text
stringlengths
4
2.78M
meta
dict
--- abstract: 'Despite the widespread use of gradient-based algorithms for optimizing high-dimensional non-convex functions, understanding their ability of finding good minima instead of being trapped in spurious ones remains to a large extent an open problem. Here we focus on gradient flow dynamics for phase retrieval from random measurements. When the ratio of the number of measurements over the input dimension is small the dynamics remains trapped in spurious minima with large basins of attraction. We find analytically that above a critical ratio those critical points become unstable developing a negative direction toward the signal. By numerical experiments we show that in this regime the gradient flow algorithm is not trapped; it drifts away from the spurious critical points along the unstable direction and succeeds in finding the global minimum. Using tools from statistical physics we characterize this phenomenon, which is related to a BBP-type transition in the Hessian of the spurious minima.' author: - Stefano Sarao Mannelli - Giulio Biroli - Chiara Cammarota - Florent Krzakala - | \ Pierfrancesco Urbani - Lenka Zdeborová title: | Complex Dynamics in Simple Neural Networks:\ Understanding Gradient Flow in Phase Retrieval --- Introduction ============ In many machine learning applications one optimizes a non-convex loss function; this is often achieved using simple descending algorithms such as gradient descent or its stochastic variations. The positive results obtained in practice are often hard to justify from the theoretical point of view, and this apparent contradiction between non-convex landscapes and good performance of simple algorithms is a recurrent problem in machine learning. A successful line of research has studied the geometrical properties of the loss landscape, distinguishing between good minima - that lead to good generalization error - and spurious minima - associated with bad generalization error. The results showed that in some regimes, for several problems from matrix completion [@ge2016matrix] to wide neural networks [@kawaguchi2016deep; @soudry2016no], spurious minima disappear and consequently under weak assumptions [@lee2016gradient] gradient descent will converge to good minima. However, these results do not justify numerous other results showing that good and spurious minima are present, but systematically gradient descent works [@safran2017spurious; @LPA19]. In [@sarao2019passed] it was theoretically shown that in a toy model - the spiked matrix-tensor model - it is possible to find good minima with high probability in a regime where exponentially many spurious minima are provably present. In [@mannelli2019afraid] it was shown that this is due to the presence of the so-called threshold states in the landscape, that play a key role in the dynamics of the gradient flow [@cugliandolo1993analytical; @antenucci2018glassy]: at first attracting it, and successively triggering the converge towards lower minima under certain conditions [@watkin1994optimal; @sarao2018marvels]. However, the spiked matrix-tensor model is an unsupervised learning model and it remained open whether the picture put forward in [@sarao2019passed; @mannelli2019afraid] happens also in learning with neural networks. In this work we thus study learning with a simple single-layer neural network on data stemming from the well-known phase retrieval problem - that consists of reconstructing a hidden vector having access only to the absolute value of its projection onto random directions. The problem emerges naturally in a variety of imaging applications where the intensity is easier to access than the phase [@walther1963question; @millane1990phase; @harrison1993phase; @bunk2007diffractive] but it appears also in acoustics [@balan2006signal] and quantum mechanics [@corbett2006pauli]. The phase retrieval problem considered here leads to a high-dimensional and non-convex optimization problem defined as follows. [**Phase retrieval.**]{} Consider $\alpha N$ $N$-dimensional sensing vectors $\pmb x_m$ with unitary norm and generated according to a centered Gaussian distribution, and the true labels $y_m^2$ with $y_m = \langle \pmb x_m, \pmb W^* \rangle$ and $\pmb W^*$ an unknown teacher-vector (the signal in the phase retrieval literature) from $\mathbb{S}^{N-1}(\sqrt{N})$. Given the dataset, the goal is to build an estimator $\pmb W$ of $\pmb W^*$ by minimizing the loss function $$\label{eq:Hamiltonian} \mathcal{L}(\pmb W) = \frac12\sum_{m=1}^{\alpha N} \left(\langle \pmb x_m, \pmb W\rangle^2 - \langle \pmb x_m, \pmb W^*\rangle^2\right)^2 = \frac12\sum_{m=1}^{\alpha N} \ell(\hat{y}_m,y_m),$$ with $\hat{y}_m = \langle \pmb x_m, \pmb W \rangle$ and $\ell(\hat{y},y)=(\hat{y}^2-y^2)^2$ is a modified square loss commonly used in the literature [@candes2015phase; @chen2019gradient] that ensures a smoother landscape compared to the square loss with the absolute values. The loss function, Eq. , is minimized using gradient-descent flow on the sphere starting from random initialization $$\begin{split} & \dot{\pmb W}(t) = -\nabla_{\pmb W}\mathcal{L}(t) + \mu(t) \pmb W(t), \\ & W_i(t=0) \sim\mathcal{N}(0,1)\quad\forall\ i, \end{split}$$ with $\mu(t)$ the Lagrange multiplier that enforces the spherical constraint during the dynamics. The value of $\mu(t)$ can be readily evaluated by taking the scalar product of the gradient of the loss with $\pmb W$ and dividing by $N$ $$\label{eq:mu} \mu(t) = \frac1{2N} \sum_{m=1}^{\alpha N} \frac{\partial}{\partial \hat{y}_m} \ell(\hat{y}_m(t),y_m) \hat{y}_m(t).$$ We can finally define the Hessian of the problem, denoting $\delta_{i,j}$ the Kronecker delta, it reads $$\label{eq:Hessian} \mathcal{H}_{i,j}(\pmb W) = \frac12\sum_{m=1}^{\alpha N} \frac{\partial^2}{\partial \hat{y}_m^2} \ell(\hat{y}_m,y_m) x_{m,i} x_{m,j} - \mu \delta_{i,j}.$$ [**Related work and our main contributions.**]{} Numerous algorithms can be applied to achieve a good reconstruction of the hidden signal (teacher vector) [@schniter2014compressive; @candes2015phase]. For random i.i.d. data and labels generated by a teacher the information-theoretically optimal generalization error has been analyzed in [@barbier2019optimal], showing that for $\alpha < \alpha_{\rm WR} = 0.5$ in the limit of $N\to \infty$ no estimator is able to obtain a generalization error better than a random guess. On the other hand for $\alpha > \alpha_{\rm IT} = 1$ algorithms (not necessarily polynomial ones) exist that are able to achieve zero generalization error. While the weak-recovery threshold is achievable with efficient algorithms [@mondelli2019fundamental], the information-theoretic threshold is conjectured not to be and an algorithmic gap has been conjectured to exist, with perfect recovery achievable with the approximate message passing algorithm only for $\alpha > \alpha_{\rm alg} = 1.13$ [@barbier2019optimal]. In the present paper we are interested in the algorithmic performance of the gradient descent algorithm. The motivation of the present work is not to compare vanilla gradient descent to other algorithms but rather use this well studied phase-retrieval problem to get insights on the properties of the gradient descent algorithm in high-dimensional non-convex optimization. The spurious minima are known to disappear in phase retrieval if the number of samples in the dataset scales as $O(N \log^3 N)$ with $N$ the input dimension [@sun2018geometric]. Later [@chen2019gradient] showed that randomly initialized plain gradient descent will solve the phase retrieval problem with $O(N \, {\rm poly}(\log N))$ samples. An open theoretical question is whether randomly initialized gradient descent is able to solve the phase retrieval problem with $O(N)$ samples. We explore this question in the present work. Our main contributions can be summarized in the following points: - We show empirically that in the phase retrieval problem, randomly initialized gradient flow is attracted by the so-called threshold states. We show that as the number of samples increases, the geometry of the threshold states changes from minima to saddles with the slope pointing toward the teacher-vector. This transition is akin to the BBP phase transition known in random matrix theory [@baik2005phase]. This transition affects gradient descent that, following the slope, achieves zero generalization error. - We obtain a close formula for the number of samples per dimension $\alpha_c$ at which this transition happens. This depends on the joint probability distribution of the labels of the student and the teacher at the threshold. - Using the replica theory from statistical physics as a non-rigorous proxy we characterize the approximate distribution of the labels of the student and the teacher leading to a approximate prediction for threshold $\alpha_c = 13.8$, notably suggesting that the additional logarithmic factors in the previous works might not be needed. BBP on the threshold states =========================== ![Evolution of the training loss for systems of size $N=16384$ with number of samples $\alpha N$. The left panel has labels created using a teacher, while the right panel has random labels constructed using a Gaussian distribution with variance matching the teacher-labels. The left figure shows that also the simulations with a teacher approach the threshold energy before transiting to the global minimum, if $\alpha$ is large enough. []{data-label="fig:energy_planted_vs_unplanted"}](img/{PR_b2.0_energy_th_N16384}.pdf "fig:"){width=".4\linewidth"} ![Evolution of the training loss for systems of size $N=16384$ with number of samples $\alpha N$. The left panel has labels created using a teacher, while the right panel has random labels constructed using a Gaussian distribution with variance matching the teacher-labels. The left figure shows that also the simulations with a teacher approach the threshold energy before transiting to the global minimum, if $\alpha$ is large enough. []{data-label="fig:energy_planted_vs_unplanted"}](img/{PR_b2.0_energy_th_N16384_unplanted}.pdf "fig:"){width=".4\linewidth"} In Fig. \[fig:energy\_planted\_vs\_unplanted\] (left) we show the loss as a function of iteration time for the phase retrieval problem, defined above, with varying number of samples to dimension ratio $\alpha$ in many different runs (full lines). In the right hand side of the figure we show the loss but this time for labels that do not come from a teacher network, but that are randomly reshuffled. We see that in that case the loss converges after a very long time towards a value marked by the dotted line (reproduced also in the left part), that we defined to be the so-called threshold energy. We see that for small $\alpha$, e.g. $\alpha=6$ the train loss on the phase retrieval problem does not decrease to zero (nor the test one), while for the larger values $\alpha=18$ and $24$ it does very rapidly. The value $\alpha=12$ is close to a critical regime where some realization find perfect generalization and other do not, with the dynamics staying for a long time close to the threshold energy. In a different model, the spiked matrix-tensor, Ref. [@mannelli2019afraid] described exactly this phenomenology and showed that gradient flow starting from random initial conditions has a transition when the Hessian of the spurious minima that trap the dynamics, the so-called threshold states, display a BBP transition [@baik2005phase]. This leads to the emergence of a descending direction toward the informative minimum which is correlated with the ground truth signal $\pmb W^*$. In Fig. \[fig:BBP\_example\] we argue that the same mechanism is at play in the phase retrieval problem and based on these insights we derive an analytic equation for the corresponding threshold in Sec. \[sec:theory\]. ![Properties of the Hessian for phase retrieval of a system of size $N=2048$ at $\alpha=10$. On the left figure, we show the evolution of the density of the bulk of the eigenvalues, from zero density in white to high density in blue, the smallest eigenvalue in orange, and the second smallest in green. The picture shows that a BBP transition occurs when the training loss approaches the threshold energy. The right panel depicts the evolution of the training loss (purple) and the generalization loss (dark blue) in time. The training loss rapidly approaches a plateau at the level of the threshold states (black dashed line) and converges towards the teacher after the BBP transition. Namely, when the smallest eigenvalue detaches from the rest of the bulk. The bottom panels show the distribution of the eigenvalues at five different instants in the dynamics. At iteration $\approx 1210$ we cross the threshold energy and we observe an isolated eigenvalue detaching from the bulk. []{data-label="fig:BBP_example"}](img/{hessian_N2048_a10.0_seed1-1}.pdf "fig:") ![Properties of the Hessian for phase retrieval of a system of size $N=2048$ at $\alpha=10$. On the left figure, we show the evolution of the density of the bulk of the eigenvalues, from zero density in white to high density in blue, the smallest eigenvalue in orange, and the second smallest in green. The picture shows that a BBP transition occurs when the training loss approaches the threshold energy. The right panel depicts the evolution of the training loss (purple) and the generalization loss (dark blue) in time. The training loss rapidly approaches a plateau at the level of the threshold states (black dashed line) and converges towards the teacher after the BBP transition. Namely, when the smallest eigenvalue detaches from the rest of the bulk. The bottom panels show the distribution of the eigenvalues at five different instants in the dynamics. At iteration $\approx 1210$ we cross the threshold energy and we observe an isolated eigenvalue detaching from the bulk. []{data-label="fig:BBP_example"}](img/{hessian_N2048_a10.0_seed1-4_energy}.pdf "fig:") ![Properties of the Hessian for phase retrieval of a system of size $N=2048$ at $\alpha=10$. On the left figure, we show the evolution of the density of the bulk of the eigenvalues, from zero density in white to high density in blue, the smallest eigenvalue in orange, and the second smallest in green. The picture shows that a BBP transition occurs when the training loss approaches the threshold energy. The right panel depicts the evolution of the training loss (purple) and the generalization loss (dark blue) in time. The training loss rapidly approaches a plateau at the level of the threshold states (black dashed line) and converges towards the teacher after the BBP transition. Namely, when the smallest eigenvalue detaches from the rest of the bulk. The bottom panels show the distribution of the eigenvalues at five different instants in the dynamics. At iteration $\approx 1210$ we cross the threshold energy and we observe an isolated eigenvalue detaching from the bulk. []{data-label="fig:BBP_example"}](img/{hessian_N2048_a10.0_seed1-1_spectra}.pdf){width=".75\linewidth"} Theory for the BBP threshold {#sec:theory} ---------------------------- Based on the numerical results just presented, we aim to obtain an equation determining the value of threshold $\alpha_c$ such that for $\alpha>\alpha_c$ the BBP transition occurs, whereas for $\alpha < \alpha_c$ it does not. For $\alpha<\alpha_c$ the system is at long times trapped in the threshold states and not able to recover, even weakly, the signal. We define $P(\hat{y},y)$ the long-time limit of the distribution of the estimated labels and the true labels; $P(\hat{y},y)$ allows us to study the Hessian of the threshold states, which is the random matrix defined in (\[eq:Hessian\]) with $\hat y_\mu$ and $y_\mu$ distributed following the law $P(\hat{y},y)$.\ A type of random matrix $\mathcal{M}_{i,j}$ with similar structure as the contribution to the Hessian coming from the Loss function, $\mathcal{H}_{i,j}=-\mathcal{M}_{i,j}-\mu\delta_{i,j}$, has been studied recently in [@lu2017phase] and the convergence in probability for large $N$ of the largest and the second largest eigenvalues of such matrix was proven. We applied the results of the Theorem 1 of [@lu2017phase] to determine the behavior of the smallest and second smallest eigenvalue of the Hessian. Call $$\label{eq:lu} \Psi_\alpha(\lambda) = \lambda \left[\frac1\alpha - \EE_{\hat y,y}\left(\frac{\alpha \partial_{\hat{y}}^2\ell(\hat{y},y)}{2\lambda + \alpha \partial_{\hat{y}}^2\ell(\hat{y},y)}\right)\right], \quad\quad \Phi(\lambda) = - \lambda \EE_{\hat y,y}\left(\frac{\alpha \partial_{\hat{y}}^2\ell(\hat{y},y) y^2}{2\lambda + \alpha \partial_{\hat{y}}^2\ell(\hat{y},y)}\right);$$ let $\bar\lambda = \arg\min\Psi_\alpha(\lambda)$ and $\xi_\alpha(\lambda)=\Psi_\alpha(\max\{\lambda,\bar\lambda\})$ then the largest eigenvalue of $\mathcal{M}_{i,j}$ is $ \lambda_1 \rightarrow_\mathbb{P} \xi_\alpha(\lambda_\alpha^*) $ with $\lambda_\alpha^*$ being the solution of $\xi_\alpha(\lambda) = \Phi(\lambda)$. The second largest is $ \lambda_2 \rightarrow_\mathbb{P} \Psi_\alpha(\bar\lambda_\alpha) $. A BBP transition occurs at $\alpha_{\rm BBP}$, the largest $\alpha$ such that $\lambda_1 = \lambda_2$. Following [@lu2017phase] this leads to an equation on $\bar\lambda$ and $\alpha_{\rm BBP}$ which reads: $$\label{eq:lbar} \frac{1}{\alpha_{\rm BBP}} = \EE_{\hat y,y} \left(\frac{\alpha_{\rm BBP} \partial_{\hat y}^2 \ell(\hat{y},y)}{2\bar\lambda+\alpha_{\rm BBP}\partial_{\hat y}^2 \ell(\hat{y},y)}\right)^2.$$ We now use an additional assumption, which comes from studies of gradient descent dynamics of mean-field spin-glasses [@cugliandolo2002dynamics], that the threshold states are marginal, i.e. the smallest eigenvalue of their Hessian is null. As the smallest eigenvalue of the Hessian is determined by the largest eigenvalue of $\mathcal{M}_{i,j}$, this imposes the additional condition $ \lambda_1=\lambda_2=-\mu$. Using the second Eq. in (\[eq:lu\]) to enforce this last condition and the definition of $\mu$ in Eq. (\[eq:mu\]) one finds (see Appendix Sec. \[si:bbp\] for more details): $$\label{eq:BBP_condition} \mu = \bar \lambda \EE_{\hat y,y}\left(\frac{\alpha_{\rm BBP} \partial_{\hat{y}}^2\ell(\hat{y},y) y^2}{2\bar \lambda + \alpha_{\rm BBP} \partial_{\hat{y}}^2\ell(\hat{y},y)}\right) \ ,\quad\quad \mu=\frac {\alpha_{\rm BBP}}{ 2}\EE_{\hat y,y}\left(\frac{\partial}{\partial \hat{y}} \ell(\hat{y},y) \hat{y}\right) \ .$$ Together the three eqs. (\[eq:lbar\]), and the two in (\[eq:BBP\_condition\]) allow to determine the three unknown $\alpha_{\rm BBP},\bar\lambda,\mu$. Assuming that $\alpha_c=\alpha_{\rm BBP}$, they therefore provide an equation for the algorithmic threshold once $P(\hat{y},y)$ at threshold states is known. ![ The three images show the occurring of a BBP transition at the moment when the training loss crosses the threshold energy. The images investigate the BBP as the input dimension is increased form $N=512$ to $N=8192$. At the BBP transition the smallest eigenvalue pops out of the bulk of eigenvalues and the associated eigenvector contains information on the signal. The left figure is the difference of the smallest eigenvalues (the second with the first in solid line, the third with the second in dotted line). The central image shows the overlap of the first eigenvector (full) and second eigenvector (dotted) with the signal. The transition appears to shift as in the left image. Finally on the right the fluctuations of the overlap first eigenvector with the teacher are shown, the peak corresponds to the transition. []{data-label="fig:BBP_analysis"}](img/analysis_BBP.pdf){width="\linewidth"} Further numerical justifications -------------------------------- We now illustrate and test the theory by numerical simulations. Fig. \[fig:BBP\_example\] shows that the training loss slowly tends to the threshold energy, before departing in direction of the global minimum. According to the theory described in the previous section, the phenomenon occurring is a BBP transition. In order to confirm this we characterize the spectrum of the Hessian and focus on the smallest eigenvalues. On left figure of Fig. \[fig:BBP\_example\] we show the density of eigenvalues in blue and we show separately the smallest eigenvalue in orange and the second smallest eigenvalue in green. During the dynamics the spectrum moves compactly forming a bulk - except for the largest eigenvalues that do not play role in the transition. Approaching the threshold states, for $\alpha > \alpha_c$ the dynamics feels the presence of a descending direction associated with the smallest eigenvalue. This becomes increasingly strong and when the negative eigenvalue pops out of the bulk, the dynamics will follow and will converge to the global minimum. In the lower panel of Fig. \[fig:BBP\_example\] we show $5$ snapshots of the spectral density during the dynamics. In Fig. \[fig:BBP\_analysis\] we study the spectral properties of the Hessian at the time when the training loss hits the threshold energy. In the first panel from the left, we show the difference between the second smallest eigenvalue and the smallest eigenvalue (solid line) and the difference between the third smallest and the second smallest eigenvalue (dotted). The two quantities are close, and very small, until a certain value value of $\alpha$ - that depends on the size and is reported in the caption - after which the former increases linearly whereas the latter remains small. The second property that we investigate, central figure, is the overlap of the first and second eigenvectors with the teacher-vector $\pmb W^*$ (respectively solid and dotted lines). This overlap is zero before the transition, then as $\lambda_2-\lambda_1$, it suddenly increases and finally saturates. The first two panels are provide a strong evidences that a BBP transition is taking place. To further corroborate this findings, in the right panel we consider the fluctuation of the overlap (Fig. \[fig:BBP\_analysis\] right panel). In statistical physics terms, the overlap plays the role of order parameter, and its fluctuations diverge at the BBP transition. We indeed find that at the value of $\alpha$ at which the BBP transition seems to take place fluctuations peak (the more so the larger is $N$ as expected for a phase transition). Finally, we also note the presence of strong finite size effects that shift the effective value of $\alpha$ at which the transition (cross-over for finite $N$) takes place. Characterization of threshold states ==================================== ![On the left panel is the loss of the threshold states from the simulations and from the analysis with the 1RSB replica method, evaluated for different values of $\alpha$. The errorbars given by mean and standard deviation of 1000 simulations with input dimension $N=2048$ and shuffled labels. On the right we compare the moments of the label distribution Eq.  (solid lines) with the moments obtained in the simulations (circles). []{data-label="fig:replica_th_energy"}](img/{PR_b2.0_energy}.pdf "fig:"){width=".4\linewidth"} ![On the left panel is the loss of the threshold states from the simulations and from the analysis with the 1RSB replica method, evaluated for different values of $\alpha$. The errorbars given by mean and standard deviation of 1000 simulations with input dimension $N=2048$ and shuffled labels. On the right we compare the moments of the label distribution Eq.  (solid lines) with the moments obtained in the simulations (circles). []{data-label="fig:replica_th_energy"}](img/comparison_moments.pdf "fig:"){width=".4\linewidth"} The pivotal point of our analysis is that the gradient-flow dynamics is that the spurious minima trapping the dynamics and hindering weak recovery are the threshold states. The study of threshold states has been started in statistical physics of disordered systems, where it has been conjectured and verified that they play a prominent role in the gradient descent dynamics of such random systems [@cugliandolo1993analytical]. according to this theory, threshold states are the most numerous minima and the Hessian associated to the threshold states has a spectrum that is positive semi-definite and gap-less. In the previous section we have obtained a closed set of equations that allow us to obtain $\alpha_{\rm BBP}$ once the distribution $\PP(\hat{y},y)$ is known. Such distribution can be derived using tools from statistical physics, in particular replica theory [@monasson1995structural; @antenucci2018glassy; @franz2017universality]. The main quantity of interest is the partition function, which is defined as the normalization constant of the Gibbs distribution associated with the loss Eq. , $$\label{eq:partition_function} {\mathcal{Z}} = \int_{\SS^{\alpha(N-1)}(\sqrt{N})} e^{-\beta\mathcal{L}(\pmb W)} d\pmb W,$$ where $\beta$ is a parameter associated to the inverse temperature in the physical analog of the problem. We consider the $\beta\to\infty$ limit to study the minima of $\mathcal{L}$ that are relevant in gradient-flow dynamics. The disordered systems approach focuses on the average properties of the systems. In order to ensure concentration properties as the input dimension goes to infinity, the quantity to average is the logarithm of the partition function. This quantity in general is hard to compute and we resort to replica theory. The idea is to apply replica method, described below and in more detail in the Appendix Sec. \[si:1rsb\], to move from the moments of the partition function to average of its logarithmic value. The analysis leads to the free energy density, Eq. , that we use to compute the distribution of the labels. We start writing the moments of Eq.  $$\label{eq:avg_partition_function} \overline{\mathcal{Z}^n} = \int_{\SS^{\alpha(N-1)}(\sqrt{N})} \EE_{(\pmb x, y)} e^{-\beta n\mathcal{L}(\pmb W)} d\pmb W,$$ where we represent the average over the possible datasets with the overbar. The high-dimensional integral in Eq.  can be written in term of the overlap matrix $Q_{a,b}=\langle \pmb W_a \cdot \pmb W_b\rangle/N$ that encodes the similarity between two configurations extracted with the Gibbs measure, of the labels $\hat{y}_\mu$ and $y_\mu$, $\overline{\mathcal{Z}^n} = \int e^{N n S(\pmb Q)} \prod_{a\ge b=0}^n dQ_{a,b}$ with $$\label{eq:action} S(\pmb Q) = \frac12\log\det\pmb Q + \alpha\log \exp \left[\frac12\sum_{a,b=0}^{n}Q_{a,b}\frac{\partial^2}{\partial h_a\partial h_b}\right] \exp\left[ -\frac\beta2\sum_{a}\ell(h_a,h_0)\right]\Bigg|_{\{h_a=0\ \forall a\}} \, .$$ We perform a so-called *1-step replica symmetry breaking* (1RSB) analysis [@MPV87] that allows us to obtain an explicit expression for the distribution of $(\hat{y},y)$. Using the 1RSB ansatz on this problem we reduce the number of parameters in Eq.  to $\chi$ and $z$: where $\chi$ is related to amplitude of the fluctuations in minimum when this is perturbed, and $z$ is a Lagrange multiplier that we use to enforce that the minima in consideration have a Hessian with gapless spectrum, i.e. they are threshold states. The general prescription for finding the global minimum of this kind of problem would require to optimize over $\chi$ and $z$, however, the same formalism can be used to characterize the threshold states by imposing a value for $z$ and optimizing only over $\chi$ [@monasson1995structural; @zamponi2010mean]. We evaluate the integral of the partition function Eq.  at the threshold using Laplace’s method on Eq. . Finally we consider the replica method, by taking the analytic continuation of the moments of the partition function and the limit $n\rightarrow0^+$. With this considerations we obtain the free energy density in low temperature limit $f(\chi|\alpha,z) =-\lim_{\beta\rightarrow\infty}\lim_{N\rightarrow\infty}\frac1{\beta N}\overline{\log\mathcal Z} / (\beta N)$ that reads $$\label{eq:1RSB_f} \begin{split} f(\chi|\alpha,z) &= -\frac1{2z}\log\frac{\chi + z}{\chi} - \frac{\alpha}z \int_{\mathbb{R}} dy \frac1{\sqrt{2\pi}}e^{-\frac{y^2}2} \log \gamma_{1}\star e^{\frac{-z}2 V(\hat{y},y|\chi)}\Bigg|_{\hat{y}=0}, \end{split}$$ where $V(\hat{y},y|\chi) \doteq \min_{\tilde y} \frac{(\hat{y}-\tilde{y})^2}\chi + \ell(\tilde{y},y)$, the $\star$ symbol represent a convolution and we write $\gamma_1$ as a compact notation to represent a centered Gaussian distribution with unit variance. The arguments $\chi$ and $z$ are given via implicit equations that respectively impose that the Laplace’s approximation on $f$ and that the free energy describes the threshold states (further details are in the Appendix Sec. \[si:1rsb\]): $$\label{eq:parisi_par} \frac1{\chi(\chi + z)} = \frac \alpha4 \int_{-\infty}^\infty \frac1{\sqrt{2\pi}} e^{-\frac{y^2}2} \frac{\int_{-\infty}^\infty \frac{1}{\sqrt{2\pi}}e^{-\frac{\hat{y}^2}{2}}e^{-\frac z2 V(\hat{y},y|\chi)} \left(\partial_{\hat{y}} V(\hat{y},y|\chi)\right)^2 d\hat{y}}{\int_{-\infty}^\infty \frac{1}{\sqrt{2\pi}}e^{-\frac{\hat{y}^2}{2}}e^{-\frac z2 V(\hat{y},y|\chi)}d\hat{y}}dy;$$ $$\label{eq:replicon} 1 = \frac\alpha4 \chi^2\int_{-\infty}^\infty \frac{1}{\sqrt{2\pi}} e^{-\frac{y^2}{2}} \frac{\int_{-\infty}^\infty \frac{1}{\sqrt{2\pi}}e^{-\frac{\hat{y}^2}{2}}e^{-\frac z2 V(\hat{y},y|\chi)} \left(\partial_{\hat{y}}^2 V(\hat{y},y|\chi)\right)^2 d\hat{y}}{\int_{-\infty}^\infty \frac{1}{\sqrt{2\pi}}e^{-\frac{\hat{y}^2}{2}}e^{-\frac z2 V(\hat{y},y|\chi)}d\hat{y}}dy \, .$$ In order to obtain $\PP(\hat{y},y)$ we follow the strategy introduced in [@franz2017universality]. The partition function in Eq.  can be written as a functional integral over the empirical labels $\rho_\text{quen}(\hat{y},y)=\frac1{\alpha N}\sum_{m=1}^{\alpha N} \delta(y-y_m) \delta(\hat{y}-\hat{y}_m)$: $$\mathcal{Z} = \int {\mathcal D}[\rho_\text{quen}(\hat{y},y)] \, e^{ -\frac{\beta \alpha N}2 \int d\hat{y}dy \rho_\text{quen}(\hat{y},y) \ell(\hat{y},y)+NS[\rho_\text{quen}(\hat{y},y)]} ,$$ where $S[\rho_\text{quen}(\hat{y},y)]$ is the entropic factor (evaluated on configurations corresponding to the threshold states). This makes clear that the free-energy can also be obtained in terms of a large-deviation principle: $$f=\text{Min}_\rho \left( \frac \alpha 2 \int d\hat{y}dy \rho_\text{quen}(\hat{y},y) \ell(\hat{y},y)-S[\rho_\text{quen}(\hat{y},y)]\right)$$ The minimizer of this variational problem corresponds by definition to $\PP(\hat{y},y)$. By taking the functional derivative of $f$ with respect to $\ell(h,h_0)$ one obtains: $$\label{eq:18} \frac{\delta f}{\delta\ell(\hat{y},y)} = \frac\alpha2 \PP(\hat{y},y).$$ Since we have an explicit expression of $f$ in terms of $\ell$, see eq. (\[eq:1RSB\_f\]), we can obtain this distribution directly. In taking the functional derivative of Eq.  we must be careful in considering the implicit dependence of $V(\hat{y},y|\chi)$ from $\ell(\hat{y},y)$. Finally inverting Eq. \[eq:18\] we obtain the label distribution $$\begin{aligned} \label{eq:gap_distribution} & \PP_{\rm 1RSB}(\hat{y},y) = \frac1{\sqrt{2\pi}}e^{-\frac{y^2}2}\frac{ \exp\left[-\frac{\hat{y}^2}2-\frac{z}2 V(\hat{y},y|\chi)\right]}{ \int_\RR \exp\left[-\frac{\tilde{y}^2}2-\frac{z}2 V(\tilde{y},y|\chi)\right]d\tilde{y} }.\end{aligned}$$ The index 1RSB denotes that the free-energy has been obtained within the 1RSB scheme. We run numerical experiments to characterize the threshold states in terms of their energy and their moments with respect to the label probability distribution. In Fig. \[fig:replica\_th\_energy\] (left) the 1RSB threshold energy is plotted together with the value of the plateau obtained from the simulations in Fig. \[fig:energy\_planted\_vs\_unplanted\]. The 1RSB ansatz appears to be a good approximation of the empirically obtained energy. In the inset we highlight the discrepancy with the empirical line, which may be due to both finite size effects as well as to the fact that the 1RSB scheme is an approximation and more involved schemes must be employed to obtain the exact distribution [@MPV87]. Finally in Fig. \[fig:replica\_th\_energy\] (right) we show the first $4$ moments of the label distribution from the 1RSB analysis and we compare them with the numerical results. The 1RSB moments give a nice agreement in relation to the empirical ones. Encouraged by the reasonable, though not perfect, accuracy of the 1RSB approximation, we use the expression   as input for the three eqs. in (\[eq:lbar\],\[eq:BBP\_condition\]). Their solutions leads to a finite threshold at $\alpha_{\rm BBP}^{\rm 1RSB} \approx 13.8$. This value could be compatible with the numerical results presented in Fig. \[fig:BBP\_analysis\] if the finite size shift saturates for yet larger sizes. Discussion ========== In Fig. \[fig:empirical\_th\] we present the fraction of runs of gradient descent that converge to zero test error in the phase retrieval problem under consideration. We see that a large fraction of simulations achieves convergence before the $\alpha_{\rm BBP}^{\rm 1RSB}$ threshold, in general before $\alpha\sim7.0$. The mechanism behind this difference is not clear to us, as well as it is not clear whether it will hold as the input dimension tends to infinity. The fact that the empirical threshold seems even smaller than the one predicted by our approximate theory strengthens our conjecture that the critical value of $\alpha$ to get weak recovery for randomly initialized gradient descent is of order one and not polynomial in $\log N$. In [@sarao2019passed] the authors showed that large finite size effects in the initialization affect the location of the BBP transition. Whether this happens also in this case is unclear and deserves further investigations. ![Fraction of simulations with gradient descent that achieve zero test error for various input dimensions $N$. The number of simulations increases with the smaller input dimension in order the account for the larger fluctuations. The number varies from $500$ simulations for $N=8192$ to $125000$ simulations for $N=32$. The learning rate is $0.006$ and the simulation runs until either the loss goes below $10^{-8}$ or the number of steps exceeds $1500\log_2 N$ iterations. The logarithmic term in the time accounts for the fact that the average initial overlap with the ground truth is $\sim1/\sqrt{N}$. []{data-label="fig:empirical_th"}](img/PR_SQ_success.pdf){width=".6\linewidth"} Let us conclude by commenting on the limitations and possible generalizations of the results presented in this paper. A key element of the phase retrieval model is the existence of a phase at small $\alpha$ in which the best achievable generalization error is not better than random guess. This arises in models that present a symmetry, such as the $\pm 1$ symmetry due to the absolute value in the phase retrieval problem. The picture shown in Fig. \[fig:energy\_planted\_vs\_unplanted\] where the threshold states are characterized using the gradient-flow dynamics in a variant of the model with randomized labels is enabled by this symmetry and the rest of our analysis relies on the simplifications stemming from this. We expect that the BBP picture presented in this work generalizes to all learning where at small sample complexity error is not better than random guess. Of course working out the details, even for two layer neural networks, is an open problem of interest for future work. In most learning problems a naive linear regression often achieves a better than random guess performance, and thus an extension of our theory for problems of that type (lacking a symmetry) would be an interesting direction for future work. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Carlo Lucibello for precious discussions. We acknowledge funding from the ERC under the European Union’s Horizon 2020 Research and Innovation Programme Grant Agreement 714608-SMiLe; from the French government under management of Agence Nationale de la Recherche (ANR) grant PAIL, and grant “Investissements d’Avenir” LabEx PALM (ANR-10-LABX-0039-PALM) (StatPhysDisSys) and (ANR-19-P3IA-0001) (PRAIRIE 3IA Institute); and from the Simons Foundation collaboration Cracking the Glass Problem (\#454935, Giulio Biroli). [10]{} Rong Ge, Jason D Lee, and Tengyu Ma. Matrix completion has no spurious local minimum. In [*Advances in Neural Information Processing Systems*]{}, pages 2973–2981, 2016. Kenji Kawaguchi. Deep learning without poor local minima. In [*Advances in Neural Information Processing Systems*]{}, pages 586–594, 2016. Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error guarantees for multilayer neural networks. , 2016. Jason D Lee, Max Simchowitz, Michael I Jordan, and Benjamin Recht. Gradient descent only converges to minimizers. In [*Conference on learning theory*]{}, pages 1246–1257, 2016. Itay Safran and Ohad Shamir. Spurious local minima are common in two-layer relu neural networks. , 2017. Shengchao Liu, Dimitris Papailiopoulos, and Dimitris Achlioptas. Bad global minima exist and sgd can reach them. , 2019. Stefano Sarao Mannelli, Florent Krzakala, Pierfrancesco Urbani, and Lenka Zdeborova. Passed & spurious: analysing descent algorithms and local minima in spiked matrix-tensor model. , 2019. Stefano Sarao Mannelli, Giulio Biroli, Chiara Cammarota, Florent Krzakala, and Lenka Zdeborov[á]{}. Who is afraid of big bad minima? analysis of gradient-flow in spiked matrix-tensor models. In [*Advances in Neural Information Processing Systems*]{}, pages 8676–8686, 2019. Leticia F Cugliandolo and Jorge Kurchan. Analytical solution of the off-equilibrium dynamics of a long-range spin-glass model. , 71(1):173, 1993. Fabrizio Antenucci, Silvio Franz, Pierfrancesco Urbani, and Lenka Zdeborov[á]{}. Glassy nature of the hard phase in inference problems. , 9(1):011020, 2019. TLH Watkin and J-P Nadal. Optimal unsupervised learning. , 27(6):1899, 1994. Stefano Sarao Mannelli, Giulio Biroli, Chiara Cammarota, Florent Krzakala, Pierfrancesco Urbani, and Lenka Zdeborov[á]{}. Marvels and pitfalls of the langevin algorithm in noisy high-dimensional inference. , 10(1):011057, 2020. Adriaan Walther. The question of phase retrieval in optics. , 10(1):41–49, 1963. Rick P Millane. Phase retrieval in crystallography and optics. , 7(3):394–411, 1990. Robert W Harrison. Phase problem in crystallography. , 10(5):1046–1055, 1993. Oliver Bunk, Ana Diaz, Franz Pfeiffer, Christian David, Bernd Schmitt, Dillip K Satapathy, and J Friso Van Der Veen. Diffractive imaging for periodic samples: retrieving one-dimensional concentration profiles across microfluidic channels. , 63(4):306–314, 2007. Radu Balan, Pete Casazza, and Dan Edidin. On signal reconstruction without phase. , 20(3):345–356, 2006. John V Corbett. The pauli problem, state reconstruction and quantum-real numbers. , 1(57):53–68, 2006. Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase retrieval via wirtinger flow: Theory and algorithms. , 61(4):1985–2007, 2015. Yuxin Chen, Yuejie Chi, Jianqing Fan, and Cong Ma. Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval. , 176(1-2):5–37, 2019. Philip Schniter and Sundeep Rangan. Compressive phase retrieval via generalized approximate message passing. , 63(4):1043–1055, 2014. Jean Barbier, Florent Krzakala, Nicolas Macris, L[é]{}o Miolane, and Lenka Zdeborov[á]{}. Optimal errors and phase transitions in high-dimensional generalized linear models. , 116(12):5451–5460, 2019. Marco Mondelli and Andrea Montanari. Fundamental limits of weak recovery with applications to phase retrieval. , 19(3):703–773, 2019. Ju Sun, Qing Qu, and John Wright. A geometric analysis of phase retrieval. , 18(5):1131–1198, 2018. Jinho Baik, G[é]{}rard Ben Arous, Sandrine P[é]{}ch[é]{}, et al. Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. , 33(5):1643–1697, 2005. Yue M Lu and Gen Li. Phase transitions of spectral initialization for high-dimensional non-convex estimation. , 2019. Leticia F Cugliandolo. Dynamics of glassy systems. , 2002. R[é]{}mi Monasson. Structural glass transition and the entropy of the metastable states. , 75(15):2847, 1995. Silvio Franz, Giorgio Parisi, Maxime Sevelev, Pierfrancesco Urbani, and Francesco Zamponi. Universality of the sat-unsat (jamming) threshold in non-convex continuous constraint satisfaction problems. , 2(3):019, 2017. Marc M[é]{}zard, Giorgio Parisi, and Miguel-Angel Virasoro. World Scientific Publishing, 1987. Francesco Zamponi. Mean field theory of spin glasses. , 2010. Hidetoshi Nishimori. , volume 111. Clarendon Press, 2001. Florent Krzakala and Lenka Zdeborov[á]{}. Hiding quiet solutions in random constraint satisfaction problems. , 102(23):238701, 2009. Li, Ker-Chau. Exact theory of dense amorphous hard spheres in high dimension. ii. the high density regime and the gardner transition. , 87(420):1025–1039, 1992. Jorge Kurchan, Giorgio Parisi, Pierfrancesco Urbani, and Francesco Zamponi. Exact theory of dense amorphous hard spheres in high dimension. ii. the high density regime and the gardner transition. , 117(42):12979–12994, 2013. BBP transition of spectra initialization {#si:bbp} ======================================== In non-convex estimation problems, such as Phase Retrieval, big advantages follow from the development of well tailored spectral methods to be used as initialization step. Recently the outcomes of one such spectral method widely used in Phase Retrieval and based on the construction of a data matrix $$\mathcal{M}_{i,j}=\frac{1}{\alpha N}\sum_{m=1}^{\alpha N} \mathcal{T}(Y_m)u_{m,i}u_{m,j}=\frac{1}{\alpha }\sum_{m=1}^{\alpha N} \mathcal{T}(Y_m)x_{m,i}x_{m,j}$$ from sensing vectors $\pmb u_m$, with elements of order one (or sensing vectors $\pmb x_m$ on the unitary sphere, according to our definition) and measurements $Y_m$ has been exactly derived [@lu2017phase]. The method involves a pre processing function $\mathcal{T}(Y_m)$ that can be optimized to further improve the results. Once the data matrix is constructed the eigenvector $\bm{v_1}$ corresponding to the largest eigenvalue $\lambda_1$ can be used as an estimator of the signal $\bm{W^*}$.\ To obtain the performances of this kind of spectral initialization it is assumed [@lu2017phase] that the measurements $Y_m$ are independently drawn according to a density function conditional on $y_m=\langle \bm{x}_m\bm{W^*}\rangle$ associated to the particular acquisition process, and it is recalled that $y_m$ are themselves Gaussian random variables, due to the definition of the problem. Finally in the large $N$ limit the empirical average used to construct the data matrix can be replaced by the expected value $\mathbb{E}_{Y,y}$ over these two distributions.\ The result [@lu2017phase] goes as follows. Given two functions defined as $$\Psi_{\alpha}(\lambda)\equiv \lambda \left[\frac{1}{\alpha}+\mathbb{E}_{Y,y}\left( \frac{\mathcal{T}(Y)}{\lambda-\mathcal{T}(Y)} \right) \right]$$ and $$\Phi(\lambda)=\lambda\mathbb{E}_{Y,y}\left(\frac{\mathcal{T}(Y)y^2}{\lambda-\mathcal{T}(Y)}\right)$$ with $\lambda > \mathcal{T}(Y)$, and given $$\xi_\alpha(\lambda)=\Psi_\alpha(\max\{\lambda,\bar\lambda\})$$ with $$\bar\lambda = \arg\min\Psi_\alpha(\lambda) \ ,$$ the two largest eigenvalues of $\mathcal{M}$, $\lambda_1$ and $\lambda_2$, are such that $$\lambda_1 \rightarrow_\mathbb{P} \xi_\alpha(\lambda_\alpha^*) \ ,$$ with $\lambda_\alpha^*$ the solution of $\xi_\alpha(\lambda) = \Phi(\lambda)$, and $$\lambda_2 \rightarrow_\mathbb{P} \Psi_\alpha(\bar\lambda).$$ A phase transition occurs at the largest $\alpha$ such that $\lambda_1=\lambda_2$, which can be evaluated by imposing $\Psi'_{\alpha}(\lambda_\alpha^*)=0$ or equivalently $$\frac 1 \alpha =\mathbb{E}_{Y,y}\left(\frac{\mathcal{T}(Y)^2}{(\bar{\lambda}-\mathcal{T}(Y))^2}\right) \ , $$ which corresponds to $\Psi'_{\alpha}(\bar{\lambda})=0$ as at that point $\lambda_\alpha^*=\bar{\lambda}$. At larger $\alpha$ the largest eigenvalue pops out from the spectrum bulk ($\lambda_1\neq\lambda_2$) and the corresponding eigenvector develops a finite correlation with the signal, in a phenomenon called BBP transition [@baik2005phase], hence the definition of $\alpha_{\rm{BBP}}$ when this occurs. It is interesting to note that the structure of the data matrix is closely reminiscent of the structure of the first term in the Hessian of our problem (see Eq. ). Indeed incidentally the original idea of this spectral method for initialization can be traced back to the study of Hessian’s principal directions [@li92]. In particular we observe that $$-\mathcal{H}_{i,j}(\pmb W) = -\frac12\sum_{m=1}^{\alpha N} \frac{\partial^2}{\partial \hat{y}_m^2} \ell(\hat{y}_m,y_m) x_{m,i} x_{m,j} + \mu \delta_{i,j}=\mathcal{M}_{i,j} + \mu \delta_{i,j} \ ,$$ provided that the pre processing function is $$\mathcal{T}(Y)=-\frac{\alpha }{2}\frac{\partial^2}{\partial \hat{y}^2}\ell(\hat{y},y)=-\frac{\alpha }{2}\partial_{\hat{y}}^2\ell(\hat{y},y) \ .$$ Note that we consider a case for which measurements $Y$ have a one to one correspondence with $y$ ([*i.e.*]{} $Y=y^2$). Moreover in the problem discussed empirical averages involve not only measurements $Y$ but also estimated $\hat{y}=\langle \bm{x}\bm{W}\rangle$ in correspondence of some $\bm{W}$ of interest. Therefore the expected value, $\mathbb{E}_{\hat{y},y}$, should be taken over the relative joint probability distribution function $P(\hat{y},y)$. In conclusion, the results just mentioned tell immediately what is the largest $\alpha$ ([*i.e.*]{} $\alpha_{\it{BBP}}$): $$\frac {1 }{\alpha_{\rm{BBP}}} =\mathbb{E}_{\hat{y},y}\left(\frac{\alpha_{\rm{BBP}}\partial_{\hat{y}}^2\ell(\hat{y},y)}{2\bar{\lambda}+\alpha_{\rm{BBP}}\partial_{\hat{y}}^2\ell(\hat{y},y)}\right)^2 \ ,$$ before the smallest eigenvalue of the Hessian pops out from the spectrum bulk, being also associated to an eigenvector with finite projection on the signal to be detected. In this case we are focusing on performance of a gradient descent dynamics, which for mean-field spin-glasses naturally gets stuck on what are called threshold states [@cugliandolo2002dynamics]. We argue that the gradient descent dynamics applied to Phase Retrieval, when retrieval fails, will also approach threshold states which are mainly characterized by their property of being marginal, [*i.e.*]{} the smallest eigenvalue of their Hessian is null. This qualifies the relevant $\bm{W}$ as a typical configuration belonging to threshold states and $P(\hat{y},y)$ as the joint probability distribution at threshold states. Moreover it allows to introduce the marginal condition $\lambda_2=\lambda_1=-\mu$, which can be re-expressed by equating $\lambda_2=\Psi_\alpha(\bar{\lambda})$ to $-\mu$: $$\mu=\bar{\lambda}\mathbb{E}_{\hat{y},y}\left(\frac{\alpha_{\rm{BBP}}\partial^2_{\hat{y}}\ell(\hat{y},y)y^2}{2\bar{\lambda}+\alpha_{\rm{BBP}}\partial^2_{\hat{y}}\ell(\hat{y},y)}\right) \ .$$ Finally the definition of the spherical parameter in the main text must be considered to close the system of equations $$\mu=\frac{\alpha_{\rm{BBP}}}{2}\mathbb{E}_{\hat{y},y}\left(\partial_{\hat{y}}\ell(\hat{y},y)\hat{y}\right)$$ to be used to determine $\alpha_{\rm{BBP}}, \bar{\lambda}, \mu$ in correspondence of $P(\hat y, y)$ from threshold states. The resulting picture is as follows. In the small $\alpha$ regime, gradient descent will systematically approach threshold states and remain stuck there. However starting from $\alpha_{\rm{BBP}}$ in the Hessian of these states, that is otherwise marginal, an isolated eigenvalue pops out from the bulk immediately becoming negative. Moreover the eigenvector associated to such negative direction, naturally followed by gradient descent, has a finite overlap with the signal. Therefore, we argue, it is from that point on that the signal should be easily retrieved. Replica Analysis {#si:1rsb} ================ The field of physics of disordered systems has developed numerous tools to deal with random systems [@MPV87; @nishimori2001statistical]. At an abstract level of thinking, using those tools means that we identify the inference problem with a physical systems that subject to a certain potential. The randomness comes from the having a dataset made of random projections. The estimator is mapped into a spherical spin and the loss function becomes the energy - or *Hamiltonian*. Finally the system the temperature in which the system lives is sent to zero and the system tend to the lowest energy, herefore minimizing the loss. The ground truth in the inference problem becomes equivalent to a minimum planted in the energetic landscape [@krzakala2009hiding]. For example this formulation is equivalent to a physical system that before the experiment is liquid, but as we cool it down it can become either a crystal - an ordered solid - or a glass - an amorphous solid. Finding the crystal means reconstructing the signal. Partition function ------------------ Moving to the mathematical aspects of the problem. We define a Gibbs distribution associated with the problem and evaluate its normalization constant - the *partition function* $\mathcal Z$. The partition function, and in particular its logarithm divided by the temperature - the free energy -, contains all the information we aim to understand in the problem. By taking the proper derivatives, and possibly add an external field, we can compute relevant macroscopic properties such as: average overlap with the ground truth, average loss achieved. In disordered system we have to consider the additional complication given by the randomness. Therefor, we need to consider the average free energy that is the average of the logarithm of a high-dimensional integral, that can be done by the simple observation: $$\label{eq:replica_trick} \log\mathcal Z = \lim_{n\rightarrow0}\frac{\mathcal{Z}^n-1}n.$$ Which for arbitrary $n$ is not simpler than computing the logarithm, but it is much simpler if $n\in\mathbb{N}$ and we perform an analytic continuation to $n\in\RR$. Under this - replica trick - the average of the logarithm is equivalent to compute the average of the $n$th moment of the partition function and take the limit. Formally the $n$th moment correspond to the partition function of $n$ identical - replicated - system that do not interact but with the same realization of the disorder. The problem is now computing the moments of the partition function which in general is prohibitive and we have to use an ansatz on the specific form of the solution, in particular we use the so called *replica symmetry breaking ansatz* [@MPV87]. This largely reduces the number of parameters. Finally the average free energy can be evaluated by set of saddle point equations. We can now move to the analysis. The partition function already defined in the main text is $$\label{eq:partition_function_2} \mathcal{Z} = \int_{\mathbb{S}^{N-1}(\sqrt{N})} \mathcal{D} \pmb W_S \, \exp\left\{ -\frac\beta2\sum_\mu\ell(\langle \pmb x_\mu, \pmb W_S\rangle,\langle \pmb x_\mu, \pmb W_T\rangle) \right\}$$ and consider its $n$th moment - the *replicated partition function* - $$\overline{\mathcal{Z}^n} = \EE_{\{\pmb x_\mu\}}\int_{\mathbb{S}^{n(N-1)}(\sqrt{N})} \prod_{a=1}^n \mathcal{D} \pmb W_a \, \exp\left\{ -\frac\beta2\sum_\mu \ell(|\langle \pmb x_\mu, \pmb W_S\rangle|,|\langle \pmb x_\mu, \pmb W_T\rangle|) \right\}\,.$$ This is formally equivalent to have $n$ independent systems. Introduce the overlaps with the projector $r_\mu^{(a)} = \langle \pmb x_\mu, \pmb W_a \rangle$ with indices $a=0,\dots,n$ where $a=0$ is the overlap with the ground truth and the others are the overlaps with the estimators of the $n$ systems. Introduce those quantities in the replicated partition function via Dirac’s deltas using their Fourier transform $$\begin{split} \overline{\mathcal{Z}^n} &=\EE_{\{\pmb x_\mu\}}\int \prod_a\mathcal{D} \pmb W_a \int \mathcal{D}(r,\hat r)\exp\left\{ -\frac\beta2\sum_{a,\mu}\ell(r_\mu^{(a)},r_\mu^{(0)}) + i\sum_{a,\mu} \hat{r}_\mu^{(a)}r_\mu^{(a)} - i\sum_{a,\mu} \langle\pmb x_\mu, \hat{r}_\mu^{(a)}\pmb W_a\rangle \right\} = \\ &=\int \prod_a\mathcal{D} \pmb W_a \int \mathcal{D}(r,\hat r)\exp\left\{ -\frac\beta2\sum_{a,\mu}\ell(r_\mu^{(a)},r_\mu^{(0)}) + i\sum_{a,\mu} \hat{r}_\mu^{(a)}r_\mu^{(a)} - \frac1{2N}\sum_\mu\sum_{a,b=0}^n \hat{r}_\mu^{(a)} \hat{r}_\mu^{(b)}\langle\pmb W_a,\pmb W_b\rangle \right\}. \end{split}$$ Where $\mathcal{D}$ contains the normalization factor of the Fourier transform. We introduce the matrix of overlaps between estimators and ground ground truth $\pmb Q$, $Q_{ab}=\frac1N\langle\pmb W_a,\pmb W_b\rangle$. This is done using the same idea of introducing delta function and gives a contribution $\frac{N}2\log\det\pmb Q$ to the action [@kurchan2013mech; @zamponi2010mean]. The equation can now be factorized in $\mu$, so we can drop the $\mu$s from $r$ and $\hat r$. Observing that $$\label{eq:observation_dha_dhb} e^{-\frac12\sum_{ab}\hat{r}^{(a)}\hat{r}^{(b)}Q_{ab}} = e^{\frac12\sum_{ab}Q_{ab}\frac{\partial^2}{\partial h_a\partial h_b}}e^{-i\ h_a\hat{r}^{(a)}}\big|_{\{h_a=0\}_a},$$ we can integrate over $r$ and $\hat r$, and write a simplified replicated partition function $\overline{\mathcal{Z}^n}=\int \prod_{a\ge b=0}^{n}dQ_{ab} e^{N S(\pmb Q)}$ with action $$\label{eq:action_no_asatz} S(\pmb Q) = \frac12\log\det\pmb Q + \alpha\log \exp \left[\frac12\sum_{a,b=0}^nQ_{a,b}\frac{\partial^2}{\partial h_a\partial h_b}\right] \exp\left[ -\frac\beta2\sum_{a}\ell(h_0,h_a)\right]\Big|_{\{h_a=0\}_a}.$$ where the first term is an entropic term that accounts for the degeneracy of the matrix $\pmb Q$ in the space of symmetric matrices. And the second term is an energetic term that accounts for the potential acting on the system. Observe that so far we did not make any ansatz on the structure of the overlap matrix $\pmb Q$. In the next subsections we will consider the 1-step replica symmetry breaking ansatz (1RSB). 1 step replica symmetric breaking {#sec:1RSB} --------------------------------- The 1RSB scheme consists in making an ansatz on the structure of the overlap matrix $\pmb Q$ [@MPV87]. The assumption is that not all the replicated systems will have the same overlap - which correspond to the replica symmetric ansatz - but the systems are clustered. Systems inside the same cluster will have a larger overlap, systems outside the cluster will have a smaller overlap. This translates into the following parameters: $q_1$ overlaps inside the same cluster, $q_0$ overlaps in different clusters, $x$ dimension of the clusters, finally $m$ the overlap with the signal. Schematically we have $$\label{eq:1RSB_matrix} \pmb Q = \begin{pmatrix} 1 & m & m \\ m & \begin{pmatrix} 1 & q_1& q_1 \\ q_1 & 1 & q_1 \\ q_1 & q_1 & 1 \\ \end{pmatrix} & q_0 \\ m & q_0 & \begin{pmatrix} 1 & q_1& q_1 \\ q_1 & 1 & q_1 \\ q_1 & q_1 & 1 \\ \end{pmatrix} \\ \end{pmatrix}$$ with $\pmb Q$ of dimension $(n+1)\times(n+1)$, and the inner matrices of dimension $x\times x$. The analysis proceed as in a standard way: we derive the 1RSB free energy with the associated saddle point equations and move to the zero temperature limit [@MPV87; @nishimori2001statistical], we impose the marginal stability - corresponding to the threshold states - [@monasson1995structural], finally we derive the label distribution [@franz2017universality]. For notational convenience we call $\gamma_a(x)$ the probability density function of a Gaussian with zero mean and (co-)variance $a$, and we use the symbol $\star$ to indicate the convolutions, i.e. $f\star g(x) = \int_\RR f(x-t)g(t) dt$. We can plug Eq.  into Eq.  and obtain 1RSB formulation of the action. $$\begin{split} \frac1nS_{\rm 1RSB}(\pmb Q) & = \frac12 \log(1-q_1) + \frac1{2x}\log\frac{1-q_1 + x(q_1-q_0)}{1-q_1} + \frac12\frac{q_0-m^2}{1-q_1+x(q_1-q_0)} + \\ & + \alpha \int_{\mathbb{R}^2} d\hat{y}dy \gamma_{\pmb\Sigma}(\hat{y},y) \frac1x \log \left[ \gamma_{q_1-q_0}\star\left( \gamma_{1-q_1}\star e^{-\frac\beta2\ell(\hat{y},y)} \right)^x \right]. \end{split}$$ and using Eq.  the free energy is $-\frac1{n\beta} S_{\rm 1RSB}$. We can now take derivatives to find the saddle point equations. $$\begin{aligned} \label{eq:1RSB_saddle_point_q1} \begin{split} & \frac1x\left(\frac1{1-q_1}-\frac1{1-q_1+x\ (q_0-q_1)}\right) + \frac{q_0-m^2}{[1-q_1+x\ (q_0-q_1)]^2} = \\ & \quad\quad\quad = \alpha\int_{\mathbb{R}^2} d\hat{y} dy \gamma_\Sigma(\hat{y},y) e^{-x\ f(x,\hat{y},y)} \gamma_{q_1-q_0}\star \left[ e^{x\ f(1,\hat{y},y)}\left(\frac{d}{d\hat{y}} f(1,\hat{y},y) \right)^2 \right] \end{split}; \\ \label{eq:1RSB_saddle_point_q0} & \frac{q_0-m^2}{[1-q_1+x\ (q_0-q_1)]^2} = \alpha \int_{\mathbb{R}^2} d\hat{y}dy \gamma_\Sigma(\hat{y},y)\left(\frac{d}{dh} f(x,\hat{y},y) \right)^2; \\ \label{eq:1RSB_saddle_point_m} & \frac{m}{1-q_1+x\ (q_0-q_1)} = \alpha \int_{\mathbb{R}^2} d\hat{y}dy \gamma_\Sigma(\hat{y},y) \frac{d^2}{d\hat{y} dy} f(x,\hat{y},\hat{y});\end{aligned}$$ where we define $$\begin{aligned} & f(1,\hat{y},y) = \log\left[ \gamma_{1-q_1}\star e^{-\frac\beta2\ell(\hat{y},y)}\right]; \\ & f(x,\hat{y},y) = \frac1x\log\left[ \gamma_{q_1-q_0}\star\left(\gamma_{1-q_1}\star e^{-\frac\beta2\ell(\hat{y},y)}\right)^x \right] = \frac1x\log\left[ \gamma_{q_1-q_0}\star e^{x f(1,\hat{y},y)} \right]; \\ & \ell(\hat{y},y) = (\hat{y}^2-y^2)^2;\end{aligned}$$ and $\Sigma$ is a $2\times2$ covariance matrix with entries $\Sigma_{11} = 1$, $\Sigma_{12} = m$ and $\Sigma_{22} = q_0$. #### Zero temperature and parameter ansatz {#zero-temperature-and-parameter-ansatz .unnumbered} As the zero temperature goes to zero the order parameters $q_1$ and $x$ needs to be rescaled with the temperature as $q_1 \approx 1 - \chi T$ and $x \approx z T$ with $\chi, z \sim O(1)$. Instead $m$ and $q_0$ will not be affected by the limit. The reason why some parameters need to be rescaled is that as long as there is a positive temperature the replicated systems exploit this thermal energy to fluctuate in the basin of the minimum, therefor their overlap $q_1$ is given by average overlap in the basin of attraction. When the temperature drops to zero the thermal goes to zero and all the system shrink to a point. $\chi$ represent the fluctuation that systems can have when they receive an infinitesimal amount of thermal energy. With the same physical reason, also the cluster itself shrinks to a point. The rate of convergence to the point is given by $z$. Under those observation we obtain the 1RSB free energy $$\label{eq:1RSB_free_energy} \begin{split} - f(\pmb Q) & \approx_{\beta\gg1} \frac1{2z}\log\frac{\chi + z(1-q_0)}{\chi} + \frac12\frac{q_0}{\chi+z(1-q_0)} + \\ &\quad+\frac{\alpha}y \int_{\mathbb{R}} dy \gamma_{\pmb\Sigma}(\hat{y},y) \log \gamma_{1-q_0}\star e^{\frac{-y}2 V(\hat{y},y|\chi)}\Bigg|_{\hat{y}=0} . \end{split}$$ and the corresponding saddle point equations from Eqs. (\[eq:1RSB\_saddle\_point\_q1\]-\[eq:1RSB\_saddle\_point\_m\]). However, the solution of those equation will lead to the global minimum of the loss, while we are interested in the threshold states. We follow the idea of [@monasson1995structural] where $z$ is used as a Lagrange multiplier that selects the threshold states. $z$ is fixed by imposing the marginal stability condition, i.e. that the spectrum of the Hessian of the minima in consideration touches zero, following [@franz2017universality] this is given by $$1 = \frac\alpha4 \chi^2 \int_{\mathbb{R}} dy \frac1{\sqrt{2\pi}}e^{-\frac{y^2}2} \frac{\gamma_1\star\left[V''(\hat{y},y|\chi)^2 e^{-\frac{z}2 V(\hat{y},y|\chi)} \right]}{\gamma_1\star e^{-\frac{z}2 V(\hat{y},y|\chi)}}\Bigg|_{\hat{y}=0}.$$ At the threshold the overlap with the signal is zero, $m=0$, and for symmetries of the problem also $q_0=0$. In fact $m=0$ is always a solution of Eq. , and if $m=0$ then $q_0=0$ is also solution of the second saddle point equation, Eq. , as in the RHS the Gaussian becomes degenerate in $h$ and $d_h f(x,h,h_0)|_{h=0}$ being an odd function in $h$. Therefore we can restrict the equations to just the one for $\chi$, Eq.  that becomes $$\begin{split} & \frac{\beta^2}{z} \left(\frac1{\chi}-\frac1{\chi + z}\right) = \alpha\int_{\mathbb{R}} dy \gamma_1(y) e^{-x\ f(x,0,y)} \int_{\mathbb{R}} d\hat{y} \gamma_{1}(\hat{y}) \left[ e^{x\ f(1,\hat{y},y)}\left(\frac{d}{d\hat{y}} f(1,\hat{y},y) \right)^2 \right] \end{split}.$$ Rewriting these equation explicifying the convolutions we obtain the equations presented in the main text. With those element we can obtain the label distribution presented in the main text.
{ "pile_set_name": "ArXiv" }
--- abstract: | The purpose of this paper is to study singular integrals whose kernels $k(x;\xi)$ are variable, i.e. they depend on some parameter $x\in {{\Bbb R}}^n$ and in $\xi\in{{\Bbb R}}^n\setminus\{0\}$ satisfy mixed homogeneity condition of the form $k(x;\mu^{\alpha_1}\xi_1,\ldots,\mu^{\alpha_n}\xi_n)= \mu^{-\sum_{i=1}^n \alpha_i}k(x;\xi)$ with positive real numbers $\alpha_i\geq 1$ and $\mu>0.$ The continuity of these operators in $L^p({{\Bbb R}}^n)$ is well studied by Fabes and Rivière. Our goal is to extend their results in generalized Morrey spaces with a weight satisfying suitable dabbling and integral conditions. A special attention is paid also of the commutators of the kernel with functions of bounded and vanishing mean oscillation. address: ' Bulgarian Academy of Sciences, Institute of Mathematics and Informatics, Acad. G. Bonchev Str. bl. 8, 1113 Sofia, Bulgaria' author: - 'Lubomira G. Softova' title: Singular Integrals and Commutators in Generalized Morrey Spaces --- =13.5cm =-1cm 1362 ¶[[P]{}]{} [H]{} Ł[[L]{}]{} §[[S]{}]{} ł \[subsection\] \[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Remark]{} Introduction {#s1} ============ We consider the following integral operators $$\label{Kf} \K f(x):= P.V. \int_{{{\Bbb R}}^n} k(x;x-y)f(y)dy$$ and its commutators with essentially bounded functions $$\begin{aligned} \nonumber \C[a,k]f(x):= &\ P.V.\int_{{{\Bbb R}}^n} k(x;x-y)[a(y)-a(x)]f(y)dy\\ \label{Cakf} = &\ \K(af)(x)-a(x)\K f(x).\end{aligned}$$ The generating kernel $k(x;\xi) : {{\Bbb R}}^n\times{{\Bbb R}}^n\setminus\{0\}\to {{\Bbb R}}$ is variable, i.e. it depends on some parameter $x$ and possesses “good” properties with respect to the second variable $\xi.$ This class of kernels is firstly studied by Fabes and Rivière in [@FR]. They generalize the classical kernels of Calderón and Zygmund $k(\xi)=\Omega(\xi)/|\xi|^{n}$ having homogeneity of degree $-n$ and those studied by Jones in [@BJ] and satisfying homogeneity property of the form $k(\lambda \xi,\lambda^m \tau)=\lambda^{-n-m}k(\xi,\tau),$ $\xi\in {{\Bbb R}}^n,$ $\tau\in(0,\infty),$ $m\geq 1.$ Introducing a new metric $\rho,$ Fabes and Rivière study in $L^p({{\Bbb R}}^n),$ where ${{\Bbb R}}^n$ is endowed with the topology induced by $\rho$ and defined by ellipsoids. Thus, the unite sphere with respect to $\rho$ coincides with the unite sphere $\Sigma_n$ with respect to the Euclidean metric. This fact allows to impose on the kernel $k$ the Calderón-Zygmund conditions on the unite sphere, in spite of the lack of “symmetry” of $k$ with respect to the variables $\xi_i,$ $i=1,\ldots,n.$ Let we note, that the standard parabolic metric $\tl\rho=\sup\{|x|,\sqrt t\},$ $x\in{{\Bbb R}}^n,t\in(0,\infty),$ for instance, does not permit to define the mentioned above conditions on kernels having homogeneity of parabolic type. Using the Fourier transform in $L^2({{\Bbb R}}^n)$ and the Marcinkiewicz interpolation theorem, Fabes and Rivière obtained that the integral operators are continuous in $L^p({{\Bbb R}}^n),$ $p\in(1,\infty).$ In the present work we study the continuity of these operators in the generalized Morrey spaces $L^{p,\omega}({{\Bbb R}}^n)$ where the function $\omega$ satisfies suitable conditions. A special attention is paid also to the commutators $\C[a,k]$ of the kernel $k$ and functions $a$ having bounded or vanishing mean oscillation. In this case we impose of the results of Coifman-Rochberg-Weiss ([@CRW]) and Bramanti-Cerutti ([@BC]) treating continuity in $L^p({{\Bbb R}}^n)$ of commutators with constant kernels. The technique we used is the one elaborated by Calderón and Zygmund and consisting of expansion of the kernel into spherical harmonics and restricting the considerations on integral operators with constant kernels. Definitions and preliminary results {#s2} =================================== Let $\alpha_1,\ldots,\alpha_n$ be real numbers, $\alpha_i\geq 1$ and define $\alpha=\sum_{i=1}^n\alpha_i.$ Following Fabes and Rivière ([@FR]), the function $ F(x,\rho)=\sum_{i=1}^{n} x_i^2 \rho^{-2\alpha_i}, $ considered for any fixed $x\in{{\Bbb R}}^n,$ is a decreasing one with respect to $\rho>0$ and the equation $F(x,\rho)=1$ is unique solvable in $\rho(x).$ It is a simple matter to check that $\rho(x-y)$ defines a distance between any two points $x,y\in{{\Bbb R}}^n.$ Thus ${{\Bbb R}}^n,$ endowed with the metric $\rho$ results a homogeneous metric space ([@FR Remark 1], [@BC]). The balls with respect to $\rho(x),$ centered at the origin and of radius $r$ are simply the ellipsoids $${\cal E}_r(0)=\left\{x\in{{\Bbb R}}^n\colon\quad \frac{x_1^2}{r^{2\alpha_1}}+\ldots+ \frac{x_n^2}{r^{2\alpha_n}} <1\right\}$$ with Lebesgue measure $|\E_r|=C(n) r^\alpha.$ It is easy to see that the unite sphere with respect to this metric coincides with the unite sphere $\Sigma_n$ with respect to the Euclidean one. \[d1\] The function $ k(x;\xi) \colon{{\Bbb R}}^n\times\{{{\Bbb R}}^{n}\setminus \{0\}\}\to {{\Bbb R}}$ is called a [*variable kernel with mixed homogeneity*]{} if: - for every fixed $x$ the function $k(x;\cdot)$ is a [*constant kernel*]{} satisfying - $k(x;\cdot)\in C^\infty({{\Bbb R}}^{n}\setminus \{0\});$ - $k(x;\mu^{\alpha_1}\xi_1,\ldots, \mu^{\alpha_n}\xi_n)=\mu^{-\alpha}k(x;\xi),$ $\forall$ $\mu>0,$ $\alpha_i\geq 1,$ $\alpha=\sum_{i=1}^n\alpha_i;$ - $ \ds\int_{\Sigma_n} k(x;\xi) d\sigma_\xi =0$and$\ds \int_{\Sigma_n}|k(x;\xi)|d\sigma_\xi<\infty;$ - for every multiindex $\beta:$ $\ds \sup\limits_{\xi\in\Sigma_n} \left|D^\beta_\xi k(x;\xi)\right|\leq C(\beta)$ independently of $x$. Let us note that in the special case $\alpha_i=1$ and thus $\alpha =n,$ Definition \[d1\] gives rise to the classical Calderón-Zygmund kernels. One more example is when $\alpha_1=\ldots=\alpha_{n-1}=1,$ $\alpha_n=\bar\alpha\geq 1.$ In this case we obtain the kernels studied by Jones in [@BJ] and discussed in [@FR]. For the sake of completeness we recall the definitions and some properties of the spaces we are going to use. \[dBMO\] For $f\in L^1_{\rm loc}({{\Bbb R}}^n)$ and any ellipsoid $\E\subset{{\Bbb R}}^n$ centered at $x\in{{\Bbb R}}^n$ and of radius $r>0$ set $$\label{modul} \gamma_f(R):=\sup_{r\leq R} \frac{1}{|{\E}|}\int_{\E} |f(y)-f_{\E}|dy\quad for\ every \ R>0,$$ where $f_{\E}=\frac{1}{|\E|}\int_{\E} f(y) dy $ and $|\E|$ is the Lebesgue measure of $\E,$ comparable to $r^\alpha.$ Then: - $f\in BMO$ $($[*bounded mean oscillation*]{}$)$ if $ \|f\|_\ast:=\sup_R \gamma_f(R)<\infty.$ The quantity $\|f\|_\ast$ is a norm in $BMO$ modulo constant function under which $BMO$ results a Banach space $($see [@JN]$)$; - $f\in VMO$ $($[*vanishing mean oscillation*]{}$)$ with $VMO$-modulus $\gamma_f(R)$ if $f$ belongs to $ BMO$ and $\gamma_f(R)\to 0$ as $R\to 0$ $($see [@S]$)$. For a bounded domain $\Omega\subset {{\Bbb R}}^n,$ we define $BMO(\Omega)$ and $VMO(\Omega)$ taking $f\in L^1(\Omega)$ and $\E\cap \Omega$ instead of $\E$ in $(\ref{modul})$. Let $\omega\colon {{\Bbb R}}^n\times{{\Bbb R}}_+\to {{\Bbb R}}_+$ and for any ellipsoid $\E$ we write $\omega(x,r)=:\omega(\E).$ \[dPMS\] A function $f\in L^p_{\text{loc}}({{\Bbb R}}^n),$ $p\in(1,\infty)$ belongs to the generalized Morrey space $L^{p,\omega}({{\Bbb R}}^n)$ if the following norm is finite $$\label{PMS} \|f\|_{p,\omega}:=\left(\sup_{\E} \frac{1}{\omega(\E)} \int_{\E}|f(y)|^pdy \right)^{1/p}.$$ The space $L^{p,\omega}(\Omega)$ and the norm $\|f\|_{p,\omega;\Omega}$ are defined by taking $f\in L^p(\Omega)$ and $\E\cap \Omega$ instead of $\E$ in . For $\omega(x,r) = 1$ we get the Lebesgue space $L^p({{\Bbb R}}^n)$ and for $\omega(x,r)=r^\lambda,$ $\lambda\in(0,\alpha),$ $L^{p,\omega}({{\Bbb R}}^n)$ coincides with the Morrey space $L^{p,\lambda}({{\Bbb R}}^n)$ when ${{\Bbb R}}^n$ is endowed with the metric $\rho.$ However, there exist weight functions, as $\omega(x,r)=r^\lambda\ln(r+2),$ $\lambda\in (0,\alpha)$ for which $L^{p,\omega}({{\Bbb R}}^n)$ does not coincides with any Morrey space. For a given measurable function $f\in L^1_{\rm loc}({{\Bbb R}}^n)$ define the [*Hardy–Littlewood maximal operator*]{} $Mf$ and the [*sharp maximal operator*]{} $f^{\sharp}$ as $$Mf(x):=\sup_{x\ni\E}\frac{1}{|\E|}\int_{\E}|f(y)|dy, \quad f^{\sharp}(x):=\sup_{x\ni\E}\frac{1}{|\E|}\int_{\E}|f(y)-f_{\E}|dy$$ almost everywhere in ${{\Bbb R}}^n$ and the supremum is taken over all ellipsoids $\E$ centered at $x.$ Define also the operator $M_sf(x):=(M|f|^s(x))^{1/s}$ for $s\geq 1.$ The next results are weighted variants of the well-known maximal and sharp inequalities obtained in Lebesgue and Morrey spaces (see [@St], [@CF], [@DPR]). \[l1\] [(Maximal inequality)([@Na])]{} Assume that there are constants $C_1$ and $C_2$ such that for any $x_0\in {{\Bbb R}}^n$ and for any $r>0$ $$\begin{aligned} \label{Na1} r\leq t\leq 2r \Longrightarrow C_1\leq \frac{\omega(x_0,t)}{\omega(x_0,r)} \leq C_2,\\ \label{Na2} \int_r^\infty \frac{\omega(x_0,t)}{t^{\alpha+1}} dt\leq C\frac{\omega(x_0,r)}{r^\alpha}.\end{aligned}$$ For $1\leq s<p<\infty,$ there is a constant $C_{p,s}>0 $ such that for $f\in L^{p,\omega}({{\Bbb R}}^n)$ $$\|M_sf\|_{p,\omega}\leq C_{p,s} \|f\|_{p,\omega}.$$ \[l2\] [(Sharp inequality)]{} Let $0<\sigma\leq 1$ and $\E$ be an ellipsoid centered at $x_0\in {{\Bbb R}}^n$ of radius $r.$ Suppose that $\omega(x_0,r)$ satisfies and $$\int_r^\infty \frac{\omega(x_0,t)}{t^{\sigma \alpha+1}} dt\leq C\frac{\omega(x_0,r)}{r^{\sigma\alpha}}.$$ Then for $p\in(1,\infty)$ and $f\in L^{p,\omega}({{\Bbb R}}^n)$ there exists a constant $C$ independent of $f$ such that $$\|f\|_{p,\omega}\leq C \|f^{\sharp}\|_{p,\omega}.$$ Let $\chi_{\E}$ be the characteristic function of the ellipsoid and denote by $2\E$ an ellipsoid centered at $x_0$ and of radius $2r.$ It is easy to verify that $ M\chi_{\E}(x) \leq r^\alpha/(\rho(x-x_0)-r)^\alpha \leq 1,$ for all $x\in{{\Bbb R}}^n. $ Further, for any $x\in 2^{k+1}\E\setminus 2^k\E, $ $k=1,2,\ldots$ the maximal function of $\chi_{\E}$ could be estimated by $ r^\alpha / (2^{k+1}r-r)^\alpha \leq M\chi_{\E}(x) \leq r^\alpha / (2^{k}r-r)^\alpha $ which gives a reason to compare $M\chi_{\E}(x)$ with $2^{-k\alpha}$ for any $x$ as above. From the properties of the maximal function, that is $|f|\leq Mf$ and $Mf\leq f^\#$ (see [@GR p. 410]) follows $$\begin{aligned} J&=\int_{\E}|f(y)|^p dy=\int_{{{\Bbb R}}^{n}}|f(y)|^p \chi_{\E}(y)dy\\ &\leq\int_{{{\Bbb R}}^{n}}|Mf(y)|^p \big(M\chi_{\E}(y)\big)^{\sigma}dy \leq C\int_{{{\Bbb R}}^{n}}|f^{\#}(y)|^p \big(M\chi_{\E}(y)\big)^{\sigma}dy\\ &\leq C\Big\{\int_{2\E}|f^{\#}(y)|^pdy\\ &\qquad +\sum_{k=1}^{\infty}\int_{2^{k+1}\E \setminus 2^k\E } |f^{\#}(y)|^p\Big(\frac{r}{\rho(y-x_0)-r} \Big)^{\sigma \alpha}dy\Big\}\\ &\leq C\Big\{ \omega(2\E) \frac{1}{\omega(2\E)} \int_{2\E}|f^{\#}(y)|^pdy\\ &\qquad + r^{\sigma \alpha}\sum_{k=1}^\infty \frac{ \omega(2^{k+1}\E)}{(2^{k}r)^{\sigma \alpha}} \frac{1}{\omega(2^{k+1}\E)} \int_{2^{k+1}\E} |f^{\#}(y)|^p dy \Big\}\\ & \leq C r^{\sigma \alpha} \sum_{k=0}^\infty \frac{\omega(2^k\E)}{(2^{k}r)^{\sigma \alpha}} \|f^\#\|_{p,\omega}^p. \end{aligned}$$ From the properties of the function $\omega(x_0,t)$ follows $$\frac{\omega(2^k\E)}{(2^{k}r)^{\sigma \alpha}} \sim \int_{2^k r}^{2^{k+1}r} \frac{\omega(x_0,t)}{t^{\sigma \alpha+1}}dt.$$ Hence $$\int_\E |f(y)| dy\leq C r^{\sigma \alpha} \int_r^\infty \frac{\omega(x_0,t)}{t^{\sigma \alpha+1}} dt \|f\|^p_{p,\omega} \leq C\omega(x_0,r) \|f\|^p_{p,\omega}.$$ \[l3\] [(John-Nirenberg type lemma)]{} Let $a\in BMO$ and $p\in(1,\infty).$ Then for any ellipsoid $\E$ holds $$\left(\frac{1}{|\E|} \int_{\E}\left|a(y)-a_{\E}\right|^pdy \right)^{1/p}\leq C(p)\|a\|_\ast.$$ One more background we need is that for [*spherical harmonics*]{} and their properties (see for instance [@CZ], [@FR], [@CFL]). Recall that any homogeneous polynomial $P\colon\ {{\Bbb R}}^n\to{{\Bbb R}}$ of degree $m$ that satisfies $\Delta P(x)=0$ is called an [*$n$-dimensional solid harmonic of degree $m.$*]{} Its restriction to the unit sphere $\Sigma_n$ will be called an [*$n$-dimensional spherical harmonic of degree $m.$*]{} Denote by $\Upsilon_m$ the space of all $n$-dimensional spherical harmonics of degree $m.$ In general it results a finite-dimensional linear space with $g_m=\dim\Upsilon_m$ such that $g_0=1,$ $g_1=n$ and $$\label{*} g_m={m+n-1\choose n-1}-{m+n-3 \choose n-1} \leq C(n) m^{n-2}, \ m\geq 2.$$ Further, let $\{Y_{sm}(x)\}_{s=1}^{g_m}$ be an [*orthonormal base*]{} of $\Upsilon_m.$ Then $\{Y_{sm}\}_{s=1}^{g_m}{}_{m=0}^{\infty}$ is a [*complete orthonormal system*]{} in $L^2(\Sigma_n)$ and $$\label{Y1} \sup\limits_{x\in\Sigma_n} \left|D^\beta_x Y_{sm}(x) \right| \leq C(n) m^{|\beta|+(n-2)/2},\quad m=1,2,\ldots.$$ If, for instance, $\phi\in C^\infty(\Sigma_n)$ then $\sum_{s,m} b_{sm} Y_{sm}(x) $ is the Fourier series expansion of $\phi(x)$ with respect to $\{Y_{sm}(x)\}_{s,m}$ $\left(\sum_{s,m}\right.$ substitutes $\left.\sum_{m=0}^{\infty} \sum_{s=1}^{g_m}\right)$ and $$\label{Y3} b_{sm}= \int_{\Sigma_n}\phi ( y) Y_{sm}( y)d\sigma,\quad |b_{sm}|\leq C(n,l) m^{-2l}\sup\limits_{\overset{|\beta|=2l}{y\in\Sigma_n}} \left|D^\beta_y\phi(y) \right|$$ for any integer $l.$ In particular, the expansion of $\phi$ into spherical harmonics converges uniformly to $\phi.$ For the proof of the above results see [@CZ]. Singular integral estimates {#s3} =========================== Let $k(x;\xi)$ be a kernel in the sense of Definition \[d1\]. In order to ensure the existence of the operators and in $L^p({{\Bbb R}}^n)$ we restrict our considerations to functions $f\in L^p({{\Bbb R}}^n),$ $1<p<\infty$ for which the norm is finite. For the sake of convenience we still denote these spaces by $L^{p,\omega}({{\Bbb R}}^n).$ Having in mind this we define the operators $\K_\varepsilon f$ and $\C_\varepsilon [a,k] f$ for $a\in BMO$ and $f\in L^{p,\omega}({{\Bbb R}}^n)$ with $p\in(1,\infty)$ and $\omega$ satisfying and , by $$\begin{aligned} \K_\varepsilon f(x):=&\ \ds \int_{\rho(x-y)>\varepsilon} k(x;x-y)f(y)dy,\\ \C_\varepsilon [a,k]f(x):=&\ \ds \K_\varepsilon (af)(x) -a(x)\K_\varepsilon f(x)\\ \phantom{:}=&\ \ds \int_{\rho(x-y)>\varepsilon} k(x;x-y)[a(y)-a(x)]f(y)dy.\end{aligned}$$ We are going to prove that $\K_\varepsilon$ and $\C_\varepsilon [a,k]$ are bounded and continuous from $ L^{p,\omega}({{\Bbb R}}^n)$ into itself uniformly in $\varepsilon.$ This along with the properties of the kernel $k(x;\xi)$ will enable to let $\varepsilon \to0$ obtaining as limits in the $L^{p,\omega}({{\Bbb R}}^n)$-topology the [*singular*]{} integrals $$\begin{aligned} \K f(x):=&\ \ds P.V. \int_{{{\Bbb R}}^n} k(x;x-y)f(y)dy=\lim_{\varepsilon \to 0} \K_\varepsilon f(x)\\ \C[a,k]f(x):=&\ \ds P.V.\int_{{{\Bbb R}}^n} k(x;x-y)[a(y)-a(x)]f(y)dy= \lim_{\varepsilon \to 0} \C_\varepsilon[a,k]f(x).\end{aligned}$$ Moreover, we shall show that the last ones are also continuous in $L^{p,\omega}({{\Bbb R}}^n).$ Let us note assuming $f\in L^p({{\Bbb R}}^n),$ $p\in(1,\infty)$ Fabes-Rivière ([@FR]) show that $\K f$ exists in $L^p({{\Bbb R}}^n)$ for $p\in(1,\infty)$ as a limit of $\K_\varepsilon f$ when $\varepsilon\to 0$ in the $L^p$-norm. Moreover, the operator $\K\colon\ L^p({{\Bbb R}}^n)\to L^p({{\Bbb R}}^n)$ is continuous and this leads also to continuity in $L^p({{\Bbb R}}^n)$ of $\C[a,k]f$ if $a(x)$ is essentially bounded. As it concerns to the commutator we are going to derive a result similar to that of Coifman-Rochberg-Weiss ([@CRW Theorem 1]), which asserts: if $\K$ is Calderón-Zygmund operator in $L^p({{\Bbb R}}^n),$ $p\in(1,\infty)$ and $a\in BMO$ than the commutator $\C[a,\cdot]$ is a well defined linear continuous operator from $L^p({{\Bbb R}}^n)$ into itself. Later, this result has been extended by Bramanti-Cerutti ([@BC]) in the framework of homogeneous spaces. Based on this background about Calderón-Zygmund operators, we are going to obtain continuity in $L^{p,\omega}({{\Bbb R}}^n)$ and boundedness in terms of $\|a\|_\ast$ for the commutators having kernel of more general type. \[th1\] Let $k(x;\xi)$ be a variable kernel of mixed homogeneity, $f\in L^{p,\omega}({{\Bbb R}}^n),$ $p\in (1,\infty),$ $\omega$ satisfies and , and $a\in BMO.$ Then there exist the integrals $\K f,\ \C[a,k]f\in L^{p,\omega}({{\Bbb R}}^n)$ as limits of $\K_\varepsilon f$ and $\C_\varepsilon [a,k]f$ when $\varepsilon \to0$ with respect to the $L^{p,\omega}({{\Bbb R}}^n)$-norm. The operators $\K$ and $\C [a,k]$ are bounded from $ L^{p,\omega}({{\Bbb R}}^n)$ into itself and $$\|\K f\|_{p,\omega} \leq C\|f\|_{p,\omega},\qquad \|\C[a,k]f \|_{p,\omega} \leq C\|a\|_\ast \|f\|_{p,\omega}$$ where the constants depend on $n,p,\alpha$ and $k$ through the constant $C(\beta).$ Let $x,y\in {{\Bbb R}}^{n}$ and $\ol{y}={y}/{\rho(y)}\in \Sigma_n.$ From the properties of the kernel with respect to the second variable and the completeness of $\{Y_{sm}(x)\}_{s,m}$ in $L^2(\Sigma_n )$ it follows $$k(x;x-y)=\rho(x-y)^{-\alpha} k(x;\ol{x-y})=\rho(x-y)^{-\alpha}\sum_{s,m} b_{sm}(x) Y_{sm}(\ol{x-y}).$$ This way, the Definition \[d1\] $ii)$ and imply $$\label{Y4} \|b_{sm}\|_\infty \leq C(n,l,k)m^{-2l}$$ for any integer $l>1.$ Replacing the kernel with its expansion, we get $$\begin{aligned} \label{K_eps} \K_\varepsilon f(x)=&\ \ds \int_{\rho(x-y)>\varepsilon} \sum_{s,m}b_{sm}(x) \H_{sm}(x-y) f(y)dy,\\ \nonumber \C_\varepsilon [a,k]f(x) =&\ \ds \int_{\rho(x-y)>\varepsilon} \sum_{s,m}b_{sm}(x) \H_{sm}(x-y) [a(y) - a(x)]f(y)dy\end{aligned}$$ with $\H_{sm}(x-y)$ standing for $Y_{sm}(\ol{x-y}) \rho(x-y)^{-\alpha}.$ It is easy to check that $\H_{sm}(\cdot)$ is a constant kernel in the sense of Definition \[d1\] $i)$. Indeed, $i_a)$ and $i_b)$ are trivial while $i_c)$ follows from the fact that $Y_{sm}(x)$ is a harmonic homogeneous polynomial and the property of integral mean on sphere for harmonic functions (i.e. $Y_{sm}(0)=0$). In order to get series expansions of $\K_\varepsilon f$ and $\C_\varepsilon[a,k]f,$ we let $x\in {{\Bbb R}}^n$ and $y\in{{\Bbb R}}^n$ to be such that $\rho(x-y)>\varepsilon.$ Then , and yield $$\left|\sum_{m=1}^N \sum_{s=1}^{g_m} b_{sm}(x)\frac{Y_{sm}(\overline{x-y})}{ \rho(x-y)^\alpha} f(y)\right| \leq C(n) \frac{|f(y)|}{\rho(x-y)^\alpha} \sum_{m=1}^{\infty} m^{n-2+(n-2)/2-2l}$$ where $|f(\cdot)|\rho(x-\cdot)^{-\alpha}\in L^1({{\Bbb R}}^n)$ for a.a. $x\in {{\Bbb R}}^n$ and the integer $l$ is preliminary chosen greater than $(3n-4)/4.$ Similar inequality holds also for the commutator $\C_\varepsilon[a,k]f.$ Thus, by the dominated convergence theorem $$\begin{aligned} \label{EK} \K_\varepsilon f(x)= & \sum_{s,m}b_{sm}(x) \K_{sm,\varepsilon}(x),\\ \nonumber \C_\varepsilon [a,k]f(x) = & \sum_{s,m}b_{sm}(x) \C_{sm,\varepsilon}[a,k]f(x)\end{aligned}$$ with $$\begin{aligned} \K_{sm,\varepsilon}(x):= & \int_{\rho(x-y)>\varepsilon} \H_{sm}(x-y) f(y)dy,\\ \C_{sm,\varepsilon} [a,k]f(x):= & \int_{\rho(x-y)>\varepsilon} \H_{sm}(x-y) [a(y) - a(x)]f(y)dy. \end{aligned}$$ This way instead of the operators $\K f$ and $\C[a,k]f$ we shall study the existence and boundedness in $L^{p,\omega}({{\Bbb R}}^n)$ of the singular integrals $$\begin{aligned} \K_{sm}f(x):= & P.V. \ds\int_{{{\Bbb R}}^{n}}\H_{sm}(x-y) f(y)dy,\\ \C_{sm}[a,k]f(x):= & P.V \ds\int_{{{\Bbb R}}^{n}} \H_{sm}(x-y)[a(y)-a(x)] f(y) dy\end{aligned}$$ with constant kernels $\H_{sm}(\cdot).$ For what concern boundedness of $\K_{sm}$ in $L^p({{\Bbb R}}^n)$ we dispose of [@FR Theorem II.1] and this implies, through [@BC Theorem 2.5], boundedness in $L^p({{\Bbb R}}^n)$ of $\C_{sm}[a,k]$ as well. The cited results however require the kernel to have some “integral continuity”, called [*the Hörmander condition*]{}. It turns out that $\H_{sm}(\cdot)$ satisfies even stronger condition as shows the following lemma. \[l5\] [(Pointwise Hörmander condition)]{} Let $\E$ and $2\E$ be ellipsoids centered at $x_0$ and of radius $r$ and $2r,$ respectively. Then $$\label{PHE} \left|\H_{sm}(x-y)- \H_{sm}(x_0-y) \right| \leq C(n,\alpha) m^{n/2} \frac{\rho(x_0-x)}{\rho(x_0-y)^{\alpha+1}}$$ for each $x\in \E$ and $y\notin 2\E.$ We shall apply the mean value theorem to $\H_{sm}$ and therefore decay estimate for the gradient $\nabla \H_{sm}$ is needed. Let $x\in{{\Bbb R}}^n\setminus\{0\}$ be an arbitrary point. The implicit function theorem applied to the equation $F(x,\rho(x))=1$ gives an expression for the gradient $\nabla \rho(x)$ and straightforward calculations imply $$\begin{aligned} \ds \frac{\partial\H_{sm}}{\partial x_i}(x) = \ds \frac{1}{\rho(x)^{\alpha+\alpha_i}}&\Big( -\alpha Y_{sm}(\ol{x}) \frac{x_i}{\rho(x)^{\alpha_i} \sum_{j=1}^n \alpha_j x_j^2 \rho(x)^{-2\alpha_j}} \ds + \frac{\partial Y_{sm}}{\partial \ol{x}_i}(\ol{x})\\ \ds & - \sum_{k=1}^n \alpha_k \frac{\partial \ds Y_{sm}}{\partial \ol{x}_k}(\ol{x}) \frac{x_ix_k}{\rho(x)^{\alpha_i} \rho(x)^{\alpha_k} \sum_{j=1}^n \alpha_j x_j^2 \rho(x)^{-2\alpha_j}}\Big).\end{aligned}$$ Since $\ol{x}\in\Sigma_n$ and taking into account , $x_i /\rho(x)^{\alpha_i}\leq |\ol x|\leq 1$ and $\min\alpha_i \leq \sum_{j=1}^n \alpha_j x_j^2 / \rho(x)^{2\alpha_j}\leq \max\alpha_i,$ we get $$\label{Hx} \left|\frac{\partial\H_{sm}}{\partial x_i}(x)\right| \leq C(n,\alpha) \frac{m^{n/2}}{\rho(x)^{\alpha+\alpha_i}}\qquad \forall\ x\in {{\Bbb R}}^n\setminus\{0\}.$$ Now, applying the mean value theorem to the left-hand side of we get $$\label{gradH} \H_{sm}(x-y)- \H_{sm}(x_0-y) =\sum_{i=1}^n \frac{\partial \H_{sm}}{\partial x_i} (x_0-\xi) (x_0-x)_i$$ with $\xi=y-t(x-x_0)$ and $t\in(0,1).$ Obviously $\rho(y-\xi)=t\rho(x_0-x)\leq r$ which along with $y\not\in 2\E$ gives that $\xi$ does not belong to $\E$ and $\rho(x_0-\xi)\geq \frac{1}{2}\rho(x_0-y).$ Having in mind $(x_0-x)_i\leq \rho(x_0-x)^{\alpha_i},$ and we obtain $$\begin{aligned} &\left|\H_{sm}(x-y)- \H_{sm}(x_0-y) \right| \leq C(n,\alpha) m^{n/2} \sum_{i=1}^n \frac{\rho(x_0-x)^{\alpha_i}}{\rho(x_0-\xi)^{\alpha+\alpha_i}}\\ &\quad \leq C(n,\alpha) m^{n/2} \frac{\rho(x_0-x)}{\rho(x_0-y)^{\alpha+1}} \sum_{i=1}^n \frac{\rho(x_0-x)^{\alpha_i-1}}{\rho(x_0-\xi)^{\alpha_i-1}}\\ &\quad \leq C(n,\alpha) m^{n/2} \frac{\rho(x_0-x)}{\rho(x_0-y)^{\alpha+1}}\end{aligned}$$ where we have used that $\alpha_i\geq 1$ and $\rho(x_0-x)<\frac{1}{2}\rho(x_0-y)\leq \rho(x_0-\xi)$ from which follows immediately the last sum is no greater than $n.$ \[remarkHIC\]*This result ensures the kernel $\H_{sm}$ satisfies the Hörmander integral condition (see [@FR (1.1)]) $$\int_{\{y\in{{\Bbb R}}^n\colon \rho(y)\geq 4\rho(x)\}} \left|\H_{sm}(y-x)-\H_{sm}(y)\right|dy \leq C$$ with a constant independent of $x.$* In view of the cited above results there exist $\K_{sm}f,\ \C_{sm}[a,k]f\in L^p({{\Bbb R}}^n)$ such that $$\lim_{\varepsilon\to0} \left\| \K_{sm,\varepsilon}f-\K_{sm}f\right\|_{L^p({{\Bbb R}}^n)}= \lim_{\varepsilon\to0} \left\| \C_{sm,\varepsilon}[a,k]f-\C_{sm}[a,k]f\right\|_{L^p({{\Bbb R}}^n)}=0.$$ Our goal is to show that this convergence is fulfilled also with respect to the $L^{p,\omega}({{\Bbb R}}^n)$-norm. The proof is broken up into several Lemmas. \[l6\] The singular integrals $\K_{sm}f$ and their commutators $\C_{sm}[a,k]f$ satisfy $$\begin{aligned} \label{Ksharp} (\K_{sm}f)^{\sharp}(x)\leq & C m^{n/2} \big(M(|f|^p)(x)\big)^{1/p},\\ \nonumber (\C_{sm}[a,k]f)^{\sharp}(x)\leq & C \|a\|_\ast\Big\{\big(M(|\K_{sm}f|^p)(x)\big)^{1/p} + m^{n/{2}} \big(M(|f|^p)(x)\big)^{1/p} \Big\},\end{aligned}$$ where the constant depends on $n,$ $p$ and $\alpha$ but not on $f.$ For arbitrary $x_0\in {{\Bbb R}}^{n},$ set $\E$ for the ellipsoid $\E$ centered at $x_0$ and of radius $r.$ Let we consider the expression $$I:=\frac{1}{|\E|} \int_{\E} \left|\K_{sm}f(y) -(\K_{sm}f)_{\E}\right|dy.$$ Adding and extracting $\K_{sm,2r}f(x_0)$ to the function under the sign of the integral we obtain $$I \leq \frac{2}{|\E|} \int_{\E} \left|\K_{sm}f(y) -\K_{sm,2r}f(x_0)\right| dy :=2 I(x_0,\E).$$ Set $(2\E)^c={{\Bbb R}}^{n}\setminus 2\E$ and write $f=f\chi_{2\E}+f\chi_{(2\E)^c}=f_1+f_2$ with $\chi$ being the characteristic function of the respective set. Hence $$\begin{aligned} I(x_0,\E)\leq&\ \frac{1}{|\E|} \int_{\E} |\K_{sm}f_1(y)| dy\\ &\ + \frac{1}{|\E|} \int_{\E} |\K_{sm}f_2(y) -\K_{sm,2r}f(x_0)| dy=:I_1(x_0,\E)+I_2(x_0,\E).\end{aligned}$$ From the boundedness of $\K_{sm}$ in $L^p({{\Bbb R}}^n)$ ([@FR Theorem II.1]) follows $$\begin{aligned} I_1(x_0,\E)\leq&\ \frac{1}{|\E|}\left(\int_{\E}1 dy \right)^{1/p'}\left(\int_{\E} |\K_{sm}f_1(y)|^pdy \right)^{1/p}=\frac{1}{|\E|^{1/p}} \|\K_{sm}f_1\|_p\\ \leq&\ \frac{C(p,\alpha)}{|\E|^{1/p}} \|f_1\|_p \leq C(p,\alpha) \big(M(|f|^p)(x_0)\big)^{1/p}\end{aligned}$$ with $1/p'+1/p=1.$ About $I_2(x_0,\E),$ we have $$\begin{aligned} I_2(x_0,\E)\leq&\ \frac{1}{|\E|} \int_{\E}\left( \int_{(2\E)^c} \left| \H_{sm}(y-\xi) -\H_{sm}(x_0-\xi) \right| |f(\xi)| d\xi\right) dy\\ \leq&\ C(n,\alpha)m^{n/2} \frac{1}{|\E|} \int_{\E} \left(\ \int_{(2\E)^c} \frac{\rho(x_0-y)}{\rho(x_0-\xi)^{\alpha+1}} |f(\xi)| d\xi\right) dy\\ \leq&\ C(n,\alpha)m^{n/2} r \sum_{k=1}^{\infty} \int_{2^{k+1}\E\setminus 2^k\E} \frac{|f(\xi)|}{\rho(x_0-\xi)^{\alpha+1}} d\xi\\ \leq&\ C(n,\alpha) m^{n/2} \frac{1}{r^{\alpha}} \sum_{k=1}^{\infty} \frac{1}{2^{k(\alpha+1)}}|2^{k+1}\E| \left(\frac{1}{|2^{k+1}\E|} \int_{2^{k+1}\E} |f(\xi)|^pd\xi \right)^{1/p}\\ \leq&\ C(n,\alpha) m^{n/2} \big(M(|f|^p)(x_0)\big)^{1/p},\end{aligned}$$ after applying Lemma \[l5\] for $y\in\E$ and $\xi\in (2\E)^c.$ Taking $\sup_{\E}I(x_0,\E)$ and heaving in mind the arbitrarity of $x_0,$ we obtain for any $x\in {{\Bbb R}}^{n}.$ To estimate the sharp function of the commutator we shall employ the idea of Stromberg (see [@To]) which consists of expressing $\C_{sm}[a,k]f$ as a sum of integral operators and estimating their sharp functions. Precisely, $$\begin{aligned} &\C_{sm}[a,k]f(x)= \K_{sm}(a -a_{\E})f(x)-(a(x)-a_{\E})\K_{sm}f(x)\\ &\qquad = \K_{sm}(a -a_{\E})f_1(x) + \K_{sm}(a -a_{\E})f_2(x) -(a(x)-a_{\E})\K_{sm}f(x) \\ &\qquad =: J_1(x)+J_2(x) +J_3(x)\end{aligned}$$ where we have used the same truncation for the function $f$ as in $I(x_0,\E).$ Before proceed further, let us point out the obvious inequality $$\label{a1} |a_{2\E}-a_{\E}|\leq C(n,\alpha)\|a\|_\ast\qquad \forall\ a\in BMO({{\Bbb R}}^n)$$ and its by-product $$\label{a2} |a_{2^k\E}-a_{\E}|\leq C(n,\alpha)k\|a\|_\ast$$ following from by running induction. Now, for arbitrary $p\in(1,\infty)$ and $q\in (1,p),$ we have $$\begin{aligned} G_1(x_0,\E)&:= \frac{1}{|\E|}\int_{\E} |J_1(x)-(J_1)_{\E}| dx \leq\frac{2}{|\E|} \int_\E |\K_{sm}(a -a_{\E})f_1(x)|dx\\ \leq&\ \frac{2}{|\E|}\left(\ \int_{\E} |\K_{sm}(a -a_{\E})f_1(x)|^qdx \right)^{1/q} \left(\ \int_{\E} 1 dx \right)^{1/q'}\\ \leq&\ \frac{C(q,\alpha)}{|\E|^{1/q}} \left(\ \int_\E \big|(a(x)-a_\E)f_1(x)\big|^q dx \right)^{1/q}\\ \leq&\ \frac{C(q,\alpha)}{|\E|^{1/q}} \left(\ \int_{2\E}|f(x)|^pdx \right)^{1/p} \left(\ \int_{2\E}|a(x)-a_{\E}|^{pq/(p-q)} dx\right)^{(p-q)/pq}.\end{aligned}$$ Further, and Lemma \[l3\] applied to the second integral yield $$\begin{aligned} &\int_{2\E} |a(x)-a_\E|^{pq/(p-q)}dx \leq C(p,q)\left(\ \int_{2\E} |a(x)-a_{2\E}|^{pq/(p-q)}dx\right.\\ &\qquad\qquad \qquad\qquad\qquad \left.+ \int_{2\E} |a_{2\E}-a_\E|^{pq/(p-q)}dx\right)\\ &\quad\leq C(p,q)\left(|2\E|\frac{1}{|2\E|} \int_{2\E} |a(x)-a_{2\E}|^{pq/(p-q)}dx+ |2\E| C(n,\alpha) \|a\|_{\ast}^{pq/(p-q)}\right)\\ &\quad\leq C(n,p,q,\alpha)|2\E|\|a\|_{\ast}^{pq/(p-q)}. \end{aligned}$$ Therefore, $$G_1(x_0,\E)\leq C \|a\|_\ast \left(\frac{1}{|2\E|}\int_{2\E}|f(y)|^pdy \right)^{1/p}\leq C\|a\|_\ast \big(M(|f|^p)(x_0)\big)^{1/p}.$$ To estimate the sharp function of $J_2(x),$ we proceed analogously as we already did for $I_2(x_0,\E).$ Precisely, $$G_2(x_0,\E):=\frac{1}{|\E|} \int_{\E} |J_2(x)-(J_2)_{\E}|dx \leq \frac{2}{|\E|}\int_{\E} |J_2(x)-J_2(x_0)| dx$$ and the integrand satisfies $$\begin{aligned} &|J_2(x)-J_2(x_0)| \leq \int_{(2\E)^c}\left| \H_{sm}(x-y)- \H_{sm}(x_0-y)\right| |a(y)-a_{\E}|\, |f(y)| dy\\ &\ \leq C(n,\alpha) m^{n/2}\rho(x_0-x) \int_{(2\E)^c} \frac{|a(y)-a_{\E}| |f(y)|}{\rho(x_0-y)^{\alpha+1}}dy\\ &\ \leq C(n,\alpha) m^{n/2} r \left(\ \int_{(2\E)^c} \frac{|f(y)|^p}{\rho(x_0-y)^{\alpha+1}}dy\right)^{1/p} \left(\ \int_{(2\E)^c} \frac{|a(y)-a_{\E}|^{p'}}{\rho(x_0-y)^{\alpha+1}}dy\right)^{1/p'}\end{aligned}$$ where $1/p+1/p'=1$ and we have applied the Hörmander pointwise estimate (Lemma \[l5\]). Later on, $$\int_{(2\E)^c} \frac{|f(y)|^p}{\rho(x_0-y)^{\alpha+1}}dy = \sum_{k=1}^{\infty} \int_{2^{k+1}\E\setminus 2^k\E} \frac{|f(y)|^p}{\rho(x_0-y)^{\alpha+1}}dy \leq\frac{2^{\alpha+1}}{r}M(|f|^p)(x_0),$$ while and Lemma \[l3\] imply $$\begin{aligned} &\int_{(2\E)^c} \frac{|a(y)-a_{\E}|^{p'}}{\rho(x_0-y)^{\alpha+1}}dy= \sum_{k=1}^{\infty} \int_{2^{k+1}\E\setminus 2^k \E} \frac{|a(y)-a_{\E}|^{p'}}{\rho(x_0-y)^{\alpha+1}}dy\\ &\quad\leq \sum_{k=1}^{\infty} \frac{1}{(2^k r)^{\alpha+1}} \int_{2^{k+1}\E}|a(y)-a_{\E}|^{p'}dy\\ &\quad \leq \sum_{k=1}^\infty \frac{2^{p'-1}}{(2^kr)^{\alpha+1}} \int_{2^{k+1}\E}\left(|a(y)-a_{2^{k+1}\E}|^{p'}+ |a_{2^{k+1}\E}-a_\E|^{p'} \right) dy\\ &\quad \leq C\sum_{k=1}^\infty \frac{|2^{k+1}\E|}{(2^kr)^{\alpha+1}} (1+k^{p'})\|a\|_\ast^{p'} \leq C \frac{\|a\|_\ast^{p'}}{r}\end{aligned}$$ and the constant depends on $n,$ $p$ and $\alpha.$ Hence $$G_2(x_0,\E)\leq C(n,p,\alpha) m^{n/2} \|a\|_\ast \big(M(|f|^p)(x_0)\big)^{1/p}.$$ Finally, $$\begin{aligned} G_3(x_0,\E):=&\ \frac{1}{|\E|} \int_{\E} |J_3(x)-(J_3)_{\E}| dx \leq \frac{2}{|\E|} \int_{\E} |a(x)-a_{\E}| |\K_{sm}f(x)|dx\\ \leq&\ 2\left(\ \frac{1}{|\E|}\int_{\E} |a(x)-a_{\E}|^{p'}dx\right)^{1/p'} \left(\ \frac{1}{|\E|}\int_{\E} |\K_{sm}f(x)|^{p}dx\right)^{1/p}\\ \leq&\ C(p) \|a\|_\ast \big(M(|\K_{sm}f|^p)(x_0)\big)^{1/p}.\end{aligned}$$ Summing up $G_1(x_0,\E),$ $G_2(x_0,\E)$ and $G_3(x_0,\E)$ and taking the supremum with respect to $\E$ and rendering in account the arbitrarity of the point $x_0$ we get the desired estimate for the commutator. \[l6a\] The operators $\K_{sm}$ and $\C_{sm}[a,k]$ are continuous acting from $ L^{p,\omega}({{\Bbb R}}^n)$ into itself and $$\label{Kpl} \|\K_{sm}f\|_{p,\omega} \leq C m^{n/2}\|f\|_{p,\omega},\ \|\C_{sm}[a,k]f\|_{p,\omega} \leq C m^{n/2}\|a\|_\ast\|f\|_{p,\omega}$$ with constants depending on $n,$ $p,$ and $\alpha.$ First of all we shall estimate the $L^{p,\omega}$-norms of the sharp functions of the considered operators. Since the expression for $(\K_{sm}f)^\sharp$ in holds true for any $q\in(1,p)$ as well, the maximal inequality (Lemma \[l1\]) with $s=1$ asserts $$\begin{aligned} \int_\E |(\K_{sm}f)^{\sharp}(x)|^p dx & \leq Cm^{pn/2} \int_\E |M(|f|^q)(x)|^{p/q}dx\\ & \leq Cm^{pn/2} \omega(\E) \|M(|f|^q)\|^{p/q}_{p/q,\omega}\\ &\leq Cm^{pn/2} \omega(\E) \| |f|^q \|^{p/q}_{p/q,\omega} \leq Cm^{pn/2} \omega(\E) \| f \|^{p}_{p,\omega}.\end{aligned}$$ Dividing of $\omega(\E)$ and taking $\sup_{\E},$ we arrive at $$\|(\K_{sm}f)^{\sharp} \|_{p,\omega}\leq C m^{n/2} \|f\|_{p,\omega}$$ which implies the first inequality in through Lemma \[l2\]. The $L^{p,\omega}$-estimate for the commutator follows in the same manner. \[c1\] The operators $\K_{sm,\varepsilon}$ and $\C_{sm,\varepsilon}[a,k]$ are continuous acting from $L^{p,\omega}({{\Bbb R}}^n)$ into itself and satisfy $$\label{Kpl1} \|\K_{sm,\varepsilon}f\|_{p,\omega} \leq C m^{n/2}\|f\|_{p,\omega},\ \|\C_{sm,\varepsilon}[a,k]f\|_{p,\omega} \leq C m^{n/2}\|a\|_\ast\|f\|_{p,\omega}$$ with constants depending on $n,$ $p$ and $\alpha.$ Let $\E_\varepsilon$ and $\E_{\varepsilon/2}$ be ellipsoids centered at $x\in{{\Bbb R}}^n$ and of radius $\varepsilon$ and $\varepsilon/2,$ respectively. Writing $f=f\chi_{\E_\varepsilon}+f\chi_{(\E_\varepsilon)^c}=f_1+f_2$ we obtain $$\begin{aligned} &\K_{sm,\varepsilon} f(x)\leq \frac{1}{|\E_{\varepsilon/2}|}\int_{\E_{\varepsilon/2}} |\K_{sm,\varepsilon}f(x)|dy \leq \frac{1}{|\E_{\varepsilon/2}|} \int_{\E_{\varepsilon/2}}|\K_{sm}f(y)|dy\\ &\qquad +\frac{1}{|\E_{\varepsilon/2}|}\int_{\E_{\varepsilon/2}} |\K_{sm,\varepsilon}f(x)-\K_{sm} f(y)|dy\\ &\leq \frac{2}{|\E_{\varepsilon/2}|}\int_{\E_{\varepsilon/2}} |\K_{sm}f_1(y)|dy + \frac{1}{|\E_{\varepsilon/2}|}\int_{\E_{\varepsilon/2}} |\K_{sm,\varepsilon}f(x)-\K_{sm} f_2(y)| dy\\ & := 2I_1(x,\E_{\varepsilon/2})+I_2(x,\E_{\varepsilon/2})\end{aligned}$$ where $I_1$ and $I_2$ stand for the terms introduced at the proof of Lemma \[l6\], and the same arguments as therein lead to $$|\K_{sm,\varepsilon}f(x)|\leq C(n,p,\alpha) m^{n/2}\left( M(|f|^q)(x) \right)^{1/q}$$ for any $q\in(1,\infty).$ It remains to take the $L^{p,\omega}$-norms of the both sides above for $1<q<p$ and to apply Lemma \[l1\] in order to get . The commutator estimate follows analogously. Returning to the series expansions , we are in a position now to complete the proof of Theorem \[th1\]. First of all, note that $$\begin{aligned} \sum_{m=1}^{\infty} \sum_{s=1}^{g_m} \|b_{sm}\K_{sm,\varepsilon}f\|_{p,\omega} \leq&\ C(n,p,\alpha,k) \|f\|_{p,\omega} \sum_{m=1}^{\infty} m^{-2l+n-2+n/2},\\ \sum_{m=1}^{\infty} \sum_{s=1}^{g_m} \|b_{sm}\C_{sm,\varepsilon}[a,k]f\|_{p,\omega} \leq&\ C(n,p,\alpha,k) \|a\|_\ast \|f\|_{p,\omega} \sum_{m=1}^{\infty} m^{-2l+n-2+n/2} \end{aligned}$$ as it follows from , and Lemma \[c1\]. Choosing $l>(3n-2)/4$ the series in result totally convergent in $L^{p,\omega}({{\Bbb R}}^n),$ uniformly in $\varepsilon>0,$ whence $$\|\K_{\varepsilon}f\|_{p,\omega} \leq C \|f\|_{p,\omega},\quad \|\C_{\varepsilon}[a,k]f\|_{p,\omega} \leq C \|a\|_\ast\|f\|_{p,\omega}.$$ Setting $$\K f(x) := \sum_{s,m} b_{sm}(x)\K_{sm}f(x),\qquad \C[a,k]f(x) := \sum_{s,m} b_{sm}(x)\C_{sm}[a,k]f(x),$$ we obtain as above $$\|\K f\|_{p,\omega} \leq C \|f\|_{p,\omega},\qquad \|\C[a,k]f\|_{p,\omega} \leq C \|a\|_\ast\|f\|_{p,\omega}$$ through , and Lemma \[l6a\]. Finally, the total convergence in $L^{p,\omega}({{\Bbb R}}^n)$ of the series expansions , uniformly in $\varepsilon>0,$ gives $$\begin{aligned} &\lim_{\varepsilon\to0} \K_\varepsilon f(x) = \sum_{s,m} b_{sm}(x) \lim_{\varepsilon\to0} \K_{sm,\varepsilon}f(x)= \sum_{s,m} b_{sm}(x) \K_{sm}f(x)=\K f(x),\\ &\lim_{\varepsilon\to0} \C_\varepsilon [a,k]f(x) = \sum_{s,m} b_{sm }(x) \lim_{\varepsilon\to0} \C_{sm,\varepsilon}[a,k]f(x)=\C[a,k]f(x)\end{aligned}$$ and this completes the proof of Theorem \[th1\]. It is worth noting that singular integrals like and appear in the representation formulas for the solutions of linear elliptic and parabolic partial differential equations. To make the obtained here results applicable to the study of regularizing properties of these operators we need of some additional local results. \[crl1\] Let $\Omega$ be a bounded domain in ${{\Bbb R}}^n$ and $k(x;\xi)\colon\ \Omega\times ({{\Bbb R}}^n\setminus\{0\})\to{{\Bbb R}}$ be a variable kernel of mixed homogeneity, $a\in BMO(\Omega),$ $p\in (1,\infty)$ and $\omega$ satisfies and . Then, for any $f\in L^{p,\omega}(\Omega)$ and almost all $x\in\Omega,$ the singular integrals $$\begin{aligned} \K f(x)= & P.V. \int_{\Omega} k(x;x-y)f(y)dy\\ \C[a,k]f(x)= & P.V. \int_{\Omega} k(x;x-y)[a(y)-a(x)]f(y)dy\end{aligned}$$ are well defined in $L^{p,\omega}(\Omega)$ and $$\|\K f\|_{p,\omega;\Omega} \leq C\|f\|_{p,\omega;\Omega},\quad \|\C[a,k]f \|_{p,\omega;\Omega} \leq C\|a\|_\ast \|f\|_{p,\omega;\Omega}$$ with $C=C(n,p,\alpha,\Omega,k).$ To obtain the above assertion it is sufficient to extend $k(x;\cdot)$ and $f(\cdot)$ as zero outside $\Omega.$ One more necessary extension preserving the norm is that of $a$ in $BMO({{\Bbb R}}^n)$ and we have it according to the results of Jones [@PJ] and Acquistapace [@A] (see [@CFL] for details). Another consequence of Theorem \[th1\] is the “good behavior” of the commutator for $VMO$ functions $a.$ \[crl2\] Suppose $a\in VMO$ with $VMO$-modulus $\gamma_a.$ Then, for each $\varepsilon>0$ there exists $r_0=r_0(\varepsilon,\gamma_a)>0$ such that for any $\varrho\in(0,r_0)$ and any ellipsoid $\E_\varrho$ of radius $\varrho$ one has $$\label{vmo-omega} \|\C[a,k]f \|_{p,\omega;\E_\varrho} \leq C \varepsilon \|f\|_{p,\omega;\E_\varrho}$$ for all $f\in L^{p,\omega}(\E_\varrho).$ From the properties of the $VMO$ functions [@S Theorem 1] it follows that for any $\varepsilon>0$ there exists $r_0=r_0(\varepsilon,\gamma_a)$ and continuous and uniformly bounded function $g$ with modulus of continuity $\omega_g(r_0)< \varepsilon/2$ such that $\|a-g\|_\ast<\varepsilon/2.$ Let $\E_\varrho$ be an ellipsoid centered at $x_0$ and of radius $\varrho<r_0.$ Following [@CFL] we construct a function $$h(x)=\begin{cases} g(x) & x\in\E_\varrho\\ g\left({x_0}_1+\varrho^{\alpha_1}\frac{x_1-{x_0}_1}{\rho(x-x_0)^{\alpha_1}},\ldots, {x_0}_n+\varrho^{\alpha_n}\frac{x_n-{x_0}_n}{\rho(x-x_0)^{\alpha_n}}\right) &x\in\E_\varrho^c \end{cases}$$ which is uniformly continuous in ${{\Bbb R}}^n.$ Whence the oscillation of $h$ in ${{\Bbb R}}^n$ is no greater than the oscillation of $g$ in $\E_{r_0}.$ Then $$\begin{aligned} \|\C[a,k]f \|_{p,\omega;\E_\varrho} & \leq \|\C[a-g,k]f \|_{p,\omega;\E_\varrho}+ \|\C[g,k]f\|_{p,\omega;\E_\varrho}\\ & \leq C \|a-g\|_\ast \|f\|_{p,\omega;\E_\varrho} + C\|h\|_\ast\|f\|_{p,\omega;\E_\varrho}\\ & \leq C\left(\|a-g\|_\ast +\omega_g(r_0) \right)\|f\|_{p,\omega;\E_\varrho} < C\varepsilon \|f\|_{p,\omega;\E_\varrho}.\end{aligned}$$ [MPSS]{} P. Acquistapace, [*On $BMO$ regularity for linear elliptic systems*]{}, Ann. Mat. Pura Appl. (4) [**161**]{} (1992), 231–269. [**286**]{} (1978), A139–A142. M. Bramanti, M.C. Cerutti, [*Commutators of singular integrals on homogeneous spaces,*]{} Boll. Un. Mat. Ital. B (7) [**10**]{} (1996), 843–883. A.P. Calderón, A. Zygmund, [*Singular integral operators and differential equations,*]{} Amer. J. Math. [**79**]{} (1957), 901–921. F. Chiarenza, M. Frasca, [*Morrey spaces and Hardy-Littlewood maximal functions,*]{} Rend. Mat. Appl. (7) [**7**]{} (1987), 273–279. F. Chiarenza, M. Frasca, P. Longo, [*Interior $W^{2,p}$ estimates for nondivergence elliptic equations with discontinuous coefficients*]{}, Ricerche Mat. [**40**]{} (1991), 149–168. R. Coifman, R. Rochberg, G. Weiss, [*Factorization theorems for Hardy spaces in several variables,*]{} [Ann. of Math. (2)]{} [**103**]{} (1976), 611–635. G. di Fazio, D.K. Palagachev, M.A. Ragusa, [*Global Morrey regularity of strong solutions to Dirichlet problem for elliptic equations with discontinuous coefficients*]{}, J. Funct. Anal. [**166**]{} (1999), 179–196. E.B. Fabes, N. Rivière, [*Singular integrals with mixed homogeneity*]{}, Studia Math. [**27**]{} (1966), 19–38. J. Garcia-Cuerva, J.L. Rubio De Francia, Weighted Norm Inequalities and Related Topics, [*North-Holand Math. Studies,*]{} Vol. 116, North-Holand, Amsterdam, 1985. F. John, L. Nirenberg, [*On functions of bounded mean oscillation,*]{} [Comm. Pure Appl. Math.]{} [**14**]{} (1961), 415–426. B.F. Jones, [*On a class of singular integrals,*]{} Amer. Jour. Math. [**86**]{} (1964), 441–462. P.W. Jones, [*Extension theorems for $BMO$*]{}, Indiana Univ. Math. J. [**29**]{} (1980), 41–66. E. Nakai, [*Hardy-Littlewood Maximal Operator, Singular Integral Operators and the Riesz Potentials on Generalized Morrey Spaces,*]{} Math.Nachr. [**166**]{} (1994) 95-103. D. Sarason, [*Functions of vanishing mean oscillation*]{}, Trans. Amer. Math. Soc. [**207**]{} (1975), 391–405. E.M. Stein, Singular Integrals and Differentiability Properties of Functions, [*Princeton University Press*]{}, Princeton, New Jersey, 1970. A. Torchinsky, Real-Variable Methods in Harmonic Analysis, Pure Appl. Math. [**123**]{}, Academic Press, Orlando, FL, 1986.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The paper presents a general theory of coupling of eigenvalues of complex matrices of arbitrary dimension depending on real parameters. The cases of weak and strong coupling are distinguished and their geometric interpretation in two and three-dimensional spaces is given. General asymptotic formulae for eigenvalue surfaces near diabolic and exceptional points are presented demonstrating crossing and avoided crossing scenarios. Two physical examples illustrate effectiveness and accuracy of the presented theory.' author: - 'A. A. Mailybaev, O. N. Kirillov and A. P. Seyranian' date: | Institute of Mechanics, Moscow State Lomonosov University,\ Michurinskii pr. 1, 119192 Moscow, Russia\ E-mail: mailybaev{kirillov,seyran}@imec.msu.ru title: | Coupling of eigenvalues of complex matrices\ at diabolic and exceptional points --- PACS numbers: 02.10.Yn, 02.30.Oz, 42.25.Bs Introduction ============ Behavior of eigenvalues of matrices dependent on parameters is a problem of general interest having many important applications in natural and engineering sciences. Probably, [@Hamilton1] was the first who revealed an interesting physical effect associated with coincident eigenvalues known as conical refraction, see also [@Berry1999]. In modern physics, e.g. quantum mechanics, crystal optics, physical chemistry, acoustics and mechanics, singular points of matrix spectra associated with specific effects attract great interest of researchers since the papers [@Neumann; @Herring; @Teller]. These are the points where matrices possess multiple eigenvalues. In applications the case of double eigenvalues is the most important. With a change of parameters coupling and decoupling of eigenvalues with crossing and avoided crossing scenario occur. The crossing of eigenvalue surfaces (energy levels) is connected with the topic of geometrical phase, see [@Berry1]. In recent papers, see e.g. [@Heiss2000; @Dembowsky; @Berry2; @Dembowsky2003; @KKM; @KKM1; @Heiss; @Stehmann], two important cases are distinguished: the diabolic points (DPs) and the exceptional points (EPs). From mathematical point of view DP is a point where the eigenvalues coalesce (become double), while corresponding eigenvectors remain different (linearly independent); and EP is a point where both eigenvalues and eigenvectors merge forming a Jordan block. Both the DP and EP cases are interesting in applications and were observed in experiments, see e.g. [@Dembowsky; @Dembowsky2003; @Stehmann]. In early studies only real and Hermitian matrices were considered while modern physical systems require study of complex symmetric and non-symmetric matrices, see [@MH; @Berry2; @KKM]. Note that most of the cited papers dealt with specific $2\times2$ matrices depending on two or three parameters. Of course, in the vicinity of an EP (and also DP) the $m$-dimensional matrix problem becomes effectively two-dimensional, but finding the corresponding two-dimensional space for a general $m$-dimensional matrix family is a nontrivial problem [@Arnold2]. In this paper we present a general theory of coupling of eigenvalues of complex matrices of arbitrary dimension smoothly depending on multiple real parameters. Two essential cases of weak and strong coupling based on a Jordan form of the system matrix are distinguished. These two cases correspond to diabolic and exceptional points, respectively. We derive general formulae describing coupling and decoupling of eigenvalues, crossing and avoided crossing of eigenvalue surfaces. We present typical (generic) pictures showing movement of eigenvalues, the eigenvalue surfaces and their cross-sections. It is emphasized that the presented theory of coupling of eigenvalues of complex matrices gives not only qualitative, but also quantitative results on behavior of eigenvalues based only on the information taken at the singular points. Two examples on propagation of light in a homogeneous non-magnetic crystal possessing natural optical activity (chirality) and dichroism (absorption) in addition to biaxial birefringence illustrate basic ideas and effectiveness of the developed theory. The presented theory is based on previous research on interaction of eigenvalues of real matrices depending on multiple parameters with mechanical applications. In [@Seyranian1991; @Seyranian1993] the important notion of weak and strong coupling (interaction) was introduced for the first time. In the papers [@SP1993; @SLO1994; @MS1999; @SKliem2001; @SM2001; @KS2002; @SM2003; @K2004; @KS2004], and in the recent book [@SeyMai2004] significant mechanical effects related to diabolic and exceptional points were studied. These include transference of instability between eigenvalue branches, bimodal solutions in optimal structures under stability constraints, flutter and divergence instabilities in undamped nonconservative systems, effect of gyroscopic stabilization, destabilization of a nonconservative system by infinitely small damping, which were described and explained from the point of view of coupling of eigenvalues. An interesting application of the results on eigenvalue coupling to electrical engineering problems is given in [@Dobson]. The paper is organized as follows. In Section 2 we present general results on weak and strong coupling of eigenvalues of complex matrices depending on parameters. These two cases correspond to the study of eigenvalue behavior near diabolic and exceptional points. Section 3 is devoted to crossing and avoided crossing of eigenvalue surfaces near double eigenvalues with one and two eigenvectors. Two physical examples are presented in Section 4, and finally we end up with the conclusion in Section 5. Coupling of eigenvalues ======================= Let us consider the eigenvalue problem $$\mathbf{A}\mathbf{u} = \lambda\mathbf{u} \label{eq1.1}$$ for a general $m\times m$ complex matrix $\mathbf{A}$ smoothly depending on a vector of $n$ real parameters $\mathbf{p} = (p_1,\ldots,p_n)$. Assume that, at $\mathbf{p} = \mathbf{p}_0$, the eigenvalue coupling occurs, i.e., the matrix $\mathbf{A}_0 = \mathbf{A}(\mathbf{p}_0)$ has an eigenvalue $\lambda_0$ of multiplicity $2$ as a root of the characteristic equation $\det(\mathbf{A}_0-\lambda_0\mathbf{I}) = 0$; $\mathbf{I}$ is the identity matrix. This double eigenvalue can have one or two linearly independent eigenvectors $\mathbf{u}$, which determine the geometric multiplicity. The eigenvalue problem adjoint to (\[eq1.1\]) is $$\mathbf{A}^*\mathbf{v} = \eta\mathbf{v}, \label{eq1.1b}$$ where $\mathbf{A}^* = \overline{\mathbf{A}}^T$ is the adjoint matrix operator (Hermitian transpose), see e.g. [@Lancaster]. The eigenvalues $\lambda$ and $\eta$ of problems (\[eq1.1\]) and (\[eq1.1b\]) are complex conjugate: $\eta = \overline{\lambda}$. Double eigenvalues appear at sets in parameter space, whose codimensions depend on the matrix type and the degeneracy (EP or DP). In Table \[tab1\], we list these codimensions based on the results of the singularity theory [@Neumann; @Arnold2]. In this paper we analyze general (nonsymmetric) complex matrices. The EP degeneracy is the most typical for this type of matrices. In comparison with EP, the DP degeneracy is a rare phenomenon in systems described by general complex matrices. However, some nongeneric situations may be interesting from the physical point of view. As an example, we mention complex non-Hermitian perturbations of symmetric two-parameter real matrices, when the eigenvalue surfaces have coffee-filter singularity, see [@MH; @Berry2; @KKM]. A general theory of this phenomenon will be given in our companion paper [@KMS]. matrix type codimension of DP codimension of EP ---------------------- ------------------- ------------------- real symmetric 2 non-existent real nonsymmetric 3 1 Hermitian 3 non-existent complex symmetric 4 2 complex nonsymmetric 6 2 : Codimensions of eigenvalue degeneracies.[]{data-label="tab1"} Let us consider a smooth perturbation of parameters in the form $\mathbf{p} = \mathbf{p}(\varepsilon)$, where $\mathbf{p}(0) = \mathbf{p}_0$ and $\varepsilon$ is a small real number. For the perturbed matrix $\mathbf{A} = \mathbf{A}(\mathbf{p}(\varepsilon))$, we have $$\begin{array}{c} \mathbf{A} = \mathbf{A}_0+\varepsilon\mathbf{A}_1 +\frac12\varepsilon^2\mathbf{A}_2+o(\varepsilon^2),\\[12pt] \displaystyle \mathbf{A}_0 = \mathbf{A}(\mathbf{p}_0),\quad \mathbf{A}_1 = \sum_{i = 1}^n \frac{\partial\mathbf{A}}{\partial p_i} \frac{dp_i}{d\varepsilon},\quad \mathbf{A}_2 = \sum_{i = 1}^n \frac{\partial\mathbf{A}}{\partial p_i} \frac{d^2p_i}{d\varepsilon^2} +\sum_{i,j = 1}^n \frac{\partial^2\mathbf{A}}{\partial p_i\partial p_j} \frac{dp_i}{d\varepsilon}\frac{dp_j}{d\varepsilon}. \end{array} \label{eq1.2}$$ The double eigenvalue $\lambda_0$ generally splits into a pair of simple eigenvalues under the perturbation. Asymptotic formulae for these eigenvalues and corresponding eigenvectors contain integer or fractional powers of $\varepsilon$ [@Vishik]. Weak coupling of eigenvalues ---------------------------- Let us consider the coupling of eigenvalues in the case of $\lambda_0$ with two linearly independent eigenvectors $\mathbf{u}_1$ and $\mathbf{u}_2$. This coupling point is known as a diabolic point. Let us denote by $\mathbf{v}_1$ and $\mathbf{v}_2$ two eigenvectors of the complex conjugate eigenvalue $\eta = \overline{\lambda}$ for the adjoint eigenvalue problem (\[eq1.1b\]) satisfying the normalization conditions $$(\mathbf{u}_1,\mathbf{v}_1) = (\mathbf{u}_2,\mathbf{v}_2) = 1,\quad (\mathbf{u}_1,\mathbf{v}_2) = (\mathbf{u}_2,\mathbf{v}_1) = 0, \label{eq1.NC}$$ where $(\mathbf{u},\mathbf{v}) = \sum_{i = 1}^n u_i\overline{v}_i$ denotes the Hermitian inner product. Conditions (\[eq1.NC\]) define the unique vectors $\mathbf{v}_1$ and $\mathbf{v}_2$ for given $\mathbf{u}_1$ and $\mathbf{u}_2$ [@SeyMai2004]. For nonzero small $\varepsilon$, the two eigenvalues $\lambda_+$ and $\lambda_-$ resulting from the bifurcation of $\lambda_0$ and the corresponding eigenvectors $\mathbf{u}_\pm$ are given by $$\lambda_\pm = \lambda_0+\mu_\pm\varepsilon+o(\varepsilon),\quad \mathbf{u}_\pm = \alpha_\pm\mathbf{u}_1+\beta_\pm\mathbf{u}_2+o(1). \label{eq1.3}$$ The coefficients $\mu_\pm$, $\alpha_\pm$, and $\beta_\pm$ are found from the $2\times 2$ eigenvalue problem\ (see e.g. [@SeyMai2004]) $$\left(\begin{array}{cc} (\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_1) & (\mathbf{A}_1\mathbf{u}_2,\mathbf{v}_1) \\[3pt] (\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_2) & (\mathbf{A}_1\mathbf{u}_2,\mathbf{v}_2) \end{array}\right) \left(\begin{array}{cc} \alpha_\pm \\[2pt] \beta_\pm \end{array}\right) = \mu_\pm \left(\begin{array}{cc} \alpha_\pm \\[2pt] \beta_\pm \end{array}\right). \label{eq1.4}$$ Solving the characteristic equation for (\[eq1.4\]), we find $$\mu_\pm = \frac{(\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_1) +(\mathbf{A}_1\mathbf{u}_2,\mathbf{v}_2)}{2}\pm \sqrt{\frac{((\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_1) -(\mathbf{A}_1\mathbf{u}_2,\mathbf{v}_2))^2}{4} +(\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_2) (\mathbf{A}_1\mathbf{u}_2,\mathbf{v}_1)}. \label{eq1.4b}$$ We note that for Hermitian matrices $\mathbf{A}$ one can take $\mathbf{v}_1 = \mathbf{u}_1$ and $\mathbf{v}_2 = \mathbf{u}_2$ in (\[eq1.4\]), where the eigenvectors $\mathbf{u}_1$ and $\mathbf{u}_2$ are chosen satisfying the conditions $(\mathbf{u}_1,\mathbf{u}_1) = (\mathbf{u}_2,\mathbf{u}_2) = 1$ and $(\mathbf{u}_1,\mathbf{u}_2) = 0$, and obtain the well-known formula, see [@Hilbert]. As the parameter vector passes the coupling point $\mathbf{p}_0$ along the curve $\mathbf{p}(\varepsilon)$ in parameter space, the eigenvalues $\lambda_+$ and $\lambda_-$ change smoothly and cross each other at $\lambda_0$, see Figure \[fig1\]a. At the same time, the corresponding eigenvectors $\mathbf{u}_+$ and $\mathbf{u}_-$ remain different (linearly independent) at all values of $\varepsilon$ including the point $\mathbf{p}_0$. We call this interaction *weak coupling*. By means of eigenvectors, the eigenvalues $\lambda_\pm$ are well distinguished during the weak coupling. We emphasize that despite the eigenvalues $\lambda_\pm$ and the eigenvectors $\mathbf{u}_\pm$ depend smoothly on a single parameter $\varepsilon$, they are nondifferentiable functions of multiple parameters at $\mathbf{p}_0$ in the sense of Frechét [@AnalysisTextbook]. ![Eigenvalue coupling: (a) weak, (b) strong.[]{data-label="fig1"}](fig1.eps){width="95.00000%"} Strong coupling of eigenvalues ------------------------------ Let us consider coupling of eigenvalues at $\mathbf{p}_0$ with a double eigenvalue $\lambda_0$ possessing a single eigenvector $\mathbf{u}_0$. This case corresponds to the exceptional point. The second vector of the invariant subspace corresponding to $\lambda_0$ is called an associated vector $\mathbf{u}_1$ (also called a generalized eigenvector [@Lancaster]); it is determined by the equation $$\mathbf{A}_0\mathbf{u}_1 = \lambda_0\mathbf{u}_1+\mathbf{u}_0. \label{eq1.5}$$ An eigenvector $\mathbf{v}_0$ and an associated vector $\mathbf{v}_1$ of the matrix $\mathbf{A}^*$ are determined by $$\mathbf{A}_0^*\mathbf{v}_0 = \overline{\lambda}_0\mathbf{v}_0,\quad \mathbf{A}_0^*\mathbf{v}_1 = \overline{\lambda}_0\mathbf{v}_1+\mathbf{v}_0,\quad (\mathbf{u}_1,\mathbf{v}_0) = 1,\quad (\mathbf{u}_1,\mathbf{v}_1) = 0, \label{eq1.6}$$ where the last two equations are the normalization conditions determining $\mathbf{v}_0$ and $\mathbf{v}_1$ uniquely for a given $\mathbf{u}_1$. Bifurcation of $\lambda_0$ into two eigenvalues $\lambda_\pm$ and the corresponding eigenvectors $\mathbf{u}_\pm$ are described by (see e.g. [@SeyMai2004]) $$\begin{array}{rcl} \lambda_\pm & = & \lambda_0\pm\sqrt{\mu_1\varepsilon} +\mu_2\varepsilon+o(\varepsilon),\\[5pt] \mathbf{u}_\pm & = & \mathbf{u}_0\pm\mathbf{u}_1\sqrt{\mu_1\varepsilon} +(\mu_1\mathbf{u}_0+\mu_2\mathbf{u}_1 -\mathbf{G}^{-1}\mathbf{A}_1\mathbf{u}_0)\varepsilon +o(\varepsilon), \end{array} \label{eq1.7}$$ where $\mathbf{G} = \mathbf{A}_0-\lambda_0\mathbf{I}+\mathbf{u}_1\mathbf{v}_1^*$. The coefficients $\mu_1$ and $\mu_2$ are $$\mu_1 = (\mathbf{A}_1\mathbf{u}_0,\mathbf{v}_0),\quad \mu_2 = \big((\mathbf{A}_1\mathbf{u}_0,\mathbf{v}_1) +(\mathbf{A}_1\mathbf{u}_1,\mathbf{v}_0)\big)/2. \label{eq1.8}$$ With a change of $\varepsilon$ from negative to positive values, the two eigenvalues $\lambda_\pm$ approach, collide with infinite speed (derivative with respect to $\varepsilon$ tends to infinity) at $\lambda_0$, and diverge in the perpendicular direction, see Figure \[fig1\]b. The eigenvectors interact too. At $\varepsilon = 0$, they merge to $\mathbf{u}_0$ up to a scalar complex factor. At nonzero $\varepsilon$, the eigenvectors $\mathbf{u}_\pm$ differ from $\mathbf{u}_0$ by the leading term $\pm\mathbf{u}_1\sqrt{\mu_1\varepsilon}$. This term takes the purely imaginary factor $i$ as $\varepsilon$ changes the sign, for example altering from negative to positive values. We call such a coupling of eigenvalues as *strong*. An exciting feature of the strong coupling is that the two eigenvalues cannot be distinguished after the interaction. Indeed, there is no natural rule telling how the eigenvalues before coupling correspond to those after the coupling. Crossing of eigenvalue surfaces =============================== Double eigenvalue with single eigenvector ----------------------------------------- Let, at the point ${\bf p}_0$, the spectrum of the complex matrix family ${\bf A}({\bf p})$ contain a double complex eigenvalue $\lambda_0$ with an eigenvector ${\bf u}_0$ and an associated vector ${\bf u}_1$. The splitting of the double eigenvalue with a change of the parameters is governed by equations (\[eq1.7\]) and (\[eq1.8\]). Introducing the real $n$-dimensional vectors $\bf f$, $\bf g$, $\bf h$, $\bf r$ with the components $${f}_s={\rm Re}\left(\frac{\partial \bf A}{\partial p_s} {\bf u}_0, {\bf v}_0\right),~~ {g}_s={\rm Im}\left(\frac{\partial \bf A}{\partial p_s} {\bf u}_0, {\bf v}_0\right), \label{eq2.1}$$ $${h}_s = {\rm Re} \left(\left(\frac{\partial \bf A}{\partial p_s} {\bf u}_0, {\bf v}_1\right)+ \left(\frac{\partial \bf A}{\partial p_s} {\bf u}_1, {\bf v}_0\right)\right),~~ {r}_s = {\rm Im} \left(\left(\frac{\partial \bf A}{\partial p_s} {\bf u}_0, {\bf v}_1\right)+ \left(\frac{\partial \bf A}{\partial p_s} {\bf u}_1, {\bf v}_0\right)\right), \label{eq2.1a}$$ $$s=1,\ldots,n.$$ and neglecting higher order terms, we obtain from (\[eq1.7\]) an asymptotic formula $${\rm Re}\Delta\lambda+i{\rm Im}\Delta\lambda= \pm\sqrt{\langle {\bf f}, \Delta {\bf p} \rangle+ i\langle {\bf g},\Delta {\bf p}\rangle}+ \frac{1}{2}(\langle {\bf h},\Delta {\bf p}\rangle+ i\langle {\bf r},\Delta {\bf p} \rangle), \label{eq2.2}$$ where $\Delta \lambda = \lambda_{\pm}-\lambda_0$, $\Delta{\bf p} = {\bf p}-{\bf p}_0$, and angular brackets denote inner product of real vectors: $\langle {\bf a}, {\bf b} \rangle=\sum_{s=1}^na_sb_s$. From equation (\[eq2.2\]) it is clear that the eigenvalue remains double in the first approximation if the two following equations are satisfied $$\label{eq2.2a} \langle {\bf f}, \Delta {\bf p}\rangle=0,~~ \langle {\bf g},\Delta {\bf p} \rangle=0.$$ This means that the double complex eigenvalue with the Jordan chain of length 2 has codimension 2. Thus, double complex eigenvalues occur at isolated points of the plane of two parameters, and in the three-parameter space the double eigenvalues form a curve [@Arnold2]. Equations (\[eq2.2a\]) define a tangent line to this curve at the point ${\bf p}_0$. Taking square of (\[eq2.2\]), where the terms linear with respect to the increment of parameters are neglected, and separating real and imaginary parts, we derive the equations $$({\rm Re}\Delta \lambda)^2-({\rm Im}\Delta \lambda)^2= \langle {\bf f},\Delta {\bf p} \rangle,~~ 2{\rm Re}\Delta \lambda\,{\rm Im}\Delta\lambda= \langle {\bf g},\Delta {\bf p} \rangle. \label{eq2.3}$$ Let us assume that $f_1^2+g_1^2 \ne 0$, which is the nondegeneracy condition for the complex eigenvalue $\lambda_0$. Isolating the increment $\Delta p_1$ in one of the equations (\[eq2.3\]) and substituting it into the other one we get $$g_1({\rm Re}\Delta\lambda)^2- 2f_1{\rm Re}\Delta\lambda\,{\rm Im}\Delta\lambda- g_1({\rm Im}\Delta\lambda)^2=\gamma, \label{eq2.4}$$ where $\gamma$ is a small real constant $$\gamma=\sum_{s=2}^n(f_sg_1-f_1g_s)\Delta p_s. \label{eq2.5}$$ Equation (\[eq2.4\]) describes hyperbolic trajectories of the eigenvalues $\lambda_{\pm}$ in the complex plane when only $\Delta p_1$ is changed and the increments $\Delta p_2$, $\ldots$, $\Delta p_n$ are fixed. Of course, any component of the vector $\Delta {\bf p}$ can be chosen instead of $\Delta p_1$. Let us study movement of eigenvalues in the complex plane in more detail. If $\Delta p_j=0$, $j=2,\ldots,n$, or if they are nonzero but satisfy the equality $\gamma=0$, then equation (\[eq2.4\]) yields two perpendicular lines which for $g_1 \ne 0$ are described by the expression $$g_1{\rm Re}(\lambda-\lambda_0)- \left(f_1\pm\sqrt{f_1^2+g_1^2}\right) {\rm Im}(\lambda-\lambda_0)=0. \label{eq2.6}$$ These lines intersect at the point $\lambda_0$ of the complex plane. Due to variation of the parameter $p_1$ two eigenvalues $\lambda_{\pm}$ approach along one of the lines (\[eq2.6\]), merge to $\lambda_0$ at $\Delta p_1 = 0 $, and then diverge along the other line (\[eq2.6\]), perpendicular to the line of approach; see Figure \[fig2\]b, where the arrows show motion of eigenvalues with a monotonous change of $p_1$. Recall that the eigenvalues that born after the coupling cannot be identified with the eigenvalues before the coupling. If $\gamma \ne 0$, then equation (\[eq2.4\]) defines a hyperbola in the complex plane. Indeed, for $g_1 \ne 0$ it is transformed to the equation of hyperbola $$(g_1{\rm Re}(\lambda-\lambda_0)- f_1{\rm Im}(\lambda-\lambda_0))^2- ({\rm Im}(\lambda-\lambda_0))^2(f_1^2+g_1^2)= \gamma g_1 \label{eq2.7}$$ with the asymptotes described by equation (\[eq2.6\]). As $\Delta p_1$ changes monotonously, two eigenvalues $\lambda_+$ and $\lambda_-$ moving each along its own branch of hyperbola come closer, turn and diverge; see Figure \[fig2\]a,c. Note that for a small $\gamma$ the eigenvalues $\lambda_{\pm}$ come arbitrarily close to each other without coupling that means [*avoided crossing*]{}. When $\gamma$ changes the sign, the quadrants containing hyperbola branches are changed to the adjacent. ![Crossing and avoided crossing of eigenvalues.[]{data-label="fig2"}](fig2.eps){width="80.00000%"} Expressing ${\rm Im}\Delta\lambda$ from the second of equations (\[eq2.3\]), substituting it into the first equation and then isolating ${\rm Re}\Delta\lambda$, we find $${\rm Re}\lambda_{\pm}=\lambda_0+\frac{1}{2}\langle {\bf h}, \Delta {\bf p}\rangle\pm \sqrt{\frac{1}{2}\left(\langle {\bf f}, \Delta {\bf p}\rangle+ \sqrt{\langle {\bf f},\Delta {\bf p}\rangle^2+\langle {\bf g},\Delta {\bf p}\rangle^2}\right)}. \label{eq2.8}$$ Similar transformation yields $${\rm Im}\lambda_{\pm}=\lambda_0+\frac{1}{2}\langle {\bf r}, \Delta {\bf p}\rangle\pm \sqrt{\frac{1}{2}\left(-\langle {\bf f}, \Delta {\bf p}\rangle+ \sqrt{\langle {\bf f},\Delta {\bf p}\rangle^2+\langle {\bf g},\Delta {\bf p}\rangle^2}\right)}. \label{eq2.9}$$ Equations (\[eq2.8\]) and (\[eq2.9\]) describe behavior of real and imaginary parts of eigenvalues $\lambda_{\pm}$ with a change of the parameters. On the other hand they define hypersurfaces in the spaces $(p_1,p_2,\ldots,p_n,{\rm Re}\lambda)$ and $(p_1,p_2,\ldots,p_n,{\rm Im}\lambda)$. The sheets ${\rm Re}\lambda_+(\bf p)$ and ${\rm Re}\lambda_-(\bf p)$ of the eigenvalue hypersurface (\[eq2.8\]) are connected at the points of the set $$\label{eq2.10} {\rm Re}\Delta\lambda=\frac{1}{2}\langle {\bf h}, \Delta {\bf p}\rangle,~~ \langle {\bf g}, \Delta {\bf p}\rangle=0,~~\langle {\bf f}, \Delta {\bf p}\rangle \le 0,$$ where the real parts of the eigenvalues $\lambda_{\pm}$ coincide: ${\rm Re}\lambda_-={\rm Re}\lambda_+$. Similarly, the set $$\label{eq2.11} {\rm Im}\Delta\lambda=\frac{1}{2}\langle {\bf r}, \Delta {\bf p}\rangle,~~ \langle {\bf g}, \Delta {\bf p}\rangle=0,~~\langle {\bf f}, \Delta {\bf p}\rangle \ge 0,$$ glues the sheets ${\rm Im}\lambda_+(\bf p)$ and ${\rm Im}\lambda_-(\bf p)$ of the eigenvalue hypersurface (\[eq2.9\]). To study the geometry of the eigenvalue hypersurfaces we look at their two-dimensional cross-sections. Consider for example the functions ${\rm Re}\lambda(p_1)$ and ${\rm Im}\lambda(p_1)$ at fixed values of the other parameters $p_2,p_3,\ldots,p_n$. When the increments $\Delta p_s=0$, $s=2,3,\ldots,n$, both the real and imaginary parts of the eigenvalues $\lambda_{\pm}$ cross at $p_1=p_1^0$, see Figure \[fig2\]b. The crossings are described by the double cusps defined by the equations following from (\[eq2.8\]) and (\[eq2.9\]) as $$\label{eq2.12} {\rm Re}\Delta\lambda = \pm\sqrt{ \frac{f_1\pm\sqrt{f_1^2+g_1^2}}{2}\Delta p_1}+\frac{h_1}{2}\Delta p_1, ~~ {\rm Im}\Delta\lambda = \pm\sqrt{ \frac{-f_1\pm\sqrt{f_1^2+g_1^2}}{2}\Delta p_1}+\frac{r_1}{2}\Delta p_1.$$ For the fixed $\Delta p_s\ne0$, $s=2,3,\ldots,n$, either real parts of the eigenvalues $\lambda_{\pm}$ cross due to variation of $p_1$ while the imaginary parts avoid crossing or vice-versa, as shown in Figure \[fig2\]a,c. Note that these two cases correspond to level crossing and width repulsion or vice-versa studied in [@Heiss2000]. The crossings, which occur at $p_1^{\times}=p_1^0-\sum_{s=2}^n (g_s/g_1)\Delta p_s$ and $$\label{eq2.15} {\rm Re}\lambda_h={\rm Re}\lambda_0-\frac{1}{2g_1}\sum_{s=2}^n(h_1g_s-g_1h_s)\Delta p_s,~~{\rm Im}\lambda_r={\rm Im}\lambda_0-\frac{1}{2g_1}\sum_{s=2}^n(r_1g_s-g_1r_s)\Delta p_s,$$ are described by the equations (\[eq2.8\]) and (\[eq2.9\]). In the vicinity of the crossing points the tangents of two intersecting curves are $$\label{eq2.13} {\rm Re}\lambda={\rm Re}\lambda_h+ \left(\frac{h_1}{2}\pm\frac{g_1}{2}\sqrt{\frac{g_1}{\gamma}} \right)(p_1-p_1^{\times}),$$ $$\label{eq2.14} {\rm Im}\lambda={\rm Im}\lambda_r+ \left(\frac{r_1}{2}\pm\frac{g_1}{2}\sqrt{-\frac{g_1}{\gamma}} \right)(p_1-p_1^{\times}),$$ where the coefficient $\gamma$ is defined by equation (\[eq2.5\]). Lines (\[eq2.13\]) and (\[eq2.14\]) tend to the vertical position as $\gamma\rightarrow 0$ and coincide at $\gamma=0$. The avoided crossings are governed by the equations (\[eq2.8\]) and (\[eq2.9\]). ![Crossing of eigenvalue surfaces near the double eigenvalue with single eigenvector.[]{data-label="fig3"}](fig3.eps){width="85.00000%"} If the vector of parameters consists of only two components ${\bf p}=(p_1, p_2)$, then in the vicinity of the point ${\bf p}_0$, corresponding to the double eigenvalue $\lambda_0$, the eigenvalue surfaces (\[eq2.8\]) and (\[eq2.9\]) have the form of the well-known Whitney umbrella; see Figure \[fig3\]. The sheets of the eigensurfaces are connected along the rays (\[eq2.10\]) and (\[eq2.11\]). We emphasize that these rays are inclined with respect to the plane of the parameters $p_1$, $p_2$. The cross-sections of the eigensurfaces by the planes orthogonal to the axis $p_2$, described by the equations (\[eq2.12\])–(\[eq2.14\]), are shown in Figure \[fig2\]. Note that the rays (\[eq2.10\]), (\[eq2.11\]) and the point ${\bf p}_0$ are well-known in crystal optics as [*the branch cuts*]{} and [*the singular axis*]{}, respectively [@Berry2]. We emphasize that the branch cut is a general phenomenon: it always appears near the EP degeneracy. In general, branch cuts may be infinite or end up at another EP. The second scenario is always the case when the complex matrix $\mathbf{A}(\mathbf{p})$ is a small perturbation of a family of real symmetric matrices [@KMS]. Consider the movement of the eigenvalues in the complex plane near the point ${\bf p}_0$ due to cyclic variation of the parameters $p_1$ and $p_2$ of the form $\Delta p_1=a+r\cos{\varphi}$ and $\Delta p_2=b+r\sin{\varphi}$, where $a$, $b$, and $r$ are small parameters of the same order. From equations (\[eq2.3\]) we derive $$(g_1{\rm Re}\Delta\lambda^2-2f_1{\rm Re}\Delta\lambda{\rm Im}\Delta\lambda- g_1{\rm Im}\Delta\lambda^2-b(f_2g_1-f_1g_2))^2+$$ $$\label{eq2.16} +(g_2{\rm Re}\Delta\lambda^2-2f_2{\rm Re}\Delta\lambda{\rm Im}\Delta\lambda- g_2{\rm Im}\Delta\lambda^2-a(f_1g_2-g_1f_2))^2= (f_2g_1-f_1g_2)^2r^2.$$ ![Movement of eigenvalues due to cyclic evolution of the parameters.[]{data-label="fig4"}](fig5.eps){width="100.00000%"} Movement of eigenvalues on the complex plane governed by equation (\[eq2.16\]) is shown in Figure \[fig4\]. If the contour encircles the point ${\bf p}_0$, then the eigenvalues move along the curve (\[eq2.16\]) around the double eigenvalue $\lambda_0$ in the complex plane, see Figure \[fig4\]c. Indeed, in this case $a^2+b^2<r^2$ and the loop (\[eq2.16\]) crosses the lines ${\rm Re}\lambda={\rm Re}\lambda_0$ and ${\rm Im}\lambda={\rm Im}\lambda_0$ at the four points given by the equations $$\label{eq2.18} ({\rm Im}\Delta\lambda)^2 = \frac{(f_2g_1-f_1g_2)\left(g_2 a -g_1 b\pm \sqrt{(g_2 a-g_1 b)^2+(r^2-a^2-b^2)(g_1^2+g_2^2)}\right)}{g_1^2+g_2^2}$$ and $$\label{eq2.17} ({\rm Re}\Delta\lambda)^2 = \frac{(f_2g_1-f_1g_2) \left(g_1 b -g_2 a \pm \sqrt{(g_1 b - g_2 a)^2+(r^2-a^2-b^2)(g_1^2+g_2^2)}\right)}{g_1^2+g_2^2},$$ respectively. When $a^2+b^2=r^2$ the loop overlaps at the double eigenvalue and its form depends on the sign of the quantity $\sigma=(f_2g_1-f_1g_2)(g_1b-g_2a)$. If $\sigma<0$ the eigenvalues cross the line ${\rm Re}\lambda={\rm Re}\lambda_0$ (Figure \[fig4\]b), otherwise they cross the line ${\rm Im}\lambda={\rm Im}\lambda_0$ (Figure \[fig4\]d). Eigenvalues strongly couple at the point $\lambda_0$ in the complex plane. For $a^2+b^2>r^2$ the circuit in the parameter plane does not contain the point ${\bf p}_0$ and the eigenvalues move along the two different closed paths (“kidneys”, [@Arnold1989]) in the complex plane, see Figure \[fig4\]a,e. Each eigenvalue crosses the line ${\rm Re}\lambda={\rm Re}\lambda_0$ twice for $\sigma<0$ (Figure \[fig4\]a), and for $\sigma>0$ they cross the line ${\rm Im}\lambda={\rm Im}\lambda_0$ (Figure \[fig4\]d). Note that the “kidneys” in the complex plane were observed in [@KKM1] for the specific problem of Stark resonances for a double $\delta$ quantum well. Double eigenvalue with two eigenvectors --------------------------------------- Let $\lambda_0$ be a double eigenvalue of the matrix $\mathbf{A}_0 = \mathbf{A}(\mathbf{p}_0)$ with two eigenvectors $\mathbf{u}_1$ and $\mathbf{u}_2$. Under perturbation of parameters $\mathbf{p} = \mathbf{p}_0+\Delta\mathbf{p}$, the bifurcation of $\lambda_0$ into two simple eigenvalues $\lambda_+$ and $\lambda_-$ occurs. Using (\[eq1.3\]) and (\[eq1.4b\]), we obtain the asymptotic formula for $\lambda_\pm$ under multiparameter perturbation as $$\lambda_\pm = \lambda_0+\frac{\langle\mathbf{d}_{11}+\mathbf{d}_{22}, \Delta\mathbf{p}\rangle}{2}\pm \sqrt{\frac{\langle\mathbf{d}_{11}-\mathbf{d}_{22}, \Delta\mathbf{p}\rangle^2}{4}+\langle\mathbf{d}_{12}, \Delta\mathbf{p}\rangle\langle\mathbf{d}_{21}, \Delta\mathbf{p}\rangle}, \label{eq3.1}$$ where $\mathbf{d}_{ij} = (d_{ij}^1,\ldots,d_{ij}^n)$ is a complex vector with the components $$d_{ij}^k = \left(\frac{\partial\mathbf{A}}{\partial p_k} \mathbf{u}_i,\mathbf{v}_j\right), \label{eq3.2}$$ and $\langle\mathbf{d}_{ij},\Delta\mathbf{p}\rangle = \langle\mathrm{Re}\,\mathbf{d}_{ij},\Delta\mathbf{p}\rangle +i\langle\mathrm{Im}\,\mathbf{d}_{ij},\Delta\mathbf{p}\rangle$. In the same way as we derived formulae (\[eq2.8\]) and (\[eq2.9\]), we obtain from (\[eq3.1\]) the expressions for real and imaginary parts of $\lambda_\pm$ in the form $$\mathrm{Re}\,\lambda_\pm = \mathrm{Re}\,\lambda_0 +\mathrm{Re}\,\langle\mathbf{d}_{11}+\mathbf{d}_{22}, \Delta\mathbf{p}\rangle/2\pm \sqrt{(|c|+\mathrm{Re}\,c)/2}, \label{eq3.3}$$ $$\mathrm{Im}\,\lambda_\pm = \mathrm{Im}\,\lambda_0 +\mathrm{Im}\,\langle\mathbf{d}_{11}+\mathbf{d}_{22}, \Delta\mathbf{p}\rangle/2\pm \sqrt{(|c|-\mathrm{Re}\,c)/2}, \label{eq3.4}$$ where $$c = \langle\mathbf{d}_{11}-\mathbf{d}_{22}, \Delta\mathbf{p}\rangle^2/4+\langle\mathbf{d}_{12}, \Delta\mathbf{p}\rangle\langle\mathbf{d}_{21}, \Delta\mathbf{p}\rangle. \label{eq3.5}$$ Considering the situation when $\lambda_0$ remains double under perturbation of parameters, i.e. $\lambda_+ = \lambda_-$, we obtain the two independent equations $$\mathrm{Re}\,c = 0,\quad\mathrm{Im}\,c = 0. \label{eq3.6}$$ By using (\[eq1.3\])–(\[eq1.4b\]), one can show that the perturbed double eigenvalue $\lambda_+ = \lambda_-$ possesses a single eigenvector $\mathbf{u}_+ = \mathbf{u}_-$, i.e., the weak coupling becomes strong due to perturbation, see [@SeyMai2004]. The perturbed double eigenvalue has two eigenvectors only when the matrix in the left-hand side of (\[eq1.4\]) is proportional to the identity matrix. This yields the equations $$\langle\mathbf{d}_{11},\Delta\mathbf{p}\rangle = \langle\mathbf{d}_{22},\Delta\mathbf{p}\rangle,\quad \langle\mathbf{d}_{12},\Delta\mathbf{p}\rangle = \langle\mathbf{d}_{21},\Delta\mathbf{p}\rangle = 0, \label{eq3.7}$$ Conditions (\[eq3.7\]) imply (\[eq3.6\]) and represent six independent equations taken for real and imaginary parts. Thus, weak coupling of eigenvalues is a phenomenon of codimension 6, which generically occurs at isolated points in 6-parameter space, see [@Arnold2; @MH]. First, let us study behavior of the eigenvalues $\lambda_+$ and $\lambda_-$ depending on one parameter, say $p_1$, when the other parameters $p_2,\ldots,p_n$ are fixed in the neighborhood of the coupling point $\lambda_+(\mathbf{p}_0) = \lambda_-(\mathbf{p}_0) = \lambda_0$. In case $\Delta p_2 = \cdots = \Delta p_n = 0$, expression (\[eq3.1\]) yields $$\lambda_\pm = \lambda_0+\left(\frac{d_{11}^1+d_{22}^1}{2} \pm\sqrt{\frac{(d_{11}^1-d_{22}^1)^2}{4}+d_{12}^1d_{21}^1} \right)\Delta p_1. \label{eq3.8}$$ The two eigenvalues couple when $\Delta p_1 = 0$ with the double eigenvalue $\lambda_0$, see Figure \[fig3.1\]a. As we showed in Section 2, the eigenvalues $\lambda_+$ and $\lambda_-$ behave as smooth functions at the coupling point; they possess different eigenvectors, which are smooth functions of $\Delta p_1$ too. ![Weak coupling of eigenvalues and avoided crossing.[]{data-label="fig3.1"}](figDP.eps){width="80.00000%"} If the perturbations $\Delta p_2,\ldots,\Delta p_n$ are nonzero, the avoided crossing of the eigenvalues $\lambda_\pm$ with a change of $p_1$ is a typical scenario. We can distinguish different cases by checking intersections of real and imaginary parts of $\lambda_+$ and $\lambda_-$. By using (\[eq3.3\]), we find that $\mathrm{Re}\,\lambda_+ = \mathrm{Re}\,\lambda_-$ if $$\mathrm{Im}\,c = 0,\quad \mathrm{Re}\,c < 0. \label{eq3.9}$$ Analogously, from (\[eq3.4\]) it follows that $\mathrm{Im}\,\lambda_+ = \mathrm{Im}\,\lambda_-$ if $$\mathrm{Im}\,c = 0,\quad \mathrm{Re}\,c > 0. \label{eq3.10}$$ Let us write expression (\[eq3.5\]) in the form $$c = c_0+c_1\Delta p_1+c_2(\Delta p_1)^2, \label{eq3.11}$$ where $$\begin{array}{c} c_0 = \displaystyle\sum_{k,l = 2}^n \left[(d_{11}^k-d_{22}^k)(d_{11}^l-d_{22}^l)/4 +d_{12}^kd_{21}^l\right]\Delta p_k\Delta p_l, \\[15pt] c_1 = \displaystyle\sum_{k = 2}^n \left[(d_{11}^1-d_{22}^1)(d_{11}^k-d_{22}^k)/2 +(d_{12}^1d_{21}^k+d_{12}^kd_{21}^1)\right]\Delta p_k, \\[12pt] c_2 = (d_{11}^1-d_{22}^1)^2/4+d_{12}^1d_{21}^1. \end{array} \label{eq3.12}$$ If the discriminant $D = (\mathrm{Im}\,c_1)^2-4\mathrm{Im}\,c_0\mathrm{Im}\,c_2 > 0$, the equation $\mathrm{Im}\,c = 0$ yields two solutions $$\Delta p_1^a = \frac{-\mathrm{Im}\,c_1-\sqrt{D}}{2\mathrm{Im}\,c_2},\quad \Delta p_1^b = \frac{-\mathrm{Im}\,c_1+\sqrt{D}}{2\mathrm{Im}\,c_2}. \label{eq3.13}$$ There are no real solutions if $D < 0$, and the single solution corresponds to the degenerate case $D = 0$. At the points $p_1^a = p_1^0+\Delta p_1^a$ and $p_1^b = p_1^0+\Delta p_1^b$ the values of $c$ are real, and we denote them by $c_a$ and $c_b$, respectively. According to (\[eq3.9\]) and (\[eq3.10\]), the sign of $c_{a,b}$ determines whether the real or imaginary parts of $\lambda_\pm$ coincide at $p_1^{a,b}$. In the nondegenerate case $D \ne 0$, there are four types of avoided crossing shown in Figure \[fig3.1\]b–e. The first case corresponds to $D < 0$ when both real and imaginary parts of the eigenvalues $\lambda_\pm$ are separate at all $p_1$, see Figure \[fig3.1\]b. In other cases $D > 0$, so that there are two separate points $p_{1}^a$ and $p_{1}^b$. For the second type we have $c_a > 0$ and $c_b < 0$, when both real and imaginary parts of $\lambda_\pm$ have a single intersection, see Figure \[fig3.1\]c. The equivalent situation when $c_a < 0$ and $c_b > 0$ is obtained by interchanging the points $p_{1}^a$ and $p_{1}^b$ in Figure \[fig3.1\]c. The third type is represented by $c_{a,b} < 0$, when the real parts of $\lambda_\pm$ have two intersections and $\mathrm{Im}\,\lambda_\pm$ do not intersect, see Figure \[fig3.1\]d. Finally, if $c_{a,b} > 0$, when the real parts of $\lambda_\pm$ do not intersect and $\mathrm{Im}\,\lambda_\pm$ intersect at both $p_{1}^a$ and $p_{1}^b$, see Figure \[fig3.1\]e. The last column in Figure \[fig3.1\] shows behavior of the eigenvalues $\lambda_\pm$ on the complex plane. In each of the cases b–e, the trajectories of eigenvalues on the complex plane may intersect and/or self-intersect, which can be studied by using expression (\[eq3.1\]). Note that intersections of the eigenvalue trajectories on the complex plane do not imply eigenvalue coupling since the eigenvalues $\lambda_+$ and $\lambda_-$ pass the intersection point at different values of $p_1$. The small loops of the eigenvalue trajectories on the complex plane, shown in Figure \[fig3.1\]b,e, shrink as the perturbations of the parameters $\Delta p_2,\Delta p_3,\ldots,\Delta p_n$ tend to zero. Finally, we mention that the case of Figure \[fig3.1\]c is the only avoided crossing scenario when the eigenvalues follow the initial directions on the complex plane after interaction. In the other three cases (b,d, and e) the eigenvalues interchange their directions due the interaction. Let us consider a system depending on two parameters $p_1$ and $p_2$ with the weak coupling of eigenvalues at $p_1 = p_1^0$ and $p_2 = p_2^0$. The double eigenvalue $\lambda_0$ bifurcates into a pair $\lambda_\pm$ under perturbation of the parameters $\Delta p_1$ and $\Delta p_2$. Conditions (\[eq3.9\]) and (\[eq3.10\]) determine the values of parameters, at which the real and imaginary parts of $\lambda_\pm$ coincide. Let us write expression (\[eq3.5\]) in the form $$c = c_{11}(\Delta p_1)^2+c_{12}\Delta p_1\Delta p_2+c_{22}(\Delta p_2)^2, \label{eq3.14}$$ where $$\begin{array}{c} c_{11} = (d_{11}^1-d_{22}^1)^2/4+d_{12}^1d_{21}^1,\quad c_{22} = (d_{11}^2-d_{22}^2)^2/4+d_{12}^2d_{21}^2,\\[5pt] c_{12} = (d_{11}^1-d_{22}^1)(d_{11}^2-d_{22}^2)/2+d_{12}^1d_{21}^2 +d_{12}^2d_{21}^1. \end{array} \label{eq3.15}$$ If the discriminant $D' = (\mathrm{Im}\,c_{12})^2-4\mathrm{Im}\,c_{11}\mathrm{Im}\,c_{22} > 0$, the equation $\mathrm{Im}\,c = 0$ yields the two crossing lines $$\begin{array}{l} l_a: \ \ 2\mathrm{Im}\,c_{11}\Delta p_1+(\mathrm{Im}\,c_{12}+\sqrt{D'})\Delta p_2 = 0,\\[5pt] l_b: \ \ 2\mathrm{Im}\,c_{11}\Delta p_1+(\mathrm{Im}\,c_{12}-\sqrt{D'})\Delta p_2 = 0. \end{array} \label{eq3.16}$$ There are no real solutions if $D' < 0$, and the lines $l_a$ and $l_b$ coincide in the degenerate case $D' = 0$. On the lines $l_{a,b}$ the values of $c$ are real numbers of the same sign; we denote $\gamma_a = \mathrm{sign}\,c$ for the line $l_a$, and $\gamma_b = \mathrm{sign}\,c$ for the line $l_b$. According to (\[eq3.9\]) and (\[eq3.10\]), the real or imaginary parts of $\lambda_\pm$ coincide at $l_{a,b}$ for negative or positive $\gamma_{a,b}$, respectively. One can distinguish four types of the graphs for $\mathrm{Re}\,\lambda_\pm(p_1,p_2)$ and $\mathrm{Im}\,\lambda_\pm(p_1,p_2)$ shown in Figure \[fig3.2\]. In nondegenerate case $D' \ne 0$, the eigenvalues $\lambda_+$ and $\lambda_-$ are different for all parameter values except the initial point $p_{1,2} = p_{1,2}^0$. If $D' < 0$, the eigenvalue surfaces are cones with non-elliptic cross-section, see Figure \[fig3.2\]a. Other three types correspond to the case $D' > 0$. If $\gamma_a < 0$ and $\gamma_b > 0$ then there is an intersection of the real parts along the line $l_a$ and an intersection of the imaginary parts along the line $l_b$ (in case $\gamma_a > 0$ and $\gamma_b < 0$ the lines $l_a$ and $l_b$ are interchanged), see Figure \[fig3.2\]b. If $\gamma_a < 0$ and $\gamma_b < 0$ then the real parts intersect along the both lines $l_a$ and $l_b$ forming a “cluster of shells”, while there is no intersections for the imaginary parts, see Figure \[fig3.2\]c. Finally, if $\gamma_a > 0$ and $\gamma_b > 0$ then there is no intersections for the real parts, while the imaginary parts intersect along the both lines $l_a$ and $l_b$, see Figure \[fig3.2\]d. ![Eigenvalue surfaces near a point of weak coupling.[]{data-label="fig3.2"}](figDPS.eps){width="80.00000%"} Example ======= As a physical example, we consider propagation of light in a homogeneous non-magnetic crystal in the general case when the crystal possesses natural optical activity (chirality) and dichroism (absorption) in addition to biaxial birefringence, see [@Berry2] for the general formulation. The optical properties of the crystal are characterized by the inverse dielectric tensor $\boldsymbol \eta$. The vectors of electric field $\bf E$ and displacement $\bf D$ are related as [@LLP] $$\label{eq2.19a} {\bf E}={\boldsymbol \eta}{\bf D}.$$ The tensor $\boldsymbol \eta$ is described by a non-Hermitian complex matrix. The electric field $\bf E$ and magnetic field $\bf H$ in the crystal are determined by Maxwell’s equations [@LLP] $$\label{eq2.19} {\rm rot} {\bf E}=-\frac{1}{c}\frac{\partial {\bf H}}{\partial t}, ~~ {\rm rot} {\bf H}=\frac{1}{c}\frac{\partial {\bf D}}{\partial t},$$ where $t$ is time and $c$ is the speed of light in vacuum. A monochromatic plane wave of frequency $\omega$ that propagates in a direction specified by a real unit vector ${\bf s}=(s_1,s_2,s_3)$ has the form $$\label{eq2.20} {\bf D}({\bf r},t) = {\bf D}({\bf s})\exp i\omega\!\left(\frac{n({\bf s})}{c} {\bf s}^T{\bf r}-t\right),\ {\bf H}({\bf r},t) = {\bf H}({\bf s})\exp i\omega\!\left(\frac{n({\bf s})}{c} {\bf s}^T{\bf r}-t\right),$$ where $n({\bf s})$ is a refractive index, and ${\bf r} = (x_1,x_2,x_3)$ is the real vector of spatial coordinates. Substituting the wave (\[eq2.20\]) into Maxwell’s equations (\[eq2.19\]), we find $$\label{eq2.21} {\bf H} = n [{\bf s}, {\boldsymbol \eta}{\bf D}], ~~ {\bf D} = - n [{\bf s}, {\bf H}],$$ where square brackets indicate cross product of vectors [@LLP]. With the vector $\bf H$ determined by the first equation of (\[eq2.21\]), the second equation of (\[eq2.21\]) yields [@Berry2] $$\label{eq2.22} -[{\bf s},[{\bf s}, {\boldsymbol \eta} {\bf D}({\bf s})]]={\boldsymbol \eta}{\bf D}({\bf s})-{\bf s}({\bf s}^T{\boldsymbol \eta}{\bf D}({\bf s}))=\frac{1}{n^2({\bf s})}{\bf D}({\bf s}).$$ Multiplying equation (\[eq2.22\]) by the vector $\mathbf{s}^T$ from the left we find that for plane waves the vector ${\bf D}$ is always orthogonal to the direction $\bf s$, i.e., ${\bf s}^T{\bf D}({\bf s})=0$. Since the quantity ${\bf s}^T{\boldsymbol \eta}{\bf D}({\bf s})$ is a scalar, we can write (\[eq2.22\]) in the form of an eigenvalue problem for the complex non-Hermitian matrix $\bf A({\bf s})$ dependent on the vector of parameters ${\bf s}=(s_1,s_2,s_3)$: $$\label{eq2.24} {\bf A}{\bf u}=\lambda{\bf u},~~{\bf A}({\bf s})=({\bf I}-{\bf s}{\bf s}^T){\boldsymbol \eta}({\bf s}),$$ where $\lambda=n^{-2}$, ${\bf u}={\bf D}$, and ${\bf I}$ is the identity matrix. Multiplying the matrix $\bf A$ by the vector $\bf s$ from the left we conclude that ${\bf s}^T{\bf A}=0$, i.e., the vector ${\bf s}$ is the left eigenvector with the eigenvalue $\lambda=0$. Zero eigenvalue always exists, because $\det({\bf I}-{\bf s}{\bf s}^T)\equiv 0$, if $\|{\bf s}\|=1$. The matrix ${\bf A}({\bf s})$ defined by equation (\[eq2.24\]) is a product of the matrix ${\bf I}-{\bf s}{\bf s}^T$ and the inverse dielectric tensor ${\boldsymbol \eta}({\bf s})$. The symmetric part of ${\boldsymbol \eta}$ constitutes the anisotropy tensor describing the birefringence of the crystal. It is represented by the complex symmetric matrix $\bf U$, which is independent of the vector of parameters $\bf s$. The antisymmetric part of ${\boldsymbol \eta}$ is determined by the optical activity vector ${\bf g}({\bf s}) = (g_1,g_2,g_3)$, describing the chirality (optical activity) of the crystal. It is represented by the skew-symmetric matrix $$\label{eq2.25} {\bf G} = i\left(% \begin{array}{ccc} 0 & -g_3 & g_2 \\ g_3 & 0 & -g_1 \\ -g_2 & g_1 & 0 \\ \end{array}% \right).$$ The vector $\bf g$ is given by the expression ${\bf g}({\bf s})={\boldsymbol \gamma}{\bf s}$, where $\boldsymbol \gamma$ is the optical activity tensor represented by a symmetric complex matrix. Thus, the matrix $\mathbf{G}(\mathbf{s})$ depends linearly on the parameters $s_1$, $s_2$, $s_3$. In the present formulation, the problem was studied analytically and numerically in [@Berry2]. Below we present two specific numerical examples in case of a non-diagonal matrix $\boldsymbol\gamma$, for which the structure of singularities was not fully investigated. Unlike [@Berry2], where the reduction to two dimensions was carried out, we work with the three-dimensional form of problem (\[eq2.24\]). Our intention here is to give guidelines for using our theory by means of the relatively simple $3\times 3$ matrix family, keeping in mind that the main area of applications would be higher dimensional problems. As a first example, we choose the inverse dielectric tensor in the form $$\label{eq2.26} {\boldsymbol \eta} = \left(% \begin{array}{ccc} 3 & 0 & 0\\ 0 & 1 & 0\\ 0 & 0 & 2\\ \end{array}% \right)+ i\left(% \begin{array}{ccc} 0 & 1 & 2\\ 1 & 0 & 0\\ 2 & 0 & 0\\ \end{array}% \right)+ i\left(% \begin{array}{lll} 0 & -s_1 & 0\\ s_1 & 0 & -s_3\\ 0 & s_3 & 0\\ \end{array}% \right)$$ where $s_3=\sqrt{1-s_1^2-s_2^2}$. The crystal defined by (\[eq2.26\]) is dichroic and optically active with the non-diagonal matrix $\boldsymbol\gamma$. When $s_1=0$ and $s_2=0$ the spectrum of the matrix ${\bf A}$ consists of the double eigenvalue $\lambda_0=2$ and the simple zero eigenvalue. The double eigenvalue possesses the eigenvector ${\bf u}_0$ and associated vector ${\bf u}_1$: $$\label{eq2.27} {\bf u}_0=\left(% \begin{array}{c} i \\ -1 \\ 0 \\ \end{array}% \right),~~ {\bf u}_1=\left(% \begin{array}{c} 0 \\ 1 \\ 0 \\ \end{array}% \right).$$ The eigenvector ${\bf v}_0$ and associated vector ${\bf v}_1$ corresponding to the double eigenvalue $\lambda_0=2$ of the adjoint matrix ${\bf A}^*$ are $$\label{eq2.28} {\bf v}_0=\left(% \begin{array}{c} i \\ 1 \\ 1+i/2 \\ \end{array}% \right),~~ {\bf v}_1=\left(% \begin{array}{c} i \\ 0 \\ 1/2-i/4 \\ \end{array}% \right).$$ The vectors ${\bf u}_0$, ${\bf u}_1$ and ${\bf v}_0$, ${\bf v}_1$ satisfy the normalization and orthogonality conditions (\[eq1.6\]). Calculating the derivatives of the matrix ${\bf A}(s_1,s_2)$ at the point ${\bf s}_0=(0,0,1)$ we obtain $$\label{eq2.29} \frac{\partial {\bf A}}{\partial s_1}=\left(% \begin{array}{ccc} -2i & -2i & -2 \\ i & 0 & 0 \\ -3 & -i & -2i \\ \end{array}% \right),~~ \frac{\partial {\bf A}}{\partial s_2}=\left(% \begin{array}{ccc} 0 & 0 & 0 \\ -2i & -i & -2 \\ -i & -1 & i \\ \end{array}% \right).$$ ![Eigensurfaces of the crystal (\[eq2.26\]) and their approximations.[]{data-label="fig7"}](fig7.eps){width="100.00000%"} Substitution of the derivatives (\[eq2.29\]) together with the vectors given by equations (\[eq2.27\]) and (\[eq2.28\]) into the formulae (\[eq2.1\]) and (\[eq2.1a\]) yields the vectors ${\bf f}$, ${\bf g}$ and $\bf h$, $\bf r$ as $$\label{eq2.30} {\bf f}=(0, 4),~~{\bf g}=(-4, 0),~~{\bf h}=(0,0),~~{\bf r}=(-4,0).$$ With the vectors (\[eq2.30\]) we find from (\[eq2.8\]) and (\[eq2.9\]) the approximations of the eigensurfaces ${\rm Re}\lambda(s_1,s_2)$ and ${\rm Im}\lambda(s_1,s_2)$ in the vicinity of the point ${\bf s}_0=(0,0,1)$: $$\label{eq2.31} {\rm Re}\lambda_{\pm} = 2 \pm \sqrt{2s_2 + 2\sqrt{s_1^2+s_2^2}},~~ {\rm Im}\lambda_{\pm} = -2s_1\pm\sqrt{-2s_2+2\sqrt{s_1^2+s_2^2}}.$$ Calculation of the exact solution of the characteristic equation for the matrix $\bf A$ with the inverse dielectric tensor $\boldsymbol \eta$ defined by equation (\[eq2.26\]) shows a good agreement of the approximations (\[eq2.31\]) with the numerical solution, see Figure [\[fig7\]]{}. One can see that the both surfaces of real and imaginary parts have a Whitney umbrella singularity at the coupling point; the surfaces self-intersect along different rays, which together constitute a straight line when projected on parameter plane. As a second numerical example, let us consider the inverse dielectric tensor as $${\boldsymbol\eta} = \left(\begin{array}{ccc} 1 & 0 & 1 \\ 0 & 1 & 0 \\ 1 & 0 & 4 \end{array}\right) +i\left(\begin{array}{ccc} 5 & 0 & 4 \\ 0 & 5 & 2 \\ 4 & 2 & 0 \end{array}\right) +4i\left(\begin{array}{ccc} 0 & -s_1-is_2 & is_3 \\ s_1+is_2 & 0 & -s_3 \\ -is_3 & s_3 & 0 \end{array}\right). \label{ex2.1}$$ At $\mathbf{s} = (0,0,1)$, the matrix $\mathbf{A}$ has the double eigenvalue $\lambda_0 = 1+5i$ with two eigenvectors and the simple zero eigenvalue. The eigenvectors $\mathbf{u}_1$, $\mathbf{u}_2$ of $\lambda_0$ and the eigenvectors $\mathbf{v}_1$, $\mathbf{v}_2$ of $\overline{\lambda}_0$ for the adjoint matrix $\mathbf{A}^*$ are $$\mathbf{u}_1 = \left(\begin{array}{c} 1 \\ 0 \\ 0 \end{array}\right),\quad \mathbf{u}_2 = \left(\begin{array}{c} 0 \\ 1 \\ 0 \end{array}\right),\quad \mathbf{v}_1 = \left(\begin{array}{c} 1 \\ 0 \\ \frac{-3-4i}{1-5i} \end{array}\right),\quad \mathbf{v}_2 = \left(\begin{array}{c} 0 \\ 1 \\ \frac{2i}{1-5i} \end{array}\right). \label{ex2.2}$$ These vectors satisfy normalization conditions (\[eq1.NC\]). Taking derivatives of the matrix $\mathbf{A}$ with respect to parameters $s_1$ and $s_2$, where $s_3 = \sqrt{1-s_1^2-s_2^2}$, and using formula (\[eq3.2\]), we obtain $$\mathbf{d}_{11} = (-2-8i, 0),\ \mathbf{d}_{12} = (6i, -9-4i),\ \mathbf{d}_{21} = (-10i, 7-4i),\ \mathbf{d}_{22} = (0, -4i). \label{ex2.3}$$ Using (\[ex2.3\]) in formulae (\[eq3.3\])–(\[eq3.5\]), we find approximations for real and imaginary parts of two nonzero eigenvalues $\lambda_\pm$ near the point $\mathbf{s} = (0,0,1)$ as $$\mathrm{Re}\,\lambda_\pm = 1-s_1\pm \sqrt{(|c|+\mathrm{Re}\,c)/2},\ \ \mathrm{Im}\,\lambda_\pm = 5-4s_1-2s_2\pm \sqrt{(|c|-\mathrm{Re}\,c)/2}, \label{ex2.4}$$ where $c = (45+8i)s_1^2+128is_1s_2+(-83+8i)s_2^2$. Approximations of eigenvalue surfaces (\[ex2.4\]) and the exact solutions are presented in Figure \[fig4.1\]. The eigenvalue surfaces have intersections both in $(s_1,s_2,\mathrm{Re}\,\lambda)$ and $(s_1,s_2,\mathrm{Im}\,\lambda)$ spaces. These intersections are represented by two different lines $l_a$ and $l_b$ in parameter space, see Figure \[fig3.2\]b. ![Eigensurfaces of the crystal (\[ex2.1\]) and their approximations.[]{data-label="fig4.1"}](fig8.eps){width="100.00000%"} Conclusion ========== A general theory of coupling of eigenvalues of complex matrices smoothly depending on multiple real parameters has been presented. Diabolic and exceptional points have been mathematically described and general formulae for coupling of eigenvalues at these points have been derived. This theory gives a clear and complete picture of crossing and avoided crossing of eigenvalues with a change of parameters. It has a very broad field of applications since any physical system contains parameters. It is important that the presented theory of coupling gives not only qualitative, but also quantitative results on eigenvalue surfaces based only on the information at the diabolic and exceptional points. This information includes eigenvalues, eigenvectors and associated vectors with derivatives of the system matrix taken at the singular points. We emphasize that the developed methods provide a firm basis for analysis of spectrum singularities of matrix operators. Acknowledgement =============== The work is supported by the research grants RFBR-NSFC 02-01-39004, RFBR 03-01-00161, CRDF-BRHE Y1-MP-06-19, and CRDF-BRHE Y1-M-06-03. [11]{} Hamilton W. R. 1833. On a general Method of expressing the Paths of Light, and of the Planets, by the Coefficients of a Characteristic Function. *Dublin University Review and Quarterly Magazine* [**1**]{}, 795–826 Berry M., Bhandari R. and Klein S. 1999. Black plastic sandwiches demonstrating biaxial optical anisotropy. [*Eur. J. Phys.*]{} **20**, 1–-14 Von Neumann J. and Wigner E. P. 1929. Über das Verhalten von Eigenwerten bei adiabatischen Prozessen. *Zeitschrift für Physik* [**30**]{}, 467–470 Herring C. 1937. Accidental Degeneracy in the Energy Bands of Crystals. *Phys. Rev.* **52**, 365–373 Teller E. 1937. The Crossing of Potential Surfaces. *J. Phys. Chemistry* [**41**]{}, 109–116 Berry M. V. and Wilkinson M. 1984. Diabolical points in the spectra of triangles. *Proc. R. Soc. Lond. A* [**392**]{}, 15–43 Heiss W. D. 2000. Repulsion of resonance states and exceptional points, *Phys. Rev.* **61**, 929–932 Dembowsky C., Gräf H.-D., Harney H. L., Heine A., Heiss W. D., Rehfeld H. and Richter A. 2001. Experimental observation of the topological structure of exceptional points. *Phys. Rev. Lett.* [**86**]{}, 787–790 Berry M. V. and Dennis M. R. 2003. The optical singularities of birefringent dichroic chiral crystals. *Proc. R. Soc. Lond. A* [**459**]{}, 1261–1292 Dembowski C., Dietz B., Gräf H.-D., Harney H. L., Heine A., Heiss W. D. and Richter A. 2003. Observation of a Chiral State in a Microwave Cavity. [*Phys. Rev. Lett.*]{} **90**, 034101 Keck F., Korsch, H. J. and Mossmann S. 2003. Unfolding a diabolic point: a generalized crossing scenario. *J. Phys. A: Math. Gen.* [**36**]{}, 2125–2137 Korsch H. J. and Mossmann S. 2003. Stark resonances for a double $\delta$ quantum well: crossing scenarios, exceptional points and geometric phases. *J. Phys. A: Math. Gen.* [**36**]{}, 2139–2153 Heiss W. D. 2004. Exceptional points of non-Hermitian operators, *J. Phys. A: Math. Gen.* **37**, 2455–2464 Stehmann T., Heiss W. D. and Scholtz F. G. 2004. Observation of exceptional points in electronic circuits, *J. Phys. A: Math. Gen.* **37**, 7813–7819 Mondragon A. and Hernandez E. 1993. Degeneracy and crossing of resonance energy surfaces. *J. Phys. A: Math. Gen.* [**26**]{}, 5595–5611 Seyranian A. P. 1991. Sensitivity analysis of eigenvalues and the development of instability. *Stroinicky Casopic* **42**, 193–208 Seyranian A. P. 1993. Sensitivity analysis of multiple eigenvalues, [*Mech. Struct. Mach.*]{} **21**, 261–284 Seyranian A. P. and Pedersen P. 1993. On Interaction of Eigenvalue Branches in Non-Conservative Multi-Parameter Problems. *Proc. of ASME Conf. “Dynamics and Vibration of Time Varying Systems and Structures”* (eds. Sinha and Evan-Iwanowski), DE-Vol. **56**, 19–30 Seyranian A. P., Lund E. and Olhoff N. 1994. Multiple eigenvalues in structural optimization problems. *Structural and Multidisciplinary Optimization* [**8**]{}, 207–227 Mailybaev A. A. and Seyranian A. P. 1999. On singularities of a boundary of the stability domain. [*SIAM J. Matrix Anal. Appl.*]{} **21**, 106–128 Seyranian A. P. and Kliem W. 2001. Bifurcations of eigenvalues of gyroscopic systems with parameters near stability boundaries. [*Trans. ASME, J. Appl. Mech.*]{} **68**, 199–205 Seyranian A. P. and Mailybaev A. A. 2001. On stability boundaries of conservative systems. [*Z. Angew. Math. Phys.*]{} **52**, 669–679 Kirillov O. N. and Seyranian A. P. 2002. Metamorphoses of characteristic curves in circulatory systems. *J. Appl. Math. Mech.* [**66**]{}, 371-385 Seyranian A. P. and Mailybaev A. A. 2003. Interaction of eigenvalues in multi-parameter problems. [*J. Sound Vibration*]{} **267**, 1047–1064 Kirillov O. N. 2004. Destabilization paradox. *Doklady Physics* [**49**]{}, 239–245 Kirillov O. N. and Seyranian A. P. 2004. Collapse of the Keldysh chains and stability of continuous non-conservative systems. *SIAM J. Appl. Math.* [**64**]{}, 1383–1407 Seyranian A. P. and Mailybaev A. A. 2003. *Multiparameter Stability Theory with Mechanical Applications* (Singapore: World Scientific) Dobson I., Zhang J., Greene S., Engdahl H. and Sauer P. W. 2001. Is strong modal resonance a precursor to power system oscillations? *IEEE Transactions On Circuits And Systems I: Fundamental Theory And Applications* [**48**]{}, 340–349 Vishik M. I. and Lyusternik L. A. 1960. Solution of Some Perturbation Problems in the Case of Matrices and Selfadjoint or Non-selfadjoint Equations. *Russian Math. Surveys* [**15**]{}, 1–73 Courant C. and Hilbert D. 1953. *Methods of Mathematical Physics* (New York: Wiley) Schwartz L. 1967. *Cours d’Analyse* (Paris: Hermann) Lancaster P. 1969. *Theory of Matrices* (New York: Academic Press) Arnold V. I. 1983. *Geometrical Methods in the Theory of Ordinary Differential Equations* (New York: Springer) Arnold V. I. 1989. Spaces of functions with moderate critical points, *Funct. Anal. i Pril.* **23**, 1–10 Kirillov O. N., Mailybaev A. A. and Seyranian A. P. 2004. Unfolding of eigenvalue surfaces near a diabolic point due to a complex perturbation. *J. Phys. A: Math. Gen.* (submitted), ArXive:math-ph/0411006 Landau L. D., Lifshitz E. M. and Pitaevskii L. P. 1984. *Electrodynamics of continuous media* (Oxford: Pergamon)
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper we investigate the $L^p$-boundedness of certain classes of periodic pseudo-differential operators. The operators considered arise from the study of symbols on $\mathbb{T}^n\times\mathbb{Z}^n$ with limited regularity.' address: '$^{1}$ Department of Mathematics, Universidad de los Andes,Bogotá- Colombia.' author: - Duván Cardona$^1$ date: 'Received:; Revised:; Accepted:; $^{1}$ Universidad de los Andes, Bogotá - Colombia' title: 'On the boundedness of periodic pseudo-differential operators' --- Introduction ============ In this work we obtain $L^p-$boundedness theorems for pseudo-differential operators with symbols defined on $\mathbb{T}^n\times\mathbb{Z}^n.$ For some recent work on boundedness results of periodic pseudo-differential operators in $L^p$-spaces we refer the reader to [@Duvan3; @Profe; @s2; @Ruz; @Ruz-2]. Pseudo-differential operators on $\mathbb{R}^n$ are generalizations of differential operators and singular integrals. They are formally defined by $T_{\sigma}u(x)=\int e^{i2\pi\langle x,\eta \rangle}\sigma(x,\eta)\widehat{u}(\eta)d\eta.$ The function $\sigma$ is usually called the symbol of the corresponding operator $T_{\sigma}.$ Symbols are classified according to their behavior and the behavior of their derivatives. For $m\in \mathbb{R}$ and $0\leq \rho,\delta\leq 1,$ the $(\rho,\delta)-$Hörmander class $S^{m}_{\rho,\delta}(\mathbb{R}^n\times \mathbb{R}^n)$ consists of those functions which are smooth in $(x,\eta)$ and which satisfy symbols inequalities $|\partial^{\beta}_{x}\partial_{\eta}^{\alpha}\sigma(x,\eta)|\leq C_{\alpha,\beta}\langle \eta\rangle^{m-\rho|\alpha|+\delta|\beta|}.$ The corresponding set of operators with symbols in $(\rho,\delta)-$classes will be denoted by $\Psi^{m}_{\rho,\delta}(\mathbb{R}^n\times \mathbb{R}^n).$ A remarkable result due to A. P. Calderón and R. Vaillancourt, gives us that a pseudo-differential operator with symbol in $S^{0}_{\rho,\rho}(\mathbb{R}^n\times \mathbb{R}^n)$ for some $0\leq \rho<1$ is bounded on $L^2(\mathbb{R}^n),$ (see [@ca1; @ca2]). Their result is false when $\rho=1:$ there exists symbols in $S^{0}_{1,1}(\mathbb{R}^n\times \mathbb{R}^n)$ whose associated pseudo-differential operators are not bounded on $L^2(\mathbb{R}^n),$ (see [@Duo]). For general $1<p<\infty$, we have the following theorem (see Fefferman [@fe]): if $m\leq -m_{p}= -n(1-\rho)|\frac{1}{p}-\frac{1}{2}|,$ $m_{p}<n(1-\rho)/2,$ then $T_{\sigma}\in \Psi^{m}_{\rho,\delta}(\mathbb{R}^n\times \mathbb{R}^n)$ is a $L^p$- bounded operator. It is well known that for every $m>m_p$ there exists $T_{\sigma}\in \Psi^{-m}_{\rho,\delta}(\mathbb{R}^n\times \mathbb{R}^n)$ which is not bounded on $L^{p}(\mathbb{R}^n).$ The historical development of the problem about the $L^p$-boundedness of pseudo-differential operators on $\mathbb{R}^n$ can be found in [@Na3; @LW].\ \ Pseudo-differential operators with symbols in Hörmander classes can be defined on $C^{\infty}$- manifolds by using local charts. In 1979, Agranovich (see [@ag]) gives a global definition of pseudo-differential operators on the circle $\mathbb{S}^1,$ (instead of the local formulation on the circle as a manifold). By using the Fourier transform Agranovich’s definition was readily generalizable to the $n-$dimensional torus $\mathbb{T}^n.$ It is a non-trivial result, that the definition of pseudo-differential operators with symbols on $(\rho,\delta)-$classes by Agranovich and Hörmander are equivalent. This fact is known as the equivalence theorem of McLean (see [@Mc]). Important consequences of this equivalence are the following periodic versions of the Calderón-Vaillancourt theorem and the Fefferman theorem: \[TFC\] Let $0\leq \delta<\rho\leq 1$ and let $a:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ be a symbol with corresponding periodic pseudo-differential operator defined on $C^{\infty}(\mathbb{T}^n)$ by $$a(x,D)f(x)=\sum_{\xi\in\mathbb{Z}^n } e^{ i2\pi \langle x,\xi \rangle}a(x,\xi)\widehat{f}(\xi).$$ If $\alpha, \beta\in\mathbb{N}^n,$ under one of the following two conditions, - $($Calderon-Vaillancourt condition, periodic version$).$ $$|\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}a(x,\xi) |\leq C_{\alpha,\beta} \langle \xi\rangle^{\rho(|\beta|-|\alpha|)},\,\,\,(x,\xi)\in\mathbb{T}^n\times \mathbb{Z}^n,$$ - $($Fefferman condition, periodic version$).$ $$|\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}a(x,\xi) |\leq C_{\alpha,\beta} \langle \xi\rangle^{-m_{p}-\rho|\alpha|+\delta|\beta| },\,\,\,(x,\xi)\in\mathbb{T}^n\times \mathbb{Z}^n,\,\delta<1,$$ the operator $a(x,D)$ is bounded on $L^p(\mathbb{T}^n)$, for all $1<p<\infty.$ Theorem \[TFC\] is a consequence of the McLean equivalence theorem, the Calderón-Vaillancourt theorem and the Fefferman theorem. Moreover, from the proofs of these results it is easy to see that symbols with limited regularity satisfying conditions in Theorem \[TFC\] give rise to $L^p$-bounded pseudo-differential operators (under the condition $0\leq \delta<\rho\leq 1$) (see [@va]). With this in mind, in this paper we provide some new results about the $L^p$ boundedness of periodic pseudo-differential operators associated to toroidal symbols with limited regularity. Now, we first recall some recent $L^p$-theorems of periodic operators and later we present our results. [@s2 Wong-Molahajloo]. Let $a(x,D)$ be a periodic pseudo-differential operator on $\mathbb{T}^1\equiv\mathbb{S}^1 $. If $\sigma:\mathbb{T}\times\mathbb{Z}\rightarrow \mathbb{C}$ satisfies $$|\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}a(x,\xi) |\leq C_{\alpha,\beta}\langle \xi\rangle^{-|\alpha|},\,\,\,(x,\xi)\in\mathbb{T}^n\times \mathbb{Z}^n,$$ with $|\alpha|,|\beta|\leq 1,$ then, $a(x,D):L^p(\mathbb{T}^1)\rightarrow L^p(\mathbb{T}^1)$ is a bounded operator for all $1<p<\infty$. \[l2\][@Ruz Ruzhansky-Turunen] Let $n\in \mathbb{N},$ $k:=[\frac{n}{2}]+1$ and $a:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ such that $|\partial^{\beta}_{x}a(x,\xi)|\leq C_{\beta},\,\,\, |\beta|\leq k.$ Then, $a(x,D):L^2(\mathbb{T}^n)\rightarrow L^2(\mathbb{T}^n)$ is a bounded operator. We note that compared with several well-known theorems on $L^2-$boundedness of pseudo-differential operators (Calderon-Vaillancourt theorem, for example), the Theorem \[l2\] does not require any regularity with respect to the $\xi-$variable. As a consequence of real interpolation and the $L^2-$estimate by Ruzhansky and Turunen, Delgado [@Profe] established the following sharp $L^p-$theorem. \[p1\][@Profe Delgado] Let $0<\varepsilon <1$ and $k:=[\frac{n}{2}]+1,$ let $a:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ be a symbol such that $|\Delta_{\xi}^{\alpha}a(x,\xi)|\leq C_{\alpha}\langle \xi \rangle^{-\frac{n}{2}\varepsilon-(1-\varepsilon)|\alpha|},$ $|\partial_{x}^{\beta}a(x,\xi)|\leq C_{\beta}\langle \xi \rangle^{-\frac{n}{2}\varepsilon},$ for $|\alpha|,|\beta|\leq k.$ Then $a(x,D):L^p(\mathbb{T}^n)\rightarrow L^p(\mathbb{T}^n)$ is a bounded linear operator, for $2\leq p<\infty.$ In the recent paper [@DR4] the authors have considered the $L^{p}$-boundedness of pseudo-differential operators on Compact Lie groups. Although, the original theorem is valid for general compact Lie groups, we present the following periodic version. [@DR4 Delgado-Ruzhansky]\[RuzDelg\] Let $0\leq \delta,\rho\leq 1$ and $1<p<\infty.$ Denote by $\kappa$ the smallest even integer larger than $\frac{n}{2}.$ Let $\sigma(x,D)$ be a pseudo-differential operator with symbol $\sigma$ satisfying $$\label{DRR} |\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\sigma(x,\xi) |\leq C\langle \xi\rangle^{-m_0-\rho|\alpha|+\delta|\beta|},$$ for all $|\alpha|\leq \kappa$ and $|\beta|\leq [\frac{n}{p}]+1,$ where $m_{0}:=\kappa (1-\rho) |1/p-1/2|+\delta([\frac{n}{p}]+1).$ Then $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n)$ for all $1<p<\infty.$ Now, we present the main results of this paper. Our starting point is the following result which consider the problem about the $L^{p}(\mathbb{T}^n)$-boundedness for $1\leq p<\infty$, in contrast with the previous results where the case $p=1$ can be not considered because, as it is well known the boundedness of pseudo-differential operators with symbols in $(\rho,\delta)$- classes fails on $L^1.$ We denote $m_{p}:=-n(1-\rho)|\frac{1}{p}-\frac{1}{2}|,$ for all $1\leq p<\infty,$ $n\in\mathbb{N}$ and $0\leq \rho\leq 1.$\ \ **Theorem IA.**[ *Let $0<\varepsilon\leq 1,$ $0\leq \rho\leq 1$ and let $\omega:[0,\infty)\rightarrow [0,\infty) $ be a non-decreasing and bounded function such that $$\int_{0}^1\omega(t)t^{-1}dt<\infty.$$ Let $\sigma(x,\xi):\mathbb{T}^n\times \mathbb{Z}^n\rightarrow\mathbb{C}$ be a symbol satisfying $$|\partial_{x}^{\beta}\Delta^{\alpha}_{\xi}\sigma(x,\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{-\frac{n}{2}(1-\rho)-\rho|\alpha|}$$ for $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ then the corresponding periodic operator $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n)$ for all $1\leq p<\infty.$*]{}\ \ The following theorem is related with the result proved by Delgado and Ruzhansky mentioned above. We discuss such relation in Remark \[mainremark\].\ \ **Theorem IB.** [ *Let $0< \rho\leq 1, $ and $1< p<\infty.$ Let $\sigma:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ be a symbol satisfying $$\label{interpola} |\partial_{x}^\beta\Delta^{\alpha}_{\xi}\sigma(x,\xi)|\leq C\langle \xi \rangle^{-m_p-\rho|\alpha|},$$ for all $\alpha$ and $\beta$ in $\mathbb{N}^n$ such that $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ then the corresponding periodic operator $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n).$*]{}\ We observe that from classical results (e.g., Lemma 2.2 of [@Na3]) together with suitable versions of the McLean equivalence theorem (see Proposition \[eqcc\] below) the $L^p$-boundedness of a periodic operator satisfying for $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ $2\leq p<\infty$ can be proved. For the case of $\mathbb{R}^n$ it is well known that symbols with $[\frac{n}{2}]+1$-derivatives in $\xi$ satisfying $(\rho,\delta)$-estimates does not always imply $L^p$-boundedness, $1<p\leq 2.$ However, in order to assure boundedness it is suffice to consider for this case $(n+1)$-derivatives in $\xi$. However, with our approach, we only need $k:=[\frac{n}{2}]+1$-derivatives in $\xi$ for a periodic symbol improving the immediate result that we could have if we apply suitable versions of the McLean equivalence theorem (see Proposition \[eqcc\]) (together with Lemma 2.2 of [@Na3]). It is important to mention that the previous result can be not deduced from Theorem \[p1\]. In the following result we consider the boundedness of multipliers on $L^{\nu},$ $0<\nu\leq 1.$ **Theorem IC.** [ *Let $k:=[\frac{n}{2}]+1$. Let $\sigma(\xi)$ be a periodic symbol on $\mathbb{Z}^n$ satisfying $$|\Delta^{\alpha}_{\xi}\sigma(\xi)|\leq C_{\alpha}\langle \xi \rangle^{-|\alpha|}, \,\,\,|\alpha|\leq k,$$ then, the corresponding periodic operator $\sigma(D)$ is bounded from $L^1(\mathbb{T}^n)$ into $L^p(\mathbb{T}^n)$ for all $0<p<1.$*]{} Interpolation via Riesz-Thorin Theorem allow us to obtain the next $(L^p,L^r)$-theorem: **Theorem II.** [ *Let $2\leq p<\infty,$ $0<\varepsilon <1$ and $k:=[\frac{n}{2}]+1.$ Let $a:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ be a symbol such that $|\Delta_{\xi}^{\alpha}a(x,\xi)|\leq C_{\alpha}\langle \xi \rangle^{-\frac{n}{2}\varepsilon-(1-\varepsilon)|\alpha|},$ $|\partial_{x}^{\beta}a(x,\xi)|\leq C_{\beta}\langle \xi \rangle^{-\frac{n}{2}\varepsilon},$ for $|\alpha|,|\beta|\leq k$, then $\sigma(x,D):L^p(\mathbb{T}^n)\rightarrow L^r(\mathbb{T}^n)$ is a bounded linear operator for all $1<q\leq r\leq p<\infty,$ where $\frac{1}{p}+\frac{1}{q}=1.$*]{} An operator $T$ is positive if $f\geq 0$ implies that, the function $Tf$ is non-negative. In the following theorem we study positive and periodic amplitude operators (see equation for the definition of amplitude operator): **Theorem III.** [ Let $0<\varepsilon,\delta <1.$ If $a(x,y,D)$ is a positive amplitude operator with symbol satisfying the following inequalities $$|\partial_{x}^{\beta}\partial_{y}^{\alpha}a(x,y,\xi)|\leq C_{\alpha,\beta}\langle \xi\rangle^{\delta|\beta|-\varepsilon|\alpha|},\,\,\,|\alpha|,|\beta|\leq\mu:=[\frac{n}{2}(1-\delta)^{-1}]+1,$$ then $a(x,y,D)$ is bounded on $L^1(\mathbb{T}^n).$ ]{}\ Same as in Theorem \[l2\], symbols considered in Theorem III does not require any regularity condition on the Fourier variable. It is important to mention that there exists a connection between the $L^p$ boundedness of periodic operators and its continuity on Besov spaces. This relation has been studied by the author on general compact Lie groups in[@Duvan4 Section 3]. Although some results in this paper consider the $L^p$-boundedness of pseudo-differential operators on the torus (for $1\leq p<\infty$), this problem has been addressed on general compact Lie groups in the references [@DR4; @Fischer2] for all $1<p<\infty.$ Finally, we refer the reader to the references [@Ghaemi; @Ghaemi2; @Ghaemi3; @m; @Wong2] for other properties of periodic operators on $L^p$-spaces. Preliminaries ============= We use the standard notation of pseudo-differential operators (see [@f; @Hor1; @Hor2; @Ruz; @Wong]). The Schwartz space $\mathcal{S}(\mathbb{Z}^n)$ denote the space of functions $\phi:\mathbb{Z}^n\rightarrow \mathbb{C}$ such that $$\forall M\in\mathbb{R}, \exists C_{M}>0,\, |\phi(\xi)|\leq C_{M}\langle \xi \rangle^M,$$ where $\langle \xi \rangle=(1+|\xi|^2)^{\frac{1}{2}}.$ The toroidal Fourier transform is defined for any $f\in C^{\infty}(\mathbb{T}^n)$ by $\widehat{f}(\xi)=\int_{}e^{-i2\pi\langle x,\xi\rangle}f(x)dx,\,\,\xi\in\mathbb{Z}^n.$ The inversion formula is given by $f(x)=\sum_{}e^{i2\pi\langle x,\xi \rangle }\widehat{u}(\xi),\,\,x\in\mathbb{T}^n.$ The periodic Hörmander class $S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{R}^n), \,\, 0\leq \rho,\delta\leq 1,$ consists of those functions $a(x,\xi)$ which are smooth in $(x,\xi)\in \mathbb{T}^n\times \mathbb{R}^n$ and which satisfy toroidal symbols inequalities $$\label{css} |\partial^{\beta}_{x}\partial^{\alpha}_{\xi}a(x,\xi)|\leq C_{\alpha,\beta}\langle \xi \rangle^{m-\rho|\alpha|+\delta|\beta|}.$$ Symbols in $S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{R}^n)$ are symbols in $S^m_{\rho,\delta}(\mathbb{R}^n\times \mathbb{R}^n)$ (see [@Hor1; @Ruz]) of order $m$ which are 1-periodic in $x.$ If $a(x,\xi)\in S^{m}_{\rho,\delta}(\mathbb{T}^n\times \mathbb{R}^n),$ the corresponding pseudo-differential operator is defined by $$\label{hh} a(x,D)u(x)=\int_{\mathbb{T}^n}\int_{\mathbb{R}^n}e^{i2\pi\langle x-y,\xi \rangle}a(x,\xi)u(y)d\xi dy,\,\, u\in C^{\infty}(\mathbb{T}^n).$$ The set $S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n),\, 0\leq \rho,\delta\leq 1,$ consists of those functions $a(x, \xi)$ which are smooth in $x$ for all $\xi\in\mathbb{Z}^n$ and which satisfy $$\label{cs} \forall \alpha,\beta\in\mathbb{N}^n,\exists C_{\alpha,\beta}>0,\,\, |\Delta^{\alpha}_{\xi}\partial^{\beta}_{x}a(x,\xi)|\leq C_{\alpha,\beta}\langle \xi \rangle^{m-\rho|\alpha|+\delta|\beta|}.$$ The operator $\Delta_\xi^\alpha$ in is the difference operator which is defined as follows. First, if $f:\mathbb{Z}^n\rightarrow \mathbb{C}$ is a discrete function and $(e_j)_{1\leq j\leq n}$ is the canonical basis of $\mathbb{R}^n,$ $$(\Delta_{\xi_{j}} f)(\xi)=f(\xi+e_{j})-f(\xi).$$ If $k\in\mathbb{N},$ denote by $\Delta^k_{\xi_{j}}$ the composition of $\Delta_{\xi_{j}}$ with itself $k-$times. Finally, if $\alpha\in\mathbb{N}^n,$ $\Delta^{\alpha}_{\xi}= \Delta^{\alpha_1}_{\xi_{1}}\cdots \Delta^{\alpha_n}_{\xi_{n}}.$ The toroidal operator (or periodic operator) with symbol $a(x,\xi)$ is defined as $$\label{aa} a(x,D)u(x)=\sum_{\xi\in\mathbb{Z}^n}e^{i 2\pi\langle x,\xi\rangle}a(x,\xi)\widehat{u}(\xi),\,\, u\in C^{\infty}(\mathbb{T}^n).$$ There exists a process to interpolate the second argument of symbols on $\mathbb{T}^n\times \mathbb{Z}^n$ in a smooth way to get a symbol defined on $\mathbb{T}^n\times \mathbb{R}^n.$ \[eq\] Let $0\leq \delta \leq 1,$ $0< \rho\leq 1.$ The symbol $a\in S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n)$ if only if there exists a Euclidean symbol $a'\in S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{R}^n)$ such that $a=a'|_{\mathbb{T}^n\times \mathbb{Z}^n}.$ The proof can be found in [@Mc; @Ruz]. It is a non trivial fact, however, that the definition of pseudo-differential operator on a torus given by Agranovich (equation \[aa\] ) and Hörmander (equation \[hh\]) are equivalent. McLean (see [@Mc]) prove this for all the Hörmander classes $S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n).$ A different proof to this fact can be found in [@Ruz], Corollary 4.6.13. $\label{eqc}($Equality of Operators Classes$).$ For $0\leq \delta \leq 1,$ $0<\rho\leq 1$ we have $\Psi^{m}_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n)=\Psi^{m}_{\rho,\delta}(\mathbb{T}^n\times \mathbb{R}^n).$ A look at the proof (based in Theorem 4.5.3 of [@Ruz]) of the Proposition \[eqc\] shows us that a more general version is still valid for symbols with limited regularity as follows (see Corollary 4.5.7 of [@Ruz]): \[eqc’\] Let $0\leq \delta \leq 1,$ $0\leq \rho<1.$ Let $a:\mathbb{T}^n\times \mathbb{R}^n\rightarrow \mathbb{C}$ satisfying $(\rho,\delta)-$inequalities for $|\alpha|\leq N_{1}$ and $|\beta|\leq N_{2}.$ Then the restriction $\tilde{a}=a|_{\mathbb{T}^n\times \mathbb{Z}^n}$ satisfies $(\rho,\delta)-$estimates for $|\alpha|\leq N_{1}$ and $|\beta|\leq N_{2}.$ The converse holds true, i.e, if every symbol on $\mathbb{T}^n\times \mathbb{Z}^n$ satisfying $(\rho,\delta)$-inequalities (as in ) is the restriction of a symbol on $\mathbb{T}^n\times \mathbb{R}^n$ satisfying $(\rho,\delta)$ inequalities as in . Let us denote $\Psi^{m}_{\rho,\delta,N_1,N_2}(\mathbb{T}^n\times \mathbb{Z}^n)$ to the set of operators associated to symbols satisfying for all $|\alpha|\leq N_{1}$ and $|\beta|\leq N_2,$ and $\Psi^{m}_{\rho,\delta,N_1,N_2}(\mathbb{T}^n\times \mathbb{R}^n)$ defined similarly. Then we have (see Theorem 2.14 of [@Profe]): $\label{eqcc}($Equality of Operators Classes$).$ For $0\leq \delta \leq 1,$ $0<\rho\leq 1$ we have $\Psi^{m}_{\rho,\delta,N_1,N_2}(\mathbb{T}^n\times \mathbb{Z}^n)=\Psi^{m}_{\rho,\delta,N_1,N_2}(\mathbb{T}^n\times \mathbb{R}^n).$ The toroidal calculus is closed under adjoint operators. The corresponding announcement is the following. \[ad\] Let $0\leq \delta <\rho\leq 1.$ Let $a(x,D)$ be a operator with symbol $a(X,\xi)\in S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n).$ Then, the adjoint $a^{*}(x,D)$ of $a(x,D),$ has symbol $a^{*}(x,\xi)\in S^m_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n).$ The symbol $a^{*}(x,\xi)$ has the following asymptotic expansion: $$\sigma^{*}(x,\xi)\approx \sum_{\alpha \geq 0}\frac{1}{\alpha !}\Delta^{\alpha}_{\xi}D^{(\alpha)}_{x}\overline{\sigma(x,\xi)}.$$ Moreover, if $n_{0}\in \mathbb{N},$ then $$\sigma^{*}(x,\xi)- \sum_{|\alpha|<n_{0} }\frac{1}{\alpha !}\Delta^{\alpha}_{\xi}D^{(\alpha)}_{x}\overline{\sigma(x,\xi)}\,\,\, \in S^{m-n_{0}(\rho-\delta)}_{\rho,\delta}(\mathbb{T}^n\times \mathbb{Z}^n).$$ In order to establish our result on positive operators, we introduce amplitude operators. The periodic amplitudes are functions $a(x,y,\xi)$ defined on $\mathbb{T}^n\times\mathbb{T}^n\times\mathbb{Z}^n.$ The corresponding amplitude operators are defined as $$\label{amplitude}a(x,y,D)u(x)=\int\limits_{\mathbb{T}^n\times \mathbb{R}^n} e^{i2\pi \langle x-y,\xi \rangle}a(x,y,\xi)u(y)dy\,d\xi.$$ If the symbol depends only on $(x,\xi)-$variables then $p(x,y,D)=p(x,D).$ Moreover, if a symbol $\sigma(x,\xi)=\sigma(\xi)$ depends only on the Fourier variable $\xi,$ the corresponding pseudo-differential operator $\sigma(x,D)=\sigma(D)$ is called a Fourier multiplier. An instrumental result on Fourier multipliers in the proof of our main results is the following: (see, Theorem 3.8 of Stein [@Eli]). \[stein\] Suppose $1\leq p\leq \infty$ and $T_\sigma$ be a Fourier multiplier on $ \mathbb{R}^n$ with symbol $\sigma(\xi)$. If $\sigma(\xi)$ is continuous at each point of $\mathbb{Z}^n$ then the periodic operator defined by $$\sigma(D)f(x)=\sum_{\xi \in \mathbb{Z}^n}e^{i2\pi \langle x,\xi\rangle}\sigma(\xi)\widehat{u}(\xi),$$ is a bounded operator from $L^p(\mathbb{T}^n)$ into $L^p(\mathbb{T}^n).$ The following results will help clarify the nature of the conditions that will be imposed on periodic symbols in order to obtain $L^p$ theorems for periodic operators with symbols of limited regularity. \[remi1\] Let $0\leq \rho\leq 1$ and $0<\varepsilon \leq 1,$ and suppose that the symbol $\sigma(x,\xi)$ on $\mathbb{R}^n\times \mathbb{R}^n$ satisfies $$|\partial_{\xi}^{\alpha}\sigma(x,\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon} )\langle \xi \rangle^{-\frac{n}{2}(1-\rho)-\rho|\alpha|},\,\,\,\,|\alpha|\leq [\frac{n}{2}]+1,$$ where $\omega$ is a non-decreasing, bounded and non-negative function on $[0,\infty)$ satisfying $$\int_{0}^{1}\omega(t)t^{-1}dt <\infty.$$ Then $T_{\sigma}$ is a bounded operator on $L^p(\mathbb{R}^n)$ for all $2\leq p\leq \infty.$ See Theorem 4.4 and Corollary 4.2 of [@va]. The following is a version of the Fefferman theorem but symbols are considered with limited smoothness. \[remi2\] Let $2\leq p<\infty$ and $0\leq \delta\leq \rho\leq 1,$ $\delta<1.$ Let $\sigma(x,\xi)$ be a symbol satisfying $$|\partial_{x}^{\beta}\partial_{\xi}^{\alpha}\sigma(x,\xi)|\leq C\langle \xi \rangle^{-m_{p}-\rho|\alpha|+\delta|\beta|},\,\,\,\,|\alpha|,|\beta|\leq [\frac{n}{2}]+1,$$ where $m_{p}=n(1-\rho)|\frac{1}{p}-\frac{1}{2}|.$ Then $T_\sigma:L^p(\mathbb{R}^n)\rightarrow L^p(\mathbb{R}^n)$ is a bounded operator. See Theorem 5.1 and Corollary 5.1 of [@va]. The following theorem is the particular case of one proved in [@Ruz3] for compact Lie groups. \[ruz3\] Let $k>\frac{n}{2}$ be an even integer. If $\sigma(\xi)$ is a periodic symbol on $\mathbb{Z}^n$ satisfying $$|\Delta^{\alpha}_{\xi}\sigma(\xi)|\leq C_{\alpha}\langle \xi \rangle^{-|\alpha|}, \,\,\,|\alpha|\leq k,$$ then, the corresponding periodic operator $\sigma(D)$ is of weak type (1,1) and $L^p$-bounded for all $1<p<\infty.$ Also, weak(1,1) boundedness of periodic operators has been considered by the author in [@Duvan3]. We end this section with the following result proved in [@DR4]. Although, the original theorem is valid for general compact Lie groups, we present the periodic version for simplicity. \[Ruz-Delg\] Let $0\leq \delta,\rho\leq 1$ and $1<p<\infty.$ Denote by $\kappa$ the smallest even integer larger than $\frac{n}{2}.$ Let $\sigma(x,D)$ be a pseudo-differential operator with symbol $\sigma$ satisfying $$\label{DRR} |\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\sigma(x,\xi) |\leq C\langle \xi\rangle^{-m_0-\rho|\alpha|+\delta|\beta|},$$ for all $|\alpha|\leq \kappa$ and $|\beta|\leq [\frac{n}{p}]+1,$ where $m_{0}:=\kappa (1-\rho) |1/p-1/2|+\delta([\frac{n}{p}]+1).$ Then $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n)$ for all $1<p<\infty.$ Main Results-Proofs =================== In this section we prove our main results. we discuss that conditions on the periodic symbol $\sigma(x,\xi)$ guarantee the $L^p-$boundedness of the corresponding pseudo-differential operator.\ \[LE1\] Let $0<\varepsilon\leq 1,$ $0\leq \rho\leq 1, $ $1\leq p\leq\infty$ and $\omega:[0,\infty)\rightarrow [0,\infty) $ be a non-decreasing function such that $$\int_{0}^1\omega(t)t^{-1}dt<\infty.$$ If $\sigma_{1}: \mathbb{R}^n\rightarrow \mathbb{C}$ is a symbol satisfying $$|\partial^{\alpha}_{\xi}\sigma_{1}(\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{-\frac{n}{2}(1-\rho)-\rho|\alpha|}$$ for all $\alpha$ with $|\alpha|\leq [\frac{n}{2}]+1,$ then the corresponding periodic operator $\sigma(D)$ with symbol $\sigma(\xi)=\sigma_{1}(\xi)|_{ \mathbb{Z}^n}$ is a bounded operator on $L^p(\mathbb{T}^n).$ Proposition \[remi1\] provides the $L^p(\mathbb{R}^n)$-boundedness of $T_{\sigma_1}.$ Now, by Proposition \[stein\], the pseudo-differential operator with symbol $\sigma(\xi)$ is $L^{p}(\mathbb{T}^n)-$bounded. \[IA\] Let $0<\varepsilon\leq 1,$ $0\leq \rho\leq 1, $ $1\leq p<\infty$ and $\omega:[0,\infty)\rightarrow [0,\infty) $ be a bounded and non-decreasing function such that $$\int_{0}^1\omega(t)t^{-1}dt<\infty.$$ If $\sigma:\mathbb{T}^n\times \mathbb{R}^n\rightarrow \mathbb{C}$ is a symbol satisfying $$|\partial^{\alpha}_{\xi} \partial_{x}^\beta\sigma(x,\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{-\frac{n}{2}(1-\rho)-\rho|\alpha|}$$ for $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ then the corresponding periodic operator $\sigma(x,D)$ is a bounded linear operator on $L^p(\mathbb{T}^n).$ Let us consider the Schwartz kernel $K(x,z)$ of $\sigma(x,D)$ which is given by $K(x,y)=r(x-y,x)$ where $ r(x,z)=\int_{\mathbb{R}^n}e^{i2\pi \langle x,\xi \rangle }\sigma(z,\xi), $ is understood in the distributional sense. For every $z\in\mathbb{T}^n$ fixed, $r(z)(\cdot)=r(\cdot,z)$ is a distribution and the map $f\mapsto f\ast r(\cdot,z)=f\ast r(z)(\cdot)$ is a pseudo-differential operator with symbol $\sigma_{z}:\xi\mapsto \sigma(z,\xi).$ If $x\in\mathbb{T}^n$ and $f\in C^{\infty}(\mathbb{T}^n),$ $\sigma(x,D)f(x)=(f\ast r(x))(x).$ Moreover, for any $\beta\in \mathbb{N}^n$ with $|\beta|\leq [\frac{n}{p}]+1,$ the pseudo-differential operator $f\ast [\partial^{\beta}_{x}r(\cdot,x)]|_{x=z}:=f\ast [\partial^{\beta}_{z}r(\cdot,z)]$ is a pseudo-differential operator with symbol $[\partial^{\beta}_{x}\sigma(\cdot,x)]|_{x=z}:=\partial^{\beta}_{z}\sigma(z,\cdot).$ By Lemma \[LE1\], every pseudo-differential operator $\sigma_{z,\beta}(D)$ with symbol $\sigma_{z,\beta}(\xi)=\partial^{\beta}_{z}\sigma(z,\xi)$ is $L^p(\mathbb{T}^n)$-bounded for $1\leq p< \infty$. Now, by the Sobolev embedding theorem, for $1\leq p<\infty,$ $$\begin{aligned} |\sigma_(x,D)f(x)|^p &\leq \sup_{z\in\mathbb{T}^n} |(f\ast r(z))(x)|^p \leq \sum_{|\beta|\leq [\frac{n}{p}]+1}\int_{\mathbb{T}^n}| (f\ast \partial^{\beta}_{z}r(z))(x) |^pdz\end{aligned}$$ Hence, by application of the Fubini theorem we get $$\begin{aligned} \Vert \sigma(x,D)f \Vert^{p}_{L^p(\mathbb{T}^n)} & \leq C^p \sum_{|\beta|\leq [\frac{n}{p}]+1}\int_{\mathbb{T}^n}\int_{\mathbb{T}^n}| (f\ast \partial^{\beta}_{z}r(z))(x) |^pdzdx\\ & \leq C^p\sum_{|\beta|\leq [\frac{n}{p}]+1}\int_{\mathbb{T}^n}\int_{\mathbb{T}^n}| (f\ast \partial^{\beta}_{z}r(z))(x) |^pdxdz\\ &\leq C^p \sum_{|\beta|\leq [\frac{n}{p}]+1 }\sup_{z\in\mathbb{T}^n} \int_{\mathbb{T}^n}| (f\ast \partial^{\beta}_{z}r(z))(x) |^pdx\\&= C^p\sum_{|\beta|\leq [\frac{n}{p}]+1 }\sup_{z\in\mathbb{T}^n}\Vert \sigma_{z,\beta}(D)f \Vert^p_{L^p(\mathbb{T}^n)} \\ &\leq C^p\left( \sum_{|\beta|\leq [\frac{n}{p}]+1 }\sup_{z\in\mathbb{T}^n}\Vert \sigma_{z,\beta}(D) \Vert^p_{B(L^p,L^p)} \right)\Vert f\Vert^p_{L^p(\mathbb{T}^n)}.\end{aligned}$$ Thus, $$\Vert \sigma(x,D)f \Vert_{L^p(\mathbb{T}^n)}\leq C \left( \sum_{|\beta|\leq [\frac{n}{p}]+1 }\sup_{z\in\mathbb{T}^n}\Vert \sigma_{z,\beta}(D) \Vert^p_{B(L^p,L^p)} \right)^{\frac{1}{p}}\Vert f \Vert_{L^p(\mathbb{T}^n)}.$$ \[lemmaIA\] Let $k\in\mathbb{R},$ $\varepsilon>0,$ $\omega$ be a function as in Proposition \[remi1\] and $\sigma:\mathbb{T}^n\times \mathbb{R}^n\rightarrow \mathbb{C}$ be a symbol satisfying $$\label{w1} |\partial^{\alpha}_{\xi} \partial_{x}^\beta\sigma(x,\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{k},$$ for all $|\alpha|\leq N_1$ and $|\beta|\leq N_2.$ Let $\tilde{a}(x,\xi):=\sigma(x,\xi)|_{\mathbb{T}^{n}\times \mathbb{Z}^n}.$ Then $$\label{w2} |\Delta^{\alpha}_{\xi} \partial_{x}^\beta\tilde{\sigma}(x,\xi)|\leq C'\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{k},$$ for all $|\alpha|\leq N_1$ and $|\beta|\leq N_2.$ Moreover, every symbol satisfying is the restriction of a symbol on $\mathbb{T}^n\times \mathbb{R}^n$ satisfying . Let us consider $\sigma$ as in . By the mean value theorem, if $|\alpha|=1$ we have $$\begin{aligned} \Delta_{\xi}^{\alpha}\partial^{\beta}_{x}\tilde{\sigma}(x,\xi)=\Delta_{\xi}^{\alpha}\partial^{\beta}_{x}{\sigma}(x,\xi)=\partial_{\xi}^{\alpha}\partial^{\beta}_{x}{\sigma}(x,\xi)|_{\xi=\eta}\end{aligned}$$ where $\eta$ is on the line $[\xi,\xi+\alpha].$ For a general multi-index $\alpha\in\mathbb{N}^n,$ it can proved by induction that $$\begin{aligned} \Delta_{\xi}^{\alpha}\partial^{\beta}_{x}\tilde{\sigma}(x,\xi)=\partial_{\xi}^{\alpha}\partial^{\beta}_{x}{\sigma}(x,\xi)|_{\xi=\eta}\end{aligned}$$ for some $\eta\in Q:=[\xi_{1}\times \xi_{1}+\alpha_1]\times \cdots [\xi_{n}\times \xi_{n}+\alpha_n]. $ Hence, we have $$\begin{aligned} |\Delta_{\xi}^{\alpha}\partial_{x}^{\beta}\tilde{\sigma}(x,\xi)| &=|\partial_{\xi}^{\alpha}\partial^{\beta}_{x}{\sigma}(x,\xi)|_{\xi =\eta\in Q}|\\ &\leq C\omega(\langle \eta \rangle^{-\varepsilon})\langle \eta \rangle^{k}\leq C'\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{k}.\end{aligned}$$ So, we have proved the first part of the theorem. Now, let us consider a symbol $\tilde{\sigma}$ on $\mathbb{T}^n\times \mathbb{Z}^n$ satisfying . Let us consider $\theta$ as in Lemma 4.5.1 of [@Ruz]. Define the symbol $\sigma$ on $\mathbb{T}^n\times \mathbb{R}^n$ by $$\sigma(x,\xi)=\sum_{\eta\in\mathbb{Z}^n}(\mathcal{F}_{\mathbb{R}^n}\theta)(\xi-\eta)\tilde{\sigma}(x,\eta).$$ Same as in the proof of Theorem 4.5.3 of [@Ruz], pag. 359, we have $$\begin{aligned} |\partial_{x}^{\beta}\partial_{\xi}^{\alpha}\sigma(x,\xi)|=\left| \sum_{\eta\in\mathbb{Z}^n}\phi_{\alpha}(\xi-\eta)\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\tilde{\sigma}(x,\eta) \right|\end{aligned}$$ where every $\phi_{\alpha}\in\mathcal{S}(\mathbb{R}^n)$ is a function as in Lemma 4.5.1 of [@Ruz]. So, we obtain $$|\partial_{x}^{\beta}\partial_{\xi}^{\alpha}\sigma(x,\xi)|\leq \sum_{\eta\in\mathbb{Z}^n}|\phi_{\alpha}(\eta)\omega(\langle \xi-\eta \rangle^{-\varepsilon})\langle \xi-\eta\rangle^{k}|$$ Since $\omega$ is bounded, for some $M>0,$ and all $\alpha\geq 1$ we have $$\label{omega}|\omega(\alpha t)|\leq \alpha\omega(t),\,\,t>M.$$ By , the Peetre inequality and using that $\omega$ is increasing we have $$\begin{aligned} |\partial_{x}^{\beta}\partial_{\xi}^{\alpha}\sigma(x,\xi)| &\leq \sum_{\eta\in\mathbb{Z}^n}|\phi_{\alpha}(\eta)\omega(\langle \xi-\eta \rangle^{-\varepsilon})\langle \xi-\eta\rangle^{k} |\\ &\leq \sum_{\eta\in\mathbb{Z}^n}|\phi_{\alpha}(\eta)\omega(\langle \xi\rangle^{-\varepsilon}\langle\eta \rangle^{\varepsilon})\langle \xi\rangle^{k}\langle \eta\rangle^{k} |\\ &\leq C \omega(\langle \xi\rangle^{-\varepsilon})\langle \xi\rangle^{k}\sum_{\eta\in\mathbb{Z}^n}|\phi_{\alpha}(\eta)\langle\eta \rangle^{\varepsilon+k}|.\end{aligned}$$ Since every $\phi_{\alpha}$ is a function in the Schwartz class we obtain $$|\partial_{x}^{\beta}\partial_{\xi}^{\alpha}\sigma(x,\xi)| \leq C'\omega(\langle \xi\rangle^{-\varepsilon})\langle \xi\rangle^{k}.$$ So, we end the proof. As a consequence of the results above we obtain the following result.\ \ **Theorem IA.**[ *Let $0<\varepsilon\leq 1,$ $0\leq \rho\leq 1$ and let $\omega:[0,\infty)\rightarrow [0,\infty) $ be a non-decreasing and bounded function such that $$\int_{0}^1\omega(t)t^{-1}dt<\infty.$$ Let $\sigma(x,\xi):\mathbb{T}^n\times \mathbb{Z}^n\rightarrow\mathbb{C}$ be a symbol satisfying $$|\partial_{x}^{\beta}\Delta^{\alpha}_{\xi}\sigma(x,\xi)|\leq C\omega(\langle \xi \rangle^{-\varepsilon})\langle \xi \rangle^{-\frac{n}{2}(1-\rho)-\rho|\alpha|}$$ for $\alpha,\beta\in\mathbb{N}^n$ such that $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ then the corresponding periodic operator $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n)$ for all $1\leq p<\infty.$*]{}\ \ First, let us denote by $\Psi^{m,\omega}_{\rho,\delta, N_1,N_2}(\mathbb{T}^n\times \mathbb{R}^n)$ and $\Psi^{m,\omega}_{\rho,\delta, N_1,N_2}(\mathbb{T}^n\times \mathbb{Z}^n)$ to the set of operators with symbols satisfying and respectively. If we combine the Theorem 4.6.12 of [@Ruz] and Lemma \[lemmaIA\] we obtain the equality of classes $$\Psi^{m,\omega}_{\rho,\delta, N_1,N_2}(\mathbb{T}^n\times \mathbb{R}^n)=\Psi^{m,\omega}_{\rho,\delta, N_1,N_2}(\mathbb{T}^n\times \mathbb{Z}^n).$$ Hence, in order to proof the boundedness of $\sigma(x,D)$ we only need to proof that $$\Psi^{-\frac{n}{2}(1-\rho),\,\omega}_{\rho,0, [\frac{n}{2}]+1,[\frac{n}{p}]+1}(\mathbb{T}^n\times \mathbb{R}^n)\subset\mathcal{L}(L^{p}(\mathbb{T}^n)),$$ but, this fact has been proved in Theorem \[IA\]. \[LE2\] Let $0\leq \rho\leq 1, $ and $1< p<\infty.$ Let $\sigma_{1}: \mathbb{R}^n\rightarrow \mathbb{C}$ be a symbol satisfying $$|\partial^{\alpha}_{\xi}\sigma_{1}(\xi)|\leq C\langle \xi \rangle^{-m_p-\rho|\alpha|}$$ for all $\alpha$ and $\beta$ with $|\alpha|\leq [\frac{n}{2}]+1,$ then the corresponding periodic operator $\sigma(D)$ with symbol $\sigma(\xi)=\sigma_{1}(\xi)|_{ \mathbb{Z}^n}$ is a bounded operator on $L^p(\mathbb{T}^n).$ From Proposition \[remi2\], we deduce the $L^p(\mathbb{R}^n)-$boundedness of $T_{\sigma_1}.$ Finally, by Proposition \[stein\], $\sigma(D)$ is a bounded operator on $L^{p}(\mathbb{T}^n).$ **Theorem IB.** [ *Let $0< \rho\leq 1, $ and $1< p<\infty.$ Let $\sigma:\mathbb{T}^n\times \mathbb{Z}^n\rightarrow \mathbb{C}$ be a symbol satisfying $$\label{cardona} |\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\sigma(x,\xi) |\leq C\langle \xi\rangle^{-m_p-\rho|\alpha|},$$ for all $\alpha$ and $\beta$ with $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1,$ then $\sigma(x,D)$ extends to a bounded operator on $L^p(\mathbb{T}^n).$*]{}\ By Proposition \[eqcc\] we only need to proof that $$\Psi^{-m_p}_{\rho,0,[\frac{n}{2}]+1,[\frac{n}{p}]+1}(\mathbb{T}^n\times \mathbb{R}^n)\subset \mathcal{L}(L^{p}(\mathbb{T}^n)),\,\,1<p<\infty.$$ Hence, let us consider a symbol $\sigma_{1}$ on $\mathbb{T}^n\times \mathbb{R}^n$ satisfying $$\label{interpolac} |\partial_{x}^\beta\partial^{\alpha}_{\xi}\sigma_{1}(x,\xi)|\leq C\langle \xi \rangle^{-m_p-\rho|\alpha|},$$for all $|\alpha|\leq [\frac{n}{2}]+1,$ $|\beta|\leq [\frac{n}{p}]+1.$ From Lemma \[LE2\], we obtain the $L^{p}-$boundedness of every operator ${\sigma_1}_{z,\beta}(D)$ with symbol ${\sigma_1}_{z,\beta}(\xi)=(\partial_{x}^{\beta}\sigma_{1}(x,\xi))|_{x=z},$ $z\in\mathbb{T}^n.$ As in the proof of Theorem IA, an application of the Sobolev embedding theorem gives $$\Vert \sigma_{1}(x,D)f \Vert_{L^p(\mathbb{T}^n)}\leq C \left( \sum_{|\beta|\leq [\frac{n}{p}]+1 }\sup_{z\in\mathbb{T}^n}\Vert {\sigma_1}_{z,\beta}(D) \Vert^p_{B(L^p,L^p)} \right)^{\frac{1}{p}}\Vert f \Vert_{L^p(\mathbb{T}^n)}.$$ Hence $\sigma_1(x,D)$ is $L^p$-bounded. So we end the proof. \[mainremark\] We observe that conditions on the number of derivatives in Theorem \[Ruz-Delg\] and Theorem IB agree, but the behavior of the symbol derivatives in every theorem is different. In fact, can be written as $$|\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\sigma(x,\xi) |\leq C\langle \xi\rangle^{-n(1-\rho)|1/p-1/2|-\rho|\alpha|},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,$$ in contrast with : $$|\partial_{x}^{\beta}\Delta_{\xi}^{\alpha}\sigma(x,\xi) |\leq C\langle \xi\rangle^{-\kappa(1-\rho)|1/p-1/2|-\rho|\alpha|+\delta(|\beta|-([\frac{n}{p}]+1))},$$ where $|\beta|-([\frac{n}{p}]+1)\leq 0.$ \[Kol1\] $($Kolmogorov’s lemma$).$ Given an operator $S$ which is weak(1,1), $0<v<1,$ and a set $E$ of finite measure, there exists a $C>0$ such that $$\int_{E}|Sf(x)|^v\,dx\leq C\mu (E)^{1-v}\Vert f\Vert^v_{L^1(\mathbb{R}^n)}.$$ For the proof of this result, see Lemma 5.16 of [@Duo]. **Theorem IC.** [ *Let $k:=[\frac{n}{2}]+1$. If $\sigma(\xi)$ is a periodic symbol on $\mathbb{Z}^n$ satisfying $$|\Delta^{\alpha}_{\xi}\sigma(\xi)|\leq C_{\alpha}\langle \xi \rangle^{-|\alpha|}, \,\,\,|\alpha|\leq k,$$ then, the corresponding periodic operator $\sigma(D)$ is bounded from $L^1(\mathbb{T}^n)$ into $L^p(\mathbb{T}^n)$ for all $0<p<1.$*]{} By Theorem \[eqcc\] we only need to prove that an operator $\sigma_{1}(D)$ with symbol $\sigma_{1}$ on $\mathbb{T}^n\times \mathbb{R}^n$ satisfying $$|\partial^{\alpha}_{\xi}\sigma_1(\xi)|\leq C_{\alpha}\langle \xi \rangle^{-|\alpha|}, \,\,\,|\alpha|\leq k,$$ is a bounded operator from $L^1(\mathbb{T}^n)$ into $L^p(\mathbb{T}^n)$ for all $0<p<1.$ The multiplier $\sigma(D)$ is of weak type (1,1) on $\mathbb{R}^n$ (see [@Ste; @Eli]). Now, by Kolmogorov’s lemma with $E=\mathbb{T}^n,$ for $f\in L^{p}(\mathbb{T}^n)$ (which can be considered as a function on $\mathbb{R}^n$ equal to zero in $\mathbb{R}^n-\mathbb{T}^n$) we have $$\int_{\mathbb{T}^n}|\sigma_{1}(x,D)f(x)|^p\,dx\leq C\Vert f\Vert^p_{L^1(\mathbb{R}^n)}=C\Vert f\Vert^p_{L^1(\mathbb{T}^n)}$$ which proves the boundedness of $\sigma_{1}(x,D).$ So, we end the proof. \[lemma22\] Let $2\leq p<\infty,$ and $k=[\frac{n}{2}]+1,$ let $p(x,\xi)$ be a symbol such that $$|\partial_{x}^{\beta}p(x,\xi)|\leq C_{k},\,\,|\beta |\leq k.$$ Then $p(x,D):L^p(\mathbb{T}^n)\rightarrow L^q(\mathbb{T}^n)$ is a bounded operator, where $\frac{1}{p}+\frac{1}{q}=1.$ By the definition of periodic pseudo-differential operator, integration by parts and inversion formula, we have $$\begin{aligned} \langle p(x,D_{x})u, g \rangle_{L^2(\mathbb{T}^n)} &=\int_{T^n}\sum_{\xi\in\mathbb{Z}^n}e^{i2\pi x\xi}p(x,\xi)\widehat{u}(\xi)g(x)dx\\ &=\int_{T^n}\sum_{\xi\in\mathbb{Z}^n}\sum_{\eta\in\mathbb{Z}^n}e^{i2\pi x(\xi-\eta)}p(x,\xi)\widehat{u}(\xi)\overline{\widehat{g}(\eta)}dx\\ &=\sum_{\xi\in\mathbb{Z}^n}\sum_{\eta\in\mathbb{Z}^n}\int_{T^n}e^{i2\pi x(\xi-\eta)}p(x,\xi)\widehat{u}(\xi)\overline{\widehat{g}(\eta)}dx\\ &=\sum_{\xi\in\mathbb{Z}^n}\sum_{\eta\in\mathbb{Z}^n}\int_{T^n}\langle \xi-\eta\rangle^{-2k}e^{i2\pi x(\xi-\eta)}L^{k}_{x}p(x,\xi)\widehat{u}(\xi)\overline{\widehat{g}(\eta)}dx,\end{aligned}$$ where $L^q_{x}=(I-\frac{1}{4\pi^2}\mathcal{L}_{x})^q$ and $\mathcal{L}_{x}$ is the Laplacian in $x-$variables. Using the Young Inequality, we get $$\begin{aligned} |\langle p(x,D_{x})u, g \rangle_{L^2(\mathbb{T}^n)}| &\leq \sum_{\xi\in\mathbb{Z}^n}\sum_{\eta\in\mathbb{Z}^n}\langle \xi-\eta\rangle^{-2k}|L^{k}_{x}p(x,\xi)||\widehat{u}(\xi)||\widehat{g}(\eta)|\\ &\leq \sum_{\xi\in\mathbb{Z}^n}\sum_{\eta\in\mathbb{Z}^n}\langle \xi-\eta\rangle^{-2k}C_{k}\widehat{u}(\xi)||\widehat{g}(\eta)|\\ &= \sum_{\xi\in\mathbb{Z}^n}|\widehat{u}(\xi)|\sum_{\eta\in\mathbb{Z}^n}\langle \xi-\eta\rangle^{-2k}|\widehat{g}(\eta)|C_{k}\\ &= \sum_{\xi\in\mathbb{Z}^n}C_{k}|\widehat{u}(\xi)|\langle \cdot\rangle^{-2q}*|\widehat{g}(\cdot)|(\xi)C_{k}\\ &\leq C_{k} \Vert \widehat{u}\Vert_{L^2(\mathbb{Z}^n)} \Vert\langle \cdot\rangle^{-2k}|\widehat{g}| \Vert_{L^2(\mathbb{Z}^n)}\\ &\leq C_{k}\Vert u\Vert_{L^2(\mathbb{T}^n)} \Vert\langle \cdot\rangle^{-2k} \Vert_{L^1(\mathbb{Z})}\Vert{g} \Vert_{L^2(\mathbb{T}^n)}\end{aligned}$$ If $p\geq 2$ then the inclusion map $i:L^p(\mathbb{T}^n)\rightarrow L^2(\mathbb{T}^n)$ is continuous. So, for some $C>0$ we have $\Vert \cdot\Vert_{L^2(\mathbb{T}^n)}\leq C\Vert \cdot\Vert_{L^p(\mathbb{T}^n)}.$ So, we get $$|\langle p(x,D_{x})u, g \rangle_{L^2(\mathbb{T}^n)}| \leq CC_{k} \Vert\langle \cdot\rangle^{-2k} \Vert_{L^1(\mathbb{Z})} \Vert u \Vert_{L^p(\mathbb{T}^n)} \Vert g \Vert_{L^p(\mathbb{T}^n)}.$$ Finally, $$\begin{aligned} \Vert p(x,D_{x})u \Vert_{L^q(\mathbb{T}^n)} &=\sup \{ |\langle p(x,D_{x})u,g \rangle|:{\Vert g \Vert_{L^{p}(\mathbb{T}^n)}}\leq 1 \} \\ &\leq CC_{k}\Vert\langle \cdot\rangle^{-2k} \Vert_{L^1(\mathbb{Z})} \Vert u \Vert_{L^p(\mathbb{T}^n)} .\end{aligned}$$ By using the previous lemma we can prove the following theorem on $(L^p,L^r)$-boundedness of periodic operators.\ \ **Theorem II.** [ *Let $2\leq p<\infty.$ If $\sigma(x,\xi)$ satisfies the hypotheses of Theorem \[p1\], then $\sigma(x,D):L^p(\mathbb{T}^n)\rightarrow L^r(\mathbb{T}^n)$ is a bounded linear operator for all $1<q\leq r\leq p<\infty,$ where $\frac{1}{p}+\frac{1}{q}=1.$*]{} Theorem \[p1\] implies that, $\sigma(x,D):L^p(\mathbb{T}^n)\rightarrow L^p(\mathbb{T}^n)$ is a bounded operator for $1<p<\infty.$ That $\sigma(x,D):L^p(\mathbb{T}^n)\rightarrow L^q(\mathbb{T}^n)$ is a bounded operator is a consequence of Lemma \[lemma22\]. Now, by Riesz-Thorin interpolation theorem we deduce that $\sigma(x,D):L^p(\mathbb{T}^n)\rightarrow L^r(\mathbb{T}^n)$ is a bounded operator for all $q\leq r\leq p<\infty,$ with $2\leq p<\infty.$ In the following theorem we analyze the boundedness of periodic amplitude operators on $L^1.$\ \ **Theorem III.** [ Let $0<\varepsilon,\delta <1.$ If $a(x,y,D)$ is a positive amplitude operator with symbol satisfying the following inequalities $$|\partial_{x}^{\beta}\partial_{y}^{\alpha}a(x,y,\xi)|\leq C_{\alpha,\beta}\langle \xi\rangle^{\delta|\beta|-\varepsilon|\alpha|},\,\,\,|\alpha|,|\beta|\leq[\frac{n}{2}(1-\delta)^{-1}]+1,$$ then $a(x,y,D)$ is bounded on $L^1(\mathbb{T}^n).$ ]{}\ If the operator $a(x,y,D)$ is positive and $f\geq 0$ we have that $$a(x,y,D)f\geq 0.$$ Hence, $$\begin{aligned} |a & (x,y,D)f(x)|=a(x,y,D)f(x)\\ &=\int\limits_{\mathbb{T}^n \times \mathbb{R}^n}e^{i2\pi (x-y)\xi}a(x,y,\xi)f(y)d\xi\,dy\\ &=\int\limits_{\mathbb{T}^n\times \mathbb{R}^n\times \mathbb{R}^n}e^{i2\pi (x-y)\xi}e^{i2\pi(\eta)y}a(x,y,\xi)d\xi\,dy\widehat{f}(\eta)d\eta\\ &=\int\limits_{\mathbb{T}^n\times \mathbb{R}^n\times \mathbb{R}^n}e^{i2\pi (x)\xi}e^{i2\pi(\eta-\xi)y}a(x,y,\xi)d\xi\,dy\widehat{f}(\eta)d\eta.\end{aligned}$$ Clearly $\langle \xi \rangle^{-2q}L^q_{x}e^{i2\pi x\xi}=e^{i2\pi x\xi},$ where $L^q_{x}=(I-\frac{1}{4\pi^2}\mathcal{L}_{x})^q$ and $\mathcal{L}_{x}$ is the Laplacian in $x-$variables. Using integration by parts successively we obtain, $$\begin{aligned} \int\limits_{\mathbb{T}^n}| & a(x,y,D)f(x)|dx=\int\limits_{\mathbb{T}^n\times \mathbb{T}^n \times \mathbb{R}^n\times \mathbb{R}^n}e^{i2\pi (x)\xi}e^{i2\pi(\eta-\xi)y}a(x,y,\xi)d\xi\,dy\widehat{f}(\eta)d\eta\,dx\\ &=\int\limits_{\mathbb{T}^n\times \mathbb{T}^n \times \mathbb{R}^n\times \mathbb{R}^n}\langle \xi \rangle^{-2q}\langle \eta-\xi \rangle^{-2q} e^{i2\pi (x)\xi}e^{i2\pi(\eta-\xi)y} L^q_{x}L^{q}_{y}a(x,y,\xi)d\xi\,dy\widehat{f}(\eta)d\eta\,dx \\ &\leq \int\limits_{\mathbb{T}^n\times \mathbb{T}^n \times \mathbb{R}^n\times \mathbb{R}^n}\langle \xi \rangle^{-2q}\langle \eta-\xi \rangle^{-2q} |L^q_{x}L^{q}_{y}a(x,y,\xi)|d\xi\,dy|\widehat{f}(\eta)|d\eta\,dx\\ &\leq \int\limits_{\mathbb{T}^n\times \mathbb{T}^n \times \mathbb{R}^n\times \mathbb{R}^n}\langle \xi \rangle^{-2q}\langle \eta-\xi \rangle^{-2q} C_{q}\langle \xi\rangle^{\delta\cdot 2q} d\xi\,dy|\widehat{f}(\eta)|d\eta\,dx\\ &\leq \int\limits_{\mathbb{T}^n\times \mathbb{T}^n \times \mathbb{R}^n\times \mathbb{R}^n}\langle \xi \rangle^{2q(\delta-1)}\langle \eta-\xi \rangle^{-2q} C_{q} d\xi\,dy\Vert f\Vert_{L^1(\mathbb{T}^n)} d\eta\,dx\\ &\leq C \Vert f\Vert_{L^1(\mathbb{T}^n)},\\\end{aligned}$$ where $C=\int\limits_{\mathbb{R}^n} (\langle \cdot \rangle^{2q(\delta-1)}*\langle \cdot \rangle^{-2q} )(\eta) C_{q} d\eta.$ Clearly $C$ is finite for $\frac{n}{2}(1-\delta)^{-1}<q\leq[\frac{n}{2}(1-\delta)^{-1}]+1$. Our proof of the $L^1-$boundedness is for non-negative $f,$ but this is sufficient since an arbitrary real function can be decomposed into its positive and negative parts, and complex functions into its real and imaginary parts. Positive operators in the form of Theorem $\textnormal{IV}$ arise with Bessel potentials of order $m,$ $m\in \mathbb{R}.$ We recall that for every $m\in\mathbb{R},$ the Bessel potential of order $m,$ denoted by $\langle D_{x}\rangle^m$ is the pseudo-differential operator with symbol $\sigma_{m}(\xi)=\langle \xi\rangle^m,$ $\xi\in \mathbb{Z}^n.$ There exists a connection between the $L^{p}$ boundedness of Fourier multipliers on compact Lie groups and its continuity on Besov spaces $B^r_{p,q}$. This fact was proved by the author in Theorem 1.2 of [@Duvan5]. In fact, the Lie group structure of the torus $\mathbb{T}^n$ implies that every periodic Fourier multiplier (i.e. a periodic pseudo-differential operator with symbol depending only on the dual variable $\xi\in\mathbb{Z}^n$) bounded from $L^{p_{1}}$ into $L^{p_{2}}$ also is bounded from $B^{r}_{p_1,q}$ into $B^{r}_{p_2,q},$ $r\in\mathbb{R}$ and $0<q\leq \infty.$ So, restrictions of our results to the case of symbols $\sigma(\xi)$ associated to Fourier multipliers give the boundedness of these operators on every Besov space $B^r_{p,q}(\mathbb{T}^n),$ $1<p<\infty. $ **Acknowledgments:** I would like to thank the anonymous referee for his/her remarks which helped to improve the manuscript. This project was partially supported by Universidad de los Andes, Mathematics Department, Bogotá-Colombia. [99]{} Agranovich, M. S.: [Spectral properties of elliptic pseudodifferential operators on a closed curve]{} Funct. Anal. Appl. **13**, 279-281 (1971) Ashino, R., Nagase, M., Vaillancourt, R. [Pseudodifferential operators on $L^p(\mathbb{R}^n)$ spaces]{} Cubo., Vol 6, N 3. pp. 91-129. (2004) Calderón, A., Vaillancourt, R.: [On the boundedness of pseudo-differential operators.]{} J. Math. Soc. Japan **23**, 374-378 (1971) Calderón, A., Vaillancourt, R.: [A class of bounded pseudo-differential operators]{} Proc. Nat. Acad. Sci. Usa. **69**, 1185-1187 (1972) Cardona, D.: [Estimativos $L^2$ para una clase de operadores pseudodiferenciales definidos en el toro]{} Rev. Integr. Temas Mat., **31**(2), 147-152 (2013) Cardona, D.: Weak type (1, 1) bounds for a class of periodic pseudo-differential operators. J. Pseudo-Differ. Oper. Appl., 5(4), (2014) 507-515. Cardona, D.: Hölder-Besov boundedness for periodic pseudo-differential operators, to appear, J. Pseudo-Differ. Oper. Appl. 2016 Cardona, D.: Besov continuity for Multipliers defined on compact Lie groups. Palest. J. Math. Vol. 5(2) 35–44 (2016) Delgado, J.: [$L^p$ bounds for pseudo-differential operators on the torus ]{} Operators Theory, advances and applications. **231**, 103-116 (2012) Delgado, J. Ruzhansky, M. $L^p$-bounds for pseudo-differential operators on compact Lie groups. arXiv:1605.07027 Duoandikoetxea, J.: [Fourier Analysis]{}, Amer. Math. Soc. (2001) Folland, G.B.: Harmonic Analysis in phase space. Princeton University Press, (1989) Fefferman, C.: $L^p-$bounds for pseudo-differential operators. Israel J. Math. **14**, 413-417 (1973) Fischer, V.: Hörmander condition for Fourier multipliers on compact Lie groups. arXiv:1610.06348. Ghaemi, M. B., Nabizadeh M. E.: A study on the inverse of pseudo-differential operators on S1. J. Pseudo-Differ. Oper. Appl. 7 (2016), no. 4, 511–517. Ghaemi, M. B., Jamalpour B. M., Nabizadeh M. E.: A study on pseudo-differential operators on S1 and Z. J. Pseudo-Differ. Oper. Appl. 7 (2016), no. 2, 237–247. Ghaemi, M. B., Nabizadeh M. E., Jamalpour B. M.: A study on the adjoint of pseudo-differential operators on S1 and Z. J. Pseudo-Differ. Oper. Appl. 6 (2015), no. 2, 197–203. Hörmander, L.: [ Pseudo-differential Operators and Hypo-elliptic equations ]{} Proc. Symposium on Singular Integrals, Amer. Math. Soc. **10**, 138-183 (1967) Hörmander, L.: [ The Analysis of the linear partial differential operators]{} Vol. III. Springer-Verlag, (1985) Mclean, W.M.: [Local and Global description of periodic pseudo-differential operators,]{} Math. Nachr. **150**, 151-161 (1991) Molahajloo, S.: [A characterization of compact pseudo-differential operators on $\mathbb{S}^1$ ]{} Oper. Theory Adv. Appl. Birkhüser/Springer Basel AG, Basel. **213**, 25-29 (2011) Molahajloo, S., Wong, M.W.: [Pseudo-differential Operators on $\mathbb{S}^1$. ]{} New developments in pseudo-differential operators, Eds. L. Rodino and M.W. Wong. 297-306 (2008) Molahajloo, S., Wong, M. W. .: [ Ellipticity, Fredholmness and spectral invariance of pseudo-differential operators on $\mathbb{S}^1.$]{} J. Pseudo-Differ. Oper. Appl. **1** 183-205 (2010) Nagase, M.: [ The $L^p$ boundedness of pseudo-differential operators with non-regular symbols.]{} Comm. Partial Differential Equations. **2**, 1045-1061 (1977) Nagase, M.: [ On a class of $L^p-$bounded pseudo-differential operators.]{} Sci. Rep. College Gen. Ed. Osaka Univ. 1-7, (1984) Nagase, M.: On some classes of Lp-bounded pseudodifferential operators. Osaka J. Math. 23 (1986), no. 2, 425–440. Ruzhansky, M., Turunen, V.: [Pseudo-differential Operators and Symmetries: Background Analysis and Advanced Topics ]{} Birkhaüser-Verlag, Basel, (2010) Ruzhansky, M., Turunen, V.: [Quantization of Pseudo-Differential Operators on the Torus]{} J Fourier Annal Appl. Vol. 16. pp. 943-982. Birkhäuser Verlag, Basel, (2010) Ruzhansky, M. Wirth, J.: $L^p$ Fourier multipliers on compact Lie groups, Mathematische Zeitschrift, Vol 280, 621-642, (2015) Turunen, V., Vainikko, G.: [On symbol analysis of periodic pseudodifferential operators]{} Z. Anal. Anwendungen. **17**, 9-22 (1998) Stein, E.: Harmonic Analysis. Princeton University Press. (1993) Stein, E.: Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals. Princeton University Press, 1993. Wang, L. Pseudo-differential operators with rough coefficients. Thesis (Ph.D.)–McMaster University (Canada). ProQuest LLC, Ann Arbor, MI, 1997. 66 pp. ISBN: 978–0612–30120–7 Wong, M. W.: [An introduction to Pseudo-Differential Operators ]{} Second Edition, World Scientific, (1999) Wong, M. W.: [ Discrete Fourier Analysis ]{} Birkhäuser, (2011)
{ "pile_set_name": "ArXiv" }
Ł[[L]{}]{} [$\tilde{\phantom{a}}$]{} [**The Rolling Motion of a Disk on a Horizontal Plane**]{} Alexander J. McDonald [*Princeton High School, Princeton, New Jersey 08540*]{} Kirk T. McDonald [*Joseph Henry Laboratories, Princeton University, Princeton, New Jersey 08544*]{} [email protected] (March 28, 2001) [**Abstract**]{} Recent interest in the old problem of the motion of a coin spinning on a tabletop has focused on mechanisms of dissipation of energy as the angle $\alpha$ of the coin to the table decreases, while the angular velocity $\Omega$ of the point of contact increases. Following a review of the general equations of motion of a thin disk rolling without slipping on a horizontal surface, we present results of simple experiment on the time dependence of the motion that indicate the dominant dissipative power loss to be proportional to the $\Omega^2$ up to and including the last observable cycle. Introduction ============ This classic problem has been treated by many authors, perhaps in greatest detail but very succinctly by Routh in article 244 of [@Routh]. About such problems, Lamb [@Lamb] has said, “It is not that the phenomena, though familiar and often interesting, are held to be specially important, but it was regarded rather as a point of honour to shew how the mathematical formulation could be effected, even if the solution should prove to be impracticable, or difficult of interpretation." Typically, the role of friction was little discussed other than in relation to “rising” [@Routh; @Lamb; @HT; @Jellett; @Gallop; @Gray00]. The present paper is motivated by recent discussions [@Moffatt; @Engh; @Moffatt2] of friction for small angles of inclination of a spinning disk to the horizontal supporting surface. The issues of rolling motion of a disk are introduced in sec. 2 in the larger context of non-rigid-body motion and rolling motion on curved surfaces, using the science toy “Euler’s Disk" as an example. In our analysis of the motion of a rigid disk on a horizontal plane we adopt a vectorial approach as advocated by Milne [@Milne]. The equations of motion assuming rolling without slipping are deduced in sec. 3, steady motion is discussed in secs. 4 and 5, and oscillation about steady motion is considered in sec. 7. The case of zero friction is discussed in secs. 8 and 9, and effects of dynamic friction are discussed in secs. 6, 10 and 11. Section 12 presents a brief summary of the various aspects of the motions discussed in secs. 3-11. The Tangent Toy “Euler’s Disk" ============================== An excellent science toy that illustrates the topic of this article is “Euler’s Disk", distributed by Tangent Toy Co. [@Tangent]. Besides the disk itself, a base is included that appears to be the key to the superior performance exhibited by this toy. The surface of the base is a thin, curved layer of glass, glued to a plastic backing. The base rests on three support points to minimize rocking. As the disk rolls on the base, the latter is noticeably deformed. If the same disk is rolled on a smooth, hard surface such as a granite surface plate, the motion dies out more quickly, and rattling sounds are more prominent. It appears that a small amount of flexibility in the base is important in damping the perturbations of the rolling motion if long spin times are to be achieved. Thus, high-performance rolling motion is not strictly a rigid-body phenomenon. However, we do not pursue the theme of elasticity further in this paper. The concave shape of the Tangent Toy base helps center the rolling motion of the disk, and speeds up the reduction of an initially nonzero radius $b$ to the desirable value of zero. An analysis of the motion of a disk rolling on a curved surface is more complex than that of rolling on a horizontal plane [@sphere]. For rolling near the bottom of the sphere, the results as very similar to those for rolling on a plane. A possibly nonintuitive result is that a disk can roll stably on the inside of the upper hemisphere of a fixed sphere, as demonstrated in the motorcycle riding act “The Globe of Death” [@globeofdeath]. The Equations of Motion for Rolling Without Slipping ==================================================== ![\[fig1\] A disk of radius $a$ rolls without slipping on a horizontal plane. The symmetry axis of the disk is called axis $\hat{\bf 1}$, and makes angle $\alpha$ to the $\hat{\bf z}$ axis, which is vertically upwards, with $0 \leq \alpha \leq \pi$. The line from the center of the disk to the point of contact with the plane is called axis $\hat{\bf 3}$, which makes angle $\alpha$ to the horizontal. The horizontal axis $\hat{\bf 2}$ is defined by $\hat{\bf 2} = \hat{\bf 3} \times \hat{\bf 1}$, and the horizontal axis $\hat{\bf r}$ is defined by $\hat{\bf r} = \hat{\bf 2} \times \hat{\bf z}$. The angular velocity of the disk about axis $\hat{\bf 1}$ is called $\omega_1$, and the angular velocity of the axes $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$ about the vertical is called $\Omega$. The motion of the point of contact is instantaneously in a circle of radius $r$. The distance from the axis of this motion to the center of mass of the disk is labeled $b$. ](rollingdisk.eps){width="3.0in"} In addition to the $\hat {\bf z}$ axis which is vertically upwards, we introduce a right-handed coordinate triad of unit vectors $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$ related to the geometry of the disk, as shown in Fig. \[fig1\]. Axis $\hat {\bf 1}$ lies along the symmetry axis of the disk. Axis $\hat {\bf 3}$ is directed from the center of the disk to the point of contact with the horizontal plane, and makes angle $\alpha$ to that plane. The vector from the center of the disk to the point of contact is then $${\bf a} = a\hat {\bf 3}. \label{e1}$$ Axis $\hat {\bf 2} = \hat {\bf 3} \times \hat {\bf 1}$ lies in the plane of the disk, and also in the horizontal plane. The sense of axis $\hat {\bf 1}$ is chosen so that the component $\omega_1$ of the angular velocity vector $\vec\omega$ of the disk about this axis is positive. Consequently, axis $\hat{\bf 2}$ points in the direction of the velocity of the point of contact. (For the special case where the point of contact does not move, $\omega_1 = 0$ and analysis is unaffected by the choice of direction of axis $\hat{\bf 1}$.) Before discussing the dynamics of the problem, a considerable amount can be deduced from kinematics. The total angular velocity $\vec\omega$ can be thought of as composed of two parts, $$\vec\omega = \vec\omega_{123} + \omega_{\rm rel} \hat{\bf 1}, \label{e2}$$ where $\vec\omega_{123}$ is the angular velocity of the triad $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$, and $\omega_{\rm rel} \hat{\bf 1}$ is the angular velocity of the disk relative to that triad; the relative angular velocity can only have a component along $\hat{\bf 1}$ by definition. The angular velocity of the triad $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$ has component $\dot\alpha$ about the horizontal axis $\hat{\bf 2}$ (where the dot indicates differentiation with respect to time), and is defined to have component $\Omega$ about the vertical axis $\hat{\bf z}$. Since axis $\hat{\bf 2}$ is always horizontal, $\vec\omega_{123}$ has no component along the (horizonal) axis $\hat{\bf 2} \times \hat{\bf z} \equiv \hat{\bf r}$. Hence, the angular velocity of the triad $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$ can be written $$\vec\omega_{123} = \Omega \hat{\bf z} + \dot\alpha \hat{\bf 2} = - \Omega \cos \alpha \hat{\bf 1} + \dot\alpha \hat{\bf 2} - \Omega \sin\alpha \hat{\bf 3}, \label{e4}$$ noting that $$\hat{\bf z} = - \cos\alpha \hat{\bf 1} - \sin\alpha \hat{\bf 3}, \label{e3}$$ as can be seen from Fig. \[fig1\]. The time rates of change of the axes are therefore $$\begin{aligned} {d\hat{\bf 1} \over dt} & = & \vec\omega_{123} \times \hat{\bf 1} = -\Omega \sin\alpha \hat{\bf 2} - \dot\alpha \hat{\bf 3}, \label{e5} \\ {d\hat{\bf 2} \over dt} & = & \vec\omega_{123} \times \hat{\bf 2} = \Omega \sin\alpha \hat{\bf 1} - \Omega \cos\alpha \hat{\bf 3}, = - \Omega \hat{\bf r}, \label{e6} \\ {d\hat{\bf 3} \over dt} & = & \vec\omega_{123} \times \hat{\bf 3} = \dot\alpha \hat{\bf 1} + \Omega \cos\alpha \hat{\bf 2}, \label{e7}\end{aligned}$$ where the rotating horizontal axis $\hat{\bf r}$ is related by $$\hat{\bf r} = \hat{\bf 2} \times \hat{\bf z} = - \sin\alpha \hat{\bf 1} + \cos\alpha \hat{\bf 3}, \label{e25b}$$ whose time rate of change is $${d\hat{\bf r} \over dt} = \Omega \hat{\bf 2}. \label{e250}$$ Combining eqs. (\[e2\]) and (\[e4\]) we write the total angular velocity as $$\vec\omega = \omega_1 \hat{\bf 1} + \dot\alpha \hat{\bf 2} - \Omega \sin\alpha \hat{\bf 3}, \label{e8}$$ where $$\omega_1 = \omega_{\rm rel} - \Omega \cos \alpha. \label{e9}$$ The (nonholonomic) constraint that the disk rolls without slipping relates the velocity of the center of mass to the angular velocity vector $\vec\omega$ of the disk. In particular, the instantaneous velocity of the point contact of the disk with the horizontal plane is zero, $${\bf v}_{\rm contact} = {\bf v}_{\rm cm} + \vec\omega \times {\bf a} = 0. \label{e10}$$ Hence, $${\bf v}_{\rm cm} = {d {\bf r}_{\rm cm} \over dt} = a\hat {\bf 3} \times \vec\omega = - a \dot\alpha \hat{\bf 1} + a \omega_1 \hat{\bf 2}, \label{e11}$$ using eqs. (\[e1\]) and (\[e8\]). Additional kinematic relations can be deduced by noting that the point of contact between the disk and the horizontal plane can always be considered as moving instantaneously in a circle whose radius vector we define as ${\bf r} = r \hat{\bf r}$ with $r \ge 0$, as shown in Fig. \[fig1\], and whose center is defined to have position $x_A \hat{\bf x} + y_A \hat{\bf y}$ where $\hat{\bf x}$ and $\hat{\bf y}$ are fixed horizontal unit vectors in the lab frame. Then, the center of mass of the disk has position $${\bf r}_{\rm cm} = x_A \hat{\bf x} + y_A \hat{\bf y} + r \hat{\bf r} - a \hat{\bf 3}, \label{e10a}$$ and $${d {\bf r}_{\rm cm} \over dt} = \dot x_A \hat{\bf x} + \dot y_A \hat{\bf y} + \dot r \hat{\bf r} - a \dot\alpha \hat{\bf 1} + (r - a \cos\alpha) \Omega \hat{\bf 2}. \label{e10b}$$ In the special case of steady motion, $\dot x_A = \dot y_A = \dot r = \dot \alpha = 0$, eqs. (\[e11\]) and (\[e10b\]) combine to give $$\omega_{1} = {b \over a} \Omega, \label{e11a}$$ where $$b = r - a \cos\alpha \label{e11b}$$ is the horizontal distance from the axis of the circular motion to the center of mass of the disk. Thus, for steady motion the “spin” angular velocity $\omega_1$ is related to the “precession" angular velocity $\Omega$ according to eq. (\[e11a\]). While $\omega_1$ is defined to be nonnegative, length $b$ can be negative if $\Omega$ is negative as well. Except for axis $\hat {\bf 1}$, the rotating axes $(\hat {\bf 1},\hat {\bf 2},\hat {\bf 3})$ are not body axes, but the inertia tensor $I_{ij}$ is diagonal with respect to them in view of the symmetry of the disk. We write $$I_{11} = 2kma^2, \qquad I_{22} = kma^2 = I_{33}, \label{e12}$$ which holds for any thin, circularly symmetric disk according to the perpendicular axis theorem; $k = 1/2$ for a disk with mass $m$ concentrated at the rim, $k = 1/4$ for a uniform disk,  The angular momentum ${\bf L}_{\rm cm}$ of the disk with respect to its center of mass can now be written as $${\bf L}_{\rm cm} = \vec{\vec I} \cdot \vec \omega = kma^2 (2 \omega_1 \hat {\bf 1} + \dot\alpha \hat{\bf 2} - \Omega \sin\alpha \hat{\bf 3}). \label{e13}$$ Turning at last to the dynamics of the rolling disk, we suppose that the only forces on it are $- m g \hat{\bf z}$ due to gravity and [**F**]{} at the point of contact with the horizontal plane. For now, we ignore rolling friction and friction due to the air surrounding the disk. The equation of motion for the position ${\bf r}_{\rm cm}$ of the center of mass of the disk is then $$m {d^2 {\bf r}_{\rm cm} \over dt^2} = {\bf F} - mg \hat {\bf z}. \label{e14}$$ The torque equation of motion for the angular momentum ${\bf L}_{\rm cm}$ about the center of mass is $${d {\bf L}_{\rm cm} \over dt} = {\bf N}_{\rm cm} = {\bf a} \times {\bf F}. \label{e15}$$ We eliminate the unknown force ${\bf F}$ in eq. (\[e15\]) via eqs. (\[e1\]) and (\[e14\]) to find $${1 \over ma} {d {\bf L}_{\rm cm} \over dt} + {d^2 {\bf r}_{\rm cm} \over dt^2} \times \hat {\bf 3} = g\hat {\bf 3} \times \hat {\bf z}. \label{e17}$$ This can be expanded using eqs. (\[e3\]), (\[e5\])-(\[e7\]), (\[e11\]) and (\[e13\]) to yield the $\hat{\bf 1}$, $\hat{\bf 2}$ and $\hat{\bf 3}$ components of the equation of motion, $$\begin{aligned} (2k + 1) \dot\omega_1 + \dot\alpha \Omega \sin\alpha & = & 0, \label{e18} \\ k \Omega^2 \sin\alpha \cos\alpha + (2k + 1) \omega_1 \Omega \sin\alpha - (k + 1) \ddot\alpha & = & {g \over a} \cos\alpha, \label{e19} \\ \dot\Omega \sin\alpha + 2 \dot\alpha \Omega \cos\alpha + 2 \omega_1 \dot\alpha & = & 0. \label{e20}\end{aligned}$$ These equations agree with those of sec. 244 of Routh [@Routh], noting that his $A$, $C$, $\theta$, $\dot\psi$ and $\omega_3$ are expressed as $k a^2$, $2 k a^2$, $\alpha$, $\Omega$ and $- \omega_1$, respectively, in our notation. Besides the coordinates ($x_A,y_A)$ of the center of motion, we can readily identify only one other constant of the motion, the total energy $$\begin{aligned} E & = & T + V = {1 \over 2} m v_{\rm cm}^2 + {1 \over 2} \vec \omega \cdot \vec{\vec I} \cdot \vec \omega + m g z \nonumber \\ & = & {m a^2 \over 2} \left[ (2 k + 1) \omega_1^2 + (k + 1) \dot\alpha^2 + k \Omega^2 \sin^2\alpha + {2 g \over a} \sin\alpha \right]. \label{e20a}\end{aligned}$$ The time derivative of the energy is consistent with the equations of motion (\[e18\])-(\[e20\]), but does not provide any independent information. Steady Motion ============= For steady motion, $\dot\alpha = \ddot\alpha = \dot\Omega = \dot\omega_1 = 0$, and we define $\alpha_{\rm steady} = \alpha_0$, $\Omega_{\rm steady} = \Omega_0$ and $\omega_{1,\rm steady} = \omega_{10}$. The equations of motion (\[e18\]) and (\[e20\]) are now trivially satisfied, and eq. (\[e19\]) becomes $$k \Omega_0^2 \sin\alpha_0 \cos\alpha_0 + (2k + 1) \omega_{10} \Omega_0 \sin\alpha_0 = {g \over a} \cos\alpha_0, \label{e21}$$ A special case of steady motion is $\alpha_0 = \pi / 2$, corresponding to the plane of the disk being vertical. In this case, eq. (\[e21\]) requires that $\omega_{10} \Omega_0 = 0$. If $\Omega_0 = 0$, the disk rolls along a straight line and $\omega_{10}$ is the rolling angular velocity. If $\omega_{10} = 0$, the disk spins in place about the vertical axis with angular velocity $\Omega_0$. For $\alpha_0 \neq \pi/2$, the angular velocity $\Omega_0 \hat{\bf z}$ of the axes about the vertical must be nonzero. We can then replace $\omega_{10}$ by the radius $b$ of the horizontal circular motion of the center of mass using eqs. (\[e11a\])-(\[e11b\]): $$\omega_{10} = {b \over a} \Omega_0 = \Omega_0 \left( {r \over a} - \cos\alpha_0 \right). \label{e22}$$ Inserting this in (\[e21\]), we find $$\Omega_0^2 = {g \cot\alpha_0 \over k a \cos\alpha_0 + (2k + 1) b} = {g \cot\alpha_0 \over (2k + 1) r - (k + 1) a \cos\alpha_0 }. \label{e23}$$ For $\pi / 2 < \alpha_0 < \pi$ the denominator of eq. (\[e23\]) is positive, since $r$ is positive by definition, but the numerator is negative. Hence, $\Omega_0$ is imaginary, and steady motion is not possible in this quadrant of angle $\alpha_0$. For $0 < \alpha_0 < \pi / 2$, $\Omega_0$ is real and steady motion is possible so long as $$b > - {a k \cos\alpha_0 \over 2 k + 1}. \label{e23a}$$ In addition to the commonly observed case of $b > 0$, steady motion is possible with small negative values of $b$ A famous special case is when $b = 0$, and the center of mass of the disk is at rest. Here, eq. (\[e23\]) becomes $$\Omega_0^2 = {g \over a k \sin\alpha_0}, \label{e24}$$ and $\omega_{10} = 0$ according to eq. (\[e22\]), so that $$\omega_{\rm rel} = \Omega_0 \cos\alpha_0, \label{e25}$$ recalling eq. (\[e9\]). Also, the total angular velocity becomes simply $\vec\omega = - \Omega_0 \sin\alpha_0 \hat{\bf 3}$ according to eq. (\[e8\]), so the instantaneous axis of rotation is axis $ {\bf 3}$ which contains the center of mass and the point of contact, both of which are instantaneously at rest. Shorter Analysis of Steady Motion with $b = 0$ ============================================== The analysis of a spinning disk (for example, a coin) whose center is at rest can be shortened considerably by noting at the outset that in this case axis $\hat{\bf 3}$ is the instantaneous axis of rotation. Then, the angular velocity is $\vec\omega = \omega \hat{\bf 3}$, and the angular momentum is simply $${\bf L} = I_{33} \omega \hat{\bf 3} = k m a^2 \omega \hat{\bf 3}. \label{e26}$$ Since the center of mass is at rest, the contact force [**F**]{} is just $mg \hat{\bf z}$, so the torque about the center of mass is $${\bf N} = a \hat{\bf 3} \times mg \hat{\bf z} = {d{\bf L} \over dt}. \label{e27}$$ We see that the equation of motion for [**L**]{} has the form $${d{\bf L} \over dt} = \vec\Omega_0 \times {\bf L}, \label{e28}$$ where $$\vec\Omega_0 = - {g \over a k \omega} \hat{\bf z}. \label{e29}$$ Thus, the angular momentum, and the coin precesses about the vertical at rate $\Omega_0$. A second relation between $\vec\omega$ and $\vec\Omega_0$ is obtained from eqs. (\[e2\]) and (\[e4\]) by noting that $\vec\omega_{123} = \vec\Omega_0$, so that $$\vec\omega = (- \Omega_0 \cos\alpha_0 + \omega_{\rm rel} )\hat{\bf 1} - \Omega_0 \sin\alpha_0 \hat{\bf 3} = \omega \hat{\bf 3}, \label{e30}$$ using eq. (\[e3\]). Hence, $$\omega = - \Omega_0 \sin\alpha_0, \label{e31}$$ and the angular velocity $\omega_1$ about the symmetry axis vanishes, so that $$\omega_{\rm rel} = \Omega_0 \cos\alpha_0. \label{e32}$$ Combining eqs. (\[e29\]) and (\[e31\]), we again find that $$\Omega_0^2 = {g \over a k \sin\alpha_0}, \label{e33}$$ As $\alpha_0$ approaches zero, the angular velocity of the point of contact becomes very large, and one hears a high-frequency sound associated with the spinning coin. However, a prominent aspect of what one sees is the rotation of the figure on the face of the coin, whose angular velocity $\Omega_0 - \omega_{\rm rel} = \Omega_0 (1 - \cos\alpha_0)$ approaches zero. The total angular velocity $\omega$ also vanishes as $\alpha_0 \to 0$. Radial Slippage During “Steady” Motion ====================================== The contact force [**F**]{} during steady motion at a small angle $\alpha_0$ is obtained from eqs. (\[e6\]), (\[e11\]), (\[e14\]), (\[e22\]) and (\[e24\]) as $${\bf F} = mg \hat{\bf z} - {b \over a k \sin\alpha_0} mg \hat{\bf r}. \label{e25a}$$ The horizontal component of force [**F**]{} is due to static friction at the point of contact. The coefficient $\mu$ of friction must therefore satisfy $$\mu \ge {\abs{b} \over a k \sin\alpha_0}, \label{e25c}$$ otherwise the disk will slip in the direction opposite to the radius vector [**b**]{}. Since coefficient $\mu$ is typically one or less, slippage will occur whenever $a k \sin\alpha_0 \lsim \abs{b}$. As the disk loses energy and angle $\alpha$ decreases, the slippage will reduce $\abs{b}$ as well. The trajectory of the center of the disk will be a kind of inward spiral leading toward $b = 0$ for small $\alpha$. If distance $b$ is negative, it must obey $\abs{b} < a k \cos\alpha_0 / (2 k + 1)$ according to eq. (\[e23a\]). In this case, eq. (\[e25c\]) becomes $$\mu \ge {\cot\alpha_0 \over 2 k + 1}, \label{e25d}$$ which could be satisfied for a uniform disk only for $\alpha_0 \gsim \pi / 3$. Motion with negative $b$ is likely to be observed only briefly before large radial slippage when $\alpha_0$ is large reduces $b$ to zero. Once $b$ is zero, the contact force is purely vertical, according to eq. (\[e25a\]). Surprisingly, the condition of rolling without slipping can be maintained in this special case without any friction at the point of contact. Hence, an analysis of the motion with $b = 0$ could be made with the assumption of zero friction, as discussed in secs. 8-9. In practice, there will always be some friction, aspects of which are further discussed in secs. 10-11. From the argument here, we see that if $b = 0$, there is no frictional force to oppose the radial slippage that accompanies a change in angle $\alpha$. Small Oscillations about Steady Motion ====================================== We now consider oscillations at angular frequency $\varpi$ about steady motion, assuming that the disk rolls without slipping. We suppose that $\alpha$, $\Omega$ and $\omega_{1}$ have the form $$\begin{aligned} \alpha & = & \alpha_0 + \epsilon \cos\varpi t, \label{e41} \\ \Omega & = & \Omega_0 + \delta \cos\varpi t, \label{e42} \\ \omega_1 & = & \omega_{10} + \gamma \cos\varpi t, \label{e43}\end{aligned}$$ where $\epsilon$, $\delta$ and $\gamma$ are small constants. Inserting these in the equation of motion (\[e20\]) and equating terms of first order of smallness, we find that $$\delta = - {2 \epsilon \over \sin\alpha_0} (\Omega_0 \cos\alpha_0 + \omega_{10}). \label{e44}$$ From this as well as from eq. (\[e41\]), we see that $\epsilon / \sin\alpha_0 \ll 1$ for small oscillations. Similarly, eq. (\[e18\]) leads to $$\gamma = - \epsilon {\Omega_0 \sin\alpha_0 \over 2 k + 1}, \label{e45}$$ and eq. (\[e19\]) leads to $$\begin{aligned} \epsilon \varpi^2 (k + 1) & = & - (2 k + 1) (\epsilon \omega_{10} \Omega_0 \cos\alpha_0 + \gamma \Omega_0 \sin\alpha_0 + \delta \omega_{10} \sin\alpha_0) + \epsilon k \Omega_0^2 (1 - 2 \cos^2\alpha_0) \nonumber \\ & & \ - 2 \delta k \Omega_0 \sin\alpha_0 \cos\alpha_0 - \epsilon {g \over a} \sin\alpha_0. \label{e46}\end{aligned}$$ Combining eqs. (\[e44\])-(\[e46\]), we obtain $$\begin{aligned} \varpi^2 (k + 1) & = & \Omega_0^2 (k (1 + 2 \cos^2\alpha_0) + \sin^2\alpha_0) + (6k + 1) \omega_{10} \Omega_0 \cos\alpha_0 \nonumber \\ & & + 2 (2 k + 1) \omega^2_{10} - {g \over a} \sin\alpha_0, \label{e47}\end{aligned}$$ which agrees with Routh [@Routh], noting that our $k$, $\Omega_0$, and $\omega_{10}$ are his $k^2$, $\mu$, and $- n$. For the special case of a wheel rolling in a straight line, $\alpha_0 = \pi/2$, $\Omega_0 = 0$, and $$\varpi^2 (k + 1) = 2 (2 k + 1) \omega^2_{10} - {g \over a}. \label{e48}$$ The rolling is stable only if $$\omega^2_{10} > {g \over 2 (2 k + 1) a}. \label{e49}$$ Another special case is that of a disk spinning about a vertical diameter, for which $\alpha_0 = \pi / 2$ and $\omega_{10}$ and $b$ are zero. Then, eq. (\[e47\]) indicates that the spinning is stable only for $$\abs{\Omega_0} > \sqrt{g \over a (k + 1)}, \label{e49a}$$ which has been called the condition for “sleeping”. Otherwise, angle $\alpha$ decreases when perturbed, and the motion of the disc becomes that of the more general case. Further discussion of this special case is given in the following section. Returning to the general analysis of eq. (\[e47\]), we eliminate $\omega_{10}$ using eq. (\[e22\]) and replace the term $(g / a) \sin\alpha_0$ via eq. (\[e23\]) to find $$\begin{aligned} {\varpi^2 \over \Omega_0^2} (k + 1) & = & 3 k \cos^2\alpha_0 + \sin^2\alpha_0 + {b \over a} \left( (6k + 1) \cos\alpha_0 - (2 k + 1) {\sin^2\alpha_0 \over \cos\alpha_0} \right) \nonumber \\ & & \qquad +\ 2 {b^2 \over a^2} (2 k + 1). \label{e50}\end{aligned}$$ The term in eq. (\[e50\]) in large parentheses is negative for $\alpha_0 > \tan^{-1} \sqrt{(6 k + 1) / (2 k + 1)}$, which is about $60^\circ$ for a uniform disk. Hence, for positive $b$ the motion is unstable for large $\alpha_0$, and the disk will appear fall over quickly into a rolling motion with $\alpha_0 \lsim 60^\circ$, after which $\alpha_0$ will decrease more slowly due to the radial slippage discussed in sec. 5, until $b$ becomes very small. The subsequent motion at small $\alpha_0$ is considered further in sec. 11. The motion with negative $b$ is always stable against small oscillations, but the radial slippage is large as noted in sec. 5. For motion with $b \ll a$, such as for a spinning coin whose center is nearly fixed, the frequency of small oscillation is given by $${\varpi \over \Omega_0} = \sqrt{3 k \cos^2\alpha_0 + \sin^2\alpha_0 \over k + 1}. \label{e51}$$ For small angles this becomes $${\varpi \over \Omega_0} \approx \sqrt{3 k \over k + 1}. \label{e52}$$ For a uniform disk with $k = 1/4$, the frequency $\varpi$ of small oscillation approaches $\sqrt{3/5} \Omega_0 = 0.77 \Omega_0$, while for a hoop with $k = 1/2$, $\varpi \to \Omega_0$ as $\alpha_0 \to 0$. The effect of this small oscillation of a spinning coin is to produce a kind of rattling sound during which the frequency sounds a bit “wrong". This may be particularly noticeable if a surface imperfection suddenly excites the oscillation to a somewhat larger amplitude. The radial slippage of the point of contact discussed in sec. 5 will be enhanced by the rattling, which requires a larger peak frictional force to maintain slip-free motion. As angle $\alpha_0$ approaches zero, the slippage keeps the radius $b$ of order $ a \sin\alpha_0$. For small $\alpha_0$, $b \approx \alpha_0 a$ and eq. (\[e50\]) gives the frequency of small oscillation as $$\varpi \approx \Omega_0 \sqrt{3 k + (6 k + 1) \alpha_0 \over k + 1}. \label{e53}$$ For a uniform disk, $k = 1/4$, and eq. (\[e53\]) gives $$\varpi \approx \Omega_0 \sqrt{3 + 10 \alpha_0 \over 5}. \label{e54}$$ When $\alpha_0 \approx 0.2$ rad, the oscillation and rotation frequencies are nearly identical, at which time a very low frequency beat can be discerned in the nutations of the disk. Once $\alpha_0$ drops below about 0.1 rad, the low-frequency nutation disappears and the disk settles into a motion in which the center of mass hardly appears to move, and the rotation frequency $\Omega_0 \approx \sqrt{g / a k \alpha_0}$ grows very large. For a hoop ($k = 1/2$), the low-frequency beat will be prominent for angles $\alpha$ near zero. Disk Spinning About a Vertical Diameter ======================================= When a disc is spinning about a vertical diameter the condition of contact with the horizontal surface is not obviously that of rolling without slipping, which requires nonzero static friction. Olsson has suggested that there is zero friction between the disk and the surface in this case [@Olsson]. If there is no friction, all forces on the disc are vertical. Then, the center of mass moves only vertically, and there is no vertical torque component about the center of mass, so the vertical component $L_z$ of angular momentum is constant. The equations of motion in the absence of friction can be found by the method of sec. 3, writing the position of the center of mass as $${\bf r}_{\rm cm} = a \sin\alpha \hat{\bf z}. \label{e401}$$ Using this in eq. (\[e17\]), the $\hat{\bf 1}$, $\hat{\bf 2}$ and $\hat{\bf 3}$ components of the equation of motion are $$\begin{aligned} \dot\omega_1 & = & 0, \label{e402} \\ (k \Omega^2 + \dot\alpha^2) \sin\alpha \cos\alpha + 2k \omega_1 \Omega \sin\alpha - (k + \cos^2\alpha) \ddot\alpha & = & {g \over a} \cos\alpha, \label{e403} \\ \dot\Omega \sin\alpha + 2 \dot\alpha \Omega \cos\alpha + 2 \omega_1 \dot\alpha & = & 0. \label{e404}\end{aligned}$$ According to eq. (\[e402\]), the angular velocity $\omega_1$ about the symmetry axis of the disk is constant. Then, eq. (\[e404\]) can be multiplied by $k m a^2 \sin\alpha$ and integrated to give $$k m a^2 (\Omega \sin^2\alpha - 2 \omega_1 \cos\alpha) = L_z =\ {\rm constant}, \label{e405}$$ recalling eq. (\[e13\]). In the case of motion of a disk with no friction we find five constants of the motion, $x_{\rm cm}$, $y_{\rm cm}$, $\omega_1$, $L_z$ and the total energy $E$, in contrast to the case of rolling without slipping in which the only (known) constants of the motion are the energy $E$ and the coordinates ($x_A,y_A$) of the center of motion. For spinning about a vertical diameter, $\alpha = \pi /2$ and $\omega_1 = 0$. For small perturbations about this motion we write $\alpha = \pi / 2 - \epsilon$, and for small $\epsilon$ eq. (\[e403\]) becomes $$\ddot\epsilon + \left( \Omega^2 - {g \over a k} \right) \epsilon = 0. \label{e406}$$ Hence, in the case of no friction, spinning about a vertical diameter is stable for $$\abs{\Omega} > \sqrt{g \over a k}\, . \label{e407}$$ For a uniform disk with $k = 1/4$, this stability condition is that $\abs{\Omega} > 2\sqrt{g/a}$. In contrast, the stability condition (\[e49a\]) for a uniform disk that rolls without slipping is that $\abs{\Omega} > 2\sqrt{g/5a} \approx 0.9 \sqrt{g/a}$. As the stability conditions (\[e49a\]) and (\[e407\]) differ by more than a factor of two for a uniform disk, there is hope of distinguishing between them experimentally. We conducted several tests in which a U.S.quarter dollar was spun initially about a vertical diameter on a vinyl floor, on a sheet of glossy paper, and on the glass surface of the base of the Tangent Toy Euler’s Disk. (The Euler’s Disk is so thick that when spun about a vertical diameter it comes to rest without falling over.) We found it essentially impossible to spin a coin such that there is no motion of its center of mass. Rather, the center of mass moves slowly in a spiral before the coin falls over into the “steady” motion with small $b$ described in sec. 5. A centripetal force is required for such spiral motion, and so friction cannot be entirely neglected. The occasional observation of “rising”, as discussed further in sec. 10, is additional evidence for the role of friction in nearly vertical spinning. Analysis of frames taken with a digital video camera [@Panasonic] at 30 frames per second with exposure time 1/8000 s did not reveal a sharp transition from spinning of a coin nearly vertically about a diameter to the settling motion of sec. 5, but in our judgment the transition point for $\Omega$ in several data sets was in the range 1.5-3$\sqrt{g/a}$. This suggests that during the spinning about a nearly vertical diameter friction plays only a small role, as advocated by Olsson [@Olsson]. Small Oscillations About Steady Motion with No Friction ======================================================= It is interesting to pursue the consequences of the equations of motion (\[e402\])-(\[e404\]), deduced assuming no friction, when angle $\alpha$ is different from $\pi / 2$. For motion that has evolved from $\alpha = \pi / 2$ initially, we expect the constant $\omega_1$ to be zero still. Then, eq. (\[e403\]) indicates that the value of $\Omega_0$ for steady motion at angle $\alpha_0$ is the same as that of eq. (\[e24\]). This was anticipated in sec. 5, where it was noted that for $b = 0$ no friction is required to enforce the condition of rolling without slipping. We consider small oscillations about steady motion at angle $\alpha_0$ of the form $$\begin{aligned} \alpha & = & \alpha_0 + \epsilon \cos\varpi t, \label{e408} \\ \Omega & = & \Omega_0 + \delta \cos\varpi t, \label{e409}\end{aligned}$$ where $\epsilon$ and $\delta$ are small constants. Inserting these in the equation of motion (\[e404\]) and equating terms of first order of smallness, we find that $$\delta = - 2 \epsilon \Omega_0 \cot\alpha_0, \label{e410}$$ which is the same as eq. (\[e44\]) with $\omega_1 = 0$, since eqs. (\[e20\]) and (\[e404\]) are the same. Similarly, eq. (\[e403\]) leads to $$\begin{aligned} \epsilon \varpi^2 (k + \cos^2\alpha_0) & = & \epsilon k \Omega_0^2 (1 - 2 \cos^2\alpha_0) - 2 \delta k \Omega_0 \sin\alpha_0 \cos\alpha_0 - \epsilon {g \over a} \sin\alpha_0. \label{e411}\end{aligned}$$ Combining eqs. (\[e410\])-(\[e411\]), we obtain $$\varpi^2 (k + \cos^2\alpha_0) = k \Omega_0^2 (1 + 2 \cos^2\alpha_0) - {g \over a} \sin\alpha_0 = 3 k \Omega_0^2 \cos^2\alpha_0, \label{e412}$$ using eq. (\[e24\]). The ratio of the frequency $\varpi$ of small oscillations to the frequency $\Omega_0$ of rotation about the vertical axis for $\alpha_0 < \pi / 2$ is $${\varpi \over \Omega_0} = \sqrt{3 k \over k + \cos^2\alpha_0} \cos\alpha_0, \label{e413}$$ which differs somewhat from the result (\[e51\]) obtained assuming rolling without slipping. For very small $\alpha_0$, both eq. (\[e51\]) and (\[e413\]) take on the same limiting value (\[e52\]). Because of the similarity of the results for small oscillations about steady motion with $b = 0$ for either assumption of no friction or rolling without slipping, it will be hard to distinguish experimentally which condition is the more realistic, but the distinction is of little consequence. “Rising” of a Rotating Disk When Nearly Vertical ($\alpha \approx \pi / 2$) ================================================================= A rotating disk can exhibit “rising" when launched with spin about a nearly vertical diameter, provided there is slippage at the point of contact with the horizontal plane. That is, the plane of the disc may rise first towards the vertical, before eventually falling towards the horizontal. The rising of tops appears to have been considered by Euler, but rather inconclusively. The present explanation based on sliding friction can be traced to a note by “H.T.” in 1839 [@HT]. Briefly, we consider motion that is primarily rotation about a nearly vertical diameter. The angular velocity about the vertical is $\Omega > \sqrt{g/a k}$, large enough so that “sleeping” at the vertical is possible in the absence of friction. The needed sliding friction depends on angular velocity component $\omega_1 = b \Omega / a$ being nonzero, which implies that the center of mass moves in a circle of radius $b \ll a$ in the present case. Then, $\omega_1 \ll \Omega$, and the angular momentum (\[e13\]) is ${\bf L} \approx - \Omega \hat{\bf 3}$, which is almost vertically upwards (see Fig. \[fig1\]). Rising depends on slippage of the disk at the point of contact such that the lowermost point on the disk is not at rest but moves with velocity $- \epsilon a \omega_1 \hat{\bf 2}$, which is opposite to the direction of motion of the center of mass. Corresponding to this slippage, the horizontal surface exerts friction $F_s \hat{\bf 2}$ on the disk, with $F_s > 0$. The related torque, ${\bf N}_s = a \hat{\bf 3} \times F_s \hat{\bf 2} = - a F_s \hat{\bf 1}$, pushes the angular momentum towards the vertical, and the center of mass of the disk rises. The torque needed for rising exists in principle even for a disk of zero thickness, provided there is sliding friction at the point of contact. The most dramatic form of rising motion is that of a “tippe” top, which has recently been reviewed by Gray and Nickel [@Gray00]. Friction at Very Small $\alpha$ =============================== In practice, the motion of a spinning disk appears to cease rather abruptly for a small value of the angle $\alpha$, corresponding to large precession angular velocity $\Omega$. If the motion continued, the velocity $\Omega a$ of the point of contact would eventually exceed the speed of sound. This suggests that air friction may play a role in the motion at very small $\alpha$, as has been discussed recently by Moffatt [@Moffatt; @Engh; @Moffatt2]. When the rolling motion ceases, the disk seems to float for a moment, and then settle onto the horizontal surface. It appears that the upward contact force ${\bf F}_z$ vanished, and the disk lost contact with the surface. From eqs. (\[e11\]) and (\[e14\]), we see that for small $\alpha$, ${\bf F}_z \approx mg + m a \ddot \alpha$. Since the height of the center of mass above the surface is $h \approx a \alpha$ for small $\alpha$, we recognize that the disk loses contact with the surface when the center of mass is falling with acceleration $g$. Moffatt invites us to relate the power $P$ dissipated by friction to the rate of change $dU/dt$ of total energy of the disk. For a disk moving with $b = 0$ at a small angle $\alpha(t)$, $$U = {1 \over 2} m \dot h^2 + {1 \over 2} I_{33} \omega^2 + m g h \approx {1 \over 2} m a^2 \dot\alpha^2 + {3 \over 2} m a g \alpha, \label{e101}$$ using eq. (\[e31\]) and assuming that eq. (\[e33\]) holds adiabatically. Then, $$P = {d U \over dt} \approx m a^2 \dot\alpha \ddot\alpha + {3 \over 2} m a g \dot\alpha \approx {5 \over 2} m a g \dot\alpha, \label{e102}$$ where the second approximation holds when ${\bf F}_z \approx 0$ and $m a \ddot\alpha \approx m g$. For the dissipation of energy we need a model. First, we consider rolling friction, taken to be the effect of inelastic collisions between the disk and the horizontal surface. For example, suppose the surface has small bumps with average spacing $\delta$ and average height $h =\epsilon \delta$. We also suppose that the disk dissipates energy $m g h = m g \epsilon \delta$ when passing over a bump. The time taken for the rotating disk to pass over a bump is $\delta / a \Omega$ (at small $\alpha$), so the rate of dissipation of energy to rolling friction is $$P = - {m g \epsilon \delta \over \delta / a \Omega} = - \epsilon m a g \Omega. \label{e103}$$ A generalized form of velocity-dependent friction could be written as $$P = - \epsilon m a g \Omega^\beta, \label{e103a}$$ where the drag force varies with (angular) velocity as $\Omega^{\beta - 1}$. A rolling frictional force proportional to the velocity of the contact point corresponds to $\beta = 2$; an air drag force proportional to the square of the velocity corresponds to $\beta = 3$. The model of Moffatt [@Moffatt] emphasizes the viscous shear of the air between the disk and the supporting horizontal surface, and corresponds to $\beta = 4$. A revised version of Moffatt’s model reportedly [@Moffatt2] corresponds to $\beta = 2.5$. Equating the frictional power loss to the rate of change (\[e102\]) of the energy of the disk, we find $$\dot\alpha = - {2 \epsilon \over 5} \Omega^\beta \approx - {2 \epsilon \over 5} \left( {g \over ak} \right)^{\beta/2} {1 \over \alpha^{\beta/2}}\, , \label{e104}$$ using $\Omega^2 \approx g / a k \alpha$ from eq. (\[e24\]), which integrates to give $$\alpha^{(\beta + 2)/2} = {\epsilon (\beta + 2) \over 5} \left( {g \over ak} \right)^{\beta/2} (t_0 - t), \label{e105}$$ and $$\alpha = \left( {\epsilon (\beta + 2) \over 5} \right)^{2/(\beta + 2)} \left( {g \over ak} \right)^{\beta/(\beta + 2)} (t_0 - t)^{2/(\beta + 2)}. \label{e106}$$ In this model, the angular velocity $\Omega$ obeys $$\Omega = \left( {5 g / \epsilon (\beta + 2) a k \over t_0 - t} \right)^{1/(\beta + 2)} \equiv \left( {C \over t_0 - t} \right)^{1/(\beta + 2)}, \label{e107}$$ which exhibits what is called by Moffatt a “finite-time singularity" at time $t_0$ [@Moffatt] for any value of $\beta$ greater than $-2$. A premise of this analysis is that it will cease to hold when the disk loses contact with the surface, , when $F_z = 0$, at which time $\ddot\alpha = - g/a$, or equivalently $d^2(1/\Omega^2)/dt^2 = - k$. Taking the derivative of eq. (\[e107\]), the maximum angular velocity is $$\Omega_{\rm max} = \left( {k (\beta + 2)^2 \over 2 \beta} \right)^{1 / 2 (\beta + 1) } C^{1 / (\beta + 1)}, \label{e108}$$ which occurs at time $t_{\rm max}$ given by $$t_0 - t_{\rm max} = \left( {2 \beta \over k (\beta + 2)^2 } \right)^{(\beta + 2) / 2 (\beta + 1)} C^{-1 / (\beta + 1)} . \label{e109}$$ In Moffatt’s model based on viscous drag of the air between the disc and the surface [@Moffatt], $\beta = 4$, $$\alpha = \left( {2 \pi \eta a \over m} (t_0 - t) \right)^{1/3}, \label{e111}$$ where $\eta = 1.8 \times 10^{-4}$ g-cm$^{-1}$-s is the viscosity of air, and $$\Omega = \sqrt{g \over ak} \left( {m / 2 \pi \eta a \over t_0 - t} \right)^{1/6}. \label{e112}$$ This model is notable for having no free parameters. The main distinguishing feature between the various models for friction is the different time dependences (\[e107\]) for the angular velocity $\Omega$ as angle $\alpha$ decreases. A recent report [@Engh] indicates that the total times of spin of coins in vacuum and in air are similar, which suggests that air drag is not the dominant mechanism of energy dissipation. Such results do not preclude that air drag could be important for disks of better surface quality, and hence lower rolling friction, or that air drag becomes important only during the high-frequency motion as time $t$ approaches $t_0$. To help determine whether any of the above models corresponds to the practical physics, we have performed an experiment using a Tangent Toy Euler’s Disk [@Tangent]. The spinning disk was illuminated by a flashlight whose beam was reflected off the surface of the disk onto a phototransistor [@Taos] whose output was recorded by a digital oscilloscope [@Tek] at 5 kS/s. The record length of 50,000 samples permitted the last 10 seconds of the spin history of the disk to be recorded, as shown in Figs. \[tek04\]-\[tek02\]. ![\[tek04\] A 10-s record at 5 kS/s of the spinning of a Tangent Toy Euler’s Disk [@Tangent] as observed by a phototransistor that detected light reflected off the disk. ](tek0t4.eps){width="3.0in"} ![\[tek01\] The last 0.25 s of the history of the spinning disk shown in Fig. \[tek04\]. ](tek0t1.eps){width="3.0in"} ![\[tek02\] The last 0.06 s of the history of the spinning disk shown in Figs. \[tek04\] and \[tek01\]. ](tek0t2.eps){width="3.0in"} The analysis of the data shown in Figs. \[tek04\]-\[tek02\] consisted of identifying the time $t_i$ of the peak of cycle $i$ of oscillation as the mean of the times of the rising and falling edges of the waveform at roughly one half the peak height. The average angular frequency for each cycle was calculated as $\Omega_i = 2 \pi / (t_{i+1} - t_i)$, as shown in Fig. \[Omega\], and the rate of change of angular frequency was calculated as $d\Omega_i /dt = 2(\Omega_{i+1} - \Omega_i) / (t_{i+2} - t_i)$. The angular frequency of the last analyzable cycle was $\Omega_{\rm max} = 680$ Hz. ![\[Omega\] $\Omega$  $t$ deduced from the data shown in Figs. \[tek04\]-\[tek02\]. ](tek0_Omega.eps){width="3.0in"} The data can be conveniently compared to the result (\[e107\]) in the form $${1 \over \Omega} = \left( {t_0 - t \over C} \right)^{1/(\beta + 2)} \label{e120}$$ via a log-log plot of $1/\Omega$  $t_0 - t$, given an hypothesis as to $t_0.$ Inspection of Fig. \[tek02\] suggests that $t_0$ is in the range 7.26-7.28 s for our data sample. Figures \[tek726\] and \[tek728\] show plots of $1/\Omega$  $t_0 - t$ for $t_0 = 7.26$ and 7.28 s, respectively. The straight lines are not fits to the data, but illustrate the behavior expected according to eq. (\[e107\]) for various values of parameter $\beta$. ![\[tek726\] $1/\Omega$  $t_0 - t$ for $t_0 = 7.26$ s, using the data shown in Figs. \[tek04\]-\[tek02\]. The straight lines illustrate the behavior expected according to eq. (\[e107\]) for various values of parameter $\beta$. ](tek0_726.eps){width="3.0in"} ![\[tek728\] $1/\Omega$  $t_0 - t$ for $t_0 = 7.28$ s, using the data shown in Figs. \[tek04\]-\[tek02\]. ](tek0_728.eps){width="3.0in"} A larger value of $t_0$ has the effect of lowering the apparent value of parameter $\beta$ for the last few cycles of the motion. Figure \[tek02\] suggests that $t_0$ could hardly be less than $7.26$ s, for which case a value of $\beta = 2$ would fit the entire data sample rather well. In view of the uncertainty in assigning a value to the time $t_0$, it is interesting to ask at what time $t$ does the time remaining equal exactly one period of the motion, , when does $t_0 - t = 2 \pi / \Omega(t)$? For $\beta = 2$, the answer from eq. (\[e107\]) is when $\Omega = (C / 2 \pi)^{1/3}$. From Fig. \[tek726\] we estimate that $C^{-1/4} = 0.0055$, and hence $\Omega \approx 560$ Hz when the remaining time of the motion is $2 \pi / \Omega = 0.011$ s. Recall that the last cycle analyzable in our data sample yielded a value of 680 Hz for $\Omega$. The preceding analysis tells us that time $t_0$ cannot be more than about 0.01 s after the last observable peak in the data, which suggests that $t_0$ is closer to 7.26 than to 7.28 s, and that Fig. \[tek726\] is the relevant representation of the experiment. For $\beta = 2$, the spinning disk is predicted by eq. (\[e109\]) to lose contact with the horizontal surface when $t_0 - t = C^{-1/3} = 0.001$ s for $C^{-1/4} = 0.0055$. The instantaneous angular frequency at that time is predicted by eq. (\[e108\]) to be $\Omega = C^{1/3} = 1030$ Hz. These values are, of course, beyond those for the last analyzable cycle in the data. The question as to the value of $t_0$ can be avoided by noting [@Chatterjee] that the time derivative of eq. (\[e107\]) yields the relation $${d \Omega \over dt} \propto \Omega^{\beta + 3}. \label{e121}$$ However, $d\Omega / dt$ must be calculated from differences of differences of the times of the peaks in the data, so is subject to greater uncertainty than is $\Omega$. Figure \[Odot\] shows a log-log plot of $d\Omega/dt$ $\Omega$ together with straight lines illustrating the expected behavior for various values of $\beta$. Again, $\beta = 2$ is a consistent description of the entire data sample. The value of $\beta = 0$ suggested by Fig. \[tek728\] based on $t_0 = 7.28$ s is quite inconsistent with Fig. \[Odot\]. ![\[Odot\] $d\Omega/dt$  $\Omega$ for the data shown in Figs. \[tek04\]-\[tek02\]. The straight lines illustrate the behavior expected according to eq. (\[e121\]) for various values of parameter $\beta$. ](OdotvsO_0.eps){width="3.0in"} The results of our experiment on the time history of the motion of a spinning disk are not definitive, but are rather consistent with the dissipated power being proportional to the square of the velocity of the point of contact, and hence the dissipative force varying linearly with velocity. Our experiment cannot determine whether during the 0.01 s beyond the last full cycle of the motion an additional dissipative mechanism such as air friction with power loss proportional to the fourth power of the velocity [@Moffatt] became important. Summary of the Motion of a Disk Spun Initially About a Vertical Diameter ======================================================================== If a uniform disk is given a large initial angular velocity about a vertical diameter, and the initial horizontal velocity of the center of mass is very small, the disk will “sleep” until friction at the point of contact reduces the angular velocity below that of condition (\[e407\]) (secs. 8). The disk will then appear to fall over rather quickly into a rocking motion with angle $\alpha < 90^\circ$ (sec. 9). After this, the vertical angular velocity $\Omega$ will increase ever more rapidly, while angle $\alpha$ decreases (sec. 5), until the disk loses contact with the table at a value of $\alpha$ of a few degrees (sec. 11). The disk then quickly settles on to the horizontal surface. One hears sound at frequency $\Omega / 2 \pi$, which becomes dramatically higher until the sound abruptly ceases. But if one observes a figure on the face of the disk, this rotates every more slowly and seems almost to have stopped moving before the sounds ceases (sec. 4). If the initial motion of the disk included a nonzero initial velocity in addition to the spin about a vertical diameter, the center of mass will initially move in a circle (sec. 4). If the initial vertical angular velocity is small, the disc will roll in a large circle, tilting slightly inwards until the rolling angular velocity $\omega_1$ drops below that of condition (\[e49\]). While in most cases the angle $\alpha$ of the disk will then quickly drop to $60^\circ$ or so (sec. 6), occasionally $\alpha$ will rise back towards $90^\circ$ before falling (sec. 9). As the disk rolls and spins, the center of mass traces an inward spiral on average (sec. 5), but nutations about this spiral can be seen, often accompanied by a rattling sound (sec. 6). The nutation is especially prominent for $\alpha \approx 10-15^\circ$ at which time a very low beat frequency between that of primary spin and that of the small oscillation can be observed (sec. 6). As $\alpha$ decreases below this, the radius of the circle traced by the center of mass becomes very small, and the subsequent motion is that of a disk without horizontal center of mass motion (secs. 4 and 8). The authors thank A. Chatterjee, C.G. Gray, H.K. Moffatt, M.G. Olsson and A. Ruina for insightful correspondence on this topic. [34]{} E.J. Routh, [*The Advanced Part of a Treatise on the Dynamics of a System of Rigid Bodies*]{}, 6th ed. (Macmillan, London, 1905; reprinted by Dover Publications, New York, 1955). H. Lamb, [*Higher Mechanics*]{} (Cambridge U. Press, Cambridge, 1920), p. 156. H.T., [*Note on the Theory of the Spinning Top*]{}, Camb. Math. J. [**1**]{}, 42-44 (1839). A reprint edition of this journal, published in 1845 under the supervision of W. Thomson, includes an index that identifies the author of this paper as Archibald Smith, who was the first editor of the Camb. Math. J. (T. Crilly, private communication). This paper is alluded to in ref. 1 of [@Gray00]. J.H. Jellett, [*A Treatise on the Theory of Friction*]{} (Hodges, Foster and Co., Dublin; Macmillan, London 1872), Chap. VIII, Sec. I. E.G. Gallop, [*On the Rise of a Spinning Top*]{}, Trans. Camb. Phil. Soc. [**19**]{}, 356-373 (1904). C.G. Gray and B.G. Nickel, [*Constants of the motion for nonslipping tippe tops and other tops with round pegs*]{}, Am. J. Phys. [**68**]{}, 821-828 (2000). H.K. Moffatt, [*Euler’s disk and its finite-time singularity*]{}, Nature [**404**]{}, 833-834 (2000). G. van den Engh, P. Nelson, J. Roach, [*Numismatic gyrations*]{}, Nature [**408**]{}, 540 (2000). H.K. Moffatt, [*Moffatt replies*]{}, Nature [**408**]{}, 540 (2000). E.A. Milne, [*Vectorial Mechanics*]{} (Interscience Publishers, New York, 1948). J. Bendik, [*The Official Euler’s Disk Website*]{}, http://www.eulersdisk.com/ Tangent Toy Co., P.O. Box 436, Sausalito, CA 94966, http://www.tangenttoy.com/ K.T. McDonald, [*Circular Orbits Inside the Globe of Death*]{}, http://xxx.lanl.gov/abs/physics/0008226 Some websites featuring the Globe of Death are http://www.globeofdeath.com/home.html “The Urias” at http://www.ringling.com/bios/ http://www.sciencejoywagon.com/physicszone/lesson/03circ/sphear/sphear.htm M.G. Olsson, [*Coin Spinning On a Table*]{}, Am. J. Phys. [**40**]{}, 1543-1545 (1972). Panasonic Model PVDV910D, Matsushita Electric Corporation of America, One Panasonic Way, Secaucus, NJ 07094, http://www.panasonic.com/ Model 253, Taos Inc., Plano, TX 75074, http://www.taosinc.com/pdf/TSL253.pdf Model TDS744A, Tektronix Inc., Beaverton, OR 97077, http://www.tektronix.com/ A. Chatterjee, private communication. He performed (unpublished) experiments in which a 20-cm-diameter disk was photographed with a high-speed camera, yielding measurements of angle $\alpha$ and time $t$ for each cycle. A plot of $d\alpha / dt$  $\Omega$ suggested that $\beta$ was considerably less than 4.
