\r\infty$ the black hole non-extremal energy is $E_0$ and the charge is $P$; we can find $n_p, \bar n_p$ using (\[ep\]). For finite $L$ there is a lowering of the energy due to the interaction with images, by an amount $V_{int}$. Thus for the same total energy, there will be higher values of $n_p, \bar n_p$, and a correspondingly higher entropy. This increase in the entropy can be checked against the gravity prediction. This is a ‘mixed’ approach in the sense that $V_{int}$ is computed from gravity, and then fed into the microscopic computation through the values of $n_p, \bar n_p$. We will also follow such a mixed approach, but proceed in a slightly different way that will allow us to better extract some quantities of interest. At $L\r\infty$ the near horizon region of the D1-D5 system is dual to a 1+1 dimensional CFT, in which the left and right movers do not interact; thus there is no term coupling $n_p$ to $ \bar n_p$. Because the AdS region ends and goes over to flat space, there is a breaking of conformal symmetry and a coupling of left and right movers to the graviton field; this cubic interaction was responsible for Hawking emission from the near-extremal D1-D5 system [@dm1]. In the present problem we will let the charge radii of the D1,D5 charges be much larger than $L$, so this changeover from AdS to flat space is not relevant, and the above coupling is not the leading order interaction between left and right movers. But what we do have is a compact $z$ direction, and KK monopole pairs can wrap around this circle. When we integrate out the effect of these pairs we will be left with an interaction between left and right movers. We will now try to understand these effective interactions in more detail, obtaining information about them from gravity, the Smarr scaling relation, the behavior under boost etc. We see from (\[vintcharged\]), using (\[npnpbar\]), that the interaction energy can be written as V\_[int]{}= -[8 ]{} [G\_N\^[(5)]{} m\_p\^2L\^2]{} n\_p [\|n]{}\_p We observe that the interaction energy vanishes in the BPS case ($n_p=0$ or $\bar n_p=0$), as expected. The total energy is (to this order) E = E\_0+V\_[int]{}= m\_p (n\_p+\|n\_p)-[8 ]{} [G\_N\^[(5)]{} m\_p\^2L\^2]{} n\_p [\|n]{}\_p \[e20\] This relation gives the energy of the microscopic model to second order in $n_p, \bar n_p$. The horizon is not deformed to leading order, so we still have S=2(+) \[entropy\] How do we continue the study of interactions to higher order? In the $L\r\infty$ limit we had a free CFT and we determined $n_p, \bar n_p$ from the relations (\[ep\]). In the interacting theory there will be an ambiguity in defining $n_p, \bar n_p$ for a given state. From the gravity perspective we can split the corrections to the thermodynamics into two kinds of terms: those that come from nonlinearlities of general relativity and those that arise because the horizon of the black hole is distorted by the compactification. The latter set of terms do not start till order $r_0^{10}$, where $r_0$ is the horizon radius. The former set of terms includes the interaction energy $V_{int}$ derived above as the leading correction, and other similar terms that follow at higher order, all in integer powers of $r_0^2$ [@rothstein]. We will define $n_p, \bar n_p$ at all orders in the interaction through the relations P=[Q\_pm\_p]{}=n\_p-\|n\_p,S=2(+) \[relations\] In view of the comments above, this definition is natural to all orders where the horizon is not distorted (since the internal degrees of freedom giving the entropy $S$ are not affected), and at higher orders there is no particularly natural scheme to define $n_p, \bar n_p$, so we might as well use (\[relations\]). The nontrivial information about the system will then be in the expression for $E$ in terms of $n_p, \bar n_p$. Note that a priori we have $E=E(S,Q,L)$. Roughly speaking, the Smarr scaling relation will set the $L$ dependence, and the behavior under boosts will set the $Q$ dependence. Thus the freedom in $E$ will be equivalent to ‘one function of one variable’. The coefficients describing this remaining function we will take from the gravity calculation – in principle they can all be determined numerically, but of course we will be looking at only the first few coefficients, and these can be understood analytically. Interactions at second order ---------------------------- This section is somewhat long, so we first summarize the steps in our computation. In section (\[gravhole\]) we recall the relations in gravity between the neutral system and the charged one, but in a notation specially adapted to the study of the black hole branch. In section (\[defdim\]) we introduce dimensionless variables to simplify notation. In section (\[smarrhole\]) we argue that because of the Smarr relation, we can write the dimensionless energy $\epsilon$ as a function only of rescaled versions of the parameters $n_p, \bar n_p$ (these rescaled parameters are called $\nu, \bar \nu$). In section (\[secondhole\]) we derive our main result: by using the fact that the tension ${\cal T}$ must remain unchanged under boosting, we derive the term in $\epsilon(\nu, \bar \nu)$ which takes us to one order higher in $\nu, \bar\nu$ than the expression (\[e20\]). In section (\[largenu\]) we extend this result to the case where we keep the first order in $\bar\nu$, but all orders in $\nu$. ### Gravity variables for the small black hole limit {#gravhole} Let us recall the relation (\[erelation\]) E=[TS2]{}2\_p +[[T]{}L2]{} \[erelationq\] Since we are working in the phase where we have a small black hole, it will be convenient to use variables that relate directly to the structure of the hole. The neutral hole in 4+1 dimensions has the metric ds\^2=-(1-[r\_0\^2r\^2]{})dt\^2+[dr\^21-[r\_0\^2r\^2]{}]{}+r\^2d\_3\^2 \[metricsymm\] We have M=[38 G\_N\^[(5)]{}]{}r\_0\^2, S=[\^22 G\_N\^[(5)]{}]{}r\_0\^3, T=[12r\_0]{}, TS=[4 G\_N\^[(5)]{}]{} r\_0\^2 \[mts\] After we compactify the transverse circle $z$ the metric will no longer have the spherically symmetric form (\[metricsymm\]). But we will use the last of the relations (\[mts\]) to motivate the following definition of $r_0$ even after compactification TS r\_0\^2 Let us substitute this in (\[erelationq\]). We find E=[4 G\_N\^[(5)]{}]{} [r\_0\^22]{} 2\_p +[[T]{} L2]{} \[etwo\] The tension ${\cal T}$ to leading order has been found in (\[tensionone\]). Putting this in (\[etwo\]) we get E 2\_p+[2G\_N\^[(5)]{}L\^2]{} ([M3]{})\^2 \[ethree\] The Smarr relation for the neutral system gives M [2-n3]{} =TS r\_0\^2 \[formulaqq\] When $L\r \infty$ we have $n\r 0$, so for small black holes we have r\_0\^2 Substituting this is (\[ethree\]) we get E=[4 G\_N\^[(5)]{}]{} [r\_0\^22]{} 2\_p+[2G\_N\^[(5)]{}L\^2]{} ([8 G\_N\^[(5)]{}]{})\^2 r\_0\^4+O(r\_0\^6) \[efour\] ### Defining dimensionless variables {#defdim} The above expression can be simplified by factoring out constants to define dimensionless variables. Thus write =[16G\_N\^[(5)]{}L\^2]{} E, q=[16G\_N\^[(5)]{}L\^2]{} Q\_p, \_0=[2L]{} r\_0 Then (\[efour\]) becomes =[\_0\^22]{}2\_p+[\_0\^432]{}+O(\_0\^6) \[eeten\] and we have (using (\[qpcharge\]),(\[formulaqq\])) q=[\_0\^22]{}2\_p \[eeel\] We also define scaled microscopic values: =[16G\_N\^[(5)]{}L\^2]{} m\_p n\_p, \|=[16G\_N\^[(5)]{}L\^2]{} m\_p \|n\_p \[relationsn\] ### The Smarr relation in microscopic variables {#smarrhole} We are interested in looking at the general expression for $E$ as a power series in $n_p, \bar n_p$; this expression will encode the effect of interactions in the CFT arising from the compactification of the direction $z$. We will now argue that in terms of the dimensionless variables introduced above, we will have =(, \|) \[eenine\] The argument proceeds in three steps: \(A) First recall eq. (\[enew\]). For the neutral system, we have n=n(M,L) \[eetwo\] Using (\[qpcharge\]) to express $\alpha_p$, we find that we can write E=E(M,L,Q\_p) \[eeone\] Now consider S=S\_[neutral]{}\_1\_5\_p where S\_[neutral]{}=S\_[neutral]{}(M,L) Let us define S Then we find, for large $D1$ and $D5$ charges, S=[S\_[neutral]{}\_1\_5\_pM[2-n3]{}]{} and we find that we can again write S=S(M,L,Q\_p) \[eethree\] We can now eliminate $M$ between (\[eeone\]) and (\[eethree\]) to write E=E(S, L, Q\_p) \[eefour\] (B)Now consider the Smarr scaling for the 4+1 dimensional gravity solution. The solution is invariant under EL\^2E, Q\_1L\^2 Q\_1, Q\_5L\^2 Q\_5, Q\_pL\^2 Q\_p, SL\^3 S \[eefourbis\] Note that under this scaling we will have SLS Thus we can write (\[eefourbis\]) as =[EL\^2]{}([SL]{}, [Q\_pL\^2]{}) In the above scalings we have regarded $G_N^{(5)}$ as a constant, but since this constant has units, its useful to see how it appears in the physical quantities of interest. All energies appear in the gravity theory through the combinations G\_N\^[(5)]{}E, G\_N\^[(5)]{}Q\_1, G\_N\^[(5)]{}Q\_5, G\_N\^[(5)]{}Q\_p etc. The entropy appears in the gravity theory through the area of a horizon, i.e. as G\_N\^[(5)]{} S\~A, S=[G\_N\^[(5)]{}S]{} We can thus write (\[eefour\]) as =[G\_N\^[(5)]{}EL\^2]{}([SL]{}, [G\_N\^[(5)]{}Q\_pL\^2]{}) \[eesix\] \(C) Now note that Q\_p=-\|\[eesev\] Also, S=[S]{}=[2(+)]{}= [2(+)]{} where $m_1, m_5, m_p$ will be used to denote the masses of individual quanta of D1, D5, P charges. Now note that m\_1m\_5m\_p=[4 G\_N\^[(5)]{}]{} and we find = + \[eeeig\] Returning to (\[eesix\]) we see that the LHS is just $\epsilon$ (upto a numerical constant), and the arguments on the RHS are (from (\[eesev\]),(\[eeeig\])) just functions of $\nu, \bar \nu$. Thus we establish (\[eenine\]). ### The second order correction in the microscopic picture {#secondhole} We have defined the microscopic parameters $n_p, \bar n_p$ through (\[relations\]), and their rescaled versions $\nu, \bar \nu$ through (\[relationsn\]). We now wish to work out the expansion of $\epsilon(\nu, \bar \nu)$ in powers of its arguments, using what we know from gravity. At leading order in $\nu, \bar \nu$ we have =+\|, q=-\|We note some other general properties of $\epsilon$. First, it must be symmetric in $\nu, \bar \nu$. Second, if $\bar \nu=0$ then the system is BPS, and the energy is not corrected by the compactification length $L$ being finite. This means that for $\bar \nu=0$ we must have precisely $\epsilon=\nu$; there are thus no terms in $\epsilon$ of type $\nu^2, \nu^3, \dots$, and similarly there are no terms $\bar \nu^2, \bar \nu^3, \dots$. The gravity calculation of the next to leading correction to the energy (eq. (\[e20\])) is written in dimensionless form as =+\|-[12]{} \|At next order we would have =+\|-[12]{}\|+ \|(+\|) \[mueq\] We will now see that we can determine the constant $\gamma$ just by using the fact that the tension does not change under boosting. Note that boosting changes the charge $Q_p$, and will also change $\nu, \bar \nu$. But since the tension is to be left invariant, we get a constraint on expansion coefficients like $\gamma$ in (\[mueq\]). Let us carry out this computation. The tension is given through L[T]{}=L[L]{}E=L[L]{} The $L$ derivative is done by keeping $S$ and $Q_p$ fixed; from (\[relations\]) this is equivalent to keep $n_p$ and $\bar n_p$ fixed, and hence $\nu, \bar \nu$ scale as $L^{-2}$. It follows that for a term in $\epsilon$ of the form $\nu^a\bar \nu^b$ we will have the contribution to the tension L[T]{}-2(a+b-1)\^a\|\^b Thus for the terms in (\[mueq\]) we will get =\|-4 \|(+\|) \[taurel\] Let us now convert to gravity variables, using (\[eeten\]),(\[eeel\]). We have =+\|-[12]{}\|=[\_0\^22]{}2\_p+[\_0\^432]{} q=-\|=[\_0\^22]{}2\_p where we have written the $\nu, \bar \nu$ expansion only to the orders that we will need.[^8] These relations invert to give =[\_0\^24]{}e\^[2\_p]{}+[\_0\^432]{}, \|=[\_0\^24]{}e\^[-2\_p]{}+[\_0\^432]{} \[eenn\] Substituting in (\[taurel\]) we find =[\_0\^416]{}+[\_0\^664]{} (1-8) 2\_p +O(\_0\^8) Requiring that $\tau$ not depend on $\alpha_p$ then gives =[18]{} \[gamma\] and =[\_0\^416]{} +O(\_0\^8) \[taurelbis\] The relation (\[etwo\]) rewritten in rescaled variables appears as =[\_0\^22]{}2\_p + [2]{} Substituting (\[taurelbis\]) we get = [\_0\^22]{}2\_p+[\_0\^432]{}+ C \_0\^6+O(\_0\^8) with C=0 Thus we find that the coefficient of $\rho_0^6$ in $\epsilon$ vanishes. The determination of this $O(\rho_0^6)$ term is the main result of the computation above. Determination of this term has given us the next correction to the energy beyond (\[e20\]). This term was determined by an explicit and detailed gravity computation in [@cincinnati; @rothstein]. It may appear remarkable that we have found it here by a simple argument involving the invariance of the tension under boosting, and one might wonder where the physics of this correction has been fed into our analysis. This physical input is actually hidden in the assumption that we can expand $\epsilon$ in integer powers of $\nu, \bar \nu$ to the order that we are working. In other words, the vanishing of the $O(\rho_0^6)$ term in the energy does not seem to have a simple interpretation in the gravity computation, but it does have a simple interpretation in the microscopic description. Thus we find that the numbers $n_p \bar n_p$ defined through (\[relations\]) give a good definition of ‘quasi-particle excitations’ to the first few orders in the interaction, and using an expansion in integral powers of these numbers automatically reproduces the $O(\rho_0^6)$ term in the energy. ### First order in $\bar \nu$, all orders in $\nu$ {#largenu} In the above we have worked to second order in the excitations $\nu, \bar \nu$ which correspond to left and right movers respectively. We will now extend our microscopic expression for the energy $\epsilon$ to include terms with arbitrary orders in $\nu$, while keeping only the first order in $\bar \nu$. This means that we allow a large net P charge, since the left moving excitation $\nu$ is no longer small. Our technique will be to again use the fact that the tension ${\cal T}$ does not depend on the boost parameter $\alpha_p$, but now we will take our limits in a slightly different way, as follows. We will still let $\rho_0$ be our small parameter, but we are interested in large boosts $\alpha_p$. Thus we will let \_00, \_p, \_0\^2 e\^[2\_p]{} [finite]{} \[limitnew\] Note that we will thus have \_0\^2e\^[-2\_p]{}\~O(\_0\^4) etc. We now have = +\|-[\|2]{}I(),q= -\|\[theequations\] and =\|(I()+I’())+O(\|\^2) \[taufirst\] To understand the limits that we are taking, look at the gravity description of these variables. First consider the expression for energy in the expansion where $\rho_0$ is small, and $\alpha_p$ is finite. From the way that energy transforms under boosting, we know that =[\_0\^2]{}\^2\_p+\_0(\_0) where $\tilde\epsilon_0(\rho_0)$ is the energy of the system without the P charge; thus $\tilde\epsilon_0(\rho_0)$ does not depend on $\alpha_p$. We can thus write = [\_0\^22]{}2\_p + [\_0\^432]{}+O(\_0\^6), q = [\_0\^22]{}2\_p \[gravityfirst\] where we have used $\cosh 2\alpha_p=2\sinh^2\alpha_p+1$ and the terms $O(\rho_0^6)$ do not depend upon $\alpha_p$. The latter fact then tells us that even when we take the limit (\[limitnew\]) the terms dropped in $\epsilon$ are $O(\rho_0^6)$. The expression for $q$ is exact. We now find q e\^[2\_p]{}q\_0, e\^[-2\_p]{}(1+[q\_04]{}) \[eqnsol\] We will in general keep terms upto order $\rho_0^4$, but we have written the expression for $q$ only to leading order since we will not need higher orders. We invert (\[theequations\]) to find $\nu, \bar \nu$ to the required order =q+[-q2]{}(1-[q4]{}I(q))\^[-1]{}+O((-q)\^2), \|=[-q2]{}(1-[q4]{}I(q))\^[-1]{}+O((-q)\^2) \[nnbarfirst\] Using (\[eqnsol\]) and then substituting in (\[taufirst\]) we have =[\_0\^416]{}(1+[q\_04]{}) (1-[q\_04]{}I(q\_0))\^[-1]{}(I(q\_0)+q\_0 I’(q\_0))+O(\_0\^8) \[tupto8\] Invariance under boost requires (1+[q\_04]{}) (1-[q\_04]{}I(q\_0))\^[-1]{}(I(q\_0)+q\_0 I’(q\_0))=const. Moreover when $q_0=0$ one should recover the previous result (\[taurelbis\]), so the constant in the above equation should be unity. The differential equation (1+[4]{}) (1-[4]{}I())\^[-1]{}(I()+I’())=1 with the boundary condition $I(0)=1$, has a unique solution given by I()=(1+[4]{})\^[-1]{} \[iofn\] (\[iofn\]) is the main result of this subsection. Substituting back in $\epsilon$ in (\[theequations\]) we find that we have the energy to first order in $\bar n_p$ but to all order in $n_p$. We have again obtained this result by a simple argument; the essential physics here was that $n_p$ can be changed by boosting (which adds P charge), and properties of black objects change in simple ways under boosting. Note that the first order expansion of $I(\nu)$ in powers of $\nu$ implies $\gamma={1\over 8}$, in agreement with our earlier result (\[gamma\]). Discussion ========== Let us summarize the computations of this paper. We have noted that gravity computations reveal an interesting phase diagram for the black hole – black string transition, with three phases for a range of $r_0/L$. We can add charges to the neutral system by ‘boosting plus dualities’. This does not add or lose any information as far as the gravity computation is concerned, since the neutral and boosted variables are related in a simple and invertible way. But it does map the problem to a near extremal one, where we can look for microscopic models. For $r_0/L$ very small we get the near extremal 2-charge system in 4+1 noncompact dimensions, and for $r_0/L$ very large we get the near extremal 2-charge system in 3+1 noncompact dimensions. Both these systems have been given a microscopic description in earlier work on the subject. We propose a simple model that interpolates between these two microscopic models; the nontrivial step is guessing what kind of ‘fractionation’ will describe the D1-D5 degrees of freedom in the interpolating regime. We find that the three-phase nature of the phase diagram is reproduced if we assume that the the two large charges $n_1, n_5$ bind to make $N=n_1n_5$ units of an effective charge, and this product $N$ is then partitioned between subsystems that behave like one or the other of the two limiting systems mentioned above. We suggest a physical reason why this kind of ‘fractionation’ is reasonable for this problem. Note that even though the microscopic model looks like a superposition of two subsystems (eq.(\[major\])), we have not just superposed a black hole and a black string; the fact that the [product]{} $N=n_1n_5$ partitions between the two subssytems forces the entire state to be a nontrivial bound state of all the excitations. The non-uniform string appears as a saddle point between two maxima of the entropy; this accords with the fact that the entropy of this branch is always lower than the entropy of the other two phases, so it is always unstable. The $n-\mu$ graph describing the tension of this branch appears to have a profile $n\sim 1/\mu$ from the gravity computation; the microscopic computation also gives the functional behavior $n\sim 1/\mu$. The microscopic computation based on such a simple model does not give any tension for the black hole branch. We have argued that this tension will be seen only if we include interactions like the ‘twist interaction’ that are present in the CFT when we move away from the orbifold point (where we have a ‘free CFT’) to a point in moduli space that has the couplings appropriate to the actual gravity dual. We did not look at such interaction terms directly, but instead followed a ‘mixed’ approach similar to that in [@costaperry; @hkor]. Thus we used the knowledge of the leading order tension in the gravity computation to see what effective interactions we obtain in the CFT. We were then able to reproduce the next order term in the tension by using the invariance of the tension under ‘boosting’. Since boosting adds charge it changes $n_p, \bar n_p$ and the non-extremal energy $E$, and we get a nontrivial condition when we require invariance of the tension ${\cal T}={\p E\over \p L}$. This next order term in the tension is equivalent to a corresponding correction term in the energy, and this correction is seen to agree with detailed computations of the energy carried out in [@cincinnati; @rothstein]. The essential physics we used was to assume that we have an expansion of the energy in integer powers of $n_p, \bar n_p$ to the first few orders; this suggest that the $n_p, \bar n_p$ (defined through (\[relations\])) are good definitions of ‘quasi-particles’ for the system to this order in the interaction. While our microscopic model gives some general qualitative agreements with the gravity phase diagram, there are also some differences. First of course is the issue of tension for the black hole branch mentioned above. Next, consider for example the point where the non-uniform black string branch starts off from the uniform string branch. In the microscopic model this branch starts at a mass value that is $\approx .57$ times the value seen in the gravity diagram. It would also be nice to see in the microscopic model how the black hole branch bends around to join with the non-uniform string branch, and nature of the physics and the joining point. In principle all features of the gravity phase diagram are contained in the correct microscopic dual, so we look upon this phase diagram as a large source of data on the interactions of branes that make up black holes and other black objects. The phase diagram offers us continuous curves that the microscopics must fit; further, we are probing a domain where the black object is changing character from one type to another, so we learn about aspects of the CFT that we do not encounter when looking at just near extremal entropy or low energy Hawking radiation. Thus we consider the model proposed here as just a first pass at understanding the microscopics of branes dual to black objects. Acknowledgements {#acknowledgements .unnumbered} ================ We thank T. Harmark, N. Obers and B. Kol for explaining many aspects of their work to us, and A. Saxena, K. Skenderis and Y. Srivastava for many helpful comments. This work was supported in part by DOE grant DE-FG02-91ER-40690. Map between neutral and 2-charge hole {#first} ===================================== In this appendix, starting from the neutral geometry (\[neutral\]), we generate a geometry carrying D1-D5 charges: the D1 extends along the direction $y$ and the D5 along $y$ and $z_a$, with $a=1,\ldots,4$. Of the remaining $4+1$ directions, one, denoted by $z$, is compactified on a circle of length $L$. We are thus left with $d=4$ non-compact directions. The desired 2-charge solution is generated by acting on (\[neutral\]) with the following sequence of boosts and dualities (we denote by $\mathcal{B}_y$ a boost along $y$ and by $T_{ij\ldots}$ a T-duality in the directions $i,j,\ldots$) && P\_y NS1\_y NS1\_y, P\_y D1\_y, P\_y D5\_[y1234]{}, P\_y\ && NS5\_[y1234]{}, P\_y NS5\_[y1234]{},NS1\_y D5\_[y1234]{}, D1\_y We list the geometries obtained at each step. The notation is as follows: $ds^2$ denotes the string metric, $\Phi$ is the dilaton, $B^{(2)}$ the NS-NS B-field and $H^{(3)}$ its field strength, $C^{(p)}$ is the RR $p$-form gauge potential and $G^{(p+1)}$ the corresponding field strength, $\star_B$ denotes the Hodge dual with respect to $ds^2_B$. - $\mathcal{B}_y$: ds\^2 &=& -H\_5\^[-1]{} U dt\^2 + H\_5 (dy-(H\_5\^[-1]{}-1)\_5 dt)\^2+ds\^2\_B + dz\_a dz\_a\ H\_5&=& 1+(1-U)\^2\_5 - $T_y$: ds\^2 &=& H\_5\^[-1]{} (-U dt\^2+dy\^2) +ds\^2\_B + dz\_a dz\_a\ B\^[(2)]{}\_[ty]{}&=&-(H\_5\^[-1]{}-1)\_5,e\^[2]{}=H\_5\^[-1]{} - $\mathcal{B}_y$: ds\^2 &=& H\_5\^[-1]{} \[-H\_1\^[-1]{} U dt\^2+ H\_1(dy-(H\_1\^[-1]{}-1)\_1dt)\^2\]\ && +ds\^2\_B + dz\_a dz\_a\ B\^[(2)]{}\_[ty]{}&=&-(H\_5\^[-1]{}-1)\_5,e\^[2]{}=H\_5\^[-1]{}\ H\_1&=& 1+(1-U)\^2\_1 - $S$: ds\^2 &=& H\_5\^[-1/2]{} \[-H\_1\^[-1]{} U dt\^2+ H\_1(dy-(H\_1\^[-1]{}-1)\_1dt)\^2\]\ &&+H\_5\^[1/2]{}(ds\^2\_B + dz\_a dz\_a)\ C\^[(2)]{}\_[ty]{}&=&-(H\_5\^[-1]{}-1)\_5,e\^[2]{}=H\_5 - $T_{1234}$: ds\^2 &=& H\_5\^[-1/2]{} \[-H\_1\^[-1]{} U dt\^2+ H\_1(dy-(H\_1\^[-1]{}-1)\_1dt)\^2\]\ &&+H\_5\^[1/2]{}ds\^2\_B + H\_5\^[-1/2]{}dz\_a dz\_a\ C\^[(6)]{}\_[tyz\_1…z\_4]{}&=&-(H\_5\^[-1]{}-1)\_5, e\^[2]{}=H\_5\^[-1]{}\ G\^[(7)]{}\_[rtyz\_1…z\_4]{}&=& -\_r H\_5\^[-1]{}\_5,G\^[(3)]{}=-U\^[-1/2]{} \_r H\_5 \_5 \_B dr - $S$: ds\^2 &=& -H\_1\^[-1]{} U dt\^2+ H\_1(dy-(H\_1\^[-1]{}-1)\_1dt)\^2\ &&+H\_5 ds\^2\_B + dz\_a dz\_a\ H\^[(3)]{}&=&-U\^[-1/2]{} \_r H\_5 \_5 \_B dr, e\^[2]{}=H\_5 - $T_{y1}$: ds\^2 &=& H\_1\^[-1]{}(- U dt\^2+ dy\^2) +H\_5 ds\^2\_B + dz\_a dz\_a\ H\^[(3)]{}&=&-U\^[-1/2]{} \_r H\_5 \_5 \_B dr-\_r H\_1\^[-1]{}\_1 drdtdy\ e\^[2]{}&=&[H\_5H\_1]{} - $S$: ds\^2 &=& (H\_5H\_1)\^[-1/2]{}(- U dt\^2+ dy\^2) + (H\_5H\_1)\^[1/2]{} ds\^2\_B + ([H\_1H\_5]{})\^[1/2]{} dz\_a dz\_a\ G\^[(3)]{}&=&-U\^[-1/2]{} \_r H\_5 \_1 \_B dr-\_r H\_1\^[-1]{}\_1 drdtdy\ e\^[2]{}&=&[H\_1H\_5]{} Converting from string to Einstein frame, we find ds\^2\_E &=& H\_1\^[1/4]{}H\_5\^[-1/4]{}\[H\_1\^[-1]{}(- U dt\^2+ dy\^2) + H\_5 ds\^2\_B + dz\_a dz\_a\]\ G\^[(3)]{}&=&-U\^[-1/2]{} \_r H\_5 \_5 \_B dr-\_r H\_1\^[-1]{}\_1 drdtdy\ e\^[2]{}&=&[H\_1H\_5]{},H\_1 = 1+(1-U)\^2\_1,H\_5 = 1+(1-U)\^2\_5 \[chargedEApp\] Energy and tension are derived from the large $r$ behavior of the Einstein metric. Assume that the neutral metric has the following asymptotic expansion -g\_[tt]{}=1-[c\_tr]{},g\_[zz]{}=1+[c\_zr]{}, g\_[yy]{}=1, g\_[z\_a z\_a]{}= 1\ Let $M$ and $\mathcal{T}$ be the mass and tension of the neutral system, and define the relative tension n =[LM]{} \[relativetens\] Einstein’s equations relate these quantities to the $c_t$ and $c_z$ defined above: M= [L16G\_N\^[(5)]{}]{} \_2 (2 c\_t -c\_z),=[\_216G\_N\^[(5)]{}]{}(c\_t - 2 c\_z),n = [c\_t - 2 c\_z2 c\_t -c\_z]{} \[neutralmass\] Let $g'_{\mu\nu}$ denote the components of the metric in (\[chargedEApp\]), $C^{(2)}$ the RR 2-form gauge field and $C^{(6)}$ the dual 6-form field. Their asymptotic expansions are of the form &&-g’\_[tt]{}=1-[c’\_tr]{},g’\_[zz]{}=1+[c’\_zr]{}, g’\_[yy]{}=1+[c’\_yr]{}, g’\_[z\_a z\_a]{}= 1+[c’\_ar]{}\ &&C\^[(2)]{}\_[ty]{}=[c’\_1r]{},C\^[(6)]{}\_[t y z\_1…z\_4]{}=[c’\_5r]{} Let $M'$ and $\mathcal{T}'$ be the mass and tension of the charged system and $Q_1$ and $Q_5$ denote the D1 and D5 charges (with a normalization such that $M' = Q_1 + Q_5$ for the extremal hole). The relations linking mass, tension and charges to the asymptotic fall-off of the metric are &&M’ =[L16G\_N\^[(5)]{}]{} \_2 (2 c’\_t -c’\_z-c’\_y -4 c’\_a),’= [\_216G\_N\^[(5)]{}]{}(c’\_t - 2 c’\_z -c’\_y - 4 c’\_a)\ &&Q\_1 = [L16G\_N\^[(5)]{}]{} \_2 c’\_1,Q\_5 =[L16G\_N\^[(5)]{}]{} \_2 c’\_5 \[chargedmass\] From (\[chargedEApp\]) we find the relations between the coefficients $c_i$ and $c'_i$: c’\_t &=& c\_t +[34]{}c\_t \^2\_1 + [14]{}c\_t \^2\_5, c’\_z = c\_z +[14]{}c\_t \^2\_1 + [34]{}c\_t \^2\_5\ c’\_y &=& -[34]{}c\_t \^2\_1 - [14]{}c\_t \^2\_5, c’\_u = [14]{}c\_t \^2\_1 - [14]{}c\_t \^2\_5\ c’\_1 &=& c\_t \_1\_1,c’\_5 = c\_t \_5\_5 \[c’s\] We deduce from (\[neutralmass\]), (\[chargedmass\]) and (\[c’s\]) that M’ = M(1+[2-n3]{}(\^2\_1+\^2\_5)),’=, Q\_i =M [2-n3]{}\_i \_i(i=1,5) \[2charge\] Let us now consider temperature and entropy. Temperature is proportional to the surface gravity $\kappa$, which is given by \^2 = -[14]{} \[g\^[tt]{}g\^[rr]{}(\_r g\_[tt]{})\^2\]\|\_[r=r\_0]{} \[surfacegravity\] with $r=r_0$ the location of the horizon. From (\[chargedE\]), we find that the components of the 2-charge metric are related as follows to the components of the neutral metric (\[neutral\]) g’\_[tt]{}=(H\_1\^[1/4]{}H\_5\^[-1/4]{})H\_1\^[-1]{}g\_[tt]{},g’\_[rr]{}=(H\_1\^[1/4]{}H\_5\^[-1/4]{})H\_5 g\_[rr]{} The surface gravity of the 2-charge system is thus ’\^2 &=& -[14]{} \[g’\^[tt]{}g’\^[rr]{}(\_r g’\_[tt]{})\^2\]\|\_[r=r\_0]{}\ &=&\^2 (H\_1 H\_5)\^[-1]{}\|\_[r=r\_0]{}=[\^2\^2\_1\^2\_2]{} where we have used the fact that $g_{tt}=0$ at the horizon and that H\_i\|\_[r=r\_0]{}=\^2\_i(i=1,5) Then the temperatures of the neutral and charged holes ($T$ and $T'$) are related as T’ = [T\_1\_5]{} \[temperature\] Entropy is proportional to the horizon area. Using again (\[chargedE\]), we see that the horizon areas before and after the boost ($A_h$ and $A'_h$) are related as A’\_h = (H\_1 H\_5)\^[1/2]{} A\_h and thus the respective entropies ($S$ and $S'$) scale as S’ = S \_1 \_5 \[entropyApp\] Note that T S= T’ S’ [99]{} R. Gregory and R. Laflamme, Phys. Rev. Lett. [70]{}, 2837 (1993) \[arXiv:hep-th/9301052\]; R. Gregory and R. Laflamme, Nucl. Phys. B [428]{}, 399 (1994) \[arXiv:hep-th/9404071\]. S. S. Gubser, Class. Quant. Grav. [19]{}, 4825 (2002) \[arXiv:hep-th/0110193\]. T. Harmark and N. A. Obers, JHEP [0205]{}, 032 (2002) \[arXiv:hep-th/0204047\]. B. Kol, JHEP [0510]{}, 049 (2005) \[arXiv:hep-th/0206220\]. T. Wiseman, Class. Quant. Grav. [20]{}, 1137 (2003) \[arXiv:hep-th/0209051\]; T. Wiseman, Class. Quant. Grav. [20]{}, 1177 (2003) \[arXiv:hep-th/0211028\]. B. Kol and T. Wiseman, Class. Quant. Grav. [20]{}, 3493 (2003) \[arXiv:hep-th/0304070\]. T. Harmark and N. A. Obers, Class. Quant. Grav. [21]{}, 1709 (2004) \[arXiv:hep-th/0309116\]. B. Kol, E. Sorkin and T. Piran, Phys. Rev. D [69]{}, 064031 (2004) \[arXiv:hep-th/0309190\]. T. Harmark and N. A. Obers, Nucl. Phys. B [684]{}, 183 (2004) \[arXiv:hep-th/0309230\]. E. Sorkin, B. Kol and T. Piran, Phys. Rev. D [69]{}, 064032 (2004) \[arXiv:hep-th/0310096\]. H. Kudoh and T. Wiseman, Prog. Theor. Phys. [111]{}, 475 (2004) \[arXiv:hep-th/0310104\]; H. Kudoh and T. Wiseman, Phys. Rev. Lett. [94]{}, 161102 (2005) \[arXiv:hep-th/0409111\]. T. Harmark, Phys. Rev. D [69]{}, 104015 (2004) \[arXiv:hep-th/0310259\]. E. Sorkin, Phys. Rev. Lett. [93]{}, 031601 (2004) \[arXiv:hep-th/0402216\]. D. Gorbonos and B. Kol, JHEP [0406]{}, 053 (2004) \[arXiv:hep-th/0406002\]. O. Aharony, J. Marsano, S. Minwalla and T. Wiseman, Class. Quant. Grav. [21]{}, 5169 (2004) \[arXiv:hep-th/0406210\]. B. Kol and E. Sorkin, Class. Quant. Grav. [21]{}, 4793 (2004) \[arXiv:gr-qc/0407058\]; B. Kol and E. Sorkin, Class. Quant. Grav. [23]{}, 4563 (2006) \[arXiv:hep-th/0604015\]. T. Harmark and N. A. Obers, JHEP [0409]{}, 022 (2004) \[arXiv:hep-th/0407094\]; T. Harmark and N. A. Obers, Nucl. Phys. B [742]{}, 41 (2006) \[arXiv:hep-th/0510098\]. D. Karasik, C. Sahabandu, P. Suranyi and L. C. R. Wijewardhana, Phys. Rev. D [71]{}, 024024 (2005) \[arXiv:hep-th/0410078\]. B. Kol, Phys. Rept. [422]{}, 119 (2006) \[arXiv:hep-th/0411240\]. T. Harmark and N. A. Obers, arXiv:hep-th/0503020. D. Gorbonos and B. Kol, Class. Quant. Grav. [22]{}, 3935 (2005) \[arXiv:hep-th/0505009\]. J. L. Hovdebo and R. C. Myers, Phys. Rev. D [73]{}, 084013 (2006) \[arXiv:hep-th/0601079\]. Y. Z. Chu, W. D. Goldberger and I. Z. Rothstein, JHEP [0603]{}, 013 (2006) \[arXiv:hep-th/0602016\]. B. Kleihaus, J. Kunz and E. Radu, JHEP [0606]{}, 016 (2006) \[arXiv:hep-th/0603119\]. T. Harmark, K. R. Kristjansson, N. A. Obers and P. B. Ronne, arXiv:hep-th/0606246. A. Murugan and V. Sahakian, arXiv:hep-th/0608103. C. G. . Callan and J. M. Maldacena, Nucl. Phys. B [472]{}, 591 (1996) \[arXiv:hep-th/9602043\]. J. M. Maldacena and A. Strominger, Phys. Rev. D [55]{}, 861 (1997) \[arXiv:hep-th/9609026\]. J. M. Maldacena, Nucl. Phys. B [477]{}, 168 (1996) \[arXiv:hep-th/9605016\]. G. T. Horowitz, J. M. Maldacena and A. Strominger, Phys. Lett. B [383]{}, 151 (1996) \[arXiv:hep-th/9603109\]. G. T. Horowitz, D. A. Lowe and J. M. Maldacena, Phys. Rev. Lett. [77]{}, 430 (1996) \[arXiv:hep-th/9603195\]. S. R. Das and S. D. Mathur, Nucl. Phys. B [478]{}, 561 (1996) \[arXiv:hep-th/9606185\]. I. R. Klebanov and S. D. Mathur, Nucl. Phys. B [500]{}, 115 (1997) \[arXiv:hep-th/9701187\]. S. R. Das and S. D. Mathur, Phys. Lett. B [375]{}, 103 (1996) \[arXiv:hep-th/9601152\]. J. M. Maldacena and L. Susskind, Nucl. Phys. B [475]{}, 679 (1996) \[arXiv:hep-th/9604042\]. M. S. Costa and M. J. Perry, Nucl. Phys. B [591]{}, 469 (2000) \[arXiv:hep-th/0008106\]. S. D. Mathur, Nucl. Phys. B [529]{}, 295 (1998) \[arXiv:hep-th/9706151\]. S. D. Mathur, Class. Quant. Grav. [23]{}, R115 (2006) \[arXiv:hep-th/0510180\]. A. Strominger and C. Vafa, Phys. Lett. B [379]{}, 99 (1996) \[arXiv:hep-th/9601029\]. I. R. Klebanov, J. M. Maldacena and C. B. Thorn, JHEP [0604]{}, 024 (2006) \[arXiv:hep-th/0602255\]. O. Lunin and S. D. Mathur, Commun. Math. Phys. [219]{}, 399 (2001) \[arXiv:hep-th/0006196\]; O. Lunin and S. D. Mathur, Commun. Math. Phys. [227]{}, 385 (2002) \[arXiv:hep-th/0103169\]. [^1]: E-mail: [[email protected]]{}. [^2]: E-mail: [[email protected]]{}. [^3]: E-mail: [[email protected]]{}. [^4]: It is easy to check that the same result is obtained if we start with the definition $\mathcal{T}=\Big ({\partial E\over \partial L}\Big )_{S, Q_i}$ and compute the requited variation of $E$ keeping $S, Q_i$ fixed. [^5]: In the second equation in (\[smarrGL\]) we have used the value of the uniform string tension in $d$-dimensions: $n_\mathrm{us}=1/(d-2)$. [^6]: A review is given in [@review2]. [^7]: This can be seen as follows. For the 3-charge system the energy $M'$ and the charges are given by M’&=&M(1+[2-n3]{}(\^2\_1+\^2\_5+\^2\_p))\ Q\_i &=& M [2-n3]{} \_i\_i,i=1,5,p Taking all the charges to be large ($\alpha_i\gg 1$) one finds M’=Q\_1+Q\_5+Q\_p+M [n2]{} E = M [n2]{} Since the absolute tension of the neutral system is the same of the one of the 3-charge system, the rescaled tension is r = [ME]{} n = 2 [^8]: Since $\nu,\bar \nu\sim \rho_0^2$, if follows from (\[taurel\]) that to compute $\tau$ up to order $\rho_0^6$ we only need $\epsilon$ and $q$ up to order $\rho_0^4$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The multi-stream paradigm of audio processing, in which several sources are simultaneously considered, has been an active research area for information fusion. Our previous study offered a promising direction within end-to-end automatic speech recognition, where parallel encoders aim to capture diverse information followed by a stream-level fusion based on attention mechanisms to combine the different views. However, with an increasing number of streams resulting in an increasing number of encoders, the previous approach could require substantial memory and massive amounts of parallel data for joint training. In this work, we propose a practical two-stage training scheme. Stage-1 is to train a Universal Feature Extractor (UFE), where encoder outputs are produced from a single-stream model trained with all data. Stage-2 formulates a multi-stream scheme intending to solely train the attention fusion module using the UFE features and pretrained components from Stage-1. Experiments have been conducted on two datasets, DIRHA and AMI, as a multi-stream scenario. Compared with our previous method, this strategy achieves relative word error rate reductions of 8.2–32.4%, while consistently outperforming several conventional combination methods.' address: \| $^1$Center for Language and Speech Processing, The Johns Hopkins University, USA\ $^2$Human Language Technology Center of Excellence, The Johns Hopkins University, USA\ $^3$Speech and Dialog Research Group, Microsoft, USA bibliography: - 'strings.bib' - 'refs.bib' title: \| A Practical Two-Stage Training Strategy for multi-stream\ end-to-end speech recognition --- End-to-End Speech Recognition, Multi-Stream, Multiple Microphone Array, Two-Stage Training introduction {#sec:intro} ============ The multi-stream paradigm in speech processing considers scenarios where parallel streams carry diverse or complementary task-related knowledge. In these cases, an appropriate strategy to fuse streams or select the most informative source is necessary. One potential source of inspiration in this setting is from the observations of parallel processing in the human auditory system, and resulting innovations have been successfully applied to conventional automatic speech recognition (ASR) frameworks [@mallidi2018practical; @hermansky2013multistream; @mallidi2016novel; @hermansky2018coding]. For instance, multi-band acoustic modeling was formulated to address noise robustness [@mallidi2018practical; @mallidi2016novel]. [@meyer2016performance] investigated several performance measures in spatial acoustic scenes to choose the most reliable source for hearing aids. The multi-modal applications combine visual [@palaskar2018end] or symbolic [@renduchintala2018multi] inputs together with audio signal to improve speech recognition. The work that follows considers far-field ASR using multiple microphone arrays, a specific case of multi-stream paradigm. Without any knowledge of speaker-array distance or orientation, it is still challenging to speculate which array is most informative or least corrupted. The common methods of utilizing multiple arrays in conventional ASR are posterior combination [@wang2018stream; @xiong2018channel], ROVER [@fiscus1997post], distributed beamformer [@yoshioka2019meeting], and selection based on Signal-to-Noise/Interference Ratio (SNR/SIR) [@du2018theustc]. In recent years, with the increasing use of Deep Neural Networks (DNNs) in ASR, End-to-End (E2E) speech recognition approaches, which directly transcribe human speech into text, have received greater attention. The E2E models combine several disjoint components (acoustic model, pronunciation model, language model) from hybrid ASR into one single DNN for joint training. Three dominant E2E architectures for ASR are Connentionist Temporal Classification (CTC) [@graves2006connectionist; @graves2014towards; @miao2015eesen], attention-based encoder decoder [@chan2015listen; @chorowski2015attention], and Recurrent Neural Network Transducer (RNN-T) [@graves2012sequence; @graves2013speech]. Coupled with a CTC network within a multi-task scheme, the joint CTC/Attention model [@kim2016joint_icassp2017; @hori2017advances; @watanabe2017hybrid] outperforms the attention-based model by addressing misalignment issues, achieving the state-of-the-art E2E performance on several benchmark datasets [@watanabe2017hybrid]. In [@li2019multistream], we proposed a novel multi-stream model based on a joint CTC/Attention E2E scheme, where each stream is characterized by a separate encoder and CTC network. A Hierarchical Attention Network (HAN) [@li2018multiencoder; @wang2019stream] acts as a fusion component to dynamically assign higher weights for streams carrying more discriminative information for prediction. The Multi-Encoder Multi-Array (MEM-Array) framework was introduced in [@li2019multistream] to improve the robustness of distant microphone arrays, where each array is represented by a separate encoder. While substantial improvements were reported within a two-stream configuration, there are two concerns when more streams are involved. First, during training, fitting all parallel encoders in device computing memory is potentially impractical for joint optimization, as the encoder is typically the largest component by far, i.e., 88% of total parameters in this work. Second, due to the data-hungry nature of DNNs and the expensive cost of collecting parallel data, training multiple models with excess degrees of freedom is not optimal. In this paper, we present a practical two-stage training strategy on the MEM-Array framework targeting the aforementioned issues. The proposed technique has the following highlights: 1. In Stage-1, a single-stream model is trained using all data for better model generalization. The encoder will then acts as a Universal Feature Extractor (UFE) to process parallel data individually to generate a set of high-level parallel features. 2. Initializing components (CTC, decoder, frame-level attention) from Stage-1, Stage-2 training only optimizes the HAN component operating directly on UFE parallel features. The resulting memory and computation savings greatly simplify training, potentially allowing for more hyperparameter exploration or consideration of more complicated architectures. 3. Lack of adequate volume of data, specially parallel data, leads to overfit or is hard to tackle unseen data. The proposed two-stage strategy better defines the data augmentation scheme. Augmentation in Stage-1 aims to extract more discriminative high-level features and provides well-pretrained modules for Stage-2, whereas Stage-2 could focus on improving the robustness of information fusion. MEM-Array End-to-End Speech Recognition {#sec:memarr} ======================================= In this session, we review the joint CTC/Attention framework and the extended MEM-Array model, one realization of the multi-stream approach with focus on distant multi-array scenario. Joint CTC/Attention {#ssec:ctc/att} ------------------- The joint CTC/Attention architecture, illustrated in Stage-1 of Fig. \[fig:2stage\], takes advantage of both CTC and attention-based models within a Multi-Task Learning (MTL) scheme. The model directly maps a $T$-length sequence of $D$-dimensional speech vectors, $X=\{\textbf{x}_{t}\in \mathbb{R}^{D}\|t = 1,2,...,T\}$, into an $L$-length label sequence, $C=\{c_{l}\in \mathcal{U}\|l = 1,2,...,L\}$. Here $\mathcal{U}$ is a set of distinct labels. The encoder transforms the acoustic sequence $X$ into a higher-level feature representation $H=\{\textbf{h}_{1},..., \textbf{h}_{\lfloor T/s\rfloor}\}$, which is shared for the use of CTC and attention-based models. Here, $\lfloor T/s\rfloor$ time instances are generated at the encoder-output level with a subsampling factor of $s$. The loss function to be optimized is a logarithmic linear combination of CTC and attention objectives, i.e., $p_\textrm{ctc}(C\|X)$ and $p_\textrm{att}^{\dagger}(C\|X)$: $$\label{f:mtl} \mathcal{L}_\textrm{MTL}=\lambda\log p_\textrm{ctc}(C\|X)+(1-\lambda)\log p_\textrm{att}^{\dagger}(C\|X),$$ where $\lambda \in [0,1]$ is a hyper parameter. Note that $p_\textrm{att}^{\dagger}(C\|X)$ is an approximated letter-wise objective where the probability of a prediction is conditioned on previous true labels. During inference, a label-synchronous beam search is employed to predict the most probable label sequence $\hat{C}$: $$\begin{aligned} \label{f:jointdec} \hat{C}=\arg\max_{C\in \mathcal{U}^{}} &\{\lambda \log p_\textrm{ctc}(C\|X)+(1-\lambda)\log p_\textrm{att}(C\|X) \nonumber \\ &+\gamma \log p_\textrm{lm}(C)\}, \end{aligned}$$ where $\log p_\textrm{lm}(C)$ is evaluated from an external Recurrent Neural Network Language Model (RNN-LM) with a scaling factor $\gamma$. MEM-Array Model {#ssec:memarr} --------------- An end-to-end ASR model addressing the general multi-stream setting was introduced in [@li2019multistream]. As one representative framework, MEM-Array concentrates on cases of far-field microphone arrays to handle different dynamics of streams. The architecture of $N$ streams is shown in Stage-2 of Fig. \[fig:2stage\]. Each encoder operates separately on a parallel input $X^{(i)}$ to extract a set of frame-wise hidden vectors $H^{(i)}$: $$\label{eq:enc} H^{(i)}=\textrm{Encoder}^{(i)}(X^{(i)}), i\in\{1, ..., N\},$$ where we denote superscript $i$ as the index for stream $i$, and $H^{(i)}=\{\textbf{h}^{(i)}_{1},..., \textbf{h}^{(i)}_{\lfloor T^{(i)}/s\rfloor}\}$. A frame-level attention mechanism is designated to each encoder to carry out the stream-specific speech-label alignment. For stream $i$, the letter-wise context vector $\textbf{r}_{l}^{(i)}$ is computed via a location-based attention network [@NIPS2015_5847] as follows: $$\label{f:cv1} \textbf{r}_{l}^{(i)}={\sum}_{t=1}^{\lfloor T^{(i)}/s^{(i)}\rfloor}a_{lt}^{(i)}\textbf{h}_{t}^{(i)},$$ $$\label{f:cv2} a^{(i)}_{lt} = \textrm{Attention}(\{a_{l-1}^{(i)}\}^{T^{(i)}}_{t=1}, \textbf{q}_{l-1}, \textbf{h}^{(i)}_t),$$ where ${a}^{(i)}_{lt}$ is the attention weight, a soft-alignment of $\textbf{h}^{(i)}_t$ for output $c_{l}$, and $\textbf{q}_{l-1}$ is the previous decoder state. In the multi-stream setting, the contribution of each stream changes dynamically. Hence, a secondary stream attention, the HAN component, is exploited for the purpose of robustness. The fusion context vector $\textbf{r}_l$ is obtained as a weighted summation of $\{\textbf{r}_l^{(i)}\}^{N}_{i=1}$: $$\label{f:han} \textbf{r}_{l}={\sum}_{i=1}^{N}\beta_{l}^{(i)}\textbf{r}_{l}^{(i)},$$ $$\label{f:l2att} \beta_{l}^{(i)}=\textrm{HierarchicalAttention}(\textbf{q}_{l-1}, \textbf{r}_l^{(i)}), i\in\{1, ..., N\}.$$ where in this work, a content-based attention network [@NIPS2015_5847] is applied here, and $\beta_{l}^{(i)}$ is a Softmax output across $\{i\}^{N}_{1}$ from the HAN component, a stream-level attention weight for array $i$ of prediction $c_{l}$. In addition, a separate CTC network is active for each encoder to enhance the stream diversity instead of sharing a CTC across all streams. In this setting, the MEM-Array model follows Eq. (\[f:mtl\]) and (\[f:jointdec\]) with a modified CTC objective: $$\log p_\textrm{ctc}(C\|X)=\frac{1}{N}{\sum}_{i=1}^{N}\log p_{\textrm{ctc}^{(i)}}(C\|X),$$ where joint CTC loss is the average of per-encoder CTCs. Proposed Training Strategy ========================== In this section, we present a practical two-stage training strategy for the MEM-Array model, depicted in Fig. \[fig:2stage\]. The details of each stage are discussed in the following sections. \[sec:twostage\] ![Proposed Two-Stage Training Strategy. Color “green” indicates the components are trainable; Color “blue” means parameters of the components are frozen. []{data-label="fig:2stage"}](streamattention.pdf){width="9.1cm"} Stage 1: Universal Feature Extractor {#ssec:stage1} ------------------------------------ The intent of Stage-1 is to obtain a single well-trained encoder, which we refer to as Universal Feature Extractor (UFE), to prepare a new set of high-level features for Stage-2. Encoder in E2E model can be viewed as an acoustic modeling that generates sequences $H=\{\textbf{h}_{1},..., \textbf{h}_{\lfloor T/s\rfloor}\}$ with more discrinimative power for prediction. We denote the encoder outputs $H$ as the UFE features. In general, the majority of the overall parameters are contained in the encoder. In Stage-1, a single-stream joint CTC/Attention model is optimized as shown in Fig. \[fig:2stage\]. Audio features from all available streams are used to train the model. After training, we extract UFE features $H^{(i)}=\{\textbf{h}^{(i)}_{1},..., \textbf{h}^{(i)}_{\lfloor T^{(i)}/s\rfloor}\}$ for each stream $i$, separately. Since subsampling mitigates the increased dimension of UFE features, it is possible to save the UFE features at a similar size to the original speech features. Moreover, byproducts in Stage-1, such as decoder, CTC and attention, can be used for initialization in Stage-2. Stage 2: Parallel-Stream Fusion {#ssec:stage2} ------------------------------- As illustrated in Fig. \[fig:2stage\], Stage-2 focuses on training the fusion component within the multi-stream context. The MEM-Array model uses parallel encoders as information streams. The previous strategy uses joint training with multiple large encoders, which is expensive in memory and time for more complex models or more streams. Taking advantage of UFE features greatly alleviates this complication. In Stage-2, we formulate a multi-stream scheme on UFE features $\{H^{(i)}\}^N_{i=1}$ as parallel inputs. In this model, parameters of all components, except the stream attention module, are initialized from Stage-1 and frozen during optimization. The stream fusion component is randomly initialized, and is the only trainable element in Stage-2. Without any involvement of encoders, frame-level attention directly operates on UFE features. This setup not only reduces the amount of required parallel data, but it also greatly reduces memory and time requirements, allowing for more thorough hyperparameter exploration or utilization of more complex architectures. Experimental Setup {#sec:exptsetup} ================== We demonstrated the two-stage training strategy using two datasets: DIHRA English WSJ [@ravanelli2016realistic] and AMI Meeting Corpus [@carletta2005ami]. The DIRHA English WSJ is part of the DIRHA project, which focuses on speech interaction in domestic scenes via distant microphones. There are in total 32 microphones placed in an apartment setting with a living room and a kitchen. We chose a 6-mic circular array (Beam Circular Array) and an 11-mic linear array (Beam Linear Array) in the living room for experiments with two parallel streams. Additionally, a single microphone (L1C) was picked to serve as a third stream in 3-stream experiments. Training data was created by contaminating original Wall Street Journal data (WSJ0 and WSJ1, 81 hours per stream), providing room impulse responses for each stream. Simulated WSJ recordings with domestic background noise and reverberation were used as the development set for cross validation. The evaluation set has 409 WSJ recordings uttered in real domestic scenario. The AMI Meeting Corpus was collected in three instrumented rooms with meeting conversations. Each room has two microphone arrays to collect 100 hours of far-field signal-synchronized recordings. With no speakers overlapping, the training, development and evaluation set have 81 hours, 9 hours and 9 hours of meeting recordings, respectively. No close-talk microphone recordings are used here. We designed both 2-stream and 3-stream settings for DIRHA and 2-stream experiments for AMI. Note that for each array, the multi-channel input was synthesized into single-channel audio using Delay-and-Sum beamforming with BeamformIt [@anguera2007acoustic]. Experiments were conducted using a Pytorch back-end on ESPnet [@watanabe2018espnet] configured as described in Table \[tab:config\]. [ll]{} & [80-dim log-mel filter bank + 3-dim pitch]{}\ \ [Encoder type]{} & [VGGBLSTM [@hori2017advances; @cho2018multilingual] (subsampling factor: 4)]{}\ [Encoder layers]{} & [6(CNN)+2(BLSTM)]{}\ [Encoder units ]{} & [320 cells (BLSTM layers)]{}\ [Encoder projection]{} & [320 cells (BLSTM layers)]{}\ [Frame-level Attention]{} & [320-cell Content-based]{}\ [Stream Attention]{} & [320-cell Location-based]{}\ [Decoder type]{} & [1-layer 300-cell LSTM]{}\ \ [Optimizer]{} & [AdaDelta (Batch size: 15)]{}\ [Training Epoch]{} & [30 epochs (patience:3 epochs)]{}\ [CTC weight $\lambda$]{} & [0.2 (train); 0.3 (decode)]{}\ [Label Smoothing]{} & [Type: Unigram [@pereyra2017regularizing], Weight: 0.05]{}\ \ [Type]{} & [Look-ahead Word-level RNNLM [@hori2018end]]{}\ [Train data]{} & [AMI:AMI; DIRHA:WSJ0-1+extra WSJ text]{}\ [LM weight $\gamma$]{} & [AMI:0.5; DIRHA:1.0]{}\ results and discussions {#sec:expts} ======================= Firstly, we examined UFE features in a single-stream setting. Next, the full proposed strategy was analyzed in comparison to the previous approach as well as to several conventional fusion methods on DIRHA 2-stream case. Results on AMI and extension with more streams on DIRHA were explored as well. Lastly, we considered the potential benefits of data augmentation in this framework. Effectiveness of Two-Stage Training ----------------------------------- In this section, we discuss the results on 2-stream DIRHA to demonstrate the value of proposed strategy. First, to evaluate Stage-1 training, Character/Word Error Rates (CER/WER) results on single stream systems are summarized in Table \[tab:single-stream\]. Training the model using data from both streams improves performance substantially on the individual arrays, i.e., $37.6\%\rightarrow{}33.9\%$ and $39.2\%\rightarrow{}30.7\%$. The UFE features are the outputs of an encoder trained with this improved strategy. In our setup, 320-dimensional UFE features took slightly smaller space than 83-dimensional acoustic frames since the subsampling factor $s=4$. ------------------------------ ---------- ---------- ---------- ---------- Train Data CER(%) WER(%) CER(%) WER(%) $\text{Arr}_1$ 22.3 37.6 – – $\text{Arr}_2$ – – 23.0 39.2 $\text{Arr}_1, \text{Arr}_2$ 20.1* 33.9 17.9 30.7 ------------------------------ ---------- ---------- ---------- ---------- : Stage-1 results on 2-stream DIRHA. \[tab:single-stream\] Table \[tab:two-stage\] illustrates several training strategies in Stage-2. Since Stage-2 operates on UFE features directly, its training only involves, at most, frame-level attention (ATT), decoder (DEC), hierarchical attention (HAN) and CTC. These experiments considered which of these components should be initialized from their Stage-1 counterparts, as well as which components should be fine-tuned or frozen during Stage-2 updates. In both cases of fine-tuning or freezing Pre-Trained (PT) modules in Stage-2, more noticeable improvements were reported with introducing more pretraining knowledge, i.e., $32.9\%\rightarrow{}28.4\%$ and $31.8\%\rightarrow{}26.8\%$, respectively. Moreover, keeping all PT components frozen during Stage-2 and training solely the fusion module showed relative WER reduction of 5.6% ($28.4\%\rightarrow{}26.8\%$) with only 0.2 million active parameters. Overall, a substantial improvement of 18.8% relative WER reduction ($33.0\%\rightarrow{}26.8\%$) was observed compared to jointly training a massive model, including encoders, from scratch. --------------------- ----------- ------------------ Initialization with Fine-tune Freeze PT Comp. PT Comp. PT Comp. [No Two-Stage]{} Baseline – 33.0 (21.82M) [Two-Stage]{} – 32.9 31.8 (1.78M) CTC 34.4 30.7 (1.75M) ATT 33.3 30.6 (1.37M) ATT, DEC 29.0 27.4 (0.23M) ATT, DEC, CTC 28.4 26.8 (0.20M) --------------------- ----------- ------------------ : WER(%) Comparison among various Stage-2 training strategies on 2-stream DIRHA. Note that components with random initialization in Stage-2 are listed in parentheses of first column. The amount of trainable parameters in Stage-2 when freezing Stage-1 Pre-Trained (PT) components is stated in parentheses of last column. \[tab:two-stage\] Multi-Stream v.s. Conventional Methods -------------------------------------- In Table \[tab:more\], the MEM-Array model with our two-stage training strategy consistently outperforms the baseline model which needs joint training after random initialization. 18.8%, 32.4%, and 8.2% relative WER reductions are achieved in 2-stream DIRHA, 3-stream DIRHA, and 2-stream AMI, respectively. Note that AMI experients were conducted using VGGBLSTM with 2-layer BLSTM layers without any close-talk recordings and data perturbations. It is worth mentioning that those reductions in WERs were accomplished while simultaneously significantly decreasing the number of unique parameters in training by avoiding costly multiples of the large encoder component (10 million parameters per stream, in this case). In addition, results from several conventional fusion strategies are shown in Table \[tab:more\]: signal-level fusion via WAV alignment and average; feature-level frame-by-frame concatenation; word-level prediction fusion using ROVER. For fair comparison, single-level and word-level fusion models utilized Stage-1 pre-trained models as their initialization. Note that word-level fusion operates on decoding results from pretrained single-stream from Stage-1. Still, our proposed strategy consistently perform better than all other fusion methods in all conditions. [lcccc]{} & Unique &\ & Params & &AMI\ Model & (in million) & 2 & 3 &2\ \ Baseline [@li2019multistream] &21.8(2),32.1(3) &33.0& 32.1& 59.5\ Proposed Strategy &11.6& 26.8&21.7& 54.6\ \ WAV Align.& Avg. &11.4 & 32.4& 30.1& 55.9\ Frame Concat.& 16.9(2),23.8(3) &33.7 &33.8 &59.4\ ROVER & 11.4 & 34.2& 23.6&58.0\ \[tab:more\] Discussion on Data Augmentation ------------------------------- The two-stage training strategy provides various opportunities for data augmentation. Stage-1 does not consider parallel data, so any augmentation technique for regular E2E ASR could be applied in this stage to improve the robustness of the UFE. Stage-2 augmentation, on the other hand, would be expected to improve robustness of the combination of corrupted individual streams. In this study, we employed a simple data augmentation technique called SpecAugment [@Park2019SpecAugmentAS], which randomly removes sections of the input signal in a structured fashion, to demonstrate the potential of this direction. Table \[tab:specaug\] shows the improvements from applying SpecAugment on two separate training stages. The best performance was from data augmentation on Stage-1 when freezing all Stage-1 pretrained components. With additional Stage-2 SpecAugment, there was not a noticeable difference in terms of WERs ($22.6\%$ v.s. $22.4\%$ and $22.6\%$ v.s. $22.5\%$). 10% absolute WER reduction was achived in AMI with two stage augmentation. However, it is important to remember that, while the performance gap from fine-tuning versus freezing pre-trained components is narrowed with Stage-2 augmentation, the reductions in Stage-2 memory and computation requirements are still substantially better with frozen parameters. -------------------- ----------- ---------- ---------- Fine-tune Freeze Model PT Comp. PT Comp. AMI [Augmentation]{} no SpecAugment 28.4 26.8 59.5 Stage-1 22.6 22.4 55.8 Stage-1, Stage-2 22.5 22.6 49.2 -------------------- ----------- ---------- ---------- : Perfermance (WER(%)) investigation of two-stage data augmentation using SpecAugment on 2-stream DIRHA and AMI. \[tab:specaug\] conclusions {#sec:concl} =========== In this work, we proposed a practical two-stage training strategy to improve multi-stream end-to-end ASR. A universal feature extractor is trained in Stage-1 with all available data. In Stage-2, a set of high-level UFE features are used to train a multi-stream model without requiring highly-parameterized parallel encoders. This two-stage strategy remarkably alleviates the burden of optimizing a massive multi-encoder model while still substantially improving the ASR performance. This work shows great potential and value for this approach, but numerous directions remain for future exploration. More sophisticated data augmentation techniques beyond the single method considered here should be explored. Stage-2 training could also possibly benefit from stream-specific knowledge by exploiting more complex stream attention. Strategies for adding new streams to an existing model would also be worth investigating.	{ "pile_set_name": "ArXiv" }
-2.0truecm It has been claimed that pure cold dark matter (CDM) leads to a larger baryon fraction ($\Omega_{b}$) than predicted by big bang nucleosynthesis (BBN) if the observed hot X-ray-emitting gas represents a fair sample of the universe [@strickland97]. An admixture of hot dark matter (HDM) with CDM shifts the estimates of the baryon fraction closer to that by BBN. In addition, this mixed cold + hot dark matter model (CHDM) has been shown to agree well with the cosmic microwave background (CMB) spectrum measured by COBE, and galaxy group properties such as the number density of clusters [@primack95]. Neutrinos are the best candidate for HDM and a total neutrino mass of 5 eV, or $m_{\nu_\mu}\sim m_{\nu_\tau}\sim 2.4$ eV ($\Omega_{\nu}\simeq 0.2$) may be a solution consistent with all available observations. If HDM consists of cosmological background neutrinos (CBN) with $\sim$ eV mass [and]{} there exist cosmic neutrino beams reaching to $\sim 10^{22}$ eV, the interactions of these extremely high energy (EHE) cosmic neutrinos with the CBN during their propagation can become significant [@roulet93] due to the enhanced interaction probability at the Z boson resonance. The resulting neutrino cascade causes modifications such as a bump and a dip in the EHE neutrino spectrum at Earth [@yoshida97]. The cascade contains several hadronic decay channels that produce mostly pions which in turn reproduce neutrinos through their decay [@yoshida94], but also $\gamma$-rays and some nucleons. Hence this mechanism has been proposed [@weiler97; @fargion97] as an explanation of the highest energy cosmic rays (HECRs) whose flux above $\sim5\times10^{19}$ eV is severely attenuated by photopion production on the cosmic microwave background (CMB) [@yoshida98] in case of nucleon primaries, forming the Greisen-Zatsepin-Kuzmin (GZK) cutoff [@greisen66], The evolution of cascades initiated by EHE cosmic neutrinos is determined by very complex chains of interactions: The neutrinos undergo $\nu\nu$ reactions which involve the exchange of W and Z bosons and hadronization of their strongly interacting decay products. The produced photons, electrons and protons collide with the CMB, the infrared and optical background (IR/O) and the universal radio background (URB) [@sigl97; @protheroe96] initiating electromagnetic cascades. Electrons are also subject to synchrotron cooling in extragalactic magnetic fields (EGMF). The final particle fluxes after propagation depend on all these interactions and solving the relevant transport equations is inevitable for an accurate evaluation of the consequences of this scenario. In this letter we present numerical calculations of the “primary” EHE neutrinos and the “secondary” $\gamma$-rays and protons that may contribute a sizable fraction of the observed HECRs above $\simeq10^{19}$ eV, under different assumptions concerning neutrino mass and local density enhancement of the HDM. Several observable signatures to confirm or rule out this scenario are discussed followed by a summary. [Cascading Calculation.]{}– Our numerical calculations combine simulation codes for neutrino cascades [@yoshida97; @yoshida94], and for electromagnetic cascades and nucleon propagation [@sigl97; @lee96]. Detailed accounts of these codes are provided in Refs. [@yoshida97; @lee96]. The following processes are included: Inelastic and elastic $\nu\nu$ collisions involving an exchange of either a W or a Z boson on the CBN; the subsequent decays of produced $\pi$’s, $\mu$’s, and $\tau$’s, hadronization of quark jets, all of which eventually feed into the electromagnetic, neutrino, and nucleon channels; $\gamma\gamma\to e^+e^-$ on the CMB, the IR/O and the URB; inverse Compton scattering on the same backgrounds; triplet pair production and double pair production on the CMB; synchrotron cooling in the EGMF; the nucleon interactions on the CMB (pair production and photopion production), and neutron decay; redshifting and evolution of the black body temperature due to expansion of the universe. For the IR/O we used recent data [@fixsen97], and for the (unmeasured) URB we used the highest prediction of Ref. [@protheroe96-2] yielding conservatively low $\gamma$-ray fluxes around $10^{20}$ eV for which the URB is the most important target for pair production. We neglect interactions of EHE neutrinos with the CMB photons which are of comparable importance to those with the CBN only for neutrino energies above the Z resonance [@seckel98]. The hadronic decay of Z bosons resonantly produced by neutrinos of energy $$E_{\rm res}=M_z^2/2m_{\nu}= 4\times10^{21} \left({m_{\nu}\over 1eV}\right)^{-1}eV, \label{resonance}$$ with the CBN is the most important neutrino process for production of $\gamma$-rays and nucleons whose spectra are determined by the hadron fragmentation function. At the energy range around the Z pole, this has been measured accurately by the LEP at CERN. We implemented empirical functions using the MLLA approximation [@khoze97] in our code, which have been fitted by measurements of the inclusive production rates of $\pi^\pm$ and $p{\bar p}$ with the OPAL detector [@akers94]. This constitutes the major revision of the original codes in Refs. [@yoshida97; @yoshida94]. The dominant contribution to secondary particle fluxes from resonant Z production can be estimated analytically, for example for the produced $\gamma$-ray spectrum, $${dN_{\gamma}\over dE_{\gamma}dL}\sim \sqrt{2}\pi {\Gamma_Z\over M_Z}{1\over \lambda_Z} {dN_{\nu}\over dE_{\nu}}\vert_{E_{\nu}=E_{res}} {dn_h\over dx}\vert_{x=E_\gamma/E_{res}}, \label{rate}$$ where $\Gamma_Z\simeq0.03M_Z$ is the decay width of the Z boson whose mass is $M_Z$ , $dn_h/dx$ is the $\pi^0$ fragmentation spectrum, and $\lambda_Z\simeq$ 38 Gpc is the mean free path of neutrinos at $E_{\rm res}$ given by Eq. (\[resonance\]). The number of photons above $3\times 10^{19}/(m_\nu/\,{\rm eV})$ eV is then $0.012\,E_{\rm res}(dN_\nu/dE_\nu)(E_{\rm res}) \,{\rm Mpc}^{-1}$, compared to 0.03 from our full numerical calculation which includes contributions from all channels and uses the more accurate MLLA formula for the fragmentation function. [The particle fluxes]{}.– For a general discussion we consider a homogeneous distribution of sources radiating EHE neutrinos with a constant differential spectrum $\propto E^{-q}$ and a luminosity per comoving volume that scales as $(1+z)^m$ between $z=z_{\rm min}$ and $z=z_{\rm max}$ [@yoshida97], with $m$ characterizing source evolution. Because of the small neutrino absorption probability, the results are essentially independent of $z_{\rm min}\lesssim1$. We assume a flat universe with a Hubble constant of $H_0= 65$ km sec$^{-1}$ Mpc$^{-1}$ which is consistent with the CHDM picture of the universe [@strickland97; @primack95]. We use $q=1$, a typical spectral index expected for neutrinos produced from photopion production by accelerated protons [@stecker96; @halzen97; @waxman98]. Results are, however, rather insensitive to $q$ for $q\lesssim2$. The ratio of emitted $\nu_{\mu}$ and $\nu_e$ fluxes is assumed to be $\simeq1.86$, as expected from charged pion decay. Furthermore, for the moment we assume that the source luminosity in $\gamma$-rays and nucleons is negligible compared to the neutrino luminosity. HDM is usually expected to cluster locally and the Fermi distribution with a velocity dispersion $v$ yields an overdensity $f_\nu\lesssim v^3 m_\nu^3/(2\pi)^{3/2}/\bar{n_\nu}\simeq 330\,(v/500\,{\rm km}\,{\rm sec}^{-1})^3\,(m_\nu/{\rm eV})^3$ over the uniform CBN [@peebles]. If clustering occurs on a scale $l_\nu$ smaller than the typical attenuation length $l_{\rm att}\sim$ few Mpc of nucleons and $\gamma$-rays around $10^{20}$ eV, the ratio of their fluxes produced on that scale to the ones produced on the uniform background is $\simeq f_\nu l_\nu/l_{\rm att}$. Therefore, while clustering in the galactic halo or in a nearby galaxy cluster is unlikely to contribute to the HECR flux [@waxman98a], neutrinos clustering in the local Supercluster may have $f_\nu\sim100$, $l_\nu\sim$ few Mpc. In Fig. \[fig:fig1\] we show the calculated spectra for the following typical case: $m=3$, $z_{\rm min}=0$, $z_{\rm max}=3$, an EGMF $\lesssim10^{-12}$ Gauss, $m_{\nu_e}=m_{\nu_\mu}=m_{\nu_\tau}=1$ eV, $f_\nu\simeq300$, and $l_\nu=5$ Mpc. It can be seen clearly that the predicted fluxes are consistent with the measurement of the diffuse $\gamma$-ray flux by EGRET [@sreekumar97] and with upper limits on neutrino fluxes by Frejus [@rhode96] and Fly’s Eye [@baltrusaitis]. Typically, the energy content in the produced low energy cascade $\gamma$-rays is a few percent of the neutrino energy which agrees with a rough analytical estimate giving $\sim10/(H_0\lambda_Z)\Gamma_Z/M_Z$. By scaling the cosmologically produced low energy $\gamma$-ray flux in Fig. \[fig:fig1\] with $l_{\rm att}/(f_\nu l_\nu)$, the EGRET constraint on diffuse $\gamma$-rays requires $f_\nu\gtrsim20\,(l_\nu/5\,{\rm Mpc})^{-1}$. The EHE part of the secondary $\gamma$-rays and protons possibly constitute a hard component of the observed HECRs without a GZK cutoff. The energy content in this “visible” HECR component is about $(\Gamma_Z/M_Z)(f_\nu l_\nu/\lambda_Z) E_{\rm res}^2(dN_\nu/dE_\nu)(E_{\rm res})$, again consistent with the fluxes shown in Fig. \[fig:fig1\]. The collisions of the EHE cosmic neutrinos with the HDM can be responsible for $\sim10\%$ of the observed cosmic rays above $10^{19}$ eV with dominant contributions of $\gamma$-rays above the GZK cutoff. The fluxes deviate at most by 50% for $m_{\nu_e}\ll m_{\nu_\mu}$. Fig. \[fig:fig2\] shows the high energy part of the resultant spectra above $10^{18}$eV as in Fig. \[fig:fig1\], but for the case of the lower local enhancement of the neutrino dark matter, $f_\nu=20$ over a scale $l_\nu=5$ Mpc, the lowest possible $f_\nu$ allowed by the EGRET bound. As compared to the case for stronger clustering shown in Fig. \[fig:fig1\], the required EHE neutrino intensity is 10 times larger. In general, models based on photopion production predict an integrated photon source luminosity $L_\gamma$ that is comparable to the total neutrino luminosity, $L_\gamma\simeq\frac{13}{3}L_\nu$ [@halzen97; @mannheim95]. In this case, the EGRET constraint translates into the more stringent requirement $f_\nu\gtrsim 10^3(l_\nu/5\,{\rm Mpc})^{-1}$, as can be seen by applying the above mentioned scaling to the integrated neutrino luminosity from Fig. \[fig:fig1\]. Fig. \[fig:fig3\] shows the fluxes for this optimistic case of strong clustering. This bound on $f_\nu$ can be relaxed if most of $L_\gamma$ does not appear at EGRET energies, but is dominantly released in the TeV range. This could be a detectable signature from individual point sources [@aharonian94], in addition to the secondary $\gamma$-rays from neutrino interactions appearing at EGRET energies. Furthermore, the scenario discussed here requires sources that are optically thick for accelerated protons with respect to photopion production because otherwise the observable proton flux below the GZK cutoff would be comparable to the neutrino flux [@waxman98]. [Discussion.]{}– The EHE neutrino scenario we explored here is quite solid in terms of the particle physics because the interactions with the cosmological backgrounds occur in the well measured LEP energy range. No physics beyond the Standard Model is involved except neutrino mass. The major uncertainty arises in the question whether any astronomical objects are capable of producing neutrinos with energies of $\sim E_{\rm res}$. In the conventional models invoking the decay of photoproduced pions, primary protons must be accelerated to $\sim20E_{\rm res}\sim 10^{23}/(m_\nu/{\rm eV})$ eV in order to generate neutrinos of energy $E_{\rm res}$. Furthermore, the sources would need a dense photon target to supply high neutrino luminosity and to absorb protons and $\gamma$-rays. Thus, a new model for the neutrino beam sources may be necessary [@takahashi98]. Interestingly, the energy generation rate of $\sim E_{\rm res}$ neutrinos for the scenario shown in Fig. \[fig:fig1\], $1.8\times 10^{45}$ erg Mpc$^{-3}$ yr$^{-1}$ divided by the the rate of cosmological Gamma Ray Bursts (GRBs), $3\times 10^{-8}$ Mpc$^{-3}$ yr$^{-1}$ [@mao92], yields $\sim 6\times 10^{52}$ erg, and is comparable to the observed energy release including afterglow from a typical GRB in the BATSE range [@waxman97]. The EHE neutrino scenario has several advantages to explain the HECR observation. The observed relatively hard spectrum without GZK cutoff [@takeda] is reasonably reconstructed in our model which is determined mainly by the well-measured hadron fragmentation function at the Z pole and the energy loss process in the cosmological backgrounds, regardless of the nature of the EHE neutrino sources. The highest energy events above the GZK cutoff can originate from very distant powerful objects because neutrinos propagate without significant energy loss. For example, the AGN 3C147 at a redshift of 0.545 is a candidate for the Fly’s Eye $3\times 10^{20}$ eV event [@elbert95]. For the same reason, it is natural that we found no [nearby]{} powerful astronomical objects in directions of the possible event clusters observed by AGASA [@hayashida96]. Because the EHE neutrino beams can be responsible for a sizable fraction of cosmic rays above $10^{19}$ eV, this scenario can explain the observational fact that the intensity of the events observed above the GZK cutoff is consistent with the extrapolation of the flux from lower energies. Among the observable signatures of the neutrino scenario are the primary EHE neutrinos whose flux should be detectable, as projected sensitivities of future experiments such as the Pierre Auger [@capelle98] and NASA’s OWL [@ormes98] projects suggest. Correlation of the arrival direction of the EHE neutrino and the secondary HECR showers may also be observable. As opposed to conventional models with nucleon primaries, our model predicts that some of the observed HECRs should originate in sources at cosmological distances. At energies beyond the GZK cutoff, the correlation of arrival directions of HECR showers with sources at distances $\gg l_{\rm att}$ should be easy to detect since the background from a conventional nucleon component should be suppressed due to the GZK effect, whereas the component proposed here continues as a relatively flat spectrum. Finally, this scenario predicts a $\gamma$-ray domination above $10^{20}$ eV, and next generation experiments should settle the question whether observed HECR are consistent with $\gamma$-ray primaries [@elbert95]. In summary, we have seen that collisions of EHE cosmic neutrino beams with $\sim$ eV mass neutrino dark matter would explain the observed HECR energy spectrum, regardless of the nature of the neutrino sources if the maximum neutrino energy reaches to the Z boson pole region and the dark matter is clustered on the Supercluster scale by amounts consistent with expectations. Although EHE neutrino sources require very high efficiency of converting the energy to neutrino flux which may require a new production mechanism of neutrinos, the necessary neutrino intensity can be consistent with observed diffuse $\gamma$-ray fluxes and the GRB energy release rate. The EHE neutrino scenario is a way of producing a relatively flat component of nucleons and $\gamma$-rays that provides a significant fraction of the HECR flux above $10^{19}$ eV, dominating above the GZK cutoff, without invoking physics beyond the Standard Model except neutrino mass. Future observations of HECRs lead to indirect search for signatures of neutrino dark matter. We thank Eli Waxman and Makoto Sasaki for helpful discussions and advice and Tom Weiler and Pijush Bhattacharjee for encouragement. Paolo Coppi is acknowledged for earlier collaboration on electromagnetic cascades. At the University of Chicago this research was supported by the DoE, NSF, and NASA. Electronic Address: [email protected] R. W. Strickland and D. N. Schramm, Astrophys. J. [481]{}, 571 (1997). J. R. Primack [et al.]{}, Phys. Rev. Lett. [74]{}, 2160 (1995). E. Roulet, Phys. Rev. D [47]{}, 5247 (1993). S. Yoshida [et al.]{}, Astrophys. J. [479]{}, 547 (1997). S. Yoshida, Astropart. Phys. [2]{}, 187 (1994). T. J. Weiler, hep-ph/9710431, to be published in Astropart. Phys. D. Fargion, B. Mele, and A. Salis, astro-ph/9710029. For reviews, see S. Yoshida and H. Dai, J. Phys. G [24]{}, 905 (1998). K. Greisen, Phys. Rev. Lett. [16]{}, 748 (1966); G. T. Zatsepin and V. A. Kuzmin, JETP Lett. [4]{}, 178 (1966). G. Sigl [et al.]{}, Phys. Lett. B [392]{}, 129 (1997). R. J. Protheroe and T. Stanev, Phys. Rev. Lett. [77]{}, 3708 (1996); erratum Phys. Rev. Lett. [78]{}, 2420 (1997). S. Lee, Phys. Rev. D [58]{}, 043004 (1998). D. J. Fixsen [et al.]{}, Astrophys. J. [490]{}, 482 (1997). R. J. Protheroe and P. L. Biermann, Astropart. Phys. [6]{}, 45 (1996). D. Seckel, Phys. Rev. Lett. [80]{}, 900 (1998). V. A. Khoze and W. Ochs, Int. J. Mod. Phys. A [12]{}, 2949 (1997). R. Akers [et al]{}., Z. Phys. [C63]{}, 181 (1994). F. W. Stecker and M. H. Salamon, Space. Sc. Rev. [75]{}, 341 (1996). F. Halzen and E. Zas, Astrophys. J. [488]{}, 669 (1997). E. Waxman and J. Bahcall, hep-ph/9807282, submitted to Phys. Rev. D. see, e.g., P. J. E. Peebles, [Principles of Physical Cosmology]{}, Princeton University Press, New Jersey, 1993. E. Waxman, astro-ph/9804023, submitted to Astropart. Phys. P. Sreekumar [et al.]{}, Astrophys. J. [494]{}, 523 (1998). W. Rhode [et al.]{}, Astropart. Phys. [4]{}, 217 (1996). R. M. Baltrusaitis [et al.]{}, Phys. Rev. D [31]{}, 2192 (1985). D. J. Bird [et al.]{}, Phys. Rev. Lett. [71]{}, 3401 (1993). M. Takeda [et al.]{} Phys. Rev. Lett. [81]{}, 1163 (1998). K. S. Capelle, J. W. Cronin, G. Parente, and E. Zas, Astropart. Phys. [8]{}, 321 (1998). J. F. Ormes [et al.]{}, in Proc. 25th International Cosmic Ray Conference (Durban, 1997), eds.: M. S. Potgieter [et al.]{} K. Mannheim, Astropart. Phys. [3]{}, 295 (1995). F.A. Aharonian, P.S.Coppi, and H.J.Völk, Astrophys.J. [423]{}, L5 (1994). One possible mechanism to produce neutrinos efficiently without EHE proton radiation is the collision of protons directly accelerated by Wakefield and Snowplow plasma at initial GRB fireball with thick $\gamma$-rays. See Y. Takahashi in Proceedings of Workshop on Observing Giant Air Showers from Space, edited by J. F. Krizmanic [et al.]{} (The American Institute of Physics, 1998), p. 469. S. Mao and B. Paczyński, Astrophys. J. [388]{}, L45 (1992); E. Cohen and T. Piran, Astrophys. J. [444]{}, L25 (1995). E. Waxman and J. Bahcall, Phys. Rev. Lett. [78]{}, 2292 (1997) and references therein; M. Vietri, Phys. Rev. Lett. [80]{}, 3690 (1998). J. W. Elbert and P. Sommers, Astrophys. J. [441]{}, 151 (1995); F. Halzen [et al]{}., Astropart. Phys. [3]{}, 151 (1995). N. Hayashida [et al]{}., Phys. Rev. Lett. [77]{}, 1000 (1996).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We propose a distributed algorithm, named D-ADMM, for solving separable optimization problems in networks of interconnected nodes or agents. In a separable optimization problem, the cost function is the sum of all the agents’ private cost functions, and the constraint set is the intersection of all the agents’ private constraint sets. We require the private cost function and constraint set of a node to be known by that node only, during and before the execution of the algorithm. The application of our algorithm is illustrated with problems from signal processing and control, namely average consensus, compressed sensing, and support vector machines. It is well known that communicating in distributed environments is the most energy/time-demanding operation. Thus, algorithms using less communications are more prone to make networks live longer, e.g., sensor networks, or to execute faster, e.g., in supercomputing platforms. Through simulations for several network types and problems, we show that our algorithm requires less communications than the state-of-the-art algorithms.' author: - 'João F. C. Mota, João M. F. Xavier, Pedro M. Q. Aguiar, and Markus Püschel [^1]' bibliography: - 'paper.bib' title: 'D-ADMM: A Communication-Efficient Distributed Algorithm For Separable Optimization' --- Distributed algorithms, alternating direction method of multipliers, consensus, compressed sensing, machine learning, sensor networks. Introduction {#Sec:intro} ============ Recently, there has been a growing interest in distributed methods for data processing. Such interest was triggered both by the need of processing large quantities of data in distributed platforms, such as supercomputers, and by applications where data comes from spatially different locations and central processing is costly or impractical, as in sensor networks. On the other hand, over the last decades, convex optimization theory has proven to be an invaluable tool for deriving data processing algorithms [@Boyd:ConvexOpti]. Standard algorithms like interior-point methods are, however, unable to handle either large-scale or distributed problems. To solve these problems we thus need new algorithms. In this paper, we aim to solve separable optimization problems in a distributed, decentralized way. In a separable optimization problem, the cost function can be decomposed as a sum of $P$ functions $f_p$ and the constraint set can be decomposed as the intersection of $P$ sets $X_p$: $$\label{Eq:IntroSeparableProb} \begin{array}{ll} \underset{x}{\text{minimize}} & f_1(x) + f_2(x) + \cdots + f_P(x) \\ \text{subject to} & x \in X_1 \cap X_2 \cap \cdots \cap X_P\,, \end{array}$$ where $x \in \mathbb{R}^n$ is the global optimization variable. We will denote any solution of with $x^\star$. We associate with problem a network of $P$ nodes, where only node $p$ has access to its private cost function $f_p$ and to its private constraint set $X_p$. This situation is illustrated in Figure \[Fig:IntroFig7Nodes\]. Each node can communicate with its neighbors using the network infrastructure. A method that does not use any kind of special or central node will be called distributed algorithm. Thus, in a distributed algorithm, aggregating data is not allowed; instead, each node has to exchange information only with its neighbors in order to achieve the common goal: arriving at a solution of . We propose a novel distributed algorithm based on the alternating direction method of multipliers (ADMM) for solving . In applications such as sensor networks, nodes are battery-operated devices and thus energy is a scarce resource. It is well known that communication is the most energy-consuming operation in such a device [@Akyildiz02-WirelessNetworksASurvey]; therefore, algorithms using less communications can increase the lifespan of the network. Similarly, communication is known to be the algorithmic bottleneck in applications with more controlled environments, for example, in network protocols [@Kurose05-ComputerNetworking] or supercomputing platforms [@Graham05-FutureOfSuperComputing]. Algorithms using few communications are thus required for such applications. We use our proposed distributed algorithm to solve several instances of from several areas in engineering, namely average consensus, compressed sensing problems, and support vector machines. Our simulations show that, for each of these problems and for almost all network types, our proposed algorithm requires significantly less communication than prior state-of-the-art algorithms, including those that are specifically designed for a particular problem and are not applicable to the entire problem class . Next we formally state the problem we solve; then we give an overview of related work. [[Formal problem statement.]{}]{} Given a network with $P$ nodes, we associate $f_p$ and $X_p$ in with the $p$th node of the network, $p=1,\ldots,P$. We make the following assumptions: \[Ass:FunctionsAndSets\] Each $f_p: \mathbb{R}^n \xrightarrow{} \mathbb{R}$ is a convex function over $\mathbb{R}^n$, and each set $X_p$ is closed and convex. \[Ass:ProblemSolvable\] Problem is solvable. \[Ass:Network\] The network is connected and its topology does not vary with time. Assumption \[Ass:FunctionsAndSets\] involves the concepts of convexity of a function and convexity of a set. We assume the reader is familiar with these concepts, as well as with Lagrangian duality [@Boyd:ConvexOpti; @Bertsekas:Convex]. The second assumption states that there is at least one vector $x^\star$ that solves . This implies that the optimal cost $f^\star := \sum_{p=1}^P f_p(x^\star)$ is finite (and that $\cap_{p=1}^P X_p \neq \emptyset$). In Assumption \[Ass:Network\], a network is connected if there is a path between every pair of nodes. Under the previous assumptions, we solve the following problem: given a network, design a distributed algorithm that solves . By distributed algorithm, we mean that there is no notion of a central or special node; also, no node in the network, except node $p$, has access to $f_p$ or $X_p$ at any time before or during the algorithm. [[Related work.]{}]{} The first method addressing the generic problem was based on an average consensus algorithm coupled with a subgradient method [@Lobel10-DistributedMultiAgentOptimization; @CXVOptimInSPandComm]. Such method has been extensively studied and is known to be robust to noise and link failures; however, it generally requires too many iterations to converge. Another approach for solving in a distributed way is via augmented Lagrangian methods, which consist of two loops: an outer loop updating the dual variables, and an inner loop updating the primal variables. There are some options for the algorithms in each loop. For the outer loop, the most common is the gradient method [@Ruszczynski; @Spars09; @Jakovetic11-CooperativeConvexOptimization], which degenerates in the so-called method of multipliers [@Bertsekas97-ParallelDistributed]. For the inner loop, common choices are the nonlinear Gauss-Seidel method [@Spars09; @Jakovetic11-CooperativeConvexOptimization] and Jacobi type algorithms, such as the diagonal quadratic approximation [@Ruszczynski]. In [@Mota11ICASSP] we applied fast gradient algorithms [@BeckTeboulleFISTA] in both loops. We mention that, although [@Ruszczynski; @Spars09; @Mota11ICASSP] do not address specifically, they can be easily adapted to solve it, as explained in section \[Sec:RelatedAlgs\]. A special augmented Lagrangian-based algorithm, which uses one loop only, is the alternating direction method of multipliers (ADMM) [@Glowinski75-ADMMFirst; @Gabay76-ADMMFirst; @Bertsekas97-ParallelDistributed]. In fact, ADMM is the method of multipliers concatenated with one iteration of the nonlinear Gauss-Seidel, and therefore, it consists of just one loop. ADMM is not directly applicable to : one has to reformulate that problem first. There are several options for such reformulation. For example, [@Zhu09-DistributedInNetworkChannelCoding] proposes a distributed algorithm for solving specific instance of , based on a particular reformulation (the works [@Bazerque10-DistributedSpectrumSensing; @Kim11-CooperativeSpectrumSensingKrigedKalmanFiltering; @Forero10-ConsensusBasedDistributedSVMs; @Erseghe11-FastConsensusByADMM] apply the same algorithm to other instances of ). Another possible reformulation of is explored in [@Schizas08-ConsensusAdHocWSNsPartI; @Schizas08-ConsensusAdHocWSNsPartII], which also proposes an ADMM-based algorithm, but in contrast with [@Zhu09-DistributedInNetworkChannelCoding], it uses two communication steps per iteration, or in other words, each node uses information from its second-order neighborhood. A comparison between [@Zhu09-DistributedInNetworkChannelCoding] and [@Schizas08-ConsensusAdHocWSNsPartI] for the average consensus problem is provided in [@Erseghe11-FastConsensusByADMM]; it is reported that both methods have similar convergence rates (and thus [@Schizas08-ConsensusAdHocWSNsPartI] uses twice as much communications as [@Zhu09-DistributedInNetworkChannelCoding]), while [@Zhu09-DistributedInNetworkChannelCoding] shows more resilience to noise. The algorithm we propose in this paper is also based on ADMM, but applied to a new reformulation of . We show through extensive simulations that the proposed algorithm requires less communications than any of the previous approaches. All the above algorithms can solve in a distributed way and will be briefly described in section \[Sec:RelatedAlgs\]. There are, however, other algorithms that can solve , but are not distributed. For example, [@Boyd11-ADMM] applies ADMM to distributed scenarios, but the resulting algorithms are not distributed in the sense that they require a central node or a special network topology. Our algorithm and the ones described before, in contrast, are decentralized and can run on any connected network topology. [[Contributions.]{}]{} We propose a distributed algorithm for solving . This algorithm is an extension of our prior work [@Mota12-DistributedBP], where we solved a particular instance of , namely Basis Pursuit [@AtomicDecompBP]. Here, we first show that the algorithm in [@Mota12-DistributedBP] can be generalized to the problem class . Then we apply it to following problems: average consensus, two compressed sensing problems (different from the Basis Pursuit), and support vector machines. We also provide extensive simulations of several distributed algorithms solving . Our simulations show that the proposed algorithm outperforms, in terms of the number of used communications, any of the previous algorithms. One of the above problems, the average consensus, has been extensively studied due to its importance and simplicity. For example, many distributed algorithms, which cannot be trivially generalized to solve , have been proposed to solve it. Our algorithm, designed to solve the generic problem , can also solve the average consensus problem and, as our simulations show, it outperforms a state-of-the-art consensus algorithm [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging], for many networks of interest. [[Organization.]{}]{} The paper is organized as follows. In section \[Sec:AlgDerivation\] we derive the algorithm, by manipulating the problem and then applying ADMM. Section \[Sec:RelatedAlgs\] is dedicated to related algorithms; in particular, we describe the competing algorithms and show how we compare them with ours. In section \[Sec:Applications\] we take problems from several areas and see how they can be recast as . Additionally, we analyze the resulting algorithm for each one of these problems. Finally, in section \[Sec:Experiments\], we describe and show the results of our simulations. In Appendix \[App:ADMMProof\] we provide a proof of convergence of the particular version of ADMM we use, making this paper self-contained. Algorithm Derivation {#Sec:AlgDerivation} ==================== In this section, we propose a distributed algorithm for solving . We begin by introducing some graph theory concepts and notation. Next, we address the scenario of a bipartite network, where a proof of convergence is given and the exposition is simpler; then, we generalize the algorithm for arbitrary networks. In this case, although we do not provide a proof of convergence, the algorithm always converges in practice, as shown in section \[Sec:Experiments\]. [[Network notation.]{}]{} We represent a network as an undirected graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where $\mathcal{V} = \{1,2,\ldots,P\}$ is the set of nodes and $\mathcal{E} \subset \mathcal{V} \times \mathcal{V}$ is the set of edges. The cardinality of these sets is represented respectively by $P$, the number of nodes, and $E$, the number of edges. An edge is represented by $\{i,j\} = \{j,i\}$, and $\{i,j\} \in \mathcal{E}$ means that nodes $i$ and $j$ can exchange data with each other. We define the neighborhood $\mathcal{N}_p$ of a node $p$ as the set of nodes connected to node $p$, but excluding it; the cardinality of this set, $D_p := \|\mathcal{N}_p\|$, is the degree of node $p$. [[Node coloring.]{}]{} Given a network $\mathcal{G}$, a (node) coloring scheme is an assignment of numbers, which we call colors, to the nodes of the network such that no adjacent nodes have the same color. The colors will be represented as a set $\mathcal{C} = \{1,2,\ldots,C\}$, where $C$ is the total number of colors. For each color $c$, the set $\mathcal{C}_c \subset \mathcal{V}$ represents the nodes with color $c$, and $C_c$ is the number of nodes with that color, i.e., $\|\mathcal{C}_c\| = C_c$. A network $\mathcal{G}$ can have several coloring schemes, and $\chi(\mathcal{G})$ denotes the minimum number of colors required to color it; it is called its chromatic number. Finding a coloring scheme that uses only $\chi(\mathcal{G})$ colors is NP-hard for $\chi(\mathcal{G}) > 2$. Node-coloring is very important in wireless networks, since it is used in some medium access (MAC) protocols; consequently, many distributed node-coloring algorithms have been proposed, e.g., [@Kuhn06ComplexityDistributedGraphColoring; @Leith06DistributedLearningAlgorithms; @Duffy09ComplexityAnalysisDecentralizedGraphColoring; @Linial92LocalityDistributedGraphAlgorithms]. Henceforth, we will assume that the network $\mathcal{G}$ is given together with a coloring scheme $\mathcal{C}$. [[Problem manipulations.]{}]{} We start by assuming the network $\mathcal{G}$ is bipartite or, equivalently, that $\chi(\mathcal{G}) = 2$, i.e., it can be colored with two colors. An example of a bipartite network is a grid. We start with this assumption for two reasons: the exposition is simpler, and we prove that our algorithm converges for this case. Without loss of generality, assume the nodes are ordered such that $\mathcal{C}_1 = \{1,2,\ldots,C_1\}$ and $\mathcal{C}_2 = \{C_1+1,C_1+2,\ldots,C\}$. A common technique to decouple problem is to assign copies of the global variable $x$ to each node and then constrain all copies to be equal. Denoting the copy held by node $p$ with $x_p \in \mathbb{R}^n$, problem is written equivalently as $$\label{Eq:BipClonedProb} \begin{array}{cl} \underset{\bar{x} = (x_1,\ldots,x_P)}{\text{minimize}} & f_1(x_1) + f_2(x_2) +\cdots + f_P(x_P) \\ \text{subject to} & x_p \in X_p\,,\quad p = 1,\ldots,P \\ & x_i = x_j \,,\quad \{i,j\} \in \mathcal{E}\,, \end{array}$$ where $\bar{x}= (x_1,\ldots,x_P) \in (\mathbb{R}^n)^P$ is the optimization variable. Problem is no longer coupled by the intersection of the sets $X_p$ or by a common variable in all $f_p$’s, as , but instead by the new equations $x_i=x_j$, for all pairs of edges in the network $\{i,j\} \in \mathcal{E}$. These equations enforce all copies to be equal since the network is connected, due to Assumption \[Ass:Network\]. Note that they can be written more compactly as $(B^\top \otimes I_n) \bar{x} = 0$, where $B \in \mathbb{R}^{P\times E}$ is the node arc-incidence matrix of the graph, $I_n$ is the identity matrix in $\mathbb{R}^n$, and $\otimes$ is the Kronecker product. Each column of $B$ is associated with an edge $\{i,j\} \in \mathcal{E}$ and has $1$ and $-1$ in the $i$th and $j$th entry, respectively; the remaining entries are zeros. We can partition $B$ into two matrices $B_1$ and $B_2$, where $B_1$ contains the first $C_1$ rows and $B_2$ contains the remaining rows. Therefore, $(B^\top \otimes I_n)\bar{x} = (B_1^\top \otimes I_n)\bar{x}_1 + (B_2^\top \otimes I_n)\bar{x}_2$, where $\bar{x}_1 = (x_1,\ldots,x_{C_1}) \in (\mathbb{R}^{n})^{C_1}$ and $\bar{x}_2 = (x_{C_1 + 1},\ldots,x_{C}) \in (\mathbb{R}^{n})^{C_2}$. This enables rewriting as $$\label{Eq:BipClonedProb2} \begin{array}{ll} \underset{\bar{x}_1,\bar{x}_2}{\text{minimize}} & \sum_{p \in \mathcal{C}_1}f_p(x_p) + \sum_{p \in \mathcal{C}_2}f_p(x_p) \\ \text{subject to} & \bar{x}_1 \in \bar{X}_1\,\,,\,\, \bar{x}_2 \in \bar{X}_2\\ & (B_1^\top \otimes I_n) \bar{x_1} + (B_2^\top \otimes I_n) \bar{x_2} = 0\,, \end{array}$$ where $\bar{X}_i = \bigcap_{p \in \mathcal{C}_i} X_p$, for $i=1,2$. Problem can be solved with ADMM, explained next. [[ADMM.]{}]{} The alternating direction method of multipliers (ADMM) was proposed in the seventies by [@Glowinski75-ADMMFirst; @Gabay76-ADMMFirst]. Given two functions $g_1$ and $g_2$, two sets $X_1$ and $X_2$, and two matrices $A_1$ and $A_2$, ADMM solves $$\label{Eq:ProblemSolvedByADMM} \begin{array}{cl} \underset{x_1,x_2}{\text{minimize}} & g_1(x_1) + g_2(x_2) \\ \text{subject to} & x_1 \in X_1\,\,,\,\, x_2\in X_2 \\ & A_1x_1 + A_2x_2 = 0\,. \end{array}$$ It uses the method of multipliers concatenated with an iteration of the Gauss-Seidel algorithm [@Bertsekas97-ParallelDistributed], i.e., it iterates on $k$: $$\begin{aligned} x_1^{k+1} &\in \arg\min_{x_1 \in X_1} L_{\rho}(x_1, x_2^{k};\lambda^k) \label{Eq:ADMMAlg1} \\ x_2^{k+1} &\in \arg\min_{x_2 \in X_2} L_{\rho}(x_1^{k+1}, x_2; \lambda^k) \label{Eq:ADMMAlg2} \\ \lambda^{k+1} &= \lambda^k + \rho(A_1x_1^{k+1} + A_2x_2^{k+1})\,, \label{Eq:ADMMAlg3} \end{aligned}$$ where is the augmented Lagrangian of , $\lambda$ is the dual variable, and $\rho>0$ is a predefined parameter. More information about ADMM, including variations, applications, and related algorithms can be found, for example, in [@Bertsekas97-ParallelDistributed; @Boyd11-ADMM; @Combettes10-ProximalSplittingInSP]. The following theorem establishes its convergence. \[Teo:ConvergenceADMM\] Assume $g_i:\mathbb{R}^{n_i}\xrightarrow{} \mathbb{R}$ is a convex function over $\mathbb{R}^{n_i}$, $X_i \subset \mathbb{R}^{n_i}$ is a closed convex set, and $A_i$ is a full column-rank matrix, for $i=1,2$. Also, assume problem is solvable. Then, the sequence $\{(x_1^k,x_2^k,\lambda^k)\}$ generated by - converges to $(x_1^\star, x_2^\star, \lambda^\star)$, where 1. $(x_1^\star,x_2^\star)$ solves 2. $\lambda^\star$ solves the dual problem of : $$\label{Eq:ADMMDualProb} \underset{\lambda}{\text{minimize}}\,\,\, G_1(\lambda) + G_2(\lambda)\,,$$ where $G_i(\lambda) = \inf_{x_i \in X_i} (g_i(x_i) + \lambda^\top A_ix_i)$, $i=1,2$. This theorem is more general than the versions in [@Boyd11-ADMM] and [@Bertsekas97-ParallelDistributed Prop.4.2]. Namely, [@Boyd11-ADMM] only proves objective convergence, but since it does not assume $A_1$ and $A_2$ to be full column-rank, the sequence $\{(x_1^k,x_2^k,\lambda^k)\}$ might not even converge. On the other hand, [@Bertsekas97-ParallelDistributed] proves claims $1)$ and $2)$ only for the special case when one of the $A_i$’s is the identity. Since we could not find a proof of this version of the theorem in literature, we provide it in Appendix \[App:ADMMProof\]. The proof is, however, very similar to the ones in [@Bertsekas97-ParallelDistributed; @Boyd11-ADMM]. [[Applying ADMM.]{}]{} Clearly, has the same format as : just make the following associations $$\label{Eq:BipAssociations} g_i(x_i) = \sum_{p \in \mathcal{C}_i} f_p(x_p)\,,\quad X_i = \bar{X}_i\,,\quad A_i = B_i^\top \otimes I_n\,,$$ for $i=1,2$. The following theorem guarantees that, under Assumptions \[Ass:FunctionsAndSets\]-\[Ass:Network\], problem satisfies all the conditions of Theorem \[Teo:ConvergenceADMM\]; thus, we can use ADMM to solve . \[Teo:BipApplyADMM\] Let Assumptions \[Ass:FunctionsAndSets\]-\[Ass:Network\] hold and consider problem . Then, for $i=1,2$, the function $\sum_{p \in \mathcal{C}_i} f_p(x_p)$ is convex over $\mathbb{R}^n$, the set $\bar{X}_i$ is closed and convex, and the matrix $B_i^\top \otimes I_n$ has full column-rank. Furthermore, problem is solvable. All conclusions of Theorem \[Teo:BipApplyADMM\], except that $B_i^\top \otimes I_n$ has full column-rank, are derived straightforwardly Assumptions \[Ass:FunctionsAndSets\] and \[Ass:ProblemSolvable\] and the equivalence between and . To see how Assumption \[Ass:Network\] implies that each $B_i^\top \otimes I_n$ has full column-rank for $i=1,2$, note that it is sufficient to prove that $B_i^\top$ has full column-rank. If we prove that $B_iB_i^\top$ has full rank, then the results follows because $\text{rank}(B_iB_i^\top) = \text{rank}(B_i^\top)$. Now note that $B_iB_i^\top$ is a diagonal matrix, where in the diagonal are the degrees of the nodes belonging to the subnetwork composed by the nodes in $\mathcal{C}_i$. Assumption \[Ass:Network\] states that the network is connected and thus no node can have degree $0$; consequently, $B_iB_i^\top$ has full rank. [[Algorithm for bipartite networks.]{}]{} We now apply ADMM, i.e., - directly to . By analyzing the augmented Lagrangian , we will see that yields $C_1$ optimization problems that can be solved in parallel. Similarly, yields $C_2$ optimization problems that can be solved in parallel. In fact, developing the squared term in we obtain for the $x_1$-minimization : $$\label{Eq:BipMinx1} x_1^{k+1} \in \arg\min_{x_1 \in X_1} g_1(x_1) + \bar{\eta}_1^\top x_1 + \frac{\rho}{2} x_1^\top (A_1^\top A_1) x_1\,,$$ where $\bar{\eta}_1 :=A_1^\top \lambda^k + \rho (A_1^\top A_2)x_2^k$. Using , the quadratic term in becomes $(\rho/2)\bar{x}_1^\top(B_1B_1^\top \otimes I_n) \bar{x}_1$. As we had seen before, $B_1B_1^\top$ is a diagonal matrix with the degree of node $p$, $D_p$, in the $(pp)$th entry. This quadratic term can then be rewritten as $(\rho/2)\sum_{p \in \mathcal{C}_1}D_p\\|x_p\\|^2$. Regarding the linear term, $$\begin{aligned} \bar{\eta}_1^\top \bar{x}_1 &= \Bigl((B_1\otimes I_n)\lambda^k + \rho(B_1B_2^\top \otimes I_n)\bar{x}_2^k\Bigr)^\top \bar{x}_1 \notag \\ &= \sum_{p \in \mathcal{C}_1} \Bigl( \sum_{j \in \mathcal{N}_p}\text{sign}(j-p) \lambda_{\{p,j\}}^k - \rho\sum_{j \in \mathcal{N}_p}x_j^k \Bigr)^\top x_p\,, \label{Eq:BipGamma} \end{aligned}$$ where $\text{sign}(w) = 1$ if $w \geq 0$ and $\text{sign}(w) = -1$ otherwise. We decomposed the dual variable $\lambda$ as $(\ldots,\lambda_{\{i,j\}},\ldots)$, where $\lambda_{\{i,j\}}= \lambda_{\{j,i\}}$ is associated with the edge $\{i,j\} \in \mathcal{E}$. To see why holds, note that the $(ij)$th entry of $B_1B_2^\top$ is $-1$ if $\{i,j\} \in \mathcal{E}$ and $0$ otherwise. In conclusion, using on , we get $C_1$ optimization problems that can be solved in parallel: $$x_p^{k+1} \in \arg\min_{x_p \in X_p} f_p(x_p) + ({\gamma_p^k}^\top -\rho\sum_{j \in \mathcal{N}_p}x_j^k)^\top x_p + \frac{\rho D_p}{2}\\|x_p\\|^2\,,$$ where $\gamma_p^k := \sum_{j \in \mathcal{N}_p}\text{sign}(j-p) \lambda_{\{p,j\}}^k$, for $p \in \mathcal{C}_1$, i.e., all nodes with color $1$. A similar analysis for yields the same problem for each node in $\mathcal{C}_2$, but using the neighbors’ estimates at $k+1$. Note that the optimization problem each node solves depends on the dual variable $\lambda$ through $\gamma_p$. In fact, we can derive a formula for updating $\gamma_p$ from : write edge-wise, $\lambda_{\{i,j\}}^{(k+1)} = \lambda_{\{i,j\}}^{(k)} + \rho \,\text{sign}(j-p)(x_i^{k+1} - x_j^{k+1})$, and insert this formula in the definition of $\gamma_p^k$; the result is $ \gamma_p^{k+1} = \gamma_p^k + \rho \sum_{j \in \mathcal{N}_p}(x_p^{k+1} - x_j^k)\,. $ In sum, we arrive at the following algorithm, named D-ADMM, after Distributed-ADMM. \[Alg:ADMMBipartite\] for all $p \in \mathcal{V}$, set $\gamma_{p}^{1} = x_p^{1} = 0$ and $k=1$ \[SubAlg:ADMMBipLoop\] Set $v_p^{k} = \gamma_p^{k} -\rho \sum_{j \in \mathcal{N}_p} x_j^{k}$ and find \[SubAlg:ADMMBipOptimProb\] $$x_p^{k+1} = \begin{array}[t]{cl} \textrm{argmin} & f_p(x_p) + {v_p^{k}}^\top x_p + \frac{D_p \rho}{2}\\|x_p\\|^2\\ \textrm{s.t.} & x_p \in X_p \end{array}$$ Send $x_p^{k+1}$ to $\mathcal{N}_p$ \[SubAlg:ADMMBipCommunication\] \[SubAlg:ADMMBipEndLoop\] Repeat \[SubAlg:ADMMBipLoop\]-\[SubAlg:ADMMBipEndLoop\] for all $p \in \mathcal{C}_2$, replacing $x_j^{k}$ by $x_j^{k+1}$ in $v_p^k$ $ \gamma_p^{k+1} = \gamma_p^{k} + \rho \sum_{j \in \mathcal{N}_p} (x_p^{k+1} - x_j^{k+1}) $ \[SubAlg:ADMMBipDualVarUp\] $k \gets k+1$ In Algorithm \[Alg:ADMMBipartite\], there are two groups of nodes: $\mathcal{C}_1$ and $\mathcal{C}_2$. In each group, the nodes are not neighbors between themselves and operate at the same time. In particular, one node receives $x_j^k$ or $x_j^{k+1}$ from all its neighbors and then solves the optimization problem of step \[SubAlg:ADMMBipOptimProb\], finding its estimate $x_p^{k+1}$; it then broadcasts $x_p^{k+1}$ to its neighbors. After having received all the estimates from its neighbors, node $p$ can update $\gamma_p$ as in step \[SubAlg:ADMMBipDualVarUp\]. Algorithm \[Alg:ADMMBipartite\] only applies to bipartite networks and its convergence is guaranteed by Theorems \[Teo:ConvergenceADMM\] and \[Teo:BipApplyADMM\]. Next, we generalize it for any type of network. [[Algorithm for arbitrary networks.]{}]{} So far we assumed the network was bipartite, i.e., it could be colored with just two colors: $\chi(\mathcal{G})=2$. Now, suppose the network is colored with $C\geq 2$ colors. We derived problem from problem by partitioning the node-arc incidence matrix $B$ into $\begin{bmatrix}B_1^\top & B_2^\top\end{bmatrix}^\top$, where $B_1$ and $B_2$ contain the rows corresponding to nodes in $\mathcal{C}_1$ and $\mathcal{C}_2$, respectively. For convenience, we had assumed the nodes were ordered such that $\mathcal{C}_1 = \{1,2,\ldots,C_1\}$ and $\mathcal{C}_2 = \{C_1+1,C_1+2,\ldots,C\}$. Without loss of generality, assume a similar ordering for the nodes in the current case, i.e., $\mathcal{C}_1 = \{1,2,\ldots,C_1\}$, $\mathcal{C}_2 = \{C_1+1,\ldots,C_1+C_2\}$, …, $\mathcal{C}_C = \{\sum_{c=1}^{C-1}C_c+1,\ldots,C\}$. This induces a natural partition of $B$ as $\begin{bmatrix}B_1^\top & B_2^\top & \cdots & B_C^\top\end{bmatrix}^\top$. The counterpart of for this case is then $$\label{Eq:GenNetCloned} \begin{array}{ll} \underset{\bar{x}_1,\ldots,\bar{x}_C}{\text{minimize}} & \sum_{c=1}^C \sum_{p \in \mathcal{C}_p} f_p(x_p) \\ \text{subject to} & \bar{x}_c \in \bar{X}_c\,,\quad c = 1,\ldots,C \\ & \sum_{c=1}^C (B_c^\top \otimes I_n)\bar{x}_c= 0\,, \end{array}$$ where the variable is $(\bar{x}_1,\ldots,\bar{x}_C) \in (\mathbb{R}^n)^P$. To solve , we cannot apply ADMM - directly, but its generalization to solve $$\begin{array}{cl} \underset{x_1,\ldots,x_C}{\text{minimize}} & g_1(x_1) + \cdots + g_C(x_C) \\ \text{subject to} & x_c \in X_c\,,\quad c=1,\ldots,C \\ & A_1x_1 + \cdots + A_Cx_C = 0\,. \end{array}$$ Such a generalization is straightforward, but there are no convergence guarantees in the sense that there is no known proof of Theorem \[Teo:ConvergenceADMM\] for the resulting algorithm. Yet, experimental results, as the ones presented here, suggest that Theorem \[Teo:ConvergenceADMM\] might still hold. The following algorithm is a straightforward generalization of Algorithm \[Alg:ADMMBipartite\] to solve . \[Alg:DADMMGeneral\] for all $p \in \mathcal{V}$, set $\gamma_{p}^{1} = x_p^{1} = 0$ and $k=1$ $$v_p^{k} = \gamma_p^{k}- \rho \sum_{\begin{subarray}{c} j \in \mathcal{N}_p \\ j < p \end{subarray} }x_j^{k+1} - \rho \sum_{\begin{subarray}{c} j \in \mathcal{N}_p \\ j > p \end{subarray} }x_j^{k}$$ \[SubAlg:GeneralDADMMv\] and find $$x_p^{k+1} = \begin{array}[t]{cl} \textrm{argmin} & f_p(x_p) + {v_p^{k}}^\top x_p + \frac{D_p \rho}{2}\\|x_p\\|^2\\ \textrm{s.t.} & x_p \in X_p \end{array}$$ \[SubAlg:ADMMGenProb\] Send $x_p^{k+1}$ to $\mathcal{N}_p$ \[SubAlg:ADMMGenComm\] $ \gamma_p^{k+1} = \gamma_p^{k} + \rho \sum_{j \in \mathcal{N}_p} (x_p^{k+1} - x_j^{k+1}) $ \[SubAlg:ADMMGenDualVarUp\] $k \gets k+1$ In Algorithm \[Alg:DADMMGeneral\] there are $C$ groups of nodes, arranged by colors. As in the bipartite case, in each group, the nodes are not neighbors between themselves and operate at the same time. In practice, they cannot operate at the same time because there is no central coordination. As seen in the expression for $v_p^k$ (step \[SubAlg:GeneralDADMMv\]), one node, knowing the color of its neighbors, can operate only after having received the estimates $x_k^{k+1}$ from the nodes with lower color. Excepting that, the algorithm is very similar to Algorithm \[Alg:DADMMGeneral\]. As said before, although there is no convergence guarantee for Algorithm \[Alg:DADMMGeneral\], it converges in practice for several kinds of problems, and we have never seen a nonconvergent example. We point out that, using a reasoning similar to that of the discussion after Theorem \[Teo:BipApplyADMM\], each matrix $B_i^\top$ can be proved to be full column-rank. Related Algorithms {#Sec:RelatedAlgs} ================== We now review some algorithms for solving . We start with describing the performance measure we will use and its practical relevance. [[Performance measure: communication steps.]{}]{} We divide the algorithms that solve in two kinds: single-looped and double-looped. The single-looped algorithms have just one (nested) loop, and in each iteration, each node receives the neighbors’ estimates, computes its on estimate, and broadcasts it to its neighbors. The double-looped algorithms have two nested loops and they are based on augmented Lagrangian methods (see ). The outer loop serves to update the dual variable of the augmented Lagrangian; the inner loop serves for each node to receive the neighbors’ estimates, compute its own estimate, and broadcast it to the neighbors, as one iteration of a single-looped algorithm. The natural way of comparing all algorithms is then through a communication step, which will denote one iteration (for the single-looped algorithms) or one inner loop iteration (for the double-looped algorithms). We say that an algorithm is more efficient than another one if it uses less communication steps for the same solution accuracy. Communication steps is a fair measure because it compares, not only similar computational operations at each node (as we will soon see), but also the amount of communications each node requires to achieve a solution with a given precision. Furthermore, one of the main concerns in designing algorithms and protocols for wireless networks is reducing the number of communications; the reason is because communicating is the slowest and most energy-consuming task [@Akyildiz02-WirelessNetworksASurvey; @Krishnamachari05-NetworkingWirelessSensors]. The same concern is shared in supercomputing environments [@Graham05-FutureOfSuperComputing]. In contrast with execution time, communication steps are application and implementation independent. For example, given a distributed algorithm, we get different execution times if we change any of the following: medium-access protocol, transmission medium, scheduling, platform for the nodes, or even the way the algorithm is coded. Communication steps are, on the other hand, invariant to any of these. Note that, given a network with $E$ edges, if we multiply $2E$ by the number of communication steps, we obtain the total number of communications in the network. We point out, however, that our algorithm, D-ADMM, is asynchronous, i.e., not all nodes perform the same tasks at the same time, in contrast with most of the algorithms presented next. More concretely, D-ADMM assumes a coloring for the nodes, and each node operates according to its color. Other algorithms do not rely on any coloring scheme, enabling them to be synchronous, i.e., all nodes can perform the same tasks at the same time.[^2] Therefore, if one communication step in D-ADMM takes $T$ time units, it would take $T/C$ units in a synchronous algorithm, where $C$ is the number of colors, in case the nodes were able to transmit messages at the same time. [[Gradient/Subgradient method.]{}]{} The classical approach for solving is to use a subgradient method combined with an averaging consensus algorithm. This approach was taken in [@Lobel10-DistributedMultiAgentOptimization], and consists of iterating in $k$ $$\label{Eq:OASubgradAlg} x_p^{k+1} = \Bigl[a_{pp}^{k} x_p^{k} + \sum_{j \in \mathcal{N}_p} a_{pj}^{k} x_j^{k} - \alpha^{k} g_p^{k}\Bigr]_{X_p}^+\,,$$ for each node $p$. In , $[s]_S$ represents the projection of the point $s$ onto the (convex) set $S$, $x_p^k$ the estimate of node $p$ at iteration $k$, $g_p^k$ the gradient (or a subgradient, in case the objective is not differentiable) of $f_p$ at the point $a_{pp}^{k} x_p^{k} + \sum_{j \in \mathcal{N}_p} a_{pj}^{k} x_j^{k}$, and $a_{ij}^k$ represents the influence node $j$ exerts on node $i$ at iteration $k$. There are some conditions the weights $a_{ij}^k$ have to satisfy [@Lobel10-DistributedMultiAgentOptimization], for example, for each $k$, the matrix with entries $a_{ij}^k$ should be doubly stochastic. In each iteration, node $p$ receives the estimates $x_{j}^k$ from its neighbors, computes , and then broadcasts $x_p^{k+1}$ to its neighbors. It is thus a (synchronous) single-looped algorithm. Although this algorithm is provably robust to noise, it is known to perform very slowly, requiring lots of iterations to converge. [[ADMM-based algorithms.]{}]{} We now describe two algorithms that are based on ADMM, like ours. Both of them are based on different reformulations of , yielding two different algorithms. The first algorithm is [@Zhu09-DistributedInNetworkChannelCoding] (it also appears in [@Bazerque10-DistributedSpectrumSensing; @Kim11-CooperativeSpectrumSensingKrigedKalmanFiltering; @Forero10-ConsensusBasedDistributedSVMs; @Erseghe11-FastConsensusByADMM]), described in Algorithm \[Alg:Giannakis\]. \[Alg:Giannakis\] for all $p \in \mathcal{V}$, set $\gamma_{p}^{1} = x_p^{1} = 0$ and $k=1$ set $ v_p^{k} = \gamma_p^{k}- \rho \sum_{j \in \mathcal{N}_p \cup \{p\}} x_j^{k} $ and find $$x_p^{k+1} = \begin{array}[t]{cl} \textrm{argmin} & f_p(x_p) + {v_p^{k}}^\top x_p + \rho D_p\\|x_p\\|^2\\ \textrm{s.t.} & x_p \in X_p \end{array}$$ \[SubAlg:Giannakis-Optm\] Send $x_p^{k+1}$ to $\mathcal{N}_p$, and receive $x_j^{k+1}$, $j \in \mathcal{N}_p$ $ \gamma_p^{k+1} = \gamma_p^{k} + \rho \sum_{j \in \mathcal{N}_p} (x_p^{k+1} - x_j^{k+1}) $ \[SubAlg:ADMMGenDualVarUp\] $k \gets k+1$ Although derived from a different reformulation of , Algorithm \[Alg:Giannakis\] algorithm is very similar to D-ADMM. The three main differences are: 1. Algorithm \[Alg:Giannakis\] is synchronous, i.e., all the nodes can perform the same tasks at the same time; 2. while in Algorithm \[Alg:Giannakis\] node $p$ uses $x_p^k$ to construct the vector $v_p^k$ (see step \[SubAlg:Giannakis-Optm\]), D-ADMM does not; 3. Algorithm \[Alg:Giannakis\] is proven to converge for any connected network, and D-ADMM is only proven to converge in bipartite networks (although in practice it converges in any connected network). We will see that D-ADMM requires almost always less communication steps to achieve a given accuracy than Algorithm \[Alg:Giannakis\], for several instances of . Another ADMM-based algorithm is [@Schizas08-ConsensusAdHocWSNsPartI] (also appearing in [@Schizas08-ConsensusAdHocWSNsPartII; @Erseghe11-FastConsensusByADMM]). That algorithm, in contrast with D-ADMM and Algorithm \[Alg:Giannakis\], uses two communication steps per iteration. Since the number of iterations it takes to achieve a given accuracy in the average consensus problem is similar to Algorithm \[Alg:Giannakis\] [@Erseghe11-FastConsensusByADMM], it takes the double of the communication steps. [[Double-looped algorithms.]{}]{} Several double-looped algorithms were proposed to solve particular optimization problems, which as D-ADMM, can be generalized to solve . In particular, they can solve , by solving the following dual problem of : $$\label{Eq:DualProblemCloned} \underset{\lambda}{\text{maximize}} \,\,\, L_{\rho}(\lambda)\,,$$ where $L_\rho(\lambda)$ is the (augmented) dual function $$\label{Eq:LagrangianProblemCloned} L_\rho(\lambda) = \begin{array}[t]{cl} \underset{\bar{x}}{\inf} & \sum_{p=1}^P f_p(x_p) + \sum_{\{i,j\}\in \mathcal{E}} \phi_{\lambda_{\{i,j\}}}^\rho (x_i - x_j) \\ \text{s.t.} & x_p \in X_p\,,\quad p=1,\ldots,P\,, \end{array}$$ and $\phi_{v}^\rho(\eta) := v^\top \eta + \frac{\rho}{2}\\|\eta\\|^2$. It can be proved that $L_\rho(\lambda)$ is differentiable, and thus can be solved with a gradient method, or even with a fast (Nesterov) gradient method. The outer loop consists of applying such algorithm: it updates the dual variables $\lambda_{\{i,j\}}$. However, to compute the gradient of $L_\rho(\lambda)$, one needs to solve the optimization problem in ; and this requires another iterative algorithm, consisting of the second nested loop. Several approaches have been used to solve that problem, e.g., nonlinear Gauss-Seidel methods [@Spars09; @Jakovetic11-CooperativeConvexOptimization], Jacobi-type methods [@Ruszczynski], and Nesterov methods [@Mota11ICASSP]. Of all the algorithms described in this section, only gradient/subgradient method does not solve an optimization problem at each iteration. All the remaining algorithms solve the same problem as D-ADMM (step \[SubAlg:ADMMGenProb\] of Algorithm \[Alg:DADMMGeneral\]), but possibly with different parameters. As illustrated in section \[Sec:Experiments\], in practice, our D-ADMM performs generally better than any of algorithms described in this subsection. The only true competitor is Algorithm \[Alg:Giannakis\]. Applications {#Sec:Applications} ============ We now take some optimization problems from signal processing and see how they can be recast as . We also review some of the conventional methods to solve them in a distributed way. Then, in section \[Sec:Experiments\], we present simulation results of D-ADMM and other algorithms applied to these problems. [[Consensus.]{}]{} Consensus is one of the most fundamental problems in networks. Given a network with $P$ nodes, node $p$ generates a number or measurement, say $\theta_p$, and the goal is to compute the average $\theta^\star = (1/P)\sum_{p=1}^P \theta_p$ in every node. The classical approach to consensus [@DeGroot74ReachingConsensus; @Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging] is $$\label{Eq:ApplicationConsensusLinearSys} x^{k+1} = A x^k\,,\qquad x^0 = \theta\,,$$ where the $p$th entry of $x^k = (x_1^k,\ldots,x_P^k) \in \mathbb{R}^P$ contains the estimate of node $p$ at time $k$ and $\theta = (\theta_1,\ldots,\theta_P)\in \mathbb{R}^P$ is the vector collecting the measurements. The matrix $A$ reflects the topology of the network by having $a_{ij} = 0$ for $\{i,j\}\not\in \mathcal{E}$. Thus, is equivalent to $x_p^{k+1} = \sum_{j \in \mathcal{N}_p \cup \{p\}}a_{pj} x_j^k$, i.e., the estimate of node $p$ at $k+1$ will be a weighted average of its estimate and the neighbors’ estimates at $k$. There are some constraints on choosing the weights, for example, $\sum_{j \in \mathcal{N}_p \cup \{p\}}a_{pj} = 1$. See [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging] for a recent survey on consensus algorithms using the approach . Another approach to consensus, recently adopted in [@Erseghe11-FastConsensusByADMM], is to solve the optimization problem $$\label{Eq:ApplicationConsensusOptimProb} \underset{x}{\text{minimize}} \,\,\, \sum_{p=1}^P (x - \theta_p)^2$$ in a distributed way. The solution of is $\theta^\star = (1/P)\sum_{p=1}^P\theta_p$. Clearly, has the format of by assigning $f_p(x) = (x-\theta_p)^2$, $X_p = \mathbb{R}$; thus, it can be solved by D-ADMM or any of the methods described in the last section. The work [@Erseghe11-FastConsensusByADMM] applies Algorithm \[Alg:Giannakis\] for solving . In all algorithms, node $p$ has to solve at each iteration $$\underset{x}{\text{minimize}} \,\,\, (x - \theta_p)^2 + v^\top x + c x^2\,,$$ which has the closed-form solution $x^\star = (2\theta_p - v)/(2(1+c))$. [[Compressed sensing problems.]{}]{} Compressed sensing is a new paradigm for the acquisition of signals. For an introductory survey see, for example, [@IntroCompressiveSamplingCandesWakin]. In compressed sensing, a signal is compressed at the same time as it is acquired, in contrast to the classical acquire-then-compress paradigm. Although compressed sensing makes the acquisition process simpler, the reconstruction of a compressed signal is more complicated, since an optimization problem has to be solved. The goal of this optimization problem is to find the sparsest solution of a linear system $Ax = b$, where $A \in \mathbb{R}^{m\times n}$ is a matrix related with the acquisition process and $b \in \mathbb{R}^m$ is the acquired/compressed signal. If $A$ satisfies some technical conditions [@IntroCompressiveSamplingCandesWakin], then the solution to the Basis Pursuit (BP) [@AtomicDecompBP], $$\label{Eq:BP} \begin{array}{ll} \underset{x}{\text{minimize}} & \\|x\\|_1 \\ \text{subject to} & Ax = b\,, \end{array}$$ recovers the sparsest solution of $Ax = b$. In , $\\|x\\|_1 = \|x_1\| + \cdots + \|x_n\|$ is the $\ell_1$-norm of $x$. A variation of is the LASSO [@IntroCompressiveSamplingCandesWakin]: $$\label{Eq:Lasso} \begin{array}{ll} \underset{x}{\text{minimize}} & \\|x\\|_1 \\ \text{subject to} & \\|Ax - b\\| \leq \sigma\,, \end{array}$$ where $\sigma > 0$ is a fixed parameter and $\\|\cdot\\|$ the $\ell_2$-norm. Problem is more appropriate when there is noise in the acquisition process and we have an upper bound $\sigma$ on the norm of the noise. Closely related with , is $$\label{Eq:BPDN} \underset{x}{\text{minimize}} \,\,\, \\|Ax - b\\|^2 + \beta\\|x\\|_1\,,$$ called Basis Pursuit Denoising (BPDN) [@AtomicDecompBP]. The parameter $\beta > 0$ regulates the tradeoff between the sparsity of the signal and the weight of the noise. Here, we are interested in solving distributed versions of -. Namely, we assume the matrix $A$ is partitioned into $P$ blocks, either by rows or by columns, as depicted in Figure \[Fig:PartitionOfA\]. We assign the $p$th block to the $p$th node of a network with $P$ nodes. For applications of this problem see, for example, our previous work [@Mota12-DistributedBP]. In [@Mota12-DistributedBP], we solved BP with D-ADMM for both the row and the column partition. In this paper, we solve for the column partition, and for the row partition. The reverse cases, i.e., for the row partition or for the column partition cannot be trivially recast as . [[LASSO: column partition.]{}]{} Assume the matrix $A$ is partitioned by columns, $A = \begin{bmatrix}A_1 & \cdots & A_P\end{bmatrix}$, where the $p$th block of $A$ is only known at node $p$. Assume vector $b$, parameter $\sigma$, and the number of nodes $P$ are available at all nodes. This problem cannot be directly recast as ; we will have to do it through duality. However, if we solve the “simple dual” of , we will not be able to recover solution to the primal problem afterwards. The reason is that the objective of is not strictly convex. We thus start by regularizing by adding a small term to its objective, making it strictly convex: $$\label{Eq:LassoRegularized} \begin{array}{ll} \underset{x}{\text{minimize}} & \\|x\\|_1 + \frac{\delta}{2} \\|x\\|^2 \\ \text{subject to} & \\|Ax - b\\| \leq \sigma\,, \end{array}$$ where $\delta >0$ is sufficiently small. This regularization is inspired by [@Friedlander07ExactRegularizationConvexPrograms], which establishes some conditions for regularizations of this type to be exact. By exact, we mean that there exists $\bar{\delta} > 0$ such that the solution of is always a solution of , for $\delta \leq \bar{\delta}$. One of the conditions for exact regularization is that the objective is linear and the constraint set is the intersection of a linear system with a closed polyhedral cone. LASSO can be recast as $$\label{Eq:LassoManipulated} \begin{array}{ll} \underset{x,t,u,v}{\text{minimize}} & 1_n^\top t \\ \text{subject to} & \\|u\\| \leq v \\ & u = Ax - b\,,\,\,\, v = \sigma \\ & x \leq t\,,\,\,\, -x\leq t\,, \end{array}$$ where $1_n \in \mathbb{R}^n$ is the vector of ones, and $(x,t,u,v)$ the variable. Although the objective of is linear and its constraint set is the intersection of a linear system with a closed convex cone $\mathcal{K} := \{(u,v)\,:\, \\|u\\|\leq v\}$, there is not a proof of exact regularization for , because $\mathcal{K}$ is not polyhedral. However, experimental results in [@Friedlander07ExactRegularizationConvexPrograms] suggest that exact regularization might occur for non-polyhedral cones. We will also see in the results of our simulations that approximating with can still yield very accurate solutions. Before computing the dual of , we first introduce a new variable $y \in \mathbb{R}^{m}$ and make the column partition explicit: $$\label{Eq:LassoWithY} \begin{array}{ll} \underset{x,y}{\text{minimize}} & \sum_{p=1}^P (\\|x_p\\|_1 + \frac{\delta}{2}\\|x_p\\|^2) \\ \text{subject to} & \\|y\\| \leq \sigma \\ & y = \sum_{p=1}^P A_p x_p - b\,. \end{array}$$ The dual problem of , by dualizing the last constraint only, can be shown to be equivalent to $$\underset{\lambda}{\text{minimize}} \,\,\, \sum_{p=1}^P g_p(\lambda)\,,$$ where $g_p(\lambda) := \frac{1}{P}(b^\top \lambda + \sigma\\|\lambda\\|) - \inf_{x_p}(\\|x_p\\|_1 + (A_p^\top \lambda)^\top x_p + \frac{\delta}{2}\\|x_p\\|^2)$ is the function associated to node $p$. Each function $g_p$ is convex because it is the pointwise supremum of convex functions [@Bertsekas:Convex Prop.1.2.4]. The problem each node has to solve in each iteration is addressed in Appendix \[App:ProbLasso\]. We point out that problem can be ill-conditioned due to the small value that has to be used for $\delta$, if we want to get a good approximation of the original problem. In particular, the problem solved at each node can be very time-consuming, since it may require lots of internal iterations. We used $\delta = 10^{-3}$ in our simulations. While such value for $\delta$ required a considerable amount of iterations for solving the internal problem of each node (we set $500$ as the number of maximum iterations), it also allowed to get solutions of the original problem with relative errors of $0.1\%$, using few communications. BPDN, addressed next, handles the same type of problems as LASSO, but its distributed implementation does not require solving ill-conditioned problems. [[BPDN: row partition.]{}]{} In BPDN , we assume that both the matrix $A$ and the vector $b$ are partitioned by rows: $A = \begin{bmatrix}A_1^\top & \cdots & A_P^\top\end{bmatrix}^\top$, $b = \begin{bmatrix}b_1^\top & \cdots & b_P^\top\end{bmatrix}^\top$. Therefore, can be readily rewritten as $$\label{Eq:BPDN2} \underset{x}{\text{minimize}} \,\,\, \sum_{p=1}^P \Bigl( \\|A_px - b_p\\|^2 + \frac{\beta}{P}\\|x\\|_1\Bigr)\,.$$ The identification between and is immediate: set $f_p(x) := \\|A_px - b_p\\|^2 + \frac{\beta}{P}\\|x\\|_1$. The problem that has to be solved at node $p$, $$\label{Eq:BPDNAtEachNode} \underset{x}{\text{minimize}}\,\,\, \\|A_px - b_p\\|^2 + \frac{\beta}{P}\\|x\\|_1 + v^\top x + c\\|x\\|^2\,,$$ can be solved, for example, with the GPSR solver [@GPSR]. [[Distributed support vector machines.]{}]{} In a Support Vector Machine (SVM) [@Bishop06-PatternRecognitionMachineLearning Ch.7], we are given data in a matrix $A \in \mathbb{R}^{m \times n}$, where each row of $A$ represents a data point in $\mathbb{R}^n$. Each data point can be classified as belonging to group $\mathcal{X}$ or to group $\mathcal{Y}$. If data point $a_i$ belongs to $\mathcal{X}$ (resp. $\mathcal{Y}$), we set the $i$th entry of a vector $d \in \mathbb{R}^m$ to $+1$ (resp. $-1$). The goal is to find the parameters $(s,r) \in \mathbb{R}^n\times \mathbb{R}$ of a linear separator $z(a) = s^\top a + r$ such that $z(a) > 0$ if $a \in \mathcal{X}$, and $z(a) < 0$ if $a \in \mathcal{Y}$. Once the pair $(s,r)$ is found, we can attempt to determine if a new point $a$ belongs to $\mathcal{X}$ or $\mathcal{Y}$ just by analyzing the sign of $z(a)$. There are several ways of finding $(s,r)$; one of them is solving the quadratic program $$\label{Eq:SVMProb} \begin{array}{ll} \underset{s,r}{\text{minimize}} & \\|s\\|^2 \\ \text{subject to} & D(As - 1_m r) \geq 1_m\,, \end{array}$$ where $D = \text{Diag}(d_1,\ldots,d_m)$, and $1_m$ is the vector of ones in $\mathbb{R}^m$. Many times, the data is distributed among several entities, and processing it in a distributed way is prohibitive for complexity or privacy reasons. Here, we assume the model of [@Mangasarian07-PrivacyPreservingClassificationHP; @Forero10-ConsensusBasedDistributedSVMs]: there are $P$ entities, and each entity $p$ has $m_p$ data points, or equivalently, a block $A_p \in \mathbb{R}^{m_p \times n}$ of rows of $A$, and their corresponding classification, i.e., the $p$th diagonal block $D_p\in \mathbb{R}^{m_p \times m_p}$ of $D$, such that $m_1 + \cdots + m_P = m$. This enables us to rewrite as $$\label{Eq:SVMProb2} \begin{array}{ll} \underset{s,r}{\text{minimize}} & \sum_{p=1}^P\\|s\\|^2 \\ \text{subject to} & D_p(A_ps - 1_{m_p}r) \geq 1_{m_p}\,,\quad p = 1,\ldots,P\,. \end{array}$$ Problem has the format of : just set $f_p(s,r) = \\|s\\|^2$ and $X_p = \{(s,r) \in \mathbb{R}^n\times \mathbb{R}\,:\, D_p(A_ps - 1_{m_p}r) \geq 1_{m_p}\}$. The problem each node has to solve at each iteration is a quadratic program, which we solve with Matlab’s `quadprog` function. Algorithm \[Alg:Giannakis\] was applied in [@Forero10-ConsensusBasedDistributedSVMs] to solve a version of that takes into account possible violations of the linear separability of the data. Here, for simplicity, we just consider and assume the data can be separated linearly. Experimental Results {#Sec:Experiments} ==================== In this section, we show some simulation results of D-ADMM, and the algorithms from section \[Sec:RelatedAlgs\], applied to the problems just described. Of these applications, consensus is the one that assesses the algorithms the best, since neither the problem solution, nor the problem each node solves at each iteration, requires iterative algorithms or any kind of solver: they have closed-form solutions. We will, therefore, give more emphasis to consensus than to the other applications. We start with describing the network models. Name Parameters Description ------------------------------------------------------- -- ------------ -- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Erdős-Rényi [@Erdos59RandomGraphs] $p$ Every pair of nodes $\{i,j\}$ is connected or not with probability $p$ Watts-Strogatz [@Watts98CollectiveDynamicsSmallWorld] $(n,p)$ First, it creates a lattice where every node is connected to $n$ nodes; then, it rewires every link with probability $p$. If link $\{i,j\}$ is to be rewired, it removes the link, and connects node $i$ or node $j$ (chosen with equal probability) to another node in the network, chosen uniformly. Barabasi-Albert [@Barabasi99EmergenceOfScaling] —— It starts with one node. At each step, one node is added to the network by connecting it to $2$ existing nodes: the probability to connect it to node $p$ is proportional to $D_p$. Geometric [@Penrose] $d$ It drops $P$ points, corresponding to the nodes of the network, randomly in a $[0,1]^2$ square; then, it connects nodes whose (Euclidean) distance is less than $d$. Lattice —— Creates a lattice of dimensions $m\times n$; $m$ and $n$ are chosen to make the lattice as square as possible. [@clc>r>rrrrrrr@]{} Number & Model & Parameters &\ & & &\ (r)[4-11]{} & & & 10 & 50 & 100 & 200 & 500 & 700 & 1000 & 2000\ 1 & Erdős-Rényi & $0.25$ & & & & & & & &\ 2 & Erdős-Rényi & $0.75$ & & & & & & & &\ 3 & Watts-Strogatz & $(2,0.8)$ & & & & & & & &\ 4 & Watts-Strogatz & $(4,0.6)$ & & & & & & & &\ 5 & Barabasi-Albert & —— & & & & & & & &\ 6 & Geometric & $0.2$& & & & & & & &\ 7 & Lattice & ——& & & & & & & &\ [[Network models.]{}]{} We generated networks according to the models described in Table \[Tab:NetworkModels\]. All networks were generated using the igraph package in R [@igraph; @R]. Table \[Tab:NetworkParametersADNC\] shows the $56$ different networks we generated: given a model and fixed its parameters, we generated $8$ networks with different number of nodes, ranging from $10$ to $2000$; we used $7$ different combinations of models and parameters. All networks, except Lattice networks, were generated randomly, and we had to guarantee that all networks were connected (cf. Assumption \[Ass:Network\]). So, every time we would get a non-connected network, we would generate another network with slightly different parameters, changed in the direction to make the next network connected with greater probability. Therefore, there were some exceptions to the parameters described in Table \[Tab:NetworkParametersADNC\]. For example, network number $1$ with $10$ nodes was generated with $p=0.27$, instead of with $p=0.25$, or network number $6$ with $10$ nodes was generated with $d=0.36$, instead of with $d=0.2$. Table \[Tab:NetworkParametersADNC\] also shows the average node degree and the number of colors used for each network. The average node degree can be as low as $3$, for example in the Lattice network with $50$ nodes, or as high as $1499$, in the Erdős-Rényi network with $2000$ nodes. The number of colors each network was colored with is also varied. Note that only the Lattice networks were colored with $2$ colors, and thus these are the only networks for which D-ADMM is proven to converge. We use all networks of Table \[Tab:NetworkParametersADNC\] for the consensus problem, since we study that problem in more detail. To illustrate the algorithms for the remaining applications, we will only use the networks with $10$ and $50$ nodes. [[Choosing .]{}]{} In section \[Sec:RelatedAlgs\] we reviewed the distributed algorithms in literature that can solve . All these algorithms, including D-ADMM, but excluding the gradient/subgradient method, are based on augmented Lagrangian duality. This means that, for each algorithm, we have to choose the parameter $\rho$ of the augmented Lagrangian (see, e.g., ). We are unaware of any scheme that chooses this parameter in an optimal way before the execution of the algorithm. There are, however, schemes for updating it during the execution of the algorithm [@Boyd11-ADMM; @He00ADMWithSlefAdaptativePenaltyForVIs], but it is not straightforward to implement them in a distributed scenario. We will thus adopt the following scheme for choosing $\rho$: given a fixed network, we execute several times all the algorithms that depend on $\rho$, each for a different $\rho$, chosen from the set $\{10^{-4},10^{-3},10^{-2},10^{-1},10^{0},10^{1},10^{2}\}$; then, we pick the $\rho$ that leads to the best result, i.e., the least number of communication steps. This scheme was adopted in all experiments. [[Consensus.]{}]{} We now present simulation results for the consensus problem. The data was generated the following way: node $p$ generates, independently of the other nodes, one number $\theta_p$ from a Gaussian distribution with mean $10$ and standard deviation $100$. We chose such a large standard deviation to make sure that the nodes’ numbers differed significantly. Figure \[Fig:ConsAllAlgs50nodes\] shows the results of the simulations of D-ADMM, the algorithms described in section \[Sec:RelatedAlgs\], and the additional algorithm from [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging], which uses the approach . Note that [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging] only solves the consensus problem and cannot be trivially generalized to solve . Figure \[Fig:ConsAllAlgs50nodes\] depicts the number of communication steps, which is equal to the number of iterations for the single-looped algorithms and to the number of inner iterations for the double-looped algorithms, for all the networks with $50$ nodes. All algorithms stopped after reaching a relative error of $10^{-2} \%$, i.e., $\\|x^k - 1_P\theta^\star\\|/(\sqrt{P}\|\theta^\star\|) \leq 10^{-4}$, or after reaching the maximum number of $1000$ communication steps. The optimal solution $\theta^\star$ had been computed before. It can been seen from the plot of Figure \[Fig:ConsAllAlgs50nodes\] that D-ADMM is the algorithm requiring the least communication steps to converge, i.e., to achieve a precision of $10^{-2}\%$, for all networks except the first two, where [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging] was better. Figure \[Fig:ConsAllAlgs50nodes\] shows the error evolution along the communication steps for network $3$ (cf. with point $3$ of Figure \[Fig:ConsAllAlgs50nodes\]). It can be observed that D-ADMM required always the least number of communication steps to reach any accuracy between $10\%$ and $10^{-2}\%$. We only consider the three best algorithms (D-ADMM, [@Zhu09-DistributedInNetworkChannelCoding], and [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging]) for a more thorough analysis, in Figure \[Fig:ConsSize\]; here, we simulate these algorithms in all the networks of Table \[Tab:NetworkParametersADNC\]. In each of the plots of Figure \[Fig:ConsSize\] it is shown the number of communication steps required to achieve $10^{-2}\%$ of accuracy as a function of the number of nodes. It can be seen that D-ADMM always required less communication steps than [@Zhu09-DistributedInNetworkChannelCoding], except for the smallest network in almost all scenarios. D-ADMM also required less communication steps than [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging], except for the Erdős-Rényi networks (Figures \[Fig:ConsSize\] and \[Fig:ConsSize\]). [@Olshevsky11-ConvergenceSpeedDistributedConsensusAveraging] was the only algorithm achieving the maximum number of $1000$ iterations, which occurred for the Geometric and Lattice networks (Figures \[Fig:ConsSize\] and \[Fig:ConsSize\]). We also note a curious fact from these plots: the number of communication steps is almost invariant to the size of the networks (for the considered range). We thus conclude that D-ADMM is a competing algorithm for solving the consensus problem, since it requires, in general, less communication steps to converge than other state-of-the-art algorithm, and it has approximately the same behavior for very different networks. [[BPDN.]{}]{} We also executed some experiments comparing D-ADMM and [@Zhu09-DistributedInNetworkChannelCoding] (Algorithm \[Alg:Giannakis\]), the two best algorithms, solving BPDN . Figure \[Fig:BPDN\] shows the results of such experiments, where we only used the networks with $50$ nodes. For generating data, we used the Sparco toolbox [@Sparco], namely problems with ID’s $7$ and $902$. Problem with ID $7$ contains a $600\times 2560$ matrix $A$ with Gaussian entries, and problem with ID $902$ contains a $200\times 1000$ partial DCT matrix. In both problems, we added a small amount of noise to the provided vector $b$, in order to make the application of more realistic. The trade-off parameter $\beta$ was set to $1$ for problem ID $7$ and to $0.3$ for problem ID $902$. We ran the algorithms until they reached a relative error of $10^{-2}\%$ or the maximum number of iterations $1000$, only reached by [@Zhu09-DistributedInNetworkChannelCoding] in network $6$. At each iteration, each node solved problem with GPSR [@GPSR]. The results for problems with ID’s $7$ and $902$ are shown respectively in Figures \[Fig:BPDN\] and \[Fig:BPDN\]. These show the number of communication steps for each network model. Again, each point in the plots is the best result for $\rho \in \{10^{-4},10^{-3},\ldots,10^2\}$. It can be observed that D-ADMM always required less communication steps than [@Zhu09-DistributedInNetworkChannelCoding]; however, there were some cases where both algorithms required almost the same communication steps, namely, for networks $2$ and $3$ of Figure \[Fig:BPDN\], and for networks $2$ and $4$ of Figure \[Fig:BPDN\]. The similarity of the shape of the curves of Figures \[Fig:BPDN\] and \[Fig:BPDN\] indicates that changing the network models causes more variability in the difficulty of the problem than the problem data. [[LASSO.]{}]{} The results of our experiments for LASSO are shown in Figure \[Fig:LASSO\]. Again, we just compared the best two algorithms: D-ADMM and [@Zhu09-DistributedInNetworkChannelCoding]. For the matrix $A$ and vector $b$, we used the same data as in BPDN: problems $7$ and $902$ from the Sparco toolbox. The bound for the noise was set to $\sigma=0.5$ for problem ID $7$, and to $\sigma=0.1$ for problem ID $902$. In all experiments, which were executed for the networks with $P=10$ nodes, the regularizing parameter in was always $\delta = 10^{-3}$. With such a choice, both algorithms were able to achieve the relative error of $10^{-1}\%$ in all instances (the maximum number of iterations was never achieved). Note that the error was measured with respect to the solution of the original problem , which was obtained beforehand using the algorithm SPGL1 [@spgl1]. In Figure \[Fig:LASSO\], we can see that, in spite of solving a slightly ill-conditioned problem, D-ADMM and [@Zhu09-DistributedInNetworkChannelCoding] required considerably few communication steps to converge. Also note that the behavior of the algorithms is approximately uniform over all the types of networks. In contrast with BPDN, it was changing the data what produced different curves: the curves for ID $902$ are on top of the curves for ID $7$, which reveals that the former was harder to solve. For each problem, D-ADMM always took less communications steps than [@Zhu09-DistributedInNetworkChannelCoding] to converge. [[SVM.]{}]{} Quite surprisingly, the problem that required the largest number of communication steps was SVM. For this problem, we used data from the UCI machine learning repository [@UCIMachineLearningRepository], namely the Iris dataset, which contains two clusters of points that are linearly separable. Each point has $n = 4$ features and there are $m = 100$ total points. We first solved with the Matlab function `quadprog`. Next, we executed D-ADMM and [@Zhu09-DistributedInNetworkChannelCoding], which stopped after achieving a relative error of $10^{-2}\%$ or $10^4$ communication steps. Figure \[Fig:SVM\] shows the results for the networks with $50$ nodes. Again, D-ADMM required always less communication steps than [@Zhu09-DistributedInNetworkChannelCoding] to converge; however, both algorithms required an amount of communications one order of magnitude larger than all the previous applications. And this happened with a relatively small problem. This might indicate that problem is inherently difficult to solve in a distributed way. Conclusions =========== We proposed an algorithm for solving separable problems in a distributed way. This means that, given a connected network, each node has a cost function and a constraint set associated. While keeping their costs and constraints sets private all the time, all nodes cooperate in order to solve a global optimization problem, which minimizes the sum of all the nodes’ costs, and constrains the solution to be in the intersection of all sets. We showed how several problems from signal processing and control can be solved with the proposed algorithm, named D-ADMM, and presented extensive simulation results. These results show that D-ADMM requires almost always less communications to converge than any other state-of-the-art distributed optimization algorithm. Proof of Theorem \[Teo:ConvergenceADMM\] {#App:ADMMProof} ======================================== First, note that the assumptions imply that strong duality holds for . This follows directly from [@Bertsekas:Convex Prop.6.4.2] and it means that the optimal cost of and of have the same (finite) value, which will be denoted by $p^\star$. It also means that is solvable. Let then $(x_1^\star,x_2^\star,\lambda^\star)$ denote any primal-dual solution. We will use the notation $p^k = g_1(x_1^k) + g_2(x_2^k)$, $r^k = A_1x_1^k + A_2x_2^k$. Using only strong duality and our assumptions on $X_1$ and $X_2$, we can prove that $$\begin{aligned} &p^\star - p^{k+1} \leq {\lambda^{\star}}^\top r^{k+1} \label{Eq:ADMMProofIneq1} \\ &p^{k+1} - p^{\star} \leq -(\lambda^{k+1})^{\top} r^{k+1} {\ifthenelse{\boolean{draft}}{}{\notag \\ &\phantom{aaaaa}}}-\rho (A_2(x_2^{k+1} - x_2^k))^{\top}(A_2(x_2^{k+1} - x_2^\star) - r^{k+1}) \label{Eq:ADMMProofIneq2} \\ &V^{k+1} \leq V^k - \rho \\|r^{k+1}\\|^2 - \rho \\|A_2(x_2^{k+1} - x_2^k)\\|^2\,, \label{Eq:ADMMProofIneq3} \end{aligned}$$ where $V^k$ is the Lyapunov function $$V^k := \frac{1}{\rho}\\|\lambda^k - \lambda^\star\\|^2 + \rho\\|A_2(x_2^k - x_2^\star)\\|^2\,.$$ This is done in [@Boyd11-ADMM] and we omit the proof here. In the course of that proof, the following fact is used: if $\phi$ and $\Psi$ are convex functions and $\Psi$ is continuously differentiable, then $x^\star \in \arg\min_x\{\phi(x) + \Psi(x)\,:\, x \in X\}$ implies $x^\star \in \arg\min_x\{\phi(x) + \nabla \Psi(x^\star)^\top (x - x^\star)\,:\, x \in X\}$, for any closed convex set $X$. Using $\lambda^{k} = \lambda^{k-1} + \rho r^{k}$ and applying the previous fact to and , we get, respectively, $$\begin{aligned} x_1^{k} &\in \arg\min_{x_1 \in X_1} g_1(x_1) + (\lambda^{k} - \rho A_2(x_2^{k} - x_2^{k-1}))^\top A_1 x_1 \label{Eq:ADMMProofIneq4} \\ x_2^{k} &\in \arg\min_{x_2 \in X_2} g_2(x_2) + {\lambda^{k}}^\top A_2 x_2\,. \label{Eq:ADMMProofIneq5} \end{aligned}$$ [[Limit points of are primal optimal.]{}]{} From , we see that $0 \leq V^k \leq V^0$ for any $k$, and thus $\{\lambda^k\}$ and $\{A_2x_2^k\}$ are bounded, showing that they have limit points. Since $A_2$ has full column-rank, $\{x_2^k\}$ is also bounded and thus has limit points. To show existence of limit points of $\{x_1^k\}$, we first have to show that $r^k \xrightarrow{} 0$. Iterating we get $$\rho \sum_{k=0}^\infty (\\|r^{k+1}\\|^2 + \\|A_2(x_2^{k+1} - x_2^k)\\|^2) \leq V^0\,,$$ from which we conclude $r^k \xrightarrow{} 0$ and $A_2(x_2^{k+1} - x_2^k) \xrightarrow{} 0$ as $k \xrightarrow{} 0$. Now, $r^k = A_1x_1^k + A_2x_2^k \xrightarrow{} 0$, together with the boundedness of $\{A_2x_2^k\}$, shows that $\{A_1x_1^k\}$ is bounded and thus also $\{x_1^k\}$, because $A_1$ has full column-rank. Therefore, $\{x_1^k\}$ has limit points. We showed the existence of limit points of $\{(x_1^k,x_2^k)\}$; now, we show their optimality. In fact, taking $k \xrightarrow{} \infty$ in and , using the boundedness of $\{A_2x_2^k\}$ and $\{\lambda^k\}$, and the facts that $r^k \xrightarrow{} 0$ and $A_2(x_2^{k+1} - x_2^k) \xrightarrow{} 0$, we conclude that $p^k = g_1(x_1^k) + g_2(x_2^k) \xrightarrow{} p^\star$. [[Limit points of are dual optimal.]{}]{} We had seen that $\{\lambda^k\}$ has limit points because it is bounded. Let $\bar{\lambda}$ be such limit point, and let $\mathcal{K}$ be the set of indices such that $\{\lambda^k\}_{k \in \mathcal{K}} \xrightarrow{} \bar{\lambda}$ and also that $\{(x_1^k,x_2^k)\}_{k \in \mathcal{K}}\xrightarrow{}(\bar{x}_1,\bar{x}_2)$, where $(\bar{x}_1,\bar{x}_2)$ is a limit point of $\{(x_1^k,x_2^k)\}$. Now, define $\hat{\lambda}^k = \lambda^k - \rho A_2(x_2^k - x_2^{k-1})$. Since $A_2(x_2^k - x_2^{k-1}) \xrightarrow{} 0$, we have $\lim_{k\xrightarrow{} \infty}\{\hat{\lambda}^k\}_{k \in \mathcal{K}} = \lim_{k\xrightarrow{} \infty}\{\lambda^k\}_{k \in \mathcal{K}} = \bar{\lambda}$. Now, using the definition of $G_1$ and $G_2$, and -, we see that $$\begin{aligned} G_1(\hat{\lambda}^{k}) &= \inf_{x_1 \in X_1} g_1(x_1) + (\hat{\lambda}^{k})^\top A_1x_1 \notag \\ &= g_1(x_1^{k}) + (\hat{\lambda}^{k})^\top A_1x_1^{k} \label{Eq:ADMMProofDualProb1} \\ &\leq g_1(x_1) + (\hat{\lambda}^{k})^\top A_1x_1\,, \qquad \forall_{x_1 \in X_1}\,, \label{Eq:ADMMProofDualProb2} \\ G_2(\lambda^{k}) &= \inf_{x_2 \in X_2} g_2(x_2) + (\lambda^{k})^\top A_2x_2 \notag \\ &= g_2(x_2^{k}) + (\lambda^{k})^\top A_2x_2^{k} \label{Eq:ADMMProofDualProb3} \\ &\leq g_2(x_2) + (\lambda^{k})^\top A_2x_2\,, \qquad \forall_{x_2 \in X_2}\,. \label{Eq:ADMMProofDualProb4} \end{aligned}$$ Adding and and taking the limit $k\xrightarrow{} +\infty$ ($k \in \mathcal{K}$), where the second-to-last equality follows from primal optimality (and feasibility) of $(\bar{x}_1,\bar{x}_2)$, and the last equality follows from strong duality. Adding and and taking the limit $k\xrightarrow{} +\infty$ ($k \in \mathcal{K}$), $$\lim_{\mathcal{K}\ni k\xrightarrow{}\infty} (G_1(\hat{\lambda}^{k}) + G_2(\lambda^{k})) \leq \sum_{i=1}^2 (g_i(x_i) + \bar{\lambda}^\top A_ix_i) \,,$$ for all $x_i \in X_i$, $i=1,2$. In particular, taking the infimum on the right-hand side: $$\label{Eq:ADMMProofDualProb6} \lim_{\mathcal{K}\ni k\xrightarrow{}\infty} (G_1(\hat{\lambda}^{k}) + G_2(\lambda^{k})) \leq G_1(\bar{\lambda}) + G_2(\bar{\lambda})\,.$$ From and we conclude that $G_1(\bar{\lambda}) + G_2(\bar{\lambda}) \geq L(\lambda^\star)$, showing that $\bar{\lambda}$ is dual optimal. [[The sequence converges.]{}]{} We have seen that $\{(x_1^k,x_2^k,\lambda^k)\}$ has limit points and they are primal-dual optimal. Now, we will see that this sequence has only one limit point, and thus converges. Let $(\bar{x}_1,\bar{x}_2,\bar{\lambda})$ be a limit point of $\{(x_1^k,x_2^k,\lambda^k)\}$. Since it is primal-dual optimal, - hold with $(x_1^\star,x_2^\star,\lambda^\star)$ replaced by $(\bar{x}_1,\bar{x}_2,\bar{\lambda})$. We had seen that $V^k$ was convergent because it was bounded and nonincreasing. With the previous replacement, we now see that its limit is $0$. Therefore, $\lambda^k \xrightarrow{} \bar{\lambda}$ and $x_2^k \xrightarrow{} \bar{x}_2$ (because $A_2$ has full column-rank). Since $r^k = A_1(x_1 - \bar{x}_1) + A_2(x_2 - \bar{x}_2) \xrightarrow{} 0$, we also have $x_1^k \xrightarrow{} \bar{x}_1$ (because $A_1$ has full-column rank), i.e., $\{(x_1^k,x_2^k,\lambda^k)\}$ converges. Problem For Each Node in LASSO {#App:ProbLasso} ============================== When LASSO is solved in the column partition case, node $p$ has to solve, in each iteration, $$\label{Eq:AppLassoProb} \underset{\lambda}{\text{minimize}} \,\,\, g_p(\lambda) + v^\top \lambda + c\\|\lambda\\|^2\,,$$ for some vector $v$ and scalar $c$, and with $g_p(\lambda) = \frac{1}{P}(b^\top \lambda + \sigma\\|\lambda\\|) - \inf_{x_p}(\\|x_p\\|_1 + (A_p^\top \lambda)^\top x_p + \frac{\delta}{2}\\|x_p\\|^2)$. Define $$\label{Eq:AppLasso1} \phi_p(\lambda) = -\inf_{x_p} \,\, \\|x_p\\|_1 + (A_p^\top \lambda)^\top x_p + \frac{\delta}{2}\\|x_p\\|^2\,,$$ and let $x_p(\lambda)$ denote the solution of the optimization problem in for a given $\lambda$. Since the objective in that problem is strictly convex, $\phi_p$ is differentiable and its gradient is given by $\nabla \phi_p(\lambda) = -A_px_p(\lambda)$. Furthermore, each component of $x_p(\lambda)$ can be found in closed-form: $-(r_i + 1)/\delta$ if $r_i < -1$, $-(r_i - 1)/\delta$ if $r_i > 1$, and $0$ otherwise, where $r_i$ denotes the $i$th row of $A_p^\top \lambda$. Problem then becomes equivalent to $$\underset{\lambda}{\text{minimize}} \,\,\, \phi_p(\lambda) + \frac{1}{P}b^\top\lambda + \frac{\sigma}{P}\\|\lambda\\| + v^\top \lambda + c\\|\lambda\\|^2\,,$$ or introducing an epigraph variable $t$, $$\label{Eq:AppLasso2} \begin{array}{ll} \underset{x,t}{\text{minimize}} & \phi_p(\lambda) + \frac{1}{P}b^\top\lambda + \frac{\sigma}{P}t + v^\top \lambda + c\\|\lambda\\|^2 \\ \text{subject to} & \\|\lambda\\| \leq t\,. \end{array}$$ Problem can be solved using a projected gradient method. We will use FISTA [@BeckTeboulleFISTA §4], whose complexity is $O(1/k^2)$. FISTA solves $$\label{Eq:FistaProb} \underset{\lambda}{\text{minimize}} \,\,\, f(\lambda) + g(\lambda)\,,$$ where $f$ is a convex continuously differentiable function with a Lipschitz continuous gradient $\nabla f$ with constant $L$ (i.e., $\\|\nabla f(\eta) - \nabla f(\lambda)\\| \leq L\\|\eta - \lambda\\|$ for any $\lambda,\eta$), and $g$ is a convex function. FISTA consists of iterating $ \lambda^{k+1} = \bigl[\lambda^k - \frac{1}{L} \nabla f(\lambda^k))\bigr]^+\,, $ where $$\label{Eq:FistaProjection} [\eta]^+ = \arg\min_z \,\, g(z) + \frac{L}{2} \\|z - \eta\\|^2\,.$$ We set $f(\lambda)$ to be the objective of , and $h(\lambda)$ to be the indicator function of the set $\mathcal{S} := \{(\lambda,t)\,:\, \\|\lambda\\|\leq t\}$. In this case, becomes the projection onto $\mathcal{S}$, which can be shown to be $$[(\lambda,t)]^+_{\mathcal{S}} = \left\{ \begin{array}{lcl} (\lambda,t) &,& t\geq \\|\lambda\\| \\ (0,0) &,& t \leq -\\|\lambda\\| \\ \frac{t+\\|\lambda\\|}{2} (\frac{\lambda}{\\|\lambda\\|},1) &,& -\\|x\\|<t< \\|x\\|\,. \end{array} \right.$$ The Lipschitz constant of $\nabla f$ can also be shown to be $\sigma_{\max}^2(A_p)/\delta + 2c$, where $\sigma_{\max}(A_p)$ is the largest singular value of $A_p$. [^1]: This work was supported by the following grants from Fundação para a Ciência e Tecnologia (FCT): CMU-PT/SIA/0026/2009, PTDC/EEA-ACR/73749/2006, PEst-OE/EEI/LA0009/2011, and SFRH/BD/33520/2008 (through the Carnegie Mellon/Portugal Program managed by ICTI). [^2]: At our abstraction level, we are overlooking packet collision and other medium-access problems that may potentially destroy synchronism.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The Hyperbolic Nonlinear Schrödinger equation () arises as a model for the dynamics of three–dimensional narrowband deep water gravity waves. In this study, the Petviashvili method is exploited to numerically compute bi-periodic time-harmonic solutions of the equation. In physical space they represent non-localized standing waves. Non-trivial spatial patterns are revealed and an attempt is made to describe them using symbolic dynamics and the language of substitutions. Finally, the dynamics of a slightly perturbed standing wave is numerically investigated by means a highly acccurate Fourier solver. : Hyperbolic equations; NLS equation; wave patterns; deep water waves address: - 'LAMA, UMR 5127 CNRS, Université de Savoie, Campus Scientifique, 73376 Le Bourget-du-Lac Cedex, France' - 'University College Dublin, School of Mathematical Sciences, Belfield, Dublin 4, Ireland LAMA, UMR 5127 CNRS, Université de Savoie, Campus Scientifique, 73376 Le Bourget-du-Lac Cedex, France' - 'School of Civil and Environmental Engineering & School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, USA' author: - Laurent Vuillon - 'Denys Dutykh$^$' - Francesco Fedele bibliography: - 'biblio.bib' title: Some special solutions to the Hyperbolic NLS equation --- [^1] Introduction ============ The celebrated cubic Nonlinear Schrödinger (NLS) equation is one of the most important equations in nonlinear science [@Sulem1999]. For example, it arises in plasma physics [@Sen1978] and in normally dispersive optical waveguide arrays modeling [@Conti2003; @Lahini2007]. In the context of water waves the equation is the leading order model of the wave envelope evolution. It was derived for the first time by V. <span style="font-variant:small-caps;">Zakharov</span> (1968) [@Zakharov1968] and rediscovered later by several other authors [@Chu1970; @Hasimoto1972]. In one dimension (1-D), the NLS equation for unidirectional water waves is integrable [@Zakharov1972] and of focusing type. As a result, a steady or periodic balance of the cubic nonlinearities and wave dispersion can be attained, and this yields the formation of localized traveling waves (solitons) or homoclinic orbits to a plane wave (breathers). Analytical solutions for solitons follow via the inverse scattering transform [@Zakharov1972; @John; @Ablowitz1979; @Pelinovsky2001a] and breathers can be easily obtained via the Darboux transformation [@Osborne2010]. The two-dimensional (2-D) propagation of deep water narrowband waves is instead governed by the 2-D Hyperbolic NLS equation [@Ablowitz1979; @Yuen1982; @Sulem1999; @Pelinovsky2003]. In this case, energy can spread along the transversal direction to the main propagation. A consequence of this defocusing is that localized traveling waves cannot occur. Indeed, their non-existence was proved in ([@Ghidaglia1996a], see also [@Ghidaglia1993]). However, this does not exclude the existence of nontrivial nonlocalized* travelling wave patterns that may arise due to a balance between nonlinearities and wave dispersion in both directions under toric constraints. An extended equation was derived by <span style="font-variant:small-caps;">Trulsen</span> & <span style="font-variant:small-caps;">Dysthe</span> (1996) [@Trulsen1996] and additionally to the classical hyperbolic (D’Alembert) operator, it contains also higher order dispersive and nonlinear terms. The generalization to the finite depth case leads to the Davey–Stewartson equations [@Benney1967; @Davey1974; @Ghidaglia1990]. Several important analytical solutions to the the Davey–Stewartson model were derived in [@Ebadi2011; @Ebadi2011a; @Yildirim2012; @Jafari2012]. For mathematical/numerical studies it is often convenient to restrict the attention to a particular class of solutions. For example, the very first mathematical description of plane permanent waves is known at least since G. <span style="font-variant:small-caps;">Stokes</span> (1847) [@Stokes1847]. Permanent waves can be periodic or localized in space. In this study we focus on bi-periodic time-harmonic solutions of the form $A = F(x - c_g t, y){\mathrm{e}}^{{\mathrm{i}}\omega t}$, where $A$ is the complex envelope and $F$ a complex function of its argument. These are stationary solutions in the frame of reference moving with the group speed $c_g$. In the physical domain, the associated wave surface displacements is that of standing waves, which have been the subject of many studies. In particular, radial standing solutions of the equation were investigated in [@Kevrekidis2011]. The existence of standing waves in deep waters were proved in [@Iooss2005]. In the shallow-water regime standing wave patterns have been found in the context of the Boussinesq equations by M. <span style="font-variant:small-caps;">Chen</span> & G. <span style="font-variant:small-caps;">Iooss</span> [@Chen2005; @Chen2006; @Chen2008]. 2-D bi-periodic travelling wave solutions to the Euler equations were studied by W. <span style="font-variant:small-caps;">Craig</span> & D. <span style="font-variant:small-caps;">Nicholls</span> (2002) [@Craig2002]. Recently, 2-D wave patterns of the free surface were investigated experimentally by D. <span style="font-variant:small-caps;">Henderson</span> et al. (2010) [@Henderson2010] along with a theoretical stability analysis. In this work, symbolic dynamics and associated techniques of substitutions are exploited to investigate the structure of standing wave patterns (see [@Morse1940; @Lind1995; @Berthe2000]). In physics, their application in studies of dynamical systems led to unveiling the structure of quasi-periodic tilings (see [@Senechal1995]). Symbolic dynamics allows coding the nonlinear behavior of a complex system and pattern formation by means of 1-D or 2-D words of a finite alphabet. Such approach was applied to code the non-periodic trajectories on a unit circle with a particular partition on two intervals (see [@Morse1940]). Coding of finite, periodic and non-periodic infinite patterns using 2-D words and tilings was done in [@Berthe2000; @DeBruijn1981]. The numerical standing waves investigated in this work are periodic in space, thus the associated patterns are described up to toric constraints. This is the first step for developing a theory for the description of periodic or quasi-periodic patterns associated to trajectories $\{a_n(t)\}$ of the dynamics in the infinite phase space spanned by, for example, generalized Fourier basis $\phi_n(x,y)$ associated to the formal series for $A(x,y,t) = \sum_n a_n(t)\phi_n(x,y)$ on a periodic domain. The present study is organized as follows. First, the equation is introduced in the context of deep water waves. Then, the Petviashvili method [@Petviashvili1976] used to compute a class of special standing wave solutions is presented. Symbolic dynamics is then applied to describe the associated spatial patterns using words and substitutions. Finally, a highly accurate Fourier-based solver is exploited to investigate the dynamics of a perturbed standing wave. The Mathematical Model {#sec:math} ====================== Consider a three-dimensional fluid domain with a free surface. The water is assumed to be infinitely deep. The Cartesian coordinate system $Oxyz$ is chosen such that the undisturbed water level corresponds to $z = 0$, and the free surface elevation is $z = \eta(x,y,t)$. The Euler equations that describe the irrotational flow of an ideal incompressible fluid of infinite depth with a free surface are of fundamental relevance in fluid mechanics, ocean sciences and both pure and applied mathematics (see for example, [@Stoker1957; @Zakharov1968; @Johnson1997]). The structure of the Euler equations is given in terms of the free-surface elevation $\eta(x, y, t)$ and the velocity potential $\varphi(x, y, t) = \phi(x, y, z = \eta(x,y,t), t)$ evaluated at the free surface of the fluid. In late 70s Dysthe used the method of multiple scales to derive from the Euler equations a modified Nonlinear Schrödinger (NLS) equation [@Dysthe1979] for the time evolution of the unidirectional narrowband envelope $A$ of the velocity potential $\varphi$ with carrier wave ${\mathrm{e}}^{{\mathrm{i}}(k_0 x - \omega_0 t)}$, $k_0$ and $\omega_0$ being, respectively, the wavenumber and frequency of the carrier wave. The equation for $A$ can be formulated in a frame moving with the group velocity $c_g = \omega_0/(2k_0)$ as follows. Define ${\varepsilon}$ as a small parameter, $a_0$ as a characteristic wave amplitude and rescale space, time and the envelope as $x \to k_0 x-c_g t$, $y \to k_0 y$, $t \to \omega_0 t$ and $A \to {\varepsilon}a_0 A$ respectively. Then, the 2-D Dysthe equation for $A$ is given by [@Trulsen1996]: $$\begin{gathered} \label{eq:Dysthe} {\mathrm{i}}A_t = \frac{1}{8}A_{xx} - \frac{1}{4}A_{yy} + \frac{1}{2}\|A\|^2A \\ - {\mathrm{i}}{\varepsilon}\bigl(\frac{1}{16}A_{xxx} - \frac{3}{8}A_{xyy} - \frac{3}{2}\|A\|^2A_x + \frac{1}{4}A^2A^{\star}_x - {\mathrm{i}}A{\mathcal{H}}[\|A\|^2]\bigr),\end{gathered}$$ where ${\mathcal{H}}$ is the Hilbert transform and the subscripts $A_t = \partial_t A$, $A_x = \partial_x A$ denote partial derivatives with respect to $x$, $y$ and $t$ respectively, and $A^{\star}$ denotes complex conjugation. The envelope of the free surface $\eta$ relates to $A$ by a simple transformation that involves only $A$ and its derivatives (see [@Dysthe1979]). To ${\mathcal{O}}(1)$ in ${\varepsilon}$, the 2-D Dysthe equation reduces to the equation: $$\label{eq:hNLS} {\mathrm{i}}A_t = \frac{1}{8}A_{xx} - \frac{1}{4}A_{yy} + \frac{1}{2}\|A\|^2A.$$ By rescaling $A \to 2A$, $x \to 2\sqrt{2} x$ and the transverse coordinate $y \to 2y$, takes the form [@Sulem1999; @Pelinovsky2003]: $$\label{eq:adim} {\mathrm{i}}A_t = A_{xx} - A_{yy} + 2A\|A\|^2.$$ This equation admits the three invariants ${\mathbb{H}}$, ${\mathbb{A}}$ and ${\mathbb{M}}$, which have the meaning of energy, wave action and momentum respectively: $$\begin{aligned} {\mathbb{H}}&=& \iint_{{\mathds{R}}^2}\Bigl\{\|A_y\|^2 - \|A_x\|^2 + \|A\|^4\Bigr\}\,{\mathrm{d}}x\,{\mathrm{d}}y, \\ {\mathbb{A}}&=& \iint_{{\mathds{R}}^2}\|A\|^2\,{\mathrm{d}}x\,{\mathrm{d}}y, \\ {\mathbb{M}}&=& \frac{{\mathrm{i}}}{2}\iint_{{\mathds{R}}^2}\Bigl\{A{\boldsymbol{\nabla}}A^\star + A^\star{\boldsymbol{\nabla}}A\Bigr\}\,{\mathrm{d}}x\,{\mathrm{d}}y.\end{aligned}$$ Note that the total energy ${\mathbb{H}}$ is also the Hamiltonian for the equation. In the following, we will solve for a special class of standing wave solutions to . Standing wave patterns {#sec:gs} ---------------------- Consider the ansatz for standing wave solutions that oscillate harmonically in time $$A(x,y,t) = {\mathrm{e}}^{-{\mathrm{i}}\omega t}B(x,y),$$ where the real function $B$ describes the spatial pattern of a standing wave. According to the unscaled equation , $B$ satisfies the following real nonlinear hyperbolic PDE $$\label{eq:Beq} \omega B + B_{xx} - B_{yy} = 2B^3,$$ which can be solved numerically using the classical Petviashvili method [@Petviashvili1976; @Pelinovsky2004; @Lakoba2007]. To do so, is rewritten in the operator form $${\mathcal{L}}\cdot B = {\mathcal{N}}(B), \qquad {\mathcal{L}}:= \omega + \partial_{xx} - \partial_{yy}, \qquad {\mathcal{N}}(B) := 2B^3.$$ and the iteration scheme is given by $$B_{n+1} = {\ensuremath{\mathcal{S}}}^\gamma {\mathcal{L}}^{-1}\cdot {\mathcal{N}}(B_n), \quad {\ensuremath{\mathcal{S}}}= \frac{\langle B_n, {\mathcal{L}}\cdot B_n\rangle}{\langle B_n, {\mathcal{N}}(B_n)\rangle},$$ where ${\ensuremath{\mathcal{S}}}$ is the so-called stabilyzing factor and the exponent $\gamma$ is usually defined as a function of the degree of nonlinearity $p$ ($p = 3$ for the equation). The rule of thumb prescribes the following formula $\gamma = \frac{p}{p-1}$. The scalar product is defined in the $L_2$ space. The inverse operator ${\mathcal{L}}^{-1}$ can be efficiently computed in Fourier space using the Fast Fourier Transform (FFT) (see, for example, [@Frigo2005]), and spatial periodicity across the 2D computational box boundaries is implicitly imposed. The iterative process converges for a large class of smooth initial guesses[^2]. Convergence is attained when the $L_\infty$ norm between two successive iterations is less than a prescribed tolerance $\varepsilon$ (usually of the order of machine precision). Additionally, the residual error $E_r$ in approximating the nonlinear equation is checked by substituting in the converged numerical solution. In the present work, the convergence of the algorithm is checked numerically in the extended floating point arithmetics using $30$ significant digits [@MATLAB2012]. For example, consider the domain $\Omega := [-\ell_x/2, \ell_x/2] \times [-\ell_y/2, \ell_y/2] \subset {\mathds{R}}^2$ to be a square with the side length equal to $210$ ($\Omega = [-105, 105]^2$). For $\omega = 0.012$, the Petviashvili method on a $1024 \times 1024$ Fourier grid yields the strictly periodic regular pattern shown in Figure \[fig:f0\](a). The convergence is attained in $N\sim 80$ iterations as clearly seen in Figure \[fig:f0\](b) and the associated error $L_\infty \sim {\mathcal{O}}(10^{-33})$, and the residual $E_r \sim {\mathcal{O}}(10^{-32})$. Note that $L_\infty \sim~{\mathrm{e}}^{-0.095 N}$ decays exponentially in agreement with the theoretical geometric convergence rate [@Alvarez2014]. For the sake of efficiency, the numerical solutions presented below are computed on the same Fourier grid using standard double-precision arithmetics. It is verified that residual errors are within the prescribed tolerance parameter $\varepsilon \sim 10^{-15}$ and never exceed $10\varepsilon$. ![A bi-periodic wave pattern $B(x,y)$, $\Omega = [-30, 30]\times[-15, 15]$ and $\omega = 0.18$. $B$ can be described by the $1\times 4$ discrete pattern $E = {}^t(r, b, r, b)$ on the two-letter alphabet $\{r, b\}$ (red spots code negative values far from zero and blue spots refer to positive values far from zero).[]{data-label="fig:f1"}](figs/gstates/60x30_0_18_select.eps){width="75.00000%"} ![A bi-periodic wave pattern $B(x,y)$, $\Omega = [-30, 30]^2$ and $\omega = 0.18$. The box delimits the elementary $6\times 10$ discrete pattern of blue ($b$), red ($r$) and white ($w$, denoting the values around zero) spots identified to describe $B$ by substitutions on the three-letter alphabet $\{r, b, w\}$. For symbolic coding see caption of Figure \[fig:f0\].[]{data-label="fig:f2"}](figs/gstates/60x60_0_18_select.eps){width="75.00000%"} [figs/gstates/60x40\_0\_18\_select.eps]{} (24,54.5) [$\sigma(b)$]{} (43,54.5) [$\sigma(a)$]{} (24,47.5) [$\sigma(c)$]{} (43,47.5) [$\sigma(d)$]{} (24,39.5) [$\sigma(\bar{b})$]{} (43,39.5) [$\sigma(\bar{a})$]{} (24,30) [$\sigma(\bar{a})$]{} (43,30) [$\sigma(\bar{b})$]{} (24,22) [$\sigma(d)$]{} (43,22) [$\sigma(c)$]{} (24,14) [$\sigma(a)$]{} (43,14) [$\sigma(b)$]{} ![A bi-periodic wave pattern $B(x,y)$, $\Omega = [-30, 30]\times[-60, 60]$ and $\omega = 0.18$. The box delimits the elementary $32\times 66$ discrete pattern of blue ($b$) red ($r$) and white ($w$) spots identified to describe $B$ by substitutions on the three-letter alphabet $\{r, b, w\}$. For symbolic coding see Figure \[fig:f0\].[]{data-label="fig:f4"}](figs/gstates/60x120_0_18_select.eps){width="99.00000%"} ![A bi-periodic wave pattern $B(x,y)$, $\Omega = [-\frac{55}{2}, \frac{55}{2}]\times[-\frac{89}{2}, \frac{89}{2}]$ and $\omega = 0.51$.[]{data-label="fig:f5"}](figs/gstates/55x89_0_51.eps){width="99.00000%"} ![(Right) a bi-periodic wave pattern $B(x,y)$, $\Omega = [-105, 105]^2$ and $\omega = 1.3295$; (Left) zoom on the sub-region $[-40, -10] \times [30, 80]$.[]{data-label="fig:f6"}](figs/gstates/210x210_1_3295_select.eps){width="99.00000%"} Symbolic coding {#sec:descr} =============== The theory of bi-dimensional words and tilings (see [@Berthe2000; @DeBruijn1981]) can be exploited to describe standing wave patterns of the equation. The key tool in tiling theory is the substitution operation that consists in replacing letters by group of letters and constructing iteratively words on a finite alphabet. For example, the Fibonacci substitution $\sigma(a) = ab$ and $\sigma(b) = a$ constructs a fixed point (limit of infinite substitutions) with a non-periodic structure ([@Arnoux1991; @Lind1995; @Allouche2003]. This word constitutes a model for 1-D quasicrystals in physics (see [@Senechal1995]) and a coding of a discrete line with irrational slope in discrete geometry (see [@Vuillon2003]). The method of substitutions empowers the construction of either non-periodic or periodic words $G$ as fixed points of the mapping $G=\sigma^n(G)$ as $n\to +\infty$ (see [@Lind1995]). On the one hand, non-periodic infinite words can be constructed iteratively by the use of morphism properties. In particular, a substitution $\sigma$ applied to a word $w=w_1w_2\cdots w_n$ is in fact the image of the substitution applied to each letter of $w$, that is $\sigma(w_1w_2\cdots w_n) = \sigma(w_1) \sigma(w_2) \cdots \sigma(w_n)$, where $w_i$ are letters of the alphabet. As an example, the first few iterations on the letter $a$ using the Fibonacci substitution are given by $\sigma^1(a) = \sigma(a) = ab$, $\sigma^2(a) = \sigma(\sigma(a)) = \sigma(ab) = \sigma(a) \sigma(b) = aba$, $\sigma^3(a) = \sigma(\sigma^2(a)) = \sigma(aba) = \sigma(a)\sigma(b)\sigma(a) = abaab$, and so on (see [@Arnoux1991; @Vuillon2003]). Note that each word $\sigma^{n}(a)$ is the beginning of the next word $\sigma^{n+1}(a)$. Thus, as $n\to +\infty$, a non-periodic fixed point is constructed, viz. $$\sigma(F) = F = abaababaabaababaababaabaababaabaababaababaabaababaababaabaababaaba\cdots.$$ On the other hand, a periodic word can be constructed by means of the substitution $\sigma(a)=ab$ and $\sigma(b)=ab$ by repetition of the pattern $ab.$. Indeed $\sigma^1(a)=\sigma(a)=ab, $$\sigma^2(a) = \sigma(\sigma(a)) = \sigma(ab) = \sigma(a)\sigma(b) = abab$, $\sigma^3(a) = \sigma(\sigma^2(a)) = \sigma(abab) = \sigma(a)\sigma(b)\sigma(a)\sigma(b) = abababab,$ and so on. The fixed point is periodic and given by $$\sigma(P) = P = abababababababababababababababababababababababababababa\cdots.$$ In order to code periodic spatial patterns of the equation, one can exploit bi-dimensional words (see [@Frougny2005; @Berthe2000; @Lind1995]) and proper substitutions (see [@Arnoux1991; @Berthe2000; @Frougny2005]). For example, the bi-dimensional Thue–Morse substitution (see [@Allouche2003]) is defined by $$\sigma(a) = \begin{pmatrix} a & b \\ b & a \end{pmatrix} \quad \mbox{and} \quad \sigma(b) = \begin{pmatrix} b & a \\ a & b \end{pmatrix}.$$ One can just iterate a finite number of substitutions, viz. $\sigma^j(a)$ with $j$ finite, and obtain a finite world. For example, the first two iterations of the Thue-Morse substitution yield, respectively, the two words $$\sigma^1(a) = \sigma(a) = \begin{pmatrix} a & b \\ b & a \end{pmatrix}$$ and $$\sigma^2(a) = \sigma \begin{pmatrix} a & b \\ b & a \end{pmatrix} = \begin{pmatrix} \sigma(a) & \sigma(b) \\ \sigma(b) & \sigma(a) \end{pmatrix} = \begin{pmatrix} a & b & b & a \\ b & a & a & b \\ b & a & a & b \\ a & b & b & a \end{pmatrix}$$ In the limit of infinite iterations, an infinite word is obtained in the form of the non-periodic fixed point $$\sigma(M) = M = \begin{pmatrix} a & b & b & a & b & a \cdots \\ b & a & a & b & a & b \cdots \\ b & a & a & b & a & b \cdots \\ a & b & b & a & b & a \cdots \\ b & a & a & b & a & b \cdots \\ a & b & b & a & b & a \cdots \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots ~ \ddots \end{pmatrix}$$ Hereafter, a discrete description of several spatial patterns $B(x,y)$ of the equation obtained via the Petviashvili method is presented. Each continuous pattern is coded on a finite alphabet using proper substitutions iterated on a discrete elementary pattern of letters. Consider the bi-periodic pattern $B$ of the equation shown in Figure \[fig:f0\]. This can be easily described by repetition of the elementary 3x3 discrete pattern delimited by a box in the same Figure. To do so, coding with a three–letter alphabet is used since the pattern is characterized by a discrete structure with three different elementary spots. The letters $r$,$b$ and $w$ are used to code red spots (negative values far from zero), blue spots (positive values far from zero) and white spots (values around zero) respectively. As clearly seen in Figure \[fig:f0\], the three spots are arranged on a rectangular sub-region of the ${\mathbb{Z}}^2$ grid. Thus, the continuous periodic pattern $B$ can be easily described by bi-dimensional words constructed using the two simple substitutions $$\sigma(r) = \begin{pmatrix} r & b \\ b & r \end{pmatrix} \quad \mbox{and}\quad \sigma(b) = \begin{pmatrix} r & b \\ b & r \end{pmatrix}.$$ By iterating the above substitutions yields the strictly periodic word $$\sigma(R) = R = \begin{pmatrix} r & b & r & b \cdots\\ b & r & b & r \cdots\\ r & b & r & b \cdots \\ b & r & b & r \cdots\\ \vdots & \vdots & \vdots & \vdots ~ \ddots \end{pmatrix}$$ which describes the continuous wave pattern of Figure \[fig:f0\]. Consider now the wave pattern solutions for $B(x,y)$ obtained for $\omega = 0.18$ on various rectangular grid sizes. For $\Omega = [-30, 30]\times [-15, 15]$, $B$ can be described by a $1 \times 4$ discrete pattern delimited by a box in Figure \[fig:f1\] and coded as $E = {}^t(r, b, r, b)$, where the superscript ${}^t$ denotes matrix transposition. $E$ can be decomposed as $\sigma(e) = {}^{t}(r, b)$ and the initial pattern $E$ is given by the substitution $\sigma$ applied to $E = {}^t(e, e)$ that is: $$\sigma(E) = {}^t\bigl(\sigma(e), \sigma(e)\bigr) = {}^t\bigl(r, b, r, b\bigr).$$ For $\Omega = [-30, 30]\times [-30, 30]$, the associated continuous pattern $B$ is shown in Figure \[fig:f2\]. A box delimits the elementary $6 \times 10$ discrete pattern $F$ that can describe $B$ by successive substitutions, viz. $$F = \begin{pmatrix} b & w & b & w & b & w \\ r & w & r & w & r & w \\ b & r & b & r & b & r \\ r & b & r & b & r & b \\ w & r & w & r & w & r \\ w & b & w & b & w & b \\ w & r & w & r & w & r \\ r & b & r & b & r & b \\ b & r & b & r & b & r \\ r & w & r & w & r & w \end{pmatrix}.$$ This discrete pattern can be decomposed as $$\sigma(f) = {}^t\begin{pmatrix} w & w & r & b & r & b & r & b & r & w \\ b & r & b & r & w & w & w & r & b & r \end{pmatrix}.$$ As a result, the initial pattern $F$ is given by the substitution $\sigma$ applied to $F = (f~f~f)$, that is $$\sigma(F) = \bigl(\sigma(f) ~ \sigma(f) ~ \sigma(f)\bigr) = \begin{pmatrix} b & w & b & w & b & w \\ r & w & r & w & r & w \\ b & r & b & r & b & r \\ r & b & r & b & r & b \\ w & r & w & r & w & r \\ w & b & w & b & w & b \\ w & r & w & r & w & r \\ r & b & r & b & r & b \\ b & r & b & r & b & r \\ r & w & r & w & r & w \end{pmatrix}.$$ For the larger domain $\Omega = [-30, 30]\times [-20, 20]$, the associated standing wave pattern $B(x,y)$ is reported in Figure \[fig:f3\]. An elementary $32\times 22$ discrete pattern of blue ($b$), red ($r$) and white ($w$) spots is identified and delimited by a box to describe $B$. This can be obtained by four iterations of the following 6 substitutions on the three-letter alphabet $\{r, b, w\}$, that is $$\sigma(a) = \begin{pmatrix} r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(\bar{a}) = \begin{pmatrix} w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(b) = \begin{pmatrix} w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(\bar{b}) = \begin{pmatrix} b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(c) = \begin{pmatrix} r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \end{pmatrix}$$ $$\sigma(d) = \begin{pmatrix} w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \end{pmatrix}$$ As described above, this substitution is applied four times under toric constraints starting from the finite $4 \times 6$ pattern $G$ given by $$G = \begin{pmatrix} b & a & b & a \\ c & d & c & d \\ \bar{b} & \bar{a} & \bar{b} & \bar{a} \\ \bar{a} & \bar{b} & \bar{a} & \bar{b} \\ d & c & d & c \\ a & b & a & b \end{pmatrix}.$$ Here, the difference between $\sigma(a)$ and $\sigma(\bar{a})$ resides in the interchange of the letters $r$ and $w$ in the first 3 rows of the pattern except the last column, and similarly for $\sigma(b)$ and $\sigma(\bar{b})$. This switching between components is exactly the typical dynamical property that we expect for symbolic coding of patterns. For example, the first iteration $\sigma(G)$ yields $$\sigma(G) = \sigma\left(\begin{array}{llllllllllllllll} b & a & b & a \\ c & d & c & d \\ \bar{b} & \bar{a} & \bar{b} & \bar{a} \\ \bar{a} & \bar{b} & \bar{a} & \bar{b} \\ d & c & d & c \\ a & b & a & b \end{array}\right) = \begin{pmatrix} \sigma(b) & \sigma(a) & \sigma(b) & \sigma(a) \\ \sigma(c) & \sigma(d) & \sigma(c) & \sigma(d) \\ \sigma(\bar{b}) & \sigma(\bar{a}) & \sigma(\bar{b}) & \sigma(\bar{a}) \\ \sigma(\bar{a}) & \sigma(\bar{b}) & \sigma(\bar{a}) & \sigma(\bar{b}) \\ \sigma(d) & \sigma(c) & \sigma(d) & \sigma(c) \\ \sigma(a) & \sigma(b) & \sigma(a) & \sigma(b) \end{pmatrix}.$$ The whole discrete description of the continuous pattern $B(x,y)$ follows after four iterations as $$\sigma(G) = \left(\begin{array}{lllllllllllllllllllllllllllllllllll} w & b & w & b & w & b & w & w & r & w & r & w & r & w & r & w \cdots \\ b & w & b & w & b & w & b & w & w & r & w & r & w & r & w & w \cdots \\ w & b & w & b & w & b & w & w & r & w & r & w & r & w & r & w \cdots \\ w & w & w & w & w & w & w & w & w & w & w & w & w & w & w & w \cdots \\ r & w & r & w & r & w & r & w & w & b & w & b & w & b & w & w \cdots \\ w & r & w & r & w & r & w & w & b & w & b & w & b & w & b & w \cdots \\ r & w & r & w & r & w & r & w & w & b & w & b & w & b & w & w \cdots \\ b & w & b & w & b & w & b & w & w & r & w & r & w & r & w & w \cdots \\ w & b & w & b & w & b & w & w & r & w & r & w & r & w & r & w \cdots \\ b & w & b & w & b & w & b & w & w & r & w & r & w & r & w & w \cdots \\ w & w & w & w & w & w & w & w & w & w & w & w & w & w & w & w \cdots \\ w & r & w & r & w & r & w & w & b & w & b & w & b & w & b & w \cdots \\ r & w & r & w & r & w & r & w & w & b & w & b & w & b & w & w \cdots \\ w & r & w & r & w & r & w & w & b & w & b & w & b & w & b & w \cdots \\ w & w & w & w & w & w & w & w & w & w & w & w & w & w & w & w \cdots \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots ~ \ddots \end{array}\right)$$ For $\Omega = [-30, 30]\times [-60, 60]$, the associated water pattern $B$ can be described by the elementary $32 \times 66$ discrete pattern $K$ delimited by a box and shown in Figure \[fig:f4\]. In this case, the symbolic description of $K$ follows from the substitution $K=\sigma(H)$, where $\sigma$ is defined as $$\sigma(a) = \begin{pmatrix} r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(\bar{a}) = \begin{pmatrix} w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(b) = \begin{pmatrix} w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(\bar{b}) = \begin{pmatrix} b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & w & w & w & w & w & w & w \end{pmatrix},$$ $$\sigma(c) = \begin{pmatrix} r & w & r & w & r & w & r & w \\ w & r & w & r & w & r & w & w \\ r & w & r & w & r & w & r & w \end{pmatrix}$$ $$\sigma(d) = \begin{pmatrix} w & b & w & b & w & b & w & w \\ b & w & b & w & b & w & b & w \\ w & b & w & b & w & b & w & w \end{pmatrix}$$ and applied to the following $4 \times 12$ pattern $H$ under toric constraints, viz. $$H = \begin{pmatrix} G \\ G \end{pmatrix} = {}^t\left( \begin{array}{llllllllllll} a & d & \bar{a} & \bar{b} & c & b & a & d & \bar{a} & \bar{b} & c & b \\ b & c & \bar{b} & \bar{a} & d & a & b & c & \bar{b} & \bar{a} & d & a \\ a & d & \bar{a} & \bar{b} & c & d & a & d & \bar{a} & \bar{b} & c & b \\ b & c & \bar{b} & \bar{a} & d & a & b & c & \bar{b} & \bar{a} & d & a \end{array}\right)$$ Note that $H$ is made of two copies of the discrete structure $G$ identified for the pattern $B(x,y)$ relative to the domain $\Omega = [-30, 30]\times [-20, 20]$, see Figure \[fig:f3\]. Thus, $\sigma(H)$ is exactly the substitution applied to two copies of $G$, viz. $\sigma(H) = {}^t\bigl(\sigma(G), \sigma(G)\bigr)$. This suggests that for a given value of $\omega$ the patterns associated to various domains $\Omega$ have in common the same substitutions. These capture the structure of the discrete patterns and the nature of the dynamical system. Future research aims at finding a sequence of increasing domain sizes whose associated discrete patterns share the same substitution. This leads to construct periodic orbits and fixed points of the dynamics. For example, Figure \[fig:f5\] shows a patterns that may be generated by substitutions of elementary cell patterns. Moreover, for given domain size, one could explore which frequencies $\omega$ gives the same kind of substitutions and try to explain this regularity by the arithmetic nature of $\omega$. More precisely, suppose that for a given $\omega$ one identifies two discrete patterns generated by repeated substitutions $\sigma^k(g)$ and $\sigma^\ell(g)$, respectively, with $k$ and $\ell$ as the number of repeations. Then, one may find a sequence of increasing domain sizes $(\Omega_i)_{i \in I}$ and try to associate substitutive patterns of the form $\sigma^i(g).$ This could yield characteristic scales with invariance of patterns and could suggest that the dynamics is given by a coding of iterated substitutions of the form $(\sigma^i(g)))_{i \in I}$. A fixed point $G$ such that $\sigma(G) = G$ could then exist (see [@Morse1940; @Lind1995]). Furthermore, note that there is a link between the complexity of the patterns and the decomposition in prime factors of the domain $\Omega$ size. For example, the domain $\Omega = [-\ell/2, \ell/2] \times [-\ell/2, \ell/2]$ with length $\ell = 2 \times 3 \times 5 \times 7 = 210$ is the product of the four first prime numbers and the associated complex pattern can be seen in Figure \[fig:f6\]. The apparent complexity of the solution may be explained by a combinations of well chosen substitutions. One expects that the larger domain $\Omega$ with length $\ell = 2 \times 3 \times 5 \times 7 \times 11 = 2310$ leads to even more complicated pattern solutions. Indeed, this decomposition in prime factors increases the number of divisors of the pattern size and the complexity of patterns. To summarize, this part presents the first step for a symbolic dynamic coding of patterns arising from non linear waves. We will use in next works these substitutions in order to understand the complexity of such patterns. Dynamics of pertubed standing waves {#sec:descr} =================================== Pseudo-spectral scheme ---------------------- A highly accurate Fourier-type pseudo-spectral method [@Trefethen2000; @Boyd2000] will be used to solve the unsteady equation in order to investigate the dynamics of a perturbed standing wave pattern. To do so, equation is recast in the following form: $$\label{eq:oper} A_t + {\mathrm{i}}{\mathcal{L}}\cdot A = {\mathcal{N}}(A),$$ where the operators ${\mathcal{L}}$ and ${\mathcal{N}}$ are defined such as: $${\mathcal{L}}:= \partial_{xx} - \partial_{yy}, \qquad {\mathcal{N}}(A) = -2{\mathrm{i}}\|A\|^2A.$$ With this setting, is discretized by applying the 2-D Fourier transform in the spatial variables $(x, y)$. The nonlinear terms are computed in physical space, while spatial derivatives are computed spectrally in Fourier space. The standard $3/2$ rule is applied for anti-aliasing [@Trefethen2000; @Clamond2001; @Fructus2005]. The transformed variables will be denoted by $\hat{A}(t, {\boldsymbol{k}}) = {\mathcal{F}}\{A(t, x, y)\}$, with ${\boldsymbol{k}}= (k_x, k_y)$ being the Fourier transform parameter. In order to improve the stability of the time discretization procedure, the linear part of the operator is integrated exactly by a change of variables [@Milewski1999; @Fructus2005] that yields $$\hat{{\mathcal{A}}}_t = {\mathrm{e}}^{(t-t_0){\mathcal{L}}}\cdot{\mathcal{N}}\Bigl\{{\mathrm{e}}^{-(t-t_0){\mathcal{L}}}\cdot\hat{{\mathcal{A}}}\Bigr\}, \qquad \hat{{\mathcal{A}}}(t) := {\mathrm{e}}^{(t-t_0){\mathcal{L}}}\cdot\hat{A}(t), \qquad \hat{{\mathcal{A}}}(t_0) = \hat{A}(t_0).$$ The exponential matrix $\hat{{\mathcal{L}}}$ is explicitly computed in Fourier space as $${\mathrm{e}}^{(t-t_0)\hat{{\mathcal{L}}}} = {\mathrm{e}}^{-{\mathrm{i}}(k_x^2 - k_y^2)(t-t_0)}.$$ Finally, the resulting system of ODEs is discretized in space by the Verner’s embedded adaptive 9(8) Runge–Kutta scheme [@Verner1978]. The step size is chosen adaptively using the so-called H211b digital filter [@Soderlind2003; @Soderlind2006] to meet the prescribed error tolerance, set as of the order of machine precision. Numerical results ----------------- Consider the bi-periodic pattern computed for $\Omega=[-30, 30]^2$ and $\omega = 0.30$ and shown in Figure \[fig:f30\]. To simulate the dynamics the pseudo-spectral method described above is used with $256\times 256$ Fourier modes. The numerical solver independently confirmed that the time-harmonic standing wave $A(x,y,t) = B(x,y){\mathrm{e}}^{-{\mathrm{i}}\omega t}$ associated to the spatial pattern $B(x,y)$ of Figure \[fig:f30\] computed using the Petviashvili scheme is effectively a solution of the original equation . Then, another experiment was performed where the initial condition for the solver was set as $A(x,y,0) = B(x,y) + w(x,y)$, where $w(x,y)$ is approximatively a $7\%$ double-periodic perturbation with the wavelength four times smaller than that of the unperturbed pattern $B(x,y)$. The simulations were carried out up to the dimensionless time $T = 14.0$. The energy (Hamiltonian) ${\mathbb{H}}$ and action ${\mathbb{A}}$ were conserved with 12 digits accuracy during the whole simulation. The momentum ${\mathbb{M}}$ was preserved to machine precision. On short time scales slight oscillations around the unperturbed solution occur. However, on a much longer time scale a transition to another solution is observed, which is quite similar in shape to the $B(x,y)$, but slightly shifted in space. A few snapshots taken from the dynamical simulation are depicted in Figure \[fig:dyn\] (see also video [@Vuillon2014]). To visualize the dynamics, $A$ is projected onto the subspace ${\ensuremath{\mathcal{S}}}= \operatorname{span}\bigl\{\phi_1, \phi_2, \phi_3\bigr\}$ spanned by the first three leading Karhunen–Loève (KL) eigenmodes $\phi_j$, $j=1,\ldots, 3$ (see, for example [@Ghanem2003]). These are estimated from the numerical simulations using the method of snapshots (see [@Sirovich1987; @Rowley2005]) after the time average is removed. The associated trajectory $\gamma(t)$ in ${\ensuremath{\mathcal{S}}}$ and the three KL modes are shown in Figure \[fig:dyn\]. The first two modes represent the most energetic structures of the imposed perturbation, whereas the 3[$\,^{\mathrm{\small\textsf{rd}}}$]{} mode arises due to the nonlinear interaction between the unperturbed standing wave and the perturbation. Note that $\gamma$ lies approximately on a cylindrical manifold. The motion is circular on the $x_1$–$x_2$ plane with oscillations in the vertical $x_3$ axis. In physical space the dynamical wave patterns smoothly vary between the 1[$\,^{\mathrm{\small\textsf{st}}}$]{} and 2[$\,^{\mathrm{\small\textsf{nd}}}$]{} KL mode in a periodic fashion while being modulated by the 3[$\,^{\mathrm{\small\textsf{rd}}}$]{} mode. A new dynamical state is reached, which is not a standing wave. ![A bi-periodic wave pattern, $\Omega = [-30, 30]^2$ and $\omega = 0.30$.[]{data-label="fig:f30"}](figs/gstates/60x60_0_30.eps){width="99.00000%"} Conclusions {#sec:concl} =========== In this study a class of special solutions to the hyperbolic NLS equation have been investigated. In particular, bi-periodic standing waves are obtained numerically using the iterative Petviashvili scheme [@Petviashvili1976; @Pelinovsky2004; @Lakoba2007]. Non-trivial wave patterns are revealed by varying the computational domain $\Omega$ and the frequency $\omega$ of the standing wave. These are described by means of symbolic dynamics and the language of substitutions. For given value of $\omega$, the patterns associated to different domains $\Omega$ have in common the same substitution rule $\sigma$. The dynamics of a perturbed standing wave is also numerically investigated by means a highly acccurate Fourier solver in the reduced state space $S$ spanned by the first three dominant KL eigenmodes. The trajectory in $S$ lies approximately on a cylindrical manifold and in physical space the wave pattern appears to vary both in space and time. The discrete symbolic construction is the first step for developing a whole theory of description of periodic patterns of the equations. The theoretical explanation of steady and unsteady wave patterns via coding remains a major challenge and also a perspective opened by this study. Acknowledgements {#acknowledgements .unnumbered} ================ D. <span style="font-variant:small-caps;">Dutykh</span> acknowledges the support from ERC under the research project ERC-2011-AdG 290562-MULTIWAVE. The authors would like to thank Angel <span style="font-variant:small-caps;">Duran</span>[^3] for helpful discussions on the Petviashvili method and Pavel <span style="font-variant:small-caps;">Holodoborodko</span> for providing us with the MC Toolbox for Matlab. [^1]: $^*$ Corresponding author [^2]: In our computations we just took an initial localized bump in the center of the domain and the iterative method turned it into a pattern. [^3]: University of Valladolid, Spain	{ "pile_set_name": "ArXiv" }