{ "pile_set_name": "ArXiv" }
#### Introduction. {#introduction. .unnumbered} Recently there has been considerable interest in two dimensional systems containing both electrons and holes in the presence of a strong magnetic field.[@kheng; @shields; @chen; @wojs1; @rashba; @palacios; @wojs2; @lerner] In such systems, neutral ($X^0$) and charged excitons ($X^-$) and larger exciton complexes ($X_k^-$, $k$ neutral $X^0$’s bound to an electron) can occur. The excitonic ions $X_k^-$ are long-lived Fermions,[@palacios; @wojs2] whose energy spectra contain Landau level structure.[@wojs1; @wojs2] In this paper we investigate by exact numerical diagonalization small systems containing $N_e$ electrons and $N_h$ holes ($N_e\ge N_h$), confined to the surface of a Haldane sphere.[@haldane1] For $N_h=1$ these systems serve as simple guides to understanding photoluminescence.[@kheng; @shields; @chen; @wojs1; @rashba] For larger values of $N_h$ it is possible to form a multi-component plasma containing electrons and $X_k^-$ complexes.[@wojs2] We propose a model [@halperin] for determining the incompressible quantum fluid states[@laughlin] of such plasmas, and confirm the validity of the model by numerical calculations. In addition, we introduce a new generalized composite Fermion (CF) picture[@jain] for the multi-component plasma and use it to predict the low lying bands of angular momentum multiplets for any value of the magnetic field. #### Bound States. {#bound-states. .unnumbered} In a sufficiently strong magnetic field, the only bound electron-hole complexes are the neutral exciton $X^0$ and the spin-polarized charged excitonic ions $X_k^-$ (electron $e^-\equiv X_0^-$, charged exciton $X^-\equiv X_1^-$, charged biexciton $X_2^-$, etc.). [@palacios; @wojs2] All other complexes found at weaker magnetic fields (e.g. spin-singlet charged exciton[@kheng] or spin-singlet biexciton) unbind.[@lerner] The angular momenta of complexes $X^0$ and $X_k^-$ on a Haldane sphere[@haldane1] with monopole strength $2S$ are $l_{X^0}=0$ and $l_{X_k^-}=|S|-k$.[@wojs2] The binding energies of an exciton, $\varepsilon_0=-E_{X^0}$, and of excitonic ions, $\varepsilon_k=E_{X_{k-1}^-}+E_{X^0}-E_{X_k^-}$ ($E_A$ is the energy of complex $A$) are listed in Tab. \[tab1\] for several different values of $2S$. $2S$ $\varepsilon_0$ $\varepsilon_1$ $\varepsilon_2$ $\varepsilon_3$ ------ ----------------- ----------------- ----------------- ----------------- 10 1.3295043 0.0728357 0.0411069 0.0252268 15 1.3045679 0.0677108 0.0395282 0.0262927 20 1.2919313 0.0647886 0.0381324 0.0260328 : Binding energies $\varepsilon_0$, $\varepsilon_1$, $\varepsilon_2$, and $\varepsilon_3$ of $X^0$, $X^-$, $X_2^-$, and $X_3^-$, respectively, in the units of $e^2/\lambda$. \[tab1\] It is apparent that $\varepsilon_0>\varepsilon_1>\varepsilon_2> \varepsilon_3$. Depending on the ratio $N_e\!:\!N_h$, we expect to find different combinations of complexes that have the largest total binding energy. When $N_e=N_h$ we expect $N_h$ neutral excitons $X^0$ to form. When $N_e\ge2N_h$ the low lying states will contain $N_h$ charged excitons $X^-$ and $N_e-2N_h$ free electrons $e^-$. For $N_h<N_e<2N_h$ we expect to find larger charged exciton complexes. #### Pseudopotentials. {#pseudopotentials. .unnumbered} Whether the states with largest binding energy form the lowest energy band of the electron-hole system depends on the interaction between charged complexes $X_k^-$. The interaction of a pair of charged particles $A$ and $B$ of angular momentum $l_A$ and $l_B$ can be described by a pseudopotential $V_{AB}(L)$ where $\hat{L}=\hat{l}_A+\hat{l}_B$ is the total pair angular momentum.[@haldane2] It is convenient to plot pseudopotentials as a function of the relative angular momentum ${\cal R}=l_A+l_B-L$.[@wojs3] Fig. \[fig1\] shows $V_{AB}({\cal R})$ for the pairs $e^-e^-$, $e^-X^-$, $X^-X^-$, and $e^-X_2^-$, at the monopole strength $2S=17$. =3.35in Roughly, the pseudopotential parameters $V_{AB}({\cal R})$ calculated for different pairs $AB$ and for a given $2S$ lie on the same curve. Small differences between energies $V_{AB}$ calculated for different pairs at the same ${\cal R}$ are due to different values of $l_A$ and $l_B$ and to the finite size and polarization of composite particles. Only the latter effect, important at small ${\cal R}$, persists for $2S\rightarrow\infty$, i.e. in the planar geometry. The major and critical difference between four plotted pseudopotentials lies in the allowed values of ${\cal R}$. If all $A$ and $B$ were point charges, the allowed pair angular momenta for two identical Fermions ($A=B$) would be $L=2l_A-j$, where $j$ is an odd integer, i.e. ${\cal R}=1$, 3, … and ${\cal R}\le2l_A$. For two distinguishable particles ($A\ne B$), the values of $L$ would satisfy $|l_A-l_B|\le L\le l_A+l_B$, i.e. ${\cal R}=0$, 1, 2, …and ${\cal R}\le2\min(l_A,l_B)$. However, if $A$ or $B$ is a composite particle, one or more pair states with largest $L$ (smallest ${\cal R}$) are forbidden, and the corresponding pseudopotential parameters are effectively infinite ($AB$ repulsion has a hard core). For $A=X_{k_A}^-$ and $B=X_{k_B}^-$, the smallest allowed ${\cal R}$ can be deduced from the mapping[@lerner] between the electron-hole and two-spin systems, $${\cal R}_{AB}^{\rm min}=2\min(k_A,k_B)+1. \label{eq1}$$ Thus, in Fig. \[fig1\], ${\cal R}_{e^-X^-}\ge1$, ${\cal R}_{X^-X^-}\ge3$, etc. Low lying states of a system of $N_e$ electrons and $N_h$ holes can contain a number of charged complexes $X_k^-$ ($X^-$ and possibly larger ones) interacting with one another and with electrons through appropriate pseudopotentials. It has been shown[@wojs3] that the Laughlin $\nu=1/m$ state occurs in the gas of (identical) Fermions if the pseudopotential increases faster than linearly as a function of $L(L+1)$ in the vicinity of ${\cal R}=m$. As seen in the inset in Fig. \[fig1\], this is true for both $V_{e^-e^-}$ and $V_{X^-X^-}$, and also (at even values of ${\cal R}$) for $V_{e^-X^-}$ and $V_{e^-X_2^-}$. In Ref.  we found Laughlin states of one-component $X^-$ gas formed at $N_e=2N_h$. In the present note we concentrate on a more general situation, where more than one kind of charged particles occur in an electron-hole system, and find incompressible fluid states of such multi-component plasma. #### Numerical Results. {#numerical-results. .unnumbered} As an illustration, we present first the results of exact diagonalization performed for the system with $N_e=8$ and $N_h=2$. We expect low lying bands of states containing the following combinations of complexes: (i) $4e^-+2X^-$, (ii) $5e^-+X_2^-$, (iii) $5e^-+X^-+X^0$, and (iv) $6e^-+2X^0$. All groupings (i)–(iv) contain an equal number $N=N_e-N_h$ of singly charged complexes, however, both the angular momenta of involved complexes and the relevant hard cores are different. The total binding energies are: $\varepsilon_{\rm i}=2\varepsilon_0+2\varepsilon_1$, $\varepsilon_{\rm ii}=2\varepsilon_0+\varepsilon_1+\varepsilon_2$, $\varepsilon_{\rm iii}=2\varepsilon_0+\varepsilon_1$, and $\varepsilon_{\rm iv}=2\varepsilon_0$. Clearly, $\varepsilon_{\rm i}>\varepsilon_{\rm ii}>\varepsilon_{\rm iii}> \varepsilon_{\rm iv}$. However, which of the groupings contains the (possibly incompressible) ground state depends upon not only the total binding energy, but the interactions between all the charged particles which depends on $2S$. In Fig. \[fig2\], we show the low energy spectra of the $8e+2h$ system at $2S=9$ (a), $2S=13$ (c), and $2S=14$ (e). =3.35in Filled circles mark the non-multiplicative states, and the open circles and squares mark the multiplicative states with one and two decoupled excitons, respectively. In frames (b), (d) and (f) we plot the low energy spectra of different charge complexes interacting through appropriate pseudopotentials (see Fig. \[fig1\]), corresponding to four possible groupings (i)–(iv). By comparing left and right frames, we can identify low lying states of type (i)–(iv) in the electron-hole spectra. In general, energies calculated from pseudopotentials $V_{AB}$ in Fig. \[fig2\] underestimate energies of the corresponding electron-hole system if $N$ and $2S$ are large. This can be partially understood in terms of polarization effects in the two-particle pseudopotentials. For a particular grouping and value of $2S$, it is possible to calculate pseudopotentials that give a very good fit to the electron-hole spectrum. The “correct” pseudopotentials for the $8e+2h$ system are close to those of a pair of point charges with appropriate angular momenta $l_A$ and $l_B$, except for the hard cores.[@wojs2] It is unlikely that a system containing a large number of different species (e.g. $e^-$, $X^-$, $X_2^-$, etc.) will form the absolute ground state of the electron-hole system. However, different charge configurations can form low lying excited bands. An interesting example is the $12e+6h$ system at $2S=17$. The $6X^-$ grouping (v) has the maximum total binding energy $\varepsilon_{\rm v}=6\varepsilon_0+6\varepsilon_1$. Other expected low lying bands correspond to the following groupings: (vi) $e^-+5X^-+X^0$ with $\varepsilon_{\rm vi}=6\varepsilon_0+5 \varepsilon_1$ and (vii) $e^-+4X^-+X_2^-$ with $\varepsilon_{\rm vii} =6\varepsilon_0+5\varepsilon_1+\varepsilon_2$. Although we are unable to perform an exact diagonalization for the $12e+6h$ system in terms individual electrons and holes, we can use appropriate pseudopotentials and binding energies of groupings (v)–(vii) to obtain the low lying states in the spectrum. The results are presented in Fig. \[fig3\]. =3.35in There is only one $6X^-$ state (the $L=0$ Laughlin $\nu_{X^-}=1/3$ state[@wojs2]) and two bands of states in each of groupings (vi) and (vii). A gap of 0.0626 $e^2/\lambda$ separates the $L=0$ ground state from the lowest excited state. #### Generalized Laughlin Wavefunction. {#generalized-laughlin-wavefunction. .unnumbered} It is known that if the pseudopotential $V({\cal R})$ decreases quickly with increasing ${\cal R}$, the low lying multiplets avoid (strongly repulsive) pair states with one or more of the smallest values of ${\cal R}$.[@wojs3; @wojs4] For the (one-component) electron gas on a plane, avoiding pair states with ${\cal R}<m$ is achieved with the factor $\prod_{i<j}(x_i-x_j)^m$ in the Laughlin $\nu=1/m$ wavefunction. For a system containing a number of distinguishable types of Fermions interacting through Coulomb-like pseudopotentials, the appropriate generalization of the Laughlin wavefunction will contain a factor $\prod(x^{(a)}_i-x^{(b)}_j)^{m_{ab}}$, where $x^{(a)}_i$ is the complex coordinate for the position of $i$th particle of type $a$, and the product is taken over all pairs. For each type of particle one power of $(x^{(a)}_i-x^{(a)}_j)$ results from the antisymmetrization required for indistinguishable Fermions and the other factors describe Jastrow type correlations between the interacting particles. Such a wavefunction guarantees that ${\cal R}_{ab}\ge m_{ab}$, for all pairings of various types of particles, thereby avoiding large pair repulsion.[@halperin; @haldane2] Fermi statistics of particles of each type requires that all $m_{aa}$ are odd, and the hard cores defined by Eq. (\[eq1\]) require that $m_{ab}\ge{\cal R}_{ab}^{\rm min}$ for all pairs. #### Generalized Composite Fermion Picture. {#generalized-composite-fermion-picture. .unnumbered} In order to understand the numerical results obtained in the spherical geometry (Figs. \[fig2\] and \[fig3\]), it is useful to introduce a generalized CF picture by attaching to each particle fictitious flux tubes carrying an integral number of flux quanta $\phi_0$. In the multi-component system, each $a$-particle carries flux $(m_{aa}-1) \phi_0$ that couples only to charges on all other $a$-particles and fluxes $m_{ab}\phi_0$ that couple only to charges on all $b$-particles, where $a$ and $b$ are any of the types of Fermions. The effective monopole strength [@chen; @jain; @wojs3; @sitko] seen by a CF of type $a$ (CF-$a$) is $$2S_a^*=2S-\sum_b(m_{ab}-\delta_{ab})(N_b-\delta_{ab}) \label{eq2}$$ For different multi-component systems we expect generalized Laughlin incompressible states (for two components denoted as $[m_{AA},m_{BB}, m_{AB}]$) when all the hard core pseudopotentials are avoided and CF’s of each kind fill completely an integral number of their CF shells (e.g. $N_a=2l_a^*+1$ for the lowest shell). In other cases, the low lying multiplets are expected to contain different kinds of quasiparticles (QP-$A$, QP-$B$, …) or quasiholes (QH-$A$, QH-$B$, …) in the neighboring incompressible state. Our multi-component CF picture can be applied to the system of excitonic ions, where the CF angular momenta are given by $l_{X_k^-}^*=|S_{X_k^-}^*|-k$. As an example, let us first analyze the low lying $8e+2h$ states in Fig. \[fig2\]. At $2S=9$, for $m_{e^-e^-}=m_{X^-X^-}=3$ and $m_{e^-X^-}=1$ we predict the following low lying multiplets in each grouping: (i) $2S_{e^-}^*=1$ and $2S_{X^-}^*=3$ gives $l_{e^-}^*=l_{X^-}^*=1/2$. Two CF-$X^-$’s fill their lowest shell ($L_{X^-}=0$) and we have two QP-$e^-$’s in their first excited shell, each with angular momentum $l_{e^-}^*+1=3/2$ ($L_{e^-}=0$ and 2). Addition of $L_{e^-}$ and $L_{X^-}$ gives total angular momenta $L=0$ and 2. We interpret these states as those of two QP-$e$’s in the incompressible \[331\] state. Similarly, for other groupings we obtain: (ii) $L=2$; (iii) $L=1$, 2, and 3; and (iv) $L=0$ ($\nu=2/3$ state of six electrons). At $2S=13$ and 14 we set $m_{e^-e^-}=m_{X^-X^-}=3$ and $m_{e^-X^-}=2$ and obtain the following predictions. First, at $2S=13$: (i) The ground state is the incompressible \[332\] state at $L=0$; the first excited band should therefore contain states with one QP-QH pair of either kind. For the $e^-$ excitations, the QP-$e^-$ and QH-$e^-$ angular momenta are $l_{e^-}^*=3/2$ and $l_{e^-}^*+1=5/2$, respectively, and the allowed pair states have $L_{e^-}=1$, 2, 3, and 4. However, the $L=1$ state has to be discarded, as it is known to have high energy in the one-component (four electron) spectrum.[@sitko] For the $X^-$ excitations, we have $l_{X^-}^*=1/2$ and pair states can have $L_{X^-}=1$ or 2. The first excited band is therefore expected to contain multiplets at $L=1$, $2^2$, 3, and 4. The low lying multiplets for other groupings are expected at: (ii) $L=2$ and 3; (iii) $2S_{X_2^-}^*=3$ gives no bound $X_2^-$ state; setting $m_{e^-X^-}=1$ we obtain $L=2$; and (iv) $L=0$, 2, and 4. Finally, at $2S=14$ we obtain: (i) $L=1$, 2, and 3; (ii) incompressible \[3\*2\] state at $L=0$ ($m_{X^-X^-}$ is irrelevant for one $X^-$) and the first excited band at $L=1$, 2, 3, 4, and 5; (iii) $L=1$; and (iv) $L=3$. For the $12e+6h$ spectrum in Fig. \[fig3\] the following CF predictions are obtained: (v) For $m_{X^-X^-}=3$ we obtain the Laughlin $\nu=1/3$ state with $L=0$. Because of the hard core of $V_{X^-X^-}$, this is the only state of this grouping. (vi) We set $m_{X^-X^-}=3$ and $m_{e^-X^-}=1$, 2, and 3. For $m_{e^-X^-}=1$ we obtain $L=1$, 2, $3^2$, $4^2$, $5^3$, $6^3$, $7^3$, $8^2$, $9^2$, 10, and 11. For $m_{e^-X^-}=2$ we obtain $L=1$, 2, 3, 4, 5, and 6. For $m_{e^-X^-}=3$ we obtain $L=1$. (vii) We set $m_{X^-X^-}=3$, $m_{e^-X_2^-}=1$, $m_{X^-X_2^-}=3$, and $m_{e^-X^-}=1$, 2, or 3. For $m_{e^-X^-}=1$ we obtain $L=2$, 3, $4^2$, $5^2$, $6^3$, $7^2$, $8^2$, 9, and 10. For $m_{e^-X^-}=2$ we obtain $L=2$, 3, 4, 5, and 6. For $m_{e^-X^-}=3$ we obtain $L=2$. In groupings (vi) and (vii), the sets of multiplets obtained for higher values of $m_{e^-X^-}$ are subsets of the sets obtained for lower values, and we would expect them to form lower energy bands since they avoid additional small values of ${\cal R}_{e^-X^-}$. However, note that the (vi) and (vii) states predicted for $m_{e^-X^-}=3$ (at $L=1$ and 2, respectively) do not form separate bands in Fig. \[fig3\]. This is because the $V_{e^-X^-}$ pseudopotential increases more slowly than linearly as a function of $L(L+1)$ in the vicinity of ${\cal R}_{e^-X^-}=3$ (see Fig. \[fig1\]). In such case the CF picture fails.[@wojs3] The agreement of our CF predictions with the data in Figs. \[fig2\] and \[fig3\] (marked with lines) is really quite remarkable and strongly indicates that our multi-component CF picture is correct. We were indeed able to confirm predicted Jastrow type correlations in the low lying states by calculating their coefficients of fractional parentage.[@wojs3; @shalit] We have also verified the CF predictions for other systems that we were able to treat numerically. If exponents $m_{ab}$ are chosen correctly, the CF picture works well in all cases. #### Summary. {#summary. .unnumbered} Charged excitons and excitonic complexes play an important role in determining the low energy spectra of electron-hole systems in a strong magnetic field. We have introduced general Laughlin type correlations into the wavefunctions, and proposed a generalized CF picture to elucidate the angular momentum multiplets forming the lowest energy bands for different charge configurations occurring in the electron-hole system. We have found Laughlin incompressible fluid states of multi-component plasmas at particular values of the magnetic field, and the lowest bands of multiplets for various charge configurations at any value of the magnetic field. It is noteworthy that the fictitious Chern–Simons fluxes and charges of different types or colors are needed in the generalized CF model. This strongly suggests that the effective magnetic field seen by the CF’s does not physically exist and that the CF picture should be regarded as a mathematical convenience rather than physical reality. Our model also suggests an explanation of some perplexing observations found in photoluminescence, but this topic will be addressed in a separate publication. We thank P. Hawrylak and M. Potemski for helpful discussions. AW and JJQ acknowledge partial support from the Materials Research Program of Basic Energy Sciences, US Department of Energy. KSY acknowledges support from the Korea Research Foundation (Project No. 1998-001-D00305). K. Kheng, R. T. Cox, Y. Merle d’Aubigne, F. Bassani, K. Saminadayar, and S. Tatarenko, Phys. Rev. Lett. [**71**]{}, 1752 (1993); H. Buhmann, L. Mansouri, J. Wang, P. H. Beton, N. Mori, M. Heini, and M. Potemski, Phys. Rev. B [**51**]{}, 7969 (1995). A. J. Shields, M. Pepper, M. Y. Simmons, and D. A. Ritchie, Phys. Rev. B [**52**]{}, 7841 (1995); G. Finkelstein, H. Shtrikman, and I. Bar-Joseph, Phys. Rev. B [**53**]{} 1709, (1996). X. M. Chen and J. J. Quinn, Phys. Rev. B [**50**]{}, 2354 (1994); ibid. [**51**]{}, 5578 (1995). A. Wójs and P. Hawrylak, Phys. Rev. B [**51**]{} 10 880, (1995). E. I. Rashba and M. E. Portnoi, Phys. Rev. Lett. [**70**]{}, 3315 (1993); V. M. Apalkov, F. G. Pikus, and E. I. Rashba, Phys. Rev. B [**52**]{}, 6111 (1995). J. J. Palacios, D. Yoshioka, and A. H. MacDonald, Phys. Rev. B [**54**]{}, 2296 (1996). A. Wójs, P. Hawrylak, and J. J. Quinn, Physica B [**256–258**]{}, 490 (1998); Phys. Rev. Lett. (submitted, cond-mat/9810082) I. V. Lerner and Yu. E. Lozovik, Sov. Phys. JETP [**53**]{}, 763 (1981); A. H. MacDonald and E. H. Rezayi, Phys. Rev. B [**42**]{}, 3224 (1990). F. D. M. Haldane, Phys. Rev. Lett. [**51**]{}, 605 (1983); T. T. Wu and C. N. Yang, Nucl. Phys. B [**107**]{}, 365 (1976). B. I. Halperin, Helv. Phys. Acta [**56**]{}, 75 (1983). R. Laughlin, Phys. Rev. Lett. [**50**]{}, 1395 (1983). J. Jain, Phys. Rev. Lett. [**63**]{}, 199 (1989). F. D. M. Haldane and E. H. Rezayi, Phys. Rev. Lett. [**60**]{}, 956 (1988). A. Wójs and J. J. Quinn, Solid State Commun. [**108**]{}, 493 (1998); ibid. [**110**]{}, 45 (1999); Phys. Rev. B (submitted, cond-mat/9903145). A. Wójs and J. J. Quinn, Physica E [**3**]{}, 181 (1998). P. Sitko, S. N. Yi, K.-S. Yi, and J. J. Quinn, Phys. Rev. Lett. [**76**]{}, 3396 (1996). A. de Shalit and I. Talmi, [*Nuclear Shell Theory*]{}, Academic Press, New York and London 1963.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In flat spacetime, quantum fluctuations in dark matter, as described as a Bose-Einstein condensate, are stable and display a relativistic Bogoliubov dispersion relation. In the weak gravitational field limit, both relativistic and nonrelativistic models self-gravitating dark matter suggest the formation of structures as the result of a dynamical (Jeans) instability. Here, we show that in the presence of spontaneous symmetry breaking of the dark matter field, the gravitational wave is damped for wave-lengths larger than the Jeans length. Such energy is converted to the Bogoliubov modes of the BEC that in their turn become unstable and grow, leading to the formation of structures even in the absence of expansion. Remarkably, this compensated attenuation/amplification mechanism is the signature of a discrete $\cal PT$ symmetry-breaking of the system.' author: - Catarina Bastos - Hugo Terças title: ' Violation of the $\cal PT-$symmetry and structure formation in the dark matter $-$gravitational wave interaction ' --- [*Introduction.*]{} Scalar field theories in curved spacetimes are on the basis of modern advances in cosmology and astrophysics [@kolb_1989; @zee_2003; @dodelson_2003], as they constitute important candidates to explain the behaviour of dark-matter [@chavanis_2011; @harko_2011a; @suarez_2014; @chavanis_2015a]. Their relativistic dynamics is governed by the Klein-Gordon (KG) equation [@klein_gordon], historically emerging as a first attempt to unify quantum mechanics with the special relativity theory, to obtain a unified theory to explain our universe. The Klein-Gordon-Einstein (KGE) equations, that involve a coupling between the KG equations and gravity, were first used to study boson stars [@bosonstars]. Although, the first models of these stars do not assumed that bosons could have a self interaction potential, it was shown then that self-interaction can significantly change the physical dimensions of the boson stars and make them clearly more interesting as an astrophysical object [@colpi_1986]. Furthermore, models of dark matter (DM) halos were proposed based on scalar fields that are described by KGE equations [@suarez_2014]. These DM halos can be explained through Schrödinger-Poisson (SP) equation or Gross-Pitaevskii-Poisson (GPP) equations, since the Newtonian limit is valid at the galactic scale. In this sense, we can think about DM halos as gigantic quantum objects made of Bose-Einstein condensates (BECs). Furthermore, these models are a tentative of solving the problem of Cold Dark Matter (CDM), as the wave properties of bosons can stabilize the system against gravitational collapse. At the cosmological level it is quite important to study the implications of these scalar field models. It was shown by Matos et al. [@matos_2009] that when a spatially homogenous interacting real scalar field competes with baryonic matter, radiation and dark energy in terms of cosmological evolution, these real scalar fields can reproduce quite well the cosmological predictions of the $\Lambda$-CDM model. A perturbative analysis then showed the formation of structures corresponding to DM halos. Finally, Chavanis [@chavanis_2011] has considered the case of a complex self-interacting scalar field in the context of Newtonian cosmology and based on the GPP equations. The formation of structures has been recently studied through the Jeans instability of an homogeneous self-gravitating BEC in a static background [@chavanis_2015]. Basically, the so-called BECDM have shown that perturbations grow faster than in a $\Lambda$-CDM model. Some relativistic models have been then analyzed [@chavanis_2011; @harko_2011b]. In the last year, gravitational waves (i.e. fluctuations in the metric) generated by accelerated mass distributions, like massive black holes, were finally detected by the LIGO collaboration [@LIGO2015]. The theory of general relativity predicts that the amplitude of these gravitational is extremely small, which harnessed their detection for a long time. Although gravitational waves produced by black hole collisions could be detected, finding experimental evidence of primordial gravitational waves remains elusive. Nevertheless, the advent of table-top, high-sensitivity devices based on quantum technologies revived their interest. Also, ESA is now developing the eLiSA, a space-based interferometer that will be used to detect gravitational waves in other range of frequencies, the low-frequency band [@amaro-seoane2012]. It is expected that it could detect waves coming from other sources rather than merging black holes. On the other hand, a lot of attention has been drawn to the study of space-time effects in quantum systems as, for instance, in phononic fields [@ahmadi2013; @sabin2014]. It was shown in that gravitational waves can create phonons in a BEC [@sabin2014], a features that is motivating a new generation of gravitational-wave detectors using matter waves, which may become a reality in a medium-term timescale. In this work, we investigate the self-consistent dynamics of an interacting complex scalar field (BEC) - described by a nonlinear Klein-Gordon equation - evolving in a fluctuating space-time. In particular, we show that gravitational waves (obtained from Einstein’s equation in the weak field limit) can couple to scalar field fluctuations, leading the dynamics of the latter unstable. Remarkably, the presente instability mechanism appears to be associated with the violation of the discrete parity-time ($\mathcal{PT}$) symmetry. The latter, initially proposed as a concept in quantum mechanics [@bender_98], is now being extensively studied in optics [@zyablovsky_2014; @suchkov_2015; @konotop_2016] and, more recently, in acoustics [@zhu_2014]. In such systems, any spatial region with a loss is mirrored by a region of gain. Therefore, the processes of light (or sound) absorption and amplification can be compensated, and the frequencies of the eigen optical (acoustic) modes can be real. When $\mathcal{PT}$-symmetry is broken, the eigenmodes appear in complex conjugate pairs. In our case, complex eigenmodes appear for hybrid modes made of the mixture between gravity and Bogoliubov (sound) BEC modes. This suggests that the formation of structures, as described by the Jeans self-gravitating instability, due to primordial gravitational waves is a consequence of the breaking of the $U(1)\times \mathcal{PT}$ symmetry. Our findings show that the gravitational wave is damped for wave lengths larger than the Jeans length and the energy is converted to the Bogoliubov modes of the BEC, which grow in time. This will turn the system unstable, leading to the formation of primordial cosmological structures even in the absence of an expanding universe. Moreover, we argue that this particular form of the space-time$-$field interaction may be an important mechanism preventing the detection of primordial gravitational waves, as their energy is transferred to the matter field originating structures in the universe. [*Minimally coupled theory.*]{} The dynamics of a complex scalar field (SF) $\varphi(x_\mu)$ in a curved spacetime of curvature $R$ is governed by the following minimal-coupling action $$\mathcal{S}=\frac{c^4}{16\pi G}\int d^4x \sqrt{-g} R+\int d^4 x\sqrt{-g} \mathcal{L_\varphi}, \label{eq_action}$$ where $x_\mu=(-c t, \textbf{x})$ is the four vector, $g_{\mu \nu}$ is the metric tensor, $g=g^\mu_\mu$ denotes its trace, and $\mathcal{L}_\varphi$ is the SF Lagrangian $$\mathcal{L}_\varphi=\frac{1}{2}g^{\mu\nu}\partial_\mu \varphi^*\partial_\nu \varphi-V\left(\vert \varphi\vert^2\right).$$ Here, $V\left(\vert \varphi\vert^2\right)$ contains the KG rest mass term and the self-interaction potential, $$V\left(\vert \varphi\vert^2\right)=\frac{m^2 c^2}{2\hbar^2}\vert \varphi\vert^2+\frac{1}{4}\lambda \vert \varphi\vert^4,$$ where $\lambda=8 \pi a_s m/\hbar^2$ is the coupling constant, $a_s$ is the scattering length and $m$ is the field mass. The minimization of Eq. with respect to $\varphi$ provides the Euler-Lagrange equation $\nabla_\mu \left[ \frac{\partial \mathcal{L}_\varphi}{\partial \left( \partial_\mu \varphi\right)^*}-\frac{\partial \mathcal{L}_\varphi}{\partial\varphi^*}\right]=0$, which in turn yields the following generalized KG equation $$\square_g \varphi+V'\left(\vert \varphi \vert^2\right)_{,\varphi^*}=0, \label{eq_KG1}$$ where $\square_g\equiv g^{\mu\nu}\partial_\mu \partial_\nu-g^{\mu\nu}\Gamma_{\mu\nu}^\alpha \partial_\alpha$ is the generalized d’Alembert operator and $\Gamma_{\mu \nu}^\alpha$ denotes the Christoffel symbol [@schutz]. In Eq. , we made use of the parallel transport of the metric $\nabla_\mu g^{\mu \nu}=0$. A similar minimization procedure with respect to the metric $g_{\mu \nu}$ leads to the Einstein field equations $$R_{\mu \nu}-\frac{1}{2}g_{\mu\nu} R=\kappa T_{\mu\nu}, \label{eq_E1}$$ with $\kappa= 8\pi G/c^4$ and $T_{\mu\nu}$ is the energy-momentum tensor $$\begin{array}{ll} T_{\mu \nu}=\frac{1}{2} &\left( \partial_\mu \varphi^* \partial_\nu \varphi +\partial_\nu \varphi^* \partial_\mu \varphi\right)\\ &-g_{\mu\nu}\left[ \frac{1}{2} g^{\alpha \beta}\partial_\alpha \varphi^*\partial_\beta \varphi-V\left(\vert \varphi\vert^2\right)\right]. \end{array}$$ [*Perturbative analysis.*]{} We now assume a perturbation around the Minkowski space time of the form $g_{\mu\nu}=\eta_{\mu\nu}+h_{\mu\nu}$, where $h_{\mu\nu}\ll \eta_{\mu\nu}$ is the spacetime ripple and $\eta_{\mu\nu}={\rm diag}(-,+,+,+)$. To first order in $h_{\mu\nu}$, Eq. reads $$\square \varphi +V'\left(\vert \varphi \vert^2\right)_{,\varphi^*}+h^{\mu\nu}\partial_\mu\partial_\nu\varphi -\eta^{\mu \nu}\gamma_{\mu\nu}^\alpha \partial_\alpha \varphi=0, \label{eq_KG2}$$ where $\gamma_{\mu\nu}^\alpha=\frac{1}{2}\left( \partial^\alpha h_{\mu\nu}+\partial_\nu h_{\mu}^{\alpha}-\partial_\mu h_{\nu}^{\alpha}\right)$, with $h_{\mu}^{\nu}=\eta^{\alpha \nu}h_{\mu\alpha}$. Making use of the trasnverse-traceless (TT) gauge, $\partial_\mu h_{\nu}^{\mu}=0$ and $h\equiv h_{\mu}^{\mu}=0$, the last term in Eq. vanishes and the KG equation explicitly reads $$\square \varphi+\frac{m^2 c^2}{\hbar ^2}\varphi+\lambda \vert \varphi\vert^2 \varphi+h^{\mu\nu}\partial_{\mu}\partial_{\nu}\varphi=0, \label{eq_KG3}$$ where we made use of the property $h^{\mu\nu}=h_{\mu\nu}$. Similarly, the weak-field limit of Eq. describes spacetime radiation (gravitational waves) in the presence of matter $$\square h_{\mu\nu}=-2\kappa T_{\mu\nu}, \label{eq_E2}$$ with $\square=-(1/c^2)\partial_t^2+\nabla^2$. In what follows, we introduce quantum fluctuations around the homogeneous scalar field (i.e. the vacuum expectation value $\langle \varphi \rangle =\sqrt{n_0}$ spontaneously breaking the continuous $U(1)$ symmetry) in the form $$\varphi(x_\mu)=\sqrt{n_0} e^{-i\mu t/\hbar}\left[ 1+\sum _k\left( u_k e^{ik_\mu x^\mu}+v_k^* e^{-ik_\mu x^\mu}\right)\right], \label{eq_phi}$$ where $\mu$ is the chemical potential of the condensate. By dividing the metric fluctuation into its time and space components, $h_{\mu,\nu}=h_{00}+h_{ij}\equiv 2\phi/c^2+h_{ij}$, we can obtain purely transverse solutions satisfying the condition $k^jh_{ij}=0$ as $$\left(\square -\frac{m_{\rm eff}^2 c^2}{\hbar^2}\right) h_{ij}=0,$$ where $m_{\rm eff}^2=2\hbar ^2\kappa V(n_0)/c^2$ is the square of the effective graviton mass. Assuming plane-wave solutions of the form $h_{ij}=\sum_k \chi_{ij}e^{ik_\mu x^\mu}$, we obtain the dispersion relation $$\omega^2 =\omega_p^2+c^2k^2,\quad \omega_p^2=\frac{m_{\rm eff }^2 c^4}{\hbar^2}. \label{eq_plasma}$$ The latter is obtained by making use of the equation of state obtained at the zeroth order, which fixes the chemical potential of the BEC as $$\mu=\sqrt{m^2 c^4+\hbar^2 \lambda n_0 c^2}=mc^2\sqrt{1+\frac{c_s^2}{c^2}},$$ where $c_s=\hbar /(m\sqrt{\lambda n_0 })$ is the BEC sound speed. The dispersion relation in Eq. is analogous of that of an electromagnetic wave propagating in a charged medium characterized by a plasma frequency $\omega_p$, where the photon also acquires an effective mass. Consequently, the KG equation decouples from Einstein’s equations and the Bogoliubov modes for the scalar field can be obtained from Eqs. and . Their dispersion relation can be found from plugging Eq. in Eq. and separating it into its particle (anti-particle) coefficients $\propto e^{ik_\mu x^\mu}$ ($\propto e^{-ik_\mu x^\mu}$). The resulting secular equation contains two real solutions $\omega_{\pm}$ (and the corresponding “anti-mode" solutions $\omega^*_{\pm}=-\omega_{\pm}$) $$\omega_\pm^2=2\omega_0^2+c^2 k^2 \pm 2\sqrt{\omega_0^4+c^2k^2\omega_0^2}, \label{eq_modes}$$ where $\omega_0=mc^2\left(1+\beta^2\right)/\hbar$, with $\beta^2=3 c_s^2/(2c^2)$, is the cut-off frequency. In the long wavelength limit $k\ll \xi (1+\beta^2)$ - with $\xi=\hbar/m c_s$ denoting the healing length - the lower (Goldstone) mode is [*gapeless*]{}, $\omega_-\simeq c_s(1+\beta^2)k+\xi c_s k^2/2(1+\beta^2)^3$, reducing to the usual Bogoliubov dispersion in the non-relativistic limit $\beta \rightarrow 0$ [@pitaevskii_book; @pethick_book]. The [*gapped*]{} mode, corresponding to the massive Higgs mode of mass $M=\hbar \omega_0/c^2=m(1+\beta^2)$, reads $\omega_+\simeq \sqrt{M^2 c_s^4/\hbar^2 +c_s^2 k^2}$. Although with a different notation, the dispersion modes of Eq. have first been discussed in Ref. [@fagnocchi]. ![Dispersion relation of the various modes present in the dynamics. Top panel: mode dispersion in the absence of coupling ($G=0$). We observe that the Bogoliubov-Goldstone modes are $\mathcal{PT}$-symmetric. Bottom panel: when the gravity is switched on, the Goldstone modes lose their symmetry. The gravity mode is damped to favour unstable (growing) modes in the BEC. The imaginary part of the frequencies goes to zero at the Jeans mode $k_J$, exhibiting the usual signature of $\mathcal{PT}-$ symmetry breaking. For illustration purposes, we use $c_s=c/3$.[]{data-label="fig_disp"}](Fig1.pdf) The situation changes if we consider perturbations in the time-time components of Einstein’s equations , i.e., for gravitational waves of the form $h_{\mu\nu}\simeq h_{00}=2\phi/c^2$. This amounts to generalize the usual self-gravitating problem, as described by the Klein-Gordon-Poisson system [@chavanis_2015], to the study of propagation of gravitational radiation in a symmetry-broken quantum vacuum. As we are about to see, the formation of structures emerges in this case as a consequence of the violation of the discrete $\mathcal{PT}-$ symmetry. Putting Eqs. and together, and keeping terms to the first order in the Fourier components of the vector $\bm{V}_k=(u_k, v_k, \phi_k)$, we obtain the eigenvalue problem $\mathcal{L}_k \bm{V}_k=0$, where $$\mathcal{L}_k=\left[ \begin{array}{ccc} \epsilon_1^2 &-m^2 c^2 c_s^2 &- 2\mu^2\\ -m^2 c^2 c_s^2 & \epsilon_2^2 & 0 \\ -\tilde\kappa \left(\mu -\hbar \omega \right)^2 & -\tilde\kappa \left(\mu +\hbar \omega \right)^2 &\epsilon_0^2 \end{array}\right],$$ with $\tilde \kappa= 8\pi G n_0/c^4$, $\epsilon_{1,2}^2=\left( \mu\pm \hbar \omega\right)^2-\hbar^2 c^2 k^2-m^2c^4-2m^2 c_s^4$, and $\epsilon_0^2=\hbar^2(\omega^2-c^2k^2)$. Nontrivial solutions are obtained by solving the secular equation $\det (\mathcal{L}_k)=0$ in respect to $\omega$, for which we obtain six solutions (three for positive-energy and other three for negative-energy excitations). For zero gravity-matter coupling ($G=0$), the positive-energy modes are the BEC Goldstone-Bogoliubov and the Higgs, as described in Eq. , and the gravity mode $\omega=ck$. In this situation, all modes are real and therefore dynamically stable (see Fig. \[fig\_disp\] a). In the presence of gravity, however, the Bogoliubov and the gravity modes hybridize and collapse, exhibiting an imaginary part for $k-$modes below the Jeans wave vector $k_J$ that satisfy the condition $\Im(\omega)=0$, for which we obtain the equation. $$\hbar ^4k_J^4 + 2 \hbar ^2 k_J^2 m^2 c_s^2 -\tilde \kappa m^4 (c^2 + c_s^2)^2=0.$$ Remarkably, the imaginary part of the Bogoliubov and the gravity modes have opposite signs, suggesting that the formation of dark-matter cosmological structures (triggered by the long-wavelength dynamical instability) is accompanied by the damping of space-time perturbations. In other words, the gravitational waves transfer their energy to the BEC modes so the latter can grow. Because $Re(\omega)>0$, a $I_o$-type of instability [@cross_93] is responsible for the formation of large structures in flat spacetimes. Also, we observe that the positive and negative Bogoliubov modes are not symmetric, i.e. $\omega_-\neq -\omega_-^*$, indicating violation of the $\mathcal{PT}-$symmetry. These features are depicted in Fig. \[fig\_disp\] b). ![One-dimensional illustration of the structure formation dynamics at early stages of the instability onset. Panel a) shows the initial ($t=0$) plane-wave superposition solution for the gravitational wave $\phi(x)$ (black line) and the BEC $\varphi(x)$ (lighter line). Panels b), c) and d) depict their evolution at $t=1.5 \hbar/mc^2$, $t=4.5 \hbar/mc^2$ and $t=5.5 \hbar/mc^2$, respectively. The shadowed region represent the Jeans length $\lambda_J=2\pi/k_J$. We use $c_s=c/3$. []{data-label="fig_structure"}](Fig2.pdf) In order to illustrate how the $\mathcal{PT}$-symmetry breaking affects the formation of structures, we perform one-dimensional simulations of Eqs. and for the early stages of the Jeans instability. As depicted in Fig. \[fig\_structure\], an initial linear superposition of plane gravitational and Bogoliubov waves (Fig. \[fig\_structure\] a)) lead to the formation of 1d structures in the BEC sector. Short after the onset the instability, the long wavelength structures of the BEC start to grow, leading to the formation of structures of typical size $\lambda_J=2\pi/k_J$. Simultaneously, the gravitational modes in the same wavelength modes attenuate, eventually vanishing out for longer times. We notice that our calculations are valid near the onset of the instability only, i.e for $t\ll \hbar/mc_s^2$, for which a quasi-linear approximation of Eqs. and is valid. A more accurate, quantitative discussion of our results would involve taking into account saturation effects. [*Conclusions.*]{} In this work, we have studied the coupling between a gravitational wave in a Minkowski spacetime with dark matter modelled by a self-interacting complex scalar field (Bose-Einstein condensate). Considering perturbations in the spatial components of the metric only, the gravitational wave dispersion relation is analogous to that of an electromagnetic wave propagating in a charged medium characterized by a plasma frequency $\omega_p$, where the photon also acquires an effective mass. In this case, the two modes (the gravity mode and the Bogoliubov mode) are decoupled. However, when we consider perturbations in the temporal component of the metric, the gravity and the Bogoliubov modes hybridize and become dynamically unstable. Because of the local breaking of the $\mathcal{PT}$- symmetry, the modes form conjugate pairs, in such a way that there is a transfer of energy from the gravitational wave (damping) to the BEC field (growth). In short, this means that the instability mechanism triggering the formation of large dark-matter structures is accompanied by the breaking of the $U(1)\times {\cal PT}$ symmetry. Remarkably, our findings may also constitute an alternative explanation why primordial gravitational waves are quite hard to detect: they just vanish and give away their energy to the formation of large-scale structures. In a near future, our work could strongly benefit from numerical GR tools, both in weak and strong gravity scenarios, which could correctly describe the saturation at later stages due to the nonlinearity in the Klein-Gordon equation and, eventually, the effects of curvature due to the presence of massive objects. [*Acknowledgments.*]{} One of the authors (H. T.) acknowledges the Security of Quantum Information Group for the hospitality and for providing the working conditions, and financial support from Fundação para a Ciência e a Tecnologia (Portugal) through grant number SFRH/BPD/110059/2015. The work of C.B. is supported by the European Research Council (ERC-2010-AdG Grant 267841). [9]{} E. W. Kolb and M. S. Turner, [*The Early Universe*]{}, Addison-Wesley, Reading, MA (1989). A. Zee, [*Quantum Field Theory in a Nutshell*]{}, Princeton University Press, Princeton, NJ (2003). S. Dodelson, [*Modern Cosmology*]{}, Academic Press, New York (2003). P. H. Chavanis and L. Delfini, Phys. Rev. D [**84**]{}, 043531 (2011). T. Harko, J. Cosmol. Astropart. Phys. [**6**]{}, 022 (2011). A. Suar' ez, V. H. Robles, and T. Matos, [*Accelerated Cosmic Expansion*]{}, Astrophysics and Space Science Proceedings [**38**]{}, 107 (2014). P. H. Chavanis, [*Quantum Aspects of Black Holes*]{}, Fundamental Theories of Physics [**178**]{}, Springer, New York (2015). O. Klein, Z. Phys. [**37**]{}, 895 (1926); W. Gordon, Z. Phys. [**40**]{}, 117 (1926). D.J. Kaup, Phys. Rev. [**172**]{}, 1331 (1968); R. Ruffini and S. Bonazzola, Phys. Rev. [**187**]{}, 187 (1969). M. Colpi, S.L. Shapiro and I. Wasserman, Phys.Rev. Lett. [**57**]{}, 2485 (1986). T. Matos, A. Vázquez-González, and J. Magaña, Mon. Not. R. Astron. Soc. [**393**]{}, 1359 (2009). A. Suárez and T. Matos, Mon. Not- R. Astron. Soc. [**416**]{}, 87 (2011). A. Su' arez and P.-H. Chavanis, J. Phys.: Conf. Series [**654**]{}, 012008 (2015); A. Su' arez and P.-H. Chavanis, Phys. Rev. D [**92**]{}, 023510 (2015). T. Harko, Mon. Not. R. Astron. Soc. [**413**]{}, 3095 (2011). LIGO Scientific Collaboration, Phys. Rev. Lett. [**116**]{}, 061102; Phys. Rev. D [**93**]{}, 042005 (2016). A.-Seoane et al., Class. Quant. Grav. [**29**]{} (2012), 124016. M. Ahmadi, D.E. Bruschi, N. Friis, C. Sabin, G. Adesso and I. Fuentes, Sci. Rep. [**4**]{}, 4996 (2014). C. Sabin, D.E. Bruschi, M. Ahmadi and I. Fuentes, New J. Phys [**16**]{} (2014) 085003. C. M. Bender and S. Boettcher, Phys. Rev. Lett. [**80**]{}, 5243Ð5246 (1998). A. A. Zyablovsky, A. P. Vinogradov, A. A. Pukhov, A. V. Dorofeenko, and A. A. Lisyansky, Usp. Fiz. Nauk [**184**]{}, 1177Ð1198 (2014). S. V. Suchkov, A. A. Sukhorukov, J. Huang, S. V. Dmitriev, C. Lee, and Y. S. Kivshar, [`a`rXiv:1509.03378]{} (2005). V. V. Konotop, J. Yang, and D. A. Zezyulin, [`a`rXiv:1603.06826]{} (2016). X. Zhu, H. Ramezani, C. Shi, J. Zhu, and X. Zhang, Phys. Rev. X [**4**]{}, 031042 (2014). B. Schutz, [*A First Course in General Relativity*]{}, Cambridge University Press, 2nd Ed. (2009). L. P. Pitaevskii and S. Stringari, [*BoseÐEinstein Condensation*]{} , Oxford: Clarendon Press (2003). C. J. Pethick and H. Smith, [*BoseÐEinstein Condensation in Dilute Gases*]{}, Cambridge: Cambridge University Press (2008). S. Fagnocchi, S. Finazzi, S. Liberati, M. Kormos, and A. Trombettoni, New J. Phys [**12**]{}, 095012, (2010). M. C. Cross and P. C. Hohenberg, Rev. Mod. Phys. [**65**]{}, 851 (1993).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We employ a certain labeled finite graph, called a chart, in a closed oriented surface for describing the monodromy of a(n achiral) Lefschetz fibration over the surface. Applying charts and their moves with respect to Wajnryb’s presentation of mapping class groups, we first generalize a signature formula for Lefschetz fibrations over the $2$–sphere obtained by Endo and Nagami to that for Lefschetz fibrations over arbitrary closed oriented surface. We then show two theorems on stabilization of Lefschetz fibrations under fiber summing with copies of a typical Lefschetz fibration as generalizations of a theorem of Auroux.' address: - | Department of Mathematics\ Tokyo Institute of Technology\ 2-12-1 Oh-okayama\ Meguro-ku\ Tokyo 152-8551\ Japan - | Ministry of Health\ Labour and Welfare\ 1-2-2 Kasumigaseki\ Chiyoda-ku\ Tokyo 100-8916\ Japan - | Department of Mathematics\ Osaka City University\ 3-3-138 Sugimoto\ Sumiyoshi-ku\ Osaka 558-8585\ Japan - | Department of Mathematics\ Tokyo Gakugei University\ 4-1-1 Nukuikita-machi\ Koganei-shi\ Tokyo 184-8501\ Japan author: - Hisaaki Endo - Isao Hasegawa - Seiichi Kamada - Kokoro Tanaka title: 'Charts, signatures, and stabilizations of Lefschetz fibrations' --- Introduction ============ Matsumoto [@Matsumoto1986] proved that every Lefschetz fibration of genus one over a closed oriented surface is isomorphic to a fiber sum of copies of a holomorphic elliptic fibration on $\mathbb{CP}^2\# 9\overline{\mathbb{CP}}^2$ and a trivial torus bundle over the surface if it has at least one critical point. This result played a crucial role in completing the classification of diffeomorphism types of elliptic surfaces (see Gompf and Stipsicz [@GS1999 Section 8.3]). Although such a classification has not been established for Lefschetz fibrations of higher genus, Auroux [@Auroux2003] proved a stabilization theorem for Lefschetz fibrations of genus two, which states that every Lefschetz fibration of genus two over the $2$–sphere becomes isomorphic to a fiber sum of copies of three typical fibrations after fiber summing with a holomorphic fibration on $\mathbb{CP}^2\# 13\overline{\mathbb{CP}}^2$. Auroux [@Auroux2005] gave a generalization of this theorem for Lefschetz fibrations of higher genus, which states that two Lefschetz fibrations of the same genus over the $2$–sphere which have the same signature, the same numbers of singular fibers of each type, and admit sections of the same self-intersection number become isomorphic after fiber summing the same number of copies of a ‘universal’ Lefschetz fibration. Kamada [@Kamada1992; @Kamada1996] introduced charts, which are labeled finite graphs in a disk, to describe monodromies of surface braids (see also a textbook [@Kamada2002] of Kamada). Kamada, Matsumoto, Matumoto, and Waki [@KMMW2005] considered a variant of chart for Lefschetz fibrations of genus one to give a remarkably simple proof of the above result of Matsumoto. Furthermore Kamada [@Kamada2012], and Endo and Kamada [@EK2013; @EK2014] made use of generalized charts to give a simple proof of the above theorem of Auroux for Lefschetz fibrations of genus two, and to investigate a stabilization theorem and an invariant for hyperelliptic Lefschetz fibrations of arbitrary genus. See also Baykur and Kamada [@BK2010], and Hayano [@Hayano2011] for applications of charts to broken Lefschetz fibrations. In this paper we introduce a chart description for Lefschetz fibrations of genus greater than two over closed oriented surfaces of arbitrary genus to show a signature formula and two theorems on stabilization for such fibrations. In Section 2 we introduce charts and chart moves with respect to Wajnryb’s presentation of mapping class groups to examine monodromies of Lefschetz fibrations. After a short survey of Meyer’s signature cocycle, we generalize a signature formula [@EN2004] of Endo and Nagami for Lefschetz fibrations over the $2$–sphere to that for Lefschetz fibrations over a closed oriented surface of arbitrary genus in Section 3. Section 4 is devoted to proofs of two theorems on stabilization of Lefschetz fibrations under fiber summing with copies of a ‘universal’ Lefschetz fibration. In particular the first of our stabilization theorems is a generalization of the theorem of Auroux [@Auroux2005]. We make several comments on variations of chart description and propose some possible directions for future research in Section 5. Chart description for Lefschetz fibrations ========================================== In this section we review a definition and properties of Lefschetz fibrations and introduce a chart description for Lefschetz fibrations of genus greater than two. Lefschetz fibrations and their monodromies ------------------------------------------ In this subsection we review a precise definition and basic properties of Lefschetz fibrations. More details can be found in Matsumoto [@Matsumoto1996] and Gompf and Stipsicz [@GS1999]. Let $\Sigma_g$ be a connected closed oriented surface of genus $g$. \[LF\] Let $M$ and $B$ be connected closed oriented smooth $4$–manifold and $2$–manifold, respectively. A smooth map $f\co M\rightarrow B$ is called a [*Lefschetz fibration*]{} of genus $g$ if it satisfies the following conditions: \(i) the set $\Delta\subset B$ of critical values of $f$ is finite and $f$ is a smooth fiber bundle over $B-\Delta$ with fiber $\Sigma_g$; \(ii) for each $b\in\Delta$, there exists a unique critical point $p$ in the [*singular fiber*]{} $F_b:=f^{-1}(b)$ such that $f$ is locally written as $f(z_1,z_2)=z_1z_2$ or $\bar{z}_1z_2$ with respect to some local complex coordinates around $p$ and $b$ which are compatible with orientations of $M$ and $B$; \(iii) no fiber contains a $(\pm 1)$–sphere. We call $M$ the [*total space*]{}, $B$ the [*base space*]{}, and $f$ the [*projection*]{}. We call $p$ a critical point of [*positive type*]{} (resp. of [*negative type*]{}) and $F_b$ a singular fiber of [*positive type*]{} (resp. of [*negative type*]{}) if $f$ is locally written as $f(z_1,z_2)=z_1z_2$ (resp. $f(z_1,z_2)=\bar{z}_1z_2$) in (ii). For a regular value $b\in B$ of $f$, $f^{-1}(b)$ is often called a [*general fiber*]{}. A Lefschetz fibration in this paper is called an [*achiral*]{} Lefschetz fibration in many other papers. Let $f\co M\rightarrow B$ and $f'\co M'\rightarrow B$ be Lefschetz fibrations of genus $g$ over the same base space $B$. We say that $f$ is [*isomorphic*]{} to $f'$ if there exist orientation preserving diffeomorphisms $H\co M\rightarrow M'$ and $h\co B\rightarrow B$ which satisfy $f'\circ H=h\circ f$. If we can choose such an $h$ isotopic to the identity relative to a given base point $b_0\in B$, we say that $f$ is [*strictly isomorphic*]{} to $f'$. Let $\mathcal{M}_g$ be the mapping class group of $\Sigma_g$, namely the group of all isotopy classes of orientation preserving diffeomorphisms of $\Sigma_g$. We assume that $\mathcal{M}_g$ acts on the [*right*]{}: the symbol $\varphi\psi$ means that we apply $\varphi$ first and then $\psi$ for $\varphi, \psi \in \mathcal{M}_g$. We denote the mapping class group of $\Sigma_g$ acting on the left by $\mathcal{M}_g^*$. Hence the identity map $\mathcal{M}_g\rightarrow \mathcal{M}_g^*$ is an anti-isomorphism. Let $f\co M\rightarrow B$ be a Lefschetz fibration of genus $g$ as in Definition \[LF\]. Take a base point $b_0\in B$ and an orientation preserving diffeomorphism $\Phi\co\Sigma_g\rightarrow F_0:=f^{-1}(b_0)$. Since $f$ restricted over $B-\Delta$ is a smooth fiber bundle with fiber $\Sigma_g$, we can define a homomorphism $$\rho\co\pi_1(B-\Delta,b_0)\rightarrow \mathcal{M}_g$$ called the [*monodromy representation*]{} of $f$ with respect to $\Phi$. Let $\gamma$ be the loop based at $b_0$ consisting of the boundary circle of a small disk neighborhood of $b\in\Delta$ oriented counterclockwise and a simple path connecting a point on the circle to $b_0$ in $B-\Delta$. It is known that $\rho([\gamma])$ is a Dehn twist along some essential simple closed curve $c$ on $\Sigma_g$, which is called the [*vanishing cycle*]{} of the critical point $p$ on $f^{-1}(b)$. If $p$ is of positive type (resp. of negative type), then the Dehn twist is right-handed (resp. left-handed). A singular fiber is said to be of [*type [I]{}*]{} if the vanishing cycle is non-separating and of [*type ${\rm II}_h$*]{} for $h=1,\ldots , [g/2]$ if the vanishing cycle is separating and it bounds a genus–$h$ subsurface of $\Sigma_g$. A singular fiber is said to be of [*type ${\rm I}^+$*]{} (resp. [*type ${\rm I}^-$*]{} and [*type ${\rm II}_h^+$*]{}, [*type ${\rm II}_h^-$*]{}) if it is of type I and of positive type (resp. of type I and of negative type, of type ${\rm II}_h$ and of positive type, of type ${\rm II}_h$ and of negative type). We denote by $n_0^{+}(f)$, $n_0^{-}(f)$, $n_h^{+}(f)$, and $n_h^{-}(f)$, the numbers of singular fibers of $f$ of type ${\rm I}^+$, ${\rm I}^-$, ${\rm II}_h^+$, and ${\rm II}_h^-$, respectively. A Lefschetz fibration is called [*irreducible*]{} if every singular fiber is of type I. A Lefschetz fibration is called [*chiral*]{} if every singular fiber is of positive type. Suppose that the cardinality of $\Delta$ is equal to $n$. A system ${\cal A}= (A_1, \dots, A_n)$ of arcs on $B$ is called a [*Hurwitz arc system*]{} for $\Delta$ with base point $b_0$ if each $A_i$ is an embedded arc connecting $b_0$ with a point of $\Delta$ in $B$ such that $A_i \cap A_j= \{b_0\}$ for $i \ne j$, and they appear in this order around $b_0$ (see Kamada [@Kamada2002]). When $B$ is a $2$-sphere, the system ${\cal A}$ determines a system of generators of $\pi_1(B- \Delta, b_0)$, say $(a_1, \dots, a_n)$. We call $( \rho(a_1), \dots, \rho(a_n) )$ a [*Hurwitz system*]{} of $f$. Chart description and Wajnryb’s presentation -------------------------------------------- In this subsection we introduce a chart description for Lefschetz fibrations of genus greater than two by employing Wajnryb’s finite presentation [@Wajnryb1983] of mapping class groups. General theories of charts for presentations of groups were developed independently by Kamada [@Kamada2007] and Hasegawa [@Hasegawa2006]. We use the terminology of chart description in Kamada [@Kamada2007]. We first review a finite presentation of the mapping class group of a closed oriented surface due to Wajnryb. For $i=0,1,\ldots ,2g$, let $\zeta_i$ be a right-handed Dehn twist along the simple closed curve $c_i$ on $\Sigma_g$ depicted in Figure \[curves\]. $c_1$ \[r\] at 0 155 $c_2$ \[t\] at 45 139 $c_3$ \[t\] at 70 146 $c_4$ \[t\] at 98 139 $c_0$ \[r\] at 88 182 $c_{2g}$ \[t\] at 242 139 $c_2$ \[t\] at 44 26 $c_{2h}$ \[t\] at 105 26 $c_{2h+2}$ \[t\] at 182 26 $c_{2g}$ \[t\] at 243 26 $s_{h}$ \[r\] at 135 72 ![Simple closed curves on $\Sigma_g$[]{data-label="curves"}](curves "fig:") \[pres\] Suppose that $g$ is greater than two. The mapping class group $\mathcal{M}_g$ is generated by elements $\zeta_0,\zeta_1,\zeta_2,\ldots ,\zeta_{2g}$ and has defining relations: [$$\tag*{$\bullet$ (Far--commutation)} \zeta_i\zeta_j =\zeta_j\zeta_i \;\; (1\leq i<j-1\leq 2g-1), \quad \zeta_0\zeta_j =\zeta_j\zeta_0 \;\; (j=1,2,3,5,\ldots ,2g),$$ $$\tag*{$\bullet$ (Braid relation)} \zeta_i\zeta_{i+1}\zeta_i =\zeta_{i+1}\zeta_i\zeta_{i+1} \;\; (i=1,\ldots ,2g-1), \quad \zeta_0\zeta_4\zeta_0 =\zeta_4\zeta_0\zeta_4;$$ $$\tag*{$\bullet$ ($3$-chain relation)} (\zeta_3\zeta_2\zeta_1)^4 =\zeta_0 \zeta_4^{-1}\zeta_3^{-1}\zeta_2^{-1}\zeta_1^{-2}\zeta_2^{-1}\zeta_3^{-1}\zeta_4^{-1} \zeta_0\zeta_4\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\zeta_4;$$ $$\tag*{$\bullet$ (Lantern relation)} \delta_3\zeta_1\zeta_3\zeta_5 = \zeta_0\tau_2\zeta_0\tau_2^{-1} \tau_1\tau_2\zeta_0\tau_2^{-1}\tau_1^{-1}, \quad {\sl where}$$ $$\begin{aligned} & \tau_1:=\zeta_2\zeta_3\zeta_1\zeta_2, \quad \tau_2:=\zeta_4\zeta_5\zeta_3\zeta_4, \quad \mu:=\zeta_5\zeta_6\tau_2\zeta_0\tau_2^{-1}\zeta_6^{-1}\zeta_5^{-1}, \\ & \nu:=\zeta_1\zeta_2\zeta_3\zeta_4\zeta_0\zeta_4^{-1}\zeta_3^{-1}\zeta_2^{-1}\zeta_1^{-1}, \quad \delta_3:=\zeta_6^{-1}\zeta_5^{-1}\zeta_4^{-1}\zeta_3^{-1}\zeta_2^{-1} \mu^{-1}\nu\mu\zeta_2\zeta_3\zeta_4\zeta_5\zeta_6;\end{aligned}$$ $$\tag*{$\bullet$ (Hyperelliptic relation)} \zeta_{2g}\cdots\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\cdots\zeta_{2g} \delta_g=\delta_g \zeta_{2g}\cdots\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\cdots\zeta_{2g}, \quad {\sl where}$$ $$\begin{aligned} & \tau_1:=\zeta_2\zeta_3\zeta_1\zeta_2, \quad \tau_i:=\zeta_{2i}\zeta_{2i-1}\zeta_{2i+1}\zeta_{2i}, \\ & \nu_1:=\zeta_4^{-1}\zeta_3^{-1}\zeta_2^{-1}\zeta_1^{-2}\zeta_2^{-1}\zeta_3^{-1}\zeta_4^{-1} \zeta_0\zeta_4\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\zeta_4, \quad \nu_i:=\tau_{i-1}\tau_i\nu_{i-1}\tau_i^{-1}\tau_{i-1}^{-1}, \\ & \mu_1:=\zeta_2\zeta_3\zeta_4\nu_1\zeta_1^{-1}\zeta_2^{-1}\zeta_3^{-1}\zeta_4^{-1}, \quad \mu_i:=\zeta_{2i}\zeta_{2i+1}\zeta_{2i+2}\nu_i\zeta_{2i-1}^{-1}\zeta_{2i}^{-1} \zeta_{2i+1}^{-1}\zeta_{2i+2}^{-1}, \\ & \delta_g :=\mu_{g-1}^{-1}\cdots \mu_2^{-1}\mu_1^{-1}\zeta_1\mu_1\mu_2\cdots\mu_{g-1}\end{aligned}$$]{} for $i=2,\ldots ,g-1$. We make use of the presentation above to introduce a notion of chart which gives a graphic description of monodromy representations of Lefschetz fibrations. We set [$$\begin{aligned} \mathcal{X} & :=\{\zeta_0,\zeta_1,\ldots ,\zeta_{2g}\}, \\ \mathcal{R} & :=\{r_F(i,j)\, |\, 1\leq i<j-1\leq 2g-1\}\cup \{r_F(0,j)\, |\, j=1,2,3,5,\ldots ,2g\} \\ & \quad \cup \{r_B(i)\, |\, i=0,1,\ldots ,2g-1\}\cup \{r_C,r_L,r_H \}, \\ \mathcal{S} & :=\{\ell_0(i)^{\pm 1}\, |\, i=0,1,\ldots ,2g\} \cup \{\ell_h^{\pm 1}\, |\, h=1,\ldots ,[g/2]\}, \end{aligned}$$]{} for $g\geq 3$, where [$$\begin{aligned} & r_F(i,j) :=\zeta_i\zeta_j\zeta_i^{-1}\zeta_j^{-1}, \quad r_B(0):=\zeta_0\zeta_4\zeta_0\zeta_4^{-1}\zeta_0^{-1}\zeta_4^{-1}, \\ & r_B(i) :=\zeta_i\zeta_{i+1}\zeta_i\zeta_{i+1}^{-1}\zeta_i^{-1}\zeta_{i+1}^{-1} \;\; (i=1,\ldots ,2g-1), \\ & r_C:=(\zeta_3\zeta_2\zeta_1)^4 \zeta_4^{-1}\zeta_3^{-1}\zeta_2^{-1}\zeta_1^{-2}\zeta_2^{-1}\zeta_3^{-1}\zeta_4^{-1} \zeta_0^{-1}\zeta_4\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\zeta_4\zeta_0^{-1}, \\ & r_L:=\delta_3\zeta_1\zeta_3\zeta_5\tau_1\tau_2\zeta_0^{-1}\tau_2^{-1}\tau_1^{-1} \tau_2\zeta_0^{-1}\tau_2^{-1}\zeta_0^{-1}, \\ & r_H:=\zeta_{2g}\cdots\zeta_3\zeta_2\zeta_1^2\zeta_2\zeta_3\cdots\zeta_{2g} \delta_g\zeta_{2g}^{-1}\cdots\zeta_3^{-1}\zeta_2^{-1}\zeta_1^{-2} \zeta_2^{-1}\zeta_3^{-1}\cdots\zeta_{2g}^{-1}\delta_g^{-1}, \\ & \ell_0(i):=\zeta_i \;\; (i=0,1,\ldots ,2g), \quad \ell_h:=(\zeta_1\zeta_2\cdots\zeta_{2h})^{4h+2} \;\; (h=1,\ldots ,[g/2]), \end{aligned}$$]{} and $\delta_3, \tau_1, \tau_2, \delta_g$ are defined as in Theorem \[pres\]. Let $B$ be a connected closed oriented surface and $\Gamma$ a finite graph in $B$ such that each edge of $\Gamma$ is oriented and labeled with an element of $\mathcal{X}$. We denote the label $\zeta_i$ by $i$ for short. Choose a simple path $\gamma$ which intersects with edges of $\Gamma$ transversely and does not intersect with vertices of $\Gamma$. For such a path $\gamma$, we obtain a word $w_{\Gamma}(\gamma)$ in $\mathcal{X}\cup\mathcal{X}^{-1}$ by reading off the labels of intersecting edges along $\gamma$ with exponents as in Figure \[intersection\] (a). We call the word $w_{\Gamma}(\gamma)$ the [*intersection word*]{} of $\gamma$ with respect to $\Gamma$. Conversely, we can specify the number, orientations, and labels of consecutive edges in $\Gamma$ by indicating a (dashed) arrow intersecting the edges transversely together with the intersection word of the arrow with respect to $\Gamma$ (see Figure \[intersection\] (b) and (c)). $\gamma$ \[r\] at 0 18 $1$ \[t\] at 30 0 $2$ \[t\] at 59 0 $1$ \[t\] at 87 0 $3$ \[t\] at 116 0 $2$ \[t\] at 143 0 \[t\] at 85 -15 $w$ \[l\] at 255 18 \[t\] at 235 -15 $w$ \[b\] at 313 33 \[t\] at 313 -15 ![Intersection word $w_{\Gamma}(\gamma)=w=\zeta_1\zeta_2^{-1}\zeta_1^{-1}\zeta_3\zeta_2$[]{data-label="intersection"}](intersection "fig:") For a vertex $v$ of $\Gamma$, a small simple closed curve surrounding $v$ in the counterclockwise direction is called a [*meridian loop*]{} of $v$ and denoted by $m_v$. The vertex $v$ is said to be [*marked*]{} if one of the regions around $v$ is specified by an asterisk. If $v$ is marked, the intersection word $w_{\Gamma}(m_v)$ of $m_v$ with respect to $\Gamma$ is well-defined. If not, it is determined up to cyclic permutation. See Kamada [@Kamada2007] for details. \[def:chart\] A [*chart*]{} in $B$ is a finite graph $\Gamma$ in $B$ (possibly being empty or having [*hoops*]{} that are closed edges without vertices) whose edges are labeled with an element of $\mathcal{X}$, and oriented so that the following conditions are satisfied (see Figure \[verticesA\], Figure \[verticesB\], and Figure \[verticesC\]): - the vertices of $\Gamma$ are classified into two families: [*white vertices*]{} and [*black vertices*]{}; - if $v$ is a white vertex (resp. a black vertex), the word $w_{\Gamma}(m_v)$ is a cyclic permutation of an element of $\mathcal{R}\cup\mathcal{R}^{-1}$ (resp. of $\mathcal{S}$). A white vertex $v$ is said to be of [*type*]{} $r$ (resp. of [*type*]{} $r^{-1}$) if $w_{\Gamma}(m_v)^{-1}$ is a cyclic permutation of $r\in\mathcal{R}$ (resp. of $r^{-1}\in\mathcal{R}^{-1}$). A black vertex $v$ is said to be of [*type*]{} $s$ if $w_{\Gamma}(m_v)$ is a cyclic permutation of $s\in\mathcal{S}$. A chart $\Gamma$ is said to be [*marked*]{} if each white vertex (resp. black vertex) $v$ is marked and $w_{\Gamma}(m_v)$ is exactly an element of $\mathcal{R}\cup\mathcal{R}^{-1}$ (resp. of $\mathcal{S}$). If a base point $b_0$ of $B$ is specified, we always assume that a chart $\Gamma$ is disjoint from $b_0$. A chart consisting of two black vertices and one edge connecting them is called a [*free edge*]{}. $i$ \[tr\] at 1 21 $j$ \[br\] at 1 52 $i$ \[bl\] at 31 52 $j$ \[tl\] at 31 21 $i$ \[tr\] at 95 18 $i\negthinspace +\negthinspace 1$ \[r\] at 88 35 $i$ \[br\] at 96 54 $i\negthinspace +\negthinspace 1$ \[bl\] at 133 54 $i$ \[l\] at 139 35 $i\negthinspace +\negthinspace 1$ \[tl\] at 132 18 $3$ \[b\] at 222 73 $2$ \[b\] at 231 73 $1$ \[b\] at 239 73 $3$ \[b\] at 248 73 $2$ \[b\] at 256 73 $1$ \[b\] at 265 73 $3$ \[b\] at 273 73 $2$ \[b\] at 282 73 $1$ \[b\] at 291 73 $3$ \[b\] at 300 73 $2$ \[b\] at 309 73 $1$ \[b\] at 317 73 $4$ \[t\] at 346 0 $3$ \[t\] at 337 0 $2$ \[t\] at 330 0 $1$ \[t\] at 321 0 $1$ \[t\] at 312 0 $2$ \[t\] at 303 0 $3$ \[t\] at 295 0 $4$ \[t\] at 287 0 $0$ \[t\] at 278 0 $4$ \[t\] at 270 0 $3$ \[t\] at 261 0 $2$ \[t\] at 252 0 $1$ \[t\] at 243 0 $1$ \[t\] at 235 0 $2$ \[t\] at 227 0 $3$ \[t\] at 219 0 $4$ \[t\] at 210 0 $0$ \[t\] at 201 0 ![Vertices of type $r_F(i,j)$, $r_B(i)\; (i\ne 0)$, $r_C$[]{data-label="verticesA"}](verticesA "fig:") $\delta_3$ \[b\] at 29 90 $1$ \[b\] at 47 84 $3$ \[b\] at 56 84 $5$ \[b\] at 64 84 $\tau_1$ \[b\] at 81 90 $\tau_2$ \[b\] at 107 90 $0$ \[t\] at 128 9 $\tau_2^{-\negthinspace 1}$ \[t\] at 111 4 $\tau_1^{-\negthinspace 1}$ \[t\] at 85 4 $\tau_2$ \[t\] at 62 4 $0$ \[t\] at 42 9 $\tau_2^{-\negthinspace 1}$ \[t\] at 26 4 $0$ \[t\] at 8 9 $2g$ \[b\] at 218 94 $1$ \[b\] at 244 94 $1$ \[b\] at 253 94 $2g$ \[b\] at 278 94 $\delta_g$ \[l\] at 321 47 $2g$ \[t\] at 278 0 $1$ \[t\] at 253 0 $1$ \[t\] at 244 0 $2g$ \[t\] at 218 0 $\delta_g^{-1}$ \[r\] at 178 47 ![Vertices of type $r_L$ and $r_H$[]{data-label="verticesB"}](verticesB "fig:") $i$ \[l\] at 46 34 $i$ \[l\] at 46 8 $1$ \[b\] at 223 45 $2h$ \[b\] at 202 45 $1$ \[b\] at 193 45 $2h$ \[b\] at 174 45 $1$ \[b\] at 139 45 $2h$ \[b\] at 118 45 ![Vertices of type $\ell_0(i)^{\pm 1}$ and $\ell_h$[]{data-label="verticesC"}](verticesC "fig:") It would be worth noting that the intersection word of a ‘clockwise’ meridian of a white vertex of type $r$ is equal to $r$, while that of a ‘counterclockwise’ meridian of a black vertex of type $s$ is equal to $s$ in this paper. This notation is different from those of Kamada [@Kamada2007] and Hasegawa [@Hasegawa2006], who always consider ‘counterclockwise’ meridians for both white and black vertices. We next introduce several moves for charts. Let $\Gamma$ and $\Gamma'$ be two charts on $B$ and $b_0$ a base point of $B$. Let $D$ be a disk embedded in $B-\{b_0\}$. Suppose that the boundary $\partial D$ of $D$ intersects $\Gamma$ and $\Gamma'$ transversely. We say that $\Gamma'$ is obtained from $\Gamma$ by a [*chart move of type W*]{} if $\Gamma\cap (B-{\rm Int}\, D)=\Gamma'\cap (B-{\rm Int}\, D)$ and that both $\Gamma\cap D$ and $\Gamma'\cap D$ have no black vertices. We call chart moves of type W shown in Figure \[movesA\] (a), (b), and (c), a [*channel change*]{}, a [*birth/death of a hoop*]{}, and a [*birth/death of a pair of white vertices*]{}, respectively. $i$ \[tr\] at 0 66 $i$ \[bl\] at 41 106 $i$ \[tr\] at 80 66 $i$ \[bl\] at 121 106 (a) \[t\] at 61 84 $i$ \[tl\] at 184 76 empty at 258 86 (b) \[t\] at 209 84 $r$ at 57 20 $r^{\negthinspace -\negthinspace 1}$ at 87 23 (c) \[t\] at 132 19 ![Chart moves of type W[]{data-label="movesA"}](movesA "fig:") Let $s$ and $s'$ be elements of $\mathcal{S}$. Suppose that there exists a word $w$ in $\mathcal{X}\cup\mathcal{X}^{-1}$ such that two words $s'$ and $wsw^{-1}$ determine the same element of $\mathcal{M}_g$. If a chart $\Gamma$ contains a black vertex of type $s$, then we can change a part of $\Gamma$ near the vertex by using a local replacement depicted in Figure \[movesB\] to obtain another chart $\Gamma'$. We say that $\Gamma'$ is obtained from $\Gamma$ by a [*chart move of transition*]{}. Note that the blank labeled with ${\rm T}$ can be filled only with edges and white vertices. $s$ \[b\] at 35 81 $s'$ \[b\] at 205 81 $w$ \[bl\] at 245 86 $w$ \[tl\] at 245 40 $s$ \[bl\] at 257 81 T at 232 64 ![Chart move of transition[]{data-label="movesB"}](movesB "fig:") We say that $\Gamma'$ is obtained from $\Gamma$ by a [*chart move of conjugacy type*]{} if $\Gamma'$ is obtained from $\Gamma$ by a local replacement depicted in Figure \[movesC\]. $b_0$ \[l\] at 7 22 $b_0$ \[l\] at 117 22 $i$ \[l\] at 135 22 $b_0$ \[l\] at 197 22 $b_0$ \[l\] at 304 22 $i$ \[l\] at 323 22 ![Chart moves of conjugacy type[]{data-label="movesC"}](movesC "fig:") Let $\Gamma$ be a chart in $B$ with base point $b_0$ and $\Delta_{\Gamma}$ the set of black vertices of $\Gamma$. For a loop $\gamma$ in $B-\Delta_{\Gamma}$ based at $b_0$, the element of $\mathcal{M}_g$ determined by the intersection word $w_{\Gamma}(\gamma)$ of $\gamma$ with respect to $\Gamma$ does not depend on a choice of representative of the homotopy class of $\gamma$. Thus we obtain a homomorphism $\rho_{\Gamma}\co\pi_1(B-\Delta_{\Gamma},b_0)\rightarrow \mathcal{M}_g$, which is called the [*homomorphism determined by $\Gamma$*]{}. We now state a classification of Lefschetz fibrations in terms of charts and chart moves. Let $B$ be a connected closed oriented surface. \[correspondence\] Suppose that $g$ is greater than two. [(1)]{} Let $f$ be a Lefschetz fibration of genus $g$ over $B$ and $\rho$ a monodromy representation of $f$. Then there exists a chart $\Gamma$ in $B$ such that the homomorphism $\rho_{\Gamma}$ determined by $\Gamma$ is equal to $\rho$. [(2)]{} For every chart $\Gamma$ in $B$, there exists a Lefschetz fibration $f$ of genus $g$ over $B$ such that a monodromy representation of $f$ is equal to the homomorphism $\rho_{\Gamma}$ determined by $\Gamma$. We call such $\Gamma$ as in Proposition \[correspondence\] (1) a chart [*corresponding to $f$*]{}, and such $f$ as in Proposition \[correspondence\] (2) a Lefschetz fibration [*described by $\Gamma$*]{}. Instead of giving a proof of Proposition \[correspondence\], we show an example of a chart and describe the correspondence of the chart to a Hurwitz system of a Lefschetz fibration. Let $B$ be a $2$–sphere. We consider a chart $\Gamma$ in $B$ with base point $b_0$ and a system $(\gamma_1,\gamma_2,\gamma_3,\gamma_4)$ of loops based at $b_0$, which is determined by a Hurwitz arc system $\mathcal{A}$ for the set $\Delta_{\Gamma}$ of black vertices of $\Gamma$, as in Figure \[exampleA\]. The intersection words of the loops with respect to $\Gamma$ are $$\begin{aligned} w_{\Gamma}(\gamma_1) & =\zeta_1^{-1}\zeta_2^{-1}\zeta_1\zeta_2\zeta_1, \quad w_{\Gamma}(\gamma_2) =\zeta_1^{-1}\zeta_3\zeta_1, \\ w_{\Gamma}(\gamma_3) & =\zeta_2^{-1}\zeta_3^{-1}\zeta_2^{-1}\zeta_3\zeta_2, \quad w_{\Gamma}(\gamma_4) =\zeta_2^{-1}, \end{aligned}$$ each of which represents the image $\rho_{\Gamma}(a_i)$ of the homotopy class $a_i$ of $\gamma_i$ under the homomorphism $\rho_{\Gamma}:\pi_1(B-\Delta_{\Gamma},b_0)\rightarrow \mathcal{M}_g$. Since the group $\pi_1(B-\Delta_{\Gamma},b_0)$ has a presentation $\langle a_1,a_2,a_3,a_4\, |\, a_1a_2a_3a_4=1 \rangle$, $\rho_{\Gamma}$ is determined by the system $(\rho_{\Gamma}(a_1),\rho_{\Gamma}(a_2),\rho_{\Gamma}(a_3), \rho_{\Gamma}(a_4))$, which is a Hurwitz system of a certain Lefschetz fibration of genus $g$ over $B$ because each $\rho_{\Gamma}(a_i)$ is a Dehn twist. Note that the product $w_{\Gamma}(\gamma_1)w_{\Gamma}(\gamma_2) w_{\Gamma}(\gamma_3)w_{\Gamma}(\gamma_4)$ of the intersection words represents the identity of $\mathcal{M}_g$. $b_0$ \[t\] at 175 184 $1$ \[b\] at 30 156 $2$ \[b\] at 46 95 $1$ \[b\] at 112 77 $2$ \[b\] at 186 77 $1$ \[r\] at 154 103 $3$ \[t\] at 137 129 $3$ \[b\] at 191 115 $3$ \[r\] at 208 25 $2$ \[r\] at 243 50 $2$ \[t\] at 283 114 \[r\] at 70 135 \[t\] at 152 166 \[t\] at 233 142 \[b\] at 278 162 ![Monodromy of a chart $\Gamma$[]{data-label="exampleA"}](exampleA "fig:") \[classification\] Suppose that $g$ is greater than two. Let $f$ and $f'$ be Lefschetz fibrations of genus $g$ over $B$, and $\Gamma$ and $\Gamma'$ charts corresponding to $f$ and $f'$, respectively. Then $f$ is strictly isomorphic to $f'$ if and only if $\Gamma$ is transformed to $\Gamma'$ by a finite sequence of chart moves of type W, chart moves of transitions, chart moves of conjugacy type, and ambient isotopies of $B$ relative to $b_0$. Proposition \[correspondence\] and Theorem \[classification\] follow from a classification theorem of Lefschetz fibrations due to Kas [@Kas1980] and Matsumoto [@Matsumoto1996] together with fundamental theorems on charts and chart moves by Kamada [@Kamada2007 Sections 4–8]. We end this subsection with a definition and chart description of fiber sums of Lefschetz fibrations. Let $f\co M\rightarrow B$ and $f'\co M'\rightarrow B'$ be Lefschetz fibrations of genus $g$. Take regular values $b_0\in B$ and $b_0'\in B'$ of $f$ and $f'$, and small disks $D_0\subset B-\Delta$ and $D'_0\subset B-\Delta'$ near $b_0$ and $b'_0$, respectively. Consider general fibers $F_0:=f^{-1}(b_0)$ and $F'_0:=f'^{-1}(b_0')$ and orientation preserving diffeomorphisms $\Phi\co \Sigma_g\rightarrow F_0$ and $\Phi'\co\Sigma_g\rightarrow F'_0$, respectively. Let $\Psi\co \Sigma_g\rightarrow \Sigma_g$ be an orientation preserving diffeomorphism and $r\co \partial D_0\rightarrow \partial D'_0$ an orientation reversing diffeomorphism. The new manifold $M\#_F M'$ obtained by glueing $M-f^{-1}({\rm Int}\, D_0)$ and $M'-f'^{-1}({\rm Int}\, D'_0)$ by $(\Phi'\circ\Psi\circ\Phi^{-1})\times r$ admits a Lefschetz fibration $f\#_{\Psi}\, f'\co M\#_F M'\rightarrow B\# B'$ of genus $g$. We call $f\#_{\Psi}\, f'$ the [*fiber sum*]{} of $f$ and $f'$ with respect to $\Psi$. Although the diffeomorphim type of $M\#_F M'$ and the isomorphism type of $f\#_{\Psi}\, f'$ depend on a choice of the diffeomorphism $\Psi$ in general, we often abbreviate $f\#_{\Psi}\, f'$ as $f\#\, f'$. Let $\Gamma$ and $\Gamma'$ be charts corresponding to $f$ and $f'$, and $D_0$ and $D'_0$ small disks near $b_0$ and $b_0'$ disjoint from $\Gamma$ and $\Gamma'$, respectively. Connecting $B-{\rm Int}\, D_0$ with $B'-{\rm Int}\, D'_0$ by a tube, we have a connected sum $B\# B'$ of $B$ and $B'$. Let $w$ be a word in $\mathcal{X}\cup\mathcal{X}^{-1}$ which represents the mapping class of $\Psi$ in $\mathcal{M}_g$. Let $\Gamma\#_w\Gamma'$ be the union of $\Gamma$, $\Gamma'$, and hoops on the tube representing $w$ (see Figure \[fibersum\]). Then the fiber sum $f\#_{\Psi}\, f'$ is described by this new chart $\Gamma\#_w\Gamma'$ in $B\# B'$ with base point $b_0$. If the word $w$ is trivial, then the chart $\Gamma\#_w\Gamma'$ is denoted also by $\Gamma\oplus\Gamma'$, which is called a [*product*]{} of $\Gamma$ and $\Gamma'$. $b_0$ \[r\] at 128 38 $b'_0$ \[l\] at 210 38 $B$ \[br\] at 3 57 $B'$ \[bl\] at 330 59 $w$ \[t\] at 170 13 $\Gamma$ at 71 40 $\Gamma'$ at 270 40 ![Chart $\Gamma\#_w\Gamma'$ in $B\# B'$[]{data-label="fibersum"}](fibersum "fig:") Signature of Lefschetz fibrations ================================= In this section we review the signature cocycle discovered by Meyer and prove a signature theorem for Lefschetz fibrations. Meyer’s signature cocycle ------------------------- In this subsection we give a brief survey on Meyer’s signature cocycle. We begin with the definition of the signature cocycle. Let $g$ be a positive integer. \[cocycle\] For $A,B\in {\rm Sp}(2g,\Bbb{Z})$, we consider the vector space $$V_{A,B}:=\{ (x,y)\in \Bbb{R}^{2g}\times\Bbb{R}^{2g}\, |\, (A^{-1}-I_{2g})x+(B-I_{2g})y=0\}$$ and the bilinear form $\langle \; ,\; \rangle_{A,B}\co V_{A,B}\times V_{A,B}\rightarrow\Bbb{R}$ defined by $$\langle (x_1,y_1),(x_2,y_2)\rangle_{A,B}:=(x_1+y_1)\cdot J(I_{2g}-B)y_2,$$ where $\cdot$ is the standard inner product of $\Bbb{R}^{2g}$ and $J=\left(\begin{smallmatrix} 0 & I_g \\ -I_g & 0 \end{smallmatrix}\right)$. Since $\langle \; ,\; \rangle_{A,B}$ is symmetric, we can define an integer $\tau_g(A,B)$ to be the signature of $(V_{A,B}, \langle \; ,\; \rangle_{A,B})$. The map $\tau_g\co {\rm Sp}(2g,\Bbb{Z})\times {\rm Sp}(2g,\Bbb{Z})\rightarrow \Bbb{Z}$ is called the [*signature cocycle*]{}. Let $P$ be a compact connected oriented surface of genus $0$ with three boundary components and $\pi\co E\rightarrow P$ a fiber bundle over $P$ with fiber $\Sigma_g$ and structure group ${\rm Diff}_+\Sigma_g$. The fundamental group $\pi_1(P,*)$ of $P$ with base point $*$ is a free group generated by two loops $a$ and $b$ depicted in Figure \[pants\]. If we take an orientation preserving diffeomorphism $\Sigma_g\rightarrow \pi^{-1}(*)$, we obtain the monodromy representation $\pi_1(P,*)\rightarrow \mathcal{M}_g$ which sends $a$ to $\alpha$ and $b$ to $\beta$. Since $\mathcal{M}_g^*$ acts on $H:=H_1(\Sigma_g;\Bbb{Z})$ preserving the intersection form, we have a representation $\mathcal{M}_g^*\rightarrow {\rm Sp}(2g,\Bbb{Z})$ by fixing a symplectic basis on $H$. Let $A$ and $B$ denote matrices corresponding to $\alpha$ and $\beta$, respectively. $P_1$ \[br\] at 1 56 $P_2$ \[bl\] at 117 53 $*$ \[b\] at 58 59 $a$ \[tl\] at 41 16 $b$ \[tr\] at 75 16 ![Pair of pants $P$[]{data-label="pants"}](pants "fig:") Meyer closely studied the signature of the total space $E$ to obtain the following theorem. \[meyer\] The signature $\sigma(E)$ of $E$ is equal to $-\tau_g(A,B)$. Theorem \[meyer\] and Novikov’s additivity implies that $\tau_g$ is a $2$–cocycle of ${\rm Sp}(2g,\Bbb{Z})$. We recall a Maslov index for a triple of Lagrangian subspaces and Wall’s non-additivity theorem, which are used in the proof of Theorem \[meyer\]. Let $V$ be a real vector space of dimension $2n$, $\omega\in \Lambda^2V^*$ a symplectic form on $V$, and $\Lambda(V,\omega)$ the Lagrangian Grassmannian of $(V,\omega)$, which is the set of Lagrangian subspaces of $(V,\omega)$. For $L_1,L_2,L_3\in \Lambda(V,\omega)$, the bilinear form $$\begin{aligned} \Psi & \co (L_3+L_1)\cap L_2\times (L_3+L_1)\cap L_2\rightarrow \Bbb{R}\co (v,w)\mapsto \omega(v,w_3) \\ & (v,w\in (L_3+L_1)\cap L_2, \, w=w_1+w_3 \; (w_1\in L_1,w_3\in L_3)) \end{aligned}$$ is symmetric. We define an integer $i(L_1,L_2,L_3)$ to be the signature of $((L_3+L_1)\cap L_2, \Psi)$, which is called the [*ternary Maslov index*]{} of the triple $(L_1,L_2,L_3)$. Let $M_1, M_2$ be compact oriented smooth $4$–manifolds, $X_1,X_2,X_3$ compact oriented smooth $3$–manifolds, and $\Sigma$ a closed oriented smooth $2$–manifold. We assume that $M=M_1\cup M_2, \, \partial M_1=X_1\cup X_2,\, \partial M_2=X_2\cup X_3,\, \partial X_1=\partial X_2=\partial X_3=\Sigma$, and the orientations of these manifolds satisfy $$\begin{gathered} [M]=[M_1]+[M_2], \; \partial_*[M_1]=[X_2]-[X_1], \; \partial_*[M_2]=[X_3]-[X_2], \\ \partial_*[X_1] =\partial_*[X_2]=\partial_*[X_3]=[\Sigma].\end{gathered}$$ Let $\omega\co V\times V\rightarrow \Bbb{R}$ be the intersection form on $V:=H_1(\Sigma;\Bbb{R})$ and $L_i$ the kernel of the homomorphism $V\rightarrow H_1(X_i;\Bbb{R})$ induced by the inclusion $\Sigma\rightarrow X_i$ for $i=1,2,3$. Since $L_i\in\Lambda(V,\omega)$ for $i=1,2,3$, we can define the ternary Maslov index $i(L_1,L_2,L_3)$ of the triple $(L_1,L_2,L_3)$. \[wall\] $\sigma(M)=\sigma(M_1)+\sigma(M_2)-i(L_1,L_2,L_3)$. Gambaudo and Ghys [@GG2005] (and independently the first author) made use of Theorem \[wall\] to give the following proof of Theorem \[meyer\]. See also Gilmer and Masbaum [@GM2011]. Consider $P$ to be a boundary sum of two annuli $P_1$ and $P_2$ (see Figure \[pants\]). We set $M:=E$, $M_i:=\pi^{-1}(P_i)\, (i=1,2)$, $X_2:=M_1\cap M_2$, $X_1:=\partial M_1-{\rm Int}\, X_2$, $X_3:=\partial M_3-{\rm Int}\, X_2$, and $\Sigma:=\partial X_2$. Applying Theorem \[wall\] to these manifolds, we have $$\sigma(E)=\sigma(M_1)+\sigma(M_2)-i(L_1,L_2,L_3) =-i(L_1,L_2,L_3)$$ because each of $M_1$ and $M_2$ is a product of a mapping torus with an interval, which has signature zero. Since the bordered component of $X_i$ is diffeomorphic to $I\times\Sigma_g$ for $i=1,2,3$, we put $V:=H\oplus H$, $\omega:=\mu\oplus (-\mu)$, and obtain $$\begin{aligned} L_1 & =\{ (-\xi, \alpha_*^{-1}(\xi))\in V\, |\, \xi\in H\}, \quad L_2=\{ (-\xi,\xi)\in V\, |\, \xi\in H\}, \\ L_3 & =\{ (-\xi, \beta_*(\xi))\in V\, |\, \xi\in H\}, \end{aligned}$$ where $H$ is the first homology $H_1(\Sigma_g;\Bbb{R})$ of $\Sigma_g$ and $\mu\co H\times H\rightarrow\Bbb{R}$ is the intersection form of $\Sigma_g$. It is easily seen that the subspace $(L_1+L_3)\cap L_2$ is written as $$(L_1+L_3)\cap L_2 =\{ (-\xi-\eta, \alpha_*^{-1}(\xi)+\beta_*(\eta))\in V \, |\, \xi+\eta=\alpha_*^{-1}(\xi)+\beta_*(\eta)\, (\xi,\eta\in H) \}$$ and the symmetric bilinear form $\Psi$ on $(L_1+L_3)\cap L_2$ is written as $$\Psi((-\xi-\eta,\alpha_*^{-1}(\xi)+\beta_*(\eta)),(-\xi'-\eta',\alpha_*^{-1}(\xi')+\beta_*(\eta'))) = \mu(\xi+\eta,({\rm id}-\beta_*)(\eta')).$$ We consider the vector space $$U_{\alpha,\beta}:=\{ (\xi,\eta)\in V\, |\, (\alpha_*^{-1}-{\rm id})(\xi)+(\beta_*-{\rm id})(\eta)=0 \}$$ and the symmetric bilinear form $\langle \; ,\;\rangle_{\alpha,\beta}$ on $U_{\alpha,\beta}$ defined by $$\langle (\xi,\eta),(\xi',\eta')\rangle_{\alpha,\beta} :=\mu(\xi+\eta,({\rm id}-\beta_*)(\eta')) \quad ((\xi,\eta),(\xi',\eta')\in U_{\alpha,\beta}).$$ Since the linear map $U_{\alpha,\beta}\rightarrow (L_1+L_3)\cap L_2 \co (\xi,\eta)\mapsto (-\xi-\eta,\xi+\eta)$ is compatible with the bilinear forms, the signature of $((L_1+L_3)\cap L_2,\Psi)$ is equal to that of $(U_{\alpha,\beta},\langle \; ,\; \rangle_{\alpha,\beta})$, which is isomorphic to $(V_{A,B},\langle \; ,\; \rangle_{A,B})$ under a choice of a symplectic basis of $H$. Therefore we conclude that $i(L_1,L_2,L_3)=\tau_g(A,B)$. It is known that $\tau_g$ is a normalized, symmetric $2$–cocycle of ${\rm Sp}(2g,\Bbb{Z})$ and invariant under conjugation. The cohomology class $[\tau_g]\in H^2({\rm Sp}(2g,\Bbb{Z});\Bbb{Z})$ corresponds to $-4c_1$ under homomorphisms: $$H^2({\rm Sp}(2g,\Bbb{Z});\Bbb{Z})\leftarrow H^2(B{\rm Sp}(2g,\Bbb{R});\Bbb{Z})\cong H^2(BU(g);\Bbb{Z})\cong \Bbb{Z}.$$ For more details see Meyer [@Meyer1973], Turaev [@Turaev1987], Barge and Ghys [@BG1992], and Kuno [@Kuno2012]. A signature formula ------------------- In this subsection we describe the signature of a Lefschetz fibration of genus greater than two in terms of charts. Let $g$ be an integer greater than two. Let $B$ be a connected closed oriented surface and $\Gamma$ a chart in $B$. We denote the number of white vertices of type $r_F(i,j)$ (resp. $r_B(i),\, r_C,\, r_L,\, r_H$) minus the number of white vertices of type $r_F(i,j)^{-1}$ (resp. $r_B(i)^{-1},\, r_C^{-1},\, r_L^{-1},\, r_H^{-1}$) included in $\Gamma$ by $n_F(i,j)(\Gamma)$ (resp. $n_B(i)(\Gamma),\, n_C(\Gamma),\, n_L(\Gamma),\, n_H(\Gamma)$). Similarly, we denote the number of black vertices of type $\ell_0(i)^{\pm 1}$ (resp. $\ell_h^{\pm 1}$) included in $\Gamma$ by $n_0^{\pm}(i)(\Gamma)$ (resp. $n_h^{\pm}(\Gamma)$), and set $n_0(i)(\Gamma):=n_0^+(i)(\Gamma)-n_0^-(i)(\Gamma)$ (resp. $n_h(\Gamma):=n_h^+(\Gamma)-n_h^-(\Gamma)$) and $n_0^{\pm}(\Gamma):=\sum_{i=0}^{2g}n_0^{\pm}(i)(\Gamma)$. \[sigma\] The number $$\sigma(\Gamma):= -6\, n_C(\Gamma)-n_L(\Gamma)+\sum_{h=1}^{[g/2]}(4h(h+1)-1)\, n_h(\Gamma)$$ is called the [*signature*]{} of $\Gamma$. Let $f\co M\rightarrow B$ be a Lefschetz fibration of genus $g$ and $\Gamma$ a chart in $B$ corresponding to $f$. The purpose of this subsection is to show the following theorem. \[signature\] The signature $\sigma(M)$ of $M$ is equal to $\sigma(\Gamma)$. It immediately follows from Theorem \[signature\] that $\sigma(\Gamma)$ is invariant under chart moves of type W and chart moves of transition. Although any combinatorial proof of this fact does not seem to be known, Hasegawa [@Hasegawa2006] proved that $\sigma(\Gamma)$ is invariant under chart moves of transitions by a purely combinatorial method on the assumption that it is invariant under chart moves of type W. Let $\tilde{\mathcal{X}}$ be the set of right-handed Dehn twists along simple closed curves in $\Sigma_g$ and $\tilde{\mathcal{R}}$ the set of words in $\tilde{\mathcal{X}}\cup\tilde{\mathcal{X}}^{-1}$ representing an element of the kernel of the natural epimorphism from the free group generated by $\tilde{\mathcal{X}}$ to $\mathcal{M}_g$. For a word $w=\alpha_1\cdots \alpha_n\in\tilde{\mathcal{R}}$, we define an integer $$I_g(w):=-\sum_{j=1}^{n-1} \tau_g (\overline{\alpha_{n-j}},\overline{\alpha_{n-j+1}}\cdots\overline{\alpha_n}) -s(w),$$ where $\tau_g$ is the signature cocycle (Definition \[cocycle\]), $\overline{\alpha}$ is the image of $\alpha\in\tilde{\mathcal{X}}\cup\tilde{\mathcal{X}}^{-1}$ under the composition of the natural map $\tilde{\mathcal{X}}\cup\tilde{\mathcal{X}}^{-1}\rightarrow\mathcal{M}_g$ and a natural epimorphism $\mathcal{M}_g^*\rightarrow {\rm Sp}(2g,\Bbb{Z})$, and $s(w)$ is the number of Dehn twists along separating simple closed curves included in $w$. Suppose that $B$ is a $2$–sphere. If we choose a monodromy representation $\rho$ and a Hurwitz arc system $\mathcal{A}$ for $\Delta$ with base point $b_0$, we have a Hurwitz system $(\alpha_1,\ldots ,\alpha_n)\in (\mathcal{M}_g)^n$ of $f$. Since $\alpha_1,\ldots ,\alpha_n$ are Dehn twists and $\alpha_1\cdots \alpha_n=1$ in $\mathcal{M}_g$, we think $(\alpha_1,\ldots ,\alpha_n)$ as a word $w:=\alpha_1\cdots \alpha_n$ in $\tilde{\mathcal{R}}$. Theorem \[meyer\] and Novikov’s additivity for signature imply the next theorem. \[nagami\] The signature $\sigma(M)$ of $M$ is equal to $I_g(w)$. We are now ready to prove Theorem \[signature\]. Choose a base point $b_0\in B-\Gamma$ and a disk $D$ in $B-\Gamma$ centered at $b_0$. We denote the set of edges of $\Gamma$ by $E(\Gamma)$. For each $e\in E(\Gamma)$, we choose a point $b_e$ in a region of $B-\Gamma$ adjacent to $e$, and a simple path $\gamma_e$ from $b_e$ to $b_0$ which intersects with edges of $\Gamma$ transversely and does not intersect with vertices of $\Gamma$. Let $w_e$ be the intersection word of $\gamma_e$ with respect to $\Gamma$ and $i_e\in\{0,1,\ldots ,2g\}$ the label of $e$. We choose a family $\{D_e\}_{e\in E(\Gamma)}$ of mutually disjoint disks included in $D$ and put the chart $\Gamma_e$ depicted in Figure \[chartE\] in $D_e$ for each $e$. $w_e^{-1}$ \[r\] at 0 34 $i_e$ \[t\] at 62 31 ![Chart $\Gamma_e$[]{data-label="chartE"}](chartE "fig:") Taking the union of $\Gamma$ with $\Gamma_e$ for all $e\in E(\Gamma)$, we obtain a new chart $\Gamma_1$ in $B$, which describes a fiber sum $f_1\co M_1\rightarrow B$ of $f$ with Lefschetz fibrations over $S^2$ described by a free edge. For each $e\in E(\Gamma)$, we apply channel changes as in Figure \[channelC\] to let a free edge pass through the edges intersecting with $\gamma_e$. We then apply a channel change as in Figure \[channelD\] to ‘cut’ $e$ into two edges. Thus we obtain a new chart $\Gamma_2$ in $B$. $e$ \[b\] at 1 199 $b_e$ \[b\] at 16 158 $i_e$ \[l\] at 7 128 $\gamma_e$ \[b\] at 40 160 $w_e$ \[tl\] at 85 131 $w_e^{-\negthinspace 1}$ \[b\] at 111 157 $b_0$ \[b\] at 90 158 $i_e$ \[t\] at 172 152 $\Gamma_e$ \[l\] at 230 156 $e$ \[b\] at 1 84 $i_e$ \[l\] at 7 13 $w_e$ \[tl\] at 84 16 $i_e$ \[t\] at 172 38 $w_e^{-\negthinspace 1}$ \[b\] at 172 85 ![Channel change[]{data-label="channelC"}](channelC "fig:") $e$ \[b\] at 3 86 $i_e$ \[l\] at 7 13 $i_e$ \[l\] at 70 41 $i_e$ \[l\] at 176 13 $i_e$ \[l\] at 176 72 ![Channel change[]{data-label="channelD"}](channelD "fig:") Since each component of $\Gamma_2$ is a tree, a Lefschetz fibration $f_2\co M_2\rightarrow B$ corresponding to $\Gamma_2$ is a fiber sum of a Lefschetz fibration $f_3\co M_3\rightarrow S^2$ with a trivial $\Sigma_g$–bundle over $B$. Drawing a copy of $\Gamma_2$ in $S^2$, we have a chart $\Gamma_3$ corresponding to $f_3$. The signature of a Lefschetz fibration over $S^2$ described by a free edge is equal to zero because $\tau_g(A,A^{-1})=0$ for any $A\in {\rm Sp}(2g,\Bbb{Z})$ (see Meyer [@Meyer1973 Section 2]). Hence we have $$\sigma(M)=\sigma(M_1)=\sigma(M_2)=\sigma(M_3)+\sigma(\Sigma_g\times B) =\sigma(M_3)$$ by Theorem \[classification\] and Novikov’s additivity. Since we did not change the numbers of white vertices and black vertices of type $\ell_h^{\pm}$ to make $\Gamma_3$ from $\Gamma$, we see $\sigma(\Gamma_3)=\sigma(\Gamma)$. Hence we only have to show $\sigma(M_3)=\sigma(\Gamma_3)$ in order to conclude $\sigma(M)=\sigma(\Gamma)$. Applying chart moves of transition to each component of $\Gamma_3$ as in Figure \[transition\], we remove white vertices of type $r_F(i,j)^{\pm 1}, r_B(i)^{\pm 1},r_H^{\pm 1}$ to obtain a union of copies of $L_0(i), L_h, L_h^*, R_C, R_C^*, R_L, R_L^*$, where $L_0(i), L_h, R_C, R_L$ are charts depicted in Figure \[chartB\] and Figure \[chartF\], and $L_h^*$ (resp. $R_C^*$, $R_L^*$) is the mirror image of $L_h$ (resp. $R_C$, $R_L$) with edges orientation reversed. For the proof of $\sigma(M_3)=\sigma(\Gamma_3)$, it is enough to show that the signature of a Lefschetz fibration described by each of these charts coincides with the signature of the chart. $j$ \[br\] at 34 192 $i$ \[bl\] at 64 192 $i$ \[tr\] at 34 161 $j$ \[tl\] at 64 161 $j$ \[br\] at 103 192 $i$ \[bl\] at 134 192 $j$ \[tl\] at 134 161 $i$ \[br\] at 172 195 $i\negthinspace +\negthinspace 1$ \[bl\] at 209 195 $i\negthinspace +\negthinspace 1$ \[r\] at 165 177 $i$ \[l\] at 216 177 $i$ \[tr\] at 171 157 $i\negthinspace +\negthinspace 1$ \[tl\] at 209 157 $i$ \[br\] at 273 195 $i\negthinspace +\negthinspace 1$ \[bl\] at 309 195 $i\negthinspace +\negthinspace 1$ \[t\] at 267 175 $i$ \[l\] at 317 177 $i\negthinspace +\negthinspace 1$ \[tl\] at 310 157 $2g$ \[b\] at 43 125 $1$ \[b\] at 69 125 $1$ \[b\] at 78 125 $2g$ \[b\] at 103 125 $u^{-\negthinspace 1}$ \[l\] at 144 81 $1$ \[l\] at 142 62 $u$ \[l\] at 144 44 $2g$ \[t\] at 103 0 $1$ \[t\] at 78 0 $1$ \[t\] at 69 0 $2g$ \[t\] at 43 0 $u^{-\negthinspace 1}$ \[r\] at 2 44 $1$ \[r\] at 6 63 $u$ \[r\] at 2 80 $1$ \[b\] at 210 64 $u$ \[b\] at 238 119 $2g$ \[b\] at 255 114 $1$ \[b\] at 281 114 $1$ \[b\] at 290 114 $2g$ \[b\] at 316 114 $u^{-\negthinspace 1}$ \[b\] at 332 119 ![Chart moves of transition[]{data-label="transition"}](transition "fig:") $i$ \[b\] at 25 28 $1$ \[b\] at 218 50 $2h$ \[b\] at 199 50 $1$ \[b\] at 188 50 $2h$ \[b\] at 169 50 $1$ \[b\] at 135 50 $2h$ \[b\] at 113 50 ![Charts $L_0(i)$ and $L_h$[]{data-label="chartB"}](chartG "fig:") $1$ \[b\] at 124 83 $2$ \[b\] at 115 83 $3$ \[b\] at 107 83 $1$ \[b\] at 98 83 $2$ \[b\] at 91 83 $3$ \[b\] at 82 83 $1$ \[b\] at 73 83 $2$ \[b\] at 64 83 $3$ \[b\] at 56 83 $1$ \[b\] at 48 83 $2$ \[b\] at 39 83 $3$ \[b\] at 31 83 $0$ \[t\] at 9 5 $4$ \[t\] at 18 5 $3$ \[t\] at 27 5 $2$ \[t\] at 35 5 $1$ \[t\] at 43 5 $1$ \[t\] at 52 5 $2$ \[t\] at 61 5 $3$ \[t\] at 69 5 $4$ \[t\] at 78 5 $0$ \[t\] at 86 5 $4$ \[t\] at 94 5 $3$ \[t\] at 103 5 $2$ \[t\] at 112 5 $1$ \[t\] at 120 5 $1$ \[t\] at 129 5 $2$ \[t\] at 138 5 $3$ \[t\] at 146 5 $4$ \[t\] at 154 5 $\delta_3$ \[b\] at 216 88 $1$ \[b\] at 234 83 $3$ \[b\] at 243 83 $5$ \[b\] at 251 83 $\tau_1$ \[b\] at 267 88 $\tau_2$ \[b\] at 293 88 $0$ \[t\] at 315 4 $\tau_2^{-\negthinspace 1}$ \[t\] at 297 1 $\tau_1^{-\negthinspace 1}$ \[t\] at 273 1 $\tau_2$ \[t\] at 248 1 $0$ \[t\] at 230 4 $\tau_2^{-\negthinspace 1}$ \[t\] at 214 1 $0$ \[t\] at 196 4 ![Charts $R_C$ and $R_L$[]{data-label="chartF"}](chartF "fig:") Let $\Gamma_4$ be one of $L_0(i), L_h, L_h^*, R_C, R_C^*, R_L, R_L^*$ drawn in $S^2$ and $f_4\co M_4\rightarrow S^2$ a Lefschetz fibration described by $\Gamma_4$. If $\Gamma_4$ is equal to $L_0(i)$, it is easily seen that $\sigma(M_4)=\sigma(\Gamma_4)$. If $\Gamma_4$ is equal to $L_h$, the word $\ell_h^{-1}\sigma_h$ corresponds to a Hurwitz system of $f_4$ (see Figure \[chartB\]), where $\sigma_h$ is a right-handed Dehn twist along the curve $s_h$ depicted in Figure \[curves\]. Thus we have $$\sigma(M_4)=I_g(\ell_h^{-1}\sigma_h)=4h(h+1)-1=\sigma(\Gamma_4)$$ from Definition \[sigma\], Theorem \[nagami\], and explicit computations for $I_g$ due to Endo and Nagami [@EN2004 Lemma 3.5, Proposition 3.9]. If $\Gamma_4$ is equal to $R_C$ (resp. $R_L$), the word $r_C$ (resp. $r_L$) corresponds to a Hurwitz system of $f_4$ (see Figure \[chartF\]). Thus we have $$\sigma(M_4)=I_g(r_C)=-6=\sigma(\Gamma_4) \quad ({\rm resp.} \;\sigma(M_4)=I_g(r_L)=-1=\sigma(\Gamma_4))$$ from Definition \[sigma\], Theorem \[nagami\], and formulae of Endo and Nagami [@EN2004 Lemma 3.5, Remark 3.7, Propositions 3.9 and 3.10]. Suppose that $\Gamma_4$ is equal to one of $L_h^*, R_C^*, R_L^*$. The mirror image $\Gamma_4^*$ of $\Gamma_4$ with edges orientation reversed corresponds to the Lefschetz fibration $f_4\co -M_4\rightarrow S^2$ with total space orientation reversed. Hence we have $$\sigma(M_4)=-\sigma(-M_4)=-\sigma(\Gamma_4^*)=\sigma(\Gamma_4)$$ because we have already shown that $\sigma(M_4)=\sigma(\Gamma_4)$ is valid for $\Gamma_4=L_h, R_C, R_L$. This completes the proof of Theorem \[signature\]. Stabilization theorems ====================== In this section we prove two theorems on stabilization of Lefschetz fibrations under taking fiber sums with copies of a fixed Lefschetz fibration. Following Auroux [@Auroux2005], we first introduce a notion of universality for Lefschetz fibrations. Suppose that $g$ is greater than two. A Lefschetz fibration of genus $g$ over $S^2$ is called [*universal*]{} if it is irreducible, chiral, and it contains $2g+1$ singular fibers of type ${\rm I}^+$ whose vanishing cycles $a_0,a_1,\ldots ,a_{2g}\subset \Sigma_g$ satisfies the following conditions: (i) $a_i$ and $a_{i+1}$ intersect transversely at one point for every $i\in\{1,\ldots ,2g-1\}$; (ii) $a_0$ and $a_4$ intersect transversely at one point; (iii) $a_i$ and $a_j$ does not intersect for other pairs $(i,j)$. A Lefschetz fibration over $S^2$ is universal if and only if it is described by a chart $\Gamma_0$ depicted in Figure \[universal\] by virtue of Proposition \[correspondence\], where the blank labeled with ${\rm T}_0$ is filled only with edges, white vertices, and black vertices of type $\ell_0(i)$. $0$ \[b\] at 16 57 $1$ \[b\] at 25 57 $2g$ \[b\] at 53 57 $T_0$ at 36 13 \[l\] at 75 45 ![Universal chart $\Gamma_0$[]{data-label="universal"}](universal "fig:") A universal Lefschetz fibration does exist for every $g$ greater than two. For example, Lefschetz fibrations $f_g^0, \, f_g^A, \, f_g^B, \, f_g^C, \, f_g^D$ constructed by Auroux [@Auroux2005] are universal except $f_g^D$ for $g=3$. There would be many universal Lefschetz fibrations of genus $g$ for a fixed $g$. We now state the first of our main theorems. Let $B$ be a connected closed oriented surface and $f_0\co M_0\rightarrow S^2$ a universal Lefschetz fibration of genus $g$. \[main1\] Let $f\co M\rightarrow B$ and $f'\co M'\rightarrow B$ be Lefschetz fibrations of genus $g$. There exists a non-negative integer $N$ such that $f\# Nf_0$ is isomorphic to $f'\# Nf_0$ if and only if the following conditions hold: (i) $n_0^{\pm}(f)=n_0^{\pm}(f')$; (ii) $n_h^{\pm}(f)=n_h^{\pm}(f')$ for every $h=1,\ldots ,[g/2]$; (iii) $\sigma(M)=\sigma(M')$. Auroux [@Auroux2005] proved the ‘if’ part of Theorem \[main1\] for chiral Lefschetz fibrations over $S^2$ under the assumption that $f$ and $f'$ have sections with the same self-intersection number. Hasegawa [@Hasegawa2006] gave another proof of Auroux’s theorem by using chart description. Moreover he removed the assumption about existence and self-intersection number of sections in Auroux’s theorem. The isomorphism class of a fiber sum $f\#_{\Psi}\, f_0$ of a Lefschetz fibration $f$ with a universal Lefschetz fibration $f_0$ does not depend on a choice of an orientation preserving diffeomorphism $\Psi$ (see Proof of Theorem \[main1\]). We first prove the ‘if’ part. Assume that $f$ and $f'$ satisfy the conditions (i), (ii), and (iii). Let $\Gamma$ and $\Gamma'$ be charts in $B$ corresponding to $f$ and $f'$, respectively. We suppose that $f_0$ is described by a chart $\Gamma_0$ depicted in Figure \[universal\]. Since every edge has two adjacent vertices, the sum of the signed numbers of adjacent edges for all vertices of $\Gamma$ is equal to zero: $$10n_C(\Gamma)+n_L(\Gamma)-\sum_{i=0}^{2g}n_0(i)(\Gamma) -4\sum_{h=1}^{[g/2]}h(2h+1)\cdot n_h(\Gamma)=0.$$ A similar equality for $\Gamma'$ also holds. Interpreting the conditions (i) and (ii) as conditions on $\Gamma$ and $\Gamma'$, we have $\sum_{i=0}^{2g}n_0(i)(\Gamma)=\sum_{i=0}^{2g}n_0(i)(\Gamma')$ and $n_h(\Gamma)=n_h(\Gamma')$ for $h=1,\ldots ,[g/2]$. Thus we obtain $$10n_C(\Gamma)+n_L(\Gamma)=10n_C(\Gamma')+n_L(\Gamma').$$ On the other hand, we have $$-6\, n_C(\Gamma)-n_L(\Gamma)=-6\, n_C(\Gamma')-n_L(\Gamma')$$ by the condition (iii), Theorem \[signature\], and $n_h(\Gamma)=n_h(\Gamma')$ for $h=1,\ldots ,[g/2]$. Hence $n_C(\Gamma)=n_C(\Gamma')$ and $n_L(\Gamma)=n_L(\Gamma')$. Let $N$ be an integer larger than both of the number of edges of $\Gamma$ and that of $\Gamma'$. Choose a base point $b_0\in B-(\Gamma\cup\Gamma')$. The fiber sum $f\# Nf_0$ is described by a chart $(\cdots ((\Gamma\#_{w_1}\Gamma_0)\#_{w_2}\Gamma_0)\cdots )\#_{w_N}\Gamma_0$ for some words $w_1,\ldots ,w_N$ in $\mathcal{X}\cup\mathcal{X}^{-1}$. Since hoops surrounding $\Gamma_0$ can be removed by use of the edges of $\Gamma_0$ as in Figure \[removeA\], the chart is transformed into a product $\Gamma\oplus N\Gamma_0$ by channel changes. Similarly, the fiber sum $f'\# Nf_0$ is described by a product $\Gamma'\oplus N\Gamma_0$. $i$ \[b\] at 41 71 $0$ \[b\] at 23 51 $i$ \[b\] at 43 52 $2g$ \[b\] at 61 51 $T_0$ at 43 11 $i$ \[b\] at 188 72 $0$ \[b\] at 142 51 $2g$ \[b\] at 179 51 $T_0$ at 161 11 $0$ \[b\] at 257 51 $i$ \[b\] at 275 52 $2g$ \[b\] at 293 51 $T_0$ at 276 11 ![Removing a hoop[]{data-label="removeA"}](removeA "fig:") We choose and fix $2g+1$ edges of $\Gamma_0$ which are labeled with $0,1,\ldots ,2g$ and adjacent to black vertices. We apply chart moves only to these edges in the following. Since $\Gamma_0$ can pass through any edge of $\Gamma$ as shown in Figure \[pass\], we can move $\Gamma_0$ to any region of $B-\Gamma$ by channel changes. $i$ \[b\] at 35 69 $0$ \[b\] at 18 51 $i$ \[b\] at 36 52 $2g$ \[b\] at 55 51 $T_0$ at 35 11 \[t\] at 37 0 $i$ \[b\] at 106 72 $i$ \[b\] at 160 72 $0$ \[b\] at 115 52 $2g$ \[b\] at 151 52 $T_0$ at 134 11 \[t\] at 134 0 $i$ \[t\] at 200 66 $i$ \[t\] at 255 66 $2g$ \[t\] at 210 86 $0$ \[t\] at 246 86 $T_0$ at 228 126 \[t\] at 228 0 $i$ \[t\] at 322 66 $2g$ \[t\] at 303 86 $i$ \[t\] at 321 85 $0$ \[t\] at 341 86 $T_0$ at 322 126 \[t\] at 321 0 ![Passing through an edge[]{data-label="pass"}](pass "fig:") For each edge of $\Gamma$, we move a copy of $\Gamma_0$ to a region adjacent to the edge and apply a channel change to the edge and $\Gamma_0$ as in Figure \[pass\] (a) and (b). Applying chart moves of transition to each component of the chart as in Figure \[transition\], we remove white vertices of type $r_F(i,j)^{\pm 1}, r_B(i)^{\pm 1},r_H^{\pm 1}$ to obtain a union of copies of $L_0(i), \tilde{L}_h, L_h^*, \tilde{R}_C, \hat{R}_C, \tilde{R}_L, \hat{R}_L, \Gamma_0$ shown in Figures \[chartH\], \[chartC\], \[chartD\], where we use a simplification of diagrams as in Figure \[symbol\]. $1$ \[b\] at 113 50 $2h$ \[b\] at 92 50 $1$ \[b\] at 82 50 $2h$ \[b\] at 64 50 $1$ \[b\] at 30 50 $2h$ \[b\] at 8 50 ![Chart $\tilde{L}_h$[]{data-label="chartH"}](chartH "fig:") $1$ \[b\] at 124 78 $2$ \[b\] at 115 78 $3$ \[b\] at 107 78 $1$ \[b\] at 98 78 $2$ \[b\] at 91 78 $3$ \[b\] at 82 78 $1$ \[b\] at 73 78 $2$ \[b\] at 64 78 $3$ \[b\] at 56 78 $1$ \[b\] at 48 78 $2$ \[b\] at 39 78 $3$ \[b\] at 31 78 $0$ \[t\] at 9 0 $4$ \[t\] at 18 0 $3$ \[t\] at 27 0 $2$ \[t\] at 35 0 $1$ \[t\] at 43 0 $1$ \[t\] at 52 0 $2$ \[t\] at 61 0 $3$ \[t\] at 69 0 $4$ \[t\] at 78 0 $0$ \[t\] at 86 0 $4$ \[t\] at 94 0 $3$ \[t\] at 103 0 $2$ \[t\] at 112 0 $1$ \[t\] at 120 0 $1$ \[t\] at 129 0 $2$ \[t\] at 138 0 $3$ \[t\] at 146 0 $4$ \[t\] at 154 0 $1$ \[b\] at 226 79 $2$ \[b\] at 235 79 $3$ \[b\] at 244 79 $1$ \[b\] at 252 79 $2$ \[b\] at 260 79 $3$ \[b\] at 269 79 $1$ \[b\] at 278 79 $2$ \[b\] at 286 79 $3$ \[b\] at 295 79 $1$ \[b\] at 302 79 $2$ \[b\] at 311 79 $3$ \[b\] at 320 79 $0$ \[t\] at 341 0 $4$ \[t\] at 333 0 $3$ \[t\] at 324 0 $2$ \[t\] at 316 0 $1$ \[t\] at 308 0 $1$ \[t\] at 299 0 $2$ \[t\] at 291 0 $3$ \[t\] at 281 0 $4$ \[t\] at 274 0 $0$ \[t\] at 265 0 $4$ \[t\] at 256 0 $3$ \[t\] at 247 0 $2$ \[t\] at 239 0 $1$ \[t\] at 231 0 $1$ \[t\] at 222 0 $2$ \[t\] at 214 0 $3$ \[t\] at 206 0 $4$ \[t\] at 197 0 ![Charts $\tilde{R}_C$ and $\hat{R}_C$[]{data-label="chartC"}](chartC "fig:") $\delta_3$ \[b\] at 29 88 $1$ \[b\] at 46 83 $3$ \[b\] at 55 83 $5$ \[b\] at 64 83 $\tau_1$ \[b\] at 80 88 $\tau_2$ \[b\] at 105 88 $0$ \[t\] at 128 3 $\tau_2^{-\negthinspace 1}$ \[t\] at 110 0 $\tau_1^{-\negthinspace 1}$ \[t\] at 85 0 $\tau_2$ \[t\] at 60 0 $0$ \[t\] at 43 3 $\tau_2^{-\negthinspace 1}$ \[t\] at 26 0 $0$ \[t\] at 9 3 $0$ \[t\] at 296 3 $\tau_2$ \[t\] at 279 0 $0$ \[t\] at 262 3 $\tau_2^{-\negthinspace 1}$ \[t\] at 245 0 $\tau_1$ \[t\] at 220 0 $\tau_2$ \[t\] at 195 0 $0$ \[t\] at 177 3 $\tau_2^{-\negthinspace 1}$ \[b\] at 197 88 $\tau_1^{-\negthinspace 1}$ \[b\] at 223 88 $5$ \[b\] at 240 85 $3$ \[b\] at 249 85 $1$ \[b\] at 258 85 $\delta_3^{-\negthinspace 1}$ \[b\] at 273 88 ![Charts $\tilde{R}_L$ and $\hat{R}_L$[]{data-label="chartD"}](chartD "fig:") $i$ \[b\] at 3 45 $0$ \[b\] at 89 57 $i$ \[b\] at 107 71 $2g$ \[b\] at 126 57 $T_0$ at 108 14 $=$ at 40 28 ![Simplification of diagram[]{data-label="symbol"}](symbol "fig:") If there is a pair of $\tilde{R}_C$ and $\hat{R}_C$, we remove them by a death of a pair of white vertices to obtain many copies of $\Gamma_0$. Similarly, we remove a pair of $\tilde{R}_L$ and $\hat{R}_L$. Since there is at least one $\Gamma_0$, any copy of $L_0(i)$ can be transformed into $L_0(1)$ as in Figure \[label\]. $j$ \[l\] at 28 152 $0$ \[b\] at 83 186 $i$ \[b\] at 84 201 $2g$ \[b\] at 120 186 $T_0$ at 101 143 $j$ \[l\] at 199 152 $i$ \[l\] at 212 164 $0$ \[b\] at 255 186 $j$ \[b\] at 255 201 $2g$ \[b\] at 292 186 $T_0$ at 274 143 $j$ \[l\] at 370 152 $i$ \[l\] at 382 150 $j$ \[l\] at 390 142 $0$ \[b\] at 426 186 $j$ \[b\] at 444 186 $2g$ \[b\] at 463 186 $T_0$ at 444 143 $i$ \[b\] at 26 91 $j$ \[t\] at 26 46 $i$ \[b\] at 26 23 $j$ \[t\] at 26 13 $0$ \[b\] at 82 73 $j$ \[b\] at 100 73 $2g$ \[b\] at 119 73 $T_0$ at 100 29 $i$ \[l\] at 202 82 $j$ \[tl\] at 216 8 $i$ \[b\] at 201 10 $0$ \[b\] at 254 73 $j$ \[b\] at 273 73 $2g$ \[b\] at 291 73 $T_0$ at 273 29 $i$ \[l\] at 372 44 $0$ \[b\] at 425 73 $j$ \[b\] at 444 73 $2g$ \[b\] at 463 73 $T_0$ at 443 29 ![Changing a label ($j=i+1$ or $(i,j)=(4,0)$)[]{data-label="label"}](label "fig:") Thus we have a union $\Gamma_1$ of $n_0^-(\Gamma)$ copies of $L_0(1)$, $n_h^+(\Gamma)$ copies of $\tilde{L}_h$, $n_h^-(\Gamma)$ copies of $L_h^*$, $|n_C(\Gamma)|$ copies of $\tilde{R}_C$ (or $\hat{R}_C$), $|n_L(\Gamma)|$ copies of $\tilde{R}_L$ (or $\hat{R}_L$), and $k$ copies of $\Gamma_0$ for some $k$. A similar argument implies that $\Gamma'\oplus N\Gamma_0$ is transformed into a union $\Gamma'_1$ of $n_0^-(\Gamma')$ copies of $L_0(1)$, $n_h^+(\Gamma')$ copies of $\tilde{L}_h$, $n_h^-(\Gamma')$ copies of $L_h^*$, $|n_C(\Gamma')|$ copies of $\tilde{R}_C$ (or $\hat{R}_C$), $|n_L(\Gamma')|$ copies of $\tilde{R}_L$ (or $\hat{R}_L$), and $k'$ copies of $\Gamma_0$ for some $k'$ by chart moves of type W and chart moves of transition. By virtue of the conditions (i) and (ii) together with $n_C(\Gamma)=n_C(\Gamma')$, $n_L(\Gamma)=n_L(\Gamma')$, $n_0^+(\Gamma\oplus N\Gamma_0)=n_0^+(\Gamma_1)$, and $n_0^+(\Gamma'\oplus N\Gamma_0)=n_0^+(\Gamma'_1)$, we conclude that $k=k'$ because of $n_0^+(\Gamma_0)\ne 0$. Hence $\Gamma_1$ is transformed into $\Gamma'_1$ by an ambient isotopy of $B$ relative to $b_0$, which means that $\Gamma\oplus N\Gamma_0$ is transformed into $\Gamma'\oplus N\Gamma_0$ by chart moves of type W, chart moves of transition, and ambient isotopies of $B$ relative to $b_0$. Therefore $f\# Nf_0$ is (strictly) isomorphic to $f'\# Nf_0$ by Theorem \[classification\]. We next prove the ‘only if’ part. Take a non-negative integer $N$ so that $f\# Nf_0$ is isomorphic to $f'\# Nf_0$. Since an isomorphism preserves numbers and types of vanishing cycles and signatures, we have $n_0^{\pm}(f\# Nf_0)=n_0^{\pm}(f'\# Nf_0)$, $n_h^{\pm}(f\# Nf_0)=n_h^{\pm}(f'\# Nf_0)$ for every $h=1,\ldots ,[g/2]$, and $\sigma(M\#_F NM_0)=\sigma(M'\#_F NM_0)$. The conditions (i), (ii), (iii) follows from additivity of $n_0^{\pm}, n_h^{\pm}, \sigma$ under fiber sum. A Lefschetz fibration of genus $g$ over $S^2$ is called [*elementary*]{} if it contains exactly two singular fibers of type ${\rm I}^+$ and of type ${\rm I}^-$ which have the same vanishing cycles. A chart $L_0(i)$ in $S^2$ corresponds to an elementary Lefschetz fibration. Two elementary Lefschetz fibrations of genus $g$ are isomorphic to each other. The total space of an elementary Lefschetz fibration of genus $g$ is diffeomorphic to $\Sigma_{g-1}\times S^2\# S^1\times S^3$. We state the second of our main theorems. Let $B$ be a connected closed oriented surface and $f_{\star}\co M_{\star}\rightarrow S^2$ an elementary Lefschetz fibration of genus $g$. \[main2\] Let $f\co M\rightarrow B$ and $f'\co M'\rightarrow B$ be Lefschetz fibrations of genus $g$. There exists a non-negative integer $N$ such that a fiber sum $f\# Nf_{\star}$ is isomorphic to a fiber sum $f'\# Nf_{\star}$ if and only if the following conditions hold: (i) $n_0^{\pm}(f)=n_0^{\pm}(f')$; (ii) $n_h^{\pm}(f)=n_h^{\pm}(f')$ for every $h=1,\ldots ,[g/2]$; (iii) $\sigma(M)=\sigma(M')$. In contrast to Theorem \[main1\], the isomorphism class of a fiber sum $f\#_{\Psi}\, f_{\star}$ of a Lefschetz fibration $f$ with an elementary Lefschetz fibration $f_{\star}$ depends on a choice of an orientation preserving diffeomorphism $\Psi$ in general. We only show the ‘if’ part. The ‘only if’ part is the same as that of the proof of Theorem \[main1\]. Assume that $f$ and $f'$ satisfy the conditions (i), (ii), and (iii). Let $\Gamma$ and $\Gamma'$ be charts in $B$ corresponding to $f$ and $f'$, respectively. It follows from the same argument as in the proof of Theorem \[main1\] that $n_C(\Gamma)=n_C(\Gamma')$ and $n_L(\Gamma)=n_L(\Gamma')$. Let $N$ be an integer larger than both of the number of edges of $\Gamma$ and that of $\Gamma'$. We construct the chart $\Gamma_e$ in $B$ for each $e\in E(\Gamma)$ as in the proof of Theorem \[signature\]. Taking the union of $\Gamma$ with $\Gamma_e$ for all $e\in E(\Gamma)$ and with $N-\#E(\Gamma)$ copies of $L_1(1)$, we obtain a new chart $\Gamma_1$ in $B$, which describes a fiber sum $f\# Nf_{\star}$. Applying channel changes as in the proof of Theorem \[signature\] and deaths of pairs of white vertices appropriately, we obtain a union $\Gamma_2$ of $n_h^+(\Gamma)$ copies of $L_h$, $n_h^-(\Gamma)$ copies of $L_h^*$, $|n_C(\Gamma)|$ copies of $R_C$ (or $R_C^*$), $|n_L(\Gamma)|$ copies of $R_L$ (or $R_L^*$), and $k_i$ copies of $L_0(i)$ for some $k_i$, Similarly, $\Gamma'$ is transformed into $\Gamma'_1$, which describes a fiber sum $f'\# Nf_{\star}$, and then a union $\Gamma'_2$ of $n_h^+(\Gamma')$ copies of $L_h$, $n_h^-(\Gamma')$ copies of $L_h^*$, $|n_C(\Gamma')|$ copies of $R_C$ (or $R_C^*$), $|n_L(\Gamma')|$ copies of $R_L$ (or $R_L^*$), and $k'_i$ copies of $L_0(i)$ for some $k'_i$. A similar argument on the number $n_0^+$ as in the proof of Theorem \[main1\] implies that $k_0+k_1+\cdots +k_{2g}=k'_0+k'_1+\cdots +k'_{2g}$. Adding $|k_i-k'_i|$ copies of $L_0(i)$ to either $\Gamma_2$ or $\Gamma'_2$ if necessary, we may assume that $k_i=k'_i$ for every $i\in\{0,1,\ldots ,2g\}$. Hence $\Gamma_2$ is transformed into $\Gamma'_2$ by an ambient isotopy of $B$ relative to $b_0$, which means that $f\# Nf_{\star}$ is (strictly) isomorphic to $f'\# Nf_{\star}$ by Theorem \[classification\]. Let $g$ be an integer greater than two and $B_1,\ldots ,B_r$ connected closed oriented surfaces. We consider a Lefschetz fibration $f_i\co M_i\rightarrow B_i$ of genus $g$ for each $i\in\{1,\ldots ,r\}$, and a universal Lefschetz fibration $f_0\co M_0\rightarrow S^2$ of genus $g$. \[sums\] For (possibly different) fiber sums $f$ and $f'$ of $f_1,\ldots ,f_r$, fiber sums $f\# f_0$ and $f'\# f_0$ are isomorphic to each other. Let $\Gamma$ and $\Gamma'$ be charts corresponding to $f$ and $f'$. Since hoops surrounding a component of $\Gamma$ (and $\Gamma'$) can be removed by use of the edges of $\Gamma_0$ as in Figure \[removeA\], $\Gamma\#\Gamma_0$ and $\Gamma'\#\Gamma_0$ are transformed into the same chart. Proposition \[sums\] implies that there are many examples of non-isomorphic Lefschetz fibrations with the same base, the same fiber, and the same numbers of singular fibers of each type which become isomorphic after one stabilization. For example, the Lefschetz fibration on $E(n)_K$ constructed by Fintushel and Stern [@FS2004 Theorem 14] (see also Park and Yun [@PY2009]) for a fibered knot $K$ becomes isomorphic to that on $E(n)_{K'}$ for another fibered knot $K'$ of the same genus after one stabilization. Similar results hold for Lefschetz fibrations on $Y(n;K_1,K_2)$ constructed by Fintushel and Stern [@FS2004 §7] (see also Park and Yun [@PY2011]) as well as fiber sums of (generalizations of) Matsumoto’s fibration studied by Ozbagci and Stipsicz [@OS2000], Korkmaz [@Korkmaz2001; @Korkmaz2009], and Okamori [@Okamori2011]. Variations and problems ======================= In this section we discuss possible variations of chart description for Lefschetz fibrations. If we replace the triple $(\mathcal{X},\mathcal{R},\mathcal{S})$ defined in Section 2 with other triples, we obtain various chart descriptions for Lefschetz fibrations (see Kamada [@Kamada2007] and Hasegawa [@Hasegawa2006]). We first choose large $\mathcal{X},\mathcal{R}$, and $\mathcal{S}$. Let $\mathcal{X}$ be the set of right-handed Dehn twists along simple closed curves in $\Sigma_g$ and $\mathcal{S}$ the set of Dehn twists along non-trivial simple closed curves in $\Sigma_g$. By virtue of a theorem of Luo [@Luo1997], $\langle \mathcal{X}\, |\,\mathcal{R}\rangle$ gives an infinite presentation of $\mathcal{M}_g$ for the set $\mathcal{R}$ of the following four kinds of words: (0) trivial relator $r_T:=a$, where $a$ is the Dehn twist along a trivial simple closed curve on $\Sigma_g$; (1) primitive braid relator $r_P:=b^{-1}abc^{-1}$, where $a,b,c\in\mathcal{X}$ and the curve for $c$ is the image of the curve for $a$ by $b$; (2) $2$–chain relator $r_C:=(c_2c_1)^6d^{-1}$, where $c_1,c_2,d\in \mathcal{X}$ and the curves for $c_1$ and $c_2$ intersect transversely at one point and the curve for $d$ is the boundary curve of a regular neighborhood of the union of the curves for $c_1$ and $c_2$; (3) lantern relator $r_L:=cbad_4^{-1}d_3^{-1}d_2^{-1}d_1^{-1}$, where $a,b,c,d_1,d_2,d_3,d_4\in\mathcal{X}$ and the curves for $a$ and $b$ intersect transversely at two points with algebraic intersection number zero, the curve for $c$ is obtained by resolving the intersections of these two curves, and the curves for $d_1,d_2,d_3,d_4$ are the boundary curves of a regular neighborhood of those for $a,b,c$. Let $B$ be a connected closed oriented surface. Charts in $B$ for the triple $(\mathcal{X},\mathcal{R},\mathcal{S})$ defined above have white vertices of type $r_T^{\pm 1}, r_P^{\pm 1}, r_C^{\pm 1}, r_L^{\pm 1}$ (see Figure \[verticesD\]). For a chart $\Gamma$ in $B$, we denote the number of white vertices of type $r_X$ minus the number of white vertices of type $r_X^{-1}$ included in $\Gamma$ by $n_X(\Gamma)$, where $X=T,P,C,L$. $a$ \[b\] at 4 54 $a$ \[br\] at 40 50 $b$ \[bl\] at 70 50 $b$ \[tr\] at 40 24 $c$ \[tl\] at 70 24 $c_2$ \[b\] at 114 73 $c_1$ \[b\] at 123 73 $c_2$ \[b\] at 132 73 $c_1$ \[b\] at 141 73 $c_2$ \[b\] at 150 73 $c_1$ \[b\] at 159 73 $c_2$ \[b\] at 167 73 $c_1$ \[b\] at 176 73 $c_2$ \[b\] at 185 73 $c_1$ \[b\] at 194 73 $c_2$ \[b\] at 203 73 $c_1$ \[b\] at 212 73 $d$ \[t\] at 162 0 $c$ \[b\] at 263 73 $b$ \[b\] at 271 73 $a$ \[b\] at 281 73 $d_1$ \[t\] at 257 0 $d_2$ \[t\] at 267 0 $d_3$ \[t\] at 277 0 $d_4$ \[t\] at 287 0 ![Vertices of type $r_T$, $r_P$, $r_C$, $r_L$[]{data-label="verticesD"}](verticesD "fig:") \[signature2\] The signature $\sigma(M)$ of the total space $M$ of a Lefschetz fibration $f\co M\rightarrow B$ described by $\Gamma$ is equal to $-n_T(\Gamma)-7n_C(\Gamma)+n_L(\Gamma)$. It is seen by a similar argument to the proof of Theorem \[signature\]. Let $B$ be a connected closed oriented surface of genus $2$ and $(\mathcal{X},\mathcal{R},\mathcal{S})$ the triple defined above for $g=3$. Let $a,b,c,d_1,d_2,d_3,d_4,c_1,c_2,c_3$ be right-handed Dehn twists along simple closed curves of the same names on $\Sigma_3$ depicted in Figure \[twists\]. We present $B$ as an octagon with opposite sides identified and consider a chart $\Gamma$ and loops $\gamma_1,\gamma_2,\gamma_3,\gamma_4,\gamma_5$ based at $b_0$ in $B$ as in Figure \[exampleB\]. We use a simplification of diagrams as in Figure \[verticesE\] (a) if the curves for $x,y\in\mathcal{X}$ intersect transversely at one point, and that as in Figure \[verticesE\] (b) if the curves for $x$ and $y$ are disjoint. Since the intersection words of $\gamma_1,\gamma_2,\gamma_3,\gamma_4,\gamma_5$ with respect to $\Gamma$ are $$\begin{aligned} w_{\Gamma}(\gamma_1) & =d_2^{-1}, \quad w_{\Gamma}(\gamma_2) =c_3^{-1}c^{-1}d_2^{-1}c_3^{-1}, \\ w_{\Gamma}(\gamma_3) & =c_1^{-1}b^{-1}c_2^{-1}d_3^{-1} a^{-1}c_2^{-1}d_4^{-1}c_1^{-1}, \quad w_{\Gamma}(\gamma_4) =d_4a^{-1}, \quad w_{\Gamma}(\gamma_5) =d_1, \end{aligned}$$ a Lefschetz fibration $f:M\rightarrow B$ of genus $3$ described by $\Gamma$ is isomorphic to the Lefschetz fibration constructed by Korkmaz and Ozbagci [@KO2001 Theorem 1.2]. $f$ has only one singular fiber and it is of type ${\rm I}^+$. We can compute the signature $\sigma(M)$ of the total space $M$ by Proposition \[signature2\]: $$\sigma(M)=n_L(\Gamma)=-1,$$ which coincides with the value computed in [@EKKOS2002 Proposition 14]. $d_1$ \[b\] at 16 53 $d_4$ \[b\] at 70 53 $d_3$ \[b\] at 124 53 $d_2$ \[b\] at 177 53 $a$ \[r\] at 88 25 $b$ \[t\] at 132 17 $c_1$ \[b\] at 247 60 $c_2$ \[b\] at 302 60 $c_3$ \[b\] at 356 60 $c$ \[t\] at 284 16 ![Simple closed curves on $\Sigma_3$[]{data-label="twists"}](twists "fig:") $b_0$ \[t\] at 234 450 \[b\] at 145 427 \[b\] at 23 305 \[t\] at 24 144 \[t\] at 146 26 \[t\] at 323 33 \[t\] at 430 142 \[b\] at 432 306 \[b\] at 310 427 \[r\] at 116 361 $d_2$ \[b\] at 188 400 $d_1$ \[t\] at 158 358 $c$ \[r\] at 153 304 $b$ \[r\] at 193 289 $a$ \[t\] at 228 328 $d_4$ \[b\] at 259 404 $d_3$ \[b\] at 230 384 $c_3$ \[b\] at 97 324 $c$ \[b\] at 73 287 $d_2$ \[b\] at 58 253 $c_3$ \[b\] at 41 224 $c_3$ \[r\] at 104 219 $d_2$ \[b\] at 59 158 $c_3$ \[l\] at 161 212 $c$ \[r\] at 135 193 $d_2$ \[l\] at 113 143 $c_3$ \[r\] at 88 113 $c_1$ \[b\] at 240 300 $b$ \[b\] at 275 295 $c_2$ \[r\] at 278 245 $d_3$ \[l\] at 298 239 $a$ \[r\] at 319 198 $c_2$ \[l\] at 342 187 $d_4$ \[t\] at 412 260 $c_1$ \[l\] at 424 223 $a$ \[t\] at 420 184 $c_2$ \[b\] at 334 315 $d_3$ \[t\] at 353 310 $a$ \[r\] at 370 289 $c_2$ \[t\] at 383 270 $c_2$ \[t\] at 343 277 $c_1$ \[t\] at 292 171 $c_1$ \[r\] at 205 183 $b$ \[l\] at 224 164 $d_4$ \[r\] at 341 102 $c_1$ \[l\] at 364 82 $d_4$ \[t\] at 388 120 \[b\] at 203 343 ![Chart for Lefschetz fibration of Korkmaz and Ozbagci[]{data-label="exampleB"}](exampleB "fig:") $x$ \[b\] at 4 56 $y$ \[b\] at 26 56 $x$ \[b\] at 49 56 $y$ \[t\] at 4 0 $x$ \[t\] at 26 0 $y$ \[t\] at 49 0 $:=$ \[b\] at 70 22 $x$ \[b\] at 91 56 $y$ \[b\] at 113 56 $x$ \[b\] at 135 56 $y$ \[t\] at 91 0 $x$ \[t\] at 113 0 $y$ \[t\] at 135 0 $z$ \[b\] at 117 29 \[t\] at 70 -15 $x$ \[br\] at 180 44 $y$ \[bl\] at 210 44 $y$ \[tr\] at 180 12 $x$ \[tl\] at 210 12 $:=$ \[b\] at 228 22 $x$ \[br\] at 243 44 $y$ \[bl\] at 273 44 $y$ \[tr\] at 243 12 $x$ \[tl\] at 273 12 \[t\] at 228 -15 ![Simplification of vertices[]{data-label="verticesE"}](verticesE "fig:") Study various properties of Lefschetz fibrations by using chart description for the triple $(\mathcal{X},\mathcal{R},\mathcal{S})$ defined above. We next mention chart description for Lefschetz fibrations with bordered base and fiber. Kamada [@Kamada2007] gave a general theory for charts in a compact oriented surface with boundary. Various presentations of mapping class groups of surfaces with boundary have been investigated by researchers including Gervais [@Gervais2001], Labruère and Paris [@LP2001], Margalit and McCammond [@MM2009]. Combining these two kinds of studies, one can immediately obtain a chart description for Lefschetz fibrations with bordered base and fiber. Make use of chart description to study PALFs and Stein surfaces. It would be worth considering compositions of monodromy representations with appropriate homomorphisms and charts corresponding to the compositions. For example, Hasegawa [@Hasegawa2006a; @Hasegawa2006] adopted a homomorphism from the $m$–string braid group $B_m$ to the semi-direct product $(\mathbb{Z}_2)^m\times S_m$, while Endo and Kamada [@EK2014] used a standard epimorphism from the hyperelliptic mapping class group of a closed oriented surface of genus $g$ to the mapping class group of a sphere with $2g+2$ marked points. Consider chart descriptions for ‘nice’ representations of mapping class groups to study invariants and classifications of Lefschetz fibrations. Theorem \[main1\] and Theorem \[main2\] tell us that the numbers of singular fibers of all types and the signature of the total space completely determine the stable isomorphism class of a Lefschetz fibration with given base and fiber. Thus any numerical invariant of Lefschetz fibrations which is additive under fiber sum is determined by these invariants in principle. Construct numerical invariants of Lefschetz fibrations which are [*not*]{} additive under fiber sum. Nosaka [@Nosaka2014] has recently defined an invariant which is not additive under fiber sum. Non-numerical invariants such as monodromy group would be also useful (see Matsumoto [@Matsumoto1996] and Park and Yun [@PY2009; @PY2011]). [**Acknowledgements**]{}    The authors would like to thank the referees for their helpful suggestions and corrections. The first author was partially supported by JSPS KAKENHI Grant Numbers 21540079, 25400082. The third author was partially supported by JSPS KAKENHI Grant Numbers 21340015, 26287013. The fourth author was partially supported by JSPS KAKENHI Grant Numbers 21740042, 26400082. [99]{} D. Auroux, [*Fiber sums of genus $2$ Lefschetz fibrations*]{}, Turkish J. Math. [**27**]{} (2003) 1–10. D. Auroux, [*A stable classification of Lefschetz fibrations*]{}, Geom. Topol. [**9**]{} (2005), 203–217. J. Barge and É. Ghys, [*Cocycles d’Euler et de Maslov*]{}, Math. Ann. [**294**]{} (1992), 235–265. R. I. Baykur and S. Kamada, [*Classification of broken Lefschetz fibrations with small fiber genera*]{}, arXiv:1010.5814, to appear in J. Math. Soc. Japan. H. Endo and S. Kamada, [*Chart description for hyperelliptic Lefschetz fibrations and their stabilization*]{}, arXiv:1306.2707, to appear in Topology Appl. H. Endo and S. Kamada, [*Counting Dirac braids and hyperelliptic Lefschetz fibrations*]{}, in preparation. H. Endo, M. Korkmaz, D. Kotschick, B. Ozbagci, and A. Stipsicz, [*Commutators, Lefschetz fibrations and the signatures of surface bundles*]{}, Topology [**41**]{} (2002), 961–977. H. Endo and S. Nagami, [*Signature of relations in mapping class groups and non-holomorphic Lefschetz fibrations*]{}, Trans. Amer. Math. Soc. [**357**]{} (2004), 3179–3199. R. Fintushel and R. J. Stern, [*Families of simply connected $4$–manifolds with the same Seiberg-Witten invariants*]{}, Topology [**43**]{} (2004), 1449–1467. J.-M. Gambaudo and É. Ghys, [*Braids and signatures*]{}, Bull. Soc. Math. France [**133**]{} (2005), 541–579. S. Gervais, [*A presentation of the mapping class group of a punctured surface*]{}, Topology [**40**]{} (2001), 703–725. P. M. Gilmer and G. Masbaum, [*Maslov index, lagrangians, mapping class groups and TQFT*]{}, Forum Math. [**25**]{} (2011), 1067–1106. R. E. Gompf and A. I. Stipsicz, [*$4$–manifolds and Kirby calculus*]{}, Grad. Stud. Math. [**20**]{}, Amer. Math. Soc., Providence, RI, 1999. I. Hasegawa, [*A certain linear representation of the classical braid group and its application to surface braids*]{}, Math. Proc. Camb. Phil. Soc. [**141**]{} (2006), 287–301. I. Hasegawa, [*Chart descriptions of monodromy representations on oriented closed surfaces*]{}, thesis, Univ. of Tokyo, March 23, 2006. K. Hayano, [*On genus–$1$ simplified broken Lefschetz fibrations*]{}, Algebr. Geom. Topol. [**11**]{} (2011), 1267–1322. S. Kamada, [*Surfaces in $R^4$ of braid index three are ribbon*]{}, J. Knot Theory Ramifications [**1**]{} (1992), 137–160. S. Kamada, [*An observation of surface braids via chart description*]{}, J. Knot Theory Ramifications [**4**]{} (1996), 517–529. S. Kamada, [*Braid and knot theory in dimension four*]{}, Math. Surveys Monogr. [**95**]{}, Amer. Math. Soc., Providence, RI, 2002. S. Kamada, [*Graphic descriptions of monodromy representations*]{}, Topology Appl. [**154**]{} (2007), 1430–1446. S. Kamada, [*Chart description for genus-two Lefschetz fibrations and a theorem on their stabilization*]{}, Topology Appl. [**159**]{} (2012), 1041–1051. S. Kamada, Y. Matsumoto, T. Matumoto and K. Waki, [*Chart description and a new proof of the classification theorem of genus one Lefschetz fibrations*]{}, J. Math. Soc. Japan [**57**]{} (2005), 537–555. A. Kas, [*On the handlebody decomposition associated to a Lefschetz fibration*]{}, Pacific J. Math. [**89**]{} (1980), 89–104. M. Korkmaz, [*Noncomplex smooth $4$–manifolds with Lefschetz fibrations*]{}, Int. Math. Res. Not. IMRN, Vol. 2001, pp. 115–128. . M. Korkmaz, [*Lefschetz fibrations and an invariant of finitely presentable groups*]{}, Int. Math. Res. Not. IMRN, Vol. 2009, pp. 1547–1572. M. Korkmaz and B. Ozbagci, [*Minimal number of singular fibers in a Lefschetz fibration*]{}, Proc. Amer. Math. Soc. [**129**]{} (2001), 1545–1549. Y. Kuno, [*Meyer functions and the signature of fibered $4$–manifolds*]{}, arXiv:1204.1701. C. Labruère and L. Paris, [*Presentations for the punctured mapping class groups in terms of Artin groups*]{}, Algebr. Geom. Topol. [**1**]{} (2001), 73–114. F. Luo, [*A presentation of the mapping class groups*]{}, Math. Res. Lett. [**4**]{} (1997), 725–739. D. Margalit and J. McCammond, [*Geometric presentations for the pure braid group*]{}, J. Knot Theory Ramifications [**18**]{} (2009), 1–20. Y. Matsumoto, [*Diffeomorphism types of elliptic surfaces*]{}, Topology [**25**]{} (1986), 549–563. Y. Matsumoto, [*Lefschetz fibrations of genus two — A topological approach*]{}, in “Topology and Teichmüller spaces” S. Kojima et al, eds., Proc. the 37-th Taniguchi Sympo., pp. 123–148, World Scientific Publishing, River Edge, NJ, 1996. W. Meyer, [*Die Signatur von Flächenbündeln*]{}, Math. Ann. [**201**]{} (1973), 239–264. T. Nosaka, [*Bilinear-form invariants of Lefschetz fibrations over the $2$–sphere*]{}, preprint. K. Okamori, [*A genus $2$ Lefschetz fibration on an exotic $\Bbb{CP}^2\# 9\overline{\Bbb{CP}}^2$*]{}, master’s thesis (in Japanese), Osaka University, February 2011. B. Ozbagci and A. I. Stipsicz, [*Noncomplex smooth $4$–manifolds with genus-$2$ Lefschetz fibrations*]{}, Proc. Amer. Math. Soc. [**128**]{} (2000), 3125–3128. J. Park and K.-H. Yun, [*Nonisomorphic Lefschetz fibrations on knot surgery $4$–manifolds*]{}, Math. Ann. [**345**]{} (2009), 581–597. J. Park and K.-H. Yun, [*Lefschetz fibration structure on knot surgery $4$–manifolds*]{}, Michigan Math. J. [**60**]{} (2011), 525–544. V. G. Turaev, [*First symplectic Chern class and Maslov indices*]{}, J. Soviet Math. [**37**]{} (1987), 1115–1127. B. Wajnryb, [*A simple presentation for the mapping class group of an orientable surface*]{}, Israel J. Math. [**45**]{} (1983), 157–174. B. Wajnryb, [*An elementary approach to the mapping class group of a surface*]{}, Geom. Topol. [**3**]{} (1999), 405–466. C. T. C. Wall, [*Non-additivity of the signature*]{}, Invent. Math. [**7**]{} (1969), 269–274.
{ "pile_set_name": "ArXiv" }
--- abstract: | Lately, there is a growing interest in dynamic string matching problems. Specifically - the dynamic Longest Common Factor problem has been reserched and some interesting results has been reached. In this paper we examine another classic string problem in a dynamic setting - finding the longest palindrome substring of a given string. We show that the longest palindrome can be maintained in polylogarithmic time per symbol edit. author: - 'Amihood Amir [^1]' - 'Itai Boneh[^2]' bibliography: - 'paper.bib' title: 'Dynamic Palindrome Detection [^3]' --- Introduction {#s:int} ============ Palindrome recognition is one of the fundamental problems in computer science. It is among the first problems assigned in a programming course, and it reigns at tests and assignments for the automata and language courses, since it is a good example of a context-free language that is non-regular. It visits complexity courses as an example of a problem that is solved in linear time by a two-tape Turing machine [@slis:73] but requires quadratic time in a single-tape machine [@maass:1984]. Seeking all palindromes in a string is also a good example for usages of subword trees. Apostolico, Breslauer and Galil [@abg:92] considered parallel algorithms for the problem. Manacher [@man:75] and Galil [@g-75] showed how to use DPDAs for recognizing palindrome prefixes of a string that is input online. Amir and Porat [@ap:cpm14] showed how to recognize [*approximate*]{} palindrome prefixes of a string that is being input online. In addition to its myriad theoretical virtues, the palindrome also plays an important role in nature. Because the DNA is double stranded, its base pair representation offers palindromes in hairpin structures, for example. Many restriction enzymes recognize and cut specific palindromic sequences. In addition palindromic sequences play roles in methyl group attachments and in T cell receptors. For some examples of the varied roles of palindromes in Biology see, e.g. [@gk:97; @f:03; @lssnz:05; @sr:12]. Due to the importance, both theoretical and practical, of palindromes, it is surprising that the problem of finding palindromes in a dynamic text has not been studied. Clearly, one can re-run a palindrome detection algorithm after every change in the text, but this is obviously a very inefficient way of handling the problem. In the 1990’s the active field of dynamic graph algorithms was started, with the motive of answering questions on graphs that dynamically change over time. For an overview see [@degi:2010]. Recently, there has been a growing interest in dynamic pattern matching. This natural interest grew from the fact that the biggest digital library in the world - the web - is constantly changing, as well as from the fact that other big digital libraries - genomes and astrophysical data, are also subject to change through mutation and time, respectively. Historically, some dynamic string matching algorithms had been developed. Amir and Farach [@AF-FOCS-91] introduced dynamic dictionary matching, which was later improved by Amir et al. [@AFILS-92-journal]. Idury and Scheffer [@Idury-Schaeffer-92:cpm] designed an automaton-based dynamic dictionary algorithm. Gu et al. [@Gu:F:Beigel:94] and Sahinalp and Vishkin [@Sah:Vish:96] developed a dynamic indexing algorithm, where a dynamic text is indexed. Amir et al. [@ALLS07] showed a pattern matching algorithm where the text is dynamic and the pattern is static. The last few years saw a resurgence of interest in dynamic string matching. In 2017 a theory began to develop with its nascent set of tools. Bille et al. [@Bille2017] investigated dynamic relative compression and dynamic partial sums. Amir et al. [@acipr:17] considered the longest common factor (LCF) problem. They investigated the case after one error. Special cases of the dynamic LCF problem were discussed by Amir and Boneh [@ab:18]. The fully dynamic LCF problem was tackled by Amir et al. [@acpr:18]. Amir and Kondratovsky [@ak:18] made a first step toward a fully dynamic string matching algorithm by considering a dynamic pattern and text that is changing in a limited fashion. In this paper we consider the problem of finding the longest palindrome in a dynamic string. The changes to the string are character replacements. The contributions of this paper are: 1. We present a deterministic algorithm for computing the longest palindrome in a dynamic text in time $\tilde{O}(1)$ per substitution. 2. We reinforce the dynamic LCP as an important tool for dynamic string matching algorithms. 3. We prove some novel combinatorial properties of palindromes and periodic palindromes. This deeper understanding of the nature of palindromes enables the efficient dynamic longest palindrome detection algorithm. This paper is organized as follows. Section \[s:pre\] gives the basic pattern matching definitions and tools and can be safely skipped by the practitioner. Section \[s:dlcp\] summarizes the known tecniques for dynamic LCP. Section \[s:dlps\] gives the dynamic algorithm for finding the longest palindrome in a changing sequence. We conclude with some open problems and future directions. Preliminaries {#s:pre} ============= We begin with basic definitions and notation generally following [@AlgorithmsOnStrings]. Let $S=S[1]S[2]\ldots S[n]$ be a *string* of length $|S|=n$ over a finite ordered alphabet $\Sigma$ of size $|\Sigma|=\sigma= O(1)$. By $\varepsilon$ we denote an exmpty string. For two positions $i$ and $j$ on $S$, we denote by $S[i.. j]=S[i].. S[j]$ the *factor* (sometimes called *substring*) of $S$ that starts at position $i$ and ends at position $j$ (it equals $\varepsilon$ if $j<i$). We recall that a [*prefix*]{} of $S$ is a factor that starts at position $1$ ($S[1.. j]$) and a [*suffix*]{} is a factor that ends at position $n$ ($S[i..n]$). We denote the [*reverse*]{} string of $S$ by $S^R$, i.e. $S^R=S[n]S[n-1]\ldots S[1]$. We say that string $S$ is a [*palindrome*]{} if $S=S^R$. Let $S$ be a string, $Y$ a factor of $S$. We say that $Y$ is a [*palindromic factor*]{} if $Y$ is a palindrome. $Y$ is a [*longest palindromic factor*]{} if there is no palindromic factor $F$ of $S$ where $|F|>|Y|$. Given two strings $S$ and $T$, the string $Y$ that is a prefix of both is the [*longest common prefix (LCP)*]{} of $S$ and $T$ if there is no longer prefix of $T$ that is also a prefix of $S$. Let $Y$ be a string of length $m$ with $0<m\leq n$. We say that there exists an *occurrence* of $Y$ in $S$, or, more simply, that $Y$ *occurs in* $S$, when $Y$ is a factor of $S$. Every occurrence of $Y$ can be characterised by a starting position in $S$. Thus we say that $Y$ occurs at the *starting position* $i$ in $S$ when $Y=S[i .. i + m - 1]$.\ We say that string $S$ of size $n$ has a *period* $p$ if for every $i$ such that $1 \ le i \le n - p$ , it’s satisfied that $S[i] = S[i+p]$ for some $1 \le p \le \frac{n}{2}$. The *period* of $S$ is the minimal $p$ for which that condition holds.\ We say that a substring of $S$,denoted as $A = S[a..b]$ is a *run* with period $p$ if it’s period is $p$, but $S[a-1] \neq S[a - 1 + p]$ and $S[b+1] \neq S[b+1 - p]$. Meaning that every substring containing $A$ doesn’t have a period $p$. Suffix tree and suffix array. ----------------------------- The *suffix tree* $\mathcal{T}(S)$ of a non-empty string $S$ of length $n$ is a compact trie representing all suffixes of $S$. The branching nodes of the trie as well as the terminal nodes, that correspond to suffixes of $S$, become [*explicit*]{} nodes of the suffix tree, while the other nodes are [*implicit*]{}. Each edge of the suffix tree can be viewed as an upward maximal path of implicit nodes starting with an explicit node. Moreover, each node belongs to a unique path of that kind. Thus, each node of the trie can be represented in the suffix tree by the edge it belongs to and an index within the corresponding path. We let $\mathcal{L}(v)$ denote the *path-label* of a node $v$, i.e., the concatenation of the edge labels along the path from the root to $v$. We say that $v$ is path-labelled $\mathcal{L}(v)$. Additionally, $\mathcal{D}(v)= |\mathcal{L}(v)|$ is used to denote the *string-depth* of node $v$. Node $v$ is a *terminal* node if its path-label is a suffix of $S$, that is, $\mathcal{L}(v) = S[i .. n]$ for some $1 \leq i \leq n$; here $v$ is also labelled with index $i$. It should be clear that each factor of $S$ is uniquely represented by either an explicit or an implicit node of $\mathcal{T}(S)$, called its *locus*. In standard suffix tree implementations, we assume that each node of the suffix tree is able to access its parent. Once $\mathcal{T}(S)$ is constructed, it can be traversed in a depth-first manner to compute the string-depth $\mathcal{D}(v)$ for each node $v$. It is known that the suffix tree of a string of length $n$, over a fixed-sized ordered alphabet, can be computed in time and space $O(n)$ [@AlgorithmsOnStrings]. The suffix array of a string $S$, denoted as $SA(S)$, is an integer array of size $n+1$ storing the starting positions of all (lexicographically) sorted non-empty suffixes of $S$, i.e. for all $1 < r \le n+1$ we have $S[SA(S)[r-1] .. n] < S[SA(S)[r] .. n]$. Note that we explicitly add the empty suffix to the array. The suffix array of $S$ corresponds to a pre-order traversal of all the leaves of the suffix tree of $S$. The inverse $iSA(S)$ of the array $SA(S)$ is defined by $iSA(S)[SA(S)[r]] = r$, for all $1 \le r \le n+1$. The Karp-Rabin Algorithm ------------------------ Karp and Rabin developed a randomized linear time algorithm for finding all occurrences of a pattern in a text (pattern matching) [@KR-87]. The main idea of their algorithm is computing a numeric signature of the pattern, then sliding the pattern over the text and comparing the signature of the text substring that is tested against the pattern, to the pattern signature. Any signature that is updated in constant time per shift is a good candidate. Such a signature is also called a [*rolling hash function*]{}. For example, assume the alphabet is $\{1,...,k\}$. The hash of a substring of length $n$ would be the representation of the substring as a number is base $p$ taken modulo $q$, for some prime numbers $p$ and $q$. Clearly the computation of a hash in a single shift can be done in constant time, and a hash equality implies a substring equality with high probability. Dynamic Longest Common Prefix queries {#s:dlcp} ===================================== #### Definition and implementation qualities [*Dynamic Longest Common Prefix*]{} queries are a fundamental and powerful tool for maintaining properties of a dynamic string. [*\[The Dynamic LCP problem\]*]{} Let $D$ be a text string over alphabet $\Sigma$., A [*Dynamic Longest Common Prefix (LCP) algorithm*]{} supports two queries: 1. $LCP(i,j)$ - Return the longest common prefix of $D[i..n]$ and $D[j..n]$. 2. $Update(i,\sigma)$ - Change the symbol in $D[i]$ to be $\sigma$. The quality of an implementation for D-LCP can be measured by various parameters: 1) update time, 2) time for LCP query on the current text, and 3) whether the algorithm is deterministic or randomized. Note that since Static LCP can be done with linear time preprocessing and constant time query , any solution in which the update time is not sublinear will not be better than doing the static LCP prerocessing from scratch after every update. The Deterministic Implementation -------------------------------- There are a number of algorithms that yield a polylogarithmic computation of an LCP query on a dynamic text, following a polylogarithmic processing per change. We mention Mehlhorn et al. [@msu:94]. Using their algorithm with appropriate deamortization, one can compute the LCP in time $\lambda(n) \in O(\log^3 n \log^* n)$, and $O(\log^2 n \log^* n)$ per text change. Randomized Implementation ------------------------- It is a folklore fact that dynamic LCP can be achieved via Rabin-Karp methods. In this case, the LCP of two indices can be computed in time $\lambda(n) \in O(\log(n))$ with high probability. Dynamic Longest Palindrome Substring {#s:dlps} ==================================== The Algorithm’s Idea -------------------- The goal is to maintain a data structure containing all the maximal palindromes. [*Maximal*]{} in this context means that the palindrome can not be expanded around its center. The longest palindrome substring is obviously a maximal palindrome, so as long as we keep track of the maximal palindromes in the text - we have the longest palindrome substring as well. Given an index in $T$ , We can find the maximal palindrome centered in this index using a single LCP query on $T\$T^{R}$.\ After a text update - some maximal palindromes may be cut and some may be extended. We should query the relevant centers for the updated sizes of the affected maximal palindromes.\ In the worst case - a single update can affect $O(n)$ maximal palindromes. So checking every single affected palindrome will not result in sublinear time. We make several observations on maximal palindromes that allow us to reduce the amount of contested palindromes to $O({\rm polylog}(n))$. Locally Maximal Palindromes --------------------------- Let $D$ be a text. A [*Locally Maximal Palindrome*]{} of $D$ is defined to be a substring $D[i..j]$ so that $D[i..j]$ is a palindrome and $D[i-1] \ne D[j + 1]$. Meaning that the palindrome can not be extended to the sides from its center.\ The first observation we make about locally maximal palindromes gives an upper bound on the amount of similar sized maximal palindromes within a given distance from each other’s starting points. For this purpose, we pick some constant $\epsilon > 0$. We denote $a = 1 + \epsilon$. We partition the maximal palindromes of some text $D$ to $O(\log(n))$ classes. Class $i$ contains the palindromes whose size is $s$ such that $a^{i} \le s < a^{i+1}$. Let $p_{1} = D[s_{1}..e_{1}]$ and $p_{2} = D[s_{2}..e_{2}]$ be two maximal palindromes in class $i$ that also satisfy $d=|s_{1} - s_{2}| < \epsilon a^{i}$ and neither of them contains the other. Assume w.l.o.g that $s_{1} < s_{2}$. Then $D[s_{1} .. e_{2}]$ has a period. [**Proof:**]{} Consider the overlap between $p_{1}$ and $p_{2}$ : $p_{o} = D[s_{2} .. e_{1}]$. Since it is contained in $p_{1}$ , which is a palindrome, its reverse appears in the symmetric place in $p_{1}$. So we have $p_{o}^{R} = D[s_{1} .. e_{1} - s_{2}]$. Symmetrically, the reverse of the overlap should also appear in $p_{2}$. So $p_{o}^{R} = D[s_{2} + e_{2} - e_{1} .. e_{2}]$. We, therefore, have two instances of the same string, $p_{o}^{R}$, starting in two different indices in the text. If the difference between the indices is smaller than half the size of the substring - then this substring has a period. The size of the overlap is $|p_{0}^{R}| = |p_{1}| - d$ since $d$ is the chunk of $p_{1}$ that is not participating in the overlap. $|p_{1}|$ is at least $a^{i}$ since $p_{1}$ belongs to class $i$, and $d<\epsilon a^{i}$. So $|p_{o}^{R}| \ge a^{i} - \epsilon a^{i}$. The difference between the starting indices is $s_{2} - s_{1} + e_{2} - e_{1}$. We already have a bound for $s_{2} - s_{1}$. As for $e_{2} - e_{1}$, note that $e_{i} = s_{i} + |p_{i}|$. Therefore the difference between the ending indices equals the difference between the starting indices plus the difference between the palindromes’ sizes. The difference between the sizes is bounded by $a^{i+1} - a^{i} = \epsilon a^{i}$ so overall we have $e_{2} - e_{1} \le 2\epsilon a^{i}$ for the difference between the ending indicies. It is also the case that $ \epsilon a_{i} + 2\epsilon a_{i} = 3\epsilon a^{i}$ for the difference between the two instances of $p_{o}^{R}$. For a period, we need $3\epsilon a^{i} \le \frac{a^{i} - \epsilon a^{i}}{2}$, which is satisfied for $\epsilon \le \frac{1}{7}$. We fix $\epsilon = \frac{1}{7}$ from now on.\ \ The main implication of the above theorem is the following: \[l:2\] At most two locally maximal palindromes in class $i$ can start in an interval of size $\epsilon a^{i}$ (unless one of them is contained within the other). [**Proof:**]{} Consider 3 maximal palindromes $p_{1}$, $p_{2}$, and $p_{3}$, ordered by the starting index, such that $s_{3} - s_{1} < \epsilon a^{i}$ and none of them is contained within the other. According to the previous lemma, The entire interval containing $p_{1}$ and $p_{3}$ is a periodic string with period size $x$. Note that according to our proof, for an appropriate choice of $\epsilon$ we get that $x \le 3\epsilon a_{i} \le a_{i}$. So $x$ is smaller than any palindrome in class $i$ , including $p_{2} = D[i..j]$ which is fully contained in this periodic interval. From periodicity we have $D[i - 1] = D[i - 1 + x]$ and $D[j + 1] = D[j + 1 - x]$ . From the fact that $D[i..j]$ is a palindrome we have $D[i - 1 + x] = D[j + 1 - x]$ ,as those are symmetric indicies in the palindrome (they must be included in $p_{2}$ since $|p_{2}| > x$). Transitivity now yields $D[i - 1] = D[j + 1]$ in contradiction to $p_{2}$’s maximality. \ \ This is the key observation for lowering the amount of necessary LCP queries.\ \ Maintaining all the LMPs ------------------------ The data structure we use for maintaining all the LMPs consists of $O(\log(n))$ priority queues. $Q_{i}$ contains the LMP with size $s$ such that $a^{i} \le s < a^{i+1}$, i.e. the $i$’th class palindromes. The values kept in $Q_{i}$ are the start and end indicies of each palindrome, sorted by the value of the starting index. We also maintain extra data about the maximum size of a palindrome within $Q_{i}$. When an index is changed, we need to update every LMP that touches that index. We first observe a [*simplified case*]{} in which $Q_{i}$ does not contain palindromes that are fully contained in another palindrome.\ \ Given an update in index $x$ - we need to update all the palindromes that touched $x$. Palindromes that fully extend $x$ are cut in index $x$ (and in the symmetrical index as well), as the equality was destroyed. Palindromes that end just before $x$ may have been extended. Their new value is checked using LCP queries.\ We start by checking all the palindromes starting in $x + 1$ for extension. Since we are assuming that $Q_{i}$ does not contain palindromes that are fully contained in each other, we have at most one palindrome starting in $x+1$ in every $Q_{i}$. Now, we want to find all the other palindromes that are affected by the update. We do it by considering exponentially growing intervals of distances of the palindromes’ starting index from $x$. At step $i$, we look at all the palindromes that start within distance $d$ from $x$ such that $ a^{i} \le d < a^{i+1}$. Note that the palindromes that start in the said distances from $x$ must be at least in class $i$ (otherwise - they will not reach all the way to $x$ from that distance). Also, the size of this distances interval is $a^{i+1} - a^{i} = \epsilon a^{i}$. Our lemma directly implies that in step $i$, every size class with index larger than $i$ consists of no more than 2 palindromes in the contested interval. So, for every value of $d$, we have to inspect each priority queue a constant number of times.\ \ [**To conclude:**]{} There are $O(\log(n))$ priority queues and each of them is queried a constant number of times for every exponential interval. There are $O(\log(n))$ such intervals so the time complexity for this simplified case is $O(\log^{2}(n)(\log(n) + \lambda(n)))$, where $\lambda(n)$ is the LCP query time.\ \ Sadly, our simplification is far for being true. Palindromes of the same size class can be included in each other in great quantities. For example, consider the text $D = ab^{n}a$. The whole text is a palindrome. and every single index is the beginning of a LMP that is contained in the LMP starting in the previous index. We need to enhance both our data structure and understanding of locally maximal palindromes to deal with these cases. Central Periodic Palindromes ---------------------------- The example we presented for many palindromes of similar sizes that are contained in each other actually demonstrates the structure of palindromes of that type. The following Theorem is the key to handling those palindromes: \[t:1\] Let $p_{b} = D[s_{b}..e_{b}]$ and $p_{s} = D[s_{s}..e_{s}]$ be two LCPs in size class $i$ . If $p_{s}$ is contained in $p_{b}$ then $D[s_{s} .. e_{b} - s_{s} +s_{b}]$ has a period of size $|p_{b}| - |p_{s}| - 2d$, where $d$ is the difference between the starting indices. Note that the period is at most $\epsilon a^{i}$. \ **Proof:** Since $p_{s}$ is a substring of $p_{b}$ and is also a LMP, it can not share a center with $p_{b}$. Therefore its reverse appears in the symmetrical indices in $p_{b}$. But since $p_{s}$ is also a palindrome then it is equal to its reverse. Therefore, we have two instances of $p_{s}$. Because of our choice of $\epsilon$ and because $p_s$ and $p_b$ are in the same size class, then $p_s$ is periodic and the size of its period is the difference between the starting indices of the instances of $p_{s}$.\ \ Note that the formula only works if the initial $D[s_{s}..e_{s}]$ is the left side instance of $p_{s}$ with respect to the center of $p_{b}$. If we are given the right side instance - we can calculate the left side instance and proceed to apply the formula.\ \ We call the periodic palindromes that is created as a result of a LMP that is contained within another LMP, $P$, in the same size class the [*Central Periodic Palindrome*]{} of $P$, or the [*CPP*]{} of $P$. We call the period of a CPP the [*periodic seed*]{} of the CPP. We call a maximal run of the periodic seed a **periodic palindromes cluster**. We point out two important substrings of a periodic palindromes cluster: - **The maximal palindrome prefix:** The longest prefix of the cluster that is a palindrome. - **The maximal palindrome suffix:** The logest suffix of the cluster that is a palindrome. It is possible for a CPP to be both a prefix CPP and a suffix CPP. We call a maximal run of the periodic seed a [*periodic palindromes cluster*]{}. Note that for the maximal palindrome prefix, all the prefixes of size $i*p + r$ are LMPs (with $p$ and $r$ being the period and the remainder of the largest palindrome prefix, respectively). The same applies to the maximal palindrome suffix with suffixes of the same sizes. The mentioned LMPs are [*represented*]{} by the cluster. Meaning that if we know the starting and ending position of $C$, its periodic seed, its maximal palindrome prefix and its maximal palindrome suffix then the existence of all those LMPs is implied.\ The periodic palindrome clusters and their components are our key ingredient for efficient palindrome detection. They have several properties that make them comfortable to work with. For example: [*all the LMPs that are contained in a cluster are either represented by the cluster or smaller than twice the size of the periodic seed.*]{} More formally: \[l:3\] Let $C[1..n]$ be a periodic palindromes cluster with period $|A| = p$. Let $MPP=A^{k}A'$ be the maximal palindrome prefix of $C$ with a remainder $r_{p} = |A'|$, and Let $MPS=B^{k}B'$ be the maximal palindrome suffix of $C$ with a remainder $r_{s} = |B'|$ . A substring $P$ of $C$ with $|P| \ge 2p$ is a locally maximal palindrome only if $P \in D = \{C[1.. i \cdot p + r_{p}] | i \in \{0.. k\}\} \cup \{C[ n -(i \cdot p + r_{s})..n] | i \in \{0..k\}\}$. We call $D$ the set of LMPs that are represented by $C$. **Proof:** First, we show that every element in $D$ is a palindrome. From periodicity, we have $C[1 .. i \cdot p + r_{p}] = C[(k - i) \cdot p ..k \cdot p + r_{p}]$ for $i \in \{0.. k\}$. From $MPP$’s symmetry as a palindrome we have $C[1 .. i \cdot p + r_{p}] ^{R} = C[(k - i) \cdot p ..k \cdot p + r_{p}]$ for $i \in \{0.. k\}$. Transitivity now yieldws that $C[1 .. i \cdot p + r_{p}] = C[1 .. i \cdot p + r_{p}]^{R}$, which makes this interval a palindrome. Symmetrical arguments can be made to show that $C[ n -(i \cdot p + r_{s})..n]$ is a palindrome too.\ Now, Let $P = C[i..j]$ be a LMP within $C$ with $|P| \ge 2p$ such that $P$ is not in $D$. If $i \neq 1$ and $j \neq n$, then $P$ can be extended around its center due to similar arguments as in the proof of lemma \[l:2\]. Otherwise, We can assume that $i=1$ (The proof for the case where $j=n$ is symmetrical). Let $k^{*}$ be the minimal value of $k$ such that $k \cdot p + r_{p} > j$ and $s^{*}=k^{*} \cdot p + r$. Note that $s^{*} \ge 2p$ and $s^{*}-j< p$. Since $C[1..s^{*}]$ is a palindrome containing the palindrome $P$ , $P$ appears in the symmetrical place in $C[1..s^{*}]$. The difference between the starting indices of these two instances will be $s^{*}-j$ , which is smaller than $p$. That yields a period smaller than $p$ for the prefix of size $2p$ of $C$ , which indicates that $C$ has a period smaller than $p$.\ \ The above lemma implies that if we have a cluster $C$ in class size $i$ in our data structure, and we find a LMP $P$ in the same class that is contained in $C$ - we do not need to explicitly store it. $P$ can not be smaller than twice the size of the period. So according to the lemma it is implicitly represented by $C$.\ The contained LCPs that are smaller than twice the size of the period can be handled within smaller exponential size classes.\ Two other important properties of the periodic palindromes clusters are: - Two clusters within the same exponential size class **can not** be contained within each other. - Two clusters in size class $i$ **can not** have starting indices with distance smaller than $\epsilon a^{i}$ from each other. These properties can be proved by observing that if two clusters violate any of them - The run of one of the clusters can be extended. Extension and cuts of CPPs -------------------------- Our algorithm represents LMPs under substitutions using periodic palindromes clusters. We, therefore, need to understand how clusters act under substitution. We wish to maintain every cluster along with its periodic seed and its maximal palindrome prefix and suffix.\ \ We start by examining the case in which a cluster is cut by a substitution in index $x$. For a clearer exposition, we denote the cut cluster to be $C[1..c]$ (rather than representing $C$ as some substring $S[i..j]$). Let its period be $p$ and the remainders of the contained prefix CPP and the contained suffix CPP be $r_{p}$ and $r_{s}$ respectively. The substitution splits $C$ into two periodic palindromes clusters: $C[1..x-1]$ and $C[x + 1.. c]$. We show how the implied LMPs that are centered in the left side of $x$ are affected and deduce the resulting maximal palindrome prefix and maximal palindrome suffix of the cluster $C[1..x-1]$. Symmetrical arguments can be made for the LMPs centered in the right side of $x$ and $C[x+1 .. c]$.\ The LMPs consistent with the period that are centered in the left side of $x$ can be sorted into two groups: 1. **LMPs that are not touched by $x$ :** The implied LMPs can be either prefixes of the maximal palindrome prefix or suffixes of the maximal palindrome suffix. Since we are considering LMPs with centers in the left side of $x$ that were not touched by $x$, these can only be prefixes. Those are $C[1.. k \cdot p + r_{p}]$ for every $k$ such that $ k \cdot p + r_{p} \le x - 1$. We observe that the largest LMP in this set is $C_{1} = C[1..k^{*} \cdot p + r_{p}]$ with $k^{*}$ being the maximal $k$ that satisfies the previous constraint. We point out that $C_{1}$ contains all the other LMPs in that set. Assuming that $C[1..x-1] > 2\cdot p$, there is no palindrome prefix larger than $C_{1}$ according to Lemma \[l:3\]. So $C_{1}$ is the maximal palindrome prefix of $C[1.. x- 1]$. The assumption that $C[1..x-1] > 2\cdot p$ implies that it is still periodic. If not - we won’t keep $C[1..x-1]$ as a cluster, But all the LMPs that are represented by it instead. Since in this case, the cluster is not periodic - the amount of represented LMPs is bounded by a constant factor. 2. **LMPs that are touched by $x$:** In this case we may consider both LMPs that are prefixes of $C$ and LMPs that are suffixes of $C$.\ [*The relevant prefixes*]{} are $C[1..k \cdot p + r_{p}]$ for every $k$ such that $ k \cdot p + r_{p} \ge x $ and $\frac{k\cdot p + r_{p}}{2} < x$. The first constraint implies that the prefix is indeed touched by $x$, and the second constraint implies the location of the center. The smallest $k$ that satisfies these two constraints will yield the represented prefix LMP that extends farthest to the left after the change. Denote this pivot value of $k$ as $k'$, and Denote $r' = (k'\cdot p + r_{p}) - (x-1)$. $r'$ is the size of the suffix that was cut from the pivotal LMP. A prefix of the same size should be removed, so the new largest LMP that touches $x$ from the left will be $C[r'..x-1]$ (among the LMPs that are prefixes of the original maximal palindrome prefix).\ **The relevant suffixes:** Actually, the only possible candidate to extend the farthest to the left after a substitution in index $x$ is the maximal palindrome suffix denoted as $C[s..c]$, since its center is the farthest to the left from all its suffixes. If it is indeed cut by $x$ and its center is in the left side of $x$, the resulting LMP after the substitution will be $C[s + c -x + 1..x-1]$.\ Out of these two candidates, the one that extends farthest to the right will be the maximal palindrome suffix of $C[1..x-1]$. All of the above can be calculated in constant time given $C,x,p,r_{p}$ and $r_{s}$.\ \ We now analyse periodic palindromes clusters that are touched by $x$ in their ends:\ Again, for ease of exposition, we denote the cluster as $C[1..c]$. We assume that $x = c + 1$. The case where $x$ touches the left side of $C$ is treated symmetrically. Let the periodic seed of $C$ be $p$. Denote the remainder of $C$’s prefix CPP as $r_{p}$. Finally, denote $c' = LCP(1, 1+ p)$. The LCP query in the last notation indicates the maximal extension of the run. Therefore the new updated interval for the cluster is $C[1..c']$. It can be proven by induction that given a prefix CPP $S=A^{l}A'$ with period $p = |A|$, the string $A^{k}A'$ is a palindrome for every $k\geq 0$. If we take the maximal $k^{*}$ such that $k \cdot p + r_{p} \le c'$ , we get the maximal palindrome prefix of $C[1..c']$. Denote this prefix as $C_{p} = [1..k^{*} \cdot p + r_{p}] = [1..s]$.\ As for the maximal palindrome suffix - consider the suffix of size $s-p$ of $C_{p}$ $C' = C[p+1 .. s]$. According to the previous claim this is a palindrome but it is not necessarily a LMP. Since it is within the periodic cluster, it may be extended around the center as long as the resulting extension is within the cluster. $C'$ can be extended by $c-s$ to the right and $p$ to the left. Since $p > c-s$ , this extension will result in the LMP $C_{s} = C[p+1+s-c .. c']$. $C_{s}$ is the maximal palindrome suffix cluster $C = [1..c']$.\ \ All of the above can be calculated in constant time given $C,x,p,r_{r}$ and $c'$. $c'$ can be calculated using a dynamic LCP data structure. Adding CPPs to the maximal palindrome algorithm ----------------------------------------------- We enhance our algorithm to maintain $CPP$ collections i addition to the LMP collections. The maintained invariant in this setting is that every LMP in the text is either represented explicitly as an LMP or implicitly as a part of a cluster. In addition to the $Q_{i}$ priority queues, we also define $CPP_{i}$. $CPP_{i}$ contains periodic palindromes clusters in the $i$’th exponential size class. Every cluster is stored along with its corresponding period , prefix CPP and suffix CPP. As in $Q_{i}$, they are sorted by increasing value of the starting index. We maintain the invariant that each of those priority queues does not contain any element that is contained in another element in the queue. In $CPP_{i}$ this is naturally preserved as long as we maintain valid clusters due to properties of clusters. Preserving this condition in $Q_{i}$ requires more sensitive care.\ Given a substitution in index $x$, $x$ needs to be tested against every $Q_{i}$ and $CPP_{i}$ in every exponential distance level. First, we extract all the affected LMPs and clusters from the priority queues. We treat every LMP we extracted from some $Q_{i}$ as in the simplified case - we cut it in a symmetrical manner if it was cut by $x$ and check for extension if $x$ touches one of its ends. We save the results of those extensions and cuts in some temporary list $L$. As for the extracted $CPP$s, we treat them as described in the previous section. Additionally, for every cluster created as a product of an extension or a cut in the process, we query the new center of the cluster for the LMP centered in this index. This is due to the fact that it is [*the only candidate among the LMPs represented by the cluster to be extendible beyond the cluster’s range*]{}. We add the results of these queries to $L$ as well. We also add the updated cluster to $CPP_{i}$ with respect to its size. If a cluster was cut to the point when it is no longer periodic then we do not add it to $CPP_{i}$. Instead, we add all the LMPs that are implied by the cut CPP to $L$. Since the cut CPP is no longer periodic then the amount of implied LMPs is bounded by a constant.\ \ Note that the size of $L$ is at most $O(\log^{2}(n))$, since we add an element to $L$ only once an element that is cutting $x$ is met in either $Q_{i}$ or $CPP_{i}$. This happens a constant number of times in every exponential distance level. So we have $O(\log(n))$ queues multiplied by $O(\log(n))$ exponential distance levels.\ \ At this stage, every LMP in the new text is represented either in $Q_{i},CPP_{i}$ or in $L$. The next natural step would be adding every LMP saved in $L$ to the appropriate $Q_{i}$. But this may violate our invariant that $Q_{i}$ does not contain two elements such that one of them is containing the other. We handle it by deducing the existence of a cluster and adding the cluster to our data structure instead of the contained elements. This is done as follows:\ For every $P =[i..j] \in L$, we query $Q_{i}$ that matches $P$’s size for the predecessor of $i$. We denote the returned LMP as $P'$.\ \[l:4\] If there is an element in $Q_{i}$ that contains $P$, $P'$ must contain $P$. additionally, there are at most two elements in $Q_{i}$ that contains $P$ **Proof:** $P'=[i'..j']$ has the largest value of $i'$ that is less than $i$, so every successor of $P'$ can not contain $P$. Assume that $P'$ does not contain $P$. That necessarily means that $j' < j$. Assume that there is another element in $Q_{i}$ that contains $P$. Denoted it as $P^{*}$. As seen, $P^{*}=[i^{*}..j^{*}]$ must be a predecessor of $P'$, so $i^{*} < i'$. $P^{*}$ also contains $P$ Therefore $j<j^{*}$. Transitivity yields that $j'<j^{*}$ and $P^{*}$ contains $P'$, In contradiction to $Q_{i}$ not containing two elements that contain each other.\ As for the existence of no more than two including elements, $P$ belongs to the $i$’th size class, meaning that its size is at least $a^{i}$. If we have three LMPs in the $i$th class that share an interval of size $a^{i}$, the distance between their starting indices won’t be more than $\epsilon a^{i}$. Lemma \[l:2\] suggests that this is impossible.\ \ Given lemma \[l:4\] we can find $P'$ and check if it contains $P$. If it does then we calculate the $CPP$ derived from $P$ and $P'$ and a period of this CPP using the formula provided in theorem \[t:1\], and check for the extension of the cluster in both directions using LCP queries. We proceed to ignore $P$ and add the new cluster to the appropriate $CPP_{i}$.\ \ Another violation that may result in adding $P$ to $Q_{i}$ is that $P$ contains elements in $Q_{i}$. This case can be tested in a similar way. Query $Q_{i}$ for the successor of $i$ and denote it as $P'$. As in lemma \[l:4\], it can be proven that if any element in $Q_{i}$ is contained within $P$ then $P'$ most be contained within $P$. Also, there are at most two elements that are contained within $P$ in $Q_{i}$. We can add $P$, remove $P'$ (and its successor, if necessary), and proceed to calculate the cluster as in the previous case.\ \ It may seem like we are done at this point, but there is another subtle detail that we need to take care of. Since the period of a periodic cluster is required in order to calculate the resulting cluster after an extension or a cut, The periods of the new clusters most be calculated. For the clusters created as a result of an extension or a cut this is done as described in the previous section. For clusters that were created to solve $Q_{i}$ violations we need to work harder. Note that the formula in theorem \[t:1\] yields [*some*]{} period of the CPP (or cluster), which is valid for checking the extension, but is not necessarily the smallest period. We present a subroutine *FindCppPeriod* to solve that problem. *FindCppPeriod* is meant to be used when all the LMPs are either represented explicitly or implicitly. Therefore running it after we fully evaluated $L$ will be sufficient. *FindCppPeriod* also assumes that we know the starting and ending indices of the $CPP$ of $P$. This assumption is valid because*FindCppPeriod* is called only after two LMPs within the same size class with one of them containing the other are met. In this settings, theorem \[t:1\] can be used to calculate the periodic interval that defines the $CPP$.\ \ Denote the containing LMP $P = S[i..j]$, and its size class as $s$. The following lemma is the key to finding the period:  \[l:5\] 1. The CPP of $P$ is unique in the sense that every LMP $P'$ in class size $s$ will yield the same CPP using the formula from theorem \[t:1\] on $P$ and $P'$. 2. Let $C=P[a..b]$ be the $CPP$ of $P$. There is a LMP $L$ in the size class $s$ that is contained in $P$ such that if we apply the formula from theorem \[t:1\] on $P$ and $L$ then we will get the minimal period of the $CPP$. **Proof:** For the first part of the lemma, Assume that we have two different LMPs, $P[i_{1}..j_{1}]$ and $P_{2}=P[i_{2}..j_{2}]$, and the CPPs derived from their existence are $C_{1}$ and $C_{2}$ respectively. We can assume, without loss of generality, that both $P[i_{1}..j_{1}]$ and $P[i_{2}..j_{2}]$ are the left side instances with respect to the center of $P$ of $P_{1}$ and $P_{2}$, respectively. If $i_{1}=i_{2}$ then the resulting CPPs would be the same since the formula from theorem \[t:1\] depends only on the starting index of the inner LMP. Otherwise assume, w.l.o.g., that $i_{1} < i_{2}$. According to the formula, that would mean that $C_{1}$ fully contains $C_{2}$. Specifically it fully contains $P_{2}$. So we have that $P_{2}$ is a LMP in size class $s$ that is contained in the CPP $C_{1}$ but is not consistent with the period of $C_{1}$ (since it is neither a prefix or a suffix of $C_{1}$). That is a contradiction to Lemma \[l:3\].\ \ For the second part of the lemma, consider the unique CPP $C=P[a..b]$ with a periodic seed $p$. The existence of $C$ is a result of some LMP $P'=[a'..b']$ in size class $s$ that is contained within $P$. $P'$ is either a proper prefix or a proper suffix of $C$. We will prove the lemma assuming that it is a proper prefix of $C$. The proof for the case in which $P'$ is a proper suffix of $C$ is symmetrical. Since $P'$ is a proper suffix of $C$, and an LMP, the period of $C$ can not be extended to the left. Otherwise, $P'$ would have been extendable around its center. So $a$ is the beginning of a prefix $CPP$ with period $p$ and remainder $r$, and every substring of the form $P[a..a+ k \cdot p + r]$ is a LMP. Let $k^{*}$ be the maximal $k$ such that $a + k\cdot p + r < b$. Let $P^{*} = S[a..a + k^{*} \cdot p + r]$. $P^{*}$ is a LMP contained in $P$. Since $P'$ is of the said form too, $P^{*}$ is at least as long as $P'$, implying that it is in class size $s$. Since $k^{*}$ is the maximal value of $k$ that satisfies $a + k\cdot p + r < b$, we have $b - (a + k\cdot p + r) \le p$. The size of $P$ is $b-a$, the size of $P^{*}$ is $k^{*} \cdot p + r$ and the distance between the starting locations is $a'-a$. Applying the formula from theorem \[t:1\], we get $(b - a) - (k^{*} \cdot p + r) - 2(a' - a) = b - (a + k\cdot p + r) -2(a'-a) \le p -2(a'-a) \le p$. If the formula yields a number smaller than $p$, it will be a contradiction to $p$ being the minimal period. So it most yield $p$.\ \ To conclude: in order to compute the periodic seed of the cluster, we need to find that $P^{*}$ in our collection and apply theorem \[t:1\]. Thus the subroutine will work as follows:\ \ *FindCppPeriod*\ Let $P=S[i..j]$ be the palindrome of interest in size class $s$ and let $C$ be the CPP of $P$. First, query $Q_{s}$ for the successor of $s$, Denoted as $P'=[i'..j']$. $P'$ is a candidate for being $P^{*}$ If it is contained within $P$, and is in the size class $s$. If $P'$ is a candidate for being $P^{*}$, apply theorem \[t:1\] and get a *candidate* for the cluster’s period. Do the same for the second Successor of $a$, provided that it is a candidate for being $P^{*}$. Proceed to check all the LMPs in $L$. Apply theorem \[t:1\] on every LMP in $L$ that is a $P^{*}$ candidate. As we previously claimed, there are no more than two LMPs IN $Q_{s}$ that are contained in $P$, so we didn’t missed any candidates in $Q_{s}$.\ Assuming that $P^{*}$ is represented explicitly, we already have the right value of $p$ in hand. But what if it is represented as a part of a cluster? In this case we use the fact that $P^{*}$ is in the size class $s$, meaning that its start index in in the interval $A =[i..i+ \epsilon a^{s}]$ and its ending index is in the interval $B=[j-\epsilon a^{s}..j]$. If $P^{*}$ is indeed represented as a part of a cluster, it is either a prefix of a cluster or a suffix of one. Since $P^{*}$ is in the size class $s$, The cluster containing it must be in a size class $s'$ such that $s' \ge s$. But as we previously showed, there are no more than three clusters in class size $s'$ with starting indices in an interval smaller than $\epsilon a^{s'}$. This is at least $\epsilon a^{s}$ - the size of $A$ and $B$. So every priority queue $CPP_{i}$ with $i \ge s$ will yield at most four candidates for the cluster that implies the existence of $P^{*}$. These are at most two clusters that start within $A$ , located with two successor queries on $a$, and at most two clusters that end within $B$ , located by two predecessor queries on $b$. Given a cluster, its period, and the remainder of both the prefix CPP and the suffix CPP, we can find the largest implied prefix LMP contained in $P$ in constant time. We can also find the largest implied suffix LMP that is contained in $P$. These will be the only possible candidates for $P^{*}$ from this cluster since, as implied from the proof of lemma \[l:5\], $P^{*}$ is the largest extension of a run that is contained in $C$. To conclude this case - we iterate through every queue $CPP_{s'}$ with $s' \ge s$ and get the candidate clusters for containing $P^{*}$. From every candidate cluster, we get at most two LMPs that are candidates for being $P^{*}$. We apply theorem \[t:1\] on every $P^{*}$ that is encountered in this process.\ \ After that process, it is guaranteed that one of the candidates that we tested was indeed $P^{*}$. We take the minimal value of $p$ that was collected. This value must be the periodic seed.\ \ [**Complexity of finding the period:**]{} In the worst case, we do a constant number of predecessor and successor queries on every one of the $O(\log(n))$ priority queues. This may take $O(\log^{2}(n))$ time. we also go though $L$, and $|L| \in O(\log^{2}(n)$. Every one of these queries may yield a candidate for $P^{*}$ and we do constant work to produce a candidate for $p$ from each candidate. So the complexity is $O(\log^{2}(n)$.\ \ With this, our algorithm is finally complete. We try to add every LMP in $L$ to its appropriate $Q_{i}$. If the insertion results in two LMPs in $Q_{i}$ containing each other, the contained element is removed, the periodic palindrome cluster is calculated, and *FindCppPeriod* is invoked to compute its periodic seed. The prefix CPP and the suffix CPP of the cluster can be deduced from $p$ and from $C$, the CPP of the containing LMP.\ \ **Complexity:** Finding, extending and cutting all the affected LMPs and CPPs takes a constant amount of priority queue queries and LCP queries per exponential distance level for each priority queue. Resulting in $O(\log^{2}(n) \cdot (\log(n) + \lambda(n)))$ time, where $\lambda(n)$ is the time for LCP computation. In the worst case, we activate *FindCppPeriod* for every LMP in $L$ when adding it to $Q_{i}$, resulting in $O(\log^{4}(n))$ time. Overall, the complexity is $O(\max(\log^{4}(n), \lambda(n) \cdot \log^{2}(n))$. Since dynamic LCP queries can be computed in polylogarithmic time, this is $\tilde{O}(1)$. Conclusion and Open Problems ============================ We presented a dynamic algorithm for maintaining the longest palindromic subsequence in a changing text. This can be done in time $\tilde{O}(1)$ per change.\ \ We made heavy use of a polylogarithmic time dynamic LCP algorithm. It would be interesting to tighten up the dynamic LCP time as much as possible, and thus achieve logarithmically better time.\ \ The field of dynamic string matching is re-emerging in recent years. It would be interesting to study various string problems in a dynamic setting, such as finding the longest periodic substring, and finding various motifs. [^1]: Department of Computer Science, Bar Ilan University, Ramat Gan 52900, Israel. E-mail: [email protected]. [^2]: Department of Computer Science, Bar Ilan University, Ramat Gan 52900, Israel. E-mail: [email protected]. This work is part of this author’s Ph.D. dissertation. [^3]: This work was partially supported by ISF grant 1475/18 and BSF grant 2014028.
{ "pile_set_name": "ArXiv" }
[**The Finite-Difference Analysis and Time Flow**]{}\ Ashot Yu. Shahverdian\ (Yerevan Physics Institute)\ [**1.   Introduction**]{}\ The present paper introduces a method of analysis of one dimensional systems. Its application to study of Poincare recurrence time flow is considered. The approach suggested consists of reducing the research of a given system’s orbits ${\bar X}=(x_{i})_{i=1}^{\infty}$ to analysis of alternations of the monotone increase and decrease of higher order absolute finite differences, taken from $\bar X$. Through some special representation of finite orbits, we associate with ${\bar X}$ some numerical sequence, is called the conjugate orbit. We formulate two conditions, when the study of the ${\bar X}$ is reduced to analysis of the special orbits’ asymptotical intersections with some base set from numerical interval $(0,1)$ of zero Lebesgue measure. Hence, the approach can be characterized as an “asymptotical” analogy of Poincare’s classical “section”method. The method distinguishes two cases: the continuous, when some numerical quantities $\rho$ can take arbitrary values from interval $(0,1)$, and the discrete case, when all of them belong to a finite set. The continuous was applied in Ref. 1 to study of neural activity. In the present work, we are more interested in the discrete case, which has an additional specific. Thus, as soon as we have stated the method’s applicability, the mechanism generating time series $\bar X$, is replaced by the return map $\cal R$, generating the conjugate series. While the actual exact description of this mechanism may remain unknown (e.g., for earthquake time series or neuron spike trains), the function $\cal R$ does not depend on it and has a simple analytic shape. This approach is applicable to the study of various time series, arising in modern applied science (e.g., medicine, astronomy). The innovation consists of a new way to measure the fast oscillating time series – this measure is the Hausdorff dimension of some thin spaces $\cal A$, to which the conjugate series are attracted. As an example of such an application, the computational analysis of the sequences of fractional parts is presented. We show that the results obtained testify on the possibility of Cantorian structure of our usual time flow. In this context, we recall some well known (qualitative) conclusions on irregular and intricate time flow, as indicated in Bergson’s philosophical treatments [@hb] and in some works on quantum physics [@ms; @di].\ [**2.   Description of the Method**]{}\ [**2.1   Finite differences and conjugate orbits**]{}\ We deal with one dimensional, deterministic or stochastic systems, generating numerical sequences ${\bar X} = (x_{i})_{i=1}^{\infty}$ ($0\leq x_{i}\leq 1$). We impose no restrictions on the system, and it may also possess different inner states changing with time. Below, we set two conditions on a given orbit $\bar X$, under which it should be regarded as chaotic. We introduce the notion of conjugate (with $\bar X$) orbit $\bar \nu$, in terms of which the irregular nature of $\bar X$ can be established and studied. Let us first give some special representation of finite orbit ${\bar{X}}_{k}=(x_{i})_{i=1}^{k}$, which reflects its higher order differential structure. For the sequence ${\bar{X}}_{k}$ and number $1\leq s\leq k-1$ we let $$\Delta_{i}^{(0)} = x_{i}, \qquad \Delta_{i}^{(s)} = |\Delta_{i+1}^{(s-1)} - \Delta_{i}^{(s-1)}| \quad (i=1, 2, \ldots, k-s).$$ It is not difficult to obtain $$\Delta_{i}^{(s-1)} = \mu_{k,s-1} + \sum_{p=1}^{i-1}(-1)^{\delta_{p}^{(s)}} \Delta_{p}^{(s)} - \min_{0\leq i\leq k-s}(\sum_{p=1}^{i} (-1)^{\delta_{p}^{(s)}}\Delta_{p}^{(s)})$$ where $$\delta_{p}^{(s)} = \left \{ \begin{array}{ll} 0 & \Delta_{p+1}^{(s)} \geq \Delta_{p}^{(s)} \\ 1 & \Delta_{p+1}^{(s)} < \Delta_{p}^{(s)} \end{array} \right. \quad \mbox{(it is supposed that} \quad \sum_{1}^{0} = 0) \ ,$$ and $$\mu_{k,s} = \min \{ \Delta_{i}^{(s)}: 1\leq i \leq k-s\} \ .$$ From Eq. (1) it is easy to see, that if for a given $m$, $0\leq m\leq k$ all of the finite binary sequences $$\delta_{1}^{(s)}, \delta_{2}^{(s)}, \ldots, \delta_{k-s}^{(s)} \qquad (s = 0, 1, \ldots, m)$$ as well as all of the numbers $$\mu_{k,1}, \ \mu_{k,2}, \ \ldots , \mu_{k,m} \quad \mbox{and} \quad \Delta_{1}^{(m)}, \Delta_{2}^{(m)}, \ldots, \Delta_{k-m}^{(m)}$$ are known, one can completely restore the initial finite orbit ${\bar X}_{k}=(x_{i})_{i=1}^{k}$. Hence, it follows from (1), we can consider that finite orbits with length $k$ are given namely in the next special representation $${\bar \zeta}_{k} = (r_{1}^{(k)},r_{2}^{(k)},\ldots,r_{m}^{(k)}; \mu_{k,1}, \ \mu_{k,2}, \ \ldots , \mu_{k,m}; \rho_{1}, \ \rho_{2}, \ \ldots, \rho_{k-m})$$ where $$r_{s}^{(k)} = 0.\delta_{1}^{(s)}\delta_{2}^{(s)}\cdots \delta_{k-s}^{(s)} \qquad (1\leq s\leq m)$$ are the rational numbers, and $$\mu_{k,1}, \ \mu_{k,2}, \ \ldots , \mu_{k,m} \quad \mbox{and} \quad \rho_{1}, \ \rho_{2}, \ \ldots , \rho_{k-m} \quad (\mbox{where} \quad \rho_{i} = \Delta_{i}^{(m)})$$ are some numbers from interval $[0,1]$. Here $m=m_{k}$ and $m_{k}$ tend to $\infty$ as $k\to \infty$. It is easy to see, that after applying the recurrent procedure (1) the sequence ${\bar{X}}_{k}$ is completely recovered by ${\bar{\zeta}}_{k}$. Now, we can introduce the basic tool of the method – the notion of conjugate orbit: we say, that numerical sequence ${\bar \nu} = (\nu_{s})_{s=1}^{\infty}$ is the conjugate orbit, associated with the given orbit ${\bar X}= (x_{i})_{i=1}^{\infty}$, if for each $s\geq 1$ the terms $\nu_{s}$ are defined as follows $$\nu_{s} = \lim_{k\to \infty}r_{s}^{(k)} \qquad (= 0.\delta_{1}^{(s)}\delta_{2}^{(s)} \delta_{3}^{(s)}\ldots)$$ were $r_{s}^{(k)}$ are the $r$-coordinates from (2). Let us introduce a thin set $\cal B$ and its subsets ${\cal B}_{k}$, are the base sets, mentioned in Sec. 1. We consider the numbers $0<x<1$ represented in the form of binary expansion, $$x=0.\delta_{1}\delta_{2}\delta_{3}\ldots \qquad (=\sum_{k=1}^{\infty}2^{-k}\delta_{k}, \quad \delta_{k}=0, 1) \ ;$$ further, we also operate with the corresponding binary sequences $\bar x$, $${\bar x}= (\delta_{1}, \delta_{2}, \delta_{3}, \ldots, \delta_{n},\ldots) \ .$$ For a given natural $k\geq 2$ we define the subsets ${\cal B}_{k}$ of numerical interval $(0,1)$: ${\cal B}_{k}$ is the set all of those real numbers $x\in (0,1)$ for each of which $$n_{i+1}-n_{i} \leq k \qquad (i=1, 2, 3, \ldots)$$ where $n_{i}$ denote all the consecuitive positions where the changes of binary symbol from (4) occur, $\delta_{n_{i+1}} = 1-\delta_{n_{i}}$. Let also $\cal B$ be the union of all ${\cal B}_{k}$ ($k\geq 2$). All of the sets ${\cal B}_{k}$, $ 2\leq k \leq \infty $, have zero Lebesgue measure (see Ref. 1) (hereafter, in order to reduce some formulations, we use also the notation ${\cal B}_{\infty} \ (\equiv {\cal B})$). Let us now suppose, that the orbit $\bar X$ is such, that each of the sequences $${\bar X}^{s} = (\delta_{1}^{(s)}, \delta_{2}^{(s)}, \ldots, \delta_{n}^{(s)}, \ldots) \qquad (s = 0, 1, \ldots)$$ where $s$ is fixed, has bounded lengths of series with the same binary symbol. In other words, we assume that for increasing sequence of indeces $n_{i}^{(s)}$, designating all of those positions in natural series, where the changes of binary symbol occur, $\delta_{n_{i+1}^{(s)}}^{(s)} = 1-\delta_{n_{i}^{(s)}}^{(s)}$, we have $$n_{i+1}^{(s)}-n_{i}^{(s)}\leq K_{s} < \infty \qquad (i = 1,2,\ldots) \ .$$ Hence, taking both numbers $k$ and $k-m$ large enough (for determinity it can be chosen $m$ is equal to entire part of $k/2$, $m=[k/2]$), we can consider each of the sequences (5) as the binary expansion of some number $\nu_{s}\in {\cal B}_{K}$, $$\nu_{s} = 0. \delta_{1}^{(s)} \delta_{2}^{(s)} \ldots \delta_{n}^{(s)}\ldots$$ or, in other words, each of the sequences $(r_{s}^{(k)})_{s=1}^{\infty}$ of the form (3) converges (as $k\to \infty$) to a number $\nu_{s}$ from ${\cal B}_{K}$. Here, $2\leq K \leq \infty$, and according to (6) it should be taken $K = \sup \ \{ K_{s}: s\geq 1 \}$.\ [*Continuous Case*]{}.  Now, we asume also, that the quantities $$\mu_{k} = \sum_{i=1}^{m_{k}}\mu_{k,i}^{2} + \sum_{i=1}^{k-m_{k}}\rho_{i}^{2}$$ converge to zero, $$\lim_{k\to \infty}\mu_{k} = 0 \ .$$ In such a way, the limitations we impose on the time series $\bar X$ are the following: > $(C_{1})$:   for every $k\geq 1$ the sequence ${\bar X}^{k} \in {\cal B}$ > $(C_{2})$:   the quantities $\rho_{i,k}$ and $\mu_{i,k}$ are such, that: $\mu_{k}=o(1)$ as $k\to \infty$. Then, one can see [@av], these rwo restrictions imply that the transformed sequence ${\bar \zeta}_{k}$ from (2) tends to space ${\cal B}^{\infty} \ (={\cal B}\times{\cal B}\times\ldots)$, $$||{\bar{\zeta}}_{k} - {\cal B}^{\infty}|| = o\:(1) \qquad (k \to \infty)$$ where $||.||$ is the usual metric in Euclidean spaces $R^{k}$, $||x|| = (\sum_{i=1}^{k}x_{i}^{2})^{1/2} $. Particularly, (8) implies that for every $s$-th ($1\leq s\leq m$) coordinate of ${\bar{\zeta}}_{k}$ we have $$\lim_{k\to\infty}{\bar{\zeta}}_{k}^{(s)} = \lim_{k\to\infty}r_{s}^{(k)} = \nu_{s} \in {\cal B}$$ – namely this relation was meant when we mentioned the “asymptotical” intersections (of transformed into a special form orbits) with a thin space.\ [*Discrete Case*]{}.  In this case we deal with one dimensional time series with upper bounded terms and with restricted measurement accuracy, known in advance. If it is some number of the form $10^{-m}$, $m\geq 1$, then multiplying all of the terms of our series by this quantity, we will obtain some time series $\bar X$, consisting of natural numbers bounded by a pregiven $N$, $${\bar X}= (n_{1}, n_{2}, n_{3}, \ldots, n_{k}, \ldots), \qquad n_{k}\in [0, N] \quad (k = 1, 2, 3, \ldots) \ .$$ In other words, we deal with the random sequences [@me]; particularly, such kind of sequences arise as a result of the realizations of hazard games. However, we consider only a special class of sequences (9) – as above, we impose two limitations (below $d\geq 1$ is some natural number; we may assume $d=1$) on the differential structure of the series $\bar X$: > $(D_{1})$:   for every $k\geq 0$ the sequence ${\bar X}^{k}\in {\cal B}_{K}$ > $(D_{2})$:   for all large enough $k\geq 1$ and all $i\geq 1$ the quantities $\rho_{i}^{k} \in \{0, d\}$. It should be noticed, that condition $(D_{1})$ (and $(C_{1})$) implies the dimensionality limitations (see next section), while the $(D_{2})$, an analogy of $(C_{2})$, is a requirement on certain regularity of the process, generating $\bar X$. The restriction $(D_{2})$ means, that we consider such processes (9), which are reduced to binary processes $\bar X$ (with components $x_{i}=0, 1$). Let us call the binary sequence $\bar X$, satisfying the conditions $(D_{1})$ and $(D_{2})$ as $\beta_{K}$-sequence and denote ${\hat{\cal B}}_{K}$ the collection all of the numbers $x\in (0,1)$ for which its binary expansion $\bar x$ is ${\beta}_{K}$-sequence. From the definition, we have the next characteristic of these sets; below $x_{n}=\Delta^{(n)}(x)$ and $\Delta$ is the shift transformation (or the “tent map”, see Ref. 6): $$\Delta(x) = \left \{ \begin{array}{ll} 2x \ & 0<x<1/2 \\ 2x-1 \ & 1/2 \leq x <1 \ . \end{array} \right.$$ The number $x\in {\hat {\cal B}}_{\infty}$ iff for some numbers $0<K_{p}<\infty$ the relation $$|x_{n}-\Delta^{(p)}(x_{n})| \geq 2^{-K_{p}}$$ holds for every $n\geq 0$ and $p\geq 1$. The number $x\in {\hat {\cal B}}_{K}$,  $2\leq K <\infty$, iff the inequality (10) holds with $K_{p}\equiv K$ and for every $n\geq 0$ and $p\geq 1$. According to Poincare recurrence theorem (see, e.g., Ref. 6), the relation $${\lim\inf}_{p\to \infty} |x-\Delta^{(p)}(x)| = 0$$ holds for each $0<x<1$ except some set of zero Lebesgue measure. In the case $K_{p}\equiv K$ the condition (10) means recurrence break (by analogy with “ergodicity break” from Ref. 7) for all points $x_{n}$. Hence, all the sets ${\hat{\cal B}}_{k}$ are found within exceptional set, mentioned in just quoted Poincare theorem. Note also certain resemblance of the relation (10) with the basic inequalities in Li - Yorke known theorems [@ly].\ [**2.2  Conjugate attractor and return map**]{}\ Let us now denote ${\cal A}$ the closure of conjugate orbit $(\nu_{s})_{s=1}^{\infty}$ , i.e. ${\cal A} \subset [0,1]$ is the union of the set $\{\nu_{s}: s\geq 1\}$ with the collection of all its cluster points. Then this orbit should be considered as a chaotic one, whenever $\cal A$ fills in either an interval or a Cantor set. We note, that for the most actual systems the constants from Eq. (6) for all $s\geq 1$ are upper bounded by the same number $K\geq 2$ and, consequently, $${\cal A}\subset {\cal B}_{K} \quad \mbox{and} \quad dim ({\cal A}) \leq dim ({\cal B}_{K}) \ .$$ Hence, just mentioned criterion of chaosity simply means that $\cal A$ is a Cantor set. If $\cal A$ appears to be the same set for almost each orbit $\bar X$, then we can conclude that $\cal A$ is the attractor of the system. The numerical values of Hausdorff dimension of the sets ${\cal B}_{k}$ ($2\leq k\leq \infty $) can be obtained from the formula [@av] $$z_{k}(2^{dim({\cal B}_{k})}) = 1 \quad \mbox{where} \quad z_{k}(s)=\sum_{n=1}^{k}s^{- n} \qquad (s>0)$$ – hence, (11) implies an upper estimate of Hausdorff dimension of attractor $\cal A$. If we consider the numbers $x$ from unit interval are given in the form (4), then it is not difficult to obtain the next statement (below, $a\oplus b = |a-b|$ is the logical sum of the binary symbols $a, b \in \{0, 1\}$): If $\bar X$ is a binary sequence, then the return map ${\cal R}: \nu_{k}\to \nu_{k+1}$ of the conjugate orbit $\bar\nu$ has the form $${\cal R} (x) = \sum_{n=1}^{\infty} 2^{-n}(\epsilon_{n} \oplus \epsilon_{n+1})$$ where $\epsilon_{n}$ are the coefficients of binary expansion of number $x$. One can find (in the operator form) the function $\cal R$ in Sinai’s early constructions on weak isomorphism of dynamical systems [@pb; @yg]. For more details on this function see Ref. 9, where a fractal dynamical system is introduced.\ [**3.   Sequences of Fractional Parts and Time Flow**]{}\ Some works in quantum mechanics [@ms] introduce the concept “Cantorian Space-Time”. The computational study of the flow of our usual (but in some sense discretized) time, implemented in this section, confirms again the possibility of such a concept. We are based on the method introduced – an approach completely different from those considered in other works on time flow [@rh; @mf; @gm]. Let us start from the statement and computational analysis of the following number-theoretical problem. Let us have two numbers $0<\alpha, \lambda <1$, $\alpha $ is irrational and let $$\{ \alpha \}, \ \{2\alpha\}, \ \{3\alpha\}, \ldots, \{n\alpha\}, \ldots$$ be the sequence of the fractional parts, generated by $\alpha$ (hereafter $\{x\}$ denotes the fractional part of number $x$). We are interested in the behavior of the time series $${\bar X} = (n_{k})_{k=1}^{\infty} \quad \mbox{where} \quad n_{k}=m_{k+1}-m_{k} \qquad (k\geq 1) \ ;$$ here $m_{k}$ are all of the natural numbers, arranged in increasing order, $m_{k+1}>m_{k}$, such that $$\{ \alpha m_{k} \} \in (0,\lambda) \ .$$ According to H. Weyl theorem [@pb; @hn] the sequence (12) is uniformly distributed in the interval $(0,1)$, so that we have infinite number of $m_{k}$, satisfying (13). If we take into attention the ergodicity of the transformation $\pi$, $$\pi(x)=\{ x + \alpha \} \ , \qquad \pi: (0,1) \to (0,1) \ ,$$ this statement can also be deduced (for a.e. irrational $\alpha$) from Poincare’s recurrence theorem. From numerous theoretical results on the sequences of fractional parts (see details in Ref. 14), we note the following Slater-Florec theorem: for given irrational $\alpha$ there exist two natural numbers $p$ and $q$, such that for all $k\geq 1$ we have $ n_{k}\in \{p, q, p+q\} $. We may also note a formula, obtained in Ref. 15 – it expresses the exact value of the total number of indeces $m_{k}$, satisfying (13) and not exceeding a given $N$, through the coefficients of some expansions of the numbers $N$ and $\lambda$. Our basic conclusion in the problem stated, deduced from computational experiments, can be formulated in the next form: for each irrational $0< \alpha <1$ there exist the numbers $0<\lambda <1$ such that the conjugate to $\bar X$ orbit $\bar\nu$ is asymptotically close to a Cantor set; and conversely, for each $0< \lambda <1$ there exist irrational numbers $0<\alpha <1$, such that the conjugate to $\bar X$ orbit $\bar\nu$ is asymptotically close to a Cantor set. Some of these computational results, on the Fig. 2 are presented. Here, the numerical value $\lambda=0.13$ and several values of parameter $\alpha$ are considered. In dependence of control parameter $\alpha$ the conjugate orbits $\bar\nu$ demonstrate all possible types of motion: the periodic, when they are attracted to some finite set, Cantorian, when the attractor $\cal A$ has Cantorian self-similar structure, and the completely chaotic motion, when the orbits fill in an interval from $(0,1)$. For instance, for $\alpha=0.1250002$ the Cantorian attractor in Fig. 2.b is shown. The graph Fig. 2.a has been obtained for the value $\alpha = 0.124999$; in this case the orbit $\bar\nu$ consists of 13 straight lines (some of them are not shown). In some respect, the behavior (in dependence of control parameter $\alpha$) of conjugate orbits in this problem reminds Feigenbaum’s transitions (see e.g., Ref. 6), and this needs further detailed computational study. To end, let us consider the following interpretation of the above introduced number-theoretical problem; in particular, this will explain the title of present section. Let us have in a two dimensional plain a circle $E$ and a point $S$, situated outside of disc, bounded by $E$. Let $\omega$ be the angle, under which the circle $E$ from the point $S$ is seen, and let also $L$ be an arc of $E$ with the length $|L|=\lambda <\omega$ with the center at the point of intersection of circle $E$ with the rectilinear segment, connecting its center with $S$. Let us assume that the point $M\in E$ is rotated in discrete time $T = \{ 1, 2, 3, \ldots. \}$ with the frequency $\alpha$. Then, the consecuitive moments of falling the point $M$ in the arc $L$ constistute a sequence of natural numbers $m_{k} \in T$, coinciding with the sequence $\bar X$ from Eq. (13). Also, it is not difficult to see, that the sequence of fractional parts (12) coincides with the sequence of iterates of the function $\pi (x)$ from Eq. (14), i.e. $$\{\alpha n \} = \pi^{(n)}(\alpha) \qquad (n\geq 1)$$ and, consequently, the numbers $m_{k}$ represent the Poincare recurrence (of point $M$ into the arc $L$) time. Now, let us interpret $E$ as “Earth” and $S$ as “Sun”. Then, replacing the continuous rotation of Earth by discrete motion, we obtain that $m_{k}$ appears to be an approximate of the flow of our usual time. In such a way, the above described computations demonstrate, that in principle, it is possible that our time possesses the Cantorian structure. The numerical values of the quantities $\alpha$, $\lambda$, and $\omega$ can be specified for this “astronomical” case, but here we do not dwell on this point.\ [**4.   Some Remarks and Discussions**]{}\ Let us discuss two problems, relating to the brain activity. In the paper Ref. 1, through the method above, applied to experimental data of actual neurons, the Cantorian strucuture of brain activity has been deduced. The method permits further development [@ay] by means of introducing some dynamical system ${\cal F} = ({\cal A}, T, \mu)$ on the Cantor spaces $\cal A$. The evolution operator $T$ is easly expressed through the return map $\cal R$, while the attractor $\cal A$ and invariant (singular) measure $\mu$ remain the main quantities, distinguishing different processes. In practical applications, the measure $\mu$ can be calculated through frequency analysis, accepted in mathematical-linguistic works (see, e.g. Ref. 5). Hence, we can operate with the theoretical-information characteristics of system $\cal F$. Whether there exists an analogy of known Billingsley-Eggleston formula [@pb] $$Dimension = Entropy,$$ relating to Eggleston thin spaces, for the fractal dynamical systems $\cal F$? Note also Mandelbrot’s work [@bm], who considered the applications of this formula to problems of turbulence. We know [@ep], that there exists some mechanism of time perception in the brain. Its activity, it follows from our previous results [@av], must be found in some thin (multidimensional) Cantor space. How to relate the measurable spaces ${\cal F}_{t}$ and ${\cal F}_{p}$, which correspond to time flow and time perception, with each other? Particularly, can we compare the values of $dim ({\cal A}_{t})$ and $dim ({\cal A}_{p})$? It seems, these problems can be resolved after detailed computational frequency analysis, in order to compute the invariant measures $\mu$ (and hence another characteristics, say, entropy) of both time flow and brain activity.\ [**Acknowledgements**]{}\ The author is grateful to Dr. A. V. Apkarian from State University of New York Health Science Center, Syracuse, N-Y, USA, for support and discussions of main topics of this work. [99]{} SHAHVERDIAN A.YU. & APKARIAN A.V., [*Fractals*]{} [**7**]{}, 93-103 (1999). BERGSON H. [*L’Evolution Creatrice*]{} (Paris). El NASCHIER M.S, [*Chaos, Solitons & Fractals*]{} [**1**]{}, 485-87 (1991). BLOKHINTSEV D.I., [*Space and Time in Microworld*]{} (Nauka, Moscow, 1984), in Russian. MARTIN N. F. G. & ENGLAND J. W., [*Mathematical Theory of Entropy*]{} (Addison-Wesley, Massachusetts, 1981). SCHUSTER H. G., [*Deterministic Chaos*]{} (Physik-Verlag, Weinheim, 1984). HILFER R., [*Chaos, Solitons & Fractals*]{} [**5**]{}, 1475-1484 (1995). LI T.-Y & YORKE J. A., [*Amer. Math. Monthly*]{} [**82**]{}, 985-992 (1975). SHAHVERDIAN A. YU., (Fractals, will be submitted, 1999). BILLINGSLEY P., [*The Ergodic Theory and Information*]{} (Wiley, N-Y, 1965). SINAI YA.G., [*Matemat. Sbornik*]{} [**63**]{}(105), 23-42 (1963), in Russian. SHLESINGER M. F., [*Ann. Rev. Phys. Chem.*]{} [**39**]{}, 269-90 (1988). ZASLAVSKY G. M. & TIPPET M. K., [*Phys. Rev. Lett.*]{} [**67**]{}, 3251 (1991). KUIPERS L. & NIEDERREITER H., [*Uniform Distribution of Sequences*]{} (Wiley, London, 1976). SHAHVERDIAN A. YU. & NIEDERREITER H., (will be submitted, 1999). MANDELBROT B., [*Lect. Not. in Math.*]{} [**615**]{}, 83-93 (Springer, 1977). POPPEL E. & SCHILL K., in [*The Handbook of Brain Theory and Neural Networks*]{}, ed. Michael A. Arbib (The MIT Press, Massachusetts, 1995). [**Figure Captions**]{}\ Figure 1.  The graph of inverse function ${\cal M}(x) = {\cal R}^{-1}(x)$. The graph is constructed through 15,000 points, computed by means of random numbers generator. From definition of the function $\cal R$ it is easy to see that $\cal M$ is a multivalent function, defined on the interval $(0,1)$. The graph of the return map $\cal R$ can be obtained as a result of symmetric reflection of this “fractal letter $M$” in respect of diagonal $(0,0)$, $(1,1)$.\ Figure 2.  Two different types of conjugate orbit’s behavior in dependence of control parameter $\alpha$. The graph a) demonstrates the case when the attractor $\cal A$ consists of the finite set, while in the graph b) these orbits are attracted to some Cantor set. The graph c) shows the self-similar structure of the Cantor set $\cal A$ from the previous graph.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A brief review of the attempts to define “elements of reality” in the framework of quantum theory is presented. It is noted that most definitions of elements of reality have in common the feature to be a definite outcome of some measurement. Elements of reality are extended to pre- and post-selected systems and to measurements which fulfill certain criteria of weakness of the coupling. Some features of the newly introduced concepts are discussed.' author: - Lev Vaidman title: ' WEAK-MEASUREMENT ELEMENTS OF REALITY' --- [*School of Physics and Astronomy\ Raymond and Beverly Sackler Faculty of Exact Sciences\ Tel Aviv University, Tel-Aviv 69978, Israel.\ *]{} INTRODUCTION {#int} ============ Experiments performed in laboratories of twenties century tell us that the picture of reality we have from our everyday experience cannot be a true representation of nature. This leads us to develop a new language which will be more adequate for the description of our (quantum) world. The fact that our world is quantum seems indisputable because the predictions of quantum theory have been confirmed with incredible precision in all experiments which have been performed until today. There have been numerous attempts to describe quantum reality, but a consensus has not been reached. I do not have the pretension to give here [*the*]{} correct definition of elements of reality in quantum theory. I believe, however, that the concepts which I introduce are helpful tools for developing intuition to see and to understand quantum phenomena. Einstein, Podolsky and Rosen (EPR) were the pioneers in their attempt to define [*elements of reality*]{} in quantum theory. The difficulty of considering elements of reality in the framework of quantum theory which they discussed in their seminal paper[@EPR] led Bohr to claim that the only elements of reality in quantum theory are results of (quantum) measurements. (This is, of course, an oversimplification. Max Jammer has illuminating writings on this subject clarifying this important historical issue[@Jamm1; @Jamm2].) The position of Bohr is certainly consistent, but, I believe, is not very fruitful: it implies that no reality is associated with the quantum system between measurements. This approach refuses to consider various concepts which are helpful tools to see the bizarre features of our quantum world. The EPR definition of element of reality is: > “If, without in any way disturbing the system, we can predict with certainty (i.e. with probability equal to unity) the value of a physical quantity, then there exists an element of physical reality corresponding to this physical quantity.” The key to understanding of the meaning of their definition is the interpretation of the phrase “without in any way disturbing the system”. They assumed a strong version of relativistic causality, i.e. that no operation in a space-like separated region can disturb the system. I would say that Bell [@Bell64] showed that the existence of EPR elements of reality is inconsistent with predictions of quantum theory. However there are many important historical aspects of that issue as Max Jammer [@Jamm2] has taught us. The next definition of element of reality I want to bring here was inspired by EPR but it makes no assumptions about relativistic causality. This is the definition of element of reality due to Redhead [@R] > “If we can predict with certainty, or at any rate with probability one, the result of measuring a physical quantity at time $t$, then, at time $t$, there exists an element of reality corresponding to this physical quantity and having value equal to the predicted measurement result.” Although the definitions look very similar, they are very different conceptually. The EPR element of reality is defined by the mere possibility of finding the outcome “without disturbing the system in any way”, i.e., from some measurements, while according to Redhead, there is an element of reality when we [*know*]{} the outcome of the measurement without performing any other measurement. For example, if we consider an EPR pair of spin 1/2 particles (i.e. the two particles in the singlet state) then, according to EPR, the spin components of each particle in all directions are elements of reality since they can be found from the spin measurement performed on the other particle (and the assumption of no action at a distance ensures that this measurement does not disturb the first particle). According to Redhead, however, there is not any (local) element of reality for a spin variable of any particle; we cannot predict the result of such measurement. The three alternative definitions of element of reality which I presented above are all related to definite results of measurements. In the first (spirit of Bohr) these are the results of measurements which were actually performed, in the last (Redhead) these are the results of measurements which (if performed) are known with certainty, and in the second (EPR) these are the results of measurements which we can ascertain with certainty using other measurements (performed in a space-like separated region). Thus, it is natural to adopt the following principle:[*For any definite result of a measurement there is corresponding element of reality.*]{} ELEMENTS OF REALITY AS A DEFINITE SHIFT OF THE PROBABILITY DISTRIBUTION OF THE POINTER VARIABLE {#ershift} =============================================================================================== In order to understand the meaning of “definite result of measurement” we have to analyze the concept of quantum measurement. The standard definition of quantum measurement is due to von Neumann [@vN]: The measurement of a physical variable $A$ is described by the Hamiltonian: $$\label{ham} H = g(t) P A~~~~,$$ where $P$ is a canonical momentum conjugate to the pointer variable $Q$ of the measuring device. The function $g(t)$ is nonzero only for a very short time interval corresponding to the measurement, and is normalized so that $\int g(t)dt=1$. During the time of this impulsive measurement, the Hamiltonian (2) dominates the evolution of the measured system and the measuring device. Since $[A , H] =0$, the variable $A$ does not change during the measuring interaction. The initial state of the pointer variable is usually modeled by a Gaussian centered at zero: $$\label{phi-in} \Phi _{in} (Q) =(\Delta ^2 \pi )^{-1/4} e^{ -{{Q ^2} /{2\Delta ^2}}}.$$ The pointer is in the “zero” position before the measurement, i.e. its initial probability distribution is $$\label{prob} prob(Q) = (\Delta ^2 \pi )^{-1/2} e^{ -{{Q ^2} /{\Delta ^2}}}.$$ If the initial state of the system is an eigenstate $ |\Psi_1 \rangle = |a_i \rangle $, then after the interaction (1), the state of the system and the measuring device is: $$\label{mdstate} (\Delta ^2 \pi )^{-1/4} |a_i \rangle e^{ -{{(Q-a_i)^2} /{2\Delta ^2}}}.$$ The probability distribution of the pointer variable, $(\Delta ^2 \pi )^{-1/2} e^{ -{{(Q-a_i)^2} /{\Delta ^2}}}$ remained unchanged in its shape, but it is shifted by the eigenvalue $a_i$. This eigenvalue is considered to be the element of reality. Thus we can translate the meaning of “definite result of quantum measurement” as [*definite shift of the probability distribution of the pointer variable*]{}. I suggest to take this property to be the definition: [*If we are certain that a procedure for measuring a certain variable will lead to a definite shift of the unchanged probability distribution of the pointer, then there is an element of reality: the variable equal to this shift.*]{} In an ideal measurement, the initial probability distribution of the pointer is well localized around zero, and thus the final distribution is well localized around the eigenvalue. Thus, the reading of the pointer variable in the end of the measurement almost always yields the value of the shift (the eigenvalue of the variable). The generalization I suggest in the above definition is applicable also to the situations in which the initial probability distribution of the pointer variable has a significant spread. Although, in this case, the reading of the measuring device at the end of the measurement is not definite, the shift of the distribution is. In such a case, a measurement performed on a single system does not yield the value of the shift (the element of reality), but such measurements performed on large enough ensemble of identical systems yield the shift with any desirable precision. If the initial state of the system is a superposition $ |\Psi_1 \rangle = \Sigma \alpha_i |a_i \rangle $, then after the interaction (2) the state of the system and the measuring device is: $$\label{state} (\Delta ^2 \pi )^{-1/4} \Sigma \alpha_i |a_i \rangle e^{ -{{(Q-a_i)^2} /{2\Delta ^2}}}.$$ The probability distribution of the pointer variable corresponding to the state (5) is $$\label{prob2} prob(Q) =(\Delta ^2 \pi )^{-1/2} \Sigma |\alpha_i|^2 e^{ -{{(Q-a_i) ^2} /{\Delta ^2}}}.$$ In case of ideal measurement this is a weighted sum of the initial probability distribution localized around various eigenvalues. It is not a single shifted original probability distribution, and therefore, according to our definition in this case there is no element of reality for the value of the measured variable. (If one adds a “collapse” to one of the positions of the pointer, he ends up with one of the eigenvalues. So, [*a la*]{} Bohr, after the measurement is completed, there is an element of reality, but here we are looking for an element of reality which can be predicted with certainty before the measurement.) In general, when the initial probability distribution is not very well localized, the final distribution (6) is very different from the original distribution. However, as I will show in the next section, even in this case there is a well defined limit in which the final distribution converges to the unchanged initial distribution shifted by a well defined value. I will suggest to call this well defined shift an element of reality of a new type. WEAK-MEASUREMENT ELEMENT OF REALITY {#wmer} =================================== I propose to consider the standard measuring procedure (1) in which we weaken the interaction in such a way that the state of the quantum system is not changed significantly during the interaction. Usually, the measuring interaction leads to a very large uncertain change of the system due to a large uncertainty of the variable $P$. Indeed, in the standard measurement we require that the pointer shows zero before the measurement, i.e., $\Delta$ is very small for the initial state of the measuring device (2). This requires large uncertainty in $P$, and therefore the Hamiltonian (1) causes a large uncertain change. I propose to take the initial state of the measuring device in which $\langle P \rangle = 0$ and the uncertainty in $P$ is small, and I will show that this is enough to ensure that the pointer probability distribution after the measuring interaction is essentially equal to the shifted initial distribution. In this case the interaction Hamiltonian (1) is small and this is why we call such procedure a [*weak measurement*]{}[@AV-weak]. The limit of weak measurement corresponds to $\Delta \gg a_i$ for all eigenvalues $a_i$. Then, we can perform the Taylor expansion of the sum (6) around $Q=0$ up to the first order and rewrite the probability distribution of the pointer in the following way: $$\begin{aligned} \label{prob2} prob(Q) =(\Delta ^2 \pi )^{-1/2} \Sigma |\alpha_i|^2 e^{ -{{(Q-a_i)^2} /{\Delta ^2}}} =~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\ (\Delta ^2 \pi )^{-1/2} \Sigma |\alpha_i|^2 (1 -{{(Q-a_i) ^2} /{\Delta ^2}}) = (\Delta ^2 \pi )^{-1/2} e^{ -{{(Q-\Sigma|\alpha_i|^2a_i) ^2} /{\Delta ^2}}} . \end{aligned}$$ But this is exactly the initial distribution shifted by the value $\Sigma|\alpha_i|^2a_i$. Thus we will say that there is here a [*weak-measurement element of reality*]{} $A_w = \Sigma|\alpha_i|^2a_i$. The mathematical expression of this weak-measurement of reality is not something new. This is the expectation value: $A_w = \Sigma|\alpha_i|^2a_i =\langle \Psi |A|\Psi\rangle$. The weak value is obtained from statistical analysis of the readings of the measuring devices of the measurements on an ensemble of identical quantum systems. But it is different conceptually from the standard definition of expectation value which is a mathematical concept defined from the statistical analysis of the [*ideal*]{} measurements of the variable $A$ all of which yield one of the eigenvalues $a_i$. I have showed that expectation values fall under the definition of the elements of reality as a definite shift of unchanged probability distribution of the pointer variable in the limit of weak-coupling measurement ($\Delta$ large). The advantage of this definition is that it is applicable for any quantum system in any pure or mixed state and for any quantum variable of this system. The disadvantage is that usually we need an ensemble of identical systems in order to find the expectation values of their quantum variables with good precision. It is important to mention that sometimes the [*weak*]{} measurement need not be too weak and the expectation value can be found with relatively good precision from a single such measurement. This is the case when the uncertainty of $A$ is small. The latter, in particular, is a generic property of “average” variables of large composite systems. ELEMENTS OF REALITY OF PRE- AND POST-SELECTED SYSTEMS {#erpps} ====================================================== The concept of the weak-measurement element of reality yields novel results when considered on pre- and post-selected systems. But the ideal-measurement element of reality of the pre- and post-selected system also have novel features and I will consider them first. For the pre- and post-selected systems it is fruitful to consider a modification of Redhead’s definition when I replace “predict” by “infer”[@V-prl; @V-er]. > “If we can infer with certainty, or at any rate with probability one, the result of measuring a physical quantity at time $t$, then at time $t$, there exists an element of reality corresponding to this physical quantity and having a value equal to the predicted measurement result.” Essentially, Redhead’s definition says that $A=a$ if and only if the system is in the appropriate eigenstate or mixture of such eigenstates. In pre-and post-selected situations we might have $A=a$ even if the system is not in an eigenstate of $A$; we obtain the inference both from the preparation and from the post-selection measurement. Elements of reality in the pre- and post-selected situations might be very peculiar. One such example is a single particle inside three boxes $A, B$ and $C$, with two elements of reality: “the particle is in box 1” and “the particle is in box 2”. This is the case with the pre-selection of the state of the particle $|\Psi_1 \rangle = 1/\sqrt 3 ~(|A\rangle + |B\rangle + |C\rangle)$ which was found later in the state $|\Psi_2 \rangle = 1/\sqrt 3 ~(|A\rangle + |B\rangle -|C\rangle)$. If, in the intermediate time it was searched for in box $A$ it has to be found there with probability one, and if, instead, it was searched for in box $B$, it has to be found there too with probability one. (Indeed, not finding the particle in box $A$ would project the initial state $|\Psi_1 \rangle$ onto $ 1/\sqrt 2 ~( |B\rangle + |C\rangle)$ which is orthogonal to the final state $|\Psi_2\rangle$.) This example shows that the “product rule” does not hold for elements of reality of pre- and post-selected systems. Indeed $\Pi_A =1$ is an element of reality and $\Pi_B =1$ is an element of reality, but $\Pi_A \Pi_B =1$ is not an element of reality. In fact, $\Pi_A \Pi_B =0$ is an element of reality. The meaning of this equation is a trivial point that the probability to find the particle both in $A$ and $B$ is equal to zero. The “sum rule” does not hold either. Indeed, there is no element of reality $\Pi_A + \Pi_B =2$. In fact, there is no any element of reality for the value of the sum $\Pi_A + \Pi_B$. This means that when we perform a measurement which tells us that the particle is inside one of the boxes $A$ or $B$, but without telling in which one, the probability to find it is neither zero nor one. (The probability is equal 2/3, but this does not correspond to any element of reality.) Note, however, that there is an element of reality for the sum of the three projection operators: $\Pi_A + \Pi_B + \Pi_C=1$; clearly, the measurement testing the existence of the particle in the three boxes will say yes with probability one. The elements of reality for pre- and post-selected quantum systems have unusual and counterintuitive properties. But, maybe this is not because of the illness of their definition, but due to bizarre features of quantum systems which goes against the intuition built during thousands of years, when the results of quantum experiments were not known. WEAK-MEASUREMENT ELEMENTS OF REALITY OF PRE- AND POST-SELECTED SYSTEMS {#wmerpps} ====================================================================== The next natural step is to consider the limit of weak measurements performed on pre- and post-selected quantum systems. Again we consider the measuring Hamiltonian (1), the initial state of the measuring device (2) and the limit of large $\Delta$. In general, this will lead to a very low precision of the measurement so we consider an ensemble of identical systems with such measurements. The difference here from the weak measurement of Sec. 3 is that now, before reading the outcomes of the measuring devices, we post-select a certain state of the system and discard the readings of measuring devices corresponding to the systems for which the post-selection was not successful. I will not repeat here the calculations, they can be found in Ref.(7). The above procedure is called [*weak measurement*]{} and it is indeed converges to a well defined value. At the limit of large $\Delta$, the probability distribution of the final state of the measuring device converges to the initial distribution shifted by the real part of the [*weak value*]{} of the variable $A$: $$\label{a-weak} A_w = {{\langle \Psi_2|A|\Psi_1\rangle} \over {\langle \Psi_2|\Psi_1\rangle}}.$$ Our definition of elements of reality, i.e., a definite shift of the probability distribution of the pointer variable yields for pre-and post-selected systems the weak value (8). Even the imaginary part of the weak value falls under this definition, but only for a particular case of the Gaussian as the initial measuring device state. In order to see that we have to consider, instead of the pointer position distribution, the distribution of the conjugate momentum. It turns out[@AV-weak] that the original Gaussian in $P$ does not change its shape (again, in the limit of a large $\Delta$) and is shifted by the value $Im(A_w)$. The weak-measurement elements of reality of pre-selected only systems, the expectation values, are a particular case of the weak values. Indeed, we can consider a future measurement which tests that we are in the initial state $|\Psi_1\rangle$. The weak measurement, by assumption, does not change the state of the system significantly, and therefore, this future measurement [*must*]{} yield the state $|\Psi_1\rangle$. But then, the definition of weak value (8) yields the expectation value. The advantage of weak-measurement elements of reality is that they are well defined for any situation and for all variables. It also have some classical type features. The “sum rule” holds. Indeed, if $C = A + B$ then $C_w = A_w + B_w$. Therefore, if $A_w =a$ is a weak-measurement element of reality and $B_w = b$ is a weak-measurement element of reality, then $(A+B)_w=a +b$ is also a weak-measurement element of reality. The “product rule”, however, does not hold. From $C = A B$ does not follow that $C_w = A_w B_w$. The main disadvantage of weak-measurement elements of reality is again that usually we need an ensemble of identical pre- and post-selected systems in order to find the weak values of their quantum variables with a good precision. However, there are certain important cases in which the [*weak*]{} measurement need not be too weak and the “weak” value can be found with a relatively good precision from a single such measurement. For example if a spin $N$ particle is prepared in the state $S_x=N$ and later found in the state $S_y=N$, then, at the intermediate time the weak value of the spin component in the direction $\hat{\xi}$ which bisects $\hat{x}$ and $\hat{y}$ is larger than $N$. Indeed, $(S_\xi)_w =\sqrt{2} N$. An experimenter can repeatedly see this “forbidden” value in a standard measurement with precision of order $\sqrt{N}$. Note, however, that for any “unusual” weak values the probability to obtain the required result of the post-selection is extremely small. (In the last example this probability is equal $2^{-N}$.) The concept of weak-measurement elements of reality is a generalization of the usual concept of the (strong-measurement) element of reality. Indeed, if we know with certainty that a strong measurement of $A$ will yield $A=a$ with probability one, then we know that the weak measurement will also yield $A_w= a$ [@foot2]. Thus, all (strong) elements of reality are also weak elements of reality. The class of weak elements of reality is much wider; it is defined for all variables for any realizable pre- and post-selected situation (as well as for the pre-selection only situation). In contrast, the strong elements of reality are defined only for some variables in each situation and sometimes they do not exist at all. (There is no any [*local*]{} element of reality for spin variables of an EPR pair.) Let us analyze the example of the three boxes we have introduced above. Since we know several (ideal-measurement) elements of reality we can immediately write down corresponding weak-measurement elements of reality for the discussed pre- and post-selected particle: $$\label{psi1} (\Pi_A)_w =1,~~ (\Pi_B)_w =1,~~ (\Pi_A+ \Pi_B +\Pi_C)_w =1 .$$ Now using the sum rule we obtain another weak-measurement element of reality: $$\label{psi2} (\Pi_C)_w = (\Pi_A+ \Pi_B +\Pi_C)_w- (\Pi_A)_w -(\Pi_B)_w = -1 .$$ To say that there is a “reality” of having $-1$ particle in a box sounds paradoxical. However, when we test this reality weakly this is what we see. We cannot see this “reality” for one particle because the uncertainty of the appropriate weak measurement has to be much larger than 1, but if we have a large number of such pre- and post-selected particles in the three boxes, then a realistic measurement of the pressure in the boxes will yield $p$ for boxes $A$ and $B$ and the negative pressure, $-p$, for the measurement in the box $C$. Note, however, that the probability to obtain in the post-selection measurement the state $|\Psi_2\rangle$ for a macroscopic number of particle is extremely small. CONCLUSION {#CONC} ========== The name “element of reality” suggests an ontological meaning. However, historically, and in the present paper, the element of reality is an epistemological concept. It is better to have ontological elements of reality, but there are severe difficulties in constructing them. Probably, the most serious attempt in this direction is the causal interpretation[@Bohm] which introduces “real” point-like particles moving according to a simple (but nonlocal) law. However, it seems that we are forced to accept that the total quantum wave in the causal interpretation has also ontological status[@AV-Bohm]. I feel that if we do consider the quantum wave function of the universe as an ontological reality, we need not add anything else, e.g. Bohmian particles. I am perfectly ready to accept that the reality of our physical universe [*is*]{} its wave function. But since that reality is very far from what we experience, I think it is fruitful to define epistemological reality as it has been done here. I certainly see a deficiency of weak-measurement elements of reality defined above in the situations in which they cannot be measured on a single system. Still, I do not think that the fact that a weak value cannot be measured on a single system prevents it from being a “reality”. We know that the measuring device shifts its pointer exactly according to the weak value, even though we cannot find it because of the large uncertainty of the pointer position. We can verify this knowledge performing measurements on an ensemble. In the cases when weak values can be found with good precision on a single system, the concept of weak-measurement elements of reality is fully justified. In fact, the word “weak” is not exactly appropriate, since the measurements in question are the usual one. (However, they are not “ideal” in the von Neumann sense, since the measuring Hamiltonian has to be bounded.) Another example of measurements which are “good measurements” in the sense that they yield a measured quantity with a good precision and which yield weak-measurement elements of reality are [*protective measurements*]{}. For pre-selected protected systems these elements of reality are again expectation values[@AV-prot]. It has been shown that even the two-state vector can be protected and that the weak values can be measured on any single (appropriately protected) system with a good precision[@AV-nyas]. Finally, it has been shown recently[@AMPTV], that adiabatic measurements performed on decaying systems which were post-selected not to decay, yield one of the (non-trivial, i.e. not expectation value) weak values even without specific pre- and post-selection, manifesting again the physical meaning of the weak values. The concept of elements of reality for pre- and post-selected systems fits very well the many-worlds interpretation[@mwi-eve] which I believe is the best available interpretation today[@mwi-V]. It answers in a very convincing way the following difficulty. Consider a present moment of time $t$. We can assign a list of elements of reality and weak-measurement elements of reality based on the quantum state at that time. It includes the eigenvalues of variables for which the quantum state is an eigenstate and expectation values for all variables. In particular, let us assume that a spin $N$ particle has an element of reality $S_x =N$, and consequently it has no element of reality regarding the value of $S_y$, but it has weak-measurement elements of reality $(S_y)_w =0$. Let us assume that at a later time the spin in the $y$ direction was measured and the (very improbable for large $N$) result $S_y =N$ was obtained. If now we assign elements of reality for the time $t$, we see that the list is different. Indeed, we have to add some elements of reality, e.g., $S_y =N$ in our case, and we have to change some weak-measurement elements of reality, e.g., $(S_y)_w =N$. How can we associate different elements of reality for the same moment of time? The answer is natural in the framework of the many worlds interpretation. We discuss here epistemological element of reality of conscious beings. Before the measurement of $S_y$ we considered one world corresponding to a certain quantum state and in this world we had certain elements of reality. The measurement of $S_y$ generated $2N +1$ new worlds corresponding to different outcomes of the measurement. The conscious beings (the experimenters) in these different worlds have, not surprisingly, different sets of elements of reality. The research was supported in part by grant 614/95 of the Basic Research Foundation (administered by the Israel Academy of Sciences and Humanities). [99]{} A. Einstein, B. Podolsky, and N. Rosen, [*Phys. Rev.*]{} [**47**]{}, 777 (1935) M. Jammer, [*The Conceptual Development of Quantum Mechanics,*]{} (McGraw-Hill, New York, 1966). M. Jammer, [*The Philosophy of Quantum Mechanics,*]{} (Wiley, New York, 1974). J.S. Bell, [*Physics*]{} [**1**]{}, 195 (1964). J. von Neumann, [*Mathematical Foundations of Quantum Theory*]{}, (Princeton, University Press, New Jersey (1983). Y.Aharonov and L. Vaidman, [*Phys. Rev.*]{} [**A 41**]{}, 11 (1990). L. Vaidman [*Phys. Rev. Lett.*]{} [**70**]{}, 3369-72 (1993). Y. Aharonov and L. Vaidman [*Jour. Phys.*]{} [**A 24**]{}, 2315-28 (1991). Y. Aharonov and L. Vaidman, [*Phys. Lett.*]{} [**A178**]{}, 38 (1993). Y.Aharonov and L. Vaidman, [*Ann. NY Acad. Sci.*]{} [**755**]{}, 361 (1995). Y.Aharonov, S. Massar, S. Popescu, J. Tollaksen, and L. Vaidman, TAUP 2315-96.
{ "pile_set_name": "ArXiv" }
--- author: - Zhonghui Fan - Siming Liu - Qiang Yuan - Lyndsay Fletcher date: 'Received 8 June 2010 / Accepted 5 July 2010' title: 'Lepton models for TeV emission from SNR RX J1713.7-3946' --- [SNR RX J1713.7-3946 is perhaps one of the best observed shell-type supernova remnants with emissions dominated by energetic particles accelerated near the shock front. The nature of the TeV emission, however, is an issue still open to investigation.]{} [We carried out a systematic study of four lepton models for the TeV emission with the Markov chain Monte Carlo method.]{} [It is shown that current data already give good constraints on the model parameters. Two commonly used parametric models do not appear to fit the observed radio, X-ray, and $\gamma$-ray spectra. Models motivated by diffusive shock acceleration and by stochastic acceleration by compressive waves in the shock downstream give comparably good fits. The former has a sharper spectral cutoff in the hard X-ray band than the latter. Future observations with the [[*HXMT*]{}]{} and [[*NuSTAR*]{}]{} may distinguish these two models.]{} Introduction ============ Energetic particles produce radiations over a broad energy range under typical astrophysical circumstances. To study the energetic particle population in high-energy astrophysical sources, observations over a broad spectral range are needed. Multi-wavelength observations of individual sources usually give sparse spectral data points, which can be fitted reasonably well with a simple power-law or broken power-law function. A featureless power-law distribution implies the lack of characteristic scales or distinct processes in the system. The nature of particle acceleration processes are still a matter of debate nearly a century after the discovery of high-energy particles from the outer space [@b09]. Recent progress in observations of high-energy astrophysical sources has resulted in more detailed spectral coverage, and spectral features start to emerge. These features are produced by the underlying physical processes with well-defined characteristic scales in space, time, and/or energy, and they play essential roles in advancing our understanding of the acceleration mechanism [@b88]. Diffusive astrophysical shocks are considered as one of the most important particle accelerators [@t80; @d83; @b96]. The acceleration of particles by compressive motions of astrophysical flows is also a very generic process [@p88]. Both processes may be important in the acceleration of particles dominating the radiative characteristics of the shock associated with SNR RX J1713.7-394 [@f10]. Detailed X-ray and $\gamma$-ray observations of this source have revealed several features: 1) the X-ray and $\gamma$-ray emissions are well correlated in space [@a06; @p08]; 2) the emission spectrum shows a clear high-energy cutoff in both the X-ray and TeV bands [@a07], and the cutoff in hard X-rays appears to be sharp [@t08], implying a sharp cutoff in the electron distribution producing the observed radio to X-ray spectrum through the synchrotron process; 3) the $\gamma$-ray spectrum has a broad convex shape [@f09]; 4) there are bright X-ray filaments with a width of $\sim0.1$ light year varying on a timescale of about a year [@u07]. @l08 showed that these emissions may be attributed to a single population of electrons with the $\gamma$-rays produced via the inverse Compton scattering of the background photons by relativistic electrons. In light of recent high-resolution observations with [*Suzaku*]{} and [*Fermi*]{}, we use the Markov chain Monte Carlo (MCMC) method to constrain parameters of four possible lepton models. It is shown that simple models with a power-law electron distribution, which cuts off at TeV energies, give a poor fit to either the X-ray or $\gamma$-ray spectrum. Motivated by the mechanism of diffusive shock acceleration (DSA) and stochastic acceleration (SA) by compressive motions, we introduce two alternative models, both of which give acceptable fits to the observed spectra (Section \[models\]). In Section \[dis\], we discuss the implications of these results and future work needed to improve these models. Conclusions are drawn in Section \[con\]. Spectral fits with lepton models {#models} ================================ In the lepton scenario, the radio to X-ray emissions are produced through the synchrotron process of relativistic electrons, and the $\gamma$-ray emission is produced through the inverse Compton scattering of the background radiation fields by the same electron population. If the background magnetic fields are relatively uniform, similar to the background radiation fields, the model can naturally explain the spatial correlation between the X-rays and $\gamma$-rays [@p08]. As pointed out by @l08, the magnetic field required to produce the X-ray flux agrees with what is required to produce an X-ray spectral cutoff by an electron population cutting off in the TeV energy range [@a07; @e10]. There are at least four parameters in a specific model. In this work we employ an MCMC technique suitable for multi-parameter determination to search all the model parameters. The Metropolis-Hastings sampling algorithm is adopted when determining the jump probability from one point to the next in the parameter space [@m03]. The MCMC approach, which is based on the Bayesian statistics, is superior to the grid approach with a more efficient sampling of the parameter space of interest, especially for high dimensions. The algorithm ensures that the probability density functions (PDF) of model parameters can be asymptotically approached with the number density of samples. Starting with an initial parameter set ${\bf P_0}$, which should lie in a reasonable parameter space, one obtains the likelihood function of ${\bf P_0}$: $\mathcal{L}_0({\bf P_0})\propto \exp \left(-{\chi^2({\bf P_0})}/{2}\right)$. Then another parameter set ${\bf P_1}$ is randomly selected with the corresponding likelihood function $\mathcal{L}_1$. This parameter set is accepted with the probability $\alpha=\min\{1,{\mathcal{L}_1}/{\mathcal{L}_0}\}$. If the new point is accepted, it becomes the starting point for the next step, otherwise, we reset ${\bf P_1}={\bf P_0}$. This procedure is repeated to derive a parameter chain that determines the PDF of model parameters. More details of this procedure can be found in Neil (1993), Gamerman (1997), and MacKay (2003). [**I:**]{} Without detailed modeling of the particle acceleration processes, a power-law electron distribution with an exponential high-energy cutoff is often assumed to fit the observed spectrum: $$N(\gamma)\propto \gamma^{-p}\exp({-\gamma/\gamma_c})$$ where $p$ is the spectral index, $\gamma_c$ the Lorentz factor of the high-energy cutoff $E_c=\gamma_c m_ec^2$, $m_e$ and $c$ are the electron mass and the speed of light, respectively. The integration of $N(\gamma)$ over the Lorentz factor $\gamma$ gives the particle number density. It has been shown that this model gives a sufficient fit to previous observations of SNR RX J1713.7-3946 when the interstellar radiation field is considered [@p06]; however, more detailed studies suggest that such a model systematically underestimates the photon flux at the low-energy end of HESS observations [@t08; @za10]. In this paper, we adopt the interstellar radiation field spectrum of @p06. There are then four model parameters. Besides the magnetic field $B$, $p$, and $\gamma_c$, we need a normalization for the relativistic electrons $E_{\rm {tot}}$, which is the total energy of electrons with $\gamma>10$. Figures \[spec\] and \[prob\] give the best fit of this model to the data and the probability density of the model parameters. All four model parameters are well constrained. However, detailed examination of residuals of the spectral fit shows that there are several data points below 1 and near 10 TeV with normalized residuals significantly greater than 3. [**II:**]{} @l08 suggested that the fit to the $\gamma$-ray spectrum can be improved by considering a more gradual cutoff in the electron distribution: $$N(\gamma)\propto \gamma^{-p}\exp{-(\gamma/\gamma_c)^{1/2}}\,.$$ Several mechanisms can cause variations in the shape of the high-energy cutoff [@b88; @bld06; @za07; @b10]. However, given the sharp cutoff of the X-ray spectrum, @t08 claim that the electron distribution must have a very steep cutoff. Indeed, by using their data and performing the MCMC fit, we obtain radio fluxes nearly one order of magnitude below the observed values. The model also significantly overestimates the hard X-ray fluxes. To increase the weight of the radio and $\gamma$-ray data in the model fit, we artificially increase the errors of the X-ray data from [*Suzaku*]{} by a factor of 2 [^1] (Figs. \[spec\] and $\ref{prob}$). The normalized residuals are indeed improved significantly with a reduced $\chi^2$ of 1.2, in agreement with Liu et al. (2008), where cruder X-ray data from [*ASCA*]{} observations were used. However, as emphasized by @t08, the sharp cutoff in the X-ray spectrum is significant and the error estimate of the [*Suzaku*]{} observation appears to be robust [@tak08]. This simple model therefore cannot fit the X-ray spectrum. [**III:**]{} The particle distribution is determined by the injection process and the spatial diffusion coefficient $\chi$ in the DSA model. The time-dependent electron distribution with a constant injection rate may be approximated as [@fd83; @b88; @d91] $$N(\gamma)\propto \gamma^{-p} \exp [-9\chi (\gamma)/U^2{\rm T_{\rm life}}]\,, \label{f2}$$ where $T_{\rm life}$ is the supernova lifetime since the onset of particle acceleration and $U$ the shock speed. A very sharp high-energy cutoff can be produced when gyro-radii of high-energy particles exceed the characteristic length of the magnetic field in the background plasma $l_d$. Particles at such high energies do not perform gyro-motions around the chaotic magnetic field and instead interact randomly with the background fields. The spatial diffusion coefficient of these particles increases as the square of the relativistic Lorentz factor in the absence of waves on scales greater than $l_d$. Consequently, they can readily cross the acceleration site without interactions. We assume a spatial diffusion coefficient $$\chi(\gamma) = \eta \left\{ \begin{array}{ll} {(\gamma/\gamma_c)^{\beta} \gamma_c m_ec^3/(3eB)}\ \ \ \ \ &{\rm for}\;\;\;\; \gamma \leq \gamma_c \,, \\ {\gamma^2 m_ec^3/(3eB\gamma_c)}\ \ \ \ \ &{\rm for}\;\;\;\; \gamma \geq \gamma_c \,,\\ \end{array} \right. \label{dsaxi}$$ where $e$ is the elementary charge, $\eta>1$ depends on the turbulence intensity, $\beta=1/2$ and 1/3 for particle interactions with a spectrum of plasma waves following the @k65 and @k41 phenomenology for the turbulence spectral evolution, respectively. The characteristic length of the magnetic field is then given by the gyro-radius of electrons at $\gamma_c$: $l_d= \gamma_c m_e c^2/e B$. Compared with the previous two models, an extra parameter is introduced here: $\eta/T_{\rm life}U^2$, which is proportional to the electron acceleration timescale. Figures \[spec\] and \[prob\] show the results with $\beta=1/2$. The probability density of $\eta(1.6 {\rm kyr}/T_{\rm life})(4 {\rm Mm\ s}^{-1}/U)^2$ peaks near a typical value of 12. Since $\eta$ has to be greater than 1, we have $T_{\rm life}U^2>0.02 c$ lyr. The gyro-radius of electrons at the cutoff Lorentz factor $\gamma_c\sim 2\times 10^8$ is $\sim$0.01 pc comparable to the width of the observed highly variable X-ray filaments. The X-ray variability may then be attributed to spatial diffusion of X-ray emitting electrons from a high-density region formed presumably via an intermittent process. It is interesting to note that the best fit spectral index $p=1.92$ is less than $2$, implying acceleration by multiple shocks or strong nonlinear effects [@za10]. The model with $\beta=1/3$ gives a similar fit with a slightly higher value of $\eta$ and sharper hard X-ray cutoff. [**IV:**]{} @f10 have carried out detailed modeling of electron acceleration by compressive motions in the shock downstream and argued that it might also explain these observations. Here we follow the same treatment and also consider the effect of incompressive motions, which can dominate the spatial diffusion and therefore the escape of energetic electrons from the acceleration region. The acceleration rate is given by [@bt93] $$\tau_{ac}^{-1}= D_{\gamma\gamma}/\gamma^2 = {4\pi \chi\over 9}\int_{k_m}^{k_d}dk {k^4 S(k)\over v_F^2+ \chi^2k^2}\,, \label{tac}$$ where $\chi$, $k$, $v_F$, and $S$ are the spatial diffusion coefficient, the wave-number, speed, and intensity of compressive waves, respectively. The overall wave intensity $\int S(k) 4\pi k^2dk= \xi v_F^2$ with $\xi<1$ for subsonic turbulence. [^2] The turbulence is generated on a scale $L$ near the shock front, and $l_d$ corresponds to the scale where the turbulence speed is comparable to the Alfvén speed $v_{\rm A}$. Relativistic electrons with a gyro-radius less than $l_d$ have a mean-free path comparable to $l_d$ since waves on smaller scales are subject to strong transit time damping by thermal background ions. We then have the diffusion coefficient $$\chi(\gamma,x) = \left\{ \begin{array}{ll} {cl_d(x)/3}\ \ \ \ \ &{\rm for}\;\;\;\; \gamma \leq \gamma_c(x) \,, \\ {\gamma m_ec^3/(3 eB)}\ \ \ \ \ &{\rm for}\;\;\;\; \gamma \geq \gamma_c(x) \,,\\ \end{array} \right. \label{saxi}$$ where $x$ indicates the spatial location in the downstream, $l_d = \gamma_c m_ec^2/eB$, and we have assumed Bohm diffusion for $\gamma>\gamma_c$. Due to gradual decay of the turbulence intensity, related quantities evolve with $x$ in the downstream. Equation (\[saxi\]) is quite different from equation (\[dsaxi\]), where particle scattering by the background turbulence is parameterized without considering details of the turbulence structure. Considering the presence of strong turbulence with $u>v_{\rm A}$ between the scales of $l_d$ and $L$, Bohm diffusion may be appropriate, and as we show below, faster diffusion may not be able to accelerate electrons efficiently to account for the observations. The intensity of the incompressive motions $T(k)$ can be determined from the energy conservation: $\int [2T(k)+S(k)]4\pi k^2dk=3u^2$, where $3u^2/2$ is the total turbulence energy density per unit mass. Further downstream, where $v_F$ is high ($\ge 6^{1/2} u$), the incompressive motions vanish completely. To derive the escape time $T_{esc}$ of electrons from the acceleration region, we need to consider the enhancement of the spatial diffusion coefficient by the turbulence motions. According to Eq. (4.30) in @bt93, the effective diffusion coefficient $\chi_*$ can be derived from $$\begin{aligned} \chi_* &=& \chi + \frac{8\pi}{3\chi_*}\int_{k_m}^{k_d} T(k) dk + \\\nonumber &&\frac{4\pi \chi_*}{3}\int_{k_m}^{k_d} \left[\frac{1}{k^2\chi_*^2+v_F^2}-\frac{2(k^2\chi_*^2-v_F^2)} {(k^2\chi_*^2+v_F^2)^2}\right]k^2 S(k) dk\,,\end{aligned}$$ where we have ignored the decay of incompressive motions. Then we have $$T_{esc} = L^2/\chi_*.$$ With the above treatment, we can study particle acceleration by compressive motions all the way to the shock front, which is different from the case studied by @f10, where only the subsonic phase with $v_F>u$ is considered. Near the shock front, we assume $\xi =1/2$ so that equation (\[tac\]) is applicable. Since the compressive wave intensity is lower than in @f10 and the acceleration timescale of TeV electrons becomes comparable to the supernova lifetime, a time-dependent treatment is necessary. [^3] Both the acceleration and escape timescale explicitly depend on the diffusion coefficient $\chi$, which is independent of $\gamma$ for $\gamma<\gamma_c$. Assuming that electrons are injected at the Lorentz factor $\gamma_{inj}=10$ with a constant rate, the time-dependent results of @bld06 ([-@bld06]; see also @cs84 [@pp95]) can be used to describe the evolution of the electron distribution function at energies below $\gamma_c$. Near the cutoff energy $\gamma_c$, the nontrivial dependence of the acceleration and escape timescales on the Lorentz factor demand numerical solutions. We assume the time-dependent distribution function has a high-energy cutoff given by $\exp[-(\tau_{ac}/T_{esc})^{1/2}]$ [For details, see @f10]. Figures (\[spec\]) and (\[prob\]) show the results for $U=4000$ km/s with the Kraichnan phenomenology. The quality of this fit is comparable to the DSA model. The deduced model parameters also agree with those obtained in @f10. ![Model fit to the spectrum of SNR RX J1713.7-3946. The TeV data is from @a07, the [*Fermi*]{} data is from @f09, and the X-ray data is from @t08 (Thanks to Jun Fang). The thick solid, dashed line and the thin solid, dashed line in the upper panel correspond to the diffusive shock, stochastic acceleration model, and the model with an exponential, more gradual cutoff in the electron distribution, respectively. The low and high-energy spectral hump are produced by relativistic electrons through the synchrotron and inverse Comptonization process, respectively. The lower panels show the normalized residuals. From top to bottom, they are for the model with an exponential and more gradual cutoff, the DSA, and SA model. For the model with a gradual cutoff, the X-ray errors have been artificially increased by a factor of 2 to improve the fit. The reduced $\chi^2$ of these fits are indicated in the top-left corner. See the text for details. The instrumental sensitivity of [*NuSTAR*]{} and [*HXMT*]{} are indicated by the dotted and dot-dashed lines, respectively. \[spec\]](f1.eps){width="9"} ![Probability density function of model parameters normalized at the peak value. Different line types correspond to different models as explained in Figure \[spec\] and the text. \[prob\]](f2.eps){width="9"} Discussion {#dis} ========== The SA model (Fig. \[spec\]) predicts a higher hard X-ray flux than the DSA model since the diffusion coefficient of the DSA model has a stronger dependence on the electron energies leading to a sharper high-energy cutoff. The electron distribution in the SA model includes contributions from broad regions in the downstream, which also makes the overall electron distribution broader. Future observations with [*NuSTAR*]{} and [*HXMT*]{} may be able to distinguish these two models. In actual turbulent astrophysical shocks, both the DSA and SA mechanisms may contribute to the electron acceleration [@a90; @bt93]. The distinction between the two resides in the dissipation structure [@b88]. Supersonic shock fronts dominate the acceleration in the DSA model. The SA model dominates in the subsonic phase. It is possible that the shock downstream of SNR RX 1713.7-3946 has both a supersonic and subsonic phase turbulence with the former closer to the shock front. The electron acceleration is therefore a continuous process in the turbulent downstream. In the SA model of this paper, we intentionally remove the DSA by requiring that the compressive waves be subsonic in the downstream. A self-consistent treatment of the turbulence evolution in the shock downstream is needed for particle acceleration near astrophysical shock fronts. All the models appear to systematically underproduce $\gamma$-ray flux in the [*Fermi*]{} energy band and overproduce flux near 10 TeV. The latter could be caused by inhomogeneities of the background magnetic field, which implies lower cutoff energy of the electron distribution and less high-energy $\gamma$-ray flux [@t08]. The former requires a broader electron distribution from the GeV to TeV energy range. A combination of DSA and SA model with the former dominating the higher energy particle acceleration and the latter enhancing the lower energy ones may address this issue. Such a scenario is possible if both supersonic and subsonic turbulence are produced by the supernova shock [@b88; @bt93]. Contributions to GeV–TeV emission from decay of neutral pions produced by inelastic collisions of relativistic protons with the background ions may also explain these residuals. However, as shown by @e10 and @kw08, relativistic protons can not be the dominant TeV emission component [@za10]. Including this component is beyond the scope of the current investigation but may give a good constraint on the relative acceleration efficiency of relativistic electrons and protons [@kw08]. Conclusions {#con} =========== With recent high spectral-resolution observations of SNR RX 1713.7-3946 in X-rays and $\gamma$-rays, we show that, in general, models with simple parametric descriptions of energetic particle distribution give a poor fit to the broadband spectrum. Details of relativistic electron acceleration may be probed with advanced models. Although both the DSA and SA model give acceptable fits to the spectrum, the SA model predicts a higher hard X-ray flux than the DSA model, which may be tested by future observations. The spatial distribution of energetic electrons in the SA and DSA models is also different, and the source structure can be used to distinguish these two models as well. The results agree with the scenario where energetic electrons are accelerated by interacting diffusively with a turbulent electromagnetic field and produce nearly all of the observed emissions from the shell of SNR RX 1713.7-3946. Turbulence evolution at astrophysical shocks appears to play a key role in advancing our understanding of particle acceleration processes. This work is supported in part by the SOLAIRE research and training network at the University of Glasgow (MTRN-CT-2006-035484), the National Science Foundation of China (grants 10963004 and 10778702), Yunnan Provincial Science Foundation of China (grant 2008CD061) and SRFDP of China (grant 20095301120006). SL thanks Eduard Kontar for helpful discussion. QY thanks Jie Liu for helping to develop the MCMC code adapted from the COSMOMC code of @l02. Achterberg, A. 1990, A&A, 231, 251 Aharonian, F. A., et al. 2006, A&A, 449, 223 Aharonian, F. A., et al. 2007, A&A, 464, 235 Becker, P. A., Le, T., Demer, C. D. 2006, ApJ, 647, 539B Berezhko, E. G. 1996, Astroparticle Physics, 5, 367 Berezhko, E. G., & Krymsky, G. F. 1988, Sov. Phys. Usp., 31, 27 Blasi, P. 2010, MNRAS, 402, 2807 Butt, Y. M. 2009, Nature, 460, 701 Bykov, A. M., & Toptygin, I. N. 1993, Phys. Usp., 36, 1020 Cassam-Chenaï. G., Decourchelle, A., Ballet, J., et al. 2004, A&A, 427, 199 Cowsik, R., & Sarkar, S. 1984, MNRAS, 207, 745 Drury, L. 1983, SSRv, 36, 57 Drury, L. 1991, MNRAS, 251, 340 Ellison, D. C., Patnaude, D. J., Slane, P., & Raymond, J. 2010, ApJ, 712, 287 Fan, Z. H., Liu, S. M., Fryer, C. L. 2010, MNRAS, doi:10.1111/j.1365-2966.2010.16767.x Forman, M. A., & Drury, L. O. 1983, ICRC, 2, 267 Funk, S. 2009, Fermi Symposium. Gamerman, D., [*Markov Chain Monte Carlo: Stochastic Simulation for Bayesian Inference*]{}, Chapman and Hall, London, 1997 Katz, B., & Waxman, E. 2008, JCAP, 01, 018 Kolmogorov A., 1941, Dokl. Akad. Nauk SSSR, 30, 301 Kraichnan 1965, Phys. Fluids, 8, 1385 Lewis, A., & Bridle, S. 2002, Phys. Rev. D 66, 103511 Liu, S. M., Fan, Z. H., Fryer, C. L., Wang, J. M., Li, H., 2008, ApJ, 683, L163 MacKay, D., [*Information Theory, Inference, and Learning Algorithms*]{}, Cambridge University Press, 2003 Neal, R. M., [*Probabilistic Inference Using Markov Chain Monte Carlo Methods*]{}, Technical Report CRG-TR-93-1, Department of Computer Science, University of Toronto, 1993 Park, B. T., & Petrosian, V. 1995, ApJ, 446, 699 Plaga, R. 2008, NewA, 13, 73 Porter, T. A., Moskalenko, I. V., & Strong, A. W. 2006, ApJ, 648, L29 Ptuskin, V. S. 1988, Soviet Astron. Lett., 14, 255 Takahashi, T., et al. 2008, PASJ, 60, 131 Tanaka, T., et al. 2008, ApJ, 685, 988 Toptygin, I. N. 1980, SSRv, 26, 157 Uchiyama, Y., et al. 2007, Nature, 449, 576 Zirakashvili, V. N., & Aharonian, F. A. 2007, A&A, 465, 695 Zirakashvili, V. N., & Aharonian, F. A. 2010, ApJ, 708, 965 [^1]: Given the complexity of the instrumental background and calibration, and unknowns, such as the estimated cosmic X-ray background, such an artificial increase may well be reasonable. [^2]: Here we adopt the formula given by @bt93, which gives an acceleration rate a factor of 2 lower than that of @p88. [^3]: At the point, where $v_F=u$, the compressive wave intensity used here is 3 times lower than in @f10.