--- abstract: 'Determining accurate redshift distributions for very large samples of objects has become increasingly important in cosmology. We investigate the impact of extending cross-correlation based redshift distribution recovery methods to include small scale clustering information. The major concern in such work is the ability to disentangle the amplitude of the underlying redshift distribution from the influence of evolving galaxy bias. Using multiple simulations covering a variety of galaxy bias evolution scenarios, we demonstrate reliable redshift recoveries using linear clustering assumptions well into the non-linear regime for redshift distributions of narrow redshift width. Including information from intermediate physical scales balances the increased information available from clustering and the residual bias incurred from relaxing of linear constraints. We discuss how breaking a broad sample into tomographic bins can improve estimates of the redshift distribution, and present a simple bias removal technique using clustering information from the spectroscopic sample alone.' author: - \| Samuel J. Schmidt$^{1}$[^1], Brice Ménard$^{2}$[^2], Ryan Scranton$^{1}$, Christopher Morrison$^{1}$, Cameron K. McBride$^{3}$\ $^{1}$Department of Physics, University of California, One Shields Avenue, Davis, CA 95616, USA\ $^{2}$Department of Physics and Astronomy, Johns Hopkins University, 3400 N. Charles Street, Baltimore, MD, USA\ $^{3}$Harvard–-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA bibliography: - 'paper\_v3.bib' title: 'Recovering Redshift Distributions with Cross-Correlations: Pushing The Boundaries' --- large-scale structure of the Universe—cosmology: observations—methods: data analysis—methods: statistical Introduction {#intro} ============ Cosmological measurements require distance estimates in order to map the large scale structure of the universe. In the past this has most often been done on an object by object basis by obtaining spectroscopic redshifts for individual sources. Surveys of large samples of galaxies such as the Sloan Digital Sky Survey and its extensions [@York:00; @sdss] have been instrumental in improving cosmological measurements. However, a number of current and upcoming missions (DES, LSST, etc...) will attempt to measure the fundamental properties of cosmology, and particularly dark energy, using a variety of methods (e.g. weak gravitational lensing, baryon acoustic oscillations, etc.). Fundamental to all of these surveys is the assumption that the millions, or even billions, of galaxies observed by these instruments will be separable into redshift bins, despite the fact that the number of objects involved makes spectroscopic follow-up wildly impractical. Photometric redshift techniques show a good deal of promise towards this goal [e. g. @Con:95; @Ben:00; @Cun:09], but there remain questions as to whether or not they can meet the stringent requirements outlined in these surveys and avoid systematic biases that could leak into dark energy constraints [e. g. @Ma:06; @Cun:12]. In this paper, we examine a technique that uses clustering between spectroscopic and photometric samples to accurately determine a photometric sample’s redshift distribution. The applications of such a technique are much more general than the aforementioned large surveys: this technique can be used to estimate the redshift distribution of nearly any data set. Even single-band detections that lack photometric redshift estimates can be used, as long as they have reliable astrometric information for the calculation of cross-correlation functions. The technique described in this paper uses the physical associations due to large scale clustering to probe redshift distributions. Such ideas are not new: @Seld:79 cross-correlated quasars and galaxy counts to test for physical association, though they found no trend with redshift. @Rob:79 similarly cross-correlated quasars and rich galaxy clusters. More recently, @Qua:10 counted pairs of galaxies at small angular separations between photometric redshift selected samples, taking advantage of physically associated pairs of galaxies in order to determine an empirical measure of the photometric redshift errors. Similarly, @Ben:10 cross-correlated photometric redshift bins to determine the relative contamination fraction between redshift bins based on the magnitude of the induced angular cross-correlations. These previous techniques do not require any spectroscopic sample and rely solely on photometric redshift information. @Sch:06 discuss using cross-correlations of objects sorted into redshift bins in order to determine their redshift distribution. They mention that having a subset of objects with more accurately determined redshifts would enable tighter constraints than photometric redshifts alone. Expanding on this idea, @New:08 (hereafter N08) and @Matt:10 [@Matt:11] describe a technique that requires a spectroscopic sample that spans the redshift range of interest. In simple terms, the method measures the amount of overlap between the spectroscopic sample divided into redshift bins and an unknown sample (which we will refer to as the “photometric” sample, though photometric redshifts are not necessary for sample selection). As galaxies cluster on all scales, if a spectroscopic bin overlaps in redshift with the photometric sample, we expect to see an excess number of objects, whereas if there is no overlap we expect to simply see the average number of objects that overlap spatially due to projection. By measuring the strength of the spatial cross-correlations as a function of redshift we can recover the redshift distribution of the photometric sample. A major component of the N08 technique is an iterative method to correct for bias evolution that may occur in the sample. @Sch:10 implemented a very similar technique on mock data and reported difficulty in distentangling the galaxy bias from the redshift distribution, a point which we will examine in this paper. It is only very recently that this technique has been used with real data [@Mit:12; @Nik:12], including the exact technique described in this work [@Mor:12 Ménard et al, in preparation]. The N08 technique is designed to work with large scale correlations where the galaxy bias can be treated as linear. However, the increasing amount of power in galaxy correlation functions due to large scale clustering means that there is considerable signal that is not being fully utilized at smaller scales (this is particularly relevant given the small angular extent of many deep spectroscopic surveys). Further study of the spectroscopic sample’s non-linear bias properties may enable us to account for the effects of bias in the redshift recovery. In this paper we explore the impact of retaining the linear assumptions while expanding the procedure to include smaller physical scales, where the galaxy bias becomes non-linear in the density field. In this regime evolution of the galaxy bias will modulate the amplitude of the recovered redshift distribution. We test the efficacy of using the linear assumptions well into the non-linear regime. Additionally, in some instances we are only interested in the existence or absence of galaxies in a redshift interval, and the detailed shape of the redshift distribution is not the main concern. As an example, when selecting objects based on photometric redshifts, parameter degeneracies can lead to inclusion of a secondary population of objects far outside the intended redshift range. In such a case the exact shape of the redshift distribution, which can be distorted by the presence of evolving galaxy bias, may be secondary to detecting the presence or absence of an interloper population. For these reasons we examine the relative amount of information contained at a range of physical scales around our spectroscopic samples. The layout of the paper is as follows: in §\[method\] we discuss the algorithms used to determine the redshift distributions. A summary of the mock datasets is given in §\[sims\]. Results are presented in §\[results\]. We conclude and present future work in §\[future\]. Method ====== Our main goal is to measure the redshift probability density function (pdf), $\phi(z)$, for a specific sample of objects that we will refer to as the “photometric” sample. As mentioned in Section \[intro\], we do not necessarily need a photometric redshift measurement for our samples. However, as we will discuss, a sample that is selected to cover a narrow range in redshift leads to a more accurate recovery than that of a broad distribution. So, while the method can be applied to almost any arbitrary data set, in practice it will most often use samples selected with photometric redshifts. We estimate these redshift distributions by measuring the amplitude of the cross-correlation signal between our photometric sample and a sample of objects with known redshifts. The angular cross-correlation between the photometric sample and spectroscopic sample, $w_{sp}(\theta,z)$ is defined in terms of the mean density of objects an anglular distance $\theta$ from objects in the spectroscopic sample: $$\label{sigmaeqn} \langle \Sigma(\theta,z) \rangle = \Sigma_{p}(1 + w_{sp}(\theta,z))$$ where $\Sigma_{p}$ is the mean surface density of photometric objects. In practice, rather than measuring the correlation function in multiple angular bins and fitting a power law form, we measure the density of “photometric” sources in a single physical annulus around each individual spectroscopic source, from a minimum radius ([$r_{min}$]{}) to a maximum radius ([$r_{max}$]{}), measured in units of comoving $kpc$. We subdivide the spectroscopic sample into bins of redshift and measure the mean (over)density of “photometric” objects around each spectroscopic source within each redshift bin. After subtraction of the average density expected from points randomly placed in the survey geometry and normalization, this is equivalent to the Natural Estimator, $DD/RR\,-\,1$ [@Ker:00], where $DD$ is the number of cross-correlated pairs within our annulus and $RR$ is the number of correlated pairs from a dataset with randomized positions in the same survey footprint. We use the amplitude of this “one bin” estimate of the excess clustering as our estimator of $\phi(z)$. In addition to calculating the density with uniform weight within the annulus, we also also compute a density measure where we weight each object proportional to the inverse of the spatial distance from the spectroscopic object. We will compare these estimators in the Appendix. All calculations in the body of the paper will use the inverse weighted estimator. In computing the projected overdensities, proper treatment of the survey area, including complicated selection and masks, is essential. To accomplish this goal we develop code which employs the [astro-stomp]{} software package[^3]. The software uses a pixelization of the sky to encode both the galaxy positions as well as the survey footprint for fast computation of galaxy density and has the ability to encode complex masking and selection. Bias Correction {#bias} --------------- Measuring the overdensity of galaxies around objects in the spectroscopic sample does not immediately give us the underlying redshift distribution: cross-correlations measure the object overdensity within a fixed real space annulus, but the clustering length of both the spectroscopic and “photometric” samples are not necessarily constant with redshift. We must account for any such evolution in order to recover the redshift distribution. We will examine using the clustering length calculated from the spectroscopic sample alone to account for bias evolution in Section \[approx\], but will mainly use the iterative method introduced by N08. This technique describes an iterative correction to the redshift distribution using estimates of the mean clustering length for the photometric sample and the (presumably known) clustering evolution of the spectroscopic sample. Thus, the method requires that we have sufficient data to calculate the clustering evolution of the spectroscopic sample, and assumes that the bias in the photometric sample varies linearly with the bias in the spectroscopic sample. This assumption will break down as we include measurements at small radii as the clustering moves in to the non-linear regime. Finding the scales at which the bias evolution becomes too great for effective correction is the subject of this paper. As in N08 and @Matt:10 (see these references for the full derivation), we have a relation between the cross-correlation function of the spectroscopic and photometric samples and the normalized redshift distribution, $\phi(z)$ of the sample that we are attempting to estimate, given by: $$\label{eq:newman1} w_{sp}(\theta,z) = \frac{\phi(z)H(\gamma)r_{\rm{0,sp}}^{\gamma_{sp}}\theta^{1-\gamma_{sp}}D_{\rm{A}}^{1-\gamma_{sp}}}{dl/dz}$$ where $\gamma$ is the power law slope of the correlation function, $D_{\rm{A}}$ is the transverse comoving distance, $H(\gamma)=\Gamma(1/2)\Gamma[(\gamma-1)/2]/\Gamma(\gamma/2)$ and l(z) is the comoving distance to redshift z. If we assume that the correlation function is a power law of the form $w_{sp}(\theta)=A_{sp}\,\theta^{1-\gamma}$ between [$r_{min}$]{} and [$r_{max}$]{} then our one bin measurement of the overdensity is proportional to the amplitude of the cross-correlation signal, $A_{sp}$. The unknown quantities in Equation \[eq:newman1\] are $\phi(z)$ and $r_{\rm{0,sp}}^{\gamma_{sp}}$. We cannot evaluate $r_{0,sp}$ and $\gamma_{sp}$ as a function of redshift directly, as we do not know the redshift distribution of the “photometric” sample, though we can estimate them using other measured quantities. Assuming that the galaxy bias is linear, we can estimate the cross-correlation parameters from power law fits to the autocorrelation functions of the spectroscopic and photometric samples, $\gamma_{sp}=(\gamma_{ss}+\gamma_{pp})/2$ and $r_{0,sp}^{\gamma_{sp}}=(r_{0,pp}^{\gamma_{pp}}r_{0,ss}^{\gamma_{ss}})^{1/2}$, where $\gamma_{ss}$ and $r_{0,ss}$ are measured from the projected correlation function, $w_{p}(r_{p})$ of the spectroscopic sample, and $\gamma_{pp}$ is the measured power law slope of the photometric sample angular autocorrelation function. Rearranging Equation \[eq:newman1\], we find: $$\label{eq:newman2} \phi(z)=A_{sp}(z)\frac{dl/dz}{H(\gamma)r_{0,sp}^{\gamma}D_{\rm{A}}^{1-\gamma}}$$ with the estimated $r_{0,sp}$ entering in the denominator, diminishing the effect of the bias. The updated $\phi(z)$ can now be used in Equation \[eq:newman1\] to obtain an updated value for $r_{0,sp}$, and the process can be iterated until convergence is reached. Note that this iterative procedure simply estimates a single, average value for $r_{0,pp}$ and assumes that $r_{0,sp}$ scales linearly with $r_{0,ss}$ to improve the redshift recovery. If $r_{0,sp}$ is evolving in a non-linear fashion with redshift this method will not correct the redshift distribution appropriately. The shape of the redshift distribution also impacts the effectiveness of the iterative correction: as we are assuming that the clustering length of the photometric sample is proportional to that of the spectroscopic sample, the optimal iterative solution will work best at the mean redshift of the sample. For a compact and peaked redshift distribution the linear bias assumption will be a good approximation of the true bias. For a broad, or multiply peaked distribution, deviations from the linear approximation will become more problematic. If we can break the photometric sample into narrow subsets in redshift it is possible to mitigate this problem. This will be examined in Section \[tomography\]. Choice of $r_{min}$ and $r_{max}$ {#rminrmax} --------------------------------- The photometric overdensity is measured over a constant range of projected physical scales around each spectroscopic object, bounded by an inner and an outer radius, [$r_{min}$]{} and [$r_{max}$]{}. The choice of these radii affects the recovery in a number of ways and the values for [$r_{min}$]{} and [$r_{max}$]{} serve as the primary tuning parameters for the fidelity of the recovered redshift distribution. For cases where the photometric catalog contains some fraction of the spectroscopic catalog a complication can arise. Excess signal from the cross-matches between objects in the catalogs may boost the recovery signal if we allow for [$r_{min}$]{}= 0, or if astrometric uncertainties lead to mismatch of spectroscopic and photometric objects. For our simulations, we have explicity excluded such objects, so small radius matches are unaffected by such contamination. More broadly, at small physical separations (below $\sim\,1$ Mpc) the clustering of objects will become stronger and increasingly non-linear. This increased amplitude is a strong indicator that there is more information to be extracted from the cross-correlation, albeit at the cost of bending some of the linear assumptions described in §\[bias\]. We will discuss this issue further in Section \[future\]. For the outer boundary of the annulus, as [$r_{max}$]{} increases, more physically associated galaxies will be included in the annulus, but so will an increasing number of unassociated background sources. The clustering signal declines as radius increases, so increasing [$r_{max}$]{} can degrade the signal to noise ratio of the measurement. The optimum $r_{max}$ will depend on both the clustering of the sample and the density of the photometric source catalog. Simulated Galaxy Catalogs {#sims} ========================= The redshift recovery procedure is sensitive to the evolution of galaxy bias. In order to test this sensitivity we employ two sets of simulations: mocks based on Millennium light cones with a limited field of view [@Sper:05; @Cro:06] and larger area mocks based on LasDamas simulations (McBride et al, in prep). To cover a wide variety of possible scenarios we will examine four mock data sets: 1. No galaxy bias evolution. 2. Evolving galaxy bias as expected for a realistic, magnitude limited sample. 3. A magnitude limited selection with an additional stellar mass cut used to recover the magnitude limited sample, as might be expected if our spectroscopic catalog was a particular galaxy type. 4. A mixed case with constant bias at low redshift and a magnitude limited selection at higher redshift used to recover a sample with smooth bias evolution. Such a distribution might arise from multiple populations or complex selection criteria. The first two cases have spectroscopic and photometric data drawn from the same underlying distribution, and thus identical galaxy bias properties. Cases iii and iv have different bias properties for the spectroscopic and observed samples, as will be discussed in the following subsections. We use the LasDamas simulations for the constant bias (item i) and mixed evolution samples (item iv), and the Millennium simulations for the magnitude selected and stellar mass selected samples (items ii & iii ). LasDamas Based Mock Catalogs {#damas} ---------------------------- The LasDamas catalogs used in this paper are a customized galaxy data set generated from the dark matter simulations of the LasDamas project (McBride et al, in prep)[^4]. These galaxy mocks were constructed for testing this method, and do not explicitly fit to observed SDSS data, as is done in the full LasDamas simulations. This enabled us to extend the redshift range beyond $z > 1$, with samples spanning $0.03\leq z\leq 1.33$, and covering a $9\times14$ degree patch of sky. The galaxy mocks are constructed from a static redshift output of one of the four large LasDamas boxes (the Carmen simulations) at $z = 0.5$. We defined friends-of-friends halos [@fof] with a linking length of 20% of the mean inter-particle separation. We assigned mock galaxies based on a simple $3$ parameter halo occupation distribution model [i.e. HOD; @berlind:02]. To achieve the variable bias, the HOD is varied to reduce the number density as a function of redshift (thereby increasing the bias). The LasDamas simulations assume a flat [$\Lambda\mathrm{CDM}$]{} cosmology with $\Omega_{m}=0.25$, $\Omega_{\Lambda}=0.75$, $h=0.70$, and $\sigma_8 = 0.8$. We construct two spectroscopic catalogs, populating the same dark matter halo catalogs with two different types of galaxy bias applied to generate the data: - A constant bias for the whole redshift range consisting of approximately 120,000 galaxies over 125 deg$^{2}$ - A constant bias over ($0.2\,\leq\,z\,\leq\,0.77$,) then mimicking an apparent magnitude limited selection for $z>0.77$ with about 235,000 galaxies over 500 deg$^{2}$. We refer to the first as the “constant bias” sample and the second as the “mixed bias evolution” sample. For the photometric samples we create distributions drawn from the same constant bias case, as well as an additional dataset with bias is chosen such that the density decreases linearly with redshift over the range $0.2\,\leq\,z\,\leq1.33$ (referred to as the “linear density evolution” sample. For these samples we create: - A bimodal sample with galaxies in the ranges $0.4 < z < 0.6$ and $0.8 < z < 1.1$ containing about $350,000$ for the constant bias sample and about $592,000$ galaxies for the linear density evolution sample. - A Gaussian centered at $z=0.75$ and width $\sigma_{z}=0.10$ with $\sim 145,000$ for the constant bias case and $\sim 410,000$ galaxies for the linear density evolution sample. Millennium Galaxy Mock Catalogs {#millen} ------------------------------- The Millennium Simulation galaxy mock catalogs of @Cro:06[^5] are light cones populated with galaxies generated from semi-analytic models that follow the prescriptions of @Cro:06 and @Kitz:07. We use the four 2$\times$2 degree “DLS” cones, designed to match the footprints of the Deep Lens Survey [@Witt:02]. The light cones contain 17.4 million galaxies with redshifts spanning $0<z<3$ over 16 deg$^{2}$ with a magnitude limited r-band depth of $r=29.0$, which cover a redshift range large enough that significant galaxy bias evolution will occur. Areal coverage is limited enough that sample variance will be a significant factor in some measurements. The Millennium simulation assumes cosmological parameters $\Omega_{m}=0.25$, $\Omega_{\Lambda}=0.75$, $h=0.73$, and $\sigma_{8}=0.9$. We construct two “spectroscopic” catalogs with known redshifts, and two “photometric” catalogs, where no redshift information is retained. For the spectroscopic sets: - We randomly select $\sim2\%$ ( approximately 325,000 galaxies) of the magnitude limited sample that will have a galaxy bias evolution matching that of the underlying sample. - We design a galaxy sample of roughly the same size as the previous sample (335,000 galaxies) which contains all galaxies with stellar mass greater than $2.3\times10^{10}\ M_{\sun}$. We refer to the first as the “bias evolution” sample and the second as the “masscut” sample. Each has a surface density of approximately 5.6 galaxies/arcmin$^{2}$. For the photometric samples, we draw from the same underlying magnitude limited distribution used in the evolving bias scenario, thus they have identical bias evolution properties to the evolving bias case. We construct two samples from the simulation data: - A bimodal distribution with 1.9 million galaxies covering the ranges $0.9 < z < 1.1$ and $1.9 < z < 2.1$. - A Gaussian distribution of about 690,000 galaxies centered at $z = 1.5$ with width $\sigma_z = 0.15$. As will be shown in later sections, despite similar redshift spans and identical spectroscopic catalogs, the effects of evolving bias on the recovered redshift distribution for these two samples can be quite dissimilar. The stellar mass selection gives our spectroscopic sample different bias properties than the photometric samples, enabling us to test the efficacy of the recovery algorithm in the presence of stronger, non-representative bias. The evolving bias scenario has identical bias in the spectroscopic and photometric samples. The iterative procedure should perfectly recover the redshift distribution when the biases are the same, however, we will see that this is not the case when we measure the clustering length based on the large linear regime scales but use non-linear clustering information to reconstruct the distribution. Clustering Measurements {#clusterstuff} ----------------------- The recovery procedure requires fits to the projected correlation function of the spectroscopic datasets, as well as the slope and amplitude of the two point autocorrelation functions of the photometric samples. For the Millennium data sets the projected correlation functions of the spectroscopic samples are well fit by a power law form, and lack a strong 1-halo break. As the slope of the power law shows little variation, we fix it at $\gamma_{ss}=1.8$ when fitting for the correlation length, $r_{0,ss}$ and use only projected separations greater than $r_{p}>300kpc$. Thus, non-linear evolution in the one halo regime will not be reflected in the clustering lengths used in the iterative corrections. We fit a parabolic form to the $r_{0,ss}$ data to smooth small scale redshift dependence induced by the measurement errors. For the angular correlation function fits of the photometric samples, we measure best fit power law slopes of $\gamma_{pp}=1.75\pm0.13$ for bimodal and $\gamma_{pp}=1.68\pm0.11$ for the Gaussian distribution. For the LasDamas data, both the constant bias and mixed bias data sets show a strong one halo component and a break in their projected correlation functions, with slopes of $\gamma_{ss}=1.45$ for the constant bias data set and $\gamma_{ss}=1.8$ for the mixed bias case, nearly independent of redshift, at $r_{p}>300$ kpc beyond the one-halo break. Once again we fit for $r_{0,ss}$ using only this "quasi-linear” regime. ![image](Fig1.ps){width=".99\hsize"} ![image](Fig2.ps){width=".99\hsize"} Results ======= The effectiveness of cross-correlation methods in recovering redshift distributions is dependent on many factors. Some of these (e.g. spectroscopic completeness or galaxy bias evolution) are determined by the survey data itself. Others (the scale used for the cross-correlations, the redshift binning of the spectroscopic samples and weighting of the cross-correlation pairs) are nearly free parameters that can be used to tune the recovery. Of these free parameters, the choice of scale has the most significant effect on the recovery due to its direct linkage to the galaxy bias dependence of the recovery. In the N08 iterative technique, measurements are made on large enough scales (several Mpc) that the linear bias can be removed. For the purposes of our analyis, we consider a much broader range of scales, from the linear to the quasi-linear ($\sim 1$ Mpc) down to the non-linear scales of a few kiloparsecs. Our full analysis tests a wide range of scales, covering $3 \le$[$r_{min}$]{}$\le 3000$ kpc and $10 \le$[$r_{max}$]{}$\le 5000$ kpc, but for illustrative purposes we will show results only for three representative decade-width scales: $3 < r < 30$ kpc, $30 < r < 300$ kpc and $300 < r < 3000$ kpc. Before examining the redshift recoveries for these scales, a word about the smallest scales: Since we are using simulated data with perfect astrometry, we can distinguish perfectly between galaxies at scales where real data sets with noise from astrometric calibration and atmospheric blurring would likely struggle. Applying these techniques to real data on those scales would likely mean that cross-contamination between the spectroscopic and photometric samples would dominate the recovered signal. We have experimented with differing levels of cross-contamination between our simulated samples and find that the behavior of the recovered distributions is highly dependent on the choice of simulated spectroscopic sample. To avoid this additional complication, we have chosen to eliminate all spectroscopic objects from our photometric catalogs and vice versa and defer further exploration of this effect to a future publication. We estimate errors on the redshift distributions with a spatial jackknife. This consists of subdividing the sample into $N$ contiguous regions of the sky, each with approximately equal area. We then perform each measurement $N$ times, each time leaving out one region. We then estimate the jackknife variance as: $$Var(x)=\frac{N-1}{N}\sum_{i=1}^{N} ({x_{i}-\bar{x}})^{2}$$ where $x_{i}$ is the measurement for the ith region and $\bar{x}$ is the mean for the entire sample. Recovery Scales and Populations {#scales} ------------------------------- In practial terms, several different bias scenarios may arise depending on the type of data selected. We might select a population with very little expected bias evolution (e. g. Luminous Red Galaxies), slowly evolving bias over a broad redshift interval (e. g. field galaxies), or complex evolution due to the presence of multiple populations (e. g. a tomographic redshift bin with outliers). For this reason we study the redshift reconstruction in several bias scenarios. The shape of the redshift distribution of the photometric sample also plays a role: even a sample with strong galaxy bias evolution will not show significant relative bias change if the redshift interval of the recovery is sufficiently narrow. Conversely, even slight bias evolution over a broad redshift interval may become significant when we include additional information from non-linear scales. We test the recovery algorithm on two types of distributions to explore these effects: a centrally peaked Gaussian distribution and a bimodal distribution. Figure \[Gauss\_plot\] shows the recovered redshift distributions of the Gaussian photometric samples for both the LasDamas constant bias scenario and the Millennium evolving bias at our three representative scales. Red points show the distribution before the iterative correction of Equations \[eq:newman1\] and \[eq:newman2\] is applied, while black points show the results after the correction. In the constant bias case the iterative correction should have almost no effect, which is seen in the small difference between the pre- and post-iteration recoveries. Interestingly, the method performs extremely well in the absence of bias evolution, down to the smallest scales, and including scales that span the break in the LasDamas correlation function. The centrally peaked Gaussian distribution, with most galaxies close to the mean redshift where our $r_{0,pp}$ estimate is most accurate, shows little sensitivity to effects of the evolving bias. More compact and symmetric photometric distributions will be less affected by galaxy bias evolution, which enables us to push the recovery to smaller scales. We will discuss this further in Sections \[tomography\] and \[purity\]. For the constant bias scenario the best fit is $50 \le r \le 100$ kpc with a reduced $\chi^{2}=1.42$, though the change in $\chi^{2}$ is not particularly sensitive to the exact values of [$r_{min}$]{} and [$r_{max}$]{}(i. e. the likelihood surface is very flat). For the evolving bias the best fit occurs for $100 \le r \le 300$ kpc with a reduced $\chi^{2}=0.107$. The $\chi^{2}$ value significantly below 1.0 shows that we are overestimating our error bars for the Millennium sample, which is not unexpected: with only 16 square degrees available in the Millennium light cones we use only 32 jackknife samples to estimate the errors on fits for 99 bins. This appears to only affect the amplitude of the errors, and not the overall structure of the covariance matrix. The centrally peaked distribution shows little sensitivity to bias evolution even down to the smallest scales probed, and we accurately recover the distribution at all scales. Figure \[bimodal\_plot\] shows the recovered redshift distributions of the bimodal samples for the constant bias and evolving bias spectroscopic samples. The best fit for the bimodal sample occurs at $10 \le r \le 50$ kpc with a reduced $\chi^{2}$ of 1.16. The $\chi^{2}$ values are similar for the bimodal and Gaussian distributions, showing that in the absence of bias evolution the recovery performs accurately regardless of the shape of the redshift distribution. The most notable feature in the evolving bias scenario is the relative amplitude of the two peaks. The bimodal sample shows a clear bias before the iterative correction is applied, with larger discrepancies as the annulus moves to smaller physical scales. This is as expected, since this bimodal configuration is particularly sensitive to bias evolution. Because the iterative correction estimates a single, average value of $r_{0,pp}$ for the sample the bias correction is most accurate near the mean redshift of the photometric distribution. The bimodal sample has a mean redshift of $z=1.59$, between the two peaks where no galaxies are located. Also of note is the fact that even with identical bias evolution we introduce error into the recovered distributions even after iterative correction. This is mainly due to the fact that we (purposely) measure the clustering length using only large ($>300\,kpc$) scales, and to a lesser extent due to the empirical estimation of the clustering length with finite samples that can also introduce errors. We note, however, that the iterative technique does accurately recover the distributions when only the large scale clustering information is used, as expected. The effectiveness of the iteration in correcting for the bias is obviously reduced as the radius of the annulus decreases, though errors due to covariance between bins also increase as [$r_{max}$]{} grows and more unassociated galaxies are included in the estimate. The best fit values are for intermediate scales, with a minumum at $200 \le r \le 300$ kpc and $\chi^{2}=0.334$. The best fits at intermediate scales balance the increasing influence of the bias at small scales with the concurrent increase in signal to noise and decreased bin to bin covariance that comes with smaller physical apertures. It is clear that small scale information greatly increases bias in the recovery, and should not be used to recover broad redshift distributions. Table \[chi\_table\] lists the reduced $\chi^{2}$ values for the representative scale distributions both before and after the N08 iteration is applied. The success of the iterative technique in aiding the recovery procedure is varied. The iteration improves the recovery for every case in the evolving bias scenario, mixed results in the constant bias and mass cut scenarios, and mainly degrades results in the mixed/linear case. Annulus $\chi^{2}$/$N_{D}$ pre-iteration $\chi^{2}$/$N_{D}$ post-iteration $\chi^{2}$/$N_{D}$ pre-iteration $\chi^{2}$/$N_{D}$ post-iteration -------------------------- ---------------------------------- ----------------------------------- ---------------------------------- ----------------------------------- [LD Constant Bias]{} 3-30 kpc 3.34 3.11 1.55 1.59 30-300 kpc 1.99 1.47 2.12 3.33 300-3000 kpc 1.86 1.64 1.35 2.06 3-3000 kpc 1.97 1.79 1.43 2.61 [Evolving Bias]{} 3-30 kpc 1.12 0.67 0.274 0.201 30-300 kpc 0.566 0.258 0.229 0.148 300-3000 kpc 0.400 0.216 0.264 0.215 3-3000 kpc 0.496 0.270 0.237 0.207 [Mass Cut]{} 3-30 kpc 7.11 7.20 0.279 0.276 30-300 kpc 1.99 2.04 0.156 0.155 300-3000 kpc 0.381 0.376 0.215 0.204 3-3000 kpc 0.743 0.783 0.514 0.521 [LD Mixed Bias]{} 3-30 kpc 228.2 513.1 366.08 122.8 30-300 kpc 10.99 27.90 3.26 4.92 300-3000 kpc 5.98 10.01 2.05 1.78 3-3000 kpc 131.6 340.3 20.71 34.44 \[chi\_table\] For another more quantitative measure of the fidelity of the redshift recovery, we calculate the sample mean (average redshift) and standard deviation (square root of the sample variance, i. e. the “sample width”) for each distribution. Figure \[all\_moments\] shows the deviation from the true redshift mean $\bar{z}_{tr}$ and true sample width $\sigma_{tr}$ for the two cases shown in Figures \[Gauss\_plot\] and \[bimodal\_plot\]. We show the three decade width bins and also results using information encompassing all three scales, with $3 \le r \le 3000$ kpc as a gray shaded ellipse. As the information in each of the three annuli is independent, combining all scales should provide a higher signal-to-noise measurement of the redshift distribution. The top two panels in Figure \[all\_moments\] show that with little or no expected bias evolution using all scales works extremely well at recovering the photometric sample distributions. For the evolving data set, the presence of even modest bias evolution results in a misestimation of the mean redshift for the bimodel distribution, due to the relative amplitudes of the recovered peaks. However, the sample width is largely unaffected, as the lack of cross-correlation signal outside of the two bimodal bins provides a strong constraint on $\sigma$. In the centrally peaked Gaussian distribution the mean is more accurately recovered, but the uncertainty in the sample width is increased. This is due to the fact that the tails of the distribution are now more affected by the difference in bias at low and high redshift. ![Measured deviation from the true redshift mean and width of the bimodal (left) and Gaussian (right) distributions for all four spectroscopic data sets. The truth is shown as the black dot, $3-30$ kpc (red), $30-300$ kpc (green), $300-3000$ kpc (blue), and $3-3000$ kpc (gray shaded) are shown for comparison.\[all\_moments\]](Fig3.ps){width=".99\hsize"} Unlike the top two rows, the bottom of Figure \[all\_moments\] use simulated spectroscopic samples with very different bias profiles from their respective photometric samples. The stellar mass cut sample has stronger bias evolution than the magnitude limited sample that comprises the observed sample. The effects on the bimodel distribution in the evolving bias scenario are exacerbated by the stronger bias in the mass cut sample. Once again the mean redshift for the bimodal sample is significantly skewed. To illustrate the differing bias of the three LasDamas samples we calculate the linear bias explicitly and show them in Figure \[biasfig\]. The mixed/linear bias case uses the “mixed” bias data for the spectrscopic sample and the linear density sample for the photometric sample. In this case, the normal tendency of the method to over-estimate signal at lower redshifts is counteracted by the more rapid bias evolution in the spectroscopic sample, resulting in a mean recovered redshift near the expected value for all scales. However, the difference in bias on the two sides of the Gaussian distribution results in greatly increased scatter in the recovered width. Note that we did not have access to this information when computing the recovered distributions, in fact the bias is never explicitly calculated in the N08 iteration. Instead, the clustering length of the spectroscopic sample, empirically measured from the correlation functions, is used to iteratively determine the best value for the photometric sample clustering length. ![Linear galaxy bias as a function of redshift for the three LasDamas samples described in Section \[damas\]. Red indicates the constant bias sample, green the sample with linear density evolution, and blue the sample with “mixed” bias evolution. .\[biasfig\]](Fig4.ps){width=".99\hsize"} Overall, we see several trends: - [In absence of bias evolution, recovery S/N is always highest at the smallest scales.]{} - [In the presence of modest bias evolution, intermediate scales ($\sim100 < r < 500$ kpc) offer the most reliable, highest S/N recovery.]{} - [For extreme bias evolution, larger scales ($1000 < r < 3000$) offer the cleanest recovery.]{} - [For all cases, a centrally peaked redshift distribution is far less sensitive to bias evolution, although outliers can affect the recovered distribution width.]{} - [Small scale information should not be used to recover broad redshift distributions when the bias is known to evolve.]{} Tomographic Binning {#tomography} ------------------- The previous section discussed the recovery of broad redshift distributions. However, most upcoming surveys will focus on determining the redshift distribution for narrow tomographic redshift bins for the purposes of measuring weak lensing and baryonic acoustic oscillations. The main limitation of the iterative method is that it relies on a single estimated value for $r_{0,pp}$. If the clustering length of the photometric sample evolves differently than the spectroscopic sample, then the assumption that $r_{0,sp}$ scales as $r_{0,sp}^{\gamma_{sp}}=(r_{0,pp}^{\gamma_{pp}}r_{0,ss}^{\gamma_{ss}})^{1/2}$ will not hold. The iterative method essentially finds the best single value for $r_{0,pp}$ given the data. However, if we can further subdivide our photometric sample in redshift, e. g. with some photometric redshift algorithm or color selection, we benefit in several ways: First, we may now determine a best fit $r_{0,pp}$ over a smaller redshift range for each subsample, over which the bias presumably evolves less. Second, having two values of $r_{0,pp}$ to estimate gives an additional free parameter. Third, the signal to noise of the measurement increases, as by breaking our initial photometric dataset into multiple samples in redshift, we have removed a large number of physically unassociated galaxies from the correlation measurement that were adding to the background and diluting the signal. Figure \[millen\_tomog\] shows the result of splitting the bimodal sample for the evolving bias dataset (the same as shown in the top right panel of Figure \[bimodal\_plot\]) into two distinct redshift bins and computing the recovered distributions for each individual bin. This is done for [$r_{min}$]{} and [$r_{max}$]{} values of 3 kpc and 30 kpc, far into the non-linear regime and at much smaller scales than the best fit in the previous section. The red points show the recovery for the low redshift bin, while the blue points show the recovery for the high redshift bin (overlapping points have been omitted for clarity). The top panel shows the results of the same reconstruction using a single photometric sample. The change in the size of the errors on the two individual recoveries is related to the normalization factor enabled by the two independent estimates of the best fit clustering length, which boosts/lowers the signal-to-noise of the low/high redshift recoveries. In general, the more narrow the redshift range can be restricted, the smaller the optimum recovery scale will be, which results in an increase in S/N. However, the presence of catastrophic outliers in certain photometric redshift ranges may be a concern when applying tomographic selections. We address this in the following Section. ![Top: Recovered redshift distribution for the bimodal Millennium light cone sample for an annulus of $3-30kpc$. Bottom: the same sample split into two redshift bins (overlapping points omitted for clarity). The bottom panel shows that the amplitudes of the bimodal recovered low redshift (red) and high redshift (blue) samples are significantly less biased than the union of the two samples recovered at once.\[millen\_tomog\]](Fig5.ps){width=".99\hsize"} Redshift Outlier Detection {#purity} -------------------------- Selection of tomographic redshift bins in cosmological analyses, for instance with color cuts or photometric redshift cuts, can include data sets where degeneracies exist that include an unrelated population far outside the intended redshift range. The most prominent example in optical photometric surveys is the common Lyman/Balmer break degeneracy, where low redshift ($z \sim 0.2-0.3$) blue galaxies are mistaken for very high redshift ($z \sim 2-3$) blue galaxies and vice-versa, due to their similar optical colors. Such “catastrophic outlier” populations often result in two bimodal peaks widely separated in redshift, which we have shown (§\[scales\]) can be problematic for accurate redshift recovery. However, we can use small scale information to diagnose the presence of such outlier populations. In nearly all cases examined previously, reconstructing the redshift distribution at smaller physical scales results in smaller uncertainties, albeit at the expense of increased bias sensitivity. This is expected, given the power law form of the correlation function we expect more signal on smaller scales. We also note that the method does an excellent job at returning a null signal in areas where there is no overlap between the spectroscopic and photometric samples, e. g. we see signal consistent with zero and small error bars outside the bimodal bins in Figure \[bimodal\_plot\]. We can use these features to test for the presence of interlopers, as done in @Mor:12, where the authors cross-correlate a high redshift luminous blue galaxy sample with spectroscopically confirmed galaxies to test for the presence of intermediate redshift elliptical galaxies with similar expected colors. We construct several data sets to test the influence of recovery scales on sensitivity to outlier populations. Using the LasDamas mixed bias evolution data set, we construct samples where we have a primary peak at the redshift of interest ($0.4\le z \le 0.6$) and a secondary peak due to color degeneracies ($0.8\,\leq\,z\,\leq\,1.0$) that contains between 0.5% and 10% of the total number of galaxies. Figure \[contamfig\] shows detection significance (in terms of $\sigma$ determined from the $\chi^{2}$) as a function of contamination fraction. The inset shows one recovered distribution as an example, with 10% contamination and using an annulus of 30 $\le r \le$ 300 kpc. Using the $300 \le r \le 3000$ kpc annulus we cannot reliably detect the secondary peak, however at smaller scales we see nearly all bins outside of the two peaks consistent with zero, and clearly detect non-zero signal in the range $0.8\,\leq\,z\,\leq\,1.0$ for contamination fractions above 2%. The ability to detect secondary peaks will depend on both the redshift evolution of the bias and the amount of separation between the two peaks in redshift space. The influence of the bias evolution when using small scale information can cause us to misestimate the overall contamination fraction, though detection of any contaminants at all may be the goal. While the recovery method does not directly inform us of which galaxies are degenerate, we can use the method to tailor photometric redshift cuts that lead to maximum purity in the sample by testing variations of the cuts and choosing those that minimize sample contamination. ![Detection significance of the secondary peak as a function of contamination fraction using the LasDamas mixed bias data set for three annuli. The inset shows an example recovery with 10% of the galaxies in the second peak. The higher S/N per bin for smaller annuli enables us to detect contaminating objects at much greater significance than when using large scale information alone. \[contamfig\]](Fig6.ps){width=".99\hsize"} Alternative Bias Removal Technique {#approx} ---------------------------------- The application of the full iterative procedure discussed so far requires calculation of the photometric sample angular autocorrelation functions. In actual surveys, complex selection and masking often make estimation of the correlation functions difficult. We can simplify our analysis by, instead of assuming a linear relation between the spectroscopic and photometric samples, assume that the two samples have the same bias as calculated from $r_{0,ss}$ (or measurements from the literature). We take the estimates of $r_{0,ss}$ estimated for $r_{p}\,>\,300kpc$ discussed in in Section \[millen\] and calculate the bias evolution of the spectroscopic sample as a function of redshift. In place of the full iterative procedure, we then simply divide our initial estimate of $\phi(z)$ by this relative bias and renormalize. The top panel of Figure \[approxbias\] shows a comparison between the initial estimate (black), the final iterative correction (red), and this alternative bias removal (blue) for the Millennium light cone simulation with [$r_{min}$]{}$\,=\,30kpc$ and [$r_{max}$]{}$\,=\,300kpc$ (though the conclusions hold at both smaller and larger scales as well). The simple bias correction actually outperforms the iterative solution, with a $\chi^{2}=0.40$, compared to $\chi^{2}=0.61$ for the iterative method. In retrospect, this is not unexpected: The photometric samples from the Millennium simulation used in Figure \[approxbias\] were drawn from the same underlying population as the spectrocopic sample, and thus have the same galaxy bias properties. The linear bias approximation used in the iterative correction, calculating the correlation length $r_{0,sp}$ as the geometric mean of the spectroscopic correlation length and a single, constant value for the average photometric correlation length, actually lessens the predicted redshift evolution of the bias, particularly when using very wide redshift baseline for the photometric sample. This is related to the improvements gained from splitting the sample into subsets in redshift, where we gain both in a smaller relative evolution in bias over the shorter redshift interval, and in the ability to estimate multiple values of $r_{0,pp}$ in the different redshift intervals. Applying a stellar mass selection to the spectroscopic sample will change the galaxy bias evolution properties. The bottom panel of Figure \[approxbias\] shows the recovery of the same photometric sample using the mass selected spectroscopic sample and corresponding bias estimate. While the difference is less pronounced, using the spectroscopic bias to correct the amplitude again outperforms the iterative method, with $\chi^{2}=3.4$ versus the $\chi^{2}=5.1$ for the iterative method. In practical terms, removing the bias evolution of the estimated sample using an estimate based solely on the spectroscopic sample can provide results competitive or better than those obtained from the full iterative technique. ![Comparing the iterative and approximate bias removal techniques for 30 $\le r \le$ 300 kpc. The top panel shows the recovered distributions for the Millennium sample with evolving bias using the full iterative technique (red), and using only the spectroscopic bias (blue). The bottom panel shows the same recovery using the stellar mass selected spectroscopic sample. The simple bias approximation can provide as good or better estimates of the redshift distribution when the bias evolution of the two samples is similar. \[approxbias\]](Fig7.ps){width=".99\hsize"} Discussion and Future Work {#future} ========================== In this paper we have presented a study of a simple but powerful redshift recovery algorithm applied to realistic mock datasets, testing the inclusion of information from the non-linear clustering regime. Our galaxy density estimator is equivalent to a one bin measurement of the cross-correlation function between user adjustable physical scales of [$r_{min}$]{} and [$r_{max}$]{}. We have shown that non-linear scales contain a wealth of information that can be exploited to increase S/N in the determination of redshift distributions compared to using only large scale information, and that the iterative technique used to mitigate the effects of bias evolution works well beyond the linear regime used to date for narrow distributions. Due to the wide variety of bias scenarios that may be present in real data we are limited to somewhat qualitative assessments. However, these general conclusions are informative in future applications of the method. We successfully recover the redshift distributions for several evolving and non-evolving galaxy bias configurations. However, the non-linear biasing does incur increasing amounts of error as we push to smaller and smaller radii. The optimum scale depends on the details of the photometric dataset, both in terms of bias properties and extent in redshift: narrow redshift distributions and those with little expected bias evolution can exploit clustering signal well into the non-linear regime, while broad redshift distributions or complex galaxy bias should be restricted to the more conservative limits at larger scales. Furthermore, in our one bin treatment, larger values of [$r_{max}$]{} lead to increasing covariance between redshift recovery bins, which increases the relative error when using large scales. One must find the balance between the increased signal-to-noise and the accompanying increased sensitivity to galaxy bias when performing the recovery. The iterative correction suggested by N08 and @Matt:10 and employed in this paper has limitations. The assumption that the bias of the photometric sample scales linearly with the spectroscopic sample allows us to determine only a single value for the average cluster scaling between the spectroscopic and photometric samples via equation \[eq:newman2\]. The technique works well at correcting for galaxy bias when used for large scales, but begins to fail, as expected, as the non-linear information is included. We explored using an approximation of simply dividing by the bias of the spectroscopic sample in section \[approx\], and found that this works well in many cases, though the same caveats that apply to the use of the iterative corrections apply. The iterative correction works best at the mean redshift of the photometric sample, thus narrow redshift distributions peaked near the mean redshift are recovered much more accurately than broad distributions, as illustrated by the relative performance of the Gaussian and bimodal samples shown in Figures \[Gauss\_plot\] and \[bimodal\_plot\]. If we are able to subdivide the distribution we wish to recover into narrower redshift ranges then we can recover the distribution more accurately, as the bias should evolve less over the smaller redshift interval (assuming a smoothly varying bias evolution). This was illustrated in the simple example of breaking one of our bimodal samples into two bins in Section\[tomography\]. This is in line with the direction of the large future surveys (DES, LSST, etc...), where the strategy for determining cosmological parameters hinges on precisely determining the redshift distribution for a number of relatively narrow tomographic photometric redshift bins. The tomographic bins planned for such surveys are ideal samples for including non-linear information in redshift recovery. However, extra care will have to be taken if bins include any “catastrophic outlier” galaxies, where photometric redshift degeneracies cause some portion of the sample to lie at very different redshifts than that targeted by the selection. Such distributions will be susceptible to biasing, particularly when including information from non-linear scales. In such cases, reverting to large [$r_{min}$]{} values may be necessary. Even in such cases, the non-linear regime can be used to accurately assess the presence or absence of catastropic outliers in the sample, as illustrated in the tests of sample contamination discussed in section \[purity\]. We plan to carry out tests on more realistic tomographic photometric redshift bins based on improved simulations in an upcoming paper. The method could be further improved by extending beyond the simple one-bin treatment used in this paper, particularly in cases where there is obvious non-power law form to the correlation functions, or where the slope of the power law changes substantially. For example, in the determination of $r_{0,ss}$ we used only information at scales greater than 300 kpc, beyond the break in the correlation function, even when testing the recovery at the smallest scales. An explicit fit to both the one halo and two halo portions of the correlation function would enable a more precise recovery. However, the non-linear relation between galaxies and underlying dark matter at small scales will still leave the method susceptible to the influence of galaxy bias evolution. Having shown that the methods discussed in this paper can accurately recover redshift distributions using small scale clustering information, we will follow up with analyses using real data sets for both known and wholly novel redshift distributions (Ménard et. al, in preparation). This powerful technique will be an important and useful tool for both current and future photometric surveys. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by NSF Grant AST-1009514. Brice M´enard is supported by the NSF and the Alfred P. Sloan foundation. We thank the anonymous referee for suggestions that improved the content of the paper. We used mock catalogs based on the LasDamas project; we thank the LasDamas collaboration for providing us with this data. We thank Darren Croton for making the Millennium Simulation light cones used in this work publicly available. Additional Recovery Parameters {#recov_params} ============================== While the choice of the physical annulus defined by [$r_{min}$]{} and [$r_{max}$]{} is the dominant factor in determining the redshift recovery, we have the freedom to choose both an additional radial weighting of our pixelized aperture and the bin width of the spectroscopic sample. Annulus Weighting {#weight} ----------------- The power law form of the correlation function shows that there is an increasing amount of clustering information at smaller angular and physical scales, however there are also fewer galaxies due to the decreasing area of the annulus. Similarly, while larger apertures decrease shot noise in density estimates, they also increase the number of unassociated galaxies that are included due solely to line of sight projection. In addition to a “uniform” density estimator, where we simply divide the number of galaxies within the pixelized annulus by the area in physical units of Mpc, we test an “inverse” weighted density estimate, calculating the density in the pixelized annulus and weighting the density in each pixel by the inverse of the distance from the spectroscopic object. Errors are a combination of variations due to large-scale-structure (which becomes a more serious issue for surveys with small areal footprint), and Poisson shot noise. We estimate errors on the recovery empirically with a spatial jackknife, which captures both sources of error. For illustration, Figure \[weight\_bimod\] shows the uniform vs. inverse weighting recovered redshift distributions of the constant bias bimodal sample with a large outer radius of [$r_{max}$]{}=3000kpc. There is a clear reduction of error when using the inverse weighting, which is observed at nearly all scales and in all samples tested, thus we employ this inverse weighted estimator throughout the paper. However, because the smaller scales are now weighted more heavily, this estimator is more sensitive to evolving galaxy bias. Therefore, caution should be used when using the inverse weighting at very small scales when galaxy bias evolution is known to be large. ![Recovered redshift distribution for bimodal sample of galaxies for the constant bias sample for the “uniform” density weight (left) and “inverse” density weight (right). The magenta histogram shows the actual redshift distribution of the photometric sample. The inverse weighting produces smaller error estimates, but is more sensitive to the effects of non-linear bias evolution.\[weight\_bimod\]](FigA1.ps){width=".99\hsize"} Redshift Binning {#binning} ---------------- The choice of binning for the spectroscopic sample is an additional free parameter that we must choose. To construct our redshift distribution we take each spectroscopic galaxy and estimate the density of sources within the physical aperture defined by [$r_{min}$]{} and [$r_{max}$]{}. Then, we bin all spectroscopic objects within a redshift interval $\Delta\,z$ and take the mean of the density estimates within each bin to determine the amplitude of the redshift distribution estimate. Several factors influence the uncertainties resulting from a specific choice of redshift binning. Errors are a combination of Poisson fluctuations, i. e. the number of spectrocopic galaxies included in the bin, sample variance, and the fractional error in the amplitude of the cross-correlation function. The sample variance is fixed by large scale structure and the areal coverage of the survey. The amplitude of the cross-correlation function depends on the width of the redshift bin, as using broader redshift bins lowers the amplitude of the cross-correlation signal. Narrow redshift bins lead to a stronger cross correlation signal, however this must be balanced with Poisson noise from small samples within the bin. In practice, the total signal-to-noise is not a strong function of bin-width choice for small bins. However, the total signal-to-noise is significantly lower when using a small number of very broad bins. Using a small number of bins effectively throws out information unnecessarily. Figure \[binsize\] shows the jackknife error estimates for several bin size choices using the LasDamas based mock dataset with no bias evolution and $30 \le r \le 300$ kpc. For the constant bias sample used in Figure \[binsize\] the optimal scale occurs at $\Delta\,z\approx\,0.005$ with $\sim200-800$ galaxies per redshift bin used to determine the mean density. We expect adjacent bins to be increasingly correlated as $\Delta\,z$ decreases, as shared large scale structure near the bin boundaries should become more important. All bins are correlated with each other, as expected, since the density estimate of background galaxies samples the distribution over the entire projected redshift range, with many galaxies falling within the physical annulus surrounding a spectroscopic object multiple times. This leads to a distinct correlation matrix structure: a strong diagonal and all off diagonal elements correlated at a similar “floor” level, the amplitude of which is determined by the size of the annulus and the width of the recovery. A redshift bin of width $\Delta\,z=0.005$ corresponds to $\sim10-20$Mpc in comoving distance for $0.2\leq\,z\,\leq\,1.33$ probed in the recovery, much larger than the weighted physical distance, so the correlation matrix shows a strong diagonal component, but adjacent bins do not show excess correlation compared to widely separated bins. The projected nature of the measurement leads to highly correlated bins, and the full covariance matrix must be used for proper error estimation. ![The effect of spectroscopic bin size on the recovered redshift distribution in a data set with no bias evolution. Errors are a combination of large scale structure fluctuations and Poisson noise in the average density estimate. The effect of Poisson fluctuations can be seen as the number of galaxies per bin decreases at $\Delta\,z=0.005$. \[binsize\]](FigA2.ps){width=".99\hsize"} [^1]: Email: [email protected] [^2]: Alfred P. Sloan fellow [^3]: available at: http://code.google.com/p/astro-stomp/ [^4]: We note that these mocks are [not]{} part of the “publicly available” mocks accessible from the LasDamas website [ `http://lss.phy.vanderbilt.edu/lasdamas/`]{} [^5]: Available at:\ [`http://web.me.com/darrencroton/Homepage/SDSS-DEEP2.html`]{}	{ "pile_set_name": "ArXiv" }
--- abstract: \| We propose a second order differential calculus to analyze the regularity and the stability properties of the distribution semigroup associated with McKean-Vlasov diffusions. This methodology provides second order Taylor type expansions with remainder for both the evolution semigroup as well as the stochastic flow associated with this class of nonlinear diffusions. Bismut-Elworthy-Li formulae for the gradient and the Hessian of the integro-differential operators associated with these expansions are also presented. The article also provides explicit Dyson-Phillips expansions and a refined analysis of the norm of these integro-differential operators. Under some natural and easily verifiable regularity conditions we derive a series of exponential decays inequalities with respect to the time horizon. We illustrate the impact of these results with a second order extension of the Alekseev-Gröbner lemma to nonlinear measure valued semigroups and interacting diffusion flows. This second order perturbation analysis provides direct proofs of several uniform propagation of chaos properties w.r.t. the time parameter, including bias, fluctuation error estimate as well as exponential concentration inequalities.\ Keywords : Nonlinear diffusions, mean field particle systems, variational equations, logarithmic norms, gradient flows, Taylor expansions, contraction inequalities, Wasserstein distance, Bismut-Elworthy-Li formulae. Mathematics Subject Classification : 65C35, 82C80, 58J65, 47J20. author: - 'M. Arnaudon' - 'P. Del Moral' title: 'A second order analysis of McKean-Vlasov semigroups' --- Introduction ============ Description of the models {#ref-desciption} ------------------------- For any $n\geq 1$ we let $P_n({\mathbb{R}}^d)$ be the convex set of probability measures $\eta,\mu$ on ${\mathbb{R}}^d$ with absolute $n$-th moment and equipped with the Wasserstein distance of order $n$ denoted by $\WW_n(\eta,\mu)$. Also let $b_t(x_1,x_2)$ be some Lipschitz function from ${\mathbb{R}}^{2d}$ into ${\mathbb{R}}^d$ and let $W_t$ be an $d$-dimensional Brownian motion defined on some filtered probability space $(\Omega,({\mathbb{F}}_t)_{t\geq 0},{\mathbb{P}})$. We also consider the Hilbert space ${\mathbb{H}}_t({\mathbb{R}}^d):={\mathbb{L}}_2((\Omega,{\mathbb{F}}_t,{\mathbb{P}}),{\mathbb{R}}^d)$ equipped with the ${\mathbb{L}}_2$ inner product $\langle{\mbox{\LARGE .}},{\mbox{\LARGE .}}\rangle_{{\mathbb{H}}_t({\mathbb{R}}^d)}$. Up to a probability space enlargement there is no loss of generality to assume that ${\mathbb{H}}_t({\mathbb{R}}^d)$ contains square integrable ${\mathbb{R}}^d$-valued variables independent of the Brownian motion. For any $\mu\in P_2({\mathbb{R}}^d)$ and any time horizon $s\geq 0$ we denote by $X_{s,t}^{\mu}(x)$ the stochastic flow defined for any $t\in [s,\infty[$ and any starting point $x\in {\mathbb{R}}^d$ by the McKean-Vlasov diffusion $$\label{diff-st-ref-general} dX^{\mu}_{s,t}(x)=b_t\left(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu)\right)~dt+ dW_t\quad \mbox{\rm with}\quad b_t\left(x,\mu\right):=\int~\mu(dy)~b_t(x,y)$$ In the above display, $\phi_{s,t}$ stands for the evolution semigroup on $P_2({\mathbb{R}}^d)$ defined by the formulae $$\phi_{s,t}(\mu)(dy)=\mu P^{\mu}_{s,t}(dy):=\int~\mu(dx)~P^{\mu}_{s,t}(x,dy)\quad\mbox{\rm with}\quad P^{\mu}_{s,t}(x,dy):={\mathbb{P}}(X^{\mu}_{s,t}(x)\in dy)$$ [We denote by $L_{t,\phi_{s,t}(\mu)}$ the generator of the stochastic flow $X^{\mu}_{s,t}(x)$. ]{} The existence of the stochastic flow $X^{\mu}_{s,t}(x)$ is ensured by the Lipschitz property of the drift function see for instance [@graham; @huang]. To analyze the smoothness of the semigroup $\phi_{s,t}$ we need to strengthen this condition. We shall assume that the function $b_t(x_1,x_2)$ is differentiable at any order with uniformly bounded derivatives. In addition, the partial differential matrices w.r.t. the first and the second coordinate are uniformly bounded; that is for any $i=1,2$ we have $$\label{def-H} \Vert b^{[i]}\Vert_2:=\sup_{t\geq 0}~\sup_{(x_1,x_2)\in{\mathbb{R}}^{2d}}\Vert b_t^{[i]}(x_1,x_2)\Vert_2<\infty\quad\mbox{\rm with}\quad b^{[i]}_t(x_1,x_2):=\nabla_{x_i}b_t (x_1,x_2)$$ In the above display, $\Vert A\Vert_{2}:=\lambda_{\tiny max}(AA^{\prime})^{1/2}$ stands for the spectral norm of some matrix $A$, where $A^{\prime}$ stands for the transpose of $A$, $\lambda_{\tiny max}({\mbox{\LARGE .}})$ and $\lambda_{\tiny min}({\mbox{\LARGE .}})$ the maximal and minimal eigenvalue. In the further development of the article, we shall also denote by $A_{\tiny sym}=(A+A^{\prime})/2$ the symmetric part of a matrix $A$. [ In the further development of the article we represent the gradient of a real valued function as a column vector, or equivalently as the transpose of the differential-Jacobian operator which is, as any cotangent vector, represented by a row vector. The gradient and the Hessian of a column vector valued function as tensors of type $(1,1)$ and $(2,1)$, see for instance (\[grad-def\]). ]{} The mean field particle interpretation of the nonlinear diffusion (\[diff-st-ref-general\]) is described by a system of $N$-interacting diffusions $\xi_t=(\xi^i_t)_{1\leq i\leq N}$ defined by the stochastic differential equations $$\label{diff-st-ref-general-mf} d\xi_t^i=b_t(\xi^i_t,m(\xi_t))~dt+ dW^i_t\quad \mbox{\rm with}\quad 1\leq i\leq N \quad \mbox{\rm and}\quad m(\xi_t):=\frac{1}{N}\sum_{1\leq j\leq N}\delta_{\xi^i_t}$$ In the above display, $\xi_0^i$ stands for $N$ independent random variables $\xi_0^i$ with common distribution $\mu_0$, and $W^i_t$ are $N$ independent copie of the Brownian motion $W_t$. McKean-Vlasov diffusions and their mean field type particle interpretations arise in a variety of application domains, including in porous media and granular flows [@bene-1; @bene-2; @cattiaux; @toscani], fluid mechanics [@mckean-1; @mckean-2; @otto; @villani], data assimilation [@Bishop/DelMoral:2016; @d-2013; @DelMoral/Tugaut:2016], and more recently in mean field game theory [@benoussan; @cardaliaguet; @peng-17; @carmona-delarue; @carmona-delarue-2; @carmona-delarue-3; @huang-0; @gueant], and many others. The origins of this subject certainly go back to the beginning of the 1950s with the article by Harris and Kahn [@harris] using mean field type splitting techniques for estimating particle transmission energies. We also refer to the pioneering article by Kac [@kac-1; @kac-2] on particle interpretations of Boltzmann and Vlasov equations, and the seminal articles by McKean [@mckean-1; @mckean-2] on mean field particle interpretations of nonlinear parabolic equations arising in fluid mechanics. Since this period, the analysis of this class of mean field type nonlinear diffusions and their discrete time versions have been developed in various directions. For a survey on these developments we refer to [@carmona-delarue; @d-2013; @sznitman], and the references therein. The McKean-Vlasov diffusions discussed in this article belong to the class of nonlinear Markov processes. One of the most important and difficult research questions concerns the regularity analysis and more particularly the stability and the long time behavior of these stochastic models. In contrast with conventional Markov processes, one of the main difficulty of these Markov processes comes from the fact that the evolution semigroup $\phi_{s,t}(\mu)$ is nonlinear w.r.t. the initial condition $\mu$ of the system. The additional complexity in the analysis of these models is that their state space is the convex set of probability measures, thus conventional functional analysis and differential calculus on Banach space cannot be directly applied. The main contribution of this article is the development of a second order differential calculus to analyze the regularity and the stability properties of the distribution semigroup associated with McKean-Vlasov diffusions. This methodology provides second order Taylor type expansions with remainder for both the evolution semigroup as well as the stochastic flow associated with this class of nonlinear diffusions. We also provide a refined analysis of the norm of these integro-differential operators with a series of exponential decays inequalities with respect to the time horizon. The article is organized as follows: The main contributions of this article are briefly discussed in section \[sec-statement-intro\]. The main theorems are stated in some detailed in section \[sec-statements\]. Section \[spr-sec\] provides some pivotal results on tensor integral operators and on integro-differential operators associated with the second order Taylor expansions of the semigroup $\phi_{s,t}(\mu)$. Section \[tangent-sec\] is dedicated to the analysis of the tangent process associated with the nonlinear diffusion flow. We presents explicit Dyson-Phillips expansions as well as some spectral estimates. The last section, section \[tangent-sec\] is mainly concerned with the proofs of the first and second order Taylor expansions. The proof of some technical results are collected in the appendix. Detailed comparisons with existing literature on this subject are also provided in section \[comparison-sec\]. Statement of some main results {#sec-statement-intro} ------------------------------ One of the main contribution of the present article is the derivation of a second order Taylor expansion with remainder of the semigroup $\phi_{s,t}$ on probability spaces. For any pair of measures $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$, these expansions take basically the following form: $$\label{TT-intro} \phi_{s,t}(\mu_1)\simeq \phi_{s,t}(\mu_0)+(\mu_1-\mu_0)D_{\mu_0}\phi_{s,t}+\frac{1}{2}~(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_0}\phi_{s,t}$$ In the above display, $D^k_{\mu_0}\phi_{s,t}$ stands some first and second order operators, with $k=1,2$. A more precise description of these expansions and the remainder terms is provided in section \[sec-taylor-intro\]. Section \[ae-expansion\], also provides an almost sure second order Taylor expansions with remainder of the random state $X^{\mu}_{s,t}(x)$ of the McKean diffusion w.r.t. the initial distribution $\mu$. These almost sure expansions take basically the following form $$\label{TT-intro-a-e} \begin{array}{l} \displaystyle X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)\simeq \int (\mu_1-\mu_0)(dy)~D_{\mu_0}X^{\mu_0}_{s,t}(x,y)+\frac{1}{2}~~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~D^2_{\mu_0}X^{\mu_0}_{s,t}(x,z) \end{array}$$ for some random functions $D^k_{\mu_0}X^{\mu_0}_{s,t}$ from ${\mathbb{R}}^{(1+k)d}$ into ${\mathbb{R}}^d$, with $k=1,2$. A more precise description of these almost sure expansions is provided in section \[ae-expansion\] (see for instance (\[a-e-T1\]) and theorem \[theo-ae-taylor\]). Given some random variable $Y\in {\mathbb{H}}_s({\mathbb{R}}^d)$ with distribution $\mu\in P_2({\mathbb{R}}^d)$, observe that the stochastic flow $ \psi_{s,t}(Y):=X^{\mu}_{s,t}(Y)$ satisfies the ${\mathbb{H}}_t({\mathbb{R}}^d)$-valued stochastic differential equation $$\label{eq-diff} d \psi_{s,t}(Y):=B_t( \psi_{s,t}(Y))~dt+dW_t$$ In the above display, $B_t$ stands for the drift function from ${\mathbb{H}}_t({\mathbb{R}}^d)$ into itself defined by the formula $$B_t(X):={{\mathbb{E}}}\left(b_t(X,\overline{X})~\|~X\right)$$ In the above display, $\overline{X}$ stands for an independent copy of $X$. The above Hilbert space valued representation of the McKean-Vlasov diffusion (\[diff-st-ref-general\]) readily implies that for any $Y_1,Y_0\in {\mathbb{H}}_s({\mathbb{R}}^d)$ we have the exponential contraction inequality $$\Vert \psi_{s,t}(Y_1)-\psi_{s,t}(Y_0)\Vert_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}\leq e^{-\lambda (t-s)}~\Vert Y_1-Y_0\Vert_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}$$ for some $\lambda>0$, as soon as the following condition is satisfied $$\label{Hilbert-condition} \left\langle X_1-X_0, B_t(X_1)- B_t(X_0)\right\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}\leq-2\lambda~ \Vert X_1-X_0\Vert_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}^2$$ for any $t\geq 0$ and any $X_1,X_0\in {\mathbb{H}}_t({\mathbb{R}}^d)$. In addition, in this framework the first order differential $ \partial \psi_{s,t}(Y)$ of the stochastic flow coincides with the conventional Fréchet derivative of functions from an Hilbert space into another. In addition, we shall see that the gradient of first order operator $D_{\mu}\phi_{s,t}$ coincides with the dual of the tangent process associated with the Hilbert space-valued representation (\[eq-diff\]) of the McKean-Vlasov diffusion (\[diff-st-ref-general\]); that is, for any smooth function $f$ we have that the dual tangent formula $$\label{ref-intro-dual-D} \partial \psi_{s,t}(Y)^{\star}\cdot\nabla f(\psi_{s,t}(Y))=\nabla D_{\mu}\phi_{s,t}(f)(Y)$$ A more precise description of the Fréchet differential $ \partial \psi_{s,t}(Y)$ and the dual operator is provided in section \[sec-H-tangent-intro\] and section \[tangent-sec\]. A proof of the above formula is provided in theorem \[theo-0\]. The Taylor expansions discussed above are valid under fairly general and easily verifiable conditions on the drift function. For instance, the regularity condition (\[def-H\]) is clearly satisfied for linear drift functions. As it is well known, dynamical systems and hence stochastic models involving drift functions with quadratic growth require additional regularity conditions to ensure non explosion of the solution in finite time. Of course the expansions (\[TT-intro\]) and (\[TT-intro-a-e\]) will be of rather poor practical interest without a better understanding of the differential operators and the remainder terms. To get some useful approximations, we need to quantify with some precision the norm of these operators. A important part of the article is concerned with developing a series of quantitative estimates of the differential operators $D^k_{\mu_0}\phi_{s,t}$ and the remainder term; see for instance theorem \[theo-intro-3\] and theorem \[theo-intro-4\]. To avoid estimates that grow exponentially fast with respect to the time horizon, we need to estimate with some precision the operator norms of the differential operators in (\[TT-intro\]). To this end, we shall consider an additional regularity condition:\ [$(H)$ : There exists some $\lambda_0>0$ and $\lambda_1>\Vert b^{[2]}\Vert_2$ such that for any $(x_1,x_2)\in{\mathbb{R}}^{2d}$ and any time horizon $t\geq 0$ we have $$\label{def-HS-0} A_t(x_1,x_2)_{\tiny sym}\leq -\lambda_0~I\quad \mbox{and}\quad b_t^{[1]}(x_1,x_2)_{\tiny sym}\leq -\lambda_1~I$$ In the above display, $I$ stands for the identity matrix and $A_t$ the matrix-valued function defined by $$\label{def-lambda-1-2} A_t(x_1,x_2):=\left[\begin{array}{cc} b_t^{[1]}(x_1,x_2)&b_t^{[2]}(x_2,x_1)\\ b_t^{[2]}(x_1,x_2)&b_t^{[1]}(x_2,x_1) \end{array} \right]\quad \mbox{and we set}\quad \lambda_{1,2}:= \lambda_1-\Vert b^{[2]}\Vert_2$$ ]{} Whenever (\[def-HS-0\]) and (\[def-lambda-1-2\]) are met for some parameters $\lambda_0$ and $\lambda_1\in {\mathbb{R}}$ all the exponential estimates stated in the article remains valid but they grow exponentially fast with respect to the time horizon. More detailed comments on the above regularity conditions, including illustrations for linear drift and gradient flow models, as well as comparisons with related conditions used in the literature on this subject are also provided in section \[sec-H-comments-intro\]. Under the above condition, we shall develop several exponential decays inequalities for the norm of the differential operators $D_{\mu_0}^k\phi_{s,t}$ as well as for the remainder terms in the Taylor expansions. The first order estimates are given in (\[est-nabla-D-intro\]), the ones on the Bismut-Elworthy-Li gradient and Hessian extension formulae are provided in (\[def-lambda-hat-bismut-intro\]) and (\[def-lambda-hat-bismut-intro-hessian\]). Second and third order estimates can also be found in (\[estimate-D2-nabla\]) and (\[D3-estimation\]). The second order differential calculus discussed above provides a natural theoretical basis to analyze the stability properties of the semigroup $\phi_{s,t}$ and the one of the mean field particle system discussed in (\[diff-st-ref-general-mf\]). For instance, a first order Taylor expansion of the form (\[TT-intro\]) already indicates that the sensitivity properties of the semigroup w.r.t. the initial condition $\mu$ are encapsulated in the first order differential operator $D_{\mu}\phi_{s,t}$. Roughly speaking, whenever $(H)$ is satisfied, we show that there exists some parameter $\lambda>0$ such that $$\label{R-intro-0} \vee_{k=1,2}{{\vert\kern-0.25ex\vert\kern-0.25ex\vert D_{\mu_0}^k\phi_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\simeq e^{-\lambda (t-s)}\quad \mbox{\rm and therefore}\quad {{\vert\kern-0.25ex\vert\kern-0.25ex\vert \phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0) \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\simeq e^{-\lambda (t-s)}$$ for some operator norms ${{\vert\kern-0.25ex\vert\kern-0.25ex\vert {\mbox{\LARGE .}}\vert\kern-0.25ex\vert\kern-0.25ex\vert}}$. For a more precise statement we refer to theorem \[theo-intro-2\] and the discussion following the theorem. The second order expansion (\[TT-intro\]) also provides a natural basis to quantify the propagation of chaos properties of the mean field particle model (\[diff-st-ref-general-mf\]). Combining these Taylor expansions with a backward semigroup analysis we derive a a variety of uniform mean error estimates w.r.t. the time horizon. This backward second order analysis can be seen a second order extension of the Alekseev-Gröbner lemma [@alekseev; @grobner] to nonlinear measure valued and stochastic semigroups. For a more precise statement we refer to theorem \[theo-ag-mf\]. As in (\[R-intro-0\]), one of the main feature of the expansion (\[TT-intro\]) is that it allows to enter the stability properties of the limiting semigroup $\phi_{s,t}$ into the analysis of the flow of empirical measures $m(\xi_t)$. Roughly speaking, this backward perturbation analysis can be interpreted as a second order variation-of-constants technique applied to nonlinear equations in distribution spaces. As in the Ito’s lemma, the second order term is essential to capture the quadratic variation of the processes, see for instance the recent articles [@dm-singh; @hudde] in the context of conventional stochastic differential equation, as well as in [@mp-18; @dm-2003] in the context of interacting jump models. The discrete time version of this backward perturbation semigroup methodology can also be found in chapter 7 in [@d-2004], a well as in the articles [@dm-g-99; @guionnet; @dm-2000] and [@dmrio-09; @dm-hu-wu] for general classes of mean field particle systems. The central idea is to consider the telescoping sum on some time mesh $t_n\leq t_{n+1}$ given by the interpolating formula $$m_{t_n}-\phi_{t_0,t_n}(m_{t_0})=\sum_{1\leq k\leq n}\left[\phi_{t_k,t_n}(m_{t_k})-\phi_{t_{k},t_n}\left(\phi_{t_{k-1},t_k}(m_{t_{k-1}})\right)\right] \quad\mbox{\rm with}\quad m_{t_k}:=m(\xi_{t_k})$$ Applying (\[TT-intro\]) and whenever $(t_{k}-t_{k-1})\simeq 0$ we have the second order approximation $$m_{t_n}-\phi_{t_0,t_n}(m_{t_0})\simeq\frac{1}{\sqrt{N}}\sum_{1\leq k\leq n} \Delta M_{t_{k}} D_{m_{t_{k-1}}}\phi_{t_{k},t_n}+\frac{1}{2N}\sum_{1\leq k\leq n} (\Delta M_{t_{k}})^{\otimes 2}D^2_{m_{t_{k-1}}}\phi_{t_k,t}$$ with the local fluctuation random fields $$\Delta M_{t_{k}}:=\sqrt{N}~\left(m_{t_k}-\overline{m}_{t_{k}}\right)\quad \mbox{\rm and}\quad \overline{m}_{t_{k}}:=\phi_{t_{k-1},t_k}\left(m_{t_{k-1}}\right)\simeq m_{t_{k-1}}$$ For discrete generation particle systems, $\xi_{t_k}^i$ are defined by $N$ conditionally independent variables given the system $\xi_{t_{k-1}}$. For a more rigorous analysis we refer to section \[ips-sec\]. The above decomposition shows that the first order operator $D_{\mu}\phi_{s,t}$ reflects the fluctuation errors of the particle measures, while the second order term encapsulates their bias. In other words, estimating the norm of second order operator $D^2_{\mu}\phi_{s,t}$ allows to quantify the bias induced by the interaction function, while the estimation of first order term is used to derive central limit theorems as well as ${\mathbb{L}}_p$-mean error estimates. As in (\[R-intro-0\]), these estimates take basically the following form. For $n\geq 1$ and any sufficiently regular function $f$ we have $$\label{ln-est} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert D_{\mu_0}\phi_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\simeq e^{-\lambda (t-s)}\Longrightarrow \vert{\mathbb{E}}\left[\Vert m_t(f)-\phi_{0,t}(m_0)(f)\Vert^n\right]^{1/n}\vert\leq c_n/\sqrt{N}$$ In addition, we have the uniform bias estimate w.r.t. the time horizon $$\label{bias-intro-est} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert D^2_{\mu_0}\phi_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\simeq e^{-\lambda (t-s)}\Longrightarrow \vert{\mathbb{E}}\left[m_t(f)-\phi_{0,t}(m_0)(f)\right]\vert\leq c/N$$ In the above display, ${{\vert\kern-0.25ex\vert\kern-0.25ex\vert {\mbox{\LARGE .}}\vert\kern-0.25ex\vert\kern-0.25ex\vert}}$ stands for some operator norm, and $(c,c_n)$ stands for some finite constants whose values doesn’t depend on the time horizon. We emphasize that the above results are direct consequence of a second order extension of the Alekseev-Gröbner type lemma for particle density profiles. For more precise statements we refer to theorem \[theo-ag-mf\] and the discussion following the theorem. Some basic notation {#sec-notation} ------------------- Let $\mbox{\rm Lin}({ {\cal B }}_1,{ {\cal B }}_2)$ be the set of bounded linear operators from a normed space ${ {\cal B }}_1$ into a possibly different normed space ${ {\cal B }}_2$ equipped with the operator norm ${{\vert\kern-0.25ex\vert\kern-0.25ex\vert {\mbox{\LARGE .}}\vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{ {\cal B }}_1\rightarrow { {\cal B }}_2}$. When ${ {\cal B }}_1={ {\cal B }}_2$ we write $\mbox{\rm Lin}({ {\cal B }}_1)$ instead of $\mbox{\rm Lin}({ {\cal B }}_1,{ {\cal B }}_1)$. With a slight abuse of notation, we denote by $I$ the identity $(d\times d)$-matrix, for any $d\geq 1$, as well as the identity operator in $\mbox{\rm Lin}({ {\cal B }}_1,{ {\cal B }}_1)$. We also denote by $\Vert{\mbox{\LARGE .}}\Vert$ any (equivalent) norm on some finite dimensional vector space over ${\mathbb{R}}$. We also use the conventional notation $\partial_{\epsilon}$, $\partial_{x_i}$, $\partial_s$, $\partial_t$ and so on for the partial derivatives w.r.t. some real valued parameters $\epsilon$, $x_i$, $s$ and $t$. We let $\nabla f(x)=\left[\partial_{x_i}f(x)\right]_{1\leq i\leq d}$ be the gradient column vector associated with some smooth function $f(x)$ from ${\mathbb{R}}^d$ into ${\mathbb{R}}$. Given some smooth function $h(x)$ from ${\mathbb{R}}^d$ into ${\mathbb{R}}^d$ we denote by $\nabla h=\left[\nabla h^1,\ldots,\nabla h^d\right]$ the gradient matrix associated with the column vector function $h=(h^i)_{1\leq i\leq d}$. We also let $(\nabla\otimes\nabla)$ be the second order differential operator defined for any twice differentiable function $g(x_1,x_2)$ on ${\mathbb{R}}^{2d}$ by the Hessian-type formula $$\begin{aligned} \left((\nabla\otimes\nabla)g\right)_{i,j}&=&(\nabla_{x_1}\otimes\nabla_{x_2})(g)_{i,j}=(\nabla_{x_2}\otimes\nabla_{x_1})(g)_{j,i}=\partial_{x^i_1}\partial_{x^j_2}g\label{def-nabla-nabla}\end{aligned}$$ We consider the space ${ {\cal C }}^n({\mathbb{R}}^d)$ of $n$-differentiable functions and we denote by ${ {\cal C }}^n_m({\mathbb{R}}^d)$ the subspace of functions $f$ such that $$\sup_{0\leq k\leq n}{\Vert\nabla^k f(x)\Vert}\leq c~w_m(x)\quad\mbox{\rm with the weight function}\quad w_m(x)=(1+\Vert x\Vert)^{m}\quad \mbox{\rm for some}\quad m\geq 0.$$ We equip ${ {\cal C }}^n_m({\mathbb{R}}^d)$ with the norm $$\Vert f\Vert_{{ {\cal C }}^n_m({\mathbb{R}}^d)}:=\sum_{0\leq k\leq n}\Vert \nabla^kf/w_{m}\Vert_{\infty} \quad\mbox{\rm with}\quad \Vert \nabla^kf/w_{m}\Vert_{\infty}=\sup_{x\in {\mathbb{R}}^d}\Vert \nabla^kf(x)/w_{m}(x)\Vert$$ When there are no confusions, we drop to lower symbol $\Vert{\mbox{\LARGE .}}\Vert_{\infty}$ and we write $\Vert f\Vert$ instead of $\Vert f\Vert_{\infty}$ the supremum norm of some real valued function. We let $e(x):=x$ be the identify function on ${\mathbb{R}}^d$ and for any $\mu\in P_n({\mathbb{R}}^d)$ and $n\geq 1$ we set $$\Vert e\Vert_{\mu,n}:=\left[\int~\Vert x\Vert^n~\mu(dx)\right]^{1/n}$$ For any $\mu_1,\mu_2\in P_n({\mathbb{R}}^d)$, we also denote by $\rho_n(\mu_1,\mu_2)$ some polynomial function of $\Vert e\Vert_{\mu_i,n}$ with $i=1,2$. When $\mu_1=\mu_2$ we write $\rho_n(\mu_1)$ instead of $\rho_n(\mu_1,\mu_1)$. Under our regularity conditions on the drift function, using elementary stochastic calculus for any $n\geq 2$ and $\mu\in P_n({\mathbb{R}}^d)$ we check the following estimates $$\label{ref-moments} {\mathbb{E}}\left(\Vert X_{s,t}^{\mu}(x)\Vert^n\right)^{1/n}\leq c_{n}(t)~\left(\Vert x\Vert+\Vert e\Vert_{\mu,2}\right) ~~ \mbox{\rm which implies that}~~ \phi_{s,t}(\mu)(\Vert e\Vert^n)^{1/n}~\leq c_{n}(t)~\Vert e\Vert_{\mu,n}$$ In the above display and throughout the rest of the article, we write $c(t), c_{\epsilon}(t), c_n(t), c_{n,\epsilon}(t), c_{\epsilon,n}(t)$ and $c_{m,n}(t)$ with $m,n\geq 0$ and $\epsilon\in [0,1]$ some collection of non decreasing and non negative functions of the time parameter $t$ whose values may vary from line to line, but which only depend on the parameters $m,n, \epsilon$, as well as on the drift function $b_t$. Importantly these contants do not depend on the probability measures $\mu$. We also write $c,c_{\epsilon},c_n,c_{n,\epsilon},$ and $c_{m,n}$ when the constant do not depend on the time horizon. Statement of the main theorems {#sec-statements} ============================== First variational equation on Hilbert spaces {#sec-H-tangent-intro} -------------------------------------------- As expected, the Fréchet differential $\partial \psi_{s,t}(Y)$ of the stochastic flow $\psi_{s,t}(Y)$ associated with the stochastic differential equation (\[eq-diff\]) satisfies an Hilbert space-valued linear equation (cf. (\[eq-diff-lin-eq\])). The drift-matrix of this evolution equation is given by the Fréchet differential $\partial B_{t}(\psi_{s,t}(Y))$ of the drift function $B_t$ evaluated along the solution of the flow. Mimicking the exponential notation of the solution of conventional homogeneous linear systems, the evolution semigroup (a.k.a. propagator) associated with the first variational equation is written as follows $$\partial \psi_{s,t}(Y)=e^{\oint_s^t\partial B_{u}(\psi_{s,u}(Y))\,du}\in \mbox{\rm Lin}({\mathbb{H}}_s({\mathbb{R}}^d),{\mathbb{H}}_t({\mathbb{R}}^d))$$ The above exponential is understood as an operator valued Peano-Baker series [@Peano]. A more detailed presentation of these models is provided in section \[tangent-sec\]. The ${\mathbb{H}}_t({\mathbb{R}}^d)$-log-norm of an operator $T_t\in \mbox{\rm Lin}({\mathbb{H}}_t({\mathbb{R}}^d),{\mathbb{H}}_t({\mathbb{R}}^d))$ is defined by $$\gamma(T_t):= \sup_{\Vert Z\Vert_{{\mathbb{H}}_t({\mathbb{R}}^d)}=1}{\langle Z, (T_t+T^{\star}_t)/2\cdot Z\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}}$$ Our first main result is an extension of an inequality of Coppel [@coppel] to tangent processes associated with Hilbert-space valued stochastic flows. \[theo-intro-1\] For any time horizon $t\geq s$ and any $Y\in{\mathbb{H}}_s({\mathbb{R}}^d)$ we have the log-norm estimate $$\label{theo-intro-1-eq-log-norm} - \int_s^t\gamma\left(-\partial B_{u}(\psi_{s,u}(Y))\right)~du\leq \frac{1}{t}\log{ {{\vert\kern-0.25ex\vert\kern-0.25ex\vert e^{\oint_s^t\partial B_{u}(\psi_{s,u}(Y))\,du} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{\mathbb{H}}_t({\mathbb{R}}^d)\rightarrow {\mathbb{H}}_t({\mathbb{R}}^d)}}\leq \int_s^t\gamma\left(\partial B_{u}(\psi_{s,u}(Y))\right)~du$$ In addition, we have $$\label{theo-intro-1-eq} (H)\Longrightarrow \partial B_t(X)_{\tiny sym}\leq -\lambda_0~I\Longrightarrow\frac{1}{t}\log{ {{\vert\kern-0.25ex\vert\kern-0.25ex\vert e^{\oint_s^t\partial B_{u}(\psi_{s,u}(Y))\,du} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{\mathbb{H}}_t({\mathbb{R}}^d)\rightarrow {\mathbb{H}}_t({\mathbb{R}}^d)}}\leq -\lambda_0$$ The proof of the above theorem in provided in section \[spectral-sec\]. Let $Y_0,Y_1\in {\mathbb{H}}_s({\mathbb{R}}^d)$ be a pair of random variables with distributions $(\mu_0,\mu_1)\in P_2({\mathbb{R}}^d)^2$. Also let $\mu_{\epsilon}$ be the probability distribution of the random variable $$\label{def-Y-epsilon} Y_{\epsilon}:=(1-\epsilon)~Y_0+\epsilon ~Y_1\Longrightarrow\partial_{\epsilon}\psi_{s,t}(Y_{\epsilon})=e^{\oint_s^t\partial B_{u}(\psi_{s,u}(Y))\,du}\cdot(Y_1-Y_0)$$ This observation combined with the above theorem yields an alternative and more direct proof of an exponential Wasserstein contraction estimate obtained in [@mp-var-18]. Namely, using (\[theo-intro-1-eq\]) we readily check the $\WW_2$-exponential contraction inequality $$\label{ref-stab-W-1} { \partial B_t(X)_{\tiny sym}\leq -\lambda_0~I\quad \Longrightarrow}\quad\WW_2(\phi_{s,t}(\mu_1),\phi_{s,t}(\mu_0))\leq e^{-\lambda_0 (t-s)}~\WW_2(\mu_0,\mu_1)$$ For any function $f\in { {\cal C }}^1({\mathbb{R}}^d)$ with bounded derivative we also quote the first order expansion $$\left[\phi_{s,t}(\mu_1)-\phi_{s,t}(\mu_0)\right](f)=\int_0^1~\langle \partial\psi_{s,t}(Y_{\epsilon})^{\star}\cdot\nabla f(\psi_{s,t}(Y_{\epsilon})), (Y_1-Y_0)\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}~d\epsilon$$ In the above display, $\langle {\mbox{\LARGE .}},{\mbox{\LARGE .}}\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}$ stands for the conventional inner product on ${\mathbb{L}}_2((\Omega,{\mathbb{F}}_t,{\mathbb{P}}),{\mathbb{R}}^d)$. The above assertion is a direct consequence of theorem \[theo-0\]. Taylor expansions with remainder {#sec-taylor-intro} -------------------------------- The first expansion presented in this section is a first order linearization of the measure valued mapping $\phi_{s,t}$ in terms of a semigroup of linear integro-differential operators. \[theo-intro-2\] For any $m,n\geq 1$ and $\mu_0,\mu_1\in P_{m\vee 2}({\mathbb{R}}^d)$, there exists a semigroup of linear operators $D_{\mu_1,\mu_0}\phi_{s,t}$ from ${ {\cal C }}^n_m({\mathbb{R}}^d)$ into itself such that $$\label{a-1} \phi_{s,t}(\mu_1)= \phi_{s,t}(\mu_0)+(\mu_1-\mu_0)D_{\mu_1,\mu_0}\phi_{s,t}$$ In addition, when $(H)$ is satisfied we have the gradient estimate $$\label{est-nabla-D-intro} \Vert \nabla D_{\mu_1,\mu_0}\phi_{s,t}(f)\Vert\leq c~e^{-\lambda(t-s)}~\Vert\nabla f\Vert\quad \mbox{for some $\lambda>0$}$$ The proof of the above theorem with a more explicit description of the first order operators $D_{\mu_1,\mu_0}\phi_{s,t}$ are provided in section \[g-sg-sec\]. [In (\[est-nabla-D-intro\]) we can choose $\lambda=\lambda_{1,2}$, with the parameter $\lambda_{1,2}$ introduced in (\[def-lambda-1-2\])]{}. The semigroup property is a consequence of theorem \[theo-1\] and the gradient estimates is a reformulation of the operator norm estimate discussed in (\[commutation-D-etimate\]). We also provide Bismut-Elworthy-Li-type formulae that allow to extend the gradient and Hessian operators $\nabla^k D_{\mu_1,\mu_0}\phi_{s,t}$ with $k=1,2$ to measurable and bounded functions. When the condition $(H)$ is satisfied we show the following exponential estimates $$\label{def-lambda-hat-bismut-intro} \Vert \nabla D_{\mu_1,\mu_0}\phi_{s,t}(f)\Vert\leq c~\left(1\vee1/\sqrt{t-s}\right)~e^{- \lambda(t-s)}~\Vert f\Vert\quad \mbox{for some $\lambda>0$}$$ In addition, we have the Hessian estimate $$\label{def-lambda-hat-bismut-intro-hessian} \Vert \nabla^2 D_{\mu_1,\mu_0}\phi_{s,t}(f)\Vert\leq c~\left(1\vee 1/(t-s)\right)~e^{- \lambda(t-s)}~\Vert f\Vert \quad \mbox{for some $\lambda>0$}$$ The proof of the first assertion can be found in remark \[rmk-1\] on page . The proof of the Hessian estimates is a consequence of the decomposition of $\nabla^2 D_{\mu_0,\mu_1}\phi_{s,t}$ discussed in (\[dD-W\]) and the Hessian estimates (\[bismut-est-P2\]) and (\[def-lambda-hat-bismut\]). It is worth mentioning that the semigroup property is equivalent to the chain rule formula $$\label{sg-chain-rule} D_{\mu_1,\mu_0}\phi_{s,t}=D_{\mu_1,\mu_0}\phi_{s,u}\circ D_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}$$ which is valid for any $s\leq u\leq t$. Without further work, theorem \[theo-intro-2\] also yields the exponential $\WW_1$-contraction inequality $$\label{ref-stab-W-1-1} \WW_1(\phi_{s,t}(\mu_1),\phi_{s,t}(\mu_0))\leq c~e^{-\lambda (t-s)}~\WW_1(\mu_0,\mu_1)$$ with the same parameter $\lambda$ a in (\[est-nabla-D-intro\]). In the same vein, the estimate (\[def-lambda-hat-bismut-intro\]) yields the total variation estimate $$\Vert \phi_{s,t}(\mu_1)-\phi_{s,t}(\mu_0)\Vert_{\tiny tv}\leq c~\left(1\vee1/\sqrt{t-s}\right)~e^{- \lambda(t-s)}~\Vert\mu_0-\mu_1\Vert_{\tiny tv}$$ with the same parameter $\lambda$ a in (\[def-lambda-hat-bismut-intro\]). [In all the inequalities discussed above we can choose any parameter $\lambda>0$ such that $\lambda<\lambda_{1,2}$, with the parameter $\lambda_{1,2}$ introduced in (\[def-lambda-1-2\]). In the $\WW_1$-contraction inequality (\[ref-stab-W-1-1\]) we can choose $\lambda=\lambda_{1,2}$. A more refined estimate is provided in section \[sec-H-comments-intro\].]{} Next theorem provides a first order Taylor expansion with remainder. \[theo-intro-3\] For any $m,n\geq 0$ and $\mu_0,\mu_1\in P_{m+2}({\mathbb{R}}^d)$, there exists a linear operators $D^2_{\mu_1,\mu_0}\phi_{s,t}$ from ${ {\cal C }}^{n+2}_m({\mathbb{R}}^d)$ into ${ {\cal C }}^{n}_{m+2}({\mathbb{R}}^{2d})$ such that $$\label{s-o-r} \phi_{s,t}(\mu_1)= \phi_{s,t}(\mu_0)+(\mu_1-\mu_0)D_{\mu_0}\phi_{s,t}+\frac{1}{2}~(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}$$ with the first order operator $ D_{\mu_0}\phi_{s,t}:=D_{\mu_0,\mu_0}\phi_{s,t}$ introduced in theorem \[theo-intro-2\]. In addition, when $(H)$ is satisfied we also have the estimate $$\label{estimate-D2-nabla} \Vert (\nabla\otimes\nabla) D^2_{\mu_1,\mu_0}\phi_{s,t}(f)\Vert\leq c~e^{-\lambda(t-s)}~\sup_{i=1,2}\Vert \nabla^i f\Vert\quad \mbox{for some $\lambda>0$}$$ [ The proof of the above theorem in provided in section \[sec-first-order-remainder\]. A more precise description of the second order operator $D^2_{\mu_1,\mu_0}\phi_{s,t}$ is provided in (\[def-D2\]) and (\[nabla-2-D-2\]). Using (\[s-o-r\]) and arguing as in the proof of proposition 2.1 in [@mp-18], for any twice differentiable function $f$ with bounded derivatives we check the backward evolution equation $$\label{ref-backward-eq} \partial_s\phi_{s,t}(\mu)(f)=-\mu L_{s,\mu}\left(D_{\mu}\phi_{s,t}(f)\right)$$ with the first order operator $ D_{\mu}\phi_{s,t}$ introduced in theorem \[theo-intro-3\]. The above equation is a central tool to derive an extended version of the Alekseev-Gröbner lemma [@alekseev; @grobner] to measure valued semigroups and interacting diffusions (cf. theorem \[theo-ag-mf\]).]{} Next theorem provides a second order Taylor expansion with remainder. \[theo-intro-4\] For any $m,n\geq 1$ and $\mu_0,\mu_1\in P_{m+4}({\mathbb{R}}^d)$, there exists a linear operators $D^3_{\mu_1,\mu_0}\phi_{s,t}$ from ${ {\cal C }}^{n+3}_m({\mathbb{R}}^d)$ into ${ {\cal C }}^{n}_{m+4}({\mathbb{R}}^{3d})$ such that $$\label{taylor-2-intro} \begin{array}{l} \phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0)\\ \\ \displaystyle=(\mu_1-\mu_0)D_{\mu_0}\phi_{s,t}+\frac{1}{2} (\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_0}\phi_{s,t}+(\mu_1-\mu_0)^{\otimes 3}D^3_{\mu_0,\mu_1}\phi_{s,t} \end{array}$$ with the second order operator $ D^2_{\mu_0}\phi_{s,t}:=D^2_{\mu_0,\mu_0}\phi_{s,t}$ introduced in theorem \[theo-intro-3\]. In addition, when $(H)$ is satisfied we have the third order estimate $$\label{D3-estimation} \begin{array}{l} \vert (\mu_1-\mu_0)^{\otimes 3}D^3_{\mu_0,\mu_1}\phi_{s,t}(f)\vert\\ \\ \leq c~ e^{-\lambda(t-s)}~\left(\vee_{i=1,2,3}\Vert\nabla^i f\Vert\right)~\WW_2(\mu_0,\mu_1)^3\quad \mbox{for some $\lambda>0$} \end{array}$$ The proof of the first part of the above theorem in provided in section \[sec-o-analysis\]. [We can choose in (\[D3-estimation\]) any parameter $\lambda>0$ such that $\lambda<\lambda_{1,2}$, with the parameter $\lambda_{1,2}$ introduced in (\[def-lambda-1-2\])]{}. The proof of the third order estimate (\[D3-estimation\]) is rather technical, thus it is provided in the appendix, on page . Illustrations {#illustration-sec} ------------- The first part of this section states with more details the almost sure expansions discussed in (\[TT-intro-a-e\]). Up to some differential calculus technicalities, this result is a more or less direct consequence of the Taylor expansions with remainder presented in theorem \[theo-intro-3\] and theorem \[theo-intro-4\] combining with a backward formula presented in [@mp-var-18]. The second part of this section is concerned with a second order extension of the Alekseev-Gröbner lemma to nonlinear measure valued semigroups and interacting diffusion flows. This second order stochastic perturbation analysis is also mainly based on the second order Taylor expansion with remainder presented in theorem \[theo-intro-4\] . In the further development of this section without further mention we shall assume that condition $(H)$ is satisfied. ### Almost sure expansions {#ae-expansion} We recall the backward formula $$\label{stoch-interpolation-bis} X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)=\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_1}_{s,u}(x))^{\prime}~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_1}_{s,u}(x),{\mbox{\LARGE .}})) ~du$$ The above formula combined with (\[ref-stab-W-1\]) and the tangent process estimates presented in section \[var-eq-sec\] yields the uniform almost sure estimates $$\label{estimate-X-mu} \Vert X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)\Vert\leq e^{-(\lambda_0\wedge\lambda_1)(t-s)}~\WW_2(\mu_0,\mu_1)$$ The above estimate is a consequence of (\[ref-stab-W-1\]) and conventional exponential estimates of the tangent process $\nabla X^{\mu}_{s,t}$ (cf. for instance (\[tau-1-estimates\])). A detailed proof of this claim and the backward formula (\[stoch-interpolation-bis\]) can be found in [@mp-var-18]. We extend the operators $D^k_{\mu}\phi_{s,t}$ introduced in theorem \[theo-intro-4\] to tensor valued functions $f=(f_{i})_{i\in [n]}$ with $i=(i_1,\ldots,i_n)\in [n]:=\{1,\ldots,d\}^n$ by considering the same type tensor function with entries $$\label{def-D-k-extension-intro} D_{\mu}^k\phi_{s,t}(f)_i:=D_{\mu}^k\phi_{s,t}(f_{i})\quad\mbox{\rm and we set}\quad d_{s,t}^{\mu}(x,y):=D_{\mu}\phi_{s,t}(b_t(x,{\mbox{\LARGE .}}))(y)$$ for any $(x,y)\in {\mathbb{R}}^{2d}$. A brief review on tensor spaces is provided in section \[not-tensor-sec\]. We also consider the function $$D_{\mu}X^{\mu}_{s,t}(x,y):=\int_s^t\left[\nabla X^{\phi_{s,u}(\mu)}_{u,t}\right]({X}^{\mu}_{s,u}(x))^{\prime}~d_{s,u}^{\mu}(X^{\mu}_{s,u}(x),y)~du$$ Combining the first order formulae stated in theorem \[theo-intro-3\] with conventional Taylor expansions we check the following theorem. For any $x\in{\mathbb{R}}^d$, $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ and $s\leq t$ we have the almost sure expansion $$\label{a-e-T1} \displaystyle X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)=\int~(\mu_1-\mu_0)(dy)~D_{\mu_0}X^{\mu_0}_{s,t}(x,y)+\Delta^{[2],\mu_0,\mu_1}_{s,t}(x)$$ with the second order remainder function $\Delta^{[2],\mu_0,\mu_1}_{s,t}$ such that $$\Vert \Delta^{[2],\mu_0,\mu_1}_{s,t}\Vert\leq c~e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^2 \quad \mbox{for some $\lambda>0$}$$ The detailed proof of the above theorem is provided in the appendix, on page . Second order expansions are expressed in terms of the functions defined for any $(x,y)\in {\mathbb{R}}^{2d}$ and for any $z\in {\mathbb{R}}^{2d}$ by the formulae $$d_{s,t}^{[1,1],\mu}(x,y):=D_{\mu}\phi_{s,t}(b_t^{[1]}(x,{\mbox{\LARGE .}})^{\prime})(y)\quad\mbox{\rm and}\quad d_{s,t}^{[2],\mu}(x,z):=D_{\mu}^2\phi_{s,t}(b_t(x,{\mbox{\LARGE .}}))(z)$$ We associate with these objects the function $D^2_{\mu_0}X^{\mu_0}_{s,t}$ defined by $$\begin{aligned} \displaystyle D^2_{\mu}X^{\mu}_{s,t}(x,z)&:=&\int_s^t\left[\nabla X^{\phi_{s,u}(\mu)}_{u,t}\right]({X}^{\mu}_{s,u}(x))^{\prime}~\left[ d_{s,u}^{[2],\mu}(X^{\mu}_{s,u}(x),z)+D^{[1,1]}_{\mu}X^{\mu}_{s,u}(x,z)\right]~du\\ &&\hskip4cm \displaystyle+\int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu)}_{u,t}\right]({X}^{\mu}_{s,u}(x))^{\prime}~D^{[2,1]}_{\mu}X^{\mu}_{s,u}(x,z)~du\end{aligned}$$ In the above display, $D^{[i,1]}_{\mu}X^{\mu}_{s,u}$ stands for the functions given by $$\begin{array}{l} \displaystyle D^{[1,1]}_{\mu}X^{\mu}_{s,u}(x,z):=\left[d_{s,u}^{[1,1],\mu}(X^{\mu}_{s,u}(x),z_2)~ D_{\mu}X^{\mu}_{s,u}(x,z_1)+d_{s,u}^{[1,1],\mu}(X^{\mu_0}_{s,u}(x),z_1)~ D_{\mu}X^{\mu}_{s,u}(x,z_2)\right]~\\ \\ \displaystyle D^{[2,1]}_{\mu_0}X^{\mu}_{s,u}(x,z) := \left[D_{\mu}X^{\mu}_{s,u}(x,z_1)~d_{s,u}^{\mu}(X^{\mu}_{s,u}(x),z_2)+D_{\mu}X^{\mu}_{s,u}(x,z_2)~d_{s,u}^{\mu}(X^{\mu}_{s,u}(x),z_1)\right] \end{array}$$ We are now in position to state the main result of this section. \[theo-ae-taylor\] For any $x\in{\mathbb{R}}^d$, $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ and $s\leq t$ we have the almost sure expansion $$\label{ae-taylor} \begin{array}{l} \displaystyle X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)\\ \\ \displaystyle=\int (\mu_1-\mu_0)(dy)~D_{\mu_0}X^{\mu_0}_{s,t}(x,y)+\frac{1}{2}~~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~D^2_{\mu_0}X^{\mu_0}_{s,t}(x,z) + \Delta^{[3],\mu_0,\mu_1}_{s,t}(x) \end{array}$$ with a third order remainder function $\Delta^{[3],\mu_1,\mu_0}_{s,t}$ such that $$\Vert \Delta^{[3],\mu_0,\mu_1}_{s,t}\Vert\leq c~e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^3\quad \mbox{for some $\lambda>0$}$$ The proof of the above theorem is provided in the appendix, on page . [In the remainder term estimates presented in the above theorems, we can choose any parameter $\lambda>0$ such that $\lambda<\lambda_{1,2}$, with the parameter $\lambda_{1,2}$ introduced in (\[def-lambda-1-2\])]{}. ### Interacting diffusions {#ips-sec} For any $N\geq 2$, the $N$-mean field particle interpretation associated with a collection of generators $L_{t,\eta}$ is defined by the Markov process $\xi_t=\left(\xi_t^i\right)_{1\leq i\leq N}\in ({\mathbb{R}}^d)^N$ with generators $ \Lambda_t$ given for any sufficiently smooth function $F$ and any $x=(x^i)_{1\leq i\leq N}\in ({\mathbb{R}}^d)^N$ by $$\label{ref-mean-field-FK} \Lambda_t(F)(x)=\sum_{1\leq i\leq N}~L_{t,m(x)}(F_{x^{-i}})(x^i)$$ with the function $$F_{x^{-i}}(y):=F\left(x^1,\ldots,x^{i-1},y,x^{i+1},\ldots,x^N\right)\quad \mbox{\rm and the measure}\quad m(x)=\frac{1}{N}~\sum_{1\leq i\leq N}~\delta_{x^i}$$ We extend $L_{t,\mu}$ to symmetric functions $F(x^1,x^2)$ on ${\mathbb{R}}^{2d}$ by setting $$L_{t,\mu}^{(2)}(F)(x^1,x^2):=L_{t,\mu}(F(x^1,{\mbox{\LARGE .}}))(x^2)+L_{t,\mu}(F({\mbox{\LARGE .}},x^2))(x^1)$$ In this notation, in our context we readily check that $$\label{backward-D-phi-mf} \begin{array}{rclcrcl} { {\cal F }}(x)&=&m(x)(f)&\Longrightarrow&\displaystyle \Lambda_t({ {\cal F }})(x)&=& m(x)L_{t,m(x)}(f)\\ && &&&&\\ { {\cal F }}(x)&=&m(x)^{\otimes 2}(F)&\Longrightarrow&\displaystyle \Lambda_t({ {\cal F }})(x)&=& \displaystyle m(x)^{\otimes 2}L_{t,m(x)}^{(2)}(F)+\frac{1}{N}~m(x)\left[\Gamma(F)\right] \end{array}$$ for any symmetric function $F(x^1,x^2)=F(x^2,x^1)$, with the function $\Gamma(F)$ on ${\mathbb{R}}^d$ defined for any $y\in{\mathbb{R}}^d$ by the formula $$\begin{array}{l} \displaystyle\Gamma(F)(y) :=\tr\left(\left(\left[\nabla\otimes\nabla\right]F\right)(y,y)\right)=\sum_{1\leq i\leq d}~\left(\partial_{x_1^i}\,\partial_{x_2^i}\, F\right)(y,y)\\ \\ \displaystyle\Longrightarrow \Gamma\left(f\otimes g\right)(y)=\sum_{1\leq k\leq d}\partial_{y_k}f(y)~\partial_{y_k}g(y)=\tr\left(\nabla f(y)\nabla g(y)^{\prime}\right) \end{array}$$ A proof of the above formula is provided in the appendix, on page . Applying Ito’s formula, for any smooth function $g:t\in [0,\infty[\mapsto g_t\in { {\cal C }}^2_b({\mathbb{R}}^d)$ we prove that $$m_t:=m(\xi_t)\Longrightarrow dm_t(g_t)=\left[m_t\left(\partial_tg_t\right)+m_tL_{t,m_t}(g_t)\right]~dt+\frac{1}{\sqrt{N}}~dM_t(g)$$ In the above display, $g\mapsto M_t(g)$ stands for a martingale random field with angle bracket $$\partial_t\langle M(f),M(g)\rangle_t:=m_t\left(\Gamma(f\otimes g)\right)\Longrightarrow \partial_t\langle M(g)\rangle_t=\int~m_t(dx)~\Vert \nabla g(x)\Vert^2$$ The above evolution equation is rather standard in mean field type interacting particle system theory, a detailed proof can be found in [@dm-geom-2013] (see for instance section 4.3). In the same vein, with some obvious abusive notation, using (\[backward-D-phi-mf\]) we have $$\begin{aligned} dm_s^{\otimes 2}(F)&=& \left[m_s\otimes dm_s+ dm_s\otimes m_s +(dm_s\otimes dm_s)\right](F)\\ &=& \left[m_s^{\otimes 2}L_{s,m_s}^{(2)}(F)+\frac{1}{N}~m_s\left[\Gamma(F)\right]\right]~ds+\mbox{\rm martingale increment}\\ &&\Longrightarrow \left[dm_s\otimes dm_s\right](F)=\frac{1}{N}~m_s\left[\Gamma(F)\right]~ds\end{aligned}$$ We fix a final time horizon $t\geq 0$ and we denote by $$s\in [0,t]\mapsto M_{s}\left(D_{m_{{\mbox{\LARGE .}}}}\phi_{{\mbox{\LARGE .}},t}(f)\right)$$ the martingale associated with the predictable function $$s\in [0,t]\mapsto g_s=D_{m_s}\phi_{s,t}(f)$$ Combining the Itô formula with the tensor product formula (\[backward-D-phi-mf\]) and with the backward formula (\[ref-backward-eq\]) we obtain $$d\,\phi_{s,t}(m_s)(f)=-m_s L_{s,m_s}\left(D_{m_s}\phi_{s,t}(f)\right)~ds+(dm_s)\left(D_{m_s}\phi_{s,t}(f)\right)+\frac{1}{2}~(dm_s\otimes dm_s)(D^2_{m_s}\phi_{s,t}(f))~ds$$ This implies that $$d\,\phi_{s,t}(m_s)(f)=\frac{1}{2}~(dm_s\otimes dm_s)(D^2_{m_s}\phi_{s,t}(f))~ds+ \frac{1}{\sqrt{N}}~dM_{s}\left(D_{m_{{\mbox{\LARGE .}}}}\phi_{{\mbox{\LARGE .}},t}(f)\right)$$ This yields the following theorem. \[theo-ag-mf\] For any time horizon $t\geq 0$, the interpolating semigroup $s\in [0,t]\mapsto\phi_{s,t}(m_s)$ satisfies for any $f\in { {\cal C }}^2({\mathbb{R}}^d)$ with $\sup_{k=1,2}\Vert \nabla^kf\Vert\leq 1$ the evolution equation $$d\,\phi_{s,t}(m_s)(f)=\frac{1}{2N}~m_s\left[\Gamma\left(D^2_{m_s}\phi_{s,t}(f)\right)\right]~ds+ \frac{1}{\sqrt{N}}~dM_{s}\left(D_{m_{{\mbox{\LARGE .}}}}\phi_{{\mbox{\LARGE .}},t}(f)\right)$$ The above theorem can be seen as a second order extension of the Alekseev-Gröbner lemma [@alekseev; @grobner] to nonlinear measure valued and stochastic semigroups. This result also extends the perturbation theorem obtained in [@mp-18] (cf. theorem 3.6) in the context of interacting jumps processes to McKean-Vlasov diffusions. The discrete time version of the backward perturbation analysis described above can also be found in [@dm-g-99; @guionnet; @dm-2000] in the context of Feynman-Kac particle models (see also [@d-2004; @d-2013; @dm-2003]). We end this section with some direct consequences of the above theorem. Firstly, using (\[est-nabla-D-intro\]) and (\[estimate-D2-nabla\]) we have the almost sure estimates $$\begin{aligned} \vert\partial_s\langle M_{{\mbox{\LARGE .}},t}\left(D_{m_{{\mbox{\LARGE .}}}}\phi_{{\mbox{\LARGE .}},t}(f)\right) \rangle_s\vert&\leq &c~e^{-2\lambda(t-s)}~\Vert\nabla f\Vert^2 \\ \mbox{\rm and}\quad\Vert m_s\left[\Gamma\left(D^2_{m_s}\phi_{s,t}(f)\right)\right]\Vert&\leq& c~e^{-\lambda(t-s)}~\sup_{i=1,2}\Vert \nabla^i f\Vert\quad \mbox{for some $\lambda>0$}\end{aligned}$$ Without further work, the above inequality yields the uniform bias estimate stated in the r.h.s. of (\[bias-intro-est\]), for any twice differentiable function $f$ with bounded derivatives. Using well known martingale concentration inequalities (cf. for instance lemma 3.2 in [@nishi]), there exists some finite parameter $c$ such that for any $t\geq 0$ and any $\delta\geq 1$ the probability of the following event $$\vert m_t(f)- \phi_{0,t}(m_0)(f)-\frac{1}{2N}\int_0^t~m_s\left[\Gamma\left(D^2_{m_s}\phi_{s,t}(f)\right)\right]~ds\vert\leq c~\sqrt{\frac{\delta}{N}}$$ is greater than $1-e^{-\delta}$. In addition, using the Burkholder-Davis-Gundy inequality, for any $n\geq 1$ we obtain the time uniform estimates stated in the r.h.s. of (\[ln-est\]). On the other hand, using (\[a-1\]) and (\[est-nabla-D-intro\]) we have the almost sure exponential contraction inequality $$\WW_1(\phi_{0,t}(m_0),\phi_{0,t}(\mu_0))\leq c~e^{-\lambda t}~\WW_1(m_0,\mu_0)\quad \mbox{for some $\lambda>0$}$$ This yields the bias estimates $$\vert{\mathbb{E}}\left[m_t(f)-\phi_{0,t}(\mu_0)(f)\right]\vert\leq \frac{c_1}{N}+\frac{c_2}{N^{1/d}}~e^{-\lambda t}~$$ for any twice differentiable function $f$ with bounded derivatives. The r.h.s. estimate comes from well known estimates of the average of the Wassertein distance for occupation measures, see for instance [@dudley-69] and the more recent studies [@fournier; @lei-18]. The above inequality yields the following uniform bias estimate $$\sup_{t\geq \frac{d-1}{d\lambda}\log{N}}\vert{\mathbb{E}}\left[m_t(f)-\phi_{0,t}(\mu_0)(f)\right]\vert\leq \frac{c}{N}$$ Comments on the regularity conditions {#sec-H-comments-intro} ------------------------------------- We discuss in this section the regularity condition $(H)$ introduced in (\[def-HS-0\]). We illustrate these spectral conditions for linear-drift and gradient flow models. Comparisons with related conditions presented in other works are also provided. Firstly, we mention that the condition stated in (\[def-HS-0\]) has been introduced in the article [@mp-var-18] to derive several Wasserstein exponential contraction inequalities as well as uniform propagation of chaos estimates w.r.t. the time horizon. Using the log-norm triangle inequality and recalling that the log-norm is dominated by the spectral norm we check that $$\lambda_{\tiny max}(A_t(x_1,x_2)_{\tiny sym})\leq \lambda_{\tiny max}(b_t^{[1]}(x_1,x_2)_{\tiny sym})+2^{-1}~\Vert b_t^{[2]}(x_2,x_1)+b_t^{[2]}(x_1,x_2)^{\prime}\Vert_2$$ Choosing $\lambda_0$ and $\lambda_1$ as the supremum of the maximal eigenvalue functional of the matrices $ A_t(x_1,x_2)_{\tiny sym}$ and $b_t^{[1]}(x_1,x_2)_{\tiny sym}$, the Cauchy interlacing theorem (see for instance [@lancaster] on page 294) yields $ \lambda_1\geq \lambda_0\geq \lambda_{1,2} $. For linear drift functions $$\label{lin-case-intro} b_t(x_1,x_2)=B_1\,x_1+B_2\,x_2$$ the matrix $A_t(x_1,x_2)_{\tiny sym}$ reduces to the two-by-two block partitioned matrix $$\label{lin-case} A_t(x_1,x_2)_{\tiny sym}=\left[\begin{array}{cc} (B_1)_{\tiny sym}& (B_2)_{\tiny sym}\\ (B_2)_{\tiny sym}& (B_1)_{\tiny sym} \end{array}\right]\Longrightarrow \lambda_0\geq \lambda_1=- \lambda_{\tiny max}((B_1)_{\tiny sym})\quad\mbox{\rm and}\quad \Vert b^{[2]}\Vert_2= \Vert B_2\Vert_2$$ [In this situation the diffusion flow $X_{s,t}^{\mu}(x)\in {\mathbb{R}}^d$ is given by the formula $$X^{\mu}_{s,t}(x)=e^{(t-s)B_1}(x-\mu(e))+e^{(t-s)[B_1+B_2]}~\mu(e)+\int_s^te^{B_1(t-u)}~dW_u$$ In the one dimensional case we have $$B_1<0<B_2\quad \Longrightarrow\quad B_1=-\lambda_1\leq B_1+B_2=-\lambda_{1,2}=-\lambda_0$$]{} Nonlinear Langevin diffusions are associated with the drift function $$\begin{array}{l} b(x_1,x_2):=-\nabla U(x_1)-\nabla V(x_1-x_2)\\ \\ \Longrightarrow b^{[1]}(x_1,x_2)=-\nabla^2 U(x_1)-\nabla^2 V(x_1-x_2) \quad\mbox{\rm and}\quad b^{[2]}(x_1,x_2)=\nabla^2 V(x_1-x_2) \end{array}$$ some confinement type potential function $U$ (a.k.a. the exterior potential) and some interaction potential function $V$. In this context we have $$\begin{array}{l} -A_t(x_1,x_2)_{\tiny sym}=\left[\begin{array}{cc} \nabla^2 U(x_1)& 0\\ 0& \nabla^2 U(x_2) \end{array}\right]\\ \\ \hskip2cm+\left[\begin{array}{cc} \nabla^2 V(x_1-x_2)& -(\nabla^2 V(x_2-x_1)+\nabla^2 V(x_1-x_2))/2\\ -(\nabla^2 V(x_2-x_1)+\nabla^2 V(x_1-x_2))/2& \nabla^2 V(x_2-x_1) \end{array}\right] \end{array}$$ When the potential function $V$ is even and convex we have $$A_t(x_1,x_2)_{\tiny sym}\leq -\left[\begin{array}{cc} \nabla^2 U(x_1)& 0\\ 0& \nabla^2 U(x_2) \end{array}\right]$$ In the reverse angle, when the function $V$ is odd we have the formula $$A_t(x_1,x_2)_{\tiny sym}=-\left[\begin{array}{cc} \nabla^2 U(x_1)+\nabla^2 V(x_1-x_2)& 0\\ 0& \nabla^2 U(x_2)+ \nabla^2 V(x_2-x_1) \end{array}\right]$$ In both situations, condition $(H)$ is satisfied when the strength of the confinement type potential dominates the one of the interaction potential; that is when we have that $$\nabla^2 U(x_1)+\nabla^2 V(x_2)\geq \lambda_1>\Vert\nabla^2 V\Vert_{2}$$ [ The decay rate $\lambda_0$ in the $\WW_2$-contraction inequality (\[ref-stab-W-1\]) is larger than the decay rate $\lambda_{1,2}$ in the $\WW_1$-contraction inequality (\[ref-stab-W-1-1\]). In addition, the $\WW_1$-exponential stability requires that $\lambda_0$ dominates the spectral norm of the matrix $b^{[2]}$. Next we provide a more refined analysis based on the proof of the $\WW_2$-contraction inequality presented in [@mp-var-18]. Using the interpolating paths $(Y_{\epsilon},\mu_{\epsilon})$ introduced in (\[def-Y-epsilon\]) we set $$\label{def-mu-epsilon} X^{\epsilon}_{s,t}:=X^{\mu_{\epsilon}}_{s,t}(Y_{\epsilon})\quad \mbox{\rm and}\quad \overline{X}^{\epsilon}_{s,t}:=\overline{X}^{\mu_{\epsilon}}_{s,t}(\overline{Y}_{\epsilon})$$ In the above display $(\overline{X}^{\mu_{\epsilon}}_{s,t}(x),\overline{Y}_{\epsilon})$ stands for an independent copy of $(X^{\mu_{\epsilon}}_{s,t}(x),Y_{\epsilon})$. Arguing as in [@mp-var-18] we have $$\partial_t{\mathbb{E}}(\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert)={\mathbb{E}}\left[\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert^{-1}\left( \langle \partial_{\epsilon}X^{\epsilon}_{s,t},b^{[1]}(X^{\epsilon}_{s,t},\overline{X}^{\epsilon}_{s,t}) \partial_{\epsilon}X^{\epsilon}_{s,t}\rangle+ \langle \partial_{\epsilon}\overline{X}^{\epsilon}_{s,t},b^{[2]}(X^{\epsilon}_{s,t},\overline{X}^{\epsilon}_{s,t}) \partial_{\epsilon}X^{\epsilon}_{s,t}\rangle\right)\right]$$ We consider the symmetric and anti-symmetric matrices $$\begin{aligned} b^{[2]}_t(x_1,x_2)_{\tiny sym}&:=&\frac{1}{2}~ \left(b^{[2]}_t(x_1,x_2)+ b^{[2]}_t(x_2,x_1)^{\prime}\right) \\ b^{[2]}_t(x_1,x_2)_{\tiny asym}&:=&\frac{1}{2}~ \left(b^{[2]}_t(x_1,x_2)- b^{[2]}_t(x_2,x_1)^{\prime}\right) \end{aligned}$$ and we set $$(U^{\epsilon}_{s,t},\overline{U}^{\epsilon}_{s,t}):=\left(\frac{\partial_{\epsilon}X^{\epsilon}_{s,t}}{\sqrt{\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert}},\frac{\partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}}{\sqrt{\Vert \partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}\Vert}}\right) \quad \mbox{\rm and}\quad (V^{\epsilon}_{s,t},\overline{V}^{\epsilon}_{s,t}):=\left(\frac{\partial_{\epsilon}X^{\epsilon}_{s,t}}{\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert},\frac{\partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}}{\Vert \partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}\Vert}\right)$$ By symmetry arguments and using some elementary manipulations we check the formula $$\begin{array}{l} 2\, \partial_t\,{\mathbb{E}}(\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert) ={\mathbb{E}}\left(\left\langle\left( \begin{array}{c} U^{\epsilon}_{s,t}\\ \overline{U}^{\epsilon}_{s,t} \end{array} \right),A_t(X^{\epsilon}_{s,t},\overline{X}^{\epsilon}_{s,t})\left( \begin{array}{c} U^{\epsilon}_{s,t}\\ \overline{U}^{\epsilon}_{s,t} \end{array} \right)\right\rangle\right.\\ \\ \hskip3cm\left.+ \left(\sqrt{\Vert \partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}\Vert}-\sqrt{\Vert \partial_{\epsilon}{X}^{\epsilon}_{s,t}\Vert}\,\right)^2~ \left\langle \overline{V}^{\epsilon}_{s,t}, b^{[2]}_t(X^{\epsilon}_{s,t},\overline{X}^{\epsilon}_{s,t})_{\tiny sym}\,V^{\epsilon}_{s,t}\right\rangle\right.\\ \\ \hskip4cm\left.+\left(\Vert \partial_{\epsilon}\overline{X}^{\epsilon}_{s,t}\Vert-\Vert \partial_{\epsilon}{X}^{\epsilon}_{s,t}\Vert\,\right)~ \left\langle \overline{V}^{\epsilon}_{s,t}, b^{[2]}_t(X^{\epsilon}_{s,t},\overline{X}^{\epsilon}_{s,t})_{\tiny asym}\,V^{\epsilon}_{s,t}\right\rangle \,\right) \end{array}$$ This shows that $$\partial_t\,{\mathbb{E}}(\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert)\leq -\widehat{\lambda}_{1,2}~{\mathbb{E}}(\Vert \partial_{\epsilon}X^{\epsilon}_{s,t}\Vert)$$ with the parameter $\widehat{\lambda}_{1,2}$ given by $$-\widehat{\lambda}_{1,2}:=\sup_{x_1,x_2}{\left[\lambda_{\tiny max}(A_t(x_1,x_2))+\Vert b^{[2]}_t(x_1,x_2)_{\tiny sym}\Vert_2+\Vert b^{[2]}_t(x_1,x_2)_{\tiny asym}\Vert_2\right]}\leq -\lambda_{1,2}$$ We conclude that the $\WW_1$-contraction inequality (\[ref-stab-W-1-1\]) is met with $\lambda=\widehat{\lambda}_{1,2}$.]{} In a more recent article [@wang] the author presents some Wasserstein contraction inequalities of the same form as in (\[ref-stab-W-1\]) with $\lambda_0$ replaced by some parameter $\lambda^{-}_0=(\kappa_1-\kappa_2)$, under the assumption $$\langle x_1-y_1,b_t(x_1,\mu_1)-b_t(y_1,\mu_2)\rangle\leq -\kappa_1~\Vert x_1-y_1\Vert^2+\kappa_2~\WW_2(\mu_1,\mu_2)^2\quad\mbox{\rm for some}\quad \kappa_1>\kappa_2$$ Taking Dirac measures $\mu_1=\delta_{x_2}$ and $\mu_2=\delta_{y_2}$ we check that the above condition is equivalent to the fact that $$\langle x_1-y_1,b_t(x_1,x_2)-b_t(y_1,y_2)\rangle\leq -\kappa_1~\Vert x_1-y_1\Vert^2+\kappa_2~\Vert x_2-y_2\Vert^2$$ By symmetry arguments this implies that $$\label{Hilbert-condition-sym} \langle x_1-y_1,b_t(x_1,x_2)-b_t(y_1,y_2)\rangle+\langle x_2-y_2,b_t(x_2,x_1)-b_t(y_2,y_1)\rangle\leq -\lambda_0^-~[\Vert x_1-y_1\Vert^2+\Vert x_2-y_2\Vert^2]$$ For the linear drift model discussed in (\[lin-case\]) the above condition reads $$\left[\begin{array}{cc} (B_1)_{\tiny sym}& (B_2)_{\tiny sym}\\ (B_2)_{\tiny sym}& (B_1)_{\tiny sym} \end{array}\right]\leq -\lambda_0^-~I\quad \mbox{\rm which is implies that}\quad \lambda_0\geq \lambda_0^-$$ We also have $(\ref{Hilbert-condition-sym})\Longrightarrow (\ref{Hilbert-condition})$ with $\lambda=\lambda^-_0$. Comparisons with existing literature {#comparison-sec} ------------------------------------ The perturbation analysis developed in the article differs from the Otto differential calculus on $(P_2({\mathbb{R}}^d),\WW_2)$ introduced in [@otto] and further developed by Ambrosio and his co-authors [@ambrosio-1; @ambrosio-2] and Otto and Villani in [@otto-2]. These sophisticated gradient flow techniques in Wasserstein metric spaces are based on optimal transport theory. The central idea is to interpret $P_2({\mathbb{R}}^d)$ as an infinite dimensional Riemannian manifold. In this context, the Benamou-Brenier formulation of the Wasserstein distance provides a natural way to define geodesics, gradients and Hessians w.r.t. the Wasserstein distance. The details of these gradient flow techniques are beyond the scope of the semigroup perturbation analysis considered herein. This methodology is mainly used to quantify the entropy dissipation of Langevin-type nonlinear diffusions. Thus, it cannot be used to derive any Taylor expansion of the form (\[TT-intro\]) nor to analyze the stability properties of more general classes of McKean-Vlasov diffusions. Besides some interesting contact points, the methodology developed in the present article doesn’t rely on the more recent differential calculus on $(P_2({\mathbb{R}}^d),\WW_2)$ developed by P.L. Lions and his co-authors in the seminal works on mean field game theory [@cardaliaguet; @gueant]. In this context, the first order Lions differential of a smooth function from $P_2({\mathbb{R}}^d)$ into ${\mathbb{R}}$ is defined as the conventional derivative of lifted real valued function acting on the Hilbert space of square integrable random variables. In this interpretation, for a given test function, say $f$ the gradient $\nabla D_{\mu}\phi_{s,t}(f)(Y)$ of the first order differential in (\[TT-intro\]) can be seen as the Lions derivative $(\delta u_{s,t}/\delta\mu)(Y)$ of the lifted scalar function $Y\mapsto u_{s,t}(Y):={\mathbb{E}}(f(X_{s,t}^{\mu}(Y)))$, for some random variable $Y$ with distribution $\mu$. In the recent book [@carmona-delarue], to distinguish these two notions, the authors called the random variable $D_{\mu}\phi_{s,t}(f)(Y)$ the linear functional derivative. For a more thorough discussion on the origins and the recent developments in mean field game theory, we refer to the book [@carmona-delarue] as well as the more recent articles [@peng-17; @chassagneux; @crisan-18] and the references therein. To the best of our knowledge, most of the literature on Lions’ derivatives is concerned with existence theorems without a refined analysis of the exponential decays of these differentials w.r.t. the time parameter. Last but not least, from the practical point of view all differential estimates we found in the literature are rather quite deceiving since after carefully checking, they grow exponentially fast with respect to the time horizon (cf. for instance [@peng-17; @chassagneux; @chaudru; @crisan-18]). Taylor expansions of the form (\[TT-intro\]) have already been discussed in the book [@d-2013] for discrete time nonlinear measure valued semigroups (cf. for instance chapters 3 and 10). We also refer to the more recent article [@mp-18] in the context of continuous time Feynman-Kac semigroups. In this context, we emphasize that the semigroup $\phi_{s,t}(\mu)$ is explicitly given by a normalization of a linear semigroup of positive operators. Thus, a fairly simple Taylor expansion yields the second order formula (\[TT-intro\]). In contrast with Feynman-Kac models, McKean-Vlasov semigroups don’t have any explicit form nor an analytical description. As a result, none of above methodologies cannot be used to analyze nonlinear diffusions. The second order perturbation analysis discussed in this article has been used with success in [@dm-g-99; @guionnet; @dm-2000] to analyze the stability properties of Feynman-Kac type particle models, as well as the fluctuations and the exponential concentration of this class of interacting jump processes; see also [@dmrio-09; @dm-hu-wu] for general classes of discrete generation mean field particle systems, a well as chapter 7 in [@d-2004] and [@mp-18; @dm-2003] for continuous time models. These second order perturbation techniques have also been extended in the seminal book by V.N. Kolokoltsov [@kolo-10] to general classes of nonlinear Markov processes and kinetic equations. Chapter 8 in [@kolo-10] is dedicated to the analysis of the first and the second order derivatives of nonlinear semigroups with respect to initial data. The use of the first and the second order derivatives in the analysis of central limit theorems and propagation of chaos properties respectively is developed in Chapters 9 and Chapter 10 in [@kolo-10]. We underline that these results are obtained for diffusion processes as well as for jump-type processes and their combinations, see also [@kol-troeva-1; @kol-troeva-2]. Nevertheless none of these studies apply to derive non asymptotic Taylor expansions (\[taylor-2-intro\]) and (\[ae-taylor\]) with exponential decay-type remainder estimates for McKean-Vlasov diffusions nor to estimate the stability properties of the associated semigroups. In addition, to the best of our knowledge the stochastic perturbation theorem \[theo-ag-mf\] is the first result of this type for mean field type interacting diffusions. Last but not least, the idea of considering the flow of empirical measures $m(\xi_t)$ of a mean field particle model as a stochastic perturbation of the limiting flow $\phi_{0,t}(\mu_0)$ certainly goes back to the work by Dawson [@dawson], itself based on the martingale approach developed by Papanicolaou, Stroock and Varadhan in [@pap-77], published in the end of the 1970’s. These two works are mainly centered on fluctuation type limit theorems. They don’t discuss any Taylor expansion on the limiting semigroup $\phi_{s,t}$ nor any question related to the stability properties of the underlying processes. Some preliminary results {#spr-sec} ======================== The first part of this section provides a review of tensor product theory and Fréchet differential on Hilbert spaces. Section \[not-tensor-sec\] is concerned with conventional tensor products and Fréchet derivatives. Section \[tensor-integral-op-sec\] provides a short introduction to tensor integral operators. In the second part of this section we review some basic tools of the theory of stochastic variational equations, including some differential properties of Markov semigroups. Section \[var-eq-sec\] is dedicated to variational equations. Section \[bel-formula-sec\] discusses Bismut-Elworthy-Li extension formulae. We also provide some exponential inequalities for the gradient and the Hessian operators on bounded measurable functions. The differential operator arising in the Taylor expansions (\[TT-intro\]) are defined in terms of tensor integral operators that depend on the gradient of the drift function $b_t(x_1,x_2)$ of the nonlinear diffusion. These integro-differential operators are described in section \[integro-diff-sec\]. The last section, section \[s-diff-form-sec\] provides some differential formulae as well as some exponential decays estimates of the norm of these operators w.r.t. the time horizon. Fréchet differential {#not-tensor-sec} -------------------- We let $[n]$ stands for the set of $n$ multiple indexes $i=(i_1,\ldots,i_n)\in { {\cal I }}^n$ over some finite set ${ {\cal I }}$. Notice that $[n_1]\times[n_2]=[n_1+n_2]$. We denote by ${ {\cal T}}_{p,q}({ {\cal I }})$ the space of $(p,q)$-tensor $X$ with real entries $(X_{i,j})_{(i,j)\in [p]\times [q]}$. Given a $(p_1,q_1)$-tensor $X$ and a $(p_2,q_2)$-tensor $Y$ we denote by $( X\otimes Y)$ the $((p_1+q_1),(p_2+q_2))$-tensor defined by $$( X\otimes Y)_{(i,j),(k,l)}:=X_{i,k}~Y_{j,l}$$ For a given $(p_1,q)$-tensor $X$ and a given $(q,p_2)$ tensor $Y$, the product $XY$ and the transposition $Y^{\prime}$ are the $(p_1,p_2)$ and $(p_2,q)$ tensors with entries $$\forall (i,j)\in [p_1]\times [p_2]\qquad (XY)_{i,j}:=\sum_{k\in [q]}X_{i,k}Y_{k,j}\quad \mbox{\rm and}\quad Y^{\prime}_{j,k}=Y_{k,j}$$ We equip ${ {\cal T}}_{p,q}({ {\cal I }})$ with the Frobenius inner product $$\langle X,Y\rangle:=\tr(XY^{\prime}):=\sum_{i\in [p]}(XY^{\prime})_{i,i}\quad \mbox{\rm and the norm}\quad \Vert X\Vert_{\tiny Frob}:=\sqrt{\tr(XX^{\prime})}$$ Identifying $(1,0)$-tensors ${ {\cal T}}_{1,0}({ {\cal I }})={\mathbb{R}}^{{ {\cal I }}}$ with column vectors $(X_{i})_{i\in { {\cal I }}}\in {\mathbb{R}}^{{ {\cal I }}}$ the above quantities coincide with the conventional Euclidian inner product and norm on the product space ${\mathbb{R}}^{{ {\cal I }}}$. When ${ {\cal I }}=\{1,\ldots,d\}$ we simplify notation and we set ${\mathbb{R}}^{d}$ instead of ${\mathbb{R}}^{\{1,\ldots,d\}}$. For any tensors $X$ and $Y$ with appropriate dimensions, using Cauchy-Schwartz inequality we check that $$\langle X,Y\rangle^2\leq \Vert X\Vert_{\tiny Frob}~\Vert Y\Vert_{\tiny Frob}\quad \mbox{\rm and}\quad \Vert XY\Vert_{\tiny Frob}\leq \Vert X\Vert_{\tiny Frob}~\Vert Y\Vert_{\tiny Frob}$$ Let ${\mathbb{H}}({ {\cal T}}_{p,q}({ {\cal I }})):={\mathbb{L}}_2((\Omega,{\mathbb{F}},{\mathbb{P}}),{ {\cal T}}_{p,q}({ {\cal I }}))$ be the Hilbert space of ${ {\cal T}}_{p,q}({ {\cal I }})$-valued random variables defined on some probability space $(\Omega,{\mathbb{F}},{\mathbb{P}})$, equipped with the inner product $$\langle X,Y\rangle_{\,{\mathbb{H}}({ {\cal T}}_{p,q}({ {\cal I }}))}={\mathbb{E}}(\langle X,Y\rangle)\quad\mbox{\rm and the norm}\quad \Vert X\Vert_{\,{\mathbb{H}}({ {\cal T}}_{p,q}({ {\cal I }}))}:=\langle X,X\rangle_{\,{\mathbb{H}}({ {\cal T}}_{p,q}({ {\cal I }}))}^{1/2}$$ induced by the inner product $\langle X,Y\rangle$ on ${ {\cal T}}_{p,q}({ {\cal I }})$. We denote by ${\mathbb{E}}(X)={\mathbb{E}}(X_{i,j})_{(i,j)\in [p]\times [q]}$ the entry-wise expected value of a $(p,q)$-tensor. When ${ {\cal I }}=\{1,\ldots,d\}$ and $(p,q)=(1,0)$ the space ${\mathbb{H}}({ {\cal T}}_{p,q}({ {\cal I }}))$ coincides with be the Hilbert space ${\mathbb{H}}({\mathbb{R}}^d)={\mathbb{L}}_2((\Omega,{\mathbb{F}},{\mathbb{P}}),{\mathbb{R}}^d)$ of square integrable ${\mathbb{R}}^d$-valued and ${\mathbb{F}}$-measurable random variables. We denote by $${\mathbb{H}}_n({ {\cal T}}_{p,q}({ {\cal I }})):={\mathbb{L}}_2((\Omega,{\mathbb{F}}_n,{\mathbb{P}}),{ {\cal T}}_{p,q}({ {\cal I }}))$$ the non decreasing sequence of Hilbert spaces associated with some increasing filtration ${\mathbb{F}}_n\subset {\mathbb{F}}_{n+1}$. In Landau notation, we recall that a function $$F:X\in{\mathbb{H}}_1({ {\cal T}}_{p_1,q_1}({ {\cal I }}))~\mapsto ~F(X)\in{\mathbb{H}}_2({ {\cal T}}_{p_2,q_2}({ {\cal J }}))$$ is said to be Fréchet differentiable at $X$ if there exists a continuous map $$X\in{\mathbb{H}}_1({ {\cal T}}_{p,q}({ {\cal I }}))~\mapsto ~ \partial F(X)\in \mbox{\rm Lin}({\mathbb{H}}_1({ {\cal T}}_{p_1,q_2}({ {\cal I }})),{\mathbb{H}}_2({ {\cal T}}_{p_2,q_2}({ {\cal J }})))$$ such that $$F(X+Y)=F(X)+\partial F(X)\cdot Y+\mbox{\rm o}\left(Y\right)$$ Tensor integral operators {#tensor-integral-op-sec} ------------------------- Let ${ {\cal B }}(E,{ {\cal T}}_{p,q}({ {\cal I }}))$ be the set of bounded measurable functions from a measurable space $E$ into some tensor space ${ {\cal T}}_{p,q}({ {\cal I }})$. Signed measures $\mu$ on $E$ act on bounded measurable functions $g$ from $E$ into ${\mathbb{R}}$. We extend these integral operators to tensor valued functions $g=(g_{i,j})_{(i,j)\in [p]\times [q])}\in{ {\cal B }}(E,{ {\cal T}}_{p,q}({ {\cal I }}))$ by setting for any $(i,j)\in [p]\times [q]$ $$\mu(g)_{i,j}= \mu(g_{i,j}):=\int~\mu(dx)~g_{i,j}(x)\quad \mbox{\rm and we set}\quad\mu(g): =\int~\mu(dx)~g(x)$$ Let $(E,{ {\cal E }})$ and $(F,{ {\cal F }})$ be some pair of measurable spaces. A $(p,q)$-tensor integral operator $${ {\cal Q }}~:~g\in{ {\cal B }}(F,{ {\cal T}}_{q,r}({ {\cal I }}))\mapsto { {\cal Q }}(g)\in{ {\cal B }}(E,{ {\cal T}}_{p,r}({ {\cal I }}))$$ is defined for $r\geq 0$ and $g\in { {\cal B }}({ {\cal F }},{ {\cal T}}_{q,r}({ {\cal I }}))$ by the tensor valued and measurable function ${ {\cal Q }}(g)$ with entries given $x\in E$ and $(i,j)\in ([p]\times [r])$ by the integral formula $${ {\cal Q }}(g)_{i,j}(x)=\sum_{k\in [q]}~\int_{F}~{ {\cal Q }}_{i,k}(x,d\overline{x})~g_{k,j}(\overline{x})$$ for some collection of integral operators ${ {\cal Q }}_{i,k}(x_1,dx_2)$ from ${ {\cal B }}(E,{\mathbb{R}})$ into ${ {\cal B }}(F,{\mathbb{R}})$. We also consider the operator norm $${{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal Q }}\vert\kern-0.25ex\vert\kern-0.25ex\vert}}:=\sup_{\Vert g\Vert\leq 1}{\Vert { {\cal Q }}(g)\Vert}\quad \mbox{\rm for some tensor norm $\Vert{\mbox{\LARGE .}}\Vert$}$$ The tensor product $({ {\cal Q }}^1\otimes { {\cal Q }}^2)$ of a couple of $(p_i,q_i)$-tensor integral operators $${ {\cal Q }}^i~:~g\in{ {\cal B }}(F_i,{ {\cal T}}_{q_i,r_i}({ {\cal I }}))\mapsto { {\cal Q }}(g)\in{ {\cal B }}(E_i,{ {\cal T}}_{p_i,r_i}({ {\cal I }}))\quad \mbox{\rm with}\quad i=1,2$$ is a $(p,q)$-tensor integral operator $${ {\cal Q }}^1\otimes { {\cal Q }}^2~:~h\in{ {\cal B }}(F,{ {\cal T}}_{q,r}({ {\cal I }}))\mapsto { {\cal Q }}(g)\in{ {\cal B }}(E,{ {\cal T}}_{p,q}({ {\cal I }}))$$ with the product spaces $$(E,F):=(E_1\times E_2,F_1\times F_2)\quad \mbox{\rm and}\quad ( p,q,r)=(p_1+p_2,q_1+q_2, r_1+r_2)$$ The entries of $({ {\cal Q }}^1\otimes { {\cal Q }}^2)(h)$ are given for any $x=(x_1,x_2)$ and any pair of multi-indices $i=(i_1,i_2)\in ([p_1]\times [p_2])$, $j=(j_1,j_2) \in ([r_1]\times [r_2])$ by the integral formula $$({ {\cal Q }}^1\otimes { {\cal Q }}^2)(h)_{i,j}(x)=\sum_{k\in ( [q_1]\times[q_2])}~\int_{F_1\times F_2}~({ {\cal Q }}^1\otimes { {\cal Q }}^2)_{i,k}(x,dy)~h_{k,j}(y)$$ with the tensor product measures defined for any $k=(k_1,k_2)\in ([q_1]\times [q_2])$ and any $y=(y_1,y_2)$ by $$({ {\cal Q }}^1\otimes { {\cal Q }}^2)_{(i_1,i_2),(k_1,k_2)}((x_1,x_2),d(y_1,y_2)):= { {\cal Q }}^1_{i_1,k_1}(x_1,dy_1)~{ {\cal Q }}^2_{i_2,k_2}(x_2,dy_2)$$ Variational equations {#var-eq-sec} --------------------- The gradient and the Hessian of a multivariate smooth function $h(x)=(h_{i}(x))_{i\in [p]}$ is defined by the $(1,p)$ and $(2,p)$ tensors $\nabla h(x)$ and $\nabla^2 h(x)$ with entries given for any $1\leq k,l\leq d$ and $i\in [p]$ by the formula $$\label{grad-def} \nabla h(x)_{k,i}=\partial_{x_k}h_{i}(x)\quad \mbox{\rm and}\quad \nabla^2 h(x)_{(k,l),i}=\partial_{x_k}\partial_{x_l}h_{i}(x)$$ We consider the tensor valued functions $ b_t^{[k_1,k_2]}$ and $b_t^{[k_1,k_2,k_3]}$ defined for any $k_1,k_2,k_3=1,2$ by $$b_t^{[k_1,k_2]}:=(\nabla_{x_{k_1}}\otimes\nabla_{x_{k_2}})b_t\quad\mbox{and}\quad b_t^{[k_1,k_2,k_3]}:=(\nabla_{x_{k_1}}\otimes\nabla_{x_{k_2}}\otimes\nabla_{x_{k_3}})b_t$$ with the $(2,1)$ and $(3,1)$-tensor valued functions $$\left( b_t^{[k_1,k_2]}\right)_{(i_1,i_2),j}=\partial_{x_{k_1}^{i_1}}\partial_{x_{k_2}^{i_2}}b^j_t \quad\mbox{\rm and}\quad \left( b_t^{[k_1,k_2,k_3]}\right)_{(i_1,i_2,i_3),j}=\partial_{x_{k_1}^{i_1}}\partial_{x_{k_2}^{i_2}}\partial_{x_{k_3}^{i_3}}b^j_t$$ In the above display, $\partial_{x_{k}^{i}}b^j_t(x_1,x_2)$ stands for the partial derivative of the scalar function $b_t^j(x_1,x_2)$ w.r.t. the coordinate $x_k^i$, with the drift function $b_t(x_1,x_2)$ from ${\mathbb{R}}^{2d}$ into ${\mathbb{R}}^d$ introduced in section \[ref-desciption\], In the same vein, $\partial_{x_{k_1}^{i_1}}\partial_{x_{k_2}^{i_2}}b^j_t(x_1,x_2)$ and $\partial_{x_{k_1}^{i_1}}\partial_{x_{k_2}^{i_2}}\partial_{x_{k_3}^{i_3}}b^j_t(x_1,x_2)$ stands for the second and third partial derivatives of $b_t^j(x_1,x_2)$ w.r.t. the coordinates $x_{k_1}^{i_1}$, $x_{k_2}^{i_2}$ and $x_{k_3}^{i_3}$ with $k_1,k_2,k_3\in \{1,2\}$. For any $\mu\in P_2({\mathbb{R}}^d)$ and $x_1\in{\mathbb{R}}^d$ we also consider the tensor functions $$b_t^{[1]}(x_1,\mu)_{i,j}:=\int\mu(dx_2)~\partial_{x_1^i} b_t^j(x_1,x_2)\qquad b_t^{[1,1]}(x_1,\mu)_{(i_1,i_2),j}:=\int\mu(dx_2)~\partial_{x_{1}^{i_1}}\partial_{x_{1}^{i_2}}b^j_t(x_1,x_2)$$ Recalling that $b_t(x,\phi_{s,t}(\mu))$ has continuous and uniformly bounded derivatives up to the third order, the stochastic flow $ x\mapsto X_{s,t}^{\mu}(x)$ is a twice differentiable function of the initial state $x$. In addition, when $(H)$ holds the gradient $\nabla X^{\mu}_{s,t}(x)$ of the diffusion flow $X^{\mu}_{s,t}(x)$ satifies the $(d\times d)$-matrix valued stochastic diffusion equation $$\label{tau-1-estimates} \partial_t \,\nabla X^{\mu}_{s,t}(x)=\nabla X^{\mu}_{s,t}(x)~b_t^{[1]}\left(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu)\right)~\Longrightarrow~ \Vert\nabla X^{\mu}_{s,t}(x)\Vert_2\leq e^{-\lambda_1(t-s)}$$ The above estimate is a direct consequence of well known log-norm estimates for exponential semigroups, see for instance [@coppel] as well as section 1.3 in the recent article [@Bishop/DelMoral:2019]. We have the stochastic tensor evolution equation $$\begin{array}{l} \partial_t\, \nabla^2 X^{\mu}_{s,t}(x)\\ \\ \displaystyle=\nabla^2 X^{\mu}_{s,t}(x)~b_t^{[1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu)) +\left[\nabla X^{\mu}_{s,t}(x)\otimes \nabla X^{\mu}_{s,t}(x)\right]~b_t^{[1,1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu))\end{array}$$ [This implies that $$\partial_t\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert^2_{\tiny Frob}\leq~-2\lambda_1~\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^2+2\Vert b^{[1,1]}\Vert_{\tiny Frob}~\Vert\nabla X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^2~\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}$$ from which we check that $$\partial_t\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\leq~-\lambda_1~\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}+\Vert b^{[1,1]}\Vert_{\tiny Frob}~\Vert\nabla X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^2$$ Using (\[tau-1-estimates\]), this yields the estimate $$\label{tau-2-estimates} \Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\leq c_1~e^{-\lambda_1(t-s)}~\int_s^t~e^{\lambda_1(u-s)}~\Vert\nabla X^{\mu}_{s,u}(x)\Vert_{\tiny Frob}^2~du\leq c_2~e^{-\lambda_1(t-s)}$$ More generally, using the multivariate version of the de Faà di Bruno derivation formula [@ma-2009] (see also formula (\[faa-di-bruno-ap\]) in the appendix), for any $n\geq 1$ we also check the uniform estimate $$\label{tau-n-estimates} \Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\leq c_{n}~e^{-\lambda_1(t-s)}$$]{} A detailed proof is provided in the appendix, on page . Differential of Markov semigroups --------------------------------- We have the commutation formula $$\label{commutation-tensor} \nabla \circ P_{s,t}^{\mu}={ {\cal P }}_{s,t}^{\mu}\circ \nabla$$ with the $(1,1)$-tensor integral operator ${ {\cal P }}_{s,t}^{\mu}$ defined for any $x\in {\mathbb{R}}^d$ and any differentiable function $f$ on ${\mathbb{R}}^d$ by the formula $$\label{def-Pa} { {\cal P }}_{s,t}^{\mu}(\nabla f)(x):= {\mathbb{E}}\left[\nabla X_{s,t}^{\mu}(x)~\nabla f(X_{s,t}^{\mu}(x))\right]$$ The tensor product of ${ {\cal P }}_{s,t}^{\mu}$ is also given by the $(2,2)$-tensor integral operator $$\left({ {\cal P }}_{s,t}^{\mu}\right)^{\otimes 2}(h)(x_1,x_2):= {\mathbb{E}}\left[\left[\nabla X_{s,t}^{\mu}(x_1)\otimes\nabla \overline{X}_{s,t}^{\mu}(x_2)\right]~h\left(X_{s,t}^{\mu}(x_1),\overline{X}_{s,t}^{\mu}(x_2)\right)\right]$$ In the above display, $\overline{X}_{s,t}^{\mu}(x)$ stands for an independent copy of $X_{s,t}^{\mu}(x)$ and $h=(\nabla\otimes \nabla) g$ stands for the matrix valued function defined in (\[def-nabla-nabla\]). We also have the commutation formula $$\left({ {\cal P }}_{s,t}^{\mu}\right)^{\otimes 2}\circ (\nabla\otimes \nabla)=(\nabla\otimes \nabla)\circ \left(P^{\mu_0}_{s,t}\right)^{\otimes 2}$$ In the same vein, we have the second order differential formula $$\label{ref-nabla-Pa} \nabla^2P^{\mu}_{s,t}(f)={ {\cal P }}^{[2,1],\mu}_{s,t}(\nabla f)+{ {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)$$ with the $(2,1)$ and $(2,2)$-tensor integral operators $$\begin{aligned} { {\cal P }}^{[2,1],\mu}_{s,t}(\nabla f)(x)&:=& {\mathbb{E}}\left[\nabla^2 X_{s,t}^{\mu}(x)~\nabla f(X_{s,t}^{\mu}(x))\right]\nonumber\\ { {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)(x)&:=& {\mathbb{E}}\left[\left(\nabla X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\right)~\nabla^2 f(X_{s,t}^{\mu}(x))\right] \label{def-U-V}\end{aligned}$$ Iterating the above procedure, the $n$-th differential of $P^{\mu}_{s,t}(f)$ at any order $n\geq 1$ takes the form $$\nabla^nP^{\mu}_{s,t}(f)=\sum_{1\leq k\leq n}{ {\cal P }}^{[n,k],\mu}_{s,t}(\nabla^k f)$$ for some integral operators ${ {\cal P }}^{[n,k],\mu}_{s,t}$. For instance, we have the third order differential formula $$\label{ref-nabla-3-Pa} \nabla^3P^{\mu}_{s,t}(\nabla f)={ {\cal P }}^{[3,1],\mu}_{s,t}(\nabla f)+{ {\cal P }}^{[3,2],\mu}_{s,t}(\nabla^2 f)+{ {\cal P }}^{[3,3],\mu}_{s,t}(\nabla^3 f)$$ with the $(2,1)$ and $(2,2)$-tensor integral operators $$\begin{aligned} { {\cal P }}^{[3,1],\mu}_{s,t}(\nabla f)(x)&:=& {\mathbb{E}}\left[\nabla^3 X_{s,t}^{\mu}(x)~\nabla f(X_{s,t}^{\mu}(x))\right]\nonumber\\ { {\cal P }}^{[3,2],\mu}_{s,t}(\nabla^2 f)(x)&:=& {\mathbb{E}}\left[\left(\nabla^2 X_{s,t}^{\mu}(x)\frownotimes \nabla X_{s,t}^{\mu}(x)\right)~\nabla^2 f(X_{s,t}^{\mu}(x))\right]\nonumber\\ { {\cal P }}^{[3,3],\mu}_{s,t}(\nabla^3 f)(x)&:=&{\mathbb{E}}\left[\left(\nabla X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\right)~\nabla^3 f(X_{s,t}^{\mu}(x))\right] \label{def-nabla-3}\end{aligned}$$ with the $\frownotimes$-tensor product of type $(3,2)$ given for any $i=(i_1,i_2,i_3)$ and $l=(l_1,l_2)$ by $$\begin{array}{l} \left(\nabla^2 X_{s,t}^{\mu}(x)\frownotimes \nabla X_{s,t}^{\mu}(x)\right)_{i,l} := \left(\nabla^2 X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\right)_{((i_1,i_2),i_3),l}\\ \\ \hskip3cm+ \left(\nabla^2 X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\right)_{((i_2,i_3),i_1),l} +\left(\nabla^2 X_{s,t}^{\mu}(x)\otimes \nabla X_{s,t}^{\mu}(x)\right)_{((i_3,i_1),i_2),l} \end{array}$$ The above formulae remains valid for any column vector multivariate function $f=(f_{i})_{1\leq i\leq d}$. An explicit description of the integral operators ${ {\cal P }}^{[n,k],\mu}_{s,t}$ for any $1\leq k\leq n$ can be obtained using multivariate derivations and combinatorial manipulations, see for instance the multivariate version of the de Faà di Bruno derivation formulae (\[faa-di-bruno-ap\]) and (\[faa-di-bruno-ap-2\]) in the appendix. Following the proof of (\[tau-n-estimates\]) we also check the uniform estimates $$\label{estimates-U-V} \sup_{1\leq k\leq n}{{{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal P }}^{[n,k],\mu}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}}\leq c_n~e^{-\lambda_1(t-s)}$$ Using the moment estimates (\[ref-moments\]) for any $\mu\in P_2({\mathbb{R}}^d)$, $m,n\geq 0$, and any $s\leq t$, we also check the rather crude estimate $$\label{crude-est-rev} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert P^{\mu}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{ {\cal C }}^n_m({\mathbb{R}}^d)\rightarrow { {\cal C }}^n_m({\mathbb{R}}^d)}\vee {{\vert\kern-0.25ex\vert\kern-0.25ex\vert (P^{\mu}_{s,t})^{\otimes 2} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{ {\cal C }}^n_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_m({\mathbb{R}}^{2d})}\leq c_{m,n}(t)~\left[1+\Vert e\Vert_{\mu,2}\right]^{m}$$ For instance, using the de Faà di Bruno derivation formula (\[faa-di-bruno-ap-2\]) for any function $f\in{ {\cal C }}^n_m({\mathbb{R}}^d)$ such that $\Vert f\Vert_{{ {\cal C }}^n_m({\mathbb{R}}^d)}\leq 1$ and for any $0\leq k\leq n$ we check that $$\Vert\nabla^kP^{\mu}_{s,t}(f)(x)\Vert=\Vert{\mathbb{E}}\left(\nabla^k(f\circ X^{\mu}_{s,t})(x)\right)\Vert\leq c_{n,m}(t)~{\mathbb{E}}\left((1+\Vert X_{s,t}^{\mu}(x)\Vert)^m\right)$$ The estimates (\[ref-moments\]) implies that $$\Vert\nabla^kP^{\mu}_{s,t}(f)(x)\Vert\leq c_{n,m}(t)~\left(\Vert x\Vert+\Vert e\Vert_{\mu,2}\right)^m\leq c_{n,m}(t)~(1+\Vert x\Vert)^m~ \left(1\vee \Vert e\Vert_{\mu,2}\right)^m$$ from which we conclude that $${{\vert\kern-0.25ex\vert\kern-0.25ex\vert P^{\mu}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}_{{ {\cal C }}^n_m({\mathbb{R}}^d)\rightarrow { {\cal C }}^n_m({\mathbb{R}}^d)}\leq c_{m,n}(t)~\left[1+\Vert e\Vert_{\mu,2}\right]^{m}$$ Bismut-Elworthy-Li extension formulae {#bel-formula-sec} ------------------------------------- We have the Bismut-Elworthy-Li formula $$\label{bismut-omega} \nabla P^{\mu}_{s,t}(f)(x)= {\mathbb{E}}\left(f(X_{s,t}^{\mu}(x))~ \tau^{\mu,\omega}_{s,t}(x)\right)\quad \mbox{\rm with}\quad \tau^{\mu,\omega}_{s,t}(x):=\int_s^t~ \partial_u \omega_{s,t}(u)~\nabla X^{\mu}_{s,u}(x)~dW_u$$ The above formula is valid for any function $\omega_{s,t}:u\in [s,t]\mapsto \omega_{s,t}(u)\in {\mathbb{R}}$ of the following form $$\label{bismut-omega-varphi} \omega_{s,t}(u)=\varphi\left((u-s)/(t-s)\right)~\Longrightarrow \partial_u \omega_{s,t}(u)=\frac{1}{t-s}~\partial\varphi\left((u-s)/(t-s)\right)~$$ for some non decreasing differentiable function $\varphi$ on $[0,1]$ with bounded continuous derivatives and such that $$(\varphi(0),\varphi(1))=(0,1)\Longrightarrow \omega_{s,t}(t)-\omega_{s,t}(s)=1$$ In the same vein, for any $s\leq u\leq t$ we have $$\label{bismut-omega-2} \nabla^2 P^{\mu}_{s,t}(f)(x)= {\mathbb{E}}\left(f(X_{s,t}^{\mu}(x))~ \left[\tau^{[2],\mu,\omega}_{s,u}(x)+\nabla X_{s,u}^{\mu}(x)~\tau^{\phi_{s,u}(\mu),\omega}_{u,t}(X_{s,u}^{\mu}(x))\,\tau^{\mu,\omega}_{s,u}(x)^{\prime}\right]\right)$$ with the stochastic process $$\tau^{[2],\mu,\omega}_{s,t}(x):=\int_s^t~ \partial_u \omega_{s,t}(u)~\nabla^2 X^{\mu}_{s,u}(x)~dW_u$$ Besides the fact that $X^{\mu}_{s,t}(x)$ is a nonlinear diffusion, the proof of the above formula follows the same proof as the one provided in [@aht-03; @bismut; @Elworthy; @xm-li; @thompson] in the context of diffusions on differentiable manifolds. For the convenience of the reader, a detailed proof is provided in the appendix on page . Using (\[bismut-omega\]), for any $f$ s.t. $\Vert f\Vert\leq 1$ we check that $$\begin{aligned} \Vert \nabla P^{\mu}_{s,t}(f)\Vert^2&\leq& {\mathbb{E}}\left( \Vert\tau^{\mu,\omega}_{s,t}(x)\Vert^2 \right)\\ &\leq& \int_s^t~e^{-2\lambda_1 (u-s)}~ \Vert \partial_u \omega^{s,t}(u)\Vert^2~du=~\frac{1}{t-s}~ \int_0^1~e^{-2\lambda_1 (t-s)v}~ \left( \partial\varphi(v)\right)^2~dv\end{aligned}$$ Let $\varphi_{\epsilon}$ with $\epsilon\in ]0,1[$ be some differentiable function on $[0,1]$ null on $[0,1-\epsilon]$ and such that $\vert \partial\varphi_{\epsilon}(u)\vert\leq c/\epsilon$ and $(\varphi_{\epsilon}(1-\epsilon),\varphi(1))=(0,1)$, for instance we can choose $$\varphi(u)=\left\{ \begin{array}{ccl} 0&\mbox{\rm if}&u\in [0,1-\epsilon]\\ \displaystyle1+\cos{\left(\left(1+\frac{1-u}{\epsilon}\right)\frac{\pi}{2}\right)}&\mbox{\rm if}&u\in [1-\epsilon,1] \end{array} \right.$$ In this situation, we find the rather crude uniform estimate $$\label{bismut-est} \Vert \nabla P^{\mu}_{s,t}(f)\Vert^2\leq \left(\frac{c}{\epsilon}\right)^2\frac{1}{t-s}~ \int_{1-\epsilon}^1~e^{-2\lambda_1 (t-s)v}~dv \Longrightarrow \Vert \nabla P^{\mu}_{s,t}(f)\Vert\leq \frac{c}{\epsilon}~\frac{1}{\sqrt{t-s}}~e^{-\lambda_1(1-\epsilon) (t-s)}$$ In the same vein, combining (\[bismut-omega-2\]) with the estimate (\[tau-2-estimates\]) for any $\epsilon\in ]0,1[$ and $u\in ]s,t[$ we also check the rather crude uniform estimate $$\Vert \nabla^2 P^{\mu}_{s,t}(f)\Vert\leq \frac{c_1}{\epsilon}~\frac{1}{\sqrt{u-s}}~e^{-\lambda_1 (u-s) (1-\epsilon)}+\frac{c_2}{\epsilon^2}\frac{1}{\sqrt{(t-u)(u-s)}}~e^{-\lambda_1 (u-s) }~e^{-\lambda_1 (t-s) (1-\epsilon)}$$ Choosing $u=s+(1-\epsilon)(t-s)$ in the above display we readily check that $$\label{bismut-est-P2} \Vert \nabla^2 P^{\mu}_{s,t}(f)\Vert\leq \frac{c_1}{\epsilon\sqrt{1-\epsilon}}~\frac{1}{\sqrt{t-s}}~e^{-\lambda_1 (1-\epsilon)^2(t-s)}+\frac{c_2}{\epsilon^2}\frac{1}{\sqrt{\epsilon(1-\epsilon)}}~\frac{1}{t-s}~ e^{-2\lambda_1 (t-s) (1-\epsilon)}$$ Integro-differential operators {#integro-diff-sec} ------------------------------ Let $\BB^{\mu}_{s,t}(x_0,x_1)$ be the matrix-valued function defined for any $(x_0,x_1)\in {\mathbb{R}}^{2d}$, $\mu\in P_2({\mathbb{R}}^d)$ and any $s\leq t$ by the formulae $$\label{def-BB-s-t} \BB^{\mu}_{s,t}(x_0,x_1):=\nabla_{x_0}b_{s,t}^{\mu}(x_0,x_1)\quad\mbox{\rm with}\quad b_{s,t}^{\mu}(x_0,x_1):={\mathbb{E}}\left[b_{t}\left(x_1,X_{s,t}^{\mu}(x_0)\right)\right]$$ For instance, for the linear model discussed in (\[lin-case-intro\]) we have $$\BB^{\mu}_{s,t}(x_0,x_1)^{\prime}=B_2~e^{(t-s)B_1}\quad\mbox{\rm and}\quad b_{s,t}^{\mu}(x_0,x_1)=B_1x_1+B_2\left[e^{(t-s)B_1}(x_0-\mu(e))+e^{(t-s)[B_1+B_2]}~\mu(e)\right]$$ We also consider the collection Weyl chambers $ [s,t]_n$ defined for any $n\geq 1$ by $$\begin{aligned} [s,t]_n&:=&\left\{u=(u_1,\ldots,u_n)\in [s,t]^n~:~s\leq u_1\leq\ldots\leq u_n\leq t\right\}\quad \mbox{\rm and set}\quad du:=du_1\ldots du_n\end{aligned}$$ We consider the space-time Weyl chambers $$\begin{aligned} \Delta_{s,t}&:=&\cup_{n\geq 1}\Delta_{s,t}^n\quad \mbox{\rm with}\quad \Delta_{s,t}^n:=[s,t]_n\times {\mathbb{R}}^{nd} \label{def-Delta-s-t}\end{aligned}$$ The coordinates of a generic point $(u,y)\in \Delta_{s,t}^n$ for some $n\geq 1$ are denoted by $$u=(u_1,\ldots,u_n)\in [s,t]_n\quad \mbox{\rm and} \quad y=(y_1,\ldots,y_n)\in{\mathbb{R}}^{nd}$$ We also use the convention $u_0=s$ and $u_{n+1}=t$. We consider the measures $\Phi_{s,u}(\mu)$ on $\Delta_{s,t}$ given on every set $\Delta_{s,t}^n$ and any $n\geq 1$ by $$\Phi_{s,u}(\mu)(d(u,y))=\phi_{s,u}(\mu)(dy)~du$$ with the tensor product measures $$\phi_{s,u}(\mu)(dy):=\phi_{s,u_1}(\mu)(dy_1)\ldots\phi_{s,u_n}(\mu)(dy_n)$$ Let $b^{\mu}_{s,u}(x,y)$ be the function defined for any $\mu\in P_2({\mathbb{R}}^d)$, $x\in{\mathbb{R}}^d$, and any $(u,y)\in \Delta_{s,t}^n$ and $n\geq 1$ by the formula $$\label{defi-tau} b^{\mu}_{s,u}(x,y)^{\prime}:= b_{s,u_1}^{\mu}\left(x,y_1\right)^{\prime}~ \prod_{1\leq k<n}\BB^{\phi_{s,u_k}(\mu)}_{u_k,u_{k+1}}(y_k,y_{k+1})$$ In the above display the product of matrices is understood as a directed product from $k=1$ to $k=(n-1)$. For instance, for the linear model discussed in (\[lin-case-intro\]) we have $$b^{\mu}_{s,u}(x,y)=B_2~e^{(u_n-u_{n-1})B_1}\ldots B_2~e^{(u_2-u_1)B_1}~b^{\mu}_{s,u_1}(x,y_1)$$ For any $x\in{\mathbb{R}}^d$, and any $(u,y)\in \Delta_{s,t}^n$ and $n\geq 1$ we also set $$\label{def-T} \BB_{u,t}^{\phi_{s,u}(\mu)}(y,x):=\BB_{u_n,t}^{\phi_{s,u_n}(\mu)}(y_n,x)\quad \mbox{\rm and}\quad { {\cal P }}^{\,\phi_{s,u}(\mu)}_{u,t}(\nabla f)(y):= { {\cal P }}^{\,\phi_{s,u_n}(\mu)}_{u_n,t}(\nabla f)(y_n)$$ For any $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ and $s\leq t$ we let $Q^{\mu_1,\mu_0}_{s,t}$ be the operator defined on differentiable functions $f$ on ${\mathbb{R}}^d$ by $$\label{def-Q-Qa} Q^{\mu_1,\mu_0}_{s,t}(f):= { {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)$$ with the $(0,1)$-tensor integral operator ${ {\cal Q }}^{\mu_1,\mu_0}_{s,t}$ defined by the integral formula $${ {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)(x):=\int_{\Delta_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(x,y)^{\prime}~{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)(y)$$ Recall that $b_{t}(x,y)$ is differentiable at any order with uniformly bounded derivatives. Thus, using the estimates (\[ref-moments\]) and (\[tau-n-estimates\]), for any $m,n\geq 0$, $\mu_0,\mu_1\in P_{m\vee 2}({\mathbb{R}}^d)$ we have $$\label{est-Q-1} \Vert Q^{\mu_1,\mu_0}_{s,t}\Vert_{{ {\cal C }}^1_m({\mathbb{R}}^d)\rightarrow { {\cal C }}^n_1({\mathbb{R}}^d)}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ Let $p^{\mu_1,\mu_0}_{s,t}$ be the function defined for any $s\leq t$ and $x,z\in {\mathbb{R}}^d$ by the formula $$\begin{aligned} p_{s,t}^{\mu_1,\mu_0}(x,z)^{\prime} &=&b^{\mu_0}_{s,t}(x,z)^{\prime}+\int_{\Delta_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(x,y)^{\prime} ~ \BB^{\,\phi_{s,u}(\mu_0)}_{u,t}(y,z)\label{p-def-al} \end{aligned}$$ In this notation, we readily check the following proposition. \[prop-ref-rev\] The $(0,1)$-tensor integral operator ${ {\cal Q }}^{\mu_1,\mu_0}_{s,t}$ can be rewritten as follows: $${ {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)(x)=\int_{\Delta_{s,t}^1}~\Phi_{s,u}(\mu_1)(d(u,y))~p^{\mu_1,\mu_0}_{s,u}(x,y)^{\prime}~{ {\cal P }}^{\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)(y)$$ For instance, for the linear model discussed in (\[lin-case-intro\]) the function $p_{s,t}^{\mu_1,\mu_0}(x,z)$ defined in (\[p-def-al\]) reduces to $$\label{linear-example-d} \begin{array}{l} \displaystyle p_{s,t}^{\mu_1,\mu_0}(x,z)=B_1\,z+B_2\,e^{(t-s)(B_1+B_2)}\,x\\ \\ \displaystyle +B_2~\left[\int_s^te^{(t-u)(B_1+B_2)}\,B_1\,e^{(u-s)(B_1+B_2)}~du~\mu_1(e)+\int_s^te^{(t-u)(B_1+B_2)}\,B_2\,e^{(u-s)(B_1+B_2)}~du~\mu_0(e)\right] \end{array}$$ We check this claim expanding in (\[p-def-al\]) the exponential series coming from the integration over the set $\Delta_{s,t}$. A detailed proof of the above formula is provided in the appendix on page . Some differential formulae {#s-diff-form-sec} -------------------------- The matrix $\nabla_{y_0}b_{s,t}^{\mu}(y_0,y_1)$ defined in (\[def-BB-s-t\]) can alternatively be written as follows $$\nabla_{y_0}b_{s,t}^{\mu}(y_0,y_1)={ {\cal P }}^{\mu}_{s,t}\left(b^{[2]}_{t}(y_1,{\mbox{\LARGE .}})\right)(y_0)={\mathbb{E}}\left[\nabla X_{s,t}^{\mu}(y_0)~b^{[2]}_{t}(y_1,X_{s,t}^{\mu}(y_0))\right]$$ We also have the $(2,1)$ and $(3,1)$-tensor formulae $$\begin{aligned} \nabla_{y_0}^2b_{s,t}^{\mu}(y_0,y_1)&=&{ {\cal P }}^{[2,1],\mu}_{s,t}(b^{[2]}_{t}(y_1,{\mbox{\LARGE .}}))(y_0)+{ {\cal P }}^{[2,2],\mu}_{s,t}(b^{[2,2]}_{t}(y_1,{\mbox{\LARGE .}}))(y_0)\\ \nabla^3_{y_0}b_{s,t}^{\mu}(y_0,y_1)&=&{ {\cal P }}^{[3,1],\mu}_{s,t}(b^{[2]}_{t}(y_1,{\mbox{\LARGE .}}))(y_0)+{ {\cal P }}^{[3,2],\mu}_{s,t}(b^{[2,2]}_{t}(y_1,{\mbox{\LARGE .}}))(y_0)+{ {\cal P }}^{[3,3],\mu}_{s,t}(b^{[2,2,2]}_{t}(y_1,{\mbox{\LARGE .}}))(y_0)\end{aligned}$$ For any $(u,y)\in \Delta_{s,t}^n$ with $n\geq 1$ and for any $k\geq 1$ we have the $(k,1)$-tensor formulae $$\label{def-BB-k} \nabla^{k}_{y_0}b^{\mu}_{s,u}(y_0,y)=\BB_{s,u}^{[k],\mu}(y_0,y):=\nabla^k_{y_0}b_{s,u_1}^{\mu}(y_0,y_1)\prod_{1\leq k<n}\BB^{\phi_{s,u_k}(\mu)}_{u_k,u_{k+1}}(y_k,y_{k+1})$$ We consider the $(n,1)$-tensor valued function $$\begin{aligned} q_{s,t}^{[n],\mu_1,\mu_0}(x,z) &:=& \BB^{[n],\mu_0}_{s,t}(x,z) +\int_{\Delta_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~\BB_{s,u}^{[n],\mu_0}(x,y) ~ \BB^{\,\phi_{s,u}(\mu_0)}_{u,t}(y,z)\end{aligned}$$ and we use the convention $$\BB^{[0],\mu_0}_{s,t}(x,z)=b^{\mu_0}_{s,t}(x,z)^{\prime} \quad \mbox{\rm so that}\quad q_{s,t}^{[0],\mu_1,\mu_0}(x,z)=p_{s,t}^{\mu_1,\mu_0}(x,z)^{\prime}$$ [For instance, for the linear model discussed in (\[lin-case-intro\]) and (\[linear-example-d\]) the above objects reduce to $$q_{s,t}^{[1],\mu_1,\mu_0}(x,y)^{\prime}=B_2\,e^{(B_1+B_2)(t-s)}\quad\mbox{\rm and}\quad \forall n\geq 2\quad q_{s,t}^{[n],\mu_1,\mu_0}(x,y)=0$$ ]{} In this notation, we have the following proposition. \[prop-d-n-Q\] For any $n\geq 0$ the $n$-th differential of the operator $Q^{\mu_1,\mu_0}_{s,t}$ is given by the formula $$\nabla^n Q^{\mu_1,\mu_0}_{s,t}(f)= { {\cal Q }}^{[n],\mu_1,\mu_0}_{s,t}(\nabla f)$$ with the $(n,1)$-tensor integral operator given by $$\label{def-Qa-n} { {\cal Q }}^{[n],\mu_1,\mu_0}_{s,t}(\nabla f)(x):= \int_{\Delta^1_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~q^{[n],\mu_1,\mu_0}_{s,u}(x,y)~{ {\cal P }}^{\mu_0}_{u,t}(\nabla f)(y)$$ In addition, when condition $(H)$ is satisfied for any $n\geq 1$ we have the exponential estimates $$\label{def-lambda-hat-2} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal Q }}^{[n],\mu_1,\mu_0}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\leq c_n~e^{- \lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ The proof of the first assertion follows from (\[p-def-al\]). More precisely, using (\[p-def-al\]) we have $$\nabla^n_x \,p_{s,t}^{\mu_1,\mu_0}(x,y) =q^{[n],\mu_1,\mu_0}_{s,t}(x,y)$$ On the other hand, by proposition \[prop-ref-rev\] we also have $$\begin{aligned} \nabla^n Q^{\mu_1,\mu_0}_{s,t}(f)(x)&=& \nabla^n{ {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)(x)\\ &=&\int_{\Delta_{s,t}^1}~\Phi_{s,u}(\mu_1)(d(u,y))~\nabla^n_x\,p^{\mu_1,\mu_0}_{s,u}(x,y)~{ {\cal P }}^{\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)(y)= { {\cal Q }}^{[n],\mu_1,\mu_0}_{s,t}(\nabla f)(x)\end{aligned}$$ This ends the proof of the first assertion. When condition $(H)$ is satisfied, for any $x\in {\mathbb{R}}^d$ and $(u,y)\in \Delta_{s,t}^n$ we have $$\label{estimate-BB} \Vert \BB^{\mu}_{s,t}(y_0,y_1)\Vert_2\leq \Vert b^{[2]}\Vert_2~e^{-\lambda_1(t-s)}\quad\mbox{\rm and}\quad \Vert \BB_{s,u}^{\mu}(x,y)\Vert_2\leq \Vert b^{[2]}\Vert_2^n~e^{-\lambda_1(u_n-s)}$$ Using (\[tau-n-estimates\]) we also check the uniform estimate $$\label{estimate-q-2} \Vert q_{s,t}^{[n],\mu_1,\mu_0}(x,y) \Vert\leq c_n~e^{- \lambda_{1,2}(t-s)}$$ The end of the proof is now a consequence of (\[tau-1-estimates\]). [\ ]{} For any $n\geq 0$ any bounded function $f$ on ${\mathbb{R}}^d$ and for any function $\omega$ of the form (\[bismut-omega-varphi\]) we have the Bismut-Elworthy-Li formula $$\label{def-Q-bismut-ref} \nabla^n Q^{\mu_1,\mu_0}_{s,t}(f)= \int_{\Delta^1_{s,t}}~\Phi_{s,u}(\mu)(d(u,y))~q^{[n],\mu_1,\mu_0}_{s,u}(x,y)~ {\mathbb{E}}\left(f(X^{\mu_0}_{u,t}(y))~ \tau^{\mu_0,\omega}_{u,t}(y)\right)$$ In the above display, $\tau^{\mu,\omega}_{u,t}(y)$ stands for the stochastic process defined in (\[bismut-omega\]). In addition, when condition $(H)$ is satisfied we have the exponential estimates $$\label{def-lambda-hat-bismut} \Vert \nabla^n Q^{\mu_1,\mu_0}_{s,t}(f)\Vert\leq c_n~e^{- \lambda(t-s)}~\Vert f\Vert\quad \mbox{for some $\lambda>0$}$$ The proof of the first assertion is a direct application of the Bismut-Elworthy-Li formula (\[bismut-omega\]). More precisely, using (\[bismut-omega\]) we have $${ {\cal P }}^{\mu_0}_{u,t}(\nabla f)(y)= {\mathbb{E}}\left(f(X^{\mu_0}_{u,t}(y))~ \tau^{\mu_0,\omega}_{u,t}(y)\right)$$ The formula (\[def-Q-bismut-ref\]) is now a direct consequence of (\[def-Qa-n\]). We check (\[def-lambda-hat-bismut\]) combining (\[bismut-est\]) with (\[estimate-q-2\]). This ends the proof of the proposition. [\ ]{} When $n=1$ we drop the upper index and we write $\left(\BB_{s,u}^{\mu},q_{s,t}^{\mu_1,\mu_0}\right)$ instead of $\left( \BB_{s,u}^{[1],\mu},q_{s,t}^{[1],\mu_1,\mu_0}\right)$. The operators discussed above are indexed by a pair of measures $(\mu_0,\mu_1)$. To simplify notation, when $\mu_1=\mu_0=\mu$ we suppress one of the indices and we write $ ( Q^{\mu}_{s,t},{ {\cal Q }}^{[n],\mu}_{s,t})$ and $(p^{\mu}_{s,t},q^{[n],\mu}_{s,t})$ instead of $(Q^{\mu,\mu}_{s,t},{ {\cal Q }}^{[n],\mu,\mu}_{s,t})$ and $(p^{\mu,\mu}_{s,t}, q^{[n],\mu,\mu}_{s,t})$. Tangent processes {#tangent-sec} ================= The tangent process associated with the diffusion flow $\psi_{s,t}(Y)$ introduced in (\[eq-diff\]) is given for any $U\in {\mathbb{H}}_s({\mathbb{R}}^d)$ by the evolution equation $$\label{eq-diff-lin-eq} \partial_t(\partial \psi_{s,t}(Y)\cdot U)=\partial B_t(\psi_{s,t}(Y))\cdot (\partial \psi_{s,t}(Y)\cdot U)$$ In the above display, $\partial B_t(X)\in \mbox{\rm Lin}({\mathbb{H}}_t({\mathbb{R}}^d),{\mathbb{H}}_t({\mathbb{R}}^d))$ stands for the Fréchet differential of the drift function $B_t$ defined for any $Z\in {\mathbb{H}}_t({\mathbb{R}}^d)$ by $$\partial B_t(X)\cdot Z={\mathbb{E}}\left( \nabla_{x_1}\,b_t(X,\overline{X})^{\prime}~Z+\nabla_{x_2}\,b_t(X,\overline{X})^{\prime}~\overline{Z}~\|~{\mathbb{F}}_t\right)$$ where $(\overline{X},\overline{Z})$ stands for an independent copy of $(X,Z)$. Spectral estimate {#spectral-sec} ----------------- This section is mainly concerned with the proof of theorem \[theo-intro-1\]. For any pair of random variables $Z_1,Z_2\in {\mathbb{H}}_t({\mathbb{R}}^d)$ we have the duality formula $$\langle Z_1, \partial B_t(X)\cdot Z_2\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}=\langle \partial B_t(X)^{\star}\cdot Z_1, Z_2\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}$$ with the dual operator $\partial B_t(X)^{\star}$ defined by the formula $$\partial B_t(X)^{\star}\cdot Z_1:={\mathbb{E}}\left(b_t^{[1]}(X,\overline{X})~Z_1+b_t^{[2]}(\overline{X},X)~\overline{Z}_1~\vert~{\mathbb{F}}_t\right)$$ In the above display, $(\overline{X},\overline{Z}_1)$ stands for an independent copy of $(X,Z_1)$. The symmetric part of $ \partial B_t(X)$ is given by the formula $$\partial B_t(X)_{\tiny sym}:= \frac{1}{2}\left[\partial B_t(X)+\partial B_t(X)^{\star}\right]$$ We are now in position to prove theorem \[theo-intro-1\]. The first assertion is a direct consequence of the evolution equation $$\begin{aligned} 2^{-1} \partial_t\,\Vert \partial \psi_{s,t}(Y)\cdot U\Vert^2_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}&=&\langle (\partial \psi_{s,t}(Y)\cdot U), \partial B_t(\psi_{s,t}(Y))_{\tiny sym}\cdot (\partial \psi_{s,t}(Y)\cdot U)\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)} \end{aligned}$$ Whenever $(H)$ is met we have $ \partial B_t(X)_{\tiny sym}\leq -\lambda_0~I$ for some $\lambda_0>0$. In this situation, the r.h.s. estimate in (\[theo-intro-1-eq\]) is a direct consequence of (\[theo-intro-1-eq-log-norm\]). Given an independent copy $(\overline{X},\overline{Z}_2) $ of $(X,Z_2)$ we have $$\begin{aligned} 2~\langle Z_1,\partial B_t(X)^{\star}\cdot Z_2\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}&=&{\mathbb{E}}\left(\left\langle \left[\begin{array}{c} Z_1\\ \overline{Z}_1 \end{array}\right],A_t(X,\overline{X})\left[\begin{array}{c} Z_2\\ \overline{Z}_2 \end{array}\right]\right\rangle\right)\\&=&2~\langle \partial B_t(X)\cdot Z_1, Z_2\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}\end{aligned}$$ This yields the log-norm estimate $$A_t(X,\overline{X})_{\tiny sym}\leq -\lambda_0~I\Longrightarrow ~ \partial B_t(X)_{\tiny sym}\leq -\lambda_0~I$$ The proof of theorem \[theo-intro-1\] is now completed. [\ ]{} Dyson-Phillips expansions ------------------------- In the further development of this section we shall denote by $$(\overline{\psi}_{s,t},\overline{U},\overline{X}^{\mu}_{s,t},\overline{Y})\quad\mbox{\rm and} \quad(\overline{\psi}^n_{s,t},\overline{U}^n,\overline{X}^{\mu,n}_{s,t},\overline{Y}^n)_{n\geq 0}$$ a collection of independent copies of the stochastic flows $(\psi_{s,t},X^{\mu}_{s,t})$ and some given $U,Y\in{\mathbb{H}}_s({\mathbb{R}}^d)$. To simplify notation, we also set $$X_{s,t}:=\psi_{s,t}(Y)\qquad\overline{X}_{s,t}:=\overline{\psi}_{s,t}(\overline{Y})\quad\mbox{\rm and}\quad\overline{X}^n_{s,t}:=\overline{\psi}^n_{s,t}(\overline{Y}^n)$$ We are now in position to state and prove the main result of this section. The tangent process $\partial \psi_{s,t}$ is given for any $U\in{\mathbb{H}}_s({\mathbb{R}}^d)$ and any $Y\in{\mathbb{H}}_s({\mathbb{R}}^d)$ with distribution $\mu\in P_2({\mathbb{R}}^d)$ by the Dyson-Phillips series $$\label{tangent-hilbert} \begin{array}{l} \displaystyle\partial \psi_{s,t}(Y)\cdot U=\nabla X^{\mu}_{s,t}(Y)^{\prime}~U\\ \\ \hskip.3cm\displaystyle+\sum_{n\geq 1}\int_{[s,t]_n}~ \left(\nabla X^{\phi_{s,u_n}(\mu)}_{u_n,t}\right)\left(X_{s,u_n}\right)^{\prime} ~ {{\mathbb{E}}}\left( \left[ \prod_{1\leq k\leq n}{\BB}_{u_{k-1},u_k}^{\phi_{s,u_{k-1}}(\mu)}\left(\overline{X}^{k-1}_{s,u_{k-1}},\overline{X}^k_{s,u_k}\right) \right]^{\prime} \overline{U}~\vert~{\mathbb{F}}_{u_n}\right)~du \end{array}$$ with the boundary conventions $$u_0=s\qquad \overline{X}^0_{s,u_1}=\overline{X}_{s,u_1}\quad \mbox{and}\quad \overline{X}^n_{s,u_n}=X_{s,u_n}\quad \mbox{for any $n\geq 1$}$$ For any $s\leq u\leq t$ and $x\in{\mathbb{R}}^d$ we have $$\partial_t \,\nabla X^{\mu}_{s,t}(x)^{-1}=-~b_t^{[1]}\left(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu)\right)~\nabla X^{\mu}_{s,t}(x)^{-1}$$ and $$\nabla X^{\mu}_{s,t}(x)=\nabla X^{\mu}_{s,u}(x)~\left(\nabla X_{u,t}^{\phi_{s,u}(\mu)}\right)\left(X^{\mu}_{s,u}(x)\right)$$ In addition, for any $s\leq u\leq t$ and $x_0,x_1\in{\mathbb{R}}^d$ we have $$\nabla_{x_0} b_t(x_1,X_{u,t}^{\phi_{s,u}(\mu)}(x_0))=\nabla X_{u,t}^{\phi_{s,u}(\mu)}(x)~b^{[2]}_t(x_1,X_{u,t}^{\phi_{s,u}(\mu)}(x_0))$$ Combining the above with (\[eq-diff-lin-eq\]) we check that $$\begin{array}{l} \displaystyle\partial_t\left(\left(\nabla X^{\mu}_{s,t}(Y)^{-1}\right)^{\prime}~(\partial \psi_{s,t}(Y)\cdot U)\right)\\ \\ \displaystyle=\left(\nabla X^{\mu}_{s,t}(Y)^{-1}\right)^{\prime}~{{\mathbb{E}}}\left( \nabla b_t\left(\psi_{s,t}(Y),\overline{X}^{\mu}_{s,t}({\mbox{\LARGE .}})\right)(\overline{Y})^{\prime}~\left(\nabla \overline{X}^{\mu}_{s,t}(\overline{Y})^{-1}\right)^{\prime} (\partial \overline{\psi}_{s,t}(\overline{Y})\cdot \overline{U})~\vert~{\mathbb{F}}_t \right) \end{array}$$ In the above display, $\nabla b_t\left(\psi_{s,t}(Y),\overline{X}^{\mu}_{s,t}({\mbox{\LARGE .}})\right)(\overline{Y})=\nabla h(\overline{Y})$ stands for the gradient of the random function $$h~:~x\mapsto h(x)=b_t\left(\psi_{s,t}(Y),\overline{X}^{\mu}_{s,t}(x)\right) \quad\mbox{\rm evaluated at $x=\overline{Y}$.}$$ Equivalently, we have $$\begin{array}{l} \displaystyle\left(\nabla X^{\mu}_{s,t}(Y)^{-1}\right)^{\prime}~(\partial \psi_{s,t}(Y)\cdot U)\\ \\ \displaystyle =U+\int_s^t\left(\nabla X^{\mu}_{s,u}(Y)^{-1}\right)^{\prime}~{{\mathbb{E}}}\left( \nabla b_u\left(\psi_{s,u}(Y),\overline{X}^{\mu}_{s,u}({\mbox{\LARGE .}})\right)(\overline{Y})^{\prime}~\left(\nabla \overline{X}^{\mu}_{s,u}(\overline{Y})^{-1}\right)^{\prime} (\partial \overline{\psi}_{s,u}(\overline{Y})\cdot \overline{U})~\vert~{\mathbb{F}}_u \right)~du \end{array}$$ and therefore $$\begin{array}{l} \displaystyle\partial \psi_{s,t}(Y)\cdot U=\left(\nabla X^{\mu}_{s,t}(Y)\right)^{\prime}U+\int_s^t~\left(\left(\nabla X^{\phi_{s,u}(\mu)}_{u,t}\right)\left(X_{s,u}^{\mu}(Y)\right)\right)^{\prime}\\ \\ \displaystyle~\hskip3cm\times {{\mathbb{E}}}\left( \nabla b_u\left(\psi_{s,u}(Y),\overline{X}^{\mu}_{s,u}({\mbox{\LARGE .}})\right)(\overline{Y})^{\prime}~\left(\nabla \overline{X}^{\mu}_{s,u}(\overline{Y})^{-1}\right)^{\prime} (\partial \overline{\psi}_{s,u}(\overline{Y})\cdot \overline{U})~\vert~{\mathbb{F}}_u \right)~du \end{array}$$ Now, the end of the proof of (\[tangent-hilbert\]) follows a simple induction, thus it is skipped. [\ ]{} \[cor-tangent\] For any $V\in{\mathbb{H}}_t({\mathbb{R}}^d)$ and for any $Y\in{\mathbb{H}}_s({\mathbb{R}}^d)$ with distribution $\mu\in P_2({\mathbb{R}}^d)$ we have $$\label{tangent-hilbert} \begin{array}{l} \displaystyle\partial \psi_{s,t}(Y)^{\star}\cdot V={\mathbb{E}}\left(\nabla X^{\mu}_{s,t}(Y) V~\|~{\mathbb{F}}_s\right)\\ \\ \hskip.3cm\displaystyle+\sum_{n\geq 1}\int_{[s,t]_n}~{{\mathbb{E}}}\left( \left[ \prod_{1\leq k\leq n} \BB_{u_{k-1},u_k}^{\phi_{s,u_{k-1}}(\mu)}\left(\overline{X}^{k-1}_{s,u_{k-1}},\overline{X}^k_{s,u_k}\right) \right]\left(\nabla \overline{X}^{\phi_{s,u_n}(\mu)}_{u_n,t}\right)\left(\overline{X}_{s,u_n}\right) \overline{V}~\|~{\mathbb{F}}_s\right)~du \end{array}$$ with the boundary conditions $$u_0=s\quad \mbox{and}\quad \overline{X}^0_{s,u_1}=\psi_{s,u_1}(Y) \quad \mbox{and}\quad \overline{X}^n_{s,u_n}=\overline{X}_{s,u_n}$$ Gradient semigroup analysis {#g-sg-sec} --------------------------- This section is concerned with a gradient semigroup description of the dual of the tangent process. For any $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ and $s\leq t$ we let $D_{\mu_1,\mu_0}\phi_{s,t}$ be the operator defined on differentiable functions $f$ on ${\mathbb{R}}^d$ by $$\label{def-D} D_{\mu_1,\mu_0}\phi_{s,t}:=P^{\mu_0}_{s,t}+Q^{\mu_1,\mu_0}_{s,t}$$ In the above display, $Q^{\mu_1,\mu_0}_{s,t} $ stands for the operator defined in (\[def-Q-Qa\]). [Rewritten in terms of expectation operators we have $$\begin{array}{l} D_{\mu_1,\mu_0}\phi_{s,t}(f)(x)\\ \\ \displaystyle={\mathbb{E}}\left[(f\circ X^{\mu_0}_{s,t})(x)\right]+\sum_{n\geq 1}\int_{\Delta_{s,t}^n}~\Phi_{s,u}(\mu_1)(d(u,y))~{\mathbb{E}}\left[b^{\mu_0}_{s,u}(x,y)^{\prime}~\nabla (f\circ X^{\phi_{s,u_n}(\mu_0)}_{u_n,t})(y_n)\right] \end{array}$$]{} Recall that $b_{t}(x,y)$ is differentiable at any order with uniformly bounded derivatives. Thus, arguing as in the proof of (\[crude-est-rev\]) and (\[est-Q-1\]) for any $m,n\geq 1$, $\mu_0,\mu_1\in P_{m\vee 2}({\mathbb{R}}^d)$ we have $$\Vert D_{\mu_1,\mu_0}\phi_{s,t}\Vert_{{ {\cal C }}^n_m({\mathbb{R}}^d)\rightarrow { {\cal C }}^n_m({\mathbb{R}}^d)}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)\label{nabla-D}$$ In the same vein, we check that $$\label{nabla-D-times-2} \Vert \left(D_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}\Vert_{{ {\cal C }}^{n+1}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+1}({\mathbb{R}}^{2d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ The proof of the above estimate is rather technical, thus it is housed in the appendix on page . \[rmk-bismut\] Using the Bismut-Elworthy-Li formula (\[def-Q-bismut-ref\]), we extend the operators $D_{\mu_1,\mu_0}\phi_{s,t}$ with $s<t$ to non necessarily differentiable and bounded functions. We also extend the operator $D_{\mu_1,\mu_0}\phi_{s,t}$ to tensor functions $f=(f_{i})_{i\in [n]}$ by considering the tensor function with entries $$\label{def-D-extension} D_{\mu_1,\mu_0}\phi_{s,t}(f)_i=D_{\mu_1,\mu_0}\phi_{s,t}(f_{i})$$ In this situation, the function $p^{\mu_1,\mu_0}_{s,t}$ introduced in (\[p-def-al\]) takes the form $$p_{s,t}^{\mu_1,\mu_0}(x,z)=D_{\mu_1,\mu_0}\phi_{s,t}(b_t(z,{\mbox{\LARGE .}}))(x)$$ Let $G_{t,\mu_1}$ be the collection of integro-differential operators indexed by $\mu_1\in P_2({\mathbb{R}}^d)$ defined by $$G_{t,\mu_1}(f)(x_2):=\int\mu_1(dx_1)~b_t(x_1,x_2)^{\prime}\,\nabla f(x_1)$$ We also set $$H_{t,\mu_0,\mu_1}:=L_{t,\mu_0}+G_{t,\mu_1}\quad \mbox{\rm and}\quad H_{t,\mu_0}:=L_{t,\mu_0}+G_{t,\mu_0}$$ In this notation, we have the first order expansion $$\label{ref-rev-key} \mu_1L_{t,\mu_1}-\mu_0L_{t,\mu_0}=(\mu_1-\mu_0)L_{t,\mu_0}+(\mu_1-\mu_0)~G_{t,\mu_1}=(\mu_1-\mu_0)H_{t,\mu_0,\mu_1}$$ \[theo-1\] For any $m,n\geq 1$ and any $\mu_0,\mu_1\in P_{m\vee 2}({\mathbb{R}}^d)$ the operator $D_{\mu_1,\mu_0}\phi_{s,t}$ coincides with the evolution semigroup of the integro-differential operator $ H_{t,\phi_{s,t}(\mu_0),\phi_{s,t}(\mu_1)} $; that is, we have the forward evolution equation $$\label{forward-eq} \partial_tD_{\mu_1,\mu_0}\phi_{s,t}=D_{\mu_1,\mu_0}\phi_{s,t}\circ H_{t,\phi_{s,t}(\mu_0),\phi_{s,t}(\mu_1)}\quad \mbox{on}\quad { {\cal C }}^{n\vee 2}_m({\mathbb{R}}^d)$$ In addition, for any $s\leq u< t$ we have the backward evolution equation $$\label{backward-eq} \partial_uD_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}=-H_{u,\phi_{s,u}(\mu_0),\phi_{s,u}(\mu_1)}\circ D_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}\quad \mbox{on}\quad { {\cal C }}^n_m({\mathbb{R}}^d)$$ The proof of the forward equation (\[forward-eq\]) is a direct consequence of the forward evolution equation $$\partial_tP^{\mu_0}_{s,t}=P_{s,t}^{\mu_0}L_{t,\phi_{s,t}(\mu_0)}$$ associated with the Markov semigroup $P^{\mu_0}_{s,t}$, thus it is skipped. The semigroup property (\[sg-chain-rule\]) yields $$\partial_u\left(D_{\mu_1,\mu_0}\phi_{s,u}\circ D_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}\right)=0$$ Combining the above with the forward equation (\[forward-eq\]) we check that $$D_{\mu_1,\mu_0}\phi_{s,u}\circ \partial_uD_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}=-D_{\mu_1,\mu_0}\phi_{s,u}\circ H_{u,\phi_{s,u}(\mu_0),\phi_{s,u}(\mu_1)}\circ D_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}$$ This implies that $$\left[\partial_uD_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}\right]_{u=s}=-H_{s,\mu_0,\mu_1}D_{\mu_1,\mu_0}\phi_{u,t}$$ from which we conclude that $$\begin{array}{l} \left[\partial_uD_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}\right]_{u=v}\\ \\ =\left[\partial_uD_{\phi_{v,u}\left(\phi_{s,v}(\mu_1)\right),\phi_{v,u}\left(\phi_{s,v}(\mu_0)\right)}\phi_{u,t}\right]_{u=v}=-H_{s,\phi_{s,v}(\mu_0),\phi_{s,v}(\mu_1)}D_{\phi_{s,v}(\mu_1),\phi_{s,v}(\mu_0)}\phi_{v,t} \end{array}$$ This yields the backward evolution equation (\[backward-eq\]). This ends the proof of the theorem. [\ ]{} Next proposition is a direct consequence of (\[def-D\]) combined with the formulae (\[commutation-tensor\]) and (\[def-Qa-n\]). We have the commutation formula $$\label{commutation-D} \nabla \circ D_{\mu_1,\mu_0}\phi_{s,t}= { {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\circ \nabla$$ with the $(1,1)$-tensor integral operator given by the column vector function $$\label{def-Da} { {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}(\nabla f)(x):= { {\cal P }}^{\mu_0}_{s,t}(\nabla f)(x)+\int_{\Delta^1_{s,t}}\Phi_{s,v}(\mu_1)(d(v,y))~q^{\mu_1,\mu_0}_{s,v}(x,y)~{ {\cal P }}^{\phi_{s,v}(\mu_0)}_{v,t}(\nabla f)(y)$$ In addition, when condition $(H)$ is satisfied we have $$\label{commutation-D-etimate} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal D }}_{\mu_1,\mu_0}\phi_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\leq c~e^{-\lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ \[rmk-1\] Following remark \[rmk-bismut\], using the Bismut-Elworthy-Li formula (\[def-Q-bismut-ref\]), we extend the gradient operators $\nabla D_{\mu_1,\mu_0}\phi_{s,t}$ with $s<t$ to measurable and bounded functions. The exponential estimate stated in (\[def-lambda-hat-bismut\]) are a direct consequence of the estimates presented in (\[def-lambda-hat-bismut\]). By (\[def-D-extension\]) the commutation formula (\[commutation-D\]) is also satisfied for multivariate column functions $f$. In this situation $ { {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}(\nabla f)$ is a $(d\times d)$-matrix valued function. The proof of theorem \[theo-intro-2\] is now a consequence of the estimate (\[commutation-D-etimate\]) and the fact that $$\partial_t \left[\phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0)\right]=\left[\phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0)\right]\circ H_{t,\phi_{s,t}(\mu_0),\phi_{s,t}(\mu_1)}$$ More precisely, using (\[ref-rev-key\]) the above formula implies that $$\begin{array}{l} \partial_u\left( \left[\phi_{s,u}(\mu_1)- \phi_{s,u}(\mu_0)\right]D_{\phi_{s,u}(\mu_1),\phi_{s,u}(\mu_0)}\phi_{u,t}\right)=0\\ \\ \Longrightarrow \phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0) =(\mu_1-\mu_0)D_{\mu_1,\mu_0}\phi_{s,t} \end{array}$$ The operators discussed above are indexed by a pair of measures $(\mu_0,\mu_1)$. To simplify notation, when $\mu_1=\mu_0=\mu$ we suppress one of the parameter and we write $(D_{\mu}\phi_{s,t},{ {\cal D }}_{\mu}\phi_{s,t}) $ instead of $(D_{\mu,\mu}\phi_{s,t},{ {\cal D }}_{\mu,\mu}\phi_{s,t})$. \[theo-0\] For any $m,n\geq 1$, any function $f\in { {\cal C }}^n_m({\mathbb{R}}^d)$ and any $Y\in{\mathbb{H}}_s({\mathbb{R}}^d)$ with distribution $\mu\in P_2({\mathbb{R}}^d)$ we have the gradient formula $$\partial \psi_{s,t}(Y)^{\star}\cdot\nabla f(\psi_{s,t}(Y))=\nabla D_{\mu}\phi_{s,t}(f)(Y)={ {\cal D }}_{\mu}\phi_{s,t}(\nabla f)(Y)$$ Given a smooth function $f$ on ${\mathbb{R}}^d$ we have $$\langle \nabla f(\psi_{s,t}(Y)),\partial \psi_{s,t}(Y)\cdot U\rangle_{\,{\mathbb{H}}_t({\mathbb{R}}^d)}= \langle \partial \psi_{s,t}(Y)^{\star}\cdot\nabla f(\psi_{s,t}(Y)), U\rangle_{\,{\mathbb{H}}_s({\mathbb{R}}^d)}$$ Replacing $V$ by $\nabla f(\psi_{s,t}(Y))$ in (\[tangent-hilbert\]) we check that $$\begin{array}{l} \displaystyle\partial \psi_{s,t}(Y)^{\star}\cdot\nabla f(\psi_{s,t}(Y))\\ \\ \displaystyle=\nabla D_{\mu}\phi_{s,t}(f)(Y)={\mathbb{E}}\left( \nabla \left(f\circ X^{\mu}_{s,t}\right)(Y)~\|~Y\right)\\ \\ \hskip.3cm\displaystyle+\sum_{n\geq 1}\int_{[s,t]_n}~\overline{{\mathbb{E}}}\left( \left[ \prod_{0\leq k< n} \BB_{u_{k},u_{k+1}}^{\phi_{s,u_{k}}(\mu)}\left(\overline{X}^{k}_{s,u_{k}},\overline{X}^{k+1}_{s,u_{k+1}}\right) \right] \nabla \left(f\circ X^{\phi_{s,u_n}(\mu)}_{u_n,t}\right)\left(\overline{X}_{s,u_n}\right)~\|~Y\right)~du \end{array}$$ This ends the proof of the theorem [\ ]{} Taylor expansions {#sec-taylor} ================= This section is mainly concerned with the proof of the first and second order Taylor expansions stated in theorem \[theo-intro-3\] and theorem \[theo-intro-4\] . Section \[sec-some-diff\] presents some preliminary differential formulae used in the proof of the theorems. Some differential formulae {#sec-some-diff} -------------------------- The commutation formula (\[commutation-D\]) takes the form $$\nabla D_{\mu_1,\mu_0}\phi_{s,t}(f)= { {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}(\nabla f)$$ Combining (\[def-D\]) with proposition \[prop-d-n-Q\] and the second order formula (\[ref-nabla-Pa\]) we also have $$\nabla^2 D_{\mu_1,\mu_0}\phi_{s,t}(f)= \nabla^2 P^{\mu_0}_{s,t}(f)+ { {\cal Q }}^{[2],\mu_1,\mu_0}_{s,t}(\nabla f) ={ {\cal P }}^{[2,1],\mu}_{s,t}(\nabla f)+{ {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)+ { {\cal Q }}^{[2],\mu_1,\mu_0}_{s,t}(\nabla f)$$ In summary, we have the first and second order differential formulae $$\begin{aligned} \nabla D_{\mu}\phi_{s,t}(f) &=& { {\cal D }}_{\mu}\phi_{s,t}(\nabla f) \nonumber\\ \nabla^2D_{\mu}\phi_{s,t}(f) &=&{ {\cal D }}_{\mu}\phi_{s,t}^{[2,1]}(\nabla f)+{ {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)\quad \mbox{with}\quad { {\cal D }}_{\mu}\phi_{s,t}^{[2,1]}={ {\cal P }}^{[2,1],\mu}_{s,t}+{ {\cal Q }}^{[2],\mu}_{s,t}\label{dD-W}\end{aligned}$$ Similar formulae for $ \nabla D_{\mu_0,\mu_1}\phi_{s,t}$ and $ \nabla^2 D_{\mu_0,\mu_1}\phi_{s,t}$ can easily be found. In the same vein, using (\[ref-nabla-3-Pa\]) we check the third order differential formula $$\label{dD-W-3} \begin{array}{l} \nabla^3D_{\mu}\phi_{s,t}(f) \\ \\ ={ {\cal D }}_{\mu}\phi_{s,t}^{[3,1]}(\nabla f)+{ {\cal P }}^{[3,2],\mu}_{s,t}(\nabla^2 f)+{ {\cal P }}^{[3,3],\mu}_{s,t}(\nabla^3 f)\quad \mbox{with}\quad { {\cal D }}_{\mu}\phi_{s,t}^{[3,1]}:={ {\cal P }}^{[3,1],\mu}_{s,t}+{ {\cal Q }}^{[3],\mu}_{s,t} \end{array}$$ In addition, when condition $(H)$ is satisfied we have the exponential estimates $$\label{def-lambda-hat-3} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal D }}_{\mu}\phi_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\vee {{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal D }}_{\mu}\phi_{s,t}^{[2,1]} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\vee{{\vert\kern-0.25ex\vert\kern-0.25ex\vert { {\cal D }}_{\mu}\phi_{s,t}^{[3,1]} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\leq c~e^{- \lambda(t-s)} \quad \mbox{for some $\lambda>0$}$$ We let $S_{s,t}^{\mu}$ be the operator defined for any differentiable function $f$ on ${\mathbb{R}}^d$ by $$S_{s,t}^{\mu}(f)={ {\cal S }}_{s,t}^{\mu}(\nabla f)$$ with the $(0,1)$-tensor integral operator ${ {\cal S }}_{s,t}^{\mu}$ defined by the formula $$\begin{aligned} { {\cal S }}_{s,t}^{\mu}(\nabla f)(x_1,x_2) &:=& b_s(x_1,x_2)^{\prime}~{ {\cal D }}_{\mu}\phi_{s,t}(\nabla f)(x_1)+b_s(x_2,x_1)^{\prime}~{ {\cal D }}_{\mu}\phi_{s,t}(\nabla f)(x_2)\label{def-S}\end{aligned}$$ Using (\[nabla-D\]) and (\[def-Da\]) for any $m,n\geq 0$ and $\mu\in P_{m\vee 2}({\mathbb{R}}^d)$ we check that $$\label{estimate-S-m} \Vert S^{\mu}_{s,t}\Vert_{{ {\cal C }}^{n+1}_m({\mathbb{R}}^d)\rightarrow { {\cal C }}^{n}_{m+1}({\mathbb{R}}^{2d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu)$$ We also have the differential formula $$\label{nabla-2-S} (\nabla\otimes\nabla)\left(S^{\mu}_{s,t}(f)\right)={\mathbb{S}}^{[2,1],\mu}_{s,t}(\nabla f)+{\mathbb{S}}^{[2,2],\mu}_{s,t}(\nabla^2 f)$$ with the matrix valued functions $$\begin{array}{l} {\mathbb{S}}^{[2,1],\mu}_{s,t}(\nabla f)(x_1,x_2)=b_s^{[1,2]}(x_1,x_2)~{ {\cal D }}_{\mu}\phi_{s,t}(\nabla f)(x_1)+b_s^{[2,1]}(x_2,x_1)~{ {\cal D }}_{\mu}\phi_{s,t}(\nabla f)(x_2)\\ \\ \displaystyle\hskip5cm+b_s^{[2]}(x_2,x_1)~{ {\cal D }}_{\mu}\phi_{s,t}^{[2,1]}(\nabla f)(x_2)^{\prime}+ { {\cal D }}_{\mu}\phi_{s,t}^{[2,1]}(\nabla f)(x_1)~b_s^{[2]}(x_1,x_2)^{\prime}\\ \\ {\mathbb{S}}^{[2,2],\mu}_{s,t}(\nabla^2 f)(x_1,x_2):=b_s^{[2]}(x_2,x_1)~{ {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)(x_2)^{\prime}+ { {\cal P }}^{[2,2],\mu}_{s,t}(\nabla^2 f)(x_1)~b_s^{[2]}(x_1,x_2)^{\prime} \end{array}$$ When condition $(H)$ is satisfied we also have the exponential estimates $$\label{3s-2} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert {\mathbb{S}}^{[2,1],\mu}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}}\vee {{\vert\kern-0.25ex\vert\kern-0.25ex\vert {\mathbb{S}}^{[2,2],\mu}_{s,t} \vert\kern-0.25ex\vert\kern-0.25ex\vert}} \leq c~e^{-\lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ In addition, using the Bismut-Elworthy-Li extension formulae and the estimates (\[def-lambda-hat-bismut-intro\]) and (\[def-lambda-hat-bismut-intro-hessian\]), or any bounded measurable function $f$ on ${\mathbb{R}}^d$ we check that $$\Vert (\nabla\otimes\nabla)\left(S^{\mu}_{s,t}(f)\right)\Vert\leq c~\left(1\vee 1/(t-s)\right)~e^{- \lambda(t-s)}~\Vert f\Vert\quad \mbox{for some $\lambda>0$}$$ A first order expansion {#sec-first-order-remainder} ----------------------- This section is mainly concerned with the proof of theorem \[theo-intro-3\]. The next technical lemma is pivotal. For any $m\geq 1$ for any $\mu_0,\mu_1\in P_{m+1}({\mathbb{R}}^d)$ we have the second order expansion $$\label{a-2} \begin{array}{l} \phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0)\\ \\ \displaystyle=(\mu_1-\mu_0)D_{\mu_0}\phi_{s,t}+\frac{1}{2}~ \int_s^t~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]^{\otimes 2}\circ S^{\phi_{s,u}(\mu_0)}_{u,t}~du \quad \mbox{on}\quad { {\cal C }}^{n+1}_m({\mathbb{R}}^d) \end{array}$$ Combining (\[ref-rev-key\]) with the backward evolution equation (\[backward-eq\]) we check that $$\begin{array}{l} \displaystyle \partial_u\left\{\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\circ D_{\phi_{s,u}(\mu_0)}\phi_{u,t}\right\}\\ \\ =\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\circ \left[ H_{u,\phi_{s,u}(\mu_0),\phi_{s,u}(\mu_1)}-H_{u,\phi_{s,u}(\mu_0)}\right]\circ D_{\phi_{s,u}(\mu_0)}\phi_{u,t}\\ \\ =\displaystyle\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\circ \left[ G_{u,\phi_{s,u}(\mu_1)}-G_{u,\phi_{s,u}(\mu_0)}\right]\circ D_{\phi_{s,u}(\mu)}\phi_{u,t} \end{array}$$ On the other hand, we have $$\left[ G_{u,\phi_{s,u}(\mu_1)}-G_{u,\phi_{s,u}(\mu_0)}\right](x_2):=\int \left(\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right)(dx_1)~b_u(x_1,x_2)^{\prime}\,\nabla f(x_1)$$ Integrating $u$ from $u=s$ to $u=t$ we obtain the formula $$\begin{array}{l} \displaystyle \left[\phi_{s,t}(\mu_1)-\phi_{s,t}(\mu_0)-(\mu_1-\mu_0)D_{\mu_0}\phi_{s,t}\right](f)\\ \\ \displaystyle=\frac{1}{2} \int_s^t~\int~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]^{\otimes 2}(d(x_1,x_2))~\\ \\ \hskip3cm\left[ b_u(x_1,x_2)^{\prime}~\nabla D_{\phi_{s,u}(\mu_0)}\phi_{u,t}(f)(x_1)+b_u(x_2,x_1)^{\prime}~\nabla D_{\phi_{s,u}(\mu_0)}\phi_{u,t}(f)(x_2)\right]~du \end{array}$$ The end of the lemma is now completed. [\ ]{} Combining the above lemma with (\[nabla-D-times-2\]) and (\[estimate-S-m\]) we check (\[s-o-r\]) with the operator $D^2_{\mu_1,\mu_0}\phi_{s,t}$ defined for any $m,n\geq 0$ and $\mu_0,\mu_1\in P_{m+2}({\mathbb{R}}^d)$ by $$\label{def-D2} D^2_{\mu_1,\mu_0}\phi_{s,t}:=\int_s^t~\left(D_{\mu_1,\mu_0}\phi_{s,u}\right)^{\otimes 2}\circ S^{\phi_{s,u}(\mu_0)}_{u,t}~du\in\mbox{\rm Lin}\left({ {\cal C }}^{n+2}_m({\mathbb{R}}^d),{ {\cal C }}^{n}_{m+2}({\mathbb{R}}^{2d})\right)$$ The second order term in (\[s-o-r\]) can alternatively be expressed in terms of the Hessian of the semigroup $D^2_{\mu_1,\mu_0}\phi_{s,t}$; that is, we have that $$\label{nabla-equiv-2} \begin{array}{l} (\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}(f)\\ \\ \displaystyle=\int_{[0,1]^2}~{\mathbb{E}}\left(\langle \left[ (\nabla\otimes\nabla) D^2_{\mu_1,\mu_0}\phi_{s,t}(f)\right](Y_{\epsilon,\overline{\epsilon}}),(Y_1-Y_0)\otimes (\overline{Y}_1-\overline{Y}_0)\rangle\right)~d\epsilon~ d\overline{\epsilon} \end{array}$$ with the interpolating path $$Y_{\epsilon,\overline{\epsilon}}:=(Y_0+\epsilon (Y_1-Y_0), \overline{Y}_0+\overline{\epsilon}(\overline{Y}_1-\overline{Y}_0))$$ In the above display, $ (\overline{Y}_1,\overline{Y}_0)$ stands for an independent copy of a pair of random variables $(Y_{0},Y_1)$ with distribution $(\mu_{0},\mu_1)$. Also observe that $$(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}=(\mu_1-\mu_0)^{\otimes 2}\overline{D}^2_{\mu_1,\mu_0}\phi_{s,t}$$ with the centered second order operator $$\begin{array}{l} \overline{D}^2_{\mu_1,\mu_0}\phi_{s,t}(f)(x_1,x_2)\\ \\ \displaystyle:=\left[(\delta_{x_1}-\mu_0)\otimes (\delta_{x_2}-\mu_0)\right]D^2_{\mu_0}\phi_{s,t}(f)\\ \\ \displaystyle=\int_{[0,1]^2}~{\mathbb{E}}\left(\langle \left[ (\nabla\otimes\nabla) D^2_{\mu_1,\mu_0}\phi_{s,t}(f)\right](Y_{\epsilon,\overline{\epsilon}}(x_1,x_2)),(x_1-Y_0)\otimes (x_2-\overline{Y}_0)\rangle\right)~d\epsilon ~d\overline{\epsilon} \end{array}$$ In the above display, $Y_{\epsilon,\overline{\epsilon}}(x_1,x_2)$ stands for the interpolating path $$Y_{\epsilon,\overline{\epsilon}}(x_1,x_2):=(Y_0+\epsilon (x_1-Y_0), \overline{Y}_0+\overline{\epsilon}(x_2-\overline{Y}_0))$$ We have commutation formula $$\label{commutation-D-Da} (\nabla\otimes\nabla)\circ\left(D_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}=\left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}\circ (\nabla\otimes\nabla)$$ In addition, we have the estimate $$\label{estimate-D2} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert \left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2} \vert\kern-0.25ex\vert\kern-0.25ex\vert}} \leq c~ e^{-\lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ The proof of the first assertion is a consequence of the commutation formula (\[commutation-D\]). Letting $h=(\nabla\otimes\nabla) g$ we have $$\begin{array}{l} \left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}( h)(x_1,x_2)= \left( { {\cal P }}^{\mu_0}_{s,t}\right)^{\otimes 2}(h)(x_1,x_2)\\ \\ \displaystyle \hskip1cm+\int_{\Delta_{s,t}^1}\Phi_{s,v}(\mu_1)(d(u,y))~q^{\mu_1,\mu_0}_{s,u}(x_2,y)~\left({ {\cal P }}^{\mu_0}_{s,t}\otimes{ {\cal P }}^{\phi_{s,u}(\mu_0)}_{u,t}\right)(h)(x_1,y)\\ \\ \displaystyle \hskip1cm+\int_{\Delta_{s,t}^1}\Phi_{s,v}(\mu_1)(d(u,y))~q^{\mu_1,\mu_0}_{s,u}(x_1,y)~\left({ {\cal P }}^{\phi_{s,u}(\mu_0)}_{u,t}\otimes { {\cal P }}^{\mu_0}_{s,t}\right)(h)(y,x_2)\\ \\ \displaystyle \hskip1cm+\int_{\Delta_{s,t}^1\times \Delta_{s,t}^1}\Phi_{s,u}(\mu_1)(d(u,y))~\Phi_{s,v}(\mu_1)(d(v,z))\\ \\ \displaystyle \hskip3cm~\times~\left[q^{\mu_1,\mu_0}_{s,u}(x_1,y)\otimes q^{\mu_1,\mu_0}_{s,v}(x_2,z)\right]\left({ {\cal P }}^{\phi_{s,u}(\mu_0)}_{u,t}\otimes{ {\cal P }}^{\phi_{s,v}(\mu_0)}_{v,t}\right)(h)(y,z) \end{array}$$ The proof of (\[estimate-D2\]) now follows the same arguments as the ones we used in the proof of (\[commutation-D-etimate\]), thus it is skipped. This ends the proof of the proposition. [\ ]{} Combining (\[nabla-2-S\]) with the commutation formula (\[commutation-D-Da\]), for any twice differentiable function $f$ and any $s\leq t$ and $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ we check that $$\label{nabla-2-D-2} (\nabla\otimes\nabla)D^2_{\mu_0,\mu_1}\phi_{s,t}(f):=\int_s^t~\left({ {\cal D }}_{\mu_0,\mu_1}\phi_{s,u}\right)^{\otimes 2}\left( {\mathbb{S}}^{[2,1],\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)+{\mathbb{S}}^{[2,2],\phi_{s,u}(\mu_0)}_{u,t}(\nabla^2 f)\right) ~du$$ with the operators ${\mathbb{S}}^{[2,k],\mu}_{s,t}$ discussed in (\[nabla-2-S\]). The proof of (\[estimate-D2-nabla\]) is a direct consequence of (\[3s-2\]) and (\[estimate-D2\]). The proof of theorem \[theo-intro-3\] is now completed. Second order analysis {#sec-o-analysis} --------------------- This short section is mainly concerned with the proof of the first part of theorem \[theo-intro-4\]. \[lem-tex\] For any $m\geq 1$ and $\mu_0,\mu_1\in P_{m+3}({\mathbb{R}}^d)$ and $s\leq t$ we have the tensor product formula $$\begin{array}{l} \left(\phi_{s,t}(\mu_1)-\phi_{s,t}(\mu_0)\right)^{\otimes 2}\\ \\ \displaystyle=(\mu_1-\mu_0)^{\otimes 2}\left(D_{\mu_0}\phi_{s,t}\right)^{\otimes 2}+(\mu_1-\mu_0)^{\otimes 3}\,{ {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}\quad\mbox{on}\quad { {\cal C }}^{n+2}_m({\mathbb{R}}^{2d}) \end{array}$$ for some third order linear operator ${ {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}$ such that $$\Vert { {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}\Vert_{{ {\cal C }}^{n+2}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+3}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m+2}(\mu_0,\mu_1)$$ The proof of the above lemma is rather technical, thus it is housed in the appendix, on page . Combining the above lemma with (\[a-2\]) we readily check the second order decomposition (\[taylor-2-intro\]) with a the remainder linear operator $D^3_{\mu_0,\mu_1}\phi_{s,t}$ such that $$\Vert D^3_{\mu_0,\mu_1}\phi_{s,t}\Vert_{{ {\cal C }}^{n+3}_m({\mathbb{R}}^{d})\rightarrow { {\cal C }}^n_{m+4}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m+3}(\mu_0,\mu_1)$$ This ends the proof of the first part of theorem \[theo-intro-4\]. The proof of the second part of the theorem is provided in the appendix, on page . ### Acknowledgements {#acknowledgements .unnumbered} The authors are supported by the ANR Quamprocs on quantitative analysis of metastable processes. P. Del Moral is also supported in part from the Chair Stress Test, RISK Management and Financial Steering, led by the French Ecole polytechnique and its Foundation and sponsored by BNP Paribas. Appendix {#appendix .unnumbered} ======== Proof of (\[backward-D-phi-mf\]) {#backward-D-phi-mf-proof .unnumbered} -------------------------------- It is easy to check that this first assertion is true for any collection of generators $L_{t,\mu}$, thus we skip the details. The proof of the second assertion is a also a direct consequence of a more general result which is valid for any collection of generators and non necessarily symmetric functions. For any $N\geq 2$ and $x=(x^i)_{1\leq i\leq N}\in ({\mathbb{R}}^d)^N$ we set $$m(x)^{\odot 2}:=\frac{1}{N(N-1)}\sum_{1\leq i\not j\leq N}~\delta_{(x^i,x^j)}\quad\mbox{\rm and}\quad { {\cal F }}(x)=m(x)^{\otimes 2}(F)$$ We extend $L_{t,\mu}$ to functions $F(x^1,x^2)$ on ${\mathbb{R}}^{2d}$ by setting $$\begin{array}{l} L_{t,\mu}^{(2)}(F)(x^1,x^2)\\ \\ \displaystyle=\frac{1}{2}\left(L_{t,\mu}(F(x^1,{\mbox{\LARGE .}}))(x^2)+L_{t,\mu}(F({\mbox{\LARGE .}},x^2))(x^1)+L_{t,\mu}(F(x^2,{\mbox{\LARGE .}}))(x^1)+L_{t,\mu}(F({\mbox{\LARGE .}},x^1))(x^2)\right) \end{array}$$ For any function $F(x^1,x^2)$ on ${\mathbb{R}}^{2d}$ we have $$m(x)^{\otimes 2}(F)=\left(1-\frac{1}{N}\right)~m(x)^{\odot 2}(F)+\frac{1}{N}~m(x)(C(F))= m(x)^{\odot 2}\left(\left(1-\frac{1}{N}\right)~F+\frac{1}{N}~C^{(2)}(F)\right)$$ with $$C(F)(x)=F(x,x)\quad\mbox{\rm and}\quad C^{(2)}(F)(x^1,x^2)=\frac{1}{2}~\left(C(F)(x^1)+C(F)(x^2)\right)$$ This implies that $$\Lambda_t({ {\cal F }})(x)=\left(1-\frac{1}{N}\right)~m(x)^{\odot 2}\left(L_{t,m(x)}^{(2)}(F)\right)+\frac{1}{N}~m(x)^{\odot 2}\left(L_{t,m(x)}^{(2)}(C^{(2)}(F))\right)$$ Recalling that $$\displaystyle m(x)^{\odot 2}(F)=\frac{N}{N-1}~m(x)^{\otimes 2}(F)-\frac{1}{N-1}~m(x)^{\odot 2}(C^{(2)}(F))$$ we conclude that $$\Lambda_t({ {\cal F }})(x)=m(x)^{\otimes 2}\left(L_{t,m(x)}^{(2)}(F)\right)+\frac{1}{N}~m(x)^{\odot 2}\left(\Gamma^{(2)}_{L_{t,m(x)}}(F)\right)$$ with the operator $$\Gamma^{(2)}_{L_{t,m(x)}}=L_{t,m(x)}^{(2)}\circ C^{(2)}-C^{(2)}\circ L_{t,m(x)}^{(2)}$$ Observe that $$\begin{array}{l} \displaystyle\Gamma^{(2)}_{L_{t,m(x)}}(F)(x^1,x^2)\\ \\ \displaystyle=\frac{1}{2}\left(L_{t,m(x)}(C(F))(x^1)+L_{t,m(x)}(C(F))(x^2)\right)\\ \\ \displaystyle-\frac{1}{2}~\left(L_{t,m(x)}(F(x^1,{\mbox{\LARGE .}}))(x^1)+L_{t,m(x)}(F({\mbox{\LARGE .}},x^1))(x^1)+ L_{t,m(x)}(F(x^2,{\mbox{\LARGE .}}))(x^2)+L_{t,m(x)}(F({\mbox{\LARGE .}},x^2))(x^2) \right) \end{array}$$ This yields the formula $$\Gamma^{(2)}_{L_{t,m(x)}}(F)(x^1,x^2)=\frac{1}{2}\left(C(\Gamma^{(2)}_{L_{t,m(x)}}(F))(x^1)+C(\Gamma^{(2)}_{L_{t,m(x)}}(F))(x^2)\right)$$ from which we conclude that $$\Lambda_t({ {\cal F }})(x)=m(x)^{\otimes 2}\left(L_{t,m(x)}^{(2)}(F)\right)+\frac{1}{N}~m(x)\left(\Gamma_{L_{t,m(x)}}(F)\right)$$ with the function $\Gamma_{L_{t,m(x)}}(F)$ defined for any $y\in{\mathbb{R}}^d$ by $$\begin{aligned} \Gamma_{L_{t,m(x)}}(F)(y)&=&C\left(\Gamma^{(2)}_{L_{t,m(x)}}(F)\right)(y,y)\\ &=& L_{t,m(x)}(C(F))(y)-L_{t,m(x)}(F(y,{\mbox{\LARGE .}}))(y)-L_{t,m(x)}(F({\mbox{\LARGE .}},y))(y)\end{aligned}$$ The above formula readily implies (\[backward-D-phi-mf\]) as soon as $L_{t,\mu}$ is the collection of generators associated with the stochastic flow defined in (\[diff-st-ref-general\]). This ends the proof of (\[backward-D-phi-mf\]).[\ ]{} Proof of (\[tau-n-estimates\]) {#tau-n-estimates-proof .unnumbered} ------------------------------ For any given $1\leq m\leq n$, we denote by $\Pi_{n,m}$ the set of partitions $\pi=\{\pi_1,\ldots,\pi_m\}$ of the set $\{1,\ldots,n\}$ with $m$ blocks $\pi_i$ of size $\vert \pi_i\vert$, with $i\in \{1,\ldots, m\}$. We also let $\Pi_{n}$ the set of partitions of the set $\{1,\ldots,n\}$ and $\flat(\pi)$ the number of blocks in a given partition $\pi$, and $\Pi^+_{n}$ the subset of partitions $\pi$ s.t. $\flat(\pi)>1$. Let $[n]$ be the set of $m$ multiple indexes $i=(i_1,\ldots,i_n)\in \{1,\ldots,d\}^n$. For any given $i\in [n]$ and any subset $S=\{j_1,\ldots,j_s\}\subset \{1,\ldots,n\}$ we set $$i_{S}=(i_{j_1},\ldots,i_{j_s})$$ For any $x=(x^1,\ldots,x^d)\in{\mathbb{R}}^d$ and any multiple index $i\in [n]$ we write $ \partial_{i} $ instead of $\partial_{x^{i_1},\ldots x^{i_n}}=\partial_{x^{i_1}}\ldots\partial_{ x^{i_n}}$ the $n$-th partial derivatives w.r.t. the coordinates $(x^{i_1},\ldots x^{i_n})$. Let $f$ and $X$ be a couple of smooth function from ${\mathbb{R}}^d$ into itself. In this notation for any $i\in [n]$ and $1\leq j\leq d$ we have the multivariate Faà di Bruno derivation formula $$\partial_i(f^j\circ X)=\sum_{1\leq m\leq n}~\sum_{k\in [m]}~\partial_{k}f^j(X)~\sum_{\pi\in \Pi_{n,m}}(\nabla^{\pi}X)_{i,k}$$ with the $\pi$-gradient tensor $$(\nabla^{\pi}X)_{i,k}:=(\nabla^{\vert\pi_1\vert}X)_{i_{\pi_1},k_1}\ldots (\nabla^{\vert\pi_m\vert}X)_{i_{\pi_m},k_m}$$ We check the above formula by induction w.r.t. the parameter $n$. In a more compact we have checked the following lemma. For any $n\geq 1$ we have the Faà di Bruno derivation formula $$\label{faa-di-bruno-ap} \nabla^n(f\circ X)=\sum_{\pi\in \Pi_{n}}(\nabla^{\pi}X)~(\nabla^{\flat(\pi)} f)(X)$$ Whenever $X(x)$ is a random function we have $$\label{faa-di-bruno-ap-2} P(f)(x):={\mathbb{E}}((f\circ X)(x))\Longrightarrow \nabla^nP(f)=\sum_{1\leq m\leq n}~{ {\cal P }}^{[n,m]}(\nabla^m f)$$ with the collection of integral operators $${ {\cal P }}^{[n,m]}(\nabla^m f)(x):=\sum_{\pi\in \Pi_{n,m}}~{\mathbb{E}}\left((\nabla^{\pi}X(x))~\nabla^m f(X(x))\right)$$ Using the above lemma we also check the stochastic tensor evolution equation $$\begin{array}{l} \partial_t\, (\nabla^n X^{\mu}_{s,t}(x))_{i,j}\\ \\ \displaystyle= (\nabla^n X^{\mu}_{s,t}(x)~b_t^{[1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu)))_{i,j}+\sum_{1<m\leq n}\sum_{k\in [m]}\sum_{\pi\in \Pi_{n,m}}(\nabla^{\pi}X^{\mu}_{s,t}(x))_{i,k}~ b_t^{[1]_{\flat(\pi)}}(x,\phi_{s,t}(\mu))_{k,j}: \end{array}$$ with $$b_t^{[1]_m}(x,\mu)_{(k_1,\ldots,k_m),j}:=\partial_{k_1,\ldots,k_m}b^j_t(x,\mu)$$ In a more compact form we have $$\partial_t\, \nabla^n X^{\mu}_{s,t}(x)= \nabla^n X^{\mu}_{s,t}(x)~b_t^{[1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu))+\sum_{\pi\in \Pi^+_{n}}\nabla^{\pi}X^{\mu}_{s,t}(x) ~b_t^{[1]_{\flat(\pi)}}(x,\phi_{s,t}(\mu))$$ This implies that $$\begin{array}{l} \displaystyle\partial_t\, \nabla^n X^{\mu}_{s,t}(x)\nabla^n X^{\mu}_{s,t}(x)^{\prime}\\ \\ \displaystyle=\nabla^n X^{\mu}_{s,t}(x)~\left(b_t^{[1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu))+b_t^{[1]}(X^{\mu}_{s,t}(x),\phi_{s,t}(\mu))^{\prime}\right)\nabla^n X^{\mu}_{s,t}(x)^{\prime}\\ \\ \displaystyle +\sum_{\pi\in \Pi^+_{n}}\nabla^{\pi}X^{\mu}_{s,t}(x) ~\left(b_t^{[1]_{\flat(\pi)}}(x,\phi_{s,t}(\mu))+b_t^{[1]_{\flat(\pi)}}(x,\phi_{s,t}(\mu))^{\prime}\right)~\nabla^n X^{\mu}_{s,t}(x)^{\prime} \end{array}$$ Taking the trace in the above display, we check that $$\partial_t\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert^2_{\tiny Frob}\leq~-2\lambda_1~\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^2+2\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob} ~ \sum_{\pi\in \Pi^+_{n}}\Vert b^{[1]_{\flat(\pi)}}\Vert_{\tiny Frob}~\Vert\nabla^{\pi} X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}$$ This yields the rather crude estimate $$\begin{array}{l} \partial_t\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert^2_{\tiny Frob}\\ \\ \displaystyle \leq~-2\lambda_1~\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^2+c_{n}~\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob} ~ \sum_{\pi\in \Pi^+_{n}}~\Vert\nabla^{\vert{\pi_1\vert}} X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\ldots \Vert\nabla^{\vert{\pi_{\flat(\pi)}\vert}} X^{\mu}_{s,t}(x)\Vert_{\tiny Frob} \end{array}$$ from which we check that $$\begin{array}{l} \partial_t\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\\ \\ \displaystyle \leq~-\lambda_1~\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}+c_{n}~ ~ \sum_{}~\Vert\nabla X^{\mu}_{s,t}(x)\Vert^{l_1}_{\tiny Frob}~\Vert\nabla^2 X^{\mu}_{s,t}(x)\Vert^{l_2}_{\tiny Frob}\ldots \Vert\nabla^{n-1} X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}^{l_{n-1}} \end{array}$$ The summation in the above display is taken over all indices $l_1,\ldots,l_{n-1}$ such that $l_1+\ldots+l_{n-1}=m$ and $l_1+2l_2+\ldots+(n-1)l_{n-1}=n$ and $1<m\leq n$. Assume that (\[tau-n-estimates\]) has been checked up to rank $(n-1)$. In this case, we have $$\Vert\nabla^n X^{\mu}_{s,t}(x)\Vert_{\tiny Frob}\leq c_{n,1}~e^{-\lambda_1(t-s)}~\int_s^t~e^{\lambda_1(u-s)}~e^{-2\lambda_1(u-s)}~du\leq c_{n,2}~e^{-\lambda_1(t-s)}$$ This ends the proof of (\[tau-n-estimates\]).[\ ]{} Proof of (\[bismut-omega\]) and (\[bismut-omega-2\]) {#bismut-omega-proof .unnumbered} ---------------------------------------------------- We recall the backward formula $$P^{\mu}_{s,t}(f)(x)=f(x)+\int_s^tL_{u,\phi_{s,u}(\mu)}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(x)~du$$ A detailed proof of the above formula based on backward stochastic flows can be found in theorem 3.1 in the article [@mp-var-18]. This implies that $$d\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)(X^{\mu}_{u,t}(x))\right)=\left(\nabla P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(X^{\mu}_{u,t}(x))^{\prime}~dW_u$$ from which we check that $$f(X^{\mu}_{s,t}(x))= P^{\mu}_{s,t}(f)(x)+\int_s^t\left(\nabla P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(X^{\mu}_{u,t}(x))^{\prime}~dW_u$$ This yields the formula $$\begin{array}{l} {\mathbb{E}}\left(f(X_{s,t}^{\mu}(x))~ \tau^{\mu,\omega}_{s,t}(x)\right)\\ \\ \displaystyle={\mathbb{E}}\left(\left(\int_s^t\left(\nabla P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(X^{\mu}_{u,t}(x))^{\prime}~dW_u\right)\left(\int_s^t~ \partial_u \omega_{s,t}(u)~\nabla X^{\mu}_{s,u}(x)~dW_u \right)\right)\\ \\ \displaystyle={\mathbb{E}}\left(\int_s^t~\nabla \left(P^{\phi_{s,u}(\mu)}_{u,t}(f)\circ X^{\mu}_{u,t}\right)(x)~ \partial_u \omega_{s,t}(u)~du\right) \end{array}$$ We conclude that $$\begin{array}{l} {\mathbb{E}}\left(f(X_{s,t}^{\mu}(x))~ \tau^{\mu,\omega}_{s,t}(x)\right)\\ \\ \displaystyle=\nabla P^{\mu}_{s,t}(f)(x)~{\mathbb{E}}\left(\int_s^t \partial_u \omega_{s,t}(u)~du\right)=\nabla P^{\mu}_{s,t}(f)(x)~(\omega_{s,t}(t)-\omega_{s,t}(s))=\nabla P^{\mu}_{s,t}(f)(x) \end{array}$$ This ends the proof of (\[bismut-omega\]). For any $s\leq u\leq t$ applying (\[bismut-omega\]) to the function $P^{\phi_{s,u}(\mu)}_{u,t}(f)$ we have $$\begin{aligned} \nabla P^{\mu}_{s,t}(f)(x)&=&\nabla P^{\mu}_{s,u}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(x)\\ &=&{\mathbb{E}}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)(X^{\mu}_{s,u}(x))~\int_s^u~ \partial_v \omega_{s,u}(v)~\nabla X^{\mu}_{s,v}(x)~dW_v \right)\end{aligned}$$ This implies that $$\begin{array}{l} \partial_{x_j,x_i} P^{\mu}_{s,t}(f)(x)\\ \\ =\displaystyle\sum_{1\leq l\leq d}{\mathbb{E}}\left(\partial_{x_j}X^{\mu,l}_{s,u}(x)~\partial_{x_l}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(X^{\mu}_{s,u}(x))~ \int_s^u~ \partial_v \omega_{s,u}(v)~\partial_{x_i} X^{\mu,k}_{s,v}(x)~dW^k_v\right)\\ \\ \hskip3cm\displaystyle+{\mathbb{E}}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)(X^{\mu}_{s,u}(x))~\int_s^u~ \partial_v \omega_{s,u}(v)~\partial_{x_j,x_i} X^{\mu,k}_{s,v}(x)~dW^k_v\right)\ \end{array}$$ Applying (\[bismut-omega\]) to the first term we check that $$\begin{array}{l} \displaystyle{\mathbb{E}}\left(\partial_{x_j}X^{\mu,l}_{s,u}(x)~\partial_{x_l}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)\right)(X^{\mu}_{s,u}(x))~ \tau^{\mu,\omega}_{s,u}(x)_i\right)\\ \\ =\displaystyle \displaystyle{\mathbb{E}}\left(f(X^{\mu}_{s,t}(x))~~\partial_{x_j}X^{\mu,l}_{s,u}(x)~ \tau^{\mu,\omega}_{s,u}(x)_i \left(\sum_{1\leq m\leq d} \int_u^t~\partial_v \omega_{u,t}(v)~\left(\partial_{x_l}X_{u,v}^{\phi_{s,u}(\mu),m}\right)(X_{s,u}^{\mu}(x))~dW^m_v \right) \right)\\ \\ =\displaystyle {\mathbb{E}}\left(f(X^{\mu}_{s,t}(x))~~\partial_{x_j}X^{\mu,l}_{s,u}(x)~ ~\tau^{\phi_{s,u}(\mu),\omega}_{u,t}(X_{s,u}^{\mu}(x))_l~\tau^{\mu,\omega}_{s,u}(x)_i\right) \end{array}$$ We conclude that $$\begin{array}{l} \nabla^2P^{\mu}_{s,t}(f)(x)_{i,j}=\nabla^2P^{\mu}_{s,t}(f)(x)_{j,i}\\ \\ =\displaystyle{\mathbb{E}}\left(f(X^{\mu}_{s,t}(x))~~\nabla X^{\mu}_{s,u}(x)_{j,l}~ ~\tau^{\phi_{s,u}(\mu),\omega}_{u,t}(X_{s,u}^{\mu}(x))_l~\tau^{\mu,\omega}_{s,u}(x)_i\right)\\ \\ \hskip3cm\displaystyle+{\mathbb{E}}\left(P^{\phi_{s,u}(\mu)}_{u,t}(f)(X^{\mu}_{s,u}(x))~\int_s^u~ \partial_v \omega_{s,u}(v)~\nabla^2 X^{\mu}_{s,v}(x)_{(i,j),k}~dW^k_v\right)\ \end{array}$$ This ends the proof of (\[bismut-omega-2\]).[\ ]{} Proof of (\[linear-example-d\]) {#linear-example-d-proof .unnumbered} ------------------------------- We have $$\begin{array}{l} \displaystyle p_{s,t}^{\mu_1,\mu_0}(x,z)=b^{\mu_0}_{s,t}(x,z) +\sum_{n\geq 1} \int_{[s,t]_n}~B_2~e^{(t-u_n)B_1} B_2~e^{(u_n-u_{n-1})B_1}\ldots~B_2~e^{(u_2-u_1)B_1}~\\ \\ \displaystyle\times\left(B_1~\phi_{s,u_1}(\mu_1)(e)+ B_2\left(e^{(u_1-s)B_1}(x-\mu_0(e))+e^{(u_1-s)(B_1+B_2)}\mu_0(e)\right) \right)~du_1\ldots du_n~\end{array}$$ Recalling that $$\begin{aligned} \phi_{s,u_1}(\mu_1)(e)&=& e^{(u_1-s)[B_1+B_2]}~\mu_1(e)\\ b_{s,t}^{\mu_0}(x,z)&=&B_1z+B_2\left[e^{(t-s)B_1}(x-\mu_0(e))+e^{(t-s)[B_1+B_2]}~\mu_0(e)\right]\end{aligned}$$ and using the rather well known exponential formulae $$\begin{aligned} e^{(t-s)(B_1+B_2)}&=&e^{(t-s)B_1}+\int_s^t~e^{(t-u)B_1}\,B_2\,e^{(u-s)(B_1+B_2)}~du\\ &=&e^{(t-s)B_1}+\int_s^t~e^{(t-u)(B_1+B_2)}\,B_2\,e^{(u-s)B_1}~du\end{aligned}$$ we check that $$\begin{array}{l} \displaystyle p_{s,t}^{\mu_1,\mu_0}(x,z)=b^{\mu_0}_{s,t}(x,z) \\ \\ \displaystyle+B_2~\int_s^t~e^{(t-u_1)(B_1+B_2)}~\left(B_1~\phi_{s,u_1}(\mu_1)(e)+ B_2\left(e^{(u_1-s)B_1}(x-\mu_0(e))+e^{(u_1-s)(B_1+B_2)}\mu_0(e)\right) \right)~du_1~\end{array}$$ from which we find that $$\begin{array}{l} \displaystyle p_{s,t}^{\mu_1,\mu_0}(x,z)=B_1z+B_2\left[e^{(t-s)B_1}(x-\mu_0(e))+e^{(t-s)[B_1+B_2]}~\mu_0(e)\right]\\ \\ \displaystyle +B_2~\left[\int_s^t e^{(t-u_1)(B_1+B_2)}~B_1~e^{(u_1-s)[B_1+B_2]}~du_1\right] ~\mu_1(e)~~\\ \\ \displaystyle +B_2~\left[\int_s^t e^{(t-u_1)(B_1+B_2)}~B_2~e^{(u_1-s)(B_1+B_2)}~du_1\right]~ \mu_0(e) +B_2~\left[e^{(t-s)(B_1+B_2)}-e^{(t-s)B_1}\right] ~(x-\mu_0(e)) \end{array}$$ This ends the proof of (\[linear-example-d\]).[\ ]{} Proof of (\[nabla-D-times-2\]) {#nabla-D-times-2-poof .unnumbered} ------------------------------ We have the tensor product formula $$\left(D_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}:=\left(P^{\mu_0}_{s,t}\right)^{\otimes 2}+\left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}+Q^{\mu_1,\mu_0}_{s,t}\otimes P^{\mu_0}_{s,t}+P^{\mu_0}_{s,t}\otimes Q^{\mu_1,\mu_0}_{s,t}$$ We also have $$\begin{array}{l} \displaystyle \left(Q^{\mu_1,\mu_0}_{s,t}\otimes P^{\mu_0}_{s,t}\right)(g)(x,\overline{x})\\ \\ \displaystyle=\int_{\Delta_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~(b^{\mu_0}_{s,u}(x,y)^{\prime}\otimes I)~\left({ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}\otimes P^{\mu_0}_{s,t}\right)(\nabla_{x_1}g)(y,\overline{x})\end{array}$$ Recall that $b_{t}(x,y)$ is differentiable at any order with uniformly bounded derivatives. Thus all differentials of the above function w.r.t. the coordinate $x$ have uniformly bounded derivatives. On the other hand, the mapping $x\mapsto b_{t}(x,y)$ has at most linear growth. Thus, using the estimates (\[ref-moments\]) and (\[tau-n-estimates\]), for any $m\geq 0$ we check that $$\Vert Q^{\mu_1,\mu_0}_{s,t}\otimes P^{\mu_0}_{s,t}\Vert_{{ {\cal C }}^{n+1}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+1}({\mathbb{R}}^{2d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ In the same vein, we have the tensor product formula $$\begin{array}{l} \displaystyle \left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}(g)(x,\overline{x})=\left({ {\cal Q }}^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}((\nabla\otimes\nabla) g)(x,\overline{x})\\ \\ \displaystyle:=\int_{\Delta_{s,t}\times \Delta_{s,t}}~\left[\Phi_{s,u}(\mu_1)\otimes \Phi_{s,u}(\mu_1)\right](d((u,y),(\overline{u},\overline{y})))~\\ \displaystyle\hskip7cm~\widehat{b}^{\,\mu_0}_{s,u,\overline{u}}((x,\overline{x}),(y,\overline{y}))^{\prime}~\widehat{{ {\cal P }}}^{\,\phi_{s,u,\overline{u}}(\mu_0)}_{u,\overline{u},t}((\nabla\otimes\nabla) g)(y,\overline{y}) \end{array}$$ with $$\widehat{b}^{\,\mu_0}_{s,u,\overline{u}}((x,\overline{x}),(y,\overline{y}))^{\prime}:=b^{\,\mu_0}_{s,u}(x,y)^{\prime}\otimes b^{\,\mu_0}_{s,\overline{u}}(\overline{x},\overline{y})^{\prime} \quad\mbox{\rm and}\quad \widehat{{ {\cal P }}}^{\,\phi_{s,u,\overline{u}}(\mu_0)}_{u,\overline{u},t}:={ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}\otimes { {\cal P }}^{\,\phi_{s,\overline{u}}(\mu_0)}_{\overline{u},t}$$ Arguing as above and using the estimates (\[ref-moments\]) and (\[tau-n-estimates\]) for any $m\geq 0$ we check that $$\Vert \left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}\Vert_{{ {\cal C }}^{2}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{2}({\mathbb{R}}^{2d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ Proof of lemma \[lem-tex\] {#lem-tex-proof .unnumbered} -------------------------- Using the decomposition $$\begin{array}{l} \phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\\ \\ \displaystyle=\sum_{1\leq l\leq n}\left[\phi_{s,u_1}(\mu_1)\otimes\ldots\otimes \phi_{s,u_{l-1}}(\mu_1) \right]\otimes \left[\phi_{s,u_l}(\mu_1)-\phi_{s,u_l}(\mu_0)\right]\otimes\left[\phi_{s,u_{l+1}}(\mu_{0})\otimes\ldots\otimes \phi_{s,u_{n}}(\mu_0) \right] \end{array}$$ which is valid for any $\mu_0,\mu_1\in P_2({\mathbb{R}}^d)$ and any $u=(u_1,\ldots,u_n)\in [s,t]_n$ with $n\geq 1$, for any function $$(u,y)\in \Delta_{s,t}\mapsto h_u(y)\in {\mathbb{R}}$$ we check that $$\label{decomp-Phi-k} \int_{\Delta_{s,t}}~\left[\Phi_{s,u}(\mu_1)-\Phi_{s,u}(\mu_0)\right](d(u,y))~h_u(y)=\int_{\Delta^1_{s,t}}~\left[\Phi_{s,v}(\mu_1)-\Phi_{s,v}(\mu_0)\right](d(v,z))~\overline{h}_{v}(z)$$ with the function $$\begin{array}{l} \displaystyle\overline{h}_{v}(z):=h_v(z)+\int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~h_{u,v}(y,z)+\int_{\Delta_{v,t}}~\Phi_{v,u}\left(\phi_{s,v}(\mu_0)\right)(d(u,y)) ~h_{v,u}(z,y) \\ \\ \hskip3cm \displaystyle+\int_{\Delta_{s,v}\times \Delta_{v,t}}~\Upsilon^{\mu_1,\mu_0}_{s,t}((v,z),d((u,y),(\overline{u},\overline{y})))~h_{(u,v,\overline{u})}(y,z,\overline{y}) \end{array}$$ In the above display, $\Upsilon^{\mu_1,\mu_0}_{s,t}$ stands for the tensor product measures $$\Upsilon^{\mu_1,\mu_0}_{s,t}((v,z),d((u,y),(\overline{u},\overline{y})))=\Phi_{s,u}(\mu_1)(d(u,y))~\Phi_{v,\overline{u}}\left(\phi_{s,v}(\mu_0)\right)(d(\overline{u},\overline{y}))$$ We also have the tensor product formula $$\begin{array}{l} \displaystyle\left(D_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}-\left(D_{\mu_0}\phi_{s,t}\right)^{\otimes 2}\\ \\ \displaystyle=\left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}-\left(Q^{\mu_0}_{s,t}\right)^{\otimes 2}+\left(Q^{\mu_1,\mu_0}_{s,t}-Q^{\mu_0}_{s,t}\right)\otimes P^{\mu_0}_{s,t}+P^{\mu_0}_{s,t}\otimes\left(Q^{\mu_1,\mu_0}_{s,t}-Q^{\mu_0}_{s,t}\right) \end{array}$$ This yields the decomposition $$\begin{array}{l} \displaystyle \left(\left[Q^{\mu_1,\mu_0}_{s,t}-Q^{\mu_0}_{s,t}\right]\otimes P^{\mu_0}_{s,t}\right)(g)(x,\overline{x})\\ \\ \displaystyle:=\int_s^t~\int~\left[\phi_{s,v}(\mu_1)-\phi_{s,v}(\mu_0)\right](d\widehat{x})~{ {\cal I }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},\widehat{x})~dv\end{array}$$ with the integral operator $$\begin{array}{l} \displaystyle{ {\cal I }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},\widehat{x})\\ \\ :=b^{\mu_0}_{s,v}(x,\widehat{x})^{\prime}~\left({ {\cal P }}^{\,\phi_{s,v}(\mu_0)}_{v,t}\otimes P^{\mu_0}_{s,t}\right)(\nabla_{x_1}g)(\widehat{x},\overline{x})\\ \\ \displaystyle+\int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u,v}(x,y,\widehat{x})^{\prime}~\left({ {\cal P }}^{\,\phi_{s,v}(\mu_0)}_{v,t}\otimes P^{\mu_0}_{s,t}\right)(\nabla_{x_1}g)(\widehat{x},\overline{x})\\ \\ \displaystyle+\int_{\Delta_{v,t}}~\Phi_{v,u}\left(\phi_{s,v}(\mu_0)\right)(d(u,y))~b^{\mu_0}_{s,v,u}(x,\widehat{x},y)^{\prime}~\left({ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}\otimes P^{\mu_0}_{s,t}\right)(\nabla_{x_1}g)(y,\overline{x})\\ \\ +\displaystyle\int_{\Delta_{s,v}\times \Delta_{v,t}}~\Upsilon^{\mu_1,\mu_0}_{s,t}((v,z),d((u,y),(\overline{u},\overline{y})))~b^{\mu_0}_{s,u,v,\overline{u}}(x,y,\widehat{x},\overline{y})^{\prime}~\left({ {\cal P }}^{\,\phi_{s,\overline{u}}(\mu_0)}_{\overline{u},t}\otimes P^{\mu_0}_{s,t}\right)(\nabla_{x_1}g)(\overline{y},\overline{x})\end{array}$$ Arguing as in the proof of (\[crude-est-rev\]) and (\[nabla-D\]) we check that $$\Vert { {\cal I }}_{s,v,t}^{\mu_0,\mu_1}\Vert_{{ {\cal C }}^{n+1}_{m}({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+2}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ In the same vein, we have $$\begin{array}{l} \displaystyle \left[\left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}-\left(Q^{\mu_0}_{s,t}\right)^{\otimes 2}\right](g)(x,\overline{x})\\\ \\ =\displaystyle\int_{\Delta_{s,t}}~\left[\Phi_{s,u}(\mu_1)-\Phi_{s,u}(\mu_0)\right](d(u,y))~\left[\Theta_{s,u,t}^{\mu_1,\mu_0}+\overline{\Theta}_{s,u,t}^{\mu_1,\mu_0}\right](g)(x,\overline{x},y)~dv \end{array}$$ with $$\overline{\Theta}_{s,u,t}^{\mu_1,\mu_0}(g)(x,\overline{x},y):= \int_{\Delta_{s,t}}~\Phi_{s,\overline{u}}(\mu_1)(d(\overline{u},\overline{y}))~ \widehat{b}^{\,\mu_0}_{s,u,\overline{u}}((x,\overline{x}),(y,\overline{y}))^{\prime}~\widehat{{ {\cal P }}}^{\,\phi_{s,u,\overline{u}}(\mu_0)}_{u,\overline{u},t}((\nabla\otimes\nabla) g)(y,\overline{y})$$ and $$\Theta_{s,u,t}^{\mu_1,\mu_0}(g)(x,\overline{x},y):= \int_{\Delta_{s,t}}~\Phi_{s,\overline{u}}(\mu_0)(d(\overline{u},\overline{y}))~ \widehat{b}^{\,\mu_0}_{s,\overline{u},u}((x,\overline{x}),(\overline{y},y))^{\prime}~\widehat{{ {\cal P }}}^{\,\phi_{s,\overline{u},u}(\mu_0)}_{\overline{u},u,t}((\nabla\otimes\nabla) g)(\overline{y},y)$$ This yields the formula $$\begin{array}{l} \displaystyle \left[\left(Q^{\mu_1,\mu_0}_{s,t}\right)^{\otimes 2}-\left(Q^{\mu_0}_{s,t}\right)^{\otimes 2}\right](g)(x,\overline{x})\\ \\ \displaystyle=\int_s^t~\left[\phi_{s,v}(\mu_1)-\phi_{s,v}(\mu_0)\right](d\widehat{x})~{ {\cal J }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},\widehat{x})~dv \end{array}$$ with the integral operator $$\begin{array}{l} \displaystyle{ {\cal J }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},\widehat{x})\\ \\ \displaystyle:=\left[\Theta_{s,v,t}^{\mu_1,\mu_0}+\overline{\Theta}_{s,v,t}^{\mu_1,\mu_0}\right](g)(x,\overline{x},\widehat{x})\\ \\ \displaystyle+\int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~\left[\Theta_{s,(u,v),t}^{\mu_1,\mu_0}+\overline{\Theta}_{s,(u,v),t}^{\mu_1,\mu_0}\right](g)(x,\overline{x},(y,\widehat{x}))\\ \\ \displaystyle+\int_{\Delta_{v,t}}~\Phi_{v,u}\left(\phi_{s,v}(\mu_0)\right)(d(u,y))~\left[\Theta_{s,(v,u),t}^{\mu_1,\mu_0}+\overline{\Theta}_{s,(v,u),t}^{\mu_1,\mu_0}\right](g)(x,\overline{x},(\widehat{x},y))\\ \\ \displaystyle+\int_{\Delta_{s,v}\times \Delta_{v,t}}~\Upsilon^{\mu_1,\mu_0}_{s,t}((v,z),d((u,y),(\overline{u},\overline{y})))~\left[\Theta_{s,(u,v,\overline{u}),t}^{\mu_1,\mu_0}+\overline{\Theta}_{s,(u,v,\overline{u}),t}^{\mu_1,\mu_0}\right](g)(x,\overline{x},(y,\widehat{x},\overline{y})) \end{array}$$ Arguing as above, we check that $$\Vert { {\cal J }}_{s,v,t}^{\mu_0,\mu_1}\Vert_{{ {\cal C }}^{2}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{2}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ Combining the above decompositions we find that $$\begin{array}{l} \left[\left(D_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}-\left(D_{\mu_0}\phi_{s,t}\right)^{\otimes 2}\right](g)(x,\overline{x})\\ \\ \displaystyle=\int_s^t~\left[\phi_{s,v}(\mu_1)-\phi_{s,v}(\mu_0)\right](d\widehat{x})~{ {\cal K }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},\widehat{x})~dv\quad\mbox{\rm with}\quad{ {\cal K }}_{s,v,t}^{\mu_0,\mu_1}:= 2~{ {\cal I }}_{s,v,t}^{\mu_0,\mu_1}+{ {\cal J }}_{s,v,t}^{\mu_0,\mu_1} \end{array}$$ For any $n\geq 2$ and $m\geq 0$ we have $$\Vert { {\cal K }}_{s,v,t}^{\mu_0,\mu_1}\Vert_{{ {\cal C }}^{n+1}_{m}({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+2}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m\vee 2}(\mu_0,\mu_1)$$ We conclude that $$\left(\phi_{s,t}(\mu_1)-\phi_{s,t}(\mu_0)\right)^{\otimes 2}=(\mu_1-\mu_0)^{\otimes 2}\left(D_{\mu_0}\phi_{s,t}\right)^{\otimes 2}+(\mu_1-\mu_0)^{\otimes 3}{ {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}$$ with the operator $$\begin{array}{l} { {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}(g)(x,\overline{x},\widehat{x})\\ \\ \displaystyle :=\int_s^t \left[\int~P^{\mu_0}_{s,v}(\widehat{x},dz)~{ {\cal K }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},z)+\int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(\widehat{x},y)^{\prime}~{ {\cal L }}_{s,u,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},y)\right]~dv \end{array}$$ In the above display, ${ {\cal L }}_{s,u,v,t}^{\mu_0,\mu_1}$ stands for the integral operator operator $$\begin{aligned} { {\cal L }}_{s,u,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},y) &=&{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}\left(\nabla_{x_3}{ {\cal K }}_{s,v,t}^{\mu_0,\mu_1}(g)(x,\overline{x},{\mbox{\LARGE .}})\right)(y) $$ We also check that $$\Vert { {\cal R }}_{\mu_1,\mu_0}\phi_{s,t}\Vert_{{ {\cal C }}^{n+2}_m({\mathbb{R}}^{2d})\rightarrow { {\cal C }}^n_{m+3}({\mathbb{R}}^{3d})}\leq c_{m,n}(t)~\rho_{m+2}(\mu_1,\mu_2)$$ This ends the proof of the lemma. [\ ]{} Proof of the estimate (\[D3-estimation\]) {#D3-estimation-lem-proof .unnumbered} ----------------------------------------- For any $x=(x_1,x_2)\in {\mathbb{R}}^{2d}$ we set $\sigma(x_1,x_2):=\sigma(x_2,x_1)$. In this notation, for any matrix valued function $h(x)=(h_{i,j}(x))_{1\leq i,j\leq d}$ we have the tensor product formula $$\begin{array}{l} \displaystyle\left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}( h)(x)\\ \\ \displaystyle = \left( { {\cal P }}^{\mu_0}_{s,t}\right)^{\otimes 2}(h)(x) +\int_{\Delta_{s,t}}\Phi_{s,v}(\mu_1)(d(u,y))~ \left[ {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)+{\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(\sigma(x),y)\right]\\ \\ \displaystyle \hskip3cm+\int_{\Delta_{s,t}\times \Delta_{s,t}}\Phi_{s,u}(\mu_1)(d(u,y))~\Phi_{s,v}(\mu_1)(d(v,z))~ {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z) \end{array}$$ with the matrix valued functions $ {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)$ and $ {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)$ given for any $(u,y)\in \Delta^n_{s,t}$ and $(v,z)\in \Delta^m_{s,t}$ by the formula $$\begin{aligned} {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)&:=&\BB^{\mu_0}_{s,u}(x_1,y)~\left({ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\mu_0}_{s,t}\right)(h)(y_n,x_2)\\ {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z)&:=&\left[\BB^{\mu_0}_{s,u}(x_1,y)\otimes \BB^{\mu_0}_{s,v}(x_2,z)\right]\left({ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_m}(\mu_0)}_{v_m,t}\right)(h)(y_n,z_m) \end{aligned}$$ Using (\[ref-nabla-Pa\]) we have $$\nabla{ {\cal P }}^{\mu}_{s,t}(g)={ {\cal P }}^{[2,1],\mu}_{s,t}(g)+{ {\cal P }}^{[2,2],\mu}_{s,t}(\nabla g)$$ from which we check the formula $$\begin{array}{l} \nabla_{y_n} \left({ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_m}(\mu_0)}_{v_m,t}\right)(h)(y_n,z_m)\\ \\ =\left[{ {\cal P }}^{[2,1],\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_n}(\mu_0)}_{v_n,t}\right](h)(y_n,z_m)+\left[{ {\cal P }}^{[2,2]\phi_{s,u}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_m}(\mu_0)}_{v_m,t}\right](\nabla_{x_1}h)(y_n,z_m) \end{array}$$ By symmetry arguments, we also have $$\begin{array}{l} \nabla_{z_m} \left({ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_m}(\mu_0)}_{v_m,t}\right)(h)(y_n,z_m)\\ \\ =\left[{ {\cal P }}^{[2,1],\phi_{s,v_m}(\mu_0)}_{v_m,t}\otimes{ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\right](h)(z_m,y_n)+\left[{ {\cal P }}^{[2,2],\phi_{s,v_m}(\mu_0)}_{v_m,t}\otimes{ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\right](\nabla_{x_1}h)(z_m,y_n) \end{array}$$ Using (\[estimates-U-V\]) for any differentiable matrix valued function $h(x_1,x_2)$ such that $\Vert h\Vert\vee \Vert \nabla_{x_1}h\Vert\leq 1$ we have the uniform estimate $$\Vert \nabla_{y_n} \left({ {\cal P }}^{\phi_{s,u_n}(\mu_0)}_{u_n,t}\otimes{ {\cal P }}^{\phi_{s,v_m}(\mu_0)}_{v_m,t}\right)(h)(y_n,z_m)\Vert\leq c_1~e^{-\lambda_1[(t-u_n)+(t-v_m)]}$$ In the same vein, we have $$\begin{array}{l} \BB^{\mu_0}_{s,u}(x,y)= \BB_{s,u}^{[1],\mu_0}(x,y)\\ \\ \displaystyle={\mathbb{E}}\left[\nabla X_{s,u_1}^{\mu_0}(x)~b^{[2]}_{u_1}(y_1,X_{s,u_1}^{\mu_0}(x))\right]\prod_{1\leq l<n}{\mathbb{E}}\left[\nabla X_{u_l,u_{l+1}}^{\phi_{s,u_l}(\mu_0)}(y_l)~b^{[2]}_{u_{l+1}}(y_{l+1},X_{u_l,u_{l+1}}^{\phi_{s,u_l}(\mu_0)}(y_l))\right] \end{array}$$ Using the gradient and the Hessian estimates (\[tau-1-estimates\]) and (\[tau-2-estimates\]) for any $1\leq k\leq n$ we check that $$\Vert \nabla_{y_k}\BB^{\mu_0}_{s,u}(x_1,y)\Vert\leq c_2~\Vert b^{[2]}\Vert_2^n~e^{-\lambda_1(u_n-s)}~$$ Combining the above estimates with (\[estimate-BB\]) we check that $$\begin{array}{l} \Vert \nabla_{y_n} {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)\Vert\\ \\ \displaystyle\leq c_{3}~\Vert b^{[2]}\Vert_2^n~\left[ e^{-\lambda_1(u_n-s)}~e^{-\lambda_1[(t-u_n)+(t-s)]}+e^{-\lambda_1[(u_n-s)]}~e^{-\lambda_1[(t-u_n)+(t-s)]}\right]\\ \\ \displaystyle\leq c_{4}~\Vert b^{[2]}\Vert_2^n~ e^{-2\lambda_1(t-s)}~ \end{array}$$ In addition, for any $1\leq k<n$ we have $$\begin{array}{l} \Vert \nabla_{y_k} {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)\Vert\\ \\ \displaystyle \leq c_{5}~\Vert b^{[2]}\Vert_2^n~e^{-\lambda_1(u_n-s)}~e^{-\lambda_1[(t-u_n)+(t-s)]} \leq c_{5}~\Vert b^{[2]}\Vert_2^n~e^{-2\lambda_1(t-s)} \end{array}$$ We conclude that $$\label{estimate-II} \sup_{1\leq k\leq n}\Vert \nabla_{y_k} {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)\Vert \leq c~\Vert b^{[2]}\Vert_2^n~e^{-2\lambda_1(t-s)}$$ Arguing as above, for any $1\leq k<n$ we have $$\begin{array}{l} \Vert \nabla_{y_k} {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z)\Vert\\ \\ \displaystyle \leq c_{1}~ \Vert b^{[2]}\Vert_2^{m+n}~~e^{-\lambda_1(u_n-s)}~e^{-\lambda_1(v_m-s)}~e^{-\lambda_1[(t-u_n)+(t-v_m)]} \leq c_{2}~ \Vert b^{[2]}\Vert_2^{m+n}~~e^{-2\lambda_1(t-s)} \end{array}$$ In addition, for $k=n$ we have $$\begin{array}{l} \Vert \nabla_{y_n} {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z)\Vert\\ \\ \displaystyle \leq c_{3}~ \Vert\nabla_{x_2}b\Vert_2^{m+n}~\left[ e^{-\lambda_1(u_n-s)}~e^{-\lambda_1(v_m-s)}~e^{-\lambda_1[(t-u_n)+(t-v_m)]}\right.