{ "pile_set_name": "ArXiv" }
--- abstract: 'By carrying out appropriate continuous quantum measurements with a family of projection operators, a unitary channel can be approximated in an arbitrary precision in the trace norm sense. In particular, the quantum Zeno effect is described as an application. In the case of an infinite dimension, although the von Neumann entropy is not necessarily continuous, the difference of the entropies between the states, as mentioned above, can be arbitrarily made small under some conditions.' author: - | Toru Fuda[^1]\ Department of Mathematics\ Hokkaido University\ Sapporo 060-0810\ Japan title: | **Convergence Conditions of Mixed States and their von Neumann Entropy\ in Continuous Quantum Measurements** --- Introduction ============ The quantum Zeno effect (QZE) is a quantum effect which was shown by Misra and Sudarshan in [@MS]. This effect demonstrates that, in quantum mechanics, continuous measurements can freeze a state. Of course, this effect is peculiar to quantum mechanics. Such an effect is not observed in classical mechanics. The QZE has been extensively investigated by many researchers since its discovery. Recently, some general mathematical aspects of quantum Zeno effect were investigated in [@AF]. In particular, continuous measurements of a state along a certain curve in a Hilbert space were considered. Roughly speaking, continuous measurements made along a curve prescribed in advance change the initial state to the final state with probability $1$. This fact includes the QZE as a special case. However, in the paper [@AF], it is assumed that states under consideration are vector states. In this paper, we show that a result similar to one in [@AF] holds with respect to mixed states too. By considering a mixed state, its von Neumann entropy can also be considered. In the case where the Hilbert space under consideration is infinite dimensional, the von Neumann entropy is not necessarily continuous with respect to the trace norm. Hence, by continuous measurements, even if the initial state converges to the final state in the trace norm sense, it does not always mean that the von Neumann entropy converges too. Moreover, the set of density operators with finite entropy is a first category [@We]. Hence, it is meaningful to investigate convergence conditions of the von Neumann entropy in our continuous measurements. In Section 2, we begin with defining the “continuous quantum measurements" as a certain type of quantum channel. We use two types of quantum channels and a combination of them. By doing so, a concept of “continuous quantum measurements" are defined clearly. We consider conditions for pointwise convergence and trace norm convergence. We apply obtained results to the QZE. In Section 3, we consider the von Neumann entropy in infinite dimension. We show that the convergence conditions of the von Neumann entropy in continuous quantum measurement which considered in Section 2. Here, Simon’s convergence theorem [@LR] plays a central role. Continuous measurements for mixed states ======================================== Preliminaries ------------- Let $\mathcal{H}$ be a separable Hilbert space of state vectors of a quantum system $\mathcal{S}$. We denote the inner product and the norm of ${{\cal H}}$ by ${\left\langle}\,\cdot\,,\,\cdot\,{\right\rangle}$ (anti-linear in the first variable and linear in the second) and $\|\,\cdot\,\|$, respectively. Let $d (\leq \infty)$ be the dimension of $\mathcal{H}$. We denote all bounded linear operators, all compact operators, all trace-class operators, all density operators, and all unitary operators on $\mathcal{H}$ by $\mathfrak{B}(\mathcal{H}), \mathfrak{C}(\mathcal{H}), \mathfrak{T}(\mathcal{H}), \mathfrak{S}(\mathcal{H}), and \ \mathfrak{U}(\mathcal{H})$, respectively. A mixed state of $\mathcal{S}$ is represented as an element of $\mathfrak{S}(\mathcal{H})$. We denote the trace norm by $\|\cdot \|_1:=\mathrm{Tr}|\cdot|$. The Hamiltonian of the quantum system $\mathcal{S}$ is given by a self-adjoint operator $H$ which is time independent. The domain of $H$ is denoted as $D(H)$. Let us consider the following two maps on $\mathfrak{S}(\mathcal{H})$: 1. [(**Unitary channel**)]{} Let $U$ be a unitary operator on $\mathcal{H}$ and $\mathcal{E}_U$ be a map on $\mathfrak{S}(\mathcal{H})$ which is given by $$\mathcal{E}_U\rho := U\rho U^*, \ \forall \rho \in \mathfrak{S}(\mathcal{H}).$$ In particular, in the case $U=e^{-itH}\ (t\in \mathbb{R})$, we denote $\mathcal{E}_{e^{-itH}}$ by $\mathcal{E}_t$. 2. [(**Projection channel**)]{} Let $\mathfrak{P}:=\{P_n\}_n$ be a family of projection operators on $\mathcal{H}$ with $P_m\perp P_n\ (m \neq n), I=\sum_nP_n$, and $\mathcal{E}_{\mathfrak{P}}$ be a map on $\mathfrak{S}(\mathcal{H})$ which is given by $$\mathcal{E}_{\mathfrak{P}}\rho := \sum_n P_n\rho P_n, \ \forall \rho \in \mathfrak{S}(\mathcal{H}).$$ Now, consider a state $\rho \in \mathfrak{S}(\mathcal{H})$ fixed and suppose that one of the Schatten decompositions is given by $$\rho = \sum_{n=1}^d \lambda _n|\Psi_n \rangle \langle \Psi_n|, \label{schatten}$$ where, for all $\Psi , \Phi \in \mathcal{H}$, we denote the operator $\langle \Psi , \cdot \ \rangle \Phi$ by $| \Phi \rangle \langle \Psi |$. In (\[schatten\]), we allow $\lambda_n = 0$ to take $\Psi_n$ such that $\{ \Psi_n\}_{n=1}^d$ is a complete orthonormal system (CONS). We remark that it is not necessarily $\lambda_n \geq \lambda_{n+1}$ in this representation. Let us consider a time interval $[0,\tau]$ with $\tau>0$. For the decomposition (\[schatten\]), consider a CONS of $\mathcal{H}$ denoted by $\{\Psi_n(t)\}_{n=1}^d$ which is parametrized by $t\in [0, \tau]$ with $\Psi_n(0)=\Psi_n \ (1\leq \forall n \leq d)$. If $n\in \mathbb{N}$ is fixed, then $\Psi_n (\cdot)$ is a map from $[0, \tau]$ to $\mathcal{H}$. We define $$\mathfrak{P}(t) :=\{|\Psi_n (t) \rangle \langle \Psi_n (t)| \}_{n=1}^d, \quad (t\in [0, \tau]).$$ Let $\Delta: t_0, t_1,\cdots, t_N$ ($t_j \in [0,\tau], j=0,\cdots,N$) be an arbitrary partition of the interval $[0,\tau]$: $$0=t_0 < t_1 < \cdots < t_{N-1} < t_N = \tau.$$ We set $$\Delta _k :=t_k -t_{k-1}, \quad (k=1,\cdots ,N), \qquad |\Delta| :=\max_{1 \leq k \leq N}\Delta _k,$$ and define $$\begin{aligned} \rho_{\Delta}(\tau) :=\mathcal{E}_{\mathfrak{P}(t_N)} \circ \mathcal{E}_{\Delta_N}\circ \mathcal{E}_{\mathfrak{P}(t_{N-1})} \circ \mathcal{E}_{\Delta_{N-1}} \circ \cdots \circ \mathcal{E}_{\mathfrak{P}(t_1)}\circ \mathcal{E}_{\Delta_1}\rho.\end{aligned}$$ In the context of quantum mechanics where $\rho_{\Delta}(\tau)$ is interpreted as the posterior state that, in the successive measurements at time $t_1 , \cdots , t_N$ by using the family of projection operators $\mathfrak{P}(t_1), \cdots , \mathfrak{P}(t_N)$, respectively. We remark that $\rho_{\Delta}(\tau)$ is dependent on the form of decomposition (\[schatten\]). If $\rho_{\Delta}(\tau)$ converges with respect to $|\Delta|\to 0$ in a certain sense, we call such a measurements of a series “continuous quantum measurements". By direct computations, we have $$\rho_{\Delta}(\tau) = \sum_{k}\lambda_{\Delta, k} \left|\Psi_k(\tau) \rangle \langle \Psi_k(\tau)\right|$$ with $$\begin{aligned} \lambda_{\Delta, k} :=\sum_{k_0,\cdots,k_{N-1}} \lambda_{k_0}\prod _{j=1}^N\left|\langle \Psi_{k_j}(t_j), e^{-i\Delta_j H} \Psi_{k_{j-1}}(t_{j-1})\rangle \right|^2, \quad (k_N = k).\end{aligned}$$ Pointwise convergence --------------------- Let us consider a convergence condition of $\lambda_{\Delta, k}$ in the case $|\Delta |\to 0$. Let $$\begin{aligned} \gamma_{\Delta, k} &:=&\prod_{j=1}^{N}\left|\langle \Psi_{k}(t_j), e^{-i\Delta_j H} \Psi_{k}(t_{j-1})\rangle \right|^2 ,\\ \epsilon_{\Delta, k} &:=&\sum_{\stackrel{k_0, \cdots, k_{N-1}} {\exists l\in\{0,\cdots,N-1\}, k_l\neq k}} \lambda_{k_0}\prod _{j=1}^N \left|\langle \Psi_{k_j}(t_j), e^{-i\Delta_j H} \Psi_{k_{j-1}}(t_{j-1})\rangle \right|^2,\end{aligned}$$ so that $$\lambda_{\Delta, k} = \lambda_k \gamma_{\Delta, k} +\epsilon_{\Delta, k}.\label{decomep}$$ \[saisho\] Assume that there exists $k\in \mathbb{N}$ such that the following conditions hold: $$\begin{aligned} &&\forall \lambda \in [0, \tau],\quad \Psi_k(\lambda) \in D(H), \label{A0} \\ &&\xi_k := \sup_{0\leq \lambda\leq \tau} \|H\Psi_k(\lambda)\|<\infty, \label{A1} \\ &&\eta_k :=\sup_{\stackrel{\lambda, \nu \in [0, \tau]} {\lambda \neq \nu}}\frac{\|\Psi_k(\lambda)-\Psi_k(\nu)\|} {|\lambda - \nu|}<\infty,\label{A2}\\ &&\lim_{|\Delta|\to 0}\sum_{j=1}^N {{\rm Re}\,}{\left\langle}\Psi_k(t_j)-\Psi_k(t_{j-1}), \Psi_k(t_{j-1}){\right\rangle}= 0. \label{A3} \end{aligned}$$ Then we have $$\lim_{|\Delta|\to 0}\lambda_{\Delta, k}=\lambda_k \label{A4}.$$ Condition (\[A2\]) implies that $\|\Psi_k(\lambda)-\Psi_k(\nu)\|\leq \eta_k |\lambda-\mu|, \forall \lambda,\mu\in [0,\tau]$ (Lipschitz continuity). In particular, $\Psi_k(\cdot)$ is strongly continuous, so that the mapping $\Psi_k(\cdot):[0,t]\to{{\cal H}}$ is a curve in ${{\cal H}}$. [*Proof*]{}. By using [@AF THEOREM 4.2], the assumptions (\[A0\])–(\[A3\]) imply that $$\lim_{|\Delta |\to 0}\gamma_{\Delta, k}=1. \label{gamma}$$ On the other hand, we can estimate $\epsilon_{\Delta, k}$ as follows. $$\begin{aligned} \epsilon_{\Delta, k} &=&\sum_{l=0}^{N-1}\sum_{\stackrel{k_0, \cdots, k_{N-1}} {\forall i > l, k_i = k , k_l \neq k}} \lambda_{k_0}\prod _{j=1}^N\left|\langle \Psi_{k_j}(t_j), e^{-i\Delta_j H} \Psi_{k_{j-1}}(t_{j-1})\rangle \right|^2\\ &=&\sum_{l=0}^{N-1}\left\{ \prod_{j=l+2}^{N} \left|\langle \Psi_{k}(t_j), e^{-i\Delta_j H} \Psi_{k}(t_{j-1})\rangle \right|^2 \sum_{k_l, k_l \neq k} \left|\langle \Psi_{k}(t_{l+1}), e^{-i\Delta_{l+1} H} \Psi_{k_l}(t_{l})\rangle \right|^2\right. \notag \\ &&\left. \times \sum_{k_{l-1}}\left|\langle \Psi_{k_l}(t_l), e^{-i\Delta_l H} \Psi_{k_{l-1}}(t_{l-1})\rangle \right|^2 \times \cdots \times \sum_{k_{0}}\left|\langle \Psi_{k_1}(t_1), e^{-i\Delta_1 H} \Psi_{k_{0}}(t_{0})\rangle \right|^2\lambda_{k_0} \right\}, \notag \\ \label{epsilon}\end{aligned}$$ in the case where $l=0, N-1$, $\{\ \cdots\}$ in (\[epsilon\]) is given by $$\begin{aligned} &&\prod_{j=2}^{N} \left|\langle \Psi_{k}(t_j), e^{-i\Delta_j H} \Psi_{k}(t_{j-1})\rangle \right|^2 \sum_{k_0, k_0 \neq k} \left|\langle \Psi_{k_1}(t_1), e^{-i\Delta_1 H} \Psi_{k_{0}}(t_{0})\rangle \right|^2\lambda_{k_0}, \\ &&\sum_{k_{N-1}, k_{N-1} \neq k} \left|\langle \Psi_{k}(t_{l+1}), e^{-i\Delta_{l+1} H} \Psi_{k_l}(t_{l})\rangle \right|^2 \sum_{k_{N-2}}\left|\langle \Psi_{k_{N-1}}(t_{N-1}), e^{-i\Delta_{N-1} H} \Psi_{k_{l-2}}(t_{l-2})\rangle \right|^2 \notag \\ &&\cdots \times \sum_{k_{0}}\left|\langle \Psi_{k_1}(t_1), e^{-i\Delta_1 H} \Psi_{k_{0}}(t_{0})\rangle \right|^2\lambda_{k_0},\end{aligned}$$ respectively. By the Schwarz inequality, we have $$\begin{aligned} \prod_{j=l+2}^{N} \left|\langle \Psi_{k}(t_j), e^{-i\Delta_j H} \Psi_{k}(t_{j-1})\rangle \right|^2 &\leq& \prod_{j=l+2}^{N}\|\Psi_{k}(t_j)\|^2\cdot \| e^{-i\Delta_j H} \Psi_{k}(t_{j-1})\|^2 \\ &\leq& 1, \qquad \forall l \in \{0, \cdots ,N-2\}.\end{aligned}$$ For all $l\geq 1$, $$\begin{aligned} &&\sum_{k_{l-1}}\left|\langle \Psi_{k_l}(t_l), e^{-i\Delta_l H} \Psi_{k_{l-1}}(t_{l-1})\rangle \right|^2 \times \cdots \times \sum_{k_{0}}\left|\langle \Psi_{k_1}(t_1), e^{-i\Delta_1 H} \Psi_{k_{0}}(t_{0})\rangle \right|^2\lambda_{k_0}\\ &\leq& \sum_{k_{l-1}}\left|\langle \Psi_{k_l}(t_l), e^{-i\Delta_l H} \Psi_{k_{l-1}}(t_{l-1})\rangle \right|^2 \times \cdots \times \sum_{k_{0}}\left|\langle e^{i\Delta_1 H} \Psi_{k_1}(t_1), \Psi_{k_{0}}(t_{0})\rangle \right|^2\\ &\leq&\sum_{k_{l-1}}\left|\langle \Psi_{k_l}(t_l), e^{-i\Delta_l H} \Psi_{k_{l-1}}(t_{l-1})\rangle \right|^2 \times \cdots \times \| e^{i\Delta_1 H} \Psi_{k_1}(t_1)\|^2\\ &\leq&\dots \leq 1.\end{aligned}$$ Thus (\[epsilon\]) implies that $$\epsilon_{\Delta, k}\leq \sum_{l=0}^{N-1} \sum_{k_l, k_l \neq k} \left|\langle \Psi_{k}(t_{l+1}), e^{-i\Delta_{l+1} H} \Psi_{k_l}(t_{l})\rangle \right|^2. \label{hyouka215}$$ In the case where $k_l\neq k$, we have ${\left\langle}\Psi_k(t_l), \Psi_{k_l}(t_l){\right\rangle}=0$. Hence $$\begin{aligned} &&\sum_{k_l, k_l \neq k} \left|\langle \Psi_{k}(t_{l+1}), e^{-i\Delta_{l+1} H} \Psi_{k_l}(t_{l})\rangle \right|^2\notag \\ &=& \sum_{k_l, k_l \neq k} \left| \langle \Psi_{k}(t_{l+1}), (e^{-i\Delta_{l+1} H}-1) \Psi_{k_l}(t_{l})\rangle + \langle \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l}), \Psi_{k_l}(t_{l})\rangle \right|^2\notag \\ &\leq& 2\sum_{k_l, k_l \neq k}\left\{ \left| \langle (e^{i\Delta_{l+1} H}-1)\Psi_{k}(t_{l+1}), \Psi_{k_l}(t_{l})\rangle \right|^2 + \left| \langle \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l}), \Psi_{k_l}(t_{l})\rangle \right|^2 \right\}\notag \\ &\leq& 2\left\{ \| (e^{i\Delta_{l+1} H}-1)\Psi_{k}(t_{l+1})\|^2 + \| \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l})\|^2 \right\}.\label{hyouka_o}\end{aligned}$$ Let $E_H(\cdot)$ be the spectral measure of Hamiltonian $H$. By the spectral theorem, we have $$\begin{aligned} \| (e^{i\Delta_{l+1} H}-1)\Psi_{k}(t_{l+1})\|^2 &=&\int_{\mathbb{R}}|e^{i\Delta_{l+1} x}-1|^2d\|E_H(x)\Psi_{k}(t_{l+1})\|^2 \notag \\ &\leq&\int_{\mathbb{R}}\Delta_{l+1}^2x^2d\|E_H(x)\Psi_{k}(t_{l+1})\|^2 \notag \\ &\leq&\Delta_{l+1}^2\|H\Psi_{k}(t_{l+1})\|^2.\label{hyouka_hami}\end{aligned}$$ The assumptions (\[A0\])–(\[A2\]) imply that $$\|H\Psi_{k}(t_{l+1})\|^2\leq \xi_k^2, \quad \| \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l})\|^2\leq \Delta_{l+1}^2\eta_k^2. \label{hyouka_la}$$ Therefore, (\[hyouka215\]), (\[hyouka\_o\]), (\[hyouka\_hami\]) and (\[hyouka\_la\]) implies that $$\begin{aligned} \epsilon_{\Delta, k} &\leq& 2\sum_{l=0}^{N-1}\left\{ \| (e^{i\Delta_{l+1} H}-1)\Psi_{k}(t_{l+1})\|^2 + \| \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l})\|^2 \right\}\notag \\ &\leq& 2\sum_{l=0}^{N-1}\left\{ \Delta_{l+1}^2 \|H\Psi_{k}(t_{l+1})\|^2+\| \Psi_{k}(t_{l+1})-\Psi_{k}(t_{l})\|^2\right\}\notag \\ &\leq &2(\xi_k^2+\eta_k^2)\sum_{l=1}^N\Delta_l^2. \label{hyoukaep}\end{aligned}$$ By [@AF LEMMA 2.2], $$\lim_{|\Delta|\to 0}\sum_{l=1}^N\Delta_l^2=0.$$ Thus (\[hyoukaep\]) implies that $\lim_{|\Delta |\to 0}\epsilon_{\Delta, k}=0$. Hence, by (\[decomep\]) and (\[gamma\]), we obtain (\[A4\]) Assume that the conditions of Theorem \[saisho\] hold. Let $a>1$ be a constant and take $|\Delta |$ such that $$(\xi_k^2+2\xi_k\eta_k)|\Delta |^2 + 2\eta_k |\Delta | \leq \frac{\log a}{a}. \label{lok}$$ By the proof of [@AF THEOREM 4.2], $$\begin{aligned} &&\exp \left[-a \left\{(\xi_k^2+2\xi_k \eta_k)\sum_{l=1}^N\Delta_l^2 -2\sum_{l=1}^N{{\rm Re}\,}{\left\langle}\Psi_k(t_l)-\Psi_k(t_{l-1}), \Psi_k(t_{l-1}){\right\rangle}\right\}\right] \notag \\ &&\leq \gamma_{\Delta , k} \leq 1.\label{loga}\end{aligned}$$ Then, by (\[hyoukaep\]) and (\[loga\]), we have $$\begin{aligned} &&| \lambda_{\Delta , k}-\lambda_k | =|\lambda_k(\gamma_{\Delta , k}-1)+\epsilon_{\Delta, k}| \leq \lambda_k(1-\gamma_{\Delta , k})+\epsilon_{\Delta, k}\notag \\ &\leq &\lambda_k\left( 1- \exp \left[-a \left\{(\xi_k^2+2\xi_k \eta_k)\sum_{l=1}^N\Delta_l^2 -2\sum_{l=1}^N{{\rm Re}\,}{\left\langle}\Psi_k(t_l)-\Psi_k(t_{l-1}), \Psi_k(t_{l-1}){\right\rangle}\right\}\right]\right)\notag \\ &&+2(\xi_k^2+\eta_k^2)\sum_{l=1}^N\Delta_l^2.\end{aligned}$$ The following corollary can be easily proven by using [@AF COROLLARY 4.4]. \[E1\] Assume that there exists $k\in \mathbb{N}$ such that the following conditions hold: $$\begin{aligned} &&\Psi_k(\cdot):[0,\tau]\to \mathcal{H} \quad \text{is a strongly differentiable mapping}, \label{E11}\\ &&\forall \lambda \in [0,\tau], \quad \Psi_k(\lambda)\in D(H), \label{E12}\\ &&\xi_k < \infty , \label{E13}\\ &&\sup_{0\leq \lambda \leq \tau}\|\Psi_k'(\lambda) \|< \infty, \label{E14}\\ &&\text{where} \ \Psi_k'(\cdot) \ \text{denotes the strong derivative of} \ \Psi_k(\cdot).\notag\end{aligned}$$ Then (\[A0\])–(\[A3\]) hold. Therefore, by Theorem \[saisho\], (\[A4\]) holds. \[E2\] Let $A$ be a self-adjoint operator on $\mathcal{H}$. Assume that there exists $k\in \mathbb{N}$ such that the following conditions hold: $$\begin{aligned} &&\Psi_k \in D(A) \cap \bigcap_{0\leq\lambda\leq \tau}D(He^{-i\lambda A}), \label{E21}\\ &&\sup_{0\leq\lambda\leq \tau}\|He^{-i\lambda A}\Psi_k\| < \infty , \label{E22}\\ &&\forall \lambda \in [o,\tau], \ \Psi_k(\lambda) = e^{-i\lambda A} \Psi_k . \label{E23}\end{aligned}$$ In this case, by [@AF EXAMPLE 4.5], (\[E11\])–(\[E14\]) hold. Then by using Corollary \[E1\], (\[A0\])–(\[A4\]) hold. Trace norm convergence ---------------------- For the decomposition (\[schatten\]), we define $$\rho(t) := \sum_n \lambda_n|\Psi_n(t)\rangle \langle \Psi_n(t)|, \quad \forall t \in [0,\tau].$$ Let us consider conditions of convergence from $\rho_{\Delta}(\tau)$ to $\rho(\tau)$ in the trace norm sense. \[nibannme\] Assume that the conditions (\[A0\])–(\[A3\]) hold for all $k \in \mathbb{N}$ satisfying $\lambda_k >0$. Then we have $$\begin{aligned} \lim_{|\Delta |\to 0}\| \rho_{\Delta}(\tau) - \rho(\tau)\|_1=0.\label{trace1}\end{aligned}$$ [*Proof*]{}. By definition of $\rho_{\Delta}(\tau)$, $\rho(\tau)$, and equation (\[decomep\]), we have $$\begin{aligned} \| \rho_{\Delta}(\tau) - \rho(\tau)\|_1 &=&\sum_k\langle \Psi_k(\tau), |\rho_{\Delta}(\tau) - \rho(\tau)| \Psi_k(\tau) \rangle \notag \\ &=&\sum_k|\lambda_{\Delta, k}-\lambda_k| \label{koyuuchi} \notag \\ &=&\sum_k|\lambda_k(\gamma_{\Delta,k}-1)+\epsilon_{\Delta,k}|\notag \\ &\leq&\sum_k\lambda_k(1-\gamma_{\Delta,k}) +\sum_k\epsilon_{\Delta,k}\notag \\ &=&\sum_k\lambda_k(1-\gamma_{\Delta,k}) +\sum_k(\lambda_{\Delta, k}-\lambda_k\gamma_{\Delta,k})\notag \\ &=& 2-2\sum_k\lambda_k\gamma_{\Delta,k}. \label{rholeq}\end{aligned}$$ Note that $$|\lambda_k\gamma_{\Delta,k}|\leq \lambda_k \ (\forall k \in \mathbb{N}), \quad \sum_k\lambda_k=1.$$ The assumptions (\[A0\])–(\[A3\]) imply that $$\lim_{|\Delta|\to0}\lambda_k\gamma_{\Delta,k}=\lambda_k \ (\forall k \in \mathbb{N}).$$ Hence, by using Lebesgue’s dominated convergence theorem, we have $$\lim_{|\Delta|\to0}\sum_k\lambda_k\gamma_{\Delta,k}=1.$$ Therefore, by (\[rholeq\]), we obtain (\[trace1\]). Assume that the conditions of Theorem \[nibannme\] hold and that $\sup_{k, \lambda_k\neq 0}\xi_k < \infty$ and $\sup_{k, \lambda_k\neq 0}\eta_k < \infty$ hold. Then, for $a>1$, we can take $|\Delta |$ such that (\[lok\]) holds for all $k$ with $\lambda_k \neq 0$. Then we have (\[loga\]) for all $k\in \mathbb{N}$ with $\lambda_k \neq 0$. Hence, by (\[rholeq\]), for all $k\in \mathbb{N}$, we obtain the following estimation: $$\begin{aligned} &&|\lambda_{\Delta , k}-\lambda_k | \leq \| \rho_{\Delta}(\tau) - \rho(\tau)\|_1 \notag \\ &\leq& 2-2\sum_k\lambda_k \exp \left[-a \left\{(\xi_k^2+2\xi_k \eta_k)\sum_{l=1}^N\Delta_l^2 -2\sum_{l=1}^N{{\rm Re}\,}{\left\langle}\Psi_k(t_l)-\Psi_k(t_{l-1}), \Psi_k(t_{l-1}){\right\rangle}\right\}\right].\notag \\\end{aligned}$$ The following corollary and example can be easily proven by using Corollary \[E1\], Example \[E2\], and Theorem \[nibannme\]. \[c3\] Assume that the conditions (\[E11\])–(\[E14\]) hold for all $k \in \mathbb{N}$ with $\lambda_k >0$. Then we have (\[trace1\]). \[c4\] Let $A$ be a self-adjoint operator on $\mathcal{H}$. Assume that the conditions (\[E21\])–(\[E23\]) hold for all $k \in \mathbb{N}$ with $\lambda_k >0$. Then we have (\[trace1\]). In Example \[c4\], let us consider the case of $d<\infty$. It is easy to see that the assumptions (\[E21\])–(\[E22\]) are satisfied. On the other hand, by Stone’s theorem, for all $U \in \mathfrak{U}(\mathcal{H})$, there exists a self-adjoint operator $A$ such that $U = e^{-i\tau A}$. Since $\rho(\tau)= U\rho U^*$, we have $ \lim_{|\Delta |\to 0}\| \rho_{\Delta}(\tau) - U\rho U^*\|_1=0. $ This fact shows that, in the case $d<\infty$, an arbitrary state in $\{U\rho U^*\ | \ U \in \mathfrak{U}(\mathcal{H}) \}$ can be approximated (in the trace norm sense) by states obtained after an appropriate continuous measurements. In other words, in this case, we can approximate an arbitrary unitary channel by continuous quantum measurements. Application to quantum Zeno effect for mixed states --------------------------------------------------- Let $\Psi_k \in D(H)$ and $\Psi_k(\lambda)=\Psi_k \ (\forall \lambda \in [0,\tau])$ holds for all $k\in \mathbb{N}$ with $\lambda_k >0$. This is the case where $A=0$ in Example \[c4\]. Then (\[A0\])–(\[A3\]) hold for all $k\in \mathbb{N}$ with $\lambda_k >0$. Hence, we have (\[trace1\]). This means that, by the series of measurement with respect to the family of the projection operators $\{|\Psi_k \rangle \langle \Psi_k | \}_k$, transitions to states different from the initial state are hindered. This can be interpreted as a quantum Zeno effect for mixed states. Convergence condition of the von Neumann entropy ================================================ Let $\varphi:[0,\infty )\ni \lambda \mapsto -\lambda \log \lambda \in [0,\infty)$, where $\varphi (0):=0$. Then $\varphi$ is continuous, concave, and subadditive. Let $S(\rho)$ be the von Neumann entropy of $\rho\in\mathfrak{S}(\mathcal{H})$. i.e. $$S(\rho):= \mathrm{Tr}\varphi(\rho).$$ In the case $d < \infty$, by Fannes’ inequality, we have for all $\rho_1, \rho_2 \in \mathfrak{S}(\mathcal{H})$ $$\|\rho_1 -\rho_2 \|_1 \leq 1/e \Longrightarrow |S(\rho_1)-S(\rho_2)|\leq \|\rho_1 -\rho_2 \|_1\log d +\varphi(\|\rho_1 -\rho_2 \|_1).$$ Therefore the von Neumann entropy is continuous with respect to the trace norm. On the other hand,in the case $d= \infty$, although the von Neumann entropy is lower semi-continuous with respect to the trace norm (i.e. $\lim_{n\to \infty}\|\rho_n -\rho \|_1 = 0 \Rightarrow S(\rho)\leq \liminf_{n\to \infty}S(\rho_n)$), it is not necessarily continuous. Moreover, it is known that the set $\{ \rho \in \mathfrak{S}(\mathcal{H})\ | \ S(\rho) < \infty \}$ is of the first category [@We]. In what follows, we deal with the case where $d = \infty$ only. For $\rho_{\Delta}(\tau)$ and $\rho$ considered in the section 2, conditions of the convergence $S(\rho_{\Delta}(\tau))\to S(\rho)$ are given by the following theorem. \[vNentropy\] Assume that the conditions (\[A0\])–(\[A2\]) hold for all $k \in \mathbb{N}$, and that the condition (\[A3\]) holds for all $k \in \mathbb{N}$ with $\lambda_k >0$. Suppose that the following conditions hold: $$\begin{aligned} &&\xi_k \to 0, \quad \eta_k \to 0 \ (k\to 0), \label{main1}\\ &&S(\rho)< \infty, \label{main2}\\ &&\sum_k \varphi (\xi_k^2) <\infty , \quad \sum_k \varphi (\eta_k^2) <\infty. \label{main3}\end{aligned}$$ Then $$\begin{aligned} \lim_{|\Delta |\to 0}S(\rho_{\Delta}(\tau))=S(\rho (\tau))=S(\rho).\end{aligned}$$ \[vNrem1\] The function $\varphi$ is monotone increasing on $[0,1/e]$ and $$\xi_k^2=\sup_{0\leq \lambda \leq \tau}\| H\Psi_k(\lambda )\|^2 =\sup_{0\leq \lambda \leq \tau}\int_{\mathbb{R}}x^2d\|E_H(x)\Psi_k(\lambda)\|^2.$$ Hence, $\xi_k\to0\ (k\to \infty)$ implies that there exists $N_0 \in \mathbb{N}$ such that, for all $k>N_0$, $$\varphi(\xi_k^2)\geq \sup_{0\leq \lambda \leq \tau} \varphi \left( \int_{\mathbb{R}}x^2d\|E_H(x)\Psi_k(\lambda)\|^2 \right).$$ By Jensen’s inequality, we have $$\varphi \left( \int_{\mathbb{R}}x^2d\|E_H(x)\Psi_k(\lambda)\|^2 \right) \geq \int_{\mathbb{R}}\varphi(x^2)d\|E_H(x)\Psi_k(\lambda)\|^2.$$ Hence, for all $k> N_0$, $$\varphi(\xi_k^2) \geq \sup_{0\leq \lambda \leq \tau} \int_{\mathbb{R}}\varphi(x^2)d\| E_H(x)\Psi_k(\lambda)\|^2.$$ Then, we have $$\forall k>N_0, \ \forall \lambda \in [0,\tau], \ \Psi_k(\lambda)\in D(\sqrt{\varphi(H^2)}), \ \varphi(\xi_k^2)\geq \sup_{0\leq \lambda \leq \tau} \|\sqrt{\varphi(H^2)}\Psi_k(\lambda) \|^2.$$ Moreover, using the estimate that $$\begin{aligned} \sum_k\varphi(\xi_k^2)=\sum_{k=1}^{N_0}\varphi(\xi_k^2) +\sum_{k=N_0+1}^{\infty}\varphi(\xi_k^2) \geq \sum_{k=1}^{N_0}\varphi(\xi_k^2) + \sum_{k=N_0+1}^{\infty} \sup_{0\leq \lambda \leq \tau} \|\sqrt{\varphi(H^2)}\Psi_k(\lambda) \|^2 \\ \geq \sum_{k=1}^{N_0}\varphi(\xi_k^2) + \sup_{0\leq \lambda \leq \tau}\sum_{k=N_0+1}^{\infty} \|\sqrt{\varphi(H^2)}\Psi_k(\lambda) \|^2,\end{aligned}$$ we obtain $$\xi_k\to 0 \ (k\to \infty), \ \sum_k\varphi(\xi_k^2)<\infty \Longrightarrow \exists N_0 \in \mathbb{N}, \ \sup_{0\leq \lambda \leq \tau}\sum_{k=N_0+1}^{\infty} \|\sqrt{\varphi(H^2)}\Psi_k(\lambda) \|^2 < \infty.\label{varphi}$$ Particularly, in the case $H\in \mathfrak{B}({{\cal H}})$, we have, for all $\Phi \in \mathcal{H}$, $$\int_{\mathbb{R}}\varphi(x^2)d\|E_H(x)\Phi\|^2 \leq \sup_{x\in \sigma(H)}\varphi(x^2)\int_{\mathbb{R}}d\|E_H(x)\Phi \|^2 =\sup_{x\in \sigma(H)}\varphi(x^2)\cdot \|\Phi \| < \infty.$$ Hence, we obtain $\sqrt{\varphi(H^2)}\in \mathfrak{B}({{\cal H}})$. Therefore, by (\[varphi\]), we have $$\xi_k\to 0 \ (k\to \infty), \ \sum_k\varphi(\xi_k^2)<\infty \Longrightarrow \varphi(H^2)\in \mathfrak{T}({{\cal H}}).$$ We remark that, in this case, if Hamiltonian $H$ is represented as a density operator, then $\varphi(H^2)\in \mathfrak{T}({{\cal H}})$ means $S(H^2)<\infty $. [*Proof*]{}. The assumption of this theorem and Theorem \[nibannme\] imply that $ \lim_{|\Delta |\to 0}\| \rho_{\Delta}(\tau) - \rho(\tau)\|_1=0. $ Hence we have $\text{w}\mbox{-}\lim_{|\Delta |\to 0}\rho_{\Delta}(\tau)=\rho (\tau)$, where $\text{w}\mbox{-}\lim$ means weak limit. By (\[decomep\]), (\[hyoukaep\]) and $\gamma_{\Delta , k}\leq 1$, we have $$\lambda_{\Delta, k}\leq \lambda_{k}+2(\xi_k^2+\eta_k^2)\sum_{l=1}^{N}\Delta_l^2.$$ By $\lim_{|\Delta |\to 0}\sum_{l=1}^{N}\Delta_l^2=0$, there exists $\delta >0$ such that, for all $\Delta$, $ |\Delta |<\delta \Rightarrow \sum_{l=1}^{N}\Delta_l^2 <1/2. $ Thus $$\lambda_{\Delta, k}\leq \lambda_{k}+\xi_k^2+\eta_k^2 \quad (|\Delta |<\delta).\label{lren}$$ We set $$\begin{aligned} \sigma := \sum_k(\lambda_k + \xi_k^2 +\eta_k^2)|\Psi_k(\tau)\rangle \langle \Psi_k(\tau)|.\end{aligned}$$ By the assumption of this theorem, $\sigma \in \mathfrak{C}({{\cal H}})$. On the other hand, (\[lren\]) implies that $$\rho_{\Delta}(\tau)\leq \sigma \quad (|\Delta |<\delta).$$ Moreover, by the assumption of this theorem and subadditivity of $\varphi$, we have $$\begin{aligned} S(\sigma)&=&\sum_k \varphi(\lambda_k +\xi_k^2+\eta_k^2)\\ &\leq& S(\rho)+\sum_k\varphi(\xi_k^2)+\sum_k\varphi(\eta_k^2)<\infty.\end{aligned}$$ Hence, by Simon’s dominated convergence theorem for entropy [@LR THEOREM A.3], we have $$\lim_{|\Delta |\to 0}S(\rho_{\Delta}(\tau))=S(\rho (\tau)).$$ It is obvious that $S(\rho (\tau))=S(\rho)$ holds. In the proof of Theorem \[vNentropy\], we used that $$S(\rho)<\infty, \ \sum_k\varphi(\xi_k^2)<\infty, \ \sum_k\varphi(\eta_k^2)<\infty \Longrightarrow \sum_k \varphi(\lambda_k +\xi_k^2+\eta_k^2) <\infty.$$ Conversely, we can show that, under condition (\[main1\]), $$\sum_k \varphi(\lambda_k +\xi_k^2+\eta_k^2) <\infty \Longrightarrow S(\rho), \ \sum_k\varphi(\xi_k^2), \ \sum_k\varphi(\eta_k^2)<\infty \label{follow}$$ as follows. By $\lambda_k+\xi_k^2+\eta_k^2 \to 0 \ (k\to \infty)$, we have $$\exists N_0 \in \mathbb{N}, \forall k > N_0, \max \{\lambda_k, \xi_k^2, \eta_k^2 \}\leq \lambda_k+\xi_k^2+\eta_k^2 < 1/e.$$ Hence, by the fact that $\varphi$ is a monotone increasing function on $[0,1/e]$, we obtain $$\max \left\{ \sum_{k=N_0+1}^{\infty}\varphi(\lambda_k), \sum_{k=N_0+1}^{\infty}\varphi(\xi_k^2), \sum_{k=N_0+1}^{\infty}\varphi(\eta_k^2) \right\} \leq \sum_{k=N_0+1}^{\infty}\varphi(\lambda_k + \xi_k^2+ \eta_k^2).$$ Therefore, we have (\[follow\]). Thus, in Theorem \[vNentropy\], we can replace the condition (\[main2\]) and (\[main3\]) with $\sum_k \varphi(\lambda_k +\xi_k^2+\eta_k^2) <\infty $. \[saigo\] Let $A$ be a self-adjoint operator on $\mathcal{H}$. Assume that $A, H \in \mathfrak{C}({{\cal H}})$, and that $A$ and $H$ are strongly commuting. Moreover, we assume that $$\begin{aligned} &&\forall k \in \mathbb{N}, \ \forall \lambda \in [0, \tau], \ \Psi_k(\lambda)=e^{-i\lambda A}\Psi_k , \label{vNe1}\\ &&S(\rho)< \infty , \ \sum_{k}\varphi(\| H\Psi_k \|^2)< \infty , \ \sum_{k}\varphi(\| A\Psi_k \|^2)< \infty . \label{vNe2}\end{aligned}$$ Then, the compactness, the strong commutativity of $A$ and $H$, and (\[vNe1\]) imply that $\xi_k=\| H\Psi_k \| \to 0 , \ \eta_k=\| A\Psi_k \| \to 0 \ (k\to \infty)$. Hence, the assumption of Theorem \[vNentropy\] is satisfied. Hence, we have $S(\rho_{\Delta}(\tau)) \to S(\rho)\ (|\Delta |\to 0)$. In Example \[saigo\], let us consider the case of $A=0$. The following fact can be easily seen: $$\begin{aligned} &&H \in \mathfrak{C}({{\cal H}}), \ \Psi_k(\lambda)=\Psi_k \ (\forall k \in \mathbb{N}, \ \forall \lambda \in [0, \tau]), \ S(\rho)< \infty , \ \sum_{k}\varphi(\| H\Psi_k \|^2)< \infty \notag \\ &&\Longrightarrow \lim_{|\Delta |\to 0}S(\rho_{\Delta}(\tau)) = S(\rho).\label{compact1}\end{aligned}$$ This is the case of QZE. We remark that, if $\{\Psi_k \}_k$ is a sequence of eigenvectors of $H$, we have $\sum_{k}\varphi(\| H\Psi_k \|^2)=\mathrm{Tr}\varphi(H^2)<\infty$. Then, in (\[compact1\]), we can replace the condition $\sum_{k}\varphi(\| H\Psi_k \|^2)<\infty$ with $\varphi (H^2)\in \mathfrak{T}(\mathcal{H})$. Acknowledgments {#acknowledgments .unnumbered} =============== The author would like to thank Professor Asao Arai for valuable comments. [99]{} A. Arai, Mathematical Principles of Quantum Statistical Mechanics, Kyoritsu Shuppan, 2008. (in Japanese). A. Arai and T. Fuda, Some mathematical aspects of quantum Zeno effect, [*Lett. Math. Phys.*]{} [**100**]{} (2012), 245–260. M. Fannes, A continuity property of the entropy density for spin lattice systems, [*Comm. Math. Phys.*]{} [**31**]{} (1973), 291–294. E. H. Lieb and M. B. Ruskai, Proof of the strong subadditivity of quantum-mechanical entropy (with an appendix by B. Simon), [*J. Math. Phys.*]{} [**14**]{} (1973), 1938–1941. B. Misra and E. C. G. Sudarshan, The Zeno’s paradox in quantum theory, [*J. Math. Phys.*]{} [**18**]{} (1977), 756–763. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, Cambridge Univ. Press, Cambridge, 2000. M. Reed and B. Simon, Methods of Modern Mathematical Physics Vol. I, Academic Press, New York, 1972. H. Umegaki and M. Ohya, Quantum Mechanical Entropy, Kyoritsu Shuppan, 1984. (in Japanease). J. von Neumann, Die Mathematische Grundlagen der Quantenmechanik, Springer, Berlin, 1932. Reprint:1981. A. Wehrl, Three theorems about entropy and convergence of density matrices, [*Rep. Math. Phys.*]{} [**10**]{} (1976), 159–163. [^1]: E-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'We apply geometric incidence estimates in positive characteristic to prove the optimal $L^2 \to L^3$ Fourier extension estimate for the paraboloid in the four-dimensional vector space over a prime residue field. In three dimensions, when $-1$ is not a square, we prove an $L^2 \to L^{\frac{32}{9} }$ extension estimate, improving the previously known exponent $\frac{68}{19}.$' address: - 'Misha Rudnev, Department of Mathematics, University of Bristol, Bristol BS8 1TW, United Kingdom' - 'Ilya D. Shkredov, Steklov Mathematical Institute, Division of Number Theory, ul. Gubkina, 8, Moscow, Russia, 119991 and IITP RAS, Bolshoy Karetny per. 19, Moscow, Russia, 127994 and MIPT, Institutskii per. 9, Dolgoprudnii, Russia, 141701' author: - Misha Rudnev - 'Ilya D. Shkredov' title: On the restriction problem for discrete paraboloid in lower dimension --- Introduction ============ Let $F=\F_p$, the prime residue field of characteristic $p>2$, where $p$ is viewed as asymptotic parameter, consider the vector space $V=F^d$, $d\ge 3$, with the standard scalar product. Define the paraboloid $\P$ as $$\mathcal{P} = \{ (\u x=(x_1,\ldots, x_{d-1}),\, \u x\cdot \u x) ~:~ \u x\in F^{d-1} \} = \{ (x_1,\ldots x_{d-1}, x^2_1 + \ldots + x^2_{d-1}) \} \subset V \,.$$ The Fourier extension problem is bounding some Lebesgue norm on $V$ of the inverse Fourier transform of a complex-valued function $f$ on $\mathcal{P} \subset V^*$, the paraboloid in the Fourier space (or more generally some degree $2$ or higher irreducible codimension-one variety) in terms of some Lebesgue norm of the function $f$ on $\mathcal{P}$. An equivalent question, with dual Lebesgue exponents, is bounding the norm of the restriction to $\mathcal P$ of the Fourier transform of a function $g$ on $V$ in terms of the norm of $g$. We consider the specific case where one of the Lebesgue norms is $L^2$, which can be reduced nicely to geometric incidence combinatorics, and dimensions three and four. In $d=4$ we prove the optimal $L^2 \to L^3$ extension estimate. Optimality, in view of normalisation described below, is readily verified by calculating the inverse Fourier transform of the characteristic function of the lift of an isotropic line in the base space $F^3$ on $\mathcal{P}$. Indeed, the paraboloid $\mathcal P$ contains, in particular, the [*null sphere*]{} – alias [*isotropic cone*]{} – that is the set $S_0\times\{0\}\subset F^4$, where $$S_0 = \{\u x \in F^3:\, \u x\cdot \u x =0\}.$$ Clearly, $S_0$ is a cone, for if $\u x\in S_0$, then so is its multiple by a scalar. Such a three-vector $\u x$ is called [*isotropic*]{}, as well as any line, whose direction vector is isotropic. By non-degeneracy of the dot product, if $\u x,\,\u y$ are nonzero isotropic vectors in $F^3$, with $\u x\cdot \u y=0$, then one is a scalar multiple of the other. In $d=3$ two isotropic directions in the base space $F^2$ exist if $-1$ is a square in $F$, that is $p\equiv 1 \pmod 4$ and do not exist if $p\equiv 3 \pmod 4$. In the former case, by taking the inverse Fourier transform of an isotropic line, one concludes that the best possible extension Fourier exponent from $L^2 $ is $4,$ which is also the Stein-Tomas exponent (which means that it follows from the basic Fourier estimates and alone). However, if $-1$ is not a square, the best possible extension exponent from $L^2 $ is an open question. Mockenhaupt and and Tao [@MoT] proved a threshold $L^2 \to L^{\frac{18}{5}}$ estimate (but for the endpoint) and conjectured that the best exponent should be $3$. Their exponent $\frac{18}{5}$ was improved to $\frac{18}{5}-\delta$, for $0<\delta<\frac{1}{1035}$ by Lewko [@Lew]. Stevens and de Zeeuw [@SdZ] remark that their new incidence theorem – presented here as Lemma \[lem:2d\] – would justify $0<\delta<\frac{2}{95}$ in Lewko’s argument, thus claiming the exponent $\frac{68}{19}$. Here we push it a little further, namely $0<\delta\leq \frac{2}{45}$, proving an $L^2 \to L^{\frac{32}{9}}$ extension estimate, owing to a more meticulous application of the Stevens-de Zeeuw theorem. The restriction problem over the real field has a reputed history, which we do not aim to present; since the 2000s, after having been set up by Mockenhaupt and Tao [@MoT], the question has also been studied in the finite field setting. Introduction-level discussion of the discrete paraboloid case can be found in Green’s lecture notes [@Green Sections 6–9]. Heuristically, restriction phenomena are closely linked to geometric incidence laws, governing intersections of lower-dimensional affine subspaces in a vector space. Recent progress is due to Lewko [@Lew], [@Lewko], as well as Iosevich, Koh and Lewko [@IKL]. These works – see especially [@Lew], [@Lewko] – also introduce the subject at length and breadth and contain exhaustive reference lists. Remarkably, in [@IKL] optimal extension/restriction estimates for the paraboloid in even dimensions $\geq 6$ (but for the endpoint exponent in dimension $6$) were established by using a somewhat crude geometric incidence machinery based entirely on discrete Fourier analysis, involving Gauss and Kloosterman sum estimates. By somewhat crude we mean, heuristically, that this machinery works well, provided that the sets of geometric objects involved are sufficiently large in terms of the cardinality of the finite field. This approach in [@IKL] turns out to be powerful enough to have led to optimal estimates in higher dimension, however in dimension $4$ it only allows for a partial result. In this note, if the dimension $d=4$, we use sharper incidences results from the first author’s paper [@Rud], which turn out to be ideally suited to study the restriction problem on the paraboloid in this dimension. This allows us to settle the question. We must admit in comparison that the methods of [@IKL]-[@Lewko] work in any finite field $\F_q$ of odd characteristic, while we are forced to confine ourselves to $F=\F_p$. The obstacle is that the characteristic $p$ does appear in the statement of the point-plane incidence theorem, Lemma \[lem:Misha+\], that we fetch from the first author’s work [@Rud] – see the latter paper for discussion why replacing (under some constraint) $p$ by $q$ is likely to be a difficult structural problem. The argument in dimension $3$ follows roughly the same lines if instead of Rudnev’s point-plane incidences theorem one uses the point-line Szemerédi-Trotter type incidence theorems in $F^2$ due to Stevens and de Zeeuw [@SdZ]. The fact that the theorem is useful for the restriction problem was observed in the latter paper, claiming the extension exponent $\frac{68}{19}$. Here we develop a more elaborate way of applying the latter point-line incidence bound along the lines of the well-known, and sharp, application of the Szemerédi-Trotter theorem to the repeated angle in the plane bound, developed by Pach and Sharir [@PS]. We remark that the Stevens-de Zeeuw theorem is in some sense a corollary of Rudnev’s theorem, which in turn can be viewed as an adaptation of the breakthrough development of the polynomial method by Guth and Katz [@GK] (at this point we stop the genealogical detour) which has also inspired progress in the restriction problem over the reals, due to Guth [@Guth1], [@Guth2]. As far as the notation is concerned, we follow [@IKL], using the counting norm on $V$, thus defining the Fourier transform of a function $g:V\to \mathbb C$ as $$\hat g(\xi) := \sum_{x\in V} g(x) e(-x\cdot \xi)\,,$$ where $e$ is a non-trivial additive character. In the inverse Fourier transform summation is replaced by taking the expectation. As in the real prototype of the question, for a function $f$ on $\mathcal P,$ its inverse Fourier transform is denoted/defined as $$(fd\sigma)^\lor(x) := \frac{1}{p^{d-1}} \sum_{\xi\in \mathcal P} f(\xi) e(x\cdot\xi)\,,$$ the notation $d\sigma$ standing for the normalised “surface area” on $\mathcal P$, assigning to each point on $\mathcal P$ the mass $|\mathcal P|^{-1}=1/p^{d-1}$ ($|\cdot|$ standing in particular, as usual, for finite cardinality). Thus a Lebesgue $L^q$-norm of $f$ on $\mathcal P$ is defined as $$\|f\|_{L^q(\mathcal P,d\sigma)}:= \left( \frac{1}{p^{d-1}} \sum_{\xi\in \mathcal P} |f(\xi) |^q \right)^{\frac{1}{q}}\,,$$ while for a function $g$ on $V$ it is $$\|g\|_{L^q(V)}:= \left( \sum_{x\in V} |g(x) |^q \right)^{\frac{1}{q}}\,.$$ Finally, we use the symbols $\ll$ (as well as $O(\,)$) and $\gg$ in the standard way to suppress absolute constants, as well as $\sim$ for an approximate equality of two quantities within a constant factor. Our main results are as follows. Let $f$ be a function on $\mathcal P\subset \mathbb F_p^4$. Then $$\|(fd\sigma)^\lor \|_{L^3(\F_p^4)} \ll \|f\|_{L^2(\mathcal P,d\sigma)}\,.$$ \[t:rest\] Let $f$ be a function on $\mathcal P\subset \mathbb F_p^3$ and $p\equiv 3 \pmod 4$. Then $$\|(fd\sigma)^\lor \|_{L^{\frac{32}{9}}(\F_p^3)} \ll \|f\|_{L^2(\mathcal P,d\sigma)}\,.$$ \[t:rest3\] Preliminaries ============= For a set $G\subseteq V$, with the fourth coordinate $h\in F$, define $G_h \subseteq \mathcal{P}$ as the horizontal $h$-slice of $G$, lifted to $\mathcal P$, that is $$G_h := \{ (\u x,\u x\cdot \u x): \,(\u x,h)\in G\}\,.$$ For a finite set $X$ in an abelian group, we define the usual additive energy as $$\label{energy} \E(X) = |\{(x,y,z,u)\in X^4:\,x+y=z+u\}|\,.$$ We begin with the key preliminary consideration, originating in [@MoT], which reduces restriction bounds to estimating additive energy of sets on $\mathcal P$. We essentially quote [@IKL Lemma 2.1], where a thorough sketch of the proof is given. See [@Lewko Corollary 25, Lemma 29] for more details and references as to the claim. \[l:g\^-energy\] Let $g: V \to \mathbb{C}$ be a function such that $\| g\|_\infty \leq 1$ on its support $G$. Then $$\label{e:g^-energy} \| \hat{g} \|_{L^2 (\P,d\sigma)} \ll |G|^{\frac{1}{2}} + |G|^{\frac{3}{8}} p^{-\frac{d-2}{8}} \left( \sum_{h\in F} \E^{\frac{1}{4}} (G_h) \right)^{\frac{1}{2}} \,.$$ Note that in the left-hand side of the paraboloid $\mathcal P$ lives in the Fourier space $V^*$, while in the right-hand side the sets $G_h$ arise on $\mathcal P\subset V$. This can be seen by inspection of the proof of Lemma \[l:g\^-energy\], which uses Gauss sums and the Hölder inequality. In addition, we will use two more basic Stein-Tomas type estimates, also obtained in this vein for $\| \hat{g} \|_{L^2 (\P,d\sigma)}$, see [@IKL formulae (1.6), (1.7)]: $$\label{f:g_hat_2} \| \hat{g} \|_{L^2 (\P,d\sigma)} \ll |G|^{\frac{1}{2}} + |G| p^{-\frac{d-1}{4}}$$ and $$\label{f:g_hat_3} \| \hat{g} \|_{L^2 (\P,d\sigma)} \ll p^{\frac{1}{2}} |G|^{\frac{1}{2}} \,.$$ Proof of Theorem \[t:rest\] =========================== In this this section we set $d=4$, so $V=F^4$. For a vector $x = (x_1,\dots, x_4)\in V$ we write $x=(\u x,h)$, so $\underline{x}=(x_1,x_2,x_3)$ and $h=x_4$. We complement the bounds in the previous section by a key incidence bound in the next lemma, as follows. \[l:par\_energy\] Let $A\subseteq \mathcal{P}\subset V$. Then $$\label{f:E_par_lemma} \E(A) \ll \frac{|A|^3}{p} + |A|^{\frac{5}{2}} + p^2|A| \,.$$ The proof of Lemma \[l:par\_energy\] is presented in a separate section. We now put together the estimates we have so far in this section and show how they imply Theorem \[t:rest\]. Let $g: V \to \mathbb{C}$ be a function such that $\| g\|_\infty \leq 1$ on its support $G$. Then $$\| \hat{g} \|_{L^2 (\P,d\sigma)} \ll \left\{ \begin{array}{ll} |G|^{\frac{1}{2}} + |G| p^{-\frac{3}{4}} & \textrm{for } \;1\le |G| \le p^{\frac{9}{4}}\,, \\ |G|^{\frac{1}{2}} p^{\frac{3}{8}} & \textrm{for } \;p^{\frac{9}{4}}\le |G| \le p^{\frac{7}{3}}\,, \\ |G|^{\frac{11}{16}} p^{-\frac{1}{16}} & \textrm{for } \;p^{\frac{7}{3}} \le |G| \le p^3\,, \\ |G|^{\frac{1}{2}} p^{\frac{1}{2}} & \textrm{for } \;p^3 \le |G| \le p^4\,. \end{array} \right.$$ \[cor:g\_hat\_bounds\] By Lemma \[l:g\^-energy\] and Lemma \[l:par\_energy\], we have $$\| \hat{g} \|_{L^2 (\P,d\sigma)} \ll |G|^{\frac{1}{2}} \; + \;|G|^{\frac{3}{8}} p^{-\frac{1}{4}} \left( \sum_{h\in F} p^{-\frac{1}{4}}|G_h|^{\frac{3}{4}} + |G_h|^{\frac{5}{8}} + p^{\frac{1}{2}}|G_h|^{\frac{1}{4}} \right)^{\frac{1}{2}} \,.$$ Hence, by the Hölder inequality, $$\| \hat{g} \|_{L^2 (\P,d\sigma)} \ll |G|^{\frac{3}{4}} p^{-\frac{1}{4}} + |G|^{\frac{11}{16}} p^{-\frac{1}{16}} + |G|^{\frac{1}{2}} p^{\frac{3}{8}}\,.$$ Combining this with inequalities , completes the proof of Corollary \[cor:g\_hat\_bounds\]. We are now ready to prove Theorem \[t:rest\]. By duality, for all $g : V \to \mathbb{C}$, it suffices to establish the restriction estimate $\| \hat{g} \|_{L^2 (\P,d\sigma)} \ll 1$, given that $$\label{f:p_restriction} \sum_{x\in V} |g(x)|^{\frac{3}{2}} = 1 \,.$$ We start out with dyadic decomposition of the set of level sets of $g$. Namely, for $i=0,1,\ldots,L\ll \log p,$ define $G_i$ to be the set of $x\in V$, where $g(x)\sim 2^{-i}$ up to a factor of $2$, and $g_i$ is the restriction of $g$ to $G_i$. Clearly, from one has $|\supp g_i| \ll 2^{\frac{3}{2}i}$. By the triangle inequality, followed by invoking Corollary \[cor:g\_hat\_bounds\], we obtain $$\begin{aligned} \| \hat{g} \|_{L^2 (\P,d\sigma)} & \ll \;\;\;\;\;\;\;\sum_{i=0}^L 2^{-i} \| \hat{g}_i \|_{L^2 (\P,d\sigma)} \\ & \ll \;\;\sum_{i :~ 2^{\frac{3}{2}i} \le p^{\frac{9}{4}}} 2^{-i} \left( 2^{\frac{3}{4}i} + p^{-\frac{3}{4}} 2^{\frac{3}{2}i} \right) \;\; + \;\;\sum_{i :~ p^{\frac{9}{4}}\le 2^{\frac{3}{2}i} \le p^{\frac{7}{3}}} 2^{-i}2^{\frac{3}{4}i}p^{\frac{3}{8}} \\ & + \sum_{i :~ p^{\frac{7}{3}} \le 2^{\frac{3}{2}i} \le p^3} 2^{-i} 2^{\frac{33}{32} i} p^{-\frac{1}{16}} \;\;\;\;\;\;\;\;\;\;\;+\;\;\;\;\; \sum_{i :~ p^3 \le 2^{\frac{3}{2}i} \le p^4} 2^{-i} 2^{\frac{3}{4}i} p^{\frac{1}{2}} \\ & \ll \;\;\;\;1 + 1+ 1 + 1 + 1 \\ & \ll \;\;\;1, \end{aligned}\label{long}$$ as required. Proof of Lemma \[l:par\_energy\] -------------------------------- We use the point-plane incidence theorem of the first author  [@Rud Theorem 3, 3\*] (see also [@MPR-NRS Theorem 8] and the proof of [@MuPet2 Corollary 2]), which combined with the incidence bound from [@MuPet1 Section 3] leads to the following asymptotic formula. \[lem:Misha+\] Let $Q \subseteq \F^3$ be a set of points and $\Pi$ a set of planes in $F^3$. Suppose that $|Q| \le |\Pi|$ and that $k$ is the maximum number of collinear points in $Q$. Then $$\sum_{q \in Q} \sum_{\pi \in \Pi} \pi (q) - \frac{|Q| |\Pi|}{p} \ll |Q|^{\frac{1}{2}} |\Pi| + k |Q| \,.$$ Moreover, if $L$ is a set of lines in $F^3$ and one excludes incidences $(q,\pi)\in Q\times \Pi$, such that $q\in l\subset \pi$ for some $l$ in $L$, then $k$ gets replaced by the maximum number of points of $Q$ on a line not in $L$. Clearly, if $x,y,z\in \P$, then in order that $x-z+y$ also be in $\P$, one must have $(\u x- \u z+\u y)\cdot (\u x- \u z+\u y) = \u x \cdot \u x - \u z \cdot \u z + \u y \cdot \u y$, which means $$\label{f:inc_d-1} (\u x - \u z) \cdot (\u z - \u y) = 0\,.$$ Permuting the variables in the definition of energy we arrive in a simple criterion: a quadruple $(x,y,z,u)\in \mathcal P^4$ satisfies if and only if $(\u x,\u y, \u z,\u u)\in (F^3)^4$ is a [*rectangle*]{}, as opposed to generally being a parallelogram to form an additive quadruple just in $(F^3)^4$. Namely at each vertex $\u x,\u y, \u z,\u u$, the dot products of adjacent difference vectors is zero. Note that condition should hold at every vertex of the rectangle, hence finding all solutions of the latter relation applies to each geometric rectangle at least four times. In the sequel we will use this freedom to chose the convenient “corner” $(x,y,z)$ of the rectangle and the fact that the energy equals four times the number of “geometric’ rectangles. The rectangles one encounters in $d=4$ are of three types. The first type is [*ordinary*]{} ones, that is lines along all the four sides are non-isotropic. These rectangles “look like” the Euclidean ones. The opposite case is [*degenerate*]{} rectangles, namely when both sides, adjacent to a vertex, are isotropic vectors. This implies that $\u x,\u y, \u z,\u u$ all lie on the same isotropic line, for isotropic lines through a point in $F^3$ form a copy of the cone $S_0$, rather than a plane. The intermediate case comprises [*semi-degenerate*]{} rectangles, namely when the lines along one pair of opposite sides are isotropic, but for the other two sides – non-isotropic. Given the line $l$ containing an isotropic side of such a rectangle (by which we mean that $l$ contains two adjacent vertices), the rectangle must lie in the unique [*semi-isotropic plane*]{} $l^\perp$, containing $l$. Heuristically, every parallelogram in $l^\perp$, one of whose sides is parallel to $l$ is a rectangle. We further use $\u A$ to denote the projection of $A\subseteq \mathcal P$ on the first three variables: $A$ is a graph over $\u A$, and $|\u A|=|A|.$ We seek to count solutions of equations , when the variables are in $\u A$. Obviously, a solution of equation (\[f:inc\_d-1\]) can be interpreted as a point-plane incidence in $F^3$: $\u x$ being the point and the plane $\pi$ being the one passing through $\u z$ and with the normal vector $\u y- \u z$. In the sequel we assume that $\u y\neq \u z$, for otherwise we have $|A|^2$ trivial solutions to the energy equation. The set of planes is, in fact, a multiset. Observe that linearity of the estimate of Lemma \[lem:Misha+\] in $|\Pi|$ enables the set of planes $\Pi$ to be a multiset as well: the estimate of the lemma remains valid, with $m$ counting planes with multiplicities, as long as the number of distinct planes, defined by $\Pi$ is $\gg |Q|$. To this end, as to , we set $m=|A|^2$, $n=|A|$ and need to show that the number of distinct planes, defined by the latter relation is $\gg n$. We do this assuming $|A|\gg Cp$, with a sufficiently large absolute $C$, the number of distinct planes, defined by pairs $(\u y,\u z)$ is $\gg |A|$, for otherwise, if $|A|\ll p$, the energy $\E(A)$ can be trivially bounded by the last term in , whereupon nothing remains left to prove. Observe that given $\u y$, each value of $\u z$ yields a different plane, unless the vector $\u z - \u y$ is isotropic. Hence, by the pigeonhole principle, we have $\gg|A|$ distinct planes, provided that the number of pairs $(\u y, \u z)$, such that $\u z - \u y$ is non-isotropic is $\gg|A|^2$. Let us show that this is certainly the case if $|A|\gg Cp$ for a sufficiently large $C$. Indeed, we can build a graph $\Gamma$ on the vertex set $\u A$, connecting vertices $(\u y, \u z)$ by an edge if $\u z - \u y$ is isotropic, and we shall show that $\Gamma$ needs $\gg |A|^2$ additional edges to be turned into a complete graph. One cannot have a non-trivial triangle (that is when its three distinct vertices are non-collinear on an isotropic line) with vertices $\u a,\,\u b,\,\u c\,\in\, F^3$, so that each side of the triangle is isotropic. Indeed, $$\u a-\u c = (\u a-\u b) - (\u c-\u b),$$ hence if both sides $\u a-\u b$ and $\u c-\u b$ are isotropic, then $\u a-\u c$ is not, for otherwise we would have $(\u a-\u b) \cdot (\u c-\u b)=0$, and hence the whole plane in $F^3$, defined by the points $\u a,\,\u b,\,\u c$ would be isotropic, which cannot happen. If $|A|\geq Cp$, for some sufficiently large absolute constant $C$, then the number of collinear point triples in $A$, lying on isotropic lines is at most $|A|p(p+1)\ll \frac{|A|^3}{C^2}$. (From each point there are at most $p+1$ isotropic directions, however we could use a cruder estimate $|A|^2p$). A collinear triple on an isotropic line is the only way to have a triangle in the graph $\Gamma$, so $\Gamma$ has a small triangle density $\ll C^{-2}$. On the other hand, if by the pigeonhole principle the edge density of $\Gamma$ were as large as $(1-c)$, for a sufficiently small constant $c>0$ – which would mean that $\u z -\u y$ were non-isotropic for only $\ll c|A|^2$ pairs $(\u y, \u z)$ – then the triangle density would be at least $1-O(c)$. In fact, one can use a much more fine-tuned asymptotic formula by Razborov [@Raz], which tells that that the triangle density would be at least $1-3c + O(c^2)$. We continue under the assumption that $|A|\gg p$ and applying Lemma \[lem:Misha+\] obtain an intermediate bound for energy $$\E(A) \ll \frac{|A|^3}{p} + |A|^{\frac{5}{2}} + k|A|^2 \,,$$ where $k$ is the maximum number of collinear points in $\u A$. Clearly, if $k\ll |A|^{\frac{1}{2}}$, we are done. Otherwise, we proceed as follows. Let $L$ be the set of all lines in $F^3$, supporting, say $\geq 10 |A|^{\frac{1}{2}}$ points of $\u A$. By excluding the incidences along the lines in $L$ in the application of Lemma \[lem:Misha+\], we succeed in counting all the rectangles in $\u A$, such that at least one line containing the side of the rectangle is not in $L$. The number $\E_1$ of such rectangles is $$\label{f:E_par1} \E_1 \ll \frac{|A|^3}{p} + |A|^{\frac{5}{2}} \,,$$ Let $L'\subseteq L$ be the subset of non-isotropic lines and $\u A'$ the subset of $\u A$, supported on the union of these lines, and $A'$ its lift on $\mathcal P$. We now bound $\E(A',A)$. This count will include all ordinary and semi-degenerate rectangles, the line supporting one of whose sides contains $\geq 10 |A|^{\frac{1}{2}}$ points of $\u A$. Let $\E_2\leq \E(A',A)$ be the number of such rectangles. By the exclusion-inclusion principle, $\u A'$ is supported on some minimal set $L'$ of $|L'|\ll |A'| |A|^{-\frac{1}{2}}$ non-isotropic lines with at least $10 |A|^{\frac{1}{2}}$ points of $\u A$ on each line. Let $\u A_l := l \cap \u A'$, for a line $l\in L'$, $A_l$ denoting the lift of $\u A_l$ on $\mathcal P$. Let us show that $$\label{mix}\E(A_l,A) = |\{ ( x, y, z, u) \in A\times A_l\times A\times A_l:\; x+ y = z + u\}|\ll |A_l||A|\,.$$ Geometrically speaking, if we fix a diagonal of a rectangle, whose one side lies on a non-isotropic line, this fixes the remaining two vertices. Analytically, without loss of generality, the line $l$ can be parametrised as $x_1=t$, $x_2=\a t+ \beta$, $x_3=\gamma t +\delta$ for some $(\a,\beta,\gamma,\delta)\in F^4$. Since $A \subseteq \mathcal{P}$, then a quadruple satisfying the energy equation in is described by the system of equations $$\begin{array}{rcl} t^2 (1+\a^2 +\gamma^2) + 2 t(\a \beta + \gamma \d ) + \u x \cdot \u x &=& s^2 (1+\a^2 +\gamma^2) + 2s(\a \beta + \gamma \d ) + \u z \cdot \u z\,, \\ \hfill \\ \u x - \u z & = & (s-t) (1,\a, \gamma) \,.\end{array}$$ The second equation is vacuous if $s=t$ and $\u x = \u z$, yielding $|A| |A_l|$ trivial solutions to . Otherwise we eliminate $\u x$, getting $$t^2 (1+\a^2 +\gamma^2) + t(\a \beta + \gamma \d ) - ts (1+\a^2 +\gamma^2) = s(\a \beta + \gamma \d ) + (t-s) \u z \cdot (1,\a, \gamma) \,,$$ a quadratic equation in $t$. Since the line $l$ is non-isotropic, the leading coefficient $1+\a^2 +\gamma^2 \neq 0$. This gives at most $2|A_l| |A|$ nontrivial solutions to the equation in . It follows from by the Cauchy-Schwarz inequality and since $|L'|\ll |A|^{\frac{1}{2}}$ that $$\label{f:E_par2} \E_2 \le \left( \sum_{l\in L'} \E(A,A_l)^{\frac{1}{2}} \right)^{2} \ll |A| \left(\sum_{l\in L'} |A_l|^{\frac{1}{2}} \right)^{2} \le |A| |L'| |A'| \; \ll \; |A|^{\frac{5}{2}}.$$ It remains to count the degenerate rectangles on lines in $L$, whose number we denote as $\E_3$. It’s easy to see that $\E_3 \ll |A|p^2$. Indeed, take a dyadic subset of isotropic lines in $L$, each supporting $\sim k$ points of $\u A$. Then the maximum number of degenerate rectangles supported on these lines is trivially bounded as $\ll \frac{|A|}{k} k^3$. Since $k\leq p$, the dyadic sum is dominated by the term with the largest value of $k$, that is $k=p$. Combining this bound with bounds , , since $\E(A)\ll\E_1+\E_2+\E_3$, completes the proof of Lemma \[l:par\_energy\]. Proof of Theorem \[t:rest3\] ============================ We set $d=3$, $V=F^3$ and assume that the base plane $F^2$ has no isotropic directions, that is $p\equiv 3 \pmod 4$. Whether or not there are isotropic directions, interpolating just between formulae and yields a variant of Corollary \[cor:g\_hat\_bounds\], which results in the Stein-Tomas extension estimate $L^2\to L^4$, which the presence of isotropic lines makes the best possible one if $p\equiv 1 \pmod 4$. Since the structure of the proof of Theorem \[t:rest3\] is similar to that of the proof of Theorem \[t:rest\], we omit some tedious straightforward calculations in the sequel. We first present an incidence statement that will play the same role as Lemma \[l:par\_energy\] played in the case $d=4$. Actually, the second bound of Lemma \[l:corners\] coincides with [@IKL Lemma 4.1] but for us it is an unconditional bound of relatively little consequence. \[l:corners\] For $A\subseteq \mathcal{P}\subset V,$ one has $$\label{f:E2} \E(A) \;\ll \;\left\{ \begin{array}{lll} |A|^{\frac{17}{7}} &for & |A|\leq p^{\frac{26}{21}}\,,\\ \hfill \\ \frac{|A|^3}{p} + |A|^2\sqrt{p}\,.\end{array}\right.$$ The proof of Lemma \[l:corners\] is presented in a separate section. Substituting these bounds as $\E(G_h)$ in the estimate of Lemma \[l:g\^-energy\] and optimising with the Stein-Tomas inequalities , yields the following statement, an an analogue of Corollary \[cor:g\_hat\_bounds\]. Let $g: V \to \mathbb{C}$ be a function such that $\| g\|_\infty \leq 1$ on its support $G$. Then $$\| \hat{g} \|_{L^2 (\P,d\sigma)} \ll \left\{ \begin{array}{ll} |G|^{\frac{1}{2}} + |G| p^{-\frac{1}{2}} & \textrm{for } \;1\le |G| \le p^{\frac{16}{9}}\,, \\ |G|^{\frac{19}{28}} p^{\frac{1}{14}}+|G|^{\frac{7}{8}}p^{-\frac{125}{336}} & \textrm{for } \;p^{\frac{16}{9}} \le |G| \le p^{\frac{47}{21}}\,, \\ |G|^{\frac{5}{8}} p^{\frac{3}{16}} & \textrm{for } \;p^{\frac{47}{21}} \le |G| \le p^{\frac{5}{2}}\,,\\ |G|^{\frac{1}{2}} p^{\frac{1}{2}} & \textrm{for } \;p^{\frac{5}{2}} \le |G| \le p^3\,. \end{array} \right.$$ \[cor:g\_hat\_bounds2d\] We start with the last estimate, namely the bound , which is better than the unconditional bound in the second line of for $|G|\geq p^{\frac{5}{2}}.$ Furthermore, when $p^{\frac{47}{21}} \le |G| \le p^{\frac{5}{2}}$ we use the unconditional bound in the second line of : substituting it for each $\E(G_h)$ in formula and applying the Hölder inequality yields an estimate $$\begin{aligned} |G|^{\frac{3}{8}} p^{-\frac{1}{8}}\left( p^{-\frac{1}{8}} \left( \sum_{h\in F} |G_h|^{\frac{3}{4}} \right )^{\frac{1}{2}} + p^{\frac{1}{16}} \left( \sum_{h\in F} |G_h|^{\frac{1}{2}} \right )^{\frac{1}{2}}\right) \\ \leq \;\;\; |G|^{\frac{3}{4}}p^{-\frac{1}{8}} + |G|^{\frac{5}{8}}p^{\frac{3}{16}} \;\; \ll \;\;|G|^{\frac{5}{8}}p^{\frac{3}{16}}\,, \end{aligned}$$ given the above range of $|G|$. We now go back to the other end of the claim of Corollary \[cor:g\_hat\_bounds2d\], and for $ |G| \le p^{\frac{16}{9}}$ use the Stein-Tomas bound . At the upper end of the latter interval of $|G|$ the Stein-Tomas bound meets the following estimate, for which we use the first bound of for $|G_h|$ as long as $|G_h|\leq p^{\frac{26}{21}}$ and otherwise the second bound, but the number of such $h$ is at most $p^{\frac{16}{9}-\frac{26}{21}} < p^{\frac{4}{7}}$ – in the following calculation we index these slices of $G$ by $i$. Substitute these in , apply the Hölder inequality and pick the dominating term, which comes from the first bound of – this yields a bound $$\begin{aligned} \;\;\;|G|^{\frac{3}{8}} p^{-\frac{1}{8}}\left( \left( \sum_{h\in F} |G_h|^{\frac{17}{28}} \right )^{\frac{1}{2}} + p^{-\frac{1}{8}}\left(\sum_{i=1,\ldots,p^{\frac{4}{7}}} |G_i|^{\frac{3}{4}} \right )^{\frac{1}{2}} + p^{\frac{1}{16}} \left( \sum_{i=1,\ldots,p^{\frac{4}{7}}} |G_i|^{\frac{1}{2}} \right )^{\frac{1}{2}}\right) \\ \leq \;\;\; |G|^{\frac{19}{28}} p^{\frac{1}{14}}+|G|^{\frac{3}{4}}p^{-\frac{5}{28}} + |G|^{\frac{5}{8}}p^{\frac{9}{112}} \;\; \ll \;\;|G|^{\frac{19}{28}} p^{\frac{1}{14}}\,, \end{aligned}$$ in the above range of $|G|$. The latter estimate, naturally, meets estimate when $ |G| \sim p^{\frac{16}{9}}$. Finally, for $p^{\frac{16}{9}} \le |G| \le p^{\frac{47}{21}}$ we repeat the latter calculation, only with the range of $i$, marking slices of $G$ with $|G_h|\geq p^{\frac{26}{21}}$, ranging from $1$ to $|G|p^{-\frac{26}{21}}$. This yields the estimate $$\begin{aligned} |G|^{\frac{19}{28}} p^{\frac{1}{14}}+|G|^{\frac{3}{8}} p^{-\frac{1}{8}}\left( p^{-\frac{1}{8}}\left(\sum_{i=1,\ldots,|G|p^{-\frac{26}{21}}} |G_i|^{\frac{3}{4}} \right )^{\frac{1}{2}} + p^{\frac{1}{16}} \left( \sum_{i=1,\ldots,|G|p^{-\frac{26}{21}}} |G_i|^{\frac{1}{2}} \right )^{\frac{1}{2}}\right) \\ \leq \;\;\; |G|^{\frac{19}{28}} p^{\frac{1}{14}} +|G|^{\frac{7}{8}}p^{-\frac{1}{4}-\frac{13}{84}} + |G|^{\frac{7}{8}}p^{-\frac{1}{16}-\frac{13}{42}} \;\;\ll \;\; |G|^{\frac{19}{28}} p^{\frac{1}{14}}+|G|^{\frac{7}{8}}p^{-\frac{125}{336}}\,. \end{aligned}$$ The statement of the theorem now follows from the bounds of Corollary \[cor:g\_hat\_bounds2d\], similar to in the proof of Theorem \[t:rest\]. We only present a part of the calculation, which yields the worst extension exponent. This is contributed, naturally, by the second range in Corollary \[cor:g\_hat\_bounds2d\], which has come from the Stevens-de Zeeuw Theorem. Assume that for some $r \in (1,3/2)$ one has $$\sum_{x\in V} |g(x)|^{r} = 1 \,.$$ We need to ensure, for the corresponding group of dyadic pieces $g_i$ of $g$ (with $g(x)\sim 2^{-i}$ on a dyadic piece) that the following estimate holds: $$1\gg \sum_{i :~ p^{\frac{16}{9}} \le 2^{ri} } 2^{-i} \left( 2^{\frac{19}{28} ri} p^{\frac{1}{14}}+2^{\frac{7}{8} ri}p^{-\frac{125}{336}}\right).$$ This means $r$ should satisfy $\frac{1}{r} = \frac{19}{28}+ \frac{1}{14}\cdot\frac{9}{16}$ (the second term in the above estimate can be easily seen to play no role), whence $r=\frac{224}{161}.$ Therefore, the dual Fourier extension exponent is $\frac{32}{9}.$ This concludes the proof of Theorem \[t:rest3\]. Proof of Lemma \[l:corners\] ---------------------------- We start off with a geometric incidence bound to replace Lemma \[lem:Misha+\]. \[lem:2d\] The number of incidences $I(n)$ between $m$ lines and $n$ points in $F^2$ is bounded as $$\label{ins} I(n) \ll \left\{ \begin{array}{lll} (mn)^{\frac{11}{15}} + m+n & for\;\;\;n^{13}m^{-2}\leq p^{15}\,, \\ \hfill \\ \frac{mn}{p} + (mnp)^{\frac{1}{2}}\,.\end{array}\right.$$ Lemma \[lem:2d\] is an amalgamation of incidence theorems of Stevens and de Zeeuw [@SdZ Theorem 3] and a well-known theorem of Vinh [@Vinh Theorem 3]. The former theorem, which contributes the first line in estimate is valid for small sets. The latter theorem, proved just by linear algebra, is nontrivial if both $m,n>p.$ We restate equation in the proof of Lemma \[l:par\_energy\], aiming to bound the number of solutions of the equation $$(\u x- \u z)\cdot(\u z- \u y)=0:\;\;\u x,\u y, \u z\in \u A\,, \label{right}$$ where $\u A$ is the projection of $A$ on the $(x_1,x_2)-$plane. Let us set $|A|=|\u A|=n$. Equation is a well-known problem of counting the maximum number of right triangles with vertices in the plane point set $\u A$, which in the real case was given a sharp answer by Pach and Sharir [@PS], by using the Szemerédi-Trotter theorem. Here we adapt the argument in order to use the Stevens-de Zeeuw incidence bound – the first estimate in – instead. Note that the second bound in always provides a universal bound $$\label{uni} \E(A)\ll \frac{|A|^3}{p} + |A|^2\sqrt{p}.$$ We next recast the Stevens-de Zeeuw incidence bound in the usual way, aiming at the cardinality $m_k$ of the set of $k$-rich lines, that is lines, supporting $\geq k$ points of a $n$-point set: $$m_k \ll \frac{n^{\frac{11}{4}}} {k^{\frac{15}{4}}} + \frac{n}{k} + \frac{n^{\frac{13}{2}}}{p^{\frac{15}{2}}}\,,$$ The third term in the bound arises as the alternative to the constraint $n^{13}m_k^{-2}\leq p^{15}$ of Lemma \[lem:2d\] when the Stevens-de Zeeuw theorem may not apply. Let us for technical purposes loosen this bound by subsuming the last term in the increased second term (clearly, $k\leq p$) as follows: $$\label{k-rich} m_k\ll \frac{n^{\frac{11}{4}}} {k^{\frac{15}{4}}} + \frac{n^{\frac{5}{4}}}{k} \qquad for \;\;\; n\leq p^{\frac{26}{21}}\,.$$ This apparently crude step is of little consequence but eases some calculations in the sequel: its positive effect is making the range of $n$ when the next key estimate is applicable somewhat wider, while the negative effect is that following estimate ends up being worse than it may be, yet still better than the desired . Next we are going to show that the number of nontrivial solutions (that is with $\u x\neq \u z$ and $\u y\neq \u z$) $N$ of equation satisfies the following bound: $$\label{thin} for \;\;\; n\leq p^{\frac{26}{21}}, \qquad N\ll n^{2+\frac{3}{7}}.$$ Assume $n\leq p^{\frac{26}{21}}$, let us express the quantity $N$ as follows. For $\u z\in A$, define $L_z$ as the set of all the $p+1$ lines in $F^2$ incident to $z$. For any line $l$ in $F^2$, let $n(l)$ be the number of points of $\u A$, supported on $l$ [*minus 1*]{}. Then $$\label{f:N} N = \sum_{\u z \in \u A}\, \sum_{l\in L_z} n(l) n(l^\perp),$$ where $l^\perp$ is the line perpendicular to $l$. Let us set up a cut-off value $k_*= n^{\frac{3}{7}}$ of $n(l)$ to be justified. Partition, for every $z$, the lines $l\in L_z$ to poor ones, that is those with $n(l) \leq k_*$, and otherwise rich. Accordingly partition $N=N_{poor} + N_{rich}$, where the term $N_{poor}$ means that at least one of $l,l^\perp$ under summation in is poor, hence the alternative $N_{rich}$ is when both $l,l^\perp$ are rich. Clearly $$N_{poor} \leq 2k_* n^2.$$ Let us now bound $N_{rich}$. Observe that the two terms in estimate meet when $k\sim n^{\frac{6}{11}}$. Let us call the lines with $n(l) \geq n^{\frac{6}{11}}$ [*very rich*]{} and partition $$N_{rich} = N_{very-rich}+N_{just-rich}\,,$$ the first term corresponding, for each $\u z$, to the sub-sum, corresponding to the case when one of $l,l^\perp$ is very rich. Then one can bound $N_{very-rich}$ trivially, using the second term in and dyadic summation in $k\geq n^{\frac{6}{11}}$ as $$\label{e:vr} N_{very-rich}\ll n \sum_{l:\,n(l)\geq n^{\frac{6}{11}}} n(l) \ll n^\frac{9}{4}\log n,$$ which is better than . Indeed, given a very-rich line $l$ we count all triangles with vertices $\u x, \u z,\u y$, such that $\u z\in l$ and $\u y$ is any point outside $l$; the two will determine the third vertex $\u x\in l$. What is left to consider is the case of the summation in when both $n^{\frac{3}{7}}\leq n(l),\,n(l^\perp)\leq n^{\frac{6}{11}}$. We apply Cauchy-Schwarz to obtain $$N_{just-rich} \leq \sum_{\u z \in \u A} \;\;\sum_{l\in L_z:\,p^{\frac{3}{7}}\leq n(l) \leq n^{\frac{6}{11} } } n^2(l) .$$ The expression in the right-hand side counts collinear triples of points in $\u A$ on rich, but not very rich lines. The number of such lines with $n(l)\sim k$, for the range of $n(l)$ in question is bounded by the first term in estimate . Multiplication of the latter term by $k^3$ followed by dyadic summation in $k$ yields $$N_{just-rich} \ll n^{\frac{11}{4}} k_*^{-\frac{3}{4}}\,,$$ optimising with $N_{poor}\leq n^2k_*$ justifies the choice of $k_*=n^{\frac{3}{7}}$, and proves . Together with the better bound this completes the proof of Lemma \[l:corners\]. Final remark {#final-remark .unnumbered} ------------ Mockenhaupt and Tao conjectured that in $d=3$, the sharp extension exponent from $L^2(\mathcal P,d\sigma)$ should be $3$. However, the strategy in this paper (as well as [@MoT], [@Lew], etc.) can only yield an extension exponent $>\frac{10}{3}$ (which one would get with the full strength Szemerédi-Trotter theorem). Indeed, as we have seen, the energy $\E(G_h)$ under summation in the right-hand side of equals the number of rectangles on the corresponding horizontal slice of the support $G$ of a bounded function $g$ on $F^3$. Choosing $g\equiv 1$ on its support and each horizontal slice of $G$ as a translate or dilate of the Cartesian product of the interval $[1,\ldots,N]$ with itself, the number of rectangles on each slice is $\gg N^4\log N$. An easy calculation with, say, choosing $N\sim p^{\frac{1}{3}}$ and having $p$ such slices of $G$ shows that the minimum of the estimates , and is always greater than $p^{\frac{7}{6}}$ times a power of $\log p$. This means that the restriction exponent is $<\frac{10}{7}$, hence the extension exponent from $L^2(\mathcal P,d\sigma)$ is $>\frac{10}{3}.$ [99]{} Cambridge Part III course notes, http://people.maths.ox.ac.uk/greenbj/papers/rkp.pdf, 2013. , [*A restriction estimate using polynomial partitioning,*]{} J. Amer. Math. Soc. [**29**]{}:2 (2016), 371-413. , arXiv:1603.04250v3 \[math.CA\] 2 Nov 2017. L. Guth, N. H. Katz. [*On the Erdős distinct distance problem in the plane*]{}. Ann. of Math. (2) [**181**]{} (2015), no. 1, 155–190. *Finite field restriction estimates for the paraboloid in high even dimensions,* arXiv:1712.05549v1 \[math.CA\] 15 Dec 2017. Adv. Math. [**270**]{} (2015), 457–479. *Finite field estimates based on Kakeya maximal operator estimates,* to appear in J. Eur. Math. Soc., arXiv:1401.8011v5 \[math.CA\] 2 Oct 2016. , Duke Math. J. [**121**]{}:1 (2004), 35–74. , Mosc. J. Comb. Number Theory, [**6**]{} (2016), 64–95. *A second wave of expanders in finite fields,* In: Nathanson M. (eds) Combinatorial and Additive Number Theory II. CANT 2015, CANT 2016. Springer Proceedings in Mathematics & Statistics, vol 220, 2017. *New results on sum-product type growth over fields,* arXiv:1702.01003v3 \[math.CO\] 9 Mar 2017. J. Combinatorial Theory Ser. A [**59**]{} (1992), 12–22. Combin., Prob., Computing [**17**]{}:4 (2008), 603–618. Combinatorica [**38**]{}:1 (2018), 219–254; https://doi.org/10.1007/s00493-016-3329-6 Bull. LMS [**49**]{} (2017), 842–858. The Szemerédi-Trotter type theorem and the sum-product estimate in finite fields, Eur. J. Combinatorics, [**32**]{}:8 (2011), 1177–1181.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We have mapped out a detailed phase diagram that shows the ground state structure of a spin-1 condensate with magnetic dipole-dipole interactions. We show that the interplay between the dipolar and the spin-exchange interactions induces a rich variety of quantum phases that exhibit spontaneous magnetic ordering in the form of intricate spin textures.' author: - 'S. Yi$^{1,2}$ and H. Pu$^1$' title: Spontaneous spin textures in dipolar spinor condensates --- Recent achievement of chromium condensate provides a platform for exploring the effects of magnetic dipolar interaction in dilute atomic Bose-Einstein condensates (BECs) [@Cr]. On the theoretical side, early studies on dipolar condensates [@scalar; @dipreview] concentrated on the scalar atomic BECs where the dipole moments are assumed to be polarized by, e.g., external magnetic fields. To unlock the magnetic dipole moments, one has to resort to optically confined spinor condensates [@dan; @spinor]. A widely used and very powerful theoretical tool in treating spinor condensates is the single mode approximation (SMA) in which different spin states are assumed to share the same spatial wave function [@yi1; @law]. An important implication, as well as a limitation, of the SMA is the lack of spatially varying spin textures, i.e., the spins are uniformly oriented under the SMA. The SMA is, however, expected to be valid only when the total spin-dependent interaction (spin-exchange and dipolar) is sufficiently weak compared to the spin-independent interaction [@yi2], and hence cannot cover the whole spectrum of interesting quantum spin phenomena. The SMA becomes particularly questionable in the presence of dipolar interactions — from the studies of ferromagnetic [@magdom] and ferroelectric [@ivan] materials, it is known that even a relatively weak dipolar interaction may spontaneously induce spatially varying dipole moments, a fact which can be attributed to the long-range anisotropic nature of the dipolar interaction. In the present work, we investigate the ground state wave function and spin structure of a dipolar spinor condensate by directly minimizing the mean-field energy functional without the assumption of the SMA. The main result is summarized in Fig. \[phase\] where a phase diagram in the parameter space of the relative dipolar strength and the trap aspect ratio is present. It can be seen that the competition between the long-range dipolar and the short-range exchange interactions gives rise to a rich variety of spin textures. These spin textures emerge [*spontaneously*]{} in the ground state, in contrast to the case of non-dipolar spinor condensates, where spin textures can only be induced dynamically [@yip; @matt; @stoof2]. We consider an $F=1$ condensate with three magnetic sublevels $m_F=1,0,-1$. In the mean-field treatment, the grand-canonical energy functional of the system can be expressed as $$\begin{aligned} E[\phi_\alpha] &=&\int d{\mathbf r}\left[\sum_\alpha\frac{\hbar^2 |\nabla\phi_\alpha|^2}{2m}-\mu n +Vn+\frac{c_0n^2}{2}+\frac{c_2{\mathbf S}^2}{2}\right]\nonumber\\ &&+\frac{c_d}{2}\int \frac{d{\mathbf r}d{\mathbf r}'}{|{\mathbf R}|^3}\left[{\mathbf S}({\mathbf r})\cdot{\mathbf S}({\mathbf r}')-\frac{3{\mathbf S}({\mathbf r})\cdot{\mathbf R}\,{\mathbf S}({\mathbf r}')\cdot{\mathbf R}}{|{\mathbf R}|^2}\right],\end{aligned}$$ where $\phi_\alpha({\mathbf r})=\sqrt{n_\alpha({\mathbf r})}\,e^{i\Theta_\alpha({\mathbf r})}$ is the macroscopic wave function of the $m_F=\alpha$ atoms with $n_\alpha$ being the density and $\Theta_\alpha$ being the phase of the component $\alpha$; $\mu$ is the chemical potential, $n=\sum_\alpha n_\alpha$ the total density, $V=m\omega_0^2(x^2+y^2+\lambda^2z^2)/2$ is trapping potential with $\lambda$ being the trap aspect ratio, ${\mathbf S}=\sum_{\alpha\beta}\phi^*_\alpha{\mathbf F}_{\alpha\beta}\phi_\beta$ is the density of spin with ${\mathbf F}$ being the angular momentum operator, and ${\mathbf R}={\mathbf r}-{\mathbf r}'$. The collisional interactions include a spin-independent part $c_0=4\pi\hbar^2(a_0+2a_2)/(3m)$ and a spin-exchange part $c_2=4\pi\hbar^2(a_2-a_0)/(3m)$ with $a_f$ ($f=0,2$) being the $s$-wave scattering length for two spin-1 atoms in the combined symmetric channel of total spin $f$ [@spinor]. For the two experimentally realized spin-1 condensates $^{87}$Rb [@chapman] and $^{23}$Na [@kurn], the spin-exchange interactions are ferromagnetic ($c_2<0$) and anti-ferromagnetic ($c_2>0$), respectively. The strength of the dipolar interaction is characterized by $c_d=\mu_0\mu_B^2g_F^2/(4\pi)$ with $\mu_0$ being the vacuum magnetic permeability, $\mu_B$ the Bohr magneton, and $g_F$ the Landé g-factor. The condensate wave function is found by numerically minimizing the energy functional $E[\phi_\alpha,\phi_\alpha^*]$ subject to the constraint of fixed total number of atoms $N$ using the steepest descent method. In our calculation, we fix the value of $c_0$ to be positive and choose $c_2/c_0=-0.01$ and $0.03$ for the case of the ferromagnetic and anti-ferromagnetic spin-exchange interaction, respectively. These two ratios correspond to the scattering parameters of $^{87}$Rb and $^{23}$Na, respectively [@kemp]. To focus on the dipolar effects, we allow $c_d$ to vary and introduce $q\equiv c_d/|c_2|\geq0$ to measure the relative strength of the dipolar interaction. The trap aspect ratio $\lambda$ (which is varied between $0.1$ and $10$ in our calculations) is another control knob that allows us to tune the shape of the condensate and hence the effective dipolar interaction [@dipreview; @yi1]. Finally, in all numerical results presented in this paper, we use $N=10^5$ and $\omega_0=2\pi\times100$Hz. Figure \[phase\] summarizes the main result of this paper in the form the phase diagram of the dipolar spin-1 condensate in the $q$-$\lambda$ parameter space. For the SMA-I, II, and III phases, the SMA is found to be valid and the spins are uniformly oriented towards a direction determined by the dipolar interaction. In the phases labelled by C, S, and P, the spatial wave functions are spin-dependent and the system exhibits rich spin textures. ![Phase diagram of dipolar spin-1 condensates of ferromagnetic (a) and anti-ferromagnetic (b) spin-exchange interaction. ${\bf s}={\bf S}/n$ is the normalized local spin vector.[]{data-label="phase"}](phases.eps){width="3.3in"} [*SMA ground states*]{} — The SMA implies that the normalized local spin vector ${\mathbf s}\equiv{\mathbf S}/n$ is a constant. For a ferromagnetic system in the absence of the dipolar interaction, the vector ${\bf s}$ has a length of unity and the ground state is degenerate about the orientation of $\mathbf s$ due to the SO(3) symmetry possessed by the order parameter [@spinor]. When the dipolar interaction is present, the SO(3) symmetry is broken and, as shown in Fig. \[phase\](a), the SMA is valid only when $q$ is below a $\lambda$-dependent critical value. In the SMA regions, the direction of ${\bf s}$ is determined by the trapping geometry: ${\mathbf s}$ is perpendicular to the $z$-axis for a pancake-shaped ($\lambda>1$) condensate (SMA-I) and parallel to the $z$-axis for a cigar-shaped ($\lambda<1$) condensate (SMA-II), reminiscent of a quantum ferromagnet with easy-plane and easy-axis anisotropy, respectively. This result agrees with the quantum mechanical calculation under the SMA [@yi1]. As shown in Fig. \[phase\](b), for an anti-ferromagnetic system, the SMA is again valid for small $q$. When $q$ is below a critical value that is rather insensitive to $\lambda$, the system exhibits a vanishing spin with ${\mathbf s}=0$ (SMA-III). For $q$ above this critical value and a cigar-shaped trapping potential with $\lambda \lesssim 0.4$, the dipolar interaction may overwhelm the spin-exchange interaction and make the condensate effectively ferromagnetic, and the system enters the SMA-II phase with spin oriented along the $z$-axis. We do not find in this case the SMA-I phase with ${\bf s} \perp {\bf z}$. It is also worth mentioning that transitions between different SMA phases are all first order. Figure \[phase\] shows that the critical dipolar strength at which the SMA becomes invalid decreases as $\lambda$. For a ferromagnetic system, $q_{\rm cr}=0.1$ at $\lambda=10$, which can be reached in $^{87}$Rb [@yi1]. The dipolar effects thus become more prominent as the condensate becomes more two-dimensional (2D) pancake-like. This is consistent with the study in solid magnetic materials, where the dipolar interaction, normally weak enough to be ignored in bulk materials, plays an essential role in stabilizing long-range magnetic order in 2D systems, e.g., in magnetic thin films [@film]. Our work shows that it is feasible to observe non-SMA ground state in a pancake $^{87}$Rb condensate. ![(Color online). Ground state of the P phase for ferromagnetic coupling with $\lambda=2$ and $q=1$. (a) Density of each spin component along the $x$-axis. The total density is denoted by the dashed line. (b) and (c) are the respective phase images for $\phi_1$ and $\phi_{-1}$ in the $z=0$ plane. The phase of $\phi_0$ is a constant and is not shown here. The phases in the $z\neq 0$ plane are similar. (d) The vector plot of ${\mathbf S}$ in the $xy$-plane with the color map corresponding to the total density. Here, as well as in Figs. \[nsma2\] and \[nsma3\], we adopt $a_{\rm ho}=(\hbar/m\omega_0)^{1/2}$ and $N/a_{\rm ho}^3$ as the units for length and density, respectively.[]{data-label="nsma1"}](nsma1.eps){width="2.4in"} [*Non-SMA ground states*]{} — The SMA regime is analogous to the single-domain regime in micromagnetics, where the exchange energy dominates over all other spin-dependent energy terms and a uniform magnetization forms as a result [@magdom]. As $q$ increases, the SMA eventually breaks down. At large $q$, the dipolar interaction will dominate the spin-dependent interaction and the difference between the ferromagnetic and anti-ferromagnetic spin-exchange interaction becomes insignificant, as can be seen from Fig. \[phase\]. The non-SMA region is further divided into three distinct phases P, C and S, which stands for pancake, cigar and spherical, respectively (see below). Figure \[nsma1\] shows a typical wave function in the P phase which occurs in pancake geometry. The density profiles of all spin components are axially symmetric with $n_1({\bf r})= n_{-1}({\bf r})$ and hence $s_z=0$, i.e., the spins are planar and lie in the $xy$-plane. The phases $\Theta_\alpha$ take the form $$\begin{aligned} \Theta_\alpha=w_\alpha\varphi+\varphi_\alpha,\label{phase}\end{aligned}$$ where $\varphi$ is the azimuthal angle, $\varphi_\alpha$ is a constant phase that satisfies $$\varphi_1+\varphi_{-1}-2\varphi_0 = 0 \,, \label{relphase}$$ and $w_\alpha$ is the phase winding number with the values $$\langle w_1,w_0,w_{-1}\rangle=\langle-1,0,1\rangle\,.$$ Therefore the two spin components, $\phi_1$ and $\phi_{-1}$, are in vortex states with opposite winding numbers. Due to the presence of the non-rotating $\phi_0$ component, the total density does not vanish at the vortex core. Such a ground state represents a coreless skrymion. It is easy to show that the spin vector can be expressed as $$\begin{aligned} {\mathbf S}=2\sqrt{2n_0 n_1}\,(\sin\varphi,\;-\cos\varphi,\;0)\,.\nonumber\end{aligned}$$ As illustrated in Fig. \[nsma1\](d), the spin continuously curls around the $z$-axis to form a vortex such that the system possesses a persistent spin current but no net density current. The spin vortex state is analogous to the flux-closure magnetic states in micromagnetics [@magdom]. In the vicinity of the $z$-axis, to reduce the exchange energy, the magnitude of the spin gradually decreases and vanishes on the $z$-axis. This differs from the magnetic vortex observed in nanoscale ferromagnets [@buss] where, due to the conservation of local spin moment, the magnetization in the vortex core turns toward the $z$-axis [@magdom]. ![(Color online). (a) The streamline plot of the spins for ferromagnetic coupling in the C phase with $\lambda=0.4$ and $q=1.6$. The color map represents $|{\mathbf S}_\perp|$ on these cross-sections. (b) The vector plot of ${\mathbf s}_\parallel$ in the $xz$-plane. Same parameters as in (a). (c) The same as (b) except for $q=6$.[]{data-label="nsma2"}](nsma2.eps){width="2.5in"} In the C phase which occurs in a cigar-shaped trap, $n_\alpha$ remains to be cylindrically symmetric and $\Theta_\alpha$ can also be expressed in the form of Eq. (\[phase\]) where the phases $\varphi_\alpha$ are now $z$-dependent but still satisfy Eq. (\[relphase\]), and the winding numbers are now given by $$\langle w_1,w_0,w_{-1}\rangle=\langle 0,1,2\rangle\,.$$ The spin vector in the C phase takes the form $$\begin{aligned} {\mathbf S}=\big[\Delta \sin(\varphi+\delta),\; -\Delta \cos(\varphi+\delta),\;n_1-n_{-1}\big]\,,\nonumber\end{aligned}$$ where $\Delta \equiv \sqrt{2n_0}(\sqrt{n_1}+\sqrt{n_{-1}})$, and $\delta(z)\equiv\varphi_0(z)-\varphi_1(z)$ is the spin twisting angle which is a monotonically increasing function of $z$ with $\delta(z=0)=0$. A typical spin texture plot in the C phase is shown in Fig. \[nsma2\](a) where one can see that the spins twist around the $z$-axis, tracing out a helical pattern. The total twisting angle $\delta(\infty)-\delta(-\infty)$ increases with $q$ and approaches $\pi$ at large $q$. The corresponding structures of the planar spin ${\mathbf S}_\parallel\equiv(S_x,S_z)$ are plotted in Fig. \[nsma2\](b) and (c), which resembles the magnetization of a bar magnet. Note that, in the absence of the spin-exchange interaction ($c_2=0$), the C phase does not exist and is replaced by the SMA-II phase with spins uniformly oriented along $z$. ![(Color online). The spin structure in the S phase for anti-ferromagnetic coupling. (a) Spin projection in the $z=2.29$ plane. $\lambda=0.4$ and $q=1$. (b) Spin projection in the $y=0$ plane. $\lambda=0.8$ and $q=1$. (c) The spin structure for $\lambda=0.8$ and $q=1$. (d)-(f) The dipolar strength dependence of the toroidal moments for $\lambda=1.2$, $0.8$, and $0.4$, respectively.[]{data-label="nsma3"}](nsma3.eps){width="2.5in"} In between the P and C phases, is the S phase which occurs when the trapping potential is close to spherical ($\lambda \approx 1$). A distinct feature of the S phase is that $n_\alpha$ becomes non-axisymmetric, signalling the broken of the cylindrical symmetry of the spatial wave functions. To further quantify the spin texture, we define the spin toroidal moment vector as [@ivan] $${\mathbf G}\equiv\int d{\mathbf r}\,\left[{\mathbf r}\times{\mathbf S}({\bf r})\right]\,.$$For all the SMA phases, we have ${\bf G}=0$. For both the P and C phases, ${\bf G}$ points along the $z$-axis. The S phase, by contrast, features finite $G_x$ and $G_y$. The toroidal moments as functions of $q$ is plotted in Fig. \[nsma3\](d)-(f) for three different values of $\lambda$. For small $q$, ${\bf G}=0$ and the system is in the SMA regime. As $q$ exceeds a threshold, ${\bf G} \neq 0$ and the S phase is reached. The continuity of ${\bf G}$ suggests, in this mean-field calculation, the transitions between SMA and non-SMA, and also among non-SMA, phases are of second order. In Fig. \[nsma3\](d), we also see that for $q>0.6$, the system enters the P phase. Fig. \[nsma3\](a)-(c) show examples of the spin structure in the S phase. For this particular set of parameters, the condensate displays two domains: one has $S_z>0$ and and the other $S_z<0$. [*Collapse*]{} — Condensate collapse will occur when the dipolar interaction strength exceeds a critical value $c_d^*$. For both ferromagnetic and anti-ferromagnetic coupling, we have found that $c_d^*\approx0.24c_0$ and is nearly independent of the trap aspect ration $\lambda$. This critical value agrees with that of a highly elongated dipolar scalar condensate with dipole moments polarized along the axial direction [@scalar]. As is well known, the critical dipolar strength for scalar condensate is very sensitive to $\lambda$ [@scalar]. In the spinor system studied here, the dipoles are free to rearrange themselves to minimize the dipolar interaction, which renders the insensitivity of $c_d^*$ with respect to the trapping geometry. [*Dipole induced spin-orbital coupling*]{} — In the SMA regime, the total orbital angular momentum [**L**]{} vanishes, and the ground state wave functions $\phi_\alpha$ can be taken to be real. In the non-SMA regime, the spin and orbital degrees of freedom are intimately coupled and $\phi_\alpha$ must be described by complex functions. Here the spin (${\bf S}^2$) and orbital (${\bf L}^2$) angular momenta are not separately conserved, only the total angular momentum $({\bf S}+{\bf L})^2$ is conserved. It is not difficult to show that, in the non-SMA regime, if the set of wave functions $\{\phi_\alpha\}$ minimizes the energy functional, then the set $\{\phi_\alpha^*\}$ does not. On the other hand, if we make the transformation: $ (\phi_1, \phi_0, \phi_{-1}) \rightarrow (\phi_{-1}^*, -\phi_0^*, \phi_{1}^*)$, then the new set still minimizes the energy. The above transformation amounts to inverting ${\bf S}$ and ${\bf L}$ [*simultaneously*]{} and hence preserving the relative orientation between them. This inter-connection between ${\bf S}$ and ${\bf L}$ is well known in superfluid $^3$He and the related phenomenon has been termed the spontaneously broken spin-orbit symmetry [@he3; @legg]. The dipole induced spin-orbital coupling is recently explored theoretically in the dynamics of $^{52}$Cr condensate and is predicted to manifest in the Einstein-de Haas effect [@haas]. $^{52}$Cr atom features a spin-3 ground state [@diener] whose short-range collisional interaction is characterized by four scattering lengths and not all of them are accurately known. Although we have studied, in this work, a relatively simpler spin-1 system, we expect much of the essential physics can be applied to higher spin systems. In conclusion, we have provided a detailed phase diagram which shows the ground state structure of a dipolar spin-1 condensate beyond the SMA. In this system, the spin and orbital degrees of freedom are intimately coupled together. The interplay between the long-range dipolar and the short-range exchange interactions gives rise to a variety of quantum phases characterized by distinctive spin textures. As such, dipolar spinor condensates represent an intriguing quantum magnetic system whose properties are highly tunable. The work here assumes zero magnetic field, an assumption valid for $^{87}$Rb and $^{52}$Cr when the external magnetic field strength is less than $0.1\,$mG [@yi1]. This constraint poses an experimental challenge but is definitely within the reach with current technology. We thank Profs. Michael Chapman, Jason Ho, Carl Rau and Li You, and Dr. Jian Li for many insightful discussions. HP acknowledges the hospitality of the Aspen Center for Physics, and support from ORAU and NSF. A. Griesmaier [*et al.*]{}, Phys. Rev. Lett. [**94**]{}, 160401 (2004). S. Yi and L. You, Phys. Rev. A [**61**]{}, 041604(R) (2000); [*ibid*]{}. [**63**]{}, 053607 (2001); K. Goral [*et al*]{}., [*ibid*]{}. [**61**]{}, 051601(R) (2000); L. Santos [*et al*]{}., Phys. Rev. Lett. [**85**]{}, 1791 (2000). For a review, see, M. Baranov [*et al.*]{}, Phys. Scripta. [**T102**]{}, 74 (2002). D. Stamper-Kurn [*et al*]{}., Phys. Rev. Lett. [**80**]{}, 2027 (1998). T.-L. Ho, Phys. Rev. Lett. [**81**]{}, 742 (1998); T. Ohmi and K. Machida, J. Phys. Soc. Jap. [**67**]{}, 1822 (1998). S. Yi, L. You, and H. Pu, Phys. Rev. Lett. [**93**]{}, 040403 (2004); S. Yi and H. Pu, Phys. Rev. A [**73**]{}, 023602 (2006). C. K. Law, H. Pu, and N. P. Bigelow, Phys. Rev. Lett. [**81**]{}, 5257 (1998). S. Yi [*et al.*]{}, Phys. Rev. A [**66**]{}, 011601 (2002). A. Hubert and R. Schäfer, [*Magnetic domains*]{} (Springer, Berlin, 1998). I. I. Naumov, L. Bellaiche, and H. Fu, Nature (London) [**432**]{}, 737 (2004); A. A. Gorbatsevich and Yu. V. Kopaev, Ferroelectrics [**161**]{}, 321 (1994). S.-K. Yip, Phys. Rev. Lett. [**83**]{}, 4677 (1999); T. Mizushima, K. Machida, and T. Kita, [*ibid*]{}. [**89**]{}, 030401 (2002); E. N. Bulgakov and A. F. Sadreev, [*ibid*]{}. [**90**]{}, 200401 (2003). M. R. Matthews [*et al*]{}., Phys. Rev. Lett. [**83**]{}, 2498 (1999); A. E. Leanhardt [*et al*]{}., [*ibid*]{}. [**90**]{}, 140403 (2003). H. T. C. Stoof, E. Vliegen, and U. Al Khawaja, Phys. Rev. Lett. [**87**]{}, 120407 (2001); J. Ruostekoski and J. R. Anglin, Phys. Rev. Lett. [**91**]{}, 190402 (2003); H. Zhai [*et al*]{}., Phys. Rev. A [**68**]{}, 043602 (2003). M. D. Barret, J. A. Sauer, and M. S. Chapman, Phys. Rev. Lett. [**87**]{}, 010404 (2001). D. M. Stamper-Kurn [*et al*]{}., Phys. Rev. Lett. [**80**]{}, 2027 (1998); J. Stenger [*et al*]{}., Nature (London) [**396**]{}, 345 (1998). E.G.M. van Kempen [*et al*]{}., Phys. Rev. Lett. [**88**]{}, 093201 (2002); A. Crubellier [*et al*]{}., Eur. Phys. J. D [**6**]{}, 211 (1999). K. De’Bell, A. B. Maclsaac and J. P. Whitehead, Rev. Mod. Phys. [**72**]{}, 225 (2000). K. Bussman[*et al*]{}., Appl. Phys. Lett. [**75**]{}, 2476 (1999); J. Raabe [*et al*]{}., J. Appl. Phys. [**88**]{}, 4437 (2000); T. Shinjo [*et al*]{}., Science [**289**]{}, 930 (2000); A. Wachoviak [*et al*]{}., Science [**298**]{}, 577 (2002). D. Vollhardt and P. Wölfle, [*The superfluid phases of Helium 3*]{} (Taylor & Francis, London, 1990). A. J. Leggett, Rev. Mod. Phys. [**47**]{}, 331 (1975). L. Santos and T. Pfau, cond-mat/0510634; Y. Kawaguchi, H. Saito and M. Ueda, Phys. Rev. Lett. [**96**]{}, 080405 (2006). R. B. Diener and T.-L. Ho, cond-mat/0511751.
{ "pile_set_name": "ArXiv" }
--- abstract: | [We present a mechanism to generate the baryon asymmetry of the Universe which preserves the net baryon number created in the Big Bang. If dark matter particles carry baryon number $B_X$, and $\sigma^{\rm ann}_{\bar{X}} < \sigma^{\rm ann}_{X } $, the $\bar{ X}$’s freeze out at a higher temperature and have a larger relic density than $X$’s. If $m_X {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}4.5 \,B_X \,$GeV and the annihilation cross sections differ by $\mathcal{O}$(10%) or more, this type of scenario naturally explains the observed $\Omega_{DM} \approx 5\, \Omega_b$. ]{} author: - 'Glennys R. Farrar and Gabrijela Zaharijas' title: Dark Matter and the Baryon Asymmetry of the Universe --- The abundance of baryons and dark matter (DM) in our Universe poses several challenging puzzles:\ $\bullet$ Why is there a non-zero net nucleon density and what determines its value?\ $\bullet$ What does dark matter consist of?\ $\bullet$ Is it an accident that the dark matter density is roughly comparable to the nucleon density, $\rho_{DM} = 5 ~\rho_N$? As pointed out by Sakharov[@sakharov], baryogensis requires three conditions: non-conservation of baryon number, violation of C and CP, and a departure from thermal equilibrium. The last is provided by the expansion of the Universe and the first two are naturally present in unified theories and even in the standard model at temperatures above the electroweak phase transition. However in most approaches the origins of DM and the BAU are completely unrelated and their densities could naturally differ by many orders of magnitude. In this Letter we propose a new type of scenario, in which the observed baryon asymmetry is due to the separation of baryon number between ordinary matter and dark matter and not to a net change in the total baryon number since the Big Bang. (See [@other] for other papers with this aim.) Thus the abundances of nucleons and dark matter are related. The first Sakharov condition is not required, while the last two remain essential. We give explicit examples in which anti-baryon number is sequestered at temperatures of order 100 MeV. The CPT theorem requires that the total interaction rate of any ensemble of particles and antiparticles is the same as for the conjugate state in which each particle is replaced by its antiparticle and all spins are reversed. However individual channels need not have the same rate so, when CP is violated, the annihilation rates of the CP reversed systems are not in general equal. A difference in the annihilation cross section, $\sigma^{\rm ann}_{\bar{X}} < \sigma^{\rm ann}_{X } $, means that the freeze out temperature for $X$’s ($T_X$) is lower than for $\bar{X}$’s ($T_{\bar{X}}$). After the $\bar{X}$’s freeze out, the $X$’s continue to annihilate until the temperature drops to $T_{X}$, removing $B_X$ antinucleons for each $X$ which annihilates. Assuming there are no other significant contributions to the DM density, the present values $n_{o\, N}$, $n_{o\, X}$ and $n_{o\, \bar{X}}$ are determined in terms of $m_X$, $B_X$ and the observables $ \frac{\Omega_{DM}}{\Omega_b}$ and $\frac{n_{o\,N}}{n_{o\, \gamma}} \equiv \eta_{10} \,10^{-10}$ or $\rho_{\rm crit}$. (Following common usage, the subscript $b$ refers to the baryon number in ordinary matter.) From WMAP, $\eta_{10} = 6.5^{+0.4}_{-0.3}$, $\Omega_m h^2 = 0.14 \pm 0.02$, $\Omega_b h^2 = 0.024 \pm 0.001$[@WMAP], so $\frac{\Omega_{DM}}{\Omega_b} = 4.83 \pm 0.87$. Given the values of these observables, we can “reverse engineer" the process of baryon-number-segregation. For brevity, suppose there is only one significant species of DM particle. Let us define $\epsilon = \frac{n_X}{n_{\bar{X}} }$. Then the total energy density in $X$’s and $\bar{X}$’s is $\rho_{DM} = m_X n_{\bar{X}} (1 + \epsilon)$. By hypothesis, the baryon number density in nucleons equals the antibaryon number density in $X $ and $\bar{X}$’s, so $ B_X n_{\bar{X}} (1-\epsilon) = (n_N - n_{\bar{N}}) = \frac{\rho_b}{m_N}$. Thus $$\label{kappa} \frac{\Omega_{DM}}{\Omega_b} = \left( \frac{1 + \epsilon}{1 - \epsilon} \right) \frac{m_X}{m_N B_X}.$$ As long as the DM particle mass is of order hadronic masses and $\epsilon $ is not close to 1, this type of scenario naturally accounts for the fact that the DM and ordinary matter densities are of the same order of magnitude. Furthermore, since $ \frac{1 + \epsilon}{1 - \epsilon} \ge 1$, the DM density in this scenario must be [*greater*]{} than the nucleonic density as observed, unless $m_X < m_N B_X$. Given the parameters of our Universe, we can instead write (\[kappa\]) as an equation for the DM mass $ m_X = \left( \frac{1 - \epsilon}{1 + \epsilon} \right) \frac{\Omega_{DM}}{\Omega_b} \, B_X m_N . $ For low baryon number, $B_X = 1\, (2)$, this implies $m_X {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}4.5 \,(9)\,$GeV. If dark matter has other components in addition to the $X$ and $\bar{X}$, the $X$ must be lighter still. The observed BAU can be due to baryon number sequestration with heavy DM only if $B_X$ is very large, e.g., strangelets or Q-balls. However segregating the baryon number in such cases is challenging. As an existence proof and to focus our discussion of the issues, we present two concrete scenarios. In the first, $X$ is a particle already postulated in QCD, the $H$ dibaryon ([*uuddss*]{}). New particle physics is necessary, however, because the CP violation of the Standard Model via the CKM matrix cannot produce the required $\mathcal{O}$(20%) difference in annihilation cross sections, since only the first two generations of quarks are involved. (Readers interested in why the $H$ may have escaped detection in particle physics searches and why its lifetime could be of order the age of the Universe are refered to [@fz:H] and references therein.) The second scenario postulates a new particle with mass ${\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}4.5$ GeV, which couples to quarks through dimension-6 operators coming from beyond-the-standard-model physics. In this case CP violation is naturally large enough, $\mathcal{O}$(10%), because all three quark generations are involved and, in addition, the new interactions in general violate CP. After deducing the properties required of these particles by cosmology, we discuss detection methods and then briefly mention particle physics aspects. As we shall show, the $H,\,\bar{H}$ scenario can already be ruled out by limits on the heat production in Uranus. The annihilation rate of particles of type $j$ with particles of type $i$ is $\Gamma^{ann}_{j}(T) = \Sigma_i ~n_i(T) <\sigma_{ij}^{ann} v_{ij}>$, where $<...>$ indicates a thermal average and $v_{ij}$ is the relative velocity. As the Universe cools, the densities of all the particle species $i$ decrease and eventually the rate of even the most important annihilation reaction falls below the expansion rate of the Universe. The temperature at which this occurs is called the freezeout temperature $T_j$ defined by $ \Gamma^{ann}_{j}(T_j) = H(T_j) = 1.66 \sqrt{g_*} ~ T_j^2/ M_{Pl} $, where $g_*$ is the effective number of relativistic degrees of freedom[@kolbTurner]. Between a few MeV and the QCD phase transition only neutrinos, $e^\pm$ and $\gamma$’s are in equillibrium and $g_* = 10.75$. Above the QCD phase transition which is presumably around 150 MeV, light quarks and antiquarks ($q, \, \bar{q}$) and $c, \, \bar{c}$, $\mu^\pm$ are also in equilibrium, giving $g_* = 56.25$. Above $T_b \approx 170$ MeV, $b$ and $\bar{b}$ quarks are kept in equilibrium through $ q~\bar{q} \leftrightarrow b ~ \bar{b} $ giving $g_* = 66.75$; below this temperature production cannot keep up with decay. The equilibrium density at freeze out temperature, $n_j(T_j)$, is a good estimate of the relic abundance of the $j$th species[@kolbTurner]. A key element of baryon-number sequestration is that self-annihilation cannot be important for maintaining equilibrium prior to freeze out. This is easily satisfied as long as $\sigma_{\bar{X}X}^{\rm ann}$ is not much greater than $\sigma_{\bar{X}N}^{\rm ann}$ , since at freezeout in the $H,\,\bar{H}$ and “$X_4$" scenarios, $n_{H ,\, \bar{H} } {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}10^{-5} n_{N ,\, \bar{N} }$ and $n_{X_4 ,\, \bar{X_4 } }\sim 10^{-11} n_{d ,\, \bar{d} }$. Given $m_X,\, B_X$ and $g_X$ (the number of degrees of freedom of the $X$ particle) and associated densities $n_{\{X,\bar{X}\}}$, the temperature $T_{\bar{X}}$ at which $\bar{X}$’s must freeze out of thermal equilibrium satisfies: $$\label{Xbarfo} \frac{n_{\bar{X}} - n_X}{n_{\bar{X}}} \frac{n_{\bar{X}}}{n_\gamma } = (1-\epsilon) \frac{ \pi^2 g_X x_{\bar{X}}^{3/2} e^{-x_{\bar{X}} } }{2 \zeta(3) (2 \pi)^{3/2} }= \frac{10.75}{3.91}\frac{\eta_{10} 10^{-10} }{B_X} ,$$ where $x_{\bar{X}} \equiv m_X/T_{\bar{X}}$. $ \frac{10.75}{3.91}$ is the factor by which $\frac{n_b}{n_\gamma}$ increases above $e^\pm$ annihilation. The equation for $X$ freezeout is the same, with $(1-\epsilon) \rightarrow (1-\epsilon)/\epsilon $. Freezeout parameters for our specific models are given in Table I; $\tilde{\sigma} \equiv \langle \sigma^{\rm ann} v \rangle / \langle v \rangle$ is averaged over the relevant distribution of c.m. kinetic energies, thermal at $\approx 100$ MeV for freezeout, or the degraded halo distribution with $v_H {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}11$km/s for $H,\,\bar{H}$ DM detection discussed next. If $X$s interact weakly with nucleons, standard WIMP searches constrain the low energy scattering cross section $\sigma_{DM} \equiv (\sigma^{\rm el}_{\bar{X} N} + \epsilon \sigma^{\rm el}_{XN})/(1+ \epsilon)$. However if the $X$ is a hadron, multiple scattering in the earth or atmosphere above the detector can cause a significant fraction to reflect or be degraded to below threshold energy before reaching a deep underground detector. Scattering also greatly enhances DM capture by Earth, since only a small fraction of the halo velocities are less than $v_{\rm esc}^{E} = 11$ km/s. Table I gives the total fluxes and the factor $f_{\rm cap}$ by which the flux of captured $\bar{X}$’s is lower, for the two scenarios. These are the result of integrating the conservative halo velocity distribution[@zf:window]. A comprehensive reanalysis of DM cross section limits including the effect of multiple scattering has recently been given in ref. [@zf:window]. A window in the DM exclusion was discovered for $m_X {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}2.4$ GeV and $ \tilde{\sigma}_{DM} \approx 0.3 - 1\, \mu $b; otherwise, if the DM mass ${\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}5$ GeV, $\tilde{\sigma}_{DM}$ must be $ {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}10^{-38} {\rm cm}^2$. Since $\sigma_{\{X,\bar{X} \}N}$ is negligible compared to $\sigma_{NN}$ and the $X,\,\bar{X}$ do not bind to nuclei[@fz:nucbind], nucleosynthesis works the same in these scenarios as with standard CDM. Primordial light element abundances constrain the [*nucleon*]{} – not [*baryon*]{} – to photon ratio! \[table\] Model $T_{\bar{X}}$ MeV $T_{X}$ $\tilde{\sigma}^{\rm ann}_{\bar{X}}$ cm$^2$ $\tilde{\sigma}^{\rm ann}_{{X}}$ $\Phi_{\bar{X}} ({\rm cm}^2 s)^{-1}$ $f^E_{\rm cap}$ ------- ------------------- --------- --------------------------------------------- ---------------------------------- -------------------------------------- ------------------- H 86.3 84.5 $2.2~10^{-41}$ $2.8~10^{-41}$ $3.7 \times 10^5$ 0.20 $X_4$ 180 159 $3.3~10^{-45}$ $3.7~10^{-45}$ $ 2.9 \times 10^5$ $2 \, 10^{-6} $ : Freezeout parameters, solar system average $\bar{X}$ flux, and fraction of flux captured in Earth, in two models. The CPT theorem requires that $\sigma^{\rm ann}_{X} + \sigma^{\rm non-ann}_{X} = \sigma^{\rm ann}_{\bar{X}} + \sigma^{\rm non-ann}_{\bar{X}}$. Therefore a non-trivial consistency condition in this scenario is $ \sigma^{\rm ann}_{X} - \sigma^{\rm ann}_{\bar{X}} \le \sigma^{\rm non-ann}_{\bar{X}} $. The value of the LHS needed for B-sequestration from Table I is compatible with the upper limits on the RHS from DM searches, and $\sigma^{\rm non-ann}_{\bar{X}} \ge \sigma^{\rm el}_{\bar{X}} $, so no fine-tuning is required to satisfy CPT. B-sequestration scenarios imply the possiblity of detectable annihilation of $\bar{X}$’s with nucleons in the Earth, Sun or galactic center. The rate of $\bar{X}$ annihilation in an Earth-based detector is the $\bar{X}$ flux at the detector, times $\sigma_{\bar{X}N}^{\rm ann}$, times (since annihilation is incoherent) the number of target nucleons in the detector, $6 \times 10^{32} $ per kton. The final products of $\bar{X} N$ annihilation are mostly pions and some kaons, with energies of order 0.1 to 1 GeV. The main background in SuperK at these energies comes from atmospheric neutrino interactions whose level is $\sim100$ events per kton-yr[@SKatmneutrinos]. Taking $\Phi^{SK}_{\bar{H}} \approx f_{\rm cap} \Phi_{\bar{H}}$ and $\Phi^{SK}_{\bar{X_4}} = \Phi_{\bar{X_4}}$ from Table I, the annihilation rate in SuperK is lower than the background if $\tilde{\sigma}^{\rm ann}_{\bar{H}N} \le 6 \times 10^{-44}\, {\rm cm}^2$ and $\tilde{\sigma}^{\rm ann}_{\bar{X_{4}}N} \le 2 \times 10^{-44}\, {\rm cm}^2$. The total energy release of $m_X + B_X m_N $ should give a dramatic signal, so it should be possible for SuperK to improve this limit. Note that for the $H,\,\bar{H}$ scenario this limit is already uncomfortable, since it is so much lower than the effective cross section required at freezeout. However this cannot be regarded as fatal, until one can exclude with certainty the possibility that the annihilation cross section is very strongly energy dependent. Besides direct observation of annihilation with nucleons in a detector, constraints can be placed from indirect effects of $\bar{X}$ annihilation in concentrations of nucleons. We first discuss the photons and neutrinos which are produced by decay of annihilation products. The signal is proportional to the number of nucleons divided by the square of the distance to the source, so Earth is a thousand-fold better source for a neutrino signal than is the Sun, all other things being equal. Since $\gamma$’s created by annihilation in the Earth or Sun cannot escape, the galactic center is the best source of $\gamma$’s; we do not pursue this here because [*i)*]{} the constraints above imply the signal is several orders of magnitude below present detector capabilities and [*ii)*]{} $\sigma^{\rm ann}_{\bar{X_4}N}$ is model dependent and may be very small, as discussed below. The rate of observable neutrino interactions in SuperK is $$\label{neutintsSK} \Gamma_{\nu{ \rm SK}} = N_{{\rm SK}}\, \Sigma_i \int{ \frac{d n_{\nu_i}}{dE} \sigma^{\rm eff}_{\nu_i N} \Phi_{\nu_i} dE },$$ where the sum is over neutrino types, $N_{{\rm SK}}$ is the total number of nucleons in SuperK, $\frac{d n_{\nu_i}}{dE}$ is the spectrum of $i$-type neutrinos from an $\bar{X}$ annihilation, $\sigma^{\rm eff}_{\nu_i N} $ is the neutrino interaction cross section summed over observable final states (weighted by efficiency if computing the rate of observed neutrinos), and $ \Phi_{\nu_i} $ is the $\nu_i$ flux at SK. This last is $f_{\nu_i}$, the mean effective number of $\nu_i$’s produced in each $\bar{X}$ annihilation discussed below, times the total rate of $\bar{X}$ annihilation in the source, $\Gamma^{\rm ann}_{\bar{X},s}$, divided by $\approx 4 \pi R_s^2$, where $R_s$ is the distance from source to SuperK; $R_s \approx R_E$ for annihilation in Earth. In general, computation of the annihilation rate $\Gamma^{\rm ann}_{\bar{X},s}$ is a complex task because it involves solving the transport equation by which DM is captured at the surface, migrates toward the core and annihilates, eventually evolving to a steady state distribution. However if the characteristic time for a DM particle to annihilate, $\tau^{\rm ann} = \,<\sigma^{\rm ann} n_N v>^{-1}$, is short compared to the age of the system, equilibrium between annihilation and capture is established (evaporation can be neglected for $M_{DM} {\,\raisebox{-0.13cm}{$\stackrel{\textstyle>} {\textstyle\sim}$}\,}\mathcal{O}$(GeV)[@krauss]) so $\Gamma^{\rm ann}_{\bar{X},E}$ equals $f_{\rm cap} \Phi_{\bar{X}} 4 \pi R_E^2$. Then the neutrino flux (\[neutintsSK\]) is independent of $\sigma^{\rm ann}_{\bar{X}N},$ because the annihilation rate per $\bar{X}$ is proportional to it but the equilibrium number of $\bar{X}$’s in Earth is inversely proportional to it. For Earth, the equilibrium assumption is applicable for $\tilde{\sigma}^{\rm ann} {\,\raisebox{-0.13cm}{$\stackrel{\textstyle>} {\textstyle\sim}$}\,}5 \times 10^{-49} {\rm cm}^2$, while for the Sun it is applicable if, roughly, $\tilde{\sigma}^{\rm ann} {\,\raisebox{-0.13cm}{$\stackrel{\textstyle>} {\textstyle\sim}$}\,}10^{-52} {\rm cm}^2$. For lower annihilation cross sections, transport must be treated to get an accurate estimate, but the equilibrium rate is an upper limit. The final state in $\bar{H} N$ annihilation is expected to contain $\bar{\Lambda}$ or $\bar{\Sigma}$ and a kaon, or $\bar{\Xi}$ and a pion, and perhaps additional pions. In a dense environment such as the core of the Earth, the antihyperon annihilates with a nucleon, producing pions and at least one kaon. In a low density environment such as the Sun, the antihyperon decay length is typically much shorter than its interaction length. In Earth, pions do not contribute significantly to the neutrino flux because $\pi^0$’s decay immediately to photons, and the interaction length of $\pi^\pm$’s is far smaller than their decay length so they eventually convert to $\pi^0$’s through charge exchange reactions; similarly, the interaction lengths of $K^{0}_L$’s and $K^\pm$’s are much longer than their decay lengths, so through charge exchange they essentially all convert to $K^0_S$’s before decaying. The branching fraction for production of $\nu_{e,\mu}$ and $\bar{\nu}_{e,\mu}$ from $K_S^0 \rightarrow \pi l^\pm \nu$ is $3.4 \times 10^{-4}$ for each, so $f_{\nu_i} \ge 2(3.4\times 10^{-4})$ for $\bar{H}$ annihilation in Earth. Since the Sun has a paucity of neutrons, any kaons in the annihilation products are typically $K^+$ and furthermore their charge exchange is suppressed by the absence of neutrons. The branching fraction for $K^+ \rightarrow \mu^+ \nu_\mu$ is 63% and the $\nu_\mu$ has 240 MeV if the kaon is at rest. If the final states of $\bar{H}$ annihilation typically contain kaons, then $f_\nu $ is $\mathcal{O}$(1). However if annihilation favors $\bar{\Xi}$ production, $f_\nu$ could be as low as $\approx 3 \cdot 10^{-4}$ for production of $\bar{\nu}_{e}$’s and $\bar{\nu}_\mu$’s above the charged current threshold. Thus the predicted neutrino signal from $\bar{H}$ annihilation in the Sun is far more uncertain than in Earth. Neutrinos from $\bar{X}$ annihilation can be detected by SuperK, with a background level and sensitivity which depends strongly on neutrino energy and flavor. Taking the captured $\bar{X}$ flux on Earth from Table I, assuming the neutrinos have energy in the range 20-80 MeV for which the background rate is at a minimum, and taking the effective cross section with which $\nu$’s from the kaon decays make observable interactions in SuperK to be $10^{-42} {\rm cm}^2$, (\[neutintsSK\]) leads to a predicted rate of excess events from annihilations in Earth of $ \Gamma_{\nu{ \rm SK}} \approx 2$/(kton yr) in the $\bar{H}$ scenario. This is to be compared to the observed event rate in this energy range $\approx 3$/(kton yr)[@SKloEneutrinos], showing that SuperK is potentially sensitive. If a detailed analysis of SuperK’s sensitivity were able to establish that the rate is lower than this prediction, it would imply either that the $H, \bar{H}$ model is wrong or that the annihilation cross section is so low that the equilibrium assumption is invalid, i.e., $\sigma^{\rm ann}_{\bar{H}N} {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}2 \times 10^{-48} {\rm cm}^2$. The analogous calculation for the Sun gives $ \Gamma_{\nu{ \rm SK}} \approx 130 f_\nu$/(kton yr) for energies in the sub-GeV atmospheric neutrino sample, for which the rate is $\approx 35$ events/(kton yr)[@SKatmneutrinos].[^1] Thus if $f_\nu$ were large enough, SuperK could provide evidence for the $H,\,\bar{H}$ scenario via energetic solar neutrinos. However the absence of a solar neutrino signal cannot be taken as excluding the $H,\,\bar{H}$ scenario, given the possibility that $f_\nu \le 10^{-3}$. Fortunately, there is a clean way to see that the DM cannot contain a sufficient density of $ \bar{H}$’s to account for the BAU. When an $\bar{X}$ annihilates, an energy $m_X + B_X m_N$ is released, almost all of which is converted to heat. Uranus provides a remarkable system for constraining such a possibility, because of its large size and extremely low level of heat production, $42 \pm 47 \, {\rm erg ~ cm}^{-2} s^{-1}$ [@uranusVoyager]. When annihilation is in equilibrium with capture as discussed above, the heat flux supplied by annihilation is $ f_{\rm cap}^U \Phi_{\bar{X}} (m_X + B_X m_N).$ For the $\bar{H}$, $f_{\rm cap}^U \approx 0.2$ as for Earth, so the heat flux generated in Uranus should be $470 \, {\rm erg ~ cm}^{-2}s^{-1}$, which definitively excludes the $H,\,\bar{H}$ scenario. However the $\bar{X_4}$ scenario is safe by many orders of magnitude, because $ f_{\rm cap}^U = 3 \cdot 10^{-5}$ with the “conservative" DM halo velocity distribution[@zf:window], giving a heat flux of $0.07 \, {\rm erg ~ cm}^{-2}s^{-1}$ or even less if equilibrium is not established. We now turn to the particle physics of a new, light fundamental particle with $B_X = 1$ and $m_X {\,\raisebox{-0.13cm}{$\stackrel{\textstyle<} {\textstyle\sim}$}\,}4.5$ GeV. Such a low mass suggests it is a fermion whose mass is protected by a chiral symmetry. Various dimension-6 interactions with quarks could appear in the low energy effective theory after high scale interactions, e.g., those responsible for the family structure of the Standard Model, have been integrated out. These include $$\label{Xbcd} \frac{g}{M^2} (\bar{X} b \, \bar{d^c} c - \bar{X} c \, \bar{d^c} b) + h.c.,$$ where the $b$ and $c$ fields are left-handed SU(2) doublets combined to form an SU(2) singlet and $d^c$ is the charge conjugate of the SU(2) singlet field $b_R$. The suppressed color and spin indices are those of the antisymmetric operator $\tilde{O}^{\dot{a}}$ given in equation (10) of ref. [@peskin79]. The hypercharge of the left-handed quarks is +1/3 and that of $d_R$ is -2/3, so the $X$ is a singlet under all standard model interactions and its only interaction with fields of the standard model are through operators such as (\[Xbcd\]). Dimension-6 operators involving only third generation quarks can be constructed; supplemented by $W$ exchange or penguins, they could also be relevant. Prior to freezeout, $\bar{X}$’s stay in equilibrium through scattering reactions like $$\label{dXbar} d + \bar{X} \leftrightarrow \bar{b}~\bar{c}.$$ The coupling $g$ in (\[Xbcd\]) is in general complex and a variety of diagrams involving all three generations and including both W exhange and penquins contribute to generating the effective interaction in (\[dXbar\]), so the conditions necessary for a sizeable CP-violating asymmetry between $\sigma_{X}^{ \rm ann} $ and $ \sigma_{\bar{X}}^{ \rm ann}$ are in place. An interaction such as (\[Xbcd\]) gives rise to $ \sigma_{\bar{X} d \rightarrow \bar{b} \bar{c}} \sim (g/M^2)^2 m_X T_{\rm f.o.}.$ Therefore $g/M^2 \approx 10^{-9} - 10^{-10}\, {\rm GeV}^{-2}$ is needed for freezeout to occur at the correct temperature. It is not possible to use this estimate of $(g/M^2)$ to obtain $\sigma^{\rm el}_{XN \{\bar{X} N\} }$ without understanding how the interaction (\[Xbcd\]) is generated, since it is not renormalizable. A naive guess $\sigma^{\rm el}_{XN \{\bar{X} N\} } \sim (g/M^2)^4 m_X^6 \,10 \, {\rm mb} \approx 10^{-40} \,{\rm to} \, 10^{-43} {\rm cm}^2$ is well below the present limit of $\approx 10^{-38} {\rm cm}^2$ for a 4 GeV particle, but the actual value depends on the high-scale physics and could be significantly larger or smaller. If (\[Xbcd\]) is the only coupling of $X$ to quarks, the effects of annihilation in Earth, Sun, Uranus and galactic center are unobservably small. This is because $m_X + m_N < m_B + m_D$, so an additional factor $\approx G_F^2 |V_{bc}|^2 {\rm GeV}^4$ is required below the chiral phase transition leading to $\sigma^{\rm ann}_{\bar{X}N} \approx 5 \times 10^{-58} {\rm cm}^2$. However there could be other interactions which are subdominant at freezeout but are kinematically allowed in the low temperature phase, e.g., (\[Xbcd\]) with $b \rightarrow d$, so a search for an unexpected neutrino or gamma-ray flux is worthwhile, even though null results would not exclude the model. To summarize, we have shown that the dark matter and baryon asymmetry puzzles may be related. We presented two concrete scenarios, in which the observed values of $\Omega_{DM}$ and $\Omega_b$ are explained. One of them entails a long-lived H dibaryon, but it is excluded by limits on heat production in Uranus. The other involves a new $\sim 4$ GeV particle with dimension-6 couplings to quarks. This scenario can arise naturally from beyond-the-standard-model particle physics, but it suggests that the DM particle may be elusive.\ [**Acknowledgements:** ]{} This research was supported in part by NSF-PHY-0101738. We have benefited from discussions with many colleagues; special thanks go to S. Mitra for bringing the strong limits on heating in Uranus to our attention, and to S. Nussinov for discussions about ways to exclude the H dibaryon scenario, in particular the idea of using white dwarf heating which although unsuccessful planted the idea we used here with Uranus. GRF also thanks the Departments of Physics and Astronomy of Princeton University for their support and hospitality during the course of this research. [10]{} A. Sakharov 5: 32,1967. D. B. Kaplan, , 68:741–743, 1992; S. Dodelson and L. Widrow, , D42:326–342, 1990; S. M. Barr et al. , B241:387–391, 1990; D. Oaknin and A. Zhitnitsky, hep-ph/0309086. D. N. Spergel et al. 148:175, 2003. G. R. Farrar and G. Zaharijas , in press (hep-ph/0308137); and hep-ph in preparation. E. W. Kolb and M. S. Turner. The Early Universe (1990). G. Zaharijas and G. R. Farrar ast-ph/0406531. G. R. Farrar and G. Zaharijas B559:223, 2003. Y. Fukuda B433:9, 1998. L. M. Krauss et al. , D33:2079, 1986. Y. Suzuki. , A718:83, 2003. I. Goldman and S. Nussinov hep-ph/0406259. J. C. Pearl et al. ,84:12 (1990). M. E Peskin , B88:128, 1979. [^1]: This estimate disagrees with that of Goldman and Nussinov (GN)[@GN], independently of the question of the value of $f_\nu$. GN use an $\bar{H}$ flux in the solar system which is eight times larger than our value in Table I from integrating the normal component of the halo velocity distribution, due to poor approximations and taking a factor-of-two larger value for the local DM density. We include a factor 0.35 for the loss of $\nu_\mu$’s due to oscillation, we account for the fact that only neutrons in SuperK are targets for CC events, and we avoid order-of-magnitude roundup. Note that the discussion of the particle physics of the $H$ in [@GN] applies to the case of an absolutely stable $H$, which we discussed but discarded in [@fz:nucbind; @fz:H].
{ "pile_set_name": "ArXiv" }
--- author: - 'R. Czyżykiewicz$^{a,b}$, P. Moskal$^a$, H.-H. Adam$^c$, A. Budzanowski$^d$, E. Czerwiński$^a$, D. Gil$^a$, D. Grzonka$^b$, M. Janusz$^{a}$, L. Jarczyk$^a$, B. Kamys$^a$, A. Khoukaz$^c$, K. Kilian$^b$, P. Klaja$^a$, B. Lorentz$^b$, J. Majewski$^{a,b}$, W. Oelert$^b$, C. Piskor-Ignatowicz$^a$, J. Przerwa$^a$, J. Ritman$^b$, H. Rohdjess$^e$, T. Rożek$^{b,f}$, R. Santo$^c$, T. Sefzick$^b$, M. Siemaszko$^f$, J. Smyrski$^a$, A. Täschner$^c$, K. Ulbrich$^e$, P. Winter$^{b\dagger}$, M. Wolke$^b$, P. Wüstner$^g$, Z. Zhang$^b$, W. Zipper$^f$' title: ' THE ANALYZING POWER FOR THE $\vec{p}p\to pp\eta$ REACTION AT Q=10 MeV ' --- [ $^a$Institute of Physics, Jagellonian University, 30-059 Cracow, Poland\ $^b$IKP, Forschungszentrum Jülich, 52425 Jülich, Germany\ $^c$IKP, Westfälische Wilhelms-Universität, 48149 Münster, Germany\ $^d$Institute of Nuclear Physics, 31-342 Cracow, Poland\ $^e$Institut für Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universität, 53115 Bonn, Germany\ $^f$Institute of Physics, University of Silesia, 40-007 Katowice, Poland\ $^g$ZEL, Forschungszentrum Jülich, 52425 Jülich, Germany\ $^\dagger$present address: Paul Scherrer Institute, 5232 Villigen, Switzerland\ ]{} Introduction {#sec:intr} ============ The close-to-threshold measurements of the total cross section for the $pp\to pp\eta$ reaction [@bergdolt; @chiavassa; @calen1; @calen2; @hibou; @smyrski], the measurements of the differential cross sections [@calen3; @tatischeff; @abdel], the high statistic investigations of the pp$\eta$ final state dynamics [@moskal; @hab] as well as the results of the first ever analyzing power experiment for the $\vec{p}p\to pp\eta$ reaction [@winter] are in agreement with the hypothesis that the main contribution to the production of the $\eta$ meson in the proton-proton collisions comes from the mesonic excitation of the S$_{11}$(1535) resonance and it’s further decay into the p$\eta$ pair [@wilkin; @hanhart], as depicted in Fig. \[S11\]. However, it is still not settled what are the relative contributions to the production mechanism originating from each meson exchange denoted in Fig. \[S11\]. A possible way for elucidation of this problem is a precise measurement of the polarization observables, as they are very sensitive to the model assumptions on the type of the meson being exchanged in order to excite the S$_{11}$ resonance. For example, from Fig. \[modele\], where the predictions of the analyzing power for the $\vec{p}p\to pp\eta$ reaction at the excess energy Q = 10 MeV are shown, it can be noticed that the analyzing power function strongly depends on whether we deal with pseudoscalar [@nakayama] (full line) or vector meson exchange model [@wilkin] (dotted line). ![ Predictions of the analyzing power for the $\vec{p}p\to pp\eta$ reaction at Q=10 MeV as a function of the center-of-mass (CM) polar angle of $\eta$ ($\theta_{CM}$). Full line are the predictions of the pseudoscalar meson exchange model [@nakayama], whereas the dotted line represents the results of the calculations, based on the vector meson exchange [@wilkin]. []{data-label="modele"}](S11.eps "fig:"){width="2.5in"}\ ![ Predictions of the analyzing power for the $\vec{p}p\to pp\eta$ reaction at Q=10 MeV as a function of the center-of-mass (CM) polar angle of $\eta$ ($\theta_{CM}$). Full line are the predictions of the pseudoscalar meson exchange model [@nakayama], whereas the dotted line represents the results of the calculations, based on the vector meson exchange [@wilkin]. []{data-label="modele"}](ay_10_gole.eps "fig:"){width="2.5in"}\ Therefore, in order to verify the validity of these two models, an experiment devoted to a determination of the analyzing power for the $\vec{p}p\to pp\eta$ reaction has been performed at the storage ring COSY in the Research Center Jülich in Germany. In this article we present preliminary results obtained at excess energy Q = 10 MeV, where the predictions of the models differ at most. Experimental method {#sec:experimental} ==================== In the measurements a polarized proton beam [@prasuhn; @stockhorst] with the momentum of p$_{beam}$ = 2010 MeV/c has been used. This beam momentum value corresponds to the excess energy $Q~=~10$ MeV. The beam of protons has been scattered on the $H_2$ clusters, which were produced by the hydrogen cluster target [@dombrowski]. The COSY-11 detection setup [@brauksiepe; @jurek] has been used for the detection of the reaction products. As the COSY-11 system is a one-arm detection setup, in order to register both types of events: scattered to the right and to the left side (with respect to the polarization plane) there is a need to perform the experiment with the separate cycles, in which the spin of the proton beam has to be flipped, as schematically shown in Fig. \[fig:vectors\]. In this figure we also define what is meant by the terms: “scattering to the right” and “to the left”. This definition strictly follows the Madison convention [@XXX]. ![Definition of the scattering to the right (a) and to the left side (b) with respect to the polarization plane. $\vec{P}$ is a polarization vector of the beam of protons, whereas $\vec{p}_{in}$ and $\vec{p}_{out}$ denote the momenta of incoming and outgoing particles, respectively. The polarization vector of Fig. \[fig:vectors\].a points out of the paper, whereas the polarization vector of Fig. \[fig:vectors\].b points into the paper. []{data-label="fig:vectors"}](vectors.eps "fig:"){width="4.5in"}\ In the following we define the polarization plane as a plane spanned by the beam momentum vector $\vec{p}_{beam}\equiv [0,0,p^z_{beam}]$ and a polarization vector $\vec{P}=[0,P,0]$. If the incident beam consists of spin $\frac{1}{2}$ particles with a transverse polarization, which was the case in our experiment, the formula for the differential cross section $\sigma(\theta,\phi)$ for a scattering into the solid angle around $\theta$ and $\phi$ angles reads: $$\sigma(\theta,\phi) = \sigma_0(\theta) (1+A_y(\theta)\vec{P} \cdot \hat{n}), \label{eq:cross}$$ where $\sigma_0(\theta)$ denotes the differential cross section for a scattering of an unpolarized beam, A$_y$($\theta$) is an analyzing power of the reaction, and $\hat{n}$ is a unit vector along $\vec{p}_{in} \times \vec{p}_{out}$. $\phi$ is an angle between $\vec{P}$ and $\hat{n}$, i.e. $\vec{P} \cdot \hat{n} = P \cos \phi \sin \theta$ and $\theta$ denotes the angle between $\vec{p}_{in}$ and $\vec{p}_{out}$. In the experiment we were restricted to the plane corresponding to $\phi=0^\mathrm{o}$ in case of spin “down” (or $\phi=180^\mathrm{o}$ for spin “up”), where the detection efficiency of the COSY-11 setup is the highest and it decreases drastically when going outside this particular scattering plane (see Fig. \[katfi\]). Therefore, in what follows we will consider the scattering in this plane solely. ![Monte-Carlo simulation of the distribution of the events as a function of the angle $\phi$. []{data-label="katfi"}](kat_fi_mc.ps){width="3.5in"} The total production rate to the right N$_R$ and to the left side N$_L$ can be expressed as follows: $$N_R(\theta) = N(\theta,\phi=\pi) = \sigma_R(\theta,\pi) E_R(\theta,\pi) \int{L_R dt_R} = \sigma_0(\theta) (1-A_y(\theta) P_R) E_R(\theta,\pi) \int{L_R dt_R}, \label{eq:crossR}$$ $$N_L(\theta)= N(\theta,\phi=0) = \sigma_L(\theta,0) E_L(\theta,0) \int{L_L dt_L} = \sigma_0(\theta) (1+A_y(\theta) P_L) E_L(\theta,0) \int{L_L dt_L}, \label{eq:crossL}$$ where the E$_R(\theta,\pi)\equiv$ E$_L(\theta,0)$ is the function of the efficiency of the COSY-11 detection system, L$_{R/L}$ stands for the luminosity for spin up/down, and t$_{R/L}$ denotes the time of the measurement with spin up/down, respectively. The polarization degrees P$_R$ for spin up and P$_L$ for spin down are equal within the 5% of accuracy, as has been shown in the previous measurements by means of the EDDA polarimeter [@altmeier]. Therefore for the further considerations we will assume that P$_L$=P$_R\equiv$ P. Dividing equation \[eq:crossL\] by \[eq:crossR\], and introducing a relative luminosity: $$L_{rel} \equiv \frac{\int{L_R dt_R}}{\int{L_L dt_L}}, \label{relative}$$ results in the following formula for the analyzing power: $$A_{y}(\cos\theta)=\frac{1}{P} \frac{L_{rel} N_L(\cos\theta) - N_R(\cos\theta)}{L_{rel} N_L(\cos\theta) + N_R(\cos\theta)}, \label{eq:analysing}$$ Therefore, the values of the relative luminosity L$_{rel}$, the spin-averaged polarization degree P, together with the production rates to the right N$_R$, and to the left side N$_L$ are the physical observables to be found in order to determine the analyzing power function A$_y$. In the following subsections the methods used to obtain these quantities will be presented. Calculations of the relative luminosity and polarization {#subsec:polarisation} --------------------------------------------------------- The experiment has been performed with the 300 s long cycles, for which the spin of the incident proton beam has been flipped from cycle to cycle. This method was intended to reduce the systematical uncertainties caused by the change of the beam parameters due to the variation of the target densities. The duration of the cycle has been set up in such a manner, that it was significantly shorter than the time scale (10 hours) for the substantial changes of the density of the target. For the determination of the relative luminosity a detection system consisting of two pairs of the scintillators placed in the polarization plane have been used [@czyzyk]. Due to the parity invariance, a cross section for any scattering into this plane does not depend on the polarization degree. Thus, the number of the coincidences $n$ between scintillators has been used as a measure of the absolute luminosity. Following the definition given in eq. \[relative\], we can express the relative luminosity as a ratio of the number of coincidences in the polarization plane during the cycles with spin “up” ($n_R$) and “down” ($n_L$): $$L_{rel} = \frac{n_R}{n_L}. \label{gowienko}$$ The relative luminosity determined using eq. \[gowienko\] was found to be L$_{rel}$=0.96468$\pm$ 0.00065. The detailed method of determination of the averaged polarization value has been described in [@czyzyk]. For the calculations the assymetry of the elastically scattered events in the plane perpendicular to the polarization plane has been found for different scattering angles in the center-of-mass system (CM). In order to determine the values of the analyzing powers for the beam momentum of p$_{beam}$ = 2010 MeV, the linear interpolations between available experimental results for the proton-proton elastic scattering at $p_{beam}~=~1995$ and 2025 MeV/c have been used. Two data sets of the analyzing powers for the $\vec{p}p\to pp$ reaction for the mentioned above beam momenta, as measured by the EDDA collaboration [@altmeier] at different CM scattering angles, are presented in Fig. \[edda\]. ![ Average polarization degree versus the time of the measurement. Open circles are the results obtained by means of the COSY polarimeter, whereas the full dots are the polarization values determined using the method described in [@czyzyk]. Data obtained using the independent methods are in line. []{data-label="polar"}](anl_power_edda.eps "fig:"){width="2.5in"}\ ![ Average polarization degree versus the time of the measurement. Open circles are the results obtained by means of the COSY polarimeter, whereas the full dots are the polarization values determined using the method described in [@czyzyk]. Data obtained using the independent methods are in line. []{data-label="polar"}](polaryzacja_caly_pomiar6.eps "fig:"){width="2.5in"}\ Fig. \[polar\] depicts the variation of the beam polarization during the time of the measurement. Data sets obtained using the COSY polarimeter [@bauer](open circles) are confronted with the results of the measurements with the COSY-11 polarimeter [@isabela; @kowina; @czyzyk](full circles). An agreement between both sets of data is visible in Fig. \[polar\]. The average value of the beam polarization for the whole period of the measurement was found to be P=0.680$\pm$ 0.010. Determination of the background free production rates for the $\vec{p}p\to pp\eta$ reaction {#subsec:rl} -------------------------------------------------------------------------------------------- A typical missing mass spectrum for the $\vec{p}p\to pp\eta$ reaction as measured using the COSY-11 detection setup is presented in Fig. \[missmass\]. Figure shows the spin-averaged missing mass spectrum, gathered during the whole time of the experiment. Over the wide multi-pionic background a clear peak originating from the $\eta$ meson production is visible. ![ Spin-averaged missing mass spectrum for the $\vec{p}p\to pp\eta$ reaction at the excess energy Q = 10 MeV, as measured by means of the COSY-11 detector setup. []{data-label="missmass"}](miss_mass_final.eps "fig:"){width="2.5in"}\ Optimizing the statistics and the expected shape of the analyzing power function, the range of the center-of-mass polar angles of the $\eta$ meson emission has been divided into four bins. To separate the actual production rates from the background, the reactions with multi-pionic production as well as the events with $\eta$ production have been simulated, using a program based on the GEANT3 [@geant] code. Generated events which fulfilled conditions equivalent to the experimental trigger have been analyzed in the same way as the experimental data. In such a way we have obtained the missing mass shapes of the background reactions ($pp\to pp2\pi, 3\pi, 4\pi$) as well as the shape of the signal ($pp\to pp\eta$). In order to perform credible simulations of the missing mass spectra the geometry of the drift chambers (used for the momentum reconstruction) as well as the position and geometrical parameters of the target have to be known precisely. The angle and the relative position of the COSY-11 drift chambers (2 parameters), a target position (3 parameters) and a relative shift between the beam and a target (1 parameter) have been optimized using the elastically scattered events and the method described in [@NIM; @hab]. The multi-dimensional fit of the simulated missing mass spectra to the corresponding experimental histograms has been performed using the MINUIT [@minuit] minimization package. The fit has been performed simultaneously to 8 histograms (see Fig. \[up\_and\_down\]), each with 60 points. The $\chi^2$ of the fit have been minimized as a function of six parameters: scaling factors of the generated background and signal reactions (altogether 4 parameters) and 2 parameters responsible for the spread and the absolute value of the beam momentum. The detailed process of the search for the best values of these parameters will be described elsewhere [@czyzyk-phd], here we would only like to report that the minimum value of $\chi^2$ was found to be 1.62, corresponding to the spread of the beam momentum equal to 0.2 MeV/c and the shift from the nominal beam momentum equal to 2.1 MeV/c. These values are in a good agreement with results of previous COSY-11 experiments [@NIM]. In Fig. \[up\_and\_down\] the missing masses for the individual $\theta_{CM}$ subranges, separately for spin up (upper panel) and down (lower panel), are shown. Full dots correspond to the experimental data, the dotted line depicts the shape of the background, whereas the solid line represents the best fit of the Monte-Carlo data to the experimental spectra. ![ Missing mass spectra for different ranges of $\cos\theta_{CM}$ for spin up (two upper panels) and spin down (two lower panels). Dots represent the experimental points. Dotted line shows the generated multi-pionic background. The full line is a best fit of the sum of the signal and background to the experimental data. []{data-label="up_and_down"}](missssing.ps){width="3.7in"} Results ======== Preliminary results of the analyzing power function A$_y(\cos\theta)$ are presented in Fig. \[fig-analysing\] as the full dots. The dotted line shows the predictions determined according to the vector meson exchange model [@wilkin], whereas the solid line refers to the pseudoscalar meson exchange model [@nakayama]. The errors are of the statistical nature only. ![ Preliminary analyzing power function for the $\vec{p}p\to pp\eta$ reaction at Q=10 MeV. Vertical bars denote the statistical error, whereas the horizontal bars stand for the ranges of averaging. []{data-label="fig-analysing"}](ay_10_krakow.eps "fig:"){width="2.5in"}\ Conclusions ============ The preliminary analyzing power results are bared with rather large statistical uncertainties, therefore a statement on the mechanism of the $\eta$ meson production relying on this set of data is currently unclear. As can be seen from Fig. \[fig-analysing\] the data point at $\cos \theta = 0.75$ tend to prefer the pseudoscalar meson exchange model [@nakayama], whereas the data point at $\cos \theta = -0.75$ is slightly more in line with the predictions of the vector meson exchange model. The results show rather small values of the analyzing power in the close-to-threshold region, which may be the indication of the $\eta$ production to the $s$-wave final state, solely. However, for this statement the exact partial wave analysis remains to be done. We acknowledge the support of the European Community-Research Infrastructure Activity under the FP6 “Structuring the European Research Area” programme (HadronPhysics, contract number RII3-CT-2004-506078), of the FFE grants (41266606 and 41266654) from the Research Centre J[ü]{}lich, of the DAAD Exchange Programme (PPP-Polen), of the Polish State Committe for Scientific Research (grant No. PB1060/P03/2004/26),\ and of the RII3/CT/2004/506078 - Hadron Physics-Activity -N4:EtaMesonNet. [99]{} P. Moskal: [*Hadronic interaction of $\eta$ and $\eta^{\prime}$ mesons with protons*]{}, Habilitation Dissertation No. 364, Jagellonian University Press (2004); e-Print Archive: hep-ph/0408162 ; erratum-ibid. [**B553**]{} (2003) 339 H. Stockhorst: [*Polarized Proton and Deuteron Beams at COSY*]{} in the [*Proceedings of the second symposium on threshold meson production in pp and pd interaction*]{}, ed. P. Moskal; Schriften des Forschungszentrum Jülich, Matter and Materials [**21**]{} (2004) pp. 106-121; e-Print Archive: physics/0411148 Madison Convention, in [*Polarisation Phenomena in Nuclear Reactions*]{}, University of Wisconsin Press, Madison (1971) pp. XXV R. Czy[ż]{}ykiewicz: [*COSY-11 as a polarimeter*]{} in the [*Proceedings of the second symposium on threshold meson production in pp and pd interaction*]{}, ed. P. Moskal; Schriften des Forschungszentrum Jülich, Matter and Materials [**21**]{} (2004) pp. 122-137. I.-A. Pellmann: Diploma Thesis (1999), Berichte des FZ-J[ü]{}lich, J[ü]{}l-3686. P. Kowina et al.: contribution to the [*Annual Report of IKP*]{} (2001) pp. 14, Berichte des FZ-J[ü]{}lich, J[ü]{}l-3978. GEANT — Detector Description and Simulation Tool, CERN Program Library Long Writeups W5013 (1994) MINUIT — Minimization Package, CERN Program Library Long Writeups D506 (1994) R. Czy[ż]{}ykiewicz: PhD Dissertation at the Jagellonian University, in preparation
{ "pile_set_name": "ArXiv" }
--- abstract: 'We calculate Curie temperatures and study the stability of ferromagnetism in diluted magnetic materials, taking as a model for the exchange between magnetic impurities a damped Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction and a short range term representing the effects of superexchange. To properly include effects of spin and thermal fluctuations as well as geometric disorder, we solve the effective Heisenberg Hamiltonian by means of a recently developed semi-analytical approach. This approach, “self-consistent local Random Phase Approximation (SC-LRPA)”, is explained. We show that previous mean-field treatments, which have been widely used in the literature, largely overestimate both the Curie temperatures and the stability of ferromagnetism as a function of carrier density. The discrepancy when compared to the current approach was that effects of frustration in RKKY oscillations had been strongly underestimated by such simple mean-field theories. We argue that the use, as is frequent, of a weakly-disordered RKKY exchange to model ferromagnetism in diluted III-V systems is inconsistent with the observation of ferromagnetism over a wide region of itinerant carrier densities. This may be puzzling when compared to the apparent success of calculations based on [*ab-initio*]{} estimates of the coupling; we propose a resolution to this issue by taking RKKY-like interactions between resonant states close to the Fermi level.' author: - 'Richard  Bouzerar$^{1}$, Georges  Bouzerar$^{2}$ and Timothy Ziman$^{3}$' title: '**Why RKKY exchange integrals are inappropriate to describe ferromagnetism in diluted magnetic semiconductors.**' --- Introduction ============ After the discovery by Ohno et al. [@Ohno] that the doping by a small amount of magnetic impurities in GaAs could lead to relatively high Curie temperature in III-V semiconductors[@Edmonds; @Potashnik], materials of this class have attracted considerable interest from both experimentalists and theoreticians. It is now accepted that the ferromagnetism is mediated by the “itinerant carriers” introduced after substitution of Ga, of nominal valence ${3+}$, by Mn$^{2+}$. Among theories of the ferromagnetism in these materials, the simplest and most commonly used approach considers that the exchange between localized magnetic impurities is of standard RKKY type [@Dietl; @DasSarma]. The exchange integrals are, to second order in perturbation theory, quadratic with respect to the local coupling of the itinerant carrier to the localized magnetic impurity spin $J_{pd}$. $$\begin{aligned} J_{ij}=-{J_{pd}^2 \over \pi}(\Im \ \chi({\bf R}_{ij}))\end{aligned}$$ where $\chi$ is the magnetic susceptibility. The density of carriers enters in the magnetism via the dependence of the susceptibility but in the simplest approach the effects of disorder (multiple scattering of the itinerant carrier due to the magnetic impurities) are not taken into account. Once the effective magnetic Heisenberg Hamiltonian has been determined, it has commonly been solved by treating fluctuations within mean field theory, and disorder within a virtual crystal approximation (VCA), giving an approximation we shall refer to as (MF-VCA)[@Dietl; @DasSarma; @Matsukura; @Sato; @Jungwirth] and is commonly referred to loosely as Zener Mean-Field theory. It has often been stated in the literature[@Dogma] that such an approach gives good account of the experimental situation in both III-V and II-VI doped semiconductors and that it can be relied on for quantitative prediction. This apparent consensus on the applicability of such a picture is, however, seriously in doubt, as we shall discuss. From an experimental point of view, it is observed that the Curie temperatures are very sensitive to the method of preparation. Indeed the Curie temperatures often vary greatly when measured on samples as they are grown, and after they have been annealed. In the Zener Mean-Field Theory, the Curie temperature for fixed concentration of impurities varies with a simple power law (T$_C \propto \rho^{\frac{1}{3}})$ on the density of itinerant carriers. In reality, other aspects are important: such as configurational disorder, thermal fluctuations, and detailed compensating mechanisms[@Wang; @Kirby]. Calculations that treat the effects of band structure realistically from first principles show that the RKKY description is inaccurate [@Sandraskii; @DDDasSarma] in that the strength of interactions do not simply depend on distance, but also on lattice direction. Furthermore simple one-band model calculations of the exchange integrals treating the disorder by a Coherent Potential Approximation (CPA) has shown that the RKKY behaviour is in fact restricted to very small value of the ratio $J_{pd}/W$ [@Bouzerar1] where $W$ is the bandwidth of the carrier band. To higher order the oscillatory behaviour is strongly reduced. Calculations which, in addition, treat the effect of disorder on the itinerant carriers within CPA, show clearly that the exchange integrals (i) do not oscillate and (ii) are exponentially damped with distance[@Bouzerar3] as in the model calculations[@Bouzerar-unpublished]. A different approach was taken by Brey et al[@Brey] who took a [**k . p**]{} description of the band structure with a [*non-local*]{} J$_{pd}$ coupling treated perturbatively. In this case the non-locality of the J$_{pd}$ coupling lead to suppression of oscillations. Another attempt[@Timm] to use a realistic (Slater-Koster) band structure with non-local couplings between p- and d-orbitals treated perturbatively, lead, however, to very different magnetic couplings which oscillated strongly. Both calculations included spin-orbit couplings but differed in their conclusions as to its importance. Recent work[@Zhou] including Monte-Carlo simulations concluded that the effects of spin-orbit interactions on the Curie temperature are weak, unless the anisotropy induced is very large which seems unlikely for III-V semiconductors. Fiete et al[@Fiete] found effects of non-collinearity from spin-orbit coupling but they were significant at parameters corresponding to very low concentrations of dilute Ga(Mn)As. Recent theoretical studies have shown that if the exchange integrals includes both the effects of disorder and realistic band structure, MF-VCA treatment of the Heisenberg model leads to larger Curie temperature than experimentally observed[@Josef]. In contrast, it was shown that using a self-consistent local Random Phase Approximation (SC-LRPA) it is possible to attain quantitatively accurate estimates of ferromagnetism in diluted magnetic semiconductors for annealed [@Bouzerar2] and partially annealed samples [@Bouzerar4]. This method will be made explicit in the next section, but the essential improvement over MF-VCA is that both disorder and transverse fluctuations are properly treated. For the disorder this is by avoiding any effective medium approximation and, instead, using sampling over randomly generated geometries. The tranverse fluctuations are included by preserving the rotational spin symmetry in the decoupling. The success of this semi-analytical approach is fully supported in the limit of large spin where classical Monte Carlo calculations can be performed[@mc1; @mc2] For finite spin we can compare only for the case of regular lattices, but by comparison to series expansions the Callen approach is known to give critical temperatures accurate to a few per cent[@Callen] even for couplings of short range. Note that, compared to Monte-Carlo simulations, the SC-LRPA has several advantages: it provides a direct expression for the Curie temperature and the computing cost is extremely low, allowing systematic examination of the space of couplings, and effects of cut-off and so forth. In addition it can include quantum fluctuations for finite spin S for which Monte Carlo methods encounter problems of sign. The disadvantage of using “realistic” values of the exchange integrals taken from first principle calculations is that they are strongly material dependent, and do not provide a simple understanding of the parameters which control the exchange integrals. Thus, attempts to find simplified models where the exchange integrals between magnetic impurities would depend on a few physical parameters are still attractive and should be pursued. The purpose of this paper is then to return to the simple RKKY approach whose validity has been argued on the basis of over-simplified calculation of the thermodynamics, but now to study the the influence of disorder and transverse fluctuations using a more reliable calculation. Thus we shall examine the stability of ferromagnetism and Curie temperature for exchange integrals in the effective Heisenberg model of RKKY type. We will also, in the following, analyze both the effect of including a nearest-neighbour antiferromagnetic (AF) superexchange and damping of the exchange coupling. The damping mimics in a simple manner the effect of multiple scattering of the itinerant carriers over the magnetic impurities. As we have mentioned, while this corresponds to the [*average*]{} exchange in an approximation of weak coupling and weak disorder, it is not a completely general form for strong coupling. The effective Heisenberg Hamiltonian reads, $$\begin{aligned} H &=& {-} \sum_{ij} {J_{ij} \bf{\Large S}_{i} \cdot \bf{\Large S}_{j}}\end{aligned}$$ where $S_i$ are quantum spins randomly distributed on a lattice, we denote x the density of magnetic impurities. For comparison with Ga(Mn)As we take the lattice to be face centered cubic (fcc) corresponding to simple substitution of Mn on Ga sites. $J_{ij}$ is the exchange interaction coupling between two impurities located at site i and j, it reads $$\begin{aligned} J_{\rm ij} &=& J_{0} \exp({-r \over r_{0}}) \frac{({sin(2k_{F}r)-2k_{F}rcos(2k_{F}r)})} {(r/a)^{4}} + J_{ij}^{AF}\end{aligned}$$ where $r=R_{i}-R_{j}$ is the distance between two impurities labeled i and j. The first term is the damped RKKY exchange coupling, $r_0$ is the damping length. The second term is the direct antiferromagnetic superexchange interaction. For simplicity we restrict ourselves to short range SE term, $J_{ij}^{AF}=J_{AF}$ if i and j are nearest neighbours (NN) and vanishes otherwise. We consider the diluted regime for magnetic impurities and itinerant carriers; thus we assume a spherical Fermi surface and $k_{\rm F} = (3 \pi ^2 n_{c})^{1\over 3}$ where $n_{c}$ is the density of carriers. Note that we shall restrict ourselves to the simplest RKKY form corresponding to the asymptotic form of a single parabolic band of carriers. Our model has four parameters : $x$,$n_{c}$,$r_0$, and $J_{AF}/J_{0}$. In the following calculations they are treated as being independent, although of course physically if we vary the doping $x$ to model a series of compounds of different doping, the other parameters would be functions of $x$. A derived parameter useful for comparison with experiment is the doping concentration [*per*]{} doping impurity $\gamma=n_{c}/x$. In the simplest model of doping this would be 1, in practice it is usually much less and varies with sample history. As we treat the disorder fully and thermal fluctuations accurately, we will show that previous treatments of the same Hamiltonian (Equation 2) by MF-VCA[@DasSarma] lead to wrong conclusions. Furthermore by understanding this simple model, we shall show that we can demonstrate the importance of qualitative aspects. In particular, the presence or absence of oscillations will be seen to be crucial to stabilize ferromagnetism at finite temperatures. In MF-VCA this was obscured by the simplified treatment. This may be significant in more complicated forms of model band structures, e.g. 6 band Kohn-Luttinger forms [@Dietl97; @Dietl; @Abolfath; @Brey; @Timm]. In addition, in different [*ab-initio*]{} treatments, for example local-density-approximation (LDA) or extensions such as LDA+U approaches, there are changes in the nature of the states close to the Fermi surface[@DDDasSarma; @Shick] which will affect long-range oscillations of the exchange couplings and thus the region of stability of ferromagnetism. One aspect we do [*not*]{} consider is the role of anisotropic couplings in spin-space coming from spin-orbit couplings. As we have noted above, there is some debate[@Timm; @Brey; @Zarand; @Fiete] as to whether these make significant contributions to the exchange. We argue that for the Curie temperature this is not of primary importance especially as any spin anisotropy is random in direction[@Fiete]. Of course spin-orbit effects must be considered for calculations of magnetic anisotropies. We will argue that for the Curie temperature, which is the subject of the present paper, what is much more significant is the bias in the sub-asymptotic form of the isotropic RKKY-like couplings, coming from resonant effects. This will be explained in Section IV. The importance of such effects, rather than anisotropies, is supported by the success[@Bouzerar2; @Bouzerar4] of the non-relativistic [*ab-initio* ]{} calculations which include the effects of hybridization and resonance. Method ====== Let us now summarize the main steps of the approach we use. We define the following retarded Green’s function $G_{ij}(t)$ for localized spins at sites $i$ and $j$, $$\begin{aligned} G_{ij}(t)&=& -i\theta (t) < [S_{i}^{+}(t) ; S_{j}^{-}(0)] >\end{aligned}$$ The important point is that we shall decouple in real space, and within a Random Phase Approximation, the equation of motion for the frequency-transformed Green’s functions $G_{ij}(\omega)$: $$\begin{aligned} (\omega -h^{eff}_{i})G_{ij}(\omega) &=&2 \langle S_{i}^{z} \rangle \delta_{ij}-\left[\langle S_{i}^{z} \rangle \sum_lJ_{il}G_{lj}(\omega)\right]\end{aligned}$$ where $\langle S_{i}^{z} \rangle$ is the local magnetization at site i. For a given configuration of impurities and temperature, $\langle S_{i}^{z} \rangle$ should be determined self-consistently at each impurity site. $h^{eff}_{i}$ is the local effective field at site i, $$\begin{aligned} h^{eff}_{i} = \sum_lJ_{il} \langle S_{l}^{z} \rangle\end{aligned}$$ To determine $\langle S_{i}^{z} \rangle$ [*self-consistently*]{} we use the Callen expression: [@Callen], $$\begin{aligned} \langle S_{i}^{z} \rangle=\frac{(S-\Phi_{i})(1+\Phi_{i})^{2S+1} + (S+1+\Phi_{i}) \Phi_{i}^{ 2S+1}} {(1+\Phi_{i})^{2S+1}- \Phi_{i}^{ 2S+1}} \label{calleneq}\end{aligned}$$ where $\Phi_{i}$ is a local effective boson occupation number. $$\begin{aligned} \Phi_{i}= -\frac{1}{2 \pi}\frac{1} {\langle S_{i}^{z} \rangle}\int_{-\infty}^{+\infty} \frac{ImG_{ii}(\omega)}{\exp(\omega/kT)-1} d\omega \end{aligned}$$ At each temperature, (5),(7) and (8) form a closed set of equations. These may be solved to determine the temperature dependence of the local magnetization at each site and dynamical properties. To determine the critical temperature T$_{C}$, we take the limit of vanishing $ \langle S_{i}^{z} \rangle \rightarrow 0$ in the previous set of equations which leads to $$\begin{aligned} \langle S_{i}^{z} \rangle=\frac{1}{3}S(S+1)\frac{1}{\phi_i} \end{aligned}$$ We obtain, $$\begin{aligned} { k_{B} T_C} &=& {1\over 3 N_{imp}} S(S+1)\sum_{i} {1\over F_{i}} \label{eqtcrpa}\end{aligned}$$ The local quantity $F_{i}$ is $$\begin{aligned} F_i \equiv \int^{+\infty}_{-\infty}{A_{ii}(E) \over {E}} dE\end{aligned}$$ where the reduced variable $E=\frac{\omega}{m}$, m is the averaged magnetization over impurity sites, $$\begin{aligned} \lambda_{i} = lim_{T \rightarrow T_C} \frac{ \langle S_{i}^{z} \rangle}{m}\end{aligned}$$ The local spectral function $A_{ii}(E)$ is, $$\begin{aligned} A_{ii}(E) = -{1\over 2\pi} Im({G_{ii}(E)\over{\lambda_i}})\end{aligned}$$ We note that by the above argument, the dependence of T$_{C}$ on quantum spin $S$ is entirely in the factor $S(S+1)$ in equation (\[eqtcrpa\]). For the magnetization in the ordered phase the dependence is more complex as equation (\[calleneq\]) must be used at each stage of iteration. In the following, to evaluate the effect of both disorder and thermal fluctuations beyond mean field theory we will compare our self-consistent local-RPA expression of the Curie temperature T$_C$ to the often-quoted MF-VCA expression. $$\begin{aligned} {\rm{T}_C}^{MF-VCA} &=& {S(S+1)\over {3}} x \sum_{i} {N_{i} J(r_{i})} \label{eqtcmfvca}\end{aligned}$$ $N_{i}$ (resp. $r_{i}$) is the number (resp. distance) of the $i^{th}$ nearest neighbour (summation over all the host crystal sites). We note that this expression is equivalent to equation 4 of ref[@DasSarma], and is the basis of the results of that paper. While in that paper sampling over different configurations of disorder was performed, the mean-field treatment of thermal fluctuations effectively replaced the complex geometry of random substitution by an homogeneous crystal lattice. In contrast, the SC-LRPA simultaneously treats the effects of random geometry explicitly and includes thermal fluctuations beyond molecular field theory. In this case the disordered medium can no longer be reduced to an effective crystal. In the following we will be able to see [*quantitatively*]{} how inaccurate such mean-field lattice approximations may be. Numerical results ================= Results for Self-Consistent local RPA for the Curie Temperature as a function of carrier density ------------------------------------------------------------------------------------------------ In this section we present self-consistent local-RPA calculations for the RKKY interaction for parameters, in particular concentration of interest for III-V semiconductors. We shall compare to the results of MF-VCA treatment and attempt to understand the differences when they appear. We first consider the RKKY interaction in the absence of AF superexchange term, i.e. we fix $J_{AF}=0$, and assume strong damping, $r_{0}=2$, in units of the cubic lattice parameter. The effects of varying these two parameters will be considered in the following sections. In Figure \[curiegamma\] we have plotted the Curie temperature for fixed concentration $x=5 \%$ of Mn impurities as a function of $\gamma=n_{c}/x$ for both the self-consistent local-RPA and MF-VCA treatment. Note that for the SC-LRPA an average is made over approximately 100 configurations of disorder for each set of parameters. For each configuration the magnetic impurities location have been randomly generated on a LxLxL fcc lattice, where $L \ge 16$. In order to show that finite size effects are negligible we have plotted in the same figure the calculations corresponding respectively to 744 and 1987 magnetic impurities distributed randomly over the fcc lattice. By comparing results from the different sizes it is seen that finite size effects are negligible for the sizes considered here. It is seen that MF-VCA systematically overestimates the Curie temperature, and agreement between MF-VCA and SC-LRPA is only observed for very low carrier densities. Whilst the MF-VCA curve is monotonically increasing, the SC-LRPA curve exhibits a maximum at relatively low density of carriers $n_{c} \approx 0.1 x$ and ferromagnetism disappears for $\gamma \ge 0.25$. Thus the better treatment of disorder and thermal fluctuations leads to a narrow region of stability for ferromagnetism. What is the reason for this? In the RKKY form of the exchange integrals for sufficiently large distances the exchange integrals become antiferromagnetic, leading to frustration of the ferromagnetism. The effects of this frustration are strongly underestimated within a VCA mean field treatment This is illustrated in Fig.  \[distribution\] , where we plot the distribution $P(h^{eff})$ of the local fields at $T=0~K$. For a given configuration of disorder the local field at site i is $$\begin{aligned} h_{i}^{eff} = \sum_{l} J_{il} \langle S^{z}_{l} \rangle\end{aligned}$$ In Figure \[distribution\] the distributions are plotted for $\gamma = 0.025 , 0.1 , 0.2, 0.3$, corresponding to (a) , (b) , (c) and (d) respectively, in Figure \[curiegamma\]. In cases (a) and (b) we observe that the distribution of local fields is completely ferromagnetic: there is no antiferromagnetic part $P(h)=0$ for $h\le 0$. The distribution is very narrow in the case (a), this explains why in Figure \[curiegamma\] MF-VCA and SC-LRPA are very close. The distribution broadens asymmetrically in case (b) and the average value $\langle h^{\rm eff} \rangle = \int_{-\infty}^{+\infty} hP(h)$ increases, leading to higher T$_C$ in MF-VCA than in SC-LRPA. For higher density of carriers,(c) and (d), we observe a negative tail in the local field distribution. Note that, although this tail is very small, its effect on the Curie temperature is dramatic. Figures \[curiegamma\] and  \[distribution\] show clearly that the ferromagnetism in diluted systems is very sensitive to frustration effects coming from the exchange couplings at large distances. In a VCA the effect of the tail is lost in the average over different sites of the lattice. In the SC-LRPA we see clearly an instability to bulk uniaxial ferromagnetism. It may seem surprising that a relatively small proportion of sites with local negative molecular fields can suppress the ferromagnetism; however by analogy to the response of a ferromagnet to random [*external fields*]{} we can make an Imry-Ma argument [@ImryMa] that in three dimensions the rotational symmetry allows for complete break-up of ferromagnetism even for a small proportion of random fields. Thus it seems likely that the resultant state is either paramagnetic or of spin-glass type. In Figure \[exchangereal\] we mark the variation of the Heisenberg exchanges (normalized to omit the $1\over {r^3}$ decrease) as a function of distance for the same parameters $\gamma = 0.025, 0.1, 0.2, 0.3$, corresponding to (a) , (b) , (c) and (d) respectively, in Figure \[curiegamma\]. Unlike the distribution of local fields, this gives no obvious clue as to the instability of ferromagnetism. This is perhaps surprising if we remark that the average distance between magnetic impurities at this concentration is only slightly larger (r=1.07 in the units of Fig. \[exchangereal\]) than the lattice parameter of the fcc lattice. Effect of the exponential cut-off. ---------------------------------- We now discuss the effect of the exponential cut-off on the domain of ferromagnetic stability. The use of such a cut-off dates back to deGennes [@deGennes] who calculated the average coupling in weakly disordered systems. There has been much discussion of whether the difference between this average coupling and the undamped [*typical*]{} couplings that may determine the characteristic scale for spin-glass behaviour in the very dilute limit but for ferromagnetic transitions this has been assumed, without absolute proof to be appropriate. In Figure \[curiedamping\] we have plotted the Curie temperature as a function of the carrier density for different values of $r_{0}$:$r_{0}$= 2; 5; 10; $\infty$. Note that $r_{0}=\infty$ corresponds to the pure undamped RKKY case. We defer the question of the influence of superexchange $J_{AF}=0$ and consider a fixed density of magnetic impurities $x=5 \%$. The SC-LRPA shows that in the undamped RKKY limit the region of stability is extremely narrow. We observe that the ferromagnetic stability domain increases significantly when $r_{0}$ is relatively small, ie of the order of the average distance between impurities. For instance for $r_{0}$=2, the region of stability of ferromagnetism is five times broader than in the pure undamped case. This illustrates the strong influence of the oscillating tail on ferromagnetism. This figure also shows clearly that the undamped RKKY exchange integrals often used in model calculations can not explain the ferromagnetism often observed in III-V diluted materials for the carrier densities $n_c \approx x$ or $\gamma \approx 1$ for annealed samples which exhibit the highest T$_{C}$. Even for strong disorder ( $r_{0}$=2), ferromagnetism is not possible for $n_c \ge 0.25 x$. From this figure and Figure \[curiegamma\] , we conclude that the use of RKKY exchange integrals often used in simplified theories is questionable and its often-cited “success” is due to the unrealistic MF-VCA approximation which overestimates the Curie temperature and misses the instability seen here. The difference with the agreement with first principle calculations is that there, as in model calculations with the disorder treated in CPA[@Bouzerar-unpublished], not only are the exchange integrals are strongly damped but, in addition, the anti-ferromagnetic contributions are almost completely suppressed[@Josef]. Influence of the superexchange. ------------------------------- Let us now turn to the influence on ferromagnetism of a short-range superexchange contribution. This is important for understanding materials as it is essentially independent of the longer range RKKY exchange and sensitive to local effects. In mean field theories a very large ferromagnetic coupling at small distances has been used to argue, for example in doped Ga(Mn)N, for potentially high Curie temperatures. For simplicity, and because superexchange are short range, we consider the case where superexchange modifies nearest-neighbour magnetic impurities only. In Figure \[curiesuperex\] we have plotted the dependence of the SC-LRPA Curie temperature as a function of $J_{AF}$ for different carrier density and fixed magnetic impurity concentration and $r_{0}$=2. First, in contrast to the mean-field treatment (see inset) , we observe that in the region dominated by the RKKY term (nearest-neighbour ferromagnetic) that the Curie temperature is [*insensitive*]{} to $J_{AF}$. This can be understood as following: since the system is diluted, we might expect that T$_C$ should be controlled mainly by exchange integrals for typical distances between magnetic impurities. In fact, as we have discussed, it is not sufficient to consider only these distances: the long-range tail is also important. Our local approximation builds in all these features. In contrast the MF-VCA, in which the Curie temperature is linear in all couplings, could not capture these geometric effects. In our calculations, when the antiferromagnetic superexchange becomes dominant the Curie temperature vanishes abruptly. The reason for this instability is that the system becomes frustrated, but now from the short-range exchange, not the long range RKKY oscillations controlling the disappearance observed in Figures \[curiegamma\] and \[curiedamping\]. Variation of T$_C$ with concentration ------------------------------------- Let us now discuss the influence of the impurity concentration on the Curie temperature. For simplicity, we consider only the case $J_{AF}=0$. From the previous section, the results change little if we add a term of short-range superexchange, provided its strength is insufficient to give the instability of Figure \[curiesuperex\]. In Figure \[curieconc\] we have plotted T$_{C}$ as a function of x for different values of the carrier density per impurity $n_c = \gamma x$. For MF-VCA calculations we observe that T$_{C}$ increases monotonically with the density of both magnetic impurities and itinerant carriers. The SC-LRPA, however, predicts much smaller Curie temperatures than the MF-VCA results and that T$_C$ systematically [*decreases*]{} with increasing the carrier density $\gamma$. In contrast to MF-VCA, we also observe that as we increase $\gamma$, a maximum in T$_{C}$ appears. The location of this maximum is shifted to lower impurity concentration as the carrier density per magnetic impurity increases. For sufficiently high densities ($\gamma \ge 0.25$) of carriers we observe complete suppression of ferromagnetism. We have already discussed previously the reason for this: the frustrating effect of the RKKY “tail”. Thus again, we see that better treatment of disorder and fluctuations, give very different trends from the predictions of simple effective medium theories. The immediate question raised is then, how can we understand, from a simple model point of view, the observed ferromagnetism? For example that observed at carrier densities where from Figure  \[curieconc\] we now predict $T_C=0$ and we argue that that predictions from MF-VCA were invalid. This will be the subject of the next section. The solution: generalization of RKKY to treat resonances ======================================================== We have seen that simple RKKY forms of magnetic interaction leads to a region for stable ferromagnetic order that is very narrow. How can the experimental results be reconciled with this? One approach, which seems to be successful for the III-V is to abandon a simple model approach and use [*ab-initio*]{} results for the effective Hamiltonian. This works, as mentioned in the introduction, as the couplings do not display the changes of sign characteristic of the RKKY interaction. It is interesting to look more carefully at the physical reason behind this. While the non-spherical geometry of the Fermi surface of the host material affects the oscillations in exchange coupling it will not suppress the changes of sign. The significant point seems to be, in for example Ga(Mn)As, that the states at the top of the valence band are strongly hybridized with the impurity states. This gives rise to resonant peaks in the density of states close to the Fermi level. In consequence the states contributing to magnetic exchange must be modified by the hybridization between impurity and host bands. In this case the appropriate model is that of interacting resonances[@Caroli; @AlexanderAnderson]. As shown by Caroli and others [@Caroli; @Price; @Levy; @ZhangLevy], the simplest way to include resonant effects in order to generalize the regular RKKY expression, is to replace the term $J_{pd}$ by a frequency-dependent element ${{| V_{pd} |}^2} \over{(\omega-\omega{+})}$. In perturbation theory, the interactions between two resonant impurities can be separated into two contributions, the first an “RKKY-like” term generated by particle-hole exchange between the two spins. While asymptotically the standard RKKY interaction is recovered, there is an important “pre-asymptotic” regime where subleading corrections to the exchange change the qualitative form of the interaction. In particular in the first “RKKY-like” contribution there is a ferromagnetic “bias”, ie in the pre-asymptotic region, the interaction oscillates with distance with period $2k_f$ around a smoothly decaying value and does not change sign. The second contribution is the antiferromagnetic superexchange term coming from particle-particle excitations, ie with intermediate states where the moments are unoccupied. For a perfect gas the superexchange term compensates the ferromagnetic bias of the RKKY oscillation giving a a total exchange term that oscillates in sign[@Levy]. With the effects of disorder and interaction the compensation need not occur[@Levy] and this may provide a qualitative explanation of the form of the exchange interactions from [*ab-initio*]{} approaches. Because the [*ab-initio*]{} calculations are non-perturbative, they do not allow an easy separation into RKKY and superexchange contributions. As the [*ab-initio*]{} calculations include interactions and, to some extent, the disorder, the similarity between the forms seen there ( see for example Fig. 3 of Ref. [@Josef]) and the perturbative RKKY-like contribution to interactions between resonances supports the idea that it makes sense to exclude the superexchange contribution other than the short-range term. Thus an appropriate model Hamiltonian may be to take the perturbative “RKKY-like” interaction between resonances. While this will be more fully investigated in future work, we illustrate how this can restore ferromagnetism in a range where, as we have noted above, standard RKKY interactions fails to predict ferromagnetism. In Figure \[Fig.resonant\] we show results of calculation taking interactions which include the resonant RKKY-like contribution [@ZhangLevy; @Levy]. The resonances are defined by a $k$-vector $k_i$, i.e. $E_{imp}=\hbar{\frac {k_i^2} {2m}} $ which can be varied continuously relative to a spherical Fermi surface of fixed Fermi Energy $E_{f}=\hbar{\frac {k_f^2}{2m}}$. Perturbative calculations between narrow resonant levels interacting by free electrons give an RKKY-like contribution to the exchange[@ZhangLevy; @Levy], in momentum space, $$\begin{aligned} J_{RKKY}[q] = \frac{J_0}{4\pi q} \left[ \frac{\log | (\frac{-2\, {k_f} + q}{2\,{k_f} + q} )|}{{k_i}\, \left( 2\,{k_i} + q \right) } -\frac{ \log | (\frac{-2\, {k_f} + q}{2\,{k_f} + q} )|}{{k_i}\, \left( -2\,{k_i} + q \right) } - \frac{\log | (\frac{2\, {k_f} + q}{-2\,{k_f} + q} )|}{{{k_f}}^2 - {{k_i}}^2} \right. \nonumber \\ \left. - \frac{ \log | (\frac{{k_f} - {k_i} + q}{{k_f} - {k_i}}) | }{{k_i}\, \left( -2\,{k_i} + q \right) } + \frac{\log | (\frac{{k_f} + {k_i} + q}{{k_f} + {k_i}})|}{{k_i}\, \left( 2\,{k_i} + q \right) } + \frac{\log | (\frac{-{{k_i }}^2 + {\left( {k_f} + q \right) }^2} {{{k_f}}^2 - {{k_i}}^2})|} {{{k_f}}^2 - {{k_i}}^2} \right]\end{aligned}$$ Interactions in real space are obtained by Fourier transform of this analytical expression and an exponential damping term to make clear comparison with the results for standard RKKY (see Figure \[Fig.resonant\], inset, for the couplings as a function of distance). We chose the value of doping $\gamma=0.3$ and the same damping $r_0=2$, where after Figure \[curiegamma\] (see point ($d$) of that figure)there was no ferromagnetism for well developed moments. It is seen that for resonances well below the Fermi level there is no ferromagnetism, but as the resonance level increases ferromagnetism appears, and the Curie temperature rises rapidly as the impurity approaches the Fermi level. This is clearly a simplified calculation, and the divergence for $k_i\rightarrow k_f$ should be suppressed if superexchange is fully included. In addition we assume a moment which is fixed while this will, in fact, decrease as the resonance approaches the Fermi level. Nevertheless it illustrates the point that it is the resonant nature of the exchange that can resolve the apparent contradiction of observation, both from experiment and calculation based on [*ab-initio*]{} approach, of ferromagnetism at doping levels approaching large values of $\gamma$. We note that in the sense of the underlying model to take, this agrees with previous work emphasizing the influence of proximity to the Fermi level of the resonant states on Curie temperature[@Inoue; @Bouzerar1; @Kikoin]. Our calculations of the ferromagnetism are different, however, and, we argue, more precise. Conclusions =========== To conclude, we have studied the effects of both transverse fluctuations, and disorder, on ferromagnetism for diluted Heisenberg models assuming an RKKY type for the exchange integrals. We have shown that previous MF-VCA treatments are inappropriate to study ferromagnetism in diluted magnetic systems leading to strong over-estimates of Curie Temperatures. The apparent success of theories starting from RKKY couplings, which are often cited in the literature as being “qualitatively correct”, is in fact due to the oversimplified approximations (MF-VCA) used to treat the effective Heisenberg model. We have seen from Figure \[curiegamma\] that there is agreement only in the limit of extremely small carrier density. The long-range tail of the RKKY interaction destabilizes ferromagnetism over all but very narrow ranges of parameter values. Even damping this tail is insufficient to explain ferromagnetism at the high doping densities of the materials showing the highest Curie temperature. For experimentally interesting densities, we have seen that the randomness is essential, as seen clearly if Fig. \[distribution\]. Thus including the random geometry and transverse spin fluctuations, as we do, is much more significant than adding corrections to the continuum version of mean-field theory coming from the lattice version[@DasSarma]. It is appealing to have a simple phenomenological picture of ferromagnetism of diluted magnetic semiconductors rather than having to rely on [*ab-initio*]{} calculations of exchange which must be performed material by material and where the underlying physics may be obscured by the fact that several aspects (hybridization, band structure, correlations, disorder...) are included but it is not easy to separate their effects individually. While we have shown that a picture of magnetic moments interacting with standard RKKY interactions is [*not*]{} compatible with experiments, we argue that the generalization to include the [*resonant*]{} nature of the exchange may be the correct phenomenology. H. Ohno, Science [**281**]{},951 (1998). K. W. Edmonds, P. Boguslawski,K.Y. Wang, R.P. Campion, S. N. Novikov, N.R.S. Farley, B.L. Gallagher, C.T. Foxon, M. Sawicki, T. Dietl, M. Buongiorno Nardelli and J. Bernholc, Phys. Rev. Lett. [**92**]{}, 037201 (2004). S.J. Potashnik, K. C. Ku, S. C. Chun,J. J. Berry,N. Samarth,and P. Schiffer, Appl. Phys. Lett. [**79**]{}, 1495 (2001). T. Dietl, H. Ohno, and F. Matsukura, Phys. Rev. B [**63**]{}, 195205 (2001). D.J. Priour, E.H. Hwang and S. Das Sarma, Phys Rev. Lett. [**92**]{}, 117201 (2004). F. Matsukura, H.Ohno, A. Shen, and Y. Sugawara, Phys. Rev. B [**57**]{}, R2037 (1998). K. Sato, P. H. Dederichs and H. Katayama-Yoshida, Europhys Lett., [**61**]{}, 403 (2003). T. Jungwirth, W. A. Atkinson, B. H. Lee, and A. H. MacDonald Phys. Rev. B [**59**]{}, 9818 (1999). See for example References [@Dietl; @Jungwirth] and references therein. K.Y. Wang,K. W. Edmonds, R.P. Campion, B.L. Gallagher,N.R.S. Farley, C.T. Foxon, M. Sawicki, P. Boguslawski, T. Dietl Journal of Applied Physics [**95**]{}, 6512 (2004). B.J. Kirby, J.A.Borchers, , J.J. Rhyne, S.G.E. teVelthuis, A. Hoffmann, K.V. O’Donovan, T. Wojtowicz, X. Liu, W.L. Lim and J.K. Furdyna, Phys. Rev. B [**69** ]{}, 081307(R) (2004). L.M. Sandratskii and P. Bruno, Phys. Rev. B [**66**]{}, 134435 (2002). P. Mahadevan, A. Zunger and D.D. Sarma, Phys Rev. Lett. [**93**]{}, 177201 (2004). G. Bouzerar, J. Kudrnovsky and P. Bruno, Phys. Rev. B [**68**]{}, 205311 (2003). G. Bouzerar, J. Kudrnovsky, L. Bergqvist and P. Bruno, Phys. Rev. B [**68**]{}, 081203(R) (2003). G. Bouzerar (unpublished). L. Brey and G. Gomez-Santos, Phys. Rev. B [**68**]{}, 115206 (2003). C. Timm and A. H. MacDonald, Phys. Rev. B [**71**]{}, 155206 (2005). A. B. Shick, J. Kudrnovský, and V. Drchal Phys. Rev. B [**69**]{}, 125207 (2004). C. Zhou, M. P. Kennett, X. Wan, M. Berciu, R. N. Bhatt B[**69**]{} 144419 (2004). G. Fiete, G. Zarand, B. Janko, P. Redlinski, and C. P. Moca, Phys. Rev. B[**71**]{} 115202 (2005). J. Kudrnovský, I. Turek, V. Drchal, F. Maca, P. Weinberger, P. Bruno Phys. Rev. B [**69**]{}, 115208 (2004). G. Bouzerar, T. Ziman and J. Kudrnovsky, Europhys. Lett., [**69**]{}, 812-818 (2005); G. Bouzerar, T. Ziman and J. Kudrnovsky, Appl. Physics Lett. [**85**]{} 4941 (2004). Georges Bouzerar, Timothy Ziman and Josef Kudrnovsky, Phys. Rev. B [**72**]{}, 125207 (2005). L. Bergqvist, O. Eriksson, J. Kudrnovsky, V. Drchal, P. Korzhavyi and I. Turek, Phys. Rev. Lett., [**93**]{} 137202 (2004). K. Sato, W. Schweika, P.H. Dederichs and H. Katayama-Yoshida, Phys. Rev. B [**70**]{}, 201202(R) (2004). H.B. Callen, Phys. Rev. [**130**]{}, 890 (1963). T. Dietl, A. Haury, and Y. Merle d’Aubigné, Phys. Rev. B [**55**]{}, R3347-R3350 (1997). M. Abolfath, T. Jungwirth, J. Brum, and A. H. MacDonald, Phys. Rev. B [**63**]{}, 054418 (2001). G. Zarand, B. Janko, Phys. Rev. Lett. [**89**]{}, 047201 (2002). Y. Imry and S.-K. Ma Phys. Rev. Lett. [**35**]{}, 1399-1401 (1975). P. G. de Gennes, J. Phys. Radium [**23**]{}, 630 (1962). B. Caroli, J. Phys. Chem. Solids, [**28**]{}, 1427 (1967). S. Alexander, P. W. Anderson, Phys. Rev. 133, A1594-A1603 (1964). D. C. Price, J. Phys. F [**8**]{}, 933 (1978). Z-P Shi, P. M. Levy, J.L. Fry, Phys. Rev. B [**49**]{}, 15159 (1994). Q. Zhang, P.M. Levy, Phys. Rev. B [**34**]{}, 1884 (1986). J-I. Inoue, Phys. Rev. B [**67**]{}, 125302 (2003), J. Inoue, S. Nonoyama, and H. Itoh, Phys. Rev. Lett.[**85**]{}, 4610 (2000). P. M. Krstajic, F. M. Peeters, V. A. Ivanov, V. Fleurov, and K. Kikoin, Phys. Rev. B [**70**]{}, 195215 (2004).
{ "pile_set_name": "ArXiv" }
--- abstract: | In this note we show that a set is provably $\Delta^{0}_{2}$ in the fragment $I\Sigma_{n}$ of arithmetic iff it is $I\Sigma_{n}$-provably in the class $D_{\alpha}$ of $\alpha$-r.e. sets in the Ershov hierarchy for an $\alpha<_{\varepsilon_{0}}\omega_{1+n}$, where $<_{\varepsilon_{0}}$ denotes a standard $\varepsilon_{0}$-ordering. In the Appendix it is shown that a limit existence rule $(LimR)$ due to Beklemishev and Visser becomes stronger when the number of nested applications of the inference rule grows. author: - | Toshiyasu Arai\ Graduate School of Science, Chiba University\ 1-33, Yayoi-cho, Inage-ku, Chiba, 263-8522, JAPAN\ [email protected] title: 'Provably $\Delta^{0}_{2}$ and weakly descending chains [^1] ' --- Introduction ============ Thoroughout this paper, we identify a predicate $A$ with its characteristic function $$A(x_{1},\ldots,x_{n})=\left\{ \begin{array}{ll} 0 & \mbox{{\rm if }} A(x_{1},\ldots,x_{n}) \\ 1 & \mbox{{\rm otherwise}} \end{array} \right.$$ Natural numbers $c$ are identified with the sets $\{n\in\mathbb{N}: n<c\}$. The following Limit Lemma due to Shoenfield is a classic in computability theory. \[th:limitlemma\][(Limit Lemma)]{}\ A set $A$ of natural numbers is $\Delta^{0}_{2}$ iff there is a binary (primitive) recursive predicate $f:\omega\times\omega\to 2=\{0,1\}$ such that $$\forall c[\lim_{w\to\infty}f(c,w)=A(c)] .$$ Moreover the theorem is provable uniformly in $B\Sigma^{0}_{1}\subseteq I\Sigma^{0}_{1}$, cf. [@H-P], pp. 89-91. Let us call the predicate $f$ a [*witnessing predicate*]{} for $A\in\Delta^{0}_{2}$. In this paper we address a problem asking what can we say about the rate of convergences of the predicate $f$ under the assumption that the set $A$ is [*provably*]{} $\Delta^{0}_{2}$ in a formal (sound) theory T? This is a problem on a hierarchy. The class of $\Delta^{0}_{2}$-sets is classified in the Ershov hierarchy, [@Ershov2]. A recent article [@alcSYY] due to F. Stephan, Y. Yang and L. Yu is a readable contribution to the hierarchy, to which we refer as a standard text. The $\alpha$-th level of the Ershov hierarchy is denoted $D_{\alpha}$ for notations $\alpha$ of constructive ordinals, and a set in $D_{\alpha}$ is said to be an $\alpha$-r.e. set. It is known, as usual in hierarchic problems indexed by constructive ordinals, that $D_{\alpha}$ depends heavily on notations $\alpha$, i.e., the order type of $\alpha$ does not determine the set $D_{\alpha}$. By reason of this dependency let us fix a standard elementary recursive well ordering $<_{\alpha}$ of type $\alpha$. I don’t want to discuss here what is a ’standard ordering’ or a ’natural well ordering’. We assume that [EA]{}=$\mbox{I}\Delta^{0}_{0}+exp$, Elementary Recursive Arithmetic, proves some algebraic facts on the ordering $<_{\alpha}$. For the case $\alpha=\varepsilon_{0}$, what we need on $<_{\varepsilon_{0}}$ can be found in, e.g., [@Sommer]. In what follows let us drop the subscript $\alpha$ in $<_{\alpha}$ when no confusion likely occurs. \[df:alcSYY4.1\][(Stephan-Yang-Yu [@alcSYY])]{}\ [Let]{} $K\in dom(<)$[, the domain of the order]{} $<$. [A set]{} $A$ [of natural numbers is]{} $K$-r.e. with respect to $<$ [iff there exist a binary]{} recursive [predicate]{} $f$[, and a]{} recursive [function]{} $h:\omega \times\omega\to K=\{\beta\in dom(<): \beta<K\}$ such that 1. $$\label{th:delta2pra.1SYY} \mbox{{\rm (weakly descending)} } K>h(c,w)\geq h(c,w+1)$$ 2. $$\label{th:delta2pra.2SYY} \mbox{{\rm (lowering)} } f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. $$\label{th:delta2pra.3SYY} \forall c[\lim_{w\to\infty}f(c,w)=A(c)]$$ [Roughly speaking, a set is]{} $K$[-r.e. if the convergence of its witnessing predicate follows from the fact that weakly decreasing functions in]{} $K$ [have to be constant eventually.]{} Now suppose that we have a proof-theoretic analysis of a formal (and sound) theory T, e.g., a cut-elimination through a transfinite induction along a standard well ordering $<$. It, then, turns out that $A$ is provably $\Delta^{0}_{2}$ in T iff T proves the fact that $A\in D_{K}$ with respect to $<$ for a $K\in dom(<)$. Though, in this paper, we restrict our attention to $\mbox{T}=I\Sigma^{0}_{n}$ of fragments of first order arithmetic as a concrete example, where the order $<$ denotes a standard well ordering of type $\varepsilon_{0}$, it is easy to see that our proof works also for stronger theories, e.g., second order arithmetic $\Pi^{1}_{1}\mbox{-CA}_{0}$ and fragments of set theories. In Section \[sect:ISigman\] it is shown that for each $n\geq 1$, a set is provably $\Delta^{0}_{2}$ in the fragment $I\Sigma_{n}$ iff it is $I\Sigma_{n}$-provably in the class $D_{\alpha}$ for an $\alpha<_{\varepsilon_{0}}\omega_{1+n}$ (Theorem \[th:delta2Isigman\]). Also any provably $\Sigma^{0}_{2}$-function has a Skolem function $F(c)= \lim_{w\to\infty}f(c,w)$ as limits of an $f$, whose convergence is ensured by weakly descending chains of ordinals (Theorem \[th:sigma2witness\]). Moreover the 2-consistency $\mbox{RFN}_{\Pi^{0}_{3}}(I\Sigma^{0}_{n})$ is seen to be equivalent over Primitive Recursive Arithmetic [PRA]{} to the fact that every primitive recursive weakly descending chain of ordinals$<\omega_{1+n}$ has a limit(Theorem \[th:2con\]). In Section \[sect:PRA\] it is shown that a set is provably $\Delta^{0}_{2}$ in Elementary Recursive Arithmetic [EA]{} iff it is [EA]{}-provably in the class $D_{n}$ of a finite level (Theorem \[th:delta2pra\]). Our proof seems to be a neat application of the Herbrand’s theorem. The Appendix \[appendix\] contains another application of Herbrand’s theorem. We consider, over [EA]{}, an inference rule $(LimR)$ in [@Lev], which concludes the convergence of an elementary recursive series $\{h(n)\}_{n}$ under the assumption that the series is weakly decreasing almost all $n$. Note that $(LimR)$ is an inference rule, and not an axiom(sentence). On the other side, let $L\Sigma_{1}^{-(k)}$ denote the schema in [@KPD], saying that any non-empty $\Sigma^{0}_{1}$ $k$-ary predicate has the least tuple, which is least with respect to the lexicographic ordering on $\mathbb{N}^{k}$. It is shown that $L\Sigma_{1}^{-(k)}$ is equivalent to the $k$-nested applications of $(LimR)$. In [@KPD], Corollary 2.11 it was shown that $\{L\Sigma_{1}^{-(k)}\}_{k}$ forms a proper hierarchy, i.e., $L\Sigma_{1}^{-(k+1)}\vdash\mbox{Con}(L\Sigma_{1}^{-(k)})$. Hence we conclude that a $(k+1)$-nested application of $(LimR)$ proves the consistency of the $k$-nested applications of $(LimR)$. Provably $\Delta^{0}_{2}$ in $I\Sigma^{0}_{n}$ {#sect:ISigman} ============================================== Let [LEA]{} \[[EA]{}\] denote the Lower Elementary Recursive Arithmetic \[Elementary Recursive Arithmetic\], which is a first-order theory in the language having function constants for each code(algorithm) of lower elementary recursive function \[function constants for each code of elementary recursive function\], resp. Cf. [@Rose] and [@Skolem] for these classes of subrecursive functions. Induction schema is restricted to quantifier-free formulas in the language. The axioms of the theories [LEA]{}, [EA]{} are purely universal ones. Let $I\Sigma^{0}_{n}$ denote the fragment of arithmetic, which is a first-order theory in the language of [LEA]{}, and Induction schema is restricted to $\Sigma^{0}_{n}$ formulas. Here a $\Sigma^{0}_{0}$ formula is a quantifier-free formula. $I\Sigma^{0}_{0}$ is another name for [LEA]{}. Let $<_{\varepsilon_{0}}$ denote a standard $\varepsilon_{0}$-ordering. We assume that [EA]{} proves some algebraic facts on the ordering $<_{\varepsilon_{0}}$. What we need on $<_{\varepsilon_{0}}$ can be found in, e.g., [@Sommer]. In what follows let us drop the subscript $\varepsilon_{0}$ in $<_{\varepsilon_{0}}$ when no confusion likely occurs. For a class $\Phi$ of formulas and an ordinal $\alpha$ let $TI(\Phi,\alpha)$ denote the schema of transfinite induction up to $\alpha$ and applied to a formula $\varphi\in\Phi$: $$\forall \beta[ \forall\gamma<\beta \varphi(\gamma) \to \varphi(\beta) ] \to \forall \beta<\alpha\varphi(\beta) .$$ Let $$\omega_{0}:=1, \: \omega_{1+n}:=\omega^{\omega_{n}} .$$ Here is a folklore result on provability of the restricted transfinite induction schemata in fragments of arithmetic. \[th:Sommer\][(See, e.g., [@Sommer])]{}\ For each $n\geq 0$, $I\Sigma^{0}_{n} \vdash TI(\Pi^{0}_{1},\alpha)$ iff $\alpha<\omega_{1+n}$. The following Theorem \[th:delta2Isigman\] states that for positive integers $n$, a set is provably $\Delta^{0}_{2}$ in $I\Sigma^{0}_{n}$ iff it is $I\Sigma^{0}_{n}$-provably in the class $D_{\alpha}$ of $\alpha$-r.e. sets in the Ershov hierarchy for an $\alpha<_{\varepsilon_{0}}\omega_{1+n}$. Moreover (weakly descending) and (lowering) are provable in [EA]{}. \[th:delta2Isigman\] For positive integers $n$, the following are equivalent for quantifier-free $A,B$ and a free variable $c$. 1. $I\Sigma^{0}_{n}$ proves $$\label{eq:delta2} \forall x\exists y A(x,y,c) \leftrightarrow \exists z\forall u B(z,u,c)$$ 2. There exists a binary [elementary]{} recursive predicate $f$, an ordinal $K<\omega_{1+n}$ and an [elementary]{} recursive function $h:\omega \times\omega\to K$ such that 1. [(weakly descending)]{} $$\mbox{{\rm {\sf EA}}}\vdash K>h(c,w)\geq h(c,w+1)$$ 2. [(lowering)]{} $$\mbox{{\rm {\sf EA}}}\vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. $$\begin{aligned} \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=0 \to \exists z\forall u B(z,u,c) \\ \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=1 \to \exists x\forall y \lnot A(x,y,c) \\ I\Sigma^{0}_{n} & \vdash & \exists z\forall u B(z,u,c) \to \forall x\exists y A(x,y,c)\end{aligned}$$ where the ordering $<$ denotes a standard $\varepsilon_{0}$-ordering $<_{\varepsilon_{0}}$. First note that by Theorem \[th:Sommer\] we have $\Sigma^{0}_{1}$-minimization up to each ordinal less than $\omega_{1+n}$ in $I\Sigma^{0}_{n}$. Hence $\exists\alpha<_{\varepsilon_{0}}K[\alpha=\min_{<_{\varepsilon_{0}}}\{\beta: \exists w[\beta=h(c,w)]\}]$. Pick a $w$ so that the least $\alpha=h(c,w)$. Assuming that [[EA]{}]{} (a fortiori $I\Sigma^{0}_{n}$) proves (weakly descending) and (lowering), we have $$I\Sigma^{0}_{n} \vdash \forall u\geq w[f(c,u)=f(c,w)] .$$ Therefore the convergence of the predicate $f$ is shown in $I\Sigma^{0}_{n}$. Also $$I\Sigma^{0}_{n} \vdash \forall x\exists y A(x,y,c)\leftrightarrow \lim_{w\to \infty}f(c,w)=0 .$$ The converse follows from the following Reduction Theorem \[th:reduction\]. The theorem says that if a disjunction $\exists x\forall y \lnot A(x,y,c)\lor \exists z\forall u B(z,u,c)$ of $\Sigma^{0}_{2}$-formulas is provable in $I\Sigma^{0}_{n}$, then one can construct an elementary recursive predicate $f$ whose limit tells us which disjunct is true. The convergence of $f$ is ensured by a descending function $h$ in ordinals$<\omega_{1+n}$. Moreover these are all provable in [EA]{}. Assuming the convergence of $f$(, which is provable in $I\Sigma^{0}_{n}$) this is a provable version of the classical Reduction Property of $\Sigma^{0}_{2}$ sets to $\Delta^{0}_{2}$ sets. The point is that the $\Delta^{0}_{2}$ sets $\{c:\lim_{w\to\infty}f(c,w)=0\}$ are in a level $D_{<\omega_{1+n}}$ of Ershov hierarchy, demonstrably in $I\Sigma^{0}_{n}$. \[th:reduction\][(Reduction Property)]{} Let $n\geq 1$. Suppose $I\Sigma^{0}_{n}\vdash\exists x\forall y \lnot A(x,y,c)\lor \exists z\forall u B(z,u,c)$ for quantifier-free $A,B$. Then there exists an [elementary]{} recursive predicate $f$, an ordinal $K<\omega_{1+n}$ and an [elementary]{} recursive function $h$ such that 1. [(weakly descending)]{} $$\mbox{{\rm {\sf EA}}}\vdash K>h(c,w)\geq h(c,w+1)$$ 2. [(lowering)]{} $$\mbox{{\rm {\sf EA}}}\vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. [(reduction)]{} $$\begin{aligned} \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=0 \to \exists z\forall u B(z,u,c) \\ \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=1 \to \exists x\forall y \lnot A(x,y,c)\end{aligned}$$ In what follows, given a $I\Sigma^{0}_{n}$-proof of $\exists x\forall y \lnot A(x,y,c)\lor \exists z\forall u B(z,u,c)$ let us construct a predicate $f$, an ordinal $K<\omega_{1+n}$ and a function $h$ enjoying (weakly descending), (lowering) and (reduction). Let $p(x,y,c)$ denote the characteristic function of the predicate $$A((x)_{0},(y)_{0},c)\to B((x)_{1},(y)_{1},c) ,$$ where $(x)_{i}\, (i=0,1)$ denotes the projections of the pairing function. Then $$\exists x\forall y[ p(x,y,c)=0]$$ is provable in $I\Sigma^{0}_{n}$. Infinitary derivations ---------------------- In what follows let us consider (finite or infinite) derivations in one-sided sequent calculi. Given a finite derivation of $\exists x\forall y[ p(x,y,c)=0]$ in $I\Sigma^{0}_{n}$, first eliminate cut inferences partially to get a derivation of the same formula in which any cut formula is $\Sigma^{0}_{n}$. Next embed the derivation into an infinite derivation of the sentence $$\exists x\forall y[ p(x,y,\bar{c})=0]$$ with the $c$-th numeral $\bar{c}$. Then eliminate cut inferences to get a cut-free derivation $P_{c}$ of the same sentence. As usual the depth of $P_{c}$ is bounded by an ordinal $K<\omega_{1+n}$ uniformly, i.e., $\forall c[\mbox{dp}(P_{c})<K]$. In the derivation $P_{c}$, the [*initial sequents*]{} are $$(Int)\:\: \Gamma, E$$ for true equation $E$. The equation $E$ is called the [*main formula*]{} of the initial sequent. In what follows we identify the closed term $t$ with the numeral $\bar{n}$ of its value $n=val(t)$. Note that the value of closed terms and truth values of equations in [LEA]{} are elementary recursively computable. The initial sequents are regarded as inference rules with empty premiss (upper sequent), and with the empty list of side formulas. The [*inference rules*]{} are $(\exists), (\forall)$, and the repetition rule $(Rep)$. These are standard ones. $$\infer[(\exists)] {\Gamma, \exists x B(x)} {\Gamma, B(\bar{n})} \: ;\: \infer[(\forall)] {\Gamma, \forall x B(x)} { \cdots & \Gamma,B(\bar{n}) & \cdots (n\in\omega) } \: ;\: \infer[(Rep)] {\Gamma}{\Gamma}$$ where $\exists x B(x)$ in the $(\exists)$ and $\forall x B(x)$ in the $(\forall)$ are the [*main formula*]{} of the inference, and $B(\bar{n})$ are [*side formulas*]{} of the inferences. The inference $(Rep)$ has no main nor side formulas. Our infinitary derivations are equipped with additional informations as in [@Mintsfinite]. \[df:derivation\] [An]{} infinitary derivation [is a sextuple]{} $$D=(T,Seq,Rule,Mfml,Sfml,ord)$$ [which enjoys the following conditions. The naked tree of]{} $D$ [is denoted]{} $T=T(D)$. 1. $T\subseteq{}^{<\omega}\omega$ [is a tree with its root]{} $\emptyset$ [such that]{} $$a*\langle n\rangle\in T \,\&\, m<n \Rightarrow a*\langle m\rangle\in T .$$ 2. $Seq(a)$ [for]{} $a\in T$ [denotes the sequent situated at the node]{} $a$. [If]{} $Seq(a)$ [is a sequent]{} $\Gamma$[, then it is denoted]{} $$a:\Gamma .$$ 3. $Rule(a)$ [for]{} $a\in T$ [denotes the name of the inference rule with its lower sequent]{} $Seq(a)$. 4. $Mfml(a)$ [for]{} $a\in T$ [denotes the]{} main formula [of the inference rule]{} $Rule(a)$. [When]{} $Rule(a)=(Rep)$[, then]{} $Mfml(a)=\emptyset$. 5. $Sfml(a*\langle n\rangle)$ [for]{} $a*\langle n\rangle\in T$ [denotes the]{} side formula [of the inference rule]{} $Rule(a)$[, which is in the]{} $n$[-th upper sequent, i.e.,]{} $Sfml(a*\langle n\rangle)\in Seq(a*\langle n\rangle)$. [When]{} $Rule(a)=(Rep), (Int)$[, then]{} $Sfml(a*\langle n\rangle)=\emptyset$. 6. $ord(a)$ [for]{} $a\in T$ [denotes the ordinal]{}$<_{\varepsilon_{0}}K$ [attached to]{} $a$. 7. [The sextuple]{} $(T,Seq,Rule,Mfml,Sfml,ord)$ [has to be locally correct with respect to inference rules of the infinitary calculus and for being well founded tree]{} $T$. In a derivation each inference rule except $(Int)$ receives the following nodes: $$\infer[(\exists)] {a: \Gamma, \exists x B(x)} {a*\langle 0\rangle : \Gamma, B(\bar{n})} \: ;\: \infer[(\forall)] {a: \Gamma, \forall x B(x)} { \cdots & a*\langle n\rangle : \Gamma,B(\bar{n}) & \cdots (n\in\omega) } \: ;\: \infer[(Rep)] {a: \Gamma}{a*\langle 0\rangle: \Gamma}$$ The ordinals $ord_{c}(a)$ in the inference $(\forall)$ $$\infer[(\forall)] {a: \Gamma, \forall x B(x)} { \cdots & a*\langle n\rangle : \Gamma,B(\bar{n}) & \cdots (n\in\omega) }$$ enjoys $$\label{eq:ordass} ord_{c}(a)>ord_{c}(a*\langle n\rangle)=ord_{c}(a*\langle m\rangle)$$ for any $n,m$. As in [@Mintsfinite] we see that the function $c\mapsto P_{c}$ is elementary recursive. We denote $P_{c}=(T_{c},Seq_{c},Rule_{c},Mfml_{c},Sfml_{c},ord_{c})$. Searching witnesses of $\Sigma^{0}_{2}$ in derivations ------------------------------------------------------ Let us define a tracing function $\sigma(c,i)\in T_{c}=T(P_{c})$. The function $\{\sigma(c,w)\}_{w}$ indicates the trail in the proof tree $T_{c}$ in which we go through in searching a witness $x_{a}$ of $\exists x\forall y[ p(x,y,\bar{c})=0]$, and verifying $\forall y[p(x_{a},y,\bar{c})=0]$. 1. $\sigma(c,0)=\emptyset$(root). In what follows let $a=\sigma(c,w)$. 2. Until $Seq_{c}(a)$ is an upper sequent of an $(\forall)$, go to the leftmost branch: $$\sigma(c,w+1)=a*\langle 0\rangle .$$ For example $$\infer[(\exists)] {a: \Gamma,\exists x\forall y[p(x,y,\bar{c})=0]} { a*\langle 0\rangle : \Gamma,\exists x\forall y[p(x,y,\bar{c})=0],\forall y[p(x_{a},y,\bar{c})=0] }$$ 3. The case when $Rule_{c}(b)=(\forall)$ with $a=b*\langle n\rangle$. Namely $Seq_{c}(a)$ is the $n$-th upper sequent of an $(\forall)$. $$\infer[(\forall)] {\Gamma,\forall y[p(x_{a},y,\bar{c})=0]} { \cdots & a: \Gamma,p(x_{a},y_{a},\bar{c})=0 & \cdots }$$ $x_{a},y_{a}$ are closed terms. 1. If $p(x_{a},y_{a},\bar{c})=0$ is a TRUE equation, $\sigma(c,w+1)=a\oplus 1$, the next right to the $a$: $$\infer[(\forall)] {\Gamma,\forall y[p(x_{a},y,\bar{c})=0]} { \sigma(c,w) : \Gamma,p(x_{a},y_{a},\bar{c})=0 & \sigma(c,w+1) : \Gamma,p(x_{a},y_{a}+1,\bar{c})=0 }$$ where for an $a=(a_{0},\ldots,a_{n-2},a_{n-1})\in{}^{<\omega}\omega$ $$a\oplus 1=(a_{0},\ldots,a_{n-2},a_{n-1}+1)$$ if $lh(a)=n>0$. $\emptyset\oplus 1$ is defined to be $\emptyset$. 2. Otherwise $\sigma(c,w+1)=a*\langle 0\rangle$, i.e., go to the leftmost branch from $a$. $$\infer[(\forall)] {\Gamma,\forall y[p(x_{a},y,\bar{c})=0]} { \cdots & \infer{\sigma(c,w) : \Gamma,p(x_{a},y_{a},\bar{c})=0}{\sigma(c,w+1):\Delta & \cdots} & \cdots }$$ It is easy to see that the function $(c,w)\mapsto \sigma(c,w)$ is elementary recursive since $\max(\{(\sigma(c,w))_{i}: i<lh(\sigma(c,w))\}\cup\{lh(\sigma(c,w))\})\leq w$. Once $\sigma(c,w)$ is on an $(\forall)$, the tracing function goes through the upper sequents as long as the equations $p(x_{a},y_{a},\bar{c})=0$ is TRUE. It is intuitively clear that after a finite number of steps, the sequence $\{\sigma(c,w)\}_{w}$ goes through the upper sequents of an $(\forall)$: $$\infer[(\forall)] {\Gamma,\forall y[p(x_{a},y,\bar{c})=0]} { \sigma(c,w_{0}) : \Gamma,p(x_{a},0,\bar{c})=0 & \cdots & \sigma(c,w_{0}+y) : \Gamma,p(x_{a},\bar{y},\bar{c})=0 & \cdots }$$ since $\forall y[p(x_{a},y,\bar{c})=0]$ is true for an $x_{a}$. We will know at the limit the fact, i.e., for $x=(x_{a})_{0}$ and $z=(x_{a})_{1}$ $$\exists y A(\bar{x},y,\bar{c})\to \forall u B(\bar{z},u,\bar{c})$$ is true. Now let us define an elementary recursive predicate $f$ as follows. 1. $f(c,0)=1$. 2. Alternate values $f(c,w+1)=1-f(c,w)$ if $Seq_{c}(\sigma(c,w+1))$ is an upper sequent of an inference other than $(\forall)$. 3. Suppose $Seq_{c}(\sigma(c,w+1))$ is the $n$-th upper sequent of an $(\forall)$, and $\sigma(c,w+1)=b*\langle n\rangle$. $$\infer[(\forall)] {b: \Gamma,\forall y,u[A(x_{b},y,\bar{c})\to B(z_{b},u,\bar{c})]} { \cdots & b*\langle n\rangle : \Gamma, A(x_{b},(\bar{n})_{0},\bar{c})\to B(z_{b},(\bar{n})_{1},\bar{c}) & \cdots }$$ $f(c,w+1)=0$ iff $A(x_{b},(n)_{0},\bar{c})\to B(z_{b},(n)_{1},\bar{c})$ is true, and the following condition holds: $$\exists k\leq n [A(x_{b}, (\bar{k})_{0},\bar{c})]$$ Namely $$\begin{aligned} && f(c,w+1)=0 \Leftrightarrow \\ && [A(x_{b},(n)_{0},\bar{c})\to B(z_{b},(n)_{1},\bar{c})] \,\&\, \exists k\leq n [A(x_{b}, (\bar{k})_{0},\bar{c})]\end{aligned}$$ Suppose $\sigma(c,w)$ is on an $(\forall)$. Until a witness $k$ such that $A(x_{b}, (\bar{k})_{0},\bar{c})$ is found, $f(c,w)=1\, (w<k)$. After a witness $k$ has been found, $f(c,w)=0\, (w\geq k)$ as long as $A(x_{b},(\bar{n})_{0},\bar{c})\to B(z_{b},(\bar{n})_{1},\bar{c})$ is true. Therefore if the tracing function $\sigma(c,w)$ goes through the upper sequents of the $(\forall)$, then either $\lim_{w\to\infty}f(c,w)=1$ and $\forall y \lnot A(x_{b}, y,\bar{c})$, or $\lim_{w\to\infty}f(c,w)=0$ and $\forall u B(z_{b},u,\bar{c})$. \[prp:delta2Isigman.1\] 1. \[prp:delta2Isigman.11\] Suppose that $b*\langle n\rangle=\sigma(c,w+1)$ and $Seq_{c}(b*\langle n\rangle)$ is the $n$-th upper sequent of an inference $(\forall)$. Then $\{f(c,u): \sigma(c,u)=b*\langle m\rangle, m\leq n\}$ changes the values at most twice. Moreover if $f(c,u)=0$ and $f(c,v)=1$ for some $u<v\leq w+1$, then $v=w+1$ and $\sigma(c,v+1)=\sigma(c,v)*\langle 0\rangle$, i.e., $Seq_{c}(\sigma(c,v))$ is the last upper sequent of the inference $(\forall)$ in the tracing function $\sigma$. 2. \[prp:delta2Isigman.12\][(Reduction)]{} $$\begin{aligned} \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=0 \to \exists z\forall u B(z,u,c) \\ \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=1 \to \exists x\forall y \lnot A(x,y,c)\end{aligned}$$ [**Proof**]{}. Recall that an inference rule $(\forall)$ in $P_{c}$ is of the form: $$\infer[(\forall)] {b: \Gamma, \forall y [p(x_{b}, y,\bar{c})=0]} { \ldots & b*\langle n\rangle: \Gamma,p(x_{b*\langle n\rangle},\bar{n},\bar{c})=0 & \ldots }$$ where $$p(x_{b*\langle n\rangle},\bar{n},\bar{c})=0 \leftrightarrow [A(((x)_{b*\langle n\rangle})_{0}, (\bar{n})_{0},\bar{c}) \to B((x)_{b*\langle k\rangle})_{1}, (\bar{n})_{1},\bar{c}) ]$$ Let $u$ be such that $\sigma(c,u)=b*\langle m\rangle$ with an $m\leq n$. Then by the definition of the tracing function $\sigma$, we have for $m<n$ $p(x_{b*\langle m\rangle},\bar{m},\bar{c})=0$, i.e., $$A((x_{b*\langle m\rangle})_{0},(\bar{m})_{0},\bar{c})\to B((x_{b*\langle m\rangle})_{1},(\bar{m})_{1},\bar{c}) .$$ Suppose there exists a $u\leq w+1$ such that $f(c,u)=0$, and let $u$ denote the minimal such one. Then for any $v$ with $u\leq v<w+1$, we have $f(c,v)=0$. Therefore if $f(c,v)=1$ for a $v>u$, it must be the case $v=w+1$. This means that for some $k\leq n-1$ with $(k)_{0}=(n)_{0}$ $$A((x_{b})_{0},(k)_{0},c)\land \lnot B((x_{b})_{1},(n)_{1},c) .$$ Hence $p(x_{b},n,c)\neq 0$, and $\sigma(c,v+1)=\sigma(c,v)*\langle 0\rangle$. $\Box$ Next define $h$ as follows. 1. $$h(c,0)=3\cdot ord_{c}(\emptyset) .$$ In what follows put $a=\sigma(c,w+1)$ and let $Seq_{c}(a)$ be an upper sequent of an inference $Rule_{c}(b)$ with $a=b*\langle n\rangle$. 2. The case when $Rule_{c}(b)$ is an inference rule other than $(\forall)$. $$h(c,w+1):=3\cdot ord_{c}(\sigma(c,w+1)) .$$ By Proposition \[prp:delta2Isigman.1\].\[prp:delta2Isigman.11\] we know that the $f(c,u)$ changes the values at most twice in the upper sequents of an $(\forall)$. 3. The case when $n=0$ and $Rule_{c}(b)=(\forall)$. $$h(c,w+1):=3\cdot ord_{c}(\sigma(c,w+1))+2 .$$ 4. The case when $n>0$, $Rule_{c}(b)=(\forall)$. $$\infer[(\forall)] {b: \Gamma,\forall y,u[A(x_{b},y,\bar{c})\to B(z_{b},u,\bar{c})]} { \cdots & b*\langle n\rangle : \Gamma, A(x_{b},(\bar{n})_{0},\bar{c})\to B(z_{b},(\bar{n})_{1},\bar{c}) & \cdots }$$ We have by (\[eq:ordass\]) $$ord_{c}(\sigma(c,w))=ord_{c}(\sigma(c,w+1)) .$$ 1. The case when $f(c,w+1)=f(c,w)$. $$h(c,w+1):=h(c,w) .$$ where $\sigma(c,w+1)=\sigma(c,w)\oplus 1$. 2. The case when $f(c,w)=1\,\&\, f(c,w+1)=0$. Then $\sigma(c,w+1)=\sigma(c,w)\oplus 1$ and $n=\min\{k: A(x_{b}, (\bar{k})_{0},\bar{c})\}$. Let $$h(c,w+1):=3\cdot ord_{c}(\sigma(c,w+1))+1 .$$ 3. The case when $f(c,w)=0\,\&\, f(c,w+1)=1$. This means that $A(x_{b}, (\bar{n})_{0},\bar{c})\to B(z_{b}, (\bar{n})_{1},\bar{c})$ is FALSE and\ $\sigma(c,w+2)=\sigma(c,w+1)*\langle 0\rangle$. $$h(c,w+1):=3\cdot ord_{c}(\sigma(c,w+1)) .$$ Obviously $h$ is elementary recursive. \[prp:delta2Isigman.2\] $$\mbox{{\rm (weakly descending) }} \mbox{{\sf EA}} \vdash 3K>h(c,w)\geq h(c,w+1)$$ $$\mbox{{\rm (lowering) }} \mbox{{\sf EA}} \vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ [**Proof**]{}. (weakly descending) is obvious. Consider the case when $\sigma(c,w+1)=a$ and $Seq_{c}(a)$ is an upper sequent of an inference $Rule_{c}(b)=(\forall)$ with $a=b*\langle n\rangle$. If $n=0$, then $$h(c,w+1)=3\cdot ord_{c}(\sigma(c,w+1))+2<3\cdot ord_{c}(\sigma(c,w))\leq h(c,w)$$ since $Seq_{c}(\sigma(c,w))$ is the lower sequent of $Seq_{c}(a)$ with $b=\sigma(c,w)$. Assume $n>0$. Using Proposition \[prp:delta2Isigman.1\].\[prp:delta2Isigman.11\] we see $h(c,w+1)\in\{3\cdot ord_{c}(\sigma(c,w+1))+i: i<3\}$. Moreover if $\sigma(c,w+2)=\sigma(c,w+1)*\langle 0\rangle$, then $$h(c,w+1)\geq 3\cdot ord_{c}(\sigma(c,w+1))>3\cdot ord_{c}(\sigma(c,w+1))+2\geq h(c,w+2) .$$ $\Box$ This completes a proof of Theorems \[th:reduction\] and \[th:delta2Isigman\]. Provably $\Sigma^{0}_{2}$-functions ----------------------------------- If $\exists z\forall u B(z,u,c)$ is provable for quantifier-free $B$, then we can find a witness $z= \lim_{w\to\infty}f(c,w)$ as limits of an $f$, whose convergence is ensured by weakly descending chains of ordinals. \[th:sigma2witness\] Suppose $I\Sigma^{0}_{n}\vdash \exists z\forall u B(z,u,c)$ for quantifier-free $B$. Then there exist [elementary]{} recursive functions $f$, $h$ and an ordinal $K<\omega_{1+n}$ such that 1. $$\mbox{{\rm (weakly descending) }} \mbox{{\sf EA}} \vdash K>h(c,w)\geq h(c,w+1)$$ 2. $$\mbox{{\rm (lowering) }} \mbox{{\sf EA}} \vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. $$\mbox{{\sf EA}} \vdash \lim_{w\to\infty}f(c,w)=z \to \forall u B(z,u,c)$$ [**Proof**]{}. As in the proof of Theorem \[th:reduction\], let us define a tracing function $\sigma$. $\sigma(c,w)$ goes on the leftmost branch up to an $(\forall)$. $\sigma(c,w)$ goes through the upper sequents of $(\forall)$ as long as side formulas $B(z_{a},\bar{n},\bar{c})$ is TRUE. If a FALSE side formula $B(z_{a},\bar{n},\bar{c})$ is found, then throw $z_{a}$ away and go on the leftmost branch. Now $h$ is defined by $h(c,w):=ord_{c}(\sigma(c,w))$. $f$ is defined obviously. $f(c,w)=z_{a}$ if $Seq_{c}(\sigma(c,w))$ is an upper sequent of an $(\forall)$ with its side formula $B(z_{a},\bar{n},\bar{c})$. Otherwise $f(c,w)$ is arbitrary, say $f(c,w)=0$. $\Box$\ \ It is well known that the 1-consistency $\mbox{RFN}_{\Pi^{0}_{2}}(I\Sigma^{0}_{n})$ is equivalent over Primitive Recursive Arithmetic [PRA]{} to the fact that there is no primitive recursive descending chain of ordinals$<\omega_{1+n}$. \[th:2con\](Cf. [@AraiMints] for another form of the 2-consistency of arithmetic.) The 2-consistency $\mbox{{\rm RFN}}_{\Pi^{0}_{3}}(I\Sigma^{0}_{n})$ is equivalent over [PRA]{} to the fact that every primitive recursive weakly descending chain of ordinals$<\omega_{1+n}$ has a limit, or equivalently to the fact that for any primitive recursive sequence $\{h(c,w)\}_{w}$ of ordinals$<\omega_{1+n}$ the least ordinal $\min_{<_{\varepsilon_{0}}}\{h(c,w): w\in\omega\}$ exists. [**Proof**]{}. Over [PRA]{}, $\mbox{RFN}_{\Pi^{0}_{3}}(I\Sigma^{0}_{n})$ yields the existence of the least ordinal $\min_{<_{\varepsilon_{0}}}\{h(c,w)<\omega_{1+n}: w\in\omega\}$ since $\alpha=\min_{<_{\varepsilon_{0}}}\{\beta: \exists w[\beta=h(c,w)]\}$ is a $\Sigma^{0}_{2}$-formula. Conversely let $f(c,w)<2$ be defined as follows: 1. $c$ is not a Gödel number of an $I\Sigma^{0}_{n}$-proof of a $\Sigma^{0}_{2}$-sentence: Then $f(c,w)=0$ for any $w$. 2. $c$ is a Gödel number of an $I\Sigma^{0}_{n}$-proof of a $\Sigma^{0}_{2}$-sentence $\exists z\forall u B_{c}(z,u)$:\ $f(c,w)$ is defined as in Theorem \[th:sigma2witness\] for a cut free infinite derivation $P_{c}$ of $\exists z\forall u B_{c}(z,u)$. Note that $f$ is non-elementary since it involves cut elimination for predicate logic. Also let $h(c,w):=ord_{c}(\sigma(c,w))$. Then 1. $$\mbox{{\rm (weakly descending) }} \mbox{{\sf PRA}} \vdash \omega_{1+n}>h(c,w)\geq h(c,w+1)$$ 2. $$\mbox{{\rm (lowering) }} \mbox{{\sf PRA}} \vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. $$\mbox{{\sf PRA}} \vdash \lim_{w\to\infty}f(c,w)=0 \to \mbox{{\rm Prov}}_{I\Sigma^{0}_{n}}(c,\lceil \exists z\forall u B_{c}(z,u)\rceil)\to \exists z\forall u B_{c}(z,u)$$ Therefore $$\mbox{{\sf PRA}}\vdash \forall c [\exists w\forall u\geq w \{h(c,u)=h(c,w)\} \to \exists\ell\{ \lim_{w\to\infty}f(c,w)=\ell\} ]$$ and $$\mbox{{\sf PRA}}\vdash \forall c\exists\ell[ \lim_{w\to\infty}f(c,w)=\ell]\to \mbox{RFN}_{\Pi^{0}_{3}}(I\Sigma^{0}_{n}) .$$ $\Box$ Provably $\Delta^{0}_{2}$ in [EA]{} {#sect:PRA} ==================================== In this section we consider the $\Delta^{0}_{2}$-sets provably in [EA]{}. The following Theorem \[th:delta2pra\] states that for a set is provably $\Delta^{0}_{2}$ in [EA]{} iff it is [EA]{}-provably in the class $D_{n}$ of a finite level in the Ershov hierarchy. The finite levels $\{D_{n} : n<\omega\}$ are called the difference (or Boolean) hierarchy, and by a result due to H. Putnam(Theorem 2 in [@Putnam]) we see that a set is provably $\Delta^{0}_{2}$ in [EA]{} iff it is equivalent to a Boolean combination of $\Sigma^{0}_{1}$-formulas, provably in [EA]{}. This answers to a problem of L. Beklemishev. \[th:delta2pra\] The following are equivalent for quantifier-free $A,B$ and a free variable $c$. 1. [[EA]{}]{} proves $$\renewcommand{\theequation}{\ref{eq:delta2}} \forall x\exists y A(x,y,c) \leftrightarrow \exists z\forall u B(z,u,c)$$ 2. There exists a binary [elementary]{} recursive predicate $f$, a natural number $K<\omega$ and an [elementary]{} recursive function $h:\omega \times\omega\to K$ such that 1. [(weakly descending)]{} $$\mbox{{\rm {\sf EA}}}\vdash K>h(c,w)\geq h(c,w+1)$$ 2. [(lowering)]{} $$\mbox{{\rm {\sf EA}}}\vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. [(reduction)]{} $$\begin{aligned} \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=0 \to \exists z\forall u B(z,u,c) \\ \mbox{{\sf EA}} & \vdash & \lim_{w\to\infty}f(c,w)=1 \to \exists x\forall y \lnot A(x,y,c) \\ \mbox{{\sf EA}} & \vdash & \exists z\forall u B(z,u,c) \to \forall x\exists y A(x,y,c)\end{aligned}$$ for the usual ordering $<$ on $\omega$. [**Proof**]{}. Assume [EA]{} proves (weakly descending) and (lowering) for a natural number $K$. Then [EA]{} also proves the convergence of $f$: $$\mbox{{\rm {\sf EA}}}\vdash\exists \ell[\lim_{w\to\infty}f(c,w)=\ell] ,$$ (reduction) yields $$\mbox{{\rm {\sf EA}}}\vdash \forall x\exists y A(x,y,c)\leftrightarrow \lim_{w\to \infty}f(c,w)=0 .$$ Conversely suppose that [EA]{} proves (\[eq:delta2\]). Then so is the $\exists\forall$-formula $$\exists x\exists z\forall y\forall z[A(x,y,c)\to B(z,u,c)] .$$ By the Herbrand’s theorem there exist a list of variables $\{a_{i},b_{i}: i\leq r\}$ and a list of terms $\{t_{i},s_{i}:i\leq r\}$ such that $$\label{eq:Herbrand} \bigvee\{ A(t_{i}, a_{i},c) \to B(s_{i},b_{i},c) :i\leq r\}$$ is provable in [EA]{}, and variables occurring in $t_{i},s_{i}$ are among $a_{j},b_{j}$ for $j<i$ besides the parameter $c$. For simplicity consider the case when $r=1$. Then we have $$\label{eq:Herbrand1} \mbox{{\sf EA}}\vdash \lnot A(t_{0}, a_{0},c) \lor B(s_{0},b_{0},c) \lor \lnot A(t_{1}(a_{0},b_{0}), a_{1},c) \lor B(s_{1}(a_{0},b_{0}),b_{1},c)$$ Let $f$ denote the elementary recursive predicate $$f(c,w):=\left\{ \begin{array}{ll} 0 & [\{t_{0}\leq w \land \exists y\leq w A(t_{0},y,c)\} \land \{ s_{0}\leq w \land \forall u\leq w B(s_{0},u,c)\}] \lor \\ & [\exists a_{0},b_{0}\leq w\{A(t_{0},a_{0},c)\land \lnot B(s_{0},b_{0},c) \land \\ & \exists a_{1}\leq w A(t_{1}(a_{0},b_{0}), a_{1},c) \land \forall b_{1}\leq w B(s_{1}(a_{0},b_{0}),b_{1},c)\}] \\ 1 & \mbox{ otherwise} \end{array} \right.$$ For the number $$K:=1+2r+2 (=5 \mbox{ if } r=1) ,$$ let $h:\omega\times\omega\to K$ denote the elementary recursive function $$h(c,0)=K-1$$ and $$h(c,w+1):=\left\{ \begin{array}{ll} h(c,w) & \mbox{ if } f(c,w+1)=f(c,w) \\ h(c,w)\dot{-}1 & \mbox{ if } f(c,w+1)\neq f(c,w) \end{array} \right.$$ \[lem:1\] [[EA]{}]{} proves the facts (weakly descending), (lowering) and (reduction). [**Proof**]{}. Argue in [EA]{}. (weakly descending) is obvious. Suppose $$\lim_{w\to\infty}f(c,w)=\ell$$ for an $\ell=0,1$. By (\[eq:delta2\]) we have either $\exists z\forall u B(z,u,c)$ or $\exists x\forall y \lnot A(x,y,c)$. First consider the case when $\exists z\forall u B(z,u,c)$. Then $\forall x\exists y A(x,y,c)$. Hence by (\[eq:Herbrand1\]) either $\forall b_{0} B(s_{0},b_{0},c)$ or $\forall b_{1} B(s_{1}(a_{0},b_{0}),b_{1},c)$ for some $a_{0}, b_{0}$ with $A(t_{0},a_{0},c)\land \lnot B(s_{0},b_{0},c)$. If $\forall b_{0} B(s_{0},b_{0},c)$, then $f(c,w)=0$ for any $w\geq\max\{t_{0},s_{0},y_{0}\}$, where $y_{0}=\mu y. A(t_{0},y.c)$. Therefore $\ell=0$. Moreover $f(c,w)=1$ for $w<\max\{t_{0},s_{0},y_{0}\}$. Next assume $\forall b_{1} B(s_{1}(a_{0},b_{0}),b_{1},c)$ for the minimal $a_{0}.b_{0}$ such that $A(t_{0},a_{0},c)\land \lnot B(s_{0},b_{0},c)$. Then let $a_{1}$ denote the minimal $a_{1}$ such that $A(t_{1}(a_{0},b_{0}),a_{1},c)$. We have $f(c,w)=0$ for any $w\geq\max\{a_{0},b_{0},a_{1}\}$, and hence $\ell=0$. Now consider $w<\max\{a_{0},b_{0},a_{1}\}$. Then $f(c,w)=0$ iff $\max\{t_{0},s_{0},a_{0}\}\leq w<b_{0}$. Therefore $\lambda w.f(c,w)$ changes its values at most three times (when $\max\{t_{0},s_{0},a_{0}\}<b_{0}<a_{1}$). Next consider the case when $\exists x\forall y \lnot A(x,y,c)$. We have $\forall z\exists u\lnot B(z,u,c)$. Then $f(c,w)=1$ for any $w$ if $\forall a_{0}\lnot A(t_{0},a_{0},c)$, and $f(c,w)=1$ for any $w\geq\max\{b_{0},b_{1}\}$ if $\forall a_{1}\lnot A(t_{1}(a_{0},b_{0}),a_{1},c)$ for the minimal $a_{0}, b_{0},b_{1}$ such that $A(t_{0},a_{0},c)\land \lnot B(s_{0},b_{0},c)$ and $\lnot B(s_{1}(a_{0},b_{0}),b_{1},c)$. Therefore $\ell=1$. Finally assume $\forall a_{1}\lnot A(t_{1}(a_{0},b_{0}),a_{1},c)$, and consider $w<\max\{b_{0},b_{1}\}$. Then $f(c,w)=0$ iff $\max\{t_{0},s_{0},a_{0}\}\leq w<b_{0}$. Therefore $\lambda w.f(c,w)$ changes its values at most two times in this case. In any cases, (reduction) was shown, and $\lambda w.f(c,w)$ changes its values at most $1+2r(=3 \mbox{ {\rm if }} r=1)$ times for any $c$, i.e., $$\forall (w_{0}<w_{1}<\cdots<w_{1+2r})\exists i\leq 1+2r [ f(c, w_{i})=f(c, w_{i}+1)] .$$ Hence (lowering) follows. $\Box$ Lemma \[lem:1\] with a result due to H. Putnam(Theorem 2 in [@Putnam]) yields the \[th:1\] Suppose that [EA]{} proves $$\forall x\exists y A(x,y,c) \leftrightarrow \exists z\forall u B(z,u,c)$$ for quantifier-free $A,B$. Then over [EA]{} $\exists z\forall u B(z,u,c)$ is equivalent to a Boolean combination of $\Sigma^{0}_{1}$-formulas. [**Proof**]{}.(cf. [@Putnam].) Let $r$ be as in (\[eq:Herbrand\]). Define $\Sigma^{0}_{1}$ $Y_{k}(c)\, (k\leq 1+2r)$ and $N_{i}(c)$ by $$\begin{aligned} Y_{k}(c) & :\Leftrightarrow & \exists (w_{0}<w_{1}<\cdots<w_{k-1})\forall i<k [ f(c, w_{i})\neq f(c, w_{i}+1) \\ && \land f(c,w_{k-1})=0] \\ N_{k}(c) & :\Leftrightarrow & \exists (w_{0}<w_{1}<\cdots<w_{k-1})\forall i<k [ f(c, w_{i})\neq f(c, w_{i}+1) \\ && \land f(c,w_{k-1})=1]]\end{aligned}$$ for $k>0$, and $Y_{0}(c):\Leftrightarrow f(c,0)=0$, $N_{0}(c):\Leftrightarrow f(c,0)=1$. Also put $N_{2r+2}(c):\Leftrightarrow 0=1$. Then [EA]{} proves that $$\exists z\forall u B(z,u,c) \leftrightarrow \bigvee\{Y_{k}(c)\land\lnot N_{k+1}(c): k\leq 1+2r\} .$$ $\Box$ As in Theorems \[th:sigma2witness\], \[th:2con\] we see the following theorems. \[th:sigma2witnesspra\] Suppose ${\sf EA}\vdash \exists z\forall u B(z,u,c)$ for quantifier-free $B$. Then there exist [elementary]{} recursive functions $f$, $h$ and a natural number $K<\omega$ such that 1. $$\mbox{{\rm (weakly descending) }} \mbox{{\sf EA}} \vdash K>h(c,w)\geq h(c,w+1)$$ 2. $$\mbox{{\rm (lowering) }} \mbox{{\sf EA}} \vdash f(c,w)\neq f(c,w+1) \to h(c,w)>h(c,w+1)$$ 3. $$\mbox{{\sf EA}} \vdash \lim_{w\to\infty}f(c,w)=z \to \forall u B(z,u,c)$$ \[th:2conpra\] The 2-consistency $\mbox{{\rm RFN}}_{\Pi^{0}_{3}}({\sf EA})$ is equivalent over [PRA]{} to the fact that every primitive recursive weakly descending chain of natural number$<\omega$ has a limit, or equivalently to the fact that for any primitive recursive sequence $\{h(c,w)\}_{w}$ of natural number$<\omega$ the least number $\min_{<}\{h(c,w)<\omega: w\in\omega\}$ exists. [**Remark**]{}. Obviously Theorems \[th:delta2pra\], \[th:1\] and \[th:sigma2witnesspra\] hold for any purely universal extension of [EA]{}, eg., [EA]{}+CON([EA]{}), [PRA]{}. Nested limit existence rules {#appendix} ============================ Every fragment in the Appendix is an extension of Elementary Recursive Arithmetic [EA]{}. In [@Lev], Beklemishev and Visser gave an elegant axiomatization of $\Sigma^{0}_{2}$-consequences of $I\Sigma^{0}_{1}$ in terms of the inference rule $(LimR)$ for limit existence principle: $$\infer[(LimR)]{\exists m\forall n\geq m \, h(n)=h(m)}{\exists m\forall n\geq m\, h(n+1)\leq h(n)}$$ Moreover unnested applications of $(LimR)$ is shown to be equivalent to $I\Pi_{1}^{-}$ (over [EA]{}). This reminds us another axiomatization of $\Sigma^{0}_{2}$-consequences of $I\Sigma^{0}_{1}$ in [@KPD]. Namely $I\Sigma^{0}_{1}$ is a $\Sigma^{0}_{2}$ conservative extension of $L\Sigma_{1}^{-(\infty)}=\bigcup_{k}L\Sigma_{1}^{-(k)}$, where $L\Sigma_{1}^{-(k)}$ denotes the schema $$\begin{aligned} && \exists x_{1}\cdots\exists x_{k} \theta(x_{1},\ldots,x_{k}) \to \\ && \exists x_{1}\cdots\exists x_{k} \bigwedge_{i=1}^{k}[ \exists\vec{y} \theta(x_{1},\ldots,x_{i},\vec{y}) \land \forall z<x_{i}\forall\vec{y}\lnot\theta(x_{1},\ldots,x_{i-1},z,\vec{y})]\end{aligned}$$ for $\theta\in\Sigma^{0}_{1}$ without parameters. For example $L\Sigma_{1}^{-(0)}={\sf EA}$ and $L\Sigma_{1}^{-(1)}=L\Sigma_{1}^{-}=I\Pi_{1}^{-}$. In this Appendix we show that $L\Sigma_{1}^{-(k)}$ is equivalent to the $k$-nested applications of $(LimR)$. To be precise, let $(LimR)^{(k)}\vdash$ denote the derivability in the $k$-nested applications of $(LimR)$: $(LimR)^{(0)}\vdash$ is nothing but ${\sf EA}\vdash$, and if $(LimR)^{(k)}\vdash\exists m\forall n\geq m\, h(n+1)\leq h(n)$, then $(LimR)^{(k+1)}\vdash\exists m\forall n\geq m \, h(n)=h(m)$. \[th:klim\] $(LimR)^{(k)}\vdash \varphi \Leftrightarrow L\Sigma_{1}^{-(k)}\vdash\varphi$ for any $\varphi$. This is shown by induction on $k$. The proof is obtained by a slight modification of proofs in [@Lev]. First consider $$(LimR)^{(k)}\vdash L\Sigma_{1}^{-(k)} .$$ Let $<^{(k)}\, (k\geq 1)$ denote the lexicographic order on $k$-tuples of natural numbers. Also $\langle x_{1},\ldots,x_{k}\rangle^{(k)}$ denotes a(n elementary recursive) bijective coding of $k$-tuples with its inverses $(n)_{i}^{(k)}\, (1\leq i\leq k)$. In what follows the super scripts $(k)$ are omitted. Then $L\Sigma_{1}^{-(k)}$ says that if there exists an $x$ satisfying $\varphi(x)\equiv\theta((x)_{1},\ldots,(x)_{k})$, then there exists a minimal such $x$ with respect to $<^{(k)}$. We can assume that [EA]{} proves $$\label{eq:lex} \exists i[\forall j\neq i(x_{j}=y_{j}) \land x_{i}<y_{i}] \to \langle x_{1},\ldots,x_{k}\rangle< \langle y_{1},\ldots,y_{k}\rangle$$ Now given a $\Delta^{0}_{0}$-formula $\varphi(x_{1},\ldots,x_{k},x_{k+1})$ without parameters, we want to show $L\Sigma_{1}^{-(k)}$ with $\theta\equiv \exists x_{k+1}\varphi$. As in [@Lev] some elementary functions $g_{1}, g, h, h'$ are defined successively as follows. $$g_{1}(n)= \left\{ \begin{array}{ll} n & \mbox{if } \forall y\leq n \lnot\varphi((y)_{1},\ldots,(y)_{k},(y)_{k+1}) \\ \langle (y)_{1},\ldots,(y)_{k}\rangle & \mbox{otherwise with } y=\mu y\leq n \varphi((y)_{1},\ldots,(y)_{k},(y)_{k+1}) \end{array} \right.$$ $$g(n)= \left\{ \begin{array}{ll} \langle (n)_{1},\ldots,(n)_{k}\rangle & \mbox{if } \exists u\leq (n)_{k+1} \varphi((n)_{1},\ldots,(n)_{k}, u) \\ g_{1}(n) & \mbox{otherwise} \end{array} \right.$$ $h(0)=g(0)$ and $$h(n+1)= \left\{ \begin{array}{ll} g(n+1) & \mbox{if } \forall k,m\leq n(k\neq m \to g(k)\neq g(m)) \\ g(n+1) & \mbox{if } \exists m\leq n (g(n+1)=g(m)) \mbox{ and } g(n+1)<^{(k)}h(n) \\ h(n) & \mbox{otherwise} \end{array} \right.$$ Observe that $h(n)\leq\max\{g(m):m\leq n\}$, and hence $h$ is elementary. $$h'(x)= \left\{ \begin{array}{ll} h(x) & \mbox{ if } \exists n\leq x \varphi((n)_{1},\ldots,(n)_{k}, (n)_{k+1}) \\ 0 & \mbox{otherwise} \end{array} \right.$$ Then [EA]{} proves that $h'$ is eventually decreasing with respect to $<^{(k)}$: $\exists m\forall n\geq m (h'(n+1)\leq^{(k)}h'(n))$. Therefore $h'_{1}(n)=(h'(n))_{1}$ is eventually decreasing. Hence $\exists y_{1}[y_{1}=\lim_{x\to\infty}h_{1}'(x)]$ in $(LimR)^{(1)}$. This in turn implies that $\langle (h'(n))_{2},\ldots,(h'(n))_{k}\rangle$ is eventually decreasing with respect to $<^{(k-1)}$. Therefore $h'_{2}(n)=(h'(n))_{2}$ is eventually decreasing demonstrably in $(LimR)^{(1)}$. Hence $\exists y_{2}[y_{2}=\lim_{x\to\infty}h_{2}'(x)]$ in $(LimR)^{(2)}$, and so on. Therefore $\exists y[y=\lim_{x\to\infty}h'(x)]$ in $(LimR)^{(k)}$. Now assuming $\exists x_{1}\cdots\exists x_{k}\exists x_{k+1} \theta(x_{1},\ldots,x_{k}, x_{k+1})$, we see as in [@Lev] that $y=\lim_{x\to\infty}h'(x)=\lim_{x\to\infty}h(x)$, and the limit $y$ is the minimum of $\{\langle x_{1},\ldots,x_{k}\rangle: \exists x_{k+1} \theta(x_{1},\ldots,x_{k}, x_{k+1})\}$ with respect to the lexicographic order $<^{(k)}$ as desired.\ \ Next assume by IH that $$L\Sigma_{1}^{-(k)}\vdash \exists m\forall n\geq m\, h(n+1)\leq h(n) .$$ We need to show $$L\Sigma_{1}^{-(k+1)}\vdash \exists m\forall n\geq m \, h(n)=h(m) .$$ For simplicity consider the case $k=1$, and assume that [EA]{} proves that $$\begin{aligned} && \{\exists x_{1}\varphi_{1}(x_{1})\to\exists x_{1}[\varphi_{1}(x_{1}) \land \forall z<x_{1}\lnot\varphi_{1}(z)]\} \land \\ && \{\exists x_{2}\varphi_{2}(x_{2})\to\exists x_{2}[\varphi_{2}(x_{2}) \land \forall z<x_{2}\lnot\varphi_{2}(z)]\} \\ && \to \exists m\forall n\geq m\, h(n+1)\leq h(n)\end{aligned}$$ Let $\varphi_{i}(x_{i})\equiv \exists y \theta_{i}(x_{i},y)$. By the Herbrand’s Theorem there exists a sequence of terms $m_{0}(a_{1},a_{2},b_{1},b_{2})$, $m_{1}(x_{0},a_{1},a_{2},b_{1},b_{2})$, $m_{2}(x_{0},x_{1},a_{1},a_{2},b_{1},b_{2}),\ldots,m_{k}(x_{0},\ldots,x_{k-1},a_{1},a_{2},b_{1},b_{2})$ such that the following disjunction is provable in [EA]{}: $$\begin{aligned} && \{\exists x_{1}\varphi_{1}(x_{1})\land [\lnot\theta_{1}(a_{1},b_{1}) \lor \exists z<a_{1} \varphi_{1}(z)]\} \lor \label{eq:dis} \\ && \{\exists x_{2}\varphi_{2}(x_{2})\land [\lnot\theta_{2}(a_{2},b_{2}) \lor \exists z<a_{2} \varphi_{2}(z)]\} \lor \nonumber \\ && (x_{0}\geq m_{0}(a_{1},a_{2},b_{1},b_{2}) \to h(x_{0})\geq h(x_{0}+1))\lor \nonumber \\ && (x_{1}\geq m_{1}(x_{0}, a_{1},a_{2},b_{1},b_{2}) \to h(x_{1})\geq h(x_{1}+1))\lor \nonumber \\ && \cdots \nonumber \\ && (x_{k}\geq m_{k}(x_{0}, \ldots,x_{k-1},a_{1},a_{2},b_{1},b_{2}) \to h(x_{k})\geq h(x_{k}+1)) \nonumber \end{aligned}$$ First assume $\exists x_{1}\varphi_{1}(x_{1})\land\exists x_{2}\varphi_{2}(x_{2})$, and pick a minimal $a=\langle a_{1},a_{2}\rangle$ such that $\varphi_{1}(a_{1})\land\varphi_{2}(a_{2})$ and $\forall b<a\lnot[\varphi_{1}((b)_{1})\land\varphi_{2}((b)_{2})]$. Then from (\[eq:lex\]) we see that $\forall z<a_{i}\lnot\varphi_{i}(z)$ for $i=1,2$. Now let $$y_{0}=\mu y_{0}[\exists b_{1}, b_{2}\exists x_{0}\geq m_{0}(a_{1},a_{2},b_{1},b_{2}) \{\theta_{1}(a_{1},b_{1}) \land \theta_{2}(a_{2},b_{2}) \land h(x_{0})=y_{0}\}]$$ by $L\Sigma_{1}^{-(2)}$. If $\forall x_{0}\geq m_{0}(a_{1},a_{2},b_{1},b_{2})(h(x_{0})\geq h(x_{0}+1))$ for some $b_{1},b_{2}$ with $\theta_{1}(a_{1},b_{1}) \land \theta_{2}(a_{2},b_{2})$, then $y_{0}=\lim_{x\to\infty}h(x)$. Otherwise let $$\begin{aligned} && y_{1}=\mu y_{1}[\exists b_{1}, b_{2}\exists x_{0}\geq m_{0}(a_{1},a_{2},b_{1},b_{2}) \exists x_{1}\geq m_{1}(x_{0}, a_{1},a_{2},b_{1},b_{2}) \\ && (\theta_{1}(a_{1},b_{1}) \land \theta_{2}(a_{2},b_{2}) \land h(x_{0})< h(x_{0}+1) \land h(x_{1})=y_{1})]\end{aligned}$$ If $\forall x_{1}\geq m_{1}(x_{0}, a_{1},a_{2},b_{1},b_{2}) (h(x_{1})\geq h(x_{1}+1))$ for some $b_{1},b_{2}$ with $\theta_{1}(a_{1},b_{1}) \land \theta_{2}(a_{2},b_{2})$, and an $x_{0}\geq m_{0}(a_{1},a_{2},b_{1},b_{2})$, then $y_{1}=\lim_{x\to\infty}h(x)$, and so on. If $\lnot\exists x_{1} \varphi_{1}(x_{1})$, then substitute $0$ for $a_{1}, b_{1}$ in (\[eq:dis\]). [99]{} T. Arai and G. Mints, Extended normal form theorems for logical proofs from axioms, Theor. Comp. Sci. 232 (2000), 121-132. L. D. Beklemishev and A. Visser, On the limit existence principles in elementary arithmetic and $\Sigma^{0}_{n}$-consequences of theories, Ann. Pure Appl. Logic 136(2005), 56-74. Y. l. Ershov, A certain hierarchy of sets. II. Algebra i Logika 7(1968), 15-47. P. Hájek and P. Pudlák, Metamathematics of First Order Arithmetic, Springer, 1993. R. Kaye, J. Paris and C. Dimitracopoulos, On parameter free induction schemas, Jour. Symb. Logic 53(1988), 1082-1097. G.E. Mints, Finite investigations of transfinite derivations, in: Selected Papers in Proof Theory (Bibliopolis, Napoli, 1992) 17-72. H. Putnam, Trial and error predicates and the solution to a problem of Mostowski, Jour. Symb. Logic 30(1965), 51-57. H. E. Rose, Subrecursion:Functions and hierarchies. Oxford Logic Guides 9. Oxford University Press 1984. Th. Skolem, Proof of some theorems on recursively enumerable sets. Notre Dame J. of Formal Logic 3, 65-74 (1962). R. Sommer, Transfinite inducton within Peano arithmetic, Ann. Pure Appl. Logic 76 (1995), 231-289. F. Stephan, Y. Yang and L. Yu, Turing degrees and the Ershov hierarchy, In: T. Arai, J. Brendle, C. T. Chong, R. Downey, Q. Feng, H. Kikyo and H. Ono (eds.), Proceedings of the 10th Asian Logic Conference, 2009, World Scientific, Singapore, pp. 300-321. [^1]: Dedicated to the occasion of Chong Chi Tat’s 60th birthday
{ "pile_set_name": "ArXiv" }
--- bibliography: - 'Thesis.bib' ---
{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $F$ be a number field, and ${{\mathbb {A}}}={{\mathbb {A}}}_F$. In this paper, first, we provide a family of global Arthur parameters confirming all parts of a general conjecture on the relation between the structure of Fourier coefficients and the structure of global Arthur parameters, given by Jiang in 2012. Then we extend a correspondence between certain automorphic forms on $Sp_{4n}({{\mathbb {A}}})$ and ${\widetilde}{Sp}_{2n}({{\mathbb {A}}})$, given by Ginzburg, Jiang and Soudry in 2012, to certain automorphic forms on $Sp_{4mn}({{\mathbb {A}}})$ and ${\widetilde}{Sp}_{4mn \pm 2n}({{\mathbb {A}}})$, using the same idea of considering compositions of automorphic descent maps.' address: | Department of Mathematics\ University of Utah\ 155 S 1400 E Room 233, Salt Lake City, UT 84112-0090, USA. author: - B L date: 'August, 2013' title: 'OG-J-S $Sp_{4mn}({{\mathbb {A}}})$ ${\widetilde}{Sp}_{4mn \pm 2n}({{\mathbb {A}}})$' --- Introduction ============ Let $F$ be a number field, and ${{\mathbb {A}}}={{\mathbb {A}}}_F$. Fourier coefficients play important roles in the study automorphic forms. For example, a basic and fundamental result in the theory of automorphic forms for $GL_n({{\mathbb {A}}})$ is that cuspidal automorphic forms are globally generic, that is, have non-vanishing Whittaker-Fourier coefficients, due to Shalika ([@S74]) and Piatetski-Shapiro ([@PS79]) independently. This result has been extended to the discrete spectrum of $GL_n({{\mathbb {A}}})$ in [@JL13a]. As another example, for classical groups, Ginzburg, Rallis and Soudry ([@GRS11]) developed the theory of automorphic descent by studying certain Fourier coefficients of special type residual representations, producing the inverse of special cases of Langlands functorial transfers from classical groups to the general linear groups. The idea of automorphic descent has been generalized to explicit constructions of endoscopy transfers for classical groups which can be found in [@G12] and [@J12]. For general connected reductive groups, there is a framework of attaching Fourier coefficients to nilpotent orbits (for symplectic groups, see [@GRS03], [@JL13b]). For classical groups $G$, nilpotent orbits are parametrized by partitions and certain non-degenerate quadratic forms ([@W01]). For any irreducible automorphic representation $\pi$ of $G({{\mathbb {A}}})$, let $\mathfrak{p}^m(\pi)$ be the set of maximal partitions (under the natural ordering of partitions) providing non-vanishing Fourier coefficients for $\pi$ (for precise definition, see [@J12] and [@JL13b]). On the other hand, given a general connected reductive group $G$, a main theme in the theory of automorphic forms is to study the discrete spectrum, which consists of cuspidal spectrum and residual spectrum. For $G=GL_n$, the residual spectrum was constructed explicitly by Moeglin and Waldspurger ([@MW89]). For symplectic and special orthogonal groups, the discrete spectrum was classified by Arthur ([@Ar13]) up to automorphic $L^2$-packets parametrized by global Arthur parameters. Towards understanding the Fourier coefficients information of members in the automorphic $L^2$-packets, Jiang ([@J12]) made a conjecture, which relates the structure of the global Arthur parameter of an irreducible automorphic representation $\pi$ which is in the discrete spectrum, to the structure of $\mathfrak{p}^m(\pi)$. We recall the symplectic case of this conjecture as follows. Let $G_n = Sp_{2n}$, with symplectic form $$\begin{pmatrix} 0 & v_n\\ -v_n & 0 \end{pmatrix},$$ where $v_n$ is an $n \times n$ matrix with $1$’s on the second diagonal and $0$’s elsewhere. Fix a Borel subgroup $B=TU$, where the maximal torus $T$ consists of elements of the following form $$\diag(t_1,\cdots,t_n;t_n^{-1},\cdots,t_1^{-1})$$ and the unipotent radical $U$ consists of all upper unipotent matrices in $G_n$. The set of global Arthur parameters for the discrete spectrum of $G_n=Sp_{2n}$ is denoted by ${\widetilde}{\Psi}_2(Sp_{2n})$, the elements of which are of the form $$\label{psin} \psi:=\psi_1\boxplus\psi_2\boxplus\cdots\boxplus\psi_r,$$ where $\psi_i$ are pairwise different simple global Arthur parameters of orthogonal type and have the form $\psi_i=(\tau_i,b_i)$ with $\tau_i\in{{\mathcal {A}}}_{{\mathrm{cusp}}}(a_i)$, $2n+1 = \sum_{i=1}^r a_ib_i$ (the dual group of $Sp_{2n}$ is $SO_{2n+1}({{\mathbb {C}}})$), and $\prod_i \omega_{\tau_i}^{b_i} = 1$ (the condition on the central characters of the parameter), following [@Ar13 Section 1.4]. More precisely, $\psi_i=(\tau_i,b_i)$ satisfies the following conditions: if $\tau_i$ is of symplectic type (i.e., $L(s, \tau_i, \wedge^2)$ has a pole at $s=1$), then $b_i$ is even; if $\tau_i$ is of orthogonal type (i.e., $L(s, \tau_i, {{\mathrm{Sym}}}^2)$ has a pole at $s=1$), then $b_i$ is odd. For each global Arthur parameter $\psi\in{\widetilde}{\Psi}_2(Sp_{2n})$ there defines a global Arthur packet ${\widetilde}{\Pi}_\psi$. The discrete spectrum of $Sp_{2n}({{\mathbb {A}}})$ has the following decomposition $$L^2_{{\mathrm{disc}}}(Sp_{2n}(F)\bks Sp_{2n}({{\mathbb {A}}})) \cong\oplus_{\psi\in{\widetilde}{\Psi}_2(Sp_{2n})}\oplus_{\pi\in{\widetilde}{\Pi}_\psi(\epsilon_\psi)}m_\psi\pi,$$ where ${\widetilde}{\Pi}_\psi(\epsilon_\psi)$ denotes the subset of ${\widetilde}{\Pi}_\psi$ consisting of members which occur in the discrete spectrum. As in [@J12], ${\widetilde}{\Pi}_\psi(\epsilon_\psi)$ is called the automorphic $L^2$-packet attached to $\psi$. For $\psi$ of the form in , let $\underline{p}(\psi)=[(b_1)^{(a_1)}\cdots(b_r)^{(a_r)}]$. For $\pi\in{\widetilde}{\Pi}_\psi(\epsilon_\psi)$, the structure of the global Arthur parameter $\psi$ deduces constraints on the structure of $\frak{p}^m(\pi)$, which is given by the following conjecture. \[cubmfc\] For any $\psi\in{\widetilde}{\Psi}_2(Sp_{2n})$, let ${\widetilde}{\Pi}_{\psi}(\epsilon_\psi)$ be the automorphic $L^2$-packet attached to $\psi$. Then the following hold. 1. Any symplectic partition ${\underline}{p}$ of $2n$, if ${\underline}{p}>\eta_{{\frak{g}^\vee,\frak{g}}}(\underline{p}(\psi))$, does not belong to $\frak{p}^m(\pi)$ for any $\pi\in{\widetilde}{\Pi}_{\psi}(\epsilon_\psi)$. 2. For a $\pi\in{\widetilde}{\Pi}_{\psi}(\epsilon_\psi)$, any partition ${\underline}{p}\in\frak{p}^m(\pi)$ has the property that ${\underline}{p}\leq \eta_{{\frak{g}^\vee,\frak{g}}}({\underline}{p}(\psi))$. 3. There exists at least one member $\pi\in{\widetilde}{\Pi}_{\psi}(\epsilon_\psi)$ having the property that $\eta_{{\frak{g}^\vee,\frak{g}}}({\underline}{p}(\psi))\in \frak{p}^m(\pi)$. Here $\eta_{{\frak{g}^\vee,\frak{g}}}$ denotes the Barbasch-Vogan duality map $($see [@BV85 Definition A1] and [@Ac03 Section 3.5]$)$ from the partitions for the dual group $G^\vee({{\mathbb {C}}})$ to the partitions for $G$. We refer to [@J12 Section 4] for more discussion on this conjecture and related topics. Part (1) of Conjecture \[cubmfc\] is completely proved in [@JL13c]. In the first part of this paper, we provide a family of global Arthur parameters confirming all parts of Conjecture \[cubmfc\]. Let $\tau$ be an irreducible unitary cuspidal automorphic representation of $GL_{2n}({{\mathbb {A}}})$, with the properties that $L(s, \tau, \wedge^2)$ has a simple pole at $s=1$, and $L(\frac{1}{2}, \tau) \neq 0$. By Theorem 2.3 of [@GJS12], there is an irreducible representation ${\widetilde}{\pi}$ of ${\widetilde}{Sp}_{2n}({{\mathbb {A}}})$, which is $\psi^{1}$-generic, lifts weakly to $\tau$ with respect to $\psi$. Let $\Delta(\tau, m)$ be a Speh representation in the discrete spectrum of $GL_{2mn}({{\mathbb {A}}})$. For more information about Speh representations, we refer to [@MW89], or the Section 1.2 of [@JLZ12]. Let $P_r=M_rN_r$ be the maximal parabolic subgroup of $Sp_{2l}$ with Levi subgroup $M_r$ isomorphic to $GL_r \times Sp_{2l-2r}$. Using the normalization in [@Sh10], the group $X_{M_{r}}^{Sp_{2l}}$ of all continues homomorphisms from $M_{r}({{\mathbb {A}}})$ to ${{\mathbb {C}}}^{\times}$, which is trivial on $M_{r}({{\mathbb {A}}})^1$ (see [@MW95]), will be identified with ${{\mathbb {C}}}$ by $s \rightarrow \lambda_s$. Let ${\widetilde}{P_r}({{\mathbb {A}}})$ be the pre-image of $P_r({{\mathbb {A}}})$ in ${\widetilde}{Sp}_{2l}({{\mathbb {A}}})$. For any $\phi \in A(N_{2mn}({{\mathbb {A}}})M_{2mn}(F) {\backslash}Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,m)}$, following [@L76] and [@MW95], an residual Eisenstein series can be defined by $$E(\phi,s)(g)=\sum_{\gamma\in P_{2mn}(F)\bks Sp_{4mn}(F)}\lambda_s \phi(\gamma g).$$ It converges absolutely for real part of $s$ large and has meromorphic continuation to the whole complex plane ${{\mathbb {C}}}$. By [@JLZ12], this Eisenstein series has a simple pole at $\frac{m}{2}$, which is the right-most one. Denote the representation generated by these residues at $s=\frac{m}{2}$ by ${{\mathcal {E}}}_{\Delta(\tau, m)}$. This residual representation is square-integrable. By Section 6.2 of [@JLZ12], the global Arthur parameter of ${{\mathcal {E}}}_{\Delta(\tau,m)}$ is $\psi=(\tau, 2m) \boxplus (1_{GL_1}, 1)$. Our first main result can be stated as follows. \[main1\] Assume that $F$ is any number field. $$\mathfrak{p}^m({{\mathcal {E}}}_{\Delta(\tau, m)}) = [(2n)^{2m}].$$ This theorem was discussed by Ginzburg in [@G08] with a quite sketchy argument. A fully detailed proof will be given in Section 2. By Theorem 1.3, Proposition 6.4 and Remark 6.5 of [@JL13c], Parts (1) and (2) of Conjecture \[cubmfc\] hold for these global Arthur parameters $\psi=(\tau, 2m) \boxplus (1_{GL_1}, 1)$. Note that in this case $$\eta_{{so_{2n+1}({{\mathbb {C}}}),sp_{2n}({{\mathbb {C}}})}}({\underline}{p}(\psi)) =\eta_{so_{2n+1}({{\mathbb {C}}}),sp_{2n}({{\mathbb {C}}})}([(2m)^{2n}(1)])= [(2n)^{2m}].$$ Combining Theorem \[main1\], we can see that all parts of Conjecture \[cubmfc\] are confirmed for this family of global Arthur parameters. In [@GRS03], for any irreducible cuspidal automorphic representation $\pi$ of symplectic groups or their double covers, Ginzburg, Rallis and Soudry found a maximal partition which has only even parts, providing non-vanishing Fourier coefficients for $\pi$. We denote this partition by ${\underline}{p}(\pi)$. Next, we assume that $F$ is not totally imaginary, and consider ${\overline}{{{\mathcal {N}}}}_{Sp_{4mn}}$, the set of irreducible cuspidal automorphic representations $\pi$ which are nearly equivalent to ${{\mathcal {E}}}_{\Delta(\tau, m)}$ and $${\underline}{p}(\pi) = [(2n)^{2m-1}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}],$$ with $2n \geq 2n_1 > 2n_2 > \cdots > 2n_k$, $k \geq 1$. Note that these partitions are less than or equal to $[(2n)^{2m}]$. ${\overline}{{{\mathcal {N}}}}_{Sp_{4mn}}$ can be naturally decomposed into a disjoint union of two sets ${{\mathcal {N}}}_{Sp_{4mn}} \cup {{\mathcal {N}}}'_{Sp_{4mn}}$, where ${{\mathcal {N}}}_{Sp_{4mn}}$ consists of elements having a nonzero Fourier coefficient $FJ_{\psi^{-1}_{n-1}}$ (for definition, see [@GRS11 Section 3.2]), while ${{\mathcal {N}}}'_{Sp_{4mn}}$ consists of elements having no nonzero Fourier coefficients $FJ_{\psi^{-1}_{n-1}}$. For any ${\widetilde}{\phi} \in A(N_{2kn}({{\mathbb {A}}}){\widetilde}{M}_{2kn}(F) {\backslash}{\widetilde}{Sp}_{4kn+2n}({{\mathbb {A}}}))_{\mu_{\psi} \Delta(\tau,k) \otimes {\widetilde}{\pi}}$, following [@L76] and [@MW95], an residual Eisenstein series can be defined by $${\widetilde}{E}({\widetilde}{\phi},s)(g)=\sum_{\gamma\in P_{2kn}(F)\bks Sp_{4kn+2n}(F)}\lambda_s {\widetilde}{\phi}(\gamma g).$$ It converges absolutely for real part of $s$ large and has meromorphic continuation to the whole complex plane ${{\mathbb {C}}}$. By similar argument as that in [@JLZ12], this Eisenstein series has a simple pole at $\frac{k+1}{2}$, which is the right-most one. Denote the representation generated by these residues at $s=\frac{k+1}{2}$ by ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, k) \otimes {\widetilde}{\pi}}$. This residual representation is square-integrable. Let ${{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$ be the set of irreducible genuine cuspidal automorphic representations ${\widetilde}{\sigma}_{4(m-1)n+2n}$ of ${\widetilde}{Sp}_{4(m-1)n+2n}({{\mathbb {A}}})$, which are nearly equivalent to the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, m-1) \otimes {\widetilde}{\pi}}$, have no nonzero Fourier coefficients $FJ_{\psi^{1}_{n-1}}$, and $${\underline}{p}({\widetilde}{\sigma}_{4(m-1)n+2n})=[(2n)^{2(m-1)}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}],$$ with $2n \geq 2n_1 > 2n_2 > \cdots > 2n_k$, $k \geq 1$. Let ${{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn+2n}}(\tau, \psi)$ be the set of irreducible genuine cuspidal automorphic representations ${\widetilde}{\sigma}_{4mn+2n}$ of ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$, which are nearly equivalent to the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, m) \otimes {\widetilde}{\pi}}$, have a nonzero Fourier coefficient $FJ_{\psi^{1}_{n-1}}$, and $${\underline}{p}({\widetilde}{\sigma}_{4mn+2n})=[(2n)^{2m}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}],$$ with $2n \geq 2n_1 > 2n_2 > \cdots > 2n_k$, $k \geq 1$. For any ${\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, for any $${\widetilde}{\phi} \in A(N_{2mn}({{\mathbb {A}}}){\widetilde}{M}_{2mn}(F) {\backslash}{\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}}))_{\mu_{\psi} \tau \otimes {\widetilde}{\sigma}_{4(m-1)n+2n}},$$ by similar calculation as in Pages 996-997 of [@GJS12], it is easy to see that the corresponding Eisenstein series has a simple pole at $s=m$. Let ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$ be the residual representation of ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$ generated by the corresponding residues. This residual representation is square-integrable. For any $\sigma_{4mn} \in {{\mathcal {N}}}'_{{Sp}_{4mn}}(\tau, \psi)$, for any $${\phi} \in A(N_{2mn}({{\mathbb {A}}}){M}_{2mn}(F) {\backslash}{Sp}_{4mn+2n}({{\mathbb {A}}}))_{\tau \otimes {\sigma}_{4mn}},$$ also by similar calculation as in Pages 996-997 of [@GJS12], it is easy to see that the corresponding Eisenstein series has a simple pole at $s=\frac{2m+1}{2}$. Let ${{\mathcal {E}}}_{\tau, \sigma_{4mn}}$ be the residual representation of $Sp_{4(m+1)n}({{\mathbb {A}}})$ generated by the corresponding residues. This residual representation is square-integrable. For any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, let ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$ be the $\psi^{-1}$-descent of $\sigma_{4mn}$ from ${Sp}_{4mn}({{\mathbb {A}}})$ to ${\widetilde}{Sp}_{4(m-1)n+2n}({{\mathbb {A}}})$ (defined in Chapter 3 of [@GRS11]). Note that by the tower property (see Theorem 7.10 of [@GRS11]), ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$ is cuspidal. For any ${\widetilde}{\sigma}_{4mn+2n} \in {{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn+2n}}(\tau, \psi)$, let ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{\sigma}_{4mn+2n})$ be the $\psi^{1}$-descent of ${\widetilde}{\sigma}_{4mn+2n}$ from ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$ to ${Sp}_{4mn}({{\mathbb {A}}})$ (defined in Chapter 3 of [@GRS11]). Note that by the tower property (see Theorem 7.10 of [@GRS11]), ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{\sigma}_{4mn+2n})$ is also cuspidal. Our second main result is that there are correspondences between ${{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$ and ${{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, and between ${{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn+2n}}(\tau, \psi)$ and ${{\mathcal {N}}}'_{{Sp}_{4n}}(\tau, \psi)$, as follows. \[thm1\] Assume that $F$ is a number field which is not totally imaginary. \(1) There is a surjective map $$\Psi: {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi) \rightarrow {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$$ $$\sigma_{4mn} \mapsto {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}).$$ \(2) If for any ${\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$ is irreducible, then $\Psi$ is also injective. Note that the case of $m=1$ has already been proved by Ginzburg, Jiang and Soudry in [@GJS12]. We use the same idea here to extend the correspondence to higher ranks. Among others, one key idea is to consider compositions of automorphic descent maps. Also note that they include ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ in the domain of the map $\Psi$. For simplicity, we just let the domain of the map $\Psi$ consist of only irreducible cuspidal automorphic representations, and consider the descent of ${{\mathcal {E}}}_{\Delta(\tau, m)}$ separately in Section 7. \[thm2\] Assume that $F$ is a number field which is not totally imaginary. \(1) There is a surjective map $$\Psi: {{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn+2n}}(\tau, \psi) \rightarrow {{\mathcal {N}}}'_{{Sp}_{4n}}(\tau, \psi)$$ $${\widetilde}{\sigma}_{4mn+2n} \mapsto {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{\sigma}_{4mn+2n}).$$ \(2) If for any $\sigma_{4mn} \in {{\mathcal {N}}}'_{{Sp}_{4mn}}(\tau, \psi)$, ${{\mathcal {E}}}_{\tau, {\sigma}_{4mn}}$ is irreducible, then $\Psi$ is also injective. Due to the similarity of the proofs of Theorem \[thm1\] and Theorem \[thm2\], we only give the proof for Theorem \[thm1\]. Theorem \[thm1\] and Theorem \[thm2\] together give us the following diagram about correspondences between various sets of irreducible cuspidal automorphic representations: \[fig1\] (60,36) (4,0)[$\vdots$]{} (0,8)[${\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}})$]{} (0,16)[$Sp_{4mn}({{\mathbb {A}}})$]{} (0,24)[${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$]{} (4,32)[$\vdots$]{} (16.2,0)[$\vdots$]{} (16,4)[$\downarrow$]{} (18,4)[${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn-2n}$]{} (12,8)[${\widetilde}{{{\mathcal {N}}}}_{4mn-2n}$]{} (21,8)[$\dot\bigcup$]{} (24,8)[${\widetilde}{{{\mathcal {N}}}}_{4mn-2n}'$]{} (28,12)[$\downarrow$]{} (30,12)[${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}$]{} (26,16)[${{\mathcal {N}}}_{4mn}$]{} (34,16)[$\dot\bigcup$]{} (38,16)[${{\mathcal {N}}}_{4mn}'$]{} (40,20)[$\downarrow$]{} (42,20)[${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}$]{} (36,24)[${\widetilde}{{{\mathcal {N}}}}_{4mn+2n}$]{} (45,24)[$\dot\bigcup$]{} (48,24)[${\widetilde}{{{\mathcal {N}}}}_{4mn+2n}'$]{} (52,28)[$\downarrow$]{} (54,28)[${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn+4n}$]{} (52.2,32)[$\vdots$]{} In the above diagram, for short, we write that ${{\mathcal {N}}}_{{4mn}} := {{\mathcal {N}}}_{Sp_{4mn}}$, ${{\mathcal {N}}}_{{4mn}}' := {{\mathcal {N}}}_{Sp_{4mn}}'$, ${\widetilde}{{{\mathcal {N}}}}_{{4mn \pm 2n}} := {{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn \pm 2n}}$, ${\widetilde}{{{\mathcal {N}}}}_{{4mn \pm 2n}}' := {{\mathcal {N}}}_{{\widetilde}{Sp}_{4mn \pm 2n}}'$. \[rmk1\] In Theorem \[thm1\] and Theorem \[thm2\], we assume that $F$ is a number field which is not totally imaginary, the reason is that when $F$ is a totally imaginary number field, then our construction will “stop at some point, and can not go to higher levels". The explicit explanation of this phenomenon will appear elsewhere. From Theorem \[main1\], for the residual representation ${{\mathcal {E}}}_{\Delta(\tau, m)}$, we know that $\mathfrak{p}^m({{\mathcal {E}}}_{\Delta(\tau, m)}) = [(2n)^{2m}]$. From its proof, and by Lemma 2.6 of [@GRS03] or Lemma 3.1 of [@JL13b], we can see that it has a non-vanishing Fourier coefficient attached to the partition $[(2n)1^{4mn-2n}]$ with respect to the character $\psi_{[(2n)1^{4mn-2n}], -1}$. In Section 7, for any number field $F$, we show that both ${{\mathcal {E}}}_{\Delta(\tau, m)}$ and ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ are irreducible. The result can be stated as follows. \[irre2\] Assume that $F$ is any number field. \(1) ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ is square-integrable and is in the discrete spectrum. \(2) Both ${{\mathcal {E}}}_{\Delta(\tau, m)}$ and ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ are irreducible. Note that in general, it is difficult to prove the irreducibility of certain descent representations. The case of $m=1$ of Theorem \[irre2\] was proved in Theorem 4.1 of [@GJS12], noting that by Theorem 2.5 of [@GJS12] ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ is irreducible. Also note that, the irreducibility of ${{\mathcal {D}}}^{4n}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, 1)})$ actually has already been proved by Jiang and Soudry in [@JS03], using different methods. At the end of this introduction, we discuss the contents by section. In Section 2, we will show Theorem \[main1\], whose proof is reduced to that of Lemma \[constantterm\], which will be given in Section 3. The Section 4, we will prove Part (1) of Theorem \[thm1\], whose proof is reduced to that of Theorem \[thm6\], which will be given in Section 5. In Section 6, we completes the proof of Theorem \[thm1\], by proving its Part (2). In Section 7, we prove Theorem \[irre2\]. We assume that $F$ is not totally imaginary only in Sections 4–6. Acknowledgement. The author would like to take this opportunity to express his deepest gratitude to his advisor Prof. Dihua Jiang, for introducing the author to the topics of Fourier coefficients of automorphic forms, automorphic descent, and constructions of square-integrable automorphic representations, for sharing with the author his wonderful ideas and insights to various problems of mathematics, for his constant encouragement and support. The author also would like to thank Prof. David Soudry for very helpful conversations on related topics of automorphic descent. Proof of Theorem \[main1\] ========================== In this section, we prove Theorem \[main1\], which was discussed by Ginzburg in [@G08] with a quite sketchy argument. To be complete, we give the full details here. $F$ is any number field in this section, Sections 3 and 7. \[main1part2\] $$\mathfrak{p}^m({{\mathcal {E}}}_{\Delta(\tau, m)}) = [(2n)^{2m}].$$ By Theorem 1.3, Proposition 6.4 and Remark 6.5 of [@JL13c], we only have to show that ${{\mathcal {E}}}_{\Delta(\tau, m)}$ has a nonzero Fourier coefficient attached to $[(2n)^{2m}]$. We will prove this by induction on $m$. For $m=1$, this is proved in the book [@GRS11]. Note that when $m=1$, ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ has a nonzero Fourier coefficient attached to the partition $[(2n)1^{2n}]$, and the descent to $\widetilde{Sp}_{2n}$ is generic (see Theorem 3.1 of [@GRS11]). Therefore, ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ has a nonzero Fourier coefficient attached to the composite partition $[(2n)1^{2n}] \circ [(2n)]$, which implies that ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ has a nonzero Fourier coefficient attached to the partition $[(2n)^{2}]$ by Lemma 2.6 of [@GRS03], or Lemma 3.1 of [@JL13b]. For definition of composite partitions, we refer to Section 1 of [@GRS03]. Now we assume that the theorem is true for the case of $m-1$, and consider the case of $m \geq 2$. Take any $\varphi \in {{\mathcal {E}}}_{\Delta(\tau, m)}$, its Fourier coefficients attached to ${\underline}{p} = [(2n)^{2m}]$ are of the following forms $$\label{main1equ1} \varphi^{\psi_{{\underline}{p}, {\underline}{a}}}(g) = \int_{[V_{{\underline}{p},2}]} \varphi(vg) \psi_{{\underline}{p}, {\underline}{a}}^{-1}(v)dv,$$ where ${\underline}{a} = \{a_1, a_2, \ldots, a_{2m}\} \subset (F^*/ (F^*)^2)^{2m}$. For definitions of the unipotent group $V_{{\underline}{p},2}$ and its character $\psi_{{\underline}{p}, {\underline}{a}}$, see Section 2 of [@JL13b]. Assume that $T$ is the maximal split torus in $Sp_{4mn}$, consists of elements $$\diag(t_1, t_2, \ldots, t_{2mn}, t_{2mn}^{-1}, \ldots, t_2^{-1}, t_1^{-1}).$$ Let $\omega_1$ be the Weyl element of $Sp_{4mn}$, sending elements $t \in T$ to the following torus elements: $$\label{main1equ2} t'=\diag(t^{(0)}, t^{(1)}, t^{(2)}, \ldots, t^{(n)}, t^{(n),*}, \ldots, t^{(2),*}, t^{(1),*}, t^{(0),*}),$$ where $$\begin{aligned} t^{(0)} = & \diag(t_1, t_{n+1}, t_2, t_{n+2}, \ldots, t_i, t_{n+i}, \ldots, t_n, t_{2n})\\ t^{(j)} = & \diag(t_{2n+j}, t_{3n+j}, \ldots, t_{in+j}, \ldots, t_{(2m-1)n+j}),\end{aligned}$$ for $1 \leq j \leq n$. Identify $Sp_{4(m-1)n}$ with its image in $Sp_{4mn}$ under the embedding $g \mapsto \diag(I_{2n}, g, I_{2n})$. Denote the restriction of $\omega_1$ to $Sp_{4(m-1)n}$ by $\omega_1'$. Conjugating cross by $\omega_1$, the Fourier coefficient $\varphi^{\psi_{\underline{p},\underline{a}}}$ becomes: $$\label{main1equ3} \int_{[U_{{\underline}{p},2}]} \varphi(u \omega_1 g) \psi_{\underline{p},\underline{a}}^{\omega_1}(u)^{-1} du,$$ where $U_{{\underline}{p},2} = \omega_1 V_{{\underline}{p},2} \omega_1^{-1}$, and $\psi_{\underline{p},\underline{a}}^{\omega_1}(u) = \psi_{\underline{p},\underline{a}}(\omega_1^{-1}u\omega_1)$. Now, we describe the structure of elements in $U_{{\underline}{p},2}$. Any element in $U_{\underline{p},2}$ has the following form: $$\label{main1equ4} u = \begin{pmatrix} z_{2^n} & q_1 & q_2\\ 0 & u' & q_1^*\\ 0 & 0 & z_{2^n}^* \end{pmatrix} \begin{pmatrix} I_{2n} & 0 & 0\\ p_1 & I_{(4m-4)n} & 0\\ 0 & p_1^* & I_{2n} \end{pmatrix},$$ where $z_{2^n} \in V_{2^n}$, the unipotent radical of the parabolic $Q_{2^n}$ of $GL_{2n}$ with Levi isomorphic to $GL_2^{\times n}$; $u' \in U_{[(2n)^{2m-2}],2} := \omega_1' V_{[(2n)^{2m-2}],2} \omega_1'^{-1}$; $q_1 \in M_{2n\times (4m-4)n}$, $p_1 \in M_{(4m-4)n \times 2n}$, satisfy certain conditions, which we do not specify at this moment; $q_2 \in M_{(2n)\times (2n)}$, such that $q_2^t v_{2n} - v_{2n} q_2 = 0$, where $v_{2n}$ is a matrix only with ones on the second diagonal. Note that $$\begin{aligned} & \psi_{\underline{p},\underline{a}}^{\omega_1} (\begin{pmatrix} z_{2^n} & q_1 & q_2\\ 0 & I_{(4m-4)n} & q_1^*\\ 0 & 0 & z_{2^n}^* \end{pmatrix}) \\ = & \psi(z_{2^n}(1,3)+\cdots + z_{2^n}(i,i+2) + \cdots + z_{2^n}(2n-2,2n))\\ \cdot & \psi(a_1 q_2(2n-1, 2) + a_2 q_2(2n,1)),\end{aligned}$$ where $a_1, a_2$ come from the ${\underline}{a} = \{a_1, a_2, \ldots, a_{2m}\}$ occurred in the Fourier coefficient $\varphi^{\psi_{\underline{p},\underline{a}}}$. Since to show that $\mathcal{E}_{\Delta(\tau, m)}$ has a nonzero Fourier coefficient attached to the partition ${\underline}{p} = [(2n)^{2m}]$, we only need to show that it has a nonzero Fourier coefficient $\varphi^{\psi_{\underline{p},\underline{a}}}$ for some ${\underline}{a}$, we consider the following special type of ${\underline}{a}$: $${\underline}{a} = \{1, -1, a_3, \ldots, a_{2m}\},$$ where $a_3, \ldots, a_{2m}$ are arbitrary elements in $F^*/(F^*)^2$. Let $A = \begin{pmatrix} 1 & -1\\ 1 & 1 \end{pmatrix}$, and $\epsilon=\diag(A, \ldots, A; I_{(4m-4)n}; A^*, \ldots, A^*)$, as in (2.31) of [@GJS12]. Conjugating cross the integral in by $\epsilon$, it becomes: $$\label{main1equ5} \int_{[U^{\epsilon}_{{\underline}{p},2}]} \varphi(u \epsilon \omega_1 g) \psi_{\underline{p},\underline{a}}^{\omega_1, \epsilon}(u)^{-1} du,$$ where $U^{\epsilon}_{{\underline}{p},2} = \epsilon U_{{\underline}{p},2} \epsilon^{-1}$ whose elements have the same structure as $U_{{\underline}{p},2}$ (see ), and $\psi_{\underline{p},\underline{a}}^{\omega_1,\epsilon}(u) = \psi_{\underline{p},\underline{a}}^{\omega_1}(\epsilon^{-1} u \epsilon)$. Note that now $$\begin{aligned} & \psi_{\underline{p},\underline{a}}^{\omega_1, \epsilon} (\begin{pmatrix} z_{2^n} & q_1 & q_2\\ 0 & I_{(4m-4)n} & q_1^*\\ 0 & 0 & z_{2^n}^* \end{pmatrix}) \\ = & \psi(z_{2^n}(1,3)+\cdots + z_{2^n}(i,i+2) + \cdots + z_{2^n}(2n-2,2n))\\ \cdot & \psi(q_2(2n-1, 1)).\end{aligned}$$ Note that the $a$ in (2.18) and (2.35) of [@GJS12] is $-1$ here. Let $\nu$ be the following Weyl element of ${{\mathrm{Sp}}}_{4n}$ which is defined on Page 14 of [@GJS12], also in (4.9) of [@GRS99]: $$\begin{aligned} \label{main1equ6} \begin{split} \nu_{i,2i-1} & = 1, i = 1, \ldots, 2n,\\ \nu_{2n+i,2i} & = -1, i = 1, \ldots, n,\\ \nu_{2n+i,2i} & = 1, i = n+1, \ldots, 2n,\\ \nu_{i,j} & = 0, \text{ otherwise}. \end{split}\end{aligned}$$ Write $\nu$ as $\begin{pmatrix} \nu_1 & \nu_2\\ \nu_3 & \nu_4 \end{pmatrix}$, where $\nu_i$’s are of size $2n \times 2n$. Let $\omega_2 = \begin{pmatrix} \nu_1 & & \nu_2\\ & I_{(4m-4)n} & \\ \nu_3 & & \nu_4 \end{pmatrix}$, a Weyl element of ${{\mathrm{Sp}}}_{4mn}$. Conjugating cross the integral in by $\omega_2$, it becomes: $$\label{main1equ7} \int_{[U^{\epsilon, \omega_2}_{{\underline}{p},2}]} \varphi(u \omega_2 \epsilon \omega_1 g) \psi_{\underline{p},\underline{a}}^{\omega_1, \epsilon, \omega_2}(u)^{-1} du,$$ where $U^{\epsilon, \omega_2}_{{\underline}{p},2} = \omega_2 U^{\epsilon}_{{\underline}{p},2} \omega_2^{-1}$, and $\psi_{\underline{p},\underline{a}}^{\omega_1,\epsilon, \omega_2}(u) = \psi_{\underline{p},\underline{a}}^{\omega_1, \epsilon}(\omega_2^{-1} u \omega_2)$. Now let’s describe the structure of elements in $U^{\epsilon, \omega_2}_{{\underline}{p},2}$. Any element in $U^{\epsilon, \omega_2}_{{\underline}{p},2}$ has the following form: $$\label{main1equ8} u = \begin{pmatrix} z_{2n} & q_1 & q_2\\ 0 & u' & q_1^*\\ 0 & 0 & z_{2n}^* \end{pmatrix} \begin{pmatrix} I_{2n} & 0 & 0\\ p_1 & I_{(4m-4)n} & 0\\ p_2 & p_1^* & I_{2n} \end{pmatrix},$$ where $z_{2n} \in V_{2n}$, the standard maximal unipotent subgroup of $GL_{2n}$ as before; $u' \in U_{[(2n)^{2m-2}],2} := \omega_1' V_{[(2n)^{2m-2}],2} \omega_1'^{-1}$; $q_1 \in M_{2n\times (4m-4)n}$, with $q_1(i,j) = 0$, for $j \leq (2m-2)i$; $q_2, p_2 \in M_{(2n)\times (2n)}$, with $q_2(i,j) = p_2(i,j) = 0$, for $j \leq i$; $p_1 \in M_{(4m-4)n \times 2n}$, with $p_1(i,j) = 0$, for $i \geq (2m-2)(j-1)$. Note that now $$\begin{aligned} \label{main1equ9} \begin{split} & \psi_{\underline{p},\underline{a}}^{\omega_1,\epsilon, \omega_2} (\begin{pmatrix} z_{2n} & q_1 & q_2\\ 0 & I_{(4m-4)n} & q_1^*\\ 0 & 0 & z_{2n}^* \end{pmatrix}) \\ = & \psi(z_{2n}(1,2)+\cdots + z_{2n}(n,n+1) \\ & - z_{2n}(n+1, n+2) - \cdots - z_{2n}(2n-1,2n)). \end{split}\end{aligned}$$ To proceed, we need to define some unipotent subgroups. Let $R^1_i = \prod_{j=1}^i R^1_{i,j}$, for $1 \leq i \leq n$, where $R^1_{i,j} = \prod_{s=1}^{2m-2} X_{\alpha^i_{j,s}}$, with $\alpha^i_{j,s} = e_i - e_{2n+(2m-2)(i-j+1)-s+1}$. Let $R^1_i = \prod_{j=1}^i R^1_{i,j}$, for $n+1 \leq i \leq 2n-1$, where $R^1_{i,j} = \prod_{s=1}^{2m-2} X_{\alpha^i_{j,s}}$, with $\alpha^i_{j,s} = e_i - e_{2n+(2m-2)(i-j+1)-s+1}$ if $j \geq i-n+1$, and $\alpha^i_{j,s} = e_i + e_{2mn-(2m-2)(i-n)+(2m-1)(j-1)+s}$ if $j \leq i-n$. Let $C^1_i = \prod_{j=1}^i C^1_{i,j}$, for $1 \leq i \leq n$, where $C^1_{i,j} = \prod_{s=1}^{2m-2} X_{\beta^i_{j,s}}$, with $\beta^i_{j,s} = e_{2n+(2m-2)(i-j+1)-s+1} - e_{i+1}$. Let $C^1_i = \prod_{j=1}^i C^1_{i,j}$, for $n+1 \leq i \leq 2n-1$, where $C^1_{i,j} = \prod_{s=1}^{2m-2} X_{\beta^i_{j,s}}$, with $\beta^i_{j,s} = e_{2n+(2m-2)(i-j+1)-s+1}-e_{i+1}$ if $j \geq i-n+1$, and $\beta^i_{j,s} = -e_{2mn-(2m-2)(i-n)+(2m-1)(j-1)+s}-e_{i+1}$ if $j \leq i-n$. Let $R^2_i = \prod_{j=1}^i X_{\alpha^i_j}$, for $1 \leq i \leq n$, with $\alpha^i_j = e_i+e_{2n-i+j}$, and $R^2_i = \prod_{j=1}^{2n-i} X_{\alpha^i_j}$, for $n+1 \leq i \leq 2n-1$, with $\alpha^i_j = e_i+e_{i+j}$. Let $C^2_i = \prod_{j=1}^i X_{\beta^i_j}$, for $1 \leq i \leq n$, with $\beta^i_j = -e_{2n-i+j}-e_{i+1}$, and $C^2_i = \prod_{j=1}^{2n-i} X_{\beta^i_j}$, for $n+1 \leq i \leq 2n-1$, with $\beta^i_j = -e_{i+j}-e_{i+1}$. Now, we are ready to apply Lemma 2.3 of [@JL13b] repeatedly to the integration over $\prod_{i=1}^{2n-1} C^2_i C^1_i$. We will consider the following pairs of groups: $$\begin{aligned} & (R^2_1 R^1_1, C^2_1 C^1_1), \ldots, (R^2_n R^1_n, C^2_n C^1_n);\\ & (R^2_{n+1}, C^2_{n+1}), (R^1_{n+1}, C^1_{n+1});\\ & \ldots;\\ & (R^2_{2n-1}, C^2_{2n-1}), (R^1_{2n-1}, C^1_{2n-1}). \end{aligned}$$ Let ${\widetilde}{U}^{\epsilon, \omega_2}_{{\underline}{p},2}$ be the subgroup of $U^{\epsilon, \omega_2}_{{\underline}{p},2}$ with $p_1$ and $p_2$-parts zero. Then, $U^{\epsilon, \omega_2}_{{\underline}{p},2} = {\widetilde}{U}^{\epsilon, \omega_2}_{{\underline}{p},2} \prod_{i=1}^{2n-1} C^2_i C^1_i$. Let $W = U^{\epsilon, \omega_2}_{{\underline}{p},2}$, ${\widetilde}{W} = {\widetilde}{U}^{\epsilon, \omega_2}_{{\underline}{p},2}$, and $\psi_W = \psi_{\underline{p},\underline{a}}^{\omega_1, \epsilon, \omega_2}$. First, we apply Lemma 2.3 of [@JL13b] to $(R^2_1 R^1_1, C^2_1 C^1_1)$. For this, we need to consider the quadruple $ ({\widetilde}{W} \prod_{i=2}^{2n-1} C^2_i C^1_i, \psi_W, X_{\alpha^1_1} R^1_{1,1}, X_{\beta^1_1} C^1_{1,1})$. It is easy to see that this quadruple satisfies all the conditions for this lemma. By this lemma, the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ10} \int_{[R^2_1 R^1_1 {\widetilde}{W} \prod_{i=2}^{2n-1} C^2_i C^1_i]} \varphi(rwc \omega_2 \epsilon \omega_1 g) \psi_W(w)^{-1} dcdwdr.$$ Note that $R^2_1 = X_{\alpha^1_1}$, and $R^1_1 = R^1_{1,1}$. Then we apply Lemma 2.3 of [@JL13b] to $(R^2_2 R^1_2, C^2_2 C^1_2)$. For this, we need to consider the following sequence of quadruples: $$\begin{aligned} & (R^2_1 R^1_1 {\widetilde}{W} \prod_{i=3}^{2n-1} C^2_iC^1_i X_{\beta^2_2}R^1_{2,2}, \psi_W, X_{\alpha^2_1}R^1_{2,1}, X_{\beta^2_1}C^1_{2,1}),\\ & (R^2_1 R^1_1 X_{\alpha^2_1} R^1_{2,1} {\widetilde}{W} \prod_{i=3}^{2n-1} C^2_iC^1_i, \psi_W, X_{\alpha^2_2}R^1_{2,2}, X_{\beta^2_2}C^1_{2,2}).\end{aligned}$$ Applying this lemma twice, we get that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ11} \int_{[R^2_1R^2_2R^1_1R^1_2 {\widetilde}{W} \prod_{i=3}^{2n-1} C^2_iC^1_i]} \varphi(rwc \omega_2 \epsilon \omega_1 g) \psi_W(w)^{-1} dcdwdr.$$ Then we continue the above procedure, applying Lemma 2.3 of [@JL13b] to pairs $(R^2_3 R^1_3, C^2_3 C^1_3), \ldots, (R^2_n R^1_n, C^2_n C^1_n)$. For the pair $(R^2_n R^1_n, C^2_n C^1_n)$, we need to consider the following sequence of quadruples: $$\begin{aligned} & (\prod_{i=1}^{n-1} R^2_i R^1_i {\widetilde}{W} \prod_{i=n+1}^{2n-1} C^2_iC^1_i \prod_{s=2}^{n} X_{\beta^n_s}R^1_{n,s}, \psi_W, X_{\alpha^n_1}R^1_{s,1}, X_{\beta^n_1}C^1_{n,1}),\\ & \cdots,\\ & (\prod_{i=1}^{n-1} R^2_i R^1_i \prod_{s=1}^{n-1} X_{\alpha^l_s}R^1_{n,s} {\widetilde}{W} \prod_{i=n+1}^{2n-1} C^2_iC^1_i, \psi_W, X_{\alpha^n_n}R^1_{n,n}, X_{\beta^n_n}C^1_{n,n}).\end{aligned}$$ Applying Lemma 2.3 of [@JL13b] repeatedly, we get that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ12} \int_{[\prod_{i=1}^{n} R^2_i R^1_i {\widetilde}{W} \prod_{i=n+1}^{2n-1} C^2_iC^1_i]} \varphi(rwc \omega_2 \epsilon \omega_1 g) \psi_W(w)^{-1} dcdwdr.$$ Before applying Lemma 2.3 of [@JL13b] to pairs $(R^2_s, C^2_s), (R^1_s, C^1_s)$, $s = n+1, n+2, \ldots, 2n-1$, we need to take Fourier expansion along the one-dimensional root subgroup $X_{e_s+e_s}$. And then we need to consider the pair $(R^2_s, C^2_s)$ first, then $(R^1_s, C^1_s)$. For example, for $s=n+1$, we first take the Fourier expansion of the integral in along the one-dimensional root subgroup $X_{e_s+e_s}$. Under the action of $GL_1$, we get two kinds of Fourier coefficients corresponding to the two orbits of the dual of $[X_{e_s+e_s}]$: the trivial one and the non-trivial one. For the Fourier coefficients attached to the non-trivial orbit, we can see that there is an inner integral $\varphi_{[(2n+2)1^{4mn-2n-2}],\{a\}}$, which is identically zero by Proposition 6.4 and Remark 6.5 of [@JL13c]. Therefore only the Fourier coefficient attached to the trivial orbit, which actually equals to the integral in , survives. Then, we can apply Lemma 2.3 of [@JL13b] to pairs $$(R^2_{n+1}, C^2_{n+1}), (R^1_{n+1}, C^1_{n+1}).$$ We need to consider the following sequence of quadruples: $$\begin{aligned} & (\prod_{i=1}^{n} R^2_i R^1_i {\widetilde}{W} X_{e_{n+1}+e_{n+1}} \prod_{i=n+2}^{2n-1} C^2_i \prod_{t=n+1}^{2n-1} C^1_t \prod_{s=2}^{n-1} X_{\beta^n_s}, \psi_W, X_{\alpha^{n+1}_1}, X_{\beta^{n+1}_1}),\\ & \cdots,\\ & (\prod_{i=1}^{n} R^2_i R^1_i \prod_{s=1}^{n-2} X_{\alpha^n_s} {\widetilde}{W} X_{e_{n+1}+e_{n+1}} \prod_{i=n+2}^{2n-1} C^2_i \prod_{t=n+1}^{2n-1} C^1_t, \psi_W, X_{\alpha^{n+1}_{n-1}}, X_{\beta^{n+1}_{n-1}}),\\ & (\prod_{i=1}^{n} R^2_i R^1_i R^2_{n+1} {\widetilde}{W} X_{e_{n+1}+e_{n+1}} \prod_{i=n+2}^{2n-1} C^2_i \prod_{t=n+2}^{2n-1} C^1_t \prod_{s=2}^{n+1} C^1_{n+1,s}, \psi_W, R^1_{n+1,1}, \\ & C^1_{n+1,1}),\\ & \cdots,\\ & (\prod_{i=1}^{n} R^2_i R^1_i R^2_{n+1} \prod_{s=1}^{n} R^1_{n+1,s}{\widetilde}{W} X_{e_{n+1}+e_{n+1}} \prod_{i=n+2}^{2n-1} C^2_i \prod_{t=n+2}^{2n-1} C^1_t, \psi_W, R^1_{n+1,n+1}, \\ & C^1_{n+1,n+1}),\end{aligned}$$ Applying Lemma 2.3 of [@JL13b] $2n$ times, we get that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ13} \int_{[\prod_{i=1}^{n+1} R^2_i R^1_i {\widetilde}{W} X_{e_{n+1}+e_{n+1}} \prod_{i=n+2}^{2n-1} C^2_iC^1_i]} \varphi(rwc \omega_2 \epsilon \omega_1 g) \psi_W(w)^{-1} dcdwdr.$$ After repeating the above procedure to the pairs $(R^2_s, C^2_s), (R^1_s, C^1_s)$, $s = n+1, n+2, \ldots, 2n-1$, we get that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ14} \int_{[\prod_{i=1}^{2n-1}R^2_i R^1_i {\widetilde}{W} \prod_{j=n+1}^{2n-1}X_{e_j+e_j}]} \varphi(rwx \omega_2 \epsilon \omega_1 g) \psi_W(w)^{-1} dxdwdr.$$ Now, let’s see the structure of elements in $\prod_{i=1}^{2n-1}R^2_i R^1_i {\widetilde}{W}\prod_{j=n+1}^{2n-1}X_{e_j+e_j}$. Any element in $\prod_{i=1}^{2n-1}R^2_i R^1_i {\widetilde}{W}\prod_{j=n+1}^{2n-1}X_{e_j+e_j}$ has the following form: $$w = \begin{pmatrix} z_{2n} & q_1 & q_2\\ 0 & u' & q_1^*\\ 0 & 0 & z_{2n}^* \end{pmatrix},$$ where $z_{2n} \in V_{2n}$, the standard standard maximal unipotent subgroup of $GL_{2n}$, $u' \in U_{[(2n)^{2m-2}],2} := \omega_1' V_{[(2n)^{2m-2}],2} \omega_1'^{-1}$, $q_1 \in M_{(2n)\times (4m-4)l}$, with the last row zero, $q_2 \in M_{(2n)\times (2n)}$, with $q_2(2n,1)=0$. And by , the restriction of $\psi_W$ to the $z_{2n}$-part gives a Whittaker character of $GL_{2n}$. Note that $\psi_W(\begin{pmatrix} z_{2n} & 0 & 0\\ 0 & I_{4mn-4n} & 0\\ 0 & 0 & z_{2n}^* \end{pmatrix})=\psi(z_{2n}(1,2)+\cdots + z_{2n}(n,n+1) - z_{2n}(n+1, n+2) - \cdots - z_{2n}(2n-1,2n)).$ It is clear that the integral in \[main1equ14\] is non-vanishing if and only if the following integral is non-vanishing: $$\label{main1equ17} \int_{[\prod_{i=1}^{2n-1}R^2_i R^1_i {\widetilde}{W} \prod_{j=n+1}^{2n-1}X_{e_j+e_j}]} \varphi(rwx \omega_2 \epsilon \omega_1 g) \psi_W'(w)^{-1} dxdwdr,$$ where $\psi_W'(\begin{pmatrix} z_{2n} & 0 & 0\\ 0 & I_{4mn-4n} & 0\\ 0 & 0 & z_{2n}^* \end{pmatrix})=\psi(\sum_{i=1}^{2n-1} z_{2n}(i,i+1)).$ Then it is easy to see that the integral in has an inner integral which is exactly $\varphi^{\psi_{N_{1^{2n-1}}}}$, using notation in Lemma \[lem2\]. On the other hand, we know that by Lemma \[lem2\], $\varphi^{\psi_{N_{1^{2n-1}}}} = \varphi^{{\widetilde}{\psi}_{N_{1^{2n}}}}$. Therefore, the integral in becomes $$\label{main1equ15} \int_{[{\widetilde}{U}]} \varphi(u \omega_2 \epsilon \omega_1 g) \psi_{{\widetilde}{U}}(u)^{-1} du,$$ where any element in ${\widetilde}{U}$ has the following form: $$u = u(z_{2l}, u',q_1, q_2) = \begin{pmatrix} z_{2n} & q_1 & q_2\\ 0 & u' & q_1^*\\ 0 & 0 & z_{2n}^* \end{pmatrix},$$ where $z_{2n} \in V_{2n}$, the standard maximal unipotent subgroup of $GL_{2n}$, $u' \in U_{[(2n)^{2m-2}],2} := \omega_1' V_{[(2n)^{2m-2}],2} \omega_1'^{-1}$, $q_1 \in M_{(2n)\times (4m-4)n}$, $q_2 \in M_{(2n)\times (2n)}$, such that $q_2^t v_{2n} - v_{2n} q_2 = 0$, where $v_{2n}$ is a matrix only with ones on the second diagonal. Hence, the integral in can be written as $$\label{main1equ16} \int_{u(z_{2n}, u',0, 0)} \varphi(u \omega_2 \epsilon \omega_1 g)_{P_{2n}} \psi_{{\widetilde}{U}}(u)^{-1} du,$$ where $\varphi_{P_{2n}}$ is the constant term of $\varphi$ along the parabolic subgroup $P_{2n}=M_{2n}U_{2n}$ of ${{\mathrm{Sp}}}_{4mn}$ with Levi isomorphic to $GL_{2n} \times Sp_{(4m-4)n}$. By Lemma \[constantterm\], there is an automorphic function $$f \in A(N_{2n}({{\mathbb {A}}})M_{2n}(F){\backslash}{{\mathrm{Sp}}}_{4mn}({{\mathbb {A}}}))_{\tau \lvert \cdot \rvert^{-\frac{2m-1}{2}} \otimes {{\mathcal {E}}}_{\Delta(\tau,m-1)}},$$ such that $$\varphi(g)_{P_{2n}} = f(g), \forall g \in Sp_{4mn}({{\mathbb {A}}}).$$ Therefore, $\varphi(u \omega_2 \epsilon \omega_1 g)_{P_{2n}}$ is an automorphic form in $\tau \lvert \cdot \rvert^{-\frac{2m-1}{2}} \otimes \mathcal{E}_{\Delta(\tau,m-1)}$. Note that the restriction of $\psi_{{\widetilde}{U}}$ to the $z_{2n}$-part gives us a Whittaker coefficient, and the restriction to the $u'$-part gives a Fourier coefficient of $\mathcal{E}_{\Delta(\tau,m-1)}$ attached to the partition $[(2n)^{2m-2}]$ up to a conjugation of the Weyl element $\omega_1'$. On the other hand, $\tau$ is generic, and by induction assumption, $\mathcal{E}_{\Delta(\tau,m-1)}$ has a nonzero Fourier coefficient attached to the partition $[(2n)^{2m-2}]$. Therefore, we can make the conclusion that $\mathcal{E}_{\Delta(\tau,m)}$ has a nonzero $\psi_{\underline{p},\underline{a}}$-Fourier coefficient attached to the partition $[(2n)^{2m}]$, for some ${\underline}{a} = \{1, -1, a_3, \ldots, a_{2m}\}$, $a_i \in F^*/(F^*)^2$, for $3 \leq i \leq 2m$. Hence, $\mathcal{O}(\mathcal{E}_{\Delta(\tau,m)}) = [(2n)^{2m}]$. Since the proof of Lemma \[lem2\] will be reduced to that of Lemma \[constantterm\], we have completed the proof of the theorem, up to Lemma \[constantterm\]. \[rmk3\] By similar argument, we can also prove that for the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, k) \otimes {\widetilde}{\pi}}$ of ${\widetilde}{Sp}_{4kn+2n}({{\mathbb {A}}})$, $$\mathfrak{p}^m({\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, k) \otimes {\widetilde}{\pi}}) = [(2n)^{2k+1}].$$ We will not give the details here. The following lemma is an ${{\mathrm{Sp}}}$-analogue of the Lemma 4.1 of [@JL13a]. It gives formulas for certain constant terms of automorphic forms in ${{\mathcal {E}}}_{\Delta(\tau, m)}$. The proof of this lemma will be given in the next section. \[constantterm\] Let $P_{2ni} = M_{2ni} N_{2ni}$, $1 \leq i \leq m$ be the parabolic subgroup of $Sp_{4mn}$ with Levi part $M_{2ni} \cong GL_{2ni} \times Sp_{4n(m-i)}$. Let $\varphi$ be an arbitrary automorphic form in ${{\mathcal {E}}}_{\Delta(\tau, m)}$. Denote by $\varphi(g)_{P_{2ni}}$ the constant term of $\varphi$ along $P_{2ni}$. Then for $1 \leq i \leq m-1$, there is an automorphic function $$f \in A(N_{2ni}({{\mathbb {A}}})M_{2ni}(F){\backslash}Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau, i)\lvert \cdot \rvert^{-\frac{2m-i}{2}} \otimes \mathcal{E}_{\Delta(\tau,m-i)}},$$ such that $$\varphi(g)_{P_{2ni}} = f(g), \forall g \in Sp_{4mn}({{\mathbb {A}}}).$$ For $i=m$, there is an automorphic function $$f \in A(N_{2mn}({{\mathbb {A}}})M_{2mn}(F){\backslash}Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau, m)\lvert \cdot \rvert^{-\frac{m}{2}}},$$ such that $$\varphi(g)_{P_{2mn}} = f(g), \forall g \in Sp_{4mn}({{\mathbb {A}}}).$$ Next, we prove two important lemmas, which can be viewed as an ${{\mathrm{Sp}}}$-analogue of Lemmas 6.1 and 6.2 of [@JL13a]. \[lem1\] Let $N_{1^{2mn}}$ be the standard maximal unipotent subgroup of $Sp_{4mn}$ consisting of upper triangular matrices. For $p \geq 2n$, define a character of $N_{1^{2mn}}$ as follows: $$\begin{aligned} \label{lem1equ1} \begin{split} \psi_p^{\underline{\epsilon}}(n) := & \psi(n_{1,2} + \cdots n_{p-1,p} + n_{p,p+1})\\ \cdot & \psi(\epsilon_1 n_{p+1, p+2} + \cdots + \epsilon_{2mn-p}n_{2mn,2mn+1}), \end{split}\end{aligned}$$ where $n \in N_{1^{2mn}}$, $\epsilon_i \in \{0, 1\}$, for $1 \leq i \leq 2mn-p-1$, $\epsilon_{2mn-p} \in F/(F^*)^2$, $\underline{\epsilon}= \{\epsilon_1 ,\ldots, \epsilon_{2mn-p}\}$. Then for any automorphic form $\varphi \in {{\mathcal {E}}}_{\Delta(\tau,m)}$, the following $\psi_p^{\underline{\epsilon}}$-Fourier coefficient is identically zero: $$\label{lem1equ2} \varphi^{\psi_p^{\underline{\epsilon}}}(g) := \int_{[N_{1^{2mn}}]} \varphi(ng)\psi_p^{\underline{\epsilon}}(n)^{-1} dn \equiv 0.$$ If $\epsilon_i \neq 0$, $\forall 1 \leq i \leq 2mn-p$, then $\psi_p^{\underline{\epsilon}}$ is a generic character of $G_n$. Since $\mathcal{E}_{\Delta(\tau,m)}$ is not generic, it has no nonzero $\psi_p^{\underline{\epsilon}}$-Fourier coefficients, i.e., $\varphi^{\psi_p^{\underline{\epsilon}}} \equiv 0$, $\forall \varphi \in \mathcal{E}_{\Delta(\tau,m)}$. Assume that $1 \leq i \leq 2mn-p$ is the first number such that $\epsilon_i = 0$. Then $\varphi^{\psi_p^{\underline{\epsilon}}}$ has an inner integral $\varphi_{P_{p+i}}$, which is the constant term of $\varphi$ along $P_{p+i}$, the parabolic subgroup of $Sp_{4mn}$ with Levi isomorphic to $GL_{p+i} \times Sp_{4mn-2(p+i)}$. Note that $p+i > 2n$. By the cuspidal support of $\mathcal{E}_{\Delta(\tau,m)}$, we can see that $\varphi_{P_{p+i}} = 0$ unless $p+i = 2ns$ with $2 \leq s \leq m$. By Lemma \[constantterm\], there is an automorphic function $$f \in A(N_{2ns}({{\mathbb {A}}})M_{2ns}(F){\backslash}Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau, s)\lvert \cdot \rvert^{-\frac{2m-s}{2}} \otimes \mathcal{E}_{\Delta(\tau,b-s)}},$$ such that $$\varphi(g)_{P_{2ns}} = f(g), \forall g \in Sp_{4mn}({{\mathbb {A}}}),$$ Therefore, after taking the constant term, $\varphi(g)_{P_{2ns}}$ is an automorphic function over $GL_{2ns}({{\mathbb {A}}}) \times Sp_{4n(m-s)}({{\mathbb {A}}})$. Note that the character $\psi(n_{1,2}+\cdots+n_{2ns-1,2ns})$ gives a Whittaker character of $GL_{2ns}$. Since by Proposition 2.1 of [@JL13a], $\Delta(\tau,s)$ is not generic for $s > 1$, i.e., it has no nonzero Whittaker Fourier coefficients, we conclude that $\varphi^{\psi_p^{\underline{\epsilon}}} \equiv 0$. This completes the proof of the lemma. \[lem2\] Let $N_{1^{p}}$ be the unipotent radical of the parabolic subgroup $P_{1^{p}}$ of $Sp_{4mn}$ with Levi part isomorphic to $GL_1^{\times p} \times Sp_{4mn-2p}$. Let $$\psi_{N_{1^{p}}}(n) := \psi(n_{1,2}+ \cdots + n_{p,p+1}),$$ and $$\widetilde{\psi}_{N_{1^{p}}}(n) := \psi(n_{1,2}+ \cdots + n_{p-1,p}),$$ be two characters of $N_{1^{p}}$. For any automorphic form $\varphi \in \mathcal{E}_{\Delta(\tau,m)}$, define $\psi_{N_{1^{p}}}$ and $\widetilde{\psi}_{N_{1^{p}}}$-Fourier coefficients as follows: $$\label{lem2equ1} \varphi^{\psi_{N_{1^{p}}}}(g) := \int_{[N_{1^{p}}]} \varphi(ng)\psi_{N_{1^{p}}}(n)^{-1} dn,$$ $$\label{lem2equ2} \varphi^{\widetilde{\psi}_{N_{1^{p}}}}(g) := \int_{[N_{1^{p}}]} \varphi(ng)\widetilde{\psi}_{N_{1^{p}}}(n)^{-1} du.$$ Then $\varphi^{\psi_{N_{1^{p}}}} \equiv 0, \forall p \geq 2n$; $\varphi^{\psi_{N_{1^{2n-1}}}} = \varphi^{\widetilde{\psi}_{N_{1^{2n}}}}$. First we assume that $p \geq n$. Let $X_{e_{p+1}+e_{p+1}}$ be the root subgroup corresponding to the root $e_{p+1}+e_{p+1}$. Since the conjugating action of $X_{e_{p+1}+e_{p+1}}$ normalizes $N_{1^{p}}$, and preserves the character $\psi_{N_{1^{p}}}$, we can take the Fourier expansion of $\varphi^{\psi_{N_{1^{p}}}}$ along $X_{e_{p+1}+e_{p+1}}$. Under the action of $GL_1$, we will get Fourier coefficients corresponding the two orbits of the dual of $X_{e_{p+1}+e_{p+1}}$: the trivial one and the non-trivial one. Note that the non-trivial orbit gives us Fourier coefficients which are exactly the Fourier coefficients attached to $[(2(p+1))1^{4mn-2q-2}]$, with $p+1 >n$. On the other hand, by Proposition 6.4 and Remark 6.5 of [@JL13c], all the Fourier coefficients attached to the non-trivial orbit vanish, and the Fourier coefficient attached to the trivial orbit, i.e., the Fourier coefficient with respect to the trivial character, survives. Hence, $\varphi^{\psi_{N_{1^{p}}}}$ becomes: $$\label{lem2equ3} \int_{[X_{e_{p+1}+e_{p+1}}]} \int_{[N_{1^{p}}]} \varphi(nxg)\psi_{N_{1^{p}}}(n)^{-1} dn dx.$$ Let $R_{i}$, $p+1 \leq i \leq 2mn-1$ be the following subgroup of $N_{1^{2mn}}$: $$\begin{aligned} R_{i} := & \{n \in N_{1^{2mn}}| n(j,l)=0, \forall (j,l)\neq (p+1,w), \\ & p+2 \leq w \leq 4mn-p-1\}.\end{aligned}$$ Since $R_{p+1}$ normalizes the group $X_{e_{p+1}+e_{p+1}} N_{1^{p}}$, and preserves the character $\psi_{N_{1^{p}}}$, we can take the Fourier expansion of the integral in along $R_{p+1}$. Since the subgroup $I_{p+1} \times Sp_{4mn-2p-2}$ (image of $Sp_{4mn-2p-2}$ under the embedding $g \mapsto \diag(I_{p+1}, g, I_{p+1})$ to $Sp_{4mn}$) of $Sp_{4mn}$ acts on the dual space of $[R_{p+1}]$ with two orbits: the trivial one and the non-trivial one, the integral in becomes: $$\label{lem2equ4} \sum_{\gamma} \varphi^{\psi_{N_{1^{p+1}}}}(\gamma g) + \varphi^{\widetilde{\psi}_{N_{1^{p}}}}(g),$$ where $\gamma$ is in some quotient space which we will not specify here. Then we continue the above process of Fourier expansion for the two kinds of Fourier coefficients $\varphi^{\psi_{N_{1^{p+1}}}}$ and $\varphi^{\widetilde{\psi}_{N_{1^{p+1}}}}$ along the pair of groups $(X_{e_{p+2}+e_{p+2}}, R_{p+2})$. We will get four kinds of Fourier coefficients: $$\int_{[N_{1^{p+2}}]} \varphi(ng)\psi_{N_{1^{p+2}}}^{\underline{\epsilon}}(n)^{-1} dn,$$ where $$\begin{aligned} \psi_{N_{1^{p+2}}}^{\underline{\epsilon}}(n) := & \psi(n_{1,2} + \cdots n_{p-1,p} + n_{p,p+1})\\ \cdot & \psi(\epsilon_1 n_{p+1, p+2} + \epsilon_2 n_{p+2, p+3}),\end{aligned}$$ and $\underline{\epsilon} = \{\epsilon_1, \epsilon_2\}$, $\epsilon_1, \epsilon_2 \in \{0,1\}$. Then we can continue the above process of Fourier expansion for each of these four kinds of Fourier coefficients along the pair of groups $(X_{e_{p+3}+e_{p+3}}, R_{p+3})$. We will get six kinds of Fourier coefficients: $$\int_{[N_{1^{p+3}}]} \varphi(ng)\psi_{N_{1^{p+3}}}^{\underline{\epsilon}}(n)^{-1} dn,$$ where $$\begin{aligned} \psi_{N_{1^{p+3}}}^{\underline{\epsilon}}(n) := & \psi(n_{1,2} + \cdots n_{p-1,p} + n_{p,p+1})\\ \cdot & \psi(\epsilon_1 n_{p+1, p+2} + \epsilon_2 n_{p+2, p+3} + \epsilon_2 n_{p+3, p+4}),\end{aligned}$$ and $\underline{\epsilon} = \{\epsilon_1, \epsilon_2, \epsilon_3\}$, $\epsilon_1, \epsilon_2, \epsilon_3 \in \{0,1\}$. Keep repeating the above procedure, we will get Fourier coefficients of the following form: $\varphi^{\psi_p^{\underline{\epsilon}}}$, with $\underline{\epsilon} = \{\epsilon_1, \ldots, \epsilon_{2mn-p}\}$, $\epsilon_i\in \{0,1\}$, $1 \leq i \leq 2mn-p-1$, and $\epsilon_{2mn-p} \in F/(F^*)^2$. By Lemma \[lem1\], all Fourier coefficients of this kind are zero, if $p \geq 2n$. Therefore, $\varphi^{\psi_{N_{1^{p}}}} \equiv 0$, if $p \geq 2n$. For $p=2n-1$, by , we can see that $\varphi^{\psi_{N_{1^{2n-1}}}} = \varphi^{\widetilde{\psi}_{N_{1^{2n}}}}$, since $\varphi^{\psi_{N_{1^{2n}}}} \equiv 0$ by above discussion. This completes the proof of the lemma. Proof of Lemma \[constantterm\] =============================== In this section, we prove Lemma \[constantterm\]. Before we start, we recall the definition of the Eisenstein series in Section 1: $$E(\phi, s)(g) = \sum_{\gamma\in P_{2mn}(F)\bks Sp_{4mn}(F)}\lambda_s \phi(\gamma g),$$ where $\phi \in A(N_{2mn}({{\mathbb {A}}})M_{2mn}(F) {\backslash}Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,m)}$. Assume that $\varphi = {{\mathrm{Res}}}_{s=\frac{m}{2}} E(\phi, s)$. To compute $\varphi_{P_{2ni}} = ({{\mathrm{Res}}}_{s=\frac{m}{2}} E(\phi, s))_{P_{2ni}}$, we use the fact that the residue operator and the constant term operator are interchangeable. So, first, we are going to calculate the constant term of $E(\phi,s)$ along $P_{2ni}$. We follow the calculation in Section 2 of [@JLZ12]. $$\begin{aligned} \label{ctequ1} \begin{split} & E_{P_{2ni}}(\phi,s)(g)\\ = & \int_{N_{2ni}(F)\bks N_{2ni}({{\mathbb {A}}})} E(\phi,s)(ug)du\\ = & \sum_{\omega^{-1} \in P_{2mn}\bks Sp_{4mn}/P_{2ni}}\sum_{\gamma \in M_{2ni}^{\omega}(F){\backslash}M_{2ni}(F)}\int_{[N_{2ni}^{\omega}]} \int_{N_{2ni, \omega}({{\mathbb {A}}})}\lam_s\phi({\omega}^{-1}\gamma u' u'' g)\\ & d u' d u'' \end{split}\end{aligned}$$ where we define $M_{2ni}^{\omega}:=\omega P_{2mn}{\omega}^{-1}\cap M_{2ni}$ and $N_{2ni}^{\omega}:=\omega P_{2mn}{\omega}^{-1}\cap N_{2ni}$ and $[N_{2ni}^{\omega}]:=N_{2ni}^{\omega}(F)\bks N_{2ni}^{\omega}({{\mathbb {A}}})$. Note that the unipotent radical $N_{2ni}$ can be decomposed as a product $N_{2ni, \omega}N_{2ni}^{\omega}$, where $N_{2ni,\omega}$ satisfies $N_{2ni,\omega}\cap N_{2ni}^{\omega}=\{1\}$ and $N_{2ni}=N_{2ni,\omega}N_{2ni}^{\omega}=N_{2ni}^{\omega}N_{2ni,\omega}$. Using the explicit calculation about the generalized Bruhat decomposition $P_{2mn}\bks Sp_{4mn}/P_{2ni}$ (see Chapter 4 of [@GRS11]), we can see that all the double cosets are killed by the cuspidal support of the Eisenstein series except two, which have the following representatives: $\omega_0 = Id$, and $$\omega_1 = \begin{pmatrix} 0&0&I_{2ni}&0\\ I_{2n(m-i)}&0&0&0\\ 0&0&0&I_{2n(m-i)}\\ 0&-I_{2ni}&0&0 \end{pmatrix}.$$ Then $$\label{ctequ2} E_{P_{2ni}}(\phi,s)(g)=E_{P_{2ni}}(\phi,s)_{\omega_0}(g)+E_{P_{2ni}}(\phi,s)_{\omega_1}(g),$$ where $$\begin{aligned} & E_{P_{2ni}}(\phi,s)_{\omega_j}(g)\\ = & \sum_{\gamma \in M_{2ni}^{\omega_j}(F){\backslash}M_{2ni}(F)}\int_{[N_{2ni}^{\omega_j}]} \int_{N_{2ni, \omega_j}({{\mathbb {A}}})}\lam_s\phi({\omega_j}^{-1}\gamma u' u'' g)d u' d u'',\end{aligned}$$ $j=0,1$. We will consider these two terms separately in the next two subsections. $\omega_0$-term --------------- Write elements in $N_{2ni}$ as follows: $$u(X,Z,W)=\begin{pmatrix} I_{2ni}&X&Z&W\\ &I_{2n(m-i)}&&Z^{'}\\ &&I_{2n(m-i)}&X^{'}\\ &&&I_{2ni} \end{pmatrix}.$$ Note that $P_{2mn}\cap M_{2ni}\bks M_{2ni}\cong P_{2n(m-i)}\bks Sp_{4n(m-i)}$, where $P_{2n(m-i)}$ is the parabolic subgroup of $Sp_{4n(m-i)}$ with Levi isomorphic to $GL_{2n(m-i)} \times 1_{Sp_0}$. Then the $\omega_0$-term of the constant term can be written as $$\label{ctequ3} E_{P_{2ni}}(\phi,s)_{\omega_0}(g) =\sum_{\gamma\in P_{2n(m-i)}(F)\bks Sp_{4n(m-i)}(F)}\int_{[N_{2ni}]}\lam_s\phi(\gamma u g)du.$$ The integral can be calculated as follows: $$\begin{aligned} & \int_{[N_{2ni}]}\lam_s\phi(\gamma u g)du\\ = & \int_{[N_{2ni}]}\lam_s\phi(u \gamma g)du\\ = & \int_{[M_{2ni\times 2n(m-i)}]}\int_{[N_{2mn}\cap U_{2ni}]}\lam_s\phi( u'u(X)\gamma g)d u' d X\\ = & \int_{[M_{2ni\times 2n(m-i)}]}\lam_s\phi(u(X)\gamma g)d X,\end{aligned}$$ where $u(X)=u(X,0,0)$ with $X \in M_{2ni\times 2n(m-i)}$. As in Section 2.2 of [@JLZ12], the integral $\int_{[M_{2ni\times 2n(m-i)}]}\lam_s\phi(u(X)\gamma g)d X$ can be viewed as the constant term of the automorphic function in $\Delta(\tau,m)$: $x \mapsto \phi(\diag(x,x^*)g)$, along the maximal parabolic subgroup $Q_{2ni,2n(m-i)}$ of $GL_{2mn}$ with Levi isomorphic to $GL_{2ni} \times GL_{2n(m-i)}$. We will denote it by $\phi_{Q_{2ni,2n(m-i)}}$. Let $P_{2ni,2n(m-i)} = M_{2ni,2n(m-i)}N_{2ni,2n(m-i)}$ be a standard parabolic subgroup of $Sp_{4mn}$, whose Levi $M_{2ni,2n(m-i)} \cong {{\mathrm{GL}}}_{2ni}\times {{\mathrm{GL}}}_{2n(m-i)}\times 1_{Sp_0}$. The following lemma is parallel to Lemma 2.1 of [@JLZ12]. \[omega0term\] The constant term $\lam_s\phi_{Q_{2ni,2n(m-i)}}$ belongs to the following space $$A(N_{2ni,2n(m-i)}({{\mathbb {A}}})M_{2ni,2n(m-i)}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,i) |\cdot|^{s-\frac{m-i}{2}}\otimes \Delta(\tau,m-i)|\cdot|^{s+\frac{i}{2}}\otimes 1_{Sp_0}}.$$ The proof is omitted here, since it is almost the same as that of Lemma 2.1 of [@JLZ12], word-by-word. Therefore, by and Lemma \[omega0term\], we can see that $E_{P_{2ni}}(\phi,s)_{\omega_0}$ belongs to the following space $$A(N_{2ni}({{\mathbb {A}}})M_{2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,i) |\cdot|^{s-\frac{m-i}{2}}\otimes (\Delta(\tau,m-i)|\cdot|^{s+\frac{i}{2}}\rtimes 1_{Sp_0})}.$$ Hence, the residue operator will kill $E_{P_{2ni}}(\phi,s)_{\omega_0}$ , since $s=\frac{m}{2}$ is not a pole of the Eisenstein series on $Sp_{4n(m-i)}$ with inducing data $\Delta(\tau,m-i)|\cdot|^{s+\frac{i}{2}}\otimes 1_{Sp_0}$. $\omega_1$-term --------------- Note that $N_{2ni}^{\omega_1}=\{ u(Z)=u(0,Z,0)\mid Z\in M_{2ni\times 2n(m-i)}\}$. and $M_{2ni}^{\omega_1}(F)\bks M_{2ni}(F)$ is isomorphic to $P_{2n(m-i)}(F)\bks Sp_{4n(m-i)}(F)$. Therefore, we have $$\begin{aligned} \label{ctequ4} \begin{split} &E_{P_{2ni}}(\phi,s)_{\omega_1}(g)\\ =&\sum_{\gamma\in P_{2n(m-i)}(F)\bks Sp_{4n(m-i)}}\int_{N_{2ni,\omega_1}({{\mathbb {A}}})} \int_{[M_{2ni\times 2n(m-i)}]}\lam_s\phi(\omega_1^{-1}\gamma u(Z) u g)d Z d u\\ =&\sum_{\gamma\in P_{2n(m-i)}(F)\bks Sp_{4n(m-i)}}\int_{N_{2ni,\omega_1}({{\mathbb {A}}})} \int_{[M_{2n(m-i)\times2ni}]}\lam_s\phi(u'(X) \omega_1^{-1} u \gamma g)d X du, \end{split}\end{aligned}$$ where $$u'(X)= \begin{pmatrix} I_{2n(m-i)}&X&&\\ &I_{2ni}&&\\ &&I_{2ni}&X'\\ &&&I_{2n(m-i)} \end{pmatrix} \text{ for } X\in M_{2n(m-i)\times 2ni}.$$ As in case of $\omega_0$, the integral $\int_{[M_{2n(m-i)\times2ni}]} \phi(u'(X)g)d X$ can be viewed as the constant term of the automorphic function in $\Delta(\tau,m)$: $x \mapsto \phi(\diag(x,x^*)g)$, along the maximal parabolic subgroup $Q_{2n(m-i),2ni}$ of $GL_{2mn}$ with Levi isomorphic to $GL_{2n(m-i)} \times GL_{2ni}$. We will denote it by $\phi_{Q_{2n(m-i),2ni}}$. Let $P_{2n(m-i),2ni} = M_{2n(m-i),2ni}U_{2n(m-i),2ni}$ be a standard parabolic subgroup of $Sp_{4mn}$, whose Levi $M_{2n(m-i),2ni} \cong {{\mathrm{GL}}}_{2n(m-i) \times {{\mathrm{GL}}}_{2ni}}\times 1_{Sp_0}$. Then, by Lemma \[omega0term\], the constant term $\lam_s\phi_{Q_{2n(m-i),2ni}}$ belongs to the following space $$A(N_{2n(m-i),2ni}({{\mathbb {A}}})M_{2n(m-i),2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,m-i) |\cdot|^{s-\frac{i}{2}}\otimes \Delta(\tau,i)|\cdot|^{s+\frac{m-i}{2}}\otimes 1_{Sp_0}}.$$ Note that the outer integral in is the intertwining operator corresponding to $\omega_1$, which maps $$A(N_{2n(m-i),2ni}({{\mathbb {A}}})M_{2n(m-i),2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau,m-i) |\cdot|^{s-\frac{i}{2}}\otimes \Delta(\tau,i)|\cdot|^{s+\frac{m-i}{2}}\otimes 1_{Sp_0}}$$ to $$A(N_{2ni,2n(m-i)}({{\mathbb {A}}})M_{ai,a(b-i)}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{ \Delta(\tau,i)|\cdot|^{-(s+\frac{m-i}{2})} \otimes \Delta(\tau,m-i) |\cdot|^{s-\frac{i}{2}}\otimes 1_{Sp_0}}.$$ Note that $\tau$ is self-dual. Therefore, by , and the the above discussion, we can see that $E_{P_{2ni}}(\phi,s)_{\omega_1}$ belongs to the following space $$A(N_{2ni}({{\mathbb {A}}})M_{2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{ \Delta(\tau,i)|\cdot|^{-(s+\frac{m-i}{2})} \otimes (\Delta(\tau,m-i) |\cdot|^{s-\frac{i}{2}}\rtimes 1_{Sp_0})}.$$ And for $1 \leq i \leq m-1$, after taking the residue operator, ${{\mathrm{Res}}}_{s=\frac{m}{2}} E_{P_{2ni}}(\phi,s)_{\omega_1}$ belongs to the following space $$A(N_{2ni}({{\mathbb {A}}})M_{2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau, i)\lvert \cdot \rvert^{-\frac{2m-i}{2}} \otimes \mathcal{E}_{\Delta(\tau,m-i)}},$$ since $s=\frac{m}{2}$ is the rightmost simple pole of the (normalized) Eisenstein series with inducing data $\Delta(\tau,m-i) |\cdot|^{s-\frac{i}{2}}\otimes 1_{Sp_0}$, and it is not a pole of the intertwining operator corresponding to $\omega_1$. Therefore, for $1 \leq i \leq m-1$, $$\begin{aligned} & \varphi_{P_{2ni}} \\ = & ({{\mathrm{Res}}}_{s=\frac{m}{2}} E(\phi,s))_{P_{2ni}} \\ = & {{\mathrm{Res}}}_{s=\frac{m}{2}} (E_{P_{2ni}}(\phi,s)) \\ = & {{\mathrm{Res}}}_{s=\frac{m}{2}} (E_{P_{2ni}}(\phi,s)_{\omega_1}),\end{aligned}$$ belongs to the following space $$A(N_{2ni}({{\mathbb {A}}})M_{2ni}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{\Delta(\tau, i)\lvert \cdot \rvert^{-\frac{2m-i}{2}} \otimes \mathcal{E}_{\Delta(\tau,m-i)}}.$$ For $i=m$, $\varphi_{P_{2mn}} = {{\mathrm{Res}}}_{s=\frac{m}{2}} (E_{P_{2mn}}(\phi,s)_{\omega_1})$ belongs to the following space $$A(N_{2mn}({{\mathbb {A}}})M_{2mn}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{ \Delta(\tau,m)|\cdot|^{-\frac{m}{2}} \otimes 1_{Sp_0}},$$ since $E_{P_{2mn}}(\phi,s)_{\omega_1}$ belongs to the following space $$A(N_{2mn}({{\mathbb {A}}})M_{2mn}(F)\bks Sp_{4mn}({{\mathbb {A}}}))_{ \Delta(\tau,m)|\cdot|^{-s} \otimes 1_{Sp_0}},$$ and $s=\frac{m}{2}$ is a simple pole of the intertwining operator corresponding to $\omega_1$. This completes the proof of Lemma \[constantterm\]. Proof of Part (1) of Theorem \[thm1\] ===================================== In this section, we will prove that $\Psi$ is surjective, which comes from the computation of composition of two descents $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}),$$ where ${\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, and $${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \subset {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi).$$ It turns out that there is a similar identity: $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})={\widetilde}{\sigma}_{4(m-1)n+2n},$$ as in Proposition 5.2 of [@GJS12]. We will prove in the next section that $\Psi$ is well-defined, i.e., ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$ is irreducible (see Theorem \[thm6\]). Note that, from this section to Section 6, we assume that $F$ is a number field which is not totally imaginary, unless specified. To start, we recall Proposition 4.1 of [@JL13c] which generalizes Theorem 6.3 of [@GRS11] and is true for any number field. \[prop1\] Assume that $F$ is any number field. \(1) Let $\mu_i$, $1 \leq i \leq r$, be characters of $F_v^*$, $a \in F_v^*$, then $$\begin{aligned} \label{prop1equ1} \begin{split} & FJ_{\psi^a_{k-1}} ({{\mathrm{Ind}}}_{P_{m_1, \ldots, m_k}}^{Sp_{2n}} \nu^{\alpha_1} \chi_1 (det_{GL_{m_1}}) \otimes \cdots \otimes \nu^{\alpha_k} \chi_k(det_{GL_{m_k}}))\\ \cong & {{\mathrm{Ind}}}_{P_{m_1-1, \ldots, m_k-1}}^{{\widetilde}{Sp}_{2n-2k}} \mu_{\psi^{-a}} \nu^{\alpha_1} \chi_1 (det_{GL_{m_1-1}}) \otimes \cdots \otimes \nu^{\alpha_k} \chi_k(det_{GL_{m_k-1}}). \end{split}\end{aligned}$$ \(2) Let $\mu_i$, $1 \leq i \leq r$, be characters of $F_v^*$, $a, b \in F_v^*$, then $$\begin{aligned} \label{prop1equ2} \begin{split} & FJ_{\psi^b_{k-1}} ({{\mathrm{Ind}}}_{P_{m_1, \ldots, m_k}}^{{\widetilde}{Sp}_{2n}} \mu_{\psi^{a}}\nu^{\alpha_1} \chi_1 (det_{GL_{m_1}}) \otimes \cdots \otimes \nu^{\alpha_k} \chi_k(det_{GL_{m_k}}))\\ \cong & {{\mathrm{Ind}}}_{P_{m_1-1, \ldots, m_k-1}}^{{Sp}_{2n-2k}} \chi_{\frac{b}{a}} \nu^{\alpha_1} \chi_1 (det_{GL_{m_1-1}}) \otimes \cdots \otimes \nu^{\alpha_k} \chi_k(det_{GL_{m_k-1}}), \end{split}\end{aligned}$$ where $\chi_{\frac{b}{a}}$ is a quadratic character of $F_v^*$ defined by Hilbert symbol as follows: $\chi_{\frac{b}{a}}(x)=(\frac{b}{a}, x)$, and $\nu=\lvert \det(\cdot) \rvert$. Note that when $a=b=1$, Proposition \[prop1\] is exactly Theorem 6.3 of [@GRS11]. Next, we prove the equality mentioned at the beginning of this section, which will imply later that $\Psi$ is surjective. \[thm7\] (1) $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \neq 0.$$ \(2) $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})={\widetilde}{\sigma}_{4(m-1)n+2n}.$$ The proof of Part (1) is similar to that of Theorem \[main1part2\] and to that of Theorem 2.1 of [@GJS12]. The proof of Part (2) is similar to those of Theorem 5.1 and Proposition 5.2 of [@GJS12]. **Proof of Part (1)**. Note that descents ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}$ and ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}$ are defined in Section 3.2 of [@GRS11]. By Corollary 2.4 of [@JL13b] or Lemma 1.1 of [@GRS03], and the discussion at the end of Section 1 of [@GRS03], ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \neq 0$ if and only if the following integral is non-vanishing: $$\begin{aligned} \label{thm7equ1} \begin{split} & \int_{[Y_2 V_{[(2n)1^{4mn-2n}],2}]} \int_{[Y_1 V_{[(2n)1^{4mn}],2}]} {\widetilde}{\varphi}(v v_1 g) \\ & \psi_{[(2n)1^{4mn}],1}^{-1}(v) \psi_{[(2n)1^{4mn-2n}],-1}^{-1}(v_1) dv dv_1, \end{split}\end{aligned}$$ where ${\widetilde}{\varphi} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, $g \in {\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}})$, embedded into ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$ via the map $g \mapsto \diag(I_{2n}, g, I_{2n})$; $Y_1$, $Y_2$ are the groups defined in (2.5) of [@JL13b] corresponding to the partitions $[(2n)1^{4mn}]$ and $[(2n)1^{4mn-2n}]$ respectively; $V_{[(2n)1^{4mn-2n}],2}$ and $V_{[(2n)1^{4mn}],2}$ are defined in Section 2 of [@JL13b]. Explicitly, let $N_{1^n}$ be the unipotent radical of the parabolic ${P}_{1^n}$ of $Sp_{4mn+2n}$ with the Levi subgroup isomorphic to $GL_1^n \times {Sp}_{4mn}$, then $Y_1 V_{[(2n)1^{4mn}],2}$ is a subgroup of $N_{1^n}$ consists of elements $v$ with $v_{n, j}=0$, for $n+1 \leq j \leq 2mn+n$. $\psi_{[(2n)1^{4mn}],1}(v) = \psi(\sum_{i=1}^{n-1} v_{i,i+1} + v_{n,4mn+n+1})$. Identify $Sp_{4mn}$ with its embedding into $Sp_{4mn+2n}$ via the map $g \mapsto \diag(I_n, g, I_n)$. Let $N_{1^n}$ be the unipotent radical of the parabolic ${P}_{1^n}$ of $Sp_{4mn}$ with the Levi subgroup isomorphic to $GL_1^n \times {Sp}_{4mn-2n}$, then $Y_1 V_{[(2n)1^{4mn-2n}],2}$ is a subgroup of $N_{1^n}$ consists of elements $v$ with $v_{n, j}=0$, for $n+1 \leq j \leq 2mn$. $\psi_{[(2n)1^{4mn-2n}],-1}(v) = \psi(\sum_{i=1}^{n-1} v_{i,i+1} - v_{n,4mn-n+1})$. Let ${\widetilde}{\omega}$ be the Weyl element of $GL_{2n}$ defined in (4.31) of [@GRS99]: $$\begin{aligned} {\widetilde}{\omega}_{2i,i}=1, & \text{ } i=1, \ldots, n,\\ {\widetilde}{\omega}_{2i-1,i+n}=1, & \text{ } i=1, \ldots, n,\\ {\widetilde}{\omega}_{i,j}=0, & \text{ } \text{otherwise.}\end{aligned}$$ Let $$\label{thm7equ2} \omega_1 = \begin{pmatrix} {\widetilde}{\omega} & & \\ & I_{4mn-2n} &\\ && {\widetilde}{\omega}^* \end{pmatrix} \in Sp_{4mn+2n}(F).$$ As in [@GJS12], we identify $Sp_{4mn+2n}(F)$ with the subgroup $Sp_{4mn+2n}(F) \times 1$ of ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$. Conjugating cross the integral in \[thm7equ1\] by $\omega_1$, it becomes: $$\begin{aligned} \label{thm7equ3} \int_{[W_1]} {\widetilde}{\varphi}(w \omega_1 g) \psi_{W_1}^{-1}(w) dw,\end{aligned}$$ where $W_1 = \omega_1 Y_2 V_{[(2n)1^{4mn-2n}],2} Y_1 V_{[(2n)1^{4mn}],2} \omega_1^{-1}$, if $\omega_1^{-1} w \omega_1 = v_1 v$, $v_1 \in Y_2 V_{[(2n)1^{4mn-2n}],2}$, $v \in Y_1 V_{[(2n)1^{4mn}],2}$, then $$\psi_{W_1}(w)=\psi_{[(2n)1^{4mn}],1}(v) \psi_{[(2n)1^{4mn-2n}],-1}(v_1).$$ Note that the metaplectic cover splits over $W_1({{\mathbb {A}}})$, and $W_1({{\mathbb {A}}}) \times 1$ is a subgroup of ${\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}})$. We identify $W_1$ with $W_1 \times 1$. Elements in $W_1$ are of the following form $$\label{thm7equ29} w=\begin{pmatrix} Z & q_1 & q_2\\ & I_{4mn-2n} & q_1^*\\ &&Z^* \end{pmatrix},$$ where $q_1(i,j)=0$, for $i=2n-1, 2n$, $1 \leq j \leq 2mn-n$, $Z \in GL_{2n}$ has the form (4.34) of [@GRS99]. Write $Z$ as an $n \times n$ matrix of $2 \times 2$ block matrices $Z = ([Z]_{i,j})$, $1 \leq i, j \leq n$, then $[Z]_{n,1} = \cdots = [Z]_{n,n-1}=0$, $[Z]_{n,n}=I_2$; $[Z]_{i,i}$ is lower unipotent, for $i < n$; $[Z]_{i,j}$ is lower triangular, for $i < j$; $[Z]_{i,j}$ is lower nilpotent, for $j < i < n$. And $$\label{thm7equ4} \psi_{W_1}(w) = \psi(\sum_{i=1}^{n-1} {{\mathrm{tr}}}([Z]_{i,i+1})+(q_2(2n,1)-q_2(2n-1,2))).$$ Let $R^1_i=\prod_{j=1}^{i} R^1_{i,j}$, for $1 \leq i \leq n-1$, with $R^1_{i,j} = X_{\alpha_{i,j}}$, the root subgroup corresponding to the root $\alpha_{i,j} = e_{2i} -e_{2(j-1)+1}$. Let $R^1 = \prod_{i=1}^{n-1} R^1_i$. Actually $R^1$ is the subgroup of $W_1$ consists of lower unipotent matrices. Write $W_1 = R^1 {\widetilde}{W}_1$, with $R^1 \cap {\widetilde}{W}_1 = \{1\}$. Let $C^1_i=\prod_{j=1}^{i} C^1_{i,j}$, for $1 \leq i \leq n-1$, with $C^1_{i,j} = X_{\beta_{i,j}}$, the root subgroup corresponding to the root $\beta_{i,j} = e_{2(j-1)+1} -e_{2(i+1)}$. We consider the quadruple $$\label{thm7equ5} ({\widetilde}{W}_1 \prod_{i=1}^{n-2} R^1_i \prod_{j=2}^{n-1} R^1_{n-1,j}, \psi_{W_1}, R^1_{n-1,1}, C^1_{n-1,1}).$$ It is easy to see that this quadruple satisfies all the conditions for Lemma 2.3 of [@JL13b]. Hence, by Lemma 2.3 of [@JL13b], the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\begin{aligned} \label{thm7equ6} \int_{[C^1_{n-1,1}{\widetilde}{W}_1 \prod_{i=1}^{n-2} R^1_i \prod_{j=2}^{n-1} R^1_{n-1,j}]} {\widetilde}{\varphi}(cwr \omega_1 g) \psi_{W_1}^{-1}(w) drdwdc.\end{aligned}$$ We continue to consider the following sequence of quadruples: $$\begin{aligned} \label{thm7equ7} \begin{split} & (C^1_{n-1,1}{\widetilde}{W}_1 \prod_{i=1}^{n-1} R^1_i \prod_{j=3}^{n-1} R^1_{n-1,j}, \psi_{W_1}, R^1_{n-1,2}, C^1_{n-1,2}),\\ & (\prod_{k=1}^2 C^1_{n-1,k}{\widetilde}{W}_1 \prod_{i=1}^{n-2} R^1_i \prod_{j=4}^{n-1} R^1_{n-1,j}, \psi_{W_1}, R^1_{n-1,3}, C^1_{n-1,3}),\\ & \cdots,\\ & (\prod_{k=1}^{n-2} C^1_{n-1,k}{\widetilde}{W}_1 \prod_{i=1}^{n-2} R^1_i , \psi_{W_1}, R^1_{n-1,n-1}, C^1_{n-1,n-1}). \end{split}\end{aligned}$$ Applying Lemma 2.3 of [@JL13b] $(n-1)$ times, we get that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\begin{aligned} \label{thm7equ8} \int_{[C^1_{n-1} {\widetilde}{W}_1 \prod_{i=1}^{n-2} R^1_i]} {\widetilde}{\varphi}(cwr \omega_1 g) \psi_{W_1}^{-1}(w) drdwdc.\end{aligned}$$ Then, we repeat the above procedure for the pairs $$(R^1_{n-2}, C^1_{n-2}), \ldots, (R^1_1, C^1_1).$$ For example, after repeating the above procedure for the pairs $$(R^1_{n-2}, C^1_{n-2}), \ldots, (R^1_{s+1}, C^1_{s+1}),$$ we need the following sequence of quadruples for the pair $(R^1_s, C^1_s)$: $$\begin{aligned} \label{thm7equ9} \begin{split} & (\prod_{i=s+1}^{n-1} C^1_i {\widetilde}{W}_1 \prod_{i=1}^{s-1} R^1_i \prod_{j=2}^{s} R^1_{s,j}, \psi_{W_1}, R^1_{s,1}, C^1_{s,1}),\\ & (\prod_{i=s+1}^{n-1} C^1_i C^1_{s,1} {\widetilde}{W}_1 \prod_{i=1}^{s-1} R^1_i \prod_{j=3}^{s} R^1_{s,j}, \psi_{W_1}, R^1_{s,2}, C^1_{s,2}),\\ & \cdots,\\ & (\prod_{i=s+1}^{n-1} C^1_i \prod_{k=1}^{s-1} C^1_{s,k} {\widetilde}{W}_1 \prod_{i=1}^{s-1} R^1_i, \psi_{W_1}, R^1_{s,s}, C^1_{s,s}). \end{split}\end{aligned}$$ After applying Lemma 2.3 of [@JL13b] $s$ times, the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\begin{aligned} \label{thm7equ10} \int_{[\prod_{i=s}^{n-1} C^1_i {\widetilde}{W}_1 \prod_{j=1}^{s-1} R^1_j]} {\widetilde}{\varphi}(cwr \omega_1 g) \psi_{W_1}^{-1}(w) drdwdc.\end{aligned}$$ After repeating the above procedure for all the pairs $$(R^1_{n-2}, C^1_{n-2}), \ldots, (R^1_1, C^1_1),$$ we will see that the integral in is non-vanishing if and only if the following integral is non-vanishing: $$\begin{aligned} \label{thm7equ11} \int_{[\prod_{i=1}^{n-1} C^1_i {\widetilde}{W}_1]} {\widetilde}{\varphi}(cw \omega_1 g) \psi_{W_1}^{-1}(w) dwdc.\end{aligned}$$ Note that $\prod_{i=1}^{n-1} C^1_i {\widetilde}{W}_1 = \omega_1 V_{[(2n)^2 1^{4mn-2n}],2} \omega_1^{-1}$, and $\psi_{\prod_{i=1}^{n-1} C^1_i{\widetilde}{W}_1}(cw)=\psi_{W_1}(w)=\psi_{[(2n)^2 1^{4mn-2n}], \{-1, 1\}} (\omega_1^{-1} cw \omega_1)$. Let $A = \begin{pmatrix} 1 & -1\\ 1 & 1 \end{pmatrix}$, and $\epsilon=\diag(A, \ldots, A; I_{(4m-4)n}; A^*, \ldots, A^*)$, as in (2.31) of [@GJS12]. Conjugating cross the integral in by $\epsilon$, it becomes: $$\begin{aligned} \label{thm7equ12} \int_{[W_2]} {\widetilde}{\varphi}(w \epsilon \omega_1 g) \psi_{W_2}^{-1}(w) dw,\end{aligned}$$ where $W_2 = \epsilon \prod_{i=1}^{n-1} C^1_i{\widetilde}{W}_1 \epsilon^{-1}$, $\psi_{W_2}(w)=\psi_{\prod_{i=1}^{n-1} C^1_i{\widetilde}{W}_1}(\epsilon^{-1} w \epsilon)$. Elements in $W_2$ are of the following form $$\label{thm7equ35} w=\begin{pmatrix} Z & q_1 & q_2\\ & I_{4mn-2n} & q_1^*\\ &&Z^* \end{pmatrix},$$ where $q_1(i,j)=0$, for $i=2n-1, 2n$, $1 \leq j \leq 2mn-n$, $Z$ is in the unipotent radical of the parabolic subgroup of $GL_{2n}$ with the Levi subgroup isomorphic to $GL_2^n$. And $$\label{thm7equ13} \psi_{W_2}(w) = \psi(\sum_{i=1}^{2n-2} Z_{i,i+2}+q_2(2n-1,1)).$$ Let $\nu=\begin{pmatrix} \nu_1 & \nu_2\\ \nu_3 & \nu_4 \end{pmatrix}$ be the Weyl element in . Let $$\label{thm7equ37} \omega_2 = \begin{pmatrix} \nu_1 & & \nu_2\\ & I_{4mn-2n} & \\ \nu_3 & & \nu_4 \end{pmatrix},$$ a Weyl element of ${{\mathrm{Sp}}}_{4mn+2n}$. Conjugating cross the integral in by $\omega_2$, it becomes: $$\label{thm7equ14} \int_{[W_3]} {\widetilde}{\varphi}(w \omega_2 \epsilon \omega_1 g) \psi_{W_3}^{-1}(w) dw,$$ where $W_3 = \omega_2 W_2 \omega_2^{-1}$, $\psi_{W_3}(w)=\psi_{W_2}(\omega_2^{-1} w \omega_2)$. Elements in $W_3$ have the following form: $$\label{thm7equ15} w = \begin{pmatrix} Z & q_1 & q_2\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & Z^* \end{pmatrix} \begin{pmatrix} I_{2n} & 0 & 0\\ p_1 & I_{4mn-2n} & 0\\ p_2 & p_1^* & I_{2n} \end{pmatrix},$$ where $Z$ is in the standard maximal unipotent subgroup of $GL_{2n}$; $q_1(i,j)=0$, for $n+1 \leq i \leq 2n$, and for $i=n$, $1 \leq j \leq 2mn-n$; $q_2(i,j)=p_2(i,j)=0$, for $1 \leq j \leq i \leq 2n$; $p_1(i,j)=0$, for $1 \leq j \leq n$, and for $j=n+1$, $n+1 \leq i \leq 2n$. Let ${\widetilde}{C}$ be the unipotent subgroup consisting of elements of the following form: $$\begin{pmatrix} I_{2n} & 0 & 0\\ p_1 & I_{4mn-2n} & 0\\ 0 & p_1^* & I_{2n} \end{pmatrix},$$ where $p_1(i,j)=0$ for any $i,j$, except $j=n+1$, $1 \leq i \leq 2mn-n$. Let ${\widetilde}{R}$ be the unipotent subgroup consisting of elements of the following form: $$\begin{pmatrix} I_{2n} & q_1 & 0\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & I_{2n} \end{pmatrix},$$ where $q_1(i,j)=0$ for any $i,j$, except $i=n$, $1 \leq j \leq 2mn-n$. Write $W_3 = {\widetilde}{C} {\widetilde}{W}_3$, with ${\widetilde}{C} \cap {\widetilde}{W_3}=\{1\}$. Consider the quadruple $({\widetilde}{W}_3, \psi_{W_3}, {\widetilde}{R}, {\widetilde}{C})$. It is easy to see that this quadruple satisfies all the conditions of Lemma 2.3 of [@JL13b]. Hence, by Lemma 2.3 of [@JL13b], the integral is non-vanishing if and only if the following integral is non-vanishing: $$\label{thm7equ16} \int_{[W_4]} {\widetilde}{\varphi}(w \omega_2 \epsilon \omega_1 g) \psi_{W_4}^{-1}(w) dw,$$ where $W_4={\widetilde}{R} {\widetilde}{W}_3$, for $w= rw'$, $r \in {\widetilde}{R}$, $w \in {\widetilde}{W}_3$, $\psi_{W_4}(rw') = \psi_{W_3}(w')$. Let $R^2_i = \prod_{j=1}^i X_{\alpha^i_j}$, for $1 \leq i \leq n$, with $\alpha^i_j = e_i+e_{2n-i+j}$, and $R^2_i = \prod_{j=1}^{2n-i} X_{\alpha^i_j}$, for $n+1 \leq i \leq 2n-1$, with $\alpha^i_j = e_i+e_{i+j}$. Let $C^2_i = \prod_{j=1}^i X_{\beta^i_j}$, for $1 \leq i \leq n$, with $\beta^i_j = -e_{2n-i+j}-e_{i+1}$, and $C^2_i = \prod_{j=1}^{2n-i} X_{\beta^i_j}$, for $n+1 \leq i \leq 2n-1$, with $\beta^i_j = -e_{i+j}-e_{i+1}$. For $n+1 \leq i \leq 2n-1$, let $R^3_i$ be the unipotent subgroup consisting of elements of the following form: $$\begin{pmatrix} I_{2n} & q_1 & 0\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & I_{2n} \end{pmatrix},$$ where $q_1(k,j)=0$ for any $k,j$, except $k=i$, $1 \leq j \leq 4mn-2n$. For $n+1 \leq i \leq 2n-1$, let $C^3_i$ be the unipotent subgroup consisting of elements of the following form: $$\begin{pmatrix} I_{2n} & 0 & 0\\ p_1 & I_{4mn-2n} & 0\\ 0 & p_1^* & I_{2n} \end{pmatrix},$$ where $p_1(k,j)=0$ for any $k,j$, except $j=i+1$, $1 \leq k \leq 4mn-2n$. Write $W_4 = \prod_{i=1}^{2n-1} C^2_i \prod_{j=n+1}^{2n-1} C^3_j {\widetilde}{W}_4$, with $\prod_{i=1}^{2n-1} C^2_i \prod_{j=n+1}^{2n-1} C^3_j \cap {\widetilde}{W}_4 = \{1\}$. Then, we apply Lemma 2.3 of [@JL13b] to the pairs $$(R^2_1, C^2_1), (R^2_2, C^2_2), \ldots, (R^2_n, C^2_n).$$ For example, for $(R^2_s, C^2_s)$, $1 \leq s \leq n$, we need to consider the following sequence of quadruples: $$\begin{aligned} &(\prod_{i=1}^{s-1} R^2_i \prod_{j=n+1}^{2n-1} C^3_j {\widetilde}{W}_4 \prod_{t=s+1}^{2n-1} C^2_t \prod_{l=2}^s X_{\beta^s_l}, \psi_{W_4}, X_{\beta^s_1}, X_{\alpha^s_1}),\\ & (\prod_{i=1}^{s-1} R^2_i X_{\alpha^s_1} \prod_{j=n+1}^{2n-1} C^3_j {\widetilde}{W}_4 \prod_{t=s+1}^{2n-1} C^2_t \prod_{l=3}^s X_{\beta^s_l}, \psi_{W_4}, X_{\beta^s_2}, X_{\alpha^s_2}),\\ & \cdots,\\ & (\prod_{i=1}^{s-1} R^2_i \prod_{k=1}^{s-1} X_{\alpha^s_k} \prod_{j=n+1}^{2n-1} C^3_j {\widetilde}{W}_4 \prod_{t=s+1}^{2n-1} C^2_t, \psi_{W_4}, X_{\beta^s_s}, X_{\alpha^s_s}).\\\end{aligned}$$ After applying Lemma 2.3 of [@JL13b] to all the pairs $$(R^2_1, C^2_1), (R^2_2, C^2_2), \ldots, (R^2_n, C^2_n),$$ we get that the integral is non-vanishing if and only if the following integral is non-vanishing: $$\label{thm7equ17} \int_{[\prod_{i=1}^{n} R^2_i {\widetilde}{W}_4 \prod_{t=n+1}^{2n-1} C^2_t C^3_t]} {\widetilde}{\varphi}(rwc \omega_2 \epsilon \omega_1 g) \psi_{W_4}^{-1}(w) dcdwdr,$$ Next, we apply Lemma 2.3 of [@JL13b] to the pairs $$(R^2_{n+1}, C^2_{n+1}), (R^3_{n+1}, C^3_{n+1}); \cdots; (R^2_{2n-1}, C^2_{2n-1}), (R^3_{2n-1}, C^3_{2n-1}).$$ Note that before applying Lemma 2.3 of [@JL13b] to each pair $(R^2_{s}, C^2_{s})$, $n+1 \leq s \leq 2n-1$, we need to take the Fourier expansion along the one-dimensional root subgroup $X_{e_s+e_s}$, as in the proof of Theorem \[main1part2\]. For example, for $s=n+1$, we first take the Fourier expansion of the integral in along the one-dimensional root subgroup $X_{e_s+e_s}$. Under the action of $GL_1$, we get two kinds of Fourier coefficients corresponding to the two orbits of the dual of $[X_{e_s+e_s}]$: the trivial one and the non-trivial one. For any Fourier coefficient attached to the non-trivial orbit, we can see that there is an inner integral $\varphi_{[(2n+2)1^{4mn-2}],\{a\}}$, which is identically zero by (1) in the proof of Theorem \[thm10\]. Therefore only the Fourier coefficient attached to the trivial orbit, which actually equals to the integral in , survives. After applying Lemma 2.3 of [@JL13b] to all the pairs $$(R^2_{n+1}, C^2_{n+1}), (R^3_{n+1}, C^3_{n+1}); \cdots; (R^2_{2n-1}, C^2_{2n-1}), (R^3_{2n-1}, C^3_{2n-1}),$$ the integral is non-vanishing if and only if the following integral is non-vanishing: $$\label{thm7equ18} \int_{[\prod_{i=1}^{2n-1} R^2_i \prod_{t=n+1}^{2n-1} R^3_t X_{e_t+e_t} {\widetilde}{W}_4]} {\widetilde}{\varphi}(rxw \omega_2 \epsilon \omega_1 g) \psi_{W_4}^{-1}(w) dwdxdr,$$ Note that elements in $\prod_{i=1}^{2n-1} R^2_i \prod_{t=n+1}^{2n-1} R^3_t X_{e_t+e_t} {\widetilde}{W}_4$ have the following form: $$\label{thm7equ19} w = \begin{pmatrix} Z & q_1 & q_2\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & Z^* \end{pmatrix},$$ where $Z$ is in the standard maximal unipotent subgroup of $GL_{2n}$; the last row of $q_1$ is zero. And $\psi_{W_4}(\begin{pmatrix} Z & 0 & 0\\ 0 & I_{4mn-2n} & 0\\ 0 & 0 & Z^* \end{pmatrix}) = \psi(\sum_{i=1}^{n-1}Z_{i,i+1} - \sum_{j=n}^{2n-1} Z_{i,i+1})$. Clearly the integral is non-vanishing if and only if the following integral is non-vanishing: $$\label{thm7equ20} \int_{[\prod_{i=1}^{2n-1} R^2_i \prod_{t=n+1}^{2n-1} R^3_t X_{e_t+e_t} {\widetilde}{W}_4]} {\widetilde}{\varphi}(rxw \omega_2 \epsilon \omega_1 g) \psi_{W_4}^{'-1}(w) dwdxdr,$$ where $\psi_{W_4}'(\begin{pmatrix} Z & 0 & 0\\ 0 & I_{4mn-2n} & 0\\ 0 & 0 & Z^* \end{pmatrix}) = \psi(\sum_{i=1}^{2n-1}Z_{i,i+1})$. Note that the integral is exactly ${\widetilde}{\varphi}^{\psi_{N_{1^{2n-1}}}}$, using notation in Lemma \[lem2\]. On the other hand, we know that by Lemma \[lem2\], ${\widetilde}{\varphi}^{\psi_{N_{1^{2n-1}}}} = {\widetilde}{\varphi}^{{\widetilde}{\psi}_{N_{1^{2n}}}}$. Note that Lemma \[lem2\] also applies to metaplectic groups. Therefore, the integral in becomes $$\label{thm7equ21} \int_{[{\widetilde}{U}]} {\widetilde}{\varphi}(u \omega_2 \epsilon \omega_1 g) \psi_{{\widetilde}{U}}(u)^{-1} du,$$ where any element in ${\widetilde}{U}$ has the following form: $$u = u(Z, q_1, q_2) = \begin{pmatrix} Z & q_1 & q_2\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & Z^* \end{pmatrix},$$ where $Z$ is in the standard maximal unipotent subgroup of $GL_{2n}$, $q_1 \in M_{(2n)\times (4m-4)n}$, $q_2 \in M_{(2n)\times (2n)}$, such that $q_2^t v_{2n} - v_{2n} q_2 = 0$, where $v_{2n}$ is a matrix only with ones on the second diagonal. $\psi_{{\widetilde}{U}}(u) = \psi(\sum_{i=1}^{2n-1} u_{i,i+1})$. Hence, the integral in can be written as $$\label{thm7equ22} \int_{u(Z, 0, 0)} {\widetilde}{\varphi}(u \omega_2 \epsilon \omega_1 g)_{{\widetilde}{P}_{2n}} \psi_{{\widetilde}{U}}(u)^{-1} du,$$ where ${\widetilde}{\varphi}_{P_{2n}}$ is the constant term of ${\widetilde}{\varphi}$ along the pre-image of the parabolic subgroup $P_{2n}=M_{2n}U_{2n}$ of ${{\mathrm{Sp}}}_{4mn+2n}$ with Levi isomorphic to $GL_{2n} \times Sp_{(4m-2)n}$. By the similar calculation as in the proof of Lemma \[constantterm\], or the calculation at the end of Theorem 2.1 of [@GJS12], there is an automorphic function $$f \in A(N_{2n}({{\mathbb {A}}}){\widetilde}{M}_{2n}(F){\backslash}{\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}}))_{\tau \lvert \cdot \rvert^{-m} \otimes {\widetilde}{\sigma}_{4(m-1)n+2n}},$$ such that $${\widetilde}{\varphi}(g)_{{\widetilde}{P}_{2n}} = f(g), \forall g \in {\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}}).$$ Therefore, the integral is the Whittaker Fourier coefficient of an element in $\tau$, hence not identically zero. This completes the proof of Part (1). **Proof of Part (2)**. By definition of Fourier-Jacobi coefficients ((3.14) of [@GRS11]), for $\phi_1 \in {{\mathcal {S}}}({{\mathbb {A}}}^{2mn-n})$, $\phi_2 = \phi_{21} \otimes \phi_{22}$, $\phi_{21} \in {{\mathcal {S}}}({{\mathbb {A}}}^n)$, $\phi_{22} \in {{\mathcal {S}}}({{\mathbb {A}}}^{2mn-n})$, we need to compute the composition of two Fourier-Jacobi coefficients $FJ^{\phi_1}_{\psi^{-1}_{n-1}}$ and $FJ^{\phi_2}_{\psi^{1}_{n-1}}$: $$\begin{aligned} \label{thm7equ23} \begin{split} & FJ^{\phi_1}_{\psi^{-1}_{n-1}} \circ FJ^{\phi_2}_{\psi^{1}_{n-1}}({\widetilde}{\xi})(g)\\ = & \int_{[V_{[(2n)1^{4mn-2n}],1}]} \int_{[V_{[(2n)1^{4mn}],1}]} {\widetilde}{\varphi}(uvg) \theta^{2mn, \phi_2}_{\psi^{-1}}(l_2(u)vg) \psi_{[(2n)1^{4mn}], 1}^{-1}(u)du\\ & \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_1(v)g) \psi_{[(2n)1^{4mn-2n}], -1}^{-1}(v)dv, \end{split}\end{aligned}$$ where ${\widetilde}{\varphi} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, $g \in {\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}})$, the theta series are defined in Section 1.2 [@GRS11]. $V_{[(2n)1^{4mn-2n}],1}$ and $V_{[(2n)1^{4mn}],1}$, $\psi_{[(2n)1^{4mn}], 1}$ and $\psi_{[(2n)1^{4mn-2n}], -1}$ are defined in Section 2 of [@JL13b]. $V_{[(2n)1^{4mn-2n}],1}$ and $V_{[(2n)1^{4mn}],1}$ are as $N_n$ in (3.14) of [@GRS11]. Explicitly, $V_{[(2n)1^{4mn}],1}$ is the unipotent radical of the parabolic subgroup $P_{1^n}^{4mn+2n}$ of $Sp_{4mn+2n}$ with Levi subgroup isomorphic to $GL_1^n \times Sp_{4mn}$. $V_{[(2n)1^{4mn-2n}],1}$ is the unipotent radical of the parabolic subgroup $P_{1^n}^{4mn}$ of $Sp_{4mn}$, with Levi subgroup isomorphic to $GL_1^n \times Sp_{4mn-2n}$. Note that $Sp_{4mn}$ is embedded into $Sp_{4mn+2n}$ via the map $g \mapsto \diag(I_n, g, I_n)$, and we identify it with the image. Then for $u \in V_{[(2n)1^{4mn}],1}$, $\psi_{[(2n)1^{4mn}], 1}(u) = \psi(\sum_{i=1}^{n-1} u_{i,i+1})$, $l_2(u) = \prod_{i=1}^{2mn} X_{\alpha_i}(u_{n,n+i})$, with $\alpha_i = e_n-e_{n+i}$. And for $v \in V_{[(2n)1^{4mn-2n}],1}$, $\psi_{[(2n)1^{4mn-2n}], -1}(v) = \psi(\sum_{i=1}^{n-1} v_{n+i, n+i+1})$, $l_1(v) = \prod_{i=1}^{2mn-n} X_{\beta_i}(v_{2n, 2n+i})$, with $\beta_i = e_{2n}-e_{2n+i}$. First, we want to unfold the theta series $\theta^{2mn, \phi_2}_{\psi^{-1}}(l_2(u)vg)$. Write $l_2(u)$ as $l_2(u)=(q_1, q_2, q_3; z)$, where $q_1, q_3 \in {{\mathbb {A}}}^n$, $q_2 \in {{\mathbb {A}}}^{4mn-2n}$, $z \in {{\mathbb {A}}}$. Then $$\begin{aligned} \label{thm7equ24} \begin{split} & \theta^{2mn, \phi_2}_{\psi^{-1}}(l_2(u)vg)\\ = & \sum_{\xi \in F^{2mn}} \omega_{\psi^{-1}}^{2mn} (l_2(u)vg)\phi_2(\xi)\\ = & \sum_{\xi_1 \in F^n, \xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((q_1, q_2, q_3; z)vg)\phi_2(\xi_1, \xi_2)\\ = & \sum_{\xi_1 \in F^n, \xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((\xi_1, 0, 0; 0)(q_1, q_2, q_3; z)vg)\phi_2(0, \xi_2)\\ = & \sum_{\xi_1 \in F^n, \xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((q_1 + \xi_1, q_2, q_3; z + \xi_1 \nu_n q_3')vg)\phi_2(0, \xi_2)\\ = & \sum_{\xi_1 \in F^n, \xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((0, q_2, q_3; z + {\widetilde}{\xi}_1 )(q_1 + \xi_1, 0, 0; 0)vg)\phi_2(0, \xi_2) \end{split}\end{aligned}$$ $$\begin{aligned} = & \sum_{\xi_1 \in F^n, \xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((0, {\widetilde}{q}_2, q_3; z + {\widetilde}{\xi}_1 )vg(q_1 + \xi_1, 0, 0; 0))\phi_2(0, \xi_2)\\\end{aligned}$$ where $\nu_n$ is the matrix only with 1’s on the second diagonal, ${\widetilde}{\xi}_1 = 2 \xi_1 \nu_n q_3'+ q_3 \nu_n q_1'$, ${\widetilde}{q}_2 = q_2 + (q_1+\xi_1)(n)v$, $(q_1+\xi_1)(n)$ is the $n$-th coordinate of the vector $q_1+\xi_1$. Note that $(q_1 + \xi_1, 0, 0; 0))$ commutes with $g$. Plugging into the integral in , collapsing the summation over $\xi_1$ with the integration over $q_1$, changing variables for $q_2$ and $z$, we will have $$\begin{aligned} \label{thm7equ25} \begin{split} & FJ^{\phi_1}_{\psi^{-1}_{n-1}} \circ FJ^{\phi_2}_{\psi^{1}_{n-1}}({\widetilde}{\xi})(g)\\ = & \int_{[V_{[(2n)1^{4mn-2n}],1}]} \int_{[V_{[(2n)1^{4mn}],1}']} \int_{{{\mathbb {A}}}^n} {\widetilde}{\varphi}(uvg\hat{q}_1) \\ &\sum_{\xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} (l_2(u)vg(q_1, 0, 0; 0))\phi_2(0, \xi_2) \psi_{[(2n)1^{4mn}], 1}^{-1}(u)du\\ & \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_1(v)g) \psi_{[(2n)1^{4mn-2n}], -1}^{-1}(v)dq_1dv, \end{split}\end{aligned}$$ where $\hat{q}_1 = \prod_{i=1}^{n} X_{\alpha_i}(q_1(n,n+i))$, with $\alpha_i = e_n-e_{n+i}$; $V_{[(2n)1^{4mn}],1}'$ is a subgroup of $V_{[(2n)1^{4mn}],1}$ consisting of elements $u$ with $u_{n,n+i}=0$, for $1 \leq i \leq n$; $l_2(u) = (0, q_2, q_3; z)$. Note that, $V_{[(2n)1^{4mn}],1}'$ is actually $Y V_{[(2n)1^{4mn}],2}$, where $Y$ is defined in (2.5) of [@JL13b] corresponding to the partition $[(2n)1^{4mn}]$, and $V_{[(2n)1^{4mn}],2}$ is defined Section 2 of [@JL13b]. For short, we will write $q_1$ for $(q_1, 0, 0; 0)$. By Formula (1.4) [@GRS11], $$\begin{aligned} \label{thm7equ26} \begin{split} & \sum_{\xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((0, q_2, q_3; z)vg(q_1, 0, 0; 0))\phi_2(0, \xi_2)\\ = & \sum_{\xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((0, q_2, 0; z)vg(q_1, 0, 0; 0))\phi_2(0, \xi_2)\\ = & \sum_{\xi_2 \in F^{2mn-n}} \omega_{\psi^{-1}}^{2mn} ((0, q_2, 0; z)vg(q_1, 0, 0; 0))\phi_2(q_1, \xi_2)\\ = & \theta^{2mn-n, \phi_{22}}_{\psi^{-1}}((q_2, z)vg) \phi_{21}(q_1), \end{split}\end{aligned}$$ since we assumed that $\phi_2 = \phi_{21} \otimes \phi_{22}$. Let $l_3(u) = (q_2, z)$, for $u \in V_{[(2n)1^{4mn}],1}'$. Then, the integral in becomes $$\begin{aligned} \label{thm7equ27} \begin{split} & \int_{[V_{[(2n)1^{4mn-2n}],1}]} \int_{[V_{[(2n)1^{4mn}],1}']} \int_{{{\mathbb {A}}}^n} {\widetilde}{\varphi}(uvg\hat{q}_1) \phi_{21}(q_1) dq_1\\ &\theta^{2mn-n, \phi_{22}}_{\psi^{-1}}(l_3(u)vg) \psi_{[(2n)1^{4mn}], 1}^{-1}(u)du \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_1(v)g) \psi_{[(2n)1^{4mn-2n}], -1}^{-1}(v)dv\\ = & \int_{[V_{[(2n)1^{4mn-2n}],1}]} \int_{[V_{[(2n)1^{4mn}],1}']} {\widetilde}{\varphi}'(uvg) \theta^{2mn-n, \phi_{22}}_{\psi^{-1}}(l_3(u)vg) \psi_{[(2n)1^{4mn}], 1}^{-1}(u)du \\ & \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_1(v)g) \psi_{[(2n)1^{4mn-2n}], -1}^{-1}(v)dv, \end{split}\end{aligned}$$ where ${\widetilde}{\varphi}'(uvg) = \int_{{{\mathbb {A}}}^n} {\widetilde}{\varphi}(uvg\hat{q}_1) \phi_{21}(q_1) dq_1$. Note that we still have that ${\widetilde}{\varphi}' \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Let $\omega_1$ be the Weyl element of $Sp_{4mn+2n}$ in . Conjugating cross by $\omega_1$, the integral in becomes $$\begin{aligned} \label{thm7equ28} \begin{split} & \int_{[W_1']} {\widetilde}{\varphi}'(w \omega_1 g) \theta^{2mn-n, \phi_{22}}_{\psi^{-1}}(l_4(w)g) \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_5(w)g) \psi_{W_1'}^{-1}(w)dw, \end{split}\end{aligned}$$ where $W_1'=\omega_1 V_{[(2n)1^{4mn-2n}],1} V_{[(2n)1^{4mn}],1}' \omega_1^{-1}$, its elements have the form as in , except that there is no requirement that $q_1(i,j)=0$, for $i=2n-1, 2n$, $1 \leq j \leq 2mn-n$. For $w=\begin{pmatrix} Z & q_1 & q_2\\ & I_{4mn-2n} & q_1^*\\ &&Z^* \end{pmatrix}$ as in , $l_4(w) = (q_1(2n-1), q_2(2n-1,2))$, $l_5(w)=(q_1(2n), q_2(2n,1))$, $q_1(2n-1)$, $q_1(2n)$ are the $(2n-1)$-th, $(2n)$-th rows of $q_1$, respectively. And $\psi_{W_1'}(w)= \psi(\sum_{i=1}^{n-1} {{\mathrm{tr}}}([Z]_{i,i+1})$, with notation as in . Next, we repeat the steps from to , and use Lemma 2.5 of [@JL13b] whenever Lemma 2.3 of [@JL13b] is used. We will get that the integral in becomes $$\begin{aligned} \label{thm7equ30} \begin{split} & \int_{[W_1'']} {\widetilde}{\varphi}''(w \omega_1 g) \theta^{2mn-n, \phi_{22}}_{\psi^{-1}}(l_4(w)g) \theta^{2mn-n, \phi_1}_{\psi^{1}}(l_5(w)g) \psi_{W_1''}^{-1}(w)dw, \end{split}\end{aligned}$$ where $W_1''$ is unipotent radical of the parabolic subgroup $P_{2^n}^{4mn+2n}$ of $Sp_{4mn+2n}$ with Levi subgroup isomorphic to $GL_2^n \times Sp_{4mn-2n}$. ${\widetilde}{\varphi}'' \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. And for $w \in W_1''$, $\psi_{W_1''}(w)= \psi(\sum_{i=1}^{2n-2} w_{i,i+2})$. Next, we want to unfold the two theta series as before. To do this, we need to use certain property of theta series as in (5.9) [@GJS12]: $$\begin{aligned} \label{thm7equ31} \begin{split} & \theta^{2mn-n, \phi_{22}}_{\psi^{-1}}((x_1, y_1; z_1)g) \theta^{2mn-n, \phi_1}_{\psi^{1}}((x_2, y_2; z_1)g)\\ = & \theta^{4mn-2n, \phi_{22} \otimes \phi_1}_{\psi} ((x_1, x_2, y_2, -y_1; z_2-z_1) {\widetilde}{g}), \end{split}\end{aligned}$$ where for $w \in W_1'$, we write $l_4(w) = (x_1, y_1; z_1)$, $l_5(w) = (x_2, y_2; z_2)$, $x_1, y_1, x_2, y_2 \in {{\mathbb {A}}}^{2mn-n}$, and $${\widetilde}{g} = \begin{pmatrix} A & & & -B\\ &A&B&\\ &C&D&\\ -C& & &D \end{pmatrix},$$ if we write $g=\left( \begin{pmatrix} A&B\\ C&D \end{pmatrix}, \varepsilon \right)$. Let $\phi_3 = \phi_{22} \otimes \phi_1$. Let $$\gamma = \begin{pmatrix} I_{2mn-n} & 0 & \frac{1}{2} I_{2mn-n} & 0\\ I_{2mn-n} & 0 & -\frac{1}{2} I_{2mn-n} & 0\\ 0 & I_{2mn-n} & 0 & \frac{1}{2} I_{2mn-n}\\ 0 & - I_{2mn-n} & 0 &\frac{1}{2} I_{2mn-n} \end{pmatrix} \in Sp_{8mn-4n}.$$ Then, $$\begin{aligned} \label{thm7equ32} \begin{split} (x_1, x_2, y_2, -y_1) \gamma & = (x_1+x_2, y_1+y_2, \frac{1}{2}(x_1-x_2), -\frac{1}{2}(y_1-y_2)),\\ \hat{g} & := \gamma^{-1} {\widetilde}{g} \gamma = \left( \begin{pmatrix} g &\\ & g^* \end{pmatrix}, 1 \right). \end{split}\end{aligned}$$ Therefore, by , the right hand side of becomes: $$\begin{aligned} \label{thm7equ33} \begin{split} & \theta^{4mn-2n, \phi_3}_{\psi} ((x_1, x_2, y_2, -y_1; z_2-z_1) {\widetilde}{g})\\ = & \theta^{4mn-2n, \phi_3'}_{\psi} ((x_1+x_2, y_1+y_2, \frac{1}{2}(x_1-x_2), -\frac{1}{2}(y_1-y_2); z_2-z_1) \hat{g}), \end{split}\end{aligned}$$ where $\phi_3' = \omega^{4mn-2n}_{\psi} (\gamma^{-1}) \phi_3$. Let $A = \begin{pmatrix} 1 & -1\\ 1 & 1 \end{pmatrix}$, and $\epsilon=\diag(A, \ldots, A; I_{(4m-4)n}; A^*, \ldots, A^*)$, as in (2.31) of [@GJS12]. Conjugating cross the integral in by $\epsilon$, it becomes: $$\begin{aligned} \label{thm7equ34} \begin{split} & \int_{[W_2']} {\widetilde}{\varphi}''(w \epsilon \omega_1 g) \theta^{4mn-2n, \phi_3'}_{\psi} (l_6(w)\hat{g}) \psi_{W_2'}^{-1}(w)dw\\ = & \int_{[W_2']} {\widetilde}{\varphi}''(w g \epsilon \omega_1) \theta^{4mn-2n, \phi_3'}_{\psi} (l_6(w)\hat{g}) \psi_{W_2'}^{-1}(w)dw\\ = & \int_{[W_2']} {\widetilde}{\varphi}'''(w g) \theta^{4mn-2n, \phi_3'}_{\psi} (l_6(w)\hat{g}) \psi_{W_2'}^{-1}(w)dw, \end{split}\end{aligned}$$ where $W_2' = \epsilon W_1'' \epsilon^{-1}$, and elements in $W_2'$ are as in , except that there is no requirement that $q_1(i,j)=0$, for $i=2n-1, 2n$, $1 \leq j \leq 2mn-n$, still ${\widetilde}{\varphi}''' \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Note that $g$ commutes with $\epsilon \omega_1$. And $\psi_{W_2'}(w) = \psi(\sum_{i=1}^{2n-2} w_{i,i+2}+w_{2n-1,4mn+1})$. For $w=\begin{pmatrix} Z & q_1 & q_2\\ & I_{4mn-2n} & q_1^*\\ &&Z^* \end{pmatrix}$ as in , for $i=2n-1, 2n$, write the $i$-th row of $q_1$ as $q_1(i)=(x_i, y_i)$, then $$l_6(w) = (x_{2n}, y_{2n}, \frac{1}{2}x_{2n-1}, -\frac{1}{2}y_{2n-1}, -q_2(2n-1,1)).$$ Then, we unfold the theta series $\theta^{4mn-2n, \phi_3'}_{\psi} (l_6(w)\hat{g})$ as in , the integral in becomes: $$\begin{aligned} \label{thm7equ36} \begin{split} & \int_{[W_3']} \int_{{{\mathbb {A}}}^{4mn-2n}} {\widetilde}{\varphi}'''(w g \widehat{\xi}) \phi_3(\xi) d\xi\psi_{W_2'}^{-1}(w)dw\\ = & \int_{[W_3']} {\widetilde}{\varphi}^{(4)}(w g)\psi_{W_3'}^{-1}(w)dw \end{split}\end{aligned}$$ where $\widehat{\xi} = \prod_{i=1}^{2mn-n} X_{\alpha_i} (\xi(i)) \prod_{j=1}^{2mn-n} X_{\beta_j} (\xi(2mn-n+j))$, with $\alpha_i = e_{2n}-e_{2n+i}$, $\beta_j = e_{2n} + e_{2mn+n-j+1}$, and ${\widetilde}{\varphi}^{(4)}(w g)=\int_{{{\mathbb {A}}}^{4mn-2n}} {\widetilde}{\varphi}'''(w g \widehat{\xi}) \phi_3(\xi)$, still ${\widetilde}{\varphi}^{(4)} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. And $W_3'$ is the subgroup of $W_2'$ consisting of elements $w=\begin{pmatrix} Z & q_1 & q_2\\ & I_{4mn-2n} & q_1^*\\ &&Z^* \end{pmatrix}$, with $q_1(2n)=0$. Conjugate cross the integral in by the Weyl elements $\omega_2$ in , it becomes: $$\begin{aligned} \label{thm7equ38} \begin{split} & \int_{[W_4]} {\widetilde}{\varphi}^{(4)}(w \omega_2 g)\psi_{W_4}^{-1}(w)dw\\ = & \int_{[W_4]} {\widetilde}{\varphi}^{(4)}(w g \omega_2)\psi_{W_4}^{-1}(w)dw\\ = & \int_{[W_4]} {\widetilde}{\varphi}^{(5)}(w g)\psi_{W_4}^{-1}(w)dw, \end{split}\end{aligned}$$ where $W_4$, $\psi_{W_4}$ are exactly as in . Now, we repeat the steps from to , and use Lemma 2.5 of [@JL13b] whenever Lemma 2.3 of [@JL13b] is used. Then, we get that the integral in becomes: $$\begin{aligned} \label{thm7equ39} \begin{split} \int_{[W_5]} {\widetilde}{\varphi}^{(5)}(w g)\psi_{W_5}^{-1}(w)dw, \end{split}\end{aligned}$$ where $W_5=\prod_{i=1}^{2n-1} R^2_i \prod_{t=n+1}^{2n-1} R^3_t X_{e_t+e_t} {\widetilde}{W}_4$ as in . And given $w = \begin{pmatrix} Z & q_1 & q_2\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & Z^* \end{pmatrix}$ as in , $\psi_{W_5}(w) = \psi(\sum_{i=1}^{n-1}Z_{i,i+1} - \sum_{j=n}^{2n-1} Z_{i,i+1})$. Let $t = \diag(I_{n-1}, -I_{n+1}; I_{4mn-2n}; -I_{n+1}, I_{n-1}) \in Sp_{4mn+2n}(F)$. Since ${\widetilde}{\varphi}^{(5)}$ is automorphic, the integral in becomes: $$\begin{aligned} \label{thm7equ40} \begin{split} & \int_{[W_5]} {\widetilde}{\varphi}^{(5)}(tw g)\psi_{W_5}^{-1}(w)dw\\ = & \int_{[W_5]} {\widetilde}{\varphi}^{(5)}(w g)\psi_{W_5}^{',-1}(w)dw, \end{split}\end{aligned}$$ after changing variable $w \mapsto t^{-1} w$, and $\psi_{W_5}^{'}(w)=\psi(\sum_{i=1}^{2n-1}Z_{i,i+1})$, for $w = \begin{pmatrix} Z & q_1 & q_2\\ 0 & I_{4mn-2n} & q_1^*\\ 0 & 0 & Z^* \end{pmatrix} \in W_5$. Note that the integral on the right hand side of the identity in is exactly ${\widetilde}{\varphi}^{\psi_{N_{1^{2n-1}}}}$, using notation in Lemma \[lem2\]. On the other hand, we know that by Lemma \[lem2\], ${\widetilde}{\varphi}^{\psi_{N_{1^{2n-1}}}} = {\widetilde}{\varphi}^{{\widetilde}{\psi}_{N_{1^{2n}}}}$. Note that Lemma \[lem2\] also applies to metaplectic groups. Therefore, the integral in becomes $$\label{thm7equ41} \int_{[{\widetilde}{U}]} {\widetilde}{\varphi}^{(5)}(u g) \psi_{{\widetilde}{U}}(u)^{-1} du,$$ where ${\widetilde}{U}$ and $\psi_{{\widetilde}{U}}$ are exactly as in . Now, it follows easily from the end of the proof of Part (1) that as a function of $g \in {\widetilde}{Sp}_{4mn-2n}$, the integral in gives a section in ${\widetilde}{\sigma}_{4(m-1)n+2n}$. Since starting from the integral in , we always get equalities, $FJ^{\phi_1}_{\psi^{-1}_{n-1}} \circ FJ^{\phi_2}_{\psi^{1}_{n-1}}({\widetilde}{\xi}) \in {\widetilde}{\sigma}_{4(m-1)n+2n}$. Therefore, $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \subset {\widetilde}{\sigma}_{4(m-1)n+2n}.$$ On the other hand, by Part (1), $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \neq 0.$$ Since ${\widetilde}{\sigma}_{4(m-1)n+2n}$ is irreducible, we have that $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) = {\widetilde}{\sigma}_{4(m-1)n+2n}.$$ This finishes the proof of Part (2), and completes the proof of the theorem. \[rmk2\] Note that in the proof of Part (2), we could easily get a similar identity as in Theorem 5.1 [@GJS12], but for simplicity, we did not write it down explicitly. Theorem \[thm7\] easily implies the following result. \[thm10\] ${\underline}{p}=[(2n)^{2m}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$ is a maximal partition providing non-vanishing Fourier coefficients for ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. By Theorem \[thm7\], $$[(2n)1^{4mn}] \circ [(2n)1^{4mn-2n}] \circ [(2n)^{2m-2}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$$ is a composite partition providing non-vanishing Fourier coefficients for ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. By Lemma 2.6 of [@GRS03] or Lemma 3.1 of [@JL13b], $[(2n)^{2m}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$ is a partition providing non-vanishing Fourier coefficients for ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Since by Theorem \[thm7\], ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})={\widetilde}{\sigma}_{4(m-1)n+2n},$ and ${\underline}{p}({\widetilde}{\sigma}_{4(m-1)n+2n})=[(2n)^{2(m-1)}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}],$ to show the maximality of ${\underline}{p}$, we just have to show that \(1) at the step of taking ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}$, ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$ has no nonzero Fourier coefficients attached to the symplectic partitions $[(2l)1^{4mn+2n-2l}]$ for any $l \geq n+1$, or $[(2l+1)^21^{4mn+2n-4l-2}]$, for any $l \geq n$; \(2) at the step of taking ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}$, ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ has no nonzero Fourier coefficients attached to the symplectic partitions $[(2l)1^{4mn-2l}]$ for any $l \geq n+1$, or $[(2l+1)^21^{4mn-4l-2}]$, for any $l \geq n$. We will show (1) and (2) using calculations of unramified local components. Let $v$ be a finite place such that ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n},v}$ is unramified. This means that both $\tau_v$ and ${\widetilde}{\sigma}_{4(m-1)n+2n, v}$ are also unramified. Assume that $\tau_v = \times_{i=1}^n \nu^{\alpha_i} \chi_i \times \times_{i=1}^n \nu^{-\alpha_i} \chi_i^{-1}$, where $\nu^{\alpha_i}(\cdot) = \lvert det(\cdot) \rvert^{\alpha_i}$, $0 \leq \alpha_i < \frac{1}{2}$, and $\chi_i$’s are unitary unramified characters of $F^*_v$, for $1 \leq i \leq n$. Since ${\widetilde}{\pi}$ lifts weakly to $\tau$ with respect to $\psi$, ${\widetilde}{\pi}_v = \mu_{\psi} \times_{i=1}^n \nu^{\alpha_i} \chi_i \rtimes 1_{{\widetilde}{Sp}_0}$. Since by the definition of the set ${{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, ${\widetilde}{\sigma}_{4(m-1)n+2n}$ is an irreducible cuspidal automorphic representation of ${\widetilde}{Sp}_{4(m-1)n+2n}({{\mathbb {A}}})$, which is nearly equivalent to the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, m-1) \otimes {\widetilde}{\pi}}$, similarly as Lemma 3.1 of [@GRS05], it is easy to see that ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n},v}$ is the unique unramified component of the following induced representation $$\label{thm10equ1} {{\mathrm{Ind}}}^{{\widetilde}{Sp}_{4mn+2n}(F^*_v)}_{{\widetilde}{P}_{(2m+1)^n}(F^*_v)} \mu_{\psi} \otimes_{i=1}^n \lvert \cdot \rvert^{\alpha_i} \chi_i(det_{GL_{2m+1}}) \otimes 1_{{\widetilde}{Sp}_0},$$ where ${P}_{(2m+1)^n}$ is the parabolic subgroup of ${Sp}_{4mn+2n}$ with Levi isomorphic to $GL_{2m+1}^n \times 1_{Sp_0}$, and ${\widetilde}{P}_{(2m+1)^n}$ is its full pre-image in ${\widetilde}{Sp}_{4mn+2n}$. By Lemma 3.2 of [@JL13c], we can easily see that (1) holds. By of Proposition \[prop1\], $$\begin{aligned} \label{thm10equ2} \begin{split} & FJ_{\psi^1_{n-1}} ({{\mathrm{Ind}}}^{{\widetilde}{Sp}_{4mn+2n}(F^*_v)}_{{\widetilde}{P}_{(2m+1)^n}(F^*_v)} \mu_{\psi} \otimes_{i=1}^n \lvert \cdot \rvert^{\alpha_i} \chi_i(det_{GL_{2m+1}}) \otimes 1_{{\widetilde}{Sp}_0})\\ \cong & {{\mathrm{Ind}}}^{{Sp}_{4mn}(F^*_v)}_{{P}_{(2m)^n}(F^*_v)} \otimes_{i=1}^n \lvert \cdot \rvert^{\alpha_i} \chi_i(det_{GL_{2m}}) \otimes 1_{{Sp}_0}, \end{split}\end{aligned}$$ which is actually irreducible, by results in [@Jan96], and is an unramified local component of ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$. Again, by Lemma 3.1 of [@JL13c], we can easily see that (2) also holds. Therefore, we have shown that ${\underline}{p}=[(2n)^{2m}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$ is a maximal partition providing non-vanishing Fourier coefficients for ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, which completes the proof of this theorem. To continue, we prove that ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ is a cuspidal representation, every component of which is inside ${{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$. \[thm5\] ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}) \subset {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$. We will prove that \(1) ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ is a cuspidal representation; \(2) every irreducible component of ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ is inside ${{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$. To prove (1), we will show that the constant terms of elements in ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ along all maximal parabolic subgroups of ${Sp}_{4mn}$ are all zero. Recall that $P^{4mn}_r = M^{4mn}_r N^{4mn}_r$ (with $1 \leq r \leq 2mn$) is the standard parabolic subgroup of ${Sp}_{4mn}$ with Levi part $M^{4mn}_r$ isomorphic to $GL_r \times Sp_{4mn-2r}$, $N^{4mn}_r$ is the unipotent radical. Take any ${\widetilde}{\xi} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, we will calculate the constant term of $FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi})$ along $P^{4mn}_r$, which is denoted by ${{\mathcal {C}}}_{N^{4mn}_r}(FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi}))$. By Theorem 7.8 of [@GRS11], $$\begin{aligned} \label{thm5equ1} \begin{split} & {{\mathcal {C}}}_{N^{4mn}_r}(FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi}))\\ = & \sum_{k=0}^r \sum_{\gamma \in P^1_{r-k, 1^k}(F) {\backslash}GL_r(F)} \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^1_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} ({\widetilde}{\xi}))(\hat{\gamma} \lambda \beta) d \lambda, \end{split}\end{aligned}$$ where $N^{4mn+2n}_{r-k}$ is the unipotent radical of the parabolic subgroup $P^{4mn+2n}_{r-k}$ of ${Sp}_{4mn+2n}$ with Levi isomorphic to $GL_{r-k} \times Sp_{4mn+2n-2r+2k}$, and it is identified with it’s image in ${\widetilde}{Sp}_{4mn+2n}$; $P^1_{r-k, 1^k}$ is a subgroup of $GL_r$ consisting of matrices of the form $\begin{pmatrix} g & x\\ 0 & z \end{pmatrix}$, with $z \in U_k$, the standard maximal unipotent subgroup of $GL_k$; for $a \in GL_j$, $j \leq 2mn+n$, $\hat{a}=\diag(a, I_{4mn+2n-2j}, a^*)$; $L$ is a unipotent subgroup, consisting of matrices of the form $\lambda = \begin{pmatrix} I_r & 0\\ x & I_n \end{pmatrix}^{\wedge}$, and $i(\lambda)$ is the last row of $x$; $\beta=\begin{pmatrix} 0 & I_r\\ I_n & 0 \end{pmatrix}^{\wedge}$; $\phi = \phi_1 \otimes \phi_2$, with $\phi_1 \in {{\mathcal {S}}}({{\mathbb {A}}}^r)$, $\phi_2 \in {{\mathcal {S}}}({{\mathbb {A}}}^{2mn-r})$; $$FJ^{\phi_2}_{\psi^1_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} ({\widetilde}{\xi}))(\hat{\gamma} \lambda \beta):= FJ^{\phi_2}_{\psi^1_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\widetilde}{\xi}))(I),$$ with $\rho(\hat{\gamma} \lambda \beta)$ denoting the right translation by $\hat{\gamma} \lambda \beta$; ${{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\widetilde}{\xi})$ is restricted to ${\widetilde}{Sp}_{4mn+2n-2r+2k}({{\mathbb {A}}})$, then we apply the Fourier-Jacobi coefficient $FJ^{\phi_2}_{\psi^1_{n-1+k}}$, taking automorphic forms on ${\widetilde}{Sp}_{4mn+2n-2r+2k}({{\mathbb {A}}})$ to $Sp_{4mn-2r}({{\mathbb {A}}})$. By the cuspidal support of ${\widetilde}{\xi}$, ${{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} ({\widetilde}{\xi})$ is identically zero, unless $r=k$ or $r-k = 2n$. When $r=k$, the corresponding term is zero, because $FJ^{\phi_2}_{\psi^1_{n-1+r}} ({\widetilde}{\xi})$ is zero, by Theorem \[thm10\]. When $r-k=2n$, the restriction of ${{\mathcal {C}}}_{N^{4mn+2n}_{2n}} ({\widetilde}{\xi})$ to ${\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}})$ is actually a vector in ${\widetilde}{\sigma}_{4(m-1)n+2n}$. Hence, $FJ^{\phi_2}_{\psi^1_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+2n}_{r-k}} ({\widetilde}{\xi}))$ is identically zero, for $0 \leq k \leq r$, because ${\widetilde}{\sigma}_{4(m-1)n+2n}$ has no nonzero Fourier coefficients $FJ_{\psi^{1}_{n-1}}$, and ${\underline}{p}({\widetilde}{\sigma}_{4(m-1)n+2n})= [(2n)^{2(m-1)}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$. So, when $r-k=2n$, the corresponding term is also zero. Therefore, we have shown that ${{\mathcal {C}}}_{N^{4mn}_r}(FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi}))$ is identically zero for any $1 \leq r \leq 2mn$, and for any ${\widetilde}{\xi} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, which implies that ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$ is a cuspidal representation. This completes the proof of (1). To prove (2), we need to show that for every irreducible component $\pi$ of ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$, (2-1) ${\underline}{p}(\pi)=[(2n)^{2m-1}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$; (2-2) $\pi$ is nearly equivalent to the residual representation ${{\mathcal {E}}}_{\Delta(\tau, m)}$; (2-3) $\pi$ has a nonzero Fourier coefficient $FJ_{\psi^{-1}_{n-1}}$. (2-1) is obvious by Theorem \[thm10\], and by Lemma 2.6 of [@GRS05] or Lemma 3.1 of [@JL13b]. (2-2) follows easily from , because the induced representation on the right hand side of is also an unramified component of ${{\mathcal {E}}}_{\Delta(\tau, m)}$. To show (2-3), as in the proof of Proposition 3.4 of [@GJS12], we need to consider the following integral $$\label{thm5equ2} \langle \varphi_{\pi}, FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi}) \rangle =\int_{[Sp_{4mn}]} \varphi_{\pi}(h){\overline}{FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{\xi})(h)} dh,$$ which is nonzero for some data $\varphi_{\pi} \in \pi$, ${\widetilde}{\xi} \in {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, since $\pi$ is an irreducible component of ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$. Assume that ${\widetilde}{\xi} = {{\mathrm{Res}}}_{s=\frac{m+1}{2}} {\widetilde}{E}(\phi_s, \cdot)$, then from , we know that the following integral is also nonzero for some choice of data: $$\label{thm5equ3} \langle \varphi_{\pi}, FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{E}(\phi_s, \cdot)) \rangle =\int_{[Sp_{4mn}]} \varphi_{\pi}(h){\overline}{FJ^{\phi}_{\psi^1_{n-1}}({\widetilde}{E}(\phi_s, \cdot))(h)} dh.$$ Then, by the unfolding in Theorem 3.3 of [@GJRS11] (take $m=2n$, $r=n$ there), the non-vanishing of the integral in implies the non-vanishing of $FJ_{\psi^{-1}_{n-1}}(\pi)$. This finishes the proof of (2), and completes the proof of the theorem. The next theorem implies that $\Psi$ is well-defined. We will prove it in the next section. \[thm6\] ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$ is irreducible, and $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi),$$ for any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$. Now by Theorems \[thm7\], \[thm5\], \[thm6\], we are able to conclude the Part (1) of Theorem \[thm1\]. \[thm11\] There is a surjective map $$\Psi: {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi) \rightarrow {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$$ $$\sigma_{4mn} \mapsto {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}).$$ Theorem \[thm6\] implies that $\Psi$ is well-defined. Theorem \[thm7\] and Theorem \[thm5\] imply that for any $ {\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, take any irreducible component $\pi$ of ${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$, which is inside ${{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, then $$\begin{aligned} & \Psi(\pi)\\ = & {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} (\pi)\\ \subseteq & {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})\\ = & {\widetilde}{\sigma}_{4(m-1)n+2n}.\end{aligned}$$ Theorem \[thm5\] also implies that ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} (\pi) \neq 0$. Since ${\widetilde}{\sigma}_{4(m-1)n+2n}$ is irreducible, actually we have ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} (\pi) = {\widetilde}{\sigma}_{4(m-1)n+2n}$. Hence $\Psi$ is surjective. Proof of Theorem \[thm6\] ========================= For any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, we know that the Eisenstein series corresponding to $${{\mathrm{Ind}}}_{P_{2n}({{\mathbb {A}}})}^{Sp_{4(m+1)n}({{\mathbb {A}}})} \tau \lvert \cdot \rvert^{s} \otimes \sigma_{4mn}$$ has a simple pole at $s=\frac{m+1}{2}$. Let ${{\mathcal {E}}}_{\tau, \sigma_{4mn}}$ be the residual representation of $Sp_{4(m+1)n}({{\mathbb {A}}})$ generated by the corresponding residues. First, we have a similar result as that in Theorem \[thm7\]. \[thm8\] (1) $${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} \circ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}}) \neq 0.$$ \(2) $${{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} \circ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})=\sigma_{4mn}.$$ The proof is very similar to that of Theorem \[thm7\]. We omit it here. Next, we have a similar result as that in Theorem 5.8 of [@GJS12]. \[thm9\] For any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, the automorphic representation ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ is square-integrable. Moreover, there is an irreducible representation ${\widetilde}{\sigma}_{4(m-1)+2n}$, which is a component of $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) \subset {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi),$$ such that the representation space of $ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ has a non-trivial intersection with the representation space of ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. We follow the constant term calculation in the proof of the Theorem \[thm5\]. Recall that $P^{4mn+2n}_r = M^{4mn+2n}_r N^{4mn+2n}_r$ (with $1 \leq r \leq 2mn+n$) is the standard parabolic subgroup of ${Sp}_{4mn+2n}$ with Levi part $M^{4mn+2n}_r$ isomorphic to $GL_r \times Sp_{4mn+2n-2r}$, $N^{4mn+2n}_r$ is the unipotent radical, and ${\widetilde}{P}^{4mn+2n}_r$ is the pre-image of $P^{4mn+2n}_r$ in ${\widetilde}{Sp}_{4mn+2n}$. Take any $\xi \in {{\mathcal {E}}}_{\tau, \sigma_{4mn}}$, we will calculate the constant term of $FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi})$ along ${\widetilde}{P}^{4mn+2n}_r$, which is denoted by ${{\mathcal {C}}}_{N^{4mn+2n}_r}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))$. By Theorem 7.8 of [@GRS11], $$\begin{aligned} \label{thm9equ1} \begin{split} & {{\mathcal {C}}}_{N^{4mn+2n}_r}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ = & \sum_{k=0}^r \sum_{\gamma \in P^1_{r-k, 1^k}(F) {\backslash}GL_r(F)} \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} ({\xi}))(\hat{\gamma} \lambda \beta) d \lambda, \end{split}\end{aligned}$$ where $N^{4mn+4n}_{r-k}$ is the unipotent radical of the parabolic subgroup $P^{4mn+4n}_{r-k}$ of ${Sp}_{4mn+4n}$ with Levi isomorphic to $GL_{r-k} \times Sp_{4mn+4n-2r+2k}$; $P^1_{r-k, 1^k}$ is a subgroup of $GL_r$ consisting of matrices of the form $\begin{pmatrix} g & x\\ 0 & z \end{pmatrix}$, with $z \in U_k$, the standard maximal unipotent subgroup of $GL_k$; for $a \in GL_j$, $j \leq 2mn+2n$, $\hat{a}=\diag(a, I_{4mn+4n-2j}, a^*)$; $L$ is a unipotent subgroup, consisting of matrices of the form $\lambda = \begin{pmatrix} I_r & 0\\ x & I_n \end{pmatrix}^{\wedge}$, and $i(\lambda)$ is the last row of $x$; $\beta=\begin{pmatrix} 0 & I_r\\ I_n & 0 \end{pmatrix}^{\wedge}$; $\phi = \phi_1 \otimes \phi_2$, with $\phi_1 \in {{\mathcal {S}}}({{\mathbb {A}}}^r)$, $\phi_2 \in {{\mathcal {S}}}({{\mathbb {A}}}^{2mn+n-r})$; $$FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} ({\xi}))(\hat{\gamma} \lambda \beta):= FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\xi}))(I),$$ with $\rho(\hat{\gamma} \lambda \beta)$ denoting the right translation by $\hat{\gamma} \lambda \beta$; ${{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\xi})$ is restricted to ${Sp}_{4mn+4n-2r+2k}({{\mathbb {A}}})$, then we apply the Fourier-Jacobi coefficient $FJ^{\phi_2}_{\psi^{-1}_{n-1+k}}$, taking automorphic forms on ${Sp}_{4mn+4n-2r+2k}({{\mathbb {A}}})$ to ${\widetilde}{Sp}_{4mn+2n-2r}({{\mathbb {A}}})$. By the cuspidal support of ${\xi}$, ${{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} ({\xi})$ is identically zero, unless $r=k$ or $r-k = 2n$. When $r=k$, the corresponding term is zero, because $FJ^{\phi_2}_{\psi^{-1}_{n-1+r}} ({\xi})$ is zero, by Theorem \[thm10\]. When $r-k=2n$, the restriction of ${{\mathcal {C}}}_{N^{4mn+4n}_{2n}} ({\xi})$ to ${Sp}_{4mn}({{\mathbb {A}}})$ is actually a vector inside ${\sigma}_{4mn}$. Hence, $FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn+4n}_{r-k}} ({\xi}))$ is not zero for $k=0$, and is identically zero for $1 \leq k \leq r$, because ${\sigma}_{4mn}$ has a nonzero Fourier coefficient $FJ_{\psi^{-1}_{n-1}}$, and ${\underline}{p}({\sigma}_{4mn})= [(2n)^{2m-1}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$. Therefore, $$\begin{aligned} \label{thm9equ2} \begin{split} & {{\mathcal {C}}}_{N^{4mn+2n}_{2n}}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ = & \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^{-1}_{n-1}} ({{\mathcal {C}}}_{N^{4mn+4n}_{2n}} ({\xi}))(\lambda \beta) d \lambda. \end{split}\end{aligned}$$ By similar calculation as in the proof of Lemma \[constantterm\], when restricting to $GL_{2n}({{\mathbb {A}}}) \times {Sp}_{4mn}({{\mathbb {A}}})$, $${{\mathcal {C}}}_{N^{4mn+4n}_{2n}} ({\xi}) \in \delta_{P^{4mn+4n}_{2n}}^{\frac{1}{2}} \lvert \det \rvert^{-\frac{2m+1}{2}} \tau \otimes \sigma_{4mn}.$$ As in the proof of Theorem 2.5 [@GJS12], we need to calculate the automorphic exponents attached to this non-trivial constant term (for definition see [slowromancap1@]{}.3.3 [@MW95]). We consider the action of $$\hat{g}=\diag(g, I_{4mn-2n}, g^*) \in GL_{2n}({{\mathbb {A}}}) \times {\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}}).$$ Since $r=2n$, $\beta=\begin{pmatrix} 0 & I_{2n}\\ I_n & 0 \end{pmatrix}^{\wedge}$. $\beta \diag(I_n, \hat{g}, I_n) \beta^{-1} = \diag(g, I_{4mn}, g^*) =: {\widetilde}{g}$. Then changing variables in $\lambda \mapsto {\widetilde}{g} \lambda {\widetilde}{g}^{-1}$ will give a Jacobian $\lvert \det(g) \rvert^{-n}$. On the other hand, by Formula (1.4) [@GRS11], the action of $\hat{g}$ on $\phi_1$ gives $\gamma_{\psi}(\det(g)) \lvert \det(g) \rvert^{\frac{1}{2}}$. Therefore, the $\hat{g}$ acts by $\tau(g)$ with character $$\begin{aligned} & \delta_{P^{4mn+4n}_{2n}}^{\frac{1}{2}} \lvert \det(g) \rvert^{-\frac{2m+1}{2}} \lvert \det(g) \rvert^{-n} \gamma_{\psi}(\det(g)) \lvert \det(g) \rvert^{\frac{1}{2}}\\ = & \gamma_{\psi}(\det(g)) \delta_{P^{4mn+2n}_{2n}}^{\frac{1}{2}} \lvert \det(g) \rvert^{-m}.\end{aligned}$$ Hence, by Langlands square-integrability criterion (Lemma [slowromancap1@]{}.4.11 [@MW95]), the automorphic representation ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ is square-integrable. And as a representation of $GL_{2n}({{\mathbb {A}}}) \times {\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}})$, $$\label{thm9equ3} {{\mathcal {C}}}_{N^{4mn+4n}_{2n}}({{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})) = \gamma_{\psi} \delta_{P^{4mn+2n}_{2n}}^{\frac{1}{2}} \lvert \det \rvert^{-m} \tau \otimes {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} (\sigma_{4mn}).$$ By , it is easy to see that any non-cuspidal irreducible subrepresentation of ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ must be an irreducible subrepresentation of ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, for some irreducible subrepresentation ${\widetilde}{\sigma}_{4(m-1)+2n}$ of ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$. To prove ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) \subset {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, we need to show that for every irreducible component $\sigma$ of ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn})$, \(1) $\sigma$ is cuspidal; \(2) ${\underline}{p}(\sigma)=[(2n)^{2m-2}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}]$; \(3) $\sigma$ is nearly equivalent to the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau, m-1) \otimes {\widetilde}{\pi}}$; \(4) $\sigma$ has no nonzero Fourier coefficient $FJ_{\psi^{1}_{n-1}}$. \(1) follows easily from the tower property (Theorem 7.10 [@GRS11]). (2) is implied by Lemma 2.6 [@GRS03] or Lemma 3.1 [@JL13b]. (3) can be read out easily from the right hand side of by of Proposition \[prop1\]. Note that the right hand side of is an unramified component of ${{\mathcal {E}}}_{\Delta(\tau, m)}$, hence unramified component of $\sigma_{4mn}$. By Theorem 5.2 of [@JL13b], as a cuspidal representation, $\sigma_{4mn}$ has no nonzero Fourier coefficient with respect to character $\psi_{[(2n)^{2m-1}(2n_1)^{s_1}(2n_2)^{s_2} \cdots (2n_k)^{s_k}], {\underline}{a}}$, where ${\underline}{a}=\{-1, 1\} \cup {\underline}{a}'$. Now (4) follows easily from Lemma 3.1 [@JL13b]. Therefore, the representation space of $ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ has a non-trivial intersection with the representation space of ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$, for some component ${\widetilde}{\sigma}_{4(m-1)n+2n}$ of $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) \subset {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi).$$ This completes the proof of the theorem. [***Proof of Theorem \[thm6\].***]{} For any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, by Theorem \[thm8\], $$\label{thm6equ1} {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} \circ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})=\sigma_{4mn}.$$ By Theorem \[thm9\], there is an irreducible representation ${\widetilde}{\sigma}_{4(m-1)+2n}$, which is a component of ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) \subset {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi),$ such that the representation space of $ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\tau, \sigma_{4mn}})$ contains an irreducible subrepresentation $\pi$ of ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Since $\sigma_{4mn}$ is irreducible, by , $$\label{thm6equ2} \sigma_{4mn} = {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}(\pi) \subset {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}).$$ Therefore, $$\begin{aligned} {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) & \subset {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{+1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})\\ & = {\widetilde}{\sigma}_{4(m-1)n+2n},\end{aligned}$$ by Theorem \[thm7\]. Hence, ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) = {\widetilde}{\sigma}_{4(m-1)n+2n}$, irreducible as an element in ${{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$. This completes the proof of Theorem \[thm6\], showing that $\Psi$ is well-defined. $\square$ Proof of Part (2) of Theorem \[thm1\] ===================================== In this section, we will prove that $\Psi$ is injective. For this, we need to assume that for any ${\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$ is irreducible. For any $\sigma_{4mn} \in {{\mathcal {N}}}_{{Sp}_{4mn}}(\tau, \psi)$, by Theorem \[thm6\], ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}(\sigma_{4mn}) = {\widetilde}{\sigma}_{4(m-1)n+2n} \in {{\mathcal {N}}}'_{{\widetilde}{Sp}_{4(m-1)n+2n}}(\tau, \psi)$, which is irreducible. To show $\Psi$ is injective, we only need to show that $\sigma_{4mn}$ is uniquely determined by ${\widetilde}{\sigma}_{4(m-1)n+2n}$. By , $\sigma_{4mn} = {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}(\pi) \subset {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$, where $\pi$ is an irreducible subrepresentation of ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Since we assume that ${\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$ is irreducible, we have that $\pi = {\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}}$. Hence $\sigma_{4mn} = {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n}({\widetilde}{{{\mathcal {E}}}}_{\tau, {\widetilde}{\sigma}_{4(m-1)n+2n}})$, which means that $\sigma_{4mn}$ is uniquely determined by ${\widetilde}{\sigma}_{4(m-1)n+2n}$. This completes the proof of Part (2) of Theorem \[thm1\]. Irreducibility of Certain Descent Representations ================================================= In Theorem \[main1part2\], for the residual representation ${{\mathcal {E}}}_{\Delta(\tau, m)}$, we have proved that $\mathfrak{p}^m({{\mathcal {E}}}_{\Delta(\tau, m)}) = [(2n)^{2m}]$. From the proof, and by Lemma 2.6 [@GRS03] or Lemma 3.1 [@JL13b], we can see that it has a nonzero Fourier coefficient attached to the partition $[(2n)1^{4mn-2n}]$ with respect to the character $\psi_{[(2n)1^{4mn-2n}], -1}$. In this section, for any number field $F$, we show that both ${{\mathcal {E}}}_{\Delta(\tau, m)}$ and ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ are irreducible. The result can be stated as follows. \[irre\] Assume that $F$ is any number field. \(1) ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ is square-integrable and is in the discrete spectrum. \(2) Both ${{\mathcal {E}}}_{\Delta(\tau, m)}$ and ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ are irreducible. **Proof of Part (1)**. As in Theorem \[thm9\], we follow the constant term calculation in the proof of the Theorem \[thm5\]. Recall that $P^{4mn-2n}_r = M^{4mn-2n}_r N^{4mn-2n}_r$ (with $1 \leq r \leq 2mn-n$) is the standard parabolic subgroup of ${Sp}_{4mn-2n}$ with Levi part $M^{4mn-2n}_r$ isomorphic to $GL_r \times Sp_{4mn-2n-2r}$, $N^{4mn-2n}_r$ is the unipotent radical, and ${\widetilde}{P}^{4mn-2n}_r$ is the pre-image of $P^{4mn-2n}_r$ in ${\widetilde}{Sp}_{4mn-2n}$. Take any $\xi \in {{\mathcal {E}}}_{\Delta(\tau, m)}$, we will calculate the constant term of $FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi})$ along ${\widetilde}{P}^{4mn-2n}_r$, which is denoted by ${{\mathcal {C}}}_{N^{4mn-2n}_r}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))$. By Theorem 7.8 of [@GRS11], $$\begin{aligned} \label{irreequ3} \begin{split} & {{\mathcal {C}}}_{N^{4mn-2n}_r}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ = & \sum_{k=0}^r \sum_{\gamma \in P^1_{r-k, 1^k}(F) {\backslash}GL_r(F)} \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn}_{r-k}} ({\xi}))(\hat{\gamma} \lambda \beta) d \lambda, \end{split}\end{aligned}$$ where $N^{4mn}_{r-k}$ is the unipotent radical of the parabolic subgroup $P^{4mn}_{r-k}$ of ${Sp}_{4mn}$ with Levi isomorphic to $GL_{r-k} \times Sp_{4mn-2r+2k}$; $P^1_{r-k, 1^k}$ is a subgroup of $GL_r$ consisting of matrices of the form $\begin{pmatrix} g & x\\ 0 & z \end{pmatrix}$, with $z \in U_k$, the standard maximal unipotent subgroup of $GL_k$; for $a \in GL_j$, $j \leq 2mn$, $\hat{a}=\diag(a, I_{4mn-2j}, a^*)$; $L$ is a unipotent subgroup, consisting of matrices of the form $\lambda = \begin{pmatrix} I_r & 0\\ x & I_n \end{pmatrix}^{\wedge}$, and $i(\lambda)$ is the last row of $x$; $\beta=\begin{pmatrix} 0 & I_r\\ I_n & 0 \end{pmatrix}^{\wedge}$; $\phi = \phi_1 \otimes \phi_2$, with $\phi_1 \in {{\mathcal {S}}}({{\mathbb {A}}}^r)$, $\phi_2 \in {{\mathcal {S}}}({{\mathbb {A}}}^{2mn-n-r})$; $$FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn}_{r-k}} ({\xi}))(\hat{\gamma} \lambda \beta):= FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\xi}))(I),$$ with $\rho(\hat{\gamma} \lambda \beta)$ denoting the right translation by $\hat{\gamma} \lambda \beta$; ${{\mathcal {C}}}_{N^{4mn}_{r-k}} (\rho(\hat{\gamma} \lambda \beta){\xi})$ is restricted to ${Sp}_{4mn-2r+2k}({{\mathbb {A}}})$, then we apply the Fourier-Jacobi coefficient $FJ^{\phi_2}_{\psi^{-1}_{n-1+k}}$, taking automorphic forms on ${Sp}_{4mn-2r+2k}({{\mathbb {A}}})$ to ${\widetilde}{Sp}_{4mn-2n-2r}({{\mathbb {A}}})$. By the cuspidal support of ${\xi}$, ${{\mathcal {C}}}_{N^{4mn}_{r-k}} ({\xi})$ is identically zero, unless $r=k$ or $r-k = 2ln$, $1 \leq l \leq m-1$. When $r=k$, the corresponding term is zero, because $FJ^{\phi_2}_{\psi^{-1}_{n-1+r}} ({\xi})$ is zero, by Theorem \[thm10\]. When $r-k=2ln$, $1 \leq l \leq m-1$, $FJ^{\phi_2}_{\psi^{-1}_{n-1+k}} ({{\mathcal {C}}}_{N^{4mn}_{r-k}} ({\xi}))$ is not zero for $k=0$, and is identically zero for $1 \leq k \leq r$, because $\mathfrak{p}^m({{\mathcal {E}}}_{\Delta(\tau, m)}) = [(2n)^{2m}]$. Therefore, ${{\mathcal {C}}}_{N^{4mn-2n}_r}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi})) \neq 0$, only for $r = 2ln$, $1 \leq l \leq m-1$. And for $1 \leq l \leq m-1$, $$\begin{aligned} \label{irreequ4} \begin{split} & {{\mathcal {C}}}_{N^{4mn-2n}_{2ln}}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ = & \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^{-1}_{n-1}} ({{\mathcal {C}}}_{N^{4mn}_{2ln}} ({\xi}))(\lambda \beta) d \lambda. \end{split}\end{aligned}$$ To prove square-integrability of ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$, it turns out we only need to consider $r=2(m-1)n$, which will be clear from the following discussion. For $r = 2(m-1)n$, $$\begin{aligned} \label{irreequ5} \begin{split} & {{\mathcal {C}}}_{N^{4mn-2n}_{2(m-1)n}}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ = & \int_{L({{\mathbb {A}}})} \phi_1(i(\lambda)) FJ^{\phi_2}_{\psi^{-1}_{n-1}} ({{\mathcal {C}}}_{N^{4mn}_{2(m-1)n}} ({\xi}))(\lambda \beta) d \lambda. \end{split}\end{aligned}$$ By Lemma \[constantterm\], when restricted to $GL_{2(m-1)n}({{\mathbb {A}}}) \times Sp_{4n}({{\mathbb {A}}})$, $${{\mathcal {C}}}_{N^{4mn}_{2(m-1)n}} ({\xi}) \in \delta_{P^{4mn}_{2(m-1)n}}^{\frac{1}{2}} \lvert \det \rvert^{-\frac{m+1}{2}} \Delta(\tau,m-1) \otimes {{\mathcal {E}}}_{\Delta(\tau,1)}.$$ As in the proof of Theorem 2.5 [@GJS12], to calculate the automorphic exponent attached to this non-trivial constant term (for definition see [slowromancap1@]{}.3.3 [@MW95]), we need to consider the action of $$\hat{g}=\diag(g, I_{2n}, g^*) \in GL_{2(m-1)n}({{\mathbb {A}}}) \times {\widetilde}{Sp}_{2n}({{\mathbb {A}}}).$$ Since $r=2(m-1)n$, $\beta=\begin{pmatrix} 0 & I_{2(m-1)n}\\ I_n & 0 \end{pmatrix}^{\wedge}$. $\beta \diag(I_n, \hat{g}, I_n) \beta^{-1} = \diag(g, I_{4n}, g^*) =: {\widetilde}{g}$. Then changing variables in $\lambda \mapsto {\widetilde}{g} \lambda {\widetilde}{g}^{-1}$ will give a Jacobian $\lvert \det(g) \rvert^{-n}$. On the other hand, by Formula (1.4) [@GRS11], the action of $\hat{g}$ on $\phi_1$ gives $\gamma_{\psi}(\det(g)) \lvert \det(g) \rvert^{\frac{1}{2}}$. Therefore, the $\hat{g}$ acts by $\tau(g)$ with character $$\begin{aligned} & \delta_{P^{4mn}_{2(m-1)n}}^{\frac{1}{2}} \lvert \det(g) \rvert^{-\frac{m+1}{2}} \lvert \det(g) \rvert^{-n} \gamma_{\psi}(\det(g)) \lvert \det(g) \rvert^{\frac{1}{2}}\\ = & \gamma_{\psi}(\det(g)) \delta_{P^{4mn-2n}_{2(m-1)n}}^{\frac{1}{2}} \lvert \det(g) \rvert^{-\frac{m}{2}}.\end{aligned}$$ Therefore, as a function on $GL_{2(m-1)n}({{\mathbb {A}}}) \times {\widetilde}{Sp}_{2n}({{\mathbb {A}}})$, $$\begin{aligned} \label{irreequ6} \begin{split} & {{\mathcal {C}}}_{N^{4mn-2n}_{2(m-1)n}}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))\\ \in & \gamma_{\psi} \delta_{P^{4mn-2n}_{2(m-1)n}}^{\frac{1}{2}} \lvert \det \rvert^{-\frac{m}{2}} \Delta(\tau,m-1) \otimes {{\mathcal {D}}}^{4n}_{2n,\psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau,1)}). \end{split}\end{aligned}$$ By Theorem 2.3 [@GJS12], we know that ${{\mathcal {D}}}^{4n}_{2n,\psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau,1)})$ is an irreducible, genuine, $\psi$-generic, cuspidal automorphic representation of ${\widetilde}{Sp}_{2n}({{\mathbb {A}}})$, which lifts to $\tau$ with respect to $\psi$. Hence, as a representation of $GL_{2(m-1)n}({{\mathbb {A}}}) \times {\widetilde}{Sp}_{2n}({{\mathbb {A}}})$, $$\begin{aligned} \label{irreequ7} \begin{split} & {{\mathcal {C}}}_{N^{4mn-2n}_{2(m-1)n}}({{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)}))\\ = & \gamma_{\psi} \delta_{P^{4mn-2n}_{2(m-1)n}}^{\frac{1}{2}} \lvert \det \rvert^{-\frac{m}{2}} \Delta(\tau,m-1) \otimes {{\mathcal {D}}}^{4n}_{2n,\psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau,1)}). \end{split}\end{aligned}$$ Since, the cuspidal exponent of $\Delta(\tau,m-1)$ is $\{(\frac{2-m}{2}, \frac{4-m}{2}, \ldots, \frac{m-2}{2})\}$, the cuspidal exponent of ${{\mathcal {C}}}_{N^{4mn-2n}_{2(m-1)n}}(FJ^{\phi}_{\psi^{-1}_{n-1}}({\xi}))$ is $\{(\frac{2-2m}{2}, \frac{4-2m}{2}, \ldots, -1)\}$. Hence, by Langlands square-integrability criterion [@MW95 Lemma [slowromancap1@]{}.4.11], the automorphic representation ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ is square-integrable and is in the discrete spectrum. This completes the proof of Part (1). **Proof of Part (2)**. The proof of irreducibility of ${{\mathcal {E}}}_{\Delta(\tau, m)}$ is similar to that of ${{\mathcal {E}}}_{\Delta(\tau, 1)}$ which is given on Page 982 of [@GJS12] and in the proof of Theorem 2.1 of [@GRS11]. To show the square-integrable residual representation ${{\mathcal {E}}}_{\Delta(\tau, m)}$ is irreducible, it suffices to show that at each local place $v$, $$\label{equirr1} {{\mathrm{Ind}}}_{P_{2mn}(F_v)}^{Sp_{2mn}(F_v)} \Delta(\tau_v, m) \lvert \cdot \rvert^{\frac{m}{2}}$$ has a unique quotient. Since $\Delta(\tau_v, m)$ is the unique quotient of the following induced representation $${{\mathrm{Ind}}}_{Q_{(2n)^m}(F_v)}^{GL_{2mn}(F_v)}\tau_v \lvert \cdot \rvert^{\frac{m-1}{2}} \otimes \tau_v \lvert \cdot \rvert^{\frac{m-3}{2}} \otimes \cdots \otimes \tau_v \lvert \cdot \rvert^{\frac{1-m}{2}},$$ where $Q_{(2n)^m}$ is the parabolic subgroup of $GL_{2mn}$ with Levi subgroup isomorphic to $GL_{2n}^m$. We just have to show that the following induced representation has a unique quotient $$\label{equirr2} {{\mathrm{Ind}}}_{P_{(2n)^m}(F_v)}^{Sp_{4mn}(F_v)} \tau_v \lvert \cdot \rvert^{\frac{2m-1}{2}} \otimes \tau_v \lvert \cdot \rvert^{\frac{2m-3}{2}} \otimes \cdots \otimes \tau_v \lvert \cdot \rvert^{\frac{1}{2}},$$ where $P_{(2n)^m}$ is the parabolic subgroup of $Sp_{4mn}$ with Levi subgroup isomorphic to $GL_{2n}^m$. Since $\tau_v$ is generic and unitary, by [@T86] and [@V86], $\tau_v$ is full parabolic induction from its Langlands data with exponents in the open interval $(-\frac{1}{2}, \frac{1}{2})$. Explicitly, we can assume that $$\tau_v \cong \rho_1 \lvert \cdot \rvert^{\alpha_1} \times \rho_2 \lvert \cdot \rvert^{\alpha_2} \times \cdots \times \rho_r \lvert \cdot \rvert^{\alpha_r},$$ where $\rho_i$’s are tempered representations, $\alpha_i \in {{\mathbb {R}}}$ and $\frac{1}{2} > \alpha_1 > \alpha_2 > \cdots > \alpha_r > -\frac{1}{2}$. Therefore, the induced representation in can be written as follows: $$\begin{aligned} & \rho_1 \lvert \cdot \rvert^{\frac{2m-1}{2}+ \alpha_1} \times \rho_2 \lvert \cdot \rvert^{\frac{2m-1}{2}+\alpha_2} \times \cdots \times \rho_r \lvert \cdot \rvert^{\frac{2m-1}{2}+\alpha_r}\\ \times & \rho_1 \lvert \cdot \rvert^{\frac{2m-3}{2}+ \alpha_1} \times \rho_2 \lvert \cdot \rvert^{\frac{2m-3}{2}+\alpha_2} \times \cdots \times \rho_r \lvert \cdot \rvert^{\frac{2m-3}{2}+\alpha_r}\\ \times & \cdots\\ \times & \rho_1 \lvert \cdot \rvert^{\frac{1}{2}+ \alpha_1} \times \rho_2 \lvert \cdot \rvert^{\frac{1}{2}+\alpha_2} \times \cdots \times \rho_r \lvert \cdot \rvert^{\frac{1}{2}+\alpha_r} \rtimes 1_{Sp_0}.\end{aligned}$$ Since $\alpha_i \in {{\mathbb {R}}}$ and $\frac{1}{2} > \alpha_1 > \alpha_2 > \cdots > \alpha_r > -\frac{1}{2}$, we can easily see that the exponents satisfy $$\begin{aligned} & \frac{2m-1}{2}+ \alpha_1 > \frac{2m-1}{2}+\alpha_2 > \cdots > \frac{2m-1}{2}+\alpha_r \\ > & \frac{2m-3}{2}+ \alpha_1 > \frac{2m-3}{2}+\alpha_2 > \cdots > \frac{2m-3}{2}+\alpha_r \\ > & \cdots \\ > & \frac{1}{2}+ \alpha_1 > \frac{1}{2}+\alpha_2 > \cdots > \frac{1}{2}+\alpha_r > 0.\end{aligned}$$ By Langlands classification, it is easy to see that the induced representation in has a unique quotient which is the Langlands quotient. This completes the proof of irreducibility of ${{\mathcal {E}}}_{\Delta(\tau, m)}$. The proof of irreducibility of ${{\mathcal {D}}}^{4mn}_{2n, \psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau, m)})$ is similar to that in Theorem \[thm6\]. We just sketch all the steps needed. Recall that $P^{4mn+4n}_{2n} = M^{4mn+4n}_{2n}N^{4mn+4n}_{2n}$ is the parabolic subgroup of $Sp_{4mn+4n}$ with Levi subgroup $M^{4mn+4n}_{2n}$ isomorphic to $GL_{2n} \times Sp_{4mn}$. For any $\phi \in A(N^{4mn+4n}_{2n}(BA)M^{4mn+4n}_{2n}(F) {\backslash}Sp_{4mn+4n}({{\mathbb {A}}})_{\tau \otimes {{\mathcal {E}}}_{\Delta(\tau, m)}}$, the corresponding Eisenstein series defined as follows has a pole at $s=\frac{m+1}{2}$: $$E(\phi,s)(g)=\sum_{\gamma\in P^{4mn+4n}_{2n}(F)\bks Sp_{4mn+4n}(F)}\lambda_s \phi(\gamma g).$$ The resulting residual representation generated by all the residues is actually ${{\mathcal {E}}}_{\Delta(\tau, m+1)}$. Then, by a similar argument as in the proof of Theorem \[thm7\], we get that $$\begin{aligned} \label{irreequ1} \begin{split} {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} \circ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\Delta(\tau, m+1)}) & \neq 0,\\ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} \circ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4(m+1)n} ({{\mathcal {E}}}_{\Delta(\tau, m+1)})& ={{\mathcal {E}}}_{\Delta(\tau, m)}. \end{split}\end{aligned}$$ Note that, as indicated at the end of the proof of Theorem \[thm7\], the irreducibility of ${{\mathcal {E}}}_{\Delta(\tau, m)}$ plays an essential role in proving the equality in . From Part (1), we see that ${{\mathcal {D}}}^{4mn}_{2n, \psi^1}({{\mathcal {E}}}_{\Delta(\tau, m)})$ is square-integrable and is in the discrete spectrum. For any irreducible component $\pi$ of ${{\mathcal {D}}}^{4mn}_{2n, \psi^1}({{\mathcal {E}}}_{\Delta(\tau, m)})$, for any $\phi \in A(N^{4mn+2n}_{2n}({{\mathbb {A}}}){\widetilde}{M}^{4mn+2n}_{2n}(F) {\backslash}{\widetilde}{Sp}_{4mn+2n}({{\mathbb {A}}}))_{\mu_{\psi}\tau \otimes \pi}$, the corresponding Eisenstein series defined as follows has a pole at $s=m$. Denote the residual representation generated by all the residues by ${\widetilde}{{{\mathcal {E}}}}_{\tau, \pi}$. Since $\pi$ is irreducible, also by a similar argument as in the proof of Theorem \[thm7\], we get that $$\begin{aligned} \label{irreequ2} \begin{split} {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, \pi}) & \neq 0,\\ {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, \pi})& =\pi. \end{split}\end{aligned}$$ Then, using a similar argument as in the proof of Theorem \[thm9\], we have that there is an irreducible component $\pi$ of ${{\mathcal {D}}}^{4mn}_{2n, \psi^1}({{\mathcal {E}}}_{\Delta(\tau, m)})$, such that the representation space of ${{\mathcal {D}}}^{4mn+4n}_{2n, \psi^1}({{\mathcal {E}}}_{\Delta(\tau, m+1)})$ has a non-trivial intersection with the representation space of ${\widetilde}{{{\mathcal {E}}}}_{\tau, \pi}$. Let $\pi'$ be an irreducible subrepresentation of ${\widetilde}{{{\mathcal {E}}}}_{\tau,\pi}$ which is in this intersection. Since ${{\mathcal {E}}}_{\Delta(\tau, m)}$ is irreducible, by the identity in we have ${{\mathcal {E}}}_{\Delta(\tau, m)} = {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} (\pi') \subseteq {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, \pi})$. Therefore, $${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}({{\mathcal {E}}}_{\Delta(\tau, m)}) \subseteq {{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn} \circ {{\mathcal {D}}}_{2n, \psi^{1}}^{4mn+2n} ({\widetilde}{{{\mathcal {E}}}}_{\tau, \pi})=\pi,$$ by . Hence, ${{\mathcal {D}}}_{2n, \psi^{-1}}^{4mn}({{\mathcal {E}}}_{\Delta(\tau, m)}) = \pi$, irreducible. This completes the proof of the theorem. \[rmk4\] Write ${\widetilde}{\pi} = {{\mathcal {D}}}^{4n}_{2n,\psi^{-1}}({{\mathcal {E}}}_{\Delta(\tau,1)})$. For $${\widetilde}{\phi} \in A(N^{4mn-2n}_{2(m-1)n}({{\mathbb {A}}}) {\widetilde}{M}^{4mn-2n}_{2(m-1)n}{\backslash}{\widetilde}{Sp}_{4mn-2n}({{\mathbb {A}}}))_{\mu_{\psi} \Delta(\tau,m-1) \otimes {\widetilde}{\pi}},$$ it is easy to see that the corresponding Eisenstein series has a simple pole at $\frac{m}{2}$. Denote the residual representation by ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}}$. From the proof of Part (1) of Theorem \[irre\] (in particular, ), it is easy to see that if the residual representation ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}}$ is irreducible, then actually we have proved that ${{\mathcal {D}}}^{4mn}_{2n, \psi^1}({{\mathcal {E}}}_{\Delta(\tau, m)}) = {\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}}$. And, with the assumption that ${\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}}$ is irreducible, using similar argument as that in Theorem \[irre\], we can also prove that ${{\mathcal {D}}}^{4mn-2n}_{2n, \psi^{-1}}({\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}})$ is irreducible, square-integrable and is in the discrete spectrum. Furthermore, since ${{\mathcal {E}}}_{\Delta(\tau, m-1)}$ is also irreducible by Theorem \[irre\], we actually have $${{\mathcal {D}}}^{4mn-2n}_{2n, \psi^{-1}}({\widetilde}{{{\mathcal {E}}}}_{\Delta(\tau,m-1), {\widetilde}{\pi}}) = {{\mathcal {E}}}_{\Delta(\tau, m-1)}.$$ [CAP]{} P. Achar, [*An order-reversing duality map for conjugacy classes in Lusztig’s canonical quotient.*]{} Transform. Groups **8** (2003), no. 2, 107–145. J. Arthur, [*The endoscopic classification of representations: Orthogonal and Symplectic groups*]{}. Colloquium Publication Vol. 61, 2013, American Mathematical Society. D. Barbasch and D. Vogan, [*Unipotent representations of complex semisimple groups.*]{} Ann. of Math. (2) **121** (1985), no. 1, 41–110. D. Ginzburg, [*Endoscopic lifting in classical groups and poles of tensor L-functions.*]{} Duke Math. J. **141** (2008), no. 3, 447–503. D. Ginzburg, [*Constructing automorphic representations in split classical groups.*]{} Electron. Res. Announc. Math. Sci. **19** (2012), 18–32. D. Ginzburg, D. Jiang, S. Rallis and D. Soudry, [*L-functions for symplectic groups using Fourier-Jacobi models.*]{} Arithmetic geometry and automorphic forms, 183–207, Adv. Lect. Math. (ALM), **19**, Int. Press, Somerville, MA, 2011. D. Ginzburg, D. Jiang and D. Soudry, [*On correspondences between certain automorphic forms on [$Sp_{4n}$]{} and [$\widetilde{Sp}_{2n}$]{}.*]{} Published online by Israel J. of Math. (2012). DOI: 10.1007/s11856-012-0058-4. D. Ginzburg, S. Rallis and D. Soudry, [*On a correspondence between cuspidal representations of [$GL_{2n}$]{} and [${\widetilde}{Sp}_{2n}$]{}.*]{} J. Amer. Math. Soc. **12** (1999), no. 3, 849–907. D. Ginzburg, S. Rallis and D. Soudry, [*On Fourier coefficients of automorphic forms of symplectic groups.*]{} Manuscripta Math. **111** (2003), no. 1, 1–16. D. Ginzburg, S. Rallis and D. Soudry, [*Contruction of CAP representations for Symplectic groups using the descent method.*]{} Automorphic representations, L-functions and applications: progress and prospects, 193–224, Ohio State Univ. Math. Res. Inst. Publ., 11, de Gruyter, Berlin, 2005. D. Ginzburg, S. Rallis and D. Soudry, [*The descent map from automorphic representations of [${\rm GL}(n)$]{} to classical groups.*]{} World Scientific, Singapore, 2011. v+339 pp. C. Jantzen, [*Reducibility of certain representations for symplectic and odd-orthogonal groups.*]{} Compositio Math. **104** (1996), no. 1, 55–63. D. Jiang, [*Integral transforms and endoscopy correspondences for classical groups.*]{} Preprint, 2012. arXiv 1212.6525. To appear in Proceedings of the Conference on Automorphic Forms and Related Geometry: Assessing the Legacy of I. I. Piatetski-Shapiro. Edited by: J. Cogdell, F. Shahidi, and D. Soudry. Contemporary Mathematics, AMS. D. Jiang and B. Liu, [*On Fourier coefficients of automorphic forms of ${\rm GL}(n)$*]{}. Int. Math. Res. Not. 2013 (17): 4029–4071. doi: 10.1093/imrn/rns153. D. Jiang and B. Liu, [*On special nilpotent orbits and Fourier coefficients for automorphic forms on symplectic groups*]{}. Submitted. 2013. D. Jiang and B. Liu, [*Arthur parameters and Fourier coefficients for automorphic forms on symplectic groups.*]{} Submitted. 2013. D. Jiang, B. Liu and L. Zhang, [*Poles of certain residual Eisenstein series of classical groups*]{}. Pacific J. of Math. Vol. **264** (2013), No. 1, 83–123 D. Jiang and D. Soudry, [*The local converse theorem for SO(2n+1) and applications*]{}. Ann. of Math., **157** (2003), 743–806. R. Langlands, [*On the functional equations satisfied by Eisenstein series*]{}. Springer Lecture Notes in Math. 544, 1976. C. Moeglin and J.-P. Waldspurger, [*Le spectre residuel de [${\rm GL}(n)$.]{}*]{} Ann. Sci. École Norm. Sup. (4) **22** (1989), no. 4, 605–674. C. Moeglin and J.-P. Waldspurger, [*Spectral decomposition and Eisenstein series.*]{} Cambridge Tracts in Mathematics, 113. Cambridge University Press, Cambridge, 1995. I. Piatetski-Shapiro, *Multiplicity one theorems*. Automorphic forms, representations and $L$-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 1, pp. 209–212, Proc. Sympos. Pure Math., XXXIII, Amer. Math. Soc., Providence, R.I., 1979. F. Shahidi, [*Eisenstein series and automorphic L-functions,*]{} volume 58 of American Mathematical Society Colloquium Publications. American Mathematical Society, Providence, RI, 2010. ISBN 978-0-8218- 4989-7. J. Shalika, *The multiplicity one theorem for ${\rm GL}_{n}$*. Ann. of Math. (2) **100** (1974), 171–193. M. Tadić, [*Classification of unitary representations in irreducible representations of general linear group (non-Archimedean case).*]{} Ann. Sci. École Norm. Sup. (4) **19** (1986), no. 3, 335–382. D. Vogan, [*The unitary dual of $GL(n)$ over an Archimedean field.*]{} Invent. Math. **83** (1986), no. 3, 449–505. J.-L. Waldspurger, [*Intégrales orbitales nilpotentes et endoscopie pour les groupes classiques non ramifiés*]{}. Astérisque 269, 2001.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Ion acceleration in the MeV range can be routinely achieved with table-top laser technology. One of the current challenges is to improve the energy coupling from the laser to the proton beam without increasing the laser peak power. Introducing nanostructures at the front target surface was shown to be beneficial for an efficient transfer of energy to the electrons. In this manuscript, we study by using full-scale three-dimensional particle-in-cell simulations and finite laser pulses, the process when a proposed optimal target with triangular nanostructures (previously found to allow 97% laser energy absorption) is used . We demonstrate that the absorbed laser energy does not depend on the dimensionality in the range of parameters presented. We also present an analytical model for laser absorption that includes deviations from the ideal conditions. This is supported by a numerical parameter study that establishes the tolerance with respect to the nanostructure size, use of different ion species, existence of preplasma, etc. We found that altering the target thickness or using different ions does not affect the absorption, but it does affect the energy redistribution among the different plasma species. The optimal configuration ($h = 1~\lambda,~ w = 0.7~ \lambda$) is robust with respect to the target fabrication errors. However, high contrast laser pulses are required, because a pre-plasma layer with a thickness on the order of 0.5 lambda is enough to lower the laser absorption by more than a $10\%$ in a non-optimal scenario.' address: - '$^1$ Grupo Photonics4Life, Departamento de Física Aplicada, Facultade de Física, Universidade de Santiago de Compostela, Campus Vida s/n, Santiago de Compostela, E15782, Spain' - '$^2$ GoLP/IPFN, Instituto Superior Tecnico, Universidade de Lisboa, Lisbon, Portugal' author: - 'M. Blanco$^1$, M.T. Flores-Arias$^1$ and M. Vranic$^2$' title: Suitability and robustness of triangular nanostructured targets for proton acceleration --- [99]{} A. Macchi, M. Borghesi, and M. Passoni. Rev. Mod. Phys., 85:751–793, 2013. H. Daido, M. Nishiuchi, and A. S. Pirozhkov. Rep. Prog. Phys., 75(5):056401, 2012. A. Stockem Novo, M. C. Kaluza, R. A. Fonseca, and L. O. Silva. Sci. Rep., 6:29402, 2016. S. C. Wilks, A. B. Langdon, T. E. Cowan, M. Roth, M. Singh, S. Hatchett, M. H. Key, D. Pennington, A. MacKinnon, and R. A. Snavely. Phys. Plasmas, 8(2):542–549, 2001. R. A. Snavely et al. Phys. Rev. Lett., 85:2945–2948, 2000. P. Mora. Phys. Rev. Lett., 90:185002, 2003. J. Fuchs et al. Phys. Rev. Lett., 94:045004, 2005. S. Zheng-Ming, W. Su-Ming, Y. Lu-Le, W. Wei-Min, C. Yun-Qian, C. Min, and Z. Jie. Chin. Phys. B, 24(1):015201, 2015. A. Bigongiari, M. Raynaud, C. Riconda, A. Héron, and A. Macchi. Phys. Plasmas, 18(10):102701, 2011. A. Andreev, N. Kumar, K. Platonov, and A. Pukhov. Phys. Plasmas, 18(10):103103, 2011. A. Andreev and K. Platonov. Contrib. Plasma Phys., 53(2):173–178, 2013. A. Andreev, K. Platonov, J. Braenzel, A. Lübcke, S. Das, H. Messaoudi, R. Grunwald, C. Gray, E. McGlynn, and M. Schnürer. Plasma Phys. Control. Fusion, 58(1):014038, 2016. M Blanco, M T Flores-Arias, C Ruiz, and M Vranic. New J. Phys., 19(3):033004, 2017. O. Klimo, J. Psikal, J. Limpouch, J. Proska, F. Novotny, T. Ceccotti, V. Floquet, and S. Kawata. New J. Phys., 13(5):053028, 2011. A. Sgattoni et al. Proc. SPIE, 8779:87790L–87790L–7, 2013. S. Kahaly, S. K. Yadav, W. M. Wang, S. Sengupta, Z. M. Sheng, A. Das, P. K. Kaw, and G. R. Kumar. Phys. Rev. Lett., 101:145001, 2008. D. Margarone et al. Phys. Rev. Lett., 109:234801, 2012. A. Zigler et al. Phys. Rev. Lett., 110:215004, 2013. A. Brantov and V. Bychenkov. Contrib. Plasma Phys., 53(10):731–735, 2013. A. Andreev, K. Platonov, J. Braenzel, A. Lübcke, S. Das, H. Messaoudi, R. Grunwald, C. Gray, E. McGlynn, and M. Schnürer. Plasma Phys. Control. Fusion, 58(1):014038, 2016. D. Margarone et al. Phys. Rev. AB, 18:071304, 2015. M. Dalui, W.-M. Wang, T. M. Trivikram, S. Sarkar, S. Tata, J. Jha, P. Ayyub, Z. M. Sheng, and M. Krishnamurthy. Sci. Rep., 5:11930 EP, 2015. A. Zigler et al. Phys. Rev. Lett., 106:134801, 2011. A. Bigongiari, M. Raynaud, C. Riconda, and A. Héron. Phys. Plasmas, 20(5):052701, 2013. T. Ceccotti et al. Phys. Rev. Lett., 111:185001, 2013. M. Raynaud, J. Kupersztych, C. Riconda, J. C. Adam, and A. Héron. Phys. Plasmas, 14(9):092702, 2007. Luca Fedeli, Arianna Formenti, Lorenzo Cialfi, Andrea Sgattoni, Giada Cantono, and Matteo Passoni. Plasma Physics and Controlled Fusion, 2017. Xia-Zhi Li, Hong-Bin Zhuo, De-Bin Zou, Shi-Jie Zhang, Hong-Yu Zhou, Na Zhao, Yue Lang, and De-Yao Yu. Chinese Physics Letters, 34(9):094201, 2017. L. Fedeli, A. Sgattoni, G. Cantono, and A. Macchi. Applied Physics Letters, 110(5):051103, 2017. A. Sgattoni, L. Fedeli, G. Cantono, T. Ceccotti, and A. Macchi. Plasma Phys. Control. Fusion, 58(1):014004, 2016. L. Fedeli et al. Phys. Rev. Lett., 116:015001, 2016. J. Kupersztych, M. Raynaud, and C. Riconda. Phys. Plasmas, 11(4):1669–1673, 2004. C. Riconda, M. Raynaud, T. Vialis, and M. Grech. Phys. Plasmas, 22(7):073103, 2015. A. Bigongiari, M. Raynaud, and C. Riconda. Phys. Rev. E, 84:015402, 2011. Seungchul Kim, Jonghan Jin, Young-Jin Kim, In-Yong Park, Yunseok Kim, and Seung-Woo Kim. Nature, 453(7196):757–760, June 2008. M Sivis, M Duwe, B Abel, and C Ropers. Nature, 485(7397):E1–E3, May 2012. N Pfullmann, C Waltermann, M Noack, S Rausch, T Nagy, C Reinhardt, M Kovačev, V Knittel, R Bratschitsch, D Akemeier, A Hütten, A Leitenstorfer, and U Morgner. New J. Phys., 15(9):093027, 2013. S Han, H Kim, Y W Kim, Y-J Kim, S Kim, I-Y Park, and S-W Kim. Nat. Commun., 7:13105, 2016. G Vampa, B G Ghamsari, S Siadat Mousavi, T J Hammond, A Olivieri, E Lisicka-Skrek, A Yu Naumov, D M Villeneuve, A Staudte, P Berini, and P B Corkum. Nat. Phys., 64:39–5, April 2017. G Vampa, B G Ghamsari, S Siadat Mousavi, T J Hammond, A Olivieri, E Lisicka-Skrek, A Yu Naumov, D M Villeneuve, A Staudte, P Berini, and P B Corkum. Nat. Phys., 64:39–5, April 2017. M. Blanco, C. Hernández-García, A. Chacón, M. Lewenstein, M. T. Flores-Arias, and L. Plaja. Opt. Express, 25(13):14974–14985, Jun 2017. E. d’Humières, A. Brantov, V. Yu. Bychenkov, and V. T. Tikhonchuk. Phys. Plasmas, 20(2):023103, 2013. A. Sgattoni, P. Londrillo, A. Macchi, and M. Passoni. Phys. Rev. E, 85:036405, 2012. R. A. Fonseca, L. O. Silva, F. S. Tsung, V. K. Decyk, W. Lu, C. Ren, W. B. Mori, S. Deng, S. Lee, T. Katsouleas, and J. C. Adam, “OSIRIS: A Three-Dimensional, Fully Relativistic Particle in Cell Code for Modeling Plasma Based Accelerators,” in *Lecture Notes in Computer Science, vol 2331,* P. M. A. Sloot, A. G. Hoekstra, C. J. K. Tan, J. J. Dongarra, eds. (Springer, Berlin, Heidelberg, 2002), pp. 342-351. R. A. Fonseca, J. Vieira, F. Fiuza, A. Davidson, F. S. Tsung, W. B. Mori, and L. O. Silva, “Exploiting multi-scale parallelism for large scale numerical modelling of laser wakefield accelerators,” Plasma Phys. Control. Fusion **55**, 124011 (2013). R. A. Fonseca, S. F. Martins, L. O. Silva, J. W. Tonge, F. S. Tsung, and W. B. Mori, “One-to-one direct modeling of experiments and astrophysical scenarios: pushing the envelope on kinetic plasma simulations,” Plasma Phys. Control. Fusion **50**, 124034 (2008). M. Allen, P.K. Patel, A. Mackinnon, D. Price, S. Wilks, and E. Morse. Phys. Rev. Lett., 93:265004, 2004. K.W.D. Ledingham, P. McKenna, T. McCanny, S. Shimizu, J.M. Yang, L. Robson, J. Zweit, J.M. Gillies, J. Bailey, G.N. Chimon. J. Phys. D: Appl. Phys. 37: 2341, 2004. S. Fritzler, V. Malka,G. Grillon, J.P. Rousseau, F. Burgy, E. Lefebvre, E. d’Humières, P. McKenna and K.W.D. Ledingham Appl.Phys.Lett. 83:3039–3041,2003. W. Luo, M. Bobeica, I. Gheorghe, D.M. Filipescu, D. Niculae, D.L. Balabanski Appl.Phys.B. 122:8, 1–11 2016. M. Hegelich, S. Karsch, G. Pretzler, D. Habs, K. Witte, W. Guenther, M. Allen, A. Blazevic, J. Fuchs, J. C. Gauthier, M. Geissel, P. Audebert, T. Cowan, and M. Roth. Phys. Rev. Lett., 89:085002, 2002. B. M. Hegelich, B. Albright, P. Audebert, A. Blazevic, E. Brambrink, J. Cobble, T. Cowan, J. Fuchs, J. C. Gauthier, C. Gautier, M. Geissel, D. Habs, R. Johnson, S. Karsch, A. Kemp, S. Letzring, M. Roth, U. Schramm, J. Schreiber, K. J. Witte, and J. C. Fernández. Phys. Plasmas, 12(5):056314, 2005. S. Busold, D. Schumacher, C. Brabetz, D. Jahn, F. Kroll, O. Deppert, U. Schramm, T.E. Cowan, A. Blažević, V. Bagnoud, and M. Roth. Sci. Rep., 5:12459, 2015. L.A. Gizzi, C. Altana, F. Brandi, P. Cirrone, G. Cristoforetti, A. Fazzi, P. Ferrara, L. Fulgentini, D. Giove, P. Koester, L. Labate, G. Lanzalone, P. Londrillo, D. Mascali, A. Muoio, D. Palla, F. Schillaci, S. Sinigardi, S. Tudisco, and G. Turchetti. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 829:144 – 148, 2016. Paul McKenna, Filip Lindau, Olle Lundh, David Neely, Anders Persson, and Claes-G[ö]{}ran Wahlstr[ö]{}m. Phil. Trans. R. Soc. Lond. A, 364(1840):711–723, 2006. J. Peebles, M.S. Wei, A.V. Arefiev, C. McGuffey, R.B. Stephens, W. Theobald, D. Haberberger, L.C. Jarrott, A. Link, H. Chen, H.S. McLean, A. Sorokovikova, S. Krasheninnikov, and F.N. Beg. New J. Phys., 19(2):023008, 2017. T.Zh. Esirkepov, J.K. Koga, A. Sunahara, T. Morita, M. Nishikino, K. Kageyama, H. Nagatomo, K. Nishihara, A. Sagisaka, H. Kotaki, T. Nakamura, Y. Fukuda, H. Okada, A.S. Pirozhkov, A. Yogo, M. Nishiuchi, H. Kiriyama, K. Kondo, M. Kando, and S.V. Bulanov. Nucl. Instr. Meth. Phys. Res. Sect. A, 745:150 – 163, 2014. C. Courtois, A. Compant La Fontaine, O. Landoas, G. Lidove, V. Méot, P. Morel, R. Nuter, E. Lefebvre, A. Boscheron, J. Grenier, M. M. Aléonard, M. Gerbaux, F. Gobet, F. Hannachi, G. Malka, J. N. Scheurer, and M. Tarisien. Physics of Plasmas, 16(1):013105, 2009. J. A. Cobble, R. P. Johnson, N. A. Kurnit, D. S. Montgomery, and J. C. Fernández. Rev. Sci. Instr., 73(11):3813–3817, 2002. N.E. Andreev, M.E. Povarnitsyn, L.P. Pugachev, and P.R. Levashov. J. Phys.: Conf. Series, 653(1):012006, 2015. K. Adumi, K. A. Tanaka, T. Matsuoka, T. Kurahashi, T. Yabuuchi, Y. Kitagawa, R. Kodama, K. Sawai, K. Suzuki, K. Okabe, T. Sera, T. Norimatsu, and Y. Izawa. Physics of Plasmas, 11(8):3721–3725, 2004. S. Kahaly, S. Monchocé, H. Vincenti, T. Dzelzainis, B. Dromey, M. Zepf, Ph. Martin, and F. Quéré. Phys. Rev. Lett., 110:175001, 2013. F. Wagner, S. Bedacht, A. Ortner, M. Roth, A. Tauschwitz, B. Zielbauer, and V. Bagnoud. Opt. Express, 22(24):29505–29514, 2014. V. Stankevič, G. Račiukaitis, F. Bragheri, X. Wang, E. G. Gamaly, R. Osellame, and S. Juodkazis. Sci. Rep., 7:39989, 2017. K. Miyazaki and G. Miyaji. J. Appl. Phys., 114(15):153108, 2013. W. Wan, W. Qiao, W. Huang, M. Zhu, Z. Fang, D. Pu, Y. Ye, Y. Liu, and L. Chen. Opt. Express, 24(6):6203–6212, 2016. L. Yuan and P. R. Herman. Nanoscale, 7:19905–19913, 2015. H. Cao, Z. Yu, J. Wang, J. O. Tegenfeldt, R. H. Austin, E. Chen, W. Wu, and S. Y. Chou. Appl. Phys. Lett., 81(1):174–176, 2002. S. K. Lynch, C. Liu, N. Y. Morgan, X. Xiao, A. A. Gomella, D. Mazilu, E. E. Bennett, L. Assoufid, F. de Carlo, and H. Wen. J. Micromech. Microeng., 22(10):105007, 2012. M.-S. Jo, K.-W. Choi, M.-H. Seo, and J.-B. Yoon. Micro Nano Syst. Lett., 5(1):19, 2017. Introduction {#sec:I} ============ Ion acceleration via laser-plasma interaction represents a promising technology for the future, and as such it has been a very active research field in the last two decades [@acceleration_Macchi; @Daido; @acceleration_several]. Ions can be accelerated efficiently in a controllable and reproducible manner into the MeV regime with table-top laser sources and metallic thin foils, via Target Normal Sheath Acceleration (TNSA) [@TNSA_1; @TNSA_2; @TNSA_mech; @TNSA_proton]. Several properties of the laser pulse and/or the target can be varied to improve different features of the accelerated proton beam, such as the total number of particles, the maximum energy and the energy conversion efficiency. It was recently shown that nanostructuring the front target surface with a periodic pattern can lead to an improved efficiency of the laser energy absorption, and therefore increase the energy of the accelerated particles [@absorption_structured; @absorption_structured_2; @structured_1; @structured_2; @structured_10; @structured_0; @structured_3; @structured_8; @structured_11; @structured_4; @structured_5; @structured_6; @andreev; @nanoSphere; @structured_9; @structured_12; @structured_13; @structured_14; @plasmon_4; @structured_15]. Theoretical models to explain this improved proton acceleration were proposed [@structured_1; @structured_2; @structured_10; @structured_0] and further parameter studies revealed that the geometry of the structure is a key factor to be taken into account [@structured_1; @structured_2; @structured_10; @structured_0; @structured_3; @structured_8; @absorption_structured; @absorption_structured_2; @structured_11]. Apart from ion acceleration, the nanostructured targets were also used to explore new paths for controlling high harmonic generation (HHG) and the associated attosecond pulse production [@structured_15; @HHG_grating_1; @HHG_grating_2]; to explore electron acceleration through surface plasma waves (SPW) [@plasmon_1; @plasmon_2; @plasmon_3; @plasmon_4; @plasmon_5; @plasmon_6] or as a nanoantenna array to locally enhance the intensity of low power laser pulses [@nanoAntenna_1; @nanoAntenna_2; @nanoAntenna_3; @nanoAntenna_4; @nanoAntenna_5; @nanoAntenna_6; @nanoAntenna_7]. The variety of recent studies dealing with periodic nanostructures illustrates that their interaction with high-power table-top laser sources is of interest for several research communities. Previous numerical and theoretical studies of TNSA proton acceleration using nanostructured targets assumed ideal conditions: perfect nanostructures, no existence of pre-plasma and no laser intensity variations. Some aspects of the interaction are bound to change when one uses real targets, as the fabrication techniques have a limited precision. In addition, the laser pre-pulse generates a pre-plasma and the laser peak intensity on-target could fluctuate. Furthermore, many studies were based on 2D geometry, while it is known that removing one spatial dimension leads to an overestimation of the final proton energy [@2D3D; @2D3D_2]. To our knowledge, studies addressing how deviations from ideal conditions affect the outcome of TNSA with nanostructured targets are still missing. The aim of this article is to establish how robust the experimental configurations using nanostructured targets are with respect to differences in fabrication, preplasma and laser intensity fluctuations. The final purpose is to give a more realistic estimate on the requirements of the target manufacturing and the laser contrast that both affect the final cost of obtaining the energetic ions. In a previous work published in Ref. [@structured_0], some of the authours have shown through an analytical model and numerical simulations that there is an optimal configuration for laser absorption with ideal triangular nanostructured targets. In the present manuscript, we perform full-scale 3D particle-in-cell (PIC) simulations for the optimal conditions to give a plausible estimate on the accelerated proton energy. We show that by using the optimal target, ions with tens of MeV and charge on the level of nC can be obtained using table-top laser systems $(I \sim 10^{19}~$W/cm${}^2)$. We then proceed to analyze how sensitive these results are to changes in the target nanostructure and laser parameters via a series of 2D PIC simulations. We also extend our analytical model from Ref. [@structured_0] to predict the laser absorption in the conditions where the nanostructures on the target front surface deviate from the ideal periodic configuration. Our study reveals that the high-level of laser absorption is robust with respect to target fabrication used with the state-of-the-art techniques , but it is sensitive to preplasma and therefore laser contrast is the most restrictive requirement. This paper is structured as follows: the results arising from full-scale 3D PIC simulations are presented in section \[sec:II\], the analysis of the effect of non-ideal conditions on the laser energy absorption is shown in section \[sec:III\]. We consider the presence of different ion species in the target, the variation of the laser pulse peak intensity, the existence of a pre-plasma and the presence of irregularities in the periodicity of the nanostructures. Finally, conclusions are presented in section \[sec:IV\]. All the simulations were performed with the PIC code OSIRIS [@osiris1; @osiris2; @osiris3]. Proton acceleration via TNSA in optimal nanostructured targets {#sec:II} ============================================================== In this section we use an ideal target design to model the TNSA acceleration process for nanostructured targets with triangular nanostructures. The majority of the previous numerical studies of TNSA are performed using a two-dimensional geometry since 3D PIC modeling requires computational capabilities on the order of millions of CPU hours. Two-dimensional simulations can provide a valid qualitative description of the laser-plasma interaction. However, due to the elimination of one spatial dimension, the accelerating field at the rear surface decays at a slower rate spatially compared to a realistic 3D geometry, thus yielding to an overestimation of the proton energy. To tackle this problem and provide an estimate of what can be expected in an experiment, we perform 3D simulations for the optimal target design found in Ref. [@structured_0], with a triangular structure height and width of $h=\lambda$ and $w=0.7\lambda$. Our results show that the optimal configuration retains the previously predicted enhancement in laser absorption even in a full-scale 3D geometry using a finite Gaussian laser pulse. The setup consists of a laser pulse impinging normally onto the structured surface of a solid target. The laser parameters correspond to the STELA laser at the *L2A2* facility at the *University of Santiago de Compostela* (USC): a Ti:Sapphire laser $(\lambda = 800~$nm$)$ with a peak power of $45~$TW and a high contrast $(> 10^{10} ~$at$~ 5 ~$ps$)$. We consider a dimensionless peak amplitude of $a_0=4$, that corresponds to an intensity of $3.46 \times 10^{19} ~$W/cm$^2$. We use targets with bulk thickness of $1$, $2$ and $3~\mu$m, respectively. The targets are made of electrons and heavy ions with a number density of $n = 90 n_c$, where $n_c = \frac{m_e \omega^2}{4\pi q_e^2}$ is the critical plasma density. The assumed ion mass-to-charge ratio is 9 times that of a proton, and it means to represent, on average, the mass of the different compounds that would be present in a real scenario, all with different ionization levels. Behind the ion plasma, there is a thin layer of protons and electrons with a thickness of $0.15\lambda$ [@TNSA_contamination_3], where $\lambda$ is the laser wavelength. The plasma density has a steep profile. Boundary conditions are open for particles and fields in the direction of laser propagation and periodic in the transverse directions. The simulation box is $26.7\lambda$ wide and $52.5\lambda$ long. The spatial resolution is $\delta \sim 0.008\lambda$ in all directions, which corresponds to $126$ cells per wavelength. The number of particles per cell is $12$ per species. At $t=0$, the laser pulse front is nearly touching the tip of the nanostructures. The laser has a $\sin^2$ temporal envelope, with a FWHM of $25~$fs. The transverse profile is Gaussian, focused on a $5~\mu$m spot (FWHM). The laser pulse is linearly polarized in the transverse plane, such that it is always p-polarized in relation to the nanostructures. The simulation advances in timesteps of $\sim 0.004 T$, where $T$ is the laser field period. The reflected energy is measured right after the interaction finishes ( $t = 18.75 T$), by integrating the reflected field energy density. The simulation proceeds until the accelerated protons reach a constant cutoff energy, at the time $t=238.7 T$. Energy conservation has been verified throughout all the simulations. Figure \[fig3D:fig1\] depicts the acceleration process in 3D, by showing the spatial distribution of charge density for each particle species in the target. Panel a) shows the target with initial triangular nanostructures before the interaction, while panel b) shows the final stage of the acceleration, where the protons initially located at the rear of the target have been accelerated. ![Scheme of the particle species distribution inside the simulation box at the a) initial instant $(t=0)$ and b) final stage of the acceleration process $(t=1200/\omega)$. The yellow isosurface corresponds to the electron density at $0.01n_c$, the blue isosurface represents ions with a density of $10n_c$ and the proton isosurface at a density of $0.02n_c$ is shown in red.[]{data-label="fig3D:fig1"}](fig1.eps){width="100.00000%"} We have performed 3D simulations for three different target thicknesses ($1~\mu$m, $2~\mu$m, and $3~\mu$m). Figure \[fig3D:fig2\] compares the properties of the accelerated proton beam for the 3D simulations and their 2D counterparts. It is observed that increasing the target thickness causes a decrease in the energy of the obtained protons, being detrimental for the acceleration process. For the target with a bulk thickness of $1~\mu$m, it is possible to obtain protons with a maximum energy of $12.8~$MeV. In this case, the total charge being accelerated above $3~$MeV is $1.04~$nC, that corresponds to $6.5\cdot 10^{9}$ protons. Although the phasespaces for 2D and 3D simulations exhibit similar features, the energies achieved in the 2D case are significantly higher that those in 3D. Using a different target thickness affects the final proton energy more substantially in 3D geometry. ![PIC simulations of targets with different thicknesses with an optimally nanostructured front: a) proton energy spectra for several targets; b) ,e) proton spatial distribution, c) ,f) proton momentum space $p_1-p_2$; d),g) proton beam energy distribution as a function of $x_1$. Panels b)-d) correspond to 3D simulations, while panels e)-g) correspond to 2D simulations. All phasespaces b) - g) correspond to the $1~\mu$m thick target.[]{data-label="fig3D:fig2"}](results_temp.eps){width="100.00000%"} In all the cases mentioned above (the three chosen target bulk thicknesses, 2D and 3D geometry), $\sim 97\%$ of laser energy is absorbed into the target. The agreement between the 2D and 3D results for the laser absorption is expected, as the absorption is controlled by the motion of electrons across the vacuum gaps in the laser field polarization plane. This is intrinsically a 2D problem, as the grating structures possess a translational symmetry in the third spatial direction and the relevant electron dynamics responsible for the enhanced absorption is well described in 2D. It is also not surprising that the bulk thickness does not affect the laser absorption, as changing the thickness does not change the geometry of the front surface. The thickness of the target, however, affects the redistribution of the absorbed energy among the different particle species. A summary of the proton beam properties for each one of the thickness target used in the example is shown in table \[tab3D:tab1\], where two main characteristics can be observed. On one hand, the maximum energy of the protons and the proton count above $3~$MeV, are strongly affected by the bulk thickness of the target. A decrease of $\sim 50 \%$ on the maximum energy is observed for the target $3~\mu$m thick in comparison with that of $1~\mu$m. On the other hand, the proton temperature and beam divergence are almost unchanged. This hints that the change of the bulk target thickness does not significantly alter the acceleration mechanism, but affects the efficiency by changing the way that electrons propagate through the target and distribute themselves at the target rear surface to accelerate the protons. The proton temperature is calculated by fitting the most energetic part of the spectrum to a Maxwell-Boltzmann distribution, and the beam divergence is obtained by calculating the average of $\langle\theta\rangle = \langle\tan^{-1} (p_{\perp}/p_{\parallel})\rangle$, where the subindexes indicate the perpendicular and parallel direction with respect to the laser propagation, that is $p_{\parallel} \equiv p_1$ and $p_{\perp} \equiv (p_2^2+p_3^2)^{1/2}$. This equation is applied for the $20\%$ of the most energetic particles in the beam. Thickness Cutoff energy Charge Temperature Divergence ------------ --------------- ----------- ------------- ---------------- $1$ $\mu$m $12.83~$MeV $1.04~$nC $5.31~$MeV ${6.04}^\circ$ $2$ $\mu$m $9.95~$MeV $0.20~$nC $5.67~$MeV ${6.51}^\circ$ $3$ $\mu$m $6.72~$MeV $0.04~$nC $4.55~$MeV ${7.21}^\circ$ : Properties of the accelerated proton beam for target thicknesses of figure 2a). These results are for a target with triangular nanostructures with height $h=\lambda$ and width $w = 0.7\lambda$.[]{data-label="tab3D:tab1"} In summary, 3D PIC simulations show that using a table-top laser source interacting with an optimal target design, TNSA proton beams with energies in the tens of MeV range can be obtained without increasing the laser peak intensity above $3.5 \cdot 10^{19}~$W/cm${}^2$. This is a promising result for applications such as the production of more specific tracers for Positron Emission Tomography (PET) such as ${}^{18}$F or ${}^{11}$C, that require protons with energies around $10~$MeV, according to their cross section[@radioisotope_1; @radioisotope_2; @radioisotope_3]. However, to move in this direction, one needs to consider how the acceleration changes when the conditions are less than ideal. The effect of deviations from ideal conditions to the laser energy absorption {#sec:III} ============================================================================= The previous section demonstrated that by using an optimal structured target, one can exploit 97 % of the laser energy for proton acceleration and obtain tens of MeV with table-top laser systems. In this section, we analyze how the deviations from the optimal configuration depicted in Ref. [@structured_0] affects the laser energy absorption. The optimal conditions were obtained in Ref. [@structured_0] using a purely electron-proton target. Here we focus on solid targets with different ion species, we vary the peak intensity of the laser pulse, consider the existence of a pre-plasma region and the effect of non-perfectly periodic nanostructures. A composite target with different ion species {#sec:31} --------------------------------------------- In a typical TNSA scenario the targets are manufactured from different materials that do not necessarily contain hydrogen. The accelerated protons then come from a thin layer at the target rear surface, originated from contamination with atmospheric hydrogen [@acceleration_Macchi; @Daido]. The heavier ions from the target can also be accelerated if the contamination layer of hydrogen is removed [@TNSA_contamination_1; @TNSA_contamination_2; @TNSA_contamination_3]. Typical inexpensive metallic foils are formed of a mixture of elements that usually come as derivatives of the fabrication process, where the main element appears in a higher concentration than others. It is, therefore, important to address the effect of having a different ion species on the laser energy absorption. As the energy absorption in nanostructured targets is controlled by the electron motion within the structure vacuum gaps, we expect that the energy absorption cannot be significantly affected by changing the ion species if the ion motion is negligible on the timescale of the interaction of the laser pulse with the target. Peak field corresponding to the laser intensity $(I \sim 10^{19}~$W/cm${}^2)$ applied without interruptions cannot displace protons further than $0.1~\mu$m from their initial position during 30 fs (which is the laser pulse duration). As the particles do not feel the peak field at all times, and the field periodically reverses its polarity, the actual displacement is even smaller. Heavier ions are slower than protons, so the ion motion begins to be significant only after the laser has been reflected or absorbed by the target. The target design is meant to resemble that of the simulations performed in Ref. [@structured_0]. The targets are made of electrons and protons with a number density of $n = 40 n_c$, with a bulk thickness of $0.5 \lambda$, and 25 particles per cell per species. The density has a steep profile, as here we consider that the laser pulse has a contrast good enough to avoid the existence of a pre-plasma. The simulations are performed in slab geometry, with open boundary conditions in the longitudinal limits of the simulation box and periodic boundary conditions in the transverse direction. The dimensions of the box are: the width of $7\lambda$ and the length of $25.5\lambda$, with a spatial resolution of $\delta \sim 0.003\lambda$ in both axes, which corresponds to $314$ cells per wavelength. At the initial instant, the laser pulse is located at a distance of $0.6\lambda$ away from the target. The laser pulse propagates towards the target with a timestep of $\sim 0.0016 T$, where $T$ is the laser field period. The laser pulse has a total temporal width of 7 laser periods, with a pulse peak amplitude and polarization equal to the ones employed in the previous section. The reflected laser energy is measured right after the interaction finishes, at the time $t = 9.5 T$. We considered several different composite targets, varying the charge density and considering fully ionized ion species. Table \[tab:tab1\] shows the absorption levels for three targets composed of electrons and different ions: hydrogen, aluminum and copper. Each configuration is considered with a plasma density of 40, 80 and $120~ n_c$, respectively. Simulation results show that the laser energy absorption is around 96 % for all the cases, with a maximum deviation of $2 \%$. At the lowest plasma density of $40~n_c$, Hydrogen shows the largest deviation (2 %) from the results obtained with heavier ions. However, as the density of the plasma becomes higher, the difference in the absorption reduces to values below $1\%$. According to these results, the different mixtures of ion species in composite targets are therefore not likely to change the laser absorption levels for structured targets. --------------------------- ---------- ---------- ---------- Ion species $\rightarrow$ Hydrogen Aluminum Copper Density $\downarrow$ ($Z=1$) ($Z=13$) ($Z=29$) $40 n_c$ $94.8\%$ $95.9\%$ $96.0\%$ $80 n_c$ $96.1\%$ $96.8\%$ $96.8\%$ $120 n_c$ $95.9\%$ $96.5\%$ $96.6\%$ --------------------------- ---------- ---------- ---------- : Absorption percentages for plasmas composed of electrons and different ion species at different densities. These results are for a target with triangular structures with height $h=\lambda$ and width $w = 0.7\lambda$.[]{data-label="tab:tab1"} Variations in the laser intensity --------------------------------- A variation in the peak laser intensity can originate from fluctuations between different laser shots and laser defocusing. Due to this deviation, experiments of TNSA proton acceleration show discrepancies between different shots even in simple configurations with planar targets [@structured_11; @fluctuation]. As typical fluctuations differ among different laser systems, we will analyze cases from small to substantial variations in laser intensity to cover a wide range of possibilities. In the considered optimal target design where the structure height is $h=\lambda$ and width is $w=0.7\lambda$, we can observe that a $\pm30\%$ variation in the peak laser amplitude translates into a maximum $\sim 1-2 \%$ variation in the relative absorption, as can be confirmed in Fig. \[fig:fig1\]. The simulation parameters considered are the same as in the previous subsection. This demonstrates that fluctuations in the laser intensity (in otherwise identical conditions) do not significantly alter the fraction of the transmitted laser energy. As in an experiment the total energy in a laser system is fixed, the expected value of the nanostructure-induced laser absorption enhancement is robust with respect to small fluctuations in intensity at the focal plane. ![Laser energy absorption for different laser intensities, sweeping a $\pm 30 \%$ variation around the original laser field amplitude of $a_0=4$.[]{data-label="fig:fig1"}](figure1.eps){width="50.00000%"} The existence of pre-plasma --------------------------- Pre-plasmas can be created through expansion of electrons and ions into vacuum, caused by the interaction with a pre-pulse or an energy pedestal before the main laser pulse hits the target. In the pre-plasma region, plasma density rises from zero to the maximum density gradually, which can result in a different interaction with a laser pulse compared to a steep density profile characteristic for the solid targets, as considered in previous sections. TNSA particle acceleration using flat targets was previously shown to be sensitive to the existence of pre-plasmas [@preplasma_8; @preplasma_3]. However, in the case of structured targets there are still only a few studies addressing its effect [@structured_3; @structured_1; @structured_2; @structured_0]. These studies have indicated that in order to obtain high levels of laser energy absorption with nanostructures (and the increased final proton energy accordingly), the contrast of the laser pulse must be high enough to avoid the existence of a long pre-plasma that would destroy the structures before the arrival of the main laser pulse. It is not clear from previous studies whether specific features of the nanostructures can mitigate the effect of pre-plasma on the high levels of energy absorption. In this subsection we analyze how the energy absorption depends on the scale-length of the pre-plasma for different structure dimensions, and we show that some structures may be more affected than others by the existence of the preplasma. The main simulation parameters for this subsection are the same as in the previous subsection, apart from the corresponding pre-plasma. We have assumed the existence of an exponential pre-plasma profile [@preplasma_1; @preplasma_2; @preplasma_3; @preplasma_4; @preplasma_5; @preplasma_6; @preplasma_7] at the target front surface and studied its effect on the laser energy absorption for different lengths. The equation that describes the density profile is given by $n = n_0 \left[ \exp \left(\log(2) \frac{x - (x_0 - L)}{L}\right) - 1 \right]$, where $n_0$ is the plasma bulk density. The parameter $L$, called scale-length, determines the length of the pre-plasma. This equation is defined for the spatial interval $x \in [x_0 - L,~ x_0]$, where $x$ is the spatial coordinate and $x_0$ is the position of the front of the undisturbed target. Figure \[fig:fig2\] shows how the presence of a short pre-plasma affects the laser energy absorption for targets with different sizes of nanostructures. For the optimized structure with height and width of $h=\lambda$ and $w=0.7\lambda$, the existence of a pre-plasma with a scale-length of $0.5\lambda$ causes a decrease of a $\sim 6 \%$ in the energy absorption. If the structure differs from the optimal, and especially if the triangles have a lower height, the effect of the pre-plasma becomes more prominent. This is evident for a structure with height $h = 0.2\lambda$ and width $w = 0.35\lambda$, where a decrease of a $\sim 20\%$ in the absorption percentage is found for the same scale-length of $0.5\lambda$. ![Effect of the pre-plasma on laser absorption. Each panel corresponds to a different structure height a) $h=0.2\lambda$, b) $h=0.5\lambda$ and c) $h=\lambda$. Diamonds correspond to the optimal structure width. Targets with smaller structures experience a higher sensitivity to the existence of pre-plasma.[]{data-label="fig:fig2"}](figure2.eps){width="\textwidth"} The results shown in Fig. \[fig:fig2\] demonstrate that to preserve the enhanced absorption of these structured targets, especially for a non-optimal design, very high contrast pulses are needed. Taking as a reference the expansion of pre-plasmas in flat targets [@preplasma_9; @preplasma_10] and assuming that in these structured targets the expansion will be similar, we can conclude that a laser pulse with an ASE contrast higher than $\sim 10^{10}$, and a pre-pulse with a peak intensity not higher than $\sim 10^{16}~$Wcm${}^{-2}$ at a distance shorter than $\sim 5~$ps would be needed to have a pre-plasma shorter than $\sim 0.2\lambda$. Below this scale-length, provided that the target design is according to the optimal conditions, the effects of preplasma on the energy absorption are below a $3\%$. Non-periodic structures ----------------------- The last challenge addressed in this study is related to the non perfect periodicity of the nanostructures. This will set a requirement for the precision of its fabrication, and therefore for the fabrication costs of these targets. The fabrication of periodic nanostructured targets can be achieved by several methods, such as laser direct writing [@fabrication_laser_1; @fabrication_laser_2], lithographic methods [@fabrication_lithography_1; @fabrication_lithography_2; @fabrication_lithography_3; @fabrication_lithography_4] or chemical etching [@fabrication_etching_1; @fabrication_lithography_4]. The precision, velocity and costs of fabrication vary substantially depending on the chosen fabrication method and the materials employed. State of the art techniques allow the fabrication of the targets studied here with high precision. However, we aim to address what would happen if there were deviations form a perfect structure in the fabrication process, as may happen with more affordable fabrication techniques. To analyze the effect of irregularities in the structure periodicity, we consider targets where each triangle could, in principle, be different than its immediate neighbors. We generate such targets starting from an optimal configuration by randomly selecting the deviation in size for every individual triangle of the target. This process is controlled by a variation percentage, that sets the maximum allowed deviation from the ideal height or width for each triangle. For example with a variation percentage of $50\%$ and for an ideal structure height of $h$, all the triangles on the target would have a random height in the interval $[0.5h,~1.5h]$. Figure \[fig:fig3\] illustrates this. ![Examples of structured targets. a) Ideal optimal structure with height and width of $h=\lambda$ and $w = 0.7\lambda$. b) An example of a target where each individual triangle was initialized with a random deviation from the optimal value (in this example, the maximum deviation is $50\%$). c) Same as b), but the random deviation up to $50\%$ is applied to the height of each triangle. The dotted lines indicate the shape of the original optimal design of triangles centered at the same position as the ones in the current target design, to facilitate comparison.[]{data-label="fig:fig3"}](figure3.eps){width="\textwidth"} Such targets are bound to have different local absorption levels than the original, optimal target. An analytical estimate for laser absorption in such a target can be obtained by extending our model presented in Ref. [@structured_0], where the equations can be tailored in order to address the local changes in geometry. The starting equations to consider are the momentum and displacement of an electron under the influence of a laser field with amplitude $a_0$: $$\begin{aligned} p_1 = m_e c \frac{a_0^2}{2} \sin^2(\varphi) & {~~~~} & p_2 = m_e c a_0 \sin(\varphi) \\ \Delta x_1 = \lambda \frac{a_0^2}{8 \pi} \left(\varphi - \frac{\sin(2 \varphi)}{2} \right) & {~~~~} & \Delta x_2 = \lambda \frac{a_0}{\pi} \sin^2\left(\frac{\varphi}{2}\right) \end{aligned}$$ where the labels $1$ and $2$ refer to the longitudinal and transverse dimensions, respectively, and $\varphi$ is the field phase. The relation between the initial height of an electron at the triangle surface and the oscillation phase at the position of reentry into the target is given by equation if the particle lands on a neighboring triangle ($\Delta x_1 \geq w / 2$ when $\Delta x_2 = h$) or by equation in case the electron reenters the target at the same triangle it originated from ($\Delta x_1 < w / 2$ when $\Delta x_2 = h$). $$\begin{aligned} h_0 = \frac{\Delta x_1}{2} + \frac{h}{w} \Delta x_2 \label{eq:004} \\ \Delta x_1 = \Delta x_2 \frac{2 h}{w} \label{eq:005} \end{aligned}$$ If we assume that the electrons carrying the laser field energy are the ones from the top of the structure ($h_0 = h$), which are the first to feel the laser field, then we solve the previous equations to obtain the phase at reentry, use this information to calculate the momentum at reentry and thus the energy they carry. By comparing that energy with a possible maximum energy attainable by the particle in a plane wave (corresponding to a reentry phase of $\varphi = \pi / 2$), one can obtain the absorbed fraction of the laser energy. This is summarized in the following equation: $$A = \frac{E(\varphi)}{E(\pi / 2)} = \left(\sqrt{1 + \frac{p_1^2(\varphi) + p_2^2(\varphi)}{(m_e c^2)^2}} - 1\right) \frac{2}{a_0^2} \label{eq:006}$$ In the case that contiguous triangles have different heights and widths, equation would remain the same, however equation would be rewritten as: $$h_0 = \frac{h_A w_B}{h_B w_A} \frac{1}{\frac{h_A w_B}{h_B w_A} + 1} \left[\Delta x_1 + \frac{2 h_A}{w_A} \Delta x_2 \right] \label{eq:007}$$ where indexes $A$ and $B$ indicate the departure and arrival triangle, respectively. It is straightforward to verify that if $h_A=h_B=h$ and $w_A=w_B=w$, equation is recovered. Table \[tab:tab2\] shows a direct comparison between this model and PIC simulations, for four different targets. The ideal target design from which the random deviations were applied has triangles with height $h=\lambda$ and width $w = 0.7\lambda$, respectively. Maximum variation Model PIC ------------------- --------------- --------------- $30 \%$ (height) $~~94.98\%~~$ $~~94.70\%~~$ $30 \%$ (width) $~~95.61\%~~$ $~~93.89\%~~$ $60 \%$ (height) $~~95.03\%~~$ $~~93.96\%~~$ $60 \%$ (width) $~~94.51\%~~$ $~~94.21\%~~$ : Laser absorption predicted by the analytical model and PIC simulations (for the exact same target) for four different examples in which each individual triangle size is randomly selected within the limits of the maximum allowed variation. The original (i.e., for ideal target before the random variation) height and width of the structures are $h=\lambda$ and $w = 0.7\lambda$, respectively.[]{data-label="tab:tab2"} Fig. \[fig:fig4\_5\] shows the results obtained with the analytical model taking as reference the optimized structure with dimensions $h=\lambda$ and $w=0.7\lambda$. For each variation percentage, $200$ random structures were generated and the laser absorption was computed. These individual results can be grouped to generate a probability density (panels a) and b)), that illustrates the probability for absorbing a specific fraction of the laser energy as a function of tolerance on the target fabrication (given by the maximum allowed structure size variation). The model is verified with a series of PIC simulations (main simulation parameters are given in section \[sec:31\]) whose results are shown in Fig. \[fig:fig4\_5\]c)-d). For each value of maximum allowed variation, 10 randomly generated structured targets were considered. The results from PIC simulations are compared with the analytically predicted average and $3 \sigma$ intervals, obtained from the analytical model predictions of the probability density above. There is an agreement between the simulation data and the model predictions. ![Results of the analytical calculations and PIC simulations for structures versus variation in height and width. The panels show the a)-b) probability density for obtaining a specific laser absorption percentage for a given maximum allowed percentage of variation of the a) width and b) height. This is accomplished by initializing 200 random configurations for each value of maximum individual size variation, and then applying the equation to evaluate the expected laser absorption. The c)-d) average expected absorption and $3\sigma$ intervals predicted by the analytical model are shown along with the laser energy absorption obtained from PIC simulations for different variations of the c) width and d) height.[]{data-label="fig:fig4_5"}](figure4_5.eps){width="80.00000%"} The results shown in this section prove the robustness of these targets for energy absorption in terms of the fabrication techniques, both from PIC simulations and analytical calculations. It can be observed that even with a $50 \%$ variation in height and width, the average decrease in the energy absorption is below $10\%$. This suggests that it is possible to use cheap fabrication techniques and relax the precision requirements for the fabrication of these targets, thus reducing the potential cost of an experiment. Conclusions {#sec:IV} =========== Triangular nanostructured targets for TNSA proton acceleration represent a very promising candidate for obtaining high energy proton sources using the table-top laser technology, as shown by the full-scale 3D simulations. An enhanced laser absorption is associated with a more efficient proton acceleration in otherwise identical conditions. The enhanced laser absorption in nanostructured targets, has proven to be very robust with respect to deviations of the topography of the structures, in terms of regularity and homogeneity from ideal fabrication. On one hand, the effect of several factors that occur in a realistic setup, such as different ion species, variations of the laser peak intensity or defects in the regularity of the nanostructures, are found not to significantly affect the absorption for structures with height and width that we previously found to be optimal. On the other hand, it has been confirmed that in order to obtain the high absorption percentages, high contrast laser pulses must be used, especially if the target design is not optimized. These results offer a strong motivation to use nanostructured targets in experiments and for industrial applications, as the efficiency of TNSA can be reliably increased without a notable increase in the cost of the experiment. Funding {#funding .unnumbered} ======= Xunta de Galicia (ED431E 2018/08) and Xunta de Galicia/FEDER (ED431B 2017/64). Spanish Ministry of Economy and Competitiveness (MINECO) (MAT2015-71119-R, FIS2015-71933-REDT). Spanish Ministry of Education, Culture and Sports (MECD) (FPU14/00289). Portuguese Foundation for Science and Technology (FCT - SFRH/BPD/119642/2016). Laserlab-Europe (EU-H2020 654148). Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to acknowledge the OSIRIS Consortium, consisting of UCLA and IST (Lisbon, Portugal) for the use of OSIRIS, for providing access to the OSIRIS framework. The authors thankfully acknowledge the computer resources at *Mare Nostrum 4* (RES-FI-2017-3-0029)
{ "pile_set_name": "ArXiv" }