\\ \\ \hskip3cm\left.+e^{-\lambda_1(u_n-s)}~e^{-\lambda_1(v_m-s)}~e^{-\lambda_1[(t-u_n)+(t-v_m)]} \right] \end{array}$$ This implies that $$\label{estimate-JJ} \sup_{1\leq k\leq n} \Vert \nabla_{y_k} {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z)\Vert \leq c~ \Vert b^{[2]}\Vert_2^{m+n}~~e^{-2\lambda_1(t-s)}$$ On the other hand, we have the decomposition $$\displaystyle\left[\left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}-\left({ {\cal D }}_{\mu_0}\phi_{s,t}\right)^{\otimes 2}\right]( h)(x) =\int_{\Delta_{s,t}}\left[\Phi_{s,v}(\mu_1)-\Phi_{s,v}(\mu_0)\right](d(u,y))~ {\mathbb{K}}^{\,\mu_0,\mu_1}_{s,u,t}(h)(x,y)$$ with the matrix valued function $$\begin{array}{l} \displaystyle{\mathbb{K}}^{\,\mu_0,\mu_1}_{s,u,t}(h)(x,y):= {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(x,y)+ {\mathbb{I}}^{\,\mu_0}_{s,u,t}(h)(\sigma(x),y) +\int_{\Delta_{s,t}}~\Phi_{s,v}(\mu_1)(d(v,z)) ~ {\mathbb{J}}^{\,\mu_0}_{s,u,v,t}(h)(x,y,z)\\ \\ \hskip7cm \displaystyle+\int_{\Delta_{s,t}}\Phi_{s,v}(\mu_0)(d(v,z)) ~ {\mathbb{J}}^{\,\mu_0}_{s,v,u,t}(h)(x,z,y) \end{array}$$ Using the estimates (\[estimate-II\]) and (\[estimate-JJ\]), for any $(u,y)\in \Delta^n_{s,t}$ we check that $$\label{estimate-KK} \begin{array}{l} \displaystyle \sup_{1\leq k\leq n} \Vert \nabla_{y_k} {\mathbb{K}}^{\,\mu_0,\mu_1}_{s,u,t}(h)(x,y)\Vert\\ \\ \leq c_{1}~\Vert b^{[2]}\Vert_2^n~e^{-\lambda_1(t-s)}~\left[e^{-\lambda_1(t-s)}+\left(e^{\Vert b^{[2]}\Vert_2(t-s)}-1\right) e^{-\lambda_1(t-s)}\right]\\ \\ \leq c_{2}~\Vert b^{[2]}\Vert_2^n~e^{-\lambda_1(t-s)}~e^{- \lambda_{1,2}(t-s)} \end{array}$$ Using the decomposition (\[decomp-Phi-k\]) we also check that $$\begin{array}{l} \displaystyle\left[\left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}-\left({ {\cal D }}_{\mu_0}\phi_{s,t}\right)^{\otimes 2}\right]( h)(x) =\int_{s}^t~\left[\phi_{s,v}(\mu_1)-\phi_{s,v}(\mu_0)\right](dz)~ \overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x,z) ~dv \end{array}$$ with the matrix valued function $$\begin{array}{l} \displaystyle \overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,x_3)\\ \\ \displaystyle={{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,x_3) + \int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~ {\mathbb{K}}^{\,\mu_0,\mu_1}_{s,u,v,t}(h)(x_1,x_2,(y,x_3)) \\ \\ \hskip4cm \displaystyle +\int_{\Delta_{v,t}}~\Phi_{v,u}\left(\phi_{s,v}(\mu_0)\right)(d(u,y)) ~ {\mathbb{K}}^{\,\mu_0,\mu_1}_{s,v,u,t}(h)(x_1,x_2,x_3,y) \\ \\ \hskip4cm \displaystyle+\int_{\Delta_{s,v}\times \Delta_{v,t}}~\Upsilon^{\mu_1,\mu_0}_{s,t}((v,z),d((u,y),(\overline{u},\overline{y})))~ {\mathbb{K}}^{\,\mu_0,\mu_1}_{s,u,v,\overline{u},t}(h)(x_1,x_2,(y,x_3,\overline{y})) \end{array}$$ Using (\[estimate-KK\]) we find the uniform estimates $$\label{estimate-over-KK} \begin{array}{l} \Vert \nabla_{x_3}\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,x_3)\Vert\\ \\ \displaystyle \leq c_{1}~\left[ e^{-2 \lambda_{1,2}(t-s)} +\left(e^{\Vert b^{[2]}\Vert_2 (t-s)}-1\right) ~e^{-\lambda_1(t-s)}~e^{- \lambda_{1,2}(t-s)} \right]\leq c_{2}~e^{-2 \lambda_{1,2}(t-s)} \end{array}$$ On the other hand, using (\[def-D\]) and (\[a-1\]) we have $$\left[\phi_{s,t}(\mu_1)- \phi_{s,t}(\mu_0)\right](f) = (\mu_1-\mu_0)P^{\mu_0}_{s,t}(f)+(\mu_1-\mu_0){ {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)$$ Thus, recalling that $${ {\cal Q }}^{\mu_1,\mu_0}_{s,t}(\nabla f)(z):=\int_{\Delta_{s,t}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(z,y)^{\prime}~{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)(y)$$ we check that $$\begin{array}{l} \displaystyle\left[\left({ {\cal D }}_{\mu_1,\mu_0}\phi_{s,t}\right)^{\otimes 2}-\left({ {\cal D }}_{\mu_0}\phi_{s,t}\right)^{\otimes 2}\right]( h)(x) =\int~(\mu_1-\mu_0)(dz)~\int_{s}^t~P^{\mu_0}_{s,v}\left(\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x,{\mbox{\LARGE .}})\right)(z) ~dv\\ \\ \hskip.3cm \displaystyle+ \int~(\mu_1-\mu_0)(dz)~\int_{s}^t~ \int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(z,y)^{\prime}~{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,v}(\nabla_{x_3}\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x,{\mbox{\LARGE .}}))(y)~dv \end{array}$$ This implies that $$\begin{array}{l} (\nabla\otimes\nabla)D^2_{\mu_0,\mu_1}\phi_{s,t}(f)(x_1,x_2)- (\nabla\otimes\nabla)D^2_{\mu_0}\phi_{s,t}(f)(x_1,x_2)\\ \\ \displaystyle= \int~(\mu_1-\mu_0)(dx_3)~\int_s^t~ {\mathbb{L}}^{\mu_1,\mu_0}_{s,u}\left( {\mathbb{S}}^{[2,1],\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)+{\mathbb{S}}^{[2,2],\phi_{s,u}(\mu_0)}_{u,t}(\nabla^2 f)\right) (x_1,x_2,x_3) ~du \end{array}$$ with the tensor integral operator $$\begin{array}{l} {\mathbb{L}}^{\mu_1,\mu_0}_{s,t}(h) (x_1,x_2,x_3) \displaystyle:=~\int_{s}^t~P^{\mu_0}_{s,v}\left(\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,{\mbox{\LARGE .}})\right)(x_3) ~dv\\ \\ \hskip.3cm \displaystyle+ ~\int_{s}^t~ \int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~b^{\mu_0}_{s,u}(x_3,y)^{\prime}~{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,v}(\nabla_{x_3}\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,{\mbox{\LARGE .}}))(y)~dv \end{array}$$ On the other hand, using (\[nabla-equiv-2\]) $$\begin{array}{l} (\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}(f)-(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_0}\phi_{s,t}(f)\\ \\ \displaystyle=\int_{[0,1]^3}~\int_s^t~{\mathbb{E}}\left(\langle \nabla_{x_3}{\mathbb{L}}^{\mu_1,\mu_0}_{s,u}\left( {\mathbb{S}}^{[2,1],\phi_{s,u}(\mu_0)}_{u,t}(\nabla f)+{\mathbb{S}}^{[2,2],\phi_{s,u}(\mu_0)}_{u,t}(\nabla^2 f)\right)({ {\cal Y }}_{\epsilon}),({ {\cal Y }}_1-{ {\cal Y }}_0)^{\otimes 3}\rangle\right)~du~d\epsilon\end{array}$$ with the interpolating path $$\epsilon=(\epsilon_1,\epsilon_2,\epsilon_3)\mapsto { {\cal Y }}_{\epsilon}:=\left(\overline{Y}^1_0+\epsilon_1(\overline{Y}^1_1-\overline{Y}^1_0), \overline{Y}^2_0+\epsilon_2(\overline{Y}^2_1-\overline{Y}^2_0),\overline{Y}^3_0+\epsilon_3(\overline{Y}^3_1-\overline{Y}^3_0)\right)$$ and $$({ {\cal Y }}_1-{ {\cal Y }}_0)^{\otimes 3}:=(\overline{Y}^1_1-\overline{Y}^1_0)\otimes (\overline{Y}^2_1-\overline{Y}^2_0)\otimes (\overline{Y}^3_1-\overline{Y}^2_0)$$ In the above display, $ (\overline{Y}^i_1,\overline{Y}^i_0)_{i=1,2,3}$ stands for independent copies of a pair of random variables $(Y_{0},Y_1)$ with distribution $(\mu_{0},\mu_1)$. Using the commutation formula (\[commutation-tensor\]) we check that $$\begin{array}{l} \nabla_{x_3} {\mathbb{L}}^{\mu_1,\mu_0}_{s,t}(h) (x_1,x_2,x_3) \displaystyle:=~\int_{s}^t~{ {\cal P }}^{\mu_0}_{s,v}\left(\nabla_{x_3}\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,{\mbox{\LARGE .}})\right)(x_3) ~dv\\ \\ \hskip.3cm \displaystyle+ ~\int_{s}^t~ \int_{\Delta_{s,v}}~\Phi_{s,u}(\mu_1)(d(u,y))~\BB^{\mu_0}_{s,u}(x_3,y)~{ {\cal P }}^{\,\phi_{s,u}(\mu_0)}_{u,v}(\nabla_{x_3}\overline{{\mathbb{K}}}^{\,\mu_0,\mu_1}_{s,v,t}(h)(x_1,x_2,{\mbox{\LARGE .}}))(y)~dv \end{array}$$ Using (\[estimate-over-KK\]) for any differentiable matrix valued function $h(x_1,x_2)$ such that $\Vert h\Vert\vee \Vert \nabla_{x_1}h\Vert\leq 1$ and for any $\epsilon\in ]0,1[$ we check that $$\begin{array}{l} \Vert \nabla_{x_3} {\mathbb{L}}^{\mu_1,\mu_0}_{s,t}(h) (x_1,x_2,x_3)\Vert\\ \\ \displaystyle \leq c_{1}~ e^{-2 \lambda_{1,2}(t-s)}~\left[\int_s^te^{-\lambda_1(v-s)}~dv+ ~\int_{s}^t~ \left(e^{\Vert b^{[2]}\Vert_2(v-s)}-1\right)e^{-\lambda_1(v-s)}~dv\right] \leq c_{2}~ e^{-2 \lambda_{1,2}(t-s)}~ \end{array}$$ On the other hand, we have $$\begin{array}{l} \nabla_{x_1}\left[{\mathbb{S}}^{[2,1],\mu}_{s,t}(\nabla f)+ \nabla_{x_1}{\mathbb{S}}^{[2,2],\mu}_{s,t}(\nabla^2 f)\right](x_1,x_2)\\ \\ \displaystyle= b^{[1,1,2]}_s(x_1,x_2)~\nabla D_{\mu}\phi_{s,t}(f)(x_1)+b^{[2,2,1]}_s(x_2,x_1)~\nabla D_{\mu}\phi_{s,t}(f)(x_2)\\ \\ +\nabla^3D_{\mu}\phi_{s,t}(f)(x_1)~ b^{[2]}_s(x_1,x_2)^{\prime}+b^{[2,2]}_s(x_2,x_1)~\nabla^2D_{\mu}\phi_{s,t}(f)(x_2)+\nabla^2D_{\mu}\phi_{s,t}(f)(x_1)\star b^{[1,2]}_s(x_1,x_2) \end{array}$$ with the $\star$-tensor product $$\begin{array}{l} \left[\nabla^2D_{\mu}\phi_{s,t}(f)(x_1)\star b^{[1,2]}_s(x_1,x_2)\right]_{k,i,j}\\ \\ \displaystyle =\sum_{1\leq l\leq d}\left[ \nabla^2D_{\mu}\phi_{s,t}(f)(x_1)_{k,l}~ b^{[1,2]}_s(x_1,x_2)^{\prime}_{l,i,j}+ b^{[1,2]}_s(x_1,x_2)^{\prime}_{k,j,l} \nabla^2D_{\mu}\phi_{s,t}(f)(x_1)^{\prime}_{l,j}\right] \end{array}$$ Using (\[def-lambda-hat-3\]) we check that $$\Vert \nabla_{x_1}\left[{\mathbb{S}}^{[2,1],\mu}_{s,t}(\nabla f)+ \nabla_{x_1}{\mathbb{S}}^{[2,2],\mu}_{s,t}(\nabla^2 f)\right]\Vert\leq c\,e^{- \lambda(t-s)}~\sup_{k=1,2,3}\Vert\nabla^k f\Vert\quad \mbox{for some $\lambda>0$}$$ We conclude that for any function $f\in { {\cal C }}^3({\mathbb{R}}^d)$ s.t. $\sup_{k=1,2,3}\Vert\nabla^k f\Vert\leq 1$ $$\vert (\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}(f)-(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_0}\phi_{s,t}(f)\vert\leq c~ e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^3\quad \mbox{for some $\lambda>0$}$$ The last assertion comes from the formula $$\frac{1}{2}~(\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_1,\mu_0}\phi_{s,t}=\frac{1}{2} (\mu_1-\mu_0)^{\otimes 2}D^2_{\mu_0}\phi_{s,t}+(\mu_1-\mu_0)^{\otimes 3}D^3_{\mu_0,\mu_1}\phi_{s,t}$$ [\ ]{} Proof of theorem \[theo-ae-taylor\] {#theo-ae-taylor-proof .unnumbered} ----------------------------------- We extend the operators $D^k_{\mu_1,\mu_0}\phi_{s,t}$ introduced in theorem \[theo-intro-4\] to tensor functions $f=(f_{i})_{i\in [n]}$ by considering the tensor function with entries $$\label{def-D-k-extension} D_{\mu_1,\mu_0}^k\phi_{s,t}(f)_i=D_{\mu_1,\mu_0}^k\phi_{s,t}(f_{i})$$ By theorem \[theo-intro-4\] we have $$\label{k-deco} \begin{array}{l} \displaystyle\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}}))\\ \\ \displaystyle=\int~(\mu_1-\mu_0)(dy)~d_{s,u}^{[1],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),y)\\ \\ \displaystyle=\int~(\mu_1-\mu_0)(dy)~d_{s,u}^{[1],\mu_0}(X^{\mu_0}_{s,u}(x),y)+\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~d_{s,u}^{[2],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),z)\\ \\ \displaystyle=\int~(\mu_1-\mu_0)(dy)~d_{s,u}^{[1],\mu_0}(X^{\mu_0}_{s,u}(x),y)+\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~d_{s,u}^{[2],\mu_0}(X^{\mu_0}_{s,u}(x),z)\\ \\ \displaystyle\hskip3cm+\int~(\mu_1-\mu_0)^{\otimes 3}(dz)~d_{s,u}^{[3],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),z) \end{array}$$ with the functions $$\begin{aligned} d_{s,t}^{[1],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),y)&:=&D_{\mu_1,\mu_0}\phi_{s,t}(b_t(X^{\mu_0}_{s,t}(x),{\mbox{\LARGE .}}))(y)\\ d_{s,t}^{[2],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),(z_1,z_2))&:=&D_{\mu_1,\mu_0}^2\phi_{s,t}(b_t(X^{\mu_0}_{s,t}(x),{\mbox{\LARGE .}}))(z_1,z_2)\\ d_{s,u}^{[3],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),(z_1,z_2,z_3))&:=&D_{\mu_1,\mu_0}^3\phi_{s,t}(b_t(X^{\mu_0}_{s,t}(x),{\mbox{\LARGE .}}))(z_1,z_2,z_3)\end{aligned}$$ We also write $d_{s,t}^{[k],\mu}$ instead of $d_{s,t}^{[k],\mu,\mu}$. Using (\[estimate-D2-nabla\]) and (\[commutation-D-etimate\]) we check that $$\Vert \nabla_y\, d_{s,t}^{[1],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),y)\Vert \leq c_1~~e^{- \lambda(t-s)}$$ as well as $$\label{R-ok-4-3} \Vert (\nabla_{z_1}\otimes\nabla_{z_2})\,d_{s,t}^{[2],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),z_1,z_2)\Vert\leq c_2~e^{-\lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ Using (\[D3-estimation\]) we also have $$\label{R-ok-3} \vert\int~(\mu_1-\mu_0)^{\otimes 3}(dz)~d_{s,t}^{[3],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),z)\vert \leq c_3~ e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^3\quad \mbox{for some $\lambda>0$}$$ On the other hand, we have the second order expansions $$\begin{array}{l} \left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_1}_{s,u}(x))^{\prime}-\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}\\ \\ \displaystyle=\int_0^1 \left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(y)-X^{\mu_0}_{s,u}(x))\right)^{\prime}~ [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~d\epsilon\\ \\ =\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]\\ \\ \displaystyle+\int_0^1 (1-\epsilon)~\left[\nabla^3 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(y)-X^{\mu_0}_{s,u}(x))\right)^{\prime}~ [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]^{\otimes 2}~d\epsilon \end{array}$$ In the same vein, we have $$\begin{array}{l} b_u(X^{\mu_1}_{s,u}(x),y)-b_u(X^{\mu_0}_{s,u}(x),y)\\ \\ \displaystyle=\int_0^1 ~b_u^{[1]}\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)),y\right)^{\prime}~ [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~d\epsilon\\ \\ \displaystyle=b_u^{[1]}(X^{\mu_0}_{s,u}(x),y)^{\prime}~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]\\ \\ \displaystyle+\int_0^1 (1-\epsilon)~b_u^{[1,1]}\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)),y\right)^{\prime}~ [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]^{\otimes 2}~d\epsilon \end{array}$$ This implies that $$\begin{array}{l} \displaystyle X^{\mu_1}_{s,t}(x)-X^{\mu_0}_{s,t}(x)\\ \\ \displaystyle=\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}}))~du+\sum_{k=2,3}R^{[k],\mu_0,\mu_1}_{s,t}(x) \end{array}$$ with the second order remainder term $$\begin{array}{l} \displaystyle R^{[2],\mu_0,\mu_1}_{s,t}(x)\\ \\ \displaystyle:= \int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}}))~du\\ \\ \displaystyle+\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~ \left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\left(b_u^{[1]}(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}})^{\prime}\right)~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~du \end{array}$$ and the third order remainder term $$\begin{array}{l} \displaystyle R^{[3],\mu_0,\mu_1}_{s,t}(x)\\ \\ \displaystyle:=\int_0^1\int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~\\ \\ \displaystyle\hskip.3cm\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\left(b_u^{[1]}\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(x)-X^{\mu_1}_{s,u}(x)),{\mbox{\LARGE .}}\right)^{\prime}~\right) [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]~d\epsilon~du\\\ \\ \displaystyle+\int_0^1 (1-\epsilon)~\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~\\ \\ \displaystyle\hskip.3cm\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right] \left(b_u^{[1,1]}\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(x)-X^{\mu_1}_{s,u}(x)),{\mbox{\LARGE .}}\right)^{\prime}~ \right)~[X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]^{\otimes 2}~du~d\epsilon\\ \\ \displaystyle+\int_0^1~(1-\epsilon)~ \int_s^t \left[\nabla^3 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]\left(X^{\mu_0}_{s,u}(x)+\epsilon(X^{\mu_1}_{s,u}(y)-X^{\mu_1}_{s,u}(x))\right)^{\prime}~\\ \\ \displaystyle\hskip3cm [X^{\mu_1}_{s,u}(x)-X^{\mu_0}_{s,u}(x)]^{\otimes 2} ~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_1}_{s,u}(x),{\mbox{\LARGE .}}))~du~~d\epsilon\\ \\ \end{array}$$ Combining (\[tau-n-estimates\]) with (\[ref-stab-W-1\]) and (\[estimate-X-mu\]) for any $k=1,2$ we check the uniform estimate $$\label{R-k-e} \Vert R^{[k],\mu_0,\mu_1}_{s,t}(x)\Vert\leq c~e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^k\quad \mbox{for some $\lambda>0$}$$ We check (\[a-e-T1\]) using (\[R-ok-4-3\]) and (\[k-deco\]). Using (\[def-lambda-hat-3\]) we also have the estimate $$\Vert \nabla_y\,D_{\mu_0}X^{\mu_0}_{s,t}(x,y)\Vert\leq c_3~e^{-\lambda(t-s)}\quad \mbox{for some $\lambda>0$}$$ Observe that $$\begin{array}{l} \displaystyle\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u^{[1]}(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}})^{\prime})=\int~(\mu_1-\mu_0)(dy)~d_{s,u}^{[1,1],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),y)\\ \\ \displaystyle=\int~(\mu_1-\mu_0)(dy)~d_{s,u}^{[1,1],\mu_0}(X^{\mu_0}_{s,u}(x),y)+\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~d_{s,u}^{[2,1],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),z) \end{array}$$ with the matrix valued functions $$\begin{aligned} d_{s,t}^{[1,1],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),y)&:=&D_{\mu_1,\mu_0}\phi_{s,t}(b_t^{[1]}(X^{\mu_0}_{s,t}(x),{\mbox{\LARGE .}})^{\prime})(y)\\ d_{s,t}^{[2,1],\mu_1,\mu_0}(X^{\mu_0}_{s,t}(x),z_1,z_2)&:=&D_{\mu_1,\mu_0}^2\phi_{s,t}(b_t^{[1]}(X^{\mu_0}_{s,t}(x),{\mbox{\LARGE .}})^{\prime})(z_1,z_2)\end{aligned}$$ We also write $d_{s,t}^{[1,1],\mu}$ instead of $d_{s,t}^{[1,1],\mu,\mu}$. Observe that $$\begin{array}{l} \displaystyle R^{[2],\mu_0,\mu_1}_{s,t}(x)\\ \\ \displaystyle=\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 2}(dz)~\int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~D^{[2,1]}_{\mu_0}X^{\mu_0}_{s,u}(x,z)~du\\ \\ \displaystyle+\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 2}(dy)~\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~D^{[1,1]}_{\mu_0}X^{\mu_0}_{s,u}(x,z)~du+R^{[3,2],\mu_0,\mu_1}_{s,t}(x) \end{array}$$ with $$\begin{array}{l} \displaystyle R^{[3,2],\mu_0,\mu_1}_{s,t}(x)\\ \\ \displaystyle=\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 3}(dy)~\int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~ D_{\mu_0}X^{\mu_0}_{s,u}(x,y_1)~d_{s,u}^{[2],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),(y_2,y_3))~du\\ \\ \displaystyle+\frac{1}{2}~\int~(\mu_1-\mu_0)^{\otimes 3}(dy)~\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~ d_{s,u}^{[2,1],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),(y_2,y_3))~ D_{\mu_0}X^{\mu_0}_{s,u}(x,y_1)~du\\ \\ \displaystyle+ \int_s^t\left[\nabla^2 X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~ { {\cal R }}^{[2],\mu_0,\mu_1}_{s,u}(x)~\left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right](b_u(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}}))~du\\ \\ \displaystyle+\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~ \left[\phi_{s,u}(\mu_1)-\phi_{s,u}(\mu_0)\right]\left(b_u^{[1]}(X^{\mu_0}_{s,u}(x),{\mbox{\LARGE .}})^{\prime}\right)~{ {\cal R }}^{[2],\mu_0,\mu_1}_{s,u}(x)~du \end{array}$$ Observe that $$\label{R-k-32-e} \Vert R^{[3,2],\mu_0,\mu_1}_{s,t}(x)\Vert\leq c~e^{-\lambda(t-s)}~\WW_2(\mu_0,\mu_1)^3 \quad \mbox{for some $\lambda>0$}$$ This yields the second order decompositionn (\[ae-taylor\]) with the remainder term $$\begin{array}{l} \displaystyle { {\cal R }}^{\mu_1,\mu_0}_{s,t}(x):=R^{[3],\mu_0,\mu_1}_{s,t}(x)+R^{[3,2],\mu_0,\mu_1}_{s,t}(x)\\ \\ \displaystyle \hskip3cm+\int~(\mu_1-\mu_0)^{\otimes 3}(dz)~\int_s^t\left[\nabla X^{\phi_{s,u}(\mu_0)}_{u,t}\right]({X}^{\mu_0}_{s,u}(x))^{\prime}~d_{s,u}^{[3],\mu_1,\mu_0}(X^{\mu_0}_{s,u}(x),z)~du \end{array}$$ The end of the proof of is now a consequence of the estimates (\[R-ok-3\]), (\[R-k-e\]) and (\[R-k-32-e\]). The proof of the theorem is completed. [\ ]{} [99]{} V.M. Alekseev, An estimate for the perturbations of the solutions of ordinary differential equations. II, Vestnik Moskov. Univ. Ser. I Mat. Mech, vol. 3 , pp. 3–10 (1961). L. Ambrosio, N. Gigli, Construction of parallel transport in the Wasserstein space, Methods and Applications of Analysis, no. 15 , vol.1, pp. 1–30 (2008). L. Ambrosio, N. Gigli, and G. Savare, Gradient flows in metric spaces and in spaces of probability measures, Birkäuser, 2005. M. Arnaudon, P. Del Moral. A duality formula and a particle Gibbs sampler for continuous time Feynman-Kac measures on path spaces. [ArXiv:1805.05044](https://arxiv.org/abs/1805.05044) (2018). M. Arnaudon, P. Del Moral. A variational approach to nonlinear and interacting diffusions. [ArXiv:1812.04269 (2018)](https://arxiv.org/abs/1812.04269). Stochastic Analysis and Applications DOI: 10.1080/07362994.2019.1609985 (2019). M. Arnaudon, H. Plank, A. Thalmaier. A Bismut type formula for the Hessian of heat semigroups. C. R. Math. Acad. Sci. Paris, vol. 336, no. 8, pp. 661–666 (2003). D. Benedetto, E. Caglioti, M. Pulvirenti. A kinetic equation for granular media. RAIRO Modèl. Math. Anal. Numér. vol. 31, no. 5, pp. 615–641 (1997). D. Benedetto, E. Caglioti, E., Carrillo, M. Pulvirenti. A non-Maxwellian steady distribution for one-dimensional granular media. J. Statist. Phys.vol. 91, pp. 979–990 (1998). A. Bensoussan, J. Frehse, P. Yam. The Master Equation in Mean Field Theory. http://arxiv.org/abs/1404.4150 (2014). A.N. Bishop and P. Del Moral. On the Stability of Kalman-Bucy Diffusion Processes. SIAM Journal on Control and Optimization. vol. 55, no. 6. pp 4015–4047 (2017); arxiv e-print [arXiv:1610.04686](https://arxiv.org/abs/1610.04686) updated. A.N. Bishop and P. Del Moral Stability Properties of Systems of Linear Stochastic Differential Equations with Random Coefficients. SIAM Journal on Control and Optimization. vol. 57, no. 2, pp. 1023–1042 (2019). J.M. Bismut. Large deviations and the Malliavin calculus. Birkhauser Prog. Math. 45 (1984). R. Buckdahn, J. Li, S. Peng, C. Rainer. Mean-field stochastic differential equations and associated PDEs. Ann. Probab., vol. 45 , no. 2, pp. 824–878 (2017). P. Cardaliaguet, F. Delarue, J.M. Lasry, P.L. Lions. (2015). The master equation and the convergence problem in mean field games. ArXiv preprint arXiv:1509.02505. Princeton University Press (2019). R. Carmona, F. Delarue. Probabilistic Theory of Mean Field Games with Applications I-II. Springer Nature (2018). R. Carmona and F. Delarue. Mean-field forward-backward stochastic differential equations. Electron. Commun. Probab., vol. 18, no. 68, pp. 1–15 (2013). R. Carmona and F. Delarue. Probabilistic analysis of mean field games. SIAM J. Control Optim. vol. 51, pp. 2705–2734. (2013). P. Cattiaux, A. Guillin, and F. Malrieu. Probabilistic approach for granular media equations in the non uniformly convex case. Prob. Theor. Rel. Fields, vol. 140, no. 1-2, pp. 19–40 (2008). J.F. Chassagneux, D. Crisan, F. Delarue. A probabilistic approach to classical solutions of the master equation for large population equilibria. ArXiv preprint arXiv:1411.3009 (2014). P.E. Chaudru de Raynal, N. Frikha. Well-posedness for some non-linear diffusion processes and related PDE on the Wasserstein space. ArXiv preprint arXiv:1811.06904 (2018). G. Constantine, T. Savits. A multivariate Faà di Bruno formula with applications. Transactions of the American Mathematical Society, vol. 348, no. 2, pp. 503–520 (1996). W. A. Coppel, Stability and asymptotic behavior of differential equations, D. C. Heath, Boston, Mass. (1965). D. Crisan, E. McMurray. Smoothing properties of Mckean-Vlasov sdes. Probability Theory and Related Fields vol. 171, no. 1-2, pp. 97–148 (2018). D.A. Dawson. Critical dynamics and fluctuations for a mean-field model of cooperative behavior. Journal of Statistical Physics, vol. 31, no. 1, pp. 29–85 (1983). P. [Del Moral]{}. [Feynman-[K]{}ac formulae](http://www.math.u-bordeaux1.fr/~delmoral/gips.html). Genealogical and interacting particle systems with applications. Probability and its Applications (New York). (573p.) Springer-Verlag, New York (2004). P. [Del Moral]{}. [Mean field simulation for Monte Carlo integration.](http://www.math.u-bordeaux1.fr/~pdelmora/Intro+Refs-Mean-Field-Simulation.pdf) [Chapman & Hall. Monographs on Statistics & Applied Probability](http://www.crcpress.com/product/isbn/9781466504059) (2013). P. Del Moral, A. Guionnet. On the stability of measure valued processes with applications to filtering. Comptes Rendus de l’Académie des Sciences-Series I-Mathematics, vol. 329, no. 5, pp. 429–434 (1999). P. Del Moral and A. Guionnet. On the stability of interacting processes with applications to filtering and genetic algorithms. Ann. Inst. Henri Poincaré, vol. 37, no. 2, pp. 155–194 (2001). P Del Moral, P. Jacob, A. Lee, L. Murray, G.W. Peters. Feynman-Kac particle integration with geometric interacting jumps. Stochastic Analysis and Applications. vol. 31, no. 5, pp. 830–871 (2013). P. Del Moral and L. Miclo. Branching and interacting particle systems approximations of [F]{}eynman-[K]{}ac formulae with applications to non-linear filtering. In [Séminaire de [P]{}robabilités, [XXXIV]{}]{}, volume 1729, [Lecture Notes in Math.]{}, pages 1–145. Springer, Berlin (2000). P. Del Moral and L. Miclo. Particle approximations of Lyapunov exponents connected to Schrödinger operators and Feynman-Kac semigroups. [ESAIM: Probability and Statistics](https://www.esaim-ps.org/articles/ps/pdf/2003/01/ps104.pdf), vol. 7, 171–208 (2003). P. Del Moral, F. Patras, S. Rubenthaler. Convergence of U-statistics for interacting particle systems. Journal of Theoretical Probability vol. 24, no. .4, p. 1002 (2011). P. Del Moral, F. Patras, S. Rubenthaler. Coalescent tree based functional representations for some Feynman-Kac particle models. Annals of Applied Probability, Vol. 19, No. 2, pages 1–50 (2009) P. Del Moral, E. Rio. Concentration inequalities for mean field particle models. [HAL INRIA RR-6901 (2009)](https://hal.inria.fr/file/index/docid/378753/filename/RR-6901.pdf). The Annals of Applied Probability. vol. 21, no. 3, pp. 1017-1052 (2011). P. Del Moral, S.S. Singh. A forward-backward stochastic analysis of diffusion flows. [arXiv preprint arXiv:1906.09145](https://arxiv.org/abs/1906.09145) (2019). P. Del Moral and J. Tugaut. On the stability and the uniform propagation of chaos properties of ensemble Kalman-Bucy filters. Annals of Applied Probability. vol. 28, no. 2. pp 790–850 (2018). P. Del Moral, S. Hu, L. Wu. Moderate Deviations for Interacting Processes. Statistica Sinica. [HAL-00687827 (2012)](https://hal.inria.fr/hal-00687827) vol. 25, no. 3, pp. 921–951 (2015) R.M. Dudley, The speed of mean Glivenko-Cantelli convergence, The Annals of Mathematical Statistics vol. 40, no. 1, pp. 40–50 (1969). K.D. Elworthy, X.M. Li. Formulae for the Derivative of Heat Eemigroups. Journal of Functional Analysis 125, pp. 252–286 (1994). N. Fournier, A. Guillin. On the rate of convergence in Wasserstein distance of the empirical measure Probab. Theory Relat. Fields, vol. 162, pp. 707-738 (2015). C. Graham. McKean-Vlasov, Ito-Skorohod equations and nonlinear diffusions with discrete jumps. Stochastic Processes and their Applications. vol. 40, pp. 69-82 (1992). W. Gröbner, Die Lie-Reihen und ihre Anwendungen, VEB Deutscher Verlag der Wissenschaften, Berlin, (1967). O. Guéant, J.M. Lasry, and P.L. Lions. Mean field games and applications. In R. Carmona et al., editor, Paris Princeton Lectures in Mathematical Finance IV, volume 2003 of Lecture Notes in Mathematics. Springer Verlag (2010). E. Hairer, S. Norsett, and G. Wanner. Solving Ordinary Differential Equations I. Nonstiff Problems. Springer-Verlag, Berlin, second edition (1993). T.E. Harris, H. Kahn. Estimation of particle transmission by random sampling. Natl.Bur. Stand., Appli. Math.Ser., vol.12, pp. pp. 27-30 (1951) M. Huang, P.E. Caines, and R.P. Malhamé, Large population stochastic dynamic games: closed loop McKean-Vlasov systems and the Nash certainty equivalence principle. Communications in Information and Systems, vol. 6, pp. 221–252 (2006). X. Huang, M. Röckner, F.Y. Wang. Nonlinear Fokker-Planck equations for probability measures on path space and path-distribution dependent SDEs. arXiv preprint arXiv:1709.00556 (2017). A. Hudde, M. Hutzenthaler, A. Jentzen, S. Mazzonetto. On the It\^ o-Alekseev-Gr" obner formula for stochastic differential equations. ArXiv preprint arXiv:1812.09857 (2018). R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck equation. SIAM Journal on Mathematical Analysis, vol.29, no. 1, pp. 1–17 (1998). M. Kac. Foundations of kinetic theory. In Proceedings of the 3rd Berkeley Symposium on Mathematical Statistics and Probability, volume 3, 171-197 (1956). M. Kac. Probability and Related Topics in the Physical Sciences. Interscience Publishers, New York (1958). V.N. Kolokoltsov. Nonlinear Markov processes and kinetic equations (Vol. 182). Cambridge University Press (2010). V.N. Kolokoltsov, M. Troeva, W. Yang. On the rate of convergence for the mean-field approximation of controlled diffusions with large number of players. Dyn. Games Appl. vol. 4, no. 2, pp. 208–230 (2014). V. N. Kolokoltsov, M. Troeva, W. Yang. Mean field games based on the stable like processes (In Russian). Mathematical Theory of Games and its Applications 5 (2013), pp. 33–65. Engl. transl. in Journal Automation and Remote Control, vol. 77, no. 11, pp. 2044–2064 (2016). P. Lancaster and M. Tismenetsky, The Theory of Matrices, 2nd ed., Academic Press, New York (1985). J. Lei. Convergence and concentration of empirical measures under Wasserstein distance in unbounded functional spaces. arXiv preprint arXiv:1804.10556 (2018). X.M. Li. Doubly Damped Stochastic Parallel Translations and Hessian Formulas. International Conference on Stochastic Partial Differential Equations and Related Fields. Springer, Cham (2016). H. P. McKean. A class of Markov processes associated with nonlinear parabolic equations. Proc. Nat. Acad. Sci. U.S.A., vol. 56, pp. 1907–1911 (1966). H. P. McKean. Propagation of chaos for a class of non-linear parabolic equations. In Stochastic Differential Equations, Lecture Series in Differential Equations, Session 7, Catholic Univ. (1967), pp. 41–57. Air Force Office Sci. Res., Arlington, Va., (1967). Y. Nishiyama. Some central limit theorems for ${ {\cal L }}^{\infty}$-valued semimartingales and their applications. Probability Theory and Related Fields, vol. 108, pp. 459-494 (1997). F. Otto. The geometry of dissipative evolution equations: the porous medium equation. Comm. Partial Differential Equations, vol. 26, pp. 101–174 (2001). F. Otto, V. Villani. Generalization of an inequality by Talagrand, and links with the logarithmic Sobolev inequality. J. Funct. Anal. vol. 173, pp. 361–400 (2000). G.C. Papanicolaou, D. Stroock, S.R.S. Varadhan. Martingale approach to some limit theorems. Statistical Mechanics, Dynamical Systems. Duke Turbulence Conference. D. Ruelle editor; Duke univ. series, vol.3 (1977) G. Peano. Intégration par séries des équations différentielles linéaires. Mathematische Annalen. vol. 32, no. 3. pp. 450–456 (1888). Alain-Sol Sznitman. Topics in propagation of chaos. In École d’Été de Probabilités de Saint-Flour XIX-1989, volume 1464 of Lecture Notes in Math., pages 165–251. Springer, Berlin (1991). J. Thompson. Derivatives of Feynman-Kac semigroups. Journal of Theoretical Probability, vol. 32, no. 2, pp. 950–973 (2019). G. Toscani. One-dimensional kinetic models of granular flows. RAIRO Modèl. Math. Anal. Numér. no. 34, no. 6, pp. 1277–1291 (2000). C. Villani. A survey of mathematical topics in the collisional kinetic theory of gases. Handbook of mathematical fluid dynamics, vol. 1, no 71–305, p. 3–8 (2002). F.Y. Wang. Diffusions and PDEs on Wasserstein space [ArXiv:1903.02148](https://arxiv.org/pdf/1903.02148.pdf) (2019).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We examine the behaviour of a charged particle in a two dimensional confining potential, in the presence of a magnetic field. The confinement serves to remove the otherwise infinite degeneracy, but additional ingredients are required to produce sensible results. We treat both circular and square geometries, and in the latter we explicitly demonstrate the gauge invariance of the energy levels and wave function amplitudes. Both bulk states and edge states are examined, and in the latter case, with sufficiently high quantum numbers we achieve significant differences in the square and circular geometries. Results are achieved using straightforward matrix mechanics, in a manner that is accessible to novices in the field.' author: - Asadullah Bhuiyan and Frank Marsiglio bibliography: - 'References.bib' title: 'Landau Levels and the Issue of Gauge Invariance in Confined Spaces' --- [Introduction]{} ================ The motion of a charged particle in the presence of a magnetic field is a subject of considerable interest in many areas of physics. In particular, in condensed matter the presence of the magnetic field alters the particle’s behaviour in a profound way. Ignoring the spin degree of freedom, it is known from classical physics that a vector potential is required to include the effect of a magnetic field, and more than one choice of vector potential is possible. Each of these choices constitutes a particular gauge. For a uniform magnetic field, say in the $z$ direction, common choices in Cartesian coordinates are the Landau Gauge, $\vec{\bf A}_1= xB \hat{\bf y}$ (or $\vec{\bf A}_2 = -yB \hat{\bf x}$), or the so-called Symmetric gauge, $\vec{\bf A}_S = -yB/2 \hat{\bf x} + xB/2 \hat{\bf y}$; all of these choices yield the same field $\vec{\bf B} \equiv \vec{\bf \nabla} \times \vec{\bf A} = B {\hat{\bf z}}$. Naturally all physical quantities obtained in a calculation should be independent of the gauge choice. Typically a gauge choice is made to take advantage of some symmetry in the problem, so that one can proceed analytically. Just as often the physical system is altered to take advantage of the symmetry afforded by the gauge choice. A good example is the seminal paper by Laughlin, [@laughlin1981quantized] where he wanted to address the quantization of the Hall conductivity in a two-dimensional metal. To this end he chose a ribbon bent into a loop (i.e. open boundary conditions in one direction and periodic boundary conditions in the other). The magnetic field is perpendicular to the plane of the metal, which implies the physically questionable notion of a field either emanating from or converging to an axis at the centre of the loop. That this configuration did not coincide precisely with the experiment was immaterial — the concept of quantization was properly established and what mattered was that the Landau gauge allowed Laughlin to proceed analytically with the argument. Similarly, Halperin [@halperin1982quantized] followed up with a thin film with annular geometry, and adopted the gauge choice $\vec{\bf A} = [B_0r/2 + \Phi/(2 \pi r)] \hat{\bf \theta}$ to produce a uniform field $B_0$ in the $z$-direction along with a central flux of magnitude $\Phi$. Other, “physical” geometries, suitable to so-called “quantum dots,” were adopted thereafter — see for example, Refs. \[\] and \[\]. In this paper we will review and expand upon some of these calculations, representing charged particles in a confined geometry subjected to a constant magnetic field. A very good reference for the quantum dot work is Chakraborty’s book, [Quantum Dots]{},[@chakraborty99] which also contains a number of very useful reprinted articles. As mentioned above, we will focus on the orbital motion of the charged particle (hereafter these will be electrons) and neglect the spin degree of freedom. Spin will cause a breaking of degeneracy in the presence of a magnetic field, owing to the direct coupling with field through the so-called Zeeman term. We begin in the next section with a textbook review of the free particle, from which the idea of Landau levels first arise. In particular we will explore both the Landau and the Symmetric gauges, and seemingly derive very different results for the wave functions. We further note that confinement within a parabolic trap can also be treated analytically, following Rontani.[@rontani99] Next we will review and expand upon Lent’s treatment of the circular quantum dot,[@lent91] and finally, we will show numerical results for the square quantum dot in both aforementioned gauges. This will hopefully make clear the role of degeneracy, and we explicitly illustrate the equivalence of the results in the two gauge formulations by removing the degeneracy. We will present the results of calculations for simple properties only, such as probability density and current density. [Brief Review of Landau Quantization in Free Space]{} ===================================================== In the presence of a magnetic field $\mathbf{B}$, the canonical momentum $\mathbf{p}$ is shifted by the magnetic vector potential $\mathbf{A}$ to give us a Hamiltonian of the form: $$H=\frac{\boldsymbol{\pi}^2}{2m}=\frac{(\mathbf{p}-q\mathbf{A})^2}{2m} \label{ham_original}$$ where $q$ is the charge and $m$ is the mass of the particle of interest, and $\boldsymbol{\pi}=\mathbf{p}-q\mathbf{A}$ is the kinetic momentum operator. Hereafter we adopt $q = -e$ for the electron, where $e>0$ is the magnitude of the charge of the electron. It is important to note that in the presence of a magnetic field, $\boldsymbol{\pi}$ represents the true momentum of the particle rather than $\mathbf{p}$. In this problem, we consider an electron free to move in a 2-dimensional system in a uniform magnetic field, $\mathbf{B}=B\boldsymbol{\hat{z}}$. For such a magnetic field, there are three common gauge choices, as already mentioned in the introduction. The two Landau gauges are handled very similarly mathematically, and so we will focus only on $\mathbf{A}_1$. However, $\mathbf{A}_1$ and $\mathbf{A}_S$ are handled very differently mathematically and we now discuss each in turn. A number of resources are available that provide considerable detail for this problem. These include textbooks like Refs. \[\], \[\], \[\] and \[\] and online lecture notes by Tong[@qheTong] and Murayama.[@Murayama] [The Symmetric Gauge]{} ----------------------- Substituting $\mathbf{A}_S$ into the Hamiltonian (\[ham\_original\]) leads to $$H=\frac{\mathbf{p}^2}{2m}-i\frac{\hbar\omega_c}{2}(x\partial_y-y\partial_x)+\frac{1}{8}m\omega_c^2(x^2+y^2), \label{ham_symm1}$$ where $\partial_x \equiv \partial/\partial x$, and similarly for $y$ and later for radial coordinates as well, We have introduced the classical cyclotron frequency $\omega_c \equiv {eB}/{m}$. Converting to polar coordinates, we get $$H=-\frac{\hbar^2}{2m}(\partial_r^2+\frac{1}{r}\partial_r)+\frac{L_z^2}{2mr^2}+\frac{\omega_c}{2}L_z+\frac{1}{8}m\omega_c^2r^2, \label{ham_symm2}$$ where $L_z= -i\hbar (x\partial_y - y \partial_x) = -i\hbar \partial_{\phi}$ is the standard operator for the $z$ component of the angular momentum. Note that this operator is gauge-dependent and its eigenvalues should not be considered to represent the real, physical angular momentum of the electrons in this system. To consider the true angular momentum of the system, we must take into account the effect of the magnetic field. Just like the canonical momentum is shifted to the kinetic momentum in the presence of a magnetic field, we can also define a kinetic angular momentum $\mathbf{\Lambda}$, $$\mathbf{L}=\mathbf{r}\cross\mathbf{p}\rightarrow\mathbf{\Lambda}=\mathbf{r}\cross\boldsymbol{\pi}=\mathbf{r}\cross(\mathbf{p}+e\mathbf{A})=\mathbf{L}+\mathbf{r}\cross e\mathbf{A}. \label{kin_ang_mom}$$ Since we are working in 2 dimensions, the kinetic angular momentum is perpendicular to the plane and is given by $$\mathbf{\Lambda}(\mathbf{r})=\left(L_{z}+\frac{1}{2}m\omega_cr^2\right)\boldsymbol{\hat{z}}=\Lambda_z(r)\boldsymbol{\hat{z}}, \label{kin_ang_mom2}$$ where $\Lambda_z(r)$ is the magnitude of the kinetic angular momentum in the $\boldsymbol{\hat{z}}$ direction. We can also rewrite Eq. (\[ham\_symm2\]) in terms of $\Lambda_z(r)$, $$H=-\frac{\hbar^2}{2m}(\partial_r^2+\frac{1}{r}\partial_r)+\frac{\Lambda_z^2(r)}{2mr^2}. \label{ham_symm3}$$ Analogously, this new term can be interpreted as a kinetic centrifugal potential $V_{l}(r)$. To calculate the eigenstates and eigenvalues of Eq. (\[ham\_symm2\]), we first note that since $[H,L_z]=0$, we can write the eigenstates $\Psi(r,\phi)$ as $$\Psi(r,\phi)=e^{i \ell \phi}\psi(r), \label{separable}$$ where $\hbar \ell$ is the eigenvalue of the $L_z$ operator. Using periodic boundary conditions, $\Psi(r,\phi+2\pi)=\Psi(r,\phi)$, requires $\ell$ to be an integer, and the eigenvalues of $L_z$ are thus quantized, with $\ell$ the ‘azimuthal’ or ‘angular momentum’ quantum number. This results in the radial ordinary differential equation, $$\left[-\frac{\hbar^2}{2m}\left(\frac{d^2}{dr^2}+\frac{1}{r}\frac{d}{dr}\right)+\frac{\hbar^2 \ell^2}{2mr^2}+\frac{1}{2}m\left(\frac{\omega_c}{2}\right)^2r^2 +\frac{\hbar \omega_c\ell}{2} \right]\psi(r) = E\psi(r). \label{ham_rad}$$ But Eq. (\[ham\_rad\]) is clearly just the 2D radial quantum harmonic oscillator Hamiltonian with frequency ${\omega_c}/{2}$ and with an additional energy shift by the constant term ${\hbar \omega_c\ell}/{2}$; the normalized 2D radial eigenstates are given by:[@jain2007composite] $$\Psi_{n_r,\ell}(r,\phi)=\frac{1}{\ell_B}\sqrt{\frac{n_r!}{\left(n_r+\|\ell\|\right)!}}\left(\frac{r}{\sqrt{2}\ell_B}\right)^{\|\ell\|}e^{-\frac{r^2}{4\ell_B^2}}L^{\|\ell\|}_{n_r}\left[\frac{r^2}{2\ell_B^2}\right]\frac{e^{i\ell\phi}}{\sqrt{2\pi}} \label{psi_rad}$$ where $L^{\|\ell\|}_{n_r}[x]$ are the associated Laguerre polynomials and $\ell_B \equiv \sqrt{{\hbar}/{eB}}$ is the magnetic length. The magnetic length plays an important role in what follows, and represents the characteristic length for the wave function of a charged particle in a magnetic field. For $B \approx 1$ Tesla, $\ell_B \approx 26$ nm. The quantum number $n_r$ is the ‘radial quantum number’ and simply counts the number of nodes that the radial part of the wave function has ($n_r$ is a non-negative integer). The corresponding probability density takes the appearance of concentric rings that are radially localized increasingly outwards from $r=0$ with increasing $\|\ell\|$. The wave function has $n_r$ nodes and the probability density $\|\Psi\|^2$ depends on $\|\ell\|$ and not the sign of $\ell$. The corresponding eigenenergies are given by: $$E_{n_r,\ell} = \frac{\hbar\omega_c}{2}(2n_r+\|\ell\|+1) + \frac{\hbar \omega_c\ell}{2} =\hbar\omega_c\left(n_r+\frac{\|\ell\|+\ell}{2}+\frac{1}{2}\right),$$ so for all $\ell \leq 0 \ $, $E_{n_r,\ell} = \hbar\omega_c\left(n_r+\frac{1}{2}\right)$, and the energy spectrum is infinitely degenerate. Sometimes a new quantum number $n \equiv n_r+{(\|l\|+l)}/{2}$ is defined, and then the energies are simply $$E_{n}=\hbar\omega_c\left(n+\frac{1}{2}\right). \label{energy_landau_level}$$ Clearly we see that further degeneracies are possible through the use of $\ell > 0$ for $\ell \leq n$. This is shown in the table below, where all energies have an infinite degeneracy due to the negative $\ell$ states that carry on indefinitely to the right for each energy. Additionally, however, as $n$ increases, positive $\ell$ states are also degenerate in growing number as indicated. $$\begin{aligned} {\rm Energy} (\hbar \omega_c/2) \phantom{aaaaaaaa} n \phantom{aaaaaaaaaaaaaaaaaaaaa} (n_r,\ell) \ \ \phantom{aaaaaaaaaaaaaa} \cr \vdots \phantom{aaaaaaaaaaaaaaaa} \vdots \phantom{aaaaaaaaaaaaa} \phantom{(0,2) \ \ (1,1) \ \ } \vdots \phantom{\ \ (2,-1) \ \ (2,-2)} \dots \cr 5 \phantom{aaaaaaaaaaaaaaaa} 2 \phantom{aaaaaaaaaa} (0,2) \ \ (1,1) \ \ (2,0) \ \ (2,-1) \ \ (2,-2) \dots \cr 3 \phantom{aaaaaaaaaaaaaaaa} 1 \phantom{aaaaaaaaaa} \phantom{(0,2)} \ \ (0,1) \ \ (1,0) \ \ (1,-1) \ \ (1,-2) \dots \cr 1 \phantom{aaaaaaaaaaaaaaaa} 0 \phantom{aaaaaaaaaa} \phantom{(0,2) \ \ (1,1)} \ \ (0,0) \ \ (0,-1) \ \ (0,-2) \dots \cr \end{aligned}$$ [Degeneracy of the Lowest Landau Level (LLL) Wave Function]{} ------------------------------------------------------------- To better understand the wave functions and issues associated with them we focus on the lowest Landau level (LLL). For the LLL, $n_r = 0$ and $\ell \leq0$. Then $E_0 = \hbar \omega_c/2$ and the associated Laguerre polynomial is identically unity. So an infinite set of degenerate LLL wave functions is given simply by $$\Psi_{LLL}(r,\phi)=\frac{e^{i\ell\phi}}{\sqrt{2\pi}}\left(\frac{r}{\sqrt{2}l_B}\right)^{\|\ell\|}\frac{e^{-\frac{r^2}{4l_B^2}}}{l_B\sqrt{\|\ell\|!}} \ \ \ \ \ \ell \le 0. \label{lll_wavefunction}$$ The corresponding probability density $\|\Psi_{LLL}\|^2$ is a radial Gaussian function centred around $r=0$ but with a maximum at $r_{\ell}=l_B\sqrt{2\|\ell\|}$ with a spread about this maximum of $\sim\mathcal{O}(l_B)$. As $\|\ell\|$ increases, the location of the maximum of $\|\Psi_{LLL}\|^2$, i.e. $r_{\ell}$, moves outwards [from zero]{}. On the other hand, increasing the magnetic field $B$ (thus decreasing $l_B$) leads to a decrease in the position of the maximum in $\|\Psi_{LLL}\|^2$. It is important to note that these results are all gauge-dependent, even given that we use the Symmetric gauge. In particular, the choice of Symmetric gauge with $A_s$ defined in the first paragraph of the Introduction represents just one possible choice (out of infinitely many) where the vector field is referred to the origin. Since for this problem there is no preferred origin the wave functions will reflect this choice and be gauge-dependent. For this reason we highlighted “[from zero]{}” in italics above. An alternative would have been to use $$\mathbf{A}_S=\frac{B}{2}[(x-x_0)\boldsymbol{\hat{y}}-(y-y_0)\boldsymbol{\hat{x}}], \label{gauge_as}$$ and then the wave functions would have been different, even though this represents the same problem, i.e. a charged particle in free space in the presence of a uniform magnetic field $B$ (see, for example, Ref. \[\] for a discussion of this point). Returning to the particular gauge choice given in Eq. (\[gauge\_as\]) with $x_0=y_0=0$, the values of $r_{\ell}$ also reveal that the amount of magnetic flux $\Phi_B$ associated with each wave function in the LLL is quantized in units of the magnetic flux quantum $\Phi_0={h}/{e}$. This is easily seen: consider the magnetic flux $\Phi_B$ between a circle of radius $r_{\ell}$ and that of (smaller) radius $r_{\ell + 1}$ (recall that $\ell \leq 0$, so that $r_{\ell} > r_{\ell +1}$ ): $$\Phi_B=B\pi (r_{\ell}^2-r_{\ell+1}^2) = 2 \pi B \ell_B^2(\|\ell\| - \|\ell+1\|) = 2 \pi B\frac{\hbar}{eB}=\Phi_0. \label{flux_quantum}$$ Thus, there is exactly a single quantum of magnetic flux between successive (negative) $\ell$ states in the LLL. There are troublesome issues here. The infinite degeneracy is difficult, but even more so is the dependency of the wave function on gauge choice. There are some gauge invariant properties in such a system, for example, the spectrum of energies. The wave function, on the other hand, is [not]{} gauge invariant, and for a change in gauge, $\vec{\bf A} \rightarrow \vec{\bf A^\prime}$, there is a corresponding expected change in wave function,[@griffiths18] $$\vec{\bf A^\prime} \equiv \vec{\bf A} + \vec{\bf \nabla}\Lambda \Longrightarrow \Psi^\prime \equiv e^{iq \Lambda/\hbar} \Psi, \label{psi_transform}$$ where $\Lambda$ is an arbitrary real function of position and time, and $q$ is the charge of the particle. So the probability density, $\|\Psi\|^2$ is expected to be gauge invariant. And yet, even within the symmetric gauge, Eq. (\[gauge\_as\]), different choices of $(x_0,y_0)$ seemingly result in wave functions that differ by far more than a phase (and therefore probability densities that [do differ]{} from one another). The reason for this problem is that a degeneracy exists in the solution. An explicit calculation of the wave function transformation corresponding to that in Eq. (\[psi\_transform\]) between the Landau gauge (to be reviewed below) and the Symmetric gauge has been given recently in Ref. \[\], and this problem was first highlighted in Ref. \[\], where the author presented a more general gauge condition to replace the simple wave function relation in Eq. (\[psi\_transform\]) when degeneracy is present. We will return to this issue later. For the physical samples considered in this paper, the system is finite, and the electron is confined to some extent, so there must be some finite number of $\ell$’s that constitute the degeneracy number for a given Landau level (since the wave functions grow outward with increasing $\|\ell \|$). The standard argument (already provided above) is that the expected degeneracy $G$ of the $LLL$, given a sample of area $A$, is the ratio of the total magnetic flux $\Phi$ to the flux quantum, $$G=\frac{\Phi}{\Phi_0}=\frac{BA}{h/e}. \label{degeneracy}$$ As demonstrated later, this is not quite true, as sample edges confine the electron and result in higher energies for states near the edges. Also note that the difference between centres of two successive $\ell$ eigenstates $\Psi_{n_r,\ell}$ and $\Psi_{n_r,\ell+1}$ close to the edge will be much less than the spread $l_B$ \[see Eq. (\[lll\_wavefunction\])\] and will decrease like $\sim\sqrt{\|\ell\|}-\sqrt{\|\ell + 1 \|}$, implying that the corresponding eigenstates will increasingly overlap with one another as $\|\ell \|$ increases. [The Kinetic Centrifugal Potential]{} ------------------------------------- Returning to the free electron case (still in the Symmetric gauge) note that the Landau energies clearly depend on the sign of $\ell$, and yet the wave functions (\[psi\_rad\]) clearly depend only on their absolute value, $\|\ell\|$. The situation where two identical probability densities correspond to completely different energies is not as strange as one might initially think; for example, the difference between a stationary vs. travelling wave packet is a simple phase factor and the energies of these two configurations differ by an arbitrary amount, as the phase factor determines the average velocity of the wave packet, and yet the probability distributions are identical in the two cases. Here, two eigenstates with oppositely signed $\ell$ values ($\|\ell\|$ and $-\|\ell\|$), but with the same value of $n_r$, have energies $E_{n_r,\|\ell\|}=\left(n_r+\|\ell\|+\frac{1}{2}\right)\hbar\omega_c$ and $E_{n_r,-\|\ell\|}=\left(n_r+\frac{1}{2}\right)\hbar\omega_c$, respectively. Yet $\|\Psi_{n_r,\|\ell\|}\|^2=\|\Psi_{n_r,-\|\ell\|}\|^2$. To understand why, note that the effective potential, given by the kinetic centrifugal potential (\[ham\_symm3\]) is $$V_{\ell}(r)=\frac{\Lambda_z^2(r)}{2mr^2}=\frac{\left(\hbar \ell+\frac{1}{2}m\omega_cr^2\right)^2}{2mr^2},\\[5pt] \label{kin_cent_pot}$$ where $L_z$ in the original definition of $\Lambda_z(r)$, Eq. (\[kin\_ang\_mom2\]), is replaced by its eigenvalue $\hbar \ell$. Rewriting this as ($x \equiv r/r_{\ell}$)\ $$V_{\ell}(r)=\frac{\hbar \omega_c \|\ell\|}{4} \left( x^2 + 2\frac{\ell}{\|\ell\|} + \frac{1}{x^2} \right) \\[16pt] \label{kin_cent_pot2}$$ allows one to easily show that the minimum occurs at $x = 1$ ($r = r_\ell$) and the wave function (not surprisingly) is maximized where the effective potential is a minimum, [independent of the sign of $\ell$]{}. The value of the minimum potential is zero for all $\ell \le 0$ (independent of the magnitude of $\ell$), while it is $\hbar \omega_c \|\ell\|$ for $\ell > 0$ (grows linearly with $\ell$). In fact the difference between positive and negative $\ell$ is more general than this, and one can write that for [all]{} $r$, $$V_{\|\ell\|}(r) = V_{-\|\ell\|}(r) + \hbar\omega_c \|\ell\|, \label{pot_shift}$$ so there is a constant shift in potential values. The constant shift is consistent with attaining the same probability density for different energies. ![ $V_\ell(r)/(\hbar \omega_c)$ vs. $r/\ell_B$ for various values of $\ell$, both negative and positive. Solid lines correspond to negative $\ell$ while dashed lines correspond to positive $\ell$, and solid and dashed lines of the same colour correspond to the same magnitude, $\|\ell\|$ and hence have the same $r_\ell$ at which their minima occur. As explained in the text, for negative $\ell$ the minimum value of $V_\ell$ is always at zero while for positive $\ell$, the minimum value of $V_\ell$ increases linearly with $\ell$.[]{data-label="eff_pot"}](Kin-cent-potential.pdf) Physically, electrons circulating with positive angular momentum in the presence of a magnetic field directed in the positive $z$ direction cost energy, whereas those with negative angular momentum do not. This is consistent with the classical idea that electrons speed up or slow down depending on whether they are circulating with positive or negative angular momentum, respectively.[@griffiths99] This effective potential is illustrated in Fig. (\[eff\_pot\]). [Fock-Darwin States]{} ---------------------- Fock first addressed the problem of the eigenstates of a charged particle in a uniform magnetic field in 1928.[@fock28] He discovered a generally non-degenerate spectrum, particularly because the charged particle was confined to a parabolic potential. Somewhat later Landau[@landau30] examined a similar (simpler) problem in free space, and found a degenerate spectrum given in Eq. (\[energy\_landau\_level\]), which now bears his name. A year later, Darwin[@darwin31] independently obtained results similar to those of Fock, and figures displaying energy levels as a function of applied field are now called Fock-Darwin Spectra. The presence of a confining potential in the form of a parabolic trap requires very little additional work, so we now describe this work, also featured, for example, in Ref. \[\]. In the Symmetric gauge, with an additional confining potential of the form $m\omega_0^2r^2/2$, the Hamiltonian given in Eq. (\[ham\_symm1\]) becomes $$H=\frac{{\mathbf{p}}^2}{2m}+\frac{\omega_c}{2}L_z+\frac{1}{2}m\Omega^2r^2, \label{ham_parabolic}$$ where $\Omega \equiv \sqrt{\omega_0^2+\omega_c^2/4}$. With this parabolic confinement, the eigenenergies are $$E_{n_r,l}=2\hbar\Omega \biggl(n_r + \frac{1}{2} [ \|\ell \| + \frac{\omega_c}{2 \Omega} \ell] + \frac{1}{2} \biggr). \label{energies_parabolic}$$ Figure \[fock\_darwin\] shows a typical Fock-Darwin plot. A variation on this plot is shown in Fig. (\[fock\_darwin\_fm\]). As in Fig. \[fock\_darwin\], Eq. (\[energies\_parabolic\]) is used for the energies, but now they are normalized to the energy $\hbar \omega_c$, and are plotted as a function of the (normalized) confinement potential. ![Fock-Darwin spectra obtained from using the symmetric gauge in a parabolic trap, i.e. Eq. (\[energies\_parabolic\]). The crossing of energy levels is the defining feature of the Fock-Darwin spectra. The dashed lines correspond to the free Landau levels. Note however, that there is no condensation of levels as the field increases, in contrast to what happens with more confining traps, to be discussed later. []{data-label="fock_darwin"}](FD-spec-fig-2.pdf) ![The same Fock-Darwin spectra obtained as in the previous figure by confining the electron with a parabolic trap, in the Symmetric gauge, but now $E_{n_r,\ell}/(\hbar \omega_c)$ is plotted vs $\omega_0/\omega_c$ for some selected levels. The numbers in brackets label the $(n_r,\ell)$ quantum numbers. The curves of a given colour all emerge from the horizontal dashed line drawn for that same colour; these denote the infinitely degenerate free Landau level energies ($\omega_0 = 0$), while the labelled curves of the same colour indicate how the degeneracy is broken with increasing confinement (increasing $\omega_0$). It is clear from those levels drawn here that the confinement plays a more important role for negative values of $\ell$ and increasing values of $\| \ell \|$.[]{data-label="fock_darwin_fm"}](fig3fd-parab.pdf) Note that we could include all negative $\ell$ states, and these would rise even more than those shown as the confinement potential becomes stiffer. But at the far left of Fig. \[fock\_darwin\_fm\] ($\omega_0 \rightarrow 0$), we approach the free space limit, where the levels all become degenerate (this limit is indicated by the dashed horizontal lines). [The Landau Gauge]{} ==================== Later on we will illustrate explicitly that correct gauge-invariant properties are obtained for various gauge choices. Here we will not demonstrate this explicitly, but instead point out that differences are apparent in the free space case, and what the root causes for these differences are. We therefore proceed in this section with the Landau gauge discussed in the Introduction, ${\vec{\bf A}}_1 = xB \hat{\bf y}$. Now our Hamiltonian is $$H=\frac{1}{2m}\left[p_x^2+\pi_y^2\right]=\frac{1}{2m}\left[p_x^2+(p_y+eBx)^2\right]$$ where $\pi_y$ is the kinetic momentum operator in the $\boldsymbol{\hat{y}}$ direction. Immediately, we see that $[H,p_y]=0$, so we can separate our eigenstate $\Psi$ as $$\Psi(x,y)\propto e^{ik_yy}\psi(x),$$ where the operator $p_y$ is replaced by its eigenvalue $\hbar k_y$. Substituting this into the Schrödinger Equation leads to the one dimensional differential equation $$\left[\frac{-\hbar^2}{2m}\partial_x^2+\frac{1}{2}m\omega_c^2(x+k_yl_B^2)^2\right]\psi(x)=E\psi(x),$$ which is the usual harmonic oscillator (albeit not centred at zero) equation. Our solution is therefore $$\Psi_{n,k_y}(x,y)=e^{ik_yy}\frac{1}{\sqrt{2^{n}n! \sqrt{\pi}l_B}}e^{-\frac{(x+k_yl_B^2)^2}{2l_B^2}}H_{n}\left[\frac{x+k_yl_B^2}{l_B}\right] \label{psi_landau}$$ where $H_n(z)$ are the Hermite polynomials. This method of solution has introduced two quantum numbers, $k_y$, a real number, and $n$, a non-negative integer. The corresponding energies are as before, in Eq. (\[energy\_landau\_level\]). Note that the wave functions are one-dimensional-like and certainly don’t resemble those obtained previously. The reason this is possible is because of the degeneracy present (energy does not depend on $k_y$), which we now discuss further. [Degeneracy and the LLL Wavefunction in the Landau Gauge]{} ----------------------------------------------------------- In the Landau gauge, the $LLL$ wavefunction must have $n=0$ for any $k_y\in\mathbb{R}$, giving us: $$\Psi_{LLL}(x,y)=e^{ik_yy}\frac{e^{-\frac{(x+k_yl_B^2)^2}{2l_B^2}}}{\sqrt{\sqrt{\pi}l_B}}.$$ Usually periodic boundary conditions are imposed in the $y$-direction, resulting in the requirement that $k_y = 2\pi m_y/a$, where $a$ is the length of the sample (i.e. “ribbon”) in the $y$-direction and now $m_y$ is an integer. We can estimate the degeneracy here as well. The magnetic flux $\Phi_B$ between the centres of two successive $m_y$ eigenstates $\Psi_{n,m_y+1}$ and $\Psi_{n,m_y}$ is $$\Phi_B=Ba(x_{k+1}-x_k)=Ba\left(\frac{2\pi (m_y+1)}{a}l^2_B-\frac{2\pi m_y}{a}l^2_B\right)=B\frac{2\pi\hbar}{eB}=\frac{h}{e}=\Phi_0, \label{degeneracy_landau}$$ so, as in the Symmetric gauge, each Landau gauge eigenstate contains roughly a single quantum of flux, $\Phi_0$. Thus, the expected degeneracy $G$ in a finite system of area $A$ is $G = BA/(h/e) = \Phi/\Phi_0$, as before. The infinite degeneracy in both cases allows the wave functions in the two gauges to differ by more than the simple “textbook” phase, as discussed by Swenson[@swenson89] and Wakamatsu et al.[@wakamatsu18] In summary, the two gauges discussed both represent the same uniform magnetic field $\mathbf{B}=B\boldsymbol{\hat{z}}$. They result in the same energy spectrum, and yet they produce two very different-looking eigenstates (see Fig. \[fig: fs symm vs landau\]). There are gauge-invariant similarities as well; both gauges produce infinitely degenerate Landau levels in free space and both gauges have degeneracies that are expected to be approximately $G \equiv BA/\Phi_0$ in a finite space with area $A$. Eq. (\[psi\_transform\]) suggests that the two eigenstates should differ by a simple phase factor. That does not happen here, because of the degeneracy associated with the eigenstates. The dilemma is solved by taking an appropriate linear combination of the degenerate eigenstates (which is still an eigenstate), and these can be chosen to result in a state that differs from one in the other gauge by a simple phase factor, as illustrated in Refs. \[\] and \[\]. We will encounter this dilemma once more in a situation where the degeneracy has (nominally) been lifted, and outline a resolution by means different than that used in Refs. \[\] and \[\]. ![Contour plots of the free space probability density vs. $x$ and $y$, in length units of $\ell_B$, for (a) the Symmetric gauge centred about the origin, with $\ell=5$ and (b) the Landau gauge, with $k_y\ell_B=5$. The probability densities in each gauge have strikingly different spatial structure. It is clear that these eigenstates are not related by just a simple phase factor.[]{data-label="fig: fs symm vs landau"}](free-space-landau-symm.pdf) [Circular Confinement: Symmetric Gauge]{} {#circ} ========================================= We will revisit the issue of degeneracy in different gauges in Section \[square\] when we discuss an electron confined to a square dot. First, however, partly to review results due to Lent,[@lent91] and partly to introduce some technical aspects and more general results to be used later, we present a discussion of results for circular confinement. We will tend to use dimensionless units wherever possible. However, it is important to convert our results to physical values. As we have already seen, the flux quantum plays an important role in these problems; its value in SI units is $h/e \approx 4140$ T (nm)$^2$. We will use a dimensionless measure of flux, $G = BA/(h/e)$, where $A$ is the area of the sample through which a (constant) magnetic field penetrates. For circular geometry, we use $\gamma \equiv B\pi a^2/(h/e)$, where $a$ is the radius. So, for example, applying a $10$ T magnetic field through a sample with radius 100 nm gives $\gamma \approx 76$. [The Infinite Circular Well]{} ------------------------------ Let us first briefly review the problem of a particle confined to an infinite circular well. The infinite circular well potential $V(r)$ is simply given by: $$V(r) = \left\{ \begin{array}{ll} 0 & \quad r \leq a \\ \infty & \quad r > a \end{array} \right. \label{circ_pot}$$ where $a$ is the radius of the well. With this potential, the Hamiltonian in polar coordinates becomes: $$H=-\frac{\hbar^2}{2m}(\partial_r^2+\frac{1}{r}\partial_r)+\frac{L_z^2}{2mr^2}. \label{ham_polar}$$ Since $[H,L_z]=0$, we again write the eigenstates $\Psi(r,\phi) $ of the Hamiltonian as a separable product, $\Psi=e^{i\ell\phi}\psi(r)$, where $\hbar \ell$ is the same $L_z$ eigenvalue mentioned earlier and $\ell$ must be an integer. Substituting this into the Schrödinger Equation and introducing the wave vector $k=\frac{\sqrt{2mE}}{\hbar}$ results in $$\frac{d^2\psi}{dr^2}(r)+\frac{1}{r}\frac{d\psi}{dr}(r)+\left(k^2-\frac{l^2}{r^2}\right)\psi(r)=0. \label{bess_eqn}$$ Equation (\[bess\_eqn\]) is the Bessel equation, with solutions given by $$\psi(r)=AJ_\ell(kr)+BN_\ell(kr) \label{bess_fn}$$ where $A$ and $B$ are arbitrary constants, and $J_{\ell}(r)$ and $N_{\ell}(r)$ are the $\ell^{\rm th}$ order Bessel and Neumman functions, respectively. Because the Neumman functions are non-normalizable at the origin, we discard this solution, and apply the boundary condition: $AJ_{\ell}(ka)=0$. The solution is given by $ka=\beta_{n,\ell}$, where $\beta_{n,\ell}$ is the $n$th zero of the $\ell$th order Bessel function. We thus obtain the eigenvalues, $$E_{n,\ell}=\frac{\hbar^2\beta_{n,\ell}^2}{2ma^2}, \label{bess_eigen}$$ where $n$ is a positive integer that enumerates the zeros for the $\ell^{\rm th}$ order Bessel function. Notice the lack of degeneracy for all $n,\ell$. Bessel functions of the same order are both normalizable and orthogonal for different zeros over $r \in [0,a]$ for all integer $\ell$,[@Weisstein1] $$\int_{0}^aJ_\ell\left(\beta_{n,\ell}\frac{r}{a}\right)J_\ell\left(\beta_{n',\ell}\frac{r}{a}\right)rdr = \delta_{n,n'} \frac{a^2}{2}J_{\ell+1}^2\left(\beta_{n,\ell}\right). \label{ortho}$$ Because of the orthogonality of $e^{i\ell \phi}$, we need not worry about the orthogonality relations for Bessel functions of different order. These solutions imply that a suitable basis set for the “empty” infinite circular well are the functions $$\Psi_{n,\ell}(r,\phi)=\frac{e^{i\ell\phi}}{\sqrt{2\pi}}\frac{\sqrt{2}}{a}\frac{J_\ell\left(\beta_{n,\ell}\frac{r}{a}\right)}{J_{\ell+1}(\beta_{n,\ell})}. \label{circ_basis}$$ Since $J_{-\ell}(r)=(-1)^\ell J_\ell (r)$ then wave functions with opposite angular momentum are degenerate, as expected in the absence of a magnetic field. [Matrix Mechanics with Bessel Functions]{} ------------------------------------------ Matrix mechanics techniques have been utilized for problems like this before, albeit for a study of disordered systems, with a different gauge and confining potential.[@ohtsuki1988numerical] Here, when we include a magnetic field with the Symmetric gauge, we obtain $$H = \frac{\mathbf{p}^2}{2m} + \frac{\omega_c}{2}L_z+\frac{1}{8}m\omega_c^2r^2. \label{ham_circ}$$ To solve the Schrödinger equation with this Hamiltonian we follow the usual procedure of expanding in a basis (see, for example, Refs. \[,\]). We use $$\ket{\Psi}=\sum_{n^\prime=1}^{\infty}\sum_{\ell^\prime= -\infty}^{\infty}c_{n^\prime,\ell^\prime} \ket{n^\prime,\ell^\prime}$$ where $\ket{n',\ell'}$ is the ket corresponding to the circular well basis state, Eq. (\[circ\_basis\]). Acting on this state with the Hamiltonian, Eq. (\[ham\_circ\]), and taking the inner product with an arbitrary basis state $\ket{n,\ell}$ results in the matrix equation: $$\sum_{n^\prime=1}^{\infty}\sum_{\ell^\prime=-\infty}^{\infty} H_{n \ell, n^\prime \ell^\prime} c_{n^\prime,\ell^\prime} =E_{n,\ell}c_{n,\ell} \label{matrix1}$$ where $$H_{n \ell, n^\prime,\ell^\prime}=\bra{n,\ell}H\ket{n^\prime,\ell^\prime}, \label{hmatrix}$$ and, using Eq. (\[ham\_circ\]), $$H_{n \ell, n^\prime\ell^\prime}= \delta_{n,n^\prime}\delta_{\ell,\ell^\prime}E^{(0)}_{n,l}+\frac{1}{2}\omega_c\bra{n,\ell}L_z\ket{n^\prime,\ell^\prime}+\frac{1}{8}m\omega_c^2\bra{n,\ell}r^2\ket{n^\prime,\ell^\prime}. \label{hmatrix2}$$ Here, $E_{n,l}^{(0)}$ is simply the infinite circular well eigenenergy Eq. (\[bess\_eigen\]) and $\delta_{ij}$ is the standard Kronecker delta function. It should now be apparent that the entire matrix is diagonal in $\ell$, i.e. $ H_{n \ell,n^\prime \ell^\prime} = \delta_{\ell, \ell^\prime} H_{n,n^\prime}$, and the eigenvalue problem [for a given $\ell$ only]{} needs be solved. While a diagonalization is now required for each $\ell$ separately, this greatly reduces the computational cost. Evaluating the last two inner products of Eq. (\[hmatrix2\]), we obtain all the matrix elements for the Hamiltonian, $$H_{n,n^\prime}=\delta_{n,n^\prime}\left[\frac{\hbar^2\beta_{n,\ell}^2}{2ma^2}+\frac{\hbar\omega_c\ell}{ {2}}\right]+\frac{1}{8}m\omega_c^2a^2\rho^2_{n,n^\prime}. \label{hmatrix3}$$ Using energy units of $\hbar^2/(2ma^2)$, we can write this in dimensionless form $$\frac{H_{n,n^\prime}}{\hbar^2/(2ma^2)} \equiv h_{n,n^\prime} = \delta_{n,n^\prime}\left[\beta_{n,\ell}^2+2{\gamma} \ell \right]+ {\gamma^2}\rho^2_{n,n^\prime} \label{hmatrix4}$$ with $$\rho^2_{n,n^\prime}=\frac{\bra{n}r^2\ket{n^\prime}}{a^2}=\frac{2\int_{0}^1J_\ell(\beta_{n^\prime,\ell}\rho)\rho^2J_\ell(\beta_{n,\ell}\rho)\rho d\rho}{J_{\ell+1}(\beta_{n^\prime,\ell})J_{\ell+1}(\beta_{n,\ell})}. \label{rho_defn}$$\ Note that we have suppressed the dependency on $\ell$ in the matrix labels, but it is carried as a parameter in all of these matrices. Recall that the constant $G$ is related to the ratio of flux to the quantum of flux (given by $\gamma$), through $$\gamma = \pi G = \frac{B\pi a^2}{h/e}=\frac{\Phi}{\Phi_0}. \label{gamma_defn}$$ As was argued earlier, we expect $\gamma$ to roughly determine the degeneracy in this system. The integral in Eq. (\[hmatrix\]) can be evaluated analytically; we get $$\rho^2_{n,n^\prime} = \delta_{n,n^\prime} \Bigl[ \frac{\beta^2_{n,\ell} + 2\ell^2 -2} {3 \beta^2_{n,\ell}} \Bigr] + (1 - \delta_{n,n^\prime}) \Bigl[\frac{8\beta_{n^\prime,\ell}\beta_{n,\ell}}{(\beta_{n^\prime,\ell}^2-\beta_{n,\ell}^2)^2} \frac{J_{\ell-1}(\beta_{n^\prime,\ell})J_{\ell-1}(\beta_{n,\ell})}{J_{\ell+1}(\beta_{n^\prime,\ell})J_{\ell+1}(\beta_{n,\ell})} \Bigr] \label{bess_int1}$$ We now have all the matrix elements analytically, to insert into Eq. (\[matrix1\]) for diagonalization. The big advantage of having adopted the symmetric gauge with this geometry is that the resulting Hamiltonian is diagonal in one of the quantum numbers, $\ell$. This represents a tremendous computational gain. We mention it here because we will [not]{} have this possibility when we study the square geometry in the next section. There we will have to diagonalize $N^2 \times N^2$ matrices, where $N$ is a large number (here typically something like $50$ or $200$. Also, for circular geometry we can monitor the quantum number $\ell$ precisely, which means we can track the radial quantum number $n_r$ as well. In the Symmetric gauge the Landau level quantum number $n$ is given simply by $n = n_r + (\ell + \|\ell\|)/2$. We will not have this possibility with the square geometry. [Results and Discussion]{} -------------------------- To be clear, in practice in this section we diagonalize $N \times N$ matrices, and only show results that have converged as $N$ increases (recall that $N$ should be infinite, but is chosen to be finite and represents the number of radial basis states (i.e. number of $n$) for a given $\ell$ in Eq. (\[circ\_basis\]) to be used in the matrix diagonalization). We will do this for a number of different $\ell$’s (typically $N$ of these), again so that all presented properties are converged as a function of basis size. ### [EigenEnergies]{} Having carried out this programme, we can now order the eigenvalues in increasing value. We do this to mimic the result we would have achieved had we not recognized that the matrix is diagonal in $\ell$ (as will be the case with the square geometry). The eigenvalue results for a few values of $\gamma$ are shown in Fig. \[circular\_eigen\], and show in all cases plateaus (i.e. degenerate levels) separated by regions in which the energies increase to the next plateau. The number of points from the start of one plateau to the start of the next plateau is essentially $\gamma$. ![Dimensionless eigenvalues ($E_n/(\hbar^2/(2ma^2))$) as a function of total quantum number $n$ for various values of $\gamma$. Here, $n$ is used as a label for the horizontal axis to indicate the energies are plotted in ascending order and does not refer to any previously defined quantum numbers also labelled with $n$. The truncation sizes of the matrices produced are given by $N = 200$ in all cases. Note that these plots were produced as a result of generating and diagonalizing matrices for $\ell\in [-200,200]$ followed by aggregating the resulting eigenenergies acquired from all matrices and organizing them in ascending order. This results in over 80000 eigenvalues, and the lowest 200 of these are shown here. Note that the degeneracy increases with the value of $\gamma$ according to our expectations, and in fact there are precisely $\gamma$ eigenvalues occurring before the next plateau begins. Also note that the plateaus occur at dimensionless energies given by those expected from the free space Landau levels: $E_{n}/(\hbar^2/(2ma^2)) = 2\gamma(2n+1)$.[]{data-label="circular_eigen"}](circ-confine-eigenenergies.pdf) The so-called “plateaus” are not degenerate to infinite precision, but they are degenerate to the precision of our computer (16 digit accuracy). These plateaus appear to be the Landau levels, originally infinitely degenerate, but now with a degeneracy of order $\gamma$, and in practice less. The increase in eigenvalues towards the end of the plateau regions is indicative that these eigenfunctions can feel the edge of the confining potential, and these are usually referred to as “edge states”, because, as we shall see, their probability density is concentrated there. One might ask, where have the rest of the (originally) infinite states in a given Landau level gone? Fig. \[circular\_eigen\] gives the impression that there were $\gamma$ such states, and about three quarters of them have remained degenerate, while the remaining one quarter have had their energies elevated due to the edge. A clue that this interpretation is incorrect arises from something barely discernible on this plot, which is that the increase in eigenvalues at the ends of the plateaus become increasingly less smooth for successive plateaus. The reason for this is made clear in Fig. (\[circular\_eigen\_ell\]a), where now the energy levels are plotted as a function of $\ell$. Again we emphasize that we are fortunate that this is possible here, because different matrices were diagonalized for individual values of $\ell$. In Fig. (\[circular\_eigen\_ell\]b) we show the same energy levels vs. $r_\ell/a \equiv \sqrt{\|\ell\|/\gamma}$, to show how the maximum in the wave function moves towards the edge as $\ell$ increases (with the cautionary proviso that $r_\ell$ represents the maximum only for cases well away from the wall – which is why $r_\ell/a$ ends up exceeding unity in this plot — but the qualitative trend is correctly portrayed). Figure (\[circular\_eigen\_ell\]) illustrates that the eigenenergies are smoothly increasing when plotted vs. $\ell$. The raggedness that exists in Fig. (\[circular\_eigen\]) is due to the toggling back and forth between edges states of similar energy from the various Landau levels, uniquely identified in Fig. (\[circular\_eigen\_ell\]) \[(and not in Fig. (\[circular\_eigen\])\]. The Landau levels are distinct and clear, except now they carry on indefinitely, to the left in Fig. (\[circular\_eigen\_ell\]a), and to the right in Fig. (\[circular\_eigen\_ell\]b). ![Lowest 1000 Eigenenergies plotted vs. (a) $\ell$ and (b) $r_\ell/a=\sqrt{\|\ell\|/\gamma}$, for $\gamma=150$ and matrix truncation size, $\text{N}=200$. $\ell$ is the angular momentum quantum number and $r_\ell$ is the centre of the probability density of the corresponding eigenstate as long as the eigenstate is not too close to the walls. This means that $r_\ell$ gives us a good idea about where the eigenstates corresponding to the eigenergies are located in real space, as long as $r_\ell/a { { \atop <} \atop {\sim \atop }} 0.9$. Values of $r_\ell/a>1$ should be interpreted as corresponding to eigenstates piling up near the boundary. This plot (as opposed to Fig. \[circular\_eigen\]), makes clear that the original Landau level with infinite degeneracy ([all]{} negative values of $\ell$) has the very large $\|\ell\|$ states pile up near the boundary, with increasingly higher energies, thus breaking the original free Landau (infinite) degeneracy. As is clear from the figure, though, a quasi-degeneracy of order $\gamma$ (proportional to the applied magnetic field) remains.[]{data-label="circular_eigen_ell"}](E-vs-rl-l.pdf) ![The first 100 levels of the Fock-Darwin spectrum for the circular well. Also shown (with red dashed lines) are the free Landau levels. In this case, in contrast to that shown in Fig. (\[fock\_darwin\]) with parabolic confinement, the levels “condense” to these free Landau levels as the magnetic field increases, as illustrated by Lent (his Fig. 1).[@lent91][]{data-label="fock_darwin1"}](FD-spec-circ-confine.pdf) As already stated and further corroborated below, the increase in energy is due to the confinement of the edge. The ‘toggling’ is evident if one examines the states at a dimensionless energy value like $\approx 1400$ in Fig. (\[circular\_eigen\_ell\]a). The next energy level that would be placed in Fig. (\[circular\_eigen\]) would come from one of the two branches either emanating from the first or the second Landau level, and these would go back and forth. At a higher value, say $\approx 2000$, toggling would occur between three branches, and so on, leading to increasing “raggedness” in Fig. (\[circular\_eigen\]) as the energy goes up, which is precisely what we observed. Finally, we also include a Fock-Darwin spectrum, as in Lent,[@lent91] shown in Fig. (\[fock\_darwin1\]), where the so-called “condensation” of levels occurs.[@robnik86] As $\gamma$ increases, more and more levels coalesce as the free space Landau levels become more applicable since the wave functions become more compact and away from the edge. The thick dashed red lines indicate the expected Landau levels indicative of free space. To the far left, with the exception of $\ell = 0$ the eigenvalues come in pairs, corresponding to positive and negative $\ell$ which are degenerate in the absence of a magnetic field. Such condensation is absent with parabolic confinement, which is presumably why these are called ‘Landau’ levels and not ‘Fock’ levels. ### [Probability Density]{} We have been discussing edge states, and using the parameter $r_\ell \equiv \ell_B\sqrt{2\|\ell\|}$ (recall $\ell_B \equiv \sqrt{\hbar/(eB)}$ is the magnetic length) to indicate the location of the wave function. Of course $r_\ell$ is the radius of the maximum of the wave function (from some specified origin) in free space only, so here we examine the actual probability densities. Recall that the probability density is gauge invariant, except in cases of degenerate wave functions. Here the degeneracy is removed because of the confining cylinder, but as already discussed, a ‘practical’ degeneracy remains, so the statement about gauge invariance is no longer true,[@swenson89] as we shall see below and in the next section. ![Contour plots of $\|\Psi_{n_r,\ell}(r,\phi)\|^2$ for various values of $n_r$ and $\ell$, for $\gamma = 150$. The number of nodes increases with increasing $n_r$, more clearly seen for low values of $\|\ell\|$, and the radius of the maximal amplitude increases with increasing $\|\ell\|$. The notion that large $\|\ell\|$ states are “edge” states is clear from this plot, as the entire probability density resides on the circumference for these states. []{data-label="prob_density1"}](circ-confine-contours-l-nr.pdf) ![Contour plots of $\|\Psi_{n}(r,\phi)\|^2$ as a function of total quantum number $n$ for various values of $\gamma$, as indicated at the top of each column. Trends similar to those shown in Fig. (\[prob\_density1\]) are evident, but in addition one can note changes in $\ell$ and $n_r$ quantum numbers. For example, $\|\Psi_{50}\|^2$ for $\gamma = 50$ has $n_r = 0$ and $\ell = +1$, and is located at the start of the 2nd plateau visible in Fig. (\[circular\_eigen\]). It is identical to $\|\Psi_2(r,\phi)\|^2$ for $\gamma = 50$ displayed here in the 2nd row. The wave function $\Psi_2(r,\phi)$ has quantum numbers $n_r=0$ and $\ell = -1$ (recall that $\|\Psi\|^2$ was insensitive to the sign of $\ell$). Note that this plot is orderly and symmetric because we have artificially broken the remaining quasi-degeneracy apparent in Figs. (\[circular\_eigen\]) and (\[circular\_eigen\_ell\]) with a [very]{} shallow parabolic trap, centred at the middle of the cylinder. This shallow trap changes nothing that is visible to the eye, but it does remove the remaining quasi-degeneracy so that a proper ordering is established. See the text and the next section for further explanation of this additional potential, whose sole purpose is to remove the degeneracy.[]{data-label="prob_density1b"}](circ-confine-contours-G-state.pdf) In Fig. (\[prob\_density1\]) we show contours of $\|\Psi_{n_r,\ell}(r,\phi)\|^2$ for $\gamma=150$ for various values of $n_r$ and $\ell$. These indicate that the wave function amplitude moves out from the centre as $\|\ell\|$ increases, as noted analytically for the free space result. We should note that Fig. (\[prob\_density1\]) looks very “orderly” when it comes to the progression of the number of nodes (as one moves to the right, increasing $n_r$) and the radius of maximum amplitude (as one moves down, increasing $\|\ell\|$). This is because we have complete control of the quantum numbers, as stressed with respect to Fig. (\[circular\_eigen\_ell\]). This scenario is different when we simply order all the states according to their energies \[this was done to produce Fig. (\[circular\_eigen\])\]. This will make the progression of states somewhat disordered, [for a reason entirely different]{} than the disorder already noted in connection with Fig. (\[circular\_eigen\]). [That]{} disorder was visible to the eye. Now we are referring to the quasi-degenerate states in the plateau regions, where the disorder [is not visible, but nonetheless exists at a level lower than than $10^{-16}$]{} (so it is invisible even to the computer, using double precision accuracy). For this reason, even in the plateau regions in Fig. (\[circular\_eigen\]), the states may be “out of order,” and a state with a higher $\|\ell\|$ may be ranked lower (according to energy) compared with a lower $\|\ell\|$ state. To avoid this we artificially include a [very shallow]{} parabolic trap, whose sole purpose is to break the remaining quasi-degeneracy, so that the progression of states is orderly. The resulting contour plots are shown in Fig. (\[prob\_density1b\]). We will discuss in more detail the introduction of this “degeneracy-breaking” potential in the next section. Including it as we do here does not change any of the results concerning energy that are visible to the eye; it does order the quasi-degenerate energy levels at the $10^{-10}$ level, so the computer “knows” which states are supposed to come first, and does not resort to a somewhat random linear combination, as happens when there is a degenerate subspace of solutions. ![Dimensionless radial probability density $(r/a)a^2\|\psi_{n_r,l}(r/a)\|^2$ vs $r/a$, for $\gamma = 150$, for various $\|\ell\|$ with $n_r=0$. We have used a matrix size to ensure that the results are converged ($N=200$). Note that only positive $\ell$ (i.e. $\|\ell\|$) values are used since the probability density does not depend on the sign of $\ell$. Also note that the area under each curve is unity, as required by normalization. Finally, note that the edge states are those with large values of $\|\ell \|$, as expected.[]{data-label="prob_density2"}](radial-prob-density-plots-circ-confine.pdf) Because of the circular symmetry, it is enlightening to examine the probability density as a function of the radial coordinate, $r$. Defining a purely radial wave function through $\Psi_{n_r,\ell}(r,\phi) \equiv \frac{e^{i\ell\phi}}{\sqrt{2\pi}}\psi_{n_r,\ell}(r)$, we plot in Fig. (\[prob\_density2\]) the radial probability density, $({r/a}) a^2\|\psi_{n_r,\ell}(r)\|^2$ vs. $r/a$ for a variety of values of $\ell$, with $n_r = 0$, for $\gamma = 150$. These are properly normalized, and show clearly that the large $\|\ell\|$ states are edge states. ### [Probability Current]{} In the Integer Quantum Hall Effect, edge states (eigenstates localized along a boundary) are often cited as a key ingredient to explain the quantization of hall conductance. [@qheTong] Classically speaking, the idea is that if an electron encounters a boundary while undergoing a cyclotron orbit, it will reflect off the wall and once again try to complete its orbit. This causes the electron to undergo ’skipping’ orbits along the boundary, resulting in chiral edge currents along the boundary of the sample. This was the motivation for examining the (probability) current associated with these edge states. One would expect that edge states would contribute non-zero current to the system while states in the bulk would not contribute any current. Lent[@lent91] illustrated such currents, and since the matrix mechanics technique allows for easy numerically exact calculations of this property, we will illustrate them here as well. The probability current density $\mathbf{J}$ of an electron in an eigenstate $\Psi$ immersed in a magnetic field represented by a gauge choice $\mathbf{A}$ is given by [@landau77; @wysin] $$\mathbf{J}=\frac{1}{2m}\left[\Psi^\mathbf{p}\Psi-\Psi\mathbf{p}\Psi^+2e\|\Psi\|^2\mathbf{A}\right]. \label{current1}$$ Since we are working in two dimensions, ${\mathbf{J}}$ is a surface current density with dimensions $IL^{-1}$. Using the symmetric gauge $\vec{\bf A}_s$, the probability current density associated with an eigenstate $\Psi_{n_r,\ell}(r,\phi)=\frac{e^{i\ell\phi}}{\sqrt{2\pi}}\psi_{n_r,\ell}(r)$ is, in polar coordinates, $$\mathbf{J}_{n_r,\ell}(r)=\frac{\|\psi_{n_r,\ell}(r)\|^2}{2\pi}\left[\frac{\hbar \ell}{mr}+\frac{1}{2}\omega_cr\right]\boldsymbol{\hat{\phi}}=\frac{\|\psi_{n_r,\ell}(r)\|^2}{2\pi}\frac{\Lambda_z(r)}{mr} \boldsymbol{\hat{\phi}} \label{current2}$$ where $\Lambda_z(r)=\hbar \ell+\frac{1}{2}m\omega_cr^2$ is the kinetic angular momentum in the $\hat{\boldsymbol{z}}$ direction. Note that this can also be expressed as $$\mathbf{J}_{n_r,\ell}(r)=\frac{\|\psi_{n_r,\ell}(r)\|^2}{2\pi}\sqrt{\frac{2V_\ell(r)}{m}}\boldsymbol{\hat{\phi}} \label{current3}$$ where $V_\ell(r)=\frac{\Lambda_z^2}{2mr^2}$ is the kinetic centrifugal potential. This form is particularly interesting because we see that ${\mathbf{J}}$ is simply the product of the radial probability density and the classical velocity of a particle in the radial potential $V_l(r)$. As can be inferred from Eq. (\[current2\]), and will be evident in the figures shown below, ${\mathbf{J}}$ has non-zero magnitude only in the $\boldsymbol{\hat{\phi}}$ direction, vanishes and has an inflection point at $r_\ell$ for $\ell<0$, is maximum at $r_\ell$ for $\ell>0$, and only vanishes at $r_\ell=0$ for $\ell=0$. To obtain the current from a surface current density, we integrate it with respect to a scalar differential ‘strip’ orthogonal to the direction of flow. The state current ${\mathbf{I}}_{n_r,\ell}$ of a given eigenstate is thus given by $${\mathbf{I}}_{n_r,\ell}=\int_{0}^{a} {\mathbf{J}}_{n_r,\ell}(r)dr. \label{current4}$$ In dimensionless length ($a$) and current density units ($\hbar/(2\pi ma^3)$, we have $$\mathbf{j}_{n_r,\ell}(x)=\|\tilde{\psi}_{n_r,\ell}(x)\|^2\left[\frac{\ell}{x}+\gamma x\right]\boldsymbol{\hat{\phi}} \label{current6}$$ where $x \equiv r/a$, $\tilde{\psi}_{n_r,\ell} \equiv a\psi_{n_r,\ell}$, and $\gamma \equiv B\pi a^2/(h/e)$ as before. The state current is now given by $${\mathbf{I}}_{n_r,\ell}/\frac{\hbar}{2\pi ma^2}={\mathbf{i}}_{n_r,\ell}=\int_{0}^{1} {\mathbf{j}}_{n_r,\ell}(x)dx. \label{current7}$$ We can estimate roughly what to expect for a state far from the edge; for example, for $n_r=0$ in free space, we have $$\|\tilde{\psi}^{(FS)}_{0,\ell}\|^2=\frac{2\gamma^{\|\ell\|+1}x^{2\|\ell\|}}{\|\ell\|!}e^{-\gamma x^2}, \label{current_fs}$$ so the free space (FS) state current magnitude $i^{(FS)}_{0,\ell}$ is $$\begin{aligned} i^{(FS)}_{0,\ell}&=\int^{\infty}_0\left[\frac{\ell}{x}+\gamma x\right]\|\tilde{\psi}^{(FS)}_{0,\ell}\|^2dx\\[5pt] &=\frac{2\gamma^{\|\ell\|+1}\ell}{\|\ell\|!}\int^{\infty}_0x^{2\|\ell\|-1}e^{-\gamma x^2}dx+g\underbrace{\int_0^{\infty}\|\tilde{\psi}^{(FS)}_{0,\ell}\|^2xdx}_{=1}\\ &=\gamma \frac{\ell}{\|\ell\|!}\underbrace{\left(\int^{\infty}_0x^{\|\ell\|-1}e^{-x}dx\right)}_{=\Gamma(\|\ell\|)=(\|\ell\|-1)!}+\gamma \\ &=\gamma [\text{sgn}(\ell)+1]. \label{current9}\end{aligned}$$ Therefore, the state current in free space is given by: $$i^{(FS)}_{0,\ell} =\begin{cases} 2\gamma \ &\text{if} \ \ell>0\\ \gamma \ &\text{if} \ \ell=0\\ 0 \ &\text{if} \ \ell<0.\\ \end{cases} \label{current10}$$ Thus, the dimensionless state current is constant in free space with a value of $2\gamma$ for $\ell>0$ and with a value of $0$ for $\ell<0$. The case $\ell=0$ in free space actually holds for all $n_r$ due to the orthonormality of the eigenstate, $$i^{(FS)}_{n_r,0}=\gamma \int_0^{\infty}\|\tilde{\psi}^{(FS)}_{n_r,0}\|^2xdx=\gamma. \label{current11}$$ This case is also exactly calculable in our confined system as well, since we can expand $\tilde{\psi}_{n_r,0}$ as a series of Bessel functions: $$i_{n_r,0}=\gamma \int_0^{1}\|\tilde{\psi}_{n_r,0}\|^2xdx=\gamma. \label{current12}$$ To better understand how the bulk current of our confined system will behave in comparison, we obtain $$\begin{aligned} i_{0,\ell}&=\int^{1}_0\left[\frac{\ell}{x}+\gamma x\right]\|\tilde{\psi}_{0,\ell}\|^2dx\\[5pt] &=\int^{1}_0\frac{\ell}{x}\|\tilde{\psi}_{0,\ell}\|^2dx+\int^{1}_0\gamma x\|\tilde{\psi}_{0,\ell}\|^2dx\\[5pt] &=\int^{1}_0\frac{\ell}{x}\|\tilde{\psi}_{0,\ell}\|^2dx+\gamma \label{current13}\end{aligned}$$ The integral term left in the last line of Eq. (\[current13\]) is a bit tricky to calculate with a series expansion; since we are interested in the current in the bulk of the well (i.e. $\|\ell\|<<\gamma$) and the free space solutions are fairly localized, we approximate $\|\tilde{\psi}_{0,\ell}\|^2$ with its free space solution Eq. (\[current\_fs\]), $$\begin{aligned} \|\tilde{\psi}_{0,\ell}\|^2&\approx\|\tilde{\psi}^{(FS)}_{0,\ell}\|^2 \ \text{for} \ \|\ell\|<<\gamma \\[5pt] \Rightarrow i_{0,\ell}&\approx\frac{2\gamma^{\|\ell\|+1}\ell}{\|\ell\|!}\int^{1}_0x^{2\|\ell\|-1}e^{-\gamma x^2}dx+\gamma \\[5pt] &=\gamma \frac{\ell}{\|\ell\|!}\left(\int^{\gamma}_0x^{\|\ell\|-1}e^{-x}dx\right)+\gamma\\[5pt] &=\gamma \left[\frac{\ell}{\|\ell \|!}\bar{\gamma}(\|\ell\|,\gamma)+1\right] \label{current14}\end{aligned}$$ where $\bar{\gamma}(\|\ell\|,\gamma)$ is the lower incomplete gamma function given by [@Weisstein2] $$\begin{aligned} &\bar{\gamma}(n,z)=\int^z_0e^{-t}t^{n-1}dt\\ &\bar{\gamma}(n,z){\underset{z\rightarrow\infty}{\sim}}\Gamma(n)=(n-1)! \label{gamma_fn}\end{aligned}$$ where we have also shown the asymptotic form for large $z$. Thus, in our confined system, we expect the state current magnitude far from the boundary to approach the free space results Eq. (\[current10\]) for large $\gamma$: $$\begin{aligned} i_{0,\ell}&{\underset{\gamma \rightarrow\infty}{\sim}}\gamma \left[\frac{\ell}{\|\ell\|!}\Gamma(\|\ell\|)+1\right]=\gamma[\text{sgn}(\ell)+1] =\begin{cases} 2\gamma \ &\text{if} \ \ell>0\\ \gamma \ &\text{if} \ \ell=0\\ 0 \ &\text{if} \ \ell<0\\ \end{cases}\nonumber \\ \Rightarrow i_{0,\ell}&{\underset{\gamma \rightarrow\infty}{\sim}}i^{(FS)}_{0,\ell}. \label{current15}\end{aligned}$$ So, the current contribution from the $l=0$ state is given by $i_{0,0}=\gamma$. This surprising result gives us both an expectation for our numerical results and implies that near the centre of the well, the probability current is nonzero. Moreover, it seems for very large $\gamma$, the current is expected to be a constant as a function of $\ell$ of a given sign. For $\ell<0$, the current contribution is approximately $0$, which is what would be expected in the bulk. However, for $\ell>0$ it is clear that the current contribution will be approximately $2\gamma$, which is unexpected because it implies that there exists non-zero current in the bulk of the sample. ![Probability current density $J$ (recall $\mathbf{J}=J\hat{\mathbf{\phi}}$) (in units of $\frac{\hbar}{2\pi ma^3}$) vs $r/a$ for (a) various negative and (b) various positive values of $\ell$, with $n_r=0$, and $\gamma=150$. Convergence was achieved with $N= 200$. The $\ell=0$ state is also shown for reference. In contrast to the negative $\ell$ states, the positive $\ell$ states all circulate in the direction expected classically (through the Lorentz force), but are energetically unfavourable. As with the negative $\ell$ states, for larger $\|\ell\|$ states, the current density is concentrated near the edge, but now positive. Note how the $\ell=0$ state has current density concentrated near the origin, while other (negative) $\ell$ states have current density located (naturally) where their probability density dominates, but with an inflection point \[as predicted from Eq. (\[current2\])\] right at $r = r_\ell$, [as long as the probability density is well away from the edge.]{}[]{data-label="prob_current2"}](radial-prob-curr-circ-confine.pdf) ![A vector field plot of the probability current for a “bulk” state in (a) ($\ell = -25$, $n_r = 0$) and an edge state in (b) ($\ell=-300$, $n_r = 0$). In both cases we used $\gamma=100$, and convergence was attained with a matrix size of $N=200$. The expanded portions show what should already be clear from Fig. (\[prob\_current2\]).[]{data-label="circ_vec_field"}](circ_confine_vec_field.pdf) Numerical results are shown in Fig. (\[prob\_current2\]) for (a) negative $\ell$ states and for (b) positive $\ell$ states. The expectations from our earlier analytical analysis are borne out by these figures. In Fig. (\[circ\_vec\_field\]) we show a vector plot of the current for a “bulk” state in (a) ($\ell = -25$) and an edge state in (b) ($\ell = -300$), with the plots on the right showing the same result in more detail. The calculation of current is shown in Figure (\[current\]); these are numerical results for the current in the $\hat{\phi}$ direction, and the agreement with the analytical estimate, Eq. (\[current15\]) is remarkable. Our numerical results are also in qualitative agreement with those shown by Lent,[@lent91] although he utilized smaller values of $\gamma$ (his $\beta \equiv 2\gamma$), and he showed pictorial vector plots for fairly small magnitudes of $\ell$ (which he calls $m$). He also provides a nice description in terms of classical orbits, which we will not repeat here; the reader is referred to Ref. \[\] for this description. ![Individual state current $I_{n_r=0,\ell}$ vs (a) $\ell$ and vs (b) $r_\ell/a$. We include (b) to illustrate how the constant current begins to be impacted by the edge, once $r_\ell$ is within a distance $\ell_B$ of the edge. Here $\ell_B \approx 0.06a$, and the deviation from the constant value predicted in Eq. (\[current15\]) starts shortly before this value is reached. Note the symmetry between negative and positive $\ell$ states. Also note how these numerical results agree very well with our analytical estimates, Eq. (\[current15\]).[]{data-label="current"}](curr-vs-rl-l.pdf) [Square Confinement]{} {#square} ====================== Relatively little work has been done to date for a square geometry. In this section we present results in both the Symmetric and Landau gauges, and illustrate the difficulties encountered for large fields (or samples) due to the practical degeneracy that remains in these cases. We also note some peculiarities with the edge states for large quantum numbers. [Symmetric Gauge]{} ------------------- ### [Matrix Element Calculation]{} The confining potential is defined as $$V(x,y)=\begin{cases} 0 \ &\text{if} \ 0< x < a \ \ {\rm and} \ \ 0<y<a\\ \infty \ &\text{otherwise} \label{pot_confining_sq} \end{cases}$$ so $a$ is the length of a side of the two-dimensional infinite square well representing the square quantum dot. Given this potential, a convenient set of normalized basis states is $$\braket{x,y}{n_x,n_y}=\frac{2}{a}\sin{\left(\frac{n_x\pi x}{a}\right)}\sin{\left(\frac{n_y\pi y}{a}\right)}, \label{basis_sq}$$ where $n_x$ and $n_y$ are positive integers. For convenience, we modify the symmetric gauge ${\mathbf{A}}_S$ such that it is centred in the potential well: $${\mathbf{A}}'_{S}=\frac{B}{2}\left[\left(x-\frac{a}{2}\right)\hat{{\mathbf{y}}}-\left(y-\frac{a}{2}\right)\hat{{\mathbf{x}}}\right]. \label{symm_shifted}$$ With this gauge ${\mathbf{A}}'_S$, our Hamiltonian takes the form: $$H=\frac{\textbf{p}^2}{2 m_e}-i\frac{\hbar\omega_c}{2}\left[\left(x-\frac{a}{2}\right)\partial_y-\left(y-\frac{a}{2}\right)\partial_x\right]+\frac{1}{8}{m_e}\omega_c^2\left[\left(x-\frac{a}{2}\right)^2+\left(y-\frac{a}{2}\right)^2\right] \label{ham_sq}$$ where we use $m_e$ as the mass of the electron to avoid confusion with matrix indices. Following the steps leading to Eq. (\[matrix1\]) we then expand the eigenstate of (\[ham\_sq\]), $\ket{\Psi}$, as: $$\ket{\Psi}=\sum_{n_x,n_y}c_{n_x,n_y}\ket{n_x,n_y} \label{expand}$$ Thus, in energy units of $E_0 \equiv \hbar^2\pi^2/(2{m_e}a^2)$, and using $n\equiv (n_x,n_y)$ and $m \equiv (m_x,m_y)$ for short, our dimensionless matrix elements, $h_{n,m} \equiv H_{n,m}/E_0$ for the Hamiltonian for this problem are given by\ $$\begin{split} h_{n,m}&=\delta_{n_x,m_x}\delta_{n_y,m_y} \left[n_x^2+n_y^2+\frac{G^2}{2}\left(\frac{1}{3}-\frac{1}{(\pi n_x)^2}-\frac{1}{(\pi n_y)^2}\right)\right]\\[5pt] &+2 \left(\frac{G}{\pi}\right)^2 \Bigl( (1-\delta_{n_x,m_x})\delta_{n_y,m_y} g_e(n_x,m_x) + (1-\delta_{n_y,m_y})\delta_{n_x,m_x} g_e(n_y,m_y) \Bigr)\\[5pt] &-i\frac{16G} {\pi^3}(1-\delta_{n_y,m_y})(1-\delta_{n_x,m_x})\left[\left(g_o(n_y,m_y)f(n_x,m_x)-g_o(n_x,m_x)f(n_y,m_y)\right)\right] \end{split} \label{ham_sq_dim}$$ with $$g_e(n,m)=\frac{1 + (-1)^{n+m}}{ 2}\left(\frac{1}{(n-m)^2}-\frac{1}{(n+m)^2}\right)=\begin{cases} 0 \ &\text{if} \ n+m=\text{odd}\\ \frac{1}{(n-m)^2}-\frac{1}{(n+m)^2} \ &\text{if} \ n+m=\text{even} \end{cases} \label{ge_def}$$ $$g_o(n,m)=\frac{1 - (-1)^{n+m}} {2}\left(\frac{1}{(n-m)^2}-\frac{1}{(n+m)^2}\right)=\begin{cases} \frac{1}{(n-m)^2}-\frac{1}{(n+m)^2} \ &\text{if} \ n+m=\text{odd}\\ 0 \ &\text{if} \ n+m=\text{even} \end{cases} \label{go_def}$$ $$f_o(n,m)=\frac{1-(-1)^{n+m}} {2}\frac{nm}{n^2-m^2}=\begin{cases} \frac{nm}{n^2-m^2} \ &\text{if} \ n+m=\text{odd}\\ 0 \ &\text{if} \ n+m=\text{even} \end{cases} \label{fo_def}$$ where $$G=\frac{Ba^2}{h/e} \equiv \frac{\Phi} {\Phi_0}$$ as defined earlier, and the subscripts ‘e’ and ‘o’ in the definitions, Eqs. (\[ge\_def\],\[go\_def\],\[fo\_def\]) serve to remind us that these are respectively non-zero for even or odd sums of the indices only. ### [Results and Discussion]{} Note that a lower truncation is [necessarily]{} used here since matrices with block matrices as elements have to be constructed and diagonalized, which is much more computationally intensive than in the problem with circular confinement. Thus a truncation of $N=100$ for one of the indices, say $n_x$, implies a total truncation size of $N^2$ for both $n_x$ and $n_y$. We will refer only to the block matrix truncation size ($N$) in the following results (so $N=100$ values of $n_x$ requires diagonalization of a $10000 \times 10000$ matrix). [Eigenenergies]{} The computed eigenvalues are shown in Fig. (\[square\_eigen\]) for various values of $G$ as indicated. This figure shares many things in common with its circular counterpart, Fig. (\[circular\_eigen\]), and so it should, since for ‘bulk-like’ states we have argued that they resemble the free space results and hence, do not “feel” the edge, i.e. the electrons don’t even know if they are confined in a circular or square geometry. ![Eigenvalues shown in ascending order vs quantum number for $G=50$, $100$, and $150$. This figure is the ‘square’ version of its circular counterpart, Fig. (\[circular\_eigen\]), and we [do not]{} have the counterpart to Fig. (\[circular\_eigen\_ell\]), which would allow us to sort out here the two quantum numbers that are present in that case. Just like its circular well counterpart, the energies plateau at free space Landau level energies. Two noticeable features are the plateaus, as in the circular case, especially for large values of $G$, and the number of states present between plateaus, precisely $G$ here, just as it was $\gamma$ for the circular case. Also noticeable is a similar “raggedness” for energy values beyond the higher plateaus, presumably due to the same “toggling” phenomenon that we could explicitly identify in the circular case. Note that the free space Landau level energies are given by: $E_{n_L}/\left(\frac{\hbar^2\pi^2}{2ma^2}\right)=\frac{4G}{\pi}\left(n_L+\frac{1}{2}\right)$ (here we use $n_L$ for the Landau level number to avoid confusion; this plot has energy levels corresponding to the lowest 2 or 3 Landau levels). We used $N=150$ here.[]{data-label="square_eigen"}](square-confine-symm-eigenenergies.pdf) However, the concept of a good quantum number $\ell$ does not exist here, at least not explicitly. So we do not have the benefit of a figure like Fig. (\[circular\_eigen\_ell\]) for the square geometry. One key similarity is the role of the ratio of the total flux to the flux quantum, which gives precisely the number of states between plateaus, regardless of geometry. In Fig. (\[square\_fock\_darwin\]) we show the Fock-Darwin spectrum for the case of square confinement. As with circular confinement the energy levels eventually “condense” and become coalesced into the degenerate Landau levels, as indicated. Both the circular and square confinement show this “condensation” phenomenon, whereas the parabolic confinement does not.\ ![The Fock-Darwin spectrum for the two dimensional infinite square well. Here $N=100$. This plot shares with its counterpart for the circular well, Fig. (\[fock\_darwin1\]) the idea that for large enough field the levels “condense” to a set of degenerate Landau levels (indicated by thicker dashed red lines for the first 3 levels. For weaker fields the levels are fairly disordered, reflecting the more complicated geometry of the square. Note that the zero field energies no longer consistently come in pairs, as the degeneracy pattern for an infinite square well is more complicated than that for an infinite circular well. []{data-label="square_fock_darwin"}](FD-spec-square-confine.pdf) [Probability Densities]{} As we just remarked, for the so-called “bulk-like” states we expect the results for the square well to be very similar to those of the circular well, simply because the magnetic field keeps the electron sufficiently contained in the central region of the well so that the electron does not “know” the geometry of the confining potential. In Fig. (\[prob\_density\_sq1\]) we show contours of the probability density for a variety of values of G. This plot is to be compared with Fig. (\[prob\_density1\]), which showed probability densities for the circular well. In that plot, however, we could specify both $n_r$ and $\ell$, whereas here we can only order the plots according to the value of the energy. Some features of this plot are immediately apparent. First, results with small quantum number and small $G$ are consistent with what we found for circular geometry. This is expected. More excited states become fairly diffuse throughout the square, and retain the symmetry of the square; this is also expected. However, as $G$ increases, the results become somewhat irregular. The problem here was already alluded to in the case of the circular well: a high degree of “practical” degeneracy remains as $G$ increases. We fixed the problem with little explanation in that case, so now we discuss this issue in more detail. ![Probability density, $\|\Psi_{n}(x,y)\|^2$ contour plots for various values of total quantum $n$, as ordered by the diagonalization subroutine. Here the confining region is a square, outlined in black. The results only look sensible in the first two columns; the low quantum number results resemble those obtained from the circular geometry, as expected, and then the probability densities become more “square-like”, reflecting the geometry of the confinement, also as expected. As one increases $G$, however, the results become less understandable, and appear to be wrong. The difficulty, as discussed in the text, is the high degree of degeneracy that remains when large values of $G$ are used. We used $N=150$ for all the square well results.[]{data-label="prob_density_sq1"}](square-confine-symm-contour-eps0.pdf) ![Probability density, $\|\Psi_{n}(x,y)\|^2$ contour plots for the same parameters as in Fig. (\[prob\_density\_sq1\]), but now with an additional confining potential given by Eq. (\[hamp\]) with $\omega_0/\omega_c =1.3\times 10^{-6}$. Now there is a more systematic progression, for example, for the ground state as one increases $G$; the wave function becomes more localized at the centre of the square as $G$ increases. In fact this could be happening in a circular geometry, and the result would be the same. As we move down, i.e. to more excited states for a given $G$, the wave function becomes more delocalized, and takes the shape of the edges in the square; these are veritable edge states. One notable exception is the result for $\|\psi_{50}\|^2$ with $G=50$, where the wave function becomes more concentrated; this is likely because we have suddenly moved on to the next Landau level (recall how we could track this explicitly in the case of the circular well). Note that in general the results compare quite well with those of the circular well \[see Fig. (\[prob\_density1b\])\], except fo the obvious “squareness” of the contours of the higher excited states.[]{data-label="prob_density_sq2"}](square-confine-symm-contour-eps1e-3.pdf) Because of the remaining “practical” degeneracy, the diagonalization routine ends up picking some linear combination of these “degenerate” states and ordering them in some fashion. The arbitrary linear combination easily leads to a wave function that does not have the symmetry of the confining potential, as is evident in Fig. (\[prob\_density\_sq1\]). We will see the same problem emerge in the Landau gauge, which [does not have]{} the symmetry of the confinement potential, so it is prudent to emphasize that this aspect (having or not having a gauge with the symmetry of the confining potential) is not so important. Here, where the gauge choice is the symmetric one, we still encounter this problem, and it is because of the “practical” degeneracy (i.e. energies that are within $<10^{-16}$ of one another) that remains in the solutions. The ‘practical’’ degeneracy remains because, for large values of $G$, there is enough space in the confined potential for the “centre” of the wave function to be arbitrarily located. We have resolved this problem by adding an additional confining potential with the symmetry of the square. Our choice is gauge invariant and is $$H' = \frac{1}{2} m_e \omega_0^2 \left[\left(x-\frac{a}{2}\right)^2+\left(y-\frac{a}{2}\right)^2\right], \label{hamp}$$ where $\omega_0$ is the characteristic frequency to characterize the perturbation potential. It is important to emphasize that this additional potential is minute: typically $(\omega_0/\omega_c)^2 \approx 10^{-6}$ for $G=50$, so that the energies are not really affected at any level the eye can detect, but the degeneracy is lifted sufficiently to allow the diagonalization algorithm to properly and unambiguously order the otherwise quasi-degenerate eigenstates. Most importantly, the centre of all wave functions is now correctly placed at the centre of the square. When this tiny potential is included we obtain the results for the probability densities shown in Fig. (\[prob\_density\_sq2\]), where now the results look correct and more sensible \[the same tiny potential was actually used in Fig. (\[prob\_density1b\])\]. For the sake of completeness, we include the required matrix elements for our current problem, $$\begin{split} {(h')}_{n,m}= \frac{H'_{n,m}}{E_0} = \frac{\pi^2 \epsilon^2}{4} &\Biggl[ \frac{1}{12} \delta_{n_x,m_x}\delta_{n_y,m_y} \left( 1 - \frac{6}{(\pi n_x)^2} + 1 - \frac{6}{(\pi n_y)^2} \right)\\[5pt] &+\frac{2}{\pi^2} \Bigl( \delta_{n_y,m_y} [1-\delta_{n_x,m_x}] g_e(n_x,m_x) + \delta_{n_x,m_x} [1-\delta_{n_y,m_y}] g_e(n_y,m_y) \Bigr) \Biggr] \end{split} \label{ham_sq_pert}$$ where $g_e(n,m)$ is as defined in Eq. (\[ge\_def\]) and $\epsilon \equiv \hbar \omega_0/E_0 = (4G/\pi)(\omega_0/\omega_c)$; with this definition, $\epsilon^2$ provides the relative energy scale of the perturbing potential. [Landau Gauge]{} ---------------- Since the square geometry is rather difficult in either gauge, we thought it worthwhile to provide results for the same geometry in the Landau gauge. That this will work illustrates the remarkable machinery of gauge invariance, as now the gauge choice results in a very non-symmetric looking potential. In fact as we will shortly see it doesn’t work for large values of $G$, for the reasons just described above, but our very slight perturbing confining potential will fix the problem (i.e. lift the degeneracy) so that results identical to the symmetric gauge are achieved. ### [Matrix Element Calculation]{} Now we use the (square well centred) Landau gauge, given by $${\mathbf{A}}'_L=B\left(x-\frac{a}{2}\right)\hat{{\mathbf{y}}} \label{landau_shifted}$$ so the Hamiltonian is $$H=\frac{\textbf{p}^2}{2{m_e}}-i\hbar\omega_c\left(x-\frac{a}{2}\right)\partial_y+\frac{1}{2}{m_e}\omega_c^2\left(x-\frac{a}{2}\right)^2. \label{ham_landau}$$ The Hamiltonian \[units of $\hbar^2\pi^2/(2{m_e}a^2)$\] becomes $$\begin{split} h_{n,m}&=\delta_{n_x,m_x}\delta_{n_y,m_y} \left[ n_x^2+n_y^2 + \frac{G^2}{3} \left( 1 - \frac{6}{(\pi n_x)^2} \right) \right]\\[5pt] &+8 \left(\frac{G}{\pi}\right)^2 (1-\delta_{n_x,m_x})\delta_{n_y,m_y} g_e(n_x,m_x) \\[5pt] &+i\frac{32G}{\pi^3}(1-\delta_{n_x,m_x})(1-\delta_{n_y,m_y}) g_o(n_x,m_x)f_o(n_y,m_y) \label{ham_landau_nm} \end{split}$$ where the functions $g_e$, $g_o$, and $f_o$ were all defined earlier. This is a very different matrix than that generated for the symmetric gauge, and moreover is very asymmetric in $x$ and $y$. And yet we expect to obtain the same results as in that gauge. ### [Results and Discussion]{} [Eigenenergies]{} Here we could plot the eigenvalues generated in the Landau gauge, as in Fig. (\[square\_eigen\]), but we will not, because they are identical. Similarly, the results for the Fock-Darwin spectrum, Fig. (\[square\_fock\_darwin\]), are identical as well. [Probability Densities]{} Without using the perturbing potential, however, the probability density does not agree. These are shown in Fig. (\[prob\_density\_sq\_lan1\]), where significant discrepancies with Fig. (\[prob\_density\_sq1\]) are apparent, particularly for larger values of $G$. ![Contour plots of the probability density, $\|\Psi_{n}(x,y)\|^2$, for various values of $n$, calculated with the Landau gauge, and as ordered by the diagonalization subroutine. The results agree with those in the symmetric gauge, Fig. (\[prob\_density\_sq1\]), only in the first two columns. Once again, the reason for the discrepancy in the remaining columns is the degeneracy that is still present and this is remedied in the same manner as with the symmetric gauge results.[]{data-label="prob_density_sq_lan1"}](square-confine-landau-contour-eps0.pdf) The results of Fig. (\[prob\_density\_sq\_lan1\]) show somewhat random character, particularly for larger values of $G$, in a manner similar to that seen earlier \[Fig. (\[prob\_density\_sq1\])\] in the symmetric gauge, and for the same reason — persistent numerical degeneracy. Nonetheless the addition of the perturbing potential, Eq. (\[hamp\]) with $\epsilon = 10^{-3}$ solves the problem as expected, and results identical (i.e. numerically, to about 5 digit accuracy) to those of Fig. (\[prob\_density\_sq2\]) are attained. These are shown in Fig. (\[prob\_density\_sq\_lan2\]). ![Contour plots of the probability density in the Landau Gauge with the perturbative trap included. While these results are identical to those obtained in the symmetric gauge in Fig. (\[prob\_density\_sq2\]), we have repeated them here to emphasize that this has been achieved with different gauges, and matrix diagonalization on completely different-looking matrices.[]{data-label="prob_density_sq_lan2"}](square-confine-landau-contour-eps1e-3.pdf) [Other Gauge Possibilities]{} ----------------------------- In the following section we present briefly our results on the probability current density. Before doing so, however, we should note that we explored other gauge choices, namely $${\mathbf{A}}^{\prime\prime}_{S}=\frac{B}{2}\left[\left(x-\frac{b_x}{2}\right)\hat{{\mathbf{y}}}-\left(y-\frac{b_y}{2}\right)\hat{{\mathbf{x}}}\right], \label{symm_shiftedb}$$ for the symmetric gauge, and $${\mathbf{A}}^{\prime \prime}_L=B\left(x-\frac{b_x}{2}\right)\hat{{\mathbf{y}}} \label{landau_shiftedb}$$ for the Landau gauge, where $b_x$ and $b_y$ are [arbitrary]{}. In particular we formulated the problem for $b_x = b_y = 0$, knowing that this was putting ourselves at a disadvantage. Nonetheless, even this choice works, but at a cost that much larger matrices are required. This is expected, since now the gauge potential is very asymmetric throughout the square — in either case the parabolic confinement arising from the applied magnetic field is centred at one corner of the square. While we refrain from showing results here, we found identical results as earlier for weak magnetic fields (where we could pursue convergence as a function of matrix size). It is noteworthy that in this regime there is no practical degeneracy and so the perturbation potential, Eq. (\[hamp\]), is not required. [Probability Current]{} ----------------------- The expression for the probability current density is given by Eq. (\[current1\]). Adopting the expansion of the wave function in the square well (\[expand\]) with basis states (\[basis\_sq\]) and arbitrary gauge choice $\mathbf{A}$, the probability current density of the $n$th total quantum state $\mathbf{J}_n$ can be expressed as: $$\begin{aligned} \mathbf{J}_{n}(x,y)&= \frac{\hbar}{m_e}\left[{\rm Im}({\Psi}_n^{\ast} {\mathbf{\Theta}}_n) +\frac{e\mathbf{A}}{\hbar}\|\Psi_n\|^2, \right] \ \ \ \ \text{where} \label{sq_prob_curr_dens}\\[5pt] \Psi_n(x,y)&={2 \over a} \sum_{m_x,m_y=1}^{\infty}c^{(n)}_{m_x,m_y} \sin\left(\frac{m_x\pi x}{a}\right)\sin\left(\frac{m_y\pi y}{a}\right), \ \ \ \ \text{and}\\[5pt] \begin{split} \mathbf{\Theta}_n(x,y)&\equiv \mathbf{\nabla}\Psi_n(x,y)={2 \over a}\sum_{m_x,m_y=1}^{\infty}c^{(n)}_{m_x,m_y}\frac{\pi m_x}{a}\cos\left(\frac{m_x\pi x}{a}\right)\sin\left(\frac{m_y\pi y}{a}\right) \ \mathbf{\hat{x}}\\[5pt] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +{2 \over a}\sum_{m_x,m_y=1}^{\infty}c^{(n)}_{m_x,m_y}\frac{\pi m_y}{a}\sin\left(\frac{m_x\pi x}{a}\right)\cos\left(\frac{m_y\pi y}{a}\right) \ \mathbf{\hat{y}}. \end{split}\end{aligned}$$ Note that the eigenvector coefficient corresponding to the $n^{\rm th}$ eigenvector, $c^{(n)}_{m_x,m_y}$, is in general a complex number. For the Symmetric \[Eq. (\[symm\_shifted\])\] and Landau \[Eq. (\[landau\_shifted\])\] gauge choices, the probability current density is given by $$\mathbf{J}^{(S)}_n(x,y)=\frac{\hbar}{m_e a}\left[a \ {\rm Im}({\Psi}_n^{\ast}{\mathbf{\Theta}}_n)+ \pi G\|{\Psi}_n\|^2\left(\left(\frac{x}{a}-\frac{1}{2}\right)\mathbf{\hat{y}}-\left(\frac{y}{a}-\frac{1}{2}\right)\mathbf{\hat{x}}\right)\right] \label{shifted_symm_prob_curr}$$ for the shifted-symmetric gauge, and $$\mathbf{J}^{(L)}_n(x,y)=\frac{\hbar}{m_e a}\left[a \ {\rm Im}({\Psi}_n^{\ast}{\mathbf{\Theta}}_n)+ 2\pi G\|{\Psi}_n\|^2\left(\frac{x}{a}-\frac{1}{2}\right)\mathbf{\hat{y}} \right] \label{shifted_landau_prob_curr}$$ for the shifted-Landau gauge. While the structure of the equations for the two gauge choices Eqs. (\[shifted\_symm\_prob\_curr\]) and (\[shifted\_landau\_prob\_curr\]) is quite different, the probability current density is a gauge-invariant quantity, and we have confirmed that both results are identical. ![A vector field plot of the probability current density with $G=100$ for a “bulk” state in (a) ($n = 10$) and an edge state in (b) ($n=95$). Further details of the region enclosed by the red square are provided in each case to the right. Convergence was attained with a matrix size of $N=150$ and $\epsilon=10^{-3}$ was used for the perturbing parabolic trap. Although this result was generated with the symmetric gauge, we confirmed that an identical result is found for the Landau gauge. In the first case an inner shell of clockwise-circulating current is followed by a concentric shell of counter-clockwise-circulating current; this is very similar to what was found with circular geometry \[see Fig. (\[circ\_vec\_field\]a)\]. However, the edge state current density is very different. As in the circular case it is [primarily]{} a clockwise-circulating current, but the corners cause a vortex-anti-vortex pair to be created, as is clear from the blow-up on the right.[]{data-label="fig: square_well_vec_field"}](square_confine_vec_field.pdf) Equipped with equations (\[sq\_prob\_curr\_dens\]-\[shifted\_landau\_prob\_curr\]) it is straightforward to evaluate the probability current density for a given eigenstate. Two vector field plots of typical results are shown in Fig. (\[fig: square\_well\_vec\_field\]) (in units of $h/(m_e a^3)$). We see that (a) is a “bulk” state while (b) is an edge state. In fact, the former result is qualitatively indistinguishable from that attained for a circular geometry in Fig. (\[circ\_vec\_field\]), as is apparent from the result. In all cases we have also used a perturbing potential as described earlier, with $\epsilon = 10^{-3}$. In (b) there is a clear difference from the circular case, and the current tends to follow the geometry of the square boundary. While one may suspect from this that edge state currents simply follow the geometry of the boundary, a closer look at the expanded portion in (b) shows very interesting structure. Rather than smoothly following the geometry of the boundary at the corners of the square well, the probability current density appears to form stationary vortices rotating in a direction opposite to that of the nearby principal current density flow. This is a very different scenario than that encountered with the circular geometry, and is reminiscent of the ‘corner modes’ noted recently for topological insulators.[@yan18] Here, these are not zero energy modes, but appear as a simple consequence of the square geometry, and occur at large quantum number. Thus they may in fact be the precursor of semi-classical behaviour (e.g. the skipping orbits described in Ref. \[\]). Further investigation is clearly required. One can also characterize the probability current density at slices through the sample. Defining $J_{x,n}\equiv\mathbf{J}_n\cdot\mathbf{\hat{x}}$, and $J_{y,n}\equiv\mathbf{J}_n\cdot\mathbf{\hat{y}}$, we show $J_{x,n}$ as a function of $y$ (a vertical slice at $x=a/2$) in Fig. (\[fig: x=a/2\_prob\_curr\_slice\]) through the centre of the square. For low quantum numbers the result will resemble that of the circle, shown in Fig. (\[prob\_current2\]). Here we cannot classify the states according to their $\ell$ quantum number, as we did in that case. Also note that Fig. (\[prob\_current2\]) displays results across a radius, i.e. [half]{} the sample, whereas for the square, Fig. (\[fig: x=a/2\_prob\_curr\_slice\]) shows results across [the entire]{} sample, and therefore has an inherent asymmetry, This is because, in spite of the [square]{} geometry, the current is [circulating]{}. For the same reason, $J_{y,n}$ along this vertical slice was found to be zero (within numerical noise). Slices related through symmetry operations yield entirely equivalent results. For example, the $J_{y,n}$ plotted across a horizontal slice at $y=a/2$ looks identical to $J_{x,n}$ at $x=a/2$. ![The $x$-component of the current density $J_{x,n}$ as a function of $y$ taken through $x=a/2$. This represents a vertical slice through the centre of the sample. We show results for various total quantum numbers $n$, as labeled. Results obtained in the Landau (Symmetric) gauge are shown as curves (points), and obviously agree with one another. Here we used $G=100$ and $N=150$. Also note that $J_{y,n}\sim0$ over this slice.[]{data-label="fig: x=a/2_prob_curr_slice"}](square-confine-prob-curr-gauge-compare-x-a-2-xcomponent-G100-N150-eps1e-3.pdf) As seen before in the circular well, the current contribution for a particular state becomes non-zero as we move from the bulk region to the edge region in Fig. (\[fig: x=a/2\_prob\_curr\_slice\]). Moving from $y=a/2$ to $y=a$, we see that the total area under the lower states (with the exception of the ground state) vanishes due to a sign change across a node. States near the edge ($J_{x,95}$, for example) do not experience this nodal sign change, resulting in non-zero area, and thus non-zero current contribution. Also note that the edge state current has opposite sign on opposite edges of the confinement, confirming its inherent chirality.\ ![$J_{y,n}$ as a function of $x$, taken along the diagonal from the upper left of the square down to the lower right. As $x$ varies from $0$ to $a$, $y = a-x$ varies also, from $a$ to $0$. As in Fig. (\[fig: x=a/2\_prob\_curr\_slice\]), results are shown for various quantum numbers and in both gauges. We used $G=100$ and $N=150$. $J_{x,n}$ mirrors this result about the vertical axis. []{data-label="fig: diagonal_prob_curr_slice"}](square-confine-prob-curr-gauge-compare-y-x-ycomponent-G100-N150-eps1e-3.pdf) In Fig. (\[fig: diagonal\_prob\_curr\_slice\]) we show results across another slice, a diagonal across the square from the upper left to the lower right. We show only $J_{y,n}$ but in this case $J_{x,n}$ is simply the mirror of $J_{y,n}$ about the vertical axis. This plot is not so different from the previous one, but can give us a glimpse of the vorticies seen only at the corners of the square. The occurrence of these vortices clearly merits further study. [Summary]{} =========== We have presented an elementary discussion of the issues concerning gauge invariance, first in free space, and, more pertinently, in confined geometries. Practical degeneracies still exist in confined geometries, or quantum dots, in particular where a level “floor” in the potential well exists; this is not the case with a parabolic trap, for example. We used a quantum dot with circular symmetry to illustrate some of these degeneracies. However, in this case it is easy to overlook some of the subtleties, as we conveniently have a good quantum number, $\ell$, that can both simplify the calculation, and can be used to organize the results in an unambiguous manner. In this way, even states whose energy difference cannot be distinguished to numerical precision will nonetheless be ordered properly through our knowledge of these quantum numbers. The circular potential also “begs” for the Symmetric Gauge to be utilized. Doing so simplifies the problem immensely, so only a one-dimensional equation requires solution. We could have used the Landau gauge, but then the problem would have been significantly more difficult. For the case of a two-dimensional square, however, the problem is more difficult right from the start, regardless of which gauge is chosen. Partly for this reason, it became a good testing ground for comparing gauge choices, especially given our method of solution, matrix mechanics. To our knowledge this problem has not been previously solved in either gauge. Choosing a geometry also highlights that a gauge choice also involves a choice of origin for the gauge. Whether we use the Symmetric or Landau gauge, the natural origin is the centre of the sample, but we showed that this is not required. Non-optimal choices (i.e. [not]{} the centre of the sample) generally require more basis states with our method, so there is an additional numerical cost for a non-optimal choice. One can compute any property desired (we showed probability and current densities in this study), since we have the eigenvalues and eigenvectors. A further study will explore the susceptibility and other properties that can readily be measured. Finally, we also suggested a simple way of removing the degeneracy at the $10^{-10}$ level (so the computer could tell the difference). As plotted, the eigenvalues will still appear to be degenerate, and this is the correct physics. However, for purposes of organization it is necessary to have a method to distinguish these from one another, and the (very) shallow gauge-invariant parabolic potential that we proposed does the trick. The results for a square geometry are new; we expected that, for the properties studied in this work, the square geometry would not produce anything qualitatively new beyond the results for the circular quantum dot. Instead, as we saw, the current density for the edge states contains vortex-anti-vortex pairs at the corners, which are completely absent for the circular dot for the same range of quantum numbers. We plan to carry out a more in-depth investigation of the conditions under which such modes appear. [Acknowledgments]{} =================== This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC). These studies were initiated through a number of former students; in particular Sophie Taylor carried out some initial studies that were very helpful to get us started. We also thank Ania Michalik for helpful calculations and discussions. Mason Protter and Joel Hutchinson were instrumental in the initial stages with assistance with numerically technical matters. We also thank Joseph Maciejko and Jorge Hirsch for very helpful discussions on various aspects of this problem.	{ "pile_set_name": "ArXiv" }
--- abstract: \| Verlinde conjectured that gravitation is an emergent entropic force. This surprising conjecture was proved in \[Physica A [505]{} (2018) 190\] within a purely classical context. Here, we appeal to a quantum environment to deal with the conjecture in the case of bosons and consider also the classical limit of quantum mechanics (QM). [Keywords]{}: Gravitation, bosons, entropic force, emergent force, Verlide’s conjecture..\ author: - \| ,\ \ \ \ \ \ \ title: '[Quantum treatment of Verlinde’s entropic force conjecture]{}' --- Introduction ============ Eight years ago, Verlinde [@verlinde] proposed to link gravity with an entropic force. The ensuing conjecture was proved recently [@p1], in a purely classical environment. According to Verlinde, gravity would emerge as a result of information about the positions of material particles, connecting a thermal treatment of gravity to ’t Hooft’s holographic principle. In this perspective, gravitation should be regarded as an emergent phenomenon. This Verlinde’s idea was the focus of much attention. For an example, see [@times; @libro]. An excellent overview on the statistical mechanics of gravitation can be found in Padmanabhan’s article [@india], and references therein. Verlinde’s work originated endeavors in cosmology, the dark energy hypothesis, cosmological acceleration, cosmological inflation, and loop quantum gravity. The associated literature is extensive [@libro]. An important contribution is that of Guseo [@guseo]. He showed that the local entropy function, related to a logistic distribution, is a catenary and vice versa, an invariance that may be interpreted through Verlinde’s conjecture regarding gravity’s origin as an entropic force. Guseo advances a novel interpretation of the local entropy in a system [@guseo]. This paper does not treat any of these issues, though. Based on the fact that we proved Verlinde’s hypothesis in a classical environment [@p1], we wish to deal here with the quantal bosonic scenario. Entropic force for bosons ========================= Quantum entropic force ---------------------- The Bose gas’ entropy is (see, for instance, [@lemons]) $$\label{eq2.1} {\cal S}=Nk_B\left[\left(\frac {n} {N}\right)\ln\left(1+\frac {N} {n}\right)+ \ln\left(1+\frac {n} {N}\right)\right],$$ where $$\label{eq2.2} n=V\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi e m} {2h^2}\right)^{\frac {3} {2}}.$$ Since we can cast the volume as $$\label{eq2.3} V=\frac {4} {3}\pi r^3,$$ we recast (\[eq2.1\]) as $${\cal S}=k_BV\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi em} {2h^2}\right)^{\frac {3} {2}} \ln\left[1+\frac {N} {V}\left(\frac {N} {E}\right)^{\frac {3} {2}} \left(\frac {2h^2} {4\pi em}\right)^{\frac {3} {2}}\right]+$$ $$\label{eq2.4} Nk_B\ln\left[1+\frac {V} {N}\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi em} {2h^2}\right)^{\frac {3} {2}}\right].$$ Now, according to [@p1] the entropic force is $$F_e=-\lambda\frac {\partial S} {\partial A}=$$ $$\frac {\lambda 3k_BN} {8\pi r^2} \frac {1} {1+\frac {3N} {4\pi r^3}\left(\frac {N} {E}\right)^{\frac {3} {2}} \left(\frac {3h^2} {4\pi e m}\right)^{\frac {3} {2}}}-$$ $$\frac {\lambda k_B} {2}\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi e m} {2h^2}\right)^{\frac {3} {2}} r \ln\left[1+\frac {3n} {4\pi r^3}\left(\frac {N} {E}\right)^{\frac {3} {2}} \left(\frac {3h^2} {4\pi e m}\right)^{\frac {3} {2}}\right]-$$ $$\label{eq2.5} \frac {\frac {\lambda k_B} {2}\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi e m} {2h^2}\right)^{\frac {3} {2}} r} {1+\frac {4\pi r^3} {3N}\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi e m} {2h^2}\right)^{\frac {3} {2}}},$$ where $\lambda$ is an arbitrary constant. We can recast the above expression as $$F_e=\frac {12\lambda k_BN\left(\pi emE\right)^{\frac {3} {2}} r} {32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}+ 3^{\frac {5} {2}}N^{\frac {5} {2}}h^3}-$$ $$\frac {4\pi\lambda k_B\left(\pi emE\right)^{\frac {3} {2}}} {\left(3N\right)^{\frac {3} {2}}h^3}r \left\{\ln\left[32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}+ (3N)^{\frac {5} {2}}h^3\right]- \ln\left[32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}\right]\right\}-$$ $$\label{eq2.6} \frac {12\lambda k_BN\left(\pi emE\right)^{\frac {3} {2}} r} {32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}+ (3N)^{\frac {5} {2}}h^3},$$ and then $$\label{eq2.7} F_e=\frac {4\pi\lambda k_B\left(\pi emE\right)^{\frac {3} {2}}} {\left(3N\right)^{\frac {3} {2}}h^3}r \left\{\ln\left[32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}+ (3N)^{\frac {5} {2}}h^3\right]- \ln\left[32\pi r^3\left(\pi emE\right)^{\frac {3} {2}}\right]\right\}.$$ Bosonic entropic force in the classical limit (CL) -------------------------------------------------- The CL is attained for [@lemons] $$\label{eq2.8} \frac {N} {n}<<1,$$ and in this limit the entropy becomes [@lemons] $$\label{eq2.9} {\cal S}=Nk_B\left[1+\ln\left(\frac {n} {N}\right)\right],$$ or $$\label{eq2.10} {\cal S}= \frac {5Nk_B} {2}+Nk_B\ln\left[\frac {V} {N}\left(\frac {E} {N}\right)^{\frac {3} {2}} \left(\frac {4\pi m} {2h^2}\right)^{\frac {3} {2}}\right].$$ Now, we have an entropic force of the form $$\label{eq2.11} F_e=-\lambda\frac {\partial S} {\partial A}= -\frac {\lambda 3Nk_B} {8\pi r^2},$$ which is indeed of the Newton appearance, so that Verlinde’s conjecture gets proved in the classical limit. Note also that the entropic force (\[eq2.11\]) can be derived as well from (\[eq2.7\]) by taking $r$ large enough. ![Here we plot $F_e/C$, $C=\frac {3Nk_B\lambda} {8\pi}$. Green line: Boson entropic force. Violet line: approximate semi-classic Bose-one. Here, on the x axis, 1 unit=$10^{- 37}$ meter and, on the y axis, 1 unit=$10^{-4}$ Newton[]{data-label="fig1"}](completa.eps) In Fig. 1 we set: $m$= the oxygen-molecule’s mass, $E=\frac {Nmv^2} {2}$, $v=1 meter/second$, and $N$ is extracted from $$32\pi(\pi e mE)^{\frac {3} {2}}=(3N)^{\frac {5} {2}}h^3$$, so that $N=2.358458\times 10^{26}$. The classic entropic force (\[eq2.11\]) can also be encountered starting from (\[eq2.7\]) and taking $r$ large enough. Appealing such $r$ values and using the equality $$\label{eq2.12} -\frac {\lambda 3Nk_B} {8\pi r^2}= -\frac {GmM} {r^2},$$ where $G$ is the gravitational constant, we detect the fact that $\lambda=\lambda (m,M,N)$. Entropic Potential Energy ------------------------- The entropic force is proportional to the derivative of the entropy with respect to the spherical area $A$. It is interesting to calculate the corresponding potential energy $E_P$. To such an end we define the new constants $a$ and $b$ in the fashion $a=(3N)^{\frac {5} {2}}h^3$ and $b=32\pi(\pi e mE)^{\frac {3} {2}}$. Using reference [@gra] we can compute the potential energy we are looking for from the entropic force’s expression. The ensuing calculation is simple but lengthy. One arrives at the result $$E_P(r)=\frac {3Nk_B\lambda} {8\pi}\frac {b} {a}\left\{\frac {r^2} {2} \ln\left(1+\frac {a} {br^3}\right)- \frac {a^{\frac {2} {3}}} {2b^{\frac {2} {3}}}\left\{\frac {1} {2}\ln \left[\frac {\left[r+\left(\frac {a} {b}\right)^{\frac {1} {3}}\right]^2} {r^2-\left(\frac {a} {b}\right)^{\frac {1} {3}}r+\left(\frac {a} {b}\right)^{\frac {2} {3}}} \right]\right.\right.+$$ $$\label{eq2.13} \left.\left.\sqrt{3}\left[\frac {\pi} {2}- \arctan\left[\frac{2r-\left(\frac {a} {b}\right)^{\frac {1} {3}}} {\sqrt{3}\left(\frac {a} {b}\right)^{\frac {1} {3}}}\right]\right] \right\}\right\},$$ where we have set $E_P(r)=0$ for $r\rightarrow\infty$. For $r$ large the potential energy adopts the appearance $$\label{eq2.14} E_P(r)= -\frac {\lambda 3Nk_B} {8\pi r},$$ which is consistent with the result (\[eq2.11\]). Conclusions =========== We have here considered Verlinde’s \[entropic force - Gravitation\] link, proved recently in a classical context [@p1], in a quantum bosonic scenario. We have seen that Verlinde’s conjecture holds in this scenario as well. Further, the quantum emergent gravitation à la Verlinde does not diverge at the origin. Such an asymptotic behaviour is rather surprising. One wonders whether this emergent gravitation-behaviour might perhaps be an artifact of not being able to include general relativity effects. Moreover, in two limits - the classical limit of QM - $r \rightarrow \infty$, the Newton $r$-dependence of the gravitation force is recovered. A natural challenge is to attack the fermionic case. This we did in reference [@fer]. For fermions, the vanishing of the entropic force at the origin is also observed. [99]{} E. Verlinde, arXiv:1001.0785 \[hep-th\]; JHEP [04]{} (2011) 29. A. Plastino, M. C. Rocca: Physica A [505]{} (2018) 190. D. Overbye, [A Scientist Takes On Gravity]{}, The New York Times, 12 July 2010; M. Calmthout, New Scientist [205]{} (2010) 6. J. Makela, arXiv:1001.3808v3; J. Lee, arXiv:1005.1347; V. V. Kiselev, S. A. Timofeev , Mod. Phys. Lett. A [25]{} (2010) 2223; T. Aaltonen et al; Mod. Phys. Lett. A [25]{} (2010) 2825. T. Padmanabhan, arXiv 0812.2610v2. R. Guseo, Physica A [464]{} (2016) 1. D. S. Lemons:“A Student’s Guide to Entropy”. Cambridge University Press (2014). I. S. Gradshteyn and I. M. Ryzhik : “Table of Integrals, Series and Products”. Academic Press, Inc (1980). <https://www.researchgate.net/publication/324831972_Spatial_cut-offs_Fermion_Statistics_and_Verlinde's_Conjecture>	{ "pile_set_name": "ArXiv" }
--- abstract: 'Over 100 trigonometric parallaxes and proper motions for masers associated with young, high-mass stars have been measured with the Bar and Spiral Structure Legacy Survey, a Very Long Basline Array key science project, the European VLBI Network, and the Japanese VERA project. These measurements provide strong evidence for the existence of spiral arms in the Milky Way, accurately locating many arm segments and yielding spiral pitch angles ranging from about $7^\circ$ to $20^\circ$. The widths of spiral arms increase with distance from the Galactic center. Fitting axially symmetric models of the Milky Way with the 3-dimensional position and velocity information and conservative priors for the solar and average source peculiar motions, we estimate the distance to the Galactic center, $\Ro$, to be $8.34\pm0.16$ kpc, a circular rotation speed at the Sun, $\To$, to be $240\pm8$ , and a rotation curve that is nearly flat ( a slope of $-0.2\pm0.4$ ) between Galactocentric radii of $\approx5$ and 16 kpc. Assuming a “universal” spiral galaxy form for the rotation curve, we estimate the thin disk scale length to be $2.44\pm0.16$ kpc. With this large data set, the parameters and are no longer highly correlated and are relatively insensitive to different forms of the rotation curve. If one adopts a theoretically motivated prior that high-mass star forming regions are in nearly circular Galactic orbits, we estimate a global solar motion component in the direction of Galactic rotation, $\V=14.6\pm5.0$ . While and $\V$ are significantly correlated, the sum of these parameters is well constrained, $\To+\V = 255.2\pm5.1$ , as is the angular speed of the Sun in its orbit about the Galactic center, $(\To+\V)/\Ro = 30.57\pm0.43$ . These parameters improve the accuracy of estimates of the accelerations of the Sun and the Hulse-Taylor binary pulsar in their Galactic orbits, significantly reducing the uncertainty in tests of gravitational radiation predicted by general relativity.' author: - 'M. J. Reid, K. M. Menten, A. Brunthaler, X. W. Zheng, T. M. Dame, Y. Xu, Y. Wu, B. Zhang, A. Sanna, M. Sato, K. Hachisuka, Y. K. Choi, K. Immer, L. Moscadelli, K. L. J. Rygl, & A. Bartkiewicz' title: \| TRIGONOMETRIC PARALLAXES OF HIGH MASS STAR FORMING REGIONS:\ THE STRUCTURE AND KINEMATICS OF THE MILKY WAY --- Introduction ============ Two major projects to map the spiral structure of the Milky Way are providing parallaxes and proper motions for water and methanol masers associated with high-mass star forming regions (HMSFRs) across large portions of the Milky Way. The Bar and Spiral Structure Legacy (BeSSeL) Survey [^1] and the Japanese VLBI Exploration of Radio Astrometry (VERA) [^2] have yielded over 100 parallax measurements with accuracies typically about $\pm20$ , and some as good as $\pm5$ . This accuracy exceeds the target of the European astrometric satellite mission Gaia, launched in December 2013 and scheduled for final results in 2021-2022 [@Eyer:13]. While Gaia aims to measure $\sim10^9$ stars, far more than practical by Very Long Baseline Interferometry (VLBI), Gaia will be limited by extinction at optical wavelengths and will not be able to freely probe the Galactic plane. In contrast, VLBI at radio wavelengths is not affected by dust extinction and can yield parallaxes for massive young stars that best trace spiral structure in other galaxies, and current parallax accuracy allows measurements for stars across most of the Milky Way. Given parallax and proper motion measurements (coupled with source coordinates and line-of-sight velocities from Doppler shifts of spectral lines), one has complete phase-space information. This provides direct and powerful constraints on the fundamental parameters of the Galaxy, including the distance to the Galactic center, , and the circular orbital speed at the Sun, . Preliminary models of the structure and dynamics of the Galaxy based on VLBI parallax and proper motions of star forming regions have been published. @Reid:09b fitted results from 16 HMSFRs and determined $\Ro=8.4\pm0.6$ kpc and $\To=254\pm16$ , assuming the solar motion in the direction of Galactic rotation, $\V$, is 5 [@Dehnen:98]. More recently @Honma:12 analyzed results from a larger sample of 52 sources, including both low-mass star forming regions and HMSFRs, and concluded that $\Ro=8.05\pm0.45$ kpc and $\To=238\pm14$ , assuming $\V=12$ [@Schoenrich:10]. Several groups have re-modeled maser parallax and proper motion data [@Bovy:09; @McMillan:10; @Bobylev:10] using different approaches and focusing on effects of parameter correlations and prior assumptions, most notably the values adopted for the solar motion (see §\[sect:priors\] and §\[sect:solar\_motion\]). With the much larger number and wider distribution of parallaxes and proper motions of HMSFRs now available, we can provide more robust estimates of the fundamental Galactic parameters. In Section \[sect:parallaxes\], we present the combined parallax data sets from the BeSSeL and VERA groups and comment on aspects of spiral structure in Section \[sect:spiral\_structure\]. We model the combined data set to obtain better estimates of and in Section \[sect:modeling\], including discussion of priors, different forms of rotation curves, and parameter correlations. Finally, in Section \[sect:discussion\], we discuss the solar motion, best values for and , and some astrophysical implications. Parallaxes and Proper Motions {#sect:parallaxes} ============================= Table \[table:parallaxes\] lists the parallaxes and proper motions of 103 regions of high-mass star formation measured with VLBI techniques, using the National Radio Astronomy Observatory’s Very Long Baseline Array (VLBA), the Japanese VERA project, and the European VLBI Network (EVN). We have include three red supergiants (NML Cyg, S Per, VY CMa) as indicative of HMSFRs, since they are high mass stars that have short lifetimes ($<10^7$ yr) and therefore cannot have migrated far from their birth locations. The locations of these star forming regions in the Galaxy are shown in Figure \[fig:parallaxes\], superposed on a schematic diagram of the Milky Way. Distance errors are indicated with error bars ($1\sigma$), but for many sources the error bars are smaller than the symbols. [llrrrrrrll]{} G348.70$-$01.04 & &17:20:04.04 &$-$38:58:30.9 & 0.296$\pm$ 0.026 & $-$0.73$\pm$ 0.19 & $-$2.83$\pm$ 0.54 & $-$7$\pm$ 6 &... &1\ G351.44$+$00.65 &NGC 6334 &17:20:54.60 &$-$35:45:08.6 & 0.744$\pm$ 0.074 & 0.40$\pm$ 0.51 & $-$2.24$\pm$ 0.64 & $-$8$\pm$ 3 &Sgr &2\ G000.67$-$00.03 &Sgr B2 &17:47:20.00 &$-$28:22:40.0 & 0.129$\pm$ 0.012 & $-$0.78$\pm$ 0.40 & $-$4.26$\pm$ 0.40 & 62$\pm$ 5 &... &3\ G005.88$-$00.39 & &18:00:30.31 &$-$24:04:04.5 & 0.334$\pm$ 0.020 & 0.18$\pm$ 0.34 & $-$2.26$\pm$ 0.34 & 9$\pm$ 3 &Sct &4\ G009.62$+$00.19 & &18:06:14.66 &$-$20:31:31.7 & 0.194$\pm$ 0.023 & $-$0.58$\pm$ 0.13 & $-$2.49$\pm$ 0.29 & 2$\pm$ 3 &4$-$k &5\ G010.47$+$00.02 & &18:08:38.23 &$-$19:51:50.3 & 0.117$\pm$ 0.008 & $-$3.86$\pm$ 0.19 & $-$6.40$\pm$ 0.14 & 69$\pm$ 5 &Con &7\ G010.62$-$00.38 &W 31 &18:10:28.55 &$-$19:55:48.6 & 0.202$\pm$ 0.019 & $-$0.37$\pm$ 0.50 & $-$0.60$\pm$ 0.25 & $-$3$\pm$ 5 &3$-$k &7\ G011.49$-$01.48 & &18:16:22.13 &$-$19:41:27.2 & 0.800$\pm$ 0.033 & 1.42$\pm$ 0.52 & $-$0.60$\pm$ 0.65 & 11$\pm$ 3 &Sgr &2\ G011.91$-$00.61 & &18:13:58.12 &$-$18:54:20.3 & 0.297$\pm$ 0.031 & 0.66$\pm$ 0.28 & $-$1.36$\pm$ 0.41 & 37$\pm$ 5 &Sct &4\ G012.02$-$00.03 & &18:12:01.84 &$-$18:31:55.8 & 0.106$\pm$ 0.008 & $-$4.11$\pm$ 0.07 & $-$7.76$\pm$ 0.27 & 108$\pm$ 5 &3$-$k &7\ G012.68$-$00.18 & &18:13:54.75 &$-$18:01:46.6 & 0.416$\pm$ 0.028 & $-$1.00$\pm$ 0.95 & $-$2.85$\pm$ 0.95 & 58$\pm$ 10 &Sct &8\ G012.80$-$00.20 & &18:14:14.23 &$-$17:55:40.5 & 0.343$\pm$ 0.037 & $-$0.60$\pm$ 0.70 & $-$0.99$\pm$ 0.70 & 34$\pm$ 5 &Sct &8\ G012.88$+$00.48 &IRAS 18089$-$1732&18:11:51.42 &$-$17:31:29.0 & 0.400$\pm$ 0.040 & 0.15$\pm$ 0.25 & $-$2.30$\pm$ 0.39 & 31$\pm$ 7 &Sct &8,10\ G012.90$-$00.24 & &18:14:34.42 &$-$17:51:51.9 & 0.408$\pm$ 0.025 & 0.19$\pm$ 0.80 & $-$2.52$\pm$ 0.80 & 36$\pm$ 10 &Sct &8\ G012.90$-$00.26 & &18:14:39.57 &$-$17:52:00.4 & 0.396$\pm$ 0.032 & $-$0.36$\pm$ 0.80 & $-$2.22$\pm$ 0.80 & 39$\pm$ 10 &Sct &8\ G013.87$+$00.28 & &18:14:35.83 &$-$16:45:35.9 & 0.254$\pm$ 0.024 & $-$0.25$\pm$ 2.00 & $-$2.49$\pm$ 2.00 & 48$\pm$ 10 &Sct &4\ G014.33$-$00.64 & &18:18:54.67 &$-$16:47:50.3 & 0.893$\pm$ 0.101 & 0.95$\pm$ 1.50 & $-$2.40$\pm$ 1.30 & 22$\pm$ 5 &Sgr &9\ G014.63$-$00.57 & &18:19:15.54 &$-$16:29:45.8 & 0.546$\pm$ 0.022 & 0.22$\pm$ 1.20 & $-$2.07$\pm$ 1.20 & 19$\pm$ 5 &Sgr &2\ G015.03$-$00.67 &M 17 &18:20:24.81 &$-$16:11:35.3 & 0.505$\pm$ 0.033 & 0.68$\pm$ 0.32 & $-$1.42$\pm$ 0.33 & 22$\pm$ 3 &Sgr &10\ G016.58$-$00.05 & &18:21:09.08 &$-$14:31:48.8 & 0.279$\pm$ 0.023 & $-$2.52$\pm$ 0.37 & $-$2.33$\pm$ 0.35 & 60$\pm$ 5 &Sct &4\ G023.00$-$00.41 & &18:34:40.20 &$-$09:00:37.0 & 0.218$\pm$ 0.017 & $-$1.72$\pm$ 0.14 & $-$4.12$\pm$ 0.33 & 80$\pm$ 3 &4$-$k &11\ G023.44$-$00.18 & &18:34:39.19 &$-$08:31:25.4 & 0.170$\pm$ 0.032 & $-$1.93$\pm$ 0.15 & $-$4.11$\pm$ 0.13 & 97$\pm$ 3 &4$-$k &11\ G023.65$-$00.12 & &18:34:51.59 &$-$08:18:21.4 & 0.313$\pm$ 0.039 & $-$1.32$\pm$ 0.20 & $-$2.96$\pm$ 0.20 & 83$\pm$ 3 &... &12\ G023.70$-$00.19 & &18:35:12.36 &$-$08:17:39.5 & 0.161$\pm$ 0.024 & $-$3.17$\pm$ 0.12 & $-$6.38$\pm$ 0.16 & 73$\pm$ 5 &4$-$k &7\ G025.70$+$00.04 & &18:38:03.14 &$-$06:24:15.5 & 0.098$\pm$ 0.029 & $-$2.89$\pm$ 0.07 & $-$6.20$\pm$ 0.36 & 93$\pm$ 5 &Sct &4\ G027.36$-$00.16 & &18:41:51.06 &$-$05:01:43.4 & 0.125$\pm$ 0.042 & $-$1.81$\pm$ 0.11 & $-$4.11$\pm$ 0.27 & 92$\pm$ 3 &Sct &10\ G028.86$+$00.06 & &18:43:46.22 &$-$03:35:29.6 & 0.135$\pm$ 0.018 & $-$4.80$\pm$ 0.30 & $-$5.90$\pm$ 0.30 & 100$\pm$ 10 &Sct &4\ G029.86$-$00.04 & &18:45:59.57 &$-$02:45:06.7 & 0.161$\pm$ 0.020 & $-$2.32$\pm$ 0.11 & $-$5.29$\pm$ 0.16 & 100$\pm$ 3 &Sct &6\ G029.95$-$00.01 &W 43S &18:46:03.74 &$-$02:39:22.3 & 0.190$\pm$ 0.019 & $-$2.30$\pm$ 0.13 & $-$5.34$\pm$ 0.13 & 98$\pm$ 3 &Sct &6\ G031.28$+$00.06 & &18:48:12.39 &$-$01:26:30.7 & 0.234$\pm$ 0.039 & $-$2.09$\pm$ 0.16 & $-$4.37$\pm$ 0.21 & 109$\pm$ 3 &Sct &6\ G031.58$+$00.07 &W 43Main &18:48:41.68 &$-$01:09:59.0 & 0.204$\pm$ 0.030 & $-$1.88$\pm$ 0.40 & $-$4.84$\pm$ 0.40 & 96$\pm$ 5 &Sct &6\ G032.04$+$00.05 & &18:49:36.58 &$-$00:45:46.9 & 0.193$\pm$ 0.008 & $-$2.21$\pm$ 0.40 & $-$4.80$\pm$ 0.40 & 97$\pm$ 5 &Sct &4\ G033.64$-$00.22 & &18:53:32.56 &$+$00:31:39.1 & 0.153$\pm$ 0.017 & $-$3.18$\pm$ 0.10 & $-$6.10$\pm$ 0.10 & 60$\pm$ 3 &... &1\ G034.39$+$00.22 & &18:53:18.77 &$+$01:24:08.8 & 0.643$\pm$ 0.049 & $-$0.90$\pm$ 1.00 & $-$2.75$\pm$ 2.00 & 57$\pm$ 5 &Sgr &13\ G035.02$+$00.34 & &18:54:00.67 &$+$02:01:19.2 & 0.430$\pm$ 0.040 & $-$0.92$\pm$ 0.90 & $-$3.61$\pm$ 0.90 & 52$\pm$ 5 &Sgr &2\ G035.19$-$00.74 & &18:58:13.05 &$+$01:40:35.7 & 0.456$\pm$ 0.045 & $-$0.18$\pm$ 0.50 & $-$3.63$\pm$ 0.50 & 30$\pm$ 7 &Sgr &14\ G035.20$-$01.73 & &19:01:45.54 &$+$01:13:32.5 & 0.306$\pm$ 0.045 & $-$0.71$\pm$ 0.21 & $-$3.61$\pm$ 0.26 & 42$\pm$ 3 &Sgr &14\ G037.43$+$01.51 & &18:54:14.35 &$+$04:41:41.7 & 0.532$\pm$ 0.021 & $-$0.45$\pm$ 0.35 & $-$3.69$\pm$ 0.39 & 41$\pm$ 3 &Sgr &2\ G043.16$+$00.01 &W 49N &19:10:13.41 &$+$09:06:12.8 & 0.090$\pm$ 0.007 & $-$2.88$\pm$ 0.20 & $-$5.41$\pm$ 0.20 & 10$\pm$ 5 &Per &15\ G043.79$-$00.12 &OH 43.8$-$0.1 &19:11:53.99 &$+$09:35:50.3 & 0.166$\pm$ 0.005 & $-$3.02$\pm$ 0.36 & $-$6.20$\pm$ 0.36 & 44$\pm$ 10 &Sgr &2\ G043.89$-$00.78 & &19:14:26.39 &$+$09:22:36.5 & 0.121$\pm$ 0.020 & $-$2.75$\pm$ 0.30 & $-$6.43$\pm$ 0.30 & 54$\pm$ 5 &Sgr &2\ G045.07$+$00.13 & &19:13:22.04 &$+$10:50:53.3 & 0.125$\pm$ 0.005 & $-$2.98$\pm$ 0.45 & $-$6.26$\pm$ 0.45 & 59$\pm$ 5 &Sgr &2\ G045.45$+$00.05 & &19:14:21.27 &$+$11:09:15.9 & 0.119$\pm$ 0.017 & $-$2.34$\pm$ 0.38 & $-$6.00$\pm$ 0.54 & 55$\pm$ 7 &Sgr &2\ G048.60$+$00.02 & &19:20:31.18 &$+$13:55:25.2 & 0.093$\pm$ 0.005 & $-$2.89$\pm$ 0.13 & $-$5.50$\pm$ 0.13 & 18$\pm$ 5 &Per &15\ G049.19$-$00.33 & &19:22:57.77 &$+$14:16:10.0 & 0.189$\pm$ 0.007 & $-$2.99$\pm$ 0.40 & $-$5.71$\pm$ 0.40 & 67$\pm$ 5 &Sgr &2\ G049.48$-$00.36 &W 51 IRS2 &19:23:39.82 &$+$14:31:05.0 & 0.195$\pm$ 0.071 & $-$2.49$\pm$ 0.14 & $-$5.51$\pm$ 0.16 & 56$\pm$ 3 &Sgr &16\ G049.48$-$00.38 &W 51M &19:23:43.87 &$+$14:30:29.5 & 0.185$\pm$ 0.010 & $-$2.64$\pm$ 0.20 & $-$5.11$\pm$ 0.20 & 58$\pm$ 4 &Sgr &17\ G052.10$+$01.04 &IRAS 19213+1723&19:23:37.32 &$+$17:29:10.5 & 0.251$\pm$ 0.060 & $-$2.60$\pm$ 2.00 & $-$6.10$\pm$ 2.00 & 42$\pm$ 5 &Sgr &18\ G059.78$+$00.06 & &19:43:11.25 &$+$23:44:03.3 & 0.463$\pm$ 0.020 & $-$1.65$\pm$ 0.30 & $-$5.12$\pm$ 0.30 & 25$\pm$ 3 &Loc &16\ G069.54$-$00.97 &ON 1 &20:10:09.07 &$+$31:31:36.0 & 0.406$\pm$ 0.013 & $-$3.19$\pm$ 0.40 & $-$5.22$\pm$ 0.40 & 12$\pm$ 5 &Loc &19,20,21\ G074.03$-$01.71 & &20:25:07.11 &$+$34:49:57.6 & 0.629$\pm$ 0.017 & $-$3.79$\pm$ 1.30 & $-$4.88$\pm$ 1.50 & 5$\pm$ 5 &Loc &21\ G075.29$+$01.32 & &20:16:16.01 &$+$37:35:45.8 & 0.108$\pm$ 0.005 & $-$2.37$\pm$ 0.11 & $-$4.48$\pm$ 0.17 & $-$58$\pm$ 5 &Out &22\ G075.76$+$00.33 & &20:21:41.09 &$+$37:25:29.3 & 0.285$\pm$ 0.022 & $-$3.08$\pm$ 0.60 & $-$4.56$\pm$ 0.60 & $-$9$\pm$ 9 &Loc &21\ G075.78$+$00.34 &ON 2N &20:21:44.01 &$+$37:26:37.5 & 0.261$\pm$ 0.030 & $-$2.79$\pm$ 0.55 & $-$4.66$\pm$ 0.55 & 1$\pm$ 5 &Loc &23\ G076.38$-$00.61 & &20:27:25.48 &$+$37:22:48.5 & 0.770$\pm$ 0.053 & $-$3.73$\pm$ 3.00 & $-$3.84$\pm$ 3.00 & $-$2$\pm$ 5 &Loc &21\ G078.12$+$03.63 &IRAS 20126+4104&20:14:26.07 &$+$41:13:32.7 & 0.610$\pm$ 0.030 & $-$2.06$\pm$ 0.50 & 0.98$\pm$ 0.50 & $-$4$\pm$ 5 &Loc &24\ G078.88$+$00.70 &AFGL 2591 &20:29:24.82 &$+$40:11:19.6 & 0.300$\pm$ 0.024 & $-$1.20$\pm$ 0.72 & $-$4.80$\pm$ 0.66 & $-$6$\pm$ 7 &Loc &25\ G079.73$+$00.99 &IRAS 20290+4052&20:30:50.67 &$+$41:02:27.5 & 0.737$\pm$ 0.062 & $-$2.84$\pm$ 0.50 & $-$4.14$\pm$ 0.70 & $-$3$\pm$ 5 &Loc &25\ G079.87$+$01.17 & &20:30:29.14 &$+$41:15:53.6 & 0.620$\pm$ 0.027 & $-$3.23$\pm$ 1.30 & $-$5.19$\pm$ 1.30 & $-$5$\pm$ 10 &Loc &21\ G080.79$-$01.92 &NML Cyg &20:46:25.54 &$+$40:06:59.4 & 0.620$\pm$ 0.047 & $-$1.55$\pm$ 0.57 & $-$4.59$\pm$ 0.57 & $-$3$\pm$ 3 &Loc &26\ G080.86$+$00.38 &DR 20 &20:37:00.96 &$+$41:34:55.7 & 0.687$\pm$ 0.038 & $-$3.29$\pm$ 0.45 & $-$4.83$\pm$ 0.50 & $-$3$\pm$ 5 &Loc &25\ G081.75$+$00.59 &DR 21 &20:39:01.99 &$+$42:24:59.3 & 0.666$\pm$ 0.035 & $-$2.84$\pm$ 0.45 & $-$3.80$\pm$ 0.47 & $-$3$\pm$ 3 &Loc &25\ G081.87$+$00.78 &W 75N &20:38:36.43 &$+$42:37:34.8 & 0.772$\pm$ 0.042 & $-$1.97$\pm$ 0.50 & $-$4.16$\pm$ 0.51 & 7$\pm$ 3 &Loc &25\ G090.21$+$02.32 & &21:02:22.70 &$+$50:03:08.3 & 1.483$\pm$ 0.038 & $-$0.67$\pm$ 1.56 & $-$0.90$\pm$ 1.67 & $-$3$\pm$ 5 &Loc &21\ G092.67$+$03.07 & &21:09:21.73 &$+$52:22:37.1 & 0.613$\pm$ 0.020 & $-$0.69$\pm$ 0.60 & $-$2.25$\pm$ 0.60 & $-$5$\pm$ 10 &Loc &21\ G094.60$-$01.79 &AFGL 2789 &21:39:58.27 &$+$50:14:21.0 & 0.280$\pm$ 0.030 & $-$2.30$\pm$ 0.60 & $-$3.80$\pm$ 0.60 & $-$46$\pm$ 5 &Per &18,28\ G095.29$-$00.93 & &21:39:40.51 &$+$51:20:32.8 & 0.205$\pm$ 0.015 & $-$2.75$\pm$ 0.20 & $-$2.75$\pm$ 0.25 & $-$38$\pm$ 5 &Per &28\ G097.53$+$03.18 & &21:32:12.43 &$+$55:53:49.7 & 0.133$\pm$ 0.017 & $-$2.94$\pm$ 0.29 & $-$2.48$\pm$ 0.29 & $-$73$\pm$ 5 &Out &27\ G100.37$-$03.57 & &22:16:10.37 &$+$52:21:34.1 & 0.291$\pm$ 0.010 & $-$3.77$\pm$ 0.60 & $-$3.12$\pm$ 0.60 & $-$37$\pm$ 10 &Per &28\ G105.41$+$09.87 & &21:43:06.48 &$+$66:06:55.3 & 1.129$\pm$ 0.063 & $-$0.21$\pm$ 1.20 & $-$5.49$\pm$ 1.20 & $-$10$\pm$ 5 &Loc &21\ G107.29$+$05.63 &IRAS 22198+6336&22:21:26.73 &$+$63:51:37.9 & 1.288$\pm$ 0.107 & $-$2.47$\pm$ 1.40 & 0.26$\pm$ 1.40 & $-$11$\pm$ 5 &Loc &29\ G108.18$+$05.51 &L 1206 &22:28:51.41 &$+$64:13:41.3 & 1.289$\pm$ 0.153 & 0.27$\pm$ 0.50 & $-$1.40$\pm$ 1.95 & $-$11$\pm$ 3 &Loc &19\ G108.20$+$00.58 & &22:49:31.48 &$+$59:55:42.0 & 0.229$\pm$ 0.028 & $-$2.25$\pm$ 0.50 & $-$1.00$\pm$ 0.50 & $-$49$\pm$ 5 &Per &28\ G108.47$-$02.81 & &23:02:32.08 &$+$56:57:51.4 & 0.309$\pm$ 0.010 & $-$2.45$\pm$ 1.00 & $-$3.00$\pm$ 0.70 & $-$54$\pm$ 5 &Per &28\ G108.59$+$00.49 & &22:52:38.30 &$+$60:00:52.0 & 0.398$\pm$ 0.031 & $-$5.55$\pm$ 0.40 & $-$3.38$\pm$ 0.40 & $-$52$\pm$ 5 &Per &28\ G109.87$+$02.11 &Cep A &22:56:18.10 &$+$62:01:49.5 & 1.430$\pm$ 0.080 & 0.50$\pm$ 1.50 & $-$3.70$\pm$ 1.00 & $-$7$\pm$ 5 &Loc &30\ G111.23$-$01.23 & &23:17:20.79 &$+$59:28:47.0 & 0.288$\pm$ 0.044 & $-$4.28$\pm$ 0.60 & $-$2.33$\pm$ 0.60 & $-$53$\pm$ 10 &Per &28\ G111.25$-$00.76 & &23:16:10.36 &$+$59:55:28.5 & 0.294$\pm$ 0.016 & $-$2.45$\pm$ 0.60 & $-$2.10$\pm$ 0.60 & $-$43$\pm$ 5 &Per &28\ G111.54$+$00.77 &NGC 7538 &23:13:45.36 &$+$61:28:10.6 & 0.378$\pm$ 0.017 & $-$2.45$\pm$ 0.24 & $-$2.44$\pm$ 0.25 & $-$57$\pm$ 5 &Per &30\ G121.29$+$00.65 &L 1287 &00:36:47.35 &$+$63:29:02.2 & 1.077$\pm$ 0.039 & $-$0.86$\pm$ 0.76 & $-$2.29$\pm$ 0.82 & $-$23$\pm$ 5 &Loc &19\ G122.01$-$07.08 &IRAS 00420+5530&00:44:58.40 &$+$55:46:47.6 & 0.460$\pm$ 0.020 & $-$3.70$\pm$ 0.50 & $-$1.25$\pm$ 0.50 & $-$50$\pm$ 5 &Per &31\ G123.06$-$06.30 &NGC 281 &00:52:24.70 &$+$56:33:50.5 & 0.355$\pm$ 0.030 & $-$2.79$\pm$ 0.62 & $-$2.14$\pm$ 0.70 & $-$30$\pm$ 5 &Per &32\ G123.06$-$06.30 &NGC 281W &00:52:24.20 &$+$56:33:43.2 & 0.421$\pm$ 0.022 & $-$2.69$\pm$ 0.31 & $-$1.77$\pm$ 0.29 & $-$29$\pm$ 3 &Per &19\ G133.94$+$01.06 &W 3OH &02:27:03.82 &$+$61:52:25.2 & 0.512$\pm$ 0.010 & $-$1.20$\pm$ 0.32 & $-$0.15$\pm$ 0.32 & $-$47$\pm$ 3 &Per &33,34\ G134.62$-$02.19 &S Per &02:22:51.71 &$+$58:35:11.4 & 0.413$\pm$ 0.017 & $-$0.49$\pm$ 0.35 & $-$1.19$\pm$ 0.33 & $-$39$\pm$ 5 &Per &35\ G135.27$+$02.79 &WB 89$-$437 &02:43:28.57 &$+$62:57:08.4 & 0.167$\pm$ 0.011 & $-$1.22$\pm$ 0.30 & 0.46$\pm$ 0.36 & $-$72$\pm$ 3 &Out &36\ G160.14$+$03.15 & &05:01:40.24 &$+$47:07:19.0 & 0.244$\pm$ 0.006 & 0.87$\pm$ 0.35 & $-$1.32$\pm$ 0.29 & $-$18$\pm$ 5 &... &1\ G168.06$+$00.82 &IRAS 05137+3919&05:17:13.74 &$+$39:22:19.9 & 0.130$\pm$ 0.040 & 0.50$\pm$ 0.24 & $-$0.85$\pm$ 0.17 & $-$27$\pm$ 5 &Out &37,38\ G176.51$+$00.20 & &05:37:52.14 &$+$32:00:03.9 & 1.038$\pm$ 0.021 & 1.84$\pm$ 1.00 & $-$5.86$\pm$ 1.00 & $-$17$\pm$ 5 &Loc &21\ G182.67$-$03.26 & &05:39:28.42 &$+$24:56:32.1 & 0.149$\pm$ 0.011 & 0.16$\pm$ 0.32 & $-$0.17$\pm$ 0.32 & $-$7$\pm$ 10 &Out &37\ G183.72$-$03.66 & &05:40:24.23 &$+$23:50:54.7 & 0.570$\pm$ 0.013 & 0.13$\pm$ 1.20 & $-$1.40$\pm$ 1.20 & 3$\pm$ 5 &Per &28\ G188.79$+$01.03 &IRAS 06061+2151&06:09:06.97 &$+$21:50:41.4 & 0.496$\pm$ 0.103 & $-$0.10$\pm$ 0.50 & $-$3.91$\pm$ 0.50 & $-$5$\pm$ 5 &Per &39\ G188.94$+$00.88 &S 252 &06:08:53.35 &$+$21:38:28.7 & 0.476$\pm$ 0.006 & 0.02$\pm$ 0.30 & $-$2.02$\pm$ 0.30 & 8$\pm$ 5 &Per &18,40\ G192.16$-$03.81 & &05:58:13.53 &$+$16:31:58.9 & 0.660$\pm$ 0.040 & 0.70$\pm$ 0.78 & $-$1.80$\pm$ 0.86 & 5$\pm$ 5 &Per &41\ G192.60$-$00.04 &S 255 &06:12:54.02 &$+$17:59:23.3 & 0.628$\pm$ 0.027 & $-$0.14$\pm$ 0.67 & $-$0.84$\pm$ 1.80 & 6$\pm$ 5 &Per &19\ G196.45$-$01.67 &S 269 &06:14:37.08 &$+$13:49:36.7 & 0.189$\pm$ 0.012 & $-$0.42$\pm$ 0.20 & $-$0.12$\pm$ 0.20 & 19$\pm$ 5 &Out &42\ G209.00$-$19.38 &Orion Nebula &05:35:15.80 &$-$05:23:14.1 & 2.410$\pm$ 0.030 & 3.30$\pm$ 1.50 & 0.10$\pm$ 1.50 & 3$\pm$ 5 &Loc &43,44,45\ G211.59$+$01.05 & &06:52:45.32 &$+$01:40:23.1 & 0.228$\pm$ 0.007 & $-$0.93$\pm$ 0.24 & 0.71$\pm$ 0.26 & 45$\pm$ 5 &... &1\ G229.57$+$00.15 & &07:23:01.84 &$-$14:41:32.8 & 0.221$\pm$ 0.014 & $-$1.34$\pm$ 0.70 & 0.81$\pm$ 0.70 & 47$\pm$ 10 &Per &28\ G232.62$+$00.99 & &07:32:09.78 &$-$16:58:12.8 & 0.596$\pm$ 0.035 & $-$2.17$\pm$ 0.38 & 2.09$\pm$ 0.60 & 21$\pm$ 3 &Loc &40\ G236.81$+$01.98 & &07:44:28.24 &$-$20:08:30.2 & 0.298$\pm$ 0.018 & $-$3.10$\pm$ 0.63 & 2.12$\pm$ 0.63 & 43$\pm$ 7 &Per &28\ G239.35$-$05.06 &VY CMa &07:22:58.33 &$-$25:46:03.1 & 0.855$\pm$ 0.057 & $-$2.80$\pm$ 0.58 & 2.60$\pm$ 0.58 & 20$\pm$ 3 &Loc &46,47\ G240.31$+$00.07 & &07:44:51.92 &$-$24:07:41.5 & 0.212$\pm$ 0.021 & $-$2.36$\pm$ 0.23 & 2.45$\pm$ 0.30 & 67$\pm$ 5 &Per &28\ \[table:parallaxes\] Both the proper motion, $\mu_x$ and $\mu_y$, and Local Standard of Rest (LSR) velocity, , values and their uncertainties are meant to apply to the central star (or stars) that excite the masers. (Note that “LSR velocities” are [defined]{} based on the Standard Solar Motion values of 20 toward $18^h$ Right Ascension and $30^\circ$ Declination in 1900 coordinates, which translate to Galactic cartesian components of $\Uo=10$, $\Vo=15$ and $\Wo=7$ .) For the values we adopted methanol maser values, when available, or CO emission values from associated giant molecular clouds. Since some of the references reporting parallax and proper motion present only measurement uncertainty, for these we estimated an additional error term associated with the uncertainty in transferring the maser motions to that of the central star. These were added in quadrature with the measurement uncertainties. For methanol masers, which typically have modest motions of $\lax10$ with respect to the central star, we estimated the additional error term to be $\pm5$ for and a corresponding value for the proper motion components at the measured distance. While some water masers have expansion motions comparable to methanol masers, others display much faster outflow motions. High velocity outflows are usually associated with water masers that have spectra rich in features, spread over many tens of . We, therefore, evaluated the richness and spread of the water spectra (with respect to the systemic velocity as indicated by CO emission) and assigned the additional error term for $\mu_x$ and $\mu_y$ values between 5 and 20 . Spiral Structure {#sect:spiral_structure} ================ Spiral arms in the Milky Way have long been recognized as presenting coherent arcs and loops in Galactic longitude–velocity ($\ell-V$) plots of atomic and molecular emissions. However, transforming velocity to distance (kinematic distances) has been problematic, owing to near-far distance ambiguities in the first and fourth Galactic quadrants and significant distance errors owing to large peculiar motions for some arm material (see, @Xu:06 [@Reid:09b]). While one cannot accurately place spiral arms on a plan view of the Milky Way from $\ell-V$ plots, one can in most cases unambiguously assign HMSFRs to spiral arms by association with CO and emission features. We have done this for the vast majority of the HMSFRs for which parallax and proper motions have been measured [@Hachi:13; @Choi:13; @Zhang:13; @Xu:13; @Wu:13; @Sato:13; @Sanna:13], as indicated in Table \[table:parallaxes\] and Figure \[fig:parallaxes\]. This avoids using the measured distances (parallaxes) and subjective judgment based on spatial location for arm assignments. There are two avenues for checking that the arm assignments are reliable. Firstly, and most straightforwardly, looking at a plan view of the Milky Way (see Fig. \[fig:parallaxes\]) on which star forming regions with parallax distances are located, one can see that the pattern of sources for any given arm traces a continuous arc that resembles a spiral arm in external galaxies. Also, there are clear inter-arm regions with few, if any, HMSFRs between the Outer, Perseus, Local, Sagittarius, and Scutum arms. However, as one looks to the inner Galaxy, the current parallax data are not adequate to clearly separate arms, presuming significant separations even exist. Secondly, once sources are assigned to arms based on $\ell-V$ information, one can then attempt to fit their radial and azimuthal locations to log-periodic spiral forms using measured distances. In the papers cited above, we fitted spiral patterns to arm segments, adopting a log-periodic spiral defined by $$\ln{(R/R_{ref})} = -(\beta - \beta_{ref}) \tan{\pa}~~,$$ where $R$ is the Galactocentric radius at a Galactocentric azimuth $\beta$ (defined as 0 toward the Sun and increasing with Galactic longitude) for an arm with a radius $R_{ref}$ at reference azimuth $\beta_{ref}$ and pitch angle $\pa$. We fitted a straight line to ($x,y$)=($\beta,\ln{(R/R_{ref})}$) using a Bayesian Markov chain Monte Carlo (McMC) procedure to estimate the parameters $R_{ref}$ and $\pa$. (The reference azimuth, $\beta_{ref}$, was arbitrarily set near the midpoint of the azimuth values for the sources in an arm). We minimized the “distance” perpendicular to the fitted straight line by rotating ($x,y$) through the angle $\pa$ to ($x_r,y_r$), $$x_r = x~\cos{\pa} + y~\sin{\pa};~~~y_r = y~\cos{\pa} - x~\sin{\pa}~~,$$ such that the best-fitting line lay in the $x_r$ axis. Uncertainties in the source parallax “map” into both coordinates and were estimated numerically by randomly drawing trial parallax values (consistent with the measured values and uncertainties) and calculating the root-mean-squares for trial $\ln{(R/R_{ref})}$ and $\beta$ values. The locations of the HMSFRs deviated from fitted spirals by more than could be explained by parallax uncertainties. This is expected for spiral arms with intrinsic widths of several hundred parsecs. In order to allow for (and estimate) the scatter in location expected from the width of the spiral arm, before calculating trial $\ln{(R/R_{ref})}$ values, we added random scatter to the trial $R$ values via $R \leftarrow R + g a_w \cos{\pa}$, where $g$ is a random number drawn from a Gaussian distribution with zero mean and unity standard deviation and $a_w$ is an arm-width parameter, adjusted to give a post-fit $\chi^2_\nu$ near unity. The uncertainties in ($\beta,\ln{(R/R_{ref})}$) were then rotated by angle $\pa$ to match the data. The sum of the squares of the residuals divided by their uncertainties in the $y_r$ direction were minimized. Since preliminary estimates of $\pa$ affect these quantities, we iterated the fitting to convergence. Final parameter values were estimated from marginalized posteriori probability density distribution functions (PDFs) for each parameter based on McMC trials that were accepted or rejected following the Metropolis-Hastings algorithm; the values reported in Table \[table:pitchangles\] assume $\Ro=8.34$ kpc (see §\[sect:modeling\]). Based on the fitted parameter values, we plot the trace of the centers and $1\sigma$ widths of each arm on Fig. \[fig:parallaxes\]. The intrinsic widths of the spiral arms, estimated from the $a_w$ parameters, show an interesting pattern in Fig. \[fig:armwidths\]. The estimated arm widths increase nearly linearly with Galactocentric radius at a rate of 42 pc kpc$^{-1}$ between radii of 5 to 13 kpc. Spiral pitch angles vary between $7^\circ$ and $20^\circ$ as listed in Table \[table:pitchangles\]. The significant range of pitch angles among arms suggests that no single value applies to all arms and, possibly, cannot be applied to the full length of an arm as it winds around the Galaxy [@Savchenko:13]. However, these pitch angles are characteristic of spiral galaxies of Sb to Sc class [@Kennicutt:81], further supporting the identification of $\ell-V$ tracks as spiral arms for the Milky Way. [lrlrrr]{} Scutum &17 &27.6 ($+3\rightarrow101$) &$5.0\pm0.1$ & $0.17\pm0.02$ &$19.8\pm2.6$\ Sagittarius &18 &25.6 ($-2\rightarrow68$) &$6.6\pm0.1$ & $0.26\pm0.02$ &$6.9\pm1.6$\ Local &25 &8.9 ($-8\rightarrow27$) &$8.4\pm0.1$ & $0.33\pm0.01$ &$12.8\pm2.7$\ Perseus &24 &14.2 ($-21\rightarrow88$) &$9.9\pm0.1$ & $0.38\pm0.01$ &$9.4\pm1.4$\ Outer &6 &18.6 ($-6\rightarrow56$) &$13.0\pm0.3$ & $0.63\pm0.18$ &$13.8\pm3.3$\ \[table:pitchangles\] The HMSFRs with measured parallaxes are clearly tracing the major spiral arms of the Milky Way (see Fig. \[fig:parallaxes\]), and details of the locations and properties of the individual arms can be found in the primary references [@Hachi:13; @Choi:13; @Zhang:13; @Xu:13; @Wu:13; @Sato:13; @Sanna:13]. Interestingly, some surprising results are already evident. We are finding that the Perseus arm, thought to be one of the major spiral arms of the Milky Way, has little massive star formation over a 6 kpc-long arc between Galactic longitudes of $50^\circ$ and $80^\circ$ [@Choi:13; @Zhang:13]. On the other hand, the Local (Orion) arm, often called a “spur” and considered a minor structure [@Blaauw:85], has comparable massive star formation to its adjacent Sagittarius and Perseus arms [@Xu:13]. Modeling the Galaxy {#sect:modeling} =================== Given measurements of position, parallax, proper motion and Doppler shift, one has complete three-dimensional location and velocity vectors relative to the Sun. One can then construct a model of the Milky Way and adjust the model parameters to best match the data. As in @Reid:09b, we model the Milky Way as a disk rotating with speed $\Theta(R)=\To+\Tdot~(R-\Ro)$, where is the distance from the Sun to the Galactic center and is the circular rotation speed at this distance. We then evaluate the effects of different forms for the rotation curve. Since all measured motions are relative to the Sun, we need to model the peculiar (non-circular) motion of the Sun, parameterized by toward the Galactic center, in the direction of Galactic rotation, and towards the north Galactic pole (NGP). Table \[table:model\] summarizes these and other parameters. [ll]{} & Distance of Sun from GC\ & Rotation Speed of Galaxy at\ & Derivative of $\Theta$ with $R$: $\Theta(R)=\To+\Tdot~(R-\Ro)$\ \ & Solar motion toward GC\ & Solar motion in direction of Galactic rotation\ & Solar motion toward NGP\ \ & Average source peculiar motion toward GC\ & Average source peculiar motion in direction of Galactic rotation\ & Average source peculiar motion toward NGP\ \[table:model\] For each source, we treated the 3-dimensional velocity components (two components of proper motion, $\mu_x$ and $\mu_y$, and the heliocentric Doppler velocity, , as data to be compared to a model. The source coordinates ($\ell,b$) and parallax distance ($1/\pars$) were treated as independent variables. This approach is slightly different than in @Reid:09b, where the parallaxes were also treated as data in the least-squares fitting. While that approach adds some extra information ( for sources near the Galactic tangent points, distance is very sensitive to Doppler velocity, but not vice versa), it brings correlated data into the fitting, which will lead to slightly underestimated parameter uncertainties. We tested the inclusion versus exclusion of parallax with simulated data sets and found little difference and no bias between the methods. However, in order to avoid the need to adjust formal parameter uncertainties, as well as subtle issues associated with resolving the near/far distance ambiguities for sources in the first and fourth Galactic quadrants, we used the more conservative “velocity-only” fitting as done, for example, by others [@Bovy:09; @McMillan:10; @Bobylev:10; @Honma:12]. Bayesian fitting {#sect:Bayesian} ---------------- We adjusted the Galactic parameters so as to best match the data to the spatial-kinematic model using a Bayesian fitting approach. The posteriori PDFs of the parameters were estimated with Markov chain Monte Carlo (McMC) trials that were accepted or rejected by the Metropolis–Hastings algorithm. While a simple axi-symmetric model for the Galaxy may be a reasonable approximation for the majority of sources, a significant minority of outliers are expected for a variety of well known reasons. For example, the gravitational potential of the Galactic bar (or bars), which extend 3 to 4 kpc from the Galactic center [@Liszt:80; @Blitz:91; @Hammersley:00; @Benjamin:05] is expected to induce large non-circular motions for sources in its vicinity. Indeed, some of these sources show large peculiar motions, although based on a nearly flat rotation curve extrapolated inward from measurements outside this region [@Sanna:13]. Therefore, we removed the eight sources within 4 kpc of the Galactic center (excluding G000.67$-$00.03, G009.62$+$00.19, G010.47$+$00.02, G010.62$-$00.38, G012.02$-$00.03, G023.43$-$00.18, G023.70$-$00.19, G027.36$-$00.16) before model fitting. In the Galaxy’s spiral arms, super-bubbles created by multiple supernovae can accelerate molecular clouds to $\approx20$ [@Sato:08]. It is probably not possible, prior to fitting, to determine which sources have been thus affected and are likely kinematically anomalous. Therefore, we initially used an “outlier-tolerant” Bayesian fitting scheme described by @Sivia:06 as a “conservative formulation,” which minimizes the effects of deviant points on estimates of the fitted parameters. For this approach, one maximizes $$\sum_{i=1}^N~\sum_{j=1}^3~{\ln\bigl(~(1-e^{-R_{i,j}^2/2})/R_{i,j}^2~\bigr)}~~,$$ where the weighted residual $R_{i,j} = (v_{i,j}~-~m_{i,j})/w_{i,j}$ ( the data ($v$) minus model ($m$) divided by the uncertainty ($w$) of the $i^{th}$ of $N$ sources and $j^{th}$ velocity component). For large residuals, this formulation assigns a $1/R^2$ probability, compared to a Gaussian probability of $e^{-R^2/2}$ which vanishes rapidly. Thus, for example, a $5\sigma$ outlier has a reasonable (4%) probability with the outlier-tolerant approach, compared to $\approx10^{-6}$ probability for Gaussian errors in the least-squares method, and will not be given excessive weight when adjusting parameters. Once the outliers were identified and removed, we assumed Gaussian data uncertainties and fitted data by maximizing $$\sum_{i=1}^N~\sum_{j=1}^3~{-R_{i,j}^2/2}~~,$$ essentially least-squares fitting. Our choice of weights ($w$) for the data in the model fitting process was discussed in detail in @Reid:09b. We include both measurement uncertainty and the effects of random (Virial) motions of a massive young star (with maser emission) with respect to the average motion of the much larger and more massive HMSFR when weighting the differences between observed and modeled components of motion. Specifically the proper motion and Doppler velocity weights were given by $w(\mu) = \sqrt{\sigma^2_\mu + \sigma^2_{Vir}/d^2_s}$ and $w(\vhelio) = \sqrt{\sigma^2_v + \sigma^2_{Vir}}$, where $\sigma^2_{Vir}$ is the expected (1-dimensional) Virial dispersion for stars in a high mass star forming region (HMSFR). We adopted $\sigma_{Vir}=5$ , appropriate for HMSFRs with $\sim10^4$ within a radius of $\sim1$ pc, and did not adjust this value. As will be seen in §\[sect:Bayesian\], the vast majority of the velocity data can be fit with a $\chi^2_\nu$ near unity with these weights. Note that we were fairly conservative when assigning motion uncertainties for individual stars based on the maser data (see §\[sect:parallaxes\]), and this may result in a slightly low $\sigma_{Vir}$ value in order to achieve unity $\chi^2_\nu$ fits. Priors {#sect:priors} ------ In order to model the observations, one needs prior constraints on the non-circular motion of our measurement “platform” ( the solar motion parameterized by , , ) and/or the average peculiar motion of the sources being measured (parameterized by , , ). Allowing for a non-zero average source peculiar motion can be thought of as a first approximation of the kinematic effects of spiral structure. In @Reid:09b, we assumed the solar motion determined by @Dehnen:98 based on Hipparcos measurements and concluded that HMSFRs lagged circular orbital speeds by 15 ($\Vsbar=-15$ ). The observed orbital lag ($\Vsbar<0$) is insensitive to the value adopted for , but it is strongly correlated with the adopted solar motion component, $\V$ [@Reid:09b; @Honma:12]. Recently, the value of the solar motion component in the direction of Galactic rotation () has become controversial. Motivated in part by the large lag in @Reid:09b, @Schoenrich:10 re-evaluated the standard “asymmetric-drift” approach used by @Dehnen:98 and concluded that it was biased by coupled metallicity/orbital-eccentricity effects. They suggested new solar motion values; specifically they argued for a substantial increase for from 5 to 12 . This change would decrease the average orbital lag of HMSFRs () by $\approx7$ to a more theoretically appealing value near 8 . Based on the first year of data from the Apache Point Observatory Galactic Evolution Experiment (APOGEE), @Bovy:12 argue that the Sun’s motion relative to a circular orbit in the Galaxy (ie, a “rotational standard of rest”) is 26 in the direction of Galactic rotation, suggesting that the entire Solar Neighborhood, which defines the local standard of rest (LSR), leads a circular orbit by 14 . Taking into account these developments, we considered a conservative prior of $\V = 15\pm10$ , that encompasses the values of from 5 to 26 within approximately the $\pm1\sigma$ range. One could argue on theoretical grounds that HMSFRs should, on average, lag circular orbits by only a few [@McMillan:10]. We observe masers in HMSFRs that are very young and the gas out of which their exciting stars formed could have responded to magnetic shocks when entering spiral arms, leading to departures from circular speeds by $\lax10$ [@Roberts:70], apportioned between components counter to rotation and toward the Galactic Center. In addition, radial pressure gradients can also reduce orbital speeds of gas slightly [@Burkert:10], contributing to a small lag of $\approx1$ . Allowing for such effects, we consider priors for of $3\pm10$ and of $-3\pm10$ as reasonable and conservative. Given the current uncertainty in a) the value for the circular () component of solar motion and b) the magnitude of the average peculiar motions of HMSFRs, we tried four sets of priors when fitting the data: - Adopting a loose prior for the component of solar motion, $\U = 11.1\pm1.2$, $\V = 15\pm10$, $\W = 7.2\pm1.1$ , and for the average peculiar motions for HMSFRs of $\Usbar = 3\pm10$ and $\Vsbar =-3\pm10$ . - Using no priors for the average peculiar motions of HMSFRs, but tighter priors for the solar motion of $\U = 11.1\pm1.2$, $\V = 12.2\pm2.1$, $\W = 7.2\pm1.1$ from @Schoenrich:10. - Using no priors for the solar motion, but tighter priors on the average peculiar motions of HMSFRs of $\Usbar = 3\pm5$ and $\Vsbar =-3\pm5$ . - Using essentially no priors for either the solar or average peculiar motions of HMSFRs, but bounding the and parameters with equal probability within $\pm20$ of the Set-A initial values and zero probability outside that range. Models A1–A4 ------------ Using the 95 sources with Galactocentric radii greater than 4 kpc[^3], the outlier-tolerant Bayesian fitting approach, and the Set-A priors as described above, we obtained the parameter estimates listed in Table \[table:fits\] under fit A1. As expected for a sample with some outliers (see discussion in §\[sect:Bayesian\]), we found a $\chisq=562.6$, greatly exceeded the 277 degrees of freedom, owing to a number of sources with large residuals. We iteratively removed the sources with the largest residuals. Using the outlier-tolerant Bayesian fitting approach (see §\[sect:Bayesian\]) minimizes potential bias, based on assumed “correct” parameter values, when editing data. However, to further guard against any residual bias, we first removed sources with $>6\sigma$ residuals, followed by re-fitting and removal of those with $>4\sigma$ residuals, and finally re-fitting and removal of those with $>3\sigma$ residuals (fits A2, A3 & A4, not listed here). In total, 15 sources[^4] were removed. [lccccc]{} Parameter Estimates\ (kpc) &$8.15\pm0.25$ &$8.34\pm0.16$ &$8.33\pm0.16$ &$8.30\pm0.19$ & $8.29\pm0.21$\ () & $238\pm11\q$ &$240\pm8\q$ &$243\pm6\q$ &$239\pm8\q$ &$238\pm15$\ ()&$-0.1\pm0.7\q$&$-0.2\pm0.4\q$ &$-0.2\pm0.4\q$ &$-0.1\pm0.4\q$ &$-0.1\pm0.4$\ \ () &$10.4\pm1.8$ &$10.7\pm1.8$ &$10.7\pm1.8$ &$ 9.9\pm3.0$ &$ 9.6\pm3.9$\ () &$15.1\pm7.3$ &$15.6\pm6.8$ &$12.2\pm2.0$ &$14.6\pm5.0$ &$16.1\pm13.5$\ () &$8.2\pm1.2$ &$\p8.9\pm0.9$ &$\p8.7\pm0.9$ &$\p9.3\pm1.0$ &$\p9.3\pm1.0$\ \ ()&$\p3.7\pm2.4$ &$\p2.9\pm2.1$ &$\p2.9\pm2.1$ &$\p2.2\pm3.0$ &$\p1.6\pm3.9$\ ()&$-2.4\pm7.4$ &$-1.6\pm6.8$ &$-5.0\pm2.1 $ &$-2.4\pm5.0$ &$-1.2\pm13.6$\ \ Fit Statistics\ $\chisq$ & 562.6 & 224.9 & 225.1 & 224.7 & 224.1\ $N_{dof}$ & 277 & 232 & 232 & 232 & 232\ $N_{sources}$ & 95 & 80 & 80 & 80 & 80\ $r_{\Ro,\To}$ &0.61 &0.46 &0.74 &0.66 &0.44\ \[table:fits\] Model A5 -------- With the resulting “clean” data set of 80 sources, we performed a least-squares fit (assuming a Gaussian PDF for the data uncertainties). We used the same loose priors (Set-A) as for model A1, namely solar motion components $\U = 11.1\pm1.2$, $\V = 15\pm10$, $\W = 7.2\pm1.1$ and average peculiar motions for HMSFRs of $\Usbar = 3\pm10$ and $\Vsbar =-3\pm10$ . This resulted in the parameter estimates listed under fit A5 in Table \[table:fits\]. This model produced a good $\chisq=224.9$ for 232 degrees of freedom and estimates of $\Ro=8.34\pm0.16$ kpc and $\To=240\pm8$ . We find $\Tdot=-0.2\pm0.4$ , indicating a very flat rotation curve for the Milky Way between radii of $\approx5$ and 16 kpc from the Galactic center. Compared to the preliminary results of @Reid:09b based on 16 sources, where the Pearson product-moment correlation coefficient for and was high, $r_{\Ro,\To}=0.87$, with the larger number of sources and a better distribution across the Galaxy, these parameters are significantly less correlated, $r_{\Ro,\To}=0.46$. The joint and marginalized PDFs for these fundamental Galactic parameters are displayed in Figure \[fig:RoToPDF\]. The circular velocity parameters are still correlated (see §\[sect:correlations\]), but linear combinations of these parameters are well determined: $\To+\V=255.2\pm5.1$ and $\V-\Vsbar=17.1\pm1.0$. Also, the angular rotation rate for the Sun’s orbit about the Galactic center is constrained to $\pm1.4$% accuracy: $(\To+\V)/\Ro=30.57\pm0.43$ . This value is consistent with the reflex of the [apparent]{} motion of Sgr A\, the assumed motionless supermassive black hole at the center of the Galaxy, which gives $30.26\pm0.12$ [@Reid:04]. The component of solar motion in the direction of Galactic rotation, $\V$, estimated to be $15.6\pm6.8$ is better constrained than the prior of $15\pm10$ . It is consistent with the [local]{} estimate (relative to Solar Neighborhood stars) of 12 [@Schoenrich:10] and the [global]{} estimate of @Bovy:12 of $26\pm3$ (relative to stars across the Milky Way). Model B1 -------- In order to explore the sensitivity of the modeling to our priors, we fit the clean data set using the Set-B priors: adopting the latest Hipparcos measurement of the solar motion of $\U = 11.1\pm1.2$, $\V = 12.2\pm2.1$, $\W = 7.2\pm1.1$ [@Schoenrich:10] and no prior information on the average peculiar motion of the HMSFRs. This resulted in parameter estimates similar to those of model A5, $\Ro=8.33\pm0.16$ kpc and $\To=243\pm6$ . The quality of fit, as measured by $\chisq=225.1$ for 232 degrees of freedom, was comparably good as for model A5. The average velocity lag of the HMSFRs relative to circular orbits, which was not constrained by priors, was $\Vsbar = -5.0 \pm 2.1$ . This is comparable to that found by @Reid:09b, after correcting for the 7 difference in the adopted solar motion values. Model C1 -------- Given the current uncertainty in the component of solar motion, we fit the data with the Set-C priors, assuming no prior information for the solar motion, but using a stronger prior than for model A5 for the average peculiar motion of the HMSFRs: $\Usbar = 3\pm5$ and $\Vsbar =-3\pm5$ . As for model B1, we found most parameter estimates to be similar to model A5, eg, $\Ro=8.30\pm0.19$ kpc and $\To=239\pm8$ . For the solar motion, we find $\U = 9.9\pm2.0$, $\V =14.6\pm5.0$, and $\W = 9.3\pm1.0$ . The value is consistent with revised @Schoenrich:10 [12 ] solar motion, but differs by $2\sigma$ from the @Bovy:12 estimate. Model D1 -------- In order to facilitate the use of the results presented here with other Galactic parameter estimates, we perfomed a fit with essentially no informative priors. We did this by taking the A5 (Set-A) initial parameter values and assuming flat priors for all parameters except for and . For these parameters we assumed equal probability for values within $\pm20$ of the initial values and zero probability outside this range in order to exclude unreasonable parameter values. The parameters that remain well determined include $\Ro=8.29\pm0.21$ kpc, $\To=238\pm15$ , $\Tdot=-0.1\pm0.4$ , $\U=9.6\pm3.9$ , $\W=9.3\pm1.0$ , and $\Usbar=1.6\pm3.9$ . The correlated velocity terms, and displayed nearly flat posteriori PDFs over their allowed ranges. However, linear combinations involving these parameters are very well constrained, $\To+\V = 253.8\pm6.4$ and $\V-\Vsbar = 17.2\pm1.2$ , as well as the angular rotation rate of the Sun about the Galactic center, $(\To+\V)/\Ro=30.64\pm0.41$ . Rotation Curves {#sect:rotationcurves} --------------- Next, we investigated the sensitivity of the fundamental Galactic parameters, and , to alternative rotation curves. When fitting, we replaced the simple linear form, $\Theta(R)=\To+\Tdot~(R-\Ro)$, with the empirically determined functions of $\Theta(R)$ of @Clemens:85, the power-law parameterization of @Brand:93, a polynomial, and the “universal” rotation curve of @Persic:96. We adopted the Set-A priors in order to facilitate comparisons with the A5 fit. Table \[table:rotationcurves\] presents the fitting results for these rotation curves. [lccccc]{} Parameter Estimates\ (kpc) &$8.36\pm0.16$ &$8.12\pm0.14$ &$8.34\pm0.16$ &$8.34\pm0.17$ &$8.31\pm0.16$\ () &$237\pm8\q$ &$221\pm8\q$ &$240\pm9\q$ &$241\pm9\q$ &$241\pm8\q$\ () &$10.1\pm1.8~$ &$10.5\pm1.8\p$&$10.5\pm1.8\p$&$10.7\pm1.7\p$ &$10.5\pm1.7\p$\ () &$19.4\pm6.8~$ &$25.0\pm6.8~$ &$15.5\pm6.8~$ &$14.7\pm6.8~$ &$14.4\pm6.8~$\ () &$8.9\pm1.0$ &$8.9\pm1.0$ &$8.8\pm1.0$ &$8.8\pm0.9$ &$8.9\pm0.9$\ ()&$2.4\pm2.1$ &$2.6\pm2.0$ &$2.8\pm2.0$ &$2.8\pm2.0$ &$2.6\pm2.1$\ ()&$+3.4\pm6.8\p$ &$+8.5\pm6.8\p$&$-1.5\pm6.8$ &$-1.4\pm6.8\p$ &$-1.4\pm6.8\p$\ () &... &... &$240\pm9~$ &$241\pm9~$ &$241\pm8~$\ &... &... &$~0.00\pm0.02$&$~~0.5\pm3.7$ &$~0.90\pm0.06$\ &... &... &... &$-15.1\pm8.4~$ &$~1.46\pm0.16$\ \ Fit Statistics\ $\chisq$ &229.7 & 248.1 & 225.2 & 221.9 &214.5\ $N_{dof}$ &233 & 233 & 231 & 230 &230\ $N_{sources}$&80 & 80 & 80 & 80 & 80\ $r_{\Ro,\To}$&0.46 &0.36 &0.48 &0.47 &0.47\ \[table:rotationcurves\] @Clemens:85 supplied two curves with different shapes: one assuming the old IAU constants (C-10) of $\Ro=10$ kpc and $\To=250$ and the other assuming the revised constants (C-8.5) of $\Ro=8.5$ kpc and $\To=220$ currently in widespread use. The C-10 model has rotational speeds that rise faster with radius than the C-8.5 model. For either model, we fitted for different values of (which we used to scale model radii) and (which we used to scale rotation speeds). @Brand:93 parameterize their rotation curve (BB) as a power law in Galactocentric radius, $R$, with potentially three adjustable parameters: $\Theta(R) = \aone (R/\Ro)^\atwo + \athr$. For a flat rotation curve ($\atwo=0$), parameters and become degenerate. Since the Galaxy’s rotation curve is nearly flat over the range of radii we sample (see, eg, model A5 above), we held at zero, solving only for and . Indeed, we find the power law exponent, $\atwo=-0.01\pm0.01$, essentially flat. For this formulation, $\aone=\To$, and in Table \[table:rotationcurves\] we copy to to facilitate comparison with other models. As an alternative to a power law rotation curve, we fitted a second-order polynomial (Poly) in $\rho=(R/\Ro)-1$: $\Theta(R) = \aone + \atwo \rho + \athr \rho^2$. The model fit parameters for this form of a rotation curve are similar to those from models C-10, BB and Univ. The universal (Univ) rotation curve of @Persic:96 includes terms for an exponential disk and a halo. It can have three adjustable parameters: , the circular rotation speed at the radius enclosing 83% of the optical light ($R_{opt}$); $\atwo = R_{opt} / \Ro$; and , a core-radius parameter for the halo contribution, nominally 1.5 for an $L^$ galaxy. With flat priors for the three rotation curve parameters, the posteriori PDF for $\atwo$ was bimodal, with the dominant peak at $\atwo=0.9$ and a second peak with 50% of the primary’s amplitude at $\atwo=0.1$. Since the secondary peak seems unlikely, we refit the data using a prior for $\atwo$ of $1.2\pm0.5$. We then obtained similar parameter values as other models (see Table \[table:rotationcurves\]), with the three adjustable rotation curve parameters of $\aone = 241\pm8$ , $\atwo = 0.90\pm0.06$, and $\athr = 1.46\pm0.16$. All but one of the rotation curve models lead to similar values for the fundamental Galactic parameters and as our A5 fit. Only the Clemens “$\Ro=8.5$ kpc; $\To=220$ ” (C-8.5) rotation curve results in a marginally significant change in estimates of and . However, this fit has a significantly poorer quality ($\chi^2=248.1$ for 233 degrees of freedom) than, for example, the A5 fit ($\chi^2=224.9$ for 232 degrees of freedom), and we do not consider this model further. We conclude that the fundamental Galactic parameters and are reasonably insensitive to a wide variety of rotation curve shapes. With full 3-dimensional location and velocity information, we can transform our heliocentric velocities to a Galactocentric reference frame and calculate the tangential (circular) speed for each HMSFR. Figure \[fig:rotationcurve\] plots these speeds for [all]{} sources in Table \[table:parallaxes\]. Most published rotation curves for the Milky Way have come from only one component of velocity (radial), often using kinematic distances and [assuming]{} a value for . As such, the data in Fig \[fig:rotationcurve\] represent a considerable advance. See also the analysis of this data set by @Xin:13. It is important to remember that the transformation from heliocentric to Galactocentric frames requires accurate values of , , and, most importantly, $\To+\V$, since the motion of the Sun has (by definition) been subtracted in the heliocentric frame. For most sources, increasing or decreasing the assumed value of $\To+\V$ would, correspondingly, move each data point up or down by about the same amount. Thus, the level of this, and essentially all published, rotation curves is determined mostly by $\To+\V$. Our results are the first to use fully 3-dimensional data to strongly constrain all three parameters: , and $\To+V$. The dashed line in Fig \[fig:rotationcurve\] represent the linear rotation curve from the A5 fit, based only on sources with $R>4$ kpc. Sources used in the fit are plotted with filled symbols and the sources not used with open symbols. The dashed line indicates the expected rotation for sources in circular Galactic orbit ( $\Usbar=\Vsbar=0$). There are now sufficient data to clearly indicate that the rotation curve drops at Galactocentric radii $\lax4$ kpc. However, given the likelihood for a significant non-axisymmetric gravitational potential within $\approx4$ kpc of the center, more measurements are needed before extending a rotation curve to this region as azimuthal terms may be needed. Peculiar Motions of HMSFRs -------------------------- Figure \[fig:peculiarmotions\] shows the peculiar (non-circular) motions of all sources in Table \[table:parallaxes\] with motion uncertainties less than 20 . Similar results were described in the primary papers presenting the parallaxes and proper motions for each arm [@Sato:13; @Wu:13; @Xu:13; @Choi:13; @Zhang:13; @Hachi:13]. For uniformity, here the motions were calculated using the A5 fit parameters (see Table \[table:fits\]), but with zero correction for the average source peculiar motions. Typical peculiar motions are $\approx10$ , but some sources have much larger values. For example, many sources in the Perseus arm in the Galactic longitude range $\approx100^\circ$ to $\approx135^\circ$ display peculiar motions $\gax20$ . Many sources within $\approx4$ kpc of the Galactic Center display even larger peculiar motions, probably indicating that the rotation curve used here is inadequate to describe their Galactic orbits, especially in the presence of the Galactic bar(s). Parameter Correlations {#sect:correlations} ---------------------- [lrrrrrrrr]{} $\Ro$ &1.000 &0.465 &0.103 &0.452 &0.023 &$-$0.003&0.517 &$-$0.002 $\To$ &0.465 &1.000 &0.136 &0.243 &$-$0.796&$-$0.009&0.171 &$-$0.809 $\Tdot$ &0.103 &0.136 &1.000 &$-$0.124&$-$0.009&0.025 &$-$0.094 &$-$0.018 $\U$ &0.452 &0.243 &$-$0.124&1.000 &$-$0.014&$-$0.017&0.839 &0.025 $\V$ &0.023 &$-$0.796&$-$0.009&$-$0.014&1.000 &0.011 &$-$0.006 &0.990 $\W$ &$-$0.003&$-$0.009&0.025 &$-$0.017&0.011 &1.000 &$-$0.002 &0.010 $\Usbar$ &0.517 &0.171 &$-$0.094&0.839 &$-$0.006&$-$0.002&1.000 &0.028 $\Vsbar$ &$-$0.002&$-$0.809&$-$0.018&0.025 &0.990 &0.010 &0.028 &1.000 \[table:correlations\] The Pearson product-moment correlation coefficients, $r$, for all parameters from fit A5 are listed in Table \[table:correlations\]. In the preliminary analysis of 16 HMSFRs with parallaxes and proper motions of @Reid:09b, the estimates of and were strongly correlated ($\rPpm=0.87$). However, with the much larger set of HMSFRs that covers a larger portion of the Galaxy, the correlation between and estimates is now moderate: $\rPpm=0.465$ for our reference A5 fit. However, there remains a significant anti-correlation between and ($r_{\To,\V}=-0.809$), as well as a strong correlation between and ($r_{\V,\Vsbar}=0.990$). As suggested by the fitted parameter values in Table \[table:fits\], our data strongly constrain the following combinations of these correlated parameters: $\To + \V=255.2\pm5.1$ and $\V - \Vsbar=17.1\pm1.0$ . Also, the combination of parameters that yield the angular orbital speed of the Sun about the Galactic center, $(\To+\V)/\Ro = 30.57 \pm 0.43$ , is more tightly constained than the individual parameters. Figure \[fig:3PDFs\] shows the marginalized PDFs for these combinations of parameters. Comparison with Other Modeling Approaches ----------------------------------------- Other groups have analyzed parallax and proper motion data sets from the BeSSeL Survey and the VERA project, focusing on different assumptions and results. @Bovy:09 confirmed the counter rotation of HMSFRs (assuming $\V=5$ ) noted by @Reid:09b and argued for a comparable value for $(246\pm30$ ), but with considerably lower significance. Alternatively, @McMillan:10 found that the -component of solar motion of 5 , provided by @Dehnen:98, should be raised to $\approx12$ , thereby reducing the estimated counter-rotation of HMSFRs. @Bobylev:10, using 28 parallaxes available at that time and a Fourier analysis technique, estimated $\To=248\pm14$ and $\V=11.0\pm1.7$ , assuming $\Ro\equiv8.0$ kpc. Finally, @Honma:12, using 52 parallaxes, including some low-mass star forming regions, estimated $\Ro=8.05\pm0.45$ kpc and $\To=238\pm14$ for $\V\equiv12$ . The @Bovy:09 re-analysis of our preliminary data employed a different approach than that of @Reid:09b. Bovy treat the elements of the velocity dispersion tensor of the HMSFRs as free parameters. These parameters give the expected deviations (variances and covariances) of the velocity data from a smooth, axi-symmetric model of Galactic rotation and are used to adjust the weights applied to the different velocity components when fitting the data. However, while Bovy found a significant trace for the tensor, the velocity dispersion parameters were only marginally constrained; formally none of the diagonal components had $>2.8\sigma$ formal significance. Also, their values for the radial and tangential components were nearly identical, suggesting that little is gained by making these free parameters versus adopting a single physically motivated value ($\sigma_{Vir}$) as we have done. Note that our value for $\sigma_{Vir}$ is comparable to the dispersion parameter ($\Delta_v$) values found by @McMillan:10, which range from about 6 to 10 , but is considerably smaller than those of @Bovy:09 of $\approx20$ . The reason for this difference is unclear, but might reflect different treatments of outlying data and/or increased parameter correlations associated with the 6 extra parameters used in solving for the tensor elements. Discussion {#sect:discussion} =========== Solar Motion {#sect:solar_motion} ------------ If one adopts the theoretically motivated prior that HMSFRs have small peculiar motions (Set-C with no prior on the solar motion), then model fit C1 indicates $\V=14.6\pm5.0$ . This is a [global]{} measure of the peculiar motion of the Sun and, as such, is relative to a “rotational standard of rest” as opposed to a Local Standard of Rest (LSR), defined relative to stellar motions in the Solar Neighborhood. If Solar Neighborhood stars (extrapolated to a zero-dispersion sample) are, on average, stationary with respect to a circular orbit, then these two solar motion systems will be the same. Our estimate of $\V$ is consistent with the $12$ value of @Schoenrich:10, measured with respect to Solar neighborhood stars, but there is some tension between our global estimate of and that of @Bovy:12 of $26\pm3$ , as these two estimates differ by about $2\sigma$. However, if one drops the prior that HMSFRs have small peculiar motions, then our result loses significance. The large counter-rotation of HMSFRs, originally suggested by @Reid:09b, was based on the initial Hipparcos result of @Dehnen:98 that $\V=5$ . As the outcome of the @Schoenrich:10 re-analysis of Hipparcos data, which gives $\V=12$ , supersedes that lower $\V$ value, it now appears that any average counter-rotation of HMSFRs is $\lax5$ . Given that we strongly constrain $\V-\Vsbar=17.1\pm1.0$ , were one to independently constrain $\V$ with $\pm2$ accuracy, the issue of HMSFR counter rotation could be clarified. While our estimate of has a large uncertainty (owing to correlations with and ), we find and are well constrained. In fit D1, in which no informative prior was used for the components of motion either toward the Galactic center or perpendicular to the Galactic plane, we find that $\U=9.6\pm3.9$ and $\W=9.3\pm1.0$ , respectively. Our estimate of the Sun’s motion toward the Galactic center is in agreement with most other estimates, $11.1\pm1.2$ by @Schoenrich:10 and $10\pm1$ by @Bovy:12; see also the compilation of estimates by @Coskunoglu:11. The solar motion component perpendicular to the Galactic plane, $\W$, is generally considered to be straight forwardly determined and recent estimates typically range between 7.2 [@Schoenrich:10] (relative to local stars within $\sim0.2$ kpc) and 7.6 [@Feast:97] (relative to stars within $\sim3$ kpc), with uncertainties of about $\pm0.5$ . We find a slightly larger value of $\W=9.3\pm1.0$ (for model D1 which used no informative priors for the solar motion), which may be significant; the difference between the locally and our globally measured value ( relative to stars across the Galaxy) is $2.1\pm1.1$ . Note that one might expect a small difference between measurements with respect to a local and a global distribution of stars were the disk of the Galaxy to precess owing to Local Group torques. Simulations of galaxy interactions in a group suggest that a disk galaxy can complete one precession cycle over a Hubble time. Were the Milky Way to do this, one would expect a vertical pecessional motion at a Galactocentric radius of the solar neighborhood of order $\Ro\Ho~\sim0.6$ . It is possible that the differences in the local and global estimates of $\W$ can, in part, be explained in this manner. Galactic Rotation Curve and Disk Scale Length --------------------------------------------- Among the various forms of rotation curves that we fit to the data, the universal curve advocated by @Persic:96 to apply to most spiral galaxies yielded the best fit (see discussion in §\[sect:rotationcurves\], Table \[table:rotationcurves\] and Fig. \[fig:rotationcurve\]). This rotation curve matches the flat to slightly declining run of velocity with Galactocentric radius from $R\approx5\rightarrow16$ kpc, as well as reasonably tracing the decline in orbital velocity for $R\lax5$ kpc. However, many of the sources near the Galactic bar(s) cannot be well modeled with any axi-symmetric rotation curve. The best fit value for our $\atwo$ parameter ($R_{opt}/\Ro$), coupled with our estimate of $\Ro=8.34\pm0.16$, locates $R_{opt}$ at $7.5\pm0.52$ kpc. The $\atwo$ parameter is sensitive to the slope of the rotation curve (near $R_{opt}$) and the radius at which it turns down toward the Galactic center. For example, setting $\atwo=0.7$ steepens the rotation curve at large radii and moves the turn down radius to $\approx3.5$ kpc, while setting $\atwo=1.1$ flattens the rotation curve and increases the turn down radius to $\approx6.5$ kpc. Given that the (thin) disk scale length, $R_D = R_{opt} / 3.2$ [@Persic:96], we estimate $R_D=2.44\pm0.16$ kpc. Estimates of $R_D$ in the literature range from $\approx 1 \rightarrow 6$ kpc [@Kent:91; @Chang:11; @McMillan:11], with most consistent with a value between $2\rightarrow3$ kpc. Our estimate is also consistent with that of @Porcel:98 who modeled the positions and magnitudes of 700,000 stars in the Two Micron Galactic Survey database and found $R_D=2.3\pm0.3$ kpc and, more recently, @Bovy:13, who modeled the dynamics of $\approx16,000$ stars from the SEGUE survey and concluded that $R_D=2.14\pm0.14$ kpc. The Distance to the Galactic Center: ------------------------------------- Models A5, B1 and C1, which used different combinations of solar motion and/or average source peculiar motion priors, have comparable $\chi^2$ values and all parameter estimates are statistically consistent. Because the priors for Model A5 are the least restrictive in keeping with current knowledge, we adopt those parameters as representative. Specifically, we find $\Ro=8.34\pm0.16$ kpc, $\To=240\pm8$ and $\Tdot=-0.2\pm0.4$ . As noted in §\[sect:correlations\] and §\[sect:rotationcurves\], with the much larger data set now available, estimates of and are no longer strongly correlated and appear fairly insensitive to the assumed nature of the rotation curve. These parameter estimates are consistent with, but significantly better than, the preliminary values of $\Ro=8.4\pm0.6$ kpc, $\To=254\pm16$ and a nearly flat rotation curve reported in @Reid:09b, based on parallaxes and proper motions of 16 HMSFRs and assuming $\V=5$ , and $\Ro=8.05\pm0.45$ kpc and $\To=238\pm14$ from @Honma:12, based on a sample of 52 sources and assuming $\V=12$ . While there are numerous estimates of the distance to the Galactic center in the literature ( @Reid:93), here we only compare those based on direct distance measurements. A parallax for the water masers in Sgr B2, a star forming region projected less than 0.1 kpc from the Galactic center, indicates $\Ro=7.9\pm0.8$ kpc [@Reid:09c], consistent with, but considerably less accurate than, our current result. More competitive estimates of come from the orbits of “S-stars” about the supermassive blackhole Sgr A\. Combining the nearly two decades of data from the ESO NTT/VLT [@Gillessen:09a] and Keck [@Ghez:08] telescopes that trace more than one full orbit for the star S2 (a.k.a. S0-2), @Gillessen:09b conclude that $\Ro=8.28\pm0.33$ kpc. Recently the Keck group, extending their time sequence of observations by only a few years, announced a value of $\Ro=7.7\pm0.4$ kpc [@Morris:12], in mild tension both with the @Gillessen:09b analysis and our parallax-based result. However, in the latest publication of the Keck group, @Do:13 combined modeling of the distribution and space velocities of stars within the central 0.5 pc of the Galactic center with the stellar orbital result for star S0-2 [@Ghez:08] and conclude that $\Ro=8.46^{+0.42}_{-0.38}$ kpc, removing any tension with our estimate and that of the ESO group. We conclude that our estimate of $\Ro=8.34\pm0.16$ kpc is consistent with that from the Galactic center stellar orbits and is likely the most accurate to date. The Circular Rotation Speed at the Sun: ---------------------------------------- Over the last four decades there have been many estimates of ranging from $\sim170 \rightarrow 270$ [@Kerr:86; @Olling:98]. Focussing the discussion to the more direct measurements, two recent studies favor a lower and one a higher value of than our estimate of $\To=240\pm8$ . @Koposov:10 model the orbit of the GD-1 stream from a tidally disrupted stellar cluster in the Milky Way halo and estimate $\To+\V=221\pm18$, where the @Dehnen:98 solar motion component of $\V=5$ was adopted. Recently, @Bovy:12 modeled line-of-sight velocities of 3365 stars from APOGEE and find $\To=218\pm6$ , but with a large value for the solar motion component in the direction of Galactic rotation, $\V=26\pm3$ . Their full tangential speed $\To+\V=242^{+10}_{-3}$ is consistent with our value of $252.2\pm4.8$ , suggesting the discrepancy between the Bovy and our results are probably caused by differences in the solar motion. However, another recent study by @Carlin:12, modeling the Sagittarius tidal stream, yields $\To$ estimates from $232 \rightarrow 264$ . Our data also strongly constrain the angular rotation of the Sun about the Galactic center, $(\To+\V)/\Ro=30.57\pm0.43$ . This value can be compared with an independent and direct estimate based on the proper motion of Sgr A\, interpreted as the reflex motion from the Sun’s Galactic orbit, of $30.24\pm0.12$ [@Reid:04]. For $\Ro=8.34\pm0.16$ kpc, the proper motion of Sgr A\* translates to $\To+\V=252.2\pm4.8$ , in good agreement with the parallax results. We conclude that $\To$ exceeds the IAU recommended value of 220 with $>95$% probability provided that $\V\lax23$ . Clearly, independent [global]{} measures of $\V$ are critical to establish $\To$ and $\Vsbar$ with high accuracy. Changing the value of would have widespread impact in astrophysics. For example, increasing by 20 with respect to the IAU recommended value of 220 reduces kinematic distances by about 10%, leading to a decrease of 20% in estimated young star luminosities, a corresponding decrease in estimated cloud masses, and a change in young stellar object ages. Estimates of the total mass of the dark matter halo of the Milky Way scale as $V_{max}^2~R_{Vir}$. Since the maximum in the rotation curve ($V_{max}$) and the Virial radius ($R_{Vir}$) scale linearly with , the mass of the halo scales as $\Theta_0^3$, leading to a 30% increase in the estimate of the Milky Way’s (dark-matter dominated) mass. This, in turn, affects the expected dark-matter annihilation signal [@Finkbeiner:09], increases the “missing satellite” problem [@Wang:12], and increases the likelihood that the Magellanic Clouds are bound to the Milky Way [@Shattow:09]. The Hulse-Taylor Binary Pulsar and Gravitation Radiation -------------------------------------------------------- An interesting example of the effects of Galactic parameters on fundamental physics comes from the Hulse-Taylor binary pulsar. The dominant uncertainty in measuring the gravitational radiation damping of the binary’s orbit comes from the need to correct for the effects of the Galactic accelerations of the Sun and the binary [@Damour:91; @Weisberg:10]. These accelerations contribute $\approx1$% to the [apparent]{} orbital period decay. In 1993 when the Nobel Prize was awarded in part for this work, the IAU recommended values were $\Ro=8.5\pm1.1$ kpc and $\To=220\pm20$ [@Kerr:86]. Using these Galactic parameters, the formalism of Damour & Taylor, improved pulsar timing data of Weisberg, Nice, & Taylor, and a pulsar distance of 9.9 kpc, the binary’s orbital period decays at a rate of $0.9994\pm0.0023$ times that prediction from general relativity (GR). Using the improved Galactic parameters from the A5 fit ($\Ro=8.34\pm0.16$ kpc and $\To=240\pm8$ ), gives a GR test value of $0.9976\pm0.0008$. This provides a three-fold improvement in accuracy. Both of these examples assumed a distance to the binary pulsar of 9.9 kpc [@Weisberg:08]. Given the improvement in the Galactic parameter values, the dominant uncertainty in the GR test now is the uncertain pulsar distance. A pulsar distance of 7.2 kpc would bring the GR test value to 1.0000 and a trigonometric parallax accurate to $\pm8$%, which is possible with in-beam calibration with the VLBA, would bring the contribution of distance uncertainty down to that of the current Galactic parameter uncertainty. Alternatively, if one assumes GR is correct, the current improvement in Galactic parameters suggests that the Hulse-Taylor binary pulsar’s distance is $7.2\pm0.5$ kpc. 0.5truein This work was partially funded by the ERC Advanced Investigator Grant GLOSTAR (247078). The work was supported in part by the National Science Foundation of China (under grants 10921063, 11073046, 11073054 and 11133008) and the Key Laboratory for Radio Astronomy, Chinese Academy of Sciences. AB acknowledges support by the National Science Centre Poland through grant 2011/03/B/ST9/00627. 0.5truein [Facilities:]{} , , Ando, K., Nagayama, T., Omodaka, T., 2011, , 63, 45 Asaki, Y., Deguchi, S., Imai, H., Hachisuka, K., Miyoshi, M. & Honma, M. 2010, , 721, 267 Bartkiewicz, A., Brunthaler, A., Szymczak, M. van Langevelde, H. J & Reid, M. J. 2008, , 490, 787 Benjamin, R. A. 2008, in “Massive Star Formation: Observations Confront Theory,” ASP Conference Series, Vol. 387, eds. H. Beuther, H. Linz & Th. Henning, p. 375 Benjamin, R. A. 2005, , 630, L149 Blaauw, A. 1985, in “The Milky Way Galaxy: Proceedings of IAU Symp. 106” H. van Woerden et al. eds., (Dordrecht, D. Reidel Pub. Co.), p. 335 Blitz, L. & Spergel,, D. N. 1991, , 379, 631 Bobylev, V. V. & Bajkova, A. T. 2010, , 408, 1788 Bovy, J., Hogg, D. W. & Rix, H.-W. 2009, , 704, 1704 Bovy, J., Prieto, C. A., Beers, T. C., 2012, , 759, 131 Bovy, J. & Rix, H.-W. 2013, , 779, 115 Brand, J. & Blitz, L. 1993, , 275, 67 Brunthaler, A., Reid, M. J., Menten, K. M., Zheng, X. W., Moscadelli, L. & Xu, Y. 2009, , 693, 424 Burkert, A., Genzel, R., Bouché, N. 2010, , 725, 2324 Carlin, J. L., Majewski, S. R., Casetti-Dinescu, D. I., Law, D. R., Girard, T. M. & Patterson, R. J. 2012, , 744, 25 Chang, C.-K., Ko, C.-M. & Peng, T.-H. 2011, , 740, 34 Choi, Y. K. , 2008, , 60, 1007 Choi, Y. K., Hachisuka, K., Reid, M. J., 2014, submitted to Clemens, D. P. 1985, , 295, 422 Coskunoglu, B, Ak, S., Bilir, S. 2011, , 412, 1237 Damour, T. & Taylor, J. H. 1991, , 366, 501 Dehnen, W., & Binney, J. J., , 1998, 298,387 Do, T., Martinez,, G. D., Yelda, S., 2013, , 779, 6 Eyer, L., Holl, B., Pourbaix, D. 2013, CEAB, 37, 115 Feast, M. & Whitelock, P. 1997, , 291, 683 Finkbeiner, D. P., Slatyer, T. R., Weiner, N., & Yavin, I. 2009, [JCAP]{}, 9, 37 Ghez, A. M., Salim, S., Weinberg, N. N., 2008, , 689, 1044 Gillessen, S., Eisenhauer, F. Trippe, S., 2009a, , 692, 1075 Gillessen, S., Eisenhauer, F., Fritz, T. K., 2009b, , 707, L114 Hachisuka, K. 2006, , 645, 337 Hachisuka, K. Brunthaler, A., Menten, K. M., 2009, , 696, 1981 Hachisuka, K., Choi, Y. K., Reid, M. J., 2014, submitted to Hammersley, P. L., Garzón, F., Mahoney, T. J., López-Corredoira, M., Torres, M. A. P. 2000, , 317, 45 Hirota, T, Ando, K., Bushimata, T., 2008, , 60, 961 Honma, M., Bushimata, T., Choi, Y. K., 2007, , 59, 889 Honma, M., Hirota, T., Kan-Ya, Y., 2007, , 63, 17 Honma,M. 2012, , 64, 136 Immer, K., Reid, M. J., Menten, K. M., Brunthaler, A., & Dame, T. M. 2013, , 553, 117 Kennicutt, R. C. Jr. 1981, , 86, 1847 Kent, S. M., Dame, T. M. & Fazio, G. 1991, , 378, 131 Kerr, F. J. & Lynden-Bell, D. 1986, , 221, 1023 Kim, M. K., Hirota, T., Honma, M., 2008, , 60, 991 Koposov, S. E., Rix, H.-W. & Hogg, D. W. 2010, , 712, 260 Kurayama, T., Nakagawa, A., Sawada-Satoh, S, 2011, , 63, 513 Liszt, H. S. & Burton, W. B. 1980, , 236, 779 McMillan, P. J. 2011, , 414, 2446 McMillan, P. J. & Binney J. J. 2010, , 402, 934 Menten, K. M., Reid, M. J., Forbrich J. & Brunthaler, A. 2007, , 474, 515 Moellenbrock, G. A., Claussen, M. J. & Goss, W. M. 2009, , 694, 192 Morris, M. R., Meyer, L & Ghez, A. 2012, [RAA]{}, 12, 995 Moscadelli, L., Reid, M. J., Menten, K. M., Brunthaler, A., Zheng, X. W. & Xu, Y. 2009, , 693, 406 Moscadelli, L., Cesaroni, R., Rioja, M. J., Dodson, R., Reid, M. J., 2011, , 526, 66 Nagayama, T., Omodaka, T., Nakagawa, A, 2011, , 63, 23 Niinuma, K. Nagayama, T., Hirota, T., 2011, , 63, 9 Oh, C. S., Kobayashi, H., Honma, M., Hirota, T., Sato, K., Ueno, Y., 2010, , 62, 101 Olling, R. P. & Merrifield, M. R. 1998, , 297,943 Persic, M., Salucci, P. & Stel, F. 1996, , 281, 27 Porcel, C., Garzon, F., Jimenez-Vicente, J. & Battaner, E. 1998, , 330, 136 Reid, M. J. 1993, , 31, 345 Reid, M. J. & Brunthaler, A., 2004, , 616, 872 Reid, M. J., Menten, K. M., Brunthaler, A., Zheng, X. W., Moscadelli, L. & Xu, Y. 2009, , 693, 397 Reid, M. J., Menten, K. M., Zheng, X. W. 2009, , 700, 137 Reid, M. J., Menten, K. M., Zheng, X. W., Brunthaler, A., & Xu, Y. 2009, , 705, 1548 Roberts, W. W. & Yuan, C. 1970, , 161, 877 Rygl, K. L. J., Brunthaler, A., Reid, M. J., Menten, K. M., van Langevelde, H. J., Xu, Y., 2010, , 511, 2 Rygl, K. L. J., Brunthaler, A., Sanna, A., 2012, , 539, 79 Sandstrom, K. M., Peek, J. E. G., Bower, G. C., Bolatto, A. D. & Plambeck, R. L., 2007, , 667, 1161 Sanna, A., Reid, M. J., Moscadelli, L., 2009, , 706, 464 Sanna, A., Reid, M. J., Dame, T., 2012, , 745, 82 Sanna, A., Reid, M. J., Menten, K. M., 2014, , 781, 108 Sato, M. 2008, , 60, 975 Sato, M., Hirota, T., Reid, M., 2010a, , 62, 287 Sato, M., Reid, M. J., Brunthaler, A. & Menten, K. M. 2010b, , 720, 1055 Sato, M., Wu, Y. W., Immer, K., 2014, submitted to Savchenko, S. S. & Reshetnikov, V. P. 2013, , 436, 1074 Schoenrich, R., Binney, J. & Dehnen, W. 2010, , 403, 1829 Shattow, G. & Loeb, A. 2009, , 392, 21 Shiozaki, S., Imai, H., Tafoya, D., 2011, , 63, 1219 Sivia, D. & Skilling, J. 2006, Data Analysis: A Bayesian Tutorial (2nd ed.; New York, Oxford Univ. Press), 168 Wang, J., Frenk, C. S., Navarro, J. F., Gao, L., Sawala, T. 2012, , 424, 2715 Weisberg, J. M., Stanimirović, S., Xilouris, K., 2008, , 674, 286 Weisberg, J. M., Nice, D. J., & Taylor, J. H. 2010, , 722, 1030 Wu, Y. W., Sato, M., Reid, M. J., 2014, submitted to Xin, X.-S. & Zheng, X.-W. 2013, [Res. Astron. Astroph.]{}, 13, 849 Xu, Y., Reid, M. J., Zheng, W. W. & Menten, K. M. 2006, [Science]{}, 311, 54 Xu, Y., Reid, M. J., Menten, K. M., Brunthaler, A., Zheng, X. W. & Moscadelli, L. 2009, , 693, 413 Xu, Y., Moscadelli, L., Reid, M. J., 2011, , 733, 25 Xu, Y., Li, J. J., Reid, M. J., 2013, , 769,15 Zhang, B., Zheng, X. W., Reid, M. J., 2009, , 693, 419 Zhang, B., Reid, M. J., Menten, K. M., & Zheng, X. W., 2012a, , 744, 23 Zhang, B., Reid, M. J., Menten, K. M., Zheng, X. W., Brunthaler, A., 2012b, , 544, 42 Zhang, B., Reid, M. J., Menten, K. M., 2013a, , 775, 79 Zhang, B., Moscadelli, L., Sato, M., 2014, , 781, 89 [^1]: http://bessel.vlbi-astrometry.org [^2]: http://veraserver.mtk.nao.ac.jp [^3]: Removing sources for which $R<4$ kpc: G$000.67-00.03$, G$009.62+00.19$, G$010.47+00.02$, G$010.62-00.38$, G$012.02-00.03$, G$023.43-00.18$, G$023.70-00.19$, G$027.36-00.16$ [^4]: Removing outlying sources: G012.68$-$00.18, G016.58$-$00.05, G023.65$-$00.12, G025.70$+$00.04, G028.86$+$00.06, G029.95$-$00.01, G031.28$+$00.06, G033.64$-$00.22, G034.39$+$00.22, G078.12$+$03.63, G108.59$+$00.49, G111.54+0.77, G122.01$-$07.08, G133.94+01.06, G176.51$+$00.20	{ "pile_set_name": "ArXiv" }
--- author: - 'Lorenzo Posti[^1], Filippo Fraternali' - Antonino Marasco bibliography: - 'refs.bib' date: 'Received XXX; accepted YYY' title: Peak star formation efficiency and no missing baryons in massive spirals --- Introduction {#sec:intro} ============ In our Universe, only about one-sixth of the total matter is baryonic, while the rest is widely thought to be in form of non-baryonic, collisionless, non-relativistic dark matter [e.g. @Planck18]. In the so-called standard $\Lambda$ Cold Dark Matter ($\Lambda$CDM) paradigm, galaxies form within extended haloes of dark matter that were able to grow enough to become gravitationally bound [e.g. @WhiteRees78]. In this scenario it is then reasonable to expect that the amount baryons present in galaxies today is roughly a fraction ${f_{\rm b}}\equiv\Omega_{\rm b}/{\Omega_{\rm c}}= 0.188$ [the “cosmological baryon fraction”, e.g. @Planck18] of the mass in dark matter. However, it was realised that the total amount of baryons that we can directly observe in galaxies (stars, gas, dust etc.) is instead at most only about 20$\%$ of the cosmological value [e.g. @PersicSalucci92; @Fukugita+98]. This became known as the “missing baryons” problem and has prompted the search for large resevoirs of baryons within the diffuse, multi-phase circumgalactic medium of galaxies [@Bregman07; @Tumlinson+17]. Arguably the most important indicator of this issue is the so-called stellar-to-halo mass relation, which connects the stellar mass ${M_\star}$ of a galaxy to its dark matter halo of mass ${M_{\rm halo}}$ . This relation can be probed observationally through many different techniques, e.g. galaxy abundance as a function of stellar mass [e.g. @ValeOstriker04; @Behroozi+10; @Moster+13], galaxy clustering [e.g. @Kravtsov+04; @Zheng+07], group catalogues [e.g. @Yang+08], weak galaxy-galaxy lensing [e.g. @Mandelbaum+06; @Leauthaud+12], satellite kinematics [e.g. @vandenBosch+04; @More+11; @WojtakMamon13] and internal galaxy dynamics [e.g. @Persic+96; @McConnachie12; @Cappellari+13; @DesmondWechsler15; @Read+17; @Katz+17 hereafter ]. Amongst all these determinations there is wide consensus on the overall shape of the relation and, in particular, on the fact that the ratio of stellar-to-halo mass ${f_\star}={M_\star}/{f_{\rm b}}{M_{\rm halo}}$, sometimes called star-formation efficiency, is a non-monotonic function of mass with a peak (${f_\star}\approx 0.2$) at ${M_{\rm halo}}\approx 10^{12}{M_\odot}$ (roughly the mass of the Milky Way). This can be interpreted as galaxies of these characteristic mass having been overall, during the course of their life, the most efficient at turning gas into stars. And yet, efficiencies of the order of $20\%$ are still relatively low, implying that most baryons are still undetected even in these systems[^2]. Several works have suggested that the exact shape of the stellar-to-halo mass relation depends on galaxy morphology [e.g. @Mandelbaum+06; @Conroy+07; @Dutton+10; @More+11; @Rodriguez-Puebla+15; @Lange+18], especially on the high-mass side ($\log\,{M_\star}/{M_\odot}\gtrsim 10$) where red, passive early-type systems appear to reside in more massive halos with respect to blue, star-forming late-type galaxies. This is intriguing, since it is suggesting that galaxies with different morphologies likely followed different evolutionary pathways that led the late-type ones, at a given M$_\star$, to live in relatively lighter halos and to have a somewhat smaller fraction of missing baryons with respect to early-type systems[^3]. However, one of the main difficulties associated to these measurements is the paucity of high-mass galaxies in the nearby Universe [e.g. @Kelvin+14], given that most of the aforementioned observational probes use statistical estimates based on on large galaxy samples. In this paper we use another, complementary approach to estimate the stellar-to-halo mass relation through accurate modelling of the gas dynamics within spiral galaxies. We use the observed [H[i]{}]{} rotation curves of a sample of regularly rotating, nearby disc galaxies to fit mass models comprising of a baryonic plus a dark matter component. We then extrapolate the dark matter profile to the virial radius, with cosmologically motivated assumptions, to yield the halo mass. A considerable advantage of this method is that each system can be studied individually and halo masses, along with their associated uncertainties, can be determined in great detail for each object. We show that this approach leads to a coherent picture of the relation between stellar and halo mass in late-type galaxies, which in turns profoundly affects our perspective on the star-formation efficiency in the high-mass regime. The paper is organised as follows: we present our sample and methodology to derive stellar and halo masses in Section \[sec:method\]; we describe our results in Section \[sec:results\] and we discuss them in detail in Section \[sec:discussion\]. Method {#sec:method} ====== Here we describe the data and methodology of our analysis. We adopt a standard $\Lambda$CDM cosmology, with parameters estimated by the [@Planck18]. In particular, we use a Hubble constant of $H_0=67.66$ km s$^{-1}$ Mpc$^{-1}$ and a cosmological baryon fraction of ${f_{\rm b}}\equiv\Omega_{\rm b}/{\Omega_{\rm c}}=0.188$. Data {#sec:data} ---- We use the sample of 175 disc galaxies with near-infrared photometry and [H[i]{}]{} rotation curves (SPARC) collected by @SPARC [hereafter ]. This sample of spirals in the nearby Universe spans more than 4 orders of magnitude in luminosity at 3.6$\mu$m and all morphological types, from irregulars to lenticulars. The galaxies have been selected to have extended, regular, high-quality [H[i]{}]{} rotation curves and measured near-infrared photometry; thus it is not volume limited. Nevertheless, it still provides a fair representation of the population of (regularly rotating) spirals at $z=0$ and most importantly is best suited for our dynamical study. The [H[i]{}]{} rotation curves are used as tracers of the circular velocity of the galaxies, while the individual contributions of the atomic gas (${V_{\rm gas}}$) and stars (${V_\star}$) to the circular velocity are derived from the [H[i]{}]{} and 3.6$\mu$m total intensity maps respectively . ${V_{\rm gas}}$ traces the distribution of atomic hydrogen, corrected for the presence of helium, while the near-infrared surface brightness is decomposed into and exponential disc (${V_{\rm disc}}$) and a spherical bulge (${V_{\rm bulge}}$). The contribution of the stars to the circular velocity is then $V_\star^2={\Upsilon_{\rm disc}}{V_{\rm disc}}^2 + {\Upsilon_{\rm bulge}}{V_{\rm bulge}}^2$, given stellar mass-to-light ratios of the disc (${\Upsilon_{\rm disc}}$) and bulge populations (${\Upsilon_{\rm bulge}}$) respectively. Model {#sec:model} ----- We model the observed rotation curve as $$\label{eq:vcirc} {V_{\rm c}}= \sqrt{{V_{\rm DM}}^2 + {V_{\rm gas}}^2 + {V_\star}^2}.$$ Here ${V_{\rm DM}}$ is the dark matter contribution to the circular velocity and, for simplicity, we have assumed that ${\Upsilon_{\rm bulge}}=1.4{\Upsilon_{\rm disc}}$, as suggested by stellar population synthesis models [e.g. @SchombertMcGaugh14], thus ${V_\star}^2 = {\Upsilon_{\rm disc}}\left({V_{\rm disc}}^2 + 1.4 {V_{\rm bulge}}^2\right)$. In Appendix \[appendix\] we explore the effect of fixing different mass-to-light ratios ${\Upsilon_{\rm disc}}$ and ${\Upsilon_{\rm bulge}}$ for disc and bulge respectively: our findings on the stellar-to-halo mass relation do not change significantly if we assume ${\Upsilon_{\rm disc}}=0.5$ and ${\Upsilon_{\rm bulge}}=0.7$, for which the scatter of the baryonic Tully-Fisher relation is minimised [@Lelli+16a]. The dark matter distribution is modelled as a @NFW [hereafter NFW] spherical halo, which is characterised by a dimensionless concentration parameter ($c$) and the halo mass (${M_{\rm halo}}$), which we take as that within a radius enclosing 200 times the critical density of the Universe. Thus our rotation curve model has three free parameters: ${M_{\rm halo}}$, $c$ and ${\Upsilon_\star}$. We compute the posterior distributions of these parameters with a Bayesian approach. We define a standard $\chi^2$ likelihood $\mathcal{P}$, given the data $\theta$, as $$\label{eq:likelihood} \begin{split} \chi^2 &= -\ln{\mathcal{P}(\theta\|{M_{\rm halo}},c,{\Upsilon_{\rm disc}})} \\ &= \sum_{i=0}^{N} \frac{1}{2}\left[\frac{{V_{{\rm obs},i}}- {V_{\rm c}}(R_i\|{M_{\rm halo}},c,{\Upsilon_{\rm disc}})}{\sigma_{{V_{{\rm obs},i}}}}\right]^2 \end{split}$$ where ${V_{{\rm obs},i}}$ is the $i$-th point of the observed rotation curve at radius $R_i$ and $\sigma_{{V_{{\rm obs},i}}}$ is its observed uncertainty. The posterior distribution of the three parameters is then given by Bayes’ theorem $$\label{eq:bayes} \mathcal{P}({M_{\rm halo}},c,{\Upsilon_{\rm disc}}\|\theta) \propto \mathcal{P}(\theta\|{M_{\rm halo}},c,{\Upsilon_{\rm disc}})\,\mathcal{P}({M_{\rm halo}},c,{\Upsilon_{\rm disc}})$$ where $\mathcal{P}({M_{\rm halo}},c,{\Upsilon_{\rm disc}})$ is the prior. We sample the posterior with an affine-invariant Markov Chain Monte Carlo method [MCMC, in particular, we use the `python` implementation by @emcee]. We use a flat prior on the stellar mass-to-light ratio ${\Upsilon_{\rm disc}}$ limited to a reasonable range, $0.01 \lesssim {\Upsilon_{\rm disc}}\lesssim 1.2$, which encompasses estimates obtained with stellar population models [@Meidt+14; @McGaughSchombert14]. In a $\Lambda$CDM Universe the halo mass and concentration are well known to be anti-correlated. Thus, in order to test whether standard $\Lambda$CDM haloes can be used to fit galaxy rotation curves and then yield a stellar-to-halo mass relation, for the halo concentration we assume a prior which follows the $c-{M_{\rm halo}}$ relation as estimated in N-body cosmological simulations [e.g. @DuttonMaccio14 hereafter ]: for each ${M_{\rm halo}}$, the prior on $c$ is lognormal with mean and uncertainty given by the $c=c({M_{\rm halo}})$ of (their Eq. 8). The prior on the dark matter halo mass ${M_{\rm halo}}$ is, instead, flat over a wide range: $6 \leq \log\,{M_{\rm halo}}/{M_\odot}\leq 15$. A non-uniform prior on the halo concentration is needed to infer reasonable constraints on the halo parameters (see e.g. ). The reason for this is that the [H[i]{}]{} rotation curves do not typically extend out enough to probe the region where the NFW density profile steepens, thus yielding only a weak inference on $c$. The $\Lambda$CDM-motivated prior on the $c-{M_{\rm halo}}$ relation proves to be enough to well constrain all the model parameters. Furthermore, we notice that the $c-{M_{\rm halo}}$ relation does not distinguish between haloes hosting late-type or early-type galaxies, so we use it under the assumption that it provides a reasonable description of the correlation for the haloes where late-type galaxies form. We summarise in Table \[tab:priors\] our choice of priors. [lcc]{}\ Parameter & Type &\ \ ${\Upsilon_\star}$ & uniform & $0.01 \leq {\Upsilon_\star}\leq 1.2$\ ${M_{\rm halo}}$ & uniform & $6 \leq \log{M_{\rm halo}}/{M_\odot}\leq 15$\ $c$ & lognormal & $c-{M_{\rm halo}}$ from\ Results {#sec:results} ======= We modelled the rotation curves and we have measured the posterior distributions of ${\Upsilon_{\rm disc}}, {M_{\rm halo}}$ and $c$ for all the 158 SPARC galaxies with inclination on the sky larger than 30 degrees – since for nearly face-on systems the rotation curves are very uncertain. For each parameter, we define the “best-value” to be the median of the posterior distribution and its uncertainty as the 16th - 84th percentiles. In Appendix \[appendix\] we provide in tabular form all the measurements and uncertainties, together with the value of the likelihood associated to the best model (Table \[tab:results\]). We also present the full rotation curve decomposition for one case as an example (NGC 3992, Figure \[fig:curve\_decomp\]), while we make available the plots of all the other galaxies online at <http://astro.u-strasbg.fr/~posti/PFM19_fiducial_fits/>. Unsurprisingly, we find that our model typically does not give very stringent constraints on the stellar mass-to-light ratio, with only 84 (68) galaxies having an uncertainty on ${\Upsilon_\star}$ smaller than 50$\%$ (30$\%$). In these cases, which are mostly for ${M_\star}>10^{10}{M_\odot}$ where the signal-to-noise is large, the $V_{\rm obs}$ and ${V_\star}$ profiles are similar enough to yield good constraints on ${\Upsilon_{\rm disc}}$. We find that these galaxies are not all maximal discs, as their ${\Upsilon_{\rm disc}}$ is homogeneously distributed in the range allowed by our prior. We find the highest mass spirals (${M_\star}\gtrsim 10^{11}{M_\odot}$) to have much better fits with a slightly larger mass-to-light ratio (${\Upsilon_{\rm disc}}\sim 0.7$) than the mean of our prior (${\Upsilon_{\rm disc}}=0.6$), consistently with previous works who found that high mass discs are close to maximal [e.g. @Lapi+18; @Starkman+18; @Li+18]. Smaller systems, instead, have typically a poorer inference on the mass-to-light ratio, with about $\sim 50$ cases in which the posterior on ${\Upsilon_{\rm disc}}$ is quite flat. Even in these extreme cases it is nevertheless useful to let the MCMC explore the full range of possible mass-to-light ratios ($0.01\leq{\Upsilon_{\rm disc}}\leq 1.2$) as opposed to just fixing a value for ${\Upsilon_{\rm disc}}$, because this provides a more realistic estimate of the uncertainty on the other parameters of the dark matter halo. In other words, when the inference on ${\Upsilon_{\rm disc}}$ is poor, it may be thought as a nuisance parameter over which the posterior distributions of the other two more interesting halo parameters is marginalised. For 137 (out of 158) galaxies we obtain a unimodal posterior distribution for the halo mass, thus we can associate a measurement and an uncertainty to ${M_{\rm halo}}$; the remaining 21 galaxies have, instead, either a multi-modal or a flat posterior on the halo mass and thus we discard them. These 21 are mostly low-mass systems (${M_\star}\lesssim 2\times 10^{9}{M_\odot}$) and their removal does not alter in any way the high-mass end of the population, which is the main focus of our work. For some of the remaining 137 galaxies, we find that the NFW halo model provides a poor fit to the observed rotation curve, as their best-fit $\chi^2$ is large. This is not surprising, since it is well known that especially low-mass discs tend to have slowly rising rotation curves, which makes them more compatible with having centrally cored haloes [e.g. @deBlok+01]. Indeed, by re-fitting all rotation curves with a cored halo model from @Burkert95, we have found 27, mostly low-mass (${M_\star}\lesssim 10^{10}{M_\odot}$), systems for which such cored profile is preferred to the NFW at a 3-$\sigma$ confidence level. For consistency we have decided to remove these 27 systems from our sample, but in Appendix \[appendix\] we demonstrate that their stellar and halo masses, derived by extrapolating the Burkert profile to the virial radius, are perfectly consistent with the picture that we present below. ![Stellar-to-halo mass relation for 110 galaxies in the SPARC sample. The points are colour coded by the ratio of [H[i]{}]{}-to-stellar mass. The stellar-to-halo mass relation estimated by [@Moster+13] using abundance matching is shown as a black dashed curve (with grey area representing the scatter of the relation). Galaxies that have converted all the available baryons in the halo into stars would lie on the long dashed line, whose thickness encompasses uncertainties on ${f_{\rm b}}$. For reference, we also show the location of the Milky Way (cross) and of the Andromeda galaxy (plus) on the plot, as given by the modelling by @PostiHelmi19 and @Corbelli+10, respectively.[]{data-label="fig:mstar_mhalo"}](figs/mstar-mhalo_colorfgas_newM13.pdf){width="50.00000%"} ![image](figs/mstar-fstar_fiducial_noBurk_whist.pdf){width="\textwidth"} In Figure \[fig:mstar\_mhalo\] we plot the ${M_\star}-{M_{\rm halo}}$ relation for the 110 SPARC galaxies in our final sample. Points are the median of the posterior distributions of ${M_{\rm halo}}$ and ${M_\star}$; the 16th-84th percentiles of the ${M_{\rm halo}}$ distribution define the errorbar, while the uncertainty on the stellar mass is calculated as in @Lelli+16a [their Eq. 5] where the uncertainty on ${\Upsilon_{\rm disc}}$ is given by the 16th-84th percentiles of its posterior. For comparison we also plot the ${M_\star}-{M_{\rm halo}}$ relation estimated by [@Moster+13] using abundance matching. In general we find that the abundance matching model is in good agreement with our measurements for ${M_\star}\lesssim 5\times 10^{10}{M_\odot}$, albeit our points have a large scatter especially at the lowest masses. The agreement is instead much poorer at high stellar masses, where the @Moster+13 model predicts significantly larger halo masses with respect to our estimates. Our measurements indicate that there is no sign of a break in the stellar-to-halo mass relation of spirals and that it is consistent with being an increasing function of mass with roughly the same slope at all masses. The tension at the high-mass end between our measurements and the abundance matching model is much clearer if one plots the stellar fraction, i.e. ${f_\star}\equiv{M_\star}/{f_{\rm b}}{M_{\rm halo}}$, also sometimes called star-formation efficiency, as a function of the stellar mass: we show this in Figure \[fig:mstar\_fstar\]. This plot highlights the two main findings of our work, the first being that ${f_\star}$ appears to increase monotonically with galaxy stellar mass with no indication of a peak in the range $10\leq\log\,{M_\star}/{M_\odot}\leq 11$, where most abundance matching models find a maximum star-formation efficiency. For instance, a galaxy with ${M_\star}=2\times 10^{11}{M_\odot}$ has ${f_\star}\simeq 0.04$ in the [@Moster+13] model, while we find ${f_\star}\simeq 0.5$. By computing the difference between the measured ${f_\star}$ and that expected in the [@Moster+13] model, normalised by the sum in quadrature of the measured uncertainty on ${f_\star}$ and of the intrinsic scatter of the model, we find the measurement for the high-mass systems to be inconsistent at $2-3\sigma$ with the model (see the colours of the points in Fig. \[fig:mstar\_fstar\]). Such a discrepancy is very robust and holds for all the tests we have run (we show in Appendix \[appendix\], Figure \[fig:tests\], the ${f_\star}-{M_\star}$ diagram in all these cases): - we have fitted the rotation curves assuming a cored [@Burkert95] instead of a cuspy (NFW) profile. In general, this yields better fits for many low-mass systems, slightly larger stellar masses and smaller halo virial masses for all galaxies; - we have used the fits recently obtained by [@Ghari+19], who used [@Einasto65] halo profiles [and distances and mass-to-light ratios from @Li+18]. In general, we typically find slightly smaller halo virial masses, but broadly consistent with our estimates with NFW profiles; - we have fixed the mass-to-light ratio of the bulge and disc components to reasonable values suggested by stellar population synthesis models [${\Upsilon_{\rm disc}}=0.5,{\Upsilon_{\rm bulge}}=0.7$, see e.g. @Meidt+14; @SchombertMcGaugh14]; - we tried allowing both ${\Upsilon_{\rm disc}}$ and ${\Upsilon_{\rm bulge}}$ to vary in our fits, with the additional constraint (${\Upsilon_{\rm disc}}\leq{\Upsilon_{\rm bulge}}$). This has an effect only on the 28 (out of 110) galaxies in our final sample that have non-negligible bulges. We find the resulting uncertainties on ${\Upsilon_{\rm disc}}$ to be typically significantly larger in this case, but never dramatic. In all these cases the final result is that the ${f_\star}-{M_\star}$ diagram is not significantly different from the one presented in Fig. \[fig:mstar\_fstar\]. Additionally, as shown by @Katz+14 [see their Fig. 20 and 23], the effect of adiabatic contraction of the dark matter halos due to the formation of stellar discs has a negligible impact on $f_\star$ for galaxies in the interested mass regime. The other main finding highlighted by Fig. \[fig:mstar\_fstar\] is even more surprising: we find that all spirals with ${M_\star}\gtrsim10^{11}{M_\odot}$ have stellar fraction very close to unity, in the range ${f_\star}\approx 0.3-1$, with a handful of them being consistent with ${f_\star}=1$ within the uncertainties. This implies that these galaxies were extremely efficient at turning gas into stars and that the amount of mass collapsed in stars is a considerable portion of the total amount of baryons expected to be associated with their haloes. In fact, if we include also the contribution of atomic and molecular hydrogen , spirals with ${M_\star}\geq 10^{11}{M_\odot}$ are found to be consistent with a cold baryon budget of $f_{\rm baryons}={f_\star}+1.4{f_{\rm HI}}+{f_{{\rm H}_2}}\approx 1$ within the uncertainties [where the factor 1.4 accounts for helium, e.g. @SPARC], with a mean value of $\sim 0.6$ and uncertainties of $[-0.3, +0.5]$. Moreover, considering that galaxies are known to be surrounded by massive, hot coronae, which are detected both in X-ray and with the Sunyaev-Zeldovich effect and account for about $0.1-0.3{f_{\rm b}}{M_{\rm halo}}$ [typically estimated statistically by stacking over many galaxies with a given stellar mass, e.g. @Planck13XI; @Bregman+18 and references therein], the total (cold+hot) baryon budget is easily compatible with unity at the high-mass end, with very little room for other baryonic components. In other words, we have found that the most massive, regularly rotating spirals in the nearby Universe have virtually no missing baryons. Discussion {#sec:discussion} ========== Our analysis allowed us to have a robust and unbiased estimate of the halo virial mass for a sample of 108 spiral galaxies in the nearby Universe using their high-quality [H[i]{}]{} rotation curves. While we find good agreement with previous determinations of the stellar-to-halo mass relation for galaxies roughly up to the mass of Milky Way (${M_\star}=5\times 10^{10}{M_\odot}$), we also find systematically smaller halo masses (factor $\sim 10$), corresponding to higher stellar-to-halo mass ratios, for the most massive spirals with respect to expectations from most up-to-date abundance matching models [e.g. @WechslerTinker18]. A possible explanation for this discrepancy is that, while the high-mass end (${M_\star}\gtrsim10^{11}{M_\odot}$) of the galaxy stellar mass function is vastly dominated by passive early-type galaxies which occupy massive (${M_{\rm halo}}\gtrsim 5\times 10^{12}{M_\odot}$) dark matter halos, there still exists a population of star-forming spirals that inhabit halos of lower masses. The presence of this second population – which is not well represented by current abundance matching models – implies the existence of different evolutionary pathways for building galaxies of a given stellar mass. This suggests that e.g. a massive system that has evolved in isolation may have had the chance to sustain star-formation unimpeded for its entire life, potentially converting most of its available baryons into stars. While this is certainly not the case for high-mass early-types galaxies, which tend to live in high-density environments, it may well be the pathway taken by the high-mass population of spirals studied in this work. In fact, also [@McGaugh+10] by simply analysing the Tully-Fisher relation of a similar sample of spirals concluded that ${f_\star}$ does not turn over at the highest masses. A discrepancy between the expected halo mass for a typical passive (red) $10^{11}{M_\odot}$ galaxy and an active (blue) one of the same ${M_\star}$, was also noted by other authors using various probes, such as satellite kinematics [e.g. @Conroy+07; @More+11; @WojtakMamon13], galaxy-galaxy weak lensing [e.g. @Mandelbaum+06; @Mandelbaum+16; @Reyes+12], abundance matching [e.g. @Rodriguez-Puebla+15] or combinations [e.g. @Dutton+10]. The works most similar to ours are those of and [@Lapi+18]. We use the same galaxy sample as in (SPARC) and we perform a similar analysis as them, but with the crucial difference that we do not impose a prior on halo mass that follows an ${M_\star}-{M_{\rm halo}}$ relation from abundance matching, which slightly biases towards higher halo masses some of the high-mass galaxies[^4]. [@Lapi+18], on the other hand, have a much larger sample of spirals than ours, but they rely on “stacked” rotation curves for their mass decompositions – i.e. they stack individual curves of galaxies in bins of absolute magnitude – whereas we focus on individual, well studied systems. Finally, we notice that, amongst the detailed studies of individual systems, i) [@Corbelli+10] measured the dynamical mass of M31 by decomposing its [H[i]{}]{} rotation cureve, to find a surprisingly high ${f_\star}\simeq 0.6$, and ii) [@Martinsson+13] decomposed the [H[i]{}]{} rotation curves of a small sample of 30 spirals from the DiskMass Survey, to find the highest star-formation efficiencies ${f_\star}\gtrsim 0.3$ for their three most massive galaxies ($\log{M_\star}/{M_\odot}\gtrsim 10.9$). While our results align with these previous works, to our knowledge we are the first to focus specifically on the ${f_\star}-{M_\star}$ relation and to highlight the fact that i) the highest-mass spirals are the most efficient galaxies at turning gas into stars, ii) that ${f_\star}$ increases monotonically with stellar mass for regularly rotating nearby discs and that iii) virtually all high-mass discs have $\gtrsim 30\%$ of the total baryons within their haloes in stars. Our analysis establishes that the most efficient galaxies at forming stars are not $L^\ast$ galaxies, as previously thought [e.g. @WechslerTinker18], but much more massive systems, some of the most massive spiral galaxies in the nearby Universe (${M_\star}\gtrsim 10^{11}{M_\odot}$). Not only the galactic star-formation efficiency peaks at much larger masses than we knew before, but we also showed that several massive discs have efficiencies ${f_\star}$ of the order unity. This result alone is of key importance since it demonstrates that there is no universal physical mechanism that sets the maximum star-formation efficiency to $20-30\%$. Furthermore, the fact that some massive galaxies with high ${f_\star}$ exist has fundamental implications for star-formation quenching. Since these galaxies live in haloes with ${M_{\rm halo}}\sim 2-5\times 10^{12}{M_\odot}$, if mass is the main driver of quenching and if a critical mass for quenching exists [e.g. as expected in scenarios where virial shock heating of the circumgalactic medium is the key process, see @BirnboimDekel03; @DekelBirnboim08], then it follows that this critical mass can not be smaller than $\sim 5\times 10^{12}{M_\odot}$, which is almost an order of magnitude higher than previously thought [e.g. @DekelBirnboim06]. Interestingly, such a high threshold is instead expected in scenarios where the accretion of cool gas is hampered (“starvation”), e.g. by the high virial temperature of the circumgalactic gas in a galactic fountain cycle [e.g. @Armillotta+16] or by the complex interplay of radiative cooling and feedback in the smooth gas accretion from cold filaments [e.g. @vandeVoort+11]. Even if we have measured high ${f_\star}$ for some massive spirals, still the vast majority of galaxies living in ${M_{\rm halo}}>10^{12}{M_\odot}$ haloes has ${f_\star}\ll 1$, which means that they managed to efficiently quench their star-formation. Our results imply that since mass can not be the major player in quenching galaxies, at least for ${M_{\rm halo}}\lesssim 5\times 10^{12}{M_\odot}$, and some other mechanism must play a fundamental role in the transition from actively to passively star-forming. One of the main suspects is clearly environment, since gas removal happens more frequently and also gas accretion is more difficult in high-density environments [e.g. @Peng+10; @vandeVoort+17]. Another is the powerful feedback from the active galactic nucleus (AGN), which can episodically suppress any gas condensation throughout the galaxy [e.g. @Croton+06; @Fabian12]. Finally, another key process is the interaction with other galaxies, with passive galaxies being hosted in haloes with an active merger history, which can result in bursty star-formation histories and subsequent suppressive stellar/AGN feedback [e.g. @Cox+06b; @Gabor+10]. This latter scenario also naturally accounts for the morphological transformation of disc galaxies, living in haloes with quiet merger histories, to spheroids, which are the dominant galaxy population at the high-mass end, where also mergers are more frequent [e.g. @Cox+06a]. This scenario is, in principle, testable both with current cosmological simulations and with a new abundance matching model which depends also on secondary halo parameters, such as merger history or formation time, and it is able to predict not only stellar masses but also other galaxy properties, such as morphology or colour. Acknowledgements {#acknowledgements .unnumbered} ================ We thank E. Corbelli, B. Famaey, A. Lapi, F. Lelli, A. Robertson, J. Sellwood and F. van den Bosch for useful discussions and A. Ghari for making their Einasto fits available to us. LP acknowledges financial support from a VICI grant from the Netherlands Organisation for Scientific Research (NWO) and from the Centre National d’Etudes Spatiales (CNES). Supplementary material {#appendix} ====================== [lcccccccccccccc]{} \ Name & $\log\,L_{[3.6]}$ & ${\Upsilon_{\rm disc}}$ & 16th & 84th & $\log\,{M_{\rm halo}}$ & 16th & 84th & $\log\,c$ & 16th & 84th & $\chi^2_{\rm red}$ & ${f_\star}$ & 16th & 84th\ \ [[ – continued from previous page]{}]{}\ Name & $\log\,L_{[3.6]}$ & ${\Upsilon_{\rm disc}}$ & 16th & 84th & $\log\,{M_{\rm halo}}$ & 16th & 84th & $\log\,c$ & 16th & 84th & $\chi^2_{\rm red}$ & ${f_\star}$ & 16th & 84th\ \ \ D512-2 & 8.51 & 0.62 & 0.22 & 1.02 & 9.91 & 9.59 & 10.26 & 0.98 & 0.86 & 1.11 & 1.05 & 0.0852 & 0.0157 & 0.2845\ DDO064 & 8.20 & 0.60 & 0.21 & 1.00 & 10.29 & 9.76 & 10.92 & 1.00 & 0.83 & 1.17 & 1.07 & 0.0237 & 0.0029 & 0.1213\ DDO170 & 8.73 & 0.38 & 0.12 & 0.80 & 10.66 & 10.58 & 10.76 & 0.82 & 0.74 & 0.88 & 2.73 & 0.0215 & 0.0052 & 0.0431\ ESO116-G012 & 9.63 & 0.44 & 0.20 & 0.69 & 11.72 & 11.49 & 12.05 & 0.89 & 0.74 & 1.01 & 2.52 & 0.0181 & 0.0064 & 0.0374\ ESO444-G084 & 7.85 & 0.60 & 0.21 & 0.99 & 11.23 & 10.93 & 11.65 & 0.92 & 0.79 & 1.03 & 0.76 & 0.0011 & 0.0003 & 0.0030\ F565-V2 & 8.75 & 0.58 & 0.19 & 1.00 & 11.14 & 10.88 & 11.51 & 0.89 & 0.73 & 1.02 & 1.10 & 0.0103 & 0.0026 & 0.0280\ F568-V1 & 9.58 & 0.70 & 0.27 & 1.05 & 11.63 & 11.30 & 12.07 & 1.01 & 0.84 & 1.15 & 0.30 & 0.0314 & 0.0077 & 0.0919\ F574-1 & 9.82 & 0.68 & 0.27 & 1.03 & 11.29 & 11.07 & 11.55 & 0.92 & 0.80 & 1.03 & 1.84 & 0.1097 & 0.0312 & 0.2466\ F583-1 & 8.99 & 0.57 & 0.18 & 0.98 & 11.08 & 10.77 & 11.42 & 0.86 & 0.73 & 0.98 & 2.11 & 0.0255 & 0.0058 & 0.0731\ F583-4 & 9.23 & 0.62 & 0.21 & 1.02 & 10.61 & 10.31 & 10.98 & 0.98 & 0.82 & 1.11 & 0.44 & 0.1377 & 0.0263 & 0.4602\ NGC0024 & 9.59 & 1.02 & 0.77 & 1.15 & 11.27 & 11.06 & 11.56 & 1.06 & 0.91 & 1.20 & 0.66 & 0.1306 & 0.0670 & 0.2269\ NGC0100 & 9.51 & 0.29 & 0.09 & 0.60 & 11.36 & 11.02 & 11.76 & 0.85 & 0.70 & 0.97 & 1.20 & 0.0208 & 0.0024 & 0.0820\ NGC0247 & 9.87 & 0.64 & 0.25 & 1.01 & 11.35 & 11.09 & 11.62 & 0.82 & 0.72 & 0.90 & 2.14 & 0.1013 & 0.0280 & 0.2969\ NGC0289 & 10.86 & 0.59 & 0.43 & 0.76 & 11.83 & 11.74 & 11.94 & 0.91 & 0.77 & 1.05 & 1.95 & 0.2833 & 0.1765 & 0.4227\ NGC0300 & 9.47 & 0.46 & 0.17 & 0.79 & 11.37 & 11.18 & 11.63 & 0.89 & 0.75 & 1.01 & 0.72 & 0.0268 & 0.0102 & 0.0573\ NGC0801 & 11.49 & 0.56 & 0.52 & 0.60 & 12.00 & 11.90 & 12.14 & 0.77 & 0.63 & 0.90 & 6.80 & 1.0564 & 0.7746 & 1.3789\ NGC1003 & 9.83 & 0.46 & 0.24 & 0.66 & 11.49 & 11.39 & 11.62 & 0.78 & 0.67 & 0.88 & 3.09 & 0.0485 & 0.0240 & 0.0746\ NGC1090 & 10.86 & 0.48 & 0.36 & 0.59 & 11.72 & 11.63 & 11.84 & 0.94 & 0.80 & 1.07 & 2.50 & 0.3931 & 0.2445 & 0.5734\ NGC1705 & 8.73 & 0.99 & 0.72 & 1.15 & 10.86 & 10.57 & 11.26 & 1.16 & 0.99 & 1.31 & 0.66 & 0.0352 & 0.0133 & 0.0712\ NGC2403 & 10.00 & 0.42 & 0.30 & 0.53 & 11.40 & 11.33 & 11.49 & 1.14 & 1.06 & 1.23 & 9.47 & 0.1012 & 0.0828 & 0.1164\ NGC2683 & 10.91 & 0.66 & 0.58 & 0.73 & 11.63 & 11.46 & 11.82 & 0.96 & 0.82 & 1.11 & 1.31 & 0.4620 & 0.2927 & 0.7192\ NGC2841 & 11.27 & 0.87 & 0.79 & 0.94 & 12.54 & 12.42 & 12.69 & 0.88 & 0.76 & 1.00 & 1.81 & 0.1796 & 0.1335 & 0.2283\ NGC2903 & 10.91 & 0.37 & 0.31 & 0.41 & 11.75 & 11.67 & 11.85 & 1.24 & 1.14 & 1.34 & 7.61 & 0.3001 & 0.1718 & 0.4448\ NGC2915 & 8.81 & 0.56 & 0.19 & 0.97 & 11.10 & 10.85 & 11.44 & 1.03 & 0.86 & 1.18 & 0.98 & 0.0106 & 0.0030 & 0.0270\ NGC2955 & 11.50 & 0.47 & 0.44 & 0.51 & 12.13 & 11.80 & 12.48 & 0.88 & 0.71 & 1.03 & 4.81 & 0.6863 & 0.3052 & 1.4911\ NGC2998 & 11.18 & 0.62 & 0.48 & 0.74 & 12.01 & 11.91 & 12.13 & 0.91 & 0.76 & 1.06 & 2.74 & 0.5532 & 0.3896 & 0.7568\ NGC3198 & 10.58 & 0.51 & 0.38 & 0.61 & 11.67 & 11.60 & 11.75 & 0.98 & 0.87 & 1.09 & 1.43 & 0.2475 & 0.1981 & 0.2979\ NGC3521 & 10.93 & 0.52 & 0.47 & 0.58 & 12.29 & 11.83 & 12.85 & 0.86 & 0.68 & 1.03 & 0.29 & 0.1212 & 0.0315 & 0.3787\ NGC3726 & 10.85 & 0.39 & 0.28 & 0.47 & 11.76 & 11.59 & 11.98 & 0.87 & 0.73 & 1.02 & 2.96 & 0.1987 & 0.1058 & 0.3469\ NGC3741 & 7.45 & 0.46 & 0.14 & 0.89 & 10.57 & 10.33 & 10.86 & 0.84 & 0.72 & 0.95 & 1.05 & 0.0013 & 0.0004 & 0.0031\ NGC3769 & 10.27 & 0.35 & 0.21 & 0.51 & 11.40 & 11.25 & 11.57 & 1.01 & 0.88 & 1.14 & 0.68 & 0.0970 & 0.0495 & 0.1719\ NGC3893 & 10.77 & 0.50 & 0.41 & 0.58 & 12.01 & 11.75 & 12.36 & 0.95 & 0.78 & 1.11 & 1.27 & 0.1227 & 0.0551 & 0.2310\ NGC3972 & 10.16 & 0.40 & 0.14 & 0.73 & 12.03 & 11.57 & 12.52 & 0.86 & 0.70 & 0.98 & 1.19 & 0.0240 & 0.0038 & 0.1250\ NGC3992 & 11.36 & 0.82 & 0.69 & 0.93 & 12.15 & 12.03 & 12.30 & 0.90 & 0.74 & 1.05 & 0.85 & 0.4339 & 0.3037 & 0.6160\ NGC4010 & 10.24 & 0.25 & 0.09 & 0.45 & 11.96 & 11.62 & 12.36 & 0.81 & 0.68 & 0.95 & 2.44 & 0.0216 & 0.0045 & 0.0764\ NGC4013 & 10.90 & 0.48 & 0.41 & 0.54 & 11.98 & 11.81 & 12.19 & 0.85 & 0.70 & 0.99 & 1.31 & 0.0776 & 0.0483 & 0.1161\ NGC4088 & 11.03 & 0.31 & 0.24 & 0.37 & 11.77 & 11.54 & 12.05 & 0.91 & 0.74 & 1.06 & 0.57 & 0.2924 & 0.1393 & 0.5817\ NGC4100 & 10.77 & 0.74 & 0.61 & 0.85 & 11.69 & 11.48 & 11.93 & 0.97 & 0.81 & 1.12 & 1.27 & 0.4199 & 0.2325 & 0.7423\ NGC4138 & 10.64 & 0.68 & 0.58 & 0.80 & 11.46 & 11.09 & 11.82 & 0.99 & 0.82 & 1.16 & 1.68 & 0.2491 & 0.1029 & 0.6693\ NGC4157 & 11.02 & 0.40 & 0.32 & 0.48 & 11.95 & 11.74 & 12.22 & 0.89 & 0.73 & 1.04 & 0.55 & 0.2388 & 0.1231 & 0.4311\ NGC4183 & 10.03 & 0.75 & 0.38 & 1.04 & 11.16 & 10.97 & 11.35 & 1.01 & 0.87 & 1.13 & 0.18 & 0.3102 & 0.1236 & 0.6488\ NGC4559 & 10.29 & 0.38 & 0.20 & 0.55 & 11.41 & 11.23 & 11.61 & 0.95 & 0.81 & 1.09 & 0.24 & 0.1513 & 0.0527 & 0.3176\ NGC5033 & 11.04 & 0.40 & 0.31 & 0.48 & 11.91 & 11.86 & 11.96 & 1.23 & 1.14 & 1.31 & 3.81 & 0.3049 & 0.1732 & 0.4383\ NGC5055 & 11.18 & 0.32 & 0.29 & 0.34 & 11.82 & 11.79 & 11.85 & 1.12 & 1.06 & 1.18 & 2.75 & 0.4220 & 0.3913 & 0.4514\ NGC5371 & 11.53 & 0.44 & 0.34 & 0.53 & 11.64 & 11.53 & 11.74 & 1.21 & 1.02 & 1.34 & 6.59 & 1.9570 & 1.1181 & 3.1110\ NGC5585 & 9.47 & 0.18 & 0.08 & 0.30 & 11.33 & 11.18 & 11.52 & 0.90 & 0.79 & 0.98 & 5.85 & 0.0142 & 0.0006 & 0.0294\ NGC5907 & 11.24 & 0.68 & 0.56 & 0.78 & 12.02 & 11.93 & 12.16 & 0.89 & 0.71 & 1.07 & 6.38 & 0.5110 & 0.4049 & 0.6183\ NGC5985 & 11.32 & 0.45 & 0.26 & 0.65 & 12.21 & 12.12 & 12.28 & 1.37 & 1.30 & 1.44 & 2.85 & 0.3156 & 0.1265 & 0.5595\ NGC6015 & 10.51 & 0.78 & 0.65 & 0.87 & 11.67 & 11.52 & 11.88 & 0.94 & 0.77 & 1.10 & 8.45 & 0.3054 & 0.1972 & 0.4377\ NGC6195 & 11.59 & 0.46 & 0.42 & 0.48 & 12.16 & 11.94 & 12.42 & 0.79 & 0.64 & 0.93 & 3.44 & 0.6961 & 0.3866 & 1.1779\ NGC6503 & 10.11 & 0.45 & 0.36 & 0.53 & 11.28 & 11.21 & 11.36 & 1.11 & 1.02 & 1.19 & 1.61 & 0.1585 & 0.1316 & 0.1883\ NGC6674 & 11.33 & 0.94 & 0.83 & 1.03 & 12.42 & 12.32 & 12.56 & 0.65 & 0.52 & 0.77 & 3.87 & 0.3996 & 0.2914 & 0.5274\ NGC6946 & 10.82 & 0.44 & 0.38 & 0.48 & 11.83 & 11.62 & 12.12 & 0.95 & 0.79 & 1.09 & 1.88 & 0.2336 & 0.1103 & 0.4250\ NGC7331 & 11.40 & 0.36 & 0.33 & 0.40 & 12.38 & 12.21 & 12.60 & 0.85 & 0.71 & 0.98 & 0.80 & 0.1527 & 0.0945 & 0.2232\ NGC7814 & 10.87 & 0.50 & 0.43 & 0.56 & 12.21 & 12.01 & 12.50 & 1.01 & 0.86 & 1.15 & 1.30 & 0.1245 & 0.0688 & 0.1869\ UGC00128 & 10.08 & 0.53 & 0.18 & 0.92 & 11.56 & 11.53 & 11.59 & 0.93 & 0.86 & 0.99 & 3.19 & 0.1058 & 0.0370 & 0.1797\ UGC00191 & 9.30 & 0.83 & 0.51 & 1.08 & 10.96 & 10.87 & 11.10 & 0.93 & 0.82 & 1.02 & 3.68 & 0.0947 & 0.0586 & 0.1368\ UGC00731 & 8.51 & 0.59 & 0.19 & 1.01 & 10.77 & 10.64 & 10.91 & 0.99 & 0.91 & 1.08 & 0.36 & 0.0176 & 0.0051 & 0.0338\ UGC02259 & 9.24 & 0.86 & 0.46 & 1.11 & 10.78 & 10.69 & 10.89 & 1.23 & 1.15 & 1.31 & 1.37 & 0.1220 & 0.0610 & 0.1851\ UGC02487 & 11.69 & 0.98 & 0.85 & 1.08 & 12.58 & 12.52 & 12.67 & 0.94 & 0.81 & 1.06 & 5.28 & 0.3968 & 0.3302 & 0.4704\ UGC02885 & 11.61 & 0.63 & 0.55 & 0.72 & 12.62 & 12.48 & 12.79 & 0.75 & 0.62 & 0.88 & 1.47 & 0.3448 & 0.2284 & 0.5073\ UGC02916 & 11.09 & 0.34 & 0.31 & 0.36 & 12.10 & 11.93 & 12.31 & 1.05 & 0.95 & 1.15 & 10.88 & 0.2354 & 0.1404 & 0.3645\ UGC02953 & 11.41 & 0.56 & 0.51 & 0.60 & 12.29 & 12.22 & 12.36 & 1.11 & 1.02 & 1.20 & 6.78 & 0.4796 & 0.3421 & 0.6312\ UGC03205 & 11.06 & 0.72 & 0.64 & 0.79 & 12.12 & 11.95 & 12.33 & 0.85 & 0.70 & 1.01 & 3.51 & 0.4040 & 0.2531 & 0.5862\ UGC03546 & 11.01 & 0.41 & 0.34 & 0.46 & 11.92 & 11.80 & 12.06 & 1.07 & 0.96 & 1.18 & 1.52 & 0.2236 & 0.1352 & 0.3344\ UGC03580 & 10.12 & 0.18 & 0.13 & 0.22 & 11.52 & 11.42 & 11.64 & 0.95 & 0.87 & 1.04 & 3.52 & 0.0459 & 0.0121 & 0.0823\ UGC04278 & 9.12 & 0.36 & 0.10 & 0.76 & 11.41 & 11.00 & 11.89 & 0.80 & 0.65 & 0.94 & 2.19 & 0.0095 & 0.0011 & 0.0430\ UGC04483 & 7.11 & 0.52 & 0.17 & 0.93 & 9.30 & 8.97 & 9.74 & 1.11 & 0.95 & 1.26 & 0.74 & 0.0160 & 0.0038 & 0.0485\ UGC04499 & 9.19 & 0.34 & 0.11 & 0.69 & 10.89 & 10.70 & 11.12 & 0.93 & 0.81 & 1.04 & 0.95 & 0.0322 & 0.0070 & 0.0839\ UGC05005 & 9.61 & 0.36 & 0.10 & 0.78 & 11.10 & 10.84 & 11.36 & 0.85 & 0.71 & 0.97 & 1.11 & 0.0718 & 0.0151 & 0.2207\ UGC05253 & 11.23 & 0.46 & 0.43 & 0.48 & 12.16 & 12.08 & 12.27 & 1.05 & 0.98 & 1.12 & 3.22 & 0.3759 & 0.2567 & 0.5165\ UGC05414 & 9.05 & 0.20 & 0.06 & 0.46 & 11.17 & 10.82 & 11.57 & 0.77 & 0.64 & 0.89 & 1.68 & 0.0061 & 0.0002 & 0.0256\ UGC05716 & 8.77 & 0.44 & 0.15 & 0.83 & 10.81 & 10.75 & 10.89 & 0.98 & 0.91 & 1.03 & 1.76 & 0.0186 & 0.0062 & 0.0312\ UGC05721 & 8.73 & 0.93 & 0.60 & 1.12 & 10.91 & 10.68 & 11.23 & 1.17 & 1.01 & 1.30 & 1.90 & 0.0317 & 0.0142 & 0.0596\ UGC05829 & 8.75 & 0.59 & 0.18 & 1.01 & 10.47 & 10.16 & 10.83 & 0.95 & 0.80 & 1.09 & 0.84 & 0.0539 & 0.0106 & 0.1593\ UGC05918 & 8.37 & 0.63 & 0.21 & 1.02 & 10.07 & 9.81 & 10.43 & 1.04 & 0.89 & 1.17 & 0.35 & 0.0580 & 0.0124 & 0.1611\ UGC06399 & 9.36 & 0.61 & 0.22 & 0.99 & 11.27 & 10.95 & 11.67 & 0.89 & 0.75 & 1.02 & 0.97 & 0.0362 & 0.0077 & 0.1135\ UGC06446 & 8.99 & 0.75 & 0.32 & 1.08 & 10.96 & 10.75 & 11.23 & 1.06 & 0.92 & 1.18 & 0.22 & 0.0385 & 0.0133 & 0.0808\ UGC06614 & 11.09 & 0.27 & 0.17 & 0.36 & 12.20 & 12.03 & 12.41 & 0.83 & 0.68 & 0.96 & 0.44 & 0.0828 & 0.0428 & 0.1474\ UGC06667 & 9.15 & 0.63 & 0.21 & 1.03 & 11.41 & 11.18 & 11.72 & 0.88 & 0.76 & 0.98 & 1.57 & 0.0113 & 0.0029 & 0.0275\ UGC06786 & 10.87 & 0.57 & 0.49 & 0.65 & 12.22 & 12.10 & 12.37 & 1.05 & 0.94 & 1.16 & 1.47 & 0.1669 & 0.1166 & 0.2240\ UGC06787 & 10.99 & 0.43 & 0.38 & 0.47 & 12.17 & 12.10 & 12.24 & 1.19 & 1.12 & 1.26 & 27.20 & 0.2041 & 0.1410 & 0.2737\ UGC06917 & 9.83 & 0.46 & 0.18 & 0.78 & 11.46 & 11.23 & 11.77 & 0.93 & 0.79 & 1.05 & 0.75 & 0.0438 & 0.0137 & 0.1163\ UGC06923 & 9.46 & 0.30 & 0.11 & 0.59 & 11.20 & 10.83 & 11.68 & 0.94 & 0.78 & 1.08 & 0.85 & 0.0194 & 0.0035 & 0.0809\ UGC06930 & 9.95 & 0.68 & 0.28 & 1.02 & 11.15 & 10.93 & 11.38 & 0.99 & 0.86 & 1.12 & 0.33 & 0.2057 & 0.0617 & 0.4919\ UGC06973 & 10.73 & 0.18 & 0.16 & 0.20 & 12.83 & 12.24 & 13.53 & 0.86 & 0.65 & 1.06 & 1.11 & 0.0032 & 0.0006 & 0.0126\ UGC06983 & 9.72 & 0.76 & 0.38 & 1.06 & 11.31 & 11.11 & 11.57 & 1.00 & 0.85 & 1.13 & 0.70 & 0.0767 & 0.0301 & 0.1557\ UGC07089 & 9.55 & 0.44 & 0.13 & 1.05 & 10.68 & 9.71 & 11.15 & 0.91 & 0.75 & 1.13 & 1.01 & 0.1587 & 0.0203 & 3.9876\ UGC07125 & 9.43 & 0.28 & 0.09 & 0.57 & 10.46 & 10.33 & 10.60 & 0.91 & 0.81 & 1.01 & 1.08 & 0.1392 & 0.0239 & 0.3678\ UGC07151 & 9.36 & 0.84 & 0.58 & 1.06 & 10.77 & 10.45 & 11.14 & 0.95 & 0.80 & 1.07 & 2.64 & 0.1613 & 0.0586 & 0.4149\ UGC07399 & 9.06 & 0.84 & 0.45 & 1.10 & 11.39 & 11.17 & 11.70 & 1.13 & 1.01 & 1.23 & 1.74 & 0.0163 & 0.0066 & 0.0315\ UGC07524 & 9.39 & 0.50 & 0.17 & 0.94 & 11.00 & 10.77 & 11.27 & 0.87 & 0.75 & 0.97 & 0.94 & 0.0657 & 0.0155 & 0.1930\ UGC07559 & 8.04 & 0.53 & 0.15 & 1.02 & 9.31 & 8.70 & 9.76 & 1.08 & 0.92 & 1.27 & 1.29 & 0.1263 & 0.0185 & 1.0510\ UGC07603 & 8.58 & 0.53 & 0.20 & 0.88 & 11.01 & 10.70 & 11.44 & 0.97 & 0.82 & 1.11 & 1.62 & 0.0084 & 0.0021 & 0.0224\ UGC07690 & 8.93 & 0.89 & 0.66 & 1.08 & 10.18 & 9.87 & 10.53 & 1.09 & 0.94 & 1.25 & 0.48 & 0.1986 & 0.0754 & 0.4751\ UGC07866 & 8.09 & 0.66 & 0.22 & 1.06 & 9.31 & 8.78 & 9.80 & 1.14 & 0.97 & 1.30 & 0.23 & 0.1754 & 0.0266 & 0.9472\ UGC08286 & 9.10 & 0.94 & 0.61 & 1.13 & 10.90 & 10.78 & 11.05 & 1.11 & 1.02 & 1.20 & 2.13 & 0.0801 & 0.0490 & 0.1160\ UGC08490 & 9.01 & 0.92 & 0.58 & 1.12 & 10.79 & 10.64 & 10.99 & 1.15 & 1.01 & 1.27 & 0.29 & 0.0746 & 0.0425 & 0.1147\ UGC08550 & 8.46 & 0.79 & 0.42 & 1.07 & 10.51 & 10.33 & 10.74 & 1.05 & 0.93 & 1.16 & 0.66 & 0.0314 & 0.0154 & 0.0546\ UGC08699 & 10.70 & 0.56 & 0.51 & 0.60 & 11.95 & 11.75 & 12.21 & 0.99 & 0.85 & 1.11 & 1.13 & 0.1982 & 0.1076 & 0.3284\ UGC09037 & 10.84 & 0.11 & 0.04 & 0.20 & 11.91 & 11.74 & 12.13 & 0.87 & 0.74 & 0.98 & 1.03 & 0.0381 & 0.0101 & 0.0852\ UGC09133 & 11.45 & 0.47 & 0.44 & 0.50 & 12.22 & 12.18 & 12.25 & 0.99 & 0.92 & 1.05 & 8.84 & 0.5423 & 0.4231 & 0.6673\ UGC10310 & 9.24 & 0.73 & 0.30 & 1.06 & 10.67 & 10.42 & 10.96 & 1.02 & 0.88 & 1.14 & 0.49 & 0.1258 & 0.0341 & 0.3281\ UGC11820 & 8.99 & 0.52 & 0.17 & 0.90 & 11.15 & 11.04 & 11.28 & 0.74 & 0.65 & 0.81 & 2.20 & 0.0221 & 0.0079 & 0.0377\ UGC11914 & 11.18 & 0.64 & 0.61 & 0.67 & 13.04 & 12.44 & 13.67 & 0.75 & 0.58 & 0.94 & 2.55 & 0.0492 & 0.0110 & 0.2009\ UGC12506 & 11.14 & 0.97 & 0.66 & 1.14 & 12.14 & 11.96 & 12.33 & 0.99 & 0.84 & 1.13 & 0.67 & 0.5698 & 0.2753 & 0.9742\ UGC12632 & 9.11 & 0.66 & 0.23 & 1.04 & 10.73 & 10.56 & 10.92 & 0.98 & 0.87 & 1.09 & 0.41 & 0.0878 & 0.0252 & 0.1817\ UGC12732 & 9.22 & 0.54 & 0.18 & 0.95 & 11.11 & 10.96 & 11.30 & 0.92 & 0.80 & 1.02 & 0.29 & 0.0361 & 0.0109 & 0.0741\ UGCA281 & 8.29 & 0.66 & 0.28 & 1.01 & 9.86 & 9.36 & 10.46 & 1.04 & 0.88 & 1.18 & 0.89 & 0.0382 & 0.0054 & 0.1712\ UGCA444 & 7.08 & 0.61 & 0.21 & 1.02 & 9.62 & 9.19 & 10.14 & 1.08 & 0.91 & 1.25 & 0.55 & 0.0088 & 0.0018 & 0.0316\ ![image](figs/NGC3992_decomp_ML.pdf){width="\textwidth"} ![image](figs/mstar-fstar_nfw_check.pdf){width="33.00000%"} ![image](figs/mstar-fstar_Burkert_check.pdf){width="33.00000%"} ![image](figs/mstar-fstar_Einasto_check.pdf){width="33.00000%"}\ ![image](figs/mstar-fstar_withMLfixed.pdf){width="33.00000%"} ![image](figs/mstar-fstar_MLB_MLD_check.pdf){width="33.00000%"} [^1]: [email protected] [^2]: Since molecular, atomic and ionized gas is typically dynamically sub-dominant in ${M_\star}>10^{10} {M_\odot}$ galaxies. [^3]: Blue galaxies also have typically larger reservoirs of cold gas with respect to red ones. However, on average, the amount cold gas is sub-dominant with respect to stars for ${M_\star}>10^{10}{M_\odot}$. [e.g. @Papastergis+12]. [^4]: Taking into account this difference in the priors used, our results are very well compatible with theirs: our conclusions sit in the middle between their case with uniform priors (their Fig. 3) and that in which they impose a prior following the @Moster+13 ${M_\star}-{M_{\rm halo}}$ relation (their Fig. 5)	{ "pile_set_name": "ArXiv" }
--- abstract: 'Throughout scientific history, overarching theoretical frameworks have allowed researchers to grow beyond personal intuitions and culturally biased theories. They allow to verify and replicate existing findings, and to link disconnected results. The notion of self-play, albeit often cited in multiagent Reinforcement Learning, has never been grounded in a formal model. We present a formalized framework, with clearly defined assumptions, which encapsulates the meaning of self-play as abstracted from various existing self-play algorithms. This framework is framed as an approximation to a theoretical solution concept for multiagent training. On a simple environment, we qualitatively measure how well a subset of the captured self-play methods approximate this solution when paired with the famous PPO algorithm. We also provide insights on interpreting quantitative metrics of performance for self-play training. Our results indicate that, throughout training, various self-play definitions exhibit cyclic policy evolutions.' bibliography: - 'IEEEabrv.bib' - 'main.bib' title: 'A Comparison of Self-Play Algorithms Under a Generalized Framework [^1]' --- Introduction {#section:introduction} ============ Related Work {#section:related-work} ============ Preliminary notation {#section:preliminary_notation} ==================== Generalized Self-Play Framework {#section:self-play-framework} =============================== Self-Play Algorithms {#section:examples-from-literature} ==================== Proposed Incremental Innovations {#section:novel-contributions} ================================ Experimental Details {#section:experimental-details} ==================== Results {#section:results} ======= Conclusions & Future Work {#section:conclusion} ========================= Acknowledgements {#acknowledgements .unnumbered} ================ [^1]: This work was funded by the EPSRC Centre for Doctoral Training in Intelligent Games and Game Intelligence (IGGI) EP/L015846/1.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The response of an interferometer changing its orientation with respect to a fixed reference frame is analyzed in terms of the beam-pattern factors and the polarization-averaged antenna power pattern. Given the antenna’s motion, the latter quantity describes the antenna’s directionality as a function of time. An interesting case is represented by the class of motions where the detector’s plane is constrained to move on the surface of a cone of constant aperture; at the same time, the two arms are rotating around a vertical axis. This picture describes, in particular, the motion of LISA, a proposed space-based laser interferometer, as well as of other planned missions. The overall sky’s coverage, and that of the galactic plane in particular, is provided as a function of the cone’s aperture. Similarly, one can consider the case of an earth-based interferometer. Using the same formalism, one can derive a simple expression for the antenna pattern, averaged over the time of arrival of the signal, as a function of the position and orientation on the earth’s surface. In particular, there turn out to be two particular values for the terrestrial latitude and the inclination angle with respect to the local parallel which render the time-averaged antenna response perfectly isotropic. In the frequency domain, the general result is that the detector’s motion introduces in the instrumental response to a long-duration continuous signal a few harmonics of the orbital frequency, whose magnitude depends on the source’s position in the sky. In particular, we describe LISA’s response to circularly polarized sinusoidal waves coming from a few known binary systems in our Galaxy. address: 'Jet Propulsion Laboratory, California Institute of Technology, Pasadena, California, 91109' author: - '[Giacomo Giampieri]{}[^1]' date: March 1997 title: On the Antenna Pattern of an Orbiting Interferometer --- Introduction ============ Gravitational waves in the low-frequency regime ($10^{-4}$ to $10^{-1}$ Hz) can only be observed from space, due to terrestrial disturbances. In space, the only technique currently available, besides pulsar timing, is based on Doppler tracking of an interplanetary spacecraft [@EW]. However, this relatively inexpensive method has not provided enough sensitivity, thus far, for a detection. While better sensitivities may be expected in the near future, with advanced spacecraft such as CASSINI, much more ambitious projects for gravitational wave observatories in space have been proposed. Among these, the most promising detectors are based on space-born laser interferometry. In particular, LISA (Laser Interferometer Space Antenna) [@LISA; @Hough], and OMEGA (Orbiting Medium Explorer for Gravitational Astrophysics) [@OMEGA] consist of six drag-free, laser-bearing spacecraft, launched in orbit around the sun (LISA) or the earth (OMEGA). The six spacecraft would be placed, in pair, at the vertex of a triangle with $5\times 10^6$ km sides for LISA, and 5 times smaller for OMEGA. At each corner, the two spacecraft are phase locked through the exchange of a laser signal, replacing in this way the central mirror of an ordinary Michelson intereferometer. Each of the two probes sends a laser beam to a probe at each of the other two equilateral points, where the tracking signal is transponded back by phase-locked lasers, and the returning beams are eventually interfered. In order to keep the triangular constellation as stable as possible, elaborated orbits have been designed. In LISA case each spacecraft is orbiting a circle of radius $3\times 10^6$ km over a period of 1 yr. The interferometer plane, at an inclination of $60^o$ with respect to the ecliptic, is also rotating around the sun with the same periodicity. In OMEGA case the orbital plane is almost coincident with the ecliptic, and the interferometer is rotating around itself with a period of 53.21 days. The complicate motion is reflected in the time evolution of the interferometer’s response to a source located in a fixed position in the sky. We will investigate the behavior of the antenna response in presence of a generic motion, and apply our results to the specific motions of interest. As a side-product of our analysis, we can also examine a terrestrial interferometer, where the motion is simply related to the earth’s rotation around its axis, and study its antenna pattern as a function of the location and orientation on the earth’s surface. In this section, we briefly recall the formalism describing the antenna response to a gravitational wave passing by, in the long-wavelength approximation [@T300]. First, we introduce the wave symmetric trace-free (STF) tensor $$\bbox{W} = h_+ \Re (\vec{m} \otimes \vec{m}) + h_\times \Im (\vec{m} \otimes \vec{m})\,,$$ where the (complex) vector $\vec{m}$ is defined in terms of the polarization vectors $\vec{e}_X$ and $\vec{e}_Y$ as $$\vec{m} = \frac{1}{\sqrt{2}} \left( \vec{e}_X + i \vec{e}_Y \right)\,.$$ The tensor $\bbox{W}$ represents the wave field as measured in the interferometer’s proper rest frame. Then we define the STF detector tensor $$\bbox{D}= \vec{n}_1 \otimes \vec{n}_1 - \vec{n}_2 \otimes \vec{n}_2\,,$$ where $\vec{n}_i$ is the unit vector along the i-th arm. The interferometer response is the scalar obtained from the contraction of the wave tensor $\bbox{W}$ with the detector tensor $\bbox{D}$ $$R(t)= W_{ij} D^{ij} \equiv F_+ h_+ + F_\times h_\times\,. \label{response}$$ The beam-pattern factors $F_+$ and $F_\times$ depend on the antenna’s orientation with respect to the wave’s propagation direction and polarization axes. We can choose the reference frame as in Fig. 1, with the x-axis of the $(x,y,z)$ frame bisecting the interferometer’s arms, so that the only non-null components of $\bbox{D}$ in this reference frame are $$D_{12} = D_{21}= \sin(2\Omega)\,,$$ where $2\Omega$ is the aperture angle. Therefore, in order to obtain $R(t)$, we just need the component $W_{12}$ in this particular frame. Fig. 1 also shows the Euler’s angles $\theta,\phi$, and $\psi$ which transform from the interferometer’s frame $(x,y,z)$ to the wave reference frame $(X,Y,Z)$. The latter is defined with the Z-axis opposite to the propagation direction, and the $X$ and $Y$ axes along $\vec{e}_X$ and $\vec{e}_Y$, respectively. It is easy to find, for the + polarization $$\begin{aligned} F_+ &=& \sin(2\Omega) \Bigl[ \cos\theta\cos(2\phi) \sin(2\psi) \Bigr.\nonumber\\ &\mbox{}& \Bigl. \mbox{\hspace{1truecm}} + \frac{1}{2} \left(1 + \cos^2\theta\right) \sin(2\phi) \cos(2\psi)\Bigr]\,. \label{F+}\end{aligned}$$ The beam-pattern factor $F_\times$ is obtained from Eq. (\[F+\]) with the substitution $\psi \rightarrow \psi+\pi/4$, a well known polarization property of gravitational waves. However, when averages over the polarization angle $\psi$ are considered, we can assume, without loss of generality, $h^+=h^\times=h$. The quantity of interest is thus the polarization-averaged antenna’s [power pattern]{} $$P\equiv \langle\left(R\over \sin(2\Omega) h\right)^2 \rangle_\psi\,, \label{power}$$ which for the interferometer in Fig. 1 reads $$P(\theta,\phi) = \frac{1}{8}\left[ 1 + \cos^4\theta + 6\cos^2\theta - \sin^4\theta \cos(4\phi)\right]\,. \label{fixed}$$ For future reference, note that this definition of the antenna pattern is not normalized to unity, the average of $P(\theta,\phi)$ over the whole sky being 2/5. A plot of $P(\theta,\phi)$ in polar coordinates is shown in Fig. 5a. Eq. (\[fixed\]) gives the instantaneous power pattern for a wave impinging from the direction $(\theta,\phi)$ in the interferometer’s reference frame. If the detector is moving with respect to the source, then, apart from Doppler effects considerations, all we need to do is simply replace $\theta$ and $\phi$ with the appropriate functions $\theta(t)$ and $\phi(t)$. For example, if the antenna is rotating around its vertical axis with angular velocity $\omega$ (Fig. 2), then we can obtain the antenna pattern at any time $t$ from Eq. (\[fixed\]), with the substitution $$\phi \rightarrow \phi - \xi_0 - \omega t\,, \label{phi}$$ where $\xi_0$ is some initial angle. As a matter of fact, this very simple case describes, with good approximation, the time evolution of OMEGA [@OMEGA]. As mentioned in the Introduction, the idea of OMEGA is essentially similar to that of LISA, except that the six spacecraft are launched into a circular earth orbit, beyond the Moon orbit. The triangle has now $10^6$ km sides, and it is rotating around itself with a period of $\sim 53$ days. However, to mask the motion of the antenna with the apparent motion of the source is not always convenient, especially when dealing with a large number of sources, or when the motion is very complicate. In this paper, we shall introduce a more useful representation, where every quantity is referred to a fixed reference frame, so that the source’s polar coordinates $\theta$ and $\phi$ remain constant, and the antenna response depends on time through the actual motion of the interferometer. Antenna Pattern for a generic motion ==================================== We now introduce an arbitrary reference frame $(x',y',z')$, with the only requirement to be stationary with respect to fixed stars[^2]. To be more explicit, when dealing with a space-born interferometer, we can adopt an Ecliptic coordinate system. In the last section we will also consider a terrestrial interferometer, which is most easily described in an Equatorial frame. The full description of the antenna response requires six Euler’s angles, defined as in Fig. 3. The orthogonal transformation from the wave’s frame $(X,Y,Z)$ to the fixed frame $(x',y',z')$ is given by the orthogonal matrix $$\bbox{A}=\left( \begin{array}{ccc} \cos\phi\cos\psi - \cos\theta \sin\phi \sin\psi & -(\cos\phi\sin\psi + \cos\theta \sin\phi \cos\psi) & \sin\theta \sin\phi \\ \sin\phi\cos\psi + \cos\theta \cos\phi \sin\psi & -\sin\phi\sin\psi + \cos\theta \cos\phi \cos\psi & -\sin\theta \cos\phi \\ \sin\theta\sin\psi & \sin\theta \cos\psi & \cos\theta \end{array} \right)\,. \label{matrix}$$ The matrix $\bbox{B}$ which transform from $(x,y,z)$ to $(x',y',z')$ is analogous to the matrix $\bbox{A}$, with the Euler’s angles $\theta,\phi,\psi$ replaced by the corresponding ones $\zeta,\eta,\xi$. Thus, the complete transformation from the wave’s frame to the detector’s one is given by $\bbox{B}^T\cdot \bbox{A}$. Actually, as we may expect from Fig. 3, the angles $\phi$ and $\eta$ appear in our results only in the combination $\delta \equiv \phi - \eta$. After a rather lengthy calculation, one ends up with the following expressions for the beam-pattern factors $F_+$ and $F_\times$ \[F\] $$\begin{aligned} F_+ &=& \sin(2\Omega) \left[ A \cos(2\xi) \cos(2\psi) + B \cos(2\xi) \sin(2\psi) +\right.\nonumber\\ &\mbox{} & \left. \mbox{\hspace{1truecm}} + C \sin(2\xi) \cos(2\psi) + D \sin(2\xi) \sin(2\psi) \right] \,, \\ F_\times &=& \sin(2\Omega) \left[ B \cos(2\xi) \cos(2\psi) - A \cos(2\xi) \sin(2\psi) +\right.\nonumber\\ &\mbox{} & \left. \mbox{\hspace{1truecm}} + D \sin(2\xi) \cos(2\psi) -C \sin(2\xi) \sin(2\psi) \right] \,.\end{aligned}$$ The coefficients $A,B,C,$ and $D$ in Eqs. (\[F\]) depend only on the angles $\zeta, \theta,$ and $\delta$. They are explicitly given in Appendix A. We recall that the power pattern $P$ is obtained squaring, and averaging over the polarization angle $\psi$, the interferometer response. From Eqs. (\[response\]), (\[power\]), and (\[F\]) we eventually obtain $$P= \sum\limits_{n=0}^4 \left\{ \left[ \lambda_n + \mu_n \cos(4\xi) \right] \cos(n\delta) + \sigma_n \sin(4\xi) \sin( n \delta)\right\} \,, \label{sum}$$ where now the coefficients $\lambda_n, \mu_n,$ and $\sigma_n$ depend only on $\theta$ and $\zeta$. These coefficients, given in Appendix B, are quite complicate trigonometric polynomials of their arguments. Nonetheless, Eq. (\[sum\]) turns out to be very useful in practice. In fact, for the planned detectors considered in the present paper, the angle $\zeta$ is constant, and thus the only possible time evolution is related to the sinusoidal functions of $\xi$ and $\delta$ which appear explicitly in Eq. (\[sum\]). Before analyzing in more detail the proposed space-born interferometers, we consider a trivial application of Eq. (\[sum\]). A simple example {#a-simple-example .unnumbered} ---------------- As a first test of Eq. (\[sum\]) we can consider, as we did at the end of Sec. I, an interferometer which is rotating around its z-axis (see Fig. 2), so that its trivial motion is described by $$\zeta=\eta=0, \qquad\xi=\xi_0+\omega t\,. \label{simple}$$ From Eq. (\[simple\]), and Eqs. (B.1)-(B.15) of Appendix B, we find that the only non-null coefficients in Eq. (\[sum\]) are $$\begin{aligned} \lambda_0 &=& \frac{1}{8}\left( 1 + \cos^4\theta + 6\cos^2\theta\right) \,, \\ \mu_4 &=& \sigma_4 = -\frac{1}{8} \sin^4\theta \,,\end{aligned}$$ so that $$P = \frac{1}{8} \left[ 1 + \cos^4\theta + 6 \cos^2\theta - \sin^4\theta \cos\left[4(\phi-\xi_0-\omega t)\right]\right] \,. \label{rot}$$ As pointed out in Sec. I, this result can be obtained much more easily directly from Eq. (\[fixed\]), with the substitution (\[phi\]). Eq. (\[rot\]) gives, with good approximation, the antenna pattern of OMEGA, with $\omega\simeq 1.4\times 10^{-6}$ sec$^{-1}$. Space interferometers: Conical motion ===================================== In the previous section, we have considered the antenna power pattern associated to an unspecified motion of the detector. We will now focus our attention to the case of a space interferometer, which presents, independently of the particular project under investigation, some very general and interesting properties. Inserting the space interferometer in its orbit and keeping the interferometer configuration stable over the mission lifetime - at least two orbital periods - is a very demanding navigation task, due principally to the perturbations from the earth and the other planets. For instance, one of the consequences of the complicate orbit is the fact that we cannot maintain equal distances between the probes. In a recent paper [@GG], we modeled the noise that is introduced into the differenced data because of the unequal arms, and showed that the final accuracy of the interferometer is not compromised. Another example of the problems we may face in a space-born interferometer is that, due to the earth disturbances, high Doppler rates would result. Hellings et al.[@RH] described a laser transmitter and receiver hardware system that provides the required readout accuracy and implements a self-correction procedure for the on-board frequency standard used for laser phase measurement. We will now consider the interferometer’s orbit, and discuss its implication on the antenna response to a wave coming from a given direction in the sky. In particular, as we mentioned in the Introduction, LISA [@LISA] will orbit the sun at the earth’s distance, as far behind the earth as possible. The plane containing the six probes, during its orbit, will remain always tangent to the surface of a cone of $60^o$ aperture, and the detector itself will rotate in this plane with same periodicity - one year - but opposite direction. Fig. 4, reproduced from [@Hough], shows LISA configuration. In this section, we will consider a LISA-like motion, characterized by a generic cone aperture. In other words, the motion of each of the three interferometers is assumed to be described by[^3] \[lisa\] $$\begin{aligned} \zeta&=& \text{const} \,, \\ \eta &=& \eta_0 + \omega t \qquad (\Rightarrow \delta= \phi-\eta_0 - \omega t) \,, \\ \xi &=& \xi_0 - \omega t \,.\end{aligned}$$ Note that $\eta$ and $\xi$ are counter-rotating. Note also that the three interferometers formed by the triangular configuration have initial values $\xi_0$ which differ from each other by $120^o$, whereas $\eta_0$ is the same for all of them. For the sake of conciseness, since we are concerned here with the antenna power pattern, we will consider only one of the interferometers, leaving the possibility of exploiting the polarization sensitivity to future works. The proposed LISA orbit has $\zeta=60^o$, a critical value for the stability of the triangular configuration. Since the behavior of the antenna, in terms of sky’s coverage, directionality, etc., is very sensitive to the inclination, we will keep $\zeta$ as a free parameter throughout this paper, and refer to Eqs. (\[lisa\]) as describing a ‘conical’ motion. For a given $\zeta$, the coefficients $\lambda_n, \mu_n,$ and $\sigma_n$ are now functions of $\theta$. They are explicitly given in Appendix C for the LISA case. In order to determine the sky’s coverage during the detector’s lifetime, we need to consider the time average of the antenna pattern $P$ over one orbital period, which gives $$\langle P \rangle_T = \lambda_0 + \frac{1}{2}\left( \mu_4+\sigma_4\right) \cos[4(\phi-\eta_0-\xi_0)] \label{time}$$ Fig. 5 shows a plot of $\langle P \rangle_T$ in polar coordinates for various values of the cone’s aperture $\zeta$. Note that $\zeta=0$ corresponds to an interferometer fixed in space; since we are taking $\xi$ and $\eta$ counter-rotating, the interferometer does not rotate at all. We know from Eq. (\[fixed\]) that an interferometer fixed in space can never detect waves impinging from four specific null-directions, given by $\theta=\pi/2$, and $\phi=k \pi/2\; ( k=0,\dots,3)$. We can now analyze what happens to these null-directions in the generic conical case, focussing our attention to the ecliptic plane $\theta=\pi/2$. It is easy to show, from the behavior of the functions $\lambda_0$ and $\|\mu_4+\sigma_4\|/2$ (see Fig. 6), that there is a tendency for the null-directions on the ecliptic plane to remain visible, although the magnitude of this effect is strongly affected by the value of $\zeta$: for $\zeta=0$ the null-directions are obviously completely preserved, while for $\zeta=\pi/2$ the $\phi$-dependence is very poor. LISA is much closer to the latter case, and actually its $\langle P\rangle_T$ is almost independent on $\theta$ as well, as we shall see in a while. Note that, if $\xi$ and $\eta$ were co-rotating, instead of counter-rotating as in LISA, then in the last term of Eq. (\[time\]) we would have to make the substitution $$\frac{1}{2}\left(\mu_4+\sigma_4\right) \rightarrow \frac{1}{2}\left(\mu_4-\sigma_4\right) \,.$$ which means that the $\phi$-term would be even smaller, in absolute value, compared to $\lambda_0$. One might infer from Fig. 5d that, if the inclination is close to $\zeta=\pi/2$, then the zeros of the antenna pattern can be found in the direction orthogonal to the ecliptic plane. However, this is not the case, since $\theta=0$ implies that $P$ is still given by Eq. (\[fixed\]), with $\theta$ and $\phi$ replaced, respectively, by $\zeta$ and $\xi$. Thus, even for $\zeta=\pi/2$ the time average is non-zero at the poles. Another important issue related to Eq. (\[time\]) is the time-averaged antenna directionality. As we can predict from Fig. 5, directionality is strongly dependent on the angle $\zeta$. To make this statement more precise, let us consider the r.m.s. deviation from isotropy, defined as $$\Delta \equiv \left[ \frac{175}{48\pi} \int\limits_{4\pi} \left(\langle P\rangle_T - 2/5\right)^2 d\Omega \right]^{1/2} \,. \label{delta}$$ The normalization factor in front of Eq. (\[delta\]) is chosen in such a way that $\Delta$ is normalized to one for a fixed interferometer. For a conical motion, inserting Eq. (\[time\]) in Eq. (\[delta\]) we easily obtain $\Delta=\Delta(\zeta)$, shown in Fig. 7. Explicitly, we find $$\begin{aligned} \Delta(\zeta) &=& \frac{1}{288} \left[ 19779 + 120 \cos\zeta - 118380 \cos^2\zeta \right.\nonumber\\ &\mbox{}&\left. + 840 \cos^3\zeta + 180690 \cos^4\zeta +840 \cos^5\zeta \right.\nonumber\\ &\mbox{}&\left. -3180 \cos^6\zeta +120 \cos^7\zeta +2115 \cos^8\zeta\right]^{1/2}\,.\end{aligned}$$ Thus, the antenna pattern is distributed more and more isotropically as we increase $\zeta$ from $\zeta=0$. After we reach a minimum at $\zeta\simeq 55^o$, the antenna’s directionality starts increasing again. Therefore, we can conclude that, in the conical case, it is impossible to get a perfectly isotropic response, i.e. we never get $\Delta=0$ (see also the above discussion about the $\phi$-dependence). However, we can get very close to this ideal situation, if we choose $\zeta$ appropriately. Remarkably, LISA’s inclination is very close to the optimal value $\zeta\simeq 55^o$, and gives $\Delta(60^o) \simeq 0.14$. This means that LISA, during its lifetime, will cover the whole sky in an approximately uniform manner. Of course, in some circumstances, directionality needs to be preserved. For instance, one may want to disentangle the isotropic component of the stochastic background from the anisotropic contribution of the galactic binaries. Directionality can always be preserved integrating over a shorter period. Fig. 8 shows LISA’s antenna pattern averaged over 3 and 6 months, respectively. Finally, we want to consider the interferometer’s responsiveness to the galactic plane, given that most of the strongest sources will lie on this plane. In our coordinate system, the galactic plane is characterized by $$\theta(\phi) = \arctan(\alpha \cos\phi + \beta\sin\phi)^{-1}\,, \label{plane}$$ where $\alpha \simeq 1.75$ and $\beta \simeq 4\times10^{-4}$. Assuming, for simplicity, that the sources are distributed isotropically on the plane, the event rate of disk’s sources is proportional to the average area of the intersection of this plane with the antenna pattern, equal to $$G = \frac{1}{2T} \int\limits_0^{2\pi} \int\limits_0^T \Bigl[ P(\theta(\phi),\phi) \Bigr]^2 dt\, d\phi \,,$$ where $\theta(\phi)$ is the function given in Eq. (\[plane\]). This area depends on the details of the detector’s motion, in our case on $\zeta$, $\xi_0$, and $\eta_0$. Fig. 9 shows the quantity $G$ as a function of $\zeta$, since the dependence on $\xi_0$ and $\eta_0$ can be neglected in a first approximation. For LISA, we find the small value $G \sim 0.46$, the exact value depending on the initial conditions $\eta_0+\xi_0$. We conclude that LISA is not particularly sensititive to the galactic disk, due to the fairly large inclination to the Ecliptic of both the detector’s plane and the Galaxy. According to this crude analysis, we can expect that OMEGA, with a smaller, almost negligible, inclination, would increase its chances of observing a signal from the disk by roughly a factor two. Note, however, that previous calculations [@Evans; @Hils] have shown that gravitational radiation from galactic binaries in the disk is comparable to that coming from the local region ($r < 200$ pc). Since the latter is isotropically distributed, the time-varying signal from the disk contributes only a fraction of the galactic binaries stochastic background. Nonetheless, this small component could make the stochastic signal distinguishable from the detector’s noise. In a future paper we will investigate the amplitude modulations introduced by the antenna motion in the confusion noise generated by different populations of galactic binary systems, and describe how to exploit this effect in order to detect the signal and to obtain information about the distribution of sources in the Galaxy. Earth-based interferometers =========================== As an additional application of Eq. (\[sum\]), let us consider a terrestrial interferometer. In this case the most convenient choice for the ‘fixed’ reference frame is the Equatorial one, with the $z'$ axis directed toward the North Pole, and the $x'$ axis toward the Vernal Equinox. In this frame, the motion of the interferometer becomes similar to the conical case previously analyzed, except that now the detector can not rotate on itself, of course. In other words we have \[earthmotion\] $$\begin{aligned} \zeta &=& \pi/2 -\ell \,, \\ \eta &=& \eta_0 +\omega t \,, \\ \xi &=& \iota \,,\end{aligned}$$ where $\ell$ is the terrestrial latitude, $\omega$ is the earth’s angular velocity of rotation, and $\iota\in[-\frac{\pi}{4},+\frac{\pi}{4}]$ is the angle between the arms’ bisector and the local parallel. For example, let us consider the average of the antenna pattern $P$ over the time of arrival of the signal. In the terrestrial case, as opposed to the conical case considered in Sec. III, the interferometer cannot rotate around its vertical axis, and therefore averaging over time or $\phi$ gives the same result, namely $$\langle P \rangle_T = \lambda_0 (\theta,\ell) + \mu_0(\theta,\ell) \cos(4\iota)\,. \label{earth}$$ For any specific value of $\theta$, for example $\theta=102^o$, corresponding to the direction of the center of the Virgo cluster, Eq. (\[earth\]) gives the square of the r.m.s. power as a function of the antenna’s position and orientation on the earth’s surface, a quantity already numerically studied in [@ST]. Instead of fixing $\theta$, we could try to answer the question: is there any particular location and orientation for which the antenna pattern, averaged over one day, is isotropic? The answer turns out to be affirmative, and the isotropic antenna is characterized by \[isotropy\] $$\begin{aligned} \ell_{is} &=& \arcsin \left(\pm \frac{1}{\sqrt{3}}\right) \simeq \pm 35^o.26438972 \,,\\ \iota_{is} &=& \frac{1}{4}\arccos \left(-\frac{1}{5}\right) \simeq \pm 25^o.38423976 \,.\end{aligned}$$ One can easily check, by inspection, that $\ell_{is}$ and $\iota_{is}$ produce $\langle P \rangle_T \equiv 2/5$ or, equivalently, $\Delta\equiv 0$. A detector located at latitude $\ell_{is}$ and oriented by $\iota_{is}$ maximizes the event rate of an isotropic population of sources. In the event that the detector’s position has already been chosen, one can still make use of Eq. (\[earth\]) in order to find the optimal orientation $\iota_$ which gives the least directional antenna pattern at that latitude. At each latitude $\ell$, we define as optimal that orientation $\iota_$ which minimize the quantity $\Delta$. Fig. 10 shows $\iota_=\iota_(\ell)$ and the corresponding minimum $\Delta_\equiv \Delta(\ell,\iota_)$. We stress that, as one may actually expect, $\Delta$ depends much more strongly on $\ell$ than $\iota$, and in particular the antenna becomes rapidly anisotropic as we move away from $\ell_{is}$, no matter how optimally we try to choose $\iota$. Moreover, we can foresee several terrestrial interferometers to be operative in the near future, so that the sky’s coverage of a single antenna is not really an issue as critical as in the space-born case previously discussed. Sinuosoidal waves from binaries. Fourier analysis. ================================================== It is generally assumed that galactic and extragalactic binary systems are the most promising sources of gravitational waves for detectors based on laser interferometry. In fact, waves from a binary star, including the effect of eccentricity, orbital inclination, and also post-Newtonian corrections, have long been studied, and are today well understood. In particular, the sensitivity of the planned space interferometers should allow the detection of waves from several known galactic binary stars. In the LISA and OMEGA frequency band, the strongest among these sources are presumably the Interacting White Dwarfs Binaries (IWDB) [@Evans; @Hils]. Table I contains the available data for five IWDB, including the amplitude and frequency of the expected gravitational waves. We have applied our results to these objects, and describe the LISA’s response to the waves originating from them. Fig. 11 shows the beam-pattern factors $F_+$ and $F_\times$ for the five IWDB in Table I, as seen from LISA over one year. Fig. 11 also shows the analogous quantities for a sinusoidal signal, of unspecified amplitude and frequency, coming from the galactic centre. We will now consider the effect of the motion in the frequency domain, for both the space-born and the terrestrial cases. We define the Fourier series as usual $$P(t) = a_0 + \sum\limits_{k=1}^{\infty} \left\{ a_k \cos\left(k\pi t\over\ L\right) + b_k \sin\left(k\pi t\over\ L\right) \right\}\,,$$ where $T=2L$ is the orbital period, and $$\begin{aligned} a_0 &=& \frac{1}{2L} \int\limits_{-L}^{+L} P(t) dt\,,\\ a_k &=& \frac{1}{L} \int\limits_{-L}^{+L} P(t) \cos\left(k\pi t\over\ L\right) dt \,,\\ b_k &=& \frac{1}{L} \int\limits_{-L}^{+L} P(t) \sin\left(k\pi t\over\ L\right) dt\,.\end{aligned}$$ For the conical case analyzed in Sec. III, using eqs.(B1)-(B14) of Appendix B, one finds that the only non-zero Fourier coefficients are given by ($\delta_0 \equiv \phi-\eta_0,\: k=1,\dots,8$): $$\begin{aligned} a_0 &=& \lambda_0 +\frac{1}{2}(\mu_4+\sigma_4) \cos[4(\delta_0-\xi_0)] \,,\\ a_k &=& \mu_0 \cos(4\xi_0) \delta_{k 4} + \sum\limits_{n=1}^{4}\Bigl\{\lambda_n \cos(n\delta_0) \delta_{nk} \Bigr.\nonumber \\ &\mbox{}& \Bigl. +\frac{1}{2} (\mu_n+\sigma_n) \cos(n\delta_0-4\xi_0) \delta_{k+n,4} \Bigr.\nonumber \\ &\mbox{}& \Bigl. +\frac{1}{2} (\mu_n-\sigma_n) \cos(n\delta_0+4\xi_0) \delta_{k-n,4}\Bigr\}\,, \\ b_k &=& \mu_0 \sin(4\xi_0) \delta_{k 4} + \sum\limits_{n=1}^{4}\Bigl\{\lambda_n \sin(n\delta_0) \delta_{nk} \Bigr.\nonumber \\ &\mbox{}& \Bigl. -\frac{1}{2} (\mu_n+\sigma_n) \sin(n\delta_0-4\xi_0) \delta_{k+n,4} \Bigr.\nonumber\\ &\mbox{}& \Bigl. +\frac{1}{2} (\mu_n-\sigma_n) \sin(n\delta_0+4\xi_0) \delta_{k-n,4}\Bigr\}\,.\end{aligned}$$ The analogous calculation for the terrestrial case, described by Eqs. (\[earthmotion\]), gives the following nine coefficients $(k=1,\dots,4)$ $$\begin{aligned} a_0 &=& \lambda_0 + \mu_0 \cos(4\iota)\,,\\ a_k &=& \left[\lambda_k +\mu_k\cos(4\iota)\right] \cos(k\delta_0) +\sigma_k \sin(4\iota) \sin(k\delta_0)\,,\\ b_k &=& \left[\lambda_k +\mu_k\cos(4\iota)\right] \sin(k\delta_0) -\sigma_k \sin(4\iota) \cos(k\delta_0)\,.\end{aligned}$$ In general, given the initial detector’s position, these Fourier coefficients depend on the source’s coordinates $\theta$ - through $\lambda_k, \mu_k, \sigma_k$ - and $\phi$ - through $\delta_0$. If the source location is known, then one can look for these spectral lines as a convincing signature about the gravitational origin of the signal. When the source’s coordinates are unknown, however, one has to deal with the complication arising from the Doppler effect [@Schutz]. The motion of the detector, besides the amplitude modulation described in the present work, also introduces a location-dependent phase modulation, in the form of a Doppler broadening of the sinusoidal signal. In the case of LISA, the magnitude of this effect, over a period $T=1$ yr, is $$\frac{\Delta f}{f} \simeq 2\times 10^{-4} \,,$$ so that, in the spectral region below $10^{-2}$ Hz, we do not have any hope of finding the aforementioned lines, separated from each other by only $1/T\simeq 3\times 10^{-8}$ Hz. For a terrestrial interferometer, the situation is analogous, only complicated by the simultaneous effects of the diurnal and annual motion of the earth, and also by the earth-moon interaction. In conclusion, for long enough observations, we need special techniques to compensate for the frequency spread over several frequency-resolution bins, and eventually to recover the amplitude modulation described in this paper. Several different strategies can be adopted to overcome this problem, although none of them is completely satisfactory, due to the large amount of computation involved. See [@Schutz] for more details. In any case, the amplitude modulation can be exploited for an independent measurement of the location of the source, and, in addition, to obtain the polarization of the wave. Conclusions =========== In the present work, we have considered a gravitational wave interferometer, in motion with respect to fixed stars, and studied the resulting amplitude modulation of a long-duration continuous signal. The general results are presented in Sec. II, where the istantaneous beam-pattern factors - Eqs. (\[F\]) - and the polarization-averaged antenna power pattern - Eq. (\[sum\]) - are given as functions of time, for a generic motion. Next, two particular cases have been analyzed: 1) the probable orbit of a space-born interferometer, with particular emphasis on LISA, and 2) the motion of a ground-based interferometer. For what concerns LISA, we have shown that its peculiar motion makes the time-averaged antenna pattern practically isotropic, thus providing an uniform coverage of the whole sky over the period of one year. For shorter integration periods directionality is mostly preserved, and can be exploited where necessary, for example in the search of a galactic binaries background. However, when we focused on the galactic disk, we found that the average antenna response is far from optimal, due to the relative orientation of the Ecliptic and the galactic plane itself. We stress that these results are not conclusive, since we have neglected the anisotropy in the distribution of the sources with respect to the sun, due to the fact that we are located near the edge of the disk. In this respect, additional work is needed. In the terrestrial case, thanks to the probable redundancy of future gravitational wave observatories, the discussion about a single antenna’s sky’s coverage is not so critical. However, we found that there are particular positions on the earth’s surface, given in Eqs.(\[isotropy\]), which render the time-averaged antenna response perfectly isotropic. For what concerns the galactic plane, since the latter makes with the Equatorial plane approximately the same angle it makes with the Ecliptic ($\sim 60^o$), the result is analogous to the conical case, with the angle $\zeta$ interpreted as $90-\ell$ in Fig.9. In other words, the response to the galactic plane increases as we move the interferometer from the equator toward the poles, with a minor role played by the orientation angle $\iota$. The author would like to thank P.Bender and R.Hellings for discussions. The research described in this paper was performed while the author held an NRC-NASA Resident Research Associateship at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. In this Appendix, we give the coefficients $A,B,C,$ and $D$ which enter in the expressions of $F_+$ and $F_\times$, Eqs. (\[F\]). They, in turn, can be expressed in terms of intermediate quantities $\alpha_1, \alpha_2, \beta_1,$ and $\beta_2$ as follows $$\begin{aligned} A &=& \alpha_1 \beta_1 + \alpha_2 \beta_2 \\ B &=& \alpha_1 \beta_2 - \alpha_2 \beta_1 \\ C &=& \frac{1}{2} \left[ \beta_1^2 - \beta_2^2 - \alpha_1^2 + \alpha_2^2\right]\\ D &=& \alpha_1 \alpha_2 + \beta_1 \beta_2 \end{aligned}$$ where $$\begin{aligned} \alpha_1 &=& \cos\delta\\ \alpha_2 &=& \cos\theta \sin\delta\\ \beta_1 &=& \cos\zeta \sin\delta\\ \beta_2 &=& \cos\zeta \cos\theta \cos\delta + \sin\zeta \sin\theta\end{aligned}$$ In this Appendix, we give the coefficients $\lambda_n, \mu_n$, and $\sigma_n\quad (n=0,\dots,4)$, defined in Eq. (\[sum\]), as functions of the angles $\theta$ and $\zeta$. $$\begin{aligned} \lambda_0 &=& \frac{1}{64} \left( 35 + 35 \cos ^4\theta \cos^4\zeta - 30 \cos^2\theta + 3\cos^4\zeta + 108 \cos^2\zeta \cos^2\theta \right. \nonumber\\ &\mbox{}& \left. - 30 \cos^2\zeta \cos^4\theta - 30 \cos^4\zeta \cos^2\theta - 30 \cos^2\zeta + 3 \cos^4\theta\right) \\ \lambda_1 &=& \frac{1}{32} \sin(2\zeta) \sin(2\theta) \left( 15 -3 \cos^2\theta -3 \cos^2\zeta + 7\cos^2\zeta \cos^2\theta \right) \\ \lambda_2 &=& \frac{1}{16} \left(7 + 7 \cos ^4\theta \cos^4\zeta - 8 \cos^2\theta + \cos^4\zeta + 16 \cos^2\zeta \cos^2\theta - 8\cos^2\zeta \cos^4\theta \right. \nonumber\\ &\mbox{}& \left. - 8 \cos^4\zeta \cos^2\theta - 8 \cos^2\zeta + \cos^4\theta\right) \\ \lambda_3 &=& \frac{1}{8} \sin^3\zeta \cos\zeta \sin^3\theta \cos\theta \\ \lambda_4 &=& \frac{1}{64} \sin^4\zeta \sin^4\theta \\ \nonumber \\ \mu_0 &=& -\frac{1}{64} \sin^4\zeta \left( 3 - 30 \cos^2\theta + 35 \cos^4\theta \right) \\ \mu_1 &=& -\frac{1}{16} \sin(2\theta) \cos\zeta \sin^3\zeta \left( 3 - 7\cos^2\theta \right) \\ \mu_2 &=& \frac{1}{16}\sin^2\zeta \sin^2\theta \left(1+\cos^2\zeta\right) \left(1 -7 \cos^2\theta\right) \\ \mu_3 &=& \frac{1}{16} \sin(2\zeta) \sin^3\theta \cos\theta \left(3 + \cos^2\zeta\right) \\ \mu_4 &=& -\frac{1}{64} \sin^4\theta \left( 1 + \cos^4\zeta + 6\cos^2\zeta\right) \\ \nonumber \\ \sigma_0 &=& 0 \\ \sigma_1 &=& -\frac{1}{16} \sin(2\theta) \sin^3\zeta \left( 3 -7 \cos^2\theta \right) \\ \sigma_2 &=& \frac{1}{8} \cos\zeta \sin^2\zeta \sin^2\theta \left(1 - 7\cos^2\theta\right) \\ \sigma_3 &=& \frac{1}{8} \sin\zeta \cos\theta \sin^3\theta \left(1+3\cos^2\zeta\right) \\ \sigma_4 &=& -\frac{1}{16} \sin^4\theta \cos\zeta \left(1+\cos^2\zeta\right) \end{aligned}$$ In this Appendix, we provide the coefficients $\lambda_n, \mu_n$, and $\sigma_n\: (n=0,\dots,4)$ for LISA $$\begin{aligned} \lambda_0 &=& \frac{1}{1024} \left(443 -37 \cos^4\theta -78 \cos^2\theta\right) \\ \lambda_1 &=& \frac{\sqrt{3}}{256} \sin(2\theta)\left(57-5 \cos^2\theta\right) \\ \lambda_2 &=& \frac{9}{256} \left(9-\cos^4\theta- 8\cos^2\theta\right) \\ \lambda_3 &=& \frac{3\sqrt{3}}{128} \sin^3\theta\cos\theta \\ \lambda_4 &=& \frac{9}{1024} \sin^4\theta \\ \nonumber \\ \mu_0 &=& -\frac{9}{1024} \left(3 + 35 \cos^4\theta -30 \cos^2\theta\right) \\ \mu_1 &=& -\frac{3\sqrt{3}}{256} \sin(2\theta)\left( 3 - 7 \cos^2\theta\right) \\ \mu_2 &=& \frac{15}{256} \sin^2\theta \left(1 -7 \cos^2\theta \right) \\ \mu_3 &=& \frac{13\sqrt{3}}{128} \sin^3\theta \cos\theta \\ \mu_4 &=& -\frac{41}{1024} \sin^4\theta \\ \nonumber \\ \sigma_0 &=&0 \\ \sigma_1 &=& -\frac{3\sqrt{3}}{128} \sin(2\theta)\left( 3 -7 \cos^2\theta \right) \\ \sigma_2 &=& \frac{3}{64} \sin^2\theta\left(1 - 7 \cos^2\theta\right) \\ \sigma_3 &=& \frac{7\sqrt{3}}{64} \sin^3\theta \cos\theta \\ \sigma_4 &=& - \frac{5}{128} \sin^4\theta\end{aligned}$$ F.B. Estabrook and H.D. Wahlquist, Gen. Rel. Grav., 439 (1975). K. Danzmann et al., [LISA: Proposal for a Laser Interferometric Gravitational Wave Detector in Space]{}, Report No. MPQ 177, Max-Planck Institut für Quantenoptik (Garching bei München, Germany, 1993); P.L. Bender et al., [LISA: Presentation of Assessment Study Results]{}, ESA SCI(94)9 (1994). J. Hough et al., in [First Edoardo Amaldi Conference on Gravitational Wave Experiments]{} (World Scientific, Singapore, 1995). R.W. Hellings et al., [OMEGA: Orbiting Medium Explorer for Gravitational Astrophysics]{}, Midex Proposal (Jet Propulsion Laboratory, 1995). K.S. Thorne, in [300 Years of Gravitation]{}, eds.S.W. Hawking and W. Israel (Cambridge University Press, Cambridge, 1987); see also S.V. Dhurandhar and M. Tinto, M.N.R.A.S. [234]{}, 663 (1988). G. Giampieri, R.W. Hellings, M. Tinto, and J.E.Faller, Optics Comm. [123]{}, 669 (1996). R.W. Hellings, G. Giampieri, L. Maleki, M. Tinto, K. Danzmann, J. Hough, and D. Robertson, Optics Comm. [124]{}, 313 (1996). C.R. Evans, I. Iben, and L. Smarr, Astrophys.J. [323]{}, 129 (1987). D. Hils, P.L. Bender, and R.F. Webbink, Astrophys.J. [360]{}, 75 (1990). B.F. Schutz and M. Tinto, M.N.R.A.S. [224]{}, 131 (1987). B.F. Schutz, in [The Detection of Gravitational Waves]{}, ed. D.G. Blair (Cambridge University Press, Cambridge, 1991). ---------------- ----------- ----------- -------------- ---------------- Name $\theta$ $\phi$ GW Amplitude GW Frequency (degrees) (degrees) ($10^{-22}$) ($10^{-3}$ Hz) 52.56 260.38 5.27 1.94 [CR Boo]{} 72.10 292.27 2.82 1.34 [V803 Cen]{} 120.31 306.17 0.89 1.24 [CP Eri]{} 120.83 151.77 4.02 1.16 [GP Com]{} 67.00 277.73 1.77 0.72 ---------------- ----------- ----------- -------------- ---------------- : Data for 5 known IWDB. The first column is the name, the second and third column are, respectively, the angles $\theta$ and $\phi$ in the Ecliptic coordinate system. The last two columns gives, respectively, the predicted amplitude and frequency of the gravitational waves. [^1]: Present address: Queen Mary and Westfield College, Astronomy Unit, London E1 4NS. [^2]: Here and in the following, when we say ‘stationary’ (or ‘fixed’) we mean stationary over the characteristic time scale of the detector’s motion. [^3]: we could eliminate one of the two initial conditions $\xi_0$ or $\eta_0$ by simply rescaling the time, taking for instance the origin of time at the passage through the line of nodes ($\xi_0 =0$) or through the vernal equinox ($\eta_0=0$).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the thermoelectric transport of a small metallic island weakly coupled to two electrodes by tunnel junctions. In the Coulomb blockade regime, in the case when the ground state of the system corresponds to an even number of electrons on the island, the main mechanism of electron transport at the lowest temperatures is elastic cotunneling. In this regime, the transport coefficients strongly depend on the realization of the random impurity potential or the shape of the island. Using random-matrix theory, we calculate the thermopower and the thermoelectric kinetic coefficient and study the statistics of their mesoscopic fluctuations in the elastic cotunneling regime. The fluctuations of the thermopower turn out to be much larger than the average value.' author: - 'A. S. Vasenko' - 'D. M. Basko' - 'F. W. J. Hekking' title: 'Giant mesoscopic fluctuations of the elastic cotunneling thermopower of a single-electron transistor' --- Introduction ============ Thermoelectric transport through various nanodevices has been the subject of extensive experimental and theoretical studies for more than two decades. The coherent propagation of electron waves in clean nanostructured conductors leads to quantum size effects that strongly affect the thermoelectric transport coefficients; [@VanHouten1992] the presence of electron-electron interactions leads to additional renormalization phenomena. [@Kane1996; @Fazio1998] In low-dimensional disordered conductors, interference of diffusively scattered electron waves weakens the screening of electron-electron interactions, leading to anomalous, energy-dependent non-Fermi-liquid behavior of the thermoelectric transport coefficients. [@Catelani2005] All these effects can be made visible explicitly using the tunability of nanodevices, [e.g.]{}, by varying external gate potentials or magnetic fields. Various practical applications based on thermoelectric phenomena in nanostructures have been developed, including thermometry, and nanorefrigeration, [@Giazotto2006; @Muhonen2012; @Courtois; @Vasenko; @Ozaeta] and more generally thermoelectric nanomachines. [@DiSalvo1999; @Shakouri2011; @Whitney2014; @Sothmann] A prototypical device that manifests all the relevant aspects of electron transport in nanostructures, [i.e.]{}, quantum size effects, energy-dependent coherent propagation and electron-electron interaction effects, is the single-electron transistor (SET). It consists of an island (a quantum dot or a small metallic particle) connected to two leads (source and drain) by small tunnel junctions. The electrostatic potential on the island can be controlled externally due a capacitive coupling between the island and a nearby gate electrode with the capacitance $C_g$ (Fig. \[set\]). The electrostatic energy cost of putting an extra electron on the island is of the order of the so-called charging energy, $E_C \equiv e^2/ 2C$, where $C$ is the total capacitance of the island. When the temperature $T$ and the applied source-drain voltage $V_{sd}$ are small, $eV_{sd}, T \ll E_C$, this charging effect results in the so-called Coulomb blockade of the electron transport through the island (see Ref. for a review). Moreover, at low temperatures the electronic phase coherence length is longer than the typical dimensions of the island and as a result the electronic motion is phase-coherent. ![(Color online) A sketch of a single-electron transistor. The central island is connected to the source and drain electrodes by tunnel junctions and capacitively coupled to the gate electrode via the gate capacitance $C_g$. The total capacitance $C$ of the island is given by the sum of the capacitances between the island and each electrode.[]{data-label="set"}](set){width="8cm"} The number of electrons $N$ on the island that minimizes the electrostatic energy $E_{el}(N)$, as well as the energy cost $E_{el}(N\pm{1})-E_{el}(N)$ to add/remove an electron, depends on the external gate voltage $V_g$. For each $N$, there is a particular value of $V_g$, such that $E_{el}(N)=E_{el}(N+1)$, called the degeneracy point. Then, starting from the state with $N$ electrons, one electron can tunnel from the source to the island, and then another electron tunnel from the island to the drain, restoring the number of electrons on the island to $N$. This so-called sequential tunneling mechanism (when electrons tunnel one by one in and out of the island, hence the term “single-electron transistor”) leads to a sequence of peaks in the dependence of the source-drain linear-response conductance $G$ on $V_g$, spaced at $e/C_g$, schematically shown in Fig. \[TM\]a. If $V_g$ is tuned away from the degeneracy point into the Coulomb blockade valley, the sequential-tunneling contribution to the conductance is exponentially suppressed at low temperatures,[@Kulik1975; @Beenakker] and a more important contribution to the transport arises from the so-called cotunneling mechanism. It is due to processes where an electron tunnels from the source to the drain via a virtual intermediate state on the island. The energy of this virtual state is higher than that of the initial and final states by a large amount $\sim{E}_C$, so the tunneling amplitude is small as $\sim{1}/E_C$. Yet, at low temperatures, this dominates over the exponentially small sequential-tunneling contribution $\sim\exp(-E_C/T)$. If the internal state of the island (i. e., the distribution of the $N$ electrons over the single-particle energy levels on the island) is different before and after the process, one speaks of inelastic cotunneling, in the opposite case it is called elastic cotunneling.[@Averin1989; @Glazman1990; @AN; @ANbook] As the inelastic cotunneling process involves creation of an electron-hole pair on the island, the corresponding contribution to the linear-response conductance vanishes at $T\to{0}$ ($\propto{T}^2$), while the elastic one is temperature-independent. Thus, the latter dominates the transport at very low temperatures. [@footnote] An important difference between inelastic and elastic cotunneling is that the latter is sensitive to the coherent electron motion on the island whereas the former is not. As a result, the elastic cotunneling contribution to the SET’s conductance shows strong mesoscopic sample-to-sample fluctuations, [@AG] the fluctuations being of the same order as the average conductance. Moreover, the conductance fluctuations of elastic cotunneling are so large that they dominate the inelastic mechanism even at not too low temperatures, when the average conductance value is already determined by inelastic cotunneling.[@AG] The noise of the cotunneling current through one or several tunnel-coupled quantum dots in the Coulomb blockade regime was calculated in Ref. . ![(Color online) Schematic view of the dependence of the SET’s conductance (a) and thermopower (b) on the dimensionless gate voltage $C_g V_g/e$. The three curves in panel (b) correspond to different temperatures, the black solid line corresponding to the highest temperature, the blue dash-dotted line to the lowest.[]{data-label="TM"}](TM){width="8cm"} The thermoelectric properties of SETs have been investigated in part, both theoretically and experimentally. The thermoelectric kinetic coefficient $G_T=I/\delta{T}$ is defined as the response of the source-drain electric current $I$ to a small temperature difference $\delta{T}$ between the source and the drain at zero voltage, $V_{sd}=0$. The thermopower $S=-V_{sd}/\delta{T}$ determines the voltage response to $\delta{T}$ at zero electric current, $I=0$ (that is, with disconnected external circuit in Fig. \[set\]). In the sequential tunneling regime, the thermopower was predicted to exhibit periodic sawtooth oscillations as a function of $V_g$ (see Ref. ), as shown schematically in Fig. \[TM\]b (black solid line). This sawtooth behavior has been observed experimentally, [@BSexp; @Moller1998] however deviations from it have also been seen.[@Dzurak1993; @Dzurak1997] The latter observations motivated the theoretical study of the inelastic cotunneling contribution to the thermopower.[@TM] It was shown that below some crossover temperature the thermopower in the valleys of Coulomb blockade is supressed, the sawtooth behavior (black solid line in Fig. \[TM\]b) is strongly modified at low temperatures (blue dash-dotted line in Fig. \[TM\]b). Similar behavior was later observed experimentally.[@TMexp] Taking into account the inelastic cotunneling contribution also leads to the violation of the Wiedemann-Franz law in a SET device, as was theoretically shown in Ref. . The elastic cotunneling contribution to thermopower was discussed in Refs. , but its statistics have not been properly analyzed. At the same time, given the fact that the elastic cotunneling regime gives rise to strong mesoscopic conductance fluctuations, it appears crucial to study the statistics of the thermoelectric coefficients in the elastic cotunneling regime. The purpose of the present work is to perform such a study. We consider thermoelectric transport through a small metallic island containing many electrons, whose discrete single-particle energy spectrum is characterized by the mean level spacing $\Delta\ll{E}_C$. This spectrum, as well as the coherent electron motion inside the dot, are assumed to be described by the orthogonal ensemble of the random matrix theory, corresponding to the absence of any external magnetic field. This assumption is valid as long as $E_C$ is small compared to the Thouless energy of the island. At low temperatures, $T\ll\sqrt{E_C \Delta}$, we can neglect the contribution of the inelastic cotunneling.[@AN] Under these assumptions, we determine (in the elastic cotunneling regime) the full statistics of the thermoelectric kinetic coefficient $G_T^{el}$ and of the thermopower, $S^{el}=G_T^{el}/G^{el}$, and show that the fluctuations of $S^{el}$ are much larger than the average value. The paper is organized as follows. In Section \[results\], we summarize the main results and discuss them qualitatively. We specify the model in Section \[model\]. The detailed calculations are presented in Section \[calculations\]. We summarize results in Section \[conclusions\]. Qualitative discussion and summary of the main results {#results} ====================================================== Very generally, the linear response of charge and energy currents, $I$ and $J_E$, to the voltage and temperature differences, $V$ and $\delta{T}$, is determined by the $2\times{2}$ matrix of the kinetic coefficients, $$\label{kinetic=} \left(\begin{array}{c} I \\ J_E \end{array}\right)= \left(\begin{array}{cc} G & G_T \\ K_V & K_T \end{array}\right) \left(\begin{array}{c} V \\ \delta{T} \end{array}\right),$$ where $K_V=-TG_T$ due to the Onsager symmetry holding under time-reversal symmetry, thermopower $S=G_T/G$, and the thermal conductance $K$ is given by $K=K_T-K_VG_T/G$. As will be shown in Sec. \[calculations\], for a given realization of the disorder or of the shape of the island and a given value of the gate voltage, the three independent kinetic coefficients in the regime of the elastic cotunneling can be represented in the form $$\begin{aligned} &G^{el}=\frac{G_sG_d}{4\pi{e}^2/\hbar}\,\frac\Delta{E_C}\,\tau^2(E_F), \label{G=tau}\\ &G_T^{el}=-\frac{\pi^2T}{3e}\,\frac{G_sG_d}{4\pi{e}^2/\hbar}\, \frac\Delta{E_C}\,\frac{d\tau^2(E_F)}{dE_F},\label{GT=tau}\\ &K_T^{el}=\frac{\pi^2T}{3e^2}\,\frac{G_sG_d}{4\pi{e}^2/\hbar}\, \frac\Delta{E_C}\,\tau^2(E_F).\end{aligned}$$ Here $G_s(G_d)$ is the conductance of the tunnel junction between the island and the source (drain) electrode. $\tau(E_F)$ is a smooth real dimensionless function of the Fermi energy $E_F$ in the electrodes, which depends on the microscopic realization of disorder or the island shape, such as the one shown in Fig. \[fig:tau\]. For a given realization, it varies on a typical energy scale $E_C$, and its typical value is $\sim{1}$ (provided that the gate voltage is not too close to a degeneracy point). Thus, for a given realization, the kinetic coefficients satisfy the Mott formula for the thermopower and the Wiedemann-Franz law for $K_T^{el}$. This is quite natural, as these relations hold quite generally when the electron scattering is elastic.[@Jonson1980] As $G^{el}$ and $K_T^{el}$ depend on the realization via the same quantity $\tau^2(E_F)$, their statistics is identical. It was found in Ref. , where the distribution function of the elastic cotunneling electrical conductance $G^{el}$ was found explicitly. Introducing the dimensionless variable $g=\tau^2(E_F)$, see Eq. (\[G=tau\]), it coincides with the Porter-Thomas distribution for the orthogonal ensemble,[@PT] $$P(g)=\Theta(g)\sqrt{\frac{1-4x^2}{4\pi{g}}}\, e^{-(1-4x^2)g/4},$$ where $\Theta(g)$ is the Heaviside step function. Here, $x$ is the rescaled gate voltage, such that $x=0$ corresponds to the center of the Coulomb blockade valley, and $x=\pm{1}/2$ corresponds to the two nearby degeneracy points. We restrict $x$ to the interval $-1/2<x<1/2$, outside of which the dependence on the gate voltage should be periodically repeated. The average elastic cotunneling conductance value is given by Eq.(17) in Ref. , $$\label{Gelastic} \langle G^{el} \rangle = \frac{G_sG_d}{4\pi{e}^2/\hbar}\,\frac\Delta{E_C}\, \frac{2}{1 - 4x^2},$$ and the fluctuations are indeed mesoscopically large, of the order of the average conductance, $$\label{VarGelastic} \sqrt{\langle (G^{el})^2 \rangle - \langle G^{el} \rangle^2} = \sqrt{2}\, \langle G^{el} \rangle.$$ ![(Color online) A sketch of possible behavior of the function $\tau(E)$ (upper panel) and the corresponding $\tau^2(E)$ (lower panel), entering Eq. (\[G=tau\]). The scale of the horizontal axis is $\sim{E}_C$, that of the vertical axis is $\sim{1}$.[]{data-label="fig:tau"}](tau){width="8cm"} The quantity $G_T^{el}$ depends on the realization via $d\tau^2(E_F)/dE_F$, therefore its statistics is different from that of $G^{el}$. However, these quantities are correlated. The consequence of this fact for the statistics of the thermopower can be understood from the following simple argument. The function $\tau(E)$ can have arbitrary sign, and it may even change sign at some point $E=E_0$ (Fig. \[fig:tau\]). In the vicinity of $E_0$, it can be approximated as $\tau(E)\approx{A}(E-E_0)$, where the coefficient $A$ is non-singular. If $E_0$ happens to be close to $E_F$, then $G^{el}\propto\tau^2(E_F)\propto(E_F-E_0)^2$ is very small. As $E_0$ is random and determined by the island, while $E_F$ is determined by the electrodes, $E_0$ can be assumed uniformly distributed in the vicinity of $E_F$. This immediately results in the $1/\sqrt{G^{el}}$ behavior of the distribution function of $G^{el}\propto{A}^2(E_F-E_0)^2$ at $G^{el}\to{0}$, as found in Ref. in the orthogonal ensemble. At the same time, the thermoelectric coefficient $G_T^{el}\propto{d}\tau^2(E_F)/dE_F\propto(E_F-E_0)$, so the thermopower $S^{el}=G_T^{el}/G^{el}\propto{1}/(E_F-E_0)$. For a uniformly distributed $E_0$, this gives $\alpha/(S^{el})^2$ for the asymptotics of the distribution function at $S^{el}\to\pm\infty$ with some coefficient $\alpha$. As the coefficient is the same at $S^{el}\to+\infty$ and $S^{el}\to-\infty$, such a distribution has a finite first moment $\langle{S^{el}}\rangle$, but a divergent second moment $\langle{(S^{el})}^2\rangle$, leading indeed to large mesoscopic fluctuations of $S^{el}$. As we have just seen, these large fluctuations are dominated by those realizations where $E_0$ and $E_F$ happen to be close to each other, that is, the electrical conductance is anomalously small. These simple arguments are confirmed by the explicit calculation in Sec. \[calculations\], which gives the average elastic cotunneling thermopower, $$\label{avS=} \langle S^{el} \rangle = - \frac{\pi^2 T}{3 e E_C} \frac{4x}{1 - 4x^2},$$ and divergent higher moments. Note that $\langle S^{el} \rangle=0$ at $x=0$, since in the valley center the system is electron-hole-symmetric on the average. However, for any given realization, the electronic energy spectrum on the island does not have any symmetry, so there is no reason for $S^{el}$ to vanish at $x=0$ in any specific realization. Near the degeneracy points $x=\pm 1/2$, Eq. (\[avS=\]) gives a divergence, but in this region our theory does not work any more as the dominant contribution to the transport comes from the sequential tunneling mechanism. The full distribution function of the thermopower turns out to be a simple Lorentzian, conveniently written in terms of the dimensionless variable $s$, such that $S^{el}=-[\pi^2T/(3eE_C)]s$: $$\label{Sdist=} P(s)=\frac{\sqrt{4/3}}\pi\,\frac{1-4x^2}{4/3+[(1-4x^2)s-4x]^2}.$$ We have also determined the full distribution function of $G_T^{el}$ written here in terms of a dimensionless variable $g_T=E_C\,d\tau^2(E_F)/dE_F$ \[see Eq. (\[GT=tau\])\], $$\begin{aligned} P(g_T) = {}&\frac{\sqrt{3}}{4\pi}\, (1 - 4x^2)^2 \exp\left( \frac{3}{2}\, x (1 - 4x^2)^2 g_T\right)\nonumber \\ &\times K_0\!\left( \frac{\sqrt{3}}{4}\, (1 - 4x^2)^2 \sqrt{1 + 12 x^2}\, \|g_T\| \right),\label{distrib_f1}\end{aligned}$$ where $K_0$ is the modified Bessel function. The corresponding average value is given by $$\begin{aligned} \label{G_Tav} \langle G_T^{el} \rangle = -\frac{\pi^2T}{3e}\, \frac{G_sG_d}{4\pi{e}^2/\hbar}\,\frac\Delta{E_C^2}\, \frac{4x}{(1-4x^2)^2},\end{aligned}$$ and higher moments are given by Eq. (\[XYmoments=\]). Note that $\langle{G}_T^{el}/G^{el}\rangle= 2\langle{G}_T^{el}\rangle/\langle{G}^{el}\rangle$. In Fig. \[distrib\] we plot $P(g_T)$ for two different values of the dimensionless gate voltage $x=0,0.3$. Note that changing the sign $x\to-x$ amounts to $P(g_T)\to{P}(-g_T)$. ![(Color online) The distribution function $P(g_T)$ for $x=0$ (black solid line) and $x=0.3$ (red dash-dotted line).[]{data-label="distrib"}](distrib){width="8.5cm"} As $G^{el}$ and $G_T^{el}$ are correlated random variables, the full information about their statistics at a given value of $x$ is provided by the joint distribution function, for which we have obtained the following analytical expression: $$\begin{aligned} P(g,g_T) = {}&\frac{\sqrt{3}\,\Theta(g)}{8\pi{g}}\,(1 - 4x^2)^2 \exp\left(-\frac{1-4x^2}{4}\,g\right)\nonumber\\ {}&{}\times\exp \left\{-\frac{3(1-4x^2)[4x(1-4x^2)g-g_T]^2}{16g}\right\}.\label{PggT=}\end{aligned}$$ To characterize statistical correlations at different values of $x$, one should consider $g(x)$ and $g_T(x)$ as two correlated random processes. They can be conveniently characterized in terms of $$\label{tauvep=} \tau_x=\sqrt{g(x)}=\tau(E_F),\quad \varepsilon_x=\frac{g_T(x)}{2\sqrt{g(x)}} =E_C\,\frac{d\tau(E_F)}{dE_F},$$ which turn out to be Gaussian random processes with zero averages and pair correlators $$\begin{aligned} \label{avtautau=} \langle\tau_{x_1}\tau_{x_2}\rangle={}&{}\frac{1}{2(x_1-x_2)}\, \ln\left(\frac{1+2x_1}{1-2x_1}\,\frac{1-2x_2}{1+2x_2}\right),\\ \langle\tau_{x_1}\varepsilon_{x_2}\rangle={}&{} \frac{1}{(x_1-x_2)(1-4x_2^2)}+{}\nonumber\\ &{}+\frac{1}{4(x_1-x_2)^2} \ln\left(\frac{1+2x_1}{1-2x_1}\,\frac{1-2x_2}{1+2x_2}\right),\\ \langle\varepsilon_{x_1}\varepsilon_{x_2}\rangle={}&{} \frac{1-2x_1^2-2x_2^2}{(x_1-x_2)^2(1-4x_1^2)(1-4x_2^2)}-{}\nonumber\\ &{}-\frac{1}{4(x_1-x_2)^3} \ln\left(\frac{1+2x_1}{1-2x_1}\,\frac{1-2x_2}{1+2x_2}\right). \label{avvepvep=}\end{aligned}$$ The divergence of mesoscopic fluctuations of the thermopower, found in the present work, originates from the fact that the elastic cotunneling contribution to the electrical conductance, calculated in the leading order in the tunneling couplings, has too high a probability to vanish. Indeed, as discussed in the paragraph preceding Eq. (\[avS=\]), both $G^{el}$ and $G_T^{el}$ may be small for some realizations, but it is easier for $G^{el}$ to have an anomalously small value, than for $G_T^{el}$, and then their ratio $S^{el}=G_T^{el}/G^{el}$ becomes anomalously large. One should recall, however, that the leading-order elastic cotunneling is not the only contribution to the conductance. There are other contributions (e. g., the inelastic cotunneling, or higher-order contributions to the elastic one), which work as parallel conduction channels, so the conductance never vanishes exactly. These contributions will cut off the divergence of $\langle{S}^2\rangle$. Nevertheless, the fluctuations will still be parametrically large. To estimate the magnitude of the effect, let us assume that the elastic cotunneling conductance is shunted by the inelastic one,[@AN] $G^{in}\sim(G_sG_d\hbar/e^2)(T^2/E_C)^2$. The thermopower fluctuations will be determined by those realizations which have $G^{el}\sim{G}^{in}$, that is, $g\sim{T}^2/(E_c\Delta)\ll{1}$. Then the typical value of $g_T\sim\sqrt{g}$, as seen from Eq. (\[PggT=\]). Thus, we can estimate $\sqrt{\langle{S}^2\rangle}/\langle{S}\rangle\sim{g}_T/g\sim \sqrt{E_C\Delta}/T$. This factor is large precisely in the regime when the elastic cotunneling dominates over the inelastic one. Taking the values corresponding to the experiment of Ref. , $E_C=1.5$ meV, $\Delta=0.05$ meV, $T\sim\Delta\approx{0}.6$ K, and $G_s = G_d = 0.012\,e^2/\hbar$, we have $\sqrt{E_C\Delta}/T\sim{5}$. At lower temperatures, $T\ll\Delta$, the inelastic cotunneling is suppressed even stronger, so the fluctuations of the thermopower will be even larger. At the same time, the effect of elastic cotunneling on the average thermopower is not very dramatic. Even at $T\ll\sqrt{E_C\Delta}$ when the inelastic contributions $G^{in},G_T^{in}$ are small compared to the typical values of the elastic ones, $G^{in}\ll{G^{el}}$, $\|G_T^{in}\|\ll\|G_T^{el}\|$, the ratios $S^{in}=G_T^{in}/G^{in}$ and $\langle{S}^{el}\rangle=\langle{G}_T^{el}/G^{el}\rangle$ are of the same order. Indeed, the average elastic cotunneling thermopower, given by Eq. (\[avS=\]), differs from the inelastic one, given by Eq. (23) in Ref. , $$\label{SinTM} S^{in} = - \frac{4 \pi^2 T}{5 e E_C} \frac{4x}{1 - 4x^2},$$ just by a constant factor $12/5$. To illustrate this effect, we include sequential tunneling and inelastic cotunneling contributions to the conductance ($G^{sq},G^{in}$) and to the thermoelectric kinetic coefficient ($G_T^{sq},G_T^{in}$), and calculate the average of the total thermopower, $$\label{avStot=} \langle{S}\rangle= \left\langle\frac{G_T^{sq}+G_T^{in}+G_T^{el}}{G^{sq}+G^{in}+G^{el}}\right\rangle,$$ which is straightforwardly evaluated from the joint distribution function (\[PggT=\]). Taking $G^{sq}$ and $G_T^{sq}$ from Eqs. (13) and (14) in Ref. , respectively, and $G^{in}$ and $G_T^{in}$ from Eq. (10) in Ref. and Eq. (22) in Ref. , we plot in Fig. \[FStot\] the average $\langle{S}\rangle$ for the values of the parameters from the experiment of Ref. , listed above, with and without elastic cotunneling contributions. Thus, the qualitative shape of the dependence of the average $\langle{S}\rangle$ on the gate voltage is the same as in the elastic cotunneling case, shown in Fig. \[TM\]b. However, for any specific realization of the quantum dot, the dependence of $S$ will be different. In particular, there is no reason why it would vanish exactly in the center of the valley. ![(Color online) The averaged total thermopower with (black solid line) and without (red dashed line) taking into account the elastic cotunneling contiributions. The difference of two curves is most visible in the interval $-0.2 < x < 0.2$. See the text for details.[]{data-label="FStot"}](s){width="8.5cm"} The model {#model} ========= We model the single-electron transistor using the standard Hamiltonian [@AleinerRev], $$\begin{aligned} &\hat{H}=\hat{H}_0+\hat{H}_{Ts}+\hat{H}_{Td},\\ &\hat{H}_0 = \sum_{\alpha=s,d}\sum_n \xi_{\alpha,n} \hat{c}_{\alpha,n}^\dagger \hat{c}_{\alpha,n} + \sum_k \epsilon_k \hat{b}_k^\dagger \hat{b}_k + \hat{H}_C,\label{H0=}\\ &\hat{H}_{T\alpha} = \sum_{k,n} \left( t_{\alpha,kn} \hat{c}^\dagger_{\alpha,n} \hat{b}_k + t_{\alpha,kn}^* \hat{b}^\dagger_k \hat{c}_{\alpha,n} \right). \label{tunnel}\end{aligned}$$ Here the subscript $\alpha=s,d$ labels the two electrodes (source and drain, respectively), $n$ and $k$ label the single-electron states in the leads and on the island, respectively. For the sake of compactness, we suppress the spin indices. As we will not consider spin-flip processes, all subsequent calculations can be understood as performed for a given spin projection, and the final expressions for the transport coefficients will be multiplied by 2. $\hat{c}_{\alpha,n}$ and $\hat{b}_k$ are the electron annihilation operators for the corresponding states, and the corresponding single-electron energies $\xi_{\alpha,n},\epsilon_k$ are measured from the Fermi level. $\hat{H}_C$ in Eq. (\[H0=\]) is the Coulomb interaction Hamiltonian for electrons on the island, obtained from the standard considerations.[@NBbook] Namely, the electrostatic energy $E_{el}$ is assumed to be determined by the total electric charge $Q=-Ne$ on the island, $E_{el}(Q)=\int^Q\varphi(Q')\,dQ'$, where $\varphi(Q)=Q/C+C_gV_g/C$ is the electrostatic potential on the island, and $C=C_g + C_s + C_d + C_i$ is the total capacitance of the island, given by the sum of the capacitances to the gate ($C_g$), source ($C_s$), and drain ($C_d$) electrodes, as well as the self-capacitance $C_i$ of the island. Thus, $\hat{H}_C$ can be written as $$\label{HC=} \hat{H}_C=E_C(\hat{N}^2-2\hat{N}C_g V_g/e),\quad \hat{N}=\sum_k\left[\hat{b}_k^\dagger\hat{b}_k-\Theta(-\epsilon_k)\right].$$ Here $\Theta(-\epsilon)$ is the Heaviside step function, so $\hat{N}$ is the operator of the excess number of electrons on the island, the charge of the filled Fermi sea at $\epsilon_k<0$ assumed to be compensated by the neutralizing background. The degeneracy between states with $N$ and $N+1$ on the island occurs when the dimensionless gate voltage $x$ is half-integer: $$x\equiv\frac{C_g V_g}{e}=N+1/2.$$ As the dependence of the transport coefficients on $V_g$ is periodic, we can restrict our attention to the interval $-1/2<x<1/2$, where the Coulomb energy is minimized by $N=0$. Thus, $x$ measures the relative distance from the center of the Coulomb blockade valley. It is convenient to introduce the Coulomb energy cost of adding/removing one electron to/from the island, $$\label{Epm=} E_\pm=E_{el}(N=\pm{1})-E_{el}(N=0)=E_C(1\mp{2x}).$$ The matrix elements $t_{\alpha,kn}$ describe weak tunneling coupling between the electrodes and the island. Analogously to Ref. , we assume that this coupling is due to small overlap between the wave functions in the island and in each electrode $\alpha$, dominated by the vicinity of a single point $\mathbf{r}_\alpha$, where the island touches the electrode $\alpha$. Then the tunneling Hamiltonian can be assumed to have the form $$\hat{H}_{T\alpha} = t_\alpha\, \hat\psi_\alpha^\dagger(\mathbf{r}_\alpha) \hat\Psi(\mathbf{r}_\alpha) +t_\alpha^\,\hat\Psi^\dagger(\mathbf{r}_\alpha)\, \hat\psi_\alpha(\mathbf{r}_\alpha),$$ where $\hat\Psi(\mathbf{r})$ and $\hat\psi_\alpha(\mathbf{r})$ are the fermionic field operators for the electrons on the island and in the contacts, respectively, and $t_\alpha$ are the tunneling amplitudes incorporating all necessary normalization factors. Expanding the fermionic operators in terms of the corresponding single-particle wave functions, $\Psi_k(\mathbf{r})$ and $\psi_{\alpha,n}(\mathbf{r})$, as $$\hat\Psi(\mathbf{r})=\sum_k\hat{b}_k\,\Psi_k(\mathbf{r}), \quad \hat\psi_\alpha(\mathbf{r})=\sum_n\hat{c}_{\alpha,n} \psi_{\alpha,n}(\mathbf{r}),$$ we obtain the following simple expression for the matrix elements $t_{\alpha,kn}$: $$\label{takn=} t_{\alpha,kn}=t_\alpha\,\Psi_k(\mathbf{r}_\alpha)\, \psi_{\alpha,n}^(\mathbf{r}_\alpha).$$ The energies $\xi_{\alpha,n}$ are assumed to have continuous spectra, so the leads are characterized by the local densities of states (per spin) $$\label{nualpha=} \nu_\alpha(\epsilon)=\sum_n\|\psi_{\alpha,n}(\mathbf{r}_\alpha)\|^2\,\delta(\epsilon-\xi_{\alpha,n}),$$ assumed to be self-averaging and energy-independent. The energies $\epsilon_k$ of the single-particle states on the island are discrete with the mean level spacing $\Delta$ ($2/\Delta$ being the ensemble average of the single-electron density of states on the island for both spin projections). The island wave functions $\Psi_k(\mathbf{r}_\alpha)$ are assumed to be real random variables, not correlated with the energies $\epsilon_k$, and corresponding to the elements of a random orthogonal matrix uniformly distributed in the orthogonal group. To the leading order in the matrix size, they can be treated as real independent Gaussian random variables,[@AleinerRev] whose statistics is entirely determined by the pair correlator $$\label{corrPsiPsi=} \langle\Psi_k(\mathbf{r}_\alpha)\,\Psi_{k'}(\mathbf{r}_{\alpha'}) \rangle=\delta_{kk'}\delta_{\alpha\alpha'},$$ all normalization factors being absorbed in the tunneling amplitudes $t_\alpha$ in Eq. (\[takn=\]). Instead of $t_\alpha$, it is convenient to characterize each tunneling contact by a physical quantity, such as its average conductance (including the factor of 2 from spin), $$\label{Galpha=} G_\alpha=2\,\frac{2\pi{e}^2}\hbar\,\frac{\|t_\alpha\|^2\nu_\alpha}\Delta.$$ Calculation {#calculations} =========== Transport coefficients for a given realization ---------------------------------------------- Following Refs. , we start from the Golden Rule expression for the source-drain current, which represents the difference between the rate of electron transfer from state $n$ on the source to the state $m$ on the drain and the rate of the opposite process, $$\begin{aligned} \label{current=FGR} I=&-2e\sum_{n,m}\left[f_s(\xi_{s,n})-f_d(\xi_{d,m})\right]\nonumber\\ &{}\times \|M_{s,n\to{d},m}\|^2\,\frac{2\pi}{\hbar}\,\delta(\xi_{s,n}-\xi_{d,m}),\end{aligned}$$ where we used the fact that the cotunneling matrix element $M_{d,m\to{s},n}=M_{s,n\to{d},m}^$ and took into account the spin degeneracy. In Eq. (\[current=FGR\]), $f_\alpha(\xi)$ is the average occupation probability of the state with energy $\xi$ on the electrode $\alpha$: $$f_\alpha(\xi)=\frac{1}{1+e^{(\xi-\mu_\alpha)/T_\alpha}}.$$ Focusing on the linear response to small chemical potential and temperature differences, $\mu_s-\mu_d=-eV$, $T_s-T_d=\delta{T}$, we can write $$\begin{aligned} &I=\frac{1}{\pi\hbar}\int(-e)\,\mathcal{T}(\xi)\left(-eV+\frac{\xi}{T}\,\delta{T}\right) \left(-\frac{\partial{f}_{eq}}{\partial\xi}\right)d\xi, \label{Ilinear=}\\ &\mathcal{T}(\xi)=4\pi^2\sum_{nm} \|M_{s,n\to{d},m}\|^2\,\delta(\xi_{s,n}-\xi)\,\delta(\xi_{d,m}-\xi), \label{Tdef=}\end{aligned}$$ where $f_{eq}(\xi)=1/[1+\exp(\xi/T)]$. Using the same approach, one can also find the energy current $J_E$ between the source and the drain. The corresponding expression can be obtained from Eq. (\[Ilinear=\]) by simply replacing the factor $(-e)$, which is nothing but the charge transferred in a single tunneling event, by the corresponding transferred energy $\xi$. As will be seen below \[Eq. (\[T=\])\], $\mathcal{T}(\xi)$ is a smooth function of $\xi$ varying on a typical scale of $\xi\sim{E}_C$. At the same time, $-\partial{f}_{eq}/\partial\xi$ is strongly peaked around zero on the scale $\xi\sim{T}$. Thus, at low temperatures, the elastic cotunneling kinetic coefficients, appearing in Eq. (\[kinetic=\]), can be approximated as $$\begin{aligned} &G^{el}=\frac{e^2}{\pi\hbar}\,\mathcal{T}(0),\label{GlowT=}\\ &G_T^{el}=-\frac{e}{\pi\hbar}\,\frac{\pi^2T}{3}\,\mathcal{T}'(0), \label{GTlowT=}\\ &K_T^{el}=\frac{1}{\pi\hbar}\,\frac{\pi^2T}{3}\,\mathcal{T}(0),\end{aligned}$$ where we have used $$\int\left(-\frac{\partial{f}_{eq}}{\partial\xi}\right)d\xi=1,\quad \int\xi^2\left(-\frac{\partial{f}_{eq}}{\partial\xi}\right)d\xi =\frac{\pi^2T^2}{3}.$$ Note that in order to calculate $G_T^{el}$, one has to expand the transmission function $\mathcal{T}(\xi)\approx\mathcal{T}(0)+\xi\,\mathcal{T}'(0)$, as the leading term vanishes due to parity $\xi\to-\xi$. The cotunneling matrix element $M_{s,n\to{d},m}$ is evaluated in the second-order perturbation theory in the tunneling Hamiltonian $\hat{H}_T=\hat{H}_{Ts}+\hat{H}_{Td}$ as[@Glazman1990; @NBbook] $$M_{s,n\to{d},m}=\sum_v\frac{\langle\Phi\|\hat{c}_{d,m}\hat{H}_T\|v\rangle \langle{v}\|\hat{H}_T\hat{c}^\dagger_{s,n}\|\Phi\rangle}{E_\Phi+\xi_{s,n}-E_v},$$ where $\hat{c}^\dagger_{s,n}\|\Phi\rangle$ is the initial state of the system. It is conveniently represented as an extra electron on top of some reference many-body state $\|\Phi\rangle$, defined by the occupation numbers of all single-particle states. The final state is represented as $\hat{c}^\dagger_{d,m}\|\Phi\rangle$, an extra electron on top of [the same]{} reference state $\|\Phi\rangle$, which is the characteristic of the [elastic]{} cotunneling process. The states $\|v\rangle$ are virtual intermediate states with energies $E_v$. As $\hat{H}_T$ changes the number of electrons on the island by one, the states $\|v\rangle$ can belong to two sectors: those with one more electron on the island (which is thus added to some empty single-particle level $k$), and those with one less electron (which is thus removed from some filled single-particle level $k$). Splitting the tunnel Hamiltonian Eq. (\[tunnel\]) as $$\begin{aligned} &\hat{H}_{T\alpha}=\hat{H}_{T\alpha-}+\hat{H}_{T\alpha+},\\ &\hat{H}_{T\alpha-}= \sum_{k,n} t_{\alpha,kn} \hat{c}_{\alpha,n}^\dagger \hat{b}_k =\hat{H}_{T\alpha+}^\dagger,\end{aligned}$$ we note that the first sector can be coupled to the initial state only by the $\hat{H}_{Ts+}$ term, while the second sector only by the $\hat{H}_{Td-}$ term. The energies of the intermediate states in the two sectors are given by $$E_\Phi+E_++\epsilon_k,\quad E_\Phi+E_--\epsilon_k+\xi_{s,n}+\xi_{d,m},$$ respectively, where the Coulomb energies $E_\pm$ are defined in Eq. (\[Epm=\]). Evaluation of the matrix elements gives $$\begin{aligned} \label{M=} M_{s,n\to{d},m}=&\sum_k t_{d,km}t_{s,kn}^\times\nonumber\\ &{}\times\left(\frac{f_k}{E_--\epsilon_k+\xi_{d,m}} -\frac{1-f_k}{E_++\epsilon_k-\xi_{s,n}}\right),\end{aligned}$$ where $f_k$ is the occupation number of the single-electron state $k$ in the many-body state $\|\Phi\rangle$. It should be noted that, strictly speaking, in a given reference state $\|\Phi\rangle$, $f_k$ is either 0 or 1. Then, to obtain the statistics of the transport coefficients, averaging over different reference states $\|\Phi\rangle$ should also be performed, which results in the probability of $f_k=1$ to be given by $f_{eq}(\epsilon_k)$. However, this probability is different from 0 or 1 only in the range of energies $\|\epsilon_k\|\sim{T}$, while, as will be seen later, the sum over $k$ in Eq. (\[M=\]) is contributed by a much wider range, $\|\epsilon_k\|\sim{E}_C$. Thus, we approximate $f_k=\Theta(-\epsilon_k)$. The error introduced by this approximation is small by a factor $T/E_C\ll{1}$. Substituting Eq. (\[M=\]) into Eq. (\[Tdef=\]) and using Eqs. (\[takn=\]), (\[nualpha=\]), and (\[Galpha=\]), we obtain $$\label{T=} \mathcal{T}(\xi)=\frac{\hbar^2G_sG_d}{4e^4} \left\|\sum_k\rho_k\left[ \frac{\Delta\,\Theta(\epsilon)}{E_++\epsilon-\xi} -\frac{\Delta\,\Theta(-\epsilon)}{E_--\epsilon+\xi}\right]\right\|^2,$$ where we denoted $$\rho_k=\Psi_k(\mathbf{r}_s)\,\Psi_k^(\mathbf{r}_d).$$ Using Eqs. (\[GlowT=\]) and (\[GTlowT=\]), the transport coefficients $G^{el}$ and $G_T^{el}$ can be expressed as $$\begin{aligned} &G^{el}=\frac{G_sG_d}{4\pi{e}^2/\hbar}\,X,\quad G_T^{el}=-\frac{\pi^2T}{3e}\,\frac{G_sG_d}{4\pi{e}^2/\hbar}\,Y,\\ \label{Xdef=} &X=\sum_{k,k'} \rho_k\rho_{k'}^\,F(\epsilon_k)\,F(\epsilon_{k'}),\\ \label{Ydef=} &Y=\sum_{k,k'}\rho_k\rho_{k'}^* \left[F(\epsilon_k)\,\tilde{F}(\epsilon_{k'}) +\tilde{F}(\epsilon_k)\,F(\epsilon_{k'})\right],\\ &F(\epsilon)=\frac{\Delta\,\Theta(\epsilon)}{E_++\epsilon} -\frac{\Delta\,\Theta(-\epsilon)}{E_--\epsilon},\label{Fep=}\\ &\tilde{F}(\epsilon) =\frac{\Delta\,\Theta(\epsilon)}{(E_++\epsilon)^2} +\frac{\Delta\,\Theta(-\epsilon)}{(E_--\epsilon)^2}.\label{tFep=}\end{aligned}$$ The random variables $\tau_x,\varepsilon_x$, defined in Eq.(\[tauvep=\]), are represented as $$\label{tauvep=rho} \tau_x=\sqrt{\frac{E_C}\Delta}\sum_k\rho_k\,F(\epsilon_k), \quad \varepsilon_x=\sqrt{\frac{E_C^3}\Delta} \sum_k\rho_k\,\tilde{F}(\epsilon_k).$$ Averaging over realizations --------------------------- The statistics of the transport coefficients will be obtained from the joint moments of $G^{el}$ and $G_T^{el}$, as it was done in Ref. for $G^{el}$ alone. Thus, we need to average products of $X$’s and $Y$’s \[defined in Eqs. (\[Xdef=\]), (\[Ydef=\])\] over the energies $\epsilon_k$ and wave function amplitudes $\Psi_k(\mathbf{r}_\alpha)$. This task is facilitated by the following two considerations. (i) Strictly speaking, the positions of the random energy levels $\epsilon_k$ are correlated.[@MehtaBook] However, these correlations occur on the energy scale $\Delta$. Thus, for any smooth function of energy, $\mathcal{F}(\epsilon)$, varying on the scale $\epsilon\sim{E}_C$, the sum over $\epsilon_k$ will be replaced by the integration over $\epsilon$, $$\label{sumk=} \sum_k\mathcal{F}(\epsilon_k)\to\int\mathcal{F}(\epsilon)\, \frac{d\epsilon}\Delta.$$ This approximation introduces an error which is small by a factor $\Delta/E_C$. Of particular importance for the future calculations will be the following three integrals: $$\begin{aligned} \mathcal{J}_1 &= \int F^2(\epsilon)\,\frac{d\epsilon}\Delta = \frac{\Delta}{E_+} + \frac{\Delta}{E_-}, \label{J1=}\\ \mathcal{J}_2 &= \int F(\epsilon)\, \tilde{F}(\epsilon)\, \frac{d\epsilon}\Delta = \frac{\Delta}{2E_+^2} - \frac{\Delta}{2E_-^2},\label{J2=}\\ \mathcal{J}_3 &= \int \tilde{F}^2(\epsilon)\,\frac{d\epsilon}\Delta = \frac{\Delta}{3E_+^3} + \frac{\Delta}{3E_-^3}.\label{J3=}\end{aligned}$$ (ii) When averaging over the wave function amplitudes $\Psi_k(\mathbf{r}_\alpha)$ (which is independent of the averaging over $\epsilon_k$) using Eq. (\[corrPsiPsi=\]) in the orthogonal ensemble, strictly speaking, all possible pairings should be taken, in accordance with Wick’s theorem for Gaussian random variables. However, the amplitudes enter $X$ and $Y$ via a combination $\rho_k=\Psi_k(\mathbf{r}_s)\,\Psi_k^(\mathbf{r}_d)$. Then, for a product $\rho_{k_1}\ldots\rho_{k_{2n}}$ \[a product of an odd number of factors always vanishes because $\Psi_k(\mathbf{r}_s)$ and $\Psi_{k'}(\mathbf{r}_d)$ are uncorrelated\], those pairings are more important, where $\Psi_{k_1}(\mathbf{r}_s)\ldots\Psi_{k_{2n}}(\mathbf{r}_s)$ are paired exactly in the same way as $\Psi_{k_1}(\mathbf{r}_d)\ldots\Psi_{k_{2n}}(\mathbf{r}_d)$, as it gives the minimal number of constraints on the indices. This happens because each summation is transformed into integration over a range of $\epsilon\sim{E}_C$, and thus produces a large factor $\sim{E}_C/\Delta$, as discussed in the previous paragraph. To illustrate this fact, consider the average $$\begin{aligned} \langle{Y}^2\rangle={}&4\sum_{k_1\ldots{k}_4} \langle \rho_{k_1}\rho_{k_2}\rho_{k_3}\rho_{k_4}\rangle\, F(\epsilon_{k_1})\,F(\epsilon_{k_2})\, \tilde{F}(\epsilon_{k_3})\,\tilde{F}(\epsilon_{k_4})\nonumber\\ ={}&4\sum_{k_1\ldots{k}_4} F(\epsilon_{k_1})\,F(\epsilon_{k_2})\, \tilde{F}(\epsilon_{k_3})\,\tilde{F}(\epsilon_{k_4}) \times{}\nonumber\\ &\times \left[\delta_{k_1k_2}\delta_{k_3k_4}+\delta_{k_1k_3}\delta_{k_2k_4} +2\delta_{k_1k_2}\delta_{k_1k_3}\delta_{k_1k_4}\right]\nonumber\\ ={}&4\sum_{k,k'}F^2(\epsilon_k)\,\tilde{F}^2(\epsilon_{k'}) +{}\nonumber\\ &+4\sum_{k,k'}F(\epsilon_k)\,\tilde{F}(\epsilon_k)\, F(\epsilon_{k'})\,\tilde{F}(\epsilon_{k'})+{}\nonumber\\ &+8\sum_kF^2(\epsilon_k)\,\tilde{F}^2(\epsilon_k)\nonumber\\ ={}&4(\mathcal{J}_1\mathcal{J}_3+\mathcal{J}_2^2)[1+O(\Delta/E_C)].\end{aligned}$$ The third term contains a single sum instead of a double sum, so its contribution is smaller by a factor $\Delta/E_C$. Neglecting those pairings, which produce extra constraints on the indices, is equivalent to treating $\rho_k$’s as real independent Gaussian random variables with the pair correlator $$\langle\rho_k\rho_{k'}\rangle=\delta_{kk'}.$$ Armed with this knowledge, we are now ready to calculate an arbitrary joint moment, $$\begin{aligned} \langle{X}^pY^q\rangle={}&2^q\sum_{k_1\ldots{k}_{2p+2q}} \langle\rho_{k_1}\ldots\rho_{k_{2p+2q}}\rangle \nonumber\\ &{}\times F_1\ldots{F}_{2p+q}\tilde{F}_{2p+q+1}\ldots \tilde{F}_{2p+2q},\end{aligned}$$ where we denoted $F(\epsilon_{k_i})=F_i$, $\tilde{F}(\epsilon_{k_i})=\tilde{F}_i$ for compactness. Evaluation of the average $\langle\rho_{k_1}\ldots\rho_{k_{2p+2q}}\rangle$ amounts to summation over all possible pairings of $\rho_k$’s. Each pairing of $\rho_k$’s induces a pairing of $F_k$’s and $\tilde{F}_k$’s, then the subsequent summation over the corresponding $k$ index is performed independently from other indices, thereby producing a factor $\mathcal{J}_1$ for $FF$, $\mathcal{J}_2$ for $F\tilde{F}$, and $\mathcal{J}_3$ for $\tilde{F}\tilde{F}$, according to Eqs. (\[sumk=\]) and (\[J1=\])–(\[J3=\]). ![(Color online) An example of pairing $2p+q$ $F$’s (shown by empty circles) and $q$ $\tilde{F}$’s (filled circles). []{data-label="fig:pairing"}](pairing){width="8cm"} Let us classify all possible pairings by the number $2l$ of all $\tilde{F}$’s which are paired among themselves, so that the remaining $q-2l$ $\tilde{F}$’s are paired with $F$’s, and the remaining $2p+2l$ $F$’s are also paired among themselves, as illustrated graphically in Fig. \[fig:pairing\]. Obviously, $0\leqslant{l}\leqslant{q}/2$, and all pairings with a given $l$ give a factor $\mathcal{J}_1^{p+l}\mathcal{J}_2^{q-2l}\mathcal{J}_3^l$. To determine the combinatorial coefficient, we first note that there are $q!/[(2l)!\,(q-2l)!]$ ways to choose $2l$ $\tilde{F}$’s out of $q$, and $(2p+q)!/[(q-2l)!\,(2p+2l)!]$ ways to choose $2p+2l$ $F$’s out of $2p+q$. Then, there are $(q-2l)!$ ways to pair up $q-2l$ $F$’s with $q-2l$ $\tilde{F}$’s, $(2l-1)!!=(2l)!/(2^ll!)$ ways to pair up the $2l$ $\tilde{F}$’s among themselves, and $(2p+2l-1)!!=(2p+2l)!/[2^{p+l}(p+l)!]$ ways to pair up the $2p+2l$ $F$’s. As a result, $$\label{XYmoments=} \langle{X}^pY^q\rangle=\sum_{0\leqslant{l}\leqslant{q}/2} \frac{2^{q-2l-p}\,q!\,(2p+q)!}{l!\,(q-2l)!\,(p+l)!}\, \mathcal{J}_1^{p+l}\mathcal{J}_2^{q-2l}\mathcal{J}_3^l.$$ From the joint moments, the full statistics can be reconstructed following the standard procedure (given in detail in the Appendix). Conclusions =========== We have developed a quantitative theory of thermoelectric transport in a single electron transistor, consisting of a quantum dot weakly coupled to two electronic leads and capacitively coupled to a gate electrode, in the regime of elastic cotunneling. In this regime, the transport coefficients strongly depend on the realization of the random impurity potential or the shape of the island. We assumed the quantum dot wave functions to be Gaussian random variables and used the random-matrix theory for the orthogonal ensemble (i.e. in the absence of the magnetic field). The distribution function of the conductance $G$ was previously obtained by Aleiner and Glazman in Ref. . We have extended this result and calculated the distributions of the thermopower $S$, the thermoelectric kinetic coefficient $G_T$, and the joint distribution function of the conductance and thermoelectric kinetic coefficient as functions of the gate voltage. Statistical correlations of $G$ and $G_T$ at different values of the gate potential were also calculated. Finally, we have calculated the average elastic cotunneling values of the thermopower and thermoelectric kinetic coefficient and the average values of all moments of $G_T$. We have shown that the second and higher moments of the thermopower diverge, which leads to large mesoscopic fluctuations of the elastic cotunneling thermopower. This divergence is cut off by taking into account the inelastic cotunneling contribution, or higher-order contributions to the elastic one. Nevertheless, the fluctuations will still be parametrically large. We have estimated the magnitude of these fluctuations, taking into account the experimental parameters from Ref. , $\sqrt{\langle{S}^2\rangle}/\langle{S}\rangle \sim 5$. Therefore for any specific realisation of the quantum dot, the dependence of $S$ on the gate voltage will be different from $\langle{S}\rangle$. In particular, there is no reason why it would vanish exactly in the center of the Coulomb blockade valley. Acknowledgements ================ The authors are grateful to D.V. Averin, V. Bubanja, R. Whitney, and C. Winkelmann for helpful discussions. This work was supported by European Union Seventh Framework Programme (FP7/2007–2013) under grant agreement “INFERNOS” No. 308850, as well as by Institut Universitaire de France. Statistics from joint moments ============================= From the joint moments (\[XYmoments=\]), we first reconstruct the characteristic function: $$\begin{aligned} \chi(u,v)={}&\left\langle{e}^{-iuX-ivY}\right\rangle ={}\sum_{p,q=0}^\infty\frac{(-iu)^p(-iv)^q}{p!\,q!}\, \langle{X}^pY^q\rangle\nonumber\\ ={}&\sum_{p=0}^\infty\sum_{l=0}^\infty\sum_{q=2l}^\infty \frac{(2p+q)!}{p!\,(p+l)!\,l!\,(q-2l)!}\,\nonumber\\ &{}\times(-iu\mathcal{J}_1/2)^p (-v^2\mathcal{J}_1\mathcal{J}_3)^l(-2iv\mathcal{J}_2)^{q-2l} \nonumber\\ ={}&\frac{1}{1+2iv\mathcal{J}_2} \sum_{p=0}^\infty\sum_{l=0}^\infty \frac{(2p+2l)!}{p!\,(p+l)!\,l!}\,\nonumber\\ &{}\times\left[\frac{-iu\mathcal{J}_1/2}{(1+2iv\mathcal{J}_2)^2}\right]^p \left[\frac{-v^2\mathcal{J}_1\mathcal{J}_3}{(1+2iv\mathcal{J}_2)^2}\right]^l\nonumber\\ ={}&\frac{1}{1+2iv\mathcal{J}_2}\sum_{s=0}^\infty \frac{(2s)!}{(s!)^2}\left[ \frac{-iu\mathcal{J}_1/2-v^2\mathcal{J}_1\mathcal{J}_3}{(1+2iv\mathcal{J}_2)^2}\right]^s\nonumber\\ ={}&\frac{1}{\sqrt{(1+2iv\mathcal{J}_2)^2+2iu\mathcal{J}_1 +4v^2\mathcal{J}_1\mathcal{J}_3}}.\end{aligned}$$ The sums were calculated using the following relations: $$\begin{aligned} &\sum_{n=0}^\infty\frac{(m+n)!}{n!}\,z^n=\frac{m!}{(1-z)^{m+1}},\\ &\sum_{m,n=0}\frac{\mathcal{F}(m+n)}{m!\,n!}\,x^my^n= \sum_{N=0}^\infty\sum_{k=0}^N\frac{\mathcal{F}(N)}{k!\,(N-k)!}\, x^ky^{N-k}\nonumber\\ &{}\hspace{3.5cm}= \sum_{N=0}^\infty\frac{\mathcal{F}(N)}{N!}\,(x+y)^N,\\ &\sum_{n=0}^\infty\frac{(2n)!}{(n!)^2}\,z^n= \frac{1}{\sqrt{1-4z}}.\end{aligned}$$ From the characteristic function, the probability distributions can be determined. For the conductance, $$P(X)=\int\frac{du}{2\pi}\,e^{iuX}\,\chi(u,0) =\frac{\Theta(X)\,e^{-X/(2\mathcal{J}_1)}}{\sqrt{2\pi\mathcal{J}_1X}}$$ coincides with the result of Ref. for the orthogonal ensemble. For the thermoelectric kinetic coefficient, $$\begin{aligned} P(Y)={}&\int\frac{dv}{2\pi}\,e^{ivY}\,\chi(0,v)\nonumber\\ ={}&\frac{1}{2\pi\sqrt{\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2}}\, \exp\left(\frac{\mathcal{J}_2}{\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2}\,\frac{Y}2\right) \nonumber\\ &{}\times K_0\!\left(\frac{\sqrt{\mathcal{J}_1\mathcal{J}_3}}{\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2}\,\frac{Y}2\right), \label{Ydist=}\end{aligned}$$ where $K_0$ is the modified Bessel function. Here it was important that $\lambda=\mathcal{J}_1\mathcal{J}_3/\mathcal{J}_2^2 =1+1/(12x^2)\geqslant{1}$, so the following relation could be used: $$\begin{aligned} &\int\limits_{-\infty}^{\infty} \frac{d q}{2\pi}\, \frac{e^{i q z}}{\sqrt{ (1 + iq)^2 - \lambda (iq)^2 }}\nonumber \\ &= \frac{1}{\pi \sqrt{\lambda - 1}} \exp\left( \frac{z}{ \lambda - 1} \right) K_0\!\left( \|z\| \frac{\sqrt{\lambda}}{\lambda - 1} \right).\end{aligned}$$ In combination with Eqs. (\[J1=\])–(\[J3=\]) and with the facts that $\mathcal{J}_1\mathcal{J}_3=(\Delta^2/E_C^4)(4/3)(1+12x^2)/(1-4x^2)^4$, and $\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2=(\Delta^2/E_C^4)(4/3)/(1-4x^2)^4$, Eq. (\[Ydist=\]) gives Eq. (\[distrib\_f1\]). The joint probability distribution is given by $$\begin{aligned} P(X,Y)={}&\int\frac{du}{2\pi}\,\frac{dv}{2\pi}\,e^{iuX+ivY}\, \chi(u,v)=\nonumber\\ ={}&\frac{\Theta(X)}{\sqrt{2\pi\mathcal{J}_1X}} \int\frac{dv}{2\pi}\times{}\nonumber\\ &{}\times{e}^{ivY-[(1+2iv\mathcal{J}_2)^2 +4v^2\mathcal{J}_1\mathcal{J}_3]X/(2\mathcal{J}_1)}\nonumber\\ ={}&\frac{\Theta(X)}{4\pi{X}\sqrt{\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2}}\times{}\nonumber\\ &{}\times\exp\left[-\frac{X}{2\mathcal{J}_1} -\frac{(2\mathcal{J}_2X-\mathcal{J}_1Y)^2}{8(\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2)\mathcal{J}_1X}\right].\end{aligned}$$ From this, the distribution function for the thermopower can be obtained by introducing the variable $Z=Y/X$, $$\begin{aligned} P(Z)={}&{}\int\delta(Z-Y/X)\,P(X,Y)\,dX\,dY\nonumber\\ ={}&{}\frac{(2\mathcal{J}_1/\pi) \sqrt{\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2}}{4(\mathcal{J}_1\mathcal{J}_3-\mathcal{J}_2^2) +(2\mathcal{J}_2-\mathcal{J}_1Z)^2},\end{aligned}$$ which gives Eq. (\[Sdist=\]). Finally, to describe the correlations of the random processes $\tau_x,\varepsilon_x$ at different $x$, it is sufficient to use representation (\[tauvep=rho\]), and calculate the characteristic functional $$\begin{aligned} \mathcal{X}[u(x),v(x)]={}&{} \left\langle\exp\left\{i\int[u(x)\,\tau_x+v(x)\,\varepsilon_x]\, dx\right\}\right\rangle=\nonumber\\ ={}&{}\exp\left\{-\frac{1}{2}\int{d}x_1\,dx_2\, \mathcal{K}(x_1,x_2)\right\},\label{calX=}\end{aligned}$$ $$\begin{aligned} \mathcal{K}(x_1,x_2)={}&{} u(x_1)\,u(x_2)\,\frac{E_C}\Delta\sum_k F_{x_1}(\epsilon_k)\,F_{x_2}(\epsilon_k)+{}\nonumber\\ &{}+2u(x_1)\,v(x_2)\,\frac{E_C^2}\Delta\sum_k F_{x_1}(\epsilon_k)\,\tilde{F}_{x_2}(\epsilon_k)+{}\nonumber\\ &{}+v(x_1)\,v(x_2)\,\frac{E_C^3}\Delta\sum_k \tilde{F}_{x_1}(\epsilon_k)\,\tilde{F}_{x_2}(\epsilon_k), \label{calK=}\end{aligned}$$ where the subscripts at $F(\epsilon_k)$, $\tilde{F}(\epsilon_k)$ indicate that the values $E_+,E_-$, entering in Eqs. (\[Fep=\]), Eqs. (\[tFep=\]), should be taken at the corresponding value of $x$, see Eq. (\[Epm=\]). Eqs. (\[calX=\]), (\[calK=\]) represent the characteristic functional of a pair of Gaussian random processes, whose pair correlators are given by Eqs. (\[avtautau=\])–(\[avvepvep=\]), obtained by evaluation of the sums in Eq. (\[calK=\]) using the rule (\[sumk=\]). [99]{} H. van Houten, L.W. Molenkamp, C. W. J. Beenakker, and C. T. Foxon, Semicond. Sci. Tech. 7, B215 (1992). C. L. Kane and M. P. A. Fisher, Phys. Rev. Lett. 76, 3192 (1996). R. Fazio, F. W. J. Hekking, and D. E. Khmelnitskii, Phys. Rev. Lett. 80, 5611 (1998). G. Catelani and I. L. Aleiner, Sov. Phys. JETP 100, 331 (2005). F. Giazotto, T. T. Heikkilä, A. Luukanen, A. M. Savin, and J. P. Pekola, Rev. Mod. Phys. 78, 217 (2006). J. T. Muhonen, M. Meschke, and J. P. Pekola, Rep. Prog. Phys. 75, 046501 (2012). H. Courtois, F. W. J. Hekking, H. Q. Nguyen, C. B. Winkelmann, J. Low Temp. Phys. 175, 799 (2014). A. S. Vasenko, E. V. Bezuglyi, H. Courtois, and F. W. J. Hekking, Phys. Rev. B 81, 094513 (2010). A. Ozaeta, A. S. Vasenko, F. W. J. Hekking, and F. S. Bergeret, Phys. Rev. B 85, 174518 (2012). F. J. DiSalvo, Science 285, 703 (1999). A. Shakouri, Annu. Rev. Mater. Res. 41, 399 (2011). R. S. Whitney, Phys. Rev. Lett. 112, 130601 (2014). B. Sothmann, R. Sánchez, and A. N. Jordan, Nanotechnology 26, 032001 (2015). I. L. Aleiner, P. W. Brouwer and L. I. Glazman, Phys. Rep. 358, 309 (2002). I. O. Kulik and R. I. Shekhter, Sov. Phys. JETP 41 308 (1975). C. W. J. Beenakker, Phys. Rev. B 44, 1646 (1991). D. V. Averin and A. A. Odintsov, Sov. Phys. JETP 69, 766 (1989). L. I. Glazman, K. A. Matveev, JETP Letters 51, 484 (1990). D. V. Averin and Yu. V. Nazarov, Phys. Rev. Lett. 65, 2446 (1990). D. V. Averin and Yu. V. Nazarov, in [Single Charge Tunneling: Coulomb Blockade Phenomena in Nanostructures]{}, ed. by H. Grabert and M. H. Devoret, NATO ASI Series B 294 (Plenum, New York, 1992), pp. 217-247. Provided that $N$ is even. In the case of odd $N$, the change in the internal state of the island may reduce to flipping the spin of the unpaired electron; this leads to the Kondo effect which enhances the conductance in the odd-$N$ valleys \[L. I. Glazman and M. E. Raikh, JETP Lett. 47, 452 (1988); T. K. Ng and P. A. Lee, Phys. Rev. Lett. 61, 1768 (1988)\]. I. L. Aleiner and L. I. Glazman, Phys. Rev. Lett. 77, 2057 (1996). E. V. Sukhorukov, G. Burkard, and D. Loss, Phys. Rev. B 63, 125315 (2001). C. W. J. Beenakker and A. A. M. Staring, Phys. Rev. B 46, 9667 (1992). A. A. M. Staring, L. W. Molenkamp, B. W. Alphenhaar, H. van Houten, O. J. A. Buyk, M. A. A. Mabesoone, C. W. J. Beenakker, and C. G. Foxon, Europhys. Lett. 22, 57 (1993). S. Möller, H. Buhmann, S. F. Godijn, and L. W. Molenkamp, Phys. Rev. Lett. 81, 5197 (1998). A. S. Dzurak, C. G. Smith, M. Pepper, D. A. Ritchie, J. E. F. Frost, G. A. C. Jones, and D. G. Hasko, Solid State Comm. 87, 1145 (1993). A. S. Dzurak, C. G. Smith, C. H. W. Barnes, M. Pepper, L. Martín-Moreno, C. T. Liang, D. A. Ritchie, and G. A. C. Jones, Phys. Rev. B 55, R10197 (1997). M. Turek and K. A. Matveev, Phys. Rev. B 65, 115332 (2002). R. Scheibner, E. G. Novik, T. Borzenko, M. König, D. Reuter, A. D. Wieck, H. Buhmann, and L. W. Molenkamp, Phys. Rev. B 75, 041301(R) (2007). B. Kubala, J. König, and J. Pekola , Phys. Rev. Lett. 100, 066801 (2008). J. Koch, F. von Oppen, Y. Oreg, and E. Sela, Phys. Rev. B 70, 195107 (2004). G. Billings, A. D. Stone, and Y. Alhassid, Phys. Rev. B 81, 205303 (2010). M. Jonson and G. D. Mahan, Phys. Rev. 181 1336 (1980). C. E. Porter and R. G. Thomas, Phys. Rev. 104, 483 (1956). Yu. V. Nazarov and Ya. M. Blanter, Quantum transport, (Cambridge University Press, 2009). M. L. Mehta, Random matrices (Elesevier, Amsterdam, 2004).	{ "pile_set_name": "ArXiv" }
--- abstract: 'This is an introduction to work of Hopkins, Kuhn, and Ravenel \[9\] on generalized group characters, which seems to fit very well with the theory of what physicists call higher twisted sectors in the theory of orbifolds.' address: 'Department of Mathematics, Johns Hopkins University, Baltimore, Maryland 21218' author: - Jack Morava date: 15 May 2004 title: HKR characters and higher twisted sectors --- [^1] Basic definitions ================= This paper is meant to be expository. Its first main point is that the [inertia stack]{} construction, which has been a focus of considerable attention in recent work on orbifolds and stacks, can be iterated to define a simplicial object ${{\mathbb I}}^\bullet({{\mathcal X}})$, which is a convenient device for organizing the ‘higher twisted sectors’ of the cohomology of an orbifold ${{\mathcal X}}$. The twisted sectors defined by the twice-iterated intertia stack construction are crucially important in elliptic cohomology (\[8\], cf. also \[16\]), but the significance in physics of even higher iterations is not yet clear. They are nevertheless interesting invariants, and my second main point is that these things [already]{} have a deep literature in mathematics. The study of conjugacy classes of groups is absolutely fundamental to the theory of linear representations, and it seems that conjugacy classes of commuting [tuples]{} of group elements play a systematically analogous role in the theory of actions on manifolds. This section recalls some standard facts from the ‘classical’ theory of orbifolds. The second section is concerned with the semisimplicial constructions mentioned above, and the third is a quick account of the higher character theory of Hopkins, Kuhn, and Ravenel. This paper began as a talk at the ChengDu (Sichuan) satellite ICM conference on stringy orbifolds. I want to thank Alejandro Ádem and Yongbin Ruan on one hand, and Arkady Vaintrob and the referee on the other, for their interest and helpful comments. I owe Ian Leary a special acknowledgement, for contributing an important idea at a crucial point (§2.3). This paper grew out of many conversations with Matthew Ando; in a better world he would be its coauthor. [1.0]{} I will work in a convenient [ad hoc]{} category of orbispaces. For my purposes, an orbispace is a (topological) category ${{\mathcal X}}:= [X/G]$ (with the points of a topological space $X$ as objects), defined by an action of a compact Lie group $G$ on $X$, subject to the restriction that the isotropy group $G_x$ of any point $x \in X$ be [finite]{}. Morphisms of orbispaces are to be [equivalence classes]{}, under invertible natural transformations, of (continuous) functors between categories. This class is rich enough to contain some interesting examples: [Ex 1]{} If ${{\mathcal M}}$ is a reduced $d$-dimensional orbifold, then its principal orthogonal frame ‘bundle’ ${{\mathbb O}}(M)$ is a smooth manifold upon which the orthogonal group ${{\mathbb O}}(d)$ acts with finite isotropy. By a fundamental lemma \[10 §1 (example)\] of Kawasaki, the category (or groupoid) $[{{\mathbb O}}(M)/{{\mathbb O}}(d)]$ is equivalent to the category defined by the original orbifold ${{\mathcal M}}$. [Ex 2]{} If $G$ is a [finite]{} group, then the category $[/G]$ with one object, and the set $G$ of morphisms, is an interesting [un]{}reduced orbifold. [Remarks:]{} Useful topological constructions take us out of the category of smooth objects, so it is convenient to work with a class slightly larger than the usual orbifolds. In general, I will use the [mathcal]{} typeface for an orbispace, and the usual mathematical typeface for its underlying space of objects; thus ${{\mathcal X}}:= [X/G]$ has objects $X$ and underlying quotient space $X/G$. However, there will be exceptions: [1.1]{} If $G$ is a group, and $X \in (G-{\rm spaces})$, then $$I(X) := \{(g,x) \in G \times X \;\|\: gx = x \}$$ is itself a $G$-space, with action defined by $$h \cdot (g,x) = (hgh^{-1},hx) \;.$$ $I$ is thus a functor from the category of $G$-spaces to itself. The isotropy group of $(g,x) \in I(X)$ is $$\{ h \in G \;\|\; h(g,x) = (hgh^{-1},hx) = (g,x) \} \;;$$ being a subset of $G_x$, it is finite if the latter is. It follows that if ${{\mathcal X}}= [X/G]$ is an orbispace, in the sense above, then $${{\mathbb I}}({{\mathcal X}}) := [I(X)/G]$$ is again such an orbispace; following the terminology of algebraic geometers, it is now called the [inertia stack]{} of ${{\mathcal X}}$. \[It is also the fixed-point orbispace \[10\] of the circle group, acting on the free loops in ${{\mathcal X}}$.\] The description above makes it clear that ${{\mathbb I}}$ is an endofunctor of the category of orbispaces. [1.2]{} These constructions define some useful invariants. I will call the Borel cohomology $$H^({{\mathcal X}},{{\mathbb Q}}) := H_G^(X,{{\mathbb Q}}) := H^(EG \times_G X,{{\mathbb Q}})$$ (in this paper all coefficients will be vector spaces over ${{\mathbb Q}}$) the [ordinary]{} cohomology of the orbispace ${{\mathcal X}}$: its Leray spectral sequence has as $E_2$-term, the cohomology $$H^(X/G,{\mathcal H}^(G_x,{{\mathbb Q}}))$$ of the quotient with coefficients in a sheaf whose stalk at $x$ is the group cohomology of $G_x$. Since these groups are by hypothesis finite, this sheaf is concentrated in degree zero, and the spectral sequence degenerates to an isomorphism $$H^({{\mathcal X}},{{\mathbb Q}}) \cong H^(X/G,{{\mathbb Q}}) = H^0(G,H^(X,{{\mathbb Q}}))$$ with the cohomology of the quotient space. This is interesting enough, but it is not very subtle. A more powerful invariant is defined by the equivariant $K$-theory $$K^({{\mathcal X}}) := K_G^(X)$$ of the orbispace. [1.3 Theorem:]{} There is a natural multiplicative transformation $$K^({{\mathcal X}}) = K_G(X)^* \to H^_G(I(X),{{\mathbb Q}}) = H^({{\mathbb I}}({{\mathcal X}}),{{\mathbb Q}})$$ which becomes an isomorphism after tensoring with ${{\mathbb Q}}$ on the left. [Remarks:]{} This is proved in \[2\]; nowadays this (rational) invariant is usually called the Adem-Ruan, or classical, orbifold cohomology. It is to be distinguished from the Chen-Ruan \[4\] orbifold cohomology, which has a different multiplicative structure. \[I will ignore some deep questions about the gradings of these theories, since I have nothing to say about them.\] I should note that the existence of some such generalized Chern character was also known to Baum and Brylinski \[3\]. The cohomology groups on the right have a natural decomposition \[2 §5.1\] as $$\oplus_{g \in \hat G} \; H^_{C(g)}(X^g,{{\mathbb Q}})$$ where $\hat G$ denotes the set of conjugacy classes in $G$, $X^g$ is the set of $g$-fixed points in $X$, and $C(g)$ is the centralizer of $g$ in $G$. \[More precisely: for any choice of $g$ in the appropriate conjugacy class, the cohomology group in question is well-defined under conjugation by elements of $G$.\] The contributions to this sum, indexed by conjugacy classes [other than the identity]{} are now called the [twisted sectors]{} of the cohomology. Higher inertia stacks ===================== [2.1 Definition:]{} If ${{\mathcal X}}= [X/G]$ as above, let $${{\mathbb I}}^n({{\mathcal X}}) = [I^n(X)/G] \;;$$ note that $$I^n(X) = \{ (g_1,\dots,g_n;x) \in G^n \times X \;\|\; g_i \in G_x, \; \forall i,k \; [g_i,g_k] = 1 \} \;.$$ [Proof:]{} See the argument in §1.1, and induct. [For example:]{} If $X = $ is a single point, $${{\mathbb I}}^n[/G] = {{\rm Hom}}({{\mathbb Z}}^n,G)/G$$ is the set of conjugacy classes of commuting $n$-tuples of elements in $G$. When $n=1$, this is just the classical set $\hat G$ of conjugacy classes in $G$. The construction of the inertia stack is essentially local, so more generally $${{\mathbb I}}^n[X/G] = [(\bigcup_{x \in X} I^n[\{x\}/G_x])/G] \;.$$ [2.2]{} Recall now that a simplicial object in a category ${{\mathcal C}}$ can be defined as a functor $C$ from the category of finite ordered sets to ${{\mathcal C}}$. We can think of such a functor as defined by its sets $C[n]$ of $n$-simplices, together with various face and degeneracy maps between them. A simplicial object in the category of spaces (for example, a simplicial set) has a [geometric realization]{} $$\|C\| = \coprod_{n \geq 0} (C[n] \times \Delta^n)/({\rm face \; \& \; degeneracy \; relations}) \;.$$ For example: a category ${{\mathcal C}}$ can be regarded, following Grothendieck and Segal \[17\], as a simplicial set with objects as zero-simplices, morphisms as one-simplices, and chains of $n$ composable morphisms as its $n$-simplices. The face maps are defined by composing maps, and degeneracies are defined by inserting identities. The geometric realization of this simplicial set is sometimes called the classifying space for the category; in particular, $$\|[/G]\| = BG$$ is the classifying space for the (discrete) group $G$, and more generally the geometric realization $$\|[X/G]\| = EG \times_G X$$ of (the category defined by) a transformation group is homotopy equivalent to its Borel construction. The map $$\|[X/G]\| = EG \times_G X \to * \times_G X = X/G$$ which collapses (the free contractible $G$-space) $EG$ to a point is sometimes called the ‘homotopy-to-geometric’ quotient; the arguments of §1.2 above show that for our class of orbispaces, this map induces an isomorphism on rational cohomology. [2.3]{} Conjugation by a group element defines a functor from the category $[/G]$ to itself; thus $G$ acts on $\|[/G]\| = BG$. This endofunctor is naturally equivalent to the identity: natural transformations of functors become (un-base pointed) homotopies under geometric realization, but the resulting action of $G$ on $BG$ need not be trivial. This construction generalizes to an action of $G$ on the category $[X/G]$ and its realization $\|[X/G]\|$: thus the homotopy quotient of a $G$-action is again a (not necessarily trivial!) $G$-space. I will write $X//G$ for the quotient of the homotopy quotient by $G$: the homotopy-to-geometric quotient map can thus be factored, as $$\|[X/G]\| \to \|[X/G]\|/G := X//G \to X/G \;.$$ I owe thanks to Ian Leary for help in understanding this intermediate quotient. He notes that if $G \times G$ acts on $G$ by $(g,h) \cdot k \mapsto gkh^{-1}$ then the usual construction for $EG$ (as a simplicial set with $n$-simplices $G^{n+1}$) becomes a space with $G \times G$-action; as such it is a classifying space for the family consisting of subgroups of $G \times G$ which are conjugate to a subgroup of the diagonal. It follows that $$EG/(G \times G) = BG/G = //G \;,$$ and that $\pi_1(//G) \cong H_1(G,{{\mathbb Z}})$ is the abelianization of $G$; the quotient $\|[/G]\| \to //G$ is a homotopy equivalence if and only if $G$ is abelian. [2.4]{} The simplicial set $n \mapsto {{\mathbb Z}}^n$ defining $[/{{\mathbb Z}}]$ is in fact a simplicial object in the category of abelian groups: group composition is a homomorphism when the group is abelian. It follows that the [co]{}variant functor $$n \mapsto (\check{{{\mathbb Z}}})^n := {{\rm Hom}}({{\mathbb Z}}^n,{{\mathbb Z}})$$ is, in a natural sense, a [co]{}simplicial abelian group. [Definition:]{} The functor $$n \mapsto {{\rm Hom}}((\check{{{\mathbb Z}}})^n,G)/G$$ defines the simplicial set ${{\mathbb I}}^\bullet[/G]$ of conjugacy classes of commuting tuples of elements $G$, cf. \[11 §4\]; more generally, $${{\mathbb I}}^\bullet[X/G] := [(\bigcup_{x \in X} {{\mathbb I}}^\bullet[\{x\}/G_x])/G]$$ is the [simplicial inertia stack]{} of ${{\mathcal X}}$. We can use this construction to elaborate Adem and Ruan’s construction for orbifold cohomology: $$n \mapsto H^({{\mathbb I}}^n({{\mathcal X}}),{{\mathbb Q}})$$ is a cosimplicial object in the category of graded-commutative algebras, which keeps simultaneous track of the higher inertia stacks of ${{\mathcal X}}$. [2.5 Theorem:]{} There is a natural transformation $$\|{{\mathbb I}}^\bullet[X/G]\| \to X//G$$ which is an equivalence if $G$ is abelian. The proof is by construction; it is easiest to begin in the special case when $X$ is a point. Then ${{\mathbb I}}^\bullet[/G]$ is a simplicial set with one zero-simplex; a one-simplex is a conjugacy class, a two-simplex is a conjugacy class of commuting pairs, etc. If $\langle g_1,\dots,g_n \rangle$ is an $n$-simplex, then its faces are the maps $$\langle g_1,\dots,g_n \rangle \mapsto \langle g_1,\dots,g_{i-1}g_i,\dots, g_n \rangle$$ and its degeneracies are the maps which insert identity elements. These are exactly the maps defining the classifying space of $G$; but we are working now not with group elements, but conjugacy classes. The promised map is then the quotient of the obvious equivariant inclusion $${{\rm Hom}}((\check{{{\mathbb Z}}})^\bullet,G) \to BG$$ by $G$. Because ${{\mathbb I}}^\bullet$ is a local construction, this definition now extends directly to $[X/G]$; alternately, we can display the simplicial object ${{\mathbb I}}^\bullet[X/G]$ (with most of its maps supressed) as $$\dots \to \coprod [(X^g \cap X^h)/C(g,h)] \to \coprod [X^g/C(g)] \to [X/G] \;,$$ where the $n$th coproduct is indexed by conjugacy classes of commuting $n$-tuples, and $C(g_1,\dots,g_n)$ is the centralizer of the commuting tuple. [Remarks:]{} It is tempting to think of this construction as a kind of blowup or resolution of the Borel construction; it seems analogous in some ways to Segal’s \[17\] reconstruction of a manifold, up to homotopy, from the category defined by the sets of an atlas with inclusions as morphisms. In our case, the charts are reminiscent of the complete sets of commuting observable of classical quantum mechanics. Kuhn \[11 §7\] remarks that $\|{{\mathbb I}}^\bullet[/G]\|$ is in fact a $\Gamma$-space \[though not, in general, a special $\Gamma$-space\] in the sense of Segal. I am reluctant to admit that I don’t know how a single example works out. Symmetric groups and finite subgroups of ${\rm Sl}_2({{\mathbb C}})$ are of course very interesting candidates. This construction may also be related to the theory of motivic integration: if $X$ is an algebraic variety, say over the complexes, the $n$-simplices of ${{\mathbb I}}^\bullet$ are roughly deformations of the scheme over fields of transcendence degree $n$. To make this precise would require a better understanding of the degree-shifting numbers \[4; 12 §8; 14 §2\], which do not appear in the formalism above. When $n=1$, these are locally constant ${{\mathbb Q}}$-valued functions $w$ on the fixed-point set $X^g$, which are slightly more sophisticated than the function which assigns to $g$, the number $$\log \; \det \; (g\|\nu) \;,$$ where $g\|\nu$ represents the action of $g$ on the normal bundle of $X^g$ in $X$. \[I’m assuming here that the orbifolds in question have complex structures on their tangent spaces.\] In general, the normal bundle to the fixed point set of a commuting $n$-tuple $\langle g_1,\dots,g_n \rangle$ has a flag decomposition as a sum $$X^{g_1} \cap \dots \cap X^{g_{i-1}} \subset X^{g_1} \cap \dots \cap X^{g_i}$$ of normal bundles, and it seems reasonable to expect that the degree-shifting number of this $n$-tuple will be the sum of the degree-shifting numbers of these subbundles. HKR characters ============== [3.0]{} A homomorphism from a free abelian group to a finite group $G$ factors through some finite abelian quotient group, so $${{\rm Hom}}({{\mathbb Z}}^n,G)/G = {{\rm Hom}}(\hat{{{\mathbb Z}}}^n,G)/G = {\prod}^_p {{\rm Hom}}({{\mathbb Z}}_p^n,G)/G$$ decomposes as the restricted product (with only finitely many nontrivial entries) of $p$-local contributions, indexed by primes $p$. This uses the fact that $$\hat {{\mathbb Z}}= {\prod}_p {{\mathbb Z}}_p \;,$$ where ${{\mathbb Z}}_p = \lim {{\mathbb Z}}/p^n {{\mathbb Z}}$ the $p$-adic integers. Since products of simplicial sets (and spaces) are defined coordinate-wise, ${{\mathbb I}}^\bullet [X/G]$ can be expressed as a restricted infinite fiber product (over $[X/G]$) of $p$-local objects $ {{\mathbb I}}^\bullet_p[X/G]$ built like ${{\mathbb I}}$ but with ${{\mathbb Z}}_p$ replacing ${{\mathbb Z}}$. I will ignore questions about infinite restricted products by assuming that $[X/G]$ is ‘ramified’ at a finite set of primes (dividing $\# G$, say, when the group is finite); the cohomology of the simplicial inertia stack can then be calculated from the local contributions, one prime at a time. [3.1]{} In this context, Hopkins, Kuhn, and Ravenel \[9\] provide, for each $n \geq 1$, an interpretation of $H^({{\mathbb I}}^n_p[X/G],{{\mathbb Q}})$ which reduces when $n = 1$ to the theorem of Adem and Ruan in §1.3 above. To state these results, however, requires a short digression about cobordism. Very briefly, then: cobordism is to homology as smooth manifolds are to simplices. In this theory, a $d$-dimensional chain is not some sum of nasty singular simplices, but a map, say $f: M \to X$, of a nice smooth $d$-manifold $M$ to the space $X$ of interest. Instead of boundaries of simplices, we take boundaries of manifolds; thus $\partial f : \partial M \to X$ is the boundary of $f$, which is said to be closed if $\partial M = 0$. Similarly, $f = \partial F$ if $\exists F: W \to M$ such that $\partial W = M$ and $F\|_{\partial W} = f$. The analog of the homology of $X$ is the quotient of the abelian semigroup of closed objects (cycles) by the subsemigroup of boundaries; this is well-defined, since of course $\partial \circ \partial = 0$. It is more usual to say that these groups are defined by classes of maps of smooth manifolds to $X$ under the equivalence relation defined by cobordism: which is to say that two maps of closed manifolds to $X$ are related if they are the boundary values of maps defined on a smooth manifold of one higher dimension. These groups are obviously homotopy-invariant (use the cobordism defined by a cylinder) and covariant: a map $\phi : X \to Y$ pushes the class $[f]$ to the class $[\phi \circ f]$. Atiyah’s convention is to call this (graded-abelian-group-valued, homological) functor the [bordism]{} of $X$; there is a corresponding [co]{}homological theory (contravariant under pullback or fiber product, using Thom’s theory of transversality), now usually called the [co]{}bordism of $X$. One advantage of the latter theory is a nice multiplicative structure, defined by the obvious Cartesian product, without need for any Eilenberg-Zilber foolishness. Cobordism theory has very natural connections with the theory of group actions on manifolds: the Borel construction $$EG \times_G M \to EG \times_G = BG$$ associated to a $G$-manifold $M$ is a kind of relative manifold, which defines a $(-d)$-dimensional class in the cobordism of $BG$. This is the beginnings of a rich subject; a more sophisticated approach can be found in \[7\]. HKR theory is a natural generalization of the classical theory of characters of representations of groups on vector spaces to a theory of characters for actions on manifolds. The advantages of cobordism (geometric naturality, etc.) are recognized in the Russian literature, where it is usually called ‘intrinsic homology’. Its [dis]{}advantages include the fact that there are many cobordism theories, depending on one’s favorite choice of manifold: oriented, spin, symplectic, framed …each with its own special features. A more substantial issue is that the ground ring of such a theory (ie, the value of the cohomology theory on a point) tends to be quite large. It is arguably the cobordism theory of stably almost complex manifolds (with a complex structure on the sum of the tangent bundle with some trivial bundle) which is technically most accessible; that theory, called complex cobordism, has a polynomial ground ring $$MU^* := MU^({\rm pt}) \cong MU^ \cong {{\mathbb Z}}[x_i \;\|\; i \geq 1]$$ with one generators of each even degree. \[Frank Adams’s convention is to write $ML^$ for the cobordism theory of manifolds with structure group reduced to the Lie group $L$, eg $U$ for weakly almost complex manifolds\]. Over the rationals, $$MU^({\rm pt}) \otimes {{\mathbb Q}}= {{\mathbb Q}}[{{\mathbb C}}P_n \;\|\;n \geq 1]$$ is the polynomial ring generated by the complex projective spaces; but these classes do not generate over the integers. More generally, an old argument of Dold shows that there is a natural multiplicative transformation $$MU^(X) \to H^(X, MU^* \otimes {{\mathbb Q}})$$ which factors through an isomorphism of the rationalization of the left-hand side. Over the rationals, then, there is in some sense little difference between cobordism and ordinary cohomology. \[In some sense the chromatic filtration and the Atiyah-Swan filtration of ordinary cohomology by size of supporting elementary abelian subgroups are aspects of some common underlying phenomenon; see \[5,6\] for recent work on the latter topic.\] The advantage of cobordism lies in its naturality: its cycles are geometric objects, which carry characteristic class data (for example, of the sort familiar to physicists in the theory of ‘gravitational descendents’). [3.2]{} However, These cobordism theories are often too big to be technically convenient - for example, their ground rings are not Noetherian - so topologists have developed an arsenal of techniques to make them more useful. One useful ruse is to work $p$-locally, at some fixed prime. It turns out that to understand $MU$ in general, it suffices to understand a hierarchy of cohomology theories with ground rings $$\hat E^_n = {{\mathbb Z}}_p [v_1,\dots,v_{n-1}]((v_n^{-1}))$$ indexed by integers $n \geq 1$, defined as truncations (in a suitable sense) of the $p$-completion of $MU$; here $v_k$ can be taken to be the cobordism class of a degree $p$ hypersurface in ${{\mathbb C}}P(p^k)$, and $A((x))$ is the formal Laurent series extension of a ring $A$ which allows only finitely many negative powers of $x$. When $n=1$, this theory is a version of $p$-adically completed complex $K$-theory. The study of these theories tends to involve some quite subtle number theory, and one of the main technical advances in \[9\] is the construction of a certain faithfully flat ring extension $\hat E_n \subset \hat D_n$, which is most naturally interpreted as a kind of generalized Galois extension, with Galois group ${{\rm Gl}}_n({{\mathbb Z}}_p)$. [Theorem:]{} There is a natural multiplicative transformation $$\hat E^_n(\|[X/G]\|) \to H^({{\mathbb I}}^n_p[X/G],\hat D_n \otimes {{\mathbb Q}})^{{{\rm Gl}}_n(\hat {{\mathbb Z}}_p)-{\rm inv}}$$ which factors through an isomorphism with the rationalization of the group on the left. The term on the right is the subring of invariants under the action of the Galois group ${{\rm Gl}}_n({{\mathbb Z}}_p)$, but that action requires some clarification. The point is that this group acts on the coefficient ring $\hat D_n$, but it also acts on ${{\mathbb I}}^n_p$, through its construction in terms of conjugacy classes of homomorphisms from ${{\mathbb Z}}_p^n$ to $G$. The relevant action on the right is the (diagonal) product of these two natural actions. These groups can be expressed in terms of fixedpoint sets by formulae generalizing \[2 §5.1\]. [3.3]{} This indeed restricts when $n=1$ to the theorem of §1.3, plus (a $p$-adic version of) a theorem of Artin: there is a natural multiplicative transformation $$R(G) \to {{\rm Fns}}(\hat G,{{\mathbb Q}}_{cyc})^{{\rm Gal}({{\mathbb Q}}_{cyc}/{{\mathbb Q}})-{\rm inv}}$$ which factors through an isomorphism with the rationalization of the left-hand side; where ${{\mathbb Q}}_{cyc}$ is the cyclotomic closure of the rationals (defined by adjoining all roots of unity). This natural transformation is nothing but the map which assigns to a representation, its character; this version of the theorem encompasses the fact, also due to Artin, that the values of such characters lie in ${{\mathbb Q}}_{cyc}$. The Galois group $${\rm Gal}({{\mathbb Q}}_{cyc}/{{\mathbb Q}}) \cong \hat {{\mathbb Z}}^\times$$ of this extension is the multiplicative group of profinite integers, whose $p$-local component is the $p$-adic unit group $${{\mathbb Z}}_p^\times = {{\rm Gl}}_1({{\mathbb Z}}_p) \;.$$ The statement above conceals an action of $\hat {{\mathbb Z}}^\times$ on the conjugacy classes, in which $k \in {{\mathbb Z}}$ sends the class of $g$ to the class of $g^k$ (away from the order of $g$). [3.4]{} Here are a few closing remarks: i\) The rings $\hat E_n$ classify (in a suitable sense) one-dimensional formal groups of height $n$ over $p$-adic integer rings, and the rings $\hat D_n$ classify such groups, together with a level structure: this is a preferred basis for the torsion subgroup. In the theory of algebraic stacks \[1\], the cyclotomic Galois action plays a distinguished role; the level structure is just a choice of isomorphism of ${{\mathbb Q}}_p/{{\mathbb Z}}_p$ with the group of $p$-power roots of unity. In the case of a stack defined over ${{\mathbb Q}}$, it is natural to think of the center of ${{\rm Gl}}_n(\hat Z_p)$ as acting through the determinant $$\det : {{\rm Gl}}_n({{\mathbb Z}}_p) \to {{\mathbb Z}}_p^\times$$ on the roots of unity. ii\) The $\hat E_n$’s and the $\hat D_n$’s do [not]{} fit together naturally as a (co)simplicial ring. In particular, the natural action of the symmetric group $\Sigma_n$ on ${{\mathbb I}}^n$ gets lost in the action of ${{\rm Gl}}_n$ on $\hat D_n$. This suggests that there is lots of room in the transition between chromatic levels for all sorts of gerbish orbifold twisting, and other kinds of noncommutative monkey business … [99]{} D. Abramovich, T. Graber, A. Vistoli, Algebraic orbifold quantum products, in [Orbifolds in mathematics and physics]{} 1 - 24, Contemporary Math 310, AMS (2002); available at [math.AG/00112004]{} A. Adem, Y.B. Ruan: Twisted orbifold $K$-theory, Comm. Math. Phys. 237 (2003) 535 - 556; available at [math.AT/0107168]{} P. Baum, J.L. Brylinski: Noncommutative topology: talk at AMS Winter meeting (2000) W. Chen, Y.B. Ruan, Orbifold quantum cohomology, in [ Orbifolds in mathematics and physics]{} 25 - 85, Contemporary Math 310, AMS (2002); available at [math.AG/0005198]{} D.J. Green, I. Leary, The spectrum of the Chern subring, Comment. Math. Helv. 73 (1998) 406 - 426 ———-, J. Hunton, B. Schuster, Chromatic characteristic classes in ordinary group cohomology, Topology 42 (2003) 243 - 263 J. P. C. Greenlees, N. P. Strickland: Varieties and local cohomology for chromatic group cohomology rings, Topology 38 (1999) 1093 - 1139 M. J. Hopkins: Characters and elliptic cohomology, in [Advances in homotopy theory]{} (Cortona, 1988), 87 - 104; London Math. Soc. Lecture Notes 139, Cambridge Univ. Press (1989) ——–, N. J. Kuhn, D. C. Ravenel: Generalized group characters and complex oriented cohomology theories. J. Amer. Math. Soc. 13 (2000) 553 - 594 T. Kawasaki, The signature theorem for $V$-manifolds, Topology 17 (1978) 75 - 83 N. J. Kuhn: Character rings in algebraic topology, in [Advances in homotopy theory]{} (Cortona, 1988), 111 - 126, London Math. Soc. Lecture Notes 139, Cambridge Univ. Press (1989) E. Looijenga: Motivic measures, in [Séminaire Bourbaki]{}, Asterisque 276 (2002) E. Lupercio, B. Uribe: Loop groupoids, gerbes, and twisted sectors on orbifolds, in [Orbifolds in mathematics and physics]{} 163 - 184 - 24, Contemporary Math 310, AMS (2002); available at [math.AT/0110207]{} J. Morava: Some Weil group representations motivated by algebraic topology, in [Elliptic curves and modular forms in algebraic topology]{}, Springer Lecture Notes in Mathematics 1326 (1986) M. Reid: La correspondance de McKay, in [Seminaire Bourbaki]{}, Asterisque 276 (2002) S. Norton: Appendix \[on generalized moonshine\] to G. Mason, Finite groups and modular functions, in Proc. Sympos. Pure Math., 47: [Arcata Conference*]{} 181–210 (1987) G. Segal: Classifying spaces and spectral sequences, Publ. Math. IHES 34 (1968) [^1]: The author was supported in part by the NSF	{ "pile_set_name": "ArXiv" }
--- author: - 'Akio K. Inoue$^1$' title: 'The origin of dust in galaxies revisited: the mechanism determining dust content' --- \[firstpage\] Introduction ============ Cosmic dust grains are negligible in mass in the Universe. Nevertheless, they play significant roles on a lot of astronomical, astrophysical, and astrochemical aspects: extinction (absorption and scattering) matter of radiation, an emission source in infrared wavelengths, a coolant and a heat source in the interstellar medium (ISM) and intergalactic medium (IGM), and a site of formation of molecules. Therefore, dust is one of the most important ingredients in the Universe. Dust is also important for planetary science because grains are material for planets. Dust grains are formed in rapidly cooling gas of stellar outflows (Yamamoto & Hasegawa 1977; Draine & Salpeter 1977). We call such grains ‘stardust’. Sources of the stardust are asymptotic giant branch (AGB) stars, supernovae (SNe), red supergiants, novae, Wolf-Rayet stars, and so on (e.g., Gehrz 1989). The main source of the stardust in the present Milky Way and the Magellanic Clouds is thought to be AGB stars (Gehrz 1989; Draine 2009; Matsuura et al. 2009). SNe may also produce a significant amount of stardust (Kozasa & Hasegawa 1987; Todini & Ferrara 2001; Nozawa et al. 2003, 2007; Schneider et al. 2004; see also Kozasa et al. 2009). Stardust from SNe is particularly important in the early Universe because the time for stars to evolve to the AGB phase is typically about 1 Gyr, but the cosmic time in the early Universe is shorter than it (Morgan & Edmunds 2003; Maiolino et al. 2004; Dwek et al. 2007; but see also Valiante et al. 2009). The ‘first’ stardust may also play an important role to change the mode of star formation from massive star dominated to present-day Sun-like star dominated (Schneider et al. 2003, 2006). However, dust formation by SNe remains in controversy observationally. First detections of a few $M_\odot$ dust freshly formed, which is much larger than expected, in Cassiopeia A (Cas A) and Kepler SN remnants (SNRs) by submilimeter observations with [SCUBA]{} (Dunne et al. 2003; Morgan et al. 2003) were almost contaminated by foreground dust in the ISM on the sight-lines (Krouze et al. 2004; Gomez et al. 2009). Recent infrared observations with [Spiter Space Telescope]{} and [AKARI]{} and submilimeter observations with [Herschel]{} and [BLAST]{} of Cas A and other SNRs are agreeing with theoretical expectations of 0.01–0.1 $M_\odot$ per one SN (Rho et al. 2008; Sakon et al. 2009; Nozawa et al. 2010; Barlow et al. 2010; Sibthorpe et al. 2010). Once stardust grains are injected into the ISM, they are processed there. The grains in hot gas are bombarded by thermally moving protons and sputtered (Onaka & Kamijo 1978; Draine & Salpeter 1979). SN shock waves probably destroy dust grains by grain–grain collisional shattering as well as sputtering (e.g., Dwek & Arendt 1992; Jones et al. 1994, 1996; Nozawa et al. 2006; Silvia et al. 2010). This destruction process is widely accepted and observational evidences of the destruction have been found in several SNRs, especially with [Spitzer Space Telescope]{} (Arendt et al. 1991, 2010; Borkowski et al. 2006; Williams et al. 2006; Dwek et al. 2008; Sankrit et al. 2010; but see Mouri & Taniguchi 2000). Assuming the destruction efficiency expected by theoretical works, we obtain the life-time of dust grains of the order of 100 Myr (McKee 1989; Draine 1990; Jones et al. 1994, 1996). On the other hand, the injection time of stardust is of the order of 1 Gyr (e.g., Gehrz 1989). Thus, another efficient channel of dust formation is required to keep dust content in galaxies. The most plausible mechanism is the accretion growth in the ISM (Draine 1990, 2009); in dense molecular clouds, atoms and molecules of some refractory elements and compounds accrete onto pre-existent grains and may change from the gas phase to the solid phase. Note that unlike the sticking growth of grains well studied in protoplanetary disks, this accretion growth causes an increase in dust mass. This type of growth is favored to explain the observed depletions of some elements in the gas phase of the ISM relative to the solar abundance. The correlation between the depletion degree and the density in the ISM is particularly suggestive for this process (e.g., Savage & Sembach 1996; Jenkins 2009). It is also suggested that an efficient growth is required to explain massive dust mass observed in the early Universe (Michałowski et al. 2010). Since the pioneering work by Dwek & Scalo (1980), many theoretical works on dust content evolution in galaxies have been made so far (Dwek 1998; Lisenfeld & Ferrara 1998; Edmunds & Eales 1998; Hirashita 1999a,b,c; Edmunds 2001; Hirashita et al. 2002; Inoue 2003; Morgan & Edmunds 2003; Dwek et al. 2007; Zhukovska et al. 2008; Calura et al. 2008; Valiante et al. 2009; Pipino et al. 2011; Dwek & Cherchneff 2011; Gall et al. 2011a,b; Mattsson 2011; Asano et al. 2011). These works are based on the evolutionary model of elemental abundance in galaxies called chemical evolution model (Tinsley 1980 for a review) and add some (or all) of the three processes of formation, destruction, and growth of dust to it. One of the main results from the recent works is the importance of the accretion growth. This paper presents a new interpretation of the mechanism for determining dust content in galaxies. Previous works imply that the mechanism is a balance between dust destruction by SNe and accretion growth in the ISM. However, this point has not been discussed clearly, in contrast, this paper analytically shows that it is. For this aim, a simple one-zone model is sufficient. In addition, we present new simple recipes describing stellar remnant mass and yields of elements and dust from state-of-the-art models of stellar nucleosynthesis and formation of stardust. We will start from a review of basic equations presented in §2. In §3, we present new simple recipes of stellar remnant mass and yields. In §4, we calibrate some model parameters to reproduce the observed properties of the solar neighborhood. In §5, we present our analytical interpretation of the mechanism for determining dust content in galaxies. We will present some further discussions in §6. Experts of this field may go straight to §5 which is the new result of this paper. Throughout this paper, we call elements heavier than helium ‘metal’ according to the custom of astronomy. We adopt the metal mass fraction (so-called metallicity) in the Sun of $Z_\odot=0.02$ (Anders & Grevesse 1989) conventionally, although the recent measurements suggest a smaller value of 0.0134 (Asplund et al. 2009). Chemical and dust evolution model of galaxies ============================================= Equations of chemical and dust amount evolution ----------------------------------------------- We deal with a galaxy composed of stars (including their remnants; i.e. white dwarfs, neutron stars, and black-holes) and the ISM. For simplicity, we assume the ISM to be one-zone. The ISM contains metal and dust as internal components. If we denote masses of these components as $M_$ (stars \[and remnants\]), $M_{\rm ISM}$ (ISM), $M_Z$ (metal), and $M_{\rm d}$ (dust), the equations describing their time evolutions are (e.g., Dwek 1998) $$\frac{dM_}{dt} = S(t) - R(t)\,,$$ $$\frac{dM_{\rm ISM}}{dt} = -S(t) + R(t) + I(t) - O(t)\,,$$ $$\frac{dM_Z}{dt} = -Z(t)S(t) + Y_Z(t) + I_Z(t) - O_Z(t)\,,$$ $$\frac{dM_{\rm d}}{dt} = -Z_{\rm d}(t)S(t) + Y_{\rm d}(t) - D_{\rm SN}(t) + G_{\rm ac}(t) + I_{\rm d}(t) - O_{\rm d}(t)\,,$$ where $S$ is the star formation rate, $R$ is the mass return rate from dying stars, $Y_Z$ and $Y_{\rm d}$ are the metal and dust supplying rates ‘yields’ by dying stars, respectively. $Z\equiv M_Z/M_{\rm ISM}$ is the metal mass fraction in the ISM called ‘metallicity’, and $Z_{\rm d}\equiv M_{\rm d}/M_{\rm ISM}$ is the dust mass fraction in the ISM which we call the dust-to-gas mass ratio. Note that $M_{\rm ISM}>M_Z \geq M_{\rm d}$. $I$, $I_Z$, and $I_{\rm d}$ are the ISM, metal, and dust infall rates from the IGM, respectively. $O$, $O_Z$, and $O_{\rm d}$ are the ISM, metal, and dust outflow rates to the IGM, respectively. In this paper, we do not consider any outflows ($O=O_Z=O_{\rm d}=0$), but consider only an ISM infall $I$ (no metal and dust in infalling gas: $I_Z=I_{\rm d}=0$), which is required to reproduce the metallicity distribution of stars nearby the Sun.[^1] The reason why we omit any outflows is that we do not know the transport mechanism of metal and dust from galaxies to the IGM (e.g., Bianchi & Ferrara 2005). However, this omission may be inconsistent with detections of metal and dust in the IGM (e.g., Songaila & Cowie 1996; M[é]{}nard et al. 2010).[^2] In the dust mass equation (eq. \[4\]), there are two additional terms; $D_{\rm SN}$ is the dust destruction rate by SNe and $G_{\rm ac}$ is the dust growth rate in the ISM by metal accretion. These two terms are discussed in §2.5 and §2.6 in detail. Star formation and infall rates ------------------------------- We adopt a simple recipe for star formation introduced by Schmidt (1959): $S\propto {M_{\rm ISM}}^p$ (Schmidt law). The index $p$ called Schmidt index is observationally indicated to be $p=1$–2 (e.g., Kennicutt 1998; Elmegreen 2011) and some theoretical interpretations for the value are presented (e.g., Dopita & Ryder 1994). However, the value and its origin of the index is still an open problem (Elmegreen 2011 and references therein). Fortunately, the choice of the index is not important in fact because in §4 we calibrate other model parameters so as to reproduce the observed star formation history $S(t)$ at the solar neighborhood which is essential. We here assume $p=1$ in order to solve the equations analytically in §5. In this case, we need a time-scale to give the star formation rate: star formation time-scale, $\tau_{\rm SF}$ (see Table 1 in §4 for the values). Thus, the star formation rate is given by $$S(t)=\frac{M_{\rm ISM}(t)}{\tau_{\rm SF}}\,.$$ The infall from the IGM mimics the structure formation in the Universe based on the hierarchical scenario with cold dark matter (e.g., Peacock 1999); small galaxies are first formed at density peaks of the dark matter distribution in the Universe and they grow up larger and larger as they merge each other and also obtain mass by accretion process. Here we simply assume a smooth exponential infall rate although the mass assembly of a galaxy is intrinsically episodic due to the merging process. This simplification is a kind of ensemble average of many galaxies and appropriate to examine a mean property of the galaxies. The infall rate which we adopt is $$I(t)=\frac{M_{\rm total}}{\tau_{\rm in}}\exp(-t/\tau_{\rm in})\,,$$ where $\tau_{\rm in}$ is the infall time-scale and $M_{\rm total}$ is the total mass which a galaxy obtains within the infinite time (see Table 1 in §4 for the values). Note that $M_{\rm total}$ just gives the normalization of mass of a galaxy. Stellar mass spectrum and returned mass rate -------------------------------------------- Salpeter (1955) first investigated the mass spectrum of stars in the solar neighborhood, corrected it for the modulation by stellar evolution and death, and obtained the mass spectrum of stars when they are born, called initial mass function (IMF) of stars. The Salpeter’s IMF is a power-law: $dN/dm=\phi(m)\propto m^{-q}$ with $q=2.35$. A lot of following researches confirmed that the slope was quite universal, especially for massive stars, although there was a cut-off mass for low mass stars (e.g., Kroupa 2002, Chabrier 2003 for reviews). We adopt here a simple functional from proposed by Larson (1998) which is essentially equivalent to the IMFs by Kroupa (2002) and Chabrier (2003) as $$\phi(m)\propto m^{-q}\exp(-m_{\rm c}/m)\,,$$ with a cut-off mass $m_{\rm c}$ and the range from $m_{\rm low}$ to $m_{\rm up}$. As a standard case, we adopt $p=2.35$, $m_{\rm c}=0.2$ $M_\odot$, $m_{\rm low}=0.1$ $M_\odot$, and $m_{\rm up}=100$ $M_\odot$. The cut-off mass well matches with the observed data compiled by Kroupa (2002). We normalize the IMF as $\int_{m_{\rm low}}^{m_{\rm up}}m\phi(m)dm=1$. The mass returned rate from dying stars, $R$, is given by $$R(t) = \int_{m_{\rm lf}(t)}^{m_{\rm up}} \{m-w(m,Z[t'])\}\phi(m) S(t') dm\,,$$ where $$t'=t-\tau_{\rm lf}(m)$$ is the time at which stars with mass $m$ dying at time $t$ are born, $\tau_{\rm lf}(m)$ is the stellar life-time, $w(m,Z)$ is the remnant mass of stars with mass $m$ and metallicity $Z$, and $m_{\rm lf}(t)$ is the minimum mass of stars dying at time $t$. This is the inverse function of $t=\tau_{\rm lf}(m)$. If time $t$ is less than the life-time of the star with $m_{\rm up}$, the returned rate $R=0$. We have assumed that the metallicity of a star is the same as the ISM metallicity at the time when the star is born. The stellar life-time $\tau_{\rm lf}(m)$ is calculated by the formula of Raiteri et al. (1996) which is a fitting function of Padova stellar evolutionary tracks (Bertelli et al. 1994). This formula is a function of stellar mass $m$ and metallicity $Z$. However, the $Z$-dependence is weak. Thus, we neglect it (we always set $Z=Z_\odot$ in the formula). Stellar yields of ‘metal’ and dust ---------------------------------- When stars die, they eject substantial mass of metal and dust into the ISM. The term driving the time evolution of metal mass given by equation (3) is the metal supplying rate, $Y_Z$, called metal yield. Using the IMF, $\phi(m)$, and the star formation rate, $S(t)$, we can express the metal yield as $$Y_Z(t) = \int_{m_{\rm lf}(t)}^{m_{\rm up}} m_Z(m,Z[t'])\phi(m) S(t') dm\,,$$ where $m_Z$ is the metal mass ejected from a star with mass $m$ and metallicity $Z$, and $t'$ is given by equation (9). The dust supplying rate, $Y_{\rm d}$, called dust yield can be expressed likewise: $$Y_{\rm d}(t) = \int_{m_{\rm lf}(t)}^{m_{\rm up}} m_{\rm d}(m,Z[t'])\phi(m) S(t') dm\,,$$ where $m_{\rm d}$ is the dust mass ejected from a star with mass $m$ and metallicity $Z$, and $t'$ is given by equation (9). Dust destruction by supernova blast waves ----------------------------------------- Dust grains are destroyed by SN shock waves due to shattering and sputtering (e.g., Dwek & Arendt 1992). This dust destruction are observed in some SNRs as described in §1. In this paper, we adopt the dust destruction rate by SNe proposed by Dwek & Scalo (1980) and McKee (1989): $$D_{\rm SN}(t) = \frac{M_{\rm d}(t)}{\tau_{\rm SN}(t)}\,,$$ where the destruction time-scale $\tau_{\rm SN}$ is defined as the time-scale which all the ISM is swept by ‘dust destructive’ shock waves: $$\tau_{\rm SN}(t) = \frac{M_{\rm ISM}(t)}{\epsilon m_{\rm SN} R_{\rm SN}(t)}\,,$$ where $R_{\rm SN}$ is the SN occurrence rate, $m_{\rm SN}$ is the mass swept by a single SN, and $\epsilon$ is the efficiency of the dust destruction. The SN occurrence rate is given by $$R_{\rm SN}(t) = \int_{8M_\odot}^{40M_\odot} \phi(m) S(t') dm\,,$$ where we have assumed the mass range for SNe to be 8–40 $M_\odot$ (Heager et al. 2003) and $t'$ is in equation (9). If $t<\tau_{\rm lf}(40\,M_\odot)$, $R_{\rm SN}=0$. Note that we consider only Type II SNe and neglect Type Ia SNe. The reason is discussed in §3. The effective mass swept by dust destructive shock wave, $\epsilon m_{\rm SN}$ is the important parameter. It is estimated to be $\sim1000$ $M_\odot$, namely $\epsilon\sim0.1$ and $m_{\rm SN}\sim10^4$ $M_\odot$ (McKee 1989, Nozawa et al. 2006). Recent models for starburst galaxies in the early Universe often assume an effective mass of $\epsilon m_{\rm SN}\sim100$ $M_\odot$ which is a factor of 10 smaller than our fiducial value (Dwek et al. 2007; Pipino et al. 2011; Gall et al. 2011a). Their argument is that starburst activity produces multiple SNe which make the ISM highly inhomogeneous and the dust destruction efficiency decreases in such medium. However, the solar neighborhood is not the case, and thus, we keep $\epsilon m_{\rm SN}\sim1000$ $M_\odot$. Dust growth by ‘metal’ accretion in the ISM ------------------------------------------- In the ISM, atoms of some refractory elements (or refractory molecules) may accrete onto a dust grain and may become a part of the grain. We call this process the accretion growth of dust in the ISM (Draine 1990). Note that this process does not need nucleation, and thus, can occur even in the ISM. A simple estimate of the growth rate is (e.g., Hirashita 2000) $$G_{\rm ac}(t) = X_{\rm cold} N_{\rm d}(t) \pi a^2 s_Z v_Z \rho_Z^{\rm gas}(t) \,,$$ where $X_{\rm cold}N_{\rm d}$ is the number of dust grains in cold dense clouds, $a$ is the grain radius, $s_Z$ is the sticking probability of accreting metals (atoms or molecules), $v_Z$ is the thermal velocity of the accreting metals and $\rho_Z^{\rm gas}$ is the mass density of the accreting metals in the gas-phase. Note that all the quantities except for $X_{\rm cold}N_{\rm d}$ in equation (15) are typical (or effective) values averaged over various grain radii, elements, and ISM phases. The gas-phase metal density is reduced to $\rho_Z^{\rm gas}=\rho_{\rm ISM}^{\rm eff}Z(1-\delta)$, where $\rho_{\rm ISM}^{\rm eff}$ is an effective ISM mass density. We define it as a mass-weighted average density of various ISM phases, and then, it is determined by the density of dense molecular clouds where the dust growth occurs. Note that $\delta=M_{\rm d}/M_Z$, the dust-to-metal mass ratio (the dust depletion factor is $1-\delta$). For spherical grains, $N_{\rm d}=3M_{\rm d}/(4\pi a^3 \sigma$), where $\sigma$ is the typical material density of grains. Equation (15) can be reduced to $$G_{\rm ac}(t) = \frac{M_{\rm d}(t)}{\tau_{\rm ac}(t)}\,.$$ The accretion growth time-scale $\tau_{\rm ac}$ is $$\tau_{\rm ac}(t) = \frac{\tau_{\rm ac,0}}{Z(t) (1-\delta[t])}\,,$$ where the normalization $\tau_{\rm ac,0}$ is the parameter determining the process: $$\tau_{\rm ac,0}=\frac{4a \sigma} {3X_{\rm cold}s_Z v_Z \rho_{\rm ISM}^{\rm eff}}\,.$$ This time-scale is very uncertain, but we will obtain $\tau_{\rm ac,0}=3\times10^6$ yr as the fiducial value in §4.2 in order to reproduce the dust-to-metal ratio at the solar neighborhood with the SN destruction efficiency of $\epsilon m_{\rm SN}=1000$ $M_\odot$. This value can be obtained with a set of parameters of $a=0.1$ $\mu$m (typical size in the ISM of the Milky Way), $\sigma=3$ g cm$^{-3}$ (compact silicates), $s_Z=1$, $v_Z=0.2$ km s$^{-1}$ ($^{56}$Fe as an accreting metal atom and thermal temperature of 100 K), $\rho_{\rm ISM}^{\rm eff}=1\times10^{-22}$ g cm$^{-3}$, and $X_{\rm cold}=0.2$. This set is just an example but ensures that the time-scale is not outrageous. There is a discussion about the lifetime of dense clouds (or recycling time-scale of dense gas) should be longer than the accretion growth time-scale for an efficient dust growth (Zhukovska et al. 2008; Dwek & Cherchneff 2011). According to these authors, the lifetime is long enough to realize an efficient dust growth in the Milky Way and even in starburst in the early Universe. Another issue is the effect of grain size distribution which is discussed in Hirashita (2011). Stellar remnant and ‘metal’ and dust yields =========================================== In this section, we present new simple formulas to describe the stellar remnant mass and yields of metal and dust which are useful to input into chemical evolution codes. We represent all elements heavier than helium as just a ‘metal’ in the formulas for simplicity, while yields of various elements are presented in literature. We consider three types of stellar death: white dwarfs through the AGB phase, core-collapse Type II SNe, and direct collapse toward black-hole called ‘collapser’ (Heger et al. 2003). In this paper, we assume the mass range for the SNe to be 8–40 $M_\odot$ (Heger et al. 2003). The stars with mass below or above this mass range become AGB stars or ‘collapsers’, respectively. We neglect Type Ia SNe for simplicity. This population of SNe is the major source of iron element (Iwamoto et al. 1999) and may be the source of iron dust (Calura et al. 2008). However, in respect of the total stardust mass budget, the contribution relative to SNe II is always less than 1–10% (Zhukovska et al. 2008; Pipino et al. 2011). Since we are dealing with metal and dust as each a single component, we safely neglect the contribution of SNe Ia. The remnant mass, $w(m,Z)$, is taken from model calculations of AGB stars (Karakas 2010) and SNe (Nomoto et al. 2006). Figure 1 shows the remnant mass fraction relative to the initial stellar mass, $w/m$. This depends on metallicity $Z$ because $Z$ in the stellar atmosphere determines radiation pressure through opacity and the strength of the stellar wind in the course of the stellar evolution, and affects the remnant mass. However, as shown in Figure 1, the dependence is weak, so we neglect it. We obtain the following fitting formula: $$\frac{w(m)}{m}=\cases{ 1 & ($m > 40\,M_\odot$) \cr 0.13 \left(\frac{m}{8\,M_\odot}\right)^{-0.5} & ($8\,M_\odot \leq m \leq 40\,M_\odot$) \cr 0.13 \left(\frac{m}{8\,M_\odot}\right)^{-0.7} & ($m < 8\,M_\odot$) }\,,$$ which is shown by the solid line in Figure 1. This fitting formula agrees with the values in Table 1 of Morgan & Edmunds (2003) within a $<20\%$ difference, except for $m=9$ $M_\odot$ case in which our estimate is a factor of 2 lower than that of Morgan & Edmunds (2003). ![Remnant mass fraction, $w/m$, as a function of the initial stellar mass, $m$. The data of AGB stars (1–8 $M_\odot$) are taken from Karakas (2010) and those of SNe (8–40 $M_\odot$) are taken from Nomoto et al. (2006). The different symbols indicate different metallicity $Z$: $Z=0$ (plus), $Z=0.0001$ (cross), $Z=0.001$ (diamond), $Z=0.004$ (triangle), $Z=0.008$ (circle), and $Z=0.02$ (square). The solid line is a fitting function given by equation (19).](fig1.eps){width="7.0cm"} ![Ejected metal mass fraction, $m_Z/m$, as a function of the initial stellar mass, $m$. The data of AGB stars (1–8 $M_\odot$) are taken from Karakas (2010) and those of SNe (8–40 $M_\odot$) are taken from Nomoto et al. (2006). The meaning of the symbols are the same as in Fig. 1. The solid line is a fitting function given by equation (20). The dotted lines are the cases a factor of two higher or lower than the solid line.](fig2.eps){width="7.0cm"} For the metal yield, $m_Z$, we adopt the data taken from model calculations of AGBs (Karakas 2010) and SNe (Nomoto et al. 2006). Figure 2 shows $m_Z$ relative to the initial stellar mass $m$ as a function of $m$. While the expected $m_Z$ depends on mass $m$ and metallicity $Z$ by a complex way, we approximate the data with a simple power-law of only $m$ as $$\frac{m_Z(m)}{m}=\cases{ 0 & ($m > 40\,M_\odot$) \cr f_Z \left(\frac{m}{8\,M_\odot}\right)^{2} & ($8\,M_\odot \leq m \leq 40\,M_\odot$) \cr f_Z \left(\frac{m}{8\,M_\odot}\right)^{0.7} & ($m < 8\,M_\odot$) }\,.$$ When the normalization $f_Z=0.02$, equation (20) is the solid line in Figure 2. As shown in the figure, the uncertainty of equation (20) is a factor of $\sim2$. This fitting agrees with the values in Table 1 of Morgan & Edmunds (2003) within a factor of 2 difference in the SN regime. However, in the AGB regime, the difference is as large as the model results by Karakas (2010). The effect of this large uncertainty of the yield is discussed in §6.1. ![Ejected dust mass fraction, $m_{\rm d}/m$, as a function of the initial stellar mass, $m$. The data of AGB stars (1–8 $M_\odot$) are taken from Zhukovska et al. (2008) and those of SNe (8–40 $M_\odot$) are taken from Nozawa et al. (2007). For the AGB data, the different symbols indicate different metallicity $Z$: $Z=0.001$ (diamond), $Z=0.004$ (triangle), $Z=0.008$ (circle), and $Z=0.02$ (square). For the SNe data, the different symbols indicate different ambient hydrogen density $n_{\rm H}=0.1$ cm$^{-3}$ (circle), $n_{\rm H}=1$ cm$^{-3}$ (triangle), and $n_{\rm H}=10$ cm$^{-3}$ (square). The open and filled symbols correspond to ‘mixed’ or ‘unmixed’ cases of Nozawa et al. (2007), respectively. The dotted, solid, and dashed lines are the cases of $\xi=1$, 0.1, and 0.001, respectively, in equation (21).](fig3.eps){width="7.0cm"} The dust yield, $m_{\rm d}$, calculated by Zhukovska et al. (2008) and Ferrarotti & Gail (2006) for AGBs and Nozawa et al. (2007) for SNe are shown in Figure 3. These yields are theoretical ones and do not seem to be compared with observations very much yet. As found in Figure 3, $m_{\rm d}$ depends on mass $m$ and metallicity $Z$ by a complex way as the metal yield $m_Z$ does. Moreover, the dust production by SNe is further complex because the reverse shock moving in the ejecta of a SN may destroy the dust produced in the ejecta (Bianchi & Schneider 2007, Nozawa et al. 2007, Nath et al. 2008, Silvia et al. 2010). This self-destruction depends on the material strength against the destruction[^3] and the ambient gas density which determines the strength of the reverse shock. According to Nozawa et al. (2007), we plot three cases of the ambient density and ‘mixed’ and ‘unmixed’ dust productions[^4] in Figure 3. We adopt a simple formula for $m_{\rm d}$ as $$m_{\rm d}(m) = \xi m_Z(m)\,,$$ where $\xi$ is a scaling factor and means an efficiency of condensation of metal elements. In Figure 3, we show three cases of $\xi=1$ (all metal condenses into dust: an extreme but unrealistic case), 0.1 (fiducial case), and 0.01 (a lower efficiency case). The reader may be anxious about a large uncertainty of this approximation. However, the dust mass in galaxies does not depend on $m_{\rm d}$ after the accretion growth becomes active. This is because the growth of dust is the dominant process of dust production after the activation as shown later in §6.1. Milky Way analog ================ Let us calibrate parameters in the chemical and dust evolution model of galaxies so as to reproduce the properties at the solar neighborhood in the Milky Way. There are two parameters in the chemical evolution part: the time-scales of star formation, $\tau_{\rm SF}$, and infall, $\tau_{\rm in}$. There are additional two parameters in the dust content evolution: the time-scale of the ISM accretion growth, $\tau_{\rm ac,0}$, and the efficiency of the dust destruction, $\epsilon m_{\rm SN}$. In addition, there are two parameters as uncertainties of metal and dust yields, $f_Z$ and $\xi$. Table 1 is a summary of these parameters and values. Note that we do not apply any statistical method to justify the goodness of the reproduction of the observational constraints throughout this paper because our aim is not to find the best fit solution for the constraints but to demonstrate the dust content evolution in galaxies qualitatively. This is partly due to the weakness of the observational constraints and due to large uncertainties of the dust physics itself. Parameter Fiducial value Considered values -------------------------------------------------------------- ---------------- -------------------------------------------------------------------- ($\tau_{\rm SF}$/Gyr, $\tau_{\rm in}$/Gyr) (3, 15) (1, 50), (2, 20), (3, 15), and (5, 10) ($\tau_{\rm ac,0}$/Myr, $\epsilon m_{\rm SN}$/$10^3M_\odot$) (3, 1) (1.5, 1), (1.5, 2), (3, 0.5), (3, 1), (3, 2), (6, 0.5), and (6, 1) $f_Z$ 0.02 0.01, 0.02, and 0.04 $\xi$ 0.1 0.01, 0.1, and 1 Chemical evolution at the solar neighborhood -------------------------------------------- ![Star formation history (time evolution of star formation rate) normalized by the average rate. The histogram is the observed history at the solar neighborhood reported by Rocha-Pinto et al. (2000a). The filled circles with error-bars are the average of the histogram over 4 Gyr interval and corrected by the average rate only for the age less than 12 Gyr which is the present age of the Milky Way assumed in this paper. The vertical error-bars are the standard error of the mean. The four lines correspond to the time evolutions with four different sets of time-scales of star formation and infall ($\tau_{\rm SF}$/Gyr, $\tau_{\rm in}$/Gyr) as indicated in the panel.](fig4.eps){width="7.0cm"} We here determine the time-scales of star formation and infall in the chemical evolution part. First, we constrain these time-scales by using the star formation history at the solar neighborhood reported by Rocha-Pinto et al. (2000a). Such a method was adopted by Takeuchi & Hirashita (2000). Rocha-Pinto et al. (2000a) derived the star formation history from the age distribution of 552 late-type dwarf stars at the solar neighborhood. The histogram in Figure 4 is their result and shows very stochastic nature of the history. However, our model can treat only a smooth history. Thus, we smoothed the stochastic history by averaging with 4 Gyr interval. The filled circles are the result. The vertical error-bars indicate the standard error of the mean. The average history is re-normalized by the average star formation rate for the stellar age less than 12 Gyr which is the assumed age of the Milky Way in this paper, although this choice of the age is arbitrary. We have tried four cases of $\tau_{\rm SF}$ in this paper: 1, 2, 3, and 5 Gyr which are the observed range of the time-scale (or gas consumption time-scale) for disk galaxies like the Milky Way (e.g., Larson, Tinsley, & Caldwell 1980). For each $\tau_{\rm SF}$, we have found $\tau_{\rm in}$ with which we can reproduce the smoothed history as shown in Figure 4. ![Age-metallicity relation of stars. The diamonds with error-bars are the data of stars at the solar neighborhood reported by Rocha-Pinto et al. (2000b). The solar mark ($\odot$) indicates the position of the Sun on this plot. The four lines correspond to the model relations same as Fig. 4. The present age of the Milky Way is assumed to be 12 Gyr.](fig5.eps){width="7.0cm"} Next, we adopt the observed relation between the stellar age and metallicity, so-called the age-metallicity relation, reported by Rocha-Pinto et al. (2000b) to further constrain $(\tau_{\rm SF},\tau_{\rm in})$. Rocha-Pinto et al. (2000b) derived the relation from the same 552 stars as Rocha-Pinto et al. (2000a). Their result is shown in Figure 5 by diamonds with error-bars. After comparing with our four model lines, we have found that $(\tau_{\rm SF}/{\rm Gyr},\tau_{\rm in}/{\rm Gyr})=(5,10)$ case seems the best match with the observed relation but $(\tau_{\rm SF}/{\rm Gyr},\tau_{\rm in}/{\rm Gyr})=(3,15)$ case is also acceptable. ![Time evolution of the stellar mass relative to the ISM mass. The filled circle with error-bars is an estimate at the solar neighborhood taken from Naab & Ostriker (2006) and references therein. Note that the stellar mass includes mass of remnants. The present age of the Milky Way should be less than the age of the Universe (13.7 Gyr). The four lines correspond to the model evolutions same as Fig. 4.](fig6.eps){width="7.0cm"} Finally, we adopt another constraint: the current stellar mass relative to the ISM mass. Naab & Ostriker (2006) compiled observational constraints for the solar neighborhood. From the compilation, we adopt the ratio of the stellar mass to the ISM mass at the present epoch of $3.1\pm1.4$. Note that the stellar mass includes the remnant mass (i.e. white dwarf, neutron stars, and black-holes). Figure 6 shows the comparison of the ratio with our four star formation histories. We have found that the two sets of $(\tau_{\rm SF}/{\rm Gyr},\tau_{\rm in}/{\rm Gyr})=(2,20)$ and (3,15) are consistent with the data. From these three comparisons, we finally adopt the case of $(\tau_{\rm SF}/{\rm Gyr},\tau_{\rm in}/{\rm Gyr})=(3,15)$ as the fiducial set for the Milky Way (or more precisely, for the solar neighborhood) in this paper. Dust content evolution at the solar neighborhood ------------------------------------------------ ![Time evolution of metallicity (metal mass fraction in the ISM; dashed line) and dust-to-gas mass ratio (other lines) for the fiducial case (Table 1). The dotted line corresponds to the case only with ‘stardust’ production and destruction by star formation (i.e. astration). The dot-dashed line correspond to the case with the dust destruction by SNe but without the dust growth in the ISM. The solid line is the case with all the processes. The filled circle with error-bars is an estimate of the dust-to-gas mass ratio at the solar neighborhood (see text). The present age of the Milky Way should be less than the age of the Universe (13.7 Gyr).](fig7.eps){width="7.0cm"} Here we examine the dust content evolution. First, we show the significant effect of the dust destruction and the ISM growth. Figure 7 shows the time evolution of metallicity and dust-to-gas mass ratio for the fiducial set of $\tau_{\rm SF}$ and $\tau_{\rm in}$ obtained in the previous subsection. The model curves of the dust-to-gas ratio (dotted, dot-dashed, and solid lines) are compared with the filled circle with error-bars which is an observational estimate at the solar neighborhood. This is obtained from metallicity $Z\approx Z_\odot$ (van den Bergh 2000; see also Rocha-Pinto et al. 2000b) and dust-to-metal mass ratio $\delta\approx0.5$ (Kimura et al. 2003; see below) and the uncertainty is the quadrature of uncertainties of 30%[^5] in $Z$ and 20% in $\delta$. If there is neither destruction nor accretion growth of dust, the dust-to-gas ratio evolution is just the metallicity evolution multiplied by the condensation efficiency of stardust, $\xi$, as shown by the dotted line. We have assumed $\xi=0.1$ for the line. Once the SN destruction of dust is turned on with a standard efficiency as $\epsilon m_{\rm SN}=1\times10^3\,M_\odot$ (McKee 1989; Nozawa et al. 2006), it reduces the dust amount by a factor of ten as shown by the dot-dashed line. This confirms that the dust destruction is very efficient and the stardust injection is too small to compensate the destruction (e.g., Draine 1990; Tielens 1998). Then we need the accretion growth in the ISM to reproduce the dust-to-gas ratio $\sim10^{-2}$ in the present Milky Way. If we assume the time-scale of $\tau_{\rm ac,0}=3\times10^6$ yr, the dust-to-gas ratio evolution becomes the solid line and it reaches $\simeq10^{-2}$ which is almost two orders of magnitude larger than the case without the growth after several Gyr. ![Time evolution of time-scales of the dust destruction by SNe (dashed line) and of the dust growth in the ISM (solid line) for the fiducial case (Table 1).](fig8.eps){width="7.0cm"} Figure 8 shows the time evolution of $\tau_{\rm SN}$ in equation (13) and $\tau_{\rm ac}$ in equation (17). The SN destruction time-scale $\tau_{\rm SN}$ is almost constant promptly after the first a few hundreds Myr. On the other hand, the accretion growth time-scale $\tau_{\rm ac}$ decreases gradually in the first a few Gyr. This is because $\tau_{\rm ac}$ has a metallicity dependence as shown in equation (17) and it decreases as the metallicity increases. At the time around 4 Gyr, $\tau_{\rm ac}$ becomes shorter than $\tau_{\rm SN}$, and then, the accretion growth becomes significant and the dust amount increases rapidly. As the accretion growth proceeds, the metal abundance in the gas phase decreases, i.e., the dust-to-metal ratio $\delta$ increases, then, $\tau_{\rm ac}$ becomes almost constant and balances with $\tau_{\rm SN}$. We will discuss this point in §5 more in detail. ![Time evolution of dust-to-metal mass ratio, $\delta$, for various sets of the parameters of dust growth and destruction $(\tau_{\rm ac,0},\epsilon m_{\rm SN})$. The solid line is the fiducial case with ($3\times10^6$ yr, $1\times10^3$ $M_\odot$). The two dotted lines correspond to two different sets of ($1.5\times10^6$ yr, $1\times10^3$ $M_\odot$) for the upper and ($1.5\times10^6$ yr, $2\times10^3$ $M_\odot$) for the lower. The two dot-dashed lines correspond to the sets of ($3\times10^6$ yr, $5\times10^2$ $M_\odot$) for the upper and ($3\times10^6$ yr, $2\times10^3$ $M_\odot$) for the lower. The two dashed lines correspond to the sets of ($6\times10^6$ yr, $5\times10^2$ $M_\odot$) for the upper and ($6\times10^6$ yr, $1\times10^3$ $M_\odot$) for the lower. The filled circle with error-bars is an estimate in the Local Interstellar Cloud surrounding the Sun by Kimura et al. (2003). The present age of the Milky Way should be less than the age of the Universe (13.7 Gyr).](fig9.eps){width="7.0cm"} Figure 9 shows the time evolution of the dust-to-metal ratio, $\delta$. The solid line is the fiducial case which is shown in Figures 7 and 8. This can be compared with the observed ratio in the Local Interstellar Cloud reported by Kimura et al. (2003): $\delta=0.5\pm0.1$. As shown in Figure 9, the fiducial set of $(\tau_{\rm ac,0},\epsilon m_{\rm SN})=(3\,{\rm Myr},1\times10^3\,M_\odot)$ excellently agrees with the observed data. On the other hand, other sets can also reproduce the data. For example, $(\tau_{\rm ac,0},\epsilon m_{\rm SN})=(1.5\,{\rm Myr},2\times10^3\,M_\odot)$ or $(6\,{\rm Myr},5\times10^2\,M_\odot)$. Interestingly, the $\delta$ evolutions become very similar if the product of $\tau_{\rm ac,0}$ and $\epsilon m_{\rm SN}$ is the same. We will also discuss this point in §5. Determining dust-to-metal ratio =============================== In this section, we demonstrate the mechanism for determining the dust-to-metal mass ratio, $\delta$, in galaxies. Starting from equations (3) and (4), we can obtain the time evolutionary equation of $\delta\equiv M_{\rm d}/M_Z$. Here, let us adopt the instantaneous recycling approximation (IRA) in which we neglect the finite stellar life-time and assume that stars with a mass larger than a certain threshold (the turn-off mass $m_t$) die instantly when they are formed. This approximation allows us to manage the equations analytically and is good enough to see phenomena with a time-scale longer than a Gyr (see Tinsley 1980 for more details). In the IRA, we can approximate the metal and dust yields in equations (10) and (11) as $Y_Z\approx{\cal Y}_Z S$ and $Y_{\rm d}=\xi Y_Z\approx\xi{\cal Y}_Z S$, where the effective metal yield $${\cal Y}_Z=\int_{m_t}^{m_{\rm up}} m_Z(m) \phi(m) dm = 0.024 \left(\frac{f_Z}{0.02}\right)\,,$$ where we have assumed $m_t=1$ $M\odot$. This value is not sensitive to $m_t$. We obtain ${\cal Y}_Z=0.021(f_Z/0.02)$ if $m_t=5$ $M_\odot$. Remembering the star formation rate $S=M_{\rm ISM}/\tau_{\rm SF}$ as in equation (5), then, we obtain $$\frac{1}{\delta}\frac{d\delta}{dt} \approx -\frac{{\cal Y}_Z}{\tau_{\rm SF}Z}\left(1-\frac{\xi}{\delta}\right) -\frac{1}{\tau_{\rm SN}}-\frac{1}{\tau_{\rm ac}}\,.