--- abstract: 'We perform a Bayesian analysis of the mass distribution of stellar-mass black holes using the observed masses of 15 low-mass X-ray binary systems undergoing Roche lobe overflow and five high-mass, wind-fed X-ray binary systems. Using Markov Chain Monte Carlo calculations, we model the mass distribution both parametrically—as a power law, exponential, gaussian, combination of two gaussians, or log-normal distribution—and non-parametrically—as histograms with varying numbers of bins. We provide confidence bounds on the shape of the mass distribution in the context of each model and compare the models with each other by calculating their relative Bayesian evidence as supported by the measurements, taking into account the number of degrees of freedom of each model. The mass distribution of the low-mass systems is best fit by a power-law, while the distribution of the combined sample is best fit by the exponential model. This difference indicates that the low-mass subsample is not consistent with being drawn from the distribution of the combined population. We examine the existence of a “gap” between the most massive neutron stars and the least massive black holes by considering the value, $M_{1\%}$, of the 1% quantile from each black hole mass distribution as the lower bound of black hole masses. Our analysis generates posterior distributions for $M_{1\%}$; the best model (the power law) fitted to the low-mass systems has a distribution of lower-bounds with $M_{1\%} > 4.3$ ${M_\odot}$ with 90% confidence, while the best model (the exponential) fitted to all 20 systems has $M_{1\%} > 4.5$ ${M_\odot}$ with 90% confidence. We conclude that our sample of black hole masses provides strong evidence of a gap between the maximum neutron star mass and the lower bound on black hole masses. Our results on the low-mass sample are in qualitative agreement with those of @Ozel2010, although our broad model-selection analysis more reliably reveals the best-fit quantitative description of the underlying mass distribution. The results on the combined sample of low- and high-mass systems are in qualitative agreement with @Fryer2001 although the presence of a mass gap remains theoretically unexplained.' author: - 'Will M. Farr' - Niharika Sravan - Andrew Cantrell - Laura Kreidberg - 'Charles D. Bailyn' - Ilya Mandel - Vicky Kalogera bibliography: - 'paper.bib' title: 'The Mass Distribution of Stellar-Mass Black Holes' --- Introduction {#sec:intro} ============ The most massive stars probably end their lives with a supernova explosion or a quiet core collapse, becoming stellar-mass black holes. The mass distribution of such black holes can provide important clues to the end stages of evolution of these stars. In addition, the mass distribution of stellar-mass black holes is an important input in calculations of rates of gravitational wave emission events from coalescing neutron star-black hole and black hole-black hole binaries in the LIGO gravitational wave observatory [@Abadie2010]. Observations of X-ray binaries in both the optical and X-ray bands can provide a measurement of the mass of the compact object in these systems. The current sample of stellar mass black holes with dynamically measured masses includes 15 systems with low-mass, Roche lobe overflowing donors and 5 wind-fed systems with high-mass donors. Hence, sophisticated statistical analyses of the black hole mass distribution in these systems are possible. The first study of the mass distribution of stellar-mass black holes, in @Bailyn1998, examined a sample of seven low-mass X-ray binaries thought to contain a black hole, concluding in a Bayesian analysis that the mass function was strongly peaked around seven solar masses[^1]. @Bailyn1998 found evidence of a “gap” between the least massive black hole and a “safe” upper limit for neutron star masses of 3 ${M_\odot}$ (e.g. @Kalogera1996). Such a gap is puzzling in light of theoretical studies that predict a continuous distribution of compact object supernova remnant masses with a smooth transition from neutron stars to black holes [@Fryer2001]. (We note that @Fryer2001 considered binary evolution effects only heuristically and put forward some possible explanations for the gap from @Bailyn1998 both in the context of selection effects or in connection to the energetics of supernova explosions.) Towards the end of our analysis work, we became aware of a more recent study [@Ozel2010], also in a Bayesian framework, analyzing the low-mass X-ray binary sample. Our results are largely consistent with those obtained by @Ozel2010, who examined 16 low-mass X-ray binary systems containing black holes and found a strongly peaked distribution at $7.8 \pm 1.2 \, {M_\odot}$. They used two models for the mass function: a Gaussian and a decaying exponential with a minimum “turn-on” mass (motivated by the analytical model of the black-hole mass function in @Fryer2001). We note that @Ozel2010 do not provide confidence limits for the minimum black hole mass, instead discussing only the model parameters at the peak of their posterior distributions. They also do not perform any model selection analysis; thus, they give the distribution of parameters within each of their models, but cannot say which model is more likely to correspond to the true distribution of black hole masses. Nevertheless, it appears that their analysis confirms the existence of a mass gap. @Ozel2010 discuss possible selection effects that could lead to the appearance of a mass gap, but conclude these effects could not produce the observed gap, which they therefore claim is a real property of the black hole mass distribution. We use a Bayesian Markov-Chain Monte Carlo (MCMC) analysis to quantitatively assess a wide range of models for the black hole mass function for both samples. We include both parametric models, such as a Gaussian, and non-parametric models where the mass function is represented by histograms with various numbers of bins. (Our set of models includes those of @Ozel2010 and @Bailyn1998.) After computing posterior distributions for the model parameters, we use model selection techniques (including a new technique for efficient reversible-jump MCMC [@Farr2010]) to compare the evidence for the various models from both samples. We define the “minimum black hole mass” to be the 1% quantile, $M_{1\%}$, in the black hole mass distribution (see Section \[sec:minimum-mass\]). In qualitative agreement with @Ozel2010 and @Bailyn1998, we find strong evidence for a mass gap among the best models for both samples. Our analysis gives distributions for $M_{1\%}$ implied by the data in the context of each of our models for the black hole mass distribution. In the context of the best model for the low-mass systems (a power-law), the distribution for $M_{1\%}$ gives $M_{1\%} > 4.3$ ${M_\odot}$ with 90% confidence; in the context of the best model for the combined sample of lower- and high-mass systems the distribution of $M_{1\%}$ has $M_{1\%} > 4.5$ ${M_\odot}$ with 90% confidence. Further, in the context of models with lower evidence, most also have a mass gap, with 90% confidence bounds on $M_{1\%}$ significantly above a “safe” maximum neutron star mass of 3 ${M_\odot}$ [@Kalogera1996]. We find that, for the low-mass X-ray binary sample, the theoretical model from @Fryer2001—a decaying exponential—is strongly disfavored by our model selection. We find that the low-mass systems are best described by a power law, followed closely by a Gaussian (which is the second model considered by @Ozel2010). On the other hand, we find that the theoretical model from @Fryer2001 is the preferred model for the combined sample of low- and high-mass X-ray binaries. A model with two separate Gaussian peaks also has relatively high evidence for the combined sample of systems. The difference in best-fitting model indicates that the low-mass subsample is not consistent with being drawn from the distribution of the combined population. The structure of this paper is as follows. In Section \[sec:systems\] we discuss the 15 systems that comprise the low-mass X-ray binary black hole sample and the 5 additional high-mass, wind-fed systems that make up the combined sample. In Section \[sec:models\] we discuss the Bayesian techniques we use to analyze the black hole mass distribution, the techniques we use for model selection, and the parametric and non-parametric models we will use for the black hole mass distribution. In Section \[sec:results\] we discuss the results of our analysis and model selection. In Section \[sec:minimum-mass\] we discuss the distribution of the minimum black hole mass implied by the analysis of Section \[sec:results\]. In Section \[sec:conclusion\] we summarize our results and comment on the significance of the observed mass gap in the context of theoretical models. Appendix \[sec:mcmc\] describes MCMC techniques in some detail. Appendix \[sec:reversible-jump-mcmc\] explains our novel algorithm for efficiently performing the reversible jump MCMC computations used in the model comparisons of Section \[sec:results\] (but see also @Farr2010). Systems {#sec:systems} ======= The 20 X-ray binary systems on which this study is based are listed in Table \[tab:sources\]. We separate the systems into 15 low-mass systems in which the central black hole appears to be fed by roche-lobe overflow from the secondary, and 5 high-mass systems in which the black hole is fed via winds (these systems all have a secondary that appears to be more massive than the black hole). The low- and high-mass systems undoubtedly have different evolutionary tracks, and therefore it is reasonable that they would have different black-hole mass distributions. We will first analyze the 15 low-mass systems alone (Section \[sec:results-low-mass\]), and then the combined sample of 20 systems (Section \[sec:high-mass\]). In each of these systems, spectroscopic measurements of the secondary star provide an orbital period for the system and a semi-amplitude for the secondary’s velocity curve. These measurements can be combined into the mass function, $$\label{eq:mass-function} f(M) = \frac{P K^3}{2\pi G} = \frac{M \sin^3 i}{\left( 1 + q \right)^2},$$ where $P$ is the orbital period, $K$ is the secondary’s velocity semi-amplitude, $M$ is the black hole mass, $i$ is the inclination of the system, and $q \equiv M_2 / M$ is the mass ratio of the system. The mass function defines a lower limit on the mass: $f(M) < M$. To accurately determine the mass of the black hole, the inclination $i$ and mass ratio $q$ must be measured. Ideally, this can be accomplished by fitting ellipsoidal light curves and study of the rotational broadening of spectral lines from the secondary, but even in the most studied case (see, e.g., @Cantrell2010 on A0620) this procedure is complicated. In particular, contributions from an accretion disk and hot spots in the disk can significantly distort the measured inclination and mass ratios. For some systems (e.g. GS 1354 [@Casares2009]) strong variability completely prevents determination of the inclination from the lightcurve; in these cases an upper limit on the inclination often comes from the observed lack of eclipses in the lightcurve. In general, accurately determining $q$ and $i$ requires a careful system-by-system analysis. For the purposes of this paper, we adopt the following simplified approach to the estimation of the black hole mass from the observed data. When an observable is well-constrained, we assume that the true value is normally distributed about the measured value with a standard deviation equal to the quoted observational error. This is the case for the mass function in all the systems we use, and for many systems’ mass ratios and inclinations. When a large range is quoted in the literature for an observable, we take the true value to be distributed uniformly (for the mass ratio) or isotropically (for the inclination) within the quoted range. Table \[tab:sources\] gives the assumed distribution for the observables in the 20 systems we use. We do not attempt to deal with the systematic biases in the observational determination of $f$, $q$, and $i$ in any realistic way; we are currently investigating more realistic treatments of the errors (including observational biases that can shift the peak of the true mass distribution away from the “best-fit” mass in the observations). This treatment will appear in future work. ------------------------------------------------------------------------------------------------- Source $f$ (${M_\odot}$) $q$ $i$ (degrees) References --------------- ------------------- ---------------------- ------------------ ------------------- GRS 1915 $N(9.5, 3.0)$ $N(0.0857, 0.0284)$ $N(70, 2)$ @Greiner2001 XTE J1118 $N(6.44, 0.08)$ $N(0.0264, 0.004)$ $N(68, 2)$ @Gelino2008 @Harlaftis2005 XTE J1650 $N(2.73, 0.56)$ $U(0, 0.5)$ $I(50, 80)$ [@Orosz2004] GRS 1009 $N(3.17, 0.12)$ $N(0.137, 0.015)$ $I(37, 80)$ [@Filippenko1999] A0620 $N(2.76, 0.036)$ $N(0.06, 0.004)$ $N(50.98, 0.87)$ @Cantrell2010 @Neilsen2008 GRO J0422 $N(1.13, 0.09)$ $U(0.076, 0.31)$ $N(45, 2)$ @Gelino2003 Nova Mus 1991 $N(3.01, 0.15)$ $N(0.128, 0.04)$ $N(54,1.5)$ [@Gelino2001] GRO J1655 $N(2.73,0.09)$ $N(0.3663, 0.04025)$ $N(70.2, @Greene2001 1.9)$ 4U 1543 $N(0.25, 0.01)$ $U(0.25, 0.31)$ $N(20.7,1.5)$ @Orosz2003 XTE J1550 $N(7.73,0.4)$ $U(0,0.04)$ $N(74.7, 3.8)$ @Orosz2010 V4641 Sgr $N(3.13,0.13)$ $U(0.42,0.45)$ $N(75,2)$ @Orosz2003 GS 2023 $N(6.08, 0.06)$ $U(0.056,0.063)$ $I(66, 70)$ @Charles2006 @Khargharia2010 GS 1354 $N(5.73, 0.29)$ $N(0.12,0.04)$ $I(50, 80)$ @Casares2009 Nova Oph 77 $N(4.86,0.13)$ $U(0, 0.053)$ $I(60, 80)$ @Charles2006 GS 2000 $N(5.01, 0.12)$ $U(0.035, 0.053)$ $I(43, 74)$ @Charles2006 Cyg X1 $N(0.251, 0.007)$ $N(2.778, 0.386)$ $I(23, 38)$ @Gies2003 M33 X7 $N(0.46, 0.08)$ $N(4.47, 0.61)$ $N(74.6, 1)$ @Orosz2007 NGC 300 X1 $N(2.6, 0.3)$ $U(1.05, 1.65)$ $I(60, 75)$ @Crowther2010 LMC X1 $N(0.148, 0.004)$ $N(2.91, 0.49)$ $N(36.38, 2.02)$ @Orosz2009 IC 10 X1 $N(7.64, 1.26)$ $U(0.7, 1.7)$ $I(75, 90)$ @Prestwich2007 @Silverman2008 ------------------------------------------------------------------------------------------------- : \[tab:sources\] The source parameters for the 20 X-ray binaries used in this work. The first 15 systems have low-mass secondaries that feed the black hole via Roche lobe overflow; the last five systems have high-mass secondaries ($q \gtrsim 1$) that feed the black hole via winds. In each line, $f$ is the mass function for the compact object, $q$ is the mass ratio $M_2/M$, and $i$ is the inclination of the system to the line of sight. We indicate the distribution used for the true parameters when computing the probability distributions for the masses of these systems: $N(\mu,\sigma)$ implies a Gaussian with mean $\mu$ and standard deviation $\sigma$, $U(a,b)$ is a uniform distribution between $a$ and $b$, and $I(\alpha,\beta)$ is an isotropic distribution between the angles $\alpha$ and $\beta$. From these assumptions, we can generate probability distributions for the true mass of the black hole given the observations and errors via the Monte Carlo method: drawing samples of $f$, $q$, and $i$ from the assumed distributions and computing the mass implied by Equation gives samples of $M$ from the distribution induced by the relationship in Equation . Mass distributions generated in this way for the systems used in this work are shown in Figure \[fig:low-masses\] and Figure \[fig:high-masses\]. Systems for which $i$ is poorly constrained have broad “tails” on their mass distributions. These mass distributions constitute the “observational data” we will use in the remainder of this paper. Statistical Analysis {#sec:models} ==================== In this section we describe the statistical analysis we will apply to various models for the underlying mass distribution from which the low-mass sample and the combined sample of X-ray binary systems in Table \[tab:sources\] were drawn. The results of our analysis are presented in Section \[sec:results\]. Bayesian Inference ------------------ The end result of our statistical analysis will be the probability distribution for the parameters of each model implied by the data from Section \[sec:systems\] in combination with our prior assumptions about the probability distribution for the parameters. Bayes’ rule relates these quantities. For a model with parameters ${\vec{\theta}}$ in the presence of data $d$, Bayes’ rule states $$\label{eq:Bayes-rule} p({\vec{\theta}}| d) = \frac{p(d | {\vec{\theta}}) p({\vec{\theta}})}{p(d)}.$$ Here, $p({\vec{\theta}}|d)$, called the posterior probability distribution function, is the probability distribution for the parameters ${\vec{\theta}}$ implied by the data $d$; $p(d|{\vec{\theta}})$, called the likelihood, is the probability of observing data $d$ given that the model parameters are ${\vec{\theta}}$; $p({\vec{\theta}})$, called the prior, reflects our estimate of the probability of the various model parameters in the absence of any data; and $p(d)$, called the evidence, is an overall normalizing constant ensuring that $$\int d\theta\, p({\vec{\theta}}|d) = 1,$$ whence $$\label{eq:evidence-def} p(d) = \int d{\vec{\theta}}\, p(d|{\vec{\theta}}) p({\vec{\theta}}).$$ In our context, the data are the mass distributions given in Section \[sec:systems\]: $d = \{ p_i(M)| i = 1, 2, \ldots, 20 \}$. We assume that the measurements in Section \[sec:systems\] are independent, so the complete likelihood is given by a product of the likelihoods for the individual measurements. For a model with parameters ${\vec{\theta}}$ that predicts a mass distribution $p(M|{\vec{\theta}})$ for black holes, we have $$\label{eq:likelihood-def} p(d|{\vec{\theta}}) = \prod_i \int dM\, p_i(M) p(M|{\vec{\theta}}).$$ That is, the likelihood of an observation is the average over the individual mass distribution implied by the observation, $p_i(M)$, of the probability for a black hole of that mass to exist according to the model of the mass distribution, $p(M | {\vec{\theta}})$. We approximate the integrals as averages of $p(M|{\vec{\theta}})$ over the Monte Carlo mass samples drawn from the distributions in Table \[tab:sources\] (also see Figures \[fig:low-masses\] and \[fig:high-masses\]): $$p(d|{\vec{\theta}}) \approx \prod_i \frac{1}{N_i} \sum_{j = 1}^{N_i} p(M_{ij} | {\vec{\theta}}),$$ where $M_{ij}$ is the $j$th sample (out of a total $N_i$) from the $i$th individual mass distribution. Our calculation of the likelihood of each observation does not include any attempt to account for selection effects in the observations. We simply assume (almost certainly incorrectly) that any black hole drawn from the underlying mass distribution is equally likely to be observed. The results of @Ozel2010 suggest that selection effects are unlikely to significantly bias our analysis. For a mass distribution with several parameters, $p({\vec{\theta}}| d)$ lives in a multi-dimensional space. Previous works [@Ozel2010; @Bailyn1998] have considered models with only two parameters; for such models evaluating $p({\vec{\theta}}|d)$ on a grid may be a reliable method. Many of our models for the underlying mass distribution have three or more parameters. Exploring the entirety of these parameter spaces with a grid rapidly becomes prohibitive as the number of parameters increases. A more efficient way to explore the distribution $p({\vec{\theta}}| d)$ is to use a Markov Chain Monte Carlo (MCMC) method (see Appendix \[sec:mcmc\]). MCMC methods produce a chain (sequence) of parameter samples, $\{ {\vec{\theta}}_i \, | \, i = 1, \ldots \}$, such that a particular parameter sample, ${\vec{\theta}}$, appears in the sequence with a frequency proportional to its posterior probability, $p({\vec{\theta}}|d)$. In this way, regions of parameter space where $p({\vec{\theta}}|d)$ is large are sampled densely while regions where $p({\vec{\theta}}|d)$ is small are effectively ignored. Once we have a chain of samples from $p({\vec{\theta}}|d)$, the distribution for any quantity of interest can be computed by evaluating it on each sample in the chain and forming a histogram of these values. For example, to compute the one-dimensional distribution for a single parameter obtained by integrating over all other dimensions in parameter space, called the “marginalized” distribution, one plots the histogram of the values of that parameter appearing in the chain. Priors {#sec:priors} ------ An important part of any Bayesian analysis is the priors placed on the parameters of the model. The choice of priors can bias the results of the analysis through both the shape and the range of prior support in parameter space. The prior should reflect the “best guess” for the distribution of parameters before examining any of the data. In the absence of any information about the distribution of parameters, it is best to choose a prior that is broad and uninformative to avoid biasing the posterior as much as possible. A prior that is independent of parameters, ${\vec{\theta}}$, in some region, called “flat,” results in a posterior that is proportional to the likelihood (see Equation ). A flat prior does not change the shape of the posterior. However, the choice of a flat prior is parameterization-dependent: a change of parameter from ${\vec{\theta}}$ to ${\vec{\theta}}' = \vec{f}({\vec{\theta}})$ can change a flat distribution into one with non-trivial structure. In this work, we choose priors that are flat when the parameters are measured in physical units. In particular, for the log-normal model (Section \[sec:log-normal\]) the natural parameters for the distribution are the mean, $\langle \log M \rangle$, and standard deviation, $\sigma_{\log M}$, in $\log M$, but we choose priors that are flat in $\langle M \rangle$ and $\sigma_M$. The range of prior support can also affect the results of a Bayesian analysis. Because priors are normalized, prior support over a larger region of parameter space results in a smaller prior probability at each point. Such “wide” priors are implicitly claiming that any particular sub-region of parameter space is less likely than it would be under a prior of the same shape but smaller support volume. This difference is important in model selection (Section \[sec:model-selection\]): when comparing two models with the same likelihood, one with wide priors will seem less probable than one with narrower priors. Of course, priors should be wide enough to encompass all regions of parameter space that have significant likelihood. To make the model comparison in Section \[sec:model-selection\] fair, we choose prior support in parameter space so that the allowed parameter values for each model give distributions for which nearly all the probability lies in the range $0 \leq M \leq 40 {M_\odot}$. Parametric Models for the Black Hole Mass Distribution {#sec:parametric-models} ------------------------------------------------------ Here we discuss the various parametric models of the underlying black hole mass distribution considered in this paper. ### Power-Law Models {#sec:power-law} Many astrophysical distributions are power laws. Let us assume that the BH mass distribution is given by $$\label{eq:power-law-dist} p(M|{\vec{\theta}}) = p(M|\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha \}) = \begin{cases} A M^\alpha & {M_{\textnormal{min}}}\leq m \leq {M_{\textnormal{max}}}\\ 0 & \textnormal{otherwise} \end{cases}.$$ The normalizing constant $A$ is $$A = \frac{1+\alpha}{{M_{\textnormal{max}}}^{1+\alpha} - {M_{\textnormal{min}}}^{1+\alpha}}.$$ We choose uniform priors on ${M_{\textnormal{min}}}$ and ${M_{\textnormal{max}}}\geq {M_{\textnormal{min}}}$ between 0 and $40 {M_\odot}$, and uniform priors on the exponent $\alpha$ in a broad range between $-15$ and $13$: $$p({\vec{\theta}}) = p(\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}) = \begin{cases} 2 \frac{1}{40^2}\frac{1}{28} & 0 \leq {M_{\textnormal{min}}}\leq {M_{\textnormal{max}}}\leq 40, \quad -15 \leq \alpha \leq 13 \\ 0 & \textnormal{otherwise} \end{cases}.$$ Our MCMC analysis output is a list of $\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}$ values distributed according to the posterior $$p({\vec{\theta}}|d) = p(\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}|d) \propto p(d|\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}) p(\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}),$$ with the likelihood $p(d|\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\})$ defined in Equation . ### Decaying Exponential {#sec:exponential} @Fryer2001 studied the relation between progenitor and remnant mass in simulations of supernova explosions. Combining this with the mass function for supernova progenitors, they suggested that the black-hole mass distribution may be well-represented by a decaying exponential with a minimum mass: $$\label{eq:exp-def} p(M|{\vec{\theta}}) = p(M|\{{M_{\textnormal{min}}}, M_0\}) = \begin{cases} \frac{e^{\frac{{M_{\textnormal{min}}}}{M_0}}}{M_0} \exp \left[ - \frac{M}{M_0} \right] & M \geq {M_{\textnormal{min}}}\\ 0 & \textnormal{otherwise} \end{cases}.$$ We choose a prior for this model where ${M_{\textnormal{min}}}$ is uniform between 0 and $40 {M_\odot}$. For each ${M_{\textnormal{min}}}$, we choose $M_0$ uniformly within a range ensuring that $40{M_\odot}$ is at least two scale masses above the cutoff: $40{M_\odot}\geq {M_{\textnormal{min}}}+ 2 M_0$. This ensures that the majority of the mass probability lies in the range $0 \leq M \leq 40{M_\odot}$. The resulting prior is $$p({\vec{\theta}}) = p(\{{M_{\textnormal{min}}}, M_0\}) = \begin{cases} \frac{4}{40^2} & 0 \leq {M_{\textnormal{min}}}\leq 40, \quad 0 < M_0, \quad {M_{\textnormal{min}}}+ 2 M_0 \leq 40, \\ 0 & \textnormal{otherwise} \end{cases}$$ ### Gaussian and Two-Gaussian Models {#sec:gaussian} The mass distributions in Figure \[fig:low-masses\] all peak in a relatively narrow range near $\sim 10 {M_\odot}$. The prototypical single-peaked probability distribution is a Gaussian: $$\label{eq:gaussian-def} p(M|{\vec{\theta}}) = p(M|\{\mu, \sigma\}) = \frac{1}{\sigma \sqrt{2\pi}} \exp\left[ - \left(\frac{M - \mu}{\sqrt{2} \sigma} \right)^2 \right].$$ We use a prior on the mean mass, $\mu$, and the standard deviation, $\sigma$, that ensures that the majority of the mass distribution lies below $40 {M_\odot}$: $$\label{eq:gaussian-prior-def} p(\{\mu,\sigma\}) = \begin{cases} \frac{8}{40^2} & 0 \leq \mu \leq 40, \quad \sigma \geq 0, \quad \mu + 2\sigma \leq 40 \\ 0 & \textnormal{otherwise} \end{cases},$$ where both $\mu$ and $\sigma$ are measured in solar masses. Though we do not expect to find a second peak in the low-mass distribution, we may find evidence of one when exploring the combined low- and high-mass samples. To look for a second peak in the black-hole mass distribution, we use a two-Gaussian model: $$\begin{gathered} \label{eq:two-gaussian-def} p(M|{\vec{\theta}}) = p(M|\{\mu_1, \mu_2, \sigma_1, \sigma_2, \alpha\}) = \\ \frac{\alpha}{\sigma_1 \sqrt{2\pi}} \exp\left[ - \left( \frac{M - \mu_1}{\sqrt{2}\sigma_1} \right)^2 \right] + \frac{1-\alpha}{\sigma_2 \sqrt{2\pi}} \exp\left[ - \left( \frac{M - \mu_2}{\sqrt{2}\sigma_2} \right)^2 \right].\end{gathered}$$ The probability is a linear combination of two Gaussians with weights $\alpha$ and $1-\alpha$. We restrict $\mu_1 < \mu_2$ and also impose combined conditions on $\mu_i$ and $\sigma_i$ that ensure that most of the mass probability lies below $40 {M_\odot}$ with the prior $$p(\{\mu_1, \mu_2, \sigma_1, \sigma_2, \alpha\}) = \begin{cases} 2 p(\{\mu_1, \sigma_1\}) p(\{\mu_2, \sigma_2\}) & \mu_1 \leq \mu_2, \quad 0 \leq \alpha \leq 1 \\ 0 & \textnormal{otherwise} \end{cases},$$ where the single-Gaussian prior, $p(\{\mu_i, \sigma_i\})$, is defined in Equation . ### Log Normal {#sec:log-normal} Many of the mass distributions for the systems in Figure \[fig:low-masses\] rise rapidly to a peak and then fall off more slowly in a longer tail toward high masses. So far, none of the parameterized distributions we have discussed have this property. In this section, we consider a log-normal model for the underlying mass distribution; the log-normal distribution has a rise to a peak with a slower falloff in a long tail. The log-normal distribution gives $\log M$ a Gaussian distribution with mean $\mu$ and standard deviation $\sigma$: $$\label{eq:log-normal-def} p(M|{\vec{\theta}}) = p(M|\{\mu, \sigma \}) = \frac{1}{ \sqrt{2\pi} M \sigma} \exp\left[ -\frac{\left(\log M - \mu\right)^2}{2 \sigma^2} \right].$$ The parameters $\mu$ and $\sigma$ are dimensionless; the mean mass $\langle M \rangle$ and mass standard deviation $\sigma_M$ are related to $\mu$ and $\sigma$ by $$\begin{aligned} \label{eq:avg-M} \langle M \rangle & = & \exp\left( \mu + \frac{1}{2} \sigma^2 \right) \\ \label{eq:sigma-M} \sigma_M & = & \langle M \rangle \sqrt{\exp\left( \sigma^2 \right) - 1}.\end{aligned}$$ For a fair comparison with the other models, we impose a prior that is flat in $\langle M \rangle$ and $\sigma_M$. To ensure that most of the probability in this model occurs for masses below $40 {M_\odot}$, we require $\langle M \rangle + 2 \sigma_M \leq 40$, resulting in a prior $$p({\vec{\theta}}) = p(\{\mu,\sigma\}) = \begin{cases} \frac{4}{40^2} \left| \frac{\partial \left(\langle M \rangle, \sigma_M \right)}{\partial \left( \mu, \sigma \right)} \right| & \sigma > 0, \quad \langle M \rangle + 2 \sigma_M \leq 40 \\ 0 & \textnormal{otherwise} \end{cases},$$ where $$\left| \frac{\partial \left(\langle M \rangle, \sigma_M \right)}{\partial \left( \mu, \sigma \right)} \right| = \frac{\exp\left( 2 \left( \mu + \sigma^2 \right)\right) \sigma}{\sqrt{\exp\left( \sigma^2 \right) - 1}}$$ is the determinant of the Jacobian of the map in Equations and . Non-Parametric Models for the Black Hole Mass Distribution {#sec:non-parametric-models} ---------------------------------------------------------- The previous subsection discussed models for the underlying black hole mass distribution that assumed particular parameterized shapes for the distribution. In this subsection, we will discuss models that do not assume a priori a shape for the black hole mass distribution. The fundamental non-parametric distribution in this section is a histogram with some number of bins, ${N_{\textnormal{bin}}}$. Such a distribution is piecewise-constant in $M$. One choice for representing such a histogram would be to fix the bin locations, and allow the heights to vary. With this approach, one should be careful not to “split” features of the mass distribution across more than one bin in order to avoid diluting the sensitivity to such features; similarly, one should avoid including more than “one” feature in each bin. The locations of the bins, then, are crucial. An alternative representation of histogram mass distributions avoids this difficulty. We choose to represent a histogram mass distribution with ${N_{\textnormal{bin}}}$ bins by allocating a fixed probability, $1/{N_{\textnormal{bin}}}$, to each bin. The lower and upper bounds for each bin are allowed to vary; when these are close to each other (i.e. the bin is narrow), the distribution will have a large value, and conversely when the bounds are far from each other. We assume that the non-zero region of the distribution is contiguous, so we can represent the boundaries of the bins as a non-decreasing array of masses, $w_0 \leq w_1 \leq \ldots \leq w_{{N_{\textnormal{bin}}}}$, with $w_0$ the minimum and $w_{{N_{\textnormal{bin}}}}$ the maximum mass for which the distribution has support. This gives the distribution $$\label{eq:hist-def} p(M|\theta) = p(M|\{w_0, \ldots, w_{{N_{\textnormal{bin}}}}\}) = \begin{cases} 0 & M < w_0 \textnormal{ or } w_{{N_{\textnormal{bin}}}} \leq M \\ \frac{1}{{N_{\textnormal{bin}}}} \frac{1}{w_{i+1} - w_i} & w_i \leq M < w_{i+1} \end{cases}.$$ For priors on the histogram model with ${N_{\textnormal{bin}}}$ bins, we assume that the bin boundaries are uniformly distributed between 0 and $40 {M_\odot}$ subject only to the constraint that the boundaries are non-decreasing from $w_0$ to $w_{{N_{\textnormal{bin}}}}$: $$p(\{w_0, \ldots, w_{{N_{\textnormal{bin}}}}\}) = \begin{cases} \frac{\left({N_{\textnormal{bin}}}+1\right)!}{40^{{N_{\textnormal{bin}}}+1}} & 0 \leq w_0 \leq w_1 \leq \ldots \leq w_{{N_{\textnormal{bin}}}} \leq 40 \\ 0 & \textnormal{otherwise} \end{cases}.$$ We consider histograms with up to five bins in this work. We will see that the evidence for the histogram models (see Sections \[sec:model-selection\], \[sec:low-mass-model-selection\], and \[sec:high-mass-model-selection\]) from both the low-mass and combined datasets is decreasing as the number of bins reaches five, indicating that increasing the number of bins beyond five would not sufficiently improve the fit to the mass distribution to compensate for the extra parameter-space volume implied by the additional parameters. Bayesian Model Selection {#sec:model-selection} ------------------------ In Sections \[sec:parametric-models\] and \[sec:non-parametric-models\], we discussed a series of models for the underlying black hole mass distribution. Our MCMC analysis will provide the posterior distribution of the parameters within each model, but does not tell us which models are more likely to correspond to the actual distribution. This model selection problem is the topic of this section. Consider a set of models, $\{M_i| i = 1, \ldots\}$, each with corresponding parameters ${\vec{\theta}}_i$. Re-writing Equation to be explicit about the assumption of a particular model, we have $$p({\vec{\theta}}_i | d, M_i) = \frac{p(d|{\vec{\theta}}_i, M_i) p({\vec{\theta}}_i | M_i)}{p(d|M_i)}.$$ This gives the posterior probability of the parameters ${\vec{\theta}}_i$ in the context of model $M_i$. But, the model itself can be regarded as a discrete parameter in a larger “super-model” that encompasses all the $M_i$. The parameters for the super-model are $\{M_i, {\vec{\theta}}_i\}$: a choice of model and the corresponding parameter values within that model. Each point in the super-model parameter space is a statement that, e.g., “the underlying mass distribution is a Gaussian, with parameters $\mu$ and $\sigma$,” or “the underlying mass distribution is a triple-bin histogram with parameters $w_1$, $w_2$, $w_3$, and $w_4$,” or .... The posterior probability of the super-model parameters is given by Bayes’ rule: $$\label{eq:bayes-explicit-model} p({\vec{\theta}}_i, M_i|d) = \frac{p(d|{\vec{\theta}}_i, M_i) p({\vec{\theta}}_i |M_i) p(M_i)}{p(d)},$$ where we have introduced the model prior $p(M_i)$, which represents our estimate on the probability that model $M_i$ is correct in the absence of the data $d$. The normalizing evidence is now $$\label{eq:multi-model-evidence-def} p(d) = \sum_i \int d{\vec{\theta}}_i\, p(d|{\vec{\theta}}_i, M_i) p({\vec{\theta}}_i |M_i) p(M_i) = \sum_i p(d|M_i) p(M_i),$$ writing the single-model evidence from Equation as $p(d|M_i)$ to be explicit about the dependence on the choice of model. To compare the various models $M_i$, we are interested in the marginalized posterior probability of $M_i$: $$\label{eq:model-posterior-def} p(M_i|d) \equiv \int d{\vec{\theta}}_i\, p({\vec{\theta}}_i, M_i|d).$$ This is the integral of the posterior over the entire parameter space of model $M_i$. The marginalized posterior probability of model $M_i$ can be re-written in terms of the single-model evidence, $p(d|M_i)$ (see Equations and ): $$\label{eq:model-evidence-def} p(M_i|d) = \int d{\vec{\theta}}_i\, p({\vec{\theta}}_i, M_i|d) = \frac{p(M_i)}{p(d)} \int d{\vec{\theta}}_i p(d|{\vec{\theta}}_i,M_i) p({\vec{\theta}}_i|M_i) = \frac{p(d|M_i) p(M_i)}{p(d)}.$$ Here and throughout, we assume that any of the models in Section \[sec:models\] are equally likely a priori, so the model priors are equal: $$p(M_i) = \frac{1}{N_{\textnormal{model}}}.$$ A powerful technique[^2] for computing $p(M_i|d)$ is the reversible-jump MCMC [@Green1995]. Reversible jump MCMC, discussed in more detail in Appendix \[sec:reversible-jump-mcmc\], is a standard MCMC analysis conducted in the super-model. The result of a reversible jump MCMC is a chain of samples, $\{ M_i, {\vec{\theta}}_i\, | \, i = 1, \ldots \}$, from the super-model parameter space. The integral in Equation can be estimated by counting the number of times that a given model $M_i$ appears in the reversible jump MCMC chain: $$p(M_i|d) = \int d{\vec{\theta}}_i p(M_i, {\vec{\theta}}_i|d) \approx \frac{N_i}{N},$$ where $N_i$ is the number of MCMC samples that have discrete parameter $M_i$, and $N$ is the total number of samples in the MCMC. Naively implemented reversible jump MCMCs can be very inefficient when the posteriors for a model or models are strongly peaked. In this circumstance, a proposed MCMC jump into one of the peaked models is unlikely to land on the peak by chance; since it is rare to propose a jump into the important regions of parameter space of the peaked model in a naive reversible jump MCMC, the output chain must be very long to ensure that all models have been compared fairly. We describe a new algorithm in Appendix \[sec:reversible-jump-mcmc\] that produces very efficient jump proposals for a reversible jump MCMC by exploiting the information about the model posteriors we have from the single-model MCMC samples. (See also @Farr2010.) With this algorithm, reasonable chain lengths can fairly compare all the models under consideration. We have used this algorithm to perform 10-way reversible jump MCMCs to calculate the relative evidence for both the parametric and non-parametric models in this study. These results appear in Section \[sec:results\]. Results {#sec:results} ======= In this section we discuss the results of our MCMC analysis of the posterior distributions of parameters for the models in Sections \[sec:parametric-models\] and \[sec:non-parametric-models\]. We also discuss model selection results. The results in Section \[sec:results-low-mass\] apply to the low-mass sample of systems, while those of Section \[sec:high-mass\] apply to the combined sample of systems. Low-Mass Systems {#sec:results-low-mass} ---------------- Table \[tab:low-mass-parametric\] gives quantiles of the marginalized parameter distributions of the parametric models implied by the low-mass data. Table \[tab:low-mass-non-parametric\] gives the quantiles of the histogram bin boundaries in the non-parametric analysis implied by the low-mass data. Model Parameter 5% 15% 50% 85% 95% -------------------------- -------------------------- ---------- ---------- --------- --------- --------- Power Law (Equation ) ${M_{\textnormal{min}}}$ 1.2786 4.1831 6.1001 6.5011 6.6250 ${M_{\textnormal{max}}}$ 8.5578 8.9214 23.3274 36.0002 38.8113 $\alpha$ -12.4191 -10.1894 -6.3861 2.8476 5.6954 Exponential (Equation ) ${M_{\textnormal{min}}}$ 5.0185 5.4439 6.0313 6.3785 6.5316 $M_0$ 0.7796 0.9971 1.5516 2.4635 3.2518 Gaussian (Equation ) $\mu$ 6.6349 6.9130 7.3475 7.7845 8.0798 $\sigma$ 0.7478 0.9050 1.2500 1.7335 2.1134 Two Gaussian (Equation ) $\mu_1$ 5.4506 6.3877 7.1514 7.6728 7.9803 $\mu_2$ 7.2355 7.7387 12.3986 25.2456 31.4216 $\sigma_1$ 0.3758 0.7626 1.2104 1.7981 2.3065 $\sigma_2$ 0.2048 0.6421 1.9182 5.2757 7.2625 $\alpha$ 0.0983 0.3526 0.8871 0.9792 0.9936 Log Normal (Equation ) $\langle M \rangle$ 6.7619 7.0122 7.4336 7.9159 8.2942 $\sigma_M$ 0.7292 0.8920 1.2704 1.8695 2.4069 : \[tab:low-mass-parametric\] Quantiles of the marginalized distribution for each of the parameters in the models discussed in Section \[sec:parametric-models\] implied by the low-mass data. We indicate the 5%, 15%, 50% (median), 85%, and 95% quantiles. The marginalized distribution can be misleading when there are strong correlations between variables. For example, while the marginalized distributions for the power law parameters are quite broad, the distribution of mass distributions implied by the power law MCMC samples is similar to the other models. This occurs in spite of the broad marginalized distributions because of the correlations between the slope and limits of the power law discussed in Section \[sec:power-law\]. Recall that each MCMC sample in our analysis gives the parameters for a model of the black hole mass distribution. The chain of samples of parameters for a particular model gives us a distribution of black hole mass distributions. Figure \[fig:dists\] gives a sense of the shape and range of the distributions of black hole mass distributions that result from our MCMC analysis. In Figure \[fig:dists\] we plot the median, 10% and 90% values of the black hole mass distributions that result from the MCMC chains. Because the choice of parameters that gives, for example, the median distribution value at one mass need not give the median distribution at another mass, these curves do not necessarily look like the underlying model for the mass distribution. For the same reason, they are not necessarily normalized. ### Power Law In Figure \[fig:power-law\] , we display a histogram of the resulting samples in each of the parameters ${M_{\textnormal{min}}}$, ${M_{\textnormal{max}}}$, and $\alpha$ for the power law model (see Equation ); this represents the one-dimensional “marginalized” distribution $$\label{eq:alpha-pdf} p(\alpha|d) = \int d{M_{\textnormal{min}}}\, d{M_{\textnormal{max}}}\, p(\{{M_{\textnormal{min}}}, {M_{\textnormal{max}}}, \alpha\}|d),$$ and similarly for ${M_{\textnormal{min}}}$ and ${M_{\textnormal{max}}}$. The marginalized distribution for $\alpha$ is broad, with $$-11.8 < \alpha < 6.8$$ enclosing 90% of the probability (excluding 5% on each side). We have $p(\alpha < 0) = 0.6$. The median value is $\alpha = -3.35$. The broadness of the marginalized distribution for $\alpha$ comes from the need to match the relatively narrow range in mass of the low-mass systems. When $\alpha$ is negative, the resulting mass distribution slopes down; ${M_{\textnormal{min}}}$ is constrained to be near the lowest mass of the observed black holes, while ${M_{\textnormal{max}}}$ is essentially irrelevant. Conversely, when $\alpha$ is positive and the mass distribution slopes up, ${M_{\textnormal{max}}}$ must be close to the largest mass observed, while ${M_{\textnormal{min}}}$ is essentially irrelevant. Figure \[fig:power-law-2D\] illustrates this effect, showing the correlations between $\alpha$ and ${M_{\textnormal{min}}}$ and $\alpha$ and ${M_{\textnormal{max}}}$. When we include the high-mass systems in the analysis, the long tail will eliminate this effect by bringing both ${M_{\textnormal{min}}}$ and ${M_{\textnormal{max}}}$ into play for all values of $\alpha$. ### Decaying Exponential {#decaying-exponential} Figure \[fig:exp-marginal\] displays the marginalized posterior distribution for the scale mass of the exponential, $M_0$, and the cutoff mass, ${M_{\textnormal{min}}}$ (see Equation \[eq:exp-def\]). The median scale mass is $M_0 = 1.55$, and $0.78 \leq M_0 \leq 3.25$ with 90% confidence. This model was one of those considered by @Ozel2010, whose results ($M_0 \sim 1.5$ and ${M_{\textnormal{min}}}\sim 6.5$) are broadly consistent with ours. Figure \[fig:exp-2D\] displays the MCMC samples in the ${M_{\textnormal{min}}}$, $M_0$ plane for this model. There is a small correlation between smaller ${M_{\textnormal{min}}}$ and larger $M_0$, which is driven by the need to widen the distribution to encompass the peak of the mass measurements in Figure \[fig:low-masses\] when the minimum mass is smaller. ### Gaussian Figure \[fig:gaussian\] shows the resulting marginalized distributions for the parameters $\mu$ and $\sigma$. We constrain the peak of the Gaussian between $6.63 \leq \mu \leq 8.08$ with 90% confidence. This model also appeared in @Ozel2010; they found $\mu \sim 7.8$ and $\sigma \sim 1.2$, consistent with our results here. ### Two Gaussian Figure \[fig:two-gaussian\] shows the marginalized distributions for the two-Gaussian model parameters from our MCMC runs. We find $\alpha > 0.8$ with 62% probability, clearly favoring the Gaussian with smaller mean. The distributions for $\mu_1$ and $\sigma_1$ are similar to those of the single Gaussian displayed in Figure \[fig:gaussian\], indicating that this Gaussian is centered around the peaks of the low-mass distributions. The second Gaussian’s parameter distributions are much broader. The second Gaussian appears to be sampling the tail of the mass samples. In spite of the extra degrees of freedom in this model, we find that this model is strongly disfavored relative to the single-Gaussian model for this dataset: $p(\textnormal{Gaussian}|d) / p(\textnormal{Two Gaussian}|d) \simeq 4.7$ (see Sections \[sec:model-selection\] and \[sec:low-mass-model-selection\] for discussion). ### Log Normal {#log-normal} The marginal distributions for $\langle M \rangle$ and $\sigma_M$ appear in Figure \[fig:log-normal\]. The distributions are similar to those for $\mu$ and $\sigma$ from the Gaussian model in Section \[sec:gaussian\]. ### Histogram Models The median values of the histogram mass distributions that result from the MCMC samples of the posterior distribution for the $w_i$ parameters for one-, two-, three-, four-, and five-bin histogram models are shown in Figure \[fig:dists\]. Table \[tab:low-mass-non-parametric\] gives quantiles of the marginalized bin boundary distributions for the histogram models. Bins Boundary 5% 15% 50% 85% 95% ------ ---------- --------- --------- --------- --------- --------- 1 $w_0$ 3.94488 4.55603 5.43333 6.02557 6.29749 $w_1$ 8.50844 8.69262 9.11784 9.83477 10.5128 2 $w_0$ 3.3426 4.2047 5.39132 6.18413 6.47553 $w_1$ 6.41972 6.72605 7.43421 8.2489 8.52885 $w_2$ 8.46161 8.65077 9.12694 10.1113 11.2595 3 $w_0$ 2.18176 3.54345 5.16094 6.16473 6.44697 $w_1$ 5.68876 6.14223 6.68829 7.38725 8.04235 $w_2$ 6.8297 7.22718 8.1451 8.7512 9.27296 $w_3$ 8.44307 8.67362 9.25718 12.1688 21.92 4 $w_0$ 1.32131 2.7934 4.66156 5.78459 6.17946 $w_1$ 5.20112 5.77331 6.42501 6.98427 7.44584 $w_2$ 6.41805 6.73535 7.43826 8.32958 8.64212 $w_3$ 7.40302 7.95608 8.58976 9.33897 10.3992 $w_4$ 8.56724 8.8059 10.2451 24.3573 34.2423 5 $w_0$ 0.9392 2.28789 4.33389 5.7012 6.21166 $w_1$ 4.69778 5.44302 6.26575 6.76407 7.14427 $w_2$ 6.1388 6.47155 7.00606 7.97325 8.38259 $w_3$ 6.82058 7.28677 8.22514 8.81555 9.41012 $w_4$ 8.02335 8.36993 8.94879 11.3206 17.3349 $w_5$ 8.7112 9.25208 16.2059 31.897 37.2738 : \[tab:low-mass-non-parametric\] The 5%, 15%, 50% (median), 85%, and 95% quantiles for the bin boundaries in the one- through five-bin histogram models discussed in Section \[sec:non-parametric-models\]. As the number of bins increases, the models are better able to capture features of the mass distribution, but we find that the one-bin histogram is the most probable of the histogram models for the low-mass data (see Section \[sec:low-mass-model-selection\] for discussion). This occurs because the extra fitting power does not sufficiently improve the fit to compensate for the vastly larger parameter space of the models with more bins. ### Model Selection for the Low-Mass Sample {#sec:low-mass-model-selection} We have performed a suite of 500 independent reversible-jump MCMCs jumping between all the models (both parametric and non-parametric) described in Section \[sec:models\] using the single-model MCMC samples to construct an efficient jump proposal for each model as described above (see Appendix \[sec:reversible-jump-mcmc\]). The numbers of counts in each model are consistent across the MCMCs in the suite; Figure \[fig:rj\] displays the average probability for each model across the suite, along with the 1-$\sigma$ errors on the average inferred from the standard deviation of the model counts across the suite. Table \[tab:rj\] gives the numerical values of the average probability for each model across the suite of MCMCs. The most favored model is the power law from Section \[sec:power-law\], followed by the Gaussian model from Section \[sec:gaussian\]. Interestingly, the theoretical curve from @Fryer2001 (the exponential model of Section \[sec:exponential\]) places fourth in the ranking of evidence. Model Relative Evidence ---------------------------------------------------------- ------------------- Power Law (Section \[sec:power-law\]) 0.331488 Gaussian (Section \[sec:gaussian\]) 0.288129 Log Normal (Section \[sec:log-normal\]) 0.138435 Exponential (Section \[sec:exponential\]) 0.0916218 Two Gaussian (Section \[sec:gaussian\]) 0.0662577 Histogram (1 Bin, Section \[sec:non-parametric-models\]) 0.0641941 Histogram (2 Bin, Section \[sec:non-parametric-models\]) 0.015184 Histogram (3 Bin, Section \[sec:non-parametric-models\]) 0.00332933 Histogram (4 Bin, Section \[sec:non-parametric-models\]) 0.000999976 Histogram (5 Bin, Section \[sec:non-parametric-models\]) 0.0003614 : \[tab:rj\] Relative probabilities of the various models from Section \[sec:models\] implied by the low-mass data. (See also Figure \[fig:rj\].) These probabilities have been computed from reversible-jump MCMC samples using the efficient jump proposal algorithm in Appendix \[sec:reversible-jump-mcmc\]. Though the model probabilities presented in this section have small statistical error, they are subject to large “systematic” error. The source of this error is both the particular choice of model prior (uniform across models) and the choice of priors on the parameters within each model used for this work. For example, the theoretically-preferred exponential model (Section \[sec:exponential\]) is only a factor of $\sim 3$ away from the power law model (Section \[sec:power-law\]), which does not have theoretical support. Is such support worth a factor of three in the model prior? Alternately, we may say we know (in advance of any mass measurements) that black holes must exist with mass $\lesssim 10{M_\odot}$; then we could, for example, impose a prior on the minimum mass in the exponential model (${M_{\textnormal{min}}}$) that is uniform between $0$ and $10 {M_\odot}$, which would reduce the prior volume available for the model by a factor of 4 without significantly reducing the posterior support for the model. This has the same effect as increasing the model prior by a factor of 4, which would move this model from fourth to first place. Of course, we would then have to modify the prior support for the other models to take into account the restriction that there must be black holes with $M \lesssim 10{M_\odot}$.... @Linder2008 discuss these issues in the context of cosmological model selection, concluding with a warning against over-reliance on model selection probabilities. Nevertheless, we believe that our model comparison is reasonably fair (see the discussion of priors in Section \[sec:priors\]). It seems safe to conclude that “single-peaked” models (the power-law and Gaussian) are preferred over “extended” models (the exponential or log-normal), or those with “structure” (the many-bin histograms or two-Gaussian model). Previous studies have also supported the “single, narrow peak” mass distribution [@Bailyn1998; @Ozel2010]. In this light, poor performance of the single-bin histogram is surprising. Combined Sample {#sec:high-mass} --------------- This section repeats the analysis of the models from Section \[sec:models\], but including the high-mass, wind-fed systems from Table \[tab:sources\] (see also Figure \[fig:high-masses\]) in the sample. Figure \[fig:high-mass-dists\] displays bounds on the value of the underlying mass distribution for the various models in Section \[sec:models\] applied to this data set; compare to Figure \[fig:dists\]. The inclusion of the high-mass, wind-fed systems tends to widen the distribution toward the high-mass end and, in models that allow it, produce a second, high-mass peak in addition to the one in Figure \[fig:dists\]. ### Power Law Figure \[fig:power-law-high\] presents the marginalized distribution for the three power-law parameters ${M_{\textnormal{min}}}$, ${M_{\textnormal{max}}}$, and $\alpha$ (Section \[sec:power-law\]) from an analysis including the high-mass systems. The distribution for ${M_{\textnormal{max}}}$ is quite broad because the best fit power laws slope downward ($\alpha < 0$), making this parameter less relevant. The range $-5.05 \leq \alpha \leq -1.77$ encloses 90% of the probability; the median value of $\alpha$ is -3.23. The presence of the high-mass samples in the analysis produces a distinctive tail, eliminating the correlations discussed in Section \[sec:power-law\] and displayed in Figure \[fig:power-law-2D\] for the low-mass subset of the observations. ### Decaying Exponential {#decaying-exponential-1} Figure \[fig:exp-cutoff-high\] displays the marginalized distributions for the exponential parameters ${M_{\textnormal{min}}}$ and $M_0$ (Section \[sec:exponential\]) from an analysis including the high-mass systems. The distribution for the scale mass, $M_0$, has moved to higher masses relative to Figure \[fig:exp-marginal\] to fit the tail of the mass distribution; the distribution for ${M_{\textnormal{min}}}$ is less affected, though it has broadened somewhat toward low masses. ### Gaussian Figure \[fig:gaussian-high\] displays the marginalized distributions for the Gaussian parameters (Section \[sec:gaussian\]) when the high-mass objects are included in the mass distribution. The mean mass, $\mu$, and the mass standard deviation, $\sigma$, are both increased relative to Figure \[fig:gaussian\] to account for the broader distribution and high-mass tail. ### Two Gaussian The analysis of the two-Gaussian model shows the largest change when the high-mass samples are included. Figure \[fig:two-gaussian-high\] shows the marginalized distributions for the two-Gaussian parameters (Section \[sec:gaussian\]) when the high-mass samples are included in the analysis. In stark contrast to Figure \[fig:two-gaussian\], there are two well-defined, separated peaks; the low-mass peak reproduces the results from the low-mass samples, while the high-mass peak ($13.5534 \leq \mu_2 \leq 27.9481$ with 90% confidence; median 20.3839) matches the new high-mass samples. The peak in $\alpha$ near 0.8 is consistent with approximately 4/5 the total probability being concentrated in the 15 low-mass samples. ### Log Normal {#log-normal-1} The marginalized distributions for the log-normal parameters (Section \[sec:log-normal\]) when the high-mass samples are included in the analysis are displayed in Figure \[fig:log-normal-high\]. The changes when the high-mass samples are included (compare to Figure \[fig:log-normal\]) are similar to the changes in the Gaussian distribution: the mean mass moves to higher masses, and the distribution broadens. Because the log-normal distribution is inherently asymmetric, with a high-mass tail, it does not need to widen as much as the Gaussian distribution did. The confidence limits on the parameters for the parametric models of the underlying mass distribution are displayed in Table \[tab:high-mass-parametric\] (compare to Table \[tab:low-mass-parametric\]). Model Parameter 5% 15% 50% 85% 95% -------------------------- -------------------------- ---------- ---------- ---------- ---------- ---------- Power Law (Equation ) ${M_{\textnormal{min}}}$ 4.87141 5.29031 5.85019 6.26118 6.45674 ${M_{\textnormal{max}}}$ 19.1097 23.4242 31.5726 37.7519 39.3369 $\alpha$ -5.04879 -4.30368 -3.23404 -2.31365 -1.77137 Exponential (Equation ) ${M_{\textnormal{min}}}$ 4.0865 4.60236 5.32683 5.94097 6.22952 $M_0$ 2.82924 3.41139 4.70034 6.52214 7.92979 Gaussian (Equation ) $\mu$ 7.86599 8.33118 9.20116 10.2493 10.9836 $\sigma$ 2.23643 2.58899 3.33545 4.17886 4.67881 Two Gaussian (Equation ) $\mu_1$ 6.741 7.02724 7.48174 8.0139 8.46626 $\mu_2$ 13.5534 16.202 20.3839 24.9259 27.9481 $\sigma_1$ 0.742824 0.913941 1.31244 1.94862 2.50238 $\sigma_2$ 0.511159 1.5025 4.39824 7.04612 8.25905 $\alpha$ 0.575692 0.670978 0.798227 0.891522 0.932143 Log Normal (Equation ) $\langle M \rangle$ 8.00086 8.51192 9.6264 11.1851 12.3986 $\sigma_M$ 2.19262 2.8137 4.16742 6.25101 8.11839 : \[tab:high-mass-parametric\] Quantiles of the marginalized distribution for each of the parameters in the models discussed in Section \[sec:parametric-models\] when the high-mass samples are included in the analysis (compare to Table \[tab:low-mass-parametric\]). We indicate the 5%, 15%, 50% (median), 85%, and 95% quantiles. ### Histogram Models The non-parametric (histogram; see Section \[sec:non-parametric-models\]) models also show evidence of a long tail from the inclusion of the high-mass samples. Table \[tab:high-mass-non-parametric\] displays confidence limits on the histogram parameters for the analysis including the high-mass systems; compare to Table \[tab:low-mass-non-parametric\]. Bins Boundary 5% 15% 50% 85% 95% ------ ---------- --------- --------- --------- --------- --------- 1 $w_0$ 2.22294 3.12695 4.2456 5.15132 5.58265 $w_1$ 15.93 16.2535 17.7836 20.5449 22.5836 2 $w_0$ 3.87202 4.49983 5.41234 6.08334 6.35933 $w_1$ 7.22163 8.25079 8.93669 9.71551 10.4287 $w_2$ 18.4762 19.9798 24.941 32.5972 36.8615 3 $w_0$ 3.39289 4.24509 5.41694 6.15087 6.42822 $w_1$ 6.41849 6.71984 7.47263 8.2942 8.61785 $w_2$ 8.41449 8.64664 9.17056 10.4075 12.2718 $w_3$ 18.5705 21.0481 27.1494 34.7753 38.0652 4 $w_0$ 2.42094 3.69875 5.2596 6.25449 6.54316 $w_1$ 5.83725 6.2836 6.84987 7.8033 8.27706 $w_2$ 6.94919 7.43628 8.38531 9.13401 9.91845 $w_3$ 8.50371 8.75188 9.86694 17.1848 22.1086 $w_4$ 18.5823 21.4628 28.367 35.8118 38.5278 5 $w_0$ 1.73691 3.19184 4.89769 5.9547 6.35522 $w_1$ 5.46124 5.95881 6.59431 7.26795 7.91821 $w_2$ 6.63468 6.9804 7.93239 8.60918 9.06926 $w_3$ 7.89654 8.35634 8.91766 10.6568 13.9644 $w_4$ 8.74064 9.42672 15.8004 22.7101 27.6399 $w_5$ 20.0202 22.9065 29.6307 36.6606 38.8573 : \[tab:high-mass-non-parametric\] The 5%, 15%, 50% (median), 85%, and 95% quantiles for the bin boundaries in the one- through five-bin histogram models discussed in Section \[sec:non-parametric-models\] in an analysis including the high-mass, wind-fed systems. The tails evident in Figure \[fig:high-mass-dists\] are apparent here as well; compare to Table \[tab:low-mass-non-parametric\]. ### Model Selection for the Combined Sample {#sec:high-mass-model-selection} Repeating the model selection analysis discussed in Section \[sec:low-mass-model-selection\] for the sample including the high-mass systems, we find that the model probabilities have changed with the inclusion of the extra five systems. As before, we assume for this analysis that the model priors are equal. Reversible jump MCMC calculations of the model probabilities are displayed in Figure \[fig:high-rj-evidence\]; compare Figure \[fig:rj\]. The relative model probabilities are given in Table \[tab:rj-high\]. The exponential model is the most favored model for the combined sample, with the two-Gaussian model the second-most favored. The ranking of models differs significantly from the low-mass samples. The improvement of the exponential model relative to the low-mass analysis is encouraging for theoretical calculations that attempt to model the entire population of X-ray binaries with this mass model. Note also that the increasing structure of the mass distribution favors histogram models with three bins over those with fewer bins. Model Relative Evidence ---------------------------------------------------------- ------------------- Exponential (Section \[sec:exponential\]) 0.346944 Two Gaussian (Section \[sec:gaussian\]) 0.304923 Power Law (Section \[sec:power-law\]) 0.120313 Log Normal (Section \[sec:log-normal\]) 0.102536 Histogram (3 Bin, Section \[sec:non-parametric-models\]) 0.0473464 Histogram (4 Bin, Section \[sec:non-parametric-models\]) 0.0282086 Histogram (2 Bin, Section \[sec:non-parametric-models\]) 0.0210994 Histogram (5 Bin, Section \[sec:non-parametric-models\]) 0.0179703 Gaussian (Section \[sec:gaussian\]) 0.00901719 Histogram (1 Bin, Section \[sec:non-parametric-models\]) 0.00164214 : \[tab:rj-high\] Relative probabilities of the various models from Section \[sec:models\] implied by the combined sample of systems. (See also Figure \[fig:high-rj-evidence\].) These probabilities have been computed from reversible-jump MCMC samples using the efficient jump proposal algorithm in Appendix \[sec:reversible-jump-mcmc\]. The Minimum Black Hole Mass {#sec:minimum-mass} =========================== It is interesting to use our models for the underlying black hole mass distribution in X-ray binaries to place constraints on the minimum black hole mass implied by the present sample. @Bailyn1998 addressed this question in the context of a “mass gap” between the most massive neutron stars and the least massive black holes. The more recent study of @Ozel2010 also looked for a mass gap using a subset of the models and systems presented here. Both works found that the minimum black hole mass is significantly above the maximum neutron star mass [@Kalogera1996] of $\sim 3 {M_\odot}$ (though @Ozel2010 only state their evidence for a gap in terms of the maximum-posterior parameters and not the full extent of their distributions). The distributions of the minimum black hole mass from the analysis of the low-mass samples are displayed in Figure \[fig:min-mass\]. The minimum black hole mass is defined as the 1% mass quantile, $M_{1\%}$, of the black-hole mass distribution (i.e. the mass lying below 99% of the mass distribution). (A quantile-based definition is necessary in the case of those distributions that do not have a hard cutoff mass; even for those that do, like the power-law model, it can be useful to define a “soft” cutoff in the event that the lower mass hard cutoff becomes an irrelevant parameter as discussed in Section \[sec:power-law\].) For each mass distribution parameter sample from our MCMC, we can calculate the distribution’s minimum black hole mass; the collection of these minimum black hole masses approximates the distribution of minimum black hole masses implied by the data in the context of that distribution. Figure \[fig:min-mass\] plots histograms of the minimum black hole mass samples. We find that the best-fit model for the low-mass systems (the power-law) has $M_{1\%} > 4.3$ ${M_\odot}$ in 90% of the MCMC samples (i.e. at 90% confidence). This is significantly above the maximum theoretically-allowed neutron star mass, $\sim 3 {M_\odot}$ (e.g.@Kalogera1996). Hence we conclude that the low-mass systems show strong evidence of a mass gap. The distribution of minimum black hole masses for the analysis of the combined sample (i.e. including the high-mass systems) is shown in Figure \[fig:high-min-mass\]. For the most favored model, the exponential, we find that $M_{1\%} > 4.5$ ${M_\odot}$ with 90% confidence. We therefore conclude that there is strong evidence for a mass gap in the combined sample as well. Table \[tab:mmin-quants\] gives the 10%, 50% (median), and 90% quantiles for the minimum black hole mass implied by the low-mass sample; Table \[tab:mmin-quants-high\] gives the same, but for the combined sample of systems. Model 10% 50% 90% ---------------------------------------------------------- ----- ----- ----- Power Law (Section \[sec:power-law\]) 4.3 6.1 6.6 Gaussian (Section \[sec:gaussian\]) 2.9 4.4 5.5 Log Normal (Section \[sec:log-normal\]) 3.9 4.9 5.8 Exponential (Section \[sec:exponential\]) 5.3 6.0 6.5 Two Gaussian (Section \[sec:gaussian\]) 2.4 4.2 5.5 Histogram (1 Bin, Section \[sec:non-parametric-models\]) 4.4 5.5 6.2 Histogram (2 Bin, Section \[sec:non-parametric-models\]) 4.0 5.4 6.3 Histogram (3 Bin, Section \[sec:non-parametric-models\]) 3.2 5.2 6.3 Histogram (4 Bin, Section \[sec:non-parametric-models\]) 2.4 4.7 6.0 Histogram (5 Bin, Section \[sec:non-parametric-models\]) 1.9 4.4 6.0 : \[tab:mmin-quants\] The 10%, 50% (median), and 90% quantiles for the minimum black hole mass (in units of ${M_\odot}$) implied by the low-mass sample in the context of the various models for the black hole mass distribution. The models are listed in order of preference from model selection (Section \[sec:low-mass-model-selection\], Figure \[fig:rj\], and Table \[tab:rj\]). Model 10% 50% 90% ---------------------------------------------------------- ------- ----- ----- Exponential (Section \[sec:exponential\]) 4.5 5.4 6.1 Two Gaussian (Section \[sec:gaussian\]) 2.3 4.3 5.5 Power Law (Section \[sec:power-law\]) 5.1 5.9 6.4 Histogram (3 Bin, Section \[sec:non-parametric-models\]) 4.0 5.5 6.3 Histogram (4 Bin, Section \[sec:non-parametric-models\]) 3.4 5.3 6.4 Histogram (2 Bin, Section \[sec:non-parametric-models\]) 4.4 5.5 6.2 Histogram (5 Bin, Section \[sec:non-parametric-models\]) 2.8 5.0 6.2 Gaussian (Section \[sec:gaussian\]) -0.64 1.4 3.4 Histogram (1 Bin, Section \[sec:non-parametric-models\]) 2.9 4.4 5.5 : \[tab:mmin-quants-high\] The 10%, 50% (median), and 90% quantiles for the distribution of minimum black hole masses (in units of ${M_\odot}$) implied by the combined sample in the context of the various models for the black hole mass distribution. The models are listed in order of preference from model selection (Section \[sec:high-mass-model-selection\], Figure \[fig:high-rj-evidence\], and Table \[tab:rj-high\]). Summary and Conclusion {#sec:conclusion} ====================== We have presented a Bayesian analysis of the mass distribution of stellar-mass black holes in X-ray binary systems. We considered separately a sample of 15 low-mass, Roche lobe-filling systems and a sample of 20 systems containing the 15 low-mass systems and five high-mass, wind-fed X-ray binaries. We used MCMC methods to sample the posterior distributions of the parameters implied by the data for five parametric models and five non-parametric (histogram) models for the mass distribution. For both sets of samples, we used reversible jump MCMCs (exploiting a new algorithm for efficient jump proposals in such calculations) to perform model selection on the suite of models. The consideration of a broad range of models and the model-selection analysis, along with consideration of the full posterior distribution on the minimum black hole mass, significantly expand earlier statistical analyses of black hole mass measurements [@Bailyn1998; @Ozel2010]. For the low-mass systems, we found the limits on model parameters in Tables \[tab:low-mass-parametric\] and \[tab:low-mass-non-parametric\]. The relative model probabilities from the model selection are given in Table \[tab:rj\]. The most favored model for the low-mass systems is a power law. The equivalent limits on the model parameters for the combined systems are given in Tables \[tab:high-mass-parametric\] and \[tab:high-mass-non-parametric\]. Unlike the low-mass systems, the most favored model for the combined sample is the exponential model. This difference indicates that the low-mass subsample is not consistent with being drawn from the distribution of the combined population. We found strong evidence for a mass gap between the most massive neutron stars and the least massive black holes. For the low-mass systems, the most favored, power law model gives a black hole mass distribution whose 1% quantile lies above $4.3 {M_\odot}$ with 90% confidence. For the combined sample of systems, the most favored, exponential model gives a black hole mass distribution whose 1% quantile lise above $4.5 {M_\odot}$ with 90% confidence. Although the study methodology was different, the existence of a mass gap was pointed out first by @Bailyn1998 and most recently by @Ozel2010 (who did not consider a power law model, and applied both Gaussian and exponential models to the low-mass systems, where the exponential is strongly disfavored compared to our power-law model). Theoretical expectations for the black hole mass distribution have been examined in @Fryer2001. They considered results of supernova explosion and fallback simulations [@Fryer1999] applied to single star populations; they also included a heuristic treatment of the possible effects of binary evolution on the black hole mass distribution. It is interesting that we find the most-favored model for the combined sample to be an exponential, as discussed by @Fryer2001. On the other hand, we find the most-favored model for the low-mass sample to be a power law, with the exponential model strongly disfavored for this sample. In agreement with @Bailyn1998 and @Ozel2010, we too conclude that both the low-mass and combined samples require the presence of a gap between 3 and 4–4.5 ${M_\odot}$. @Fryer1999 discussed two possible causes of such a gap: (1) a step-like dependence of supernova energy on progenitor mass or (2) selection biases. Current simulations of core collapse in massive stars may shed light on the dependence of supernova energy on progenitor mass. Selection biases can occur because the X-ray binaries with very low-mass black holes systems are more likely to be persistently Roche-lobe overflowing, preventing dynamical mass measurements. @Ozel2010 conclude that the presence of such biases is not enough to account for the gap, arguing that the number (26) of observed persistent X-ray sources not known to be neutron stars is insufficient to populate the 2–5 ${M_\odot}$ region of any black hole mass distribution that rises toward low masses. Population synthesis models incorporating sophisticated treatment of binary evolution and transient behavior (e.g. @Fragos2008 [@Fragos2009]) could help shed light on this possibility. WMF, NS, and VK are supported by NSF grants CAREER AST-0449558 and AST–0908930. AC, LK, and CB are supported by NSF grant NSF/AST-0707627. IM acknowledges support from the NSF AAPF under award AST-0901985. Calculations for this work were performed on the Northwestern Fugu cluster, which was partially funded by NSF MRI grant PHY-0619274. We thank Jonathan Gair for helpful discussions. Markov Chain Monte Carlo {#sec:mcmc} ======================== MCMC methods produce a Markov chain (or sequence) of parameter samples, $\{ {\vec{\theta}}_i \, | \, i = 1, \ldots \}$, such that a particular parameter set, ${\vec{\theta}}$, appears in the sequence with a frequency equal to its probability according to a posterior, $p({\vec{\theta}}|d)$. A Markov chain has the property that the transition probability from one element to the next, $p({\vec{\theta}}_i \to {\vec{\theta}}_{i+1})$, depends only on the value of ${\vec{\theta}}_i$, not on any previous values in the chain. One way to produce a sequence of MCMC samples is via the following algorithm, first proposed by @Metropolis1953 and used widely in the physical sciences thereafter: 1. Begin with the current sample, ${\vec{\theta}}_i$. 2. Propose a new sample, ${\vec{\theta}}_p$, by drawing randomly from a “jump proposal distribution” with probability $Q({\vec{\theta}}_i \to {\vec{\theta}}_p)$. Note that $Q({\vec{\theta}}_i \to {\vec{\theta}}_p)$ can depend on the current parameters, ${\vec{\theta}}_i$, and any other “constant” data, but cannot examine the history of the chain beyond the most recent point. This is necessary to preserve the Markovian property of the chain. 3. Compute the “acceptance” probability, $$\label{eq:paccept} p_{\textnormal{accept}} \equiv \frac{p({\vec{\theta}}_p|d)}{p({\vec{\theta}}_i|d)} \frac{Q({\vec{\theta}}_p \to {\vec{\theta}}_i)}{Q({\vec{\theta}}_i \to {\vec{\theta}}_p)}$$ 4. With probability $\min(1,p_{\textnormal{accept}})$ “accept” the proposed ${\vec{\theta}}_p$, setting ${\vec{\theta}}_{i+1} = {\vec{\theta}}_p$; otherwise set ${\vec{\theta}}_{i+1} = {\vec{\theta}}_i$. This algorithm is more likely to accept a proposed jump when it increases the posterior (the first factor in Equation ) and when it is to a location in parameter space from which it is easy to return (the second factor in Equation ); the combination of these influences in Equation ensures that the equilibrium distribution of the chain is $p({\vec{\theta}}|d)$. As $i \to \infty$ the samples ${\vec{\theta}}_i$ are distributed according to $p({\vec{\theta}}|d)$. In practice the number of samples required before the chain appropriately samples $p({\vec{\theta}}|d)$ depends strongly on the jump proposal distribution; proposal distributions that often propose jumps toward or within regions of large $p({\vec{\theta}}|d)$ can be very efficient, while poor proposal distributions can require prohibitively large numbers of samples before convergence. There is no foolproof test for the convergence of a chain. In this work, we test the convergence of our chains several ways. The most basic is by comparing the statistics calculated from the entire chain to statistics calculated from only the first half of the chain; when the chain has converged, the two calculations agree. This is a necessary, but not sufficient, condition for convergence. We also have examined the sample traces from our chains, to see that the chains have densely and randomly sampled parameter space. A representative sample trace from our MCMC for the power-law model applied to the low-mass systems appears in Figure \[fig:sample-trace\]. Sample traces from MCMCs with other models are similar. Finally, for our most quantitative test of convergence we use the [ gibbsit]{} code to implement the Raftery-Lewis convergence test for our quantile measures [@Raftery1992; @Raftery1992a; @Raftery1995]. The most extreme quantile is the most difficult to determine accurately because—by design—there will be fewer samples in the tail than in the main body of a distribution obtained from an MCMC. Accordingly, we focus on the 90% confidence limit on the minimum black hole mass. For a quantile, $q$, the Raftery-Lewis test attempts to estimate how many samples from an MCMC are needed to determine $q$ to within $\pm r$ at a confidence $s$. We use $r = 0.0125$ and $s = 0.95$. The Raftery-Lewis test approximates the MCMC chain as a two-state Markov chain, the two states being “within the quantile in question” and “outside the quantile in question.” The 2x2 transition matrix for this two-state Markov chain and the associated uncertainty can be calculated analytically [@Raftery1992; @Raftery1992a; @Raftery1995], allowing the algorithm to determine the number of sample points required before the quantile of interest is determined sufficiently accurately. For our chains, in the worst case (the power-law on the lower-mass samples, as shown in Figures \[fig:power-law\] and \[fig:sample-trace\]), we have twice as many samples as the Raftery-Lewis convergence test estimates we need to determine the 90% confidence level on the minimum mass; for all the other chains, we have about 20 times as many samples as the Raftery-Lewis criterion estimates are required. We suspect that the slow convergence of the power-law model on the lower-mass systems is due to the long tails in the mass parameters and the width of the distribution on the power-law exponent. In any case, the Raftery-Lewis test indicates that all our chains are converged sufficiently to determine the 90% quantile to within $\sim 1\%$. We begin the chain at an arbitrary point in parameter space; this is equivalent to taking a finite section of an infinite chain that begins with the chosen point. Every point in parameter space occurs in an infinite chain, and no section of the chain is better than any other, so a sufficiently long, but finite, section of the infinite chain chosen in this manner can be representative of the statistics of the chain as a whole. However, because consecutive samples in a chain are correlated with each other, the beginning of our finite chain has a “memory” of the starting point; we discard enough points at the beginning of the finite chain that we can be confident it does not retain a memory of the arbitrary starting point. The points discarded in this way are commonly called “burn-in” points. Reversible-Jump MCMC {#sec:reversible-jump-mcmc} ==================== The algorithm described here for the reversible jump MCMC we have used for the model comparison in this work is more fully described in @Farr2010. In particular, @Farr2010 demonstrates the efficiency gains from the algorithm, and shows experimentally that the algorithm indeed provides a consistent RJMCMC, with the correct relative model probabilities for the case where the Bayes factor between models can be calculated analytically. Consider the problem of model selection among a set of models, and the “super-model” that encompasses all the models under consideration. The parameter space of the super-model consists of a discrete parameter that identifies the choice of model, $M_i$, and the continuous parameters appropriate for this model, ${\vec{\theta}}_i$. We denote a point in the super-model parameter space by $\{M_i, {\vec{\theta}}_i\}$; each such point is a statement that, e.g., “the underlying mass distribution for black holes in the galaxy is a Gaussian, with parameters $\mu$ and $\sigma$,” or “the underlying mass distribution for black holes in the galaxy is a triple-bin histogram with parameters $w_1$, $w_2$, $w_3$, and $w_4$,” or .... To compare models, we are interested in the quantity (see Equation ) $$p(M_i|d) = \int d{\vec{\theta}}_i p(M_i, {\vec{\theta}}_i|d).$$ If we perform an MCMC in the super-model parameter space, then we obtain a chain of samples $\{M_i, {\vec{\theta}}_i \, | \, i = 1, \ldots\}$ distributed in parameter space with density $p(M_i,{\vec{\theta}}_i|d) d{\vec{\theta}}_i$ and we can estimate the integral as $$p(M_i|d) = \int d{\vec{\theta}}_i p(M_i, {\vec{\theta}}_i|d) \approx \frac{N_i}{N},$$ where $N_i$ is the number of samples in the chain lying in the parameter space of model $M_i$ and $N$ is the total number of samples in the chain. The fraction of samples lying in the parameter space of model $M_i$ gives the probability of that model relative to the other models under consideration. To perform the MCMC in the super-model parameter space, we must propose jumps not only between points in a particular model’s parameter space, but also between the parameter spaces of different models. For this MCMC to be efficient, proposed jumps into a model from another should favor regions with large posterior; when the posterior is highly-peaked in a small region of parameter space, proposed jumps outside this region are unlikely to be accepted, and the reversible-jump MCMC samples will require a very long chain to properly sample the “super-model” posterior. We can exploit the information we have from single-model MCMCs to generate efficient jump proposal distributions for our reversible jump MCMC. We would like to propose jumps that roughly follow the distribution of samples in the single-model MCMCs. We can do this by assigning a neighborhood to each point in the sample using an algorithm we will describe in the following paragraphs; the neighborhoods are non-overlapping, completely cover the region of parameter space with prior support, and contain only one point from the MCMC samples. To propose a jump into model $M_i$, we choose a point uniformly from that single-model MCMC and then propose a jump drawn uniformly from that point’s neighborhood. This is equivalent to drawing from a piecewise-constant approximation to the single-model posterior, where each neighborhood contributes a constant fraction, $1/N_i$, to the cumulative jump probability. In regions of high density the neighborhoods are smaller, and the jump probability density is correspondingly higher. Because the neighborhoods cover the entire region of prior support, it is possible for the proposal to propose any point in parameter space with prior support (though points in regions of low single-model posterior are of course unlikely to be proposed). To assign a neighborhood to each point in a single-model MCMC we use a data structure called a kD-tree. A kD-tree is a binary space-partitioning tree. To construct a kD-tree, we begin with the set of points from a single-model MCMC and a box in parameter space bounding the region of prior support (which must necessarily enclose all the points). The construction proceeds recursively: we choose a dimension[^3] along which to divide the points, find the median point along that dimension and its nearest neighbor, and divide the box at the midpoint between these two points, producing two sub-boxes. We then partition the points into those to the left (i.e. smaller coordinate along the given dimension) and right of the dividing line, and repeat this procedure for each subset and the corresponding bounding box, until we have only one point in each box. An example of the neighborhoods that result from a two-dimensional kD-tree constructed around a Gaussian point distribution appears in Figure \[fig:kD-tree\]. Construction of a kD-tree is an ${\mathcal{O}\left( N\log N \right)}$ operation, where $N$ is the number of points in the tree. Median finding is ${\mathcal{O}\left( n \right)}$, where $n$ is the number of points from which the median is to be obtained. At level $i$ in the tree, there are $2^i$ subsets of points, each of length ${\mathcal{O}\left( N/2^i \right)}$, so the total cost of the $2^i$ median calculations is ${\mathcal{O}\left( N \right)}$ at each level. There are ${\mathcal{O}\left( \log N \right)}$ levels in the tree, yielding a total construction cost for the tree of ${\mathcal{O}\left( N \log N \right)}$. To find the neighborhood of a point using the tree, we begin at the root of the tree, and examine the two sub-boxes at the next level down. The point will be in one of them; following that branch, we have again two sub-boxes, one of which contains the point; following that branch.... Eventually, the search terminates at a leaf of the tree, containing the point in question. The box at the leaf defines the neighborhood of the point in the jump proposal algorithm described above. The total cost for this operation is proportional to the number of levels in the tree, which is ${\mathcal{O}\left( \log N \right)}$. In addition to the validation tests in @Farr2010, we have validated our interpolation method with the following test. We imagine that we have a data set that can be fit by two models: an “egg-crate” model with likelihood $$L({\vec{\theta}}|d) = 2^N \prod_{i = 1}^N \sin^2 \left( 2\pi n \theta_i \right),$$ and a single-Gaussian model with likelihood $$L({\vec{\theta}}|d) = \frac{1}{\left( 2\pi \right)^{N/2} \prod_{i=1}^N \sigma_i } \prod_{i = 1}^N \exp\left( - \frac{\left( \theta_i - \mu_i \right)^2}{2 \sigma_i^2} \right),$$ where the number of dimensions in each parameter space is $N = 5$, and we choose $2n = 10$ peaks along each dimension for the egg-crate model. We restrict the parameter space to ${\vec{\theta}}\in [0,1]^N$, and choose a uniform prior on ${\vec{\theta}}$ within that region. We choose $\mu_i = 1/2$, and $\sigma_i = 1/(20 i)$, so the Gaussian peak is well-contained within the region of interest, being at least $10\sigma$ away from the boundaries of the region. From the point of view of an (RJ)MCMC, the precise data that could produce such unusual likelihoods from two models are irrelevant; the algorithm only cares about the form of the likelihood and prior functions in parameter space. These likelihood functions provide a good test case for our interpolation technique: the egg-crate model has a broad, multi-modal likelihood, while the Gaussian model’s likelihood is very concentrated in a small region of parameter space. Particularly for the Gaussian model, an RJMCMC without interpolation—one that proposes inter-model jumps from the prior, for example—would be extremely inefficient because the region of parameter space with significant posterior support is so small. Our choice of likelihood and prior implies that the models have equal evidence: $$p(d|M_i) = \int_{[0,1]^N} d^N {\vec{\theta}}\, L({\vec{\theta}}|d) p({\vec{\theta}}) = 1.$$ Using individual MCMC parameter samples to construct interpolations of the single-model posterior, and running a $10^6$ sample reversible-jump MCMC, we find $$\frac{p(d|\textnormal{egg crate})}{p(d|\textnormal{Gaussian})} \simeq \frac{N_{\textnormal{egg crate}}}{N_{\textnormal{Gaussian}}} = \frac{499285}{500715} \simeq 0.997,$$ which has an error of $3\times 10^{-3}$, of order $1/\sqrt{N_{\textnormal{egg crate}}} \sim 1/\sqrt{N_{\textnormal{Gaussian}}} \sim 1.4 \times 10^{-3}$, as would be expected for 1M independent samples from a binomial distribution with $p = 0.5$. We conclude, as in @Farr2010, that our interpolation method leads to accurate and efficient reversible-jump MCMCs. [^1]: A similar analysis of the neutron star mass distribution can be found in @Finn1994. [^2]: We also attempted to compute $p(M_i|d)$ using two other methods: the well-known harmonic-mean estimator and the direct integration methods described in @Weinberg2010. The harmonic mean is known to be very sensitive to outlying points in the MCMC in general, and we found this to be true in our specific application. The statistical properties of the direct integration algorithm from @Weinberg2010 are less certain, but we found that it was quite noisy in our application compared to the reversible-jump MCMC. Due to the statistical noise in the other two methods, we use the results from our reversible jump MCMC analysis for model selection. [^3]: Our algorithm chooses the dimension along which the numerical extent of the points is largest. Other choices are possible; some algorithms cycle through the dimensions in order, while others choose a random dimension for each subdivision. Our goal by picking the longest dimension is to produce neighborhoods that are “square,” at least in the chosen parametrization.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Recently, @Sahni2014 combined two independent measurements of $H(z)$ from BAO data with the value of the Hubble constant $H_0 = H(z=0)$, in order to test the cosmological constant hypothesis by means of an improved version of the $Om$ diagnostic. Their result indicated a considerable tension between observations and predictions of the $\Lambda$CDM model. However, such strong conclusion was based only on three measurements of $H(z)$. This motivated us to repeat similar work on a larger sample. By using a comprehensive data set of 29 $H(z)$, we find that discrepancy indeed exists. Even though the value of $\Omega_{m,0} h^2$ inferred from $Omh^2$ diagnostic depends on the way one chooses to make a summary statistics (weighted mean or the median), the persisting discrepancy supports the claims of @Sahni2014 that $\Lambda$CDM model may not be the best description of our Universe.' author: - 'Xuheng Ding$^{1}$, Marek Biesiada$^{1,2}$, Shuo Cao$^{1}$, Zhengxiang Li$^{1}$, and Zong-Hong Zhu$^{1}$' title: 'Is there evidence for dark energy evolution?' --- Introduction ============ The discovery of accelerating expansion of the Universe [@Riess98; @Perlmutter99] brought us a mystery which became one of the most important challenges for modern cosmology and theoretical physics. This phenomenon has been confirmed since then in manifold ways using different probes like supernovae Ia, acoustic peaks in the CMBR  [@CMB1; @CMB2], baryon acoustic oscillations (BAO)  [@BAO; @Eisenstein11] in the distribution of the large scale structure etc. By now all this observational evidence was concordant with the simplest assumption that there exists non-vanishing cosmological constant $\Lambda$. While being the simplest, such a model is not theoretically satisfactory. There is a huge discrepancy if one tries to motivate $\Lambda$ as a zero-point quantum vacuum energy. Therefore alternative explanations appeared invoking the scalar field which settled down in an attractor [@Ratra88]. This motivated to push forward the phenomenological picture of the so called dark energy described as a fluid with barotropic equation of state $p = w \rho$, where $w$ can either be constant – the so called “quintessence”[@quintessence](as a value characteristic for the fixed point attractor) or evolving in time [@Chevalier01; @Linder03] since the scalar field could be expected to evolve in time. The main drawback of such an approach is that it makes explicit assumption about the dark energy before its specific model (a quintessence or evolving equation of state) can be tested on observational data. Moreover, alternatives to dark energy such like modified gravity [@DGP; @FR1; @FR2; @FT] cannot be easily tested within this phenomenology. All observational tests of quintessence pin-point its value close to $w=-1$ (within the uncertainties) which is equivalent to the cosmological constant. On the other hand tests of varying in time cosmic equation of state are much less restrictive and do not allow to make a decisive statement whether dark energy equation of state evolved or not. Therefore, we clearly need alternative probes: capable to discriminate between the cosmological constant and evolving dark energy not relying on the dark energy assumption and its equation of state parametrization. One promising approach to such probes has been initiated by @Sahni2008 and developed further in [@Sahni2012]. By properly rearranging the equation for the Hubble function in the flat $\Lambda$CDM model: $H^2(z) = H_0^2 [ \Omega_m (1+z)^3 + 1 - \Omega_m ] $, they noticed that the so called $Om(z)$ diagnostic (where ${\tilde h} = H(z) / H_0$): $$\label{Omz} Om(z) = \frac{{\tilde h}^2(z)-1}{(1+z)^3 - 1}$$ should be constant and equal exactly to the present mass density parameter, if the $\Lambda$CDM model is the true one: $Om(z)_{\Lambda CDM} = \Omega_{m,0}$. This is remarkable and differentiates between $\Lambda$CDM and other dark energy models (including evolving dark energy). Let us remark that essentially the same idea has also been formulated by @Zunckel2008 who called it “a litmus test” for the canonical $\Lambda$CDM model. Developing this method [@Sahni2012] they also considered a generalized two-point diagnostics $$\label{Omz2} Om(z_1,z_2) = \frac{{\tilde h}^2(z_1)-{\tilde h}^2(z_2)}{(1+z_1)^3 - (1+z_2)^3}$$ which should also be equal to $\Omega_{m,0}$ within the $\Lambda$CDM model but has an advantage that having $H(z)$ measurements at $n$ different redshifts, one has $\frac{n(n-1)}{2}$ two point diagnostics, hence a considerably increased sample for inference. In their latest paper @Sahni2014 used three accurately measured values of $H(z)$ to perform this test. These were: the $H(z=0)$ measurement by [@Riess2011; @PlanckXVI], $H(z=0.57)$ measurement from SDDS DR9 [@Samushia] and the latest $H(z)=2.34$ measurement from the $Ly{\alpha}$ forest in SDSS DR11 [@Delubac]. They found that all three values of the two-point diagnostics $Om(z_1,z_2)h^2$ were in strong tension with the $\Omega_{m,0}h^2$ reported by Planck [@PlanckXVI]. It has been noticed [@Delubac; @Sahni2014] that such result could be in tension not only with the $\Lambda$CDM model but with other dark energy models based on the General Relativity. Because this conclusion could be of a paramount importance for the dark energy studies, an update of this test with a larger sample of $H(z)$ is essential. Data, Methodology and Results {#sec:method} ============================= As a basic dataset we used a sample of 29 $H(z)$ measurements taken from the compilation of @Chen2014 modified in the following way: one data point at $z=0.6$ coming from @Blake2012 was added and two data points from @Gatzanaga were withdrawn. The reason for deleting aforementioned two points is that these results have been debated in subsequent papers e.g. by @Miralda-Escude, @Kazin or @Cabre. For the sake of consistency with @Sahni2014, we have also taken the latest BAO measurement by @Delubac $H(z=2.34) = 222\pm 7$ instead of $H(z=2.3) = 224\pm 8$ of @Busca2012. After these changes our data are essentially like the ones used by @Farooq with @Busca2012 measurement replaced by @Delubac. During the analysis we also made assessments on subsamples of this biggest one, as we will explain further. Part of the data comes from cosmic chronometers – spectroscopy of galaxies assumed to evolve passively [@JimenezLoeb]. Hereafter, this differential age approach will be quoted as DA for short. The other part of the data comes from BAO – including the data points used by @Sahni2014. Data are summarized in Table 1. Then we proceed in exactly the same way as @Sahni2014, i.e. for each pair of redshifts $(z_i,z_j)$ we calculate the improved $Om$ diagnostic: $$\label{improved} Omh^2(z_i,z_j) = \frac{h^2(z_i)- h^2(z_j)}{(1+z_i)^3 - (1+z_j)^3}$$ where: $h(z) = H(z) / 100km/sec/Mpc$ is dimensionless Hubble parameter. In particular case of $\Lambda$CDM model the improved diagnostic Eq. (\[improved\]) should be equal to $\Omega_{m,0}h^2$ which luckily is the quantity best constrained by the CMBR data, e.g. from Planck [@PlanckXVI]. Because the sample of $29$ data points leads to $406$ different pairs, we summarize our calculations on Figure \[fig1\]. One can see that the distribution of inferred $\Omega_{m,0}h^2$ is skewed and centered around the value different from this reported by @Sahni2014. [llll]{} \[table1\] 0.07 & 69 & 19.6 & DA\ 0.1 & 69 & 12 & DA\ 0.12 & 68.6 & 26.2 & DA\ 0.17 & 83 & 8 & DA\ 0.179 & 75 & 4 & DA\ 0.199& 75 & 5 & DA\ 0.2& 72.9 & 29.6 & DA\ 0.27 & 77 & 14 & DA\ 0.28 & 88.8 & 36.6 & DA\ 0.35 & 82.7 & 8.4 & BAO\ 0.352 & 83 & 14 & DA\ 0.4 & 95 & 17 & DA\ 0.44 & 82.6 & 7.8 & BAO\ 0.48 & 97 & 62 & DA\ 0.57 & 92.9 & 7.8 & BAO\ 0.593& 104 & 13 & DA\ 0.6 & 87.9 & 6.1 & BAO\ 0.68 & 92 & 8 & DA\ 0.73 & 97.3 & 7 & BAO\ 0.781 & 105 & 12 & DA\ 0.875 & 125 & 17 & DA\ 0.88 & 90 & 40 & DA\ 0.9 & 117 & 23 & DA\ 1.037 & 154 & 20 & DA\ 1.3 & 168 & 17 & DA\ 1.43 & 177 & 18 & DA\ 1.53 & 140 & 14 & DA\ 1.75 & 202 & 40 & DA\ 2.34 & 222 & 7 & BAO\ ![image](fig1.eps){width="12cm"} If one is to make a summary statistics one can do it in two ways. First, straightforward way would be to calculate the weighted mean: $$\label{weighted} Omh^2_{(w.m.)} = \frac{\sum_{i=1}^{29} \sum_{j \textgreater i} Omh^2(z_i,z_j)/ \sigma^2_{Omh^2,ij}}{\sum_{i=1}^{29} \sum_{j \textgreater i} 1 / \sigma^2_{Omh^2,ij}}$$ and the standard deviation: $$\label{SD} \sigma_{(w.m.)} = \left( \sum_{i=1}^{29} \sum_{j \textgreater i} 1 / \sigma^2_{Omh^2,ij} \right)^{-1/2}$$ where: $$\sigma^2_{Omh^2,ij} = \frac{4(h^2(z_i) \sigma^2_{h(z_i)} + h^2(z_j) \sigma^2_{h(z_j)}) }{((1+z_i)^3 - (1+z_j)^3)^2}$$ and $\sigma_{h(z_i)}$ denotes the uncertainty of the $i-th$ measurement. We also assume that redshifts are measured accurately. This well known and often used approach relies, however, on several strong assumptions: statistical independence of the data, no systematic effects, Gaussian distribution of the errors. These assumptions, especially the Gaussianity of errors are not valid here. Hence the weighted mean of $$\label{weighted mean} Omh^2_{(w.m.)} = 0.1253 \pm 0.0021$$ is not a reliable measure, as one can see form the histogram on Figure \[fig1\]. We will comment more on this non-Gaussianity in a moment. Therefore we took another, much more robust approach: to calculate the median. This approach was pioneered by the paper of @GottIII and then used by the others, e.g. quite recently by [@RatraMedian]. The robustness of this method stems from the fact that if systematic effects are absent, half of the data is expected to be higher and another half - lower than the median. Then, as the number of measurements $N$ increases, calculated median approaches its true value. So the median has clear and robust meaning without need to assume anything about the error distribution. From the definition of the median, probability that any particular measurement, one of $N$ independent measurements is higher than the true median is $50 \%$. Consequently, the probability that $n$ observations out of the total number of $N$ is higher than the median follows the binomial distribution: $P = 2^{-N} N! / [n!(N-n)!]$. This allows to calculate in a simple manner the confidence regions (e.g. $68 \%$ confidence region) of the median value estimated from the sample. Proceeding this way we have obtained: $$\label{Median} Omh^2_{(median)} = 0.1550^{+0.0065}_{- 0.0072}$$ In order to facilitate comparison between the inferred values of $\Omega_{m,0}h^2$ obtained from two statistical approaches and the Planck data, we display the results in Figure \[fig2\]. ![image](fig2a.eps){width="70mm"} ![image](fig2b.eps){width="70mm"} ![image](fig2c.eps){width="70mm"} ![image](fig2d.eps){width="70mm"} Inconsistency between $Omh^2(z_1,z_2)$ diagnostic calculated on $H(z)$ data and the Planck value of $\Omega_{m,0}h^2$ as well as mutual inconsistency between weighted averaging and median statistics schemes, motivated us to make some more detailed tests. First, we recalculated $Omh^2$ for three sub – samples: excluding the highest redshift $z=2.34$ measurement, using only DA data and using only BAO data. Results are visualized on Figure \[fig2\] and shown in more details in Table 2. One can see that $z=2.34$ point had a big leverage on the weighted average value — dropping this point one achieves agreement with the $\Lambda$CDM Planck value. However, the question remains whether the weighted average scheme is appropriate. Therefore, following @Chen and @Crandall, we have drawn histograms of distribution of our measurements as a function of the number of standard deviations $N_{\sigma}$ away from central estimates (weighted mean and the median respectively). Because of limited space we do not show them here, but report in Table 2 corresponding percentage of the distribution falling within $\pm 1\sigma$ i.e. $|N_{\sigma}|<1$. One clearly sees that they strongly deviate from the Gaussian $68 \;\%$ expectation. We also tested the $N_{\sigma}$ distribution with Kolmogorov-Smirnov test which strongly rejected the hypothesis of Gaussianity in each sub-sample (with p-values ranging from $10^{-4}$ to $10^{-7}$). Therefore our conclusion is that weighted average scheme is not appropriate here and the median statistics is more reliable. Both statistical methods: weighted mean and median produce similar results for the BAO data, but with addition of DA data these two schemes give drastically different results. This may suggest the existence of some systematic error in DA data. It is not obvious by itself, because the nonlinear relation between input variables ($H(z)$) underlying the $Omh^2$ diagnostic may be the source of asymmetric uncertainties of the latter. However, the fact that BAO and DA data deviate from the $\Lambda$CDM expected result in opposite directions strongly supports suggestion of unaccounted systematics. This will be the subject of a separate study. [lllll]{} \[table2\] Full sample ($n=29$) & $0.1253 \pm 0.0021$ & $80.54 \%$ & $0.1550^{+0.0065}_{-0.0072}$ & $75.62 \%$\ $z=2.34$ excluded ($n=28$) & $0.1404 \pm 0.0047$ & $77.78 \%$ & $0.1682^{+0.0075}_{-0.0074}$ & $82.80 \%$\ DA only ($n=23$) & $0.1448 \pm 0.0057$ & $77.47 \%$ & $0.1852^{+0.0032}_{-0.0079}$ & $86.56 \%$\ BAO only ($n=6$) & $0.1231 \pm 0.0045$ & $100 \%$ & $0.1218^{+0.0011}_{-0.0035}$ & $100 \%$\ Conclusions {#sec:conclusions} =========== In this paper we attempted to assess the $Omh^2(z_1,z_2)$ diagnostic introduced and developed by @Sahni2012. The main reason to do so was the recent paper by @Sahni2014 where they claimed that recent precise measurements of expansion rates at different redshifts suggest a severe tension with the $\Lambda$CDM model. We repeated this on a much more comprehensive data set of $29$ $H(z)$ obtained by two techniques: DA and BAO. One can see from the Table 1 that uncertainties of $H(z)$ obtained by different methods are different. Even within the same methodology (DA) uncertainties are different from case to case. The $Omh^2(z_1,z_2)$ diagnostic, involving ratio of certain differences (see Eq. (\[improved\])) calculated on our data produces an asymmetric distribution. This means that the weighted mean is not a reliable summary measure. Therefore we used a more robust approach to calculate the median. Our result is that the value of $Omh^2$ inferred from $Omh^2$ diagnostic is indeed in tension with the one obtained by Planck (under assumption of the $\Lambda$CDM model). In our case, this tension is not so severe as in @Sahni2014 ($Omh^2 = 0.122 \pm 0.01$ vs. $\Omega_{m,0}h^2_{Planck} = 0.1426 \pm 0.0025$). Even though the inferred value is sensitive to the way one chooses to make a summary statistics: the weighted mean value is lower and the median value is higher than that obtained by Planck, they are both discrepant with each other. Non-Gaussianity in the data suggests that median statistics approach is more appropriate, so this tension cannot be alleviated by excluding high redshift data. This supports the claims of @Sahni2014 that the concordance model ($\Lambda$CDM) might not be the true or even the best one describing our Universe. Therefore, we also performed a quick test whether XCDM or CPL models (simplest evolving equation of state parametrization) best fitted to the Planck or WMAP9 data agree better with $H(z)$ data. In such a case the $Omh^2(z_i,z_j)$ diagnostic defined be Eq.(\[improved\]) will no longer be a single number $\Omega_{m,0}h^2$, but rather: $Omh^2(z_i,z_j) = \Omega_{m,0}h^2 + (1 - \Omega_{m,0})h^2 \frac{(1+z_i)^{3(1+w)} - (1+z_j)^{3(1+w)}}{(1+z_i)^3 - (1+z_j)^3}$ for XCDM and $Omh^2(z_i,z_j) = \Omega_{m,0}h^2 + (1 - \Omega_{m,0})h^2\\ \frac{(1+z_i)^{3(1+w_0+w_a)}e^{\frac{- 3 w_a z_i}{1+z_i}} - (1+z_j)^{3(1+w_0+w_a)}e^{\frac{- 3 w_a z_j}{1+z_j}}}{(1+z_i)^3 - (1+z_j)^3}$ for CPL parametrization. Therefore, for each pair we formed the residuals $R(z_i,z_j)$ by subtracting the right-hand sides from the left-hand sides. Because such residuals inherit non-Gaussianity from $Omh^2(z_i,z_j)$ we summarized our findings with weighted average $R_{(w.m.)}$ and the median $R_{(median)}$. If a particular model (XCDM or CPL) agreed better with the $H(z)$ data than $\Lambda$CDM, then its $R$ should have been closer to zero than $R(\Lambda CDM)_{(w.m.)} = -0.0173 \pm 0.0033$ or $R(\Lambda CDM)_{(median)} = 0.0124^{+0.0070}_{-0.0076}$. In the XCDM model we took the $w = -1.0507^{+0.0469}_{-0.0507}$ parameter according to @Cai best fit to Plack+WMAP9 data. For the CPL parametrization of the equation of state, we used the values: $w_0 = -1.17^{+0.13}_{-0.12}$, $w_a = 0.35^{+0.50}_{-0.49}$ best fitted to WMAP+CMB+BAO+$H_0$+SNe according to @WMAP9. The result is: $R(XCDM)_{(w.m.)} = -0.0176 \pm 0.0025$ and $R(XCDM)_{(median)} = 0.0151^{+0.0063}_{-0.0068}$. For the CPL varying equation of state parametrization we get $R(CPL)_{(w.m.)} = -0.0275 \pm 0.0073$ or $R(CPL)_{(median)} = - 0.0517^{+0.0149}_{-0.0077}$. We see that they do not reconcile the discrepancy but their performance is even worse than $\Lambda$CDM. However, it is not a decisive conclusion, because what remains to be done is find the values of equation of state parameters $w$ or $(w_0,w_a)$ best fitted to the $H(z)$ data (according to $Omh^2$ diagnostics). This will be a subject of a separate study. Having confirmed the discrepancy between $\Lambda$CDM and $H(z)$ data its origin should be studied in greater details. One reason could be that our phenomenological description of accelerated expansion with $\Lambda$CDM is incorrect. On the other hand we pointed out that the conclusion (more specifically the direction of this discrepancy) depends on the statistical approach taken. So one should investigate possible systematics in both methods — DA and BAO and their effect on the conclusion. This is a subject of an ongoing study. Acknowledgements {#acknowledgements .unnumbered} ================ The authors are grateful to the referees for very useful comments which allowed to improve the paper. This work was supported by the Ministry of Science and Technology National Basic Science Program (Project 973) under Grants Nos. 2012CB821804 and 2014CB845806, the Strategic Priority Research Program “The Emergence of Cosmological Structure” of the Chinese Academy of Sciences (No. XDB09000000), the National Natural Science Foundation of China under Grants Nos. 11373014 and 11073005, the Fundamental Research Funds for the Central Universities and Scientific Research Foundation of Beijing Normal University, and China Postdoctoral Science Foundation under grant Nos. 2014M550642 and 2014T70043. M.B. obtained approval of foreign talent introducing project in China and gained special fund support of foreign knowledge introducing project. He also gratefully acknowledges hospitality of Beijing Normal University. Ade, P.A.R., et al., \[Planck Collaboration\] 2013 “Planck 2013 results. XVI. Cosmological parameters”, 2013 \[arXiv:1303.5076\] Bengochea, G. R. & Ferraro, R., 2009, PRD, 79, 124019 Bernardis, P. De, et al., 2000, Nature, 404, 955 Blake, C. et al., 2012, MNRAS, 425, 405-414 \[arXiV: 1204.3674\] Busca, N.G., et al., 2013, A&A, 552, 18 \[arXiv:1211.2616\] Cabr[é]{}, A. & Gazta[ñ]{}aga, E., 2011, MNRAS, 412, L98 Cai, R.-G., Guo, Z.-K. & Tang, B., 2014 \[arXiV:1409.0223\] Chen, G., Gott, J.R. III, Ratra, B., 2003, Publ.Astron.Soc.Pac., 115, 1269-1279 \[astro-ph/0308099\] Chen Y., et al., 2015, JCAP, 02(2015)010 \[arXiv:1312.1443\] Chevalier, M. & Polarski, D., 2001, IJMPD, 10, 213 Crandall, S. & Ratra, B., 2013 \[arXiv:1311.0840v1\] Crandall, S., Houston, S. & Ratra, B., 2014 \[arXiV:1409.7332\] Delubac, T., Bautista, J.E., Busca, N.G. et al., 2015, A&A, 574, A59 \[arXiv:1404.1801\] Dvali, G., Gabadadze, G. & Porrati, M., 2000, PLB, 485, 208 Eisenstein, D. J., et al., 2005, ApJ, 633, 560 Eisenstein, D. J., et al., 2011, AJ, 142, 72 Farooq, O. & Ratra, B., 2013, ApJ, 766, L7 \[arXiv:1301.5243\] Gazta[ñ]{}aga, E., Cabr[é]{}, A. & Hui, L. (2009), MNRAS, 399, 1663 Gott, J. R., Vogeley, M.S., Podariu S. & Ratra, B., 2001, ApJ, 549, 1 Hinshaw, G., et al., 2103, Astrophys. J. Suppl., 208, 19 Kazin, E. A., et al. 2010, ApJ, 719, 1032 Linder, E. V., 2003, PRD, 68, 083503 Jimenez, R. & Loeb, A., 2002, Astrophys. J., 573, 37 \[astro-ph/0106145\] Nojiri, S. & Odintsov, S. D., 2011, Phys. Rep., 505, 59 Miralda - Escud[é]{}, J., 2009 \[arXiV: 0901.1219\] Peebles, P. J. E. & Ratra, B., 1988a, ApJL, 325, 17 Perlmutter, S., et al., 1999, ApJ, 517, 565 Ratra, B. & Peebles, P. E. J., 1988, PRD, 37, 3406 Riess, A. G., et al., 1998, AJ, 116, 1009 Riess, A. G., Macri, L. & Csertano, S., 2011, ApJ, 730, 119 Sahni, V., Shafieloo, A. & Starobinsky, A. A., 2008, PRD, 78, 103502 Samushia, L. et al., 2013, MNRAS, 429, 1514 Shafieloo, A., Sahni, V. & Starobinsky, A. A., 2012, PRD, 86, 103527 Sahni, V., Shafieloo, A. & Starobinsky, A. A., 2014, ApJL, 793, L40 Sotiriou, T. P. & Faraoni, V., 2010, Rev. Mod. Phys., 82, 451 Spergel, D. N., et al., 2003, ApJS, 148, 175 Zunckel, C. and Clarkson, C., 2008, Phys.Rev.Lett. 101, 181301
{ "pile_set_name": "ArXiv" }
--- abstract: 'Carbon and oxygen abundances in F and G main-sequence stars ranging in metallicity from \[Fe/H\] = $-1.6$ to +0.5 are determined from a non-LTE analysis of C[i]{} and O[i]{} atomic lines in high-resolution spectra. Both C and O are good tracers of stellar populations; distinct trends of \[C/Fe\] and \[O/Fe\] as a function of \[Fe/H\] are found for high- and low-alpha halo stars and for thick- and thin-disk stars. These trends and that of \[C/O\] provide new information on the nucleosynthesis sites of carbon and the time-scale for the chemical enrichment of the various Galactic components.' --- Introduction ============ Recent $\Lambda$CDM, hydrodynamical simulations of the formation of the Galaxy (e.g. Zolotov et al. [@zolotov10]; Font et al. [@font11]; McCarthy et al. [@mccarthy12]) predict the existence of two populations of halo stars. The first one is formed [*in situ*]{} in merging and dissipating gas clumps, whereas the other one is [*accreted*]{} from satellite galaxies. For the upper end of the halo metallicity distribution, Zolotov et al. ([@zolotov10]) find that   in the accreted population decreases with increasing \[Fe/H\] relative to the near-constant  in the in-situ population. This is due to a difference in star formation rate (SFR); the chemical enrichment proceeds at a slower rate in satellite galaxies, so that Type Ia SNe start contributing with iron at a lower metallicity. Evidence for two halo populations[^1] in the solar neighborhood with different  trends has been found by Nissen & Schuster ([@nissen10], hereafter NS10). The “high-alpha" stars have ${\mbox{\rm [$\alpha$/Fe]}}\simeq 0.3$, similar to thick-disk stars, and extend in metallicity up to ${\mbox{\rm [Fe/H]}}\simeq -0.4$. The “low-alpha" stars show a declining -trend from $\sim 0.3$dex at ${\mbox{\rm [Fe/H]}}\simeq -1.6$ to $\sim 0.1$dex at ${\mbox{\rm [Fe/H]}}\simeq -0.8$, which is the maximum  reached by this population. Furthermore, the Toomre diagram shows that the high-alpha stars tend to move on prograde orbits, whereas the low-alpha stars have higher space velocities with respect to the LSR and an excess of retrograde orbits. These data are consistent with a scenario where the high-alpha stars have been formed in dissipational mergers of gas clouds, and the low-alpha stars have been accreted from satellite galaxies. The high- and low-alpha halo stars also separate in , , , and , but not in  (Nissen & Schuster [@nissen11]). Furthermore, the high-alpha stars seem to be on average 2 - 3Gyr older than the low-alpha stars (Schuster et al. [@schuster12]). In this paper we report on carbon and oxygen abundances in the two halo populations and compare with abundances in stars having thin- and thick-disk kinematics. Carbon abundances ================= The abundance of carbon was determined from equivalent widths (EWs) of the high-excitation [C[i]{}]{} lines at 5052.2 and 5380.3Å as measured in high-resolution, high-S/N VLT/UVES and NOT/FIES spectra. For each star, a 1D model atmosphere was obtained from the MARCS grid (Gustafsson et al. [@gustafsson08]) by interpolating to the stellar values of , , , and . The model was used to to derive an LTE carbon abundance. Non-LTE corrections from Takeda & Honda ([@takeda05]) were afterwards applied, but they are small, i.e. less than 0.02dex. The atmospheric parameters of the stars were determined spectroscopically by analysing [Fe[i]{}]{} and [Fe[ii]{}]{} lines relative to the same lines in the spectra of two nearby thick-disk stars, HD22879 and HD76932, for which  and are well known from colour indices and Hipparcos parallaxes. This enable us to determine very precise differential parameters and abundances for stars that belong to the same region of the HR-diagram as the standard stars. The absolute values may be more uncertain. Thus we have increased the  scale by +100K relative to the scale used in NS10 to take into account the new IRFM – colour calibration by Casagrande et al. ([@casagrande10]). This has only a small effect on ,  (derived from [Fe[ii]{}]{} lines), and  (derived from neutral lines), but it has a significant effect on C and O abundances derived from high-excitation atomic lines. ![\[C/Fe\] versus \[Fe/H\] with the same classification in high- and low-alpha halo and thick-disk stars as in NS10 and with thin-disk stars from Nissen ([@nissen13]) included.[]{data-label="Nissen:fig1"}](Nissen_PE_fig1.eps){width="9cm"} The derived  values are shown in Fig. \[Nissen:fig1\]. In addition to stars from NS10, we have included stars with thin-disk kinematics from Nissen ([@nissen13]), who also used the $\lambda \lambda 5052, 5380$ [C[i]{}]{} lines (measured in high-quality HARPS spectra) to derive C abundances. The error bars shown were estimated by adding, in quadrature, the errors arising from the EWs and those corresponding to the internal uncertainties of  and . At low metallicity and low , the [C[i]{}]{} lines are very weak, i.e. $EW \sim$ 2 - 4mÅ only, so the error of  due to the uncertainty of the EW measurements becomes large. Oxygen abundances ================= Oxygen abundances were determined from equivalent widths of the $\lambda 7774$ [O[i]{}]{} triplet lines. The NOT/FIES spectra do not cover this wavelength region and only a few of the UVES spectra have the [O[i]{}]{} triplet included. We have therefore adopted EW measurements from Ramírez et al. ([@ramirez12]), who derived oxygen abundances for a subset of stars from NS10 using high-resolution spectra obtained with the 2.7m telescope at the McDonald observatory, Keck/HIRES, and the MIKE spectrograph at the Magellan Telescope. For the thin-disk stars we used FEROS spectra obtained with the ESO 2.2m telescope. Altogether, [O[i]{}]{} triplet data are available for 101 stars out of the 117 stars for which [C[i]{}]{} abundances were determined. In deriving O abundances from the 7774Å triplet, non-LTE corrections from Fabbian et al. ([@fabbian09a]) were applied as described by Nissen ([@nissen13]). These corrections are important, even for differential determinations in the -range of our sample of stars; the correction of \[O/H\] ranges from about +0.1dex at 5300K to $-0.1$dex at 6300K. The derived  values are shown in Fig. \[Nissen:fig2\]. The distribution of stars looks much the same as that obtained by Ramírez et al. ([@ramirez12], Fig. 1, lower panel), except that they find a flatter distribution of \[O/Fe\] for the thick-disk and high-alpha halo stars. This is due to a somewhat stronger dependence of their non-LTE corrections on \[Fe/H\] than those of Fabbian et al. ([@fabbian09a]). ![\[O/Fe\] versus \[Fe/H\].[]{data-label="Nissen:fig2"}](Nissen_PE_fig2.eps){width="9cm"} Discussion ========== As seen from Figs. \[Nissen:fig1\] and \[Nissen:fig2\], the low-alpha halo stars are separated in  and  from thick-disk and high-alpha halo stars. For the latter two populations, the dispersion in  and at a given metallicity can be explained in terms of the errors of the derived abundances, whereas there seems to be an additional cosmic dispersion in  and  for the low-alpha halo population. In support of a real cosmic scatter, Fig. \[Nissen:fig3\] shows  versus  (from NS10) for the metallicity range $-1.2 < {\mbox{\rm [Fe/H]}}< -0.7$, where the largest dispersion is present. As seen, there is a striking correlation between  and  except for two strongly deviating stars. As suggested by the $\Lambda$CDM simulations of Zolotov et al. ([@zolotov10]), a possible explanation of the abundance differences seen in these figures is that the high-alpha stars were born in the innermost part of the Galaxy in a deep gravitational potential with such a high SFR that Type Ia SNe did not contribute significantly with iron until a metallicity of ${\mbox{\rm [Fe/H]}}\simeq -0.4$. Later these stars were dispersed to the halo by merging satellite galaxies. The low-alpha stars, on the other hand, were formed in dwarf galaxies with a relatively shallow potential, and hence low SFR, so that Type Ia SNe started contributing with iron at a metallicity around or below ${\mbox{\rm [Fe/H]}}= -1.6$. The reason for the scattter in , , , and  at a given metallicity could then be that the various satellite galaxies from which the low-alpha stars were accreted had different masses and therefore different star formation rates. ![\[C/Fe\] versus \[Na/Fe\] for stars with $-1.2 < {\mbox{\rm [Fe/H]}}< -0.7$.[]{data-label="Nissen:fig3"}](Nissen_PE_fig3.eps){width="9cm"} The two deviating low-alpha stars in Fig. \[Nissen:fig3\], G53-41 and G150-40, require a special explanation. They have high  but low  and also very low . Hence, they share the so-called Na-O anti-correlation in second generation stars in globular clusters, which are thought to be made of gas polluted by first-generation AGB stars in which the Ne-Na cycle has occured. This suggests that some of the halo stars originate from disrupted globular clusters, as also noted by Ramírez et al. ([@ramirez12]). Figures \[Nissen:fig1\] and \[Nissen:fig2\] also give a hint for a difference in  and  between thin- and thick-disk stars in the overlapping metallicity range $-0.7 < {\mbox{\rm [Fe/H]}}< -0.2$ although we have only two thin-disk stars in this range. A systematic difference is well established in the case of  (e.g. Bensby & Feltzing [@bensby06], Fig. 11b) and have also been found for the alpha-capture elements, Mg, Si, Ca, and Ti (e.g. Adibekyan [@adibekyan12], Fig. 8), but it was not seen in the case of  by Bensby & Feltzing ([@bensby06], Fig. 11a), Clearly, we need more thin-disk stars in the overlapping metallicity range to verify the possible difference in  between the thin- and thick-disk populations. Fig. \[Nissen:fig4\] shows  versus  for the four populations studied in this paper with thin- and thick-disk stars from Bensby & Feltzing ([@bensby06]) included. Their C and O abundances are derived from the weak forbidden $\lambda 8727$ \[[C[i]{}]{}\] and $\lambda 6300$ \[[O[i]{}]{}\] lines but agree very well with our abundances derived from high-excitation [C[i]{}]{} and [O[i]{}]{} lines. ![\[C/O\] versus \[O/H\]. Thin- and thick-disk stars from Bensby & Feltzing ([@bensby06]) are included without error bars.[]{data-label="Nissen:fig4"}](Nissen_PE_fig4.eps){width="9cm"} As seen from Fig. \[Nissen:fig4\], there is no systematic shift in  between high- and low-alpha halo stars. This is perhaps surprising, because the time-scale for the chemical enrichment of the low-alpha population is long enough to allow Type Ia SNe to contribute with iron (our explanation for the low $\alpha$/Fe ratios), and one could therefore have expected that low- and intermediate-mass AGB stars had enough time to contribute with carbon and raise  in the low-alpha stars to higher values than in high-alpha stars. The explanation may be that intermediate-mass (4 - 8 $M_{\rm Sun}$) AGB stars contribute very little to $^{12}$C (Kobayashi et al. [@kobayashi11]), and that the evolution time-scale of low-mass (1 - 3 $M_{\rm Sun}$) AGB stars (which do have a high $^{12}$C yield according to Kobayashi et al.) is longer than the chemical enrichment time-scale of the low-alpha population. We conclude that carbon in both high- and low-alpha halo stars was made primarely in high-mass stars ($> 10 \, M_{\rm Sun}$), but with a metallicity-dependent yield to account for the slight increase of  with increasing . ![Part of the CH-band in a high- and a low-alpha star with similar atmospheric parameters. All unmarked lines are due to CH.[]{data-label="Nissen:fig5"}](Nissen_PE_fig5.eps){width="12cm"} As also seen from Fig. \[Nissen:fig4\], thick-disk stars have on average the same  as high-alpha halo stars, but thin-disk stars have systematically higher in the overlapping metallicity range $-0.5 < {\mbox{\rm [O/H]}}< +0.1$. This offset in  between thin- and thick-disk stars was already detected by Bensby & Feltzing ([@bensby06], Fig. 12), but is even more clearly seen in our Fig. \[Nissen:fig4\]. Probably, the higher  ratios in the Galactic thin disk are due to low-mass AGB stars, i.e. the chemical enrichment time-scale of the thin disk is long enough to allow low-mass stars to contribute. As a final remark, we note that the $\lambda \lambda 5052, 5380$ [C[i]{}]{} lines used to derive C abundances in this paper are very weak in metal-poor stars. Thus, very high S/N, high-resolution spectra are required to derive precise C abundances. The somewhat stronger [C[i]{}]{} lines in the 9000 - 9500Å region may be used instead (Fabbian et al. [@fabbian09b]) but they are blended by telluric lines. As an interesting alternative one may use the CH band. Fig. \[Nissen:fig5\] shows part of this band in NOT/FIES spectra of a high-alpha star, G31-55 (, , , ) = (5738K, 4.33, $-$1.10, 0.29) and a low-alpha star HD193901 (5756K, 4.39, $-$1.09, 0.16). As seen, the three [Fe[i]{}]{} lines have similar strengths in the two stars, whereas all CH lines are significantly weaker in the low-alpha star. Hence, it may be possible to use the CH band in large forthcoming surveys such as HERMES/GALAH and ESO/4MOST for high precision differential studies of carbon abundances in stellar populations. Funding for the Stellar Astrophysics Centre is provided by the Danish National Research Foundation (Grant agreement no.: DNRF106). This paper is based on observations made with the Nordic Optical Telescope on La Palma, and on data from the European Southern Observatory ESO/ST-ECF Science Archive Facility. 2012, *A&A*, 545, A32 2006, *MNRAS*, 367, 1181 2010, *A&A*, 512, A54 2009a, *A&A*, 500, 1221 2009b, *A&A*, 500, 1143 2011, *MNRAS*, 416, 2802 2008, *A&A*, 486, 951 2011, *MNRAS*, 414, 3231 2012, *MNRAS*, 420, 2245 2013, *A&A*, 552, A73 2010, *A&A*, 511, L10 (NS10) 2011, *A&A*, 530, A15 2012, *ApJ*, 757, 164 2012, *A&A*, 538, A21 2005, *PASJ*, 57, 65 2010, *ApJ*, 721, 738 [^1]: Selected as stars having a total space velocity $V_{\rm total} > 180$ relative to the local standard of rest (LSR).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We theoretically investigate strong-coupling properties of a Bose-Fermi mixture. In the mass- and population-balanced case, two of the authors have shown that a strong hetero-pairing interaction in this mixture brings about coupling phenomena between Fermi atomic excitations and Bose atomic and composite molecular excitations, that appear as an anomalous multiple peak structure in the single-particle spectral weight (SW). \[D. Kharga, [*et. al.*]{}, J. Phys. Soc. Jpn. [**86**]{}, 084301 (2017)\]. In this paper, extending this previous work, we show that, although these many-body phenomena are sensitive to mass and population imbalances between the Bose and Fermi components, SW still exhibits the multiple peak structure in a moderately mass-imbalanced $^{87}$Rb-$^{40}$K and $^{23}$Na-$^{40}$K mixtures. We also point out that the photoemission spectrum is a useful quantity to observe this spectral anomaly. Since a real trapped Bose-Fermi mixture is usually accompanied by mass and (local) population imbalance, our results would contribute to the study of a strongly interacting Bose-Fermi mixture, under realistic imbalanced conditions.' author: - 'Koki Manabe$^1$, Daisuke Inotani$^2$, and Yoji Ohashi$^1$' title: 'Single-particle properties of a strongly-interacting Bose-Fermi mixture with mass and population imbalance' --- Introduction ============ The high tunability of ultracold atomic gases has contributed to the development of quantum many-body physics discussed in various research fields[@Bloch; @Giorgini; @Georgescu; @Gross]: Using an optical lattice technique, Greiner and co-workers have observed the superfluid-Mott insulator transition in a $^{87}$Rb Bose gas[@Greiner]. In $^{40}$K[@Regal] and $^6$Li Fermi gases[@Zwierlein; @Kinast; @Jochim], the superfluid phase transition and the BCS (Bardeen-Cooper-Schrieffer)-BEC (Bose-Einstein condensation) crossover phenomenon[@Eagles; @Leggett; @Engelbrecht; @BCS-BEC_rev1; @BCS-BEC_rev2] have been realized, by using a tunable pairing interaction associated with a Feshbach resonance[@Chin]. Besides Bose gas and Fermi gas, a gas mixture of Bose and Fermi atoms has also extensively been studied in cold atom physics[@dg_KLi1; @dg_LiLi1; @dg_LiLi2; @dg_LiNa1; @FR_KRb1; @FR_KRb2; @FR_KRb3; @FR_KRb4; @FR_LiCs1; @FR_LiNa1; @FR_LiNa2; @FR_LiRb1; @FR_NaK1; @FR_NaK2; @Mf_KRb1; @Mf_KRb2; @Mf_KRb3; @Mf_NaK1; @Mf_NaK2; @Mf_NaLi1; @dualSF1; @dualSF2; @dualSF3; @inducedint0; @cTMA1; @cTMA2; @cTMA3; @cTMA4; @cTMA5; @cTMA6; @iTMA1; @iTMA2; @others0; @others1; @others2; @Gamma2; @Gamma3; @Gamma4; @stability1; @stability2; @Bruun1; @Bruun2; @Bruun3; @inducedint1; @inducedint3; @inducedint4]. This Bose-Fermi mixture is similar to a $^4$He-$^3$He mixture, as well as quark matter in high-energy physics[@QCD1]. Using this similarity, as well as the advantage that one can tune the strength of a Bose-Fermi pairing interaction by using a hetero-nuclear Feshbach resonance, Ref. [@QCD1] suggests that this atomic mixture may be used as a quantum simulator for the study of dense QCD matter, where a bound di-quark (boson) and an unpaired quark (fermion) form a nucleon (composite fermion). In a mass- and population-balanced Bose-Fermi mixture, two of the authors have recently shown that a strong Bose-Fermi pairing interaction causes couplings between Fermi atomic excitations and Bose atomic excitations, as well as atomic excitations and molecular excitations[@iTMA1]. As an interesting phenomenon associated with these couplings, the Fermi component of the single-particle spectral weight (SW) has been shown to exhibit a three-peak structure, consisting of two sharp peaks along the free fermion dispersion and composite molecular dispersion, and a broad downward peak being related to Bose single-particle excitations. Here, we recall that SW in a free Fermi gas only has a single peak line along the free particle dispersion. In a two-component Fermi gas in the BCS-BEC crossover region, SW is known to exhibit a two-peak structure associated with the pseudogap phenomenon originating from strong-pairing fluctuations[@FFSW1; @FFSW2; @FFSW3; @Tsuchiya2009]. Thus, the three-peak structure in the Fermi SW is expected to be a characteristic of a strongly interacting Bose-Fermi mixture. To confirm this expectation, however, one should remember that any real Bose-Fermi mixture is composed of different kinds of atoms or different isotopes, such as $^{87}$Rb-$^{40}$K and $^{23}$Na-$^{40}$K gases, so that it is always accompanied by mass imbalance. In addition, when it is trapped in a harmonic potential, bosons and fermions have different density profiles so that local population imbalance is unavoidable, even when both the components have the same number of atoms. Although a box-type trap has recently been invented[@box_Fermi; @box_Bose1; @box_Bose2] (where a trapped gas is almost uniform), the conventional harmonic trap is still used in many experiments. At this stage, it is unclear to what extent these realistic situations affect the above-mentioned many-body coupling phenomena obtained in the somehow academic mass- and population-balanced case. We also note that highly population-imbalanced Bose-Fermi mixture has recently attracted much attention in the study of Bose polaron[@Bose-pol_exp1; @Bose-pol_exp2; @Bose-pol_theo1; @Bose-pol_theo2; @Bose-pol_theo3; @Bose-pol_theo4; @Bose-pol_theo5]. In this paper, we investigate single-particle properties of a Bose-Fermi mixture with a hetero-nuclear Feshbach resonance. Extending the previous work[@iTMA1] to include mass and population imbalances, we examine how these affect strong-coupling corrections to SW. We clarify whether or not the many-body coupling phenomena obtained in the mass- and population-balanced case survive in a mass-imbalanced $^{87}$Rb-$^{40}$K and $^{23}$Na-$^{40}$K mixtures. As an observable quantity related to SW, we also deal with the photoemission spectrum[@PES_exp1; @PES_exp2; @PES_exp3; @PES_rev; @Ota]. This paper is organized as follows. In Sec. II, we explain our formulation. We separately examine effects of population imbalance and mass imbalance in Secs. III, and IV, respectively. In Sec. V, we pick up a $^{87}$Rb-$^{40}$K mixture, as well as a $^{23}$Na-$^{40}$K mixture as typical two example of mass-imbalanced Bose-Fermi mixture. Throughout this paper, we set $\hbar=k_{\rm B}=1$, and the system volume $V$ is taken to be unity, for simplicity. Formulation =========== We consider a gas mixture of single-component Bose atoms and single-component Fermi atoms, with a hetero-nuclear Feshbach resonance. This Bose-Fermi mixture is modeled by the Hamiltonian, $$\begin{aligned} H=\sum_{{\bm p},{\rm s=B,F}} \xi^{\rm s}_{\bm p} c^\dagger_{{\bm p},{\rm s}} c_{{\bm p},{\rm s}} -U_{\rm BF} \sum_{\bm{p,p',q}} c^\dagger_{\bm{p}+\bm{q},{\rm B}} c^\dagger_{\bm{p'}-\bm{q},{\rm F}} c_{{\bm p}',{\rm F}} c_{{\bm p},{\rm B}}, \label{eq.1}\end{aligned}$$ where $c^\dagger_{{\bm p},{\rm s}}$ is the creation operator of a Bose (s=B) and a Fermi (s=F) atom. $\xi_{\bm p}^{\rm s}=\varepsilon_{\bm p}^{\rm s}-\mu_{\rm s}={\bm p}^2/(2m_{\rm s})-\mu_{\rm s}$ is the kinetic energy of the s-component, measured from the chemical potential $\mu_{\rm s}$ (where $m_{\rm s}$ is an atomic mass). $-U_{\rm BF}(<0)$ is a Bose-Fermi pairing interaction, which is assumed to be tunable by a hetero-nuclear Feshbach resonance. This contact-type interaction brings about the ultraviolet divergence, which is, as usual, absorbed into the $s$-wave scattering length $a_{\rm BF}$. It is related to the bare pairing interaction $-U_{\rm BF}$ as $${4\pi a_{\rm BF} \over m}= -{U_{\rm BF} \over 1-U_{\rm BF}\sum_{\bm p}^{p_{\rm c}}{m \over {\bm p^2}}}. \label{eq.2}$$ Here, $m=2m_{\rm B}m_{\rm F}/(m_{\rm B}+m_{\rm F})$ is twice the reduced mass, and $p_{\rm c}$ is a cutoff momentum. We measure the interaction strength in terms of the inverse scattering length $(k_{\rm tot}a_{\rm BF})^{-1}$. Here $k_{\rm tot}=(3\pi^2N_{\rm tot})^{1/3}$ is the Fermi momentum in an [*assumed*]{} two-component Fermi gas with the total number $N_{\rm tot}=N_{\rm F}+N_{\rm B}$ of fermions, where $N_{\rm s=B,F}$ is the particle number in the s-component. In this scale, the interaction strength increases with increasing $(k_{\rm tot}a_{\rm BF})^{-1}$ from the negative value. $(k_{\rm tot}a_{\rm BF})^{-1}=0$ represents the unitarity limit. This paper only considers a uniform gas, ignoring effects of a harmonic trap, for simplicity. However, the local population imbalance coming from the difference of the density profile between the Bose and the Fermi components in a trap is partially examined by considering the population-imbalanced case ($N_{\rm B}\ne N_{\rm F}$). Strong-coupling corrections to Bose and Fermi single-particle excitations are conveniently described by the self-energy $\Sigma_{\rm s=B,F}({\bm p},i\omega_{\rm s})$ in the single-particle thermal Green’s functions, $$\begin{aligned} G_{\rm s=B,F}(\bm{p},i\omega_{\rm s})=\frac{1}{i\omega_{\rm s}-\xi_{\bm{p}}^{\rm s}-\Sigma_{\rm s}(\bm{p},i\omega_{\rm s})}, \label{eq.3}\end{aligned}$$ where $i\omega_{\rm B}$ ($i\omega_{\rm F}$) is the boson (fermion) Matsubara frequency. ![(a) Self-energies $\Sigma_{\rm s=B,F}$ in iTMA. (b) Bose-Fermi scattering matrix $\Gamma_{\rm BF}({\bm q},i\omega_{\rm F})$ in Eq. (\[eq.10\]), physically describing hetero-pairing fluctuations. The solid line is the bare Fermi Green’s function $G_{\rm F}^0$ in Eq. (\[eq.9\]). The wavy line is the [*modified*]{} Bose Green’s function $\tilde{G}_{\rm B}^0$ in Eq.. The dashed line is the hetero-pairing interaction $-U_{\rm BF}$. []{data-label="fig1"}](Fig1.jpg){width="10cm"} We evaluate $\Sigma_{\rm s}({\bm p},i\omega_{\rm s})$ within the framework of the $T$-matrix-type approximation developed in Ref. [@iTMA1], which is diagrammatically described in Fig.\[fig1\]. This diagrammatic structure is formally the same as that in the ordinary (non-selfconsistent) $T$-matrix approximation (TMA); however, the crucial difference is that the bare Bose Green’s function, $$G_{\rm B}^0(\bm{p},i\omega_{\rm B})= {1 \over i\omega_{\rm B}-\xi_{\bm{p}}^{\rm B}}, \label{eq.4}$$ in the ordinary TMA self-energy diagrams is now replaced by the [*modified*]{} one[@iTMA1], $$\begin{aligned} \tilde{G}_{\rm B}^0(\bm{p},i\omega_{\rm B}) =\frac{1}{i\omega_{\rm B}-\tilde{\xi}_{\bm{p}}^{\rm B}}. \label{eq.5}\end{aligned}$$ Here, $\tilde{\xi}_{\bm{p}}^{\rm B}=\xi_{\bm{p}}^{\rm B}+\Sigma_{\rm B}(\bm{0},0)\equiv \varepsilon_{\bm{p}}^{\rm B}-\tilde{\mu}_{\rm B}$ (where ${\tilde \mu}_{\rm B}=\mu_{\rm B}-\Sigma_{\rm B}(0,0)$) involves the self-energy correction at $\bm{p}=i\omega_{\rm B}=0$. This is motivated by the Hugenholtz-Pines theorem[@HP], $$\begin{aligned} \mu_{\rm B}-\Sigma_{\rm B}(\bm{p}=\bm{0},i\omega_{\rm B}=0)=0, \label{eq.6}\end{aligned}$$ stating that the Bose excitations become gapless at the BEC phase transition. The modified Green’s function ${\tilde G}_{\rm B}^0({\bm p},i\omega_{\rm B})$ in Eq. (\[eq.5\]) is chosen so as to satisfy this required condition at $T_{\rm c}$. We briefly note that the bare Bose Green’s function $G_{\rm B}^0$ in Eq. (\[eq.4\]) still has a [*gapped*]{} Bose excitation spectrum at $T_{\rm c}$, which means that the self-energy in TMA underestimates effects of low-energy Bose excitations near $T_{\rm c}$. In this [*improved*]{} $T$-matrix approximation (iTMA), the summation of the diagrams in Fig. \[fig1\] gives $$\begin{aligned} \Sigma_{\rm B}(\bm{p},i\omega_{\rm B})&= T\sum_{\bm{q},i\omega_{\rm F}'}\Gamma_{\rm BF}(\bm{q},i\omega_{\rm F}')G_{\rm F}^0(\bm{q}-\bm{p},i\omega_{\rm F}'-i\omega_{\rm B}), \label{eq.7} \\ \Sigma_{\rm F}(\bm{p},i\omega_{\rm F})&=-T\sum_{\bm{q},i\omega_{\rm F}'}\Gamma_{\rm BF}(\bm{q},i\omega_{\rm F}')\tilde{G}_{\rm B}^0(\bm{q}-\bm{p},i\omega_{\rm F}'-i\omega_{\rm F}), \label{eq.8}\end{aligned}$$ where $$G_{\rm F}^0(\bm{p},i\omega_{\rm F}) ={1 \over i\omega_{\rm F}-\xi_{\bm{p}}^{\rm F}} \label{eq.9}$$ is the bare Fermi single-particle Green’s function, and $$\begin{aligned} \Gamma_{\rm BF}(\bm{q},i\omega_{\rm F}) &=& -{U_{\rm BF} \over 1-U_{\rm BF}\Pi_{\rm BF}({\bm q},i\omega_{\rm F})} \nonumber \\ &=& \frac{1}{\frac{m}{4\pi a_{\rm BF}}+\Bigl[\Pi_{\rm BF}({\bm q},i\omega_{\rm F})-\sum_{\bm{k}}^{p_{\rm c}}\frac{m}{k^2}\Bigr]} \label{eq.10}\end{aligned}$$ is the iTMA Bose-Fermi scattering matrix, physically describing hetero-pairing fluctuations. Here, $$\begin{aligned} \Pi_{\rm BF}(\bm{q},i\omega_{\rm F}) &=-T\sum_{\bm{p},i\omega_{\rm B}}G_{\rm F}^0(\bm{q}-\bm{p},i\omega_{\rm F}-i\omega_{\rm B})\tilde{G}_{\rm B}^0(\bm{p},i\omega_{\rm B}) \nonumber \\ &=\sum_{\bm{p}}\frac{1-f(\xi_{\bm{q}-\bm{p}}^{\rm F})+n_{\rm B}(\tilde{\xi}_{\bm{p}}^{\rm B})}{\xi_{\bm{q}-\bm{p}}^{\rm F}+\tilde{\xi}_{\bm{p}}^{\rm B}-i\omega_{\rm F}} \label{eq.11}\end{aligned}$$ is the hetero-pair correlation function, where $n_{\rm B}(x)$ and $f(x)$ are the Bose and Fermi distribution function, respectively. The BEC phase transition temperature $T_{\rm c}$ is conveniently determined from the Hugenholtz-Pines condition in Eq. (\[eq.6\]). We actually solve this equation, together with the equation for the number $N_{\rm B}$ ($N_{\rm F}$) of Bose (Fermi) atoms, $$\begin{aligned} N_{\rm B}&=-T\sum_{\bm{p},i\omega_{\rm B}}G_{\rm B}(\bm{p},i\omega_{\rm B}), \label{eq.12} \\ N_{\rm F}&= T\sum_{\bm{p},i\omega_{\rm F}}G_{\rm F}(\bm{p},i\omega_{\rm F}), \label{eq.13}\end{aligned}$$ to self-consistently determine $T_{\rm c}$, $\mu_{\rm B}(T_{\rm c})$, and $\mu_{\rm F}(T_{\rm c})$. Above $T_{\rm c}$, we only deal with the number equations (\[eq.12\]) and (\[eq.13\]), to evaluate $\mu_{\rm B}(T)$, and $\mu_{\rm F}(T)$. The single-particle spectral weights (SWs) $A_{s={\rm B,F}}(\bm{p},\omega)$ are related to the analytic-continued Green’s functions as, $$\begin{aligned} A_{\rm s=B,F}(\bm{p},\omega)&=-\frac{1}{\pi}{\rm Im} \left[G_{\rm s}(\bm{p},i\omega_{\rm s}\to \omega+i\delta\equiv\omega_+) \right], \label{eq.14}\end{aligned}$$ where $\delta$ is an infinitesimally small positive number. The photoemission spectra (PESs) $I_{\rm s=B,F}({\bm p},\omega)$[@PES_exp1; @PES_exp2; @PES_exp3] are then immediately obtained from SWs as[@PES_rev; @Ota], under the assumption that the final state interaction is absent, $$\begin{aligned} I_{\rm F}(\bm{p},\omega)=2\pi t_{\rm F}^2p^2A_{\rm F}(\bm{p},\omega)f(\omega), \label{eq.15} \\ I_{\rm B}(\bm{p},\omega)=2\pi t_{\rm B}^2p^2A_{\rm B}(\bm{p},\omega)n_{\rm B}(\omega). \label{eq.16}\end{aligned}$$ Here, $t_{\rm s}$ is a transfer-matrix element from the initial atomic hyperfine state $|I\rangle$ to the final one $|F\rangle$ ($\ne |I\rangle$). Between the two, the Fermi SW will be found to be more useful for the study of many-body coupling phenomena mentioned in Sec. I. Thus, we only examine the corresponding Fermi PES in this paper. ![Calculated Bose-Einstein condensation temperature $T_c$ in a Bose-Fermi mixture with population imbalance ($N_{\rm B}\ne N_{\rm F}$). We set $m_{\rm B}=m_{\rm F}$. (a) $N_{\rm B}<N_{\rm F}$. (b) $N_{\rm B}>N_{\rm F}$. The dashed line shows ${\bar T}_{\rm c}^0$ in Eq. (\[eq.3.2\]). $T_{\rm c}^0$ is the BEC phase transition temperature in an ideal gas of $N_{\rm B}$ bosons in Eq (\[eq.3.1\]). ‘QCP’ is the quantum critical point at which $T_{\rm c}$ vanishes. []{data-label="fig2"}](Fig2.jpg){width="8.5cm"} Single-particle excitations in a population-imbalanced Bose-Fermi mixture ========================================================================= Figure \[fig2\] shows the BEC transition temperature $T_{\it c}$ in a Bose-Fermi mixture with population imbalance ($N_{\rm B}\neq N_{\rm F}$). In the weak-coupling limit $(k_{\rm tot}a_{\rm BF})^{-1}\ll -1$, since Fermi atoms do not affect the BEC phase transition, $T_{\rm c}$ is simply given by the BEC phase transition temperature in an ideal Bose gas, $$T_{\rm c}^0={2\pi \over m_{\rm B}} \left( {N_{\rm B} \over \zeta(3/2)} \right)^{2/3}, \label{eq.3.1}$$ where $\zeta(3/2)\simeq 2.613$ is the zeta function. Starting from this extreme case, when $N_{\rm B}/N_{\rm F}<1$, we see in Fig. \[fig2\](a) that the overall interaction dependence of $T_{\rm c}$ is similar to that in the population-balanced case ($N_{\rm B}=N_{\rm F}$)[@iTMA1]: $T_{\rm c}$ gradually decreases from $T_{\rm c}^0$ with increasing the interaction strength, to eventually vanish around the unitarity limit $(k_{\rm tot}a_{\rm BF})^{-1}=0$. This vanishing $T_{\rm c}$ is because of the formation of Bose-Fermi molecules in the two-body level when $(k_{\rm tot}a_{\rm BF})^{-1}>0$, and the most Bose atoms pair up with Fermi atoms to become composite molecules there. As a result, BEC of unpaired [*Bose atoms*]{} no longer occurs, when the interaction strength exceeds a quantum critical point (QCP) at $(k_{\rm tot}a_{\rm BF})^{-1}\sim 0$, as seen in Fig. \[fig2\](a). When $N_{\rm B}>N_{\rm F}$ (Fig. \[fig2\](b)), on the other hand, the BEC phase transition remains to exist ($T_{\rm c}>0)$ even in the strong-coupling limit ($(k_{\rm tot}a_{\rm BF})^{-1}\gg +1$). This is simply because the number $\Delta N_{\rm B}\equiv N_{\rm B}-N_{\rm F}(>0)$ of bosons remain unpaired in this limit. Indeed, $T_{\rm c}$ in the strong-coupling limit is well described by their BEC transition temperature ${\bar T}_{\rm c}^0$, $${\bar T}_{\rm c}^0={2\pi \over m_{\rm B}} \left( {\Delta N_{\rm B} \over \zeta(3/2)} \right)^{2/3}, \label{eq.3.2}$$ as shown in Fig.\[fig2\](b). ![Calculated intensity of single-particle spectral weights $A_{{\rm s}={\rm B,F}}(\bm{p},\omega)$ in a [*population-imbalanced*]{} unitary Bose-Fermi mixture at $T=T_{\rm c}$. (a1)-(a5) $A_{\rm F}(\bm{p},\omega)$. (b1)-(b5) $A_{\rm B}(\bm{p},\omega)\times{\rm sgn}(\omega)$. We set $m_{\rm B}=m_{\rm F}$. The spectral intensity is normalized by $\varepsilon_{\rm tot}^{-1}\equiv 2m/k_{\rm tot}^2$. []{data-label="fig3"}](Fig3.jpg){width="8.5cm"} Figures \[fig3\](a1)-(a5) show the Fermi SW $A_{\rm F}(\bm{p},\omega)$ in the unitarity limit at $T_{\rm c}$. In the population-balanced case (panel (a3)), the Fermi SW exhibits a three-peak structure as a result of the Fermi-Bose and atom-molecule couplings[@iTMA1] mentioned in Sec. I: (A) Sharp peak line along the Fermi free-particle dispersion, $\omega=\xi_{\bm{p}}^{\rm F}$. (B) Broad downward peak in the negative energy region around $\omega=-\tilde{\xi}_{\bm p}^{\rm B}$, where the Bose dispersion ${\tilde \xi}_{\bm p}^{\rm B}$ is given below Eq. (\[eq.5\]). (C) Sharp upward peak line along the composite molecular dispersion, $\omega=\xi_{\bm p}^{\rm CF}\equiv {\bm p}^2/(2M_{\rm CF})-\mu_{\rm CF}$, where $M_{\rm CF}\simeq m_{\rm B}+m_{\rm F}$ is a molecular mass and $\mu_{\rm CF}$ is the molecular chemical potential[@comment4]. When the ratio $N_{\rm B}/N_{\rm F}$ decreases from unity (Fig.s \[fig3\](a3)$\to$(a1)), the two peaks (B) and (C) are found to gradually disappear. This is simply because the system approaches a single-component free Fermi gas. In the extreme population-imbalanced case shown in Fig. \[fig3\](a1), the peak line (A) is only seen, as expected. In the opposite case, on the other hand, with increasing the ratio $N_{\rm B}/N_{\rm F}>1$, while the broad peak (B) gradually disappears, the molecular peak (C) continues to exist, in addition to the free fermion dispersion (A), even in Fig. \[fig3\](a5). To understand these population-imbalance effects on the Fermi SW, we conveniently approximate the iTMA Fermi Green’s function $G_{\rm F}({\bm p},i\omega_{\rm F})$ to[@iTMA1] $$G_{\rm F}(\bm{p},i\omega_{\rm F}\rightarrow\omega_+)\simeq {1 \over \displaystyle \omega_+-\xi_{\bm{p}}^{\rm F}-\frac{ZN_{\rm B}^0}{\omega_+-\xi_{\bm{p}}^{\rm CF}}-\left\langle\frac{ZN_{\rm CF}}{\omega_++\tilde{\xi}_{{\bm{k}}_{\rm CF}-\bm{p}}^{\rm B}}\right\rangle_{{\bm k}_{\rm CF}} }. \label{eq.3.3}$$ (For the outline of the derivation, see the Appendix.) In Eq. (\[eq.3.3\]), $N_{\rm CF}=\sum_{\bm{q}}f(\xi_{\bm{q}}^{\rm CF})$ is the number of composite Fermi molecules with the chemical potential $\mu_{\rm CF}>0$, $N_{\rm B}^0=\sum_{\bm{q}}n(\tilde{\xi}_{\bm{q}}^{\rm B})$, and $|{\bm k}_{\rm CF}|=\sqrt{2M_{\rm CF}\mu_{\rm CF}}$ gives the size of the Fermi surface in the composite Fermi molecular gas. The average $\langle\cdots\rangle_{\bm Q}$ is taken over the direction of ${\bm Q}$, and $Z$ is a positive constant. (For details of $Z$, we refer to Ref. [@iTMA1].) Equation (\[eq.3.3\]) explains that hetero-pairing fluctuations couple the Fermi atomic excitations $\omega=\xi_{\rm p}^{\rm F}$ (A) with the molecular excitations $\omega=\xi_{\rm p}^{\rm CF}$ (C) with the coupling strength $ZN_{\rm B}^0$, as well as with the Bose excitations $\omega=-{\tilde \xi}^{\rm B}_{{\bm k}_{\rm CF}-{\bm p}}$ (B) with the coupling strength $ZN_{\rm CF}$, respectively. For the latter Fermi-Bose coupling, because of the average over the direction of ${\bm k}_{\rm CF}$ in Eq. (\[eq.3.3\]), the Bose excitations gives the broad spectrum structure in the negative energy region of Fig. \[fig3\](a3). As the number $N_{\rm B}$ of Bose atoms decreases ($N_{\rm B}/N_{\rm F}<1$), both $N_{\rm B}^0$ and $N_{\rm CF}$ in the denominator in Eq. (\[eq.3.3\]) decrease, to eventually vanish in the limit $N_{\rm B}\to 0$. This immediately explains the single-peak structure in the Fermi SW in Fig. \[fig3\](a1). On the other hand, when $N_{\rm F}$ decreases ($N_{\rm B}/N_{\rm F}>1$), while $N_{\rm CF}$ vanishes in the large population-imbalance limit, $N_{\rm B}^0$ would approach the non-zero value $N_{\rm B}$. Thus, Eq. (\[eq.3.3\]) is reduced to $$G_{\rm F}(\bm{p},\omega_+)= {1 \over \displaystyle \omega_+-\xi_{\bm{p}}^{\rm F}-\frac{ZN_{\rm B}^0}{\omega_+-\xi_{\bm{p}}^{\rm CF}}}, \label{eq.3.4}$$ which has the two poles, $$E_{\bm p}^\pm= {1 \over 2} \left[ [\xi_{\bm p}^{\rm F}+\xi_{\bm p}^{\rm CF}] \pm \sqrt{[\xi_{\bm p}^{\rm F}-\xi_{\bm p}^{\rm CF}]^2+4ZN_{\rm B}^0} \right]. \label{eq.3.5}$$ Equation (\[eq.3.5\]) explains the two-peak structure in Fig.\[fig3\](a5). That is, the Fermi single-particle excitations in the highly population-imbalanced regime ($N_{\rm B}/N_{\rm F}\gg 1$) are dominated by the atom-molecule coupling phenomenon. We briefly note that this limit just corresponds to the Bose-polaron system. Applying the same approximation to the Bose component, we obtain[@iTMA1], $$G_{\rm B}(\bm{p},i\omega_{\rm F}\to\omega_+)\simeq {1 \over \displaystyle \omega_+-\xi_{\bm p}^{\rm B} - \left\langle {ZN_{\rm F}^0 \over \omega_+-\xi_{{\tilde {\bm k}}_{\rm F}-\bm{p}}^{\rm CF} } \right\rangle_{{\tilde {\bm k}}_{\rm F}} - \left\langle { ZN_{\rm CF} \over \omega_++{\xi_{{\bm k}_{\rm CF}-\bm{p}}^{\rm F}} } \right\rangle_{{\tilde {\bm k}}_{\rm CF}} }. \label{eq.3.6}$$ Here, $N_{\rm F}^0=\sum_{\bm{q}}f(\xi_{\bm{q}}^{\rm F})$, and $|\tilde{\bm{k}}_{\rm F}|=\sqrt{2m_{\rm F}\mu_{\rm F}}$. Equation (\[eq.3.6\]) shows that the Bose single-particle excitations ($\omega=\xi_{\bm p}^{\rm B}$) couple with composite Fermi molecular excitations ($\omega=\xi_{{\tilde k}_{\rm F}-{\bm p}}^{\rm CF}$), as well as Fermi hole excitations ($\omega=-\xi^{\rm F}_{{\bm k}_{\rm CF}-{\bm p}}$); however, because of the angular averages in the denominator of Eq. (\[eq.3.6\]), the three-peak structure is not clearly seen in the Bose SW $A_{\rm B}({\bm p},\omega)$, when $N_{\rm B}=N_{\rm F}$ (see Fig. \[fig3\](b3)). When the number $N_{\rm F}$ of Fermi atoms decreases ($N_{\rm B}/N_{\rm F}>1$), both $N_{\rm F}^0$ and $N_{\rm CF}$ decrease. Thus, the Bose SW is gradually reduced to the spectral weight in a free Bose gas (where the single peak line is along $\omega=\xi_{\bm p}^{\rm B}$), as seen in Figs. \[fig3\](b3)$\to$(b5). With increasing $N_{\rm F}$ ($N_{\rm B}/N_{\rm F}<1$), the system eventually reaches the situation that $N_{\rm F}^0\to N_{\rm F}$ and $N_{\rm CF}\to 0$. Because of this, the spectral structure in Figs. \[fig3\](b1) and (b2) are dominated by Bose atomic excitations and broad composite molecular excitations, but the downward broad peak associated with Fermi hole excitations becomes weak. The above discussions indicate that the Fermi SW $A_{\rm F}({\bm p},\omega)$ is more suitable than the Bose SW $A_{\rm B}({\bm p},\omega)$, for the study of the Fermi-Bose and atom-molecule coupling phenomena. ![(a1)-(a4) Calculated intensity of the Fermi SW $A_{\rm F}(\bm{p},\omega)$ in the highly population imbalanced Bose-Fermi mixture ($N_{\rm B}/N_{\rm F}=100\gg 1$) at $T_{\rm c}$. The intensity is normalized by $\varepsilon_{\rm tot}^{-1}$. (b1)-(b4) Corresponding spectrum of the Bose-Fermi scattering matrix $-{\rm Im}[\Gamma_{\rm BF}(q,i\omega_{\rm F}\to\omega_+)]$, normalized by $2\pi^2/(mk_{\rm tot})$. The inset in panel (b4) shows the spectrum in the wide energy region $-4\le (\omega+\mu_{\rm F})/\varepsilon_{\rm tot}\le 2$, to clearly show the sharp peak line associated with molecular excitations. In each left panel, the dashed line shows the pole position of $\Gamma_{\rm BF}({\bm p},i\omega_{\rm F}\to\omega_+)$. []{data-label="fig4"}](Fig4.jpg){width="10cm"} The atom-molecule coupling phenomenon seen in Figs. \[fig3\](a4) and (a5) also appears away from the unitarity limit, as shown in the left panels in Fig. \[fig4\]. In this figure, when $(k_{\rm tot}a_{\rm BF})^{-1}>0$, the appearance of the lower sharp peak associated with molecular excitations would be reasonable, because the pairing interaction is strong enough to produce two-body bound states there. On the other hand, Fig. \[fig4\](a1) shows that the Fermi SW still has a sharp molecular peak line, in spite of the absence of a two-body bound state when $(k_{\rm tot}a_{\rm BF})^{-1}=-0.5<0$. Regarding this, plotting the spectrum ${\rm Im}[\Gamma_{\rm BF}({\bm p},i\omega_{\rm F}\to\omega_+)]$ of the Bose-Fermi scattering matrix, one finds an isolated sharp peak line (which physically describes composite Fermi molecular excitations) below the continuum spectrum, not only in the strong-coupling regime (Figs. \[fig4\](b2)-(b4)), but also in the weak-coupling region (Fig. \[fig4\](b1)). The latter result implies the stabilization of a Bose-Fermi bound state by a many-body (medium) effect. To see this in a simple manner, we approximate Eq. (\[eq.10\]) at $T_{\rm c}$ to, after taking the analytic continuation $i\omega_{\rm F}\to\omega_+$, $$\Gamma_{\rm BF}(\bm{q},\omega_+)\simeq {1 \over \displaystyle \frac{m}{4\pi a_{\rm BF}} + \sum_{\bm{p}}\left[\frac{1}{\varepsilon^{\rm F}_{\bm{q}-\bm{p}}+\varepsilon^{\rm B}_{\bm{p}}-{\tilde \omega}_+}-\frac{m}{p^2} \right] -N_{\rm B}^0G_{\rm F}^0(\bm{q},\omega_+)}, \label{eq.3.7}$$ where ${\tilde \omega}=\omega+\mu_{\rm F}$, and we have used the fact that the Bose distribution function $n_{\rm B}({\tilde \xi}^{\rm B}_{\bm p})$ in Eq. (\[eq.11\]) diverges at ${\bm p}=0$ at $T_{\rm c}$. When one ignores the last term in the denominator in Eq. (\[eq.3.7\]), the pole equation of Eq. (\[eq.3.7\]) is essentially the same as the two-body bound-state equation (which has a solution only when $a_{\rm BF}>0$). The term $N_{\rm B}^0G_{\rm F}({\bm q},\omega_+)$ in Eq. (\[eq.3.7\]) physically describes medium effects. Including this at ${\bm q}=0$, we obtain the pole equation, $$1=a_{\rm BF} \left[ \sqrt{m|{\tilde \omega}|}-{4\pi N_{\rm B}^0 \over m|{\tilde \omega}|} \right]. \label{eq.3.8}$$ This indicates that a Bose-Fermi bound state is also possible, when $(k_{\rm tot}a_{\rm BF})^{-1}\le 0$ (where the two-body bound state is absent) in the many-particle case[@Gamma2; @Gamma3; @Gamma4]. At the unitarity ($a_{\rm BF}^{-1}=0$), for example, Eq. (\[eq.3.8\]) gives $${\tilde \omega}=-{(4\pi N_{\rm B}^0)^{2 \over 3} \over m}. \label{eq.3.9}$$ In Fig. \[fig4\], the lower peak line in each left panel is found to be close to the bound-state dispersion (dashed line) in the right panel, although the former is somehow pushed down, due to the coupling with the Fermi atomic excitations $\omega=\xi_{\bm p}^{\rm F}$ (see Eq. (\[eq.3.5\])). In this figure, with increasing the strength of a hetero-pairing interaction (panel (a1)$\to$(a4)), the character of a Bose-Fermi molecule continuously changes from the many-body bound state assisted by medium, to the two-body bound state. We briefly note that a similar crossover phenomenon from a polaron state to the two-body bound state (polaron-molecule crossover) has been discussed in the Bose-polaron system at $T=0$[@Bose-pol_theo2]. The above discussions are also applicable to the population-balanced case ($N_{\rm B}=N_{\rm F}$): In Fig. \[fig5\], the spectrum $-{\rm Im}[\Gamma_{\rm BF}({\bm p},i\omega_{\rm F}\to\omega_+)]$ of the Bose-Fermi scattering matrix (right panels) has an isolated sharp peak, bringing out the lower peak in the Fermi SW $A_{\rm F}({\bm p},\omega)$ (left panels). In the right panels, the peak energy is lowered as the interaction strength increases, because of the increase of the binding energy of a Bose-Fermi bound state. This tendency is the same as the highly population-imbalanced case shown in Fig. \[fig4\]. ![Same plot as Fig.\[fig4\] for the population-balanced case $N_{\rm B}=N_{\rm F}$. The inset in panel (a3) shows the logarithmic plot to show the three-peak structure discussed in Fig. \[fig3\]. \[fig5\] ](Fig5.jpg){width="10cm"} ![Intensity of the Fermi SW $A_{\rm F}(\bm{p},\omega)$, as well as the spectrum $-{\rm Im}[\Gamma_{\rm BF}(q,i\omega_{\rm F}\rightarrow\omega_+)]$ of the Bose-Fermi scattering matrix, in a highly population-imbalanced Bose-Fermi mixture ($N_{\rm B}/N_{\rm F}=100$). The normalization of the spectral intensity is the same as that in Fig. \[fig4\]. (a1)-(a3) and (b1)-(b3): Weak-coupling case $(k_{\rm tot}a_{\rm BF})^{-1}=-0.5$. (c1)-(c3) and (d1)-(d3): Strong-coupling case $(k_{\rm tot}a_{\rm BF})^{-1}=0.5$. []{data-label="fig6"}](Fig6.jpg){width="16cm"} In the normal state above $T_{\rm c}$, we expect the following two thermal effects: (1) The Bose distribution function $n_{\rm B}({\tilde \xi}^{\rm B}_{\bm q})$ no longer diverges at ${\bm q}=0$ when $T>T_{\rm c}$, so that the approximation giving the first term in the last line in Eq. (\[eq.A2\]) becomes worse. Roughly speaking, this would lead to the broadening of the peak line coming from molecular excitations in the Fermi SW $A_{\rm F}({\bm p},\omega)$. In addition, because the factor $ZN_{\rm B}^0=Z\sum_{\bm q}n_{\rm B}({\tilde \xi}^{\rm B}_{\bm q})$ decreases with increasing the temperature, the atom-molecule coupling also becomes weak. (2) When Bose-Fermi bound states thermally dissociate into unpaired atoms at high temperatures, the approximate expression for the Bose-Fermi scattering matrix $\Gamma_{\rm BF}({\bm q},i\omega_{\rm F})$ in Eq. (\[eq.A1\]) is no longer valid. Keeping these two thermal effects in mind, we find in Figs. \[fig6\](a1)-(a3) ($(k_{\rm tot}a_{\rm BF})^{-1}=-0.5<0$) that the molecular peak line soon becomes obscure with increasing the temperature above $T_{\rm c}$. In this case, because of the weak Bose-Fermi pairing interaction, the sharp spectral peak in $-{\rm Im}[\Gamma_{\rm BF}(q,i\omega_{\rm F}\rightarrow\omega_+)]$ describing molecular excitations also soon disappears above $T_{\rm c}$ (see Figs. \[fig6\](b1)-(b3)). Thus, the above-mentioned two thermal effects are considered to suppress the atom-molecular coupling in the Fermi SW $A_{\rm F}({\bm p},\omega)$ in the weak-coupling case. When $(k_{\rm tot}a_{\rm BF})^{-1}=0.5>0$ in the strong-coupling regime, Figs. \[fig6\](d1)-(d3) indicate that the molecular spectrum still remains even at $T/T_{\rm c}^0=2.5$, because of a large binding energy. In this case, thermal effects on the atom-molecule coupling are dominated by thermal effect (1) in the above discussion. Indeed, in Figs. \[fig6\](c1)-(c3), while the molecular peak line gradually becomes broad with increasing the temperature, the existence of this coupling phenomenon itself can still be confirmed in $A_{\rm F}({\bm p},\omega)$ even at $T/T_{\rm c}^0=2.5$ (panel (c3)). Figure \[fig6\] indicates that, when we use the Fermi SW to examine the crossover from the medium-assisted (many-body) bound state in the weak-coupling regime to the two-body bound state in the strong-coupling regime in a highly population-imbalanced Bose-Fermi mixture, we need to set the temperature near $T_{\rm c}$, in order to observe the former bound state. ![Calculated BEC phase transition temperature $T_{\rm c}$ in a Bose-Fermi mixture with mass imbalance. The parameter $R_{\rm m}$ is defined in Eq. (\[eq.4.1\]): $R_{\rm m}>0$ ($R_{\rm m}<0$) corresponds to the case of $m_{\rm F}>m_{\rm B}$ ($m_{\rm F}<m_{\rm B}$). The dashed line is the quantum critical point (QCP) at which $T_{\rm c}$ vanishes. $T_{\rm c}^0$ is the BEC phase transition temperature in an ideal Bose gas given in Eq. (\[eq.3.1\]). []{data-label="fig7"}](Fig7.jpg){width="10cm"} ![Critical interaction strength $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}$, which is defined as the interaction strength at which $T_{\rm c}$ vanishes, in a mass-imbalanced Bose-Fermi mixture. The dashed line shows the approximate result $(k_{\rm tot}a_{\rm BF}^{\rm c,app})^{-1}$ in Eq. (\[eq.4.3\]). []{data-label="fig8"}](Fig8.jpg){width="10cm"} Single-particle properties of mass-imbalanced Bose-Fermi mixture ================================================================ Figure \[fig7\] shows the BEC phase transition temperature $T_{\rm c}$ and effects of mass imbalance in a Bose-Fermi mixture, where mass difference is parametrized by $$R_{\rm m}={m_{\rm F}-m_{\rm B} \over m_{\rm F}+m_{\rm B}}. \label{eq.4.1}$$ In this figure, $T_{\rm c}$ gradually decreases from the ideal Bose-gas result $T_{\rm c}^0$ in Eq. (\[eq.3.1\]), with increasing the Bose-Fermi interaction strength, to eventually vanish at a certain interaction strength (QCP). Although this phenomenon has already been known in the mass balanced case[@cTMA1; @iTMA1], Fig. \[fig7\] indicates that this suppression effect becomes more (less) remarkable when $m_{\rm F}<m_{\rm B}$ ($m_{\rm F}>m_{\rm B}$). To clearly show this, we separately plot in Fig. \[fig8\] the critical interaction strength ($\equiv (k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}$) at which $T_{\rm c}$ vanishes. The overall behavior of $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}$ shown in Fig. \[fig8\] may be understood as a result of competition between (1) the Bose-Einstein condensation of unpaired bosons, and (2) the formation of Bose-Fermi hetero-pairs: When $(k_{\rm tot}a_{\rm BF})^{-1}>0$, the two-body Bose-Fermi bound state can be formed, with the binding energy, $$E_{\rm bind}^{\rm 2b}={1 \over ma_{\rm BF}^2}, \label{eq.4.2}$$ where $m$ is given below Eq. (\[eq.2\]). In the strong-coupling regime where most bosons form Bose-Fermi molecules, when the bare BEC transition temperature $T_{\rm c}^0$ is much lower than the characteristic temperature $T_{\rm bind}^{\rm 2b}\sim E_{\rm bind}^{\rm 2b}$, the BEC instability would no longer occur, because of the absence of unpaired bosons. Thus, the QCP is roughly estimated from the condition $E_{\rm bind}^{\rm 2b}\sim T_{\rm c}^0$, which approximately gives the critical interaction strength, $$\begin{aligned} (k_{\rm tot}a_{\rm BF}^{{\rm c,app}})^{-1} &= \left( {4 \over 3\sqrt{\pi}\zeta(3/2)} \right)^{1/3} \frac{1}{\sqrt{1+m_{\rm B}/m_{\rm F}}} \nonumber \\ &= 0.66\frac{1}{\sqrt{1+m_{\rm B}/m_{\rm F}}}. \label{eq.4.3}\end{aligned}$$ Figure \[fig8\] shows that Eq. (\[eq.4.3\]) is consistent with the overall $R_{\rm m}$-dependence of $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}$. Because the above simple discussion assumes the two-body bound state, $(k_{\rm tot}a_{\rm BF}^{\rm c,app})^{-1}$ in Eq. (\[eq.4.3\]) must be positive. However, a Bose-Fermi bound state actually exists also in the weak-coupling region $(k_{\rm tot}a_{\rm BF})^{-1}\le 0$, as a result of the medium effects. For example, we show in Fig. \[fig9\] the spectrum $-{\rm Im}[\Gamma_{\rm BF}({\bm p},i\omega_{\rm F}\to\omega_+)]$ of Bose-Fermi scattering matrix in the highly mass-imbalanced case ($m_{\rm B}/m_{\rm F}=20\gg 1$) at the critical interaction strength $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}=-0.25<0$ ($T=0$). In this figure, an isolated molecular branch is seen below the continuum, as in the population-imbalanced case (see Figs. \[fig4\]-\[fig6\]). This many-body bound state in the weak-coupling region ($(k_{\rm tot}a_{\rm BF})^{-1}\le 0$) naturally explains why the critical coupling $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}$ can be negative in Fig. \[fig8\]. ![Spectral intensity $-{\rm Im}[\Gamma_{\rm BF}(q,i\omega_{\rm F}\rightarrow\omega_+)]$ of the Bose-Fermi scattering matrix in a highly mass-imbalanced Bose-Fermi mixture ($m_{\rm B}/m_{\rm F}=20$, or $R_{\rm m}\simeq -0.9$). We take the interaction strength $(k_{\rm tot}a_{\rm BF})^{-1}$ to be equal to the critical value $(k_{\rm tot}a_{\rm BF}^c)^{-1}=-0.25$ at which $T_{\rm c}$ vanishes.[]{data-label="fig9"}](Fig9.jpg){width="10cm"} ![Single-particle spectral weight in a unitary Bose-Fermi mixture with mass imbalance. We take $T=T_{\rm c}$, and $N_{\rm B}=N_{\rm F}$. The left and right panels show $A_{\rm F}(\bm{p},\omega)$ and $A_{\rm B}(\bm{p},\omega)$, respectively. The second panels from the top (bottom) show the case of a $^{87}$Rb-$^{40}$K ($^{23}$Na-$^{40}$K) mixture. For eye-guide, we plot the peak positions in the left figures (dashed lines). []{data-label="fig10"}](Fig10.jpg){width="10cm"} Figure \[fig10\] shows SWs $A_{\rm s=B,F}(\bm{p},\omega)$ in a mass-imbalanced unitary Bose-Fermi mixture at $T_{\rm c}$. In the Fermi SW (left panels), starting from the mass-balanced case (panel (a3)), we find that effects of mass difference are different between the cases of $m_{\rm B}/m_{\rm F}>1$ and $m_{\rm B}/m_{\rm F}<1$: Among the two sharp peaks ((A) and (C)) and a broad peak (B) in panel (a3), the peak (C), coming from the atom-molecule coupling, gradually disappears with increasing the ratio $m_{\rm B}/m_{\rm F}>1$ (see Fig. \[fig10\](a3)$\to$(a1)). When $m_{\rm B}/m_{\rm F}$ decreases from unity, on the other hand, Figs. \[fig10\](a3)$\to$(a5) show that the broad peak (B), originating from the Fermi-Bose coupling, gradually disappears. To understand these in a simple manner, we conveniently plot in Fig. \[fig11\] $N_{\rm B}^0=\sum_{\bm p}n_{\rm B}({\tilde \xi}_{\bm p}^{\rm B})$ and $N_{\rm CF}^0\equiv N_{\rm B}-N_{\rm B}^0~(\sim N_{\rm CF}=\sum_{\bm p}f(\xi_{\bm p}^{\rm CF}))$ in a unitary Bose-Fermi mixture at $T_{\rm c}$. Noting that these quantities are directly related to the atom-molecule coupling and Fermi-Bose coupling, respectively (see Eq. (\[eq.3.3\])), we find from Fig. \[fig11\] that the former (latter) coupling phenomenon becomes dominant when $R_{\rm m}\to 1$ ($R_{\rm m}\to -1$). In addition, the momentum dependence of the Bose kinetic energy ${\tilde \xi}_{\bm p}^{\rm B}$ becomes weak with increasing $m_{\rm B}$, so that the broadening by the angular integration in the last term the denominator in Eq. (\[eq.3.3\]) is suppressed. Because of this, the broad peak (B) in the mass balanced case in Fig. \[fig10\](a3) gradually becomes sharp, as one moves from panels (a3) to (a1). ![The number $N_{\rm B}^0=\sum_{\bm p}({\tilde \xi}_{\bm p}^{\rm B})$ of unpaired Bose atoms, as well as the number $N_{\rm CF}^0\equiv N_{\rm B}-N_{\rm B}^0$ of the Bose-Fermi molecules in a unitary Bose-Fermi mixture at $T_{\rm c}$, as functions of mass imbalance parameter $R_{\rm m}$ in Eq. (\[eq.4.1\]). Since the direct evaluation of $N_{\rm CF}=\sum_{\bm p}f(\xi_{\bm p}^{\rm CF})$ in Eq. (\[eq.3.3\]) is difficult, we approximately use $N_{\rm CF}^0$ for $N_{\rm CF}$ in our discussions. []{data-label="fig11"}](Fig11.jpg){width="10cm"} Compared to the Fermi SW, the Bose SW $A_{\rm B}({\bm p},\omega)$ is not so sensitive to the mass imbalance, as shown in the right panels in Fig. \[fig10\], which is simply due to the two angular integrations in the denominator in Eq. (\[eq.3.6\]). Thus, as in the population-imbalanced case, the Fermi SW $A_{\rm F}({\bm p},\omega)$ is more suitable for the study of strong-coupling corrections to single-particle excitations in a mass-imbalanced Bose-Fermi mixture. ![Calculated Fermi SW $A_{\rm F}(\bm{p},\omega)$ in a $^{87}$Rb-$^{40}$K mixture (left panels), as well as a $^{23}$Na-$^{40}$K mixture (right panels). We set $T=T_{\rm c}$ for the upper three cases, and set $T=0.01T_{\rm c}^0$ for the lowest case. We note that QCP is at $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}=0.43$ for $^{87}$Rb-$^{40}$K mixture, and $(k_{\rm tot}a_{\rm BF}^{\rm c})^{-1}=0.7$ for $^{23}$Na-$^{40}$K. Thus, there is no BEC phase transition when $(k_{\rm tot}a_{\rm BF})^{-1}=1$ in panels (a4) and (b4). []{data-label="fig12"}](Fig12.jpg){width="10cm"} ![Calculated Fermi photoemission spectrum (PES) $I_{\rm F}(p,\omega)$. Left panels: $^{87}$Rb-$^{40}$K mixture. Right panels: $^{23}$Na-$^{40}$K mixture. The parameters are same as those in Fig. \[fig12\]. The spectral intensity is normalized by $4\pi t_{\rm F}^2m$. This normalization is also used in Figs. \[fig14\] and \[fig15\]. []{data-label="fig13"}](Fig13.jpg){width="10cm"} ![Fermi PES $I_{\rm F}({\bm p},\omega)$ in the unitarity limit above $T_{\rm c}$. Left panels: $^{87}$Rb-$^{40}$K mixture ($T_{\rm c}=0.695T_{\rm c}^0$). Right panels: $^{23}$Na-$^{40}$K mixture ($T_{\rm c}=0.828T_{\rm c}^0$). The results at $T_{\rm c}$ are shown in Figs. \[fig13\](a2) and (b2). []{data-label="fig14"}](Fig14.jpg){width="10cm"} ![Effects of population imbalance on the Fermi PES $I_{\rm F}(p,\omega)$ in a unitary Bose-Fermi mixture at $T=1.5T_{\rm c}$. Left panels: $^{87}$Rb-$^{40}$K mixture. Right panels: $^{23}$Na-$^{40}$K mixture. For the population-balanced results, see Figs. \[fig13\](a2) and (b2). []{data-label="fig15"}](Fig15.jpg){width="10cm"} Examples: $^{87}$Rb-$^{40}$K and $^{23}$Na-$^{40}$K mixtures ============================================================ Figures \[fig10\](a2) and (a4) show the cases of existing $^{87}$Rb-$^{40}$K and $^{23}$Na-$^{40}$K mixtures, respectively. These figures predict that the Fermi SW still exhibits the three-peak structure in the $^{87}$Rb-$^{40}$K case. Although the broad peak (B) is suppressed in a $^{23}$Na-$^{40}$K mixture, we can still see the atom-molecule coupling in this mixture. We emphasize that these many-body coupling phenomena still remain to some extent, away from the unitarity limit, as shown in Fig. \[fig12\]. In both the mixtures, the atom-molecule coupling is found to remain with [*decreasing*]{} the interaction strength (see the upper two panels in Fig. \[fig12\]). This is simply due to the increase of $N_{\rm B}^0$ and the decrease of $N_{\rm CF}$ in the denominator in Eq. (\[eq.3.3\]). On the other hand, as one [*increases*]{} the interaction strength (see the lowest two panels in Fig. \[fig12\]), the Fermi-Bose coupling becomes important, reflecting the decrease of $N_{\rm B}^0$ and the increase of $N_{\rm CF}$. In Figs. \[fig12\](a4) and (b4), the Fermi-Bose coupling is only seen. Figure \[fig13\] shows the Fermi photoemission spectrum (PES) $I_{\rm F}({\bm p},\omega)$ in a Bose-Fermi mixture, where the parameters in each panel are the same as those in the corresponding panel in Fig. \[fig12\]. Comparing Fig. \[fig12\] with Fig. \[fig13\], one finds that, although the spectral structure seen in the positive energy region of SW is suppressed by the Fermi distribution function in PES (see Eq. (\[eq.15\])), it can still detect the downward broad spectral structure coming from the Fermi-Bose coupling associated with hetero-pairing fluctuations. For the upward peak line along the molecular dispersion seen in the Fermi SW, since the thermal broadening of the Fermi distribution function $f(\omega)$ around $\omega=0$ weakens the suppression of the spectral intensity in the positive energy region in the Fermi PES, it gradually appears in PES with increasing the temperature, as shown in Fig. \[fig14\]. Of course, this method to observe the molecular branch is not always valid, especially for the very high-temperature region where Bose-Fermi molecules thermally dissociate into unpaired atoms. However, since the molecular binding energy is large in the strong coupling regime, the temperature region where this idea works would be wide there. Figure \[fig15\] shows effects of population imbalance on the Fermi PES $I_{\rm F}({\bm p},\omega)$. As expected from Fig. \[fig3\], the multiple peak structure gradually becomes obscure in $I_{\rm F}({\bm p},\omega)$ with decreasing the ratio $N_{\rm B}/N_{\rm F}$ from unity. Simply interpreting this result as the local PES at various spatial positions in a trapped Bose-Fermi mixture, one can imagine that the detailed spectral structure is smeared out after the spatial average of the spectrum, when the contribution from the spatial region where the Bose atomic density $\rho_{\rm B}({\bm r})$ is much smaller than the Fermi atomic density $\rho_{\rm F}({\bm r})$ is dominant. To avoid this as possible as we can, it would be a good idea to detect spectra, avoiding the spatial region where $\rho_{\rm B}({\bm r})\ll \rho_{\rm F}({\bm r})$. For this purpose, the local photoemission-type experiment developed by JILA group[@PES_exp3] would be useful. As an alternative way, a box trap[@box_Fermi; @box_Bose1; @box_Bose2] may also be promising, because an almost uniform gas is realized there. Summary ======= To summarize, we have discussed single-particle excitations and effects of mass and population imbalances in a Bose-Fermi mixture. Including hetero-pairing fluctuations associated with an attractive Bose-Fermi interaction within the framework of the improved $T$-matrix approximation developed by two of the authors, we calculated the single-particle spectral weight (SW), as well as the photoemission spectrum (PES), in the normal state above the BEC phase transition temperature $T_{\rm c}$. In the mass- and population-balanced case ($m_{\rm B}=m_{\rm F}$ and $N_{\rm B}=N_{\rm F}$), it is known that strong hetero-pairing fluctuations cause couplings between atomic excitations and composite molecular excitations (atom-molecule coupling), as well as between Fermi atomic excitations and Bose atomic excitations (Fermi-Bose coupling). These many-body phenomena bring about additional two spectral peaks in the Fermi SW. Together with the ordinary spectral peak along the single-particle Fermi dispersion, the resulting Fermi SW exhibits a three-peak structure. In the presence of population imbalance, we showed that, when $N_{\rm B}/N_{\rm F}\ll 1$, the both atom-molecule and Fermi-Bose coupling phenomena become weak, so that the Fermi SW becomes close to that in a free Fermi gas. When $N_{\rm B}/N_{\rm F}\gg 1$, on the other hand, the former coupling remains to exist, leading to a two-peak structure in the Fermi SW. This difference comes from the fact that, while the atom-molecule coupling constant is dominated by the number of unpaired Bose atoms, the Fermi-Bose coupling constant is deeply related to the number $N_{\rm CF}$ of Bose-Fermi molecules ($N_{\rm CF}\le\min(N_{\rm B},N_{\rm F})$): The both coupling constants thus become small when $N_{\rm B}/N_{\rm F}\ll 1$. On the other hand, the former coupling remains non-zero even when $N_{\rm B}/N_{\rm F}\gg 1$. We have also examined how mass difference between a Fermi atom ($m_{\rm F}$) and a Bose atom ($m_{\rm B}$) modifies many-body corrections to single-particle excitations. In both the limits $m_{\rm B}/m_{\rm F}\ll 1$ and $m_{\rm B}/m_{\rm F}\gg 1$, we found that the Fermi SW exhibits, not a three-peak, but a two-peak structure; however, their physical meanings are different. When $m_{\rm B}/m_{\rm F}\ll 1$, the Fermi-Bose coupling causes the second peak line in addition to the ordinary peak line along the free-particle dispersion. In the opposite limit, the additional peak comes from the atom-molecule coupling. This is because the strengths of these couplings differently depend on the ratio $m_{\rm B}/m_{\rm F}$. When one goes away from these limiting cases, the Fermi SW exhibits the three-peak structure as in the mass-balanced case. We explicitly confirmed this in the cases of a mass-imbalanced $^{87}$Rb-$^{40}$K ($m_{\rm B}>m_{\rm F}$) and a $^{23}$Na-$^{40}$K ($m_{\rm B}<m_{\rm F}$) mixtures. We also pointed out that these many-body coupling phenomena may be observed by the photoemission-type experiment, by explicitly evaluating the photoemission spectra for these realistic examples. In this paper, we have treated a uniform Bose-Fermi mixture, for simplicity. In a real trapped mixture in a harmonic potential, we expect that the Fermi and Bose atoms have different density profiles, leading to local population imbalance. Although this inhomogeneous effect has only partially been examined in this paper, by considering the population-imbalanced case, to fully understand strong-coupling properties of a trapped Bose-Fermi mixture, it would be necessary to explicitly treat the trapped geometry. Besides this, we have also ignored an interaction between Bose atoms, which would be crucial for the stability of this system. These problems remain as our future challenges. Since the atom-molecule and Fermi-Bose couplings are characteristic many-body phenomena in a Bose-Fermi mixture with a hereto-pairing interaction, our results would contribute to further understanding of strong-coupling properties of this novel quantum many-body system. We thank D. Kagamihara and R. Sato for discussions. This work was supported by the KiPAS project at Keio University. Y.O. was supported by a Grant-in-aid for Scientific Research from MEXT and JSPS in Japan (No.JP18K11345, No.JP18H05406, and No.JP19K03689). Derivation of Eq. (\[eq.3.3\]) ============================== We approximate the Bose-Fermi scattering matrix $\Gamma_{\rm BF}({\bm q},i\omega_{\rm F})$ in Eq. (\[eq.10\]) to the composite molecular propagator, $$\Gamma_{\rm BF}({\bm q},i\omega_{\rm F})\simeq {Z \over i\omega_{\rm F}-\xi_{\bm{q}}^{\rm CF}}, \label{eq.A1}$$ where $Z$ is a positive constant[@iTMA1], and the molecular dispersion $\xi_{\bm p}^{\rm CF}$ is given below Eq. (\[eq.3.2\]). Although Eq. (\[eq.A1\]) is, strictly speaking, only justified in the strong-coupling limit, this approximate expression is still useful to grasp the essence of strong-coupling corrections to single-particle excitations. Substituting Eq. (\[eq.A1\]) into the self-energy $\Sigma_{\rm F}({\bm p},i\omega_{\rm F})$ in Eq. (\[eq.8\]), we have, after carrying out the $\omega'_{\rm F}$-summation, $$\begin{aligned} \Sigma_{\rm F}({\bm p},i\omega_{\rm F}) &=& Z\sum_{\bm q} \left[ {n_{\rm B}({\tilde \xi}_{\bm q}^{\rm B}) \over i\omega_{\rm F}-\xi^{\rm CF}_{{\bm p}-{\bm q}} +{\tilde \xi}_{\bm q}^{\rm B}} + {f(\xi_{\bm q}^{\rm CF}) \over i\omega_{\rm F}+{\tilde \xi}_{{\bm p}-{\bm q}}^{\rm B}-\xi_{\bm q}^{\rm CF}} \right] \nonumber \\ &\simeq& {Z{\tilde N}_{\rm B}^0 \over i\omega_{\rm F} -\xi_{\bm p}^{\rm CF}} + \left\langle {ZN_{\rm CF} \over i\omega_{\rm F}+{\tilde \xi}_{{\bm k}_{\rm CF}-{\bm p}}^{\rm B}} \right \rangle_{{\bm k}_{\rm CF}} . \label{eq.A2}\end{aligned}$$ Here, the definitions of the parameters in Eq. (\[eq.A2\]) are explained in the text. In obtaining the first term in the last line, we have set ${\bm q}=0$ in the denominator by using that the Bose distribution function $n_{\rm B}({\tilde \xi}_{\bm q}^{\rm B})$ diverges in the low momentum limit at $T_{\rm c}$. For the last term in the last line in Eq. (\[eq.A2\]), we have approximated ${\bm q}$ in the denominator to ${\bm k}_{\rm CF}$, by noting that the region near the Fermi surface of the composite Fermi molecules is important. Substitution of Eq. (\[eq.A2\]) into Eq. (\[eq.3\]) gives Eq. (\[eq.3.3\]). [99]{} I. Bloch, J. Dalihard, and W. Zwerger, Rev. Mod. Phys. [**80**]{}, 885 (2008). S. Giorgini, L. Pitaevskii, and S. Stringari, Rev. Mod. Phys. [**80**]{}, 1215 (2008). I. M. Georgescu, S. Ashhab, and F. Nori, Rev. Mod. Phys. [**86**]{}, 153 (2014). C. Gross and I. Bloch, Science [**357**]{}, 955 (2017). M. Greiner, O. Mandel, T. Esslinger, T. W. Hänsch, and I. Bloch, Nature [**415**]{}, 39 (2002). C. A. Regal, M. Greiner, and D. S. Jin, Phys. Rev. Lett. [**92**]{}, 040403 (2004). M. W. Zwierlein, C. A. Stan, C. H. Schunck, S. M. F. Raupach, A. J. Kerman, and W. Ketterle, Phys. Rev. Lett. [**92**]{}, 120403 (2004). J. Kinast, S. L. Hemmer, M. E. Gehm, A. Turlapov, and J. E. Thomas, Phys. Rev. Lett. [**92**]{}, 150402 (2004). M. Bartenstein, A. Altmeyer, S. Riedl, S. Jochim, C. Chin, J. H. Denschlag, and R. Grimm, Phys. Rev. Lett. [**92**]{}, 203201 (2004). D. M. Eagles, Phys. Rev. [**186**]{}, 456 (1969). A. J. Leggett, in [*Modern Trends in the Theory of Condensed Matter*]{}, ed. A. Pekalski and J. Przystawa (Springer Verlag, Berlin, 1980), p. 14. C. A. R. Sá de Melo, M. Randeria, and J. R. Engelbrecht, Phys. Rev. Lett. [**71**]{}, 3202 (1993). M. Randeria, in [*Bose-Einstein Condensation*]{}, ed. by A. Griffin, D. W. Snoke, and S. Stringari (Cambridge University Press, New York, 1995). G. C. Strinati, P. Pieri, G. Röpke, P. Schuck, and M. Urban, Phys. Rep. [**738**]{}, 1 (2018). C. Chin, R. Grimm, P. Julienee, and E. Tiesinga, Rev. Mod. Phys. [**82**]{}, 1225 (2010). A. G. Truscott, K. E. Strecker, W. I. McAlexander, G. B. Partridge, and R. G. Hulet, Science [**291**]{}, 2570 (2001). F. Schreck, L. Khaykovich, K. L. Corwin, G. Ferrari, T. Bourdel, J. Cubizolles, and C. Salmon, Phys. Rev. Lett. [**87**]{}, 080403 (2001). Z. Hadzibabic, C. A. Stan, K. Dieckmann, S. Gupta, M. W. Zwierlein, A. Görlitz, and W. Ketterle, Phys. Rev. Lett. [**88**]{}, 160401 (2002). Y.-P. Wu, X.-C. Yao, H.-Z. Chen, X.-P. Liu, X.-Q. Wang, Y.-A. Chen, and J.-W. Pan, J. Phys. A [**50**]{}, 094001 (2017). C. A. Stan, M. W. Zwierlein, C. H. Schunck, S. M. F. Raupach, and W. Ketterle, Phys. Rev. Lett. [**93**]{}, 143001 (2004). S. Inouye, J. Goldwin, M. L. Olsen, C.Ticknor, J. L. Bohn, and D. S. Jin, Phys. Rev. Lett. [**93**]{}, 183201 (2004). F. Ferlaino, C. D’Errico, G. Roati, M. Zaccanti, M. Inguscio, G. Modugno, and A. Simoni, Phys. Rev. A [**73**]{}, 040702 (2006). M. Zaccanti, C. D’Errico, F. Ferlaino, G. Roati, M. Inguscio, and G. Mondugno, Phys. Rev. A [**74**]{}, 041605 (2006). S. Ospelkaus, C. Ospelkaus, L. Humbert, K. Sengstock, and K. Bongs, Phys. Rev. A [**97**]{}, 120403 (2006). B. Deh, C. Marzok, C. Zimmermann, and Ph. W. Courteille, Phys. Rev. A [**77**]{}, 010701 (2008). J. W. Park, C.-H. Wu, I. Santiago, T. G. Tiecke, S. Will, P. Ahmadi, and M. W. Zwierlein, Phys. Rev. A [**85**]{}, 051602(R) (2012). T. Schuster, R. Scelle, A. Trautmann, S. Knoop, M. K. Oberthaler, M. M. Haverhals, M. R. Goosen, S. J. J. M. F. Kokkelmans, and E. Tiemann, Phys. Rev. A [**85**]{}, 042721 (2012). M. Repp, R. Pires, J. Ulmanis, R. Heck, E. D. Kuhnle, and M. Weidemüller, Phys. Rev. A [**87**]{}, 010701 (2013). M.-J. Zhu, H. Yang, L. Liu, D.-C. Zhang, Y.-X. Liu, J. Nan, J. Rui, B. Zhao, J.-W. Pan, and E. Tiemann, Phys. Rev. A [**96**]{}, 062705 (2017). J. J. Zirbel, k.-K. Ni, S. Ospelkaus, J. P. D’Incao, C. E. Wieman, J. Ye, and D. S. Jin, Phys. Rev. Lett. [**100**]{}, 143201 (2008). C.-H. Wu, J. W. Park, P. Ahmadi, S. Will, and M. W. Zwierlein, Phys. Rev. Lett. [**109**]{}, 085301 (2012). F. Wang, X. He, X. Li, B. Zhu, J. Chen, and D. Wang, New. J. Phys. [**17**]{}, 035003 (2015). M. S. Heo, T. T. Wang, C. A. Christensen, T. M. Rvachocv, D. A. Cotta, J. H. Choi, Y. R. Lee, and W. Ketterle, Phys. Rev. A [**86**]{}, 021602(R) (2012). T. D. Cumby, R. A. Shewmon, M.-G. Hu, J. D. Perreault, and D. S. Jin, Phys. Rev. A [**87**]{}, 012703 (2013). L. D. Macro, G. Valtolina, K. Matsuda, W. G. Tobias, J. P. Covey, and J. Ye, Science [**363**]{}, 853 (2019). I. F.-Barbut, M. Delehaye, S. Laurent, A. T. Grier, M. Pierce, B. S. Rem, F. Chevy, and C. Salmon, Science [**345**]{}, 1035 (2014). X.-C. Yao, H.-Z Chen, Y.-P. Wu, X.-P. Liu, X.-Q. Wang, X. Jiang, Y. Deng, Y.-A. Chen, J.-W. Pan, Phys. Rev. Lett. [**117**]{}, 145301 (2016). R. Roy, A. Green, R. Bowler and S. Gupta, Phys. Rev. Lett. [**118**]{}, 055301 (2017). B. J. DeSalve, K. Patel, G. Cai, and C. Chin, Nature [**568**]{}, 61 (2019). E. Fratini and P. Pieri, Phys. Rev. A [**81**]{}, 051605(R) (2010). E. Fratini and P. Pieri, Phys. Rev. A [**85**]{}, 063618 (2012). E. Fratini and P. Pieri, Phys. Rev. A [**88**]{}, 013627 (2013). G. Bertania, E. Fratini, S. Giorgini, and P. Pieri, Phys. Rev. Lett. [**110**]{}, 115303 (2013). A. Guidini, G. Bertania, E. Fratini, and P. Pieri, Phys. Rev. A [**89**]{}, 023634 (2014). A. Guidini, G. Bertania, D. E. Galli, and P. Pieri, Phys. Rev. A [**91**]{}, 023603 (2015). D. Kharga, D. Inotani, R. Hanai, and Y. Ohashi, J. Phys. Soc. Jpn. [**86**]{}, 084301 (2017). D. Kharga, H. Tajima, P. van Wyk, D. Inotani, and Y. Ohashi, J. Phys. Soc. Jpn. [**86**]{}, 074301 (2017). M. Y. Kagan, I. V. Brodsky, D. V. Efremov, and A. K. Klaptsov, Phys. Rev. A [**70**]{}, 023607 (2004). S. Powell, S. Sachdev, H. Büchler, Phys. Rev. B [**72**]{}, 024534 (2005). D. Ludwig, S. Floerchinger, S. Moroz, and C. Wetterich, Phys. Rev. A [**84**]{}, 033629 (2011). T. Watanabe, T. Suzuki, and P. Schuck, Phys. Rev. A [**78**]{}, 033601 (2008). J.-L. Song and F. Zhou, Phys. Rev. A [**84**]{}, 013601 (2011). T. Sogo, P. Schuck, and M. Urbam, Phys. Rev. A [**88**]{}, 023613 (2013). Z.-Q. Yu, S. Zhang, and H. Zhai, Phys. Rev. A [**83**]{}, 041603(R) (2011). K. Shirasaki, E. Nakano, and H. Yabu, Phys. Rev. A [**90**]{}, 063629 (2014). Z. Wu and G. M. Bruun, Phys. Rev. Lett. [**117**]{}, 245302 (2016). J. M. Midtgaard, Z. Wu, and G. M. Bruun, Phys. Rev. A [**94**]{}, 063601 (2016). J. M. Midtgaard, Z. Wu, and G. M. Bruun, Phys. Rev. A [**96**]{}, 033605 (2017). H. Heiselberg, C. J. Pethick, H. Smith, and L. Viverit, Phys. Rev. Lett. [**85**]{}, 2418 (2000). F. Matera and A. Dellafore, Euro. Phys. J. D [**65**]{}, 515 (2011). J. J. Kinnunem, Z. Wu, and G. M. Bruun, Phys. Rev. Lett. [**121**]{}, 253402 (2018). K. Maeda, G. Baym, and T. Hatsuda, Phys. Rev. Lett. [**103**]{}, 085301 (2009). S. Tsuchiya, R. Watanabe, and Y. Ohashi, Phys. Rev. A [**80**]{}, 033613 (2009). R. Haussmann, M. Punk, and W. Zwerger, Phys. Rev. A [**80**]{}, 063612 (2009). S. Tsuchiya, R. Watanabe, and Y. Ohashi, Phys. Rev. A [**84**]{}, 043647 (2011). P. Perali, F. Palestini, P. Pieri, G. C. Strinati, J. T. Stewart, J. P. Gaebler, T. E. Drake, and D. S. Jin, Phys. Rev. Lett. [**106**]{}, 060402 (2011). B. Mukherjee, Z. Yan, P. B. Patel, Z. Hadzibabic, T. Yefsah, J. Struck, and M. W. Zwierlein, Phys. Rev. Lett. [**118**]{}, 123401 (2017). A. L. Gaunt, T. F. Schmidutz, I. Gotlibovych, R. P. Smith, and Z. Hadzibabic, Phys. Rev. Lett. [**110**]{}, 200406 (2013). R. Lopes, C. Eigen, N. Navon, D. Cl[é]{}ment, R. P. Smith, and Z. Hadzibabic, Phys. Rev. Lett. [**119**]{}, 190404 (2017). M.-G. Hu, M. J. Van de Graaff, D. Kedar, J. P. Corson, E. A. Cornell, and D. S. Jin, Phys. Rev. Lett. [**117**]{}, 055301 (2016). N. B. Jørgensen, L. Wacker, K. T. Skalmstang, M. M. Parish, J. Levinsen, R. S. Christensen, G. M. Bruun, and J. J. Arlt, Phys. Rev. Lett. [**117**]{}, 055302 (2016). F. Grusdt and E. Demler, arXiv:1510.04934. S. P. Rath and R. Schmidt, Phys. Rev. A [**88**]{}, 053632 (2013). R. S. Christensen, J. J. Arlt, and G. M. Bruun, Phys. Rev. A [**96**]{}, 063622 (2017). N.-E. Guenther, Pietro Massignan, M, Lewestein, and G. M. Bruun, Phys. Rev. Lett. [**120**]{}, 050405 (2018). A. C.-Guaridan, L. A. P. Ardila, T. Pohl, and G. M. Bruun, Phys. Rev. Lett. [**121**]{}, 013401 (2018). J. T. Stewart, J. P. Gaebler, and D. S. Jin, Nature [**454**]{}, 744 (2008). J. P. Gaebler, J. T. Stewart, T. E. Drake, D. S. Jin, A. Perali, P. Pieri, and G. C. Strinati, Nat. Phys. [**6**]{}, 569 (2010). Y. Sagi, T. E. Drake, R. Paudel, R. Chapurin, and D. S. Jin, Phys. Rev. Lett. [**114**]{}, 075301 (2015). P. Törmä, Phys. Scr. [**91**]{}, 043006 (2016). M. Ota, H. Tajima, R. Hanai, D. Inotani, and Y. Ohashi, Phys. Rev. A [**95**]{}, 053623 (2017). N. M. Hugenholtz and D. Pines, Phys. Rev. [**116**]{}, 489 (1959). Although a two-body bound state vanishes at the unitary, medium effects gives a bound state at this interaction strength, as will be explained later.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Auctions are protocols to allocate goods to buyers who have preferences over them, and collect payments in return. Economists have invested significant effort in designing auction rules that result in allocations of the goods that are desirable for the group as a whole. However, for settings where participants’ valuations of the items on sale are their private information, the rules of the auction must deter buyers from misreporting their preferences, so as to maximize their own utility, since misreported preferences hinder the ability for the auctioneer to allocate goods to those who want them most. Manual auction design has yielded excellent mechanisms for specific settings, but requires significant effort when tackling new domains. We propose a deep learning based approach to automatically design auctions in a wide variety of domains, shifting the design work from human to machine. We assume that participants’ valuations for the items for sale are independently sampled from an unknown but fixed distribution. Our system receives a data-set consisting of such valuation samples, and outputs an auction rule encoding the desired incentive structure. We focus on producing truthful and efficient auctions that minimize the economic burden on participants. We evaluate the auctions designed by our framework on well-studied domains, such as multi-unit and combinatorial auctions, showing that they outperform known auction designs in terms of the economic burden placed on participants.' author: - | Andrea Tacchetti, DJ Strouse, Marta Garnelo, Thore Graepel and Yoram Bachrach\ DeepMind, UK\ `{atacchet, strouse, garnelo, thore, yorambac}@google.com` bibliography: - 'references.bib' title: A Neural Architecture for Designing Truthful and Efficient Auctions --- Introduction {#sec:introduction} ============ Mechanism design is a field in economics that deals with setting incentives and interaction rules among self-interested agents so as to achieve desired objectives for the group as a whole. It is sometimes referred to as “inverse game theory”: in game theory we set the rules of a game, and study the behaviors that emerge, while in mechanism design we have a target behavior we wish to encourage, and we set the rules of the game so that agents acting in their own self-interest will gravitate towards that desired outcome. One prominent problem in mechanism design is engineering auction rules, as auctions account for a large proportion of economic activity, such as the sponsored search auction (the main source of revenue for search engines), e-commerce websites such as eBay, or the fine art market [@bajari2003winner; @edelman2007internet; @klemperer2018auctions]. One possible goal of an auction design is to maximize the revenue to the auctioneer [@myerson1981optimal]. In many cases, however, we are merely interested in allocating a set of goods in order to maximize the total welfare of participants (and thus minimize revenue to the auctioneer). One example are spectrum auctions [@cramton1997fcc], in which governments want to allocate the rights (licenses) to transmit signals over specific bands of the electromagnetic spectrum. The government may wish to allocate the scarce transmission rights to the firms who value these the most, with the goal of maximizing job creation, trade, and economic welfare. However, the true valuation a firm has for a spectrum band is only known to the firm, rather than to the government. If the government simply declared they would give a band to the firm who wants it most without extracting payments, then all firms who want the band a non-zero amount would be incentivized to lie and say they value the band arbitrarily highly, and the government could not ensure an optimal allocation. Thus, the auction design challenge for the government becomes: which prices should it charge in order to get truthful reports regarding firms’ valuations, and optimally allocate the spectrum bands, while still minimizing the economic burden on participants? We propose a learning approach to auction design. The point of departure from the existing economics literature is that we make the (often reasonable) assumption that bidders’ valuations for the goods up for sale cannot take any value, but rather are sampled from an unknown, but fixed, probability distribution (e.g. it is very unlikely anyone would pay \$500,000 for a burrito). Under these settings we introduce a representation of bidders’ preferences and a network architecture that can be used to learn auction rules that a) incentivize truthful reports from the participants, b) result in the social-welfare-maximizing allocation of the goods in question, and c) place minimal economic burden on the participants (i.e. extract minimal payments). We show that the proposed approach can learn truthful mechanisms under a wide variety of settings, including various “bidding languges” [@nisan2000bidding] (i.e. the set or outcomes that bidders can have preferences over), arbitrary distributions of valuations, and arbitrary numbers of participants. Moreover, the resulting payment rules *generalize* over varying number of participants. Auctions are a pillar of economics and the market protocol of choice for a significant portion of world-wide trade. Similar to what Hartford et al. [@hartford2016deep] have done for modeling human strategic behavior, here we show that, under reasonable assumptions, designing auctions that shepherd the behavior of rational participants towards desirable outcomes can be cast as a supervised function approximation problem, thus unlocking the application of modern machine learning methods, and in particular deep learning, to this field. Background and notation {#sec:bg-notation} ======================= [**Mechanism design**]{} Mechanism design deals with choosing from a set $K$ of possible alternatives, where we have a set $N=\{1,2,\ldots,n\}$ of agents who each have preferences regarding the alternatives in $K$, expressed in monetary terms. [**Auction design**]{} relates to the specific case where we have a set $I$ of items, and the alternative set $K$ consists of all the possible ways to allocate the items to the agents. We call a subset of items $B \subseteq I$ a [**bundle**]{}, and let $\mathcal{P}(I)$ be the power set of $I$ (that is, the set of all possible bundles). An [**allocation**]{} of the items is a function $k : N \rightarrow \mathcal{P}(I)$ mapping each agent $i$ to a bundle of items $k(i) \subseteq I$, such that for any $i \neq j$ we have $k(i) \cap k(j) = \emptyset$ (i.e. no item is allocated more than once). Each agent, with knowledge of their own true preferences, reports what is commonly referred to as a “type”: for each allocation $k \in K$ the agent communicates to the mechanism a valuation for that outcome $\theta_i(k) \in \Theta_i$. In particular, participants may choose to report truthfully and submit $v_i(k)$[^1]. The mechanism then selects an allocation according to a choice rule $c : \Theta_1 \times ... \times \Theta_n \rightarrow K$ and determines the agents’ payments using a payment rule $t_i : \Theta_1 \times \ldots \times \Theta_n \rightarrow \mathbb{R}$, where $t_i$ is the payment for the $i^{\text{th}}$ agent. Note that “payments” can be negative, that is the auctioneer can also pay participants. After payments are collected, each agent thus derives utility $u_i(k,t_i)=v_i(k)-t_i$ from the interaction[^2]. [**Allocation efficiency**]{} The main goal of the designs we consider is to choose an efficient outcome: allocate items to those who want them the most, maximizing social welfare. We thus fix the choice function $c$ to $c=\arg \max_{k \in K} \sum_i v_i(k) = k^*$. [**Strategic behavior**]{} Selecting the welfare maximizing allocation is difficult when the mechanism does not have access to the true preferences of each agent, but only to their reported types. This asymmetry in information leads to strategic behavior: rational participants will report whatever preference $\theta_i$ maximizes their utility under the mechanism (post payment). Let $\theta_{-i}$ indicate all reports, truthful or otherwise, from all agents but $i$; then rational agents will report: $\theta_i = \arg\max_{\theta \in \Theta_i} u_i(c(\theta_{-i}, \theta), t_i(\theta_{-i}, \theta))$. In general $\theta_i \neq v_i$. [**Truthful mechanisms**]{} In the presence of rational agents, and for our choice of allocation function, it is possible to select a payment rule that makes reporting one’s true preferences the dominant strategy. That is, for any agent $i$, and for all possible reports, or misreports, from other players $\theta_1,\ldots,\theta_{i-1},\theta_{i+1},\ldots,\theta_n$, the best course of action is to tell the truth: $\arg\max_{\theta \in \Theta_i} u_i(k^*(\theta_{-i}, \theta), t_i(\theta_{-i}, \theta)) = v_i$ (where we bypassed the explicit dependence on the choice function $c$, and let $k^*$ depend on $\theta$ directly). We restrict our attention to mechanisms that are both efficient and truthful. The only such mechanisms are members of the Groves family, and their payment rule can be written as [@groves1973incentives; @green1979incentives; @green1979social]: $$\label{eq:groves-payment} t_i( \theta_{-i}, \theta_i ) = h( \theta_{-i} ) - \sum_{j \neq i} v_j( k^*(\theta_{-i}, \theta_i)),$$ where, $h : \Theta_{-i} \rightarrow \mathbb{R}$ may be any function that only depends on the reported types of agents other than $i$, and $k^*$ is the optimal allocation defined previously. [**Individual rationality**]{} In the presence of strategic agents, we must ensure that bidders are never worse off participating in the auctions we design than not. We should guarantee that our auctions are individually rational: any agent who truthfully reports their preferences realizes a non-negative utility. That is, regardless of reports from other agents $\theta_{-i}$, we wish to have $u_i(v_i, \theta_{-i}) = u_i(k^*(v_i, \theta_{-i}), t_i(v_i, \theta_{-i})) \geq 0$. [**Weak budget balance**]{} Finally, we wish to design mechanisms that do not require a subsidy to operate. That is, we require that the sum of payments collected by the mechanism be non-negative: $\sum_i t_i \geq 0$. [**Vickrey-Clarke-Groves auctions**]{} One of the main results of mechanism design is an auction rule that satisfies all the criteria we listed above for any realization of agents’ preferences: the Vickrey-Clarke-Groves (VCG) auction. VCG is both efficient, and truthful, and as such it is a member of the Groves family. It is characterized by the choice of function $h(\theta_{-i})$ that completes the payment rule in Eq. \[eq:groves-payment\]: $h_{\text{VCG}} = \sum_{j \neq i} v_j(k^*(\theta_{-i}))$. In words, $h_{\text{VCG}}$ is the collective value realized by all other agents when agent $i$ is removed from the auction. Thus, the completed VCG payment $t_\text{VCG}$ for agent $i$ (see Eq. \[eq:groves-payment\]) is the *reduction* in the collective value realized by all other agents due to agent $i$’s participation in the auction.[^3] A special case of a VCG auction for a *single item* is the well-known second-price auction (e.g. an eBay auction with no reserve price). VCG is the most widely accepted truthful and efficient mechanism (e.g. it is used for Facebook ad auctions [@varian2014vcg]). VCG does not, however, aim to minimize the economic burden on participants, as we do here. [**Bidding languages**]{} Since we focus on efficient mechanisms, we must ensure that $k^*$ can be computed quickly, even for relatively large numbers of players (see Allocation Efficiency). We thus restrict the way in which participants may express their preferences. Such representations are called bidding languages [@nisan2000bidding]. We consider the following three languages. [**Multi-unit auctions with decreasing marginal utilities**]{} The first bidding language we examine considers selling multi-unit bundles to participants’ whose preferences depend only on the [*size*]{} of bundles but not on their component objects. This language is useful for selling multiple [*identical*]{} units of the same kind. We further impose that larger bundles cannot be valued less than smaller ones. In these auctions $k^*$ can be calculated greedily by allocating objects one by one. [**Heterogeneous objects with unit demand**]{} The second bidding language we consider is useful when players can take advantage of at most one of items they receive. For example, vacation packages for a specific week. The valuation for a bundle of items, in this case, is identical to the valuation of the best object in the bundle. The allocation function $k^*$ is found by solving the maximum-weighted bipartite matching between bidders and items, where participants’ preferences are incorporated as weights. [**Hierarchical bundles**]{} Finally, we consider a bidding language that is useful to express preferences for a hierarchy of bundles. For example, home builders might bid to be awarded the contract to develop two lots, with up to two new homes within each lot. The spatial nature of the work makes building individual homes in separate lots is less cost-effective. Thus participants can express preferences for a hierarchy of bundles: component objects are arranged as the leaves of a binary tree, and valuations can be expressed for leaf-nodes (individual objects), or for any sub-tree. The integer program required to find $k^*$ can be relaxed as a feasible linear program [@nisan2000bidding]. Problem statement and main contributions {#sec:problem-statement} ======================================== Equipped with the definitions of Sec. \[sec:bg-notation\], we can proceed to state our objective: [*To design truthful and allocatively-efficient auctions, minimizing the sum of payments collected by the mechanism, while keeping the auction individually-rational and weakly budget balanced.*]{} Minimizing the sum of payments is useful in settings such as the spectrum auction, where the goal is to allocate a scarce resource in an optimal way; the payments $t_i$ are by-products resulting from the need to elicit true reports, so it is desirable to minimize them. Additionally, we strive to have adhere to the following desiderata: we seek a payment rule that is a) [*“convolutional” over players*]{} [@lecun1995convolutional] (i.e. the same function is used to compute the payment owed by each player), b) [*invariant to the order of other participants for each player*]{} (i.e. the payment of player $2$ does not change if players $1$ and $3$ swap bids), and c) [*robust to changes in the number of participants*]{}. In pursuit of this goal, we propose three main technical contributions. a) We show how the problem of [**designing truthful and efficient auctions can be cast as supervised function approximation**]{}. b) We introduce a novel [**representation of efficient auctions**]{} as a collection of counterfactual smaller auctions. c) We propose a [**network architecture to learn Groves payment rules**]{} based on our representation which supports various bidding languages and an arbitrary (and even varying) number of bidders. Related Work {#sec:related-work} ============ Mechanism design is a relatively mature field. We rely on the framework of the Vickrey-Clarke-Groves mechanism [@clarke1971multipart] presented in the 1970s, based on earlier work on auctions that Vickrey had conducted in the 1960s [@vickrey1961counterspeculation]. The design of auctions and mechanisms has been done predominantly [*manually*]{}, where a [*person*]{} uses their experience or intuition to come up with interaction or payment rules leading to their desired objective. Economists have designed incentive schemes for various goals, such as minimizing the burden on participants while maintaining efficiency [@green1979incentives; @ausubel2006lovely] or maximizing revenue [@myerson1981optimal; @bulow1989simple; @roth2002economist]. The field of [*automated mechanism design*]{}, where we let a [*computer*]{} design an incentive scheme to meet desired objectives, is relatively new [@conitzer2002complexity; @sandholm2003automated; @conitzer2004self; @hajiaghayi2007automated; @guo2010computationally]. Early work on automated mechanism design has focused on producing incentive compatible mechanisms, where truthfulness is a Nash equilibrium [@sandholm2003automated; @conitzer2004self]. In contrast, we aim to achieve truthfulness in the strong sense of a dominant strategy, where agents opt for a truthful report no matter what other agents do. More recently, economists have given significant attention to efficient mechanisms where truthfulness is a dominant strategy, characterizing the family of Groves mechanisms as the only class of mechanisms which are truthful, efficient, individually rational and weakly budget balanced [@groves1973incentives; @nisan2001algorithmic]. They have also provided negative results, showing that it is impossible to guarantee [*full*]{} budget-balance (i.e. $\sum_i t_i=0$), in fully truthful and efficient mechanisms [@green1979incentives]. Given these results, researchers have manually constructed [*Redistribution Mechanisms*]{}, specific members in the Groves family that maximize budget-balance (i.e. minimize agent payments while requiring no subsidy) in restricted settings. We also use the general family of Groves mechanisms, or more specialized cases of Groves redistribution mechanisms, but rather than manually building incentive schemes to achieve high budget balance in specific settings, we take an automated mechanism design approach, using machine learning to identify good members of the Groves family. Closest to our work are recent approaches for automated mechanism design through machine learning, and deep learning in particular [@dutting2017optimal; @feng2018deep; @manisha2018learning]. These approaches search a family of payment functions for a mechanism with desired properties by defining a loss relating to the desired properties. While our approach is similar, we propose a more elaborate neural network architecture to capture reasonable auction rule properties, tackle the more demanding domain of combinatorial auctions under various bidding languages [@de2003combinatorial; @nisan2000bidding], and crucially we are able to learn mechanisms that are truthful in the strong sense and support arbitrary, and even variable number of bidders. Methods: mechanism design as a supervised learning problem {#sec:methods} ========================================================== ![ (best viewed in color). In this example we represent a multi-unit auction with decreasing marginal utilities with five players, and three objects. We construct the network input to compute the Groves payment rule or redistribution for player $1$. The input tensor is of size $(n-1) \times |K|\times 2|K| + 1 = 4\times3\times7$ and is constructed as shown in the figure, on the left (darker shades of red indicate higher valuations in the first channel). This representation is processed with a 2-layer CNN that extracts a per-player distributed representation of preferences and a per-player 2-Layer MLP (with shared weights across the players). The resulting embeddings are sum-pooled to build invariance to the ordering of players, and robustness to the number of participants, and decoded into a single positive number.[]{data-label="fig:auction-cnn"}](AuctionCNN-3.pdf){width="\textwidth"} Here we introduce the details of our main technical contributions: we [**show how the problem of completing the Groves payment is equivalent to supervised function approximation**]{}. We introduce our [**novel representation of efficient auctions**]{}, and we propose a [**network architecture to learn social-utility-maximizing, truthful auctions**]{}. We seek to design efficient and truthful mechanisms that are, at least in expectation, as close to budget balanced as possible. As discussed in Sec. \[sec:bg-notation\], all mechanisms that are truthful and efficient belong to the Groves family (and vice versa), so we restrict our search to this family, and effectively seek to complete the Groves payment rule by selecting a function $h: \Theta_{-i} \rightarrow \mathbb{R}$ (see Eq. \[eq:groves-payment\]). Loss function: completing the Groves payment rule ------------------------------------------------- The aim is then to complete the payment rule $t_i$ of a Groves mechanism so that, in expectation over valuation profiles sampled from $\rho$, we minimize the sum total of payments received by the mechanism. However, minimizing payments without any further constraint will result in mechanisms that require a subsidy to operate. Since this is undesirable, we incorporate a [*non-deficit*]{} constraint. Similarly, ensure that strategic players are never worse off participating in our mechanism than not, by including an [*individual rationality*]{} constraint for all players. The resulting “ideal” mechanism design problem we wish to solve is thus: $$h^* = \arg\min_{h\in\mathcal{H}}\mathbb{E}_{v_i \sim \rho} \left[\sum_{i=1}^n t_i\right] \quad \text{s.t. }\left[\sum_{i=1}^n t_i\right] \geq 0,\text{ and, } v_i(k^*) - t_i \geq 0,$$ where $t_i$ is like in Eq. \[eq:groves-payment\]. As mentioned above, we assume we do not have access to the true distribution $\rho$, so that we cannot solve this minimization analytically. We do, however, assume we have access to a data-set of $L$ realized $n$-player profiles $D = \{(v_1^l,\ldots,v_n^l | l = 1,\ldots,L\}$, sampled i.i.d. from $\rho$. We therefore use use Lagrange multipliers $\lambda_b$, and $\lambda_r$ to encode the non-deficit, and individual rationality constraints, and minimize the empirical version of our loss: $$\label{eq:loss-function} \hat{h} = \arg\min_{h\in\mathcal{H}} \sum_{l=1}^L \left(\sum_{i=1}^n t_i^l + \lambda_b\left(\min\left\{\sum_{i=1}^n t_i^l, 0\right\}\right)^2 +\lambda_r \sum_{i=1}^n\left(\left(\min\left\{v_i^l(k^*) - t_i^l, 0\right\}\right)^2\right)\right).$$ [**Selecting a Groves payment rule**]{} Concretely, we introduce two alternatives to learning a Groves payments rule $\hat{h}$: first, we investigate constructing a neural network to implement $\hat{h}$ directly and minimize the empirical loss in Eq. \[eq:loss-function\], given a data-set of realized valuation profiles. [**Learning a VCG redistribution mechanism**]{} Our second approach amounts to learning a VCG redistribution mechanism. In this case, we use a neural network to implement a redistribution function $r(v_1,\ldots,v_{i-1},v_{i+1},\ldots,v_n)$, and let $\hat{h}(\cdot) = h_{\text{VCG}}(\cdot) - r(\cdot)$[^4]. Note that in this case individual rationality can be guaranteed by simply ensuring that $r$ takes non-negative values, since VCG is individually rational and giving payments back can only increase participants’ utilities. The same representation and network architecture is used in both settings. The hypothesis space: auction representation and network architecture {#sec:hypothesis-space} --------------------------------------------------------------------- [**Representing auctions**]{} We select a hypothesis space $\mathcal{H}$ so that, in practice, we can solve the minimization problem in Eq. \[eq:loss-function\], given access to a data-set $D$ of valuation profiles. To this end, we introduce a novel representation of auctions that supports learning Groves payment rules with Deep Neural Networks. Fig. \[fig:auction-cnn\] shows an example of our representation and architecture for an auction with three objects and five participants. When computing $t_i$, the payment owed by player $i$, the function we wish to learn has access to reports from “other” players, but no knowledge of player $i$’s valuation (see Eq. \[eq:groves-payment\]). We construct our representation to highlight the magnitude of each individual bid, and preference profile, relative to the rest of the “other” players’ types. The intuition behind this choice is that by comparing the available bids to each other, a network can construct a sense of how likely it is that these will be surpassed or matched by the unseen preference profile. This is achieved as follows: first, since we focus on efficient mechanisms, we assume we are given access to an “allocation oracle” (a function that for any set of valid preferences profiles, returns the welfare maximizing allocation $k^* \in K$[^5], see Sec. \[sec:bg-notation\]). Second, we choose to represent each of the $v_{-i}$ as outcomes of $|K|$ counter-factual auctions, each for the most valuable $p$ bundles ($p=1,\ldots,|K|$), thus providing information about the relative rank of each bundle valuation and preference profile. We provide evidence that an alternative representation of the same information as a flat vector results in substantially worse auction designs. Precisely, given a data-set of realized valuation profiles $D$, and an allocation oracle, for each player $i$, we construct a tall “image” with “spatial dimensions” $|K|\times (n-1)$, and $2|K| + 1$ “channels”. The first “channel” is a matrix $V_{-i} \in \mathbb{R}_{+}^{|K| \times (n-1)}$ with non-negative entries $(m,j)$ representing the utility player $j$ would realize from receiving bundle $m$. Each successive channel $p$ is constructed by considering the outcome of a counter-factual auction where the $n-1$ players bid for the $p$ most valuable bundles. In particular, the second channel contains the allocation matrix $k_1^* \in \{0,1\}^{|K| \times (n-1)}$ with entries $(m,j) = 1$ if bidder $j$ is allocated bundle $m$, and zero otherwise. The third channel represents the amount of utility realized by each player for this allocation (the element-wise product between the first and second channels). Similarly, the fourth channel contains $k_2^*$: the allocation for two bundles, and the fifth channel contains the element-wise product between channels 1 and 4, and so on until all bundles are considered. We alter this representation slightly to in multi-unit auctions with decreasing marginal utilities. In this case the matrix $V_{-i}$ contains, for each player, the marginal utility of adding one item to their bundle and has size $|B| \times (n-1)$, with $B$ the set of available items. [**A network architecture to learn Groves payment rules**]{} Given our auction representation, we propose an architecture to learn a Groves payment rule that satisfies the desiderata outlined in Sec. \[sec:problem-statement\] and is: a) [*“convolutional” over players*]{}, b) [*invariant to the order of other participants for each player*]{}, and c) [*robust to changes in the number of participants*]{}. For each player $i$, we construct the input tensor of size $|K| \times (n-1) \times (2|K| + 1)$ described above and pass it through a $2$-layer CNN. The first layer uses $64$ filters of spatial size $1\times1$ so as to construct an embedding of each individual bid (how soon each bundle is allocated, and how much utility it realizes can be readily extracted from a single “column” in our representation). The second CNN layer has $64$ filters of size $|K|\times 1$. The CNN’s output has size $1 \times (n - 1) \times 64$, and contains an embedding of each of the $n-1$ players’ preferences. We follow our CNN with a 2-Layer $64$ hidden and output units MLP, which we apply independently to each of the $(n-1)$ player preference embeddings to produce a new embedding for each player. We then sum-pool over the $n-1$ players (which guarantees the desired robustness properties), and apply a linear decoder (with ReLU rectification) to output a single value for either $\hat{h}$ directly, or for a redistribution function $r$[^6]. Experimental procedure ---------------------- For each combination of number of participants, valuation distribution and bidding language we consider, we construct an “auction simulator” that returns sample auctions (i.e. valuation profiles for all participants, expressed in the appropriate language). We use each simulator to construct training and testing data-sets containing $100{,}000$ and $2{,}000$ auctions respectively. For each auction, we construct the representation described in Sec. \[sec:hypothesis-space\], and train the auction design network above using Adam SGD [@kingma2015adam] with a learning rate of $10^{-5}$, mini-batches of size $256$, and for $250{,}000$ iterations. In all experiments we set $\lambda_b = \lambda_r = 100$ (see Eq. \[eq:loss-function\]). After training, we use our held-out test set to report performance. The details of our auction simulators (details on the distributions we consider, how we construct bundles, and how we implement the allocation function for each bidding language) can be found in the Supplementary Material. In all experiments the number of objects for sale were as follows: with non-decreasing marginal utilities: 15 objects, with heterogeneous objects and unit-demand: 8 objects, and with hierarchical bundles: 8 component objects (resulting in 15 bundles). [**Baselines**]{} We consider four baselines when reporting our performance. 1) [*VCG auctions*]{}, the most commonly used Groves mechanism: a truthful, efficient, weakly budget balanced and individually rational auction. 2) [*Guo and Conitzer [@guo2010optimal]*]{} a provably optimal-in-expectation linear VCG redistribution mechanism, which requires $n<|K|$, analytical knowledge of $\rho$, and only handles multi-unit auctions. 3) [*Manisha et al. [@manisha2018learning]*]{} a VCG redistribution learned using a MLP architecture that requires $n<|K|$, does not support hierarchical bundles, and only works with unit-demand valuations. 4) [*MLP based architecture*]{} lastly, we compare to a 2-layer, 128-hidden-unit MLP that operates on a flattened version of the same data to empirically support our choice of representation and architecture. Results ======= Setting considered Guo et al. [@guo2010optimal] Manisha et al. [@manisha2018learning] G-CNN (ours) R-CNN (ours) ------------------------------- ------------------------------ --------------------------------------- -------------- -------------- Arbitrary distribution NO YES YES YES No knowledge of dist. NO YES YES YES Arbitrary \# of participants NO NO YES YES Varying \# of participants NO NO YES YES Multi-unit auctions YES NO YES YES Unit-demand auctions NO YES YES YES Hierarchical bundles auctions NO NO YES YES Guarantees no-deficit YES NO NO NO Guarantees indiv. rationality YES YES NO YES : . The method we propose here can be applied in more general settings than previously proposed alternatives. Models: G-CNN: learns a Groves payment rule directly using our data representation and network architecture. R-CNN: learns a VCG redistribution payment rule using our data representation and network architecture.[]{data-label="tab:comparison"} [**Qualitative comparison with alternative methods**]{} We start with a qualitative comparison with two existing alternative methods to automatically construct VCG redistribution mechanisms (see Sec. \[sec:related-work\]), and highlight how our method can be applied in more general settings in Tab. \[tab:comparison\]. A quantitative comparison with these two methods (in the settings in which they can be applied) shows how our methods also leads to better performance in practice (see the Supplementary Material). Importantly, while our method does not guarantee we will find auctions that are weakly budget balanced and individually rational, our quantitative result show that, in practice, we find zero, or next-to-zero violations of these constraints (see next paragraph). [0.49]{} ![. We report performance as the average reduction in payments collected relative to the VCG mechanism (higher is better). Displayed: average performance across $2,000$ auctions sampled from $\rho$, mean and standard deviation across $5$ training seeds. For each choice of model and bidding language we also report the fraction of auctions that resulted in a deficit (i.e. the mechanism had to be subsidized) (lower is better). The right panel shows interpolation to a [**previously unseen**]{} number of participants (note that MLP-based models do not support this, so their performance is not reported). Models: G-CNN: learn a groves payment rule directly using our data-representation and network architecture ([**ours**]{}). G-MLP: learn a groves payment rule directly using a MLP. R-CNN: Learning a VCG redistribution mechanism using our network architecture ([**ours**]{}). R-MLP: Learning a VCG redistribution mechanism using a MLP.[]{data-label="fig:quant-results"}](auction-bars-experiment1-all.pdf "fig:"){width="\linewidth"} [0.49]{} ![. We report performance as the average reduction in payments collected relative to the VCG mechanism (higher is better). Displayed: average performance across $2,000$ auctions sampled from $\rho$, mean and standard deviation across $5$ training seeds. For each choice of model and bidding language we also report the fraction of auctions that resulted in a deficit (i.e. the mechanism had to be subsidized) (lower is better). The right panel shows interpolation to a [**previously unseen**]{} number of participants (note that MLP-based models do not support this, so their performance is not reported). Models: G-CNN: learn a groves payment rule directly using our data-representation and network architecture ([**ours**]{}). G-MLP: learn a groves payment rule directly using a MLP. R-CNN: Learning a VCG redistribution mechanism using our network architecture ([**ours**]{}). R-MLP: Learning a VCG redistribution mechanism using a MLP.[]{data-label="fig:quant-results"}](auction-bars-experiment2-all.pdf "fig:"){width="\linewidth"} [**Quantitative results**]{} We illustrate quantitative results on synthetic auction data-sets in Fig. \[fig:quant-results\]. Our experiments show that auctions learned using our data representation and network architecture result in a significantly smaller economic burden on the participants than using VCG, and crucially, that we are able to learn auction rules with zero or next to zero violations of the weak budget balance constraint (i.e. mechanisms should operate without a subsidy). We highlight this by comparing our designs with auction rules based on MLP architectures trained on the same data. Fig. \[fig:quant-results\] shows results for valuations distributed as $\rho = \mathcal{N}(\mathcal{N}(10.0, 1.0), \mathcal{N}(2.0, 0.5))$ (where $\mathcal{N}(\mu, \sigma)$ is a Gaussian distribution with the appropriate parameters, clipped at $0$ from the left). We leave 2 further examples of distributions in the Supplementary Material. The left panel of Fig. \[fig:quant-results\] shows results on the three bidding languages we consider and for a fixed number of participants. Learning a Groves payment rule directly with our architecture and data representation (G-CNN) results in a reduction of payments collected, relative to VCG, by at least $60\%$, and in zero violations of the no-deficit constraint. Using our representation and architecture to learn a VCG redistribution mechanism (R-CNN) results in a higher percentage of the budget returned, and in next-to-zero violations of the weak budget balance constraint. In this case we are able to compare our architecture with MLP baselines which result in a relatively larger number of violations (some of which incur in egregious deficits of up of $50\%$ of the VCG budget, see Supplementary Material). Note how across the two design choices of learning a Groves payment rule or learning a redistribution mechanism, our CNN based architecture consistently results in fewer constraint violations. The right panel shows the case where the auction rule we learn is required to interpolate to a [**previously unseen number of participants**]{}. Again our data representation and network architecture result in a dramatic reduction of the economic burden placed on participants relative to VCG, and, when we learn a redistribution (R-CNN), in zero constraint violations. MLP baselines cannot operate on a variable number of participants so we are unable to show a comparison. In the Supplementary Material we report results from testing the same network on a varying number of participants. Note that since we strive to minimize payments, we find zero, or next to zero, violations of the individual rationality constraints in any of the models (i.e. participants never pay for a bundle more than they think it’s worth). Discussion of findings and conclusion {#sec:conclusions} ===================================== We investigated a machine-learning based approach to automated mechanism design and introduced the first truly general-purpose data representation, network architecture, and robust problem formulation to learn truthful and efficient auction rules automatically, given access only to a data-set of valuations. We introduced a novel way to represent auctions as a collection of “counter-factual” smaller auctions, and proposed a neural architecture that operates on this representation, to learn truthful and efficient mechanisms with minimal economic burden on the participants. Our methods can be applied on a wide variety of settings including arbitrary distributions, complex bidding languages and variable number of participants. Our empirical analysis shows how the resulting auctions collect only a small fraction of the VCG budget, and almost never require a subsidy. Mechanism design is a pillar of economics and social sciences and the domain of choice to study how a central authority can shape the incentives of self-interested individuals in pursuit of group metrics of success (e.g. elicit truthful reports and maximize social welfare). Nonetheless, very few attempts to apply machine learning ideas in this domain have been made. Here we show that under certain reasonable assumptions, the special case of auction design can be turned into a supervised learning problem and the modern tools of statistical learning and deep networks can be brought to bear. The recent renaissance of Artificial Intelligence points to a future where multiple artificial agents act in a shared environment to maximize individual rewards, realizing the vision of the *machina economicus* [@parkes2015economic]. In this context, it is paramount to investigate how to automatically translate high-level group-wide metrics of success, such as “social welfare maximization” and “truth-telling”, to individual-level incentive structures. This work is a first step in this direction, and builds heavily on the economics literature on the subject. Future efforts will focus on the extension of these ideas beyond auctions to more general decision problems. [^1]: One common assumption is that each agent only cares about the items allocated to them, that is $v_i = v\!\left(k\!\left(i\right)\right)$. [^2]: With a slight abuse of notation, we will sometimes drop $v_i$ and $\theta_i$’s explicit dependence on each allocation $k$, and simply denote with $v_i$ and $\theta_i$ the collections of preferences that can be held and expressed by player $i$. That is, given an arbitrary ordering of the choice set $K$, so that $K = [k_1, k_2, \ldots, k_{|K|}]$, we will use $v_i$ and $\theta_i$ to refer to the vectors $v_i = [v_i(k_1),\ldots,v_i(k_{|K|})]$, and $\theta_i = [\theta_i(k_1),\ldots,\theta_i(k_{|K|})]$. [^3]: For this reason, [@parkes2001thesis] summarizes the effect of the VCG payment rule as to “internalize the externality.” [^4]: This is referred to as a “redistribution” mechanism because it can be viewed as collecting the VCG payments and then “redistributing” some of them back to participants. [^5]: Note that this is reasonable given our choice of bidding languages. [^6]: This architecture is effectively a DeepSets network applied to a graph of $n-1$ nodes, and a single global output [@manzil2017deepsets; @battaglia2018relational]. The node functions are our CNN+MLP and the aggregator function is a sum.
{ "pile_set_name": "ArXiv" }
--- abstract: | The guarding game is a game in which several cops try to guard a region in a (directed or undirected) graph against Robber. Robber and the cops are placed on the vertices of the graph; they take turns in moving to adjacent vertices (or staying), cops inside the guarded region, Robber on the remaining vertices (the robber-region). The goal of Robber is to enter the guarded region at a vertex with no cop on it. The problem is to determine whether for a given graph and given number of cops the cops are able to prevent Robber from entering the guarded region. Fomin et al. \[Fomin, Golovach, Hall, Mihalák, Vicari, Widmayer: How to Guard a Graph? Algorithmica 61(4), 839–856 (2011)\] proved that the problem is -complete when the robber-region is restricted to a tree. Further they prove that is it -complete when the robber-region is restricted to a directed acyclic graph, and they ask about the problem complexity for arbitrary graphs. In this paper we prove that the problem is -complete for arbitrary directed graphs. game, cops and robber game, graph guarding game, computational complexity, E-completeness, games on propositional formulas author: - 'R. Šámal [^1]' - 'T. Valla [^2]' title: 'The guarding game is -complete' --- Introduction and motivation =========================== The *guarding game* $(G,V_C,c)$, introduced by Fomin et al. [@HGG], is played on a graph $G=(V,E)$ (or directed graph $\dir G=(V,E)$) by two players, the *cop-player* and the *robber-player*, each having his pawns ($c$ cops and one Robber, respectively) on $V$. There is a protected region (also called cop-region) $V_C\subset V$. The remaining region $V\setminus V_C$ is called robber-region and denoted $V_R$. Robber aims to enter $V_C$ by a move to a vertex of $V_C$ with no cop on it. The cops try to prevent this. The game is played in alternating turns. In the first turn the robber-player places Robber on some vertex of $V_R$. In the second turn the cop-player places his $c$ cops on vertices of $V_C$ (more cops can share one vertex). In each subsequent turn the respective player can move each of his pawns to a neighbouring vertex of the pawn’s position (or leave it where it is). However, the cops can move only inside $V_C$ and Robber can move only on vertices with no cops. At any time of the game both players know the positions of all pawns. The robber-player wins if he is able to move Robber to some vertex of $V_C$ in a finite number of steps. The cop-player wins if the cop-player can prevent the robber-player from placing Robber on a vertex in $V_C$ indefinitely. Note that $2|V|^{c+1}$ is the upper bound on the number of all possible positions of Robber and all cops, so after that many turns the position has to repeat. Thus, if Robber can win, he can win in less than $2|V|^{c+1}$ turns. Consequently, we may define Robber to lose if he does not win in $2|V|^{c+1}$ turns. Without loss of generality, we may assume that Robber never passes his move. For a given graph $G$ and guarded region $V_C$, the task is to find the minimum number $c$ such that cop-player wins. Note that this problem is polynomially equivalent with the problem of determining the outcome of the game for a fixed number $c$ of cops. The guarding game is a member of a big class called pursuit-evasion games, see, e.g., Alspach [@Als] or Bonato and Nowakowski [@BoNo] for introduction and survey. Pursuit-evasion games is a family of problems in which one group attempts to track down members of another group in some environment. Probably the best known variant of these games is called the Cops-and-Robber game. In this game a graph $G$ is given, together with the number $c$ of cops and one robber. The cops are placed at the vertices in the first turn, then the robber is placed in the second turn. In alternating turns, the cops may move to a neighbouring vertex, and the robber may move to a neighbouring vertex. The question is whether the cops have a strategy to capture the robber, which means at some point enter a vertex with the robber. The Cops-and-Robber game was first defined for one cop by Winkler and Nowakowski [@WiNo] and by Quilliot [@Qui]. Aigner and Fromme [@AigFro] initiated the study of the problem with several cops. The minimum number of cops required to capture the robber in a given graph is called the cop number of the graph. The guarding game is thus a natural variant of the Cops-and-Robber game. The complexity of the decision problem related to the Cops-and-Robber game was studied by Goldstein and Reingold [@GoRe]. They have shown that if the number of cops is not fixed and if either the graph is directed or initial positions are given, then the problem is -complete. Also, recently Mamino [@Mam] has shown that the Cops-and-Robber decision problem without any further restriction is -hard. Very recently, Kinnersley [@Kinn] published a proof that the computational complexity of Cops-and-Robber decision problem without further restrictions is -complete. Unfortunately, the previous papers [@GoRe; @Kinn] use a misleading notion, using  instead of . Another interesting variant is the “fast robber” game, which is studied in Fomin et al. [@FastR]. In this version, the robber may move $s$-times faster than the cops. The authors prove that the problem of computing the minimum number of cops such that the fast Cops-and-Robber game is cop-win, is -complete and the parameterised version (where the parameter is the number of cops) is -complete. See the annotated bibliography [@FoThi] for reference on further topics. A different well-studied problem, the *Eternal Domination* problem (also known as *Eternal Security*) is strongly related to the guarding game. The objective in the Eternal Domination, introduced by Burger et al. [@ED2], is to place the minimum number of guards on the vertices of a graph $G$ such that the guards can protect the vertices of $G$ from an infinite sequence of attacks. In response to an attack of an unguarded vertex $v$, at least one guard must move to $v$ and the other guards can either stay put, or move to adjacent vertices. The minimum number of guards needed to protect the vertices of $G$ is denoted by $\gamma_\infty(G)$. The Eternal Domination problem is in fact a special case of the guarding game. This can be seen as follows. Let $G$ be a graph on $n$ vertices and we construct a graph $H$ from $G$ by adding a clique $K_n$ on $n$ vertices and connecting the clique and $G$ by $n$ edges which form a perfect matching. The cop-region is $V(G)$ and the robber-region is $V(K_n)$. Now $G$ has an eternal dominating set of size $k$ if and only if $k$ cops can guard $V(G)$. Several variants of the Eternal Domination are studied and various results are known. For example in [@ED1], the variant when only one guard may move in a turn is studied, and it is shown that all graphs not having $K_4$ as a minor satisfy $\gamma^1_\infty(G)=\theta(G)$ where $\theta(G)$ denotes the clique covering number of $G$ and $\gamma^1_\infty(G)$ denotes the minimum number of guards needed to protect $G$ when only one guard may move in a turn (so called *eternal $1$-security*). Also, for the eternal 1-security, $\alpha(G)\le\gamma^1_\infty(G)\le \theta(G)$ holds for every graph $G$ where $\alpha(G)$ is the independence number, and if $\alpha(G)=2$ then $\gamma^1_\infty(G)\le 3$ [@ED3]. Klostermeyer and MacGillivray [@ED6] proved that $\gamma^1_\infty(G) \le \alpha(G)(\alpha(G)+1)/2$. In [@ED4] it is shown that there are graphs $G$ for which $\gamma^1_\infty(G)\ge{\alpha(G)+1\choose 2}$. In [@ED3] the following conjecture is posed: if $\alpha(G)=\gamma^1_\infty(G)$ for a graph $G$, then the $\alpha(G)$ is equal to the chromatic number of the complement of the graph. This conjecture was proved in [@ED7] for graphs with $\alpha(G)=2$. Various complexity-related question were also studied. For example, in [@ED5] it is shown that the decision problem of the eternal 1-security is -$\NP^\NP$-hard (in fact, there is a completeness proof in [@ED5] but later an error in the proof was found). Further graph-theory and algorithmic issues are discussed in [@ED8]: For the eternal 1-security, the relationship between the number of guards and the independence and clique covering numbers of the graph is studied, results concerning which triples of these parameters can be attained by some graph are given, and the exact value of the number of guards for graphs in certain classes is determined. For the Eternal Domination, a linear algorithm to determine the minimum number of guards necessary to defend a tree is given in [@ED8]. In our paper we focus on the complexity issues of the following decision problem: Given the guarding game $\G=(G,V_C,c)$, who has the winning strategy? Let us define the computational problem precisely. In measuring the computational complexity, our model is the deterministic one-tape Turing Machine. The *guarding decision problem* () is, given a guarding game $(\dir G,V_C,c)$ where $\dir G$ is a directed graph, to decide whether it is a cop-win game or a robber-win game. More precisely, we specify a (binary) encoding of the graph $\dir G$, subset $V_C$ and starting positions; the language  then consists of all words that encode a configuration winning for the robber-player. We may also define the *guarding problem*, which is, given a directed or undirected graph $G$ and a cop-region $V_C\subseteq V(G)$, to compute the minimum number $c$ such that the $(G,V_C,c)$ is a cop-win. It is easy to see, that the guarding problem can be in polynomial time reduced to  by trying all possible values of $c$.  was introduced and studied by Fomin et al. [@HGG]. The computational complexity of  depends heavily on the chosen restrictions on the graph $G$. In particular, in [@HGG] the authors show that if Robber’s region is only a path, then the problem can be solved in polynomial time, and when Robber moves in a tree (or even in a star), then the problem is -complete. Furthermore, if Robber is moving in a directed acyclic graph, the problem becomes -complete. Later Fomin, Golovach and Lokshtanov [@Intruder] studied the *reverse guarding game* with the same rules as in the guarding game, except that the cop-player plays first. They proved in [@Intruder] that the related decision problem is -hard on undirected graphs. Nagamochi [@Naga] has also shown that that the problem is -complete even if $V_R$ induces a 3-star and that the problem is polynomially solvable if $V_R$ induces a cycle. Also, Thirumala Reddy, Sai Krishna and Pandu Rangan proved [@TSP] that if the robber-region is an arbitrary undirected graph, then the decision problem is -hard. Fomin et al. [@HGG] asked the following question. \[q:fomin\] Is  -complete? Let us consider the class $\E=\DTIME(2^{O(n)})$ of languages recognisable by a deterministic Turing machine in time $2^{O(n)}$, where $n$ is the input size. In pursuit of Question \[q:fomin\] we prove the following result. \[t:main\]  is -complete under log-space reductions. Furthermore, $\E\logred\Guard$ via length order $n(\log n)^2$. We defer the precise technical definitions of log-space reductions and length order to the end of Section \[s:mainproof\]. Note that the log-space reduction via certain length order is a more subtle definition of reduction than the general log-space reduction. The motivation is as follows. The reduction $\mathcal{L}\logred B$ via length order $\ell(n)$ means that given some instance $A\in L\in \mathcal{L}$, we obtain an instance $B$ equivalent to $A$, which size has increased to $\ell(|A|)$ (where $|A|$ is measured in bits). Therefore, informally, given a solver $S$ for $L$ working in time $2^{O(n)}$, this proves that the time needed to solve $B$ is at least $2^{O(\ell^{-1}(n))}$ where $\ell^{-1}$ is the inverse function to $\ell$. Immediately, we get the following corollary. \[c:main\] The guarding problem is -complete under log-space reductions. Using the time-hierarchy theorem, we get the following consequence (for the proof see Section \[s:mainproof\]). \[c:maintime\] There is a constant $c>1$ so that if a deterministic one-tape Turing machine decides  within time $t(n)$, then $t(n) > c^{n/(\log n)^2}$ for infinitely many $n$. Let us explain here the relevance of Theorem \[t:main\] to Question \[q:fomin\]. Very little is known about how the class  is related to . It is only known [@Book] that $\E\neq \PSPACE$. The following corollary shows that positive answer to Question \[q:fomin\] would give a relation between these two complexity classes. This gives unexpected and strong incentive to find positive answer to Question \[q:fomin\]. (On the other hand, to the skeptics among us, it may also indicate that negative answer is more likely.) If the answer to Question \[q:fomin\] is yes, then $\E \subseteq \PSPACE$. Suppose the guarding problem is -complete. Let $L\in \E$. Then (by Theorem \[t:main\]) an instance of $L$ can be reduced by a log-space reduction to an instance of the guarding game, which we suppose to be in . Consequently, $L \in \PSPACE$. Let us mention here, that the distance between  and  is rather small in the realm of guarding games: The game, that we will construct to show -hardness, has the property that after removing a single edge, the robber-region becomes a directed acyclic graph, and for such, the decision problem is in  by a result of Fomin et al. [@HGG]. In a sense, we show analogous result for the guarding game as Goldstein and Reingold [@GoRe] have shown for the original Cops-and-Robber game. The -completeness proof ======================= An algorithm in ---------------- In order to prove -completeness of , we first note that the problem is in . We define the *guarding game with prescribed starting positions* $\G=(G,V_C,c,S,r)$, where $S:\{1,\dots,c\}\to V_C$ is the initial placement of cops and $r\in V_R$ is the initial placement Robber, and we consider the guarding game $(G,V_C,c)$ after the two initial turns where the cops were placed as described by $S$ and Robber was placed to $r$. The *guarding decision problem with prescribed starting positions* () is, given a guarding game with prescribed starting positions $(\dir G,V_C,c,S,r)$ where $\dir G$ is a directed graph, to decide whether it is a cop-win game or a robber-win game. \[l:etime\] The guarding decision problem is in . We need to show that there is an algorithm deciding the outcome of a given guarding game $\G=(G,V_C,c)$ in $2^{O(n)}$ time, where $n$ is the size of the input $\G$ in some encoding. Let us first consider the normal (i.e., not reverse) version of the problem. Consider the directed graph $H$ of all configurations of the game $\G$ where edges mark the legal moves of $\G$. We use the standard backwards-labelling algorithm: first we label all final states of the game, then all states that can be finished by a single move, etc. In this way we can decide the outcome of every configuration in time polynomial in the size of the graph $H$. It remains to show that the size of $H$ is $2^{O(n)}$. As mentioned in the introduction, the simplest upper bound on $|V(H)|$ is $2|V(G)|^{c+1}$, which is unfortunately super-exponential in $n$ if $c$ is close to $n$. To find a better upper bound, we use the fact that the cops are mutually indistinguishable. There are at most $|V(G)|$ positions of Robber. Counting the number of all positions of the cops is the classical problem of putting $c$ indistinguishable balls into $|V_C|$ bins. Then, taking into account also whose turn it is and the number of vertices $r_i$, we get that $|V(H)|$ is bounded by $$|V(H)|\le 4|V(G)|{|V_C|+c-1\choose c}\le 4n{n+c-1\choose c} \le 4n2^{n+c-1} = 2^{O(n)}.$$ Thus the total size of $H$ is $2^{O(n)}$ as well. The proof for the reverse version of the game is analogous. Reduction from a formula-satisfying game ---------------------------------------- Let us first study the problem after the second move, where both players have already placed their pawns. We reduce  from the following formula-satisfying game which was described by Stockmeyer and Chandra [@StoCha] (game $G_2$ in Section 3 of [@StoCha]). A position in the formula-satisfying game configuration is a $4$-tuple $\F=(\tau, F_R(C,R), F_C(C,R), \alpha)$ where $\tau\in\{1,2\}$, $F_R$ and $F_C$ are formulas in 12-DNF both defined on set of variables $C\cup R$, where $C$ and $R$ are disjoint and $\alpha$ is the current $(C\cup R)$-assignment. The symbol $\tau$ serves only to differentiate the positions where the first or the second player is to move. Player I (II) moves by changing the values assigned to at most one variable in $R$ ($C$); either player may pass since changing no variable amounts to a “pass”. Player I (II) wins if the formula $F_R$ ($F_C$) is true after some move of player I (II). More precisely, player I can move from $(1,F_R,F_C,\alpha)$ to $(2,F_R,F_C,\alpha')$ in one move if and only if $\alpha'$ differs from $\alpha$ in the assignment given to at most one variable in $R$ and $F_C$ is false under the assignment $\alpha$; the moves of player II are defined symmetrically. Let  denote the language of all binary encodings of formula-satisfying game configurations which are winning for player I. Note, that if the game continues infinitely, then the starting position is not in , as player I did not win. We may as well consider the game won by player II, which is consistent with the rules of the cop-game. The following result is known. \[t:stocha\]  is an -complete language under log-space reduction. Moreover, $\E\logred\FormSAT$ via length order $n\log n$. Let us first informally sketch the reduction from $\F$ to $\G$, i.e., simulating $\F$ by an equivalent guarding game $\G$. The setting of variables is represented by positions of certain cops so that only one of these cops may move at a time (otherwise the cop-player loses the game), we call this device a Variable Cell. The variables in $C$ are under the control of the cop-player. The variables in $R$ are under the control of the robber-player (even though they are also represented by cops). In the following text, we construct a lot of gadgets. First, we precisely define them and then we informally describe how particular gadgets work, together with prerequisites needed for the gadget to work properly, and with the desired outcome of each gadget. Then we describe in detail the whole construction and how all gadgets are connected together. Finally, we precisely prove that the gadgets (and the whole game) behave as we described earlier. When describing the features of various gadgets, we will often use the term *binding scenario*. By binding scenario of a certain gadget (or even the whole game) we mean a flow of the game that both players must follow: if a player deviates from it, he will lose the game. The graph will be constructed in such a way, that if the players follow the binding scenario, they will simulate the formula game $\F$. Let us define the binding scenario more precisely. Let $\G=(G,V_C,c,S,r)$ be an instance of the guarding game with prescribed positions. Let $Q=(r_1,S_1,r_2,S_2,\dots,r_k,S_k)$ be a nonempty sequence such that $r_i\to r_{i+1}$ is a valid move of Robber and $S_i \to S_{i+1}$ is a valid move of cops for every $i$. We call such $Q$ a *move sequence*. A set $\F$ of move sequences is called a *scenario*. We say that players *follow the scenario $\F$* if they make a move sequence $Q\in\F$. We say that the player $P$ *deviates from the scenario $\F$* if $P$ makes a move not permitted by $\F$, but until his move the players did follow $\F$. We call a scenario $\F$ a *binding scenario* when $\F$ satisfies the following condition: If the cop-player deviates from $\F$, he will lose in at most ten turns. If the robber-player deviates from $\F$, he ends in a losing state of the game. When describing a particular gadget, we also describe its scenario and prerequisites which must be satisfied in order for the gadget to work properly. Later, after we complete the whole construction, we prove that the presented scenarios are in fact binding, if the prerequisites are satisfied. There are four cyclically repeating phases of the game, determined by the current position of Robber. The binding scenario is as follows: Robber cyclically goes through the following phases marked by four special vertices and in different phases he can enter certain gadgets. (Robber is allowed to pass. However, as we will show, it is never to his advantage.) 1. “Robber Move” ($\var{RM}$): In this step Robber can enter a *Robber Switch* gadget, allowing him to change setting of at most one variable in $R$. The values of formula variables are represented by positions of special cops in gadgets called Variable Cells. 2. “Robber Test” ($\var{RT}$): In this step Robber may pass through the *Robber Gate* into the protected region $V_C$, provided that the formula $F_R$ is satisfied under the current setting of variables. 3. “Cop Move” ($\var{CM}$): In this step the cop player can change one variable in $C$. This is realized by a gadget called *Cop Switch*. At the end of this phase Robber ‘waits’ on vertex $\var{RW}$ for the cop to finish setting the variable. 4. “Cop Test” ($CT$): In this step, if the formula $F_C$ is satisfied under the current setting of variables, the cops are able to block the entrance to the protected region forever (by temporarily leaving the *Cop Gate* gadget unguarded and sending a cop to block the entrance to $V_C$ that is provided by the Robber Gate gadgets). To begin with the construction, we start with the basic five vertices organised to a directed cycle $(\var{RM},\var{RT},\var{CM},\var{RW},CT)$, we call it the *Basic Cycle*. The Variable Cells ------------------ To maintain the current setting of a variable $x$, we use the Variable Cell gadget, there will be a separate Variable Cell for every variable. For a variable $x\in C\cup R$ we introduce a *Variable Cell* $\VAR(x)$, which is a directed cycle $(T_x,\var{FT}_x,F_x,\var{TF}_x)$ (see Fig. \[f:varcell\]). The informal description of the function of $\VAR(x)$ is as follows. There is one cop (*variable cop*) located in every $\VAR(x)$ and the position of the cop on vertices $T_x$, $F_x$ represents the boolean values true and false, respectively. The prescribed starting position of the variable cop is $T_x$ if $\alpha(x)$ is true, and $F_x$ otherwise. All the vertices of $\VAR(x)$ belong to $V_C$. Forcing and blocking {#s:forceblock} -------------------- To make the variable cops truly represent the boolean values and allow only the changes of these values that follow the rules of the formula game, we introduce two gadgets that will be used repeatedly. We say that we *block* a set $S \subseteq V_C$ by a vertex $u$, when we add to our graph the *Blocker* gadget $\BLOCK(u,S)$. The Blocker $\BLOCK(u,S)$ consists of a directed path $(u, p_{u,S}, q_{u,S})$ together with edges $(s,q_{u,S})$ for $s \in S$. The vertex $p_{u,S}$ is in $V_R$, while $q_{u,S}$ is in $V_C$. We will call $q_{u,S}$ the *entry vertex* of the blocker. See Fig. \[f:blocking\] for illustration. Suppose we blocked $S$ by $u$ and the only edges coming to $q_{u,S}$ start in $S$. If Robber is at $u$, then we must have at least one cop in the set $S$ or on $q_{u,S}$. We prove the properties of the Blocker in Lemma \[l:blocking\]. A way to prevent cops from moving arbitrarily is called *forcing* and we will now describe it. Consider $S \subset V_R$ (we will apply this for $S$ consisting of all the “normal” positions of Robber) and $V \subset V_C$. We say that we *force $V$ by $S$*, when we add the gadget $\FORCE(S,V)$ in Fig. \[f:forcing\]: it consists of a vertex $f_{S,V}$, together with edges from each vertex in $S$ to it. Further, for any $v \in V$ we add a directed path $(f_{S,V}, \bar v, v)$. All of the new vertices $f_{S,V}$ and $\bar v$ are added to the robber-region $V_R$. We describe here shortly the desired behaviour. We say a vertex $u \in V_C$ is *guarded* if there is a cop that can reach $u$ in at most two moves. The forcing gadget ensures that before each Robber’s move, every vertex of $V$ is guarded. Later, we will apply the gadget $\FORCE(S,V)$ for $V = \{F_x\}$ and $V = \{T_x\}$ (and for one more similar pair). The desired outcome is that the variable cop is forced to stay on $T_x$ or $F_x$, unless another cop helps to guard these two vertices. (We could ensure the same behaviour using blocker gadgets, but it would lead to a larger graph – one with quadratically many edges – thus we would obtain weaker result for the running time of algorithms that solve .) The properties of forcing are precisely described and proved in Lemma \[l:forcing\]. The Robber Gate to $V_C$ ------------------------ For every clause $\varphi$ of $F_R$, we construct Robber Gate gadget $\ROBG(\varphi)$. The goal of this gadget is that it allows Robber to enter $V_C$ if and only if $\varphi$ is satisfied by the current setting of variables. Consider a clause $\varphi=(\ell_1 \& \dots \& \ell_{12})$ where each $\ell_i$ is a literal. The Robber Gate $\ROBG(\varphi)$ is the Blocker $\BLOCK(\var{RT}, V)$ with $$V = \{T_x \mid \hbox{some $l_i$ is $\neg x$}\} \cup \{F_x \mid \hbox{some $l_i$ is $x$}\}$$ For easier notation, we shall use shorter notation $z'_\varphi := p_{\var{RT},V}$ and $z_\varphi := q_{\var{RT},V}$. See Fig. \[f:robgate\] for illustration. The vertices $z'_\varphi$ and $\var{RT}$ belong to $V_R$, $z_\varphi$ belongs to $V_C$. Robber stands on $\var{RT}$ and there is exactly one cop in each $\VAR(x)$, $x\in\varphi$, standing either on $T_x$ or $F_x$. No other cop can access $z_\varphi$ in one move. It is the Robber’s turn. Robber can reach $z_\varphi$ (thus winning the game) if and only if $\varphi$ is satisfied under the current setting of variables (given by the positions of cops on Variable Cells) *and* the cop-player did not win before. Otherwise, Robber moves to $\var{CM}$. We remark that if cop-player has won before (by satisfying his formulas in the formula-game – this corresponds to moving one of his pawns to vertex $a''$ described in the next section) then there will be a cop that can reach every $z_\varphi$ in a single move. This renders the Robber Gate unfunctional. The properties of the Robber Gate are proved in Lemma \[l:robbergate\]. The Cop Gate ------------ For every clause $\psi$ of $F_C$, we will use the *Cop Gate* gadget $\COPG(\psi)$ (see Fig. \[f:copgate\]). If (and only if) $\psi$ is satisfied, $\COPG(\psi)$ allows cops to win the game. According to the rules, the game will still continue infinitely long, but Robber will never be able to enter $V_C$. This is achieved by placing a cop to a vertex, from which he controls all vertices $z_\varphi$ in the Robber Gates. The Cop Gate $\COPG(\psi)$ contains a directed cycle $(a_\psi,a'_\psi,a'',a'''_\psi)$. Further, $\COPG(\psi)$ contains edges $(a'', z_\varphi)$ for every clause $\varphi$ of the Robber’s formula $F_R$. (These edges are common for all clauses $\psi$.) Note, that the vertex $a''$ does not have the subscript $\psi$ – this vertex is shared among all clauses $\psi$. The reason for this is to keep the size of the constructed graph $\dir G$ small enough. Let $\psi=(\ell_1 \& \dots \& \ell_{12})$ where each $\ell_i$ is a literal. We add 12 blockers to our gadget, one for every literal:\ If $\ell_i=x$ then we add the blocker $\BLOCK(CT, \{T_x, a_\psi, a''\})$ to $\COPG(\psi)$.\ If $\ell_i=\neg x$ then we add the blocker $\BLOCK(CT, \{F_x, a_\psi, a''\})$ instead. The directed 4-cycle belongs to $V_C$, the blocker is as described above. There is one cop at the vertex $a_\psi$ (we call him Arnold) and there is exactly one cop in each $\VAR(x)$, $x\in\psi$, standing on either $T_x$ or $F_x$. Robber is at the vertex $CT$ and no other cop can access $\COPG(\psi)$ in less than three moves. It is the cop’s turn. Arnold is able to move to $a''$ (and therefore block all the entrances $z_\varphi$ forever) without permitting Robber to enter $V_C$ if and only if $\psi$ is satisfied under the current setting of variables (given by the position of cops in the Variable Cells). Note that while Arnold will never move voluntarily back from $a''$ to $a_\psi$, Robber may force him using the forcing gadget that we will add in Section \[s:bigpicture\] in step \[item:force\] of the construction. Therefore, the vertex $a'''_\psi$ is needed. The properties of the Cop Gate are proved in Lemma \[l:copgate\]. The Commander gadget {#s:commander} -------------------- When changing variables of $R \cup C$, we have to make sure that at most one variable is changed at a time. We already prevented all variable cops from moving at all by means of forcing, see Section \[s:forceblock\] and Lemma \[l:forcing\]. Next we describe a gadget that allows one of them to move: the gadget Commander (see Fig. \[f:commander\]). It consists of vertices $\{G_x, GF_x, GT_x \mid x\in R \cup C\}\cup\{\var{HQ}\}$. For each $x \in R \cup C$ there is directed 2-cycle $(\var{HQ},G_x)$ and directed paths $(G_x, GF_x, F_x)$ and $(G_x, GT_x, T_x)$. All of the vertices belong to $V_C$. There is one cop, Commander, whose prescribed starting position is the vertex $\var{HQ}$. In Section \[s:bigpicture\] we will add some extra edges that will force Commander to stay at $\var{HQ}$ most of the time. The desired function of $\COM$ is as follows: in order to move a variable cop from $T_x$ to $F_x$ or vice versa, Commander moves to $G_x$ to temporarily guard vertices $T_x$ and $F_x$. This will happen in two stages of the game: when Robber sets variables, he will pick a particular $x \in R$ and make Commander move to $G_x$. When the cop-player moves, Robber will only ensure that Commander moves to a $G_x$ with $x \in C$. Let $y\in R$. We describe a gadget that allows Robber to change the value of the variable $y$, i.e., to force the variable cop in $\VAR(y)$ to move from $T_y$ to $F_y$ or vice versa. The Switch of Robber’s variable $y$ ($\ROBS(y)$) consists of a directed path $(\var{RM},Sw_y,\var{RT})$, edges $(\var{RM}, \var{HQ})$, $(\var{RT}, \var{HQ})$, $(Sw_y,G_y)$, $(\var{RM},\var{RT})$ and a Blocker $\BLOCK(Sw_y, \{\var{FT}_y, \var{TF}_y\})$ (see Fig. \[f:switch\]). The new vertices $Sw_y$ belong to $V_R$, the Blocker is described earlier. Robber is at $\var{RM}$, the cop in $\VAR(y)$ is either on $T_y$ or $F_y$, Commander is at $\var{HQ}$ and no other cop can access any vertex of $\VAR(y)$. It is the Robber’s turn. By entering the vertex $Sw_y$ Robber forces Commander to move to $G_y$, and the variable cop to enter $\var{FT}_y$ or $\var{TF}_y$. Robber then enters the vertex $\var{RT}$, the variable cop finishes his move to $T_y$ (or $F_y$) and at the same time Commander returns to $\var{HQ}$. If Robber decides not to change any variable, he may go directly to $\var{RT}$; in such case the cops do not move in their next turn. The properties of the Robber Switch are proved in Lemma \[l:robs\]. The $\COPS$ gadget consists of a Blocker $\BLOCK(\var{CM},\{G_x \mid x \in C\}$. Robber is on $\var{CM}$, all variable cops are at $T_x$ or $F_x$ and Commander at $\var{HQ}$. It is the cop’s turn. The cop-player decides which variable $x \in C$ to change. The Commander moves to $G_x$, and the variable cop in $\VAR(x)$ moves to $\var{TF}_x$ or $\var{FT}_x$. Next, Robber moves to $\var{RW}$. Commander moves back to $\var{HQ}$ and the cop in $\VAR(x)$ finishes his move to $F_x$ or $T_x$. Finally, Robber moves to $CT$. Alternatively, if the cop-player does not want to change any variable, all variable-cops stay put, and just Commander moves to some $G_x$ and back to $\var{HQ}$. The big picture: putting the gadgets together {#s:bigpicture} --------------------------------------------- Let us consider an instance of the formula game $\F=(\tau, F_R(C,R), F_C(C,R), \alpha)$. We now proceed by putting all gadgets together in order to construct the instance $\G=(\dir G,V_C,c, S, r)$ of . When different gadgets contain a vertex with the same name, this means that this vertex is shared by these gadgets (this happens often for vertices $\var{RM}$, $\var{RT}$, $\var{CM}$, $CT$, $\var{HQ}$, $a''$ but also for others). To make the construction clearer, we define the following sets: $$\begin{aligned} \calT &= \{ T_x \mid x \in R \cup C\} \\ \calF &= \{ F_x \mid x \in R \cup C\} \\ \calA &= \{ a_\psi \mid \psi \in F_C \} \\\end{aligned}$$ The order of construction steps is as follows. 1. \[item:basiccycle\] Construct the Basic Cycle. 2. \[item:varcell\] For every variable $x\in C\cup R$ construct the Variable Cell gadget $\VAR(x)$. 3. \[item:robbergate\] For every clause $\varphi$ of $F_R$ construct the Robber Gate gadget $\ROBG(\varphi)$. 4. \[item:copgate\] For every clause $\psi$ of $F_C$ construct the Cop Gate gadget $\COPG(\psi)$. 5. \[item:commander\] Construct the Commander gadget $\COM$. 6. \[item:robs\] For every variable $y\in R$ construct the Robber Switch $\ROBS(y)$. 7. \[item:cops\] Construct the Cop Switch $\COPS$. 8. \[item:force\] Let $S_1 = \{ \var{RM}, \var{RT}, \var{CM}, \var{RW}\} \cup \{ Sw_y \mid y \in R \}$ and $S_2 = S_1 \cup \{ CT \}$. Add to the graph gadgets $\FORCE(S_2, \calT)$, $\FORCE(S_2, \calF)$, $\FORCE(S_1, \calA)$, and $\FORCE(S_1, \{a''\})$. We call the vertices $f_{S_2, \calT}$, $f_{S_2, \calF}$, $f_{S_1, \calA}$, and $f_{S_1, \{a''\}}$ of these gadgets the *forcing vertices*. 9. \[item:edges\] Add directed edges to $\var{HQ}$ from $\var{RM}$, $\var{RT}$, $\var{RW}$, $CT$. The prescribed starting position $r$ of Robber is the vertex $\var{RM}$. We define the starting positions $S$ of the cops as follows: - For each variable $x\in C\cup R$, the corresponding variable cop starts at the vertex $T_x$ if $\alpha(x)$ is true and at the vertex $F_x$ if $\alpha(x)$ is false. - Each $\psi$-Arnold starts at the corresponding vertex $a_\psi$. - The cop Commander starts at the vertex $\var{HQ}$. The number $c$ of cops is thus set as $c=|C\cup R|+|F_C|+1$, where $|F_C|$ denotes the number of clauses in $F_C$. We also briefly show that the construction of $\G$ based on $\F$ can be done in log-space. A log-space reduction is a reduction computable by a deterministic Turing machine (with read-only input tape, a working tape, and output tape) using logarithmic space. Conceptually, this means that while producing encoding of the instance $\G$ to the output tape the reducing Turing machine can keep (stored on the work tape) a constant number of pointers into the input tape along with a constant number of integers holding at most polynomially large values, and a logarithmic number of fixed size integers. \[l:logspace\] Let $\F$ be an instance of the formula game and let $\G$ be an instance of  constructed via the above construction from $\F$. Then the reducing Turing machine is log-space bounded and $|\G|=O(|\F|\log |F|)$ where the instance sizes are measured in bits. In order to show the log-space reducibility of our construction of $\G$, we need to observe that the whole construction can be performed by an algorithm reading the encoding of the formula game $\F$ and using only constant number of work variables holding polynomially large values. We go through the construction and briefly discuss the space needed for construction. - Step \[item:basiccycle\] of the construction produces a fixed-size gadget. - Steps \[item:varcell\], \[item:commander\], and \[item:robs\] can be realised by an iteration over the set of variables, each iteration producing a fixed-size gadget. Clearly, a constant number of log-space integers suffices. - Steps \[item:robbergate\] and \[item:copgate\] can be realised by two nested iterations over the set of clauses and the set of variables. Again, a constant number of log-space integers suffices. - Step \[item:force\] can be done by several iterations (over the set of variables and over the set of clauses). Important thing to note is that the gadget $\FORCE(S,V)$ has size (the number of vertices and edges) $O(|S|+|V|)$. The output of the cops and Robber starting positions can be done by an iteration over the set of clauses and variables. Therefore, only a constant number of log-space integers is needed during the reduction. Let $f$ be the number of clauses of $\F$ and let $p$ be the number of variables of $\F$. Then the number of bits of $\F$ is $\Omega(f\log p + p) = \Omega(f+p)$ as we need $\log p$ bits to encode the variable identifiers, each clause contains at most 12 literals and the initial assignment of variables takes $p$ bits. Next, we estimate the size of the instance $\G=(\dir G,V_C,c,S,r)$, based on our construction, starting with the number of vertices and edges of $\dir G$. Each $\VAR(y)$ has a constant size, all Variable Cells thus have total size $O(p)$. Each $\ROBG(\varphi)$ has a constant size, all Robber Gates thus have total size $O(f)$. The Cop Gate gadgets contain a shared part of size $O(f+p)$ and the non-shared part of each $\COPG(\psi)$ has a constant size. Therefore, all Cop Gate gadgets have total size $O(f+p)$. The size of the Commander gadget is clearly $O(p)$. Each $\ROBS(y)$ has a constant size, so the total size of all Robber Switches is $O(p)$. The size of $\COPS$ is clearly $O(p)$. In the last nontrivial step, step \[item:force\], $O(p+f)$ vertices and edges are added to $\dir G$. As $|V(\dir G)|+|E(\dir G)| \le O(p+f)$, we need $$O((p+f)\log(p+f)) = O(|\F| \log(|\F|) )$$ bits to encode $\dir G$. Note that we may get a linear bound on the size of $\G$, if we assume fairly reasonable condition on the formula game: all variables are used and no clause repeats within a formula. Proof of correctness -------------------- Let us start by a simple observation. While the rules allow Robber to pass (skip a move), we may assume that he never uses it. For if Robber does not move, the cops may do the same and repeating this indefinitely means a win for cops. From now on, we shall assume Robber never passes. Another assumption, that will be somewhat harder to prove: [1cm]{}[1cm]{} Assumption A: the cop-player restricts to strategy where either every $\psi$-Arnold is at his starting vertex $a_\psi$, or exactly one of them moves to $a'_\psi$, $a''$ (when Robber entered vertex $CT$), and possibly back to $a'''_\psi$ and $a_\psi$. In Theorem \[t:prescribedgame\] we will show that this restriction does not change the outcome of the game, it will make it easier to analyze, though. Before getting to the individual lemmas, we describe the main idea: if the robber-player succeeds in satisfying formula $F_R$, Robber will be able to enter the cop-region by means of Robber Gate. If the cop-player will be faster with satisfying his formula $F_C$, he will be able to send one Arnold to $a''$. The resulting state of game will be cop-win – there is an easy strategy (described in Lemma \[l:arnoldwin\]) that will make sure Robber will never enter the cop-region. Recall that we call a vertex $v \in V_C$ *guarded* if there is a cop that can reach $v$ in at most two moves. Further, a *set* $V \subseteq V_C$ is *guarded* if every vertex of $V$ is guarded and each one is guarded by a different cop. Formally, we require that there is a set $U$, $|U|=|V|$, of cops and a 1-1 mapping $m$ between $U$ and $V$ such that for each $c\in U$ the vertex $m(c)$ is guarded by the cop $c$ \[l:forcing\] Let $S$, $V$ be sets such that the gadget $\FORCE(S,V)$ is part of $\dir G$. Let us consider the game state where Robber is in $S$ and it is Robber’s turn. 1. Let there be an unguarded vertex of $V$. Then this is a robber-win state. 2. If the set $V$ is guarded and Robber enters the vertex $f_{S,V}$, the game state will be a cop-win. 3. If the set $V$ is not guarded and there is no vertex $x \in V_C$ that has more than one outneighbor in $V$ then this is a robber-win state. Suppose $v$ is unguarded. Then Robber moves along the path $(f_{S,V}, \bar v, v)$ and no cop can stop him. On the other hand, suppose the set $V$ is guarded and mapping $m$ is defined as above. If Robber decides to move to $f_{S,V}$, than each cop $c$ guarding vertex $m(c) \in V$ moves towards it (or stays in place, if they already are at it). Thus, the cops can reach all vertices of $V$ before Robber, and Robber is stuck in $f_{S,V}$ or some $\bar v$ for the rest of the game. For the third part, suppose that Robber moves to $f_{S,V}$ and the cops can somehow prevent him from winning. Then for every $v \in V$ the cops have a response to Robber’s move to $\bar v$, namely, there is a cop that can move to $v$ in one step, denote him by $d(v)$. Due to the extra condition in the third part, mapping $d$ is injective, its inverse is the required mapping $m$. We now define the notion of *normal positions* of certain pawns. \[def:normal\] We say that - Robber is in normal position if he is on $S_2$ (which was previously defined as $\{ \var{RM}, \var{RT}, \var{CM}, CT, \var{RW}\} \cup \{ Sw_y \mid y \in R \}$), - Commander is in normal position if he is on $\var{HQ}$ or on some $G_x$, - variable cops are in normal position if they are in some $\VAR(x)$, - $\psi$-Arnold is in normal position if he is at $a_\psi$ or $a''$. If one of the players leaves these positions, the game will be decided quickly. This is obvious for Robber, for if he is not in the normal position, he is no longer in the strongly connected part of the graph. So, he either reaches $V_C$ within two moves, or he will be unable to move for the rest of the game. In the following lemmas we deal with the rest of the pawns. \[l:HQ\] Consider a game state where Robber is in $S_2$ (see Definition \[def:normal\]) and it is the Robber’s turn. If Commander is not on $\var{HQ}$ or on some $G_x$, then this is a robber-win state. If Commander is elsewhere, he will not be able to reach $\var{HQ}$ again (by the construction). If Robber is in $S_2$, he will be able to get to $\var{RM}$ and from there to enter $\var{HQ}$. \[l:normalposvarcops\] Consider a game state where Robber is in $S_2$ (see Definition \[def:normal\]) and it is the Robber’s turn. If there is an $x \in R \cup C$ so that there is no cop on $F_x$, $T_x$, or $G_x$, then this is a robber-win state. Otherwise, if Robber enters vertex $f_{S_2,\calF}$ or $f_{S_2,\calT}$, he loses. By Lemma \[l:HQ\], Commander must stay on $\var{HQ}$ or some $G_x$. Consequently, there is at most one cop in each $\VAR(x)$ and no cop is at vertices $GF_x$, $GT_x$. Thus, to guard both $T_x$ and $F_x$, we need a cop at either $G_x$, $T_x$, or $F_x$ (Lemma \[l:forcing\]). The second statement follows again from Lemma \[l:forcing\], as no cop will be guarding two of the vertices $T_x$, neither two of the vertices $F_x$. \[l:normalposarnolds\] 1. Consider a game state where Robber is in $S_2$ (see Definition \[def:normal\]) and it is the Robber’s turn. Then all Arnolds are in $X=\bigcup_\psi \{a_\psi, a'_\psi, a'''_\psi\} \cup \{a''\}$, or it is a robber-win state. 2. If every $\psi$-Arnold is on $a_\psi$, with possible exception of one, who is at $a''$, then Robber may not enter vertices $f_{S_1,\calA}$, $f_{S_1,\{a''\}}$. 3. If Robber is in $S_1$, some Arnold in $a'_\psi$ and it is the Robber’s turn, then it is a robber-win state. Again, a simple consequence of Lemma \[l:forcing\]: there is no edge leading towards $X$, thus if some Arnold leaves $X$, there will be always less than $|\calA|$ cops in $X$. When (now or in the next move) Robber is in $S_1$, he can use the gadget $\FORCE(S_1, \calA)$ according to the third part of Lemma \[l:forcing\]. This proves the first part and the third part. By the second part of Lemma \[l:forcing\] follows the second part. \[l:blocking\] Let a set $S$ be blocked by a vertex $u$. (That is, the Blocker gadget $\BLOCK(u,S)$ is a part of the constructed graph $\dir G$.) Suppose that no other edges leading to $q_{u,S}$ are part of $\dir G$. Suppose Robber is on $u \in S_2$ and it is his turn to move. Then there must be at least one cop in the set $S$, otherwise the cop-player loses the game. If there is at least one cop in $S$, Robber may not enter $p_{u,S}$, otherwise he loses the game. By Lemma \[l:HQ\], \[l:normalposvarcops\] and \[l:normalposarnolds\], there is no cop in $q_{u,S}$ – all cops must be at vertices distinct from all vertices of form $q_{u,S}$. This means that Robber can reach this vertex in two moves, and the cops cannot stop him. On the other hand, if Robber moves to $p_{u,S}$ while there is a cop on $S$, then this cop moves to $q_{u,S}$, blocking Robber forever. \[l:beginning\] At the beginning of the game $\G$, the prerequisites of the Robber Switches $\ROBS(y)$, $y\in R$ are satisfied. By the construction, Robber is on $\var{RM}$ and the variable cops for each variable gadget $\VAR(x)$ are either on $T_x$ or $F_x$. Commander is on $\var{HQ}$. No other cop can ever reach $\VAR(x)$. \[l:robs\] Consider a game state satisfying the prerequisites of all Robber switches $\ROBS(y)$, $y\in R$. Then the scenarios of all $\ROBS(y)$, $y\in R$ are binding scenarios. Moreover, after execution of these binding scenarios, the prerequisites of all Robber Gates $\ROBG(\varphi)$, $\varphi\in F_R$ are satisfied, unless there is an Arnold at $a''$. If Robber moves to $Sw_y$, Commander must move to $G_y$, or Robber enters $V_C$ there. (By the construction, no other cop than Commander can be at $G_y$.) Also, by Lemma \[l:blocking\], the variable cop in $\VAR(y)$ must move to $\var{TF}_y$ or $\var{FT}_y$. By Lemma \[l:forcing\] and \[l:blocking\], Robber may not enter the forcing vertices, and thus continues to $\var{RT}$. After this, Commander must move back to $\var{HQ}$, thus the variable cop in $\VAR(y)$ to $T_y$ or $F_y$ (Lemma \[l:normalposvarcops\]). If Robber moves directly to $\var{RT}$, no cop may move (by Lemma \[l:forcing\] and \[l:blocking\]). \[l:robbergate\] Consider a game state satisfying the prerequisites of all Robber Gates $\ROBG(\varphi)$, $\varphi\in F_R$. Then the scenarios of all Robber Gates $\ROBG(\varphi)$, $\varphi\in F_R$ are binding scenarios. Moreover, after execution of these binding scenarios, the prerequisites of the Cop Switch are satisfied. If no other cop may access $z_\varphi$, Lemma \[l:blocking\] applies: Robber may enter $z_\varphi$ if and only if there is no cop in the set $V$ (see the definition of Robber Gates). By the construction, this happens precisely if $\varphi$ is satisfied by the current setting of the variables. If he may not enter $z_\varphi$, the only non-losing move for him is to move to $\var{CM}$. If the vertex $a''$ is occupied (the cop-player has “won” already), Robber may not enter $z_\varphi$ either and moves to $\var{CM}$. After the Robber’s move to $\var{CM}$, the prerequisites of the Cops Switch are satisfied, as the cops did not move yet. \[l:cops\] Let us consider a game state satisfying the prerequisites of the Cop Switch. Then the scenario of the Cop Switch is a binding scenario. Moreover, after execution of it, the prerequisites of all Cop Gates $\COPG(\psi)$, $\psi\in F_C$, are satisfied. When Robber stands on $\var{CM}$, Commander may leave $\var{HQ}$, as the edge $(CM,HQ)$ is not part of $\dir G$. In fact, Commander has to move to a vertex $G_x$ where $x \in C$ is a cop-player’s variable because of the Cop Switch gadget. If the cop-player decided to switch $x$, the variable cop in $\VAR(x)$ moves also in the first move to $\var{FT}_x$ or $\var{TF}_x$. Next, Robber moves to $\var{RW}$, after which Commander must return to $\var{HQ}$ and thus the cop in $\VAR(x)$ finishes his move to $T_x$ or $F_x$ (or keeps staying in place). Finally, Robber moves to $CT$, which leads to a game state satisfying all prerequisites of all Cop Gates. \[l:arnoldwin\] Consider a game state where $\psi$-Arnold is on the vertex $a''$ for some clause $\psi$, other cops are in normal positions, and Robber is in $S_2$ (see Definition \[def:normal\]). Then this is a cop-win state. As long as Robber is in normal position, the cop-player does not move with any of his cops until he is forced to: one Arnold stays at $a''$, the others at their initial positions – even if another clause in $F_C$ becomes satisfied. The variable-cops and Commander follow the scenarios of the Cop-Switch and Robber-Switch. Robber may force the Arnold at $a''$ to move back to his original position by using the Enforcer gadget $\FORCE(S_1,\calA)$; however, he will be trapped afterwards. It is easy to check that Robber may not enter the cop-region, as in binding scenarios he may do this only in some Robber Gate. But if he moves to some vertex $z'_\varphi$, the Arnold at $a''$ moves to $z_\varphi$, thus blocking Robber forever. \[l:copgate\] Consider a game state satisfying the prerequisites of all Cop Gates $\COPG(\psi)$, $\psi\in F_C$. Then the scenarios of all $\COPG(\psi)$, $\psi\in F_C$, are binding scenarios. Moreover, after execution of these binding scenarios, the prerequisites of all Robber Switches $\ROBS(y)$, $y\in R$, are satisfied. If $\psi$ is satisfied by the current setting of variables in the variable cells, then we do not need the $\psi$-Arnold to protect the entry vertex of any of the blockers, thus he may move to $a'_\psi$. If Robber enters some of the Blockers, then he loses by Lemma \[l:blocking\]. On the other hand, if $\psi$ is not satisfied, and $\psi$-Arnold leaves $a_\psi$, then the cop-player loses: again by Lemma \[l:blocking\]. Next, Robber moves to $\var{RM}$, which forces all vertices $a_\psi$, thus the $\psi$-Arnold either stays at $a_\psi$ or, if he left it already, finishes his move to $a''$. Therefore, the scenarios of the Cop Gates are binding. It is easy to see that the prerequisites of all Robber Switches are satisfied at the end of the binding scenario. Using Lemma \[l:arnoldwin\] we justify that if some Arnold moves to $a''$, it is a win for the cop-player. We have described the complete reduction from the formula game $\F$ to the guarding game $\G$ with prescribed starting positions, and characterised the properties of various stages of $\G$. It remains to show that both games have the identical outcome; we also list some properties of the constructed graph that will be useful later. \[t:prescribedgame\] For every instance of formula satisfying game $\F=(\tau, F_C, F_R, \alpha)$ there exists a guarding game $\G=(\dir G,V_C,c, S, r)$, $\dir G$ directed, with a prescribed starting positions such that player I wins $\F$ if and only if the robber-player wins the game $\G$. Moreover, the following properties are satisfied: 1. The size of the instance $\G$ is $|\G|=O(|\F| \log |\F|)$. 2. There is no oriented edge $e\in E(\dir G)$ with the endpoint in $r$. 3. There is at most one cop standing at each vertex $v\in V(\dir G)$. 4. Consider the set $B=\{u\in V_R \mid \exists v\in V_C,\, (u,v)\in E(\dir G)\}$ of the “border” vertices. The out-degree of each $u\in B$ is exactly 1. For every formula game $\F$ we have described in Section \[s:bigpicture\] a construction of guarding game $\G=(\dir G,V_C,c, S, r)$ with prescribed starting positions. First, we justify Assumption A. Enforcer gadget $\FORCE(S_1,\calA)$ does not allow any Arnold to move from his starting position unless Robber is on $CT$. When Robber is on $CT$, Lemma \[l:copgate\] describes the condition under which $\psi$-Arnold may move from $a_\psi$ to $a'_\psi$ and further to $a''$: it is precisely when clause $\psi$ is satisfied. The cop-player may refrain from moving more then one Arnold to $a''$ even if more then one clause in $F_C$ is satisfied, and he may also avoid moving another Arnold to $a''$, if one is already there. This will not change the outcome of the game as (because of Lemma \[l:arnoldwin\]), one Arnold on $a''$ already means the game is won by the cop-player. The above proves that using Assumption A does not change the outcome of the game. If there is an Arnold at $a''$, the cop-player has won. Otherwise, all Arnolds are at their starting positions; this will be important in Lemma \[l:robbergate\] and \[l:copgate\]. By Lemma \[l:beginning\], the assumptions of Lemma \[l:robs\], are satisfied at the beginning of the game. By Lemma \[l:robs\], the assumptions of Lemma \[l:robbergate\] are satisfied when Robber moves to $RT$. Lemma \[l:robbergate\] in turn that ensures we can apply Lemma \[l:cops\] in the next step, and similarly we continue with Lemma \[l:copgate\] which closes the cycle by guaranteeing the assumptions of Lemma \[l:robs\]. This altogether means that $\G$ precisely simulates the game $\F$: In the Robber Move phase (Lemma \[l:robs\]) $\G$ imitates setting of the variables $R$, which is due to Lemma \[l:robbergate\] followed in the Robber Test phase by test if the robber-player wins. By Lemma \[l:cops\] in the following Cop Move phase the game $\G$ imitates setting of the variables $C$, which is then due to Lemma \[l:copgate\] followed in the Cop Test phase by a test if the cop-player wins. The process then repeats. We have thus found the desired game $\G$ and proved its equivalence with $\F$. The first property is proved in Lemma \[l:logspace\]. The second property is in fact not true, but it is very easy to modify the construction, so that it is. We create a vertex $r'$ and connect it by an edge to all out-neighbours of $r$. We redefine the starting position of Robber as $r'$. It is immediate, that this modification does not change the outcome, and $r'$ has no incoming edge. The third and the fourth items are true by the construction. Forcing the starting positions and proof of the main theorem {#s:mainproof} ------------------------------------------------------------ Next we prove that we can modify our current construction so that it fully conforms to the definition of the guarding game on a directed graph, i.e., without prescribing the starting positions. \[l:force\] Let $\G=(\dir G,V_C,c, S, r)$ be a guarding game with a prescribed starting positions, such that it satisfies the following initial conditions: 1. There is no oriented edge $e\in E(\dir G)$ with the endpoint in $r$. 2. There is at most one cop standing at each vertex $v\in V(\dir G)$. 3. Consider the set $B=\{u\in V_R \mid \exists v\in V_C,\, (u,v)\in E(\dir G)\}$ of the “border” vertices. The out-degree of each $u\in B$ is exactly 1. Then there exists a guarding game $\G'=(\dir G',V'_C,c')$, $\dir G\subseteq \dir G'$, $V_C\subseteq V'_C$ such that the following holds: 1. The robber-player wins $\G'$ if and only if the robber-player wins the game $\G$. 2. If the robber-player does not place Robber on $r$ in his first move, the cops win. 3. If the cop-player does not place the cops to completely cover $S$ in his first move, he will lose. 4. The construction of $\G'$ can be made in log-space. 5. $|\G'| = O(|\G|)$ Let $m=|\{v\in V_C \mid (u,v)\in E(\dir G),\, u\in V_R\}|$ be the number of vertices from $V_C$ directly threatened (i.e., in distance 1) from robber-region. Let us define the graph $\dir G'=(V',E')$ such that $V'=V(\dir G)\cup\{r\}\cup T$ where $T=\{t_1,\dots,t_m\}$ is the set of new vertices and $E'=E(\dir G)\cup \{(r,v) \mid v\in T\cup S\}$. Consider the game $\G'=(\dir G',V'_C,c')$ where $V'_C=V_C\cup T$ and $c'=c+m$. See Fig. \[f:startforce\] for illustration. Suppose that the robber-player places Robber in the first move to some vertex $v\in V_R\setminus\{r\}$. Then there are $m$ vertices in $V_C$ directly threatened by edges going from $V_R$ and because we have at least $m$ cops available, the cops in the second move can occupy all these vertices and prevent Robber from entering $V_C$ forever. So Robber must start at the vertex $r$. Then observe, that $c$ cops must occupy the positions $S$ and $m$ cops must occupy the vertices $T$. If any cop does not start either on $T$ or $S$, Robber wins in the next move. The cops on $T$ remain there harmless to the end of the game. The cops cannot move until Robber decides to leave the vertex $r$. After that, the vertices in $V'\setminus V$ no longer affect the game, thus $\G'$ exactly imitates $\G$. The construction of $\dir G'$ from $\dir G$ can be realised by an iteration over the set of vertices and edges of $\dir G$, which means that the construction can be done in log-space. It is easy to see, that the number of vertices and edges we add is linear in the number of vertices of $\dir G$, proving the last statement. Here we give the precise definition of log-space reductions via length order. Let $\Sigma=\{0,1\}$ denote the alphabet and let $\Sigma^+$ denote all nonempty and finite words over the alphabet $\Sigma$. The function $f:\Sigma^+\to\Sigma^+$ is *logspace-computable* if there is a deterministic Turing machine with a separate two-way read-only input tape, a read/write work tape, and a one-way output tape such that, when started with any word $w\in\Sigma^+$ on the input tape, the machine eventually halts with $f(w)$ on the output tape while having visited at most $\log|w|$ cells on the work tape. Let $\ell:\N\to\R$. The function $f$ is *length $\ell(n)$ bounded* if $|f(w)|\le \ell(|w|)$ for all $w\in\Sigma^+$. Let $A,B\subseteq\Sigma^+$. *$A$ transforms to $B$ within logspace via $f$* (denoted by $A\logred B$ via $f$) if $f:\Sigma^+\to\Sigma^+$ is logspace-computable function such that $w\in A \Leftrightarrow f(w)\in B$ for all $w\in\Sigma^+$. Let $B$ be a language and let $\mathcal{L}$ be a class of languages. Then $\mathcal{L}\logred B$ if $A\logred B$ for all $A\in\mathcal{L}$. Furthermore, $\mathcal{L}\logred B$ *via length order $\ell(n)$, $\ell:\N\to\R$*, provided that for each $A\in\mathcal{L}$ there is a function $f$ and constant $b\in\N$ such that $A\logred B$ via $f$ and $f$ is length $b\ell(n)$ bounded. The language $B$ is *$\mathcal{L}$-complete under log-space reductions* if both $B\in\mathcal{L}$ and $\mathcal{L}\logred B$. By Lemma \[l:etime\], $\Guard\in\E$. Consider a problem $L \in \E$. By Theorem \[t:stocha\], $L$ can be reduced to the formula satisfying game $\F$ in log-space via length-order $n \log n$. By Theorem \[t:prescribedgame\] there exists an equivalent (in the terms of the game outcome) guarding game $\G=(\dir G,V_C,c, S,r)$ with prescribed starting positions, together with other properties as stated by Theorem \[t:prescribedgame\]. By Lemma \[l:force\] applied on the game $\G$ there is an equivalent guarding game $\G'$, $\G\subseteq \G'$, without prescribed starting positions. Thus, we reduced $L$ to . By Lemma \[l:logspace\] and Lemma \[l:force\], the whole reduction can be done in log-space, via length order $n (\log n)^2$. Therefore, Theorem \[t:main\] is proved. By the time hierarchy theorem, if $1 < a < b$, there is a problem $X$ in $\DTIME(b^n) \setminus \DTIME(a^n)$. Thus, $X$ is in $\E$ and we can use the reduction from Theorem \[t:main\]. For an instance $I$ of $X$ of size $n$, we can construct in polynomial time an instance $I'$ of  of size $m = O(n (\log n)^2)$ so that deciding $I'$ solves the instance $I$. By choice of $n$, this cannot be done in time $a^n$ for all $n$. It remains to check that $n = \Omega(m/(\log m)^2)$. Further questions {#further-questions .unnumbered} ================= As we have already mentioned, the relation of the classes  and  is unclear as we only know that $\PSPACE \ne \E$ and the current state of the art is missing some deeper understanding of the relation. Therefore, the conjecture of Fomin et al. whether  is -complete still remains open. However, we believe that the conjecture is not true. For a guarding game $\G=(G,V_C,c)$, what happens if we restrict the size of strongly connected components of $G$? If the sizes are restricted by 1, we get DAG, for which the decision problem is -complete. For unrestricted sizes we have shown that $\G$ is -complete. Is there some threshold for $\G$ to become -complete from being -complete? This may give us some insight into the original conjecture. Another interesting question is what happens if we bound the degrees of $G$ by a constant. We are also working on forcing the starting position in the guarding game on undirected graphs in a way similar to Theorem \[t:main\]. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank Ruda Stolař for drawing initial versions of some pictures and for useful discussion. We thank Peter Golovach for giving a nice talk about the problem, which inspired us to work on it. We would also like to thank Jarik Nešetřil for suggesting some of the previous open questions and to Honza Kratochvíl for fruitful discussion of the paper structure. [999]{} Aigner, M., Fromme, M.: A game of cops and robbers. Discrete Appl. Math. 8(1), 1–11 (1984) Alspach, B.: Searching and sweeping graphs: a brief survey. Matematiche (Catania) 59 (1-2), 5–37 (2006) Anderson, M., Barrientos, C., Brigham, R., Carrington, J., Vitray, R., Yellen, J.: Maximum demand graphs for eternal security. J. Combin. Math. Combin. Comput. 61, 111–128 (2007) Bonato, A., Nowakowski, R.: The game of cops and robbers on graphs. Student Mathematical Library, 61. AMS, Providence, RI, (2011). Book, R. V.: Comparing complexity classes. J. of Computer and System Sciences 9(2), 213–229 (1974) Burger, A.P., Cockayne, E.J., Grundlingh, W.R., Mynhardt, C.M., van Vuuren, J.H., Winterbach, W.: Infinite order domination in graphs. J. Combin. Math. Combin. Comput. 50, 179–194 (2004) Fomin, F., Golovach, P., Hall, A., Mihalák, M., Vicari, E., Widmayer, P.: How to Guard a Graph? Algorithmica 61(4), 839–856 (2011) Fomin, F.V., Golovach, P.A., Kratochvíl, J.: On tractability of Cops and Robbers Game. Proceedings of the 5th IFIP International Conference on Theoretical Computer Science (TCS 2008), Springer-Verlag IFIP International Federation for Information Processing 237, 171–185 (2008) Fomin, F.V., Golovach, P.A., Kratochvíl, J., Nisse, N., Suchan, K.: Pursuing a fast robber on a graph. Theor. Comp. Sci. 411, 1167–1181 (2010) Fomin, F.V., Golovach, P.A., Lokshtanov, D.: Guard Games on Graphs: Keep the Intruder Out! Theor. Comp. Sci. 412(46), 6484–6497 (2011) Fomin, F.V., Thilikos, D.M.: An annotated bibliography on guaranteed graph searching. Theor. Comp. Sci. 399, 236–245 (2008) Goddard, W., Hedetniemi, S.M., Hedetniemi, S.T.: Eternal security in graphs. J. Combin. Math. Combin. Comput. 52, 169–180 (2005) Goldstein, A.S., Reingold, E.M.: The complexity of pursuit on a graph. Theor. Comp. Sci. 143, 93–112 (1995) Goldwasser, J., Klostermeyer, W.F.: Tight bounds for eternal dominating sets in graphs. Discrete Math. 308, 2589–2593 (2008) Klostermeyer, W.F.: Complexity of Eternal Security. J. Combin. Math. Combin. Comput. 61, 135–141 (2007) Klostermeyer, W.F.,MacGillivray, G.: Eternal dominating sets in graphs. J. Combin. Math. Combin. Comput. 68, 97–111 (2009) Klostermeyer, W.F., MacGillivray, G.: Eternally Secure Sets, Independence Sets, and Cliques. AKCE International Journal of Graphs and Combinatorics, 2, 119–122 (2005) Klostermeyer, W.F., MacGillivray, G.: Eternal security in graphs of fixed independence number. J. Combin. Math. Combin. Comput. 63, 97–101 (2007) Kinnersley, W.B.: Cops and Robbers is EXPTIME-complete. arXiv:1309.5405 (2013). Mamino, M.: On the computational complexity of a game of cops and robbers. Theor. Comp. Sci. 477 (2013), 48–56 (2013). Nagamochi, H.: Cop-robber guarding game with cycle robber-region. Theor. Comp. Sci. 412, 383–390 (2011). Nowakowski, R., Winkler, P.: Vertex-to-vertex pursuit in a graph. Discrete Math. 43(2-3), 235–239 (1983) Quilliot, A.: Some results about pursuit games on metric spaces obtained through graph theory techniques. European J. Combin. 7(1), 55–66 (1986) Stockmeyer, L., Chandra, A.: Provably Difficult Combinatorial Games. SIAM J. Comput. Volume 8, Issue 2, 151–174 (1979) Šámal, R., Stolař, R., Valla, T.: Complexity of the Cop and Robber Guarding Game. Proceedings of IWOCA 2011, Springer-Verlag Lecture Notes in Computer Science vol. 7056, 361–373 (2011) Thirumala Reddy, T. V., Sai Krishna, D., Pandu Rangan, C.: The guarding problem – complexity and approximation. In IWOCA, volume 5874 of Lecture Notes in Computer Science, 460–470, Springer (2009). [^1]: Partially supported by Karel Janeček Science & Research Endowment (NFKJ) grant 201201, by grant LL1201 ERC CZ of the Czech Ministry of Education, Youth and Sports and by grant GA ČR P202-12-G061. [^2]: Supported by the Centre of Excellence – Inst. for Theor. Comp. Sci. (project P202/12/G061 of GA ČR), and by the GAUK Project 66010 of Charles University in Prague.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present the first results for the sputtering of grain mantles and cores obtained with self-consistent multifluid hydromagnetic models of C-type shocks propagating through dusty media. The threshold shock speed for mantle sputtering is about 10 km s$^{-1}$ and is independent of density. The mantles are completely vapourised in shocks with speeds of 20-25 km s$^{-1}$. At such shock speeds core sputtering commences and gas-phase SiO forms. Core destruction is not total in any C-type shock because grains are not completely destroyed in shocks with speeds near the minimum speeds at which J-type shocks appear. Due to the density-dependence of the critical shock speed for this transition, higher gas-phase SiO fractional abundances are produced behind shocks propagating in lower density gas. For shock speeds near the threshold speeds for both core and mantle sputtering, sputtering is much greater for shock velocities at smaller angles relative to the upstream magnetic field. At higher shock speeds, the angular variation is still present but less pronounced.' author: - | S. Van Loo,$^1$[^1], I. Ashmore,$^2$, P. Caselli$^2$, S. A. E. G. Falle$^3$ and T. W. Hartquist$^2$\ $^1$ Harvard-Smithsonian Center for Astrophysics, 60 Garden St., Cambridge, Massachusetts 02138, USA\ $^2$ School of Physics and Astronomy, University of Leeds, Leeds LS2 9JT, UK\ $^3$ Department of Applied Mathematics, University of Leeds, Leeds LS2 9JT, UK date: 'Accepted -. Received -; in original form -' title: 'Sputtering in oblique C-type shocks' --- \[firstpage\] MHD – shockwaves – ISM: dust – ISM: jets and outflows. Introduction {#sec:intro} ============ A number of computational studies of sputtering in self-consistent models of perpendicular C-type hydromagnetic shocks in dusty, dark regions have been performed (e. g. [@FPDF95]; [@Sal97]; [@Mal00]; [@Gal08]; [@Gu11]). However, the only numerical investigation of sputtering in oblique C-type shocks in such regions is not based on self-consistent shock models ([@CHH97]). [@PH94] attempted to construct a self-consistent, four fluid model of an oblique C-type shock in which dust grains are described as a charged fluid, as they are in the perpendicular C-type shock models of [@PHH90]. However, the integrations performed by [@PH94] failed to yield solutions corresponding to steady fast-mode shocks. [@W98] pointed out that the failure is due to the saddle-point nature of the downstream state of a fast-mode dusty shock. [@F03] developed a numerical scheme for solving the time-dependent multifluid hydromagnetic equations that enables the construction of steady fast-mode solutions even when local equilibrium does not obtain. [@vLal09] have used the scheme along with non-equilibrium chemistry and treatments of grain charge and heating and radiative losses to investigate the evolution of oblique C-type shocks propagating into media with constant upstream properties. [@Aal10] extended the applications to oblique C-type shocks propagating through media with varying upstream properties. Here we use models like those of [@vLal09] to study sputtering in oblique C-type shocks in dusty, dark molecular regions. Section 2 contains descriptions of the assumptions, the physical processes included, the chemical network, the evaluation of the sputtering rates and upstream boundary conditions. Results are presented in Section 3, and Section 4 contains a discussion of the results and conclusions. The Model {#sec:themodel} ========= The model includes neutral, ion, electron and two grain fluids. Flow variables depend on only the spatial coordinate $x$ and time $t$. The upstream magnetic field has components in the $x$ and $y$ directions, and the shock propagates in the $x$ direction. Equations  1 - 8 of [@vLal09] are solved with the scheme of [@F03]. The inertia of the charged fluids, including the grain fluids, is neglected in this approach, so that their dynamics is solely determined by the balance between the Lorentz force and collision forces with neutral particles. Additionally, the grain fluids are assumed to have zero pressure. The average charge on a grain is determined by the condition that the magnitude of the current of ions onto a grain equals the magnitude of the current of electrons onto a grain; Eqs. 10 and 11 of @HHP87 [see also Appendix \[sect:appA\]] are used to calculate these currents. The numerical method of @F03 provides a rigorous means to model steady shock structures and, as discussed by [@vLal09] and @Aal10, reproduces the results of, among others, @Del83, @PH94, @CW06 and @GPJ07. Furthermore, although it neglects grain inertia, the method can accurately model neutral subshock structures [@F03; @Aal10]. The evolution of the number density, $n(X)$, of chemical species $X$ is governed by a rate equation of the form: $$\begin{aligned} \label{eqn:chemsource} \frac{\partial n(X)}{\partial t} + \frac{\partial n(X) u_s}{\partial x} = s(X),\end{aligned}$$ $u_{s}$ is the $x$ component of the ion or neutral velocity depending on whether $X$ is an ion or a neutral. $s(X)$ is the net rate of formation or destruction of species $X$ per unit volume per unit time due to the gas phase reactions and grain surface recombinations listed in Table \[table:rateqns\] and the return of material to the gas phase due to sputtering. Though no reactions involving SiO are shown in Table \[table:rateqns\], atomic Si returned to the gas phase due to sputtering is assumed to be converted to SiO immediately. ![The fractional abundance of H$_2$O produced by gas-phase chemistry (solid) and mantle sputtering (dashed) as function of the shock speed for $n_{\rm H0} = 10^6 cm^{-3}$.[]{data-label="fig:Fig4"}](Fig4){width="\columnwidth"} --------------------------------------------------------------------------------------------------------------------------------------------------------- Num. Reaction Rate (s$^{-1}$) or Rate Coefficient (cm$^{3}$ s$^{-1}$) Ref. -------- ----------------------------------------------------------------- -------------------------------------------------------------------- --------- 1$^a$ H$_{2}$ + CR $\rightarrow$ H$_{3}^{+}$ + e + CR 3 $\times$ 10$^{-17}$ 1 2 H$_{3}^{+}$ + CO $\rightarrow$ HCO$^{+}$ + H$_{2}$ 1.7 $\times$ 10$^{-9}$ 1, 2, 3 3 H$_{3}^{+}$ + H$_{2}$O $\rightarrow$ H$_{3}$O$^{+}$ + H$_{2}$ 5.9 $\times$ 10$^{-9}$ 2, 3 4 H$_{3}^{+}$ + grain$^{-}$ $\rightarrow$ H$_{2}$ + H + grain $\left(\Gamma_{ig} + \Gamma_{is}\right)\ n({\rm H}_{3}^{+})/n_{i}$ 4 5 H$_{3}^{+}$ + O $\rightarrow$ H$_{3}$O$^{+}$ 8.4 $\times$ 10$^{-10}$ 2, 3 6 HCO$^{+}$ + e $\rightarrow$ H + CO 2.76 $\times$ 10$^{-7} 3 \left({\rm T}_{e}/300\right)^{-0.64}$ 7 HCO$^{+}$ + Mg $\rightarrow$ H + CO + Mg$^{+}$ 2.9 $\times$ 10$^{-9}$ 2 8 HCO$^{+}$ + H$_{2}$O $\rightarrow$ H$_{3}$O$^{+}$ + CO 2.5 $\times$ 10$^{-9}$ 2, 3 9 HCO$^{+}$ + grain$^{-}$ $\rightarrow$ H + CO + grain $\left(\Gamma_{ig} + \Gamma_{is}\right)\ n({\rm HCO}^{+})/n_{i}$ 4 10 Mg$^{+}$ + e $\rightarrow$ Mg + h$\nu$ 2.8 $\times$ 10$^{-12} 1, 2 \left({\rm T}_{e}/300\right)^{-0.86}$ 11 Mg$^{+}$ + grain$^{-}$ $\rightarrow$ Mg + grain $\left(\Gamma_{ig} + \Gamma_{is}\right)\ n({\rm Mg}^{+})/n_{i}$ 4 12$^b$ Mg$^{+}$ + H$_{2}$ $\rightarrow$ MgH$^{+}$ + H 3.0 $\times$ 10$^{-10}$exp(-27854/T$_{\rm eff}$) 1 13 MgH$^{+}$ + e $\rightarrow$ Mg + H 1.1 $\times$ 10$^{-7}$ (T$_{e}$/300)$^{-0.5}$ 1 14 MgH$^{+}$ + grain$^{-}$ $\rightarrow$ Mg + H + grain $(\Gamma_{ig} + \Gamma_{is})\ n({\rm MgH}^{+})/n_{i}$ 4 15$^c$ H$_{3}$O$^{+}$ + e $\rightarrow$ H$_{2}$O + H 3.29 $\times$ 10$^{-7}$(T$_{e}$/300)$^{-0.5}$ 2, 3 16 H$_{3}$O$^{+}$ + Mg $\rightarrow$ H$_{2}$O + Mg$^{+}$ + H 1.8 $\times$ 10$^{-9}$ 1 17 H$_{3}$O$^{+}$ + grain$^{-}$ $\rightarrow$ H$_{2}$O + H + grain $(\Gamma_{ig} + \Gamma_{is})\ n({\rm H}_{3}{\rm O}^{+})/n_{i}$ 4 18$^d$ O + H$_{2}$ $\rightarrow$ H$_{2}$O 3.14 $\times$ 10$^{-13}$(T$_{n}$/300)$^{2.7}$exp(-3150/T$_{n}$) 2, 3 --------------------------------------------------------------------------------------------------------------------------------------------------------- $^a$ Reaction 1 is an amalgamation of two reactions: H$_{2}$ + CR $\rightarrow$ H$_{2}^{+}$ + e + CR and H$_{2}^{+}$ + H$_{2}$ $\rightarrow$ H$_{3}^{+}$ + H where CR denotes a cosmic ray. The adopted coefficient is that of the first reaction since the proton transfer reaction occurs sufficiently rapidly. $^b$ T$_{\rm eff}$ in the rate coefficient of Reaction 12 defines a weighted kinetic temperature at the ion-neutral drift speed and is given by Eq. 43 of [@FPH85]. $^c$ Reaction 15 is an amalgamation of:\ H$_{3}$O$^{+}$ + e $\rightarrow$ OH + H$_{2}$ $(6.02 \times 10^{-8} ({\rm T}_{e}/300)^{-0.5})$,\ H$_{3}$O$^{+}$ + e $\rightarrow$ OH + H + H $(2.58 \times 10^{-7} ({\rm T}_{e}/300)^{-0.5})$ and\ H$_{3}$O$^{+}$ + e $\rightarrow$ H$_{2}$O + H $(1.08 \times 10^{-8}({\rm T}_{e}/300)^{-0.5})$.\ OH is not treated as a separate species, because the OH produced by the first two reactions quickly reacts with H$_{2}$ to yield H$_{2}$O (see $^d$). $^d$ Reaction 18 is also an amalgamation of two reactions: O + H$_{2}$ $\rightarrow$ OH + H and OH + H$_{2}$ $\rightarrow$ H$_{2}$O + H. The coefficient adopted is that of the first (slower) of the two. ![image](Fig1new){width="7.0in"} The model includes the radiative losses due to O, CO, H$_2$ and H$_2$O, which are calculated for the abundances of O and H$_2$O given by the integration of the chemical rate equations and the assumed constant fractional abundances of CO and H$_2$. Given those abundances, the cooling rates are evaluated as they were by [@vLal09]. They adopted the treatments of [@Del83] for the cooling of electrons and for OI emission, used data from [@Hel80] for the cooling of neutrals by H$_2$ and implemented the treatment of CO cooling due to [@HM79] and that of H$_2$O cooling developed by [@NM87]. Sputtering due to H$_2$, CO, O, Mg, H$_2$O and SiO projectiles is included. As mentioned above, the return of material to the gas phase due to sputtering contributes to the source terms in the rate equations governing the gas phase chemistry; these contributions to the source terms are particularly important for the model gas phase abundances of Mg, H$_2$O and SiO. A grain consists of a refractory core made of olivine (MgFeSiO$_{4}$) and a mantle of water ice. A core remains intact until all of the ice has been sputtered. The sputtering rates are calculated as they were by [@JSal08], who used Eq. 27 of [@DS79] for the sputtering rate. For all projectiles, they used Eq. 31 of [@DS79] to obtain the sputtering yields of ice at normal incidence. They used the same equation to obtain the corresponding yield for core material colliding with H$_2$. However, for projectiles other than H$_2$, they employed the results of [@Mal00] for the sputtering yields of core material at normal incidence. Following [@DS79], [@JSal08] took the angle averaged value of a yield to be twice the value of the normal incidence yield. The upstream magnetic field strength is $B_0 = 1\mu {\rm G} (n_{\rm H0}/1~{\rm cm}^{-3})^{1/2}$, where n$_{\rm H0}$ is the upstream value of n$_{\rm H}$, the number density of hydrogen nuclei. A shock propagates with a speed v$_s$ at an angle $\theta$ with respect to the upstream magnetic field (for an oblique shock, $0^\circ \lneqq \theta \lneqq 90^\circ$). The grains have radii of 0.4 $\mu$m and 0.04 $\mu$m and masses of 8.03 $\times 10^{-13}$ g and 8.03 $\times 10^{-16}$ g. In the upstream region, one percent of the mass is contained in grains with the number density of smaller grains 316 times that of the larger grains (i.e. the grains follow the grain-size distribution of @MRN77). As all the grains are silicate-based in our model, the fractional abundance of Si, i.e. 5.8 $\times 10^{-5}$, is somewhat higher than the cosmic value of 3.37 $\times 10^{-5}$ [@AG89]. The fractional abundance of CO to hydrogen nuclei has a constant value of 5 $\times$ 10$^{-5}$ ahead, throughout and behind the shock, and the upstream gas phase fractional abundances of O, Mg, H$_{2}$O and SiO are 4.25 $\times$ 10$^{-4}$, 10$^{-7}$, 0 and 0, respectively. Note that, including the fractional abundance of CO and of olivine, the total abundance of O in our model exceeds the cosmic abundance of O by less than a factor of two (the latest measured value being 3.2 $\times 10^{-4}$ [@Meyeretal98] although give 4.6 $\times 10^{-4}$). The initial fractional ionisation is calculated as it was by [@vLal09]; HCO$^+$ and Mg$+$ are initially the only ions, but the chemistry leads to the production of significant fractional abundances of H$_{3}^{+}$ and H$_{3}$O$^{+}$ before heating by the shock occurs. Although upstream of a shock there is no water in the gas phase, the upstream fractional abundance of water ice molecules on the grains is 7.25 $\times$ 10$^{-5}$ (cf. [@WD91]); the upstream value of the ratio of the total mass of ice on small grains to the total mass of ice on big grains is taken to be 3.16. We followed [@vLal09] by assuming that the initial flow corresponds to that of a J-type shock. The presented results are for the steady flows yielded by time integrations. --------------- -------------- ------------- --------------------- ------------- --------------------- -------------- ------------------- Shock Speed n$_{\rm H0}$ Fraction of Contribution to Fraction of Contribution to Fraction of Contribution to (km s$^{-1}$) (cm$^{-3}$) ice removed downstream H$_{2}$O ice removed downstream H$_{2}$O ice removed downstream H$_2$O 10 1.00(4) 1.20(-3) 3.09(-8) 3.80(-3) 2.09(-7) 3.31(-3) 2.40(-7) 12 1.00(4) 4.00(-2) 1.04(-6) 5.50(-2) 3.03(-6) 5.61(-2) 4.07(-6) 14 1.00(4) 1.75(-1) 3.05(-6) 2.00(-1) 1.10(-5) 1.94(-1) 1.41(-5) 16 1.00(4) 4.50(-1) 7.83(-6) 4.80(-1) 2.64(-5) 4.72(-1) 3.42(-5) 20 1.00(4) 1.00 1.74(-5) 1.00 5.51(-5) 1.00 7.25(-5) 8 1.00(6) 0.00 - 0.00 - 0.0 - 9 1.00(6) 0.00 - $<$1.00(-4) $\approx$ 0.00 $<$ 1.00(-4) $\approx$ 0.00 10 1.00(6) 0.00 - 5.00(-3) 2.75(-7) 3.79(-3) 2.75(-7) 12 1.00(6) 2.50(-3) 4.36(-8) 6.95(-2) 3.83(-6) 5.34(-2) 3.87(-6) 14 1.00(6) 2.50(-2) 4.36(-7) 2.63(-1) 1.45(-5) 2.06(-1) 1.49(-5) 15 1.00(6) 5.00(-2) 8.71(-7) 4.25(-1) 2.34(-5) 3.35(-1) 2.43(-5) 16 1.00(6) 7.78(-2) 1.36(-6) 6.07(-1) 3.34(-5) 4.79(-1) 3.48(-5) 18 1.00(6) 1.8(-1) 3.14(-6) 1.00 5.51(-5) 8.03(-1) 5.82(-5) 20 1.00(6) 3.4(-1) 5.93(-6) 1.00 5.51(-5) 8.42(-1) 6.10(-5) 25 1.00(6) 1.00 1.74(-5) 1.00 5.51(-5) 1.00 7.25(-5) --------------- -------------- ------------- --------------------- ------------- --------------------- -------------- ------------------- Results {#sec:results} ======= We focus on the abundances of H$_2$O and SiO as mantle and core sputtering releases these molecules into the gas phase. However, water is also produced by shock chemistry. To quantify the contributions of both processes we discuss the chemistry first and then grain sputtering. Shock chemistry --------------- The specified initial relative abundances are not in steady state chemical equilibrium. The upstream gas phase ion abundances initially evolve on a timescale short compared to the flow time through the shock and then evolve further as the abundances of other key species (e. g. H$_2$O) vary. H$_{3}$O$^{+}$ and HCO$^{+}$ are the most common molecular ions, but the Mg$^{+}$ fractional abundance is about an order of magnitude greater than the total fractional abundance of molecular ions. The ratio of the total atomic ion abundance to the total molecular abundance depends on the assumed elemental abundance of metals. The assumed upstream water fractional abundance is about three orders of magnitude below the water fractional abundances obtained with standard ion-molecule chemical models, i.e. $\sim$ 10$^{-6}$ [e.g. @Mal97; @NLM95]. However, the upstream value is close to the water abundances of $\sim$ 10$^{-8}$ inferred for quiescent molecular gas towards Orion, M17 and other sources by @Bal00. Recent Herschel observations of ortho-H$_{2}$O in a dark cloud [@Casal10] imply values of $\sim$ 10$^{-10}$. Within the shock front the gas heats up, and some reactions with barriers and endothermic reactions occur. Also, the heating of electrons leads to enhanced electron impact rates on the grains, and the drift of the ions relative to the grains contributes further to efficient ion recombination on grains, reducing the fractional ionisation within the shock. As usual, the primary sequence for H$_2$O formation in the shock involves the reactions of O and OH with H$_2$. Figure \[fig:Fig4\] shows the the fractional abundance of water produced in the shock as a function of the shock speed. At the lowest shock speed (i.e. 5 km s$^{-1}$), the shocked gas does not reach high enough temperatures to cause the conversion of much oxygen to water. At a shock speed of 10 km s$^{-1}$ a few percent of the gas phase oxygen is converted to water, but it is only at shock speeds close to 20 km s$^{-1}$ and above at which the atomic oxygen abundance falls by several orders of magnitude, and nearly all of the gas phase oxygen is in water. The H$_2$O abundance is then a few times 10$^{-4}$. Mantle sputtering ----------------- Within a C-type shock, grains experience high-speed impacts with gas particles. Sputtering leads to a continuous erosion of the icy grain mantle by which water is returned to the gas phase. Grain mantle sputtering occurs for impact speeds $> 10/\sqrt{\mu}$ km s$^{-1}$ with $\mu$ the mean mass of the impacting species in atomic mass units [e.g. @Mal00]. Since CO is the most massive abundant projectile with $\mu = 28$, mantle sputtering starts for grain-neutral drift speeds as low as $\approx 2$ km s$^{-1}$. (Note from Fig. \[fig:obliquemantlesputter\] that appreciable sputtering only occurs at drift speeds above 6 km s$^{-1}$). This value of the drift speed is independent of the grain size, because the mantle molecules are bound with the same energy. However, the threshold drift speed is reached at different shock speeds for the small and large grains. Small grains have Hall parameters[^2] that are large enough that they move with the ions and electrons. The Hall parameter for the large grains is above unity but sufficiently small that they are only moderately coupled to the magnetic field and, thus, move at velocities between those of the electrons and the neutrals. Consequently, the grain-neutral drift speed for the large grains is smaller than for the small grains at a given shock speed. No projectile induces noticeable sputtering when v$_{s}$ = 5 km s$^{-1}$, but when v$_{s}$ has a higher value, small grains are sputtered. Table \[table:mantlesputterfrac\], as well as Fig. \[fig:Fig3\], displays the fraction of the initial mantle ice that is sputtered by a shock for which $\theta = 45^{\circ}$. Results are given for various values of the shock speed and values of n$_{\rm H0}$ of 10$^{4}$ cm$^{-3}$ and 10$^{6}$ cm$^{-3}$. It also gives the enhancement in the downstream gas phase water fractional abundance due to sputtering. If we define the sputtering threshold shock speed as the shock speed for which 1% of the ice is removed [cf. @CHH97], the threshold shock speed for small grain sputtering is 10 km s$^{-1}$ and for large grain sputtering it is 10 (12) km s$^{-1}$ for an upstream number density of 10$^4$ (10$^6$) cm$^{-3}$. One also sees that the fraction of material sputtered from small grains is not very sensitive to n$_{\rm H0}$, but particularly at lower shock speeds the fraction of material sputtered from the large grains does depend on n$_{\rm H0}$. This again is a result of the coupling of the grains to the magnetic field (see above). The large grains have a larger Hall parameter in the lower density regime and, thus, have a larger neutral-grain drift speed than in the high-density regime. ![The fraction of mantle ice (blue) and core silicon (red) released in the gas-phase by sputtering as function of shock speed for values of $n_{\rm H0}$ of 10$^4$ cm$^{-3}$ (solid) and 10$^6$ cm$^{-3}$ (dashed). The squares show the sputtering fractions for shock angles of 30, 45 and 60 degrees for selected shock speeds; the sputteringefficiency decreases with increasing angle.[]{data-label="fig:Fig3"}](Fig3){width="\columnwidth"} For shocks for which $\theta = 45^{\circ}$, the highest sputtering rate per unit time per unit volume of mantles is due to collisions with CO for shock speeds less than about 10 km s$^{-1}$ and to collisions with H$_2$O for speeds above about 14 km s$^{-1}$. At intermediate speeds, collisions with O also contribute non-negligibly as the relative importance of CO and H$_2$O collisions varies. Above 30 km s$^{-1}$, H$_2$ dominates the sputtering of both grain types. However, at these shock speeds, the icy mantles of both grain species are completely eroded. The contribution of the grain sputtering to the water abundance is even then negligible compared to the contribution of the gas-phase oxygen conversion (see Fig. \[fig:Fig4\]). --------------- -------------- ------------- ----------------- ------------- ----------------- -------------- ----------------- Shock Speed n$_{\rm H0}$ Fraction of Contribution to Fraction of Contribution to Fraction of Contribution to (km s$^{-1}$) (cm$^{-3}$) Si removed downstream SiO Si removed downstream SiO Si removed downstream SiO 20 1.00(4) 0.00 - $<$1.00(-4) $\approx$0.00 $<$ 1.00(-4) $\approx$ 0.00 30 1.00(4) 3.62(-4) 1.60(-8) 4.58(-3) 6.40(-8) 1.37(-3) 8.00(-8) 35 1.00(4) 2.83(-3) 1.25(-7) 3.40(-2) 4.75(-7) 1.03(-2) 6.00(-7) 40 1.00(4) 5.59(-3) 2.47(-7) 1.40(-1) 1.95(-6) 3.78(-2) 2.20(-6) 45 1.00(4) 1.12(-2) 4.94(-7) 2.51(-1) 3.51(-6) 6.87(-2) 4.00(-6) 50 1.00(4) 2.60(-2) 1.15(-6) 3.04(-1) 4.25(-6) 9.28(-2) 5.40(-6) 20 1.00(6) 0.00 - 0.00 - 0.00 - 25 1.00(6) 0.00 - 7.37(-6) 1.03(-10) 1.77(-6) 1.03(-10) 30 1.00(6) 0.00 - 3.22(-3) 4.50(-8) 7.73(-4) 4.50(-8) 35 1.00(6) 9.76(-4) 4.31(-8) 3.26(-2) 4.56(-7) 8.57(-3) 4.99(-7) --------------- -------------- ------------- ----------------- ------------- ----------------- -------------- ----------------- Core sputtering --------------- Once the grain mantles are vapourised, the grain cores begin to erode as long as the impact speeds of the projectiles are high enough. As the binding energy of a particle in the grain core is higher than in the mantle, higher grain-neutral drift speeds are necessary. Again CO is the sputtering projectile of most importance at lower speeds, and small grains start to contribute to the SiO gas-phase abundance at shock speeds of 25 km s$^{-1}$ (see Fig. \[fig:Fig3\] for shocks for which $\theta = 45^{\circ}$). This result is similar to the one of @CHH97 and @Mal00. Collisions with CO dominate the sputtering at shock speeds up to 30 km s$^{-1}$. Above shock speeds of about 35 km s$^{-1}$ the highest sputtering rate per unit time per unit volume of elemental silicon is due to collisions with H$_2$O. Table \[table:coresputterfrac\] displays the fraction of the initial core elemental silicon that is sputtered. It also gives the enhancement in the downstream gas phase SiO fractional abundance due to sputtering. Similarly to the results for mantle sputtering, the core sputtering of the small grains is independent of $n_{\rm H0}$ (as their Hall parameter is much larger than unity). The large grains are more efficiently sputtered in the low-density regime but contribute only a negligible amount to the SiO gas-phase abundance. For a range of shock speeds, the SiO production rises with increasing speed with values similar as in Fig. 2 of @Gal08a. However, the SiO abundance reaches an upper limit as C-type shocks do not exist for shock speeds above a certain value. For higher shock speeds, shocks are of J-type. For $n_{\rm H0}=10^4~{\rm cm^{-3}}$ the transition occurs for a shock speed between 50 and 55 km s$^{-1}$, while a J-type shock occurs for a minimum shock speed between 35 and 40 km s$^{-1}$ for $n_{\rm H0} = 10^6~{\rm cm^{-3}}$. At these speeds the grain cores are not completely destroyed, i.e. only a few percent of the core is removed in the high-density case and about a third of the core is destroyed in the low-density regime. Hence, the maximum SiO fractional abundance differs by an order of magnitude between the two density regimes. Angular dependence ------------------ Up to now we have discussed the grain sputtering for an oblique shock for which $\theta = 45^\circ$. However, @vLal09 showed that the shock properties, such as the shock width, change as function of $\theta$. Therefore, variations, with angle, of the H$_2$O and SiO abundances due to sputtering is expected. ![image](Fig2new){width="7.0in"} Figure \[fig:obliquemantlesputter\] presents results for 16 km s$^{-1}$ shocks propagating into media with n$_{\rm H0}$ = 10$^{4}$ cm$^{-3}$ at angles of 30$^{\circ}$, 45$^{\circ}$ and 60$^{\circ}$ with respect to the upstream magnetic field. The fraction of the ice that is sputtered from each type of grain, rather than the gas phase abundance of water, is shown because much of the gas phase water exists due to the conversion of upstream gas phase atomic oxygen to water. The results in Fig. \[fig:obliquemantlesputter\] indicate that the H$_2$O abundance is lowest in the shock with the largest width. It is obvious that the shock width is irrelevant for the level of sputtering as the sputtering depends mainly on the neutral-grain drift speeds. As the bottom row of Fig. \[fig:obliquemantlesputter\] shows, the peak small and large grain-neutral drift speeds are smallest at large angles; they are 10.5 km s$^{-1}$ at 30$^{\circ}$, 10 km s$^{-1}$ at 45$^{\circ}$ and 9.5 km s$^{-1}$ at 60$^{\circ}$. This behaviour is consistent with the results shown in Fig. 2 of [@vLal09] who found that the $y$ and $z$ components of the velocities of charged species are greater for smaller angles. Furthermore, the length scale over which the drift speeds are higher than 6 km s$^{-1}$, i.e. the minimum drift speed for which mantle sputtering is efficient, are similar for all angles. The angular variations in the fraction of water that is sputtered are significant and range from 0.4 at 60$^\circ$ to 0.6 at 30$^\circ$ Such variations disappear with increasing shock speed because for high enough speeds vaporisation of the grain mantle is complete for all angles. Figure \[fig:obliquecoresputter\] presents results for 25 km s$^{-1}$ shocks propagating into media with n$_{\rm H0}$ = 10$^{6}$ cm$^{-3}$ at angles of 30$^{\circ}$, 45$^{\circ}$ and 60$^{\circ}$ with respect to the upstream medium. The results in Fig. \[fig:obliquecoresputter\] show that the SiO abundance too is lowest in the shock with the largest width. Again these differences can be attributed to the peak small grain-neutral drift speeds which are smallest at large angles; they are 18.2 km s$^{-1}$ at 30$^{\circ}$, 18 km s$^{-1}$ at 45$^{\circ}$ and 17 km s$^{-1}$ at 60$^{\circ}$. Contrary to the H$_2$O results, these modest differences in drift speeds have large effects on the amount of elemental silicon that is sputtered, i.e. there is a difference of two orders of magnitude between the SiO abundance at 30$^\circ$ and 60$^\circ$. However, note that the shock speed is close to the threshold speed for core sputtering. Then small variations in the drift speed can mean the difference between core sputtering or no sputtering at all. For higher shock speeds the angular differences in SiO abundance are less pronounced, but the trend of lower abundances for larger $\theta$ remains. Discussion and Conclusions ========================== In this paper we present the first self-consistent multifluid MHD simulations of grain mantle and core sputtering in oblique C-type shocks propagating through dusty media. The small grains are more tightly coupled to the magnetic field than the large grains with a strong coupling in both low- and high-density regimes. This maximises the drift speeds between the small grains and the neutral sputtering projectiles resulting in more efficient sputtering of the small grains. Mantle sputtering starts at shock speeds of about 10 km s$^{-1}$, and the mantles are completely vapourised for shock speeds around 20-25 km s$^{-1}$. The sputtering of the icy mantles returns water back to the gas phase. However, neutral-neutral chemistry dominates the production of gas phase water by an order of magnitude. Once the mantles are completely eroded, core sputtering commences and releases silicon to form SiO. The abundance of SiO in the gas phase never saturates as even at the critical speed for the transition from C- to J-type shocks not all of a grain core is destroyed. While the critical speed is 50-55 km s$^{-1}$ for $n_{\rm H0} = 10^4~{\rm cm^{-3}}$, it is only 40 km s$^{-1}$ for $n_{\rm H0} = 10^6~{\rm cm^{-3}}$. For a given shock speed the efficiency of small grain sputtering is not sensitive to the density. Consequently, the density dependence of the C- to J-type transition means that higher SiO abundances can be produced by shocks in the low-density regime. An important result is that the sputtering yields depend on the angle between the magnetic field and the shock normal. Especially, at the threshold drift speeds for mantle or core sputtering, small angular variations can induce large changes in abundances. For example, for shocks propagating at 25 km s$^{-1}$ in media with number densities of 10$^6 {\rm cm^{-3}}$, the SiO abundance differs by two orders of magnitude for $\theta = 30^\circ$ and $60^\circ$. This has important consequences for the interpretation of observations. For example, discrete emission features in SiO observations have been ascribed to different shocks being present within the telescope beam [e.g. @Izas2009]. However, such features can be explained by a bow shock interacting with an inhomogeneity in the medium. The flanks of the bow shock have lower speeds than the head and also make different angles with the magnetic field. In an inhomogeneous medium some regions of the bow-shock are then highlighted resulting in discrete SiO emission features. Such an assumption, however, needs to be further examined by extending the planar shock model to multiple dimensions. Direct comparisons of the results with those of other researchers is complicated as the only previous work ([@CHH97]) on sputtering in oblique shocks is not based on self-consistent shock models. Also, the treatment of core sputtering in that work differed from that developed by [@Mal00] and used in subsequent work. However, the results given in Table \[table:coresputterfrac\] for the fraction of elemental silicon sputtered from small grains are in harmony with the results of @CHH97 for a shock for which $\theta = 45^\circ$. Other self-consistent studies concern only perpendicular shocks ($\theta = 90^\circ$) and, thus, should find similar, although lower, values for shock speed exceeding the threshold value for sputtering. [@Gal08a], who assumed a somewhat larger upstream magnetic field strength and different grain population, found that the fraction of silicon that is sputtered is about 0.04 for n$_{\rm H0} = 10^{4}$ cm s$^{-1}$ and a shock speed 40 km s$^{-1}$. [@Gu11] who adopt a kinetic method for the dust grain dynamics in favour of a fluid approach, find a value of 0.02 for the same model as @Gal08a. The corresponding result given in Table \[table:coresputterfrac\] is 0.04. The SiO abundances found in our simulations only represent upper limits. Observations indicate that silicate-based grains only account for about 75% of the interstellar dust mass with the rest of the mass stored in carbonaceous grains or PAHs [@Draine2011]. Furthermore, the conversion of atomic silicon into SiO is not fully efficient above densities of 10$^5 {\rm cm^{-3}}$ which possibly reduces the SiO abundance due to sputtering by an order of magnitude [@Gal08]. Unlike [@Gal08a] and [@Gu11], who included grains 40 times smaller than the smallest that we have assumed to be present, we have not considered the destruction of grain material due to grain-grain collisions. Our choice to focus solely on sputtering first was based in part on a desire to understand the effects of obliqueness on sputtering before looking at the complicated way in which obliqueness, sputtering and grain-grain collisions all act together to destroy grains. Our decision to do so was informed by the estimates of [@CHH97] that showed for grain populations of the sizes that we selected, grain-grain collisions would contribute significantly to mantle removal only at shock speeds near the threshold even for high pre-shock densities. Our choice has allowed us to see that obliqueness affects sputtering near the threshold shock speed significantly. However, considering the findings of both [@CHH97] and of [@Gu11] for the return of silicon to the gas phase, as well as the results given in Figure \[fig:obliquecoresputter\], we conclude that the future inclusion of both grain-grain collisions and obliqueness in shock models is necessary to infer the threshold shock speed for the injection of significant quantities of silicon into the gas phase. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the anonymous referee for his/her constructive comments. The authors are grateful for the support provided by a studentship and successive rolling grants awarded by the UK Science and Technology Funding Council and from the SMA Postdoctoral Fellowship of the Smithsonian Astrophysical Observatory held by SVL. Particle current onto a grain {#sect:appA} ============================= @HHP87 derived general expressions for the current of charged particles (with charge $Z_j$) onto a dust grain carrying a charge of $Z_g$. For $Z_g Z_j \leq 0$, $$\begin{aligned} \Gamma_{jg} &= \pi a^2 n_j Z_j e \frac{c_j^2}{2v_{jg}} \Bigl\{\left(1+2\left(\frac{v_{jg}}{c_j}\right)^2 - 2Z_jZ_g\xi \right) \\ &\times {\rm erf} \left(\frac{v_{jg}}{c_j}\right) + \frac{2(v_{jg}/c_j)}{\sqrt{\pi}} \exp\left(-\left(\frac{v_{jg}}{c_j}\right)^2 \right) \Bigr\},\end{aligned}$$ and, for $Z_g Z_j \ge 0$, $$\begin{aligned} \Gamma_{jg} &= \pi a^2 n_j Z_j e \frac{c_j^2}{2v_{jg}} \Bigl\{\frac{1}{2}\left(1+2\left(\frac{v_{jg}}{c_j}\right)^2 - 2Z_jZ_g\xi \right) \\ &\times \left[{\rm erf}\left(\frac{v_{jg}}{c_j} + \sqrt{Z_jZ_g\xi}\right) + {\rm erf}\left(\frac{v_{jg}}{c_j} - \sqrt{Z_jZ_g\xi}\right) \right] \\ & + \frac{1}{\sqrt{\pi}} \left(\frac{v_{jg}}{c_j}+ \sqrt{ZjZ_g\xi} \right) \exp\left(-\left(\frac{v_{jg}}{c_j}-\sqrt{ZjZ_g\xi}\right)^2\right) \\ & + \frac{1}{\sqrt{\pi}} \left(\frac{v_{jg}}{c_j}- \sqrt{ZjZ_g\xi} \right) \exp\left(-\left(\frac{v_{jg}}{c_j}+\sqrt{ZjZ_g\xi}\right)^2\right) \Bigr\},\end{aligned}$$ where $a$ is the grain radius, $n_j$ the number density of the charged fluid, $v_{jg}$ the relative drift velocity between the grain and the charged fluid, $c_j$ the sound speed of the charged fluid and $\xi = e^2/ak_BT$. The charged fluid can be the ion or electron fluid. [99]{} Anders E., Grevesse N., 1989, GeCoA, 53, 197 Ashmore, I., van Loo, S., Caselli, P., Falle, S. A. E. G., & Hartquist, T. W. 2010, A&A, 511, A41 Bergin, E. A., et al.2000, ApJL, 539, L129 Bergin, E. A., & Snell, R. L. 2002, ApJL, 581, L105 Caselli, P., Hartquist, T. W., & Havnes, O. 1997, A&A, 322, 296 Caselli, P., Keto, E., Pagani, L., et al. 2010, A&A, 521, L29 Chapman J. F., Wardle M., 2006, MNRAS, 371, 513 Draine, B. T., & Salpeter, E. E.  1979, ApJ, 231, 77 Draine, B. T., Roberge, W. G., & Dalgarno, A. 1983, ApJ, 264, 485 Draine, B. T. 2011, EAS Publications Series, 46, 29 Falle, S. A. E. G. 2003, MNRAS, 344, 1210 Flower, D. R., & Pineau des For[ê]{}ts, G. 1995, MNRAS, 275, 1049 Flower, D. R., Pineau des For[ê]{}ts, G., & Hartquist, T. W. 1985, MNRAS, 216, 775 Guillet V., Pineau des For[ê]{}ts G., Jones A. P., 2007, A&A, 476, 263 Guillet, V., Jones, A. P., & Pineau des For[ê]{}ts, G. 2009, A&A, 497, 145 Guillet, V., Pineau des For[ê]{}ts, G., & Jones, A. P. 2011, A&A, 527, A123 Gusdorf A., Cabrit S., Flower D. R., Pineau Des For[ê]{}ts G., 2008a, A&A, 482, 809 Gusdorf, A., Pineau des For[ê]{}ts, G., Cabrit, S., & Flower, D. R. 2008b, A&A, 490, 695 Hartquist, T. W., Dalgarno, A., & Oppenheimer, M. 1980, ApJ, 236, 182 Havnes, O., Hartquist, T. W., & Pilipp, W. 1987, NATO ASIC Proc. 210: Physical Processes in Interstellar Clouds, 389 Hollenbach, D., & McKee, C. F.  1979, ApJS, 41, 555 Jim[é]{}nez-Serra, I., Caselli, P., Mart[í]{}n-Pintado, J., & Hartquist, T. W. 2008, A&A, 482, 549 Jim[é]{}nez-Serra, I., Mart[í]{}n-Pintado, J., Caselli, P., Viti, S., & Rodr[í]{}guez-Franco, A. 2009, ApJ, 695, 149 Mathis, J. S., Rumpl, W., & Nordsieck, K. H. 1977, ApJ, 217, 425 May, P. W., Pineau des For[ê]{}ts, G., Flower, D. R., Field, D., Allan, N. L., & Purton, J. A. 2000, MNRAS, 318, 809 Meyer D. M., Jura M., Cardelli J. A., 1998, ApJ, 493, 222 Millar, T. J., Farquhar, P. R. A., & Willacy, K. 1997, A&AS, 121, 139 Neufeld, D. A., & Melnick, G. J.  1987, ApJ, 322, 266 Neufeld, D. A., Lepp, S., & Melnick, G. J. 1995, ApJS, 100, 132 Pilipp, W., & Hartquist, T. W.  1994, MNRAS, 267, 801 Pilipp, W., Hartquist, T. W., & Havnes, O. 1990, MNRAS, 243, 685 Snow T. P., & Witt A. N., 1996, ApJ, 468, L65 Schilke, P., Walmsley, C. M., Pineau des For[ê]{}ts, G., & Flower, D. R. 1997, A&A, 321, 293 Van Loo, S., Ashmore, I., Caselli, P., Falle, S. A. E. G., & Hartquist, T. W. 2009, MNRAS, 395, 319 Wardle, M. 1998, MNRAS, 298, 507 Whittet, D. C. B., & Duley, W. W.  1991, A&A Rev., 2, 167 Woodall, J., Ag[ú]{}ndez, M., Markwick-Kemper, A. J., & Millar, T. J. 2007, A&A, 466, 1197 \[lastpage\] [^1]: E-mail: [email protected] [^2]: The Hall parameter is the ratio between the gyrofrequency of a charged particle and the collisional momentum transfer frequency of the charged particle with neutrals.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A generalization of the Onsager-Machlup theory from equilibrium to nonequilibrium steady states and its connection with recent fluctuation theorems are discussed for a dragged particle restricted by a harmonic potential in a heat reservoir. Using a functional integral approach, the probability functional for a path is expressed in terms of a Lagrangian function from which an entropy production rate and dissipation functions are introduced, and nonequilibrium thermodynamic relations like the energy conservation law and the second law of thermodynamics are derived. Using this Lagrangian function we establish two nonequilibrium detailed balance relations, which not only lead to a fluctuation theorem for work but also to one related to energy loss by friction. In addition, we carried out the functional integrals for heat explicitly, leading to the extended fluctuation theorem for heat. We also present a simple argument for this extended fluctuation theorem in the long time limit.' author: - 'Tooru Taniguchi and E. G. D. Cohen' title: 'Onsager-Machlup theory for nonequilibrium steady states and fluctuation theorems' --- 0.423cm Introduction ============ Fluctuations play an important role in descriptions of nonequilibrium phenomena. A typical example is the fluctuation-dissipation theorem, which connects transport coefficients to fluctuations in terms of auto-correlation functions. This theorem can be traced back to Einstein’s relation [@E05], Nyquist’s theorem [@J28; @N28], Onsager’s arguments for reciprocal relations [@O31a; @O31b; @C45], etc., and it was established in linear response theory in nonequilibrium statistical mechanics near equilibrium [@G51; @CW51; @K57]. Another example of fluctuation theories is Onsager-Machlup’s fluctuation theory around equilibrium [@H52; @OM53; @MO53]. It is characterized by the usage of a functional integral technique for stochastic linear relaxation processes, and leads to a variational principle known as Onsager’s principle of minimum energy dissipation. Many efforts have been devoted to obtain a generalization, for example, to the cases of nonlinear dynamics [@H76; @H77; @Y71; @HR81; @R89] and nonequilibrium steady states [@BSG01; @BSG02; @G02]. Recently, another approach to fluctuation theory leading to fluctuation theorems has drawn considerable attention in nonequilibrium statistical physics [@ECM93; @ES94; @GC95]. They are asymmetric relations for the distribution functions for work, heat, etc., and they may be satisfied even in far from equilibrium states or for non-macroscopic systems which are beyond conventional statistical thermodynamics. Originally they were proposed for deterministic chaotic dynamics, but they can also be justified for stochastic systems [@K98; @LS99; @C99; @C00]. Moreover, laboratory experiments to check these fluctuation theorems have been made [@CL98; @WSM02; @CGH04; @FM04; @GC05; @SST05]. From our accumulated knowledge on fluctuations, it is meaningful to ask for relations among the different fluctuation theories. It is already known that the fluctuation-dissipation theorem, as well as Onsager’s reciprocal relations, can be derived from fluctuation theorems near equilibrium states [@ECM93; @LS99; @G96]. The heat fluctuation theorem can also be regarded as a refinement of the second law of thermodynamics. The principal aims of this paper are twofold. First, we generalize Onsager and Machlup’s original fluctuation theory around equilibrium to fluctuations around nonequilibrium steady states using the functional integral approach. For this nonequilibrium steady state Onsager-Machlup theory we discuss the energy conservation law (i.e. the analogue of the first law of thermodynamics), the second law of thermodynamics, and Onsager’s principle of minimum energy dissipation. As the second aim of this paper, we discuss fluctuation theorems based on our generalized Onsager-Machlup theory. Since the systems we consider are in a nonequilibrium steady state, the equilibrium detailed balance condition is violated. We propose generalized forms of the detailed balance conditions for nonequilibrium steady states, which we call *nonequilibrium detailed balance relations*, and show that the fluctuation theorem for work can be derived from it. To demonstrate the efficacy of nonequilibrium detailed balance as an origin of fluctuation theorems, we also show another form of nonequilibrium detailed balance, which leads to another fluctuation theorem for energy loss by friction. We also show how a heat fluctuation theorem can be derived from our generalized Onsager-Machlup theory, by carrying out explicitly a functional integral and reducing its derivation to a previous one discussed in Refs. [@ZC03a; @ZC04]. In addition, we give a simple argument leading to the long-time ($t\rightarrow+\infty$) fluctuation theorem for heat, based on the independence between the work distribution and the energy-difference distribution. In this paper, in order to make our arguments as concrete and simple as possible, we apply our theory to a specific nonequilibrium Brownian particle model described by a Langevin equation (cf. [@S98]). It has been used to discuss fluctuation theorems [@MJ99; @TTM02; @ZC03a; @ZC03b; @ZC04], and also to describe laboratory experiments for a Brownian particle captured in an optical trap which moves with a constant velocity through a fluid [@WSM02; @TTM02], as well as for an electric circuit consisting of a resistor and capacitor [@ZCC04; @GC05]. The outline of this paper is as follows. In Sec. \[DraggedBrownianParticle\], we introduce our model and give some of its properties using a functional integral approach. In Sec. \[OnsagerMachlupSteadyStates\], we discuss a generalization of Onsager-Machlup’s fluctuation theory to nonequilibrium steady states, and obtain the energy conservation law, the second law of thermodynamics, i.e. a nonequilibrium steady state thermodynamics, and Onsager’s principle of minimum energy dissipation for such states. In Sec. \[FluctuationTheoremWork\], we introduce the concept of nonequilibrium detailed balance, and obtain a fluctuation theorem for work from it. In Sec. \[FluctuationTheoremFriction\], we discuss another type of nonequilibrium detailed balance, which leads to a fluctuation theorem for energy loss by friction. In Sec. \[FluctuationTheoremHeat\], we sketch a derivation of a fluctuation theorem for heat by carrying out a functional integral and reducing it to the previous derivation [@ZC03a; @ZC04]. In addition, we give a simple argument for the heat fluctuation theorem in the long time limit. In Sec. \[InertiaEffects\], we briefly discuss inertial effects on the fluctuation theorems, which lead to four new fluctuation theorems. In Sec. \[ConclusionsRemarks\], we summarize our results in this paper and discuss some consequences of them. Dragged Particle in a Heat Reservoir {#DraggedBrownianParticle} ==================================== The system considered in this paper is a particle dragged by a constant velocity $v$ in a fluid as a heat reservoir. The dynamics of this system is expressed as a Langevin equation [@noteIIa] $$\begin{aligned} m\frac{d^{2} x_{t}}{dt^{2}} = - \alpha \frac{d x_{t}}{dt} - \kappa \left(x_{t}-v t\right) + \zeta_{t} \label{LangeEq3}\end{aligned}$$ for the particle position $x_{t}$ at time $t$ in the laboratory frame. Here, $m$ is the particle mass, and on the right-hand side of Eq. (\[LangeEq3\]) the first term is the friction force with the friction constant $\alpha$, the second term is the harmonic potential force with the spring constant $\kappa$ to confine the particle, and the third term, due to the coupling to the heat reservoir, is a Gaussian-white noise $\zeta_{t}$, whose first two auto-correlations are given by $$\begin{aligned} \langle \zeta_{t} \rangle &=& 0 , \label{RandomForce1} \\ \langle \zeta_{t_{1}}\zeta_{t_{2}} \rangle &=& \frac{2\alpha}{\beta} \delta(t_{1}-t_{2}) \label{RandomForce2}\end{aligned}$$ with the inverse temperature $\beta$ of the reservoir and the notation $\langle \cdots\rangle$ for an initial ensemble average. The coefficient $2\alpha/\beta$ in Eq. (\[RandomForce2\]) is determined by the fluctuation-dissipation theorem, so that in the case $v=0$ the stationary state distribution function for the dynamics (\[LangeEq3\]) is expressed by a canonical distribution. A schematic illustration for this system is given in Fig. \[figA1system\]. ![Schematic illustration for a particle trapped by a harmonic potential dragged with a constant velocity $v$ in a reservoir. Here, $x$ and $y$ ($0_{l}$ and $0_{c}$) are the axes (the origins) for the laboratory and comoving frame, respectively, in the direction of the motion of the particle. The particle is at the position $y_{t}$ ($x_{t}$) at time $t$ in the comoving (laboratory) frame, respectively, which are related by $y_{t}=x_{t}-vt$. After the relaxation time $\tau$, the system will reach a nonequilibrium steady state.[]{data-label="figA1system"}](figA1system.eps){width="\widthfig"} In this paper, except in Sec. \[InertiaEffects\], we consider the over-damped case in which we neglect the inertial term $m d^{2}x_{t}/dt^{2}$, or assume simply an negligible small mass $m$. Under this over-damped assumption, the Langevin equation (\[LangeEq3\]) can be written as $$\begin{aligned} \frac{d x_{t}}{dt} = - \frac{1}{\tau} \left(x_{t}-v t\right) + \frac{1}{\alpha} \zeta_{t} \label{LangeEq1}\end{aligned}$$ with the relaxation time $\tau$ given by $\tau\equiv\alpha/\kappa$. Equation (\[LangeEq1\]) is for the position $x_{t}$ in the laboratory frame. On the other hand, it is often convenient or simpler to discuss the nonequilibrium dynamics in the comoving frame [@TM04; @ZC03b]. The position $y_{t}$ in the comoving frame for the particle in our model is simply introduced as $$\begin{aligned} y_{t} \equiv x_{t} - v t. \label{ComovPosit1}\end{aligned}$$ Using this position $y_{t}$, Eq. (\[LangeEq1\]) can be rewritten as $$\begin{aligned} \frac{d y_{t}}{dt} = -\frac{1}{\tau} y_{t} - v + \frac{1}{\alpha} \zeta_{t} , \label{LangeEq2}\end{aligned}$$ whose dynamics is invariant under the change $y_{t}\rightarrow -y_{t}$ and $v\rightarrow -v$, noting that the Gaussian-white noise property of $\zeta_{t}$ is not changed into $\zeta_{t}\rightarrow-\zeta_{t}$. Note that in the comoving Langevin equation (\[LangeEq2\]) there is no explicit $t$-dependent term in the dynamical equation, while the laboratory Langevin equation (\[LangeEq1\]) has a $t$-dependence through the term $vt$, meaning Eq. (\[LangeEq2\]) to be a little simpler than Eq. (\[LangeEq1\]). The constant term $-v$ in Eq. (\[LangeEq2\]) expresses all effects of the nonequilibrium steady state in this model. The system described by the Langevin equation (\[LangeEq2\]), or equivalently Eq. (\[LangeEq1\]), approaches a nonequilibrium steady state, because the particle will, for $t>\tau$, move steadily due to the external force that drags it through the fluid. This force is given by $-\kappa y_{t}$, so the work rate $\dot{\mathcal{W}}^{(v)}(y_{t})$ to keep the particle in a steady state is expressed as $$\begin{aligned} \dot{\mathcal{W}}^{(v)}(y_{t}) = - \kappa y_{t} v. \label{WorkRate1}\end{aligned}$$ We note that since $\dot{\mathcal{W}}^{(0)}(y_{t})=0$ for $v=0$, i.e. for the equilibrium state considered by Onsager and Machlup, there is no work done, while in the nonequilibrium steady state for $v\neq 0$ work is done [@note2a]. We consider the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$ of the particle from $y_{0}(\equiv y_{t_{0}})$ at time $t_{0}$ to $y_{t}$ at time $t$, which is introduced as a transition integral kernel for the probability distribution $f(y_{t},t)$ at the position $y_{t}$ at time $t$ as $$\begin{aligned} f(y_{t},t) = \int dy_{0} \; { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } f(y_{0},t_{0}) \label{ProbaDistrY1}\end{aligned}$$ with the initial distribution $f(y_{0},t_{0})$. We can use various analytical techniques, for example the Fokker-Planck equation, whose solution gives the probability distribution $f(y_{t},t)$ [@K92; @R89], to analyze the transition probability for the dynamics expressed by the Langevin equation (\[LangeEq2\]). As one such technique, motivated by Ref. [@OM53; @MO53], we use in this paper the functional integral technique [@R89]. Using this technique, the transition probability is represented as $$\begin{aligned} { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } = \int_{y_{0}}^{y_{t}} \mathcal{D}y_{s} \; \exp\left[ \int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)\right] \label{TransProba1}\end{aligned}$$ where $L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)$ is the Lagrangian function for this stochastic process, defined by $$\begin{aligned} L^{(v)}\!\left(\dot{y}_{s},y_{s}\right) \equiv -\frac{1}{4 D} \left( \dot{y}_{s} +\frac{1}{\tau} y_{s} +v \right)^{2} , \label{Lagra1}\end{aligned}$$ where $D$ is the diffusion constant given by the Einstein relation $D \equiv 1/(\alpha\beta)$. \[We outline a derivation of Eq. (\[TransProba1\]) from Eq. (\[LangeEq2\]) in Appendix \[TransitionProbabilityFunctionalIntegral\].\] Here, the functional integral on the right-hand side of Eq. (\[TransProba1\]) is introduced as $$\begin{aligned} &&\hspace{-0.8cm} \int_{y_{0}}^{y_{t}} \mathcal{D}y_{s} \;X_{t}(\{y_{s}\}) \nonumber \\ && = \lim_{N\rightarrow+\infty} \left(\frac{1}{4\pi D\Delta t_{N}}\right)^{N/2} \nonumber \\ && {\hspace{0.8cm}}\times \int dy_{t_{N-1}} \int dy_{t_{N-2}} \cdots \int dy_{t_{1}} \; X_{t}(\{y_{s}\}) \nonumber \\ \label{FunctInteg1}\end{aligned}$$ for any functional $X_{t}(\{y_{s}\})$, with $t_{n}\equiv t_{0} + n \Delta t_{N}$, $n=1,2,\cdots,N$, $\Delta t_{N}\equiv (t-t_{0})/N$, the initial time $t_{0}$, the final time $t_{N}=t$, the initial position $y_{0}$, and the final position $y_{t}$. Here, we use the symbol $\{y_{s}\}$ in $X_{t}(\{y_{s}\})$ to show that $X_{t}(\{y_{s}\})$ is a functional of $\{y_{s}\}$ with $s\in[t_{0},t]$. It is important to note that from the representation (\[TransProba1\]) of the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$ the functional $\exp[ \int_{t_{0}}^{t}ds\;L^{(v)}(\dot{y}_{s},y_{s})]$ can be regarded as the probability functional density of the path $\{y_{s}\}_{s\in[t_{0},t]}$. For the Lagrangian function (\[Lagra1\]), the functional integral on the right-hand side of Eq. (\[TransProba1\]) can actually be carried out using Eq. (\[FunctInteg1\]), to obtain, by a simple generalization of the well-known equilibrium ($v=0$) case, $$\begin{aligned} { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } & =& \frac{1}{\sqrt{4\pi D\mathcal{T}_{t}}} \nonumber \\ &&\times \exp\left\{- \frac{\left[y_{t}+v\tau -\left(y_{0}+v\tau\right) b_{t}\right]^{2}} {4 D\mathcal{T}_{t}}\right\} , \nonumber \\ \label{TransProba2}\end{aligned}$$ where $b_{t}$ and $\mathcal{T}_{t}$ are defined by $b_{t}\equiv \exp[-(t-t_{0})/\tau]$ and $\mathcal{T}_{t} \equiv (\tau/2) (1 - b_{t}^{2}) $ so that $\mathcal{T}_{t} = t-t_{0} +\mathcal{O}((t-t_{0})^{2})$ [@noteIIA]. Equation (\[TransProba2\]) is simply a well known form of the transition probability for the Smoluchowski process [@R89]. Inserting Eq. (\[TransProba2\]) into Eq. (\[ProbaDistrY1\]), using the normalization condition $\int dy_{0}\;f(y_{0},t_{0})=1$, and taking the limit $t\rightarrow+\infty$, we can show that for an arbitrary initial distribution $f(y_{0},t_{0})$, the probability distribution $f(y_{t},t)$ approaches to a nonequilibrium steady state (ss) distribution: $$\begin{aligned} f_{ss}(y_{t}) \equiv \lim_{t\rightarrow+\infty}f(y_{t},t) = f_{eq}\left(y_{t}+v\tau\right) \label{SteadSolut1}\end{aligned}$$ in the long time limit. Here, $f_{eq}(y)$ is the equilibrium distribution function given by $$\begin{aligned} f_{eq}(y) = \sqrt{\frac{\kappa\beta}{2\pi}} \exp\left[-\beta U(y)\right] \label{EquilDistr1}\end{aligned}$$ with the harmonic potential energy $U(y)\equiv \kappa y^{2}/2$. Equation (\[SteadSolut1\]) implies that the steady state distribution $f_{ss}(y)$ is simply given by the equilibrium canonical distribution $f_{eq}(y)$ by shifting the position $y$ to $y+v\tau$. \[Note that there is no kinetic energy term in the canonical distribution (\[EquilDistr1\]) under the over-damped assumption.\] Equation (\[SteadSolut1\]) implies that the average position of the particle is shifted from the bottom $y=0$ of the harmonic potential in the equilibrium state to the position $y=-v\tau$ in the nonequilibrium steady state. The functional integral approach has already been used to describe relaxation processes to thermal equilibrium with fluctuations and averages by Onsager and Machlup [@OM53; @MO53]. In the next section, we generalize their argument to non-equilibrium steady states for our model, and construct a nonequilibrium steady state thermodynamics. The results in Refs. [@OM53; @MO53] can always be reproduced from our results in Sec. \[OnsagerMachlupSteadyStates\] by taking $v=0$, i.e. in the equilibrium case. In this generalization, we determine the work to sustain the nonequilibrium steady state in the Onsager-Machlup theory, and also give a direct connection between the entropy production rate in the Onsager-Machlup theory and the heat discussed in Ref. [@ZC03a; @ZC04]. Onsager-Machlup Theory for Nonequilibrium Steady states {#OnsagerMachlupSteadyStates} ======================================================= In the generalized Onsager-Machlup theory, the Lagrangian $L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)$ can be written in the form $$\begin{aligned} L^{(v)}(\dot{y}_{s},y_{s}) = -\frac{1}{2k_{B}}\left[ \Phi^{(v)} (\dot{y}_{s}) + \Psi(y_{s}) - \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s}) \right] \nonumber \\ \label{Lagra2}\end{aligned}$$ where $k_{B}$ is the Boltzmann constant, and $\Phi^{(v)} (\dot{y}_{s})$, $\Psi(y_{s})$ and $\dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s})$ are defined by $$\begin{aligned} \Phi^{(v)} (\dot{y}_{s}) &\equiv& \frac{\alpha}{2T}(\dot{y}_{s}+v)^{2} , \label{DissiFunctA1} \\ \Psi(y_{s}) &\equiv& \frac{\alpha}{2T} \left(\frac{y_{s}}{\tau}\right)^{2} , \label{DissiFunctB1} \\ \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s}) &\equiv& -\frac{1}{T}\kappa y_{s}(\dot{y}_{s}+v) , \label{EntroProdu1}\end{aligned}$$ respectively, with the temperature $T\equiv (k_{B}\beta)^{-1}$. These functions $\Phi^{(v)} (\dot{y}_{s})$ and $\Psi(y_{s})$ are called dissipation functions, while we call $\dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s})$ the entropy production rate. In the next subsections \[HeatEntropyBalance\] and \[DissipationFunctions\], we discuss the physical meaning of these quantities, and justify their names. Heat and energy balance equations {#HeatEntropyBalance} --------------------------------- Using the entropy production rate $\dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s})$, we introduce the heat $\mathcal{Q}^{(v)}_{t}(\{y_{s}\})$ produced by the system in the time-interval $[t_{0},t]$ as $$\begin{aligned} \mathcal{Q}^{(v)}_{t}(\{y_{s}\}) \equiv T \int_{t_{0}}^{t}ds\; \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s}) . \label{Heat1}\end{aligned}$$ On the other hand, the work $\mathcal{W}_{t}^{(v)}(\{y_{s}\})$ done on the system to sustain it in a steady state is given by $$\begin{aligned} \mathcal{W}_{t}^{(v)}(\{y_{s}\}) \equiv \int_{t_{0}}^{t}ds\; \dot{\mathcal{W}}^{(v)}(y_{s}) . \label{Work1}\end{aligned}$$ using the work rate (\[WorkRate1\]). The heat (\[Heat1\]) and the work (\[Work1\]) are related by $$\begin{aligned} \mathcal{Q}^{(v)}_{t}(\{y_{s}\}) = \mathcal{W}_{t}^{(v)}(\{y_{s}\}) - \Delta \mathcal{U}(y_{t},y_{0}) \label{EnergyBalan1}\end{aligned}$$ with the internal (potential) energy difference $$\begin{aligned} \Delta \mathcal{U}(y_{t},y_{0}) \equiv U(y_{t}) - U(y_{0}) \label{EnergDiffe1}\end{aligned}$$ at times $t$ and $t_{0}$. The relation (\[EnergyBalan1\]) is nothing but the energy conservation law satisfied even by fluctuating quantities. It may be noted that Eq. (\[EnergyBalan1\]) is used as a “definition” of heat in Ref. [@ZC03a; @ZC04], while here it appears as a consequence of our nonequilibrium Onsager-Machlup theory. In other words, our generalization of the Onsager-Machlup theory gives a justification of the heat used in Ref. [@ZC03a; @ZC04]. For other attempts to justify the energy conservation law in stochastic processes using a Langevin equation or a master equation, see Refs. [@S98; @C00]. Dissipation functions and the entropy production {#DissipationFunctions} ------------------------------------------------ First, it follows from Eqs. (\[DissiFunctA1\]) and (\[EntroProdu1\]) that $$\begin{aligned} \Phi^{(-v)} (-\dot{y}_{s}) &=& \Phi^{(v)} (\dot{y}_{s}) , \\ \dot{S}^{(-v)}(-\dot{y}_{s},y_{s}) &=& - \dot{S}^{(v)}(\dot{y}_{s},y_{s}) , $$ implying that the dissipation function $\Phi^{(v)} (\dot{y}_{s})$ \[as well as $\Psi(y_{s})$ by Eq. (\[DissiFunctB1\])\] is invariant under the time-reversal changes $\dot{y}_{s}\rightarrow -\dot{y}_{s}$ and $v\rightarrow -v$, while the entropy production rate $\dot{S}^{(v)}(\dot{y}_{s},y_{s})$ is anti-symmetric under these changes. It is also obvious from Eqs. (\[DissiFunctA1\]) and (\[DissiFunctB1\]) that $$\begin{aligned} \Phi^{(v)} (\dot{y}_{s}) &\geq& 0 , \label{DissiFunctPosit1}\\ \Psi(y_{s}) &\geq& 0 , \label{DissiFunctPosit2}\end{aligned}$$ namely, that the dissipation functions are non-negative. One should also notice that by the definitions (\[DissiFunctA1\]) and (\[DissiFunctB1\]) the dissipation functions $\Phi^{(v)} (\dot{y}_{s})$ and $\Psi(y_{s})$ are proportional to the friction constant $\alpha$. Second, from Eqs. (\[RandomForce1\]) and (\[LangeEq2\]), the ensemble average $\langle y_{s}\rangle$ of the particle position $y_{s}$ satisfies $$\begin{aligned} \langle \dot{y}_{s} \rangle = -\frac{1}{\tau} \langle y_{s} \rangle - v , \label{LangeEq2Ave}\end{aligned}$$ with the time-derivative $\dot{y}_{s} \equiv dy_{s}/ds$ of $y_{s}$, leading to $\langle y_{s} \rangle = -v\tau + (\langle y_{0} \rangle +v\tau) \exp[-(s-t_{0})/\tau]$. Using this average position $\langle y_{s} \rangle$ and the average velocity $\langle \dot{y}_{s} \rangle$, it follows from Eqs. (\[DissiFunctA1\]), (\[DissiFunctB1\]) and (\[LangeEq2Ave\]) that $$\begin{aligned} \Phi^{(v)} (\langle \dot{y}_{s} \rangle) = \Psi(\langle y_{s} \rangle) . \label{DissiFunctRelat1}\end{aligned}$$ Namely, the two dissipation functions $\Phi^{(v)} (\dot{y}_{s})$ and $\Psi(y_{s})$ have the same value for $\langle y_{s} \rangle$ and $\langle \dot{y}_{s} \rangle$, although $\Phi^{(v)} (\dot{y}_{s})$ is a function of $\dot{y}_{s}$ and $\Psi(y_{s})$ is a function of $y_{s}$. Moreover, from Eqs. (\[DissiFunctA1\]), (\[EntroProdu1\]), (\[DissiFunctPosit1\]), (\[LangeEq2Ave\]) and (\[DissiFunctRelat1\]) we derive $$\begin{aligned} \dot{S}^{(v)}(\langle \dot{y}_{s} \rangle, \langle y_{s} \rangle) = 2\Phi^{(v)} (\langle \dot{y}_{s} \rangle) = 2\Psi (\langle y_{s} \rangle) \geq 0 , \label{SeconLaw1}\end{aligned}$$ namely, the function $2\Phi^{(v)} (\langle \dot{y}_{s} \rangle)$ \[as well as $2\Psi(\langle y_{s} \rangle)$\] gives the entropy production rate $\dot{S}^{(v)}(\langle \dot{y}_{s} \rangle,\langle y_{s} \rangle)$, justifying the name “dissipation function” for $\Phi^{(v)} (\dot{y}_{s})$ and $\Psi(y_{s})$. The inequality in (\[SeconLaw1\]) is the second law of thermodynamics in the Onsager-Machlup theory. Onsager’s principle of minimum energy dissipation and the most probable path {#OnsagerPrincipleMostProbablePath} ---------------------------------------------------------------------------- Equation (\[LangeEq2Ave\]) for the average $\langle y_{s}\rangle$ of the particle position can be derived from the variational principle $$\begin{aligned} \Phi^{(v)} (\dot{y}_{s}) + \Psi(y_{s}) - \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s}) = \mbox{minimum} , \label{VariaPrinc0}\end{aligned}$$ without using the Langevin equation (\[LangeEq2\]). This can be proved by using that $\Phi^{(v)} (\dot{y}_{s}) + \Psi(y_{s}) - \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s}) = -2k_{B}L^{(v)}(\dot{y}_{s},y_{s}) \geq 0$ and $L^{(v)}(\langle \dot{y}_{s} \rangle,\langle y_{s} \rangle) =0$, so that the left-hand side of Eq. (\[VariaPrinc0\]) takes its minimum value for $y_{s} = \langle y_{s} \rangle$, i.e. for the *average* path, which is used in Eq. (\[SeconLaw1\]). Equation (\[VariaPrinc0\]) is called the Onsager’s principle of minimum energy dissipation, and is proposed as a generalization of the maximal entropy principle for equilibrium thermodynamics [@OM53; @MO53; @TM57; @O31a; @O31b; @H76; @H77]. Another result in the Onsager-Machlup theory as a variational principle is that we can justify a variational principle to extract the special path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$, the so-called *most probable path*, which give the most significant contribution in the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$. By the expression (\[TransProba1\]) for the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$, the most probable path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$ is determined by the maximal condition on $\int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)$, in other words, the path $\{y_{s}\}_{s\in[t_{0},t]}$ satisfying $$\begin{aligned} \int_{t_{0}}^{t}ds\; \left[\Phi^{(v)} (\dot{y}_{s}) + \Psi(y_{s}) - \dot{\mathcal{S}}^{(v)}(\dot{y}_{s},y_{s})\right] = \mbox{minimum} , \nonumber \\ \label{VariaPrinc1}\end{aligned}$$ under fixed values of $y_{0}$ and $y_{t}$, noting the expression (\[Lagra2\]) for the Lagrangian function $L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)$. The condition (\[VariaPrinc1\]), or equivalently the maximal condition of $\int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)$ implies the variational principle $\delta \int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right) = 0$ for the path $\{y_{s}\}_{s\in[t_{0},t]}$, leading to the Euler-Lagrange equation [@LL69] $$\begin{aligned} \frac{d}{ds} \frac{\partial L^{(v)}(\dot{y}^{*}_{s},y^{*}_{s})} {\partial \dot{y}^{*}_{s}} - \frac{\partial L^{(v)}(\dot{y}^{*}_{s},y^{*}_{s})} {\partial y^{*}_{s}} = 0 \label{EularLagra1}\end{aligned}$$ for the most probable path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$. (The most probable path can also be analyzed by the Hamilton-Jacobi equation [@BSG01; @BSG02].) Inserting Eq. (\[Lagra1\]) into Eq. (\[EularLagra1\]) we obtain $$\begin{aligned} \frac{d^{2} y^{*}_{s}}{ds^{2}} = \frac{y^{*}_{s} + v\tau}{\tau^{2}} \label{EularLagra2}\end{aligned}$$ for our model. It is interesting to note that the ensemble average $\langle y_{s}\rangle$ also satisfies Eq. (\[EularLagra2\]), because $d^{2}\langle y_{s}\rangle/ds^{2} = (\langle y_{s}\rangle + v\tau)/\tau^{2}$ from Eq. (\[LangeEq2Ave\]). In general, the most probable path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$ with the conditions $y_{t_{0}}^{*}=y_{0}$ and $y_{t}^{*}=y_{t}$ contains a superposition of the forward average path $\Upsilon_{s}^{[+]} \equiv \mathcal{A}_{+} \exp(-s/\tau) -v\tau$ (like the average path $\langle y_{t}\rangle$) and its time-reversed path $\Upsilon_{s}^{[-]} \equiv \mathcal{A}_{-} \exp(s/\tau) +v\tau$, namely $$\begin{aligned} y^{*}_{s} = \Upsilon_{s}^{[+]} + \Upsilon_{s}^{[-]} + \mathcal{A}_{0} $$ where $\mathcal{A}_{\pm}$ and $\mathcal{A}_{0} (=-v\tau)$ are time-independent constants and are determined by the conditions $y_{t_{0}}^{*}=y_{0}$ and $y_{t}^{*}=y_{t}$ [@noteIIIC]. We now discuss a relation of the Onsager-Machlup theory with Einstein’s fluctuation formula [@Lan59]. We note that $$\begin{aligned} L^{(v)}\! \left(\dot{\Upsilon}_{s}^{[+]}, \Upsilon_{s}^{[+]}\right) &=& 0, \label{LagraForwaPath1}\\ L^{(v)}\! \left(\dot{\Upsilon}_{s}^{[-]}, \Upsilon_{s}^{[-]}\right) &=& \frac{1}{k_{B}} \dot{S}^{(v)}\!\left(\dot{\Upsilon}_{s}^{[-]}, \Upsilon_{s}^{[-]}\right) \label{LagraBackPath1}\end{aligned}$$ with $\dot{\Upsilon}_{s}^{[\pm]}\equiv d \Upsilon_{s}^{[\pm]}/ds$. Here, we used the equation $\pm d \Upsilon_{s}^{[\pm]}/ds = -\Upsilon_{s}^{[\pm]}/\tau \mp v$. Using the most probable path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$ satisfying the conditions $y_{t_{0}}^{*}=y_{0}$ and $y_{t}^{*}=y_{t}$, we can approximate the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$ as $$\begin{aligned} { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } \approx \exp\left[ \int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}^{*}_{s},y^{*}_{s}\right)\right] , \label{OMformu1}\end{aligned}$$ apart from a normalization factor. This is analogous to the classical approximation for the wave function in the Feynman path-integral approach in quantum mechanics [@FH65]. It is meaningful to mention that in the case of relaxation to an equilibrium state ($v=0$), Eq. (\[OMformu1\]) becomes $$\begin{aligned} &&\left.{ {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) }\right|_{v=0} \nonumber \\ &&{\hspace{0.8cm}}\approx \exp\left[ \frac{1}{k_{B}}\int_{t_{0}}^{t}ds\; \left. \dot{S}^{(v)}\!\left(\dot{\Upsilon}_{s}^{[-]}, \Upsilon_{s}^{[-]}\right)\right|_{v=0}\right] , \;\;\; \label{OMformu2}\end{aligned}$$ noting Eqs. (\[LagraBackPath1\]) and $L^{(v)}(\dot{y}^{*}_{s},y^{*}_{s})|_{v=0} =L^{(v)}(\dot{\Upsilon}_{s}^{[-]},\Upsilon_{s}^{[-]})|_{v=0}$ [@TM57]. Here, we remark that in Eq. (\[OMformu2\]) the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }|_{v=0}$ is expressed by the time-reversed path $\{\dot{\Upsilon}_{s}^{[-]}\}_{s\in[t_{0},t]}$ only. The quantity $\int_{t_{0}}^{t}ds\; \dot{S}$ gives the entropy, so that Eq. (\[OMformu2\]) corresponds to Einstein’s fluctuation formula in equilibrium, i.e. for $v=0$. Fluctuation Theorem for Work {#FluctuationTheoremWork} ============================ In the preceding section \[OnsagerMachlupSteadyStates\], by generalizing the Onsager-Machlup theory to nonequilibrium steady states, we discussed fluctuating quantities whose averages give thermodynamic quantities, like work and heat, etc. Since these quantities fluctuate, it is important to discuss nonequilibrium characteristics of their fluctuations. In the remaining part of this paper, we discuss such characteristics using distribution functions of work, heat, etc., by the functional integral technique. For this discussion, generalized versions of the equilibrium detailed balance, which we will call nonequilibrium detailed balance relations, play an important role, leading to fluctuation theorems. Fluctuation theorems are for nonequilibrium behavior in the case of $v\neq 0$, so there is no counterpart to the contents of this paper in Onsager and Machlup’s original papers where $v=0$ always. Nonequilibrium detailed balance relation ---------------------------------------- The equilibrium detailed balance condition expresses a reversibility of the transition probability between any two states in the equilibrium state, and is known as a physical condition for the system to relax to an equilibrium state [@K92; @R89]. This condition has to be modified for the nonequilibrium steady state, because the system does not relax to an equilibrium state but is sustained in an nonequilibrium state by an external force. This modification, or violation, of the equilibrium detailed balance in the nonequilibrium steady state is expressed quantitatively for work by $$\begin{aligned} && e^{-\beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})} e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)} f_{eq}(y_{0}) \nonumber \\ && {\hspace{0.8cm}}= f_{eq}(y_{t}) \;e^{\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(-\dot{y}_{s},y_{s}\right)} \label{DetaiBalan1}\end{aligned}$$ in our path-integral approach, which is derived from Eqs. (\[Lagra1\]), (\[EquilDistr1\]) and (\[Work1\]). We call Eq. (\[DetaiBalan1\]) a *nonequilibrium detailed balance relation* for nonequilibrium steady states in this paper [@noteIVa]. Equation (\[DetaiBalan1\]) reduces to the equilibrium detailed balance condition in the case $v=0$, because from Eqs. (\[TransProba1\]), (\[DetaiBalan1\]) and $\mathcal{W}_{t}^{(0)}(\{y_{s}\})$ $=0$, we can derive the well-known equilibrium detailed balance condition $$\begin{aligned} \left.{ {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) }\right|_{v=0} f_{eq}(y_{0}) = \left.{ {F}\!\left(\begin{array}{c} \!\! y_{0} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{t} \!\! \\ \!\! t_{0} \!\! \end{array}\right) }\right|_{v=0} f_{eq}(y_{t}) \label{DetaiBalan2}\end{aligned}$$ for the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$ in equilibrium. As discussed in Sec. \[DraggedBrownianParticle\], the term $\exp[\int_{t_{0}}^{t}ds\; L^{(v)}(\dot{y}_{s},y_{s})]$ on the left-hand side of Eq. (\[DetaiBalan1\]) is the probability functional for the forward path $\{y_{s}\}_{s\in[t_{0},t]}$. On the other hand, the term $\exp[\int_{t_{0}}^{t}ds\; L^{(-v)}(-\dot{y}_{s},y_{s})]$ on the right-hand side of Eq. (\[DetaiBalan1\]) is the probability functional of the time-reversed path. Therefore, Eq. (\[DetaiBalan1\]) means that we need the work $\mathcal{W}_{t}^{(v)}(\{y_{s}\})$ so that the particle, dragged from an equilibrium state with the velocity $v$, can move along a path $\{y_{s}\}_{s\in[t_{0},t]}$ and return back to the equilibrium state along its time-reversed path with the reversed dragging velocity $-v$. Such an additional work appears as a canonical distribution type of barrier $\exp[-\beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})]$ for the transition probability on the left-hand side of Eq. (\[DetaiBalan1\]). It should be emphasized that Eq. (\[DetaiBalan1\]) is satisfied not only for the most probable path but for *any* path $\{y_{s}\}_{s\in[t_{0},t]}$, which is crucial for the derivation of the work fluctuation theorem as we will discuss in the next subsection \[WorkFluctuationTheorem\]. Work fluctuation theorem {#WorkFluctuationTheorem} ------------------------ Now, we discuss the distribution of work. For simplicity of notation, we consider the dimensionless work $\beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})$ and its distribution $P_{w}(W,t)$ given by $$\begin{aligned} P_{w}(W,t) = { \left\langle\!\!\!\left\langle \delta\!\left(W - \beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})\right) \right\rangle\!\!\!\right\rangle_{t} } . \label{WorkDistr1}\end{aligned}$$ Here, ${ \left\langle\!\!\!\left\langle \cdots \right\rangle\!\!\!\right\rangle_{t} }$ means a functional average over all possible paths $\{y_{s}\}_{s\in[t_{0},t]}$, as well as integrals over the initial and final points of the path: $$\begin{aligned} { \left\langle\!\!\!\left\langle \; X_{t}(\{y_{s}\})\; \right\rangle\!\!\!\right\rangle_{t} } &\equiv& \int dy_{t} \! \int_{y_{0}}^{y_{t}} \mathcal{D}y_{s} \!\int dy_{0} \; e^{\int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)} \nonumber \\ &&{\hspace{0.8cm}}\times \; f(y_{0},t_{0}) \; X_{t}(\{y_{s}\}) \label{PathAvera1}\end{aligned}$$ for any functional $X_{t}(\{y_{s}\})$. It is convenient to express the work distribution $P_{w}(W,t)$ as a Fourier transform $$\begin{aligned} P_{w}(W,t) = \frac{1}{2\pi}\int_{-\infty}^{+\infty}d\lambda \; e^{i\lambda W} \mathcal{E}_{w}^{(v)}(i\lambda,t) \label{WorkDistr2}\end{aligned}$$ using the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$ defined by $$\begin{aligned} \mathcal{E}_{w}^{(v)}(\lambda,t) \equiv { \left\langle\!\!\!\left\langle e^{-\lambda\beta \mathcal{W}_{t}^{(v)}(\{y_{s}\})} \right\rangle\!\!\!\right\rangle_{t} } , \label{EFunctWork1}\end{aligned}$$ which may be regarded as a generating functional of the dimensionless work. It follows from Eqs. (\[PathAvera1\]), (\[EFunctWork1\]), $L^{(-v)}(\dot{y}_{s},y_{s}) = L^{(v)}(-\dot{y}_{s},-y_{s})$ and $\mathcal{W}_{t}^{(-v)}(\{y_{s}\}) =\mathcal{W}_{t}^{(v)}(\{-y_{s}\})$, that the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$ is invariant under the change $v\rightarrow -v$, namely $$\begin{aligned} \mathcal{E}_{w}^{(-v)}(\lambda,t) = \mathcal{E}_{w}^{(v)}(\lambda,t), \label{EFunctWorkRelat1}\end{aligned}$$ if the initial distribution $f(y_{0},t_{0})$ is invariant under spatial reflection, namely $f(-y_{0},t_{0})|_{-v}=f(y_{0},t_{0})|_{v}$. This is simply due to an invariance under space inversion of our model. In addition, as shown in Appendix \[TransientFluctuationTheoremWork\], the nonequilibrium detailed balance relation (\[DetaiBalan1\]) imposes the relation $\mathcal{E}_{w}^{(v)}(\lambda,t) = \mathcal{E}_{w}^{(-v)}(1-\lambda,t)$ on the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$. Combination of this relation with Eq. (\[EFunctWorkRelat1\]) then leads to $$\begin{aligned} \mathcal{E}_{w}^{(v)}(\lambda,t) = \mathcal{E}_{w}^{(v)}(1-\lambda,t) \label{FluctTheorWork1}\end{aligned}$$ for the equilibrium initial distribution $f(y_{0},t_{0}) =f_{eq}(y_{0})$, as a form discussed in Ref. [@LS99]. Equation (\[FluctTheorWork1\]) is equivalent to the relation $$\begin{aligned} \frac{P_{w}(W,t)}{P_{w}(-W,t)} = \exp(W). \label{FluctTheorWork2}\end{aligned}$$ for the work distribution $P_{w}(W,t)$, which is known as the transient fluctuation theorem [@ZC03b; @ES02; @noteIIIB]. \[See Appendix \[TransientFluctuationTheoremWork\] for a derivation of Eq. (\[FluctTheorWork2\]) from Eq. (\[FluctTheorWork1\]).\] As shown in Eq. (\[FluctTheorWork2\]), the transient fluctuation theorem is satisfied for any time being an identity [@CG99], but it requires that the system is in the equilibrium state at the initial time $t_{0}$. Therefore, one may ask what happens to the fluctuation theorem if we choose a nonequilibrium steady state, or any other state, as the initial condition. In the next subsection \[FunctionalCalculationWorkDistribution\], we calculate the work distribution function $P_{w}(W,t)$ explicitly by carrying out the functional integral on the right-hand side of Eq. (\[WorkDistr1\]) via Eq. (\[PathAvera1\]), in order to answer this question. Functional integral calculation of the work distribution function {#FunctionalCalculationWorkDistribution} ----------------------------------------------------------------- To calculate the work distribution function $P_{w}(W,t)$, we note first that the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$, connected to $P_{w}(W,t)$ by Eq. (\[WorkDistr2\]), can be rewritten as $$\begin{aligned} \mathcal{E}_{w}^{(v)}(\lambda,t) = \int dy_{t} \! \int dy_{0} \; \mathcal{F}(y_{t},y_{0};\lambda) f(y_{0},t_{0}) \label{EFunctWork2}\end{aligned}$$ by Eqs. (\[PathAvera1\]) and (\[EFunctWork1\]). Here, $\mathcal{F}(y_{t},y_{0};\lambda)$ is defined by $$\begin{aligned} &&\!\!\!\mathcal{F}(y_{t},y_{0};\lambda) \nonumber \\ && \equiv \int_{y_{0}}^{y_{t}} \mathcal{D}y_{s} \; \exp\left\{\int_{t_{0}}^{t}ds\; \left[ L^{(v)}\!\left(\dot{y}_{s},y_{s}\right) -\lambda\beta\dot{\mathcal{W}}^{(v)}(y_{s})\right]\right\} . \nonumber \\ \label{EFunctWorkTrnas1}\end{aligned}$$ Equation (\[EFunctWorkTrnas1\]) may be regarded as a constrained transition probability for the modified Lagrangian $L^{(v)}\!\left(\dot{y}_{s},y_{s}\right) -\lambda\dot{\mathcal{W}}^{(v)}(y_{s})$ [@noteIVc]. Here, the $v$-dependence of the function $\mathcal{F}(y_{t},y_{0};\lambda)$ has been suppressed, as it is in the rest of the paper. To calculate the function $\mathcal{F}(y_{t},y_{0};\lambda)$, we introduce the solution ${\tilde{y}^{*}\!\!}_{t}$ of the modified Euler-Lagrange equation for the modified Lagrangian $L^{(v)}\!\left(\dot{y}_{s,}y_{s}\right) -\lambda\dot{\mathcal{W}}^{(v)}(y_{s})$, namely $$\begin{aligned} \frac{d}{ds} \frac{\partial L^{(v)}({\dot{{\tilde{y}^{*}\!\!}}{\,}}_{s}, {\tilde{y}^{*}\!\!}_{s})} {\partial {\dot{{\tilde{y}^{*}\!\!}}{\,}}_{s}} - \frac{\partial L^{(v)}({\dot{{\tilde{y}^{*}\!\!}}{\,}}_{s},{\tilde{y}^{*}\!\!}_{s})} {\partial {\tilde{y}^{*}\!\!}_{s}} + \lambda\beta\frac{\partial \dot{\mathcal{W}}^{(v)}({\tilde{y}^{*}\!\!}_{s})} {\partial {\tilde{y}^{*}\!\!}_{s}} = 0 \nonumber \\ \label{EularLagra2Modif1}\end{aligned}$$ under the conditions ${\tilde{y}^{*}\!\!}_{t}=y_{t}$ and $({\tilde{y}^{*}\!\!}_{0}\equiv){\tilde{y}^{*}\!\!}_{t_{0}}=y_{0}$. By solving Eq. (\[EularLagra2Modif1\]) we obtain $$\begin{aligned} {\tilde{y}^{*}\!\!}_{s} &=& -(1-2\lambda)v\tau + A_{t-t_{0}}^{((1-2\lambda)v)}(y_{t},y_{0}) \exp\left(-\frac{t-s}{\tau}\right) \nonumber \\ &&{\hspace{0.8cm}}+ A_{t-t_{0}}^{((1-2\lambda)v)}(y_{0},y_{t}) \exp\left(-\frac{s-t_{0}}{\tau}\right) \label{EularLagra2Solve1}\end{aligned}$$ where $A_{t-t_{0}}^{(v)}(y_{t},y_{0})$ is defined by $$\begin{aligned} A_{t-t_{0}}^{(v)}(y_{t},y_{0}) \equiv\frac{(y_{t} +v\tau) - (y_{0} +v\tau) b_{t} } {1-b_{t}^{2}} . \label{ConstA1} \end{aligned}$$ \[See Appendix \[CalculationWorkDistribution\] for a derivation of Eq. (\[EularLagra2Solve1\]).\] The path $\{{\tilde{y}^{*}\!\!}_{s}\}_{s\in[t_{0},t]}$ becomes the most probable path $\{y_{s}^{*}\}_{s\in[t_{0},t]}$ in the case of $\lambda=0$ in which Eq. (\[EularLagra2Modif1\]) is equivalent to Eq. (\[EularLagra1\]). Using the solution ${\tilde{y}^{*}\!\!}_{s}$ of the modified Euler-Lagrange equation (\[EularLagra2Solve1\]), we obtain $$\begin{aligned} \mathcal{F}(y_{t},y_{0};\lambda) &=& e^{\int_{t_{0}}^{t} ds \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{\mathcal{W}}^{(v)}({\tilde{y}^{*}\!\!}_{s})\right]} \nonumber \\ &&\times \int_{\tilde{z}_{0}}^{\tilde{z}_{t}} \mathcal{D}\tilde{z}_{s} \; e^{\int_{t_{0}}^{t} ds \; L^{(0)}(\dot{\tilde{z}}_{s},\tilde{z}_{s})} \label{EFunctWorkTrnas2}\end{aligned}$$ for the function $\mathcal{F}(y_{t},y_{0};\lambda)$, where $\tilde{z}_{s}$ is introduced as $$\begin{aligned} \tilde{z}_{s} \equiv y_{s} - {\tilde{y}^{*}\!\!}_{s}, \label{VariaZ1}\end{aligned}$$ namely the deviation of $y_{s}$ from ${\tilde{y}^{*}\!\!}_{s}$, satisfying the boundary conditions $$\begin{aligned} \tilde{z}_{t} = \tilde{z}_{0} = 0 \label{VariaZBound1}\end{aligned}$$ because ${\tilde{y}^{*}\!\!}_{0}=y_{0}$ and ${\tilde{y}^{*}\!\!}_{t}=y_{t}$, where $\dot{\tilde{z}}_{s}\equiv d\tilde{z}_{s}/ds$ and $\tilde{z}_{0}\equiv \tilde{z}_{t_{0}}$. \[See Appendix \[CalculationWorkDistribution\] for a derivation of Eq. (\[EFunctWorkTrnas2\]).\] For the functional integral for $\tilde{z}_{s}$ on the right-hand side of Eq. (\[EFunctWorkTrnas2\]) we obtain $$\begin{aligned} \int_{\tilde{z}_{0}}^{\tilde{z}_{t}} \mathcal{D}\tilde{z}_{s} \; \exp\left[\int_{t_{0}}^{t} ds \; L^{(0)}(\dot{\tilde{z}}_{s},\tilde{z}_{s})\right] =\frac{1}{\sqrt{4\pi D \mathcal{T}_{t}}} , \label{PathInteg1}\end{aligned}$$ noting that the Lagrangian $L^{(0)}(\dot{\tilde{z}}_{s},\tilde{z}_{s})$ on the left-hand side of Eq. (\[PathInteg1\]) is for the case of $v=0$. \[See Appendix \[CalculationWorkDistribution\] for a derivation of Eq. (\[PathInteg1\]).\] Inserting Eqs. (\[EularLagra2Solve1\]) and (\[PathInteg1\]) into Eq. (\[EFunctWorkTrnas2\]), the function $\mathcal{F}(y_{t},y_{t_{0}};\lambda)$ is represented as $$\begin{aligned} \mathcal{F}(y,y_{0};\lambda) &=& \frac{1}{\sqrt{4\pi D \mathcal{T}_{t}}} \exp\left\{- \frac{\left[(y +v\tau) - (y_{0} +v\tau) b_{t} \right]^{2}} {4D\mathcal{T}_{t}} + \lambda \alpha \beta v \left(y + y_{0}\right) \frac{1-b_{t}} {1+b_{t}} \right.\nonumber \\ && {\hspace{0.8cm}}{\hspace{0.8cm}}{\hspace{0.8cm}}\left. - \lambda (1-\lambda) \alpha \beta v^{2} \left( t-t_{0} - 2\tau \frac{1-b_{t}} {1+b_{t}} \right) \right\} \label{FunctCalF1}\end{aligned}$$ Using Eq. (\[EFunctWork2\]) and (\[FunctCalF1\]) and carrying out the integration over $y$ we obtain $$\begin{aligned} \mathcal{E}_{w}^{(v)}(\lambda,t) = e^{ - \lambda (1-\lambda) \alpha \beta v \Omega_{t}} \int dy_{0}\; f\!\left(y_{0},t_{0}\right) e^{ \lambda \alpha\beta v \left[y_{0} - \frac{v\tau}{2} \left(1-b_{t}\right)\right] \left(1 - b_{t}\right)} \label{EFunctWork3}\end{aligned}$$ where $\Omega_{t}$ is defined by $$\begin{aligned} \Omega_{t} \equiv v \left\{ t-t_{0} - \frac{\tau}{2} \left[ 4 - \left(1-b_{t}\right)^{2} \right]\frac{1-b_{t}} {1+b_{t}} \right\} . \label{FunctOmega1}\end{aligned}$$ Equation (\[EFunctWork3\]) gives an explicit expression of the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$ for any initial distribution $f(y_{0},t_{0})$. Inserting Eq. (\[EFunctWork3\]) into Eq. (\[WorkDistr2\]), we obtain $$\begin{aligned} P_{w}(W,t) = \frac{1}{\sqrt{ 4\pi\alpha\beta v \Omega_{t}}} \int dy_{0}\; f\!\left(y_{0} ,t_{0} \right) \exp\left\{-\frac{\left\{W-\alpha\beta v \left[ \Omega_{t} -\left(1-b_{t}\right) \left[y_{0}-\frac{v\tau}{2} \left(1-b_{t}\right)\right] \right]\right\}^{2}} {4\alpha\beta v \Omega_{t}}\right\} \label{WorkDistr3}\end{aligned}$$ as a concrete form of the work distribution $P_{w}(W,t)$ satisfied by *any* initial distribution $f(y_{0},t_{0})$. From Eq. (\[WorkDistr3\]), the asymptotic form of the work distribution function $P_{w}(W,t)$ is given by $$\begin{aligned} P_{w}(W,t) \stackrel{t\rightarrow+\infty}{\sim} \frac{1}{\sqrt{4\pi \alpha\beta v^{2} t}} \exp\left[-\frac{ \left(W-\alpha\beta v^{2} t\right)^{2}} {4\alpha\beta v^{2} t} \right] \label{DistrWorkLongTime1}\end{aligned}$$ where we used the asymptotic relation $\Omega_{t}\stackrel{t\rightarrow+\infty}{\sim} vt$ by Eq. (\[FunctOmega1\]), and the normalization condition $\int dy_{0}\; f(y_{0} ,t_{0})=1$. It follows immediately from Eq. (\[DistrWorkLongTime1\]) that $$\begin{aligned} \lim_{t\rightarrow+\infty} \frac{P_{w}(W,t)}{P_{w}(-W,t)} = \exp(W) . \label{FluctTheorWork3}\end{aligned}$$ Therefore, the work fluctuation theorem is satisfied for *any* initial condition (including the nonequilibrium steady state initial distribution) in the very long time limit. Fluctuation Theorem for Friction {#FluctuationTheoremFriction} ================================ In Sec. \[FluctuationTheoremWork\], we emphasized a close relation between the nonequilibrium detailed balance relation like Eq. (\[DetaiBalan1\]) and the work fluctuation theorem (\[FluctTheorWork3\]). To show the usefulness of such a relation we discuss in this section another type of nonequilibrium detailed balance relation, and show that it leads to another fluctuation theorem related to the energy loss by friction. We consider the rate of energy loss caused by the friction force $-\alpha \dot{y}_{s}$ in the comoving frame. It is given by $- \alpha \dot{y}_{s} v$, so the total energy loss $\mathcal{R}_{t}^{(v)}$ by friction in the time interval $[t_{0},t]$ is $$\begin{aligned} \mathcal{R}_{t}^{(v)}(y_{t},y_{0}) = \int_{t_{0}}^{t}ds\; \left(- \alpha \dot{y}_{s} \right) v = - \alpha v (y_{t}-y_{0}) \label{FrictLoss}\end{aligned}$$ using $y_{0}\equiv y_{t_{0}}$. It may be noted that the energy loss $\mathcal{R}_{t}^{(v)}(y_{t},y_{0})$ by friction is determined by the particle positions at the times $t_{0}$ and $t$ only, different from the work $\mathcal{W}_{t}^{(v)}(\{y_{s}\})$, which is determined by the particle positions at *all* times $s\in[t_{0},t]$. Our starting point to discuss the fluctuation theorem for friction is the relation $$\begin{aligned} && e^{-\beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0})} e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)} f_{eq}(y_{0}) \nonumber \\ && {\hspace{0.8cm}}= f_{eq}(y_{t}) \;e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(-\dot{y}_{s},y_{s}\right)} \label{DetaiBalan3}\end{aligned}$$ derived straight-forwardly from Eqs. (\[Lagra1\]), (\[EquilDistr1\]) and (\[FrictLoss\]). It must be noted that there is a difference between Eq. (\[DetaiBalan3\]) and Eq. (\[DetaiBalan1\]) in the change (or no change) of sign of the dragging velocity $v$ in their time-reversed movement on their right-hand sides. This difference leads to different fluctuation theorems as shown later in this section. Noting this difference, Eq. (\[DetaiBalan3\]) can be interpreted as that the energy loss $\mathcal{R}_{t}^{(v)}(y_{t},y_{0})$ by friction is required to move the particle from $y_{0}$ to $y_{t}$ via the path $\{y_{s}\}_{s\in [t_{0},t]}$ and to return it back from $y_{t}$ to $y_{0}$ via its reversed path *without* changing the dragging velocity $v$. Using Eqs. (\[TransProba1\]) and (\[DetaiBalan3\]) we obtain $$\begin{aligned} e^{-\beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0})} { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } f_{eq}(y_{0}) = { {F}\!\left(\begin{array}{c} \!\! y_{0} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{t} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } f_{eq}(y_{t}) , \nonumber \\ \label{DetaiBalan4}\end{aligned}$$ where we used ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) } = \int_{y_{0}}^{y_{t}}\mathcal{D}y_{s}\; \exp[\int_{t_{0}}^{t}ds\; $ $ L^{(v)}(-\dot{y}_{s},y_{s})]$, as shown by Eqs. (\[TransProba3\]) and (\[TransProba4\]) [@RCW04]. Equation (\[DetaiBalan4\]) reduces to the equilibrium detailed balance (\[DetaiBalan2\]) in the case of $v=0$ because of $\mathcal{R}_{t}^{(0)}(y_{t},y_{0})=0$. Therefore, Eq. (\[DetaiBalan3\]) is another kind of generalization of the equilibrium detailed balance condition to the nonequilibrium steady state, like Eq. (\[DetaiBalan1\]). We introduce the distribution function $P_{r}(R,t)$ of the dimensionless energy loss $R$ by friction in the time-interval $[t_{0},t]$ as $$\begin{aligned} P_{r}(R,t) = { \left\langle\!\!\!\left\langle \delta\!\left(R - \beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0})\right) \right\rangle\!\!\!\right\rangle_{t} } \label{FrictDistr1}\end{aligned}$$ Like for the work distribution function, we represent the distribution function of energy loss by friction in the form $$\begin{aligned} P_{r}(R,t) = \frac{1}{2\pi}\int_{-\infty}^{+\infty}d\lambda \; e^{i\lambda R} \mathcal{E}_{r}(i\lambda,t) \label{DistrFrict2}\end{aligned}$$ where the Fourier transformation $\mathcal{E}_{r}(i\lambda,t)$ is given by $$\begin{aligned} \mathcal{E}_{r}(\lambda,t) &\equiv& { \left\langle\!\!\!\left\langle e^{-\lambda \beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0})} \right\rangle\!\!\!\right\rangle_{t} } \label{EFunctFrict1} \\ &=& \int dy_{t} \! \int dy_{0} \; { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } e^{-\lambda \beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0}) } \nonumber \\ &&{\hspace{0.8cm}}\times f(y_{0},t_{0}) \label{EFunctFrict2}\end{aligned}$$ with Eqs. (\[TransProba1\]) and (\[PathAvera1\]). Here, the $v$-dependence of the function $\mathcal{E}_{r}(\lambda,t)$ for friction, as well as a similar $\mathcal{E}$-function for heat introduced later, has been suppressed. It follows from Eqs. (\[DetaiBalan4\]), (\[EFunctFrict2\]) and $\mathcal{R}_{t}^{(v)}(y_{0},y_{t}) = - \mathcal{R}_{t}^{(v)}(y_{t},y_{0})$ that $$\begin{aligned} \mathcal{E}_{r}(1-\lambda,t) = \mathcal{E}_{r}(\lambda,t) \label{FluctTheorFrict1}\end{aligned}$$ *if* $f(y_{0},t_{0})=f_{eq}(y_{0})$. Or equivalently, for the distribution function $P_{r}(R,t)$ of the dimensionless energy loss by friction, using Eqs. (\[DistrFrict2\]) and (\[FluctTheorFrict1\]) we obtain $$\begin{aligned} \frac{P_{r}(R,t)}{P_{r}(-R,t)} = \exp(R) \label{FluctTheorFrict2}\end{aligned}$$ for the equilibrium initial condition. Equation (\[FluctTheorFrict2\]) is the transient fluctuation theorem for friction and is satisfied for any time $t$. If one is interested in the derivation of a fluctuation theorem for more general initial states than the equilibrium initial state, we can proceed as follows. Using Eqs. (\[TransProba1\]), (\[PathAvera1\]) and (\[FrictDistr1\]) we obtain $$\begin{aligned} P_{r}(R,t) &=& \int dy_{t}\! \int dy_{0} \; f(y_{0},t_{0}) \; \delta\!\left(R - \beta \mathcal{R}_{t}^{(v)}(y_{t},y_{0}) \right) { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } \nonumber \\ &=& \frac{1}{\sqrt{4\pi\alpha\beta v^{2}\mathcal{T}_{t}}} \int dy_{0}\;f(y_{0},t_{0}) \exp \left\{ - \frac{ \left[R-\alpha\beta v(y_{0} +v\tau) \left(1-b_{t}\right) \right]^{2}} {4\alpha\beta v^{2}\mathcal{T}_{t}} \right\} \label{DistrFrict3}\end{aligned}$$ for *any* initial distribution $f(y_{0},t_{0})$, where we used Eqs. (\[TransProba2\]), (\[FrictLoss\]) and $ \delta (R - \beta \mathcal{R}_{t}^{(v)}(y_{t},y_{0}) ) = \delta (y_{t} - y_{0}+ R/(\alpha \beta v) ) /(\alpha \beta |v|)$. To get more concrete results, in the remaining part of this section we concentrate on the initial distribution $$\begin{aligned} f(y_{0},t_{0})=f_{eq}(y_{0}+v\tau\phi), \label{InitiCondi1}\end{aligned}$$ for a constant parameter $\phi$, giving the equilibrium initial distribution for $\phi=0$ and the non-equilibrium steady state initial distribution for $\phi=1$. Inserting Eq. (\[InitiCondi1\]) into Eq. (\[DistrFrict3\]) the distribution $P_{r}(R,t)$ is given by $$\begin{aligned} &&\hspace{-0.5cm} P_{r}(R,t) \nonumber \\ && = \frac{1}{\sqrt{4\pi\alpha\beta v^{2}\tau \left(1-b_{t}\right)}} \nonumber \\ && {\hspace{0.8cm}}\times \exp \left\{ - \frac{ \left[R - \alpha\beta v^{2}\tau(1-\phi) \left(1-b_{t}\right) \right]^{2}}{4 \alpha\beta v^{2} \tau \left(1-b_{t}\right) } \right\} \;\;\; \label{DistrFrict5}\end{aligned}$$ using Eq. (\[EquilDistr1\]). It follows from Eq. (\[DistrFrict5\]) that $$\begin{aligned} \frac{P_{r}(R,t)}{P_{r}(-R,t)} = \exp[(1-\phi)R], \label{FluctTheorFrict3}\end{aligned}$$ which does not have the form of a fluctuation theorem for $\phi\neq 0$. In other words, the distribution function of energy loss by friction satisfies the transient fluctuation theorem for $\phi=0$, but not the steady state fluctuation theorem using the steady state initial condition (\[InitiCondi1\]) for $\phi=1$. Actually, for the initial condition of a nonequilibrium steady state, i.e. if $\phi=1$, its distribution $P_{r}(R,t)$ is Gaussian from Eq. (\[DistrFrict5\]) with its peak at $R=0$, therefore $P_{r}(-R,t)=P_{r}(R,t)$ then at any time. Extended Fluctuation Theorem for Heat {#FluctuationTheoremHeat} ===================================== As the next topic of this paper, we consider the distribution function of heat, which was defined in Sec. \[HeatEntropyBalance\], but now we calculate it by carrying out a functional integral. and also discuss a new simple derivation of its fluctuation theorem briefly in the long time limit. The distribution function of the dimensionless heat $Q$ corresponding to $\beta\mathcal{Q}^{(v)}_{t}(\{y_{s}\})$ using Eq. (\[Heat1\]) is given by $$\begin{aligned} P_{q}(Q,t) = { \left\langle\!\!\!\left\langle \delta\!\left(Q - \beta\mathcal{Q}^{(v)}_{t}(\{y_{s}\})\right) \right\rangle\!\!\!\right\rangle_{t} } . \label{HeatDistr1}\end{aligned}$$ The heat distribution function $P_{q}(Q,t)$ can be calculated like in the distribution function of work or energy loss by friction, namely by representing it as $$\begin{aligned} P_{q}(Q,t) = \frac{1}{2\pi}\int_{-\infty}^{+\infty}d\lambda \; e^{i\lambda Q} \mathcal{E}_{q}(i\lambda,t) \label{HeatDistr2}\end{aligned}$$ where $\mathcal{E}_{q}(\lambda,t)$ is given by $$\begin{aligned} \mathcal{E}_{q}(\lambda,t) &\equiv& { \left\langle\!\!\!\left\langle e^{-\lambda \beta\mathcal{Q}^{(v)}_{t}(\{y_{s}\}) } \right\rangle\!\!\!\right\rangle_{t} } \label{EFunctHeat1} \\ &=& \int dy_{t}\! \int dy_{0} \; e^{\lambda \beta U(y_{t}) } \mathcal{F}(y_{t},y_{0};\lambda) e^{-\lambda \beta U(y_{0}) } \nonumber \\ &&{\hspace{0.8cm}}\times f(y_{0},t_{0}) \label{EFunctHeat2}\end{aligned}$$ where we used Eqs. (\[EnergyBalan1\]), (\[EnergDiffe1\]), (\[PathAvera1\]) and (\[EFunctWorkTrnas1\]) to derive Eq. (\[EFunctHeat2\]) from Eq. (\[EFunctHeat1\]). It may be meaningful to notice that from Eqs. (\[EFunctWork2\]) and (\[EFunctHeat2\]) the function $\mathcal{E}_{q}(\lambda,t)$ for heat is different from the function $\mathcal{E}_{w}^{(v)}(\lambda,t)$ for work by the factor $\exp\{\lambda \beta [U(y_{t})-U(y_{0})] \}$ only. Inserting Eq. (\[FunctCalF1\]) into Eq. (\[EFunctHeat2\]) one obtains $$\begin{aligned} \mathcal{E}_{q}(\lambda,t) &=& \frac{1}{\sqrt{1 - \lambda \left(1-b_{t}^{2}\right)}} \exp\left[ - \lambda (1-\lambda) \alpha \beta v^{2} \left( t-t_{0} - 2\tau \frac{1-b_{t}}{1+b_{t}} \right) \right] \nonumber \\ &&{\hspace{0.8cm}}\times \int dy_{0} \; f(y_{0},t_{0}) \exp\left[ - \frac{\beta\kappa}{2} \frac{\lambda(1 - \lambda )\left(1-b_{t}^{2}\right) } {1 - \lambda \left(1-b_{t}^{2}\right)} \left(y_{0} - v\tau \frac{1-b_{t}}{1+b_{t}} \right)^{2} \right] \label{EFunctHeat3}\end{aligned}$$ for any initial distribution $f(y_{0},t_{0})$. The calculation of the heat distribution function $P_{q}(Q,t)$ from its Fourier transformation like $\mathcal{E}_{q}(i\lambda,t)$ has already done in Ref. [@ZC04] in detail, for the initial condition of the equilibrium and nonequilibrium steady state, and led to the extended fluctuation theorem for heat [@ZC03a; @ZC04]. We do not repeat their calculations and argument in this paper. Instead, in the remaining of this section we discuss the heat distribution function and its fluctuation theorem by a less rigorous but much simpler argument than in Ref. [@ZC04]. This discussion is restricted to the case of the long time limit, in which time-correlations of some quantities may be neglected. This allows to simplify considerably the derivation of the relevant distribution functions. The heat fluctuation theorem is also discussed in Refs. [@BGG05; @G06; @BJM06]. We start our argument by assuming that the particle energy is canonical-like distributed due to the presence of the fluid surrounding a Brownian particle [@ZC03a], so that the distribution $P_{e}(E)$ of the dimensionless energy $E$, i.e. the (potential) energy times the inverse temperature $\beta$, is given by $$\begin{aligned} P_{e} (E) \approx \theta(E) \exp(-E) , \label{EnergDistri2}\end{aligned}$$ where $\theta (x)$ is the Heaviside function taking the value $1$ for $x>0$ and $0$ for $x \leq 0$, and $\theta(E)$ in Eq. (\[EnergDistri2\]) guarantees that the energy $E$ is positive. \[Note that on the right-hand side of Eq. (\[EnergDistri2\]) the normalization condition $\int dE P_{e}(E)=1$ is still satisfied.\] Now, we consider the distribution function $P_{\Delta e}(\Delta E,t)$ of the dimensionless energy difference $\Delta E (=E_{t} - E_{0})$ at the initial time $t_{0}$ and the final time $t$, which is given by $$\begin{aligned} P_{\Delta e}(\Delta E,t) &\stackrel{t\rightarrow+\infty}{\sim}& \int d E_{0}\int d E_{t}\; P_{e}(E_{0}) P_{e}(E_{t}) \nonumber \\ &&{\hspace{0.8cm}}\times \delta(E_{t} - E_{0} - \Delta E) , \label{EnergDiffeDistr1}\end{aligned}$$ namely $\int d E_{0}\; P_{e}(E_{0}) P_{e}(E_{0}+\Delta E)$, in the long-time limit. Here, we have assumed that the energy $E_{0}$ at the initial time $t_{0}$ and the energy $E_{t}$ at the final time $t$ are uncorrelated in the long time limit $t\rightarrow+\infty$, so that the distribution function of the energies $E_{0}$ and $E_{t}$ is given by a multiplication of $P_{e}(E_{0})$ (the initial energy probability distribution) and $P_{e}(E_{t})$ (the final energy probability distribution). The distribution function $P_{\Delta e}(\Delta E,t)$ is given by the integral of such a distribution function of the energies $E_{0}$ and $E_{t}$ over all possible values of $E_{0}$ and $E_{t}$ under the constraint $\Delta E = E_{t}-E_{0}$, therefore by Eq. (\[EnergDiffeDistr1\]). Inserting Eq. (\[EnergDistri2\]) into Eq. (\[EnergDiffeDistr1\]) we obtain $$\begin{aligned} P_{\Delta e}(\Delta E,t) \stackrel{t\rightarrow+\infty}{\sim} \frac{1}{2}\exp(-|\Delta E|) , \label{EnergDiffeDistr2}\end{aligned}$$ meaning that the distribution function $P_{\Delta e}(\Delta E,t)$ of the energy difference $\Delta E$ decays exponentially. \[Note again that the right-hand side of (\[EnergDiffeDistr2\]) satisfies the normalization condition $\int d\Delta E P_{\Delta e}(\Delta E,t) = 1$.\] The argument leading to Eq. (\[EnergDiffeDistr2\]) is also in Ref. [@ZC03a]. On the other hand, we have already calculated the distribution function $P_{w}(W,t)$ of work $W$ in Sec. \[FunctionalCalculationWorkDistribution\] and from Eq. (\[DistrWorkLongTime1\]) we derive $$\begin{aligned} P_{w}(W,t) \stackrel{t\rightarrow+\infty}{\sim} \frac{1}{\sqrt{4\pi\overline{W}_{t}}} \exp\left[-\frac{\left(W- \overline{W}_{t} \right)^{2}} {4 \overline{W}_{t}} \right] \label{WorkDistr4}\end{aligned}$$ for any initial distribution. Here, $\overline{W}_{t}$ is the average of the (dimensionless) work $W$ for the distribution $P_{w}(W,t)$ and given by $\overline{W}_{t}\stackrel{t\rightarrow+\infty}{\sim} \alpha\beta v^{2} t$ from Eq. (\[DistrWorkLongTime1\]) in the long time limit. By Eq. (\[EnergyBalan1\]), the heat $Q$ is given by $Q=W-\Delta E$ using the work $W$ and energy difference $\Delta E$, and its distribution function $P_{q}(Q,t)$ should be represented as $$\begin{aligned} P_{q}(Q,t) & \stackrel{t\rightarrow+\infty}{\sim}& \int d W \int d \Delta E\; P_{w}(W,t) P_{\Delta e}(\Delta E,t) \nonumber \\ &&{\hspace{0.8cm}}\times \delta (W-\Delta E - Q), \label{HeatDistr3}\end{aligned}$$ namely $\int d W \; P_{w}(W,t) P_{\Delta e}(W-Q,t)$, in the long time limit. Here, we used a similar argument as in Eq. (\[EnergDiffeDistr1\]) in order to justify Eq. (\[HeatDistr3\]), namely, Eq. (\[HeatDistr3\]) is the integral of the multiplication of the work distribution $P_{w}(W,t)$ and the energy-difference distribution function $P_{\Delta e}(\Delta E,t)$ over all possible values of $W$ and $\Delta E$ under the restriction $Q=W-\Delta E$ for a given $\Delta E$. Non-correlation of the work and the energy difference in the long time limit, which is assumed in Eq. (\[HeatDistr3\]), may be justified by the fact that the work depends on the particle position over the entire time interval $[t_{0},t]$ by Eqs. (\[WorkRate1\]) and (\[Work1\]) (in which the effects at the times $t_{0}$ and $t$ are negligible in the long time limit) while the energy difference depends exclusively on the particle positions at the times $t_{0}$ and $t$ only. Inserting Eq. (\[EnergDiffeDistr2\]) and (\[WorkDistr4\]) into Eq. (\[HeatDistr3\]) we obtain $$\begin{aligned} P_{q}(Q,t) &\stackrel{t\rightarrow+\infty}{\sim}& \frac{1}{4} \left[ e^{-Q+2\overline{W}_{t}} \mbox{erfc}\left(-\frac{Q- 3 \overline{W}_{t}} {2\sqrt{\overline{W}_{t}}}\right) \right.\nonumber \\ &&{\hspace{0.8cm}}\left. + e^{Q}\mbox{erfc}\left(\frac{Q + \overline{W}_{t}} {2\sqrt{\overline{W}_{t}}} \right) \right] \label{HeatDistr4}\end{aligned}$$ with the complimentary error function $\mbox{erfc}(x)$ defined by $ \mbox{erfc}(x) \equiv (2/\sqrt{\pi}) \int_{x}^{+\infty} dz\; \exp(-z^{2}) $, satisfying the inequality $0<\mbox{erfc}(x)<2$. One may notice that the average work $\overline{W}_{t}$ is equal to the average heat in the case of the nonequilibrium steady state initial condition $f(y_{0},t_{0}) =f_{ss}(y_{0})$, because of the energy conservation law (\[EnergyBalan1\]) and the fact that average of the internal energy difference (\[EnergDiffe1\]) is zero in this case. Equation (\[HeatDistr4\]) gives the asymptotic form of the heat distribution function. The exponential factors $\exp(\pm Q)$ in Eq. (\[HeatDistr4\]) dominate the tails of the heat distribution function $P_{q}(Q,t)$ [@ZC04]. ![ Comparison of the work fluctuation theorem and the heat fluctuation theorem by plotting the function $G_{w}(X,t)$ for the work distribution and the function $G_{q}(X,t)$ for the heat distribution as functions of a scaled $X/\overline{W}_{t}$ in the long time limit $t\rightarrow +\infty$. Here, we used the asymptotic forms (\[WorkDistr4\]) and (\[HeatDistr4\]) of the work distribution function and the heat distribution function, respectively, in the case of $\overline{W}_{t}=70$. In the small $X$ region ($0\leq X/\overline{W}_{t} \leq 1$), the values of the two functions $G_{w}(X,t)$ and $G_{q}(X,t)$ appear to be consistent with $G_{w}(X,t) =G_{q}(X,t)$, while $G_{w}(X,t)$ is $X/ \overline{W}_{t}$ and $G_{q}(X,t)$ is 2 in the large $X$ region ($3\leq X/\overline{W}_{t} \leq +\infty$) in the long time limit. []{data-label="figB1heatFT"}](figB1heatFT.eps){width="\widthfig"} Now, we introduce the fluctuation functions $G_{w}(W,t)$ and $G_{q}(Q,t)$ defined by $$\begin{aligned} G_{w}(W,t) &\equiv& \frac{1}{ \overline{W}_{t}} \ln\frac{P_{w}(W,t)} {P_{w}(-W,t)} , \label{WorkFluctFunct1}\\ G_{q}(Q,t) &\equiv& \frac{1}{ \overline{W}_{t}} \ln\frac{P_{q}(Q,t)} {P_{q}(-Q,t)} \label{HeatFluctFunct1}\end{aligned}$$ for the work distribution function $P_{w}(W,t)$ and the heat distribution function $P_{q}(Q,t)$. By Eq. (\[WorkDistr4\]) the function $G_{w}(W,t)$ is given simply by $G_{w}(W,t)\stackrel{t\rightarrow+\infty}{\sim}W/ \overline{W}_{t}$ in the long time limit, characterizing the work fluctuation theorem in a proper way to compare it with the heat fluctuation theorem characterized by the function $G_{q}(Q,t)$. In Fig. \[figB1heatFT\] the functions $G_{w}(W,t)$ (broken line) and $G_{q}(Q,t)$ (solid line) are plotted as functions of $X/\overline{W}_{t}$ ($X=W$ or $Q$) using Eqs. (\[WorkDistr4\]) and (\[HeatDistr4\]) in the case of $\overline{W}_{t}=70$. In this figure we plotted only in the positive region of $W$ and $Q$, because their values in the negative region is simply given by $G_{w}(-W,t) = -G_{w}(W,t)$ and $G_{q}(-Q,t) = - G_{q}(Q,t)$. It is clear from Fig. \[figB1heatFT\] that the values of the functions $G_{w}(X,t)$ and $G_{q}(X,t)$ will coincide with each other for small $W$ and $Q$ for $\overline{W}_{t}\rightarrow +\infty$, i.e. $t\rightarrow +\infty$, meaning that the heat fluctuation theorem coincides with the work fluctuation theorem in this region. The difference between the heat and work fluctuation theorems appears in the large values of the argument, where the function $G_{w}(W,t)$ remains $W/\overline{W}_{t}$, while the function $G_{q}(Q,t)$ takes the constant value $2$ for $Q/\overline{W}_{t} > 3$ in the long time limit. For further details, we refer to Ref. [@ZC04]. Inertial Effects {#InertiaEffects} ================ So far, we have concentrated our discussions to the over-damped case and have neglected inertial effects. A generalization of our discussions to the ones including the inertia is almost straightforward. One of the features caused by introducing inertia is a kinetic term in the equilibrium and nonequilibrium steady state distribution functions. This kinetic term depends on the frame one uses, namely the comoving frame or the laboratory frame, respectively. The inertial force, like d’Alembert’s force, also appears as an inertial effect. In this section we discuss briefly these effects beyond the over-damped case. The Langevin equation including inertia is expressed as Eq. (\[LangeEq3\]) in the laboratory frame. Like in the over-damped case, we can convert Eq. (\[LangeEq3\]) for the laboratory frame to $$\begin{aligned} m\frac{d^{2} y_{t}}{dt^{2}} = -\alpha \frac{d y_{t}}{dt} - \kappa \left(y_{t} +v\tau\right)+ \zeta_{t} \label{LangeEq4}\end{aligned}$$ for the comoving frame by Eq. (\[ComovPosit1\]). Equation (\[LangeEq4\]) reduces to Eq. (\[LangeEq2\]) for the over-damped case when $m d^{2} y_{t}/dt^{2}=0$. We introduce a canonical-like distribution function as $$\begin{aligned} f_{eq}^{(\vartheta)}(\dot{y},y) \equiv \Xi^{(\vartheta)}{}^{-1}\exp\left[-\beta \mathcal{H}\left(\dot{y}+\vartheta v,y\right)\right] \label{EquilDistr2}\end{aligned}$$ where $\dot{y}$ is the time-derivative of $y$ and $\mathcal{H}\left(\dot{y},y\right)$ is defined by $$\begin{aligned} \mathcal{H}\left(\dot{y},y\right) \equiv m\dot{y}^{2}/2 + \kappa y^{2}/2, \label{HamilFunct1}\end{aligned}$$ and $\Xi^{(\vartheta)}$ is the normalization constant for the distribution function $f_{eq}^{(\vartheta)}(\dot{y},y)$. It is important to note that the particle velocity depends on the frame, and is given by $\dot{y}$ for the comoving frame and by $\dot{x}(=\dot{y}+v)$ for the laboratory frame. For that reason, the canonical distribution function $f_{eq}^{(\vartheta)}(\dot{y},y)$ including the kinetic energy depends on the frame, so that $f_{eq}^{(0)}(\dot{y},y)$ for $\vartheta=0$ refers to the comoving frame and $f_{eq}^{(1)}(\dot{y},y)$ for $\vartheta=1$ refers to the laboratory frame. By a way similar to the over-damped case, the functional probability density for path $\{y_{s}\}_{s\in[t_{0},t]}$ is given by $\exp[\int_{t_{0}}^{t} ds\; L^{(v)}(\ddot{y}_{s},\dot{y}_{s},y_{s})]$ with the Lagrangian function $$\begin{aligned} L^{(v)}\!\left(\ddot{y}_{s},\dot{y}_{s},y_{s}\right) \equiv -\frac{1}{4D}\left( \dot{y}_{s} +\frac{1}{\tau} y_{s} +v +\frac{m}{\alpha}\ddot{y}_{s} \right)^{2} \label{Lagra3}\end{aligned}$$ using $\ddot{y}_{s}\equiv d^{2}y_{s}/ds^{2}$. The Lagrangian function (\[Lagra3\]) becomes the Lagrangian function (\[Lagra1\]) in the over-damped case, where $m\ddot{y}_{s}=0$. Using Eqs. (\[EquilDistr2\]) and (\[Lagra3\]) we obtain $$\begin{aligned} && e^{-\beta \int_{t_{0}}^{t}ds\; \Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta) v} e^{\int_{t_{0}}^{t} ds\; L^{(v)}\!\left(\ddot{y}_{s},\dot{y}_{s},y_{s}\right)} f_{eq}^{(\vartheta)}\!\left(\dot{y}_{0},y_{0}\right) \nonumber \\ &&{\hspace{0.8cm}}= f_{eq}^{(\vartheta)}\!\left(\dot{y}_{t},y_{t}\right) e^{\int_{t_{0}}^{t} ds\; L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right)} \label{DetaiBalan7}\end{aligned}$$ where $\dot{y}_{0}=\dot{y}_{t_{0}}$, and $\Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta)$ is a modified “force” defined by $$\begin{aligned} \Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta) &\equiv& - \kappa y_{s} \frac{1\mp 1}{2} - \alpha \dot{y}_{s}\frac{1\pm 1}{2} \nonumber \\ && {\hspace{0.8cm}}- m\ddot{y}_{s} \left(\frac{1\mp 1}{2}-\vartheta\right). \label{GenerForce1}\end{aligned}$$ Equation (\[DetaiBalan7\]) may be regarded as a nonequilibrium detailed balance relation for the case of a potential force, friction and inertia. \[See Appendix \[NonequilibriumDetailedBalanceInertia\] for a derivation of Eq. (\[DetaiBalan7\]).\] Moreover, the signs $\pm$ in Eq. (\[DetaiBalan7\]) correspond to the case of work ($-1$), discussed in \[FluctuationTheoremWork\] and that of energy loss by friction ($+1$), respectively, discussed in Sec. \[FluctuationTheoremFriction\], and are due to the $\pm v$ signs of the Lagrangian $L^{(\pm v)} (\ddot{y}_{s},-\dot{y}_{s},y_{s} )$ on the right-hand side of Eq. (\[DetaiBalan7\]). It should be noted that the first, second and third terms on the right-hand side of Eq. (\[GenerForce1\]) are regarded as the harmonic force, the friction force, and the inertial (d’Alembert-like) force, respectively. ----------------------------------------------------------------------------------------------------------------------------------------------------------------- Frame ($\vartheta$) $\hspace{0.5cm}\pm v\hspace{0.5cm}$ Force $\Lambda$ Fluctuation Theorem --------------------- ------------------------------------- ------------------------------------------ ---------------------------------------------------------- Comoving ($0$) $-v$ $\Lambda_{-}(\ddot{y},\dot{y},y;0) $\textstyle \frac{\mathcal{P}_{-}^{(0)}(\mathcal{W},t)} =-\kappa y - m\ddot{y}$ {\mathcal{P}_{-}^{(0)}(-\mathcal{W},t)} = \exp(\mathcal{W})$ Comoving ($0$) $+v$ $\Lambda_{+}(\ddot{y},\dot{y},y;0) $\textstyle \frac{\mathcal{P}_{+}^{(0)}(\mathcal{W},t)} = - \alpha \dot{y} $ {\mathcal{P}_{+}^{(0)}(-\mathcal{W},t)} = \exp(\mathcal{W})$ Laboratory ($1$) $-v$ $\Lambda_{-}(\ddot{y},\dot{y},y;1) $\textstyle \frac{\mathcal{P}_{-}^{(1)}(\mathcal{W},t)} = - \kappa y$ {\mathcal{P}_{-}^{(1)}(-\mathcal{W},t)} = \exp(\mathcal{W})$ Laboratory ($1$) $+v$ $\Lambda_{+}(\ddot{y},\dot{y},y;1) $\textstyle \frac{\mathcal{P}_{+}^{(1)}(\mathcal{W},t)} = - \alpha \dot{y} + m\ddot{y}$ {\tilde{\mathcal{P}}_{+}^{(1)}(-\mathcal{W},t)} = \exp(\mathcal{W})$ ----------------------------------------------------------------------------------------------------------------------------------------------------------------- Now, we introduce the dimensionless modified “work” rate $\beta\Lambda_{\pm}(\ddot{y},\dot{y},y;\vartheta)v$ and its distribution function $\mathcal{P}_{\pm}^{(\vartheta)}(\mathcal{W},t)$ as $$\begin{aligned} \mathcal{P}_{\pm}^{(\vartheta)}(\mathcal{W},t) = { \left\langle\!\!\!\left\langle \delta \left(\mathcal{W}-\beta\int_{t_{0}}^{t} ds\; \Lambda_{\pm}(\ddot{y},\dot{y},y;\vartheta)v\right) \right\rangle\!\!\!\right\rangle_{t} } \label{WorkDistr5}\end{aligned}$$ where ${ \left\langle\!\!\!\left\langle \cdots \right\rangle\!\!\!\right\rangle_{t} }$ is the functional average in the inertia case, like the one given by Eq. (\[PathAvera1\]). Here, we remark that $\mathcal{W}$ in Eq. (\[WorkDistr5\]) differs from the work $\mathcal{W}_{t}^{(v)}(\{y_{s}\})$ defined by Eq. (\[Work1\]). In a way similar to derive Eqs. (\[FluctTheorWork2\]) and (\[FluctTheorFrict2\]) in the over-damped case, it follows that the distribution function $\mathcal{P}_{\pm}^{(\vartheta)}(\mathcal{W},t)$ satisfies the transient fluctuation theorem $$\begin{aligned} \frac{\mathcal{P}_{\pm}^{(\vartheta)}(\mathcal{W},t)} {\tilde{\mathcal{P}}_{\pm}^{(\vartheta)}(-\mathcal{W},t)} = \exp(\mathcal{W}) \label{FluctTheorModifWork1}\end{aligned}$$ under the condition that the initial distribution at time $t_{0}$ is given by the $f_{eq}^{(\vartheta)}(\dot{y},y)$. Here, the distribution $\tilde{\mathcal{P}}_{\pm}^{(\vartheta)}(\mathcal{W},t)$ is defined by $$\begin{aligned} &&\hspace{-0.5cm} \tilde{\mathcal{P}}_{\pm}^{(\vartheta)}(\mathcal{W},t) \nonumber \\ &&= { \left\langle\!\!\!\left\langle \delta \left(\mathcal{W}-\beta\int_{t_{0}}^{t} ds\; \Lambda_{\pm}(\ddot{y},\dot{y},y; (-1)^{\frac{1\pm 1}{2}}\vartheta)v\right) \right\rangle\!\!\!\right\rangle_{t} } , \nonumber \\ \label{WorkDistr6}\end{aligned}$$ and is simply given by $$\begin{aligned} \tilde{\mathcal{P}}_{+}^{(0)}(\mathcal{W},t) &=&\mathcal{P}_{+}^{(0)}(\mathcal{W},t), \\ \tilde{\mathcal{P}}_{-}^{(\vartheta)}(\mathcal{W},t) &=&\mathcal{P}_{-}^{(\vartheta)}(\mathcal{W},t) \label{WorkDistr7}\end{aligned}$$ in these special cases, and in order to derive Eq. (\[FluctTheorModifWork1\]) we also used the relations $L^{(v)}\!\left(\ddot{y}_{s},\dot{y}_{s},y_{s}\right) =L^{(-v)}\!\left(-\ddot{y}_{s},-\dot{y}_{s},-y_{s}\right)$, $\Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta) =-\Lambda_{\pm}(-\ddot{y}_{s},-\dot{y}_{s},-y_{s};\vartheta)$ and $$\begin{aligned} \Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta) v = - \Lambda_{\pm}(\ddot{y}_{s},-\dot{y}_{s},y_{s}; (-1)^{\frac{1\pm 1}{2}}\vartheta) (\pm v) . \nonumber \\ $$ It may be noted that the two terms $- \alpha \dot{y}$ and $m\ddot{y}$ for the force $ \Lambda_{+}(\ddot{y}_{s},\dot{y}_{s},y_{s};1)$ have different time-reversal properties than the other forces $ \Lambda_{+}(\ddot{y}_{s},\dot{y}_{s},y_{s};0)$ and $ \Lambda_{+}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta)$. From Eq. (\[FluctTheorModifWork1\]) we derived four different fluctuation theorems corresponding to the cases $(\vartheta,\pm v) = (0,-v)$, $(1,-v)$, $(0,v)$, and $(1,v)$, where $\pm v$ is the velocity appearing in the Lagrangian function $L^{(\pm v)}(\ddot{y}_{s},-\dot{y}_{s},y_{s})$ on the right-hand side of the nonequilibrium detailed balance relation (\[DetaiBalan7\]). We summarize these four fluctuation theorems in Table \[4FluctTheor\]. In the last line of this table, appearance of the function $\tilde{\mathcal{P}}_{+}^{(1)}(-\mathcal{W},t)$ for the case of $(\vartheta,\pm v) = (1,v)$ is due to the different behavior with respect to time-reversal of the two terms $- \alpha \dot{y}$ and $m\ddot{y}$ composing the modified force $\Lambda_{+}(\ddot{y}_{s},\dot{y}_{s},y_{s};1)$, while in all the other cases in Table \[4FluctTheor\] the modified forces have the unique behavior under time-reversal. Conclusions and Remarks {#ConclusionsRemarks} ======================= In this paper we discussed a generalization of Onsager-Machlup’s fluctuation theory to nonequilibrium steady states and fluctuation theorems based on nonequilibrium detailed balance relations. To that end, we used a model which consists of a Brownian particle confined by a harmonic potential which is dragged with a constant velocity through a heat reservoir. This model is described by a Langevin equation, which is a simple and exactly-solvable nonequilibrium steady state model. Our basic analytical approach is a functional integral technique, which was used in Onsager and Machlup’s original work and is effective to discuss fluctuation theorems treating quantities expressed as functionals, for example, work and heat. First, we gave an expression of the transition probability in terms of a Lagrangian function which can be written as a sum of an entropy production rate and two dissipation functions. There is a difference, though with the similar result of Onsager and Machlup’s original papers [@OM53; @MO53], in that now the entropy production rate and one of the two dissipation functions - and consequently also the Lagrangian function - depend on the dragging velocity $v$ leading to nonequilibrium steady state effects. From this property of the Lagrangian function, we constructed a nonequilibrium steady state thermodynamics by obtaining the second law of thermodynamics and the energy conservation law which involves fluctuating heat, work and an internal potential energy difference. We also discussed Onsager’s principle of minimum energy dissipation and the most probable path approximating the transition probability of the particle position. This approach is different from another attempt for an Onsager-Machlup theory for nonequilibrium steady states [@BSG01; @BSG02], where a nonlinear diffusion equation is applied to models like an exclusion model and a boundary driven zero range model. Instead, we use a stochastic model described by a Langevin equation, so that our results automatically include those of Onsager and Machlup’s original works by taking a specific equilibrium value, $v=0$, for the nonequilibrium parameter $v$, and relax Onsager and Machlup’s variable $\alpha$ and $\dot{\alpha}$ in Refs. [@OM53; @MO53] to our variables $x$ and $\dot{x}$, respectively. Second, we derived nonequilibrium detailed balance relations from the Lagrangian function to obtain not only the well-known fluctuation theorem for work but also another fluctuation theorem for energy loss by friction. We also indicated the derivation of the extended fluctuation theorem for heat by carrying out explicitly the relevant functional integral and then using Refs. [@ZC03a; @ZC04]. In addition, we gave a simple argument for the heat fluctuation theorem in the long time limit. Finally, we discussed briefly the effects of inertia, and obtained four different fluctuation theorems related to a potential force, a friction force and d’Alembert-like (or inertial) force for the comoving frame or the laboratory frame. In the remaining of this section, we give remarks for the contents in the main text of this paper. 1\) In this paper, we have emphasized a close connection between nonequilibrium detailed balance relations and fluctuation theorems, using a functional integral approach. It may be noted that in some of past works concepts of detailed balances have been mentioned for formal derivations of fluctuation theorems in various different contexts, implicitly or explicitly [@ECM93; @K98; @LS99; @C99; @C00; @CCJ06]. However, we should keep in mind that a generalization of the equilibrium detailed balance to nonequilibrium states is not unique, as shown in this paper \[cf. Eq. (\[DetaiBalan1\]) and (\[DetaiBalan3\])\]. As a remark related to this point, we should notice that even if the equilibrium detailed balance condition is violated, but another detailed balance condition for the nonequilibrium steady state still holds, namely, using the nonequilibrium steady state distribution $f_{ss}(y)$ we obtain for the over-damped case: $$\begin{aligned} e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)} f_{ss}(y_{0}) = f_{ss}(y_{t}) \;e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(-\dot{y}_{s},y_{s}\right)} , \;\;\; \label{DetaiBalan5}\end{aligned}$$ by Eqs. (\[Lagra1\]), (\[SteadSolut1\]) and (\[EquilDistr1\]), or equivalently ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) } f_{ss}(y_{0}) = { {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) } f_{ss}(y_{t})$. Here, it is essential to note that on the right-hand side of Eq. (\[DetaiBalan5\]) we do not change the sign of the dragging velocity $v$ although we change the sign of the particle velocity $\dot{y}_{s}$ in the comoving frame. We emphasize that here, there is no additional multiplying factors like $\exp[-\beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})]$ as in Eq. (\[DetaiBalan1\]) or $\exp[-\beta\mathcal{R}_{t}^{(v)}(y_{t},y_{0})]$ as in Eq. (\[DetaiBalan3\]). As a consequence we have been unable to derive fluctuation theorems from Eq. (\[DetaiBalan5\]). Since we chose the equilibrium state as the reference state for the detailed balance in this paper, our interest was mainly the work to maintain the system in a nonequilibrium steady state, i.e. the work necessary to keep the system from going to the equilibrium state. In this sense, we note that the reference state is arbitrary, for example, if we are interested in the work to go from one nonequilibrium state to another nonequilibrium state. In general, the modification of the detailed balance relation based on an arbitrary reference distribution function $f_{ref}(y)$ can be expressed as $$\begin{aligned} && e^{-\beta Y_{t}(\{y_{s}\})} e^{\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)} f_{ref}(y_{0}) \nonumber \\ && {\hspace{0.8cm}}= f_{ref}(y_{t}) \;e^{\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(-\dot{y}_{s},y_{s}\right)} \label{DetaiBalan6}\end{aligned}$$ where $Y_{t}(\{y_{s}\})$ is the functional defined by $$\begin{aligned} Y_{t}(\{y_{s}\}) \equiv \mathcal{Q}^{(v)}_{t}(\{y_{s}\}) +\beta^{-1}\ln\frac{f_{ref}(y_{0})}{f_{ref}(y_{t})} . $$ Choosing $f_{ref}(y)=f_{eq}(y)$, Eq. (\[DetaiBalan6\]) leads to Eq. (\[DetaiBalan1\]). We can also get a generalization of Eq. (\[DetaiBalan3\]) for an arbitrary reference distribution function $f_{ref}(y)$. An analogous quantity to $Y_{t}(\{y_{s}\})$ is in Ref. [@ES02] for a thermostated system with deterministic dynamics. 2\) From Eq. (\[DetaiBalan5\]) we derive $$\begin{aligned} && \int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right) + \tilde{S}_{ss}(y_{0})/k_{B} \nonumber \\ &&{\hspace{0.8cm}}= \int_{t_{0}}^{t}ds\; L^{(v)}\! \left(-\dot{y}_{s},y_{s}\right) + \tilde{S}_{ss}(y_{t})/k_{B} \label{OMSymme1}\end{aligned}$$ where $\tilde{S}_{ss}(y)$ is defined by $ \tilde{S}_{ss}(y) \equiv k_{B}\ln f_{ss}(y) $. An identity like Eq. (\[OMSymme1\]) is called an Onsager-Machlup symmetry [@BSG01; @BSG02; @G02], and is used to discuss macroscopic properties of nonequilibrium steady states. Using Eq. (\[OMSymme1\]) we can also obtain an expression like $$\begin{aligned} \frac{\exp\left[\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)\right]} {\exp\left[\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(-\dot{y}_{s},y_{s}\right)\right]} = \exp\left[\beta \tilde{Q}_{ss}(t,t_{0})\right] \label{OMSymme2}\end{aligned}$$ with $\tilde{Q}_{ss}(t,t_{0}) \equiv T [ \tilde{S}_{ss}(y_{t}) - \tilde{S}_{ss}(y_{0}) ]$. On the other hand, it can be shown from Eqs. (\[EquilDistr1\]), (\[EnergyBalan1\]) and (\[DetaiBalan1\]) \[or from Eq. (\[DetaiBalan6\])\] that $$\begin{aligned} \frac{\exp\left[\int_{t_{0}}^{t}ds\; L^{(v)}\! \left(\dot{y}_{s},y_{s}\right)\right]} {\exp\left[\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(-\dot{y}_{s},y_{s}\right)\right]} = \exp\left[\beta \mathcal{Q}^{(v)}_{t}(\{y_{s}\})\right] \label{OMSymme3}\end{aligned}$$ using the heat $\mathcal{Q}^{(v)}_{t}(\{y_{s}\})$ of Eq. (\[Heat1\]). Note that Eq. (\[OMSymme3\]) is consistent with the heat $\mathcal{Q}^{(v)}_{t}(\{y_{s}\})$ appearing in our energy conservation law (\[EnergyBalan1\]), in contrast to Eq. (\[OMSymme2\]) in which the quantity $\tilde{Q}_{ss}(t,t_{0})$ does not have such a correspondence with the heat. Thus we will restrict ourselves in the following to Eq. (\[OMSymme3\]). As we discussed in Sec. \[DraggedBrownianParticle\], the term $\exp[\int_{t_{0}}^{t}ds\; L^{(v)}\! (\dot{y}_{s},y_{s})]$ appearing in the numerator on the left-hand side of Eq. (\[OMSymme3\]) is the probability functional of the forward path $\{y_{s}\}_{s\in[t_{0},t]}$. On the other hand, the denominator on the left-hand side of Eq. (\[OMSymme3\]) is the probability functional of the corresponding time-reversed path with the dragging velocity $-v$. Therefore, Eq. (\[OMSymme3\]) implies that the logarithm of the ratio of such forward and backward probability functionals is given by the heat multiplied by the inverse temperature. In this sense it is tempting to claim Eq. (\[OMSymme3\]) as a relation leading to a fluctuation theorem [@ND04; @K05; @IP06]. However, it is important to distinguish Eq. (\[OMSymme3\]) from the fluctuation theorems discussed in the main text of this paper. First of all, although it is related to the heat, Eq. (\[OMSymme3\]) has a form rather close to a relation leading to the conventional fluctuation theorems like Eqs. (\[FluctTheorWork2\]) and (\[FluctTheorFrict2\]), which are different from the extended form for the heat fluctuation theorem discussed in Sec. \[FluctuationTheoremHeat\]. Similarly, although one may regard Eq. (\[OMSymme3\]) as a nonequilibrium detailed balance relation for heat, a derivation of the extended fluctuation theorem for heat from a nonequilibrium detailed balance relation remains an open problem. We should also notice that no initial condition dependence appears in Eq. (\[OMSymme3\]), so that we cannot discuss directly, for example, a difference between the transient fluctuation theorem and the steady state fluctuation theorem for them from Eq. (\[OMSymme3\]). 3\) Although the nonequilibrium detailed balance relations, like Eq. (\[DetaiBalan1\]), (\[DetaiBalan3\]) or (\[DetaiBalan6\]), play an essential role to derive the fluctuation theorem, it is important to note that some properties of the fluctuation theorem cannot be discussed by it. Basically, the nonequilibrium detailed balance relation can lead directly to the so-called transient fluctuation theorems, which are identically satisfied for any time [@CG99], but this relation does not say what happens to fluctuation theorems if we change the initial condition (like the equilibrium distribution) to another (like the nonequilibrium steady state discussed in the steady state fluctuation theorem). The transient fluctuation theorem can be different from the steady state fluctuation theorem for some quantities, even in the long time limit. As an example for such a difference, we showed in this paper that the energy loss by friction satisfies the transient fluctuation theorem but does not satisfy the steady state fluctuation theorem. We have discussed an initial condition dependence of fluctuation theorems by carrying out functional integrals to obtain distribution functions explicitly, and showed that the work distribution function has an asymptotic form satisfying the work fluctuation theorem, independent of the initial distribution, while the friction-loss distribution function depends on the initial condition even in the long time limit. This difference between the work and the friction-loss might come from the fact that the work is given by a time-integral of the particle position so that its contribution near the initial time can be neglected in the long time limit, while the energy loss by friction is given by the particle position at the initial and final times only. A systematic way to investigate whether a fluctuation theorem is satisfied for any initial condition without calculating a distribution function, remains an open problem [@noteHeat]. 4\) Finally we note that the analogy of the Brownian particle case, discussed here, and the electric circuit case should persist not only in the over-damped case (as shown in [@ZCC04]) but also in the case including inertia. In that case, one has to add the self-induction $L_{0}$ of the electric circuit, as the corresponding quantity of the mass $m$ of the Brownian particle. This will add the correspondence of $m$ and $L_{0}$ to Table I in Ref. [@ZCC04]. Then, the fluctuation theorems in Table \[4FluctTheor\] in Sec. \[InertiaEffects\] of this paper can, by using the extended analogy described above, also be used for electric circuits, and might be experimentally accessible (cf. [@GC05]). Acknowledgements {#acknowledgements .unnumbered} ================ One of the authors (E.G.D.C.) is indebted to Prof. G. Jona-Lasinio for a stimulating discussion at the Isaac Newton Institute for Mathematical Sciences in Cambridge, U.K, while the other author (T.T.) wishes to express his gratitude to Prof. K. Kitahara for a comment about a relation between the Onsager-Machlup theory and a fluctuation theorem near equilibrium. Both authors gratefully acknowledge financial support of the National Science Foundation, under award PHY-0501315. Transition Probability using a Functional Integral Technique {#TransitionProbabilityFunctionalIntegral} ============================================================ In this Appendix, we outline a derivation of the transition probability (\[TransProba1\]) for the stochastic process described by the Langevin equation (\[LangeEq2\]). First, we translate the Langevin equation (\[LangeEq2\]) into the corresponding Fokker-Planck equation. It can be done using the Kramers-Moyal expansion technique [@K92; @R89], and we obtain the Fokker-Planck equation $$\begin{aligned} \frac{\partial f(y,t)}{\partial t} = \hat{\mathcal{L}} f(y,t) \label{FPequat1}\end{aligned}$$ for the distribution function $f(y,t)$ of the particle position $y$ at time $t$. Here, $\hat{\mathcal{L}}$ is the Fokker-Planck operator defined by $$\begin{aligned} \hat{\mathcal{L}} \equiv \frac{\partial}{\partial y} \left( \frac{y + v\tau}{\tau} + D \frac{\partial}{\partial y} \right) \label{FPopera1}\end{aligned}$$ with $D\equiv 1/(\alpha\beta)$. The transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y} \!\! \\ \!\! \scriptstyle{t+\Delta t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y'} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array}) }$ from $y'$ at time $t$ to $y$ at time $t+\Delta t$ is given by $$\begin{aligned} { {F}\!\left(\begin{array}{c} \!\! y \!\! \\ \!\! t +\Delta t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y' \!\! \\ \!\! t \!\! \end{array}\right) } &=& e^{\hat{\mathcal{L}} \Delta t}\delta (y-y') \nonumber \\ &=& \left[ 1 + \hat{\mathcal{L}} \Delta t + \mathcal{O}\left(\Delta t^{2}\right) \right] \delta \left(y-y'\right) \nonumber \\ &=& \left( 1 + \Delta t \frac{\partial}{\partial y} \frac{y'+v\tau}{\tau} +D\Delta t \frac{\partial^{2}}{\partial y^{2}} \right) \delta \left(y-y'\right) + \mathcal{O}\left(\Delta t^{2}\right) \nonumber \\ &=& \frac{1}{2\pi} \int_{-\infty}^{+\infty} d\lambda \; \left( 1 + \Delta t \frac{y'+v\tau}{\tau} \frac{\partial}{\partial y} +D\Delta t \frac{\partial^{2}}{\partial y^{2}} \right) \exp\left[ i \lambda (y-y') \right] + \mathcal{O}\left(\Delta t^{2}\right) \nonumber \\ &=& \frac{1}{2\pi}\int_{-\infty}^{+\infty} d\lambda \; \exp\left[-D\Delta t \lambda^{2} + i \Delta t \left( \frac{y-y'}{\Delta t } + \frac{y'+v\tau}{\tau} \right) \lambda \right] + \mathcal{O}\left(\Delta t^{2}\right) \nonumber \\ &=& \frac{1}{\sqrt{4\pi D\Delta t}} \exp\left[- \frac{1}{4D} \left(\frac{y-y'}{\Delta t} + \frac{y'+v\tau}{\tau} \right)^{2} \Delta t \right] + \mathcal{O}(\Delta t^{2}) \label{TransProba3}\end{aligned}$$ where we used $y \delta \left(y-y'\right) = y' \delta \left(y-y'\right)$. On the other hand, using the Chapman-Kolmogorov equation [@K92], the transition probability for a finite time interval $t-t_{0}$ is expressed as $$\begin{aligned} { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } &=& \lim_{N\rightarrow+\infty} \int dy_{N-1} \int dy_{N-2} \cdots \int dy_{1} \; { {F}\!\left(\begin{array}{c} \!\! y_{t} \!\! \\ \!\! t \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{N-1} \!\! \\ \!\! t_{N-1} \!\! \end{array}\right) } { {F}\!\left(\begin{array}{c} \!\! y_{N-1} \!\! \\ \!\! t_{N-1} \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{N-2} \!\! \\ \!\! t_{N-2} \!\! \end{array}\right) } \cdots { {F}\!\left(\begin{array}{c} \!\! y_{1} \!\! \\ \!\! t_{1} \!\! \end{array}\right| \left. \begin{array}{c} \!\! y_{0} \!\! \\ \!\! t_{0} \!\! \end{array}\right) } \label{TransProba4}\end{aligned}$$ with $t_{n}\equiv t_{0} + n \Delta t_{N}$, $n=1,2,\cdots,N$, $\Delta t_{N}\equiv (t-t_{0})/N$, $t_{N}=t$. Inserting the expression (\[TransProba3\]) for the transition probability in a short time interval $\Delta t = \Delta t_{N}$ into Eq. (\[TransProba4\]) we obtain Eq. (\[TransProba1\]) with the functional integral (\[FunctInteg1\]). Fluctuation Theorem for Work {#TransientFluctuationTheoremWork} ============================ In this Appendix, we show the relation $\mathcal{E}_{w}^{(v)}(\lambda,t) = \mathcal{E}_{w}^{(-v)}(1-\lambda,t)$, therefore Eq. (\[FluctTheorWork2\]). We also give a derivation of Eq. (\[FluctTheorWork2\]) from Eq. (\[FluctTheorWork1\]). From Eq. (\[EFunctWork1\]) with the functional average (\[PathAvera1\]) we derive $$\begin{aligned} \mathcal{E}_{w}^{(v)}(\lambda,t) &=& \int dy_{t} \! \int_{y_{0}}^{y_{t}}\mathcal{D}y_{s} \!\int dy_{0} \; e^{\int_{t_{0}}^{t}ds\; L^{(v)}\!\left(\dot{y}_{s},y_{s}\right)} f(y_{0},t_{0}) \;e^{-\lambda \beta \mathcal{W}_{t}^{(v)}(\{y_{s}\})} \nonumber \\ &=& \int dy_{t} \! \int_{y_{0}}^{y_{t}}\mathcal{D}y_{s} \!\int dy_{0} \; f_{eq}(y_{t}) \;e^{\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(-\dot{y}_{s},y_{s}\right)} \;e^{\beta\mathcal{W}_{t}^{(v)}(\{y_{s}\})} \frac{1}{f_{eq}(y_{0})} \; f(y_{0},t_{0}) \;e^{-\lambda \beta \mathcal{W}_{t}^{(v)}(\{y_{s}\})} \nonumber \\ &=& \int dy_{0} \! \int_{y_{0}}^{y_{t}}\mathcal{D}y_{s} \! \int dy_{t} \; e^{\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(-\dot{y}_{s},y_{s}\right)} f_{eq}(y_{t}) \;e^{-(1-\lambda)\beta\mathcal{W}_{t}^{(-v)}(\{y_{s}\})} \label{FunctE2b} \\ &=& \int dy_{t} \! \int_{y_{0}}^{y_{t}}\mathcal{D}y_{s} \!\int dy_{0} \; e^{\int_{t_{0}}^{t}ds\; L^{(-v)}\! \left(\dot{y}_{s},y_{s}\right)} f_{eq}(y_{0}) \;e^{-(1-\lambda)\beta\mathcal{W}_{t}^{(-v)}(\{y_{s}\})} \label{FunctE2c} \\ &=& \mathcal{E}_{w}^{(-v)}(1-\lambda,t) \label{FluctTheor2}\end{aligned}$$ where we used Eqs. (\[DetaiBalan1\]) and $\mathcal{W}_{t}^{(-v)}(\{y_{s}\}) = - \mathcal{W}_{t}^{(v)}(\{y_{s}\})$, and the assumption $f(y_{0},t_{0})=f_{eq}(y_{0})$. Here, in the transformation from Eq. (\[FunctE2b\]) to Eq. (\[FunctE2c\]) we changed the integral variables as $y_{s}\rightarrow y_{t+t_{0}-s}$ (so that $\dot{y}_{s}\rightarrow -\dot{y}_{s}$, $y_{t}\rightarrow y_{0}$ and $y_{0}\rightarrow y_{t}$). Therefore, we obtain $\mathcal{E}_{w}^{(v)}(\lambda,t) = \mathcal{E}_{w}^{(-v)}(1-\lambda,t)$, whose combination with Eq (\[EFunctWorkRelat1\]) leads to Eq. (\[FluctTheorWork1\]). Moreover, from Eqs. (\[WorkDistr2\]) and (\[FluctTheorWork1\]) we derive $$\begin{aligned} P_{w}(W,t) &=& \frac{1}{2\pi}\int_{-\infty}^{+\infty}d\lambda \; e^{i\lambda W} \mathcal{E}_{w}^{(v)}(1-i\lambda) \nonumber \\ &=& \frac{1}{2\pi}\int_{-\infty-i}^{+\infty-i}d\mu \; e^{(1-i\mu) W} \mathcal{E}_{w}^{(v)}(i\mu) \label{WorkRate3a} \\ &=& e^{W} \frac{1}{2\pi} \int_{-\infty}^{+\infty}d\mu \; e^{i \mu (-W)} \mathcal{E}_{w}^{(v)}(i\mu) \label{WorkRate3b} \\ &=& e^{W} P_{w}(-W,t) \label{WorkRate3d}\end{aligned}$$ with $\mu \equiv - \lambda - i$. Here, in the transformation from Eq. (\[WorkRate3a\]) to Eq. (\[WorkRate3b\]) we used the fact that noting Eq. (\[EFunctWork1\]) the function $\mathcal{E}_{w}^{(v)}(i\mu)\exp[(1-i\mu) W]$ appearing in Eq. (\[WorkRate3a\]) does not have any pole in the complex region for $\mbox{Im}\{\mu\}\in [0,1]$ (here $\mbox{Im}\{\mu\}$ is the imaginary part of $\mu$). Using Eq. (\[WorkRate3d\]) we obtain Eq. (\[FluctTheorWork2\]). Functional Integral Calculation for the Work Distribution {#CalculationWorkDistribution} ========================================================= In this Appendix, we give calculation details of Eqs. (\[EularLagra2Solve1\]), (\[EFunctWorkTrnas2\]) and (\[PathInteg1\]). Inserting Eq. (\[WorkRate1\]) and (\[Lagra1\]) into Eq. (\[EularLagra2Modif1\]), we obtain $$\begin{aligned} \frac{d^{2} {\tilde{y}^{*}\!\!}_{s}}{ds^{2}} = \frac{{\tilde{y}^{*}\!\!}_{s} + (1-2\lambda) v\tau}{\tau^{2}} \label{EularLagraModif1}\end{aligned}$$ where we used the relations $\alpha = \kappa\tau$ and $D=1/(\alpha\beta)$. \[Note that Eq. (\[EularLagraModif1\]) for ${\tilde{y}^{*}\!\!}_{s}$ is Eq. (\[EularLagra2\]) for $y_{s}^{*}$ except for that Eq. (\[EularLagraModif1\]) use $(1-2\lambda) v$ instead of $v$ in Eq. (\[EularLagra2\]).\] The solution of Eq. (\[EularLagraModif1\]) is given by $$\begin{aligned} {\tilde{y}^{*}\!\!}_{s} +(1-2\lambda) v\tau = \tilde{A}_{1}\exp\left(\frac{s}{\tau}\right) + \tilde{A}_{2}\exp\left(-\frac{s}{\tau}\right) . \label{SolutLagra1}\end{aligned}$$ Here, $\tilde{A}_{1}$ and $\tilde{A}_{2}$ are constants determined by the conditions ${\tilde{y}^{*}\!\!}_{t}=y_{t}$ and ${\tilde{y}^{*}\!\!}_{0}(={\tilde{y}^{*}\!\!}_{t_{0}})=y_{0}$, namely $$\begin{aligned} &&\left(\begin{array}{c} y_{0} + (1-2\lambda)v\tau \\ y_{t} + (1-2\lambda)v\tau \end{array}\right) \nonumber\\ &&{\hspace{0.8cm}}= \left(\begin{array}{cc} \exp\left(\frac{t_{0}}{\tau}\right) & \exp\left(-\frac{t_{0}}{\tau}\right) \\ \exp\left( \frac{t}{\tau}\right) & \exp\left(-\frac{t}{\tau}\right) \end{array}\right) \left(\begin{array}{c} \tilde{A}_{1} \\ \tilde{A}_{2} \end{array}\right) . \label{SolutLagraConst1}\end{aligned}$$ Solving Eq. (\[SolutLagraConst1\]) for $\tilde{A}_{1}$ and $\tilde{A}_{2}$ we obtain $$\begin{aligned} \left(\begin{array}{c} \tilde{A}_{1} \\ \tilde{A}_{2} \end{array}\right) =\left(\begin{array}{c} A_{t-t_{0}}^{((1-2\lambda)v)}(y_{t},y_{0}) \exp\left(-\frac{t}{\tau}\right) \\ A_{-(t-t_{0})}^{((1-2\lambda)v)}(y_{t},y_{0}) \exp\left(\frac{t}{\tau}\right) \end{array}\right) \label{SolutLagraConst2}\end{aligned}$$ with the function $A_{t-t_{0}}^{(v)}(y_{t},y_{0})$ defined by Eq. (\[ConstA1\]). Further, we note $$\begin{aligned} A_{-(t-t_{0})}^{(v)}(y_{t},y_{0}) = A_{t-t_{0}}^{(v)}(y_{0},y_{t}) b_{t} . \label{SolutLagraConst3}\end{aligned}$$ which can be shown from Eq. (\[ConstA1\]). Using Eqs. (\[SolutLagra1\]), (\[SolutLagraConst2\]) and (\[SolutLagraConst3\]) we obtain Eq. (\[EularLagra2Solve1\]). Noting that the Lagrangian function $L^{(v)}(\dot{y}_{s},y_{s})$ defined by Eq. (\[Lagra1\]) and the work rate $\dot{\mathcal{W}}^{(v)}(y_{t})$ given by Eq. (\[WorkRate1\]) are the second order to $y_{t}$ and $\dot{y}_{s}$ at most, we obtain $$\begin{aligned} &&\int_{t_{0}}^{t} ds \left[L^{(v)}(\dot{y}_{s},y_{s}) -\lambda \beta\dot{W}(y_{s})\right] \nonumber \\ &&{\hspace{0.8cm}}= \int_{t_{0}}^{t} ds \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s} +\dot{\tilde{z}}_{s},{\tilde{y}^{*}\!\!}_{s}+\tilde{z}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s}+\tilde{z}_{s})\right] \nonumber \\ &&{\hspace{0.8cm}}= \int_{t_{0}}^{t} ds \; \left\{ L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s}) +\frac{\partial \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s})\right]} {\partial \dot{{\tilde{y}^{*}\!\!}}_{s}} \dot{\tilde{z}}_{s} \right. \nonumber \\ &&{\hspace{0.8cm}}{\hspace{0.8cm}}{\hspace{0.8cm}}\left. +\frac{\partial \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s})\right]} {\partial {\tilde{y}^{*}\!\!}_{s}} \tilde{z}_{s} +\frac{1}{2}\frac{\partial^{2} \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s})\right] } {\partial \dot{{\tilde{y}^{*}\!\!}}_{s}{}^{2}} \dot{\tilde{z}}_{s}^{2} \right. \nonumber \\ &&{\hspace{0.8cm}}{\hspace{0.8cm}}{\hspace{0.8cm}}\left. +\frac{\partial^{2} \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s})\right] } {\partial \dot{{\tilde{y}^{*}\!\!}}_{s} \partial {\tilde{y}^{*}\!\!}_{s}} \dot{\tilde{z}}_{s}\tilde{z}_{s} +\frac{1}{2}\frac{\partial^{2} \left[L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s})\right] } {\partial {\tilde{y}^{*}\!\!}_{s}{}^{2}} \tilde{z}_{s}^{2} \right\} \nonumber \\ &&{\hspace{0.8cm}}= \int_{t_{0}}^{t} ds \; \Biggl\{ L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s}) \left. - \left[ \frac{d}{ds} \frac{\partial L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s})} {\partial \dot{{\tilde{y}^{*}\!\!}}_{s}} -\frac{\partial L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s})} {\partial {\tilde{y}^{*}\!\!}_{s}} +\lambda \beta\frac{\partial \dot{W}({\tilde{y}^{*}\!\!}_{s})} {\partial {\tilde{y}^{*}\!\!}_{s}} \right] \tilde{z}_{s} \right. \nonumber \\ &&{\hspace{0.8cm}}{\hspace{0.8cm}}{\hspace{0.8cm}}-\frac{1}{4D}\left( \dot{\tilde{z}}_{s}^{2} +\frac{2}{\tau} \dot{\tilde{z}}_{s}\tilde{z}_{s} +\frac{1}{\tau^{2}} \tilde{z}_{s}^{2} \right)\Biggr\} \nonumber \\ &&{\hspace{0.8cm}}= \int_{t_{0}}^{t} ds \; \left[ L^{(v)}(\dot{{\tilde{y}^{*}\!\!}}_{s},{\tilde{y}^{*}\!\!}_{s}) -\lambda \beta\dot{W}({\tilde{y}^{*}\!\!}_{s}) +L^{(0)}(\dot{\tilde{z}}_{s},\tilde{z}_{s}) \right] \label{ActioDecom1}\end{aligned}$$ using a partial integral and Eqs. (\[Lagra1\]), (\[EularLagra2Modif1\]) and (\[VariaZ1\]). Inserting Eq. (\[ActioDecom1\]) into Eq. (\[EFunctWorkTrnas1\]) we obtain Eq. (\[EFunctWorkTrnas2\]). Noting $t_{n}\equiv t_{0} + n \Delta t_{N}$, $n=1,2,\cdots,N$, $\Delta t_{N}\equiv (t-t_{0})/N$, the initial time $t_{0}$, the final time $t_{N}=t$, and $\tilde{z}_{0}\equiv \tilde{z}_{t_{0}}=0$ from Eq. (\[VariaZBound1\]) we have $$\begin{aligned} && \sum_{n=0}^{k}\left( \varphi \tilde{z}_{t_{n}} + \tilde{z}_{t_{n+1}}\right)^{2} \nonumber \\ &&{\hspace{0.8cm}}= \sum_{n=1}^{k} \left[ A_{n}(\varphi)+\varphi^{2} \right] \left[ \tilde{z}_{t_{n}} + \frac{\varphi}{A_{n}(\varphi)+\varphi^{2} } \; \tilde{z}_{t_{n+1}} \right]^{2} \nonumber \\ &&{\hspace{0.8cm}}{\hspace{0.8cm}}+ A_{k+1}(\varphi)\tilde{z}_{k+1}, \label{WorkCalcuSuppl1}\end{aligned}$$ for a constant $\varphi$ and $k=1,2,\cdots$, where $A_{n}(\varphi)$ is defined by $$\begin{aligned} A_{n}(\varphi) \equiv \frac{1}{\sum_{k=0}^{n-1}\varphi^{2k}} =\frac{1-\varphi^{2}}{1-\varphi^{2n}}. \label{FuncAn1}\end{aligned}$$ We can prove Eq. (\[WorkCalcuSuppl1\]) for any integer $k$ by mathematical induction, using the fact that the function $A_{n}(\varphi)$ given by Eq. (\[FuncAn1\]) satisfies the recurrence formula $$\begin{aligned} A_{n+1}(\varphi) = \frac{A_{n}(\varphi)}{A_{n}(\varphi)+\varphi^{2}}. \label{FuncAn2}\end{aligned}$$ Using Eq. (\[WorkCalcuSuppl1\]) and $\tilde{z}_{t_{N}}=\tilde{z}_{t}=0$ from Eq. (\[VariaZBound1\]), we obtain $$\begin{aligned} && \sum_{n=0}^{N-1}\left( \varphi \tilde{z}_{t_{n}} + \tilde{z}_{t_{n+1}} \right)^{2} \nonumber \\ &&{\hspace{0.8cm}}= \sum_{n=1}^{N-1} \left[ A_{n}(\varphi)+\varphi^{2} \right] \left[ \tilde{z}_{t_{n}} + \frac{\varphi}{A_{n}(\varphi)+\varphi^{2} } \; \tilde{z}_{t_{n+1}} \right]^{2} \nonumber \\ \label{ProgTheore2}\end{aligned}$$ for a any constant $\varphi$ and $k=1,2,\cdots$. Using the functional integral (\[FunctInteg1\]), the Lagrangian function (\[Lagra1\]) for $v=0$, Eq. (\[ProgTheore2\]) for $\varphi=\varphi_{N} \equiv (\Delta t_{N}/\tau)-1 $, we obtain $$\begin{aligned} && \int_{\tilde{z}_{0}}^{\tilde{z}_{t}} \mathcal{D}\tilde{z}_{s} \; \exp\left[ \int_{t_{0}}^{t}ds\; L^{(0)}\!\left(\dot{\tilde{z}}_{s},\tilde{z}_{s}\right)\right] \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \left(\frac{1}{4\pi D\Delta t_{N}}\right)^{N/2} \int d\tilde{z}_{t_{N-1}} \int d\tilde{z}_{t_{N-2}} \cdots \int d\tilde{z}_{t_{1}} \exp\left[ \sum_{n=0}^{N-1}\Delta t_{N} L^{(0)}\!\left( \frac{\tilde{z}_{t_{n+1}}-\tilde{z}_{t_{n}}}{\Delta t_{N}}, \tilde{z}_{t_{n}} \right)\right] \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \left(\frac{1}{4\pi D\Delta t_{N}}\right)^{N/2} \int d\tilde{z}_{t_{N-1}} \int d\tilde{z}_{t_{N-2}} \cdots \int d\tilde{z}_{t_{1}} \exp\left[ -\frac{1}{4 D \Delta t_{N}} \sum_{n=0}^{N-1} \left( \varphi_{N} \tilde{z}_{t_{n}} + \tilde{z}_{t_{n+1}} \right)^{2}\right] \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \left(\frac{1}{4\pi D\Delta t_{N}}\right)^{N/2} \int d\tilde{z}_{t_{N-1}} \int d\tilde{z}_{t_{N-2}} \cdots \int d\tilde{z}_{t_{1}} \nonumber \\ &&{\hspace{0.8cm}}{\hspace{0.8cm}}{\hspace{0.8cm}}\times \exp\left\{ -\frac{1}{4 D \Delta t_{N}} \sum_{n=1}^{N-1} \left[ A_{n}(\varphi_{N})+\varphi_{N}^{2} \right] \left[ \tilde{z}_{t_{n}} + \frac{\varphi_{N}} {A_{n}(\varphi_{N})+\varphi_{N}^{2} } \; \tilde{z}_{t_{n+1}} \right]^{2}\right\} \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \frac{1}{\sqrt{4\pi D\Delta t_{N}}} \prod_{n=1}^{N-1} \frac{1}{\sqrt{A_{n}(\varphi_{N})+\varphi_{N}^{2} }} \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \frac{1}{\sqrt{4\pi D\Delta t_{N}}} \prod_{n=1}^{N-1} \sqrt{\frac{A_{n+1}(\varphi_{N})}{A_{n}(\varphi_{N})}} \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \sqrt{\frac{A_{N}(\varphi_{N})}{4\pi D\Delta t_{N}}} \nonumber \\ &&{\hspace{0.8cm}}= \lim_{N\rightarrow+\infty} \left\{2\pi D \tau\left( 1-\frac{t-t_{0}}{2\tau N}\right)^{-1} \left[1- \left(1-\frac{t-t_{0}}{\tau N}\right)^{2N} \right] \right\}^{-1/2} \nonumber \\ &&{\hspace{0.8cm}}= \frac{1}{\sqrt{2\pi D \tau \left(1- b_{t}^{2} \right)}} \label{WorkCalcuSuppl2}\end{aligned}$$ where we used Eqs. (\[FuncAn1\]), (\[FuncAn2\]) and $\exp (X) = \lim_{N\rightarrow +\infty} (1+X/N)^{N}$ for any $X$. From Eq. (\[WorkCalcuSuppl2\]) and $\mathcal{T}_{t} = (\tau/2)(1 - b_{t}^{2})$ we derive Eq. (\[PathInteg1\]). Nonequilibrium Detailed Balance Including Inertia {#NonequilibriumDetailedBalanceInertia} ================================================= In this Appendix, we give a derivation of Eq. (\[DetaiBalan7\]). Using Eq. (\[Lagra3\]) we have $$\begin{aligned} L^{(v)}\!\left(\ddot{y}_{s},\dot{y}_{s},y_{s}\right) &=& -\frac{1}{4D} \left[ -\dot{y}_{s} +\frac{1}{\tau} y_{s} \pm v +\frac{m}{\alpha}\ddot{y}_{s} +2\dot{y}_{s} +(1\mp 1)v \right]^{2} \nonumber \\ &=& -\frac{1}{4D} \left(-\dot{y}_{s} +\frac{1}{\tau} y_{s} \pm v +\frac{m}{\alpha}\ddot{y}_{s} \right)^{2} -\frac{1}{D}\left(\frac{1}{\tau} y_{s} \pm v +\frac{m}{\alpha}\ddot{y}_{s} \right) \dot{y}_{s} \nonumber \\ &&{\hspace{0.8cm}}-\frac{1}{D}\left(\dot{y}_{s} +\frac{1}{\tau} y_{s} +\frac{m}{\alpha}\ddot{y}_{s} \right) \frac{1\mp 1}{2}v \nonumber \\ &=& L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right) -\beta\left[\left( m\ddot{y}_{s}+ \kappa y_{s} \right) \left(\dot{y}_{s}+\frac{1\mp 1}{2}v\right) + \alpha \dot{y}_{s}\frac{1\pm 1}{2} v \right] \nonumber \\ &=& L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right) -\beta\left[ m\ddot{y}_{s} (\dot{y}_{s} + \vartheta v) + \kappa y_{s} \dot{y}_{s} \right] +\beta m\ddot{y}_{s} \vartheta v \nonumber \\ &&{\hspace{0.8cm}}-\beta\left( m\ddot{y}_{s} + \kappa y_{s} \right)\frac{1\mp 1}{2}v - \beta \alpha \dot{y}_{s}\frac{1\pm 1}{2} v \nonumber \\ &=& L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right) -\beta\frac{d}{ds} \left[ \frac{1}{2} m (\dot{y}_{s} + \vartheta v)^{2} + \frac{1}{2}\kappa y_{s}^{2} \right] \nonumber \\ &&{\hspace{0.8cm}}-\beta\left[ \kappa y_{s} \frac{1\mp 1}{2} + \alpha \dot{y}_{s}\frac{1\pm 1}{2} + m\ddot{y}_{s} \left(\frac{1\mp 1}{2}-\vartheta\right) \right] \nonumber \\ &=& L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right) -\beta\frac{d \mathcal{H}\!\left( \dot{y}_{s}+\vartheta v,y_{s}\right)}{ds} +\beta\Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s}, y_{s};\vartheta) v \label{DetaiBalan7sup}\end{aligned}$$ where we used Eqs. (\[HamilFunct1\]) and (\[GenerForce1\]), and $\vartheta$ is a parameter. Equation (\[DetaiBalan7sup\]) leads to $$\begin{aligned} && e^{-\beta \int_{t_{0}}^{t}ds\; \Lambda_{\pm}(\ddot{y}_{s},\dot{y}_{s},y_{s};\vartheta) v} e^{\int_{t_{0}}^{t} ds\; L^{(v)}\!\left(\ddot{y}_{s},\dot{y}_{s},y_{s}\right)} e^{-\beta \mathcal{H}\!\left( \dot{y}_{0}+\vartheta v,y_{0}\right)} \nonumber \\ &&{\hspace{0.8cm}}= e^{-\beta \mathcal{H}\!\left( \dot{y}_{t}+\vartheta v,y_{t}\right)} e^{\int_{t_{0}}^{t} ds\; L^{(\pm v)}\left(\ddot{y}_{s},-\dot{y}_{s},y_{s}\right)} . \label{ModifDetaiBalanZero}\end{aligned}$$ Equation (\[DetaiBalan7\]) is derived from Eq. (\[ModifDetaiBalanZero\]) using Eq. (\[EquilDistr2\]). [99]{} A. Einstein, Ann. Physik **17**, 549 (1905). J. B. Johnson, Phys. Rev. **32**, 97 (1928). H. Nyquist, Phys. Rev. **32**, 110 (1928). L. Onsager, Phys. Rev. **37**, 405 (1931). L. Onsager, Phys. Rev. **38**, 2265 (1931). H. B. G. Casimir, Rev. Mod. Phys. **17**, 343 (1945). M. S. Green, J. Chem. Phys. **19**, 1036 (1951). H. B. Callen and T. A. Welton Phys. Rev. **83**, 34 (1951). R. Kubo, J. Phys. Soc. Jap. **12**, 570 (1957). N. Hashitsume, Prog. Theor. Phys. **8**, 461 (1952). L. Onsager and S. Machlup, Phys. Rev. **91**, 1505 (1953). S. Machlup and L. Onsager, Phys. Rev. **91**, 1512 (1953). H. Hasegawa, Prog. Theor. Phys. **56**, 44 (1976). H. Hasegawa, Prog. Theor. Phys. **58**, 128 (1977). K. Yasue, J. Math. Phys. **19**, 1671 (1978). K. L. C. Hunt and J. Ross, J. Chem. Phys. **75**, 976 (1981). H. Risken, [*The Fokker-Planck equation : methods of solution and applications*]{} (Springer-Verlag Berlin, 1989). L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim, Phys. Rev. Lett. **87**, 040601 (2001). L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim, J. Stat. Phys. **107**, 635 (2002). G. Gallavotti, ESI preprint 1144. D. J. Evans, E. G. D. Cohen, and G. P. Morriss, Phys. Rev. Lett. **71**, 2401 (1993); **71**, 3616 (1993) \[errata\]. D. J. Evans and D. J. Searles, Phys. Rev. E **50**, 1645 (1994). G. Gallavotti and E. G. D. Cohen, Phys. Rev. Lett. **74**, 2694 (1995). J. Kurchan, J. Phys. A: Math. Gen. **31**, 3719 (1998). J. L. Lebowitz and H. Spohn, J. Stat. Phys. **95**, 333 (1999). G. E. Crooks, Phys. Rev. E **60**, 2721 (1999). G. E. Crooks, Phys. Rev. E **61**, 2361 (2000). S. Ciliberto and C. Laroche, J. Phys. IV France **8**, 215 (1998). G. M. Wang, E. M. Sevick, E. Mittag, D. J. Searles, and D. J. Evans, Phys. Rev. Lett. **89**, 050601 (2002). S. Ciliberto, N. Garnier, S. Hernandez, C. Lacpatia, J.-F. Pinton, G. R. Chavarria, Physica A **340**, 240 (2004). K. Feitosa and N. Menon, Phys. Rev. Lett. **92**, 164301 (2004). N. Garnier and S. Ciliberto, Phys. Rev. E **71**, 060101R (2005). S. Schuler, T. Speck, C. Tietz, J. Wrachtrup, and U. Seifert, Phys. Rev. Lett. **94**, 180602 (2005). G. Gallavotti, Phys. Rev. Lett. **77**, 4334 (1996). R. van Zon and E. G. D. Cohen, Phys. Rev. Lett. **91**, 110601 (2003). R. van Zon and E. G. D. Cohen, Phys. Rev. E **69**, 056121 (2004). K. Sekimoto, Prog. Theor. Phys. Suppl. **130**, 17 (1998). O. Mazonka and C. Jarzynski, e-print cond-mat/9912121. S. Tasaki, I. Terasaki and T. Monnai, e-print cond-mat/0208154. R. van Zon and E. G. D. Cohen, Phys. Rev. E **67**, 046102 (2003). R. van Zon, S. Ciliberto and E. G. D. Cohen, Phys. Rev. Lett. **92**, 130601 (2004). We note that also the equilibrium fluctuating equations of Onsager and Machlup have a Langevin form [@OM53; @MO53]. T. Taniguchi and G. P. Morriss, Phys. Rev. E **70**, 056124 (2004). Mathematically, the $v$-dependence in the Langevin equation (\[LangeEq2\]) can formally be removed by changing the variable $y_{t}$ by $y_{t}+v\tau$. N. G. van Kampen, *Stochastic processes in physics and chemistry* (Elsevier, Amsterdam, 1992). A concrete calculation process of the functional integration to derive Eq. (\[TransProba2\]) from Eq. (\[TransProba1\]) is similar to the one for the work distribution function which will be discussed in Sec. \[FunctionalCalculationWorkDistribution\]. More concretely, the transition probability ${ {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) }$ is given by $ { {F}\!(\begin{array}{c} \!\! \scriptstyle{y_{t}} \!\! \\ \!\! \scriptstyle{t} \!\! \end{array} | \begin{array}{c} \!\! \scriptstyle{y_{0}} \!\! \\ \!\! \scriptstyle{t_{0}} \!\! \end{array}) } = \mathcal{F}(y_{t},y_{0};0) $ using the function $\mathcal{F}(y_{t},y_{0}; \lambda)$ defined by Eq. (\[EFunctWorkTrnas1\]), whose functional integral is carried out for any $\lambda$ in Sec. \[FunctionalCalculationWorkDistribution\]. L. Tisza and I. Manning, Phys. Rev. **105**, 1695 (1957). L. D. Landau and E. M. Lifshitz, *Mechanics*, translated from the Russian by J. B. Sykes and J. S. Bell (Pergamon Press, Oxford, 1960). More concretely, the solution of Eq. (\[EularLagra2\]) under the conditions $y_{t_{0}}^{*}=y_{0}$ and $y_{t}^{*}=y_{t}$ is given by the case of $\lambda =0$ for ${\tilde{y}^{*}\!\!}_{s}$ in Eq. (\[EularLagra2Solve1\]) which will be discussed in Sec. \[FunctionalCalculationWorkDistribution\] later. L. D. Landau and E. M. Lifshitz, *Statistical physics*, translated from the Russian by E. Peierls and R. F. Peierls (Pergamon Press, London, 1958) Chapter XII. R. P. Feynman and A. R. Hibbs, *Quantum mechanics and path integrals* (McGraw-Hill, New York, 1965). The asymmetry in the nonequilibrium detailed balance relation appears to correspond to the asymmetry noted by Bertini *et al.* [@BSG01; @BSG02] in the creation and decay of a fluctuation in a nonequilibrium steady state. (In Refs. [@BSG01; @BSG02] such an asymmetry is called an Onsager-Machlup symmetry, and we will discuss this point more in Eq. (\[OMSymme1\]) in Sec. \[ConclusionsRemarks\].) If so, this asymmetry was applied in Refs. [@BSG01; @BSG02] to exclusion and boundary driven zero range models, while here it applies to a stochastic model using the Langevin or the Onsager-Machlup approach. D. J. Evans and D. J. Searles, Adv. Phys. **51**, 1529 (2002). In this paper we call the transient fluctuation theorem as a fluctuation theorem with the equilibrium initial condition. E. G. D. Cohen and G. Gallavotti, J. Stat. Phys. **96**, 1343 (1999). In Eq. (\[EFunctWorkTrnas1\]) the dimensionless work rate is $\beta\dot{\mathcal{W}}^{(v)}(y_{s})$, multiplied by the Lagrange multiplier $\lambda$. Similarly, the third term on the right-hand side of Eq. (\[EularLagra2Modif1\]) may be regarded as a term for the Lagrange multiplier under the restriction by the delta function in Eq. (\[WorkDistr1\]). J. C. Reid, D. M. Carberry, G. M. Wang, E. M. Sevick, and D. J. Evans, Phys. Rev. E **70**, 016111 (2004). F. Bonetto, G. Gallavotti, A. Giuliani, and F. Zamponi, e-print cond-mat/0507672. T. Gilbert, Phys. Rev. E **73**, 035102R (2006). M. Baiesi, T. Jacobs, C. Maes, and N. S. Skantzos, e-print cond-mat/0602311. V. Y. Chernyak, M. Chertkov, and C. Jarzynski, e-print cond-mat/0605471. O. Narayan and A. Dhar, J. Phys. A: Math. Gen. **37**, 63 (2004). K. Kitahara, a private communication. A. Imparato and L. Peliti, e-print cond-mat/0603506. It should be noted that the extended heat fluctuation theorem may also depend on the initial condition. \[Note that in our simple argument for heat \[based on Eq. (\[HeatDistr3\]), etc.\] in the second half of Sec. \[FluctuationTheoremHeat\] we assumed a canonical-like distribution as the initial distribution.\] Actually, if we could choose the initial distribution $f(y_{0},t_{0})$ as a constant \[although in this case $f(y_{0},t_{0})$ cannot be normalized for $y_{0}\in(-\infty,+\infty)$\] then we can show that the heat satisfies the conventional form of fluctuation theorem $P_{q}(Q,t)/P_{q}(-Q,t) = \exp(Q)$ for any time, which is derived from Eq. (\[DetaiBalan6\]) for the case of $f_{ref}(y_{0})$ to be constant, or from Eq. (\[EFunctHeat3\]) leading to the relation $\mathcal{E}_{q}(\lambda,t) = \mathcal{E}_{q}(1-\lambda,t)$ in this case.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In an effective operator approach, the full set of leading contributions to anomalous top couplings comprises various new trilinear as well as higher interaction vertices, some of which are related to one another by gauge symmetry or equations of motion. In order to study trilinear top couplings to SM gauge bosons such as $tt\gamma$, $ttZ$, $tbW$ and $ttg$, the operator set can be restricted accordingly. However, the complete basis cannot be mapped onto an on-shell parametrisation of the trilinear vertices alone. Four-fermion contact terms $qqtt$ and $udtb$ must be included if the relation to the operator basis is to be retained. In this paper, we point out how these interactions contribute to the single top search channels for anomalous trilinear $tbW$ couplings at the LHC and Tevatron, thus affecting the corresponding bounds. All results are based on full leading-order partonic matrix elements, thus automatically accounting for off-shell and interference effects as well as irreducible backgrounds. A discussion of the quantitative effects of going from on-shell tops to full matrix elements including acceptance cuts is also provided.' author: - Fabian Bach - Thorsten Ohl bibliography: - 'references.bib' title: Anomalous top couplings at hadron colliders revisited --- Introduction ============ The *Standard Model of Particle Physics* (SM) has stood its ground during the past decades with great success, consistently explaining and predicting a great variety of high energy experiments with unchallenged precision. One of the major cornerstones was the discovery of the top quark at the Tevatron in 1995 [@Abe:1995hr; @Abachi:1995iq], confirming the postulated three-family doublet structure of the SM. While the Tevatron experiments have continued to collect data and improve their measurements of top properties, most importantly its mass [@Lancaster:2011wr; @Flyagin:2012zz; @Aaltonen:2012va; @Aaltonen:2012rz], most attention is now directed to the LHC up and running at , and the results of its multi-purpose experiments ATLAS and CMS improving on top statistics by the day. By now, top pair production has been measured in different channels with remarkable accuracy [@Aad:2010ey; @Chatrchyan:2011ew; @Chatrchyan:2011nb; @ATLAS:2012aa; @Chatrchyan:2012vs; @Aad:2012xh]. Single top production has already been established for associated production despite its small cross section [@Chatrchyan2012; @ATLAS-CONF-2011-104] and even been definitely observed in the dominant $t$ channel  [@Chatrchyan:2011vp; @Aad:2012ux; @Chatrchyan1209]. The ever-growing abundance of top events at the LHC is beginning to allow the determination of more involved observables such as asymmetries, invariant pair mass distribution and top couplings to the other SM particles with high precision (cf. e.g., [@Schilling:2012dx; @Chiochia:2012vg] for an overview). On the theoretical side, the top quark takes an outstanding place among the spectrum of SM particles as a possible window to new non-SM physics because of its uniquely large mass of the order of the electroweak symmetry breaking (EWSB) scale $m_t\sim\upsilon\sim\mathcal O(\unit[100]{GeV})$, with its role within the dynamics of EWSB still unresolved. Corresponding new physics effects in the top sector may manifest themselves in the deviation of top properties from their SM values, where the main attention in this paper is directed towards the trilinear couplings to SM gauge bosons, especially the charged-current (CC) interaction $tbW$. Therefore, even before any experimental analysis, a theoretically robust parametrisation of these anomalous couplings has to be found, at the same time reducing the parameter space to an experimentally manageable minimum while staying fully general within the basis of effective operators generating these couplings at Lagrangian level. Indeed, starting from the complete set of effective dimension six operators as written down by Buchmüller and Wyler already in 1985 [@Buchmuller:1985jz], substantial effort has been put into this task in the past decades by various authors [@Arzt:1994gp; @Gounaris:1996vn; @Gounaris:1996yp; @Brzezinski:1997av; @Whisnant:1997qu; @Yang:1997iv; @Grzadkowski:2003tf; @AguilarSaavedra:2008zc; @AguilarSaavedra:2009mx; @AguilarSaavedra:2010zi; @Grzadkowski:2010es]. The crucial ingredient of most of these analyses is to employ the theorem [@Arzt:1993gz; @Weinberg:1980wa; @Gasser:1983yg; @Georgi:1991ch; @Rujula:1991se] that the field equations of motion (EOM) can safely be used at the Lagrangian level at a fixed order in the effective operator expansion in order to rewrite operators and identify redundant structures. As a result of this procedure, it is often argued that these redundancies allow for a reduction of independent couplings to be incorporated in a phenomenological analysis of anomalous top couplings. However, as has been pointed out e.g., in [@Grzadkowski:2003tf; @AguilarSaavedra:2008zc] and will also be reviewed again in more detail later on, the application of the EOM necessarily generates four-fermion contact interactions, which are nevertheless often dropped from the analyses for the sake of simplicity. We note that the latter procedure does not correspond to a *rewriting* but rather to a *redefinition* of the originally chosen operator basis, thus also departing from the full generality of the original basis with respect to the richness of structures in the trilinear couplings. Still, the operator equalities derived by various authors [@Grzadkowski:2003tf; @AguilarSaavedra:2008gt; @AguilarSaavedra:2008zc] and systematically presented in [@AguilarSaavedra:2008zc] are very useful to simplify an implementation of the most general set of trilinear top couplings into a Monte Carlo (MC) generator in a gauge invariant way, so in our approach, rather than dropping part of the physics, we make use of these equalities to implement *all* the trilinear top couplings to SM gauge bosons in the language of on-shell couplings including the required quartic contact terms into the parton-level MC event generator <span style="font-variant:small-caps;">Whizard</span> [@Kilian:2007gr], also addressing the interplay of anomalous top and bottom couplings—the latter already heavily bounded by LEP data—and the repercussions on the top couplings. Finally, we present phenomenological consequences obtained with our implementation for the parameter space of the anomalous couplings in the CC sector. This article is organized as follows: in section \[theo\] we review the procedure described above of defining a complete operator basis to generate anomalous top couplings to SM gauge bosons and applying the EOM to rewrite some of these operators, thus arriving at the most suitable form for a MC implementation. In section \[pheno\] we discuss the LHC phenomenology with a focus on single top production, including a comparison of on-shell and full matrix element approaches to retrieve the cross sections at detector level as functions of the anomalous $tbW$ couplings as well as a presentation of the physical effects and consequences of the newly added coupling structures. A discussion and summary of the main statements and results can be found in section \[sum\]. Theoretical setup {#theo} ================= In order to be self-contained in this article, we start this section by reviewing in some detail the main steps of the procedure presented in [@AguilarSaavedra:2008zc] to simplify the most general set of operators generating the trilinear anomalous top couplings to SM gauge bosons $tt\gamma$, $ttZ$, $tbW$ and $ttg$. However, although we emphasize that all of these couplings are implemented in <span style="font-variant:small-caps;">Whizard</span> in a gauge-invariant way, including all quartic terms generated by the operator rewriting, we will restrict ourselves here to the discussion of the CC sector, i. e. only those operators generating anomalous contributions to the $tbW$ interaction, for two simple reasons: 1. the complexity of the parameter space in the CC sector is increased in a minimal way compared to previous studies [@delAguila:2002nf; @AguilarSaavedra:2006fy; @AguilarSaavedra:2007rs; @AguilarSaavedra:2008gt], because it turns out that only one additional operator (and hence coupling) has to be considered; 2. at hadron colliders, the experimental access to the new effects is most straightforward, combining studies of CC single top production and of top decay products. In the following two subsections we develop the basic ingredients of the effective operator analysis, recapitulate the operator rewriting procedure and finally present our extended parameter space for the anomalous $tbW$ couplings. Effective operator approach and operator basis {#eff_op} ---------------------------------------------- There are basically two ways to tackle new physics beyond the SM in a systematic and consistent manner: Either the model building (top-down) approach, i. e. starting from a postulated Lagrangian—which incorporates a sensible UV completion—and deriving from it physical effects to which present or planned experiments might be sensitive, or the effective (bottom-up) approach, i. e. starting from the established SM symmetries and *a priori* considering *all* possible new physics effects compatible with these symmetries at the Lagrangian level, postponing the question which larger theory might generate the relevant parameters at a higher energy scale $\Lambda$. Since we follow the second approach, it shall be clarified a little further. The idea is to confront new physics completely unbiased, that is without any assumptions about the dynamical degrees of freedom generating it, and to study the effects that are manifest at a testable energy scale (considerably smaller than the resonant scale $\Lambda$) where the degrees of freedom are the well known SM particles. This corresponds to integrating out the heavy modes, thus generating effective operators $O^{(d)}_i$ of mass dimension $d>4$ which are normalized by appropriate powers of $\Lambda$. In the model-independent approach, the effective Lagrangian can be written as an expansion in $1/\Lambda$ [@Burges:1983zg; @Leung:1984ni; @Buchmuller:1985jz]: $$\mathcal{L}_{\text{eff}} = \mathcal{L}_{\text{SM}} + \sum_{d>4,i} \frac{C_i^{(d)}}{\Lambda^{d-4}} O_i^{(d)} + \text{h.c.}$$ with dimensionless operator coefficients $C^{(d)}_i$, comprising *all* possible effective operators built from SM fields and derivatives only, and compatible with all local and global SM symmetries. A complete set of these operators for $d=5,6$ can be found in [@Buchmuller:1985jz]. The only possible $d=5$ operator in this setup is a neutrino mass term [@Buchmuller:1985jz], so the leading contributions to anomalous trilinear top couplings must be $d=6$. The complete operator list at this order can be found in [@AguilarSaavedra:2008zc; @AguilarSaavedra:2009mx; @AguilarSaavedra:2010zi], of which we now quote the ones relevant to trilinear $tbW$ interactions (also adopting the nomenclature of [@AguilarSaavedra:2008zc]): \[op\_basis\] $$\begin{aligned} O^{(3,ij)}_{\phi q} &= i\big( \phi^\dagger \tau^I D_\mu \phi \big) \big( \bar{q}_{Li} \gamma^\mu \tau^I q_{Lj} \big)~,\label{Opq} \\ O^{ij}_{\phi\phi} &= i\big( \tilde{\phi}^\dagger D_\mu \phi \big) \big( \bar{u}_{Ri} \gamma^\mu d_{Rj} \big)~,\label{Opp} \\ O^{ij}_{uW} &= \big( \bar{q}_{Li} \sigma^{\mu\nu} \tau^I u_{Rj} \big) \tilde{\phi}\, W^I_{\mu\nu}~,\label{OuW} \\ O^{ij}_{dW} &= \big( \bar{q}_{Li} \sigma^{\mu\nu} \tau^I d_{Rj} \big) \phi\, W^I_{\mu\nu}~,\label{OdW} \\ O^{ij}_{qW} &= \big( \bar{q}_{Li} \gamma^\mu \tau^I D^\nu q_{Lj} \big) W^I_{\mu\nu}~,\label{OqW}\end{aligned}$$ with generation indices $i,j=1,2,3$ and the non-Abelian $\mathbf{SU}(2)_L$ field strength components $$W_{\mu\nu}^I = \partial_\mu W^I_\nu - \partial_\nu W^I_\mu -g\, \varepsilon_{IJK} W^J_\mu W^K_\nu$$ to be contracted with the Pauli matrices $\tau^I$ ($I=1,2,3$). The $q_{L(R)i}$ are left(right)-handed quark spinors in the electroweak isodoublet (isosinglet) representation, and $\phi$ is the isodoublet complex SM scalar field acquiring a vev $\langle\phi\rangle=\frac{1}{\sqrt 2}(0,\upsilon)^T$, and $\tilde\phi=i\tau^2\phi^*$. Of the other electroweak operators listed in [@AguilarSaavedra:2008zc], $O^{(1,ij)}_{\phi q}$ and $O^{ij}_{\phi u}$ as well as all those containing the hypercharge field strength $B_{\mu\nu}$ only contribute to NC interactions, whereas the operators $O^{ij}_{Du}$, $O^{ij}_{\bar{D}u}$, $O^{ij}_{Dd}$ and $O^{ij}_{\bar{D}d}$ appear to contribute to the $tbW$ vertex. However, the differences $O^{ij}_{Du(d)} - O^{ij}_{\bar{D}u(d)}$ are entirely redundant as is shown in [@AguilarSaavedra:2008zc], and the sums $O^{ij}_{Du(d)} + O^{ij}_{\bar{D}u(d)}$ are proportional to the gauge boson momentum $q^\mu=(p_i-p_j)^\mu$ so that amplitudes containing these vertices vanish either for physical on-shell $W$ or for on-shell light fermions coupling to the $W$, which is always the case at parton level for all processes to be considered for single top and top decay studies discussed here. So Eq.  represents the most general $d=6$ operator basis generating anomalous $tbW$ couplings, in which we shall therefore be complete in our phenomenological studies. We could now straightforwardly start off to find and implement all interactions generated by the operator basis which could appear in the relevant amplitudes. However, this can become a rather involved business particularly for $O^{ij}_{qW}$ which contains, apart from the trilinear coupling, also some relevant quartic terms such as e.g., $tbW\!g$ with a complicated Dirac and momentum structure. Therefore, in order to facilitate the implementation work we rather follow the operator rewriting procedure of [@AguilarSaavedra:2008zc], illustrating the main steps here for $O^{ij}_{qW}$: starting from its decomposition into hermitian and anti-hermitian parts, the hermitian part becomes $$\label{OqW_herm} \frac{1}{2} \left[ O^{ij}_{qW} + \big( O^{ji}_{qW} \big)^\dagger \right] = \frac{1}{2} \big( \bar{q}_{Li} \gamma^\mu \tau^I q_{Lj} \big) \big( D^\nu W_{\nu\mu} \big)^I$$ (dropping the total derivative), where the EOM of the $W$ field $$\label{eom_W} \big( D^\nu W_{\nu\mu} \big)^I = g \left\{ \bar{\ell}_{Li} \gamma^\mu \frac{\tau^I}{2} \ell_{Li} + \bar{q}_{Li} \gamma^\mu \frac{\tau^I}{2} q_{Li} + i\left[ \phi^\dagger \frac{\tau^I}{2} D^\mu \phi - \bigl( D^\mu \phi^\dagger \bigr) \frac{\tau^I}{2} \phi \right] \right\}$$ can be applied to replace the derivative. On the other hand, with some algebra [@Buchmuller:1985jz; @AguilarSaavedra:2008zc] the anti-hermitian part can be brought in the form $$\frac{1}{2} \left[ O^{ij}_{qW} - \big( O^{ji}_{qW} \big)^\dagger \right] = - \frac{1}{4} \left( \bar{q}_{Li} \sigma^{\mu\nu} \tau^I i\slashed D q_{Lj} W^I_{\mu\nu} - \text{h.c.} \right)$$ up to total derivatives, where the EOM of the quark field $$\label{eom_q} i\slashed D q_{Li} = Y^u_{ij} u_{Rj} \tilde\phi +Y^d_{ij} d_{Rj} \phi$$ (introducing Yukawa matrices $Y^{u/d}$) can be inserted. Joining it all together, one arrives at the operator equality \[rewrite\] $$\begin{aligned} O^{ij}_{qW} =& +\frac{g}{4} \bigl( \bar{q}_{Li} \gamma^\mu \tau^I q_{Lj} \bigr) \Bigl[ \bigl( \bar{\ell}_{Lk} \gamma_\mu \tau^I \ell_{Lk} \bigr) +\bigl( \bar{q}_{Lk} \gamma_\mu \tau^I q_{Lk} \bigr) \Bigr] \label{rewrite_contact} \\ & +\frac{g}{4} \left[ O^{(3,ij)}_{\phi q} + \bigl( O^{(3,ji)}_{\phi q} \bigr)^\dagger \right] \label{rewrite_Opq} \\ & -\frac{1}{4} \left[ Y^u_{jk}\, O^{ik}_{uW} + Y^d_{jk}\, O^{ik}_{dW} - Y^{u\dagger}_{ki} \bigl( O^{jk}_{uW} \bigr)^\dagger - Y^{d\dagger}_{ki} \bigl( O^{jk}_{dW} \bigr)^\dagger \right]~.\label{rewrite_OuWdW}\end{aligned}$$ Obviously, the terms in  and  are redundant and can be absorbed into the operators –, whereas  generates four-fermion contact interactions. Although it is clear that without further restrictions there is enough freedom within the operator basis to independently vary all the couplings emerging from – and the associated contact terms coming from the rewriting , the rewriting procedure corresponds to a shift of the original operator coefficients. Setting $i=j=3$ and dropping all generation superscripts from now on, these shifts are: $$\begin{aligned} \label{coeff_shift} \delta \text{Re}\,C_{\phi q}^{(3)} &= \frac{g}{2} \text{Re}\,C_{qW}~, \nonumber \\ \delta \text{Im}\,C_{uW} &= -\frac{m_t}{\sqrt 2\upsilon} \text{Im}\,C_{qW}~, \nonumber \\ \delta \text{Im}\,C_{dW} &= -\frac{m_b}{\sqrt 2\upsilon} \text{Im}\,C_{qW} \simeq 0~,\end{aligned}$$ assuming an approximate decoupling of the third generation in the Yukawa matrices. With this setup, we can now go on to physical states of the gauge and matter fields and write down the interaction terms generated by our operator basis. Parametrisation of anomalous charged-current couplings {#anom_top} ------------------------------------------------------ Inserting the scalar vev and physical states of the gauge fields into the operators  and forming hermitian combinations $C_xO_x+C_x^*O_x^\dagger$, one finds various trilinear interaction terms $tbW$, $ttZ$, $ttA$ but also $bbZ$ and $bbA$, as well as associated quartic interactions $ttWW$, $bbWW$, $tbW\!Z$ and $tbW\!A$ which are all necessary to maintain gauge invariance in the resulting amplitudes, and have therefore been included in our implementation. The resulting effective $tbW$ interaction Lagrangian can be written as \[L\_tbW\] $$\begin{aligned} \mathcal{L}_{tbW} =& -\frac{g}{\sqrt 2} \bar{b}\,\gamma^\mu \bigl( V_L P_L + V_R P_R \bigr)\,t\,W^-_\mu + \text{h.c.} \label{V_LR} \\ & -\frac{g}{\sqrt 2} \bar{b}\,\frac{i\sigma^{\mu\nu}q_\nu}{m_W} \bigl( g_L P_L + g_R P_R \bigr)\,t\,W^-_\mu + \text{h.c.} \label{g_LR} \\ & -\frac{g}{\sqrt 2} \bar{b}\,\gamma^\mu \frac{q^2-m_W^2}{m_W^2} \bigl( V_L^\text{off} P_L \bigr)\,t\,W^-_\mu + \text{h.c.}~, \label{V_off}\end{aligned}$$ where all couplings except for $V_L\equiv V_{tb}\simeq1$ vanish in the SM at tree level, and get the following anomalous contributions from operator coefficients: $$\begin{aligned} \label{cpl} \delta V_L &= \left( C_{\phi q}^{(3)*} + \frac{g}{2} \text{Re}\,C_{qW} \right) \frac{\upsilon^2}{\Lambda^2}~, &%\nonumber \\ \delta g_L &= \sqrt 2 C_{dW}^* \frac{\upsilon^2}{\Lambda^2}~, \nonumber \\ \delta V_R &= \frac{1}{2} C_{\phi\phi}^* \frac{\upsilon^2}{\Lambda^2}~, &%\nonumber \\ \delta g_R &= \sqrt 2 C_{uW} \frac{\upsilon^2}{\Lambda^2}~, \nonumber \\ \delta V_L^\text{off} &= \frac{g}{2} \text{Re}\,C_{qW} \frac{\upsilon^2}{\Lambda^2}~.\end{aligned}$$ The interaction terms  and  represent the on-shell parametrisation widely used in various phenomenological studies (normalization convention taken from [@AguilarSaavedra:2008zc]), which is retrieved from the operators –. The interaction  emerges from the hermitian part of $O_{qW}$, $$O_{qW} + O_{qW}^\dagger = \big( \bar{q}_{L3} \gamma^\mu \tau^I q_{L3} \big) \big( \partial^2 W_\mu^I \big) + \text{higher contact interactions,}$$ cf. Eq. , which—unlike the anti-hermitian part—cannot be completely recast into a combination of the other four operators. However, the *partial* redundance of $O_{qW}$ has been made explicit in the parametrisation  by defining its on-shell part into $V_L$ so that any contribution $\sim V_L^\text{off}$ vanishes when the $W$ goes on the mass shell. Hence it is no surprise that in $\delta V_L$ of Eq.  we find again the shift of the coefficient $C_{\phi q}^{(3)}$ already stated in Eq.  after the operator rewriting. Furthermore, by comparison to Eq. , one finds that all contributions $\sim V_L^\text{off}$ must be in one-to-one correspondence to the four-fermion contact interactions given in , which is also highlighted by the fact that in physical amplitudes the kinematic structure of the $W$ propagator is exactly cancelled by the $q$-dependent vertex. We have now isolated the non-redundant contribution of $O_{qW}$ to the $tbW$ interaction Lagrangian, and also identified the most convenient way to implement it in a gauge-invariant way, namely by adding the quartic fermion vertices $$\begin{aligned} \Delta \mathcal{L} =& \frac{g_\times}{\Lambda^2} \bigl( \bar{b} \gamma^\mu P_L t \bigr) \Bigl[ \bigl( \bar{u}_k \gamma_\mu P_L d_k \bigr) + \bigl( \bar{\nu}_k \gamma_\mu P_L e_k \bigr) \Bigr] + \text{h.c.} \\ &\qquad\text{with } g_\times = g\,\text{Re}\,C_{qW} \nonumber\end{aligned}$$ (cf. [@Grzadkowski:1997cj; @Cao:2007ea; @AguilarSaavedra:2010zi]), giving a relation of coefficients $$V_L^\text{off} = \frac{\upsilon^2}{2\Lambda^2}\,g_\times~.$$ Of course, one might ask at this point if such a coupling structure should be counted among the anomalous $tbW$ sector, but then again it must be noted that as a consequence of the common operator basis, the trilinear couplings are related to $V_L^\text{off}$ through the underlying operator coefficients. Specifically, Eq.  illustrates that a limit on $\delta V_L$ cannot be unambiguously mapped onto a limit on the operator coefficient $C_{\phi q}^{(3)}$ without also bounding $\delta V_L^\text{off}$ (or the anomalous NC sector, see below). Moreover, the operator basis  and the corresponding set of couplings  parametrise *all* anomalous diagram insertions which can interfere with the SM diagram in a minimal way, making this approach consistent at the amplitude level. Finally, as pointed out in Sec. \[VO\], the inclusion of the additional coupling also affects the interpretation of current and upcoming experimental results at the LHC. Before moving on to the phenomenological implications, let us discuss briefly the issue of anomalous bottom couplings within the effective theory approach: Since the original effective operators by construction respect the full electroweak gauge symmetry $\mathbf{SU}(2)_L\times\mathbf{U}(1)_Y$, it is no surprise to find certain relations within the set of anomalous electroweak couplings of the heavy doublet $(t,b)$ after spontaneous symmetry breaking. For example, an anomalous CC contribution $\delta V_L$ is directly related to the anomalous left-handed NC vector couplings $ttZ$ and $bbZ$, the latter one stringently constrained by LEP data, so turning on $\delta V_L$ while respecting all existing bounds necessarily implies a non-vanishing anomalous contribution to the left-handed $ttZ$ vector coupling [@AguilarSaavedra:2008zc; @Berger:2009hi], or a fine-tuned relation with $\delta V_L^\text{off}$, cf. Eq. . Similarly, $\delta g_R$ is directly related to the anomalous $ttZ$/$tt\gamma$ tensor couplings, just like $\delta g_L$ is to the $bbZ$/$bb\gamma$ ones (cf. e.g., [@AguilarSaavedra:2008zc; @Aguilar-Saavedra2012] for details). In short, it is impossible to vary the anomalous CC couplings in a consistent way within the effective operator approach without either getting anomalous NC couplings or including additional operators to fine-tune these effects away. Although these relations basically have no effect on a purely CC single top study, one should bear them in mind when addressing anomalous CC couplings (the <span style="font-variant:small-caps;">Whizard</span> implementation contains the option to automatically enforce these relations). LHC phenomenology {#pheno} ================= Apart from indirect searches using low energy observables, e.g., in flavor physics [@Grzadkowski:2008mf; @Drobnak:2011aa; @Drobnak:2011wj] or SM precision observables [@Zhang2012], there are basically two different classes of direct observables for top quark properties at the current collider experiments, namely those related to top production or top decays. While it is clear that only a combination of all available observables will deliver the best bounds on anomalous contributions, it is crucial to understand each analysis separately before the combination step. Therefore, we will focus here on the discussion of single top production cross sections, citing and using results from top decay studies to derive estimates for the most stringent bounds on the full anomalous parameter space at the end of the article. Single tops are produced at the LHC (and Tevatron) in three different channels, namely $s$ channel $tb$ production, $t$ channel $tj$ production (where $j$ denotes a light hadronic jet), and associated $tW$ production, cf. Fig. \[diags\]. While experimentalists are struggling to identify and discriminate these channels at the detector level with suitable selection criteria, the theoretical question is how the corresponding measured cross sections $\sigma^\text{det}_i$ (for final states ) are represented as functions on the anomalous parameter space, how the measurement can be converted into bounds on the parameters. In this respect, a first step may be to separate the detector response from the hard production cross section: $$\label{xsec} \sigma^\text{det}_i(\vec{g}) = \sum_j \varepsilon_{ij} \cdot \sigma^\text{part}_j(\vec{g})~,$$ summing over partonic production processes $j$ as functions of the parameter point $\vec{g}$. $\varepsilon_{ij}$ denotes the detector efficiency matrix mapping the process $j$ onto the final state selection $i$, which can be retrieved with a detector simulation. Once the functions $\sigma^\text{part}_i(\vec{g})$ are known, experimentally measured confidence intervals for $\sigma^\text{det}_i$ can be mapped onto confidence intervals for $\vec{g}$ by formal inversion of Eq. . However, the remaining question to be addressed in this approach is: Where did we put the detector acceptance $\Phi$, into $\varepsilon$ or into $\sigma^\text{part}_i$? The significance of this question is obvious, since anomalous couplings might very well affect the differential distributions, thus making $\Phi$ a function of $\vec{g}$. Therefore, the answer to that question influences the strategy as well as the efforts necessary to compute the function $\sigma^\text{part}_i(\vec{g})$, and potentially also the bounds derived from it, as we will show in the following section. Technical setup {#on-off} --------------- ### Kinematics in the on-shell limit {#on} The simplest approach is to neglect the $\vec{g}$-dependence of the acceptance entirely and pull it into $\varepsilon$, implying that the $\sigma^\text{part}_i(\vec{g})$ in Eq.  represent the set of total partonic cross sections integrated over the full phase space. Further neglecting finite width and interference effects with irreducible backgrounds enables one to decompose $\sigma^\text{part}_i(\vec{g})$ as $$\label{onshell} \sigma^\text{part}_i(\vec{g}) = \sigma^\text{prod}_i(\vec{g}) \times \prod \mathcal{BR}~,$$ where $\sigma^\text{prod}_i(\vec{g})$ denotes the full on-shell single top production cross sections, and the product of branching ratios accounts for the decays of the heavy particles, namely $t$ and one or two $W$s, depending on the production channel. Since all the on-shell production diagrams can contain only one anomalous $tbW$ vertex insertion, it is argued in [@AguilarSaavedra:2008gt] that $\sigma^\text{prod}_i(\vec{g})$ may be written as a polynomial up to second order in $\vec{g}$: $$\label{poly} \sigma^\text{prod}_i(\vec{g}) = \sigma^\text{SM}_i \sum_{k,l} \kappa^i_{\,kl}\, g_k\, g_l~,$$ where the $\sigma^\text{SM}_i$ are the total SM cross sections, and the $\kappa^i_{\,kl}$ denote the integrated products of diagrams with one insertion of $g_k$ and $g_l$ each, normalized to the SM point in each production channel $i$. Plugging  and  into , one arrives at the ansatz employed in [@AguilarSaavedra:2008gt]: $$\begin{aligned} \label{xsecAS} \sigma^\text{det}_i(\vec{g}) &= \sum_{j,k,l} \left[\varepsilon\times\prod\mathcal{BR}\right]_{ij} \cdot \sigma^\text{SM}_j\cdot\kappa^j_{\,kl}\, g_k\, g_l\nonumber \\ &\equiv \sum_{j,k,l} \varepsilon_{ij} \cdot \sigma^\text{SM}_j\cdot \kappa^j_{\,kl}\, g_k\, g_l~.\end{aligned}$$ For brevity, this will be referred to as *on-shell approach* from here on. The advantage of the formula is obvious: once $\varepsilon$ and the constant $\kappa$s are known, the conversion of measured results into bounds on $\vec{g}$ becomes very simple and efficient. However, the validity of the assumptions leading to this result shall be addressed now. ### Full matrix elements and acceptances including anomalous couplings {#off} Eq.  tells us that one should be able to vary the coupling point $\vec{g}$ within the ranges relevant for the study, with only minor effects on the detector response $\varepsilon$ in the phase space window which corresponds to a given final state selection. However, Eq.  implies even more, namely that retrieving the matrix element response as a function of $\vec{g}$ and applying acceptance cuts on the phase space should approximately commute, or equivalently, Eq.  should give the same results as e.g., $$\label{xsecW2} \sigma^\text{det}_i(\vec{g}) = \sum_j \varepsilon_{ij}^\prime \cdot \left[ \Phi^\text{part} \times \sigma^\text{part}\right]_j(\vec{g})~,$$ where the basic detector acceptance cuts such as $p_T$ and $\eta$ cuts on the partons and leptons, represented by $\Phi^\text{part}$, are applied directly to the phase space integration and hence formally included in the $\vec{g}$-dependent part of the formula, while the matrix $\varepsilon^\prime$, assumed to be constant in $\vec{g}$, denotes the efficiency of mapping the partonic final states *at the acceptance level* $\Phi^\text{part}$ onto the final state selections at detector level. To be more explicit, the idea is to accommodate as much of the acceptance cuts as possible within the $\vec{g}$-dependent part without becoming exclusive to any of the different final state selections, which are still contained in the $\vec{g}$-independent $\varepsilon^\prime$. This obviously implies that the phase space window covered by $\varepsilon^\prime$ must be fully contained within the acceptance window $\Phi^\text{part}$, leading to the notion that partonic acceptance and final state selection cuts should be adapted to each other as closely as possible. Assuming leptonic $t$ decay, we therefore apply the following acceptance cuts on the partonic phase space integration: \[acc\] $$\begin{aligned} \Phi^\text{part}:\qquad\qquad p_T\left(\ell,\nu\right) > \unit[25]{GeV}\quad & \text{and}\quad \left| \eta\left(\ell\right) \right| < 3\,, \label{phi_l} \\ p_T\left(j,b\right) > \unit[30]{GeV}\quad & \text{and}\quad \left| \eta\left(j,b\right) \right| < 5\,, \label{phi_j} \\ \unit[150]{GeV} < m_{b\ell\nu} & < \unit[225]{GeV}\,, \label{phi_mt}\end{aligned}$$ where Eq.  is required for only one of the two $b$s in the $tbj$ process to be inclusive, and all the cuts are in correspondence to the detector-level selection criteria stated below. Associated $tW$ production is entirely omitted for the time being, because modelling this process within its detector acceptance window while at the same time remaining inclusive with respect to the other processes is highly nontrivial, and only marginally affects our following statements (the main effect being the neglected contamination of the other final states at the detector level, which amounts to $\lesssim$ in the $tb$ channel and practically vanishes in the dominant $tj$ channel). Clearly, this is still not the fully correct answer at detector level, but it should be closer to the truth than entirely neglecting the $\vec{g}$-dependence of $\varepsilon$, and the consistency of the two approaches can be checked. Although it is clear that the object $\left[ \Phi^\text{part} \times \sigma^\text{part}\right]_i(\vec{g})$ to be computed is much harder to handle than the constant $\kappa$s (even more so if the full matrix element response including all off-shell and interference effects is to be taken into account), it is basically just a technical issue which can be tackled with appropriate Monte Carlo machinery and respective CPU time. For brevity, we will refer to this approach as *full matrix element (ME) approach* from now on. In the following section, we compare the results of Eq.  and Eq. , and see if the effort is justified. Comparison of the results in the on-shell limit with the full results {#comp} --------------------------------------------------------------------- For the measurement of the total cross section of a given final state at the detector level, the experimental sensitivity is given in terms of a measure for $\Delta\sigma/\sigma$, where estimations for total uncertainties are adopted from [@AguilarSaavedra:2008gt] for consistency, amounting to $\unit[20.8]{\%}$ ($tb$ sel.) and $\unit[13.5]{\%}$ ($tj$ sel.) for $\unit[10]{fb^{-1}}$ of LHC data at $\sqrt{s}=\unit[14]{TeV}$. Therefore, we separate the overall normalization, which is basically given by higher-order SM results for the total production cross sections, from the modelling of the normalized LO matrix element response as a function of the anomalous coupling set $\vec{g}$, i.e. the $\kappa_{\text{on}}$ coefficients in the on-shell approach or, more generally, a function $\Delta\sigma/\sigma(\vec{g})\equiv\kappa(\vec{g})$ for each partonic input process $i$, where \[k\_g\] $$\begin{aligned} \text{on-shell:}&&\kappa^i_\text{on}\left(\vec{g}\right) & = \sum_{k,l} \kappa^i_{\,kl}\, g_k\, g_l \label{k_on} \,, \\ \text{full ME:}&&\kappa^i_\text{full}\left(\vec{g}\right) & = \frac{\left[ \Phi^\text{part} \times \sigma^\text{part}\right]_i(\vec{g})}{\left[ \Phi^\text{part} \times \sigma^\text{part}\right]_i|_\text{SM}} \,, \label{k_off} \end{aligned}$$ cf. Eqs.  and . Since the aim is to accommodate *all* coupling dependence therein, it is fruitful to first set $\varepsilon\equiv 1$ and compare this function for the different approaches at the partonic level. ### Partonic level To be self-consistent, we essentially redo the analysis procedure presented in [@AguilarSaavedra:2008gt] for the on-shell approach, employing <span style="font-variant:small-caps;">Whizard</span> to compute the coefficients of $\kappa_\text{on}$ as well as produce parton-level samples, which are then processed with <span style="font-variant:small-caps;">Pythia</span> and <span style="font-variant:small-caps;">Delphes</span> to retrieve $\varepsilon$. The quadratic coefficients (i. e. $\sim g_i^2$) are obtained in each production channel, cf. Fig. \[diags\], by separately setting each $g_i=1$, integrating the total cross section for on-shell single top production and finally normalizing to the SM point ($V_L=1$, $V_R=g_{L,R}=0$). The interference terms are computed similarly, setting always two couplings to 1 and subtracting the quadratic parts from the result before normalizing to the SM. The implementation of the vertices and phase space integration has been checked by switching off the pdfs and comparing <span style="font-variant:small-caps;">Whizard</span> to analytical results. Using the pdfs and parameter setup quoted in [@AguilarSaavedra:2008gt], <span style="font-variant:small-caps;">Whizard</span> also reproduces the on-shell $\kappa$ coefficients stated there within numerical uncertainties. For all further <span style="font-variant:small-caps;">Whizard</span> results, we set $m_t=\unit[173.1]{GeV}$, $m_b=\unit[4.2]{GeV}$, $m_W=\unit[80.42]{GeV}$ and choose CTEQ6L1 [@Pumplin:2002vw] for pdfs. In the full ME approach, the matrix element response is modelled according to Eq.  by applying the acceptance cuts $\Phi^\text{part}$, Eq. , to the full partonic off-shell matrix elements. In this approach, taking all finite width and interference effects into account, it is *a priori* not clear that the function $\kappa_\text{full}\left(\vec{g}\right)$ obeys a simple polynomial expansion in $\vec{g}$, so rather than assuming a specific functional form, we use the <span style="font-variant:small-caps;">Whizard</span> machinery to scan the entire parameter space $\vec{g}=\left(V_L,V_R,g_L,g_R\right)$ (effects of $V_L^\text{off}$ will be addressed separately in Sec. \[VO\]) within the relevant numerical ranges $0<V_L<1.2$, $-1.2<V_R<1.2$ and $-0.6<g_{L,R}<0.6$ (cf. [@AguilarSaavedra:2008gt]), also including the full dependence of the top width on the couplings $\Gamma_t(\vec{g})$. Since $\Gamma_t$ has already been measured, the most recent bound from D$\slashed 0$ being $\Gamma_t=\unit[2.00^{+0.47}_{-0.43}]{GeV}$ [@Abazov:2012vd], it is included in our analysis as an additional observable[^1]. The numerical results can then be used to test the validity of the polynomial parametrisation, Eq. , in the following way: The normalized matrix element response $\kappa$ may always be expanded as $$\kappa\left(\vec{g}\right) = \sum_i \kappa_1\left(g_i\right) + \sum_{i,j} \kappa_2\left(g_i,g_j\right) + \sum_{i,j,k}\kappa_3\left(g_i,g_j,g_k\right) + \text{...}\,,$$ where the $\kappa_i$ are polynomials in their respective arguments. Offsets $\kappa_0$ from squared irreducible background diagrams could be considered, but are obviously independent of $\vec{g}$ and merely add to the background normalization, so we just subtract them from the scans, while keeping all interference effects (appearing as terms linear in the $g_i$ in $\kappa_1$) for completeness. Usually this series terminates after $\kappa_2$, which becomes obvious when applying the narrow width approximation, where additional coupling effects cancel in the interplay of the decay vertex insertion, width dependence and phase space integration. This basically leads to the quadratic form in the on-shell approach. However, in the special case of single tops, production and decay are interrelated via the same set of CC couplings, thus affecting production as well as decay distributions, which in combination with the detector acceptance cuts might lead to deviations from the on-shell result in some regions of the parameter space. In order to estimate the size of the moments $\kappa_3$, in our cross section scans we consider subplanes $\left(g_i,g_j\right)$ among the anomalous couplings $(V_R,g_L,g_R)$ for different fixed values of $V_L$. After subtracting all the lowest moments, $$\Delta\kappa(\vec{g}) = \kappa(\vec{g}) - \sum_{k=i,j,V_L} \kappa_1(g_k) - \sum_{k=i,j} \kappa_2(g_k,V_L) \,,$$ within the on-shell picture the remaining contribution $\Delta\kappa(\vec{g})\sim\kappa_2(g_i,g_j)$ should then be independent of the value of $V_L$, otherwise it would by definition contain some $\kappa_3(g_i,g_j,V_L)$. For illustration, we choose the coupling subspace $(g_i,g_j)=(V_R,g_L)$, because it is one of the dominant interference contributions to all production processes, and evaluate $\Delta\kappa|_{V_L=1.2}-\Delta\kappa|_{V_L=0.2}$. The resulting picture is mixed: while in the $s$ channel the result is practically 0 all over the $V_R$–$g_L$ plane, in the $t$ channel process $\bar{t}j$, which plays a central role because of its comparably large cross section, this difference amounts to $\sim -0.2$ at $V_R\simeq\pm1$ and $g_L\simeq\pm0.5$ along the interference direction $V_R\sim 2g_L$ (cf. Fig. \[dk\_VRgL\]), which is of the same size as the respective on-shell contribution $\sim-0.5\times V_R\times g_L$, (the same is true for the $V_Lg_R$ interference in the $tj$ channel). For comparison, we repeat the whole procedure selecting only the resonant single top diagrams for the scan (including the full top width dependence on $\vec{g}$), finding that background interference only plays a minor role. ![image](tjc_fullME_kappa_VRgL_VL_0p2_1p2.pdf) ![image](tbjc_fullME_kappa_VRgL_VL_0p2_1p2.pdf) ![image](tjc_resonant_kappa_VRgL_VL_0p2_1p2.pdf) ![image](tbjc_resonant_kappa_VRgL_VL_0p2_1p2.pdf) Furthermore, note that there are other interference directions also showing substantial effects, e.g., in the $g_L$–$g_R$ plane of the $s$ channel, but the $V_R g_L$ interference is the most interesting one because it is large in all channels, and respective bounds are expected to remain rather weak also from other experiments along the considered direction $V_R\sim 2g_L$ [@AguilarSaavedra:2008gt]. All in all, these numbers indicate strongly that the interference and off-shell effects contained in the full matrix elements might become important at the level of the coupling limits to be expected experimentally, at least in a stand-alone single top cross section study at the LHC, and should be checked in any case. ![image](tjc_kappa_VR_2gL_VL_0p6_1p0_1p4.pdf) ![image](tbjc_kappa_VR_2gL_VL_0p6_1p0_1p4.pdf) In order to further quantify this effect and exclude potential artefacts from unphysical regions in the parameter space, we now systematically scan the $V_R g_L$ interference along the direction $V_R\sim 2g_L$ as a function of $V_L$, including resonant off-shell diagrams for anomalous single top production and decay. We observe that the scans do show a quadratic dependence on $V_R$ and $g_L$ to a very good approximation at any value of $V_L$ (cf. Fig. \[k\_VR\_2gL\]), so the respective coefficients $\kappa_{V_R}\sim V_R^2$, $\kappa_{g_L}\sim g_L^2$ and $\kappa_{V_R g_L}\sim V_R\,g_L$ can be extracted from quadratic fits along the axes $V_R$, $g_L$ and $V_R=2g_L$, for each fixed value of $V_L$ and $g_R=0$. This is done for the full phase space integration as well as for the acceptance region defined in Eq. . As illustrated in Fig. \[k\_tjc\] for the $t$ channel processes, when integrating over the full phase space the deviations from the on-shell result are very small as expected from the narrow width approximation, whereas application of the acceptance cuts leads to substantially different fit coefficients which also depend on the value of $V_L$. ![image](tjc_kappa_VR_vs_VL_cut_none_all.pdf) ![image](tbjc_kappa_VR_vs_VL_cut_none_all.pdf) ![image](tjc_kappa_gL_vs_VL_cut_none_all.pdf) ![image](tbjc_kappa_gL_vs_VL_cut_none_all.pdf) ![image](tjc_kappa_VRgL_vs_VL_cut_none_all.pdf) ![image](tbjc_kappa_VRgL_vs_VL_cut_none_all.pdf) ![image](tjc_compare_kappas_24_VLVR_heat.pdf) ![image](tjc_compare_kappas_34_VLVR_heat.pdf) ![image](tjc_compare_kappas_24_VRgL_heat.pdf) ![image](tjc_compare_kappas_34_VRgL_heat.pdf) ![image](tjc_compare_kappas_24_gLgR_heat.pdf) ![image](tjc_compare_kappas_34_gLgR_heat.pdf) ![image](tbjc_compare_kappas_24_VLVR_heat.pdf) ![image](tbjc_compare_kappas_34_VLVR_heat.pdf) ![image](tbjc_compare_kappas_24_VRgL_heat.pdf) ![image](tbjc_compare_kappas_34_VRgL_heat.pdf) ![image](tbjc_compare_kappas_24_gLgR_heat.pdf) ![image](tbjc_compare_kappas_34_gLgR_heat.pdf) Particularly this dependence of the magnitude of the $V_R g_L$ interference on $V_L$ is an example for the breakdown of the quadratic on-shell approach. However, this dependence is rather small in $\kappa_{V_R}$ and $\kappa_{g_L}$ which come with squares of the respective couplings and therefore dominate the sensitivity of a given process to these couplings, so it might still be a good approach to expand the full $\kappa_\text{full}$ as a quadratic form in *small* anomalous deviations from the SM point $V_L=1$, namely by choosing $V_L=1$ as an origin. As explained before, the numerical values are extracted from quadratic fits to coupling scans (including acceptance cuts) in all directions of $\vec{g}$, and all interference directions $g_i g_j$. A major difference to the original on-shell quadratic form is the appearance of large linear terms in the couplings, which now encode the interference with the SM. In Figs. \[tjc\_on\_off\] and \[tbjc\_on\_off\], we compare the different quadratic parametrisations, namely $\kappa_\text{on}$ from the on-shell approach and $\kappa_\text{fit}$ inferred from the fits, to the full ME response $\kappa_\text{full}$, illustrating two statements: Firstly, $\kappa_\text{on}$ significantly deviates from $\kappa_\text{full}$ in various parts of the parameter space relevant to the analysis, particularly in the $g_L$–$g_R$ and $V_R$–$g_L$ planes. Secondly, although still neglecting the higher coupling dependences, $\kappa_\text{fit}$ inferred from quadratic fits to the full scan does show a significantly improved agreement with the full scan while still being fast and efficient. This is further illustrated in Fig. \[tj\_1sigma\_part\] and \[tbj\_1sigma\_part\] showing $\pm1\sigma$ sensitivity contours around the SM strength $\kappa=1$ for various anomalous coupling combinations and production processes: Especially when the momentum-dependent couplings $g_{L,R}$ are involved, there are remarkable shifts of the contours when going from the on-shell to the full ME approach, but generally these effects are modelled very well by the adapted quadratic parametrisation $\kappa_\text{fit}(\vec{g})$, while the machine cost reduces from a 4D scan over $\vec{g}$ to a set of 1D scans along all axes $g_i$ and interference directions $g_i g_j$ for each input process. We now go on to the detector level to quantify the impact of these different approaches on exclusion bounds on anomalous couplings from combined cross section measurements. ![image](tj_compare_1sigma_contours_k234_VLVR.pdf) ![image](tjc_compare_1sigma_contours_k234_VLgL.pdf) ![image](tj_compare_1sigma_contours_k234_VLgR.pdf) ![image](tjc_compare_1sigma_contours_k234_VRgL.pdf) ![image](tj_compare_1sigma_contours_k234_gLgR.pdf) ![image](tjc_compare_1sigma_contours_k234_gLgR.pdf) ![image](tbj_compare_1sigma_contours_k234_VLVR.pdf) ![image](tbjc_compare_1sigma_contours_k234_VLgL.pdf) ![image](tbj_compare_1sigma_contours_k234_VLgR.pdf) ![image](tbjc_compare_1sigma_contours_k234_VRgL.pdf) ![image](tbj_compare_1sigma_contours_k234_gLgR.pdf) ![image](tbjc_compare_1sigma_contours_k234_gLgR.pdf) ### Detector level In order to obtain a detector efficiency matrix in the various approaches, samples of $\unit[500]{k}$ events are produced in each partonic production process, once integrating the full off-shell matrix elements over the acceptance region, Eq. , and once integrating the resonant matrix elements over the full phase space, letting the tops decay off-shell into a $b$ quark, a charged lepton and a neutrino, analogously to [@AguilarSaavedra:2008gt]. All the parton-level samples are processed with <span style="font-variant:small-caps;">Pythia</span> and <span style="font-variant:small-caps;">Delphes</span> to obtain events at detector level. This is done for the SM point as well as the coupling configurations $$\label{cpl_points} \begin{array}{rlll} \text{A:}\hspace{30pt} & V_L=1\,,\; & V_R=0.3\,,\; & g_L=0.15\,, \\ \text{B:}\hspace{30pt} & V_L=1\,,\; & g_R=0.024\,, & \end{array}$$ taken from [@AguilarSaavedra:2008gt] to facilitate comparison. Note that in each case a consistent top width $\Gamma_t\left(\vec{g}\right)$ is calculated beforehand and the result checked to comply with experimental constraints [@Abazov:2012vd]). Again for comparison reasons, we also adopt the final state selection cuts stated in [@AguilarSaavedra:2008gt] which define the various components of $\varepsilon$: apart from requiring an isolated lepton (that is, $e$ or $\mu$) with $p_T>\unit[25]{GeV}$ and missing transverse energy $\slashed E_T>\unit[25]{GeV}$, the selection criteria for the three final state signatures are, respectively 1. for $tb$ selection: exactly two $b$ tagged jets (assuming a tagging efficiency of 0.6) with $p_T>\unit[30]{GeV}$, and neither central nor forward light jets with $p_T>\unit[15]{GeV}$. In addition, the top momentum $p_t$ is reconstructed from one of the $b$s together with the charged lepton and $\slashed E_T$ (to be identified with the neutrino $p_T$), by applying the on-shell constraint $\left( p_\ell+p_\nu \right)^2=m_W^2$ and picking the smaller of the two solutions for the longitudinal component of $p_\nu$. Finally, the resulting top mass must lie between 150 and $\unit[225]{GeV}$. 2. for $tj$ selection: at least one $b$ jet with $p_T>\unit[30]{GeV}$ (one of them reconstructing $p_t$ together with the leptons as explained above), one light forward jet with $p_T>\unit[50]{GeV}$ and $2.5<\left|\eta\right|<5$ and no more than one additional light central jet, which may have $p_T<\unit[30]{GeV}$ only. By applying every final state selection to each of the $\unit[500]{k}$ event samples corresponding to the partonic input processes and averaging over lepton flavors and charge states, we find for the samples from resonant diagrams integrated over the full phase space an efficiency matrix $\varepsilon$ (in $\%$) at the SM point: $$\label{eps_full_ps} \begin{array}{r|ccc} & tb & tj & tbj \\ \hline tb\,\text{sel.} & 0.658(6) & 0.040(1) & 0.051(2) \\ tj\,\text{sel.} & 0.165(3) & 0.647(6) & 0.531(5) \,. \end{array}$$ For the full ME approach, we run the selection criteria described above on the detector level samples from full matrix elements integrated over the acceptance region, inferring $\varepsilon^\prime$ (in $\%$) $$\label{eps_acc} \begin{array}{r|ccc} & tb & tj & tbj \\ \hline tb\,\text{sel.} & 1.28(1) & 0.039(1) & 0.031(1) \\ tj\,\text{sel.} & 0.282(4) & 1.52(1) & 1.023(7) \, \end{array}$$ at the SM point. Before moving on to detector-level coupling limits, the detector efficiency matrices can be compared between the various coupling points, Eq. , to test the assumption of a constant detector response. While the small admixture of $g_R$ in point B only has a negligible effect on the efficiencies, we observe that there are sizeable changes in the efficiencies when going from the SM point to point A, amounting to some $\unit[15]{\%}$ ($\unit[8]{\%}$) of the original values for $tb$ ($tj$) selection in the samples with full phase space integration at parton level. However, when going close to the detector acceptance region already with the partonic input, this dependence is reduced to $\sim\unit[6]{\%}$ ($\unit[4]{\%}$), thus improving on another source of systematic uncertainty. Taking approximate NNLO on-shell $s$ and $t$ channel production cross sections from the literature [@Kidonakis:2010tc; @Kidonakis:2011wy] (multiplied by a partonic acceptance efficiency corresponding to the cuts in Eq.  in the full ME approach) to normalize the SM reference point for each input process, we now have all ingredients at hand to derive limits on $\vec{g}$ from a set of cross section measurements, modelled by Eqs.  and , and compare the results. In the $t$ channel, the matrix element response for the detector-level analysis is modelled using only $t\bar{b}j$ and $\bar{t}bj$ processes for simplicity, and because it was argued that the corresponding distributions already describe the proper NLO behaviour rather well [@Sullivan:2004ie; @Boos:2006af]. Moreover, it was shown [@Falgari:2011qa; @Campbell:2012uf] that NLO corrections affect the differential distributions in $s$ and $t$ channel single top production only marginally, at the few $\%$ level, and can thus be readily accounted for by channel specific overall $K$ factors, as in our analysis. A more comprehensive analysis including coupling dependent $K$ factors, similar to the case of anomalous flavor changing gluon couplings [@Gao:2011fx] and extending an existing study on anomalous top decays at NLO [@Drobnak:2010ej], will moderately influence the numerical values of the exclusion bounds. However, our results concerning the relative importance of the quartic couplings and the need to include the full matrix elements remain valid. Now turning to the resulting bounds at the detector level, as already anticipated from the $1\sigma$ contours in Fig. \[tj\_1sigma\_part\] and \[tbj\_1sigma\_part\] the effects on $V_{L,R}$ remain small in general, while the largest differences are found when the momentum dependent couplings $g_{L,R}$ are involved, particularly in the $g_L$–$g_R$ plane illustrated in Fig. \[1sigma\_det\]. In this case, when single channels and charge states are considered separately, the different approaches tend to produce very different exclusion bounds. Fig. \[1sigma\_det\] might also suggest that after combining all channels and considering the ratio $R(\bar{t}/t)$ of cross sections for $t$ and $\bar{t}$ production in the $t$ channel as an additional observable (tentatively assuming $\unit[2]{\%}$ statistical and $\unit[3]{\%}$ systematic uncertainty as estimated for $\unit[10]{fb^{-1}}$ at $\sqrt s=\unit[14]{TeV}$ in [@AguilarSaavedra:2008gt], in the absence of a more detailed experimental assessment), one might end up with the same exclusion limits again, but indeed this depends heavily on the total uncertainty of $R$ in the actual experiment (cf. “$R_2$” in Fig. \[1sigma\_det\]). In any case, when the aim is to properly understand and quantify the sensitivities to anomalous couplings of the various final states separately, going from the on-shell approach to full matrix element responses inside the selection acceptance region produces considerable effects that should not be neglected. In that respect, the adapted quadratic parametrisation introduced above, employing quadratic fits to off-shell scans inside the acceptance region, represents a very good approximation to the full off-shell parameter scan (cf. Fig. \[tjc\_on\_off\]–\[tbj\_1sigma\_part\]). ![image](chi2_cl1_k24_n3_gL_gR_tb_effFBbd_2c.pdf) ![image](chi2_cl1_k24_n3_gL_gR_tbj_effFBbd_2c.pdf) ![image](chi2_cl1_k24_n3_gL_gR_tb_tbj_effFBbd_2c.pdf) ![image](chi2_cl1_k24_n3_gL_gR_tbj_effFBbd_R100_2c.pdf) ![image](chi2_cl1_k24_n3_gL_gR_tb_tbj_effFBbd_R100_2c.pdf) ![image](chi2_cl1_k24_n3_gL_gR_tb_tbj_effFBbd_R200_2c.pdf) Pinning down the off-shell coupling {#VO} ----------------------------------- After discussing the technical issue of modelling the LO matrix element response to anomalous top couplings at an experimentally relevant acceptance level, and validating an adapted quadratic parametrisation which simultaneously meets the demands of machine efficiency and good agreement with the full off-shell coupling scan in the previous section, the closing section of the article is devoted to the application of the new approach to a physical issue, namely a possible admixture of the additional anomalous coupling $V_L^\text{off}$ introduced in Sec. \[anom\_top\], to the single top cross sections (the total top width is also included as an observable, but its sensitivity to $V_L^\text{off}$ is kinematically suppressed compared to the other anomalous couplings, since the relevant scale $m_t$ is lower than $\sqrt{\hat s}$). Considering the experimental sensitivities to the anomalous couplings of the total cross sections stated above for the LHC (which are already dominated by systematics), it is clear that a stand-alone study of single top cross sections alone will never provide the most stringent bounds on the complete parameter space of anomalous CC couplings, including $V_L^\text{off}$ or not. Therefore, rather than just adding another direction to $\vec{g}$, the focus shall be directed here to those regions of the parameter space where single top cross sections actually become the crucial inputs to the combined limits. More explicitly, the top decay observables (mostly related to the charged lepton distribution) are very sensitive to anomalous $W$ helicity fractions generated by $V_R$, $g_L$ and $g_R$ (cf. [@AguilarSaavedra:2006fy; @AguilarSaavedra:2007rs; @AguilarSaavedra:2008gt; @Aad2012]). For example, the limit $\left|g_R\right|\lesssim 0.024$ stated in [@AguilarSaavedra:2008gt] for our LHC reference point ($\unit[10]{fb^{-1}}$ at $\unit[14]{TeV}$) is more than an order of magnitude below the sensitivity of the cross sections, so we may as well set $g_R\equiv0$ for our purposes. On the other hand, the large interference among $V_R$ and $g_L$ leads to rather poor bounds $\left|V_R\right|\lesssim 0.3$ resp. $\left|g_L\right|\lesssim 0.15$ as long as they are fine-tuned to $V_R\sim 2g_L$. Finally, since decay observables basically measure helicity fractions, they are neither sensitive to the overall vertex normalization nor to the admixture of $V_L^\text{off}$ to the left-handed vector part. This is where the cross sections come into play, delivering the most stringent direct constraints. In Fig. \[VL\_VO\], we therefore present combined limits on $V_L$ and $V_L^\text{off}$ from single top cross sections, both setting $V_R=g_L=0$ as well as varying over $-0.3\leq V_R=2g_L\leq 0.3$. The very different sensitivities of the two final states greatly help in the combined limit: the $s$ channel is very sensitive along $V_L^\text{off}$ due to the kinematics, whereas the $t$ channel basically cuts the substantial interference in the $s$ channel along $V_L$. Still, the resulting limit on $V_L$ deteriorates from $0.9<V_L<1.1$ ($V_L^\text{off}=0$) to $0.82<V_L<1.1$ ($V_L^\text{off}$ varied). Naturally, projecting over the remaining freedom in $V_R$ and $g_L$ instead of switching them off further relaxes the combined limits to $0.68<V_L<1.1$. Fig. \[VR\_gL\] displays combined bounds in the $V_R$–$g_L$ plane, switching off resp. varying over $V_L^\text{off}$. ![image](chi2_cl1_k3_n3_VL_VO_tb_vs_tbj_effFBbd_2c.pdf) ![image](chi2_cl1_k3_n3_VL_VO_tb_tbj_effFBbd_opt_R100_2c.pdf) ![image](chi2_cl1_k3_n3_VL_VO_tb_tbj_effFBbd_R100_2c_vs_3c.pdf) ![image](chi2_cl1_k3_n3_VR_gL_tb_vs_tbj_effFBbd_2c.pdf) ![image](chi2_cl1_k3_n3_VR_gL_tb_tbj_effFBbd_opt_R100_2c.pdf) ![image](chi2_cl1_k3_n3_VR_gL_tb_tbj_effFBbd_R100_2c_vs_3c.pdf) In the long run, it is perfectly clear that this ambiguity among $V_L$ and $V_L^\text{off}$ remaining in the single top cross sections can be further resolved, namely by examining differential distributions, since $V_L^\text{off}$ scales very differently with the partonic $\sqrt{s}$ than $V_L$. (In fact, $V_L^\text{off}$ behaves like, or parametrises, a heavy off-resonant new degree of freedom, cf. e.g., [@Boos:2006xe].) However, this issue will have to be tackled in the $s$ channel where the momentum of the $W$ propagator producing the top becomes timelike. Sensitive observables would obviously be the total invariant mass $m_{tb}$ of the final state or the pseudorapidity $\eta_b$ of the hard $b$ jet produced along with the single top. However, such a study is experimentally challenging, since it requires a very good isolation of the tiny $s$ channel signal from the huge $t$ channel contamination, whereas at present this signal has not even been established yet individually at the LHC (the most recent search being [@ATLAS-CONF-2011-118]). Hence, one should stay careful when stating limits on $V_L$ from measurements of the overall size of $V_{tb}$ until its kinematic behaviour is further clarified experimentally. Conclusions {#sum} =========== In this article, we have revisited the model-independent parametrisation of anomalous top couplings to SM gauge bosons within the effective operator approach, paying special attention to the charged-current sector and its phenomenological implications at current hadron colliders. More explicitly, addressing the minimal fully general set of anomalous trilinear $tbW$ couplings coming from dimension six effective operators, there is a controversy regarding the meaning of “fully general”, namely whether an off-shell interaction contained in the original operator basis should be dropped because it turns out to be related to four-fermion contact interactions after application of the equations of motion, or not. While dropping it and sticking to the usual coupling basis $(V_L,V_R,g_L,g_R)$ naturally simplifies the analysis, there are good arguments to include it: Since it emerges from the minimal gauge invariant operator basis that also generates the trilinear couplings, the respective coupling size $V_L^\text{off}$ is related to the other ones by the underlying operator coefficients. (For example, an experimental limit on $\delta V_L$ is ambiguous in the context of effective operator coefficients, requiring knowledge about either $\delta V_L^\text{off}$ or the NC sector to be resolved.) For the same reason, the coupling basis $(V_L,V_R,g_L,g_R,V_L^\text{off})$ parametrises the complete set of anomalous diagrams which interfere with the SM diagram in a minimal way, so including it is also consistent at the level of matrix elements. In the phenomenological part, the dependence of single top cross sections on anomalous $tbW$ couplings in $s$ and $t$ channel production is examined, stressing the fact that the couplings do not only affect the total cross sections but also final state distributions, which determine the selection efficiencies within the detector acceptance region. While these effects are usually considered small, working with constant detector efficiencies and modelling the whole coupling dependence on the basis of on-shell production amplitudes, we use the <span style="font-variant:small-caps;">Whizard</span> machinery to scan the full off-shell matrix element dependence on the couplings inside the acceptance window defined by the final state selection cuts. Comparing to the on-shell approach, one finds considerable deviations in some regions of the parameter space, especially where the momentum-dependent couplings $g_{L,R}$ are involved, affecting the sensitivities of the various production channels to those couplings and therefore also the limits derived from the experiment. Finally, an adapted polynomial approach of the coupling dependence is discussed, which is based on quadratic fits to the full off-shell matrix element response including detector acceptance, and turns out to parametrise the full scan result rather well while still being fast and efficient. However, it is also stated as a result of the present study that the theoretical modelling of the coupling dependence should be adapted as closely as possible to a given experimental analysis with defined selection criteria to minimize the systematic uncertainty of the derived limits. The study concludes with a short discussion of the influence of top decay observables on combined coupling limits, and the regions of the parameter space where single top cross sections still provide the crucial input to the bounds, namely the overall $tbW$ vertex normalization and the interference direction $V_R\sim2g_L$. In this respect, we address the impact of including $V_L^\text{off}$ in the coupling basis, and briefly point out the possibilities to resolve the ambiguity between $V_L$ and $V_L^\text{off}$ experimentally, using kinematic distributions in the $s$ channel. FB thanks J. A. Aguilar-Saavedra for useful discussions. FB is supported by Deutsche Forschungsgemeinschaft through the Research Training Group GRK1147 *Theoretical Astrophysics and Particle Physics*. <span style="font-variant:small-caps;">Whizard</span> development is supported in part by the Helmholtz Alliance *Physics at the Terascale*. Parts of this work are supported by the German Ministry of Education and Research (BMBF) under contract no. 05H09WWE. [^1]: Clearly the experimental analysis performed in [@Abazov:2012vd] to extract $\Gamma_t$ from data will itself also be affected by $\vec g$ dependent acceptances as discussed in the course of this paper. However, since we do not aim at a reassessment of their analysis in this respect, yet also want to exclude regions in parameter space which are completely unphysical with respect to $\Gamma_t$, we still include the observable, inferring the $\Gamma_t(\vec g)$ dependence over the full phase space. This approach is conservative, because our results indicate that including the full acceptance dependence generally tends to improve the sensitivities, cf. e.g., Fig. \[k\_tjc\]. *A posteriori*, we find that, due to the still large error bars, the current measurement of $\Gamma_t$ constrains $\vec{g}$ much less than the cross sections. The limits on $\vec{g}$ would therefore not be affected substantially by such a reassessment.
{ "pile_set_name": "ArXiv" }
--- author: - 'Yan Ge, Haiping Lu, and Pan Peng [^1]' bibliography: - 'references.bib' title: 'Mixed-Order Spectral Clustering for Networks' --- [Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{} [^1]: Yan Ge, Haiping Lu and Pan Peng are with the Department of Computer Science, University of Sheffield, U.K. (e-mail: [email protected], [email protected] and [email protected]).
{ "pile_set_name": "ArXiv" }
--- abstract: 'The interaction between a straight vortex line in a superconducting film and a soft magnetic nanodisk in the magnetic vortex state in the presence of a magnetic field applied parallel to the film surfaces is studied theoretically. The superconductor is described by London theory and the nanodisk by the Landau-Lifshitz continuum theory of magnetism, using the approximation known as the rigid vortex model. Pinning of the vortex line by the nanodisk is found to result, predominantly, from the interaction between the vortex line and the changes in the nanodisk magnetization induced by the magnetic field of the vortex line and applied field. In the context of the rigid vortex model, these changes result from the displacement of the magnetic vortex. This displacement is calculated analytically by minimizing the energy, and the pinning potential is obtained. The applied field can tune the pinning potential by controlling the displacement of the magnetic vortex. The nanodisk magnetization curve is predicted to change in the presence of the vortex line.' author: - Gilson Carneiro title: Tunable pinning of a superconducting vortex a by a magnetic vortex --- introduction {#sec.int} ============ Pinning of vortices in superconducting films by arrays of soft nanomagnets placed in the vicinity of the film is a subject that has not been much studied in the literature. Most experimental [@rev1] and theoretical [@coff; @wei; @sah; @myp1; @myp2; @pri] work carried out so far deals with permanent nanomagnets. For both soft and permanent nanomagnets, pinning results from the action of the inhomogeneous magnetic field created by their magnetization in the superconductor. However, in the case of soft nanomagnets, the magnetization depends on the magnetic fields acting on it, including the field caused by the vortex itself. The question is then how this modifies the pinning potential. Recently, this question was investigated for vortices interacting with arrays of soft magnetic nanodisks using a very simple model in which the nanodisks are approximated by point magnetic dipoles with magnetic moments free to rotate [@gmc1]. This model was suggested by the results of Cowburn et. al. [@ckaw], which show that in small nanodisks the magnetic state is a single domain one, with the magnetization of each nanodisk behaving like a giant magnetic moment free to rotate. The calculations of Ref. show that the pinning potential differs considerably from that for a permanent dipole, and that it can be tuned by a magnetic field applied parallel to the film surfaces. Similar properties are expected for the pinning potential due to soft nonomagnets in general. This paper studies the pinning of vortices by nanodisks in another magnetic state: the magnetic vortex. This state is found in larger nanodisks [@ckaw; @dsk1; @dsk2] , and is characterized by the magnetization vector circulating around a nucleus of small dimensions compared with the nanodisk radius. The motivation to consider this particular state is that it contains the basic physics of nanomagnets in general, but its theoretical description can be approximated by a simple analytic model. Here, the interaction between one straight vortex line and one nanodisk is calculated using the Landau-Lifschitz continuum theory of magnetism [@ll] for the nanodisk and London theory for the superconductor [@gmc2]. The magnetic vortex state is described by the rigid vortex model introduced by Usov and Peschany [@up], and by Guslienko and collaborators [@gus]. These authors obtained the nanodisk equilibrium magnetization using the variational principle to minimize the energy. Usov and Peschany [@up] proposed an analytic expression for the trial magnetization at zero applied field, with the magnetic vortex core at the nanodisk center, and the core radius as the variational parameter. Guslienko and collaborators [@gus] extended it to finite fields parallel to the nanodisk faces. They assumed that the only effect of the applied field is to displace the magnetic vortex rigidly, with the core moving away from the nanodisk center. The new trial magnetization is the zero field one displaced rigidly with the core, and the core displacement vector is a variational parameter. The predictions of the rigid vortex model are in reasonable agreement with numerical calculations for small displacements compared to the nanodisk radius [@up; @ckaw; @gus]. This paper assumes the validity of the rigid vortex model for the nanodisk placed in the vicinity of the superconducting film. The total energy of the nanodisk-superconductor system is obtained exactly as a function of the magnetic vortex displacement vector. The calculations carried out in the paper assume small displacements compared to the nanodisk radius, and aproximate the the total energy by its expansion to second order in the displacement. The equilibrium displacement of the magnetic vortex is obtained by minimizing the approximate total energy. It is found that, contrary to what happens for the isolated nanodisk, the equilibrium displacement is a non-linear function of the vortex line and applied fields. The pinning potential for the vortex line is also calculated from the approximate total energy. The most important contributiont to the pinning potential is found to come from the changes in the magnetization caused by the displacement of the magnetic vortex. There is also a contribution from the interaction between the vortex line and the magnetization in the magnetic vortex core, but it is negligible, due to the small dimensions of the core. It is found that the pinning potential can be tuned by the applied field. The mechanism is that the field controls the magnetic vortex displacement which, in turn, modifies the pinning potential. Applications to systems of experimental interest are considered by appropriate choice of the model pararmeters. The results suggest that the pinning potential for these systems can be estimated by applying the linear response theory of elementary magnetism to the nonodidk. The magnitude of the applied field necessary to annihilate the magnetic vortex is estimated, from which changes in the nanodisk magnetization curve caused by the vortex line can be aticipated. This paper is organized as follows. The theory for the nanodisk-superconductor system is developed in Sec. \[sec.pnp\]. First, in Sec. \[sec.ind\], the rigid vortex model for the isolated nanodisk is briefly reviewed. Then, in Sec. \[sec.nsi\], the nanodisk in the vicinity of the seperconductor is considered, and the main results of the paper are derived. Finally, in Sec. \[sec.dis\], the application of the model to systems of experimental interest is considered, and the conclusions of the paper are stated. The Appendix gives the mathematical expressions needed in the calculations carried out in Sec. \[sec.nsi\], and a brief review of their derivation. ![ Schematic view of the superconducting film with one vortex line and a nanodisk placed on top. []{data-label="fig.fig1"}](V5FIG1.eps) pinning potential {#sec.pnp} ================= The superconductor-nanodisk system is shown schematically in Fig. \[fig.fig1\]. The superconducting film is planar, of thickness $d$, isotropic and has penetration depth $\lambda$. Its surfaces are parallel to the $x-y$ plane. One straight vortex line, with vorticity $q=\pm 1$, is present in the film at a position defined by ${\mbox{\boldmath $\rho$}}_{\rm v}$ with respect to an origin in the film top surface. The vortex line core radius is $\xi$. A thin magnetic nanodisk, of radius $R$ and thickness $L_z\ll R$, is located above the film, with the faces parallel to the film surfaces, and center located above the film at $(x=0,\,y=0,\,z=z_0>0)$. A magnetic field, ${\bf H}$, is applied along the $x$-axis, that is parallel to the nanodisk faces and to the film surfaces. ![ Nanodisk magnetization in the magnetic vortex state ($q_m=1$) with core at nanodisk center (a), and displaced by ${\mbox{\boldmath $\delta$}}$ from nanodisk center (b). Also shown in a) is the cylindrical coordinate system with origin at nanodisk center used in Eq. (\[eq.rgv\]).[]{data-label="fig.fig2"}](V5FIG2.eps) isolated nanodisk {#sec.ind} ----------------- First, the rigid vortex model for the isolated nanodisk is briefly reviewed. The magnetization is written as ${\bf M}({\bf r})=M_s \hat {\bf s}({\bf r})$, $M_s$ being the saturation magnetization. The nanodisk energy is given by [@ll] $$\begin{aligned} & & E_D=\frac{M^2_s\,R^2_0}{2}\int\, d^{3}r\sum_{\alpha=x,y,z} \mid{\mbox{\boldmath $\nabla$}}s_{\alpha}({\bf r})\mid^2 \nonumber \\ & & +\frac{1}{2}\int\, d^{2}r\,\int\, d^{2}r'\, \frac{\sigma({\bf r})\sigma({\bf r}')}{\mid{\bf r} -{\bf r}'\mid} - {\bf m}\cdot{\bf H}\;. \label{eq.edk}\end{aligned}$$ In Eq. (\[eq.edk\]), the first term is the exchange energy, where $R_0$ is the exchange length. The second is the magnetostatic energy, with $\sigma({\bf r})=M_s\hat {\bf s}({\bf r})\cdot \hat {\bf n}$ being the density of magnetic charges at the surface of the nanodisk, and $\hat {\bf n}$ is the unit vector in the direction normal to the nanodisk surface. The last term is the energy of interaction with the external field, with ${\bf m}=\int\,d^3r\,{\bf M}({\bf r})$. It is assumed that the nanodisk is isotropic, and that there are no volume magnetic charges, that is ${\mbox{\boldmath $\nabla$}}\cdot {\bf M}=0$. The rigid vortex model is a variational minimization of Eq. (\[eq.edk\]). The trial magnetization is written in the cylindrical coordinate system with origin at the nanodisk center and $z$-axis perpendicular to the nanodisk faces shown in Fig. \[fig.fig2\]a. For $H=0$ Usov and Peschany [@up] proposed the following expression $$\begin{aligned} \hat {\bf s}({\mbox{\boldmath $\rho$}}) & = & q_m\, {\hat {\mbox{\boldmath $\theta$}}} \;\; (a<\rho\leq R) \nonumber \\ \hat {\bf s}({\mbox{\boldmath $\rho$}}) & = & q_m\frac{2a\rho}{a^2+\rho^2}{\hat {\mbox{\boldmath $\theta$}}} \pm \frac{a^2-\rho^2}{a^2+\rho^2} {\hat {\bf z}} \;\; (0<\rho\leq a) \;, \label{eq.rgv} \end{aligned}$$ where ${\hat {\mbox{\boldmath $\theta$}}}= {\hat {\bf z}}\times {\hat {\mbox{\boldmath $\rho$}}}$. This equation describes the magnetization vector curling around the nanodisk center ( Fig. \[fig.fig2\]a ), with counterclockwise ( clockwise) rotation for $q_m=+ 1\; (-1)$. The variational parameter $a$ is the radius of the magnetic vortex core. The magnetization outside the core is parallel to the nanodisk faces. Inside it has a $z$-component, which can be either positive or negative. The equilibrium value of $a$, obtained by Usov and Peschany, is $a\approx 0.7(R^2_0L_z)^{1/3}$ [@up]. What is important for the purposes of this paper is not the precise value of $a$, but the fact that it is small compared to $R$. This follows from the assumption that $R_0,\;L_z \ll R$. For ${\bf H}\neq 0$ the trial magnetization proposed by Guslienko and collaborators [@gus] is that in Eq. (\[eq.rgv\]) displaced rigidly by ${\mbox{\boldmath $\delta$}}$ from the nanodisk center, that is $ {\hat {\bf s}}(\mid {\mbox{\boldmath $\rho$}}-{\mbox{\boldmath $\delta$}}\mid )$ ( Fig. \[fig.fig2\]b ). The equilibrium value of ${\mbox{\boldmath $\delta$}}$ is determined by minimizing the energy. The vortex core radius is unchanged as long as the displaced core remains within the nanodisk, that is for $\delta < R-a$. The simplest treatment of the rigid vortex model assumes that the magnetic vortex displacement is small, and obtains $E_{D}$ to second order in $\delta / R$. The result is [@gus] $$E_{D}({\mbox{\boldmath $\delta$}})= E_{D}(0) +M^2_sV_D[\;\frac{\delta^2_x+\delta^2_y}{2\,\chi\, R^2} - q_m\frac{H \delta_y}{M_sR}\;]\;, \label{eq.enmd}$$ where $V_D=\pi R^2L_z$ is the nanodisk volume, and $\delta_x$ and $\delta_y$ are, respectively, the components of parallel and perpendicular to ${\bf H}$. In Eq. (\[eq.enmd\]) the term quadratic in $\delta$ comes from both the exchange and magnetostatic energies, whereas the term linear in $\delta$ comes from the interaction with the applied field. The constant $\chi$ is the nanodisk linear magnetic susceptibility , and is given by[@gus] $$\frac{1}{\chi} = -\frac{R^2_0}{R^2} + 2\frac{L_z}{R}[\ln{\frac{8R}{L_z}}-\frac{1}{2}]\;. \label{eq.chi}$$ The first term in Eq. (\[eq.chi\]) is the contribution from the exchange energy, and is small because $R\gg R_0$. The second term comes from the magnetostatic energy of the magnetic charge density generated at the nanodisk edges by the vortex displacement. The density of these charges is given by $$\sigma = -q_mM_s\frac{(\hat {\bf z}\times {\mbox{\boldmath $\delta$}})\cdot \hat {\mbox{\boldmath $\rho$}}} {(R^2-2R {\mbox{\boldmath $\delta$}}\cdot \hat {\mbox{\boldmath $\rho$}}+\delta^2)^{1/2}} \label{eq.sig}$$ The interaction between the magnetic charges in the vortex core and in the nanodisk edges is neglected. To calculate the magnetostatic energy to second order in $\delta/R$, it is sufficient to use $\sigma$ to first order, that is $$\sigma \approx \sigma^{(1)} = -q_mM_s\frac{(\hat {\bf z}\times {\mbox{\boldmath $\delta$}})\cdot \hat {\mbox{\boldmath $\rho$}}}{R} \;. \label{eq.sig1}$$ The energy of interaction with the applied field also comes from $\sigma^{(1)}$. The magnetic moment generated by the displacement of the magnetic vortex is related to $\sigma$, Eq. (\[eq.sig\]), by $${\bf m}({\mbox{\boldmath $\delta$}}) = \int \, d^2 r \,\sigma \,{\mbox{\boldmath $\rho$}} \;, \label{eq.msg}$$ where the integral is over the nanodisk edge surface. To first order in $\delta/R$, the magnetic moment is obtained by replacing $\sigma$ by $\sigma^{(1)}$ in Eq. (\[eq.msg\]). The result is is $${\bf m}^{(1)}({\mbox{\boldmath $\delta$}}) = -q_mM_sV_D\frac{(\hat {\bf z}\times {\mbox{\boldmath $\delta$}})}{R} \;. \label{eq.m1}$$ By symmetry $ {\bf m}({\mbox{\boldmath $\delta$}})$ does contain terms of second order in $\delta/R$. The equilibrium vortex displacement, obtained by minimizing $E_{D}$ with respect to ${\mbox{\boldmath $\delta$}}$, is $${\mbox{\boldmath $\delta$}}_{eq} = q_m\chi\frac{\hat {\bf z}\times {\bf H}}{M_s} \;. %\delta_{x,eq}=0; \;\;\;\; %\frac{\delta_{y,eq}}{R}=q_m\chi \frac{H}{M_s} \; . \label{eq.dtid}$$ In equilibrium, the magnetic moment is $${\bf m}_{eq}={\bf m}^{(1)}({\mbox{\boldmath $\delta$}}_{eq})=\chi V_D{\bf H}\;. \label{eq.meq}$$ This result shows that $\chi$ is the nanodisk susceptibility. According to Eq. (\[eq.dtid\]), ${\mbox{\boldmath $\delta$}}_{eq}$ is perpendicular to ${\bf H}$, and depends only on $H$ scaled by $M_s/\chi$. As will be shown in Sec. \[sec.nsi\], $M_s/\chi$ is also the scale for the combined effects of the applied field and vortex line field on the magnetic vortex. The approximation to second order in $\delta/R$ to $E_D$ is valid only for $\chi H/M_s < 1$. However, it has been used outside this range to estimate the vortex annihilation field, $H_{an}$, at which the magnetic vortex is destroyed, and the nanodisk switches to the saturated state [@gus]. This estimate assumes that at $H=H_{an}$ the magnetic vortex reaches the nanodisk edge, that is $\delta_{eq}=R$. Thus, according to Eq. (\[eq.dtid\]), $H_{an}=M_s/\chi$. In this case the nanodisk magnetic moment is the saturation moment, $M_sV_D$ in the direction of ${\bf H}$ (see Eq. (\[eq.meq\])). This estimate for $H_{an}$ is in reasonable agreement with experiments and numerical simulations [@gus; @ckaw]. nanodisk-superconductor interaction {#sec.nsi} ----------------------------------- Now the nanodisk is assumed to be in the proximity of the superconducting film, as shown in Fig. \[fig.fig1\]. The equilibrium displacement of the magnetic vortex is recalculated by minimizing the total energy of the superconducting film-nanodisk system, $E_{T}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}})$, to second order in $\delta/R$. In the London limit [@gmc2] $$E_{T}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}}) = E_{D}({\mbox{\boldmath $\delta$}})+ E_{MS}({\mbox{\boldmath $\delta$}})+E_{VM}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}}) \;, \label{eq.et1}$$ where $E_{MS}$ is the energy of interaction of the nanodisk with the superconducting film in the absence of vortices, and $E_{VM}$ is the interaction energy of the vortex line and the nanodisk. The influence of ${\bf H}$ on the superconductor is neglected. This is justifiable as long as $H$ is smaller than the lower critical field parallel to the film surfaces. The energy $E_{MS}$ results from the interaction between the nanodisk and the magnetic field of the screening current generated by it in the superconductor[@gmc2]. As shown in the Appendix, $E_{MS}$ can be written as an interaction between magnetic charges, $$E_{MS} = \frac{1}{2}\int d^{2}r\int d^{2}r'\, \sigma({\bf r})\sigma({\bf r}') U_{MS}({\bf r};{\bf r}')\; . \label{eq.ems}$$ Thus, in the presence of the superconducting film, the interaction between magnetic charges is modified by the addition of $U_{MS}$ to the Coulomb potential ( see Eq. (\[eq.edk\])). In Fig. \[fig.fig3\], $\;U_{MS}$ is shown for two magnetic charges separated by $\rho$ and at the same height $z$, and compared with the Coulomb potential. For large distances, $U_{MS}$ coincides with the Coulomb potential. Large distances meaning $\rho \gg \Lambda= 2\lambda^2/d\;$ for thin films ($d\ll \lambda$) and $\rho \gg \lambda$ for films with $d\sim \lambda$. For short distances, $U_{MS}$ is considerably smaller than the Coulomb potential. The effects of $E_{MS}$ on the magnetic vortex are to change the vortex core radius, $a$, and nanodisk susceptibility, $\chi$. However these modifications are small. The reason is that the dominant contributions to $a$ and $\chi$ come from short distances [@up; @gus], where $U_{MS}$ is much smaller than the Coulomb interaction. For instance, the contribution to $\chi^{-1}$ from $E_{MS}$ is found to be about one order of magnitude smaller than that from the Coulomb interaction, Eq. (\[eq.chi\]). Hereafter $E_{MS}$ is neglected. ![a) Interaction potential induced by superconducting film between two magnetic charges separated by $\rho$, and at same height $z$, compared with the Coulomb potential. a) $\lambda U_{MS}$ vs. $\rho$ for $d=\lambda$ and two $z$ values. Inset: magnetic charges above superconducting film. b) Ratio $U_{MS}$ to Coulomb potential, $\rho U_{MS}$, vs. $\rho$ for $z=0.2\lambda$ and several $d$. []{data-label="fig.fig3"}](V5FIG3.eps) The energy $E_{VM}$ is given by [@gmc2] $$\begin{aligned} E_{VM}&=& - \int\, d^3r' \, {\bf M}({\bf r'})\cdot\, {\bf b}({\bf r}'-{\mbox{\boldmath $\rho$}}_{\rm v}) \label{eq.evma} \\ & = &\int\, d^2r' \, \sigma({\bf r'}) \Phi({\bf r}'-{\mbox{\boldmath $\rho$}_{\rm v}}) \,. \label{eq.evmb} \end{aligned}$$ where ${\bf b}({\bf r}) = -\,{\mbox{\boldmath $\nabla$}}\,\Phi({\bf r})$ is the magnetic field created by the vortex line outside the film. There are two contributions to $E_{VM}$: one from the magnetic vortex core, and another from the magnetic charge density at the nanodisk edges, $\sigma$, Eq. (\[eq.sig\]), denoted by $E^{(e)}_{VM}$. The contribution to $E_{VM}$ from the core is negligible, as justified later. In order to obtain $E^{(e)}_{VM}$ to second order in $\delta/R$ it is necessary to use $\sigma$ to the same order. The first order term is $\sigma^{(1)}$, Eq. (\[eq.sig1\]). The second order term is, according to Eq. (\[eq.sig\]), given by $$\sigma^{(2)}= \sigma^{(1)}\frac{{\mbox{\boldmath $\delta$}}\cdot \hat {\mbox{\boldmath $\rho$}}}{R} \;\;. \label{eq.sig2}$$ The contribution from $\sigma^{(1)}$ to $E_{VM}$, denoted $E^{(e1)}_{VM}$, is obtained by noting that $\sigma^{(1)}$ can be interpreted as resulting from the uniform magnetization $${\bf M}^{(1)}=-q_mM_s\frac{\hat {\bf z}\times {\mbox{\boldmath $\delta$}}}{R} \;. \label{eq.bm1}$$ Using Eq. (\[eq.evma\]) it follows that $$E^{(e1)}_{VM}= - V_D{\bf M}^{(1)}\cdot\, {\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v}) \,, \label{eq.upm}$$ where $$\begin{aligned} {\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v}) & = & \frac{1}{V_D}\,\int_{\rm disk}\, d^3r' \, {\bf b}_{\perp}({\bf r}'-{\mbox{\boldmath $\rho$}}_{\rm v}) \nonumber \\ & = & -qB_{\perp}(\rho_{\rm v})\hat {\mbox{\boldmath $\rho$}}_{\rm v}, \label{eq.bav} \end{aligned}$$ is the average over the nanodisk volume of the component of the vortex field perpendicular to the $z$-direction ( parallel to the nanodisk faces ), ${\bf b}_{\perp}$. The argument of $ {\bf B}_{\perp}$ in Eq. (\[eq.bav\]) is $-{\mbox{\boldmath $\rho$}}_{\rm v}$ because ${\mbox{\boldmath $\rho$}}_{\rm v}$ is defined with respect to the origin shown in Fig. \[fig.fig1\], whereas in the definition of the vortex line field the origin is at the vortex line ( see Eq. (\[eq.phi\])). Thus $$E^{(e1)}_{VM} =q_m\,M_sV_D\,qB_{\perp}(\rho_{\rm v})\,\frac{\delta_T}{R} \;\;, \label{eq.evme1}$$ where $\delta_T$ denotes the component of ${\mbox{\boldmath $\delta$}}$ perpendicular to ${\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v})$ ( parallel to $-{\hat {\bf z}}\times {\hat {\mbox{\boldmath $\rho$}}}_{\rm v}$, see Fig. \[fig.fig4\]). The contribution from $\sigma^{(2)}$ is calculated in the Appendix. The result is $$E^{(e2)}_{VM}= q_m\,M_sV_D \, qB_1(\rho_{\rm v})\,\frac{\delta_L\delta_T}{R^2} \;\; , \label{eq.evme2}$$ where $\delta_L$ denotes the component of ${\mbox{\boldmath $\delta$}}$ parallel to ${\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v})$ (Fig. \[fig.fig4\]), and $B_1(\rho_{\rm v})$ has dimension of magnetic field, and is given by Eq. (\[eq.b1\]). ![Definition of the displacements $\delta_L$, $\delta_T$, and angles $\theta$, and $\gamma$. []{data-label="fig.fig4"}](V5FIG4.eps) Now the equilibrium displacement of the magnetic vortex is calculated neglecting the contribution from the vortex core to $E_{VM}$. The total energy is thus $$\begin{aligned} & &E_{T}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}}) = E_{D}(0) + M^2_sV_D[\;\frac{\delta^2_T+\delta^2_L}{2\,\chi \,R^2}+ \nonumber \\ & & q_m\frac{qB_{\perp}(\rho_{\rm v})\delta_T-H\delta_y}{M_s R} + q_m\, \frac{qB_1(\rho_{\rm v})\delta_L\delta_T}{M_sR^2}\;] \;\;. \label{eq.etmv} \end{aligned}$$ The components of ${\mbox{\boldmath $\delta$}}$ parallel and perpendicular to ${\bf H}$, $\delta_x$ and $\delta_y$, respectively, are related to $\delta_L$ and $\delta_T$ by (see Fig. \[fig.fig4\]). $$\begin{aligned} \delta_x & = &-\delta_L \cos{\theta} +\delta_T \sin{\theta} \nonumber \\ \delta_y &= & -\delta_L \sin{\theta} -\delta_T \cos{\theta} \; .\end{aligned}$$ Minimizing $E_T$ with respect to ${\mbox{\boldmath $\delta$}}$, with the vortex line held fixed at ${\mbox{\boldmath $\rho$}_{\rm v}}$, it follows that $$\begin{aligned} & & \frac{\delta_{T,eq}({\mbox{\boldmath $\rho$}_{\rm v}})}{R} = q_m\frac{q{\tilde B}_{\perp}-{\tilde H}\cos{\theta} + q_mq{\tilde B}_1{\tilde H}\sin{\theta}}{1-{\tilde B}_1^2} \; ,\nonumber \\ & &\frac{\delta_{L,eq}({\mbox{\boldmath $\rho$}_{\rm v}})}{R} = \ -q_m\frac{{\tilde H}\sin{\theta}+q_mq{\tilde B}_1 (q{\tilde B}_{\perp}-{\tilde H}\cos{\theta})}{1-{\tilde B}_1^2} \; ,\nonumber \\ & & {\tilde B}_{\perp} \equiv \frac{\chi B_{\perp}(\rho_{\rm v})}{M_s}\; , {\tilde B}_{1} \equiv \frac{\chi B_{1}(\rho_{\rm v})}{M_s}\; , {\tilde H} \equiv \frac{\chi H}{M_s}\; . \label{eq.dlt1} \end{aligned}$$ Thus ${\mbox{\boldmath $\delta$}}_{eq}$ depends only on the vortex line and applied fields scaled by $M_s/\chi$, and is non-linear in these fields. The non-linearity results from the inhomogeneity of the vortex field over the nanodisk volume, which is responsible for $B_1$ being non-zero. This solution is only valid if ${\tilde B}_{1}^2<1$. Otherwise ${\mbox{\boldmath $\delta$}}_{eq}$, Eq. (\[eq.dlt1\]), does not correspond to a minimum of $E_{T}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}})$, Eq. (\[eq.etmv\]). The consequences of Eq. (\[eq.dlt1\]) are discussed in detail in Sec. \[sec.dis\]. Now the neglect of the core contribution to $E_{VM}$, denoted $E^{(c)}_{VM}$, is justified. Since $a$ is small compared to $R$, $E^{(c)}_{VM}$ can be estimated by approximating the core by a point dipole located at the nanodisk center with the total magnetic moment of the core. According to Eq. (\[eq.rgv\]), it has only the $z$-component $m^{(c)}_z=\pm M_s\pi a^2L_z(2\ln{2}-1)$. Thus, $E^{(c)}_{VM}\approx - m^{(c)}_z\,b_z({\mbox{\boldmath $\delta$}}-{\mbox{\boldmath $\rho$}_{\rm v}}+z_0\hat{\bf z})$. The basic reason why $E^{(c)}_{VM}$ can be neglected is that it is proportional to $a^2$, which is already a small quantity. One effect of $E^{(c)}_{VM}$ is to modify the vortex core. Since it depends on $a$, its contribution must be added to $E_D$ in order to obtain the equilibrium value of $a$. However, the effect is a small one because $E_D\sim M^2_sa^4/L_z$ [@up], so that $E^{(c)}_{VM}/E_{D}\sim (L_z/a)^2(b_z/M_s)$, and $(b_z/M_s)$ is small, as shown in Sec. \[sec.dis\]. Another effect of $E^{(c)}_{VM}$ is to displace the magnetic vortex from the nanodisk center. The displacement can be estimated as follows. The force exerted by the superconducting vortex on the magnetic vortex core is $\mid {\mbox{\boldmath $\nabla$}}E^{(c)}_{VM}\mid \sim M^2_sV_D (a/R)^2 b_z/M_s\ell$, where $\ell$ is the typical scale for variations of $b_z$, namely $\ell \sim z_0$ for thin films ($d \ll \lambda$) and $\ell \sim \lambda$ for films with $d\sim \lambda$. This force must be balanced by the elastic force $M_s^2V_D \delta/\chi\,R^2$, which gives $\delta/R\sim (a/R)^2 (R/\ell) \chi \,b_z/M_s$, and is smaller than the vortex displacement caused by $E^{(e)}_{VM}$ at least by a factor $a/R$. The pinning potential for the vortex line is defined as the equilibrium total energy for the vortex line held fixed at ${\mbox{\boldmath $\rho$}}_{\rm v}$. That is $ U_p({\mbox{\boldmath $\rho$}}_{\rm v})= E_{T}({\mbox{\boldmath $\rho$}}_{\rm v},{\mbox{\boldmath $\delta$}}_{eq})$. Thus $$\begin{aligned} & & U_p({\mbox{\boldmath $\rho$}}_{\rm v})= - \frac{\chi V_D}{2(1-{\tilde B}_1^2)}[B^2_{\perp}- 2qB_{\perp}H\cos{\theta} \nonumber \\ & & +2q_mq{\tilde B}_1H\sin{\theta}(qB_{\perp}-H\cos{\theta})]\; . \label{eq.umvh} \end{aligned}$$ This result shows that the pinning potential can be tuned by the applied field. The mechanism is that, according to Eq. (\[eq.dlt1\]), ${\bf H}$ controls the equilibrium displacement of the magnetic vortex which, in turn, modifies $U_p$. In order to evaluate the consequences of the above results it is necessary to attribute values to the model parameters. This is done next. discussion {#sec.dis} ========== The results of Sec. \[sec.nsi\] are applicable to arrays of soft magnetic nanodisks placed on top of low-T$_c$ superconducting films, provided that the nanodisks are sufficiently far apart to neglect dipole-dipole interactions between them, and the vortex density is low enough to neglect vortex-vortex interactions. Realistic values for the model parameters are the following [@rev1; @ckaw; @dsk1; @dsk2]. Permalloy nanodisks: $100\,nm < R < 500\,nm$, $L_z \sim 10\, nm$, $M_s = 0.8\,kG$, and $R_0\, \sim 15\, nm$. Superconducting films at zero temperature: $\lambda = 75\, nm$, $\xi = 0.1\, \lambda $, $d \sim\, 0.2 - 2\, \lambda$. With this value for $\lambda$, the scale for the vortex fields is $\phi_0/\lambda^2= 3.6\, kG$. The distance from the nanodisk center to the film surface is chosen as the smallest possible, namely $z_0=0.2\lambda=2\xi$. This value takes into account the existence of an insulating layer between the film and the nanodisk, of thickness $\sim \xi$, to avoid the proximity effect. It is found that for these parameter values the vortex fields $B_{\perp}$ and $B_1$ are small compared to $M_s/\chi$. As consequences, ${\tilde B}_1$ is negligible in Eq. (\[eq.dlt1\]), the magnetic vortex displacement is small compared with and nanodisk radius, and its dependence on the vortex field is linear. This is discussed in detail next. ![ a) Fields $B_{\perp}$ and $B_1$ , in units of $\phi_0/\lambda^2$, vs. $\rho_{\rm v}$ for $R=2.0\lambda$. b) Maximum values of $B_{\perp}$ and $B_1$ in units of $\phi_0/\lambda^2$ vs. $R$. The right-hand scale in Gauss corresponds to left-hand one for $\lambda= 75\,nm$. Parameters: $ \xi=0.1\lambda, \,z_0=0.2\lambda,\, L_z=0.1\lambda$.[]{data-label="fig.fig5"}](V5FIG5.eps) The fields $B_{\perp}$ and $B_1$ are shown in Fig . \[fig.fig5\]. They depend only on scaled variables, with $\lambda$ as the length scale and $\phi_0/\lambda^2$ as the magnetic field scale (see Eqs. (\[eq.phi\]), (\[eq.f1k\])). Their dependencies on $\rho_{\rm v}$ are shown in Fig. \[fig.fig5\].a. Both vanish for $\rho_{\rm v}=0$, and have a maximum at $\rho_{\rm v}\sim R$, with $B_1$ smaller than $B_{\perp}$ by a factor $\sim 3$. The maximum values of $B_{\perp}$ and $B_1$ depend on $R$ as shown as in Fig. \[fig.fig5\].b. The right-hand scale in Gauss in this figure corresponds to the left-hand side one for $\phi_0/\lambda^2=3.6 \, kG$. Both $B_{\perp}$ and $B_1$ increase non-linearly with $d$, up to $d \sim 2.0 \lambda$. For larger $d$, they change little, because the vortex line field is generated by currents flowing close to the film surface. These results indicate that ${\tilde B}_1$ is small. Using $\chi \sim 2$, which corresponds to $R\sim 200\, nm,\, L_z=10\, nm$ ( see Eq. (\[eq.chi\]) ), the maximum values of ${\tilde B}_1$, obtained from the $B_1$ data in Fig . \[fig.fig5\]b, are $({\tilde B}_{1})_{max} \sim 5 \times 10^{-2}$ for $d=2.0\lambda$ and $({\tilde B}_{1})_{max} \sim 1.0 \times 10^{-2}$ for $d=0.2\lambda$. Thus ${\tilde B}_1^2 \ll 1$, and can be neglected in the denominators of Eqs. (\[eq.dlt1\]) and (\[eq.umvh\]). The terms with ${\tilde B}_{1}$ in numerator of these equations are also small, because they involve the combinations ${\tilde B}_{1}{\tilde H}$, and ${\tilde B}_{1}{\tilde B}_{\perp}$, which are in general smaller than the other terms. ![ a) Equilibrium displacements of the magnetic vortex ($q_m=1$) for , $H=0$, $R=2.0\lambda$. b) Maximum values of the equilibrium displacements vs. $R$ for $H=0$. Parameters: $ \xi=0.1\lambda, \,z_0=0.2\lambda,\, L_z=0.1\lambda$, $M_s=0.22\phi_0/\lambda^2$. []{data-label="fig.fig6"}](V5FIG6.eps) Results for $\delta_{L,eq}$ and $\delta_{T,eq}$ at $H=0$ are shown in Fig. \[fig.fig6\]. In this case, according to Eq. (\[eq.dlt1\]), $\mid \delta_{L,eq}/\delta_{T,eq}\mid = {\tilde B}_{1} $. Thus, $\mid \delta_{L,eq}/\delta_{T,eq}\mid \ll 1$ since $ {\tilde B}_{1}$ is small. The curves in Fig. \[fig.fig6\].a show this. These plots also show that $\delta_{T,eq}/R <0.15 $ for $d=2.0\lambda$ and $\delta_{T,eq}/R <0.06 $ for $d=0.2\lambda$. In Fig. \[fig.fig6\].b the maximum values of $\delta_{L,eq}$ and $\delta_{T,eq}$ are shown as a function of $R$. Both vary little with $R$. The reason is that, according to Eq. (\[eq.dlt1\]), the $R$-dependence of $\delta_{L,eq}$ and $\delta_{T,eq}$ is only through the products $\chi B_{\perp}$ and $\chi B_1$. These are nearly independent of $R$, because there is a cancelation between the increase of $\chi$ with $R$ ( see Eq. (\[eq.chi\])) and the decrease of $B_{\perp}$ and $B_1$ with $R$ ( Fig. \[fig.fig5\].b). The above discussion suggest that, for the parameter values described above, a good approximation to the equilibrium displacement is to put ${\tilde B}_{1}$ equal to zero in Eq. (\[eq.dlt1\]). In this case ${\mbox{\boldmath $\delta$}}_{eq}$, is identical to that for the isolated nanodisk in the applied field ${\bf H}_T ={\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v}) + {\bf H}$. That is, Eq. (\[eq.dlt1\]) with ${\tilde B}_{1}=0$, is identical to Eq. (\[eq.dtid\]) with ${\bf H}$ replaced by ${\bf H}_T$. In this approximation the pinning potential, Eq. (\[eq.umvh\]), reduces to $$U_p= - \frac{\chi V_D}{2}[B^2_{\perp}- 2qB_{\perp}H\cos{\theta}] \;. \label{eq.uml}$$ This result is, up to a constant, the magnetostatic energy of interaction between the magnetic moment induced by ${\bf H}_T$ in the nanodisk, ${\bf m}_{eq}= \chi V_D{\bf H}_T $, and ${\bf H}_T$ itself. That is $$U_p= - \frac{1}{2}{\bf m}_{eq}\cdot {\bf H}_T + \frac{\chi V_D H^2}{2} \;. \label{eq.umlb}$$ This simple result is the just the linear response approximation from elementary macroscopic magnetism [@ecm] applied to the nanodisk. It depends only on the nanodisk susceptibility, $\chi$, and on the macroscopic vortex field acting on it, ${\bf B}_{\perp}(-{\mbox{\boldmath $\rho$}}_{\rm v})$. Results for the pinning potential $U_p({\mbox{\boldmath $\rho$}}_{\rm v})$, based on Eq. (\[eq.uml\]), are shown in Fig. \[fig.fig7\] as two-dimensional plots for characteristic values of $H$. For $H=0$, $U_p$ has circular symmetry, and is the same for vortices ($q=1$) and anti-vortices ($q=-1$), with a degenerate minimum located in a circle of radius $\rho_{\rm v}\sim R$ (Fig. \[fig.fig7\]a). For $H\neq 0$, $U_p$ has a non-degenerate minimum and also a maximum along the direction of ${\bf H}$. The plot in Fig. \[fig.fig7\]b corresponds to $H\lesssim B_{\perp}$, whereas that in Fig. \[fig.fig7\]c is for $H > B_{\perp}$. In the latter case the pinning potential reduces to $U_p({\mbox{\boldmath $\rho$}}_{\rm v})= qV_D\chi B_{\perp}(\rho_{\rm v}) H \cos{\theta}$, which is identical to that for a nanodisk with permanent uniform magnetization $\chi {\bf H}$. Since $B_{\perp} \ll M_s/\chi$, the range of $H$ values for which the vortex line field influences the pinning potential is small. Thus, except for small $H$, the spatial dependence of the pinning potential is given by $B_{\perp}(\rho_{\rm v}) \cos{\theta}$, and the magnitude is proportional to $H$. This allows for continuous tuning of $U_p$ over a wide range. It is useful to compare the magnitude of $U_p$ to that for the pinning potential for a nanodisk with permanent magnetization, equal to the saturation one, $U\sim M_sV_DB_{\perp}$. For $H<M_s/\chi$, $U_p$ is smaller than $U$, since for $H\lesssim B_{\perp}$, $\mid U_p/U\mid \sim \chi B_{\perp}/M_s \ll 1$ , and for $M_s/\chi> H > B_{\perp}$, $\mid U_p/U\mid \sim \chi H/M_s < 1$. Using Eq. (\[eq.umlb\]) to extrapolate $U_p$ to the region $H>M_s/\chi$, it follows that $U_p\sim U$ only at $H\sim M_s/\chi$. The linear response approximation is now used to estimate the effect of the vortex line on the magnetic vortex annihilation field, $H_{an}$. As in Sec. \[sec.ind\], $H_{an}$ is estimated as the value of $H$ for which $\delta_{eq}=R$. In the linear response approximation this occurs for $H_T = M_s/\chi$. Since $B_{\perp} \ll M_s/\chi$, $H_{an}$ differs little from that predicted for isolated magnetic vortex, that is $H_{an}\sim M_s/\chi$. Consequently, the nanodisk magnetization curve is expected to differ little from that for the isolated nanodisk. ![ Pinning potential for a vortex line by a magnetic nanodisk in the vortex state, in units of $\epsilon_0 \lambda=\phi_0^2/(16\pi^2\lambda)$, for $ d=2.0\lambda,\,z_0=0.2\lambda,\,\xi=0.1\lambda,\, R=1.5\lambda,\, L_z=0.1\lambda, \,M_s=0.1\phi_0/\lambda^2$. a) $H=0$. b) $H=0.01\phi_0/\lambda^2$. c) $H=0.1\phi_0/\lambda^2$.[]{data-label="fig.fig7"}](V5FIG7.eps) In summary then, for the parameter values described earlier, the main conclusions are: i) the vortex line field causes only a small displacement of the magnetic vortex, ii) the pinning potential is the energy of interaction between the magnetic moment induced in the nanodisk by the vortex line and applied fields and the fields themselves, calculated according to the laws of macroscopic linear magnetism, iii) the vortex line has little effect on the nanodisk magnetization curve. Now the non-linear dependence of ${\mbox{\boldmath $\delta$}}_{eq}$ on the vortex line field and applied fields, predicted by Eq. (\[eq.dlt1\]), is considered. The question is whether or not there are parameter values for which ${\tilde B}_{1}$ is sufficiently large that the non-linearity is important and what are its consequences. One limitation on ${\tilde B}_{1}$ is that Eq. (\[eq.dlt1\]) only makes sense if $\delta/R<1$. This effectively restricts ${\tilde B}_{1}$ to relatively small values. One reason is that the denominators in Eq. (\[eq.dlt1\]) cannot be too small. Another reason is that $B_1$ and $B_{\perp}$ are not independent (see Eqs. (\[eq.bavb\]), (\[eq.b1\])), so that if ${\tilde B}_{1}$ is not sufficiently small, ${\tilde B}_{\perp}$ is large enough to make $\delta/R>1$. It is found that, in general, $B_1<B_{\perp}$, as exemplified in Fig. \[fig.fig5\]. To investigate this quantitatively, the equilibrium displacement of the magnetic vortex for $H=0$ is calculated using a new set of model parameters. These are chosen in order to give displacements larger than the ones obtained with the parameters mentioned at the beginning of this Section. For the superconducting film the parameters are the same as before ($\xi=0.1 \lambda, \; z_0=0.2\lambda=2\xi$), except for $\lambda$ which is now chosen as $\lambda=50\,nm$. In this case the fields $B_{\perp}$ and $B_1$ are unchanged in units of $\phi_0/\lambda^2$, but their values in Gauss change because now $\phi_0/\lambda^2=8\,kG$. The new parameters for the nanodisk are chosen as $R=3.0 \lambda,\; L_z= 0.1 \lambda, M_s=0.4 kG$. In this case $\chi=3.0$ and $M_s/\chi =0.13\, kG$. The results for the magnetic vortex displacements are shown in Fig. \[fig.fig8\]a, and compared with those obtained from Eq. (\[eq.dlt1\]) with ${\tilde B}_{1}=0$. The differences between the curves for $\delta_{T,eq}$ with ${\tilde B}_{1}\neq 0$ and ${\tilde B}_{1}=0$ are due to ${\tilde B}_{1}^2$ in the denominator of Eq. (\[eq.dlt1\]). The results show that the differences is small, even when $\delta_{T,eq}$ is a considerable fraction of $R$. For $\delta_{L,eq}$, only the curves for ${\tilde B}_{1}\neq 0$ are shown, because $\delta_{L,eq}$ vanishes for ${\tilde B}_{1}=0$. The non-zero $\delta_{L,eq}$ shown in Fig. \[fig.fig8\]b results from the non-linearity of Eq. (\[eq.dlt1\]), with the most important contribution coming from the numerator. Comparing the curves for $\delta_{L,eq}$ and $\delta_{T,eq}$ in Fig. \[fig.fig8\]a, and using $\mid \delta_{L,eq}/\delta_{T,eq}\mid = {\tilde B}_{1} $, it follows that ${\tilde B}_{1}\lesssim 0.3$. This shows that the validity of Eq. (\[eq.dlt1\]) requires small values of ${\tilde B}_{1}$. The non-linear effects in the pinning potential $U_p$, Eq. (\[eq.umvh\]), for $H=0$ come only from ${\tilde B}_{1}^2$ in the denominator, and are also small according to the above discussion. The properties of $U_p$ for the new set of parameters are similar to those described earlier. ![ a) Equilibrium displacements of the magnetic vortex ( $q_m=1$ ) for $H=0$. b) Magnetic vortex annihilation field for vortex line at position $(x_{\rm v}, y_{\rm v})$ vs. $x_{\rm v}$ for $y_{\rm v}$ constant. Parameters: $\xi=0.1\lambda, \,z_0=0.2\lambda,\,R=3.0\lambda\, L_z=0.1\lambda, \, M_s=0.05\phi_0/\lambda^2$.[]{data-label="fig.fig8"}](V5FIG8.eps) The results shown in Fig. \[fig.fig8\].a also have consequences for the magnetic vortex annihilation field, $H_{an}$. When the vortex line displaces the magnetic vortex by a significant fraction of the nanodisk radius it takes only a small applied field to annihilate the magnetic vortex by further displacing it to the nanodisk edge. In this case $H_{an}$ differs significantly from that for the isolated nanodisk. To estimate $H_{an}$, Eq. (\[eq.dlt1\]) with ${\tilde B}_{1}=0$ is used. In this case, as discussed above, the magnetic vortex is annihilated, when $H_T\sim M_s/\chi$. This estimate for $H_{an}$ is shown in Fig. \[fig.fig8\].b. A strong dependence of $H_{an}$ on the vortex line position results. The absolute minimum of $H_{an}$ occurs when the vortex line is at the pinning potential minimum for $H=0$, located at $y_{\rm v}=0$ and $x_{\rm v}\sim -R$, since in this case the displacement of the magnetic vortex is maximum. Thus, when the vortex line is in equilibrium with the nanodisk $H_{an}$ is minimum. This result indicates that the vortex line in equilibrium with the nanodisk can significantly modify the magnetization curve. To conclude then, the magnitude of the displacement of the magnetic vortex caused by the vortex line depends on the particular values of the model parameters. In the cases were the displacement is a significant fraction of the nanodisk radius, effects of the non-linear relationship between the displacement and the vortex line field are felt, but are small, and the magnetization curve for the nanodisk in equilibrium with the vortex line is predicted to differ from that for the isolated nanodisk. In all cases, the displacement of the magnetic vortex and the pinning potential can be estimated by applying the linear response theory of elementary magnetism to the nanodisk. This suggests that the pinning potential for nanomagnets with other geometrical forms for which the magnetic vortex state has been reported [@vts] can be likewise estimated. Research supported in part by the Brazilian agencies CNPq, CAPES, FAPERJ, and FUJB. London theory results {#sec.mtd} ===================== Here some results of London theory for the superconducting film, obtained in Refs., are reviewed. The objective is to write out the mathematical expressions needed to carry out the calculations mentioned in Sec. \[sec.nsi\]. nanodisk-screening current interaction {#sec.ns} -------------------------------------- In the arrangement shown in Fig. \[fig.fig1\], the nanodisk generates in the superconducting film a screening current which, in turn, creates a magnetic field at the nanodisk. The energy of interaction of the nanodisk with this field is given by $$E_{MS}= -\frac{1}{2} \int\, d^3r \, {\bf M}({\bf r})\cdot\, {\bf b}_{sc}({\bf r}) \,, \label{eq.emsb}$$ where ${\bf b}_{sc}({\bf r})$ is the field of the screening current outside the film. The field ${\bf b}_{sc}$ is can be written as $$\begin{aligned} {\bf b}_{sc}({\bf r}) &= & -{\mbox{\boldmath $\nabla$}}\Phi_{sc}({\bf r})\; , \nonumber \\ \Phi_{sc}({\bf r}) & = & \int d^3r' ({\bf M}({\bf r}')\cdot\, {\mbox{\boldmath $\nabla$}'}) U_{MS}({\bf r};{\bf r}') \,, \label{eq.psc} \end{aligned}$$ where $U_{MS}$ is given by $$\begin{aligned} & & U_{MS}({\bf r};{\bf r}')=\int\, d^2k\, e^{i{\bf k }\cdot ({\bf r}_{\perp}-{\bf r}'_{\perp})}\,e^{-k(z+z')}\;g(k)\; , \nonumber \\ & & g(k) = \frac{\sinh{\tau d}}{k[e^{-\tau d}(k-\tau)^2 -e^{\tau d}(k+\tau)^2]} \; ,\nonumber \\ \tau&=&\sqrt{k^2+\lambda^{-2}}\;, \label{eq.ums} \end{aligned}$$ where ${\bf r}_{\perp}$ denotes the component of ${\bf r}$ perpendicular to the $z$-direction. Assuming that ${\mbox{\boldmath $\nabla$}}\cdot {\bf M}=0$, and integrating by parts, Eq. (\[eq.emsb\]) can be written as Eq. (\[eq.ems\]). vortex line field {#sec.vtf} ----------------- The field created by the vortex line outside the film is given by $$\begin{aligned} {\bf b}({\bf r})& = &-\,{\mbox{\boldmath $\nabla$}}\,\Phi({\bf r})\;\;, \nonumber \\ \Phi({\bf r})& = &-q\frac{\phi_0}{\lambda^2}\int\,\frac{d^{2}k}{(2\pi)^2}\,e^{i{\bf k }\cdot{\bf r}_{\perp}}\;e^{-kz}\, F_1(k)\;\;, \label{eq.phi}\end{aligned}$$ where ${\bf r}$ is the position vector with respect to an origin at the vortex line, and $$\begin{aligned} F_1(k) = \frac{e^{-2\xi^2k^2}[(k+\tau)e^{\tau d}+(k-\tau)e^{-\tau d}-2k]} {k\tau[(k-\tau)^2e^{-\tau d}-(k+\tau)^2e^{\tau d}]}\;. \label{eq.f1k}\end{aligned}$$ Integrating Eq. (\[eq.phi\]) over the direction of ${\bf k}$, $\Phi$ can be written as $$\Phi( r_{\perp},z)=-\,q\int^{\infty}_0\,\frac{dk}{2\pi}\,k J_0(kr_{\perp})\,e^{-kz}\, F_1(k)\; , \label{eq.phb}$$ where $J_0$ is the Bessel function of first kind. Thus, the vortex field can be written as $$\begin{aligned} {\bf b}({\bf r})&=& b_{\perp}(r_{\perp},z)\,\hat {\bf r}_{\perp}+b_z(r_{\perp},z)\,\hat{\bf z}\;,\nonumber\\ b_{\perp}&=&-q\frac{\phi_0}{\lambda^2} \,\int^{\infty}_0\,\frac{dk}{2\pi}\,k^2 J_1(kr_{\perp})\,e^{-kz}\, F_1(k)\;\; , \nonumber\\ b_z&=&-q\frac{\phi_0}{\lambda^2} \,\int^{\infty}_0\,\frac{dk}{2\pi}\,k^2 J_0(kr_{\perp})\,e^{-kz}\, F_1(k)\;. \label{eq.bvl} \end{aligned}$$ Using Eq. (\[eq.bvl\]) and Eq. (\[eq.bav\]), it follows that $$\begin{aligned} B_{\perp}(\rho_{\rm v})&=&-\frac{\phi_0}{\lambda^2} \,\int^{\infty}_0\,\frac{dk}{2\pi}\,S_1(k)\,k^2 J_1(k\rho_{\rm v})\,e^{-kz_0}\, F_1(k)\;\; , \nonumber\\ S_1(k)&=&\frac{2J_1(kR)}{kR}\,\frac{2\sinh{(kL_z/2)}}{kL_z}\;. \label{eq.bavb}\end{aligned}$$ superconducting vortex-magnetic vortex interaction {#sec.svmv} -------------------------------------------------- Here the contribution from $\sigma^{(2)}$ to the vortex line - nanodisk interaction is calculated. Using Eqs. (\[eq.evmb\]), and (\[eq.phi\]), $E^{(e2)}_{VM}$ can be written as $$\begin{aligned} & & E^{(e2)}_{VM} = \int^{z_0+L_z/2}_{z_0-L_z/2}\, dz' \, \int^{2\pi}_0 d\theta'\, \sigma^{(2)}(\theta') \nonumber \\ & & (-q\frac{\phi_0}{\lambda^2})\int\,\frac{d^{2}k}{(2\pi)^2}\, e^{i({\mbox{\boldmath \small$\rho$}'}-{\mbox{\boldmath \small$\rho$}_{\rm v})}\cdot{\bf k }}\;e^{-kz}\, F_1(k)\;, \label{eq.evme2b} \end{aligned}$$ where $\mid {\mbox{\boldmath \small$\rho$}'}\mid =R$. According to Eq. (\[eq.sig2\]), $\sigma^{(2)}$ is given by $$\sigma^{(2)}(\theta')=q_mM_s(\frac{\delta}{R})^2\,\sin{(\beta'-\theta')}\cos{(\beta'-\theta')}\, \label{eq.sig2b}$$ where $\beta'$ and $\theta'$ are, respectively, the angles between ${\mbox{\boldmath $\delta$}}$ and ${\bf k}$, and between ${\mbox{\boldmath $\rho$}}'$ and ${\bf k}$. Integrating over $z'$, $\theta'$ and $\beta'$, it follows that $$E^{(e2)}_{VM}= q_m\,M_sV_D \, qB_1(\rho_{\rm v})\,\frac{\delta^2\cos{\gamma}\sin{\gamma}}{R^2} \;\; , \label{eq.evme2c}$$ where $$\begin{aligned} B_1(\rho_{\rm v}) & = & -\frac{\phi_0}{\lambda^2}\int^{\infty}_0\,\frac{dk}{2\pi}S_2(k) k^2 J_2(k\rho_{\rm v}) e^{-kz_0}F_1(k) \nonumber \;,\\ S_2(k)&=&\frac{2J_2(kR)}{kR}\,\frac{2\sinh{(kL_z/2)}}{kL_z}\;, \label{eq.b1} \end{aligned}$$ and $\gamma$ is the angle between ${\mbox{\boldmath $\delta$}}$ and ${\mbox{\boldmath $\rho$}}_{\rm v}$ (Fig. \[fig.fig4\]). Using $\delta_L=\delta\cos{\gamma}$ and $\delta_T=\delta\sin{\gamma}$, Eq. (\[eq.evme2c\]) is identical to Eq. (\[eq.evme2\]). [99]{} For a recent review, see M.J. Van Bael, L. Van Look, M. Lange, J. Bekaert, S.J. Bending, A.N. Grigorenko, K. Temst, V.V. Moshchalkov, and Y. Bruynseraede, Physica C [**369**]{}, 97 (2002), and references therein. M.W. Coffey, Phys. Rev. B [**52**]{}, R9851 (1995). J.C. Wei, J.L. Chen, and L. Horng, T.J.Yang, Phys. Rev. B [**54**]{}, 15429 (1996). R. Šašik, and T. Hwa arXiv:cond-mat/003462 v1 (2000). M.V. Milošević, S.V. Yampolskii, and F.M. Peeters, Physica C [**369**]{}, 343 (2002); Phys. Rev. B [**66**]{}, 024515 (2002). M.V. Milošević, S.V. Yampolskii, and F.M. Peeters, Phys. Rev. B [**66**]{}, 174519 (2002). D.J. Priour, Jr, and H. A. Fertig, Phys. Rev. Lett. B [**93**]{}, 057003 (2004). G. Carneiro, Europhys. Lett.[**71**]{},817 (2005); Phys. Rev. B[**72**]{}, 144514 (2005). R.P Cowburn, D.K. Kolstov, A.O. Adeyeye, M.E. Welland, and D.M. Tricker, Phys. Rev. Lett.[**83**]{}, 1042 (1999). J. Raabe, R. Pulwey, R. Sattler, T. Schweinb$ \ddot{\rm o}$ck, J.Zweck, and D. Weiss, J. Appl. Phys.[**88**]{}, 2909 (2000). M. Schneider, H. Hoffmann, and J.Zweck, Appl. Phys. Lett.[**88**]{}, 4437 (2000). L. Landau, and E. Lifschitz, Phys. Z. Sowjet. [**8**]{}, 153 (1935); [*Collected Papers of L. D. Landau*]{} (Gordon and Breach, 1967), page 101. G. Carneiro, Phys. Rev. B [**69**]{}, 214504 (2004). N. A. Usov and S. E. Peschany, J. Magn. Magn. Mater. [**118**]{}, L290 (1993); Fiz. Met.Metalloved [**12**]{}, 13 (1994). K. Y. Guslienko, and K. L. Metlov, Phys. Rev. B [**63**]{}, 100403(R) (2001); K. Y. Guslienko, V. Novosad, Y. Otani, H. Shima, and K. Fukamichi, Appl. Phys. Lett. [**78**]{}, 3848 (2001); Phys. Rev. B [**65**]{}, 024414 (2001); W. Scholz, K. Y. Guslienko, V. Novosad, D. Suess, T.Schrefl, R.W. Chantrell, and J. Fidler, J. Magn. Magn. Mater. [**266**]{}, 155 (2003). G. Carneiro, and E.H. Brandt, Phys. Rev. B [**61**]{}, 6370 (2000). L.D. Landau and E.M. Lifshitz, [*Electrodynamics of Continuous Media*]{} (Pergamon,1960). R. P. Cowburn and M. E. Welland, Appl. Phys. Lett. [**72**]{}, 2041 (1998); S. P. Li, D. Peyrade, M. Natali, A. Lebib, Y. Chen,U. Ebels, L. D. Buda, and K. Ounadjela, Phys. Rev. Lett. [**86**]{}, 1102 (2001); J. Rothman, M. Klaui, L. Lopez-Diaz,C. A. F. Vaz, A. Bleloch, J. A. C. Bland, Z. Cui, and R. Speaks, Phys. Rev. Lett. [**86**]{} 1098 (2001); M. Klaui, J. Rothman, L. Lopez-Diaz, C. A. F. Vaz, J. A. C. Bland, and Z. Cui, Appl. Phys. Lett. [**78**]{}, 3268 (2001).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We provide a co-free construction which adds elementary structure to a primary doctrine. We show that the construction preserves comprehensions and all the logical operations which are in the starting doctrine, in the sense that it maps a first order many-sorted theory into a the same theory formulated with equality. As a corollary it forces an implicational doctrine to have an extentional entailment.' author: - Fabio Pasquali bibliography: - 'biblio.bib' title: 'A co-free construction for elementary doctrines' --- Introduction {#introduction .unnumbered} ============ This paper deals with the notion of internal equality in doctrines. Doctrines were introduced by Lawvere ([@LawDiag], [@LawAdj] and [@LawEq]) and we pospone in section \[sec1\] their formal definition. For the purpose of this introduction it is enough to think of doctrines as those presheaves such that, given a theory $\mathcal{T}$ over a many-sorted relational language $\mathcal{L}$, one looks at objects and morphism of the domain category as types and terms of $\mathcal{L}$ respectively, while a well formed formula in $\mathcal{T}$ of type $A$ is an element in the fiber over $A$. Lawvere made extensive use of the language of adjoints and Jacobs [@Jacobs] described equality between terms of a given type as a formula in the fiber over the product of that type with itself, satisfying the following rule of inference $$\AxiomC{$\Gamma, x:X \mid \phi \vdash \psi[x/y]$} \doubleLine \UnaryInfC{$\Gamma, x:X, y:X \mid \phi \wedge x=_Xy \vdash \psi$} \DisplayProof$$ where the double line indicates that one of the two sequents holds exactly when the other holds. A doctrine is a first order theory with equality if it possesses a formula $=_X$, for every sort $X$, which satisfies the previous rule.\ A way to introduce higher order quantification is to consider a new type $\Omega$ in the underlying signature and thinks of terms of type $\Omega$ as propositions. From the categorical viewpoint this generates a correspondence between terms of type $\Omega$ and formulas, and therefore it makes sense to investigate how the notion of internal equality $=_{\Omega}$ is related to logical equivalence. A link is in the following rule, taken from [@BoiJoy] and [@Jacobs] $$\AxiomC{$\Gamma \mid \xi \wedge \phi \vdash \psi$} \AxiomC{$\Gamma \mid \xi \wedge \psi \vdash \phi$} \doubleLine \BinaryInfC{$\Gamma \mid \xi \vdash \phi =_{\Omega} \psi$} \DisplayProof$$ where it is implicit that if $\phi$ and $\psi$ are formulas over the context $\Gamma$, then $\phi=_{\Omega} \psi$ is still a formula over $\Gamma$. We say that a doctrine is a higher order many-sorted theory with extentional entailment if there is an object $\Omega$ in the base category and a formula $=_\Omega$ in the fiber over $\Omega \times \Omega$ which satisfies both the previous rules.\ In the present paper we provide a co-free construction that, starting from any doctrine $P$, produces a new doctrine $P_\D$ with equality. That is to say that for every object $X$ in the domain category of $P_\D$ there exists a well formed formula in $P_\D(X\times X)$ which satisfies the first one of the previous rules. We show also that if the starting doctrine $P$ is an higher order implicational theory, the resulting doctrine $P_\D$ will have an internal equality over $\Omega$ satisfying both the previous rules; in other words: $=_{\Omega}$ and logical equivalence comes to coincides.\ In section \[sec1\] we give the definitions of doctrines and some relevant examples. In section \[sec2\] we introduce the construction of Maietti and Rosolini of the category of quotients and the doctrine of descent data which is the base of the co-free construction we are going to provide in \[sec3\]. In the last section we show which properties are preserved by the construction and some applications. Doctrines {#sec1} ========= We recall those structures which we will be concerned with in the paper, see [@RM] and [@Tripinret]. A **primary doctrine** is a functor $P: \C ^{op} \arr \textbf{ISL}$, where $\textbf{ISL}$ is the subcategory of **Posets** consisting of inf-semilattices and homomorphisms and $\C$ is a category with binary products. For the rest of the paper we will write $f^*$ instead of $P(f)$, to indicate the action of the functor $P$ on a morphism $f$ of $\C$. We shall refer to $f^*$ as the reindexing functor along $f$. Left and right adjoints to reindexing functor $f^*$ will be $\exists_f$ and $\forall_f$ respectively. We say that a doctrine has finite joins if every fiber has finite joins. Analogously we say that a doctrine is implicational if every fiber has relative pseudo complements which commute with reindexing. For every pair of element $x$ and $y$ we will denote their meet by $x\wedge y$, by $x\lor y$ their join and by $x\imply y$ their relative pseudo complements. Top and bottom elements will be $\top$ and $\bot$ respectively. Joins are said to be distributive if for every $x$, $y$ and $z$ it holds that $x \wedge (y \lor z) = (x \wedge y) \lor (x \wedge z)$. \[ele\] A primary doctrine $P$ is said to be **elementary** if for every $A$ in $\C$ there exists an object $\delta_A$ in $P(A \times A)$ such that for every $X$ in $\C$ - the assignment $\pi_1^* (\alpha) \wedge \delta_A$ determines a left adjoint to $\Delta_A^*$ - the assignment $\langle \pi_1,\pi_2\rangle^*(\alpha)\wedge \langle\pi_2,\pi_3\rangle^*(\delta_A)$ determines a left adjoint to $(id_X \times \Delta_A)^*$ Primary doctrines are the objects of the 2-category **PD** in which\ \ the **1-arrows** are pairs $(F,f): P \arr R$ $$\xymatrix{ \C^{op}\ar[rrd]^-{P}="a"\ar[dd]_{F}&&\\ &&\textbf{ISL}\\ \mathbb{D}^{op}\ar[rru]_-{R}="b"&& \ar "a";"b"_{f}}$$ where the functor $F$ preserves products and $f$ is a natural transformation from the functor $P:\C^{op}\arr \textbf{ISL}$ to the functor $R\circ F:\D^{op}\arr \textbf{ISL}$\ \ the **2-arrows** are those natural transformations $\nu$ $$\xymatrix{ \C^{op}\ar[rrrd]^-{P}="a"\ar@/_1pc/ [dd]_{F}="c"\ar@/^1pc/ [dd]^{G}="d"&&&\\ &&&\textbf{ISL}\\ \mathbb{D}^{op}\ar[rrru]_-{R}="b"&&& \ar@/_1pc/"a";"b"_{f}\ar@/^1pc/"a";"b"^{g}_{\le \ \ }\ar"c";"d"^{\nu}}$$ such that, for every object $A$ in $\C$ and every $\alpha$ in $P(A)$, it holds that $\nu_A^* (f_A (\alpha)) \le g_A (\alpha)$.\ \ We call **ED** the 2-subcategory of **PD**, in which the object are elementary doctrines and the 1-arrows are those 1-arrows in **PD** such that $$f_{A\times A}(\delta_A) = <F\pi_1,F\pi_2>^*\delta_{FA}$$ for every 1-arrows $(F,f)$ and for every object $A$ in $\C$. \[uni\] A primary doctrine is called **universal** if for every projection arrows $\pi$ in $\C$ the functor $\pi^*$ has a right adjoint $\forall _{\pi}$ satisfying Beck-Chevalley condition: given a pullback diagram of the kind $$\xymatrix{ X\times Y' \ar[d]_-{id \times f}\ar[r]^-{\pi'}& Y'\ar[d]^-{f}\\ X\times Y \ar[r]_-{\pi}& Y }$$ it holds that $\forall_{\pi'} \circ (id\times f)^* = f^* \circ \forall_{\pi}$ A primary docrine is **existential** if the reindexing functors along a projection have a left adjoint satisfying Beck-Chevalley and Frobenius reciprocity: $\exists_\pi(\alpha\wedge \pi^*\beta)=\exists_\pi(\alpha)\wedge \beta$, for $\alpha$ in $P(X\times Y)$ and $\beta$ in $P(Y)$. \[GQ\] Recall from [@LawAdj; @Tripinret] that in an elementary existential doctrine $P$ for every morphism $f:A\arr B$ in the base category there exists a functor $\exists_f : P(B)\arr P(A)$ such that $\exists_f \dashv f^{*}$. Indeed if $\pi_A$ and $\pi_B$ are the projections from $A\times B$ to $A$ and $B$ respectively, for $\alpha$ in $P(A)$ $$\exists_f (\alpha) := \exists_{\pi_B}((id_B \times f)^*\delta_B\wedge \pi_A^*\alpha)$$ Such a generalized quantification satisfies Frobenius Reciprocity. For $\beta$ in $P(B)$, we have that $(id_B\times f)^*\delta_B \wedge f^*\beta = (id_B\times f)^*\delta_B \wedge \pi_B^*\beta$. Therefore $$\begin{array}{c} \exists_{\pi_B}((id_B \times f)^*\delta_B\wedge \pi_A^*\alpha \wedge f^*\beta) =\\ \exists_{\pi_B}((id_B \times f)^*\delta_B\wedge \pi_A^*\alpha \wedge \pi_B^*\beta) =\\ \exists_{\pi_B}((id_B \times f)^*\delta_B\wedge \pi_A^*\alpha)\wedge \beta \end{array}$$ And for a pullback square such as that in \[uni\] the Beck-Chevalley condition holds: $\exists_{(id_X \times f)}\ \pi^* = \pi'^*\ \exists_f$. A primary doctrine is said to have a **weak power objects** if for every $A$ in $\C$ there exists an object $\pi A$ in $\C$ and an element $\in_A$ in $P(A\times \pi A)$ such that, for every object $B$ in $\C$ and element $\phi$ in $P(A \times B)$ there exists a morphism $\{\phi\}: B\arr \pi A$ such that $\phi=(id_A \times \{\phi\})^*\in_A$. \[p1\] In the case the base category $\C$ has a terminal object $1$: the first item in the definition \[ele\] is redundant, since it becomes a particular instance of the second; when the doctrine has weak power objects, for every object $A$ in $\C$ each element $\phi$ in $P(A)$ determines a term of type $\pi1$ via the following isomorphism: $$\xymatrix{ 1\times \pi1\ar[r]^{i}&\pi1\\ 1\times A\ar[u]^{id_1\times \{j^*\phi\}} \ar[r]_{j}& A\ar[u]_{\{\phi\}} }$$ we will denote with $\epsilon_1$ the element $(i^{-1})^*\in_1$; in the case $\C$ has all comprehensions, defined to be those morphisms $\lfloor \phi \rfloor : X \arr A$ which are terminal with respect to the property that $\top_X \le \lfloor \phi \rfloor^*(\phi)$, for every $\phi$ in $A$ (see [@RM]), then $\lfloor \phi \rfloor$ is weakly classified by $\{\phi\}$, where the true arrow is $\lfloor \epsilon_1\rfloor : 1 \arr \pi1$. There are several examples of doctrines, we list a few. \[syn\] (**Syntactic**) Given a theory $\mathcal{T}$ in a first order language $\mathcal{L}$, the base category $\mathbb{V}$ has lists of distinct variables $\vec{x}=(x_1, x_2, . . . x_n)$ as objects and lists of substitutions $[\vec{t}/\vec{y}]:\vec{x}\arr \vec{y}$ as morphisms. Composition is given by simultaneous substitution. For an object $\vec{x}$ in $\mathbb{V}$, the fiber over $\vec{x}$ consists of equivalence classes of well-formed formulae of $\mathcal{L}$ with no more free variables than $x_1, x_2, . . . x_n$, with respect to reciprocal entailment of $\mathcal{T}$, see [@RM]. \[sub\] (**Subobjects**) Suppose $\C$ a small category with binary products and pullbacks. Consider the functor that assigns for every object $A$ in $\C$ the collection $\textbf{Sub}(A)$ of subojects with codomain $A$, ordered by factorization. The top element is (the equivalence class of) the identity arrow. A representative of $\alpha \wedge \beta$ is any pullback of $\alpha$ along $\beta$. Given a morphism $f$ in $\C$, $f^*\alpha$ is the class of any pullback of $\alpha$ along $f$. If $\C$ is regular, the doctrine has left adjoints of all reindexing functors. It is elementary with $\delta_A=\Delta_A : A \arr A \times A$. $\textbf{Sub}:\C^{op}\arr \textbf{ISL}$ has full comprehensions. An element $\alpha : X\arr A$ in $\textbf{Sub} (A)$ has itself as its own comprehension. Consider the following diagrams $$\xymatrix{ X\ar[d]_-{\top_X}\ar[r]^-{\top_X}&X\ar[d]^-{\alpha}\\ X\ar[r]_-{\alpha}&A } \quad \qquad \xymatrix{ Y\ar[rd]_-{\top_Y}\ar[r]^-{k}&P\ar[r]^-{p}\ar[d]^-{f^*\alpha}&X\ar[d]^-{\alpha}\\ &Y\ar[r]_-{f}&A } \quad \qquad \xymatrix{ X\ar[rd]_-{\top_X}\ar[r]^-{h}&Q\ar[r]^-{q}\ar[d]^-{\alpha^*\beta}&X\ar[d]^-{\beta}\\ &X\ar[r]_-{\alpha}&A }$$ where $P$ is a pullback of $\alpha$ along $f$ and $Q$ the pulback of $\beta$ along $\alpha$. The left one is a pullback and says that $\top_X \le \alpha^* \alpha$. The second proves that if $\top_Y \le\ f^*\alpha$ ($\le$ is $k$), then $f$ factorizes through $\alpha$. Third pullback shows that if $\top_X \iso\ \alpha^*\alpha\ \le\ \alpha^*\beta$ ($\le$ is $h$), then $\alpha\ \le\ \beta$ ($\le$ is $q \circ h$).\ A particular case is when $\C$ is a small, full subcategory of **Set** closed under binary products and subsets and the functor $\textbf{Sub}:\C^{op}\arr \textbf{ISL}$ coincides with the powerset functor \[Tripos\](**Triposes**) We refer to the definition given by Pitts in [@Tripinret].\ Given a category $\C$ with binary products, a tripos is a primary doctrine $P: \C ^{op} \arr \textbf{ISL}$ such that: (i) for every object $A$ in $\C$, $P(A)$ is a Heyting Algebra (ii) for every arrow $f$ in $\C$, $f^*$ is an homomorphism of Heyting algebras (iii) for every projection arrow $\pi$ in $\C$ the functor $\pi^*$ has left and right adjoints satisfying the Beck-Chevalley conditions (iv) $P$ has weak power objects (v) for every object $A$ in $\C$ there exists an element $\delta_A$ in $P(A \times A)$ such that, for all $\alpha$ in $P(A\times A)$, $\top_A \le \Delta_A^*(\alpha)$ if and only if $\delta_A \le \alpha$.\ \ All triposes are universal doctrine with weak power objects. They are elementary, since the assignment $\exists _{\Delta _X} (\alpha) := \pi_1 ^* (\alpha) \wedge \delta _X$ provides a left adjoint to $\Delta _X ^*$, in fact $$\AxiomC{$\alpha \leq \Delta _X ^* (\beta)$} \UnaryInfC{$\top _ X \leq \alpha \imply \Delta _X ^* (\beta)$} \UnaryInfC{$\top _ X \leq \Delta_X ^* (\pi _1 ^* (\alpha) \imply \beta)$} \UnaryInfC{$\delta _X \leq \pi _1 ^* (\alpha) \imply \beta$} \UnaryInfC{$\delta _X \wedge \pi _1 ^* (\alpha) \leq \beta$} \DisplayProof \qquad \AxiomC{$\exists_{\Delta _X}(\alpha) \leq \beta$} \UnaryInfC{$\pi_1 ^*(\alpha) \wedge \delta_X \leq \beta$} \UnaryInfC{$\alpha \wedge \Delta_X ^*(\delta_X) \leq \Delta_X ^*(\beta)$} \UnaryInfC{$\alpha \wedge \top _X \leq \Delta_X ^*(\beta)$} \UnaryInfC{$\alpha \leq \Delta_X ^*(\beta)$} \DisplayProof$$ and the assignment $\exists _e (\alpha) := \langle \pi _1, \pi _2\rangle ^* (\alpha) \wedge \langle \pi _2, \pi _3\rangle ^*(\delta _A)$ determines a left adjoint to the reindexing of $e:=id_X \times \Delta_A : X\times A \arr X\times A\times A$ $$\AxiomC{$\exists _e (\alpha) \leq \beta$} \UnaryInfC{$e^*\exists _e (\alpha) \leq e^*(\beta)$} \UnaryInfC{$\alpha \wedge \langle\pi _2, \pi _2\rangle ^*(\delta _A) \leq e^*(\beta)$} \UnaryInfC{$\alpha \wedge \pi_2 ^* \Delta_A ^* \exists _{\Delta _A} (\top _A) \leq e^*(\beta)$} \UnaryInfC{$\alpha \wedge \pi_2 ^* (\top _A) \leq e^*(\beta)$} \UnaryInfC{$\alpha \leq e^*(\beta)$} \DisplayProof \qquad \AxiomC{$\alpha \leq e^*( \beta)$} \UnaryInfC{$\top _{X\times A} \leq \alpha \imply e^*(\beta)$} \UnaryInfC{$\pi_2 ^*\top _A \leq e^* \langle\pi_1,\pi_2\rangle^*(\alpha) \imply e^*(\beta)$} \UnaryInfC{$\top _A \leq \forall_{\pi_2}e^* (\langle\pi_1,\pi_2\rangle^*(\alpha) \imply \beta)$} \UnaryInfC{$\top _A \leq \Delta _A ^* \forall _{\langle\pi_2,\pi_3\rangle} (\langle\pi_1,\pi_2\rangle^*(\alpha) \imply \beta)$} \UnaryInfC{$\delta _A \leq \forall _{\langle\pi_2,\pi_3\rangle} (\langle\pi_1,\pi_2\rangle^*(\alpha) \imply \beta)$} \UnaryInfC{$\langle\pi_2,\pi_3\rangle^* (\delta _A) \leq \langle\pi_1,\pi_2\rangle^*(\alpha) \imply \beta$} \UnaryInfC{$\langle\pi_2,\pi_3\rangle^* (\delta _A) \wedge \langle\pi_1,\pi_2\rangle^*(\alpha) \leq \beta$} \UnaryInfC{$\exists _e (\alpha) \leq \beta$} \DisplayProof$$ Similarly it can be proved that Frobenius reciprocity is verified (see also [@VanOO], pag 60). Two important examples of triposes are $\mathbb{H}^{(-)}$, for a complete Heyting algebra $\mathbb{H}$, and $\mathbb{P}(\mathcal{N})^{(-)}$, for a partial combinatory algebra over a set $\mathcal{N}$. In each case $\C$ is $\textbf{Set}$, the category of sets and functions. There is no need for a tripos to have comprehensions. But this is the case for localic triposes $\mathbb{H}^{(-)}$ and realizability triposes $\mathbb{P}(\mathcal{N})^{(-)}$. Take a set $X$ and an object $\phi : X \arr \mathbb{H}$: a comprehension of $\phi$ is given by the inclusion $\lfloor \phi \rfloor : \{x \ \epsilon \ X\mid \top \le \phi(x)\} \hookrightarrow X$. The same holds for realizability troposes, for which $\lfloor \phi \rfloor : \{x \ \epsilon \ X\mid \mathcal{N} \subseteq \phi(x)\} \hookrightarrow X$. These comprehensions can not be full. Take $\lfloor \phi \rfloor : A \hookrightarrow X$ and consider the function $\psi: X \arr \mathbb{H}$ defined by $\psi(x) = \top$ if $x\ \epsilon \ A$ and $\bot$ otherwise. For this function certainly holds $\lfloor \phi \rfloor^* (\psi) = \top$, but it is not the case that $\phi \le \psi$. \[TOP\](**Topologies**) Consider the category $\textbf{TOP}$ of topological spaces and continuos functions. For every topological space $X$, $\mathcal{O}(X)$ is the collection of its open sets, and then it possesses finite meets and arbitrary joins. Take the functor $\mathcal{O} :\textbf{TOP}^{op} \arr \textbf{ISL}$ determined by the following assignment $$\xymatrix@R=1.5ex@C=1.8ex{ (X, \mathcal{O}(X))\ar[dd]_-{f}&&\mathcal{O}(X)\\ &\mapsto&\\ (Y,\mathcal{O}(Y))&&\mathcal{O}(Y)\ar[uu]_-{f^{-1}} }$$ Even though each fiber is an Heyting algebra, and therefore it has pseudo relative complements (see [@MM], page 51), $\mathcal{O}$ is not implicational as a doctrine: given a generic continuos function $f$, we have that pseudo relative complements need not commute with reindexing (see [@SS], page 39). $\mathcal{O}$ is existential, since every projection functor has a left adjoint (see [@MM], page 58) satisfying Beck-Chevalley condition and Frobenius reciprocity (recall that projections are open functions). $\mathcal{O}$ has full comprehensions. Given a set $X$, for any open set $S$ in $O(X)$, define its comprehension to be the inclusion function $\lfloor S \rfloor : (S, O_{S}(X)) \hookrightarrow (X, O(X))$, where $O_{S}(X)$ is the topology induced by $S$. These comprehensions are also full. Suppose $Q$ in $O(X)$ such that $\lfloor S \rfloor^{-1}(Q) = S$, this means $\{x\ \epsilon \ S \mid x\ \epsilon \ Q\} = S \cap Q = S$, so $S \subseteq Q$.\ The doctrine has weak power objects. We call $\Sigma$ the Sierpinski space consisting of two points $0$ and $1$ and a third non trivial open set $\{1\}$. If a topological space $T$ is locally compact, then there exists in **TOP** the function space $\Sigma^{T}$ (see [@TopHyland] and [@topo]). $\Sigma$ extends the subobjects classifier from **Set** to **TOP** in the sense that for every $\phi$, open set of $T$, the characteristic function of the inclusion $\lfloor \phi \rfloor$ is the unique arrow making the following a pullback $$\xymatrix{ X\ar[d]_-{\lfloor \phi \rfloor}\ar[r]^-{!}&1\ar[d]^-{\top}\\ T\ar[r]_-{\chi_{\phi}}&\Sigma }$$ for which it holds that $\chi_{\phi}^{-1}(\{1\}) = \phi$. Now for every topological space $A$ consider any construction that produces a larger locally compact space $\tilde{A}$ such that the inclusion morphism $i_A: A\hookrightarrow \tilde{A}$ is continuos and open, e.g. Alexandroff compactifications, see [@topo]; the following lemma holds: if $f:A\times B \arr \Sigma$ is continuous, then the extension $\tilde{f}: \tilde{A}\times B \arr \Sigma$ is continuous, where $\tilde{f}(a,b) = f(a,b)$ if $a\ \epsilon\ A$, then $\tilde{f}(a,b) = 0$.\ To prove the lemma it suffices to note that there are no open sets in $\Sigma$ containing the point $0$ other than the top element, then $\tilde{f}^{-1}(\{1\}) = f^{-1}(\{1\})$ and the inclusion function is open. Note that $f=\tilde{f}\circ (i_A\times id)$. Now consider the diagram $$\xymatrix{ A\times \Sigma^{\tilde{A}}\ar@{^{(}->}[rr]^-{i_A \times id} &&\tilde{A}\times \Sigma^{\tilde{A}}\ar[r]^-{ev_{\tilde{A}}}&\Sigma\\ A\times B\ar[u]^-{id_{A}\times \overline{\tilde{\chi}_{\phi}}}\ar@{^{(}->}[rr]_-{i_A \times id} && \tilde{A}\times B\ar[u]^-{id_{\tilde{A}}\times \overline{\tilde{\chi}_{\phi}}} \ar[ru]_-{\tilde{\chi}_{\phi}} & }$$ define $\in_A := (ev_{\tilde{A}} \circ (i_A \times id))^{-1}(\{1\})$ and for every open set $\phi$ in $A\times B$ define $\{\phi\} := \overline{\tilde{\chi}_{\phi}}$ the exponential transpose of the extension of $\chi_{\phi}$. $(id_A \times \overline{\tilde{\chi}_{\phi}})^{-1}(\in_A) = \chi_{\phi}^{-1}(\{1\}) = \phi$.\ \ The doctrine fails to be elementary. Given a topological space $X$, we have that $\delta_X$ should be the smallest open set $U$ of $X \times X$ such that $X \subseteq \Delta^{-1}(U)$. In other words $$\delta_X = (\bigcap_{X\subseteq \Delta_X^{-1}(U)}U)^{o}$$ if $X$ is the interval $[0,1]$ with the euclidean topology, then $\delta_X$ would be empty. Quotients and descents {#sec2} ====================== Recall a construction presented in [@RM2; @RM], which is based on the notion of equivalence relation in a doctrine. \[equiv\] Given a primary doctrine $P:\C^{op}\arr \textbf{ISL}$ and an object $A$ of $\C$, an element $\rho$ in $P(A\times A)$ is said to be an **equivalence relation** on $A$ if\ \ reflexivity: $\top_A \le \Delta_A^* (\rho)$\ \ symmetry: $\rho \le \langle \pi_1, \pi_2 \rangle^*(\rho)$\ \ transitivity: $\langle \pi_1, \pi_2 \rangle^*(\rho)\wedge \langle \pi_2, \pi_3 \rangle^*(\rho) \le \langle \pi_1, \pi_3 \rangle^*(\rho)$ Note that if the doctrine $P$ is also elementary, then $\delta_A$ is an equivalence relation on $A$ for every object $A$ in $\C$.\ \ In [@RM2; @RM] the authors consider a certain category $\mathcal{Q}_P$, when $P:\C^{op}\arr \textbf{ISL}$ is elementary. In the category $\Q_P$\ \ **objects** are pairs $(A,\rho)$ such that $\rho$ is an equivalence relation on $A$\ \ **morphisms** $f:(A,\rho)\arr(B,\sigma)$ are arrows $f:A\arr B$ in $\C$ such that $\rho \le (f\times f)^*\sigma$\ \ and composition is given as in $\C$.\ \ A first remark is that the construction gives a category in the more general case of $P$ primary. The category $\mathcal{Q}_P$ has binary products: given $(A,\rho)$ and $(B,\sigma)$ in $\mathcal{Q}_P$, $(A,\rho)\times(B,\sigma)\colon =(A\times B, \rho\boxtimes \sigma)$, where $\rho \boxtimes \sigma$ is $\langle\pi_1,\pi_3\rangle^*\rho \wedge\ \langle\pi_2,\pi_4\rangle^*\sigma$. Moreover if $\C$ has a terminal object, $\mathcal{Q}_P$ has a terminal object.\ \ There is an obvious forgetful functor $\mathbb{U}:\Q_P\arr \C$, and a functor $\nabla: \C \arr \Q_P$, determined by the following assignments $$\xymatrix@R=1.5ex@C=1.8ex{ &(A, \rho)\ar[dd]_-{f}&&A\ar[dd]^-{f}\\ (\mathbb{U})&&\mapsto&\\ &(B, \sigma)&&B } \qquad\ \quad \ \qquad \xymatrix@R=1.5ex@C=1.8ex{ &A\ar[dd]_-{f}&&(A, \delta_A)\ar[dd]^-{f}\\ (\nabla)&&\mapsto&\\ &B&&(B, \delta_B) }$$ $\nabla$ is clearly a functor since, for every morphism $f$ in $\C$, $\delta_A \le (f\times f)^* \delta_B$. \[left\] Given an elementary doctrine $P:\C^{op} \arr \textbf{ISL}$, the functor $\nabla$ is left adjoint to $\mathbb{U}$. For every object $(B,\sigma)$ in $\Q_P$, the map $\varepsilon_B \colon = id_B : (B,\delta_B)\arr (B,\sigma)$ is the $B$-component of a natural transformation. This is the counite of the adjunction, since for every object $A$ in $\C$ and every arrow $f:(A,\delta_A)\arr (B,\sigma)$ in $\Q_P$ the diagram commutes $$\xymatrix{ (B,\delta_B)\ar[r]^{id_B}&(B,\sigma)\\ (A,\delta_A)\ar[u]^{f}\ar[ru]_{f}& }$$ and $f$ is the unique such arrow. Given a primary doctrine $P:\C^{op}\arr \textbf{ISL}$ and an equivalence relation $\rho$ on an object $A$ of $\C$, the poset of descent data $\mathcal{D}es_{\rho}$ is the sub-order of $P(A)$ made by those $\alpha$ such that $$\pi_1^*(\alpha)\ \wedge\ \rho\ \le\ \pi_2^*(\alpha)$$ The order $\mathcal{D}es_{\rho}$ is closed under meets and it has trivially $\top_A$, then $\mathcal{D}es_{\rho}$ is an inf-semilattice.\ The following proposition generalizes to primary doctrines a similar result given for elementary doctrine in [@RM2; @RM]. \[lefto\] Given a primary doctrine $P:\C^{op}\arr \textbf{ISL}$, the assignment $$\xymatrix@R=1.5ex@C=1.8ex{ (A,\rho)\ar[dd]_{f}&&\mathcal{D}es_{\rho}\\ &\mapsto&\\ (B,\sigma)&&\mathcal{D}es_{\sigma}\ar[uu]_{f^*} }$$ determines a primary doctrine $P_{\mathcal{D}}:\mathcal{Q}_P^{op}\arr \textbf{ISL}$. It suffices to note that, for every $\beta$ in $\D es_{\sigma}$, $f^* \beta$ is in $\D es_{\rho}$, that can be proved by taking the descent condition on $\beta$, applying to both sides $(f\times f)^*$ and use the fact that $\rho\ \le\ (f\times f)^*\sigma$. A co-free construction {#sec3} ====================== There is an obvious forgetful functor $\mathcal{U}: \textbf{ED} \arr \textbf{PD}$, which maps every elementary doctrine to itself. We shall show that the construction in \[sec2\] extends to a 2-right adjoint to it.\ \ The following lemma is a strengthening of a similar result in [@RM2]. \[elem\] Given a primary doctrine $P:\C^{op}\arr \textbf{ISL}$, the doctrine $P_{\mathcal{D}}:\mathcal{Q}_P^{op}\arr \textbf{ISL}$ built as in \[lefto\] is elementary. Consider $(A, \rho)$ in $\Q_P$. Note that $\rho$ is an element of $\mathcal{D}es_{\rho \boxtimes\rho}$, since $$\pi_1^*\rho \wedge (\rho\boxtimes\rho) = \langle \pi_1,\pi_2\rangle^*\rho \wedge\ \langle\pi_1,\pi_3\rangle^*\rho\wedge\ \langle\pi_2,\pi_4\rangle^*\rho$$ and by transitivity of $\rho$ $$\pi_1^*\rho \wedge \rho\boxtimes\rho \ \le \langle\pi_3,\pi_4\rangle^*\rho \ = \ \pi_2^*\rho$$ Let $\delta_{(A,\rho)}$ be $\rho$ and define $\exists_{\Delta_A} \alpha \colon = \pi_1^*\alpha\ \wedge \rho$. We want to prove that, for every $\alpha$ in $\D es_{\rho}$ and $\beta$ in $\D es_{\rho\boxtimes\rho}$, $\exists_{\Delta_A} \alpha \le \beta$ if and only if $\alpha \le \Delta_A^*\beta$. Suppose $\exists_{\Delta_A} \alpha \le \beta$, which means $\pi_1^*(\alpha)\ \wedge \rho\ \le \beta$, and apply $\Delta_A^*$ to both sides, to obtain $\alpha \wedge \Delta_A^*\rho\ \le \Delta_A^*\beta$. So $\alpha \le \Delta_A^*\beta$, by reflexivity of $\rho$. Assume now $\alpha \le \Delta_A^*\beta$, the descent condition for $\beta$ gives: $$\langle\pi_1,\pi_2\rangle^*\beta\wedge\langle\pi_1,\pi_3\rangle^*\rho\wedge\ \langle\pi_2,\pi_4\rangle^*\rho\ \le\ \langle\pi_3,\pi_4\rangle^*\beta$$ By reindexing along $(\Delta_A\times id_A\times id_A)^*$ and $(\Delta_A\times id_A)^*$ one obtains $$\pi_1^* \Delta_A^* \beta \ \wedge \rho\ \le \beta$$ by reflexivity of $\rho$ $$\AxiomC{$\alpha \le \Delta_A^* \beta$} \UnaryInfC{$\pi_1^*\alpha \le \pi_1^*\Delta_A^* \beta$} \UnaryInfC{$\pi_1^*\alpha \ \wedge \ \rho \ \le \ \pi_1^*\Delta_A^* \beta \ \wedge \ \rho$} \AxiomC{$\pi_1^* \Delta_A^* \beta \ \wedge \rho\ \le \beta$} \BinaryInfC{$\pi_1^*\alpha \ \wedge \ \rho \ \le \ \beta$} \UnaryInfC{$\exists_{\Delta_A}\alpha \ \le \ \beta$} \DisplayProof$$ To verify the conditions ii) of \[ele\], consider an object $(X, \tau)$ and let $e\colon = id_X \times \Delta_A$ be a morphism in $\mathcal{Q}_P$. The proof that if $\exists_e(\alpha) \le \beta$, then $\alpha \le e^*(\beta)$ is similar to that in example \[Tripos\] (where $\rho$ is $\delta_A$). The proof of the converse, is essentially as before where: $$\langle\pi_1,\pi_2,\pi_3,\rangle^*\beta \ \wedge\ \langle\pi_1,\pi_4\rangle^*\tau \ \wedge \ \langle \pi_2,\pi_5\rangle^*\rho \ \wedge \ \langle\pi_3,\pi_6\rangle^*\rho\ \le \ \langle\pi_4,\pi_5,\pi_6\rangle^* \beta$$ and reindexing along the following composition $$\xymatrix{ X\times A\times A\ar[d]^-{id_X\times \Delta_A \times id_A}&&X\times A\times A\times X\times A\times A\\ X\times A\times A \times A\ar@/_1.5pc/[dr]^-{\ \ \ \ \ \Delta_X \times id_A\times id_A \times id_A}&&X\times A\times X\times A \times A\ar[u]^-{id_X \times \Delta_A \times id_X \times id_A \times id_A}\\ &X\times X\times A\times A \times A\ar@/_1.5pc/[ru]^-{id_X \times tw \times id_A \times id_A\ \ \ \ \ \ \ }& }$$ Given a 1-morphism in $\PD$, $(F,f) : P \arr R$, consider the functor $F_\mathcal{D}$ defined by the following assignment $$\xymatrix@R=1.5ex@C=1.8ex{ (A,\rho)\ar[dd]_{q}&&(FA, <\pi_1,\pi_2>^*f_{A \times A} (\rho))\ar[dd]^{Fq}\\ &\mapsto&\\ (B,\sigma)&&(FB, <\pi_1,\pi_2>^*f_{B \times B} (\sigma)) }$$ and the $\Q_P$-indexed family of arrow $f_\mathcal{D}$ whose $(A, \rho)$-component is the restriction of $f_A: P(A)\arr R(FA)$ to $\mathcal{D}es_{\rho}$ \[elemor\] Given a 1-morphism in **PD**, $(F,f) : P \arr R$ the pair $(F_\mathcal{D},f_\mathcal{D}):P_\mathcal{D} \arr R_\mathcal{D}$ determines a 1-morphism in **ED**. First note that $<\pi_1,\pi_2>^*f_{A \times A} (\rho)$ is an equivalence relation since $\rho$ is and $f$ is natural. $Fq$ is a morphism in $\Q_P$, since $<\pi_1,\pi_2>^*f_{A \times A} (\rho)\le(Fq\times Fq)^*<\pi_1,\pi_2>^*f_{B \times B} (\sigma)= <\pi_1,\pi_2>^*F(q\times q)^*f_{B \times B} (\sigma)= <\pi_1,\pi_2>^*f_{A \times A} (q\times q^*\sigma)$, for naturality of $f$. It is left to show that the images of the restriction is $\mathcal{D}es_{<\pi_1,\pi_2>^*f_{A\times A}(\rho)}$, but this is true since, for $\alpha$ in $\D es_{\rho}$, $\pi_1^* \alpha\ \wedge \rho\ \le \pi_2^* \alpha$, then apply $f_{A\times A}$ to both sides and, recalling that $f_{A\times A}\circ \pi_1^* = \pi_1^* \circ f_A$ for naturality of $f$, one has $\pi_1^*f_A^* \alpha\ \wedge\ f_{A\times A}(\rho)\ \le\ \pi_2^*f_A^* \alpha$. Now it suffices to reindex both sides along $<\pi_1,\pi_2>$. The last step is to show that $f_\D$ preserves the elementary structure, i.e. $f_{\D(A,\rho)\times(A,\rho)}(\delta_{(A,\rho)})=<F_\D \pi_1,F_\D \pi_2>^*(\delta_{F_\D(A,\rho)})$, which reduces to the following equality $f_{A\times A}(\rho)=<F \pi_1,F \pi_2>^*(<\pi_1,\pi_2>^*f_{A \times A} (\rho))$, where $<\pi_1,\pi_2>\circ <F \pi_1,F \pi_2> = id_{F(A\times A)}$. Consider the functor $(-)_\mathcal{D}:\PD\arr \ED$ $$\xymatrix@R=1.5ex@C=1.8ex{ P\ar[dd]_-{(F,f)}&&P_\D \ar[dd]^-{(F_\D,f_\D)}\\ &\mapsto&\\ R&&R_\D }$$ For every doctrine $P: \C\arr\textbf{ISL}$ in **PD** there is a 1-morphism $\varepsilon_P$ from $P_\D$ to $P$ given by the pair $(\mathbb{U}, i)$, where $\mathbb{U}:\Q_P \arr \mathbb{C}$ is the forgetful functor defined before \[left\], while the $A$-component of $i$ is the inclusion functor $\D es_{\rho} \hookrightarrow P(A)$. The natural transformation $\varepsilon$ is the counit of an adjunction $\mathcal{U}\dashv (-)_\mathcal{D}$. Note that $\mathcal{U}(P ) = P$; given an elementary doctrine $P:\C^{op}\arr \textbf{ISL}$, a morphism $(F,f): \mathcal{U}(P ) \arr R$ in $\PD$, consider the arrow $(\overline{F}, \overline{f}): P \arr R_\D$ in $\ED$, determined by the following composition $$\xymatrix{ \C^{op}\ar@/^1pc/[rrrd]^-{P}="a"\ar[d]_-{\nabla}&&&\\ \Q_P^{op}\ar[rrr]_-{P_\D}="b"\ar[d]_-{F_\D}&&&\textbf{ISL}\\ \Q_R^{op}\ar@/_1pc/[rrru]_-{R_\D}="c"&&& \ar"a";"b"^{id_{PA}}\ar"b";"c"^-{f_\D}}$$ then $\overline{F} := F_\D \circ \nabla$ and $\overline{f} := f_\D \circ id_{PA}$. Where the natural transformation $P \arr P_\D \circ \nabla$ is the identity from the fact that $\D es_{\delta_A} = P(A)$. What is left to prove is that $(\overline{F},\overline{f})$ is the unique arrow that makes the following diagram commutes $$\xymatrix{ \Q_R^{op}\ar[rd]^-{R_\D}\ar@/^/[rr]^-{(\mathbb{U},i)}&&\mathbb{D}^{op}\ar[dl]^-{R}\\ &\textbf{ISL}&\\ &\C^{op}\ar[u]^-{P}\ar@/_1pc/[ruu]_-{(F,f)}\ar@/^1pc/[uul]^{(\overline{F},\overline{f})}& }$$ Commutativity: recall that, for an object $A$ in $\C$, $\mathbb{U}(\overline{F})(A)$ is $\mathbb{U}(F_\D(\nabla (A)))$, then follow the assignments below $$A\mapsto (A,\delta_A)\mapsto (FA, \delta_{FA}) \mapsto FA$$ moreover $(i \circ \overline{f})_A$ is $i_A \circ f_{\D A} \circ id_{PA}$, then take $\alpha$ in $P(A)$ and follow the assignments $$\alpha \mapsto f_A (\alpha) \mapsto i(f_A(\alpha))=f_A(\alpha)$$ Uniqueness is given by the fact that $(\mathbb{U}, i)$ is mono, since $\mathbb{U}$ is the identity on objects and morphism and $i$ is an inclusion functor. Applications {#sec4} ============ The co-free construction presented in the previous section preserves all the first order predicate structures which are in $P$ in the sense of the following \[des\] Given a primary doctrine $P:\C^{op} \arr \textbf{ISL}$ and the elementary doctrine $P_\D:\Q_P^{op}\arr\textbf{ISL}$ - if $P$ has finite distributive joins, so has $P_\D$ and $\varepsilon_P : P_\D \arr P$ preserves them - if $P$ is implicational, so is $P_\D$ and $\varepsilon_P$ preserves this - if $P$ existential, so is $P_\D$ and $\varepsilon_P$ preserves this - if $P$ universal, so is $P_\D$ and $\varepsilon_P$ preserves this - if $P$ is has (full) comprehensions, so has $P_\D$ and $\varepsilon_P$ preserves them \(i) Given $\alpha$ and $\beta$ in $P_\D(A,\rho)$, the join $\alpha \lor \beta$ in $P(A)$ is in $\mathcal{D}es_\rho$ by distributivity. (ii) Like before, given $\alpha$ and $\beta$ in $P_\D(A,\rho)$, take $\alpha \imply \beta$ in $P(A)$. To see this is in $P_\D(A,\rho)$, recall that, since $\rho$ is symmetric, the descent condition can be written as $\pi_2^* \alpha \wedge \rho = \pi_1^*\alpha \wedge \rho$. One has that $\pi_1^*(\alpha \imply \beta) \wedge \rho \le \pi_2^*(\alpha \imply \beta)$ if and only if $\pi_1^*(\alpha \imply \beta) \wedge \rho \wedge \pi_2^*\alpha \le \pi_2^*\beta$ if and only if $\pi_1^*(\alpha \imply \beta) \wedge \pi_1^*\alpha \wedge \rho \le \pi_2^*\beta$. (iii) For $\alpha$ in $P_\D(A\times B, \rho \boxtimes \sigma)$, we have $\pi_1^*\exists_{\pi1} (\alpha) \wedge \rho = \exists_{\langle\pi_1,\pi_2\rangle}<\pi_1,\pi_3>^*(\alpha) \wedge \rho$ by Beck-Chevalley. By Frobenius Reciprocity that is equal to $\exists_{\langle\pi_1,\pi_2\rangle}(\langle\pi_1,\pi_3\rangle^*\alpha \wedge \langle\pi_1,\pi_2\rangle^*\rho) \le \exists_{\langle\pi_1,\pi_2\rangle}\langle\pi_2,\pi_3\rangle^*\alpha = \pi_2^*\exists_{\pi1}\alpha$; (iv) we have that $\pi_1^*\forall_{\pi1} (\alpha) \wedge \rho \le \pi_2^*\forall_{\pi1} (\alpha)$ if and only if $\forall_{<\pi_1,\pi_2>} <\pi_1,\pi_3>^*(\alpha) \wedge \rho \le \forall_{<\pi_1,\pi_2>} <\pi_2,\pi_3>^*(\alpha)$. Since $<\pi_1, \pi_2>^* \dashv \forall_{<\pi_1,\pi_2>} $ the inequality holds if and only if $<\pi_1,\pi_3>^* \alpha \wedge <\pi_1,\pi_2>^*\rho \le <\pi_2,\pi_3>^*\alpha$ which is the descent condition for $\alpha$. (v) Take an element $\alpha$ in $P_\D(A,\rho)$, this is also in $P(A)$, and consider its comprehension $\lfloor\alpha \rfloor: X\arr A$, this produces a comprehension morphism $(X, (\lfloor\alpha \rfloor \times \lfloor\alpha \rfloor)^*\rho)\arr(A,\rho)$ in $\Q_P$. Fullness directly derives from that in $P$. In each case (i)-(iv) we shall show that $P_\D(A, \rho) = \mathcal{D}es_\rho \subseteq P(A)$ is closed under the relevant constructions, thus obtaining immediately preservation by $\varepsilon_P$. In Example \[TOP\] we presented a doctrine that fails to be implicational since, even though every fiber has pseudo relative complements, they do not distribute under reindexing. Moreover the doctrine is not universal: it has right adjoints along all the projections, but these do not satisfied Beck-Chevalley conditions. The next two propositions show that these two properties are gained with the co-free construction. The first is from [@LawDiag] and the second is standard. \[asaf\] If $P:\C^{op} \arr \textbf{ISL}$ is an elementary existential doctrine and every fiber has pseudo relative complements, then $P$ is implicational. Suppose $f:A\arr B$ is a morphism in $\C$, by \[GQ\] there exists $\exists_f :P(A)\arr P(B)$ statisfying Frobenius Reciprocity. $$\AxiomC{$(f^*\alpha \imply f^*\beta) \wedge f^*\alpha \le f^*\beta$} \UnaryInfC{$\exists_f((f^*\alpha \imply f^*\beta) \wedge f^*\alpha) \le \beta$} \UnaryInfC{$\exists_f(f^*\alpha \imply f^*\beta) \wedge \alpha \le \beta$} \UnaryInfC{$\exists_f(f^*\alpha \imply f^*\beta) \le \alpha \imply \beta$} \UnaryInfC{$f^*\alpha \imply f^*\beta \le f^*(\alpha \imply \beta)$} \DisplayProof$$ To prove that $f^*(\alpha \imply \beta) \le f^*\alpha \imply f^*\beta $ it suffices to use the distributivity of reindexing functors on meets. \[avidian\] If $P:\C^{op} \arr \textbf{ISL}$ is an existential elementary doctrine with right adjoints $\forall_{\pi}$ along every projection $\pi$, then $P$ is universal. As a corollary of \[des\] and \[asaf\], we have that if $P$ is an existential doctrine and every fiber has pseudo relative complements, then $P_\D$ is implicational. And, as a corollary of \[des\] and \[avidian\], if $P$ is an existential doctrine with right adjoints along every projections, then $P_\D$ is universal. In particular the doctrine $\mathcal{O}_\D$ is implicational and universal.\ \ Power objects are not preserved, but it holds that \[h1\] If $P:\C^{op}\arr \textbf{ISL}$ is universal and implicational with weak power objects, then $P_\D$ has weak power objects. A weak power object of $(A, \rho)$ in $\mathcal{Q}_P$ is $$(\pi A, \forall_{<\pi_2,\pi_3>}(<\pi_1,\pi_2>^* \in_A \sse <\pi_1,\pi_3>^* \in_A ))$$ where the membership predicate $\in_{(A,\rho)}$ is $$\in_A \wedge\ \forall_{<\pi_1, \pi_3>}(<\pi_1, \pi_2>^*\rho \imply <\pi_2, \pi_3>^* \in_A)$$. It is worth to remark that power objects as defined in \[h1\] are still weak, but they gain the property that, in $\Q_P$ if two morphisms $\{\phi\}$ and $\{\phi\}'$ classify the same element $\phi$ in the fiber over $(A,\rho) \times (B,\sigma)$, then it holds that $\top_B \le <\{\phi\},\{\phi\}'>^*\delta_{\pi(A,\rho)}$. This lead to introduce internal extentionality. We said that for an object $A$ in the base category of an elementary doctrine, $\delta_A$ provides a notion of internal equality for terms of type $A$. Certainly external equality implies internal, in the sense that given $t_1,t_2:X\arr A$, if it holds that $t_1=t_2$ (i.e they are the same morphism in $\C$) then $\top_X \le <t_1,t_2>^*\delta_A$. The converse can be forced considering the category $[\C]$, whose objects are the same as in $\C$ and the morphism are equivalence classes of morphism of $\C$ with respect to the relation: $[t_1]=[t_2]$ if and only if $\top_X\ \le <t_1,t_2>^*\delta_A$. This construction is given directly in [@RM2; @RM], and named extentional collapse of $\C$.\ \ If $P$ is an elementary doctrine with power objects in which the base category $\C$ has a terminal object $1$, then for every object $A$ in $\C$, every element $\phi$ determines (at least) a term of type $\pi1$, i.e. $\{\phi\}: A\arr \pi1$ (see \[p1\]). Hence we can use this correspondence to define a notion of internal equality for formulas $$\phi \leftrightarrow \psi\ :=\langle\{\phi\},\{\psi\}\rangle^*\delta_{\pi1}$$ which depends on a choice of the morphisms $\{\phi\}$ and $\{\psi\}$ and satisfies the following rule $$\AxiomC{$\gamma \le \phi \leftrightarrow \psi$} \UnaryInfC{$\gamma \wedge \psi \le \phi$\ \quad \ $\gamma \wedge \phi \le \psi$} \DisplayProof$$ by the fact that we have $\pi_1^*x\ \wedge\ \delta_{\pi1} \le\ \pi_2^*x$, for every $x$ in $P(\pi1)$, then reindex both sides along $\langle\{\phi\},\{\psi\}\rangle$ with $x = \epsilon_1$ to have $\phi \wedge (\phi \leftrightarrow \psi) \le \psi$, which we use in the following tree $$\AxiomC{$\gamma \le \phi \leftrightarrow \psi$} \UnaryInfC{$\gamma \wedge \phi \le (\phi \leftrightarrow \psi) \wedge \phi$} \UnaryInfC{$\gamma \wedge \phi \le \psi$} \DisplayProof$$ The converse of the previous rule does not holds in general. This motivates the following Given a primary doctrine $P:\C^{op}\arr \textbf{ISL}$ in which the base category has a terminal object $1$, a weak power object $\pi1$ and an elementary structure $\delta_{\pi1}$ in the poset over $\pi1\times \pi1$, we say that $P$ has **extentional entailment** if, for every object $A$ in $\C$ and every element $\phi$, $\psi$ and $\gamma$ in $P(A)$ the following rule $$\AxiomC{$\gamma \wedge \psi \le \phi$} \AxiomC{$\gamma \wedge \phi \le \psi$} \doubleLine \BinaryInfC{$\gamma \le <\{\phi\},\{\psi\}>^*\delta_{\pi1}$} \DisplayProof$$ is satisfied As an immediate property we have that in an elementary doctrine $P:\C^{op}\arr\textbf{ISL}$ with extentional entailment, for every formulas $\phi$ and $\psi$, it holds that $\phi \leftrightarrow \psi$ if and only if $\top \le <\{\phi\},\{\psi\}>^*\delta_{\pi1}$, which means that every classifying morphism is unique in the exentional collapse of $\C$. \[imply\] There is a connection between extentional entailment and the presence of pseudo relative complements in every fibre of a doctrine. If a doctrine $P$ has extentional entailment, then for every object $A$ in $\C$, $P(A)$ has pseudo relative complements: it suffices to define $\alpha \imply \beta\ := (\alpha \wedge \beta)\leftrightarrow\alpha$, in the spirit of logic of toposes (see [@BoiJoy]). If $P$ is elementary, then is also implicational by \[asaf\]. The converse need not to be true in the sense that, even if an elementary doctrine has pseudo relative complements over each fiber, we have that $\pi_1^*\epsilon_1\sse\pi_2^*\epsilon_1$, may not be the left adjoint to $\Delta_{\pi1}^*$, as we see, for instance, in example \[extri\]. (**Subobjects**) Let $\C$ be a finitely complete small category. The doctrine $\textbf{Sub}:\C^{op}\arr\textbf{ISL}$ has extentional entailment if and only if $\C$ has a subobjects classifier. Let $\Omega$ be the subobjects classifier of $\C$. Then $\pi1$ is $\Omega$. $\epsilon_1$ is the true arrow. To prove the converse, suppose **Sub** to have extentional entailment. Define $\Omega$ to be $\pi1$. The true arrow is $\epsilon_1: 1\arr\pi1$. Every mono $\phi:X\arr A$ is classified by $\{\phi\}$, since reindexing is given by pullbacks. $\{\phi\}$ is unique because of extentionality of entailment, which says that if $\phi = f^*\epsilon_1$ for some $f$, then $\top \le <\{\phi\}, f>^*\delta_{\Omega} $, where $\delta_{\Omega}$ is $\Delta_{\Omega}$. Under the same conditions, an immediate corollary is that $\C$ is an elementary topos if and only if **Sub** has power objects. Suppose **Sub** to have power objects. For each $A$ in $\C$, $\Omega^{A}$ is $\pi A$ and $ev_A$ is $\{\in_A\}$. For every morphism $f : A\times B\arr \Omega$, the transpose $\overline{f}$ is $\{f^*\epsilon_1\}$. Then $\{\in_A\} \circ (id_A\times\{f^*\epsilon_1\})=f$, since they both classify $f^*\epsilon_1$. The converse is proved in $\cite{Jacobs}$, page 336. (**Triposes**)\[extri\] In general a tripos need not have extentional entailment. Take the localic tripos. $1=\{*\}$. $\delta_{\pi_1}:\mathbb{H}^1\times \mathbb{H}^1 \arr \mathbb{H}$ is given by the following assignment $(f,g)\mapsto \top$ if $f=g$, then $\bot$. But $f(*)\sse g(*)$ is not necesserly $\bot$ if $f \not = g$. Analogously for realizability triposes, see [@Jacobs] page 331. In \[imply\] and \[extri\] we showed that an implicational doctrine need not have an extentional entailment. The following proposition says that this holds once the elementary structure is co-freely added to a doctrine. In other words given a doctrine $P$, the canonical inequality $\delta_{\pi1} \le \pi_1^*\epsilon_1 \sse \pi_2^*\epsilon_1$ is an equality in $P_\D$. \[h2\] If $P:\C^{op}\arr \textbf{ISL}$ is such that $\C$ has a terminal object $1$ with a weak power object $\pi1$ and pseudo relative complements in $P(\pi1)$, then $P_\D:\Q_P\arr\textbf{ISL}$ has extentional entailment. Recalling that in $\Q_P$ the terminal object is $(1, \top_{1\times1})$, define $\pi(1,\top_{1\times1})\ :=(\pi1,\pi_1^*\epsilon_1\sse\pi_2^*\epsilon_1)$. $\epsilon_1$ certainly belongs to the category of descent data, since $\pi_1^*\epsilon_1 \wedge (\pi_1^*\epsilon_1\sse \pi_2^*\epsilon_1) \le \pi_2^*\epsilon_1$. To prove that $P_\D$ has extentional entailment it is left to show that $\pi_1^*\epsilon_1\sse\pi_2^*\epsilon_1$ is $\delta_{\pi(1,\top_{1\times1})}$, which is true by proposition \[elem\], . As a final remark note that, under the hypothesis of proposition \[h1\], the elementary doctrine $P_\D$ is a tripos, since it is possible to define finite joins and existential quantifications on the basis of implicational operations and (higher order) universal quantifications (See [@Tripos], [@Jacobs] and [@McLarty]). Then a doctrine that differs from a tripos only by the lack of an elementary structure, thanks to propositions \[des\] and \[h1\] comes to be a tripos and this tripos has extentional entailment by \[h2\]. On the other hand any tripos $P$, which is known to be an interpretation of higher order many-sorted non-extentional predicate logic (see [@Jacobs] or [@VanOO]), generetes a new tripos $P_\D$ which interpretes higher order many-sorted predicate logic with extentional entailment.