$$ In the IRA, the SN destruction time-scale $\tau_{\rm SN}$ in equation (13) can be reduced to $$\tau_{\rm SN} \approx \frac{\tau_{\rm SF}}{\epsilon m_{\rm SN} n_{\rm SN}}\,,$$ where the effective number of SN per unit stellar mass is $$n_{\rm SN} = \int_{8M_\odot}^{40M_\odot} \phi(m) dm = 0.010~{M_\odot}^{-1}\,.$$ Note that $\epsilon m_{\rm SN} n_{\rm SN}$ is a non-dimensional value. The accretion growth time-scale $\tau_{\rm ac}$ is given in equation (17). Then, equation (23) is reduced to $$\frac{1}{\delta}\frac{d\delta}{dt} \approx - \frac{\alpha+\epsilon m_{\rm SN} n_{\rm SN}}{\tau_{\rm SF}} + \frac{Z(1-\delta)}{\tau_{\rm ac,0}}\,,$$ where $$\alpha = \frac{{\cal Y}_Z}{Z}\left(1-\frac{\xi}{\delta}\right)\,.$$ In the IRA, the metallicity $Z\equiv M_Z/M_{\rm ISM}$ can be obtained analytically (for example, see Dwek et al. 2007). Then, we have found that $Z\to{\cal Y}_Z$ for $t\to\infty$ when $\tau_{\rm in}>\tau_{\rm SF}$. The condensation efficiency $\xi$ is uncertain but it is of the order of 0.1 (see Figure 3). When $\delta$ is of the order of 0.1–1 as shown in Figure 9, the ratio $\xi/\delta$ is of the order of 1 or smaller. Therefore, $\alpha$ is also of the order of 1 or smaller. On the other hand, $n_{\rm SN}\sim10^{-2}$ $M_\odot^{-1}$ and $\epsilon m_{\rm SN}\sim10^3$ $M_\odot$, then, we obtain $\epsilon m_{\rm SN} n_{\rm SN} \gg \alpha$. Therefore, equation (26) is further reduced to $$\frac{1}{\delta}\frac{d\delta}{dt} \approx - a + b (1-\delta)\,,$$ where $a=\epsilon m_{\rm SN} n_{\rm SN}/\tau_{\rm SF}$ and $b=Z/\tau_{\rm ac,0}$. If we assume $Z$ to be constant (i.e. $b$ is constant), equation (28) can be solved analytically. The solution is $$\delta \approx \frac{\delta_\infty\delta_0 \exp{(b-a)t}} {(\delta_\infty - \delta_0) + \delta_0 \exp{(b-a)t}}\,,$$ where $\delta_0$ and $\delta_\infty$ are the values for $t=0$ and $t\to\infty$, respectively. The asymptotic value $\delta_\infty$ for $t\to\infty$ is realized only when $b>a$, and is given by $$1-\delta_\infty = \frac{a}{b} = \frac{\tau_{\rm ac,0}\epsilon m_{\rm SN} n_{\rm SN}}{\tau_{\rm SF} Z}\,.$$ This is the equilibrium value for equation (28) and we find $$\begin{aligned} 1-\delta_\infty = 0.5 \left(\frac{\tau_{\rm ac,0}}{3~{\rm Myr}}\right) \left(\frac{\epsilon m_{\rm SN}}{10^3~M_\odot}\right) \left(\frac{n_{\rm SN}}{10^{-2}~M_\odot^{-1}}\right) \cr \times \left(\frac{3~{\rm Gyr}}{\tau_{\rm SF}}\right) \left(\frac{0.02}{Z}\right)\,, \end{aligned}$$ which excellently agrees with the results in Figure 9. We can fully understand the $\delta$ evolution by using equation (28). At the beginning, the accretion term $b\sim0$ because $Z\sim0$. Then, only the destruction term $a$ is effective. As a result, $\delta$ decreases with the time-scale of $1/a=\tau_{\rm SN}$. As $Z$ increases, the accretion term $b$ increases and finally exceeds $a$. Then, $\delta$ increases toward $\delta_\infty$ with the evolution time-scale of $1/(b-a)$. This decreases as $Z$ increases and $b-a$ increases. Therefore, the driving force of the $\delta$ evolution is $Z$. If we call $Z$ at $b=a$ as the critical metallicity, $Z_{\rm c}$, we find $$\begin{aligned} Z_{\rm c}=\frac{\tau_{\rm ac,0}\epsilon m_{\rm SN} n_{\rm SN}} {\tau_{\rm SF}} =0.01~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \cr \times \left(\frac{\tau_{\rm ac,0}}{3~{\rm Myr}}\right) \left(\frac{\epsilon m_{\rm SN}}{10^3~M_\odot}\right) \left(\frac{n_{\rm SN}}{10^{-2}~M_\odot^{-1}}\right) \left(\frac{3~{\rm Gyr}}{\tau_{\rm SF}}\right)\,.\end{aligned}$$ When $Z>Z_{\rm c}$, the accretion growth becomes effective and $\delta$ approaches the final value $\delta_\infty$. A similar critical metallicity has been derived by Asano et al. (2011) with a different way. Equation (30) shows that the final value of $\delta$ is determined by the equilibrium between the SN destruction and the accretion growth in the ISM. The time-scale to reach the equilibrium is $1/(b-a)$. This is relatively short in the fiducial case. For example, it is 0.3 Gyr when $Z=0.02$. This means that the $\delta$ evolution proceeds with keeping the equilibrium between the SN destruction and the accretion growth, or equivalently, $\delta=\delta_\infty$ after $Z$ exceeds $Z_{\rm c}$. This behavior is also found by the comparison of the two time-scales, $\tau_{\rm SN}$ and $\tau_{\rm ac}$, in Figure 8; once $\tau_{\rm ac}$ becomes shorter than $\tau_{\rm SN}$ at about 4 Gyr at which $Z$ exceeds $Z_{\rm c}$, $\tau_{\rm ac}$ turns around and approaches $\tau_{\rm SN}$ again. This is realized by the reduction of the term $(1-\delta)$ in $\tau_{\rm ac}$ (see eq. \[17\]) when $\delta$ increases from $\sim0$ to $\delta_\infty$. Such a kind of self-regulation process determines the dust-to-metal ratio $\delta$. Discussion ========== Effect of uncertainties of yields --------------------------------- ![Time evolution of (a) metallicity, (b) dust-to-metal mass ratio, and (c) time-scales of SN destruction and ISM growth for three different stellar metal yields: $f_Z=0.01$ (dot-dashed), 0.02 (solid), and 0.04 (dotted). The horizontal short-dashed line in the panel (a) shows the critical metallicity of equation (32). The vertical short-dashed lines indicate the timing at which the metallicity exceeds the critical one for the three metal yields. The long-dashed line in the panel (c) is the SN destruction time-scale, but the other three lines are the ISM growth time-scales for the three metal yields.](fig10.eps){width="7.0cm"} Here we examine the effect of uncertainties of the normalization of metal and dust yields. As we saw in Figure 2, our simple recipe for the metal yield may contain a factor of 2 (or more) uncertainty. The parameter $f_Z$ accounts for this uncertainty. In Figure 10, we show the effect of $f_Z$. As found from the panel (a), the metallicity evolution is scaled almost lineally by $f_Z$ as expected and the timing at which $Z$ exceeds $Z_{\rm c}=0.5Z_\odot$ given by equation (32) for the fiducial set of the accretion and destruction efficiencies becomes faster as $f_Z$ is larger. From the panels (b) and (c), we find that for each case of $f_Z$, $\delta$ increases and $\tau_{\rm ac}$ becomes shorter than $\tau_{\rm SN}$ soon after the timing for $Z>Z_{\rm c}$. Therefore, the timing for $\tau_{\rm ac}<\tau_{\rm SN}$, in other words, the timing for the accretion growth activation is well traced by $Z_{\rm c}$ in equation (32) and this is not affected by uncertainty of $f_Z$. On the other hand, the timing for the activation becomes faster for larger $f_Z$. The metallicity dependence on the final value of $\delta$ is explicit as found in equation (31). ![Same as Fig. 7 but for different condensation efficiencies in the stellar ejecta as indicated in the panel.](fig11.eps){width="7.0cm"} Figure 11 shows the effect of the dust yield. As seen in Figure 3, our recipe for the stardust yield has a factor of 10 or larger uncertainty because of a large uncertainty in the adopted model calculations. In Figure 11, we show the cases with a factor of 10 larger or smaller yield than the fiducial one. Other parameters are the same as the fiducial set, so that we have the same evolutions of the metallicity and the time-scales of the SN destruction and accretion growth as shown by the solid lines in Figure 10. Before the growth activation at around 4 Gyr, the dust amounts show a large difference, however, they converge nearly the same amount after the activation. This is because the final value of $\delta$ given in equation (31) does not depend on the dust yield. Therefore, we conclude that the dust content in galaxies is independent of the stardust yield after the grain growth in the ISM becomes active, or equivalently, the metallicity exceeds the critical one. What kind of dust is formed by the ISM growth? ---------------------------------------------- We have shown that the main production channel of dust is the accretion growth in the ISM of the present-day Milky Way. This conclusion had been obtained also in the literature. For example, Zhukovska et al. (2008) argued that the mass fraction of stardusts in total dust is only 0.1–1% based on a more sophisticated chemical evolution model than this paper (see their Fig. 15); more than 99% of dust is originated from the accretion growth in the ISM. It is also well known that some interplanetary dust particles show a highly enhanced abundance of deuterium and $^{15}$N relative to the solar composition, which is a signature of the molecular cloud origin because such isotopic fractionations are expected in low temperature environment (e.g., Messenger 2000). Therefore, dust produced by the ISM accretion exists. Then, we have a very important question; what kinds of dust species are formed by the accretion growth in the ISM? In molecular clouds, many kinds of ices such as H$_2$O, CO, CO$_2$, CH$_3$OH have been detected (e.g., Gibb et al. 2000). These ices are condensed onto pre-existent grains. In these ices, some chemical reactions and ultraviolet photolysis (and cosmic rays) process the material and may make it refractory. As a result, so-called ‘core-mantle grains’ coated by refractory organics would be formed (e.g., Li & Greenberg 1997). Indeed, such a grain has been found in cometary dust: olivine particles produced by a Type II SN coated by organic matter which seems to be formed in a cold molecular cloud (Messenger et al. 2005). Therefore, the ISM dust probably has core-mantle or layered structures. Moreover, the composition can be heterogeneous: for example, graphite coated by silicate, silicate coated by graphite, silicate coated by iron, etc. The formation of such grains does not seem to be studied well. Much more experimental and theoretical works are highly encouraged. If we can find signatures of the dust accretion growth in the ISM of galaxies by astronomical observations (i.e. very distant remote-sensing), it proves the growth ubiquitous. A possible evidence already obtained is a huge mass of dust in galaxies which requires the accretion growth as discussed in this paper. It is worth studying how to distinguish stardust grains (or grain cores) and ISM dust (or mantle) by observations, e.g., spectropolarimetry, in future. Dust amount in the proto-solar nebula ------------------------------------- We have shown that the dust amount is very small before the ISM growth becomes active. For example, the dust-to-gas mass ratio is of the order of $10^{-4}$ at the first a few Gyr from the formation of the Milky Way (or the onset of the major star formation at the solar neighborhood). If the dust-to-gas ratio in the proto-solar nebula was $10^{-4}$, the planet formation might be difficult. Fortunately, the activation of the ISM growth is expected to be about 8 Gyr ago at the solar neighborhood. Thus, it is well before the solar system formation. Indeed, we expect the dust-to-gas ratio of several times $10^{-3}$ at 4–5 Gyr ago (see Figure 7). Moreover, the dust-to-gas ratio may be much enhanced in the proto-solar nebula relative to the average ISM. This is because the accretion growth is more efficient for higher density and the density in the proto-solar nebula is several orders of magnitude higher than that in molecular clouds. Therefore, even if the solar system formation is before the activation of the ISM growth globally, the dust growth may be active locally in the proto-solar nebula. In this case, the planet formation is always possible if there is enough metal to accrete onto the pre-existent seed grains, even before the global growth activation. This is an interesting issue to relate to the Galactic Habitable Zone where complex life can be formed (Lineweaver et al. 2004). We will investigate it more in future. Anders, E., Grevesse, N., Abundances of the elements - Meteoritic and solar, Geochim. Cosmochim. Acta., 53, 197–214, 1989. Arendt, R. G., Dwek, E., Petre, R., An infrared analysis of Puppis A, Astrophys. J., 368, 474–485, 1991. Arendt, R. G., Dwek, E., Blair, W. P., Ghavamian, P., Hwang, U., Long, K. S., Petre, R., Rho, J., Winkler, P. F., Spitzer Observations of Dust Destruction in the Puppis A Supernova Remnant, Astrophys. J., 725, 585–597, 2010. Asano, R., Takeuchi, T. T., Hirashita, H., Inoue, A. K., Dust formation history of galaxies: metallicity vs. grain growth, Astron. Astrophys., submitted. Asplund, M., Grevesse, N., Sauval, A. J., Scott, P., The Chemical Composition of the Sun, Ann. Rev. Astron. Astrophys., 47, 481–522, 2009. Barlow, M. J., Krause, O., Swinyard, B. M., Sibthorpe, B., Besel, M.-A., Wesson, R., Ivison, R. J., Dunne, L., et al., A Herschel PACS and SPIRE study of the dust content of the Cassiopeia A supernova remnant, Astron. Astrophys., 518, L138, 2010. Bertelli, G., Bressan, A., Chiosi, C., Fagotto, F., Nasi, E., Theoretical isochrones from models with new radiative opacities, Astron. Astrophys. Suppl., 106, 275–302, 1994. Bianchi, S., Ferrara, A., Intergalactic medium metal enrichment through dust sputtering, M.N.R.A.S., 358, 379–396, 2005. Bianchi, S., Schneider, R., Dust formation and survival in supernova ejecta, M.N.R.A.S., 378, 973–982, 2007. Borkowski, K. J., Williams, B. J., Reynolds, S. P., Blair, W. P., Ghavamian, P., Sankrit, R., Hendrick, S. P., Long, K. S., et al., Dust Destruction in Type Ia Supernova Remnants in the Large Magellanic Cloud, Astrophys. J., 642, L141–L144, 2006. Calura, F., Pipino, A., Matteucci, F., The cycle of interstellar dust in galaxies of different morphological types, Astron. Astrophys., 479, 669–685, 2008. Chabrier, G., Galactic Stellar and Substellar Initial Mass Function, Publ. Astron. Soc. Pac., 115, 763–795, 2003. Dopita, M. A., Ryder, S. D., On the law of star formation in disk galaxies, Astrophys. J., 430, 163–178, 1996. Draine, B. T., Evolution of interstellar dust, in The evolution of the interstellar medium, 13 pp, Astronomical Society of the Pacific, San Francisco, 1990. Draine, B. T., Interstellar Dust Models and Evolutionary Implications, in Cosmic Dust - Near and Far, Edited by Henning, T., Gr[ü]{}n, E., Steinacker, J., 20 pp, Astromomical Society of the Pacific, San Francisco, 2009. Draine, B. T., Salpeter, E. E., Time-dependent nucleation theory, J. of Chem. Phys., 67, 2230–2235, 1977. Draine, B. T., Salpeter, E. E., Destruction mechanisms for interstellar dust, Astrophys. J., 231, 438–455, 1979. Dunne, L., Eales, S., Ivison, R., Morgan, H., Edmunds, M., Type II supernovae as a significant source of interstellar dust, Nature, 424, 285–287, 2003. Dunne, L., Maddox, S. J., Ivison, R. J., Rudnick, L., Delaney, T. A., Matthews, B. C., Crowe, C. M., Gomez, H. L., Eales, S. A., Dye, S., Cassiopeia A: dust factory revealed via submillimetre polarimetry, M.N.R.A.S., 394, 1307–1316, 2009. Dwek, E., The Evolution of the Elemental Abundances in the Gas and Dust Phases of the Galaxy, Astrophys. J., 501, 643–665, 1998. Dwek, E., Scalo, J. M., The evolution of refractory interstellar grains in the solar neighborhood, Astrophys. J., 239, 193–211, 1980. Dwek, E., Arendt, R. G., Dust-gas interactions and the infrared emission from hot astrophysical plasmas, Ann. Rev. Astron. Astrophys., 30, 11–50, 1992. Dwek, E., Galliano, F., Jones, A. P., The Evolution of Dust in the Early Universe with Applications to the Galaxy SDSS J1148+5251, Astrophs. J., 662, 927–939, 2007. Dwek, E., Arendt, R. G., Bouchet, P., Burrows, D. N., Challis, P., Danziger, I. J., De Buizer, J. M., Gehrz, R. D., et al., Infrared and X-Ray Evidence for Circumstellar Grain Destruction by the Blast Wave of Supernova 1987A, Astrophys. J., 676, 1029–1039, 2008. Dwek, E., Cherchneff, I., The Origin of Dust in the Early Universe: Probing the Star Formation History of Galaxies by Their Dust Content, Astrophys. J., 727, 63, 2011. Edmunds, M. G., An elementary model for the dust cycle in galaxies, M.N.R.A.S., 328, 223–236, 2003. Edmunds, M. G., Eales, S. A., Maximum dust masses in galaxies, M.N.R.A.S., 299, L29–L31, 1998. Elmegreen, B. G., Star Formation on Galactic Scales: Empirical Laws, in Ecole Evry Schatzman 2010: Star Formation in the Local Universe. Lecture 1 of 5, 2011. (arXiv:1101.3108) Ferrarotti, A. S., Gail, H.-P., Composition and quantities of dust produced by AGB-stars and returned to the interstellar medium, Astron. Astrophys., 447, 553–576, 2006. Gall, C., Andersen, A. C., Hjorth, J., Genesis and evolution of dust in galaxies in the early Universe I. Modeling dust evolution in starburst galaxies, Astron. Astrophys., in press (arXiv:1011.3157) Gall, C., Andersen, A. C., Hjorth, J., Genesis and evolution of dust in galaxies in the early Universe II. Rapid dust evolution in quasars at $z>6$, Astron. Astrophys., in press (arXiv:1101.1553) Gehrz, R., Sources of Stardust in the Galaxy, in Interstellar Dust, Edited by Allamandola, L. J. and Tielens, A. G. G. M., 445 pp, International Astronomical Union. Symposium no. 135, Kluwer Academic Publishers, Dordrecht, 1989 Gibb, E. L., et al., An Inventory of Interstellar Ices toward the Embedded Protostar W33A, Astrophys. J., 536, 347–356, 2000. Gomez, H. L., Dunne, L., Ivison, R. J., Reynoso, E. M., Thompson, M. A., Sibthorpe, B., Eales, S. A., Delaney, T. M., Maddox, S., Isaak, K., Accounting for the foreground contribution to the dust emission towards Kepler’s supernova remnant, M.N.R.A.S., 397, 1621–1632, 2009. Heger, A., Fryer, C. L., Woosley, S. E., Langer, N., Hartmann, D. H., How Massive Single Stars End Their Life, Astrophys. J., 591, 288–300, 2003. Hirashita, H., Global Law for the Dust-to-Gas Ratio of Spiral Galaxies, Astrophs. J., 510, L99–L102, 1999a. Hirashita, H., Dust-to-gas ratio and phase transition of interstellar medium, Astron. Astrophys., 344, L87–L89, 1999b. Hirashita, H., Dust-to-Gas Ratio and Metallicity in Dwarf Galaxies, Astrophys. J., 522, 220–224, 1999c. Hirashita, H., Cyclic Changes in Dust-to-Gas Ratio, Astrophys. J., 531, 693–700, 2000a. Hirashita, H., Dust Growth Timescale and Mass Function of Molecular Clouds in the Galaxy, Publ. Astron. Soc. Jpn., 52, 585–588, 2000b. Hirashita, H., Effects of grain size distribution on the interstellar dust mass growth, M.N.R.A.S., submitted. Hirashita, H., Tajiri, Y. Y., Kamaya, H., Dust-to-gas ratio and star formation history of blue compact dwarf galaxies, Astron. Astrophys., 388, 439–445, 2002. Ikeuchi, S., Tomita, H., Cyclic phase changes of interstellar medium, Publ. Astron. Soc. J., 35, 77–86, 1983. Inoue, A. K., Evolution of Dust-to-Metal Ratio in Galaxies, Publ. Astron. Soc. Jpn., 55, 901–909, 2003. Inoue, A. K., Kamaya, H., Constraint on intergalactic dust from thermal history of intergalactic medium, M.N.R.A.S., 341, L7–L11, 2003. Inoue, A. K., Kamaya, H., Amount of intergalactic dust: constraints from distant supernovae and the thermal history of the intergalactic medium, M.N.R.A.S., 350, 729–744, 2004. Inoue, A. K., Kamaya, H., Intergalactic dust and its photoelectric heating, Earth, Planets, Space, 62, 69–79, 2010. Iwamoto, K., Brachwitz, F., Nomoto, K., Kishimoto, N., Umeda, H., Hix, W. R., Thielemann, F.-K., Nucleosynthesis in Chandrasekhar Mass Models for Type IA Supernovae and Constraints on Progenitor Systems and Burning-Front Propagation, Astrophys. J. Suppl., 125, 439–462, 1999. Jenkins, E. B., A Unified Representation of Gas-Phase Element Depletions in the Interstellar Medium, Astrophys. J., 700, 1299–1348, 2009. Jones, A. P., Tielens, A. G. G. M., Hollenbach, D. J., McKee, C. F., Grain destruction in shocks in the interstellar medium, Astrophys. J., 433, 797–810, 1994. Jones, A. P., Tielens, A. G. G. M., Hollenbach, D. J., Grain Shattering in Shocks: The Interstellar Grain Size Distribution, Astrophys. J., 469, 740–764, 1996. Karakas, A. I., Updated stellar yields from asymptotic giant branch models, M.N.R.A.S., 403, 1413–1425, 2010. Kennicutt, R. C., The Global Schmidt Law in Star-forming Galaxies, Astrophys. J., 498, 541–552, 1998. Kimura, H., Mann, I., Jessberger, E. K., Composition, Structure, and Size Distribution of Dust in the Local Interstellar Cloud, Astrophys. J., 583, 314–321, 2003. Kozasa, T., Hasegawa, H., Grain Formation through Nucleation Process in Astrophysical Environments. II —Nucleation and Grain Growth Accompanied by Chemical Reaction—, Progress of Theoretical Physics, 77, 1402–1410, 1987. Kozasa, T., Nozawa, T., Tominaga, N., Umeda, H., Maeda, K., Nomoto, K., Dust in Supernovae: Formation and Evolution, in Cosmic Dust - Near and Far, Edited by Henning, T., Gr[ü]{}n, E., Steinacker, J., 43 pp, Astromomical Society of the Pacific, San Francisco, 2009. Krause, O., Birkmann, S. M., Rieke, G. H., Lemke, D., Klaas, U., Hines, D. C., Gordon, K. D., No cold dust within the supernova remnant Cassiopeia A, Nature, 432, 596–598, 2004. Kroupa, P., The Initial Mass Function of Stars: Evidence for Uniformity in Variable Systems, Science, 295, 82–91, 2002. Larson, R. B., Early star formation and the evolution of the stellar initial mass function in galaxies, M.N.R.A.S., 301, 569–581, 1998. Larson, R. B., Tinsley, B. M., Caldwell, C. N., The evolution of disk galaxies and the origin of S0 galaxies, Astrophys. J., 237, 692–707, 1980. Li, A., Greenberg, j. M., A unified model of interstellar dust, Astron. Astrophys., 323, 566–584, 1997. Lineweaver, C. H., Fenner, Y., Gibson, B. K., The Galactic Habitable Zone and the Age Distribution of Complex Life in the Milky Way, Science, 303, 59–62, 2004. Lisenfeld, U., Ferrara, A., Dust-to-Gas Ratio and Metal Abundance in Dwarf Galaxies, Astrophys. J., 496, 145–154, 1998. Maiolino, R., Schneider, R., Oliva, E., Bianchi, S., Ferrara, A., Mannucci, F., Pedani, M., Roca Sogorb, M., A supernova origin for dust in a high-redshift quasar, Nature, 431, 533–535, 2004. Matsuura, M., Barlow, M. J., Zijlstra, A. A., et al., The global gas and dust budget of the Large Magellanic Cloud: AGB stars and supernovae, and the impact on the ISM evolution, M.N.R.A.S., 396, 918–934, 2009. Mattsson, L., Dust in the early Universe: Evidence for non-stellar dust production or observational errors?, M.N.R.A.S., in press (arXiv:1102.0570) McKee, C., Dust Destruction in the Interstellar Medium, in Interstellar Dust, Edited by Allamandola, L., Tielens, A. G. G. M., 14 pp, Kluwer Academic Publishers, Dordrecht, 1989. M[é]{}nard, B., Scranton, R., Fukugita, M., Richards, G., Measuring the galaxy-mass and galaxy-dust correlations through magnification and reddening, M.N.R.A.S., 405, 1025–1039, 2010. Messenger, S., Identification of melecular-cloud material in interplanetary dust particles, Nature, 404, 968–971, 2000. Messenger, S., Keller, L. P., Lauretta, D. S., Supernova Olibine from Cometary Dust,textit[Science]{}, 309, 737–741, 2005. Michałowski, M. J., Murphy, E. J., Hjorth, J., Watson, D., Gall, C., Dunlop, J. S., Dust grain growth in the interstellar medium of $5<z<6.5$ quasars, Astron. Astrophys., 522, 15, 2010. Morgan, H. L., Edmunds, M., G., Dust formation in early galaxies, M.N.R.A.S., 343, 427–442, 2003. Morgan, H. L., Dunne, L., Eales, S. A., Ivison, R. J., Edmunds, M. G., Cold Dust in Kepler’s Supernova Remnant, Astrophys. J., 597, L33–L36, 2003. Mouri, H., Taniguchi, Y., Grain Survival in Supernova Remnants and Herbig-Haro Objects, Astrophys. J., 534, L63–L66, 2000. Naab, T., Ostriker, J. P., A simple model for the evolution of disc galaxies: the Milky Way, M.N.R.A.S., 366, 899–917. 2006. Nath, B. B., Laskar, T., Shull, J. M., Dust Sputtering by Reverse Shocks in Supernova Remnants, Astrophys. J., 682, 1055–1064, 2008. Nomoto, K., Tominaga, N., Umeda, H., Kobayashi, C., Maeda, K., Nucleosynthesis yields of core-collapse supernovae and hypernovae, and galactic chemical evolution, Nuclear Physics A, 777, 424–458, 2006. Nozawa, T., Kozasa, T., Umeda, H., Maeda, K., Nomoto, K., Dust in the Early Universe: Dust Formation in the Ejecta of Population III Supernovae, Astrophys. J., 598, 785–803, 2003. Nozawa, T., Kozasa, T., Habe, A., Dust Destruction in the High-Velocity Shocks Driven by Supernovae in the Early Universe, Astrophys. J., 648, 435–451, 2006. Nozawa, T., Kozasa, T., Habe, A., Dwek, E., Umeda, H., Tominaga, N., Maeda, K., Nomoto, K., Evolution of Dust in Primordial Supernova Remnants: Can Dust Grains Formed in the Ejecta Survive and Be Injected into the Early Interstellar Medium?, Astrophys. J., 666, 955–966, 2007. Nozawa, T., et al., Formation and Evolution of Dust in Type IIb Supernovae with Application to the Cassiopeia A Supernova Remnant, Astrophys. J., 713, 356–373, 2010. Onaka, T., Kamijo, F., Destruction of interstellar grains by sputtering, Astton. Astrophys., 64, 53–60, 1978. Pagel, B. E. J., The G-dwarf problem and radio-active cosmochronology, in Evolutionary phenomena in galaxies, 23pp., Cambridge University Press, Cambridge and New York, 1989. Peacock, J. A., Cosmological Physics, pp. 704, Cambridge University Press, Cambridge, 1999. Pipino, A., Fan, X. L., Matteucci, F., Calura, F., Silva, L., Granato, G., Maiolino, R., The chemical evolution of elliptical galaxies with stellar and QSO dust production, Astron. Astrophys., 525, A61, 2011. Rho, J., et al., Freshly Formed Dust in the Cassiopeia A Supernova Remnant as Revealed by the Spitzer Space Telescope, Astrophys. J., 673, 271–282, 2008. Rocha-Pinto, H. J., Scalo, J., Maciel, W. J., Flynn, C., Chemical enrichment and star formation in the Milky Way disk. II. Star formation history, Astron. Astrophys., 358, 869–885, 2000a. Rocha-Pinto, H. J., Maciel, W. J., Scalo, J., Flynn, C., Chemical enrichment and star formation in the Milky Way disk. I. Sample description and chromospheric age-metallicity relation, Astron. Astrophys., 358, 850–868, 2000b. Sakon, I., et al., Properties of Newly Formed Dust by SN 2006JC Based on Near- to Mid-Infrared Observation With AKARI, Astrophys. J., 692, 546–555, 2009. Salpeter, E. E., The Luminosity Function and Stellar Evolution, Astrophys. J., 121, 161–167, 1955. Sankrit, R., Williams, B. J., Borkowski, K. J., Gaetz, T. J., Raymond, J. C., Blair, W. P., Ghavamian, P., Long, K. S., Reynolds, S. P., Dust Destruction in a Non-radiative Shock in the Cygnus Loop Supernova Remnant, Astrophys. J., 712, 1092–1099, 2010. Savage, B. D., Sembach, K. R., Interstellar Abundances from Absorption-Line Observations with the Hubble Space Telescope, Ann. Rev. Astron. Astrophys., 34, 279–330, 1996. Schmidt, M., The Rate of Star Formation, Astrophys. J., 129, 243–258, 1959. Schneider, R., Ferrara, A., Salvaterra, R., Omukai, K., Bromm, V., Low-mass relics of early star formation, Nature, 422, 869–871, 2003. Schneider, R., Ferrara, A., Salvaterra, R., Dust formation in very massive primordial supernovae, M.N.R.A.S., 351, 1379–1386, 2004. Schneider, R. Omukai, K. Inoue, A. K.; Ferrara, A., Fragmentation of star-forming clouds enriched with the first dust, M.N.R.A.S., 369, 1437–1444, 2006. Sibthorpe, B., Ade, P. A. R., Bock, J. J., Chapin, E. L., Devlin, M. J., Dicker, S., Griffin, M., Gundersen, J. O., et al., AKARI and BLAST Observations of the Cassiopeia A Supernova Remnant and Surrounding Interstellar Medium, Astrophys. J., 719, 1553–1564, 2010. Silvia, D. W., Smith, B. D., Shull, J. M., Numerical Simulations of Supernova Dust Destruction. I. Cloud-crushing and Post-processed Grain Sputtering, Astrophys. J., 715, 1575–1590, 2010. Songaila, A. Cowie, L. L., Metal enrichment and Ionization Balance in the Lyman Alpha Forest at $z=3$, Astron. J., 112, 335–351, 1996. Takeuchi, T. T., Hirashita, H., Testing Intermittence of the Galactic Star Formation History along with the Infall Model, Astrophys. J., 540, 217–223, 2000. Tielens, A. G. G. M., Interstellar Depletions and the Life Cycle of Interstellar Dust, Astrophys. J., 499, 267–272, 1998. Tinsley, B. M., Evolution of the Stars and Gas in Galaxies, Fundam. Cosmic Phys., 5, 287–388, 1980. Todini, P., Ferrara, A., Dust formation in primordial Type II supernovae, M.N.R.A.S., 325, 726–736, 2001. Valiante, R., Schneider, R., Bianchi, S., Andersen, A. C., Stellar sources of dust in the high-redshift Universe, M.N.R.A.S., 397, 1661–1671, 2009. van den Bergh, S., The Galaxies of the Local Group, pp. 328, Cambridge University Press, 2000. Williams, B. J., Borkowski, K. J., Reynolds, S. P., Blair, W. P., Ghavamian, P., Hendrick, S. P., Long, K. S., Points, S., et al., Dust Destruction in Fast Shocks of Core-Collapse Supernova Remnants in the Large Magellanic Cloud, Astrophys. J., 652, L33–L36, 2006. Yamamoto, T., Hasegawa, H., Grain Formation through Nucleation Process in Astrophysical Environment, Progress of Theoretical Physics, 58, 816–828, 1977. Zhukovska, S., Gail, H.-P., Trieloff, M., Evolution of interstellar dust and stardust in the solar neighbourhood, Astron. Astrophys., 479, 453–480, 2008. \[finalpage\] [^1]: $^1$Without gas infall from the intergalactic space, we expect a much larger number of low-metallicity stars at the solar neighborhood than the observed. This is called the ‘G-dwarf problem’ (e.g., Pagel 1989). [^2]: $^2$The origin of intergalactic metals and dust is galactic outflows and the amount ejected from galaxies is the same order of that remained in galaxies (e.g., M[é]{}nard et al. 2010 for dust; see also Inoue & Kamaya 2003, 2004, and 2010). Dust grains may be ejected from galaxies more efficiently than metals because the grains receive momentum through radiation pressure (Bianchi & Ferrara 2005). Even in this case, our discussion about the dust-to-metal ratio in §5 would not be affected essentially by omission of this selective removal of dust, although the set of model parameters which can reproduce the observations would change. In any case, this point would be an interesting future work. [^3]: $^3$The micro-process of the destruction considered in Nozawa et al. (2007) is sputtering by hot gas. [^4]: $^4$The ‘mixed’ and ‘unmixed’ mean the elemental mixing in the SN ejecta (Nozawa et al. 2003). In the ‘mixed’ case, there is no layer where C is more abundant than O, then, only silicate, troilite, and corundum grains can be formed. On the other hand, the ‘unmixed’ case has a C-rich layer and Fe layer and can form carbon and iron grains as well as silicate. [^5]: $^5$The difference between $Z_\odot$s by Anders & Grevesse (1989) and Asplund et al. (2009) accounts for the uncertainty.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider collision of two particles 1 and 2 near the horizon of the extremal rotating axially symmetric neutral generic black hole producing particles 3 and 4. We discuss the scenario in which both particles 3 and 4 fall into a black hole and move in a white hole region. If particle 1 is fine-tuned, the energy $E_{c.m.}$ in the centre of mass grows unbounded (the Bañados-Silk-West effect). Then, particle 3 can, in principle, reach a flat infinity in another universe. If not only $E_{c.m.}$ but also the corresponding Killing energy $E$ is unbounded, this gives a so-called super-Penrose process (SPP). We show that the SPP is indeed possible. Thus white holes turn out to be potential sources of high energy fluxes that transfers from one universe to another. This generalizes recent observaitons made by Patil and Harada for the Kerr metric. We analyze two different regimes of the process on different scales.' author: - 'O. B. Zaslavskii' title: 'Super-Penrose process for extremal rotating neutral white holes' --- Introduction ============ During last decade a lot of work has been made for investigation of properties of high energy processes near black holes. This was stimulated by the paper [@ban], where it was found that collision of two particles near rapidly rotating black holes can lead to unbounded energies $E_{c.m.}$ in the centre of mass frame. This was called the Bañados-Silk-West (BSW) effect, after the authors’ names. After its publication, it turned out, that there are also earlier works [@katz], [@ps] in which near-horizon particle collisions in the Kerr metric were investigated. Meanwhile, a typical process considered there, includes head-on collision between two arbitrary particles, where particle 1 arrives from infinity while particle 2 comes from the horizon (see eq. 2.57 of [@ps]). But as far as particle 2 is concerned, this is nothing else than a typical behavior of a particle near a white hole. If such a region is allowed in the complete space-time, the effect of unbounded $E_{c.m.}$ for head-on collisions exists even in the Schwarzschild metric [@tot], [@white]. Thus white holes can be an alternative to black ones as a source of high energy collisions. More important question is whether it is possible to gain not only unbounded $E_{c.m.}$ but also unbounded conserved Killing energies $E$ since it is the latter quantity which can be measured in the Earth laboratory, at least in principle. The collisions in the Schwarzschild background are useless for this purpose since energy cannot be extracted at all. For such an extraction, the existence of negative energies and ergoregion are required that makes it possible the Penrose process [@pen] or its collisional analogue [@col]. If the energy gain is unbounded, this is called the super-Penrose process (SPP). For black holes, the energy gain is finite, so the SPP is impossible for them (see [@is] and references therein). In this context, there is a scenario with participation of white holes, different from those in [@katz] - [@white]. Now, both particles collide near the black hole horizon in “our” part of Universe but afterwards the products of reaction leave it. Passing though the horizon, they appear inside a white whole region and, eventually, transfer energy to another universe. Or, vice verse, collision in another universe can give rise to high energy in our one. If such a process is possible, this would give astrophysical realization of high energy transfer with white holes as a source that was suggested earlier [@nov], [@dad]. The concrete process of this kind in the Kerr background was considered recently in [@ph]. The authors showed that the conserved energy of produced particles can be as large as one like. In other words, the SPP is possible. Our aim is to extend consideration to generic rotating axially symmetric stationary white holes. In doing so, we exploit the approach that, in our view, is simpler and was already used for examination of the energy extraction from generic black holes of the aforementioned type including the Kerr metric [@up], [@en]. In particular, we do not use transformation between the three frames (center of mass, locally non-rotating and stationary ones) and work in the original frame. Although there are reasons to believe that white holes are unstable (see Sec. 15 of [@fn]), motivation for consideration of such objects stems from different roots. (i) High energy process, if they are confirmed, can themselves contribute to the instability of white holes, so they are important for elucidation of the fate of such objects. (ii) The complete theory of the BSW effect should take into consideration all possible configurations and scenarios, at least for better understanding the phenomenon. We use the system of units in which the fundamental constants $G=c=1$. Basic equations =============== Let us consider the metric$$ds^{2}=-N^{2}dt^{2}+\frac{dr^{2}}{A}+g_{\phi }(d\phi -\omega dt)^{2}+g_{\theta }d\theta ^{2}\text{,}$$where $g_{\phi }\equiv g_{\phi \phi }$, $g_{\theta }\equiv g_{\phi \phi }$, all coefficients do not depend on $t$ and $\phi $. For a given energy $E$, angular momentum $L$ and mass $m$ the equation of motion in the equatorial plane read$$m\dot{t}=\frac{X}{N^{2}}\text{,}$$$$X=E-\omega L\text{,} \label{X}$$$$m\dot{\phi}=\frac{L}{g_{\phi }}$$$$m\dot{r}=\sigma P\text{, }\sigma =\pm 1\text{,}$$$$P=\sqrt{X^{2}-\tilde{m}^{2}N^{2}}\text{,} \label{P}$$$$\tilde{m}^{2}=m^{2}+\frac{L^{2}}{g_{\phi }}\text{,}$$dot denotes differentiation with respect to the proper time $\tau $. To simplify formulas, we assume that in the equatorial plane $A=N^{2}$. Otherwise, we can always achieve this equality by redefining the radial coordinate. The forward-in-time condition requires$$X\geq 0\text{.} \label{ftc}$$ In what follows, we will use the standard classification of particles. If $% X_{H}=0$, a particle is called critical. If $X_{H}=O(1)$, it is called usual. If $X_{H}=O(N_{c}),$ it is called near-critical. Here, subscripts “H” and “c” refer to the quantities calculated on the horizon and the point of collision, respectively. For the near-critical particle, we use presentation$$L=\frac{E}{\omega _{H}}(1+\delta ) \label{LE}$$exploited in [@en]. Here,$$\delta =C_{1}N_{c} \label{del}$$is a small quantity for collisions near the horizon, $C_{1}=O(1)$ is a constant. Near the horizon, we assume the Taylor expansion that for the extremal case reads [@tan]$$\omega =\omega _{H}-B_{1}N+O(N^{2})\text{.}$$ Then, we have the following approximate expressions there. The critical particle: $$X=\frac{b}{h}EN+O(N^{2})\text{,} \label{cr}$$$$P=N\sqrt{E^{2}\left( \frac{b^{2}-1}{h^{2}}\right) -\frac{1}{h^{2}}-m^{2}}.$$ A usual particle:$$X=X_{H}+B_{1}LN+O(N^{2})\text{, }$$$$X_{H}=E-\omega _{H}L\text{,}$$$$P=X+O(N^{2}). \label{us}$$ The near-critical particle:$$X=E(\frac{b}{h}-C_{1})N \label{xncr}$$$$P=N\sqrt{E^{2}[(\frac{b}{h}-C_{1})^{2}-\frac{1}{h^{2}}]-m^{2}}+O(N^{2})\text{% .} \label{ncr}$$ We introduced notations $b=B_{1}\sqrt{g_{H}}$, $h=\omega _{H}\sqrt{g_{H}}$. Scenario of collision ===================== In the point of collision we assume the validity of conservation laws for the energy and angular momentum: $$E_{0}\equiv E_{1}+E_{2}=E_{3}+E_{4}\text{,} \label{e12}$$ $$L_{0}=L_{1}+L_{2}=L_{3}+L_{4}. \label{l12}$$ It follows from (\[e12\]), (\[l12\]) that$$X_{0}\equiv X_{1}+X_{2}=X_{3}+X_{4}. \label{x12}$$ There is also the conservation law for the radial momentum:$$\sigma _{1}P_{1}+\sigma _{2}P_{2}=\sigma _{3}P_{3}+\sigma _{4}P_{4}\text{.} \label{r}$$ We assume that particles 1 and 2 fall from infinity, so $\sigma _{1}=\sigma _{2}=-1$. We are interested in high energy processes in which $E_{c.m.}$ is unbounded since this is the necessary condition for $E$ to be unbounded as well [@inf], [@wald]. To this end, we choose particle 1 to be the critical, particle 2 being usual since this gives rise to the unbounded $% E_{c.m.}$ [@ban], [@prd]. Then, one of particles (say, 3) is near-critical and the other one (4) is usual [@up], [@en]. All possible scenarios can be described by two parameters - the sign of $C_{1}$ and the value of $\sigma _{3}$ immediately after collision (OUT for $\sigma _{3}=+1$ and IN for $\sigma _{3}=-1$). As a result, we have 4 scenarios OUT$% + $, OUT$-$, IN$+$, IN$-$. The first three were already analyzed in [@en], where IN$-$ was rejected since it corresponds to both particles falling into a black hole. However, now it is this scenario which we focus on. It corresponds to high energy propagation in the white hole region (see below). Thus we have $\sigma _{3}=\sigma _{4}=-1$. We must analyze the process under discussion for $N_{c}\rightarrow 0$ on the basic of the conservation law (\[r\]). In doing so, we follow the lines of Ref. [@en] applying the corresponding approach to the case that was not considered there. Lower bounds on energy\[low\] ============================= If we collect the terms of the zeroth and first order in $N_{c}$ and take into account the approximate expressions (\[cr\]) - (\[ncr\]), we obtain$$F=-\sqrt{E_{3}^{2}[(\frac{b}{h}-C_{1})^{2}-\frac{1}{h^{2}}]-m_{3}^{2}}, \label{F}$$where$$F\equiv A+E_{3}(C_{1}-\frac{b}{h}), \label{FA}$$$$A=\frac{E_{1}b-\sqrt{E_{1}^{2}(b^{2}-1)-m_{1}^{2}h^{2}}}{h}, \label{A1}$$$$C_{1}=\frac{b}{h}-\frac{A^{2}+m_{3}^{2}+\frac{E_{3}^{2}}{h^{2}}}{2E_{3}A},$$$$F=\frac{A^{2}-m_{3}^{2}-\frac{E_{3}^{2}}{h^{2}}}{2A}.$$ We are interested in scenario IN$-$. Then, $C_{1}<0$ gives us $$E_{3}^{2}-2E_{3}hA_{1}b+h^{2}(A^{2}+m_{3}^{2})>0, \label{e2}$$that can be rewritten as$$\left( E_{3}-\lambda _{+}\right) (E_{3}-\lambda _{+})>0,$$$$\lambda _{\pm }=h[A_{1}b\pm \sqrt{A^{2}(b^{2}-1)-m_{3}^{2}}]\text{.}$$ The condition $F<0$ gives us$$E_{3}^{2}>h^{2}(A^{2}-m_{3}^{2})\equiv \lambda _{0}^{2}\text{.} \label{e0}$$ If $\lambda _{\pm }$ are real, both bounds $E_{3}>\lambda _{+}$ and $% E_{3}>\lambda _{0}$ are quite compatible with each other. If $\lambda _{\pm } $ are complex, (\[e2\]) and (\[e0\]) are mutually consistent as well. Thus there is no upper bound on $E_{3}$ and the SPP is possible. As far as particle 4 is concerned, it has $E_{4}<0$. To obey the forward-in-time condition (\[ftc\]), it must have $L_{4}=-\left\vert L_{4}\right\vert <0$. Then, $X_{4}=\left\vert L_{4}\right\vert \omega -\left\vert E_{4}\right\vert $. Assuming that there is a flat infinity, where $\omega \rightarrow 0$, we see that particle 4 either falls into singularity or oscillates between turning points $r_{1}$ and $r_{2}$. In doing so, it can intersect the horizons, thus appearing in new “universes” due to a potentially rich space-time structure inside similarly to what takes place for the Kerr metric [@ck]. However, under a rather weak and reasonable restrictions on the properties of the metric, it cannot have more than 1 turning point in the outer region, so the situation when $% r_{+}<r_{1}\leq r\leq r_{2}$ is impossible. This was shown for the Kerr metric in [@gpneg] and generalized in [@myneg]. For more information about trajectories of particles 3 and 4, the metric should be specified. The above treatment changes only slightly if we consider the Schnittman process [@sch] when the critical particle 1 does not come from infinity but moves from the horizon. Then, instead of (\[A1\]), we should take $A=% \frac{E_{1}b+\sqrt{E_{1}^{2}(b^{2}-1)-m_{1}^{2}h^{2}}}{h}$. Superenergetic particles ======================== In the above treatment, we tacitly assumed that all energies and angular momenta are finite and do not grow unbounded when $N_{c}\rightarrow 0$. The only place where $N_{c}$ appear in the relation between them are equalities (\[LE\]), (\[del\]), where it gives only small corrections. The above approximate expressions for particle characteristics (\[cr\]) - (\[ncr\]) take into account this circumstances. In particular, for a usual particle, $X=O(1)$, the second term in the radical in (\[P\]) has the order $% N_{c}^{2}$. For a near-critical one, both terms in $P$ have the order $% O(N_{c})$. Meanwhile, it turns out that there exists self-consistent scenario, in which$$L_{3}=\frac{l_{3}}{\sqrt{N_{c}}}\text{,} \label{Ll}$$where $l_{3}$ is some coefficient not containing $N_{c}$. For small $N_{c}$, $L_{4}=L_{0}-L_{3}\approx -\frac{l_{3}}{\sqrt{N_{c}}}$. It was found in [@center], where it was pointed out that it corresponds to falling both particles in a black hole, so it was put aside since we were interested in particles returning to infinity. But now, it is this case that came into play. Therefore, we take advantage of formulas already derived in [@center] but exploit them in a new context - see eqs. (\[X3\]), (\[be\]) below. If (\[Ll\]) is satisfied, the previous consideration fails and the conservation law (\[r\]) is to be analyzed anew. Now, for particles 3 and 4 the second term in (\[P\]) gives a small correction (whereas for finite $L_{3}$ both terms for particle 3 would have the same order), so$$P_{3,4}\approx \sqrt{X_{3,4}^{2}-N_{c}\frac{l_{3,4}^{2}}{\left( g_{\phi }\right) _{H}}}\approx X_{3,4}-\frac{N_{c}}{2X_{3,4}}\frac{l_{3}^{2}}{\left( g_{\phi }\right) _{H}}\text{. }$$ Then, taking into account (\[cr\]) - (\[us\]) for particles 1 and 2, (\[e12\]) - (\[x12\]) and discarding the terms $O(N_{c}^{2}$) and higher, one can show after algebraic manipulations that the following equation holds:$$\frac{l_{3}^{2}}{2\left( g_{\phi }\right) _{H}}(\frac{1}{X_{3}}+\frac{1}{% X_{4}})=A\text{.}$$Using again (\[x12\]), one can obtain$$\left( X_{3,4}\right) _{c}\approx \frac{\left( X_{0}\right) _{c}}{2}(1\mp \sqrt{1-b})\text{,} \label{X3}$$where$$b\equiv \frac{2l_{3}^{2}}{\left( g_{\phi }\right) _{H}X_{0}A}\text{.} \label{be}$$It is implied that $b<1$. It follows from definition (\[X\]) that now $$E_{3}=\left( X_{3}\right) _{H}+\omega _{H}L_{3}\approx \left( X_{3}\right) _{c}+\omega _{H}\frac{l_{3}}{\sqrt{N_{c}}}\text{.}$$ Thus we have two usual particles which come down into a black hole. This is contrasted with the standard case when particle 3 is near-critical and returns to infinity. We see that there are two energy scales for the SPP. On the first scale, $% E_{3}$ can be as large as we like but with reservation that $E_{3}\ll \frac{% const}{\sqrt{N_{c}}}$. On the second scale, $E_{3}\sim \frac{1}{\sqrt{N_{c}}}% .$ In doing so, $E_{4}=E_{0}-E_{3}$ is negative having the same order $% N_{c}^{-1/2}$. Discussion about the properties of the trajectories of such particles from Section \[low\] applies now as well. Conclusions =========== Thus we showed that particle collision on our side of Universe (near the black hole horizon) can lead to high energy fluxes on the other side. If, vice verse, collision occurs in “another world”, we can detect its consequences in our one. We did not resort to the transformation between the original stationary frame and the center of mass one. We would like to stress that high energy behavior is found for the energies $E_{3}$ that can be in principle detected in a laboratory. These results qualitatively agree with claims made in [@ph] for the Kerr metric. It is instructive to compare the situation for static charged and neutral rotating black/white holes collecting the results of the present and previous works [@rn], [@whq]. SPP $q_{3}$ $L_{3}$ ------------------------------ ----- --------- ------------------- charged black holes yes large arbitrary charged white holes yes large $O(N_{c}^{-1/2})$ neutral rotating black holes no $0$ arbitrary neutral rotating white holes yes $0$ unbounded Table 1. Conditions of the existence of the super-Penrose process. We see that for the Reissner-Nordström metric, if we compare collisions near black and white holes, the situation is partially complementary to each other. For finite $L_{i}$ (for example, with all $L_{i}=0$), there is the SPP in the black hole case. However, it fails to exist near white holes. As shown in [@whq], only if particles with angular momenta $% L_{3,4}=O(N_{c}^{-1/2})$ come into play, we obtain the SPP. Meanwhile, for the rotating case, the SPP does not exist for black holes at all. Instead, the white hole scenario opens new possibilities for the SPP, in which $E$ and $L$ of particles at infinity are unbounded. This happens on two scales: on the first one $E$ and $L$ do not contain the parameter $N_{c}^{-1}$, on the second scale they have the order $O(N_{c}^{-1/2})$ similarly to the static charged case. The concrete properties of collisions are described by somewhat different formulas and the type of energetic particles are different: in the first case particle 3 is near-critical, in the second one it is usual. However, in both cases the conserved energy $E_{3}$ is unbounded. The phenomenon under discussion is two-faced. On one hand, it shows that high energy processes are indeed possible due to white holes and poses anew the question about their potential role in nature. From the other hand, it poses also a question about backreaction of such collisions on the metric itself, including the fate of white holes. This work was supported by the Russian Government Program of Competitive Growth of Kazan Federal University. [99]{} Bañados, M., Silk, J., West, S.M.: Kerr black holes as particle accelerators to arbitrarily high energy. Phys. Rev. Lett. 103, 111102 (2009). arXiv:0909.0169 Piran T., Shaham J., Katz J.: High efficiency of the Penrose mechanism for particle collisions.* Astorphys. J.* 196, L107 (1975) T. Piran and J. Shaham.: Upper bounds on collisional Penrose processes near rotating black-hole horizons, Phys. Rev. D 16, 1615 (1977) Grib A. A., Pavlov Yu. V.: Are black holes totally black? Gravit. Cosmol. 21, 13 (2015). arXiv:1410.5736 Zaslavskii O. B.: On white holes as particle accelerators.* Gravit. Cosmol. 24,* 92 (2018) arXiv:1707.07864 Penrose R.: Gravitational Collapse: the role of general relativity. Rivista del Nuovo Cimento. Numero Speziale I, 252 (1969) Schnittman J. D..: The Collisional Penrose Process. Gen. Relat. and Gravitation. 50 (2018) 77. arXiv:1910.02800 Zaslavskii O. B.: Is the super-Penrose process possible near black holes? Phys. Rev. D 93 (2016), 024056 arXiv:1511.07501 Novikov I. D.: Delayed explosion of a part of the Friedmann Universe and quasars. Astron. Zh. 6, 1075 (1964) \[Sov. Astronomy 8, 857 (1964)\]. Narlikar J. V., Appa Rao K. M. V., Dadhich N. .: High energy radiation from white holes. Nature* *251, 590 (1974). Patil M., Harada T.: Extremal Kerr white holes as a source of ultra high energy particles. arXiv:2004.12874 T. Harada, H. Nemoto, and U. Miyamoto.: Upper limits of particle emission from high-energy collision and reaction near a maximally rotating Kerr black hole. Phys. Rev. D 86, 024 027 (2012). arXiv:1205.7088 Zaslavskii O. B.: Energetics of particle collisions near dirty rotating extremal black holes: Banados-Silk-West effect versus Penrose process. Phys. Rev. D 86, 084030 (2012), arXiv:1205.4410 Frolov, V.P., Novikov, I.D.: Physics of Black Holes. Kluwer Academic, Dordrecht (1998) Tanatarov I. V., Zaslavskii O. B.: Dirty rotating black holes: regularity conditions on stationary horizons. Phys. Rev. D 86, 044019, (2012). arXiv:1206.2580 Patil, M., Harada, T., Nakao, K., Joshi, P.S., Kimura, M.: Infinite efficiency of collisional Penrose process: can over-spinningKerr geometry be the source of ultra-high-energy cosmic rays and neutrinos? Phys. Rev. D 93, 104015 (2016). arXiv:1510.08205 Tanatarov I. V., Zaslavskii O. B.: Collisional super-Penrose process and Wald inequalities. Gen. Relativ. Gravit. 49, 119 (2017). arXiv:1611.05912 O.B. Zaslavskii.: Acceleration of particles as universal property of rotating black holes, Phys. Rev. D 82 (2010) 083004. arXiv:1007.3678 Carter B.: Global structure of the Kerr family of gravitational fields. Phys. Rev. 174, 1559 (1968) Grib A. A., Pavlov Yu. V., Vertogradov V. D.: Geodesics with negative energy in the ergosphere of rotating black holes. Mod. Phys. Lett. A 29, 1450110 (2014). arXiv:1304.7360 O. B. Zaslavskii.: On geodesics with negative energies in the ergoregions of dirty black holes, Mod. Phys. Lett. A 30, 1550055 (2015). arXiv:1412.1725 J. D. Schnittman.: Revised upper limit to energy extraction from a Kerr black hole, Phys. Rev. Lett.* *113, 261102 (2014) arXiv:1410.6446 O.B. Zaslavskii.: Center of mass energy of colliding electrically neutral particles and super-Penrose process Phys. Rev. D 100, 024050 (2019). arXiv:1904.04874 Zaslavskii O. B.: Energy extraction from extremal charged black holes due to the BSW effect. Phys. Rev.* D 86*, 124039 (2012) arXiv:1207.5209 Zaslavskii O. B.: Super-Penrose process for extremal charged white holes. arXiv:2005.11090	{ "pile_set_name": "ArXiv" }
--- abstract: 'We have recently demonstrated static trapping of ammonia isotopologues in a decelerator that consists of a series of ring-shaped electrodes to which oscillating high voltages are applied \[Quintero-Pérez et al., [Phys. Rev. Lett. 110, 133003 (2013)](http://prl.aps.org/abstract/PRL/v110/i13/e133003)\]. In this paper we provide further details about this traveling wave decelerator and present new experimental data that illustrate the control over molecules that it offers. We analyze the performance of our setup under different deceleration conditions and demonstrate phase-space manipulation of the trapped molecular sample.' author: - Paul Jansen - 'Marina Quintero-Pérez' - 'Thomas E. Wall' - 'Joost E. van den Berg' - Steven Hoekstra - 'Hendrick L. Bethlem' title: Deceleration and trapping of ammonia molecules in a traveling wave decelerator --- Introduction ============ Cold molecules offer many exciting prospects in both chemistry and physics (for recent review papers see Refs. [@Carr:NJP2009; @Hogan2011; @vandeMeerakker:ChemRev2012; @Narevicius:ChemRev2012]). The great control that can be exerted over cold molecules allows the study and manipulation of collisions and chemical reactions [@Ospelkaus2010; @Kirste23112012]. The strong dipole-dipole interactions between cold molecules make them excellent systems for quantum simulation and computation [@DeMille2002; @Micheli2006; @Andre2006]. With their rich structure, molecules can be useful systems for making precision tests of fundamental physics, such as the measurements of the electron EDM [@Hudson:Nature2011; @Campbell2013] and the search for time-variation of fundamental constants [@Veldhoven2004; @Hudson:PRL2006; @Shelkovnikov2008; @Truppe2013]. In particular this last application is of great interest to our group, and as a part of this research we are building a molecular fountain [@Bethlem2008]. A fountain requires great control over the molecules, in our case slowing them from around 300 m/s to rest, cooling them, and then launching them upwards at around 3 m/s. A tool that has been successfully used to exert control over molecules is the Stark decelerator [@Bethlem1999; @Bethlem:PRA2002; @vandeMeerakker:ChemRev2012]. While excellent at removing kinetic energy, a Stark decelerator becomes lossy for low speeds ($<$100m/s) because of its reliance on creating an effective 3D potential well. When the characteristic wavelength of the longitudinal and transverse motion becomes comparable to the periodicity of the decelerator, the approximations used to derive this effective well no longer apply and both the number of molecules and phase-space density decline with speed [@Sawyer:Thesis; @Jansen:Thesis]. To avoid these losses at low speeds we use a traveling wave decelerator. Based on the design of Osterwalder et al [@Osterwalder2010; @Meek2011], this decelerator uses a sequence of ring-shaped electrodes to which a space- and time-varying voltage is applied. In this way the traveling wave decelerator creates a genuine rather than effective 3D trap that moves along the decelerator, co-propagating with the molecules. By reducing the speed of this co-moving trap the molecules are decelerated. The lowest speed reached by Osterwalder et al was 120 m/s, limited by the lowest frequency at which the voltage could be varied. The same decelerator was used to slow YbF molecules from a buffer gas source from 300 m/s to 276 m/s, limited by the deceleration that could be applied to these heavy molecules [@Bulleid:PRA2012]. A 5m long traveling wave decelerator is currently under construction at the KVI in order to decelerate SrF molecules [@Berg:EPJD2012]. We have recently been able to slow molecules to rest, and statically trap them inside the decelerator by using high voltage amplifiers that can sweep the high voltage from 15kHz down to DC. We use a conventional Stark decelerator to slow ammonia molecules from 300 m/s to around 100 m/s, removing 90% of their initial kinetic energy over a length of only 50 cm. We then load the molecules into the traveling wave decelerator for the remaining deceleration to standstill [@QuinteroPerez2013]. In this paper we provide further details about the traveling wave decelerator, and present new experimental data that illustrate the control over molecules that is offered by this decelerator. Experimental setup ================== ![(color online) Schematic view of the vertical molecular beam machine. The inset shows the electric field magnitude (in steps of 2.5kV/cm) inside the traveling wave decelerator calculated using [Simion]{} [@Dahl:RevSciInstrum1990]. The rightmost panel shows the sinusoidal waveform (dashed line) from which the potential that is applied to each electrode (solid circle) is sampled. \[fig:setup\_ring\]](fig01-setup_v2.pdf){width="1\columnwidth"} Figure \[fig:setup\_ring\] shows a schematic view of our vertical molecular beam machine. In brief, a pulsed ($\sim$100$\mu$s) ammonia beam is released into vacuum from a solenoid valve (General Valve series 99) at a 10Hz repetition rate. By cooling the valve housing to typically -50$^{\circ}$C and seeding the ammonia molecules in xenon, the mean velocity of the beam is lowered to 300m/s. The ammonia beam is decelerated using a 101-stage Stark decelerator. Adjacent stages are 5.5mm apart. Each deceleration stage is formed by two parallel 3mm diameter cylindrical rods, spaced 2mm apart. The two opposite rods are switched to $+10$ and $-10$kV by four independent HV switches that are triggered by a programmable delay generator. A traveling wave decelerator is mounted 24 mm above the last electrode pair. This decelerator consists of 336 ring electrodes, each of which is attached to one of eight 8 mm diameter stainless steel rods, resulting in a periodic array in which every ninth ring electrode is attached to the same rod. The rods are placed on a 26mm diameter circle, forming a regular octagon. Each rod is mounted by two ceramic posts that are attached to the octagon via an adjustable aluminum bar that allows for fine-tuning of the alignment. The ring-shaped electrodes are made by bending 0.6 mm thick tantalum wire into the shape of a tennis racket with an inner diameter of 4 mm. Consecutive rings are separated by 1.5 mm (center to center) resulting in a periodic length of $L=12$mm. This combination of parameters is identical to the design of Osterwalder and co-workers [@Osterwalder2010; @Meek2011]. The voltages applied to the eight support bars are generated by amplifying the output of an arbitrary wave generator (Wuntronic DA8150) using eight fast HV-amplifiers (Trek 5/80) up to $\pm$5 kV. A 50cm long quadrupole is mounted 20mm above the traveling wave decelerator. The quadrupole can be used for focusing slow molecules and to provide an extraction field for a Wiley-McLaren type mass spectrometer setup. In the experiments described here, slow molecules are not focused and the quadrupole is used only to create the extraction field. The molecular beam overlaps with the focus of an UV laser 40mm behind the last ring electrode of the decelerator to ionize the ammonia molecules. The nascent ions are counted by an ion detector. The chamber that houses the two decelerators and quadrupole guide is differentially pumped and kept at a pressure below $3\times10^{-8}$mbar when the pulsed valve is operating. Theory\[sec:theory\] ==================== A detailed discussion of the operation principles of the traveling wave decelerator can be found elsewhere [@Osterwalder2010; @Meek2011; @vandeMeerakker:ChemRev2012]. In this section we will summarize those results that are relevant for the current paper.\ In order to describe the operation of a traveling wave decelerator, it is useful to consider the electric field inside an infinitely long hollow conducting cylinder to which a voltage is applied that is periodic in $z$. As shown in Ref. [@vandeMeerakker:ChemRev2012] the electric potential on the beam axis will in this case follow the potential applied to the cylinder, but it is reduced by a factor that depends on the radius of the cylinder and the periodic length of the applied waveform. The electric field magnitude in the longitudinal direction is given by a fully-rectified sine wave, resulting in two minima per period at the positions where the electric potential at the cylinder is maximal. At the position of these minima, the electric field magnitude in the radial direction is given by a first-order modified Bessel function of the first kind. Close to the minima the electric field increases linearly in both directions with the field gradient in the longitudinal direction being twice as large as the gradient in the radial direction. These minima will act as true 3D traps for weak-field seeking molecules. By modulating the waveform that is applied to the cylinder in time, the traps can be moved along the decelerator, while keeping a constant shape and depth. In the actual implementation of the traveling wave decelerator, 8 ring-shaped electrodes are used to sample the infinite cylinder. The inset of Fig. \[fig:setup\_ring\] shows the electric field magnitude inside the traveling wave decelerator at a given time. The voltages applied to successive ring electrodes follow a sinusoidal pattern in $z$ shown on the right-hand side of the figure. In order to move the traps, the sinusoidal waveform is modulated in time. As a consequence of using a finite number of electrodes, the trapping potential no longer maintains a constant shape and depth while it is moved. In the chosen geometry, the electric field gradients in the bottom of the well, as well as the trap depth in the longitudinal direction are nearly independent of the position of the trap minimum. The trap depth in the transverse direction, however, is 40% deeper when the trap minimum is located in the plane of a ring compared to the situation when the trap minimum is located directly between two rings. ![image](./fig02-effective_potential.pdf){width="70.00000%"} The potential $V_n$ applied to the $n$th electrode in the traveling wave decelerator can be expressed as [@Meek2011] $$V_n(t)=V_0\sin\left (-\phi(t)+\frac{2\pi n}{8}\right ), \label{eq:fields}$$ where $V_0$ is the amplitude of the sine-modulated potential and $\phi$ is a time-dependent phase offset that governs the motion of the traps. For the electric field configuration that is shown in Fig. \[fig:setup\_ring\], the phase offset has a value of $\phi=0 \pmod{{\pi}/{4}}$. The angular frequency of the wave is given by the time derivative of $\phi(t)$; $(d\phi/dt)=2\pi f(t)$, where $f(t)$ is the frequency in Hz. Since one oscillation of the waveform moves the trap over one period, the velocity of the trap is given by $v_z(t)=f(t) L$. Integrating the angular frequency with respect to time results in the following expression for the phase $$\phi(t)=\frac{2\pi}{L}\int_0^t\! v_z(\tau)\,d\tau,\quad\text{where }v_z(t)=\int_0^t\!a(\tau)\,d\tau.$$ A linear increase of $\phi$ with time results in a trap that moves with a constant positive velocity along the $z$ axis. Acceleration or deceleration of the trap can be achieved by applying a chirp to the frequency. The acceleration or deceleration of the trap changes the effective longitudinal potential experienced by the molecules. In order to account for this pseudoforce, an additional term of the form $W_{\text{acc}}=maz$, with $m$ the mass of the molecule and $a$ the acceleration along the $z$ axis, is added to the potential. In Fig. \[fig:effectivepotential\], the resulting potentials are shown for different waveforms corresponding to different accelerations as indicated. The panels on the left- and right-hand side show the potential experienced by and molecules, respectively. The lower panels show the effective potential along the $z$-axis. The inversion splitting in NH$_{3}$ is 23.8GHz, while it is only 1.6GHz in ND$_{3}$. As a result, the Stark effect in NH$_{3}$ is quadratic up to electric fields of 20kV/cm and the effective potential for NH$_{3}$ molecules is almost perfectly harmonic. The Stark effect in ND$_{3}$, on the other hand, becomes linear at much smaller fields and the effective potential for ND$_{3}$ molecules is harmonic close to the trap center only. Note that in the figure the accelerations applied to are about 3 times larger than those applied to . As seen from the figure, acceleration of the trap reduces its phase-space acceptance and thus decreases the number of molecules that can be confined. Above a certain threshold value, the potential no longer contains a minimum and no molecules can be trapped at all. In addition to reducing the trap depth, acceleration of the trap results in a shift of the effective field minimum. This shift needs to be taken into account when the acceleration is not constant but a function of time, for instance, when the molecules are decelerated to a standstill and subsequently trapped at a certain position in the decelerator. We do this by rapidly ($<20$$\mu$s) sweeping the phase of the waveform. We will refer to this sweep as a “phase jump”. Results ======= In this section we present data that demonstrate the versatility of the traveling wave decelerator and analyze the performance of the setup under different conditions. Results for guiding, deceleration and trapping of and molecules are presented in Section \[sec:guidingdectrap\]. Section \[sec:phasejumps\] discusses the use of the phase jump for and . ![(color online) Measured time-of-flight profiles for (upper panel) and (lower panel) molecules guided at 90m/s (violet curve), decelerated to 60, 30 and 0m/s (blue, green, and yellow curves respectively) and trapped for 50ms (red curve) before being accelerated back to 90m/s and detected. Each data point is the average of 90 laser shots. The time of flight traces have been centered around the expected arrival time. The inset shows the velocity of the traveling wave potential as a function of time for the different recorded TOF profiles. \[fig:tofs\]](fig03-tofs.pdf){width="1\columnwidth"} Guiding, deceleration and trapping\[sec:guidingdectrap\] -------------------------------------------------------- In Fig. \[fig:tofs\], time-of-flight (TOF) profiles of molecules (upper panel) and molecules (lower panel) are shown that were obtained by applying different waveforms to the traveling wave decelerator, as shown in the inset. Note that the horizontal axis is shifted in such a way that the origin corresponds to the expected arrival time of the molecules. For each measurement, a packet of molecules is decelerated to 90m/s using the conventional Stark decelerator and subsequently coupled into the traveling wave decelerator. The violet curves in Fig. \[fig:tofs\] are TOF profiles for packets of and molecules that are guided through the traveling wave decelerator at the injection speed of 90m/s by applying a waveform with a constant frequency of 7.5kHz to the array of electrodes. The width of the TOF profile mainly reflects the velocity spread of the guided molecules. For the moving trap is deeper than for , and its TOF profile is accordingly wider. Wings are observed at earlier and later arrival times, which are attributed to molecules that are trapped in the electric-field minima that are located $12$mm above or below the synchronous one. By adjusting the waveform that is applied to the traveling wave decelerator, the velocity of the molecules can be manipulated almost at will. The blue, green, and yellow traces in Fig. \[fig:tofs\] are time-of-flight profiles for and molecules that are decelerated from 90m/s to 60, 10 and 0m/s and immediately reaccelerated to 90m/s. This corresponds to chirping the applied waveform from 7.5kHz to 5.0, 0.8, and 0kHz and results in accelerations of $\pm$9.2, $\pm$16.4, and $\pm 16.6\times 10^{3}$m/s$^2$, respectively. The red curves in Fig. \[fig:tofs\] show TOF profiles that are obtained under almost identical conditions as the yellow curves. However, after the velocity of the applied waveform is decreased to 0m/s, the voltages are kept at a constant values for 50ms before the velocity is increased to its original value of 90m/s. It can be seen in the figure that the observed TOFs are indeed almost identical to the ones recorded when the velocities are immediately returned to their original value. This measurement demonstrates that molecules can be trapped in the laboratory frame without further losses. The observed decrease in signal for NH$_{3}$ and ND$_{3}$ at higher accelerations is greater than expected from simulations. We attribute the loss mainly to mechanical misalignments that lead to parametric amplification of the motion of the trapped molecules at low velocities. On inspection, it was noticed that one of the suspension bars was slightly displaced from its original position, which must have happened when the decelerator was placed in the vacuum chamber. Another loss mechanism comes from the fact that the phase space distribution of the packet exiting the conventional Stark decelerator is not perfectly matched to the acceptance of the traveling wave decelerator, and the alignment of the axis of the traveling wave decelerator to the axis of the Stark decelerator is not perfect. As a result, the trapped packet as a whole will perform a (damped) breathing and sloshing motion. These oscillations explain why the observed TOF profile for deceleration to 60m/s is wider and more intense than the TOF profile for guided molecules. Neither loss mechanism is fundamental, and we believe they can be eliminated in future work. ![(color online) Integrated time-of-flight distributions for (squares) and (diamonds) shown as a function of acceleration. The bold lines correspond to simulations of the 1D acceptance of the traveling wave decelerator, shown on the vertical axis on the left-hand side, while the thin lines correspond to simulations of the 3D acceptance of the traveling wave, shown on the vertical axis on the right-hand side. The and signals are scaled (by a different factor) to match the simulation. The inset displays the velocity of the traveling wave potential for three typical accelerations used in the experiment. \[fig:accvssignal\]](fig04-acceptance.pdf){width="1\columnwidth"} As discussed in Sec. \[sec:theory\], the acceptance of the traveling wave decelerator depends on the chirp that is applied to the waveform. This is illustrated by the measurements that are shown in Fig. \[fig:accvssignal\]. In this figure, the time-integrated signal of (squares) and (diamonds) molecules is plotted as a function of the applied acceleration. In these measurements, ammonia molecules are guided at 90m/s for 5mm before being accelerated to 130m/s and guided for the remaining length of the decelerator. For comparison, the bold and thin solid lines shown in Fig. \[fig:accvssignal\] correspond to simulations of the 1D and 3D acceptance of the traveling wave decelerator, respectively. The calculated acceptances are based on an average (i.e. static in the moving frame of the trap) potential well, rather than the true potential. The numbers on the right axis refer to the 3D acceptance, while the numbers on the left-hand axis refer to the 1D acceptance. The fact that the 1D and 3D simulations are almost the same, apart from a scaling factor, illustrates that the transverse acceptance is independent of the acceleration, i.e., the transverse motion of the molecules in the traveling wave decelerator is largely decoupled from the longitudinal motion. The measurements for and have been scaled to match the simulations. As observed, at high acceleration the simulations predict the measurements quite well. At lower accelerations, however, the measurements are seen to reach a constant value which is consistent with the expected longitudinal acceptance of the conventional Stark decelerator at the used phase angle of 65 degrees [@Bethlem:PRA2002]. At these accelerations the longitudinal acceptance of the traveling wave decelerator is larger than the longitudinal emittance of the packet exiting the Stark decelerator; i.e., all molecules exiting the conventional Stark decelerator are trapped in the traveling wave decelerator. Consequently, decelerations below roughly $18\times 10^3$ m/s$^{2}$ do not reduce the detected signal of and molecules. The dashed vertical line also shown in Fig. \[fig:accvssignal\] indicates the maximum acceleration that was used in the measurements shown in Fig. \[fig:tofs\]. Note that, as in our measurements we do not determine an absolute value for the phase space acceptance of the decelerator, the similarity between the measurements and simulations should be considered as a consistency check only. ![(color online) (squares) and (diamonds) signal as a function of the position in the decelerator (expressed as a phase, $\phi_0$) at which the molecules were trapped. When $\phi_0=0\pmod{{\pi}/{4}}$, the center of the trap is located in the plane of a ring electrode while for $\phi_0=\pi/8 \pmod{{\pi}/{4}}$ the trap minimum is located in the plane between two ring electrodes. The solid curves represent the trap depth for both species and are specified on the right axis. The and signals are scaled (by different factors) to match the trap depth. \[fig:positionscan\]](fig05-trapposition.pdf){width="1\columnwidth"} The position in the array of ring-shaped electrodes at which the molecules are brought to a standstill and statically trapped, is determined by the waveform and can be chosen at will. Figure \[fig:positionscan\] shows the (squares) and (diamonds) signal as a function of this position. The signal oscillates with a periodicity that corresponds to the spacing between adjacent rings; 1.5mm equivalent to a phase difference of $\pi/4$. The signal is largest when the ammonia molecules are trapped in the plane of an electrode, and smallest when the trap center is located between two electrodes. The observed modulation is related to the variation in the radial confinement discussed in Sec. \[sec:theory\]. For comparison, the solid lines in Fig. \[fig:positionscan\] show the trap depth for and in cm$^{-1}$ (1 cm$^{-1}$ corresponds to 1.4K). The and signals are scaled (by different factors) to match the trap depth. Phase jumps\[sec:phasejumps\] ----------------------------- As explained in Sec. \[sec:theory\], whenever the acceleration is changed, we rapidly sweep the phase of the waveform in order to correct for the resulting shift of the effective potential minimum. We refer to this sweep as a “phase jump”. In Fig. \[fig:phasejump\] we present measurements of molecules that are decelerated from 70 to 0m/s using an acceleration of -15000m/s$^2$, trapped for 30ms, and subsequently accelerated back to 70m/s with the opposite acceleration and detected. The upper panel of Fig. \[fig:phasejump\] shows the (squares) and (diamonds) signal as a function of the (magnitude of the) phase jumps that are applied whenever the acceleration is changed. As expected, the required phase jump for optimal signal is much larger for than for . In fact, when no phase jump is applied almost no signal is observed for , while for the signal hardly changes. The solid curves in the upper panel of Fig. \[fig:phasejump\] show the result of 1D numerical simulations. Although the simulations seem to slightly overestimate the width of the observed peaks and predict slightly greater phase jumps, the maximum and overall shape match the experimental data fairly well. The discrepancies might be caused by the fact that the amplifiers are not able to follow the rapid sweep of the phase perfectly. The lower panel of Fig. \[fig:phasejump\] shows the longitudinal acceleration as a function of the position along the beam axis for and molecules using a waveform amplitude of 5kV. The electric field increases linearly away from the center of the trap resulting in a harmonic but shallow potential for and a tight but very anharmonic potential for . The horizontal line shows the acceleration that is used in the experiment. The crossings of this line with the acceleration curves, indicated by the dashed vertical lines, correspond to the positions of the minima of the effective potential. ![(color online) The effect of phase jumps for an acceleration of -15000m/s$^2$ are illustrated in the upper panel for (squares) and (diamonds). Solid curves show the result of a 1D numerical simulation. The lower panel shows the longitudinal acceleration along the beam axis for and molecules using a waveform amplitude of 5kV. The dashed vertical lines indicate the phase jumps that are derived from the acceleration curves in the lower panel. \[fig:phasejump\]](fig06-phasejump.pdf){width="1\columnwidth"} Adiabatic cooling\[sec:cooling\] -------------------------------- As the voltages applied to the decelerator are generated by amplifying the output of an arbitrary wave generator using fast HV-amplifiers, we can change the depth (and shape) of the trap at will. This is illustrated by the measurements shown in Fig. \[fig:cooling\]. In these measurements, molecules were decelerated using the conventional Stark decelerator at phase angles of 60 and 53 degrees and coupled into the traveling wave decelerator. The traveling wave decelerator is subsequently used to slow the molecules to a standstill, trap them for a period of over 50ms, accelerate them back to their initial velocity and launch them into the detection region. While the molecules are trapped, the voltages applied to the decelerator are ramped down in 2ms, kept at a lower value for 10ms and then ramped up again to 5kV. Figure \[fig:cooling\] shows the remaining signal as a function of the reduced amplitude of the waveform for the two phase angles as indicated. Typical waveform amplitudes as a function of time are shown in the inset of Fig. \[fig:cooling\]. Lowering the voltages of the trap has two effects: (i) the trap frequency is lowered, adiabatically cooling the molecules; (ii) the trap depth is reduced, allowing the hottest molecules to escape the trap. The observed signal at a phase angle of 53 degrees is seen to drop more rapidly than that at a phase angle of 60 degrees. This is expected, as at lower phase angles a larger and hotter packet is loaded into the traveling wave decelerator. The hotter ensemble of molecules initially fills the trap almost completely while the colder molecules occupy only a fraction of the trap. Consequently, the number of molecules in the hotter ensemble is reduced even when the voltages are lowered by a small amount. The solid and dashed lines also shown in Fig. \[fig:cooling\] result from a simulation that assumes an (initial) temperature of 30mK or 100mK, respectively. Note that we use temperature here only as a convenient means to describe the distribution; the densities are too low to have thermalization on the timescales of the experiment. For comparison, Fig. \[fig:cooling\] also shows measurements where the trap voltages are abruptly (10$\mu$s) lowered. In this case no adiabatic cooling occurs and as a result the signal decreases more rapidly than when the voltages are ramped to lower voltages more slowly. The solid (30mK) and dashed (100mK) lines labeled as nonadiabatic show the results of a numerical simulation when the voltages are reduced abruptly. Similar measurements have been performed for (not shown). ![(color online) Signal of trapped molecules decelerated using a phase angle of 60 and 53 degrees as a function of the amplitude of the waveform when, after deceleration, the trap is slowly (labeled as adiabatic) or abruptly (labeled as nonadiabatic) reduced. The smooth curves also shown in the figure are simulations that assume an initial temperature of 30 (solid line) or 100mK (dashed line) for the ensemble of trapped molecules. The inset shows the amplitude of a number of typical waveforms as a function of time. \[fig:cooling\]](fig07-cooling.pdf){width="1\columnwidth"} In Fig. \[fig:tramp\], the number of molecules that remain trapped at 2kV is shown as a function of the time used for lowering and increasing the voltages. The solid lines result from a 3D numerical simulation. The measurements confirm that the and molecules follow the trap adiabatically when the ramping times are longer than 1ms. The time required for the molecules to follow the change in potential adiabatically is greater for than it is for due to the lower trap frequencies experienced by the former. Similar measurements have been performed on molecules in a microstructured trap by Englert et. al [@Englert2011]. ![(color online) Signal due to (squares) and (diamonds) molecules that remain trapped at 2kV as a function of the time used for ramping the voltages. The solid lines show the result of a 3D numerical simulation that assumes a temperature of 30 or 100mK for and , respectively. In the inset, typical waveform amplitudes are shown as a function of time. \[fig:tramp\]](fig08-tramp.pdf){width="1\columnwidth"} Conclusions =========== In this paper we have demonstrated deceleration and trapping of and molecules in a traveling wave decelerator. The observed trapping times are limited to 100ms by the current repetition rate of the experiment (10Hz). The deceleration of a supersonic beam of ammonia molecules is performed in two steps; in the first step, the molecules are decelerated from 300 to 100m/s using a conventional Stark decelerator, while in the second step, the traveling wave decelerator is used to remove the remaining kinetic energy from the molecules. The advantages of such a system are that the requirements imposed on the electronics of the traveling wave decelerator remain rather low, and that the combined length of both decelerators is only slightly above 1m. The motion of the molecules in the traveling wave decelerator is controlled completely by computer generated waveforms. As an example of the possibilities offered by this control, we have adiabatically cooled the trapped molecules by lowering the amplitude of the waveforms. In previous work, we demonstrated that the motion of the molecules in the trap can be excited by applying an small oscillatory force [@QuinteroPerez2013]. The ability to control the voltages applied to the electrodes also offers the possibility to tailor the shape of the trap – for instance, changing it into a more box-like potential. This may prove useful for collisional and spectroscopic studies, as well as for the implementation of schemes to further cool the molecules such as sisyphus cooling [@Zeppenfeld:Nature2012] and evaporative cooling [@Stuhl:Nature2012]. This research has been supported by NWO via a VIDI-grant, by the ERC via a Starting Grant and by the FOM-program ‘Broken Mirrors & Drifting Constants’. We acknowledge the expert technical assistance of Rob Kortekaas, Jacques Bouma, Joost Buijs, Leo Huisman and Imko Smid. We thank Andreas Osterwalder and Gerard Meijer for helpful discussions and Wim Ubachs for his continuing interest and support. [29]{}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 [**, ()](http://stacks.iop.org/1367-2630/11/i=5/a=055049) [, ()](\doibase 10.1039/C1CP21733J) [, ()](\doibase 10.1021/cr200349r) [, ()](\doibase 10.1021/cr2004597) [, ()](\doibase 10.1126/science.1184121) [, ()](\doibase 10.1126/science.1229549) [, ()](\doibase 10.1103/PhysRevLett.88.067901) [, ()](\doibase 10.1038/nphys287) [, ()](\doibase 10.1038/nphys386) [, ()](\doibase 10.1038/nature10104) @noop [“,” ]{} (), [, ()](http://dx.doi.org/10.1140/epjd/e2004-00160-9) [, ()](\doibase 10.1103/PhysRevLett.96.143004) [, ()](\doibase 10.1103/PhysRevLett.100.150801) @noop [“,” ]{} (), [, ()](\doibase 10.1140/epjst/e2008-00809-5) [, ()](\doibase 10.1103/PhysRevLett.83.1558) [, ()](\doibase 10.1103/PhysRevA.65.053416) @noop [Ph.D. thesis]{}, () @noop [Ph.D. thesis]{}, () [, ()](\doibase 10.1103/PhysRevA.81.051401) [, ()](\doibase 10.1063/1.3640413) [, ()](\doibase 10.1103/PhysRevA.86.021404) [, ()](\doibase 10.1140/epjd/e2012-30017-5) [, ()](\doibase 10.1103/PhysRevLett.110.133003) [, ()](\doibase 10.1063/1.1141932) [, ()](\doibase 10.1103/PhysRevLett.107.263003) [, ()](\doibase 10.1038/nature11595) [**, ()](\doibase http://dx.doi.org/10.1038/nature11718)	{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that every measurable invariant conformal structure for a Hölder continuous linear cocycle over a subshift of finite type coincides almost everywhere with a continuous invariant conformal structure. We use this result to establish Hölder continuity of a measurable conjugacy between Hölder continuous cocycles where one of the cocycles is assumed to be uniformly quasiconformal. As a special case we derive that if a Hölder linear cocycle is a measurable coboundary, then it is a Hölder coboundary. We also use the main theorem to show that a linear cocycle is conformal if none of its iterates preserve a measurable family of proper subspaces of $\mathbb{R}^{d}$. We use this to characterize closed negatively curved Riemannian manifolds of constant negative curvature by irreducibility of the action of the geodesic flow on the unstable bundle.' author: - Clark Butler bibliography: - 'measurableconformalrigidity.bib' title: Continuity of measurable invariant conformal structures for linear cocycles over hyperbolic systems --- Introduction ============ This work is motivated by a pair of questions, one of which is sourced from the geometry of negatively curved manifolds and the other from the study of the cohomology of linear cocycles over hyperbolic systems. We will first explain how the central difficulty of these two motivating questions can be viewed as two different manifestations of the same problem regarding continuity of measurable invariant conformal structures for linear cocycles over hyperbolic systems. The geometric question was originally posed by Sullivan [@Su], who conjectured that if the action of the geodesic flow of a closed negatively curved Riemannian manifold $M$ (with $\dim M \geq 3$) on the tangent spaces to expanding horospheres is measurably irreducible and the sectional curvatures of $M$ satisfy the $1/4$-pinching condition $-\frac{1}{4} > K \geq -1$ then $M$ has constant negative curvature. We let $E^{u}$ denote the unstable bundle of the geodesic flow $g^{t}$ on $T^{1}M$ which is tangent to the expanding horospheres of $M$ and we let $m$ denote the invariant Liouville volume for the geodesic flow. Measurable irreducibility in this context then means that there is no measurable family of proper subspaces of $E^{u}$ which are invariant $m$-a.e. under the derivative action of the geodesic flow. This conjecture was confirmed by combining the work of Kanai [@Kan1] and Besson-Courtois-Gallot [@BCG1], as noted by Yue [@Yue]. The $1/4$-pinching assumption on the sectional curvatures of $M$ appears superfluous at first glance in the hypotheses of this theorem, as one does not require $1/4$-pinching to formulate the notion of measurable irreducibility of the geodesic flow on expanding horospheres, and the conclusion of the theorem shows that the curvature of the manifold is actually constant. This was also observed by Yue, who proposed a proof that the measurable irreducibility hypothesis alone suffices to conclude that the curvature of $M$ is constant [@Yue]. However there appears to be a crucial gap in this proof which is discussed in detail in [@Bu1 Remark 3.2]. In our first theorem we close this gap and remove the $1/4$-pinching hypothesis, \[hyperbolic rigidity\] Let $g^{t}$ be the geodesic flow of a closed negatively curved Riemannian manifold $M$ with $\dim M \geq 3$. If $Dg^{t}\|_{E^{u}}$ is measurably irreducible then $M$ has constant negative curvature. Our second motivating question comes from the study of the cohomology of linear cocycles over hyperbolic systems. We will take subshifts of finite type as our models for hyperbolic systems. For this we recall some standard definitions. Let ${\ell}\geq 2$ be a fixed positive integer, let $Q = (q_{ij})_{1 \leq i,j \leq r }$ be an ${\ell}\times {\ell}$ matrix with $q_{ij} \in \{0,1\}$ and let $\Sigma$ be the subshift of finite type associated to the matrix $Q$, $$\Sigma = \{(x_{n})_{n \in \mathbb{Z}}: q_{x_{n}x_{n+1}} = 1 \;\text{for all $n \in \mathbb{Z}$}\}.$$ We let $f: \Sigma\circlearrowleft$ be the left shift map defined by $f(x_{n})_{n \in \mathbb{Z}} = (x_{n+1})_{n \in \mathbb{Z}}$. We equip $\Sigma$ with the family of metrics, $$\label{metric def} \rho_{\tau}(x,y)=e^{-\tau N(x,y)},\; \textrm{where} \; N(x,y)=\max \lbrace N\geq 0; x_n=y_n \; \textrm{for all} \;\|n\|<N \rbrace,$$ where $\tau \in (0,\infty)$. Let $d \geq 1$ be a fixed positive integer. \[defn: cocycle\] Let $A: \Sigma \rightarrow GL(d,{\mathbb R})$ be a measurable map. The linear cocycle over $f$ generated by $A$ is the map $$\mathcal{A}: \Sigma \times {\mathbb Z}\rightarrow GL(d,{\mathbb R}),$$ $$\mathcal{A}(x,n) = \left\{ \begin{array}{ll} A(f^{n-1}(x))\ldots A(f(x))A(x) & \mbox{if } n>0 \\ I & \mbox{if } n=0 \\ A(f^{n}(x))^{-1}\ldots A(f^{-1}(x))^{-1}& \mbox{if } n<0. \\ \end{array} \right.$$ We will write $A^{n}(x):=\mathcal{A}(x,n)$. We say that $\mathcal{A}$ is a continuous linear cocycle if the generator $A$ is continuous; we say that $\mathcal{A}$ is $\alpha$-Hölder continuous if the generator $A$ is $\alpha$-Hölder continuous with respect to the metric $\rho_{\tau}$ for a fixed choice of $\tau$. Note that it then follows that $A$ is Hölder continuous with respect to $\rho_{\tau'}$ for any other $\tau' > 0$, with a possibly different Hölder exponent. Lastly, let $\mu$ be a fixed ergodic, fully supported $f$-invariant probability measure on $\Sigma$ with local product structure. Local product structure will be defined in Section \[shift definitions\]; natural examples of measures with local product structure are equilibrium states of Hölder potentials [@Bow], such as the measure of maximal entropy. \[defn: cohomologous\] Two linear cocycles $A$ and $B$ over $f$ are (measurably, continuously) cohomologous if there is a (measurable, continuous) function $P: \Sigma \rightarrow GL(d,{\mathbb R})$ such that $$A(x) = P(f(x))B(x)P(x)^{-1},$$ for $\mu$-a.e. $x \in \Sigma$ in the measurable case, and every $x \in \Sigma$ in the continuous case. Our second main theorem shows that if a Hölder continuous linear cocycle $\mathcal{A}$ over $f$ is measurably cohomologous to the identity cocycle $B(x) \equiv Id_{{\mathbb R}^{d}}$ then $A$ is continuously cohomologous to the identity. \[coboundary\] Let $\mathcal{A}$ be an $\alpha$-Hölder continuous cocycle over $f$. Suppose that there is a measurable function $P : \Sigma \rightarrow GL(d,{\mathbb R})$ such that $$A(x) = P(f(x))P(x)^{-1}, \; \; \text{for $\mu$-a.e. $x \in \Sigma$.}$$ Then $P$ coincides $\mu$-a.e. with an $\alpha$-Hölder continuous function $\hat{P}$ satisfying the same equation for every $x \in \Sigma$. We pause to give some background on previous work relating to Theorem \[coboundary\]. There are two principal questions regarding the cohomology of linear cocycles which are still not completely understood. For both questions below we take $A$ and $B$ to be $\alpha$-Hölder continuous for some $\alpha > 0$. 1. Suppose that $A$ and $B$ have conjugate periodic data in the sense that there is a continuous map $C: \Sigma \rightarrow GL(d,{\mathbb R})$ such that $A^{m}(p) = C(p)B^{m}(p)C(p)^{-1}$ for every periodic point $p$ of $f$ of period $m$. Are $A$ and $B$ continuously cohomologous? 2. Suppose that $A$ and $B$ are measurably cohomologous. Are they then continuously cohomologous? The answers to these two questions are “yes" in the case where $d = 1$, i.e., when $A$ and $B$ can be viewed as additive real-valued cocycles. This is due to the pioneering work of Livsic [@Liv]. When $A$ and $B$ take values in a non-abelian group, the answers to these questions become considerably more subtle. In the case of Question (1) when $B(x) = Id_{{\mathbb R}^{d}}$ for every $x \in \Sigma$ a positive answer was obtained by Kalinin [@Kal]; we refer to that paper for a survey of previous work on that subject. For more general $B$ a positive answer to Question (1) has also been obtained simultaneously by Backes [@B15] and Sadovskaya [@S15] under the additional assumption that $B$ is fiber bunched (see Definition \[fiber\] below). The fiber bunching condition holds in an open neighborhood of the identity cocycle. Question (1) remains open in its full generality. As for Question (2), Sadovskaya showed in the case that $A$ is fiber bunched and $B \equiv Id_{{\mathbb R}^{d}}$ that if $A$ and $B$ are measurably cohomologous then they are continuously cohomologous [@S15]. Our Theorem \[coboundary\] improves on her result by removing the fiber bunching condition on $A$. There are examples of Hölder continuous cocycles $A$ and $B$ arbitrarily close to the identity which are measurably cohomologous but not continuously cohomologous that have been constructed by Pollicott and Walkden [@PW01] (see also [@S15]). It’s thus unclear whether a general theorem is possible in the case of Question (2). We now introduce the main theorem of this paper from which both Theorems \[hyperbolic rigidity\] and \[coboundary\] will be deduced. A conformal structure on ${\mathbb R}^{d}$ is an equivalence class of inner products on ${\mathbb R}^{d}$, where two inner products $\langle , \rangle$ and $(,)$ are considered equivalent if there is a positive real number ${\lambda}> 0$ such that $\langle v,w \rangle = {\lambda}(v,w)$ for every $v,w \in {\mathbb R}^{d}$. We denote the space of conformal structures on ${\mathbb R}^{d}$ by $\mathcal{C}^{d}$. The space $\mathcal{C}^{d}$ can be naturally identified with the Riemannian symmetric space $SL(d,{\mathbb R})/SO(d,{\mathbb R})$ of real positive definite symmetric matrices of determinant 1 which carries a left-invariant nonpositively curved Riemannian metric $g$. $GL(d,{\mathbb R})$ acts transitively on $\mathcal{C}^{d}$ via the action $$B_{}\eta = \det(B B^{T})^{-\frac{2}{d}} B^{T} \eta B$$ for $B \in GL(d,{\mathbb R})$ and a positive definite matrix $\eta \in SL(d,{\mathbb R})/SO(d,{\mathbb R})$, where $B^{T}$ denotes the transpose of $B$. This action is an isometry with respect to the metric $g$. For more details and proofs related to these facts, see [@KS10]. For a matrix $B \in GL(d,{\mathbb R})$ and a positive definite matrix $\eta \in SL(d,{\mathbb R})/SO(d,{\mathbb R})$ we define $$B[\eta] = B^{-1}_{}\eta$$ We let $\langle,\rangle$ denote the standard Euclidean inner product on ${\mathbb R}^{d}$ and $\\|\cdot \\|$ the associated norm. \[defn: preserve conformal\] Let $\mathcal{A}$ be a linear cocycle over $f$. We say that $\mathcal{A}$ preserves a measurable conformal structure if there is a measurable map $\eta : \Sigma \rightarrow \mathcal{C}^{d}$ defined $\mu$-a.e. such that $$A(x)[\eta_{x}] = \eta_{f(x)},$$ for $\mu$-a.e. $x \in \Sigma$. We say that $\mathcal{A}$ preserves a continuous conformal structure if $\eta$ is continuous and the above equality holds for every $x \in \Sigma$. We can now state the main theorem of this paper, \[conformalrigidity\] Let $\mathcal{A}$ be an $\alpha$-Hölder continuous linear cocycle over $f$. If $\mathcal{A}$ preserves a measurable conformal structure $\eta: \Sigma \rightarrow \mathcal{C}^{d}$ then $\eta$ coincides $\mu$-a.e. with an $\alpha$-Hölder continuous conformal structure $\hat{\eta}: \Sigma \rightarrow \mathcal{C}^{d}$ which is invariant under $\mathcal{A}$. To discuss previous results related to this theorem we introduce the notion of fiber bunching for a linear cocycle, \[fiber\] A linear cocycle $\mathcal{A}$ over $f: \Sigma \rightarrow \Sigma$ is fiber bunched if, for some choice of $\tau > 0$, $\mathcal{A}$ is $\alpha$-Hölder continuous with respect to the metric $\rho_{\tau}$ (as defined in ) and there are constants $0 < \xi < 1$ and $L > 0$ such that $$\\|A^{n}(x)\\|\\|A^{n}(x)^{-1}\\| e^{-\|n\|\alpha \tau} < L\xi^{\|n\|}$$ for every $n \in {\mathbb Z}$. The quantity $\\|A^{n}(x)\\|\\|A^{n}(x)^{-1}\\|$ measures the degree to which $A^{n}$ fails to be conformal with respect to the Euclidean inner product on ${\mathbb R}^{d}$. The fiber bunching assumption implies that the growth rate of this failure of conformality is uniformly dominated by the expansion/contraction rates of the base system. Fiber bunching is a common assumption among recent theorems proven about linear cocycles over hyperbolic systems: these include continuity of Lyapunov exponents [@BBB], [@BV], simplicity of Lyapunov exponents [@BV04], cohomology of Hölder continuous cocycles [@B15], [@S15], and characterizations of linear cocycles with vanishing Lyapunov exponents [@AV10]. There are examples due to Bocker-Viana and the author [@BV], [@Bu2], illustrating the necessity of the fiber bunching condition for unconditional statements regarding continuity of the Lyapunov exponents, but beyond these examples very little is known about the necessity of the fiber bunching hypothesis for these theorems. Without the fiber bunching hypothesis the conclusion of Theorem \[conformalrigidity\] was also previously known to hold under additional boundedness hypotheses on the $\mathcal{A}$-invariant conformal structure $\eta$. The best previous result in this direction is due to Sadovskaya and de la Llave [@LS], who showed that if the function $x \rightarrow g(\eta_{x}, \langle,\rangle)$ is in $L^{p}(\mu)$ for $p$ sufficiently large then $\eta$ coincides $\mu$-a.e. with a continuous $\mathcal{A}$-invariant conformal structure. We remark that all of the results of this paper apply equally well to Hölder continuous cocycles over transitive Anosov diffeomorphisms or Anosov flows where the base measure $\mu$ is taken to be the equilibrium state of a Hölder continuous potential. For transitive Anosov diffeomorphism the reduction to a Hölder continuous cocycle over a subshift of finite type via a Markov partition follows standard techniques. For transitive Anosov flows the reductions to the results of this paper are less trivial but still straightforward. The reductions in that case can be done via the techniques used in the proof of Theorem \[hyperbolic rigidity\] below. We thank Amie Wilkinson for many helpful discussions regarding the contents of this paper. We also thank the anonymous referee for numerous useful suggestions which greatly improved the quality of the paper. Reduction to Theorem \[conformalrigidity\] {#reduction} ========================================== In this section we will show how Theorem \[conformalrigidity\] can be used to obtain some corollaries of independent interest which we can use to prove Theorems \[hyperbolic rigidity\] and \[coboundary\]. We begin by using Theorem \[conformalrigidity\] to remove the fiber bunching hypothesis from one of the main theorems of a recent work of Sadovskaya[@S15]. Following Sadovskaya, a continuous linear cocycle $\mathcal{A}$ with generator $A$ over $f$ is uniformly quasiconformal if there is a constant $L > 0$ such that for every $x \in \Sigma$, $$\\|A^{n}(x)\\| \\|A^{n}(x)^{-1}\\| \leq L \;\; \text{for every $n \in {\mathbb Z}$.}.$$ If $\mathcal{A}$ preserves a continuous invariant conformal structure $\eta$ then $\mathcal{A}$ is uniformly quasiconformal: since $\eta$ is continuous and $\Sigma$ is compact there is a constant $C$ such that for all $v \in {\mathbb R}^{d}$ and $x \in \Sigma$, $$C^{-1}\\|v\\| \leq \sqrt{\eta_{x}(v,v)} \leq C\\|v\\|.$$ Since $\eta$ is invariant there is a continuous function $\psi: \Sigma \rightarrow {\mathbb R}$ such that for every $v \in {\mathbb R}^{d}$ $$\sqrt{\eta_{x}(A(x)v,A(x)v)} = \psi(x)\sqrt{\eta_{x}(v,v)}.$$ Set $\psi^{n}(x) = \prod_{i=0}^{n-1} \psi(f^{i}(x))$. Putting these inequalities together we compute that $$\\|A^{n}(x)v\\| \\|A^{n}(x)^{-1}v\\| \leq C^{2} \psi^{n}(x) \cdot (\psi^{n}(x))^{-1} \cdot \eta_{x}(v,v) \leq C^{4}\\|v\\|^{2}$$ Hence $\mathcal{A}$ is uniformly quasiconformal with constant $L = C^{4}$. The reverse implication that a uniformly quasiconformal cocycle preserves a continuous conformal structure also holds when $\mathcal{A}$ is Hölder continuous [@KS10]. \[Livsicmeasurable\] Let $\mathcal{A}$ and $\mathcal{B}$ be $\alpha$-Hölder continuous linear cocycles over $f$ such that $\mathcal{B}$ preserves a measurable conformal structure. Suppose that there is a measurable function $P : \Sigma \rightarrow GL(d,{\mathbb R})$ such that $$A(x) = P(f(x))B(x)P(x)^{-1}, \; \; \text{for $\mu$-a.e. $x \in \Sigma$.}$$ Then both $\mathcal{A}$ and $\mathcal{B}$ preserve an $\alpha$-Hölder continuous conformal structure and $P$ coincides $\mu$-a.e. with an $\alpha$-Hölder continuous function. By assumption $\mathcal{B}$ preserves a measurable conformal structure $\eta$. The cocycle $\mathcal{A}$ then preserves the measurable conformal structure $P[\eta]$. By Theorem \[conformalrigidity\], both $\eta$ and $P[\eta]$ coincide $\mu$-a.e. with a $\alpha$-Hölder continuous conformal structures $\widehat{\eta}$ and $\widehat{P[\eta]}$ which are $\mathcal{B}$-invariant and $\mathcal{A}$-invariant respectively. It follows from the remarks above that both $\mathcal{A}$ and $\mathcal{B}$ are uniformly quasiconformal and thus satisfy the hypotheses of Sadovskaya’s theorem [@S15 Theorem 2.7] from which we conclude that $P$ coincides $\mu$-a.e. with an $\alpha$-Hölder continuous function. Theorem \[coboundary\] follows from Corollary \[Livsicmeasurable\] by taking $B \equiv Id_{{\mathbb R}^{d}}$. Before proving Theorem \[hyperbolic rigidity\], we obtain a corollary of Theorem \[conformalrigidity\] for linear cocycles over a subshift of finite type which will be needed in the proof. We say that $\mathcal{A}$ is measurably reducible (with respect to the measure $\mu$) if there is some $1 \leq r \leq d-1$ and some $n \geq 1$ such that there is a measurable family of $r$-dimensional linear subspaces $E_{x} \subset {\mathbb R}^{d}$ parametrized by $x \in \Sigma$ such that $A^{n}(x)(E_{x}) = E_{f^{n}(x)}$ for $\mu$-a.e. $x \in \Sigma$. We say that $\mathcal{A}$ is measurably irreducible if it is not measurably reducible. Our next corollary gives a measure-theoretic criterion for a Hölder continuous cocycle to preserve a continuous conformal structure. \[irreducible\] Let $\mathcal{A}$ be an $\alpha$-Hölder linear cocycle over $f$. Suppose that $\mathcal{A}$ is measurably irreducible. Then $\mathcal{A}$ preserves an $\alpha$-Hölder continuous conformal structure. By Zimmer’s amenable reduction theorem [@Zim] there is a measurable map $P: \Sigma \rightarrow GL(d,{\mathbb R})$ defined $\mu$-a.e. such that $B(x) := P(f(x))^{-1} A(x) P(x)$ takes values in a subgroup $G$ of $GL(d,{\mathbb R})$ which contains an amenable subgroup $H$ of finite index. There is then an $n \geq 1$ such that $B^{n}(x) \in H$ for $\mu$-a.e. $x \in \Sigma$. According to the classification of maximal amenable subgroups of $GL(d,{\mathbb R})$, either $H$ stabilizes a proper subspace $\{0\} \subsetneq V \subsetneq {\mathbb R}^{d}$ or $H$ is a subgroup of the conformal linear automorphisms ${\mathbb R}^{\times} \times SO(d,{\mathbb R})$ (see [@Moo] for this classification and [@KS Theorem 3.4] for a more detailed discussion of this implication). In the first case we conclude that $A^{n}$ leaves $\mu$-a.e. invariant the measurable family of subspaces $E_{x} = P(x)(V)$ and thus $A$ is measurably reducible. In the second case $A$ preserves the measurable conformal structure induced by $P[\langle,\rangle]$, where we recall that $\langle, \rangle$ denotes the standard Euclidean inner product on ${\mathbb R}^{d}$. By Theorem \[conformalrigidity\] $P[\langle,\rangle]$ coincides $\mu$-a.e. with an $\alpha$-Hölder continuous conformal structure which is preserved by $\mathcal{A}$. We lastly show how to obtain Theorem \[hyperbolic rigidity\] from Corollary \[irreducible\]. We let $T^{1}M$ denote the unit tangent bundle of a closed negatively curved manifold $M$ with $\dim M \geq 3$ and let $g^{t}$ be the geodesic flow on $T^{1}M$. Since $M$ is negatively curved, the geodesic flow is an Anosov flow and thus there is a $Dg^{t}$-invariant splitting $T(T^{1}M) = E^{u} \oplus E^{c} \oplus E^{s}$ where $E^{u}$ is exponentially expanded by $Dg^{t}$, $E^{s}$ is exponentially contracted by $Dg^{t}$, and $E^{c}$ is tangent to the flow direction of $g^{t}$. We also have $\dim E^{u} = \dim M - 1 \geq 2$ and we know that $E^{u}$ is an $\alpha$-Hölder continuous subbundle of $T(T^{1}M)$ for some $\alpha > 0$. These assertions are all well-known, see [@HK] for instance for proofs. Let $\varphi: T^{1}M \rightarrow {\mathbb R}$ be a Hölder continuous potential and $\mu_{\varphi}$ the $g^{t}$-invariant probability measure which is the equilibrium state associated to $\varphi$. Equilibrium states include the Liouville volume on $T^{1}M$ and the Bowen-Margulis measure of maximal entropy; we refer to [@BR] for details. Lastly for a subshift of finite type $\Sigma$ and a continuous function $\psi: \Sigma \rightarrow (0,\infty)$ we let $\Sigma_{\psi}$ be the space $$\Sigma_{\psi} = \{(x,t) \in \Sigma \times [0,\infty]: t \leq \psi(x) \}/(x,\psi(x)) \sim (f(x),0)$$ and let $F^{t}$ be the special flow on $\Sigma_{\psi}$ defined by $F^{s}(x,t) = (x,t+s)$ (up to the equivalence relation above). Fix a Riemannian metric on $T^{1}M$ and let $\| \cdot \|$ be the norm on $E^{u}$ given by restricting the Riemannian norm to this bundle. We claim that it suffices to prove that $Dg^{t}\|_{E^{u}}$ is uniformly quasiconformal in the sense that there is a constant $C > 0$ such that for all $v \in E^{u}$ and $t \in {\mathbb R}$, $$\left\|Dg^{t}(v)\right\| \cdot \left\|(Dg^{t}(v))^{-1}\right\| \leq C \|v\|^{2}.$$ This immediately implies that for each periodic point $p$ of $g^{t}$ of period ${\ell}(p)$ the complex eigenvalues of $Dg^{{\ell}(p)}: E^{u}_{p} \rightarrow E^{u}_{p}$ are all equal in absolute value. Hence by work of the author [@Bu1]\[Theorem 1.1\] this implies that $M$ has constant negative curvature. For this proof we set $d = \dim E^{u}$. We will derive from the derivative cocycle $Dg^{t}\|_{E^{u}}$ a Hölder continuous cocycle $\mathcal{A}$ over a subshift of finite type $f: \Sigma \rightarrow \Sigma$ together with an $f$-invariant measure $\nu$ corresponding to the $g^{t}$-invariant measure $\mu_{\varphi}$ such that if $Dg^{t}\|_{E^{u}}$ is measurably irreducible with respect to $\mu_{\varphi}$ then $\mathcal{A}$ is measurably irreducible with respect to $\nu$. We then apply Corollary \[irreducible\] to $\mathcal{A}$ to derive a continuous invariant conformal structure. We will show that this implies that the cocycle $Dg^{t}\|_{E^{u}}$ is uniformly quasiconformal which completes the proof by our work above. From the remarks above it suffices to show that $Dg^{t}_{E^{u}}$ preserves a continuous conformal structure. We first recall aspects of the construction of a Markov partition for an Anosov flow [@R73]. The geodesic flow $g^{t}$ is a topologically mixing Anosov flow and thus admits a Markov partition into parallelograms $\{P_{i}: 1 \leq i \leq k\}$, whose sides are composed of orbits of $g^{t}$ and leaves of the stable and unstable foliations for $g^{t}$. Given any ${\varepsilon}> 0$ these parallelograms can be chosen such that the bases $\{Q_{i} :1 \leq i \leq k\}$ transverse to the orbits of $g^{t}$ satisfy $\text{diam}\, Q_{i} < {\varepsilon}$. We choose ${\varepsilon}$ small enough that the bundle $E^{u}$ is trivial on each base $Q_{i}$. There is then a subshift of finite type $f: \Sigma \rightarrow \Sigma$ on $k$ symbols, a Hölder continuous roof function $\psi: \Sigma \rightarrow (0,\infty)$, and a Hölder continuous surjection $h: \Sigma_{\psi} \rightarrow T^{1}M$ such that the flow $F^{t}$ on $\Sigma_{\psi}$ satisfies $h \circ F^{t} = g^{t} \circ h$. Furthermore, identifying $\Sigma$ with $\Sigma \times \{0\} \subset \Sigma_{\psi}$, each cylinder $[0;i]$, $1 \leq i \leq k$, satisfies $h([0;i]) = Q_{i}$. There is an ergodic, fully supported measure $\hat{\nu}$ on $\Sigma_{\psi}$ with local product structure such that $h_{}\hat{\nu} = \mu_{\varphi}$ and such that $h$ is a bijection from a full measure subset $\Omega$ of $\Sigma_{\psi}$ to a full measure subset $h(\Omega)$ of $T^{1}M$. We let $\nu$ be the probability measure on $\Sigma$ obtained from pushing forward the measure $\hat{\nu}$ by the natural projection $\Sigma_{\psi} \rightarrow \Sigma$. The measure $\nu$ is also ergodic, fully supported, and has local product structure[@BR]. The bundle $E^{u}$ pulls back to a Hölder vector bundle $\hat{{\mathcal{E}}}$ over $\Sigma_{\psi}$. We let ${\mathcal{E}}$ be the bundle over $\Sigma$ obtained from pulling back $\hat{{\mathcal{E}}}$ by the identification $\Sigma \rightarrow \Sigma \times \{0\} \subset \Sigma_{\psi}$. Since $E^{u}$ is a Hölder continuous bundle which is trivial over each set $Q_{i}$ there is a Hölder trivialization $L: \Sigma \times {\mathbb R}^{d} \rightarrow {\mathcal{E}}$. The restriction $Dg^{t}\|E^{u}$ of the derivative cocycle $Dg^{t}$ of $g^{t}$ to the unstable bundle $E^{u}$ is Hölder (from the Hölder continuity of $E^{u}$). We let $\hat{A}^{t}$ denote the pullback of the linear cocycle $Dg^{t}\|E^{u}$ to a linear cocycle on $\hat{\mathcal{E}}$ over $F^{t}$. This induces a Hölder cocycle $\mathcal{A}$ with generator $A: \Sigma \rightarrow GL(d,{\mathbb R})$ over $f: \Sigma \rightarrow \Sigma$ by taking the time $\psi(x)$ map $\hat{A}^{\psi}(x)$ over $x \in \Sigma \times \{0\}$ and then using the trivialization $L$ of ${\mathcal{E}}$. We claim that $\mathcal{A}$ is measurably irreducible. Suppose that there is a proper measurable invariant subbundle $V$ for $\mathcal{A}$. For each $x\in \Sigma$ at which $V$ is defined we let $V_{x} \subset {\mathbb R}^{d}$ be the subspace for this bundle at $x$. For each $x \in \Sigma$ at which $V_{x}$ is defined and each $t \in {\mathbb R}$ we define $\hat{V}_{(x,t)} = \hat{A}^{t}(L(V_{x}))$. From the local decomposition of $\hat{\nu}$ into a product of $\nu$ and Lebesgue measure on orbits together with the $\nu$-a.e. $\mathcal{A}$-invariance of $V$ it follows that $\hat{V}$ is a $\hat{\nu}$-a.e. well-defined measurable subbundle of $\hat{{\mathcal{E}}}$ which is also $\hat{A}^{t}$-invariant $\hat{\nu}$-a.e. From the construction of $\hat{{\mathcal{E}}}$ this implies that there is a proper measurable subbundle of $E^{u}$ which is defined and $Dg^{t}$-invariant $\mu_{\varphi}$-a.e. This contradicts the hypothesis that $Dg^{t}\|_{E^{u}}$ is measurably irreducible. Having shown $\mathcal{A}$ is measurably irreducible, we apply Corollary \[irreducible\] to conclude that $\mathcal{A}$ preserves a continuous invariant conformal structure $\eta$. From this we construct a continuous invariant conformal structure $\hat{\eta}$ for $\hat{A}^{t}$ by the formula $\hat{\eta}_{(x,t)} = (\hat{A}^{t} \circ L)_{}[\eta_{x}]$ for each $x \in \Sigma$ and $t \in {\mathbb R}$. Recall from the construction of a Markov partition that there is a full $\hat{\nu}$-measure subset $\Omega \subset \Sigma_{\psi}$ on which $h : \Omega \rightarrow h(\Omega)$ is a continuous bijection onto a full $\mu_{\varphi}$-measure subset $h(\Omega)$ of $T^{1}M$. Let $S: \hat{{\mathcal{E}}}\|_{\Omega} \rightarrow E^{u}\|_{h(\Omega)}$ be the continuous linear identification on fibers coming from the construction of $\hat{{\mathcal{E}}}$ as the pullback of $E^{u}$. Since $\hat{\eta}$ is continuous on $\Sigma_{\psi}$, we conclude that there is a constant $\gamma > 0$ such that for any $v \in E^{u}\|_{h(\Omega)}$ we have $$\gamma^{-1} \cdot \hat{\eta}(S^{-1}(v), S^{-1}(v)) \leq \|v\|^{2} \leq \gamma \cdot \hat{\eta}(S^{-1}(v), S^{-1}(v)).$$ Since $\hat{\eta}$ is invariant under $\hat{A}^{t}$ and $S \circ \hat{A}^{t} = Dg^{t} \circ S$, we conclude using the above inequality that for any $v \in E^{u}\|_{h(\Omega)}$ and $t \in {\mathbb R}$ we have $$\begin{aligned} \left\|Dg^{t}(v)\right\|^{2} &\leq \gamma \cdot \hat{\eta}(S^{-1}(Dg^{t}(v)), S^{-1}(Dg^{t}(v))) \\ & = \gamma \cdot \hat{\eta}(S^{-1}(v), S^{-1}(v)) \\ & \leq \gamma ^{2} \|v\|^{2}\end{aligned}$$ Since $(Dg^{t}(v))^{-1} = Dg^{-t}(g^{t}(v)$, this readily implies that $$\left\|Dg^{t}(v)\right\| \cdot \left\|(Dg^{t}(v))^{-1}\right\| \leq \gamma^{2} \|v\|^{2}$$ for all $v \in E^{u}\|_{h(\Omega)}$ and $t \in {\mathbb R}$. Since $h(\Omega)$ is a dense subset of $T^{1}M$ we conclude that this inequality actually holds for all $v \in E^{u}$ and $t \in {\mathbb R}$. We conclude that $Dg^{t}\|_{E^{u}}$ is a uniformly quasiconformal linear cocycle over $g^{t}$, as desired. Definitions and Reductions {#shift definitions} ========================== In this section we define some common objects associated to $\Sigma$ and make some preliminary reductions before beginning the proof of Theorem \[conformalrigidity\]. Definitions ----------- We define $$\Sigma^{u} = \{(x_{n})_{n \geq 0}: q_{x_{n}x_{n+1}} = 1 \;\text{for all $n \geq 0$}\},$$ $$\Sigma^{s} = \{(x_{n})_{n < 0}: q_{x_{n}x_{n+1}} = 1 \;\text{for all $n \leq -2$}\},$$ to be the sets of one-sided right and left infinite sequences respectively which are associated to $Q$. We have projections $\pi_{u}: \Sigma \rightarrow \Sigma^{u}$ and $\pi_{s}: \Sigma \rightarrow \Sigma^{s}$ obtained by dropping all of the negative coordinates and all of the nonnegative coordinates respectively of a sequence in $\Sigma$. We let $f_{u}$ denote the left shift on $\Sigma^{u}$ and $f_{s}$ denote the right shift on $\Sigma^{s}$. We define the local stable set of $x \in \Sigma$ to be $$W^{s}_{loc}(x) = \{(y_{n})_{n \in\mathbb{Z}} \in \Sigma:x_{n} = y_{n} \; \text{for all $n \geq 0$}\},$$ and the local unstable set to be $$W^{u}_{loc}(x) = \{(y_{n})_{n \in \mathbb{Z}} \in \Sigma :x_{n} = y_{n} \; \text{for all $n \leq 0$}\}.$$ For $m \in \mathbb{Z}$ and $a_{0},\dots,a_{k} \in {\mathbb N}$, we define the cylinder notation $$[m;a_{0},\dots,a_{k}] = \{x \in \Sigma: x_{m+i} = a_{m+i}, \, 0 \leq i \leq k\}.$$ For $m \geq 0$, we let $[m;a_{0},\dots,a_{k}]^{u}$ denote the corresponding cylinder in $\Sigma^{u}$ and for $m \leq -k$, $[m;a_{0},\dots,a_{k}]^{s}$ is the corresponding cylinder in $\Sigma^{s}$. For two points $x,y \in [0;i]$, $1 \leq i \leq {\ell}$, we define $[x,y]$ to be the unique point in the intersection $W^{u}_{loc}(x) \cap W^{s}_{loc}(y)$. We let $\mu$ be a fully supported ergodic $f$-invariant probability measure on $\Sigma$. We let $\mu^{u} = (\pi_{u})_{}\mu$ and $\mu^{s} = (\pi_{s})_{}\mu$ denote the projections of $\mu$ to $\Sigma^{u}$ and $\Sigma^{s}$ respectively. For a subset $E$ of $\Sigma$ we let $\mu \| E$ denote the restriction of $\mu$ to $E$. Note that for each $1 \leq i \leq {\ell}$ we have a natural homeomorphism $[0;i] \rightarrow \pi_{s}([0;i]) \times \pi_{u}([0;i]))$ given by $\pi_{s} \times \pi_{u}$. We say that $\mu$ has local product structure if there is a continuous function $\xi: \Sigma \rightarrow (0,\infty)$ such that $$\mu \| [0;i] = \xi \cdot (\mu^{s}\|\pi_{s}([0;i]) \times \mu^{u}\|\pi_{u}([0;i])).$$ We will always assume that $\mu$ has local product structure in this paper. As a consequence of the local product structure of $\mu$, the measures $\mu^{u}$ and $\mu^{s}$ admit Jacobians with respect to the dynamics $f_{u}$ and $f_{s}$ [@BV04 Lemma 2.2], \[jacobians\] There are continuous functions $J_{u}: \Sigma^{u} \rightarrow (0,\infty)$ and $J_{s}: \Sigma^{s} \rightarrow (0,\infty)$ such that for each $1 \leq i,j \leq {\ell}$ and each Borel subset $K_{u} \subseteq [0;i]^{u}$, $$\mu^{u}(K_{u}) = \int_{f_{u}^{-1}(K_{u}) \cap [0;j,i]^{u}}J_{u}\, d\mu^{u},$$ and likewise for each Borel subset $K_{s} \subseteq [0;i]^{s}$, $$\mu^{s}(K_{s}) = \int_{f_{s}^{-1}(K_{s}) \cap [-1;i,j]^{s}}J_{s} \, d\mu^{s},$$ In particular, if $\mu^{u}(K_{u}) > 0$ then $\mu^{u}(f_{u}^{-1}(K_{u})) \cap [0;j,i]^{u}) > 0$ for each $1 \leq j \leq {\ell}$, and the analogous statement is true for $\mu^{s}$ and $K_{s}$. Let $\nu$ be any $f$-invariant ergodic probability measure on $\Sigma$. The extremal Lyapunov exponents of $\mathcal{A}$ with respect to $\nu$ are defined to be $${\lambda}_{+}(\mathcal{A},\nu) = \inf_{n \geq 1} \frac{1}{n}\int_{\Sigma} \log \\|A^{n}\\| \,d\nu,$$ $${\lambda}_{-}(\mathcal{A},\nu) = \sup_{n \geq 1} \frac{1}{n}\int_{\Sigma} \log \\|A^{-n}\\|^{-1} \,d\nu$$ By Kingman’s subadditive ergodic theorem [@K68], for $\nu$-a.e. $x \in \Sigma$ we have $$\lim_{n \rightarrow \infty} \frac{1}{n}\log \\|A^{n}(x)\\| = {\lambda}_{+}(\mathcal{A},\nu),$$ and the analogous statement is true for ${\lambda}_{-}(\mathcal{A},\nu)$. We will be particularly interested in the extremal Lyapunov exponents of probability measures supported on periodic orbits of $f$. Reductions {#subsec: reduction} ---------- We make some reductions before beginning the proof of Theorem \[conformalrigidity\]. Observe that a function is $\alpha$-Hölder continuous with respect to the distance $\rho_{\tau}$ if and only if it is Lipschitz continuous with respect to $\rho_{\alpha\tau}$. Hence it suffices to prove Theorem \[conformalrigidity\] in the case that $A$ is Lipschitz continuous with respect to $\rho_{\tau}$. From now on we fix $\tau \in (0,\infty)$ and define $\rho: = \rho_{\tau}$. We next reduce to the case where $A$ takes values in $SL(d,{\mathbb R})$. If we define $$B = \det(A)^{-\frac{1}{d}} \cdot A,$$ then $\det(B(x)) = 1$ for every $x \in \Sigma$ so $B$ takes values in $SL(d,{\mathbb R})$. It is clear that $B$ is Lipschitz with respect to $\rho$ if and only if $A$ is Lipschitz with respect to $\rho$. Let $\mathcal{B}$ be the linear cocycle generated by $B$. If $\eta: \Sigma \rightarrow \mathcal{C}^{d}$ defines a measurable conformal structure on $\Sigma \times {\mathbb R}^{d}$, then it is clearly $\mu$-a.e. $\mathcal{A}$-invariant if and only if it is $\mu$-a.e. $\mathcal{B}$-invariant, since these cocycles are conformally equivalent to each other. Thus for the rest of the paper we will assume that $\mathcal{A}$ takes values in $SL(d,{\mathbb R})$. As a consequence, for $\mu$-a.e. $x \in \Sigma$ we have the equality $$\eta_{f(x)}(A(x)v,A(x)w) = \eta_{x}(v,w), \; \text{for every $v,w \in {\mathbb R}^{d}$}.$$ Lastly we reduce to the case where $f: \Sigma \rightarrow \Sigma$ is topologically mixing. By the spectral decomposition theorem for hyperbolic basic sets there is an integer $k \geq 1$ such that we can write $\Sigma = \bigsqcup_{i = 1}^{k} \Sigma_{i}$ for closed subsets $\Sigma_{i}$ of $\Sigma$ such that $f(\Sigma_{i}) = \Sigma_{i+1}$, where the index is taken mod $k$ and $f^{k}\|\Sigma_{i}$ is a topologically mixing subshift of finite type for $1 \leq i \leq k$. The normalized restriction $\mu_{i}$ of $\mu$ to $\Sigma_{i}$ is an $f^{k}$-invariant ergodic, fully supported probability measure with local product structure and $A^{k}\|\Sigma_{i}$ generates a Lipschitz continuous linear cocycle $\mathcal{A}_{i}^{k}$ over $f^{k}$. If $\mathcal{A}$ preserves a measurable conformal structure $\eta$, then $\mathcal{A}_{i}^{k}$ preserves the measurable conformal structure $\eta \| [0;i]$. Hence if Theorem \[conformalrigidity\] holds for topologically mixing subshifts of finite type, then $\eta \| [0;i]$ coincides $\mu_{i}$-a.e. with a Lipschitz continuous invariant conformal structure for $\mathcal{A}^{k}_{i}$. This holds for each $1 \leq i \leq k$; we thus conclude that $\eta$ agrees $\mu$-a.e. with a Lipschitz continuous conformal structure on $\Sigma \times {\mathbb R}^{d}$ which is necessarily $\mathcal{A}$-invariant. Proof of Main Theorem {#sec: proof} ===================== We let $f: \Sigma \rightarrow \Sigma$ be a topologically mixing subshift of finite type equipped with a metric $\rho = \rho_{\tau}$ as defined in . We let $\mathcal{A}$ be a Lipschitz continuous $SL(d,{\mathbb R})$-valued cocycle over $f$. Let $\eta: \Sigma \rightarrow SL(d,{\mathbb R})/SO(d,{\mathbb R})$ be an $\mathcal{A}$-invariant measurable family of inner products on ${\mathbb R}^{d}$ for which the equality $A(x)[\eta_{x}] = \eta_{f(x)}$ holds for $\mu$-a.e. $x$. In this sectionw we will prove that there is a Lipschitz continuous $\mathcal{A}$-invariant inner product $\hat{\eta}$ which coincides $\mu$-a.e. with $\eta$. By the reductions of Section \[subsec: reduction\] this completes the proof of Theorem \[conformalrigidity\]. For a periodic point $x$ of period $k$ we define $\nu_{x}$ to be the $f$-invariant probability measure supported on the orbit of $p$ which is given by the formula $$\nu_{x} = \frac{1}{k}\sum_{j = 0}^{k-1} \delta_{f^{j}(x)},$$ where $\delta_{p}$ denotes the Dirac probability measure supported at a point $p$. We use the shorthand ${\lambda}_{+}(p) = {\lambda}_{+}(\mathcal{A},\nu_{p})$ and ${\lambda}_{-}(p) ={\lambda}_{-} (\mathcal{A},\nu_{p})$ for the extremal Lyapunov exponents of $\mathcal{A}$ with respect to $\nu_{p}$. The key proposition we will prove in this section is the following, \[periodiczero\] Suppose that $\mathcal{A}$ preserves a measurable conformal structure. Then ${\lambda}_{+}(p) = 0$ for every periodic point $p$ of $f$. In Section \[conclusion\] at the end of the paper we apply work of Kalinin and Sadovskaya to show how Proposition \[periodiczero\] implies Theorem \[conformalrigidity\]. The plan of the proof is the following: We let $\mathcal{D}(N,\theta)$ be the set of points $x \in \Sigma$ which satisfy the two conditions $$\prod_{j = 0}^{s-1} \\|A^{N}(f^{jN}(x))\\| \\|A^{N}(f^{jN}(x))^{-1}\\| \leq e^{sN\theta} \; \; \text{for all $s \geq 1$},$$ and $$\prod_{j = 0}^{s-1} \\|A^{-N}(f^{-jN}(x))\\| \\|A^{-N}(f^{-jN}(x))^{-1}\\| \leq e^{sN\theta} \; \; \text{for all $s \geq 1$}.$$ The cocycle $\mathcal{A}$ is fiber bunched if and only if there is a $\theta < \tau$ and $N \in {\mathbb N}$ such that $\mathcal{D}(N,\theta) = \Sigma$. In general the sets $\mathcal{D}(N,\theta)$ with $\theta < \tau$ isolate compact subsets of the space $\Sigma$ on which we can treat $\mathcal{A}$ as if it were fiber bunched. A straightforward argument will show that if $\mathcal{A}$ preserves a measurable conformal structure $\eta$ then, given any $\theta > 0$, we have that for $\mu$-a.e. $x \in \Sigma$ there is an $N > 0$ such that $x \in \mathcal{D}(N,\theta)$. We show in Lemma \[smallperiodiczero\] below that any periodic point $p$ satisfying ${\lambda}_{+}(p) - {\lambda}_{-}(p) < \tau$ lies in $\mathcal{D}(N,\theta)$ for some $\theta < \tau$ and some $N$. We will show in Section \[extension\] that $\eta$ admits a canonical $\mathcal{A}$-invariant (in an appropriate sense) extension $\hat{\eta}$ from a full measure subset of $\mathcal{D}(N,\theta)$ to the entire set. In particular, letting $k$ denote the period of $p$, we conclude that $A^{k}(p)$ preserves the inner product $\hat{\eta}_{p}$ on ${\mathbb R}^{d}$. This shows that if ${\lambda}_{+}(p) - {\lambda}_{-}(p)< \tau$ then we actually have ${\lambda}_{+}(p) = 0$. Thus there are no periodic points $p$ satisfying $0 < {\lambda}_{+}(p) - {\lambda}_{-}(p)< \tau$. The essential strategy to finish the proof is to show that if there exists a periodic point $x$ satisfying ${\lambda}_{+}(x) -{\lambda}_{+}(x) \geq \tau$ then by appropriately mixing the orbit of $x$ with the orbit of a periodic point $y$ satisfying ${\lambda}_{+}(y) = 0$ we can construct a new periodic point $p$ satisfying $0 < {\lambda}_{+}(p) - {\lambda}_{-}(p)< \tau$. This would give a contradiction that shows that we must have ${\lambda}_{+}(p) = 0$ for all periodic points $p$. For technical reasons related to the difficulty of estimating Lyapunov exponents from below we will instead use a modification of this strategy below. Holonomy invariance {#subsec: holonomies} ------------------- Our first observation is that the existence of a measurable invariant conformal structure implies that the extremal Lyapunov exponents of $\mathcal{A}$ with respect to $\mu$ must be zero, \[zeroexponents\] If there is a measurable function $\eta: \Sigma \rightarrow SL(d,{\mathbb R})/SO(d,{\mathbb R})$ such that $A(x)[\eta_{x}] = \eta_{f(x)}$ for $\mu$-a.e. $x$ then ${\lambda}_{+}(\mathcal{A},\mu) = -{\lambda}_{-}(\mathcal{A},\mu) = 0$. By Lusin’s theorem we can find a compact subset $K \subseteq \Sigma$ with positive $\mu$-measure on which $\eta$ is uniformly continuous. Let $\hat{f}: K \rightarrow K$ be the first return map to $K$ and $\hat{\mathcal{A}}$ the induced cocycle over $\hat{f}$ on $K$ with generator $\hat{A}$. By continuity of $\eta$ restricted to $K$, there is a constant $C \geq 1$ such that for all $x \in K$ and $v \in {\mathbb R}^{d}$ we have $$C^{-1}\\|v\\|^{2} \leq \eta_{x}(v,v) \leq C\\|v\\|^{2}$$ For every $x \in K$, every $n \geq 1$, and each $v \in {\mathbb R}^{d}$, we know that $$C^{-1}\\|\hat{A}^{n}(x)(v)\\|^{2} \leq \eta_{\hat{f}^{n}(x)}(\hat{A}^{n}(x)(v),\hat{A}^{n}(x)(v)) \leq C\\|\hat{A}^{n}(x)(v)\\|^{2}$$ Combining this with the fact that for every $x \in K$ we have $\hat{A}_{}^{n}\eta_{x} = \eta_{\hat{f}^{n}(x)}$, we conclude that for every $x \in K$ we have $$\\|\hat{A}^{n}(x)(v)\\|^{2} \leq C^{2} \\|v\\|^{2},$$ for every $n \geq 1$. Hence ${\lambda}_{+}(\hat{\mathcal{A}},\mu_{K}) = 0$, where $\mu_{K} = \mu(K)^{-1}\mu \| K$ is the induced invariant probability measure for $\hat{f}$ on $K$. By a direct application of Oseledet’s multiplicative ergodic theorem (see [@V14] for instance) the extremal exponents of the induced cocycle $\hat{\mathcal{A}}$ are related to those of $\mathcal{A}$ by the equation ${\lambda}_{+}(\hat{\mathcal{A}},\mu_{K}) = \mu(K)^{-1}{\lambda}_{+}(\mathcal{A},\mu)$. This implies that ${\lambda}_{+}(\mathcal{A},\mu) = 0$. Since ${\lambda}_{+}(\mathcal{A},\mu) = 0$, we conclude from [@V08 Corollary 2.4] that for any given $\theta > 0$ and $\mu$-a.e. $x \in \Sigma$, there is an $N > 0$ such that $x \in \mathcal{D}(N,\theta)$. When $\theta < \tau$ the cocycle $\mathcal{A}$ carries additional structure over $\mathcal{D}(N,\theta)$, \[existenceholonomies\] Given $N,\theta$ with $\theta < \tau$, there exists $L = L(N,\theta) > 0$ such that for any $x \in \mathcal{D}(N,\theta)$ and any $y,z \in W^{s}_{loc}(x)$, $$H^{s}_{yz} = \lim_{n \rightarrow \infty} A^{n}(z)^{-1}A^{n}(y),$$ exists and satisfies $\\|H^{s}_{yz}-I\\| \leq L \cdot \rho (y,z)$ and $H^{s}_{yz} = H^{s}_{xz} \circ H^{s}_{yx}$. Similarly, if $y,z \in W^{u}_{loc}(x)$ then the limit $$H^{u}_{yz} = \lim_{n \rightarrow \infty} A^{n}(z)A^{n}(y)^{-1},$$ exists and satisfies $\\|H^{u}_{yz}-I\\| \leq L \cdot \rho (y,z)$ and $H^{u}_{yz} = H^{u}_{xz} \circ H^{u}_{yx}$. We refer to $H^{s}$ and $H^{u}$ as the stable and unstable holonomies of $\mathcal{A}$ respectively. We will show in this subsection that there is an $f$-invariant full measure subset $\Omega$ of $\Sigma$ on which $\eta$ is both $f$-invariant and holonomy invariant in the sense that a holonomy map between two points in $\Omega$ preserves $\eta$. In the next lemma we show that $\eta$ is invariant under these stable and unstable holonomies on a full measure subset of $\Sigma$, \[holonomyinvariance\] There is an $f$-invariant subset $\Omega$ of $\Sigma$, with $\mu(\Omega) = 1$, such that if $x,y \in \Omega$ and $y \in W^{}_{loc}(x)$ then $$H^{}_{xy}[\eta_{x}] = \eta_{y},$$ for $ \in \{s,u\}$, and furthermore for every $n \in {\mathbb Z}$ we have $$A^{n}(x)[\eta_{x}] = \eta_{f^{n}(x)},$$ for every $x \in \Omega$. We follow [@KS Proposition 4.4], with minor adjustments to account for the fact that the holonomies $H^{s}$ and $H^{u}$ do not necessarily vary continuously on $\Sigma$. Fix $N > 0$ and $\theta < \tau$ such that $\mu(\mathcal{D}(N,\theta)) > \frac{9}{10}$. By Lusin’s theorem we can find a compact subset $K \subset \Sigma$ with $\mu(K) > \frac{9}{10}$ on which $\eta$ is uniformly continuous and thus bounded. We note that $$\frac{\mu(\mathcal{D}(N,\theta) \cap K)}{\mu(\mathcal{D}(N,\theta))} > \frac{4}{5}.$$ Set $E = \mathcal{D}(N,\theta) \cap K$. Let $\hat{f}$ denote the first return map to $\mathcal{D}(N,\theta)$. Let $S \subset \mathcal{D}(N,\theta)$ be the subset of points of $\mathcal{D}(N,\theta)$ for which the frequency of visits to $E$ under iteration of $\hat{f}^{n}$ converges to $\frac{\mu(E)}{\mathcal{D}(N,\theta)} > \frac{1}{2}$ as $n \rightarrow \infty$. We have $\mu(S) = \mu(\mathcal{D}(N,\theta))$ by the Birkhoff ergodic theorem since $\hat{f}$ is ergodic. Hence for $\mu$-a.e. $x,y \in \mathcal{D}(N,\theta)$ with $y \in W^{s}_{loc}(x)$ there are infinitely many $n \geq 1$ such that both $\hat{f}^{n}(x)$ and $\hat{f}^{n}(y) \in E$. Fix any pair of such points $x,y$ with $y \in W^{s}_{loc}(x)$ and with the $A$-invariance property of $\eta$ for all iterates, and let $n_{i} \rightarrow \infty$ be a sequence with $x^{n_{i}}:=\hat{f}^{n_{i}}(x)$, $y^{n_{i}}:= \hat{f}^{n_{i}}(y)\in E$ for each $i$. By $\hat{A}$-invariance of $\eta$ on the orbits of $x$ and $y$ and since pushing forward by $\hat{A}$ is an isometry with respect to the metric $g$, $$\begin{aligned} g(\eta(y), H^{s}_{xy}[\eta(x)]) &= g(\hat{A}^{n_{i}}(y)[\eta(y)], \hat{A}^{n_{i}}(y)H^{s}_{xy}[\eta(x)])\\ &= g(\eta(y^{n_{i}}), H^{s}_{x^{n_{i}}y^{n_{i}}}[\eta(x^{n_{i}})]) \\ &\leq g(\eta(y^{n_{i}}), \eta(x^{n_{i}})) + g(\eta(x^{n_{i}}), H^{s}_{x^{n_{i}}y^{n_{i}}}[\eta(x^{n_{i}})])\end{aligned}$$ Since $\eta$ is uniformly continuous on $E$, $$g(\eta(x^{n_{i}}), \eta(y^{n_{i}})) \rightarrow 0,$$ as $n_{i} \rightarrow \infty$. Also as $n_{i} \rightarrow \infty$, $H^{s}_{x^{n_{i}}y^{n_{i}}}$ converges uniformly to the identity map $I$ on ${\mathbb R}^{d}$. Since $\eta$ is bounded on $E$, this implies that $$g(\eta(x^{n_{i}}), H^{s}_{x^{n_{i}}y^{n_{i}}}[\eta(x_{n_{i}})]) \rightarrow 0 \;\;\text{as $n_{i} \rightarrow \infty$}.$$ We conclude that $\eta(y) = H^{s}_{xy}[\eta(x)]$, as desired. This holds for $\mu$-a.e. $x,y \in \mathcal{D}(N,\theta)$ with $y \in W^{s}_{loc}(x)$. Note that if $s \geq 1$ then $\mathcal{D}(N,\theta) \subset \mathcal{D}(sN,\theta)$. Taking the union over all $N > 0 $ of the sets $\mathcal{D}(N,\theta)$ for a fixed $\theta < \tau$ with $\mu(\mathcal{D}(N,\theta)) > \frac{9}{10}$ and also applying this argument replacing $H^{s}$ by $H^{u}$, we obtain a full $\mu$-measure subset $\Omega'$ of $\Sigma$ such that if $x,y \in \Omega$ and $y \in W^{}_{loc}(x)$ then the holonomy $H^{}_{xy}$ is defined according to Proposition \[existenceholonomies\] and $\eta(y) = H^{}_{xy}[\eta(x)]$ for $ \in \{s,u\}$. By intersecting $\Omega'$ with the full measure set on which the equality $A(x)[\eta_{x}] = \eta_{f(x)}$ holds we can arrange that this $A$-invariance property holds on $\Omega'$. We then set $\Omega = \cap_{n \in {\mathbb Z}} f^{n}(\Omega')$ to obtain the desired $f$-invariance property of $\Omega$. Extending $\eta$ {#subsec: extension} ---------------- For each $N > 0$ and $\theta < \tau$, the stable and unstable holonomies $H^{s}$ and $H^{u}$ exist and vary continuously on $\mathcal{D}(N,\theta)$. By construction $\eta$ is defined on $\Omega \cap \mathcal{D}(N,\theta)$ and invariant under the stable and unstable holonomies. By Proposition \[existenceholonomies\] the restrictions of $H^{s}$ and $H^{u}$ to $\mathcal{D}(N,\theta)$ are Lipschitz with Lipschitz constant depending only on $N$ and $\theta$. Hence $\eta$ is uniformly continuous on $\Omega \cap \mathcal{D}(N,\theta)$. We would like to construct a continuous extension $\hat{\eta}$ of $\eta$ to $\mathcal{D}(N,\theta)$ which agrees $\mu$-a.e. with $\eta$, is invariant under $H^{s}$ and $H^{u}$ on all of $\mathcal{D}(N,\theta)$, and such that if $x, f^{n}(x) \in \mathcal{D}(N,\theta)$ then $A^{n}(x)[\hat{\eta}_{x}] = \hat{\eta}_{f^{n}(x)}$. The natural choice for such an extension, since $\eta$ is uniformly continuous on $\Omega \cap \mathcal{D}(N,\theta)$, is to take the unique extension of $\eta$ to the closure $\overline{\Omega \cap \mathcal{D}(N,\theta)}$. However this may be a proper subset of $\mathcal{D}(N,\theta)$. The best we can guarantee is that $$\text{supp}(\mu \| \mathcal{D}(N,\theta)) \subseteq \overline{\Omega \cap \mathcal{D}(N,\theta)} \subseteq \mathcal{D}(N,\theta)$$ We overcome this issue using Lemma \[Lipschitzapprox\] below which is one of the main steps in the proof of Proposition \[periodiczero\]. This lemma makes critical use of the local product structure of $\mu$. \[Lipschitzapprox\] For each $N > 0$ and $\theta < \tau$ there is an $N_{} \geq N$ and $ \theta \leq \theta_{} < \tau$ such that $$\mathcal{D}(N,\theta) \subseteq \text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$$ Let $N > 0$, $\theta < \tau$ be given such that $$\frac{1}{N}\int_{\Sigma}\log \\|A^{N}(x)\\| \cdot \\|A^{N}(x)^{-1}\\| \,d\mu < \theta,$$ and set $\psi(x) = \frac{1}{N}\log \\|A^{N}(x)\\| \cdot \\|A^{N}(x)^{-1}\\|$. For any fixed $\theta > 0$ this inequality can always be arranged by taking $N$ large enough since ${\lambda}_{+}(\mathcal{A},\mu) = 0$. Set $S_{n}\psi(x) = \sum_{j=0}^{n-1}\psi(f(^{jN}(x)))$ and observe that if $x \in \mathcal{D}(N,\theta)$ then $\frac{S_{n}\psi(x)}{n} \leq \theta$. Let $x =(x_{n})_{n \in {\mathbb Z}}\in \mathcal{D}(N,\theta)$ be given and let $0 < \gamma < \frac{1}{10}(\tau-\theta)$. Choose points $z^{i}$, $1 \leq i \leq {\ell}$, in the cylinders $[0;i]$ and for $y \in [0;i]$ recall that $[z^{i},y]$ denotes the unique point in the intersection $W^{s}_{loc}(y) \cap W^{u}_{loc}(z_{i})$. We set $\psi^{u}(y) = \psi([z^{i},y])$ for $y \in [0;i]$, $1 \leq i \leq {\ell}$, and note that $$\psi^{u} = \psi + \varphi \circ f - \varphi$$ where $\varphi(y) = \sum_{k=0}^{\infty}\psi(f^{k}(y))-\psi(f^{k}([z^{i},y]))$ for $y \in [0;i]$. Since $\psi$ is Lipschitz with respect to the distance $\rho$ and $[z^{i},y] \in W^{s}_{loc}(y)$, standard distortion estimates show that the sum defining $\varphi$ converges uniformly and hence $\varphi$ is continuous and bounded on $\Sigma$. We similarly set $S_{n}\psi^{u}(x) = \sum_{j=0}^{n-1}\psi^{u}(f(^{jN}(x)))$ and note that $$\left\\|\frac{S_{n}\psi^{u} }{n}- \frac{S_{n}\psi}{n}\right\\|_{\infty} \leq \frac{2}{n}\\|\varphi\\|_{\infty}.$$ The function $\psi^{u}: \Sigma \rightarrow {\mathbb R}$ is constant on local stable sets and hence descends to a map $\psi^{u}: \Sigma^{u} \rightarrow {\mathbb R}$. We observe that $$\int_{\Sigma^{u}}\psi^{u}\,d\mu^{u} = \int_{\Sigma}\psi\,d\mu < \theta$$ By Egorov’s theorem, given any $\delta > 0$ we can find $K \subset \Sigma^{u}$ with $\mu^{u}(\Sigma \backslash K) > \delta$ and such that $\frac{S_{n}\psi^{u} }{n}$ converges uniformly to $\int_{\Sigma^{u}}\psi^{u}\,d\mu^{u}$ on $K$. Choose $K$ corresponding to a $\delta$ with $\delta < \inf_{1 \leq i \leq {\ell}} \mu([0;i])$; this forces $K \cap [0;i] \neq \emptyset$ for each $1\leq i\leq {\ell}$. For each $m \geq 0$ let $$K_{m}^{u}(x) = f_{u}^{-m}(K) \cap [0;x_{0},x_{1},\dots,x_{m}]^{u}$$ By proposition \[jacobians\] together with the fact that $K \cap [0;i] \neq \emptyset$ for each $1\leq i\leq {\ell}$ we have $\mu^{u}(K_{m}^{u}(x)) > 0$ for every $m \geq 0$. Choose $M$ large enough that for $n \geq M$ we have for every $w \in K$, $$\frac{S_{n}\psi^{u}(w)}{n} < \int_{\Sigma^{u}}\psi^{u}\,d\mu^{u} + \gamma < \theta + \gamma$$ Now let $y \in \pi_{u}^{-1}(K_{m}^{u}(x))$ for some $m \geq 0$. For $n \geq 1$ we write $$\frac{S_{n}\psi(y)}{n} = \frac{S_{m}\psi(y)-S_{m}\psi(x)}{n} + \frac{S_{m}\psi(x) }{n}+\frac{S_{n-m}\psi^{u}(f^{m}(y))}{n}$$ with the understanding that $S_{n-m}\psi(f^{m}(y)):= 0$ if $n \leq m$. We bound each of the three terms on the right side separately. For the first term we have $$\begin{aligned} \frac{S_{m}\psi(y)-S_{m}\psi(x)}{n} &\leq \frac{S_{m}\psi^{u}(y)-S_{m}\psi^{u}(x)}{n} + \frac{2}{n}\\|\varphi\\|_{\infty} \\ &\leq \frac{\text{Lip}(\psi^{u})}{n}\left(\sum_{k=0}^{m-1}\rho(f^{kN}(x),f^{kN}(y))\right) + \frac{2}{n}\\|\varphi\\|_{\infty} \\ &\leq \frac{1}{n}\left(\frac{\text{Lip}(\psi^{u})}{1-e^{-\tau}}+ 2\\|\varphi\\|_{\infty}\right)\end{aligned}$$ where $\text{Lip}(\psi^{u})$ denotes the Lipschitz constant of $\psi^{u}$. For the second term we use the bound $$\frac{S_{m}\psi(x) }{n} \leq \left\{ \begin{array}{cc} \frac{N}{n} \\|\psi\\|_{\infty}, & m \leq N \\ \frac{m}{n}\theta, & m > N \end{array} \right.$$ Similarly, for the third term we use the bound $$\frac{S_{n-m}\psi^{u}(f^{m}(y)) }{n} \leq \left\{ \begin{array}{cc} \frac{M}{n} \\|\psi^{u}\\|_{\infty}, & n-m \leq M \\ \frac{n-m}{n}(\theta + \gamma) , & m > N \end{array} \right.$$ Choose $R$ large enough that $$\frac{1}{R}\left(\frac{\text{Lip}(\psi^{u})}{1-e^{-\tau}}+ 2\\|\varphi\\|_{\infty}\right) < \gamma$$ and also large enough that $\frac{N}{R}\\|\psi\\|_{\infty} < \gamma$ and $\frac{M}{R}\\|\psi^{u}\\|_{\infty} < \gamma$. We conclude by combining all of the above estimates that for any $m \geq 1$, $x \in \mathcal{D}(N,\theta)$, $y \in \pi_{u}^{-1}(K_{m}^{u}(x))$ and $n \geq R$ we have $$\frac{S_{n}\psi(y)}{n} < \theta+ 3\gamma.$$ By our choice of $\gamma$ we have $\theta + 3\gamma < \tau$, so we can set $\theta_{} = \theta + 3\gamma$. We conclude that for every $m \geq 1$, $x \in \mathcal{D}(N,\theta)$, $y \in \pi_{u}^{-1}(K_{m}^{u})(x)$ and $k \geq 1$ we have $$\prod_{j = 0}^{k-1} \left\\|A^{N R}(f^{ j N R}(y))\right\\| \left\\|A^{N R}(f^{j N R}(y))^{-1}\right\\| \leq e^{N \cdot S_{kR}\psi(y)} \leq e^{k N R\theta_{}}.$$ We may replace $f$ and $A$ everywhere in the above proof by $f^{-1}$ and $A^{-1}$ and run the same arguments again with $u$ everywhere replaced by $s$. We obtain an $R' \geq N$ and an analogous sequence of subsets $$K_{m}^{s}(x) \subset [-m;x_{-m},x_{-m+1},\dots,x_{-1}]^{s},$$ of $\Sigma^{s}$ with $\mu^{s}(K_{m}^{s}(x)) > 0$ for every $m \geq 1$ such that for $m \geq 1$, $x \in \mathcal{D}(N,\theta)$, $y \in \pi_{s}^{-1}(K_{m}^{s})(x)$ and $k \geq 1$ we have $$\prod_{j = 0}^{k-1} \left\\|A^{-N R'}(f^{- j N R'}(y))\right\\| \left\\|A^{-N R'}(f^{-j N R'}(y))^{-1}\right\\| \leq e^{N \cdot S_{k R'}\psi(y)} \leq e^{k N R'\theta_{}}.$$ Let $N_{} = N \cdot \text{max}(R, R')$ and for $m \geq 1$ let $$E_{m}(x) = \pi_{s}^{-1}(K_{m}^{s})(x) \cap \pi_{u}^{-1}(K_{m}^{u})(x)$$ We conclude that if $x \in \mathcal{D}(N,\theta)$ and $y \in E_{m}(x)$ for some $m \geq 1$ then $y \in \mathcal{D}(N_{},\theta_{})$. Let $x \in \mathcal{D}(N,\theta)$. From the local product structure of the measure $\mu$ on the cylinder $[0;x_{0}]$ it follows that $$\mu\left(E_{m}(x)\right) \geq \left(\inf_{\Sigma}\xi\right)\mu^{u}(K_{m}^{u}(x)) \mu^{s}(K_{m}^{s}(x)) > 0$$ for every $m \geq 1$. Furthermore it is easy to check from the construction that $E_{m}(x)$ is contained in the $\rho$-ball of radius $e^{-m\tau}$ centered at $x$ in $\Sigma$. Since $E_{m}(x) \subset \mathcal{D}(N_{},\theta_{})$ for every $m \geq 1$ and the $\rho$-balls of radius $e^{-m\tau}$ form a neighborhood basis of $x$ as $m \rightarrow \infty$, we conclude that $x \in \text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$ as desired. We are now able to construct the desired extension $\hat{\eta}$ of $\eta$ to $\mathcal{D}(N,\theta)$ for each $N > 0$ large enough and $\theta < \tau$. Fix $\theta < \tau$ and take $N$ large enough that for $1 \leq i \leq {\ell}$ we have $\mu(\mathcal{D}(N,\theta) \cap [0;i]) > 0$. Choose points $\omega^{i} \in \Omega\cap \mathcal{D}(N,\theta) \cap [0;i]$ for $1 \leq i \leq {\ell}$. For any point $x \in \mathcal{D}(N, \theta) \cap [0;i]$ we then define $$\hat{\eta}_{x} = H^{s}_{[\omega^{i}, x] x}H^{u}_{\omega^{i} [\omega^{i}, x] }[\eta_{\omega^{i}}]$$ \[extension\] For each $\theta < \tau$ and $N > 0$, $\hat{\eta}$ is uniformly continuous on $\mathcal{D}(N,\theta)$. If $x \in \mathcal{D}(N,\theta)$ then for any $y \in W^{}_{loc}(x)$ we have $H^{}_{xy}[\hat{\eta}_{x}] = \hat{\eta}_{y}$ for $* \in \{s,u\}$ and $$A(x)[\hat{\eta}_{x}] = \hat{\eta}_{f(x)}$$ Fix $N > 0$ and $\theta < \tau$ as in the construction of $\hat{\eta}$ above. Apply Lemma \[Lipschitzapprox\] to $N$ and $\theta$ to obtain $\theta_{} < \tau$ and $N_{}$ such that $$\mathcal{D}(N,\theta) \subseteq \text{supp}(\mu \| \mathcal{D}(N_{},\theta_{})).$$ Since $\eta$ is uniformly continuous on $\Omega \cap \mathcal{D}(N_{},\theta_{})$ and this subset has full measure in $\mathcal{D}(N_{},\theta_{})$, we conclude that $\eta$ has a unique continuous extension to the set $\text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$ and therefore a unique continuous extension to the set $\mathcal{D}(N,\theta)$. Since every point of $\text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$ is a limit point of the set $\Omega \cap \mathcal{D}(N_{},\theta_{})$ on which $\eta$ is invariant under the stable and unstable holonomies $H^{s}$ and $H^{u}$ by Lemma \[holonomyinvariance\] and these holonomies are uniformly continuous on $\mathcal{D}(N_{},\theta_{})$ we conclude that this continuous extension of $\eta$ to $\text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$ is invariant under $H^{u}$ and $H^{s}$. Since $\omega^{i} \in \text{supp}(\mu \| \mathcal{D}(N_{},\theta_{}))$ for $1 \leq i \leq {\ell}$ this implies that this continuous extension coincides with $\hat{\eta}$. Thus $\hat{\eta}$ is uniformly continuous on $\mathcal{D}(N,\theta)$ and invariant under both $H^{s}$ and $H^{u}$ for any choice of $N > 0$, $\theta < \tau$. Lastly we need to show that $\hat{\eta}$ is $A$-invariant. To do this we use the following easy proposition, \[gap\] Let ${\varepsilon}> 0$ be given. Then for every $N$ large enough we have for any $\theta > 0$, $$f^{-1}(\mathcal{D}(N,\theta)) \subseteq \mathcal{D}(N,\theta+{\varepsilon}).$$ For every $x \in \Sigma$ and $N > 0$ we have the inequality $$\\|A^{N}(f^{-1}(x))\\| \leq \\|A^{N}(x)\\| \cdot \\|A(f^{N}(x))^{-1}\\| \cdot \\|A(f^{-1}(x))\\|$$ Let $R = \sup_{x \in \Sigma} \text{max}\{\\|A(x)\\|,\\|A(x)^{-1}\\|\}$. Then we have $$\\|A^{N}(f^{-1}(x))\\| \leq R^{2}\\|A^{N}(x)\\|$$ and similarly $$\\|A^{-N}(f^{-1}(x))\\| \leq R^{2}\\|A^{-N}(x)\\|$$ for every $x \in \Sigma$. We conclude that for every $x \in \mathcal{D}(N,\theta)$, $$\prod_{j = 0}^{k-1} \\|A^{N}(f^{jN}(f^{-1}(x)))\\| \\|A^{N}(f^{jN}(f^{-1}(x)))^{-1}\\| \leq R^{4k}e^{kN\theta} \; \; \text{for all $k \geq 1$},$$ and the same inequality holds with $f^{-1}$ and $A^{-1}$ replacing $f$ and $A$. Hence we see that if $R^{4}\leq e^{N{\varepsilon}}$ then $f^{-1}(x) \in \mathcal{D}(N,\theta+{\varepsilon})$, and this clearly holds for $N$ large enough. By Proposition \[gap\] we can (by taking $N_{}$ larger if necessary) find some $0 < {\varepsilon}< \tau - \theta_{}$ such that $$f^{-1}(\mathcal{D}(N_{},\theta_{})) \subset \mathcal{D}(N_{},\theta_{}+{\varepsilon})$$ Applying the conclusions of the previous paragraph to $\mathcal{D}(N_{},\theta_{}+{\varepsilon})$ since $\theta_{}+{\varepsilon}< \tau$, we conclude that $\hat{\eta}$ is uniformly continuous on $\mathcal{D}(N_{},\theta_{}+{\varepsilon})$ and thus $x \rightarrow \hat{\eta}_{f^{-1}(x)}$ is a uniformly continuous function on $\mathcal{D}(N_{},\theta_{})$. Setting $\zeta_{x} = A(f^{-1}(x))[\hat{\eta}_{f^{-1}(x)}]$, we conclude that $\zeta$ is uniformly continuous on $\mathcal{D}(N_{},\theta_{*})$ and $\zeta_{x} = \eta_{x}$ for $\mu$-a.e. $x \in \Sigma$. Hence $\zeta$ also gives a continuous extension of $\eta$ to $\mathcal{D}(N,\theta)$. By the uniqueness of this extension we conclude that $\zeta = \hat{\eta}$ on $\mathcal{D}(N,\theta)$ for any $N > 0$, $\theta <\tau$. This then implies that for $x \in \mathcal{D}(N,\theta)$, $$\hat{\eta}_{x} = A(f^{-1}(x))[\hat{\eta}_{f^{-1}(x)}],$$ which is equivalent to the desired invariance property of $\hat{\eta}$. Vanishing of periodic exponents {#subsec: vanishing} ------------------------------- We recall that for a periodic point $p$ we denote the $f$-invariant probability measure supported on the orbit of $p$ by $\nu_{p}$ and we denote the extremal Lyapunov exponents of $\nu_{p}$ by ${\lambda}_{+}(p)$ and ${\lambda}_{-}(p)$. The following lemma marks a major step in the proof of Proposition \[periodiczero\] which comes as a consequence of our work in Section \[subsec: extension\]. We are able to rule out the existence of periodic points for which the gap between the two extremal Lyapunov exponents of $\mathcal{A}$ falls below a certain threshold. We are also able to find a periodic point $q$ at which ${\lambda}_{+}(q) = 0$; in fact the techniques of the proposition make it clear how to produce many such points, but we will only need one for our purposes. \[smallperiodiczero\] Suppose that $\mathcal{A}$ preserves a measurable conformal structure. Then, 1. There are no periodic points $p$ for which the inequality, $$0 < {\lambda}_{+}(p)-{\lambda}_{-}(p) < \tau,$$ holds. 2. There is a periodic point $q$ such that ${\lambda}_{+}(q) = 0$. Suppose we have a periodic point $p$ of $f$ which satisfies the inequality ${\lambda}_{+}(p)-{\lambda}_{-}(p) < \tau$. Choose $\theta$ satisfying the inequality $${\lambda}_{+}(p)-{\lambda}_{-}(p) < \theta < \tau.$$ We will show that there is an $N > 0$ such that $p \in \mathcal{D}(N,\theta)$. Let $k$ be the period of $p$. As a direct consequence of the defining formulas for the extremal Lyapunov exponents of $p$ we have $${\lambda}_{+}(p) - {\lambda}_{-}(p) = \lim_{n \rightarrow \infty} \frac{\log \\|A^{kn}(p)\\| \\|(A^{kn}(p))^{-1}\\|}{kn}.$$ Chooose $M$ large enough that for $n \geq M$, $$\\|A^{kn}(p)\\| \\|(A^{kn}(p))^{-1}\\| \leq e^{kn\theta}.$$ Let $N$ = $k M$. Then, using the periodicity of $p$, for each $s \geq 1$ we have, $$\begin{aligned} \prod_{j = 0}^{s-1} \\|A^{N}(f^{jN}(p))\\| \\|(A^{N}(f^{jN}(p)))^{-1}\\| &= \prod_{j = 0}^{s-1} \\|A^{N}(f^{jkM}(p))\\| \\|(A^{N}(f^{jkM}(p)))^{-1}\\| \\ &= (\\|A^{N}(p)\\| \\|(A^{N}(p))^{-1}\\|)^{s} \\ &\leq e^{sN\theta}\end{aligned}$$ Since $A^{-kM}(p) = (A^{kM}(p))^{-1}$ and $(A^{-kM}(p))^{-1} = A^{kM}(p)$ we also have $$\prod_{j = 0}^{s-1} \\|A^{-N}(f^{-jN}(p))\\| \\|(A^{-N}(f^{-jN}(p)))^{-1}\\| \leq e^{sN\theta}$$ Hence $p \in \mathcal{D}(N,\theta)$. Since $\theta < \tau$, we conclude by Lemma \[extension\] that the extension $\hat{\eta}$ of $\eta$ is defined at $p$ and satisfies $A^{k}(p)[\hat{\eta}_{p}] = \hat{\eta}_{p}$. Thus $A^{k}(p)$ preserves an inner product on ${\mathbb R}^{d}$; we conclude that ${\lambda}_{+}(p) = 0$. This proves the first part of the lemma. By Kalinin’s exponent approximation theorem [@Kal] there is at least one periodic point $q$ such that ${\lambda}_{+}(q)- {\lambda}_{-}(q)< \tau$. This is because we have an ergodic $f$-invariant probability measure $\mu$ for which ${\lambda}_{+}(\mathcal{A},\mu) = 0$ and ${\lambda}_{-}(\mathcal{A},\mu) = 0$. By the first part of the lemma we must then have ${\lambda}_{+}(q) - {\lambda}_{-}(q) = 0$. This implies that ${\lambda}_{+}(q) = 0$ since $-{\lambda}_{-}(q) > 0$ (because ${\lambda}_{+}(x) > 0$ and $\mathcal{A}$ is an $SL(d,{\mathbb R})$-valued cocycle). A prospective strategy for finishing the proof is to show that if there is some periodic point $x \in \Sigma$ with ${\lambda}_{+}(x) - {\lambda}_{-}(x) \geq \tau$ then, letting $y$ be a periodic point such that ${\lambda}_{+}(y) = 0$, we can produce a new periodic point $p$ shadowing some mixture of the orbits of $x$ and $y$ such that $0 < {\lambda}_{+}(p) - {\lambda}_{-}(p) < \tau$. This would contradict Lemma \[smallperiodiczero\]. However the quantity ${\lambda}_{+}(p) - {\lambda}_{-}(p)$ is difficult to estimate from below; this is tied to the inherent difficulty of proving that cocycles have nonzero exponents. We will take an alternative approach which does not require us to directly estimate the Lyapunov exponents of the periodic point $p$ that we construct. We assume to a contradiction that there is a periodic point $x$ of $f$ with ${\lambda}_{+}(x) > 0$. Let $y$ be a periodic point given by Lemma \[smallperiodiczero\] such that ${\lambda}_{+}(y) = 0$. Choose $k$ such that both $x$ and $y$ are fixed by $f^{k}$. For each positive integer $m$ which is divisible by $k$, we will construct a new periodic point $p^{m}$ symbolically using orbit segments of $f$ which shadow the orbits of $x$ and $y$ and which satisfies $f^{u_{m}}(p^{m}) = p^{m}$ for a certain positive integer $u_{m}$. We will show that we can construct this sequence of periodic points $\{p^{m}\}$ such that the following two statements must hold simultaneously: First, we claim that as $m \rightarrow \infty$ we have $\\|A^{u_{m}}(p^{m})\\| \rightarrow \infty$. Second, we claim that there is a positive integer $N$ and a $\theta < \tau $ such that $p^{m} \in \mathcal{D}(N,\theta)$ for all $m$ large enough. This produces a contradiction: by Lemma \[extension\] $\hat{\eta}$ is defined and uniformly continuous on the compact set $\mathcal{D}(N,\theta)$. There is thus a constant $\gamma > 0$ such that for any vector $v \in {\mathbb R}^{d}$ and any $q \in \mathcal{D}(N,\theta)$ we have $$\gamma^{-1}\\|v\\|^{2} \leq \hat{\eta}_{q}(v,v) \leq \gamma \\|v\\|^{2}$$ Since $A^{u_{m}}(p^{m})[\hat{\eta}_{p^{m}}] = \hat{\eta}_{p^{m}}$, we thus have $$\begin{aligned} \\|A^{u_{m}}(p^{m})(v)\\|^{2} &\leq \gamma \cdot \hat{\eta}_{q}(A^{u_{m}}(p^{m})(v),A^{u_{m}}(p^{m})(v)) \\ &= \gamma \cdot \hat{\eta}_{q}(v,v) \\ &\leq \gamma^{2} \\|v\\|^{2}\end{aligned}$$ Thus $\\|A^{u_{m}}(p^{m})(v)\\| \leq \gamma$ for all large enough $m$, which contradicts the previous assertion that $\\|A^{u_{m}}(p^{m})\\| \rightarrow \infty$ as $m \rightarrow \infty$. We write $x = (x_{n})_{n \in {\mathbb Z}}$, $y = (y_{n})_{n \in {\mathbb Z}}$ in the shift coordinates on $\Sigma$. Recall that $Q$ denotes the matrix defining valid words for the subshift of finite type $\Sigma$. Since $\Sigma$ is a topologically mixing subshift of finite type there is an $M > 0$ such that for all $n \geq M$ the matrix $Q^{n}$ has all positive entries, meaning that any two letters in the alphabet $\{1, \dots, {\ell}\}$ can be connected by a valid word of length $n$. By increasing $k$ if necessary we may assume that $k \geq M$ and consequently $m \geq M$ since $m$ is a multiple of $k$. Let $b$ and $c$ be two positive integers whose values will be tuned later. Set $u_{m} = (2b + c + 1)m$. Recall that we restricted $m$ to be a multiple of $k$ so that $f^{m}(x) = x$ and $f^{m}(y) = y$. We define $p^{m} = (p^{m}_{n})_{n \in {\mathbb Z}}$ where the numbers $p^{m}_{n} \in \{1, \dots,{\ell}\}$ are chosen as follows: for $-bm \leq j \leq bm$ we set $p^{m}_{j} = y_{j}$. For $(b+1)m \leq j \leq (b+ c + 1)m$ we define $p^{m}_{j} = x_{j-(b+1)m}$. For $(b+1)m < j < (b+ c + 2)m $ we let the string $(p_{(b+1)m}^{m},\dots,p_{(b+ c +1)m}^{m})$ be any valid word in the alphabet of $\Sigma$ which is of length $m$ and connects $y_{m}$ to $x_{0}$. This is possible because we chose $m$ such that $m \geq M$. Similarly for $(b+ c +1)m \leq j \leq (b+ c +2)m$ we can fill in $m$ coordinates for $p^{m}_{j}$ which connect $x_{ (b+ c +1)m}$ to $y_{0}$ in the subshift $\Sigma$. We then define the rest of the coordinates of $p^{m}$ by declaring $p^{m}$ to be periodic with period $u_{m}$. Let ${\varepsilon}> 0$ be a small parameter which will be tuned later along with $b$ and $c$. For ease of notation in what follows we set ${\lambda}:= {\lambda}_{+}(x)$ and $\zeta = \sup_{z \in \Sigma} \log \\|A(z)\\| \\|A(z)^{-1}\\|$. We set $\chi := c({\lambda}- {\varepsilon}) - 2b{\varepsilon}- 2\zeta$ and fix some $\theta$ with $0 < \theta < \tau$ (for example $\theta = \tau/2$). We first show that if $\chi > 0$ then $\\|A^{u_{m}}(p^{m})\\| \rightarrow \infty$ as $m \rightarrow \infty$. We will address later the question of how to choose $b$, $c$ and ${\varepsilon}$ together to guarantee that $\chi > 0$ while also satisfying some other necessary conditions. We will need the following lemma, \[periodicapprox\] Let $p \in \Sigma$ be a periodic point for $f$ with extremal Lyapunov exponent ${\lambda}:={\lambda}_{+}(p)$ and let ${\varepsilon}> 0$ be given. Then there is a constant $C \geq 1$ independent of $n$ such that for every $n \geq 1$, if $q \in \Sigma$ satisfies $$\rho(f^{j}(p),f^{j}(q)) \leq \text{max}\{e^{-j\tau},e^{-(n-j)\tau}\}, \; 0 \leq j \leq n,$$ then $$C^{-1}e^{n({\lambda}-{\varepsilon})}\leq \\|A^{n}(q)\\| \leq Ce^{n({\lambda}+{\varepsilon})}.$$ This lemma follows from [@Kal Lemma 3.1] together with the observation that for a probability measure $\nu_{p}$ supported on a periodic point $p$, there is a uniform comparison between the ${\varepsilon}$-Lyapunov norm and the Euclidean norm which depends only on ${\varepsilon}$. Lemma 3.1 of [@Kal] then gives that there is a $\delta > 0$ and a constant $C' \geq 1$ such that for every $n \geq 1$, if $$\rho(f^{j}(p),f^{j}(q)) \leq \delta\cdot \text{max}\{e^{-j\tau},e^{-(n-j)\tau}\}, \; 0 \leq j \leq n,$$ then $$C'^{-1}e^{n({\lambda}-{\varepsilon})}\leq \\|A^{n}(q)\\| \leq C'e^{n({\lambda}+{\varepsilon})}.$$ To obtain our desired statement, fix $m$ large enough that $e^{-m\tau} < \delta$. Then for every $n \geq 1$, if $$\rho(f^{j}(p),f^{j}(q)) \leq\text{max}\{e^{-j\tau},e^{-(n-j)\tau}\}, \; 0 \leq j \leq n,$$ then $$\rho(f^{j}(p),f^{j}(q)) \leq \delta \cdot \text{max}\{e^{-j+m\tau},e^{-(n-j+m)\tau}\} , \; m\leq j \leq n-m,$$ and consequently $$C'^{-1}e^{(n-2m)({\lambda}-{\varepsilon})}\leq \\|A^{n-m}(f^{m}(q))\\| \leq C'e^{(n-2m)({\lambda}+{\varepsilon})}.$$ We conclude that $$C^{-1}e^{n({\lambda}-{\varepsilon})}\leq \\|A^{n}(q)\\| \leq Ce^{n({\lambda}+{\varepsilon})},$$ with $C = C' e^{2m({\lambda}+{\varepsilon})}\sup_{z \in \Sigma}\\|A^{m}(z)\\|\\|A^{m}(z)^{-1}\\|$. For $0 \leq j \leq bm$ we have $$\rho(f^{j}(y),f^{j}(p^{m})) \leq \text{max}\{e^{-j \tau}, e^{-(bm-j)\tau}\}.$$ For $0 \leq j \leq cm$, $$\rho(f^{j + (b+1)m}(x),f^{j + (b+1)m}(p^{m})) \leq \text{max}\{e^{-j \tau}, e^{-(cm-j)\tau}\}.$$ Finally for $0 \leq j \leq bm$ $$\rho(f^{j+(b+c+2)m}(y),f^{j+(b+c+2)m}(p^{m})) \leq \text{max}\{e^{-j \tau}, e^{-(bm-j)\tau}\}.$$ Apply Lemma \[periodicapprox\] to the periodic points $x$ and $y$ with the ${\varepsilon}> 0$ chosen above and let $C$ be the maximum of the two constants in the output of Lemma \[periodicapprox\] for $x$ and $y$ respectively. We have the lower bound using Lemma \[periodicapprox\], $$\begin{aligned} \\|A^{u_{m}}(p^{m})\\| &\geq \\|A^{cm}(f^{(b+1)m}(p^{m}))\\| \cdot \\|A^{bm}(p^{m})^{-1}\\|^{-1} \cdot \\|A^{km}(f^{bm}(p^{m}))^{-1}\\|^{-1} \\ &\cdot \\|A^{m}(f^{(b+c+1)m}(p^{m}))^{-1}\\|^{-1} \cdot \\|A^{bm}(f^{(b+c+1)m}(p^{m}))^{-1}\\|^{-1} \\ &\geq C^{-3}\exp((c({\lambda}- {\varepsilon}) - 2b{\varepsilon}- 2\zeta)m) \\ &= C^{-3}\exp(\chi m) \end{aligned}$$ If $\chi > 0$ then it follows that $\\|A^{u_{m}}(p^{m})\\| \rightarrow \infty$ as $m \rightarrow \infty$. It remains to prove the claim that there is an $N > 0$ such that for every $m \geq 1$ large enough we have $p^{m} \in \mathcal{D}(N,\theta)$. We will first show that there is an $N >0$ such that for $m \geq 1$ large enough and every $ s \geq 1$, $$\prod_{j=0}^{s-1} \\|A^{N}(f^{jN}(p^{m}))\\| \\|A^{N}(f^{jN}(p^{m}))^{-1}\\| \leq e^{s N \theta}.$$ We will prove this first set of inequalities using the fact that upon applying $f$ the periodic point $p^{m}$ first shadows the orbit of $y$ for $bm$ iterates, then the orbit of $x$ for $cm$ iterates, then again shadows the orbit of $y$ for another $bm$ iterates (with $2m$ transitioning iterates where $f$ does not necessarily shadow $x$ or $y$). The points $p^{m}$ are constructed such that if we apply $f^{-1}$ we get the same pattern of shadowing $y$ for $bm$ iterates, $x$ for $cm$ iterates, then $y$ again for $bm$ iterates. Hence it can be checked that the same proof given for the first set of inequalities will also prove the second set of inequalities, $$\prod_{j=0}^{s-1} \\|A^{-N}(f^{-jN}(p^{m}))\\| \\|A^{-N}(f^{-jN}(p^{m}))^{-1}\\| \leq e^{s N \theta}.$$ and thus establish that $p^{m} \in \mathcal{D}(N,\theta)$ for all $m$ large enough. We proceed with the proof of the first set of inequalities. The main idea of the proof is simple: the ratio $\frac{b}{c}$ controls the ratio of the number of iterates that $p^{m}$ shadows $y$ to the number of iterates that $p^{m}$ shadows $x$. We essentially showed above that if $c$ is large enough then the influence of the positive exponent of $x$ forces $\\|A^{u_{m}}(p^{m})\\| \rightarrow \infty$ as $m \rightarrow \infty$. We will show that once $b$ is chosen such that the ratio $\frac{b}{c}$ is large enough then the time that $p^{m}$ spends shadowing the point $y$ is sufficient to cancel enough of the influence of the Lyapunov exponents of $x$ to guarantee that $p^{m}$ lands in $\mathcal{D}(N,\theta)$ for some choice of $N$ that is independent of $m$. We now specify the parameters $b$, $c$, and ${\varepsilon}$. Recall that $\theta$ denotes a fixed number satisfying $0 < \theta < \tau$ and that we defined ${\lambda}= {\lambda}_{+}(x)$ and $\zeta = \sup_{z \in \Sigma} \log \\|A(z)\\| \\|A(z)^{-1}\\|$. We also set $\xi = {\lambda}_{+}(x) - {\lambda}_{-}(x)$. We choose the integer $c$ large enough that $c {\lambda}- 2\zeta > 0$. Having chosen $c$, we now choose $b$ large enough that the following three inequalities hold: first we require that $$\label{ineq1} \zeta\left(1-\frac{b}{b+1}\right) + \frac{\theta}{10} < \frac{9}{10}\theta.$$ Second we require that $$\label{ineq2} \frac{b}{b+1}\cdot \frac{\theta}{10} + \frac{\zeta}{b+1} +\left(1-\frac{b+1}{b+c+1}\right)\left(\xi+\frac{\theta}{10}\right) < \frac{9}{10}\theta.$$ Finally we require that $$\label{ineq3} \frac{b}{b+ c + 1}\cdot \frac{\theta}{10} + \frac{\zeta}{b+ c + 1} +\frac{c}{b+c+1}\left(\xi+\frac{\theta}{10}\right) + \zeta \left(1-\frac{b+c+1}{b+c+2}\right) < \frac{9}{10}\theta.$$ It’s easily verified that, for a fixed choice of $c$, each of these inequalities holds once $b$ is large enough. Finally, having chosen $b$ and $c$, we now choose ${\varepsilon}$ small enough that ${\varepsilon}\leq \theta/10$ and $\chi = c({\lambda}- {\varepsilon}) - 2b{\varepsilon}- 2\zeta > 0$. We note that since $\frac{9}{10}\theta \leq \theta - {\varepsilon}$, the three inequalities above also hold with $\theta - {\varepsilon}$ on the right hand side instead. The source of these three inequalities will become clear from the computations below. Let $J > 0$ be large enough that for $j \geq J$, $$\\|A^{j}(y)\\| \\|A^{j}(y)^{-1}\\| \leq e^{j{\varepsilon}}, \; \\|A^{j}(x)\\| \\|A^{j}(x)^{-1}\\| \leq e^{j(\xi + {\varepsilon})},$$ Choose $r$ to be a multiple of $k$ which satisfies $r \geq J$. Let $L := \text{Lip}(\log \\|A^{r}\\| \\|(A^{r})^{-1}\\|)$ be the Lipschitz constant of $\log \\|A^{r}\\| \\|(A^{r})^{-1}\\|$. We define a constant $$C:= 2\sum_{j=0}^{\infty}e^{-j\tau} = \frac{2}{1-e^{-\tau}}.$$ Since we only need a subsequence of periodic points $p^{m_{q}}$ such that there is some $N$ for which $p^{m_{q}} \in \mathcal{D}(N,\theta)$, we can further restrict $m$ to be divisible by $r$ so that $m = qr$ for some integer $q$. For $0 \leq s \leq bq$ we then have $$\begin{aligned} \log\left(\prod_{j = 0}^{s-1}\frac{\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\|}{\\|A^{r}(f^{jr}(y))\\| \\|A^{r}(f^{jr}(y))^{-1}\\|}\right) &\leq L\sum_{j=0}^{s-1}\rho(f^{jr}(p^{m}),f^{jr}(y)) \\ &\leq L \sum_{j=0}^{s-1}\text{max}\{e^{-jN \tau}, e^{-(km-jN)\tau}\} \\ &\leq L C,\end{aligned}$$ for the constant $C$ which is independent of $m$ and $s$. For $(b+1)q \leq s \leq (b+c+1)q$, we replace $y$ by $x$ in the above estimate $$\begin{aligned} \log\left(\prod_{j = (b+1)q}^{s-1}\right.&\left.\frac{\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\|}{\\|A^{r}(f^{jr}(x))\\| \\|A^{r}(f^{jr}(x))^{-1}\\|}\right) \\ &\leq \sum_{j= (b+1)q}^{s-1}L \rho(f^{jr}(p^{m}),f^{jr}(x)) \\ &= \sum_{j= 0}^{s-1-(b+1)q}L \rho(f^{jr +(b+1)q }(p^{m}),f^{jr + (b+1)q}(x))\\ &\leq LC.\end{aligned}$$ Similar estimates give for $(b+c+2)q \leq s \leq (2b+c+2)q$, $$\log\left(\prod_{j = (b+c+2)q}^{s-1}\frac{\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\|}{\\|A^{r}(f^{jr}(y))\\| \\|A^{r}(f^{jr}(y))^{-1}\\|}\right) \leq LC$$ We conclude, using the shadowing of $p^{m}$ along the orbit of $y$ for the first $bm$ iterates, that for $0 \leq s \leq bq$, $$\begin{aligned} \prod_{j=0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p_{m}))^{-1}\\| &\leq e^{LC}\left(\prod_{j= 0}^{s-1}\\|A^{r}(f^{jr}(y))\\| \\|A^{r}(f^{jr}(y))^{-1}\\|\right) \\ &\leq \exp(LC + rs{\varepsilon}) \\ &\leq \exp(LC + rs(\theta- {\varepsilon})),\end{aligned}$$ since $\theta > 2{\varepsilon}$. For $bq \leq s \leq (b+1)q$ we combine the above estimate with the crude bound $$\prod_{j=bq}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \leq \exp(\zeta r(s-bq)),$$ (where we recall that $\zeta = \sup_{z \in \Sigma} \\|A(z)\\| \cdot \\|A(z)^{-1}\\|$) to get $$\begin{aligned} \prod_{j=0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| &\leq \exp(LC + brq{\varepsilon}+ \zeta r(s-bq)) \\ &= \exp\left(LC + rs\left(\zeta\left(1-\frac{bq}{s}\right) + \frac{bq}{s}{\varepsilon}\right)\right) \\ &\leq \exp\left(LC + rs\left(\zeta\left(1-\frac{b}{b+1}\right) + {\varepsilon}\right)\right) \\ &\leq \exp\left(LC + rs(\theta - {\varepsilon})\right),\end{aligned}$$ where we’ve used here the inequality $\eqref{ineq1}$ together with the fact that ${\varepsilon}\leq \theta/10$. We similarly use the shadowing of $p^{m}$ along the orbit of $x$ for $cm$ iterates and combine this with the estimates above to obtain for $(b+1)q \leq s \leq (b+c+1)q$, $$\begin{aligned} \prod_{j=0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| &\leq \exp(2LC + brq{\varepsilon}+ \zeta rq )\\ &\cdot \prod_{j= (b+1)q}^{s-1}\\|A^{r}(f^{jr+M}(x))\\| \\|A^{r}(f^{jr+M}(x))^{-1}\\| \\ &\leq \exp\left(2LC + brq{\varepsilon}+ \zeta rq + r\left(s-(b+1)q\right)(\xi+{\varepsilon})\right) \\ &\leq \exp\left(2LC +s(\theta - {\varepsilon}) \right)\end{aligned}$$ To obtain the final inequality above we use the fact that $(b+1)q \leq s \leq (b+c+1)q$ to write $$\begin{aligned} brq{\varepsilon}+ \zeta rq + r\left(s-(b+1)q\right)(\xi+{\varepsilon})&= rs\left(\frac{bq}{s}{\varepsilon}+ \zeta\frac{q}{s} +\left(1-\frac{(b+1)q}{s}\right)(\xi+{\varepsilon})\right) \\ &\leq rs\left(\frac{b}{b+1}{\varepsilon}+ \frac{\zeta}{b+1} +\left(1-\frac{b+1}{b+c+1}\right)(\xi+{\varepsilon})\right),\end{aligned}$$ and from the last line we use the inequality together with the fact that ${\varepsilon}\leq \theta/10$ to obtain the desired inequality. For $(b+c + 1)q \leq s \leq (b+c +2)q$ we again use the crude bound $$\prod_{j=(b+c+1)q}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \leq \exp(\zeta r(s-(b+c+1)q)),$$ to now obtain (combining also with the above estimates) $$\begin{aligned} \prod_{j=0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| &\leq \exp(2LC + brq{\varepsilon}+ \zeta rq \\ &+ rcq(\xi+{\varepsilon}) + \zeta r(s-(b+c+1)q)),\end{aligned}$$ We must prove again that the quantity in the exponent on the right is bounded above by $2LC + rs(\theta - {\varepsilon})$. This involves an application of the inequality in a manner similar to the previous verification. Using $(b+c + 1)q \leq s \leq (b+c +2)q$ we have $$\begin{aligned} brq{\varepsilon}+ \zeta rq &+ rcq(\xi+{\varepsilon}) + \zeta r(s-(b+c+1)q) \\ &= rs\left(\frac{bq}{s}{\varepsilon}+ \zeta\frac{q}{s} +\frac{cq}{s}(\xi+{\varepsilon}) + \zeta \left(1-\frac{(b+c+1)q}{s}\right) \right) \\ &\leq \left(\frac{b}{b+ c + 1}{\varepsilon}+ \frac{\zeta}{b+ c + 1} +\frac{c}{b+c+1}(\xi+{\varepsilon}) + \zeta \left(1-\frac{b+c+1}{b+c+2}\right) \right)\end{aligned}$$ The desired bound then comes from combining the inequality with the fact that ${\varepsilon}\leq \theta/10$. Finally for $(b+c+2)q \leq s \leq (2b+c+2)q$ we use the approximation by $y$ again and combine this with the previous estimates to obtain $$\begin{aligned} \prod_{j = 0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| &\\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \\ &\leq \exp(3LC + (b+c+2)q(\theta - {\varepsilon}) + r(s-(b+c+2)q){\varepsilon}) \\ &\leq \exp(3LC + rs(\theta - {\varepsilon}) ),\end{aligned}$$ once again because $\theta > 2{\varepsilon}$. Define the integer $v_{m}$ by $v_{m} := \frac{u_{m}}{r}$. Since $f^{u_{m}}(p^{m}) = p^{m}$, we can extend our arguments to $s \geq (2b+c+2)q = v_{m}$ using the periodicity of $p^{m}$: for $s \geq v_{m}$ we write $s = h v_{m} + s'$ for some positive integers $h$ and $s'$. Then using the above arguments, $$\begin{aligned} \prod_{j = 0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| &= \left(\prod_{j = 0}^{v_{m}-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \right)^{h} \\ &\cdot \prod_{j = 0}^{s'-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \\ &\leq \exp(3LC + rs(\theta-{\varepsilon}))\end{aligned}$$ Hence we have succeeded in showing that for every $s \geq 1$, $$\prod_{j = 0}^{s-1}\\|A^{r}(f^{jr}(p^{m}))\\| \\|A^{r}(f^{jr}(p^{m}))^{-1}\\| \leq \exp(3LC + rs(\theta-{\varepsilon})).$$ We use one final trick: let $N = rt$ for some integer $t$ chosen large enough that $N{\varepsilon}> 3LC$. Then it’s easy to see that as a consequence of the above inequality we also have $$\prod_{j = 0}^{s-1}\\|A^{N}(f^{jN}(p^{m}))\\| \\|A^{N}(f^{jN}(p^{m}))^{-1}\\| \leq \exp(3LC + Ns(\theta-{\varepsilon})),$$ for every $s \geq 1$. But we chose $N$ such that $$3LC + Ns(\theta-{\varepsilon}) < N{\varepsilon}+ Ns(\theta-{\varepsilon}) \leq Ns\theta$$ Thus we conclude that $p^{m} \in \mathcal{D}(N,\theta)$ for all large enough $m$. Conclusion ---------- By Proposition \[periodiczero\], ${\lambda}_{+}(p) = 0$ for every periodic point $p$ of $f$. We conclude by [@Kal Theorem 1.4] that for every $f$-invariant ergodic probability measure $\nu$ on $\Sigma$ we must have ${\lambda}_{+}(\mathcal{A},\nu) = -{\lambda}_{-}(\mathcal{A},\nu) = 0$. We then apply [@KS Proposition 4.11] with ${\varepsilon}< \tau$ to obtain a uniform estimate $$\\|A^{n}(x)\\| \\|A^{n}(x)^{-1}\\| \leq Ce^{{\varepsilon}\|n\|}, \; \; \text{for every $x \in \Sigma$, $n \in {\mathbb Z}$},$$ for some constant $C \geq 1$. We conclude that there is $N > 0$, $\theta < \tau$ such that $ \mathcal{D}(N,\theta) = \Sigma$. Hence the continuous extension $\hat{\eta}$ of $\eta$ from Lemma \[extension\] is the desired continuous family of inner products invariant under $\mathcal{A}$. Since $\hat{\eta}$ is invariant under the stable and unstable holonomies $H^{s}$ and $H^{u}$ on $\mathcal{D}(N,\theta) = \Sigma$ and by Proposition \[existenceholonomies\] these holonomies are uniformly Lipschitz, we conclude that $\hat{\eta}$ is Lipschitz continuous with respect to the distance $\rho$, as desired.	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'auto\_generated.bib' title: 'Search for direct pair production of supersymmetric partners to the lepton in proton-proton collisions at $\sqrt{s}=13\TeV$' --- =1 \[1\] Introduction {#sec:intro} ============ Supersymmetry (SUSY) [@Ramond:1971gb; @Golfand:1971iw; @Neveu:1971rx; @Volkov:1972jx; @Wess:1973kz; @Wess:1974tw; @Fayet:1974pd; @Nilles:1983ge] is a possible extension of the standard model (SM) of particle physics, characterized by the presence of superpartners for SM particles. The superpartners have the same quantum numbers as their SM counterparts, except for the spin, which differs by half a unit. One appealing feature of SUSY is that the cancellation of quadratic divergences in quantum corrections to the Higgs boson mass from SM particles and their superpartners could resolve the fine tuning problem [@Gildener:1976ai; @Veltman:1976rt; @'tHooft:1979bh; @Witten:1981nf]. Another feature is that the lightest supersymmetric particle (LSP) is stable in SUSY models with $R$-parity conservation [@Farrar:1978xj], and could be a dark matter (DM) candidate [@Goldberg:1983nd; @Ellis:1983ew; @darkmatter]. The hypothetical superpartner of the lepton, the slepton (), is the focus of the search reported in this paper. Supersymmetric models where a light is the next-to-lightest supersymmetric particle are well motivated in early universe -neutralino coannihilation models that can accommodate the observed DM relic density [@WMAP; @Griest:1990kh; @Vasquez; @King; @Battaglia:2001zp; @Arnowitt:2008bz]. The existence of a light would enhance the rate of production of final states with leptons in collider experiments [@Belanger:2012jn; @Arganda:2018hdn]. In this analysis, we study the simplified model [@Alwall:2008ag; @Alwall:2008va; @Alves:2011wf] of direct pair production shown in Fig. \[fig:diagram\]. We assume that the decays to a lepton and , the lightest neutralino, which is the LSP in this model. The search is challenging because of the extremely small production cross section expected for this signal, as well as the large backgrounds. The most sensitive previous searches for direct pair production were performed at the CERN LEP collider [@Heister:2001nk; @Abdallah:2003xe; @Achard:2003ge; @Abbiendi:2003ji], excluding masses at 95% confidence level () up to $\approx$90for neutralino masses up to 80in some models. At the LHC, the ATLAS [@Aad:2014yka; @Aad:2015eda] and CMS [@SUS14022] Collaborations have also performed searches for direct pair production using 8data, and the CMS Collaboration has reported a search for direct pair production in an initial sample of 35.9at 13collected in 2016 [@Sirunyan:2018vig]. This paper presents a significant improvement in search sensitivity, which was limited by the small signal production rates, through the incorporation of improved analysis techniques and the inclusion of the data collected in 2017. The data used correspond to a total integrated luminosity of 77.2. Events with two leptons are used. We consider both hadronic and leptonic decay modes of the lepton, in which it decays to one or more hadrons and a neutrino, or to an electron or muon and two neutrinos, respectively. Independent analyses are carried out in the final states with two hadronically decaying leptons ([$\tauh\tauh$]{}) and with one and an electron or a muon ([$\ell\tauh$]{}, where $\ell = \Pe$ or ). The presence of missing transverse momentum, which can originate from stable neutralinos as well as neutrinos from lepton decays, provides an important source of discriminating power between signal and background. We have introduced several improvements with respect to the analysis presented in Ref. [@Sirunyan:2018vig] that are applied to both 2016 and 2017 data. We make use of dedicated machine learning techniques to enhance the search sensitivity. These include the incorporation of an improved selection method that makes use of a deep neural network (DNN) for the [$\tauh\tauh$]{}analysis, and of a boosted decision tree (BDT) for event selection in the [$\ell\tauh$]{}analyses. Improvements have also been made to the background-estimation techniques and to the search region (SR) definitions. The incorporation of these enhancements is expected to improve the search sensitivity by up to 50%, where the figure of merit considered is the 95% upper limit on the cross section for pair production obtained with the data collected in 2016. The improvement is less significant than expected, since it is found that the estimated signal acceptance is reduced when the fast detector simulation that was previously used to model signal events is replaced in this search with the more realistic, full -based detector simulation [@geant4]. Differences in the signal acceptance for the fast and more accurate full detector simulations are mainly caused by differences in the reconstructed visible transverse momentum (), which is found to have larger values in the case of the fast simulation. We consider the superpartners of both left- and right-handed leptons, [$\PSGt_{\mathrm{L}}$]{}and [$\PSGt_{\mathrm{R}}$]{}. The cross section for [$\PSGt_{\mathrm{L}}$]{}pair production is expected to be about a factor of three larger than for [$\PSGt_{\mathrm{R}}$]{}pairs [@Fuks:2013lya]. The experimental acceptance is also expected to be different for left- and right-handed assignments because of the differences in the polarization of the leptons produced in [$\PSGt_{\mathrm{L}}$]{}and [$\PSGt_{\mathrm{R}}$]{}decays. The decay products of hadronically and leptonically decaying leptons originating from [$\PSGt_{\mathrm{R}}$]{}decays are predicted to have larger and smaller , respectively, than those originating from [$\PSGt_{\mathrm{L}}$]{}decays. Two simplified models are studied for direct pair production. One model involves production of only [$\PSGt_{\mathrm{L}}$]{}pairs and the other is for the degenerate case in which both [$\PSGt_{\mathrm{L}}$]{}and [$\PSGt_{\mathrm{R}}$]{}pairs are produced. No mixing is introduced between left- and right-handed states. We study models with masses ranging from 90 to 200. The LEP limits [@Heister:2001nk; @Abdallah:2003xe; @Achard:2003ge; @Abbiendi:2003ji] place strong constraints on the allowed values of the mass below this range, while the search sensitivity for masses above this range is low as a result of the decrease in production cross section with increased mass. We also consider different assumptions for the mass, namely 1, 10, and 20. The search sensitivity decreases when the mass difference between the and becomes small, since the visible decay products in such cases have lower momentum, resulting in a loss of experimental acceptance for such signals. ![\[fig:diagram\] Diagram for direct pair production, followed by decay of each to a lepton and a .](Figure_001.pdf){width="50.00000%"} The CMS detector {#sec:detector} ================ =800 The central feature of the CMS apparatus is a superconducting solenoid of 6 internal diameter, providing a magnetic field of 3.8. A silicon pixel and strip tracker, a lead tungstate crystal electromagnetic calorimeter (ECAL), and a brass and scintillator hadron calorimeter, each composed of a barrel and two endcap sections, reside within the solenoid volume. Forward calorimeters extend the pseudorapidity ($\eta$) coverage provided by the barrel and endcap detectors. Muons are detected in gas-ionization chambers embedded in the steel flux-return yoke outside the solenoid. Events of interest are selected using a two-tiered trigger system [@Khachatryan:2016bia]. The first level, composed of custom hardware processors, uses information from the calorimeters and muon detectors to select events at a rate of around 100 within a time interval of less than 4. The second level, known as the high-level trigger, consists of a farm of processors running a version of the full event reconstruction software optimized for fast processing, which reduces the event rate to about 1 before data storage. A more detailed description of the CMS detector, together with definitions of the coordinate system and kinematic variables, can be found in Ref. [@Chatrchyan:2008zzk]. Event reconstruction and simulation {#sec:evtreco} =================================== The event reconstruction uses a particle-flow (PF) algorithm [@Sirunyan:2017ulk] that combines information from the tracker, calorimeter, and muon systems to identify charged and neutral hadrons, photons, electrons, and muons in an event. The missing transverse momentum vector, , is computed as the negative of the vector sum of the of all PF candidates reconstructed in an event, and its magnitude is used in the search as a discriminator between signal and SM background. Events selected for the search are required to pass filters [@Sirunyan:2019kia] designed to remove detector- and beam-related backgrounds, and must have at least one reconstructed vertex. Usually, more than one such vertex is reconstructed because of pileup, , multiple proton-proton ($\Pp\Pp$) collisions within the same or neighboring bunch crossings. The mean number of interactions per bunch crossing was 27 in 2016, and increased to 37 in 2017, assuming a total inelastic $\Pp\Pp$ cross section of 80. The reconstructed vertex with the largest value in summed object $\pt^2$ is selected to be the primary $\Pp\Pp$ interaction vertex (PV). These objects are defined by tracks associated with a given vertex that are clustered using a jet finding algorithm [@Cacciari:2008gp; @Cacciari:2011ma], and a more restricted form of the vector missing transverse momentum that is calculated from these track-based jets. Charged particles that originate from the PV, photons, and neutral hadrons are clustered into jets using the anti-algorithm [@Cacciari:2008gp] with a distance parameter of 0.4, as implemented in the package [@Cacciari:2011ma]. The jet energies are corrected to account for the contribution from pileup interactions and to compensate for variations in the detector response [@Cacciari:2011ma; @pileup]. To mitigate issues related to noise in the ECAL endcaps that led to significantly worse modeling of the distribution, particularly for events with large values of in 2017 data, PF candidates that are clustered in jets in $2.65 < \abs{\eta} < 3.14$ with uncorrected $\pt<50\GeV$ are not used in the calculation of in 2017 data and simulation. Disagreements between the distributions in data and simulation ranging up to $>$100% for $50 < \ptmiss < 170\GeV$ in DY+jets events, in which large values of arise mainly from mismeasurements, are reduced by this modification of the calculation. The modified distributions in simulated events and data agree within uncertainties. Jets in the search are required to have their axes within the tracker volume of $\abs{\eta} < 2.4$. For the [$\tauh\tauh$]{}analysis, we use jets with $\pt>30\GeV$, while for the [$\ell\tauh$]{}analyses, we veto events containing jets with $\pt>20$to provide efficient background rejection. Jets are required to be separated in $\eta$ and azimuthal angle ($\phi$) by ${\ensuremath{\Delta R}\xspace}\equiv \sqrt{\smash[b]{(\Delta\eta)^2 + (\Delta\phi)^2} } > 0.4$ from electron, muon, or candidates in order to minimize double counting of objects. Jets originating from the hadronization of quarks are “tagged" in the [$\tauh\tauh$]{}analysis through the DNN-based combined secondary vertex algorithm (DeepCSV) [@Sirunyan:2017ezt] to reject events with quark jets that are likely to originate from backgrounds with top quarks. The efficiency for tagging quarks originating from top quark decays is about 84%, while the misidentification rates for jets from charm quarks, and from light quarks or gluons, are about 41 and 11%, respectively. In the [$\ell\tauh$]{}analyses, the CSVv2 tagger [@Sirunyan:2017ezt] is used to identify quark jets for the selection of background-enriched control regions (CRs). The working point that is used corresponds to an efficiency of 63% and misidentification rates of 12 and 0.9% for jets from charm quarks and light quarks or gluons, respectively. Electron candidates are reconstructed by first matching reconstructed tracks to clusters of energy deposited in the ECAL. Selections based on the spatial distribution of the shower, track–cluster matching criteria, and consistency between the cluster energy and the track momentum are then used in the identification of electron candidates [@Khachatryan:2015hwa]. Muon candidates are reconstructed by requiring reconstructed tracks in the muon detector to be matched to the tracks found in the inner tracker [@Sirunyan:2018]. We require the origin of electron and muon candidates to be consistent with the PV. Restrictions are imposed on the magnitude of the impact parameters of their tracks relative to the PV in the transverse plane ([$d_{xy}$]{}), and on the longitudinal displacement ([$d_{z}$]{}) of the point of closest approach. To ensure that electron or muon candidates are isolated from jet activity, we define a relative isolation quantity ([$I_{\text{rel}}$]{}) as the ratio of the scalar sum of hadron and photon PF candidates, in an $\eta$-$\phi$ cone of radius 0.3 or 0.4 around the candidate electron or muon, to the candidate , requiring it to be below an upper bound appropriate for the selection. The quantity [$I_{\text{rel}}$]{}is adjusted to account for the contributions of particles originating from pileup interactions. The electron and muon selection criteria applied in the analysis are the same as those described in Ref. [@Sirunyan:2018vig]. The candidates are reconstructed using the CMS hadrons-plus-strips algorithm [@Sirunyan:2018pgf]. The constituents of the reconstructed jets are used to identify individual lepton decay modes with one charged hadron and up to two neutral pions, or three charged hadrons. The candidate momentum is determined from the reconstructed visible lepton decay products. The presence of extra particles within the jet that are incompatible with the reconstructed decay mode is used as a criterion to discriminate jets from decays. A multivariate-analysis (MVA) based discriminant [@Sirunyan:2018pgf], which contains isolation as well as lifetime information, is used to suppress the rate for quark and gluon jets to be misidentified as candidates. We employ a relaxed (“very loose") working point of this discriminant as a preselection requirement for the candidates selected in the [$\tauh\tauh$]{}analysis, as well as in the extrapolation used to estimate the contributions of events to the background in which quark or gluon jets are misidentified as candidates. This working point corresponds to an efficiency of $\approx$70% for a genuine , and a misidentification rate of $\approx$1% for quark or gluon jets. A DNN is used to improve the discrimination of signal candidates from background, as discussed in more detail below. Two working points are used in the [$\ell\tauh$]{}analysis: a “very tight" working point for selecting signal candidates that provides stringent background rejection, and a “loose" working point for the extrapolation procedure to estimate the misidentified background that provides higher efficiency and less background rejection. These working points, respectively, typically have efficiencies close to 45 and 67% for a genuine , with misidentification rates of $\approx$0.2 and 1% for quark or gluon jets. Electrons and muons misidentified as a are suppressed via criteria specifically developed for this purpose that are based on the consistency of information from the tracker, calorimeters, and muon detectors [@Sirunyan:2018pgf]. =800 The dominant background in the [$\tauh\tauh$]{}final state originates from misidentification of jets as candidates, mainly in SM events exclusively comprising jets produced through the strong interaction of quantum chromodynamics (QCD). These are referred to as QCD multijet events in what follows. To further improve the suppression of this background while retaining high signal efficiency, we have pursued a new approach for isolation in the [$\tauh\tauh$]{}analysis that is based upon the application of a DNN that is fed information about the properties of PF candidates within an isolation cone with ${\ensuremath{\Delta R}\xspace}< 0.5$ around the candidate. We refer to this as “Deep Particle Flow" (DeepPF) isolation. Charged PF candidates consistent with having originated from the PV, photon candidates, and neutral hadron candidates with $\pt > 0.5$, 1, and 1.25, respectively, provide the inputs to the DeepPF algorithm. The list of observables incorporated for each PF candidate includes its relative to the jet, ${\ensuremath{\Delta R}\xspace}$ between the candidate and , particle type, track quality information, and [$d_{xy}$]{}, [$d_{z}$]{}and their uncertainties, $\sigma({\ensuremath{d_{xy}}\xspace})$ and $\sigma({\ensuremath{d_{z}}\xspace})$. A convolutional DNN [@Lecun98gradient-basedlearning] is trained with simulated signal and background events. Signal candidates are those that are matched to generator-level leptons from a mixture of processes that give rise to genuine leptons. Background candidates that fail the matching are taken from simulated [$\PW$+jets]{}and QCD multijet events. The DeepPF discriminator value is obtained by averaging the DNN output with the nominal MVA-based discriminant described above. The working point for DeepPF isolation is chosen to maintain a constant efficiency of $\approx$50%, 56%, and 56% as a function of for the three respective decay modes: one charged hadron, one charged hadron with neutral pions, and three charged hadrons. Since the candidate distribution in signal events depends on the and masses, this choice of discriminator and working points allows us to maintain high efficiency for pair production signals under a large range of mass hypotheses. The overall misidentification rate for jets not originating from leptons ranges from 0.15% to 0.4% depending on and decay mode. Significant contributions to the SM background originate from Drell–Yan+jets (DY+jets), [$\PW$+jets]{}, , and diboson processes, as well as from QCD multijet events, where DY corresponds to processes such as $\qqbar\to\ell^{+}\ell^{-}$. Smaller contributions arise from single top quark production and rare SM processes, such as triboson and Higgs boson production, and top quark pair production in association with vector bosons. We rely on a combination of measurements in data CRs and Monte Carlo (MC) simulation to estimate contributions of each source of background. The MC simulation is also used to model the signal. =800 The version 2.3.3 and 2.4.2 event generators [@Alwall:2014hca] are used at leading order (LO) precision to generate simulated [$\PW$+jets]{}and DY+jets events with up to 4 additional partons for the analysis of 2016 and 2017 data, respectively. Exclusive event samples binned in jet multiplicity are used to enhance the statistical power of the simulation at higher values of jet multiplicity that are relevant to the phase space probed by this search. Production of top quark pairs, diboson and triboson events, and rare SM processes, such as single top quarks or top quark pairs associated with bosons, are generated at next-to-leading order (NLO) precision with and v2 [@Nason:2004rx; @Frixione:2007vw; @Alioli:2010xd; @Re:2010bp]. Showering and hadronization of partons are carried out using the 8.205 and 8.230 packages [@Sjostrand:2014zea] for the 2016 and 2017 analyses, respectively, while a detailed simulation of the CMS detector is based on the [@geant4] package. Finally, uncertainties in renormalization and factorization scale, and parton distribution functions (PDFs) have been obtained using the <span style="font-variant:small-caps;">SysCalc</span> package [@Kalogeropoulos:2018cke]. Models of direct pair production are generated with at LO precision up to the production of leptons, with their decay modeled by 8.212 and 8.230 for the analysis of 2016 and 2017 data, respectively. The CUETP8M1 [@Khachatryan:2015pea] (CUETP8M2T4 [@CMS-PAS-TOP-16-021] for ) and CP5 [@Sirunyan:2019dfx] underlying-event tunes are used with for the 2016 and 2017 analyses, respectively. The 2016 analysis uses the NNPDF3.0LO [@Ball:2014uwa] set of PDFs in generating [$\PW$+jets]{}, DY+jets, and signal events, while the NNPDF3.0NLO PDFs are used for other processes. The NNPDF3.1NLO PDFs are used for all simulated events in the 2017 analysis. Simulated events are reweighted to match the pileup profile observed in data. Differences between data and simulation in electron, muon, and identification and isolation efficiencies, jet, electron, muon, and energy scales, and tagging efficiency are taken into account by applying scale factors to the simulation. We improve the modeling of initial-state radiation (ISR) in simulated signal events by reweighting the $\pt^{\mathrm{ISR}}$ distribution, where $\pt^{\mathrm{ISR}}$ corresponds to the total transverse momentum of the system of SUSY particles. This reweighting procedure is based on studies of the of bosons [@Chatrchyan:2013xna]. The signal production cross sections are calculated at NLO using next-to-leading logarithmic (NLL) soft-gluon resummations [@Fuks:2013lya]. The most precise calculated cross sections available are used to normalize the simulated SM background samples, often corresponding to next-to-next-to-leading order accuracy. Event selection {#sec:evtsel} =============== The search strategy in the [$\tauh\tauh$]{}final state relies on a cut-and-count analysis based on the SRs described below in Section \[sec:tautausel\], while for the [$\ell\tauh$]{}final states we make use of BDTs to discriminate between signal and background as described in Section \[sec:leptausel\]. The data used in this search are selected through triggers that require the presence of isolated electrons, muons, candidates, or . The data used for the [$\tauh\tauh$]{}analysis are collected with two sets of triggers. Events with $\ptmiss < 200$are selected using a trigger that requires the presence of two candidates, each with $\pt>35$ and $>$40in 2016 and 2017 data, respectively. We gain up to 7% additional signal efficiency for events with $\ptmiss>200$with the help of a trigger that requires the presence of substantial , with a threshold varying between 100 and 140during the 2016 and 2017 data-taking periods. For the [$\Pe\tauh$]{}final state, the trigger relies on the presence of an isolated electron satisfying stringent identification criteria and passing $\pt>25$ or $>$35in 2016 and 2017 data, respectively. For the [$\PGm\tauh$]{}final state, the trigger is based on the presence of an isolated muon with $\pt>24$ and $>$27in 2016 and 2017 data, respectively. Trigger efficiencies are measured in data and simulation. In addition to corrections mentioned in Section \[sec:evtreco\], we apply scale factors to the simulation to account for any discrepancies in trigger efficiency with data. These scale factors are parameterized in the and $\eta$ of the reconstructed electron, muon, or candidates, or the reconstructed for events selected using triggers. Event selection and search regions in the tau-tau final state {#sec:tautausel} ------------------------------------------------------------- Beyond the trigger selection, the baseline event selection for the [$\tauh\tauh$]{}analysis requires the presence of exactly two isolated candidates of opposite charge, satisfying the DeepPF selection described in Section \[sec:evtreco\], with $\abs{\eta}<2.3$ and $\pt > 40$ and $>$45in the 2016 and 2017 analysis, respectively, as well as no additional candidates with $\pt > 30\GeV$ satisfying the very loose working point of the MVA-based discriminant. We veto events with additional electrons or muons with $\pt>20\GeV$ and $\abs{\eta} < 2.5$ or $<$2.4 for electrons and muons, respectively, and reject any events with a -tagged jet to suppress top quark backgrounds. A requirement of $\abs{\Delta\phi(\tauh^{(1)},\tauh^{(2)})} > 1.5$ helps to suppress the DY+jets background, while retaining high signal efficiency. Finally, we require $\ptmiss>50\GeV$ to suppress the QCD multijet background. =900 The removal of low-jets in the forward ECAL region from the calculation in 2017 (see Section \[sec:evtreco\]) causes the background originating from DY+jets and other sources to increase in the SRs, since events with low-jet activity in that region are assigned larger values of reconstructed . We recover some of the corresponding loss in sensitivity in the 2017 analysis by placing an upper bound of 50on the scalar sum of low-jets excluded from the calculation ($\HT^{\text{low}}$). This restriction reduces the impact of background events with significant low-jet activity in the forward region, for which the would be overestimated. To ensure that the efficiency of this requirement is correctly estimated in simulation, a [$\PZ\to\PGm^{+}\PGm^{-}$]{}CR is used to extract correction factors for the $\HT^{\text{low}}$ distribution in simulation that account for discrepancies with the distribution observed in data. The correction factors range from 0.8 for $\HT^{\text{low}}<10\GeV$ to 1.4 for $\HT^{\text{low}}>60\GeV$. In addition, to avoid effects related to jet mismeasurement that can contribute to spurious , we require the to have a minimum separation of 0.25 in $\abs{\Delta \phi}$ from jets with $\pt > 30\GeV$ and $\abs{\eta} < 2.4$, as well as from those with uncorrected $\pt > 50\GeV$ in the region $2.4 < \abs{\eta} < 3.14$. Events satisfying the baseline selection criteria are subdivided into exclusive SRs using several discriminants. To improve the discrimination of signal from SM background, we take advantage of the expected presence of two in the final state of signal events and their contribution to . Their presence skews the correlations between and the reconstructed leptons to be different from background processes, even for those backgrounds with genuine . These differences can be exploited by mass observables calculated from the reconstructed lepton transverse momenta and to provide discrimination of signal from background. For a particle decaying to a visible and an invisible particle, the transverse mass () calculated from the of the visible decay products should have a kinematic endpoint at the mass of the parent particle. Assuming that the corresponds to the of the invisible particle, we calculate the observable for the visible particle q and the invisible particle as follows: $$\mT(\mathrm{q}, \ptvecmiss) \equiv \sqrt{2 \pt^{\mathrm{q}} \ptmiss [1 - \cos \Delta\phi(\ptvec^{\mathrm{q}}, \ptvecmiss)]}.\label{eq:MT}$$ We use as a discriminant the sum of the transverse masses calculated for each with , [$\Sigma\mT$]{}, given by $${\ensuremath{\Sigma\mT}\xspace}= \mT({\tauh^{(1)},\ptvecmiss}) + \mT({\tauh^{(2)},\ptvecmiss}).$$ Another variable found to be useful in the discrimination of signal from background is the “stransverse mass" [@MT2variable; @MT2variable2; @MT2variable3]. This mass variable is a generalization of in the case of multiple invisible particles. It serves as an estimator of the mass of pair-produced particles when both particles decay to a final state containing the same invisible particle. It is given by: $$\mTii = \min_{\ptvec^{\mathrm{X}(1)} + \ptvec^{\mathrm{X}(2)} = \ptvecmiss} \left[ \max \left( \mT^{(1)} , \mT^{(2)} \right) \right], \label{eq:MT2}$$ where $\ptvec^{\mathrm{X}(i)}$ (with $i$=1, 2) are the unknown transverse momenta of the two undetected particles, X(1) and X(2), corresponding to the neutralinos in our signal models, and $\mT^{(i)}$ are the transverse masses obtained by pairing either of the two invisible particles with one of the two leptons. The minimization ($\min$) is over the possible momenta of the invisible particles, taken to be massless, which are constrained to add up to the in the event. For direct pair production, with each decaying to a lepton and a , should be correlated with the mass difference between the and . A large value of is thus common in signal events for models with larger masses and relatively rare in SM background events. The SR definitions for the [$\tauh\tauh$]{}analysis, shown in Table \[tab:sr\], are based on a cut-and-count analysis of the sample satisfying the baseline selections. The regions are defined through criteria imposed on , [$\Sigma\mT$]{}, and the number of reconstructed jets in an event, [$N_{\text{j}}$]{}. The [$\Sigma\mT$]{}and distributions of events in the [$\tauh\tauh$]{}final state surviving the baseline selections are shown in Fig. \[fig:srvariables\_tautau\]. The distributions obtained for 2016 and 2017 data are combined. Separate sets of simulated events are used to model signal and background events in 2016 and 2017 data using the methods described in Section \[sec:evtreco\]. In all distributions, the last bin includes overflow events. After applying a minimum requirement of $\mTii>25\GeV$ in all SRs, we subdivide events into low (25–50) and high (${>}50\GeV$) regions, to improve the sensitivity to lower and higher mass signals, respectively. For each region, the [$\Sigma\mT$]{}distribution is exploited to provide sensitivity for a large range of mass signals. We define three bins in [$\Sigma\mT$]{}: 200–250, 250–300, and $>$300. Finally, we subdivide events in each and [$\Sigma\mT$]{}region into the categories ${\ensuremath{N_{\text{j}}}\xspace}= 0$ and ${\ensuremath{N_{\text{j}}}\xspace}\geq 1$. This binning is beneficial as background events passing the SR kinematic selections are largely characterized by additional jet activity, while signal contains very few additional jets. The 0-jet category therefore provides nearly background-free SRs. However, we retain the SRs with ${\ensuremath{N_{\text{j}}}\xspace}\geq 1$ that are also expected to contain signal events with ISR or pileup jets. \[tab:sr\] -------------------------- -- --- --------- --- --------- --- --------- -- --- --------- --- --------- --- --------- -- \[\] [$\Sigma\mT$]{}\[\] \[\] [$N_{\text{j}}$]{} 0 $\geq$1 0 $\geq$1 0 $\geq$1 0 $\geq$1 0 $\geq$1 0 $\geq$1 -------------------------- -- --- --------- --- --------- --- --------- -- --- --------- --- --------- --- --------- -- ![\[fig:srvariables\_tautau\] Distributions in [$\Sigma\mT$]{}() and () for events in the combined 2016 and 2017 data sets passing the baseline selection in the [$\tauh\tauh$]{}final state, along with the corresponding prediction for the SM background and three benchmark models for [$\PSGt_{\mathrm{L}}$]{}pair production with $m({\ensuremath{\PSGt_{\mathrm{L}}}\xspace})=100$, 125, and 200, $m(\PSGczDo)=1\GeV$. The numbers within parentheses in the legend correspond to the masses of the [$\PSGt_{\mathrm{L}}$]{}and in . The last bin includes overflow events in each case. The shaded uncertainty bands represent the combined statistical and systematic uncertainties in the background.](Figure_002-a.pdf "fig:"){width="48.00000%"} ![\[fig:srvariables\_tautau\] Distributions in [$\Sigma\mT$]{}() and () for events in the combined 2016 and 2017 data sets passing the baseline selection in the [$\tauh\tauh$]{}final state, along with the corresponding prediction for the SM background and three benchmark models for [$\PSGt_{\mathrm{L}}$]{}pair production with $m({\ensuremath{\PSGt_{\mathrm{L}}}\xspace})=100$, 125, and 200, $m(\PSGczDo)=1\GeV$. The numbers within parentheses in the legend correspond to the masses of the [$\PSGt_{\mathrm{L}}$]{}and in . The last bin includes overflow events in each case. The shaded uncertainty bands represent the combined statistical and systematic uncertainties in the background.](Figure_002-b.pdf "fig:"){width="48.00000%"} Event selection in the lep-tau final states {#sec:leptausel} ------------------------------------------- The baseline event selections for the [$\ell\tauh$]{}analyses require either an electron with $\pt > 26\,(35)$and $\abs{\eta} < 2.1$ or a muon with $\pt > 25\,(28)$and $\abs{\eta}<2.4$ for the 2016 (2017) data, and a candidate with $\pt > 30$and $\abs{\eta} < 2.3$. Electrons, muons, and candidates are required to have $\abs{{\ensuremath{d_{z}}\xspace}} < 0.2\cm$, and electrons and muons are also required to have $\abs{{\ensuremath{d_{xy}}\xspace}} < 0.045\cm$. Electrons and muons have to satisfy ${\ensuremath{I_{\text{rel}}}\xspace}<0.15$ and $<$0.1, respectively. Backgrounds from and [$\PW$+jets]{}are greatly reduced by vetoing events that contain jets with $\pt > 20\GeV$. Events from the [$\PW$+jets]{}background are further reduced by requiring the transverse mass $\mT(\ell, \ptvecmiss)$, calculated using the electron or muon momentum vector and , to be between 20 and 60or above 120. A significant background from DY+jets events is reduced by requiring the invariant mass of the electron or muon and the , $m_{\ell \tauh}$ to be above 50. To reduce background from QCD multijet events, we require $2.0< \Delta R(\ell, \tauh) <3.5$. =800 With these preselection criteria in place, we train several BDTs corresponding to different signal hypotheses to classify signal and background events. The input variables are the of the electron or muon, the of the candidate, , $\mT(\ell, \ptvecmiss)$, $\Delta \eta(\ell,\tauh)$, $\Delta \phi(\ell, \ptvecmiss)$, $\Delta \phi(\tauh, \ptvecmiss)$, $\Delta R(\ell, \tauh)$, $m(\ell \tauh)$, and $\mT^{\text{tot}}\equiv\sqrt{\smash[b]{\mT^2(\ell, \ptvecmiss)+\mT^2(\tauh, \ptvecmiss)}}$. We also include and the contransverse mass ($m_{\mathrm{CT}}$) [@Tovey:2008ui; @Polesello:2009rn], computed from the visible decay products and defined as $$m_{\mathrm{CT}} \equiv \sqrt{2 \pt^{\ell} \pt^{\tauh} [1 + \cos\Delta\phi(\ell,\tauh)]}.