{ "pile_set_name": "ArXiv" }
--- address: - '$^{\dag}$CREST, Department of Physics, Tohoku University, Sendai 980-8578, Japan' - '$^{\ddag}$Neutron Scattering Laboratory, I.S.S.P., University of Tokyo, Tokai, Ibaraki, 319-1106, Japan' - '$^{\P}$Center for Integrated Research in Science and Engineering, Nagoya University, Nagoya, 464-01, Japan' author: - 'H. Fujioka,$^{\dag}$ M. Kubota,$^{\ddag}$ K. Hirota,$^{\dag}$ H. Yoshizawa,$^{\ddag}$ Y. Moritomo$^{\P}$ and Y. Endoh,$^{\dag}$' date: 'September 23, 1998' title: 'Spin Dynamical Properties of the Layered Perovskite La$_{1.2}$Sr$_{1.8}$Mn$_{2}$O$_{7}$' --- BoxedEPS Introduction ============ The layered perovskite Mn oxide La$_{2-2x}$Sr$_{1+2x}$Mn$_{2}$O$_{7}$ (LSMO327), in which MnO$_{2}$ double layers and (La,Sr)$_{2}$O$_{2}$ blocking layers are stacked alternatively, attracts much attention as another class of colossal magnetoresistance (CMR) system. Possibly due to the reduced dimensionality, this system exhibits an extremely large MR at the hole concentration $x=0.4$[@Y.Moritomo_96] and a tunneling MR phenomenon at $x=0.3$[@T.Kimura_96] around $T_{C}$. Neuron-scattering study on LSMO327 single crystals ($x=0.40-0.48$) by Hirota [*et al.*]{}[@K.Hirota_98] has revealed that the low-temperature magnetic structure consists of planar ferromagnetic (FM) and A-type antiferromagnetic (AFM) components, indicating a canted AF structure, where the canting angle between neighboring planes changes from 6.3$^{\circ}$ at $x=0.40$ (nearly planner FM) to 180$^{\circ}$ at $x=0.48$ (A-type AF).[@K.Hirota_98] The existence of the canted AFM structure is consistent with previous studies focusing upon the structural properties.[@J.F.Mitchell_96; @D.N.Argyriou_97] Kubota [*et al.*]{}[@M.Kubota_98] carried out a comprehensive powder neutron-scattering work and established the magnetic phase diagram for $x=0.30-0.50$; there is a planar FM phase between $x=0.32$ and $0.38$, which is smoothly connected to the canted AFM region. To understand the magnetic properties of LSMO327 in more detail, it is necessary to study the excitation spectra, from which one can determine magnetic interaction strengths and spin-spin correlation lengths. Theoretical Model ================= Figure \[Fig:Structure\] shows the magnetic spin arrangement on Mn ions in the tetragonal $I4/mmm$ cell of La$_{1.2}$Sr$_{1.8}$Mn$_{2}$O$_{7}$. Although there is a small canting between neighboring layers within a double-layer at $x=0.40$, we assume a simple planar ferromagnet, which is sufficient to consider the magnetic interactions between $x= \leq 0.40$. We expect that the dominant spin-spin interactions should occur between nearest-neighbor (NN) Mn atoms, though the in-plane interaction $J_{\parallel}$ and the [*intra*]{}-bilayer interaction $J_{\perp}$ might be different. Although there is no super-exchange coupling between layers belonging to different double-layers, there will be the [*inter*]{}-bilayer interaction $J'$ through a direct exchange. However, it is supposed to be much weaker than $J_{\parallel}$ and $J_{\perp}$, thus we neglect it in our simple model calculation. Let $l=A,B,C,D$ label the four different layers as indicated in Fig. \[Fig:Structure\]. The spin Hamiltonian can then be written in the Heisenberg form as $$\begin{aligned} {\cal H} & = & \frac{1}{2}\sum_{i}\sum_{l}\sum_{\delta} J_{il\delta}{\bf S}_{i}^{l}\cdot {\bf S}_{\delta}^{l} \nonumber \\ & = & \frac{1}{2}\sum_{i}\sum_{\delta}\left[ {\bf S}_{i}^{A}\left\{J_{\parallel}{\bf S}_{\delta}^{A}+J_{\perp}{\bf S}_{\delta}^{B}\right\}+ {\bf S}_{i}^{B}\left\{J_{\parallel}{\bf S}_{\delta}^{B}+J_{\perp}{\bf S}_{\delta}^{A}\right\}\right.\nonumber \\ & &\left. +{\bf S}_{i}^{C}\left\{J_{\parallel}{\bf S}_{\delta}^{C}+J_{\perp}{\bf S}_{\delta}^{D}\right\} +{\bf S}_{i}^{D}\left\{J_{\parallel}{\bf S}_{\delta}^{D}+J_{\perp}{\bf S}_{\delta}^{C}\right\} \right], \label{eq:Hamiltonian}\end{aligned}$$ where $i$ denotes a unit cell and $\delta$ indicates NN sites corresponding to a particular interaction, $J_{\parallel}$ or $J_{\perp}$. Following the standard approach, we make the Holstein-Primakoff transformation[@T.Holstein_40] to boson creation and annihilation opperators $a_{n}^{\dagger},a_{n}, b_{n}^{\dagger},b_{n}, c_{n}^{\dagger},c_{n}, d_{n}^{\dagger},d_{n}$, which correspond to $A, B, C, D$ layers. By Fourier transforming to reciprocal space and performing diagonalization, we obtain the following dispersion relations as the eigenvalues: $$\begin{aligned} \hbar\omega ({\bf q}) & = & -2J_{\parallel}S\left(2-\cos aq_{x}-\cos aq_{y}\right) \nonumber \\ & & - J_{\perp}S\left\{1 \mp |\exp\left(-i2zcq_{z}\right)|\right\} \nonumber \\ & = & -2J_{\parallel}S\left(2-\cos aq_{x}-\cos aq_{y}\right) - J_{\perp}S\left(1 \mp 1\right), \label{eq:Dispersion}\end{aligned}$$ where $a$ and $c$ are the lattice constants and $2zc$ is the distance between layers within a double-layer. Although there should be four different modes, these are classified to two modes, i.e, acoustic (A) and optical (O), when the inter-bilayer coupling $J'$ is neglected. Note that both $J_{\parallel}$ and $J_{\perp}$ are negative because they are FM interactions. By using a unitary matrix diagonalizing the Hamiltonian Eq. \[eq:Hamiltonian\], we obtain the differential scattering cross section for spin waves in LSMO327 with $x=0.40$ $$\begin{aligned} \frac{d^{2}\sigma}{d\Omega dE_{f}} & = & \left(\frac{\gamma e^{2}}{mc^{2}}\right)^{2} \frac{k_{f}}{k_{i}} \left\{\frac{1}{2} g f(Q)\right\}^{2} \left(1+\hat{Q}_{x}^{2}\right) e^{-2W(Q)} \nonumber \\ & \times & \frac{S}{2} \frac{(2\pi)^{3}}{v_{0}}\frac{4}{N} \sum_{m}\sum_{\bf qG} \left(n_{q}^{(m)}+\frac{1}{2}\pm\frac{1}{2}\right) \\ & \times & \delta (\hbar\omega_{q}^{(m)}\mp\hbar\omega_{q}) \delta ({\bf Q}\mp{\bf q}-{\bf G}) \left\{1\pm\cos(2zc\cdot Q_{z})\right\} \nonumber \protect\label{eq:CrossSection}\end{aligned}$$ where $\gamma$ is the gyromagnetic ratio of the neutron, $f(Q)$ is the magnetic form factor for a Mn ion, $\exp[-2W(Q)]$ is a Debye-Waller factor, $\hat{Q}_{x}=Q_{x}/Q$, $n_{q}^{(m)}$ is the bose factor and $m$ denotes a mode. Since $2zc$ is very close to $a \approx c/5$, the A-branch has maximum intensity at $l=5n$ ($n$:integer), while the phase of O-branch is shifted by $\pi$. Experimental ============ La$_{1.2}$Sr$_{1.8}$Mn$_{2}$O$_{7}$ powder was prepared by solid-state reaction using prescribed amounts of pre-dried La$_{2}$O$_{3}$ (99.9%), Mn$_{3}$O$_{4}$ (99.9%), and SrCO$_{3}$ (99.99%). The powder mixture was calcined in the air for 3 days at 1250$^{\circ}$C –1400$^{\circ}$C with frequent grindings. The calcined powder was then pressed into a rod and heated at 1450$^{\circ}$C for 24 h. Single crystals were melt-grown in flowing 100% O$_{2}$ in a floating zone optical image furnace with a travelling speed of 15 mm/h. To check the sample homogeneity, we powderized a part of single crystal and performed x-ray diffraction, which shows no indication of impurities. Neutron-scattering measurements were carried out on the triple-axis spectrometer TOPAN located in the JRR-3M reactor of the Japan Atomic Energy Research Institute (JAERI). The $(0\ 0\ 2)$ reflection of pyrolytic graphite (PG) was used to monochromate and analyze the neutron beam, together with a PG filter to eliminate higher order contamination. The spectrometer was set up in two conditions in the standard triple-axis mode, typically with the fixed final energy at 13.5 meV and the horizontal collimation of B-100$'$-S-100$'$-B. The sample was mounted in an Al can so as to give the ($h\ 0\ l$) zone in the tetragonal $I4/mmm$ notation. We studied the same crystal used in Ref.  (F-40), which is a single grain with mosaic spread of $\sim 0.3^{\circ}$ full width at half maximum (FWHM). Results and Discussions ======================= The spin-wave dispersions along $[h\ 0\ 0]$ were measured at 10 K around $(1\ 0\ 0)$ and $(1\ 0\ 5)$ for the A-branch, and $(1\ 0\ 2.5)$ for the O-branch, as shown in Fig. \[Fig:Dispersion\]. Error bars correspond to the FWHM of peak profiles. By fitting all the data points for $0 < q \leq 0.25$ r.l.u.simultaneously, we obtain $-J_{\parallel}S = 10.1$ meV and $-J_{\perp}S=3.1$ meV. To quantitatively examine the present model, we measured the $l$-dependence of the spin-wave intensities of A and O branches at a fixed transfer energy $\Delta E=E_{i}-E_{f}=5$ meV. As shown in Fig. \[Fig:CrossSection\], the differencial scattering crosssection Eq. 3 is in an excellent agreement with both the A $(1.1\ 0\ l)$ and O $(1.0\ 0\ l)$ branches. Note that we do not use any fitting parameters except for intensity scaling. The results show that spin-spin correlations are significantly anisotropic. The inter-bilayer interaction is as small as we can not detect. The [*intra*]{}-bilayer interaction compared with in-plane interaction, $J_{\perp} /J_{\parallel}$ is about 0.31. We speculate that $x^{2}-y^{2}$ orbital is dominant in the Mn $e_{g}$ band, which enhances the double-exchange, i.e., ferromagnetic, interactions within a plane. A close relation between the magnetism and the Mn $e_{g}$ orbital degree of freedom has been also pointed out by recent studies.[@K.Hirota_98; @Y.Moritomo_97; @Y.Moritomo_98] The in-plane spin wave stiffness constant $D=-J_{\parallel}Sa^2$ is about 151 meVÅ$^2$, which is corresponding to the nearly cubic perovsikte La$_{1-x}$Sr$_{x}$MnO$_{3}$($x=0.2-0.3$) whose $D$ are 188 meV($x=0.3$,$T_{C}=370$ K) and 120 meV($x=0.2$, $T_{C}=310$ K)  [@Hirota_97; @Martin_96]. $T_{C}$ (120 K in our material) is very reduced, indicating a large renormalization due to low dimensionality. We noticed that the energy-width of constant-$Q$ scan profile becomes broad at large $q$, particularly $q>0.25$ r.l.u. In the same high $q$ range, the peak position starts deviating from the dispersion curve obtained from small $q$ data using a conventional Heisenberg model Eq. \[eq:Hamiltonian\]. Similar kind of broadening and deviation are seen in other CMR systems, such as Nd$_{0.7}$Sr$_{0.3}$MnO$_{3}$,[@Baca_98] which has a narrower electronic band-width than La$_{0.7}$Sr$_{0.3}$MnO$_{3}$. Although it is not clear that electron-phonon coupling plays a significant role in such anomalies in LSMO327 as suggested in Nd$_{0.7}$Sr$_{0.3}$MnO$_{3}$, it would be interesting to study the relation between structural and magnetic properties, particularly in their dynamics. Acknowledgments =============== The authors thank S. Ishihara and S. Okamoto for constructing the theoretical model. This work was supported by a Grant-in-Aid for Scientific Research of MONBUSHO. Y. Moritomo, A. Asamitsu, H. Kuwahara and Y. Tokura, Nature [**308**]{} (1996) 141. T. Kimura, Y. Tomioka, H. Kuwahara, A. Asamitsu, M. Tamura and Y. Tokura, Science [**274**]{} (1996) 1698. K. Hirota, Y. Moritomo, H. Fujioka, M. Kubota, H. Yoshizawa and Y. Endoh, J. Phys. Soc. Jpn. [**67**]{} (1998) 3380. J. F. Mitchell, D. N. Argyriou, J. D. Jorgensen, D. G. Hinks, C. D. Potter and S. D. Bader, Phys. Rev. B [**55**]{} (1996) 63. D. N. Arygyriou, J. F. Mitchell, J. B. Goodenough, O. Chmaissem, S. Short and J. D. Jorgensen, Phys. Rev. Lett. [**78**]{} (1997) 1568. M. Kubota, H. Fujioka, K. Ohoyama, K. Hirota, Y. Moritomo, H. Yoshizawa and Y. Endoh (unpublished). T. Holstein and H. Primakoff, Phys. Rev. [**58**]{} (1940) 1094. J. A. Fernandez-Baca, P. Dai H. Y. Hwang, C. Kloc and S-W. Cheng, Phys. Rev. Lett. [**80**]{} (1998) 4012. K. Hirota, N. Kaneko, A. Nishizawa, Y. Endoh, M. C. Martin, and G. Shirane, Physica B [**237-238**]{} (1998) 36. M. C. Martin, G. Shirane, Y. Endoh, K. Hirota, Y. Moritomo, and Y. Tokura, Phys. Rev. B [**53**]{} (1996) R14285. Y. Moritomo, Y. Maruyama, T. Akimoto, and A. Nkamura, Phys. Rev. B [**56**]{} (1997) R7057. Y. Moritomo, Y. Maruyama, T. Akimoto, and A. Nkamura, J.Phys. Soc. Jpn. [**67**]{} (1998) 405.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The [*Fermi*]{} satellite has been reporting the detailed temporal properties of gamma-ray bursts (GRBs) in an extremely broad spectral range, 8 keV - 300 GeV, in particular, the unexpected delays of the GeV emission onsets behind the MeV emission of some GRBs. We focus on GRB 080916C, one of the [*Fermi*]{}-LAT GRBs for which the data of the delayed high-energy emission are quite extensive, and we show that the behavior of the high-energy emission of this burst can be explained by a model in which the prompt emission consists of two components: one is the MeV component due to the synchrotron-self-Compton radiation of electrons accelerated in the internal shock of the jet and the other is the high-energy component due to inverse Compton scattering of the photospheric X-ray emission of the expanding cocoon off the same electrons in the jet. Such an external inverse Compton effect could be important for other [*Fermi*]{}-LAT GRBs, including short GRBs as well. In this model, the delay timescale is directly linked to the physical properties of GRB progenitor.' author: - 'K. Toma$^{1,2}$, X. F. Wu$^{1,2,3}$, and P. Mészáros$^{1,2,4}$' title: 'An External Inverse Compton Emission Model of Gamma-Ray Burst High-Energy Lags' --- Introduction ============ Gamma-ray bursts (GRBs) were only sparsely observed in the $> 100$ MeV range, until the [*Fermi*]{} satellite was launched on June 11 2008 [@atwood09]. Now [*Fermi*]{} provides extremely broad energy coverage, $8~{\rm keV}-300~{\rm GeV}$, with high sensitivity for GRBs, and is accumulating a wealth of new data which open a completely new window on the physics of GRBs. The high-energy temporal and spectral data provided by [*Fermi*]{} can severely constrain the physical parameters of the GRB emission region and the circumburst environment, which will lead to a deeper understanding of the central engine and the GRB progenitors, and will also constrain models of high-energy cosmic ray acceleration [@meszaros06; @falcone08]. GRB 080916C has the largest isotropic $\gamma$-ray energy release so far, $E_{\gamma,{\rm iso}} \simeq 8.8 \times 10^{54}\;$erg (with redshift $z\simeq 4.35$). [*Fermi*]{} LAT obtained its high-energy emission data quite extensively, showing several important new properties [@abdo09]: - The time-resolved spectra (with resolution $\sim 5-50$ s) are well fitted by a smoothly broken power-law function (the so-called Band function) from 8 keV up to a photon with energy $\approx 13.2$ GeV. - The $\varepsilon > 100$ MeV emission is not detected together with the first $\varepsilon \lesssim 1$ MeV pulse and the onset of the $\varepsilon>100$ MeV emission coincides with the rise of the second pulse ($\approx 5$ s after the trigger). - Most of the emission in the second pulse shifts towards later times as higher energies are considered. - The $\varepsilon>100$ MeV emission lasts at least 1400 s, while photons with $\varepsilon<100$ MeV are not detected past 200 s. Some other [*Fermi*]{}-LAT GRBs also display high-energy lags, similar to the properties (ii) and/or (iii) [@abdo09; @abdo09b], and then they should be very important to understand the prompt emission mechanism of GRBs. We will call the $\varepsilon \lesssim 1$ MeV emission and the $\varepsilon>100$ MeV emission “MeV emission” and “high-energy emission”, respectively. A simple physical picture for the property (i) is that the prompt emission consists of a single emission component, such as synchrotron radiation of electrons accelerated in internal shocks of a relativistic jet. In this picture, the peak of the MeV pulse could be attributed to the cessation of the emission production (i.e., the shock crossing of the shell) and the subsequent emission could come from the high latitude regions of the shell. Thus the observed high-energy lag for the second pulse (property (iii)) requires that the electron energy spectrum should be harder systematically in the higher latitude region. This would imply that the particle acceleration process should definitely depend on the global parameters of the jet, e.g., the angle-dependent relative Lorentz factor of the colliding shells, but such a theory has not been formulated yet. The property (ii) could be just due to the fact that the two pulses originate in two internal shocks with different physical conditions for which the electron energy spectrum of the second internal shock is harder than that of the first one [@abdo09]. Another picture is that the prompt emission consists of the MeV component and a delayed high-energy component. The latter component could be produced by hadronic effects (i.e., photo-pion process and proton synchrotron emission) [@asano09; @razzaque09], but they require extremely large total energy budget [@wang09]. In this paper, we discuss a different two-component emission picture in which the delayed high-energy component is produced by leptonic process (i.e., electron inverse Compton scattering). We focus on the effect that the ambient radiation up-scattered by the accelerated electrons in the jet can have a later peak than that of the synchrotron and synchrotron-self-Compton (SSC) emission of the same electrons (corresponding to the property (iii)) [@wang06; @fan08; @beloborodov05]. Provided that the seed photons for the Compton scattering come from the region behind the electron acceleration region of the jet (see Figure \[fig:geometry\]), the up-scattered high-energy photon field is highly anisotropic in the comoving frame of the jet, i.e., the emissivity is much larger for the head-on collisions of the electrons and the seed photons. As a result, a stronger emission is observed from the higher latitude regions, and thus its flux peak lags behind the synchrotron and SSC emission. Here we propose that the seed photons may be provided by the photospheric emission of an expanding cocoon. GRB 080916C is a long GRB, and it may originate from the collapse of a massive star. The relativistic jet produced in the central region penetrates the star and deposits most of its energy output into a thermal bubble, or cocoon, until it breaks out of the star [@meszaros01; @matzner03; @wzhang03; @morsony07]. The cocoon can store an energy comparable or larger than the energy of the prompt emission of the jet, and thus it may make an observable signature outside the star [@ramirez02; @peer06]. The cocoon escaping from the star will emit soft X-rays, and these can be up-scattered by the accelerated electrons in the jet into the high-energy range. The optical thinning of the expanding cocoon may be delayed behind the prompt emission of the jet, so that the onset of the high-energy emission is delayed behind the MeV emission (corresponding to the property (ii)). Thus this model has the potential for explaining the two delay timescales; the delayed onset of the high-energy photons (property (ii)) is due to the delayed emission of the cocoon, while the high-energy lag within the second pulse (property (iii)) is due to the anisotropic inverse Compton scattering. We also show that the combination of the time-averaged spectra of the SSC and the up-scattered cocoon (UC) emission is roughly consistent with the observed smooth power-law spectrum (property (i)) (see Figure \[fig:spectrum\]). ![ Geometry of our model. The jet with opening angle $\theta_j \simeq 0.01$ and the cocoon with opening angle $\theta_c \simeq 0.8$ are ejected from the collapsing star. The cocoon becomes optically thin at $r = r_{\rm ph}$ some time ($\lesssim 5$ s) after the burst trigger. The second pulse of the prompt emission is produced by the accelerated electrons in the internal shock of the jet at $r=r_i$, and the cocoon X-ray photons are up-scattered by the same electrons at $r=r_i$ into the GeV range. []{data-label="fig:geometry"}](f1.eps){width="65mm"} As we will explain below, the UC emission is short-lived and may not account for the whole high-energy emission, which lasts much longer than the MeV emission (property (iv)). It is natural that the high-energy emission in later times is related to the afterglow, i.e., produced by the external shock which propagates in the circumburst medium. This possibility is studied by [@kumar09]. They claim that even the onset phase of the high-energy emission is produced by the external shock. However, the rise of the flux of GRB 080916C in the LAT energy range ($\sim t^6$) is too steep for the external shock to reproduce it. Thus at least the first part of the delayed high-energy component of this burst should be related to the prompt emission. Although we focus on GRB 080916C to examine whether the UC emission is viable for its properties in the high energy range in this paper, our modeling is generic, and the UC emission could be important for some other [*Fermi*]{}-LAT GRBs (including short GRBs). We will discuss the temporal and spectral properties of the cocoon emission (in § 2), the synchrotron and SSC emission of the jet (in § 3), and the UP emission of the jet (in § 4). We summarize our model and its implications in § 5. For more details, see [@toma09]. Cocoon Emission {#sec:cocoon_emission} =============== We take the total energy and the total mass of the cocoon and the stellar radius as parameters, $E_c$, $M_c$, and $r_*$, respectively. After the jet breaks out of the star, the cocoon expands with the sound speed, $c_s = c/\sqrt{3}$, where $c$ is the speed of light. The cocoon expands by its own thermal pressure in the comoving frame as expected in the standard fireball model [@meszaros06]. Its expansion speed in the comoving frame suddenly becomes close to $c$, and then the opening angle of the cocoon measured in the central engine frame is given by $\theta_c \simeq \Gamma_{c,{\rm ex}}^{-1}$, where $\Gamma_{c,{\rm ex}}$ is the Lorentz factor corresponding to $c_s$. Thus we obtain $\theta_c \simeq \sqrt{2/3} \simeq 0.8$. This rough estimate is consistent with the results of the numerical simulations [@morsony07]. The cocoon material accelerates as $\Gamma \propto r$ and reaches the terminal Lorentz factor $\Gamma_c = E_c/M_c c^2$ at $r_s = r_* \Gamma_c = 5.0\times10^{12}~r_{*,11} (\Gamma_c/50)~$cm. (We adopt the notation $Q = 10^x Q_x$ in cgs units throughout this paper.) For $r>r_s$, the cocoon material is dominated by the kinetic energy. The radiation stored in the cocoon is released when the opacity for the electron scattering becomes less than unity. The photosphere radius is given by $$\begin{aligned} r_{\rm ph} &\simeq& \left[\frac{E_c \sigma_T}{2\pi (1-\cos\theta_c) \Gamma_c m_p c^2}\right]^{1/2} \nonumber \\ &\simeq& 2.1 \times 10^{14}~E_{c,52}^{1/2} \left(\frac{\Gamma_c}{50}\right)^{-1/2}~{\rm cm}, \label{eq:photospheric_radius}\end{aligned}$$ where $\sigma_T$ is Thomson cross section. The cocoon may become optically thin later than the onset of the first MeV pulse (i.e., the burst trigger time). Let $t=0$ be the photon arrival time at the earth emitted at the stellar surface at the jet breakout. The first MeV pulse produced within the jet at $r=r_i$ is observed at $t \simeq \Delta t_i \equiv r_i(1+z)/(2c \Gamma_j^2)$, where $\Gamma_j$ is the bulk Lorentz factor of the jet and $z$ is the source redshift. This timescale is comparable to the angular spreading timescale of the pulse, and we can take $\Delta t_i \simeq 2$ s for GRB 080916C. The second MeV pulse is observed $\simeq 5$ s after the burst trigger, i.e., at $t \simeq \Delta t_i + 5 \simeq 7$ s. On the other hand, the cocoon photospheric emission is observed at $$t \simeq \Delta t_c \equiv \frac{r_{\rm ph}}{2c \Gamma_c^2} (1+z) \simeq 7.5~E_{c,52}^{1/2} \left(\frac{\Gamma_c}{50}\right)^{-5/2}~{\rm s}, \label{eq:delta_t_c}$$ where we have used $z \simeq 4.35$. The cocoon photospheric emission may be observed from $t=\Delta t_c$ to $t = \Delta t_c + \Delta t_d$, where $\Delta t_d \sim \Delta t_c$ is the time during which the cocoon will be adiabatically cooled. If internal dissipation occurs in the jet at $r = r_i > r_{\rm ph}$ making the second MeV pulse within the duration of the cocoon emission, the cocoon photons may be up-scattered to higher energies by the energetic electrons within the dissipation region of the jet, which may be observed along with the second MeV pulse. Therefore we require a condition $\Delta t_i < \Delta t_c \lesssim \Delta t_i + 5$. This condition puts a constraint to the physical parameters of the cocoon, $$0.3 < E_{c,52}^{1/2} \left(\frac{\Gamma_c}{50}\right)^{-5/2} \lesssim 0.9 \label{eq:co_constraint1}$$ We adopt the parameters $E_c \simeq 10^{52}$ erg and $\Gamma_c \simeq 50$ for the purposes of calculating the flux of the cocoon emission. The comoving temperature of the cocoon when its opening angle becomes $\theta_c$ is approximately given by $T'_{\rm init} \simeq [E_c/(2\pi (1-\cos\theta_c) r_*^3 a)]^{1/4}$, where $a$ is the Stefan constant. Then the temperature at the photosphere is estimated to be $T'_{\rm ph} = T'_{\rm init} (r_s/r_*)^{-1} (r_{\rm ph}/r_s)^{-2/3}$. Non-thermal electrons injected by internal shocks within the photosphere may make the emission be quasi-thermal. Its spectum is written by $$F^{\rm co}_{\varepsilon} = F^{\rm co}_{\varepsilon_{\rm ph}} \times \left\{ \begin{array}{lc} \left(\frac{\varepsilon}{\varepsilon^{\rm co}_{\rm ph}}\right)^2 & {\rm for}~\varepsilon < \varepsilon^{\rm co}_{\rm ph}, \\ \left(\frac{\varepsilon}{\varepsilon^{\rm co}_{\rm ph}}\right)^{\beta} & {\rm for}~ \varepsilon^{\rm co}_{\rm ph} < \varepsilon < \varepsilon^{\rm co}_{\rm cut}, \end{array} \right. \label{eq:cocoon_spectrum}$$ where $\varepsilon_{\rm ph}^{\rm co}$ and $F_{\varepsilon_{\rm ph}}^{\rm co}$ are given by $$\varepsilon_{\rm ph}^{\rm co} \simeq 2.82~ kT'_{\rm ph} \frac{2\Gamma_c}{1+z} \simeq 1.2~E_{c,52}^{-1/12} r_{*,11}^{-1/12}~{\rm keV} \left(\frac{\Gamma_c}{50}\right) \label{eq:cocoon_nu}$$ $$\begin{aligned} F_{\varepsilon_{\rm ph}}^{\rm co} & \simeq & \frac{(1+z)^3}{d_L^2} \frac{2\pi (\nu_{\rm ph}^{\rm co})^2}{c^2} kT'_{\rm ph} \Gamma_c \left(\frac{r_{\rm ph}}{\Gamma_c}\right)^2 \nonumber \\ & \simeq & 31~E_{c,52}^{3/4} r_{*,11}^{-1/4}~{\rm keV}~{\rm cm}^{-2}~{\rm s}^{-1}~{\rm keV}^{-1}, \label{eq:cocoon_fnu}\end{aligned}$$ where $\nu_{\rm ph}^{\rm co} = \varepsilon_{\rm ph}^{\rm co}/h$ ($h$ is the Planck constant), and we have taken the luminosity distance of GRB 080916C as $d_L \simeq 1.2 \times 10^{29}$ cm. Some numerical calculations of the radiative processes in the cocoon show $\beta \sim -1$ and $\varepsilon_{\rm cut}^{\rm co} \sim 30 \times \varepsilon_{\rm ph}^{\rm co}$ [@peer06]. The observation of GRB 080916C shows that there is no excess from the Band spectrum at the X-ray band, $\gtrsim 10$ keV, and we obtain a rough upper limit of the cocoon X-ray emission $\varepsilon_{\rm ph}^{\rm co} F_{\varepsilon_{\rm ph}}^{\rm co} \lesssim 40~{\rm keV}~{\rm cm}^{-2}~{\rm s}^{-1}$. This limit leads to another constraint on the cocoon parameters, $$r_{*,11} \gtrsim 0.8~E_{c,52}^2 \left(\frac{\Gamma_c}{50}\right)^3. \label{eq:co_constraint3}$$ Synchrotron and SSC Emission {#sec:emission_formalism} ============================ We can constrain, from the [*Fermi*]{} observation, the global physical parameters of the jet: the bulk Lorentz factor $\Gamma_j$, the emission radius of the second pulse $r_i$, and the opening angle $\theta_j$. First of all, from the absence of a $\gamma\gamma$ absorption cutoff, we obtain a lower limit on the bulk Lorentz factor of the jet, $\Gamma_j \gtrsim 870$ [@abdo09]. Since the angular spreading timescale of the pulse is $\Delta t_i \simeq 2$ s for the second pulse, similar to the first pulse, the emission radius is estimated by $$r_i \simeq 2c \Gamma_j^2 \frac{\Delta t_i}{1+z} \simeq 2.2 \times 10^{16}~\Gamma_{j,3}^2 \left(\frac{\Delta t}{2~{\rm s}}\right)~{\rm cm}.$$ Since GRB 080916C is so bright, it is probable that the jet is viewed on-axis, and we adopt this assumption as a simplification (see Figure \[fig:geometry\]). In this case, the cocoon is viewed off-axis, since the jet cone will not be filled with the cocoon material. The cocoon emission is thus less beamed, but this off-axis effect is not significant because the opening angle of the jet can be estimated to be small. The isotropic $\gamma$-ray energy of this burst is $8.8 \times 10^{54}$ erg. To obtain a realistic value of the collimation-corrected $\gamma$-ray energy, $\lesssim 10^{51}$ erg, the jet opening angle is constrained by $\theta_j \lesssim 0.015$. We adopt $\theta_j \simeq 0.01$, and having adopted a nominal value of $\Gamma_c \simeq 50$ in accord with the observed time delay of the high-energy emission (equation \[eq:co\_constraint1\]), we obtain $$\Gamma_c \theta_j \simeq 0.5 < 1.$$ Thus the off-axis dimming and softening effects are not significant for the cocoon emission. We assume that the jet is dominated by the kinetic energy of protons and we estimate the physical parameters of the jet dissipation region for the second MeV pulse. In the scenario where the dissipation is due to internal shocks [@meszaros06], the collisionless shock waves can amplify the magnetic field and accelerate electrons to a power-law energy distribution, which then produce synchrotron radiation and SSC radiation. At $r=r_i$, the comoving number density of the jet is estimated by $n'= L \Delta t_i/(4\pi r_i^3 m_p c^2 (1+z))$, where $L$ is the isotropic-equivalent luminosity of the jet. The internal energy density produced by the internal shock is given by $u' = n' \theta_p m_p c^2$, where $\theta_p$ is a factor of order unity. Assuming that a fraction $\epsilon_B$ of the internal energy of the protons is carried into the magnetic field, the field strength is estimated by $B' = (8\pi \epsilon_B u')^{1/2}$. Assuming that a fraction $\epsilon_e$ of the internal energy of the protons is given to the electrons, the minimum Lorentz factor of the electrons is given by $\gamma_m = [(p-2)/(p-1)](m_p/m_e)\epsilon_e \theta_p$, where $p$ is the index of the electron energy distribution and we have assumed that it is similar to the value inferred from the spectrum of the first pulse, $p \simeq 3.2$. We find that the cooling electron Lorentz factor $\gamma_c \sim \gamma_m$ for fitting the observational data. In this case we obtain $\gamma_c \simeq 360 (\epsilon_B/10^{-5})^{-1/3} (\tau/4\times10^{-4})^{-2/3}$, where $\tau = \sigma_T n' r_i/\Gamma_j$ is the Thomson optical depth. The synchrotron characteristic energy and the synchrotron peak flux (at the synchrotron energy corresponding to $\gamma_c$) are estimated by $$\begin{aligned} \varepsilon_m &\simeq& \frac{3h e B'}{4\pi m_e c}\gamma_m^2 \frac{2\Gamma_j}{1+z} \simeq 2.7~L_{55}^{1/2} \Gamma_{j,3}^{-2} \left(\frac{\Delta t_i}{2~{\rm s}}\right)^{-1} \nonumber \\ &\times& \theta_p^{5/2} \left(\frac{\epsilon_B}{10^{-5}}\right)^{1/2} \left(\frac{\epsilon_e}{0.4}\right)^2~{\rm eV}, \label{eq:syn_num}\end{aligned}$$ $$\begin{aligned} F_{\varepsilon_c} &\simeq& \frac{\sqrt{3} e^3 B' N}{m_e c^2} \frac{2\Gamma_j (1+z)}{4\pi d_L^2} \simeq 1.3 \times 10^4~L_{55}^{3/2} \Gamma_{j,3}^{-3} \nonumber \\ &\times& \theta_p^{1/2} \left(\frac{\epsilon_B}{10^{-5}}\right)^{1/2} ~{\rm keV}~{\rm cm}^{-2}~{\rm s}^{-1}~{\rm keV}^{-1}~ \label{eq:syn_fnuc}\end{aligned}$$ where $N = [L \Delta t_i/(1+z)]/(\Gamma_j m_p c^2)$. The 1st-order SSC characteristic energy and the SSC peak flux are approximately $$\begin{aligned} \varepsilon_m^{\rm SC} &\simeq& 4\gamma_m^2 \varepsilon_m \simeq 1.7~L_{55}^{1/2} \Gamma_{j,3}^{-2} \left(\frac{\Delta t_i}{2~{\rm s}}\right)^{-1} \nonumber \\ &\times& \theta_p^{9/2} \left(\frac{\epsilon_B}{10^{-5}}\right)^{1/2} \left(\frac{\epsilon_e}{0.4}\right)^4~{\rm MeV}, \label{eq:SSC_num}\end{aligned}$$ $$\begin{aligned} F_{\varepsilon_c}^{\rm SC} &\simeq& \tau F_{\varepsilon_c} \simeq 3.4~L_{55}^{5/2} \Gamma_{j,3}^{-8} \left(\frac{\Delta t_i}{2~{\rm s}}\right)^{-1} \nonumber \\ &\times& \theta_p^{1/2} \left(\frac{\epsilon_B}{10^{-5}}\right)^{1/2} ~{\rm keV}~{\rm cm}^{-2}~{\rm s}^{-1}~{\rm keV}^{-1}. \label{eq:SSC_fnuc}\end{aligned}$$ We find that the 1st-order SSC radiation can account for the observed MeV emission of this burst. The emission just below $\varepsilon_c$ is suppressed by synchrotron self-absorption effect, and the 2nd-order SSC emission is suppressed by the Klein-Nishina effect. Upscattered Cocoon (UC) emission {#sec:aic} ================================ Here we derive the observed spectrum of the UC emission as a function of the polar angle $\theta$ of the emitting region on the shell (see Figure \[fig:geometry\]). The spectrum of radiation scattered at an angle $\theta'_{\rm sc}$ relative to the direction of the photon beam in the Thomson scattering regime is given by [@brunetti01; @wang06; @fan08] $$j'_{\varepsilon'}(\theta'_{\rm sc}) = \frac{3}{2} \sigma_T (1-\cos\theta'_{\rm sc})\int d\gamma N'(\gamma) \int^1_0 dy J'_{\varepsilon'_s}(1-2y+2y^2), \label{eq:aic_emissivity}$$ where $y=\varepsilon'/[2\gamma^2 \varepsilon'_s (1-\cos\theta'_{\rm sc})]$. This is the scattered radiation emissivity in the jet comoving frame, $N'(\gamma)$ is the electron energy spectrum, and $J'_{\varepsilon'_s}$ is the intensity of the seed photons averaged over solid angle, i.e., the mean intensity. The term $\xi \equiv 1-\cos\theta'_{\rm sc}$ describes the anisotropy of the spectrum, and this is due to the fact that the IC scattering is strongest for the head-on collisions between electrons and seed photons. This implies that the UC emission in the observer frame is stronger from the high-latitude region of the shell, so that its flux peak lags behind the onset of the synchrotron and SSC emission of the same electrons, which have isotropic energy distribution in the comoving frame of the jet. In order to concentrate on the time-averaged spectrum including the high-latitude emission, we calculate the flux of the UC emission by neglecting the radial structure of the emitting shell for simplicity. The peak energy and the peak flux of the UC emission are reduced as functions of the angle parameter $q(\theta) \equiv \Gamma_j^2\theta^2$: $$\begin{aligned} \varepsilon_m^{\rm UC} &=& 2\gamma_m^2 \varepsilon_{\rm ph}^{\rm co} \frac{\xi(\theta)}{1+\Gamma_j^2 \theta^2} \nonumber \\ &\simeq& 160~{\rm MeV}~\left[\frac{4 q}{(1+q)^2}\right] \left(\frac{\gamma_m}{400}\right)^2 \left(\frac{\varepsilon_{\rm ph}^{\rm co}}{1~{\rm keV}}\right), \label{eq:co_nu} \\ \varepsilon_m^{\rm UC} F_{\varepsilon_m}^{\rm UC} &=& 3 \tau \gamma_m \gamma_c \varepsilon_{\rm ph}^{\rm co} F_{\varepsilon_{\rm ph}}^{\rm co} \frac{\xi^2(\theta)}{[1+\Gamma_j^2\theta^2]^3} \nonumber \\ &\simeq& 580~{\rm keV}~{\rm cm}^{-2}~{\rm s}^{-1}~ \left[\frac{40 q^2}{(1+q)^5}\right] \left(\frac{\tau}{4\times10^{-4}}\right) \nonumber \\ &\times& \left(\frac{\gamma_m}{400}\right) \left(\frac{\gamma_c}{400}\right) \left(\frac{\varepsilon_{\rm ph}^{\rm co} F_{\varepsilon_{\rm ph}}^{\rm co}} {30~{\rm keV}{\rm cm}^{-2}{\rm s}^{-1}}\right), \label{eq:co_fnu}\end{aligned}$$ where the functions in the brackets $[~]$ both have values of zero at $q=0$ and $q=\infty$ and have peaks of $1$ at $q=1$ and $\simeq 1.4$ at $q=2/3$, respectively. This means that the UC flux has a peak at $q \simeq 1$, or $\theta \simeq \Gamma_j^{-1}$, i.e., the peak time of the UC emission lags behind that of the SSC emission on the angular spreading timescale, $\Delta t_i \simeq 2$ s. This is consistent with the observed lag of the GeV emission onset behind the MeV emission peak of the second pulse of GRB 080916C. Here the values of the jet parameters $\tau = 4\times10^{-4}$ and $\gamma_m =\gamma_c=400$ are applicable for the 1st-order SSC emission of the jet being consistent with the observed MeV emission component (see § 2). This indicates that the UC emission of the electrons accelerated in the internal shock of the jet, emitting the observed MeV emission, can naturally explain the observed flux in the GeV range. ![ Time-averaged spectrum of the second pulse calculated in the up-scattered cocoon emission model. The 1st-order SSC component plus 2nd-order SSC component without including the Klein-Nishina effect ([*dashed line*]{}), the cocoon photospheric emission ([*dot-dashed line*]{}), and the UC emission ([*dotted line*]{}) are shown. The [*thick solid line*]{} represents the combination of these, taking account of the Klein-Nishina effect, which is roughly consistent, at $\varepsilon \gtrsim 1$ MeV, with the Band model spectrum ([*thin solid line*]{}) with 95% confidence errors ([*dot-short-dashed lines*]{}) (from the LAT/GBM group of [*Fermi*]{}). The bump at $\sim 30$ GeV is so dim as not to be detected. The adopted parameters are listed in equation (\[eq:cocoon\_parameter\]) and (\[eq:jet\_parameter\]). []{data-label="fig:spectrum"}](f2.eps){width="65mm"} If the flux of the cocoon photospheric X-rays is given, i.e., $E_c$, $\Gamma_c$, and $r_*,$ are given, the fluxes of the UC and SSC emission of the jet are determined by the jet parameters $L, \Gamma_j, \Delta t_i, \epsilon_B$, and $\epsilon_e$. Since $\Delta t_i \sim 2$ s is roughly given by the observations, and this value is necessary to explain the observed high-energy lag timescale, we have four free parameters. On the other hand, we have four characteristic observables; the peak fluxes and peak photon energies of the SSC component and the UC component. Therefore the jet parameters are expected to be constrained tightly. Figure \[fig:spectrum\] shows the result of the time-averaged spectrum of the second pulse for the cocoon parameters $$E_{c,52} = 1.0,~~ \Gamma_c = 52,~~ r_{*,11} = 2.5, \label{eq:cocoon_parameter}$$ and $\beta = -1.2$. These values satisfy the constraints on the cocoon parameters, equations (\[eq:co\_constraint1\]) and (\[eq:co\_constraint3\]). This figure shows that our model is roughly consistent with the observed spectrum at $\varepsilon \gtrsim 1$ MeV. The adopted values of the jet parameters are $$L_{55} = 1.1,~~ \Gamma_{j,3} = 0.93,~~ \Delta t_i = 2.3~{\rm s},~~ \epsilon_B = 10^{-5},~~\epsilon_e = 0.4, \label{eq:jet_parameter}$$ and $p = 3.2$. The corresponding values of the optical depth for electron scattering and the characteristic electron Lorentz factors are $\tau = 3.5 \times 10^{-4}, \gamma_m = 400, \gamma_c = 390$. Figure \[fig:lightcurve1\] shows the results of the multi-band lightcurves for the same parameters. Each lightcurve is normalized to a peak flux of unity. This clearly displays the lag of the high-energy emission peak. ![ Photon number fluxes of the second pulse in several frequency ranges calculated in the up-scattered cocoon emission model. Each lightcurve is normalized to a peak flux of unity. The peak of the GeV lightcurve is delayed behind that of the MeV lightcurve. []{data-label="fig:lightcurve1"}](f3.eps){width="65mm"} Discussion {#sec:summary} ========== We have discussed a model in which the prompt emission spectrum consists of an MeV component produced by the SSC emission of electrons accelerated in internal shocks in the jet and the high-energy component produced by up-scattering of the cocoon X-rays off the same electrons (UC emission), and we have shown that this model can explain the above three observed properties (i), (ii), and (iii) listed in § 1. The expanding cocoon may become optically thin some time later than the first internal shock of the jet (equation \[eq:delta\_t\_c\]), so that the first MeV pulse may not be associated with the UC emission while the second MeV pulse may be associated with it (property (ii)). The UC emission has an anisotropic energy distribution in the comoving frame of the jet so that the observed UC emission is stronger from the higher-latitude region of the shell. This results in the lag of the flux peak of the UC emission behind the MeV emission onset on the angular spreading timescale (property(iii)). Figure \[fig:spectrum\] shows that the combination of the SSC and UC emission can reproduce the observed high-energy spectral data (property (i)). The UC emission is short-lived (roughly for $\Delta t_c \simeq 7.5\;$s) and may not account for the whole high-energy emission which lasts longer than the MeV emission (property (iv)). It is natural that the high-energy emission in the later times is related to the afterglow. This has been shown by [@kumar09]. However, the early portion of the high-energy emission should be the UC emission, because the external shock cannot reproduce the observed steep rise of the flux. We have focused on GRB 080916C because of its extensive data on the high-energy emission. However our model is generic, and it could apply to other [*Fermi*]{}-LAT GRBs with typical parameters $L \simeq 10^{53}~{\rm erg}~{\rm s}^{-1}$, $\theta_j \simeq 0.1$, and $\Gamma_j \simeq 300$ for which the prompt emission in the soft $\gamma$-ray range is produced by the 1st-order SSC radiation of electrons [@toma09]. Some short GRBs might originate from the collapses of the massive stars [@zhang09], and thus they could have the delayed UC emission. Even if other short GRBs are produced by the compact star mergers, it might be possible that the jet is accompanied by the delayed disc wind [@metzger08] and that the emission from the disc wind is up-scattered by the electrons accelerated in the jet. For either progenitor models, the delayed high-energy emission associated with short GRBs, if any, would provide an interesting tool to approach their origins. A simple prediction of our model is that prompt emission spectra of some GRBs would have an excess above the Band spectrum around $\sim 1$ keV due to the cocoon photospheric emission, and this excess should have a different temporal behavior from that of the MeV emission. In addition, we expect GRB 080916C to have had bright synchrotron emission in the optical band, like the “naked-eye” GRB 080319B. If our model is correct, we can constrain the parameter range for which hadronic effects are important on the high-energy emission of GRBs, and we can also constrain the models of high-energy cosmic ray acceleration. Also, the delay time of the onset of the high-energy emission is directly linked to the optical-thinning time of the expanding cocoon, which constrains the physical parameters of the progenitor star and the cocoon material of GRBs. For GRB 080916C, the stellar radius $r_*$ and the total energy $E_c$ and mass $M_c$ of the cocoon are constrained to be $0.3 < E_{c.52}^{-2} (M_c/10^{-4}M_{\odot})^{2.5} \lesssim 0.9$ and $r_{*,11} \gtrsim 0.8~E_{c,52}^5 (M_c/10^{-4}M_{\odot})^{-3}$ (see equations \[eq:co\_constraint1\] and \[eq:co\_constraint3\]). The cocoon energy and the cocoon mass come from the jet energy released within the star and the stellar mass swept by the jet, respectively. These constraints therefore provide potential tools for investigating the structure of the progenitor star just before the explosion, as well as the physical conditions of the jet propagating inside the stellar envelope through either analytical [@meszaros01; @matzner03] or numerical [@wzhang03; @morsony07] approaches. We acknowledge NASA NNX09AT72G, NASA NNX08AL40G, and NSF PHY-0757155 for partial support. XFW was supported by the National Natural Science Foundation of China (grants 10503012, 10621303, and 10633040), National Basic Research Program of China (973 Program 2009CB824800), and the Special Foundation for the Authors of National Excellent Doctorial Dissertations of P. R. China by Chinese Academy of Sicences. [9]{} Atwood, W. B., et al. 2009, ApJ, 697, 1071 Mészáros, P. 2006, Rep. Prog. Phys., 69, 2259 Falcone, A. D., et al. 2008, GRB section of the white paper on the status and future of ground-based TeV gamma-ray astronomy (arXiv:0810.0520) Abdo, A. A., et al. 2009a, Science, 323, 1688 Abdo, A. A., et al. 2009b, ApJ, 706, L138 Asano, K., Inoue, S., & Mészáros, P. 2009, ApJ, 699, 953 Razzaque, S., Dermer, C., & Finke, J. D. 2009, arXiv:0908.0513 Wang, X. Y., Li, Z., Dai, Z. G., & Mészáros, P. 2009, ApJ, 698, L98 Wang, X. Y., & Mészáros, P. 2006, ApJ, 643, L95 Fan Y. Z., Piran, T., Narayan, R., & Wei, D. M. 2008, MNRAS, 384, 1483 Beloborodov, A. M. 2005, ApJ, 618, L13 Mészáros, P., & Rees, M. J. 2001, ApJ, 556, L37 Matzner, C. D. 2003, MNRAS, 345, 575 Zhang, W., Woosley, S., & MacFadyen, A. I. 2003, ApJ, 586, 356 Morsony, B. J., Lazzati, D., & Begelman, M. C. 2007, ApJ, 665, 569 Ramirez-Ruiz, E., Celloti, A., & Rees, M. J. 2002, MNRAS, 337, 1349 Pe’er, A., Mészáros, P., & Rees, M. J. 2006, ApJ, 652, 482 Kumar, P., & Barniol Duran, R. 2009, MNRAS in press (arXiv:0905.2417) Toma, K., Wu, X.-F., & Mészáros, P. 2009, ApJ, 707, 1404 Brunetti, G. 2001, Astropart. Phys., 13, 107 Zhang, B., et al. 2009, ApJ, 703, 1696 Metzger, B. D., Piro, A. L., & Quataert, E. 2008, MNRAS, 390, 781
{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the Helmholtz decomposition of the vector field of folding fluxes in a two-dimensional space of collective variables, a potential of the driving force for protein folding is introduced. The potential has two components. One component is responsible for the source and sink of the folding flows, which represent, respectively, the unfolded states and the native state of the protein, and the other, which accounts for the flow vorticity inherently generated at the periphery of the flow field, is responsible for the canalization of the flow between the source and sink. The theoretical consideration is illustrated by calculations for a model $\beta$-hairpin protein.' author: - 'Sergei F. Chekmarev' title: 'Protein folding: Complex potential for the driving force in a two-dimensional space of collective variables' --- Introduction ============ Protein folding is a reaction in which a protein attains its functional (native) state; the unfolded state of the protein plays a role of the reactant, and the native state plays a role of the product. In contrast to simple chemical reactions, where the reactant, transition state and product are well defined species, the folding reaction is characterized by ensembles of unfolded and transition states. Accordingly, folding pathways are numerous and diverse. The understanding of this difference between the protein folding and chemical reactions has led to a statistical view of protein folding [@BryngelsonWolynes87; @ShakhnovichFarztdinovGutinKarplus91; @LeopoldMontalOnuchic92; @BryngelsonOnuchicSocciWolynes95; @DillChan97; @Shakhnovich97; @OnuchicLuthey_SchultenWolynes97; @DobsonSaliKarplus98; @DSSDK00; @SheaBrooks01; @SchulerEverettLipmanEaton02; @Gruebele2002; @MelloBarrick04; @Karplus11] (for review, see Refs. and ). According to this view, the folding process is governed by the interplay of the protein potential energy and conformation entropy, forming a “funnel-like” energy surface. The energy and entropy both decrease from the unfolded states of the protein to its native state but in different manner, so that a free energy barrier is formed that separates the native-like states from the unfolded states. A tempting and widely used approach to study the folding reaction is to construct a free energy profile as a function of a single reaction coordinate [@BerezhkovskiiSzabo05; @RheePande05; @BestHummer2006; @KrivovKarplus06; @BeckDaggett07; @ChoderaPande11; @PetersBolhuisMullenShea13]. However, although the diversity of folding pathways is taken into account (as the entropy part of the free energy), this approach is generally limited to the case of a single reaction channel, when the folding pathways can be organized in a “transition tube” [@E_Vanden-Eijnden10]. A more powerful method to characterize the diversity of the pathways is the construction of the free energy surface (FES) of the folding reaction as a function of two collective variables [@SheaBrooks01; @SocciOnuchicWolynes98; @BoczkoBrooks95; @ChekmarevKrivovKarplus05]. One variable is usually chosen to describe the protein compaction during folding (e.g., the radius of gyration, $r_{\mathrm {g}}$), and the other its proximity to the native conformation (e.g., the fraction of the native contacts, $f_{\mathrm {nat}}$). To construct the FES, the probability for the system to be at a current point of the conformation space $P(f_{\mathrm {nat}},r_{\mathrm {g}})$ is calculated and then, using the Boltzmann hypothesis, is converted into the free energy $$\label{eq0}F(f_{\mathrm {nat}},r_{\mathrm {g}})= -k_{\mathrm{B}}T \ln P(f_{\mathrm {nat}},r_{\mathrm {g}})$$ where $T$ is the temperature, and $k_{\mathrm{B}}$ is the Boltzmann constant. A shortcoming of the FES thus determined is that it does not show the direction of the motion, i.e., the protein can have the same probability to be at some point of the conformation space when it goes towards the native state or requires partial unfolding to reach the native state. To gain a closer insight into folding dynamics, we have recently introduced a “hydrodynamic” description of the folding process [@ChekmarevPalyanovKarplus08]. In this approach, similar to the FESs, the folding process is considered in a reduced space of collective variables, and, similar to Markov state models (MSMs) [@Schutte99; @SwopePiteraSuits04; @WeberPande11], disconnectivity graphs (DGs) [@BeckerKarplus97; @KrivovKarplus04; @Wales03] and protein folding networks (PFNs) [@RaoCaflisch04; @Noe-Weikl; @BowmanVoelzPande11], the calculated folding trajectories are used to determine probabilities of transitions between the protein states. On the one hand, in comparison to the MSMs, this allows consideration of more complex (non-Markovian) kinetics, and on the other, in comparison to the DGs and PFNs, to arrange the protein states according to their distribution in the conformation space [@KalginCaflischChekmarevKarplus13]. The probabilities of transitions are organized as local flows between the points of the (reduced) conformation space. Given the flows, it is possible to construct the vector field and streamlines of the folding flows, similar to how it is done in hydrodynamics [@LandauLifshitz87]. Then, the process of protein folding can be viewed as a motion of a folding “fluid”, with the density of the fluid being proportional to the probability for the system to be at the current point of the conformation space. In equilibrium conditions, the local flows of transitions become small, or vanish, due to detailed balance [@KalginCaflischChekmarevKarplus13]. Therefore, the hydrodynamic description is most efficient for nonequilibrium conditions, particularly if detailed balance is absent on the overall scale. Such is the case when the native state is essentially stable; then the unfolding events are rare, and the process of folding is reduced to the first-passage folding. Correspondingly, the folding reaction is represented by a steady flow of the folding fluid from the unfolded states of the protein to its native state. For the equilibrium conditions, when the protein repeatedly folds and unfolds, the segments of the equilibrium trajectory between these states can be selected for the first-passage paths from the unfolded to native state, as, for example, in the transition-path theory (TPT) [@E_Vanden-Eijnden06; @E_Vanden-Eijnden10]. The hydrodynamic approach has been successfully applied to the study of folding dynamics of several model proteins - an $\alpha$-helical hairpin (a lattice model) [@ChekmarevPalyanovKarplus08], a SH3 domain (a C$_{\alpha}$ model) [@KalginKarplusChekmarev09; @KalginChekmarev11], and beta3s-miniprotein (all-atom simulations) [@KalginCaflischChekmarevKarplus13]. It has been found that although the general behavior of the folding flow is consistent with the FES landscape, i.e., the flow is directed from the unfolded states to the native state and mostly concentrated at the bottom of the valley that connects these states, the distribution of the local flows is very different from what the FES suggests [@ChekmarevPalyanovKarplus08; @KalginKarplusChekmarev09]. Moreover, local flow vortices can be formed that do not necessarily leave fingerprints on the FES [@ChekmarevPalyanovKarplus08; @KalginKarplusChekmarev09; @KalginChekmarev11; @KalginCaflischChekmarevKarplus13], and such vortical flows can have many properties of turbulent flows of a fluid [@KalginChekmarev11]. In other words, the FES does not present a true potential for folding flows. The knowledge of the local flows offers a possibility to determine such a potential on the basis of the Helmholtz decomposition [@ArfkenWeber95] of the vector field of the flows. In the present paper, considering the first-passage folding of a model $\beta$-hairpin protein taken as an example, we show that this potential intrinsically has two components. One component is responsible for the source and sink of the folding flow and the other for the canalization of the flow between the source and sink. The paper is organized as follows. Section 2 describes the protein model, the simulation method, the hydrodynamic approach, and the choice of collective variables. Also, it presents some results of the simulations to show that the present protein model leads to a representative picture of protein folding (the FES as a function of $f_{\mathrm {nat}}$ and $r_{\mathrm {g}}$, the mean first-passage time as a function of temperature, and the distribution of the first-passage times). Section 3 presents the main results of the work and their discussion: the FES and the vector flow field in the collective variables (3.1), the potential functions for folding fluxes (3.2), and the interpretation of these functions (3.3). Section 4 summarizes the results of the work. Protein Model, Simulation Method and Preliminary Results ======================================================== Determining the potential with the Helmholtz decomposition requires a smooth vector field of folding flows and, accordingly, a large number of folding trajectories to be simulated. To perform simulations at a reasonable cost, we considered a short fast-folding protein in the framework of a minimalist model. The goal was to have a representative picture of the folding dynamics rather than to describe folding of a particular protein. Specifically, a twelve-residue $\beta$-hairpin protein 2evq [@2evq] (Fig. \[native\]) was chosen as a model system, and the interactions between the residues were characterized using a C$_{\alpha}$-based Gō-like model, i.e., the residues were represented by monomers (beads) centered at the C$_{\alpha}$ atoms, with the interaction between the monomers determined by the structure of the native state of the protein [@Go83]. Specifically, the interaction potential of Ref. was used $$\begin{aligned} \nonumber U=\sum_{1}^{N-1}[k_{1}(r_{i,i+1}-d_{0})^{2}+k_{2}(r_{i,i+1}-d_{0})^{4}]\\+\sum_{i+1<j}^{\mathrm{NAT}} 4 \varepsilon \left[\left(\frac{\sigma_{ij}}{r_{ij}}\right)^{12}-\left(\frac{\sigma_{ij}}{r_{ij}}\right)^{6} \right] \nonumber \\+\sum_{i+1<j}^{\mathrm{NON}} 4 \varepsilon \left[\left(\frac{\sigma_{0}}{r_{ij}}\right)^{12}-\left(\frac{\sigma_{0}}{r_{ij}}\right)^{6} +\frac{1}{4} \right]\Delta(r_{ij}-d_{\mathrm{nat}}) \nonumber\end{aligned}$$ where $N$ is the number of monomers (residues), $r_{ij}$ is the current distance between two monomers $i$ and $j$, and $\varepsilon$ is the characteristic attractive energy. The first term represents rigidity of the backbone; here $d_{0}=3.8$, $k_{1}=\varepsilon/{\text{\AA}}^2$, and $k_{2}=100 \varepsilon/{\text{\AA}}^4$. The second and third terms describe contributions of the native and non-native contacts, respectively. The contact between two monomers $i$ and $j$ was considered as a native contact if $|j-i|>1$ and the distance between these monomers in the native state was less than 7.5. In the second term $\sigma_{ij}=2^{1/6}d_{ij}$, where $d_{ij}$ is the distance between monomers $i$ and $j$ in the native state, and in the third term $d_{\mathrm{nat}}=\langle d_{ij}\rangle$, $\sigma_{0}=2^{1/6}d_{\mathrm{nat}}$, and $\Delta(r_{ij}-d_{\mathrm{nat}})$ is the cutoff function which is equal to 1 for $r_{ij}<d_{\mathrm{nat}}$ and 0 otherwise. The number of native contacts is equal to 27. The simulations were performed using a constant-temperature molecular dynamics based on the Langevin equation $$\label{eq1}m\frac{d^{2}\mathbf{r}_{i}}{dt^{2}}+\gamma\frac{d\mathbf{r}_{i}}{dt}=-\frac{\partial U}{\partial \mathbf{r}_{i}}+\mathbf{\Phi}_{i}(t)$$ where $\mathbf{r}_{i}$ is the radius-vector of $i$ monomer representing a residue, $m$ is the monomer mass, $U$ is the potential energy of the system, $\mathbf{\Phi}_{i}$ is a random force from the surroundings (a solvent that is not considered explicitly), and $\gamma$ is the friction coefficient that introduces viscosity of the surroundings to balance the random force and dissipation. The random forces have the Gaussian distribution with zero mean and variance $\langle \Phi_{i}^{j}(t)\Phi_{i'}^{j'}(t+\tau) \rangle=2\gamma k_{\mathrm{B}}T\delta_{ii'}\delta_{jj'}\delta(\tau)$, where the angular brackets denote an ensemble average, the upper index at $\Phi$ stands for the vector component, and $\delta_{kk'}$ and $\delta (\tau)$ are the Kronecker and Dirac delta functions, respectively. The equation was numerically integrated using the algorithm of Ref. with the time step $\Delta t=0.005\tau$ and $\gamma=3m/\tau$. With the length scale $l=7.5$ and the attractive energy $\varepsilon \sim 1{\mathrm{kcal/mol}}$, the characteristic time scale $\tau=(ml^2/\varepsilon)^{1/2}$ is $\sim 1$ps, where $m$=110 Da. To characterize protein conformations, we employed a set of the bond distances between the monomers which are not immediate neighbors along the protein chain; they formed a 55-dimensional conformation space. Using the principal component analysis (PCA) [@Jolliffe02], this space was reduced to a two-dimensional space of collective variables $\mathbf{g}=(g_1,g_2)$. The variable $g_1$ was directed along the eigenvector for the largest eigenvalue, and the variable $g_2$ along the vector calculated as the linear combination of the rest of the eigenvectors, which contributed with the weights corresponding to their eigenvalues. Although the second principal component could also be chosen as $g_2$, because the first two eigenvalues are well separated from the others (Fig. \[eigenvalues\]), the present choice makes it possible to take into account the effects missed by the first component. The hydrodynamic description of protein folding [@ChekmarevPalyanovKarplus08] is based on the calculation of the transitions in a space of collective variables (the space $\mathbf{g}$ in the present case). These transitions are organized into transition probability local flows (fluxes) $\mathbf{j(g)}$. In the case of two variables, $\mathbf{g}=(g_1, g_2)$, the $g_1$-component of the flux at a point $\mathbf{g}$ is determined as $$\begin{aligned} \label{eq2}j_{g_1}({\bf g})=[\sum^{{g_1}''-{g_1}'>0}_{{\bf g'},{\bf g''}({\bf g} \subset {\bf g}^\ast)}n({\bf g''},{\bf g'}) \nonumber \\ -\sum^{{g_1}''-{g_1}'<0}_{{\bf g'},{\bf g''}({\bf g} \subset {\bf g}^\ast)}n({\bf g''},{\bf g'}) ]/(M\bar{t}_{\mathrm{f}})\end{aligned}$$ where $M$ is the total number of simulated trajectories, $\bar{t}_{\mathrm{f}}$ is the mean first-passage time (MFPT), $n({\bf g''},{\bf g'})$ is the number of transitions from state ${\bf g'}$ to ${\bf g''}$, and ${\bf g} \subset {\bf g}^\ast$ is a symbolic designation of the condition that the transitions included in the sum have the straight line connecting points ${\bf g'}$ to ${\bf g''}$, which crosses the line $g_{1}={\mathrm {const}}$ within the segment of the length of $\Delta g_2$ centered at the point ${\bf g}$. The $g_2$-component of ${\bf j}(\bf{g})$ is determined in a similar way, except that one selects the transitions crossing the line ${g_2}={\mathrm {const}}$. The calculations were performed on a grid with discretization $\Delta g_1=\Delta g_2=0.12$. There were simulated $1 \times 10^4$ folding trajectories at a temperature equal to $T=0.17$ (in units $\varepsilon/k_{\mathrm{B}}$, where $k_{\mathrm{B}}$ is the Boltzmann constant). Each trajectory started at an unfolded state of the protein and was terminated upon reaching the native state. The unfolded states were prepared by thermalization of the native conformation at $T=0.5$ for $1 \times 10^4$ time steps; if the number of native contacts did not exceed 4, the conformation was accepted, if exceeded, the thermalization was continued. The native state was considered to be reached when the root-mean square deviation (RMSD) of the protein conformation from the native conformation was less than 1. The simulations have shown that the given protein model provides a sufficiently representative picture of protein folding, both from the “thermodynamic” and kinetic viewpoints. The FES calculated as a function of the fraction of native contacts and the radius of gyration is “L-shaped”, which is characteristic of a wide family of proteins in which an early collapse is observed [@SheaBrooks01], including beta-proteins [@PandeRokhsar99; @DinnerLazaridisKarplus99] (Fig. \[fes\_nat\_rg\]), the dependence of the MFPT on the temperature exhibits the well-known U-shape behavior found in experiments and simulations [@OlivebergYanFersht95; @Karplus97; @ChekmarevKrivovKarplus05] (Fig. \[mfpt\]), and the folding kinetics are essentially single-exponential at the “optimal” folding temperature at which the MFPT has the minimum value [@ChekmarevKrivovKarplus05] (Fig. \[ftd\]). Results and Discussion ====================== Free Energy Surface and Folding Flow Field in the PCA Variables --------------------------------------------------------------- The FES as a function of the collective PCA variables $g_1$ and $g_2$ is shown in Fig. \[fes\]. It was calculated similar to the above shown FES in “physical” variables (Fig. \[fes\_nat\_rg\]), i.e., using the equation similar to Eq. (\[eq0\]). The FES has a characteristic “bean-like” shape typically observed if one employs the PCA coordinates as the collective variables, e.g., Refs. and . At the same time, the present FES retains two essential properties of the FES of Fig. \[fes\_nat\_rg\], in that the global landscape of the surface has the form of a valley connecting the unfolded states and the native state, and there exists a free energy barrier separating these states. Figure \[arrows\]$\bf{a}$ depicts the distribution of $\mathbf{j(g)}$ in the form of vector field (for illustrative purpose, the lengths of the vectors are equally increased by factor $3.5 \times 10^2$). As is seen from the comparison of Fig. \[arrows\]$\bf{a}$ with Fig. \[fes\], the flow field lies within the free energy valley that connects the unfolded states to the native state, and the flow is concentrated at the bottom of the valley. Integration of the $g_1$-component of $\mathbf{j(g)}$ over $g_2$ in each cross-section $g_1={\mathrm{const}}$ shows that the total flow from the unfolded states to the native state, $G(g_1)=\int j_{g_1}(g_1,g_2)dg_2$, is essentially constant in the region between the source and sink of the flow (Fig. \[discharge\]). Figure \[div\_vor\] shows how the divergence of the folding flow ($q=\partial j_{g_1}/\partial g_1+\partial j_{g_2}/\partial g_2$, panel [**a**]{}) and its vorticity ($\omega=\partial j_{g_2}/\partial g_1-\partial j_{g_1}/\partial g_2$, panel [**b**]{}) are distributed in the $(g_1,g_2)$ space. Except for two localized regions at the unfolded and native states, which represent, respectively, the source and sink of the folding flow, the flow is divergence-free. The vorticity, in contrast, spans the entire flow field. It arises because the intensity of the flow decreases towards both sides of the free energy valley (Fig. \[arrows\]$\bf{a}$); correspondingly, the vorticity has different signs on the different sides of the valley (see the above equation for vorticity). This decrease of the flow intensity towards the valley sides is a natural phenomenon, because the lower the probability to visit some region of the conformation space (in the present case, a side of the valley), the smaller the flows in this region; similar nonuniform distributions of the flows have been previously observed for an $\alpha$-helical hairpin [@ChekmarevPalyanovKarplus08] and SH3 domain [@KalginKarplusChekmarev09]. Therefore, the vorticity generated on the periphery of the folding flow field presents an intrinsic property of the folding dynamics. Comparison of Fig. \[arrows\]$\bf{a}$ to Fig. \[fes\] shows that the folding flux distribution is consistent with the FES landscape in that the overall folding flow follows the valley of the FES that connects the unfolded states to the native state. At the same time, the corresponding free energy $F(\mathbf{g})$ does not present a true potential of folding fluxes, i.e., determining the fluxes as $j_{g_1}=-\partial F(\mathbf{g})/\partial g_1$ and $j_{g_2}=-\partial F(\mathbf{g})/\partial g_2$ leads to a vector flow field drastically different from that of Fig. \[arrows\]$\bf{a}$ (Fig. \[arrows\_from\_fes\]). Calculation of the corresponding total flow shows that the flow is not constant but fluctuates around zero value (Fig. \[discharge\]). Potentials for Folding Fluxes ----------------------------- To determine an actual potential for the folding fluxes, we use the Helmholtz decomposition theorem [@ArfkenWeber95], according to which any smooth vector field can be uniquely represented as a sum of two terms: a curl-free field and a divergence-free field. Then $$\label{eq3} \mathbf{j}={\mathbf{j}}_{\mathrm{cf}}+{\mathbf{j}}_{\mathrm{df}}$$ where ${\mathbf{j}}_{\mathrm{cf}}$ is the curl-free component, and ${\mathbf{j}}_{\mathrm{df}}$ is the divergence-free component, i.e., $\nabla \times {\mathbf{j}}_{\mathrm{cf}}=0$ and $\nabla \cdot {\mathbf{j}}_{\mathrm{df}}=0$, respectively. These conditions allow introducing the potentials of the vector fields. In the case of two dimensions, vectors ${\mathbf{j}}_{\mathrm{cf}}$ and ${\mathbf{j}}_{\mathrm{df}}$ can be written as $$\begin{aligned} \label{eq3_1} {\mathbf{j}}_{\mathrm{cf}}=-\frac{\partial \Phi}{\partial g_1}{\mathbf{k}}_{1}-\frac{\partial \Phi}{\partial g_2}{\mathbf{k}}_{2}\\ \nonumber {\mathbf{j}}_{\mathrm{df}}=\frac{\partial \Psi}{\partial g_2}{\mathbf{k}}_{1}-\frac{\partial \Psi}{\partial g_1}{\mathbf{k}}_{2}\end{aligned}$$ where $\Phi=\Phi({\mathbf{g}})$ is the potential for the curl-free component, $\Psi=\Psi({\mathbf{g}})$ is the potential for the divergence-free component, and ${\mathbf{k}}_{1}$ and ${\mathbf{k}}_{2}$ are the unit vectors for the variables $g_1$ and $g_2$, respectively. The sets of the equipotential lines $\Phi(\mathbf{g})=\mathrm{const}$ and $\Psi(\mathbf{g})=\mathrm{const}$ are not mutually orthogonal because $\nabla \Psi \cdot \nabla \Phi=\partial \Psi/\partial g_1 \cdot \partial \Phi/\partial g_1 + \partial \Psi/\partial g_2 \cdot \partial \Phi/\partial g_2$ is generally nonzero. This is the effect of the flow vorticity: if the flow were both divergence- and curl-free, the flow flux ${\mathbf{j}}$ could be written in either form of Eq. (\[eq3\_1\]). Then, the potential functions $\Phi({\mathbf{g}})$ and $\Psi({\mathbf{g}})$ would satisfy the relations $\partial \Psi/\partial g_2=-\partial \Phi/\partial g_1$ and $\partial \Psi/\partial g_1=\partial \Phi/\partial g_2$ characteristic of the potential flows (the Cauchy-Riemann conditions) [@LandauLifshitz87], in which case $\nabla \Psi \cdot \nabla \Phi=0$. Substituting the given expressions for ${\mathbf{j}}_{\mathrm{cf}}$ and ${\mathbf{j}}_{\mathrm{df}}$ into Eq. (\[eq3\]) and regrouping the terms, one obtains $$\nonumber \mathbf{j}=j_{g_1}{\mathbf{k}}_{1}+j_{g_2}{\mathbf{k}}_{2}$$ where $$\label{eq4} j_{g_1}=-\frac{\partial \Phi}{\partial g_1}+\frac{\partial \Psi}{\partial g_2}, \hspace{0.5cm} j_{g_2}=-\frac{\partial \Phi}{\partial g_2}-\frac{\partial \Psi}{\partial g_1}$$ To find functions $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$, the functional $$\begin{aligned} \label{eq5} Q=\int [\left(j_{g_1}+\frac{\partial \Phi}{\partial g_1}-\frac{\partial \Psi}{\partial g_2}\right)^2 \nonumber \\+\left(j_{g_2}+\frac{\partial \Phi}{\partial g_2}+\frac{\partial \Psi}{\partial g_1}\right)^2]dg_1dg_2\end{aligned}$$ was minimized with respect to $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$; here $j_{g_1}$ and $j_{g_2}$ are the folding fluxes obtained with the molecular dynamics simulations. Integration in Eq. (\[eq5\]) was performed numerically on a grid with the same discretization as for Eq. (\[eq2\]). To avoid boundary effects, the region of integration was extended to $-21.8 \le g_1 \le 7.0$ and $-13.7 \le g_2 \le 12.7$ ([*[cf.]{}*]{} Fig. \[fes\]). At the boundaries of the region, no-flux conditions were imposed, i.e., at the left and right boundaries ($g_1=\mathrm{const}$) it was assumed $j_{g_1}=0$, and at the lower and upper boundaries ($g_2=\mathrm{const}$) $j_{g_2}=0$. The minimization was performed using the steepest-descent method with a variable step-size. Starting with $\Phi(\mathbf{g})=\Psi(\mathbf{g})=0$ as an initial guess for these functions, in which case $Q \approx 1 \times 10^{-3}$, the process was continued until $Q$ was as small as $\approx 1 \times 10^{-11}$, i.e., even in the worst case, when the deviation of the fluxes determined by Eqs. (\[eq4\]) from the fluxes obtained in the simulations was concentrated at a single point, it would not exceed $\sim 0.01 \%$. The results of the calculations are shown in Fig. \[phi\_psi\]. In agreement with the Helmholtz decomposition [@ArfkenWeber95], the functions $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ are characteristically different in that the former accounts for the source and sink of the flows, and the latter for the vorticity effects. More specifically, as can be seen from comparison of Eq. (\[eq4\]) with Fig. \[phi\_psi\], the function $\Phi(\mathbf{g})$ determines the intensity of the flow in the vicinity of the source and sink, and the function $\Psi(\mathbf{g})$ determines it in the region between the source and sink, providing the canalization of the flow within the free energy valley that connects the unfolded states to the native state. Thus, both functions, $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$, are equally important for a correct description of the folding dynamics and should be considered as the intrinsic components of the complex potential that determines the driving force. Interpretation of the Potentials -------------------------------- In hydrodynamic terms [@LandauLifshitz87], the functions $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ can be associated, respectively, with the scalar potential of the flow and the vector potential, or more specifically, with the component of the vector potential that is orthogonal to the two-dimensional plane under consideration. The latter plays a role of the stream function, for which the equation $\Psi(\mathbf{g})=\mathrm{const}$ determines the streamline of the flow, i.e., the line that is tangent to the local directions of the $\mathbf{j(g)}$ vectors (Fig. \[arrows\]$\bf{b}$). The difference between the stream functions for two streamlines determines the fraction of the total flow in the stream tube between the streamlines (see also Refs. and ). Similar to a steady potential flow of an inviscid fluid in a two-dimensional space [@LandauLifshitz87], the functions $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ can also be written as a complex function $\Theta({\mathbf{g}})=\Phi(\mathbf{g})+i\Psi(\mathbf{g})$, where $i$ is the imaginary unit; then ${\mathbf{j}}({\mathbf{g}})=j_{g_1}+ij_{g_2}=-\nabla_{c} \Theta({\mathbf{g}})$, where $\nabla_{c}=\partial/\partial g_1+i\partial/\partial g_2$ is the gradient operator in the complex number space $(g_1,ig_2)$. However, in contrast to the potential flow of inviscid fluid, the function $\Theta({\mathbf{g}})$ is not [*analytic*]{} function, i.e., is not a function of complex variable $g_1+ig_2$, because the Cauchy-Riemann conditions, which represent the necessary and sufficient condition for a function to be analytic and require the flow to be divergence- and curl-free [@ArfkenWeber95], are not satisfied (Fig. \[div\_vor\]). Therefore, the methods of the theory of analytic functions of complex variable, which are successfully used in hydrodynamics [@LandauLifshitz87], are not applicable here. Another interpretation of the potentials $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ stems from the kinetic theory [@Kampen81]. According to Eq. (\[eq4\]), the velocity of motion is proportional to the forces produced by these potentials, i.e., the local flows are assumed to be [*drift*]{} flows. In other words, although the inertia of the monomers was present in the Langevin equation we used for the simulations, the potentials $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ determined from the resulted fluxes $j_{g_1}$ and $j_{g_2}$ account for an [*overdamped*]{} motion. The approximation of the overdamped motion, which neglects the inertia term in the Langevin equation (\[eq1\]), is rather common to characterize the protein folding dynamics [@RohrdanzZhengClementi13]. The corresponding kinetic equation is the Smoluchowski equation $\partial p/\partial t+\nabla \cdot \mathbf{J}=0$, where $p(\mathbf{g},t)$ is the probability density, $\nabla$ is the gradient operator in the coordinate space $\mathbf{g}$, and $\mathbf{J}$ is the probability current, which can be written as $\mathbf{J}=-D(\mathbf{g})\nabla p+D(\mathbf{g})/(k_{\mathrm B}T){\mathbf F}(\mathbf{g})p$, where $-D(\mathbf{g})\nabla p$ is the diffusion flux, $D(\mathbf{g})/(k_{\mathrm B}T){\mathbf F}(\mathbf{g})p$ is the drift flux, $D(\mathbf{g})$ is the diffusion tensor, and ${\mathbf F}(\mathbf{g})$ is the driving force. If the probability current were zero, i.e., detailed balance existed, the equality $\mathbf{J}=0$ would give the [*equilibrium*]{} (Boltzmann) distribution $p(\mathbf{g}) \sim \exp[-G(\mathbf{g})]/k_{\mathrm B}T$, where $G(\mathbf{g})$ is the free energy that exerts the driving force ${\mathbf F}(\mathbf{g})=-\nabla G(\mathbf{g})$. This case would correspond to a curl-free drift flow, with the potential of the driving force being determined by a single function in the form of the free energy $G(\mathbf{g})$. However, if the probability current is nonzero, as in the present case, i.e., when the steady flow from the unfolded state to the native state exists, the stationary solution is determined by the condition $\nabla \cdot \mathbf{J}=0$, so that $\mathbf{J}$ can have a curl component [@Kampen81]. Such flow is [*non-equilibrium*]{} and is characterized by “irreversible circulation” or “cyclic balance”, which can be considered as a measure of deviation from detailed balance [@TomitaTomita74; @Graham77; @EyinkLebowitzSpohn96]. In our case the circulating flow is represented by the flux vector ${\mathbf{j}}_{\mathrm{df}}$ and, according to Eq. (\[eq3\_1\]), is generated by the potential $\Psi(\mathbf{g})$, with the factor $D(\mathbf{g})/(k_{\mathrm B}T)$ being included into the potential. In general case, the driving force can be written as $\mathbf{F}=-\nabla \Phi+\nabla \times \mathbf{A}$, where $\Phi$ and $\mathbf{A}$ are the scalar and vector potentials, respectively (in the case of two dimensions, only the component of $\mathbf{A}$ that is orthogonal to the plane is involved). The potentials of this type have recently been used to study the dynamics of Brownian particles in corrugated channels [@MartensStraubeSchmidSchimansky-GeierHaanggi13; @AiHeLiZhong13]. As an anonymous reviewer of the manuscript remarked, the resulting picture of the folding process in terms of the $\Psi$ and $\Phi$ potentials is very similar to the picture obtained for the reaction dynamics in the TPT [@E_Vanden-Eijnden06], although the approaches are apparently different. In the latter, the committor probability functions are calculated for the transition paths from the reactant to the product, and the spatial distribution of the probability currents of the reaction paths (fluxes) between these states is determined [@E_Vanden-Eijnden10]. The currents are normal to the isocommittor surfaces and organized in the form of the reaction tubes that connect the reactant and product wells, so that the boundaries of the tubes represent the flow lines (streamlines) of the currents. Then the equipotential lines $\Psi(\mathbf{g})=\mathrm{const}$ and $\Phi(\mathbf{g})=\mathrm{const}$ (Fig. \[phi\_psi\_2d\]) can be associated with the flow lines and the isocommittor surfaces, respectively. In particular, the line $\Phi(\mathbf{g})=0$, shown by the dashed curve, which separates the source and sink regions of the flow, can play a role of the isocommittor surface of probability of 1/2. Very interesting from physical and methodological viewpoints, this analogy, however, does not probably extend beyond qualitative resemblance because, in contrast to the flow lines and the isocommittor surfaces, the sets of the equipotential lines $\Phi(\mathbf{g})=\mathrm{const}$ and $\Psi(\mathbf{g})=\mathrm{const}$ are not mutually orthogonal (as has been previously indicated). As is seen from Fig. \[phi\_psi\_2d\], they possess the orthogonality property only approximately. Conclusions =========== In summary, we have shown that the potential of the driving force for protein folding in a two-dimensional space of collective variables has two components. One component accounts for the source and sink of the folding flow, representing, respectively, the unfolded and native states of the protein, and the other accounts for the vorticity generated at the periphery of the flow field and provides the canalization of the flow within the free energy valley that connects the unfolded and native states. Since the Helmholtz decomposition theorem [@ArfkenWeber95] guarantees that any vector field can be uniquely represented by a sum of divergence-free and curl-free fields, the present approach is equally applicable to more complex folding dynamics that are characterized by multiple reaction channels and formation local vortices of the folding flows, as, for example, for folding of SH3 domain [@KalginKarplusChekmarev09; @KalginChekmarev11]. In this case, the functions $\Phi(\mathbf{g})$ and $\Psi(\mathbf{g})$ will be not as regular as in Fig. \[phi\_psi\] because they should take into account this additional complexity of the folding flow, but the general pattern of these functions is expected to be preserved because of the presence of the source and sink of the folding flow and its vorticity inherently generated at the periphery of the flow field. The requirement of smoothness of the vector field, which is necessary for the Helmholtz decomposition [@ArfkenWeber95], currently restricts the possibilities of analysis of folding of proteins of practical interest on atomic level of resolution, however, for small proteins, such, e.g., as the beta3s miniprotein, the calculation of smooth vector fields is quite feasible [@KalginCaflischChekmarevKarplus13]. Also note that the application of the present approach is not limited to the case of the folding reaction. Similar complex potentials can be expected for other systems that involve multiple reaction pathways, e.g., for nanoclusters, in which the transition between the structures of alternative morphology is characterized by a multi-funnel energy landscape [@Wales03]. Acknowledgments =============== I would like to thank an anonymous reviewer for the remark on the similarity between the hydrodynamic and TPT pictures of the reaction. J. D. Bryngelson and P. G. Wolynes, Proc. Natl. Acad. Sci. USA **84**, 7524 (1987). E. Shakhnovich, G. Farztdinov, A. M. Gutin, and M. Karplus, Phys. Rev. Lett. **67**, 1665 (1991). P. E. Leopold, M. Montal, and J. N. Onuchic, Proc. Natl. Acad. Sci. USA **89**, 8721 (1992). J. D. Bryngelson, J. N. Onuchic, N. D. Socci, and P. G. Wolynes, Proteins: Structure, Function, and Bioinformatics **21**, 167 (1995). K. A. Dill and H. S. Chan, Nat. Struct, Biol. **4**, 10 (1997). E. I. Shakhnovich, Curr. Opin. Struct. Biol. **7**, 29 (1997). J. N. Onuchic, Z. Luthey-Schulten, P. G. Wolynes, Annu. Rev. Phys. Chem. **48**, 545 (1997). C. M. Dobson, A. Sali, and M. Karplus, Angew. Chem. Int. Ed. **37**, 869 (1998). A. R. Dinner, A. Sali, L. J. Smith, C. M. Dobson, and M. Karplus, Trends Biochem. Sci. **25**, 331 (2000). J.-E. Shea and C. L. Brooks III, Annu. Rev. Phys. Chem. **52**, 499 (2001). B. Schuler, E. A. Lipman, and W. A. Eaton, Nature **419**, 743 (2002). M. Gruebele, Curr. Opinion Struct. Biol. **12**, 161 (2002). C. C. Mello and D. Barrick, Proc. Natl. Acad. Sci. USA **101**, 14102 (2004). M. Karplus, Nat. Chem. Biol. [**7**]{}, 401 (2011). K. A. Dill, S. B. Ozkan, M. S. Shell, and T. R. Weikl, Annu. Rev. Biophys. **37**, 289–316 (2008). K. A. Dill and J. L. MacCallum, Science **338**, 1042 (2012). A. V. Finkelstein and O. Ptitsyn, [*Protein Physics: A Course of Lectures*]{} (Academic Press: London, 2002). A. Berezhkovskii and A. Szabo, J. Chem. Phys. **122**, 014503 (2005). Y. M. Rhee and V. S. Pande, J. Phys. Chem. B **109**, 6780 (2005). R. B. Best and G. Hummer, Proc. Natl. Acad. Sci. U.S.A. **102**, 6732 (2005). S. V. Krivov and M. Karplus, J. Phys. Chem. B **110**, 12689 (2006). D. A. C. Beck and V. Daggett, Biophys. J. **93**, 3382 (2007). J. D. Chodera and V. S. Pande, Phys. Rev. Lett. **107**, 098102 (2011). B. Peters, P. G. Bolhuis, R. G. Mullen, and J.-E. Shea, J. Chem. Phys. **138**, 054106 (2013). W. E and E. Vanden-Eijnden, Annu. Rev. Phys. Chem. **61**, 391 (2010). E. M. Boczko and C. L. Brooks III, Science **269**, 393 (1995). N. D. Socci, J. N. Onuchic, and P. G. Wolynes, Proteins: Structure, Function, and Genetics **32**, 136 (1998). S. F. Chekmarev, S. V. Krivov, and M. Karplus, J. Phys. Chem. B [**109**]{}, 5312 (2005). S. F. Chekmarev, A. Yu. Palyanov, and M. Karplus, Phys. Rev. Lett. [**100**]{}, 018107 (2008). C. Schütte, Habilitation thesis, Department of Mathematics and Computer Science, Freie Universitat Berlin, 1999. W. C. Swope, J. W. Pitera, and F. Suits, J. Phys. Chem. B [**108**]{}, 6571 (2004). J. K. Weber and V. S. Pande, J. Chem. Theory Comput. [**7**]{}, 3405 (2011). O. M. Becker and M. Karplus, J. Chem. Phys. [**106**]{}, 1495 (2007). S. V. Krivov and M. Karplus, Proc. Natl. Acad. Sci. U.S.A. **101**, 14766 (2004). D. J. Wales, [*Energy Landscapes: Applications to Clusters, Biomolecules and Glasses*]{} (Cambridge University Press: Cambridge, 2003). F. Rao and A. Caflisch, J. Mol. Biol. [**342**]{}, 299 (2004). F. Noé, C. Schütte, E. Vanden-Eijnden, L. Reich, and T. R. Weikl, Curr. Opin. Struct. Biol. [**106**]{}, 19011 (2009). G. R. Bowman, V. A. Voelz, and V. S. Pande, Proc. Natl. Acad. Sci. USA [**21**]{}, 4 (2011). I. V. Kalgin, A. Caflisch, S. F. Chekmarev, and M. Karplus, J. Phys. Chem. B [**117**]{}, 6092 (2013). L. D. Landau and E. M. Lifshitz, [*Fluid Mechanics*]{} (Pergamon: New York, 1987). I. V. Kalgin, M. Karplus, and S. F. Chekmarev, J. Phys. Chem. B [**113**]{}, 12759 (2009). I. V. Kalgin and S. F. Chekmarev, Phys. Rev. E [**83**]{}, 011920 (2011). W. E and E. Vanden-Eijnden, J. Stat. Phys. **123**, 503 (2006). G. B. Arfken and H. J. Weber, [*Mathematical Methods for Physicists*]{}, 4th ed. (Academic Press: San Diego, 1995). N. H. Andersen, K. A. Olsen, R. M. Fesinmeyer, X. Tan, F. M. Hudson, L. A. Eidenschink, and S. R. Farazi, J. Am. Chem. Soc. **128**, 6101 (2006). N. Gō, Annu. Rev. Biophys. Bioeng. **12**, 183 (1983). T. X. Hoang and M. Cieplak, J. Chem. Phys. [**112**]{}, 6851 (2000). R. Biswas and D. R. Hamann, Phys. Rev. B [**34**]{}, 895 (1986). I. T. Jolliffe, [*Principal Component Analysis*]{}; 2nd ed. (Springer: New York, 2002). V. S. Pande and D. S. Rokhsar, Proc. Natl. Acad. Sci. USA **94**, 9062 (1999). A. R. Dinner, T. Lazaridis and M. Karplus, Proc. Natl. Acad. Sci. USA **96**, 9068 (1999). M. Oliveberg, Y.-J. Yan, and A. R. Fersht, Proc. Natl. Acad. Sci. U.S.A. [**92**]{}, 8926 (1995). M. Karplus, Fold. Des. [**2**]{}, 569 (1997). Y. Levy, J. Jortner, and O. M. Becker, Proc. Natl. Acad. Sci. U.S.A. **98**, 2188 (2001). N. G. van Kampen, [*Stochastic Processes in Physics and Chemistry*]{} (North-Holland: Amsterdam, 1981), p. 294. M. A. Rohrdanz, W. Zheng, and C. Clementi, Annu. Rev. Phys. Chem. **64**, 295 (2013). K. Tomita and H. Tomita, Prog. Theor. Phys. **51**, 1731 (1974). R. Graham, Z. Phys. B **26**, 397 (1977). G. L. Eyink, J. L. Lebowitz, and H. Spohn, J. Stat. Phys. **83**, 385 (1996). S. Martens, A. V. Straube, G. Schmid, L. Schimansky-Geier, and P. Haänggi, Phys. Rev. Lett. **110**, 010601 (2013). B.-Q. Ai, Y.-F. He, F.-G. Li, and W.-R. Zhong, J. Chem. Phys. **138**, 154107 (2013).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Optical communication systems, operating in C–band, are reaching their theoretically achievable capacity limits. An attractive and economically viable solution to satisfy the future data rate demands is to employ the transmission across the full low–loss spectrum encompassing O, E, S, C and L band of the single mode fibers (SMF). Utilizing all five bands offers a bandwidth of up to $\sim$53.5 THz (365 nm) with loss below 0.4 dB/km. A key component in realizing multi–band optical communication systems is the optical amplifier. Apart from having an ultra–wide gain profile, the ability of providing arbitrary gain profiles, in a controlled way, will become an essential feature. The latter will allow for signal power spectrum shaping which has a broad range of applications such as the maximization of the achievable information rate $\times$ distance product, the elimination of static and lossy gain flattening filters (GFF) enabling a power efficient system design, and the gain equalization of optical frequency combs. In this paper, we experimentally demonstrate a multi–band (S+C+L) programmable gain optical amplifier using only Raman effects and machine learning. The amplifier achieves $>$1000 programmable gain profiles within the range from 3.5 to 30 dB, in an ultra–fast way and a very low maximum error of $1.6 \cdot 10^{-2}$ dB/THz over an ultra–wide bandwidth of 17.6–THz (140.7–nm).' author: - 'Uiara Celine de Moura, Md Asif Iqbal, Morteza Kamalian, Lukasz Krzczanowicz, Francesco Da Ros, Ann Margareth Rosa Brusin, Andrea Carena, Wladek Forysiak, Sergei Turitsyn, and Darko Zibar [^1] [^2] [^3]' bibliography: - 'IEEEabrv.bib' - 'bibliography.bib' title: 'Multi–band programmable gain Raman amplifier' --- optical communications, multi–band systems, optical amplifiers, machine learning, neural networks. Introduction {#sec:intro} ============ the past two decades, a great evolution of optical communication systems, in terms of spectral efficiency$\times$distance product, has been enabled by the advances in digital coherent detection. So far, most of the efforts, on reaching the capacity of the nonlinear fiber–optic channel, have been focusing on the C–band only [@agrell2016roadmap]. However, squeezing the information inside this transmission window will soon reach its theoretical limit [@Dar14]. To cope with the constant demand for higher throughput, novel solutions must be explored. Optical communication systems operating across multi–band transmission, are an attractive solution for providing the future capacity scaling. They can provide up to 10$\times$ higher capacity, compared to the C–band [@Ferrari20], on the already deployed SMF fiber infrastructure. To make multi–band systems commercially deployable in the near future, large research efforts in terms of components, system and network design are needed [@ciena; @infinera; @Doerr2016; @Timurdogan19; @Messner20; @Tummidi20; @Sarwar20; @Wang20]. One of the main challenges in realizing multi–band systems is the development of optical amplifiers that are able to provide sufficiently high gains over such a wide bandwidth. Additionally, a novel feature that may become essential is the ability to provide arbitrary gain profiles in a controlled and ultra–fast way. This is because different signal channels in a multi–band system are unevenly impacted by the interaction between the Kerr nonlinearity, amplified spontaneous emission (ASE) noise and stimulated Raman scattering (SRS) [@Ferrari20]. Consequently, for the maximization of the achievable information rate (AIR) $\times$ distance product, non–flat signal channel power profiles are needed. Depending on the system configuration, signal channel power profiles will be a result of a complex optimization and may assume arbitrary shapes. Moreover, to address the future requirements on high capacity optical networks, ultra–fast gain profile re–configurability is needed [@Napoli18]. A current and by far the most dominant approach for performing programmable signal channel power profile shaping is by leveraging the use of wavelength selective switches (WSSs) whose primary function is to router the signals throughout the optical network. However, this approach is highly power inefficient since it adjusts the channel powers by means of attenuation. A novel approach for realizing signal channel power shaping is by employing optical amplifiers with programmable (arbitrary) gain profiles. Such amplifiers could be a potential game changer as they would be able to simultaneously amplify the optical data signal and perform gain shaping. This has many impactful applications ranging from compensation of wavelength–dependent loss in devices such as modulators and frequency combs to gain–shaping in fixed-gain profile amplifiers. Especially, if integrated–combs are targeted for multi-channel sources, an efficient approach for gain shaping would be desirable. This is because for integrated–combs there is a large variation in power of their frequency components. Finally, optical amplifiers providing arbitrary gain profiles can be used in hybrid approaches to complement the gain, and overcome the limitations of other optical amplifier technologies [@Fukuchi01; @Gordienko16; @Ionescu19; @Galdino19; @Arnould20; @Ye20]. There are several approaches and technologies for realizing optical amplifiers covering multiple bands. To date, works on multi–band optical amplifiers have focused on: rare–earth–doped fiber amplifiers (xDFAs) covering 17.56 THz over O+E–band [@Wang20] and 10.7 THz over S+C–band [@Sakamoto06], semiconductor optical amplifiers (SOAs) for 12.7 THz on S+C+L–band [@Renaudier18], optic parametric amplifiers (OPAs) with 10 THz of bandwidth on S+C+L–band [@Kobayashi20], Raman amplifiers (RA) in combination with EDFAs, SOAs and OPAs achieving bandwidths ranging from 10.7 to 14 THz on C+L and S+C+L–band [@Fukuchi01; @Ye20; @Ionescu19; @Galdino19; @Arnould20; @Gordienko16], and pure RAs with bandwidths of up to 19.1–THz S+C+L–band [@Rottwitt99; @Zhou06; @Chen18; @Emori01; @Iqbal20]. So far, the majority of works in [@Fukuchi01; @Sakamoto06; @Wang20; @Renaudier18; @Kobayashi20; @Gordienko16; @Ionescu19; @Galdino19; @Arnould20; @Rottwitt99; @Zhou06; @Chen18; @Emori01; @Iqbal20] have focused on realizing flat gain profiles in C+L and S+C+L–band. Recently, an amplifier that relies on a hybrid SOA/Raman configuration has been demonstrated to achieve arbitrary loss/gain profile generation in S+C+L–band in 12.3 THz of bandwidth [@Ye20]. Among all different solutions, RAs are most suitable for realizing arbitrary gain profiles, in a controlled way. This is because the RAs allow for a flexible gain profile design by adjusting the pump powers and wavelengths, and provide gain availability across a broad range of wavelengths, when operated in multi-pump configurations. The challenge with Raman amplifier design is on the selection of pump powers and wavelengths that would result in a targeted gain profile. Several solutions to this optimization problem have been reported in the literature but have mainly focused on realizing flat gain profiles [@Ferreira11; @Zhou01; @Chen18; @Perlin02; @Iqbal20; @Mowla08; @Jiang10; @Emori01; @Ania07]. Recently, a machine learning framework for the ultra–fast configuration of the pump powers and wavelengths has been theoretically proposed and as a proof–of–principle experimentally demonstrated in C–band only [@Zibar20; @deMoura20]. The proposed approach can be used for the design of Raman amplifiers, where an arbitrary gain profile is achievable in a controlled way. However, moving from C–band to multi–band and realizing wider gain profiles is significantly more challenging. This is partly due to the increased number of pumps that need to be controlled and also the increased nonlinearity given the higher overall powers in the optical fiber. In this paper, we use the proposed machine learning framework for the experimental realization of multi–band RAs that can provide arbitrary gains, in a controlled way, in C+L and S+C+L–band. Up to 8 pumps are employed to provide more than 5000 arbitrary gain profiles over up to 17.6-THz of bandwidth. We achieve a highly–accurate programmable set of gain profiles with a very low average maximum error, (defined between the target and realized gain profiles), per bandwidth, $E_{MAX}/BW$, of $1.6 \cdot 10^{-2}$ dB/THz. Machine learning for broadband gain optimisation is a topic of growing interest, which is reflected in the recent work [@Ye20] reporting a root mean squared error per bandwidth, $RMSE/BW$, of 0.033 dB/THz in a 12.3 THZ bandwidth SOA/distributed Raman link scenario. In this work, we achieve a close to an order of magnitude lower $RMSE/BW$ of 0.0045 dB/THz for transmission with discrete S+C+L–band Raman amplifier over a larger bandwidth of 17.6 THz. The presented work is first time demonstration of a programmable multi–band discrete Raman amplifier setting a benchmark, to be beaten, for the maximum error per bandwidth. The structure of the paper is as following: Section \[sec:exp\_setups\] describes the experimental setup for realizing Raman amplifiers operating in C+L and S+C+L–band. We also give a brief overview of the ML framework used to obtain programmable arbitrary gain profiles. Section \[sec:results\] presents, discusses and evaluates the experimental results. In Section \[sec:conc\] conclusions and future work are presented and outlined. Experimental setup {#sec:exp_setups} ================== ![image](figures/SingleSetup_v4.pdf){width="100.00000%"} The experimental setup for realizing the multi–band RA is shown in Fig. \[fig:setup\](a). By selecting path 1 or 2, the operation in either C+L (1) or S+C+L–band (2) can be enabled. To achieve gains in the C+L and S+C+L–band, 5 and 8 pump lasers are employed, respectively. Fig. \[fig:setup\](c) illustrates the spectral pump allocation and their individual contribution to the overall Raman gain. We only consider counter propagating pumps whose wavelengths are fixed and shown in Table \[tab:pumps\]. $P_1$ $P_2$ $P_3$ $P_4$ ------------------- ------- ------- ------- ------- Wavelength \[nm\] 1508 1485 1465 1445 Frequency \[THz\] 198.8 201.9 204.6 207.5 $P_5$ $P_6$ $P_7$ $P_8$ Wavelength \[nm\] 1425 1405 1385 1365 Frequency \[THz\] 210.4 213.4 216.5 221.1 : **Pump lasers wavelengths and frequencies** \[tab:pumps\] The gain profile control is performed by only adjusting the pump powers. Pump lasers $P_1...P_7$ are semiconductor laser diodes. Their output power is controlled by adjusting the driving currents. The corresponding power going into the RA is in the range from $\sim$16 dBm to $\sim$27 dBm. Pump laser $P_8$ is a Raman–based fiber laser and is controlled by adjusting its voltage. It provides power to the RA ranging from $\sim$20 dBm to $\sim$27 dBm. C+L–band Raman amplifier ------------------------ The C+L–band RA can either be operated as a discrete, (7.5 km of inverse dispersion fiber (IDF)) or distributed (75 km span of standard SMF) amplifier. An input optical signal covering the C+L–band, for testing the performance of the RA, is generated by using two ASE sources for C and L bands channelized through a WSS to generate 90 lines placed at 100 GHz ITU-T grid covering a 9.4 THz (77 nm) bandwidth. The corresponding optical spectrum is shown in Fig. \[fig:setup\](b) (inside bracket 1) and is measured with a resolution of $\Delta \lambda = 0.1$ nm. The gaps between the C and L signal bands are due to the different ASE sources for these two bands. An isolator is placed at the input to the IDF to prevent pump powers entering the C+L–band signal source and to minimize the double Rayleigh backscattering induced multipath interference [@Iqbal19JLT]. Finally, an optical spectrum analyser (OSA) is used to capture the optical spectrum. S+C+L–band Raman amplifier -------------------------- The S+C+L–band RA is implemented as a two–stage sequential discrete RA. The first stage is responsible for providing the gain in the S–band and it consists of 7.5 km of IDF and three pump lasers, $P_6...P_8$ used to control the gain profiles. The second stage is the same as the one used for the C+L–band RA. Note that distributing the pumps into two sequential stages reduces the strong depletion of shorter wavelength pumps [@Krummrich01]. The multi–band input optical signal (17.6 THz/140.7 nm) is generated by combining the optical signal from the C+L–band with a supercontinuum S–band source [@El-Taher:s] and a single frequency laser operating at 185 THz. The resulting signal has a total of 148 frequency lines at 100 GHz ITU-T grid. The corresponding optical spectrum is shown in Fig. \[fig:setup\](b) (inside bracket 2). Due to the amplifier configuration, two pumps from the first stage ($P_{1-2}$) fall within the S-band signal. This means that some channels from the S–band need to be removed to avoid overlapping with the Rayleigh backscattered components of the pumps, leaving the gaps as shown in Fig. \[fig:setup\](b) [@Iqbal:20]. Pump power control ------------------ The objective is to determine pump power settings that result in user defined target gain profiles such as: tilted gain, flat gain or an arbitrary gain. These settings are achieved off–line using the machine learning framework presented and then later applied on–line for the pump laser currents and voltage control [@Zibar20]. As the framework in [@Zibar20] is based on supervised learning, a data–set is required. This is achieved by varying the currents and the voltage of the pump lasers and measuring the corresponding gain profiles. The gain profiles are measured on a 100-GHz grid, as the difference in power between the output optical spectrum when the pump lasers are turned on and off, also known as the on–off gain. As the currents and the voltage, $I_1,...I_7,V_8$, are drawn from a uniform distribution, we refer to the corresponding gain profiles as arbitrary. In Fig. \[fig:datasets\], the measured on–off gain profiles, $G$, obtained for the C+L and the S+C+L–band are shown. We measure $M=5600$ and $M=4025$ gain profiles, each with $K=90$ and $K=148$ data points per gain profile, for C+L and S+C+L–band, respectively. We denote the respective data–sets as: $\mathcal{D}^{M\times (K+5)}_{C+L}=\{(G_1^i,...,G_K^i,I^i_1,...I^i_5),|i=1,...,M\}$ and $\mathcal{D}^{M\times (K+8)}_{S+C+L}=\{(G_1^i,...,G_K^i,I^i_1,...I^i_7,V^i_8),|i=1,...,M\}$. ![image](figures/train_CL_SMF_onoffGain2.pdf){width="32.00000%"} ![image](figures/train_CL_IDF_onoffGain.pdf){width="32.00000%"} ![image](figures/train_SCL_onoffGain.pdf){width="32.00000%"} To find the machine learning model with the lowest prediction error, we allocate 3400 and 3000 data points, for C+L and S+C+L–band, correspondingly. We employ 10–fold cross–validation, which means that we use 90% for training (includes hyperparameter optimization) and 10% for testing as described [@Zibar20; @bishop2006]. For a more detailed explanation on the training of the employed machine learning model, see the Appendix Section. The remaining data points are later used for the final validation of the machine learning model for the pump laser current prediction of arbitrary gains. The procedure of obtaining pump current configuration is then as follows: 1) a single-layer neural network, $NN_{inv}$, is employed to learn the mapping between the target gain profiles and currents and voltage – inverse system learning, 2) once the neural network has learned the inverse mapping, given a set of target gain profiles, the corresponding pumps currents and voltages are predicted, 3) the predicted currents and voltages are then applied to the second multi–layer neural network, $NN_{fwd}$, that has learned the forward mapping between pump currents/voltage and gain profiles. The $NN_{fwd}$ thereby predicts the gain profile given the pump currents and the voltage. If the error between the predicted and targeted gain profile is not satisfactory pump currents and voltages are adjusted accordingly, i.e. fine–optimization. The fine–optimization uses iterative gradient descent by backpropagating the error through $NN_{fwd}$ to adjust the currents and voltage as described in [@Zibar20], 4) the obtained currents and voltages are and applied to the pump lasers in the experimental set–up, and new sets of measurements are performed, and 5) finally, to investigate the accuracy of the predicted pump currents and voltage, we calculate the maximum absolute error between the target and the newly measured gain profiles (i.e. $E_{MAX}$) and normalize it with the bandwidth ($BW$). The optimized topologies of the employed neural networks $NN_{fwd}$ and $NN_{dir}$, as well as their performance evaluation, are found in the Appendix Section. Results and discussion {#sec:results} ====================== Arbitrary gain profiles ----------------------- ![image](figures/pdf_cdf_distributed_CL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} ![image](figures/pdf_cdf_discrete_CL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} ![image](figures/pdf_cdf_discrete_SCL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} ![image](figures/RMSE_pdf_cdf_distributed_CL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} ![image](figures/RMSE_pdf_cdf_discrete_CL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} ![image](figures/RMSE_pdf_cdf_discrete_SCL_newFOMinTHz_same_xlim.pdf){width="32.00000%"} Fig. \[fig:results\_arb\](a)–(c) show the probability, (PDF), and the cumulative, (CDF), density functions of the $E_{MAX}/BW$ for the C+L–band (distributed and discrete) and S+C+L–band (discrete) Raman amplifiers. The error is defined between the targeted arbitrary gain profiles, taken directly from the data–set (not used for training the machine learning framework), and the predicted gain profiles obtained from the measurement using the pump currents and voltage allocation provided by the machine learning framework. We use 2100, 2600 and 1025 target arbitrary gain profiles for the distributed C+L–band, discrete C+L–band and discrete S+C+L–band validation, respectively. We compare the accuracy of allocating pump currents and voltage, by using only the inverse mapping multi–layer neural network, $(NN_{inv})$, and both the inverse and forward mapping multi–layer neural networks, $(NN_{inv}+NN_{fwd})$, which allows for fine–optimization of pump currents and the voltage. ![image](figures/DistributedRA_CL_onoffgain2_FOMinTHz_new.pdf){width="32.00000%"} ![image](figures/DiscreteRA_CL_onoffgain_FOMinTHz_new.pdf){width="32.00000%"} ![image](figures/DiscreteRA_SCL_onoffgain_FOMinTHz_new.pdf){width="33.00000%"} The PDFs shown in Fig. \[fig:results\_arb\](b)–(c), illustrate that for the discrete RA, highly–accurate pump current predictions, resulting in a low mean and standard deviation, can be obtained using only $NN_{inv}$. Thus, the currents and the voltage prediction is obtained in an ultra–fast way as $NN_{inv}$ only involves matrix computations. We notice that the mean and standard deviations are decreased by a factor of $\sim$2 when going from C+L to S+C+L–band. This is mainly because these two schemes have the same performance in terms of $E_{MAX}$ and S+C+L–band has almost two times wider bandwidth. However, qualitatively the results for C+L and S+C+L–band are comparable. If $NN_{inv}+NN_{fwd}$ is used a slight increase in the mean and the standard deviation is observed. This is because the $NN_{inv}$ has already found pump current configuration that minimizes the mean square error. Applying the fine–optimization introduces some small random deviations around this minimum and worsens the performance. For both discrete RA schemes, the CDF shows that most of the cases already present an $E_{MAX}/BW$ lower than $6\cdot10^{-2}$ dB/THz, before the fine–optimization, i.e. 97% of the cases for the C+L–band and $\sim$100% for the S+C+L–band. Compared to the discrete RA, the resulting PDF for the distributed RA (Fig. \[fig:results\_arb\](a)) has a higher mean and standard deviation when considering only $NN_{inv}$. On the other hand, a significant reduction can be obtained after applying fine–optimization $NN_{inv}+NN_{fwd}$, as also illustrated by the CDF. Indeed, the fine–optimization significantly increases the number of cases with $E_{MAX}/BW$ lower than $6\cdot10^{-2}$ dB/THz, i.e. from 18.7% to 95.4%. Finally, in Fig. \[fig:results\_arb\_rmse\](a)-(c), the resulting PDF and CDF of the RMSE per bandwidth is plotted for the distributed and discrete amplifiers. The Figure shows that very low mean and standard deviation values are achievable. Flat and tilted gain profiles ----------------------------- Next, we investigate the ability of the machine learning framework to predict accurate pump current and voltage allocations for the design of flat and tilted gain profiles using the discrete and distributed RAs, in C+L and S+C+L–band. Flat gains ranging from 6 to 12 dB (C+L–band distributed RA), 7 to 15 dB (C+L–band discrete RA), and 14 to 20 dB (S+C+L–band discrete RA) are evaluated in steps of 1 dB. For the tilted profiles, slopes of approximately 0.24 dB/THz (C+L–band RAs) and 0.20 dB/THz (S+C+L–band RA) are considered. These values were chosen to provide an overall tilt of around 1 dB on each band. Fig. \[fig:results\_flat\_tilted\], shows the predicted and target flat ((a)-(c)) and tilted ((d)-(f)) gain profiles, as a function of frequency, for the distributed and the discrete RA operating in C+L and S+C+L–band. Just a subset of gains (2 dB step) are shown for better visualization. The corresponding $E_{MAX}/BW$ for all gains under consideration is shown in Fig. \[fig:results\_flat\_tilted\](g)-(i). We only show results obtained after using $NN_{inv}+NN_{fwd}$ as the fine–optimization significantly reduced the error for all the amplifier schemes and their evaluated gains. A general trend observed in Fig. \[fig:results\_flat\_tilted\](a)–(f), is that the predicted gain oscillates around the target gain profile. The magnitude of the oscillations has a tendency to increase for increasing gains. Moreover, for the S+C+L–band RA, the oscillation amplitude increases with the frequency, achieving up to 2 dB of maximum error compared to the target. To understand what is happening, it is worth mentioning that it was observed some power instabilities on the supercontinuum S–band source and the Raman-based fiber laser $P_8$. Additionally, recall that the broadband and nonuniform Raman gain spectrum for a single pump, with a peak located near 12.5 THz below the pump frequency for the IDF, is partially overlapped in the multiple-pump configurations considered in this work as illustrated in Fig. \[fig:setup\](c). On the S–band, besides pumps $P_{6-8}$, there are also contributions of pumps $P_{1-5}$ because the S–band lies within the Raman gain spectrum bandwidth of all these pumps. This makes the design more complex on this region. Thus, although it is expected that the machine learning framework is able to deal with these broadband effects when adjusting the pumps (once the two stages on the S+C+L–band discrete RA are jointly trained), it is also expected to achieve a higher error on the S–band. It is observed in Fig. \[fig:results\_flat\_tilted\](h)-(i) that the $E_{MAX}/BW$ for the discrete RA in C+L and S+C+L–band is similar for the flat and the tilted gain profiles. The $E_{MAX}/BW$ is kept below $1.1\cdot10^{-1}$ and $0.9\cdot10^{-1}$ dB/THz for the design of flat and tilted gain profiles, respectively. On the other hand, the $E_{MAX}/BW$ for the distributed RA shown in Fig. \[fig:results\_flat\_tilted\](g) is higher for the design of the flat gains, but it is still kept below $1.3\cdot10^{-1}$ dB/THz. The reason may be related to the pump distributions, i.e. the number of pumps and wavelength being more suitable to provide a tilted gain profile. This can be observed on the experimental data–set gain profiles shown in Fig. \[fig:datasets\]. The same analysis does not apply for the S+C+L–band, since there it no clear flat/tilted profile trend on its data–set gain curves. Therefore, we also need to take into account that there will be a limitation on the theoretically achievable gain tilt and flatness given experimental set–up that has fixed wavelengths of pump lasers. Fig. \[fig:results\_flat\_tilted\](j)-(l) shows $RMSE/BW$ and it observed that the trends are very similar to as for $E_{MAX}/BW$. To put the presented work in the perspective, in Fig. \[fig:record\], $E_{MAX}/BW$, is plotted for various experimental demonstrations of multi–band amplifiers. It is observed that the presented work results in a low–error and broad bandwidth by means of machine learning. ![$E_{MAX}/BW$ as a function of amplifier bandwidth.[]{data-label="fig:record"}](figures/record_17_6THz_with_refs.pdf){width="47.00000%"} Conclusion {#sec:conc} ========== A multi–band programmable gain Raman amplifier operating in C+L and S+C+L–band is experimentally demonstrated. The key enabling technique is the machine learning framework that allows for ultra–fast and highly–accurate prediction of the pump currents and voltage for providing the targeted gain profiles. The ability to generate arbitrary gain profiles in a controlled and fast way, may provide novel approaches for the intelligent utilization of the ultra–wideband spectrum and become a key feature for future optical communication systems. Moreover, the programmable gain optical amplifier may advance other areas of fundamental science requiring spectral shaping, such as optical frequency combs. Appendix {#appendix .unnumbered} ======== The machine learning framework used in this paper to achieve highly accurate Raman amplifier (RA) programmable gains is based on two artificial neural networks. The first neural network $NN_{inv}$ models the RA inverse mapping, i.e. the mapping between gain profiles and pump lasers’ currents/voltage. Whereas the forward mapping, i.e., the mapping between the pump lasers’ currents/voltage and gain profiles, is learned by a second neural network $NN_{fwd}$. Here, we describe how these two NNs are trained for the different RA schemes considered in this paper. We also validate their prediction accuracy. Training and validation are performed on disjoint experimental data–sets, whose total number of elements are shown in Table \[tab:datasets\]. RA scheme C+L dist. C+L disc. S+C+L disc. ------------ ----------- ----------- ------------- Training 3464 3000 3000 Validation 2100 2600 1025 : **Experimental data–set distribution** \[tab:datasets\] Neural networks training {#sec:training} ------------------------ $NN_{inv}$ is trained using random projection (RP). This training algorithm, also known as extreme learning machine (ELM) [@Huang2011], initializes the weights of the hidden layers randomly, according to a normal distribution with mean zero and a certain standard deviation $\sigma_{NN_{init}}$, corresponding to NN initialization variance. This random weight assignment is independent from the training data–set and requires a high number of hidden nodes as these weights are kept untrained. The training data–set is used to optimize only the last layer weight by regularized least squares, with a regularization parameter $\lambda$. Since it is performed in a single step, the training time is drastically reduced when compared to standard approaches that updates all the weights in a numerical interactive routine. $NN_{inv}$ models for each RA scheme are shown in Table \[tab:NNinv\], where $f_{act}$ is the nonlinear activation function for all nodes (except the ones on the last layer, which use linear functions), $numHL$ is the number of hidden layers, and $numHN$ is the number of hidden nodes. To reduce the impact of the randomly initialized weights on the RP method, 20 parallel and independent $NN_{inv}$ are trained and the pump configuration prediction is the average of the 20 $NN_{inv}$ outputs [@Zibar20]. In Table \[tab:NNinv\], $f_{act}$, $numHN$, $\sigma_{NN_{init}}$ and $\lambda$ were obtained after a hyperparameter optimization routine using k-fold cross validation [@bishop2006]. RA scheme C+L dist. C+L disc. S+C+L disc. ---------------------- --------------------- --------------------- --------------------- Training alg. RP RP RP $f_{act}$ logsig sine sine $numHL$ 1 1 1 $numHN$ 760 500 500 $\sigma_{NN_{init}}$ $6.0 \cdot 10^{-3}$ $2.6 \cdot 10^{-2}$ $1.0 \cdot 10^{-2}$ $\lambda$ $1.0 \cdot 10^{9}$ $1.0 \cdot 10^{3}$ $1.0 \cdot 10^{4}$ : **Neural network models for $NN_{inv}$** \[tab:NNinv\] $NN_{fwd}$ is trained differently for each RA scheme. For the C+L–band RA (discrete and distributed), $NN_{fwd}$ is trained traditionally updating all weights on the NN interactively by using the Levenberg-Marquadt (LM) method. However, the high input and output dimensions of the S+C+L–band RA scheme makes the use of LM optimization challenging due to the long convergence time. Thus, RP is applied again only for this scheme. Table \[tab:NNdir\] summarizes $NN_{fwd}$ parameters for each RA scheme, where only the RP parameters $f_{act}$, $numHN$, $\sigma_{NN_{init}}$ and $\lambda$ were obtained after a hyperparameter optimization routine. Table \[tab:NNdir\] also shows that the RP faster training comes with the cost of having a larger network, with 500 hidden nodes instead of 20 when using LM. RA scheme C+L dist. C+L disc. S+C+L disc. ---------------------- ----------- ----------- --------------------- Training alg. LM LM RP $f_{act}$ tanh tanh tanh $numHL$ 2 2 1 $numHN$ 10 10 500 $\sigma_{NN_{init}}$ \* \* $1.0 \cdot 10^{-3}$ $\lambda$ 0 0 $1.0 \cdot 10^{8}$ : **Neural network models for $NN_{fwd}$** \[tab:NNdir\] Neural networks validation -------------------------- $NN_{inv}$’s performance in predicting pump currents/voltage is presented in Fig. \[fig:NNinv\]. The metric used is the absolute error relative to the maximum current/voltage excursion for each pump laser. Fig. \[fig:NNinv\] shows the probability density functions (PDF) and the cumulative density functions (CDF) over all the cases on the validation data–set and all pump lasers. Notice that the errors are kept bellow 2% for 95% of the cases for all the RA schemes. The prediction performance for the $NN_{fwd}$ is evaluated in terms of root mean squared error ($RMSE^{P}$) and maximum absolute error ($E^{P}_{MAX}$) between predicted $G^P$ and target $G^T$ gain profiles, extracted from the $K$ WDM points (spectrum), given by $$RMSE^{P} = \sqrt{\frac{1}{K}\sum_{i=1}^{N}(G_i^P-G_i^T)^2}, \label{eq:mse}$$ $$E^{P}_{MAX} = max\{|G_1^P-G_1^T|, |G_2^P-G_2^T|, \cdots, |G_N^P-G_N^T| \} \label{eq:maxE}$$ we use index $P$ for prediction to differentiate from the experimental validation errors shown in Section \[sec:results\] and recall that $K=90$ and $K=148$ for C+L and S+C+L-band RAs, respectively. Fig. \[fig:NNdir\] shows the PDF for $RMSE^{P}$ and $E^{P}_{MAX}$ over all the cases on the validation data–set. In Fig. \[fig:NNdir\], the overall $NN_{fwd}$ performances for both C-L–band RAs are consistent with the ones obtained in [@unpublishedBrusinJLT2020], which also considers a C+L–band RA (distributed scheme only) with same NN model and training algorithms. On the other hand, the worst performance obtained here by the S+C+L–band RA scheme in terms of $E^{P}_{MAX}$ can be explained by its more complex mapping relating more pumps to the gain over a wider bandwidth. S+C+L–band RA scheme was also the only model that used RP, but the same study presented in [@unpublishedBrusinJLT2020] showed that, for the Raman amplifier case, the performance of the LM only overcomes the RP for higher number of hidden nodes, which requires even more time to train. ![Probability density function (PDF) and cumulative density function (CDF) of the $NN_{inv}$ pump current/voltage prediction error, with indication of mean, $\mu$ and standard deviation, $\sigma$.[]{data-label="fig:NNinv"}](figures/inverseNN_all.pdf){width="40.00000%"} ![Probability density function (PDF) of the $NN_{fwd}$ gain prediction error: (a) $RMSE^{P}$ and (b) $E_{MAX}^{P}$ with indication of mean, $\mu$, standard deviation, $\sigma$, and maximum (max) values.[]{data-label="fig:NNdir"}](figures/directNN_all_rmse_maxE_test_set.pdf){width="40.00000%"} The errors $RMSE^{P}$ and $E^{P}_{MAX}$ are non-convex and unknown functions of the pump configuration that might not share the same local minimums, i.e. the pump configuration that minimizes $RMSE^{P}$ might not minimize $E^{P}_{MAX}$. However, since the fine–optimization is a gradient-based procedure, it needs to use a differentiable cost function with respect to the pump parameters, which makes the $MSE$ the only candidate for this. When the pdf curves in Fig. \[fig:NNdir\](a) and (b) present similar shapes, like for the C+L–band RAs, it might be an indication that minimums of these two errors occur for similar pump configurations and, consequently, minimizing $MSE^P$ (which is proportional to the $RMSE^P$), may also minimizes $E^{P}_{MAX}$. For the S-C-L–band RA, on the other hand, where $E^{P}_{MAX}$ and $RMSE^{P}$ pdf curves have completely different shapes, it is more likely that minimizing $MSE^P$ is not the same as minimizing $E^{P}_{MAX}$. Acknowledgment {#acknowledgment .unnumbered} ============== This work was supported by the European Union’s H2020 program (Marie Skłodowska-Curie grant 754462 and MSCA-ITN WON grant 814276), the European Research Council (ERC CoG FRECOM grant 771878), the Villum Foundations (VYI OPTIC-AI grant no. 29344), and the UK EPSRC grants EP/M009092/1 and EP/R035342/1. [^1]: U. C. de Moura, F. Da Ros and D. Zibar are with DTU Fotonik, Department of Photonics Engineering, Technical University of Denmark, DK-2800, Kgs. Lyngby, Denmark (e-mail: [email protected]). [^2]: Md A. Iqbal, M. Kamalian, L. Krzczanowicz, W. Forysiak and S. Turitsyn are with Aston Institute of Photonic Technologies (AIPT), Aston University, Birmingham, B4 7ET, United Kingdom. [^3]: A. M. Rosa Brusin and A. Carena are with Dipartimento di Elettronica e Telecomunicazioni (DET), Politecnico di Torino, Corso Duca degli Abruzzi, 24 - 10129, Torino, Italy.
{ "pile_set_name": "ArXiv" }
--- abstract: | Infrastructure-as-a-Service (IaaS) providers need to offer richer services to be competitive while optimizing their resource usage to keep costs down. Richer service offerings include new resource request models involving bandwidth guarantees between virtual machines (VMs). Thus we consider the following problem: given a VM request graph (where nodes are VMs and edges represent virtual network connectivity between the VMs) and a real data center topology, find an allocation of VMs to servers that satisfies the bandwidth guarantees for every virtual network edge—which maps to a path in the physical network—and minimizes congestion of the network. Previous work has shown that for arbitrary networks and requests, finding the optimal embedding satisfying bandwidth requests is $\mathcal{NP}$-hard. However, in most data center architectures, the routing protocols employed are based on a spanning tree of the physical network. In this paper, we prove that the problem remains $\mathcal{NP}$-hard even when the physical network topology is restricted to be a tree, and the request graph topology is also restricted. We also present a dynamic programming algorithm for computing the optimal embedding in a tree network which runs in time $O(3^kn)$, where $n$ is the number of nodes in the physical topology and $k$ is the size of the request graph, which is well suited for practical requests which have small $k$. Such requests form a large class of web-service and enterprise workloads. Also, if we restrict the requests topology to a clique (all VMs connected to a virtual switch with uniform bandwidth requirements), we show that the dynamic programming algorithm can be modified to output the minimum congestion embedding in time $O(k^2n)$. author: - | Debojyoti Dutta\ Cisco Inc.\ <[email protected]> - | Michael Kapralov\ Stanford University\ <[email protected]> - | Ian Post\ Stanford University\ <[email protected]> - | Rajendra Shinde\ Stanford University\ <[email protected]> bibliography: - 'DP\_algorithm.bib' title: 'Optimal bandwidth-aware VM allocation for Infrastructure-as-a-Service' --- Introduction {#sec:intro} ============ Infrastructure-as-a-Service (IaaS) providers like Amazon [@amazon], Rackspace [@rackspace] and Go-grid [@gogrid] provide computing and other services on demand and charge based on usage. This has resulted in the commoditization of computing and storage. Typically, these providers provide service level agreements (SLA) [@rackspace-sla] where they guarantee the type of virtual machines (VMs) that they provide and the amount of disk space available to these VMs. Although some providers offer additional services like dedicated firewalls and load-balancers, no network performance guarantees are provided, which are critical for workloads like content distribution networks, desktop virtualization, etc. Given the rapid growth and innovation in these services [@aws-growth], it is important for service providers (SPs) to offer innovative service models for differentiation, e.g., by offering richer network SLAs to be competitive while optimizing their resource usage to keep costs down. Next generation cloud services will require improved quality of service (QoS) guarantees for application workloads. For example, multi-tier enterprise applications [@tier-apps] require network isolation and QoS guarantees such as bandwidth guarantees, and for over-the-top content distribution using a cloud infrastructure, bandwidth, jitter and delay guarantees are important in determining performance. Similar guarantees are necessary for MapReduce-based analytics workloads too. Moreover, networking costs are currently a significant fraction of the total infrastructure cost in most data center (DC) designs [@ibm-dc-network; @cisco-dc-guide] since servers are cheap compared to core switches and routers. Thus, in order to provide richer network SLAs, it is important for SPs to ensure that networking resources are efficiently utilized while at the same time ensuring low congestion (that leads to better load balancing and more room for overprovisioning). In this paper we consider a virtualization request model in which clients can request bandwidth guarantees between pairs of virtual machines (VMs) [@SecondNet] for which SPs will allocate resources within their infrastructure. This naturally leads us to study the following resource allocation problem: given a VM request graph—where nodes are VMs and edges represent virtual network connectivity between the VMs—and a real data center topology, find an allocation of VMs to servers that satisfies the bandwidth guarantees for every virtual network edge and minimizes congestion of the network. Note that in this setting, each virtual edge maps to a path in the physical network topology. The above request graph model is driven by application workloads that execute on top of network infrastructure provided by the SPs. Common workloads include enterprise applications [@tier-apps], MapReduce [@mapreduce], and web hosting, and different workloads can lead to different service models. For instance, many web services request a small number of VMs to implement the web servers, the application servers, and the database. The VM implementing the web server receives a request and forwards it to an application server VM, which in turn queries the database server VMs. In such cases, specific bandwidth guarantees between the outside world and the web server, the web server and the application server, and so on, are important to ensure QoS. In MapReduce workloads on the other hand, it has been shown that network optimization can yield better results than adding machines [@mr-network], but in this setting since all the VMs implementing map and reduce tasks communicate with each other via data shuffle, the aggregate bandwidth available to the VMs may determine the application performance. A number of metrics have been studied to measure the network load including congestion, jitter, delay, hop count, or a combination of the above. Here we focus on minimizing congestion, but we also note that our algorithmic techniques are generic and can easily be adapted to optimize other metrics. It has been shown previously that the problem of embedding virtual requests in arbitrary networks is $\mathcal{NP}$-hard [@vne; @SecondNet]. However in most data center networks, routing protocols used rely on a spanning tree of the physical network [@cisco-dc-guide]. Hence, in this paper we study the problem of minimizing network congestion while allocating virtual requests when the network topology is restricted to be a tree. Our Contributions ----------------- First, we prove that optimally allocating VMs remains $\mathcal{NP}$-hard even when both the physical network topology and request topology are highly restricted. We show that if the network topology is a tree then even for simple request topologies like weighted paths with the weights signifying the amount of bandwidth required between the corresponding VMs, it is $\mathcal{NP}$-hard to approximate the minimum congestion to a factor better than $O(\theta)$, where $\theta$ is the ratio of the largest to smallest bandwidth requirements in the path request. We also show that in the unweighted case (or uniform bandwidth requirement on all edges) the problem is $\mathcal{NP}$-hard to approximate to within a factor of $O(n^{1-\epsilon})$ for any $\epsilon\in (0, 1)$, even for the case when the request topology is a tree. Given these complexity results, we cannot hope for an efficient algorithm for all instances of the problem. However, we note that in practice, many workloads consist of a small number of VMs allocated in a huge datacenter. Accordingly, our second result is a dynamic programming algorithm (Algorithm \[alg:cong\]) for computing the minimum congestion embedding of VMs in a tree network for any request graph, which satisfies the pairwise bandwidth requirements and runs in time $O(3^kn)$, where $n$ is the number of nodes in the physical topology and $k$ is the number of VMs in the request graph. Enterprise workloads often consist of small requests with specific bandwidth requirements between VMs, and for these instances the exponential $O(3^k)$ term is quite small, and can thus be optimally served using our algorithm whose run time is only linear in the network size. Third, workloads like Map-Reduce jobs have too many VMs to use an algorithm with a runtime of $O(3^kn)$, but these have uniform bandwidth requirements between the VMs [@TrafficAware], and we show that the exponential dependence on $k$ can be removed when the request network is uniform. For the special case in where the requests are restricted to be cliques or [*virtual clusters*]{} [@sig11], we propose an algorithm that finds the minimum congestion embedding in $O(k^2n)$ time (Algorithm \[alg:virtual-cluster\]). Hence our algorithms yield the minimum congestion embeddings of virtualization requests for several common use cases. We also present simulations which validate our results for common request models and practical network configurations. Outline of the paper -------------------- The paper is organized as follows. We first review previous work in Section \[subsec:related\] and formally define the problem and notation in Section \[sec:prelim\]. We prove the hardness results in Section \[sec:hardness\] followed the algorithms in Section \[sec:algo\]. In Section \[sec:simu\] we provide simulations, which validate the running time and correctness of our algorithms. Finally, we conclude and point to future work in Section \[sec:conclusion\]. Related Work {#subsec:related} ============ Previous work has shown that the problem of embedding virtual request graphs in arbitrary physical networks is $\mathcal{NP}$-hard [@vne; @SecondNet]. A number of heuristic approaches have been proposed including mapping VMs to nodes in the network greedily and mapping the flows between VMs to paths in the network via shortest paths and multi-commodity flow algorithms [@fan; @zhu]. However these approaches do not offer provable guarantees and may lead to congested networks in some circumstances. The authors of [@vne] assume network support for path-splitting [@rethinking] in order to use a multi-commodity flow based approach for mapping VMs and flows between them to the physical network, but this approach is not scalable beyond networks containing hundreds of servers [@SecondNet]. Guo [[*et al.*]{}]{}[@SecondNet] proposed a new architectural framework, Secondnet, for embedding virtualization requests with bandwidth guarantees. This framework considers requests with bandwidth guarantees $f_{ij}$ between every pair of VMs $(v_i, v_j)$. This framework provides rigorous application performance guarantees and hence is suitable for enterprise workloads but at the same time also establishes hardness of the problem of finding such embeddings in arbitrary networks. Our results employ the SecondNet framework but restrict attention to tree networks. Very recently, Ballani [[*et al.*]{}]{}[@sig11] have described a [*virtual cluster*]{} request model, which consists of requests of the form $\langle k,B\rangle$ representing $k$ VMs each connected to a virtual switch with a link of bandwidth $B$. A request $\langle k,B\rangle$ can be interpreted (although not exactly) as a clique request on $k$ VMs with a bandwidth guarantee of $B/(k-1)$ on each edge of the clique. They describe a novel VM allocation algorithm for assigning such requests on a tree network with the goal of maximizing the ability to accommodate future requests. For each $v$ in the tree network $T$, they maintain an interval of values that represents the number of VMs that can be allocated to $T_v$ without congesting the uplink edge from $v$ and allocate VMs to sub-trees greedily. We generalize this approach to the case of virtualization requests in the Secondnet framework and we use a dynamic programming solution in order to find the optimal minimum congestion embedding. By restricting the requests to [*virtual clusters*]{}, [@sig11] offers a tradeoff to the providers between meeting specific tenant demands and flexibility of allocation schemes. In this work, we explore this tradeoff further and show that it is possible to formulate flexible allocation schemes even in the Secondnet framework for small requests. The problem of resource allocation has also been studied in the virtual private network (VPN) design setting where bandwidth guarantees are desired between nodes of the virtual network [@duffield99; @gupta01]. In this setting, a set of nodes of the physical network representing the VPN endpoints is provided as the input, and the task is to reserve bandwidth on the edges of the network in order to satisfy pairwise bandwidth requirements between VPN endpoints. The fixed location of VPN endpoints makes this problem significantly different from that of embedding virtualization requests in a network, since the latter involves searching over all possible embeddings of the VMs in the network. Preliminaries {#sec:prelim} ============= An instance of our problem consists of a datacenter network and a request network. The datacenter network $N$ is a tree on $n$ nodes rooted at a gateway node $g$. Edges in $N$ have capacities $c_e$ representing their bandwidth. Let $L$ denote the set of leaves of $N$. The request network $G_R$ is an arbitrary, undirected graph on $k+1$ nodes. Nodes in $G_R$ consist of a set $V$ of $k$ virtual machines $v_1,\ldots, v_k$ and a special gateway node $g$. Edges $e$ in the request graph specify bandwidth guarantees $f_e$ (flow requirements) and are divided into two types: edges of type-[slowromancap1@]{} have the form $e=(v_i,g)$ and specify a requirement for routing $f_e$ flow between $v_i$ and the gateway node $g$ (uplink bandwidth to the outside world), and edges of type-[slowromancap2@]{} have the form $e=(v_i,v_j)$ and specify flows between two virtual machines $v_i$ and $v_j$ (“chatter” bandwidth between virtual machines). We use $R^{{\expandafter\@slowromancap\romannumeral 1@}}$ and $R^{{\expandafter\@slowromancap\romannumeral 2@}}$ to denote the sets of type-[slowromancap1@]{} and type-[slowromancap2@]{} edges and $R = R^{{\expandafter\@slowromancap\romannumeral 1@}} \cup R^{{\expandafter\@slowromancap\romannumeral 2@}}$ to denote all edges. A solution consists of an embedding $\pi: V \rightarrow L$ mapping virtual machines onto leaves in the datacenter network. For simplicity we will assume only a single VM can be mapped to each leaf, although it is easy to modify our algorithm so that each datacenter node $v$ can support up to $n_v$ VMs. The gateway node $g$ in $G_R$ is always mapped to the gateway in $N$. If $\pi$ maps the endpoints of edge $e=(v_i,v_j)$ (equivalently $e=(v_i,g)$) onto $\pi(v_i)$ and $\pi(v_j)$, then $e$ contributes $f_e$ flow to every edge along the path $P_{\pi(v_i),\pi(v_j)}$ between $\pi(v_i)$ and $\pi(v_j)$ in $N$. The congestion of an edge $e$ in $N$ under embedding $\pi$ is $${\textrm{Cong}}(\pi,e) = \frac{1}{c_e} \sum_{(u,v) \in R \textnormal{ s.t. } e \in P_{\pi(u),\pi(v)}} f_{(u,v)}$$ and our goal is to find $\pi$ minimizing $\max_{e \in N} {\textrm{Cong}}(\pi,e)$. Hardness results {#sec:hardness} ================ In this section we show that the embedding problem is $\mathcal{NP}$-hard even with the restricted topologies of the host and request graphs. In particular, we show that the problem of embedding a [*weighted path request*]{}, which is perhaps the simplest weighted request topology, is $\mathcal{NP}$-hard to approximate to a factor better than $O(\theta)$, where $\theta$ is the ratio of the largest to smallest bandwidth requirements. Furthermore, we show that in the unweighted case the problem is $\mathcal{NP}$-hard to approximate to a factor smaller than $O(n^{1-{\epsilon}})$ for any constant ${\epsilon}\in (0, 1)$, where $n$ is the number of VMs in the request, even for the case when the request topology is a tree. Both of our reductions are from 3-partition. An instance of 3-partition consists of a multiset $S = \{s_1, \ldots, s_{3m}\}$ of $3m$ integers summing to $mB$, and the goal is to determine whether $S$ can be partitioned into $m$ subsets $S_1, \ldots, S_m$ such that the sums of elements in each of the $S_i$ are equal to $B$ and $|S_i|=3$ for all $i$. Crucially, 3-partition remains $\mathcal{NP}$-complete even when the size of the integers are bounded by a polynomial in $m$: The 3-partition problem is strongly $\mathcal{NP}$-complete, even when $B/4<s_i<B/2$ for all $i$, forcing any partition to consist of triples. Weighted topologies ------------------- \[thm:hardness-1\] The embedding problem is $\mathcal{NP}$-complete even when restricted to instances where the request graph is a weighted path, and the host network is a tree. Moreover, it is $\mathcal{NP}$-hard to approximate to within a factor better than $\theta/6$, where $\theta$ is the ratio of the largest to smallest weight in the request graph. First, the problem is in $\mathcal{NP}$, since given a candidate embedding, it is easy to verify that its congestion is at most $1$. Now, let $S = \{s_1, \ldots, s_{3m}\}$ be a multiset of $3m$ integers summing to $mB$, constituting an instance of 3-partition, such that $B/4<s_i<B/2$ for all $i$. Let $T$ be a tree of height two. The root/gateway $g$ has $m$ children labeled $S_1, \ldots, S_m$, each of which has $B$ children of its own. Since 3-partition is strongly $\mathcal{NP}$-complete, we may assume that $B$ is bounded by a polynomial in $m$, so $T$ has polynomial size. All edges from $g$ to the $S_i$ have capacity 6. Each node $S_i$ is connected to each of its $B$ children by edges of capacity $W>6$. Let $R=R^I\cup R^{II}$ be defined as follows. Let $V=\{v_1,\ldots, v_{mB}\}$ be a set of VMs. For $j=1,\ldots, 3m+1$ let $q_{j}=\sum_{i=0}^{j-1} s_i$, where we set $s_0=0$ for convenience (note that $q_1=0$ and $q_{3m+1}=mB$). Further, define [*heavy intervals*]{} as $I_j=\{v_{q_j+1},\ldots, v_{q_{j+1}}\},j=1,\ldots, 3m$, so that $|I_{j}|=s_j$. Define chatter bandwidth requests $f_{ij}$ by setting $$f_{ij}=\left\lbrace \begin{array}{cc} W,&\text{if $\{i, j\}\subseteq I_k$ for some $k$}\\ 1&\text{otherwise.} \end{array} \right.$$ Define uplink bandwidths as $f_i=1$ for $i=1$ and $f_i=0$ otherwise. Thus, the requests form a path with the first node on the path connected to the gateway node. The path is partitioned into intervals of length $s_i$, such that the bandwidth requirement between consecutive nodes in each interval is high and the requirement between adjacent nodes on the path that belong to different intervals is low. We refer to the edges of weight $W$ as [*heavy edges*]{} and the edges of weight 1 as [*light edges*]{}. If $S$ has a 3-partition, then the heavy intervals $I_j$ can be divided into $m$ sets ${\mathcal{P}}_1, \ldots, {\mathcal{P}}_m$ of 3 intervals each, such that the sum of the lengths within each ${\mathcal{P}}_i$ is exactly $B$. We can map all VMs in ${\mathcal{P}}_i$ to the children of node $S_i$. Each edge $(g,S_i)$ carries flow from at most 2 light edges on the border of each of the 3 heavy intervals in ${\mathcal{P}}_i$, and each edge connecting $S_i$ to its children has load at most $W$, for a congestion of 1. Now suppose that $S$ does not have a 3-partition. Then since by assumption $B/4<s_i<B/2$, in any feasible allocation of VMs at least one heavy interval $I_k$ must be divided between children of different nodes $S_i$ and $S_j$, and hence at least one heavy edge must congest the edge $(r, S_i)$, yielding congestion at least $W/6$. Thus, it is $\mathcal{NP}$-hard to distinguish between instances with an optimal congestion of $1$ and $W/6=\theta/6$, where $\theta$ is the ratio of largest and the smallest weight in the request graph, i.e. $\theta=W/1$. Unweighted topologies --------------------- \[thm:hardness-2\] Let $n$ denote the number of leaves in the host tree. The embedding problem is $\mathcal{NP}$-complete and $\mathcal{NP}$-hard to approximate to within a factor better than $\Omega(n^{1-{\epsilon}})$, for any ${\epsilon}\in (0, 1)$, when the set of requests forms an unweighted tree. As before, we first note that the problem is in $\mathcal{NP}$, since given a candidate embedding, it is easy to verify that its congestion is at most $1$. We use a reduction to 3-partition similar to the reduction to Maximum Quadratic Assignment used in [@qa5]. Let $S = \{s_1, \ldots, s_{3m}\}$ be a multiset of $3m$ integers summing to $mB$, constituting an instance of 3-partition, such that $B/4<s_i<B/2$ for all $i$. Let $T$ be a tree of height two. The root $g$ has $m$ children labeled $S_1, \ldots, S_m$, each of which has $3+B\cdot M$ children of its own, where $M=(5 m B)^{\lceil (1-{\epsilon})/{\epsilon}\rceil}$. Since 3-partition is strongly $\mathcal{NP}$-complete, we may assume that $B$ is bounded by a polynomial in $m$, so $T$ has polynomial size. Each node $S_i$ is connected to each of its $3+B\cdot M$ children by links of capacity $B\cdot M+2$, and the root is connected to each of $S_i$ by links of capacity $6$. We now define $R=R^I\cup R^{II}$. Let $V=V^1\cup V^2$, where $V^1=\{v_1^1,\ldots, v^1_{m}\}$ and $V^2=\{v^2_1,\ldots, v^2_{mBM}\}$ be a set of VMs organized in a tree as follows. First for $j=1,\ldots, 3m+1$ let $q_{j}=\sum_{i=0}^{j-1} s_i$, where we set $s_0=0$ for convenience. We now define bandwidth requirements between VMs in $V$. Each $v^1_i\in V^1$ requires chatter connections of bandwidth 1 to $v^2_{M\cdot (q_i+1)}, v^2_{M\cdot (q_i+1)+1},\ldots, v^2_{M\cdot q_{i+1}}$. Also, $v^1_i$ requires a chatter connection to $v^1_{i-1}$ if $i>1$ and $v^1_{i+1}$ if $i<m$. Finally, both $v^1_1$ and $v^1_m$ require uplink connections to gateway $g$ of bandwidth 1. Thus, the request topology is a tree consisting of stars on $s_i\cdot M$ nodes with centers $v^1_{i}$, for each $i=1,\ldots, m$. Adjacent centers of stars (i.e. $v^1_i$ and $v^1_j$ for $|i-j|=1$) are connected to each other. If $S$ admits a 3-partition, then there exists an embedding of congestion at most $1$: assign the corresponding three centers and their children to the children of $S_j$ for $j=1,\ldots, m$, which is possible since each $S_j$ has exactly $3+B\cdot M$ children. The congestion is at most $1$ since the edges of $T$ incident on the nodes where the centers are mapped will carry load exactly $B\cdot M+2$ ($B\cdot M$ unit bandwidth connections to the children as well as two connections to neighboring centers or uplink connections), and the edges $(S_j, g)$ will carry at most $2$ units from each of the 3 centers mapped to the children of $S_j$, yielding congestion at most $1$. Now suppose that $S$ does not admit a 3-partition. Consider the node $S_j\in T$ with the maximum number of centers mapped to its children. Denote these centers by $v^1_{c_1},\ldots, v^1_{c_k}$, where $k>3$. We then have $\sum_{j=1}^k s_{c_j}\geq B+1$, and hence at least $M$ children of $v^1_{c_1},\ldots, v^1_{c_k}$ are mapped outside the set of children of $S_j$. Hence, at least $M$ edges from the centers $v^1_{c_1},\ldots, v^1_{c_k}$ to these children congest the edge $(S_i, g)$, where $g$ is the root of $T$. Thus, the congestion is at least $M/6$. The number of vertices in the tree $T$ is $n=1+m(3+B\cdot M)\leq 1+(3+B)m\cdot M\leq (5mB)\cdot M\leq M^{{\epsilon}/(1-{\epsilon})+1}=M^{1/(1-{\epsilon})}$. Hence, the congestion is at least $M/6\geq n^{1-{\epsilon}}/6$. We have shown that it is $\mathcal{NP}$-hard to distinguish between instances of the problem where the minimum congestion is $1$ and $\Omega(n^{1-{\epsilon}})$, thus completing the proof. Algorithm {#sec:algo} ========= Next we present our algorithmic results and show that despite the $\mathcal{NP}$-completeness results in the previous section, many practical instances can still be solved efficiently. Creation of binary tree ----------------------- We first convert the tree $N$ into a binary tree $T$ with not many additional nodes in a way that preserves the congestion of all solutions. This step is purely for convenience in simplifying the presentation of the algorithm that follows. We simply replace each degree $d$ node with a complete binary tree on $d$ nodes. Algorithm \[binary\_tree\_alg\] describes the procedure Create-Binary-Tree($N,g$) more formally. \[binary\_tree\_alg\] Let $u_1,\ldots, u_d$ be the children of $v$, and $e_1,\ldots, e_d$ the edges connecting $v$ to $u_i$ Replace $e_1,\ldots,e_d$ with a binary tree rooted at $v$ with leaves $u_1,\ldots, u_d$ Set the capacity of parent edges of $u_i$ to be $c_{e_i}$ and that of all other new edges to be $\infty$ Let $T$ be the resulting binary tree. We first show that the congestion of embedding into $T$ and $N$ is equal: \[lem:equivalence\] The congestion of embedding any request graph $G_R$ into a tree $N$ rooted at node $g$ is equal to the congestion of embedding $G_R$ into the binary tree $T$ constructed by the procedure Create-Binary-Tree($N,g$) Consider any embedding $\pi$ of $G_R$ into $N$. Since the auxiliary nodes inserted are not leaves, $\pi$ defines an embedding of $G_R$ into $T$ as well. Let $u,v \in N \cap T$ and $P_{u,v}^N$, $P_{u,v}^T$ be the edges on the unique paths between $u$ and $v$ in $N$ and $T$. Observe that $P_{u,v}^N \subseteq P^{T}_{u,v}$, and that all edges in $P^{T}_{u,v} \setminus P_{u,v}^N$ have infinite capacity and contribute nothing to the congestion. Hence the congestion of embedding in $N$ and $T$ is equal. Next, we show that $T$ is not much bigger than $N$: \[lem:size\] The number of nodes is $T$ is at most $2n$ and the height of $T$ is $O(H\log \Delta)$ where $\Delta$ is maximum degree in $N$ and $H$ denotes the height of $N$. We replace each node $v$ of degree $d$, with a complete binary tree on $d$ leaves, which has at most $2d$ nodes. Therefore, the number of nodes in $T$ is at most $2n$. Also by this replacement, we stretch sub-trees of height $1$ by a factor at most $\lceil \log \Delta \rceil$ which shows that the height of $T$ is $O(H\log \Delta)$. Minimum congestion of embedding requests in a binary tree --------------------------------------------------------- Now we present our primary algorithmic result and show that if the request graph is small—which is true in many practical instances—then the optimal embedding can be found efficiently. Before describing the algorithm, we introduce some notation. For any node $u \in T$ we use the symbol $e_u$ to denote the link joining the parent of node $u$ to $u$ and $T_u$ to denote the subtree of $T$ rooted at $u$. If $u$ is not a leaf, we refer to the “left” and “right” children of $u$ in $T$ as $u_l$ and $u_r$ respectively. In this section we assume that the tree $T$ rooted at $g$ is binary and of height $H$. Let $L^j$ denote the set of vertices in $T$ at distance $j$ from $g$, so $L^0 = \{g\}$, while $L^H$ denotes the leaves at the lowest level. The algorithm is straightforward dynamic programming. Starting at the leaves of $T$, and moving upwards towards the root, for each node $u \in T$ and set $S \subseteq V$ we calculate the congestion of the optimal embedding of the VMs in $S$ into $T_u$ using the congestion of embeddings into $u$’s children. Let ${\textrm{Flow}}[S]$ denote sum of the bandwidth requirements crossing the cut $(S,V\cup\{g\}\setminus S)$ in $G_R$, and ${\textrm{Cong}}[u,S]$ denote the optimal congestion of the edges of $T_u$ when embedding the subgraph of $G_R$ spanned by $S$ into $T_u$. Then ${\textrm{Cong}}[u,S]$ satisfies the recurrence $${\textrm{Cong}}[u,S] = \min_{S_l \subseteq S} \max\left\{ {\textrm{Cong}}[u_l,S_l],{\textrm{Cong}}[u_r,S\setminus S_l], {\textrm{Flow}}[S_l]/c_{e_l},{\textrm{Flow}}[S\setminus S_l]/c_{e_r} \right\}$$ That is, it is the minimum over all partitions $(S_l,S\setminus S_l)$ of $S$ of the congestion of embedding $S_l$ into $T_{u_l}$ and $S\setminus S_l$ into $T_{u_r}$. The terms ${\textrm{Flow}}[S_l]/c_{e_l}$ and ${\textrm{Flow}}[S\setminus S_l]/c_{e_r}$ are the congestion on the edges connecting $u$ to its children. The base case is when $u$ is a leaf, in which case $${\textrm{Cong}}[u,S] = \begin{cases} 0 & \textnormal{if } |S| \le 1 \\ \infty & \textnormal{if } |S| > 1 \end{cases}$$ assuming for simplicity that each server can support at most a single VM. By changing this equation, we can easily allow a server $v \in T$ to support up to $n_v$ VMs. After computing these recurrences, the algorithm outputs ${\textrm{Cong}}[g,V]$. Note that $L^0 = \{g\}$ and that it suffices to compute ${\textrm{Cong}}[g,V]$ (i.e., ${\textrm{Cong}}[g,S]$ for subsets $S \subset V$ is not needed). Algorithm \[alg:cong\] shows the procedure in more detail. Binary tree $T$ rooted at $g$, request graph $G_R$ Minimum congestion in embedding $V$ into $T$ such that requirements $R$ are satisfied $\begin{aligned} {\textrm{Flow}}[S] \gets \sum_{(v,g) \in R^{{\expandafter\@slowromancap\romannumeral 1@}}, v \in S} f_{(v,g)} + \sum_{(u,v) \in R^{{\expandafter\@slowromancap\romannumeral 2@}}, u\in S, v\notin S} f_{(u,v)} \end{aligned}$ ${\textrm{Cong}}[u,S] \gets 0 \text{ if } |S| \leq1, \infty \text{ otherwise}$ $t_{\min} \gets \infty$ $t \gets \max \left\{ {\textrm{Cong}}[u_l, S_l], {\textrm{Cong}}[u_r, S \setminus S_l], {\textrm{Flow}}[S_l]/c_{e_l}, {\textrm{Flow}}[S\setminus S_l]/c_{e_r} \right\}$ $t_{\min} \gets t$ $S_{\min} \gets S_l$ ${\textrm{Cong}}[v,S] \gets t_{\min}$ ${\textrm{Part}}[u,S] \gets (S_{\min}, S \backslash S_{\min})$ ${\textrm{Cong}}[g,V]$ When we update ${\textrm{Cong}}[u,S]$ we also store the partition $(S_l,S\setminus S_l)$ that realizes this optimal congestion in a partition table ${\textrm{Part}}[u,S]$. After the execution of the algorithm, we can recover the optimal embedding by working backwards in the standard fashion for dynamic programs: starting at $g$ we read the optimal partition $(V_1, V\setminus V_1)$ from ${\textrm{Part}}[g,V]$. Now we find the optimal partitions of $V_1$ with root $g_l$ and $V\setminus V_1$ with root $g_r$, and so on. Now we analyze the correctness and runtime: \[lem:correctness\] Algorithm \[alg:cong\] finds the minimum congestion of embedding request $G_R$ in a tree network $N$. By Lemma \[lem:equivalence\], optimizing the congestion on $N$ is equivalent to optimizing it on $T$. The optimal congestion of an embedding restricted to $T_u$ requires using an optimal partition into subsets embedded into left and right subtrees of $T_u$, and Algorithm \[alg:cong\] recursively computes the optimal embedding for all possible partitions of the VMs, thus retrieving the congestion of the optimal embedding. Algorithm \[alg:cong\] has running time $O(3^kn)$. We first calculate ${\textrm{Flow}}[S]$ for every set $S \subseteq V$. There are $2^k$ such sets, and each requires summing over at most $k^2$ edges in $R$, for a runtime of $O(k^22^k)$, which is $O(3^k)$ for large enough $k$. In the main loop, for each $u$ in $T$ we compute ${\textrm{Cong}}[u,S]$ for all sets $S \subseteq V$. If $|S| = i$, computing ${\textrm{Cong}}[u,S]$ requires looking at all $2^i$ subsets of $S$ and doing $O(1)$ work for each one. Summing over all $O(n)$ nodes and all sets $S$, this requires $O(n)O(\sum_{i=0}^k {k \choose i} 2^i) = O(3^kn)$ work total. Other Objective Functions and Request Models {#subsec:virtual_cluster} -------------------------------------------- The basic form of our algorithm is not specific to congestion, and the recurrence in Algorithm \[alg:cong\] can easily be modified to optimize for any objective function for which we can write a similar recurrence. For instance, if each edge in $T$ has a delay and bandwidth capacity, we can minimize the average or maximum latency between VMs subject to satisfying bandwidth constraints (with a slightly more complex recurrence). In practice it may not be desirable to allow request graphs to have arbitrary topologies and edge weights. If a request graph is sufficiently simple and uniform, then the complexity results of Section \[sec:hardness\] no longer apply, and we no longer need to consider all $2^k$ cuts of $G_R$ at each node. For instance, if $G_R$ is a clique with equal bandwidth on all edges, then the congestion of embedding a set of VMs $S$ into $T_u$ is dependent only on the size of $S$, so we only need to consider $k+1$ subproblems for each node in $T$. Ballani et al. [@sig11] describe a [*virtual cluster*]{} request model, which consists of requests of the form $\langle k,B\rangle$ representing $k$ VMs each connected to a virtual switch with a link of bandwidth $B$. Such a request $\langle k,B\rangle$ is similar (but not identical) to a request consisting of a clique on $k$ VMs and a bandwidth guarantee of $B/(k-1)$ on each edge of the clique in our setting. We show that when restricted to [*virtual cluster*]{} requests, a modified version of Algorithm \[alg:cong\] finds the minimum congestion embedding in time $O(nk^2)$. For the sake of completeness and comparison with their work, we present Algorithm \[alg:virtual-cluster\]. Similar adjustments could be made to handle other request models for which considering all $2^k$ cuts of the request graph is unnecessary. ${\textrm{Cong}}[u,i] \gets 0 \text{ if } z \le 1, \infty \text{ otherwise}$ $t_{\min} \gets \infty$ $f_l \gets i\cdot(k-i)\cdot B/(k-1)$ $ f_r \gets (z-i) \cdot (k-z+i) \cdot B/(k-1)$ $ t \gets \max \left\{ {\textrm{Cong}}[u_1, i], {\textrm{Cong}}[u_2, z-i], f_l/c_{e_l}, f_r/c_{e_r} \right\}$ $t_{\min} \gets t$ $i_{\min} \gets i$ ${\textrm{Cong}}[u,z] \gets t_{\min}$ ${\textrm{Part}}[u,z] \gets (i_{\min}, z-i_{\min})$ ${\textrm{Cong}}[g,k]$ The correctness of Algorithm \[alg:virtual-cluster\] can be inferred from the correctness of Algorithm \[alg:cong\] by noting that under the [*virtual cluster*]{} request model, all subsets of equal size embed in a subtree with same congestion, i.e. for any $S_1, S_2 \subseteq V$ such that $|S_1| = |S_2|$, we have ${\textrm{Cong}}[u,S_1] = {\textrm{Cong}}[u, S_2]$ for all $u \in T$. For every node $u$ and for all $z \in 0 \ldots k$, Algorithm \[alg:virtual-cluster\] calculates ${\textrm{Cong}}[u,z]$ by optimizing over $z+1$ possible splits of the $z$ VMs among its children. A simple recursive calculation shows that this computation has complexity $\sum_{z = 0}^k (z+1) = O(k^2) $. This shows that the running time of Algorithm \[alg:virtual-cluster\] is $O(nk^2)$. Simulations {#sec:simu} =========== In this section we present simulations which verify the correctness and scaling properties of Algorithm \[alg:cong\] in both the pairwise bandwidth guarantees model, as well as [*virtual cluster*]{} request model \[subsec:virtual\_cluster\]. We perform all simulations using an unoptimized python implementation of Algorithm \[alg:cong\] on Intel Sandy Bridge Quad Core machine having $4$ GB of RAM using the networkx graph package [@networkx] to simulate the physical network. Network configuration --------------------- In order to test our algorithm on a realistic networks, we simulate a typical three tier data center network [@alfares] with servers housed in racks which are connected to a Top-Of-Rack (TOR) switch (tier I). The TOR switches connect the racks to other parts of the network via Aggregation Switches (AS, tier II). The AS switches have uplinks connecting them to the Core Switch (CS, tier III). We assume that TOR’s are connected to the servers with $10$ GBps links while the uplinks from TORs to the AS’s are $40$ GBps and from the AS’s to the CS’s are $100$ GBps. We construct a tree topology over these elements, recalling that common routing protocols used in practice employ a spanning tree of the physical network. We model existing traffic in the data center network using random residual capacities for each link. We choose the residual capacity for edge $e$ independently of all other edges and uniformly at random from $[0,c(e)]$ where $c(e)$ denotes the bandwidth capacity of edge $e$. The choice of random residual link capacities is forced on us due to lack of models describing realistic network flows in a data center. We note that Algorithm \[alg:cong\] finds the optimal congestion embedding for any of the distribution of residual capacities on the network links and any choices of bandwidth capacities of the links. Linear scan over all possible VM allocations -------------------------------------------- By implementing a linear scan over all possible VM allocations in the network, we verify the correctness of Algorithm \[alg:cong\] by finding the allocation that minimizes congestion. Note that this implementation requires scanning ${n \choose k} \cdot k! = O(n^k)$ feasible VM allocations where $n$ denotes the number of servers in the network and $k$ denotes the request size. Hence we choose small network and request sizes $n \in \{50,75,100\}$ and $k = 4$ and verify correctness of the algorithm for different request topologies and randomly generated residual capacities on the network links. We observe that this procedure requires hours or even days to finish even for very small network and request sizes like $n = 125$ and $k = 4$ as seen in Table \[table:brute-force\] and hence is infeasible for modern data centers containing hundreds of thousands of servers. In contrast, Algorithm \[alg:cong\] has complexity $O(3^kn)$, which is linear in the network size $n$, and as shown in the next sub section, finishes in order of seconds on our simulation setup for small values of $k$. n k Time (hours) ----- --- -------------- 50 4 2.2 75 4 18.8 100 4 80 : Linear scan for VM allocation: run time[]{data-label="table:brute-force"} Pairwise bandwidth requirements ------------------------------- Next, we verify the scaling properties of Algorithm \[alg:cong\] with respect to parameters $n$ and $k$. First, we fix a request of size $k = 5$, and plot the running time for increasing values of $n$, the number of servers, from $n = 200$ to $n = 2000$ in Figure \[fig:scale\](a) which illustrates the linear variation of run time with respect to $n$. Next, we fix the network size to $n = 100$ and plot the run time for [*path requests*]{} with lengths from $k = 4$ to $k = 10$ in Figure \[fig:scale\]. This figure shows that the run time increases exponential with respect to $k$. Virtual Cluster Request Model ----------------------------- We also verify the scaling properties when the requests are restricted to the virtual cluster model [@sig11]. For a fixed request $\langle k,B\rangle$ where $k = 100$ and $B = 100$Mbps, we plot the running time for increasing values of $n$, from $n = 200$ to $n = 2000$ in Figure \[fig:scale:vcluster\](a) which illustrates the linear variation of run time with $n$. Next, we fix the network size to $n = 1000$ and plot the run time for $k$ in the range $10$ to $100$. These results show that for [*virtual cluster*]{} requests, our algorithm finds the minimum congestion embedding in time $O(nk^2)$. As mentioned before, a number of heuristic approaches have been formulated to perform VM allocation. However, lack of models of the existing network flows inside a data center, especially in the context of enterprise workloads, hinders the evaluation and comparison of their performance in realistic settings. In particular, we observe that by congesting particular edges in the network, it is possible to make the greedy heuristics for VM mapping perform significantly worse than the optimal embedding (output by Algorithm \[alg:cong\]). However, a thorough comparison with heuristics requires models of flow in a data center serving enterprise requests, and we leave this to future work. Conlusion and Future Work {#sec:conclusion} ========================= In this paper we study the problem of allocating a graph request within a tree topology and we present a $O(3^kn)$ dynamic programming algorithm that embeds the resource request graph of size $k$ into the data center topology (tree) of size $n$ to minimize congestion. We believe this is useful in enterprise workloads when the request size $k$ is small. For clique requests, we present a $O(n^2k)$ dynamic programming algorithm to allocate clusters of size $k$ in a tree of size $n$ for minimizing congestion, which could be useful for MapReduce-like workloads. We believe that it would also be possible to extend our results to hybrid workloads involving tiers of VMs, with both inter-tier as well as intra-tier bandwidth guarantees. We also provide hardness results and show that the problem of finding minimum congestion embedding in a network remains in $\mathcal{NP}$-hard even under the restriction to tree network. We focus on minimizing congestion as our objective function, but we believe our methods are applicable to a wider class of metrics and objective functions.
{ "pile_set_name": "ArXiv" }
--- abstract: | Let $(M=G/H, g)$ be a reductive homogeneous Riemannian manifold, where $g$ is a $G$-invariant metric and let $T_oM$ be the tangent space of $M$ at $o=eH$. We study the differential equation $ \nabla_{\dot{x}}{\dot{x}}=kI(\dot{x})$ ($k\in\mathbb{R}$ and $I$ an endomorphism of $T_oM$), whose solution $x(t)$ represents the motion of a charged particle in $M$ under the electromagnetic field $kI$. If $k=0$ then $x(t)$ is a geodesic. We solve such an equation in a Riemannian fibration $K/H\to G/H\to G/K$, where $G$ is a Lie group with a bi-invariant Riemannian metric and $H\subset K\subset G$ are closed connected subgroups of $G$. To this end, we prove a more general result which has its own interest. Namely, we study the motion of a charged particle in a homogeneous space $G/H$ with reductive decomposition ${\mathfrak}{g} = {\mathfrak}{h}\oplus{\mathfrak}{m}$ (${\mathfrak}{g}, {\mathfrak}{h}$ the Lie algebras of $G$ and $H$ respectively), under the following conditions: (i) the tangent space $T_oM\cong{\mathfrak}{m}$ admits a decomposition ${\mathfrak}{m}={\mathfrak}{m}_1\oplus\cdots\oplus{\mathfrak}{m}_s$ into $\operatorname{Ad}(H)$-modules, orthogonal with respect to an $\operatorname{Ad}$-invariant inner product of ${\mathfrak}{g}$ and the $G$-invariant metric $g$ is diagonal on ${\mathfrak}{m}$, (ii) there exist subspaces ${\mathfrak}{m}_a, {\mathfrak}{m}_b\subset{\mathfrak}{m}$ ($a, b=1, \dots ,s$) such that $[{\mathfrak}{m}_a, {\mathfrak}{m}_b]\subset {\mathfrak}{m}_a$ and (iii) the restriction of the endomorphism $I$ into ${\mathfrak}{m}_a\oplus{\mathfrak}{m}_b$ is determined by an element in the center of ${\mathfrak}{h}$. [*Mathematics Subject Classification.*]{} Primary 53C25; Secondary 53C30. [*Keywords*]{}: Charged particle; electromagnetic field; geodesic; homogeneous space; fibration; generalized flag manifold; Hopf bundle address: 'University of Patras, Department of Mathematics, GR-26500 Patras, Greece' author: - Andreas Arvanitoyeorgos and Nikolaos Panagiotis Souris title: Motion of charged particles in homogeneous fibrations --- Introduction and statement of results ===================================== Let $(M=G/H, g)$ be a homogeneous Riemannian manifold, where $g$ is a $G$-invariant metric with corresponding Levi-Civita connection $\nabla$. The aim of the present paper is the study of the “charged particle" differential equation $$\label{cp} \nabla_{\dot{x}}{\dot{x}}=kI(\dot{x}),$$ where $k$ is some real constant and $I$ an endomorphism of $T_oM$, the tangent space of $M$ at $o=eH$. Equation (\[cp\]) appears in a more general context in general relativity as follows ([@MTW]). Let $(M,g)$ be a Riemannian manifold, $F$ a closed $2$-form, and $X$ a vector field on $M$. We denote by $\iota _X:\Lambda ^p(M)\to\Lambda ^{p-1}(M)$ the interior product operator induced by $X$, and by $\mathcal{L}: TM\to T^*M$ the Legendre transformation defined by $u\mapsto\mathcal{L}(u)$, $\mathcal{L}(u)(v)=g(u,v)$ ($v\in TM$). A curve $x(t)$ in $M$ is called a [*motion of a charged particle under electromagnetic field $F$*]{} if it satisfies the differential equation $$\nabla _{\dot{x}}\dot{x}=-\mathcal{L}^{-1}(\iota _{\dot{x}}F),$$ where $\nabla$ is the Levi-Civita connection of $M$. When $F=0$ then $x(t)$ is a geodesic in $M$. In particular, if $M$ is a Kähler manifold with complex structure $J$ there is a natural choice of an electromagnetic field $F$, namely a scalar multiple of the Kähler form $\omega$, defined by $\omega (X,Y)=g(X, JY)$. Since $-\mathcal{L}^{-1}(\iota _X\omega)=JX$, a curve $x(t)$ is a motion of a charged particle under electromagnetic field $\kappa\omega$ if and only if $ \nabla _{\dot{x}}\dot{x}=kJ(\dot{x})$. We also refer to [@KST] and [@Li] for other relevant applications in physics. Differential equation (\[cp\]) has been studied by O. Ikawa for various homogeneous spaces (cf. [@Ik1], [@Ik2], [@Ik3], [@Ik4]). A class of homogeneous spaces considered in [@Ik3] were generalized flag manifolds (or Kähler C-spaces) with two isotropy summands. These spaces were classified by the first author and I. Chrysikos in [@Arv-Chr], hence obtaining a concrete class of homogeneous spaces where equation (\[cp\]) can be solved. In the present article we solve differential equation (\[cp\]) for a large class of homogeneous spaces, which are described as follows. Let $M=G/H$ be a homogeneous space with reductive decomposition ${\mathfrak}{g}={\mathfrak}{h}\oplus {\mathfrak}{m}$ with respect to an $\operatorname{Ad}$-invariant inner product $B$ of ${\mathfrak}{g}$ and assume that the Lie algebra ${\mathfrak}{h}$ has non trivial center ${\mathfrak}{z}({\mathfrak}{h})$. The tangent space $T_oM$ at $o=eH$ can be identified to ${\mathfrak}{m}$. Let $\pi: G\rightarrow G/H$ be the projection and for $p\in G$, let $\tau_p:G/H\rightarrow G/H$ be the left translation in $G/H$ by $p$. We assume that the following conditions are satisfied: \(i) The tangent space ${\mathfrak}{m}$ admits an $\operatorname{Ad}(H)$-invariant and $B$-orthogonal decomposition $$\label{dec}{\mathfrak}{m}={\mathfrak}{m}_1\oplus \cdots \oplus {\mathfrak}{m}_s.$$ We endow $G/H$ with the $G$-invariant Riemannin metric $g$ corresponding to the positive definite inner product $$\label{metric}\langle \ , \ \rangle=\lambda_1B|_{{\mathfrak}{m}_1}+\cdots + \lambda_sB|_{{\mathfrak}{m}_s}, \quad \lambda_i>0 \quad (i=1,\dots,s),$$ and let $\nabla$ be the corresponding Riemannian connection of $g$. \(ii) There exist subspaces ${\mathfrak}{m}_a, {\mathfrak}{m}_b$, ($a,b=1,\dots , s$) of ${\mathfrak}{m}$ such that $$\label{mod}[{\mathfrak}{m}_a,{\mathfrak}{m}_b]\subset {\mathfrak}{m}_a.$$ \(iii) For $W\in {\mathfrak}{z}({\mathfrak}{h})$ we define the endomorphism $I_o:{\mathfrak}{m}\rightarrow {\mathfrak}{m}$ such that $$\label{end}I_o|_{{\mathfrak}{m}_a\oplus {\mathfrak}{m}_b}=\operatorname{ad}(W)|_{{\mathfrak}{m}_a}+\frac{1}{\lambda}\operatorname{ad}(W)|_{{\mathfrak}{m}_b}, \quad \lambda=\frac{\lambda_b}{\lambda_a}.$$ Since $\operatorname{Ad}(h)I_o=I_o\operatorname{Ad}(h)$ for all $h\in H$, we can extend $I_o$ to a $G$-invariant (1,1)-tensor $I$ on $G/H$ by defining $$\label{tensor}\quad I_{\pi(p)}V_{\pi(p)}:=((\tau_p)_*\circ I_o \circ (\tau_{p^{-1}})_*)V_{\pi(p)}, \quad V_{\pi(p)}\in T_{\pi(p)}(G/H).$$ We denote a homogeneous space satisfying conditions (i) – (iii) by $(G/H, g, I, \lambda)$. For $k$ a real constant a curve $x(t)$ is called a motion of a charged particle under the electromagnetic field $kI$ if it is a solution of the differential equation $$\label{diff} \nabla_{\dot{x}}{\dot{x}}=kI(\dot{x}).$$ Note that if $k=0$ then $x(t)$ is a geodesic. We prove the following: \[maintheorem\] Let $(G/H, g, I, \lambda)$ be a Riemannian homogeneous space satisfying conditions (i), (ii) and (iii). Let $x(t)$ be the motion of a charged particle given by [(\[diff\])]{.nodecor} with initial conditions $$\label{in}x(0)=o \quad \makebox{and} \quad \dot{x}(0)=X_a+X_b,$$ where $X_a\in{\mathfrak}{m}_a, X_b\in{\mathfrak}{m}_b$. Then the curve $x(t)$ is given by $$\label{curve}x(t)=\exp t(X_a+\lambda X_b+kW) \exp t(1-\lambda)(X_b+\frac{k}{\lambda}W)\cdot o.$$ The above Theorem generalizes Ikawa’s result. As a consequence, we obtain the following description of corresponding motion in homogeneous fibrations: \[sub\] Let $G$ be a Lie group admitting a bi-invariant Riemannian metric and let $B$ be the corresponding $\operatorname{Ad}$-invariant positive definite inner product on ${\mathfrak}{g}$. Let $K,H$ be closed and connected subgroups of $G$, such that $H\subset K\subset G$ and such that the Lie algebra of $H$ has non trivial center. We identify the tangent spaces $T_o(G/H),T_o(G/K)$ and $T_o(K/H)$ with corresponding subspaces ${\mathfrak}{m},{\mathfrak}{m}_1$ and ${\mathfrak}{m}_2$ of ${\mathfrak}{g}$, such that ${\mathfrak}{m}={\mathfrak}{m}_1\oplus {\mathfrak}{m}_2$. We endow $G/H$ with a $G$-invariant Riemannian metric $g_{\lambda}$ corresponding to the $\operatorname{Ad}(H)$-invariant positive definite inner product $$\langle \ ,\ \rangle=\left.B\right|_{{\mathfrak}{m}_1}+\lambda \left.B\right|_{{\mathfrak}{m}_2}, \quad \lambda>0,$$ on ${\mathfrak}{m}$. Moreover, for $W\in {\mathfrak}{z}({\mathfrak}{h})$, let $I^W$ be the $G$-invariant $(1,1)$-tensor on $G/H$ such that $I^W_o=\operatorname{ad}(W)|_{{\mathfrak}{m}_1}+\frac{1}{\lambda}\operatorname{ad}(W)|_{{\mathfrak}{m}_2}$. Let $X=X_1+X_2\in {\mathfrak}{m}$ with $X_i\in {\mathfrak}{m}_i$, $i=1,2$. Then the motion of a charged particle in $G/H$ under electromagnetic field $kI$ with initial conditions $x(0)=o$ and $\dot{x}(0)=X_1+X_2$ is the curve $x:\mathbb R\rightarrow G/H$ given by $$\label{curve1}x(t)=\exp t(X_1+\lambda X_2+kW) \exp t(1-\lambda)(X_2+\frac{k}{\lambda}W)\cdot o.$$ We will prove Theorems \[maintheorem\] and \[sub\] in Section 2. In Section 3 we will give examples of homogeneous spaces satisfying the conditions of Theorem \[sub\]. The work was supported by Grant $\#E.037$ from the Research Committee of the University of Patras (Programme K. Karatheodori). Proof of the main results ========================= We need to show that $$\label{have}g(\nabla_{\dot{x}}{\dot{x}},V)=g(kI(\dot{x}),V),$$ for any vector field $V$ in $G/H$. By using Koszul’s formula the left-hand side of (\[have\]) is given by $$\label{koz1} g(V,\nabla_{\dot{x}}{\dot{x}})=\dot{x}g(V,\dot{x})+g(\dot{x},[V,\dot{x}])-\frac{1}{2}Vg(\dot{x},\dot{x}).$$ We set $X=X_a+\lambda X_b+ kW$, $Y=(1-\lambda)(X_b+\frac{k}{\lambda}W)$, and $\alpha:\mathbb R\rightarrow G \quad \makebox{with} \quad \alpha(t)=\exp tX \exp tY$ so $x=\pi\circ \alpha$. We also consider the one-parameter family of automorphisms $$\label{set} T:\mathbb R\rightarrow \operatorname{Aut}(G) \quad \makebox{with} \quad T(t)=\operatorname{Ad}(\exp(-tY)).$$ We will need the following: \[lem\] The following relations are satisfied: 1\)  $T(t)X_a\in {\mathfrak}{m}_a$, $t\in \mathbb R$, 2\)  $\dot{x}(t)=(\tau_{\alpha(t)})_*(T(t)X_a+X_b)$. To prove 1) we recall that $$\label{an}T(t)X_a=\sum_{n=0}^{\infty}{\frac{t^n}{n!}\operatorname{ad}^n(-Y)X_a}.$$ Moreover, by taking into account relation (\[mod\]) and the $\operatorname{ad}({\mathfrak}{h})$-invariance of the subspace ${\mathfrak}{m}_a$ we deduce that $[Y,{\mathfrak}{m}_a]\subset {\mathfrak}{m}_a$. Therefore, for any $N\in \mathbb N$, it is $\sum_{n=0}^{N}{\frac{t^n}{n!}\operatorname{ad}^n(-Y)X_a}\in {\mathfrak}{m}_a$, so by taking the limit $N\rightarrow \infty$ in (\[an\]) we obtain that $T(t)X_a\in {\mathfrak}{m}_a$. We now prove 2). For $p\in G$, let $L_p,R_p$ be the left and right translations respectively in $G$ by $p$. We compute: $$\begin{aligned} \dot{x}(t)&=&\pi_*(\dot{\alpha}(t))=\pi_*(\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\alpha(t+s)})=\pi_*(\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp (t+s)X \exp (t+s)Y})\\ &=&\pi_*(\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp (t+s)X \exp tY} +\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp tX \exp (t+s)Y})\\ &=& \pi_*(\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp sX \exp tX \exp tY} + \displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp tX \exp tY \exp sY}) \\ &=&\pi_*(\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\exp sX \alpha(t)}+ \displaystyle{\left.\frac{d}{ds}\right|_{s=0}\alpha(t) \exp sY})=\pi_*(\displaystyle{(R_{\alpha(t)})_*(X)+(L_{\alpha(t)})_*Y})\\ &=&\displaystyle{(\pi\circ L_{\alpha(t)}\circ L_{\alpha(t)^{-1}})_*(\displaystyle{(R_{\alpha(t)})_*(X)+(L_{\alpha(t)})_*Y})=(\pi\circ L_{\alpha(t)})_*(\operatorname{Ad}(\alpha(t)^{-1})X+Y)}\\ &=&\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\alpha(t)^{-1})X+Y)}=\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY)\exp(-tX))X+Y)}\\ &=&\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY))X+Y)}\\ &=&\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY))X_a+\operatorname{Ad}(\exp(-tY))(\lambda X_b+kW)+Y)}\\ &=&\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY))X_a+\operatorname{Ad}(\exp(-tY))(\frac{\lambda}{1-\lambda}Y)+Y)}\\ &=&\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY))X_a+\frac{\lambda}{1-\lambda}Y+Y)}=\displaystyle{(\tau_{\alpha(t)}\circ \pi)_*(\operatorname{Ad}(\exp(-tY))X_a+X_b+\frac{k}{\lambda}W)}\\ &=&\displaystyle{(\tau_{\alpha(t)})_*(\operatorname{Ad}(\exp(-tY))X_a+X_b)}=\displaystyle{(\tau_{\alpha(t)})_*(T(t)X_a+X_b)}.\end{aligned}$$ For any $p\in G$ we observe that the vector field $$\label{ext}\pi(p)\mapsto (\tau_p\circ \pi)_*(\operatorname{Ad}(p^{-1})X+Y),$$ is a well defined local extension of $\dot{x}(t)$ in $G/H$, therefore, the vector field $\nabla_{\dot{x}}\dot{x}$ is well defined. Moreover, in view of equation (\[koz1\]), since $T_{\pi(p)}(G/K)$ admits a basis of left invariant vectors, it suffices to take $V$ as $$\label{vf}V_{\pi(p)}=(\tau_p)_*Z, \quad Z\in {\mathfrak}{m}.$$ [*Proof of Theorem \[maintheorem\].*]{} We will first simplify each term of the right-hand side of equation (\[koz1\]). By using Lemma \[lem\], equations (\[vf\]) and (\[metric\]) as well as the $G$-invariance of the metric $g$, the first term of the right-hand side of (\[koz1\]) becomes $$\begin{aligned} \displaystyle{\dot{x}(t)g(V,\dot{x})}&=&\displaystyle{\left.\frac{d}{ds}\right|_{s=0}g((\tau_{\alpha(t+s)})_*Z,(\tau_{\alpha(t+s)})_*(T(t+s)X_a+X_b))}\nonumber \\ &=& \displaystyle{\left.\frac{d}{ds}\right|_{s=0}\langle Z,T(t+s)X_a+X_b\rangle}=\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\langle Z,T(s)T(t)X_a+X_b\rangle}\nonumber \\ &=& \displaystyle{\langle Z,[T(t)X_a,Y] \rangle}=\displaystyle{(1-\lambda)\langle Z,[T(t)X_a,X_b+\frac{k}{\lambda}W] \rangle}\nonumber\\ &=& \displaystyle{(1-\lambda)\lambda_aB(Z,[T(t)X_a,X_b+\frac{k}{\lambda}W])}\nonumber \\ &=& \displaystyle{(\lambda_a-\lambda_b)B(Z,[T(t)X_a,X_b+\frac{k}{\lambda}W])}.\label{t1}\end{aligned}$$ Similarly, and by using the $B$-orthogonality of ${\mathfrak}{h}$ and ${\mathfrak}{m}$ as well as the $\operatorname{Ad}$-invariance of $B$, the second term of the right-hand side of (\[koz1\]) becomes $$\begin{aligned} \displaystyle{g(\dot{x}(t),[V,\dot{x}]_{x(t)})}&=&\displaystyle{g((\tau_{\alpha(t)})_*(T(t)X_a+X_b),[(\tau_{\alpha(t)})_*Z,(\tau_{\alpha(t)})_*(T(t)X_a+X_b)]_{{\mathfrak}{m}})}\nonumber\\ &=& \displaystyle{g((\tau_{\alpha(t)})_*(T(t)X_a+X_b),(\tau_{\alpha(t)})_*[Z,T(t)X_a+X_b]_{{\mathfrak}{m}})}\nonumber \\ &=& \displaystyle{\langle T(t)X_a+X_b,[Z,T(t)X_a+X_b]_{{\mathfrak}{m}} \rangle}\nonumber \\ &=& \displaystyle{\lambda_aB( T(t)X_a,[Z,T(t)X_a]_{{\mathfrak}{m}})}+\displaystyle{\lambda_aB( T(t)X_a,[Z,X_b]_{{\mathfrak}{m}}})\nonumber\\ && +\displaystyle{\lambda_bB(X_b,[Z,T(t)X_a]_{{\mathfrak}{m}})}+\displaystyle{\lambda_bB( X_b,[Z,X_b]_{{\mathfrak}{m}})}\nonumber \\ &=& \displaystyle{\lambda_aB( T(t)X_a,[Z,T(t)X_a])}+\displaystyle{\lambda_aB( T(t)X_a,[Z,X_b])}\nonumber \\ && +\displaystyle{\lambda_bB(X_b,[Z,T(t)X_a])}+\displaystyle{\lambda_bB( X_b,[Z,X_b])}\nonumber \\ &=& \displaystyle{(\lambda_b-\lambda_a)B(Z,[T(t)X_a,X_b])} \label{t2},\end{aligned}$$ where for the last equality we used the $\operatorname{Ad}$-invariance of $B$.\ Finally, for the third term of the right hand side of equation (\[koz1\]), we use the local extension (\[ext\]) of $\dot{x}$ as well as (\[vf\]). We have that $$\begin{aligned} \displaystyle{V_{x(t)}g(\dot{x},\dot{x})}&=&\displaystyle{((\tau_{\alpha(t)})_*Z)g(\dot{x},\dot{x})}\nonumber \\ &=& \displaystyle{\left.\frac{d}{ds}\right|_{s=0}g(((\tau_{p^{-1}})_*\circ\pi_*)(\operatorname{Ad}(p^{-1})X+Y),((\tau_{p^{-1}})_*\circ\pi_*)(\operatorname{Ad}(p^{-1})X+Y))}\nonumber \\ &=& \displaystyle{\langle \pi_*(\operatorname{Ad}(p^{-1})X+Y),\pi_*(\operatorname{Ad}(p^{-1})X+Y) \rangle},\label{dom} \end{aligned}$$ where $p=\alpha(t)\exp sZ$. Moreover, we have that $$\begin{aligned} \displaystyle{\left.\frac{d}{ds}\right|_{s=0}\pi_*(\operatorname{Ad}(p^{-1})X)}&=&\displaystyle{\left.\frac{d}{ds}\right|_{s=0}\pi_*(\operatorname{Ad}(\exp(-sZ))T(t)X)}\nonumber \\ &=& \displaystyle{\pi_*(\left.\frac{d}{ds}\right|_{s=0}(\operatorname{Ad}(\exp(-sZ))T(t)X))}\nonumber \\ &=& \displaystyle{[T(t)X,Z]_{{\mathfrak}{m}}}=\displaystyle{[T(t)X_a+\lambda X_b+ \frac{k}{\lambda}W,Z]_{{\mathfrak}{m}}}.\label{en}\end{aligned}$$ By substituting equation (\[en\]) into equation (\[dom\]) and by using Lemma \[lem\] and the $\operatorname{Ad}$-invariance of $B$ we obtain that $$\begin{aligned} \displaystyle{\frac{1}{2}V_{x(t)}g(\dot{x},\dot{x})}&=&-\displaystyle{ \langle [T(t)X_a+\lambda X_b+ \frac{k}{\lambda}W,Z]_{{\mathfrak}{m}}, T(t)X_a+X_b \rangle}\nonumber \\ &=& -\displaystyle{\langle [T(t)X_a,Z]_{{\mathfrak}{m}},T(t)X_a \rangle-\langle [T(t)X_a,Z]_{{\mathfrak}{m}},X_b \rangle-\lambda \langle [X_b,Z]_{{\mathfrak}{m}},T(t)X_a \rangle}\nonumber \\ && -\displaystyle{\lambda \langle [X_b,Z]_{{\mathfrak}{m}},X_b \rangle-\frac{k}{\lambda} \langle [W,Z]_{{\mathfrak}{m}},T(t)X_a \rangle-\frac{k}{\lambda}\langle [W,Z]_{{\mathfrak}{m}},X_b \rangle}\nonumber \\ &=& -\displaystyle{\lambda_aB( [T(t)X_a,Z]_{{\mathfrak}{m}},T(t)X_a)-\lambda_bB( [T(t)X_a,Z]_{{\mathfrak}{m}},X_b)-\lambda \lambda_aB( [X_b,Z]_{{\mathfrak}{m}},T(t)X_a)}\nonumber \\ && -\displaystyle{\lambda \lambda_bB( [X_b,Z]_{{\mathfrak}{m}},X_b)-\frac{k}{\lambda}\lambda_aB( [W,Z]_{{\mathfrak}{m}},T(t)X_a)-\frac{k}{\lambda}\lambda_bB( [W,Z]_{{\mathfrak}{m}},X_b)}\nonumber \\ &=& -\displaystyle{\lambda_aB( [T(t)X_a,Z],T(t)X_a)-\lambda_bB( [T(t)X_a,Z],X_b)-\lambda \lambda_aB( [X_b,Z],T(t)X_a)}\nonumber \\ && -\displaystyle{\lambda \lambda_bB( [X_b,Z],X_b)-\frac{k}{\lambda}\lambda_aB( [W,Z],T(t)X_a)-\frac{k}{\lambda}\lambda_bB( [W,Z],X_b)}\nonumber \\ &=& -\displaystyle{\frac{k}{\lambda}\lambda_aB( [W,Z],T(t)X_a)-\frac{k}{\lambda}\lambda_bB( [W,Z],X_b)}.\nonumber \\ &=& -\displaystyle{\frac{k}{\lambda}\lambda_aB(Z,[T(t)X_a,W])-\frac{k}{\lambda}\lambda_bB(Z,[X_b,W])}.\label{t3}\end{aligned}$$ By substituting equations (\[t1\]), (\[t2\]) and (\[t3\]) in equation (\[koz1\]) and by adding these we obtain that $$\label{lh}g(V,\nabla_{\dot{x}}\dot{x})=\displaystyle{-k\lambda_aB(Z,[TX_a+X_b,W])}.$$ Next, we simplify the right-hand side of equation (\[have\]). By using relations (\[metric\]), (\[end\]) and (\[tensor\]), Lemma \[lem\], the $G$-invariance of $g$ and the $\operatorname{Ad}$-invariance of $B$, we obtain that $$\begin{aligned} \displaystyle{g(kI(\dot{x}),V)}&=&\displaystyle{kg(((\tau_{\alpha(t)})_*\circ I_o)(T(t)X_a+X_b),(\tau_{\alpha(t)})_*Z)}\nonumber \\ &=& \displaystyle{k\langle I_o(T(t)X_a+X_b),Z \rangle}=\displaystyle{k\langle [W,T(t)X_a],Z \rangle+\frac{k}{\lambda}\langle [W,X_b],Z \rangle}\nonumber \\ &=& \displaystyle{k\lambda_aB([W,T(t)X_a],Z)+\frac{k}{\lambda}\lambda_bB([W,X_b],Z)}\nonumber \\ &=& \displaystyle{-k\lambda_aB(Z,[TX_a+X_b,W])}\nonumber,\end{aligned}$$ which, by equation (\[lh\]) is equal to $g(V,\nabla_{\dot{x}}\dot{x})$. Therefore, relation (\[have\]) holds and the proof of Theorem \[maintheorem\] is completed. $\Box$ *Proof of Theorem \[sub\]* Let ${\mathfrak}{h},{\mathfrak}{k},{\mathfrak}{g}$ be the Lie algebras of the groups $H,K,G$ respectively. The subspaces ${\mathfrak}{m}_1$ and ${\mathfrak}{m}_2$ can be obtained from the $B$-orthogonal decompositions $$\label{123}{\mathfrak}{g}={\mathfrak}{k}\oplus {\mathfrak}{m}_1 \quad \makebox{and} \quad {\mathfrak}{k}={\mathfrak}{h}\oplus {\mathfrak}{m}_2,$$ such that $$\label{124}\operatorname{Ad}(K){\mathfrak}{m}_1\subset {\mathfrak}{m}_1, \quad \operatorname{Ad}(H){\mathfrak}{m}_2\subset {\mathfrak}{m}_2.$$ Therefore, the decomposition ${\mathfrak}{m}={\mathfrak}{m}_1\oplus {\mathfrak}{m}_2$ is $\operatorname{Ad}(H)$-invariant and $B$-orthogonal. Moreover, since $\operatorname{Ad}(K){\mathfrak}{m}_1\subset {\mathfrak}{m}_1$, we have that $[{\mathfrak}{m}_1,{\mathfrak}{k}]\subset {\mathfrak}{m}_1$, therefore, $$\label{ta}[{\mathfrak}{m}_1,{\mathfrak}{m}_2]\subset [{\mathfrak}{m}_1,{\mathfrak}{k}]\subset {\mathfrak}{m}_1.$$ By taking into account relation (\[ta\]), it is straightforward to check that the space\ $(G/H,g_{\lambda},I^W,\lambda)$ satisfies the conditions of Theorem \[maintheorem\], and this completes the proof. Examples ======== Lie groups ---------- Let $G$ be a Lie group admitting a bi-invariant Riemannian metric $g$. Let $K$ be a connected subgroup of $G$ and let ${\mathfrak}{g},{\mathfrak}{k}$ be the Lie algebras of $G,K$ respectively. The bi-invariant metric $g$ corresponds to an $\operatorname{Ad}$-invariant positive definite inner product $B$ on ${\mathfrak}{g}$, which induces an orthogonal decomposition ${\mathfrak}{g}={\mathfrak}{k}\oplus {\mathfrak}{m}$. This decomposition is $\operatorname{Ad}(K)$-invariant. Indeed, for any $X_{{\mathfrak}{k}},Y_{{\mathfrak}{k}}\in {\mathfrak}{k}$ and $X_{{\mathfrak}{m}}\in {\mathfrak}{m}$, we set $Z_{{\mathfrak}{k}}=[X_{{\mathfrak}{k}},Y_{{\mathfrak}{k}}]\in {\mathfrak}{k}$. Then $$B([X_{{\mathfrak}{k}},X_{{\mathfrak}{m}}],Y_{{\mathfrak}{k}})=-B(X_{{\mathfrak}{m}},[X_{{\mathfrak}{k}},Y_{{\mathfrak}{k}}])=-B(X_{{\mathfrak}{m}},Z_{{\mathfrak}{k}})=0.$$ It follows that $[X_{{\mathfrak}{k}},X_{{\mathfrak}{m}}]\in {\mathfrak}{m}$, and since $K$ is connected then $\operatorname{Ad}(K){\mathfrak}{m}\subset {\mathfrak}{m}$, which in turn gives that ${\mathfrak}{g}={\mathfrak}{k}\oplus {\mathfrak}{m}$. We view $G$ as the homogeneous space $G/\left\{ {e} \right\}$ and we endow $G$ with the left invariant metric $g_{\lambda}$ corresponding to the positive definite inner product $$\langle \ ,\ \rangle=\left.B\right|_{{\mathfrak}{m}}+\lambda \left.B\right|_{{\mathfrak}{k}}, \quad \lambda>0.$$ By taking ${\mathfrak}{m}_1={\mathfrak}{m}$ and ${\mathfrak}{m}_2={\mathfrak}{k}$ in Theorem \[sub\], then we obtain the motion of a charged particle in a Lie group $G$. Ikawa’s result -------------- Let $M=G/H$ be a homogeneous space satisfying the conditions of Theorem \[maintheorem\] with ${\mathfrak}{m}={\mathfrak}{m}_1\oplus{\mathfrak}{m}_2$. Then we obtain Ikawa’s result [@Ik3 Theorem 1.1] under the only condition $[{\mathfrak}{m}_1, {\mathfrak}{m}_2]\subset {\mathfrak}{m}_1$. Hopf bundles ------------ Let $G=\operatorname{U}(n+1)$, $H=\operatorname{U}(n)$ and $K=\operatorname{U}(n)\times \operatorname{U}(1)$. Then the fibration $K/H \rightarrow G/H \rightarrow G/K$ is the homogeneous Hopf bundle $$\label{fi}\mathbb S^1 \rightarrow \mathbb S^{2n+1} \rightarrow \mathbb CP^n.$$ Since $\operatorname{U}(n+1)$ is compact, it admits a bi-invariant metric corresponding to an $\operatorname{Ad}(\operatorname{U}(n+1))$-invariant positive definite inner product $B$ on ${\mathfrak}{u}(n+1)$. We identify each of the spaces $T_o\mathbb S^{2n+1}= T_o(G/H),T_o\mathbb CP^n= T_o(G/K)$, and $T_o\mathbb S^1= T_o(K/H)$ with corresponding subspaces ${\mathfrak}{m},{\mathfrak}{m}_1$, and ${\mathfrak}{m}_2$ of ${\mathfrak}{u}(n+1)$. Consider the one parameter family of metrics $g_{\lambda}$ on $\mathbb S^{2n+1}$ corresponding to the positive definite inner products $$\label{o}\langle \ ,\ \rangle=\left.B\right|_{{\mathfrak}{m}_1}+\lambda\left.B\right|_{{\mathfrak}{m}_2}, \quad \lambda>0$$ on ${\mathfrak}{m}={\mathfrak}{m}_1\oplus {\mathfrak}{m}_2$. Note that for $\lambda=1$ the inner product (\[o\]) gives the standard metric $g_1$ on $\mathbb S^{2n+1}$. Then the hypotheses of Theorem \[sub\] are satisfied, hence the curve (\[curve1\]) describes the motion of a charged particle in $(\mathbb S^{2n+1},g_{\lambda})$. However, it is known (\[Ya\]) that $G/H=\operatorname{U}(n+1)/\operatorname{U}(n)$ is a weakly symmetric space, hence in this case it can be shown that motion (\[curve1\]) reduces to $x(t)=\exp t(a+X_1+X_2+kW)\cdot o$, $X_i\in{\mathfrak}{m}_i$ $(i=1,2)$ and $a\in {\mathfrak}{h}$, which depends on $X_1, X_2$. Twistor fibrations ------------------ For $G$ semisimple let $F=G/H$ be a generalised flag manifold, i.e. an adjoint orbit of an element $W$ in ${\mathfrak}{g}$. It is known ([@Bu-Ra]) that any flag manifold $F=G/H$ can be fibered over a compact inner symmetric space $G/K$ ($H\subset K$) under the twistor fibration $\pi:G/H\rightarrow G/K$. The [*normal metric*]{} of $G/H$ is the $G$-invariant metric induced by the negative of the Killing form ${\mathfrak}{g}$, denoted by $B$. We endow $G/H$ with the “deformation" of the normal metric along the fibers $K/H$ of the twistor fibration , given by $$\langle \ ,\ \rangle=\left. B\right|_{T_o(G/K)}+\lambda \left. B\right|_{T_o(K/H)}, \quad \lambda>0.$$ Then, by virtue of Theorem \[sub\], the equation of motion of a charged particle in $(F,\langle \ ,\ \rangle)$ under electromagnetic field $kI$, is given by (\[curve1\]). [50]{} A. Arvanitoyeorgos and I. Chrysikos: [*Motion of charged particles and homogeneous geodesics in Kähler C-Spaces with two isotropy summands*]{}, Tokyo J. Math. 32(2) (2009) 487–500. F.E. Burstall and J.H. Rawnsley: [*Twistor Theory for Riemannian Symmetric Spaces*]{}, Lecture Notes in Math. 1424 (1990), Springer. O. Ikawa: [*Hamiltonian dynamics of a charged particle*]{}, Hokkaido Math. J. 32 (3) (2003) 661–671. O. Ikawa: [*Motion of charged particles in homogeneous spaces*]{}, Proc. Seventh Intern. Workshop in Diff. Geometry 7 (2003) 29–40. O. Ikawa: [*Motion of charged particles in Kähler $C$-spaces*]{}, Yokohama Math. J. 50 (2003) 31–39. O. Ikawa: [*Motion of charged particles in homogeneous Kähler and homogeneous Sasakian manifolds*]{}, Far East J. Math. Sci.(FJMS) 14 (3) (2004) 283–302. M. Kološ, Z. Stuchlík and A. Tursunov: [*Quasi-harmonic oscillatory motion of charged particles around a Schwarzschild black hole immersed in a uniform magnetic field*]{}, Classical Quantum Gravity, 16 165009, 22pp. Y-K. Lim: [*Motion of charged particles around a magnetized/electrified black hole*]{}, Phys. Rev. D 91(2) (2015) 024048, 15pp. C.W. Misner, K.S. Thorne and J.A. Wheeler: [*Gravitation*]{}, Freeman, San Francisco 1973. O.S. Yakimova: [*Weakly symmetric Riemannian manifolds with a reductive isometry group*]{}, Sb. Math. 195 (2004) 599-–614.