$$ For signal events, $m_{\mathrm{CT}}$ is expected to have an endpoint near ${(m(\PSGt)^2-m(\PSGczDo)^2)}/m(\PSGt)$. Finally, we include the variable ${\ensuremath{D_{\zeta}}\xspace}=\ptvecmiss \cdot \vec \zeta - 0.85 (\ptvec^{\ell}+\ptvec^{\tauh}) \cdot \vec \zeta$, with $\vec{\zeta}$ being the bisector of the directions of the transverse momenta of the electron or muon and the candidate [@CuencaAlmenar:2008zza; @Khachatryan:2014wca]. The value of 0.85 reflects an optimization to efficiently distinguish DY+jets events from other backgrounds and the signal. Figure \[fig:srvariables\_leptau\] shows the distributions of events passing the baseline selections in the [$\PGm\tauh$]{}final state in two of the BDT input variables that provide the highest discriminating power, and $\mT^{\text{tot}}$. The distributions observed in the [$\Pe\tauh$]{}final state are similar. Since the signal kinematics depend on mass, we train BDTs for signals with masses of 100, 150, and 200. In all cases we use a mass of 1. As the results of the training depend critically on the number of input events, we relax the MVA-based isolation criteria and reduce the threshold for the to 20for the training sample in order to increase the number of training and test events. The “very tight" isolation and a threshold of 30for the are applied in the final analysis. For a given signal hypothesis, we choose the BDT trained with the same mass for models with masses of 100, 150, and 200, or the one that provides optimal sensitivity for models with other mass values. For signal models with masses of 90 and 125, we use the BDT trained for $m(\PSGt)=100\GeV$, while for those with a mass of 175, we use the BDT trained for $m(\PSGt)=200\GeV$. While signal events are largely expected to have high BDT output values, we include the full BDT distribution in a binned fit for the statistical interpretation of the analysis as described in Section \[sec:results\]. The binning is chosen to optimize signal significance. ![\[fig:srvariables\_leptau\] Distributions in () and $\mT^{\text{tot}}$ () for events in the combined 2016 and 2017 data passing the baseline selections in the [$\PGm\tauh$]{}final state, along with the corresponding prediction for SM background and three benchmark models of [$\PSGt_{\mathrm{L}}$]{}pair production with $m({\ensuremath{\PSGt_{\mathrm{L}}}\xspace})=100$, 125, and 200and $m(\PSGczDo)=1\GeV$. The numbers within parentheses in the legend correspond to the masses of the [$\PSGt_{\mathrm{L}}$]{}and in . The last bin includes overflow events in each case. The shaded uncertainty bands represent the combined statistical and average systematic uncertainties in the background.](Figure_003-a.pdf "fig:"){width="48.00000%"} ![\[fig:srvariables\_leptau\] Distributions in () and $\mT^{\text{tot}}$ () for events in the combined 2016 and 2017 data passing the baseline selections in the [$\PGm\tauh$]{}final state, along with the corresponding prediction for SM background and three benchmark models of [$\PSGt_{\mathrm{L}}$]{}pair production with $m({\ensuremath{\PSGt_{\mathrm{L}}}\xspace})=100$, 125, and 200and $m(\PSGczDo)=1\GeV$. The numbers within parentheses in the legend correspond to the masses of the [$\PSGt_{\mathrm{L}}$]{}and in . The last bin includes overflow events in each case. The shaded uncertainty bands represent the combined statistical and average systematic uncertainties in the background.](Figure_003-b.pdf "fig:"){width="48.00000%"} Background estimation {#sec:bkgest} ===================== =800 Our most significant backgrounds are from DY+jets, [$\PW$+jets]{}, QCD multijet, , and diboson processes. They have relative contributions that vary with final state. For the [$\tauh\tauh$]{}final state, the dominant background arises from the misidentification of jets as candidates in QCD multijet and [$\PW$+jets]{}events, constituting $\approx$65% of background after the baseline selection. For the [$\ell\tauh$]{}final states after the baseline selection, the main backgrounds are from DY+jets ($\approx$50%), [$\PW$+jets]{}($\approx$30%), and QCD multijet ($\approx$10%) events. The DY+jets contribution, which is also a major background in the [$\tauh\tauh$]{}final state ($\approx$20%), usually consists of events with two prompt leptons. This background is determined with simulation samples after applying corrections to match the normalization and to be consistent with variable distributions in collider data. The [$\PW$+jets]{}and QCD multijet backgrounds usually contain one or more jets misidentified as and their contributions are determined via methods that rely on data. Finally, we have smaller contributions from other SM processes such as the production of Higgs bosons, dibosons, and top quark pairs with or without vector bosons. These are estimated via MC simulation with appropriate correction factors applied as described in Section \[sec:evtreco\]. For the [$\ell\tauh$]{}analyses, dedicated CRs that are each enriched in one of the major background processes are used to validate the modeling of the BDT distribution and to extract uncertainties that are used to account for any potential mismodeling of the distributions in simulation. These CRs are described in the following subsections below. Estimation of background from misidentified jets {#sec:fakeestimation} ------------------------------------------------ ### Misidentified jets in the tau-tau final state =800 After requiring two candidates with high , events with misidentified candidates are the dominant background in the [$\tauh\tauh$]{}final state. This background, which originates predominantly from QCD multijet and [$\PW$+jets]{}production, is predicted by extrapolating the event count in a data sample selected with a relaxed isolation requirement into the SR. The fraction of non-prompt or misidentified candidates selected with the very loose MVA-based isolation working point that also pass the tight DeepPF isolation requirement is measured in a QCD multijet-enriched sample of same-charge [$\tauh\tauh$]{}events. The same-charge [$\tauh\tauh$]{}events are collected with the same [$\tauh\tauh$]{}trigger as opposite-charge [$\tauh\tauh$]{}events to avoid additional trigger-related biases. We also require to be low ($<$40) to reduce potential contributions from signal events. We find that roughly 20% of the same-charge events with misidentified candidates selected with very loose isolation also pass the tight isolation requirement. However, the rate depends on the and decay mode (one- or three-prongs) of the candidate, as well as the jet flavor, , whether the misidentified jet originates from the hadronization of light-flavor quarks, heavy-flavor quarks, or gluons. The misidentification rate is therefore measured in bins of and decay mode to mitigate the dependence on these factors. The measurement is also binned in the number of primary vertices ([$N_{\mathrm{PV}}$]{}) to capture the effects of pileup. From studies performed with MC simulation samples, a systematic uncertainty of $\approx$30% is assigned to account for the dependence of the misidentification rate on jet flavor. Since the isolation efficiency for prompt candidates is only around 70–80%, processes containing genuine candidates can enter the sideband regions in events that are selected with the relaxed isolation requirement. To take this into account when calculating the final background estimate, we define three categories of events with at least two loosely isolated candidates: (i) events in which both candidates pass the tight DeepPF isolation requirement, (ii) events in which one passes and one fails the tight isolation requirement, and (iii) events in which both candidates fail the tight isolation requirement. We then equate the count of events in each of these three event categories to the sum of expected counts for the events with two prompt candidates, two jets misidentified as candidates, or one prompt candidate and one jet misidentified as a candidate, that contribute to each category. The contributions from backgrounds with one or two jets misidentified as candidates in the SRs are then determined analytically by solving a set of linear equations. ### Misidentified jets in the e-tau and mu-tau final states The misidentification of jets as candidates also gives rise to a major source of background in the [$\Pe\tauh$]{}and [$\PGm\tauh$]{}final states that arises mainly from [$\PW$+jets]{}events with leptonic boson decays. We estimate this background from a sideband region in data selected using the SR selection criteria, with the exception that the candidates are required to satisfy the loose isolation working point and not the very tight working point. A transfer factor for the extrapolation of event counts from this -isolation range into the tight isolation range of the SR is determined with a [$\PW$+jets]{}CR selected from events with one muon and at least one candidate that passes the loose isolation requirement. In events with more than one candidate, the candidate with the highest value of the MVA-based isolation discriminant is used. To increase the purity of [$\PW$+jets]{}events in this region, we reduce the contribution from and QCD multijet events by requiring $60 < \mT(\ell,\ptvecmiss) < 120\GeV$, $\ptmiss > 40\GeV$, no more than two jets, and an azimuthal separation of at least 2.5 radians between any jet and the boson reconstructed from the muon and ($\Delta \phi(\PW,\text{jet})>2.5$). We also reject events with additional electrons or muons satisfying looser identification criteria. The remaining sample has an expected purity of $\approx$85% for [$\PW$+jets]{}events. The transfer factor, $R$, is then determined from this control sample after subtracting the remaining non-[$\PW$+jets]{}background contributions estimated from simulation, as follows: $$\label{eq:tranferFactor} R = \frac{N^{\mathrm{CR}}_{\text{data}}({\mathrm{VT}})-N^{\mathrm{CR}}_{\text{MC no \PW}}({\mathrm{VT}})}{N^{\mathrm{CR}}_{\text{data}} (\mathrm{L}\overline{\mathrm{VT}})-N^{\mathrm{CR}}_{\text{MC no \PW}} (\mathrm{L}\overline{\mathrm{VT}})},$$ where $N^{\mathrm{CR}}_{\text{data}}$ corresponds to the number of events in the CR in data. The parenthetical argument $\mathrm{VT}$ denotes events in which the candidate satisfies the very tight isolation working point, while $\mathrm{L}\overline{\mathrm{VT}}$ denotes those that satisfy the loose, but not the very tight requirement. Transfer factors are determined separately in bins of and $\eta$ of candidates in order to achieve an accurate description of the background. =800 The contribution of the background originating from a jet misidentified as a candidate in the SR is then determined from the corresponding sideband in data: $$N(\text{jet} \to \tauh) = R \, (N^{\text{sideband}}_{\text{data}} - N^{\text{sideband}}_{\mathrm{MC}, \PGt}),$$ where $N^{\text{sideband}}_{\text{data}}$ is the number of events in the sideband in data, from which $N^{\text{sideband}}_{\mathrm{MC}, \PGt}$, the number with genuine leptons as estimated with MC simulation by generator-level matching, is subtracted. We validate the estimation of jets misidentified as in a CR requiring $60 < \mT(\ell,\ptvecmiss) < 120\GeV$ and $\Delta \phi(\PW,\text{jet})<2.5$ to ensure that the region is independent of the region described above that is used to estimate the background. Estimation of background from Drell–Yan+jets {#sec:dyestimation} -------------------------------------------- The DY+jets background comes primarily from [$\PZ\to\PGt^{+}\PGt^{-}$]{}decays. We estimate this contribution via simulation, after applying corrections based on CRs in data. Mismodeling of the boson mass or distribution in simulation can lead to significant differences between data and simulation in kinematic discriminant distributions, especially when considering the large values of these variables that are relevant for the [$\tauh\tauh$]{}SRs. We therefore use a high-purity [$\PZ\to\PGm^{+}\PGm^{-}$]{}CR to compare the dimuon mass and spectra between data and simulation and use the observed differences to correct the simulation in the SRs with weights parameterized by generator-level boson mass and . The correction factors range up to 30% for high-mass and high-values. Because these factors are intended to compensate for missing higher-order effects in the simulation, we assign the differences between the generator-level boson mass and distributions in LO and NLO simulated events as systematic uncertainties. The differences between data and simulation are taken into account through the use of scale factors, as described in Section \[sec:evtreco\]. The uncertainties in these corrections are propagated to the final background estimate. The corrected simulation is validated in the [$\tauh\tauh$]{}final state using a [$\PZ\to\PGt^{+}\PGt^{-}$]{}CR selected by inverting the and [$\Sigma\mT$]{}requirements used to define the SRs. In addition, requiring a of at least 50for the [$\tauh\tauh$]{}system reduces the QCD multijet background and improves the purity of this CR. This choice makes it possible to increase the statistical power of this region by removing the $\ptmiss>50\GeV$ requirement. The visible mass distribution of the [$\tauh\tauh$]{}system shown in Fig. \[fig:ztautau\] () demonstrates that the corrected simulation agrees with the data within experimental uncertainties. For the analysis in the [$\ell\tauh$]{}final states, a normalization scale factor, as well as corrections to the distribution of the boson in simulation are obtained from a very pure [$\PZ\to\PGm^{+}\PGm^{-}$]{}CR in data. These events are selected by requiring two isolated muons and no additional leptons, at most one jet, no -tagged jets, and a dimuon mass in a window of 75–105, to increase the probability to $>$99% that they originate from [$\PZ\to\PGm^{+}\PGm^{-}$]{}decays. After subtracting all other contributions estimated from simulation, a normalization scale factor of $0.96\pm0.05$, which is compatible with unity, is extracted from the ratio of data to simulated events. The uncertainty in the scale factor is determined by varying systematic uncertainties associated with objects such as the muon efficiency and jet energy uncertainties. =800 To validate the DY+jets background prediction in the [$\ell\tauh$]{}analyses, we construct a CR in [$\PGm\tauh$]{}events with $ \mT(\PGm,\ptvecmiss) < 20\GeV$, $ 50 < m({\ensuremath{\PGm\tauh}\xspace}) < 80\GeV$, and ${\ensuremath{N_{\text{j}}}\xspace}= 0$. These requirements are chosen to obtain a [$\PZ\to\PGt^{+}\PGt^{-}$]{}sample with good purity. The $m({\ensuremath{\PGm\tauh}\xspace})$ range is chosen to select the boson peak, low $ \mT(\PGm,\ptvecmiss)$ helps to remove [$\PW$+jets]{}and potential signal contamination while the 0-jet requirement helps remove other backgrounds. The distribution of these events is shown in Fig. \[fig:ztautau\] (). We observe good agreement between data and the predicted background. ![\[fig:ztautau\] Visible-mass spectra of lepton pairs in [$\tauh\tauh$]{}events () and distribution in [$\PGm\tauh$]{}events () in data and the corresponding prediction for SM background in the combined 2016 and 2017 DY+jets validation regions. The last bin includes overflow events in each case. The shaded uncertainty band represents the statistical and systematic uncertainties in the background prediction. For the [$\PGm\tauh$]{}distribution, the systematic uncertainty included in each bin corresponds to a single common average value.](Figure_004-a.pdf "fig:"){width="48.00000%"} ![\[fig:ztautau\] Visible-mass spectra of lepton pairs in [$\tauh\tauh$]{}events () and distribution in [$\PGm\tauh$]{}events () in data and the corresponding prediction for SM background in the combined 2016 and 2017 DY+jets validation regions. The last bin includes overflow events in each case. The shaded uncertainty band represents the statistical and systematic uncertainties in the background prediction. For the [$\PGm\tauh$]{}distribution, the systematic uncertainty included in each bin corresponds to a single common average value.](Figure_004-b.pdf "fig:"){width="48.00000%"} Estimation of other backgrounds ------------------------------- Smaller contributions are expected from other SM backgrounds, including diboson, triboson, and Higgs boson production. There are also contributions from and single top quark production, or top quark pair production in association with a vector boson. These are estimated via MC simulation after application of efficiency and energy-scale corrections. Experimental and theoretical uncertainties are evaluated as described below in Section \[sec:sysunc\]. For the [$\ell\tauh$]{}analyses, we check the BDT distribution in a -enriched CR that is defined by requiring the event selection to be the same as in the SR, except for a requirement of one or two -tagged jets. To validate the background prediction, we construct a CR of events with oppositely charged muon-electron pairs that have $m_{\PGm \Pe} > 90\GeV$ and ${\ensuremath{N_{\text{j}}}\xspace}= 0$. We obtain systematic uncertainties for the normalization of the corresponding backgrounds and any potential mismodeling of the BDT distribution in these CRs. The latter is done by constructing a $\chi^2$ test for all CRs with the BDT modeling taken into account by including an additional floating uncertainty that is determined by requiring a $p$-value [@pvalue] of at least 68% in all CRs. In this way, the BDT shape uncertainty is estimated to be 9%. Systematic uncertainties {#sec:sysunc} ======================== The dominant uncertainties in this analysis are the statistical uncertainties resulting from limited event counts in data sidebands or in simulated event samples used to obtain background estimates and the systematic uncertainties in the estimated rates for jets to be misidentified as candidates. We rely on an extrapolation in isolation to obtain an estimate of the background originating from jets misidentified as candidates. In the [$\tauh\tauh$]{}analysis, the uncertainty in this extrapolation is dominated by the dependence of isolation on jet flavor. It also includes the statistical uncertainty associated with the CR samples from which the extrapolation factors are obtained, which can be significant in the case of search regions with limited event counts that are defined with stringent kinematic requirements. The uncertainty in the combined identification and isolation efficiency for prompt candidates is also propagated to the final estimated uncertainty. In the [$\ell\tauh$]{}analyses, we estimate a transfer factor for the extrapolation in isolation from a [$\PW$+jets]{}-enriched CR. The purity of [$\PW$+jets]{}events this region is $\approx$85% as determined from simulation. We therefore propagate a relative uncertainty of 15% to account for contamination from other sources. =800 We use simulation to obtain estimates of the yields from other background contributions and to estimate the potential signal contributions. We propagate uncertainties related to the tagging, trigger, and selection efficiencies, the renormalization and factorization scales, PDFs, jet energy scale and resolution, unclustered energy contributing to , and the energy scales of electrons, muons, and candidates. The correction factors and the corresponding uncertainties for the energy scale in simulation are derived from [$\PZ\to\PGt^{+}\PGt^{-}$]{}events in the [$\ell\tauh$]{}final states by fits to distributions of the reconstructed mass and the visible mass of the [$\ell\tauh$]{}system [@Sirunyan:2018pgf]. The systematic uncertainties corresponding to energy scale variations can be significant in the [$\tauh\tauh$]{}search regions defined with stringent kinematic requirements, which are affected by large statistical uncertainties, because of potentially large event migrations. For the DY+jets background, we have an additional uncertainty associated with the corrections applied to the mass and distributions. We assign a 15% normalization uncertainty in the [$\tauh\tauh$]{}final state for the cross sections of processes estimated from simulation, namely DY+jets, , diboson, and rare SM processes, based on the results of CMS differential cross section measurements [@Sirunyan:2018owv; @Sirunyan:2018ucr]. For the [$\ell\tauh$]{}analyses, we extract normalization uncertainties of 5, 5, and 20% for the DY+jets, , and backgrounds, respectively, based on the estimated impurity of the corresponding process-enriched CRs. An additional uncertainty of 9% is assigned to cover potential mismodeling of the BDT distribution in simulation that is based on studies in CRs. The categorization of events in the [$\tauh\tauh$]{}final state by the number of reconstructed jets induces sensitivity to the modeling of ISR in the signal simulation. The $\pt^{\mathrm{ISR}}$ distribution of simulated signal events is reweighted to improve the ISR modeling. The reweighting factors are obtained from studies of boson events. We take the deviation of the reweighting factors from unity as a systematic uncertainty. The uncertainty in the integrated luminosity is taken into account in all background estimates for which we do not extract normalization scale factors in dedicated data CRs, as well as for signal estimates. This uncertainty corresponds to 2.5% [@CMS-PAS-LUM-17-001] and 2.3% [@CMS-PAS-LUM-17-004] for the 2016 and 2017 data, respectively. With the exception of statistical uncertainties, most other uncertainties are of similar size between the 2016 and 2017 analyses. The main systematic uncertainties for signal and background are summarized in Table \[tab:systematics\]. In general, we treat all statistical uncertainties as uncorrelated. In addition, all systematic uncertainties arising from statistical limitations in the 2016 and 2017 data are assumed to be uncorrelated while systematic uncertainties from similar sources are treated as correlated or partially correlated across the various background and signal predictions. For the combination of the [$\tauh\tauh$]{}and [$\ell\tauh$]{}analyses, we correlate uncertainties related to object reconstruction, with the exception of the selection efficiency, which is treated as uncorrelated because of the use of different isolation algorithms. Results and interpretation {#sec:results} ========================== The results of the search in the [$\tauh\tauh$]{}final state are presented in Fig. \[fig:tautauresults\] and summarized in Tables \[tab:results2016\] and \[tab:results2017\]. The background predictions resulting from a maximum likelihood fit to the data under the background-only hypothesis are shown in the lower row of Fig. \[fig:tautauresults\]. The BDT distributions corresponding to a training for a mass of 100and a mass of 1are shown before and after the maximum-likelihood fit to the data in Figs. \[fig:fitplots\_mutau\] and \[fig:fitplots\_etau\] for the [$\PGm\tauh$]{}and [$\Pe\tauh$]{}final states, respectively. The data are consistent with the prediction for SM background. The predicted and observed event yields in the last, most sensitive BDT bins are summarized in Tables \[tab:resultsSemilep2016\] and \[tab:resultsSemilep2017\] for [$\ell\tauh$]{}final states. For the statistical interpretation of these results, the normalization uncertainties affecting background and signal predictions are generally assumed to be log-normally distributed. For statistical uncertainties limited by small event counts in data or simulation, we use a $\Gamma$ distribution. The results are used to set upper limits on the cross section for the production of pairs in the context of simplified models [@Simp; @Alwall:2008ag; @Alwall:2008va; @Alves:2011wf] using all of the exclusive [$\tauh\tauh$]{}SRs and the [$\ell\tauh$]{}BDT distributions in a full statistical combination. The limits are evaluated using likelihood fits with the signal strength, background event yields, and nuisance parameters corresponding to the uncertainties in the signal and background estimates as fitted parameters. The nuisance parameters are constrained within their uncertainties in the fit. We assume that the decays with 100% branching fraction to a lepton and a . The 95% upper limits on SUSY production cross sections are calculated using a modified frequentist approach with the criterion [@Junk:1999kv; @Read:2002hq]. An asymptotic approximation is used for the test statistic [@CMS-NOTE-2011-005; @Cowan:2010js], $q_{\mu} = -2\ln \mathcal{L}_{\mu}/\mathcal{L}_{\text{max}}$, where $\mathcal{L}_{\text{max}}$ is the maximum likelihood determined by allowing all fitted parameters, including the signal strength $\mu$, to vary, and $\mathcal{L}_{\mu}$ is the maximum likelihood for a fixed signal strength. Figure \[fig:limitsleft\] shows the limits obtained for purely left-handed pair production, while Fig. \[fig:limitsdegenerate\] shows the limits obtained for the degenerate model in which both left- and right-handed pairs are produced. The [$\tauh\tauh$]{}analysis makes the dominant contribution to the search sensitivity. A slight excess of events over the background expectation in the [$\tauh\tauh$]{}SRs results in an observed limit that is weaker than the expected limit. The strongest limits are observed in the case of a nearly massless . In general, the constraints are weaker for higher values of the mass because of smaller experimental acceptances. For masses above $\approx$150, however, the sensitivity does not degrade significantly when the mass increases up to 20. In the purely left-handed model, the strongest limits are observed for a mass of 125where we exclude a pair production cross section of 132. This value is a factor of 1.14 larger than the theoretical cross section. In the degenerate model we exclude masses between 90 and 150under the assumption of a nearly massless . ![image](Figure_005-a.pdf){width="48.00000%"} ![image](Figure_005-b.pdf){width="48.00000%"}\ ![image](Figure_005-c.pdf){width="48.00000%"} ![image](Figure_005-d.pdf){width="48.00000%"} \[tab:results2016\] \[tab:results2017\] \[tab:resultsSemilep2017\] ![image](Figure_006-a.pdf){width="48.00000%"} ![image](Figure_006-b.pdf){width="48.00000%"}\ ![image](Figure_006-c.pdf){width="48.00000%"} ![image](Figure_006-d.pdf){width="48.00000%"}\ ![image](Figure_007-a.pdf){width="48.00000%"} ![image](Figure_007-b.pdf){width="48.00000%"}\ ![image](Figure_007-c.pdf){width="48.00000%"} ![image](Figure_007-d.pdf){width="48.00000%"}\ ![image](Figure_008-a.pdf){width="48.00000%"} ![image](Figure_008-b.pdf){width="48.00000%"} ![image](Figure_008-c.pdf){width="48.00000%"} ![image](Figure_009-a.pdf){width="48.00000%"} ![image](Figure_009-b.pdf){width="48.00000%"} ![image](Figure_009-c.pdf){width="48.00000%"} Summary {#sec:summary} ======= A search for direct slepton () pair production has been performed in proton-proton collisions at a center-of-mass energy of 13in events with a lepton pair and significant missing transverse momentum. Search regions are defined using kinematic observables that exploit expected differences in discriminants between signal and background. The data used for this search correspond to an integrated luminosity of 77.2collected in 2016 and 2017 with the CMS detector. No excess above the expected standard model background has been observed. Upper limits have been set on the cross section for direct pair production for simplified models in which each decays to a lepton and the lightest neutralino, with the latter being assumed to be the lightest supersymmetric particle. For purely left-handed pair production, the analysis is most sensitive to a mass of 125when the neutralino is nearly massless. The observed limit is a factor of 1.14 larger than the expected production cross section in this model. The limits observed for left-handed pair production are the strongest obtained thus far for low values of the mass. In a more optimistic, degenerate production model, in which both left- and right-handed pairs are produced, we exclude masses up to 150, again under the assumption of a nearly massless neutralino. These results represent the first exclusion reported for this model for low values of the mass between 90 and 120. We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centers and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: BMBWF and FWF (Austria); FNRS and FWO (Belgium); CNPq, CAPES, FAPERJ, FAPERGS, and FAPESP (Brazil); MES (Bulgaria); CERN; CAS, MoST, and NSFC (China); COLCIENCIAS (Colombia); MSES and CSF (Croatia); RPF (Cyprus); SENESCYT (Ecuador); MoER, ERC IUT, PUT and ERDF (Estonia); Academy of Finland, MEC, and HIP (Finland); CEA and CNRS/IN2P3 (France); BMBF, DFG, and HGF (Germany); GSRT (Greece); NKFIA (Hungary); DAE and DST (India); IPM (Iran); SFI (Ireland); INFN (Italy); MSIP and NRF (Republic of Korea); MES (Latvia); LAS (Lithuania); MOE and UM (Malaysia); BUAP, CINVESTAV, CONACYT, LNS, SEP, and UASLP-FAI (Mexico); MOS (Montenegro); MBIE (New Zealand); PAEC (Pakistan); MSHE and NSC (Poland); FCT (Portugal); JINR (Dubna); MON, RosAtom, RAS, RFBR, and NRC KI (Russia); MESTD (Serbia); SEIDI, CPAN, PCTI, and FEDER (Spain); MOSTR (Sri Lanka); Swiss Funding Agencies (Switzerland); MST (Taipei); ThEPCenter, IPST, STAR, and NSTDA (Thailand); TUBITAK and TAEK (Turkey); NASU and SFFR (Ukraine); STFC (United Kingdom); DOE and NSF (USA). Individuals have received support from the Marie-Curie program and the European Research Council and Horizon 2020 Grant, contract Nos. 675440, 752730, and 765710 (European Union); the Leventis Foundation; the A.P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation à la Recherche dans l’Industrie et dans l’Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the F.R.S.-FNRS and FWO (Belgium) under the “Excellence of Science – EOS" – be.h project n. 30820817; the Beijing Municipal Science & Technology Commission, No. Z181100004218003; the Ministry of Education, Youth and Sports (MEYS) of the Czech Republic; the Lendület (“Momentum") Program and the János Bolyai Research Scholarship of the Hungarian Academy of Sciences, the New National Excellence Program ÚNKP, the NKFIA research grants 123842, 123959, 124845, 124850, 125105, 128713, 128786, and 129058 (Hungary); the Council of Science and Industrial Research, India; the HOMING PLUS program of the Foundation for Polish Science, cofinanced from European Union, Regional Development Fund, the Mobility Plus program of the Ministry of Science and Higher Education, the National Science Center (Poland), contracts Harmonia 2014/14/M/ST2/00428, Opus 2014/13/B/ST2/02543, 2014/15/B/ST2/03998, and 2015/19/B/ST2/02861, Sonata-bis 2012/07/E/ST2/01406; the National Priorities Research Program by Qatar National Research Fund; the Ministry of Science and Education, grant no. 3.2989.2017 (Russia); the Programa Estatal de Fomento de la Investigaci[ó]{}n Cient[í]{}fica y T[é]{}cnica de Excelencia María de Maeztu, grant MDM-2015-0509 and the Programa Severo Ochoa del Principado de Asturias; the Thalis and Aristeia programs cofinanced by EU-ESF and the Greek NSRF; the Rachadapisek Sompot Fund for Postdoctoral Fellowship, Chulalongkorn University and the Chulalongkorn Academic into Its 2nd Century Project Advancement Project (Thailand); the Welch Foundation, contract C-1845; and the Weston Havens Foundation (USA). The CMS Collaboration \[app:collab\] ==================================== =5000=500=5000	{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by a recent experiment at MIT, we consider the collision of two clouds of spin-polarized atomic Fermi gases close to a Feshbach resonance. We explain why two dilute gas clouds, with underlying attractive interactions between its constituents, bounce off each other in the strongly interacting regime. Our hydrodynamic analysis, in excellent agreement with experiment, gives strong evidence for a novel metastable many-body state with effective repulsive interactions.' author: - 'Edward Taylor, Shizhong Zhang, William Schneider and Mohit Randeria' date: 'January 18, 2012' title: 'Colliding clouds of strongly interacting spin-polarized fermions' --- 0[\|[z]{}\_]{} 0[\|\_]{} Introduction ============ Ultracold atomic gases [@Bloch08; @Giorgini08] open up new frontiers in the study of non-equilibrium dynamics of strongly interacting quantum systems. In contrast to electronic systems, the Fermi energy in quantum gases is of the order of a few kilohertz, so that questions about metastability and equilibration in quantum systems can now be explored in real time in the laboratory. Exploring these new regimes and developing new theoretical tools to gain insight into the non-equilibrium dynamics of many-body systems is an exciting challenge at the intersection of condensed matter and atomic-molecular-optical physics. The recent experiment of Ref. [@Sommer11] offers a beautiful example of non-equilibrium dynamics that is completely unexpected and, at first sight, counterintuitive. Two clouds of ultracold fermions prepared in different hyperfine-Zeeman (“spin") states $\sigma = {\uparrow}$ and ${\downarrow}$ are initially separated using a Stern-Gerlach field; see Fig. \[schematicfig\]. Once the field is turned off, the clouds are impelled together by the confining harmonic potential and they collide. In the strongly interacting unitary regime these very dilute clouds are observed to bounce off each other, almost as if they were colliding billiard balls! This is truly remarkable in view of the fact that the underlying atomic interaction is attractive, and the equilibrium ground state is known to be a paired superfluid. Why then do the clouds behave as though the interactions are repulsive? Two additional aspects of Ref. [@Sommer11] are also noteworthy. Once the short-time bounce is damped out, the centers of mass of the clouds remain separated for a considerable duration at intermediate times ($\sim 100$ ms), before eventually merging. In fact, the analysis in Ref. [@Sommer11] focused on precisely the slow merging of the clouds, or spin diffusion, at long times ($\sim 0.5$ s). Finally, the observed dynamics changes qualitatively as one moves out of the strongly interacting regime, with the clouds merging rapidly for weak interactions. In this paper we gain insight into three aspects of this remarkable dynamics using a hydrodynamic approach in the strongly interacting regime. (1) We explain why the clouds repeatedly bounce off each other at short times. (2) We also explain why the centers of mass of the two clouds remain separated at intermediate time scales. We show that these results are natural consequences of a metastable many-body state on the “upper branch" of the Feshbach resonance where the effective interactions are repulsive [@Pilati10; @Chang11]. (3) Finally, we provide insight into how the dynamics changes as a function of the strength of the interaction. Hydrodynamics ============= Our analysis is based on two hypotheses: (i) The collisions between atoms are sufficiently rapid to establish local thermodynamic equilibrium when the clouds overlap. (ii) The loss from the upper branch scattering state is slow in the experiment of Ref. [@Sommer11]. At unitarity, the two-body collision rate is very large, $1/\tau_2 \sim \epsilon_F$ (with $\hbar = k_B = 1$), for a range of temperatures $0.1 \lesssim T/\epsilon_F \lesssim 0.3$ [@Kinast05; @Massignan05]. With a Fermi energy $\e_F \sim 10^4$Hz and an axial trap frequency $\omega_z \sim 10$Hz [@Sommer11], the gas in the overlap region of the clouds will reach local thermodynamic equilibrium within $\sim 10^{-3}$ trap periods. Thus the clouds behave hydrodynamically in the overlap region [@nonoverlap] and their dynamics reflect the (metastable) equation of state of the gas. (Hydrodynamics also describes well the very different behavior observed in colliding clouds of [spin-balanced]{} superfluid gases at unitarity [@Joseph11].) As the two clouds come together in the overlap region, the atoms are initially in scattering states, and the formation of two-body spin-singlet bound states requires three-body collisions in order to satisfy kinematic constraints. The time scale $\tau_3$ for such processes is not well understood for strongly interacting gases, however, there are indications that it is enhanced close to unitarity [@Zhang11]. In addition, the $\tau_3$ relevant to Ref. [@Sommer11] is further enhanced relative to the microscopic decay time since three-body collisions are limited to a small overlap region for a fraction of the oscillation period $\omega_z^{-1}$. Now, for the experimentally relevant time interval $\tau_2 \ll t \ll \tau_3$, the rapid two-body collisions have already established thermal equilibrium, while the slow three-body recombination process has yet to drive the system into the lower branch. This allows us to model the cloud dynamics using Euler’s equations + = - ( + V\_) - \_,\[Euler\] + (n\_\_) = 0,\[continuity\] with an appropriate energy density ${\cal{E}}[n_\sigma({\bf r})]$ for the scattering states (see below). Here $\bv_{\sigma}$ and $n_{\sigma}$ are the velocity and density of the $\sigma$ fermions, $V_{\mathrm{trap}} = m(\omega^2_{\perp}\rho^2 + \omega^2_z z^2)/2$ is the trap potential. We include a phenomenological $\gamma$ to account for strong spin-current damping in the hydrodynamic regime [@Vichi99; @Bruun08]. Viscosity, describing the damping of in-phase current ($\bv_{\uparrow}=\bv_{\downarrow}$), is ignored since the motion of the colliding clouds is primarily out-of-phase [@Massignan05; @Cao11]. As a result of tight confinement in the radial direction, the two modes with the lowest energies are the spin-dipole mode and axial breathing mode. Hence, we concentrate on the centers-of-mass of the clouds and their widths along the $\hat{z}$-direction and reformulate (\[Euler\]) and (\[continuity\]) in terms of those variables. Using the notation $\langle \cdots\rangle_{\sigma}\equiv 1/N_{\sigma}\int d^3{\bf r} n_{\sigma}(\cdots)$, we define the center-of-mass positions $\z0 \equiv \langle z\rangle_{\sigma}$ and widths $\dz \equiv \sqrt{8\langle (z-\z0)^2\rangle_{\sigma}}$. The $\sqrt{8}$ ensures that $\dz$ coincides with the axial Thomas-Fermi (TF) radius $R_z$ in equilibrium. (\[Euler\]) and (\[continuity\]) then lead to the exact equations of motion \_ + \^2\_z 0 &=& -\_z(/n\_)\_ - v\_[z]{}\_\ &&+\_v\_[z]{} - \_z v\^2\_/2\_\[zm\]and &&\ &&-(z-0) \_z(/n\_)\_ - (z-0)v\_[z]{}\_\ &&+(z-0)\[\_v\_[z]{} - \_z v\^2\_/2\]\_. \[dzm\] The problem can be further simplified while still retaining the essential physics of the (nonlinear) coupling between the spin-dipole and axial breathing modes by using a TF ansatz: n\_(,t) = \^[3/2]{},\[ansatz\]where $R_{\alpha}\equiv \sqrt{2\epsilon_F^0/m\omega^2_{\alpha}}$ is the TF radius along the $\alpha$-axis and $\epsilon_F^0 = (\omega^2_{\perp}\omega_z)^{1/3}(3N)^{1/3}$ is the chemical potential of an ideal two-component Fermi gas ($N_{\uparrow}=N_{\downarrow}=N/2$). This ansatz allows for a time-dependent center-of-mass $\z0$ as well as axial compression of the clouds. The continuity equation (\[continuity\]) leads to the velocity field ${\bf v}_\sigma=v_\sigma\hat{z}$ with $v_{\sigma}(z,t) = \dot{\bar{z}}_{\sigma}- \z0\dot{\dz}/\dz + z\dot{\dz}/\dz$. Axial and radial symmetry lets us set $\bar{z}_{\uparrow}\equiv \bar{z} = -\bar{z}_{\downarrow}$ and $\delta z_{\uparrow} = \delta z_{\downarrow}\equiv \delta z$. Using (\[ansatz\]) in (\[zm\]) and (\[dzm\]) gives the coupled nonlinear integro-differential equations: (+\^2\_z)\|[z]{}(t) = -d\^3r - v\_\_\[barz\],and (+\^2\_z)z(t) &=& -d\^3r\ && - (z-\|[z]{}(t))v\_\_.\[deltaz\] In the unitarity regime, the spin-current damping [@Bruun08] assumes a particularly simple form when the relative velocity is of order of the Fermi velocity, as is the case at early times for two clouds initially separated by $R_z$. In the overlap region, we take = ,where $\widetilde{\gamma}$ is of order unity at unitarity and $\epsilon_{F\sigma}({\bf r})=(6\pi^2 n_{\sigma}(\br,t))^{2/3}/2m$ is the local Fermi energy. This simple choice of damping is, if anything, an overestimate, and we have checked that our results are robust against reasonable modification of the damping parameter. Energy Functional ================= We now need to specify the energy functional ${\cal E}[n_\sigma]$ relevant for dynamics at times $\tau_2 \ll t \ll \tau_3$. The ground state (“lower branch” of the Feshbach resonance) for scattering length $a_s > 0$ necessarily involves bound pairs. But for $t \ll \tau_3$, the three-body processes required to relax to this state have not yet occurred. The system has, however, developed two-body correlations characteristic of the metastable “upper branch" state, studied theoretically in Refs. [@LeBlanc09; @Pilati10; @Chang11], motivated by an earlier experiment [@Jo09]. The approximate many-body wavefunction in the upper branch is =\_S({ \_[i]{}})\_S({\_[j]{}}), \[jastrow\] where $\Phi_S(\{ \br_{i\sigma}\})$’s are Slater determinants and the Jastrow factor $f(r)$ describes the short-range correlations between fermions. The effective repulsion between $\uparrow$ and $\downarrow$ atoms in the upper branch is crucially related to the node in $f(r)$, in contrast to the nodeless Jastrow factor for the lower branch. For small $a_s>0$, the node occurs at $a_s$, with $f(r)\sim (1-a_s/r)$ similar to the two-body problem. For large $a_s$, the node saturates [@Chang11] to $\sim 1/k_F$, the only length scale at unitarity. Quantum Monte Carlo (QMC) studies [@Pilati10; @Chang11] of the upper-branch wavefunction (\[jastrow\]) reveal that the system undergoes (ferromagnetic) phase separation for sufficiently strong interactions $k_Fa_s\gtrsim 1$. We need ${\cal E}[n_\sigma]$ for arbitrary polarization, which has not been studied by QMC. We thus use the lowest order constraint variational (LOCV) approximation [@Pand7173], which has been used for the upper branch [@Heiselberg10] and is in close agreement with QMC data [@Pilati10; @Chang11]. As shown in the Appendix, the upper branch LOCV energy density is =\_[F]{}n\_+\_[F]{}n\_+(n\_\_+n\_\_).\[ELOCV\] The interaction energies $\lambda_\uparrow$ and $\lambda_\downarrow$ are functions of $k_Fa_s$ and $x = n_\downarrow/n_\uparrow$. Hydrodynamics requires $T\gtrsim 0.1\epsilon_F$ [@Massignan05]; for simplicity, we use the zero temperature LOCV energy functional to study the dynamics, expecting it to qualitatively describe the physics at low temperatures. It is important to note here that while we rely on the approximate LOCV calculation to obtain the equation of state, the underlying physics of the effective repulsive interaction is independent of the particular scheme used and a more precise calculation would only leads to quantitative improvement without changing the crucial physical picture of the bounce. Dynamics at unitarity ===================== We solve (\[barz\]) and (\[deltaz\]), using (\[ELOCV\]) and (\[ansatz\]); for details see Appendix \[Appendixnumerical\]. The results at unitarity are shown in Fig. \[upperbouncefig\]. The two clouds bounce off each other for several oscillations due to the repulsive nature of the upper branch functional. The period of the initial bounce is roughly $0.56(2\pi/\omega_z)$, slightly less than the experimental value $0.61(2\pi/\omega_z)$ [@Sommer11]. We attribute this difference to the simple ansatz (\[ansatz\]) used to model the dynamics. The bounce persists for several cycles in spite of the very large damping ($\epsilon_F^0/\omega_z\gg 1$) because the overlap between the two clouds is small. Once the oscillation is damped out, the centers of mass of the two clouds remain segregated, with a final $(\bar{z}_{\uparrow} - \bar{z}_{\downarrow}) \simeq 0.4$ times the initial separation. This intermediate time behavior reflects the tendency for system to phase segregate in the upper branch at unitarity. Collisional dynamics away from unitarity ======================================== Away from unitarity, it is harder to reach the hydrodynamic regime. If $\omega_z$ is sufficiently small, however, it will always be the case that the dynamics in this direction are hydrodynamic. The greater challenge at finite $a_s > 0$ is that $\tau_3$ may not be much greater than $\tau_2$. As $a_s$ decreases from unitarity, one expects that $\tau_3$ reaches a minimum for $k_F a_s\sim 1$ and then becomes large again [@Petrov03], with $\tau_3\sim (na^3_s)^{-2}\epsilon^{-1}_F$ for $k_Fa_s \ll 1$. In contrast, the cross-section $a^2_s$ determines $\tau_2\sim (na^3_s)^{-2/3}\epsilon^{-1}_F$. Hence, for small $k_Fa_s$, again $\tau_3\gg \tau_2$. However, when $k_Fa_s\sim 1$, it is conceivable that $\tau_3\sim \tau_2$. With these caveats in mind, we solve (see Appendix \[Appendixnumerical\] for details) our hydrodynamic equations away from unitarity to understand the relationship between the intermediate-time dynamics, after the bounce is damped out, and the upper branch equation of state: Fermi liquid ($k_Fa_s\lesssim 1$) versus phase separated ($k_Fa_s\gtrsim 1$). The solutions of (\[barz\]) and (\[deltaz\]) at $k_F(0)a_s=2$ and $0.5$ (using the same initial conditions and $\epsilon_F^0$ as in Fig. \[upperbouncefig\]), are shown in Figs. \[bounce2fig\] and \[bounce3fig\]. Here $k_F(0) = \sqrt{2m\epsilon_F^0}$ is the ideal gas Fermi wavevector at the cloud center. Appropriate for the smaller interaction strength, we use an expression for the spin-current damping obtained from kinetic theory [@Vichi99]. When symmetrized in the spin components, it takes the simple form [@comment1] =(4/9)(a\_s)\^2 . Even at small $k_F(0)a_s$, the clouds exhibit a weak bounce due to compressional recoil, which damps out very quickly. We see, however, a clear difference between Fig. \[bounce2fig\] ($k_Fa_s\gtrsim 1$), where the centers of mass remain separated, and Fig. \[bounce3fig\] ($k_Fa_s\lesssim 1$), where they merge. This behavior is qualitatively similar to the experimental results shown in Supplementary Fig. 1 of Ref. [@Sommer11]. Discussion ========== We now compare our results in detail with experiments. As seen from Fig. \[upperbouncefig\], our results at unitarity – the short-time bounce, the damping of the oscillations and the separation of the clouds at intermediate times – are all in very good agreement with Ref. [@Sommer11]. Away from unitarity, at finite values of $k_F a_s > 0$, our results are qualitatively similar to the data in Supplementary Fig. 1 of Ref. [@Sommer11]. For times $ \omega_z^{-1}\ll t \ll \tau_3$, say, $t=200$ ms, the system remains phase segregated for large $k_Fa_s$ \[Figs. 1(f,g)\], while completely mixed for small $k_Fa_s$ \[Figs. 1(c,d,e)\]. In fact, a very rough estimate for the critical $k_F a_s$ can be read off from the experimental data: it is between $0.26$ \[Fig. 1(e)\] and $1.2$ \[Fig. 1(f)\]. The long time behavior, not described here, will be dominated by spin diffusion [@Sommer11] at non-zero temperatures and three-body processes. Finally, we comment on the experiment of Jo [et al.]{} [@Jo09] in which a spin-balanced mixture, initially at a small $a_s>0$, is swept close to resonance, generating strong interactions in the upper branch. It appears, however, that rapid losses render it unstable to the lower branch within a very short time ($\tau_{3,\mathrm{micro}}\sim 1$ms) [@Zhang11; @Pekker11; @Sanner11] and phase separated ferromagnetic domains are not observed. In contrast, the specific initial configuration in the colliding cloud experiment [@Sommer11] proves to be crucial for the metastability of the upper branch. Three-body loss is limited spatially within the overlap region between different spins at the trap center, and temporally to a fraction of the oscillation period. As such, the effective $\tau_3$ is greatly enhanced ($\tau_3\gg \tau_{3,\mathrm{micro}}$) and one can study the metastable upper branch. Note that while $\tau_{3,\mathrm{micro}}$ is small, $\sim 10/\epsilon_F$ [@Sanner11], it is still an order of magnitude greater than $\tau_2\sim \epsilon_F$, and thus local thermodynamic equilibrium can be established in the upper branch. In summary, we have investigated the short-time bounce dynamics of the recent MIT experiment [@Sommer11]. The bounce frequency we found is in good agreement with the experiments. Furthermore, the experimental observation that the two clouds sit side-by-side for as long as $\sim 100$ms validates our assumption of a long $\tau_3$. We argue that the intermediate time behavior of the two colliding clouds indicates that the system favors a phase separated metastable state for larger value of $k_Fa_s$. We cannot see how a lower-branch energy functional, with attractive interactions forming bound pairs, could lead to the observed dynamics. [Note added:]{} Recently, a Boltzmann approach was used to investigate the same problem in the high-$T$ regime [@Goulko11]. The system is found to be strongly hydrodynamic for $k_Fa_s \gtrsim 1$. We acknowledge discussions with S. Stringari, S-Y. Chang, T.L. Ho, N. Trivedi and M. Zwierlein. We gratefully acknowledge support from ARO W911NF-08-1-0338 (ET), NSF-DMR 0907366 (SZ, ET), NSF-DMR 1006532 (WS, MR). Upper branch and LOCV {#Appendix} ===================== In this Appendix, we present a detailed discussion of the so-called “upper branch” of a Feshbach scattering resonance in two-component Fermi gases [@Pilati10; @Chang11] as well as the lowest order constraint variational (LOCV) method [@Pand7173], which we use to calculate the upper branch equation of the state [@Heiselberg10]. Upper Branch ------------ The concept of the upper branch of a Feshbach resonance comes from the two-body problem (either in a finite box or in a harmonic potential), where it is completely well-defined for all values of $a_s$, positive or negative. The two-body wavefunction in the upper branch is a scattering state with a single node that makes it orthogonal to the ground state (lower branch) wave-function. This notion has been generalized to the many-body case [@Pilati10; @Chang11] by writing a Jastrow-Slater wavefunction, where the Jastrow correlation factor has a node; see Eq. (9) of the main paper and the discussion immediately following this equation. To summarize our definition of the upper branch in the many-body problem, any sensible definition must [at least]{} satisfy the following conditions [@Chang11]:\ (1) The many-body wavefunction includes, apart from the nodes introduced by the Pauli principle, one additional node for any pair of fermions with opposite spin.\ (2) The wavefunction should reduce, in the limit of the two-body problem, to that of the scattering states with one node in the relative wavefunction.\ (3) The energy of the system must be larger than that of the non-interacting Fermi gas and, it should reduce to the perturbative result in the weakly interacting regime $0 < k_F a_s \ll 1$. We emphasize that orthogonality with the many-body ground state (of the BCS-BEC crossover) is not sufficient to be on the upper branch. LOCV ---- LOCV is an approximation scheme to evaluate the energy of the Jastrow-Slater state that originated in nuclear many-body physics [@Pand7173] and has been used for various quantum fluids. Within LOCV, the energy of the Jastrow-Slater state \[Eq. (9) of paper\] is given by \[ELOCV1\] =\_[F]{}n+d\^3r f\^\(r)f(r), where $v(r)$ is the short-range two-body potential, $n=n_\uparrow+n_\downarrow$ is the total density, and $\e_F=(3\pi^2 n)/2m$ is the Fermi energy. The effect of a zero-range contact potential may be written in terms of the Bethe-Peierls boundary condition $\lim_{r\to 0}(rf(r))'/(rf(r))=-1/a_s$. In addition, within LOCV, the Jastrow function $f(r)$ satisfies the following conditions: $f(r\geq d)=1$ and $f'(d)=0$. The “healing length" $d$ in turn is defined so that 2 n\_0\^[d]{} dr r\^2 f\^2(r)=1. \[cons\] Variation of the energy (\[ELOCV1\]) with respect to $f(r)$, while taking into account the constraint (\[cons\]) by a Lagrange multiplier $\lambda$, gives us the “Schrödinger” equation f([r]{})=f([r]{}). Retaining only the $s$-wave part of this equation, we find the general solution with one node is given by f(r)= where $\kappa=\sqrt{m\lambda}$. We find that $f(d)=1$ and $f'(d)=0$ lead to $\kappa d =\tan (\kappa (d - b))$ and the Bethe-Peierls boundary condition gives $\kappa a_s=\tan\kappa b$. With the normalization condition (\[cons\]), we can solve for $\kappa$, $d$ and $b$ for each value of scattering length $a_s$. The energy of the system is given simply by =\_[F]{}n+n. In the case of our interest, however, the system is not necessarily balanced, so that $x\equiv n_{\downarrow}/n_{\uparrow}$ need not be unity. We have to extend the LOCV calculation to the spin-imbalanced case in which the Slater determinants will have different sizes. We find the energy of the system is given by &=&\_[F]{}n\_(1+x\^)\ &+&n\_n\_d\^3r f\^\(r)f(r). Now, in general, there is no unique way to enforce the normalization conditions as in (\[cons\]). A natural extension is to use both normalizations 4 (n-n\_)\_0\^[d\_]{} dr r\^2 f\_\^2(r)=1,\[NORM\] which introduces two healing lengths, $d_{\uparrow}$ and $d_{\downarrow}$. Accordingly, we shall introduce two Lagrange multipliers $\lambda_\uparrow$ and $\lambda_\downarrow$ and the energy of the system can then be written as =\_[F]{}n\_+\_[F]{}n\_+(n\_\_+n\_\_).\[ELOCV2\] The above energy functional reduces to the usual perturbative result in the weak coupling limit and also, as we shall show later, reproduces the phase diagram of the interacting fermion system given by Monte Carlo methods [@Pilati10; @Chang11] with very good agreement. We note that the same energy functional (\[ELOCV2\]) was analyzed recently by Heiselberg [@Heiselberg10], in the context of the experiment of ref. [@Jo09]. In Fig. \[sup\_fig\] [(a)]{}, we show the upper branch energy (\[ELOCV2\]) as a function of polarization $P \equiv (1-x)/(1+x)$ for various values of $k_Fa_s$. For small $k_Fa_s$ (in fact, $k_Fa_s<0.91$), the minimum energy is attained at $P=0$. For larger values, $k_Fa_s\geq 1.124$, the minimum value is attained at $P=1$. For intermediate values of $k_Fa_s$, the minimum value occurs at a value of $0<P_c<1$. In Fig. \[sup\_fig\] [(b)]{}, we show the upper branch energy as a function of $1/k_Fa_s$ for various values of the polarization $P$. Note that for $P=1$, [i.e.]{}, completely polarized, the energy is completely flat, since there is no interaction between the polarized fermions. In Fig. \[sup\_fig\] [(c)]{}, we show the phase diagram of the system as obtained from LOCV. For $1/(k_Fa_s)>1.124$, the system is a homogeneous mixture of the two hyperfine-Zeeman states. We call this a Fermi liquid (FL) state. For $1.124>1/(k_Fa_s)>0.91$, the system is in the partially polarized phase (PP). Here the transition as predicted by the LOCV method is second order and accompanied by a divergent spin susceptibility [@Heiselberg10]. Note that the value of $1/(k_Fa_s)$ at the transition is very close to that predicted by Monte Carlo calculations [@Pilati10; @Chang11] and compares favorably with the calculation in Ref. [@Recati11]. Lastly, for $1/(k_Fa_s)<0.91$, the system is a fully polarized, non-interacting Fermi gas. Numerical details {#Appendixnumerical} ================= In this section, we explain how we use the LOCV energy in the numerical solution of the hydrodynamic equations (6) and (7) of the main text. The important quantity that enters both equations is =\_[F]{}+(\_+ n\_ + n\_). We express this in dimensionless form as \_=1+(\_-x -+x\_-x ). \[chem\] Here $\xi = 1/(k_F a_s)$ and all energies are scaled by $\epsilon_{F\uparrow}$, so that $\widetilde{\lambda}_\uparrow \equiv {\lambda_\uparrow}/{\epsilon_{F\uparrow}}$ etc. We note that at unitarity, $\xi=0$, and the expression involves only a single derivative ${\partial \widetilde{\lambda}_\downarrow}/{\partial x}$ that has to be evaluated numerically. For arbitrary $a_s$, however, the expression involves two additional numerical derivatives. To simplify the numerics at “small” $a_s$, we find that we can make a polynomial fit to the equation (\[chem\]) as a function of $x$ and $1/\xi$. we find that, for the values of $k_{F\uparrow}a_s \lesssim 2 $ we are considering, a good fit is provided by the expression \_=1+x( k\_[F]{}a\_s+ 0.27 (k\_[F]{}a\_s)\^2)+where the coefficient $0.27$ is very close to the value obtained by Galitskii [@galitskii], $\frac{4}{15\pi^2}(11-2\ln 2)\approx 0.259$. [99]{} I. Bloch, J. Dalibard, and W. Zwerger, Rev. Mod. Phys. 80, 885 (2008). S. Giorgini, L. P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. 80, 1215 (2008). A. Sommer, M. Ku, G. Roati, and M. W. Zwierlein, Nature 472, 201 (2011). S. Pilati, G. Bertaina, S. Giorgini, and M. Troyer, Phys. Rev. Lett. 105, 030405 (2010). S.-Y. Chang, M. Randeria, and N. Trivedi, Proc. Nat. Acad. Sci., 108, 51 (2011). J. Kinast, A. Turlapov, and J. E. Thomas, Phys. Rev. Lett. 94, 170404 (2005). P. Massignan, G. M. Bruun, and H. Smith, Phys. Rev. A 71, 033607 (2005). While hydrodynamics is valid in the overlap region, the atoms in the non-overlap region are non-interacting. However, due to Pauli repulsion, the non-overlapping regions essentially execute a rigid-body dipole oscillation, insensitive [@Vichi99] to the choice of a collisional (hydrodynamic) or collisionless description. J. A. Joseph, J. E. Thomas, M. Kulkarni, and A. G. Abanov, Phys. Rev. Lett. 106, 150401 (2011). S. Zhang and T.-L. Ho, New J. Phys. 13, 055003 (2011). L. Vichi and S. Stringari, Phys. Rev. A 60, 4734 (1999). G. M. Bruun, A. Recati, C. J. Pethick, H. Smith, and S. Stringari, Phys. Rev. Lett. 100, 240406 (2008). C. Cao, E. Elliott, J. Joseph, H. Wu, J. Petricka, T. Schäfer, and J. E. Thomas, Science 331, 58 (2011). L. J. Le Blanc, J. H. Thywissen, A. A. Burkov, and A. Paramekanti, Phys. Rev. A 80, 013607 (2009). G.-B. Jo, Y.-R. Lee, J.-H. Choi, C. A. Christensen, T. H. Kim, J. H. Thywissen, D. E. Pritchard, and W. Ketterle, Science 325, 1521 (2009). V. R. Pandharipande, Nucl. Phys. A [174]{}, 641 (1971); V. R. Pandharipande and H. A. Bethe, Phys. Rev. C [7]{}, 1312 (1973). H. Heiselberg, Phys. Rev. A 83, 053635 (2011). D. S. Petrov, Phys. Rev. A 67, 010703(R), (2003). We choose $T\approx 0.2 T_F$ leading to $F(T)=1$ in eq. (20) of Ref. [@Vichi99]. Note that $F(T)$ is weakly $T$-dependent for $0.2 T_F< T < T_F$. D. Pekker, M. Babadi, R. Sensarma, N. Zinner, L. Pollet, M. W. Zwierlein, and E. Demler, Phys. Rev. Lett. [106]{}, 050402 (2011). C. Sanner, E. J. Su, W. Huang, A. Keshet, J. Gillen, and W. Ketterle, arXiv:1108.2017v1. O. Goulko, F. Chevy, and C. Lobo, arXiv:1106.5773v2. A. Recati and S. Stringari, Phys. Rev. Lett. 106, 080402 (2011). V. M. Galitskii, Sov. Phys. JETP [7]{}, 104 (1958).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We derive constraints on corrections to Newtonian gravity of the Yukawa type and light elementary particles from two recently performed measurements of the gradient of the Casimir force. In the first measurement the configuration of two Au surfaces has been used, whereas in the second a nonmagnetic metal Au interacted with a magnetic metal Ni. In these configurations one arrives at different, respectively, similar theoretical predictions for the Casimir force when the competing theoretical approaches are employed. Nevertheless, we demonstrate that the constraints following from both experiments are in mutual agreement and in line with constraints obtained from earlier measurements. This confirms the reliability of constraints on non-Newtonian gravity obtained from measurements of the Casimir force.' author: - 'G. L. Klimchitskaya,${}^1$ U. Mohideen,${}^2$ and V. M. Mostepanenko' title: 'Constraints on non-Newtonian gravity and light elementary particles from measurements of the Casimir force by means of dynamic AFM' --- Introduction ============ Deviations from Newton’s gravitational law are presently discussed in many different aspects in connection with various extensions of the Standard Model and the problem of dark matter. The Yukawa-type corrections to Newtonian gravity in the wide interaction range from nanometers to $10^{15}\,$cm are predicted [@1; @2] by the extra-dimensional models with a low-energy compactification scale of the order of 1TeV [@3; @4; @5]. The modified Newtonian dynamics is discussed [@6] as an alternative to the proposed existence of dark matter which should comprise about 23% of the Universe mass. On the other hand, if it is granted that the most realistic according to astrophysics, cold, dark matter exists in nature, the question arises as to what are its constituents. One of the answers to this question is that the cold dark matter consists of non-thermally produced axions [@8]. The exchange of such type of particles between atoms of two macrobodies generates an effective Yukawa-type correction to Newton’s gravitational law with an interaction range $\lambda$ varying from 1[Å]{} to hundreds of meters or even larger depending on the mass of a particle $m=\hbar/(\lambda c)$. Constraints on corrections to Newtonian gravity are under investigation for many years [@9]. Strong constraints within the interaction range from millimeters to meters were obtained from the Cavendish- and Eötvos-type experiments [@10; @11]. In the submillimeter interaction range stronger constraints on the Yukawa-type corrections were currently derived from Cavendish-type experiments [@12; @13; @14]. The strength of these constraints, however, decreases with decreasing $\lambda$. For $\lambda$ smaller than a few micrometers the strongest constraints on the Yukawa-type corrections to Newtonian gravity follow from measurements of the Casimir force [@15; @16] which becomes the dominant background force in place of gravitation at sufficiently short separations between the test bodies (see review [@15a]). Measurements of the lateral Casimir force between sinusoidally corrugated surfaces of a sphere and a plate by means of an atomic force microscope (AFM) [@17; @18] and the gradient of the Casimir force between smooth surfaces of a sphere and a plate by means of a micromachined oscillator [@19; @20] have already resulted [@21] in up to several orders of magnitude stronger constraints. Constraints on the parameters of non-Newtonian gravity are commonly obtained from the measure of agreement between the experimental data for the Casimir force and theoretical predictions based on the Lifshitz theory [@15]. It is necessary to stress, however, that theory-experiment comparison in the field of the Casimir effect has led to unexpected results which are yet to be fully explained. Specifically, it was found that for metallic test bodies the theoretical predictions are in agreement with the experimental data if the relaxation properties of conduction electrons are omitted [@15; @16; @19; @20; @22; @23; @24; @25]. An experiment [@26], showing agreement between the data and theory with the relaxation properties included, measured the sum of the Casimir force and up to an order of magnitude larger residual electric force supposedly originating from large patches. It was not possible to independently measure this residual electric force. Because of this, the model description for it has been used containing two fitting parameters. The values of fitting parameters were determined from the fit between the experimental data and theory. As was pointed out in the literature [@27; @28], at short separations Ref. [@26] neglects the role of surface imperfections of a spherical lens with centimeter-size radius of curvature. It was also shown [@29; @30] that even with account of patch potentials, at large separations the experimental data of Ref. [@26] for the total force are in better agreement with theory neglecting the contribution of charge carriers. Note that the issue concerning the role of patch potentials has yet not been finally solved. The experimental data of Refs. [@22; @23; @24; @25] are consistent with theory of Ref. [@30a], predicting negligibly small contribution of patches in the configurations of these experiments. For an alternative theory of patches see Ref. [@30b] and its discussion in Ref. [@24]. For dielectric test bodies the theoretical predictions were found in agreement with the data only if the contribution of free charge carriers is omitted [@31; @32; @33; @34; @35; @36]. Keeping in mind that many of the experiments mentioned above were used, first, to make a selection between different theoretical approaches and, second, to constrain corrections to Newtonian gravity from the measure of agreement between the data and the predictions following from a selected approach, the constraints obtained were sometimes claimed to be of dubious merit [@37]. This claim is questionable [@38] because the difference between the excluded and confirmed theoretical approaches to the Casimir force cannot be modeled by the correction to Newtonian gravity of Yukawa type. It would be desirable, however, to have an independent confirmation of the previously obtained constraints that is not connected with a selection between different competing models of the Casimir force. In this paper we derive constraints on non-Newtonian gravity and on the parameters of possible constituents of dark matter following from two recent measurements of the gradient of the Casimir force by means of dynamic AFM operated in the frequency-shift technique [@24; @25; @39]. The first of these experiments [@24; @25] deals with a hollow, Au-coated glass sphere oscillating in close proximity to an Au-coated sapphire plate. In some sense the experiment [@24; @25] is similar to experiments [@19; @20; @22; @23], but it is performed using quite different laboratory setup (an AFM instead of a micromachined oscillator). The experimental data were compared with different theoretical approaches to the Casimir force and again were found to be in favor of the approach with the relaxation properties of conduction electrons omitted [@24; @25]. Here we obtain constraints on non-Newtonian gravity following from this experiment and demonstrate that they are in agreement with those obtained in Refs. [@19; @20] (but slightly weaker in the same proportion as the ratio of experimental errors in both experiments). In the second experiment considered in this paper the gradient of the Casimir force acting between an Au-coated hollow glass sphere and a Si plate coated with a ferromagnetic metal Ni was measured by means of a dynamic AFM [@39]. This configuration is of outstanding interest for constraining corrections to Newtonian gravity because, as was shown in Refs. [@40; @41], within the range of experimental separations one obtains almost coincident theoretical Casimir forces, irrespective of whether the relaxation properties are included or omitted. Thus, we arrive at constraints on the parameters of non-Newtonian gravity which are independent of a selection between different theoretical approaches to the Casimir force. These constraints turn out to be slightly stronger than those obtained from the AFM experiment with two Au surfaces, because the magnetic experiment extends to smaller separation distances. At the same time, the constraints obtained here from the magnetic experiment are slightly weaker than those obtained previously in Refs. [@19; @20] due to relatively lower precision of AFM measurements, as compared with measurements performed by means of a micromachined oscillator. Thus, the strongest constraints on non-Newtonian gravity derived in Refs. [@19; @20] receive further independent substantiation. The paper is organized as follows. In Sec. II we obtain constraints on non-Newtonian gravity following from the dynamic AFM experiment with two Au surfaces. In Sec. III the same is done using the measurement data of the AFM experiment with ferromagnetic plate. Section IV contains our conclusions and discussion. Constraints from the dynamic AFM experiment with two Au test bodies =================================================================== The experiment [@24; @25] considered in this section is in some analogy to earlier performed experiments of Refs. [@19; @20; @22; @23], but it uses a dynamic AFM instead of a micromechanical torsional oscillator as a measurement device. The gradient of the Casimir force was measured between an Au-coated hollow glass microsphere, attached to the cantilever of an AFM, and an Au-coated sapphire plate within the range of separations from 235 to 500nm. The thickness and density of the glass spherical envelope were $\Delta^{\!(s)}=5\,\mu$m and $\rho_{s}=2.5\times 10^{3}\,\mbox{kg/m}^3$. For technological purposes the glass sphere was coated first with a layer of Al having a thickness $\Delta_{\rm Al}^{\!(s)}=20\,$nm and density $\rho_{\rm Al}=2.7\times 10^{3}\,\mbox{kg/m}^3$ and then with a layer of Au having the thickness $\Delta_{\rm Au}^{\!(s)}=280\,$nm and density $\rho_{\rm Au}=19.28\times 10^{3}\,\mbox{kg/m}^3$. The external radius of a coated sphere was measured to be $R=41.3\,\mu$m [@24; @25]. The sapphire plate of density $\rho_{p}=4.1\times 10^{3}\,\mbox{kg/m}^3$, which can be considered as infinitely thick, was coated with an Al layer of thickness $\Delta_{\rm Al}^{\!(p)}=\Delta_{\rm Al}^{\!(s)}$ and with an external Au layer of thickness $\Delta_{\rm Au}^{\!(p)}=\Delta_{\rm Au}^{\!(s)}$. The Casimir interaction between the sphere and the plate measured in the experiment [@24; @25] coexists with Newtonian gravitation and possible corrections to it. For calculations of the Casimir force both test bodies can be considered as made of bulk Au [@15]. The gravitational force and corrections to it should be calculated, however, taking into account the layer structure of the sphere and plate. We admit that the corrections to Newtonian gravitation has the Yukawa form, so that the additional interaction to the Casimir interaction between the two point-like masses $m_1$ and $m_2$ can be described by the potential $$V(r)=V_N(r)+V_{\rm Yu}(r)=-\frac{Gm_1m_2}{r}\left( 1+\alpha e^{-r/\lambda}\right). \label{eq1}$$ Here, $r$ is the separation distance between the masses, $G$ is the Newtonian gravitational constant, $\alpha$ and $\lambda$ are the strength and interaction range of the Yukawa interaction. It can be easily seen [@42] that in the experimental configurations under consideration the Newtonian gravitational force is much smaller than the error in the measurement of the Casimir force and can be neglected. The total Yukawa interaction energy is obtained by the integration over the volumes of a sphere $V_s$ and a plate $V_p$: $$V_{\rm Yu}(a)=-G\alpha\int_{V_p}d^3r_1 \rho_p(\mbox{\boldmath$r$}_1)\int_{V_s}d^3r_2 \rho_s(\mbox{\boldmath$r$}_2) \frac{e^{-\|{\scriptsize{\mbox{\boldmath$r$}_1- \mbox{\boldmath$r$}_2}}\|/\lambda}}{\|\mbox{\boldmath$r$}_1- \mbox{\boldmath$r$}_2\|}, \label{eq2}$$ where $\rho_p(\mbox{\boldmath$r$}_1)$ and $\rho_s(\mbox{\boldmath$r$}_2)$ are the respective mass densities and $a$ is the closest separation between the bodies. Then the Yukawa force and its gradient are given by $$F_{\rm Yu}(a)=-\frac{\partial V_{\rm Yu}(a)}{\partial a}, \qquad \frac{\partial F_{\rm Yu}(a)}{\partial a} =-\frac{\partial^2 V_{\rm Yu}(a)}{\partial a^2}. \label{eq3}$$ After performing the integration in Eq. (\[eq2\]) with account of the layer structure of a sphere and a plate, one obtains from Eq. (\[eq3\]) [@43] $$\frac{\partial F_{\rm Yu}(a)}{\partial a}= -4\pi^2G\alpha\lambda^2e^{-a/\lambda}X^{(s)}(\lambda)X^{(p)}(\lambda), \label{eq4}$$ where the following notations are introduced: $$\begin{aligned} && X^{(p)}(\lambda)=\rho_{\rm Au}-(\rho_{\rm Au}-\rho_{\rm Al}) e^{-\Delta_{\rm Au}^{\!(p)}/\lambda} \nonumber \\ &&~~~~ -(\rho_{\rm Al}-\rho_{p}) e^{-(\Delta_{\rm Au}^{\!(p)}+\Delta_{\rm Al}^{\!(p)})/\lambda}, \nonumber \\ && X^{(s)}(\lambda)=\rho_{\rm Au}\Phi(R,\lambda)- (\rho_{\rm Au}-\rho_{\rm Al})\Phi(R-\Delta_{\rm Au}^{\!(s)},\lambda) e^{-\Delta_{\rm Au}^{\!(s)}/\lambda} \nonumber \\ &&~~ -(\rho_{\rm Al}-\rho_{s})\Phi(R-\Delta_{\rm Au}^{\!(s)}-\Delta_{\rm Al}^{\!(s)},\lambda) e^{-(\Delta_{\rm Au}^{\!(s)}+\Delta_{\rm Al}^{\!(s)})/\lambda} \nonumber \\ &&~~ -\rho_{s}\Phi(R-\Delta_{\rm Au}^{\!(s)}-\Delta_{\rm Al}^{\!(s)}-\Delta^{\!(s)},\lambda) e^{-(\Delta_{\rm Au}^{\!(s)}+ \Delta_{\rm Al}^{\!(s)}+\Delta^{\!(s)})/\lambda}, \nonumber \\ && \Phi(r,\lambda)=r-\lambda+(r+\lambda)e^{-2r/\lambda}. \label{eq5}\end{aligned}$$ The experimental data for the gradient of the Casimir force were found [@24; @25] to exclude the theoretical approach using the Lifshitz theory and the dielectric permittivity obtained from the tabulated optical data of Au extrapolated to zero-frequency by means of the Drude model. The same force data turned out to be consistent with theory when the simple plasma model is used for the extrapolation to zero frequency within the limits of experimental errors $\Delta_{F^{\prime}}(a)$ in the measurement of the force gradient which were determined at the 67% confidence level [@24; @25]. In the limits of these errors no interactions of Yukawa type were observed. Thus, the constraints on the parameters of Yukawa interaction $\alpha$ and $\lambda$ can be obtained from the inequality $$\left\|\frac{\partial F_{\rm Yu}(a)}{\partial a}\right\| \leq\Delta_{F^{\prime}}(a). \label{eq6}$$ Equations (\[eq4\]) and (\[eq5\]) were substituted in inequality (\[eq6\]). It was shown that the strongest constraints follow at the shortest separation $a=235\,$nm. In the measurement scheme with applied compensating voltage $\Delta_{F^{\prime}}(a)=0.50\,\mu$N/m [@24]. The resulting constraints are shown as line 1 in Fig. 1, where the region of $(\lambda,\alpha)$ above the line is prohibited and below the line is allowed by the results of this experiment. The interaction region from $\lambda_{\min}=20\,$nm to $\lambda_{\max}=3\,\mu$m shown in Fig. 1 corresponds to the masses of a hypothetical particle (axion, for instance) in the region from 66meV to 9.9eV. This overlaps with the typical mass scale from $m_a=1\,\mu$eV to $m_a=1\,$eV allowed for an axion [@44] and includes part of the region allowed by the cosmological bound for relic thermal axions $m_a<0.42\,$eV [@45], which are also considered in a cosmological context along with non-thermally produced axions. From the line 1 of Fig. 1 the allowed interaction strength at $\lambda=\lambda_{\min}=20\,$nm is $\alpha<5.9\times 10^{18}$. With increasing $\lambda$ (or, respectively, decreasing axion mass) the constraints shown by line 1 become stronger. At $\lambda=\lambda_{\max}=3\,\mu$m one obtains $\alpha<3.6\times 10^{10}$. In the next section the constraints of line 1 in Fig. 1 are compared with other constraints following from measurements of the Casimir force. In the end of this section we note that the constraints of line 1 in Fig. 1 were obtained using the exact expressions (\[eq4\]), (\[eq5\]) for the gradient of the Yukawa force between a sphere and a plate [@43]. Nearly the same results can be obtained in a more simple way by calculating the gradient of the Yukawa force in the framework of the proximity force approximation (PFA) $$\frac{\partial F_{\rm Yu}(a)}{\partial a}= -2\pi RP_{\rm Yu}(a), \label{eq7}$$ where $P_{\rm Yu}(a)$ is the Yukawa pressure between two parallel plates having the same layer structure as a plate and a sphere in the experiment under consideration. The approximate Eq. (\[eq7\]) is applicable under the conditions $$\frac{\lambda}{R}\ll 1, \qquad \frac{\Delta_{\rm Au}^{\!(s)}+\Delta_{\rm Al}^{\!(s)}+ \Delta^{\!(s)}}{R}\ll 1, \label{eq8}$$ which are satisfied in our case with a wide safety margin. Using Eq. (\[eq7\]), one returns back to Eqs. (\[eq4\]) and (\[eq5\]) where the function $\Phi(r,\lambda)$ with any argument $r$ is replaced with $R$ [@43; @46]. Now we compare the strength of constraints obtained using the exact Yukawa force and the approximate one derived using the PFA. Thus, at $\lambda$ equal to 1 and $3\,\mu$m (recall that at larger $\lambda$ the accuracy of the PFA is lower) the exact constraints are given by $\alpha<2.19\times 10^{11}$ and $\alpha<3.62\times 10^{10}$, respectively. These should be compared with respective PFA results $\alpha<2.17\times 10^{11}$ and $\alpha<3.51\times 10^{10}$. The comparison shows that the use of the PFA results in only 0.9% and 3.0% relative errors at $\lambda=1$ and $3\,\mu$m, respectively. Constraints from the dynamic AFM experiment with a magnetic plate ================================================================= The experiment described in the previous section was first used for a selection between two different theoretical approaches to the Casimir force. Then we have used the measure of agreement between the selected approach and the measurement data for obtaining constraints on non-Newtonian gravity of the Yukawa type. This procedure is in fact well justified because the difference between the two approaches to the Casimir force cannot be modeled by the Yukawa interaction with some $\lambda$ and $\alpha$. It would be interesting, however, to independently verify the constraints obtained in such a way by using the measurement data consistent with the Lifshitz theory of the Casimir force without additional conditions. A good opportunity for this is provided by the recent measurement of the gradient of the Casimir force between the hollow Au-coated glass sphere and Si plate coated with ferromagnetic metal Ni [@39]. The experiment on measuring the gradient of the Casimir force between magnetic and nonmagnetic metals was performed by means of a dynamic AFM in the configuration of a sphere and a plate. The layer structure of a sphere was the same as in the experiment of Sec. II, but its external radius was equal to $R=64.1\,\mu$m [@39]. The Si plate of density $\rho_{\rm Si}=2.33\times 10^{3}\,\mbox{kg/m}^3$ was coated with a Ni layer of thickness $\Delta_{\rm Ni}^{\!(p)}=154\,$nm and density $\rho_{\rm Ni}=8.9\times 10^{3}\,\mbox{kg/m}^3$ with no additional intermediate layer. As a result, the exact expression for the gradient of the Yukawa force is again given by Eq. (\[eq4\]), where $X^{(s)}(\lambda)$ is contained in Eq. (\[eq5\]) and $X^{(p)}(\lambda)$ takes a more simple form $$X^{(p)}(\lambda)=\rho_{\rm Ni}-(\rho_{\rm Ni}-\rho_{\rm Si}) e^{-\Delta_{\rm Ni}^{\!(p)}/\lambda}. \label{eq9}$$ The measurement data for the gradient of the Casimir force over the entire separation region from 220 to 500nm were found consistent with the Lifshitz theory combined with the dielectric permittivities of Au and Ni. The latter were obtained by using the available optical properties of both metals extrapolated to zero frequencies by means of either the Drude or the plasma models. The important characteristic feature of this case is that, by coincidence, over the region of experimental separations the predictions of the Drude model approach almost coincide with the predictions of the plasma model approach. Thus, the experimental data for the gradient of the Casimir force were found in agreement with the Lifshitz theory with no additional selection process among the theoretical models. This makes possible to use the magnetic experiment as an additional independent test for the constraints on corrections to Newtonian gravitation obtained previously. The strongest constraints follow at the shortest separation distance $a=220\,$nm. They are obtained from Eq. (\[eq6\]), where now $\Delta_{F^{\prime}}(a)=0.79\,\mu$N/m (the larger experimental error determined at the same 67% confidence level as in the experiment of Sec. II is connected with the larger value of the sphere radius), by the substitution of Eq. (\[eq4\]), the quantity $X^{(s)}$ defined in Eq. (\[eq5\]) and $X^{(p)}$ defined in Eq. (\[eq9\]). The results are shown with line 2 in Fig. 1. As can be seen in Fig. 1, both the lines 1 and 2 show qualitatively similar constraints. In the region from $\lambda=\lambda_{\min}=20\,$nm to $\lambda=\lambda_0=0.5\,\mu$m the constraints of line 2 (the experiment with magnetic metal) are slightly stronger than the constraints of line 1 obtained from the experiment with nonmagnetic metals. Furthermore, in the separation region from $\lambda=\lambda_0=0.5\,\mu$m to $\lambda=\lambda_{\max}=3\,\mu$m the constraints of the line 2 are slightly weaker than the constraints of the line 1. This is explained by different minimum separations, where these experiments have been performed, different sphere radii, and different materials of the plate used. There are, however, no qualitative differences that might be connected with the fact that the experimental data of the measurement with nonmagnetic bodies were used for making a selection between the two different theoretical approaches to the Casimir force, whereas the experimental data of the measurement with a magnetic plate were not. From line 2, at $\lambda=20\,$nm (the axion mass $m_a=9.9\,$eV) the interaction strength of the Yukawa-type correction to Newtonian gravity is constrained by $\alpha<4.2\times 10^{18}$ and at $\lambda=3\,\mu$m (the axion mass $m_a=66\,$meV) by $\alpha<5.6\times 10^{10}$. For comparison purposes in Fig. 2 we again plot line 2 representing the constraints on the Yukawa interaction obtained from the experiment using a magnetic plate [@39] independently of any selection between the different theoretical approaches to the Casimir force. In the same figure, the constraints following from the dynamic determination of the Casimir pressure by means of a micromachined oscillator [@19; @20] are shown by line 3, from the Casimir-less experiment [@47], where the Casimir force was compensated, are shown by line 4, from measurement of the Casimir-Polder force between rubidium atoms belonging to the Bose-Einstein condensate and SiO${}_2$ plate [@33] are shown by line 5 [@21], and from measurement of the Casimir force between smooth Au-coated sphere and Si plate covered with nanoscale trapezoidal corrugations [@48] are shown by line 6 [@49]. The constraints of Ref. [@49a] obtained from an experiment [@26] are not shown in Fig. 2 because they are stronger than those of lines 4 and 5 only at large interaction ranges ${\rm log}_{10}[\lambda\,\mbox{(m)}] > - 6.38$ and are not characterized by some definite confidence level. Note that the constraints of lines 3 and 4 were determined at a 95% confidence level, and the constraints of lines 5 and 6, as well as of line 2, were found at a 67% confidence level. From Fig. 2 it can be seen that although the constraints of line 2 are not the strongest ones, they are quite competitive within some interaction range and, as it is independent of a selection process between different theoretical approaches to the Casimir force, they provide additional support to the constraints obtained from other experiments. Conclusions and discussion ========================== In the foregoing, we have obtained constraints on the Yukawa-type corrections to Newton’s gravitational law and light elementary particles which follow from the two recently performed experiments on the Casimir force. In the first of these experiments the gradient of the Casimir force between two Au surfaces has been measured by means of a dynamic AFM [@24; @25]. This experiment, as well as the previous measurement performed using another laboratory technique [@19; @20], was used for both the selection between two competing theoretical approaches to the Casimir force and for constraining corrections to Newtonian gravity. In the second recently performed experiment the dynamic AFM was used to measure the gradient of the Casimir force between a nonmagnetic metal Au and a ferromagnetic metal Ni [@39]. The unique feature of this experiment is that in the region of experimental separations both competing approaches to the theoretical description of the Casimir force lead to nearly coincident results. For this experiment the unambiguous theoretical prediction was found in agreement with the experimental data and this fact has been used for obtaining constraints on the corrections to Newtonian gravity. Good agreement between the constraints obtained from both recent experiments and from other experiments on measuring the Casimir force performed earlier was demonstrated. This allows to conclude that in spite of the widely known problems in theory-experiment comparison discussed in the literature, measurements of the Casimir force in laboratory remain a reliable source of constraints on non-Newtonian gravity of the Yukawa-type and light elementary particles within the interation range from nanometers to micrometers. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by the DFG grant No. BO 1112/21–1 (G.L.K. and V.M.M.) and by the NSF Grant No. PHY0970161 (U.M.). G.L.K. and V.M.M. are grateful to the Institute for Theoretical Physics, Leipzig University for their kind hospitality. [99]{} E. G. Floratos and G. K. Leontaris, Phys. Lett. B [465]{}, 95 (1999). A. Kehagias and K. Sfetsos, Phys. Lett. B [472]{}, 39 (2000). I. Antoniadis, N. Arkani-Hamed, S. Dimopoulos, and G. Dvali, Phys. Lett. B [436]{}, 257 (1998). N. Arkani-Hamed, S. Dimopoulos, and G. Dvali, Phys. Lett. B [429]{}, 263 (1998). N. Arkani-Hamed, S. Dimopoulos, and G. Dvali, Phys. Rev. D [59]{}, 086004 (1999). J. D. Bekenstein, Contemp. Phys. [47]{}, 387 (2006). R. D. Peccei and H. R. Quinn, Phys. Rev. Lett. [38]{}, 1440 (1977). E. Fischbach and C. L. Talmadge, [The Search for Non-Newtonian Gravity]{} (Springer, New York, 1999). E. G. Adelberger, B. R. Heckel, C. W. Stubbs, and W. F. Rogers, Ann. Rev. Nucl. Part. Sci. [41]{}, 269 (1991). E. G. Adelberger, B. R. Heckel, and A. E. Nelson, Ann. Rev. Nucl. Part. Sci. [53]{}, 77 (2003). S. J. Smullin, A. A. Geraci, D. M. Weld, J. Chiaverini, S. Holmes, and A. Kapitulnik, Phys. Rev. D [72]{}, 122001 (2005). A. A. Geraci, S. J. Smullin, D. M. Weld, J. Chiaverini, and A. Kapitulnik, [Phys. Rev. D]{} [78]{}, 022002 (2008). E. G. Adelberger, J. H. Gundlach, B. R. Heckel, S. Hoedl, and S. Schlamminger, Progr. Part. Nucl. Phys. [62]{}, 102 (2009). M. Bordag, G. L. Klimchitskaya, U. Mohideen, and V. M. Mostepanenko, [Advances in the Casimir Effect]{} (Oxford University Press, Oxford, 2009). G. L. Klimchitskaya, U. Mohideen, and V. M. Mostepanenko, Rev. Mod. Phys. [81]{}, 1827 (2009). R. Onofrio, New J. Phys. [8]{}, 237 (2006). H.-C. Chiu, G. L. Klimchitskaya, V. N. Marachevsky, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [80]{}, 121402(R) (2009). H.-C. Chiu, G. L. Klimchitskaya, V. N. Marachevsky, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [81]{}, 115417 (2010). R. S. Decca, D. López, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, and V. M. Mostepanenko, Phys. Rev. D [75]{}, 077101 (2007). R. S. Decca, D. López, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, and V. M. Mostepanenko, Eur. Phys. J. C [51]{}, 963 (2007). V. B. Bezerra, G. L. Klimchitskaya, V. M. Mostepanenko, and C. Romero, Phys. Rev. D [81]{}, 055003 (2010). R. S. Decca, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, D. López, and V. M. Mostepanenko, Phys. Rev. D [68]{}, 116003 (2003). R. S. Decca, D. López, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, and V. M. Mostepanenko, Ann. Phys. (N.Y.) [318]{}, 37 (2005). C.-C. Chang, A. A. Banishev, R. Castillo-Garza, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [85]{}, 165443 (2012). C.-C. Chang, A. A. Banishev, R. Castillo-Garza, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Int. J. Mod. Phys.: Conf. Ser. [14]{}, 270 (2012). A. O. Sushkov, W. J. Kim, D. A. R. Dalvit, and S. K. Lamoreaux, [Nature Phys.]{} [7]{}, 230 (2011). V. B. Bezerra, G. L. Klimchitskaya, U. Mohideen, V. M. Mostepanenko, and C. Romero, [Phys. Rev. B]{} [83]{}, 075417 (2011). G. L. Klimchitskaya and V. M. Mostepanenko, Int. J. Mod. Phys. A [26]{}, 3944 (2011). G. L. Klimchitskaya, M. Bordag, E. Fischbach, D. E. Krause, and V. M. Mostepanenko, Int. J. Mod. Phys. A [26]{}, 3918 (2011). G. L. Klimchitskaya, M. Bordag, and V. M. Mostepanenko, Int. J. Mod. Phys. A [27]{}, 1260012 (2012). C. C. Speake and C. Trenkel, Phys. Rev. Lett. [90]{}, 160403 (2003). R. O. Behunin, F. Intravaia, D. A. R. Dalvit, P. A. Maia Neto, and S. Reynaud, Phys. Rev. A [85]{}, 012504 (2012). F. Chen, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Optics Express [15]{}, 4823 (2007). F. Chen, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [76]{}, 035338 (2007). J. M. Obrecht, R. J. Wild, M. Antezza, L. P. Pitaevskii, S. Stringari, and E. A. Cornell, Phys. Rev. Lett. [98]{}, 063201 (2007). G. L. Klimchitskaya and V. M. Mostepanenko, J. Phys. A: Math. Theor. [41]{}, 312002 (2008). C.-C. Chang, A. A. Banishev, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. Lett. [107]{}, 090403 (2011). A. A. Banishev, C.-C. Chang, R. Castillo-Garza, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [85]{}, 045436 (2012). A. Lambrecht, A. Canaguier-Durand, R. Guérout, and S. Reynaud, In: [Casimir Physics]{}, Lecture Notes in Physics, [834]{}, p.97, eds. D. A. R. Dalvit, P. W. Milonni, D. C. Roberts, and F. S. S. Rosa (Springer, Heidelberg, 2011). V. M. Mostepanenko, V. B. Bezerra, G. L. Klimchitskaya, and C. Romero, Int. J. Mod. Phys. A [27]{}, 1260015 (2012). A. A. Banishev, C.-C. Chang, G. L. Klimchitskaya, V. M. Mostepanenko, and U. Mohideen, Phys. Rev. B [85]{}, 195422 (2012). B. Geyer, G. L. Klimchitskaya, and V. M. Mostepanenko, Phys. Rev. B [81]{}, 104101 (2010). G. L. Klimchitskaya, B. Geyer, and V. M. Mostepanenko, Int. J. Mod. Phys. A [25]{}, 2293 (2010). M. Bordag, B. Geyer, G. L. Klimchitskaya, and V. M. Mostepanenko, [Phys. Rev. D]{} [62]{}, 011701(R) (2000). R. S. Decca, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, D. López, and V. M. Mostepanenko, Phys. Rev. D [79]{}, 124021 (2009). G. G. Raffelt, J. Redondo, and N. V. Maira, Phys. Rev. D [84]{}, 103008 (2011). A. Melchiorri, O. Mena, and A. Slosar, Phys. Rev. D [76]{}, 041303(R) (2007). E. Fischbach, G. L. Klimchitskaya, D. E. Krause, and V. M. Mostepanenko, Eur. Phys. J. C [68]{}, 223 (2010). R. S. Decca, D. López, E. Fischbach, D. E. Krause, and C. R. Jamell, Phys. Rev. Lett. [94]{}, 240401 (2005). Y. Bao, R. Guérout, J. Lussange, A. Lambrecht, R. A. Cirelli, F. Klemens, W. M. Mansfield, C. S. Pai, and H. B. Chan, [Phys. Rev. Lett.]{} [105]{}, 250402 (2010). V. B. Bezerra, G. L. Klimchitskaya, V. M. Mostepanenko, and C. Romero, Phys. Rev. D [83]{}, 075004 (2011). A. O. Sushkov, W. J. Kim, D. A. R. Dalvit, and S. K. Lamoreaux, Phys. Rev. Lett. [107]{}, 171101 (2011). ![(color online). Constraints on the parameters of Yukawa-type correction to Newton’s gravitational law obtained from measurements of the gradient of the Casimir force by means of a dynamic AFM in the configuration of an Au-coated sphere and an Au-coated plate (line 1) and in the configuration of an Au-coated sphere and a magnetic Ni-coated plate (line 2). The regions of $(\lambda,\alpha)$ plane below each line are allowed and above each line are prohibited. ](figYuD-1.ps) ![(color online). Constraints on the parameters of Yukawa-type correction to Newton’s gravitational law obtained from experiments on the Casimir force with a magnetic plate performed by means of a dynamic AFM [@39] (line 2) and with two Au bodies performed by means of micromachined oscillator [@19; @20] (line 3), from the Casimir-less experiment [@47] (line 4), from measurements [@33] of the Casimir-Polder force [@21] (line 5), and from measurements of the Casimir force between an Au sphere and corrugated Si plate [@48; @49] (line 6). The regions of $(\lambda,\alpha)$ plane below each line are allowed and above each line are prohibited. ](figYuD-2.ps)	{ "pile_set_name": "ArXiv" }
--- abstract: 'Stability of recurrent models is closely linked with trainability, generalizability and in some applications, safety. Methods that train stable recurrent neural networks, however, do so at a significant cost to expressibility. We propose an implicit model structure that allows for a convex parametrization of stable models using contraction analysis of non-linear systems. Using these stability conditions we propose a new approach to model initialization and then provide a number of empirical results comparing the performance of our proposed model set to previous stable RNNs and vanilla RNNs. By carefully controlling stability in the model, we observe a significant increase in the speed of training and model performance.' author: - \| \ \ Australian Center for Field Robotics,\ Sydney Institute for Robotics and Intelligent Systems,\ The University of Sydney,\ NSW, 2006. Australia bibliography: - 'Refs\_edited.bib' title: \| Contracting Implicit Recurrent Neural Networks:\ Stable Models with Improved Trainability --- System Identification, Contraction, Stability, Recurrent Neural Network, Vanishing Gradient, Exploding Gradient, Nonlinear Systems, Echo State Network	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recently, a framework has been developed to study form factors of two-hadron states probed by an external current. The method is based on relating finite-volume matrix elements, computed using numerical lattice QCD, to the corresponding infinite-volume observables. As the formalism is complicated, it is important to provide non-trivial checks on the final results and also to explore limiting cases in which more straightforward predications may be extracted. In this work we provide examples on both fronts. First, we show that, in the case of a conserved vector current, the formalism ensures that the finite-volume matrix element of the conserved charge is volume-independent and equal to the total charge of the two-particle state. Second, we study the implications for a two-particle bound state. We demonstrate that the infinite-volume limit reproduces the expected matrix element and derive the leading finite-volume corrections to this result for a scalar current. Finally, we provide numerical estimates for the expected size of volume effects in future lattice QCD calculations of the deuteron’s scalar charge. We find that these effects completely dominate the infinite-volume result for realistic lattice volumes and that applying the present formalism, to analytically remove an infinite-series of leading volume corrections, is crucial to reliably extract the infinite-volume charge of the state.' author: - 'Raúl A. Briceño' - 'Maxwell T. Hansen' - 'Andrew W. Jackura' bibliography: - 'bibi.bib' title: \| Consistency checks for two-body finite-volume matrix elements:\ I. Conserved currents and bound states --- Introduction {#sec:intro} ============ One of the overarching goals of modern-day nuclear physics is the characterization and fundamental understanding of the low-lying strongly-interacting spectrum. There is, by now, overwhelming evidence that the detailed properties of all low-lying states are governed by the dynamics of quark and gluon fields in the mathematical framework of quantum chromodynamics (QCD). But still, it remains a significant challenge to extract low-energy predictions from the underlying theory. The vast majority of QCD states emerge as either bound states or resonances of multi-hadron configurations. An example is the deuteron, a shallow bound state of the isoscalar proton-neutron channel with a binding energy of $m_n + m_p - M_{d} \approx 2.2$ MeV. The deuteron has long been hypothesized to be a molecular state of the two nucleons [@Weinberg:1962hj] and similar pictures have been proposed for a variety of other QCD states. (See Ref. [@Guo:2017jvc] for a recent review.) However, in many cases a straightforward interpretation is unavailable. For example, the isoscalar $f_0(980)$ resonance couples strongly to $\pi\pi$ and $K\overline{K}$ states, and has been postulated to be both a tetraquark [@Jaffe:1976ig] and a $K\overline{K}$ molecule [@Weinstein:1990gu].[^1] The challenge of resolving the inner structure of composite hadrons is twofold: First, QCD is non-perturbative, so that systematic low-energy calculations are challenging. This has been addressed with substantial success using low-energy effective theories, methods based in amplitude analysis and numerical calculations using lattice QCD (LQCD). In contrast to the first two methods, LQCD has the unique advantage of relating the fundamental QCD lagrangian to low-energy predictions. Second, composite states generally manifest as dynamical enhancements of multi-hadron scattering rates, meaning that the detailed observation depends on the production mechanism and decay channel of the resonance in question. This ambiguity is resolved, at least in principle, by recognizing that across all production and decay channels, a given resonance always leads to the same pole in an analytic continuation of scattering amplitudes to complex energies. These two points have motivated the community to develop a systematic framework for extracting hadronic scattering amplitudes via LQCD. From the energy dependence of such amplitudes one can then quantitatively describe the bound and resonant states of the theory. In addition, by extracting transition amplitudes involving external currents, one can in principle access structural information of these states. In this work, we focus on an example in the latter class of the amplitudes, namely ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ transition amplitudes. We consider a method, first introduced in Refs. [@Briceno:2015tza; @Baroni:2018iau], that allows one to determine such quantities from numerical LQCD. The primary formal challenge arises from the fact that LQCD calculations are necessarily performed in a finite Euclidean spacetime, where the definition of asymptotic states is obscured. One of the leading methods to overcome this issue is to derive and apply non-perturbative mappings between finite-volume energies and matrix elements (directly calculable via numerical LQCD) and infinite-volume scattering and transition amplitudes.[^2] This approach was first introduced by Lüscher [@Luscher:1986pf; @Luscher:1991n1], in seminal work relating the spectrum of two-particle states in a cubic volume with periodicty $L$, to the corresponding infinite-volume amplitudes. The idea has since been extended for arbitrary two-particle scattering [@Rummukainen:1995vs; @Kim:2005gf; @He:2005ey; @Davoudi:2011md; @Hansen:2012tf; @Briceno:2012yi; @Briceno:2013lba; @Briceno:2014oea; @Romero-Lopez:2018zyy] and more recently to three particles [@Hansen:2014eka; @Hansen:2015zga; @Mai:2017bge; @Hammer:2017kms; @Briceno:2017tce; @Briceno:2018aml; @Mai:2018djl; @Briceno:2018mlh; @Guo:2018zss; @Blanton:2019igq], with the latter currently limited to identical scalars (or pseudoscalars). The two-particle relations have made possible the determination of hadronic scattering amplitudes for a wide range of particle species [@Dudek:2010ew; @Beane:2011sc; @Pelissier:2012pi; @Dudek:2012xn; @Liu:2012zya; @Beane:2013br; @Orginos:2015aya; @Berkowitz:2015eaa; @Lang:2015hza; @Bulava:2016mks; @Hu:2016shf; @Alexandrou:2017mpi; @Bali:2017pdv; @Bali:2017pdv; @Wagman:2017tmp; @Andersen:2017una; @Brett:2018jqw; @Werner:2019hxc; @Mai:2019pqr; @Wilson:2019wfr], including energies where multiple channels are kinematically open [@Wilson:2014cna; @Dudek:2014qha; @Wilson:2015dqa; @Dudek:2016cru; @Briceno:2016mjc; @Moir:2016srx; @Briceno:2017qmb; @Woss:2018irj; @Woss:2019hse]. Most recently, the first LQCD calculations to constrain three-particle interactions using excited states were performed in Refs. . Electroweak interactions involving scattering states can also be accessed using LQCD, via a generalization of the methods described above. The seminal example in this sector is the work of Ref. [@Lellouch:2000pv], providing a formal method for determining the electroweak decay, $K \to \pi \pi$. More generally, in processes for which the effects of the electroweak sector can be treated perturbatively, the relevant amplitudes are given via the evaluation of QCD matrix elements, built from the appropriate currents together with multi-particle external states. These ideas have been successfully developed for the case that either the initial or the final state couples strongly to two-particle scattering states [@Lellouch:2000pv; @Kim:2005gf; @Christ:2005gi; @Hansen:2012tf; @Briceno:2014uqa; @Briceno:2015csa; @Agadjanov:2016fbd] and implemented in a number lattice QCD studies, most prominently to determine the $K \to \pi \pi$ decay amplitudes as well as the electromagnetic process $\pi \gamma^\star \to \pi \pi$ [@Feng:2014gba; @Briceno:2015dca; @Briceno:2016kkp; @Feng:2018pdq; @Alexandrou:2018jbt]. This progress motivates the consideration of more complicated electroweak transitions, in particular those with two hadrons in both the initial and final state. As we discuss in detail in Sec. \[sec:FFs\], ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ transition amplitudes allow one to extract elastic form factors of bound states and resonances, thereby providing direct information on the structure of these states and possibly resolving which models are most descriptive [@Kaplan:1998sz; @Chen:1999tn; @Albaladejo:2012te]. As compared to the transitions described in the preceding paragraph, the necessary formalism for these quantities is significantly more complicated [@Bernard:2012bi; @Briceno:2012yi; @Briceno:2015tza; @Baroni:2018iau].[^3] Building on previous work, in Ref. [@Briceno:2015tza] two of us derived a model-independent relation between the corresponding finite-volume matrix elements, schematically denoted $\langle 2\|\mathcal{J}\|2\rangle_L$ (where $L$ indicates the side-length of the periodic cubic volume), and the ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ transition amplitude, $\mathcal{W}$. In Ref. [@Baroni:2018iau] we improved the method by simplifying technical details relating to the on-shell projection of the single-particle form factor and by using Lorentz covariant poles in the various finite-volume kinematic functions that arise. We stress that the two approaches are equivalent and only differ in the exact definitions of unphysical, intermediate quantities. The results are derived to all orders in the perturbative expansion of a generic relativistic field theory, for any type of two-scalar channels, with generalizations to spin and coupled channels left to future work. Details of this formalism are reviewed in Sec. \[sec:FV\_fcns\]. The purpose of this work is to provide two non-trivial checks on the general relations of Refs. [@Briceno:2015tza; @Baroni:2018iau], and also to demonstrate their predictive power even in simplified special cases. As a first check, in Sec. \[sec:WTI\] we demonstrate that the method is consistent with the consequences of the conserved vector current. In particular, the formalism predicts that the charge of a two-hadron finite-volume state is exactly equal to the sum of the constituent charges and independent of $L$. This relies on non-trivial relations between various $L$-dependent geometric functions, and a relation between the ${\mathbf{2}}\to {\mathbf{2}}$ and ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitudes that follows from the Ward-Takahashi identity. The second check, presented in Sec. \[sec:FVME\], considers the analytic continuation of the formalism below two-particle threshold, for theories with an $S$-wave bound state. We show that the finite- and infinite-volume matrix elements coincide (once normalization factors are accounted for) up to term scaling as $e^{- \mathcal \kappa_{\text{B}} L}$, where $\kappa_{\text{B}}^2 = m^2 - M_{\text{B}}^2/4$ defines the binding momentum for two constituents of mass $m$, binding to a mass of $M_{\text{B}}$. Presently, LQCD calculations of light nuclei properties are being performed at unphysically heavy quark masses, for which the binding momenta exceed their real-world values [@Beane:2012vq; @Yamazaki:2015asa; @Berkowitz:2015eaa; @Wagman:2017tmp; @Francis:2018qch]. In addition the properties of states can be shifted, e.g. the dineutron, in nature a virtual bound state, is found to be a standard bound state for $m_\pi \gtrsim 450 {\ensuremath{{\mathrm{\,Me\kern -0.1em V}}}\xspace}$ [@Beane:2011iw; @Yamazaki:2011nd; @Yamazaki:2012hi; @Beane:2013br; @Orginos:2015aya]. The increased binding suppresses finite-volume effects and this has permitted exploratory calculations of matrix elements of these states [@Beane:2014ora; @Chang:2017eiq; @Savage:2017prl; @Winter:2017bfs; @Tiburzi:2017iux], in which volume effects are ignored. As LQCD calculations of multi-nucleon systems move towards physical quark masses, the binding momenta of the nuclei decrease and it is well-known that finite-volume effects of the naively extracted states can become a dominant source of systematic uncertainty [@Davoudi:2011md; @Briceno:2013bda; @Briceno:2013hya]. In the case of spectroscopy, an infinite series of $e^{- \kappa_{\text{B}} L}$ corrections can be removed by applying the Lüscher formalism, as was done in [@Francis:2018qch; @Berkowitz:2015eaa] as well as in a wide variety of mesonic channels where bound states appear [@Wilson:2014cna; @Dudek:2014qha; @Lang:2015hza; @Briceno:2016mjc; @Moir:2016srx; @Bali:2017pdv]. The results of this work stress that it is important to pursue the same paradigm for matrix elements of loosely bound states, using the formalism of Refs. [@Briceno:2015tza; @Baroni:2018iau] to non-perturbatively remove binding-momentum-enhanced finite-volume artifacts. To illustrate this point, in Sec. \[sec:FVME\] we determine the leading $e^{- \kappa_{\text{B}} L}$ corrections and compare these to the full result, which holds up to $e^{- m L}$. Finally, in Sec. \[sec:num\_expectation\] we present a numerical example, meant to model the deuteron at physical pion masses, and show that the full formalism is needed to reliably remove the $L$-dependence for box sizes in the region of $mL \approx 4 - 7$. Otherwise the $e^{- \kappa_{\text{B}} L}$ corrections can become comparable in size with the infinite-volume result and thereby dominate the systematic uncertainties. Though largely addressed above, we close here with a brief summary of the remaining sections. After reviewing basic properties of the infinite-volume ${\mathbf{2}}\to {\mathbf{2}}$ and ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitudes in Sec. \[sec:FFs\], in Sec. \[sec:FV\_fcns\] we described the corresponding finite-volume formalism for each type of amplitude. Then, in a very compact Sec. \[sec:WTI\], we demonstrate that the finite-volume ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ formalism gives the expected results for matrix elements of a conserved current. Section \[sec:BS\_FV\] is dedicated to volume effects on a two-particle bound state, including a check that the $L \to \infty$ limit gives the required result, a calculation of the leading $\mathcal O(e^{- \kappa_{\text{B}} L})$ corrections, and a numerical exploration intended to guide future LQCD calculations of the deuteron’s scalar charge. We briefly conclude in Sec. \[sec:conclusion\]. In addition, this article includes three appendices, providing proofs of various technical results used in the main text. Infinite-volume amplitudes and bound states {#sec:FFs} =========================================== In this section we review the definitions and key properties of the infinite-volume ${\mathbf{2}}\to{\mathbf{2}}$ and ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ amplitudes, with particular attention to the expressions relevant for an $S$-wave bound state. For simplicity, we focus on systems composed of two scalar particles, with degenerate mass $m$, distinguished by their charge with respect to an external current $\mathcal J^\mu$. One of the particles carries charge ${\mathrm Q_0}$, while the other is neutral. Here we have in mind a scalar analog of the proton-neutron system. ${\mathbf{2}}\to{\mathbf{2}}$ amplitudes and bound-state poles {#sec:2to2} -------------------------------------------------------------- In a general Lorentz frame, the two-particle system has a total energy-momentum denoted by $P = (E,{\mathbf{P}})$. Boosting to the center-of-momentum frame (CMF) we define $P^\star = (E^\star, {\mathbf{0}})$, which is related to the Mandelstam variable $s$ and a generic $P$ by $$\label{eq:s_inv} E^{\star\,2} \equiv s \equiv P_\mu P^\mu = E^2 - {\mathbf{P}}^2.$$ Two-particle scattering is described by $s$, as well as the back-to-back momentum orientations of the initial and final states in the CMF: ${\mathbf{\hat{ {\mathbf{k}}}}}_{i}^{\star}$ and ${\mathbf{\hat{ {\mathbf{k}}}}}_{f}^{\star}$, respectively.[^4] Using these coordinates we can introduce the scattering amplitude and its partial wave expansion$$\label{eq:M_expand} {\mathcal{M}}(s,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star},{\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) = 4\pi \sum_{\ell, m_{\ell}} Y_{\ell m_{\ell}}({\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}) \, {\mathcal{M}}_{\ell} (s) \,Y_{\ell m_{\ell}}^{}({\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \,.$$ We have used that total angular momentum, $\ell$, is conserved, and that the partial-wave amplitude is independent of the projection, $m_{\ell}$, both consequences of rotational symmetry. In the following we are interested in the case of a scalar bound state, appearing as a sub-threshold pole in ${\mathcal{M}}_{\ell = 0} (s)$. We therefore restrict attention to the $S$-wave ($\ell = 0$) amplitude and do not write the angular momentum index on the partial wave amplitude for the rest of this section. The elastic ${\mathbf{2}}\to{\mathbf{2}}$ scattering amplitude can be represented in terms of the K matrix, which is an on-shell representation that enforces S matrix unitarity explicitly below the inelasticity threshold [@Martin:102663], $$\label{eq:Kmat_rep} {\mathcal{M}}(s) = {\mathcal{K}}(s) \frac{1}{1 - i\rho(s) \, {\mathcal{K}}(s)}.$$ Here, $\rho(s)$ is the two-body phase space, encoding the on-shell propagation of two particles. It is defined as $$\label{eq:rho} \rho(s) = \frac{q^{\star}}{8\pi E^{\star}} = \frac{1}{16\pi} \sqrt{1 - \frac{4m^2}{s}} ,$$ where $q^{\star}$ is the relative momentum of the two particles in the CMF, $q^{\star} \equiv \sqrt{s/4 - m^2}$. This square root introduces a branch cut in the complex $s$ plane, illustrated in Fig. \[fig:diag\_bs\_complex\_plane\]. Bound states are then defined as subthreshold poles on the first Riemann sheet, the sheet for which $\text{Im} \, q^\star > 0$.[^5] The K matrix, ${\mathcal{K}}(s)$, is a real function describing all of the dynamics of the system. It can be written in terms of the scattering phase shift, $\delta(s)$, via $$\label{eq:Kmat_phase} {\mathcal{K}}^{\,-1}(s) \equiv \rho(s) \cot\delta(s).$$ Unlike $\mathcal M(s)$, the K matrix is an analytic function of $s$ in a domain around $s = (2m)^2$ set by the nearest left-hand cut. It follows that the effective range expansion $$\label{eq:ERE} q^{\star} \cot\delta(s) = -\frac{1}{a} + \frac{1}{2} r q^{\star 2} + {\mathcal{O}}(q^{\star 4}) \,,$$ has a finite radius of convergence, and gives a useful description of $\mathcal K(s)$ and $\mathcal M(s)$ near threshold. The parameters $a$ and $r$ are called the scattering length and effective range, respectively. Using Eqs. -, the condition for a bound state (a real sub-threshold pole on the first Riemann sheet) can be expressed as $$\label{eq:BS_pole} q^{\star} \, \cot\delta(q^{\star}) \big\rvert_{q^{\star} = i\kappa_{\textrm{B}}} + \kappa_{\textrm{B}} = 0 \,,$$ where $\kappa_{\textrm{B}}$ is the binding momentum, related to the pole position $s_{\text{B}}$ via $ s_{\textrm{B}} = 4(m^2 - \kappa_{\textrm{B}}^2)$ where the mass of the bound state is given as $M_{\textrm{B}} = \sqrt{s_{\textrm{B}}}$. Going beyond the pole, information about the nature and structure of the bound state is also contained in its coupling to two-particle scattering states, $g$, defined as the residue of the pole $$\label{eq:M_BS_pole} {\mathcal{M}}(s) = \frac{(ig)^2}{s - s_{\textrm{B}} } \big [1 + \mathcal O(s - s_{\textrm{B}}) \big ] \,.$$ As we review in Sec. \[eq:BS\_FV\_energies\], $g$ governs the prefactor of the bound state’s leading finite-volume effects. ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ amplitudes and form factors {#sec:2Jto2} ----------------------------------------------------------------------- We now turn to the less standard ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ transition amplitude, defined via $$\label{eq:W_matrix} \bra{P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}, \text{out}} {\mathcal{J}}^{\mu} \ket{P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}, \text{in}}_{\textrm{conn.}} \equiv {\mathcal{W}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \,.$$ Here the initial and final states have kinematics as in Sec. \[sec:2to2\] and the current, ${\mathcal{J}}^{\mu}$, is a local operator evaluated in position space at the origin. Since the current can inject energy and momentum, the initial and final states carry different total four-momenta, $P_i$ and $P_f$ respectively. It is also convenient to define the squared momentum transfer, $Q^2 \equiv - (P_f - P_i)^2$, where the overall minus is included so that $Q^2 > 0$ for spacelike $P_f - P_i$. The amplitude ${\mathcal{W}}^\mu$ can be defined for local currents with any Lorentz structure and in Sec. \[sec:bs\_scalarc\] we also consider specific results for a scalar current. Here, for concreteness we focus on a conserved vector current ${\mathcal{J}}^{\mu}(x)$ satisfying $$\partial_\mu {\mathcal{J}}^{\mu}(x) = 0 \,.$$ Our first aim is to connect this amplitude to the bound-state form factor, defined via $$\label{eq:FBdef} \bra{P_f , {\text{B}}} {\mathcal{J}}^\mu \ket{P_i , {\text{B}}} = (P_f + P_i)^\mu F_{\text{B}}(Q^2) \,,$$ where $\ket{P_i , {\text{B}}}$ is the bound state, normalized as $\braket{P_f , {\text{B}} \| P_i , {\text{B}}} =(2 \pi)^3 \, 2 \omega_{{\mathbf{P}}_i} \delta^3(\textbf P_f - \textbf P_i)$ with energy $\omega_{{\mathbf{P}}} = \sqrt{ s_{\text{B}} + \textbf P^2 }$. Eq. is related to the $S$-wave projection of ${\mathcal{W}}^\mu$, analytically continued below threshold to the bound-state pole: $$\label{eq:W_BS} {\mathcal{W}}^{\mu}(P_f,P_i) = (P_i + P_f)^{\mu} \, F_{\textrm{B}}(Q^2) \frac{i^2 (ig)^2}{( s_f - s_{\textrm{B} }) ( s_i - s_{\textrm{B} }) } \big [1 + \mathcal O(s_{i,f} - s_{\text{B}}) \big ] \,,$$ where $s_{i,f} = P_{i,f}^2$. We prove this result in Appendix \[app:Wpole\]. As discussed in some detail in Refs. [@Briceno:2015tza; @Baroni:2018iau], the analytic structure of $\mathcal W^\mu$ is significantly more complicated than that of $\mathcal M$. One can identify three generic sources of non-analyticity in the ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitude: (i) Exactly as for $\mathcal M$, diagrams with on-shell two-particle intermediate states lead to factors of $\rho(s)$. $\mathcal W^\mu$ thus exhibits the same branch cut and multi-sheet structure as $\mathcal M$. (ii) Isolated poles arise due to the subtracted diagrams in Fig. \[fig:2Jto2\_W\_amp\](b), in which the current is attached to an external leg. (iii) The triangle diagram shown in Fig. \[fig:2Jto2\_W\_amp\](c) induces a new class of singularities, first described by Landau in Ref. [@Landau:1959fi]. To relate the ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitude to a physical scattering rate, it is important to recall that the former arises from perturbatively expanding the weakly-interacting sector (encoded in $\mathcal J^\mu$) while keeping the strong dynamics non-perturbative. In fact, both the triangle singularities and the isolated poles can be understood as artifacts, resulting from truncating the expansion at a fixed order. Of course, the more standard single-particle form factors also arise from such an expansion, but happen to exhibit more straightforward analytic structure in the kinematic region considered. Indeed, as we will see below, analytic continuation to the bound-state pole removes all three of the non-analyticities we have identified. As first explained in Ref. [@Briceno:2015tza], finite-volume matrix elements are more directly related to a subtracted amplitude, denoted ${\mathcal{W}}_{{\textrm{df}}}$, from which the isolated poles, item (ii) above, have been removed. Here the subscript ‘df’ stands for ‘divergence free’. The definition, depicted also in Fig. \[fig:2Jto2\_W\_amp\](b), reads $$\begin{gathered} \label{eq:Wdf_def} {\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \equiv {\mathcal{W}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \\ - i \overline{{\mathcal{M}}}(P_f,k',k) \, \frac{i}{(P_f - k)^2 - m^2} \, w^{\mu}(P_f, P_i , k) - w^{\mu} (P_f, P_i , k') \, \frac{i}{(P_i - k')^2 - m^2} \, i\overline{{\mathcal{M}}}(P_i,k',k) \,,\end{gathered}$$ where $w^{\mu}$ is the single-particle matrix element, $$\begin{aligned} \label{eq:single_w} w^{\mu} (P_f, P_i , k) & \equiv \bra{P_f - k , \text{Q}_0 } {\mathcal{J}}^{\mu}\ket{P_i - k, \text{Q}_0} \,, \\ & = (P_f+P_i - 2 k )^{\mu} \, f\big ( Q^2 \big) \,,\end{aligned}$$ and we have also introduced the corresponding form factor, $f$. Here we adopt the convention that the charged particles carry momenta $P_i - k$ (incoming) and $P_f - k'$ (outgoing).[^6] The overline in $\overline {\mathcal M}$ denotes a slight modification to the definition of ${\mathcal{M}}$ to account for the off-shell leg. This is described in Ref. [@Baroni:2018iau] and, since the distinction is irrelevant for the $S$-wave amplitude, we do not discuss the issue further in this work. The momentum directions within ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ can be projected to definite angular momentum as done in Eq. for $\mathcal M$ $$\label{eq:Wdf_pwe} {\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) = 4\pi \sum_{\substack{ \ell_f,m_f \\ \ell_i,m_i} } Y_{\ell_f m_f}({\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}) \, {\mathcal{W}}_{{\textrm{df}}, \ell_f m_f, \ell_i m_i}^{\mu}(P_f,P_i) \,Y_{\ell_i m_i}^{}({\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \,.$$ As with the scattering amplitude, for the remainder of this section we restrict attention to the $S$-wave component of ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$, i.e. the component containing our bound state. Note that, in contrast to $\mathcal M$, $ {\mathcal{W}}_{{\textrm{df}}}$ has off diagonal elements in angular momentum space, due to the angular momentum injected by the external current. A crucial observation that will guide our later analysis is that the $S$-wave component of ${\mathcal{W}}_{{\textrm{df}}}$ exactly satisfies Eq. above, i.e. ${\mathcal{W}}^{\mu}$ and ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ have the same bound-state double-pole with the same residue: $$\label{eq:Wdf_BS} {\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_f,P_i) = (P_i + P_f)^{\mu} \, F_{\textrm{B}}(Q^2) \frac{i^2 (ig)^2}{( s_f - s_{\textrm{B} }) ( s_i - s_{\textrm{B} }) } \big [1 + \mathcal O(s_{i,f} - s_{\text{B}}) \big ] \,.$$ This equivalence holds because the subtracted terms in Eq. only have a single pole, and thus cannot modify the leading divergence. The bound-state poles within ${\mathcal{W}}_{{\textrm{df}}}^\mu$ motivate us to introduce a new object, ${\mathcal{F}}^{\mu}(P_f,P_i)$, given by $$\label{eq:Wdf_on} {\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_f,P_i) = {\mathcal{M}}(s_f) \, {\mathcal{F}}^{\mu}(P_f,P_i) \, {\mathcal{M}}(s_i) \,.$$ The $S$-wave scattering amplitudes on each side remove the poles from ${\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_f,P_i)$, implying $$\begin{aligned} \label{eq:FF_K} \lim_{s_i, s_f \to s_{\text{B}}} {\mathcal{F}}^{\mu}(P_f,P_i) & = (P_f + P_i)^\mu \frac{F_{\textrm{B}}(Q^2)}{g^2} \,.\end{aligned}$$ In addition, Eq. factorizes the $\rho(s)$ branch cuts from ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ so that $\mathcal F^\mu$ does not contain this class of singularities.[^7] Nonetheless, ${\mathcal{F}}^{\mu}$ is in general complex, due to the triangle diagram of Fig. \[fig:2Jto2\_W\_amp\](c). This diagram can only contribute complexity (as well as non-analyticity) when at least one of the two-particle cuts goes on shell, i.e. when such an intermediate state can physically propagate. In particular, for subthreshold energies, and thus for some domain around the bound-state energy, the triangle integral is real and analytic. At this stage we have argued that each of the three non-analyticities listed above is irrelevant near the bound-state pole: First, the external-leg poles are removed in the conversion from ${\mathcal{W}}$ to ${\mathcal{W}}_{{\textrm{df}}}$, second, the on-shell threshold cuts in $\rho(s)$ are removed in the relation between ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ and $ {\mathcal{F}}^{\mu}$, and finally, the triangle singularity (still contained in ${\mathcal{F}}^\mu$) is avoided by the subthreshold continuation. We close this section with an important property of ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ that follows from the Ward-Takahashi identity. As we sketch in Appendix \[app:WTI\] and derive in detail in Ref. [@analytic], ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ satisfies the following simple relation to the scattering amplitude $$\label{eq:WTI_W} {\mathcal{W}}_{{\textrm{df}}}^{\mu} (P) = {\mathrm Q_0} \frac{\partial}{\partial P_{\mu}} {\mathcal{M}}(s) = 2P^{\mu} {\mathrm Q_0} \frac{\partial}{\partial s}{\mathcal{M}}(s) \,.$$ This identity is crucial for the analysis presented in Sec. \[sec:Ltoinf\]. Finite-volume formalism for two-particle systems {#sec:FV_fcns} ================================================ Before giving detailed expressions for the finite-volume effects on a two-body bound state, in this section we briefly review the general formalism describing the finite-volume energies and matrix elements of two-particle systems. In the following, we work in a cubic, periodic volume of length $L$ with infinite temporal extent. The total momentum of the system in the finite-volume frame is allowed to take on any value consistent with the periodicity: ${\mathbf{P}}= 2\pi {\mathbf{n}}/ L$ with ${\mathbf{n}}\in {\mathbbm{Z}}^{3}$. Finite-volume energies {#sec:FV_spectrum} ---------------------- In the window of energies for which only two particles can propagate, the finite-volume spectrum is related to the infinite-volume partial-wave amplitudes, defined in Eq. , via the Lüscher quantization condition [@Luscher:1991n1; @Rummukainen:1995vs; @Kim:2005gf]. Generally, the quantization condition is a determinant over angular momentum space. If we neglect waves higher than $\ell = 0$, however, it reduces to a simple algebraic relation $$\label{eq:Luscher} {\mathcal{M}}^{-1}(s_n) = -F(P_n,L) + {\mathcal{O}}(e^{- m L}) \,,$$ where $s_n = P_n^2 = E_n(L)^2 - {\mathbf{P}}^2$ corresponds to the eigenenergy of the $n$th finite-volume two-particle state. Here $F(P,L)$ is a known finite-volume function, $$\begin{aligned} \label{eq:FV_F} F(P,L) & = \bigg[ \frac{1}{L^{3}} \operatorname{ \mathchoice {\ooalign{$\displaystyle\sum$\cr\hidewidth$\displaystyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.14\height}{\scalebox{.7}{$\textstyle\sum$}}\cr\hidewidth$\textstyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} }_{{\mathbf{k}}} \bigg] \frac{1 }{2\omega_{{\mathbf{k}}} 2\omega_{{\mathbf{P}}{\mathbf{k}}} (E - \omega_{{\mathbf{k}}} - \omega_{{\mathbf{P}}{\mathbf{k}}} + i\epsilon)} \,, $$ where $\omega_{{\mathbf{k}}} = \sqrt{m^2 + {\mathbf{k}}^2}$ and $\omega_{{\mathbf{P}}{\mathbf{k}}} = \sqrt{m^2 + ({\mathbf{P}}-{\mathbf{k}})^2}$ are the on-shell energies of the two particles, and $$\bigg[ \frac{1}{L^{3}} \operatorname{ \mathchoice {\ooalign{$\displaystyle\sum$\cr\hidewidth$\displaystyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.14\height}{\scalebox{.7}{$\textstyle\sum$}}\cr\hidewidth$\textstyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} }_{{\mathbf{k}}} \bigg] \equiv \frac{1}{L^{3}} \sum_{{\mathbf{k}}\in (2\pi/L) {\mathbbm{Z}}^{3} } - \int \frac{ {\textrm{d}}^{3}{\mathbf{k}}}{(2\pi)^{3} }.$$ Equation holds up to corrections associated with higher partial waves and only for $s_n$ below the first inelastic threshold. Finite-volume matrix elements {#sec:FV_matrix} ----------------------------- Similarly, one can relate finite-volume matrix elements of two-particle systems to infinite-volume ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ transition amplitudes. Here, the relevant formalism was first derived in Ref. [@Briceno:2015tza] using an all-orders perturbative expansion based in a generic relativistic effective field theory. Recently, in Ref. [@Baroni:2018iau], the formal approach was improved in two ways: First, by rearranging the separation of finite-volume effects, we were able to show that the extracted infinite-volume transition amplitudes are manifestly Lorentz covariant. Second, we re-organized the analysis so that single-particle matrix elements enter via standard form factors (rather than a non-standard spherical harmonic decomposition used in the first publication). While the two representations are formally equivalent, the work of Ref. [@Baroni:2018iau] is expected to be significantly more convenient in numerical applications going forward. Of course, all expressions used here are taken from the improved approach. Again, assuming all but the $\ell=0$ partial waves are negligible, the matrix elements of the vector current for two-particle states can be related to ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$, defined in Eq. , as follows $$\label{eq:BH_eqns} L^{3} \bra{P_{n,f},L} {\mathcal{J}}^{\mu} \ket{P_{n,i},L} = {\mathcal{W}}_{L,{\textrm{df}}}^{\mu}(P_{n,f},P_{n,i},L) \, \sqrt{ {\mathcal{R}}(P_{n,f},L) {\mathcal{R}}(P_{n,i},L) },$$ where $P_{n,i} = (E_{n,i},{\mathbf{P}}_i)$ and $P_{n,f} = (E_{n,f},{\mathbf{P}}_f)$. Here ${\mathcal{W}}_{L,{\textrm{df}}}^{\mu}$ is an $L$-dependent function related to ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ in a manner detailed in the following paragraph. In addition, ${\mathcal{R}}$ is a generalization of the Lellouch-Lüscher factor [@Lellouch:2000pv], first introduced in Ref. [@Briceno:2014uqa] $$\begin{aligned} \label{eq:Rdef} {\mathcal{R}}(P_n,L) & = \left[ \frac{\partial}{\partial E} \left( F^{-1}(P,L) + {\mathcal{M}}(s) \right) \right]^{-1}_{E = E_n} \,, \\ & = \label{eq:Rdef2} - {\mathcal{M}}^{-2}(s_n) \left[ \frac{\partial}{\partial E} \left( F(P,L) + {\mathcal{M}}^{-1}(s) \right) \right]^{-1}_{E = E_n} \,,\end{aligned}$$ where we have given a second form that will be particularly useful for this work. In general, ${\mathcal{R}}$ is a matrix over all two-particle degrees of freedom, but in the case considered it reduces to a simple derivative of the functions shown. Before defining ${\mathcal{W}}_{L,{\textrm{df}}}^{\mu}(P_f,P_i,L)$, we need to introduce a second $L$-dependent kinematic function, $G^{\mu_1\cdots\mu_n}$, first introduced in Refs. [@Briceno:2015tza; @Baroni:2018iau]. For the $\ell = 0$ truncation it takes the form $$\label{eq:FV_G} G^{\mu_1 \cdots \mu_n} (P_f,P_i,L) = \bigg[ \frac{1}{L^{3}} \operatorname{ \mathchoice {\ooalign{$\displaystyle\sum$\cr\hidewidth$\displaystyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.14\height}{\scalebox{.7}{$\textstyle\sum$}}\cr\hidewidth$\textstyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} }_{{\mathbf{k}}} \bigg] \frac{ k^{\mu_1} \cdots k^{\mu_n} }{2\omega_{{\mathbf{k}}} ((P_{f} - k)^2 - m^2 + i\epsilon) ((P_{i} - k)^2 - m^2 + i\epsilon) } \Big\rvert_{k^{0} = \omega_{{\mathbf{k}}}}.$$ In this work we will specifically need the scalar and vector $G$-functions, denoted $G$ and $G^{\mu}$ respectively. These are defined by keeping zero or one factor, respectively, of $k^\mu$ in the numerator of the integrand. With these in hand, ${\mathcal{W}}_{L,{\textrm{df}}}^{\mu}$ can be defined via its relation to ${\mathcal{W}}^\mu_{{\textrm{df}}}$ as follows: $$\begin{aligned} \label{eq:WLdf_vector} {\mathcal{W}}^{\mu}_{L,{\textrm{df}}}(P_{f},P_{i} , L) & = {\mathcal{W}}^{\mu}_{{\textrm{df}}}(P_{f}, P_{i}) + f(Q^2) {\mathcal{M}}(s_f) \Big [ (P_f + P_i)^{\mu} G(P_{f},P_{i},L) - 2 G^{\mu}(P_{f},P_{i},L) \Big ] {\mathcal{M}}(s_i) \, . \end{aligned}$$ Here $f(Q^2)$ is the form factor of the charged particle while the form factor of the neutral particle, which vanishes identically at $Q^2=0$, is assumed negligible for all values of momentum transfer. Matrix elements of the conserved vector current {#sec:WTI} =============================================== Having introduced the general formalism, we proceed to perform the checks outlined in the introduction. The first check is to show that, for any finite-volume state, the matrix element with respect to the charge operator $$\widehat{\mathrm{Q}} \equiv \int {\textrm{d}}^3 \textbf x \, \mathcal J^0(x) \,,$$ is predicted by the formal mapping to be $L$-independent and equal to the charge of the state. To demonstrate this, we first introduce another expression for the Lellouch-Lüscher factor. Evaluating the energy derivative of $F$ in Eq. , one can show $$\begin{aligned} \label{eq:R_Kmat} {\mathcal{R}}(P_n(L),L) & = \frac{1}{{\mathcal{M}}^2(s_n(L))} \left[ -\frac{\partial}{\partial E} {\mathcal{M}}^{-1}(s) + 2 E \, G(P ,L) - 2G^{\mu=0}(P ,L) \right]^{-1}_{P = P_n(L)} \,.\end{aligned}$$ Here we have also adopted the shorthand $G(P ,L) \equiv G(P, P,L)$, i.e. we do not repeat the total momentum argument when it is the same for the incoming and outgoing states. Note that, in this subsection, we are considering not only the finite-volume bound state but also excited states. We do continue to restrict attention to the $S$-wave only. Substituting this result into Eq. , and also taking the relation between ${\mathcal{W}}_{{\textrm{df}}}^\mu$ and $\mathcal F^\mu$ \[Eq. \], we find $$\begin{gathered} \label{eq:fv_matrix_complicated} L^3 \bra{P_{n,f},L} {\mathcal{J}}^{\mu} \ket{P_{n,i},L} = \\ \frac{ {\mathcal{F}}^{\mu}(P_f, P_i) + f(Q^2) \big[ (P_i + P_f)^{\mu}G(P_f,P_i,L) - 2G^{\mu}(P_f,P_i,L) \big] }{ \sqrt{\big [- {\partial}_{E_i} {\mathcal{M}}^{-1}(s_i) + 2E_i G(P_i,L) - 2G^{\mu=0}(P_i,L) \big ] \big [-{\partial}_{E_f} {\mathcal{M}}^{-1}(s_f) + 2E_f G(P_f,L) - 2G^{\mu=0}(P_f,L) \big ] }} \bigg \vert_{P_{i,f} = P_{i,f}(L)} \,.\end{gathered}$$ This result will prove very powerful in the following derivations. To see the consequences of this for the charge operator we set $\mu=0$ in the vector current and also set the initial and final-states to coincide. This yields $$\label{eq:fv_matrix_simple} \bra{P_{n},L} \widehat {\mathrm{Q}} \ket{P_{n},L} = \\ \frac{ {\mathcal{F}}^{0} (P) + f(0) \big[ 2 E G(P ,L) - 2G^{0} (P ,L) \big] }{ - {\partial}_{E } {\mathcal{M}}^{-1}(s ) + 2E G(P ,L) - 2G^{0} (P ,L) } \bigg \vert_{P = P_n(L)} \,,$$ where we have used the $\textbf x$-independence of the matrix element to replace $L^3 \mathcal J^0\!(0) \to \widehat {\mathrm{Q}} $ and have defined ${\mathcal{F}}^{0}(P) \equiv {\mathcal{F}}^{0}(P,P)$ as a convenient shorthand for systems with identical initial and final momenta. This can be further simplified via the identity $$\label{eq:F_constraint} {\mathcal{F}}^{0}(P) =\frac{ {\mathrm Q_0} }{ {\mathcal{M}}^{2}(s)} \frac{\partial}{\partial E} {\mathcal{M}}(s) = - {\mathrm Q_0} \frac{\partial}{\partial E} {\mathcal{M}}^{-1}(s) \,,$$ which immediately follows from Eqs. and . Substituting this into the numerator of Eq. and also using $f(0) = \mathrm Q_0$, we recover a very satisfying cancellation of all terms to deduce $$\bra{P_{n},L} \widehat {\mathrm{Q}} \ket{P_{n},L} = \mathrm Q_0 \,,$$ as expected. This is a highly non-trivial verification that the general ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ finite-volume formalism is consistent the consequences of current conservation. The derivation relies on two unexpected identities: First, the fact that the energy-derivative of $F(P,L)$ can be expressed using the $G$-functions, as shown in Eq. , and second, that the Ward-Takhashi identity relates $\mathcal F^{0}(P)$ to the scattering amplitude, Eq. . Bound state in a finite volume {#sec:BS_FV} ============================== We now turn to the implications of the general formalism for bound-state matrix elements in a finite volume. Volume effects on the energies {#eq:BS_FV_energies} ------------------------------ We start by reviewing results for finite-volume effects in the energy level, $E^{\textbf P}_{\textrm{B}}(L)$, defined to coincide with the moving bound state in the infinite-volume limit, $$\lim_{L \to \infty} E^{\textbf P}_{\textrm{B}}(L) = E^{\textbf P}_{\textrm{B}} \equiv \sqrt{ M_{\textrm{B}}^2 + {\mathbf{P}}^2 } \,.$$ Boosting these energies to the rest frame, we also define $$s^{\textbf P}_{\textrm{B}}(L) \equiv E^{\textbf P}_{\textrm{B}}(L)^2 - \textbf P^2 \equiv s_{\textrm{B}} + \delta s^{\textbf P}_{\textrm{B}}(L) \,,$$ with $s_{\textrm{B}} = \lim_{L \to \infty} s^{\textbf P}_{\textrm{B}}(L) = M_{\textrm{B}}^2$. Note that the finite-volume energies depend on $\textbf P$, even after boosting back to the rest frame. In the following, we give expressions for the volume-induced shift, $\delta s^{\textbf P}_{\textrm{B}}(L)$, for two values of total momentum. This represents a small subset of the more general expressions derived in Ref. [@Davoudi:2011md].[^8] The quantization condition, Eq. , is satisfied only at the finite-volume energies, e.g. at $P_{\textrm{B}}(L)\equiv( E^{\textbf P}_{\textrm{B}}(L), \textbf P )$. We are thus strictly interested in $F(P,L)$ only when it is evaluated at these points. However, taking $\delta s_{\textrm{B}}$ as a small parameter, we note $$F (P_{\textrm{B}}(L), L) = F (P_{\textrm{B}} , L) + {\mathcal{O}}(\delta s_{\textrm{B}}) \,,$$ where $P_{\textrm{B}} \equiv ( E^{\textbf{P}}_{\textrm{B}}, \textbf P )$ is the infinite-volume bound-state momentum in a moving frame. As is discussed in detail in Ref. [@Davoudi:2011md] and reviewed in Appendix \[app:cont\], the subthreshold $L$-dependence of the $F$-function is governed by the binding momentum: $ \kappa_{\text{B}}^2 \equiv m^2 - M_B^2/4$. In particular from Eqs. and we find $$\label{eq:FV_F_kappa} F(P_{\textrm{B}},L) = -\frac{1}{8\pi M_{\textrm{B}} } \sum_{{\mathbf{m}}\ne{\mathbf{0}}} \, e^{iL {\mathbf{m}}\cdot {\mathbf{P}}/ 2} \, \frac{ e^{- \kappa_{\text{B}} L \lvert{\mathbf{m}}'\rvert}}{ L \lvert{\mathbf{m}}'\rvert},$$ where $${\mathbf{m}}' \equiv {\mathbf{m}}+ (\gamma - 1) \frac{{\mathbf{m}}\cdot {\mathbf{P}}}{\lvert {\mathbf{P}}\rvert^2} {\mathbf{P}}\,, \label{eq:mp}$$ and $\gamma = E^{\textbf{P}}_{\textrm{B}}/M_{\textrm{B}}$. This result is to be combined with the inverse scattering amplitude, also evaluated at $s^{\textbf P}_{\textrm{B}}(L)$, but then expanded in powers of $\delta s_{\textrm{B}}$ to yield $$\begin{aligned} {\mathcal{M}}^{-1}(s^{\textbf P}_{\textrm{B}}(L)) &= \delta s_{\textrm{B}}\frac{\partial}{\partial s} {\mathcal{M}}^{-1}(s) \Big\rvert_{s = s_{\textrm{B}}} + \mathcal O(\delta s_{\textrm{B}}^2) \,, \\[3pt] &= - \delta s_{\textrm{B}} /g^2 + \mathcal O(\delta s_{\textrm{B}}^2) \,, \label{eq:Mexp}\end{aligned}$$ where we have used ${\mathcal{M}}^{-1}(s_{\textrm{B}}) = 0$. Combining Eqs. and then yields the elegant result $$\begin{aligned} \label{eq:delta_sB} \delta s^{\textbf P}_{\textrm{B}}(L) &= g^2 F(P_{\textrm{B}},L) + \mathcal O(e^{- 2 \kappa_{\text{B}} L})\, ,\end{aligned}$$ which shows that the leading shift to the finite-volume bound state is given directly by the $F$-function, evaluated at the infinite-volume bound-state energy. To close this section we think it useful to unpack Eq. for a two specific cases. First, in the case of vanishing momentum in the finite-volume frame, the three universal orders are given by $$\delta s^{[000]}_{\textrm{B}} = -\frac{6 g^2}{8\pi M_{\textrm{B}} L } \bigg [ e^{-\kappa_{\textrm{B}} L } + \sqrt{2} e^{- \sqrt{2} \kappa_{\textrm{B}} L } + \frac{4}{3 \sqrt{3}} e^{- \sqrt{3} \kappa_{\textrm{B}} L } \bigg ] + \mathcal O(e^{- 2 \kappa_{\text{B}} L}) \,.$$ At $\mathcal O(e^{- 2 \kappa_{\text{B}} L})$ higher derivatives of the inverse amplitude enter, requiring information beyond the coupling, $g$. For nonzero momenta, the expressions are complicated by the relation between $\textbf m'$ and $\textbf m$, and by the volume dependence entering $\gamma$ through $E^{\textbf{P}}_B = \sqrt{ M_{\text{B}}^2 + (2 \pi/L)^2 {\mathbf{n}}^2 }$. Useful results can be reached, however, by expanding in all $L$ dependence. Performing such an expansion, and neglecting terms scaling as $ e^{-\kappa_{\textrm{B}} L }/L^2$ and $ e^{- \sqrt{2} \kappa_{\textrm{B}} L }$, we find $$\begin{aligned} \delta s^{[00n]}_{\textrm{B}} & = -\frac{ g^2 [4 + 2 \cos(n \pi)] e^{-\kappa_{\textrm{B}} L }}{8\pi M_{\textrm{B}} L } \bigg [1 - \frac{ n \cos ( n \pi)}{4 + 2 \cos ( n \pi)} \frac{4 \pi ^2 \kappa_{\textrm{B}}}{ M_{\textrm{B}}^2 L} \bigg ] \,.\end{aligned}$$ To compare these results to those in Ref. [@Davoudi:2011md] we note that, in the earlier work, the authors introduce an $L$-dependent binding momentum, defined via $\kappa_{\text{B}}(L)^2 = m^2 - s_{\text{B}}(L)/4$. Then the finite-volume shift, $\delta \kappa_{\text{B}}(L) \equiv \kappa_{\text{B}}(L) - \kappa_{\text{B}}$ satisfies the relation $$\delta s^{\textbf P}_{\textrm{B}}(L) = -8 \kappa_{\textrm{B}} \,\delta\kappa^{\textbf P}_{\textrm{B}}(L) + \mathcal O(\delta s_{\textrm{B}}^2 )\,.$$ Combining this with the relation between the coupling and the scattering phase $$\frac{1}{g^2} = \frac{ 1 }{64 \pi \kappa_{\text{B}} M_{\text{B}}} \bigg (1 - 2 \kappa_{\text{B}} \frac{{\textrm{d}}}{{\textrm{d}}q^{\star\,2}} q^{\star} \cot \delta(q^{\star} ) \bigg )_{\!\! s = s_{\text{B}}} \,,$$ yields Eq. (9) of Ref. [@Davoudi:2011md]. In closing we comment that, due to the reduction of rotational symmetry, higher partial waves do induce finite-volume corrections to the scalar bound state and corresponding matrix elements. In particular, for ${\mathbf{P}}= 0$, $\ell=0$ mixes with $\ell =4, 6, \ldots\,$, as can be seen by the fact that the corresponding off-diagonal components of $F$ are nonzero. These additional angular-momentum contributions are, in fact, not volume-suppressed relative to the $S$-wave contributions, but are suppressed by powers of the binding momentum in units of the scattering-length analogs appearing in higher-partial waves. For example the $\ell=4$ phase shift satisfies an expansion analogous to Eq. $$q^{\star} \cot\delta_{\ell=4}(s) = \frac{M_4^9}{q^{\star 8}} + \mathcal O(q^{\star -6}) \,,$$ where $M_4$ has units of energy. In the case of zero spatial momentum in the finite-volume frame, one can show that the first non $S$-wave contribution to $s_{\text{B}}^{[000]}(L)$ is suppressed relative to the leading shift by a factor of $ \kappa_{\text{B}}^{8} \, M_{\text{B}} \, / \, M_4^9$. Having reproduced the known expansion for the binding energy [@Davoudi:2011md], we now turn to the finite-$L$ corrections of the bound-state matrix element. Volume effects on the matrix elements\[sec:FVME\] ------------------------------------------------- ### Matrix elements in the $L\to\infty$ limit[\[sec:Ltoinf\]]{} We begin by confirming that, in the $ L \to\infty$ limit, the finite-volume bound-state matrix element (as described by the general ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ formalism) coincides with its infinite-volume counterpart. Here it is important to stress that the various quantities we consider have a well-defined $L \to \infty$ limit, only because we are considering them at sub-threshold kinematics and thus away from a set of finite-volume poles that becomes arbitrarily dense. We begin with Eq. , the relation between ${\mathcal{W}}_{L,{\textrm{df}}}^{\mu}$ and ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$. For $L \to \infty$, these two quantities coincide because the $G$ function defining their difference vanishes. This is the case because the sum within $G$ ([cf.]{} Eq. ) is transformed to an integral in the limit and is exactly canceled by the second, subtracted integral. Equation thus becomes $$\label{eq:BH_eqns_Linf} \lim_{L \to \infty} L^3 \bra{P_{\text{B},f},L} {\mathcal{J}}^{\mu} \ket{P_{\text{B},i},L} = \lim_{L \to \infty} {\mathcal{W}}_{{\textrm{df}}}^{\mu}(P_{\text{B},f}(L),P_{\text{B},i}(L) ) \, \sqrt{ {\mathcal{R}}(P_{\text{B},f}(L),L) \, {\mathcal{R}}(P_{\text{B},i}(L),L) } \,.$$ The next step is to expand $\mathcal R$, evaluated at the finite-volume bound-state energy, about large $L$. Using the form given by Eq. (\[eq:Rdef2\]), one readily finds $$\begin{aligned} {\mathcal{R}}(P_{\text{B}}(L),L) & = - {\mathcal{M}}^{-2}(s_{\textrm{B}}^{{\mathbf{P}}}(L)) \bigg [ \frac{2 E_{\text{B}}}{g^2} + {\mathcal{O}}(e^{- \kappa_{\text{B}} L})\bigg ]^{-1} \,, \\ & = - \frac{ \big ( s_{\textrm{B}}^{{\mathbf{P}}}(L) - s_{\textrm{B}} \big )^{2} }{2E_{\textrm{B}} g^2} \Big [ 1 + \mathcal O(e^{- \kappa_{\text{B}} L})\Big ] \,. \label{eq:Rexp}\end{aligned}$$ We are now in position to evaluate the limit. The only subtlety is that a double-zero, arising from the Lellouch-Lüscher factors, is exactly canceled by the double pole in ${\mathcal{W}}_{{\textrm{df}}}^\mu$. Substituting Eqs. and into Eq. , we reach $$\label{eq:BS_deep_matrix} \lim_{L \to \infty} 2 \sqrt{E_{\text{B},i} E_{\text{B},f} } L^3 \bra{P_{\text{B},f},L} {\mathcal{J}}^{\mu} \ket{P_{\text{B},i},L} = \bra{P_f , {\text{B}}} {\mathcal{J}}^\mu \ket{P_i , {\text{B}}} = (P_{\text{B},i} + P_{\text{B},f})^{\mu} F_{\textrm{B}}(Q^2) \,.$$ This is exactly the desired result, with the extra factors on the left-hand side accounting for the different normalization conventions of finite- and infinite-volume states. In this derivation we did not make reference to the Lorentz structure of the current, only to the fact that the ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitude, $\mathcal W$, must have a double pole structure associated with the initial and final bound states. As a result, in general the formalism fulfills the expectation that for an arbitrary current ${\mathcal{J}}_{\mu_1\ldots \mu_n}$ $$\label{eq:BS_deep_matrix_anycurrent} \lim_{L \to \infty} 2 \sqrt{E_{\text{B},i} E_{\text{B},f} } L^3 \bra{P_{\text{B},f},L} {\mathcal{J}}^{\mu_1\ldots \mu_n} \ket{P_{\text{B},i},L} = \bra{P_f , {\text{B}}} {\mathcal{J}}^{\mu_1\ldots \mu_n} \ket{P_i , {\text{B}}} \,.$$ ### Large $L$ expansion of the bound-state matrix element {#sec:bs_scalarc} As shown in Sec. \[sec:WTI\], the conserved vector current leads to volume-independent matrix elements at zero momentum transfer. Thus, to reach an interesting large-$L$ expansion, in this section we turn to a scalar current $\mathcal J$ and define $$\begin{aligned} \label{eq:BH_scalar} g^{{\mathbf{P}}}_{S,\text{B}}(L) & \equiv 2 E_B(L) L^3 \bra{P_{\text{B}},L} {\mathcal{J}}\ket{P_{\text{B}},L} \,,\end{aligned}$$ where the subscript indicates that this matrix element defines the scalar charge of the bound state. The infinite-volume bound-state scalar charge is recovered in the $L\to \infty$ limit, [i.e.]{}$g_{S,\text{B}} \equiv \lim_{L \to \infty} g^{\mathbf{P}}_{S,\text{B}}(L)$. In direct analogy to Eq. above, we observe $$\label{eq:fv_matrix_simple_scalar} g^{{\mathbf{P}}}_{S,\text{B}}(L) = \frac{ \mathcal F(s) + g_{S} G(P,L) }{ - \partial_s {\mathcal{M}}^{-1}(s) + G(P,L) - G^{ 0} (P,L)/E } \Big\rvert_{P = P_{\text{B}}(L)} \,.$$ Here ${\mathcal{F}}(s) \equiv {\mathcal{M}}^{-2}(s) {\mathcal{W}}_{{\textrm{df}}}(P,P)$ with ${\mathcal{W}}_{{\textrm{df}}}$ given by Eq. , in which the vector current is replaced by a scalar. Note that the numerator includes only the scalar $G$-function, reflecting the scalar current considered. However, the denominator remains identical to the vector case since the Lellouch-Lüscher factors are independent of the current. We have also introduced $g_{S}$ as the scalar charge of the single-particle state, $g_S \equiv f(0)$, where $f$ is the single particle form-factor $$\begin{aligned} f(Q^2) & \equiv \bra{ P_f, g_S } {\mathcal{J}}\ket{ P_i, g_S } \,.\end{aligned}$$ As above, we take the coupling of the current to the other constituent particle to be negligible. With these ingredients in hand it is straightforward to expand Eq. about $L \to \infty$ to reach $$\label{eq:BH_scalar} \frac{ g^{{\mathbf{P}}}_{S,\text{B}} (L)} { g_{S,\text{B}}} = 1 + \delta s^{{\mathbf{P}}}_{\text{B}}(L) \frac{\partial}{\partial s} \bigg [ \frac{F_{\text{B}}(s)}{g_{S,\text{B}} } + g^2 \frac{\partial}{\partial s} {\mathcal{M}}^{-1}(s) \bigg ]+ \frac{g^2 ( g_{S} - g_{S,\text{B}} )}{g_{S,\text{B}}} G(P_{\text{B}},L) + \frac{ g^2 G^{ 0} (P_{\text{B}},L)}{ E_{\text{B}}} + \mathcal O(e^{- \sqrt{2} \kappa_{\text{B}} L}) \,.$$ We note that a great deal of structural information enters the leading finite-volume correction. The $\delta s_{\text{B}}(L)$-dependent term is the correction induced from the energy shift and is thus proportional to energy derivatives of both the inverse amplitude and the ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ transition amplitude (entering via $\mathcal F(s)$). The second term in Eq. arises due to a mismatch between the scalar charge of the bound-state and the summed charges of its constituents. The final term in Eq. is a direct consequence of the triangle diagram, Fig. \[fig:2Jto2\_W\_amp\](c). We close this section with a final, more explicit result for the leading-volume correction in the case where the CMF and finite-volume frames coincide, i.e. $\textbf P = {\mathbf{0}}$. Substituting the leading result for $\delta s_{\text{B}}^{[000]}$, and results from Appendix \[app:cont\] for the $G$-functions, one finds $$\label{eq:BH_scalarv2} \frac{ g^{[000]}_{S,\text{B}} (L)} { g_{S,\text{B}}} = 1 + g^2 \frac{3 e^{-\kappa_{\textrm{B}} L } }{32 \pi M_{\text{B}} \kappa_{\text{B}} } \bigg [1+ 2 \frac{ g_{S} - g_{S,\text{B}} }{g_{S,\text{B}}} + \frac{4 \kappa_{\text{B}}}{ M_{\text{B}}^2 L} - \frac{ 8 \kappa_{\text{B}}}{ L } \frac{\partial}{\partial s} \bigg ( \frac{F_{\text{B}}(s)}{g_{S,\text{B}} } + g^2 \frac{\partial}{\partial s} {\mathcal{M}}^{-1}(s) \bigg )_{\! \! s = s_{\text{B}}} \bigg ] + \mathcal O(e^{- \sqrt{2} \kappa_{\text{B}} L}) \,.$$ The leading $1$ in the square brackets arrises from the triangle diagram, Fig. \[fig:2Jto2\_W\_amp\](c), and will be the dominant finite-volume effect provided $\|g_S-g_{S,B}\|\leq \|g_{S,B}\|/2$. Numerical expectations for finite-volume dependence {#sec:num_expectation} --------------------------------------------------- \[fig:fig-energies\] (-490,80) \[fig:fig-matrix\] (-490,80) In this section, we use the full ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ formalism to explore the finite-volume corrections to a bound-state matrix element in an example with scattering parameters chosen to mimic the deuteron. As above we consider the simplest case of a scalar current and an $S$-wave bound state and examine the finite-volume corrections given by Eqs. and . For the ${\mathbf{2}}\to {\mathbf{2}}$ scattering amplitude, we use use the phenomenological values for the $pn$ scattering length ($a = 5.425$ fm) and effective range ($r = 1.749$ fm) to describe the scattering amplitude and spectrum. With the nucleon mass at $m = 934$ [${\mathrm{\,Me\kern -0.1em V}}$]{}[^9], the deuteron bound-state pole lies at $\sqrt{s_{\textrm{B}}} = 1875.63$ [${\mathrm{\,Me\kern -0.1em V}}$]{}with a coupling of $g = 5370.7$ [${\mathrm{\,Me\kern -0.1em V}}$]{}. This corresponds to a binding momentum of $\kappa_{\textrm{B}} = 45.58$ [${\mathrm{\,Me\kern -0.1em V}}$]{}, with a binding energy $-2.21$ [${\mathrm{\,Me\kern -0.1em V}}$]{}. The effective range expansion is shown in Fig. \[fig:fig-qcot\_deut\] as a function of $q^{\star\,2}$ and the location of the bound state is indicated. We make two assumptions to simplify the numerical exercise: First, we assume that the infinite-volume bound-state form factor is constant, [i.e.]{}$F_{\textrm{B}}(s) = g_{S,\textrm{B}}$. As seen in the fourth term of the brackets in Eq. , this contribution is suppressed by $1/L$, thus it is reasonable that this approximation will not strongly alter the prediction. Second, we assume that difference $g_{S} - g_{S,\textrm{B}}$ is numerically small and set $g_{S} = g_{S,\textrm{B}}$. Within this set-up one can numerically evaluate Eqs. and and compare the results. The first step is to determine the finite-volume bound-state energy, using the effective-range description of the $pn$ scattering amplitude in the quantization condition, Eq. . Figure \[fig:fig-results\](a) shows the bound-state energy as a function of $L$ for both $\lvert{\mathbf{d}}\rvert = 0$ and 1. The solid lines represents the full solution obtained from Eqs. and with the $pn$-scattering parameters. The dashed lines correspond to the leading-order approximation using Eq. for the same momenta. These results reproduce those of Ref. [@Davoudi:2011md], and we see that for lattice calculations performed at $m_{\pi}L \sim 4$, deviations between the exact and approximated forms are significant. Turning to the two-particle matrix elements, Fig. \[fig:fig-results\](b) shows the ratio of the finite-volume bound-state matrix element to the infinite-volume scalar charge, $g_{S,\textrm{B}}^{{\mathbf{P}}}(L) \, / \, g_{S,\textrm{B}}$. Solid lines represent the full solution, using Eq. evaluated at the finite-volume energy, and the dashed lines give the leading-order shift of Eq. . Again, significant deviations arise between the full prediction, the leading-order expansion, and the infinite-volume result. This illustrates that, to reliably extract infinite-volume matrix elements of shallow bound states like the deuteron, it is highly beneficial to use the full formalism which removes an infinite series of terms scaling as powers of $e^{- \kappa_{\text{B}}L}$. In the present example, only at $m_{\pi }L \sim 8$ do corrections scale to the percent level. Conclusion {#sec:conclusion} ========== In this work, we have provided strong consistency checks on, and also explored various consequences of, the formalism derived in Refs. [@Briceno:2015tza; @Baroni:2018iau], which gives a relation between finite-volume matrix elements, schematically denoted $\langle 2 \vert {\mathcal{J}}\vert 2 \rangle_L$, and the corresponding infinite-volume ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ amplitudes. First, in the case of the conserved vector current, we have shown that resulting prediction for the two-particle matrix element of the charge operator, $\langle 2 \vert \widehat {\textrm{Q}} \vert 2 \rangle_L$, behaves as expected. Specifically, the matrix element is $L$-independent and equal to the sum of the constituent charges. Though it is clear that this relation must hold, the way it arises in the mapping is highly non-trivial, relying on an identity relating various $L$-dependent geometric functions \[Eq. \] as well as a relation between the ${\mathbf{2}}\to{\mathbf{2}}$ and ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ amplitudes that follows from the Ward-Takahashi identity \[Eq. \]. Second, for a generic local current, we have demonstrated that the mapping of Refs. [@Briceno:2015tza; @Baroni:2018iau], reproduces the expected behavior in the case of an $S$-wave two-particle bound state. By analytically continuing the formal relations below threshold to the bound-state pole, we have confirmed that the finite- and infinite-volume matrix elements are equal up to volume corrections scaling as $e^{- \kappa_{\text{B}} L}$, where $\kappa_{\text{B}}$ is the binding-momentum of the state. This is an expected extension of the well-known result for the $L$ dependence of the bound-state energy. These two checks give confidence that our admittedly complicated formalism correctly describes two-particle finite-volume states and is ready to be implemented in a LQCD calculation, with the first application likely being the $(\pi \pi)_{\text{I=1}} + {\mathcal{J}}_\mu \to (\pi \pi)_{\text{I=1}}$ transition amplitude, allowing one to extract the electromagnetic form factors of the $\rho$. As an additional example of the utility of the general approach, we have determined the full functional form of the leading, $\mathcal O(e^{- \kappa_{\text{B}} L})$ correction to the bound-state matrix element of a local scalar current. The result, Eq. , shows that the coefficient of the leading exponential depends on the bound state’s coupling to the two-particle asymptotic state, the scalar charges of both the bound state and its constituents, and also on derivatives of both the ${\mathbf{2}}\to {\mathbf{2}}$ scattering amplitude and the bound-state form factor. While the structure of this relatively simple prediction is instructive, we stress that in practice it is more useful to use the general relation of Refs. [@Briceno:2015tza; @Baroni:2018iau] to extract the ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ over a range of energies, including in a neighborhood around the bound-state pole. Doing so removes an infinite series of terms scaling as powers of $e^{- \kappa_{\text{B}} L}$ and, for shallow bound-states, allows one to control an otherwise dominant source of systematic uncertainty. To stress this point, as a final exercise, we have presented numerical comparisons of the leading $e^{- \kappa_{\text{B}} L}$ correction with the full finite-volume shift, for a toy set-up mimicking a LQCD calculation of the deuteron’s scalar charge. For physical pion masses and volumes in the range $m_{ {\pi}}L \sim 4$ to $7$, we find that the finite-$L$ correction will dominate the infinite-volume charge and that removing only the leading exponential also does not give a reliable extraction. Thus, we conclude that the full method must be used to gain a reliable result for the form factors of shallow bound states as well as resonances. This work makes use of identities that will be presented in a companion article that outlines, in detail, the analytic structure of the generalized form factors considered here [@analytic]. An additional check is underway to reproduce analytic expressions for the $1/L$ expansion presented in Ref. [@Detmold:2014fpa], for the threshold-state matrix element of a scalar current in a weakly-coupled system. Acknowledgements ================ We thank Alessandro Baroni, Felipe Ortega-Gama, and Akaki Rusetsky for useful discussions. RAB is supported in part by USDOE grant No. DE-AC05-06OR23177, under which Jefferson Science Associates, LLC, manages and operates Jefferson Lab. RAB also acknowledges support from the USDOE Early Career award, contract de-sc0019229. Proof of Eq. \[app:Wpole\] ========================== In this appendix we demonstrate that, in theories with a two-particle bound state, the ${\mathbf{2}}+ {\mathcal{J}}\to {\mathbf{2}}$ amplitude satisfies Eq. , repeated here for convenience $${\mathcal{W}}^{\mu}(P_f,P_i) = (P_i + P_f)^{\mu} \, F_{\textrm{B}}(Q^2) \frac{i^2 (ig)^2}{( s_f - s_{\textrm{B} }) ( s_i - s_{\textrm{B} }) } \big [1 + \mathcal O(s_{i,f} - s_{\text{B}}) \big ] \,.$$ To show this, it is necessary to return to the matrix element definition of the amplitude, $\mathcal W^\mu$, given in Eq. . Inserting a complete set of states on either side of the current, $\mathcal J^\mu$, we reach $$\label{eq:W_compstates} \int \! \! \frac{{\textrm{d}}^3 \textbf P''}{(2 \pi)^3 2 E_{\text{B}}^{\textbf{P}''} } \, \int \! \! \frac{{\textrm{d}}^3 \textbf P'}{(2 \pi)^3 2 E_{\text{B}}^{\textbf{P}'} } \, \langle {P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}, \text{out}} \vert {P'', \text{B}} \rangle \langle {P'', \text{B}} \vert \, {\mathcal{J}}^{\mu} \, \vert P', \text{B} \rangle \langle P', \text{B} \vert {P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}, \text{in}} \rangle = {\mathcal{W}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) + \cdots \,,$$ where we have kept only the bound-state sector of the Fock space, as this will be sufficient to identify the pole that we are after. Three additional subtleties arise here: (1) To properly implement the normalization of the bound-state, $$\langle P', \text{B} \vert P, \text{B} \rangle \equiv (2 \pi)^3 2 E_{\text{B}}^{\textbf{P}} \delta^3(\textbf{P}' - \textbf{P}) \,,$$ we must integrate over all spatial momenta with the standard Lorentz-invariant factor as shown. (2) Since the spectral decomposition can only be performed on the full matrix element we have dropped the “conn” subscript that appears in Eq. . To preserve the definition we have included the ellipsis on the right-hand side, which is understood to represent all disconnected contributions. These will, however, play no role, since they do not contain the bound-state pole. (3) The expression we are after requires the analytic continuation of $P_f$ and $P_i$ to the sub-threshold region. This is subtle at the level of Fock states, and is more easily understood by rewriting the result in terms of operators projected to definite momentum. This, in turn, reveals that the time-ordering of the operators must be carefully treated, as we explain in more detail below. The next step is to substitute $$\begin{aligned} \langle {P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}, \text{out}} \vert {P'', \text{B}} \rangle & \equiv (2 \pi)^4 \delta^4(P_f - P''_{\text{B}} ) \, i g \,, \\ \langle P', \text{B} \vert {P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}, \text{in}} \rangle & \equiv (2 \pi)^4 \delta^4(P_i - P'_{\text{B}}) \, i g \,.\end{aligned}$$ Here the four-dimensional delta function arises in direct analog to the standard relation between T matrix and scattering amplitude and leads to the definition of the bound-state coupling $g$. Using the spatial delta functions to evaluate the integrals in Eq. , we reach $$\int_{-\infty}^\infty {\textrm{d}}x_0'' \, e^{i x_0'' (E_f - E_{\text{B}}^{\textbf P''}) } \, \int_{-\infty}^\infty {\textrm{d}}x_0' \, e^{-i x_0' (E_i - E_{\text{B}}^{\textbf P'})} \, \frac{i g }{{2 E_{\text{B}}^{\textbf{P}''} } } \langle {P'', \text{B}} \vert \, {\mathcal{J}}^{\mu} \, \vert P', \text{B} \rangle \, \frac{ i g }{{2 E_{\text{B}}^{\textbf{P}'} }} = {\mathcal{W}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) + \cdots \,,$$ where it is understood that one must set $\textbf P'' \to \textbf P_f$ and $\textbf P'' \to \textbf P_i$. Here we have also written the remaining temporal delta functions as integrals over time. Introducing the integrals over $x_0''$ and $x_0'$ allows us to address the subtlety mentioned as point (3) above. Studying the correlation functions reveals that the above expression does not correctly treat all time orderings. For the present case, this is resolved by restricting the integral over $x_0''$ from $0$ to $\infty$ and similarly that over $x_0'$ from $-\infty$ to $0$. Doing so, and also including the $i \epsilon$ prescription required to project the external states in the correlator to the vacuum, one can evaluate both integrals to reach $$\frac{i(ig) }{2 E_{\text{B}}^{\textbf{P}''} (E_f - E_{\text{B}}^{\textbf P''}) } \, \frac{i (ig) }{2 E_{\text{B}}^{\textbf{P}'} (E_i - E_{\text{B}}^{\textbf P'}) } \, \langle {P'', \text{B}} \vert \, {\mathcal{J}}^{\mu} \, \vert P', \text{B} \rangle \, = {\mathcal{W}}^{\mu}(P_f,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P_i, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) + \cdots \,.$$ This is the result that we had aimed to prove. Up to the $\mathcal O((s_{i,f} - s_{\text{B}})^0 )$ terms that we neglect, one can replace each pole with the covariant form and also drop the ellipses. Projecting both sides to the $S$-wave, and substituting Eq. , we deduce Eq. . Ward-Takahashi identity for ${\mathbf{2}}+{\mathcal{J}}\to{\mathbf{2}}$ amplitudes {#app:WTI} ================================================================================== In this appendix, we demonstrate how Eq. follows from the Ward-Takahashi identity. A consequence of current conservation, the Ward-Takahashi identity relates a given $n$-point Green function, coupled to an external conserved current, to the corresponding $(n-1)$-point Green function in which the current is omitted. Let ${\mathcal{C}}^{\mu}$ be a 5-point function coupling the conserved vector current, $\mathcal J^\mu$, to two neutral and two charged mesons. The Ward-Takahashi identity then reads $$\begin{aligned} \label{eq:WTI_green} q_{\mu} {\mathcal{C}}^{\mu}(p',k';p,k) & = {\mathrm Q_0} \Big[ {\mathcal{C}}(p' + q,k' ; p,k) - {\mathcal{C}}(p',k';p-q,k) \Big] \,,\end{aligned}$$ where $q^{\mu} = (p'+k')^\mu - (p+k)^\mu = P'^{\mu} - P^{\mu}$, with the second equality introducing notation for the total momenta of the outgoing and incoming two-meson states. We have also introduced ${\mathcal{C}}$ (with no index) as the four-point function without the current insertion. We further define $k$ and $k'$ as the initial- and final-state momenta of the neutral particles, respectively, and $p = P- k$ and $p' = P' - k'$ as the corresponding momenta for the particles carrying the charge, $ {\mathrm Q_0} $. The ${\mathbf{2}}+ {\mathcal{J}}\to{\mathbf{2}}$ and ${\mathbf{2}}\to{\mathbf{2}}$ amplitudes, $\mathcal W^\mu$ and $\mathcal M$ respectively, are related to the Green functions by amputating the external meson propagators and placing them on the mass shell, [i.e.]{}, $$\begin{aligned} {\mathcal{C}}_{\textrm{amp}} & \xrightarrow[\textrm{on-shell}]{} {\mathcal{M}}\,, \\ {\mathcal{C}}_{\textrm{amp}}^{\mu} & \xrightarrow[\textrm{on-shell}]{} {\mathcal{W}}^{\mu},\end{aligned}$$ where the amputated Green functions are defined as $$\label{eq:C_amp} {\mathcal{C}}_{\textrm{amp}} (p',k';p,k) \equiv (p'^2 - m^2) (k'^2 - m^2) (p^2 - m^2) (k^2 - m^2) \, {\mathcal{C}}(p',k';p,k) \,,$$ and the same with the $\mu$ index included on both sides. Considering only the amputation at this stage and substituting Eq. into , we find $$\label{eq:WTI_amp} q_{\mu}{\mathcal{C}}_{\textrm{amp}}^{\mu}(p',k';p,k) = {\mathrm Q_0} \bigg[ {\mathcal{C}}_{\textrm{amp}}(p',k';p+q,k) \frac{p^2 - m^2}{(p+q)^2 - m^2} - \frac{p'^2 - m^2}{(p'-q)^2 - m^2} {\mathcal{C}}_{\textrm{amp}}(p'-q,k';p,k) \bigg] \,,$$ where the ratios of amputation factors arise since the Ward-Takahashi identity changes the momenta carried by the mesons on the two sides of the equation. In the limit where $p', k', p$ and $k$ go on shell, the numerators on the right hand side of Eq. vanish but the denominators do not, yielding the well-known Ward identity: $q_{\mu} {\mathcal{W}}^{\mu} = 0$. In addition, the long-range pieces that define the difference between $\mathcal W^\mu$ and ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ \[see Fig. \[fig:2Jto2\_W\_amp\](b)\] are proportional to $(P'+P)^\mu$ and therefore also vanish when contracted with $q_\mu$. (Equivalently they are proportional to the single-particle matrix element of $\mathcal J^\mu$ and must therefore also satisfy the Ward identity.) It follows that ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ itself satisfies the identity: $q_{\mu} {\mathcal{W}}^{\mu}_{{\textrm{df}}} = 0$. Returning to the off-shell relation, Eq. , we re-express all functions in terms of $P, P', k$ and $k'$ to write $$\begin{gathered} \label{eq:WTI_amp} {q_{\mu}{\mathcal{C}}_{\textrm{amp}}^{\mu}(P' - k',k';P-k,k)} =\\ \mathrm Q_0 \bigg [ {\mathcal{C}}_{\textrm{amp}}(P' - k',k'; P' - k,k) \frac{(P-k)^2 - m^2}{(P'-k)^2 - m^2} - \frac{(P' - k ')^2 - m^2}{(P-k')^2 - m^2} {\mathcal{C}}_{\textrm{amp}}(P - k',k';P - k,k) \bigg ] \,.\end{gathered}$$ Applying a $P'_\nu$ derivative on the left-hand side then gives $$\begin{aligned} \label{eq:WTI_amp} \frac{\partial}{\partial P'_{\nu}} [\text{LHS}] = { {\mathcal{C}}_{\textrm{amp}}^{\nu}(P' - k',k';P-k,k)} + q_{\mu} \frac{\partial \, {\mathcal{C}}_{\textrm{amp}}^{\mu}(P' - k',k';P-k,k)} {{\partial P'_{\nu}} } \,,\end{aligned}$$ and, applying the same to the right-hand side, one finds $$\begin{aligned} \label{eq:WTI_amp} \frac{\partial}{\partial P'_{\nu}} [\text{RHS}] &= \mathrm Q_0 \frac{\partial \, {\mathcal{C}}_{\textrm{amp}}(P' - k',k'; P' - k,k) }{\partial P'_\nu} \frac{(P-k)^2 - m^2}{(P'-k)^2 - m^2} {\nonumber}\\ & - 2 \mathrm Q_0(P' - k)^{\nu} {\mathcal{C}}_{\textrm{amp}}(P - k',k'; P - k,k) \frac{(P-k)^2 - m^2}{[(P'-k)^2 - m^2]^2} {\nonumber}\\ & - \frac{ 2 \mathrm Q_0 (P' - k)^{\nu}}{(P-k')^2 - m^2} {\mathcal{C}}_{\textrm{amp}}(P - k',k';P - k,k) \,.\end{aligned}$$ Next, before equating the two sides, we take the zero-momentum-transfer limit ($P' \to P$) and substitute $$w^{\mu}(P-k;P-k) = 2(P-k)^{\mu} {\mathrm Q_0} \,,$$ for the ${\mathbf{1}}+{\mathcal{J}}\to{\mathbf{1}}$ matrix element at zero momentum transfer. This then gives $$\begin{aligned} {\mathcal{C}}_{\textrm{amp}}^{\mu} (P - k',k';P,P-k) & = {\mathrm Q_0} \frac{\partial}{\partial P_{\mu}} {\mathcal{C}}_{\textrm{amp}}(P - k',k';P-k,k) {\nonumber}\\ & + i{\mathcal{C}}_{\textrm{amp}}(P - k',k';P-k,k) \, \frac{i}{(P-k)^2 - m^2} \, w^{\mu}(P-k;P-k) {\nonumber}\\ & + w^{\mu}(P-k';P-k') \, \frac{i}{(P-k')^2-m^2} \, i{\mathcal{C}}_{\textrm{amp}} (P-k',k';P-k,k) \,.\end{aligned}$$ Setting $P = P'$ has greatly simplified the expressions, but care must be taken as the second and third terms on the right-hand side will diverge when we set $p', k', p$ and $k$ to their on shell values. Indeed these are the same divergences that appear in the difference between ${\mathcal{W}}^{\mu}$ and ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$, with the only subtlety that they were first defined in on-shell amplitudes at $P' - P \neq 0$. Fortunately, in the present case the distinction is unimportant because, when applied to the divergence-free amplitude, the zero-momentum-transfer and on-shell limits commute. We can thus move the second and third terms to the left-hand side and take $p', k', p,k$ on shell to conclude $${\mathcal{W}}_{{\textrm{df}}}^{\mu}(P,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star}; P, {\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) = {\mathrm Q_0} \frac{\partial}{\partial P_{\mu} } {\mathcal{M}}(s,{\mathbf{\hat{ {\mathbf{k}}}}}_f^{\star},{\mathbf{\hat{ {\mathbf{k}}}}}_i^{\star}) \,.$$ This remarkable result gives a clear interpretation to ${\mathcal{W}}_{{\textrm{df}}}^{\mu}$ in the forward limit. Finally, since the derivative is with respect to total momenta, we can easily project both sides to definite angular momentum. This leads to $${\mathcal{W}}_{{\textrm{df}}, \ell' m' , \ell m}^{\mu}(P) = \delta_{\ell' \ell} \delta_{m' m } \, {\mathrm Q_0} \frac{\partial}{\partial P_{\mu} }\, {\mathcal{M}}_{\ell}(s) \,,$$ as claimed in Eq. for the special case of $S$-wave systems. Analytic continuation of finite-volume functions below threshold {#app:cont} ================================================================ In this section we give results for the analytic continuations of the $F$- and $G$-functions below threshold. Specifically we require results for $F(P,L)$, $G(P,L)$ and $G^{\mu = 0}(P,L)$, where we recall that a single momentum argument within $G$ indicates that the initial- and final-state four-momenta are equal. Each of these can be written in terms of a class of functions naturally extending those defined in Refs. [@Luscher:1986pf; @Luscher:1991n1; @Rummukainen:1995vs; @Kim:2005gf]: [^10] $$\label{eq:cJKM} c_{JM}^{(n)}(P,L) = \bigg[ \frac{1}{L^{3}} \operatorname{ \mathchoice {\ooalign{$\displaystyle\sum$\cr\hidewidth$\displaystyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.14\height}{\scalebox{.7}{$\textstyle\sum$}}\cr\hidewidth$\textstyle\int$\hidewidth\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} {\ooalign{\raisebox{.2\height}{\scalebox{.6}{$\scriptstyle\sum$}}\cr$\scriptstyle\int$\cr}} }_{{\mathbf{k}}} \bigg] \frac{\omega_{\mathbf{k}}^{\star}}{\omega_{{\mathbf{k}}}} \frac{\sqrt{4\pi} \, k^{\star\,J} Y_{JM}({\mathbf{\hat{ {\mathbf{k}}}}}^{\star})}{(q^{\star\,2} - k^{\star\,2} + i\epsilon)^{n}} \,.$$ The relations to the finite-volume functions that we require are then given by $$\begin{aligned} \label{eq:Fc1_def} F(P,L) & = \frac{1}{2 E^{\star}} c^{(1)}_{00}(P,L) \,, \\ G(P,L) & = \frac{1}{4E^{\star}} c^{(2)}_{00}(P,L) \,, \\ G^{\mu = 0}(P,L) & = -\frac{E}{4E^{\star\,3}} c_{00}^{(1)}(P,L) + \frac{ E }{8 E^{\star}} c_{00}^{(2)}(P,L) + \frac{1}{4\sqrt{3}} \frac{P_{z}}{E^{\star\,2}} c_{10}^{(2)}(P,L) \,,\end{aligned}$$ where the last result also assumes that $\textbf P$ is parallel to the $\hat {\textbf z}$ axis. For $P^2 < (2m)^2$, the summand of $c^{(n)}$ is a smooth function of $\textbf k^\star$ with a finite region of analyticity. As a result, the sum and integral must become exponentially close to each other, with the scale in the exponential given by the grid-spacing of the sum (set by $L$) and the size of the analytic domain (set by $4 m^2 - P^2$). To make this explicit, we apply the Poisson summation formula to $c_{JM}^{(n)}$, evaluated at a generic sub-threshold four-momentum, $P_{\kappa}$, satisfying $m^2 - P_{\kappa}^2/4 = \kappa^2 $. We find $$\begin{aligned} \label{eq:KSS_general} c_{JM}^{(n)}(P_{\kappa},L) &= (-1)^{n}\sum_{{\mathbf{m}}\ne{\mathbf{0}}}\int \frac{{\textrm{d}}^3{\mathbf{k}}^{\star}}{(2\pi)^{3}} \frac{\sqrt{4\pi} k^{\star\,J} Y_{JM}({\mathbf{\hat{ {\mathbf{k}}}}}^{\star})}{(\kappa^{2} + k^{\star\,2} )^{n}} e^{i L {\mathbf{m}}\cdot {\mathbf{k}}} \,,\end{aligned}$$ where we have used the fact that the integration measure is a Lorentz invariant, ${\textrm{d}}^3 {\mathbf{k}}/ \omega_{{\mathbf{k}}} = {\textrm{d}}^3 {\mathbf{k}}^{\star} / \omega_{{\mathbf{k}}}^{\star}$. The kinematic variables in the CMF are related to the moving frame variables via standard Lorentz transformations, $$\begin{aligned} {\mathbf{k}}_{\|\|}^{\star} & = \gamma({\mathbf{k}}_{\|\|} - \omega_{{\mathbf{k}}} \, {\boldsymbol{ \beta }}) , {\nonumber}\\ {\mathbf{k}}_{\perp}^{\star} & = {\mathbf{k}}_{\perp} \,, {\nonumber}\\ \omega_{{\mathbf{k}}}^{\star} & = \gamma(\omega_{{\mathbf{k}}} - {\boldsymbol{ \beta }} \cdot {\mathbf{k}}),\end{aligned}$$ where ${\mathbf{k}}_{\perp} = {\mathbf{k}}- {\mathbf{k}}_{\|\|}$, ${\mathbf{k}}_{\|\|} = ({\mathbf{k}}\cdot {\mathbf{\hat{ {\boldsymbol{ \beta }} }}}) {\mathbf{\hat{ {\boldsymbol{ \beta }} }}}$, ${\boldsymbol{ \beta }} = {\mathbf{P}}/ E$ is the velocity, and $\gamma = E / E^{\star}$ the Lorentz factor. We can then write the phase factor in terms of the CMF momenta, $${\mathbf{m}}\cdot {\mathbf{k}}= {\mathbf{m}}' \cdot {\mathbf{k}}^{\star} + \frac{\omega_{{\mathbf{k}}}^{\star}}{E^{\star}} \, {\mathbf{m}}\cdot {\mathbf{P}},$$ with ${\mathbf{m}}'$ defined in Eq. . With these relations in hand we can write the integrand solely in terms of ${\mathbf{k}}^{\star}$, $$\begin{aligned} c_{JM}^{(n)}(P_{\kappa},L) & = \frac{(-1)^{n}}{(2\pi)^{3}} \sum_{{\mathbf{m}}\ne {\mathbf{0}}} \int_{0}^{\infty} {\textrm{d}}k^{\star} \, \frac{(k^{\star})^{J + 2} }{(k^{\star\,2} + \kappa^2)^{n} } e^{i L \omega_{{\mathbf{k}}}^{\star} {\mathbf{m}}\cdot{\mathbf{P}}/ E^{\star}} \int {\textrm{d}}{\mathbf{\hat{ {\mathbf{k}}}}}^{\star} \, \sqrt{4\pi} Y_{J M}({\mathbf{\hat{ {\mathbf{k}}}}}^{\star}) e^{iL {\mathbf{m}}' \cdot {\mathbf{k}}^{\star}}. \label{eq:CJMangularint}\end{aligned}$$ Next, we evaluate the angular piece by introducing spherical Bessel functions and making use of the standard plane wave expansion, $$\begin{aligned} e^{iL {\mathbf{m}}' \cdot {\mathbf{k}}^{\star}} & = 4\pi \sum_{\ell = 0}^{\infty} i^{\ell} j_{\ell}(L \lvert{\mathbf{m}}'\rvert k^{\star}) \sum_{m_{\ell} = -\ell}^{\ell} Y_{\ell m_{\ell}}({\mathbf{\hat{ {\mathbf{m}}}}}') Y_{\ell m_{\ell}}^{*}({\mathbf{\hat{ {\mathbf{k}}}}}^{\star}) \,,\end{aligned}$$ where $j_{\ell}(z)$ is the spherical Bessel function of the first kind. The angular integral in Eq. becomes $$\begin{aligned} \mathcal{I}_{JM}(L k^{\star}\|{\mathbf{m}}'\|,{\mathbf{m}}) &\equiv \int {\textrm{d}}{\mathbf{\hat{ {\mathbf{k}}}}}^{\star} \, \sqrt{4\pi} Y_{J M}({\mathbf{\hat{ {\mathbf{k}}}}}^{\star}) e^{iL {\mathbf{m}}' \cdot {\mathbf{k}}^{\star}} \,,\\ & = (4\pi)^{3/2} i^{J} j_{J}(L \lvert {\mathbf{m}}' \rvert k^{\star}) Y_{JM}({\mathbf{\hat{ {\mathbf{m}}}}}') \,.\end{aligned}$$ For the cases considered here we require only $$\begin{aligned} \mathcal{I}_{00}(L k^{\star}\|{\mathbf{m}}'\|,{\mathbf{m}}) &= \frac{4\pi\,\sin\left(L k^{\star}\|{\mathbf{m}}'\|\right)}{L k^{\star}\|{\mathbf{m}}'\|}, \\ \mathcal{I}_{10}(L k^{\star}\|{\mathbf{m}}'\|,{\mathbf{m}}) &= i4\pi \sqrt{3} \left(\frac{\sin\left(L k^{\star}\|{\mathbf{m}}'\|\right)}{\left(L k^{\star}\|{\mathbf{m}}'\|\right)^2} -\frac{\cos\left(L k^{\star}\|{\mathbf{m}}'\|\right)}{L k^{\star}\|{\mathbf{m}}'\| } \right)\, \frac{ {{\mathbf{m}}}' \cdot {\mathbf{P}}}{ \vert {\mathbf{m}}' \vert \vert {\mathbf{P}}\vert} \,.\end{aligned}$$ To evaluate the remaining integral over $k^\star$, we express the sinusoidal functions in $\mathcal{I}_{JM}$ in terms of exponentials and then divide the function $\mathcal{I}_{JM}$ into two terms, denoted $\mathcal{I}^{(+)}_{JM}$ and $\mathcal{I}^{(-)}_{JM}$: $\mathcal{I}^{(+)}_{JM}$ is defined by replacing the sinusoidal functions with the part of their exponential representation that decays as $k^\star \to i\infty$, e.g. $\sin(x) \to e^{ix}/(2i)$, and $\mathcal{I}^{(-)}_{JM}$ is defined in the same way for the part that decays as $k^\star \to - i\infty$, e.g. $\sin(x) \to - e^{-ix}/(2i)$. This leads to a decomposition of $c_{JM}^{(n)}$ into two integrals $$\begin{aligned} c_{JM}^{(n)}(P_{\kappa},L) & = \frac{(-1)^{n}}{(2\pi)^{3}} \sum_{x = \pm} \sum_{{\mathbf{m}}\ne {\mathbf{0}}} \int_{0}^{\infty} {\textrm{d}}k^{\star} \, \frac{(k^{\star})^{J + 2} }{(k^{\star\,2} + \kappa^2)^{n} } e^{i L \omega_{{\mathbf{k}}}^{\star} {\mathbf{m}}\cdot{\mathbf{P}}/ E^{\star}} \mathcal{I}^{(x)}_{JM}(L k^{\star}\|{\mathbf{m}}'\|,{\mathbf{m}}) \,.\end{aligned}$$ Furthermore, one can show that the $\mathcal{I}^{(\pm)}_{JM}$ factors dominate the behavior at large, imaginary $k^\star$. It follows that the $\mathcal{I}^{(+)}_{JM}$ ($\mathcal{I}^{(-)}_{JM}$) integral can be evaluated by closing the contour in the upper (lower) half of the complex plane. In addition to the $k^\star=\pm i \kappa$ pole, the integrand has branch cuts starting at $k^\star=\pm i m$, associated with the square root in $\omega_{\mathbf{k}}$. These lead to exponential corrections of the order of $\mathcal{O}(e^{-mL})$ that are ignored throughout, i.p. already in deriving the formalism considered in this work. Therefore these contributions should also dropped in the present evaluations. Keeping only the contribution from the $\kappa$ pole, we deduce $$\begin{aligned} \label{eq:cn100_asymp} c_{00}^{(1)}(P_{\kappa},L) & = -\sum_{{\mathbf{m}}\ne{\mathbf{0}}} \, e^{iL {\mathbf{m}}\cdot {\mathbf{P}}/ 2} \, \frac{ e^{-\kappa L \lvert{\mathbf{m}}'\rvert}} {4\pi L \lvert{\mathbf{m}}'\rvert} \,, \\ \label{eq:cn200_asymp} c_{00}^{(2)}(P_{\kappa},L) & = \sum_{{\mathbf{m}}\ne {\mathbf{0}}} e^{iL {\mathbf{m}}\cdot{\mathbf{P}}/ 2} \frac{e^{-\lvert{\mathbf{m}}'\rvert \kappa L}}{8\pi} \left( \frac{1}{\kappa} + 2\frac{i}{E^{\star 2} } \frac{{\mathbf{m}}\cdot {\mathbf{P}}}{\lvert{\mathbf{m}}'\rvert} \right) \,, \\ \label{eq:cn200_asymp} c_{10}^{(2)}(P_{\kappa},L) & = \sqrt{3} \sum_{{\mathbf{m}}\ne {\mathbf{0}}} e^{iL {\mathbf{m}}\cdot{\mathbf{P}}/ 2} \frac{ e^{-\lvert{\mathbf{m}}'\rvert \kappa L}}{8\pi} \frac{ {{\mathbf{m}}}' \cdot {\mathbf{P}}}{ \vert {\mathbf{m}}' \vert \vert {\mathbf{P}}\vert} \,\left( i - 2\frac{{\mathbf{m}}\cdot {\mathbf{P}}}{ E^{\star 2} \, \lvert{\mathbf{m}}'\rvert } \left( \kappa +\frac{1}{L\,\lvert{\mathbf{m}}'\rvert} \right) \right) \,.\end{aligned}$$ When $\textbf P=0$ these expressions simplify significantly $$\begin{aligned} \label{eq:cn100_asympP000} c_{00}^{(1)}(P_{\kappa},L) & = -\sum_{{\mathbf{m}}\ne{\mathbf{0}}} \, \frac{ e^{-\kappa L \lvert{\mathbf{m}}\rvert}} {4\pi L \lvert{\mathbf{m}}\rvert} \,, \\ \label{eq:cn200_asympP000} c_{00}^{(2)}(P_{\kappa},L) & = \sum_{{\mathbf{m}}\ne {\mathbf{0}}} \frac{e^{-\lvert{\mathbf{m}}\rvert \kappa L}}{8\pi \kappa} \,, \\ \label{eq:cn200_asympP000} c_{10}^{(2)}(P_{\kappa},L) & = 0 \,.\end{aligned}$$ [^1]: Similar outstanding puzzles are present in the heavy quark sector; see Refs. [@Lebed:2016hpi] and [@Brambilla:2019esw] for recent reviews. [^2]: We point the reader to Refs. [@Briceno:2017max] and [@Hansen:2019nir] for recent reviews detailing the progress of the field. See also Refs. [@Hansen:2017mnd] and [@Bulava:2019kbi] for alternative methods to determine rates and amplitudes, which require significantly larger volumes as well as techniques to regulate the inverse Laplace transform. [^3]: Reference [@Detmold:2004qn] was the first work to consider the coupling of an external current to finite-volume two-hadron states. In that publication, the authors consider the use of background fields in the context of a fixed-order expansion in a particular effective field theory, an alternative to the matrix elements of currents discussed here. [^4]: We denote CMF quantities with a $\star$ superscript throughout. [^5]: Virtual bound states (e.g. the dineutron) and resonances (e.g. the $\rho$) arise on the second sheet, for which $\text{Im} \, q^\star < 0$. [^6]: In these expressions we are dropping terms that depend on the form factor of the neutral particle. These can be readily included and only introduce slight technical complications in various equations. Such contributions are numerically suppressed compared to the charged particle’s form factor, and vanish identically as $Q^2 \to 0$. [^7]: We leave a detailed proof of this claim and an analysis of its consequences to future work [@analytic]. [^8]: Related results for bound states can be found in Refs. [@Beane:2003da; @Bour:2011ef; @Briceno:2013bda; @Briceno:2013hya]. [^9]: We work here with isospin symmetry, approximating $m_p = m_n$. [^10]: The $c_{JM}^{(n)}$ are proportional to the dimensionless functions denoted by ${\mathcal{Z}}_{JM}^{(n)}$ in Ref. [@Baroni:2018iau] $$c_{JM}^{(n)}(P,L) = \frac{L^{2n-3}}{(2\pi)^{2n}} \left( \frac{2\pi}{L} \right)^{J} {\mathcal{Z}}_{JM}^{(n)}(P,L) \,.$$ For the analytic work presented here, the dimensionful versions prove slightly more convenient.	{ "pile_set_name": "ArXiv" }
--- abstract: \| Abstract. The class of Church-Rosser congruential languages has been introduced by , Narendran, and Otto in 1988. A language $L$ is Church-Rosser congruential (belongs to CRCL), if there is a finite, confluent, and length-reducing semi-Thue system $S$ such that $L$ is a finite union of congruence classes modulo $S$. To date, it is still open whether every regular language is in CRCL. In this paper, we show that every star-free language is in CRCL. In fact, we prove a stronger statement: For every star-free language $L$ there exists a finite, confluent, and subword-reducing semi-Thue system $S$ such that the total number of congruence classes modulo $S$ is finite and such that $L$ is a union of congruence classes modulo $S$. The construction turns out to be effective. Keywords. String rewriting; Church-Rosser system; star-free language; aperiodic monoid; local divisor. author: - 'Volker Diekert[^1]' - 'Manfred Kufleitner[^2]' - 'Pascal Weil[^3]' date: 'November 17, 2011' title: 'Star-Free Languages are Church-Rosser Congruential' --- Introduction {#intro} ============ Church-Rosser congruential languages (CRCL) are a nonterminal-free form of [Church-Rosser]{}languages (CRL). Both classes have been defined in [@McNaughtonNO88], and it was shown there that CRCL forms a proper subclass in CRL. Languages in CRL enjoy various nice properties. For example their word problem is decidable in linear time. A detailed discussion with links to further references can be found in the PhD-thesis of Niemann [@NiemannPhD02], see also [@NiemannO05]. We content ourselves to define CRCL: A language $L \in A^$ is called a Church-Rosser congruential language, if there is a finite, [length-reducing]{}, and confluent semi-Thue system $S {\subseteq}A^ \times A^$ such that $L$ is a finite union of congruence classes modulo $S$. This means that $L$ contains a finite set $F$ of shortest words such that we have $w \in L$ [if and only if]{}every rewriting procedure starting on $w$ and using $S$ terminates in one of the finitely many words in $F$. It was also shown in [@McNaughtonNO88] that all deterministic context-free languages are Church-Rosser. However, surprisingly it is not known whether all regular languages are CRCL. The general conjecture is “yes”, but so far only partial results have been established as in [@NiemannW02]. The most advanced result has been announced by Reinhardt and Thérien [@reinhardtT03]: According to their manuscript, if a regular language has a group as its syntactic monoid, then this language is in CRCL. In this note we consider the complementary class of group-free regular languages; and we show that they belong to CRCL. A regular language is group-free* if its syntactic monoid is group-free. This means it is [aperiodic]{}. There are many other characterizations for this class. A fundamental result of [Schützenberger]{}says that the class of aperiodic language ${\mathrm{AP}}(A)$ is exactly the same as the class of star-free languages ${\mathrm{SF}}(A)$ [@sch65sf:short]. It is the class where the Krohn-Rhodes decomposition leads to a wreath product of the three-element commutative idempotent reset-monoid $U_2$ [@kr65tams]. It is also the class ${{\mathrm{FO}}}(A,<) $ of languages definable in first-order logic [@mp71:short]; and this is the same as the class ${{\mathrm{LTL}}}(A)$ of languages definable in the linear temporal logic [@kam68]. A proof that ${{\mathrm{FO}}}(A,<) = {\mathrm{SF}}(A) = {\mathrm{AP}}(A) = {{\mathrm{LTL}}}(A)$ can be conveniently arranged in a cycle. The inclusion ${{\mathrm{FO}}}(A,<) {\subseteq}{\mathrm{SF}}(A)$ can be explained very nicely with [Ehrenfeucht-Fra[ï]{}ss[é]{}]{}-games [@Ehrenfeucht61]. The inclusion ${\mathrm{SF}}(A) {\subseteq}{\mathrm{AP}}(A)$ follows [Schützenberger]{}’s original idea. The inclusion ${\mathrm{AP}}(A) {\subseteq}{{\mathrm{LTL}}}(A)$ is done in the survey [@dg08SIWT:short] with the concept of local divisors which play a prominent role here, too. The final inclusion ${{\mathrm{LTL}}}(A) {\subseteq}{{\mathrm{FO}}}(A,<)$ is trivial. Coming back to the class of Church-Rosser congruential languages, our main result shows ${\mathrm{SF}}(A) {\subseteq}$ CRCL. Actually, we prove a much stronger result. First we define [subword-reducing]{}semi-Thue systems which are a proper subclass of finite [length-reducing]{}semi-Thue systems. For every language $L \in {\mathrm{AP}}(A)$ we effectively construct a finite [subword-reducing]{}confluent semi-Thue system $ S {\subseteq}A^* \times A^$ such that the total number of congruence classes modulo $S$ is finite and $L$ is a union of such classes, see [Theorem \[thm:main\]]{}. A main tool in our proof is the notion of local divisor, see [Section \[sec:ld\]]{} for a definition. In the final section of this paper, [Section \[sec:alg\]]{}, we explain our constructions in a rigorously algebraic framework. This part is mainly intended for possible future work. In order to give a complete positive solution to the conjecture that all regular languages are CRCL, it remains to combine our approach with the one in [@reinhardtT03]. There are however quite a number of obstacles for a fruitful combination. So, we leave the general conjecture as a challenging research problem. Preliminaries and Notation {#pre} ========================== In the following $A$ means a finite alphabet, an element of $A$ is called a [letter]{}, and $A^$ denotes the [free monoid]{} generated by $A$. It is the set of [words]{} over $A$. The empty word is denoted by 1. The [length]{} of a word $u$ is denoted by ${\left\|\mathinner{u}\right\|}$. We have ${\left\|\mathinner{u}\right\|} = n$ for $u= a_1 \cdots a_n$ where $a_i \in A$. The empty word has length $0$. We carefully distinguish between the notion of factor and subword. Let $u,v \in A^$. The word $u$ is called a [factor]{} of $v$ if there is a factorization $v= xuy$. It is called a [subword]{} of $v$ if there is a factorization $v= x_0u_1x_1 \cdots u_kx_k$ such that $u= u_1 \cdots u_k$. A subword is also sometimes called a [scattered subword]{} in the literature. A [semi-Thue system]{} over $A$ is a subset $S{\subseteq}A^\times A^$. The elements are called [rules]{}. We frequently write $\ell {\overset{}{\longrightarrow}} r$ for rules $(\ell,r)$. A system $S$ is called [[length-reducing]{}]{} if we have ${\left\|\mathinner{\ell}\right\|} > {\left\|\mathinner{r}\right\|} $ for all rules $(\ell,r) \in S$. It is called [[subword-reducing]{},]{} if $r$ is a subword of $\ell$ and $\ell \neq r$ for all rules $(\ell,r) \in S$. Every [subword-reducing]{}system is [length-reducing]{}, but not vice versa. Every system $S$ defines the rewriting relation ${{\underset{S}{\Longrightarrow}}} {\subseteq}A^ \times A^$ by $$\begin{aligned} u {\underset{S}{\Longrightarrow}} v \;\text{ if } \; u=p\ell q, \; v= prq \; \text{ for some rule } \; (\ell,r) \in S.\end{aligned}$$ By ${\overset{}{\underset{S}{\Longrightarrow}}}$ we mean the reflexive and transitive closure of ${\underset{S}{\Longrightarrow}}$. By ${\overset{}{\underset{S}{\Longleftrightarrow}}}$ we mean the symmetric, reflexive, and transitive closure of ${\underset{S}{\Longrightarrow}}$. We also write $u {\overset{}{\underset{S}{\Longleftarrow}}}v$ whenever $v{\overset{}{\underset{S}{\Longrightarrow}}}u$. The system $S$ is [confluent]{} if for all $u{\overset{}{\underset{S}{\Longleftrightarrow}}}v$ there is some $w$ such that $u{\overset{}{\underset{S}{\Longrightarrow}}}w{\overset{}{\underset{S}{\Longleftarrow}}}v$. Note that $u {\overset{}{\underset{S}{\Longrightarrow}}} v$ implies that ${\left\|\mathinner{u}\right\|} > {\left\|\mathinner{v}\right\|}$ for [length-reducing]{}systems. For [subword-reducing]{}systems it implies that the set of subwords in $v$ is a proper subset of the set of subwords in $u$. By ${\mathrm{IRR}}(S)$ we denote the set of irreducible words, i.e., the set of words where no left-hand side occurs as any factor. The relation ${{\overset{}{\underset{S}{\Longleftrightarrow}}}} {\subseteq}A^ \times A^$ is a congruence, hence the congruence classes $[u]_S = \{v \in A^\mid u {\overset{}{\underset{S}{\Longleftrightarrow}}} v\}$ form a monoid which is denoted by $A^/S $. A finite semi-Thue system $S$ can be viewed as a finite set of defining relations. Hence, $A^/S $ becomes a finitely presented monoid. A semi-Thue system $S$ is called a [Church-Rosser system]{} if it is [length-reducing]{}and confluent. A language $L {\subseteq}A^$ is called a [ Church-Rosser congruential language]{} if there is a finite Church-Rosser system $S$ such that $L$ can be written as a finite union of congruence classes $[u]_S$. \[rem:basic\] A semi-Thue system $S$ is a Church-Rosser system [if and only if]{}(1) it is length-reducing and (2) every congruence class has exactly one irreducible element. Let $\pi: A^* \to A^/S, \; u \mapsto [u]_S$ be the canonical [homomorphism]{} and $S$ be a finite Church-Rosser system. Then $\pi^{-1}(K)$ is a Church-Rosser congruential language as soon as $K$ is finite. \[con:mcnno\] Every regular language is a Church-Rosser congruential language. \[exa:1\] Consider the language $L = (bc)^+$. A Church-Rosser system for $L$ is given by the one-rule semi-Thue system $S = {\left\{\mathinner{cbc \longrightarrow c}\right\}}$. The monoid $ {\left\{b,c\right\}}^/S$ is infinite. However $L = [bc]_S$; and hence $u \in L$ if and only if $u {\overset{}{\underset{S}{\Longrightarrow}}} bc$. [$\Diamond$]{} A manuscript of Reinhardt and Thérien [@reinhardtT03] says that [Conjecture \[con:mcnno\]]{} is true in case that the syntactic monoid of the regular language is a group. Here, we are going to prove an even stronger result for aperiodic languages, i.e., for languages where the syntactic monoid is group-free. As our proof uses subword-reducing systems in the induction hypothesis, we cannot incorporate the statement of Reinhardt and Thérien (using length-reducing rather than subword-reducing systems) as base in our induction scheme. So the status of [Conjecture \[con:mcnno\]]{} remains open, in general. \[def:cr\] Let ${\varphi}: A^ \to M$ be a [homomorphism]{}to a finite monoid $M$. We say ${\varphi}$ [factorizes through a finite Church-Rosser monoid $A^/ S$]{} if there is a finite Church-Rosser system $S$ such that $A^/ S$ is a finite monoid and $[u]_S {\subseteq}{\varphi}^{-1}({\varphi}(u))$ for all $u \in A^$. A classical fact states that a language $L {\subseteq}A^$ is regular [if and only if]{}it is [recognizable]{}, i.e., there is a [homomorphism]{}${\varphi}: A^* \to M$ to a finite monoid $M$ such that $L = {\varphi}^{-1}({\varphi}(L))$. We also say that ${\varphi}$ (or that $M$) recognizes $L$. Recall that a finite monoid $M$ is called [aperiodic]{} if there exists some $n\in {\mathbb{N}}$ such that $x^n = x^{n+1}$ for all $x \in M$. Accordingly, a language $L {\subseteq}A^$ is called [aperiodic]{} if it is recognized by some finite aperiodic monoid $M$. Note that if ${\varphi}$ factorizes through a finite Church-Rosser monoid, then we have $${\varphi}: A^ \overset{\pi}\to A^/S \overset{\psi}\to M,$$ where $S$ is a Church-Rosser system such that $A^/ S$ is a finite. \[con\] Let ${\varphi}: A^* \to M$ be a [homomorphism]{}to a finite monoid $M$. Then ${\varphi}$ factorizes through a finite Church-Rosser monoid. [Conjecture \[con\]]{} is stronger than [Conjecture \[con:mcnno\]]{}. However, we believe that a positive solution to [Conjecture \[con:mcnno\]]{} comes through a proof of [Conjecture \[con\]]{}. Actually, the result in [@reinhardtT03] also announces that [Conjecture \[con\]]{} is true for finite groups. We are going to show here that an even stronger statement than [Conjecture \[con\]]{} holds for finite aperiodic monoids. \[exa:2\] Consider again the language $L = (bc)^+$ from [Example \[exa:1\]]{}. Another Church-Rosser system for $L$ is given by $$\begin{aligned} {2} S = \{ && bbb &\longrightarrow bb,\ bbc \longrightarrow bb,\ cbb \longrightarrow bb, \\ && ccc &\longrightarrow bb,\ ccb \longrightarrow bb,\ bcc \longrightarrow bb \\ && bcb &\longrightarrow b, \ cbc \longrightarrow c \}. \end{aligned}$$ As in [Example \[exa:1\]]{} we have $L = [bc]_S$; but here, the monoid ${\left\{b,c\right\}}^/S$ is finite. It has $7$ elements: $[1]_S$, $[b]_S$, $[c]_S$, $[bc]_s$, $[cb]_s$, $[bb]_S$, and $[cc]_S$. Note that $S$ is not subword-reducing. [$\Diamond$]{} Local divisors {#sec:ld} ============== The notion of [local divisor]{} dates back to a technical report of Meyberg where he introduced this concept in commutative algebra, see [@FeTo02; @Mey72]. In finite semigroup theory and formal language theory the explicit definition of a local divisor appeared first in [@dg06IC]. Since then it turned out to be a very useful tool for simplifying classical proofs like in [@dg08SIWT:short; @DiekertKS11] or in finding new results like in this paper. The definition of a local divisor extends the definition of a Schützenberger group for the $\mathcal H$-class of an arbitrary element, [@cp67; @rs09qtheory]. A category generalization is being used by Steinberg and Costa in the context of symbolic dynamics (unpublished). In this paper we use local divisors for aperiodic monoids, only. Let $M$ be a monoid and let $c \in M$. We put on the subsemigroup $cM \cap Mc$ a monoid structure by defining a new multiplication $\circ$ as follows: $$xc \circ cy = xcy.$$ It is straightforward to see that $\circ$ is well-defined and $(cM \cap Mc, \circ)$ is a monoid with neutral element $c$. The following observation is crucial: If the monoid $M$ is finite and aperiodic, then ${\left\|\mathinner{cM \cap Mc}\right\|} < {\left\|\mathinner{M}\right\|}$ whenever $c \neq 1$. This is clear, because $1 \in {cM \cap Mc}$ implies that $c$ is a unit of $M$, but $c \neq 1$ and there are no non-trivial units in aperiodic monoids. The set $M' = {\left\{x\mathrel{\left\|\vphantom{x}\vphantom{cx \in Mc}\right.}cx \in Mc\right\}}$ is a submonoid of $M$, and $c{\cdot}: M' \to cM \cap Mc : x \mapsto cx$ is a surjective homomorphism. In particular, if $M$ is aperiodic, then $(cM \cap Mc, \circ)$ is aperiodic, too. Since $(cM \cap Mc, \circ)$ is the homomorphic image of a submonoid it is a divisor of $M$. We therefore call $(cM \cap Mc, \circ)$ the [local divisor]{} of $M$ at $c$. Note that if $c = c^2$ is an idempotent, then $(cM \cap Mc, \circ) = (cMc,\cdot)$ is the usual [local monoid]{} defined by the subsemigroup $cMc$ of $M$. Thus, the notion of local divisor generalizes the notion of local monoid from idempotents to arbitrary elements. [Conjecture \[con\]]{} holds for aperiodic monoids {#sec:con} ================================================== We have the following result. \[thm:main\] Let ${\varphi}: A^ \to M$ be a [homomorphism]{}to a finite aperiodic monoid $M$. Then ${\varphi}$ factorizes through a finite aperiodic Church-Rosser monoid $A^* / S$ where $S$ is subword-reducing. The rest of this section is devoted to the proof of [Theorem \[thm:main\]]{}. The proof is by induction on the parameter $({\left\|\mathinner{M}\right\|},{\left\|\mathinner{A}\right\|})$ with lexicographic order. The result is true if $\varphi(A^)$ is trivial. Note that this covers $M = {\left\{1\right\}}$ as well as $A = \emptyset$. In the remaining case there is a letter $c \in A$ such that ${\varphi}(c) \neq 1$. We let $B= A{\setminus}{\left\{c\right\}}$, and for better reading we identify $c$ and ${\varphi}(c)\in M$. Since $c \neq 1 \in M$ and $M$ is aperiodic, $c$ is not a unit. Hence $M_c= cM \cap Mc$ has less elements than $M$. Since ${\left\|\mathinner{B}\right\|} < {\left\|\mathinner{A}\right\|}$ we find, by induction, a finite [subword-reducing]{}Church-Rosser system $R {\subseteq}B^ \times B^$ such that the restriction ${\varphi}\|_{B^}: B^\to M$ factorizes through a finite Church-Rosser monoid $B^/ R$. In particular, $(\ell, r ) \in R$ implies ${\varphi}(\ell) = {\varphi}(r)$. For $u \in B^$ let ${\widehat{u}\,}$ denote the unique word such that ${\widehat{u}\,}\in {\mathrm{IRR}}(R)$ and $u{\overset{}{\underset{R}{\Longrightarrow}}} {\widehat{u}\,}$. The subset $K = {\mathrm{IRR}}(R)c {\subseteq}A^$ is a finite code. This means that $K^$ is freely generated, as a submonoid of $A^$, by the finite set $K$. Note that $K^+ {\subseteq}A^c$. Consider the [homomorphism]{}$\psi: K^\to (M_c, \circ)$ which is given by $\psi({\widehat{u}\,} c) = c{\varphi}(u)c$. We have $c{\varphi}(u)c = {\varphi}(c {\widehat{u}\,} c)$. In particular, $\psi$ is well-defined. By induction $\psi: K^\to (M_c, \circ)$ factorizes through a finite aperiodic Church-Rosser monoid $K^/ T$, where $T{\subseteq}K^ \times K^$ is a finite [subword-reducing]{}Church-Rosser system. Consider a rule $(\ell, r) \in T$. It has the form $${\widehat{u_1}\,}c \cdots {\widehat{u_m}\,}c {\overset{}{\longrightarrow}}{\widehat{v_1}\,}c \cdots {\widehat{v_n}\,}c$$ where the ${\widehat{u_i}\,}c$ and ${\widehat{v_j}\,}c$ are letters in $K$, every right-hand side ${\widehat{v_1}\,}c \cdots {\widehat{v_n}\,}c\in K^$ is a proper subword of ${\widehat{u_1}\,}c \cdots {\widehat{u_m}\,}c\in K^+$. Since $K^* {\subseteq}A^$ we can read $T$ as a semi-Thue system over $A$ as well. Next, we define a new system ${\widehat{T}\,} {\subseteq}A^ \times A^$ as follows: $${\widehat{T}\,} = {\left\{c\ell {\overset{}{\longrightarrow}} c{r}\mathrel{\left\|\vphantom{c\ell {\overset{}{\longrightarrow}} c{r}}\vphantom{(\ell, r) \in T}\right.}(\ell, r) \in T\right\}}.$$ We collect some important properties of ${\widehat{T}\,}$ in a remark: \[rem:wht\] The semi-Thue system ${\widehat{T}\,}{\subseteq}A^ \times A^$ satisfies the following assertions. 1. ${\widehat{T}\,}$ is [subword-reducing]{}, because $T$ has this property. This is crucial. Knowing only that $T$ is [length-reducing]{}as a system over $K^$ would not be enough to conclude that ${\widehat{T}\,}$ is [length-reducing]{}as a system over $A^$. 2. ${\widehat{T}\,}$ is confluent. For this it is crucial that we added a letter $c$ on the left. This allows to read the words ${\widehat{u}\,} c$ as letters in $K$ and the confluence of $T$ transfers to the confluence of ${\widehat{T}\,}$. If there was no $c$ on the left, then $T$ could contain rules $abc {\overset{}{\longrightarrow}} 1$ and $bc {\overset{}{\longrightarrow}} 1$, but $a$ is no left-hand side in $T$. Over $K$ the words $abc$ and $bc$ are letters, hence there is no overlap in $K^$. 3. $c\ell {\overset{}{\longrightarrow}} c{r} \in {\widehat{T}\,}$ implies ${\varphi}(c\ell) = {\varphi}(cr)$. This is a straightforward calculation in local divisors: Let $c\ell = c u_1 c \cdots u_m c$ and $cr = c v_1 c \cdots v_n c$ with $u_i, v_i \in {\mathrm{IRR}}(R)$. By induction, we have $\psi(\ell) = \psi(r)$ and thus $$\begin{aligned} \varphi(c \ell) &= \varphi(cu_1c) \circ \cdots \circ \varphi(cu_mc) \\ &= \psi(u_1c) \circ \cdots \circ \psi(u_mc) \\ &= \psi(u_1c \cdots u_m c) = \psi(v_1c \cdots v_n c) \\ &= \psi(v_1c) \circ \cdots \circ \psi(v_n c) \\ &= \varphi(cv_1c) \circ \cdots \circ \varphi(cv_nc) = \varphi(c r). \end{aligned}$$ The proof of [Theorem \[thm:main\]]{} is now a direct consequence of the following lemma which shows that the system $S = R \cup {\widehat{T}\,}$ has the desired properties. \[lem:main\] The semi-Thue system $S = R \cup {\widehat{T}\,}$ over $A$ satisfies the following assertions. 1. \[i\] $S$ is [subword-reducing]{}. 2. \[ii\] $S$ is confluent. 3. \[iii\] $\ell {\overset{}{\longrightarrow}} {r} \in S $ implies ${\varphi}(\ell) = {\varphi}(r)$. 4. \[iv\] $A^/S$ is a finite aperiodic monoid. Assertion \[i\] is clear, because $R$ and ${\widehat{T}\,}$ are [subword-reducing]{}. Assertion \[ii\] is clear, because there is no overlap of left-hand sides between rules of $R$ and ${\widehat{T}\,}$. Assertion \[iii\] is clear, because $R$ and ${\widehat{T}\,}$ have this property. It remains to show \[iv\]. By induction $K^/ T$ is finite. Hence there is a maximal value $\mu$ such that every word in $K^$ of length at least $\mu$ is reducible. We conclude that: $${\mathrm{IRR}}(S) {\subseteq}{\left\{{\widehat{u_0}\,} c {\widehat{u_1}\,}\cdots c{\widehat{u_m}\,}\mathrel{\left\|\vphantom{{\widehat{u_0}\,} c {\widehat{u_1}\,}\cdots c{\widehat{u_m}\,}}\vphantom{ {\widehat{u_i}\,} \in {\mathrm{IRR}}(R) \wedge 0 \leq m \leq \mu}\right.} {\widehat{u_i}\,} \in {\mathrm{IRR}}(R) \wedge 0 \leq m \leq \mu\right\}}.$$ Since ${\mathrm{IRR}}(R)$ is finite, we see that ${\mathrm{IRR}}(S)$ is a subset of a finite set, and thus the finiteness of ${\mathrm{IRR}}(S)$ and of $A^/S$ follow. This leaves us to show that $A^/S$ is aperiodic. We have to show that there exists some $n\in {\mathbb{N}}$ such that for all $u = {\widehat{u_0}\,}c {\widehat{u_1}\,}\cdots c {\widehat{u_m}\,} \in {\mathrm{IRR}}(S)$ we have $u^{n+1} {\overset{}{\underset{S}{\Longleftrightarrow}}} u^{n}$. Let $v = {\widehat{u_1}\,}c\cdots {\widehat{u_m\, u_0}\,}c$. Then $u^{n+1} {\overset{}{\underset{R}{\Longleftrightarrow}}} p c v^n q$ and $u^n {\overset{}{\underset{R}{\Longleftrightarrow}}} p c v^{n-1} q$ for some $p,q \in A^$. Therefore, it is enough to show that $cv^{n} {\overset{}{\underset{S}{\Longleftrightarrow}}} cv^{n-1}$ whenever $n$ is large enough. The ${\widehat{u_i}\,}c$’s are code words of $K$, hence letters in the alphabet $K$ and we can read $v \in K^$. Here we can use induction, and we know $v^{n} {\overset{}{\underset{T}{\Longleftrightarrow}}} v^{n-1}$ if $n$ is large enough, because $K^/ T$ is aperiodic. This implies $cv^{n} {\overset{}{\underset{{\widehat{T}\,}}{\Longleftrightarrow}}} cv^{n-1}$ and hence the result. This completes the proof of [Theorem \[thm:main\]]{}. Consider again the language $L = (bc)^+$ from [Example \[exa:1\]]{} and [Example \[exa:2\]]{}. Its syntactic monoid is $M = {\left\{\mathinner{1,b,c,bc,cb,0}\right\}}$ with $bb = cc = 0$, $bcb = b$, $cbc = c$, $1$ is neutral, and $0$ is a zero element. In particular, $bc$ and $cb$ are idempotent. Here, the syntactic homomorphism $\varphi_L : {\left\{\mathinner{b,c}\right\}}^* \to M$ is induced by $b \mapsto b$ and $c \mapsto c$. We apply the above algorithm for obtaining a [Church-Rosser]{}monoid factorizing $\varphi_L$. First we choose to localize at $c$. Then $N = {\left\{\mathinner{1,b,0}\right\}}$ is the submonoid generated by $b$. The restriction of $\varphi_L$ to $b^$ factorizes through the [Church-Rosser]{}monoid defined by the system $$R = {\left\{\mathinner{bbb \longrightarrow bb}\right\}}.$$ This leads to the irreducible elements ${\mathrm{IRR}}(R) = {\left\{\mathinner{1,b,bb}\right\}}$. Now, the homomorphism $\psi : {\left\{\mathinner{c,bc,bbc}\right\}}^ \to M_c$ is defined by $x \mapsto cx$ for $x \in {\left\{\mathinner{c,bc,bbc}\right\}}$. Note that we consider ${\left\{\mathinner{c,bc,bbc}\right\}}$ as a three-letter alphabet. In particular, $M_c = {\left\{\mathinner{c,0}\right\}}$ and $c \mapsto 0$, $bbc \mapsto 0$, and $bc \mapsto c$. For $\psi$ we obtain the rules $$\begin{aligned} {2} T = \{\ && (c)(c) &\longrightarrow (c), \\ && (bc) &\longrightarrow 1,\\ && (bbc)(bbc) &\longrightarrow (bbc), \\ && (c)(bbc)(c) &\longrightarrow (c) \qquad \} \end{aligned}$$ The parenthesis are for identifying letters of the alphabet of $\psi$. This leads to the system $$\begin{aligned} {2} \widehat{T} = \{\ && ccc &\longrightarrow cc, \\ && cbc &\longrightarrow c,\\ && cbbcbbc &\longrightarrow cbbc, \\ && ccbbcc &\longrightarrow cc \qquad \} \end{aligned}$$ and $S = R \cup \widehat{T}$ is the system for $\varphi$. In ${\mathrm{IRR}}(S)$ there are $65$ irreducible elements and $bbcbbccbbcbb$ is the longest one. [$\Diamond$]{} Algebraic constructions {#sec:alg} ======================= The aim of this section is to place the explicit constructions from the previous [Section \[sec:con\]]{} into a broader algebraic context. It shows that the quotient monoid $A^/S$ in [Lemma \[lem:main\]]{} has an algebraic interpretation. Rees-extension monoids and Church-Rosser systems ------------------------------------------------ Let $\rho: P\to Q$ be a mapping between two monoids $P$ and $Q$. We are going to define the Rees-extension monoid* of $\rho$ which we shall denote by $E(\rho)$. If $\rho$ is chosen properly, then $E(\rho)$ coincides with the monoid $A^/S$ where $S {\subseteq}A^ \times A^$ is the [subword-reducing]{}confluent semi-Thue system of [Lemma \[lem:main\]]{}, see [Proposition \[WasSollDieseSection\]]{}. As a carrier set for the monoid $E(\rho)$ we choose the disjoint union $P {\mathbin{\dot{\cup}}}(P\times Q \times P)$. The multiplication is as follows: $$\begin{aligned} {2} u \cdot v &= uv &\quad&\text{for $u,v\in P$.} \\ x \cdot (u,q, v)\cdot y &=(xu,q, vy) &\quad&\text{for $x,u,v,y\in P$ and $q \in Q$.}\\ (u,q, v)\cdot(x,r, y)&=(u,q\,\rho(vx)\,r,y ) &\quad&\text{for $u,v,x,y\in P$ and $q,r \in P$.}\end{aligned}$$ Now, $P$ is a submonoid of $E(\rho)$ and $P \times Q \times P$ is an ideal. As a semigroup, $P \times Q \times P$ is a special case of the Rees-matrix construction, see e.g. [@cp67; @rs09qtheory]: The mapping $\rho$ defines a $P \times P$ matrix $\mathcal{R}$ with coefficients in $Q$ by $\mathcal{R}(v,x) = \rho(vx)$; and the multiplication in $P \times Q \times P$ can be written as $(u,p, v)\cdot(x,q, y) = (u,p\,\mathcal{R}(v,x)\,q,y)$. In the following we let $c = \rho(1)\in Q$. Multiplying triples $(1,q,1)$ and $(1,r,1)$ yields $(1,q,1)\cdot (1,r,1) = (1,q\,\rho(1)\, r,1)= (1,qcr,1)$. In particular, the sandwich construction $(Q,\#_c)$ appears as a subsemigroup, where $\#_c$ denotes the standard sandwich-multiplication defined by $q\mathop{\#_c}r = qcr$. We have $(1,1,1)^n = (1,c^{n-1},1)$ and, more general, $(u,q,v)^n= \big( u, \big(q\, \rho(vu)\big)^{n-1}q, v\big)$ for all $n \geq 1$. It follows that $E(\rho)= P {\mathbin{\dot{\cup}}}(P\times Q \times P)$ is aperiodic if both $P$ and $Q$ are aperiodic. For [Proposition \[WasSollDieseSection\]]{} below, we apply the Rees-extension monoid to the setting in [Section \[sec:con\]]{}. We start with a [homomorphism]{}${\varphi}: A^ \to M$ to a finite aperiodic monoid $M$, the alphabet $A$ is the disjoint union of $B$ and ${\left\{c\right\}}$, $P$ is the quotient $B^* / R$, $Q$ is the quotient $K^* / T$ for $K = {\mathrm{IRR}}(R) c$. Since we can identify $P = B^* / R$ and ${\mathrm{IRR}}(R)$, we define $\rho : P \to Q$ by $\rho(u) = [uc]_T$ for $u \in {\mathrm{IRR}}(R)$. Now, $A^* / S$ from [Section \[sec:con\]]{} and $E(\rho)$ coincide: \[WasSollDieseSection\] In the situation above, $A^* / S$ and $E(\rho)$ are isomorphic. Let $\sigma : {\mathrm{IRR}}(S) \to E(\rho)$ be defined by $$\begin{aligned} \sigma(u_0) &= u_0 \qquad \text{and} \\ \sigma(u_0 c u_1 \cdots c u_{k+1}) &= \big(u_0,\rho(u_1) \cdots \rho(u_{k}), u_{k+1}\big) \end{aligned}$$ for $k \geq 0$ and $u_i \in B^* \cap {\mathrm{IRR}}(R)$. Here, we indentify $P$ with ${\mathrm{IRR}}(R)$, and $Q$ with ${\mathrm{IRR}}(T)$. In particular, by definition of $P$ and $Q$, the mapping $\sigma$ is surjective. Suppose $\sigma(u_0 c u_1 \cdots c u_{k+1}) = \sigma(v_0 c v_1 \cdots c v_{\ell+1})$ for $k,\ell \geq 0$ and $u_i, v_i \in B^* \cap {\mathrm{IRR}}(R)$. Then $u_0 = v_0$ and $u_{k+1} = v_{\ell + 1}$. Moreover, $c u_1 c \cdots u_k c \in {\mathrm{IRR}}(S)$ and thus $(u_1c) \cdots (u_k c) \in K^* \cap {\mathrm{IRR}}(T)$. Similarly, $(v_1c) \cdots (v_\ell c) \in K^* \cap {\mathrm{IRR}}(T)$. Now, $\rho(u_1) \cdots \rho(u_k) = \rho(v_1) \cdots \rho(v_\ell)$ implies $(u_1c) \cdots (u_k c) = (v_1 c) \cdots (v_\ell c)$ in $K^$ and thus $cu_1c \cdots u_k c = cv_1c \cdots v_\ell c$ in $A^$. This shows $u_0 c u_1 \cdots c u_{k+1} = v_0 c v_1 \cdots c v_{\ell + 1}$. We conclude that $\sigma$ is injective. It remains to show that $\sigma$ is a homomorphism. Let $u, v \in {\mathrm{IRR}}(S)$ and $uv {\overset{}{\underset{S}{\Longrightarrow}}} w \in {\mathrm{IRR}}(S)$, i.e., $[u]_S [v]_S = [w]_S$. If $u,v \in B^$, then $\sigma(u)\sigma(v) = w = \sigma(w)$. Let now $uv,w \in A^* c A^$ and $w = w_0 c w_1 \cdots c w_{m+1}$. If $u \in B^$, $v = v_0 c v_1 \cdots c v_{\ell+1}$ and $uv_0 {\overset{}{\underset{S}{\Longrightarrow}}} x \in {\mathrm{IRR}}(S)$, then $w_0 = x$ and $c w_1 \cdots c w_{m+1} = c v_1 \cdots c v_{\ell + 1}$. It follows $\sigma(u)\sigma(v) = (x,\rho(w_1) \cdots \rho(w_m),w_{m+1}) = \sigma(w)$. The case $v \in B^$ is symmetric. Let now $u = u_0 c u_1 \cdots c u_{k+1} \in A^* c A^$ and $v = v_0 c v_1 \cdots c v_{\ell+1} \in A^ c A^$ with $u_i, v_i \in B^$. Let $u_{k+1} v_0 {\overset{}{\underset{S}{\Longrightarrow}}} x \in {\mathrm{IRR}}(S)$. Then $u_{k+1} v_0 {\overset{}{\underset{R}{\Longrightarrow}}} x$. We have $c u_1 \cdots c u_{k+1} v_0 c \cdots v_k c {\overset{}{\underset{S}{\Longrightarrow}}} c w_1 c \cdots w_m c$. By construction of $S$ we see that $$c u_1 \cdots c u_k c x c v_1 c \cdots v_\ell c {\overset{}{\underset{{\widehat{T}\,}}{\Longrightarrow}}} c w_1 c \cdots w_m c,$$ and hence $$(u_1 c) \cdots (u_{k} c) (x c) (v_1 c) \cdots (v_\ell c) {\overset{}{\underset{T}{\Longrightarrow}}} (w_1 c) \cdots (w_m c),$$ i.e., $\rho(u_1) \cdots \rho(u_k) \rho(x) \rho(v_1) \cdots \rho(v_\ell) = \rho(w_1) \cdots \rho(w_m)$ in $Q$. We conclude that $\sigma(u) \sigma(v) = \sigma(w)$. Rees-extension monoids and local divisors ----------------------------------------- Let $\rho : P \to Q$ be arbitrary again. Observe that $c\neq 1 \in Q$, in general. In the remainder of this section, we draw a connection between local divisors and the Rees-extension monoid. We define an alphabet ${C}$ by the disjoint union ${C}= (P \setminus {\left\{1\right\}}) {\mathbin{\dot{\cup}}}{\left\{c\right\}}$. The mapping $\rho$ induces a homomorphism $\tau : {C}^ \to E(\rho)$ by defining $\tau(x) = x$ for $x \in P \setminus {\left\{1\right\}}$ and $\tau(c) = (1,1,1)$. By considering $(P\setminus {\left\{1\right\}})^* c$ as an infinite alphabet, $\rho$ also induces a homomorphism $\sigma : \big((P\setminus {\left\{1\right\}})^* c\big)^* \to Q$ by $\sigma(uc) = \rho({\varepsilon}(u))$ for $u \in (P\setminus {\left\{1\right\}})^$. Here, ${\varepsilon}: (P\setminus {\left\{1\right\}})^ \to P$ is the evaluation homomorphism. Consider a [homomorphism]{}${\gamma}: {C}^* \to M$ with ${\gamma}(c) = c \in M$. The aim is to find a condition such that ${\gamma}$ factorizes through $\tau: {C}^* \to E(\rho)$. This means we wish to write ${\gamma}= \tau \psi$ for some suitable [homomorphism]{}$\psi:E(\rho) \to M$. The condition we are looking for is statement \[lifti\] of [Proposition \[prop:lift\]]{}. \[prop:lift\] Let ${\gamma}: {C}^* \to M$ be a [homomorphism]{}with ${\gamma}(c) = c \in M$. If $Q$ is generated by $\rho(P)$, then the following assertions are equivalent. 1. \[lifti\] For $w, w'\in \big((P\setminus {\left\{1\right\}})^* c\big)^$ the equality $\sigma(w) = \sigma(w')\in Q$ implies $c{\gamma}(w) = c{\gamma}( w')\in cM \cap Mc$. 2. \[liftii\] There exists a [homomorphism]{}$\psi_c: Q \to M_c$ with $M_c = (cM \cap Mc, \circ)$ such that the following diagram commutes. (0,0) node (Pc) [$\big((P\setminus {\left\{1\right\}})^ c\big)^$]{}; (3,0) node (Q) [$Q$]{}; (0,-2) node (M) [$Mc \cup {\left\{1\right\}}$]{}; (3,-2) node (Mc) [$M_c$]{}; (Pc) – node\[above\] [$\sigma$]{} (Q); (Pc) – node\[left\] [${\gamma}$]{} (M); (Q) – node\[right\] [$\psi_c$]{} (Mc); (M) – node\[above\] [$c {\cdot}$]{} node\[below\] [$x \mapsto cx$]{} (Mc); 3. \[liftiii\] There exists a [homomorphism]{}$\psi: E(\rho) \to M$ such that the following diagram commutes. (0,0) node (C) [$C^$]{}; (3,0) node (E) [$E(\rho)$]{}; (0,-2) node (M) [$M$]{}; (C) – node\[above\] [$\tau$]{} (E); (C) – node\[left\] [${\gamma}$]{} (M); (E) – node\[below right,inner sep=0.5mm\] [$\psi$]{} (M); \[lifti\] ${\text{$\;\Rightarrow\;$}}$ \[liftii\]: We define $\psi_c({\sigma}(w)) = c{\gamma}(w)$. Condition \[lifti\] says that $\psi_c: Q \to M_c$ is well-defined. It is a [homomorphism]{}because $\gamma$ and the left-shift $c{\cdot} : Mc \cup {\left\{1\right\}} \to M_c, x \mapsto cx$ are [homomorphisms]{} and $Q{\setminus}{\left\{1\right\}} {\subseteq}{\sigma}({C}^c)$. \[liftii\] ${\text{$\;\Rightarrow\;$}}$ \[liftiii\]: For $u \in P {\subseteq}E(\rho)$ we define $\psi (u) = {\gamma}(u) = u \in P {\subseteq}M$. All other elements in $E(\rho)$ have the form $(u, {\sigma}({\alpha}),v)$ with $u,v \in P$ and . Define $\psi(u,{\sigma}({\alpha}),v) = u\psi_c({\sigma}({\alpha}))v$. This in an element in $M$ because $M_c {\subseteq}M$. Now, $\psi_c({\sigma}({\alpha}))= c {\gamma}({\alpha})$. Hence, $\psi(u,{\sigma}({\alpha}),v) = {\gamma}(uc{\alpha}v)$. Since ${\gamma}$, $\tau$ are [homomorphism]{}s and $\tau$ is surjective, $\psi$ is a [homomorphism]{}, too. \[liftiii\] ${\text{$\;\Rightarrow\;$}}$ \[lifti\]: Consider $w \in (P^c)^$. We have $\tau(cw) = (1, {\sigma}(w), 1)$. By \[liftiii\] we have ${\gamma}(cw) = \psi(1, {\sigma}(w), 1) $ . In particular, ${\sigma}(w) = {\sigma}( w')\in Q$ implies $c{\gamma}(w) = c{\gamma}( w')\in M_c$. [10]{} A. H. Clifford and G. B. Preston. , volume 1,2. American Mathematical Society, 1961,1967. V. Diekert and P. Gastin. Pure future local temporal logics are expressively complete for [M]{}azurkiewicz traces. , 204:1597–1619, 2006. Conference version in LATIN 2004, LNCS 2976, 170–182, 2004. V. Diekert and P. Gastin. First-order definable languages. In [Logic and Automata: History and Perspectives]{}, Texts in Logic and Games, pages 261–306. Amsterdam University Press, 2008. V. [Diekert]{}, M. [Kufleitner]{}, and B. [Steinberg]{}. . 1111.1585v1, Nov. 2011. A. Ehrenfeucht. An application of games to the completeness problem for formalized theories. , 49:129–141, 1961. A. [Fern[á]{}ndez L[ó]{}pez]{} and M. [Toc[ó]{}n Barroso]{}. The local algebras of an associative algebra and their applications. In J. Misra, editor, [Applicable Mathematics in the Golden Age]{}, pages 254–275. Narosa, 2002. J. A. W. Kamp. . PhD thesis, University of California, Los Angeles (California), 1968. K. Krohn and J. Rhodes. , 116:450–464, 1965. R. McNaughton, P. Narendran, and F. Otto. hurch-[R]{}osser [T]{}hue systems and formal languages. , 35(2):324–344, 1988. R. McNaughton and S. Papert. . The MIT Press, 1971. K. Meyberg. Lectures on algebras and triple systems. Technical report, University of Virginia, Charlottesville, 1972. G. Niemann. . Kassel University Press, 2002. G. Niemann and F. Otto. The [C]{}hurch-[R]{}osser languages are the deterministic variants of the growing context-sensitive languages. , 197:1–21, 2005. G. Niemann and J. Waldmann. Some regular languages that are [C]{}hurch-[R]{}osser congruential. In [Revised Papers from the 5th International Conference on Developments in Language Theory]{}, DLT ’01, pages 330–339, London, UK, 2002. Springer-Verlag. K. Reinhardt and D. Th[é]{}rien. Some more regular languages that are [C]{}hurch [R]{}osser congruential. In [13. Theorietag, Automaten und Formale Sprachen, Herrsching]{}, pages 97–103, 2003. J. Rhodes and B. Steinberg. Springer Monographs in Mathematics. Springer, 2009. M. P. Sch[ü]{}tzenberger. On finite monoids having only trivial subgroups. , 8:190–194, 1965. [^1]: Institut für Formale Methoden der Informatik, University of Stuttgart. The work on this paper has been initiated by the program Automata Theory and Applications* at the Institute for Mathematical Sciences, National University of Singapore in September 2011. The first author would like to thank NUS for the hospitality and the organizing committee chaired by Frank Stephan for the invitation. He also thanks Klaus Reinhardt for the introduction to the topic. [^2]: Institut für Formale Methoden der Informatik, University of Stuttgart, Germany. The second author was supported by the German Research Foundation (DFG) under grant . [^3]: LaBRI, Universit[é]{} de Bordeaux and CNRS, France. The third author was supported by the grant ANR 2010 BLAN 0202 01 FREC.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A complete characterization is provided of Hankel matrices commuting with Jacobi matrices which correspond to hypergeometric orthogonal polynomials from the Askey scheme. It follows, as the main result of the paper, that the generalized Hilbert matrix is the only prominent infinite-rank Hankel matrix which, if regarded as an operator on $\ell^{2}(\mathbb{N}_{0})$, is diagonalizable by application of the commutator method with Jacobi matrices from the mentioned families.' address: - '[\[]{}Franti¨ek ¦tampach[\]]{} Department of Applied Mathematics, Faculty of Information Technology, Czech Technical University in Prague, Thákurova 9, 160 00 Praha, Czech Republic' - '[\[]{}Pavel ¦ťovíček[\]]{} Department of Mathematics, Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Trojanova 13, 12000 Praha, Czech Republic' author: - Franti¨ek ¦tampach - Pavel ¦ťovíček title: On Hankel matrices commuting with Jacobi matrices from the Askey scheme --- Introduction and motivation =========================== Operators acting on $\ell^{2}({\mathbb{N}}_{0})$ determined by a Hankel matrix $$\mathcal{H}=\begin{pmatrix}h_{0} & h_{1} & h_{2} & h_{3} & \dots\\ h_{1} & h_{2} & h_{3} & h_{4} & \dots\\ h_{2} & h_{3} & h_{4} & h_{5} & \dots\\ h_{3} & h_{4} & h_{5} & h_{6} & \dots\\ \vdots & \vdots & \vdots & \vdots & \ddots \end{pmatrix}\!,\label{eq:def_hankel}$$ with $h_{k}\in{{\mathbb{R}}}$, belong to one of the basic classes of linear operators whose general theory has been systematically developed during the last century [@peller_03; @power_82]. In contrast to other well known classes such as Jacobi, Schrödinger, Toeplitz operators, etc., there are only very few known concrete Hankel matrices that admit an explicit diagonalization, i.e, whose spectral problem is explicitly solvable. In fact, concerning Hankel matrices of infinite rank, the authors were aware of only one such example - the famous (generalized) Hilbert matrix - until recently. The (generalized) Hilbert matrix is the Hankel matrix with $$h_{k}=\frac{1}{k+t},\quad k\in{\mathbb{N}}_{0},$$ where $t\notin-{\mathbb{N}}_{0}$ is a parameter. The Hilbert matrix determines a bounded operator $H_{t}$ on $\ell^{2}({\mathbb{N}}_{0})$ which can be, quite surprisingly, diagonalized fully explicitly. In full generality, the diagonalization of $H_{t}$ was obtained by Rosenblum in [@rosenblum_pams58a; @rosenblum_pams58b]. Preliminary results appeared even earlier [@magnus_ajm50; @shanker_pcps49], however. Rosenblum’s original approach relies on an idea of the so-called commutator method that was successfully applied to other operators later on, see [@kalvoda-stovicek_lma16; @stampach-stovicek_jmaa19; @yafaev_fap10]. The commutator method relies on two main steps. First, one has to find a commuting operator with simple spectrum and explicitly solvable spectral problem and, second, determine a spectral mapping. In the case of the Hilbert matrix, one may prefer to consider the transformed Hankel matrix with the entries $$\tilde{h}_{k}=\frac{(-1)^{k}}{k+t}\,,\,\ k\in{\mathbb{N}}_{0}.\label{eq:def_tilde_h}$$ Obviously, the corresponding operator $\tilde{H}_{t}$ is unitarily equivalent to $H_{t}$ via unitary transform $U={\mathop\mathrm{diag}\nolimits}(1,-1,1,-1,\dots)$. The operator $\tilde{H}_{t}$ can be readily shown to commute with a self-adjoint Jacobi operator $J_{t}$ given by a tridiagonal matrix $\mathcal{J}_{t}$ with the entries $$\left(\mathcal{J}_{t}\right)_{n,n}=2n(n+t),\quad\left(\mathcal{J}_{t}\right)_{n,n+1}=\left(\mathcal{J}_{t}\right)_{n+1,n}=(n+1)(n+t),\quad n\in{\mathbb{N}}_{0},\label{eq:def_J_t}$$ which is the first ingredient for the commutator method to be applied. Indeed, the spectrum of a self-adjoint Jacobi operator is always simple. Moreover, up to an inessential addition of a multiple of the identity operator, the Jacobi operator $J_{t}$ corresponds to a particular subfamily of orthogonal polynomials known as the continuous dual Hahn polynomials listed in the Askey hypergeometric scheme [@koekoek-etal_10]. As a result, the spectral problem of $J_{t}$ is explicitly solvable. Next, by a general fact, $\tilde{H}_{t}=h(J_{t})$ for a Borel function $h$. The second step of the commutator method is to determine $h$ which can be done with the aid of a known generating function formula for the continuous dual Hahn polynomials, see [@kalvoda-stovicek_lma16] for details. The desired diagonalization of $\tilde{H}_{t}$, and hence also of $H_{t}$, then follows. Of course, relying on the commutator method may seem rather restrictive. On the other hand, whenever the method turned out to be applicable it proved itself to be very powerful. In particular, as far as the Hilbert matrix is concerned, the authors are not aware of an alternative way of its diagonalization, other than a variant of the commutator method. The Askey scheme of hypergeometric orthogonal polynomials [@koekoek-etal_10] can be viewed as a rich source of Jacobi operators with an explicitly solvable spectral problem. For instance, the Wilson polynomials standing at the top of the Askey scheme depend on four additional parameters and hence the same is true for the corresponding Jacobi matrix. Our initial goal was to explore thoroughly this reservoir while trying to find out whether the (generalized) Hilbert matrix is the only Hankel matrix which can be diagonalized with the aid of the commutator method with Jacobi matrices from the Askey scheme or whether there are other infinite-rank Hankel matrices having such a property. And this question was the main motivation for the current paper. In this study, we distinguish nine families of Hermitian non-decomposable Jacobi matrices $$\mathcal{J}=\begin{pmatrix}\beta_{0} & \alpha_{0}\\ \alpha_{0} & \beta_{1} & \alpha_{1}\\ & \alpha_{1} & \beta_{2} & \alpha_{2}\\ & & \alpha_{2} & \beta_{3} & \alpha_{3}\\ & & & \ddots & \ddots & \ddots \end{pmatrix}\!,\label{eq:def_jacobi}$$ with $\alpha_{n}>0$ and $\beta_{n}\in{{\mathbb{R}}}$, corresponding to hypergeometric orthogonal polynomials from the Askey scheme whose measure of orthogonality is positive and supported on an infinite subset of ${{\mathbb{R}}}$. These comprise Wilson, Continuous dual Hahn, Continues Hahn, Jacobi, Meixner–Pollaczek, Meixner, Laguerre, Charlier, and Hermite polynomials. On the other hand, Racah, Hahn, dual Hahn, Krawtchouk, Pseudo-Jacobi, and Bessel polynomials are excluded for the associated Jacobi matrices are finite or non-Hermitian. For each family, our main theorem determines the set of parameters for which there exists a nonzero Hankel matrix commuting with the Jacobi matrix in question. In addition, in these quite rare situations, the space of commuting Hankel matrices is always of dimension $2$ with an explicitly described basis; see Theorem \[thm:main\] below. It may also seem that the set of Jacobi matrices, which we restrict our analysis to, is too special. However, from the view point of the commutator method, whose principle ingredient is always a commuting operator with an explicitly solvable spectral problem, this restriction is reasonable. Naturally, it would be really interesting to have a general characterization of Jacobi matrices commuting with a nontrivial Hankel matrix. Such a problem was solved by Grünbaum in [@grunbaum_laa81] for finite Jacobi matrices commuting with Toeplitz matrices. However, as far as semi-infinite Jacobi and Hankel matrices are concerned, a solution of this problem seems to be out of reach at the moment. This paper deals with the commutation equation with semi-infinite matrices on the algebraic level only. Passing to operators acting on $\ell^{2}({\mathbb{N}}_{0})$, one would have to additionally require the columns (and hence the rows) of a Hankel matrix $\mathcal{H}$ to be square summable. In that case $\mathcal{H}$ defines a densely defined operator on $\ell^{2}({\mathbb{N}}_{0})$ with the canonical basis in its domain. A consequence of our main theorem is that the (generalized) Hilbert matrix is the only infinite-rank Hankel matrix which can be diagonalized by application of the commutator method using Hermitian non-decomposable Jacobi matrices from the Askey scheme. More precisely, an immediate consequence of Theorem 1 is as follows: Let $\mathcal{H}$ be a Hankel matrix with $\ell^{2}$-columns and ${\mathop\mathrm{rank}\nolimits}\mathcal{H}>1$ and $\mathcal{J}$ any Hermitian non-decomposable Jacobi matrix from the Askey scheme (concretely specified below). If $\mathcal{H}\mathcal{J}=\mathcal{J}\mathcal{H}$ then $\mathcal{H}$ is a scalar multiple of the Hankel matrix $\tilde{\mathcal{H}}_{t}$ with entries determined in for some $t$. This statement is an answer to our original question and emphasizes even more the prominent role played by the Hilbert matrix. In a bit more detail, our analysis reveals, too, that in all cases when the prominent Hankel matrix $\tilde{\mathcal{H}}_{t}$ commutes with a Jacobi matrix $\mathcal{J}$ corresponding to orthogonal polynomials from the Askey scheme the diagonal and the off-diagonal sequence of $\mathcal{J}$ are polynomial functions of index of order 2 and 4, respectively. There is another family of orthogonal polynomials known as the Stieltjes–Carlitz polynomials with the corresponding Jacobi matrix of the same type but not included in the hypergeometric Askey scheme. Rather than to hypergeometric series these polynomials are intimately related to the Jacobian elliptic functions. This observation led the authors to apply the commutator method to Jacobi operators associated with the Stieltjes–Carlitz polynomials. This study resulted in a discovery of four new explicitly diagonalizable Hankel matrices. Their diagonalization is treated in a separate paper [@stampach-stovicek_inprep]. The article is organized as follows. In Section 2, the basic definitions of the considered Jacobi matrices from the Askey scheme as well as restrictions on the involved parameters are recalled. Then the main result is formulated in Theorem \[thm:main\]. Section 3 is devoted to an auxiliary result used in several particular proofs, especially in those corresponding to the most complicated cases. In Section \[sec:proofs\], a proof of the main result is given. The mentioned families of orthogonal polynomials are treated case by case starting from the most complicated one (Wilson) and proceeding to simpler cases following the hierarchy of the Askey scheme. A more sophisticated approach is used for the first three families depending on 3 or 4 additional parameters while the remaining cases are treated using a rather elementary computational approach. Notation and statement of the main result {#sec:main} ========================================= In the notation, the Jacobi matrix associated with a particular family of orthogonal polynomials is distinguished by a corresponding superscript, i.e., we consider nine families of Jacobi matrices $$\mathcal{J}=\mathcal{J}^{j},\quad\mbox{ for }j\in\{\text{W},\text{CdH},\text{CH},\text{J},\text{MP},\text{M},\text{L},\text{C},\text{H}\}.$$ The entries of these matrices can depend on up to four parameters. For the chosen parametrization, we strictly follow [@koekoek-etal_10]. The concrete definitions of the entries are listed below. I\) The Jacobi matrix $\mathcal{J}^{\text{W}}$ associated with the Wilson polynomials: $$\alpha_{n}^{\text{W}}=\alpha_{n}^{\text{W}}(a,b,c,d):=\sqrt{A_{n}^{\text{W}}C_{n+1}^{\text{W}}}\quad\mbox{ and }\quad\beta_{n}^{\text{W}}=\beta_{n}^{\text{W}}(a,b,c,d):=A_{n}^{\text{W}}+C_{n}^{\text{W}}-a^{2},\label{eq:def_alpha_beta_wilson}$$ where $$\begin{aligned} A_{n}^{\text{W}}=A_{n}^{\text{W}}(a,b,c,d) & :=\frac{(n+a+b)(n+a+c)(n+a+d)(n+a+b+c+d-1)}{(2n+a+b+c+d-1)(2n+a+b+c+d)}\,,\label{eq:def_A_wilson}\\ C_{n}^{\text{W}}=C_{n}^{\text{W}}(a,b,c,d) & :=\frac{n(n+b+c-1)(n+b+d-1)(n+c+d-1)}{(2n+a+b+c+d-2)(2n+a+b+c+d-1)}\,.\label{eq:def_C_wilson}\end{aligned}$$ Though not obvious at first glance in the case of $\beta_{n}^{\text{W}}$, the coefficients $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ are both symmetric functions of the parameters $a,b,c,d$. The parameters $a,b,c,d\in{{\mathbb{C}}}$ are restricted so that one of the following holds: 1. ${\mathop\mathrm{Re}\nolimits}a,{\mathop\mathrm{Re}\nolimits}b,{\mathop\mathrm{Re}\nolimits}c,{\mathop\mathrm{Re}\nolimits}d>0$ and the non-real parameters occur in complex conjugate pairs. 2. Up to a permutation of parameters, $a<0$ and $a+b$, $a+c$, $a+d$ positive or a pair of complex conjugates occur with positive real parts. II\) The Jacobi matrix $\mathcal{J}^{\text{CdH}}$ associated with the Continuous dual Hahn polynomials: $$\alpha_{n}^{\text{CdH}}=\alpha_{n}^{\text{CdH}}(a,b,c):=\sqrt{A_{n}^{\text{CdH}}C_{n+1}^{\text{CdH}}}\quad\mbox{ and }\quad\beta_{n}^{\text{CdH}}=\beta_{n}^{\text{CdH}}(a,b,c):=A_{n}^{\text{CdH}}+C_{n}^{\text{CdH}}-a^{2},$$ where $$\begin{aligned} A_{n}^{\text{CdH}}=A_{n}^{\text{CdH}}(a,b,c) & :=(n+a+b)(n+a+c)\,,\\ C_{n}^{\text{CdH}}=C_{n}^{\text{CdH}}(a,b,c) & :=n(n+b+c-1)\,.\end{aligned}$$ Hence $$\alpha_{n}^{\text{CdH}}=\sqrt{(n+1)(n+a+b)(n+a+c)(n+b+c)}$$ and $$\beta_{n}^{\text{CdH}}=2n^{2}+(2a+2b+2c-1)n+ab+ac+bc.$$ Clearly, both $\alpha_{n}^{\text{CdH}}$ and $\beta_{n}^{\text{CdH}}$ are symmetric functions of $a,b,c$. The parameters $a,b,c\in{{\mathbb{C}}}$ are restricted such that one of the following holds: 1. $a,b,c$ are positive except possibly a pair of complex conjugates with positive real parts. 2. Up to a permutation of parameters, $a<0$ and $a+b$, $a+c$, are positive or a pair of complex conjugates with positive real parts. III\) The Jacobi matrix $\mathcal{J}^{\text{CH}}$ associated with the Continuous Hahn polynomials: $$\alpha_{n}^{\text{CH}}=\alpha_{n}^{\text{CH}}(a,b,c,d):=\sqrt{-A_{n}^{\text{CH}}C_{n+1}^{\text{CH}}}\quad\mbox{ and }\quad\beta_{n}^{\text{CH}}=\beta_{n}^{\text{CH}}(a,b,c,d):=\mathrm{i}(A_{n}^{\text{CH}}+C_{n}^{\text{CH}}+a),$$ where $$\begin{aligned} A_{n}^{\text{CH}}=A_{n}^{\text{CH}}(a,b,c,d) & :=-\frac{(n+a+b+c+d-1)(n+a+c)(n+a+d)}{(2n+a+b+c+d-1)(2n+a+b+c+d)}\,,\\ C_{n}^{\text{CH}}=C_{n}^{\text{CH}}(a,b,c,d) & :=\frac{n(n+b+c-1)(n+b+d-1)}{(2n+a+b+c+d-2)(2n+a+b+c+d-1)}\,.\end{aligned}$$ In this case, $\alpha_{n}^{\text{CH}}$ and $\beta_{n}^{\text{CH}}$ are not symmetric functions of $a,b,c,d$ but remain invariant under the permutations of $(a,b,c,d)$ which are composed of two $2$-cycles ($(b,a,d,c)$, $(c,d,a,b)$, $(d,c,b,a)$). The parameters are supposed to be such that $$c=\bar{a},\;d=\bar{b}\quad\mbox{ and }\quad{\mathop\mathrm{Re}\nolimits}a>0,\;{\mathop\mathrm{Re}\nolimits}b>0.\label{eq:param_assum_chahn}$$ Then $-A_{n}^{\text{CH}}C_{n+1}^{\text{CH}}>0$ and $\beta_{n}\in\mathbb{R}$ for all $n\in{\mathbb{N}}_{0}$, indeed. IV\) The Jacobi matrix $\mathcal{J}^{\text{J}}$ associated with the Jacobi polynomials: $$\alpha_{n}^{\text{J}}=\alpha_{n}^{\text{J}}(\alpha,\beta):=\sqrt{\frac{4(n+1)(n+\alpha+1)(n+\beta+1)(n+\alpha+\beta+1)}{(2n+\alpha+\beta+1)(2n+\alpha+\beta+2)^{2}(2n+\alpha+\beta+3)}}\label{eq:def_alpha_jacobi}$$ and $$\beta_{n}^{\text{J}}=\beta_{n}^{\text{J}}(\alpha,\beta):=\frac{\beta^{2}-\alpha^{2}}{(2n+\alpha+\beta)(2n+\alpha+\beta+2)},\label{eq:def_beta_jacobi}$$ where $\alpha,\beta>-1$. V\) The Jacobi matrix $\mathcal{J}^{\text{MP}}$ associated with the Meixner–Polaczek polynomials: $$\alpha_{n}^{\text{MP}}=\alpha_{n}^{\text{MP}}(\lambda,\phi):=\frac{\sqrt{(n+1)(n+2\lambda)}}{2\sin\phi}\quad\mbox{ and }\quad\beta_{n}^{\text{MP}}=\beta_{n}^{\text{MP}}(\lambda,\phi):=-\frac{n+\lambda}{\tan\phi},\label{eq:def_alpha_beta_meixner-pollaczek}$$ where $\lambda>0$ and $\phi\in(0,\pi)$. For $\phi=\pi/2$, $\beta_{n}^{\text{MP}}$ is to be understood as zero for all $n\in{\mathbb{N}}_{0}$. VI\) The Jacobi matrix $\mathcal{J}^{\text{M}}$ associated with the Meixner polynomials: $$\alpha_{n}^{\text{M}}=\alpha_{n}^{\text{M}}(c,\beta):=\frac{\sqrt{c(n+1)(n+\beta)}}{1-c}\quad\mbox{ and }\quad\beta_{n}^{\text{M}}=\beta_{n}^{\text{M}}(c,\beta):=\frac{n+(n+\beta)c}{1-c},\label{eq:def_alpha_beta_meixner}$$ where $\beta>0$ and $c\in(0,1)$. VII\) The Jacobi matrix $\mathcal{J}^{\text{L}}$ associated with the Laguerre polynomials: $$\alpha_{n}^{\text{L}}=\alpha_{n}^{\text{L}}(\alpha):=\sqrt{(n+1)(n+\alpha+1)}\quad\mbox{ and }\quad\beta_{n}^{\text{L}}=\beta_{n}^{\text{L}}(\alpha):=2n+\alpha+1,\label{eq:def_alpha_beta_laguerre}$$ where $\alpha>-1$. VIII\) The Jacobi matrix $\mathcal{J}^{\text{C}}$ associated with the Charlier polynomials: $$\alpha_{n}^{\text{C}}=\alpha_{n}^{\text{C}}(a):=\sqrt{a(n+1)}\quad\mbox{ and }\quad\beta_{n}^{\text{C}}=\beta_{n}^{\text{C}}(a):=n+a,\label{eq:def_alpha_beta_charlier}$$ where $a>0$. IX\) The Jacobi matrix $\mathcal{J}^{\text{H}}$ associated with the Hermite polynomials: $$\alpha_{n}^{\text{H}}=\sqrt{\frac{n+1}{2}}\quad\mbox{ and }\quad\beta_{n}^{\text{H}}=0.\label{eq:def_alpha_beta_hermite}$$ Note that for any $j\in\{\text{W},\text{CdH},\text{CH},\text{J},\text{MP},\text{M},\text{L},\text{C},\text{H}\}$, $\alpha_{n}^{j}$ and $\beta_{n}^{j}$ may be both regarded as analytic functions in $n$ on a neighborhood of infinity. Hence, whenever convenient, $\alpha_{n}^{j}$ and $\beta_{n}^{j}$ are extended to non-integer values of $n$. Finally, by a commutant of $\mathcal{J}$, we mean the space of semi-infinite matrices $\mathcal{A}\in{{\mathbb{C}}}^{\infty,\infty}$ commuting with $\mathcal{J}$, i.e, $\mathcal{AJ}=\mathcal{JA}$ entrywise. Since $\mathcal{J}$ is banded the matrix products $\mathcal{AJ}$ and $\mathcal{JA}$ are well defined for any $\mathcal{A}\in{{\mathbb{C}}}^{\infty,\infty}$. The special attention is paid to Hankel matrices commuting with a given Jacobi matrix $\mathcal{J}$, therefore we define the Hankel commutant of $\mathcal{J}$ by $${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}):=\{\mathcal{H}\in{{\mathbb{C}}}^{\infty,\infty}\mid\mathcal{JH}=\mathcal{HJ}\mbox{ and }\mathcal{H}\mbox{ is a Hankel matrix}\}.$$ Clearly, ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J})$ is a linear subspace of ${{\mathbb{C}}}^{\infty,\infty}$. Our main result characterizes ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$ for $j\in\{\text{W},\text{CdH},\text{CH},\text{J},\text{MP},\text{M},\text{L},\text{C},\text{H}\}$. Only in very particular cases, ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$ is nontrivial, i.e., there exists a nonzero Hankel matrix commuting with a Jacobi matrix from the Askey scheme. \[thm:main\] For any $j\in\{\text{W},\text{CdH},\text{CH},\text{J},\text{MP},\text{M},\text{L},\text{C},\text{H}\}$, the Hankel commutant${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$ is either trivial or a two-dimensional space. Moreover, $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})=2$ if and only if $\alpha_{n}^{j}$ and $\beta_{n}^{j}$ depend polynomially on $n$ and $\alpha_{-1}^{j}=0$. Below, we characterize all situations when ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$ is nontrivial by specifying parameters case by case. In each case, with ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$ being nontrivial, we determine two sequences, $h^{(1)}$ and $h^{(2)}$, which define two Hankel matrices (as in ) that form a basis of ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{j})$. 1. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})=2$ if and only if, up to a permutation, the parameters $a,b,c,d$ fulfill $$a=\frac{3}{4},\quad b=\frac{t}{2}+\frac{1}{4},\quad c=\frac{t}{2}-\frac{1}{4},\quad d=\frac{1}{4},$$ for $t>0$, and if that is the case, we have $$h_{k}^{(1)}=\frac{(-1)^{k}}{k+t}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k},\quad k\in{\mathbb{N}}_{0}.$$ 2. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CdH}})=2$ if and only if, up to a permutation, the parameters $a,b,c$ fulfill $$a=b=\frac{1}{2},\quad c=t-\frac{1}{2},$$ for $t>0$, and $h^{(1)}$ and $h^{(2)}$ are as in the case (i). 3. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})=2$ if and only if the parameters $a,b,c,d$ fulfill $$a=\bar{c}=\frac{1}{4}+{{\rm i}}t,\quad b=\bar{d}=\frac{3}{4}+{{\rm i}}t,\quad\mbox{ or }\quad a=\bar{c}=\frac{3}{4}+{{\rm i}}t,\quad b=\bar{d}=\frac{1}{4}+{{\rm i}}t,$$ for $t\in{{\mathbb{R}}}$, and if that is the case, we have $$h_{k}^{(1)}=\sin\frac{k\pi}{2}\quad\mbox{ and }\quad h_{k}^{(2)}=\cos\frac{k\pi}{2},\quad k\in{\mathbb{N}}_{0}.$$ 4. ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{J}})$ is trivial for all $\alpha,\beta>-1$. 5. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{MP}})=2$ if and only if $\lambda=1/2$ and $\phi\in(0,\pi)$ arbitrary. If so we have $$h_{k}^{(1)}=\sin(k\phi)\quad\mbox{ and }\quad h_{k}^{(2)}=\cos(k\phi),\quad k\in{\mathbb{N}}_{0}.$$ 6. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{M}})=2$ if and only if $\beta=1$ and $c\in(0,1)$ arbitrary. If so we have $$h_{k}^{(1)}=(-1)^{k}c^{k/2}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k}c^{-k/2},\quad k\in{\mathbb{N}}_{0}.$$ 7. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{L}})=2$ if and only if $\alpha=0$ and we have $$h_{k}^{(1)}=(-1)^{k}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k}k,\quad k\in{\mathbb{N}}_{0}.$$ 8. ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{C}})$ is trivial for all $a>0$. 9. $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{H}})=2$ and $h^{(1)}$ and $h^{(2)}$ are as in the case (iii). Clearly, those restrictions of parameters which are dictated by requiring the Jacobi matrices from the Askey scheme to be Hermitian are not essential for every claim of Theorem \[thm:main\]. For example, it follows from the claim (ii) that the Jacobi matrix $\mathcal{J}_{t}$ defined in and the Hankel matrix $\tilde{\mathcal{H}}_{t}$ whose entries are given by commute for any parameter $t>0$. Obviously this assertion can be extended to all $t\in{{\mathbb{C}}}\setminus\left(-{\mathbb{N}}_{0}\right)$. An auxiliary result {#sec:aux} =================== The following lemma will be used repeatedly below. Its statement can be of independent interest. It has been also applied in [@stampach-stovicek_inprep]. \[lem:Mzw\] Let $p$ and $q$ be complex functions which are meromorphic in a neighborhood of $\infty$ and assume that the order of the pole at $\infty$ equals $2$ for both of them. Further let $\epsilon\in\mathbb{C}$, $\epsilon\neq0$, and put, for $z,w\in\mathbb{C}$ sufficiently large, $$M(z,w):=\left(\begin{array}{cc} p(z+\epsilon)-p(w-\epsilon) & q(z+\epsilon)-q(w-\epsilon)\\ p(z-\epsilon)-p(w+\epsilon) & q(z-\epsilon)-q(w+\epsilon) \end{array}\right)\!.$$ Let us write the determinant of $M(z,w)$ in the form $$\det M(z,w)=\left((z-w)^{2}-4\epsilon^{2}\right)\delta(z,w).$$ If at least one of the functions $p(z)$ and $q(z)$ is not a polynomial in $z$ of degree $2$ and the set of functions $\{1,p,q\}$ is linearly independent, then one of the following two cases happens: \(i) for every $w\in\mathbb{C}$ sufficiently large there exists $\underset{z\to\infty}{\text{lim}}\delta(z,w)\in\mathbb{C}\setminus\{0\}$, \(ii) for every $w\in\mathbb{C}$ sufficiently large there exists $\underset{z\to\infty}{\text{lim}}z\delta(z,w)\in\mathbb{C}\setminus\{0\}$.\ Consequently, for every $w\in\mathbb{C}$ sufficiently large there exists $R(w)>0$ such that for all $z\in\mathbb{C}$, $\left\|z\right\|>R(w)$, the matrix $M(z,w)$ is regular. Without loss of generality we can assume that the leading coefficient in the Laurent expansion about $\infty$ equals $1$ for both $p$ and $q$. So we can write $$p(z)=z^{2}+\sum_{k=-1}^{\infty}p_{k}z^{-k},\ q(z)=z^{2}+\sum_{k=-1}^{\infty}q_{k}z^{-k}.$$ Let $$P(x,y):=\sum_{n=1}^{\infty}p_{n}\sum_{j=0}^{n-1}x^{j}y^{n-j-1},\ Q(x,y):=\sum_{n=1}^{\infty}q_{n}\sum_{j=0}^{n-1}x^{j}y^{n-j-1},$$ and $$P_{0}(y):=P(0,y)=\sum_{n=1}^{\infty}p_{n}y^{n-1},\ Q_{0}(y):=Q(0,y)=\sum_{n=1}^{\infty}q_{n}y^{n-1}.$$ Note that $$\frac{p(x)-p(y)}{x-y}=x+y+p_{-1}-\frac{1}{xy}\,P\!\left(\frac{1}{x},\frac{1}{y}\right)\!,\ \frac{q(x)-q(y)}{x-y}=x+y+q_{-1}-\frac{1}{xy}\,Q\!\left(\frac{1}{x},\frac{1}{y}\right)\!.\label{eq:p-diff-P}$$ Clearly, $p$ and $q$ are polynomials of degree $2$ if and only if $P_{0}(y)=Q_{0}(y)=0$. Furthermore, the functions $\{1,p,q\}$ are linearly dependent if and only if $p(z)-p_{0}=q(z)-q_{0}$ for all sufficiently large $z$, and this happens if and only if $p_{-1}=q_{-1}$ and $P_{0}(y)=Q_{0}(y)$ (the latter equation can be equivalently replaced by $P(x,y)=Q(x,y)$). Thus we assume that either (1) $P_{0}(y)\neq Q_{0}(y)$ or (2) $P_{0}(y)=Q_{0}(y)\neq0$ (equivalently, $P(x,y)=Q(x,y)\neq0$) and $p_{-1}\neq q_{-1}$. Put $$\tilde{\delta}(u,v):=\delta\!\left(\frac{1}{u},\frac{1}{v}\right)\!.$$ We have to explore the properties of $\tilde{\delta}(u,v)$ for $u$ and $v$ sufficiently small. A straightforward computation based on (\[eq:p-diff-P\]) shows that $$\begin{aligned} & & \hskip-1.5em(1-\epsilon^{2}u^{2})(1-\epsilon^{2}v^{2})\tilde{\delta}(u,v)\nonumber \\ & & \hskip-1.5em=\,(u+v+p_{-1}uv)\nonumber \\ & & \quad\times\bigg((1-\epsilon u)(1+\epsilon v)Q\!\left(\frac{u}{1+\epsilon u},\frac{v}{1-\epsilon v}\right)-(1+\epsilon u)(1-\epsilon v)Q\!\left(\frac{u}{1-\epsilon u},\frac{v}{1+\epsilon v}\right)\!\bigg)\nonumber \\ & & -\,(u+v+q_{-1}uv)\nonumber \\ & & \quad\times\bigg((1-\epsilon u)(1+\epsilon v)P\!\left(\frac{u}{1+\epsilon u},\frac{v}{1-\epsilon v}\right)-(1+\epsilon u)(1-\epsilon v)P\!\left(\frac{u}{1-\epsilon u},\frac{v}{1+\epsilon v}\right)\!\bigg)\nonumber \\ & & +\,u^{2}v^{2}\Bigg(P\!\left(\frac{u}{1+\epsilon u},\frac{v}{1-\epsilon v}\right)Q\!\left(\frac{u}{1-\epsilon u},\frac{v}{1+\epsilon v}\right)\nonumber \\ & & \qquad\qquad-\,P\!\left(\frac{u}{1-\epsilon u},\frac{v}{1+\epsilon v}\right)Q\!\left(\frac{u}{1+\epsilon u},\frac{v}{1-\epsilon v}\right)\!\Bigg).\label{eq:delta_tilde}\end{aligned}$$ We shall need the following simple observation. Let $f(u)$ be a holomorphic function in a neighborhood of $0$ (and still assuming that $\epsilon\in\mathbb{C}$, $\epsilon\neq0$). Then the function $$(1+\epsilon u)f\!\left(\frac{u}{1-\epsilon u}\right)-(1-\epsilon u)f\!\left(\frac{u}{1+\epsilon u}\right)$$ vanishes identically in a neighborhood of $0$ if and only if the same is true for $f(u)$. Indeed, suppose that $f(u)$ does not vanishes identically in a neighborhood of $0$ and let $n\in{\mathbb{N}}_{0}$ be the multiplicity of the root of $f(u)$ at $u=0$. Then $f(u)=f_{n}u^{n}+O(u^{n+1})$ as $u\to0$ where $f_{n}\in\mathbb{C}$, $f_{n}\neq0$, and one immediately finds that $$(1+\epsilon u)f\!\left(\frac{u}{1-\epsilon u}\right)-(1-\epsilon u)f\!\left(\frac{u}{1+\epsilon u}\right)=2\epsilon(n+1)f_{n}u^{n+1}+O(u^{n+2}).$$ From (\[eq:delta\_tilde\]) it is seen that $$\begin{aligned} (1-\epsilon^{2}u^{2})(1-\epsilon^{2}v^{2})\tilde{\delta}(u,v) & = & \Bigg(\!(1+\epsilon v)\left(Q_{0}\!\left(\frac{v}{1-\epsilon v}\right)-P_{0}\!\left(\frac{v}{1-\epsilon v}\right)\right)\nonumber \\ & & \quad-\,(1-\epsilon v)\left(Q_{0}\!\left(\frac{v}{1+\epsilon v}\right)-P_{0}\!\left(\frac{v}{1+\epsilon v}\right)\right)\!\Bigg)v\nonumber \\ \noalign{\smallskip} & & +\,\Phi(u,v)u\label{eq:delta_tilde_u}\end{aligned}$$ where $\Phi(u,v)$ is an analytic function for $u$, $v$ belonging to a neighborhood of $0$. Hence, by the above observation, if $P_{0}(y)\neq Q_{0}(y)$ and $v\neq0$ is sufficiently small then the leading coefficient in the asymptotic expansion of the RHS of (\[eq:delta\_tilde\_u\]), as $u\to0$, is nonzero. Consider now the case $P_{0}(y)=Q_{0}(y)\neq0$ ($P(x,y)=Q(x,y)\neq0$) and $p_{-1}\neq q_{-1}$. Then equation (\[eq:delta\_tilde\]) simplifies to $$\begin{aligned} & & \hskip-2em(1-\epsilon^{2}u^{2})(1-\epsilon^{2}v^{2})\tilde{\delta}(u,v)\\ & & \hskip-1em=(p_{-1}-q_{-1})uv\\ & & \times\bigg(\!(1-\epsilon u)(1+\epsilon v)P\!\left(\frac{u}{1+\epsilon u},\frac{v}{1-\epsilon v}\right)-(1+\epsilon u)(1-\epsilon v)P\!\left(\frac{u}{1-\epsilon u},\frac{v}{1+\epsilon v}\right)\!\bigg).\end{aligned}$$ From here it is seen that $$\begin{aligned} & & \hskip-2em(1-\epsilon^{2}u^{2})(1-\epsilon^{2}v^{2})\tilde{\delta}(u,v)\nonumber \\ & & =\,(p_{-1}-q_{-1})\bigg((1+\epsilon v)P_{0}\left(\frac{v}{1-\epsilon v}\right)-(1-\epsilon v)P_{0}\left(\frac{v}{1+\epsilon v}\right)\!\bigg)vu+\Psi(u,v)u^{2}\nonumber \\ \label{eq:delta_tilde0_u}\end{aligned}$$ where $\Psi(u,v)$ is an analytic function for $u$, $v$ belonging to a neighborhood of $0$. Referring again to the above observation we conclude that the leading coefficient in the asymptotic expansion of the RHS of (\[eq:delta\_tilde0\_u\]), as $u\to0$, is nonzero for all sufficiently small $v\neq0$. Proof of the main theorem {#sec:proofs} ========================= General equations ----------------- Consider semi-infinite matrices $\mathcal{J}$ and $\mathcal{H}$ indexed by $m,n\in{\mathbb{N}}_{0}$ where $\mathcal{J}$ is a Jacobi matrix determined by two complex sequences $\{\alpha_{n}\}$ and $\{\beta_{n}\}$, with $\alpha_{n}\neq0$ for all $n\geq0$, and $\mathcal{H}$ is a Hankel matrix . It is convenient and common to set $\alpha_{-1}:=0$. Matrices $\mathcal{H}$ and $\mathcal{J}$ commute if and only if it holds true that $$(\alpha_{n}-\alpha_{m})h_{n+m+1}+(\beta_{n}-\beta_{m})h_{n+m}+(\alpha_{n-1}-\alpha_{m-1})h_{n+m-1}=0,\label{eq:a_b_H}$$ for all $m,n\in{\mathbb{N}}_{0}$ ($h_{-1}$ is arbitrary). In particular, letting $m=0$ we have $$(\alpha_{n}-\alpha_{0})h_{n+1}+(\beta_{n}-\beta_{0})h_{n}+\alpha_{n-1}h_{n-1}=0,\label{eq:H_descend}$$ for all $n\in{\mathbb{N}}$. Taking into account the descending recurrence it is clear that, for any $n\in{\mathbb{N}}_{0}$, if $h_{n}=h_{n+1}=0$, then $h_{0}=h_{1}=\ldots=h_{n}=h_{n+1}=0$. Wilson ------ We shall need the asymptotic expansions of $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$, $$\alpha_{n}^{\text{W}}=\frac{n^{2}}{4}+\frac{sn}{4}+A_{0}+O\!\left(\frac{1}{n}\right)\!,\quad\beta_{n}^{\text{W}}=\frac{n^{2}}{2}+\frac{(s-1)n}{2}+B_{0}+O\!\left(\frac{1}{n}\right)\text{ }\ \text{as}\ n\to\infty,\label{eq:alp_bet_asympt_wilson}$$ where $s:=a+b+c+d$ and $A_{0}$, $B_{0}$ are constants. Explicitly, $$\begin{aligned} A_{0}= & \frac{1}{32}\,(-3+4a+4b+4c+4d-2a^{2}-2b^{2}-2c^{2}-2d^{2}+4ab+4ac+4ad+4bc\\ & \hskip1.5em+4bd+4cd),\\ B_{0}= & \frac{1}{8}\,(-a^{2}-b^{2}-c^{2}-d^{2}+2ab+2ac+2ad+2bc+2bd+2cd).\end{aligned}$$ The following proposition is in fact an implication stated in Theorem \[thm:main\] ad (i). Its proof is more straightforward than that for the opposite direction. \[prop:polyn\_impl\_nontriv\_wilson\] If $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ depend both polynomially on $n$, then ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})$ is nontrivial if and only if $\alpha_{-1}^{\text{W}}=0$. If so, ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})={\mathop\mathrm{span}\nolimits}\left(\mathcal{H}^{(1)},\mathcal{H}^{(2)}\right)$, where $\mathcal{H}^{(1)}$ and $\mathcal{H}^{(2)}$ are Hankel matrices determined by the sequences $$h_{k}^{(1)}=\frac{(-1)^{k}}{k+s-1}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k},\quad k\in{\mathbb{N}}_{0},\label{eq:sol_h_wilson}$$ respectively. Suppose $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ are polynomials in $n$. Then the form of the polynomials is seen from the asymptotic formulas in which the Landau symbol becomes zero. Plugging these expressions into and canceling the common term $(n-m)$ one arrives at a recurrence equation for the entries of the Hankel matrix $\mathcal{H}$ in which the indices occur in the combination $m+n$ only. Writing $k$ instead of $m+n$, the recurrence reads $$(k+s)h_{k+1}+2(k+s-1)h_{k}+(k+s-2)h_{k-1}=0,\quad k\in{\mathbb{N}}.$$ Its two linearly independent solutions are given by . Consequently, any $\mathcal{H}\in{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})$ has to be a linear combination of $\mathcal{H}^{(1)}$ and $\mathcal{H}^{(2)}$. Moreover, it is easy to see that any such $\mathcal{H}$, if nontrivial, can commute with $\mathcal{J}^{\text{W}}$ if and only if $\alpha_{-1}^{\text{W}}=0$. In order to complete the proof of Theorem \[thm:main\] ad (i) we have to prove the implication opposite to that stated in Proposition \[prop:polyn\_impl\_nontriv\_wilson\]. To this end, we need the following equivalent condition for a polynomial dependence of $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ on $n$ to hold true. \[lem:alp\_bet\_polyn\_wilson\] The following two statements are equivalent: 1. $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ depend both polynomially on $n\in\mathbb{N}_{0}$. 2. There exists a constant $\omega$ such that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\omega=0,\quad\forall n\in\mathbb{N}.\label{eq:beta_alpha_alpha_wilson}$$ Moreover, statement (2) can be true only if $\omega=1/16$. If $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ depend both polynomially on $n$ then the form of the polynomials can be deduced from the asymptotic formulas in which the Landau symbols vanish identically. The implication (1)$\Rightarrow$(2) is then a matter of a straightforward calculation showing that (\[eq:beta\_alpha\_alpha\_wilson\]) holds provided we let $\omega=1/16$. The opposite implication (2)$\Rightarrow$(1) is more tedious. From one finds that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\omega=\omega-\frac{1}{16}\,+O\!\left(\frac{1}{n}\right)\text{ }\ \text{as}\ n\to\infty.$$ Hence necessarily $\omega=1/16$. Further we deduce from some necessary conditions on the parameters $a,b,c,d$. The equation implies that $$\left(\left(\beta_{n}^{\text{W}}+\omega\right)^{2}-(\alpha_{n}^{\text{W}})^{2}-(\alpha_{n-1}^{\text{W}})^{2}\right)^{\!2}-4(\alpha_{n}^{\text{W}})^{2}(\alpha_{n-1}^{\text{W}})^{2}=0,\quad\forall n\in{\mathbb{N}}.\label{eq:A_C_id_wilson_inproof}$$ We temporarily denote by $f(n)$ the left-hand side of . From , and it is seen that $f(n)$ is a rational function of $n$. Since $f(n)$ has infinitely many roots it vanishes identically as a complex function. This is equivalent to saying that the rational function $f(n)$ has no poles and vanishes at infinity. Let us check possible poles of $f(n)$. From (\[eq:A\_C\_id\_wilson\_inproof\]) and , , it is clear that the poles can occur only at the points $z_{j}=(4-j-s)/2$, $j=1,2,3,4,5$. Here we still denote $s=a+b+c+d$. As a necessary condition, we require that, looking at the Laurent expansion of $f(n)$ around each isolated singularity $z_{j}$, the coefficient corresponding to the highest possible order of the pole that can occur in the expansion vanishes. Doing so we shall confine ourselves only to points $z_{1}$ and $z_{2}$, however, since it turns out that inspection of the remaining points does not lead to additional constraints. As far as the singular point $z_{1}=(3-s)/2$ is concerned, this singularity may occur only in the term $(\alpha_{n-1}^{\text{W}})^{2}$, and it does not occur in $\beta_{n}^{\text{W}}$ and $(\alpha_{n}^{\text{W}})^{2}$. By inspection of the left-hand side of one readily obtains a condition on the residue $${\mathop\mathrm{Res}\nolimits}_{n=z_{1}}(\alpha_{n-1}^{\text{W}})^{2}=0,\ \text{equivalently,\ }{\mathop\mathrm{Res}\nolimits}_{n=z_{3}}(\alpha_{n}^{\text{W}})^{2}=0.$$ Using the explicit definition of $(\alpha_{n}^{\text{W}})^{2}$ it is straightforward to derive that the residue is zero if and only if $$\left((a+b-c-d)^{2}-1\right)\left((a-b+c-d)^{2}-1\right)\left((a-b-c+d)^{2}-1\right)(s-3)(s-1)=0.\label{eq:1st_cond_wilson_inproof}$$ As for the singular point $z_{2}=(2-s)/2$, this pole may occurred in $\beta_{n}^{\text{W}}$ with order $1$ and in $(\alpha_{n-1}^{\text{W}})^{2}$ with order $2$, but not in $(\alpha_{n}^{\text{W}})^{2}$. An inspection of the left-hand side of shows that the coefficient standing at $(n-z_{2})^{-2}$ in the Laurent expansion around $z_{2}$ of $(\beta_{n}^{\text{W}})^{2}-(\alpha_{n-1}^{\text{W}})^{2}$ must be zero. As one can check, ${\mathop\mathrm{Res}\nolimits}_{n=z_{2}}A_{n-1}^{\text{W}}=-{\mathop\mathrm{Res}\nolimits}_{n=z_{2}}C_{n}^{\text{W}}$ and therefore this condition reduces to $${\mathop\mathrm{Res}\nolimits}_{n=z_{2}}C_{n}^{\text{W}}=0.$$ This is true if and only if $$(a+b-c-d)(a-b+c-d)(a-b-c+d)(s-2)=0.\label{eq:2st_cond_wilson_inproof}$$ Equations and imply a finite number of constrains on the parameters $a,b,c,d$ which can be discussed case by case. Given that $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ are symmetric in $a,b,c,d$ it suffices to consider the following four cases covering all possibilities up to a permutation of the parameters: (i) $a+b+c+d=1$, $a+b-c-d=0$, (ii) $a+b+c+d=3$, $a+b-c-d=0$, (iii) $a+b-c-d=1$, $a+b+c+d=2$, (iv) $a+b-c-d=1$, $a-b+c-d=0$.\ In order to complete the discussion successfully we will also take into account the asymptotic behavior of $\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+1/16$ at infinity. In case (i) we have $b=1/2-a$, $d=1/2-c$. Then a computation shows that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\frac{1}{16}=\frac{(4a-1)^{2}(4c-1)^{2}}{256}\,\frac{1}{n^{2}}+O\!\left(\frac{1}{n^{3}}\right)\ \text{as}\ n\to\infty.$$ Hence we have to put either $a=1/4$ or $c=1/4$ but the two choices differ just by a permutation of parameters. We obtain $$\begin{aligned} \alpha_{n}^{\text{W}}\!\left(a,\frac{1}{2}-a,\frac{1}{4},\frac{1}{4}\right) & =\frac{1}{64}\,(4\,n+4\,a+1)(4\,n-4\,a+3),\\ \beta_{n}^{\text{W}}\!\left(a,\frac{1}{2}-a,\frac{1}{4},\frac{1}{4}\right) & =\frac{1}{32}\,(16n^{2}-16a^{2}+8a+1).\end{aligned}$$ In case (ii) we have $b=3/2-a$, $d=3/2-c$. Then a computation shows that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\frac{1}{16}=\frac{(4a-3)^{2}(4c-3)^{2}}{256}\,\frac{1}{n^{2}}+O\!\left(\frac{1}{n^{3}}\right)\ \text{as}\ n\to\infty.$$ Hence we have to put either $a=3/4$ or $c=3/4$ but again the two choices differ by a permutation of parameters. We obtain $$\begin{aligned} \alpha_{n}^{\text{W}}\!\left(a,\frac{3}{2}-a,\frac{3}{4},\frac{3}{4}\right) & =\frac{1}{64}\,(4\,n+4\,a+3)(4\,n-4\,a+9),\\ \beta_{n}^{\text{W}}\!\left(a,\frac{3}{2}-a,\frac{3}{4},\frac{3}{4}\right) & =\frac{1}{32}\,(16n^{2}+32n-16a^{2}+24a+9).\end{aligned}$$ In case (iii) we have $b=3/2-a$, $d=1/2-c$. Then a computation shows that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\frac{1}{16}=\frac{(4a-3)^{2}(4c-1)^{2}}{256}\,\frac{1}{n^{2}}+O\!\left(\frac{1}{n^{3}}\right)\ \text{as}\ n\to\infty.$$ Now we have to distinguish two sub-cases. In sub-case (iiia) we let $a=3/4$, $b=3/4$, $d=1/2-c$, and we obtain $$\begin{aligned} \alpha_{n}^{\text{W}}\!\left(\frac{3}{4},\frac{3}{4},c,\frac{1}{2}-c\right) & =\frac{1}{64}\,(4n-4c+5)(4n+4c+3),\\ \beta_{n}^{\text{W}}\!\left(\frac{3}{4},\frac{3}{4},c,\frac{1}{2}-c\right) & =\frac{1}{32}\,(16n^{2}+16n-16c^{2}+8c+5).\end{aligned}$$ In sub-case (iiib) we let $b=3/2-a$, $c=1/4$, $d=1/4$, and we obtain $$\begin{aligned} \alpha_{n}^{\text{W}}\!\left(a,\frac{3}{2}-a,\frac{1}{4},\frac{1}{4}\right) & =\frac{1}{64}\,(4n-4a+7)(4n+4a+1),\\ \beta_{n}^{\text{W}}\!\left(a,\frac{3}{2}-a,\frac{1}{4},\frac{1}{4}\right) & =\frac{1}{32}\,(16n^{2}+16n-16a^{2}+24a-3).\end{aligned}$$ In case (iv) we have $d=a-1/2$, $c=b-1/2$. Then a computation shows that $$\beta_{n}^{\text{W}}-\alpha_{n}^{\text{W}}-\alpha_{n-1}^{\text{W}}+\frac{1}{16}=\frac{(4a-3)^{2}(4b-3)^{2}}{256}\,\frac{1}{n^{2}}+O\!\left(\frac{1}{n^{3}}\right)\ \text{as}\ n\to\infty.$$ Hence we have to put either $a=3/4$ or $b=3/4$ but again the two choices differ by a permutation of parameters. We obtain $$\begin{aligned} \alpha_{n}^{\text{W}}\!\left(\frac{3}{4},b,b-\frac{1}{2},\frac{1}{4}\right) & =\frac{1}{8}\,(n+1)(2n+4b-1),\\ \beta_{n}^{\text{W}}\!\left(\frac{3}{4},b,b-\frac{1}{2},\frac{1}{4}\right) & =\frac{1}{16}\big(8n^{2}+4(4b-1)n+8b-3\big).\end{aligned}$$ We conclude that the four discussed cases show that whenever (\[eq:beta\_alpha\_alpha\_wilson\]) holds then $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ are polynomials in $n$. \[eq:alp\_bet\_polyn\_param\_wilson\] As a matter of fact, in the second part of the proof of Lemma \[lem:alp\_bet\_polyn\_wilson\] we have found all configurations of the parameters, up to a permutation, when $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ are polynomials in $n$. But not in all found cases the requirement is met that the polynomial for $\alpha_{n}^{\text{W}}$ vanishes at $n=-1$. Imposing in addition this requirement we are left with the following admissible configurations of the parameters $(a,b,c,d)$, again up to a permutation: (i) $(-1/4,3/4,1/4,1/4,)$, (ii) $(1/4,5/4,3/4,3/4)$, (iii) $(3/4,3/4,1/4,1/4)$, (iv) $(3/4,b,b-1/2,1/4)$.\ But (i), (ii) and (iii) are particular cases of (iv), possibly up to a permutation: (i) for $b=1/4$, (ii) for $b=5/4$, (iii) for $b=3/4$. Moreover, taking the constrains on the parameters $a,b,c,d$ given below equation into account, case (iv) is admissible only if $b>1/4$. By writing $b=1/4+t/2$, for $t>0$, we obtain the parametrization used in claim (i) of Theorem \[thm:main\]. Now we are ready to prove the opposite implication of Theorem \[thm:main\]. \[prop:nontriv\_impl\_polyn\_wilson\] If ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})$ is nontrivial, then $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ depend polynomially on $n$ and $\alpha_{-1}^{\text{W}}=0$. We shall proceed by contradiction. Remember, however, as observed in Proposition \[prop:polyn\_impl\_nontriv\_wilson\], if $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ depend both polynomially on $n$ but $\alpha_{-1}^{\text{W}}\neq0$ then ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})$ is trivial. Hence to get a contradiction we assume that ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{W}})$ is nontrivial and either $\alpha_{n}^{\text{W}}$ or $\beta_{n}^{\text{W}}$ is not a polynomial in $n$. Furthermore, without loss of generality we can assume that entries $h_{n}$ of a nonzero Hankel matrix $\mathcal{H}$ commuting with $\mathcal{J}^{\text{W}}$ are real. We shall also make use of the fact that $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$ may be both regarded as analytic functions in $n$ for $n$ sufficiently large. Along with (\[eq:a\_b\_H\]) we consider the equation $$(\alpha_{n-1}-\alpha_{m+1})h_{n+m+1}+(\beta_{n-1}-\beta_{m+1})h_{n+m}+(\alpha_{n-2}-\alpha_{m})h_{n+m-1}=0.\label{eq:Hprime}$$ Let again $s:=a+b+c+d$. From the asymptotic behavior $$\alpha_{n}^{\text{W}}=\frac{n^{2}}{4}+\frac{sn}{4}+O(1),\;\beta_{n}^{\text{W}}=\frac{n^{2}}{2}+\frac{(s-1)n}{2}+O(1),\;\alpha_{n-1}^{\text{W}}=\frac{n^{2}}{4}+\frac{(s-2)n}{4}+O(1),$$ as $n\to\infty$, it is clear that none of the sets of functions in $n$, $\{\alpha_{n}^{\text{W}},\beta_{n}^{\text{W}},1\}$ or $\{\beta_{n}^{\text{W}},\alpha_{n-1}^{\text{W}},1\}$ or $\{\alpha_{n}^{\text{W}},\alpha_{n-1}^{\text{W}},1\}$, is linearly dependent. Let $$\begin{aligned} \delta_{1}(n,m) & :=\det\!\left(\begin{array}{cc} \beta_{n}^{\text{W}}-\beta_{m}^{\text{W}} & \alpha_{n-1}^{\text{W}}-\alpha_{m-1}^{\text{W}}\\ \beta_{n-1}^{\text{W}}-\beta_{m+1}^{\text{W}} & \alpha_{n-2}^{\text{W}}-\alpha_{m}^{\text{W}} \end{array}\right)\!,\nonumber \\ \delta_{2}(n,m) & :=-\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{W}}-\alpha_{m}^{\text{W}} & \alpha_{n-1}^{\text{W}}-\alpha_{m-1}^{\text{W}}\\ \alpha_{n-1}^{\text{W}}-\alpha_{m+1}^{\text{W}} & \alpha_{n-2}^{\text{W}}-\alpha_{m}^{\text{W}} \end{array}\right)\!,\label{eq:def_delta123}\\ \delta_{3}(n,m) & :=\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{W}}-\alpha_{m}^{\text{W}} & \beta_{n}^{\text{W}}-\beta_{m}^{\text{W}}\\ \alpha_{n-1}^{\text{W}}-\alpha_{m+1}^{\text{W}} & \beta_{n-1}^{\text{W}}-\beta_{m+1}^{\text{W}} \end{array}\right)\!.\nonumber \end{aligned}$$ According to Lemma \[lem:Mzw\], for all $m$ sufficiently large there exists $R_{m}\in\mathbb{N}$ such that for all $n\geq R_{m}$, $\delta_{1}(n,m)\neq0$ and $\delta_{3}(n,m)\neq0$. Then, by equations (\[eq:a\_b\_H\]) and (\[eq:Hprime\]), the vectors $$(h_{n+m+1},h_{n+m},h_{n+m-1})\ \ \text{and}\ \ \big(\delta_{1}(n,m),\delta_{2}(n,m),\delta_{3}(n,m)\big)\label{eq:vecs_h_delta}$$ are linearly dependent. We can assume that $R_{m}$ is so large that $\alpha_{n}^{\text{W}}-\alpha_{m}^{\text{W}}>0$ for all $n\geq R_{m}$. Referring to (\[eq:a\_b\_H\]) and (\[eq:H\_descend\]), one can see that the former of the two vectors in (\[eq:vecs\_h\_delta\]) is necessarily nonzero, too. Otherwise $h_{n}=0$ identically. Hence $\delta_{3}(n,m)\neq0$ implies that $h_{n+m-1}\neq0$ for $n\geq R_{m}$. This in turn implies that $\delta_{2}(n,m)\neq0$ for $n\geq R_{m}$. Note that, consequently, $\alpha_{n}^{\text{W}}$ cannot be a polynomial since otherwise $\delta_{2}(n,m)=0$ identically. Fix sufficiently large $m\in{\mathbb{N}}_{0}$. Then for all $n\in{\mathbb{N}}_{0}$, $n\geq m_{0}:=R_{m}+m$, it we have $$h_{n+1}=\psi(n)h_{n},\ \,\text{with}\ \text{ }\psi(n):=\frac{\delta_{1}(n-m,m)}{\delta_{2}(n-m,m)}\,.$$ It is of importance that $\psi(n)$ can be regarded as a meromorphic function of $n$ in a neighborhood of $\infty$. Particularly, $\psi(n)$ has an asymptotic expansion to all orders as $n\to\infty$. In view of Lemma \[lem:Mzw\], there are only three possible types of asymptotic behavior of $\psi(n)$ as $n\to\infty$: $$\begin{aligned} (\text{I}) & & \psi(n)=\lambda_{1}\!\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!,\\ (\text{II}) & & \psi(n)=\lambda_{2}\,n\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!,\\ (\text{III}) & & \psi(n)=\frac{\lambda_{3}}{n}\!\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!,\end{aligned}$$ where $\lambda_{j}\neq0$ for $j=1,2,3$. From here one can deduce the asymptotic behavior of $$h_{n}=h_{m_{0}}\,\prod_{k=m_{0}}^{n-1}\psi(k).$$ In case (I) we have $$h_{n}=c_{1}\lambda_{1}^{\,n}\,n^{\sigma_{1}}\left(1+O\!\left(\frac{1}{n}\right)\!\right)\text{ }\text{as}\ n\to\infty,$$ for some $c_{1},\sigma_{1}\in\mathbb{R}$, $c_{1}\neq0$. In case (II) we have $$h_{n}=c_{2}\lambda_{2}^{\,n}\,n!\,n^{\sigma_{2}}\left(1+O\!\left(\frac{1}{n}\right)\!\right)\text{ }\text{as}\ n\to\infty,$$ for some $c_{2},\sigma_{2}\in\mathbb{R}$, $c_{2}\neq0$. In case (III) we have $$h_{n}=\frac{c_{3}\lambda_{3}^{\,n}\,n^{\sigma_{3}}}{n!}\left(1+O\!\left(\frac{1}{n}\right)\!\right)\text{ }\text{as}\ n\to\infty,$$ for some $c_{3},\sigma_{3}\in\mathbb{R}$, $c_{3}\neq0$. Rewriting (\[eq:a\_b\_H\]) and taking into the account the asymptotic behavior of $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$, we obtain $$\begin{aligned} & h_{n+1}+\frac{\beta_{n-m}^{\text{W}}-\beta_{m}^{\text{W}}}{\alpha_{n-m}^{\text{W}}-\alpha_{m}^{\text{W}}}h_{n}+\frac{\alpha_{n-m-1}^{\text{W}}-\alpha_{m-1}^{\text{W}}}{\alpha_{n-m}^{\text{W}}-\alpha_{m}^{\text{W}}}h_{n-1}\nonumber \\ & =h_{n+1}+2\!\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!h_{n}+\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!h_{n-1}=0.\label{eq:H_eq_asympt}\end{aligned}$$ It is readily seen that the asymptotic behavior of $h_{n}$ of type (II) and (III) is incompatible with (\[eq:H\_eq\_asympt\]). Hence the only admissible asymptotic behavior of $h_{n}$ is that of type (I). Without loss of generality we can suppose that $c_{1}=1$. Moreover, from (\[eq:H\_eq\_asympt\]) it is also seen that $\lambda_{1}$ should solve the equation $\lambda_{1}^{\,2}+2\lambda_{1}+1=0$ whence $\lambda_{1}=-1$. Furthermore, for the sake of simplicity we will drop the index in $\sigma_{1}$. Thus we obtain $$h_{n}=(-1)^{n}(n+1)^{\sigma}\varphi(n)\label{eq:H_subst}$$ where $$\varphi(n)=1+O\!\left(\frac{1}{n}\right)\ \ \text{as}\ n\to\infty.\label{eq:phi_asympt}$$ As a matter of fact, $\varphi(n)$ has an asymptotic expansion to all orders as $n\to\infty$. Plugging (\[eq:H\_subst\]) into (\[eq:a\_b\_H\]) we obtain $$\begin{aligned} & & \hskip-5em(\alpha_{n}^{\text{W}}-\alpha_{m}^{\text{W}})\!\left(1+\frac{1}{n+1}\right)^{\!\sigma}\varphi(n+1)-(\beta_{n}^{\text{W}}-\beta_{m}^{\text{W}})\varphi(n)\nonumber \\ & & \hskip8em+\,(\alpha_{n-1}^{\text{W}}-\alpha_{m-1}^{\text{W}})\!\left(1-\frac{1}{n+1}\right)^{\!\sigma}\varphi(n-1)=0.\label{eq:H_phi}\end{aligned}$$ The asymptotic expansion of the LHS of (\[eq:H\_phi\]) as $n\to\infty$, with $m$ being fixed but otherwise arbitrary, while taking into account (\[eq:phi\_asympt\]) and and substituting for $s$, $A_{0}$, $B_{0}$, yields the expression $$\beta_{m}^{\text{W}}-\alpha_{m}^{\text{W}}-\alpha_{m-1}^{\text{W}}+\frac{\sigma(\sigma+1)}{4}+\frac{1}{16}+O\!\left(\frac{1}{n}\right)\!.$$ Hence $$\beta_{m}^{\text{W}}-\alpha_{m}^{\text{W}}-\alpha_{m-1}^{\text{W}}+\frac{\sigma(\sigma+1)}{4}+\frac{1}{16}=0,\text{ }\ \text{for}\ \text{all}\ m\in{\mathbb{N}}_{0}.$$ Lemma \[lem:alp\_bet\_polyn\_wilson\] implies that $\beta_{n}^{\text{W}}$ and $\alpha_{n}^{\text{W}}$ depend both polynomially on $n$, a contradiction. In summary, as far as the Jacobi matrix $\mathcal{J}^{W}$ is concerned, the equivalence stated in Theorem \[thm:main\] ad (i) means two implications which are established by Propositions \[prop:polyn\_impl\_nontriv\_wilson\] and \[prop:nontriv\_impl\_polyn\_wilson\]. Along with Remark \[eq:alp\_bet\_polyn\_param\_wilson\] this concludes the proof of Theorem \[thm:main\] ad (i). Continuous dual Hahn -------------------- The approach used in the case of Wilson polynomials can be also applied in the case of Continuous dual Hahn polynomials with no essential difference. Due to the considerably simpler form of the Jacobi parameters $\alpha_{n}^{\text{CdH}}$ and $\beta_{n}^{\text{CdH}}$ (but the same order of the asymptotic behavior for $n\to\infty$), some parts of the proof simplify significantly. Observe that $\beta_{n}^{\text{CdH}}$ is a polynomial in $n$ for all values of the parameters $a,b,c$. Namely, we have $$\beta_{n}^{\text{CdH}}=2n^{2}+(2s-1)n+\tilde{s},\label{eq:bet_asympt_cdhahn}$$ where we denote $s:=a+b+c$ and $\tilde{s}:=ab+ac+bc$. In addition, $$\alpha_{n}^{\text{CdH}}=n^{2}+\frac{2s+1}{2}\,n+\frac{4s+4\tilde{s}-1}{8}+O\!\left(\frac{1}{n}\right)\!,\,\ \text{as}\ n\to\infty.\label{eq:alp_asympt_cdhahn}$$ \[prop:polyn\_impl\_nontriv\_cdhahn\] If $\alpha_{n}^{\text{CdH}}$ depends polynomially on $n$, then ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CdH}})$ is nontrivial. More precisely, if that is the case, ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CdH}})={\mathop\mathrm{span}\nolimits}\left(\mathcal{H}^{(1)},\mathcal{H}^{(2)}\right)$, where $\mathcal{H}^{(1)}$ and $\mathcal{H}^{(2)}$ are Hankel matrices determined by the sequences $$h_{k}^{(1)}=\frac{(-1)^{k}}{k+s-1/2}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k},\quad k\in{\mathbb{N}}_{0},\label{eq:sol_h_cdhahn}$$ respectively. Note that $\alpha_{n}^{\text{CdH}}$ depends polynomially on $n$ if and only if two of the parameters $a,b,c$ are equal to $1/2$. In this case, $\alpha_{-1}^{\text{CdH}}=0$ automatically. Assuming that $\alpha_{n}^{\text{CdH}}$ is a polynomial in $n$, the Landau symbol in is identically vanishing. Plugging this expression together with into , one infers that any $\mathcal{H}\in{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CdH}})$, with $\mathcal{H}_{m,n}=h_{m+n}$, if and only if $\{h_{n}\}$ is a solution of the three-term recurrence $$(k+s+1/2)h_{s+1}+(2k+2s-1)h_{k}+(k+s-3/2)h_{k-1}=0,\quad k\in{\mathbb{N}}.$$ The two linearly independent solutions of the above recurrence are given by and the proof follows. \[lem:alp\_bet\_polyn\_cdhahn\] The following two statements are equivalent: 1. $\alpha_{n}^{\text{CdH}}$ depends polynomially on $n$. 2. There exists a constant $\omega$ such that $$\beta_{n}^{\text{CdH}}-\alpha_{n}^{\text{CdH}}-\alpha_{n-1}^{\text{CdH}}+\omega=0,\quad\forall n\in{\mathbb{N}}_{0}.\label{eq:beta_alpha_alpha_cdhahn}$$ Moreover, statement (2) can be true only if $\omega=1/4$. The implication (1)$\Rightarrow$(2): If $\alpha_{n}$ is a polynomial in $n$ then, up to a permutation of the parameters, $a=b=1/2$, whence $$\alpha_{n}^{\text{CdH}}=(n+1)\!\left(n+\frac{1}{2}+c\right)$$ and $$\beta_{n}^{\text{CdH}}=n\!\left(n-\frac{1}{2}+c\right)+(n+1)\!\left(n+\frac{1}{2}+c\right)-\frac{1}{4}=\alpha_{n-1}+\alpha_{n}-\frac{1}{4}\,.$$ The implication (2)$\Rightarrow$(1): The equation means that the sequence $\{\alpha_{n}^{\text{CdH}}\}$ obeys a two-term recurrence with a right-hand side, $$\alpha_{n}^{\text{CdH}}+\alpha_{n-1}^{\text{CdH}}=\phi_{n},$$ where $\phi_{n}:=\beta_{n}^{\text{CdH}}+\omega$ is a polynomial in $n$ of degree $2$ with the leading coefficient equal to $2$. It follows that $$\alpha_{n+2}^{\text{CdH}}-\alpha_{n+1}^{\text{CdH}}-\alpha_{n}^{\text{CdH}}+\alpha_{n-1}^{\text{CdH}}=\phi_{n+2}-2\phi_{n+1}+\phi_{n}=4.$$ Hence $\alpha_{n}^{\text{CdH}}$ must be of the form $$\alpha_{n}^{\text{CdH}}=A+Bn+C(-1)^{n}+n^{2}$$ where $A$, $B$, $C$ are some constants. At the same time, we know that the square of $\alpha_{n}^{\text{CdH}}$ is a polynomial in $n$ of degree $4$. Whence $$C(A+Bn+n^{2})\!=0,\,\quad\forall n\in\mathbb{N}_{0}.$$ Necessarily, $C=0$ and so $\alpha_{n}^{\text{CdH}}=A+Bn+n^{2}$. \[prop:nontriv\_impl\_polyn\_cdhahn\] If ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CdH}})$ is nontrivial, then $\alpha_{n}^{\text{CdH}}$ depends polynomially on $n$. The proof is based on Lemma \[lem:Mzw\] and follows the same steps as the proof of Proposition \[prop:nontriv\_impl\_polyn\_wilson\]. Therefore we just briefly address several points specific for this case. For a contradiction, we suppose that a nonzero Hankel matrix $\mathcal{H}$ commuting with $\mathcal{J}^{\text{CdH}}$ exists but $\alpha_{n}^{\text{CdH}}$ is not a polynomial in $n$. It follows from that $$\alpha_{n-1}^{\text{CdH}}=n^{2}+\frac{2s-3}{2}\,n-\frac{4s-4\tilde{s}-3}{8}+O\!\left(\frac{1}{n}\right),\,\ \text{as}\ n\to\infty.$$ This together with and implies that none of the sets of functions in $n$, $\{\alpha_{n},\beta_{n},1\}$ or $\{\beta_{n},\alpha_{n-1},1\}$ or $\{\alpha_{n},\alpha_{n-1},1\}$, is linearly dependent. Now, using exactly the same reasoning as in the proof of Proposition \[prop:nontriv\_impl\_polyn\_wilson\] while considering $\alpha_{n}^{\text{CdH}}$ and $\beta_{n}^{\text{CdH}}$ instead of $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$, one can argue that is necessarily true. By Lemma \[lem:alp\_bet\_polyn\_cdhahn\], this means a contradiction. Continuous Hahn --------------- One has, as $n\to\infty$, $$\alpha_{n}^{\text{CH}}=\frac{n}{4}+\frac{a+b+c+d}{8}+O\!\left(\frac{1}{n}\right)\quad\mbox{ and }\quad\beta_{n}^{\text{CH}}=\frac{{{\rm i}}(a+b-c-d)}{4}+O\!\left(\frac{1}{n}\right)\!.\label{eq:alp_bet_asympt_chahn}$$ \[prop:polyn\_impl\_nontriv\_chahn\] If $\alpha_{n}^{\text{CH}}$ and $\beta_{n}^{\text{CH}}$ depend both polynomially on $n$, then ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})$ is nontrivial if and only if $\alpha_{-1}^{\text{CH}}=0$. If so, ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})={\mathop\mathrm{span}\nolimits}\left(\mathcal{H}^{(1)},\mathcal{H}^{(2)}\right)$, where $\mathcal{H}^{(1)}$ and $\mathcal{H}^{(2)}$ are Hankel matrices determined by the sequences $$h_{k}^{(1)}=\sin\frac{k\pi}{2}\quad\mbox{ and }\quad h_{k}^{(2)}=\cos\frac{k\pi}{2},\quad k\in{\mathbb{N}}_{0},\label{eq:sol_h_chahn}$$ respectively. The proof is again analogous to the proof of Proposition \[prop:polyn\_impl\_nontriv\_wilson\]. Assuming that $\alpha_{n}^{\text{CH}}$ and $\beta_{n}^{\text{CH}}$ are polynomials, we can deduce the form of the polynomials from , and the general equation implies that any $\mathcal{H}\in{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})$, with entries $\mathcal{H}_{m,n}=h_{m+n}$, must fulfill the simple recurrence $$h_{k+1}+h_{k-1}=0,\quad k\in{\mathbb{N}},$$ whose two linearly independent solutions are given by . Now it suffices to notice that such $\mathcal{H}$ is necessarily trivial if $\alpha_{-1}^{\text{CH}}\neq0$. On the contrary, if $\alpha_{-1}^{\text{CH}}=0$ then any solution of the above recurrence defines a Hankel matrix which commutes with $\mathcal{J}^{\text{CH}}$. \[prop:nontriv\_impl\_polyn\_chahn\] If ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})$ is nontrivial then $\alpha_{n}^{\text{CH}}$ depends polynomially on $n$. Let us proceed by contradiction and assume that $\alpha_{n}^{\text{CH}}$ is not a polynomial in $n$ but (\[eq:a\_b\_H\]) has a nonzero solution $h_{n}$. Without loss of generality we can suppose $h_{n}$ to be real. On a neighborhood of infinity we can expand $$\alpha_{n}^{\text{CH}}=\frac{n}{4}+A_{0}+A(n),\,\ \text{where}\ A(n)=\sum_{k=1}^{\infty}\frac{A_{k}}{n^{k}}\,.$$ By our assumption, $A(n)$ cannot vanish identically. Keeping notation , with $\alpha_{n}^{\text{CH}}$ and $\beta_{n}^{\text{CH}}$ instead of $\alpha_{n}^{\text{W}}$ and $\beta_{n}^{\text{W}}$, we have, for a fixed $m$, $$\begin{aligned} \delta_{2}(n,m) & := & -\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{CH}}-\alpha_{m}^{\text{CH}} & \alpha_{n-1}^{\text{CH}}-\alpha_{m-1}^{\text{CH}}\\ \alpha_{n-1}^{\text{CH}}-\alpha_{m+1}^{\text{CH}} & \alpha_{n-2}^{\text{CH}}-\alpha_{m}^{\text{CH}} \end{array}\right)\\ & = & \frac{A(m+1)-2A(m)+A(m-1)}{4}\,n+O(1),\ \ \text{as}\ n\to\infty.\end{aligned}$$ If $A(n)=A_{k}n^{-k}+O(n^{-k-1})$, with $A_{k}\neq0$, then $$A(m+1)-2A(m)+A(m-1)=A_{k}k(k+1)m^{-k-2}+O(m^{-k-3}),\,\ \text{as}\ m\to\infty.$$ Hence for all sufficiently large $m$ there exists $$\lim_{n\to\infty}\frac{\delta_{2}(n,m)}{n}=:C_{m}\in\mathbb{R\setminus}\{0\}.\label{eq:lim_delta2_chahn}$$ In particular, there exists $R_{m}\geq0$ such that for all $n\geq R_{m}$, $\delta_{2}(n,m)\neq0$. One can argue, similarly as in the proof of Proposition \[prop:nontriv\_impl\_polyn\_wilson\], that the vectors (\[eq:vecs\_h\_delta\]) are linearly dependent. Moreover, for $m$ sufficiently large and $n\geq R_{m}$ both of them are nonzero (note that $R_{m}$ can be chosen large enough so that the coefficient $\alpha_{n}-\alpha_{m}$ occurring in (\[eq:a\_b\_H\]) is nonzero for all $n\geq R_{m}$). It follows that $\beta_{n}^{\text{CH}}$ cannot be a polynomial since if so, $\beta_{n}^{\text{CH}}$ would be a constant as seen from , and then $\delta_{1}(n,m)=\delta_{3}(n,m)=0$ identically. Consequently, $h_{n}$ would vanish for all sufficiently large and hence for all $n$. Thus, we can expand $\beta_{n}$ as a function of $n$ near infinity as $$\beta_{n}^{\text{CH}}=B_{0}+B(n),\,\ \text{where}\ B(n)=\sum_{k=1}^{\infty}\frac{B_{k}}{n^{k}}\,,$$ $B_{0}={{\rm i}}(a+b-c-d)/4$ and $B(n)$ does not vanish identically. This property in particular implies that $\beta_{m+1}^{\text{CH}}-\beta_{m}^{\text{CH}}\neq0$ for all sufficiently large $m$. Now we can check the asymptotic behavior of $\delta_{1}(n,m)$ and $\delta_{3}(n,m)$ for $m$ fixed and $n$ large. We have, as $n\to\infty$, $$\begin{aligned} \delta_{1}(n,m) & :=\det\!\left(\begin{array}{cc} \beta_{n}^{\text{CH}}-\beta_{m}^{\text{CH}} & \alpha_{n-1}^{\text{CH}}-\alpha_{m-1}^{\text{CH}}\\ \beta_{n-1}^{\text{CH}}-\beta_{m+1}^{\text{CH}} & \alpha_{n-2}^{\text{CH}}-\alpha_{m}^{\text{CH}} \end{array}\right)=\,\frac{\beta_{m+1}^{\text{CH}}-\beta_{m}^{\text{CH}}}{4}\,n+O(1),\\ \noalign{\smallskip}\delta_{3}(n,m) & :=\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{CH}}-\alpha_{m}^{\text{CH}} & \beta_{n}^{\text{CH}}-\beta_{m}^{\text{CH}}\\ \alpha_{n-1}^{\text{CH}}-\alpha_{m+1}^{\text{CH}} & \beta_{n-1}^{\text{CH}}-\beta_{m+1}^{\text{CH}} \end{array}\right)=\,-\frac{\beta_{m+1}^{\text{CH}}-\beta_{m}^{\text{CH}}}{4}\,n+O(1).\end{aligned}$$ Fix $m$ sufficiently large. Then $$\lim_{n\to\infty}\frac{h_{n+1}}{h_{n-1}}=\lim_{n\to\infty}\frac{h_{n+m+1}}{h_{n+m-1}}=\lim_{n\to\infty}\frac{\delta_{1}(n,m)}{\delta_{3}(n,m)}=-1.$$ In view of (\[eq:lim\_delta2\_chahn\]), we also have $$\lim_{n\to\infty}\frac{h_{n+1}}{h_{n}}=\lim_{n\to\infty}\frac{h_{n+m+1}}{h_{n+m}}=\lim_{n\to\infty}\frac{\delta_{1}(n,m)}{\delta_{2}(n,m)}=\frac{\beta_{m+1}^{\text{CH}}-\beta_{m}^{\text{CH}}}{4C_{m}}=:D\in\mathbb{R}\setminus\{0\}.$$ Then $$\lim_{n\to\infty}\frac{h_{n+1}}{h_{n-1}}=\lim_{n\to\infty}\frac{h_{n+1}}{h_{n}}\,\frac{h_{n}}{h_{n-1}}=D^{2}.$$ Hence $D^{2}=-1$, a contradiction. The next lemma summarizes configurations of the parameters when $\alpha_{n}^{\text{CH}}$ is a polynomial in $n$. \[lem:polyn\_param\_chahn\] $\alpha_{n}^{\text{CH}}$ depends polynomially on $n$ for the following values of parameters only: 1. ${\displaystyle a=b=\frac{1}{4}+{{\rm i}}t\;\mbox{ and }\;c=d=\frac{1}{4}-{{\rm i}}t}$, for $t\in{{\mathbb{R}}}$, in which case $$\alpha_{n}^{\text{CH}}=\frac{1}{4}\left(n+\frac{1}{2}\right).$$ 2. ${\displaystyle a=b=\frac{3}{4}+{{\rm i}}t\;\mbox{ and }\;c=d=\frac{3}{4}-{{\rm i}}t}$, for $t\in{{\mathbb{R}}}$, in which case $$\alpha_{n}^{\text{CH}}=\frac{1}{4}\left(n+\frac{3}{2}\right).$$ 3. ${\displaystyle a=\frac{1}{4}+{{\rm i}}t,\,b=\frac{3}{4}+{{\rm i}}t\;\mbox{ and }\;c=\frac{1}{4}-{{\rm i}}t,\,d=\frac{3}{4}-{{\rm i}}t}$, for $t\in{{\mathbb{R}}}$, or one can interchange $a$ with $b$ and $c$ with $d$. In both cases, $$\alpha_{n}^{\text{CH}}=\frac{1}{4}\left(n+1\right).$$ Moreover, $\beta_{n}^{\text{CH}}={{\rm i}}(a+b-c-d)/4$ in each case (I)-(III). \[rem:alp-1\_chahn\] Note that $\alpha_{-1}^{\text{CH}}=0$ only in case (III). Bearing in mind the restriction , we temporarily reparameterize the coefficients as follows: $$\alpha:={\mathop\mathrm{Re}\nolimits}a,\quad\beta:={\mathop\mathrm{Re}\nolimits}b,\quad\phi:={\mathop\mathrm{Im}\nolimits}a-{\mathop\mathrm{Im}\nolimits}b,\quad\theta:={\mathop\mathrm{Im}\nolimits}a+{\mathop\mathrm{Im}\nolimits}b,$$ where $\alpha,\beta>0$ and $\phi,\theta\in\mathbb{R}$. As a shortcut we again put $s:=a+b+c+d=2\alpha+2\beta$. We have $$\left(\alpha_{n}^{\text{CH}}\right)^{2}=\frac{(n+1)(n+2\alpha)(n+2\beta)(n+s-1)(2n+s+2{{\rm i}}\phi)(2n+s-2{{\rm i}}\phi)}{4\,(2n+s-1)(2n+s)^{2}(2n+s+1)}.\label{eq:alpha_n_2_chahn}$$ If $\alpha_{n}^{\text{CH}}$ is a polynomial in $n$, then it is of degree $1$ with the leading coefficient equal to $1/4$ as is obvious from . In that case, it is immediately seen from that $\phi=0$ and hence $$\left(\alpha_{n}^{\text{CH}}\right)^{2}=\frac{(n+1)(n+2\alpha)(n+2\beta)(n+s-1)}{4\,(2n+s-1)(2n+s+1)}.$$ Then by an elementary inspection of possible cancellation and requiring the above expression to be the square of a polynomial in $n$, one readily finds that there are only four possible configurations for $\alpha$ and $\beta$. Namely, $\alpha=\beta=1/4$, $\alpha=\beta=3/4$, $\alpha=1/4$ and $\beta=3/4$, or $\alpha=3/4$ and $\beta=1/4$. Moreover, in each case, $\beta_{n}^{\text{CH}}=-\theta/2$. The lemma follows. In summary, one implication of the equivalence in Theorem \[thm:main\] ad (iii) is established by Proposition \[prop:polyn\_impl\_nontriv\_chahn\]. Conversely, if ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{CH}})$ is nontrivial then $\alpha_{n}^{\text{CH}}$ depends polynomially on $n$, by Proposition \[prop:nontriv\_impl\_polyn\_chahn\]. In this case, $\beta_{n}^{\text{CH}}$ is a polynomial in $n$, too, as it follows from Lemma \[lem:polyn\_param\_chahn\]. The proof of Theorem \[thm:main\] ad (iii) is then completed by Proposition \[prop:polyn\_impl\_nontriv\_chahn\], Lemma \[lem:polyn\_param\_chahn\], and Remark \[rem:alp-1\_chahn\]. Jacobi ------ The proof of Theorem \[thm:main\] ad (iv) asserting that ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{J}})$ is trivial for any admissible choice of the parameters $\alpha$, $\beta$ is more elementary than in the foregoing cases but computationally rather demanding. Despite the fact that the coefficients $\alpha_{n}^{\text{J}}$ and $\beta_{n}^{\text{J}}$ are given by less complicated expressions derivations of some asymptotic formulas used in the course of the proof are best done with the aid of a convenient computer algebra system (CAS). To our opinion, any commonly used CAS will do. In order to obtain slightly less complicated expressions, we use a new parametrization: $c:=\alpha+\beta$ and $d:=\beta-\alpha$. The coefficients and take the form $$\alpha_{n}^{\text{J}}=\sqrt{\frac{(n+1)(2n+c+d+2)(2n+c-d+2)(n+c+1)}{(2n+c+1)(2n+c+2)^{2}(2n+c+3)}}\label{eq:def_alpha_jacobi_reparam}$$ and $$\beta_{n}^{\text{J}}=\frac{cd}{(2n+c)(2n+c+2)},\label{eq:def_beta_jacobi_reparam}$$ where we can assume that $c>-2$ and $d\geq0$ without loss of generality. Indeed, since $\alpha_{n}^{\text{J}}(\alpha,\beta)=\alpha_{n}^{\text{J}}(\beta,\alpha)$ and $\beta_{n}^{\text{J}}(\alpha,\beta)=-\beta_{n}^{\text{J}}(\beta,\alpha)$ for all $n\in{\mathbb{N}}_{0}$, a sequence $h_{n}$ is a solution of if and only if $(-1)^{n}h_{n}$ is a solution of the same equation with interchanged parameters $\alpha$ and $\beta$. Therefore we may assume $\beta\geq\alpha$. Writing $$\alpha_{n}^{\text{J}}=\frac{1}{2}+A(n)$$ we have $$\begin{aligned} A(n) & = & \frac{1-c^{2}-d^{2}}{16\,n^{2}}-\frac{(1-c^{2}-d^{2})(c+2)}{16\,n^{3}}\nonumber \\ \noalign{\smallskip} & & +\left(\frac{(1-c^{2}-d^{2})(13\,c^{2}+d^{2}+48\,c+51)}{256}+\frac{c^{2}d^{2}}{64}\right)\!\frac{1}{n^{4}}+O\!\left(\frac{1}{n^{5}}\right)\!,\,\ \text{as}\ n\to\infty.\nonumber \\ \label{eq:AJasympt}\end{aligned}$$ Furthermore, $$\beta_{n}^{\text{J}}=\frac{cd}{4\,n^{2}}-\frac{cd\,(c+1)}{4\,n^{3}}+\frac{cd\,(3c^{2}+6\,c+4)}{16\,n^{4}}-\frac{cd\,(c^{3}+3\,c^{2}+4\,c+2)}{8\,n^{5}}+O\!\left(\frac{1}{n^{6}}\right)\!,\,\ \text{as}\ n\to\infty.\label{eq:betaJasympt}$$ We again assume that a nontrivial real solution $h_{n}$ to exists, with the coefficients $\alpha_{n}$ and $\beta_{n}$ being given by and . Moreover, we keep the notation using, however, the coefficients $\alpha_{n}^{\text{J}}$, $\beta_{n}^{\text{J}}$ rather than $\alpha_{n}^{\text{W}}$, $\beta_{n}^{\text{W}}$. The asymptotic formulas lead us to distinguish four cases. 1) Case $c^{2}+d^{2}=1$, $cd=0$. Then we have three possible configurations of the parameters: $c=0$, $d=1$ or $c=-1$, $d=0$ or $c=1$, $d=0$. In any case $\alpha_{n}^{\text{J}}=1/2$ is a constant sequence. This means that the boundary condition $\alpha_{-1}^{\text{J}}=0$ is not fulfilled and therefore ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{J}})$ is trivial. 2) Case $c^{2}+d^{2}\neq1$, $cd=0$. Then $\beta_{n}^{\text{J}}=0$ identically and equations and simplify to $$\begin{aligned} & & (\alpha_{n}-\alpha_{m})h_{n+m+1}+(\alpha_{n-1}-\alpha_{m-1})h_{n+m-1}=0,\nonumber \\ & & (\alpha_{n-1}-\alpha_{m+1})h_{n+m+1}+(\alpha_{n-2}-\alpha_{m})h_{n+m-1}=0.\label{eq:J-a_b_H-b0}\end{aligned}$$ From it is seen that $$\delta_{2}(n,m)=-\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{J}}-\alpha_{m}^{\text{J}} & \alpha_{n-1}^{\text{J}}-\alpha_{m-1}^{\text{J}}\\ \alpha_{n-1}^{\text{J}}-\alpha_{m+1}^{\text{J}} & \alpha_{n-2}^{\text{J}}-\alpha_{m}^{\text{J}} \end{array}\right)=A(m+1)A(m-1)-A^{2}(m)+O\!\left(\frac{1}{n^{2}}\right)\!,\label{eq:delta_2_asympt_jacobi}$$ as $n\to\infty$. Moreover, $$A(m+1)A(m-1)-A^{2}(m)=\frac{(1-c^{2}+d^{2})^{2}}{128\,m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,\,\ \text{as}\ m\to\infty.\label{eq:delta_2_lead_m_asympt_jacobi}$$ Consequently, whenever $c^{2}+d^{2}\neq1$ then for all sufficiently large $m\in\mathbb{N}$ there exists $R_{m}\in{\mathbb{N}}$ such that $\delta_{2}(n,m)\neq0$ for all $n\geq R_{m}$. It readily follows from that $h_{n}=0$ for all sufficiently large $n$. Then necessarily $h_{n}=0$ for all $n\in\mathbb{N}_{0}$ (see ), a contradiction. 3) Case $c^{2}+d^{2}\neq1$, $cd\neq0$. In this case, too, we have $\delta_{2}(n,m)\neq0$ for $n\geq R_{m}$ provided $m$ is sufficiently large. Moreover, the asymptotic expansions , are valid. In addition, as $n\to\infty$, we have $$\delta_{1}(n,m)=\det\!\left(\begin{array}{cc} \beta_{n}^{\text{J}}-\beta_{m}^{\text{J}} & \alpha_{n-1}^{\text{J}}-\alpha_{m-1}^{\text{J}}\\ \beta_{n-1}^{\text{J}}-\beta_{m+1}^{\text{J}} & \alpha_{n-2}^{\text{J}}-\alpha_{m}^{\text{J}} \end{array}\right)=\beta_{m}^{\text{J}}A(m)-\beta_{m+1}^{\text{J}}A(m-1)+O\!\left(\frac{1}{n^{2}}\right)\label{eq:delta_1_asympt_jacobi}$$ and $$\delta_{3}(n,m)=\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{J}}-\alpha_{m}^{\text{J}} & \beta_{n}^{\text{J}}-\beta_{m}^{\text{J}}\\ \alpha_{n-1}^{\text{J}}-\alpha_{m+1}^{\text{J}} & \beta_{n-1}^{\text{J}}-\beta_{m+1}^{\text{J}} \end{array}\right)=\beta_{m+1}^{\text{J}}A(m)-\beta_{m}^{\text{J}}A(m+1)+O\!\left(\frac{1}{n^{2}}\right)\!.\label{eq:delta_3_asympt_jacobi}$$ Using (\[eq:AJasympt\]), (\[eq:betaJasympt\]) one derives that $$\beta_{m}^{\text{J}}A(m)-\beta_{m+1}^{\text{J}}A(m-1)=\frac{cd\,(c^{2}+d^{2}-1)}{64\,m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,\,\ \text{as}\ m\to\infty,\label{eq:delta_1_lead_m_asympt_jacobi}$$ and $$\beta_{m+1}^{\text{J}}A(m)-\beta_{m}^{\text{J}}A(m+1)=\frac{cd\,(c^{2}+d^{2}-1)}{64\,m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,\,\ \text{as}\ m\to\infty.\label{eq:delta_3_lead_m_asympt_jacobi}$$ Hence for all sufficiently large $m\in\mathbb{N}$ there exists $R_{m}\in\mathbb{N}$ such that for all $n\geq R_{m}$ and $j=1,2,3$, $\delta_{j}(n,m)\neq0$. Similarly as in the proof of Proposition \[prop:nontriv\_impl\_polyn\_wilson\] we can infer from here some information about the asymptotic behavior of $h_{n}$ as $n\to\infty$. We only sketch the basic steps. Fix sufficiently large $m\in\mathbb{N}$ but otherwise arbitrary. The vectors are linearly dependent and therefore for all $n\in\mathbb{N}$, $n\geq m_{0}:=R_{m}+m$, we have $$h_{n+1}=\psi(n)h_{n},\ \,\text{with}\ \text{ }\psi(n):=\frac{\delta_{1}(n-m,m)}{\delta_{2}(n-m,m)}\,.$$ From (\[eq:delta\_2\_asympt\_jacobi\]) and (\[eq:delta\_1\_asympt\_jacobi\]) it is seen that $\psi(n)=\lambda\big(1+O(n^{-2})\big)$ for some $\lambda\neq0$. Whence $$h_{n}=h_{m_{0}}\,\prod_{k=m_{0}}^{n-1}\psi(k)=\gamma\,\lambda^{\,n}\!\left(1+O\!\left(\frac{1}{n}\right)\!\right)\!,\,\ \text{as}\ n\to\infty,\label{eq:hJ_n-asympt}$$ where $\gamma\neq0$ is a constant. Without loss of generality we can assume that $\gamma=1$. Moreover, referring to (\[eq:delta\_1\_asympt\_jacobi\]), (\[eq:delta\_3\_asympt\_jacobi\]), $$\lambda^{2}=\lim_{n\to\infty}\,\frac{h_{n+1}}{h_{n-1}}=\lim_{n\to\infty}\,\frac{h_{n+m+1}}{h_{n+m-1}}=\lim_{n\to\infty}\,\frac{\delta_{1}(n,m)}{\delta_{3}(n,m)}=\frac{\beta_{m}^{\text{J}}A(m)-\beta_{m+1}^{\text{J}}A(m-1)}{\beta_{m+1}^{\text{J}}A(m)-\beta_{m}^{\text{J}}A(m+1)}\,.\label{eq:J-lambda-2}$$ In regard of (\[eq:delta\_1\_lead\_m\_asympt\_jacobi\]), (\[eq:delta\_3\_lead\_m\_asympt\_jacobi\]), letting $m\to\infty$ we obtain $\lambda^{2}=1$. Equation (\[eq:a\_b\_H\]) in this case means that $$\big(A(n)-A(m)\big)h_{n+m+1}+(\beta_{n}^{\text{J}}-\beta_{m}^{\text{J}})h_{n+m}+\big(A(n-1)-A(m-1)\big)h_{n+m-1}=0.$$ From here, when considering the limit $n\to\infty$ while taking into account (\[eq:hJ\_n-asympt\]) and (\[eq:AJasympt\]), (\[eq:betaJasympt\]), we get $A(m)\lambda^{2}+\beta_{m}^{\text{J}}\lambda+A(m-1)=0$. Hence for $\lambda=\pm1$ we have $$\pm\beta_{m}^{\text{J}}+A(m)+A(m-1)=0,\,\ m\in\mathbb{N}_{0}.$$ Referring again to (\[eq:AJasympt\]), (\[eq:betaJasympt\]) and checking the asymptotic expansion, as $m\to\infty$, we find that $$\pm\beta_{m}^{\text{J}}+A(m)+A(m-1)=-\frac{(c\mp d)^{2}-1}{8\,m^{2}}+O\!\left(\frac{1}{m^{3}}\right)\!.$$ Hence $(c\mp d)^{2}=1$. With this equation the asymptotic expansion simplifies and we obtain $$\pm\beta_{m}^{\text{J}}+A(m)+A(m-1)=\mp\frac{3cd}{16\,m^{4}}+O\!\left(\frac{1}{m^{5}}\right)\!.$$ Owing to our assumption $cd\neq0$ we arrive at a contradiction. 4) Case $c^{2}+d^{2}=1$, $cd\neq0$. Discussion of this case is similar to the foregoing one, but some leading terms in the above asymptotic expansions disappear and this is why we have to reconsider some formulas. First of all, we now have, as $n\to\infty$, $$A(n)=\frac{c^{2}d^{2}}{64\,n^{4}}-\frac{(c+2)\,c^{2}d^{2}}{32\,n^{5}}+O\!\left(\frac{1}{n^{6}}\right)\!,$$ while the asymptotic expansion (\[eq:betaJasympt\]) of $\beta_{n}^{\text{J}}$ remains as it is. Furthermore, for a fixed $m\in\mathbb{N}_{0}$, $$\delta_{2}(n,m)=-\det\!\left(\begin{array}{cc} \alpha_{n}^{\text{J}}-\alpha_{m}^{\text{J}} & \alpha_{n-1}^{\text{J}}-\alpha_{m-1}^{\text{J}}\\ \alpha_{n-1}^{\text{J}}-\alpha_{m+1}^{\text{J}} & \alpha_{n-2}^{\text{J}}-\alpha_{m}^{\text{J}} \end{array}\right)=A(m+1)A(m-1)-A^{2}(m)+O\!\left(\frac{1}{n^{4}}\right)\!,$$ and the asymptotic expansions (\[eq:delta\_1\_asympt\_jacobi\]), (\[eq:delta\_3\_asympt\_jacobi\]) remain valid as they are. The following asymptotic expansions, as $m\to\infty$, had to be reconsidered, however: $$\begin{aligned} A(m+1)A(m-1)-A^{2}(m) & = & \frac{c^{4}d^{4}}{1024\,m^{10}}+O\!\left(\frac{1}{m^{11}}\right)\!,\nonumber \\ \beta_{m}^{\text{J}}A(m)-\beta_{m+1}^{\text{J}}A(m-1) & = & -\frac{c^{3}d^{3}}{128\,m^{7}}+O\!\left(\frac{1}{m^{8}}\right)\!,\label{eq:J-delta-asympt-lead-1}\\ \beta_{m+1}^{\text{J}}A(m)-\beta_{m}^{\text{J}}A(m+1) & = & \frac{c^{3}d^{3}}{128\,m^{7}}+O\!\left(\frac{1}{m^{8}}\right)\!.\nonumber \end{aligned}$$ From (\[eq:J-delta-asympt-lead-1\]) we conclude that in this case, too, for all sufficiently large $m\in\mathbb{N}$ there exists $R_{m}\in\mathbb{N}$ such that for all $n\geq R_{m}$ and $j=1,2,3$, $\delta_{j}(n,m)\neq0$. It follows that the asymptotic analysis (\[eq:hJ\_n-asympt\]) of $h_{n}$, as $n\to\infty$, is still applicable. Whence, up to a constant multiplier, $h_{n}=\lambda^{n}\big(1+O(n^{-1})\big)$ for some $\lambda\in\mathbb{R}$, $\lambda\neq0$. We can again compute $\lambda^{2}$ as in (\[eq:J-lambda-2\]) with the same result, namely $$\lambda^{2}=\frac{\beta_{m}^{\text{J}}A(m)-\beta_{m+1}^{\text{J}}A(m-1)}{\beta_{m+1}^{\text{J}}A(m)-\beta_{m}^{\text{J}}A(m+1)}\,.$$ This time letting $m\to\infty$ and taking into account (\[eq:J-delta-asympt-lead-1\]) we get $\lambda^{2}=-1$ which is not possible. Meixner–Pollaczek, Meixner, Laguerre, Charlier and Hermite ---------------------------------------------------------- The remaining cases will be directly treated using a necessary condition for the existence of a nontrivial commuting Hankel matrix. Shifting indices in the general commutation equation , one gets the equation and further $$(\alpha_{n-2}-\alpha_{m+2})h_{n+m+1}+(\beta_{n-2}-\beta_{m+2})h_{n+m}+(\alpha_{n-3}-\alpha_{m+1})h_{n+m-1}=0.$$ These three equations can have a nontrivial solution only if $$D(m,n):=\det\begin{pmatrix}\alpha_{n}-\alpha_{m} & \beta_{n}-\beta_{m} & \alpha_{n-1}-\alpha_{m-1}\\ \alpha_{n-1}-\alpha_{m+1} & \beta_{n-1}-\beta_{m+1} & \alpha_{n-2}-\alpha_{m}\\ \alpha_{n-2}-\alpha_{m+2} & \beta_{n-2}-\beta_{m+2} & \alpha_{n-3}-\alpha_{m+1} \end{pmatrix}=0,\label{eq:def_Dmn}$$ for all $m,n\in{\mathbb{N}}$, $n\geq2$. In particular, if $\beta_{n}=0$ for all $n\in{\mathbb{N}}_{0}$, it follows from that the necessary condition for existence of a nontrivial commuting Hankel matrix reads $$\delta_{2}(m,n):=-\det\begin{pmatrix}\alpha_{n}-\alpha_{m} & \alpha_{n-1}-\alpha_{m-1}\\ \alpha_{n-1}-\alpha_{m+1} & \alpha_{n-2}-\alpha_{m} \end{pmatrix}=0,\quad\forall m,n\in{\mathbb{N}}.\label{eq:def_dmn}$$ ### Meixner–Pollaczek Assume first that $\phi\neq\pi/2$. The leading terms of the asymptotic expansions of $\alpha_{n}^{\text{MP}}$ and $\beta_{n}^{\text{MP}}$ are $$\alpha_{n}^{\text{MP}}=\frac{n}{2\sin\phi}+O(1)\quad\mbox{ and }\quad\beta_{n}^{\text{MP}}=-\frac{n}{\tan\phi}+O(1),\,\ \text{as}\ n\to\infty.$$ Consequently, for $m\in{\mathbb{N}}$ fixed, $$D(m,n)=C_{m}(\lambda,\phi)n+O(1),\quad n\to\infty.$$ By a simple linear algebra the function $C_{m}(\lambda,\phi)$ can be readily computed, $$\begin{aligned} C_{m}(\lambda,\phi)=\frac{\cos\phi}{4\sin^{3}\phi}\big[ & (\tilde{\alpha}_{m+1}^{\text{MP}}-\tilde{\alpha}_{m}^{\text{MP}})^{2}-(\tilde{\alpha}_{m+2}^{\text{MP}}-\tilde{\alpha}_{m+1}^{\text{MP}})(\tilde{\alpha}_{m}^{\text{MP}}-\tilde{\alpha}_{m-1}^{\text{MP}})\\ & +\tilde{\alpha}_{m+2}^{\text{MP}}-3\tilde{\alpha}_{m+1}^{\text{MP}}+3\tilde{\alpha}_{m}^{\text{MP}}-\tilde{\alpha}_{m-1}^{\text{MP}}\big],\end{aligned}$$ where $\tilde{\alpha}_{n}^{\text{MP}}:=2\sin(\phi)\alpha_{n}^{\text{MP}}=\sqrt{(n+1)(n+2\lambda)}\,$. Further, the asymptotic expansion of the above expression in the square brackets equals $$-\frac{(1-2\lambda)^{4}}{32m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,\,\ m\to\infty.$$ Consequently, if ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{MP}})$ is nontrivial, then $C_{m}(\lambda,\phi)=0$ for all $m\in{\mathbb{N}}$ which particularly implies that $\lambda=1/2$. If $\phi=\pi/2$, then $\beta_{n}^{\text{MP}}=0$ for all $n\in{\mathbb{N}}_{0}$, and one proceeds similarly using instead of . The resulting formula reads $$\delta_{2}(m,n)=c_{m}(\lambda)n+O(1),\,\ \text{as}\ n\to\infty,$$ where $$c_{m}(\lambda)=\frac{(1-2\lambda)^{2}}{16m^{3}}+O\!\left(\frac{1}{m^{4}}\right)\!,\,\ \text{as}\ m\to\infty.$$ Thus one again concludes that $\lambda=1/2$, if ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{MP}})$ is nontrivial. Assume on the other hand that $\lambda=1/2$. Note that $\lambda=1/2$ if and only if $\alpha_{n}^{\text{MP}}$ is a nonzero polynomial in $n$. Then $\alpha_{-1}^{\text{MP}}=0$. Plugging $\alpha_{n}^{\text{MP}}$ and $\beta_{n}^{\text{MP}}$, given by with $\lambda=1/2$, into we arrive at the second order difference equation $$h_{k+1}-2\cos(\phi)h_{k}+h_{k-1}=0,\quad k\in{\mathbb{N}},$$ whose two linearly independent solutions are $$h_{k}^{(1)}=\sin(k\phi)\quad\mbox{ and }\quad h_{k}^{(2)}=\cos(k\phi).$$ Consequently, $\dim{\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{MP}})=2$ and the two Hankel matrices determined by the sequences $h^{(1)}$ and $h^{(2)}$ form a basis of ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{MP}})$. ### Meixner In all remaining cases, the approach is completely analogous to the case of the Meixner–Pollaczeck polynomials. Therefore we mention only several most important points. In the case of the Meixner coefficients defined by one obtains $$\lim_{n\to\infty}\frac{D(m,n)}{n}=\frac{c(c+1)(1-\beta)^{4}}{32(1-c)^{3}m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,$$ as $m\to\infty$. Since $c\in(0,1)$ necessarily $\beta=1$, if ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{M}})$ nontrivial. Assuming on the other hand that $\beta=1$ and plugging into , one arrives at the three-term recurrence $$h_{k+1}+(c^{-1/2}+c^{1/2})h_{k}+h_{k-1}=0,\quad k\in{\mathbb{N}},$$ whose two linearly independent solutions are $$h_{k}^{(1)}=(-1)^{k}c^{k/2}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k}c^{-k/2}.$$ ### Laguerre Using in , one computes $$\lim_{n\to\infty}\frac{D(m,n)}{n}=\frac{\alpha^{4}}{16m^{6}}+O\!\left(\frac{1}{m^{7}}\right)\!,$$ as $m\to\infty$. Hence ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{L}})$ can be nontrivial only for $\alpha=0$. On the other hand, if $\alpha=0$, one deduces from the recurrence $$h_{k+1}+2h_{k}+h_{k-1}=0,\quad k\in{\mathbb{N}},$$ whose two linearly independent solutions are $$h_{k}^{(1)}=(-1)^{k}\quad\mbox{ and }\quad h_{k}^{(2)}=(-1)^{k}k.$$ ### Charlier Note that $\alpha_{n}^{\text{C}}$ defined in is never a polynomial in $n$ for $a\neq0$. In that case one obtains $$\lim_{n\to\infty}\frac{D(m,n)}{n}=\frac{a}{8m^{3}}+O\!\left(\frac{1}{m^{4}}\right)\!,\,\ \text{as}\ m\to\infty.$$ Since $a\neq0$ determinant $D(m,n)$ cannot vanish identically and therefore ${\mathop\mathrm{Comm}_{H}\nolimits}(\mathcal{J}^{\text{C}})$ is trivial. ### Hermite Substituting from into the equation , one obtains the second order recurrence $$h_{k+1}+h_{k-1}=0,\quad k\in{\mathbb{N}},$$ whose two linearly independent solutions are $$h_{k}^{(1)}=\sin\frac{k\pi}{2}\quad\mbox{ and }\quad h_{k}^{(2)}=\cos\frac{k\pi}{2}.$$ Acknowledgments {#acknowledgments .unnumbered} =============== The authors acknowledge financial support by the Ministry of Education, Youth and Sports of the Czech Republic project no. CZ.02.1.01/0.0/0.0/16\_019/0000778. [10]{} F.A Gr[" u]{}nbaum, Toeplitz matrices commuting with tridiagonal matrices, Linear Algebra Appl. 40 (1981), 25–36. Tomáš Kalvoda and Pavel Šťovíček, A family of explicitly diagonalizable weighted [H]{}ankel matrices generalizing the [H]{}ilbert matrix, Linear Multilinear Algebra 64 (2016), no. 5, 870–884. Roelof Koekoek, Peter A. Lesky, and René F. Swarttouw, Hypergeometric orthogonal polynomials and their [$q$]{}-analogues, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2010, With a foreword by Tom H. Koornwinder. Wilhelm Magnus, On the spectrum of [H]{}ilbert’s matrix, Amer. J. Math. 72 (1950), 699–704. Vladimir V. Peller, Hankel operators and their applications, Springer Monographs in Mathematics, Springer-Verlag, New York, 2003. S. C. Power, Hankel operators on [H]{}ilbert space, Research Notes in Mathematics, vol. 64, Pitman (Advanced Publishing Program), Boston, Mass.-London, 1982. Marvin Rosenblum, On the [H]{}ilbert matrix. [I]{}, Proc. Amer. Math. Soc. 9 (1958), 137–140. , On the [H]{}ilbert matrix. [II]{}, Proc. Amer. Math. Soc. 9 (1958), 581–585. Hari Shanker, An integral equation for [W]{}hittaker’s confluent hypergeometric function, Proc. Cambridge Philos. Soc. 45 (1949), 482–483. František Štampach and Pavel Šťovíček, New explicitly diagonalizable [H]{}ankel matrices related to the [S]{}tieltjes–[C]{}arlitz polynomials, in preparation (2019). , Spectral representation of some weighted [H]{}ankel matrices and orthogonal polynomials from the [A]{}skey scheme, J. Math. Anal. Appl. 472 (2019), no. 1, 483–509. D. R. Yafaev, A commutator method for the diagonalization of [H]{}ankel operators, Funktsional. Anal. i Prilozhen. 44 (2010), no. 4, 65–79.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present results of first-principles calculations of the frequency-dependent polarizabilities of all alkali atoms for light in the wavelength range 300-1600 nm, with particular attention to wavelengths of common infrared lasers. We parameterize our results so that they can be extended accurately to arbitrary wavelengths above 800 nm. This work is motivated by recent experiments involving simultaneous optical trapping of two different alkali species. Our data can be used to predict the oscillation frequencies of optically-trapped atoms, and particularly the ratios of frequencies of different species held in the same trap. We identify wavelengths at which two different alkali atoms have the same oscillation frequency.' author: - 'M. S. Safronova' - Bindiya Arora - 'Charles W. Clark' title: 'Frequency-dependent polarizabilities of alkali atoms from ultraviolet through infrared spectral regions' --- Introduction ============ The trapping of two different species of ultracold gases has been an active topic of research for about ten years (see [@Lundblad] and references therein), and has led to the simultaneous production of quantum degenerate Bose-Einstein and Fermi-Dirac gases [@schreck:080403; @hadzibabic:160401; @Modugno] and recently to the observation of heteronuclear Feshbach resonances involving Bose-Einstein and Fermi-Dirac species [@Stan; @Inouye]. Two recent experiments [@Stan; @Inouye] provide a glimpse of the possibilities of mapping out the phase diagram of a Bose-Einstein/Fermi-Dirac mixture, potentially a rich system that remains to be explored. Our present work provides accurate estimates of the trapping frequencies of disparate species confined in the same optical trap. We have been attentive to finding cases in which all species, Bose-Einstein and Fermi-Dirac atoms, and the molecules formed from them, have the same trapping frequencies. This case is of particular interest because it has been shown to be optimal for a recently proposed scheme of cooling mediated by Feshbach resonances [@Miguel]. In addition, we note that while the DC polarizabilities of alkali atoms have been subject to careful experimental investigation, there is relatively little accurate experimental data on frequency-dependent polarizabilities. We draw attention to the fact that current experiments offer the possibility of measurements of unprecedented accuracy regarding the ratio of frequency-dependent polarizabilities of different species. Such measurements could provide more decisive comparison of [ab initio]{} atomic theory with experiments, than has been made to date. We begin with a summary of the calculation of frequency-dependent polarizabilities, the details of which have been presented in previous publications [@gate; @rb]. The computational method we use here has delivered agreement with experiments on static polarizabilities at the 1% level. There are few direct measurements of frequency-dependent polarizabilities - the experiments are much more demanding, and have larger uncertainties associated with modeling the intensity profile of a focused laser beam. However, the alkali metal atoms are a special case, at least in the infrared spectral region. There, the polarizability is dominated by the contribution from the longest-wavelength resonance transitions, and it is possible to perform consistency checks between static values and long-wavelength values of the frequency-dependent atomic polarizabilities. We present tabulated values of calculated frequency-dependent polarizabilities over a range of wavelengths from the ultraviolet through the infrared spectrum, with particular attention to those wavelengths (primarily in the infrared) which have been or might be employed in the present generation of optical traps. For the infrared region, we display simple formulae that allow accurate computation of frequency-dependent polarizabilities for wavelengths not explicitly reported upon here. Finally, we enumerate selected wavelengths, at which two species of optically-trapped alkali atoms would have the same frequencies of oscillation in a common optical trap. Such wavelengths offer opportunities for mapping out phase diagrams of Bose-Einstein and Fermi-Dirac mixtures in a space of reduced dimensionality. Moreover, to the extent that molecular resonances consist of weakly bound states at large internuclear separation, the oscillation frequencies of trapped molecular species should be very close to the frequencies of their atomic components. This may present an ideal system for investigation. Calculation of polarizabilities =============================== \[ht\] [rdrrrrrr]{} & & & & & & &\ 1554 & 0.02931 & 201.0(7) & 189.4(2) & 381.2(8) & 424.0(7) & 570.9(1.2) & 407.9(2.0)\ 1550 & 0.02939 & 201.3(7) & 189.6(2) & 381.8(8) & 424.7(7) & 572.2(1.2) & 408.5(2.0)\ 1540 & 0.02958 & 201.9(7) & 190.0(2) & 383.4(8) & 426.6(7) & 575.6(1.2) & 410.1(2.0)\ 1340 & 0.03399 & 218.1(8) & 201.0(2) & 428.2(9) & 479.5(7) & 673.7(1.3) & 455.0(2.1)\ 1332 & 0.03420 & 219.0(8) & 201.6(2) & 430.7(9) & 482.5(7) & 679.5(1.3) & 457.5(2.1)\ 1313 & 0.03469 & 221.1(8) & 203.0(2) & 436.9(9) & 489.9(7) & 694.1(1.3) & 463.7(2.1)\ 1240 & 0.03673 & 231.0(8) & 209.3(2) & 465.9(9) & 524.6(7) & 764.6(1.4) & 492.8(2.2)\ 1152 & 0.03954 & 247.0(8) & 219.4(2) & 516.1(1.0)& 585.7(8) & 899.9(1.6) & 543.6(2.3)\ 1090 & 0.04179 & 262.7(8) & 228.8(2) & 568.7(1.1)& 650.8(8) & 1062.3(1.7) & 597.2(2.4)\ 1064 & 0.04281 & 270.8(8) & 233.6(2) & 597.5(1.2)& 686.9(9) & 1162.1(1.9) & 627(3)\ 1060 & 0.04297 & 272.1(8) & 234.3(2) & 602.4(1.2)& 693.1(9) & 1180.0(1.9) & 632(3)\ 1053 & 0.04326 & 274.6(8) & 235.7(3) & 611.3(1.2)& 704.4(9) & 1213.2(1.9) & 641(3)\ 1047 & 0.04351 & 276.7(8) & 237.0(3) & 619.3(1.2)& 714.5(9) & 1243.8(2.0) & 650(3)\ 1030 & 0.04422 & 283.2(8) & 240.7(3) & 643.9(1.3)& 746.0(9) & 1343.6(2.1) & 675(3)\ 985 & 0.04624 & 304.0(8) & 252.0(3) & 728.3(1.4)& 855.8(1.0)& 1759(3) & 765(3)\ 980 & 0.04648 & 306.7(8) & 253.5(3) & 739.9(1.4)& 871.2(1.0)& 1828(3) & 778(3)\ 946 & 0.04815 & 327.8(9) & 264.4(3) & 836.4(1.6)& 1001.6(1.1)& 2581(4) & 886(3)\ 930 & 0.04898 & 339.6(9) & 270.3(3) & 895.9(1.7)& 1084.4(1.2)& 3325(5) & 957(3)\ 799.3&0.05699 & 549.7(1.1) & 353.8(4) & 3666(7) & 13483(14) & -2333(4) & -1474(11)\ 700 & 0.06507 & 1983(3) & 553.1(0.6) & -1394(3) & -1192.7(1.4) & -707.9(1.4) & -3787(18)\ 600 & 0.07592 & -645.9(1.3) & 4506(5) & -438.4(0.9) & -421.8(0.7) & -333.9(1.0)& -523(3)\ 500 & 0.09110 & -200.1(0.8) & -411.1(0.4) & -201.3(0.5) & -196.8(0.6) & -166.3(1.0)& -204(2)\ 400 & 0.11388 & -86.4(0.8) & -134.0(0.2) & -131.4(2.4) & -109.7(1.1) & -88.2(1.3)& -111(4)\ 300 & 0.15183 & -38.4(1.1) & -57.5(0.3) & -47.2(0.4) & -43.7(1.1) & -35(2) & -43(2)\ The frequency-dependent polarizability of an atom in its ground state may be separated into polarizability of the ionic core $\alpha_{\rm core}$ and the valence contribution $\alpha_v^{ns}$. The calculation of the core polarizability is carried out in the random-phase approximation (RPA); we verified that our static values agree with RPA values from Ref. [@jkh]. The separation of the polarizability into core and valence parts also requires the addition of the compensation term $\alpha_{vc}$ which accounts for the contribution from the excitation to the occupied valence shell that is forbidden by the Pauli exclusion principle. The calculation of the core polarizability does not exclude the excitation from the core to the valence shell and half of this contribution has to be subtracted. Thus, the polarizability contributions may be separated as $$\alpha^{ns}=\alpha_{\rm core}+\alpha_{vc}+\alpha_v^{ns}.$$ We calculate the valence part of the ac polarizability of the ground state $ns$ for the alkali-metal atoms by computing the sum over states $$\begin{aligned} \alpha_v^{ns}&=&\frac{1}{3}\sum_{n^{\prime}} \left( \frac{\delta E_{n^{\prime}p_{1/2}} \langle n^{\prime}p_{1/2}\\|D\\| ns \rangle^2 } {\delta E_{n^{\prime}p_{1/2}}^2-\omega^2}\right. \nonumber \\ &+&\left. \frac{\delta E_{n^{\prime}p_{3/2}} \langle n^{\prime}p_{3/2}\\|D\\| ns \rangle^2 } {\delta E_{n^{\prime}p_{3/2}}^2-\omega^2} \right), \label{eq1}\end{aligned}$$ where $\langle np\\|D\\| ns \rangle$ is the reduced electric-dipole matrix element, $\delta E_{n^{\prime}p_{1/2}}= E_{n^{\prime}p_{1/2}}-E_{ns}$, and $\delta E_{n^{\prime}p_{3/2}}= E_{n^{\prime}p_{3/2}}-E_{ns}$. In this formula, $\omega$ is assumed to be at least several linewidths off resonance with the corresponding transition. We use the system of atomic units, a.u., in which $e/\sqrt{4\pi\epsilon_0}, m_{\rm e}$, and the reduced Planck constant $\hbar$ have the numerical value 1, in Eq.(\[eq1\]). Polarizability in a.u. has the dimensions of volume, and its numerical values presented here are thus measured in units of $a^3_0$, where $a_0\approx0.052918$ nm is the Bohr radius. The atomic units for $\alpha$ can be be converted to SI units via $\alpha/h$ \[Hz/(V/m)$^2$\]=2.48832$\times10^{-8}\alpha$ \[a.u.\], where the conversion coefficient is $4\pi \epsilon_0 a^3_0/h$ and Planck constant $h$ is factored out. The atomic unit of frequency $\omega$ is $E_h/\hbar\approx4.1341\times10^{16}$ Hz, where $E_h$ is the Hartree energy. \[ht\] ---- --- ------------ ------------ ---------- ---------- ------- ------ Li 2 3.317(4) 4.689(5) 0.067906 0.067907 2.04 0.69 Na 3 3.5246(23) 4.9838(34) 0.077258 0.077336 1.86 0.12 K 4 4.102(5) 5.800(8) 0.059165 0.059428 6.26 0.33 Rb 5 4.231(3) 5.977(4) 0.057314 0.058396 10.54 0.60 Cs 6 4.4890(65) 6.3238(73) 0.050932 0.053456 17.35 1.00 Fr 7 4.277(8) 5.898(15) 0.055758 0.063442 24.8 1.8 ---- --- ------------ ------------ ---------- ---------- ------- ------ The sum over $n^{\prime}$ in Eq.(\[eq1\]) converges rapidly, and unless the frequency is resonant with the transitions other than the primary ones ($ns-np_{1/2}$ and $ns-np_{3/2}$), the first term is dominant. Therefore, only a few terms need to be calculated accurately. We use the experimental numbers compiled in Ref. [@relsd] together with their uncertainties for the first term, for example for $n^{\prime}=3$ term for Na. High-precision theoretical all-order values are used for the next three terms (for example $n^{\prime}=4, 5, 6$ terms for Na). We refer the reader to Refs. [@cs; @na; @relsd] for the detailed description of the single-double (SD) all-order method and its extensions. The all-order values of the matrix elements are evaluated for their accuracy, and the extensions of the SD all-order method (SDpT, which partially includes triple excitations or SD$_{sc}$, which includes semi-empirical scaling of dominant terms) are used for certain transitions where those values are expected to be of better accuracy. The uncertainty of the resulting matrix elements is evaluated based on the relative importance of certain classes of the all-order terms, extensive comparison of various atomic properties of alkali-metal atoms with experiment [@na; @relsd; @rb], and the spread of SD, SDpT, and SD$_{sc}$ values. Some of these values were published previously in Refs. [@relsd; @csus]. The experimental values from [@csus] are used for $6s-7p_{1/2}$ and $6s-7p_{3/2}$ transitions. In summary, the set of values used for the calculation of the first four terms in the sum over states of Eq.(\[eq1\]) consists of the best known values for these transitions. The experimental energies from Refs. [@NIST; @NIST1] are used for the corresponding energy levels. We refer to the total contribution from the first four terms as the main term; the remaining contributions are referred to below as the tail contribution; i.e. $$\alpha_v^{ns}=\alpha_{\rm main}(n^{\prime}=n,\dots, n+3)+\alpha_{\rm tail}(n^{\prime}>n+3).$$ ----------------------- ----------- ------------ ------------ ----------- -- 557.56(13) $^6$Li -360(1) $^{23}$Na -1375(1) \[0.2pc\] 768.648(3) $^6$Li 681.3(1.3) $^{40}$K 4525(112) 799.46(12) $^6$Li 549(1) $^{40}$K 3649(7) \[0.2pc\] 768.645(3) $^6$Li 681.3(1.3) $^{41}$K 4638(112) 798.44(12) $^6$Li 553(1) $^{41}$K 3763(7) \[0.2pc\] 786.07(1) $^6$Li 597.9(1.2) $^{87}$Rb 8639(20) 804.57(4) $^6$Li 533(1) $^{87}$Rb 7704(8) \[0.2pc\] 862.097(28) $^6$Li 412.5(1.0) $^{133}$Cs 9113(22) 904.76(4) $^6$Li 361.7(9) $^{133}$Cs 7991(12) \[0.2pc\] 549.72(18) $^7$Li -329(1) $^{23}$Na -1078(1) \[0.2pc\] 768.665(3) $^7$Li 681(1) $^{40}$K 3880(112) 806.9(2) $^7$Li 526(1) $^{40}$K 2998(6) \[0.2pc\] 768.662(3) $^7$Li 681(1) $^{41}$K 3978(112) 805.55(15) $^7$Li 530(1) $^{41}$K 3096(6) \[0.2pc\] 786.56(1) $^7$Li 595.9(1.2) $^{87}$Rb 7383(19) 807.16(4) $^7$Li 526(1) $^{87}$Rb 6510(6) \[0.2pc\] 863.436(32) $^7$Li 410.5(1.0) $^{133}$Cs 7777(20) 907.25(6) $^7$Li 359.3(9) $^{133}$Cs 6805(10) \[0.2pc\] 789.240(7) $^{23}$Na 364.9(4) $^{87}$Rb 1380(16) 946.664(5) $^{23}$Na 264.1(3) $^{87}$Rb 999(1) \[0.2pc\] 875.53(3) $^{23}$Na 295.5(3) $^{133}$Cs 1709(11) 1021.5(4) $^{23}$Na 242.6(3) $^{133}$Cs 1403(2) \[0.2pc\] 768.848(3) $^{40}$K -3893(116) $^{87}$Rb -8470(10) 784.32(1) $^{40}$K 6806(13) $^{87}$Rb 14800(27) 807.31(6) $^{40}$K 2968(6) $^{87}$Rb 6454(6) \[0.2pc\] 868.90(4) $^{40}$K 1298(3) $^{133}$Cs 4317(14) 938.6(2) $^{40}$K 862(2) $^{133}$Cs 2868(4) \[0.2pc\] 768.851(3) $^{41}$K -4013(116) $^{87}$Rb -8472(10) 784.42(1) $^{41}$K 6767(12) $^{87}$Rb 14358(26) 808.07(7) $^{41}$K 2916(6) $^{87}$Rb 6187(6) \[0.2pc\] 869.14(4) $^{41}$K 1296(2) $^{133}$Cs 4203(14) 940.9(2) $^{41}$K 854(2) $^{133}$Cs 2771(4) \[0.2pc\] 790.303(6) $^{87}$Rb -1299(17) $^{133}$Cs -1986(3) 873.39(4) $^{87}$Rb 1624(2) $^{133}$Cs 2483(12) 1158(3) $^{87}$Rb 580.5(8) $^{133}$Cs 888(2) \[0.2pc\] ----------------------- ----------- ------------ ------------ ----------- -- : \[tab3\] Selected wavelengths $\lambda_{\rm air}$ in nm, at which two species of optically-trapped alkali atoms ($^6$Li, $^7$Li, $^{23}$Na, $^{40}$K, $^{41}$K, $^{87}$Rb, and $^{133}$Cs) would have the same frequencies of oscillation in a common optical trap. The corresponding ground state frequency-dependent polarizabilities of the alkali-metal atoms (in a.u.) are also given. We evaluated the tail contributions with a complete set of basis Dirac-Hartree-Fock (DHF) wave functions generated using the B-spline method [@Bspline]. We use 40 splines of order $k=7$ for each angular momentum. The basis set orbitals are defined on a non-linear grid and are constrained to a large spherical cavity of a radius $R=75-110$ a.u. The cavity radius is chosen to accommodate first four $ns$, $np_{1/2}$, and $np_{3/2}$ valence orbitals. For consistency, we use the same basis sets for the tail calculations as we used in the all-order calculation of the main term matrix elements. The tail contribution is calculated in the DHF approximation. We list our results for the frequency-dependent polarizabilities from the ultraviolet through the infrared spectrum in Table \[tab1\]. Particular attention is given to the values for the infrared wavelengths which have been or might be employed in the present generation of optical traps. Our values of the ac polarizabilities are accurate to better than 0.5% with the exception of the values corresponding to wavelengths in the ultraviolet spectrum. For these wavelengths, non-primary resonances become dominant and the uncertainty of the polarizability values is dominated by the uncertainty in the corresponding resonance matrix elements. For example, the first resonances in Rb ($5s-5p$ transitions) occur at 795 nm and 780 nm, while next resonances ($5s-6p$) transitions are near 420 nm. As the accuracy of the $5s-6p$ matrix elements is lower than that of the $5s-5p$ matrix elements the resulting accuracy of the ac polarizabilities at 400 nm is also lower. Results ======= We have investigated the dependence on the frequency $\omega$ of all of the terms contributing to $\alpha^{ns}$, where $ns$ is the ground state. We find that the $\alpha_{core}$, $\alpha_{vc}$, $\alpha_{tail}$, and $\alpha^{ns}_v$ terms with $n^{\prime} = n+1, n+2, n+3$ depend weakly on the value of $\omega$ for wavelengths above 800 nm. In fact, it is possible to accurately parameterize the values of the ac polarizabilities for those wavelengths by the formula $$\begin{aligned} \alpha_v^{ns}(\omega)&=&\frac{1}{3} \frac{\delta E_{np_{1/2}} \langle np_{1/2}\\|D\\| ns \rangle^2 } {\delta E_{np_{1/2}}^2-\omega^2} \nonumber \\ &+&\frac{1}{3}\frac{\delta E_{np_{3/2}} \langle np_{3/2}\\|D\\| ns \rangle^2 } {\delta E_{np_{3/2}}^2-\omega^2}+ A^{ns} , \label{eq2}\end{aligned}$$ where $A^{ns}$ is independent on the value of $\omega$. All values in Eq.(\[eq2\]) are in atomic units. The values of the constants $A^{ns}$ and their uncertainties are given in Table \[tab2\]. The values of $A^{ns}$ are calculated as the average values of the totals of the $\alpha_{core}$, $\alpha_{vc}$, $\alpha_{tail}$, and $\alpha^{ns}_v$ terms with $n^{\prime} = n+1, n+2, n+3$ calculated in the range from 800 nm to 1600 nm. The values of the $\langle np\\|D\\| ns \rangle$ reduced electric-dipole matrix elements and the corresponding energy differences $\delta E$ are also given. The electric-dipole matrix elements for Li are experimental values from Ref. [@ADNDT], all other matrix elements are experimental values compiled in [@relsd]. The energy differences are taken from Refs. [@NIST; @NIST1]. The values of the ac polarizabilities obtained using Eq.(\[eq2\]) differ from our results in Table \[tab1\] in the relevant wavelength range by less than 0.1% which is significantly below the accuracy of the values themselves. We note that the formula (\[eq2\]) is not applicable for the lower wavelengths owing to the importance of the other $n^{\prime}p-ns$ resonances. For example, the 420 nm wavelength corresponds to $6p-5s$ transition in Rb; therefore, the ground state polarizability at this frequency will be dominated by the $n^{\prime}=6$ term. The frequency-matching criteria for mixed species ================================================= We have also conducted a search of the wavelength values at which two species of optically-trapped alkali atoms ($^6$Li, $^7$Li, $^{23}$Na, $^{40}$K,$^{41}$K, $^{87}$Rb, $^{133}$Cs, and $^{210}$Fr) would have the same frequencies of oscillation in a common optical trap, i.e. $\lambda$ such that $$s\equiv \sqrt{\frac{\alpha_1(\lambda)}{m_1}\frac{m_2}{\alpha_2(\lambda)}}\approx 1, \label{match}$$ where $m_i$ is the atomic weight [@mass]. The wavelengths of those selected matches and their uncertainties are listed in Table \[tab3\]. We have omitted all matches at wavelengths below 500 nm. While we had found numerous matches in the range from 300 nm to 500 nm they are difficult to place accurately since the accuracy of our calculation is lower in this region. The width of the matches in the $300 - 500$ nm region is generally very narrow and the corresponding values of the polarizabilities are small, making them of limited experimental use. We define the width of the match as the wavelength range where the parameter $s$ given by Eq. [(\[match\])]{} is between 0.9 and 1.1. The width of the match is very narrow if the corresponding wavelength is in the immediate vicinity of the resonance. We have also omitted most matches at higher wavelengths from Table \[tab3\] that are close to the resonance. Note that we listed the values of the wavelengths in the air. While the difference between air and vacuum wavelengths is small in the considered region, $0.15-0.4$ nm, it is sufficient to cause changes in the frequency-dependent polarizabilities (at some wavelengths) which exceed the uncertainties of the calculations. We used the following formula to compute the wavelength in air from the vacuum wavelength[@refractive] $$\begin{aligned} \lambda_{\rm air}&=&\lambda_{v}/(1+10^{-8}(8060.51+\frac{2480990}{132.274-\sigma ^2\times 10^6} \nonumber\\ &+&\frac{17455.7}{39.32957-\sigma ^2\times 10^6})).\end{aligned}$$ The wavenumber $\sigma$ is here to be expressed in reciprocal nm. We also list the values of the ac polarizabilities and their uncertainties in Table \[tab3\]. The uncertainty in the wavelength of the match results from the uncertainties in the values of the frequency-dependent polarizabilities $\alpha_1(\lambda)$ and $\alpha_2(\lambda)$ given in Table \[tab3\]. The uncertainties in the values of $\alpha(\lambda)$ are found to be independent of $\lambda$ within the range of the match to a very good precision but are generally different for $\alpha_1$ and $\alpha_2$. We calculate the resulting uncertainties $\delta \lambda_1$ and $\delta \lambda_2$ due to uncertainties in the values of $\alpha_1$ and $\alpha_2$ separately and add the results in quadrature to obtain the final uncertainty $\delta \lambda$. To evaluate $\delta \lambda_1$, we calculate the $\alpha^{\pm}_1(\omega)/m_1 = [\alpha_1(\lambda)\pm \delta \alpha_1] /m_1$ and then determine the wavelengths $\lambda^{\pm}$ at which $\alpha^{\pm}_1(\lambda)/m_1=\alpha_2(\lambda)/m_2$. The maximum of the differences between the initial match wavelength and $\lambda^{\pm}$ is taken to be the uncertainty $\delta \lambda_1$. The evaluation of $\delta \lambda_2$ is done in the same way. We illustrate several of the matches given in Table \[tab3\] in Fig. \[fig1\] and Fig. \[fig2\]. The ratios $\alpha(\lambda)/m$ of frequency-dependent polarizabilities to the atomic weights for Li and Na in the wavelength range 500-800 nm are plotted in Fig. \[fig1\]. The corresponding matches are shown by the arrows and the corresponding wavelengths are given on the plot. The values $\frac{\alpha}{m}$ for $^6$Li and $^{23}$Na match at three wavelengths in the wavelength range plotted on Fig. \[fig1\]. Two of the matches, at 589.397 nm and 670.79 nm, are very close to the resonances and, therefore, are not listed in Table \[tab3\] as discussed before. The match areas for those cases are extremely narrow and the polarizability of the corresponding alkali is very sensitive to very small changes in the wavelength. Fig. \[fig2\] shows similar plot to illustrate the wavelengths at which $^{40}$K and $^{87}$Rb have the same frequencies of oscillation in a common optical trap in the range 740-820 nm. Conclusions =========== In summary, we conducted a systematic study of the ground state frequency-dependent polarizabilities of the alkali atoms from ultraviolet through infrared spectral regions. The values of the ac polarizabilities and their uncertainties are calculated for a number of wavelengths, including the wavelengths of common infrared lasers. A combination of high-precision measurements of the ratios of frequency-dependent polarizabilities of different species could provide excellent tests of the current experimental and theoretical values of the electric-dipole matrix elements in alkali-metal atoms. We provide formulas and the necessary parameters for the accurate calculation of the ac polarizabilities for all alkali-metal atoms at wavelengths above 800 nm. Finally, we list selected wavelengths at which two species of alkali atoms would have the same oscillation frequencies in a common optical trap.\ This work was performed under the sponsorship of the U.S. Department of Commerce, National Institute of Standards and Technology. [20]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , , , , **, (). , , , , , , , , (). , , , , , , , , (). , , , , , , , (). , , , , (), . , , , , , (), . , , , , , (). , , , , (). , , , , (). , , , , (). , , , , (). , , , , (). , , , , (). , , , , , (). , , vol. of (, ). , , , , , , , , , (). , , , , (). , , , , **, ().	{ "pile_set_name": "ArXiv" }
--- abstract: 'A method for the computation of the radiative momentum transfer in the Pioneer 10 & 11 spacecraft due to the diffusive and specular components of reflection is presented. The method provides a reliable estimate of the thermal contribution to the acceleration of these deep space probes and allows for a Monte-Carlo analysis from which an estimate of the impact of a possible variability of the parameters. It is shown that the whole anomalous acceleration can be explained by thermal effects. The model also allows one to estimate the expected time evolution of the acceleration due to thermal effects. The issue of thermal conduction between the different components of the spacecraft is discussed and confirmed to be negligible.' author: - 'Orfeu Bertolami[^1], Frederico Francisco[^2], Paulo J. S. Gil[^3] and Jorge Páramos[^4]' bibliography: - 'Pio\_Thermal\_DyCoSS.bib' title: 'THE CONTRIBUTION OF THERMAL EFFECTS TO THE ACCELERATION OF THE DEEP-SPACE PIONEER SPACECRAFT' --- Introduction ============ The Pioneer anomaly (PA) has been an open issue in physics for over a decade: it consists of an presumably constant sun-bound acceleration on the Pioneer 10 and 11 deep-space probes, first put forward in a 1998 work by a team from the Jet Propulsion Laboratory (JPL) [@Anderson1998] and further scrutinized in a second paper, which settled the anomaly as a cosntant acceleration $a_{\rm{Pio}} = (8.74 \pm 1.33) \times 10^{-10} ~ \rm{m/s^2}$ [@Anderson2002]. Independent data analyses [@Markwardt2002; @Levy2009; @Toth2009] have confirmed the existence of the anomalous acceleration, with at least two of them presenting results consistent with a non-constant acceleration [@Markwardt2002; @Toth2009]. The PA has attracted much attention from the scientific community throughout the last decade. Solutions range from conventional [@Katz1999; @Scheffer2003] to new physics explanations [@Bertolami2004; @Reynaud2005; @Moffat2006; @Bertolami2007]. For instance, it has been shown that the Kuiper Belt cannot be responsible for the anomalous acceleration [@Anderson2002; @Bertolami2006]. Rather surprisingly, the account of systematic effects presented in Ref. [@Anderson2002] dismissed altogether a significant contribution due to thermal effects. In opposition, it has been argued, albeit on a qualitative basis, that the on-board thermal power could indeed account for the anomalous acceleration [@Katz1999; @Scheffer2003]. Thus, the need for a quantitative description of the thermal effects became evident. Three independent efforts have been undertaken for the past few years, with the first results being released in 2008 by the Lisbon team. That work was based on the distribution of point-like Lambertian and isotropic radiation sources, validated by a set of test cases [@Bertolami2008; @Bertolami2010]. The results suggest that thermal emissions of the spacecraft itself [@Bertolami2008; @Bertolami2010] could explain most if not all of the observed acceleration. These results have been confirmed by the finite-element modelling by the ZARM team [@Rievers2009; @Rievers2010]. It has also been reported that an analysis is underway by the JPL based team [@Toth2008]. Here, we consolidate our previous work by presenting a review of the results previously obtained through a direct modelling of reflection, complementing the estimates based on surface reflectivity [@Bertolami2008; @Bertolami2010] which, when fully accounted, can explain the whole PA [@Francisco2012]. In addition, we discuss the effects of heat conduction between the main components and carry out a parametric analysis in order to establish reliable bounds for the results, as also discussed in Ref. [@Francisco2012]. Radiative Momentum Transfer =========================== Point-like Source Method {#sec:radmomtransf} ------------------------ In the present study, it is paramount that the approach chosen allows for the quick and reliable analysis of different scenarios and contributions. Motivated by the limitations inherent to the characterization of the anomalous acceleration itself, we adopted an approach that maintains a high degree of computational flexibility and speed, as outlined in Ref. [@Bertolami2008]. This uncertainty extends to the fact that both a constant acceleration and a linearly decaying one are consistent with the telemetry data [@Markwardt2002; @Toth2009], with the inclusion of a so-called “jerk term” (the derivative of the acceleration) found to be compatible with the expected temporal variation of a recoil force due to heat generated on board, with a half-life of $\sim 50~yr$. This makes the hypothesis of a thermal origin for the PA as the main culprit for the anomalous acceleration, as it would inevitably lead to a decay with at least the same rate as the power available onboard, which is derived from two $Pu$ radio-thermal generators (RTGs) with a half-life of $88~yr$. Possible causes for an enhanced decay include [e.g.]{} degradation of thermo-couples, stepwise shutdown of some systems and instruments, [etc.]{} [@Markwardt2002]. With all this in mind, we designed a method that keeps the main physical features under control and available for scrutiny. The possibility that this simplicity and transparency is achieved at the expense of accuracy is offset by a series of test cases that are carried out to demonstrate the robustness of the results [@Bertolami2008; @Bertolami2010], validating the approach and showing that the level of accuracy is still much higher than that of the characterization of the acceleration itself. Furthermore, the method is further put to the test in a parametric analysis of the problem that weighs the relative importance of the different parameters involved. The modelling is based on a distribution of point-like thermal radiation sources to account for the overall emission of the spacecraft. The formulation of emission and reflection is made in terms of the Poynting vector-field. We thus begin with the vector-field descriptions for the radiation emitting surfaces, modelled as Lambertian sources. The time-averaged Poynting vector field for a Lambertian source located at $\mathbf{x}_0$ is $$\label{lambertian} \mathbf{S}(\mathbf{x})={W \cos \theta \over \pi \|\|\mathbf{x}-\mathbf{x}_0\|\|^2}{\mathbf{x}-\mathbf{x}_0 \over \|\|\mathbf{x}-\mathbf{x}_0\|\|},$$ where $W$ is the emissive power and $\theta$ is the angle with the surface normal. The value of $\cos \theta$ is the normalized inner product between the unitary emitting surface normal $\mathbf{n}$ and the emitted ray vector $(\mathbf{x}-\mathbf{x}_0)$. The amount of energy illuminating a given surface $E_{\rm ilum}$ can be obtained by computing the Poynting-vector flux through the surface: $$E_{\rm ilum} = \int \mathbf{S} \cdot \mathbf{n}_{\rm ilum}~ dA,$$ where $\mathbf{n}_{\rm ilum}$ is the normal vector of the illuminated surface. The thermal radiation (infrared radiation) illuminating a surface will yield a force on that surface. This force per unit area is the radiation pressure $p_{\rm rad}$, given by $$p_{\rm rad}={\mathbf{S} \cdot \mathbf{n}_{\rm ilum} \over c},$$ that is, the energy flux divided by the speed of light. This result should be multiplied by a factor $\alpha$ that varies between $\al =1$ for full absorption and $\al = 2$ for full reflection, which allows for an estimate of the reflection (as assessed in Refs. [@Bertolami2008; @Bertolami2010]). However, a more rigorous treatment of reflection is presented in the next sections, following Ref. [@Francisco2012]. Integrating the radiation pressure on a surface, we obtain the exerted force $$\label{force_integration} \mathbf{F} = \int {\mathbf{S} \cdot \mathbf{n}_{\rm ilum} \over c} {\mathbf{S} \over \|\|\mathbf{S}\|\|} dA.$$ The interpretation of this integration may not always be straightforward: to obtain the force exerted by the radiation on the emitting surface, the integral must be taken over a closed surface encompassing the latter. Analogously, the force exerted by the radiation on an illuminated surface requires an integration surface that encompasses it. Also, considering a set of emitting and illuminated surfaces implies the proper account of the effect of the shadows cast by the various surfaces, which is then subtracted from the estimated force on the emitting surface. One may then straightforwardly read the thermally induced acceleration, $$\mathbf{a}_{\rm th}={\sum_i \mathbf{F}_i \over m_{\rm pio}}.$$ Reflection Modelling – Phong Shading {#sec:phong} ------------------------------------ We suggest that in order to accurately model the reflections caused by the geometric configuration of the Pioneer 10 and 11 probes [@Francisco2012], one can adopt a method developed in the 1970’s by Bui Tuong Phong at the University of Utah and published in his Ph.D. thesis [@Phong] and known as Phong Shading. This includes a set of techniques and algorithms commonly used to render the illumination of surfaces in 3D computer graphics. The method comprises two distinct parts: - A reflection model including diffusive and specular reflection, known as Phong reflection model; - An interpolation method for curved surfaces modelled as polygons, known as Phong interpolation. The Phong reflection model is based on an empirical expression that yields the illumination of a given point in a surface $I_p$ as $$I_p=k_a i_a + \sum_{m \in \text{lights}} \left[k_d (\mathbf{l}_m \cdot \mathbf{n})i_d + k_s (\mathbf{r}_m \cdot \mathbf{v})^{\alpha} i_s \right],$$ where $k_a$, $k_d$ and $k_s$ are the ambient, diffusive and specular reflection constants, $i_a$, $i_d$ and $i_s$ are the respective light source intensities, $\mathbf{l}_m$ is the direction of the light source $m$, $\mathbf{n}$ is the surface normal, $\mathbf{r}_m$ is the direction of the reflected ray, $\mathbf{v}$ is the direction of the observer and $\alpha$ is a “shininess” constant (the larger it is, the more mirror-like the surface is). This method provides a simple and straightforward method to model the various components of reflection, as well as a fairly accurate accounting of the thermal exchanges between the surfaces on the Pioneer spacecraft. There is no fundamental distinction between the treatment of infrared radiation, relevant for the Pioneer problem, and visible light, for which the method was originally developed, as long as one allows for a wavelength dependence of the above constants. Given the presented thermal radiation in subsection \[sec:radmomtransf\], the Phong shading methodology can be easily adapted from a formulation based on intensities (energy per surface unit per surface unit of the projected emitting surface) to one based on the energy-flux per surface unit (the Poynting vector). This is achieved through the following expressions for the diffusive component of the reflection $$\label{diffusive_reflection} \mathbf{S}_{\text{rd}}(\mathbf{x},\mathbf{x}')={k_d \|\mathbf{S}(\mathbf{x}')\cdot \mathbf{n}\| \over \pi \|\|\mathbf{x}-\mathbf{x}'\|\|^2} (\mathbf{n} \cdot (\mathbf{x}-\mathbf{x}')) {\mathbf{x}-\mathbf{x}' \over \|\|\mathbf{x}-\mathbf{x}'\|\|},$$ and the specular component $$\label{specular_reflection} \mathbf{S}_{\text{rs}}(\mathbf{x},\mathbf{x}')={k_s \|\mathbf{S}(\mathbf{x}')\cdot \mathbf{n}\| \over {2 \pi \over 1+ \alpha} \|\|\mathbf{x}-\mathbf{x}'\|\|^2} (\mathbf{r} \cdot (\mathbf{x}-\mathbf{x}'))^{\alpha} {\mathbf{x}-\mathbf{x}' \over \|\|\mathbf{x}-\mathbf{x}'\|\|}.$$ In both cases, the reflected radiation field depends on the incident radiation field $\mathbf{S}(\mathbf{x}')$ ($\mathbf{x}'$ is a point on the reflecting surface) and on the reflection coefficients $k_d$ and $k_s$, respectively. Using Eqs. (\[diffusive\_reflection\]) and (\[specular\_reflection\]), we can compute the reflected radiation field by adding up these diffusive and specular components. From the emitted and reflected radiation, the irradiation of each surface is computed and, from it, a calculation of the force can be performed through Eq. (\[force\_integration\]). This formulation allows for the determination of the force on the whole spacecraft, accounting for radiation that is reflected and absorbed by the various surfaces, as well as that which is propagated into space. In the actual modelling of the spacecraft, once the radiation sources are in place, the first step is to compute the emitted radiation field and the respective force exerted on the emitting surfaces. This is followed by the determination of which surfaces are illuminated and the computation of the force exerted on those surfaces by the radiation. At this stage, we get a figure for the thermal force without reflections. The reflection radiation field is then computed for each surface and subject to the same steps as the initially emitted radiation field, leading to a determination of thermal force with one reflection. The only limitation to the iterative extension of this method to multiple reflection are the numerical integration algorithms and available computational power. If necessary, each step can be simplified through a discretization of the reflecting surface into point-like reflectors. Pioneer Thermal Model ===================== Model Features -------------- We developed a geometric model of the Pioneer spacecraft that is compatible with the approach outlined in the previous sections ([cf.]{} Ref. [@Bertolami2008]). It makes a compromise between detail and simplicity, includes the most important features of the Pioneer spacecraft, namely: - The parabolic high-gain antenna; - The main equipment compartment behind the antenna; - Two RTGs, cylindrical in shape, each connected to the main compartment through a truss. The geometric model with its respective dimensions is depicted in Fig. \[pioneer\_schematic\]. The model simplifies the minor surface features and details of the spacecraft. This simplification has been tested through specific test-cases discussed in Refs. [@Bertolami2008; @Bertolami2010], which show that its effect on the final result can be safely ignored for the purposes of this study. The thermal radiation emissions modelling is achieved through a distribution of a few carefully placed point-like sources that mimic the actual emissions of the spacecraft as closely as possible. The fact that the Pioneer probes are spin-stabilised, considerably simplifies the problem, since the effect of all radial emissions is cancelled out in each full revolution of the spacecraft, leaving only contributions that are along the antenna’s axis (here taken as the $z$-axis). The main thermal radiation sources on the probe can be identified as the RTGs, where the main power source of the spacecraft is located, and the main equipment compartment, where the majority of the power is consumed. The RTGs can be easily and effectively modelled by two Lambertian sources, one at each base of the cylinder, as shown in Fig. \[RTG\_sources\]. The emissions from the source facing outwards will radiate directly into space in a radial direction and its contribution will cancel-out. However, the radiation emitted towards the centre of the spacecraft will be reflected by both the high-gain antenna and the main equipment compartment. The analysis of the main equipment compartment is divided between the front, back and lateral walls. The front wall of the latter (facing away from the sun and where the heat-dissipating louvers are located) will emit radiation directly into space, not illuminating any other surface. It can thus be modelled through a single radiation source, without impact on the final result. The side walls of this compartment are each modelled by four Lambertian sources, as seen in Fig. \[equipment\_sources\]. A previously conducted convergence analysis shows that this provides a reasonable degree of accuracy [@Bertolami2008; @Bertolami2010]. This radiation will reflect mainly on the high-gain antenna. Finally, we included the contribution of the back wall of the main equipment compartment (facing the high-gain antenna). The radiation from this wall will, in a first iteration, reflect off the antenna and add a contribution to the force in the direction of the sun, as depicted in Fig. \[back\_wall\]. This back wall was modelled using a set of six Lambertian sources evenly distributed in the hexagonal shape. The relevant contributions for this analysis can be summarized in Table \[pioneer\_components\], with each of them indexed for reference in the following sections. Emitting surface Reflecting surface Index ------------------------------------- ------------------------------------- ------- Lateral surface of main compartment High-gain antenna dish 1.1 RTG High-gain antenna dish 2.1 RTG Lateral surface of main compartment 2.2 Back surface of main compartment High-gain antenna dish 3.1 Front surface of main compartment none 4 \[pioneer\_components\] A brief remark about the experimental setup section attached to one of the sides of the main compartment. The full set of scientific instruments uses 24.5 W, out of a total of 120 W of dissipated electrical power (after launch) [@Turyshev2008]. Although not negligible, the preceding discussion indicates that disregarding this difference in geometry should not change the outcome significantly. Naturally, a more detailed geometrical modelling of the spacecraft would lead to a more refined value for the overall thermal output, but this is beyond the scope of the current work. A more detailed discussion on the hypotheses used in the Pioneer model can be found in Refs. [@Bertolami2008; @Bertolami2010]. Conductive Contribution ----------------------- Since the components of the spacecraft are mechanically connected, wee should expect a certain amount of conductive heat flux between them, changing the amount of energy that each one radiates to maintain the equilibrium. The structural elements connecting different components of the spacecraft, namely, between the RTGs and the main compartment and the main compartment and the high gain antenna should be looked upon more closely. In both cases, considering the approximate thickness, conductivity of the truss bars and temperature gradients suggest that the effect of conduction between these components is always a small fraction (typically, a few Watts) of their respective powers. Let us consider what we can expect from the conductive effect. A difference in temperature is expected between different spacecraft elements e.g. the RTGs and the main compartment. Neglecting conduction, they can be considered approximately in equilibrium, with the energy dissipated from the RTGs and electrical equipment radiated away from themselves. However, there will be a certain amount of conductive heat flow between them through the truss. This would correspond to a transfer of power from the hotter to the colder elements (from the RTG to the main compartment). Realistically, the transferred heat will tend to increase the temperature of the colder elements and decrease the temperature of the hotter ones, decreasing the conductive heat flux, until a new equilibrium. This effect can be taken into account by a variation of the relative power distribution between elements, and since it is rather small (see below), it can be dealt with in the Monte-Carlo calculations. Focusing on the connection between the RTGs and the main compartment, where the effect of thermal conduction is expected to be greater, we can estimate the effect for a given reasonable difference of temperatures. Assuming a temperature near $0^\circ \rm C$ in the equipment compartment (a worst case scenario, since it is actually warmed by the electronics to $\sim 10^\circ \rm C$ [@temperature1; @temperature2]) and the RTGs near $150^\circ \rm C$, a temperature gradient of approximately $60 ~\rm K/m$ is obtained. Each truss is composed of three small diameter rods, that we estimate to have a total cross-section of the order of $10^{-4}~\rm m^2$, made out of aluminium that has a conductivity of approximately $240~\rm W/(m \cdot K)$. These figures would translate into a total conducted power of the order of $1~W$ (up to $4~W$ in more conservative estimates), which is clearly negligible, since it is two orders of magnitude below the power of the main compartment and three orders below the RTG power. The power conduction from the main compartment to the antenna would be even smaller, since the temperature gradient is also much smaller. This effect can then be safely disregarded in any further computations of the global effect and easily taken into account in the relative power distribution between spacecraft elements. Radiative and Reflective Contributions -------------------------------------- We can now compute the contribution of the individual components listed in Table \[pioneer\_components\]. This is done through the integration of Eq. (\[force\_integration\]) in three successive steps. First, the emitted radiation field given by Eq. (\[lambertian\]) is integrated along a closed surface, yielding the first-order effect of the emissions. Afterwards, the same radiation field is integrated along the illuminated surfaces, in order to subtract the shadow effect. Finally, the reflected radiation vector-field, given by Eqs. (\[diffusive\_reflection\]) and (\[specular\_reflection\]), is integrated along closed surfaces, adding the contribution from reflection. This process allows us to obtain the values for the force in terms of the emitted powers and reflection coefficients. As already pointed out, the results that follow are along the main antenna axis, since all radial components cancel-out. A positive figure indicates a sunward force. The computation of the contribution from the front surface of the main compartment (index 4) is fairly straightforward, since there are no reflections on other surfaces. For this reason, and the fact that this surface is perpendicular to the spacecraft’s spin axis, it is effectively modelled by a single radiation source, as indicated in Table \[sources\]. by replacing the position and surface normal direction in Eq. (\[lambertian\]) we obtain the emitted radiation field. The force exerted by the radiation field on the emitting surface is obtained by integrating Eq. (\[force\_integration\]) along a closed surface — in this case, chosen as a half-sphere centered at the location of the radiation source. The $z$ component of the resulting force on the emitted radiation is, as expected, given by $$F_4 = {2 \over 3} {W_{\rm front} \over c}.$$ Emitting Surface Source Position (m) Surface Normal (m) ---------------------- -------- -------------------------- -------------------- Front wall (index 4) 1 $(0,0,-0.343)$ $(0,0,-1)$ Lateral wall 1 $(0.572,0.2475,-0.172)$ $(1,0,0)$ (index 1.1) 2 $(0.572,0.0825,-0.172)$ $(1,0,0)$ 3 $(0.572,-0.0825,-0.172)$ $(1,0,0)$ 4 $(0.572,-0.2475,-0.172)$ $(1,0,0)$ RTG 1 $(2.5,0,0)$ $(-1,0,0)$ (index 2.1 & 2.2) 2 $(3.1,0,0)$ $(1,0,0)$ Back wall 1 $(0.381,0,0)$ $(0,0,1)$ (index 3.1) 2 $(0.191,0.33,0)$ $(0,0,1)$ 3 $(-0.193,0,33)$ $(0,0,1)$ 4 $(-0.381,0,0)$ $(0,0,1)$ 5 $(-0.191,-0.33,0)$ $(0,0,1)$ 6 $(0.191,-0.33,0)$ $(0,0,1)$ \[sources\] The radiation emitted from the lateral walls of the main compartment do mainly illuminate the high-gain antenna (index 1.1). Due to the already discussed axial symmetry of the problem, and neglecting the interaction with the small and far away RTGs, it is only necessary to model one of the six walls. The set of Lambertian sources used for one of these walls is indicated in Table \[sources\]. The $z$ component of the radiation field force on the emitting surface vanishes, as the emitting surface is orthogonal to the $z$-axis. Integrating Eq. (\[force\_integration\]) over the illuminated portion of the antenna dish, we get the force exerted on the illuminated surface, which accounts for the shadow effect. This gives a $z$ component of $-0.0738(W_{\rm lat} / c)$, where $W_{\rm lat}$ is the power emitted from the lateral walls — to be subtracted from the total force of the emitted radiation. The computation of diffusive reflection is made through Eq. (\[diffusive\_reflection\]), that returns the reflected Poynting vector-field $\mathbf{S}_{\text{rd}}(\mathbf{x},\mathbf{x}')$ due to the emitted radiation field $\mathbf{S}(\mathbf{x}')$, where $\mathbf{x}'$ is a point in the reflecting surface. The reflected radiation field is given for each point in the reflecting surface and must be integrated first over the reflecting surface itself, conveniently parameterized. This gives the overall reflected radiation field that is then integrated through Eq. (\[force\_integration\]) over a closed surface to obtain the force resulting from the reflected radiation. The procedure for specular reflection is entirely similar, except that Eq. (\[specular\_reflection\]) should be used to obtain the reflected radiation field before the integration. Integrating the vector-field representing radiation from the lateral walls of the main compartment reflecting on the high-gain antenna, we obtain a force result of $0.0537 k_{\rm d,ant}(W_{\rm lat} / c)$ for the diffusive component and $0.0089 k_{\rm s,ant}(W_{\rm lat} / c)$ for the specular component, where $W_{\rm lat}$ is the power emitted from the referred walls and $k_{\rm d,ant}$ and $k_{\rm s,ant}$ are the diffusive and specular reflection coefficients of the main antenna, respectively. The result for the contribution is given by adding the (vanishing) emitted radiation force, the shadow effect and both components of reflection, leading to $$F_{11} = {W_{\rm lat} \over c} (0.0738 + 0.0537 k_{\rm d,ant} + 0.0089 k_{\rm s,ant}).$$ The emissions from the RTGs were modelled by a Lambertian source at each base of each cylindrical shape RTG, as listed in Table \[sources\]. Similarly to the lateral walls, only the modelling of one RTG is required, since the effect of the radial components cancel-out with each revolution of the probe. It can also be easily shown that only the emissions from the base facing the centre of the spacecraft (source 1 of the RTG in Table \[sources\]) will impact on the acceleration along the $z$-axis. Emissions from the base facing outwards (source 2) are not reflected on any surface and its contribution vanishes when averaged over each revolution of the spacecraft. Using the same procedure, the generated force from the RTG emissions is thus given in terms of the power emitted from the RTG bases facing the centre of the spacecraft $W_{\rm RTGb}$. The force resulting from reflections on the antenna (index 2.1) is given by $$F_{21} = {W_{\rm RTGb} \over c} (0.0283 + 0.0478 k_{\rm d,ant} + 0.0502 k_{\rm s,ant}),$$ and the contribution from reflections on the lateral surfaces of the main equipment compartment is $$F_{22} = {W_{\rm RTGb} \over c} (-0.0016 + 0.0013 k_{\rm s,lat}),$$ where $k_{\rm d,ant}$, $k_{\rm s,ant}$, $k_{\rm d,lat}$ and $k_{\rm s,lat}$ are the respective reflection coefficients. As previously argued [@Bertolami2008], the contribution from radiation emitted from the back wall of the main compartment and reflecting in the space between this compartment and the antenna dish would be small. In an attempt to confirm this assumption, a computation was made using the method described above. The results ultimately show that this contribution cannot be overlooked. Considering one reflection from the antenna dish, the result in terms of the emitted power from the back wall of the main compartment $W_{\rm back}$, by $$F_{3} = {W_{\rm back} \over c} \left( -{2 \over 3} + 0.5872 + 0.5040 k_{\rm d,ant} + 0.3479 k_{\rm s,ant} \right).$$ Note that the $-{2 \over 3}{W_{\rm back} \over c}$ is the contribution from the emitted radiation and $0.5872 {W_{\rm back} \over c}$ is the effect of the antenna’s shadow. From the force computations, once the respective powers and reflection coefficients are inserted, the final result of the acceleration due to thermal dissipation mechanism follows: $$a_{\rm th} = {F_{11} + F_{21} + F_{22} + F_{3} +F_{4} \over m_{\rm Pio}},$$ where the mass of the spacecraft is taken at an approximate value $m_{\rm Pio} = 230~\rm{kg}$. This figure considers a total mass of $259~\rm{kg}$ at launch, including $36~\rm{kg}$ of hydrazine propellant that was partially consumed in the early stages of the mission [@Anderson2002]. Note that this is an approximate figure, since the actual masses for the Pioneer 10 and 11 would be slightly different due to different fuel consumptions. Available Power {#available_power} --------------- It was chosen from the earliest stages of this study to use the available onboard power as the independent variable in the computation of the thermally induced acceleration. This is justified due to the reasonably good knowledge of the available power — indeed, this is one of the few parameters with reliable data available throughout the operational lifetime of the probes — and also because power is the driving parameter determining the emitted thermal radiation. The power on board the Pioneer probes comes from the two plutonium RTGs. It is thus easy to compute the total power available, considering the $87.74~\rm{year}$ half-life of plutonium. According to Ref. [@Anderson2002], the total thermal power of the RTGs at launch was $2580~\rm{W}$, leading to a time evolution given by $$W_{\rm tot} = 2580 \exp \left(- {t \ln 2 \over 87.72} \right)~\rm{W}, \label{total_power}$$ with $t$ being the time in years from launch. Electrical power generation is ensured by a set of thermocouples located in the RTGs. Most of this power is consumed by the various systems and instruments located in the main equipment compartment, except for a small portion used by the radio signal. A good measurement of the electrical power is available through telemetry data [@Toth2008]. Knowing the total electrical power consumption, remaining unused power is mostly dissipated at the RTGs themselves, through its external surface and radiating fins. At launch, $120~\rm{W}$ of electrical power were being delivered to the main equipment compartment plus around $20~\rm{W}$ for the radio transmission to Earth, leaving $2420~\rm{W}$ of thermal power in the RTGs. It is also known from telemetry data that the electrical power decayed at a faster rate than thermal power, with its half-life being around $24~\rm{years}$. This would lead to an approximate time evolution of the electrical power in the equipment compartment given by $$W_{\rm equip} \approx 120 \exp \left(- {t \ln 2 \over 24} \right)~\rm{W}, \label{elec_power}$$ which is consistent with Fig. 11 in Ref. [@Toth2008]. The above considerations and the power values extracted from the available telemetry data for the latest stages of the mission — specifically, the reading for the twenty six years after launch (for the Pioneer 10, up to 1998) — are used as a baseline scenario. In a second stage of this study, the time evolution is taken into account, according to the reasoning developed in this section. Results and Discussion ====================== Parametric Analysis and the Effect of Conduction ------------------------------------------------ We now perform a static parametric analysis, in an attempt to establish a reliable estimate for the thermal acceleration at an instant 26 years after launch. The analysis resorts to a classic Monte-Carlo method, where a probability distribution is assigned to each variable and random values are then generated. A distribution of the final result ([i.e.]{} the acceleration) is then obtained. We establish a set of scenarios for the distribution of the emitted power throughout the different surfaces, while keeping the total power constant as $W_{\rm tot} = 2100~\rm{W}$ and the electrical power as $W_{\rm equip} = 56~\rm{W}$, leaving RTG thermal power at $W_{\rm RTG} = 2024~\rm{W}$ (assuming the power of the radio beam is still $20~\rm{W}$). These scenarios act as a baseline for the parametric analysis and are summarised in Table \[baseline\_results\]. The parameters that come into play in this setup are the power emitted from each surface, $W_{\rm RTGb}$, $W_{\rm front}$, $W_{\rm lat}$, $W_{\rm back}$, and the reflection coefficients $k_{\rm d,ant}$, $k_{\rm s,ant}$ and $k_{\rm s,lat}$. A quick analysis of Table \[baseline\_results\] allows us to draw a qualitative assessment: for instance, the of power emitted from the front wall $W_{\rm front}$ has a decisive influence in the final result. In contrast, the relevance of the specular reflection coefficient of the lateral wall $k_{\rm s,lat}$ is almost negligible. ---------- ---------------- ----------------- --------------- ---------------- ----------------- ----------------- ----------------- ------------------------- Scenario $W_{\rm RTGb}$ $W_{\rm front}$ $W_{\rm lat}$ $W_{\rm back}$ $k_{\rm d,ant}$ $k_{\rm s,ant}$ $k_{\rm s,lat}$ $a_{\rm th}$ $(\rm{W})$ $(\rm{W})$ $(\rm{W})$ $(\rm{W})$ $(10^{-10}~\rm{m/s^2})$ 1 $143.86$ $17.5$ $21$ $17.5$ $0$ $0$ $0$ $2.27$ 2 $143.86$ $40$ $8.73$ $7.27$ $0$ $0$ $0$ $4.43$ 3 $143.86$ $40$ $8.73$ $7.27$ $0.8$ $0$ $0$ $5.71$ 4 $143.86$ $40$ $8.73$ $7.27$ $0.6$ $0.2$ $0.4$ $5.69$ 5 $158.24$ $56$ $0$ $0$ $0.8$ $0$ $0.4$ $6.71$ ---------- ---------------- ----------------- --------------- ---------------- ----------------- ----------------- ----------------- ------------------------- \[baseline\_results\] For the static analysis at $t=26~\text{years}$, Scenario 4 is taken as a baseline, since it is the one more solidly based on physical arguments. The power emitted by the RTG bases facing the main compartment $W_{\rm RTGb}$ is generated from a Gaussian distribution with the mean value of $143.86~\rm{W}$ and a standard deviation of $25\%$ of this value. This allows for a significantly larger deviation than that considered in the top-bound scenario (Scenario 5), which had only a $10\%$ increase in the power of this surface. This is made so to account for any unanticipated anisotropies in the temperature distribution of the RTGs and enhance the confidence of the results. In what regards the main equipment compartment, the focus is on the power emitted by the louvers located in the front wall. The selected distribution for the parameter $W_{\rm front}$ is also normal, with the mean value at $40~\rm{W}$ (also corresponding to Scenario 4 and the standard deviation at $7.5~\rm{W}$, so that the 95% probability interval ($2 \sigma$) for the value of $W_{\rm front}$ is below the top figure of $56~\rm{W}$, which corresponds to all the power from the equipment being dissipated from the front wall. For the remaining surfaces of the equipment compartment, the power is computed at each instance so that the total power of the equipment is maintained at $56~\rm{W}$. The reflection coefficients for the antenna are allowed to take values in an uniform distributions bounded at $[0.6,0.8]$ for $k_{\rm d,ant}$ and $[0,0.2]$ for $k_{\rm s,ant}$, with the condition $k_{\rm d,ant}+k_{\rm s,ant}=0.8$, since this is the reference value for aluminium in infrared wavelengths. We also expect the specular component to be small, since the surface is not polished. Furthermore, if we allow for the possibility of surface degradation with time along the mission, the specular component would suffer a progressive reduction in favour of the diffusive component, a possibility that this analysis takes into account. We performed $10^4$ Monte Carlo iterations, which easily ensures the convergence of the result. The thermal acceleration estimate yielded by the simulation for an instant $26~\text{years}$ after launch, with a $95\%$ probability, is $$a_{\rm{th}}(t=26)=(5.8 \pm 1.3) \times 10^{-10} ~ \rm{m/s^2}.$$ This result is extracted from the approximately normal distribution shown in Fig. \[static\_hist\]. The conformity of the results to a normal distribution was confirmed by a Shapiro-Wilk normalcy test with a $p$-value $\sim 0$. These results represent $44\%$ and $96\%$ of the reported value $a_{\rm{Pio}}=(8.74 \pm 1.33) \times 10^{-10}~\rm{m/s^2}$ which was obtained under the assumption that it was constant acceleration. At the very least, this strongly suggests a preponderant contribution of thermal effects to the PA. We can now repeat the parametric analysis, but assuming that heat conduction takes place. In this case, we adjust the power distribution so that around $4~\rm W$ of thermal power are being transferred from the RTGs to the main compartment and run the simulation with the same 10000 iterations. The results show a slight increase in the overall acceleration of around $5\%$, [i.e.]{}, $$a_{\rm{th}}(t=26,{\rm cond})=(6.05 \pm 1.4) \times 10^{-10} ~ \rm{m/s^2}.$$ The resulting distribution is shown in Fig. \[static\_hist\_cond\]. This estimate represents an upper bound for the conductive heat transfer, confirming that this effect does not have a dramatic impact on the final result and does not alter any conclusion that one may or may not draw from them. Time Evolution -------------- The last issue of our analysis concerns the expected time evolution of the thermal acceleration experienced by the Pioneer 10 and 11 spacecraft. A first estimate can be obtained by extrapolating the static results with the time evolution of electric power, using Eqs. (\[total\_power\]) and (\[elec\_power\]). Results of this extrapolation are shown as the dotted line in Fig. \[time\_evolution\], with the approximate exponential decay of the available power translated into a similar trend in the evolution of the thermally induced acceleration. This extrapolation, however, does not account for the possible temporal variation of some parameters — particularly, the power distribution throughout the different surfaces or their reflection coefficients. This could be accounted by a simulation of the full span of the missions ([i.e.]{} a large number of consecutive simulations), with a specific prescription for the variability of these parameters. Such task will be addressed in the future. For now, we have chosen a somewhat simpler approach to grasp of the possibility discussed above: we apply the Monte-Carlo static analysis to only two earlier moments of the mission. Each simulation produces a central value, with top and lower bounds that are then fitted to an exponential trend, leading to an estimated time evolution of the thermal acceleration. The chosen instants for the earlier static analysis were at $t=8 ~\text{years}$ and $t=17 ~\text{years}$ after launch, corresponding, respectively, to the 1980 and 1989 values for the Pioneer 10. The 1980 date is approximately when the effect of the solar radiation pressure dropped below $5 \times 10^{-10}~\rm{m/s^2}$ [@Anderson2002]. The procedure is similar to the analysis of the previous subsection, but using the 1980 and 1989 available power values as a base for the choice of the distributions. The resulting thermal acceleration is, for $t=8 ~\text{years}$, $$a_{\rm{th}}(t=8)=(8.9 \pm 2) \times 10^{-10} ~ \rm{m/s^2},$$ corresponding to the same $95\%$ probability, and for $t=17 ~\text{years}$ $$a_{\rm{th}}(t=17)=(7.1 \pm 1.6) \times 10^{-10} ~ \rm{m/s^2}.$$ Using the three static estimates presented above, it is now possible to produce a time evolution based on a fit to an exponential decay. This is performed for the mean value, top-bound and lower-bound of the acceleration, always based on a $95\%$ probability degree. The curve fit for the mean, upper and lower values of the thermal acceleration reads $$a_{\rm th} = [(1.07 \pm 0.24) \times 10^{-9}] \exp (-0.0240t)~\rm{m/s^2},$$ with $t$ corresponding to the time after launch in $\text{years}$. The time evolution resulting from any of these scenarios corresponds to a decay with a half-life of approximately $60~\text{years}$, related to the nuclear decay of the plutonium in the RTGs and the faster decay rate of the electrical power, already discussed in Section \[available\_power\]. The graphic representation of the band of values predicted by our model is shown in Fig \[time\_evolution\] (dark grey region) and compared with the values indicated by non-constant results for the anomalous acceleration in Refs. [@Toth2009; @Markwardt2002] (light grey region). Conclusion and Outlook ====================== The method developed to account for reflection on the surfaces of the Pioneer spacecraft allows for the modelling of thermal radiation with increased accuracy, while maintaining the desired simplicity and computational speed of our previously proposed approach [@Bertolami2008]. This new tool allows for a successful modelling of the most relevant features of the Pioneer spacecraft concerning thermal effects and their impact on the resulting acceleration. The results provided by the developed method, based on Phong shading, globally confirm those previously obtained in Refs. [@Bertolami2008; @Bertolami2010] : the acceleration resulting from thermal effects has an order of magnitude that is compatible with the constant anomalous acceleration reported in Ref. [@Anderson2002]. We believe that this problem is especially suited to the chosen approach, considering its specific features. Furthermore, this Phong shading method is capable of being adapted for future studies of radiation momentum transfer in other spacecraft. The main difficulty that this problem sets has always been the lack of sufficient and reliable data for a detailed engineering modelling of the spacecraft’s flight conditions, despite various claims otherwise. We have overcome this barrier through a parametric analysis that takes into account a wide range of different scenarios. Indeed, this contribution clearly shows how an extra effect not previously taken into account, the thermal conduction between spacecraft elements, can be easily integrated in our model to account for its impact through the Monte-Carlo simulations. It turns out that this effect does not have a significant impact and does not change the overall conclusion. This further demonstrates the value of the chosen strategy that allowed us to present a range of probable values for the thermal effects, which apparently have a signature that is compatible with the Pioneer anomalous acceleration. This conclusion has also recently been confirmed by the team working in Bremen using a finite-element method [@Rievers2011]. This, we believe, brings us significantly closer to a solution for this problem in terms of onboard thermal effects. Aknowledgments ============== The work of OB and JP is partially supported by the Fundação para a Ciência e Tecnologia (FCT), through the project PTDC/FIS/111362/2009. The work of FF is sponsored by the FCT, under the grant BD 66189/2009. [^1]: Departamento de Física e Astronomia, Faculdade de Ciências, Universidade do Porto, Rua do Campo Alegre 687, 4169-007, Porto, Portugal and Instituto de Plasmas e Fusão Nuclear, Instituto Superior Técnico, Universidade Técnica de Lisboa, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal. [^2]: Instituto de Plasmas e Fusão Nuclear, Instituto Superior Técnico, Universidade Técnica de Lisboa, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal. [^3]: Departamento de Engenharia Mecânica and IDMEC - Instituto de Engenharia Mecânica, Instituto Superior Técnico, Universidade Técnica de Lisboa, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal. [^4]: Instituto de Plasmas e Fusão Nuclear, Instituto Superior Técnico, Universidade Técnica de Lisboa, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We give a new proof of a result of Rick Kenyon that the probability that an edge in the middle of an $n \times n$ square is used in a loop-erased walk connecting opposites sides is of order $n^{-3/4}$. We, in fact, improve the result by showing that this estimate is correct up to multiplicative constants.' author: - '[^1]' date: 'University of Chicago' title: --- Introduction ============ Loop-erased random walk is a process obtained by erasing loops from simple random walk. Although the process can be defined for arbitrary Markov chains, we will discuss the process only on the planar integer lattice ${{\mathbb Z}}^2 = {{\mathbb Z}}+ i {{\mathbb Z}}$. We start this paper by stating our main result. Let $$A_n = \{j+ik \in {{\mathbb Z}}+ i {{\mathbb Z}}: -n + 1 < j < n, - n < k < n\} ,$$ $$\p A_n = \{z\in {{\mathbb Z}}^2: {{\rm dist}}(z,A_n) = 1 )\}.$$ Let $\paths_n$ denote the set of nearest neighbor paths $ \omega = [\omega_0,\ldots,\omega_k] $ with $\Re[\omega_0] = -n, \Re[\omega_k] = n+1$ and $\{\omega_1,\ldots, \omega_{k-1}\} \subset A_n$. We write $\|\omega\| = k$ for the number of steps, and let $p(\omega) = 4^{-\|\omega\|}$ be the simple random walk measure. Let $$f(n) = \sum_{\omega \in \paths_n} p(\omega) .$$ It is known that $\lim_{n \rightarrow \infty} f(n) = c_1 \in (0,\infty)$ (see, e.g., [@LLimic Proposition 8.1.3]), where the constant $c_1$ can be given in terms of the Green’s function of Brownian motion on a domain bounded by a square. Let $\saws_n$ denote the set of self-avoiding walks (SAWs) $\eta = [\eta_0,\ldots,\eta_k] \in \paths_n$. For each $\omega \in \paths_n$ there is a unique path $L(\omega) \in \saws_n$ obtained by chronological loop-erasing (see [@LLimic Chapter 9] for appropriate definitions). The loop-erased measure $\hat p_n(\eta)$ is defined by $$\hat p_n(\eta) = \sum_{\omega \in \paths_n, L(\omega) = \eta } \; p(\omega) .$$ Note that $$\sum_{\eta \in \saws_n} \hat p_n(\eta) = f(n) .$$ Let $\saws_n^+$ denote the set of $\eta \in \saws_n$ that contain the directed edge $[0,1]$ and $\saws_n^-$ those that contain $[1,0]$. Let $\saws_n^* = \saws_n^- \cup \saws_n^+$ be the set of $\eta \in \saws_n$ that contain the edge $[0,1]$ in either direction. We write $a_n \asymp b_n$ to mean that $a_n/b_n$ and $b_n/a_n$ are uniformly bounded. The goal of this paper is to prove the following theorem. As $n \rightarrow \infty$, $$\label{nov20.1} \sum_{\eta \in \saws_n^} \hat p_n(\eta) \asymp n^{-3/4} .$$ With a little more work, we could establish the existence of the limit $$\lim_{n \rightarrow \infty} n^{3/4} \sum_{\eta \in \saws_n^} \hat p_n(\eta),$$ but we will not do it here. (Our argument would not give the value of the limit. While we believe we might be able to some of the relevant asymptotic constants, we definitely do not know how to compute the value of the limit in below.) Our result can be considered a strengthening of a result of Kenyon [@Kenyon] who proved that $$\label{kenyonresult} \sum_{\eta \in \saws_n^} \hat p_n(\eta) \approx n^{-3/4} ,$$ where $\approx$ indicates that the logarithms of both sides are asymptotic. His proof used the relationship between loop-erased walks and two other models, dimers and uniform spanning trees. Another proof of was given by Masson [@Masson] using the relationship between loop-erased walk and the Schramm-Loewner evolution ($SLE$). We do not need to make reference to any of these models in our proof of . There are two main steps. - A combinatorial identity is proved which writes the left-hand side of in terms of simple random walk quantities. - The simple random walk quantities are estimated. Our computation to obtain the exponent $3/4$ uses the Brownian loop measure to estimate the random walk loop measure. This is in the spirit of Kenyon’s calculations [@Kenyon] since the loop measure is closely related to the determinant of the Laplacian. Although the proof is self-contained (other than some estimates for simple random walk) it does use a key idea from Kenyon’s work as discussed in [@KWilson Section 5.7]. For each random walk path $\omega$, we let $J(\omega)$ be the number of times that the path crosses an edge of the form $[-ki,-ki+1]$ or $[-ki+1,-ki]$ where $k$ is a positive integer. Let $q(\omega) = (-1)^{J(\omega)} \, p(\omega)$. Let $Y_{+}(\omega)$ denote the number of times that $\omega$ uses the directed edge $[0,1]$, $Y_-(\omega)$ the number of times that $\omega$ uses the directed edge $[1,0]$, and $Y(\omega) = Y_+(\omega) - Y_-(\omega)$. The combinatorial identity is obtained by writing the quantity $$\label{Lambda} \Lambda_n = \sum_{\omega \in \paths_n} q(\omega) \, Y(\omega) = \sum_{\omega \in \paths_n} p(\omega) \,(-1)^{J(\omega) }\, Y(\omega) .$$ in two different ways. The paper is written using the perspective of loop-erased walk in terms of the random walk loop measure as in [@LLimic Chapter 9]. We start by reviewing this perspective in Section \[loopsec\] and then we prove the identity in Section \[identsec\]. Section \[rwloopsec\] discusses the random walk estimates. After the main result is proved, we generalize the combinatorial identity to random walk starting at any two points on $\p A_n$. The argument is essentially the same, and one can compute the dependence of the probability on the starting points. A somewhat analogous calculation was done in [@Kenyon], where the boundary was fixed but the interior point was allowed to vary. In the last section, this dependence on the boundary point is explained in terms of the scaling limit of loop-erased random walk, the Schramm-Loewner evolution with parameter $SLE_2$. Here we do not give all details. One of the main motivations for doing the estimates in this paper is to show that the loop-erased random walk converges to $SLE_2$ in the [natural parametrization]{} [@LR; @LS]. Up-to-constant estimates for the loop-erased walk probability can be viewed as a step in the program to establish this result. Random walk loop measure {#loopsec} ======================== The random walk loop measure is a measure on unrooted random walk loops. A rooted loop is a nearest neighbor path $$\label{nov21.1} l = [l_0,l_1,\ldots,l_{2k}]$$ with $k \geq 0$ and $l_0 = l_{2k}$. We call $l_0$ the root of the loop. An unrooted loop $\bar{l}$ is an equivalence class of rooted loops with $k > 0$ under the equivalence $$[l_j,l_{j+1},\ldots,l_{2k},l_1,l_2,\ldots,l_j] \sim [l_0,l_1,\ldots,l_{2k}]$$ for all $j$. Note that the orientation of the loops is maintained. The random walk loop measure $m$ is defined by $$m(\bar l) = 4^{-\|\bar l\|} \,\frac{d(\bar l)}{\|\bar l\|},$$ where $d(\bar l)$ is the number of rooted loops in the equivalence class of the unrooted loop $\overline l$. Note that $d(\bar l)$ is always an integer dividing $\|\bar l\|$. In a slight abuse of notation, if $l$ is a loop and $A \subset {{\mathbb Z}}^2$, we write $l \subset A$ to mean that the vertices of $l$ are contained in $A$ and $l \cap A$ for the set of vertices in $A$ that $l$ visits. There is an equivalent way of defining this measure that is sometimes useful. Enumerate ${{\mathbb Z}}^2 = \{v_1,v_2,\ldots,\}$ and let $V_n = \{v_1,\ldots,v_n\}$. We define a different measure on rooted loops by assigning to each (rooted) loop as in with $ k > 0, l \subset V_n ,$ and $l_0 = v_n$ measure $s^{-1} \, 4^{-2k}$ where $s = \#\{j; 1 \leq j \leq 2k, l_j = v_n\}$. This induces a measure on unrooted loops by summing over rooted loops that generate an unrooted loop. (The factor $s^{-1}$ compensates for the fact that several rooted loops give the same unrooted loop.) One can check that the induced measure on unrooted loops is the same as the loop measure above regardless of which enumeration is chosen. For computations it is often convenient to choose an enumeration in which $\|v_j\|$ is nondecreasing. If $V = \{v_1,\ldots,v_k\} \subset A \subsetneq {{\mathbb Z}}^2$, we define $$F_V(A) = \exp \left\{\sum_{\bar l \subset A, \bar l\cap V \neq \emptyset } m(\bar l)\right\} = \prod_{j=1}^k G_{U_j}(v_j,v_j) .$$ Here $U_j = A \setminus \{v_1,\ldots,v_{j-1}\}$ and $G_U $ denotes the usual random walk Green’s function in the set $U$. The loop-erased measure satisfies [@LLimic Proposition 9.5.1] $$\label{jan19.1} \hat p_n(\eta) = p(\eta) \, F_\eta(A_n) .$$ We can also define a loop measure using the signed weight $q(\omega) = (-1)^{J(\omega) } \, p(\omega) $. The quantities $J(l), Y(l)$ as defined in the introduction are functions of the unrooted loop $\bar l$. (Note that $Y(l)$ does depend on the orientation of $l$, so it is important that we are considering oriented, unrooted loops.) Let $\looper_A$ denote the set of unrooted loops $\bar l \subset A$ such that $J(\bar l)$ is odd. If $V \subset A$, let $\looper_{A,J}$ denote the set of unrooted loops $\bar l \in \looper_A$ that intersect $V$. Let $$Q_V(A) = \exp\left\{\sum_{\bar l \subset A, \bar l \cap V \neq \emptyset } m(\bar l) \, (-1)^{J(\bar l)}. \right\} = F_V(A) \, \exp \left\{-2m(\looper_{A,J}) \right\}. $$ As in the case for $F$, if $V = \{v_1,\ldots,v_k\} \subset A$, then $$\label{jan19.2} Q_V(A) = \prod_{j=1}^k g_{U_j}(v_j,v_j) ,$$ where $U_j = A \setminus \{v_1,\ldots,v_{j-1}\}$ and $$g_{U}(v_j,v_j) = \sum_{l}q(l) = \sum_{l} (-1)^{J(l)} \, p(l)$$ where the sum is over all (rooted) loops $l$ from $v_j$ to $v_j$ staying in $U$. In particular, if $\eta \in \saws_n$, then when the algebraic computation which gives is applied to $q$, we get $$\sum_{\omega \in \paths_n, L(\omega) = \eta} q(\omega) = q(\eta) \, Q_\eta(A_n) .$$ This implies that $$\sum_{\omega \in \paths_n, L(\omega) = \eta} (-1)^{J(\omega) - J(\eta)} \, p(\omega) = p(\eta) \,Q_\eta(A_n) .$$ If $V = \{0\}$ is a singleton set, then $$\lim_{n \rightarrow \infty} Q_V(A_n) = Q_V({{\mathbb Z}}^2) = \sum_{k=0}^\infty s^k = (1- s)^{-1} > 0.$$ Here $s = {{\mathbb E}}[J']$ where $J' = (-1)^{J(S[0,T_0])}$, $S$ is a simple random walk starting at the the origin, and $T_0 = \min\{n \geq 1: S_n > 0\}$. Since ${{\mathbb P}}\{J' = 1\} > 0$ and ${{\mathbb P}}\{J' = - 1\} > 0$, we see that $\|s\| < 1$ and hence the limit exists and is positive. A similar argument shows that if ${{\mathbb Z}}^2 \setminus U$ is finite and non-empty, and $v$ is in the unbounded component of $U$, then $g_U(v,v)$ is finite and strictly positive. Given this and , it is straightforward to show that if $V$ is finite, then $$\label{nov25.1} Q_V = Q_V({{\mathbb Z}}^2) = \lim_{n \rightarrow \infty} Q_V(A_n)$$ exists and is strictly positive. We will write $Q_{01}(A_n)$ for $Q_{\{0,1\}}(A_n)$ For the important computation of the random walk loop measure, we will use the Brownian loop measure as introduced in [@Lsoup]. There are several equivalent definitions. We give one here that is convenient for computational purposes and is analogous to the second definition we gave for the random walk loop measure. We start with the Brownian (boundary) bubble measure in the upper half plane $\Half$ started at the origin. It is the limit as $\epsilon \downarrow 0$ of a measure on paths from $i \epsilon$ to $0$ in $\Half$ which we now describe. For each $\epsilon$ consider the measureof total mass $\epsilon^{-1}$ on paths whose normalized probability measure is that of a Brownian $h$-process to $0$. (An $h$-process can be viewed roughly as a Brownian motion conditioned to leave $\Half$ at $0$.) As $\epsilon \downarrow 0$, the limit measure is a $\sigma$-finite measure $\nu_\Half(0)$ on loops from $0$ to $0$ otherwise in $\Half$. The normalization is such that the measure of bubbles that hit the unit circle equals one. This definition can be extended to simply connected domains with smooth boundaries either by the analogous definition or by the following conformal covariance rule: if $f: \Half \rightarrow D$ is a conformal transformation, then $$f \circ \nu_\Half(0) = \|f'(0)\|^2 \, \nu_D(f(0)) .$$ (In the definition of $f \circ \nu_\Half(0)$, we need to modify the parametrization of the curve using Brownian scaling, but the parametrization is not important in this paper.) Given the Brownian bubble measure, the Brownian loop measure restricted to curves in the unit disk $\Disk$ can be written as $$\label{loopdefinition} \frac 1 \pi \int_{0}^1 \int_{0}^{2 \pi} \nu_{r\Disk}(re^{i\theta}) \,r\, \, d\theta \, dr,$$ To be more precise, the loop measure is the measure on [unrooted]{} loops induced by the above measure on rooted loops. (This representation of the measure on unrooted loops focuses on the rooted representative with root as far from the origin as possible.) The Brownian loop measure is the scaling limit of the random walk loop measure in a sense made precise in [@Jose]. We discuss this more in Section \[rwloopsec\]. A combinatorial identity {#identsec} ======================== Let $\paths_n'$ denote the set of nearest neighbor paths $\omega = [\omega_0,\omega_1,\ldots,\omega_k]$ with $ \Re[\omega_0] = - n, \omega_k = 0$ and $\{\omega_1,\ldots, \omega_{k-1}\} \subset A_n \setminus [0,\infty)$. Let $\paths_n''$ denote the set of nearest neighbor paths $\omega = [\omega_0,\omega_1,\ldots,\omega_k]$ with $ \Re[\omega_0] = n+1, \omega_k = 1$ and $\{\omega_1,\ldots, \omega_{k-1}\} \subset A_n \setminus (-\infty,1]$. There is a natural bijection between $\paths_n'$ and $\paths_n''$ obtained by reflection about the line $\{\Re(z) = 1/2\}$. Let $$R_n = \sum_{\omega \in \paths_n'} p(\omega) = \sum_{\omega \in \paths_n''} p(\omega).$$ Note that $R_n$ equals ${{\mathbb P}}\{\Re(S_\tau) = -n\}$ where $S$ is a simple random walk starting at the origin and $\tau = \min\{j > 0: S_j \in \p A_n \cup [0,\infty)\}.$ It is known (ee e.g., [@LLimic Proposition 5.3.2]) that $$\label{nov25.2} R_n \asymp n^{-1/2} , \,\,\,\, n \rightarrow \infty .$$ The goal of this section is to prove the following combinatorial identity which relates the probability that loop-erased walk uses the undirected edge $\{0,1\}$ to some simple random walk quantities. \[combin\] $$4\sum_{\eta \in \saws_n^} \hat p_n(\eta) = Q_{01}(A_n) \, R_n^2 \, \exp\left\{ 2 m(\looper_{A_n})\right\}.$$ We claim the following: $$\label{nov19.1} (-1)^{J(\eta) } \, Y(\eta) = 1 \;\;\mbox{ if }\;\; \eta \in \saws_n^.$$ To see this consider the path $\eta$ as a continuous path from $\{\Re(z) = -n\}$ to $\{\Re(z) = n+1\}$ in the domain $D=\{x+iy \in {{\mathbb C}}: -n < x < n+1, -n < y < n\}$. Then $\eta$ is a crosscut of $D$ such that $D \setminus \eta$ consists of two components, the “top” component $D^+$ and the “bottom” component $D^-$. Each ordered edge $[w, w']$ in $\eta$ can be considered as subsets of $\p D^+$ and $\p D^-$. As we traverse from $w$ to $w'$, the left-hand side of $[w, w']$ (considered as a prime end) is in $\p D^+$ and the right-hand side is in $\p D^-$. Let $N_+$ be the set of integers $k$ such that the ordered edge $[ki,ki+1]$ is contained in $\eta$, $N_-$ be the set of integers $k$ such that the ordered edge $[ki+1,ki]$ is contained in $\eta$, and $N = N_+ \cup N_-$. We claim that if $j \in N_+$ and $k$ is the largest integer less than $j$ with $k \in N$, then $k \in N_-$. For otherwise, the open line segment from $ji + (1/2)$ to $ki + (1/2)$ would be in both $D^+$ and $D^-$. We now consider the smallest $k$ such that $k \in N$. The line segment from $-ni + (1/2)$ to $ ki + (1/2)$ is contained in $D^-$ and hence $k \in N_+$. As we continue up the line $\{\Re(z) = 1/2\}$ we see that when we intersect edges in $\eta$, they alternate being in $N_+$ or $N_-$, the first in $N_+$, the second in $N_-$, the third in $N_+$, etc. When we reach the unordered edge $\{0, 1\}$, we see that if $0 \in N_+$, then there have been an even number of edges before $\{0,1\}$ and if $0 \in N_-$, there have been an odd number of edges. In other words, $(-1)^{J(\eta)} = 1$ if $\eta \in \saws_n^+$ and $(-1)^{J(\eta)} = -1$ if $\eta \in \saws_n^{-}$. This gives . Let $\Lambda_n$ be defined as in . We claim that $$\label{nov19.2} \Lambda_n = \sum_{\omega \in \paths_n} q(\omega) \, Y(L(\omega)) = \sum_{\eta \in \saws_n^} \sum_{L(\omega) = \eta} p(\omega) \, (-1)^{J (\omega)- J(\eta)} .$$ To see this, suppose that $L(\omega) = \eta = [\eta_0,\ldots,\eta_k]$. We can write $$\omega = [\eta_0,\eta_1] \oplus l_1 \oplus [\eta_1,\eta_2] \oplus l_2 \oplus \dots\oplus [\eta_{k-2},\eta_{k-1}] \oplus l_{k-1} \oplus [\eta_{k-1},\eta_k],$$ where $l_j$ is a loop rooted at $\eta_j$ that does not enter $\{\eta_1,\ldots,\eta_{j-1}\}$. We write $$J(\omega) = J(\eta) + J_L(\omega) , \;\;\;\; Y(\omega) = Y(\eta) + Y_L(\omega) ,$$ where $J_L,Y_L$ denote the contributions from the loops. Then $$Y(\omega) = Y(\eta) + \sum_{j=1}^{k-1} Y(l_j).$$ For each loop $l_j$ there is the corresponding reversed loop $l_j^R$ for which $Y(l_j^R) = - Y(l_j)$. Since $J(l_j^R) = J(l_j)$ and $Y(l_j^R) = - Y(l_j)$, we get cancellation. Doing this for all the loops, we see that $$\sum_{\omega \in \paths_n, L(\omega) = \eta} q(\omega) \, [Y(\omega) - Y(\eta)] = 0 .$$ This gives the first equality in . The second equality uses and the fact that $Y(\eta) = 0$ if $\eta \not \in \saws_n^$. If $\eta \in \saws_n^$, then $$\begin{aligned} \sum_{L(\omega) = \eta} {p(\omega)} \, (-1)^{J(\omega)- J(\eta)} & = & p(\eta) \sum_{L(\omega) = \eta} \frac{p(\omega)}{p(\eta)} \, (-1)^{J_L(\omega)} \\ & = & p(\eta)\, Q_\eta(A_n) \\& = & p(\eta)\,\exp \left\{\sum_{\bar l \subset A_n, \bar l \cap \eta \neq \emptyset} (-1)^{J(\bar l)} \, m(\bar l) \right\}\\ & = &p(\eta)\, F_\eta(A_n) \,\exp \left\{-2\sum_{\bar l \subset A_n, \bar l \cap \eta \neq \emptyset, J(\bar l) \mbox{ odd}} m(\bar l) \right\}. \end{aligned}$$ If $J(\bar l)$ is odd, then $\bar l$ must include at least one unordered edge $\{ki, ki+1\}$ with $k \geq 0$ and at least one unordered edge $\{ki,ki+1\}$ with $k < 0$. Therefore, topological considerations imply that if $\eta \in \saws_n^$, then $\eta \cap \bar l \neq \emptyset.$ Hence $$\sum_{\bar l \subset A_n, \bar l \cap \eta \neq \emptyset, J(\bar l) \mbox{ odd}} m(\bar l) = \sum_{\bar l \subset A_n, J(\bar l) \mbox{ odd}} m(\bar l) = { m(\looper_{A_n})}.$$ Combining this with , we see that $$\label{nov19.4} \Lambda_n = \sum_{\eta \in \saws_n^} p(\eta) \, Q_\eta(A_n) = e^{-2m(\looper_{A_n}) } \sum_{\eta \in \saws_n^} p(\eta) \, F_\eta(A_n) = e^{-2m(\looper_{A_n}) } \sum_{\eta \in \saws_n^} \hat p_n(\eta) .$$ We will now compute $\Lambda_n$ as defined in a different way. Let $\omega = [\omega_0,\ldots,\omega_\tau] \in \paths_n$. If $\omega$ does not visit $0$ or $\omega$ does not visit $1$, then $Y(\omega) = 0$. Hence, we need only consider the sum over $\omega \in \paths_n$ that visit both $0$ and $1$. Let $T_0 = \min\{j: \omega_j = 0\}, T_0' = \max\{j < \tau: \omega_j = 0\}$ and define $T_1,T_1'$ similarly. Suppose that $T_0 < T_1, T_0' > T_1'$. In this case we write $$\label{jan9.6} \omega = \omega^- \oplus l \oplus \omega^+,$$ where $l$ is the loop $[\omega_{T_0},\ldots,\omega_{T_0'}]$. Note that $Y(\omega) = Y(l)$. For any such loop, there is the corresponding reversed loop $l^R = [\omega_{T_0'}, \omega_{T_0' - 1},\ldots, \omega_{T_0}]$ for which $Y(l^R) = - Y(l)$. These terms cancel and hence the sum in over $\omega$ with $T_0 < T_1, T_0' > T_1'$ is zero. Similarly, the sum over $\omega$ with $T_1 < T_0 \leq T_0' < T_1'$ is zero. Suppose that $T_0 > T_1, T_0' > T_1'$. Then we can write $\omega$ as $$\omega = \omega^- \oplus l_1 \oplus \omega' \oplus l_0 \oplus \omega^+.$$ Here $l_0$ is a loop rooted at $0$, $l_1$ is a loop rooted at $1$, $\omega'$ is a path from $1$ to $0$, $\omega^-$ is a path from $\{\Re(z) = -n\}$ to $1$ avoiding $0$, $\omega^+$ is a path from $0$ to $\{\Re(z) = n+1\}$ avoiding $1$. Let $\tilde \omega^-$ be the reflection of $\omega^-$ about the real axis. Then $J(\omega^-) + J(\tilde \omega^-) $ is odd and these terms will cancel. Hence the sum over all $\omega$ with $T_0 > T_1, T_0' > T_1'$ is zero. Let $ \paths_n^1$ be the set of paths in $\paths_n$ that visit both $0$ and $1$ and satisfy $T_0 < T_1, T_0' < T_1'$. We have shown that $$\Lambda_n = \sum_{\omega \in \paths_n^1} q(\omega) \, Y(\omega) .$$ If $\omega \in \paths_n^1$, let $\rho = \min\{j > T_0': \omega_j = 1 \}$. Then we can write $\omega$ as $$\label{nov19.5} \omega = \omega^- \oplus l_0 \oplus \omega' \oplus l_1 \oplus \omega^+.$$ Here $l_0$ is a loop rooted at $0$, $l_1$ is a loop rooted at $1$, $\omega' =[\omega_{T_0'}, \ldots,\omega_\rho]$ is a path from $0$ to $1$, $\omega^-$ is a path from $\{\Re(z) = -n\}$ to $0$, $\omega^+$ is a path from $1$ to $\{\Re(z) = n+1\}$. The paths $\omega', \omega^-,\omega^+$ do not enter $\{0,1\}$ except at their endpoints. The loop $l_1$ does not visit $0$. All points other than the endpoints must lie in $A_n$. These are all the restrictions on the paths. Note that $Y(\omega) = Y(l_0) + Y(\omega')$. As in the previous arguments, we can replace $l_0$ with the reversed loop $l_0^R$, to see that $$\sum_{\omega \in \paths_n^1} (-1)^{J(\omega)} \, Y(l_0) \, p(\omega) =0.$$ Also $Y(\omega') \in \{0,1\}$ with $Y(\omega') = 1$ if and only if $T_0'+ 1 = \rho$, that is, if $\omega' = [0,1]$. Therefore, if $\paths_n^2$ denotes the set of paths in $\paths_n^1$ with $\omega' = [0,1]$, then $$\label{nov20.4} \Lambda_n = \sum_{\omega \in \paths_n^2} (-1)^{J(\omega )}\, p(\omega) = \sum_{\omega \in \paths_n^2} (-1)^{J(\omega^-) +J(l_0) + J(l_1) + J(\omega^+)}\, p(\omega).$$ If $\omega \in \paths_n^2$, let $\xi$ be the smallest $j$ such that $\omega_j$ is on the positive real axis. Suppose for the moment that $\xi < T_0$. Then we can write $$\omega^- = \omega^{-,1} \oplus \omega^{-,2},$$ by splitting the path at time $\xi$. The path $\omega^{-,2}$ is a path from the positive real axis to $0$ that does not go through the point $1$. Hence, $J(\omega^{-,2}) + J(\tilde \omega^{-,2})$ is odd, where $\tilde \omega^{-,2}$ is the reflection of $\omega^{-,2}$ about the real axis. These terms will cancel in the sum , and hence it suffices to sum over $\omega^-$ such that $\omega^- \cap [1,\infty) = \emptyset$. For these $\omega^-$, we can see by topological reasons that $(-1)^{J(\omega^-)} = 1$. By a similar argument, it suffices to sum over $\omega^+$ satisfying $\omega^+ \cap (-\infty,0] = \emptyset$, and for these walks $(-1)^{J(\omega^+)} = 1$. Therefore, if $\paths_n^3$ denote the set of paths in $\paths_n^2$ satisfying $$\omega^- \cap [1,\infty) = \emptyset, \;\;\; \omega^+ \cap (-\infty,0] = \emptyset,$$ we see that $$\Lambda_n =\sum_{\omega \in \paths_n^3} (-1)^{ J(l_0) + J(l_1) }\, p(\omega).$$ Let us write any $\omega \in \paths_n^3$ as in . We must choose $\omega^- \in \paths_n', (\omega^+)^R \in \paths_n''$ and $\omega' = [0,1]$. Summing over all of these possibilities, gives a factor of $R_n^2/4$. The choices of $l_0,l_1$ are independent of the choices of $\omega^-$ and $\omega^+$. The only restriction is that the loops lie in $A_n$ and $l_1$ does not contain the origin. By our definition, $$\sum_{l_0,l_1} (-1)^{J(l_0) + J(l_1)} \, p(l_0) \, p(l_1) = g_{A_n}(0,0) \, g_{A_n\setminus \{0\}}(1,1) = Q_{01} (A_n) .$$ Therefore, $$\Lambda_n =\sum_{\omega \in \paths_n^3} (-1)^{ J(l_0) + J(l_1) }\, p(\omega) = \frac 14 \, R_n^2 \, Q_{01}(A_n) .$$ Comparing this with gives the theorem. Estimate on the random walk loop measure {#rwloopsec} ======================================== Using Theorem \[combin\] and the estimates and , we see that $$\sum_{\eta \in \saws_n^} \hat p_n(\eta) \asymp n^{-1} \, \exp \left\{2 m(\looper_{A_n}) \right\}.$$ The proof of is finished with the following proposition. There exists $c < \infty$ such that for all $n$, $$\left\| m(\looper_{A_n}) - \frac {1} 8 \, \log n\right\| \leq c .$$ Let $C_n = \{z \in {{\mathbb Z}}^2: \|z\| < e^n\}. $ We will prove the stronger fact that the limit $$\label{nov20.2} \lim_{n \rightarrow \infty} \left[ m(\looper_{C_n}) - \frac n 8 \right]$$ exists by showing that $$\label{nov21.7} \sum_{n=1}^\infty \left\| m(\looper_{C_{n+1}}\setminus \looper_{C_{n}}) - \frac 1 8 \right\| < \infty .$$ Let $\mu$ denote the Brownian loop measure, and let $\tilde \looper$ denote the set of unrooted Brownian loops $\gamma$ in the unit disk that intersect $\{\|z\| \geq e^{-1}\}$ and such that the winding number of $\gamma$ about the origin is odd. We will establish by showing that $\mu(\tilde \looper) = 1/8$ and $$\label{nov20.3} \left\| m(\looper_{C_{n+1}}\setminus \looper_{C_{n}}) -\mu(\tilde \looper) \right\| = O(n^{-2}).$$ For the Brownian loop measure, we do a computation similar to that in [@Annulus Proposition 3.9]. Using , we write $$\mu (\tilde \looper)= \frac 1 \pi \int_{e^{-1}}^1 \int_{0}^{2 \pi} \phi(r,\theta) \, d\theta \,r\, dr,$$ where $\phi(r,\theta)$ denotes the Brownian bubble measure of loops in $r \Disk$ rooted at $r e^{i\theta}$ with odd winding number about the origin. Rotational symmetry implies that $\phi(r,\theta) = \phi(r,0)$ and conformal covariance implies that $\phi(r,0) = r^{-2} \,\phi$ where $\phi = \phi(1,0)$. Hence, $$\label{jan7.1} \mu (\tilde \looper)= \frac{\phi}\pi \int_{e^{-1}}^1 \int_{0}^{2 \pi} r^{-2} \, d\theta \,r\, dr = 2 \phi.$$ By considering the (multi-valued) covering map $f(z) = i\log z $ which satisfies $\|f'(1)\| = 1$, we see that $$\phi = \sum_{k \mbox{ odd}} H_{\p \Half}(0, 2\pi k) ,$$ where $H_{\p \Half}$ denotes the boundary Poisson kernel in the upper half-plane $\Half$ normalized so that $H_{\p \Half}(0,x) =x^{-2}$. Therefore, $$2\phi = 2 \sum_{ k=-\infty} ^ \infty \frac{1}{[2\pi (2k+1)]^2} = \frac{1}{\pi^2} \, \left[1 + \frac 1{3^2} + \frac 1{5^2} + \cdots\right] = \frac 1 {8} .$$ If $s > 2$, let $\tilde \looper _s^$ denote the set of Brownian loops in $\Disk$ that intersect both $\{\|z\| \geq e^{-1} \}$ and $\{\|z\| \leq e^{-s}\}$. We claim that as $s \rightarrow \infty$, $$\label{nov21.6.alt} \mu( \tilde \looper_s^) = s^{-1} + O(s^{-2}),$$ $$\label{nov21.6} \mu [\tilde \looper \cap \tilde \looper_s^] = (2s)^{-1}+ O(s^{-2} ) .$$ To see this, we first consider the boundary bubble measure $\lambda_s$ of loops in $\Disk$ rooted at $1$ that enter $\{\|z\| \leq e^{-s}\}$. An exact expression is given as follows. Let $B_t$ be a Brownian motion and $\sigma_s = \inf\{t: \|B_t\|= e^{-s}\}$. Then, $$\lambda_s = \lim_{\epsilon \downarrow 0} {{\mathbb E}}^{\epsilon} \left[H_\Disk(B_{\sigma_s},1); \sigma_s < \sigma_0 \right] .$$ The Poisson kernel in the disk is well known; for our purpose we need only know that $$H_\Disk(z,1) = \frac 12 + O(\|z\|),$$ and a standard estimate for Brownian motion gives $${{\mathbb P}}^{1-\epsilon} \{\sigma_s < \sigma_0\} = \frac{\log (1-\epsilon )} {-s} \sim \frac{\epsilon}{s}.$$ Therefore, $\lambda_s = (2s)^{-1} \, + O(e^{-s})$. Using rotational invariance, and conformal convariance, if $r \geq e^{-1}$ and $\lambda(r,\theta,s)$ denotes the bubble measure of bubbles in $r\, \Disk$ rooted ant $re^{i\theta} $ that enter $\{\|z\| \leq e^{-s}\}$, then $$\lambda(r,\theta,s) = r^{-2} \, (2s)^{-1} \, [1 + O(s^{-1})].$$ If we compute as in , we get . The relation is done similarly except that we have to worry about the winding number of the loop. This gives a factor of $1/2$. Note that we have $$\label{jan7.3} \mu [\tilde \looper \cap \tilde \looper_s^] = \frac 12 \, \mu [ \tilde \looper_s^] \, [1+O(s^{-1})].$$ For each unrooted random walk loop $\bar l \in \looper_{C_{n}} \setminus \looper_{C_{n-1}}$, there is a corresponding continuous unrooted loop $\bar l^{(n)}$ in $\Disk$ obtained from linear interpolation and Brownian scaling. We will write $d(\bar l,\gamma) \leq \delta$, if we can parametrize and root the loops $ \bar l ^{(n)}$ and $\gamma$ such that the loops are within $\delta$ in the supremum norm. In [@Jose] it was shown that there exists $\alpha > 0$ and a coupling of the random walk and Brownian loop measures in $D$, restricted to loops of diameter at least $1/e$, so that the total masses agree up to $O(e^{-n \alpha})$ and such that in the coupling, except for a set of paths of size $O(e^{-n \alpha})$, we have $d(\bar l,\gamma) < e^{-n \alpha}$. We would like to say that in the coupling, the Brownian loop has odd winding number if and only $J(\bar l)$ is odd. If the loops stay away from the origin, this holds. However, if the loops are near the origin, it is possible for the winding numbers of the continuous and the discrete walks to be different. However, and this is why we can prove what we need, it is also true that if a macroscopic loop (either continuous or discrete) gets close to the origin, then it is just about equally likely to have an odd as an even winding number. Let us be more precise. Let $\beta < \alpha$ and let $\looper^n$ denote the set of loops in $\looper_{C_{n+1}} \setminus \looper _{C_n}$ that intersect $\{\|z\| \leq e^{-\beta n} \, e^{n+1}\}$ Using the coupling and $\eqref{nov21.6},$ we see that $$m (\looper^n) = \mu(\tilde \looper_{\beta n}') + O(n^{-2}) = (\beta n)^{-1} + O(n^{-2}).$$ Let us split these paths into two sets: those for which ${{\rm dist}}(0,\gamma) \leq 2e^{-n\alpha}$ and those for which ${{\rm dist}}(0,\gamma) > 2e^{-n \alpha}$. If ${{\rm dist}}(0,\gamma) > 2e^{-n\alpha}$ and $d(\bar l,\gamma) \leq e^{-n \alpha}$, then $J(\bar l) $ is odd if and only if the winding number of $\gamma$ is odd. Therefore $$m((\looper_{C_{n+1}} \setminus \looper _{C_n} ) \setminus \looper^n) = \mu(\tilde \looper \setminus \tilde \looper'_{\beta n} ) + O(n^{-2}) .$$ (The error term $O(n^{-2})$ is comparable to the measure of loops $\gamma$ such that $e^{-n \beta} \leq {{\rm dist}}(0,\gamma) \leq 2e^{-n\beta}$.) A coupling argument can be used to give a random walk analogue of , $$m \left[\looper^n \cap (\looper_{C_{n+1}}\setminus \looper_{C_{n}})\right] = \frac 12 \, m (\looper^n) \, [1 + o(n^{-1})].$$ We sketch the proof. We use the definition of the loop measure using an enumeration of ${{\mathbb Z}}^2 = \{z_1,z_2,\ldots\}$ such that $\|z_j\|$ increases. Then an unrooted loop in $\looper^n$ is obtained from a loop rooted in $C_{n+1} \setminus C_n$. Let us call the root $z_k$ and so the loops lies in $V_k = \{z_1,\ldots,z_k\}$. Let us stop the walk at the first time it reaches a point, say $z'$, in $\{\|z\| \leq e^{-\beta n}\, e^{n+1}\}$. The remainder of the loop acts like a random walk started at $z'$ conditioned to reach $z_k$ without before leaving $V_k$. Let $J'$ denote the number of times such a walk crosses the half line $\{(1/2) +iy : y < 0\}.$ We claim that the probability that $J'$ is odd equals $\frac 12 + O(e^{-u n})$ for some $u > 0 $. Indeed, we can couple two walks starting at the point so that each walk has the distribution of random walk conditioned to reach $z_k$ before leaving $V_k$ and that, except for an event of probability $O(e^{-\delta})$, the parity of $J'$ is different for the two walks. This uses a standard technique. The key estimate is the following. There exists $c > 0$ such that if $S$ is a simple random walk starting at $z \in C_{j-1}$ and $T = \min\{j: S_y \in C_j\}$, then for all $w \in \p C_j$ with $\Im(w) > 0$, $${{\mathbb P}}\{S(T) = w , \; J' \mbox{ odd} \} \geq c \, e^{-j} ,$$ $${{\mathbb P}}\{S(T) = w , \; J' \mbox{ even} \} \geq c \, e^{-j} .$$ Without the restriction of the parity of $J'$, see, for example, [@LLimic Lemma 6.3.7]. To get the result with the restriction, we just note that there is a positive probability of making a loop in the annulus $C_{j} \setminus C_{j-1}$, and this increases $J'$ by one. Hence, we can find a coupling and a $\rho > 0$ such that at each annulus there is a probability $\rho$ of a successful coupling given that the walks have not yet been coupled. Since there are of order $ {\beta} \, n$ annuli, we can couple the processes so that the probability of not being coupled is $(1-\rho)^{\beta n} = O(e^{-u n})$ for some $u$. From the last two estimates and , we see that $$\left \|\mu(\tilde \looper)- m(\looper_{C_{n+1}}\setminus \looper_{C_{n}})\right\| \leq c \, n^{-2} .$$ This gives . Different boundary conditions ============================= In this section, we generalize Theorem \[combin\] to more general boundary conditions. Let $\bar \zeta =(\zeta_1,\zeta_2)$ be an ordered pair of distinct points in $\p A_n$, and let $\paths_n(\bar \zeta)$ to be the set of nearest neighbor paths $\omega =[\omega_0,\ldots,\omega_k]$ with $\omega_0 = \zeta_1, \omega_k \in \zeta_2$ and $\{\omega_1,\ldots,\omega_{k-1}\} \subset A_n \setminus \{0,1\}$. Let $\saws_n^+(\bar \zeta), \saws_n^-(\bar \zeta), \saws_n^(\bar \zeta)$ be defined similarly with the new boundary condition. As before, if $\eta \in \saws_n^(\bar \zeta)$, then $$\hat p_n(\eta) = \sum_{\omega \in\paths_n(\bar \zeta), L(\omega) = \eta} \hat p(\omega).$$ Define $R_n(\zeta_j)$ as follows. Let $S_n$ be a simple random walk starting at the origin and let $$T = T_n = \min\{j \geq 1: S_j \in [0,\infty) \cup \p A_n \}.$$ Then $$R_n(\zeta_j) = I(\zeta_j) \, {{\mathbb P}}\{S_T = \zeta_j\},$$ where $I(\zeta_j) = -1$ if $\Re(\zeta_j) > 0, \Im(\zeta_j) < 0$ and $I(\zeta_j) = 1$ otherwise. Note that if $\omega \in K_n(\zeta_1)$ with $\omega \cap [1,\infty) = \emptyset$, then $I(\zeta_1) = (-1)^{J(\zeta_1)}.$ Define $$\Phi_n(\bar \zeta) = \left\| R_n(\zeta_1) \, R_n(1 - \zeta_2) - R_n(1-\zeta_1) \, R_n(\zeta_2) \right\|.$$ The combinatorial identity takes the following form. \[jan10theorem\] If $\bar \zeta =(\zeta_1,\zeta_2 ) \in \p A_n \times \p A_n$, then $$\label{jan9.1} \sum_{\eta \in \saws_n^(\bar \zeta)} \hat p(\eta) = \frac 14 \, {Q_{01}(A_n) \, \exp\{2 m(\looper_{A_n})\}} \, \Phi_n(\bar \zeta) .$$ Note that the factor ${Q_{01}(A_n) \, \exp\{2 m(\looper_{A_n})\}}/4$ does not depend on $\bar \zeta$. Hence the theorem implies that if $\bar \zeta, \bar \zeta' \in \p A_n\times \p A_n$, $$\frac{ \sum_{\eta \in \saws_n^(\bar \zeta)} \hat p(\eta)} { \sum_{\eta \in \saws_n^(\bar \zeta')} \hat p(\eta)} = \frac{\Phi_n(\bar \zeta)} {\Phi_n(\bar \zeta')}.$$ Let $J(\omega),Y(\omega)$ be as in the introduction. Both sides of vanish if $\zeta_1 = \zeta_2$ so we assume that $\zeta_1 \neq \zeta_2$. Both sides of are invariant under the transformation $(\zeta_1,\zeta_2) \mapsto (\zeta_2,\zeta_1).$ Let $$\label{Dndef} D_n = \{x+iy \in {{\mathbb C}}: -n < x < n+1, -n < y < n\}.$$ The boundary $\p D_n$ is a rectangle. Without loss of generality we will assume that the order $(\zeta_1,\zeta_2)$ is chosen so that the point $\frac 12 -ni$ is on the arc of $\p D_n$ from $\zeta_1$ to $\zeta_2$ in the counterclockwise direction. If $\eta \in \saws_n^(\bar \zeta) $, considered as a simple curve by linear interpolation, then $D_n \setminus \eta$ consists of two components, $D^+,D^-$ where $D^-$ denotes the component containing $\frac 12 -ni$ on its boundary. By our choice of order of $\zeta_1,\zeta_2$, as we traverse $\eta$ from $\zeta_1$ to $\zeta_2$, the right-hand side of $\eta$ is on $\p D^-$ and the left-hand side is $\p D^+$. Using this, we can use the topological argument as in to conclude that $$\label{jan9.4} (-1)^{J(\eta)} \, Y(\eta) = 1 \;\;\;\; \mbox{if} \;\;\;\; \eta \in \saws_n^(\bar \zeta).$$ Let $$\label{jan9.5} \Lambda = \Lambda_n(\bar \zeta) = \sum_{\omega \in \paths_n(\bar \zeta)} (-1)^{J(\omega)} \, p(\omega) \, Y(\omega) .$$ We claim that $$\Lambda = \exp \left\{-2 m(\looper_{A_n}) \right\} \sum_{\eta \in \saws_n^(\bar \zeta)} \hat p_n (\eta) .$$ Indeed, given , the proof is identical to that of . Hence to prove the theorem, it suffices to prove that $$\label{jan9.3} \Lambda = \frac{Q_{01}(A_n)}{4} \, \Phi_n(\bar \zeta).$$ As in the proof of Theorem \[combin\], we note that if $\omega$ does not contain the unordered edge $\{0,1\}$, then $Y(\omega) = 0$. Hence the sum in is over $\omega$ such that $T_0,T_0',T_1,T_1'$, as defined in the proof of Theorem \[combin\], are well defined. As in that proof, if $T_0 < T_1$ and $T_0' > T_1$, we can write $\omega$ as in . By considering the path with the reversed loop, we see that the sum over $\omega$ with $T_0 < T_1$ and $T_0' > T_1$ equals zero. Similarly, the sum over $\omega$ with $T_0 > T_1, T_0' < T_1'$ is zero. Hence we need only consider the sum over paths with $T_0 < T_1, T_0' < T_1'$ or $T_0 > T_1, T_0' > T_1'$. However, unlike in the proof of Theorem \[combin\], the sum over $\omega$ with $T_0 > T_1, T_0' > T_1'$ does not vanish. Let $\paths^5 = \paths^5_n(\bar \zeta)$ be the set of $\omega \in \paths_n(\bar \zeta)$ such that $T_0 < T_1$ and $S_{T_0'+1} = 1$. Let $\paths^6 = \paths^6_n(\bar \zeta)$ be the set of $\omega \in \paths_n(\bar \zeta)$ such that $T_1 < T_0$ and $S_{T_0-1} = 1$. Then the argument leading to in this case gives $$\Lambda = \sum_{\omega \in \paths^5} (-1)^{J(\omega)} \, p(\omega) - \sum_{\omega \in \paths^6} (-1)^{J(\omega)} \, p(\omega) .$$ If $\omega=[\omega_0,\ldots,\omega_\tau] \in \paths^5$, we write $$\omega = \omega_- \oplus l_0 \oplus [0,1] \oplus l_1 \oplus \omega_+$$ where $$\omega_- = [\omega_0,\ldots,\omega_{T_0}], \;\;\;\; l_0 = [\omega_{T_0},\ldots,\omega_{T_0'}],$$ $$l_1 = [\omega_{T_0'+1},\ldots,\omega_{T_1'}], \;\;\;\; \omega_+ = [\omega_{T_1'},\ldots,\omega_\tau].$$ Note that $\omega_-$ is a path from $\zeta_1$ to $0$; $l_0$ is a loop rooted at $0$; $l_1$ is a loop rooted at $1$ that avoids the origin; and $\omega_=$ is a path from $1$ to $\zeta_2$. The points in $l_0,l_1$ must lie in $A_n$. The points in $\omega_-,\omega_+$ other than the endpoints must lie in $A_n \setminus \{0,1\}$. Let $\paths^6 = \paths^6_n(\zeta_1)$ be the set of nearest neighbor paths $\omega = [\omega_0,\ldots,\omega_k]$ with $\omega_0 = \zeta_1, \omega_k = 0, \{\omega_1,\ldots,\omega_{k-1}\} \subset A_n \setminus \{0,1\}$. Let $\paths^7 = \paths^7_n(\zeta_2)$ be the set of nearest neighbor paths $\omega = [\omega_0,\ldots,\omega_k]$ with $\omega_0 = 1, \omega_k = \zeta_2, \{\omega_1,\ldots,\omega_{k-1}\} \subset A_n \setminus \{0,1\}$. Then $$\sum_{\omega \in \paths^5} (-1)^{J(\omega)} \, p(\omega) = \frac{Q_{0,1}(A_n)}{4} \, \left[\sum_{\omega \in \paths^6 } (-1)^{J(\omega)} \, p(\omega)\right] \, \left[\sum_{\omega \in \paths^7} (-1)^{J(\omega)}\, p(\omega)\right].$$ If $\omega = [\omega_0,\ldots,\omega_k] \in \paths^6$, let $j$ be the first index (if it exists) such that $\omega_j \in [1,\infty)$. If $j$ exists, we can write $$\omega = [\omega_0,\ldots,\omega_j] \oplus [\omega_j,\ldots,\omega_k].$$ The path $\omega'= [\omega_j,\ldots,\omega_k]$ is a path from the positive real line to the origin. If $\tilde \omega'$ denotes the reflection of $\omega$ about the real axis, then $J(\omega') + J(\tilde \omega')$ is odd. Hence, these terms will cancel in the sum. Therefore, $$\sum_{\omega \in \paths^6 } (-1)^{J(\omega)} \, p(\omega) = \sum_{\omega \in \paths^8 } (-1)^{J(\omega)} \, p(\omega),$$ where $\paths^8$ denotes the set of paths in $\paths^6$ that do not intersect $[1,\infty)$. If $\omega \in \paths^8$, then $\omega$ cannot hit the positive real line, so we can see that $J(\omega)$ is even unless $\omega$ lies in the quadrant $\{ \Re(\zeta_1) > 0,\Im(\zeta_1) < 0\}$ in which case $J(\omega)$ is odd. Therefore, $$\sum_{\omega \in \paths^8 } (-1)^{J(\omega)} \, p(\omega) = R_n(\zeta_1).$$ Using symmetry about the line $\{\Re(z) = 1/2\}$, we can see that $$\sum_{\omega \in \paths^7} (-1)^{J(\omega)}\, p(\omega) = R_n(1 - \zeta_2).$$ Similarly, we see that $$\sum_{\omega \in \paths^6} (-1)^{J(\omega)} \, p(\omega) = \frac{Q_{01}(A_n)}{4} \,R_n(\zeta_2) \, R_n(1-\zeta_1).$$ Scaling limit ============= The scaling limit of the loop-erased walk is known in some sense to be the Schramm-Loewner evolution with parameter $2$ ($SLE_2$), see [@LSW]. Here we would like to give a stronger conjecture about this convergence that has not been proved and then show how the estimates here could be used to prove one case of this conjecture. Although we could probably given the details of the proof, we choose not to bother since it is only a start to the main result we hope to obtain later. We start by reviewing $SLE_\kappa$ in the simple curve case. We take a “partition function” view with curves parametrized by natural parametrization as outlined in [@Annulus]. Suppose $0 < \kappa \leq 4$ and let $a = 2/\kappa \in [1/2,\infty)$. We will consider bounded simply connected domains $D \subset {{\mathbb C}}$ containing the origin with piecewise anayltic boundaries. We will consider what is sometimes called two-sided radial $SLE_\kappa$, or, more convenient for use, as chordal $SLE_\kappa$ from $z$ to $w$ conditioned to go through the origin. This is a finite measure on simple curves $\gamma:[0,T_\gamma] \rightarrow \overline D$ that go through the origin with $\gamma(0) = z, \gamma(T_\gamma) = w, \gamma(0,T_z) \subset D$. We denote this measure by $\mu_D(0;z,w)$ and write it as $$\mu_D(0;z,w)= \tmass_D(0;z,w) \, \mu_D^\#(0,z,w)$$ where $\tmass_D(0;z,w)$ denotes the total mass of the measure and $\mu_D^\#(0,z,w) $ is a probability measure. The measure is supported on curves of Hausdorff dimension $d = d_\kappa = 1 + \frac \kappa 8$. The family of measures satisfies a conformal covariance property that we now describe. Suppose $F:D \rightarrow F(D)$ is a conformal transformation with $F(0) = 0$ and such that $\p F(D)$ is locally analytic near $F(z),F(w)$. If $\gamma$ is a curve, define the curve $F \circ \gamma$ to be the image of the curve reparametrized so that the time to traverse $F(\gamma[r,s])$ is $$\int_r^s \|F'(t)\|^d \, dt .$$ If $\nu$ is a measure on curves on $D$, we define $F \circ \nu$ to be the measure on cureves on $F(D)$, $$F\circ \nu(V) = \nu\{\gamma: F \circ \gamma \in V \}.$$ Then the conformal covariance rule is $$F \circ \mu_D(0;z,w) = \|F'(z)\|^b \, \|F'(w)\|^b \, \|F'(0)\|^{2-d} \, \mu_{F(D)} (F(0);F(z), F(w)),$$ where $b = 3a-1 = \frac{6 - \kappa}{2\kappa}$. This rule can be considered as a combination of two rules, $$\label{tmassrule} \tmass_D(0;z,w) = \|F'(z)\|^b \, \|F'(w)\|^b \, \|F'(0)\|^{2-d} \,\tmass_{F(D)}(0;F(z),F(w)),$$ $$F \circ \mu_D^\#(0;z,w) = \mu_{F(D)}^\# (F(0);F(z), F(w)).$$ For each $\kappa$ there is a unique such family of measures up to two arbitrary multiplicative constants. (If $\mu_D$ satisfies the conditions above, then so does $c \, \mu_D$, and also the measure obtained by changing the unit of time, that is, replacing $\gamma(s), 0 \leq s \leq t$ with $\tilde \gamma (s) = \gamma(sr), 0 \leq s \leq t/r$.) For simply connected $D$, the total mass is given (up to multiplicative constant) by $$\tmass_D(0;z,w) = H_{\p D}(z,w)^b \, G_{D}(0;z,w)$$ where $G_D(0;z,w)$ denotes the $SLE_\kappa$ Green’s function giving the normalized probability that a chordal $SLE_\kappa$ path from $z$ to $w$ goes through the origin. Up to multiplicative constant, $$G_\Disk(0;1,e^{2i\theta}) = \|\sin \theta\|^{4a-1} .$$ and for other domains can be determined by $$G_D(0;z,w) = \|F'(0)\|^{2-d} \, G_{F(D)}(F(z),F(w)).$$ Loop-erased walk corresponds to $\kappa = 2\, (a=1,b=1,d = 5/4)$, and for the remainder of this section we fix $\kappa = 2$. If $U,V$ are closed subarcs of $\p D$, we define $$\tmass_D(0;U,V) = \int_U \int_V \tmass_D(0;z,w) \, \|dz\| \, \|dw\|,$$ $$\mu_D(0;U,V) = \int_U \int_V \mu_D(0;z,w) \, \|dz\| \, \|dw\|.$$ Since $b=1$, the scaling rule can be used to see that $$\tmass_D(0;U,V) = \|F'(0)\|^{3/4} \, \tmass_{F(D)}(0;F(U), F(V)).$$ Given this, we can extend the definition to domains with rough boundaries. We will now state a [conjecture]{} about the convergence of loop-erased walk to $SLE_2$. We will state it for simply connected domains, but we expect it to be true for finitely connected domains as well. - Suppose $D$ is a bounded, simply connected domain containing the origin whose boundary $\p D$ is a Jordan curve. - Let $U,V$ be disjoint, closed subarcs of $\p D$. - Let $\sdom = \{x+iy \in {{\mathbb C}}: \|x\|,\|y\| \leq 1/2\}$ and if $z \in {{\mathbb Z}}\times i {{\mathbb Z}}$, let $\sdom_z = z + \sdom$. - Let $A_n = A_n(D)$ be the connected component containing the origin of the set of $z \in {{\mathbb Z}}\times i {{\mathbb Z}}$ such that $\sdom_z \subset nD$. - Let $U_n$ be the set of $z \in \p A_n$ such that $\sdom_z \cap nU \neq \emptyset$. Define $V_n$ simiarly. - Let $\paths_n = \paths_n(D,U,V)$ denote the set of nearest neighbor paths $\omega = [\omega_0,\ldots,\omega_k]$ with $\omega_0 \in U_n, \omega_k \in V_n$ and $\{\omega_1, \ldots,\omega_{k-1}\} \subset A_n$. Let $\saws_n^* = \saws_n^(D,U,V)$ denote the set of self-avoiding paths in $\paths_n$ that contain the unordered edge $\{0,1\}$. As before, let $$\hat p_n(\eta) = \sum_{\omega \in \paths_n, L(\omega) = \eta} p(\omega).$$ - If $\eta = [\eta_0,\ldots,\eta_k] \in \saws_n^ $, define the scaled path $\eta^{n}(t)$ by $$\eta^{(n)}(jn^{-5/4}) = \eta_j/n, \;\;\; j=0,1,\ldots,n$$ and linear interpolation in between. Let $\mu^{(n)} = \mu^{(n)}_D(0;U,V)$ denote the measure of paths that gives measure $ \hat p_n(\eta)$ to $\eta^{(n)}$ for each $\eta \in \saws_n^(D,U,V)$. We can choose the two arbitrary constants in the definition of the measure $\mu_D$ such for every $D,U,V$ as above, $$\lim_{n \rightarrow \infty} n^{3/4} \, \mu^{(n)} = \mu_D(0;U,V) .$$ In particular, $$\label{jan20.5} \lim_{n \rightarrow \infty} \sum_{\eta \in \saws_n} \hat p_n(\eta) = \tmass_D(0;U,V).$$ We will show how Theorem \[jan10theorem\] can be interpreted as one case of the conjecture . Let $\Disk$ denote the unit disk in ${{\mathbb C}}$ and $D$ the square $$D = \{x+iy: \|x\|,\|y\| < 1 \}.$$ If $z \in \p D$, let $z_n$ be the corresponding point in $\p A_n$. (Approximately $z_n = n z$, but we need to round to the nearest integer and compensate for the fact that $A_n$ is not exactly a square centered at the origin. These are very small errors.) Then, we would expect from that $$\label{jan10.2} n\, \Phi_n(z_n,w_n ) \sim c \, \tmass_D(0;z,w) \sim c'\, H_{\p D} ( z,w) \, G_D(0;z,w) .$$ We will show why this holds. We will not calculate the right-hand side explicitly (we could by computing the map $F$ below, but it will not be necessary). However, we do know that $$H_{\p \Disk} ( e^{2i\theta_1}, e^{2i\theta_2}) \, G_D(0;e^{2i\theta_1}, e^{2i\theta_2}) \sim c \, [\sin (\theta_1-\theta_2)]^{-2} \, \|\sin (\theta_1 - \theta_2)\|^3 = c \, \|\sin(\theta_1 - \theta_2)\|.$$ There is a unique conformal transformation $$F: \Disk \rightarrow D,$$ with $F(-1) = -1, F(0) = 0, F(1) = 1$. One way to show it exists is to define $F: \Disk \cap \Half \rightarrow D \cap \Half$ to be the unique conformal transformation that fixes the boundary points $-1,0,1$, and then to extend $F$ to $\Disk$ by Schwarz reflection. Let $\Disk^+ = \Disk \setminus (-1,0], \Disk^- =\Disk \setminus [0,1), D^+ = D \setminus (-1,0], D^- = D \setminus [0,1)$, and note that $F$ conformally maps $\Disk^+$ onto $D^+$ and $\Disk^-$ onto $D^-$. Using conformal invariance, we see that $$G_D(0;F(e^{2i\theta_1}),F(e^{2i\theta_2}) )= \|F'(0)\|^{-3/4} \, G_\Disk(0;e^{2i\theta_1},e^{2i\theta_2}) = c_3 \, \|\sin(\theta_2-\theta_1)\|^3,$$ $$H_{\p D} (F(e^{2i\theta_1}),F(e^{2i\theta_2}))\, G_D(0;F(e^{2i\theta_1}),F(e^{2i\theta_2}) ) = \hspace{1in}$$ $$\label{jan10.3} \hspace{1in} c_4\, \|F'(e^{2i\theta_1})\|^{-1} \, \|F'(e^{2i\theta_2})\|^{-1}\, \|\sin(\theta_2-\theta_1)\| .$$ Let $g_+$ denote the Poisson kernel in the slit domain $\Disk^+$ defined as follows. Start a Brownian motion near the origin in $\Disk^+$ and condition the Brownian motion to exit $\Disk^+$ on the unit circle $\p \Disk$. Then $g_+$ is the conditional density of the exit distribution (taken in the limit as the starting point approaches the origin). It is not difficult to compute $g_+$ using conformal invariance $$g_+(e^{2i\theta}) = c_1 \, \sin \theta .$$ (In this section, we will use $c_1,c_2,\ldots$ for absolute constants whose values could be computed, but we will not bother to.) If $g_-$ is the corresponding density in $\Disk^-$, $$g_-(e^{2i\theta}) = c_1 \, \cos \theta .$$ Let $\hat g_+, \hat g_-$ denote the corresponding densities for the slit domains $D^+, D^-$. Conformal covariance implies that $$\label{jan10.1} \hat g_+(F(e^{i\theta})) = c_2 \, \|F'(e^{2i\theta})\|^{-1} \, \sin \theta, \;\;\;\; \hat g_-(F(e^{i\theta})) = c_2 \, \|F'(e^{2i\theta})\|^{-1} \, \cos \theta.$$ Suppose that $z = F(e^{2i\theta_1}), w = F(e^{2i\theta_2}). $ The relation suggests (and, in fact, it can be proved) that $$\| R_n(z_n) \| \sim c_6 \, n^{-1/2} \, \|F'(e^{2i\theta_1})\|^{-1} \, \sin \theta_1, \;\;\;\; \|R_n(1-z_n)\| \sim c_6 \, n^{-1/2} \, \|F'(e^{2i\theta_1})\|^{-1} \, \cos \theta_1 .$$ (Here we use the fact that $\|F'(x+iy) = \|F'(-x+iy)\|$.) Let us consider two cases. First, suppose that $z,w$ are both in the upper half plane, that is, $0 \leq \theta_2 < \theta_1 \leq \pi/2$. Then, $$\begin{aligned} n\, \Phi_n(z_n,w_n) & = & n \, \left[R_n(z_n) \, R_n(1-w_n) - R_n(1-z_n ) \, R_n(w_n) \right]\\ & \sim & c_6^2 \|F'(e^{2i\theta_1})\|^{-1} \, \|F'(e^{2i\theta_1})\|^{-1} \, \left[\sin \theta_1 \, \cos \theta_2 - \cos \theta_1 \, \sin \theta_2\right]\\ & = & c_6^2 \|F'(e^{2i\theta_1})\|^{-1} \, \|F'(e^{2i\theta_1})\|^{-1} \, \sin(\theta_1-\theta_2) \end{aligned}$$ As a second case, suppose that $z = F(e^{2i\theta_1}), w = F(e^{-2i\theta_2}). $ where we also assume that $\theta_2 \leq \pi/4$ so that $w_n$ is in the southeast quadrant. Then $R_n(w_n) < 0, R_n(1-w_n) > 0$, and $$\begin{aligned} n\, \Phi_n(z_n,w_n) & = & n \, \left[R_n(z_n) \, R_n(1-w_n) - R_n(1-z_n ) \, R_n(w_n) \right]\\ & \sim & c_6^2 \|F'(e^{2i\theta_1})\|^{-1} \, \|F'(e^{2i\theta_1})\|^{-1} \, \left[\sin \theta_1 \, \cos \theta_2 +\cos \theta_1 \, \sin \theta_2\right]\\ & = & c_6^2 \|F'(e^{2i\theta_1})\|^{-1} \, \|F'(e^{2i\theta_1})\|^{-1} \, \sin(\theta_1 +\theta_2) \end{aligned}$$ If we compare this with , we see that we get the prediction . [00]{} R. Kenyon (2000). The asymptotic determinant of the discrete Laplacian, Acta Math. [185]{}, 239–286. R. Kenyon and D. Wilson (2012) Spanning trees of graphs on surfaces and the intensity of loop-erased random walk on ${{\mathbb Z}}^2$, preprint. G. Lawler. Defining SLE in multiply connected domains using the Brownian loop measure, preprint. G. Lawler and V. Limic (2010) [Random Walk: A Modern Introduction]{}, Cambridge U. Press. G. Lawler and M. Rezaei, Minkowski content and natural parametrization for the Schramm-Loewner evolution, preprint. G. Lawler. O. Schramm, and W. Werner (2004). Conformal invariance of planar loop-erased random walks and uniform spanning trees, Annals of Probab. [32]{}, 939–995. G. Lawler and S. Sheffield (2011). A natural parametrization for the Schramm-Loewner evolution. Annals of Probab. [39]{}, 1896–1937. G. Lawler and J. Trujillo Ferreras (2007). Random walk loop soup, Trans. AMS [359]{}, 767–787. G. Lawler and W. Werner (2004). The Brownian loop soup, Probab. Th. Rel. Fields [128]{}, 565–588. R. Masson (2009). The growth exponent for planar loop-erased walk, Elect. J. Probab [14*]{}, 1012–1073. [^1]: Research supported by NSF grant DMS-0907143.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider the two related problems of detecting if an example is misclassified or out-of-distribution. We present a simple baseline that utilizes probabilities from softmax distributions. Correctly classified examples tend to have greater maximum softmax probabilities than erroneously classified and out-of-distribution examples, allowing for their detection. We assess performance by defining several tasks in computer vision, natural language processing, and automatic speech recognition, showing the effectiveness of this baseline across all. We then show the baseline can sometimes be surpassed, demonstrating the room for future research on these underexplored detection tasks.' author: - \| Dan Hendrycks[^1]\ University of California, Berkeley\ `[email protected]` Kevin Gimpel\ Toyota Technological Institute at Chicago\ `[email protected]` bibliography: - 'bibliography.bib' title: \| A Baseline for Detecting Misclassified and Out-of-Distribution Examples\ in Neural Networks --- Introduction ============ When machine learning classifiers are employed in real-world tasks, they tend to fail when the training and test distributions differ. Worse, these classifiers often fail silently by providing high-confidence predictions while being woefully incorrect [@goodfellow; @colah]. Classifiers failing to indicate when they are likely mistaken can limit their adoption or cause serious accidents. For example, a medical diagnosis model may consistently classify with high confidence, even while it should flag difficult examples for human intervention. The resulting unflagged, erroneous diagnoses could blockade future machine learning technologies in medicine. More generally and importantly, estimating when a model is in error is of great concern to AI Safety [@colah]. These high-confidence predictions are frequently produced by softmaxes because softmax probabilities are computed with the fast-growing exponential function. Thus minor additions to the softmax inputs, i.e. the logits, can lead to substantial changes in the output distribution. Since the softmax function is a smooth approximation of an indicator function, it is uncommon to see a uniform distribution outputted for out-of-distribution examples. Indeed, random Gaussian noise fed into an MNIST image classifier gives a “prediction confidence” or predicted class probability of 91%, as we show later. Throughout our experiments we establish that the prediction probability from a softmax distribution has a poor direct correspondence to confidence. This is consistent with a great deal of anecdotal evidence from researchers [@nguyen; @yu; @provost; @foolers]. However, in this work we also show the prediction probability of incorrect and out-of-distribution examples tends to be lower than the prediction probability for correct examples. Therefore, capturing prediction probability statistics about correct or in-sample examples is often sufficient for detecting whether an example is in error or abnormal, even though the prediction probability viewed in isolation can be misleading. These prediction probabilities form our detection baseline, and we demonstrate its efficacy through various computer vision, natural language processing, and automatic speech recognition tasks. While these prediction probabilities create a consistently useful baseline, at times they are less effective, revealing room for improvement. To give ideas for future detection research, we contribute one method which outperforms the baseline on some (but not all) tasks. This new method evaluates the quality of a neural network’s input reconstruction to determine if an example is abnormal. In addition to the baseline methods, another contribution of this work is the designation of standard tasks and evaluation metrics for assessing the automatic detection of errors and out-of-distribution examples. We use a large number of well-studied tasks across three research areas, using standard neural network architectures that perform well on them. For out-of-distribution detection, we provide ways to supply the out-of-distribution examples at test time like using images from different datasets and realistically distorting inputs. We hope that other researchers will pursue these tasks in future work and surpass the performance of our baselines. In summary, while softmax classifier probabilities are not directly useful as confidence estimates, estimating model confidence is not as bleak as previously believed. Simple statistics derived from softmax distributions provide a surprisingly effective way to determine whether an example is misclassified or from a different distribution from the training data, as demonstrated by our experimental results spanning computer vision, natural language processing, and speech recognition tasks. This creates a strong baseline for detecting errors and out-of-distribution examples which we hope future research surpasses. Problem Formulation and Evaluation ================================== In this paper, we are interested in two related problems. The first is error and success prediction: can we predict whether a trained classifier will make an error on a particular held-out test example; can we predict if it will correctly classify said example? The second is in- and out-of-distribution detection: can we predict whether a test example is from a different distribution from the training data; can we predict if it is from within the same distribution?[^2] Below we present a simple baseline for solving these two problems. To evaluate our solution, we use two evaluation metrics. Before mentioning the two evaluation metrics, we first note that comparing detectors is not as straightforward as using accuracy. For detection we have two classes, and the detector outputs a score for both the positive and negative class. If the negative class is far more likely than the positive class, a model may always guess the negative class and obtain high accuracy, which can be misleading [@provost]. We must then specify a score threshold so that some positive examples are classified correctly, but this depends upon the trade-off between false negatives (fn) and false positives (fp). Faced with this issue, we employ the Area Under the Receiver Operating Characteristic curve (AUROC) metric, which is a threshold-independent performance evaluation [@auroc]. The ROC curve is a graph showing the true positive rate ($\text{tpr}=\text{tp}/(\text{tp} + \text{fn})$) and the false positive rate ($\text{fpr}=\text{fp}/(\text{fp} + \text{tn})$) against each other. Moreover, the AUROC can be interpreted as the probability that a positive example has a greater detector score/value than a negative example [@ROC]. Consequently, a random positive example detector corresponds to a 50% AUROC, and a “perfect” classifier corresponds to 100%.[^3] The AUROC sidesteps the issue of threshold selection, as does the Area Under the Precision-Recall curve (AUPR) which is sometimes deemed more informative [@manning]. This is because the AUROC is not ideal when the positive class and negative class have greatly differing base rates, and the AUPR adjusts for these different positive and negative base rates. For this reason, the AUPR is our second evaluation metric. The PR curve plots the precision ($\text{tp}/(\text{tp} + \text{fp})$) and recall ($\text{tp}/(\text{tp} + \text{fn})$) against each other. The baseline detector has an AUPR approximately equal to the precision [@auprbaseline], and a “perfect” classifier has an AUPR of $100\%$. Consequently, the base rate of the positive class greatly influences the AUPR, so for detection we must specify which class is positive. In view of this, we show the AUPRs when we treat success/normal classes as positive, and then we show the areas when we treat the error/abnormal classes as positive. We can treat the error/abnormal classes as positive by multiplying the scores by $-1$ and labeling them positive. Note that treating error/abnormal classes as positive classes does not change the AUROC since if $S$ is a score for a successfully classified value, and $E$ is the score for an erroneously classified value, $\text{AUROC}=P(S>E) = P(-E>-S)$. We begin our experiments in Section \[sec:softmax-stats\] where we describe a simple baseline which uses the maximum probability from the softmax label distribution in neural network classifiers. Then in Section \[sec:auxiliary\] we describe a method that uses an additional, auxiliary model component trained to reconstruct the input. Softmax Prediction Probability as a Baseline {#sec:softmax-stats} ============================================ In what follows we retrieve the maximum/predicted class probability from a softmax distribution and thereby detect whether an example is erroneously classified or out-of-distribution. Specifically, we separate correctly and incorrectly classified test set examples and, for each example, compute the softmax probability of the predicted class, i.e., the maximum softmax probability.[^4] From these two groups we obtain the area under PR and ROC curves. These areas summarize the performance of a binary classifier discriminating with values/scores (in this case, maximum probabilities from the softmaxes) across different thresholds. This description treats correctly classified examples as the positive class, denoted “Success” or “Succ” in our tables. In “Error” or “Err” we treat the the incorrectly classified examples as the positive class; to do this we label incorrectly classified examples as positive and take the negatives of the softmax probabilities of the predicted classes as the scores. For “In,” we treat the in-distribution, correctly classified test set examples as positive and use the softmax probability for the predicted class as a score, while for “Out” we treat the out-of-distribution examples as positive and use the negative of the aforementioned probability. Since the AUPRs for Success, Error, In, Out classifiers depend on the rate of positive examples, we list what area a random detector would achieve with “Base” values. Also in the upcoming results we list the mean predicted class probability of wrongly classified examples (Pred Prob Wrong (mean)) to demonstrate that the softmax prediction probability is a misleading confidence proxy when viewed in isolation. The “Pred. Prob (mean)” columns show this same shortcoming but for out-of-distribution examples. Table labels aside, we begin experimentation with datasets from vision then consider tasks in natural language processing and automatic speech recognition. In all of the following experiments, the AUROCs differ from the random baselines with high statistical significance according to the Wilcoxon rank-sum test. Computer Vision --------------- In the following computer vision tasks, we use three datasets: MNIST, CIFAR-10, and CIFAR-100 [@cifar]. MNIST is a dataset of handwritten digits, consisting of 60000 training and 10000 testing examples. Meanwhile, CIFAR-10 has colored images belonging to 10 different classes, with 50000 training and 10000 testing examples. CIFAR-100 is more difficult, as it has 100 different classes with 50000 training and 10000 testing examples. In Table \[tab:visionerrors\], we see that correctly classified and incorrectly classified examples are sufficiently distinct and thus allow reliable discrimination. Note that the area under the curves degrade with image recognizer test error. [X \| \[5]{}[>[=.8]{}Y]{}]{} Dataset & AUROC/Base & AUPR Succ/Base & AUPR Err/Base & Pred. Prob Wrong(mean) & Test Set Error\ [MNIST]{} & 97 & 100 & 48 & 86 & 1.69\ [CIFAR-10]{} & 93 & 100 & 43 & 80 & 4.96\ [CIFAR-100]{} & 87 & 96 & 62 & 66 & 20.7\ ***** Next, let us consider using softmax distributions to determine whether an example is in- or out-of-distribution. We use all test set examples as the in-distribution (positive) examples. For out-of-distribution (negative) examples, we use realistic images and noise. For CIFAR-10 and CIFAR-100, we use realistic images from the Scene UNderstanding dataset (SUN), which consists of 397 different scenes [@SUN]. For MNIST, we use grayscale realistic images from three sources. Omniglot [@omniglot] images are handwritten characters rather than the handwritten digits in MNIST. Next, notMNIST [@notmnist] consists of typeface characters. Last of the realistic images, CIFAR-10bw are black and white rescaled CIFAR-10 images. The synthetic “Gaussian” data is random normal noise, and “Uniform” data is random uniform noise. Images are resized when necessary. The results are shown in Table \[tab:visionood\]. Notice that the mean predicted/maximum class probabilities (Pred. Prob (mean)) are above 75%, but if the prediction probability alone is translated to confidence, the softmax distribution should be more uniform for CIFAR-100. This again shows softmax probabilities should not be viewed as a direct representation of confidence. Fortunately, out-of-distribution examples sufficiently differ in the prediction probabilities from in-distribution examples, allowing for successful detection and generally high area under PR and ROC curves. [X \| \[4]{}[>[=.45]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base & AUPR In/Base & AUPR Out/Base & Pred. Prob (mean)\ [CIFAR-10/SUN]{} & 95 & 89 & 97 & 72\ [CIFAR-10/Gaussian]{} & 97 & 98 & 95 & 77\ [CIFAR-10/All]{} & 96 & 88 & 98 & 74\ [CIFAR-100/SUN]{} & 91 & 83 & 96 & 56\ [CIFAR-100/Gaussian]{} & 88 & 92 & 80 & 77\ [CIFAR-100/All]{} & 90 & 81 & 96 & 63\ [MNIST/Omniglot]{} & 96 & 97 & 96 & 86\ [MNIST/notMNIST]{} & 85 & 86 & 88 & 92\ [MNIST/CIFAR-10bw]{} & 95 & 95 & 95 & 87\ [MNIST/Gaussian]{} & 90 & 90 & 91 & 91\ [MNIST/Uniform]{} & 99 & 99 & 98 & 83\ [MNIST/All]{} & 91 & 76 & 98 & 89\ *********************** For reproducibility, let us specify the model architectures. The MNIST classifier is a three-layer, 256 neuron-wide, fully-connected network trained for 30 epochs with Adam [@adam]. It uses a GELU nonlinearity [@gelu], $x\Phi(x)$, where $\Phi(x)$ is the CDF of the standard normal distribution. We initialize our weights according to [@init], as it is suited for arbitrary nonlinearities. For CIFAR-10 and CIFAR-100, we train a 40-4 wide residual network [@wrn] for 50 epochs with stochastic gradient descent using restarts [@sgdr], the GELU nonlinearity, and standard mirroring and cropping data augmentation. Natural Language Processing --------------------------- Let us turn to a variety of tasks and architectures used in natural language processing. ### Sentiment Classification The first NLP task is binary sentiment classification using the IMDB dataset [@maas], a dataset of polarized movie reviews with 25000 training and 25000 test reviews. This task allows us to determine if classifiers trained on a relatively small dataset still produce informative softmax distributions. For this task we use a linear classifier taking as input the average of trainable, randomly initialized word vectors with dimension 50 [@fasttext; @DAN]. We train for 15 epochs with Adam and early stopping based upon 5000 held-out training reviews. Again, Table \[tab:imdberr\] shows that the softmax distributions differ between correctly and incorrectly classified examples, so prediction probabilities allow us to detect reliably which examples are right and wrong. Now we use the Customer Review [@CR] and Movie Review [@MR] datasets as out-of-distribution examples. The Customer Review dataset has reviews of products rather than only movies, and the Movie Review dataset has snippets from professional movie reviewers rather than full-length amateur reviews. We leave all test set examples from IMDB as in-distribution examples, and out-of-distribution examples are the 500 or 1000 test reviews from Customer Review and Movie Review datasets, respectively. Table \[tab:imdbood\] displays detection results, showing a similar story to Table 2. [X \| \[5]{}[>[=1]{}Y]{}]{} Dataset & AUROC/Base & AUPR Succ/Base & AUPR Err/Base & Pred. Prob Wrong(mean) & Test Set Error\ [IMDB]{} & 82 & 97 & 36 & 74 & 11.9\ * [X \| \[4]{}[>[=.4]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base & AUPR In/Base & AUPR Out/Base & Pred. Prob (mean)\ [IMDB/Customer Reviews]{} & 95 & 99 & 60 & 62\ [IMDB/Movie Reviews]{} & 94 & 98 & 80 & 63\ [IMDB/All]{} & 94 & 97 & 84 & 63\ ***** ### Text Categorization We turn to text categorization tasks to determine whether softmax distributions are useful for detecting similar but out-of-distribution examples. In the following text categorization tasks, we train classifiers to predict the subject of the text they are processing. In the 20 Newsgroups dataset [@newsgroups], there are 20 different newsgroup subjects with a total of 20000 documents for the whole dataset. The Reuters 8 [@reuters] dataset has eight different news subjects with nearly 8000 stories in total. The Reuters 52 dataset has 52 news subjects with slightly over 9000 news stories; this dataset can have as few as three stories for a single subject. For the 20 Newsgroups dataset we train a linear classifier on 30-dimensional word vectors for 20 epochs. Meanwhile, Reuters 8 and Retuers 52 use one-layer neural networks with a bag-of-words input and a GELU nonlinearity, all optimized with Adam for 5 epochs. We train on a subset of subjects, leaving out 5 newsgroup subjects from 20 Newsgroups, 2 news subjects from Reuters 8, and 12 news subjects from Reuters 52, leaving the rest as out-of-distribution examples. Table \[tab:caterr\] shows that with these datasets and architectures, we can detect errors dependably, and Table \[tab:catood\] informs us that the softmax prediction probabilities allow for detecting out-of-distribution subjects. [X \| \[5]{}[>[=.6]{}Y]{}]{} Dataset & AUROC/Base & AUPR Succ/Base & AUPR Err/Base & Pred.Prob Wrong(mean) & Test Set Error\ [15 Newsgroups]{} & 89 & 99 & 42 & 53 & 7.31\ [Reuters 6]{} & 89 & 100 & 35 & 77 & 2.53\ [Reuters 40]{} & 91 & 99 & 45 & 62 & 7.55\ ***** [X \| \[4]{}[>[=.5]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base & AUPR In/Base & AUPR Out/Base & Pred. Prob (mean)\ [15/5 Newsgroups]{} & 75 & 92 & 45 & 65\ [Reuters6/Reuters2]{} & 92 & 100 & 56 & 72\ [Reuters40/Reuters12]{} & 95 & 100 & 60 & 47\ ***** ### Part-of-Speech Tagging Part-of-speech (POS) tagging of newswire and social media text is our next challenge. We use the Wall Street Journal portion of the Penn Treebank [@WSJ] which contains 45 distinct POS tags. For social media, we use POS-annotated tweets [@gimpel; @owoputi] which contain 25 tags. For the WSJ tagger, we train a bidirectional long short-term memory recurrent neural network [@lstm] with three layers, 128 neurons per layer, with randomly initialized word vectors, and this is trained on $90\%$ of the corpus for 10 epochs with stochastic gradient descent with a batch size of 32. The tweet tagger is simpler, as it is two-layer neural network with a GELU nonlinearity, a weight initialization according to [@init], pretrained word vectors trained on a corpus of 56 million tweets [@owoputi], and a hidden layer size of 256, all while training on 1000 tweets for 30 epochs with Adam and early stopping with 327 validation tweets. Error detection results are in Table \[tab:poserr\]. For out-of-distribution detection, we use the WSJ tagger on the tweets as well as weblog data from the English Web Treebank [@engwebtb]. The results are shown in Table \[tab:posood\]. Since the weblog data is closer in style to newswire than are the tweets, it is harder to detect whether a weblog sentence is out-of-distribution than a tweet. Indeed, since POS tagging is done at the word-level, we are detecting whether each word is out-of-distribution given the word and contextual features. With this in mind, we see that it is easier to detect words as out-of-distribution if they are from tweets than from blogs. [X \| \[5]{}[>[=1]{}Y]{}]{} Dataset & AUROC/Base & AUPR Succ/Base & AUPR Err/Base & Pred. Prob Wrong(mean) & Test Set Error\ [WSJ]{} & 96 & 100 & 51 & 71 & 3.68\ [Twitter]{} & 89 & 98 & 53 & 69 & 12.59\ *** [X \| \[4]{}[>[=.6]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base & AUPR In/Base & AUPR Out/Base & Pred. Prob (mean)\ [WSJ/Twitter]{} & 80 & 98 & 41 & 81\ [WSJ/Weblog\]{} & 61 & 88 & 30 & 93\ **** Automatic Speech Recognition ---------------------------- Now we consider a task which uses softmax values to construct entire sequences rather than determine an input’s class. Our sequence prediction system uses a bidirectional LSTM with two-layers and a clipped GELU nonlinearity, optimized for 60 epochs with RMSProp trained on $80\%$ of the TIMIT corpus [@timit]. The LSTM is trained with connectionist temporal classification (CTC) [@ctc] for predicting sequences of phones given MFCCs, energy, and first and second deltas of a 25ms frame. When trained with CTC, the LSTM learns to have its phone label probabilities spike momentarily while mostly predicting blank symbols otherwise. In this way, the softmax is used differently from typical classification problems, providing a unique test for our detection methods. We do not show how the system performs on correctness/incorrectness detection because errors are not binary and instead lie along a range of edit distances. However, we can perform out-of-distribution detection. Mixing the TIMIT audio with realistic noises from the Aurora-2 dataset [@aurora], we keep the TIMIT audio volume at 100% and noise volume at 30%, giving a mean SNR of approximately 5. Speakers are still clearly audible to the human ear but confuse the phone recognizer because the prediction edit distance more than doubles. For more out-of-distribution examples, we use the test examples from the THCHS-30 dataset [@chinese], a Chinese speech corpus. Table \[tab:ctctimitood\] shows the results. Crucially, when performing detection, we compute the softmax probabilities while ignoring the blank symbol’s logit. With the blank symbol’s presence, the softmax distributions at most time steps predict a blank symbol with high confidence, but without the blank symbol we can better differentiate between normal and abnormal distributions. With this modification, the softmax prediction probabilities allow us to detect whether an example is out-of-distribution. [X \| \[4]{}[>[=.4]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base & AUPR In/Base & AUPR Out/Base & Pred. Prob (mean)\ [TIMIT/TIMIT+Airport]{} & 99 & 99 & 99 & 59\ [TIMIT/TIMIT+Babble]{} & 100 & 100 & 100 & 55\ [TIMIT/TIMIT+Car]{} & 98 & 98 & 98 & 59\ [TIMIT/TIMIT+Exhibition]{} & 100 & 100 & 100 & 57\ [TIMIT/TIMIT+Restaurant]{} & 98 & 98 & 98 & 60\ [TIMIT/TIMIT+Street]{} & 100 & 100 & 100 & 52\ [TIMIT/TIMIT+Subway]{} & 100 & 100 & 100 & 56\ [TIMIT/TIMIT+Train]{} & 100 & 100 & 100 & 58\ [TIMIT/Chinese]{} & 85 & 80 & 90 & 64\ [TIMIT/All]{} & 97 & 79 & 100 & 58\ ***************** Abnormality Detection with Auxiliary Decoders {#sec:auxiliary} ============================================= Having seen that softmax prediction probabilities enable abnormality detection, we now show there is other information sometimes more useful for detection. To demonstrate this, we exploit the learned internal representations of neural networks. We start by training a normal classifier and append an auxiliary decoder which reconstructs the input, shown in Figure \[fig:decoder\]. Auxiliary decoders are sometimes known to increase classification performance [@swwae]. The decoder and scorer are trained jointly on in-distribution examples. Thereafter, the blue layers in Figure \[fig:decoder\] are frozen. Then we train red layers on clean and noised training examples, and the sigmoid output of the red layers scores how normal the input is. Consequently, noised examples are in the abnormal class, clean examples are of the normal class, and the sigmoid is trained to output to which class an input belongs. After training we consequently have a normal classifier, an auxiliary decoder, and what we call an abnormality module*. The gains from the abnormality module demonstrate there are possible research avenues for outperforming the baseline. TIMIT ----- We test the abnormality module by revisiting the TIMIT task with a different architecture and show how these auxiliary components can greatly improve detection. The system is a three-layer, 1024-neuron wide classifier with an auxiliary decoder and abnormality module. This network takes as input 11 frames and must predict the phone of the center frame, 26 features per frame. Weights are initialized according to [@init]. This network trains for 20 epochs, and the abnormality module trains for two. The abnormality module sees clean examples and, as negative examples, TIMIT examples distorted with either white noise, brown noise (noise with its spectral density proportional to $1/f^2$), or pink noise (noise with its spectral density proportional to $1/f$) at various volumes. We note that the abnormality module is not* trained on the same type of noise added to the test examples. Nonetheless, Table \[tab:frametimit\] shows that simple noised examples translate to effective detection of realistically distorted audio. We detect abnormal examples by comparing the typical abnormality module outputs for clean examples with the outputs for the distorted examples. The noises are from Aurora-2 and are added to TIMIT examples with 30% volume. We also use the THCHS-30 dataset for Chinese speech. Unlike before, we use the THCHS-30 training examples rather than test set examples because fully connected networks can evaluate the whole training set sufficiently quickly. It is worth mentioning that fully connected deep neural networks are noise robust [@seltzer], yet the abnormality module can still detect whether an example is out-of-distribution. To see why this is remarkable, note that the network’s frame classification error is 29.69% on the entire test (not core) dataset, and the average classification error for distorted examples is 30.43%—this is unlike the bidirectional LSTM which had a more pronounced performance decline. Because the classification degradation was only slight, the softmax statistics alone did not provide useful out-of-distribution detection. In contrast, the abnormality module provided scores which allowed the detection of different-but-similar examples. In practice, it may be important to determine whether an example is out-of-distribution even if it does not greatly confuse the network, and the abnormality module facilitates this. [X \| \[6]{}[>[=.45]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base Softmax & AUROC/Base AbMod & AUPR In/Base Softmax & AUPR In/Base AbMod & AUPR Out/Base Softmax & AUPR Out/Base AbMod\ [TIMIT/+Airport]{} & 75 & 100 & 77 & 100 & 73 & 100\ [TIMIT/+Babble]{} & 94 & 100 & 95 & 100 & 91 & 100\ [TIMIT/+Car]{} & 70 & 98 & 69 & 98 & 70 & 98\ [TIMIT/+Exhib.]{} & 91 & 98 & 92 & 98 & 91 & 98\ [TIMIT/+Rest.]{} & 68 & 95 & 70 & 96 & 67 & 95\ [TIMIT/+Subway]{} & 76 & 96 & 77 & 96 & 74 & 96\ [TIMIT/+Street]{} & 89 & 98 & 91 & 99 & 85 & 98\ [TIMIT/+Train]{} & 80 & 100 & 82 & 100 & 77 & 100\ [TIMIT/Chinese]{} & 79 & 90 & 41 & 66 & 96 & 98\ Average & 80 & 97 & 77 & 95 & 80 & 98\ ***************** MNIST ----- Finally, much like in a previous experiment, we train an MNIST classifier with three layers of width 256. This time, we also use an auxiliary decoder and abnormality module rather than relying on only softmax statistics. For abnormal examples we blur, rotate, or add Gaussian noise to training images. Gains from the abnormality module are shown in Table \[tab:mnistround2\], and there is a consistent out-of-sample detection improvement compared to softmax prediction probabilities. Even for highly dissimilar examples the abnormality module can further improve detection. [X \| \[6]{}[>[=.36]{}Y]{}]{} In-Distribution /Out-of-Distribution & AUROC/Base Softmax & AUROC/Base AbMod & AUPR In/Base Softmax & AUPR In/Base AbMod & AUPR Out/Base Softmax & AUPR Out/Base AbMod\ [MNIST/Omniglot]{} & 95 & 100 & 95 & 100 & 95 & 100\ [MNIST/notMNIST]{} & 87 & 100 & 88 & 100 & 90 & 100\ [MNIST/CIFAR-10bw]{} & 98 & 100 & 98 & 100 & 98 & 100\ [MNIST/Gaussian]{} & 88 & 100 & 88 & 100 & 90 & 100\ [MNIST/Uniform]{} & 99 & 100 & 99 & 100 & 99 & 100\ Average & 93 & 100 & 94 & 100 & 94 & 100\ ********* Discussion and Future Work ========================== The abnormality module demonstrates that in some cases the baseline can be beaten by exploiting the representations of a network, suggesting myriad research directions. Some promising future avenues may utilize the intra-class variance: if the distance from an example to another of the same predicted class is abnormally high, it may be out-of-distribution [@giryes]. Another path is to feed in a vector summarizing a layer’s activations into an RNN, one vector for each layer. The RNN may determine that the activation patterns are abnormal for out-of-distribution examples. Others could make the detections fine-grained: is the out-of-distribution example a known-unknown or an unknown-unknown? A different avenue is not just to detect correct classifications but to output the probability of a correct detection. These are but a few ideas for improving error and out-of-distribution detection. We hope that any new detection methods are tested on a variety of tasks and architectures of the researcher’s choice. A basic demonstration could include the following datasets: MNIST, CIFAR, IMDB, and tweets because vision-only demonstrations may not transfer well to other architectures and datasets. Reporting the AUPR and AUROC values is important, and so is the underlying classifier’s accuracy since an always-wrong classifier gets a maximum AUPR for error detection if error is the positive class. Also, future research need not use the exact values from this paper for comparisons. Machine learning systems evolve, so tethering the evaluations to the exact architectures and datasets in this paper is needless. Instead, one could simply choose a variety of datasets and architectures possibly like those above and compare their detection method with a detector based on the softmax prediction probabilities from their classifiers. These are our basic recommendations for others who try to surpass the baseline on this underexplored challenge. Conclusion ========== We demonstrated a softmax prediction probability baseline for error and out-of-distribution detection across several architectures and numerous datasets. We then presented the abnormality module, which provided superior scores for discriminating between normal and abnormal examples on tested cases. The abnormality module demonstrates that the baseline can be beaten in some cases, and this implies there is room for future research. Our hope is that other researchers investigate architectures which make predictions in view of abnormality estimates, and that others pursue more reliable methods for detecting errors and out-of-distribution inputs because knowing when a machine learning system fails strikes us as highly important. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank John Wieting, Hao Tang, Karen Livescu, Greg Shakhnarovich, and our reviewers for their suggestions. We would also like to thank NVIDIA Corporation for donating several TITAN X GPUs used in this research. Abnormality Module Example ========================== [^1]: Work done while the author was at TTIC. Code is available at [github.com/hendrycks/error-detection ](https://github.com/hendrycks/error-detection ) [^2]: We consider adversarial example detection techniques in a separate work [@detectadversarial]. [^3]: A debatable, imprecise interpretation of AUROC values may be as follows: 90%\|100%: Excellent, 80%\|90%: Good, 70%\|80%: Fair, 60%\|70%: Poor, 50%\|60%: Fail. [^4]: We also tried using the KL divergence of the softmax distribution from the uniform distribution for detection. With divergence values, detector AUROCs and AUPRs were highly correlated with AUROCs and AUPRs from a detector using the maximum softmax probability. This divergence is similar to entropy [@jacob; @williams].	{ "pile_set_name": "ArXiv" }
--- author: - 'I. L. San Sebastián [^1], O. M. Guilera' - 'M. G. Parisi' bibliography: - 'bibliografia.bib' date: 'Received xxx; accepted xxx' subtitle: dependencies with planetesimal relative velocities and compositions title: 'Planetesimal fragmentation and giant planet formation II:' --- [Most of planet formation models that incorporate planetesimal fragmentation consider a catastrophic impact energy threshold for basalts at a constant velocity of 3 km/s during all the process of the formation of the planets. However, as planets grow the relative velocities of the surrounding planetesimals increase from velocities of the order of m/s to a few km/s. In addition, beyond the ice line where giant planets are formed, planetesimals are expected to be composed roughly by 50 % of ices.]{} [We aim to study the role of planetesimal fragmentation on giant planet formation considering planetesimal catastrophic impact energy threshold as a function of the planetesimal relative velocities and compositions.]{} [We improve our model of planetesimal fragmentation incorporating a functional form of the catastrophic impact energy threshold with the planetesimal relative velocities and compositions. We also improve in our model the accretion of small fragments produced by the fragmentation of planetesimals during the collisional cascade considering specific pebble accretion rates.]{} [We find that a more accurate and realistic model for the calculation of the catastrophic impact energy threshold tends to slow down the formation of massive cores. Only for reduced grain opacity values at the envelope of the planet, the cross-over mass is achieved before the disk time-scale dissipation.]{} [While planetesimal fragmentation favors the quick formation of massive cores of $5-10~\text{M}_{\oplus}$ the cross-over mass could be inhibited by planetesimal fragmentation. However, grain opacity reduction or pollution by the accreted planetesimals together with planetesimal fragmentation could explain the formation of giant planets with low-mass cores.]{} Introduction {#sec1} ============ The standard model of giant planet formation is the core-accretion mechanism [@Mizuno1980; @Stevenson1982; @Pollack1996; @Guilera2010; @Helled2014]. According to this model, the planet forms from accretion of planetesimals onto a solid core until it has enough mass to start accreting gas from the protoplanetary disk. When the cross-over mass is reached, the mass of the envelope equals the core mass (at $\sim 10 M_{\oplus}$), and the planet starts to accrete big quantities of gas in a short period of time, this process is known as runaway gas accretion. At some point, for mechanisms not yet well understood, the accretion of gas onto the planet is limited. Finally, the planet evolves cooling and contracting at constant mass. There is an alternative model developed in the past few years, also based in the accretion of solid material, for the formation of giant planets. This model is based on the accretion of small sized particles (called pebbles) for the formation of the planet core. Unlike planetesimals, pebbles can be accreted by the full Hill sphere making their accretion rates significantly larger than planetesimal accretion rates [@OrmelKlahr2010; @LJ2012]. In the standard core accretion mechanism an important issue is the formation of a massive core before the dissipation of the gaseous component of the protoplanetary disk. In this framework, the process of accretion of planetesimals plays a fundamental role. As the planetary embryo grows, the surrounding planetesimals increase their relative velocities due to the gravitational stirring produced by the embryo, leading to disruptive collisions among them. This effect produces a cascade of collision fragments. [@Inaba2003], [@Kobayashi2011], [@OrmelKobayashi2012], and @Guilera2014 found that excessive fragmentation may cause that the smaller fragments produced by these collisions drift inwards by gas drag and a significant amount of mass could be lost stalling the oligarchic growth. Moreover, when the core reaches several Earth masses the relative planetesimal velocities become high enough to produce supercatastrophic collisions among planetesimals [@Guilera2014; @Chambers2014]. In this case, @Guilera2014 considered that the planetesimals involved in such collision are pulverized and the mass of such bodies is lost. On the other hand, [@Chambers2014] found that small planetesimals collide frequently and produce rapid embryo growth. Unlike [@Guilera2014], [@Chambers2014] considered that the mass that would go into fragments smaller than the minimum size considered in the model (referred as pebbles of second generation) is assigned to become an equivalent mass of the minimum size adopted in their model, since they assume that tiny fragments quickly coagulate into new pebbles. Despite the different planetesimal fragmentation models, during the process of planetary formation, fragmentation of planetesimals is an important effect that may inhibit or favor the formation of giant planet cores. In a collision between two planetesimals, the catastrophic impact energy threshold required to fragment and disperse fifty percent of the target mass (also known as specific impact energy) is an important function that has to be defined in models of planetesimal fragmentation. This function depends on many factors of the collision, such as the impact velocity, the planetesimal type of material and its porosity, the sizes of the planetesimals and other factors that may affect the outcome of the collision [see @Jutzi2015]. For example, @BenzAsphaug1999 found that a target of a fixed size compose by ice-material has catastrophic impact energy threshold lower than if it were composed by basalt. [@Benz2000] also found that low velocity collisions are more efficient in destroying and dispersing basaltic bodies than collisions at higher velocities. [@Jutzi2010] studied from SPH simulations the porosity of the planetesimals in a collision and found that, in the strength regime, porous targets are more difficult to disrupt that non-porous ones while in the gravity regime, the outcome of the collision for porous targets depends on gravity and porosity, unlike non-porous ones that are also controlled by the strength. Recently, @Beitz-et-al-2011 and @Bukhari-et-al-2017 showed in low-velocity impact experiments that the strength of compacted dust aggregates is much weaker than that of porous rocks. This outcome is analogous to @Jutzi2010. It is important to remark here, that most models of planet formation that incorporate the planetesimal fragmentation process use the prescription of [@BenzAsphaug1999] for the catastrophic impact energy threshold for non-porous basalts at impact velocities of $3$ km/s [@Kobayashi2011; @OrmelKobayashi2012; @Guilera2014; @Chambers2014]. However, @Guilera2014 showed that during the formation of massive cores the relative planetesimal velocities are incremented from cm/sec–m/sec to km/sec as planet grows. In our previous work [@Guilera2014], we incorporated a model of planetesimal fragmentation into our global model of giant planet formation [@Guilera2010]. We found that the formation of massive cores are only possible from an initial population of big planetesimals and massive disks. In other situations, planetesimal fragmentation tends to inhibit the formation of massive cores before the dissipation of the disk. In this new work, we incorporate to our planetesimal fragmentation model the dependence of the catastrophic impact energy threshold with material composition and planetesimal relative velocities. We also incorporate different velocity regimes for planetesimal relative velocities (Keplerian shear and dispersion dominate regimes) in the planetesimal fragmentation model, and pebbles accretion rates for the small particles product of the collision process into the global model of giant planet formation. The aim of this work is to analyze the impact onto the formation of massive cores of the new phenomena incorporated, specially planetesimals relative velocities and planetesimal compositions. This paper is organized as follows: in Sec. \[sec2\], we describe the improvements of our model of giant planet formation and the improvements of the model of planetesimal fragmentation; in Sec. \[sec3\], we present the results of our simulations for the formation of a giant planet, and finally, in Sec.\[sec4\] we present the summary and conclusions of our results. Our model of giant planet formation {#sec2} =================================== In a series of previous works [@Guilera2010; @Guilera2014] we developed a numerical model that describes the formation of giant planets immersed in a protoplanetary disk that evolves in time. In our model, the protoplanetary disk is represented by a gaseous and a solid components. Planets grow by the simultaneous accretion of solids and gas. The solid component of the disk evolves by planet accretion, radial drift due to nebular drag and collisional evolution, while the gaseous component evolves by an exponential decay (see Appendix \[sec5-1\] for technical details on the model). In this work we improve our planetesimal fragmentation model developed in @Guilera2014 incorporating a dependence of $Q^_d$, the catastrophic impact energy threshold, with the planetesimal relative velocities and the planetesimal composition considering a mix of non-porous ices and, as a proxy for the rocky material of planetesimals, non-porous basalts. We also incorporate pebble accretion rates for the small fragments product of the planetesimal fragmentation process (hereafter, pebbles of second generation), and different velocity regime models for the calculation of low and high relative planetesimal velocities. The improvements in our model are presented in the next subsections. Improvements in the solid accretion rates {#sec2-1} ----------------------------------------- In the past few years, a new model of solid particles accretion has been proposed as an alternative of the core accretion model for the formation of solid massive cores. Such model proposes that massive cores, precursors of giant planet cores, can be quickly formed by 100-1000 kilometers embryos accreting cm-sized particles known as pebbles. @OrmelKlahr2010 and @LJ2012 found that pebbles, particles with Stokes numbers below the unity ($S_{t} \lesssim1$), are strongly coupled to the gas and can be very efficiently accreted by hundred-thousand sized embryos. While planetesimals can be accreted by a fraction of the Hill radius of the growing planet, pebbles can be accreted by the full Hill radius, making pebble accretion rates significantly larger than planetesimal accretion rates. However, the dominant initial size of the accreted solids (cm-sized particles named pebbles or 1-100 km-sized bodies named planetesimals) is unknown [@Helled2014; @Johansen2014]. At the begging of our simulations, the solid component of the protoplanetary disk is composed by planetesimals of radius 100 km with the distribution given by the framework of the oligarchic growth [@KokuboIda1998; @KokuboIda2000; @KokuboIda2002; @IdaMakino1993]. Thus, we assume that there are no initial pebbles in our model. However, pebbles appear as result of the planetesimal collisional evolution, i.e., second generation pebbles. Following @Guilera2016 and @Guilera-Sandor2017, we improve the solid accretion rates by incorporating the pebble accretion rate for pebbles of second generation in addition to the planetesimals accretion. For planetesimals we adopt the accretion rates given by [@Inaba2001] and for pebbles of second generation we use the pebble accretion rates given by [@Lambrechts2014] with a reduction factor that takes into account the scale height of the pebbles in comparison with the Hill radius of the planet (see Appendix \[sec5-2\]). Improvements to our planetesimal fragmentation model {#sec2-2} ---------------------------------------------------- Our planetesimal fragmentation model, which is based in the Boulder Code ([@Morbidelli2009] and Supplementary Material), describes the collisional evolution of the planetesimal population. A brief description of the planetesimal fragmentation model and technical details are presented in Appendix \[sec6\]). ### Velocities and probabilities of collision regimes {#sec-2-2-1} Following @Morbidelli2009, the impact rate between targets and projectiles estimated in our model is given by $$\begin{aligned} \text{Impact rate}\|_{\text{A}}= \frac {\alpha V_{rel}}{4 H a (\delta a + 2 a e_p)} E \left(1 + b \frac{V_{esc}^2}{V_{rel}^2}\right) (R_T + R_P)^2, \label{eq1-sec-2-2-1}\end{aligned}$$ where $V_{rel}$ is the relative velocity between targets and projectiles, $\alpha$ is a coefficient that depends on $V_{rel}$, $H$ is the mutual scale height, $a$ and $\delta a$ are the mean semi-major axis and the width of the annulus that contains the targets and projectiles, respectively, $e_p$ is the eccentricity of the projectiles, $E$ is a coefficient that considers the deviation of the gravitational focusing corresponding to the two-body problem at low relative velocities, $b$ is also a function of $V_{rel}$, and $V_{esc}$ is the mutual escape velocity. Finally, $R_T$ and $R_P$ are the radius of the targets and projectiles, respectively [see @Morbidelli2009 for details]. It is important to note, that the relative velocity among planetesimals used in the Boulder code is the dispersion velocity (the velocity corresponding to the dispersion regime). However, when the planetesimal relative velocity tends to zero, the impact rate between them is undetermined. [@Weidenschilling2011] noticed this problem and incorporated the relative velocity corresponding to the keplerian shear. Following [@Weidenschilling2011], we incorporate in our planetesimal fragmentation model different velocity regimes and probabilities of collision for lower relative velocities given by [@Greenberg1991] to calculate collision rates more accurately. We consider three different regimes and their transitions according to [@Greenberg1991], regime A: dominance by random motion (impact rate given by Eq.\[eq1-sec-2-2-1\]); regime B: dominance by Keplerian shear motion; regime C: Keplerian shear dominance in a very thin disk (see Appendix \[sec6-2\] for details). ### Catastrophic impact energy threshold {#sec2-2-2} ![Catastrophic disruption thresholds for basalt targets at impact velocities of 20-30 m/s [@Benz2000], 3km/s and 5km/s [@BenzAsphaug1999]. The blue points correspond to the discrete data extracted from [@Benz2000]. []{data-label="basalts"}](qdbasaltos.pdf){width="\hsize"} ![Catastrophic disruption thresholds for icy targets at impact velocities of 0.5 km/s and 3 km/s [@BenzAsphaug1999]. []{data-label="ice"}](qdhielos.pdf){width="\hsize"} The catastrophic impact energy threshold per unit target mass $Q_{D}^$ is the energy needed to fragment and disperse half of the target mass in an impact, i.e., the threshold for catastrophic disruption. This quantity plays an important role in the collisional evolution of the planetesimal population. As we mentioned in Sec. \[sec1\], $Q_{D}^$ depends on many factors of the collision; particularly important are the relative velocity among planetesimals, which determines the impact velocity of the collision, and the planetesimals composition. Fig. \[basalts\] represents $Q_{D}^$ as a function of the radius of a non-porous basalt target for three different values of the impact velocity ($\sim 25$ m/s, 3 km/s and 5km/s), while Fig.\[ice\] represents this threshold for a non-porous icy target for two values of the impact velocity (0.5 km/s and 3 km/s) (see Appendix \[sec6-3\]). For basalt targets, we can see that for a fixed target’s radius, the smaller the impact velocity, the smaller $Q_{D}^$. This phenomenon can have an important effect during the formation of massive cores due to the fact that, initially, planetesimal relative velocities are low and are then incremented as the planet grows. We remark again that in most works of planetary formation that include planetesimal fragmentation it is used $Q_{d}^$ for non-porous basalt targets at an impact velocity of 3 km/s [@OrmelKobayashi2012; @Chambers2014]. Given the functional dependency of $Q_{d}^$ with target’s radius for different impact velocities given by [@BenzAsphaug1999] and [@Benz2000], for a fixed target’s radius, we implemented an interpolation between the impact velocity dependencies to get an improved value of $Q_{d}^$ as a function of the impact velocity. We also include an interpolation using the different curves of $Q^_d$ of [@BenzAsphaug1999] for different velocities for ices. We note that if the impact velocities are greater or lower than the velocities corresponding to the upper and lower curves of $Q_{d}^$ adopted, we do not extrapolate $Q_{d}^$. In that case, we adopt $Q_{d}^$ corresponding to the maximum or minimum velocity used in [@BenzAsphaug1999] and [@Benz2000]. On the other hand, according to [@Lodders2003], beyond the iceline the amount of solid mass in the protoplanetary disk increases by a factor of two meaning that fifty percent of the material behind the iceline should be condensated in the primitive Solar System. The ice-to-rock ratio derived from trans-Neptunian objects, comets and irregular satellites of giant planets [@McDonnell1987; @Stern1997; @JohnsonLunine2005] confirms this. As we are interested in the formation of giant planets behind the iceline, we assume that planetesimals are composed by ices and basalts following [@Lodders2003]. Finally, to implement the dependence of $Q^_D$ with the impact velocity and the target’s composition, we first interpolate between the curves of $Q_{d}^$ as a function of the target’s radius for the different impact velocities, obtaining the values of $Q^_D$ for each pure material (basalts and ices). Then, we make a linear combination between them, depending on the percentage of basalts and ices that we define for the solids (in this work we considered three cases, planetesimals composed purely by basalts, composed purely by ices, and composed fifty percent by basalts and fifty percent by ices). We remark here again that in all the cases $Q_{d}^$ is calculated by Eq.(\[eq1-sec2-2-2\]), but using the effective radius ($r_{\text{eff}}= 3 (M_{T}+M_{P})/4 \pi \rho$) instead of the target’s radius. It should be noted that @LyS2012 also obtain a derivation of a general catastrophic disruption law. However, their work is focus on the gravitational regime, and for bodies of different porosities, while in our model we adopt compatible laws valid in the gravitational as well as in the strength regime for non-porous bodies. Results {#sec3} ======= We aim to study the impact of the improvements on our planetesimal fragmentation model discussed above on the formation of a giant planet. We carried out a set of different simulations including one phenomenon at a time. Our simulations start at the beginning of the oligarchic growth with a moon-sized embryo located at 5 au from the central star. Initially, the embryo is immersed in an homogeneous single-sized population of planetesimals of 100 km of radius and the disk is ten times more massive than the Minimum Mass Solar Nebula [MMSN, @Hayashi1981]. The simulations stop when the mass of the envelope equals the core mass, i.e. when the cross-over mass is achieved (in this case we consider that the planet ends its formation in a short period of time after the gaseous runaway growth starts), or at 6 Myr when we consider that the gas of the disk is dissipated. Baseline model {#sec3-1} -------------- In [@Guilera2014] we included a first approach to the modelling of planetesimal fragmentation into our model of giant planet formation [@Guilera2010], and we studied, for different initial planetesimal sizes and disk masses, how the collisional evolution of the planetesimal population modified the planetary formation process. We remark that in @Guilera2014, the accretion rates of particles with Stokes number less than the unity, were calculated using the prescription derived by @Inaba2001 in the low-velocity regime. We found that only for initial large planetesimals ($r_{p}=$ 100 km) and massive disks, and only if most of the mass loss in collisions was distributed in larger fragments (see Eq.\[eq3-sec2-2\]), planetesimal fragmentation favored the relatively rapid formation of a massive core (larger than $10~M_{\oplus}$). If smaller planetesimals are considered, the planetesimal fragmentation process inhibits the formation of massive cores. We remark that these results in agreement with previous works that adopt similar hypothesis for the model of planetesimal fragmentation [@Inaba2003; @Kobayashi2011; @OrmelKobayashi2012]. Thus, we choose, as our baseline case for this work, the most favorable simulation in [@Guilera2014] that corresponds to the case of initial planetesimals of 100 km radius and a 10 MMSN initial disk mass. In Fig. [\[fiducialresult\]]{}, we show the time evolution of the core mass and envelope mass of the planet, for the baseline model and for the formation of the planet without considering the planetesimal fragmentation process. We can see that the inclusion of planetesimal fragmentation reduces by $\sim$ 7% the time to reach the cross-over mass. We can also observe that the planet reaches the cross-over mass at a lower core mass (18% lower) . However, the fragmentation model incorporated in [@Guilera2014] considered the catastrophic disruption threshold given by [@BenzAsphaug1999] for basalts at impact velocities of 3 km/s. This is clearly a simplification. On one hand, as we are studying the formation of a giant planet behind the iceline, it should be taken into account that planetesimals are composed by rocks and ices [@Lodders2003]. Also, during the process of planet formation, the planetesimal relative velocities are increased due to the gravitational perturbations produced by the growing planet. In the left panel of Fig. \[fiducialresult2\], we show the increase of eccentricities and inclinations for planetesimals of 100 km of radius, and in the right panel their relative velocities, nearby the planet meanwhile it grows. We can see that eccentricities and inclinations are quickly increased near the planet. This leads to an increase of the planetesimal relative velocities from velocities of some meters per second to velocities of about 5 km/s (right panel of Fig.\[fiducialresult2\]). These results are showing that clearly the assumption of a $Q_D^$ given for a constant impact velocity is a simplification in the model of planetesimal fragmentation, and thus, it motivated us to incorporate the following dependencies in the catastrophic disruption threshold. Dependencies of $Q^_D$ with the planetesimal impact velocities and the compositions {#sec3-2} ------------------------------------------------------------------------------------ As we mentioned above, in general, models of planet formation that include planetesimal fragmentation consider the catastrophic impact energy threshold for a fixed velocity of basaltic planetesimals. In this section, we explore how the dependencies of the catastrophic disruption threshold with the planetesimal compositions and with the planetesimal impact relative velocities impact affect the formation of the planet. To do this, we first calculate the formation of the planet considering the same initial parameters as in the baseline model (an embryo located at 5 au in a disk 10 times more massive than the MMSN and with initial planetesimals of 100 km of radius) but now we implement the dependence of $Q_D^$ with different impact velocities following the approaches described in Sec. \[sec2-2-2\]. The solid accretion rates are calculated following Eq.(\[eq1-sec2-1\]). In Fig. \[masacorevstime\], we show the time evolution of the core mass and envelope mass of the planet, for the model wherein planetesimal fragmentation is not considered, for the baseline model, and for the case where we consider basaltic planetesimals with $Q^_D$ as a function of the relative velocities among planetesimals. We can see that the cross-over mass, for the last case, is achieved in a time longer than for the baseline model. Moreover, the formation of the giant planet takes $\sim 25$% more time compared to the case wherein planetesimal fragmentation is not considered. In Fig. \[accretionratevstime1\], we plot the solid accretion rates for the different sizes of planetesimals and small particles for both models, the baseline model and the model that considers $Q^_D$ as a function of the relative velocities among planetesimals. We can see that the accretion of planetesimals smaller than 100 km and small particles generated by the collisional evolution becomes effective at a longer time for the model where $Q^_D$ is a function of the relative velocities. We can also see that in both models, the total solid accretion rate is always dominated by plantesimals of 100 km. However, while in the baseline model planetesimals between 1 km and 25 km are the products of the fragmentation process that most contribute to the total solid accretion rate (being the planetesimals of $\sim 25$ km the most important), in the model where $Q^_D$ is a function of the relative velocities, the most important contribution from the fragments is given by planetesimals between 10 km and $\sim 16$ km. Moreover, in this last case, planetesimals of 1 km has a negligible contribution to the total solid accretion rate. This is due to the fact that these bodies are near the minimum of $Q^_D$ for low impact velocities, which is for this size, one order of magnitude lower with respect to the $Q^_D$ at impact velocities of 3 km/s. We can see from Fig. \[velrel1\] that relative velocities for planetesimals of 1 m reach 20-30 m/s at 1 Myr (when the mass of the core is $\sim 1~\text{M}_{\oplus}$), and 1 km-sized planetesimals exceed these values rapidly, at $10^{-2}$ Myr (when the mass of the core is only $\sim 0.2~\text{M}_{\oplus}$) the relative velocity is over 200 m/s. It is important to remark that, in both cases, their velocities are always below 3 km/s. \ \ On the other hand, in Fig. \[masacorevstime2\] we plot the time evolution of the core mass and envelope mass of the planet, for a model that considers planetesimals purely composed by ices and adopting $Q^_D$ for ices at impact velocities of 3 km/s, and for a model wherein planetesimals are composed of 50% of basalts and 50% of ices and $Q^_D$ is calculated by a linear combination between $Q^_D$ for basalts at 3 km/s and $Q^_D$ for ices at 3 km/s. We also show, by completeness and comparison, the baseline model and the case where planetesimal fragmentation is not considered. We can see that when we considered planetesimals purely composed by ices, the cross-over mass is never achieved. This is due to the fact that for a same fixed impact velocity (in this case 3 km/s), $Q^_D$ is lower for ices than for basalts. Thus, the collisional evolution of the planetesimal population is different, small fragments are disrupted more efficiently, and the accretion rates of such fragments do not compensate the fragmentation of big bodies. Besides, if we compare $Q^_D$ from Fig.\[basalts\] and Fig.\[ice\] we can see that $Q^_D$ for ices at 3 km/s is lower than $Q^_D$ for basalts at 20-30 m/s for planetesimals of $\sim$ 100 km. Therefore, this difference can explain why in the model that considers $Q^_D$ as a function of the relative velocities for basalts the giant planet core could form during the disk lifetime and, for the model that adopts $Q^_D$ for ices at 3 km/s, the cross-over mass was not achieved in that time. When planetesimals are composed by 50% of basalts and 50% of ices, and $Q^_D$ is calculated as a linear combination of the corresponding $Q^_D$ for basalts and ices at impact velocities of 3 km/s, respectively, the cross-over mass is achieved in less than 6 Myr with a core mass of 16 $M_{\oplus}$. However, the planet achieves the cross-over mass at a longer time, by $\sim$ 12 $\%$, compared to the case without planetesimal fragmentation. All these results show that the collisional evolution of the planetesimal population and the accretion of the fragments produced in such collisional evolution play and important role in the formation of a giant planet. We point out that in this subsection, the accretion rates of particles with Stokes number less than the unity, was also computed as in the baseline model, using the prescription derived by @Inaba2001 in the low-velocity regime. Accretion of small fragments: pebble accretion of second generation {#sec3-3} ------------------------------------------------------------------- As we mentioned before, we improved in our model the solid accretion rates incorporating in this work the accretion rates of pebbles (Eq.\[eq2-sec2-1\]) derived by @Lambrechts2014 for particles with Stokes numbers less or equal to the unity. In @Guilera2014, the accretion rates of such particles were calculated using the prescription derived by @Inaba2001 in the low-velocity regime (Eq.\[eq1-sec2-1\]). In that work, we showed that the accretion rates for particles smaller that $\sim 1$ m could become greater than the pebbles accretion rates [see @Guilera2014 for the details]. Thus, in order to calculate in a more accurate way the accretion of small particles, we implemented the pebble accretion rates mentioned above. In Fig. \[masacorevstime3\], we show the time evolution of the core mass and the envelope mass for the case where planetesimals are purely composed by basalts adopting the corresponding $Q^_D$ at impact velocities of 3 km/s, and where small particles with Stokes number less or equal to the unity are accreted as pebbles using Eq.\[eq2-sec2-1\], while in the baseline model the accretion rate for planetesimals (Eq.\[eq1-sec2-1\]) is applied also for small particles. We also plot for comparison the baseline model and the case wherein planetesimal fragmentation is not considered. We can see that a more accurate treatment for the accretion of small particles delays the formation of the giant planet by about 30%, despite pebbles do not mostly contribute to the total solid accretion rate (Eq.\[eq2-sec2-1\]). This effect is discussed in @Guilera2014 where the rates of small fragments using the prescriptions of @Inaba2001 and @LambrechtsJohansen2012 are compared. Following that analysis we here obtain the accretion rates of @Inaba2001 in terms of the pebble accretion rates of @Lambrechts2014 $$\begin{aligned} \dot{M}= \begin{cases} 5.65 \sqrt{\tilde{R}_C/R_H} \dot{M}_H , \quad \text{if } ~ 0.1 \le S_t < 1, \\ \\ 5.65 \sqrt{\tilde{R}_C/R_H} \left( \frac{S_t}{0.1} \right)^{-2/3} \dot{M}_{H2} , \quad \text{if }~S_t < 0.1 , \end{cases} \label{accretionratespebblesplanetesimals}\end{aligned}$$ where $\tilde{R}_C$ is the enhanced radius due to the planet’s gaseous envelope, $\dot{M}_H=2 R_H^2 \Sigma_p \Omega_P$ and $\dot{M}_{H2}= \left( \frac{S_t}{0.1} \right)^{2/3} \dot{M}_H$, where $\dot{M}_{H}$ for $0.1 \le S_t < 1$ and $\dot{M}_{H2}$ for $S_t < 0.1$ are the pebble accretion rate $dM_C/dt$ of Eq. \[eq2-sec2-1\]. We can see from Fig. 15 of @Guilera2014 that when the planet’s core mass becomes larger than $\sim 0.2 M_{\oplus}$ planetesimal accretion rates are higher than pebble accretion rates for fragments with $r_P \lesssim 1$ m. In our work this is the case for particles with $0.1 \le S_t < 1$, but for fragments with $S_t < 0.1$, the difference between the planetesimal and pebble accretion rates is even higher due to the factor $\left( \frac{S_t}{0.1} \right)^{-2/3}$. Moreover, we introduced a factor $\beta = min (1, R_H/H_P)$ to include a reduction in the pebble accretion rates if the scale height of the pebbles becomes greater than the Hill radius of the planet, this is not taken into account in the planetesimal accretion rate for small fragments in the baseline model. In Fig. \[scaleheight\] we show $R_H/H_P$ as a function of time, where we can observe that objects between $1-10$ cm have values of $R_H/H_P$ below 1. The quantified differences are shown in Fig. \[accretionratesvstime14\] where we present the comparison of the solid accretion rates for small particles with $S_t \lesssim 1$ of the baseline case and the simulation including pebble accretion. We can see from Fig. \[accretionratesvstime14\] that the accretion rates in the baseline model are higher than in the simulation that takes into account the accretion of pebbles as explained before. Global model {#sec3-4} ------------ In this section, we compare the results for the formation of the giant planet obtained with the baseline model and with the global model. The global model includes all the improvements on the calculation of $Q^_D$, i.e., the dependencies of $Q^_D$ with the planetesimal impact velocities and their compositions described in Section \[sec3-2\], and the accretion of second generation pebbles described in Section \[sec3-3\]. In the global model where we include all the new effects on $Q^_D$ and a more accurate and realistic treatment for the accretion of the small particles, the core does not achieve the cross-over mass within the 6 Myr as shown in Fig. \[masacorevstime4\]. We point out this model also includes the different regimes for the calculation of the relative velocities (and the impact rates) discussed in Sec. \[sec-2-2-1\]. However, as the planet quickly excites the relative velocities of the near planetesimals, they are in general in the dispersion regime, and thus the keplerian shear regime does not play a relevant role. Finally, we remark that despite the planet does not achieve the cross-over mass, a solid core of a few Earth masses, up to 5 Earth masses, is formed more quickly compared to the case where planetesimal fragmentation is not considered. @Ikoma2000 and @Hubickyj2005 showed that a reduction in the grain opacity in the planet envelope as well as the pollution of the envelope [due to evaporated materials of icy planetesimals in the envelope, see @Hori-Ikoma-2011], could reduce the mass of the core at which the planet reaches the cross-over mass, and thus it reduces the formation time. We will explore this possibility in the next section. Reduced grain opacities {#sec3-5} ----------------------- In the previous sections we studied the process of giant planet formation considering fragmentation of plantesimals including for the catastrophic disruption threshold the dependencies with relative velocity and composition of the colliding bodies. We also include the accretion rates for small fragments called pebbles. We analyzed every improvement separately and finally all together and we found that, in all the cases, the formation of the giant planet core is slower. Motivated by this result we incorporate a physical phenomenon that could act in an opposite way, accelerating the formation of a giant planet, in this case a reduction in the grain opacities of the planet envelope. The grain opacities in the planet envelope play an important role on the formation of a giant planet. [@Ikoma2000] found that if the grain opacities are small enough, a small size giant planet core can capture significant amounts of gas. Later, [@Hubickyj2005] found that if the grain opacities are assumed as 2% of the interstellar medium value, the planet can reach the cross-over mass with a core mass between $5-10~M_{\oplus}$. More recently, @Hori-Ikoma-2010 showed that a core of only $\sim 2~M_{\oplus}$ is able to capture enough gas to form a giant planet if the accreting envelope is grain-free. Thus, following these works, specially the work of @Hubickyj2005, we compute a set of simulations, with and without planetesimal fragmentation, reducing the grain opacities in the envelope up to a 2% of the interstellar medium value. For the case where we consider the planetesimal fragmentation, we adopt the global model described above. In Fig. \[masacorevstime5\], we present the time evolution of the core mass and the envelope mass of the planet for the models that consider reduced grain pacities with and without planetesimal fragmentation and the baseline case. When we calculate the formation of the giant planet without considering the planetesimal fragmentation process, the time at which the planet achieves the cross-over mass is reduced in more than 50% when the grain opacities in the planet envelope are reduced. However, the cross-over mass remains practically similar. These results are in very good agreement with the previous results of @Hubickyj2005. A reduction in the envelope opacity allows the planet to release more efficiently the heat generated by the accretion of the planetesimals, and as consequence, the gas accretion becomes more efficient. We can see this effect in Fig.\[mgas-vs-mcore\] where we plot the envelope mass as a function of the core mass for the two models without planetesimal fragmentation. At a fixed core mass, the model with the reduced grain opacities has a greater value for the envelope mass. Finally, we note that when planetesimal fragmentation is considered, the planet achieves the cross-over mass at a time (1.66 Myr) and with a core mass (16.44 $M_{\oplus}$) that are lower than for the case where planetesimal fragmentation is not considered ($\sim$ 7$\%$ and $\sim$ 24$\%$ lower in time and core mass respectively). We associate this result to the fact that the amount of envelope mass in the planet at low-mass cores plays an important role, significantly enhancing the capture radius of the planet. Comparison with previous work {#sec3-6} ============================= A similar model of giant planet formation was developed by @Chambers2014. They carried out numerical simulations of oligarchic growth including pebble accretion and planetesimal fragmentation. One of the major differences between their model and ours is that they consider that all the collisions involving targets smaller than their minimum bin size are assumed to coagulate and grow rapidly approaching such minimum size, while we assume that the mass distributed under $r_p^{min}=1$ cm is lost. In our fragmentation model most of the mass is in the largest fragments. As discussed in @Guilera2014, the results are analogous when the minimum bin size $r_p^{min}$ is reduced to lower values. Also, @Chambers2014 uses a fixed value of $Q_D^$ (for basalts at 3 km/s) and includes an additional factor to reduce 10 and 100 times the strength for all the sizes of the bodies to see how this could change the growth of the embryo. He found the most significant differences in the case where $Q_D^$ is reduced by a factor of 100, changing the growth track of the embryo (shown in Fig. 20 of his paper). Even though we can’t compare directly this results with ours since our models are different, we agree that, the relative importance of pebble accretion for giant planet formation will depend strongly on the rate at which pebbles are generated from plantesimal-planetesimal collisions, depending in particular on the impact parameters. Another model of giant planet formation that includes pebble accretion is the one of @Alibert2018. They developed a model of the formation of Jupiter that could explain the constrains that the cosmochemical evidence gives about the existence of two main reservoirs of small bodies that remain separated during the early Solar System [@Kruijer2017]. They conclude that Jupiter formed in a three step process. First, Jupiter’s core grew by the accretion of pebbles. Second, the pebble accretion stopped and the core continued growing by the accretion of small planetesimals. Third, the core was massive enough for the runaway gas accretion to start. It is important to highlight that our initial conditions correspond to the beginning of the oligarchic growth where almost all the solid material is contained in planetesimals, in contrast to @Alibert2018 where they start their simulations with a first generation of pebbles. Moreover, they don’t include a fragmentation model since they estimate, for the start of their second stage, the amount of fragments that could have been produced during their first stage of pebble accretion. For this estimation, they used a fixed value of $Q_D^$ for 100 km-sized planetesimals. Different initial conditions for giant planet formation including a first generation of pebbles and a variation of the initial size of planetesimals will be studied in future works. Summary and conclusions {#sec4} ======================= We studied the formation of a giant planet considering the collisional evolution of the planetesimal population, including the dependence of the catastrophic impact energy threshold with composition and relative velocities of the colliding planetesimals. We also included the pebble accretion rates for small particles product of the collisional cascade which are pebbles of second generation strongly coupled to the gas to calculate in a more accurate way than in our baseline case. We showed that as the planet grows, the relative velocities among the planetesimals near the planet are increased from velocities of the order of m/s to velocities up to $\sim 4.5$ km/s before the planet reaches the cross-over mass. We analyzed the improvements incorporated in the calculation of $Q_{D}^$ and the inclusion of the pebble accretion rates one at each time, and finally all the improvements together. We point out that, for each improvement in $Q_{D}^$, as the collisional cascade increases, the timescale to reach the cross-over mass is delayed. We also found that in all the cases, the formation of the core is less efficient compared with the case where planetesimal fragmentation is not considered, and with our baseline model. When we include all the improvements at once there is a sum of effects that inhibit the formation of the giant planet core. On one hand, pebble accretion rates are lower than the plantesimal accretion rates for the small particles adopted in the baseline model. This effect contributes to slow down the growth of the giant planet core. On the other hand, when we include the dependence of $Q_{D}^$ with impact velocity and composition we first interpolate between the curves of $Q_{D}^$ for a given velocity obtaining $Q_{D}^$ for each pure material. Then, we make a linear combination between these values where we assume that planetesimals are half basalts and half ices. For low impact velocities basalts are weaker than than ice (except for $\sim$ 100 km-sized bodies) and for higher velocities, ice is weaker than basalts. These effects are averaged since the composition of planetesimals are 50% basalts and 50% ices. Moreover, it is important to remark that the catastrophic disruption threshold for ices at impact velocities of 0.5 km/s and 3 km/s are lower for $\sim$ 100 km-sized plantesimals than $Q_{D}^$ for basalts at any of the three impact velocities (20-30 m/s, 3 km/s and 5 km/s) for that size of bodies. Regardless of whether the planet arrives to the cross-over mass or not within the disk lifetime, in all the cases, planetesimal fragmentation favors the relative rapid growth of the giant planet core up to a few Earth masses compared to the case where plantesimal fragmentation is not considered. This last result motivated us to explore alternatives for the formation of the giant planet with less massive cores. In this way, we followed the work of @Hubickyj2005. These authors showed that if the grain opacity of the planet envelope is reduced up to a 2% of the interstellar medium value, the planet can reach the cross-over mass with a core mass between $5-10~M_{\oplus}$. Thus we computed two new simulations, reducing the grain opacity of the envelope up to a 2% of the interstellar medium value, considering planetesimal fragmentation (including all the improvements in our model) and without planetesimal fragmentation. For this last case, we found that the planet reached the cross-over mass at a time significantly lower (by a factor two) with respect to the case without planetesimal fragmentation but not considering reduced grain opacities. However, the mass of the core at which the cross-over mass is achieved remained practically similar. For the case where planetesimal fragmentation and reduced grain opacities are included, the planet reached the cross-over time at a lower time and with a less massive core compared to the case where reduced grain opacities are included but without planetesimal fragmentation. We associated this result to the fact that if the opacity of the planet envelope is reduced, the planet release more efficiently the luminosity through the envelope, and then, the gas accretion becomes more efficient. Thus the planet accretes significantly more gas for a low-mass core, enhancing the radius of capture of the planet and increasing the solid accretion rate. @Hori-Ikoma-2011 studied the formation of a giant planet where the gas envelope is polluted by icy planetesimals that are dissolved in the envelope. They found that the increase in the molecular weight and the reduction in the adiabatic temperature gradient produced by the pollution of planetesimals, significantly reduce the mass of the core at which the planet achieves the cross-over mass. More recently, @Venturini2015 [@Venturini2016] found similar results computing for the first time, self-consistent models of giant planet formation that include the effect of the envelope enrichment by the pollution of planetesimals dissolved as they income the planet envelope. The increment in the envelope molecular weight and the reduction in adiabatic temperature gradient produce that the envelope layers are compressed more efficiently and the gas accretion is significantly increased. We will incorporate this phenomenon in a future work. We would like to thank John Chambers and an anonymous referee for the comments and suggestions that helped us to improve this paper. This work was supported by the PIP 112-201501-00699CO grant from Consejo Nacional de Investigaciones Científicas y Técnicas, and by PIDT 11/G144 grant from Universidad Nacional de La Plata, Argentina. O.M.G is also supported by the PICT 2016-0023 from ANPCyT, Argentina. O.M.G. also acknowledges the hosting as invited researcher from IA-PUC. Accretion model {#sec5} =============== Evolution of the envelope and planetesimal radial migration {#sec5-1} ----------------------------------------------------------- The evolution of the envelope is calculated solving the standard equations of the stellar evolution theory. The following equations correspond to the conservation of mass, the hydrostatic equilibrium, the energetic balance and the energy transport $$\frac{\partial r}{\partial m_r} = \frac{1}{4 \pi r^2 \rho},$$ $$\frac{\partial P}{\partial m_r} = -\frac{G m_r}{4 \pi r^4},$$ $$\frac{\partial L_r}{\partial m_r} = \epsilon_{pl} - T \frac{\partial S}{\partial t} ,$$ $$\frac{\partial T}{\partial m_r} = - \frac{G m_r T }{4 \pi r^4 P} \nabla ,$$ where $G$ is the universal gravitational constant, $\rho$ is the density of the envelope, $S$ is the entropy per unit mass, $\epsilon_{pl}$ is the energy release rate due to the accretion of planetesimals, and $\nabla \equiv \frac{d ln T}{d ln P}$ is the dimensionless temperature gradient, which depends on whether the energy is carried by radiation or by convection [see @Fortier2009; @Guilera2010 for details]. In our simulations we consider 2500 radial bins logarithmically equally spaced in a protoplanetary disk defined between 0.4 AU and 20 AU. The particles migrate inward in the protoplanetary disk due to gas drag, where the radial migration velocity for each drag regime is given by $$v_{mig} = \begin{cases} - \frac{2 a \eta}{t_{stop}} \left[ \frac{S_t^2}{1+S_t^2}\right] \quad \text{ Epstein regime} \\ \\ -\frac{2 a \eta}{t_{stop}} \left[ \frac{S_t^2}{1+S_t^2}\right] \quad \text{ Stokes regime} \\ \\ -\frac{2 a \eta}{t_{stop}} \;\;\;\;\;\;\;\;\;\; \quad \text{ quadratic regime}, \end{cases} \label{eq1-sec2}$$ being $a$ the semimajor axis, $\eta$ the ratio of the gas velocity to the local Keplerian velocity, $S_t= t_{stop} \omega_k$ the Stokes number, with $t_{stop}$ the stopping time depending on the drag regime (Epstein, Stokes or quadratic regime) and $\omega_k$ the Keplerian frequency. The rest of the model is explained in detail in @Guilera2010 [@Guilera2014]. Solid accretion rates {#sec5-2} --------------------- For planetesimals, we use the accretion rates given by [@Inaba2001] $$\begin{aligned} \frac{dM_C}{dt}= R_H^2 \Sigma_P \Omega_{P} P_{\text{coll}}, \;\;\; \text{when} \;\;\; S_{t} \geq 1, \label{eq1-sec2-1}\end{aligned}$$ where $M_{C}$ is the core’s mass, $R_H$ is the planet’s Hill radius, $\Sigma_P$ is the surface density of solids at the location of the planet, $\Omega_{P}$ is the Keplerian frequency at the planet location, and $P_{\text{coll}}$ is the collision probability, which is a function of the core radius $R_C$, the Hill radius of the planet, and the relative velocity between the planetesimals and the planet $v_{\text{rel}}$, thus $P_{\text{coll}}=P_{\text{coll}}(R_C,R_H, v_{\text{rel}})$. In fact, as we also consider the drag force that planetesimals experience on entering the planetary envelope , the collision probability is function of the enhanced radius $\tilde{R}_C$ instead of $R_C$ [@Guilera2014]. For the pebbles of second generation, we use the pebble accretion rates given by [@Lambrechts2014] $$\begin{aligned} \frac{dM_C}{dt} = \begin{cases} 2 \beta R_H^2\Sigma_p \Omega_P, \quad \text{if } ~ 0.1 \le S_t < 1, \\ \\ 2 \beta \left( \frac{S_t}{0.1} \right)^{2/3} R_H^2\Sigma_p \Omega_P, \quad \text{if }~S_t < 0.1. \end{cases} \label{eq2-sec2-1}\end{aligned}$$ In Eq. (\[eq2-sec2-1\]), we introduce the factor $\beta= \text{min}(1, R_H/H_p)$ in order to take into account a reduction in the pebble accretion rates if the scale height of the pebbles, $H_p$, becomes greater than the Hill radius of the planet. This can happen if the vertical turbulent dispersion of small particles become significant. The scale height of the solids at a given distance from the central star is given by [@Youdin2007] $$\begin{aligned} H_p= H_g\sqrt{\frac{\alpha}{\alpha + S_t}}, \label{eq3-sec2-1}\end{aligned}$$ being $H_g$ the scale height of the disk, and $\alpha$ the dimensionless Shakura & Sunyaev viscosity-parameter [@Shakura-Sunyaev1973] we adopt $\alpha = 10^{-3}$. Fragmentation model {#sec6} =================== According to our model, a collision between a target of mass $M_{T}$ and a projectile of mass $M_{P}$ results in a remnant of mass $M_{R}$ given by $$M_{R} = \begin{cases} \left[ -\frac{1}{2} \left( \frac{Q}{Q_{D}^}- 1 \right) + \frac{1}{2} \right] (M_{T} + M_{P}), & \quad \text{if } Q < Q_{D}, \\ \\ \left[ -0.35 \left( \frac{Q}{Q_{D}^}- 1 \right) + \frac{1}{2} \right] (M_{T} + M_{P}), & \quad \text{if } Q > Q_{D}, \end{cases} \label{eq1-sec2-2}$$ where $Q$ is the collisional energy per unit target mass and $Q_{D}^$ is the catastrophic impact energy threshold per unit target mass required to fragment and disperse half of the target mass. Usually, $Q_{D}^$ is a function of the target’s radius. However, it is important to remark here that in our model [following @Morbidelli2009] $Q_{D}^$ has to be calculated with an effective radius $r_{\text{eff}}= 3 (M_{T}+M_{P})/4 \pi \rho$, being $\rho$ the planetesimal density. Besides this, in our model the mass loss in the collision, which we define as $(M_T + M_P - M_R)$, is distributed between the minimum mass bin considered and the mass bin corresponding to the biggest fragment $M_F$ given by $$\begin{aligned} M_F= 8 \times 10^{-3} \left[ \frac{Q}{Q_D^}~e^{-(Q/4Q_D^)^2} \right]~(M_T + M_P). \label{eq2-sec2-2}\end{aligned}$$ As in @Guilera2014, we note again that for some supercatastrophic collisions (which occur when $M_R \ll M_T+M_P$), $M_F > M_R$. For such collisions we assume that $M_F= 0.5 M_R$. Unlike [@Morbidelli2009], the fragments are distributed following a power-law distribution given by [@Kobayashi2011; @OrmelKobayashi2012] $$\frac{dn}{dm} \propto m^{-5/3}, \label{eq3-sec2-2}$$ meaning that most of the mass is distributed in the larger fragments. In [@Guilera2014], we found that massive cores formation is favored when the exponent of the power-law mass distribution is lower than 2, thus in this work we analyze the most favorable scenarios, taking the value of 5/3 for the exponent. To calculate the growth of the planet, we adopt a discrete planetesimal size distribution using 36 size (or mass) bins logarithmic equally spaced between 1 cm and 100 km, where initially all the solid mass is in the non porous planetesimals of 100 km of radius. However, when we calculate the planetesimal fragmentation process, we extrapolate the planetesimal size distribution two orders of magnitude below the minimum size $r_p^{min}$ of the main model to avoid the accumulation of spurious mass in the smaller fragments. Therefore, only the mass ejected from the collision distributed between the mass of the larger fragment and the minimum size considered ($r_p^{min}=1$ cm) is taken into account to calculate later the solid accretions rates, i.e., we assume that the mass distributed below 1 cm is lost. The feeding zone of the embryo extends to four Hill radii at either side of the embryo. Adopting 2500 radial bins along the protoplanetary disk guarantees that there are at least ten radial bins between $R_P - 4 R_H$ and $R_P + 4 R_H$ at the beginning of the simulation. We define the width of the fragmentation zone as twice the feeding zone, i.e., eight times the embryo Hill radius at both sides of the embryo. The excitation of eccentricities and inclinations decay with the distance to the embryo, specially outside the feeding zone. Then, our definition of the fragmentation zone guarantees that collisions are well determined within this zone. The amount of bins increase as the core mass grows, e.g., in the baseline case (see Sec. \[sec3-1\]), when the planet reaches the cross-over mass, the fragmentation zone has $\sim 600$ radial bins. Velocities and probabilities of collision regimes {#sec6-2} ------------------------------------------------- Following [@Greenberg1991] we adopt three different regimes and their transitions, regime A: dominance by random motion; regime B: dominance by Keplerian shear motion; regime C: Keplerian shear dominance in a very thin disk. For the dispersion regime (dominance by random motion), where keplerian behavior is unimportant, we adopt the impact rate given by Eq. (\[eq1-sec-2-2-1\]). The transition between regime A and regime B is given by $$\frac{(a_P+a_{T})}{2} \frac{(e_P+e_T)}{2}= 2.5 R_{H_T}, \label{eq2-sec-2-2-1}$$ where $R_{H_T}$ is the Hill radius of the target, and $a_P,e_P$ and $a_T,e_T$ are the semi-major axis and eccentricities of the projectile and the target, respectively. Then, for larger values of $e$ and $a$ the system is dominated by random motion (regime A) and for smaller values of these orbital parameters the system is in the keplerian shear regime (regime B). If $a_p i_p < R_{G}$, being $R_{G}$ the target’s gravitational diameter and $i_p$ the inclination of the projectile’s orbit, the particles are in regime C, wherein still dominates the keplerian shear motion but the system is two dimensional. For regimes B and C the relative velocity is given by $$v= 0.58 (2 \mu^{1/15}- 1.27)^{1/2} \Delta a, \label{eq3-sec-2-2-1}$$ where $\mu = M_p/M_{\odot}$ and $\Delta a = 2.5 R_{H_T}$. The impact rate of regime B follows $$\begin{aligned} \text{Impact rate}\|_{\text{B}}= \pi R_P^2 \left( 1 + b \frac{V_{esc}^2}{v^2} \right)^{1/2} \dfrac{\sigma \left(2.5 R_{H_T}\right)^2 1.125 \omega}{\dfrac{a_P+a_T}{2} 4 a_{P} i_{P} \mu^{2/5} M_P}, \label{eq4-sec-2-2-1}\end{aligned}$$ where $\sigma$ is the projectiles surface density and $\omega$ is the keplerian frequency using $a=(a_P+a_T)/2$. Finally, the impact rate of regime C is given by $$\text{Impact rate}\|_{\text{C}}= R_p \left( 1 + b \frac{V_{esc}^2}{v^2} \right)^{1/2} \dfrac{\sigma (2.5 R_{H_T})^2 1.125 \omega}{\dfrac{(a_P+a_T)}{2} \mu^{2/5} M_p}. \label{eq5-sec-2-2-1}$$ Catastrophic impact energy threshold {#sec6-3} ------------------------------------ From SHP simulations, [@BenzAsphaug1999] found that $Q_{D}^$ can be expressed by the functional form $$Q_{D}^* = Q_{0} \left(\frac{ R_T}{1~\text{cm}} \right)^a + B \rho \left( \frac{R_T}{1~\text{cm}} \right)^b, \label{eq1-sec2-2-2}$$ being $Q_{0}, B, a$, and $b$ parameters that depend on the properties of the material and on the impact velocity over the target, and $\rho$ the density of the non-porous planetesimals (in this work we adopted $\rho= 1.5~\text{gr}/\text{cm}^3$). They performed simulations for non-porous basalts at 3 km/s and 5 km/s, and for non-porous ices at 0.5 km/s and 3 km/s. Later, [@Benz2000], performed new SPH simulations to calculate $Q_{D}^$ for non-porous basalts at low impact velocities (between 20 m/s and 30 m/s) finding that targets impacted at such low impact velocities are weaker than targets impacted at greater velocities. Despite the fact that [@Benz2000] did not provide a functional form for $Q_{D}^$, we fit the results of their simulations adopting the same functional form proposed by [@BenzAsphaug1999]. The Tab. \[tab1-sec2-2-2\] summarizes the values of the free parameters that determine the values of $Q_{D}^*$ used in this work for different types of materials and different impact velocities. [^1]:	{ "pile_set_name": "ArXiv" }
--- abstract: 'A well-trained model should classify objects with a unanimous score for every category. This requires the high-level semantic features should be as much alike as possible among samples. To achive this, previous works focus on re-designing the loss or proposing new regularization constraints. In this paper, we provide a new perspective. For each category, it is assumed that there are two feature sets: one with reliable information and the other with less reliable source. We argue that the reliable set could guide the feature learning of the less reliable set during training - in spirit of student mimicking teacher’s behavior and thus pushing towards a more compact class centroid in the feature space. Such a scheme also benefits the reliable set since samples become closer within the same category - implying that it is easier for the classifier to identify. We refer to this mutual learning process as feature intertwiner and embed it into object detection. It is well-known that objects of low resolution are more difficult to detect due to the loss of detailed information during network forward pass (e.g., RoI operation). We thus regard objects of high resolution as the reliable set and objects of low resolution as the less reliable set. Specifically, an intertwiner is designed to minimize the distribution divergence between two sets. The choice of generating an effective feature representation for the reliable set is further investigated, where we introduce the optimal transport (OT) theory into the framework. Samples in the less reliable set are better aligned with aid of OT metric. Incorporated with such a plug-and-play intertwiner, we achieve an evident improvement over previous state-of-the-arts.' author: - \| Hongyang Li$^{1}$, Bo Dai$^{1}$, Shaoshuai Shi$^{1}$, Wanli Ouyang$^{2}$ & Xiaogang Wang$^{1}$\ $^{1}$ Multimedia-SenseTime Joint Lab, The Chinese University of Hong Kong\ `[email protected]`, `[email protected]`, `{ssshi,xgwang}@ee.cuhk.edu.hk`\ $^{2}$ SenseTime Computer Vision Research Group, The University of Sydney\ `[email protected]` bibliography: - 'bib\_library/deep\_learning.bib' - 'bib\_library/my\_pub.bib' - 'bib\_library/misc.bib' - 'bib\_library/obj\_det.bib' title: Feature Intertwiner for Object Detection --- Introduction {#sec:introduction} ============ Classifying complex data in the high-dimensional feature space is the core of most machine learning problems, especially with the emergence of deep learning for better feature embedding [@krizhevsky12_alexnet; @he2016_resnet; @li2018_capsule; @li2019_cmt; @guo2018learning] . Previous methods address the feature representation problem by the conventional cross-entropy loss, $l_1$ / $l_2$ loss, or a regularization constraint on the loss term to ensure small intra-class variation and large inter-class distance [@janocha2017_loss; @liu2017_coco_v2; @wen2016_center_loss; @liu2017_spherical_loss]. The goal of these works is to learn more compact representation for each class in the feature space. In this paper, we also aim for such a goal and propose a new perspective to address the problem. Our observation is that samples can be grouped into two sets in the feature space. One set is more reliable, while the other is less reliable. For example, visual samples may be less reliable due to low resolution, occlusion, adverse lighting, noise, blur, etc. The learned features for samples from the reliable set are easier to classify than those from the less reliable one. Our hypothesis is that the reliable set can guide the feature learning of the less reliable set, in the spirit of a teacher supervising the student. We refer to this mutual learning process as a feature intertwiner. In this paper, a plug-and-play module, namely, feature intertwiner, is applied for object detection, which is the task of classifying and localizing objects in the wild. An object of lower resolution will inevitably lose detailed information during the forward pass in the network. Therefore, it is well-known that the detection accuracy drops significantly as resolutions of objects decrease. We can treat samples with high resolution (often corresponds to large objects or region proposals) as the reliable set and samples with low resolution (small instances) as the less reliable set[^1]. Equipped with these two ‘prototypical’ sets, we can apply the feature intertwiner where the reliable set is leveraged to help the feature learning of the less reliable set. ![ (Zoom in for better view) Visualization of the features in object detection using t-SNE [@vanDerMaaten2008_tsne] (a) without and (b) with feature intertwiner on COCO. Each point is a sample mapped onto the low-dim manifold. []{data-label="fig:tsne"}](tsne){width="\textwidth"} Fig. \[fig:tsne\] on the left visualizes the learned detection features before classifier[^2]. Without intertwiner in (a), samples are more scattered and separated from each other. Note there are several samples that are far from its own class and close to the samples in other categories (e.g., class `person` in blue), indicating a potential mistake in classification. With the aid of feature intertwiner in (b), there is barely outlier sample outside each cluster. the features in the lower resolution set approach closer to the features in the higher resolution set - achieving the goal of compact centroids in the feature space. Empirically, these two settings correspond to the baseline and intertwiner experiments (marked in gray) in Table \[tab:ablation:vs\_baseline\]. The overall mAP metric increases from 32.8% to 35.2%, with an evident improvement of 2.6% for small instances and a satisfying increase of 0.8% for large counterparts. This suggests the proposed feature intertwiner could benefit both sets. Two important modifications are incorporated based on the preliminary intertwiner framework. The first is the use of class-dependent historical representative stored in a buffer. Since there might be no large sample for the same category in one mini-batch during training, the record of all previous features of a given category for large instances is recorded by a representative, of which value gets updated dynamically as training evolves. The second is an inclusion of the optimal transport (OT) divergence as a deluxe regularization in the feature intertwiner. OT metric maps the comparison of two distributions on high-dimensional feature space onto a lower dimension space so that it is more sensible to measure the similarity between two distributions. For the feature intertwiner, OT is capable of enforcing the less reliable set to be better aligned with the reliable set. We name the detection system equipped with the feature intertwiner as InterNet. Full code suite is available at [ ``](https://github.com/hli2020/feature_intertwiner). For brevity, we put the descriptions of dividing two sets in the detection task, related work (partial), background knowledge on OT theory and additional experiments in the appendix. Related Work {#sec:related-work} ============ Object detection [@dai2016_rfcn; @lin2017_FPN; @redmon2016_yolo_v2; @li2018_gradient; @lu2018_grid; @shi2018_pointrcnn] is one of the most fundamental computer vision tasks and serves as a precursor step for other high-level problems. It is challenging due to the complexity of features in high-dimensional space [@krizhevsky12_alexnet], the large intra-class variation and inter-class similarity across categories in benchmarks [@imagenet_conf; @coco]. Thanks to the development of deep networks structure [@simonyan2015_vgg; @he2016_resnet] and modern GPU hardware acceleration, this community has witnessed a great bloom in both performance and efficiency. The detection of small objects is addressed in concurrent literature mainly through two manners. The first is by looking at the surrounding context [@li16_attentive_context; @mottaghi14_context] since a larger receptive filed in the surrounding region could well compensate for the information loss on a tiny instance during down-sampling in the network. The second is to adopt a multi-scale strategy [@li2018_zoom_journal; @lin2017_FPN; @liu2015_ssd; @shrivastava2016_top_down_modulation] to handle the scale problem. This is probably the most effective manner to identify objects in various sizes and can be seen in (almost) all detectors. Such a practice is a “sliding-window” version of warping features across different stages in the network, aiming for normalizing the sizes of features for objects of different resolutions. The proposed feature intertwiner is perpendicular to these two solutions. We provide a new perspective of addressing the detection of small objects - leveraging the feature guidance from high-resolution reliable samples. Designing loss functions for learning better features. The standard cross-entropy loss does not have the constraint on narrowing down the intra-class variation. Several works thereafter have focused on adding new constraints to the intra-class regularization. Liu et al. [@liu2017_spherical_loss] proposed the angular softmax loss to learn angularly discriminative features. The new loss is expected to have smaller maximal intra-class distance than minimal inter-class distance. The center loss [@wen2016_center_loss] approach specifically learns a centroid for each class and penalizes the distances between samples within the category and the center. Our feature intertwiner shares some spirit with this work in that, the proposed buffer is also in charge of collecting feature representatives for each class. A simple modification [@liu2017_coco_v2] to the inner product between the normalized feature input and the class centroid for the softmax loss also decreases the inner-class variation and improves the classification accuracy. Our work is from a new perspective in using the reliable set for guiding the less reliable set. Feature Intertwiners for Object Detection ========================================= In this paper, we adopt the Faster RCNN pipeline for object detection [@he2016_resnet; @he2017_mask_rcnn; @ross15_fast_rcnn]. In Faster RCNN, the input image is first fed into a backbone network to extract features; a region proposal network [@ren2015_faster_rcnn] is built on top of it to generate potential region proposals, which are several candidate rectangular boxes that might contain objects. These region proposals vary in size. Then the features inside the region are extracted and warped into the same spatial size (by RoI-pooling). Finally, the warped features are used by the subsequent CNN layers for classifying whether an object exists in the region. ![ Feature intertwinter overview. Blue blobs stands for the less reliable set (small objects) and green for the reliable set (large ones). For current level $l$, feature map $P_l$ of the small set is first passed into a RoI-pooling layer. Then it is fed into a make-up layer, which fuels back the information lost during RoI; it is optimized via the intertwiner module (yellow rectangle), with aid of the reliable set (green). ‘OT’ (in red) stands for the optimal transport divergence, which aligns information between levels (for details see Sec. \[sec:optimal-transport-divergence-as-information-alignment\]). $P_{m \| l}$ is the input feature map of the reliable set for the RoI layer; $m$ indicates higher level(s) than $l$. []{data-label="fig:overview"}](pipeline){width=".8\textwidth"} Feature Intertwiner Overview {#sec:turbo-boost-module} ---------------------------- We now explicitly depict how the idea of feature intertwiner could be adapted into the object detection framework. Fig. \[fig:overview\] describes the overall pipeline of the proposed InterNet. A network is divided into several levels based on the spatial size of feature maps. For each level $l$, we split the set of region proposals into two categories: one is the large-region set whose size is larger than the output size of RoI-pooling layer and another the small-region set whose size is smaller. These two sets corresponds to the reliable and less reliable sets, respectively. For details on the generation of these two sets in object detection, refer to Sec. \[sec:proposal-assignment\] in the appendix. Feature map $P_l$ at level $l$ is fed into the RoI layer and then passed onto a make-up layer. This layer is designed to fuel back the lost information during RoI and compensate necessary details for instances of small resolution. The refined high-level semantics after this layer is robust to factors (such as pose, lighting, appearance, etc.) despite sample variations. It consists of one convolutional layer without altering the spatial size. The make-up unit is learned and optimized via the intertwiner unit, with aid of features from the large object set, which is shown in the upstream (green) of Fig. \[fig:overview\]. The feature intertwiner is essentially a data distribution measurement to evaluate divergence between two sets. For the reliable set, the input is directly the outcome of the RoI layer of the large-object feature maps $P_{m\|l}$, which correspond to samples of higher level/resolution. For the less reliable set, the input is the output of the make-up layer. Both inputs are fed into a critic module to extract further representation of these two sets and provide evidence for intertwiner. The critic consists of two convolutions that transfer features to a larger channel size and reduce spatial size to one, leaving out of consideration the spatial information. A simple $l_2$ loss can be used for comparing difference between two sets. The final loss is a combination of the standard detection losses [@ross15_fast_rcnn] and the intertwiner loss across all levels. The detailed network structure of the make-up and critic module in the feature intertwiner is shown in the appendix (Sec. \[sec:network\_feat\_inter\]). There are two problems when applying the aforementioned pipeline into application. The first is that the two sets for the same category often do not occur simultaneously in one mini-batch; the second is how to choose the input source for the reliable set, i.e., feature map $P_{m \| l}$ for the large object set. We address these two points in the following sections. Class Buffer ------------ The goal of the feature intertwiner is to have samples from less reliable set close to the samples within the same category from the reliable set. In one mini-batch, however, it often happens that samples from the less reliable set are existent while samples of the same category from the reliable set are non-existent (or vice versa). This makes it difficult to calculate the intertwiner loss between two sets. To address this problem, we use a [buffer]{} $\mathcal{B}$ to store the representative (prototype) for each category. Basically the representative is the mean feature representation from large instances. Let the set of features from the large-region object on all levels be $\mathbf{F}_{\text{critic}}^{(\texttt{large})}$; each sample consisting of the large set $\mathbf{F}$ be $\bm{f}^{(j)}$, where $j$ is the sample index and its feature dimension is $d$. The buffer could be generated as a mapping from sample features to class representative: $$\begin{gathered} \mathcal{B} = [\bm{b}_1, \dots, \bm{b}_i, \dots, \bm{b}_{N_{\text{cls}}}] = \mathcal{M} \bigg[ \mathbf{F}_{\text{critic}}^{(\texttt{large}, 1)}, \ldots, \mathbf{F}_{\text{critic}}^{(\texttt{large}, l)}, \ldots, \mathbf{F}_{\text{critic}}^{(\texttt{large}, L)} \bigg], \\ { \bm{b}_{i^} = \mathcal{M}\Big[\mathbf{F}_{\text{critic}}^{(\texttt{large}, l)} \Big] = \frac{1}{Z} \sum_{l, j} \bm{f}_{\text{critic}}^{(\texttt{large}, l, j)},~~~\text{where}~\mathbf{F}_{\text{critic}}^{(\texttt{large}, l)}=\{\bm{f}_{\text{critic}}^{(\texttt{large}, l, j)}\in \mathbb{R}^{d} \}, \label{representative} }\end{gathered}$$ where the total number of classes is denoted as $N_{\text{cls}}$. Each entry $\bm{b}_i$ in the buffer $\mathcal{B}$ is referred to as the representative of class $i$. Every sample, indexed by $j$ in the large object set, contributes to the class representative $i^$ if its label belongs to $i^$. Here we denote $i^$ as the label of sample $j$; and $Z$ in Eqn. (\[representative\]) denotes the total number of instances whose label is $i^$. The representative is deemed as a reliable source of feature representation and could be used to guide the learning of the less reliable set. There are many options to design the mapping $\mathcal{M}$, e.g., the weighted average of all features in the past iterations during training within the class as shown in Eqn. (\[representative\]), feature statistics from only a period of past iterations, etc. We empirically discuss different options in Table \[tab:ablation:buffer design\]. Equipped with the class buffer, we define the intertwiner loss between two sets as: $$\mathcal{L}_{\text{inter}} = \sum_{l, j} \mathcal{D} \big( \bm{f}_{\text{critic}}^{(\texttt{small}, l, j)}, \mathcal{B} \big),$$ where $\mathcal{D} $ is a divergence measurement; $\bm{f}_{\text{critic}}^{(\texttt{small}, l, j)}$ denotes the semantic feature after critic of the $j$-th sample at level $l$ in the less reliable set (small instances). Note that the feature intertwiner is proposed to optimize the feature learning of the less reliable set for each level. During inference, the green flow as shown in Fig. \[fig:overview\] for obtaining the class buffer will be removed. Discussion on the intertwiner.* (a) Through such a mutual learning, features for small-region objects gradually encode the affluent details from large-region counterparts, ensuring that the semantic features within one category should be as much similar as possible despite the visual appearance variation caused by resolution change. The resolution imperfection of small instances inherited from the RoI interpolation is compensated by mimicking a more reliable set. Such a mechanism could be seen as a teacher-student guidance in the self-supervised domain [@chen2017_obj_det_data_dis]. (b) It is observed that if the representative $\bm{b}_i$ is detached in back-propagation process [(i.e., no backward gradient update in buffer)]{}, performance gets better. The buffer is used as the guidance for less reliable samples. As contents in buffer are changing as training evolves, excluding the buffer from network update would favorably stabilize the model to converge. Such a practice shares similar spirit of the replay memory update in deep reinforcement learning. (c) The buffer statistics come from all levels. Note that the concept of “large” and “small” is a relative term: large proposals on current level could be deemed as “small” ones on the next level. However, the level-agnostic buffer would always receive semantic features for (strictly) large instances. This is why there are improvements across all levels (large or small objects) in the experiments. Choosing Best Feature Map for Large Objects using Optimal Transport {#sec:optimal-transport-divergence-as-information-alignment} -------------------------------------------------------------------- How to acquire the input source, denoted as $P^{(\texttt{large}, l)}$, i.e., feature maps of large proposals, to be fed into the RoI layer on current level $l$? The feature maps, denoted by $P_l$ or $P_m$, are the output of ResNet at different stages, corresponding to different resolutions. Altogether we use four stages, i.e., $P_2$ to $P_5$; $P_2$ corresponds to feature maps of the highest resolution and $P_5$ has the lowest resolution. The inputs are crucial since they serve as the guidance targets to be learned by small instances. There are several choices, which is depicted in Fig. \[fig:OT\_design\]. ![image](OT_design){width="\textwidth"} \[fig:OT\_design\] [l \| l \|x[20]{}x[20]{}x[20]{} \| x[20]{}x[20]{}x[20]{}]{} option & variant & AP & AP$_{50}$ & AP$_{75}$ & AP$_S$ & AP$_M$ & AP$_L$\ & 35.1 & 54.9 & 40.7& 20.2 & 38.3 & 48.5\ & 40.5 & 62.8 & 47.6 & 23.7 & 45.2 & 53.1\ & $\mathcal{F}$ bilinear & 40.6 & 62.9 & 47.6 & 23.9 & 45.4 & 53.1\ & $\mathcal{F}$ neural net\* & 41.3 & 63.5 & 48.5 & 24.6& 46.3 & 53.8\ & [+0.8]{} & [+0.7]{} & [+0.9]{} & [+0.9]{} & [+1.1]{} & [+0.7]{}\ & KL, $\mathcal{F}$ neural net & 41.0 & 63.1& 48.2& 24.5 & 45.7 & 53.4\ & $l_2$, $\mathcal{F}$ neural net & 41.8 & 64.2& 48.9& 24.7 & 46.0 & 53.8\ & optimal transport (OT)\\ & 42.5 & 65.1 & 49.4 & 25.4 & 46.6 & 54.3\ & biased version of OT & 42.5 & 65.3 & 48.6& 25.3 & 46.8 & 54.3\ & [+2.0]{} & [+2.3]{} & [+1.8]{} & [+1.7]{} & [+1.4]{} & [+1.2]{}\ \[tab:OT\_design\] Option (a): $P^{(\texttt{large}, l)}=P_l$. The most straightforward manner would be using features on current level as input for large object set. This is inappropriate since $P_l$ is trained in RPN specifically for identifying small objects; adopting it as the source could contain noisy details of small instances. Option (b):[$P^{(\texttt{large}, l)}=P_m$]{}. Here $m$ and $l$ denote the index of stage/level in ResNet and $m>l$. One can utilize the higher level feature map(s), which has the proper resolution for large objects. [ Compared with $P_l$, $P_m$ have lower resolution and higher semantics. For example, consider the large instances assigned to level $l=2$ (how to assign large and small instances is discussed in the appendix Sec. \[sec:proposal-assignment\]), $P_m$ indicates three stages $m=3,4,5$.]{} However, among these large instances, some of them are deemed as [small]{} objects on higher level $m$ - implying that those feature maps $P_m$ might not carry enough information. They would also have to be up-sampled during the RoI operation for updating the buffer on current level $l$. Take Table \[tab:anchor\_assign\] in the appendix for example, among the assigned 98 proposals on level 2, there are 31 (11 on level 3 and 20 on level 4) objects that have insufficient size (smaller than RoI’s output). [Hence it might be inappropriate to directly use the high-level feature map as well.]{} Option (c): [$P^{(\texttt{large}, l)}=P_{m\|l} \triangleq \mathcal{F}(P_m)$]{}. $P_m$ is first up-sampled to match the size at $P_l$ and then is RoI-pooled with outcome denoted as $P_{m\|l}$. The up-sampling operation aims at optimizing a mapping $\mathcal{F}: P_m \mapsto P_{m\|l}$ that can recover the information of large objects on a shallow level. $\mathcal{F}$ could be as simple as a bilinear interpolation or a neural network. These three options are empirically reported in Table \[tab:OT\_design\]. The baseline model in (b) corresponds to the default setting in cases \[tab:ablation:buffer design\], \[tab:ablation:workflow design\] of Table \[tab:ablations\], where the feature intertwiner is adopted already. There is a 0.8% AP boost from option (b) to (c), suggesting that $P_m$ for large objects should be converted back to the feature space of $P_l$. The gain from (a) to (c) is more evident, which verifies that it might not be good to use $P_l$ directly. More analysis is provided in the appendix. Option (c) is a better choice for using the reliable feature set of large-region objects. Furthermore, we build on top of this choice and introduce a better alternative to build the connection between $P_l$ and $P_{m\|l}$, since the intertwiner is designed to guide the feature learning of the less reliable set on the current level. If some constraint is introduced to keep information better aligned between two sets, the modified input source $P_{m\|l}$ for large instance would be more proper for the other set to learn. Option (d): [$P^{(\texttt{large}, l)}=\texttt{OT}(P_{l}, P_{m\|l})$]{}. The spirit of moving one distribution into another distribution optimally in the most effective manner fits well into the optimal transport (OT) domain [@peyre2018_ot_recent_book]. In this work, we incorporate the OT unit between feature map $P_l$ and $P_{m\|l}$, which serve as inputs before the RoI-pooling operation. A discretized version [@genevay2017_sinkhorn_loss; @cuturi2013_regularized_OT] of the OT divergence is employed as an additional regularization to the loss: $$\texttt{OT}(P_l, P_{m\|l}) \triangleq \mathcal{W}_{Q}( \mathds{P}_\psi, \mathds{P}_r ) \xleftarrow[]{\text{discrete}} \min_{P \in \mathds{R}_{+}^{ C_2 \times C_1}} \langle Q, P\rangle, \label{ot_loss_discrete}$$ where the non-positive $P$ serves as a proxy for the coupling and satisfies $P^\mathsf{T} \mathds{1}_{C_2} = \mathds{1}_{C_1}, P\mathds{1}_{C_1} = \mathds{1}_{C_2}$. $ \langle \cdot, \cdot\rangle$ indicates the Frobenius dot-product for two matrices and $\mathds{1}_m \coloneqq (1/m, \dots, 1/m) \in \mathds{R}_{+}^{m}$. Now the problem boils down to computing $P$ given some ground cost $Q$. We adopt the Sinkhorn algorithm [@sinkhorn1964_first] in an iterative manner to compute $\mathcal{W}_{Q}$, which is promised to have a differentiable loss function. The OT divergence is hence referred to as Sinkhorn divergence. Given features maps $P_m$ from higher level, the generator network $\mathcal{F}$ up-samples [them]{} to match the size of $P_l$ and outputs $P_{m\|l}$. The channel dimension of $P_l$ and $P_{m\|l}$ is denoted as $C$. The critic unit $\mathcal{H}$ (not the proposed critic unit in the feature intertwiner) is designed to reduce the spatial dimensionality of input to a lower dimension $k$ while keeping the channel dimension unchanged. The number of samples in each distribution is $C$. The outcome of the critic unit in OT module is denoted as $\bm{p}_l, \bm{p}_{m\|l}$, respectively. We choose cosine distance as the measurement to calculate the distance between manifolds. The output is known as the ground cost $Q_{x,y}$, where $x,y$ indexes the sample in these two distributions. The complete workflow to compute the Sinkhorn divergence is summarized in Alg. \[algorithm\]. Note that each level owns their own OT module $\mathcal{W}_{Q}^{l}(P_l, P_m)=\texttt{OT}(P_{l}, P_{m\|l})$. The total loss for the detector is summarized as: $$\mathcal{L} = \mathcal{L}_{\text{inter}} + \sum_l\mathcal{W}_{Q}^{(l)}(P_l, P_m) + \mathcal{L}_{\text{standard}},$$ where $\mathcal{L}_{\text{standard}}$ is the classification and regression losses defined in most detectors [@ross15_fast_rcnn]. Input: Feature maps on current and higher levels, $P_l, P_m$ The generator network $\mathcal{F}$ and the critic unit in OT module $\mathcal{H}$ Output: Sinkhorn loss $\mathcal{W}_{Q}^{l}(P_l, P_m)=\texttt{OT}(P_{l}, P_{m\|l})$ Upsample via generator $P_{m\|l} = \mathcal{F}(P_m)$ Feed both inputs into critic $\bm{p}_l = \mathcal{H}(P_l), \bm{p}_{m\|l} = \mathcal{H}(P_{m\|l})$ $\forall (x,y)$ in $\bm{p}_{l}, \bm{p}_{m\|l}$, define the ground cost $Q_{x,y} =\texttt{cosine\_dist}(\bm{p}_{l}, \bm{p}_{m\|l})$ Initialize coefficients $b^{(0)} = \mathds{1}_{C}$ Compute Gibbs kernel $K_{x,y}=\exp(-Q_{x,y} / \varepsilon)$ $a^{(l+1)} \coloneqq \frac{ \mathds{1}_{C} }{K^{\mathsf{T}} b^{(l)}}$, $b^{(l+1)} \coloneqq \frac{ \mathds{1}_{C} }{K^{} a^{(l)}}$, Compute the proxy matrix $P^{(L)} = \text{diag}(b^{(L)}) \cdot K \cdot \text{diag}(a^{(L)})$ Compute $\mathcal{W}_{Q}$ based on the dot-product in Eqn. (\[ot\_loss\_discrete\]): $\langle Q, P\rangle$. Why prefer OT to other alternatives. As proved in [@arjovsky2017_wgan], the OT metric converges while other variants (KL or JS divergence) do not in some scenarios. OT provides sensible cost functions when learning distributions supported by low-dim manifolds (in our case, $\bm{p}_l$ and $\bm{p}_{m\|l}$) while other alternatives do not. As verified via experiments in Table \[tab:OT\_design\], such a property could facilitate the training towards a larger gap between positive and false samples. In essence, OT metric maps the comparison of two distributions on high-dimensional feature space onto a lower dimension space. The use of Euclidean distance could improve AP by around 0.5% (see Table \[tab:OT\_design\], (d) $l_2$ case), but does not gain as much as OT does. This is probably due to the complexity of feature representations in high-dimension space, especially learned by deep models. Experimental Results ==================== We evaluate InterNet on the object detection track of the challenging COCO benchmark [@coco]. For training, we follow common practice as in [@ren2015_faster_rcnn; @he2017_mask_rcnn] and use the `trainval35k` split (union of 80k images from `train` and a random 35k subset of images from 40k `val` split) for training. The lesion and sensitivity studies are reported by evaluating on the `minival` split (the remaining 5k images from `val`). For all experiments, we use depth 50 or 101 ResNet [@he2016_resnet] with FPN [@lin2017_FPN] constructed on top. [We base the framework on Mask-RCNN [@he2017_mask_rcnn] without the segmentation branch. ]{} All ablative analysis adopt austere settings: training and test image scale only at 512; no multi-scale and data augmentation (except for horizontal flip). Details on the training and test procedure are provided in the appendix (Sec. \[sec:training\_test\_details\]). Ablation study on Intertwiner Module {#sec:ablative-analysis} ------------------------------------ Baseline comparison. Table \[tab:ablation:vs\_baseline\] lists the comparison of InterNet to baseline, where both methods shares the same setting. On average it improves by 2 points in terms of mAP. The gain for small objects is much more evident. Note that our method also enhances the detection of large objects (by 0.8%), since the last level also participates in the intertwiner update by comparing its similarity feature to the history buffer, which requires features of the same category to be closer to each other. The last level does not contribute to the buffer update though. [Assignment strategy (analysis based on Sec. \[sec:proposal-assignment\]).]{} Table \[tab:ablation:vs\_baseline\] also investigates the effect of different region proposal allocations. ‘by RoI size’ divides proposals whose area is below the RoI threshold in Table \[tab:anchor\_assign\] as small and above as large; ‘more on higher’ indicates the base value in Eqn. (\[anchor\_ass\]) is smaller (=40); the default setting follows [@lin2017_FPN] where the base is set to 224. Preliminary, we think putting more proposals on higher levels (the first two cases) would balance the workload of the intertwiner; since the default setting leans towards too many proposals on level 2. However, there is no gain due to the mis-alignment with RPN training. The distribution of anchor templates in RPN does not alter accordingly, resulting in the inappropriate use of backbone feature maps. \ Intertwinter loss. Upper block in Table \[tab:ablation:loss\_choice\] shows a factor of 1.0 to be merged on the total loss whereas lower block depicts a specific factor that achieves better AP than others. The simple $l_2$ loss achieves slightly better than the KL divergence, where the latter is computed as $L_{\text{inter}} = b \cdot \log(b / \bm{f})$. The $l_1$ option is by around 1 point inferior than these two and yet still verifies the effectiveness of the intertwiner module compared with baseline (34.2 vs 32.8) - implying the generalization ability of our method in different loss options. How does the intertwiner module affect learning? By measuring the divergence between two sets (i.e., small proposals in the batch and large references in the buffer), we have gradients, as the influence, back-propagated from the critic to make-up layer. In the end, the make-up layer is optimized to enforce raw RoI outputs recovering details even after the loss from reduced resolution. The naive design denoted by ‘separate’ achieves 34.0% AP as shown in Table \[tab:ablation:turbo\_feat\_choice\]. To further make the influence of the intertwiner stronger, we linearly combine the features after critic with the original detection feature (with equal weights, aka 0.5; not shown in Fig. \[fig:overview\]) and feed this new combination into the final detection heads. This improves AP by 1 point (denoted as ‘linear’ in Table \[tab:ablation:turbo\_feat\_choice\]). The ‘naive add’ case with equal weights 1 does not work (loss suddenly explodes during training), since the amplitude of features among these two sources vary differently if we simply add them. Does buffer size matter? Table \[tab:ablation:buffer design\] shows that it does not. A natural thought could be having a window size of $K$ and sliding the window to keep the most recent features recorded. In general, larger size improves performance (see case ‘2000’ vs the size of ‘one epoch’ where batch size is 8, 37.3% $\rightarrow$ 38.8%). In these cases, statistics of large object features for one category cannot reflect the whole training set and it keeps alternating as network is updated. Using ‘all history’ data by running averaging not only saves memory but also has the whole picture of the data. Preliminary, we choose a decayed scheme that weighs more to recent features than ones in the long run, hoping that the model would be optimized better as training evolves. However, experiments does not accord with such an assumption: AP is better where features are equally averaged (c.f., 40.5% and 39.2%) in terms of network evolution. Unified or level-based buffer? Unified. Table \[tab:ablation:workflow design\] upper block reports such a perspective. In early experiments, we only have one unified buffer in order to let objects on the last level also involved in the intertwiner. Besides, the visual features of large objects should be irrelevant of scale variation. This achieves a satisfying AP already. We also try applying different buffers on each level[^3]. The performance improvement is slight, although the additional memory cost is minor. Other investigations. As discussed at the end of Sec. \[sec:turbo-boost-module\], detaching buffer transaction from gradient update attracts improvement (40.5% vs 40.1% in Table \[tab:ablation:workflow design\]). Moreover, we tried imposing stronger supervision on the similarity feature of large proposals by branching out a cross-entropy loss, for purpose of diversifying the critic outputs among different categories. However, it does not work and this additional loss seems to dominate the training process. Comparison to State-of-the-arts ------------------------------- Performance. We list a comparison of our InterNet with previous state-of-the-arts in Table \[tab:final\_compare\_complete\] in the appendix. Without multi-scale technique, ours (42.5%) still favorably outperforms other two-stage detectors (e.g., Mask-RCNN, 39.2%) as well as one-stage detector (SSD, 31.2%). Moreover, we showcase in Fig. \[fig:improve\_per\_class\] the per-class improvement between the baseline and the improved model after adopting feature intertwiner in Table \[tab:ablation:vs\_baseline\] (two gray rows). The most improved classes are ‘microwave’, ‘truck’ while the results in ‘couch’, ‘bat’ decrease. Most small-size categories get improved. ![ Improvement per category after embedding the feature intertwiner on COCO dataset. []{data-label="fig:improve_per_class"}](mAPchange){width=".85\textwidth"} [ As for the distinct drop for the ‘couch’ class, we find that for a large couch among samples on COCO, usually there sit a bunch of people, stuff, pets, etc. And yet the annotations in these cases would cover the whole scenario including these “noises”, making the feature representation of the large couch quite inaccurate. The less accurate features would guide the learning of their small counterparts, resulting in a lower AP for this class.]{} Model complexity and timing. The feature intertwiner only increases three light-weight conv. layers at the make-up and critic units. The usage of class buffer could take up a few GPU memory on-the-fly; however, since we adopt an ‘all-history’ strategy, the window size is just 1 instead of a much larger $K$. The additional cost to the overall model parameters is also from the OT module for each level; however, we find using just one conv. layer for the critic $\mathcal{H}$ and two conv. layers with small kernels for generator $\mathcal{F}$ is enough to achieve good result. Training on 8 GPUs with batch size of 8 takes around 3.4 days; this is slower than Mask-RCNN reported in [@he2017_mask_rcnn]. The memory cost on each card is 9.6 GB, compared with baseline 8.3 GB. The inference runs at 325ms per image (input size is 800) on a Titan Pascal X, increasing around 5% time compared to baseline (308 ms). We do [not]{} intentionally optimize the codebase, however. Conclusion and Future Work ========================== In this paper, we propose a feature intertwiner module to leverage the features from a more reliable set to help guide the feature learning of another less reliable set. This is a better solution for generating a more compact centroid representation in the high-dimensional space. It is assumed that the high-level semantic features within the same category should resemble as much as possible among samples with different visual variations. The mutual learning process helps two sets to have closer distance within the cluster in each class. The intertwiner is applied on the object detection task, where a historical buffer is proposed to address the sample missing problem during one mini-batch and the optimal transport (OT) theory is introduced to enforce the similarity among the two sets. Since the features in the reliable set serve as teacher in the feature learning, careful preparation of such features is required so that they would match the information in the small-object set. This is why we design different options for the large set and finally choose OT as a solution. With aid of the feature intertwiner, we improve the detection performance by a large margin compared to previous state-of-the-arts, especially for small instances. [ Feature intertwiner is positioned as a general alternative to feature learning. As long as there exists proper division of one reliable set and the other less reliable set, one can apply the idea of utilizing the reliable set guide the feature learning of another, based on the hypothesis that these two sets share similar distribution in some feature space. One direction in the future work would be applying feature intertwiner into other domains, e.g., data classification, if proper set division are available. ]{} ### Acknowledgments {#acknowledgments .unnumbered} We thank Buyu Li for helpful comments in a preliminary version of this work. H. Li and S. Shi are supported by the Hong Kong PhD Fellowship Scheme. This project is also supported by the Research Grants Council of Hong Kong under grant CUHK14208417, CUHK14202217, and the Hong Kong Innovation and Technology Support Programme Grant ITS/121/15FX. Appendix ======== More related work ----------------- Self-supervised learning. The buffer in the feature intertwiner can be seen as utilizing non-visual domain knowledge on a set of data to help supervise the feature learning for another set in high-dimensional space. Such a spirit falls into the self-supervised learning domain. In [@chen2017_obj_det_data_dis], Chen et al. proposed a knowledge distillation framework to learn compact and accurate object detectors. A teacher model with more capacity is designed to provide strong information and guide the learning of a lite-weight student model. The center loss [@wen2016_center_loss] is formulated to learn a class center and penalize samples that have a larger distance with the centroid. It aims at enlarging inter-class resemblance with cross-entropy (CE) loss as well as narrowing down inner-class divergence for face recognition. In our work, the feature intertwiner gradually aggregates statistics of a meta-subset and utilizes them as targets during the feature learning of a less accurate (yet holding a majority) subset. We are inspired by the proposal-split mechanism in object detection domain to learn recognition at separate scales in the network. [ The self-paced learning framework [@kumar2010_self_paced] deals with two sets as well, where the easy examples are first introduced to optimize the hidden variable and later on during training, the hard examples are involved. There is no interaction between the two sets. The division is based on splitting different samples. In our framework, the two sets mutually help and interact with each other. The goal is towards optimizing a more compact class centroid in the feature space. These are two different branches of work. ]{} Optimal transport (OT) has been applied in two important tasks. One is for transfer learning in the domain adaption problem. Lu et al. [@lu2017ot_transfer_learn] explored prior knowledge in the cost matrix and applied OT loss as a soft penalty for bridging the gap between target and source predictions. Another is for estimating generative models. In [@salimans2018improving], a metric combined with OT in primal form with an energy distance results in a highly discriminative feature representation with unbiased gradients. Genevay et al. [@genevay2017_sinkhorn_loss] presents the first tractable method to train large-scale generative models using an OT-based loss. We are inspired by these works in sense that OT metric is favorably competitive to measure the divergence between two distributions supported on low-dimensional manifolds. Assignment of large and small sets in object detection {#sec:proposal-assignment} ------------------------------------------------------ In this paper we adopt the ResNet model [@he2016_resnet] with feature pyramid dressings [@lin2017_FPN] constructed on top. It generates five levels of feature maps to serve as inputs for the subsequent RPN and detection branches. Denote the level index as $l=\{1, \dots, 5\}$ and the corresponding feature maps as $P_l$. Level $l=1$ is the most shallow stage with more local details for detecting tiny objects and level $l=5$ is the deepest stage with high-level semantics. Let $\mathcal{A}=\{a_j\}$ denote the whole set of proposals generated by RPN from $l_2$ to $l_6$ (level six is generated from $l_5$, for details refer to [@lin2017_FPN]). The region proposals are divided into different levels from $l_2$ to $l_5$: $$a_j^{(l)} \rightarrow l = a_0+\log ( \sqrt{\text{Area}(a_j)} / \texttt{base} ), \label{anchor_ass}$$ where $a_0$=4 as in [@lin2017_FPN]; $\texttt{base}$=224 is the canonical ImageNet pre-training setting. Table \[tab:anchor\_assign\] shows a detailed breakdown[^4] of the proposal allocation based on Eqn. (\[anchor\_ass\]). We can see most proposals from RPN focus on identifying small objects and hence are allocated at shallow level $l=2$. The threshold is set to be the ratio of RoI output’s area over the area of feature map. For example, threshold on $l=3$ is obtained by $(14/64)^2$, where 14 is the RoI output size as default setting. Proposals whose area is below the threshold suffer from the inherent design during RoI operation - these feature outputs are up-sampled by a simple interpolation. The information of small regions is already lost and RoI layer does not help much to recover them back. As is shown on the fourth row (“below \# / above \#”), such a case holds the majority. This observation brings in the necessity of designing a meta-learner to provide guidance on feature learning of small objects due to the loophole during the RoI layer. [c \| x[40]{} x[30]{} x[30]{} x[30]{} ]{} level $l$ & 2 & 3 & 4 & 5\ proposal \# (perc.) & 302 (75%) & 36 (9%)& 54 (14%)& 8 (2%)\ threshold & 0.012 & 0.0479 & 0.1914 & 0.7657\ below \# / above \# & 263 / 39 & 25 / 11 & 34 / 20 & 8 / 0\ intertwiner small \# & 302 & 36 & 54 & 8\ intertwiner large \# & 98 & 62 & 8 & - For level $l$ in the network, we define `small` proposals (or RoIs) to be those already assigned by (\[anchor\_ass\]) and `large` to be those above $l$: $$a^{(l, \texttt{s})} \leftarrow a_j^{(l)},~~~~a^{(l, \texttt{b})} = \bigcup_{m>l} a_j^{(m)}, \label{assign}$$ where the superscript $\texttt{s,b}$ denotes the set of small and large proposals, respectively. The last two rows in Table \[tab:anchor\_assign\] show an example of the assignment. These RoIs are then fed into the RoI-pooling layer[^5] to generate output features maps for the subsequent detection pipeline to process. One may wonder the last level do not have large objects for reference based on Eqn. (\[assign\]). In preliminary experiments, leaving proposals on the last level out of the intertwiner could already improve the overall performance; however, if the last level is also involved (since the buffer is shared across all levels), AP for large objects also improves. See the experiments in Sec. \[sec:ablative-analysis\] for detailed analysis. Sinkhorn divergence {#sec:sinkhorn-divergence} ------------------- Let $u', u$ indicate the individual sample after degenerating high-dimensional features $P_{m\|l}, P_l$ from two spaces into low manifolds. $u', u$ are vectors of dimension $k$. The number of samples in these two distributions is denoted by $C_1$ and $C_2$, respectively. The OT metric between two joint probability distributions supported on two spaces $(\mathcal{U}, \mathcal{U})$ is defined as the solution of the linear program [@cuturi2013_regularized_OT]. Denote the data and reference distribution as $\mathds{P}_\psi, \mathds{P}_r \in \text{Prob}(\mathcal{U})$[^6], respectively, we have the continuous form of OT divergence: $$\mathcal{W}_{Q}( \mathds{P}_\psi, \mathds{P}_r )= \inf_{\gamma \in \Gamma(\mathds{P}_\psi, \mathds{P}_r )} \mathds{E} \bigg[ \int_{ \mathcal{U} \times \mathcal{U} } Q(u', u) d \gamma(u', u) \bigg], \label{ot_loss} \\ $$ where $\gamma$ is a coupling; $\Gamma$ is the set of couplings that consists of joint distributions. Intuitively, $\gamma(u', u)$ implies how much “mass” must be transported from $u'$ to $u$ in order to transform the distribution $\mathds{P}_\psi$ into $\mathds{P}_r$; $Q$ is the “ground cost" to move a unit mass. Eqn. (\[ot\_loss\]) above becomes the p-Wasserstein distance (or loss, divergence) between probability measures when $\mathcal{U}$ is equipped with a distance $\mathcal{D}_{\mathcal{U}}$ and $Q=\mathcal{D}_{\mathcal{U}}(u', u)^p$, for some exponent $p$. The biased version of Sinkhorn divergence used in Table \[tab:OT\_design\] is defined by: $$2\mathcal{W}_{Q}( \mathds{P}_\psi, \mathds{P}_r ) - \mathcal{W}_{Q}( \mathds{P}_r, \mathds{P}_r ) - \mathcal{W}_{Q}( \mathds{P}_\psi, \mathds{P}_\psi ). \nonumber$$ More analysis on Table \[tab:OT\_design\]. All these options have been discussed explicitly at the beginning of Sec. \[sec:optimal-transport-divergence-as-information-alignment\]. Option (a) is inferior due to the inappropriateness of feature maps; (b) serves as the baseline and used as the default setting in Table \[tab:ablations\]. Options in (c) verifies that up-sampling feature maps from higher-level onto current level is preferable; $\mathcal{F}$ being a neural net ensures better improvement. Options in (d) illustrates the case where a supervision signal is imposed onto pair $(P_l, P_{m\|l})$ to make better alignment between them. We can observe that OT outperforms other variants in this setup. Moreover, we tried a biased version [@genevay2017_sinkhorn_loss] of the Sinkhorn divergence. However, it does not bring in much gain compared to the previous setup. Besides, it could burden system efficiency during training (although it is minor considering the total time per iteration). Such a phenomenon could result from an improper update of critic and generator inside the OT module, since the gradient flow would be iterated twice more for the last two terms above. Extending OT divergence to image classification. We also testify OT divergence on CIFAR-10 [@cifar] where feature maps between stages are aligned via OT. Test error decreases by around 1.3%. This suggests the potential application of OT in various vision tasks. Different from OT in generative models, we deem the channel dimension as different samples to compare, instead of batch-wise manner as in [@salimans2018improving]; and treat the optimization of $\mathcal{F}$ and $\mathcal{H}$ in a unified $\min$ problem, as opposed to the adversarial training [@genevay2017_sinkhorn_loss]. Comparison to state-of-the-arts on COCO and PASCAL VOC ------------------------------------------------------ Table \[tab:final\_compare\_complete\] reports the performance of our model compared with other state-of-the-arts on COCO dataset. We can observe that it outperforms all previous one-stage or two-stage detectors by a large margin. The multi-scale technique bundled with data augmentation increases detection accuracy in a more evident manner, which is commonly adopted in most detectors. [ The updated result in Mask-RCNN is reported as well. It increases the original performance from 38.2% to 43.5% by switching the backbone structure to ResNetX, an updated baseline model, ImageNet-5k pre-training and train-time augmentation. It is better than ours (42.5% without multi-scale version). This is probably mainly due to the change of network structure. Our multi-scale version (44.2%) is better than the updated Mask-RCNN result, however.]{} [ r \|c\|x[22]{}x[22]{}x[22]{}\|x[22]{}x[22]{}x[22]{}]{} & backbone & AP & AP$_{50}$ & AP$_{75}$ & AP$_S$ & AP$_M$ & AP$_L$\ \[.1em\] One-stage detector & & & & & & &\ YOLOv2 [@redmon2016_yolo_v2] & DarkNet-19 & 21.6 & 44.0 & 19.2 & 5.0 & 22.4 & 35.5\ SSD513 [@liu2015_ssd] & ResNet-101-SSD & 31.2 & 50.4 & 33.3 & 10.2 & 34.5 & 49.8\ DSSD513 [@fu2017_dssd] & ResNet-101-DSSD & 33.2 & 53.3 & 35.2 & 13.0 & 35.4 & 51.1\ Two-stage detector & & & & & & &\ F-R-CNN+++ [@he2016_resnet] & ResNet-101-C4 & 34.9 & 55.7 & 37.4 & 15.6 & 38.7 & 50.9\ F-R-CNN w FPN [@lin2017_FPN] & ResNet-101-FPN & 36.2 & 59.1 & 39.0 & 18.2 & 39.0 & 48.2\ F-R-CNN by G-RMI [@huang2017_speed_accuracy] & Incept.-ResNet-v2 & 34.7 & 55.5 & 36.7 & 13.5 & 38.1 & 52.0\ F-R-CNN w TDM [@shrivastava2016_top_down_modulation] & Incept.-ResNet-v2-TDM & 36.8 & 57.7 & 39.2 & 16.2 & 39.8 & [52.1]{}\ R-FCN [@dai2016_rfcn] & ResNet-101 & 29.9 & 51.9 & - & 10.8 & 32.8 & 45.0\ Mask RCNN [@he2017_mask_rcnn] & ResNet-101-FPN & 38.2 & 60.3 & 41.7 & 20.1 & 41.1 & 50.2\ [RetinaNet]{} [@lin2017_focal_loss] & ResNet-101-FPN & 39.1 & 59.1 & 42.3 & 21.8 & 42.7 & 50.2\ Mask RCNN, updated in [@he2017_mask_rcnn] & ResNetX-101-FPN & 43.5 & 65.9 & 47.2 & - & - & -\ InterNet (ours) & ResNet-101-FPN & 42.5 & 65.1& 49.4& 25.4 & 46.6 & 54.3\ InterNet (ours) multi-scale & ResNet-101-FPN & 44.2& 67.5& 51.1& 27.2& 50.3& 57.7\ Model Structure Training data mAP -------------------------------- ------------ --------------- ------ Fast R-CNN [@ross15_fast_rcnn] VGG-16 07 66.9 Faster R-CNN [@he2016_resnet] VGG-16 07 69.9 SSD512 [@liu2015_ssd] VGG-16 07 71.6 InterNet (ours) VGG-16 07 73.1 Faster R-CNN [@he2016_resnet] ResNet-101 07+12 76.4 R-FCN [@dai2016_rfcn] ResNet-101 07+12 80.5 InterNet (ours) ResNet-101 07+12 82.7 [To further verify the effectiveness of the feature intertwiner, we further conduct experiments on the PASCAL VOC 2007 dataset. The results are shown in Table \[tab:pascal\_voc\]. Two network structures are adopted. For ResNet-101, the division of the four levels are similar as ResNet-101-FPN on COCO; for VGG-16, we take the division similarly as stated in SSD [@liu2015_ssd]. Specifically, the output of layer ‘conv7’, ‘conv8\_2’, ‘conv9\_2’ and ‘conv10\_2’ are used for $P_2$ to $P_5$, respectively. Our method performs favorably against others in both backbone structures on the PASCAL dataset.]{} Training and test details {#sec:training_test_details} ------------------------- We adopt the stochastic gradient descent as optimizer. Initial learning rate is 0.01 with momentum 0.9 and weight decay 0.0001. Altogether there are 13 epoches for most models where the learning rate is dropped by 90% at epoch 6 and 10. We find the warm-up strategy [@goyal2017_fast_imagenet_train] barely improves the performance and hence do not adopt it. The gradient clip is introduced to prevent training loss to explode in the first few iterations, with maximum gradient norm to be 5. Batch size is set to 8 and the system is running on 8 GPUs. [ The object detector is based on Mask-RCNN (or Faster-RCNN). RoIAlign is adopted for better performance. The model is initialized with the corresponding ResNet model pretrained on ImageNet. The new proposed feature intertwiner module is trained from scratch with standard initialization. The basic backbone structure for extracting features is based on FPN network [@lin2017_FPN], where five ResNet blocks are employed with up-sampling layers. The region proposal network consists of one convolutional layer with one classification and regression layer. The classifier structure is similar as RPN’s - one convolution plus one additional classification/regression head.]{} Non-maximum suppression (NMS) is used during RPN generation and detection test phase. Threshold for RPN is set to 0.7 while the value is 0.3 during test. We do not adopt a dense allocation of anchor templates as in some literature [@liu2015_ssd; @redmon2017_yolo_v1]; each pixel on a level only has the number of anchors the same as the number of aspect ratios (set to 0.5, 1 and 2). Each level $l$ among the five stages owns a unique anchor size: 32, 64, 128, 256, and 512. Network structure in feature intertwiner {#sec:network_feat_inter} ---------------------------------------- [ The detailed network architecture on the make-up layer and critic layer are shown below. ]{} Output size Layers in the make-up module ------------------------------------ ------------------------------------------- $B \times C_l \times 14 \times 14$ conv2d($C_l, C_l, k=3, \text{padding}=1$) $B \times C_l \times 14 \times 14$ batchnorm2d($C_l$) $B \times C_l \times 14 \times 14$ relu($\cdot$) Output size Layers in the critic module ----------------------------------- ------------------------------------------------------------ $B \times 512 \times 7 \times 7$ conv2d($C_l, 512, k=3, \text{padding}=1, \text{stride}=2$) $B \times 512 \times 7 \times 7$ batchnorm2d($512$) $B \times 512 \times 7 \times 7$ relu($\cdot$) $B \times 1024 \times 1 \times 1$ conv2d($512, 1024, k=7$) $B \times 1024 \times 1 \times 1$ batchnorm1d($1024$) $B \times 1024 \times 1 \times 1$ relu($\cdot$) $B \times 1024 \times 1 \times 1$ sigmoid($\cdot$) [^1]: We use the term ‘large object/(more) reliable/high resolution set’ interchangeably in the following to refer to the same meaning; likewise for the term ‘small set/less reliable set/low-resolution set’. [^2]: Only top ten categories with the most number of instances in prediction is visualized. For each category, the high-resolution objects (reliable set) are shown in `solid` color while the low-resolution instances (less reliable set) are shown in `transparent` color with dashed boundary. [^3]: In such case, the last level adopts the buffer on level 2 since it contains the most number of large objects. [^4]: Each sample has 200 proposals with input size being 512. Batch size is 2, resulting in 400 proposals in total. Statistics are averaged per iteration, based on the output of RPN network during training. [^5]: In this paper, we opt for the RoIAlign [@he2017_mask_rcnn] option in the RoI layer; one can resort to other options nonetheless. We use term RoI layer, RoI-pooling layer, RoI operation, to refer to the same process. [^6]: $\text{Prob}(\mathcal{U})$ is the set of probability distributions over a metric space $\mathcal{U}$.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The Newton map $N_f$ of an entire function $f$ turns the roots of $f$ into attracting fixed points. Let $U$ be the immediate attracting basin for such a fixed point of $N_f$. We study the behavior of $N_f$ in a component $V$ of $\C\sm U$. If $V$ can be surrounded by an invariant curve within $U$ and satisfies the condition that for all $z\in\Cc$, $N_f^{-1}(\{z\})\cap V$ is a finite set, we show that $V$ contains another immediate basin of $N_f$ or a [virtual immediate basin]{} (Definition \[Def\_VirtualBasins\]). address: - 'School of Engineering and Science, International University Bremen, Postfach 750 561, 28725 Bremen, Germany' - 'School of Engineering and Science, International University Bremen, Postfach 750 561, 28725 Bremen, Germany' author: - Johannes Rückert - Dierk Schleicher title: 'On Newton’s Method for Entire Functions' --- Introduction ============ Newton’s method is a classical way to approximate roots of entire functions by an iterative procedure. Trying to understand this method may very well be called the founding problem of holomorphic dynamics [@Milnor p. 51]. Newton’s method for a complex polynomial $p$ is the iteration of a rational function $N_p$ on the Riemann sphere. Such dynamical systems have been extensively studied in recent years. Tan Lei [@Tan] gave a complete classification of Newton maps of cubic polynomials. In 1992, Manning [@Manning] constructed a finite set of starting values for $N_p$ that depends only on the degree of $p$, such that for any appropriately normalized polynomial with degree $d\geq 10$, the set contains at least one point that converges to a root of $p$ under iteration of $N_p$. Hubbard, Schleicher and Sutherland [@HSS] extended this by constructing a small set of starting values that depends only on the degree $d\geq 2$ and trivial normalizations and finds all roots of $p$. If $f$ is a transcendental entire function, the associated Newton map $N_f$ will generally be transcendental meromorphic, except in the special case $f=pe^q$ with polynomials $p$ and $q$ (see Proposition \[Prop\_RationalNewton\]) which was studied by Haruta [@Haruta]. Bergweiler [@Bergweiler2] proved a no-wandering-domains theorem for transcendental Newton maps that satisfy several finiteness assumptions. Mayer and Schleicher [@MS] have shown that immediate basins for Newton maps of entire functions are simply connected and unbounded, extending a result of Przytycki [@Przytycki] in the polynomial case. They have also shown that Newton maps of transcendental functions may exhibit a type of Fatou component that does not appear for Newton maps of polynomials, so called [virtual immediate basins]{} (Definition \[Def\_VirtualBasins\]) in which the dynamics converges to $\infty$. The thesis [@SebastianDiplom] investigates the Newton map of the transcendental function $z\mapsto z e^{e^z}$ and shows that it exhibits virtual immediate basins; see Figure \[Figure\_SebastianDiplom\] for an illustration. While immediate basins of roots are by definition related to zeroes of $f$ (compare Definition \[Def\_ImmediateBasin\]), under mild technical assumptions a virtual immediate basin leads to an [asymptotic]{} zero of $f$; in other words, a virtual immediate basin often contains an asymptotic path of an asymptotic value at $0$ for $f$ [@Buff]. (6,6) In this paper, we continue the work of [@MS] and investigate the behavior of Newton maps in the complement of an immediate basin. Our main result (Theorem \[Thm\_NecessaryCondition\]) is that if a complementary component can be surrounded by an invariant curve through $\infty$, then it contains another immediate basin or virtual immediate basin, unless it maps infinite-to-one onto at least one point of $\Cc$. We believe that the last “unless”-condition is unnecessary, but our methods do not allow us to show this. An immediate corollary for Newton maps of polynomials is that between any two “channels” of any root, there is always another root. This is folklore, but we do not know of a published reference. This result can be viewed as a first step towards a classification of polynomial Newton maps. Our paper is structured in the following way: In Section \[Sec\_Newton\], we give an introduction to some general properties of Newton maps. In Section \[Sec\_Curves\], we investigate homotopy classes of curves to $\infty$ in immediate basins and prove some auxiliary results. In Section \[Sec\_Lefschetz\], we prove a fixed point theorem which we will need and which might be interesting in its own right. In Section \[Sec\_Main\], we state and prove our main result. Newton’s Method as a Dynamical System {#Sec_Newton} ===================================== Immediate Basins ---------------- Let $f:\C\to\C$ be a non-constant entire function and $N_f$ its associated (meromorphic) Newton map $$N_f \;=\; \id\,-\,\frac{f}{f'}\;\;.$$ If $f$ is a polynomial, then $N_f$ extends to a rational map $\Cc\to\Cc$. If $\xi$ is a root of $f$ with multiplicity $m\geq 1$, then it is an attracting fixed point of $N_f$ with multiplier $\frac{m-1}{m}$. Conversely, every fixed point $\xi\in\C$ of $N_f$ is attracting and a root of $f$. \[Def\_ImmediateBasin\] Let $\xi$ be an attracting fixed point of $N_f$. The [basin of $\xi$]{} is $\{z\in\C:\,\lim_{n\rightarrow\infty} N_f^{\circ n}(z) = \xi\}$, the open set of points which converge to $\xi$ under iteration. The connected component $U$ of the basin that contains $\xi$ is called its immediate basin. Immediate basins are $N_f$-invariant because they are Fatou components and contain a fixed point. The following theorem is the main result (Theorem 2.7) of [@MS]. \[Thm\_ImmediateBasinSC\] If $\xi$ is an attracting fixed point of the Newton map $N_f$, then its immediate basin $U$ is simply connected and unbounded. We will use the following notation throughout the paper: If $\gamma$ is a curve, the symbol $\gamma$ denotes the mapping $\gamma:I\to\C$ from an interval into the plane as well as its image $\gamma(I)\subset\C$. By a [tail]{} of an unbounded curve we mean any unbounded connected part of its image. For $r>0$ and $z\in\C$, the symbol $B_r(z)$ designates the disk of radius $r$ centered at $z$. The [full preimage]{} of a point $z\in\Cc$ is the set $N_f^{-1}(\{z\})$. Its only accumulation point can be $\infty$ by the identity theorem. Any point $z'\in N_f^{-1}(\{z\})$ is called a [preimage]{} of $z$. Unless stated otherwise, the boundary and the closure of a set are considered in $\C$. Singular Values --------------- Since the concept of singular values is crucial for the study of dynamical systems, we give a brief reminder of the most important types. In particular, we state some properties of [asymptotic values]{}; these appear only for transcendental maps. Let $h:\C\to\Cc$ be a meromorphic function. We call a point $p\in \C$ a [regular point]{} of $h$ if $p$ has a neighborhood on which $h$ is injective. Otherwise, we call $p$ a [critical point]{}. A point $v\in\Cc$ is called a [regular value]{} if there exists a neighborhood $V$ of $v$ such that for every component $W$ of $h^{-1}(V)$, $h^{-1}\|_V:V\to W$ is a single-valued meromorphic function. Otherwise, $v$ is called a [singular value]{}. The image of a critical point is a singular value and is called a [critical value]{}. Critical points in $\C$ are exactly the zeroes of the first derivative. For a rational map, all singular values are critical values. Let $h:\C\to\Cc$ be a transcendental meromorphic function. A point $a\in\Cc$ is called an asymptotic value of $h$ if there exists a curve $\Gamma:\R_+\to\C$ with $\lim_{t\to\infty}\Gamma(t)=\infty$ such that $\lim_{t\rightarrow\infty}h(\Gamma(t))=a$. We call $\Gamma$ an asymptotic path of $a$. In general, an asymptotic value is defined by having an asymptotic path towards any essential singularity. Note that in our definition, the set of singular values is the closure of the set of critical and asymptotic values. We follow [@BergweilerEremenko] in the classification of asymptotic values. Let $h:\C\to\Cc$ be a meromorphic function and $a\in\C$ be a finite asymptotic value with asymptotic path $\Gamma$. For each $r>0$, let $U_r$ be the unbounded component of $h^{-1}(B_r(a))$ that contains an unbounded end of $\Gamma$. We say that $a$ is a direct singularity (with respect to $\Gamma$) if there is an $r>0$ such that $h(z)\neq a$ for all $z\in U_r$. We call $a$ an indirect singularity if for all $r>0$, there is a $z\in U_r$ such that $h(z)=a$ (then there are infinitely many such $z$ in $U_r$). \[Thm\_DirectSingularities\] [@Heins Theorem 5]. The set of direct singularities of a meromorphic function is always countable. It is possible however that the set of (direct and indirect) singularities is the entire extended plane: Eremenko [@Eremenko] constructed meromorphic functions of prescribed finite order whose set of asymptotic values is all of $\Cc$. \[Lem\_AsymptoticValueOnBoundary\] Let $h:\C\to\Cc$ be a meromorphic function and $B\subset\C$ a bounded topological disk whose boundary is a simple closed curve $\beta$. Suppose that $\beta$ contains no critical values and that $\tilde{B}$ is an unbounded preimage component of $B$. Then $\partial\tilde{B}$ contains an unbounded curve $\tilde{\beta}$ with $h(\tilde{\beta})\subset\beta$ such that either $h\|_{\tilde{\beta}}:\tilde{\beta}\to \beta$ is a universal covering map or $h(\tilde{\beta})$ lands at an asymptotic value on $\beta$. Let $w\in\partial\tilde{B}$. Clearly, $h(w)\in\beta$ and by assumption, $h$ is a local homeomorphism in a neighborhood of $w$. It follows that the closed and unbounded set $\partial\tilde{B}$ is locally an arc everywhere; therefore it cannot accumulate in any compact subset of $\C$ and must contain an arc $\tilde{\beta}$ that converges to $\infty$. The curve $\tilde{\beta}$ contains no critical points. If $h\|_{\tilde{\beta}}:\tilde{\beta}\to \beta$ is not a universal covering map, then it must land at an asymptotic value. Newton Maps ----------- We show that there is only one class of entire functions that have rational Newton maps. This class contains all polynomials. We give a classification of the dynamics within immediate basins for Newton maps of polynomials. First, we investigate under which conditions a meromorphic function is the Newton map of an entire function. The following proposition uses ideas of Matthias Görner and extends a similar result for rational maps (see below) and certain transcendental functions [@Bergweiler2 page 3]. We do not know if the proposition is new; however, we certainly do not know of a published reference. \[Prop\_NewtonMaps\] Let $N:\C\to\Cc$ be a meromorphic function. It is the Newton map of an entire function $f:\C\to\C$ if and only if for each fixed point $N(\xi)=\xi\in\C$, there is a natural number $m\in\N$ such that $N'(\xi)=\frac{m-1}{m}$. In this case, there exists $c\in\C\sm\{0\}$ such that $$f = c\cdot \exp\left({\int\frac{d\zeta}{\zeta-N(\zeta)}}\right)\;\;.$$ Two entire functions $f,g$ have the same Newton maps if and only if $f=c\cdot g$ for a constant $c\in\C\sm\{0\}$. We start with the last claim: $f$ and $cf$ have the same Newton map $\id-f/f'=\id-1/(\ln f)'$. Conversely, if $f$ and $g$ have the same Newton maps, then $(\ln f)'=(\ln g)'$, and the claim follows. It is easy to check that every Newton map satisfies the criterion on derivatives at fixed points. For the other direction, we construct a map $f$ such that $N_f = N$. Let $z_0\in\C$ be any base point and define $\tilde{f}(z)=\int_\gamma \frac{d \zeta}{\zeta-N(\zeta)}$, where $\gamma:[0;1]\to\C$ is any integration path from $z_0$ to $z$ that avoids the fixed points of $N$. This defines $\tilde{f}$ up to $2\pi i k$: if $\gamma'$ is another choice of integration path, the residue theorem shows that $$\frac{1}{2 \pi i}\int_{\gamma'\circ\gamma^{-1}} \frac{d \zeta}{\zeta-N(\zeta)} = \sum_{N(\xi)=\xi} \mbox{Res}_{\xi} \left(\frac{1}{\zeta-N(\zeta)}\right)\;\;,$$ where the sum is taken over the finitely many fixed points of $N$ that are contained in the compact regions bounded by the closed path $\gamma'\circ\gamma^{-1}$. Near a fixed point $\xi$, it is easy to show that $z-N(z)=\frac{1}{m}(z-\xi)+o(z-\xi)$. Hence we get $\mbox{Res}_{\xi}\left(\frac{1}{z-N(z)}\right)=m\in\N$. It follows that the map $f=\exp(\tilde{f})$ is well defined and holomorphic outside the fixed points of $N$. Near such a fixed point $\xi$, $\tilde{f}$ has the form $m\log(z-\xi)+O(1)$. Clearly, the real part of this converges to $-\infty$ for $z\to\xi$, hence setting $f(\xi)=0$ makes $f$ an entire function as desired. An easy calculation then shows that $N_f = N$. A different choice of base point $z_0$ will change $f$ by a multiplicative constant and lead to the same Newton map $N_f$. The following corollary is essentially due to Janet Head ([@head Proposition 2.1.2], [@Tan Lemma 2.2]). \[HeadTheorem\] A rational map $f:\Cc\rightarrow\Cc$ of degree $d\geq 2$ is the Newton map of a polynomial of degree at least two if and only if $f(\infty)=\infty$ and for all other fixed points $a_1,\ldots,a_d\in\C$ there exists a number $m_j\in\N$ such that $f'(a_j)=\frac{m_j-1}{m_j}<1$. Then, $f$ is the Newton map of the polynomial $$p(z)=a\prod_{j=1}^d (z-a_j)^{m_j}$$ for any complex $a\neq 0$. Let $a\in\C\sm\{0\}$. Since $N_p$ and $f$ have the same fixed points with identical multiplicities, the residuals of the maps $\tilde{f}:=(f-\id)^{-1}$ and $\tilde{N}:=(N_p-\id)^{-1}$ at their common simple poles $a_1,\dots,a_d\in\C$ agree, and thus also those at $\infty$. Hence, $\tilde{f}-\tilde{N}$ is a polynomial with $\lim_{z\to\infty}(\tilde{f}-\tilde{N})(z)=0$. Hence $\tilde{f}=\tilde{N}$ and the claim follows. We want to exclude the trivial case of Newton maps with degree one. Let $f:\C\to\C$ be an entire function such that its Newton map $N_f$ has an attracting fixed point $\xi\in\C$ with immediate basin $U=\C$. Then, there exist $d>0$ and $a\in\C$ such that $f(z)=a(z-\xi)^d$. Since $N_f$ has no periodic points of minimal period at least $2$, it cannot be transcendental [@Bergweiler Theorem 2]. Hence $N_f$ is rational and its fixed points can only be $\xi$ and $\infty$, both of which must be simple. It follows that $N_f$ has degree at most one and since it has no poles in $\C$, it is a polynomial. The claim now follows from Proposition \[Prop\_NewtonMaps\]. In the rest of this paper, we will assume that $N_f$ is not a Möbius transformation. Theorem \[Thm\_ImmediateBasinSC\] implies then that for each immediate basin $U$ of $N_f$, there exists a Riemann map $\phi:\disk\to U$ with $\phi(0)=\xi$. The following simple proposition classifies rational Newton maps of entire functions. Its first half is stated without proof in [@Bergweiler]. \[Prop\_RationalNewton\] Let $f:\C\to\C$ be an entire function. Its Newton map $N_f$ is rational if and only if there are polynomials $p$ and $q$ such that $f$ has the form $f=p\,e^q$. In this case, $\infty$ is a repelling or parabolic fixed point. More precisely, let $m,n\geq 0$ be the degrees of $p$ and $q$, respectively. If $n=0$ and $m\geq 2$, then $\infty$ is repelling with multiplier $\frac{m}{m-1}$. If $n=0$ and $m=1$, then $N_f$ is constant. If $n>0$, then $\infty$ is parabolic with multiplier $+1$ and multiplicity $n+1\geq 2$. By [@Milnor Corollary 12.7], every rational function of degree at least $2$ has a repelling or parabolic fixed point. Since $N_f$ is a Newton map, this non-attracting fixed point is unique and must be at $\infty$. In addition to this, there are finitely many attracting fixed points $a_1,\ldots,a_n\in\C$ with associated natural numbers $m_1,\ldots,m_n\in\N$ such that the multipliers satisfy $N_f'(a_i)=\frac{m_i-1}{m_i}$. Let $p(z)=\prod_{i=1}^n (z-a_i)^{m_i}$. Since attracting fixed points of $N_f$ correspond exactly to the roots of $f$, $f$ has the form $f=p\,e^h$ for an entire function $h$. If $h$ was transcendental, so would be $$N_f\,=\,\id -\frac{p\,e^h}{p'e^h+h'p\,e^h}\;=\;\id-\frac{p}{p'+h'p}\;\;,$$ a contradiction. The other direction follows by direct calculation and the rest of the proof is left to the reader. (6,6) For Newton maps of $f=pe^q$, the area of every immediate basin is finite if $\deg q\ge 3$ [@Haruta] and infinite if $p(z)=z$ and $\deg q\in\{0,1\}$ [@Figen]. The dynamics within immediate basins of Newton maps of polynomials has an easy classification, because all singular values are critical values. \[Thm\_PolyNewton\] [[@HSS]]{} Let $p$ be a polynomial of degree $d>1$, normalized so that its roots are contained in the unit disk $\disk$. Let $\xi$ be a root of $p$ and $U$ its immediate basin for $N_p$. Then, $U$ contains $0<k<d$ critical points of $N_p$ and $N_p\|_U$ is a proper self-map of degree $k+1$. Outside the disk $B_2(0)$, $N_f$ is conformally conjugate to multiplication by $\frac{d-1}{d}$. Finally, $U\sm B_2(0)$ has exactly $k$ unbounded components, so called [channels]{}, each of which maps over itself under $N_f$. Figure \[Figure\_PolyNewton\] illustrates this theorem. Accesses in Immediate Basins {#Sec_Curves} ============================ Invariant Accesses ------------------ We investigate the immediate basins of attraction for the attracting fixed points of $N_f$. If $f$ is a polynomial, we have seen in Theorem \[Thm\_PolyNewton\] that immediate basins have an easy geometric structure. In the general case, $N_f$ has an essential singularity at $\infty$ and immediate basins may well have infinitely many accesses to $\infty$. We use prime end theory to distinguish them. Under a finiteness assumption, we have some control over the image of a sequence that converges to $\infty$ through an immediate basin. \[No\_Asymptotic\_Path\] Let $U$ be an immediate basin of the Newton map $N_f$ and $U_R$ an unbounded component of $U\sm B_R(0)$ with the property that no point has infinitely many preimages in $U_R$. Then for any sequence $(z_n)\subset U_R$ with $z_n\to\infty$, all limit points of $N_f(z_n)$ are contained in $\partial U\cup\{\infty\}$. The condition is necessary, because if there exists a point $p\in U$ with infinitely many preimages $p_1,p_2,\ldots \in U_R$, we have $p_n\to\infty$ and $N_f(p_n)=p\in U$ for all $n\in\N$. Assume there exists a sequence $(z_n)\subset U_R$ that converges to $\infty$ with $N_f(z_n)\to p\in U$. Let $B\subset U$ be a closed neighborhood of $p$ inside $U$ such that its boundary $\partial B$ is a simple closed curve $\beta$ that contains no direct singularities (this is possible by Theorem \[Thm\_DirectSingularities\]) nor critical values. Suppose first that $p\not\in N_f(\partial B_R(0))$. Then we may choose $B$ small enough such that $B\cap N_f(\partial B_R(0))=\emptyset$. The image of the first finitely many $z_n$ need not be in $B$; ignoring those, each $z_n$ is contained in a component $W_n$ of $N_f^{-1}(B)\cap U$. If a $W_n$ is bounded, it maps surjectively onto $B$ under $N_f$. Therefore, by the finiteness assumption, there can be only finitely many bounded $W_n$. Each bounded $W_n$ contains finitely many $z_n$; hence there must be an $n$ such that $W_n$ is unbounded. By Lemma \[Lem\_AsymptoticValueOnBoundary\] and again because of the finiteness assumption, $\partial W_n$ contains an asymptotic path of an asymptotic value on $\beta$. But this asymptotic value must be an indirect singularity, which also contradicts the finiteness assumption. If $p\in N_f(\partial B_R(0))$, a small homotopy of the curve $\partial B_R(0)$ in a neighborhood of $p$ solves the problem. Figure \[Figure\_SebastianDiplom\] suggests that immediate basins can reach out to infinity in several different directions. We make this precise in the following definitions that generalize the concept of a channel in the polynomial case. \[DefFixedAccess\] Let $\xi$ be an attracting fixed point of $N_f$ and $U$ its immediate basin. An access to $\infty$ of $U$ is a homotopy class of curves within $U$ that begin at $\xi$, land at $\infty$ and are homotopic with fixed endpoints. An invariant access to $\infty$ is an access with the additional property that for each representative $\gamma$, its image $N_f(\gamma)$ belongs to the access as well. \[LemAccessPrimeEnd\] Let $[\gamma]$ be an access to $\infty$ in $U$. Then $[\gamma]$ induces a prime end $\mathcal{P}$ in $U$ with impression $\{\infty\}$. If $[\gamma]$ is invariant, then $N_f(\mathcal{P})=\mathcal{P}$. Let $\gamma\subset U$ be a curve representing $[\gamma]$ that starts at the fixed point $\xi$ and lands at $\infty$. For $n\in\N$, let $W_n$ be the component of $U\sm B_n(0)$ that contains a tail of $\gamma$. The $W_n$ represent a prime end $\mathcal{P}$ with impression $\infty$. Now a curve $\gamma'\subset U$ that starts at $\xi$ and lands at $\infty$ is homotopic to $\gamma$ if and only if a tail of it is contained in $W_n$ for $n$ large enough. Hence the prime end $\mathcal{P}$ of $[\gamma]$ is well-defined. The last claim follows immediately from the definition. It is clear that different accesses induce different prime ends. We state one more well-known topological fact about the boundary behavior of Riemann maps before using prime ends to characterize invariant accesses. \[Lem\_RiemannConnected\] Let $U\subsetneq \C$ be a simply connected unbounded domain and $\gamma_1,\gamma_2:\R_0^+\to U\cup\{\infty\}$ two non-homotopic curves that land at $\infty$ and are disjoint except for their common base point $z_0=\gamma_1(0)=\gamma_2(0)\in U$. Let $C$ be a component of $\C\sm (\gamma_1\cup\gamma_2)$ and $\phi:\disk\to U$ a Riemann map with $\phi(0)=z_0$. Then $\phi^{-1}(\gamma_1)$ and $\phi^{-1}(\gamma_2)$ land at distinct points $\zeta_1$ and $\zeta_2$ of $\partial\disk$. Furthermore, $\ol{\partial U \cap C}\subset\Cc$ corresponds under $\phi^{-1}$ to a closed interval on $\partial\disk$ that is bounded by $\zeta_1$ and $\zeta_2$. This follows immediately because $\phi$ extends to a homeomorphism from $\diskbar$ to the Carathéodory compactification of $U$, see [@Milnor Theorem 17.12]. If $f$ is a polynomial, it follows from Theorem \[Thm\_PolyNewton\] that every immediate basin contains a curve that lands at $\infty$, is homotopic to its image and induces an invariant access. In the general case, it is a priori not even clear that a curve that lands at $\infty$ [and]{} is homotopic within $U$ to its image induces an invariant access. The following proposition deals with this issue. \[PropInvAccesshasFiniteDegree\] Let $\gamma\subset U\cup\{\infty\}$ be a curve connecting the fixed point $\xi$ to $\infty$ such that $N_f(\gamma)$ is homotopic to $\gamma$ in $U$ with endpoints fixed. Let $W_n$ be a sequence of fundamental neighborhoods representing the prime end $\mathcal{P}$ induced by $[\gamma]$. Then $\gamma$ defines an invariant access to $\infty$ if and only if there is no $z\in\Cc$ that has infinitely many preimages in all $W_n$. Suppose that $\gamma$ defines an invariant access, i.e. if $\gamma'$ is homotopic in $U$ to $\gamma$, then $N_f(\gamma')$ is homotopic to $N_f(\gamma)$. Assume there is a point $z_0\in\Cc$ with the property that $N_f^{-1}(\{z_0\})\cap W_n$ is an infinite set for all $W_n$. Without loss of generality, we may assume that for all $n\in\N$, $W_n\sm W_{n+1}$ contains one preimage of $z_0$. Then we can find a curve $\gamma'$ with a tail contained in each $W_n$ that goes through a preimage of $z_0$ in each $W_n\sm W_{n+1}$. Clearly, $\gamma'$ is homotopic to $\gamma$, while its image does not land at $\infty$ and can therefore not be homotopic to $N_f(\gamma)$ with endpoints fixed, a contradiction. Now suppose that no point has infinitely many preimages in all $W_n$. Since the $W_n$ are nested, no point can have infinitely many preimages in any $W_n$ for $n$ sufficiently large. We uniformize $U$ to the unit disk via a Riemann map $\phi:\disk\to U$ such that $\phi(0)=\xi$ and consider the induced dynamics $g=\phi^{-1}\circ N_f\circ \phi:\disk\to \disk$. By [@Milnor Corollary 17.10], $\phi^{-1}(\gamma)$ and $\phi^{-1}(N_f(\gamma))$ land on $\partial\disk$. Since the curves are homotopic, they even land at the same point $\zeta\in\partial\disk$. Now by assumption, there exists an $\eps>0$ such that within $B_\eps(\zeta)$, no $g$-preimage of any point in $\disk$ accumulates. By Lemma \[No\_Asymptotic\_Path\] it follows that the $g$-image of any sequence converging to $\partial \disk$ inside $B_\eps(\zeta)\cap\disk$ will also converge to $\partial \disk$. Hence we can use the Schwarz Reflection Principle [@Rudin Theorem 11.14] to extend $g$ holomorphically to a neighborhood of $\zeta$ in $\C$. It follows that for the extended map, $\zeta$ is a repelling fixed point with positive real multiplier: if the multiplier was not positive real, $g$ would map points in $B_{\eps}(\zeta)\cap\disk$ out of $\disk$. Also, $\zeta$ cannot be attracting or parabolic, because in this case it would attract points in $\disk$, which all converge to $0$ under iteration. Since $\disk$ is simply connected, all curves in $\disk$ from $0$ to $\zeta$ will be homotopic to each other and their $g$-images. A curve in $\disk$ that starts at $0$ lands at $\zeta$ if and only if its $\phi$-image in $U$ is homotopic to $\gamma$ with endpoints fixed, because $\phi^{-1}(\mathcal{P})$ is a prime end in $\disk$ with impression $\zeta$. We have shown that each invariant access defines a boundary fixed point in the conjugated dynamics on the unit disk, and the dynamics can be extended to a neighborhood of this boundary fixed point, necessarily yielding a repelling fixed point. By [@Milnor Corollary 17.10] it follows that different invariant accesses induce distinct boundary fixed points. If $f$ is a polynomial, there exists a one-to-one correspondence between accesses to $\infty$ of $U$ and boundary fixed points of the induced map $g$ [@HSS Proposition 6]. \[invariantCurve\] Each invariant access has an invariant representative, i.e. a curve $\gamma:\R_0^+\to U$ that lands at $\infty$ with $\gamma(0)=\xi$ and $N_f(\gamma)=\gamma$. For the extension of $g$ to a neighborhood of $\zeta\in\partial\disk$, the multiplier of $\zeta$ is positive real. A short piece of straight line in linearizing coordinates around $\zeta$ maps over itself under $g$. Its forward orbit lands at the fixed point. Since there are uncountably many choices of such invariant curves, we can always find one that contains no critical or direct asymptotic values outside a sufficiently large disk. Virtual Basins -------------- If $f$ is a polynomial and $U\subset\C$ an invariant Fatou component of $N_f$, then $U$ is the immediate basin of a root of $f$, because the Julia set of $N_f$ is connected [@Shishikura], all finite fixed points are attracting and the fixed point at $\infty$ is repelling. If $f$ is transcendental entire, $N_f$ may possess invariant unbounded Fatou domains in which the dynamics converges to $\infty$. Such components are Baker domains or attracting petals of an indifferent fixed point at infinity. In many cases, such components contain an asymptotic path of an asymptotic value at 0 for $f$ [@Buff]. \[Def\_VirtualBasins\] An unbounded domain $V\subset\C$ is called virtual immediate basin of $N_f$ if it is maximal (among domains in $\C$) with respect to the following properties: 1. $\lim\limits_{n\rightarrow\infty} N_f^{\circ n}(z)=\infty$ for all $z\in V$; 2. \[property\_2\] there is a connected and simply connected subdomain $S_0\subset V$ such that $N_f(\ol{S_0})\subset S_0$ and for all $z\in V$ there is an $m\in\N$ such that $N_f^{\circ m}(z)\in S_0$. We call the domain $S_0$ an absorbing set for $V$. Clearly, virtual immediate basins are forward invariant. [@MS Theorem 3.4] Virtual immediate basins are simply connected. It might be possible to extend Shishikura’s theorem [@Shishikura] to show that for Newton maps of entire functions, all Fatou components are simply connected. Taixes has announced partial results in this direction, in particular he rules out the existence of cycles of Herman rings (see also Corollary \[Cor\_Herman\] below). If it were also known that Baker domains are always simply connected, then a result of Cowen [@Cowen Theorem 3.2] would imply that every invariant Fatou component of a Newton map is an immediate basin or a virtual immediate basin (see [@MS Remark 3.5]). A Fixed Point Theorem {#Sec_Lefschetz} ===================== Let $X$ be a compact, connected and triangulable real $n$-manifold and let $f:X\to X$ be continuous with finitely many fixed points. Each fixed point of $f$ has a well-defined [Lefschetz index]{}, and $f$ has a global [Lefschetz number]{}. The classical Lefschetz fixed point formula says that the sum of the Lefschetz indices is equal to the Lefschetz number of $f$, up to a factor of $(-1)^n$ [@Lefschetz; @Brown]. In [@GoldbergMilnor Lemma 3.7], Goldberg and Milnor give a version of this theorem for weakly polynomial-like mappings $f:\diskbar\to\C$. We prove a similar result for a class of maps $f:\Delta\to\Cc$, where $\Delta\subset\Cc$ is a closed topological disk. By extending the range of $f$ to $\Cc$, we allow poles and have to take more boundary components into account than Goldberg and Milnor. \[Def\_LefschetzMap\] Let $\Delta\subset\Cc$ be a closed topological disk with boundary curve $\partial \Delta$ and $f:\Delta\to\Cc$ an orientation preserving open mapping with isolated fixed points. We call $f$ a [Lefschetz map]{} if it satisfies the following conditions: - for every $z\in\Cc$, the full preimage $f^{-1}(\{z\})\subset\Delta$ is a finite set; - $f(\partial\Delta)$ is a simple closed curve so that $f\|_{\partial\Delta}:\partial\Delta\to f(\partial\Delta)$ is a covering map of finite degree; - $ f(\partial\Delta)\cap \mathring{\Delta}=\emptyset $ - if $\xi\in\partial\Delta$ is a fixed point of $f$, then $\xi$ has a neighborhood $U$ such that $f(\partial\Delta\cap U)\subset \partial\Delta$, and $f$ is expanding on $\partial\Delta\cap U$. The definition of “expanding” is with respect to the local parametrization of $\partial \Delta$ near $\xi$ so that $f\|_{\partial\Delta\cap U}$ is topologically conjugate to $x\mapsto 2x$ in a neighborhood of $0$. In this case, the map $f$ can be extended continuously to $U\cup\Delta$ so that $f$ on $U\sm \mathring\Delta$ is topologically conjugate to $z\mapsto 2z$ on the half disk $\{z\in\C\colon \|z\| < 1\mbox{ and } \Im(z)\ge 0\}$ (possibly after shrinking $U$). Such an extension will be called the [simple extension outside of $\Delta$]{} near $\xi$. \[Def\_Lefschetz\] Let $W\subset\C$ be a closed topological disk and $f\colon W\to \C$ be continuous with an isolated fixed point at $\xi\in \mathring W$. With $g(z)=f(z)-z$, we assign to $\xi$ its [Lefschetz index]{} $$\iota(\xi,f) := \lim_{\eps\searrow 0}\; \frac{1}{2\pi i}\oint_{g(\partial B_{\eps}(\xi))} \frac{d\zeta}{\zeta}\;\;.$$ This is the number of full turns that the vector $f(z)-z$ makes when $z$ goes once around $\xi$ in a sufficiently small neighborhood. If $\xi\in\partial W\cap\C$ is an isolated boundary fixed point which has a simple extension outside of $W$, then we define its Lefschetz index as above for this simple extension. For an interior fixed point, it is easy to see that the limit exists and is invariant under homotopies of $f$ that avoid additional fixed points. Strictly speaking, the curve $g(\partial B_\eps(\xi))$ need not be an admissible integration path (i.e. rectifiable), but because of homotopy invariance, we may ignore this problem, and we will often do so in what follows. The Lefschetz index is clearly a local topological invariant; for boundary fixed points, it does not depend on the details of the extension. Therefore, the index is also defined if $\xi=\infty$, using local topological coordinates. Note that for boundary points, the simple extension as defined above generates the least possible Lefschetz index for all extensions of $f$ to a neighborhood of $\xi$. If $f$ is holomorphic in a neighborhood of a fixed point $\xi$, then $\iota(\xi,f)$ is the multiplicity of $\xi$ as a fixed point. \[Def:LefschetzNumber\] Let $f\colon\Delta\to\Cc$ be a Lefschetz map, let $V$ be the component of $\Cc\sm f(\partial\Delta)$ containing $\mathring\Delta$ and let $\gamma_k$ denote the components of $\partial f^{-1}(V)$. The [Lefschetz number]{} $L(f)$ of $f$ is then defined as $$L(f):=\sum_k \left\|\deg\left(f\|_{\gamma_k}\colon \gamma_k\to\gamma\right)\right\| \,\,.$$ In this definition, the orientations of all $\gamma_k$ are irrelevant. The $\gamma_k$ are exactly the components of $f^{-1}(f(\partial\Delta))$, possibly with the exception of $\partial\Delta$ itself: this latter curve is counted only if points in $\mathring\Delta$ near the boundary of $\Delta$ are mapped into $V$ (in the other case, one can imagine $\Cc\sm\ovl V$ as an omitted component of $f^{-1}(V)$, and consequently we also omit its boundary curve). As a result, $L(f)\ge 0$, with equality iff $f^{-1}(V)\cap\Delta=\emptyset$, i.e., the sum over the $\gamma_k$ is empty. The Lefschetz number is invariant under topological conjugacies. We may thus choose coordinates so that $\ovl V\subset\C$. \[Lem\_WindingNumber\] Let $\gamma\subset\C$ be a Jordan curve and $f\colon\gamma\to\C$ continuous so that $f\colon\gamma\to f(\gamma)$ is a covering map and $\gamma\sm f(\gamma)$ is contained in the bounded component of $\C\sm f(\gamma)$. If $\gamma$ contains no fixed points of $f$, then the mapping degree of $f\|_\gamma\colon\gamma\to f(\gamma)$ satisfies $$\deg(f\|_\gamma\colon\gamma\to f(\gamma))= \frac{1}{2\pi i} \oint_{g(\gamma)} \frac{d\zeta}{\zeta}\;\;, \label{Eq:WindingIntegral}$$ where $g(z)=f(z)-z$, and $\gamma$ and $f(\gamma)$ inherit their orientations from $\C$. Let $\Delta$ be the bounded component of $\C\sm\gamma$ and $w_0\in\mathring{\Delta}$ any base point. Since $\Delta$ is contractible to $w_0$ within $\mathring\Delta$, $g\|_{\partial\Delta}=(f-\id)\|_{\partial\Delta}$ is homotopic to $(f-w_0)\|_{\partial\Delta}$ in $\C\sm\{0\}$. The integral (\[Eq:WindingIntegral\]) counts the number of full turns of $f(z)-z$ as $z$ runs around $\gamma=\partial\Delta$. By homotopy invariance, this is equal to the number of full turns $f(\gamma)$ makes around $w_0$, and this equals the desired mapping degree of $f\|_\gamma$. \[Lem:LocalIndex\] Let $V\subset\C$ be a simply connected and bounded domain with piecewise $\mathcal{C}^1$ boundary and let $f\colon \ovl V\to f(\ovl V)\subset\C$ be a continuous map with finitely many fixed points, none of which are on $\partial V$. Then $$\sum_{f(\xi)=\xi} \iota(\xi,f) = \frac{1}{2\pi i}\oint_{g(\partial V)}\frac{d\zeta}{\zeta}\;,$$ where again $g=f-\id$. Break up $V$ into finitely many disjoint simply connected open pieces $V_j$ with piecewise $\mathcal{C}^1$ boundaries so that each $V_j$ either contains a single fixed point of $f$ or $f(\ovl V_j)\cap \ovl V_j=\emptyset$, and each fixed point of $f$ is contained in some $V_j$. This can be done by first choosing disjoint neighborhoods for all fixed points and then partitioning their compact complement in $\ol{V}$ into pieces of diameter less than $\theta$, where $\theta$ is chosen in such a way that $\|f-\id\|>\theta$ in this complement. Set $$c_j:=\frac{1}{2\pi i}\oint_{g(\partial V_j)}\frac{d\zeta}{\zeta} \;.$$ Then $$\frac{1}{2\pi i}\oint_{g(\partial V)}\frac{d\zeta}{\zeta} = \sum_j \frac{1}{2\pi i}\oint_{g(\partial V_j)}\frac{d\zeta}{\zeta} = \sum_j c_j \;.$$ On the pieces with $f(\ovl V_j)\cap \ovl V_j=\emptyset$, we have $c_j=0$, and on a piece $V_j$ with fixed point $\xi_j$, we have $\iota(\xi_j,f)=c_i$ by definition. The claim follows. \[Thm:Lefschetz\] Let $f:\Delta\to\Cc$ be a Lefschetz map with Lefschetz number $L(f)\in\N$. Then $$L(f) = \sum_{f(\xi)=\xi} \iota(\xi,f)\;\;.$$ Let $V$ be the component of $\Cc\sm f(\partial\Delta)$ containing $\mathring\Delta$, and choose coordinates of $\Cc$ such that $\ovl V$ is bounded. Suppose first that $f$ has no fixed points on $\partial\Delta$. Let $\{U_i\}$ be the collection of components of $f^{-1}(\Cc\sm \ovl V)$ and let $\{V_j\}$ be the collection of components of $f^{-1}(V)$. Since $f$ is open, each $U_i$ maps onto $\Cc\sm \ovl V$ and each $V_j$ maps onto $V$ as a proper map. It follows that there are only finitely many $U_i$ and $V_i$, and they satisfy $f(\partial U_i)=f(\partial V_i)=f(\partial\Delta)$ for each $U_i$ and each $V_j$. Note that every fixed point of $f$ must be in some $V_j$. Subdivide the $V_j$ into finitely many simply connected pieces so that no fixed points of $f$ are on the boundaries; call these subdivided domains $V'_{j'}$. The orientation of $\C$ induces a boundary orientation on the $V'_{j'}$. Set again $g:=f-\id$. Then, applying Lemma \[Lem:LocalIndex\] to $V'_{j'}\subset\Delta\subset\ovl{V}\subset\C$ yields $$\sum_{f(\xi)=\xi}\iota(\xi,f) = \sum_{j'} \frac{1}{2\pi i} \oint_{g(\partial V'_{j'})}\frac{d\zeta}{\zeta} = \sum_{j} \frac{1}{2\pi i} \oint_{g(\partial V_j)}\frac{d\zeta}{\zeta} \;.$$ Let $\gamma:=\partial V$ and $\gamma_k \subset\Delta$ be the components of $\partial V_k$ (these are exactly the curves from Definition \[Def:LefschetzNumber\]), with the orientation they inherit from $\partial V_k$. Then $$\sum_{j} \frac{1}{2\pi i} \oint_{g(\partial V_j)}\frac{d\zeta}{\zeta} = \sum_k \frac{1}{2\pi i} \oint_{g(\gamma_k)}\frac{d\zeta}{\zeta} \,\,.$$ The curves $\gamma_k$ come in two different kinds: those which surround the component $V_j$ of which they form part of the boundary, and those which are surrounded by their component $V_j$. The first kind has the same orientation as the orientation it would inherit as a simple closed curve in $\C$, and the second kind has the opposite orientation. On the other hand, it is easy to check that $f\|_{\gamma_k}\colon\gamma_k\to\gamma$ has positive mapping degree (with respect to this orientation of $\gamma_k$, and the standard orientation in $\C$ for $\gamma$) exactly for curves $\gamma_k$ of the first kind. As a result, Lemma \[Lem\_WindingNumber\] implies in both cases $$\frac{1}{2\pi i} \oint_{g(\gamma_k)}\frac{d\zeta}{\zeta} = \left\|\deg\left(f\|_{\gamma_k}\colon \gamma_k\to\gamma\right)\right\| \,\,.$$ This implies the claim if $f$ has no fixed points on $\partial\Delta$. If $f$ has boundary fixed points, we employ a simple extension outside of $\Delta$ in a small neighborhood of each such fixed point. In order for the extended map to be a Lefschetz map, the preimages need to be extended as well. If the extended neighborhoods are sufficiently small, this does not change the Lefschetz number of $f$. As an immediate corollary of this theorem, we observe that Newton maps do not have fixed Herman rings. Note that Taixes has announced a more general result: he uses quasiconformal surgery to rule out any periodic cycles of Herman rings for Newton maps. \[Cor\_Herman\] Newton maps of entire functions have no fixed Herman rings. By [@Shishikura], we may assume that $N:\C\to\Cc$ is a transcendental meromorphic Newton map. Suppose it has a fixed Herman ring, i.e. an invariant Fatou component $H$ such that $N\|_H$ is conjugate to an irrational rotation of an annulus of finite modulus. Then, $H$ contains an invariant and essential simple closed curve $\gamma$. Clearly, ${\rm deg}(N:\gamma\to\gamma)=+1$. Let $\Delta$ be the bounded component of $\C\setminus\gamma$. Then, $N\|_{\ol{\Delta}}$ is a Lefschetz map and by Theorem \[Thm:Lefschetz\], $\ol{\Delta}$ contains a fixed point. This is a contradiction, because all fixed points of $N$ have an unbounded immediate basin (Theorem \[Thm\_ImmediateBasinSC\]). Between Accesses of an Immediate Basin {#Sec_Main} ====================================== In this section, we state and prove our main result. Let $f:\C\to\C$ be an entire function and $N_f$ its Newton map. Let $\xi\in\C$ be a fixed point of $N_f$ and $U$ its immediate basin. Suppose that $U$ has two distinct invariant accesses, represented by $N_f$-invariant curves $\Gamma_1$ and $\Gamma_2$. Consider an unbounded component $\tilde{V}$ of $\C\sm (\Gamma_1\cup\Gamma_2)$. We keep this notation for the entire section. \[Thm\_NecessaryCondition\] If no point in $\Cc$ has infinitely many preimages within $\tilde V$, then the set $V:=\tilde{V}\sm U$ contains an immediate basin or a virtual immediate basin of $N_f$. Note that we do not assume that $V$ is connected. \[Cor\_PolynomialCase\] If $N_f$ is the Newton map of a polynomial $f$, then each component of $\C\sm U$ contains the immediate basin of another root of $f$. If $f$ is a polynomial, $N_f$ is a rational map. It has finite mapping degree and there exists $R>0$ such that all components of $U\sm B_R(0)$ contain exactly one invariant access. Furthermore, all accesses are invariant [@HSS Proposition 6]. Since $\infty$ is a repelling fixed point of $N_f$, there are no virtual immediate basins. The rest of this section will be devoted to the proof of Theorem \[Thm\_NecessaryCondition\]. This proof will be based on the fixed point formula in Theorem \[Thm:Lefschetz\]. In order to be able to use it in our setting, we will need some preliminary statements. \[Pole\_on\_Boundary\] If no point $z\in U$ has infinitely many preimages within $\tilde{V}\cap U$, then $\partial V=\partial U\cap \tilde{V}$ contains at least one pole of $N_f$. In particular, if $\partial V$ is connected or $N_f\|_U$ has finite degree, then $\partial V$ contains a pole of $N_f$. Every pole on $\partial V$ is arcwise accessible from within $U$. Let $\phi:\disk\to U$ be a Riemann map for the immediate basin $U$ with $\phi(0)=\xi$. It conjugates the dynamics of $N_f$ on $U$ to the induced map $g=\phi^{-1}\circ N_f\circ\phi:\disk\to\disk$. By Lemma \[Lem\_RiemannConnected\], the Carathéodory extension $\ol{\phi}^{-1}$ maps $\ol{\partial V}\subset\partial U\cup\{\infty\}$ to a closed interval $I\subset \partial\disk$ that is bounded by the landing points $\zeta_1$ and $\zeta_2$ of $\phi^{-1}(\Gamma_1)$ and $\phi^{-1}(\Gamma_2)$. By assumption, there is an open neighborhood of $\mathring{I}$ in $\disk$ which contains only finitely many $g$-preimages of every $z\in\disk$. By Proposition \[PropInvAccesshasFiniteDegree\], there is a neighborhood $W'$ of $I$ in $\disk$ with the same property. Consider a sequence $(z_n)\subset\disk$ whose accumulation set is in $W'\cap\partial\disk$. By Lemma \[No\_Asymptotic\_Path\], all limit points of $(g(z_n))$ are in $\partial\disk$. Hence there is a neighborhood $W$ of $I$ in $\C$ such that we can extend $g$ by Schwarz reflection to a holomorphic map $\tilde{g}:W\to\C$ that coincides with $g$ on $W\cap\disk$. The endpoints $\zeta_1$ and $\zeta_2$ of $I$ are fixed under this map, because each is the landing point of an invariant curve. They are repelling, because otherwise they would attract points from within $\disk$. Clearly, $\tilde{g}(I)\subset\partial\disk$. If $\tilde{g}(I)=I$, then $\tilde{g}$ has to have an additional fixed point on $\mathring{I}$ which is necessarily parabolic and thus attracts points in $\disk$. This is a contradiction because all points in $\disk$ converge to $0$ under iteration of $g$. If $I$ contained a critical point $c$ of $\tilde{g}$, points in $\disk$ arbitrarily close to $c$ would be mapped out of $\disk$ by $\tilde{g}$, again a contradiction. Hence $\tilde{g}:I\to\partial\disk$ is surjective and there are points $z_1, z_2\in\mathring{I}$ such that $\tilde{g}(z_1)=\zeta_1,\,\tilde{g}(z_2)=\zeta_2$. For $i=1,2$, let $\beta_i:[0;1)\to\disk$ be the radial line from $0$ to $z_i$. Then, $\phi(\beta_i)$ accumulates at a continuum $X_i\subset \partial V$ while $N_f(\phi(\beta_i))=\phi(g(\beta_i))$ lands at $\infty$ in the access of $\Gamma_i$. By continuity, $N_f(X_i)=\{\infty\}$; the identity theorem shows that $X_i=\{p_i\}$ is a pole and $\phi(\beta_i)$ lands at $p_i$. We use the following general lemma to show that $N_f\|_{\tilde{V}}$ can be continuously extended to $\infty$. \[contExtension\] Let $h:\C\to\Cc$ be a meromorphic function and $G\subset \C$ an unbounded domain. Suppose that $\partial G$ can be parametrized by two asymptotic paths of the asymptotic value $\infty$ and that no point has infinitely many preimages within $G$. Then, $h\|_{\ol{G}}$ can be continuously extended to $\infty$. Since $h(\partial G)$ is unbounded, the only possible continuous extension is to set $h(\infty)=\infty$. If $h$ cannot be continuously extended to $\infty$, there exists a sequence $z_n\to\infty$ in $G$ such that $h(z_n)\to p\in\C$. Let $S>\|p\|$ and pick $R>0$ such that $\|h(z)\|\geq S$ for all $z\in\partial G$ with $\|z\| \geq R$, and $p\notin h(\partial B_R(0))$. We may suppose that all $\|z_n\|>R$. Then we can choose a closed neighborhood $B\subset B_S(0)$ of $p$ whose boundary is a simple closed curve that contains no critical values or direct singularities and so that $B$ is disjoint from $h(\partial B_R(0))$. Now let $W_n$ be the component of $h^{-1}(B)$ that contains $z_n$. Then $W_n\subset \C\sm \ovl{B_R(0)}$. Since $z_n\in G$, it follows that all $W_n\subset G\sm \ovl{ B_R(0)}$. If all $W_n$ are bounded, each can contain only finitely many $z_k$ and there must be infinitely many such components. Since bounded $W_n$ map onto $B$, this would contradict the finiteness assumption. Hence there is an unbounded preimage component $W_0$. By Lemma \[Lem\_AsymptoticValueOnBoundary\], $G\sm \ol{B_R(0)}$ then contains an asymptotic path of an indirect singularity on $\partial B$, which also contradicts the finiteness assumption. In the next proposition, we show that $N_f\|_{\tilde{V}}$ is injective near $\infty$. For the proof, we use an extremal length argument in the half-strip $$Y:=[0,\infty)\times [0,1]\;\;,$$ in which we measure the modulus of a quadrilateral by curves connecting the left boundary arc to the right. For $x\in\R$, define $$\H_{x}:=\{z\in\C\,:\,\Re(z)\geq x\}\;\;.$$ First, we prove a technical lemma. \[Lem\_BoundedModulus\] Let $0<t\le s$, let $\beta\subset Y$ an injective curve from $(t,1)$ to $(s,0)$ and let $Q$ the bounded component of $Y\sm\beta$. Let $(0,0)$, $(s,0)$, $(t,1)$ and $(0,1)$ be the vertices of the quadrilateral $Q$. Then, $\mod(Q) \leq t+1$. Let $R\subset Y$ be the rectangle with vertices $(0,0)$, $(0,1)$, $(t+1,1)$, $(t+1,0)$. Its area and modulus are both equal to $t+1$. In particular, $\area(Q\cap R)\leq\area(R)=t+1$. Using the admissible density $\rho(x)=\frac{1}{\sqrt{\area(Q\cap R)}}\cdot\chi_{Q\cap R}(x)$, we get the estimate $$\frac{1}{\mod(Q)}\geq\frac{1}{\area(Q\cap R)} \geq \frac{1}{\area(R)}=\frac{1}{t+1}\;\;,$$ because $\int_\gamma \rho\, d\gamma\geq \frac{1}{\sqrt{\area(Q\cap R)}}$ for this density and all rectifiable curves $\gamma$ that connect the upper to the lower boundaries. \[Prop\_InvarianceNearInfinity\] Suppose that every $z\in\Cc$ has only finitely many $N_f$-preimages in $\tilde V$. Then there exists $R_0>0$ such that for all $R>R_0$, the map $N_f$ is injective on $\tilde{V}\sm B_R(0)$. Moreover, there exists $S>0$ with the property that $$N_f(\tilde{V}\sm B_R(0))\sm B_S(0)=\tilde{V}\sm B_S(0) \,\,.$$ Choose $R_0>\max\{\|z\|:z\in N_f^{-1}(\infty)\cap \ol{\tilde{V}}\}$. It follows from the open mapping principle and invariance of $\partial\tilde{V}$ that there exists $S_0>0$ such that $\partial N_f(\tilde{V}\sm B_{R_0}(0))\sm B_{S_0}(0) \subset\partial\tilde{V}$. Since there are points $z\in\tilde V$ with arbitrarily large $\|z\|$ such that $N_f(z)\in\tilde{V}$, it follows that either $N_f(\tilde{V}\sm B_R(0))\subset \tilde{V}$ or $N_f(\tilde{V}\sm B_R(0))$ contains a punctured neighborhood of $\infty$ within $\Cc$. In the first case, the claims follow easily. By way of contradiction, we may thus assume that we are in the second case. We consider the situation in logarithmic coordinates: with an arbitrary but fixed choice of branch, let $C\subset\H_{\log(R)}$ be the unique unbounded component of $\log(\ol{\tilde{V}}\sm B_R(0))$. This is a closed set whose boundary consists of two analytic curves $\gamma_1$ and $\gamma_2$ and a subset of the vertical line at real part $\log(R)$. Define a holomorphic map $g:C\to\C$ by $g(z)=\log(N_f(e^z))$, choosing the branch such that $\gamma_1\subset g(\gamma_1)$. This is possible because $\Gamma_1=e^{\gamma_1}$ is $N_f$-invariant. Since $e^{\gamma_2}$ is also $N_f$-invariant, there exists $k\in\Z$ such that with $\gamma_4:=g(\gamma_2)$, $\gamma_4=\gamma_2+2\pi i k$. Define also $\gamma_3:=\gamma_1+2\pi i k$. Since $N_f(\tilde{V}\sm B_R(0))$ contains a neighborhood of $\infty$, we get $k\neq 0$. See Figure \[Figure\_Injectivity\] for an illustration of the notations and note that $\gamma_1,\gamma_2,\gamma_3$ and $\gamma_4$ are pairwise disjoint. These four curves have a natural vertical order induced by the observation that each curve separates sufficiently far right half planes into two unbounded components. To fix ideas, suppose that $\gamma_2$ is below $\gamma_1$. Then $\gamma_4$ is below $\gamma_3$. The construction implies that $\gamma_4$ is below $\gamma_1$, and no curve is between $\gamma_1$ and $\gamma_2$. Then, the vertical order is $\gamma_1$, $\gamma_2$, $\gamma_3$, $\gamma_4$. (12,5.3) By Lemma \[contExtension\], $g$ is continuous at $+\infty$. By the open mapping principle, there exists $T>\log(R)$ such that $\partial g(C)\cap \H_T \subset \gamma_1\cup\gamma_4$. Let $C_1$ be the unique unbounded component of $C\cap \ol{\H}_T$, $C_2=C_1+2\pi i k$ and $C_T := g(C)\cap \ol{\H}_T$. Note that $C_1\cup C_2\subset C_T$. We may choose $T$ in such a way that $\partial \H_T$ does not contain any critical values of $g$. Define $C'=g^{-1}(C_T)\subset C$. Then, $g:C'\to C_T$ is a proper map and therefore has well-defined degree. Since $g$ is injective on $\partial C'$ and has no pole, this degree is one and $g$ is univalent. The idea of the proof is as follows: the curves $\gamma_2$ and $\gamma_3$ subdivide $C_T$ into three parts which are unbounded to the right. With an appropriate bound to the right, we obtain a large bounded quadrilateral consisting of three sub-quadrilaterals. Two of these sub-quadrilaterals, the upper and the lower ones, have moduli comparable to the modulus of the entire quadrilateral. This is a contradiction to the Grötzsch inequality if the right boundaries are sufficiently far out. Define a homeomorphism $\psi:Y\to C_1$ that is biholomorphic on the interior and normalized so that it preserves the boundary vertex $\infty$ and the other two boundary vertices. We denote by $\mu_x\subset Y$ the vertical line segment at real part $x$. There exists an $x_0$ such that for $x\geq x_0$, $\psi(\mu_{x})\subset C'$. For $x>x_0$, we denote by $Q_x$ the rectangle in $Y$ that is bounded by $\mu_{x_0}$ and $\mu_x$. With vertices $a=(x,1)$, $b=(x_0,1)$, $c=(x_0,0)$ and $d=(x,0)$, its modulus is equal to $x-x_0$. We denote the vertices of its image $Q'_x:=\psi(Q_x)$ by $a',b',c',d'$, respectively. Let $a''=g(a')$ and $d''=g(d')$. Since $g$ and $\psi$ are univalent, $\mod(g(Q_x'))=x-x_0$. The curve $g(\psi(\mu_x))$ is a boundary curve of $g(Q'_x)$; it connects $a''$ and $d''$ within $C_T$. Let $e_-$ be the intersection point $g(\psi(\mu_x))\cap\gamma_2$ closest to $a''$ along $g(\phi(\mu_x))$, and let $e_+$ be the intersection point furthest to the left along $\gamma_2$. Let $C'_1$ be the bounded subdomain of $C_1$ bounded by $g(\psi(\mu_x))$ between $a''$ and $e_-$, viewed as a quadrilateral with vertices $a''$ and $e_-$ and two more vertices on $\partial\H_T$. Similarly, let $C''''_1$ be the bounded subdomain of $\C$ bounded by $\partial\H_T$, $\gamma_1$, the part of $\gamma_2$ to the left of $e_+$, and the part of $g(\phi(\mu_x))$ between $a''$ and $e_+$, with right vertices $a''$ and $e_+$. Finally, let $C''_1:=C'_1\cup C''''_1$ with right vertices $a''$ and $e_-$, and let $C'''_1:=C''_1$ but with right vertices $a''$ and $e_+$ (instead of $a''$ and $e_-$). If $g(\psi(\mu_x))$ intersects $\gamma_2$ only once, then $e_-=e_+$ and $C'_1=C''_1=C'''_1=C''''_1$. In general, the three domains $C'_1,C''_1,C''''_1$ may be different. However, we have $\mod(C'_1)\ge\mod(C''_1)\ge\mod(C'''_1)\ge\mod(C''''_1)$: the first inequality holds because $C'_1\subset C''_1$, the second describes identical domains but with one boundary vertex moved, and the third follows again from the inclusion $C''''_1\subset C'''_1$, but this time the domain is extended on the “right” side of the domain, rather than on the “lower” side because the boundary vertex has moved. Pulling back under $\psi$, we find that $\Re(\psi^{-1}(a''))\le \Re(a)=x$, because the map $\psi^{-1}\circ g\circ\psi$ repels points away from $\infty$. By Lemma \[Lem\_BoundedModulus\], it follows that $\mod(C''''_1)\le\mod(C'_1)\le x+1$. Similar considerations on the left end of $C_1$, as well as for $C_2$, allow to subdivide $g(Q'_x)$ by a single curve segment of $\gamma_2$ and $\gamma_3$ into three sub-quadrilaterals, two of which have modulus at most $x+1$. But the Grötzsch inequality implies that $$\frac{1}{x-x_0}=\frac{1}{\mod(g(Q_x'))}\geq \frac{1}{x+1}+\frac{1}{x+1}\;,$$ hence $x\le 2x_0+1$ which is a contradiction for large $x$. In order to use Theorem \[Thm:Lefschetz\], we construct an injective curve that surrounds an unbounded domain in $\tilde{V}$ such that the image of the curve does not intersect this domain. Consider a Riemann map $\phi:\disk\to U$ with $\phi(0)=\xi$ and the induced dynamics $g=\phi^{-1}\circ N_f\circ\phi$ on $\disk$. By the Remark after Proposition \[PropInvAccesshasFiniteDegree\], the curves $\phi^{-1}(\Gamma_1)$ and $\phi^{-1}(\Gamma_2)$ land at points $\zeta_1,\zeta_2\in\partial\disk$, and $g$ extends to a neighborhood of $\zeta_1$ and $\zeta_2$ so that $\zeta_1$ and $\zeta_2$ become repelling fixed points. These fixed points have linearizing neighborhoods in which the curves $\phi^{-1}(\Gamma_1)$, respectively $\phi^{-1}(\Gamma_2)$, are straight lines in linearizing coordinates. If $0<r<1$ is large enough, these two curves intersect the circle at radius $r$ only once and we can join them by a circle segment at radius $r$ to an injective curve $\Gamma'\subset\disk$ in such a way that $\Gamma:=\phi(\Gamma')$ separates $V$ from $\xi$. Let $W$ be the closure in $\Cc$ of the connected component of $\C\sm\Gamma$ that contains $V$ (Figure \[Figure\_LastProof\]). Note that no component of $N_f^{-1}(\Gamma)$ that intersects $\mathring{W}$ can leave $W$: in $\disk$, any such component would have to intersect $\Gamma'$. But by the Schwarz Lemma, $g^{-1}(\Gamma')$ has greater absolute value than $r$ everywhere and $\Gamma'$ has only one $g$-preimage within the linearizing neighborhood of $\zeta_1$; this preimage is contained in $\Gamma'$. The same is true at $\zeta_2$. (8,6) By Proposition \[Prop\_InvarianceNearInfinity\], $W$ contains an unbounded preimage component $W'$ of itself such that the boundary $\partial W'$ is contained in $\Gamma_1\cup\Gamma_2$ outside a sufficiently large disk. Make $W'$ simply connected by filling in all bounded complementary components. We claim that $\partial W'$ contains at least one finite pole on $\partial U$: if it did not, then $\partial W'\subset U$ and $N_f\|_{\partial W'} : \partial W' \to \Gamma$ would be injective for all choices of $r$ above. In the limit for $r\to 1$, this would imply that $N_f\|_{\partial V}$ was injective, contradicting Proposition \[Pole\_on\_Boundary\]. Therefore, $\partial W'$ maps onto $\Gamma$ with covering degree at least $2$. If $\infty$ is not an isolated fixed point in $W'$, we are done. Otherwise it is easy to see that $N_f\|_{W'}$ is a Lefschetz map: there is a single boundary fixed point $\infty$; the conditions on this boundary fixed point are satisfied because $\Gamma_1,\Gamma_2\subset U$, where the dynamics is expanding away from $\infty$. Now Theorem \[Thm:Lefschetz\] implies that $W'$ contains fixed points of combined Lefschetz indices at least $2$, because $\partial W'$ contains a pole. If $W'$ contains a finite fixed point, we are done. If not, it follows that the fixed point at $\infty$ has Lefschetz index at least $2$. Consider a Riemann map $\psi: W' \to \H^+$ that uniformizes $W'$ to the upper half plane and maps $\infty$ to $0$; this map preserves the Lefschetz index. By Proposition \[Prop\_InvarianceNearInfinity\], the map $g=\psi\circ N_f\circ\psi^{-1}$ is defined in a relative neighborhood of $0$ in $\H^+$. If a sequence converges to $\R$ in this neighborhood, then so will the image of this sequence. Hence we can extend $g$ to a neighborhood of $0$ in $\C$ by reflection. This extension does not reduce the Lefschetz index of $0$: for a boundary fixed point, the index is defined by extending $g$ to the lower half-plane in the way which generates the least possible fixed point index (compare Definition \[Def\_Lefschetz\]). Reflection however may increase the Lefschetz index. Therefore, $0$ is a parabolic (since multiple) fixed point of the extended map, and it is easily seen that $\partial\H^+$ is in the repelling direction. By the Fatou flower theorem [@Milnor Theorem 10.5], $0$ has an attracting petal in $\H^+$ that induces a virtual immediate basin inside $V$. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank Walter Bergweiler, Xavier Buff, Matthias Görner, Alexandra Kaffl, Sebastian Mayer and Lasse Rempe for their helpful comments in many discussions and their support. We also thank the Institut Henri Poincaré, Université Paris VI, where much of the paper evolved. [99]{} , ‘Iteration of meromorphic functions’, [Bull. Amer. Math. Soc.]{} 29 (1993) 151–188. , ‘Newton’s method and a class of meromorphic functions without wandering domains’, [Ergodic Theory Dynam. Systems]{} 13 (1993) 231–247. , ‘On the singularities of the inverse to a meromorphic function of finite order’, [Rev. Mat. Iberoamericana]{} 11 (1995) 355–373. , ‘Virtual immediate basins of Newton maps and asymptotic values’, to appear in: [Int. Math. Res. Not]{}. ArXiv:math.DS/0601644. , ‘On the Lefschetz fixed point formula’, [Amer. J. Math.]{} 87 (1965) 1–10. , ‘On infinite area for complex exponential function’, [Chaos Solitons Fractals]{} 22 (2004) 1189–1198. , ‘Iteration and the solution of functional equations for functions analytic in the unit disk’, [Trans. Amer. Math. Soc.]{} 265 (1981) 69–95. , ‘The set of asymptotic values of a meromorphic function of finite order’, [Math. Notes]{} 24 5–6 (1979) 914–916. , ‘Fixed points of polynomial maps. Part II. Fixed point portraits’, [Ann. Scient. Éc.Norm. Sup. ]{} 26 (1993) 51–98. , ‘Newton’s method on the complex exponential function’, [Trans. Amer. Math. Soc.]{} 351 (1999) 2499–2513. , ‘The combinatorics of Newton’s method for cubic polynomials’, doctoral thesis, Cornell University, 1987. , ‘Asymptotic spots of entire and meromorphic functions’, [Ann. of Math.]{} 66 (1957) 430–439. , ‘How to find all roots of complex polynomials by Newton’s method’, [Invent. Math.]{} 146 (2001) 1–33. , ‘Intersections and transformations of complexes and manifolds’, [Trans. Amer. Math. Soc.]{} 28 (1926) 1–49. , ‘How to be sure of finding a root of a complex polynomial using Newton’s method’, [Bol. Soc.Brasil. Mat.]{} 22 (1992) 157–177. , ‘Newton’s method for transcendental functions’, Diplomarbeit, TU München, 2002. , ‘Immediate and virtual basins of Newton’s method for entire functions’, [Ann. Inst. Fourier (Grenoble)]{}, to appear. ArXiv:math.DS/0403336. , [Dynamics in one complex variable]{} (Vieweg, Braunschweig/Wiesbaden, 2000). , ‘Remarks on the simple connectedness of basins of sinks for iterations of rational maps’, preprint, Polish Academy of Sciences, Warsaw 1987. , [Real and complex analysis (3rd edition)]{} (McGraw-Hill, New York, 1987). , ‘The connectivity of the Julia set and fixed points’, preprint, IHES, 1990. , ‘Branched coverings and cubic Newton maps’, [Fund. Math.]{} 154 (1997) 207–260.	{ "pile_set_name": "ArXiv" }

Subsets and Splits

No community queries yet

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