{ "pile_set_name": "ArXiv" }
--- abstract: '[In this paper we revisit the derivation of a nonlocal interfacial Hamiltonian model for systems with short-ranged intermolecular forces. Starting from a microscopic Landau-Ginzburg-Wilson Hamiltonian with a double parabola potential, we reformulate the derivation of the interfacial model using a rigorous boundary integral approach. This is done for three scenarios: a single fluid phase in contact with a nonplanar substrate (i.e., wall); a free interface separating coexisting fluid phases (say, liquid and gas); and finally a liquid-gas interface in contact with a nonplanar confining wall, as is applicable to wetting phenomena. For the first two cases our approaches identifies the correct form of the curvature corrections to the free energy and, for the case of a free interface, it allows us to recast these as an interfacial self-interaction as conjectured previously in the literature. When the interface is in contact with a substrate our approach similarly identifies curvature corrections to the nonlocal binding potential, describing the interaction of the interface and wall, for which we propose a generalized and improved diagrammatic formulation. ]{}' author: - 'José M. Romero-Enrique' - Alessio Squarcini - 'Andrew O. Parry' - 'Paul M. Goldbart' title: 'Curvature corrections to the nonlocal interfacial model for short-ranged forces' --- Introduction ============ While significant progress has been made in the past few decades in understanding the statistical mechanics of inhomogeneous fluids and related interfacial phenomena [@Evans; @Dietrich; @Schick; @Forgacs], from a fundamental perspective many challenges remain for theory. Techniques based on molecular methods such as computer simulations [@Lee; @Rowley] and density-functional theory [@Evans; @Evans2] are wide spread, but under some circumstances large length scales emerge which make the use of mesoscopic models, often referred to as effective interfacial Hamiltonian or capillary-wave models, much more convenient and useful [@Buff; @Weeks; @Bedeaux]. These in fact have been pivotal in the development of the fluctuation theory of the thermal-wandering-induced roughness associated with a free or weakly pinned liquid-gas interface [@Buff] and also the classification of critical singularities occurring at continuous surface phase transitions such as wetting [@Dietrich; @Schick; @Forgacs; @BHL; @FH; @delfino1] and wedge filling [@parry10; @parry11; @parry12; @parry13; @romero1; @romero2; @romero3; @romero4; @romero5; @delfino2]. The search for a link between truly microscopic approaches and these mesoscopic descriptions can be traced back to van der Waals [@Rowlinson] and continue to this day, and in the past few years considerable effort has been invested in establishing this connection more rigorously. For example, intrinsic sampling methods use a many-body percolative, approach to identify the interfacial position from the underlying microscopic molecular configurations, and this has been extensively used in simulations [@tarazona1; @tarazona2; @tarazona3; @tarazona4; @tarazona5; @tarazona6; @tarazona7; @tarazona8; @tarazona9; @tarazona10; @tarazona11; @tarazona12; @tarazona13; @tarazona14]. A second, related, development has been the attempt to systematically derive an interfacial model for wetting transitions in settings involving short-ranged intermolecular forces from a more microscopic starting point [@parry1; @parry2; @parry3; @parry4; @parry5; @parry6; @parry7]. The need for this was originally driven by the significant discrepancy between initial predictions of strong nonuniversality for three-dimensional (3D) critical wetting, based on renormalization-group studies of local, partly phenomenologically justified, interfacial models [@BHL; @FH], and the findings of more microscopic Ising model simulation studies, which only reported minor deviations from mean-field-like behavior [@binder]. In attempting to explain this, Fisher and Jin [@Jin1; @Jin3] set out a very useful systematic basis for the derivation of an interfacial model from an underlying continuum Landau-Ginzburg-Wilson (LGW) Hamiltonian. Their idea was to introduce a constraint that specifies the interfacial configuration \[which we denote by $\ell({\mathbf{x}})$\] from that of the more microscopic order parameter $m({\mathbf{r}})$. Different options are available, such as the crossing criterion, in which $\ell({\bf{x}})$ is identified as the surface on which the order parameter takes some specified value or, alternatively, integral criteria which are generalized measures of the local adsorption. Once the interface is defined, the interfacial Hamiltonian $H[\ell]$ is identified via the partial trace: $$e^{-\beta H[\ell]}=\int D me^{-\beta \mathcal{H}_{LGW}[m]} \approx e^{-\beta \mathcal{H}_{LGW}[m_\Xi]},$$ where $m_\Xi(\bf{r})$ is the profile that *minimizes* the LGW Hamiltonian $\mathcal{H}_{LGW}[m]$, subject to the constraint and additional boundary conditions. The Fisher-Jin identification [@Jin1; @Jin3], generalized to nonplanar walls, will be the starting point for our entire investigation. Within this scheme, therefore, all that is required is the determination of the constrained profile $m_\Xi(\bf{r})$, which will be a functional of the interfacial configuration (and wall shape). Fisher and Jin obtained this for a planar wall by considering perturbations about the flat interfacial configuration. However, this perturbative approach is inadequate for the purposes at hand because it misidentifies the structure of corrections to the standard local interfacial model. Indeed, this leads to serious problems when carried forward in renormalization-group studies, where it erroneously alters the structure of the well-known global phase diagrams for wetting [@Nakanishi]. Later, it was appreciated that the solutions to the constrained mean-field-like equations for $m_{\Xi}(\bf{r})$ could be reformulated using Green’s functions [@parry1; @parry2]. This highlights immediately the [*nonlocal*]{} nature of the interaction of the interface and substrate, which has a simple diagrammatic representation. Furthermore, it resolves many of the problems associated with the fluctuation theory of critical wetting and in addition yields a description of correlation functions fully consistent with exact sum rules [@parry4; @parry5]. The predictions of this nonlocal model are also consistent with more recent Ising model simulations, which reported deviations from mean-field behavior for critical wetting [@binder2]. The nonlocal decay of order between a fluctuating interface and particles situated away from it, predicted by this approach, has been seen directly in both Ising model and molecular simulation studies [@binder3; @tarazona11]. In this paper we present an alternative and more rigorous derivation of the nonlocal interfacial model to that presented originally [@parry1; @parry2], which was still partly physically motivated. Our derivation is based on exact integral representations of the solutions of linear partial differential equations on a closed domain, which are cast as functionals of the solutions at the boundaries. This method, referred to as the boundary element method, has in fact been applied successfully to numerous engineering problems [@Brebbia; @Katsikadelis]. Here we develop an improved perturbative diagrammatic approach, which is related to the multiple reflection method used in the celebrated analysis of the wave equation yielding eigenfrequencies in a closed domain [@Balian; @Balian2; @Balian3]. When applying this methodology to the evaluation of the interfacial free energy and order-parameter profile of a fluid phase in contact with a structured substrate or of a constrained liquid-gas configuration, we recover, at leading order, the previous nonlocal model but now with curvature corrections. For the case of a constrained liquid-gas interface this leads naturally to an interfacial self-interaction precisely as has been conjectured [@PR]. The most detailed application of this method involves the rigorous determination of the binding potential functional for wetting films in contact with a nonplanar wall. Here, we identify an additional series of diagrams, not present in the original formulation, which arise when the substrate and liquid-gas interface are not parallel. These diagrams are resummed to obtain an alternative version of the nonlocal model which recovers the original version of the nonlocal model in certain limits. Our paper is arranged as follows. In Sec. II we present the theoretical framework and recall the mathematical tools used in our approach. In Sec. III we apply this to a single phase, which we take to be liquid, in contact with a wall to determine the mean-field excess free-energy functional $F_{wl}[\psi]$, which is a functional of the wall shape $\psi$. In Sec. IV we extend this to a free liquid-gas interface, and determine the interfacial Hamiltonian $H[\ell]$, which is a functional of the (constrained) interfacial configuration $\ell$. Finally, we consider the most involved situation, in which a constrained wetting film is located near to a nonplanar wall, and we determine the binding potential $W[\ell,\psi]$, which is a functional of both the interface and the wall shapes. Theoretical framework ===================== Consider the LGW Hamiltonian defined on a domain $\Omega$: $$\begin{aligned} \mathcal{H}_{LGW}[m]=\int_\Omega d\mathbf{r} \left\{\frac{1}{2} \left( \boldsymbol\nabla m\right)^2+ \Delta \phi(m)\right\}\nonumber\\ +\int_{\partial \Omega}\ \Phi_s(\mathbf{s}) d\mathbf{s}, \label{HLGW}\end{aligned}$$ where the shifted potential $\Delta \phi(m)$ corresponds to the excess contribution, with respect to the bulk, of the free-energy density, and $\Phi_s$ is a surface potential defined on the domain boundary $\partial \Omega$. Typically, $\Phi_s$ is taken to be a quadratic function of $m(\mathbf{s})$, i.e., $$\Phi_s(\mathbf{s})=-\frac{g}{2}\left[m(\mathbf{s})+\frac{h_1}{g}\right]^2,$$ where $h_1$ and $g$ are a local field and an enhancement parameter, respectively, modeling the coupling to the substrate (i.e., wall). Usually, these quantities are taken to be equal to their flat-wall counterparts, although additional curvature-induced terms may be included phenomenologically. Finally, we note that fixed (i.e., Dirichlet) boundary conditions correspond to the limit $h_1 \to \infty$ and $g\to -\infty$ with $-h_1/g = m_1$, where $m_1$ is the fixed value of the order parameter at the wall. In a zero external field (i.e., $h=0$) and below the bulk critical temperature $T_c$, the shifted potential $\Delta \phi$ has a familiar double-well structure, which we capture via the simple double-parabola (DP) approximation $$\Delta \phi (m) = \frac{\kappa^2}{2}\left(|m|-m_0\right)^2, \label{deltaphi}$$ where $\kappa$ is the inverse bulk correlation length and $m_0$ is the bulk order parameter. In this description, therefore, there are two bulk phases having order-parameter values $-m_0$ (which we regard as the gas phase) and $+m_0$ (which we regard as the liquid phase). For a general inhomogeneous situation, we will identify the phase at any point via the sign of the order parameter. Consequently, we will refer to the phase as gas if $m<0$ and liquid otherwise. Finally, we adopt a simple crossing criterion of a constrained interfacial configuration, whereby the interface is defined as the surface on which the order parameter vanishes, i.e., $m=0$ [@Jin1; @Jin3]. As the constrained profile $m_\Xi$ minimizes the LGW Hamiltonian, it satisfies the mean-field-like Euler-Lagrange equation $$\nabla^2 m_\Xi = \kappa^2(m_\Xi-m_b), \label{pde}$$ where $m_b=\pm m_0$, depending on whether the bulk phase is liquid or gas. This partial differential equation is to be solved subject to the natural boundary condition $$\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla m_\Xi (\mathbf{s})=h_1+gm_\Xi(\mathbf{s}), \label{boundary}$$ where $\mathbf{n}$ is the outward normal to the boundary of the integration domain $\Omega$. If fixed boundary conditions are applied on $\mathbf{s}$, we simply set $m_\Xi (\mathbf{s})=m_1$ instead of imposing Eq. (\[boundary\]). In order to obtain the mean-field free energy, we first consider the situation where there is only one phase in the integration domain. Multiplying Eq. (\[pde\]) by $(m_\Xi-m_b)/2$ and integrating over the domain $\Omega$, we get $$\int_\Omega d\mathbf{r} \frac{(m_\Xi-m_b)\nabla^2 m_\Xi}{2} = \int_\Omega \frac{\kappa^2}{2}(m_\Xi-m_b)^2 . \label{intpde}$$ We now use the identity $u\nabla^2 u + \boldsymbol{\nabla}u \mathbf{\cdot} \boldsymbol{\nabla} u = \boldsymbol{\nabla} \mathbf{\cdot} (u \boldsymbol{\nabla} u)$ with $u=m_\Xi-m_b$ and apply the divergence theorem to obtain $$\begin{aligned} \int_\Omega d\mathbf{r} \left\{\frac{1}{2} \left( \boldsymbol\nabla m_\Xi\right)^2+ \frac{\kappa^2}{2}(m_\Xi-m_b)^2\right\}\nonumber\\ =\oint_{\partial \Omega} d\mathbf{s} \frac{m_\Xi(\mathbf{s})-m_b}{2}[ \mathbf{n}\mathbf{\cdot}\boldsymbol\nabla m_\Xi (\mathbf{s})] . \label{mfHLGW0}\end{aligned}$$ Next we make use of Eq. (\[boundary\]) to rewrite the surface contribution to Eq. (\[HLGW\]) using $$\begin{aligned} \Phi_s(\mathbf{s})&=&-\frac{1}{2}\left(m_\Xi(\mathbf{s})+\frac{h_1}{g}\right) \left[\mathbf{n}\mathbf{\cdot}\boldsymbol\nabla m_\Xi(\mathbf{s})\right]\nonumber\\ =&-&\left[\frac{m_\Xi(\mathbf{s})-m_b}{2}+\frac{h_1+gm_b}{2g}\right][ \mathbf{n}\mathbf{\cdot}\boldsymbol\nabla m_\Xi (\mathbf{s})] .\end{aligned}$$ Hence, when evaluated at the equilibrium profile $m_\Xi$, the LGW Hamiltonian identifies the free energy as $$\begin{aligned} \mathcal{H}_{LGW}[m_\Xi]=\int_\Omega d\mathbf{r} \left\{\frac{1}{2} \left( \boldsymbol\nabla m_\Xi\right)^2+ \frac{\kappa^2}{2}(m_\Xi-m_b)^2\right\}\nonumber\\ +\oint_{\partial \Omega}\ \Phi_s(\mathbf{s}) d\mathbf{s}, \label{mfHLGWprev}\end{aligned}$$ and reduces to $$\mathcal{H}_{LGW}[m_\Xi] =-\frac{h_1+gm_b}{2g}\oint_{\partial \Omega} d\mathbf{s} \ \mathbf{n}\mathbf{\cdot} \boldsymbol\nabla m_\Xi (\mathbf{s}), \label{mfHLGW}$$ a result that will be central to our method. In the presence of a wetting layer of a different phase that intrudes between the wall and the bulk (see Fig. \[figure1\]) the domain $\Omega$ must be considered to be the union $\Omega = \cup_i \Omega_i$, where each appropriate subdomain $\Omega_i$ has boundaries $\partial \Omega_i$ that lie either on the substrate surface or at the liquid-gas interface. We define $\partial \Omega_i=\partial \Omega_{1,i} \cup \partial \Omega_{2,i}$, where the boundary condition (\[boundary\]) is satisfied in $\partial \Omega_{1,i}$ and $\partial \Omega_{2,i}$ corresponds to the gas-liquid interface. For this case, the generalization of Eq. (8) identifies the constrained free-energy functional for a given interfacial configuration as $$\begin{aligned} &&\mathcal{H}_{LGW}[m_\Xi]=\nonumber\\ &&=\sum_i \Bigg[\int_{\Omega_i} d\mathbf{r} \left\{\frac{1}{2} \left( \boldsymbol\nabla m_\Xi\right)^2+ \frac{\kappa^2}{2}[m_\Xi-(m_b)_i]^2\right\} \nonumber\\ &&\qquad\qquad\qquad\qquad+\int_{\partial \Omega_{1,i}} d\mathbf{s}\ \Phi_s(\mathbf{s}) \Bigg], \label{mfHLGW2prev}\end{aligned}$$ which reduces to $$\begin{aligned} \mathcal{H}_{LGW}[m_\Xi] =\sum_i \Bigg[-\frac{h_1+gm_b}{2g}\int_{\partial \Omega_{1,i}} d\mathbf{s} \ \mathbf{n}\mathbf{\cdot} \boldsymbol\nabla m_\Xi (\mathbf{s}) \nonumber\\ - \frac{(m_b)_i}{2}\int_{\partial \Omega_{2,i}} d\mathbf{s} [\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla m_\Xi (\mathbf{s})]\Bigg]. \label{mfHLGW2}\end{aligned}$$ The mean-field free energy corresponds to the interfacial configuration that gives the least free energy. In this sense, the free energy becomes a functional of the interfacial profile. The above results demonstrate that, for the potential, the equilibrium free energy of an interfacial configuration may be determined in terms of the normal derivatives of the order parameter at the substrate and liquid-gas interface (if present). This simplification is not so surprising, given that, for the DP potential, the whole order-parameter profile can be obtained formally in terms of the values at the boundaries. To see this, let us consider the (rescaled) Green’s function that solves $${\cal L}K (\mathbf{r},\mathbf{r}_0) \equiv \left(-\nabla^2_{\mathbf{r}} +\kappa^2\right) K (\mathbf{r},\mathbf{r}_0) =2\kappa \delta (\mathbf{r}-\mathbf{r}_0), \label{OZ}$$ and vanishes as $|\mathbf{r}-\mathbf{r}_0|\to \infty$. The subscript in $\nabla^2_{\mathbf{r}}$ denotes that the nabla operator acts on the argument $\mathbf{r}$ of $K$. Its solution is the Ornstein-Zernike correlation function $$K(\mathbf{r},\mathbf{r}_0)=\frac{\kappa}{2\pi} \frac{\exp(-\kappa|\mathbf{r}- \mathbf{r}_0|)}{|\mathbf{r}-\mathbf{r}_0|}. \label{OZ2}$$ The second Green’s identity for the Hermitian operator ${\cal L}$ states that $$\int_{\Omega} d\mathbf{r} \left[v{\cal L}u-u{\cal L}v\right]= -\int_{\partial \Omega} \left[ v (\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla u)-u(\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla v)\right] d\mathbf{s} \label{greenidentity}$$ for any domain $\Omega$ with boundary $\partial \Omega$, where the outward normal is $\mathbf{n}$ and $u$ and $v$ are arbitrary functions. If we choose $u(\mathbf{r})= K(\mathbf{r},\mathbf{r}_0)/2\kappa$ and $v(\mathbf{r})=m_\Xi(\mathbf{r})-m_b$, and taking into account Eqs. (\[pde\]) and (\[OZ\]), then $$\begin{aligned} &&[m_\Xi(\mathbf{r})-m_b]\Theta_\Omega(\mathbf{r})=\frac{1}{2\kappa}\int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{r}) [\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla m_\Xi(\mathbf{s})]\nonumber\\&&-\frac{1}{2\kappa}\int_{\partial \Omega} d\mathbf{s} [m_\Xi(\mathbf{s})-m_b] [\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{r})], \label{greenidentity2}\end{aligned}$$ where $\Theta_\Omega(\mathbf{r})$ is the characteristic function of the set $\Omega$, i.e., $\Theta_\Omega (\mathbf{r})=1$ if $\mathbf{r}\in \Omega$, and $\Theta_\Omega (\mathbf{r})=0$ otherwise (excluding in both cases the boundary $\partial \Omega$). As before, $\partial \Omega_1$ refers to the portion of the boundary where Eq. (\[boundary\]) is satisfied, while $\partial \Omega_2$ lies on the appropriate side of the gas-liquid interface (see Fig. \[figure1\]). We can then recast Eq. (\[greenidentity2\]) as $$\begin{aligned} [m_\Xi(\mathbf{r})&-&m_b]\Theta_\Omega(\mathbf{r})\nonumber\\ &=&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(\frac{h_1}{g}+m_b\right) \partial_n K(\mathbf{s},\mathbf{r})\nonumber\\ &+&\frac{m_b}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} \partial_n K(\mathbf{s},\mathbf{r}) \nonumber\\ &+&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(K(\mathbf{s},\mathbf{r})-\frac{1}{g}\partial_n K(\mathbf{s},\mathbf{r})\right) \partial_n m_\Xi(\mathbf{s})\nonumber\\&+&\frac{1}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} K(\mathbf{s},\mathbf{r}) \partial_n m_\Xi(\mathbf{s}), \label{greenidentity3}\end{aligned}$$ where $\partial_n$ denotes the normal derivative $\mathbf{n}\mathbf{\cdot} \boldsymbol\nabla_{\mathbf{s}}$. What remains is the determination of the normal derivative $\partial_n m_\Xi$ at each point along the subdomain boundaries. However, Eq. (\[greenidentity3\]) itself cannot be used to determine these, and we must use a technique to modify this appropriately. To this end, we first place $\bf{r}$ at a boundary point [*and*]{} deform the boundary near it by cutting a circular hole of radius $\epsilon$ and adding a hemispherical cap atop it, so that the point is again inside the sub-domain under consideration. We then evaluate the order parameter at $\bf{r}$ and finally take the limit $\epsilon\to 0$. Assuming the interfaces are smooth, we obtain the integral equation within each domain $$\begin{aligned} &&\frac{m_\Xi(\mathbf{s}_0)-m_b}{2}= \frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(\frac{h_1}{g}+m_b\right) \partial_n K(\mathbf{s},\mathbf{s}_0)\nonumber\\ &+&\frac{m_b}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} \partial_n K(\mathbf{s},\mathbf{s}_0) \nonumber\\ &+&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(K(\mathbf{s},\mathbf{s}_0)-\frac{1}{g}\partial_n K(\mathbf{s},\mathbf{s}_0)\right) \partial_n m_\Xi(\mathbf{s})\nonumber\\&+&\frac{1}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} K(\mathbf{s},\mathbf{s}_0) \partial_n m_\Xi(\mathbf{s}), \label{greenidentity4}\end{aligned}$$ where the normal derivative of the Green’s function $K$ acts on its first argument and integration must be interpreted as its Cauchy principal value. Consequently, if $\mathbf{s}_0\in \partial \Omega_1$, Eq. (\[greenidentity4\]) can be written as $$\begin{aligned} &&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(\frac{h_1}{g}+m_b\right) \left[\partial_n K(\mathbf{s},\mathbf{s}_0)+\kappa\delta(\mathbf{s}-\mathbf{s}_0)\right] \nonumber\\ &+&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \Bigg(K(\mathbf{s},\mathbf{s}_0)-\frac{1}{g}\partial_n K(\mathbf{s},\mathbf{s}_0)\nonumber\\&-&\frac{\kappa}{g}\delta(\mathbf{s}-\mathbf{s}_0)\Bigg) \partial_n m_\Xi(\mathbf{s}) +\frac{m_b}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} \partial_n K(\mathbf{s},\mathbf{s}_0) \nonumber\\&+&\frac{1}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} K(\mathbf{s},\mathbf{s}_0) \partial_n m_\Xi(\mathbf{s})=0. \label{greenidentity5}\end{aligned}$$ Similarly, if $\mathbf{s}_0\in \partial \Omega_2$, Eq. (\[greenidentity4\]) reads $$\begin{aligned} &&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(\frac{h_1}{g}+m_b\right) \partial_n K(\mathbf{s},\mathbf{s}_0)\nonumber\\ &+&\frac{m_b}{2\kappa}\left(\kappa+\int_{\partial \Omega_2} d\mathbf{s} \partial_n K(\mathbf{s},\mathbf{s}_0)\right) \nonumber\\ &+&\frac{1}{2\kappa}\int_{\partial \Omega_1} d\mathbf{s} \left(K(\mathbf{s},\mathbf{s}_0)-\frac{1}{g}\partial_n K(\mathbf{s},\mathbf{s}_0)\right) \partial_n m_\Xi(\mathbf{s})\nonumber\\&+&\frac{1}{2\kappa}\int_{\partial \Omega_2} d\mathbf{s} K(\mathbf{s},\mathbf{s}_0) \partial_n m_\Xi(\mathbf{s})=0. \label{greenidentity6}\end{aligned}$$ Under some circumstances, e.g., for certain boundary conditions, Eq. (\[greenidentity3\]) is not the most convenient representation of the constrained order-parameter profile. Another possible representation is the single-layer potential. Let us assume that the order parameter on the boundary $\partial \Omega$ is known. Now we determine the solutions to the Helmholtz equation inside and outside $\Omega$ with the same Dirichlet boundary conditions. We can use Eq.  (\[greenidentity2\]) for these problems, keeping in mind that the normal derivatives are different in each problem. Adding these equations, we find the representation, valid everywhere in space, $$m_\Xi(\mathbf{r})=m_b +\frac{1}{2\kappa}\int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{r}) \Psi(\mathbf{s}), \label{singlelayer}$$ where $\Psi(\mathbf{s})=[\partial_n m_\Xi(\mathbf{s})]^{+}-[\partial_n m_\Xi(\mathbf{s})]^{-}$, with the plus (minus) sign standing for the interior (exterior) problem to $\Omega$, and $\mathbf{n}(\mathbf{s})$ is the outward normal from $\Omega$. The auxiliary function $\Psi(\mathbf{s})$ can be obtained from the boundary integral equation $$m_\Xi(\mathbf{s})=m_b+\frac{1}{2\kappa} \int_{\partial \Omega} d\mathbf{s_0} K(\mathbf{s}_0,\mathbf{s})\Psi(\mathbf{s}_0). \label{singlelayer2}$$ The normal derivatives of the order-parameter profile on the boundary can be related to $\Psi$ as $$[\partial_n m_\Xi(\mathbf{s})]^{\pm}=\pm \frac{\Psi(\mathbf{s})}{2}+\frac{1}{2\kappa} \int_{\partial \Omega} d\mathbf{s_0} \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol\nabla_{\mathbf{s}} K(\mathbf{s}_0,\mathbf{s})\Psi(\mathbf{s}_0).$$ Alternatively, a double-layer potential representation of the order-parameter profile can be obtained if the normal derivative of the order parameter on the boundary $\partial \Omega$ is known. We use Eq. (\[greenidentity2\]) for the solutions to the Helmholtz equation inside and outside $\Omega$ with opposite Neumann boundary conditions. Note that the outward normal for every domain is the inward normal for the other one. By adding these equations we again obtain a representation that is valid everywhere in space: $$\delta m_\Xi(\mathbf{r})=\frac{1}{2\kappa}\int_{\partial \Omega} d\mathbf{s} \mathbf{n}(\mathbf{s})\mathbf{\cdot} \boldsymbol\nabla_\mathbf{s} K(\mathbf{s},\mathbf{r}) \overline{\Psi}(\mathbf{s}), \label{doublelayer}$$ where $\delta m_\Xi(\mathbf{r})\equiv m_\Xi(\mathbf{r})-m_b$ and $m_b=\pm m_0$ is the appropriate bulk order parameter in the region containing $\mathbf{r}$. Here the modified auxiliary function $\overline{\Psi}(\mathbf{s})=[\delta m_\Xi(\mathbf{s})]^{-}- [\delta m_\Xi(\mathbf{s})]^{+}$. The limits of the order parameter on each side of $\partial \Omega$ are related to $\overline{\Psi}$ as $$\delta m_\Xi(\mathbf{s}^{\pm})=\mp\frac{\overline{\Psi}(\mathbf{s})}{2}+\frac{1}{2\kappa} \int_{\partial \Omega} d\mathbf{s_0} \mathbf{n}(\mathbf{s}_0)\mathbf{\cdot} \boldsymbol\nabla_{\mathbf{s}_0}K(\mathbf{s}_0,\mathbf{s})\overline{\Psi}(\mathbf{s}_0). \label{doublelayer2}$$ On the other hand, $\partial_n m_\Xi(\mathbf{s})\equiv \mathbf{n}(\mathbf{s})\mathbf{\cdot} \boldsymbol\nabla_\mathbf{s} m_\Xi$ is continuous on $\partial \Omega$: $$\partial_n m_\Xi(\mathbf{s})=\mathbf{n}(\mathbf{s})\mathbf{\cdot} \boldsymbol\nabla_\mathbf{s}\left[ \int_{\partial \Omega} d\mathbf{s_0} \overline{\Psi} (\mathbf{s}_0) \mathbf{n}(\mathbf{s}_0)\mathbf{\cdot}\boldsymbol\nabla_{\mathbf{s}_0} \frac{K(\mathbf{s}_0,\mathbf{s})}{2\kappa} \right]$$ Finally, we provide some additional relations which will be useful later. On using the Green’s identity (\[greenidentity\]) for two Green’s functions, it follows that $$\begin{aligned} \int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{r}) \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{r}')\nonumber\\ =\int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{r}') \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{r}) , \label{greenidentity7}\end{aligned}$$ where $\mathbf{r}$ and $\mathbf{r}'$ are positions inside the domain $\Omega$. If $\mathbf{r}'\to \mathbf{s}'$ on the boundary $\partial \Omega$ then Eq. (\[greenidentity7\]) leads to $$\begin{aligned} &&\int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{s}') \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{r})\nonumber\\ &&=-\kappa K(\mathbf{s}',\mathbf{r})+ \int_{\partial \Omega} d\mathbf{s} K(\mathbf{s},\mathbf{r}) \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{s}') . \label{greenidentity8}\end{aligned}$$ Finally, if $\mathbf{r}\to \mathbf{s}$ on $\partial \Omega$, then $$\begin{aligned} &&\int_{\partial \Omega} d\mathbf{s} K(\mathbf{s}_0,\mathbf{s}) \mathbf{n}(\mathbf{s}_0)\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}_0} K(\mathbf{s}_0,\mathbf{s}')\nonumber\\&& = \int_{\partial \Omega} d\mathbf{s}_0 K(\mathbf{s}_0,\mathbf{s}') \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}_0} K(\mathbf{s}_0,\mathbf{s}) . \label{greenidentity9}\end{aligned}$$ In the following sections we apply this formalism to obtain the interfacial free energies relevant to wetting phenomena: (i) the interfacial free energy of a nonwetting bulk phase in contact with a rough substrate, (ii) the self-interaction corresponding to a free liquid-gas interface, and finally (iii) the binding potential for a wetting film configuration (see Fig. \[figure1\]). ![Schematic illustration of a nonplanar interfacial configuration (blue line) for a constrained wetting film of liquid at a nonplanar wall (black line). Conventions for the surface normals are shown. The inset shows the double-parabola approximation for $\Delta \phi (m)$. \[figure1\]](figure1.eps){width="8.6cm"} Interfacial free energy of a liquid phase in contact with a nonplanar wall ========================================================================== The first system that we consider is the simple case of a bulk phase in contact with a nonplanar wall when a wetting film is absent. The local height of the wall, above some reference plane (often taken to be the plane $z=0$), is written $\psi({\mathbf{x}})$, where ${\mathbf{x}}=(x,y)$ is the parallel displacement. Without loss of generality, we concentrate on the wall-liquid interface, supposing that the local surface field $h_1$ is positive so that the order parameter has the same (positive) value throughout. In this case, the domain $\Omega$ is just the set of points for which $z>\psi(\mathbf{x})$. In addition, we suppose that the substrate is chemically homogeneous, so $h_1$ and $g$ do not vary with position. The equilibrium mean-field configuration $m_\Xi(\bf{r})$ follows from the simple minimization of the LGW Hamiltonian, resulting in the Helmholtz equation (\[pde\]) and the boundary conditions $$\begin{aligned} &&\mathbf{n}_\psi\mathbf{\cdot}\boldsymbol \nabla m_\Xi (\mathbf{r})=-h_1-gm_\Xi(\mathbf{r}), \ \textrm{for}\ \mathbf{r}=(\mathbf{x},\psi(\mathbf{x})); \label{boundary2}\\ &&m_\Xi(z)\to m_b \qquad \textrm{for}\quad z\to +\infty, \label{boundary2b}\end{aligned}$$ where the bulk magnetization for the liquid phase is $m_b=m_0$. Similar results apply to the wall-gas interface, for which $h_1$ is negative and $m_b=-m_0$. Here $\mathbf{n}_\psi$ denotes the *inward* normal to the wall. Since the order parameter does not change sign in $\Omega$, Eq. (\[greenidentity5\]) can be recast as $$\begin{aligned} \int_\psi d\mathbf{s} \left[K(\mathbf{s}, \mathbf{s}^\prime) +\frac{1}{g} \partial_{n} K(\mathbf{s},\mathbf{s}^\prime) -\frac{\kappa}{g}\delta(\mathbf{s}-\mathbf{s}^\prime) \right] q(\mathbf{s})\nonumber\\ =\left(\frac{-h_1}{g}-m_b\right)\left[-\kappa+ \int_\psi d\mathbf{s} \partial_{n} K(\mathbf{s},\mathbf{s}^\prime)\right], \label{inteq1}\end{aligned}$$ where the integration $\int_\psi$ is over the substrate surface, $\partial_{n}(\mathbf{s})$ is shorthand for $\mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol{\nabla}_{\mathbf{s}}$, $q\equiv \partial_{n} \delta m_\Xi$, and $\delta m_\Xi\equiv m_\Xi-m_b$. Equation (\[inteq1\]) can only be solved exactly in a few exceptional circumstances, such as when symmetry arguments can be applied; these include planar, cylindrical, or spherical substrates, all of which have constant curvature. For example, for a planar substrate $q$ is constant over the surface and has the value $$q=-\frac{\kappa (h_1+gm_b)}{\kappa-g}. \label{qflat}$$ However, the generic solution of Eq. (\[inteq1\]) must include the *local* curvature of the substrate, and it is natural to look for a perturbative solution when the deviations from the flat case are small. To this end, let us introduce the principal curvatures $k_{1}(\mathbf{s})$ and $k_{2}(\mathbf{s})$ at a point $\mathbf{s}=(\mathbf{x},\psi(\mathbf{x}))$ on the surface. Here $R_{i}=1/\kappa_{i}$ are the corresponding radii of curvature and it is convenient to recall that $K_{G}=k_{1}k_{2}$ is the Gaussian curvature and $H=(k_{1}+k_{2})/2$ is the mean curvature (or half of the total curvature). Let us denote by $R$ the minimum of $|R_1|$ and $|R_2|$ so that $H \sim R^{-1}$. Far from the bulk critical point the bulk correlation length $\kappa^{-1}$ is microscopically small, so the substrate can be considered flat over several correlation lengths provided that $\kappa R \gg 1$. Perturbative approach {#sec_3A} --------------------- We now set out our perturbative analysis of Eq. (\[inteq1\]). The idea is to expand all elementary building blocks of Eq. (\[inteq1\]) \[the Ornstein-Zernike (OZ) kernel, its normal derivative $q$, and $d \mathbf{s}$\] on the left-hand side and right-hand side of Eq. (\[inteq1\]) in powers of the curvature $H$, which can then be equated, term by term. We suppose that locally the surface is well approximated by a paraboloid in a neighborhood of $\mathbf{s}^{\prime}$, where we locate the origin of the coordinates. Consider now a point on the substrate surface $\mathbf{s} = (\mathbf{x},\psi(\mathbf{x}))$. The vertical displacement of $\mathbf{s}$ with respect to the horizontal plane is $$\label{verticaldisplacement} \Delta\psi(\mathbf{r}_{\bot}) \equiv \frac{1}{2}k_{1}x^{2} + \frac{1}{2}k_{2}y^{2} +\cdots,$$ where we have written $\mathbf{r}_{\bot}=\mathbf{x}-\mathbf{x}_{0}\equiv(x,y)$ for the projection of the vector $\mathbf{s}-\mathbf{s}^{\prime}$ onto the horizontal plane $\pi_{0}$, tangent to the graph of $\psi(\mathbf{x})$ in $\mathbf{s}^{\prime}$. With this parametrization, the coefficients $k_{i}$ are exactly the principal curvatures $k_{i}(\mathbf{s}^{\prime})$ evaluated at $\mathbf{s}^\prime$. The ellipsis in Eq. (\[verticaldisplacement\]) stands for higher-order terms in $(x,y)$, which are coupled to higher orders of the local curvatures as well. We assume that the Taylor coefficients associated with terms $x^m y^{n-m}$ (with $0\le m \le n$ and $n>3$) scale as $R^{-n+1}$. With this property, close to the origin $\Delta \psi(\mathbf{r}_{\bot}) = R \Delta \bar{\psi} (\mathbf{r}_{\bot}R^{-1})$, i.e., $R$ is the only relevant length scale for the substrate shape. Let us consider first the OZ kernel. The two points are separated by the distance $|\mathbf{s}-\mathbf{s}^{\prime}| = \sqrt{\mathbf{r}_{\bot}^{2} + [\Delta \psi(\mathbf{r}_{\bot})]^{2}}$ and for small curvatures we can Taylor expand around the flat configuration to express the OZ kernel as a power series in the curvature: $$K(\mathbf{s},\mathbf{s}^{\prime}) = \mathcal{K}(r_{\bot}) \Biggl[ 1 - \frac{1}{2}(1+\kappa r_{\bot}) \left(\frac{\Delta\psi(\mathbf{r}_{\bot})}{r_{\bot}}\right)^{2} + {O}(R^{-4}) \Biggr] , \label{expk}$$ where $\mathcal{K}(x)=\kappa e^{-\kappa x}/2\pi x$. As the kernel $\mathcal{K}$ decays exponentially with a lengthscale $\kappa^{-1}$, Eq. (\[expk\]) is a faithful representation of $K(\mathbf{s},\mathbf{s}^{\prime})$ around $\mathbf{s}^{\prime}$ if $\kappa R \gg 1$. For the normal derivative of the kernel we have $$\partial_{n} K(\mathbf{s},\mathbf{s}^{\prime}) = \mathbf{n}(\mathbf{s}) \mathbf{\cdot}\boldsymbol \nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{s}^{\prime}) = \mathbf{n}(\mathbf{s}) \mathbf{\cdot} \frac{\mathbf{s}-\mathbf{s}^{\prime}}{|\mathbf{s}-\mathbf{s}^{\prime}|} \, \frac{\partial\mathcal{K}(|\mathbf{s}-\mathbf{s}^{\prime}|)}{\partial|\mathbf{s}-\mathbf{s}^{\prime}|} .$$ The Monge parametrization of the normal vector is $$\begin{aligned} \mathbf{n}(\mathbf{s}) &=& \frac{(-\boldsymbol\nabla_{\bot}\Delta\psi(\mathbf{r}_{\bot}),1)}{\sqrt{1+\left[\boldsymbol\nabla_{\bot}\Delta\psi(\mathbf{r}_{\bot})\right]^{2}}},\end{aligned}$$ where $\boldsymbol\nabla_\bot\equiv (\partial_{x},\partial_{y})$. Then, following the above ideas, we can show that $$\begin{aligned} \partial_{n} K(\mathbf{s},\mathbf{s}^{\prime}) = \frac{1+\kappa r_{\bot}}{r_{\bot}} \mathcal{K}(r_{\bot}) \left(\frac{ \mathbf{r}_{\bot}\mathbf{\cdot}\boldsymbol\nabla_{\bot}\Delta \psi- \Delta\psi}{r_{\bot}}\right) \nonumber \\ + {O}(R^{-3}) , \label{expk2}\end{aligned}$$ where the term in large parentheses is ${O}\left(R^{-1}\right)$ \[see Eq. (\[verticaldisplacement\])\]. The surface element $d \mathbf{s}=\sqrt{1+(\boldsymbol\nabla_{\bot}\psi)^{2}}d \mathbf{r}_{\bot}= [1+(\boldsymbol\nabla_{\bot}\Delta \psi)^{2}/2]d \mathbf{r}_{\bot} + {O}(R^{-4})$. The normal derivative of the order parameter can be expanded in a similar way; thus, $q(\mathbf{s}) = \sum_{n=1}^{\infty} q_{n}(\mathbf{s})$, where $q_{n} = {O}(R^{-n})$. Plugging the above relations into Eq. (\[inteq1\]) and identifying the corresponding terms, order by order, we find a recursive chain of integral equations for $q_{n}(\mathbf{s})$ of the form (see Appendix \[appendix\_A\]) $$\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{n}(\mathbf{r}_\bot) \left[\mathcal{K}(r_{\bot}) - \frac{\kappa}{g}\delta (\mathbf{r}_{\bot})\right] = f_n[q_0,\ldots,q_{n-1}],$$ where we have extended the integral to $\mathbb{R}^{2}$, ignoring terms exponentially decaying in $\kappa R$. In general, $f_n$ is a functional of $q_i$ for $i<n$. For $n=0$, $f_0=\kappa(h_1+gm_b)/g$, which is independent of $\mathbf{s}^\prime$, so $q_0(\mathbf{s})$ is given by Eq. (\[qflat\]) everywhere on the substrate. Following the procedure outlined in Appendix \[appendix\_A\], we find for the next-order terms $$\begin{aligned} q_1&=&q_0 \frac{g}{\kappa - g}\frac{H}{\kappa}, \label{q1}\\ q_2&=& q_0 \frac{g}{\kappa-g} \left(\frac{H^2}{2\kappa^2}\left(1+\frac{2\kappa}{\kappa-g}\right)-\frac{K_G}{2\kappa^2}\right). \label{q2}\end{aligned}$$ We are now in the position to estimate the interfacial thermodynamic properties and the order-parameter profile. The interfacial free energy $\mathcal{F}_{wl}$ of the wall-liquid interface is obtained from Eq. (\[mfHLGW\]) as $$\begin{aligned} {\mathcal{F}}_{wl}&=&\frac{h_1+gm_b}{2g}\int_\psi d\mathbf{s}\ q(\mathbf{s}) =\sigma_{wl} {\cal A} + \Delta \mathcal{F}_{wl}[\psi], \label{fsw}\end{aligned}$$ where $\sigma_{wl}\equiv(\kappa/2)(h_1+gm_b)^2/g(g-\kappa)$ is the surface tension defined for a planar wall-liquid interface and ${\cal A}$ is the total substrate area. Thus, the increment $\Delta \mathcal{F}_{wl}[\psi]$ accounts for all the curvature-related terms. For large $\kappa R$, $\Delta \mathcal{F}_{wl}$ can be expressed as $$\begin{aligned} \frac{\Delta \mathcal{F}_{wl}}{\sigma_{wl} {\cal A}} &=& \frac{g}{\kappa-g} \frac{\overline{H}}{\kappa} + \frac{1}{2}\left(\frac{g}{\kappa-g}\right)\left(1+\frac{2\kappa}{\kappa-g}\right)\frac{\overline{H^2}}{\kappa^2} \nonumber\\&-& \frac{1}{2}\left(\frac{g}{\kappa-g}\right)\frac{\overline{K_G}}{\kappa^2} + \cdots, \label{deltafwlexpansion}\end{aligned}$$ where $\overline{H}$, $\overline{H^2}$ and $\overline{K_G}$ are the averages over the substrate of the mean curvature, its square, and the Gaussian curvature, respectively. The leading order is consistent with the expression obtained in Refs. [@parry2; @parry3]. Finally, the ellipsis corresponds to higher-order curvature contributions which, in general, are nonvanishing. This feature, as well as the contribution being proportional to $\overline{H^2}$, is nonzero, which implies that the DP model does not satisfy the morphological thermodynamics hypothesis for confined fluids of hard bodies [@konig] (see also Ref. [@blokhuis] for a critical review of this proposal). Diagrammatically, the interfacial free energy can be represented as $$\begin{aligned} \label{free_energy_sub} \mathcal{F}_{wl}=\sigma_{wl}\Bigg[ {\parbox{1.5cm}{\epsfig{file=diagram1.eps,width=1.5cm}}}+\left(\frac{g}{\kappa-g}\right) {\parbox{1.5cm}{\epsfig{file=diagram2.eps,width=1.5cm}}}\nonumber\\ + \frac{1}{2}\left(\frac{g}{\kappa-g}\right)\left(1+\frac{2\kappa}{\kappa-g}\right) {\parbox{1.5cm}{\epsfig{file=diagram3.eps,width=1.5cm}}} \nonumber\\ - \frac{1}{2}\left(\frac{g}{\kappa-g}\right) {\parbox{1.5cm}{\epsfig{file=diagram4.eps,width=1.5cm}}}+ \cdots \Bigg] ,\end{aligned}$$ where the wavy line corresponds to the substrate surface and the closed circle means that one must integrate over all the positions on that surface with the appropriate infinitesimal area element. The closed triangle corresponds to integration over the surface, weighted by the local mean curvature in units of $\kappa$ (our notation for this symbol differs from that used in Ref. [@parry3] by a factor of $1/2$). Finally, for the closed square and the rhombus the weight function for the surface integrations are the squared mean curvature and the Gaussian curvature, respectively, in units of $\kappa^2$. The present treatment highlights nonzero bending rigidity and saddle-splay coefficients, which were missing in the original formulation of the nonlocal model [@parry2; @parry3]. The values of these are in agreement with those obtained from the exact solutions for the free energy of a fluid outside or inside a spherical or a cylindrical surface of radius $R$ within the DP model [@laura] (see also Appendix \[appendixe\]). General diagrammatic approach ----------------------------- We can go beyond the approach presented in the preceding section and obtain formally the full expansion of the interfacial free energy in powers of substrate curvatures. For this purpose, we return to Eq. (\[inteq1\]). The integral-equation kernel can be formally inverted as $$\begin{aligned} X(\mathbf{s},\mathbf{s}')\equiv\left(K(\mathbf{s},\mathbf{s}')+\frac{1}{g}\partial_{n} K(\mathbf{s},\mathbf{s}')-\frac{\kappa}{g} \delta(\mathbf{s}-\mathbf{s}')\right)^{-1}\nonumber\\ = \frac{g}{g-\kappa} \delta(\mathbf{s}-\mathbf{s}') \\-\frac{g}{g-\kappa} \int_{\psi} d\mathbf{s}_1 X(\mathbf{s},\mathbf{s}_1)\left(U(\mathbf{s}_1,\mathbf{s}')+ \frac{1}{g}\partial_{n_1} K(\mathbf{s}_1,\mathbf{s}')\right), \nonumber \label{seriesx}\end{aligned}$$ where $U(\mathbf{s},\mathbf{s}')\equiv K(\mathbf{s},\mathbf{s}')-\delta(\mathbf{s}-\mathbf{s}')$ is the barred kernel introduced in Ref. [@PR] and $\partial_{n_1} \equiv \mathbf{n}(\mathbf{s}_1)\mathbf{\cdot} \boldsymbol{\nabla}_{\mathbf{s}_1}$. This expression can be iterated, so we obtain the following formal expansion for $X$: $$\begin{aligned} X(\mathbf{s},\mathbf{s}')= \frac{g}{g-\kappa} \delta(\mathbf{s}-\mathbf{s}') \label{seriesx2} \\-\left(\frac{g}{g-\kappa}\right)^2 \left(U(\mathbf{s},\mathbf{s}')+ \frac{1}{g}\partial_{n} K(\mathbf{s},\mathbf{s}')\right) \nonumber\\ +\left(\frac{g}{g-\kappa}\right)^3 \int_{\psi} d\mathbf{s}_1 \left(U(\mathbf{s},\mathbf{s}_1)+ \frac{1}{g}\partial_{n} K(\mathbf{s},\mathbf{s}_1)\right) \nonumber\\\times \left(U(\mathbf{s}_1,\mathbf{s}')+ \frac{1}{g}\partial_{n_1} K(\mathbf{s}_1,\mathbf{s}')\right)+\cdots. \nonumber\end{aligned}$$ The solution to Eq. (\[inteq1\]) can be expressed as $$\begin{aligned} &&q(\mathbf{s})=-\frac{\kappa(h_1+gm_b)}{\kappa-g}\Bigg[1-\int_\psi d\mathbf{s}_1 d\mathbf{s}_2 X(\mathbf{s},\mathbf{s}_1) U(\mathbf{s}_1,\mathbf{s}_2)\nonumber\\ &&-\int_\psi d\mathbf{s}_1 d\mathbf{s}_2 X(\mathbf{s},\mathbf{s}_1) \frac{1}{\kappa}\partial_{n_1} K(\mathbf{s}_1,\mathbf{s}_2) \Bigg] . \label{qseries}\end{aligned}$$ When the expansion  (\[seriesx2\]) is introduced in the expression (\[qseries\]), we arrive at a formal series for $q$, where each term is proportional to the convolution of $n$ functions, each being either $U$ or $\partial_{n} K/\kappa$. We introduce the diagrammatic representation $$U(\mathbf{s},\mathbf{s}')={\parbox{1.5cm}{\epsfig{file=diagram5.eps,width=1.5cm}}} \quad,\quad \frac{1}{\kappa}\partial_{n} K(\mathbf{s},\mathbf{s}') ={\parbox{1.5cm}{\epsfig{file=diagram6.eps,width=1.5cm}}} ,$$ where in the latter diagram the arrow points to the position where the normal derivative is taken. In this way, the expansion terms appearing in $q$ can be represented as chainlike diagrams. For example, $$\int_\psi d\mathbf{s}_1 d\mathbf{s}_2 U(\mathbf{s},\mathbf{s}_1)\frac{1}{\kappa}\partial_{n_1} K(\mathbf{s}_1,\mathbf{s}_2) ={\parbox{1.5cm}{\epsfig{file=diagram7.eps,width=1.5cm}}},$$ where a closed circle corresponds to an integrated position and the open circle represents the point $\mathbf{s}$ where $q$ is evaluated. Thus, from Eq. (\[qseries\]) we have $$\begin{aligned} && q(\mathbf{s})=-\frac{\kappa(h_1+gm_b)}{\kappa-g}\Bigg[1+\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram8.eps,width=1.5cm}}}\nonumber\\ &&+ \frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram9.eps,width=1.5cm}}}+\frac{g}{\kappa-g}\frac{\kappa}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram10.eps,width=1.5cm}}} \nonumber\\&&+\frac{g}{\kappa-g}\frac{\kappa}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram11.eps,width=1.5cm}}} \nonumber\\&&+\left(\frac{g}{\kappa-g}\right)^2 \left({\parbox{1.5cm}{\epsfig{file=diagram7.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram12.eps,width=1.5cm}}}\right) + \cdots\Bigg] . \label{qpsidiagram}\end{aligned}$$ Here, each diagram with $n$ bonds (of which $m$ are of $\partial_{n} K/\kappa-$type) must be multiplied by a factor $$\left(\frac{g}{\kappa-g}\right)^n \left(\frac{\kappa}{g}\right)^{m-m_0}, \label{factor}$$ where the index $m_0$ is either $0$ or $1$, depending on whether the first bond on the left (i.e., the one that emerges from the closed extreme circle) is of $U$ or $\partial_{n} K/\kappa$ type, respectively. The connection with the curvature expansion is evident as, taking into account Eqs. (\[expk\]) and (\[expk2\]), we find that $$\begin{aligned} \label{dashedk} {\parbox{1.5cm}{\epsfig{file=diagram8.eps,width=1.5cm}}}&\equiv& {\parbox{1.5cm}{\epsfig{file=diagram13.eps,width=1.5cm}}}-1\\&=& \frac{1}{2} \biggl[ {\parbox{1.5cm}{\epsfig{file=diagram14.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram15.eps,width=1.5cm}}} \biggr] +{O}(R^{-4}), \nonumber\\ \label{dashedk2} {\parbox{1.5cm}{\epsfig{file=diagram16.eps,width=1.5cm}}}&=& {\parbox{1.5cm}{\epsfig{file=diagram17.eps,width=1.5cm}}} +{O}(R^{-3}),\end{aligned}$$ where by the open symbols we denote the evaluation of the corresponding weight functions at $\mathbf{s}$. Thus, a diagram with $n$, $U-$type and $m$, $\partial_{n} K/\kappa-$type bonds is of order of $R^{-(2n+m)}$. This demonstrates that in order to obtain the corrections to $q$ to order $R^{-2}$, only the first three diagrams in Eq. (\[qpsidiagram\]) are needed. By substituting Eq. (\[qpsidiagram\]) into Eq. (\[fsw\]) we obtain the diagrammatic expansion of the interfacial free energy $$\begin{aligned} \label{F_chain_1} &&\mathcal{F}_{wl}=\sigma_{wl}\Bigg[ {\parbox{1.5cm}{\epsfig{file=diagram1.eps,width=1.5cm}}}+\left(\frac{g}{\kappa-g}\right) ({\parbox{1.5cm}{\epsfig{file=diagram18.eps,width=1.5cm}}}+ {\parbox{1.5cm}{\epsfig{file=diagram19.eps,width=1.5cm}}}) \nonumber\\ &&+ \frac{g}{\kappa-g}\frac{\kappa}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram20.eps,width=1.5cm}}} +\cdots \Bigg],\end{aligned}$$ which coincides with our previous result  (\[free\_energy\_sub\]) up to $R^{-2}$ corrections. In this expansion, the factors that multiply each diagram are the same as those that multiply the corresponding diagrams in Eq. (\[qpsidiagram\]). Evaluation of the order-parameter profile ----------------------------------------- As for the interfacial free energy, we can obtain a formally exact expression for the order-parameter profile from Eq. (\[greenidentity3\]), $$\begin{aligned} \delta m_\Xi(\mathbf{r})&=& -\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} K(\mathbf{s},\mathbf{r}) q(\mathbf{s}) \label{profileprev} \\ &-&\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} \left(\frac{h_1}{g}+m_b+\frac{q(\mathbf{s})}{g}\right) \partial_{n} K(\mathbf{s},\mathbf{r}), \nonumber\end{aligned}$$ or, equivalently, $$\begin{aligned} \delta m_\Xi(\mathbf{r}) &=&-\frac{1}{2}\int_{\psi} d\mathbf{s} K(\mathbf{s},\mathbf{r}) \left(\frac{h_1}{g}+m_b+\frac{\kappa+g}{\kappa g}q(\mathbf{s})\right) \nonumber\\ &-&\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} \left(\frac{h_1}{g}+m_b+\frac{q(\mathbf{s})}{g}\right) \nonumber\\ &\times&\left[\partial_{n} K(\mathbf{s},\mathbf{r})-\kappa K(\mathbf{s},\mathbf{r})\right] \label{profile}\end{aligned}$$ by substitution of the expansion (\[qpsidiagram\]) for $q(\mathbf{s})$. However, it is more convenient to use the single-layer potential representation  (\[singlelayer\]). Note that, once we know $q$, the order parameter on the substrate can be obtained from Eq. (\[boundary2\]) as $-h_1/g-m_b - q/g$. It is instructive to derive the order parameter starting from the perturbative approach by expanding $\Psi$ in powers of the local curvature. The derivation proceeds along the same steps as the previous sections (see Appendix \[appendix\_A\]). After substitution of the expansion of $\Psi$ into Eq. (\[singlelayer\]), we find the following diagrammatic representation of the order parameter: $$\begin{aligned} \label{OP_curv_exp} \delta m_\Xi\approx \frac{h_1+gm_b}{\kappa-g}\Bigg[{\parbox{1.5cm}{\epsfig{file=diagram21.eps,width=1.5cm}}}+ \frac{\kappa}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram22.eps,width=1.5cm}}}\nonumber\\ +\left(\frac{1}{2}\frac{g}{\kappa-g}+\frac{\kappa^2}{(\kappa-g)^2}\right){\parbox{1.5cm}{\epsfig{file=diagram23.eps,width=1.5cm}}}\nonumber\\ -\frac{1}{2}\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram24.eps,width=1.5cm}}}\Bigg] .\end{aligned}$$ The presence of the curvature correction terms, which are not accounted for in the original nonlocal ansatz, can be checked again with the exact solutions known for spherical and cylindrical substrates (see Appendix \[appendixe\]). As for the free energy, we can also generate a general diagrammatic approach to the curvature corrections. For this purpose, Eq. (\[singlelayer2\]) can be formally solved as $$\Psi(\mathbf{s})=2\kappa \int_\psi d\mathbf{s}_0 K^{-1}(\mathbf{s},\mathbf{s}_0)\left(-\frac{h_1+gm_b}{g}- \frac{q(\mathbf{s}_0)}{g}\right), \label{defPsi}$$ where $K^{-1}(\mathbf{s},\mathbf{s}_0)=\lim_{g\to\-\infty} X(\mathbf{s},\mathbf{s}_0)$, i.e., $$K^{-1}(\mathbf{s},\mathbf{s}_0)=\delta(\mathbf{s}-\mathbf{s}_0)-U(\mathbf{s},\mathbf{s}_0)+\int_\psi d\mathbf{s}_1 U\mathbf{s},\mathbf{s}_1)U(\mathbf{s}_1,\mathbf{s}_0)-\cdots . \label{definvk}$$ Substituting the expansions (\[qpsidiagram\]) and (\[definvk\]) into Eq. (\[defPsi\]), and back into Eq. (\[singlelayer\]), we obtain the following expansion for the order-parameter profile: $$\begin{aligned} &&\delta m_\Xi=\frac{h_1+gm_b}{\kappa-g}\Bigg[{\parbox{1.5cm}{\epsfig{file=diagram21.eps,width=1.5cm}}}+\frac{\kappa}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram25.eps,width=1.5cm}}}\nonumber\\&+ &\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram26.eps,width=1.5cm}}} +\left(\frac{\kappa}{\kappa-g}\right)^2 {\parbox{1.5cm}{\epsfig{file=diagram27.eps,width=1.5cm}}} +\cdots\Bigg] . \label{deltamdiagram}\end{aligned}$$ The diagrams of this expansion are obtained by convolution of a chainlike diagram of $q$ from Eq. (\[qpsidiagram\]) and the Green’s function $K(\mathbf{s},\mathbf{r})$. The factor associated with each diagram \[except for the first one in Eq. (\[deltamdiagram\])\] is given by the product of two terms: the factor associated with the $q$ diagram, Eq. (\[factor\]), and either $1$ (if the chainlike diagram has only $U-$type bonds) or $1-(1-\kappa/g)^{l+1}$ otherwise, with $l$ being the number of $U-$type bonds from the last $\partial_n K/\kappa-$type bond to the extreme where the Green’s function $K(\mathbf{s},\mathbf{r})$ emerges. By using (\[dashedk\]) and (\[dashedk2\]), the diagrammatic expansion (\[deltamdiagram\]) reduces to (\[OP\_curv\_exp\]) up to $R^{-3}$ corrections. Summary and remarks ------------------- So far we have done two things. First, we devised a perturbative approach based on a small-curvature expansion of surfaces and we have applied this to a (nonwetting) bulk phase in contact with a substrate. The free energy (\[free\_energy\_sub\]) contains curvature corrections that we have identified exactly at the leading (nontrivial) order. The curvature expansions are, however, quite cumbersome, and to this end we developed a more fundamental approach, based on the formal inversion of integral equations satisfied by the order-parameter field. By using a diagrammatic approach we have found a formal expansion of the interfacial free energy \[see Eq. (\[F\_chain\_1\])\]. By following this approach we have also derived the order-parameter profile in the bulk phase \[see Eq. (\[deltamdiagram\])\]. For small interfacial or surface curvatures, the wetting diagrams entering into the formal expansions simplify and they reveal the curvature corrections in terms of local Gaussian and average curvatures; this property will be analyzed in detail in the next section for the case of an isolated liquid-gas interface. The free interface and its self-interaction energy functional ============================================================= Next we turn our attention to the liquid-gas interface. This is free in the sense that it is isolated, infinitely far from any confining walls but *constrained* so that the surface of zero magnetization adopts a given, smooth, nonplanar configuration $\ell(\bf{x})$. Overhangs are excluded and again we will suppose that the curvature is small everywhere. We follow the prescription set out by Fisher and Jin [@Jin1; @Jin3], whereby the effective interfacial model is identified as the minimum of the LGW model subject to this cross-criterion constraint together with the appropriate bulk boundary conditions, viz., that $m_\Xi(\mathbf{r})\to \mp m_0$ as $z\to\pm\infty$, i.e., gas is above and liquid is below the interface corresponding to the two domains $\Omega_1$ and $\Omega_2$. These regions are uncoupled, and in each the equilibrium constrained profile satisfies the Helmholtz equation (\[pde\]) subject to the above boundary conditions. In this case, the solution to Eq. (\[greenidentity6\]) can be written as $$\label{consistent} \int_{\ell} d \mathbf{s} \, K(\mathbf{s},\mathbf{s}^\prime) q^\pm(\mathbf{s}) = -m_{0} \Biggl( \kappa \mp \int_{\ell} d \mathbf{s} \, \partial_{n} K(\mathbf{s},\mathbf{s}^\prime) \Biggr) ,$$ where $\mathbf{n}$ is the interface normal towards the gas phase. This relation tells us that, given the interfacial profile $\ell(\mathbf{x})$ as a background, the order parameter can be found from knowledge of $q^\pm(\mathbf{s}_{\ell}) = \partial_{n_{\ell}}\delta m_{\Xi}(\mathbf{s_{\ell}^{\pm}})$, where $\delta m_\Xi=m(\mathbf{r})\pm m_0$ for $\mathbf{r}$ lying above or below the interface. Notice that the order parameter is a function of the position but also a *functional* of the interfacial shape $\ell(\mathbf{x})$. Our main goal is to determine this functional dependence. Perturbative approach {#perturbative-approach} --------------------- If we assume that the local interfacial curvatures are small, we can proceed in a similar way to the preceding section. The normal derivative of the order parameter can be expanded in powers of the minimum local curvature radius $R$; thus $q^{\pm}(\mathbf{s}_{\ell}) = \sum_{n=1}^{\infty} q_{n}^{\pm}(\mathbf{s}_{\ell})$, where $q_{n} = {O}(R^{-n})$. It follows that for a flat interface $q_{n}=0$ (for $n\geqslant1$), and the series reduces to $q^{\pm}(\mathbf{s}_{\ell})=q_{0}(\mathbf{s}_{\ell})$. Inserting the above relations into Eq. (\[consistent\]) and identifying the corresponding terms, order by order, we find a recursive chain of equations for the $q_{n}$’s. Following the scheme used in the preceding section, we can solve up to ${O}(R^{-3})$ and obtain the desired $q$’s (see Appendix \[appendix\_A\]). The DP potential allows us to write the LGW Hamiltonian in terms of surface integrals only, viz., $$\mathcal{H}_{LGW}[m_{\Xi}] \equiv H[\ell] = -\frac{m_{0}}{2} \int_{\ell} d \mathbf{s}_{\ell} \left[ q^{+}(\mathbf{s}_{\ell}) + q^{-}(\mathbf{s}_{\ell}) \right] , \label{finterface}$$ where $H[\ell]$ is the interfacial Hamiltonian. This functional can be evaluated with the perturbative expansion for $q^{\pm}$ mentioned above and, consequently, it leads to a similar expansion of the Hamiltonian that we cast in the form $H[\ell]=\sigma \mathcal{A}_{lg}+\Delta H[\ell]$, where $$\Delta H[\ell] \equiv -\sigma \sum_{n=2}^{\infty} (-1)^{n} \omega_{n}[\ell] ,$$ is the self-interaction contribution [@PR], $\sigma=\kappa m_0^2$ is the surface tension, and $\mathcal{A}_{lg}$ is the interfacial area. The functionals $\omega_{n}[\ell]$ are (ensured to be) of order ${O}(R^{-2(n-1)})$. Having in mind the approximate solutions for $q$ just derived, we find that $$H[\ell] = \kappa m_{0}^{2} \int_{\ell} d \mathbf{s}_{\ell} - \frac{m_{0}}{2} \int_{\ell} d \mathbf{s}_{\ell} \left[ q_{2}^{+}(\mathbf{s}_{\ell}) + q_{2}^{-}(\mathbf{s}_{\ell}) \right] + {O}(R^{-3}) .$$ Thus we immediately recover that $H[\ell]=\sigma\int d{\mathbf{s}}_\ell+\cdots$, which is just the standard capillary-wave Hamiltonian. The ellipsis stands for the energy corrections due to the self-interaction: The first of these corrections is $$\label{omega2} -\sigma \omega_{2}[\ell]=-\frac{m_{0}}{2} \int_{\ell} d \mathbf{s}_{\ell} \left[ q_{2}^{+}(\mathbf{s}_{\ell}) + q_{2}^{-}(\mathbf{s}_{\ell}) \right]$$ or, more explicitly, using Eq. (\[exph22\]), $$\label{ham1} H[\ell] \approx \sigma \int_{\ell} d \mathbf{s}_{\ell} - \frac{\sigma}{8} \int_{\ell} d \mathbf{s}_{\ell} \left( \frac{k_{1}(\mathbf{s}_{\ell})-k_{2}(\mathbf{s}_{\ell})}{\kappa} \right)^{2}.$$ We use the symbol $\approx$ to mean that relationships hold up to ${O}(R^{-3})$ corrections. The absence of a $1/R$ contribution, i.e., the vanishing of the Tolman length, is due to the Ising symmetry of the DP model. In the theory of lipid membranes the above functional (\[ham1\]) is commonly expressed in terms of the Gaussian and extrinsic curvatures of the interface, following Helfrich [@Helfrich], $$\label{Helfrich} H[\ell] \approx \sigma \mathcal{A}_{\ell} + \int_{\ell} d \mathbf{s}_{\ell} \left(\kappa_B H^{2} + \kappa_{G} K_G \right),$$ where the coupling constants are the bending rigidity $$\label{bendrig} \kappa_{B}=-\frac{\sigma}{2\kappa^{2}}$$ and the saddle-splay rigidity $$\label{saddlesplayrig} \kappa_{G}=\frac{\sigma}{2\kappa^{2}}.$$ Using the diagrammatic notation introduced in the preceding section, the full $H[\ell]$ can be expressed as $$H[\ell]=\sigma\Bigg[ {\parbox{1.5cm}{\epsfig{file=diagram1.eps,width=1.5cm}}}-\frac{1}{2}\left({\parbox{1.5cm}{\epsfig{file=diagram3.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram4.eps,width=1.5cm}}}\right) + \cdots \Bigg] .$$ The constrained order-parameter profiles in the gas and liquid phases can be obtained from Eq. (\[greenidentity3\]) as $$\begin{aligned} \delta m_\Xi(\mathbf{r})\equiv m_\Xi(\mathbf{r})\mp m_0 = \mp\frac{1}{2\kappa}\int_{\ell} d\mathbf{s} K(\mathbf{s},\mathbf{r}) q^{\pm}(\mathbf{s}) \nonumber\\ + \frac{m_0}{2\kappa}\int_{\ell} d\mathbf{s} \ \partial_{n} K(\mathbf{s},\mathbf{r}) , \label{profile2}\end{aligned}$$ where the upper (lower) sign must be selected when $\mathbf{r}$ is in the gas (liquid) region. Proceeding in a similar way as in the preceding section, the constrained order-parameter profiles can be expressed in the gas phase as $$\begin{aligned} \label{OP_vapor} \delta m_\Xi&\approx& m_0 \Bigg[{\parbox{1.5cm}{\epsfig{file=diagram21.eps,width=1.5cm}}} -\frac{1}{2}\left({\parbox{1.5cm}{\epsfig{file=diagram23.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram24.eps,width=1.5cm}}}\right)\Bigg] \nonumber\\\end{aligned}$$ and in the liquid phase as $$\begin{aligned} \label{OP_liquid} \delta m_\Xi&\approx& -m_0 \Bigg[{\parbox{1.5cm}{\epsfig{file=diagram28.eps,width=1.5cm}}} -\frac{1}{2}\left({\parbox{1.5cm}{\epsfig{file=diagram29.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram30.eps,width=1.5cm}}}\right) \Bigg].\nonumber\\\end{aligned}$$ General diagrammatic approach ----------------------------- We can go beyond this perturbative approach and reobtain the full set of functionals $\omega_n$. For this purpose, we define $\overline{q}(\mathbf{s})$ to be $[q^+(\mathbf{s})+q^-(\mathbf{s})]/2$. Note that the interfacial Hamiltonian  (\[finterface\]) is proportional to the surface integral of $\overline{q}$. From Eq. (\[consistent\]) it follows that $\overline{q}$ satisfies the integral equation $$\overline{q}(\mathbf{s}_\ell)=-\kappa m_0 - \int_{\ell} d \mathbf{s}\, U(\mathbf{s}_{\ell},\mathbf{s}) \overline{q}(\mathbf{s}).$$ Formally, this equation can be solved iteratively as $$\begin{aligned} \overline{q}(\mathbf{s}_\ell)=-\kappa m_0 \Bigg(1-\int_{\ell} d \mathbf{s}\, U(\mathbf{s}_{\ell},\mathbf{s}) \nonumber\\ +\int_{\ell} d \mathbf{s}\, d \mathbf{s}^\prime\, U(\mathbf{s}_{\ell},\mathbf{s}) U(\mathbf{s},\mathbf{s}^\prime) -\cdots \Bigg),\end{aligned}$$ where the $n^{\textrm{th}}$ term involves the convolution of $n$ $U-$type functions on the interface. Upon substituting this expression into Eq. (\[finterface\]) we are lead to the expansion $$\begin{aligned} &&H[\ell]=\sigma\Bigg(1-\int_{\ell} d \mathbf{s}_1 \, d \mathbf{s}_2 U(\mathbf{s}_1,\mathbf{s}_2) \nonumber\\ &&+ \int_{\ell} d \mathbf{s}_1 \, d \mathbf{s}_2 \, d \mathbf{s}_3\, U(\mathbf{s}_1,\mathbf{s}_2) U(\mathbf{s}_2,\mathbf{s}_3)-\cdots \Bigg)\end{aligned}$$ or, diagrammatically, $$\label{F_chain} H[\ell]=\sigma\Bigg[ {\parbox{1.5cm}{\epsfig{file=diagram1.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram18.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram31.eps,width=1.5cm}}} + \cdots \Bigg] .$$ From this expression, we obtain that $$\omega_n = \int_\ell d \mathbf{s}_1 \, \cdots \, d \mathbf{s}_{n} \, U(\mathbf{s}_1,\mathbf{s}_2) U(\mathbf{s}_2,\mathbf{s}_3)\cdots U(\mathbf{s}_{n-1},\mathbf{s}_n)$$ which is the expression obtained in Ref. [@PR]. In general, from Eq. (\[dashedk\]) we get that $\omega_n \sim R^{-2(n-1)}$, which connects the self-interaction contributions to the curvature corrections to the free energy. We note that this expression is general, so it is also valid for spherical bubbles, for which $\omega_n = \exp[2(n-1)\kappa R]$ [@PR] because $H^2=K_G$. We can also obtain the order-parameter profiles by considering in Eq. (\[deltamdiagram\]) the limit $h_1=0$, $g\to -\infty$, and $m_b=-m_0$ (or $+m_0$) for the gas (or liquid) phase, respectively. Thus, the order-parameter profile in the gas phase has the expansion $$\begin{aligned} \delta m_\Xi = m_0 \Bigg[{\parbox{1.5cm}{\epsfig{file=diagram21.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram32.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram33.eps,width=1.5cm}}}-\cdots\Bigg], \label{deltamdiagram2}\end{aligned}$$ and in the liquid phase has the expansion $$\begin{aligned} \delta m_\Xi = -m_0 \Bigg[{\parbox{1.5cm}{\epsfig{file=diagram28.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram34.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram35.eps,width=1.5cm}}}-\cdots \Bigg]. \label{deltamdiagram3}\end{aligned}$$ Again, these equations reduce to Eqs. (\[OP\_vapor\]) and (\[OP\_liquid\]) upon using Eq. (\[dashedk\]) up to $R^{-3}$ corrections. Once we have established the connection between the self-interaction of the fluid interface and the curvature corrections to the interfacial free energy, we can find the full functional of the interfacial shape. This task can be pursued to any desired accuracy in the curvature, which we leave at ${O}(R^{-3})$. Leaving the technical aspects aside here (see Appendix \[appendix\_B\] for details) we find that Eq. (\[Helfrich\]) reduces to the interfacial Hamiltonian $$\label{ham4} H[\ell] \approx \sigma \mathcal{A}_{\pi} + \frac{\sigma}{2} \int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{W}(x_{12}) \bigl[\ell(\mathbf{x}_{2}) - \ell(\mathbf{x}_{1})\bigr]^{2} ,$$ where the self-interaction is described by the function $$\mathcal{W}(x) \equiv \frac{\kappa}{2\pi} \frac{1+\kappa x}{x^{3}} \textrm{e}^{-\kappa x} =\frac{1+\kappa x}{x^2}\mathcal{K}(x) ,$$ thus rigorously rederiving the result first obtained in Ref. [@PR]. Clearly, for a flat interface $H[\ell]=\sigma \mathcal{A}_\pi$, with $\mathcal{A}_\pi$ being the planar (projected) area. As shown in Ref. [@PR], when the gradient is small we can expand as $\ell(\mathbf{x}_{2}) \simeq \ell(\mathbf{x}_{1}) + \mathbf{x}_{21} \mathbf{\cdot}\boldsymbol \nabla\ell(\mathbf{x}_{1})+\cdots$, in which case Eq. (\[ham4\]) reduces to $$H[\ell] \approx \sigma \mathcal{A}_{\pi}+\frac{\sigma}{2} \int d \mathbf{x} \, \bigl[\boldsymbol\nabla\ell(\mathbf{x})\bigr]^{2},$$ thus recovering the standard mesoscopic capillary-wave Hamiltonian, which can now be seen as a particular local small-gradient limit of the nonlocal functional (\[ham4\]). The present derivation of the nonlocal self-interaction improves on that presented in Ref. [@PR] inasmuch as it systematically and rigorously accounts for all curvature corrections. Summary and remarks ------------------- In this section we considered an isolated liquid-gas interface and solved the Helmholtz equations required to identify the free energy of a constrained interfacial configuration defined by a crossing criterion. We first implemented a direct perturbation expansion in the local curvature, obtaining the Helfrich-like corrections to the surface tension term and identifying the values of bending and saddle-splay rigidities for the DP potential. We then refined this expansion by considering the order-parameter profile around the interface in which the curvature corrections are explicit \[see Eqs. (\[OP\_vapor\]) and (\[OP\_liquid\])\]. This leads us naturally to express the free energy as an interfacial self-interaction that can be neatly expressed that involves a formally exact way using a diagrammatic expansion. Finally, in the limit of small curvatures this nonlocal interaction recovers the standard local capillary-wave model. Binding energy for a wetting film configuration =============================================== Having examined the wall-liquid and free (but constrained) liquid-gas interfaces, we turn to the case of a wall-gas interface where an intruding wetting layer of liquid, with positive order parameter (i.e., $m>0$), intrudes between the substrate and the bulk gas (where the order parameter is set to $-m_0$). The wall is again described by a height function $\psi(\mathbf{x})$ and the liquid-gas interface (i.e., the surface on which $m=0$) is constrained to lie along $\ell(\mathbf{x})$. No overhangs of either the interface or substrate occur; nor do these two surfaces touch. The minimum of the LGW Hamiltonian (\[HLGW\]) subject to the substrate, bulk, and crossing-criterion boundary conditions $$\begin{aligned} \mathbf{n}_\psi\mathbf{\cdot}\boldsymbol \nabla m_\Xi (\mathbf{r})=-h_1-gm_\Xi(\mathbf{r}), \ \textrm{for}\ \mathbf{r}=(\mathbf{x},\psi(\mathbf{x})), \label{boundary22}\\ m_\Xi(z)\to -m_0, \qquad \textrm{for}\quad z\to +\infty, \label{boundary2b2}\\ m_\Xi(\mathbf{r})=0,\ \textrm{for}\ \mathbf{r}=(\mathbf{x},\ell(\mathbf{x})) \label{boundary2c2}\end{aligned}$$ defines a constrained excess free energy for the wall-gas interface, which by Eq. (\[mfHLGW2\]) can be recast as $$\begin{aligned} \mathcal{F}_{wg}[\ell,\psi]=-\frac{m_0}{2}\int_{\ell} d\mathbf{s} \left[q^+_\ell(\mathbf{s}) + q^-_\ell (\mathbf{s})\right]\nonumber\\ +\frac{h_1+gm_0}{2g}\int_\psi d\mathbf{s} q_\psi(\mathbf{s}), \label{Hmin}\end{aligned}$$ where $q_\ell^{\pm}(\mathbf{s})\equiv \mathbf{n}_\ell(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_\mathbf{s} \delta m^{\pm}(\mathbf{s})$ for $\mathbf{s}$ on the liquid-gas interface, and $q_\psi\equiv \mathbf{n}_\psi(\mathbf{s})\mathbf{\cdot}\boldsymbol \nabla_\mathbf{s} \delta m(\mathbf{s})$ on the substrate. The next step is to define and identify the binding potential $W[\ell,\psi]$. By analogy with isolated interfaces, the free energy of a wetting layer can be expressed as a functional of the normal derivatives of the order parameter computed at the layer boundaries. The binding potential takes into account the interaction of the interface with the wall and is determined by subtracting the contributions arising from the isolated wall-liquid and constrained but free liquid-gas interfaces, which we have already determined. Therefore, before presenting the final result for $W[\ell,\psi]$ and its diagrammatic formulation, we need to consider the fundamental relations obeyed by the order parameter in a wetting layer. In order to do so we need some technical preliminaries. Technical preliminaries ----------------------- From Eqs. (\[greenidentity5\]) and (\[greenidentity6\]) it follows that the functions $q^+_\ell(\mathbf{s})$, $q^-_\ell(\mathbf{s})$, and $q_\psi(\mathbf{s})$ satisfy the coupled integral equations $$\begin{aligned} \int_\psi d\mathbf{s} \Bigg[K(\mathbf{s}_\psi,\mathbf{s}) +\frac{1}{g}\partial_{n} K(\mathbf{s}_\psi,\mathbf{s}) -\frac{\kappa}{g}\delta(\mathbf{s}-\mathbf{s}_\psi)\Bigg]q_\psi(\mathbf{s}) \nonumber\\-\int_\ell d\mathbf{s} K(\mathbf{s}_\psi,\mathbf{s})q_\ell^-(\mathbf{s}) \nonumber\\=\left(\frac{-h_1}{g}-m_0\right)\left(-\kappa+\int_\psi d\mathbf{s}\ \partial_n K(\mathbf{s}_\psi,\mathbf{s})\right) \nonumber\\ +m_0\int_\ell d\mathbf{s}\ \partial_n K(\mathbf{s}_\psi,\mathbf{s}); \label{inteq1a}\\ \int_\psi d\mathbf{s} \Bigg[K(\mathbf{s}_\ell,\mathbf{s}) +\frac{1}{g}\partial_n K(\mathbf{s}_\ell,\mathbf{s}) \Bigg]q_\psi (\mathbf{s})\nonumber\\ -\int_\ell d\mathbf{s} K(\mathbf{s}_\ell,\mathbf{s}) q^-_\ell (\mathbf{s}) =m_0\left(\kappa+\int_\ell d\mathbf{s}\ \partial_n K(\mathbf{s}_\ell,\mathbf{s})\right) \nonumber\\ +\left(\frac{-h_1}{g}-m_0\right)\int_\psi d\mathbf{s}\ \partial_n K(\mathbf{s}_\ell,\mathbf{s}) \label{inteq1b}\end{aligned}$$ and $$\begin{aligned} \int_\ell d\mathbf{s} K(\mathbf{s}_\ell,\mathbf{s})q_\ell^+ (\mathbf{s})=-m_0\left(\kappa- \int_\ell d\mathbf{s}\ \partial_n K(\mathbf{s}_\ell,\mathbf{s})\right), \label{inteq2}\end{aligned}$$ where $\mathbf{s}_\ell$ and $\mathbf{s}_\psi$ are on the liquid-gas interface and on the substrate, respectively. Note that these equations are linear in $q$. In order to extract the interaction between the surfaces, we obtain the equations in terms of the new fields $\Delta q^\pm_\ell(\mathbf{s})\equiv q^\pm_\ell(\mathbf{s})-q^{0,\pm}_\ell(\mathbf{s})$, and $ \Delta q_\psi(\mathbf{s})\equiv q_\psi(\mathbf{s})-q^{0}_\psi(\mathbf{s})$, where the $0$ superscript means that the corresponding normal derivative is evaluated on its isolated interface. In addition, $q^0_\psi$ and $q^{0,\pm}_\ell$ satisfy Eqs. (\[inteq1\]) and (\[consistent\]), respectively. Note that $\Delta q^{0,+}_\ell\equiv 0$ (because the gas domain is shielded from influence of the wall) and Eqs. (\[inteq1a\]) and (\[inteq1b\]) can be recast as $$\begin{aligned} & &\int_\psi d\mathbf{s} \Bigg[K(\mathbf{s}_\psi,\mathbf{s}) +\frac{1}{g}\partial_n K(\mathbf{s}_\psi,\mathbf{s}) \nonumber\\&-&\frac{\kappa}{g}\delta(\mathbf{s}_\psi-\mathbf{s}) \Bigg]\Delta q_\psi(\mathbf{s})-\int_\ell d\mathbf{s} K(\mathbf{s}_\psi,\mathbf{s})\Delta q_{\ell}^-(\mathbf{s})\nonumber\\ &=&m_0\int_\ell d\mathbf{s} \ \partial_n K(\mathbf{s}_\psi,\mathbf{s}) +\int_\ell d\mathbf{s} K(\mathbf{s}_\psi,\mathbf{s}) q_\ell^{0,-} (\mathbf{s}) \nonumber\\ &\equiv& 2\kappa \delta m_\Xi^{0,\ell}(\mathbf{s}_\psi) \label{inteq6}\end{aligned}$$ and $$\begin{aligned} & &\int_\psi d\mathbf{s} \Bigg[K(\mathbf{s}_\ell,\mathbf{s}) +\frac{1}{g}\partial_n K(\mathbf{s}_\ell,\mathbf{s})\Bigg] \Delta q_\psi (\mathbf{s})\nonumber\\ &-&\int_\ell d\mathbf{s} K(\mathbf{s}_\ell,\mathbf{s}) \Delta q_\ell(\mathbf{s}) \nonumber\\ &=&\left(\frac{-h_1}{g}-m_0\right)\int_\psi d\mathbf{s} \ \partial_n K(\mathbf{s}_\ell,\mathbf{s})\nonumber\\ &-&\int_\psi d\mathbf{s} \Bigg[K(\mathbf{s}_\ell,\mathbf{s}) +\frac{1}{g}\partial_n K(\mathbf{s}_\ell,\mathbf{s})\Bigg] q^0_\psi(\mathbf{s})\nonumber\\ &\equiv& 2\kappa \delta m_\Xi^{0,\psi}(\mathbf{s}_\ell), \label{inteq7}\end{aligned}$$ where we have identified the right-hand side of both equations as the order-parameter profile $\delta m_\Xi^{0,\ell(\psi)}(\mathbf{s})$ at the boundary point $\mathbf{s}$ due to the presence of an isolated liquid-gas (wall) interface, respectively \[see Eqs. (\[profile\]) and (\[profile2\])\]. Equations (\[inteq6\]) and (\[inteq7\]) are the basis of our perturbative approach, as we can expand $\Delta q_\psi$ and $\Delta q_\ell^-$ in powers of $K(\mathbf{s}_\ell,\mathbf{s}_\psi)$ as $\Delta q_\psi =\sum_{i=1}^\infty \Delta q_{i,\psi}$ and $\Delta q_\ell =\sum_{i=1}^\infty \Delta q_{i,\ell}$. Each term of this expansion can be formally solved as follows. At the wall, $$\begin{aligned} &&\Delta q_{1,\psi}(\mathbf{s}_\psi)=\int_\psi d\mathbf{s} X_\psi(\mathbf{s}_\psi,\mathbf{s}) 2\kappa \delta m_\Xi^{0,\ell} (\mathbf{s}), \label{formalsol}\end{aligned}$$ and otherwise $$\begin{aligned} &&\Delta q_{i>1,\psi}(\mathbf{s}_\psi)=\int_\psi d\mathbf{s} \int_\ell d\mathbf{s}' X_\psi(\mathbf{s}_\psi,\mathbf{s}) K(\mathbf{s},\mathbf{s}')\Delta q_{i-1,\ell}(\mathbf{s}') . \nonumber\\ \label{formalsol2}\end{aligned}$$ Similarly, at the interface $$\begin{aligned} &&\Delta q_{1,\ell}(\mathbf{s}_\ell)=-\int_\ell d\mathbf{s} K^{-1}_\ell(\mathbf{s}_\ell,\mathbf{s}) 2\kappa \delta m_\Xi^{0,\psi} (\mathbf{s}) \label{formalsol3}\end{aligned}$$ and otherwise $$\begin{aligned} \Delta q_{i>1,\ell}(\mathbf{s}_\ell)=\int_\ell d\mathbf{s} \int_\psi d\mathbf{s}' K^{-1}_\ell(\mathbf{s}_\ell,\mathbf{s}) \Bigg(K(\mathbf{s},\mathbf{s}')\nonumber\\ +\frac{1}{g}\partial_{n'}K(\mathbf{s}, \mathbf{s}')\Bigg)\Delta q_{i-1,\psi}(\mathbf{s}'), \label{formalsol4}\end{aligned}$$ where $X_\psi$ is the operator on the substrate defined by Eq. (\[seriesx\]) and $K^{-1}_\ell$ is the inverse operator of $K$ on the liquid-gas interface. Now using the Green’s identities (\[greenidentity7\])-(\[greenidentity9\]), it follows that $$\begin{aligned} &&\Delta q_{i>1,\ell}(\mathbf{s}_\ell)= \int_\ell d\mathbf{s} \int_\psi d\mathbf{s}' \int_\psi d\mathbf{s}'' K^{-1}_\ell(\mathbf{s}_\ell,\mathbf{s})\nonumber\\ &&\times K(\mathbf{s},\mathbf{s}')\left(\delta(\mathbf{s}'-\mathbf{s}'')+\frac{\kappa}{g}K^{-1}_\psi(\mathbf{s}',\mathbf{s}'')\right) \Delta q_{i-1,\psi}(\mathbf{s}'') \nonumber\\ &&+\frac{\kappa}{g}\int_\ell d\mathbf{s} \int_\psi d\mathbf{s}' \int_\psi d\mathbf{s}'' \int_\psi d\mathbf{s}''' K^{-1}_\ell(\mathbf{s}_\ell,\mathbf{s})K(\mathbf{s},\mathbf{s}')\nonumber\\ &&\times\frac{1}{\kappa}\partial_{n'} K(\mathbf{s}',\mathbf{s}'') K^{-1}_\psi(\mathbf{s}'',\mathbf{s}''') \Delta q_{i-1,\psi}(\mathbf{s}'''), \label{formalsol5}\end{aligned}$$ where $K^{-1}_\psi$ is the inverse operator of $K$ on the substrate, i.e., $X_\psi$ in the limit $g\to -\infty$. Taking into account the expansions  (\[seriesx2\]), (\[definvk\]), (\[deltamdiagram\]), and (\[deltamdiagram3\]), we obtain a diagrammatic expansion for $\Delta q_\ell$ and $\Delta q_\psi$, $$\begin{aligned} &&\Delta q_\psi(\mathbf{s})=-2\kappa m_0\Bigg[\frac{g}{g-\kappa}\Bigg({\parbox{1.5cm}{\epsfig{file=diagram36.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram37.eps,width=1.5cm}}}+ \cdots\nonumber\\ &&+\frac{\kappa}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram38.eps,width=1.5cm}}}+\cdots-\frac{\kappa}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram39.eps,width=1.5cm}}}\Bigg)\nonumber\\ &&+\left(\frac{g}{g-\kappa}\right)^2\left(1+\frac{\kappa}{g}\right){\parbox{1.5cm}{\epsfig{file=diagram40.eps,width=1.5cm}}}+\cdots\Bigg] \nonumber\\ &&+2\kappa \frac{h_1+gm_0}{g} \Bigg[\left(\frac{g}{g-\kappa}\right)^2\Bigg({\parbox{1.5cm}{\epsfig{file=diagram41.eps,width=1.5cm}}}\nonumber\\ &&+\frac{g}{\kappa-g} {\parbox{1.5cm}{\epsfig{file=diagram42.eps,width=1.5cm}}}+\cdots\Bigg)+\cdots\Bigg] \label{deltaqpsi}\end{aligned}$$ and $$\begin{aligned} &&\Delta q_\ell^-(\mathbf{s})= 2\kappa \frac{h_1+gm_0}{g}\Bigg[\frac{g}{g-\kappa}\Bigg({\parbox{1.5cm}{\epsfig{file=diagram43.eps,width=1.5cm}}}\nonumber\\ &&-{\parbox{1.5cm}{\epsfig{file=diagram44.eps,width=1.5cm}}} +\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram45.eps,width=1.5cm}}}+\cdots\Bigg)\Bigg]\nonumber\\ &&-2\kappa m_0 \Bigg[\frac{g+\kappa}{g-\kappa}\ \Bigg({\parbox{1.5cm}{\epsfig{file=diagram46.eps,width=1.5cm}}} +\cdots\Bigg)+\cdots\Bigg], \label{deltaql}\end{aligned}$$ where the symbols have the meanings as described above. The diagrams in this expansion have segments on alternating interfaces connected via $K$ kernels, so they can be regarded as decorated versions of the zigzag diagrams of the original nonlocal model. The segments on the substrate correspond to convolution products of $U-$ and $\partial_n K/\kappa-$type bonds on this surface, while on the liquid-gas interface only $U-$type bonds are involved. The closed extreme, which by convention we place on the left, provides the factor $-2\kappa m_0$ or $-2\kappa (h_1+gm_0)/g$, depending on whether it is located on the liquid-gas interface or on the substrate, respectively. On the other hand, the interface on which the open extreme resides indicates whether the diagram contributes to $\Delta q_\psi$ (if it is on the substrate) or $\Delta q_\ell^-$ (otherwise). The factor the multiplies each diagram can be obtained as the product of terms associated with each segment. The segments on the liquid-gas interface have a factor $(-1)^n$, where $n$ is the number of bonds (of $U$ type) in the segment. The contribution of the segments on the substrate depend on their positions. Let $n$ be the total number of bonds in the segment, and $m$ be the number of $\partial_n K/\kappa$ bonds. If the segment contains the closed extreme, its contribution is given by $$\begin{aligned} -\left(\frac{g}{\kappa-g}\right)^{n+1}\left(\frac{\kappa}{g}\right)^{m-m_0}\left[1-(1-\delta_{m,0}) \left(1-\frac{\kappa}{g}\right)^{l+1}\right], \label{factor2}\end{aligned}$$ where $\delta_{i,j}$ is the Kronecker symbol, the index $m_0$ is either $0$ or $1$ (depending on the first bond being of $U$ type or $\partial_n K/\kappa$ type, respectively), and $l$ is the number of $U-$type bonds after the last $\partial K/\kappa-$type bond. Note that this expression is the factor that multiplies the diagrams in the expansion (\[deltamdiagram\]) for the order-parameter profile above the substrate, multiplied by $g/(g-\kappa)$. The contribution of a segment on the substrate that contains the open extreme is $$-\left(\frac{g}{\kappa-g}\right)^{n+1}\left(\frac{\kappa}{g}\right)^{m}. \label{factor3}$$ Finally, any other segment on the substrate will provide the following factor: either $$-\left(\frac{g}{\kappa-g}\right)^{n+1}\left(\frac{\kappa}{g}\right)^{m}\left[2- \left(1-\frac{\kappa}{g}\right)^{l+1}\right], \label{factor4}$$ if the last bond is of $U-$type or $$\begin{aligned} -\left(\frac{g}{\kappa-g}\right)^{n+1}\left(\frac{\kappa}{g}\right)^{m}\Bigg[2-\frac{g}{\kappa}+\frac{\kappa}{g} \nonumber\\ -\left(1-\frac{g}{\kappa}\right) \left(1-\frac{\kappa}{g}\right)^{l+1}\Bigg], \nonumber\\ \label{factor5}\end{aligned}$$ with $l$ being the number of consecutive $U-$type bonds in the rightmost sequence in the segment. Binding potential functional and order parameter ------------------------------------------------ With these preliminaries behind us, we are now in a position to obtain the diagrammatic representation of the binding potential and the order-parameter profile. The binding potential functional $W[\ell,\psi]$ is defined as the substrate-interface interaction in the excess free energy: $$\mathcal{F}_{wg}=\mathcal{F}_{wl}[\psi]+H[\ell]+W[\ell,\psi], \label{bindingpotential}$$ where $\mathcal{F}_{wl}[\psi]$ is the free energy of the wall-liquid interface and $H[\ell]$ is the free liquid-gas interfacial Hamiltonian, which we have already determined. So, in terms of $\Delta q_\ell^-$ and $\Delta q_\psi$, we have $$\begin{aligned} W[\ell,\psi]=-\frac{m_0}{2}\int_{\ell} d\mathbf{s} \Delta q_\ell^{-}(\mathbf{s}) +\frac{h_1+gm_0}{2g}\int_\psi d\mathbf{s}\ \Delta q_\psi(\mathbf{s}). \nonumber\\ \end{aligned}$$ Substituting the expansions (\[deltaqpsi\]) and (\[deltaql\]) into this expression, we arrive at the diagrammatic expansion for the binding potential functional $$\begin{aligned} W[\ell,\psi]=\sum_{n=1}^\infty \Bigg(-\kappa m_0\frac{h_1+gm_0}{g} \Omega_n^n+ \kappa m_0^2\Omega_n^{n+1} \nonumber\\ +\kappa \left(\frac{h_1+gm_0}{g}\right)^2 \Omega_{n+1}^n\Bigg), \label{Hmin6}\end{aligned}$$ where $\Omega_i^j$ is the sum of all the independent diagrams that have $i$ segments on the substrate and $j$ segments on the liquid-gas interface. Note that these diagrams correspond to those obtained previously for $\Delta q_\psi$ and $\Delta q_\ell^-$, but integrating over the positions of $\mathbf{s}$, i.e., with a closed right extreme. For the first terms we have $$\begin{aligned} &&\Omega_1^1= \frac{g}{g-\kappa}\Bigg(2{\parbox{1.5cm}{\epsfig{file=diagram47.eps,width=1.5cm}}}-2{\parbox{1.5cm}{\epsfig{file=diagram48.eps,width=1.5cm}}} \nonumber\\ &&+\frac{\kappa}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram49.eps,width=1.5cm}}}+\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram50.eps,width=1.5cm}}}\nonumber\\ &&-\frac{\kappa}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram51.eps,width=1.5cm}}}+\cdots\Bigg)\label{omega11}\end{aligned}$$ while $$\begin{aligned} &&\Omega_1^2= \frac{g}{g-\kappa}\Bigg(\frac{\kappa+g}{g}{\parbox{1.5cm}{\epsfig{file=diagram52.eps,width=1.5cm}}}\nonumber\\ &&-2\frac{\kappa+g}{g}{\parbox{1.5cm}{\epsfig{file=diagram53.eps,width=1.5cm}}}+\frac{\kappa+g}{g}{\parbox{1.5cm}{\epsfig{file=diagram54.eps,width=1.5cm}}}\nonumber\\ &&+\frac{2\kappa}{g}{\parbox{1.5cm}{\epsfig{file=diagram55.eps,width=1.5cm}}}+\cdots\Bigg) \label{omega12}\end{aligned}$$ and $$\begin{aligned} &&\Omega_2^1= \left(\frac{g}{g-\kappa}\right)^2\Bigg({\parbox{1.5cm}{\epsfig{file=diagram56.eps,width=1.5cm}}}\nonumber\\ &&+\frac{g}{\kappa-g}{\parbox{1.5cm}{\epsfig{file=diagram57.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram58.eps,width=1.5cm}}}+\cdots\Bigg). \label{omega21}\end{aligned}$$ These diagrams are all decorated versions of the diagrams in the original nonlocal model. They are multiplied by a factor that is the same as the corresponding coefficient for the associated $\Delta q$ diagram with open right extreme, provided the diagram either contains $\partial_n K/\kappa-$type bonds or is symmetric under a mirror reflection, i.e., it is the same when read from left to right or the reverse. Otherwise, the factor is twice the coefficient for the associated $\Delta q$ diagram. The reason for this is that two different $\Delta q$ diagrams lead to the same contribution to $W[\ell,\psi]$. In this sense we mean that only independent diagrams are taken into account in the diagrammatic expansion of $W[\ell,\psi]$, because we discard one of the two equivalent diagrams, which are related via a mirror reflection. Now we turn to the order-parameter profile. Above the liquid-gas interface the profile is uninfluenced by the presence of the substrate, so it has a diagrammatic expansion given by Eq. (\[deltamdiagram2\]). On the other hand, the order-parameter profile within the adsorbed liquid layer is influenced by the proximity of both the wall and the liquid-gas interface and has the representation $$\begin{aligned} \delta m_\Xi(\mathbf{r})&=& -\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} K(\mathbf{s},\mathbf{r}) q_\psi(\mathbf{s}) \nonumber\\ &-&\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} \left(\frac{h_1}{g}+m_b+\frac{q_\psi(\mathbf{s})}{g}\right) \partial_{n} K(\mathbf{s},\mathbf{r}) \nonumber\\ &+&\frac{1}{2\kappa}\int_\ell d\mathbf{s} K(\mathbf{s},\mathbf{r}) q_\ell^-(\mathbf{s}) \nonumber\\ &+&\frac{m_0}{2\kappa}\int_\ell d\mathbf{s}\partial_{n} K(\mathbf{s},\mathbf{r}). \label{profile3}\end{aligned}$$ Now, writing $q_\psi=q_\psi^0+\Delta q_\psi$ and $q_\ell^-=q_\ell^{0,-}+\Delta q_\ell^-$, and making use of Eqs. (\[profile\]) and (\[profile2\]), we obtain the expression for the order parameter in the liquid layer $$\begin{aligned} \delta m_\Xi(\mathbf{r})&=& \delta m_\Xi^{0,\psi}(\mathbf{r})+\delta m_\Xi^{0,\ell}(\mathbf{r}) \nonumber\\ &+&\frac{1}{2\kappa}\int_\ell d\mathbf{s} K(\mathbf{s},\mathbf{r}) \Delta q_\ell^-(\mathbf{s}) \label{profile4} \\ &-&\frac{1}{2\kappa}\int_{\psi} d\mathbf{s} \left(K(\mathbf{s},\mathbf{r})+\frac{1}{g}\partial_{n} K(\mathbf{s},\mathbf{r})\right) \Delta q_\psi(\mathbf{s}) , \nonumber\end{aligned}$$ where the kernel connecting the substrate to the position $\mathbf{r}$ in the last term can be related to $K$ using Eq. (\[greenidentity8\]). Taking into account the expansions (\[deltaqpsi\]) and (\[deltaql\]), we find that the order-parameter profile in the liquid layer has the expansion $$\begin{aligned} &&\delta m_\Xi=-m_0\Bigg[{\parbox{1.5cm}{\epsfig{file=diagram59.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram60.eps,width=1.5cm}}}+\cdots\nonumber\\ &&-\frac{\kappa+g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram61.eps,width=1.5cm}}}+\frac{\kappa+g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram62.eps,width=1.5cm}}}+\cdots\nonumber\\ &&+\frac{\kappa+g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram63.eps,width=1.5cm}}}+\cdots\Bigg]\nonumber\\ &&-\frac{h_1+gm_0}{g}\Bigg[\frac{g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram64.eps,width=1.5cm}}}\nonumber\\ &&-\left(\frac{g}{\kappa-g}\right)^2{\parbox{1.5cm}{\epsfig{file=diagram65.eps,width=1.5cm}}} +\cdots\nonumber\\ &&-\frac{g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram66.eps,width=1.5cm}}}+\frac{g}{g-\kappa}{\parbox{1.5cm}{\epsfig{file=diagram67.eps,width=1.5cm}}}+\cdots \nonumber\\ &&+\left(\frac{g}{g-\kappa}\right)\left(\frac{\kappa+g}{g-\kappa}\right)^2{\parbox{1.5cm}{\epsfig{file=diagram68.eps,width=1.5cm}}}+\cdots\Bigg]. \label{deltamdiagram4}\end{aligned}$$ Note that, once again, these diagrams are decorated versions of those obtained in the original nonlocal model. Their prefactors are either $-m_0$ (if the left extreme is on the liquid-gas interface) or $-h_1/g-m_0$ (if it is on the substrate). The coefficient that multiplies each diagram is the product of $(-1)^{k}$, with $k$ being the number of $K-$type bonds that connect both interfaces, and the factors associated with the segments on each substrate. Sections with $n$ of the $U-$type bonds on the liquid-gas interface contribute with a factor $(-1)^n$. A segment on the substrate has a factor given by Eq. (\[factor2\]) if it contains the left diagram extreme and otherwise by Eq. (\[factor4\]) or (\[factor5\]), depending on the nature of the rightmost bond in the segment. Resummation of wetting diagrams ------------------------------- As we pointed out in the preceding sections, the curvature expansion for isolated interfaces is actually connected to the formal diagrammatic method we have developed. This connection also persists for a wetting film configuration, but it is not at all explicit. The aim of this section is to illustrate how the perturbative scheme emerges from the diagrammatic one. However, the connection is actually far from trivial. The reason is that, although the $\partial K/\kappa-$type bonds are of order $R^{-1}$ \[see Eq. (\[expk2\])\], this is not the case for the $U-$type bonds: Its integral with respect to one argument is of order of $R^{-2}$, but $U$ is of order of unity for $\kappa|\mathbf{s}-\mathbf{s}^\prime|\sim 1$. So, if we convolute $U$ with a function that varies on a length scale much larger than $\kappa^{-1}$, this is not a problem. However, in the case considered in the present section, we usually convolute $U$ with a kernel $K$ connecting both interfaces, which varies on the same length scale (i.e., $\kappa^{-1}$) as $U$. However, we will see that it is possible to resum the diagrams to obtain a diagrammatic representation of the zeroth order (in curvature) corrections. By Eq. (\[expk\]), $K(\mathbf{s},\mathbf{s}^\prime)-\mathcal{K}(r_\bot)$ is of order $R^{-2}$, where $\mathbf{r}_\bot$ is the projection of $\mathbf{s}-\mathbf{s}^\prime$ on the plane tangent to the interface at $\mathbf{s}^\prime$. Thus, we can neglect diagrams that present $\partial_n K/\kappa-$type bonds, and we replace $K(\mathbf{s},\mathbf{s}^\prime)$ by $\mathcal{K}(r_\bot)$ in the $U-$type bonds. First, we consider the convolution of $K(\mathbf{r},\mathbf{s})$ with $K(\mathbf{s},\mathbf{s}^\prime)$, where $\mathbf{s}$ and $\mathbf{s}^{\prime}$ are on the same interface and $\mathbf{r}$ is either above or below this interface. We place the origin at $\mathbf{s}^{\prime}$, neglect curvature corrections, and finally assume that $\kappa r \gg 1$. Then we have $$\begin{aligned} \int d\mathbf{s} K(\mathbf{s},\mathbf{0})K(\mathbf{s},\mathbf{r}) \approx \left(\frac{\kappa}{2\pi}\right)^2 \int_0^\infty ds \textrm{e}^{-\kappa s} \\\times \int_0^{2\pi} d\theta \frac{\exp\left[{-\kappa\sqrt{r^2+s^2-2sr\sin\alpha \cos\theta}}\right]}{\sqrt{r^2+s^2-2sr\sin\alpha \cos\theta}} , \nonumber\end{aligned}$$ where $\alpha$ is the angle between $\mathbf{r}$ and the surface normal at the origin $\mathbf{n}(\mathbf{0})$. As $\kappa r\gg 1$ but $\kappa s\lesssim 1$, we expand the distance between $\mathbf{r}$ and $\mathbf{s}$ in powers of $s/r$, so in the first approximation we have $$\begin{aligned} &&\int d\mathbf{s} K(\mathbf{s},\mathbf{0})K(\mathbf{s},\mathbf{r}) \approx \\ &&\frac{\kappa \textrm{e}^{-\kappa r}} {2\pi r} \int_0^\infty \kappa ds \textrm{e}^{-\kappa s}\frac{1}{2\pi} \int_0^{2\pi}d\theta \exp\left[\kappa s \sin\alpha \cos \theta\right]. \nonumber\end{aligned}$$ The modified Bessel function of zeroth order and the first kind $I_0$ has the integral representation $$\label{i0int} I_0(x)=\frac{1}{2\pi} \int_0^{2\pi}d\theta \exp\left[x\cos \theta\right].$$ Therefore $$\begin{aligned} \int d\mathbf{s} K(\mathbf{s},\mathbf{0})K(\mathbf{s},\mathbf{r}) &\approx& \frac{\kappa \textrm{e}^{-\kappa r}} {2\pi r} \int_0^\infty \kappa ds \textrm{e}^{-\kappa s}I_0(\kappa s \sin\alpha) \nonumber\\ &=& \frac{\kappa \textrm{e}^{-\kappa r}}{2\pi r}\frac{1}{|\cos\alpha|},\end{aligned}$$ and thus $$\int d\mathbf{s}_0 U(\mathbf{s},\mathbf{s}_0)K(\mathbf{s}_0,\mathbf{r}) \approx K(\mathbf{s},\mathbf{r}) \left(\frac{|\mathbf{r}-\mathbf{s}|}{|\mathbf{n}(\mathbf{s})\mathbf{\cdot}(\mathbf{r}-\mathbf{s})|}-1\right), \label{intuk}$$ up to corrections in powers of $(\kappa R)^{-1}$ and $(\kappa r)^{-1}$. Equation (\[intuk\]) vanishes if $\mathbf{r}$ is on the normal direction to the substrate at $\mathbf{s}$. As a consequence, when the liquid-gas interface is parallel to the substrate (e.g., for parallel planes or concentric spheres or cylinders), the nonlocal model ansatz is a good approximation to the full solution when curvature corrections are neglected [@laura; @parry2] (see also Appendix \[appendixe\]). On the other hand, a saddle-point analysis of the binding potential shows that the maximum contribution to the multiple integrals associated with each diagram in Eq. (\[Hmin6\]) arises from the neighborhood of the closest pair of points located on different interfaces, which would lie on a normal direction common to both substrates. In this sense, the binding-potential representation shown above is extremely nonlocal: in leading order only the shape of the substrate and the liquid-gas interfaces around their closest positions features. However, this is not true for the order-parameter profile at an arbitrary position $\mathbf{r}$ and, in general, corrections beyond $\delta m_\Xi^{0,\ell}+\delta m_\Xi^{0,\psi}$ will be incorrect with the original nonlocal model ansatz, even neglecting curvature corrections. In order to obtain a more local representation of the binding potential and the order-parameter profile, we note that the structure of the diagrams shows $K$ bonds that connect both interfaces, followed by a segment of the diagram on one interface. The idea is to resum the convolutions of a $K$ bond connecting the wall and the gas-liquid interface with all the possible segments either on the liquid-gas interface or on the substrate. If this is done to zeroth order in the curvature, we obtain renormalized bonds between the wall and the gas-liquid interface. For example, a renormalized bond between a (left) position on the substrate and a (right) position on the gas-liquid interface would be $$\begin{aligned} K_{\psi\to \ell} (\mathbf{s},\mathbf{r})=K(\mathbf{s},\mathbf{r})+\sum_{i=1}^\infty (-1)^i\nonumber\\\times \int_\ell d\mathbf{s}_1 \cdots d\mathbf{s}_{i} U(\mathbf{s},\mathbf{s}_1)\cdots U(\mathbf{s}_{i-1},\mathbf{s}_i) K(\mathbf{s}_i,\mathbf{r}),\label{renormbond1-prevprev}\end{aligned}$$ which is approximately given by $$\begin{aligned} &&K_{\psi\to \ell} (\mathbf{s},\mathbf{r})\approx K(\mathbf{s},\mathbf{r})\sum_{i=0}^\infty \left(1-\frac{1}{|\cos\alpha|}\right)^i\nonumber\\ &&=K(\mathbf{s},\mathbf{r})|\cos\alpha| \label{renormbond1-prev}\end{aligned}$$ and hence $$K_{\psi\to \ell} (\mathbf{s},\mathbf{r}) \approx -\frac{1}{\kappa} \partial_{n}K(\mathbf{s},\mathbf{r}), \label{renormbond1}$$ where $\partial_n K(\mathbf{s},\mathbf{r})=\mathbf{n}(\mathbf{s})\mathbf{\cdot} \boldsymbol\nabla_{\mathbf{s}} K(\mathbf{s},\mathbf{r})$. On the other hand, a renormalized bond between a (left) position on the gas-liquid interface and a (right) position on the substrate would be given by $$\begin{aligned} &&K_{\ell\to \psi} (\mathbf{s},\mathbf{r})=\frac{g+\kappa}{g-\kappa}K(\mathbf{s},\mathbf{r})\nonumber\\ &&-\sum_{i=1}^\infty \Bigg[2\left(\frac{g}{\kappa-g}\right)^{i+1}+(-1)^i\Bigg]\nonumber\\ &&\times \int_\psi d\mathbf{s}_1 \cdots d\mathbf{s}_{i} U(\mathbf{s},\mathbf{s}_1)\cdots U(\mathbf{s}_{i-1},\mathbf{s}_i) K(\mathbf{s}_i,\mathbf{r}),\label{renormbond2-prevprev}\end{aligned}$$ which is approximately $$\begin{aligned} &&K_{\ell\to \psi} (\mathbf{s},\mathbf{r})\approx-K(\mathbf{s},\mathbf{r})\sum_{i=0}^\infty \left(1-\frac{1}{|\cos\alpha|}\right)^i\nonumber\\ &&+2\frac{g}{g-\kappa}K(\mathbf{s},\mathbf{r})\sum_{i=0}^\infty \left[\left(\frac{g}{\kappa-g}\right) \left(\frac{1}{|\cos\alpha|}-1\right)\right]^i.\label{renormbond2-prev}\end{aligned}$$ Hence, at leading order, $$K_{\ell\to \psi} (\mathbf{s},\mathbf{r})=K(\mathbf{s},\mathbf{r})|\cos\alpha|\frac{g+\kappa|\cos\alpha|}{g-\kappa|\cos\alpha|} \label{renormbond2-prevpost}$$ or, equivalently, $$K_{\ell\to \psi} (\mathbf{s},\mathbf{r})\approx \frac{1}{\kappa}\partial_n K(\mathbf{s},\mathbf{r}) \frac{g+\kappa|\cos\alpha|}{g-\kappa|\cos\alpha|}, \label{renormbond2}$$ if a new $K-$type bond between the wall and the gas-liquid interface emerges from the right extreme of the diagram on the substrate segment. Otherwise $$\begin{aligned} &&K_{\ell\to \psi}^\prime (\mathbf{s},\mathbf{r})=\frac{g}{g-\kappa}\Bigg[K(\mathbf{s},\mathbf{r})\nonumber+\sum_{i=1}^\infty \left(\frac{g}{\kappa-g}\right)^i\nonumber\\ &&\times\int_\psi d\mathbf{s}_1 \cdots d\mathbf{s}_{i} U(\mathbf{s},\mathbf{s}_1)\cdots U(\mathbf{s}_{i-1},\mathbf{s}_i) K(\mathbf{s}_i,\mathbf{r})\Bigg],\label{renormbond3-prevprevprev}\end{aligned}$$ which is approximately given by $$\begin{aligned} &&K_{\ell\to \psi}^\prime (\mathbf{s},\mathbf{r}) \nonumber\\ &&\approx \frac{g}{g-\kappa}K(\mathbf{s},\mathbf{r})\sum_{i=0}^\infty \left[\left(\frac{g}{\kappa-g}\right) \left(\frac{1}{|\cos\alpha|}-1\right)\right]^i.\label{renormbond3-prevprev}\end{aligned}$$ Hence, we have $$K_{\ell\to \psi}^\prime (\mathbf{s},\mathbf{r}) \approx K(\mathbf{s},\mathbf{r})\frac{g|\cos\alpha|}{g-\kappa|\cos\alpha|} \label{renormbond3-prev}$$ and finally $$K_{\ell\to \psi}^\prime (\mathbf{s},\mathbf{r})\approx \frac{1}{\kappa}\partial_n K(\mathbf{s},\mathbf{r}) \frac{g}{g-\kappa|\cos\alpha|} . \label{renormbond3}$$ It follows that in the limit of small curvatures we can perform a resummation of a rather generic convolution of wetting diagram, the ones above providing the most important examples. The resulting diagrams, which are proportional to $\kappa^{-1}(\partial_{n} K)$, are the basic ingredients entering in the binding potential for fixed boundary conditions (i.e., where $g\rightarrow\infty$); this is what we are going to prove in the next section. Alternative representation of the binding potential functional for fixed boundary conditions -------------------------------------------------------------------------------------------- It is possible to systematically explore the curvature corrections to the binding potential through the consideration of connecting the interface and those interfacial segments involving three types of bonds: $U_\pi\equiv \mathcal{K}(\mathbf{r}_\bot)-\delta(\mathbf{r}_\bot)$, $\partial_n K/\kappa$ (only on the substrate), and $\tilde U=K(\mathbf{s},\mathbf{s}^\prime)-\mathcal{K}(\mathbf{r}_\bot)$, which are of order of $1$, $(\kappa R)^{-1}$, and $(\kappa R)^{-2}$, respectively. However, to simplify the discussion we restrict ourselves to the case in which the order parameter is fixed, to a value $m_1$, on the substrate. This will allow us to make the connection with the original nonlocal model formulation more easily. This case corresponds to the limit $g\to -\infty$ and $-h_1/g-m_0\to \delta m_1\equiv m_1-m_0$. Thus, the expansions (\[omega11\]), (\[omega12\]), (\[omega21\]) and (\[deltamdiagram4\]) only include diagrams that do not present $\partial_n K/\kappa$ bonds. On the other hand, the contributions to the coefficients of the segments on the interfaces now become $(-1)^n$, with $n$ being the number of $U-$type bonds of the diagram segment, regardless of whether or not it lies on the liquid-gas interface or substrate. This diagrammatic representation presents the same problems as mentioned above for the finite-$g$ case. However, we can rationalize them using the identities $$\begin{aligned} \label{id_1a} &&\int_\psi d\mathbf{s}_0 K_\psi^{-1}(\mathbf{s},\mathbf{s}_0)K(\mathbf{s}_0,\mathbf{r})\label{renormbond4}\\ &&=\int_\psi d\mathbf{s}_0 \left(\delta(\mathbf{s}-\mathbf{s}_0)+\frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_0)\right)^{-1}\frac{1}{\kappa}\partial_{n_0}K(\mathbf{s}_0,\mathbf{r}), \nonumber\end{aligned}$$ and $$\begin{aligned} \label{id_2a} &&\int_\ell d\mathbf{s}_0 K_\ell^{-1}(\mathbf{s},\mathbf{s}_0)K(\mathbf{s}_0,\mathbf{r})\label{renormbond5}\\ &&=-\int_\ell d\mathbf{s}_0 \left(\delta(\mathbf{s}-\mathbf{s}_0)-\frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_0)\right)^{-1}\frac{1}{\kappa}\partial_{n_0}K(\mathbf{s}_0,\mathbf{r}), \nonumber\end{aligned}$$ where $\partial_n K(\mathbf{s},\mathbf{s}_0)\equiv \mathbf{n}(\mathbf{s})\mathbf{\cdot}\boldsymbol {\nabla}_{\mathbf{s}}K(\mathbf{s},\mathbf{s}_0)$ and $\partial_{n_0} K(\mathbf{s}_0,\mathbf{r})\equiv \mathbf{n}(\mathbf{s}_0)\mathbf{\cdot}\boldsymbol {\nabla}_{\mathbf{s}_0}K(\mathbf{s}_0,\mathbf{r})$. These identities arise from Eq. (\[greenidentity8\]) or, alternatively, from the equivalence of the single- and double-layer potentials  (\[singlelayer\]) and (\[doublelayer\]) for given Dirichlet boundary conditions \[see Eqs. (\[singlelayer2\]) and (\[doublelayer2\])\]. By using the fundamental relations $\int \mathcal{O}\mathcal{O}^{-1}=\delta$ for the inverse operators $(\delta \pm \partial K/\kappa)^{-1}$ we obtain $$\begin{aligned} \label{inverseops1} &&\left(\delta(\mathbf{s}-\mathbf{s}_0)\pm \frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_0)\right)^{-1}=\delta (\mathbf{s}-\mathbf{s}_0) \\&&\mp \int d\mathbf{s}_1 \left(\delta(\mathbf{s}-\mathbf{s}_1)\pm \frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_1)\right)^{-1}\frac{1}{\kappa}\partial_{n_1}K(\mathbf{s}_1,\mathbf{s}_0), \nonumber\end{aligned}$$ which formally can be represented as $$\begin{aligned} \label{inverseops2} &&\left(\delta(\mathbf{s}-\mathbf{s}_0)\pm \frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_0)\right)^{-1}=\delta (\mathbf{s}-\mathbf{s}_0) \\&&\mp \frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_0)+ \int d\mathbf{s}_1 \frac{1}{\kappa}\partial_n K(\mathbf{s}, \mathbf{s}_1) \frac{1}{\kappa}\partial_{n_1} K(\mathbf{s}_1, \mathbf{s}_0) + \cdots . \nonumber\end{aligned}$$ It is straightforward to recognize that Eqs. (\[renormbond4\]) and (\[renormbond5\]) can be represented diagrammatically as $$\begin{aligned} \label{id_1} &&{\parbox{1.5cm}{\epsfig{file=diagram69.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram70.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram71.eps,width=1.5cm}}}-\cdots\nonumber\\ &&=-{\parbox{1.5cm}{\epsfig{file=diagram72.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram73.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram74.eps,width=1.5cm}}}-\cdots\label{identity1} \\ \label{id_2} &&{\parbox{1.5cm}{\epsfig{file=diagram69.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram75.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram76.eps,width=1.5cm}}}-\cdots\nonumber\\ &&={\parbox{1.5cm}{\epsfig{file=diagram77.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram78.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram79.eps,width=1.5cm}}}-\cdots,\label{identity2}\end{aligned}$$ where the bonds carrying arrows linking both interfaces are $\partial_n K/\kappa$ functions, and the arrow indicates the position where the normal derivative is applied. Note that the right-hand sides of these equations correspond to an expansion in powers of $(\kappa R)^{-1}$, as $\partial_n K/\kappa\sim (\kappa R)^{-1}$. On the other hand, the leading-order contributions are consistent with the renormalized bonds obtained previously. \[Consider the limit $g\to -\infty$ in Eqs. (\[renormbond1\]), (\[renormbond2\]), and (\[renormbond3\]).\] On using Eqs. (\[identity1\]) and (\[identity2\]), we obtain the alternative representation of the binding potential $$\begin{aligned} &&\frac{W[\ell,\psi]}{\kappa m_0^2}=\sum_{n=1}^\infty \Bigg(\frac{\delta m_1}{m_0} \Omega_n^n+ \Omega_n^{n+1} +\left(\frac{\delta m_1}{m_0}\right)^2 \Omega_{n+1}^n\Bigg), \nonumber\\ \label{Hmin7}\end{aligned}$$ which is now similar to the structure of the original nonlocal treatment, for example, the leading-order contribution, viz., $$\begin{aligned} \Omega_1^1=&& -{\parbox{1.5cm}{\epsfig{file=diagram80.eps,width=1.5cm}}}+ {\parbox{1.5cm}{\epsfig{file=diagram81.eps,width=1.5cm}}}\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram82.eps,width=1.5cm}}} -{\parbox{1.5cm}{\epsfig{file=diagram83.eps,width=1.5cm}}}\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram84.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram85.eps,width=1.5cm}}}+\cdots,\label{omega11-2}\end{aligned}$$ while $$\begin{aligned} \Omega_1^2=&& -{\parbox{1.5cm}{\epsfig{file=diagram86.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram87.eps,width=1.5cm}}}\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram88.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram89.eps,width=1.5cm}}}\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram90.eps,width=1.5cm}}}+\cdots, \label{omega12-2}\end{aligned}$$ and $$\begin{aligned} \Omega_2^1=&& -{\parbox{1.5cm}{\epsfig{file=diagram91.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram92.eps,width=1.5cm}}}\nonumber\\ &&-{\parbox{1.5cm}{\epsfig{file=diagram93.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram94.eps,width=1.5cm}}}\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram95.eps,width=1.5cm}}}+\cdots, \label{omega21-2}\end{aligned}$$ and so on. Each diagram has segments on each interface connected via $\partial K/\kappa-$type bonds that link the wall and the gas-liquid interface. The leftmost segment can only contain $U-$type bonds (independent of the surface on which it lies). Otherwise, they only have $\partial K/\kappa$ bonds. The coefficient associated with each diagram is now $(-1)^{l+m+o}$, with $l$ being the total number of $U-$bonds in the leftmost segment, $m$ the total number of $\partial_n K/\kappa-$type bonds on the substrate (not on the liquid-gas interface), and $o$ the number of $\partial K/\kappa$ bonds between the wall and the gas-liquid interface that emerge from the substrate and point to the gas-liquid interface. Similarly, the order-parameter profile has an alternative diagrammatic expansion $$\begin{aligned} \delta m_\Xi&&=-m_0\Bigg[{\parbox{1.5cm}{\epsfig{file=diagram59.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram60.eps,width=1.5cm}}}+\cdots\nonumber\\ &&-{\parbox{1.5cm}{\epsfig{file=diagram96.eps,width=1.5cm}}}+{\parbox{1.5cm}{\epsfig{file=diagram97.eps,width=1.5cm}}}+\cdots\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram98.eps,width=1.5cm}}}+\cdots\Bigg]\nonumber\\ &&+\delta m_1 \Bigg[{\parbox{1.5cm}{\epsfig{file=diagram64.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram65.eps,width=1.5cm}}} +\cdots\nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram99.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram100.eps,width=1.5cm}}}+\cdots \nonumber\\ &&+{\parbox{1.5cm}{\epsfig{file=diagram101.eps,width=1.5cm}}}+\cdots\Bigg] \label{deltamdiagram5}\end{aligned}$$ The diagrams start with a (left) segment either on the substrate or on the liquid-gas interface, which only can have $U-$type bonds. After that, there are $\partial_n K/\kappa-$type bonds connecting the wall and the gas-liquid interface, followed by segments on the corresponding interface that can only have $\partial_n K/\kappa-$bonds. Finally, there is a $K-$type bond connecting one interface to the position $\mathbf{r}$. Now the sign in front of each diagram is $(-1)^{l+m+o'}$, with $l$ being the number of $U-$type bonds on the leftmost segment, $m$ the number of $\partial_n K/\kappa-$type bonds on the substrate, and $o'$ the number of $\partial_n K/\kappa$ bonds that emerge from the liquid-gas interface and point to the substrate. Flat substrates --------------- In the previous sections we pointed out that the decorated diagrams constitute the different features of our formulation of the nonlocal model. In order to better appreciate these aspects, we consider the binding potential for the case of a flat substrate. The exact binding potential admits a curvature expansion, but even at leading order it differs slightly from the binding potential functional of the original formulation. We start by considering the diagrams contained in $\Omega_{1}^{1}$. From the results of the previous sections we have that $$\begin{aligned} \nonumber \label{reduction_01} \Omega_{1}^{1} & = & -{\parbox{1.5cm}{\epsfig{file=diagram102.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram103.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram104.eps,width=1.5cm}}} \\ \nonumber && -{\parbox{1.5cm}{\epsfig{file=diagram105.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram106.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram107.eps,width=1.5cm}}} +\cdots , \\\end{aligned}$$ where the flatness of the substrate has enormously simplified the diagrammatic structure. This simplification is due to the vanishing of a large class of wetting diagrams, and can be summarized by the following reduction lemmas.\ *Lemma 1*. We have $$\label{Lemma1} {\parbox{1.5cm}{\epsfig{file=diagram108.eps,width=1.5cm}}} = {\parbox{1.5cm}{\epsfig{file=diagram109.eps,width=1.5cm}}}.$$ *Lemma 2*. We have \[Lemma2\] $${\parbox{1.5cm}{\epsfig{file=diagram110.eps,width=1.5cm}}} = -\cos\alpha(\textbf{s}) {\parbox{1.5cm}{\epsfig{file=diagram111.eps,width=1.5cm}}},$$ where $\alpha(\textbf{s})$ is the angle formed by the normal vector and the vertical direction.\ *Lemma 3*. We have \[Lemma3\] $$\begin{aligned} {\parbox{3.5cm}{\epsfig{file=diagram112.eps,width=3.5cm}}} & = & 0 , \\ {\parbox{3.5cm}{\epsfig{file=diagram113.eps,width=3.5cm}}} & = & 0 .\end{aligned}$$ In addition to these rules, we recall that the decorated diagrams contribute higher-order corrections in the curvature expansion. In particular, a wetting diagram with a chain of $n$, $U$-type bonds on the fluid interface belongs to ${O}\left(R^{-2n}\right)$, while instead a chain of $m$ arrow diagrams along the substrate belongs to ${O}\left(R^{-m}\right)$, where $R$ is a typical radius of curvature. Therefore, at the leading order of a curvature expansion only the first two addends survive in Eq. (\[reduction\_01\]). Then, due to Lemmas 1 and 2, we can further simplify the leading term and we are left with $$\label{reduction_02} \Omega_{1}^{1} \approx \left(1+\langle\cos\alpha\rangle_1\right) {\parbox{1.5cm}{\epsfig{file=diagram114.eps,width=1.5cm}}},$$ where $$\langle\cos\alpha\rangle_n\equiv\frac{\int_\ell d\mathbf{s} \cos \alpha(\mathbf{s})e^{-n\kappa \ell(\mathbf{s})}}{\int_\ell d\mathbf{s} e^{-n\kappa \ell(\mathbf{s})}}, \label{averagecosalpha}$$ with $\ell(\mathbf{s})$ being the vertical distance from $\mathbf{s}$ to the substrate. Already at leading order we can appreciate the different features of this exact formulation. Indeed, in the original formulation the expansion (\[reduction\_01\]) starts with the same diagram entering in (\[reduction\_02\]) but with a factor $2$ in front. The factor $1+\cos\alpha(\mathbf{s})$ strongly depends on the local orientation of the interface with respect to the planar wall and it is clear that the two formulations coincide only for parallel interfaces. However, for interfacial configurations that have a minimum height $\tilde\ell$ with respect to the substrate, the weighted average (\[averagecosalpha\]) is near unity. More precisely, a saddle-point calculation shows that $\langle \cos \alpha \rangle_1\sim 1- (\tilde H/\kappa)$, where $\tilde H$ is the mean curvature at the interfacial position nearest the substrate. It then is straightforward to prove, using the above lemmas, that the next-to-leading diagrams appearing in (\[reduction\_02\]) are of the form $$\nonumber {\parbox{3.5cm}{\epsfig{file=diagram115.eps,width=3.5cm}}} ,\qquad {\parbox{3.5cm}{\epsfig{file=diagram116.eps,width=3.5cm}}} ,$$ with a prefactor $-1$ and $(-1)^{n}$, respectively, at ${O}\left(R^{-n}\right)$ and ${O}\left(R^{-2n}\right)$, respectively. These considerations apply also for the remaining classes of diagrams; in particular, for $\Omega_{2}^{1}$ we have $$\label{reduction_03} \Omega_{2}^{1} = - {\parbox{1.5cm}{\epsfig{file=diagram117.eps,width=1.5cm}}} - {\parbox{1.5cm}{\epsfig{file=diagram118.eps,width=1.5cm}}} - {\parbox{1.5cm}{\epsfig{file=diagram119.eps,width=1.5cm}}} - {\parbox{1.5cm}{\epsfig{file=diagram120.eps,width=1.5cm}}} - \cdots,$$ where the $n$th diagram belongs to ${O}\left(R^{1-n}\right)$. Again, by using the previous lemmas, the leading term of $\Omega_{2}^{1}$ can be written as $$-{\parbox{1.5cm}{\epsfig{file=diagram117.eps,width=1.5cm}}}=\langle \cos \alpha\rangle_2 {\parbox{1.5cm}{\epsfig{file=diagram121.eps,width=1.5cm}}},$$ where $\langle \cos\alpha \rangle_2\sim 1-\tilde H/2\kappa$ by a saddle-point calculation. The effect of the reduction is less effective for the class $\Omega_{1}^{2}$, for which the segments are located on the fluid interface. However, again a saddle-point calculation shows that, up to ${O}(R^{-1})$ terms, $$-{\parbox{1.5cm}{\epsfig{file=diagram122.eps,width=1.5cm}}}\approx {\parbox{1.5cm}{\epsfig{file=diagram123.eps,width=1.5cm}}},$$ recovering the original formulation of the nonlocal model. However, we should stress that this is a highly nonlocal formulation in the sense that the total binding potential between the wall and the gas-liquid interface is obtained. However, if we would like to characterize the influence of the substrate locally on a portion of the liquid-gas interface, we have to resort to the nonlocal model presented in this paper. In particular, the functionals $\Omega_{1}^{1}$ and $\Omega_{2}^{1}$ are local, so their contribution to the binding potential arises from $$\begin{aligned} -{\parbox{1.5cm}{\epsfig{file=diagram102.eps,width=1.5cm}}} + {\parbox{1.5cm}{\epsfig{file=diagram114.eps,width=1.5cm}}} + {O}\left(R^{-1}\right) , \\ - {\parbox{1.5cm}{\epsfig{file=diagram117.eps,width=1.5cm}}} + {O}\left(R^{-1}\right), \end{aligned}$$ where now the integration on the liquid-gas interface is restricted to the portion of the gas-liquid interface in which the binding potential is evaluated.A different feature, absent in the original formulation, emerges due to a coupling between the interface position and its orientation. However, as in the original formulation, the $\Omega_1^2$ functional is highly nonlocal and has the representation $${\parbox{1.5cm}{\epsfig{file=diagram122.eps,width=1.5cm}}}=-\int_\ell d\mathbf{s}_1 d\mathbf{s}_2 \textrm{e}^{-\kappa [\ell(\mathbf{s}_1)+\ell(\mathbf{s}_2)]}\bar{S}(x_{12},\bar{\ell}),$$ where $x_{12}$ is the projection of $\mathbf{s}_2-\mathbf{s}_1$ onto the substrate plane and $\bar{S}$ is the effective two-body interaction between the interfacial area elements located around $\mathbf{s}_1$ and $\mathbf{s}_2$. As the corresponding interaction $S\equiv S(x_{12},\bar{\ell})$ in the original nonlocal model [@parry1; @parry4; @parry5], $\bar{S}$ depends on the interfacial heights via $\bar{\ell}\equiv [\ell(\mathbf{s}_1)+\ell(\mathbf{s}_2)]/2$, which can be analyzed by using the same renormalization-group (RG) flow equations derived in Refs. [@parry1; @parry4; @parry5]. More specifically, $$\bar{S}(x_{12},\bar\ell)=\frac{e^{2\kappa \bar\ell}}{\kappa}\mathbf{n}(\mathbf{s}_2)\mathbf{\cdot}\boldsymbol \nabla_2 \mathcal{K}\left[\sqrt{x_{12}^2+(2\bar\ell)^2}\right],$$ where $\boldsymbol\nabla_2$ is the 3D gradient acting on the liquid-gas interfacial position $\mathbf{r}_2=(\mathbf{s}_2,\ell(\mathbf{s}_2))$. For large $\ell$, a saddle-point calculation shows that $$\bar{S}(x_{12},\bar\ell)\approx -\kappa\cos\alpha_2/2\pi\bar{\ell} \textrm{e}^{-\frac{\kappa x_{12}^2}{2\bar{\ell}}}= -\cos \alpha_2 S(x_{12},\bar{\ell}),$$ where $\alpha_2$ is the angle formed by the normal vector and the vertical direction at the liquid-gas interface position $\mathbf{r}_2$. Thus, as $S$, the two-body interaction $\bar{S}$ has a Gaussian form, with the nonlocal length $\xi_{NL}=\sqrt{\langle l\rangle/\kappa}$ precisely as identified in the original formulation [@parry4; @parry5]. However, our improved formulation introduces as a different feature the coupling of the two-body interaction to the surface orientation through the factor $\cos \alpha_2$. A detailed comparison of the RG flows of this effective two-body repulsion within the original and present, exact, formulations in the context of critical wetting is beyond the scope of the present paper. Summary and remarks ------------------- In this section we have applied the boundary integral diagrammatic method to determine the binding potential functional and order-parameter configuration when a wetting layer intrudes between the bulk phase and the wall. Our results are decorated versions of those appearing in the original formulation and, in particular, contain $U-$type kernels on the fluid interface, while on the substrate they show $U-$type and $\kappa^{-1}(\partial_{n}K)-$type bonds. The effect of the diagrams can be readily understood for small curvature, where the pertinent multiply embedded convolutions can be resummed, leading to renormalized diagrams involving the orientation of the surface. In this way, the full nonlocality is replaced by a weaker version, which can be used to build a more readily usable effective binding potential functional. As expected, our formulation reveals curvature corrections to the original formulation that are reliable when the substrate and fluid interfacial configurations are parallel or concentric, as in the case of spherical and cylindrical symmetry. Strictly speaking, when the interfaces are nonparallel the present improved formulation must be used; the analysis of filling transitions for fluids adsorbed in wedge geometries is a natural place for investigating this. When the surface order parameter at the wall is fixed (i.e., Dirichlet boundary conditions) and the system is at the location of the critical wetting transition ($m_1=m_0$), as pertinent to the critical isotherm, the only diagram of relevance remaining is $\Omega_1^2$. This term is strongly nonlocal and has a structure very similar to that appearing in the original nonlocal formulation. Once again, this highlights the influence of an effective two-body Gaussian interfacial interaction controlled by a nonlocal length $\xi_{NL}=\sqrt{\langle l\rangle/\kappa}$ that is missing entirely from the original local effective Hamiltonian treatments of the critical wetting phase transition. Conclusions =========== In this paper we have presented a rigorous derivation of the nonlocal effective interfacial Hamiltonian model for interfaces and wetting in systems possessing short-ranged forces. The present derivation, which is based on a boundary integral formulation, improves on the one given originally, because the boundary conditions at the interface and wall are now handled exactly rather than approximately. The first point to emphasize is that this systematic analysis can indeed be done at all, at least using a simple DP potential and the crossing criterion definition of the interface position (to which we will return later). This analysis can also be expressed diagrammatically; a glossary of the elementary diagrams from which all other diagrams follow is given in Appendix C, together with their algebraic expressions. As with the original formulation, each diagram containing a line that spans the liquid wetting layer, thus connecting the liquid-gas interface and wall, can be thought of as an interaction between these surfaces mediated by a bulklike correlation. Among other things, this rigorous formulation allows us to consider, in a systematic fashion, the nature of the curvature corrections appearing in the appropriate free energy. More specifically, we applied the boundary integral method to three situations with the following conclusions. - [*The wall-single-phase interface*]{}. First we considered a nonplanar wall-liquid interface, where a wetting layer is absent. We showed that the leading-order curvature corrections to the surface tension term involve the local mean and Gaussian curvatures, in the spirit of the Helfrich free energy, with bending and saddle-splay rigidity coefficients, respectively, the values of which are identified. However, the curvature expansion does not truncate at this, or indeed any, order and the free energy does not conform to the morphological thermodynamics hypothesis [@konig]. - [*The free liquid-gas interface*]{}. Extending this analysis to the free (but constrained) liquid-gas interface, we showed that the curvature corrections can be expressed more precisely as an interfacial self-interaction, the form of which is identical to that proposed in Ref. [@PR] using less rigorous methods. Indeed, the order-parameter profiles are also identical, lending strong support to the approximate methods used previously to discuss nonlocality. - [ *The binding potential functional*]{}. For the case in which a wetting layer is present we have derived a generalized diagrammatic representation of the binding potential and order parameter, which contains decorated versions of those diagrams appearing in the original formulation. These generate, in addition to curvature corrections, a coupling between the interfacial orientation and position, which is missing entirely in the original theory. Indeed, strictly speaking, even for small curvatures the diagrams do not converge to those of the original formulation, [*unless*]{} the interfacial configurations are nearly parallel to the substrate. However, when our formulation of the nonlocal model is applied to a flat substrate, we find features that are very similar to the original version of the nonlocal model. In particular, the contributions $\Omega_1^1$ and $\Omega_2^1$ to the binding potential functional are local, while the $\Omega_1^2$ contribution remains [*nonlocal*]{} and can be expressed as a two-body Gaussian interfacial self-interaction, mediated by the substrate, having a lateral range given by the same nonlocal lengthscale $\xi_{NL}=\sqrt{\overline{\ell}/\kappa}$. Thus, the criticism of what is missing in local interfacial Hamiltonian descriptions of critical wetting in Refs. [@BHL; @FH], including the size of the critical regime and also the paradoxical prediction of possible fluctuation-induced first-order transitions [@Jin2; @Jin4], remains unchanged (see Refs. [@parry4; @parry5]). Nevertheless, it would be interesting to include the coupling of orientation and position into renormalization group and simulation studies of the nonlocal model. Having formulated the problem exactly for the DP potential, it is possible to make extensions to more general potentials perturbatively, by using a Feynman-Hellmann theorem similar to the approximate analysis of Ref. [@parry3]. For the wall-liquid and free liquid-gas interface, this would generate further curvature corrections to the free energy, although this will not alter the diagrammatic structure only altering the values of the coefficients. However, in applications to the wetting film, the binding potential will now contain decorated versions of the $\chi$ diagram identified in Ref. [@parry3]. To identify the curvature corrections to this term, further resummation of the diagrammatic series is required, similar to the decorated diagrams in the DP model discussed here. Generalizations to heterogeneous walls are also technically possible using the boundary integral approach. Our rigorous and rather technical derivation of the nonlocal model is still subject to a number of criticisms. For example, we have assumed that the surface field $h_1$ and enhancement $g$ are not altered by the surface curvature of a structured wall, which is very probably an oversimplification. In addition, of course, the continuum LGW model (\[HLGW\]) does not in any way account for volume exclusion and local layering present when a high-density fluid is adsorbed at a wall. There are also alternative definitions of *the* interfacial position. For example, Fisher and Jin discuss integral criteria, and show that these alter the coefficients appearing in the binding potential function (see Refs. [@Jin1; @Jin3]). Hopefully, the diagrammatic structure of the binding potential functional is not altered when using a different definition of the interfacial position, although we should expect that the values of all coefficients and curvature corrections are altered. We should also mention that, of course, as soon as long-ranged forces are present all results here change dramatically [@mecke]. For example, exponential terms are replaced by algebraic terms in the binding potential. Additionally, for Lennard-Jones forces the curvature expansion of the interfacial free energy fails completely, due to nonanalytic logarithmic corrections. However, there are deeper issues concerning the connection between mesoscopic and microscopic descriptions, which highlight some of the fundamental problems still open in the theory of interfacial phenomena discussed here. For example, within the crossing criterion, for any potential $\phi(m)$, there is no escape from having a negative bending rigidity $\kappa_B$ (the positive saddle-splay rigidity plays no part since the principal radius of curvature along the wedge is infinite). However, the very meaning of having a negative bending coefficient has been questioned by Chacón and Tarazona [@tarazona13], who have argued that the continuum LGW Hamiltonian is already too coarse grained to enable a direct determination of the rigidity from a constrained minimization of the model. At a microscopic level, they argue that there must be a molecular top to the capillary-wave spectrum, which leads to a positive rigidity. While density-functional models may be consistent with this feature when we look closely at the structure of the equilibrium density-density correlation function, a constrained minimization of any model functional will not suffice. Alternatively, they propose that the constrained minimization is replaced by a weighted convolution, which smears the interface location over a region comparable with the bulk correlation length. This means, of course, that the interface position no longer has a strict crossing-criterion interpretation. In fact, it has been shown that the crossing criterion does not distinguish correctly bulk from interfacial contributions present in the mean-field correlation function and therefore cannot be used naively to determine any wave-vector-dependent corrections to the surface tension [@parry8]. These concerns must also be married with the observation that the mean-field identification [@Jin1; @Jin3] is, strictly speaking, only valid in the limit of low temperatures (i.e., $T\to 0$) where a saddle-point evaluation of the partial trace suffices. Finite-temperature corrections to the interfacial free energy, interfacial Hamiltonian, and binding potential must be present at some order. Indeed, these corrections are already allowed for implicitly when, in the application of the interfacial Hamiltonian, the mean-field value of the surface tension is replaced by its true thermodynamic value. These ideas, which are still under development, of course mean that the determination of the binding potential functional for wetting layers and the values, and indeed the signs, of the coefficients of all curvature correction terms, are much more difficult to determine. J.M.R.-E. is grateful for financial support from the Spanish Ministerio de Economía y Competitividad through Grants No. FIS2017-87117-P and No. FIS2012-32455, Junta de Andalucía through Grant No. P09-FQM-4938, all co-funded by the EU FEDER, and the Portuguese Foundation for Science and Technology under Contract No. EXCL/FIS-NAN/0083/2012. A.S. acknowledges hospitality of the University of Seville. A.O.P. acknowledges support from the EPSRC UK Grant No. EP/L020564/1 “Multiscale Analysis of Complex Interfacial Phenomena.” P.M.G. gratefully acknowledges support from NSF Grant No. DMR12-1207026. The work of P.M.G. was performed in part at the Aspen Center for Physics, which is supported by U.S. National Science Foundation Grant No. PHY-1607611. Perturbative solutions of the boundary integral equations {#appendix_A} ========================================================= In this appendix we illustrate how to solve the integral equations which emerge in our analysis of the curvature expansion. We start with the evaluation of the first terms in the curvature expansion of the normal derivative $q$ for a single phase in contact with a substrate $\psi$. After substitution of the curvature expansions of the terms which appear in (\[inteq1\]) we find a recursive chain of equations for the $q_{n}$’s up to ${O}(R^{-3})$ $$\begin{aligned} &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{0} \left[\mathcal{K}(r_{\bot}) - \frac{\kappa}{g}\delta (\mathbf{r}_{\bot})\right] = \kappa\left(\frac{h_1}{g}+m_b\right) \label{exph0}\\ \nonumber\\ &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{1} \left[\mathcal{K}(r_{\bot}) - \frac{\kappa}{g}\delta (\mathbf{r}_{\bot})\right] \nonumber\\ &&=-\frac{1}{g}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \mathcal{W}(r_{\bot}) \Delta\psi(\mathbf{r}_{\bot}) \left(q_0 +h_1+gm_b\right), \label{exph1}\end{aligned}$$ and $$\begin{aligned} &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{2} \left[\mathcal{K}(r_{\bot}) - \frac{\kappa}{g}\delta (\mathbf{r}_{\bot})\right] \nonumber\\ && =\frac{1}{2}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot}\biggl[ \mathcal{W}(r_{\bot}) \Delta\psi(\mathbf{r}_{\bot})\left(q_0 \Delta\psi(\mathbf{r}_{\bot}) - \frac{2q_1}{g}\right) \nonumber\\&&-q_0 \left(\mathcal{K}(r_{\bot}) \left(\boldsymbol\nabla_{\bot}\Delta\psi(\mathbf{r}_{\bot}) \right)^{2}+\frac{2}{g} \mathcal{W}(r_{\bot}) \chi (\mathbf{r}_{\bot})\right) \biggr]\label{exph2} \end{aligned}$$ where $\mathcal{K}(x)=\kappa \exp(-\kappa x)/2\pi x$, $\mathcal{W}(x) =(1+\kappa x)\mathcal{K}(x)/x^2$ and we have extended the integral to $\mathbb{R}^{2}$, ignoring exponentially decaying terms on $\kappa R$. Note that self-consistency means that only the leading terms of $\Delta \psi$ and $\chi\equiv {\mathbf{r}_\bot}{\mathbf{\cdot}} \boldsymbol\nabla_{\bot}\Delta\psi-2\Delta \psi$, which scale as $R^{-1}$ and $R^{-2}$, respectively, should be used. For a flat interface $q_{0}(\mathbf{s})$ is translationally invariant; therefore, it can be factorized from the integral, but since $\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \mathcal{K}(r_{\bot}) =1$ we have that $q_{0}$ is given by Eq. (\[qflat\]). However, it will be useful to develop a further technique to solve the Eqs. (\[exph0\])-(\[exph2\]). We define a parallel Fourier transform, in which only the fluctuating modes parallel to the interface are considered. The kernel reads $$K(\mathbf{s}-\mathbf{s}^{\prime}) = \int_{\mathbb{R}^{2}} \frac{d^{2} \mathbf{q}}{(2\pi)^{2}} \, \textrm{e}^{i \mathbf{q} \mathbf{\cdot} (\mathbf{s}-\mathbf{s}^{\prime})} \, \widetilde{K}(\mathbf{q}) ,$$ and with a simple complex integration, we get the inverse Fourier transform $$\label{kerneltransform} \widetilde{K}(\mathbf{q}) = \int_{\mathbb{R}^{2}} d \mathbf{s} \, \textrm{e}^{-i \mathbf{q} \mathbf{\cdot} (\mathbf{s}-\mathbf{s}^{\prime})} \, K(\mathbf{s}-\mathbf{s}^{\prime}) = \frac{\kappa}{\sqrt{\kappa^{2}+q^{2}}} .$$ With these definitions the convolution equation for $q_{0}$ becomes an algebraic equation for the Fourier modes $$\widetilde{q_{0}}(\mathbf{q}) = (2\pi)^{2} \kappa \left(\frac{h_1}{g}+m_b\right) \frac{\delta(\mathbf{q})}{\widetilde{K}(-\mathbf{q})-\frac{\kappa}{g}} ,$$ and transforming back to real space we find $q_{0}=-\kappa (h_1+gm_{b})/(\kappa-g)$. Let us consider now the equations for the ${O}(R^{-1})$ and ${O}(R^{-2})$. If the leading contributions to $\Delta \psi$ and $\chi$ in powers of $(\kappa R)^{-1}$ are used in Eqs. (\[exph1\]) and (\[exph2\]), the integrations over $\mathbf{r}_{\bot}$ on their right-hand sides can be performed in polar coordinates. After a few simple calculations we find $$\begin{aligned} \nonumber &&\int_{\psi} d \mathbf{r}_{\bot} \, q_{1} \left[\mathcal{K}(r_{\bot})-\frac{\kappa}{g}\delta(\mathbf{r}_{\bot})\right] \nonumber\\ && =\left(\frac{h_1+gm_{b}}{\kappa-g}\right) \left(\frac{k_{1}+k_{2}}{2} \right) = \frac{h_1+gm_{b}}{\kappa-g} H\end{aligned}$$ and $$\begin{aligned} &&\int_{\psi} d \mathbf{r}_{\bot} \, q_{2} \left[\mathcal{K}(r_{\bot})-\frac{\kappa}{g}\delta(\mathbf{r}_{\bot})\right] =\nonumber\\ && \kappa\frac{h_1+g m_{b}}{8(\kappa-g)} \left( \frac{k_{1} - k_{2}}{\kappa} \right)^{2} \nonumber\\ &&-\frac{1}{g}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} q_1 \mathcal{W}(r_{\bot}) \Delta\psi(\mathbf{r}_{\bot}) =\nonumber\\ && \frac{h_1+g m_{b}}{2\kappa(\kappa-g)} \biggl[ H^{2} -K_{G} \biggr] \nonumber\\ &&-\frac{1}{g}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} q_1 \mathcal{W}(r_{\bot}) \Delta\psi(\mathbf{r}_{\bot}), \label{eqq2}\end{aligned}$$ Note that the results of the integrations are expressed in terms of the mean and Gaussian curvatures of the interface, both evaluated at the origin. If we denote by $\mathcal{R}_{1,2}(\mathbf{s})$ the right-hand sides of these equations, their formal solution reads $$q_{1,2}(\mathbf{s}) = \int_{\mathbb{R}^{2}}\frac{d^{2} \mathbf{q}}{(2\pi)^{2}} \, \textrm{e}^{i \mathbf{q} \mathbf{\cdot} \mathbf{s}} \frac{\widetilde{\mathcal{R}}_{1,2}(\mathbf{q})}{\frac{1}{\sqrt{1+\frac{q^2}{\kappa^2}}}-\frac{\kappa}{g}} .$$ For our substrate $\kappa R\gg 1$, so the integral is dominated by the slow sector of Fourier modes. Hence it is reasonable to expand the square root in powers of the small parameter $q/\kappa$, thus $$\begin{aligned} \nonumber \label{q2gradient} q_{1,2}(\mathbf{s}) &\approx& \int_{\mathbb{R}^{2}}\frac{d^{2} \mathbf{q}}{(2\pi)^{2}} \, \frac{\textrm{e}^{i \mathbf{q} \mathbf{\cdot} \mathbf{s}}}{1-\frac{\kappa}{g}} \Biggl( 1 + \frac{q^{2}}{2\kappa^{2}(1-\frac{\kappa}{g})} \Biggr) \, \widetilde{\mathcal{R}}_{1,2}(\mathbf{q}) \\ \nonumber & = & -\frac{g}{\kappa-g}\Biggl(\mathcal{R}_{1,2}(\mathbf{s})+\frac{g}{2\kappa^{2}}\frac{\nabla^{2}_\bot\mathcal{R}_{1,2}(\mathbf{s})}{\kappa-g}\nonumber\\&+&{O}\left(\nabla^{4}_\bot\mathcal{R}_{1,2}(\mathbf{s})\right)\Biggr) .\end{aligned}$$ The same result is obtained if we make a Taylor expansion of $q_{1,2}$ around the origin and substitute in Eqs. (\[exph1\]) and (\[exph2\]). However, we note that $\nabla_\bot^2 \mathcal{R}_{1,2} \sim \mathcal{R}_{1,2}/R^2$, so the derivative terms contribute to higher-order curvature terms and thus they can be neglected. The solutions are then given by the leading contributions of Eq. (\[q2gradient\]), which correspond to Eqs. (\[q1\]) and (\[q2\]). In a similar way, the perturbative scheme for the computation of the normal derivatives $q$ for a free interface is $$\begin{aligned} \label{pert_q_int_1} \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{0}^{\pm} \mathcal{K}(r_{\bot}) & = & -\kappa m_{0} \\ \label{pert_q_int_2} \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{1}^{\pm} \mathcal{K}(r_{\bot}) & = & \pm m_{0} \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \mathcal{W}(r_{\bot}) \Delta\ell(\mathbf{r}_{\bot}), \end{aligned}$$ and $$\begin{aligned} \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{2}^{\pm} \mathcal{K}(r_{\bot}) & = & \frac{1}{2}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, q_{0}^{\pm} \biggl\{ \mathcal{W}(r_{\bot}) \Delta\ell(\mathbf{r}_{\bot})^{2} \label{pert_q_int_3} \nonumber\\&-&\mathcal{K}(r_{\bot}) \left[\boldsymbol\nabla_{\bot}\Delta\ell(\mathbf{r}_{\bot})\right]^{2} \biggr\}.\end{aligned}$$ These have the solutions $$\begin{aligned} \label{pert_q_sol_1} q_{0}^{\pm}& =& -\kappa m_{0},\\ \label{pert_q_sol_2} q_{1}^{\pm} & = & \pm \kappa m_{0} \frac{k_{1}+k_{2}}{2\kappa} = \pm m_{0} H(\mathbf{s}),\\ \nonumber \label{pert_q_sol_3}\end{aligned}$$ and $$\begin{aligned} q_{2}^{\pm} & = & \frac{\kappa m_{0}}{8} \left( \frac{k_{1} - k_{2}}{k} \right)^{2} \nonumber\\ & = &\frac{m_{0}}{2\kappa} \biggl[ H(\mathbf{s})^{2} -K_{G}(\mathbf{s}) \biggr] . \label{exph22}\end{aligned}$$ Finally, the curvature expansion of $\Psi$ for the single phase in contact with the substrate can be obtained from Eq. (\[singlelayer2\]). After substitution of Eq. (\[boundary2\]) and the curvature expansions of $q$, the kernel $K$, and the elementary area $d\mathbf{s}$ into Eq. (\[singlelayer2\]), we obtain the equations $$\begin{aligned} \label{psi_1} &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \Psi_{0} \mathcal{K}(r_{\bot}) = -\frac{h_1 + gm_b}{g}- \frac{q_0}{g}, \\ \label{psi_2} &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \Psi_{1} \mathcal{K}(r_{\bot}) = -\frac{q_1}{g},\end{aligned}$$ and $$\begin{aligned} &&\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \Psi_{2} \mathcal{K}(r_{\bot}) = -\frac{q_2}{g} -\frac{1}{2}\int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \frac{\Psi_0}{g} \nonumber \\&\times& \biggl\{ \mathcal{W}(r_{\bot}) \Delta\psi(\mathbf{r}_{\bot})^{2} -\mathcal{K}(r_{\bot}) \left[\boldsymbol\nabla_{\bot}\Delta\psi(\mathbf{r}_{\bot})\right]^{2} \biggr\}. \label{psi_3}\end{aligned}$$ where $\Psi_0$, $\Psi_1$, and $\Psi_2$ stand for the first terms in the curvature expansion of $\Psi$. The solutions of these integral equations are $$\begin{aligned} \label{psi_1_sol} \frac{\Psi_0}{2\kappa}&=&\frac{h_1+gm_b}{\kappa-g}, \\ \label{psi_2_sol} \frac{\Psi_1}{2\kappa}&=&\left(\frac{h_1+gm_b}{\kappa-g}\right)\left(\frac{\kappa}{\kappa-g}\right)\frac{H}{\kappa},\end{aligned}$$ and $$\begin{aligned} \frac{\Psi_1}{2\kappa}&=&\left(\frac{h_1+gm_b}{\kappa-g}\right)\Bigg[\left(\frac{1}{2}\frac{g}{\kappa-g}+\frac{\kappa^2} {(\kappa-g)^2}\right)\left(\frac{H}{\kappa}\right)^2\nonumber\\ &-&\frac{1}{2}\frac{g}{\kappa-g}\frac{K_G}{2\kappa^2}\Bigg]. \label{psi_3_sol}\end{aligned}$$ Derivation of the liquid-gas interfacial self-interaction Hamiltonian {#appendix_B} ===================================================================== ![Schematic illustration of the coordinates, vectors, and geometry appearing in the curvature expansion for a constrained interfacial configuration. The symbols are described in the text.[]{data-label="construction"}](figure2.eps){width="9cm"} In this appendix we derive (\[ham4\]). Consider two points on the interface with $\mathbf{s}_{1}$ as the origin and $\mathbf{s}_{2}$ as in Fig. \[construction\]. We supposed that the surface $\ell$ can be approximated, locally, as a paraboloid. Taking into account the right-hand side of Eq. (\[exph22\]), the interfacial free-energy functional   (\[ham1\]) can be written as $$\begin{aligned} \label{ham2} H[\ell] &\approx& \sigma \mathcal{A}_{lg} - \frac{\sigma}{2} \int_{\ell} d \mathbf{s}_1 \, \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \, \mathcal{K}(r_{\bot}) \left[\boldsymbol\nabla_{\bot}\Delta \ell(\mathbf{s},\mathbf{r}_{\bot})\right]^{2} \nonumber\\&+& \frac{\sigma}{2} \int d \mathbf{s}_{\ell} \, \int_{\mathbb{R}^{2}} d \mathbf{r}_{\bot} \mathcal{W}(r_{\bot}) \Delta \ell(\mathbf{s},\mathbf{r}_{\bot})^{2} ,\end{aligned}$$ where $\mathbf{r}_{\bot}$ is the projection of $\mathbf{s}_2-\mathbf{s}_1$ on the tangent plane $\pi_{\mathbf{s}_1}$ to the interface at $\mathbf{s}_1$ and $\Delta\ell$ is the vertical displacement from $\pi_{\mathbf{s}_1}$, $$\label{orthogonaldisplacement} \Delta \ell(\mathbf{s}_1,\mathbf{r}_{\bot}) = \mathbf{n}(\mathbf{s}_1) \mathbf{\cdot} (\mathbf{s}_2 - \mathbf{s}_1) .$$ The last step is to convert the surface integrations in integrals over the reference plane. In order to do that we need the mapping between the charts $\{\mathbf{x}_{1},\mathbf{x}_{2}\}$ and $\{\mathbf{s},\mathbf{r}_{\bot}\}$. The expressions of the mapping can be obtained from $\mathbf{s}_{2} - \mathbf{s}_{1} = \mathbf{n}(\mathbf{s}_{1}) \Delta \ell(\mathbf{s}_{1},\mathbf{r}_{\bot}) + \mathbf{r}_{\bot}$, supplemented by (\[orthogonaldisplacement\]) and $$\mathbf{n}(\mathbf{s}_{1})=\frac{1}{\sqrt{1+[\boldsymbol\nabla \ell(\mathbf{x}_{1})]^{2}}}\left(-\boldsymbol\nabla \ell (\mathbf{x}_{1}),1\right) ,$$ where $\boldsymbol\nabla$ represents the 2D gradient on the reference plane coordinates $$J = \biggl | \frac{\partial\left(\mathbf{s}_{1},\mathbf{r}_{\bot}\right)}{\partial\left(\mathbf{x}_{1},\mathbf{x}_{2}\right)} \biggr| .$$ We can show that the mapping Jacobian $J=1$ if quadratic terms on the gradients are neglected. In this limit $|\mathbf{r}_{\bot}|\approx |\mathbf{x}_{21}|$, the orthogonal displacement (\[orthogonaldisplacement\]) can be replaced with the vertical displacement $$\Delta\ell(\mathbf{s}_{1},\mathbf{r}_{\bot}) \simeq \delta\ell(\mathbf{x}_{1},\mathbf{x}_{2}) \equiv \ell(\mathbf{x}_{2})-\ell(\mathbf{x}_{1})-\mathbf{x}_{21} \mathbf{\cdot}\boldsymbol \nabla\ell(\mathbf{x}_{1}) ,$$ and taking the 2D gradient, $\boldsymbol\nabla_{\bot} \Delta\ell \simeq \boldsymbol\nabla\ell(\mathbf{x}_{1})- \boldsymbol\nabla\ell(\mathbf{x}_{1})$. Finally, $$\mathcal{A}_{lg} \simeq \mathcal{A}_{\pi} + \frac{1}{2} \int d \mathbf{x} \, \left[\boldsymbol\nabla\ell(\mathbf{x})\right]^{2} ,$$ where $\mathcal{A}_{\pi}$ is the area of the surface obtained from the projection of the surface $\ell$ onto the reference plane. The Hamiltonian (\[ham2\]) becomes $$\begin{aligned} \nonumber \label{ham3} &&H[\ell] \approx \sigma \mathcal{A}_{\pi} + \frac{\sigma}{2} \int d \mathbf{x} \, \left[\boldsymbol\nabla\ell(\mathbf{x})\right]^{2} \nonumber\\& - & \frac{\sigma}{2} \int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{K}(x_{12}) \bigl[ \boldsymbol\nabla\ell(\mathbf{x}_{1}) - \boldsymbol\nabla\ell(\mathbf{x}_{2} ) \bigr]^{2} \\ & + & \frac{\sigma}{2} \int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{W}(x_{12}) \bigl[ \ell(\mathbf{x}_{2})-\ell(\mathbf{x}_{1})-\mathbf{x}_{12}\mathbf{\cdot}\boldsymbol\nabla\ell(\mathbf{ x}_{1}) \bigr]^{2} . \nonumber\end{aligned}$$ The expressions in (\[ham3\]) can be further simplified. First we compute the squares, isolating the term proportional to the difference in vertical displacement. New terms will be created and for them we use the identities: $$\begin{aligned} &&\frac{1}{2}\int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{K}(x_{12}) \bigl[\boldsymbol \nabla\ell(\mathbf{x}_{1}) - \boldsymbol\nabla\ell(\mathbf{x}_{2} ) \bigr]^{2}\\ \nonumber &=& \int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{K}(x_{12}) \bigl\{ \left[\boldsymbol\nabla\ell(\mathbf{x}_{1})\right]^{2} - \boldsymbol\nabla\ell(\mathbf{x}_{1})\mathbf{\cdot}\boldsymbol\nabla\ell(\mathbf{x}_{2}] \bigr\} \\ \nonumber & = & \int d \mathbf{x} \, \left[\boldsymbol\nabla\ell(\mathbf{x})\right]^{2} - \int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{K}(x_{12}) \, \boldsymbol\nabla\ell(\mathbf{x}_{1})\mathbf{\cdot}\boldsymbol\nabla\ell(\mathbf{x}_{2}) \end{aligned}$$ and $$\begin{aligned} &&\int d \mathbf{x}_{1} d \mathbf{x}_{2} \, \mathcal{W}(x_{12}) \bigl[ \mathbf{x}_{12}\mathbf{\cdot}\boldsymbol{\nabla}\ell(\mathbf{ x}_{1}) \bigr]^{2} \\ \nonumber & = & \int d \mathbf{x}_{1} d \mathbf{x}_{12} \, \mathcal{W}(x_{12}) \Bigl\{ x_{21}^{2} \left[ \partial_{x_{1}}\ell(\mathbf{x}_{1}) \right]^{2} + y_{21}^{2} \left[ \partial_{y_{1}}\ell(\mathbf{x}_{1}) \right]^{2} \Bigr\} \\ \nonumber & = & \frac{1}{2} \int d \mathbf{x}_{12} \, \mathbf{x}_{12}^{2} \, \mathcal{W}(x_{12}) \int d \mathbf{x}_{1} \left[\boldsymbol\nabla\ell(\mathbf{x}_{1})\right]^{2} \\ & = & \int d \mathbf{x} \left[\boldsymbol\nabla\ell(\mathbf{x})\right]^{2} .\end{aligned}$$ There is also a term of the form $$\begin{aligned} &&\int d \mathbf{x}_{1} d \mathbf{x}_{2} \biggl\{\mathcal{K}(x_{12}) \boldsymbol\nabla\ell(\mathbf{x}_{1}) \mathbf{\cdot}\boldsymbol \nabla\ell(\mathbf{x}_{2}) \nonumber \\ &&- \mathcal{W}(x_{12}) \bigl[\ell(\mathbf{x}_{2})-\ell(\mathbf{x}_{1})\bigr] \bigl[\mathbf{x}_{21} \mathbf{\cdot}\boldsymbol \nabla\ell(\mathbf{x}_{1})\bigr]\biggr\} .\end{aligned}$$ Grouping the integral over $\mathbf{x}_{1}$ we have $$\begin{aligned} &&\int d \mathbf{x}_{1} \boldsymbol\nabla\ell(\mathbf{x}_{1}) \mathbf{\cdot} \int d \mathbf{x}_{2} \biggl\{\mathcal{K}(x_{12}) \boldsymbol\nabla\ell(\mathbf{x}_{2}) \nonumber\\ &&- \mathcal{W}(x_{12}) \bigl[\ell(\mathbf{x}_{2})-\ell(\mathbf{x}_{1})\bigr] \mathbf{x}_{21} \biggr\} ,\end{aligned}$$ but since $- \mathcal{W}(x_{12}) \mathbf{x}_{21} = \boldsymbol\nabla_{\mathbf{x}_{2}}\mathcal{K}(x_{12})$, the second integrand can be written as a gradient of a scalar function $$\int d \mathbf{x}_{1} \boldsymbol\nabla\ell(\mathbf{x}_{1}) \mathbf{\cdot} \int d \mathbf{x}_{2} \boldsymbol\nabla_{\mathbf{x}_{2}}\biggl\{\bigl[\ell(\mathbf{x}_{2})-\ell(\mathbf{x}_{1})\bigr]\mathcal{K}(x_{12})\biggr\}$$ and so it reduces to a boundary contribution, which we can neglect. Collecting all the remaining terms, we are left with (\[ham4\]). Wetting diagrams {#appendix_C} ================ In this appendix we collect the definitions for the various wetting diagrams used in the main text. The diagrams $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram1.eps,width=1.5cm}}} & = & \int \textrm{d}\mathbf{s}, \\ {\parbox{1.5cm}{\epsfig{file=diagram2.eps,width=1.5cm}}} & = & \int \textrm{d}\mathbf{s} \, H(\mathbf{s}) /\kappa, \\ {\parbox{1.5cm}{\epsfig{file=diagram3.eps,width=1.5cm}}} & = & \int \textrm{d}\mathbf{s} \, H^{2}(\mathbf{s}) /\kappa^{2}, \\ {\parbox{1.5cm}{\epsfig{file=diagram4.eps,width=1.5cm}}} & = & \int \textrm{d}\mathbf{s} \, K_{G}(\mathbf{s}) /\kappa^{2}\end{aligned}$$ involve only local interfacial properties. The circle represents the area element, while $H$ denotes the local mean curvature and $K_{G}(\mathbf{s})$ the Gaussian curvature. Open symbols such as the one appearing in (\[dashedk\]) and (\[dashedk2\]) stand for the evaluation of the corresponding weight functions at a specified point $\mathbf{s}$ on the surface. The Ornstein-Zernike kernel of (\[OZ2\]) is represented by a thick black line with two open circles at the extrema. For instance, if $\mathbf{s}$ belongs to the surface $\ell$ and $\mathbf{r}$ to the upper region we will write $${\parbox{1.5cm}{\epsfig{file=diagram21.eps,width=1.5cm}}} = \int_{\ell}\textrm{d}\mathbf{s} \, K(\mathbf{s},\mathbf{r}) \, ,$$ and similarly $${\parbox{1.5cm}{\epsfig{file=diagram24.eps,width=1.5cm}}} = \kappa^{-2} \int_{\ell}\textrm{d}\mathbf{s} \, K_{G}(\mathbf{s}) \, K(\mathbf{s},\mathbf{r}) \, .$$ Then we have the dashed and arrow diagrams $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram5.eps,width=1.5cm}}} & = & {\parbox{1.5cm}{\epsfig{file=diagram124.eps,width=1.5cm}}}-\delta(\mathbf{s}-\mathbf{s}') = U(\mathbf{s},\mathbf{s}') \, , \\ {\parbox{1.5cm}{\epsfig{file=diagram6.eps,width=1.5cm}}} & = & \frac{1}{\kappa}\partial_{n} K(\mathbf{s},\mathbf{s}') \, ,\end{aligned}$$ where in the latter diagram the arrow points to the position where the normal derivative is taken, as in the example $$\int_\ell \textrm{d}\mathbf{s}_1 \textrm{d}\mathbf{s}_2 U(\mathbf{s},\mathbf{s}_1) \frac{1}{\kappa}\partial_{n_1} K(\mathbf{s}_1,\mathbf{s}_2) \, ={\parbox{1.5cm}{\epsfig{file=diagram7.eps,width=1.5cm}}} \, .$$ The arrow diagram can also span between two interfaces, for example $${\parbox{1.5cm}{\epsfig{file=diagram80.eps,width=1.5cm}}} = \int_{\psi}\textrm{d}\mathbf{s}_{\psi}\int_{\ell}\textrm{d}\mathbf{s}_{\ell} \, \frac{1}{\kappa}\partial_{n_\ell} K(\mathbf{s}_\ell,\mathbf{s}_\psi) \, .$$ The algebraic expressions for all other diagrams can be reconstructed in terms of these elementary building blocks. Binding potential for planar, spherical and cylindrical interfacial configurations. \[appendixe\] ================================================================================================= In this Appendix we will review the known form for planar, spherical and cylindrical interfacial configurations and how they are reproduced from the nonlocal representations of the binding potential we have discussed in Sec. V. Planar interfaces ----------------- We first consider the simplest case of a planar wall ($\psi=0$) and a planar interface of constant thickness, $\ell(\mathbf{x})=\ell$. In this case, $\mathcal{F}_{wl}[\psi]=\sigma_{wl} \mathcal{A}_{wl}$ and $H[\ell]=\sigma \mathcal{A}_{lg}$, where $\mathcal{A}_{lg}=\mathcal{A}_{wl}=\mathcal{A}$ is the interfacial area and $\sigma_{wl}=(\kappa/2)(h_1+gm_b)^2/g(g-\kappa)$ and $\sigma=\kappa m_0^2$ are the surface tensions defined for the planar wall-liquid and liquid-gas interfaces, respectively. On the other hand, the binding potential is [@laura; @parry2; @parry3] $$\begin{aligned} &&\frac{W[\ell,\psi]}{\mathcal{A}}=2\kappa m_0 \left(\frac{h_1+gm_0}{\kappa-g}\right) \frac{e^{-\kappa \ell}}{1- \frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}} \nonumber\\ &&+ \kappa \left(\frac{h_1+gm_0}{\kappa-g}\right)^2 \frac{e^{-2\kappa\ell}}{1- \frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}}\nonumber\\&& + \frac{g+\kappa}{g-\kappa} \kappa m_0^2 \frac{e^{-2 \kappa \ell}} {1-\frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}}. \label{bindingplanar}\end{aligned}$$ The basic diagrams to obtain the decorated version of the original nonlocal model are $${\parbox{1.5cm}{\epsfig{file=diagram125.eps,width=1.5cm}}}={\parbox{1.5cm}{\epsfig{file=diagram126.eps,width=1.5cm}}}=e^{-\kappa \ell}$$ and $${\parbox{1.5cm}{\epsfig{file=diagram127.eps,width=1.5cm}}}={\parbox{1.5cm}{\epsfig{file=diagram128.eps,width=1.5cm}}}={\parbox{1.5cm}{\epsfig{file=diagram129.eps,width=1.5cm}}}={\parbox{1.5cm}{\epsfig{file=diagram130.eps,width=1.5cm}}}=0. \label{flatukbonds}$$ We note that these diagrams do not depend on the position associated with the open circle, so any diagram can be split into the contribution of its bonds. For example, $${\parbox{1.5cm}{\epsfig{file=diagram131.eps,width=1.5cm}}}=\left({\parbox{1.5cm}{\epsfig{file=diagram126.eps,width=1.5cm}}}\right)\times\left({\parbox{1.5cm}{\epsfig{file=diagram125.eps,width=1.5cm}}}\right)\times \left({\parbox{1.5cm}{\epsfig{file=diagram132.eps,width=1.5cm}}}\right),$$ with $${\parbox{1.5cm}{\epsfig{file=diagram133.eps,width=1.5cm}}}={\parbox{1.5cm}{\epsfig{file=diagram132.eps,width=1.5cm}}}=\mathcal{A}.$$ Due to the expression (\[flatukbonds\]), the nonvanishing diagrams are those of the original nonlocal model. In particular, Eqs. (\[omega11\]), (\[omega12\]) and (\[omega21\]) reduce, respectively, to $$\begin{aligned} \Omega_1^1 &=& \frac{2g}{g-\kappa} {\parbox{1.5cm}{\epsfig{file=diagram134.eps,width=1.5cm}}} = \frac{2g}{g-\kappa} \mathcal{A} e^{-\kappa \ell},\\ \Omega_1^2 &=& \frac{g+\kappa}{g-\kappa} {\parbox{1.5cm}{\epsfig{file=diagram131.eps,width=1.5cm}}} = \frac{g+\kappa}{g-\kappa} \mathcal{A} e^{-2\kappa \ell},\\ \Omega_2^1 &=& \left(\frac{g}{g-\kappa}\right)^2 {\parbox{1.5cm}{\epsfig{file=diagram135.eps,width=1.5cm}}} = \left(\frac{g}{g-\kappa}\right)^2 \mathcal{A} e^{-2\kappa \ell},\end{aligned}$$ which are consistent with the expressions in Refs. [@laura; @parry3], although our notation differs slightly from that used in these references. The higher-order terms in the functional can also be easily evaluated: $$\begin{aligned} \Omega_n^n &=& \Omega_1^1 \left(\frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}\right)^{n-1},\\ \Omega_n^{n+1} &=& \Omega_1^2 \left(\frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}\right)^{n-1},\\ \Omega_{n+1}^n &=& \Omega_2^1 \left(\frac{g+\kappa}{g-\kappa}e^{-2\kappa \ell}\right)^{n-1}. \end{aligned}$$ If we substitute these expressions into Eq. (\[Hmin6\]), we reobtain Eq. (\[bindingplanar\]) after a trivial resummation. Finally, the expressions obtained in Ref. [@parry2] for fixed boundary conditions on the wall are reobtained by taking the limit $g\to -\infty$ and $-h_1/g-m_0\to \delta m_1\equiv m_1-m_0$ in our equations, where $m_1$ is the order parameter on the wall. Now we turn to the formulation for the nonlocal model we have introduced in this paper for fixed boundary conditions on the wall. The basic diagrams for this formulation are $${\parbox{1.5cm}{\epsfig{file=diagram136.eps,width=1.5cm}}}=-{\parbox{1.5cm}{\epsfig{file=diagram137.eps,width=1.5cm}}}=e^{-\kappa \ell}.$$ The expressions for $\Omega_1^1$, $\Omega_1^2$, and $\Omega_2^1$ are obtained from Eqs. (\[omega11-2\]), (\[omega12-2\]) and (\[omega21-2\]), respectively, as $$\begin{aligned} \Omega_1^1 &=& {\parbox{1.5cm}{\epsfig{file=diagram138.eps,width=1.5cm}}}-{\parbox{1.5cm}{\epsfig{file=diagram139.eps,width=1.5cm}}} = 2\mathcal{A} e^{-\kappa \ell},\\ \Omega_1^2 &=& - {\parbox{1.5cm}{\epsfig{file=diagram140.eps,width=1.5cm}}} = \mathcal{A} e^{-2\kappa \ell},\\ \Omega_2^1 &=& - {\parbox{1.5cm}{\epsfig{file=diagram141.eps,width=1.5cm}}} = \mathcal{A} e^{-2\kappa \ell},\end{aligned}$$ and for higher-order contributions we get that $$\begin{aligned} \Omega_n^n &=& \Omega_1^1 e^{-2(n-1)\kappa \ell},\\ \Omega_n^{n+1} &=& \Omega_1^2 e^{-2(n-1)\kappa \ell},\\ \Omega_{n+1}^n &=& \Omega_2^1 e^{-2(n-1)\kappa \ell}.\end{aligned}$$ Substitution of these expressions into Eq. (\[Hmin7\]) leads to the expression $$\begin{aligned} &&\frac{W[\ell,\psi]}{\mathcal{A}}=\sum_{n=1}^\infty \Bigg\{2 \kappa m_0\delta m_1 e^{-(2n-1)\kappa \ell}\nonumber\\ && +\left[\kappa (\delta m_1)^2 + \kappa (m_0)^2\right] e^{-2 n \kappa \ell} \Bigg\} \label{Hmin7-2}\end{aligned}$$ which can be resummed as $$\begin{aligned} &&\frac{W[\ell,\psi]}{\mathcal{A}}=\frac{2 \kappa m_0\delta m_1 e^{-\kappa \ell}}{1-e^{-2\kappa \ell}}\nonumber\\ && +\left[\kappa (\delta m_1)^2 + \kappa (m_0)^2\right] \frac{e^{-2\kappa \ell}}{1-e^{-2\kappa \ell}}, \label{bindingplanar2}\end{aligned}$$ which is Eq. (\[bindingplanar\]) in the limit of fixed boundary conditions on the wall. Spherical interfaces -------------------- A similar calculation can be performed for the problem of wetting around a sphere. We suppose that the sphere is of radius $R$ and consider an interfacial configuration corresponding to a concentric sphere of radius $R+\ell$. In this case, $\mathcal{F}_{wl}[\psi]$ is given by [@laura] $$\mathcal{F}_{wl}[\psi]=\sigma_{wl} \mathcal{A}_{wl} \left(\frac{1+\frac{1}{\kappa R}}{1+\frac{1}{(\kappa - g)R}}\right), \label{fwlsphere}$$ where $\sigma_{wl}$ is the surface tension for the planar wall-liquid interface and the area of the sphere is $ \mathcal{A}_{wl}=4\pi R^2$. Note that the mean curvature and the Gaussian curvature on the sphere are $H=-1/R$ and $K_G=1/R^2$, respectively. Thus Eq. (\[fwlsphere\]) satisfies Eq. (\[deltafwlexpansion\]) for large $R$. It is instructive to reobtain Eq. (\[fwlsphere\]) from the diagrammatic expansion Eq. (\[F\_chain\_1\]). The relevant diagrams for this calculation are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram142.eps,width=1.5cm}}}&=&-e^{-2\kappa R},\\ {\parbox{1.5cm}{\epsfig{file=diagram143.eps,width=1.5cm}}}&=&-\frac{1}{\kappa R}\left[1-(1+\kappa R)e^{-2\kappa R}\right], \\ {\parbox{1.5cm}{\epsfig{file=diagram144.eps,width=1.5cm}}}&=&4\pi R^2,\end{aligned}$$ which are independent of the position of the open circle, as in the planar case. So, again, each diagram is just the product of its bonds. In order to sum the contributions of the diagrams, we note that each diagram is a chainlike sequence of $U$ and $\partial K/\kappa$ bonds, with coefficients given by Eq. (\[factor\]). We first sum the diagrams without $\partial K/\kappa$ bonds. Their total contribution $S_0$ to $\mathcal{F}_{wl}$ is $$\begin{aligned} S_0&=&\sigma_{wl}\mathcal{A}_{wl}\sum_{n=0}^\infty \left[\frac{g}{\kappa - g}\left(-e^{-2\kappa R}\right)\right]^n \nonumber\\ &=&\sigma_{wl}\mathcal{A}_{wl}\frac{1}{1+\frac{g}{\kappa - g}e^{-2\kappa R}}.\end{aligned}$$ Now we consider the diagrams with only one $\partial K/\kappa$ bond. Their total contribution $S_1$ to $\mathcal{F}_{wl}$ can be written as $$\begin{aligned} S_1&=&\sigma_{wl}\mathcal{A}_{wl} \left(\sum_{n_1=0}^\infty \left[\frac{g}{\kappa - g} \left[-e^{-2\kappa R}\right)\right]^{n_1}\right]\nonumber\\ &\times& \left(\frac{\kappa}{\kappa-g}\right)\left[-\frac{1}{\kappa R}\left(1-(1+\kappa R)e^{-2\kappa R}\right)\right] \nonumber\\ &\times& \left(\frac{g}{\kappa}+\sum_{n_2=1}^\infty \left[\frac{g}{\kappa - g} \left(-e^{-2\kappa R}\right)\right]^{n_2}\right),\end{aligned}$$ which can be written as $$\begin{aligned} S_1&=& \sigma_{wl}\mathcal{A}_{wl} \frac{g}{\kappa}\frac{1-e^{-2\kappa R}}{1+\frac{g}{\kappa - g}e^{-2\kappa R}} \nonumber\\ &\times& \frac{-\frac{1}{(\kappa-g)R}\left[ 1-(1+\kappa R)e^{-2\kappa R}\right]}{1+\frac{g}{\kappa - g}e^{-2\kappa R}}.\end{aligned}$$ For diagrams with $m>1$, $\partial K/\kappa$ bonds, their contribution $S_n$ to $\mathcal{F}_{wl}$ can be obtained similarly as $$\begin{aligned} S_m&=& \sigma_{wl}\mathcal{A}_{wl} \frac{g}{\kappa}\frac{1-e^{-2\kappa R}}{1+\frac{g}{\kappa - g}e^{-2\kappa R}} \nonumber\\ &\times& \left(\frac{-\frac{1}{(\kappa-g)R}\left( 1-(1+\kappa R)e^{-2\kappa R}\right)}{1+\frac{g}{\kappa - g}e^{-2\kappa R}}\right)^m .\end{aligned}$$ So, $\mathcal{F}_{wl}[\psi]$ has the expression $$\begin{aligned} \mathcal{F}_{wl}[\psi]&=&\sum_{m=0}^{\infty} S_m = \frac{\sigma_{wl}\mathcal{A}_{wl}}{1+\frac{g}{\kappa-g}e^{-2\kappa R}} \nonumber\\ &\times& \Bigg[1+\frac{g}{\kappa}\left(1-e^{-2\kappa R}\right) \nonumber\\ &\times& \frac{\frac{-1}{(\kappa-g)R}\left(\frac{1-(1+\kappa R)e^{-2\kappa R}}{1+\frac{g}{\kappa-g}e^{-2\kappa R}}\right)} {1+\frac{1}{(\kappa-g)R}\left(\frac{1-(1+\kappa R)e^{-2\kappa R}}{1+\frac{g}{\kappa-g}e^{-2\kappa R}}\right)}\Bigg],\end{aligned}$$ which after some algebra reduces to Eq. (\[fwlsphere\]). Similarly, $H[\ell]$ has the expression $$H[\ell]=\frac{\sigma\mathcal{A}_{lg}}{1-e^{-2\kappa (R+\ell)}}, \label{hsphere}$$ where $\sigma$ is the surface tension for the planar liquid-gas interface, with area $ \mathcal{A}_{lg}=4\pi (R+\ell)^2$. The diagrammatic expansion  (\[F\_chain\]) can be evaluated explicitly [@PR], where now the basic diagrams are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram142.eps,width=1.5cm}}}&=&-e^{-2\kappa (R+\ell)},\\ {\parbox{1.5cm}{\epsfig{file=diagram144.eps,width=1.5cm}}}&=&4\pi (R+\ell)^2,\end{aligned}$$ leading, after resummation, to Eq. (\[hsphere\]). We turn to the evaluation of $W[\ell,\psi]$, which has the expression $$\begin{aligned} &&W[\ell,\psi]=2\kappa m_0 \left(-\frac{h_1+gm_0}{g-\kappa-\frac{1}{R}}\right) \frac{\sqrt{\mathcal{A}_{wl}\mathcal{A}_{lg}} e^{-\kappa \ell}}{1- \frac{g+\kappa-\frac{1}{R}}{g-\kappa-\frac{1}{R}}e^{-2\kappa \ell}} \nonumber\\ && + \frac{g+\kappa-\frac{1}{R}}{g-\kappa-\frac{1}{R}} \left(\frac{\kappa m_0^2}{1-e^{-2\kappa (R+\ell)}} \right) \frac{\mathcal{A}_{lg}e^{-2 \kappa \ell}} {1-\frac{g+\kappa-\frac{1}{R}}{g-\kappa-\frac{1}{R}}e^{-2\kappa \ell}} \nonumber\\ &&+ \kappa \left(\frac{h_1+gm_0}{g-\kappa-\frac{1}{R}} \right)^2 \frac{\mathcal{A}_{wl}e^{-2\kappa\ell}}{1- \frac{g+\kappa-\frac{1}{R}}{g-\kappa-\frac{1}{R}}e^{-2\kappa \ell}} . \label{bindingspherical}\end{aligned}$$ This expression reduces to Eq. (\[bindingplanar\]) for $R\to \infty$, and it is consistent with those reported in Refs. [@laura; @parry2] if the exponential term $\exp[-2\kappa(R+\ell)]$ in Eq. (\[bindingspherical\]) is neglected. As in the planar case, we will reproduce this result within the nonlocal model. We will first consider the decorated version of the original nonlocal model. In this formalism, the basic diagrams for this model are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram145.eps,width=1.5cm}}}&=&4\pi (R+\ell)^2,\\ {\parbox{1.5cm}{\epsfig{file=diagram146.eps,width=1.5cm}}}&=&4\pi R^2,\\ {\parbox{1.5cm}{\epsfig{file=diagram147.eps,width=1.5cm}}}&=&\left(1+\frac{\ell}{R}\right)\left(1-e^{-2\kappa R}\right)e^{-\kappa\ell},\\ {\parbox{1.5cm}{\epsfig{file=diagram148.eps,width=1.5cm}}}&=&\frac{1-e^{-2\kappa R}}{1+\frac{\ell}{R}}e^{-\kappa\ell}\\ {\parbox{1.5cm}{\epsfig{file=diagram149.eps,width=1.5cm}}}&=&-e^{-2\kappa R},\\ {\parbox{1.5cm}{\epsfig{file=diagram150.eps,width=1.5cm}}}&=&-e^{-2\kappa (R+\ell)},\\ {\parbox{1.5cm}{\epsfig{file=diagram151.eps,width=1.5cm}}}&=&-\frac{1}{\kappa R}\left[1-(1+\kappa R)e^{-2\kappa R}\right].\end{aligned}$$ Note that, as for the planar case, the diagrams do not depend on the position associated with the open circle, so a general diagram can be obtained as a product of its bond contributions. In order to resum all the contributions to $\Omega_n^n$, $\Omega_{n+1}^n$, and $\Omega_{n}^{n+1}$, we have to sum the contributions of all possible segments either on the wall or on the liquid-gas interface in a similar manner as we did for the evaluation of $\mathcal{F}_{wl}$. The resummation of all the contributions of the segments of consecutive ($U$) bonds on the liquid-gas interface is $$I_\ell=\frac{1}{1-e^{-2\kappa (R+\ell)}}. \label{il}$$ On the other hand, the analogous expression for a segment of consecutive bonds on the wall depends on its position in the diagram. If the segment is on an extreme of the full diagram, its contribution is $$I_\psi^1=\frac{g}{g-\kappa-\frac{1}{R}} \frac{1}{1-e^{-2\kappa R}}. \label{ipsi1}$$ Otherwise, the contribution of the segment is $$I_\psi^2=\left(\frac{1}{1-e^{-2\kappa R}}\right)\left(1+\frac{2\kappa} {g-k-\frac{1}{R}}\frac{1}{1-e^{-2\kappa R}}\right).$$ With these results, Eqs. (\[omega11\]), (\[omega12\]) and (\[omega21\]) reduce, respectively, to $$\begin{aligned} &&\Omega_1^1 = 8\pi R(R+\ell)\left(\frac{g}{g-k-\frac{1}{R}}\right) \frac{e^{-\kappa \ell}}{1-e^{-2\kappa(R+\ell)}},\label{omega11sphere}\\ &&\Omega_1^2 = 4\pi (R+\ell)^2 \left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right)\nonumber\\ &&\times\left(1+\frac{2\kappa} {g-k-\frac{1}{R}}\frac{1}{1-e^{-2\kappa R}}\right)\frac{e^{-2 \kappa \ell}}{1-e^{-2\kappa(R+\ell)}},\label{omega12sphere}\\ &&\Omega_2^1 = 4 \pi R^2 \left(\frac{g}{g-k-\frac{1}{R}}\right)^2 \frac{e^{-2\kappa \ell}}{1-e^{-2\kappa(R+\ell)}}. \label{omega21sphere}\end{aligned}$$ For higher-order contributions $$\begin{aligned} &&\Omega_n^n = \Omega_1^1 \Bigg[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) ,\label{omegannsphere}\nonumber\\&&\times\left(1+\frac{2\kappa} {g-k-\frac{1}{R}}\frac{1}{1-e^{-2\kappa R}}\right)e^{-2\kappa \ell}\Bigg]^{n-1}\\ &&\Omega_n^{n+1} = \Omega_1^2 \Bigg[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) ,\label{omegann1sphere}\nonumber\\&&\times\left(1+\frac{2\kappa} {g-k-\frac{1}{R}}\frac{1}{1-e^{-2\kappa R}}\right)e^{-2\kappa \ell}\Bigg]^{n-1}\\ &&\Omega_{n+1}^n = \Omega_2^1 \Bigg[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) \nonumber\\&&\times\left(1+\frac{2\kappa} {g-k-\frac{1}{R}}\frac{1}{1-e^{-2\kappa R}}\right)e^{-2\kappa \ell}\Bigg]^{n-1} .\label{omegan1nsphere}\end{aligned}$$ As in the planar case, the resummation of the series  (\[Hmin7\]) leads to Eq. (\[bindingspherical\]). In order to check our formulation for the nonlocal model for fixed boundary conditions on the wall, we will make use of the diagrams $$\begin{aligned} &&{\parbox{1.5cm}{\epsfig{file=diagram152.eps,width=1.5cm}}}=\left(1+\frac{\ell}{R}\right)e^{-\kappa \ell}\nonumber\\ &&\times \left(2 -\left(1 + \frac{1}{\kappa R}\right)\left(1-e^{-2\kappa R}\right)\right), \\ &&{\parbox{1.5cm}{\epsfig{file=diagram153.eps,width=1.5cm}}}=-\frac{e^{-\kappa \ell}}{1+\frac{\ell}{R}}\nonumber\\ &&\times \left(\left(1 + \frac{1}{\kappa(R+\ell)}\right)\left(1-e^{-2\kappa R}\right)\right), \\ &&{\parbox{1.5cm}{\epsfig{file=diagram154.eps,width=1.5cm}}} = -\frac{1}{\kappa (R+\ell)}\nonumber\\ &&\times\left\{1-[1+\kappa (R+\ell)]e^{-2\kappa (R+\ell)}\right\}.\end{aligned}$$ The total contribution of the segments of the diagrams on the liquid-gas interface now depends on its positions. The leftmost one is composed of $U$ bonds and it has a contribution $I_\ell$ given by Eq. (\[il\]). Otherwise, the segments are composed of $\partial K/\kappa$ bonds, with a contribution $$I^\prime_\ell= \frac{1}{\left(1 + \frac{1}{\kappa(R+\ell)}\right)\left(1-e^{-2\kappa (R+\ell)}\right)}.$$ Similarly, the segments of the diagrams on the wall on the left extreme of the diagram contribute as $$(I^{\prime})^1_\psi= \frac{1}{1-e^{-2\kappa R}},$$ which is the limit of Eq. (\[ipsi1\]) when $g\to -\infty$, and otherwise as $$(I^{\prime})^2_\psi= \frac{1}{2-\left(1 + \frac{1}{\kappa R}\right)\left(1-e^{-2\kappa R}\right)}.$$ Thus, the expressions from Eqs. (\[omega11-2\]), (\[omega12-2\]), and (\[omega21-2\]) for $\Omega_1^1$, $\Omega_1^2$, and $\Omega_2^1$ can be resummed as $$\begin{aligned} \Omega_1^1 &=& 8\pi R(R+\ell) \frac{e^{-\kappa \ell}}{1-e^{-2\kappa(R+\ell)}},\\ \Omega_1^2 &=& 4\pi (R+\ell)^2 \left(\frac{1-e^{-2\kappa R}}{(1-e^{-2\kappa(R+\ell)})^2}\right) e^{-2 \kappa \ell},\\ \Omega_2^1 &=& 4 \pi R^2 \frac{e^{-2\kappa \ell}}{1-e^{-2\kappa(R+\ell)}},\end{aligned}$$ and for higher-order contributions we get that $$\begin{aligned} &&\Omega_n^n = \Omega_1^1 \left[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) e^{-2\kappa \ell}\right]^{n-1},\\ &&\Omega_n^{n+1} = \Omega_1^2 \left[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) e^{-2\kappa \ell}\right]^{n-1},\\ &&\Omega_{n+1}^n = \Omega_2^1 \left[\left(\frac{1-e^{-2\kappa R}}{1-e^{-2\kappa(R+\ell)}}\right) e^{-2\kappa \ell}\right]^{n-1},\end{aligned}$$ which coincide with the expressions Eqs. (\[omega11sphere\])-(\[omegan1nsphere\]) in the limit $g\to -\infty$. Cylindrical interfaces ---------------------- Finally, we will consider the problem of wetting around a cylinder of radius $R$ and length $L$ (large enough to neglect border effects), where the liquid-gas interfacial configuration is a concentric cylinder of radius $R+\ell$. In this case, $\mathcal{F}_{wl}[\psi]$ is given by $$\mathcal{F}_{wl}[\psi]=\sigma_{wl} \mathcal{A}_{wl} \frac{\frac{K_1(\kappa R)}{K_0(\kappa R)}}{1-\frac{\kappa}{\kappa-g} \left(1-\frac{K_1(\kappa R)}{K_0(\kappa R)}\right)} , \label{fwlcylinder}$$ where $\sigma_{wl}$ is the surface tension for the planar wall-liquid interface, $\mathcal{A}_{wl}=2\pi R L$ is the area of the cylinder, and $K_0$ and $K_1$ are the modified Bessel functions of the second kind and order 0 and 1, respectively. For large $\kappa R$, this expression can be approximated as $$\begin{aligned} &&\mathcal{F}_{wl}[\psi]=\sigma_{wl} \mathcal{A}_{wl}\Bigg[1-\left(\frac{g}{\kappa-g}\right)\frac{1}{2\kappa R}\nonumber\\ && + \left(\frac{g}{\kappa-g}\right)\left(1+\frac{2\kappa}{\kappa-g}\right)\frac{1}{(2\kappa R)^2}+{O}(R^{-3})\Bigg] ,\end{aligned}$$ which satisfies Eq. (\[deltafwlexpansion\]) since the mean curvature and the Gaussian curvature on the cylinder are $H=-1/2R$ and $K_G=0$, respectively. Equation (\[fwlcylinder\]) can be obtained in a similar way as in the spherical case from the diagrammatic expansion  (\[F\_chain\_1\]), where the relevant diagrams are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram142.eps,width=1.5cm}}}&=&2\kappa R I_0(\kappa R) K_0(\kappa R)-1,\\ {\parbox{1.5cm}{\epsfig{file=diagram143.eps,width=1.5cm}}}&=&2 \kappa R I_1(\kappa R)K_0(\kappa R) \nonumber\\ &=&1-2\kappa R I_0(\kappa R)K_1(\kappa R), \\ {\parbox{1.5cm}{\epsfig{file=diagram144.eps,width=1.5cm}}}&=&2\pi R L,\end{aligned}$$ where $I_0$ and $I_1$ are the modified Bessel function of the first kind and order 0 and 1, respectively. Note that as in the planar and spherical cases, they are independent of the position of the open circle. Thus, as in these previous cases, the contribution of each diagram is the product of its bonds. Similarly, $H[\ell]$ has the expression $$H[\ell]=\frac{\sigma\mathcal{A}_{lg}}{2\kappa R K_0[\kappa (R+\ell)]I_0[\kappa (R+\ell)]}, \label{hcylinder}$$ where $\sigma$ is the surface tension for the planar liquid-gas interface, with area $ \mathcal{A}_{lg}=2\pi (R+\ell)L$. For large $\kappa R$, Eq. (\[hcylinder\]) yields $$H[\ell]\approx \sigma\mathcal{A}_{lg}\left(1-\frac{1}{8(\kappa R)^2}\right),$$ in agreement with Eq. (\[Helfrich\]). The diagrammatic expansion Eq. (\[F\_chain\]) can be also evaluated explicitly in this case, where now the basic diagrams are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram142.eps,width=1.5cm}}}&=&2\kappa (R+\ell) I_0[\kappa (R+\ell)]K_0[\kappa (R+\ell)]-1,\nonumber\\ \\ {\parbox{1.5cm}{\epsfig{file=diagram144.eps,width=1.5cm}}}&=&2\pi (R+\ell)L.\end{aligned}$$ After resummation of Eq. (\[F\_chain\]), we recover Eq. (\[hcylinder\]). Finally, the binding potential $W[\ell,\psi]$ has the expression $$\begin{aligned} &&W[\ell,\psi]= \left(\frac{\pi L}{\kappa\Delta}\right) \Bigg[2\kappa m_0 \left(-\frac{h_1+gm_0}{g}\right) \nonumber\\ &&+ \kappa \left(\frac{h_1+gm_0}{g} \right)^2 \left(\frac{1}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) \left(\frac{K_0[\kappa (R+\ell)]} {K_0(\kappa R)}\right) \nonumber\\ && + \kappa m_0^2 \frac{I_0(\kappa R)}{I_0[\kappa (R+\ell)]}\left(1+\frac{\kappa}{g} \frac{I_1(\kappa R)}{I_0(\kappa R)}\right) \Bigg] \label{bindingcylindrical}\end{aligned}$$ where $\Delta$ is defined as $$\begin{aligned} \Delta&=&I_0[\kappa (R+\ell)]K_0(\kappa R)\left(1-\frac{\kappa}{g} \frac{K_1(\kappa R)}{K_0(\kappa R)}\right)\nonumber\\ &-&K_0[\kappa (R+\ell)]I_0(\kappa R) \left(1+\frac{\kappa}{g} \frac{I_1(\kappa R)}{I_0(\kappa R)}\right).\end{aligned}$$ The modified Bessel functions can be approximated asymptotically for large values of their arguments as $$\begin{aligned} &&K_0(x)\sim K_1(x) \sim \sqrt{\frac{\pi}{2x}}e^{-x}\left(1+{O} \left(\frac{1}{x}\right)\right),\nonumber\\ &&I_0(x)\sim I_1(x)\sim \sqrt{\frac{1}{2\pi x}}e^x\left(1+{O}\left(\frac{1}{x}\right)\right).\end{aligned}$$ Thus, Eq. (\[bindingcylindrical\]) reduces, for large $\kappa R$, to $$\begin{aligned} &&W[\ell,\psi]=2\kappa m_0 \left(-\frac{h_1+gm_0}{g-\kappa}\right) \frac{\sqrt{\mathcal{A}_{wl}\mathcal{A}_{lg}} e^{-\kappa\ell}}{1-\frac{g+\kappa}{g-\kappa} e^{-2\kappa \ell}} \nonumber\\ && + \frac{g+\kappa}{g-\kappa} \kappa m_0^2 \frac{\mathcal{A}_{lg} e^{-2\kappa\ell}}{1-\frac{g+\kappa}{g-\kappa} e^{-2\kappa \ell}} \nonumber\\ &&+ \kappa \left(\frac{h_1+gm_0}{g-\kappa} \right)^2 \frac{\mathcal{A}_{wl} e^{-2\kappa\ell}}{1-\frac{g+\kappa}{g-\kappa} e^{-2\kappa \ell}} \label{bindingcylindrical2}\end{aligned}$$ up to corrections of order $(\kappa R)^{-1}$ and $[\kappa(R+\ell)]^{-1}$. This expression is consistent with the expression reported in Ref. [@laura] and it reduces to Eq. (\[bindingplanar\]) for $R\to \infty$. This result can be also obtained within the nonlocal model. We will first consider the decorated version of the original nonlocal model. In this formalism, the basic diagrams model are $$\begin{aligned} {\parbox{1.5cm}{\epsfig{file=diagram145.eps,width=1.5cm}}}&=&2\pi (R+\ell) L,\\ {\parbox{1.5cm}{\epsfig{file=diagram146.eps,width=1.5cm}}}&=&2\pi R L,\\ {\parbox{1.5cm}{\epsfig{file=diagram147.eps,width=1.5cm}}}&=&2\kappa (R+\ell) I_0(\kappa R)K_0[\kappa (R+\ell)],\\ {\parbox{1.5cm}{\epsfig{file=diagram148.eps,width=1.5cm}}}&=&2\kappa R I_0(\kappa R)K_0[\kappa (R+\ell)],\\ {\parbox{1.5cm}{\epsfig{file=diagram149.eps,width=1.5cm}}}&=&2\kappa R I_0(\kappa R)K_0(\kappa R),\\ {\parbox{1.5cm}{\epsfig{file=diagram150.eps,width=1.5cm}}}&=&2\kappa (R+\ell) I_0[\kappa(R+\ell)]K_0[\kappa (R+\ell)], \\ {\parbox{1.5cm}{\epsfig{file=diagram151.eps,width=1.5cm}}}&=&2\kappa R I_1(\kappa R) K_0(\kappa R)-1.\end{aligned}$$ Again these diagrams do not depend on the position associated with the open circle. We proceed as in the spherical case to obtain the expressions of $\Omega_n^n$, $\Omega_{n+1}^n$, and $\Omega_{n}^{n+1}$. The resummation of all the contributions of the segments of consecutive ($U$) bonds on the liquid-gas interface is $$I_\ell=\frac{1}{2\kappa (R+\ell) I_0[\kappa (R+\ell)] K_0[\kappa (R+\ell)]}. \label{ilcyl}$$ On the other hand, segments on the wall contribute as $$I_\psi^1=\left(\frac{1}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}} \right) \frac{1}{2\kappa R I_0(\kappa R)K_0(\kappa R)} \label{ipsi1cyl}$$ if the segment is at any of the extremes of the diagram and otherwise $$I_\psi^2= \left(\frac{1+\frac{\kappa}{g}\frac{I_1(\kappa R)}{I_0(\kappa R)}} {1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}} \right) \frac{1}{2\kappa R I_0(\kappa R)K_0(\kappa R)}. \label{ipsi2cyl}$$ With these results, Eqs. (\[omega11\]), (\[omega12\]), and (\[omega21\]) reduce, respectively, to $$\begin{aligned} &&\Omega_1^1 = \frac{2\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]}\left(\frac{1}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) ,\label{omega11cylinder}\\ &&\Omega_1^2 = \frac{\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]} \left(\frac{I_0(\kappa R)}{I_0[\kappa(R+\ell)]} \right)\nonumber\\&&\times\left(\frac{1+\frac{\kappa}{g}\frac{I_1(\kappa R)}{I_0(\kappa R)}}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) ,\label{omega12cylinder}\\ &&\Omega_2^1 = \frac{\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]} \left(\frac{K_0[\kappa (R+\ell)]}{K_0(\kappa R)} \right)\nonumber\\&&\times\left(\frac{1}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right)^2 .\label{omega21cylinder}\end{aligned}$$ For higher-order contributions $$\begin{aligned} &&\Omega_n^n = \Omega_1^1 \Bigg[\left(\frac{I_0(\kappa R)K_0[\kappa (R+\ell)]}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) \nonumber\\ &&\times\left(\frac{1+\frac{\kappa}{g}\frac{I_1(\kappa R)} {I_0(\kappa R)}}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) \Bigg]^{n-1},\label{omeganncylinder}\\ &&\Omega_n^{n+1} = \Omega_1^2 \Bigg[\left(\frac{I_0(\kappa R)K_0[\kappa (R+\ell)]}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) \nonumber\\ &&\times\left(\frac{1+\frac{\kappa}{g}\frac{I_1(\kappa R)} {I_0(\kappa R)}}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) \Bigg]^{n-1}, \label{omegann1cylinder}\\ &&\Omega_{n+1}^n = \Omega_2^1 \Bigg[\left(\frac{I_0(\kappa R)K_0[\kappa (R+\ell)]}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) \nonumber\\ &&\times\left(\frac{1+\frac{\kappa}{g}\frac{I_1(\kappa R)} {I_0(\kappa R)}}{1-\frac{\kappa}{g}\frac{K_1(\kappa R)}{K_0(\kappa R)}}\right) \Bigg]^{n-1}, \label{omegan1ncylinder}\end{aligned}$$ which lead to Eq. (\[bindingcylindrical\]) after resummation of the series (\[Hmin7\]). For our formulation of the nonlocal model for fixed boundary conditions on the wall, we consider the diagrams $$\begin{aligned} &&{\parbox{1.5cm}{\epsfig{file=diagram152.eps,width=1.5cm}}}=2\kappa (R+\ell)I_1(\kappa R)K_0[\kappa (R+\ell)], \\ &&{\parbox{1.5cm}{\epsfig{file=diagram153.eps,width=1.5cm}}}=-2\kappa RI_0(\kappa R)K_1[\kappa (R+\ell)], \\ &&{\parbox{1.5cm}{\epsfig{file=diagram154.eps,width=1.5cm}}} = 2\kappa (R+\ell) I_1[\kappa (R+\ell)] K_0[\kappa (R+\ell)] -1. \nonumber\\\end{aligned}$$ The contributions of segments on the liquid-gas interface are given by Eq. (\[ilcyl\]) if the segment is on the left extreme and otherwise by $$I^\prime_\ell= \frac{1}{2\kappa (R+\ell) I_0[\kappa(R+\ell)] K_1[\kappa(R+\ell)]}.$$ Similarly, the segments of the diagrams on the wall on the left extreme of the diagram contribute as $$(I^{\prime})^1_\psi= \frac{1}{2\kappa R I_0(\kappa R) K_0(\kappa R)}$$ and otherwise as $$(I^{\prime})^2_\psi= \frac{1}{2\kappa R I_1(\kappa R)K_0[\kappa (R+\ell)]}.$$ The expressions from Eqs. (\[omega11-2\]), (\[omega12-2\]), and (\[omega21-2\]) for $\Omega_1^1$, $\Omega_1^2$, and $\Omega_2^1$ are $$\begin{aligned} &&\Omega_1^1 = \frac{2\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]}, \label{omega11cylinder-2}\\ &&\Omega_1^2 = \frac{\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]} \left(\frac{I_0(\kappa R)}{I_0[\kappa(R+\ell)]} \right), \label{omega12cylinder-2}\\ &&\Omega_2^1 = \frac{\pi L}{\kappa K_0(\kappa R)I_0[\kappa (R+\ell)]} \left(\frac{K_0[\kappa (R+\ell)]}{K_0(\kappa R)} \right), \label{omega21cylinder-2}\end{aligned}$$ and for higher-order contributions $$\begin{aligned} &&\Omega_n^n = \Omega_1^1 \left(\frac{I_0(\kappa R)K_0(\kappa (R+\ell))}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) ^{n-1},\label{omeganncylinder-2}\\ &&\Omega_n^{n+1} = \Omega_1^2 \left(\frac{I_0(\kappa R)K_0(\kappa (R+\ell))}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) ^{n-1}, \label{omegann1cylinder-2}\\ &&\Omega_{n+1}^n = \Omega_2^1 \left(\frac{I_0(\kappa R)K_0[\kappa (R+\ell)]}{I_0[\kappa(R+\ell)] K_0(\kappa R)}\right) ^{n-1}, \label{omegan1ncylinder-2}\end{aligned}$$ which correspond to Eqs. (\[omega11cylinder\])-(\[omegan1ncylinder\]) in the limit $g\to -\infty$. [72]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , in **, edited by (, , ), Vol. . , in **, edited by (, , ). , , , in **, edited by (, , ), Vol. . , , , ****, (). , , , ****, (). , in **, edited by (, , ). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, . , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , , ****, (). , ****, (). , ** (, , ). , ****, (). , ****, (). , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , , , ****, (). , , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ** (, , ) , ** (, , ). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, ().
{ "pile_set_name": "ArXiv" }
--- abstract: 'Machine learning is a popular tool that is being applied to many domains, from computer vision to natural language processing. It is not long ago that its use was extended to physics, but its capabilities remain to be accurately contoured. In this paper, we are interested in the prediction of 2D velocity and pressure fields around arbitrary shapes in laminar flows using supervised neural networks. To this end, a dataset composed of random shapes is built using Bézier curves, each shape being labeled with its pressure and velocity fields by solving Navier-Stokes equations using a CFD solver. Then, several U-net architectures are trained on the latter dataset, and their predictive efficiency is assessed on unseen shapes, using *ad hoc* error functions.' address: |  MINES ParisTech, PSL - Research University, CEMEF - Centre for material forming,\ CNRS UMR 7635, CS 10207 rue Claude Daunesse, 06904 Sophia-Antipolis Cedex, France. author: - Junfeng Chen - Jonathan Viquerat - Elie Hachem bibliography: - 'nn\_flow\_2d.bib' title: 'U-net architectures for fast prediction of incompressible laminar flows' --- machine learning ,neural networks , convolutional networks ,computational fluid dynamics ,immersed method Introduction ============ During the last few years, the CFD community has largely benefited from the fast-paced development of the machine learning (ML) field, and more specifically for that of the neural networks (NN) domain. In many cases, only a part of the resolution process is replaced with a trained NN, in order to reduce the computational cost. Examples for these applications are the prediction of closure terms in RANS [@Ling2016] [@Tracey2015] or LES [@Beck2018] computations. In other situations, a supervised network is directly trained to predict a flow profile: in [@Guo2016], an auto-encoder NN architecture is used to obtain steady state flow predictions around elementary and real-life shapes; in [@Jin2018], a fusion convolutional neural network (CNN) is trained to predict velocity snapshots around a cylinder in weakly turbulent flows, using the time history of pressure around the cylinder as an input. The U-net is a particular CNN architecture, which was initially proposed by Ronneberger [*et al*]{}[@Ronneberger2015] for biomedical image segmentation. As shown in figure \[fig:Unet\], U-nets are composed of (i) a contractive path based on the repetition of a convolution/max-pooling pattern, followed by (ii) an upscaling path, in which a pattern of deconvolution/concatenation/convolution is applied several times. The *skip connections* in U-nets, originally proposed by [@Long2015], are used to link the contracting path to the upscaling path, and help localize features of high resolution during the upsampling, thus improving the segmentation accuracy. ![**Standard U-net architecture**. The contractive path is based on a pattern made of two convolutional layers (in light blue) followed by a max-pooling layer (in orange). At each occurence of the pattern, the image size is divided by two, while the number of filters is doubled. In the upscaling path, a deconvolution step is first applied to the input, while the number of filters is halved. The output of this layer is then concatenated with its mirror counterpart in the contractive path using a shortcut connections (here noted with arrows). Finally, two convolution layers are applied. At the end of the last layer, a $1 \times 1$ convolution with 3 filters is applied to obtain an RGB image at the output (in red).[]{data-label="fig:Unet"}](Unet.pdf) Based on the end-to-end property of U-nets, several variations have been proposed in the recent years. In [@Newell2016], a stacked network architecture is proposed for the estimation human pose. The whole network is simply built by adding U-nets one after the other, the output of a U-net being the input of the following one. In [@Tang2018], shortcuts were added from each U-net to all the following ones, in order to help the back-propagation of gradient. Multi-paths U-nets were introduced in [@Dolz2019], where separated contracting paths were used for images holding different informations. The different paths are merged at the bottom of the U-net, and share a single expanding path. In [@Li2019], the authors report the parallel training of independent U-nets, the final prediction being obtained as a weighted sum of the outputs of the different networks. Other methodological refinements demonstrated their potential, such as multi-scale inputs and supervision [@Stawiaski2017], for example. These different variations are discussed in details in section \[section:archis\]. U-net architectures recently made their way to other physical fields, such as fluid dynamics. In [@Thuerey2018], the authors exploit U-nets to infer the velocity and pressure of turbulent flow around airfoils computed in a Reynolds-Averaged Navier-Stokes framework. In [@Fukami2018], a traditional CNN is coupled to a Downsampled Skip-Connection Multi-Scale (DSC/MS) model to reconstruct the turbulent flow computed by direct numerical simulation. The DSC/MS model, which uses a downscaling-upsampling path like U-net, efficiently recovers the detailed velocity field. In this paper, focus is made on the prediction of laminar flows around arbitrary 2D shapes. To this end, we assess the predictive capabilities of several U-net based architectures, and compare them to that of a 4-levels baseline U-net, presented in figure \[fig:Unet\]. In section \[section:archis\], the functioning of the baseline U-net is briefly covered, and advanced architectures are introduced. Details about the dataset are given in section \[section:dataset\], along with a discussion on the metrics used to evaluate the results of the different architectures. In section \[section:training\], details are given about the training, and the performances of the different networks are compared on a test subset from the dataset. Finally, the predictions of the networks are assessed on unseen configurations, such as geometrical shapes or airfoils. Existing U-net architectures {#section:archis} ============================ U-nets were introduced in 2015 [@Ronneberger2015] as an improvement over a sliding-window method for classical convolutional networks [@Ciresan2012]. In classical convolutional networks, a contracting path similar to the first branch of the U-net is used to extract high-level features from an input image. However, as the abstraction level of the features increases, localization information about these features is lost, making it hard to directly reconstruct meaningful image as an output of the network. The idea behind U-net is to follow the contracting path by an upscaling path, in which the resolution of the feature maps is re-increased symmetrically to the contracting path using deconvolution layers. After each deconvolution, localization information is re-injected in the upscaling path by concatenating information from the symmetric level of the downscaling path using a shortcut connection. The concatenation is followed by a convolution operation that assembles the two informations (features coming from the previous level of the upscaling path, and localization coming from the shortcut connection) together. [.4]{} ![**U-net architectures**.[]{data-label="fig:archis"}](archi_Unet.pdf "fig:") [.55]{} ![**U-net architectures**.[]{data-label="fig:archis"}](archi_SUnet.pdf "fig:") [.4]{} ![**U-net architectures**.[]{data-label="fig:archis"}](archi_PUnet.pdf "fig:") [.55]{} ![**U-net architectures**.[]{data-label="fig:archis"}](archi_CUnet.pdf "fig:") Recently, improved architectures based on U-nets have been proposed. Newell [*et al*]{}[@Newell2016] proposed a stacked architecture, similar to the one shown in figure \[fig:SUnets\]. The successive stacking of networks is said to consolidate the captured features, and to allow for the grasp of more complex spatial relationships. In [@Tang2018], Tang [*et al*]{}introduced coupled U-nets (CU-nets), presented in figure \[fig:CUnets\]. Based on the current trend of dense connectivities in convolutional networks [@Huang2017], the authors proposed to add a similar level of density to coupled U-nets, by introducing shortcuts between layers from same levels. They show that these shortcut help the information flow across the structure, while improving the efficiency of the gradient backpropagation. Finally, the separate training of U-nets in parallel is reported in [@Li2019]. In this contribution, the authors generate different versions of their initial dataset by adding different zero-mean gaussian noises. Then, three different U-nets are trained separately, each on a different version of the dataset. The final output is computed as a weighted combination of the outputs of the separate U-nets. Results show that this method reduces overfitting, thus inducing better performances on reduced datasets, and proved to be superior to a single U-net trained on any of the datasets (noisy or not). In the current paper, we introduce a parallel U-net (PU-net) architecture, shown in figure \[fig:PUnets\]. The input image is fed to two separate U-nets, which outputs are simply recombined using an average layer. At the difference of [@Li2019], the two networks are trained concurrently, using the same dataset. Dataset construction {#section:dataset} ==================== This section provides insights on the random shape dataset generation used to train networks in the next sections. This dataset was initially used in [@Viquerat2019], thus only the main lines are sketched here. For more details, the reader is referred to [@Viquerat2019]. First, we describe the steps to generate arbitrary shapes by means of connected Bezier curves. Then, the solving of the Navier-Stokes equations with an immersed method is presented. Finally, details about the dataset are given. Random shape generation ----------------------- In the first step, $n_s$ random points are drawn in $\left[ 0, 1 \right]^2$, and translated so their center of mass is positioned in $(0,0)$. An ascending trigonometric angle sort is then performed (see figure \[fig:shape\_generation\_1\]), and the angles between consecutive random points are then computed. An average angle is then computed around each point (see figure \[fig:shape\_generation\_2\]) using: $$\theta^*_i = \alpha \theta_{i-1,i} + (1 - \alpha) \theta_{i,i+1},$$ with $\alpha \in \left[0,1\right]$. The averaging parameter $\alpha$ allows to alter the sharpness of the curve locally, maximum smoothness being obtained for $\alpha = 0.5$. Then, each pair of points is joined using a cubic Bézier curve, defined by four points: the first and last points, $p_i$ and $p_{i+1}$, are part of the curve, while the second and third ones, $p^*_i$ and $p^{**}_i$, are control points that define the tangent of the curve at $p_i$ and $p_{i+1}$. The tangents at $p_i$ and $p_{i+1}$ are respectively controlled by $\theta^*_i$ and $\theta^*_{i+1}$ (see figure \[fig:shape\_generation\_3\]). A final sampling of the successive Bézier curves leads to a boundary description of the shape (figure \[fig:shape\_generation\_4\]). Using this method, a wide variety of shapes can be attained, as shown in figure \[fig:shape\_examples\]. [.45]{} ![**Random shape generation with cubic Bézier curves**. []{data-label="fig:shape_generation"}](shape_generation_1.pdf "fig:") [.45]{} ![**Random shape generation with cubic Bézier curves**. []{data-label="fig:shape_generation"}](shape_generation_2.pdf "fig:") [.45]{} ![**Random shape generation with cubic Bézier curves**. []{data-label="fig:shape_generation"}](shape_generation_3.pdf "fig:") [.45]{} ![**Random shape generation with cubic Bézier curves**. []{data-label="fig:shape_generation"}](shape_generation_4.pdf "fig:") Numerical resolution of Navier-Stokes equations {#section:NS} ----------------------------------------------- The flow motion of incompressible newtonian fluids is described by the Navier-Stokes (NS) equations: $$\label{eq:ns_equation1} \left\{ \begin{aligned} \rho\ (\partial_{t} {\textbf{v}} + {\textbf{v}} \cdot \nabla {\textbf{v}}) -\nabla \cdot \left( 2 \eta {\boldsymbol{\epsilon}}({\textbf{v}}) - p {\textbf{I}} \right) & = {\textbf{f}}, \\ \nabla \cdot {\textbf{v}} &= 0, \end{aligned} \right.$$ where $t \in [0,T]$ is the time, ${\textbf{v}}(x,t)$ the velocity, $p(x,t)$ the pressure, $\rho$ the fluid density, $\eta$ the dynamic viscosity and ${\textbf{I}}$ the identity tensor. In order to efficiently construct the dataset, an immersed boundary method is used for resolution instead of the usual body-fitted method, avoiding a systematic re-meshing of the whole domain for each shape. This method rely on a unified fluid-solid eulerian formulation based on level-set description of the geometry [@Bruchon2009], and leads to the following set of modified equations: $$\label{eq:ns_equation2} \left\{ \begin{aligned} \rho^* (\partial_{t} {\textbf{v}} + {\textbf{v}} \cdot \nabla {\textbf{v}}) -\nabla \cdot \left( 2 \eta {\boldsymbol{\epsilon}}({\textbf{v}}) + {\boldsymbol{\tau}} - p {\textbf{I}} \right) & = {\textbf{f}}, \\ \nabla \cdot {\textbf{v}} &= 0, \end{aligned} \right.$$ where we have introduced the following mixed quantities: $$\begin{aligned} {\boldsymbol{\tau}} & = H(\alpha) {\boldsymbol{\tau}}_{\text{s}},\\ \rho^* & = H(\alpha) \rho_{\text{s}} + (1-H(\alpha)) \rho_{\text{f}}, \end{aligned}$$ where the subscripts $f$ and $s$ respectively refer to the fluid and the solid, and $H(\alpha)$ is the Heaviside function: $$\label{eq:heavyside31} H(\alpha) = \left\{ \begin{aligned} 1 & \text{ if}\ \alpha > 0,\\ 0 & \text{ if}\ \alpha < 0. \end{aligned} \right.$$ The reader is referred to [@Hachem2013] for additional details about formulation (\[eq:ns\_equation2\]). Eventually, the modified equations (\[eq:ns\_equation2\]) are cast into a stabilized finite element formulation, and solved using a variational multi-scale (VMS) solver [@Hachem2013]. Dataset ------- The dataset (DS) is composed of 12.000 shapes, along with their steady-state velocity and pressure fields at $Re=10$ (see figure \[fig:dataset\_example\]). All the labels were computed using cimLib [@Hachem2013], following the methods exposed in section \[section:NS\]. In the following, the DS is systematically divided into three sets: 9600 shapes for the training set, 1200 shapes for the validation set, and 1200 shapes for the test set. [.3]{} [.3]{} [.3]{} Networks training {#section:training} ================= Networks -------- The baseline U-net exploited in this paper is composed of 4 blocks in the contractive path, a bottleneck block, and 4 blocks in the upscaling path (see figure \[fig:Unet\]. The initial number of kernels is set to 32 and is doubled at each pooling operation, reaching 512 in the bottleneck. It is then divided by two at each upscaling step, dropping back to 32 in the final block. The upscaling path is followed by a last $1 \times 1$ convolution with 3 filters, thus reconstructing an RGB image. All convolution kernels are of size $3 \times 3$, deconvolution kernels and strides of size $2 \times 2$, and pooling sizes and strides also of size $2 \times 2$. All activation functions are rectified linear units (ReLu). Training -------- The learning process in neural networks consists in adjusting all the biases and weights of the network in order to reduce the value of a well-chosen loss function. For regression cases, it is common to choose the mean squared error (MSE). With the loss function at hand, the optimization of the weights and biases is performed with a stochastic gradient descent that exploits the back-propagation algorithm [@Goodfellow2017]. The gradient descent is usually performed multiple times on small random subsets (called mini-batches), instead of considering a single evaluation of the gradient on the whole dataset. This method represents a compromise between an accurate gradient computation and a low computational cost. In this paper, the batch size is set to 8. The whole dataset, split in mini-batches, is re-used multiple times in random orders, a full re-use of it being called an epoch. This procedure is repeated as long as the validation accuracy (the accuracy of the prediction computed on the validation subset) increases, and is stopped when it starts decreasing, meaning that the network is overfitting ([*i.e.*]{}learning non-generic features) on the training subset. In this paper, the input images are downscaled to a constant size of $107 \times 78$ pixels to limit the training cost. To ensure fair comparisons, all random seeds are set once and for all for the training of the different considered networks. The amount of ready-to-use neural networks libraries has exploded in the recent years, most of them exploiting C++ or Python. For supervised learning, they usually include a wide range of choices regarding layer types, activation functions, losses, optimizers and so on. In this paper, we chose to use Keras [@Chollet2018] for its high level of abstraction and the ease of use provided by the Python language. Model evaluation {#section:evaluation} ---------------- The mean squared error is often used as a loss function for the train of neural networks. It is defined as: $${\varepsilon}_{\text{mse}} = \frac{1}{n(P)} \sum_{p=1}^{n(P)} \left( \frac{1}{3} \sum_{j=1}^3 \left(y_{pj} - \hat{y}_{pj}\right)^2 \right),$$ where $P$ is the set of all pixels in an image, $n(P)$ is the number of pixels in the image, $j$ is the index looping over the three channels of an RGB image, $\hat{y}_{pj}$ is the predicted value of channel $j$ at pixel $p$, and $y_{pj}$ is its corresponding label. Still, it can be hard to compare the quality of different predictions by relying only on the mean squared error, as it is directly related to the amplitude of the solution. To this end, a complementary pixel-wise relative error function is proposed for the assessment of the predictions quality: $${\varepsilon}_{\alpha} = \frac{1}{n(P)} \sum_{p=1}^{n(P)} \mathbbm{1} \left( {\varepsilon}^p_\alpha > 0 \right), \text{ with } {\varepsilon}^p_\alpha = \mathbbm{1} \left( \left| \frac{y_{p} - \hat{y}_{p}}{y_{p} + \eta} \right| > \alpha \right) \text{ and } y_p = \sum_{j=1}^3 y_{pj}.$$ In the latter expression, $\eta$ is a small positive number introduced to ensure a non-zero denominator in the case where $y_{pj} = 0$, and $\alpha$ is a tolerance for the relative error. Hence, this metric returns the proportion of pixels which prediction accuracy is lower than a given threshold level $\alpha$, without being proportional to the solution amplitude. Regarding the perturbation induced by $\eta$, we argue that only the pixels whose absolute value is close to $\eta$ will be impacted. By setting $\eta = \num{1e-6}$, the variations observed in ${\varepsilon}_\alpha$ were systematically lower than $0.1\%$, which represents a negligible error, as will be shown in later sections. Additionally, for a single image, it is possible to plot the 2D map of ${\varepsilon}^p_\alpha$ to locate the areas of low prediction accuracy. Results ======= In this section, a prediction from the test subset obtained with the baseline U-net is first presented and analyzed in details with the two metrics presented in section \[section:evaluation\]. Then, the different U-net architectures are statistically compared on the test subset. In a third time, the predictive efficiency of the different networks is assessed on several unseen shapes, including NACA airfoils. A prediction example drawn from the test subset is shown in figure \[fig:predictions\], along with its corresponding exact solutions and error maps. As can be seen, the predictions are physically sound, and, to the eye, fairly close to their targets. Looking at the pressure predictions, we observe that the network successfully predicted two minimal pressure areas at the top and bottom right of the shape. The maximal errors are located in the direct vicinity of the shape, and in areas of high gradients. By comparing the two error maps ${\varepsilon}_{0.01}^p$ and ${\varepsilon}_{0.05}^p$, one sees that most pixels are predicted within a 5% relative error range, and that the remaining pixels mostly concentrate at the surface of the shape. These error levels could be reduced by reducing the inference region of the network. [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} [.45]{} As both ${\varepsilon}_\text{mse}$ and ${\varepsilon}_\alpha$ are related to a single image, we compute their average, standard deviation and maximal values to represent these estimators over the entire test subset. The different errors, shown in figure \[fig:prediction\_errors\], are all normalized to their baseline U-net counterpart for better comparison. On the top row, the error levels for the velocity predictions indicate that, in average, advanced compositions of U-nets provide better predictions in average, up to 45% of improvement for the PU-net in the MSE sense. Regarding standard deviation and maximum error, SU-nets present a significant advantage, with a drop close to 40% for standard deviation and 50% for maximum error using MSE. These results suggest that SU-nets are able to catch outliers with more efficiency than CU- and PU-nets. Although one could expect to observe similar results for pressure predictions, the performances of SU-, CU- and PU-nets are clearly below that of the baseline U-net for average error, standard deviation and maximum error. The fact that pressure predictions cannot be improved using more complex networks can arise from an insufficient pre-processing of the target data, causing the learning problem to be ill-conditioned, as suggested in [@Thuerey2018]. Still, the improvements obtained on the velocity predictions when using more complex networks may not be worthwhile in regard of the overhead of computational effort they require. Indeed, the baseline U-net shown in figure \[fig:Unet\] contains 7.7 millions parameters, and requires 3 hours of training on a regular computer (not using GPU cards), while the SU-, CU- and PU-nets contain between 15 and 18 million parameters, leading to training times of 6 to 7 hours. For that reason, we suggest that, although improved U-net-based architectures provide better results in segmentation tasks, such superiority may not reflect in regression tasks such as field map predictions. [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](avg_err_velocity.pdf "fig:"){width=".9\textwidth"} [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](std_dev_velocity.pdf "fig:"){width=".9\textwidth"} [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](max_err_velocity.pdf "fig:"){width=".9\textwidth"} [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](avg_err_pressure.pdf "fig:"){width=".9\textwidth"} [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](std_dev_pressure.pdf "fig:"){width=".9\textwidth"} [.3]{} ![**Average errors, deviations and maximal errors obtained on the test subset using different U-net-based architectures** for velocity (top row) and pressure fields (bottom row).[]{data-label="fig:prediction_errors"}](max_err_pressure.pdf "fig:"){width=".9\textwidth"} We now focus on the prediction of field maps in realistic, unseen configurations such as geometric shapes or NACA airfoils, using the baseline U-net architecture of figure \[fig:Unet\]. As shown in figure \[fig:predictions\_unseen\], predictions provided by the trained U-net are still physically sound, and field details such as multiple high and low pressure areas in the vicinity of the obstacles are well reproduced, to the limit of the resolution of the network output. MSE error levels for these predictions are in the same order of magnitude as those obtained on the test subset, [*i.e.*]{}between and for both pressure and velocity. Further testing would require to reduce the predicted domain around the shapes to obtain better details in its vicinity, or to output predictions with higher resolution, although it would require more computational power. Conclusion ========== In the current contribution, we presented the prediction of velocity and pressure field maps around arbitrary 2D shapes in laminar flows exploiting U-net-based architectures. First, details were given about the considered U-net architectures, and how they recently started to be exploited for regression tasks. Insights were then given on the dataset construction and the networks training, followed by a discussion on the metrics used to measure the quality of the predictions. A prediction example was then analyzed on a shape drawn from the test subset, that showed good agreement with the exact solution for both velocity and pressure. Then, the predictions were assessed statistically on the entire test subset, showing that, although they perform better on segmentation tasks, advanced architectures such as stacked, coupled and parallel U-nets do not provide a clear advantage on regression tasks such as field map predictions. Finally, we showed that such predictions could be made on unseen shapes with reasonable accuracy, and that the predictions were physically sound, thanks to the rich geometrical features present in the dataset. Although this application is focused on a fairly basic case, these results give hope for the possibility of tackling more complex problems, such as turbulent viscosity maps at higher Reynolds numbers, or 3D fields, for example. Additionally, such tools can be exploited in related domains as surrogate models, for optimization purposes.
{ "pile_set_name": "ArXiv" }
--- abstract: 'For the old question whether there is always a prime in the interval $[kn, (k + 1)n]$ or not, the famous Bertrand’s postulate gave an affirmative answer for $k = 1$. It was first proved by P.L. Chebyshev in 1850, and an elegant elementary proof was given by P. Erdős in 1932 (reproduced in [@erdos pp.171-173]). M. El Bachraoui used elementary techniques to prove the case $k = 2$ in 2006 [@bachraoui]. This paper gives a proof of the case $k = 3$, again without using the prime number theorem or any deep analytic result. In addition we give a lower bound for the number of primes in the interval $[3n, 4n]$, which shows that as $n$ tends to infinity, the number of primes in the interval $[3n, 4n]$ goes to infinity.' --- **[Andy Loo]{}** St. Paul’s Co-educational College, Hong Kong \[section\] \[Theorem\][Definition]{} \[Theorem\][Corollary]{} \[Theorem\][Lemma]{} \[Theorem\][Example]{} [**Mathematics Subject Classification:**]{} 11N05\ [**Keywords:**]{} prime numbers Notations ========= Throughout this paper, we let $n$ run through the positive integers and $p$ run through the primes. We also let $\pi(n)$ be the prime counting function, which counts the number of primes not exceeding $n$. Further define $$f(x) = \sqrt{2\pi}x^{x + \frac{1}{2}}e^{-x}e^{\frac{1}{12x}}$$ and $$g(x) = \sqrt{2\pi}x^{x + \frac{1}{2}}e^{-x}e^{\frac{1}{12x + 1}}.$$ Lemmas ====== \[Lemma:oneone\] If $n \ge 8$, then $$\pi(n) \le \frac{n}{2}.$$ This is trivial since $1, 9$ and all even positive integers are not prime. \[Lemma:onetwo\] If $x$ is a positive real number, then $$\prod_{p \le x} p \le 4^x.$$ See [@erdos pp.167-168]. \[Lemma:onethree\] We have $$g(n) < n! < f(n)$$ See [@robbins]. \[Lemma:onefour\] For a fixed constant $c \ge \dfrac{1}{12}$, define the function $$h_1(x) = \frac{f(x + c)}{g(c)g(x)}.$$ Then for $x \ge \dfrac{1}{2}$, $h_1(x)$ is increasing. It suffices to prove that the function $$H_1(x) = (x + c)^{x + c + \frac{1}{2}}x^{-x - \frac{1}{2}}e^{\frac{1}{12(x + c)}- \frac{1}{12x + 1}}$$ is increasing for $x > \frac{1}{2}$. Indeed, we have $$H_1^\prime(x) = H_1(x)\Biggl(\biggl(\frac{1}{2(x + c)} - \frac{1}{12(x + c)^2} + \ln(x + c)\biggr) - \biggl( \frac{1}{2x} - \frac{1}{12\biggl(x + \dfrac{1}{12}\biggr)^2} + \ln x \biggr)\Biggr),$$ where $H_1(x) > 0$. Let $$F_1(x) = \frac{1}{2x} + \ln x.$$ As $-\dfrac{1}{12(x + c)^2} \ge \dfrac{1}{12\biggl(x + \dfrac{1}{12}\biggr)^2}$, it suffices to prove that $F_1(x)$ is increasing, so that $$\biggl(\frac{1}{2(x + c)} + \ln(x + c)\biggr) - \biggl(\frac{1}{2x} + \ln x\biggr) \ge 0.$$ We actually have $$F^\prime_1(x) = -\frac{1}{2x^2} + \frac{1}{x} = \frac{2x - 1}{2x^2},$$ which must be non-negative for all $x \ge \dfrac{1}{2}$. Therefore, the desired result follows. \[Lemma:onefive\] For a fixed positive constant $c$, define the function $$h_2(x) = \frac{f(c)}{g(x)g(c - x)}.$$ Then when $\dfrac{1}{2} \le x < \dfrac{c}{2}$, $h^\prime_2(x) > 0$; when $x = \dfrac{c}{2}$, $h^\prime_2(x) = 0$ and when $\dfrac{c}{2} < x \le c - \dfrac{1}{2}$, $h^\prime_2(x) < 0$. It suffices to prove that the function $$H_2(x) = x^{x + \frac{1}{2}}(c - x)^{c - x + \frac{1}{2}}e^{\frac{1}{12x + 1} + \frac{1}{12(c - x) + 1}}$$ has the following property: when $\dfrac{1}{2} \le x < \dfrac{c}{2}$, $H^\prime_2(x) < 0$; when $x = \dfrac{c}{2}$, $H^\prime_2(x) = 0$ and when $\dfrac{c}{2} < x \le c - \dfrac{1}{2}$, $H^\prime_2(x) > 0$. Indeed, we have $$H^\prime_2(x) = x^{x + \frac{1}{2}}(c - x)^{c - x + \frac{1}{2}}e^{\frac{1}{12x + 1} - \frac{1}{12(c - x) + 1}}(F_2(x) - F_2(c - x)),$$ where $$F_2(x) = \frac{12}{(12(c - x) + 1)^2} + \frac{1}{2x} + \ln x.$$ Clearly, for $\dfrac{1}{2} \le x < c$, $$x^{x + \frac{1}{2}}(c - x)^{c - x + \frac{1}{2}}e^{\frac{1}{12x + 1} - \frac{1}{12(c - x) + 1}} > 0.$$ Next, we actually have $$F^\prime_2(x) = \frac{288}{(12(c - x) + 1)^3} + \frac{2x - 1}{2x^2},$$ which must be positive for all $\dfrac{1}{2} \le x < c$. Thus whenever $\dfrac{1}{2} \le x \le c- \dfrac{1}{2}$, $F_2(x)$ is increasing while $F_2(c - x)$ is decreasing, implying that there is at most one value of $x$ with $\dfrac{1}{2} \le x \le c- \dfrac{1}{2}$ satisfying $F_2(x) = F_2(c - x)$. It is clear that $x = \dfrac{c}{2}$ is such a value. It follows that when $\dfrac{1}{2} \le x < \dfrac{c}{2}$, $H^\prime_2(x) < 0$ and when $\dfrac{c}{2} < x \le c - \dfrac{1}{2}$, $H^\prime_2(x) > 0$. Main Results ============ Now, suppose $n > e^{12}$. The product of all primes $p \in (3n, 4n]$, if any, must divide $\binom{4n}{3n}$. Let $\beta(p)$ be the power of $p$ in the prime factorization of $\binom{4n}{3n}$. Let $$\binom{4n}{3n} = T_1T_2T_3$$ where $$T_1 = \prod_{p \le \sqrt{4n}}p^{\beta(p)}, \quad T_2 = \prod_{\sqrt{4n} < p \le 3n }p^{\beta(p)} \quad\text{and}\quad T_3 = \prod_{3n < p \le 4n }p^{\beta(p)}.$$ Bounding each multiplicand in $T_1$ from above by $4n$ (see [@robbins p. 24]) and applying Lemma \[Lemma:oneone\], $$T_1 < (4n)^{\pi(\sqrt{4n})} \le (4n)^{\frac{\sqrt{4n}}{2}} = (4n)^{\sqrt{n}}.$$ Consider $T_2$. As the prime factorization of $\displaystyle\binom{n}{j}$ in [@robbins p. 24] manifests, for $\sqrt{4n} < p \le 3n$, $\beta(p) \le 1$. Let $x > 0$ and let $[x]$ be the greatest integer less than or equal to $x$. Define $\left\{x\right\} = x - [x]$. Let $r$ and $s$ be real numbers satisfying $s > r \ge 1$. Observe that number of integers in the interval $(s - r, s]$ is $[s] - [s - r]$, which is $[r]$ if $\left\{s\right\} \ge \left\{r\right\}$ and $[r] + 1$ if $\left\{s\right\} < \left\{r\right\}$. Let $N$ be the set of all positive integers. We define $$\begin{Bmatrix}s\\r\end{Bmatrix} = \frac{\displaystyle\prod_{k \in (s - r, s] \cap N} k}{\displaystyle\prod_{k \in (0, r] \cap N} k} = \delta(r, s)\binom{[s]}{[r]},$$ where $\delta(r, s) = 1$ if $\left\{s\right\} \ge \left\{r\right\}$ and $\delta(r, s) = [s - r] + 1$ if $\left\{s\right\} < \left\{r\right\}$. In both cases, $\delta(r, s) \le s$. Now let $A = \begin{Bmatrix}4n/3\\n\end{Bmatrix}$, $B = \begin{Bmatrix}2n\\3n/2\end{Bmatrix}$, $C = \begin{Bmatrix}4n/17\\3n/13\end{Bmatrix}$ and $D = \begin{Bmatrix}2n/7\\4n/15\end{Bmatrix}$. We have the following observations: - $\displaystyle\prod_{\sqrt{4n} < p \le \frac{n}{6}}p \le \prod_{p \le \frac{n}{6}} p \le 4^{\frac{n}{6}}$ (by Lemma \[Lemma:onetwo\]) - If $\dfrac{n}{6} < p \le \dfrac{2n}{11}$, then $$2p < \frac{n}{2} < 3p < 8p < \frac{3n}{2} < 9p < 11p \le 2n.$$ Hence $\displaystyle\prod_{\frac{n}{6} < p \le \frac{2n}{11}}p$ divides $B$. - If $\dfrac{2n}{11} < p \le \dfrac{4n}{21}$, then $$p < \frac{n}{3} < 2p < 5p < n < 6p < 7p \le \frac{4n}{3}.$$ Hence $\displaystyle\prod_{\frac{2n}{11} < p \le \frac{4n}{21}}p$ divides $A$. - If $\dfrac{4n}{21} < p \le \dfrac{n}{5}$, then $$5p \le n < 6p < 15p \le 3n < 16p < 20p \le 4n < 21p.$$ Hence $\beta(p) = 0$. - If $\dfrac{n}{5} < p \le \dfrac{2n}{9}$, then $$p < \frac{n}{3} < 2p < 4p < n < 5p < 6p \le \frac{4n}{3}.$$ Hence $\displaystyle\prod_{\frac{n}{5} < p \le \frac{2n}{9}}p$ divides $A$. - If $\dfrac{2n}{9} < p \le \dfrac{3n}{13}$, then $$4p < n < 5p < 13p < 3n < 14p < 17p < 4n < 18p.$$ Hence $\beta(p) = 0$. - $\displaystyle\prod_{\frac{3n}{13} < p \le \frac{4n}{17}}p$ divides $C$. - If $\dfrac{4n}{17} < p \le \dfrac{n}{4}$, then $$4p \le n < 5p < 12p \le 3n < 13p < 16p \le 4n < 17p.$$ Hence $\beta(p) = 0$. - If $\dfrac{n}{4} < p \le \dfrac{4n}{15}$, then $$p < \frac{n}{3} < 2p < 3p < n < 4p < 5p \le \frac{4n}{3}.$$ Hence $\displaystyle\prod_{\frac{n}{4} < p \le \frac{4n}{15}}p$ divides $A$. - $\displaystyle\prod_{\frac{4n}{15} < p \le \frac{2n}{7}}p$ divides $D$. - If $\dfrac{2n}{7} < p \le \dfrac{3n}{10}$, then $$3p < n < 4p < 10p \le 3n < 11p < 13p < 4n < 14p.$$ Hence $\beta(p) = 0$. - If $\dfrac{3n}{10} < p \le \dfrac{n}{3}$, then $$p < \frac{n}{2} < 2p < 4p < \frac{3n}{2} < 5p < 6p \le 2n.$$ Hence $\displaystyle\prod_{\frac{3n}{10} < p \le \frac{n}{3}}p$ divides $B$. - If $\dfrac{n}{3} < p \le \dfrac{4n}{9}$, then $$\frac{n}{3} < p < 2p < n < 3p \le \frac{4n}{3}.$$ Hence $\displaystyle\prod_{\frac{n}{3} < p \le \frac{4n}{9}}p$ divides $A$. - If $\dfrac{4n}{9} < p \le \dfrac{n}{2}$, then $$2p \le n < 3p < 6p \le 3n < 7p < 8p \le 4n < 9p.$$ Hence $\beta(p) = 0$. - If $\dfrac{n}{2} < p \le \dfrac{2n}{3}$, then $$\frac{n}{2} < p < 2p < \frac{3n}{2} < 3p \le 2n.$$ Hence $\displaystyle\prod_{\frac{n}{2} < p \le \frac{2n}{3}}p$ divides $B$. - If $\dfrac{2n}{3} < p \le \dfrac{3n}{4}$, then $$p < n < 2p < 4p \le 3n < 5p < 4n < 6p.$$ Hence $\beta(p) = 0$. - If $\dfrac{3n}{4} < p \le \dfrac{4n}{5}$, then $$\frac{n}{2} < p < \frac{3n}{2} < 2p \le 2n.$$ Hence $\displaystyle\prod_{\frac{3n}{4} < p \le \frac{4n}{5}}p$ divides $B$. - If $\dfrac{4n}{5} < p \le n$, then $$p \le n < 2p < 3p \le 3n < 4p \le 4n < 5p.$$ Hence $\beta(p) = 0$. - $\displaystyle\prod_{n < p \le \frac{4n}{3}}p$ divides $A$. - If $\dfrac{4n}{3} < p \le \frac{3n}{2}$, then $$n < p < 2p \le 3n < 4n < 3p.$$ Hence $\beta(p) = 0$. - $\displaystyle\prod_{\frac{3n}{2} < p \le 2n}p$ divides $B$. - If $2n < p \le 3n$, then $$n < p \le 3n < 4n < 2p.$$ Hence $\beta(p) = 0$. Therefore, to summarize, we get $$T_2 \le 4^{\frac{n}{6}}ABCD.$$ Note that by Lemma \[Lemma:onethree\], $$\begin{aligned} \binom{4n}{3n} &= \frac{(4n)!}{(3n)! n!}\\ &> \frac{g(4n)}{f(3n)f(n)}\\ &= \frac{2}{\sqrt{6\pi n}}e^{\frac{1}{48n + 1} - \frac{1}{36n} - \frac{1}{12n}}\biggl(\frac{256}{27}\biggr)^n, \end{aligned}$$ and similarly, $$\begin{aligned} A &= \begin{Bmatrix}4n/3\\n\end{Bmatrix} \le \frac{4n}{3}\binom{[4n/3]}{n} = \frac{4n}{3}\cdot\frac{\left[\dfrac{4n}{3}\right]!}{n!\biggl(\left[\dfrac{4n}{3}\right] - n\biggr)!}\\ &< \frac{4n}{3}\cdot\frac{f\biggl(\left[\dfrac{4n}{3}\right]\biggr)}{g(n)g\biggl(\left[\dfrac{4n}{3}\right] - n\biggr)}\\ &\le \frac{4n}{3}\cdot\frac{f\biggl(\dfrac{4n}{3}\biggr)}{g(n)g\biggl(\dfrac{n}{3}\biggr)}\quad\text{(by Lemma~\ref{Lemma:onefour})}\\ &= \frac{4n}{3}\sqrt{\frac{2}{\pi n}}e^{\frac{1}{16n} - \frac{1}{12n + 1} - \frac{1}{4n + 1}}\biggl(\frac{4^{\frac{4}{3}}}{3}\biggr)^n, \end{aligned}$$ $$\begin{aligned} B &= \begin{Bmatrix}2n\\3n/2\end{Bmatrix} \le 2n\binom{2n}{[3n/2]} = 2n\cdot\frac{\left[\dfrac{3n}{2}\right] + 1}{2n - \left[\dfrac{3n}{2}\right]}\binom{2n}{[3n/2] + 1}\\ &< 2n\cdot\frac{\dfrac{3n}{2} + 1}{2n - \dfrac{3n}{2}}\cdot \frac{f(2n)}{g\biggl(\left[\dfrac{3n}{2}\right] + 1\biggr)g\biggl(2n - \biggl(\left[\dfrac{3n}{2}\right] + 1\biggr)\biggr)}\\ &< (6n + 4)\cdot\frac{f(2n)}{g\biggl(\dfrac{3n}{2}\biggr)g\biggl(2n - \dfrac{3n}{2}\biggr)}\quad\text{(by Lemma~\ref{Lemma:onefive})}\\ &= \frac{12n + 8}{\sqrt{3\pi n}}e^{\frac{1}{24n} - \frac{1}{18n + 1} - \frac{1}{6n + 1}}\biggl(\frac{16}{3^{\frac{3}{2}}}\biggr)^n, \end{aligned}$$ $$\begin{aligned} C &= \begin{Bmatrix}4n/17\\3n/13\end{Bmatrix} \le \frac{4n}{17}\binom{[4n/17]}{[3n/13]} = \frac{4n}{17}\cdot\frac{\left[\dfrac{3n}{13}\right] + 1}{\left[\dfrac{4n}{17}\right] - \left[\dfrac{3n}{13}\right]} \binom{[4n/17]}{[3n/13] + 1}\\ &\le \frac{4n}{17}\cdot\frac{\dfrac{3n}{13} + 1}{\dfrac{4n}{17} - 1 - \dfrac{3n}{13}}\cdot \frac{f\biggl(\left[\dfrac{4n}{17}\right]\biggr)}{g\biggl(\left[\dfrac{3n}{13}\right] + 1\biggr) g\biggl(\left[\dfrac{4n}{17}\right] - \biggl(\left[\dfrac{3n}{13}\right] + 1\biggr)\biggr)} \end{aligned}$$ $$\begin{aligned} &< \frac{4n}{17}\cdot\frac{51n + 221}{n - 221}\frac{f\biggl(\dfrac{4n}{17}\biggr)}{g\biggl(\dfrac{3n}{13}\biggr) g\biggl(\dfrac{4n}{17} - \dfrac{3n}{13}\biggr)}\quad\text{(by Lemmas~\ref{Lemma:onefour} and \ref{Lemma:onefive})}\\ &= \frac{4n}{17}\cdot\frac{51n + 221}{n - 221}\cdot\frac{26}{\sqrt{6\pi n}} e^{\frac{17}{48n} - \frac{13}{36n + 13} - \frac{221}{12n + 221}} \biggl(221^{\frac{1}{221}}\biggl(\frac{13}{3}\biggr)^{\frac{3}{13}}\biggl(\frac{4}{17}\biggr)^{\frac{4}{17}}\biggr)^n,\\ \end{aligned}$$ and $$\begin{aligned} D &= \begin{Bmatrix}2n/7\\4n/15\end{Bmatrix} \le \frac{2n}{7}\binom{[2n/7]}{[4n/15]} = \frac{2n}{7}\cdot\frac{\left[\dfrac{4n}{15}\right] + 1}{\left[\dfrac{2n}{7}\right] - \left[\dfrac{4n}{15}\right]} \binom{[2n/7]}{[4n/15] + 1}\\ &\le \frac{2n}{7}\cdot\frac{\dfrac{4n}{15} + 1}{\dfrac{2n}{7} - 1 - \dfrac{4n}{15}}\cdot \frac{f\biggl(\left[\dfrac{2n}{7}\right]\biggr)}{g\biggl(\left[\dfrac{4n}{15}\right] + 1\biggr) g\biggl(\left[\dfrac{2n}{7}\right] - \biggl(\left[\dfrac{4n}{15}\right] + 1\biggr)\biggr)}\\ &< \frac{2n}{7}\cdot\frac{28n + 105}{2n - 105}\cdot\frac{f\biggl(\dfrac{2n}{7}\biggr)}{g\biggl(\dfrac{4n}{15}\biggr) g\biggl(\dfrac{2n}{7} - \dfrac{4n}{15}\biggr)}\quad\text{(by Lemmas~\ref{Lemma:onefour} and \ref{Lemma:onefive})}\\ &= \frac{4n^2 + 15n}{2n - 105}\cdot\frac{15}{\sqrt{2\pi n}} e^{\frac{7}{24n} - \frac{5}{16n + 5} - \frac{35}{8n + 35}} \biggl(\biggl(\frac{105}{2}\biggr)^{\frac{2}{105}} \biggl(\frac{15}{4}\biggr)^{\frac{4}{15}}\biggl(\frac{2}{7}\biggr)^{\frac{2}{7}}\biggr)^n.\\ \end{aligned}$$ Therefore $$\begin{aligned} T_3 &= \binom{4n}{3n}\frac{1}{T_1T_2} > \binom{4n}{3n}\frac{1}{(4n)^{\sqrt n}4^{\frac{n}{6}}ABCD}\\ &> \frac{\sqrt 3 \pi^{\frac{3}{2}}}{4160}e^EM^n (4n)^{-\sqrt n}\cdot \frac{n^{-\frac{3}{2}}(n - 221)(2n - 105)}{(3n + 2)(3n + 13)(4n + 15)}\\ &> \frac{\sqrt 3 \pi^{\frac{3}{2}}}{4160}e^EM^n (4n)^{-\sqrt n}\cdot \frac{n^{-\frac{3}{2}}n^2}{(4n)(4n)(5n)}\\ &= \frac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}e^EM^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}\\ \end{aligned}$$ where $$\begin{aligned} E &= \frac{1}{48n + 1} - \frac{1}{36n} - \frac{1}{12n} - \frac{1}{16n} + \frac{1}{12n + 1} + \frac{1}{4n + 1} - \frac{1}{24n} + \frac{1}{18n + 1} \\ &\hphantom{xx} + \frac{1}{6n + 1}- \frac{17}{48n} + \frac{13}{36n + 13} + \frac{221}{12n + 221} - \frac{7}{24n} + \frac{5}{16n + 5} + \frac{35}{8n + 35} \end{aligned}$$ and $$M = \frac{256}{7}\biggl(\frac{1}{4}\biggr)^{\frac{4}{3}}(3)\frac{3^{\frac{3}{2}}}{16}\biggl(\frac{1}{221}\biggr)^{\frac{1}{221}} \biggl(\frac{3}{13}\biggr)^{\frac{3}{13}} \biggl(\frac{17}{4}\biggr)^{\frac{4}{17}}\biggl(\frac{2}{105}\biggr)^{\frac{2}{105}} \biggl(\frac{4}{15}\biggr)^{\frac{4}{15}}\biggl(\frac{7}{2}\biggr)^{\frac{2}{7}}4^{-\frac{1}{6}} > 1.$$ Obviously $$\lim_{n \rightarrow \infty} e^E = 1.$$ Moreover, we have $$\ln \biggl(M^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}\biggr) = n\ln M - \sqrt n \ln (4n) - \frac{5}{2}\ln n.$$ When $n$ tends to infinity, it is easy to check that $\sqrt n \ln (4n) = o(n)$ and $\ln n = o(n)$. Thus, $\ln \biggl(M^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}\biggr)$ goes to infinity and so does $M^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}$. It follows that $$\lim_{n \rightarrow \infty} T_3 = +\infty,$$ which means that there exists some $n_0$ such that for all $n \ge n_0$, $T_3 > 1$. In fact, it is routine to check (using WolframAlpha for instance) that when $n > e^{12}$, $\dfrac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}e^EM^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}$ is always greater than $1$ and so $T_3 > 1$. Direct verification, on the other hand, ensures that there is always a prime in the interval $[3n, 4n]$ for all positive integers $n < e^{12}$. Therefore, our desired result ensues: \[Theorem:twoone\] For every positive integer $n$, there is a prime in the interval $[3n, 4n]$. Plainly, it follows that when $n \ge 2$, there is always a prime in the interval $(3n, 4n)$. If $n \ge 3$, then there is a prime in the interval $\biggl(n, \dfrac{4(n + 2)}{3}\biggr)$. If $n \equiv 0 {\rm~(mod~}3)$, then the result follows directly from Theorem \[Theorem:twoone\]. If $n \equiv 1 {\rm~(mod~}3)$, then by Theorem \[Theorem:twoone\] there exists a prime $p \in \biggl(n + 2, \dfrac{4(n + 2)}{3}\biggr)$. If $n \equiv 2 {\rm~(mod~}3)$, then by Theorem \[Theorem:twoone\] there exists a prime $p \in \biggl(n + 1, \dfrac{4(n + 1)}{3}\biggr)$. Next, we establish a lower bound for the number of primes in the interval $[3n, 4n]$. Bounding each prime in the interval from above by $4n$, we have the following For $n \ge 4$, the number of primes in the interval $(3n, 4n)$ is at least $$\log_{4n}\biggl(\dfrac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}e^EM^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}\biggr).$$ Note that $$\begin{aligned} &\log_{4n}\biggl(\dfrac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}e^EM^n (4n)^{-\sqrt n} n^{-\frac{5}{2}}\biggr)\\ &= \frac{-\dfrac{5}{2}\ln n + n\ln M - \sqrt n \ln (4n) + E + \ln\biggl(\dfrac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}\biggr)}{\ln n + \ln 4}\\ &= \frac{n\ln M - \sqrt n \ln (4n) + E + \ln\biggl(\dfrac{\sqrt 3 \pi^{\frac{3}{2}}}{332800}\biggr) + \dfrac{5}{2}\ln 4}{\ln n + \ln 4} - \frac{5}{2}\\ &> \frac{n\biggl(\ln M - \dfrac{\ln (4n)}{\sqrt n}\biggr)}{2\ln n} - \frac{5}{2}. \end{aligned}$$ Now check that $\displaystyle \lim_{n \rightarrow \infty} \frac{\ln (4n)}{\sqrt n} = 0$. Moreover, it is obvious that $$\lim_{n \rightarrow \infty} \frac{n}{\ln n} = +\infty.$$ Thus we have the following As $n$ tends to infinity, the number of primes in the interval $[3n, 4n]$ goes to infinity. In other words, for every positive integer $m$, there exists a positive integer $L$ such that for all $n \ge L$, there are at least $m$ primes in the interval $[3n, 4n]$. [**ACKNOWLEDGEMENTS.**]{} The author is deeply indebted to Dr. Kin Y. Li of the Hong Kong University of Science and Technology and Mr. C.J. Alaban for their generous and invaluable help. [99]{} International Journal of Contemporary Mathematical Sciences, Vol.1 (2006), no. 13, 617-621. Undergraduate Texts in Mathematics, Springer Verlag 2003, viii + 287pp. American Mathematical Monthly 62, 26-29, 1955.
{ "pile_set_name": "ArXiv" }
--- abstract: | X-ray flashes (XRFs) and X-ray rich gamma-ray bursts (XRGRBs) share many observational characteristics with long duration ($\gtrsim 2\;$s) GRBs, but the reason for which the spectral energy distribution of their prompt emission peaks at lower photon energies, $E_p$, is still a subject of debate. Although many different models have been invoked in order to explain the lower values of $E_p$, their implications for the afterglow emission were not considered in most cases, mainly because observations of XRF afterglows have become available only recently. Here we examine the predictions of the various XRF models for the afterglow emission, and test them against the observations of XRF 030723 and XRGRB 041006, the events with the best monitored afterglow light curves in their respective class. We show that most existing XRF models are hard to reconcile with the observed afterglow light curves, which are very flat at early times. Such light curves are, however, naturally produced by a roughly uniform jet with relatively sharp edges that is viewed off-axis (i.e. from outside of the jet aperture). This type of model self consistently accommodates both the observed prompt emission and the afterglow light curves of XRGRB 041006 and XRF 030723, implying viewing angles $\theta_{\rm obs}$ from the jet axis of $(\theta_{\rm obs}-\theta_0)\sim 0.15\theta_0$ and $(\theta_{\rm obs}-\theta_0)\sim \theta_0$, respectively, where $\theta_0\sim 3^\circ$ is the half-opening angle of the jet. This suggests that GRBs, XRGRBs and XRFs are intrinsically similar relativistic jets viewed from different angles. It is then natural to identify GRBs with $\gamma(\theta_{\rm obs}-\theta_0)\lesssim 1$, XRGRBs with $1\lesssim\gamma(\theta_{\rm obs}-\theta_0)\lesssim{\rm a\ few}$, and XRFs with $\gamma(\theta_{\rm obs}-\theta_0)\gtrsim{\rm a\ few}$, where $\gamma$ is the Lorentz factor of the outflow near the edge of the jet from which most of the observed prompt emission arises. Future observations with [ *Swift*]{} could help test this unification scheme in which GRBs, XRGRBs and XRFs share the same basic physics and differ only by their orientation relative to our line of sight. author: - 'Jonathan Granot, Enrico Ramirez-Ruiz and Rosalba Perna' title: 'Afterglow Observations Shed New Light on the Nature of X-ray Flashes' --- Introduction {#intro} ============ X-ray flashes (XRFs) are transient X-ray sources with durations ranging from several seconds to a few minutes and their distribution on the sky is consistent with it being isotropic [@Heise01; @Kippen03], similar to what is observed in long duration ($\gtrsim 2\;$sec) gamma-ray bursts (GRBs). XRFs are also similarly variable. They were first detected by the wide field camera (WFC) of [*BeppoSAX*]{} [@Heise01], and subsequently studied with [ *HETE-II*]{} [@Barraud03; @Lamb04]. In addition to XRFs, [ *HETE-II*]{} expanded the empirical classification of variable X-ray transients to include an intermediate class of events known as X-ray rich GRBs (XRGRBs). The spectrum of XRGRBs and XRFs is similar to that of GRBs [@Sakamoto04] except for the lower values of the photon energy $E_p$ at which their $\nu F_\nu$ spectrum peaks, and the lower energy output in gamma-rays and/or X-rays, $E_{\rm\gamma,iso}$, assuming isotropic emission. In all other respects XRFs, XRGRBs and GRBs seem to form a continuum. Many different models have been proposed for XRFs, most of which try to incorporate them in a unified scenario with GRBs. These models include high redshift GRBs [@Heise01], dirty (low $\gamma$) fireballs [@DCB99; @Heise01; @Huang02; @ZWM03; @ZWH04], regular GRBs viewed off-axis [@YIN02; @YIN03; @YIN04a; @YIN04b; @DDD04; @CK04; @ZWH04], photosphere dominated emission [@Drenkhahn02; @RRLR02; @Meszaros02], week internal shocks [low variability, $\Delta\gamma\ll\gamma$; @ZM02a; @Barraud03; @moch03], and large viewing angles in a structured [@Lamb05] or quasi [@Zhang04] universal jet. Most of these models mainly aim at explaining the low values of $E_p$ in XRFs, and do not address their expected afterglow properties. The afterglow evolution alone can, however, serve as a powerful test for XRF models, especially after the recent discovery of several afterglows of XRFs (020427, 020903, 030723, 040701, 040825B, 040912, 040916) and XRGRB 041006. Until a few years ago, XRFs were known predominantly as bursts of X-rays, largely devoid of any observable traces at any other wavelengths. However, a striking development in the last several years through the impetus of the [*HETE-II*]{} satellite, has been the measurement and localization of fading X-ray and optical signals from some XRFs. These afterglow observations resulted in three redshift determinations, for XRF 020903 [$z=0.251$, @Soderberg04], XRF 040701 [$z=0.2146$, @Kelson04] and XRGRB 041006 [$z=0.716$, @Fugazza04; @Price04b]. In two cases, XRF 030723 and XRGRB 041006, the afterglow light curves are reasonably well monitored from sufficiently early times so that they can be used to derive meaningful constraints on XRF models. In this paper we critically examine the different XRF models and contrast them with the afterglow observations of XRF 030723 and XRGRB 041006, as well as other available observations such as the prompt emission characteristics and the measured distances. The paper is organized as follows. §\[clas\] describes the current empirical classification and general properties of GRB, XRGRB and XRF sources. Various XRF models are considered in §\[others\] along with a brief discussion of the observations that support or undermine these schemes. All the models that are discussed in §\[others\] have at least one major flaw in common: they do not naturally produce the very flat afterglow light curve seen at early times in both XRF 030723 and XRGRB 041006. In the remainder of the paper we thus concentrate only on the class of models which naturally produce such light curves. That is, a roughly uniform jet with sufficiently sharp edges viewed outside the jet core. This class of models is discussed qualitatively in §\[off-axis\], and more quantitatively in §\[obs\], where it is also directly compared to the prompt emission and afterglow observations of XRGRB 041006 (§\[XRGRB041006\]) and XRF 030723 (§\[XRF030723\]). The role of our viewing angle as an essential parameter is given particular attention. In §\[other\_obs\] we briefly consider other XRFs and XRGRBs, and find that the data in these cases are too sparse and insufficient in order to derive meaningful constraints on the underlying model. Our conclusions are discussed in §\[conc\]. Empirical Classification of GRBs, XRGRBs & XRFs {#clas} =============================================== The operational definition of an XRF by the [*BeppoSAX*]{} team was that of a transient source, with a duration of less than $10^3\;$sec, whose flux triggered the Wide Field Camera (WFC) but not the Gamma Ray Burst Monitor (GRBM). Later, with [*HETE-II*]{}, the definition changed slightly and was based on the ratio of the fluence in the X-ray band to that in the gamma-ray band, $f_{\rm X/\gamma}=\log_{10}[S_{\rm X}(2-30\;{\rm keV})/S_\gamma(30-400\;{\rm keV})]$. In addition to XRFs, an intermediate class of X-ray rich GRBs (XRGRBs) was also introduced. According to this new empirical scheme, GRBs, XRGRBs, and XRFs correspond to $f_{\rm X/\gamma}<-0.5$, $-0.5<f_{\rm X/\gamma}<0$, and $f_{\rm X/\gamma}>0$, respectively. Although the observed peak energies, $E_p^{\rm obs}=(1+z)^{-1}E_p$ (the photon energy where $\nu F_\nu$ peaks), are on average about a factor of $\sim 10$ less than those of the “standard” GRBs ($E_{p,{\rm XRF}}^{\rm obs}\sim 25$ keV while $E_{p,{\rm GRB}}^{\rm obs}\sim 250$ keV), the spectra of XRFs are fitted by the same Band function that is commonly used to fit GRBs [@Band93], and they seem to obey the same correlation between $E_p$ (that is corrected for cosmological redshift) and the isotropic energy output seen in gamma-rays (or X-rays), $E_{\rm\gamma,iso}$: $E_p\propto E_{\rm\gamma,iso}^{1/2}$ [@Amati02; @L-RR-R02; @Lamb05]. While GRBs and XRFs have a different operational definition, they appear to form a continuum of events, rather than a bimodal distribution, with bursts varying uniformly from XRFs to XRGRBs to GRBs. The Viability of Various XRF Models {#others} =================================== In the next section we show that in order to reproduce the observed behavior seen in the afterglow light curves of both XRGRB 041006 and XRF 030723 a roughly uniform jet with sufficiently sharp edges viewed off-axis is required. This is a direct consequence of the very flat evolution of the afterglow light curve that is seen at early times. Such a behavior does not occur for a spherically symmetric outflow, or for a uniform jet that is viewed from within its aperture. The same also applies for a “structured” jet [@LPP01; @RLR02; @ZM02b] where the energy per solid angle $\epsilon$ props as the inverse square of the angle $\theta$ from the jet axis, outside of some small core angle $\theta_c$, $\epsilon\approx\epsilon_0\min[1,(\theta/\theta_c)^{-2}]$. For most models that have been proposed in the literature to explain the phenomenology of XRFs, the afterglow light curve at early times is expected to be similar to that of a spherical flow \[with $\epsilon=\epsilon(\theta_{\rm obs})$ if $\epsilon$ varies with $\theta$\], and thus behave qualitatively similar to GRB afterglow light curves. Although this early afterglow behavior alone makes most XRF models inconsistent with observations, in what follows, we give additional arguments that further undermine these various schemes. A straightforward interpretation of the low $E_p^{\rm obs}$ seen in both XRFs and XRGRBs is that they are in fact the high-redshift counterparts of long duration GRBs. While XRFs have on average lower energies than GRBs, their durations are comparable to those of GRBs [@Heise01], which argues against a high-redshift origin. Moreover, the recent redshift determination of XRF 020903 [$z=0.251$; @Soderberg04], XRF 040701 [$z=0.2146$, @Kelson04] and XRGRB 041006 [$z=0.716$; @Fugazza04; @Price04b] directly rules out this interpretation. Although some GRBs at very high redshifts may resemble XRFs, it is now clear that they do not represent the bulk of the population. In fact, recent estimates suggest that this high-redshift population may only constitute a small fraction of the total number of bursts, provided that the redshift distribution of GRBs accurately tracks the cosmic star formation rate of massive stars [e.g., @BN00; @BL02; @LRFRR02]. [@DCB99] have pointed out that “dirty fireballs”, i.e. relativistic outflows with a larger baryonic load and hence a lower initial Lorentz factor $\Gamma_0$ compared to classical GRBs, would have a smaller $E_p$ which could be in the X-rays. When XRFs were discovered it was natural to suggest this scenario as a possible way of achieving low values for $E_p$ [@Heise01; @Huang02]. We note here that while a lower $\Gamma_0$ implies a lower $E_p$ ($\propto\Gamma_0^4$) in the external shock model for the prompt emission, in the internal shocks model it would produce a higher $E_p$ ($\propto\Gamma_0^{-2}$). For the external shock model, the lower the observed $E_p=(1+z)E_p^{\rm obs}$ is, the lower the value of $\Gamma_0$ that is required to explain it. This implies that events with lower values of $E_p$ should have longer durations, since the deceleration time scales as $t_{\rm dec}\propto \epsilon^{1/(3-k)}\Gamma_0^{-2(4-k)/(3-k)}$, where $\epsilon$ is the energy per solid angle and $\rho_{\rm ext} \propto r^{-k}$. This is inconsistent with observations, as no clear trend exists between the total duration of the event and its $E_p$ [@Sakamoto04]. What is more, as is the case in any external shock model, the afterglow should be a smooth continuation of the prompt emission as both arise from the same external shock. The observations of XRF 030723 [@Fynbo04a] offer the best evidence so far against this. The presence of a dominant baryonic or shock pair photosphere within the standard fireball model was invoked by @RRLR02 and @Meszaros02 to explain the formation of XRFs. While this is a tenable scenario for producing XRGRBs, the very low $E_p<5\;$keV observed for XRFs at $z\sim 0.2$ are hard to reconcile with a low-$\Gamma_0$, pair-dominated photospheric component [@Meszaros02]. Another way of obtaining low values of $E_p$ is with a roughly constant $\Gamma_0$ between different events, but with a small contrast in the value of $\Gamma_0$ between different colliding shells in the internal shocks model, $\Delta\Gamma_0\ll\Gamma_0$ [@ZM02a; @Barraud03; @moch03]. This model, however, should produce an afterglow with an intensity that is comparable to those seen in typical GRBs. Furthermore, the isotropic equivalent kinetic energy of the afterglow shock at early times (when only the local value of $\epsilon$ along the line of sight is sampled, similar to the prompt emission) should be much larger than $E_{\rm\gamma,iso}$, because of the low radiative efficiency of the prompt emission in this scenario. An alternative model for XRFs arises in the context of the so called universal (structured) jet models. In this class of models it is assumed that all GRB jets have the same structure, where both $\epsilon$, and $\Gamma$ depend on the angle $\theta$ with respect to the jet axis [@LPP01; @RLR02; @ZM02b]. This model can reproduce the key features expected from the conventional on-axis uniform jet models, with the novelty being that the achromatic break time in the broadband afterglow light curves corresponds to the epoch during which the core of the jet becomes visible, rather than the edge of the jet as in the uniform jet model. For the internal shock model, which is thought to be the mechanism responsible for the prompt emission, $E_p \propto L_{\rm iso}^{1/2}\Gamma_0^{-2}$ [e.g., @RRLR02]. As there is no observed correlation between the duration on an event and its $E_p$, this suggests that $E_p \propto E_{\rm\gamma,iso}^{1/2}\Gamma_0^{-2}$. This reproduces the observed narrow correlation $E_p \propto E_{\rm\gamma,iso}^{1/2}$ only if $\Gamma_0$ is both independent of $\theta$ and has a very small scatter between different events. [@Lamb05] have proposed a unified description of XRFs, XRGRBs and GRBs in which either (i) the half-opening angle $\theta_0$ of a uniform jet varies over a wide range while its energy remains constant, or (ii) our viewing angle $\theta_{\rm obs}$ with respect to a universal structured jet varies over a wide range. For convenience, we shall refer to $\theta_0$ and $\theta_{\rm obs}$ in these two options, , respectively, simply as $\theta_*$. In this picture, small values of $\theta_*$ correspond to GRBs, while increasingly larger values of $\theta_*$ correspond to XRGRBs and then to XRFs. In this scenario $E_p\propto\theta_*^{-1}$, so that the large range of observed $E_p$ values, ranging from $E_p\gtrsim 1\;$MeV for bright GRBs to $E_p\lesssim 5\;$keV for dim XRFs (i.e. a range of a factor of $\gtrsim 200$), directly corresponds to a similar range in $\theta_*$. Both the inferred values of $\theta_*$ from the jet break times in the afterglow light curves, and the $\log N-\log S$ distribution of [*BATSE*]{} GRBs [@GGB05] suggest, however, a smaller range for $\theta_*$, of about $\sim 10$ ($0.05\lesssim\theta_*\lesssim 0.5$), rather than $\gtrsim 200$. Off-Axis Jet Models of GRBs & XRFs {#off-axis} ================================== The possibility that GRB outflows are collimated into narrow jets, where in many cases our line of sight would be outside of the jet aperture, resulting in no detectable prompt emission and an “orphan afterglows” at later times, was suggested by @Rhoads97. This was shortly after the first detection of a GRB afterglow and before there was compelling observational evidence for jets in GRBs. As observational evidence in favor of GRB outflows being collimated into narrow jets gradually accumulated, studies of the observational signatures of off-axis GRB jets became more common [@PL98; @WL99; @MSB00; @DGP02; @Granot02; @Levinson02; @TP02; @NPG03; @GL03; @EL04], again mostly devoted to orphan afterglows. The possibility that for viewing angles that are only slightly outside of the jet aperture the prompt emission might still be detectable, but would shift into the X-rays due to the reduced Doppler factor, has been pointed out by @WL99. That was, however, before the discovery of XRFs. After XRFs were discovered, @YIN02 suggested that GRB jets viewed slightly off-axis could naturally account for this newly discovered class of events. In later works they have significantly developed some aspects of this model [@YIN03; @YIN04a; @YIN04b], in particular those regarding the prompt emission. In this section we discuss various aspects of this model in some detail, with both the prompt and afterglow signatures being at the forefront of our attention. The Jet Structure {#jet_structure} ----------------- The usual assumption about the jet structure is that it is perfectly uniform within some finite initial half-opening angle, $\theta_0$, from the jet symmetry axis, and abruptly truncates outside of $\theta_0$ [@WL99; @YIN02]. Obviously, this is only an approximation, as physically one might expect that the jet would have a smoother outer edge, where the energy per solid angle, $\epsilon$, and the initial Lorentz factor, $\Gamma_0$, decrease smoothly with the angle $\theta$ from the jet symmetry axis, over some finite range in $\theta$, $\Delta\theta\gtrsim\Gamma_0^{-1}$. In fact, numerical simulations show that even if the jet initially has perfectly sharp edges (i.e. a ‘top hat’ jet), the interaction with the ambient medium causes its edges to become smoother with time [@Granot01]. This serves as a motivation for considering a roughly uniform jet with smooth edges, as a more realistic version of the ‘top hat’ jet. The most widely used version of such a jet is one with a Gaussian angular profile for $\epsilon$ [@ZM02b; @KG03], $\epsilon=\epsilon_0\exp(\theta^2/2\theta_0^2)$. There are also other similar jet profiles, where most of the energy resides within some finite half-opening angle $\theta_0$, and $\epsilon$ sharply drops outside of $\theta_0$. Numerical simulation of a jet boring its way through a massive star progenitor in the context of the collapsar model [@ZWH04] predict a roughly uniform jet core with $\theta_0\sim 3^\circ-5^\circ$ and wings where $\epsilon\propto\theta^{-3}$ which extend to larger angles, i.e. $\epsilon\approx\epsilon_0\min[1,(\theta/\theta_0)^{-3}]$. We consider all such models to be variants of the same basic jet structure, and when viewed from outside the jet core ($\theta_{\rm obs}>\theta_0$) they are considered as members of the same class of XRF models. The different aforementioned variants of this jet structure are considered in §\[obs\]. The Afterglow Light Curves {#AG} -------------------------- The early afterglow light curves for off-axis viewing angles ($\theta_{\rm obs}>\theta_0$) are generally flatter than those observed in typical on-axis ($\theta_{\rm obs}<\theta_0$) GRB afterglows [see @Granot02 and references therein]. For a jet structure for which $\epsilon$ and $\Gamma_0$ drop sharply with $\theta$ at $\theta>\theta_0$, we expect its early light curve to rise with time. In this case, the sharper the edge of the jet, the sharper the rise in the light curve [@Granot02]. For jets with sharp enough edges, the emission from the core of the jet (i.e. from $\theta<\theta_0$) dominates even at off-axis viewing angles ($\theta_{\rm obs}>\theta_0$), despite it being strongly beamed away from our line of sight (see Fig. \[diagram\]). This is either because there is no emitting material along the line of sight, or even if present its emission is still weaker than that arising from the jet core. As the jet sweeps up an increasing amount of external medium, it slows down and thereafter the relativistic beaming of the emission from the jet core away from our line of sight decreases. When $\gamma$ drops to $\sim (\theta_{\rm obs}-\theta_0)^{-1}$, our line of sight enters the beaming cone of the radiation from the jet core, causing the light curve to peak and subsequently decay, asymptotically approaching the light curve for an on-axis observer. If the edge of the jet is not sufficiently sharp (i.e. if $\epsilon$ and $\Gamma_0$ do not drop sufficiently sharply with $\theta$ at $\theta>\theta_0$), then the emission from material along our line of sight may dominate over that from the core of the jet for viewing angles slightly outside the edge of the jet. In this case the light curve at early times would not rise with time, but would instead simply decay more slowly when compared to the light curve seen by on-axis observers ($\theta_{\rm obs}<\theta_0$). Therefore, we conclude that the jet structure, and specifically the sharpness of its edges, can be constrained by early afterglow observations. In the context of the model discussed in this section, increasingly larger viewing angles will correspond to XRGRBs and XRFs. Such a scheme is tested against observations of XRGRB 041006 and XRF 030723 in §\[obs\]. The Prompt & Reverse Shock Emission {#prompt} ----------------------------------- The prompt emission for off-axis viewing angles ($\theta_{\rm obs}>\theta_0$) may also be dominated either by the emission from the jet core or by the emission from the material along the line of sight, depending on the viewing angle and on how sharp the edge of the jet is. If the edge of the jet is sufficiently sharp, the prompt emission is dominated by the core of the jet, and both the fluence and the peak photon energy drop sharply when compared to their on-axis values, as $[\gamma(\theta_{\rm obs}-\theta_0)]^{-6}$ and $[\gamma(\theta_{\rm obs}-\theta_0)]^{-2}$, respectively [@Granot02; @R-R05]. The prompt emission in this case arises from the same region as for on-axis viewing angles, which in this scenario correspond to GRBs. This suggests that the same physical mechanism is responsible for the prompt emission in GRBs and in XRFs (i.e. most likely internal shocks). If, on the other hand, the edges of the jet are not sharp enough, then the prompt emission will be dominated by material along our line of sight. As it might be hard to produce strong variability in the Lorentz factor of the outflow outside the core of the jet [@RR02; @ZWH04], internal shocks may not be very efficient, and the external shock due to the interaction with the external medium might dominate the prompt emission. In that case, a smooth prompt light curve consisting of a single wide peak might be expected. The ‘optical flash’ emission from the reverse shock is generally expected be weaker for off-axis observers [@FWW04]. If the reverse shock is Newtonian or only mildly relativistic, then the beaming of the prompt emission (that is attributed to internal shocks within the outflow, which occur before the ejecta is decelerated by the external medium) and the reverse shock emission would not be very different. In this case the ratio of the off-axis to on-axis flux or fluence should be roughly similar for the optical flash and the prompt emission. If the reverse shock is relativistic then it would significantly decelerate the ejecta, and the emission from the reverse shock would be less strongly beamed than the prompt emission. In this case, if the emission is dominated by the jet core (i.e. for a sharp edged jet), the ‘optical flash’ emission at off-axis viewing angles could be less suppressed compared to the prompt X-ray or gamma-ray emission. Linear Polarization {#pol} ------------------- An interesting implication of the off-axis jet model for XRGRBs and XRFs is that it predicts a higher degree of linear polarization of the prompt emission, the emission from the reverse shock, and the afterglow emission, if the polarization is dominated by the jet geometry while the magnetic field is mostly tangled in the plane of the shock, as expected from the two stream instability [@ML99]. For a shock produced magnetic field that is tangled within the shock plane, the polarization peaks at a viewing angle that satisfies $\gamma(\theta_{\rm obs}-\theta_0)\sim 1$ [@Gruzinov99; @Waxman03; @Granot03; @NPW03], since at such a viewing angle most of the observed radiation is emitted roughly along the shock plane in the rest frame of the emitting plasma, due to aberration of light effects. The peak polarization can reach up to tens of percent. This is relevant to the prompt emission, and may also be relevant for the optical flash emission. The peak of the polarization which occurs at $\gamma(\theta_{\rm obs}-\theta_0)\sim 1$ can shift to a larger viewing angle $\theta_{\rm obs}$ during the optical flash as the ejecta is decelerated by the reverse shock. For jets with sufficiently sharp edges that are viewed off-axis, the afterglow light curve initially rises at early times, and the polarization peaks around the time of the peak in the light curve, which occurs when $\gamma(\theta_{\rm obs}-\theta_0)$ decreases to $\sim 1$, as our line of sight enters the beaming cone of the emitting material [@Granot02]. Even if there is some lateral spreading of the jet, and an initially off-axis viewing angle enters into the jet aperture as the latter grows with time, then the afterglow polarization would be relatively large, as the line of sight would still be relatively close to the edge of the jet [@Sari99; @GL99]. One should keep in mind, however, that ordered magnetic fields might potentially play an important role in the polarization of the prompt emission [@Waxman03; @Granot03; @NPW03; @LPB03] as well as that of the reverse shock emission (the ‘optical flash’ and ’radio flare’) and the afterglow emission [@GKo03]. If the dominant cause of polarization is an ordered magnetic field component, instead of the jet geometry together with a shock produced magnetic field, then the viewing angle would have a smaller effect on the observed linear polarization. A recent analysis of archival ‘radio flare’ observations [@GT05] has set strong upper limits of the linear and circular polarization of this radio emission, and showed that these limits constrain the presence of an ordered magnetic field in the ejecta. The existing radio flare observations are for GRBs, which in the model considered here correspond to on-axis viewing angles ($\theta_{\rm obs}<\theta_0$). For a uniform jet with an ordered toroidal magnetic field, the polarization vanishes at the jet symmetry axis (i.e. for $\theta_{\rm obs}=0$) and strongly increases toward the edge of the jet. Therefore, the observed upper limits on the linear polarization translate to an upper limit on $\theta_{\rm obs}/\theta_0$. The best constraints so far are for GRB 991216: $P<7\%$ and $\theta_{\rm obs}/\theta_0\lesssim 0.4-0.55$, respectively. There are weaker constraints for GRBs 990123 and 020405. Tighter constraints on the presence of an ordered magnetic field in the ejecta are expected in the near future when a larger sample of radio flare polarization measurements becomes available. Such measurements for XRGRBs or XRFs are crucial when testing the off-axis jet model. This is because in this model one expects a viewing angle that is only slightly outside the edge of the jet and thus a large degree of polarization (tens of percent) for a purely ordered toroidal magnetic field in the ejecta. Description of the Numerical Model {#num} ---------------------------------- In this section we briefly describe the model that is used in §\[obs\] for describing the data. This is essentially model 1 of @GK03, similar to that used by @R-R05 for modeling the lightcurve of GRB 031203. The deceleration of the flow is calculated from the mass and energy conservation equations and the energy per solid angle $\epsilon$ is taken to be independent of time. The local emissivity is calculated using the conventional assumptions of synchrotron emission from relativistic electrons that are accelerated behind the shock into a power-law distribution of energies, $N(\gamma_e) \propto \gamma_e^{-p}$ for $\gamma_e > \gamma_m$, where the electrons and the magnetic field hold fractions $\epsilon_e$ and $\epsilon_B$, respectively, of the internal energy. The external density is taken to be a power law in the distance $r$ from the central source, $\rho_{\rm ext}=Ar^{-k}$, where $k=0$ corresponds to a uniform ISM while $k=2$ corresponds to a stellar wind of a massive star progenitor (assuming a constant ratio for the mass loss rate and the wind velocity). Another important physical parameter is the (true) energy of the jet, $E$, which is calculated assuming that the jet is double sided. The synchrotron spectrum is taken to be a piecewise power law [@SPN98]. The inverse-Compton scattering of the synchrotron photons by the same relativistic electrons, that is known as synchrotron-self Compton (SSC), is also taken into account [@PK00; @SE01]. The lateral spreading of the jet is neglected in this model. This approximation is consistent with results of numerical studies [@Granot01; @KG03] which show relatively little lateral expansion as long as the jet is relativistic. The light curves for observers located at different angles, $\theta_{\rm obs}$, with respect to the jet axis are calculated by applying the appropriate relativistic transformation of the radiation field from the local rest frame of the emitting fluid to the observer frame and integrating over equal photon arrival time surfaces [@Granot02; @R-RM04]. Observations {#obs} ============ The goal of this section is to quantitatively test the idea that a relativistic jet pointing slightly away from us could explain the observations of XRGRBs and XRFs. The modeling of radio, optical, and X-ray data is carried out in the framework of collimated ejecta interacting with the external medium. The model is described in §\[off-axis\]. In this work we focus our attention on two afterglows for which radio, optical, and X-ray light curves are available: XRGRB 041006 (§\[XRGRB041006\]) and XRF 030723 (§\[XRF030723\]). Although, as described in §\[other\_obs\], afterglow emission has also been detected for other other XRFs and XRGRBs, the data in the these cases are too sparse and insufficient in order to derive meaningful constraints on the underlying model. X-ray Rich GRB 041006 {#XRGRB041006} --------------------- XRGRB 041006 was detected by [*HETE-II*]{} [@Galassi04]. It had a fluence of $5\times 10^{-6}\;{\rm erg\;cm^{-2}}$ in the $2-30\;$keV range and $7\times 10^{-6}\;{\rm erg\;cm^{-2}}$ in the $30-400\;$keV range, corresponding to $f_{X/\gamma}\approx -0.15$ which classifies it as an XRGRB. It has a redshift of $z=0.716$ [@Fugazza04; @Price04a], which for a fluence of $f\approx 1.2 \times 10^{-5}\;{\rm erg\;cm^{-2}}$ in the $2-400\;$keV range gives $E_{\rm\gamma,iso}\approx 1.6\times 10^{52}\;$erg. It had an observed peak photon energy of[^1] $E_p^{\rm obs}=63^{+7}_{-5}\;$keV, corresponding to $E_p=109_{-9}^{+12}\;$keV. Figure \[fit\_XRGRB041006\] shows an off-axis model yielding an acceptable fit to the to the optical and X-ray afterglow observations of XRGRB 041006, which is also consistent with the upper limits at radio and sub-mm wavelengths [@Barnard04a; @Barnard04b; @SF04]. From this analysis one can conclude that a successful model for the afterglow of XRGRB 041006 is that of a collimated, misaligned jet interacting with a stellar wind external medium of mass density $\rho_{\rm ext}=Ar^{-2}$, where $r$ is the distance from the central source. The parameter values used in this fit are: $E=1.0\times 10^{51}\;$erg, $A_*\equiv A/(5\times 10^{11}\;{\rm gr\;cm^{-1}})=0.03$, $\theta_0=3^\circ$, $\theta_{\rm ons}=1.15\theta_0$, $p=2.2$, $\epsilon_e=0.1$, and $\epsilon_B=0.001$. The optical light curve is very flat at early times ($\alpha\sim 0$ at $t\lesssim 1\;$hr, where $F_\nu\propto t^{-\alpha}\nu^{-\beta}$) and becomes steeper after a few hours ($\alpha\approx 1.2$), which is a little steeper than the decay index in the X-ray at a similar time ($\alpha\approx 1$ at $t\approx 1\;$day). Also, the ratio of the flux in the optical and X-ray at $t\approx 1\;$day implies a spectral index of $\beta\approx 0.7-0.75$ assuming a single power law between them. This suggests that the cooling break frequency $\nu_c$ is above the optical after $1\;$day. Since one requires very extreme parameters to get $\nu_c$ to the X-ray range after $1\;$day (even getting $\nu_c$ to be above the optical after a day requires relatively low values of $\epsilon_B$ and of the external density), it is most likely that $\nu_c$ is between the optical and X-ray at $1\;$day, which can also explain the steeper temporal decay index in the optical (by $\Delta\alpha=0.25$) for a stellar wind environment ($k=2$). This favors a wind medium over a uniform density one, since otherwise the flux in the optical will decay more slowly than in the X-ray (also by $\Delta\alpha=0.25$), which is contrary to what is observed for XRGRB 041006. At $t\gtrsim 5\;$days there is a flattening in the optical light curve, which is probably due to an underlying SN component [@Garg04]. This explains why the observed flux is higher than that predicted by our narrow relativistic jet model. The fit to the afterglow observations does not, however, uniquely determine the model parameters. Some physical parameters are nonetheless constrained better than others. The afterglow data for XRGRB 041006 requires a stellar wind environment ($k=2$) with a low density ($A_*\sim 0.03$) and a viewing angle that is only slightly outside the edge of the jet, $(\theta_{\rm obs}-\theta_0)\sim 0.15\theta_0\sim 10^{-2}\;$rad, in order to successfully explain both the spectrum + temporal decay rates in the optical and X-ray at $\sim 1\;$day and the very flat optical light curve seen at early times. If GRB jets have well-defined edges, both the prompt gamma-ray fluence and the peak of the spectrum drop very sharply outside the opening of the jet, as[^2] $\delta^{-3}$ and $\delta^{-1}$, respectively, where[^3] $\delta\sim[\gamma(\theta_{\rm obs}-\theta_0)]^2$. Therefore, the low $E_{\rm\gamma,iso}$ of XRGRB 041006 combined with $E_{\rm k,iso} = E/(1-\cos\theta_0) \approx E(2/\theta_0^2)\approx 7.3 \times 10^{53}\;$erg implies $\delta\sim(E_{\rm k,iso}/E_{\rm\gamma,iso})^{1/3}\sim 3.6$ and $\gamma\sim(E_{\rm k,iso}/E_{\rm\gamma,iso})^{1/6}(\theta_{\rm obs}-\theta_0)^{-1}\sim 240$. This implies a (cosmological) rest frame $E_p \sim 390\;$keV, which falls closely within the observed $E_p-E_{\rm\gamma,iso}$ relationship reported by @Amati02, @L-RR-R02 and subsequently @Lamb05 using data from [*BeppoSAX*]{}, [ *BATSE*]{} and [*HETE-II*]{}, respectively. This relationship finds that in GRBs, $E_p \propto E_{\rm\gamma,iso}^{1/2}$, although a significant amount of outliers may be present due to selection effects [@Nakar05; @Band05]. Figure \[Amati\_relation\] shows the location of GRBs, XRFs and XRGRBs in the $E_p - E_{\gamma,{\rm iso}}$ plane. To date it has been difficult to extend this relationship into the XRF regime (especially at very low $E_p<10$ keV) since only one XRF in this spectral energy range that has a firmly established redshift [XRF 020903 at $z=0.251$; @Soderberg04]. On the other hand, the existence of XRF 030723 and XRF 020427 with $E_p^{\rm obs}<10\;$keV is not sufficiently constraining, since their redshift is not known. XRF 030723 {#XRF030723} ---------- XRF 030723 was also detected by the [*HETE-II*]{} satellite. It had an observed peak photon energy of $E_p^{\rm obs}=8.4^{+3.5}_{-3.4}\;$keV and a fluence of $f\approx 5.7\times 10^{-7}\;{\rm erg\; cm^{-2}}$ in the $2-400\;$keV range [@Butler04a]. No redshift determination has been made, although a firm upper limit of $z<2.3$ could be placed [@Fynbo04a]. [*Chandra*]{} observations of the X-ray afterglow were reported by @Butler04a. In the radio band, only an upper limit of $180\;\mu$Jy was reported at $8.46\;$GHz, $3.15\;$days after the event [@SBF03]. The optical transient was discovered by @Fox03, and extensive follow up in the optical and near-infrared was reported by @Fynbo04a. The well monitored R-band light curve is initially very flat[^4], with $\alpha\sim 0$ (where $F_\nu\propto t^{-\alpha}\nu^{-\beta}$). After about $1\;$day it steepens to $\alpha\approx 2$. This behavior is unusual for standard GRB light curves and allows one to constrain models of XRFs. @Fynbo04a already noted how the early time flattening of the light curve might be an indication of an off-axis jet. Between $1-4\;$days the optical spectral slope $\beta_{\rm op}$ was in the range $\sim 1.0-1.3$, which is not unusual for GRB afterglows. After about $\sim 10\;$days, a strong bump appeared in the optical light curve. This was assumed to be a SN component by @Fynbo04a, while @Huang04 interpreted it as an indication of a second jet, within the context of the two component jet model [see @RR02; @PKG05 and references therein]. The bump had a sharp rise and red colors [@Fynbo04a]. The sharp rise, with $\Delta t<t$, is hard to explain in both of these models, although @Tominaga04 were able to fit the sharp rise with models of SN light curves. The red colors arise naturally for a SN but are very hard to account for with a two component jet model, or for this matter also in other models for bumps in the afterglow light curve such as a density bump in the external medium, angular inhomogeneities in the jet (“patchy shell”), or a refreshed shock [@rm98; @pmr98; @rmr01; @wl; @enrico; @lazzati2; @heyl; @nakar; @mes98; @kumar]. Therefore, the SN explanation for the bump in the optical light curve seems to be favored by the data. The X-ray light curve consists of two points, at $3.2\;$days and $13.2\;$days. A joint fit for the spectral slope at these two epoch gives $\beta_X=0.9^{+0.3}_{-0.2}$, while the temporal index between these two points is $\alpha_X=1\pm 0.1$ [@Butler04a]. This is a significantly shallower decay compared to that in the optical prior to the bump ($\alpha_{\rm op}\approx 2$), and is therefore not easy to account for. Since the optical bump is most likely due to a SN component, the same physical component is not expected to contribute significantly to the X-ray flux. The shallower decay in the X-rays might be due to the contribution of synchrotron self-Compton (SSC) which can dominate the X-ray flux on time scales of days to weeks [@PK00; @SE01]. This would generally also decrease the value of the spectral slope, $\beta_X$, and therefore @Butler04a considered this option to be incompatible with the data. We performed a tentative fit to the data and demonstrate here that the observational constraints on the spectral slope can still be satisfied by this scenario (see Fig. \[fit XRF030723\]).[^5] The physical parameters of this fit are $z=0.8$, $E=1.0\times 10^{50}\;$erg, $n=4.5\;{\rm cm^{-3}}$, $p=2.36$, $\epsilon_B=0.012$, $\epsilon_e=0.13$, $\theta_0=2.9^\circ$, $\theta_{\rm obs}=2.03\theta_0$. We stress that the model parameters cannot be uniquely determined from the fit to the afterglow observations, and other sets of model parameters could provide an equally good fit to the data. Some features are, however, rather robust. Most noticeable is a viewing angle of $\theta_{\rm obs}\sim 2\theta_0$ which is required in order to reproduce the initially very flat part of the optical light curve. A narrow jet with $\theta_0$ of no more than a few degrees is required in order for the jet break time $t_j$ to be less than about a day, which is in turn needed in order to reproduce the steep decay in the optical light curve that starts after $\sim 1\;$day. A redshift of $z\lesssim 0.8$ is suggested by a fit of the late time bump in the optical light curve to core collapse SN light curves [@Tominaga04]. This in part motivated us to choose a redshift of $z=0.8$ for the fit that we present here, but fits for other values of $z$ are also plausible. A higher $z$ would require a higher jet energy $E$, while a lower $z$ would require a smaller jet energy. For $z\approx 0.8$, $E_{\rm\gamma,iso}\approx 9.3\times 10^{50}\;$erg which together with $E_{\rm k,iso} \approx 7.8\times 10^{52}\;$erg implies $\delta\sim 4.4$ and $\gamma\sim 40$. This would in turn imply a (cosmological) rest frame $E_p$ of $\sim 66\;$keV if viewed on-axis, which is a factor of $\sim 3$ lower than the value required to fall exactly on the Amati relation (see Fig. \[Amati\_relation\]). Given the large uncertainties associated with this relationship [@L-RR-R02; @Nakar05; @Band05], we consider this to be in good agreement with observations of on-axis GRBs. It is reasonable to expect a relatively low Lorentz factor ($\gamma\sim 40$) at the edge of the jet. Assuming $\gamma$ decreases from $\gamma_{\rm int}\gtrsim 100$ in the interior of the jet to much lower values at $\Delta\theta\gtrsim 1/\gamma_{\rm int}$ centered around $\theta_0$, then for $\gamma\lesssim 40$ the optical depth to pair production would be large, while for larger values of $\gamma$ much fewer photons would reach and off-axis observer, so that it is reasonable that the off-axis prompt emission will be dominated by $\gamma$ for which $\tau_{\gamma\gamma}$ is just smaller than 1. A similar result was obtained in a fit to GRB 031203 [@R-R05]. [^6] In addition to a uniform jet with sharp edges, we also consider other jet structures: (i) a narrow core with power law wings $\epsilon\approx\epsilon_0\min[1,(\theta/\theta_0)^{-3}]$, $\Gamma_0-1\approx 299\propto\min[1,(\theta/\theta_0)^{-2}]$, and (ii) a Gaussian jet with $\epsilon\propto\exp(-\theta^2/2\theta_0^2)$ and either a constant $\Gamma_0=200$ or a Gaussian $\Gamma_0-1$ \[i.e. $\Gamma_0=1+199\exp(-\theta^2/2\theta_0^2)$\]. The light curves for different viewing angles are shown in Fig. \[XRF030723\_4jets\]. For a jet with power law wings where $\epsilon\propto\theta^{-a}$ it is hard to reproduce the very flat light curve at early times that is observed in XRF 070323 (and in XRGRB 041006), even for $a\approx 3$ that is expected for the collapsar model [@ZWH04], and a steeper drop in $\epsilon$ (i.e. a larger value of $a$) is required. For a Gaussian jet, the light curves have a stronger dependence on the angular profile of the initial Lorentz factor, $\Gamma_0(\theta)$. If it is constant, then the deceleration time at large viewing angles is still very small, and the contribution to the observed flux from material along the line of sight dominates at early times out to reasonably large viewing angle $\theta_{\rm obs}/\theta_0\sim{\rm a\ few}$. If, on the other hand, it has a Gaussian profile, $\Gamma_0-1\propto\exp(-\theta^2/2\theta_0^2)$, then the deceleration time at large angles becomes large and the observed flux is dominated by emission from the jet core even at early times. This causes a rise in the observed flux at early times for viewing angles outside the core of the jet, similarly to a uniform jet viewed off-axis [@KG03], and in better agreement with the initially very flat light curves of XRF 070323 and XRGRB 041006. We consider a Gaussian profile for the kinetic energy per unit mass, $\Gamma_0-1$, to be more realistic than a constant $\Gamma_0$, since the latter requires a Gaussian profile for the rest mass per unit solid angle, $\mu$, that is entrained in the outflow \[since $\epsilon=(\Gamma_0-1)\mu c^2$\], while the former implies a constant $\mu$. If anything, one might expect $\mu$ to increase with $\theta$ rather than decrease with $\theta$ (since a larger amount of mass in the ejecta might be expected near the walls of the funnel). Thus, from the models we considered, a reasonable fit to the light curve of XRF 030723 (and XRGRB 041006) can be obtained either for a uniform jet with sharp edges viewed off-axis or for a Gaussian jet with a Gaussian profile in both $\epsilon$ and $\Gamma_0-1$ viewed from outside its core. The fact that the afterglow light curve of an XRGRB requires a viewing angle that is only slightly outside the edge of the jet, while the afterglow light curve of an XRF requires a larger viewing angle ($\theta_{\rm obs}\sim 2\theta_0$) provides a consistent picture where a roughly uniform jet with relatively sharp edges is viewed as a GRB from within the jet aperture \[i.e. $\gamma(\theta_{\rm obs}-\theta_0)\lesssim 1$\], as an XRGRB from slightly outside the edge of the jet \[i.e. $1\lesssim\gamma(\theta_{\rm obs}-\theta_0)\lesssim{\rm a\ few}$\], and as an XRF from yet larger off-axis viewing angles \[i.e. $\gamma(\theta_{\rm obs}-\theta_0)\gtrsim{\rm a\ few}$\]. Other events with sparse data {#other_obs} ----------------------------- Besides the two events discussed above (in §\[XRGRB041006\] and §\[XRF030723\]) there have been a few other XRFs with candidate afterglow detections. The data in these case are, however, too sparse to allow any meaningful constraint on theoretical models. XRF 020903, detected by [*HETE-II*]{}, had an exceptionally low peak energy of $\sim 5$ KeV. Detection of the optical and radio afterglow was reported by @Soderberg04, together with the identification of the likely host galaxy at $z=0.251$. Due to the large error box, and the proximity to two other transient sources (which delayed prompt identification), the optical light curve at early times was not well sampled. The first detection is at $t=0.9\;$days after the burst, while later observations are dominated by the light from the host. In contrast to the sparse optical measurements, the radio light curve was extensively monitored with VLA over the period 25-370 days. The source, which was monitored at frequencies of 1.5, 4.9, 8.5 and $22.5\;$GHz, was found to have a temporal index $\alpha$ similar to that of “standard” GRBs [@Frail03]. XRF 020427 was detected by [*BeppoSAX*]{}, and no redshift measurement is available. There is a detection of X-ray emission at $t<100\;$s and a later detection at $t\sim 1\;$day. If the last of the early time detections (at $t\sim 50\;$s) is indeed marking the begin of the afterglow [as suggested by @Amati04], then the inferred steep afterglow decline would be hard to reconcile with a sharp edge seen off-axis. However, given the lack of coverage, it is not clear whether the detection at $t\sim 50\;$s is indeed part of the afterglow or, instead, still a component of the prompt emission. In the latter situation, with only one X-ray detection available, there is not much that can be said in terms of possible models. Other cases with possible counterparts are XRF 040912, which has a candidate X-ray afterglow between $13.57\;$hr and $38.65\;$hr, and XRF 040916, which has an optical afterglow candidate but with no X-ray detection. Supernova signatures in XRFs {#SN} ---------------------------- The combined results on SN1998bw and SN2003dh offer the most direct evidence yet that typical, long-duration, energetic GRBs result from the deaths of massive stars [e.g., @Hjorth03; @Stanek03]. The lack of hydrogen lines in both spectra is consistent with model expectations that the star lost its hydrogen envelope to become a Wolf-Rayet star before exploding. The broad lines are also suggestive of an asymmetric explosion viewed along the axis of most rapid expansion [@Mazzali01; @ZWH04]. Despite the rather large uncertainty on the true event rate of GRBs, a comparison with the event rate of Type Ib/c SNe suggests that only a small fraction,[^7] $f_{\rm GRB}\lesssim 10^{-3}$, of such SNe produce GRBs. The Type Ic SNe that are firmly associated with GRBs are very bright Type Ic events, with SN 1998bw being the brightest. The lack, however, of a SN in GRBs 010921 [@p03] and 020410 [@levan04a] to a limit of $\sim$1.5 and $\sim 2$ magnitudes fainter than SN 1998bw, respectively, suggests that we may be seeing a broader luminosity function for the Type Ic SNe that are associated with GRBs. If the unification hypothesis discussed here is true (or in any model where GRBs and XRFs are intrinsically the same object), XRFs should be accompanied by a SN[@ZWH04] brightening in their afterglow light curves, as seen in GRBs. Unfortunately, the sample of XRFs with known redshifts and optical afterglows that are sufficiently well monitored is very limited, with one possible exception – XRF 030723 which has a well sampled light curve but no measured redshift. There are, nevertheless, both upper and lower limits on the redshift of XRF 030723. A lower limit of $z \gtrsim 0.3$ has been derived from the non-detection of its host galaxy [@Fynbo04a], while an upper limit of $z< 2.3$ was derived from the lack of Ly$\alpha$ absorption. @Fynbo04a obtained optical photometry and spectroscopy of XRF 030723, and found that the optical counterpart showed a “bump" in the light curve which may be the signature of a SN component. As discussed in §\[XRF030723\], the temporal and spectral energy distribution evolution are hard to reconcile with other interpretations such as a refreshed shock or a density variation in the external medium. For the redshift range $z \sim 0.3-1$, all possible SN models require a rather small mass of synthesized $^{56}$Ni [@Tominaga04]. This is because the SN brightness at this distances is $\sim 2$ magnitudes fainter than SN 1998bw. As the SN peak luminosity scales roughly linearly with its $^{56}$Ni yield, we would expect very little $^{56}$Ni production from a very faint SN. The SN associated with XRF 030723 therefore appears to have properties similar to those associated with GRB 010921 and GRB 020410, i.e. it seems to lie at the low end of the hypernova luminosity function, and is perhaps even closer in its properties to a normal Type Ic SN. This might potentially be caused by our off-axis viewing angle which resulted not only in an XRF instead of a GRB, but also in a dimmer SN as opacity effects prevented us from seeing the brightest part of the SN ejecta which lies along the rotational axis. @Nomoto03 find that for the SNe that are associated with GRBs (or hypernovae), a significant decrease in luminosity may occur only for viewing angles $\theta_{\rm obs}\gtrsim 30^\circ$. This is a direct consequence of the anisotropic distribution of the SN ejecta. We find, however, that for XRF 030723 $\theta_{\rm obs}\sim 2\theta_0\sim 6^\circ$, which is well below $30^\circ$. Thus, the SN associated with XRF 030723 is probably intrinsically dimmer than SN 1998bw. Clearly, more data on the SN-GRB/XRF connection are necessary before we can understand the full extent of the relation between these phenomena. There is already some tentative evidence that a number of XRFs (011030 and 020427), for which no optical afterglow was detected, also have no evidence for an associated SN [@levan04b]. SNe such as SN 1998bw would have been visible out to $z \sim 1.5$ in each case, while somewhat fainter SNe would have been visible to $z \sim 1$. Although it is possible that these XRFs lie at $z\gtrsim 1$, it is still puzzling given our attempt to tentatively identify GRBs, XRFs, and SNe as similar objects observed with small, medium, and large inclination, respectively. A possibility which can explain both a relatively low redshift and the absence of a SN detection is that the afterglows were heavily dust extinguished. In the off-axis jet model, prompt and intense X-ray/UV radiation from the reverse shock may efficiently destroy and clear the dust [@WD00; @FKR01; @PL02] in the circumburst cloud within the solid angle corresponding to the initial jet aperture, i.e. at $\theta<\theta_0$. This implies relatively little extinction for on-axis viewing angles, $\theta_{\rm obs}<\theta_0$ (practically no extinction of emission from $\theta<\theta_0$ and a gradual increase in the extinction as $\theta$ increases above $\theta_0$) but a relatively large extinction for off-axis viewing angles, $\theta_{\rm obs}>\theta_0$, especially for emission arising from $\theta>\theta_0$. Interestingly enough, in this case, there could be many more obscured XRF optical afterglows, compared to GRB optical afterglows. Conclusions {#conc} =========== The existing XRF models have been examined and their predictions tested against the afterglow observations of XRF 030723 and XRGRB 041006, the events with the best monitored afterglow light curves to date within their respective class. We find that most models failed to reproduce the very flat part observed in their early afterglow light curve. This behavior is, however, naturally produced by a uniform jet viewed off-axis (i.e. from $\theta_{\rm obs}>\theta_0$). The edge of the jet must be sufficiently sharp, so that the emission at early times would be dominated by the core of the jet, rather than by material along the line of sight. Even for a jet with a narrow core and wings where the energy per solid angle drops as $\epsilon\propto\theta^{-3}$, as expected in the collapsar model, the afterglow light curves at early times are not quite as flat as those observed in XRF 030723 and XRGRB 041006. A Gaussian jet can produce a sufficiently flat light curve at early times as long as both $\epsilon$ and $\Gamma_0-1$ have a Gaussian profile (but not for a constant initial Lorentz factor $\Gamma_0$; see Fig. \[XRF030723\_4jets\]). The afterglow light curve of XRGRB 041006 requires $(\theta_{\rm obs}-\theta_0)\sim 0.15\theta_0\sim 0.8\times 10^{-2}\;$rad, while that of XRF 030723 requires $(\theta_{\rm obs}-\theta_0)\sim\theta_0 \sim 3^\circ\sim 0.05\;$rad. This supports a unified picture for GRBs, XRGRBs and XRFs, where they all arise from the same narrow and roughly uniform relativistic jets with reasonably sharp edges, and differ only by the viewing angle from which they are observed. Within this scheme, GRBs, XRGRBs and XRFs correspond to $\gamma(\theta_{\rm obs}-\theta_0)\lesssim 1$, $1\lesssim\gamma(\theta_{\rm obs}-\theta_0)\lesssim{\rm a\ few}$, and $\gamma(\theta_{\rm obs}-\theta_0)\gtrsim{\rm a\ few}$, respectively. The empirical classification scheme by which an event is tagged as a GRB, XRGRB or XRF (see §\[clas\]) is rather arbitrary. Therefore there could be some cases where a jet that is viewed on-axis ($\theta_{\rm obs}<\theta_0$) will be classified as an XRGRB or XRF instead of as a GRB, or the opposite case in which a jet viewed off-axis ($\theta_{\rm obs}>\theta_0$) might be classified as a GRB instead of as an XRGRB or an XRF. A more physically motivated classification would be according to the ratio of the viewing angle $\theta_{\rm obs}$ and the jet half-opening angle $\theta_0$ \[e.g., on-axis events versus off-axis events, where off-axis events could further be classified according to the value of $\gamma(\theta_{\rm obs}-\theta_0)$\], instead of relying purely on spectral characteristics as in the present empirical scheme. Such a classification would, however, be much harder to implement as it is not a trivial task to accurately determine the viewing angle. Future observations with [*HETE-II*]{} and the recently launched [*Swift*]{} satellite, will allow us to further test this picture, and might also provide us with the necessary information to test the structure of the jet. The strongest constraints could be obtained from afterglow light curves of XRFs and XRGRBs that are well monitored from early times and at various frequencies (ranging from radio to X-rays). A useful complimentary method for constraining the jet structure is via the statistics of the observed jet break times $t_j$ in the afterglow light curves and the corresponding viewing angle $\theta_{\rm obs}$ in the universal structured jet model or the jet half-opening angle $\theta_0$ in the uniform jet model [@PSF03; @NGG04; @LWD04]. Similarly, the large statistical sample of GRBs and XRFs with redshift that will be available during the [*HETE-II/Swift*]{} era, will allow a reconstruction of the intrinsic luminosity function of the prompt emission. If GRBs, XRGRBs and XRFs are only a manifestation of the viewing angle for a structured, universal jet (whose wings are producing the XRFs), then no break would be expected in the luminosity function. On the other hand, if GRBs are the results of viewing angles that intersect the jet (whether structured or not), while XRFs and XRGRBs are off-axis events, then one would naturally expect a break in the luminosity function. @Guetta04 found that a luminosity function with a break is favored in order for the predicted rate of local bursts to be consistent with the observed rate. This also prevents the existence of an exceedingly large number of GRB remnants in the local Universe [@LP98; @PRL00]. The relative fraction of XRFs and XRGRBs to GRBs is also expected to be different in the various models [@Lamb05]. If indeed an XRF corresponds to $\gamma(\theta_{\rm obs}-\theta_0)\sim{\rm a\ few}$ and $(\theta_{\rm obs}-\theta_0)\lesssim\theta_0$, the the solid angle from which an XRF is seen scales as $\theta_0/\gamma$ or as $\theta_0$ for a constant $\gamma$ (at a constant distance to the source), while the solid angle from which a GRB is seen scales as $\theta_0^2$. Therefore, the ratio of solid angles for GRBs and XRFs scales as $\theta_0$, and more GRBs compared to XRFs would be seen for larger $\theta_0$. As the distance to the source increases, XRFs could be detected only out to a smaller off-axis viewing angle, while most GRBs would still be bright enough to be detected out to reasonably large redshifts. Therefore, the ratio of GRBs to XRFs should increase with redshift. Finally, if the true energy $E$ in the jet is roughly constant, then the maximal redshift out to which a GRB could be detected would decrease with $\theta_0$ since $E_{\rm\gamma,iso}\propto\theta_0^{-2}$. This would increase the statistical weight of narrow jets in an observed sample, as they could be seen out to a larger volume. We now briefly mention a few possible implication of the off-axis model for XRFs and XRGRBs. For sufficiently large viewing angles outside the edge of the jet, one might expect some decrease in the the variability of the prompt emission. This is since the width of an individual spike in the light curve scales as $\Delta t\propto\delta\sim[\gamma(\theta_{\rm obs}-\theta_0)]^2$ while the peak photon energy and fluence scale as $E_p\propto\delta^{-1}$ and $f\propto\delta^{-3}$, respectively. Since the interval between neighboring spikes in the light curve is typically comparable to the width of an individual spike, $\Delta t$, then if $\Delta t$ increases significantly for large viewing angles this would cause at least some overlap between different pulses which would smear out some of the variability. Thus one might expect XRFs to be somewhat less variable than GRBs, at least on average, where a lower variability might be expected for lower values of $E_p$. This may lead to a simple physical interpretation of the observed variability-luminosity relation in the prompt gamma-ray/X-ray emission [@FR00; @Reic01]. Another possible signature of the off-axis model for XRFs is in the reverse shock emission. If the reverse shock is at least mildly relativistic, then the optical flash emission would be less beamed than the prompt X-ray or gamma-ray emission, due to the deceleration of the ejecta by the passage of the reverse shock. This might cause the optical flash to be suppressed by a smaller factor relative to the gamma-ray emission, compared to the corresponding on-axis fluxes. Thus XRFs or XRGRBs might still show reasonably bright optical emission from the reverse shock, which might in some cases be almost as bright as for classical GRBs. Finally, XRFs and XRGRBs might also show a larger degree of polarization compared to GRBs (see §\[pol\]). An important conclusion from this study is that jet models in which $\epsilon$ and $\Gamma_0$ vary smoothly inside the jet, and where our lines of sight are within the jet, do not naturally reproduce the afterglow light curves of XRF 030723 and XRGRB 041006. The best example of such a model is the “universal structured jet” model [@LPP01; @RLR02; @ZM02b], where both $\epsilon$ and $\Gamma_0$ vary smoothly as a power law in $\theta$ (the usual assumption being that $\epsilon\propto\theta^{-2}$ outside of some core angle). This model fails to account for the very flat initial part of the afterglow light curves and its subsequent decay. A possible way around this problem might be to identify the flat part of the light curve with the passage of the break frequency $\nu_m$ through the optical band. This should be accompanied by a change in the optical spectral slope, and should not be observed in other frequency ranges such as the radio or X-rays. For XRGRB 041006 this may actually provide a viable explanation for the data (see Fig. \[XRGRB041006\_sph\]). For XRF 030723, however, a similar model fails because it does not reproduce both of the observed values of the temporal index $\alpha_{\rm op}$ (before or after the passage of $\nu_m$) or the observed spectral slope $\beta_{\rm op}$. One could in principle invoke both a jet break and the passage of a break frequency at roughly the same time, for a jet viewed on-axis. This would require that $\nu_m\sim\nu_c\sim\nu_{\rm op}$ at $t_0\sim t_j\sim 0.1-1\;$days, which is a large coincidence and is therefore unlikely.[^8] Even if this was the case, this assumption would be hard to reconcile with the measured optical spectral slope of $\beta_{\rm op}=0.96\pm 0.04$ at $t=1.13\;$days, as the spectral break frequencies would still be near the optical at that time, resulting in a smaller value of $\beta_{\rm op}$. The afterglow light curve of XRF 030723 therefore provides evidence against this class of models. We thank N. Butler, C. Kouveliotou, S. Woosley for useful discussions. This research was supported by US Department of Energy under contract number DE-AC03-76SF00515 (J.G.) and by NASA through a Chandra Postdoctoral Fellowship award PF3-40028 (E. R.-R.). Part of this work was done while E. R.-R. and J.G. were visiting the UCSC. [99]{} Amati, L. et al. 2002, A&A, 390, 81 Amati, L. et al. 2004, A&A submitted (astro-ph/0407166) Ayani, K., et al. 2004, GCN, 2779 Band, D. L., et al. 1993, ApJ, 413, 281 Band, D. L., & Preece R. D. 2005, submitted to ApJ (astro-ph/0501559) Barnard, V., et al. 2004a, GCN, 2774 Barnard, V., et al. 2004b, GCN, 2786 Barraud, C. et al. 2003, A&A, 400, 1021 Bikmaev, I., et al. 2004, GCN, 2826 Blain, A. W., & Natarajan, P. 2000, MNRAS, 312, L35 Bromm, V., & Loeb, A. 2002, ApJ, 575, 111 Butler, N., Dullinghan, A., Ford, P., Ricker, G., Vanderspek, R., Hurley, K., Jernigan, J., Lamb, D., Graziani, C. 2004a, ApJ in press (astro-ph/0401020) Butler, N., et al. 2004b, GCN, 2808 Covino, S., et al. 2004, GCN, 2803 Da Costa, P., & Noel, N. 2004, GCN, 2789 Dado, S., Dar, A. & De Rújula, A. 2004, A&A, 422, 381 Dalal, N., Griest, K., & Pruet, J. 2002, ApJ, 564, 209 D’Avanzo, P., et al. 2004, GCN, 2788 Dermer, C. D., Chiang J., & Böttcher, M. 1999, ApJ, 513, 656 Drenkhahn, G. 2002, A&A, 387, 714 Eichler, D., & Levinson, A. 2004, ApJ, 614, L13 Fan, Y. Z., Wei, D. M., & Wang, C. F. 2004, MNRAS, 351, L78 Fenimore, E. E., & Ramirez-Ruiz, E. 2000, (astro-ph/0004176) Ferrero, P., et al. 2004, GCN, 2777 Fox, D. B., et al. 2003, GCN, 2343 Frail, D. A., Kulkarni, S. R., Berger, E. & Wieringa, M. H. 2003, AJ, 125, 2299 Fruchter, A., Krolik, J. H., Rhoads, J. E. 2001, ApJ, 563, 597 Fugazza, D., et al. 2004, GCN, 2782 Fukushi, H., et al. 2004, GCN, 2767 Fynbo, J. P. U. et al. 2004a, preprint (astro-ph/0402240) Fynbo, J. P. U. et al. 2004b, GCN, 2802 Galassi, M., et al. 2004, GCN, 2770 Garg, A., et al. 2004, GCN, 2829 Ghirlanda, G., Ghisellini, G., & Lazzati, D. 2004, ApJ in press, astro-ph/0405602 Ghisellini, G., & Lazzati, D. 1999, MNRAS, 309, L7 Greco, G., et al. 2004, GCN, 2804 Granot, J. 2003, ApJ, 596, L17 Granot, J., & Loeb, A. 2003, ApJ, 593, L81 Granot, J., & Kónigl, A. 2003, ApJ, 594, L83 Granot, J., Kumar, P. 2003, ApJ, 591, 1086 Granot, J, Miller, M., Piran, T., Suen, W. M., & Hughes, P. A. 2001, in Gamma-Ray Bursts in the Afterglow Era, ed. E. Costa, F. Frontera, & J. Hjorth (Berlin: Springer), 312 Granot, J., Panaitescu, A., Kumar, P., Woosley, S. E. 2002, ApJ, 570, L61 Granot, J., & Sari, R. 2002, ApJ, 568, 820 Granot, J., & Taylor, G. B. 2005, ApJ in press (astro-ph/0412309) Granot, J., & Ramirez-Ruiz, E. 2004, ApJ, 609, L9 Gruzinov, A. 1999, ApJ, 525, L29 Guetta, D., Granot, J., & Begelman, M. C. 2005, ApJ in press (astro-ph/0407063) Guetta, D., Perna, R., Stella, L. & Vietri, M. 2004, ApJL, 615, 73 Heise, J., in’t Zand, J., Kippen, R. M., & Woods, P. M. 2001, in Proc. of the conference “Gamma-ray Bursts in the Afterglow Era”, 16 Heyl, J. S., & Perna, R. 2003, ApJ 586, L13 Hjorth, J., et al. 2003, Nature, 423, 847 Huang, Y. F., Dai, Z. G. & Lu, T. 2002, MNRAS, 332, 735 Huang, Y. F., Wu, X. F., Dai, Z. G., Ma, H. T., Lu, T. 2004, ApJ, 605, 300 Kahharov, B., et al. 2004, GCN, 2775 Kelson, D. D., Koviak, K., Berger, E., & Fox, D. B. 2004, GCN circ. 2627 Kinoshita, D., et al. 2004, GCN, 2785 Kippen, R. M., et al. 2003, in Gamma-Ray Bursts ans Afterglow Astronomy, AIP Conf. Proceedings 662, ed. G. R. Ricker & R. K. Vanderspek (New York: AIP), 25 Klotz, A., et al. 2004, GCN, 2784 Kouveliotou, C., et al. 2004, ApJ, 608, 872 Kumar, P., & Piran, T. 2000, ApJ 535, 152 Kumar, P., & Granot, J. 2003, ApJ, 591, 1075 Lamd, D. Q., et al. 2004, New Astron. Rev., 48, 423 Lamb, D., Q., Donaghy, T., Q. & Graziani, C. 2005, ApJ, 620, 355 Lazzati, D., Rossi, E., Covino, S., Ghisellini, G., & Malesani, D. 2002, A&A 396, L5 Levan, A. et. al. 2004, ApJ in press (astro-ph/0403450) Levan, A. et. al. 2004, ApJ in press (astro-ph/0410560) Levinson, A., Ofek, E. O., Waxman, E., & Gal-Yam, A. 2002, ApJ, 576, 923 Liang, E. W., Wu, X. F. & Dai, Z. G. 2004, MNRAS, 354, 81 Lipunov, V. M., Postnov, K. A., & Prokhorov, M. E. 2001, Astron. Rep., 45, 236 Lloyd-Ronning, N., Fryer, C., & Ramirez-Ruiz, E. 2002, ApJ, 574, 554 Lloyd-Ronning, N., & Ramirez-Ruiz, E. 2002, ApJ, 576, 101 Loeb, A. & Perna, R. 1998, ApJL, 503, 35 Lyutikov, M., Pariev, V. I., & Blandford, R. D. 2003, ApJ, 597, 998 Mazzali, P. A., Nomoto, K., Patat, F., & Maeda, K. 2001, ApJ, 559, 1047 Medvedev, M. V., & Loeb, A. 1999, ApJ, 1999, 526, 697 Mészáros, P., Rees, M. J., & Wijers, R. 1998, ApJ 499, 301 Mészáros, P., Ramirez-Ruiz, E., Rees, M. J., & Zhang, B. 2002, ApJ, 578, 812 Misra, K., & Pandey, S. B. 2004a, GCN, 2794 Misra, K., & Pandey, S. B. 2004b, GCN, 2795 Mochkovitch, R., Daigne, F., Barraud, C., & Atteia, J. L. 2003, ASP Conference Series (San Francisco ASP), in press (astro-phj/0303289) Moderski, R., Sikora, M., & Bulik, T. 2000, ApJ, 529, 151 Monfardini, A., et al. 2004, GCN, 2790 Nakar, E., Granot J. & Guetta, D. 2004, ApJ, 606, L37 Nakar, E., Piran, T., & Granot, J. 2003, NewA 8, 495 Nakar, E., & Piran, T. 2005, ApJ submitted (astro-ph/0412232) Nakar, E., Piran, T., & Granot, J. 2003, ApJ, 579, 699 Nakar, E., Piran, T., & Waxman, E. 2003, JCAP, 10, 5 Nomoto, K., Maeda, K., Mazzali, P. A., Umeda, H., Deng, J. & Iwamoto, K. 2003, to appear in “Stellar Collapse” (Astrophysics and Space Science; Kluwer) ed. C. L. Fryer Panaitescu, A., Mészáros, P., & Rees, M. J. 1998, ApJ 503, 314 Panaitescu, A., & Kumar, P. 2000, ApJ, 543, 66 Peng, F., Kónigl, A., & Granot, J. 2005 ,submitted to ApJ (astro-ph/0410384) Perna, R. & Lazzati, D. 2002, ApJ, 580, 261 Perna, R. & Loeb, A. 1998, 509L, 85 Perna, R., Raymond, J. & Loeb, A. 2000, ApJ, 533, 658 Perna, R., Sari, R. & Frail, D. 2003, ApJ, 594, 379 Price, P. et al. 2003, ApJ, 584, 931 Price, P. A., et al. 2004a, GCN, 2771 Price, P. A., et al. 2004b, GCN, 2791 Prochaska, J. X., et al. 2004, preprint (astro-ph/0402085) Ramirez-Ruiz, E., Dray, L. M., Madau, P., & Tout, C. A. 2001a, MNRAS 327, 829 Ramirez-Ruiz, E., Merloni A., & Rees M. J. 2001b, MNRAS 324, 1147 Ramirez-Ruiz, E., Celotti, A., & Rees, M. J. 2002, MNRAS, 337, 1349 Ramirez-Ruiz, E., & Lloyd-Ronning, N. M. 2002, New Astron., 7, 197 Ramirez-Ruiz, E., & Madau, E. 2004, ApJ, 608, L89 Ramirez-Ruiz, E., Granot, J., Kouveliotou, C., Woosley, S. E., Patel, S. K., & Mazzali, P. A. 2005, submitted to ApJL (astro-ph/0412145) Reichart, D., Lamb, D., Fenimore, E. E., Ramirez-Ruiz, E., Cline, T., & Hurley, K. 2001, ApJ, 552, 57 Rees, M. J., & Mészáros, P. 1998, ApJ 496, L1 Rhoads, J. E. 1997, ApJ, 487, L1 Rhoads, J. E. 2003, ApJ, 591, 1097 Richardson, D. Branch, D., Casebeer, D., Millard, J., Thomas, R.C., Baron, E. 2002, ApJ, 123, 745 Rossi, E., Lazzati, D. & Rees, M. J. 2002, MNRAS, 332, 945 Sakamoto, T., et al. 2004, submitted to ApJ (astro-ph/0409128) Sari, R. 1999, ApJ, 524, L43 Sari, R., & Esin, A. A. 2001, ApJ, 548, 787 Sari, R., Piran, T., & Narayan, R. 1998, , 497, L17 Smith, D. A., Akerlof, C. W., & Quimby, R. 2003, GCN circ. 2338 Soderberg, A. M., Berger, E. & Frail, D. A. 2003a, GCN circ. 2330 Soderberg, A. M., et al. 2004, ApJ, 606, 994 Soderberg, A. M., & Frail, D. 2004, GCN 2787 Stanek, K. Z., et al. 2003, ApJ, 591, L17 Thomsen, B., et al. 2004, preprint (astro-ph/0403451) Tominaga, N., et al. 2004, ApJ, 612, L105 Totani, T., & Panaitescu, A. 2002, ApJ, 576, 120 Wang, X., & Loeb, A. 2000, ApJ 535, 788 Waxman, E. 2003, Nature, 423, 388 Waxman, E., & Draine, B. T. 2000 , ApJ, 537, 796 Watson, D., et al. 2004, preprint (astro-ph/0401225) Woods, E., & Loeb, A. 1999, ApJ, 523, 187 Yamazaki, R., Ioka, K., & Nakamura, T. 2002, ApJ, 571, L31 Yamazaki, R., Ioka, K., & Nakamura, T. 2003, ApJ, 593, 941 Yamazaki, R., Ioka, K., & Nakamura, T. 2004a, ApJ, 606, L33 Yamazaki, R., Ioka, K., & Nakamura, T. 2004b, ApJ, 607, L103 Yost, S., et al. 2004, GCN, 2776 Zhang, B., Dai, X., Lloyd-Ronning, N. M., & Mészáros, P. 2004, ApJ, 601, L119 Zhang, B. & Mészáros, P. 2002a, ApJ, 581, 1236 Zhang, B. & Mészáros, P. 2002b, ApJ, 571, 876 Zhang, W., Woosley, S. E., & Heger, A. 2004, ApJ, 608, 365 Zhang, W., Woosley, S. E., & MacFadyen, A. I. 2004, ApJ, 608, 365 [^1]: http://space.mit.edu/HETE/Bursts/GRB041006/ [^2]: This is an approximate expression which is valid for a point source at the edge of the jet at the point closest to the line of sight, and gives reasonable off-axis light curves [@Granot02]. A more accurate calculation [e.g., @EL04] shows a more complex behavior. If one defines the local slope of the fluence, $f$, as a function of $\delta$, $a=-d\log f/d\log\delta$, then $a>3$ at very small off-axis angles $0<\gamma(\theta_{\rm obs}-\theta_0)\lesssim 1$, $a\approx 2$ at intermediate angles $\gamma^{-1}<(\theta_{\rm obs}-\theta_0)<\theta_0$, and $a\approx 3$ at $(\theta_{\rm obs}-\theta_0)\gtrsim \theta_0$. This is somewhat different from our simple power law approximation. However, since exact shape of the edge, as well as other model uncertainties, could introduce effects of similar magnitude to the difference between our simple power law approximation and the more accurate calculation, the former is sufficient for our purposes. [^3]: This is the ratio of the Doppler factor for a viewing angle along the edge of the jet (i.e. at the point where most of the off-axis emission comes from; the Lorentz factor $\gamma$ is that of the emitting fluid at the edge of the jet), $\gamma|\theta_{\rm obs}-\theta_0|\lesssim 1$, and the Doppler factor for an off-axis viewing angle $\theta_{\rm obs}$ which satisfies $\gamma(\theta_{\rm obs}-\theta_0)\gtrsim 1$. [^4]: ROTSE-III performed early unfiltered optical observation of XRF 030723 [@Smith03] and conclude that “We find no convincing evidence for a detection of the OT in the first four of our images, but the last two images do yield marginal possible detections”. Therefore, in what follows, we regard them as rough upper limits. [^5]: It is also roughly consistent with the single upper limit in the radio, since the observed frequency ($8.46\;$GHz) is somewhat below the self absorption frequency, and scintillations may further reduce the observed flux. [^6]: For XRGRB 041006 we obtain $(\theta_{\rm obs}-\theta_0)\sim 0.15\theta_0\sim 10^{-2}\;$rad and $\gamma\sim 240$ at the edge of the jet. The larger value inferred for $\gamma$ might be explained by the smaller value of the off-axis viewing angle, $\theta_{\rm obs}-\theta_0$, since such a line of sight which is significantly closer to the edge of the jet intersects the beaming cone of the emitting material near the edge of the jet up to a Lorentz factor of $\gamma\sim(\theta_{\rm obs}-\theta_0)^{-1}\sim 10^2$. [^7]: The estimates range from $f_{\rm GRB}\sim 10^{-5}$ for the universal structured jet model, to $f_{\rm GRB}\approx (0.6\pm 0.2)\times 10^{-3}$ for the uniform jet model [@GR-R04], where the latter is the relevant one for the off-axis jet model. [^8]: Such constraints would not leave enough free model parameters in order to also account for the temporal decay index in the X-rays, $\alpha_X$.
{ "pile_set_name": "ArXiv" }
--- author: - 'Andrey Katz,' - 'Joachim Kopp,' - Sergey Sibiryakov - and Wei Xue bibliography: - 'lensing.bib' title: Femtolensing by Dark Matter Revisited --- Introduction {#sec:intro} ============ The possibility that the dark matter (DM) in the Universe is made up of massive compact objects has been explored since the early days of the DM hunt. These searches have culminated in strong exclusion limits from the MACHO [@Allsman:2000kg], EROS [@Tisserand:2006zx], OGLE [@Wyrzykowski:2011tr] and HSC/Subaru [@Niikura:2017zjd] surveys. Recently, interest in compact DM objects has been revived, with the most popular candidates being primordial black holes (PBHs) [@Carr:1974nx; @Clesse:2015wea; @Carr:2016drx] or ultra-compact mini-halos (UCMHs) that are present for example in axion models [@Hogan:1988mp; @Kolb:1993zz; @Kolb:1993hw; @Hardy:2016mns]. While a wide range of experimental observations constrain compact objects with masses above $\sim 10^{-11} \, M_\odot$ (see [@Sasaki:2018dmp] for a recent review and summary of constraints), bounds on lighter objects are scarcer. Specifically for PBHs, the requirement that they should not evaporate within the lifetime of the Universe by emitting Hawking radiation sets a lower bound on their mass of about $10^{-18}M_\odot$. Moreover, the effect of their evaporation on Big Bang Nucleosynthesis and the extragalactic photon background become important already at PBH masses of order $10^{-17}\ M_\odot$, severely constraining this possibility [@Carr:2009jm]. Other bounds are based on the existence of old neutron stars [@Capela:2012jz; @Capela:2013yf; @Pani:2014rca; @Capela:2014qea] and white dwarfs [@Graham:2015apa] that would be destroyed if PBHs were abundant. Note that there is some controversy in the literature about the viability of neutron star constraints. While ref. [@Pani:2014rca] claims to identify a very efficient mechanism of PBH capture by neutron stars, the derivation was questioned in Refs. [@Capela:2014qea; @Defillon:2014wla]. On the other hand the constraints of ref. [@Capela:2013yf] hinge on the assumption that the globular clusters hosting neutron stars are embedded in overdense DM cores. While this assumption is not ruled out, it is also not experimentally supported [@Conroy:2010bs; @Ibata:2012eq]. Indeed, it has even been shown that globular clusters could be formed without any DM overdensities [@Naoz:2014bqa; @Popa:2015lkr]. None of the bounds discussed above applies directly to UCMHs, which below $10^{-11}M_\odot$ remain essentially unconstrained. To probe the mass range between $\sim 10^{-17} \, M_\odot$ and $\sim 10^{-13} \, M_\odot$, femtolensing of distant gamma-ray bursts (GRBs) has been proposed by Gould [@Gould:1991td] (see also ref. [@Nemiroff:1995ak] for a proposal along the same lines for somewhat higher masses). The basic idea behind femtolensing is that, while the two images of the GRB created by such a tiny lens cannot be resolved in space or time, their wave fronts will acquire different phases during propagation because of the different path lengths and gravitational potentials they experience. If the phase shift is of order one, one expects to see interference fringes in the frequency spectrum. Since setting meaningful limits with this method requires a sufficient number of GRBs, most works on femtolensing have been purely theoretical until very recently. An early experimental analysis based on BATSE data was presented in ref. [@Marani:1998sh], but yielded only very weak bounds. The situation has changed dramatically with the advent of Fermi data. There are now dozens of GRBs with reliably measured redshift and frequency spectrum, with hopes to significantly expand this sample in the near future. Fermi data has prompted first observational works on femtolensing of GRBs by black holes: The authors of ref. [@Barnacka:2012bm] claim to constrain primordial black holes with masses between $10^{-16}$ and $10^{-14} \, M_\odot$ to contribute no more than 10% to the total dark matter abundance in the Universe. Although the validity of these bounds has been questioned later because they are based on the assumption that the gamma-ray source is point-like in the lens plain [@Pani:2014rca; @Davidson:2016uok] (see also the relevant estimations of [@Barnacka:2014yja]), they have been widely accepted as the state of the art. In this work we critically re-analyze the idea of GRB femtolensing. We will use the techniques of [@1993ApJ...413L...7S], further elaborated in [@Matsunaga:2006uc]. One of the questions we study is the impact of the non-pointlike nature of GRBs. We will see that most GRBs are too big when projected onto the lens plane to yield meaningful femtolensing limits. Only a small population of GRBs with very fast variability might be suitable for such searches, but even their size is comparable to the Einstein radius of the lenses of interest to us. It must therefore be properly taken into account. A second question we address is to what extent femtolensing can be used to constrain non-point-like lenses, i.e., objects that extend beyond their Einstein radius. An important example are UCMHs consisting of DM. Indeed, it has been noticed in the 1990s that femtolensing might be relevant for probing UCMHs in the mass range between $10^{-16}$ and $10^{-12}$ solar masses [@Kolb:1995bu]. But this reference still treated the UCMHs as point-like, which is not necessarily in agreement with more recent estimates of their size [@Zurek:2006sy]. Femtolensing of UCMHs has not been re-analyzed since then, in spite of the significant progress that has been made both in our theoretical understanding of compact DM objects and in GRB observations. Besides accounting for the finite size of the source and lens, we will also emphasize that the geometric optics approximation used in the original proposal [@Gould:1991td] breaks down exactly in the mass range where femtolensing appears most promising. In other words, femtolensing in this mass range cannot be described by considering just two images of the source, with properties derived from the geometry of the corresponding lines of sight. Originally, the importance of wave optics effects in femtolensing had been pointed out in [@Ulmer:1994ij]. We will find that, especially because of the non-negligible size of the source, the data available to date does not constrain primordial black holes or other compact DM objects. We also compute the sensitivity of hypothetical future surveys as a function of the number of observed GRBs and their transverse size. The paper is organized as follow. In \[sec:theory\] we describe the theory of femtolensing, taking into the account all the caveats discussed above: finite size of the source, possibly finite size of the lens, breakdown of the geometric optics approximation. In \[sec:bounds\], we review the current femtolensing bounds on compact DM objects and show that they do not cover the physically interesting parameter region. We also show projections into the future, arguing that a sample of 100 GRBs with transverse size $10^9\,\text{cm}$ would be needed to exclude DM in the form of PBHs in the mass range from $10^{-16}$ to $5\times 10^{-15}$ solar masses. We conclude in Sec. \[sec:conclusions\]. Details on the size of the emission region in a GRB are relegated to the Appendix. Theory of Femtolensing {#sec:theory} ====================== In this section we review the general idea of femtolensing and the underlying formalism, and we address several important caveats. We will start by discussing the case of a point-like lens affecting light from a point-like source in the geometric optics approximation. We then introduce one-by-one the wave optics corrections, the effect of an extended source, and the possibility of an extended lens whose size exceeds its would-be Einstein radius. To the best of our knowledge, this study has not been performed before, but is strongly motivated by the appeal of ultra-compact DM miniclusters. Point-like Lens and Source, Geometric Optics Regime {#sec:basic-formalism} --------------------------------------------------- The basic femtolensing scenario, put forward in [@Gould:1991td], is based on the assumption that a gamma ray emitted by a point-like source with a non-zero impact parameter with respect to the lens–observer axis, is split by the lens into two rays, each of which is delayed with respect to the unlensed case by some time shift $\Delta t_i$ ($i=1,2$). This corresponds to a phase shift of $\Delta \phi_i \equiv \omega \Delta t_i$, where $\omega$ is the angular frequency of the photons. If the two images cannot be resolved in space and time, the two rays will interfere, producing characteristic fringes in the spectrum. In the thin lens approximation the time delay is given by [@Bartelmann:2010fz] $$\begin{aligned} \Delta t = \frac{1}{c} \frac{D_L D_S}{D_{LS}} (1 + z_L) \left( \frac{|\vec \theta - \vec \beta|^2}{2} - \psi (\vec \theta) \right) \,. \label{eq:timedelay}\end{aligned}$$ Here $D_L$, $D_S$, and $D_{LS}$ are the angular diameter distances between the observer and the lens, the observer and the source, and the lens and the source, respectively. The redshift of the lens is denoted $z_L$, while $\beta$ is the angle under which the observer would see the source in the absence of a lens, and $\theta$ is the angle under which the observer sees a given point in the lens plane. The function $\psi(\theta)$ is the lensing potential which is related to the density profile $\rho(r)$ of the lens by the Poisson equation $$\begin{aligned} \nabla^2 \psi(\theta) \equiv \frac{1}{\theta} \frac{\partial}{\partial\theta} \bigg( \theta \frac{\partial \psi}{\partial\theta} \bigg) = \frac{8\pi G}{c^2} \frac{D_{LS} D_L}{D_S} \int_{-\infty}^\infty \! d\xi \, \rho\big(\sqrt{(D_L \theta)^2 + \xi^2} \big) \,. \label{eq:lensing-potential}\end{aligned}$$ Here, the integral runs along the line of sight, and $\nabla^2$ is the two-dimensional Laplace operator, which we express in polar coordinates, with $\theta$ being the radial direction. We have also assumed a spherically symmetric lens. For a point-like lens of mass $M$,[^1] $$\begin{aligned} \psi(\theta) &= \theta_E^2 \log\theta & \text{(point-like lens)} \,,\end{aligned}$$ where $$\begin{aligned} \theta_E \equiv \left( \frac{4 G M}{c^2} \frac{D_{LS}}{D_S D_L} \right)^{1/2} \, \label{eq:theta-E}\end{aligned}$$ is the Einstein angle (i.e. the Einstein radius $R_E$ divided by $D_L$). On top of being a convenient definition, the Einstein angle has a well defined physical meaning: in the geometric optics approximation, it is the size of the Einstein ring produced by a point-like source aligned with the observer–lens line of sight ($\beta=0$). For the more general density profiles that we will discuss in \[sec:extended-lens\] the location of the ring will be different from $\theta_E$ and will be called $\theta_0$. For point-like masses, the two values coincide. In geometric optics, Fermat’s principle stipulates that the images will be seen under those angles $\theta$ for which $\Delta t$ is stationary. For point-like lens and source, this requirement leads to the lens equation[^2] $$\begin{aligned} \theta - \beta = \frac{\theta_E^2}{\theta} \,. \label{eq:lens-eq}\end{aligned}$$ Lensing is typically observable when $\theta \sim \theta_E$ (or, more generally, when $\theta\sim \theta_0$ for non-point-like lenses), otherwise one of the images becomes extremely faint. This shows that both terms in \[eq:timedelay\] scale as $\theta_E$ up to order-one factors. The time delay thus depends only on the mass of the lens and is practically insensitive to the distances either to the source or to the lens: $\Delta t \sim 4 G M/c^3 = 2 R_s / c$, where $R_s$ is the Schwarzschild radius of the lens. Assuming that the source emits gamma rays in the $10$ to $1000$ keV range, we see that order-one phase shifts occur for lens masses between $10^{-17}$ and $10^{-14}$ solar masses. For lensing at cosmological distances $\sim 1$Gpc the corresponding Einstein angles fall in the femto-arc-second range, which explains the term “femtolensing”. The solutions to the lens equation (\[eq:lens-eq\]) are $$\begin{aligned} \theta_\pm = \frac{1}{2} \left( \beta \pm \sqrt{\beta^2 + 4 \theta_E^2} \right) \,.\end{aligned}$$ The magnifications of the two images are given by [@Bartelmann:2010fz] $$\begin{aligned} \mu_\pm = \frac{y^2 + 2}{2 y \sqrt{y^2 + 4}} \pm \frac{1}{2} \,, \label{eq:magnification-geom-1}\end{aligned}$$ where we have defined $$\begin{aligned} y \equiv \beta / \theta_E\;. \label{ydef}\end{aligned}$$ Taking into account the phase shift between the images, the total intensity is proportional to [@Matsunaga:2006uc] $$\begin{aligned} \mu = \frac{y^2 +2}{y \sqrt{y^2 + 4}} + \frac{2}{y \sqrt{y^2 + 4}} \sin \left(\Omega \left[ \frac{y \sqrt{y^2 + 4}}{2} + \log \left| \frac{y + \sqrt{y^2 + 4}}{y - \sqrt{y^2 + 4}} \right| \right] \right) \,, \label{eq:magnification-geom-2}\end{aligned}$$ where we have introduced the dimensionless frequency as $$\begin{aligned} \Omega \equiv \frac{1}{c}\frac{D_S D_L }{D_{LS}} \theta_0^2 (1 + z_L ) \, \omega \equiv \Delta t_0 \, \omega \,. \label{eq:Omega_def_gen}\end{aligned}$$ Here we have also introduced a new quantity $\Delta t_0 $ which is a typical time delay between different images in the geometric optics picture. Note that this definition of the dimensionless frequency is completely generic and we are going to use it also for extended mass distributions in \[sec:extended-lens\]. For point-like lenses, however, \[eq:Omega\_def\_gen\] simplifies to $$\begin{aligned} \Omega \equiv \frac{4 G M (1 + z_L)}{c^3}\,\omega \,. \label{eq:Omega_def}\end{aligned}$$ clearly depends on the photon energy $\omega$ and varies between a maximum and minimum value as a function of $\omega$. Therefore a signal of femtolensing would be an oscillatory pattern of interference fringes in the otherwise smooth GRB spectrum. Let us list the conditions required for the validity of the above formulas. 1. The limit of geometric optics should be applicable.[^3] This is the case as long as $$\begin{aligned} \omega \, \Delta t_0 \gg 1 \label{eq:geomopt} \end{aligned}$$ (see \[sec:wave-optics\] for details). In the case of a point-like lens, this reduces (up to a factor $4 \pi (1+z_L)$) to the condition that the photon wave length should be much smaller than the Schwarzschild radius $R_s$ of the lens. Note, however, that this *does not mean* that the Schwarzschild radius can be interpreted as the effective radius of the lens, because the typical deflection distance of the photons is $R_E$ rather than $R_s$. Moreover, the relation \[eq:geomopt\] to the Schwarzschild radius is something very peculiar to point-like lenses and does not hold for more general mass distributions. A simple estimate shows that the inequality breaks down for $M \lesssim 10^{-15} M_\odot$ for a gamma ray energy of $E \sim 100$ keV, essentially invalidating the geometric optics approach in most of the parameter space that is relevant for femtolensing. The breakdown of the geometric optics approximation *does not* imply the absence of an observable femtolensing effect. It does, however, mean that a full wave optics computation is necessary (see \[sec:wave-optics\]). 2. The source should be point-like in the plane of the lens. The signals from different points of the source add up incoherently, and the condition that this does not wash out the interference fringes is $$\begin{aligned} \sigma_y \lesssim \frac{1}{\Omega}\;, \label{Deltay} \end{aligned}$$ where $\sigma_y$ is the angular size of the source normalized to $\theta_0$. This condition can be easily derived from \[eq:magnification-geom-2\]. Since interference fringes can only be observed if $\Omega \gtrsim 1$, condition also implies that the size of the source projected on the lens plane should be smaller than the Einstein radius for point-like lenses ($\theta_0$ for extended lenses). As we will see later, this condition is never fully satisfied for femtolensing of GRBs. At best, the projected size of the source can be comparable to the Einstein radius, so that finite size effects can never be neglected. Note also that the constraint on the size of the source is stricter for higher frequencies/energies. Therefore, as the size of the source increases, the interference fringes will first disappear at high frequencies. 3. The source emission must be coherent over timescale of the time delay, \[eq:timedelay\]. Usually this condition is easily satisfied because it merely requires a detector with sufficient energy resolution. Typically, we are interested in scenarios where $\omega \, \Delta t_0 \sim 1$. Given that any reasonable detector has an energy (or frequency) uncertainty $\delta\omega \ll \omega$, the Heisenberg uncertainty principle tells us that the coherence time $\delta t \sim \delta\omega^{-1}$ is much larger than $\Delta t_0$. Because the conditions (1) and (2) are never strictly satisfied, we are forced to relax some of the approximations made above. We will do so in the following subsections. Point-like Source and Lens in the Wave Optics Regime {#sec:wave-optics} ---------------------------------------------------- Gravitational lensing outside the geometric optics regime is not an unusual scenario. For example, it has been mentioned in [@Inomata:2017vxo] that the microlensing measurements of HSC Subaru are partially in the physical (wave) optics regime, invalidating the constraints on primordial black holes published in ref. [@Niikura:2017zjd] for masses below $10^{-11} M_\odot$. To the best of our knowledge, the correct interpretation of the HSC constraints in this mass range is still missing in the literature. Also, practically all discussion of lensing of gravitational waves includes wave optics effects [@Takahashi:2003ix] (for recent related works see e.g. [@Jung:2017flg; @Christian:2018vsi; @Li:2018prc]). Finally, it was noticed already in the 1990s that the geometric optics approximation is typically violated in femtolensing [@Ulmer:1994ij]. ![Comparison of the interference picture in the geometric optics approximation (dashed) and the full wave optics results (solid) in the case of a point-like source and a point-like lens. Wave effects strongly impact the magnification of the signal at low energies, while at higher energies, geometric optics provides a good approximation. The magnification is plotted against the dimensionless frequency $\Omega$ defined in Eq. .[]{data-label="fig:BH_phys_vs_geom"}](magnification-point-like){width=".48\textwidth"} In the physical optics regime, lensing is characterized by a magnification function $F(\vec y;\omega)$, which is defined as the ratio $\phi_L / \phi$, where $\phi_L$ and $\phi$ are the electromagnetic wave amplitudes with and without lensing, respectively. The magnification of the signal intensity is therefore $$\begin{aligned} \mu = |F|^2 \,, \label{eq:magnification-wave}\end{aligned}$$ The magnification function is given by [@Nakamura:review] $$\begin{aligned} F(\vec y;\omega) = \frac{\Omega}{2 \pi i} \int \! d^2\vec{x}\; e^{i \omega \Delta t(\vec x, \vec y)} \,. \label{eq:physopt}\end{aligned}$$ where the dimensionless frequency $\Omega$ has been defined in \[eq:Omega\_def\_gen\], $x \equiv \theta / \theta_0$, and the time delay $\Delta t(\vec{x}, \vec{y})$ is taken from \[eq:timedelay\]. The integral thus runs over the lens plane, with each point in the plane contributing to $F(\vec{y},\omega)$. For a point-like lens like a black hole, the integral can be evaluated analytically, leading to $$\begin{aligned} F(y, \Omega)_\text{BH} = e^{i \Omega |\vec y|^2/2} \, \left(-\frac{ i \Omega}{2}\right)^{i \Omega/2} \, \Gamma \left( 1 -\frac{i \Omega}{2} \right) L_{-1 + \frac{i \Omega}{2}} \left(-\frac{i |\vec y|^2 \Omega}{2} \right) \,, \label{eq:Fy}\end{aligned}$$ where $L_n$ is the $n$-th order Laguerre polynomial. For an extended lens, analytic expressions in general do not exist, so the integral in \[eq:physopt\] must be evaluated numerically. The geometric optics regime corresponds to the saddle point approximation of : as discussed in \[sec:basic-formalism\], the lens equation (\[eq:lens-eq\]) is obtained by imposing the stationary point condition $\vec \nabla_x \Delta t = 0$. Also the magnification $\mu$ in the geometric optics limits, \[eq:magnification-geom-1\], can be derived from the saddle point method. However, the saddle point method only provides a good approximation to the full integral if the phase is large, which explains why geometric optics only works if $\omega\, \Delta t \gg 1$. In Fig. \[fig:BH\_phys\_vs\_geom\], we compare the intensity calculated using the exact wave optics expression, \[eq:physopt\], to the one obtained in the geometric optics approximation, \[eq:magnification-geom-2\]. We see that the geometric optics approximation overestimates the amplitude of the first fringe in the spectrum, while at higher photon energies, it is applicable. We will see in \[sec:extended-lens\] that the effects of wave optics are more important for extended lens profiles. Extended Sources {#sec:extended-source} ---------------- As argued above, interference patterns due to femtolensing are only observable if the size of the source projected onto the lens plane is not much larger than the Einstein radius of the lens (assumed to be point-like in this subsection). In other words, we require that $$\begin{aligned} \sigma_y \equiv \frac{a_S}{D_S \theta_E} \ll 1 \,, \label{sigmay}\end{aligned}$$ where $a_S$ is the actual (unprojected) transverse size of the emission region. Most GRBs have an $\mathcal{O}(1)$ redshift, corresponding to $D_S \sim \text{Gpc}$. Therefore a lens of mass $M \sim 10^{-15} M_\odot$ that is also located at $z_L \sim \mathcal{O}(1)$ has an Einstein radius of order $R_E \sim 10^{9}$ cm. The Einstein radius of lighter lenses will be even smaller. Unfortunately, the emission size of most GRBs is much larger than this [@Barnacka:2014yja; @Golkhou:2015lsa]. As discussed in \[app:GRBsize\], the majority of GRBs are likely to have transverse sizes of about $a_S \sim 10^{11}\,\text{cm}$, which is two orders of magnitude larger than required for observable femtolensing. Still, the possibility that a small fraction of GRBs have much smaller sizes $a_S\lesssim 10^9 \, \text{cm}$ is not excluded. We will entertain this possibility and study the dependence of the femtolensing signal on the size of the source. Hopefully, future developments in GRB modeling, perhaps on a case-by-case basis, will lead to more precise estimates of their sizes than are currently available, and will allow us to cherry-pick events suitable for femtolensing. Note that, if such GRBs exist, they will be characterized by very fast intrinsic variability at time scales[^4] $t_{\rm var}\lesssim 0.5\times10^{-3}$ sec. To take the finite size of the source into account, we follow the formalism of ref. [@1993ApJ...413L...7S]. The observed magnification is then $$\begin{aligned} \bar\mu = \frac{\int \! d^2y \, W(\vec y; \sigma_y) \, \mu(\vec y; \Omega)} {\int \! d^2y \, W(\vec y; \sigma_y)} \,,\end{aligned}$$ where $W(y; \sigma_y)$ is a window function that describes the intensity profile of the emission and $\mu(\vec y; \Omega)$ is the magnification for a point-like source, see eqs. (\[eq:magnification-wave\]), (\[eq:physopt\]). In principle $W(\vec y)$ can be any well-behaved function that acts as a mask of size $\sigma_y$. We choose it to be a Gaussian, $$\begin{aligned} W(y; \sigma_y) = e^{-|\vec y - \vec y_0|^2/2\sigma_y^2} \,.\end{aligned}$$ Hereafter, we will use $\vec y_0$ to denote the location of the center of the emission. For Gaussian $W(y; \sigma_y)$ and a radially symmetric lens, the weighted magnification reads $$\begin{aligned} \label{mu-extended-source} \bar\mu = \frac{e^{-y_0^2/2\sigma_y^2}}{\sigma_y^2} \int_0^\infty \! dy \, y \, e^{-y^2/2\sigma_y^2}\, I_0 \bigg( \frac{y_0\, y}{\sigma_y^2} \bigg) \, \mu(y; \Omega)\end{aligned}$$ where $I_0(x)$ is the modified Bessel function of the first kind. ![ Dependence of magnification on the size of the source. Here $\sigma_y$ is the angular size of the source in units of the Einstein angle, see \[sigmay\]. The dimensionless frequency $\Omega$ is defined in \[eq:Omega\_def\]. The lens is assumed to be point-like.[]{data-label="fig:BH_phys_extended"}](magnification-extended-source){width=".48\textwidth"} We illustrate the effect of the non-zero source size in \[fig:BH\_phys\_extended\]. Clearly, the effect is mild for $\sigma_y \ll 1$, however even for these values of the emission size the oscillations in energy space are damped at high frequencies. As the emission size grows, oscillations are damped more strongly, until they eventually disappear at $\sigma_y \sim 1$. Interestingly, we observe that, even at large emission size $\sigma_y\gtrsim 1$, the asymptotic $\bar\mu$ at high frequency is larger than the value of $\bar\mu$ at low frequency. That is, even though the interference fringes are washed out, a smooth step-like feature survives. This can be understood as follows. At $\Omega \to \infty$ the point-source magnification $\mu(y;\Omega)$ is well described by the geometric optics expression (\[eq:magnification-geom-2\]). It quickly oscillates as a function of $y$ and can be replaced in the integral (\[mu-extended-source\]) by its mean value, $$\begin{aligned} \mu(y; \Omega) \mapsto \frac{y^2+2}{y\sqrt{y^2+4}} \,.\end{aligned}$$ Thus, in the high frequency limit the weighted magnification reads $$\begin{aligned} \lim_{\Omega\to\infty}\bar\mu= \frac{e^{-y_0^2/2\sigma_y^2}}{\sigma_y^2} \int_0^\infty \! dy \, \frac{y^2+2}{\sqrt{y^2+4}} \, e^{-y^2/2\sigma_y^2}\, I_0 \bigg( \frac{y_0\, y}{\sigma_y^2} \bigg) > 1\;.\end{aligned}$$ On the other hand, in the limit $\Omega \to 0$ the geometric optics approximation is not applicable and the wave optics calculation leads to $\lim_{\Omega \to 0} \bar\mu = 1$. The difference between the two limiting values leads to the step-like feature in the weighted magnification function at low dimensionless frequencies observed in \[fig:BH\_phys\_extended\]. Of course, in the absence of oscillatory fringes, one probably cannot rely on this feature to establish that a given GRB spectrum has been lensed. On the other hand, a non-observation of such feature can potentially be used to *exclude* lensing. Since our proposal to perform an exclusion based on the absence of this feature is rather speculative, we will not use it in estimating the future reach in \[sec:bounds\]. However, if proven possible, this would imply that even $\gamma$-ray sources as big as $10^{10}$cm might be used to exclude femtolensing. Extended Lens {#sec:extended-lens} ------------- We finally analyze the effects of a non-point-like lens, namely a lens whose extent is larger than its would-be Einstein radius. A notable example of such lenses would be ultracompact minihalos (UCMHs) composed of DM. These are predicted, for instance, in scenarios that involve DM in the form of QCD axions if the Peccei–Quinn phase transition happens after inflation. In this case, overdensities that arise after the phase transition due to different initial values of the axion field in different Hubble patches collapse into UCMHs around matter–radiation equality [@Hogan:1988mp; @Kolb:1993hw; @Zurek:2006sy]. The mass and radius estimates of UCMHs vary widely in the literature. For example, an average value for the mass of the QCD axion miniclusters as small as $10^{-14}~M_\odot$ was suggested in [@Enander:2017ogx]. This is two orders of magnitude lighter than the previous estimates of [@Kolb:1995bu; @Tinyakov:2015cgg; @Davidson:2016uok; @Bai:2016wpg]. Even bigger masses were suggested in [@Fairbairn:2017dmf; @Fairbairn:2017sil]. Although resolving these discrepancies is beyond the scope of this work, some part of the suggested parameter space is likely to be within the reach of femtolensing searches, as has been first pointed out in [@Kolb:1995bu]. UCMHs are not unique to QCD axions and can be formed also in many other models with axion-like particles [@Hardy:2016mns]. The parameter space explored in ref. [@Hardy:2016mns] is vast, and some of it is definitely accessible to femtolensing, in particular axion-like particles with temperature-independent ($n=0$) masses of $\sim 10^{-3}$–$10^{-5}$ eV. Another parameter subject to significant uncertainties is the radius of the UCMHs. However, what is important for our discussion is that it is most likely bigger than its would-be Einstein radius. For instance, for the QCD axion, assuming miniclusters are spherically symmetric, refs. [@Kolb:1995bu; @Zurek:2006sy; @Tinyakov:2015cgg] give, $$\begin{aligned} R_\text{UCMH} \simeq \frac{3\times 10^{12} \; \text{cm}}{\Phi(1+\Phi)^{1/3}} \left( \frac{M}{10^{-12} M_\odot} \right)^{1/3} \,, \label{eq:R-UCMH}\end{aligned}$$ where $\Phi$ is the initial axion density contrast. Most miniclusters will have $\Phi\lesssim 10$ [@Kolb:1995bu], so their radius according to \[eq:R-UCMH\] is almost two orders of magnitude larger than their corresponding Einstein radius. Only for extremely dense miniclusters with $\Phi\gtrsim 100$, the size and the Einstein radius become comparable. It is evident from these estimates that one cannot neglect the extent of the UCMHs in the lensing problem. Finally, there is no agreement in the literature on the density profile of UCMHs, which may also depend on the mechanism by which they form in the early Universe. Some scenarios suggest rather steep profiles. For example, the self-similar infall scenario motivates a density profile of the form $\rho(r)\propto r^{-9/4}$ [@Fillmore:1984wk; @Bertschinger:1985pd]. This profile has been confirmed in N-body simulations [@Zurek:2006sy; @Vogelsberger:2009bn; @Vogelsberger:2010eh], but its relevance to UCMHs was later questioned in [@Delos:2017thv]. The latter paper advocates more shallow profiles, $\rho(r) \propto r^{-3/2}$, or the Navarro–Frenk–White (NFW) profile which scales as $\rho(r) \propto r^{-1}$ at small radii. Given this uncertainty, we prefer to be agnostic about the precise shape of the UCMH profile. Rather we will assume that its inner part is described by a generic power law cusp, $$\begin{aligned} \rho(r) = \rho_0 \left( \frac{r}{r_0}\right)^{-\delta}~,~~~~\delta<3 \,, \label{eq:self-similar}\end{aligned}$$ where $r_0$ characterizes the size of the inner part of the UCMH and $\rho_0$ is the density at that distance.[^5] It is convenient to introduce the mass enclosed within radius $r_0$, which we will call the “cusp mass”, $$\begin{aligned} M_\text{cusp} = \frac{4\pi}{3-\delta}\rho_0 r_0^3 \,.\end{aligned}$$ Note that the total mass of UCMHs can, in general, be bigger than $M_\text{cusp}$ and can, in fact, even be formally divergent. The Einstein radius corresponding to $M_\text{cusp}$ is $$\begin{aligned} R_{E,\,\text{cusp}} = \bigg( \frac{4 G M_\text{cusp}}{c^2} \frac{D_{LS} D_L}{D_S} \bigg)^{1/2} \,. \label{eq:rE-UCMH}\end{aligned}$$ It is this radius that we mean when referring to the Einstein radius of the UCMH. We remind the reader that for extended lenses, the Einstein radius does not have an immediate physical interpretation. We assume $R_{E,\,\text{cusp}} < r_0$, such that the cusp cannot be treated as point-like for the purposes of gravitational lensing. It is easy to see that only profiles with $\delta>1$ can be relevant for femtolensing. Indeed, the mass enclosed within a radius $r$ grows as $M(r) \propto r^{3-\delta}$ and the corresponding Einstein radius scales as $R_E(r) \propto r^{(3-\delta)/2}$. If $\delta \leq 1$, any part of the lens is bigger than its Einstein radius, $R_E(r) < r$, so no multiple images and thus no femtolensing can arise. On the other hand, for $\delta>1$, the central part of the cusp happens to be within its Einstein radius. Thus, it acts qualitatively similar to a point-like lens leading to appearance of multiple images. We now study the case $\delta>1$ in more detail. Following \[eq:lensing-potential\], the lensing potential corresponding to the density profile \[eq:self-similar\] is $$\begin{aligned} \psi(\theta) = \frac{\theta_0^2}{3-\delta} \bigg( \frac{\theta}{\theta_0} \bigg)^{3-\delta}\,, \label{eq:psi-self-similar}\end{aligned}$$ where we have defined $$\begin{aligned} \theta_0 = \kappa(\delta)\, \frac{R_{E,\,{\rm cusp}}}{D_L} \bigg( \frac{R_{E,\,{\rm cusp}}}{r_0} \bigg)^{\frac{3-\delta}{\delta-1}}\,, \qquad \kappa(\delta)\equiv \bigg( \frac{\sqrt{\pi}\,\Gamma\big((\delta-1)/2\big)}{2\,\Gamma(\delta/2)} \bigg)^{\frac{1}{\delta-1}} \;. \label{eq:theta-0-UCMH}\end{aligned}$$ The angle $\theta_0$ is an analog of the Einstein angle for extended lenses. In the geometric optics approximation it coincides with the angular size of the Einstein ring produced by a point-like source aligned with the center of the lens ($\beta=0$ in the notations of section \[sec:basic-formalism\]). For misaligned sources ($\beta\neq 0$), $\theta_0$ sets the characteristic distance between their multiple images. The characteristic time delay $\Delta t_0$ is defined in . Note that $\theta_0$ is parametrically smaller than the naive Einstein angle $R_{E,\,\text{cusp}}/D_L$. For our subsequent discussion, we need to distinguish between the cases $\delta < 2$ and $\delta > 2$. For $1 < \delta < 2$ and small enough $\beta$, the lens equation $\vec \theta-\nabla\psi(\vec\theta)=\vec\beta$ has three solutions corresponding to three images in the geometric optics approximation. Above a certain critical value $\beta_\text{cr} = C(\delta) \theta_0$, where $C(\delta)$ is an order-one coefficient, two of the images disappear and only a single one remains. In other words, the ring $\beta = \beta_\text{cr}$ is a caustic. Thus, the appearance of interference fringes characteristic for femtolensing is possible only if the source is close enough to the line of sight passing through the center of the lens. Then, the existence of more than two images will in general lead to a complicated interference pattern (cf. ref. [@Ulmer:1994ij]). The properties of a lens with $2<\delta<3$, including the self-similar infall profile with $\delta = 9/4$, are closer to those of point-like objects like PBHs. In this case the lens equation always has two solutions corresponding to two images in the geometric optics approximation. This will give rise to the characteristic sinusoidal dependence of the magnification on frequency. The caustic shrinks to the point $\beta=0$, at which the two images turn into an Einstein ring of angular radius $\theta_0$.[^6] ![True magnification vs. geometric optics approximation for the self-similar infall profile $\rho(r) \sim r^{-9/4}$ as a function of the dimensionless frequency $\Omega=\omega \, \Delta t_0$, see \[eq:Omega\_def\_gen\]. We see that, for impact parameter $y\equiv\beta/\theta_0=1$, the picture is qualitatively quite similar to that for a point-like mass. At larger $y$, however, the geometric optics approximation significantly *underestimates* the magnitude of the lensing effect in a wide frequency range.[]{data-label="fig:94_phys_vs_geom"}](magnification-ucmh){width="48.00000%"} Wave optics effects quantitatively modify the geometric optics results at low frequencies, but do not lead to qualitative changes. We have computed the magnification numerically for the case $\delta=9/4$ using \[eq:magnification-wave,eq:physopt\]. The results are shown in \[fig:94\_phys\_vs\_geom\], and are also compared to the geometric optics approximation. As expected from the above discussion, the behavior of the curves is similar to the case of a point-like lens, but it nevertheless differs in some important details. First, wave optics corrections remain sizeable up to higher energies. Even more important, for impact parameters $y \equiv \beta/\theta_0 > 1$, we find that the geometric optics approximation significantly *underestimates* the amplitude over the first 5–7 periods of the oscillation. As the impact parameter grows, the amplitude of the oscillations falls off more slowly towards higher frequency than for a point-like lens, increasing the lensing probability for UCMHs compared to PBHs. Comparing \[fig:BH\_phys\_vs\_geom,fig:94\_phys\_vs\_geom\], we observe that the interference fringes for UCMHs are shifted to higher energies compared to the case of PBHs. This can be understood by noting from \[eq:lensing-potential\] that a photon passing the lens at a distance $\theta D_L$ is affected by DM particles residing at radii less than $\theta D_L$. The characteristic distance at which lensed photons pass an UCMH is $\theta_0 D_L$ (see \[eq:theta-0-UCMH\]). So if we call the mass contained within this radius $m(\theta_0)$, we expect the interference pattern for an UCMH to be comparable to the one for a PBH with mass $m(\theta_0)$ (modulo obvious differences in the lensing potential and the reduced deflection angle). We can thus estimate that the interference fringes in \[fig:94\_phys\_vs\_geom\] ($\delta = 9/4$) should be shifted to higher energies by a factor $$\begin{aligned} \frac{m(\theta_0)}{M_\text{cusp}} = \bigg( \frac{\theta_0 D_L}{r_0} \bigg)^{3/4} \,. \label{eq:m-over-M-UCMH}\end{aligned}$$ compared to the ones in \[fig:BH\_phys\_vs\_geom\]. For $z_S = 1$ and $z_L = 0.5$, this becomes $$\begin{aligned} \frac{m(\theta_0)}{M_\text{cusp}} = 0.037 \times \bigg( \frac{1.3 \times 10^{10}\,\text{cm}}{r_0} \bigg)^{6/5} \bigg( \frac{M_\text{cusp}}{10^{-15} M_\odot} \bigg)^{3/5} \,. \label{eq:m-over-M-UCMH-numerical}\end{aligned}$$ This estimate is indeed in good agreement with the shift observed in the plots. Revision of Current Bounds and Sensitivity Estimates {#sec:bounds} ==================================================== In this section, we will revisit the femtolensing bounds from ref. [@Barnacka:2012bm] by considering wave optics corrections (see \[sec:wave-optics\]) as well as the non-pointlike nature of the GRB sources (see \[sec:extended-source\]). From Fig \[fig:BH\_phys\_vs\_geom\] we expect that wave optics effects will modify femtolensing bounds on point-like masses by at most a few tens of per cent compared to the geometric optics approximation. The finite size of the sources, however, is expected to lead to much more dramatic modifications. As the size $a_S$ of the emission region in a GRB is very uncertain (see \[app:GRBsize\]), we will investigate the dependence of our results on $a_S$. We will find that only if $a_S \lesssim 10^8$ cm, current data is able to set meaningful limits. This is true for point-like lenses such as primordial black holes, but also for extended lenses like axion miniclusters. Unfortunately the assumption $a_S \lesssim 10^8$ cm is not realistic. We will establish these conclusions by investigating the *sensitivity* of current data, i.e. by working with simulated GRB data rather than real data. This will make our analysis more transparent, and our conclusion will be that current data is not sensitive to DM femtolensing yet. Therefore it is not necessary to go beyond a sensitivity study, and we will therefore not analyze actual Fermi GRBs data. We will, however, extrapolate our results into the future and estimate how many well observed GRBs would be needed for femtolensing to become a competitive player in the hunt for compact DM objects. In the following, we will first describe how we model GRB spectra (\[sec:grb-models\]). We will then describe the statistical methods we use (\[sec:likelihood\]) and discuss the resulting sensitivity estimates for PBHs and UCMHs (\[sec:sensitivity-results\]). GRB models {#sec:grb-models} ---------- We simulate “data” based on a phenomenological model for the unlensed GRB spectrum. In particular, we use Band’s model (BAND) [@Band:1993eg] as our baseline scenario, but we have also studied a broken power law model (BKN) as well as a simple power law model with an exponential cutoff as a cross check. The BAND model has four free parameters: an amplitude $A$, two spectral indices $\alpha_1$ (low energy) and $\alpha_2$ (high energy), and an energy scale $E_0$. In terms of these parameters, the spectrum as a function of energy $E$ is given by $$\begin{aligned} f_\text{BAND}(E) &= \begin{dcases} A \, (E/E_0)^{\alpha_1} \exp(-E/E_0) & \text{if $E \leq (\alpha_1 - \alpha_2) E_0$} \\ A \big[(\alpha_1 - \alpha_2) \big]^{\alpha_1-\alpha_2} (E/E_0)^{\alpha_2} \exp(\alpha_2 - \alpha_1) & \text{otherwise} \end{dcases} \,. \label{eq:band}\end{aligned}$$ In our sensitivity studies, we choose $A = 0.15~\text{counts}\ \text{sec}^{-1}\,\text{cm}^{-2}\,\text{keV}^{-1}$, $E_0 = 160$ keV, $\alpha_1 = -0.9$, and $\alpha_2 = -2.5$. These parameters are based on a fit to Fermi GBM data on GRB 090424 [@Bhat:2016odd], and we have normalized the spectrum to 5000 photons in the energy range from 8 keV to 550 keV on which we will focus in our analysis. This normalization is roughly based on the sample of high quality events that the Fermi GBM can collect for a typical short GRB at $z_S = 1$. We treat the detector’s effective area as a constant, $100~\text{cm}^2$, and we assume the GRB to last $1~\text{sec}$. These are typical values for short GRBs, which may appear more interesting for femtolensing because of the arguments given in \[app:GRBsize\] that indicate that the size of the emission region, $a_S$, tends to be smaller for short GRBs. Nonetheless, given the possibility of a broad distribution of $a_S$ for long GRBs, their larger abundance and higher redshift, it is unclear which type of GRB will ultimately offer the best sensitivity. The BKN model has four free parameters as well: an amplitude $A$, a characteristic energy $E_0$, and two spectral indices $\alpha_1$, $\alpha_2$: $$\begin{aligned} f_\text{BKN}(E) &= \begin{dcases} A \left( \frac{ E}{ E_0} \right)^{ -\alpha_1} & \text{if } E \leq E_0 \\ A \left( \frac{ E}{ E_0} \right)^{ -\alpha_2} & \text{if } E > E_0 \end{dcases} \,.\end{aligned}$$ The benchmark parameters for the BKN model are $A = 0.099~\text{counts}\ \text{sec}^{-1}\,\text{cm}^{-2}\,\text{keV}^{-1}$, $E_0 = 160$ keV, $\alpha_1 = -0.9$, and $\alpha_2 = -2.5$. The power law model with exponential cutoff has only three free parameters: the amplitude $A$, a spectral index $\alpha$, and an energy scale $E_0$: $$\begin{aligned} f_\text{power-exp} &= A \, E^{\alpha} \exp(-E/E_0) \,. \label{eq:power-exp}\end{aligned}$$ The benchmark parameters for this model are $A = 0.16~\text{counts}\ \text{sec}^{-1}\,\text{cm}^{-2}\,\text{keV}^{-1}$, $E_0 = 150$ keV, and $\alpha = -0.92$. This choice leads to the same normalization as for the BAND model.[^7] We choose energy-dependent bin sizes of $2 \, \delta E$, where we assume $$\begin{aligned} \label{eq:resolution} \frac{\delta E}{E} = \sqrt{ \bigg( \frac{0.05}{\sqrt{E / \text{keV}}} \bigg)^2 + (0.05)^2 } \,.\end{aligned}$$ This resolution is better than that of the Fermi GBM instrument [@Meegan:2009qu] because we will mainly be interested in the sensitivity of future observatories. In \[fig:databs\], we show a simulated data set based on these assumptions and using the BAND model. We see that the magnitude of the femtolensing effect depends crucially on the size of the source. If $a_S = 10^8$ cm, a pronounced interference pattern emerges. However already for $a_S \sim 10^9$ cm – which is still rather optimistic – it is already quite difficult to extract a clear signal, even if the lens is located at very low redshift $z_L = 0.05$, so that the projected size of the source in the lens plane is reduced compared to the case $z_L \sim 1$. Had we chosen the same $z_L$ for the orange curve as for the blue one ($a_S \sim 10^8$ cm), the amplitude of the femtolensing wiggles would be comparable to the error bars on the data or even smaller. Similarly, for larger $a_S \sim 10^{10}$ cm, the femtolensing effect disappears. Parenthetically we notice that in order to get the realistic pattern, one should further convolve it with the detector response function (see the black dotted lines). We checked explicitly, that the resolution that we have assumed in eq.  does not change qualitatively any of the curves that we show, the current Fermi resolution wipes out the effect completely for both lines. This further reinforces our point that current Fermi observation cannot even be sensitive to the emission size as small as $a_s = 10^8$ cm (which is extremely optimistic), and better resolutions of future experiments would be needed to make any further progress. ![Simulated GRB spectra based on the BAND model with parameters $A= 14.08~\text{counts}\ \text{sec}^{-1}\,\text{cm}^{-2}\,\text{keV}^{-1}$, $E_0 = 160$ keV, $\alpha_1 = -0.9$, and $\alpha_2 = -2.5$. We compare the unlensed spectrum (black) to the predicted spectra in the presence of a PBH lens with mass $M = 10^{-15}~M_\odot$ and impact parameter $y = 0.5$. Wave optics effects as well as the finite size of the GRB emission region are taken into account, where for the latter we use either $a_S = 10^8$ cm (blue) or $10^9$ cm (orange). The source is assumed to be located at redshift $z_S = 1$. For highly optimistic (small) $a_S$, a pronounced interference pattern is visible almost independently of the position of the lens, while for more realistic (larger) $a_S$, only a lens rather close to the observer may lead to an observable effect. The gray band drawn around the unlensed data represents the statistical uncertainty.The black dotted gaussians show the detector response assuming the FERMI energy resolution as well as the resolution of eq.  that we optimistically assume for future detectors. The signal that one observes in the detector would be a convolution of the detector response with the line simulated GRB spectrum.[]{data-label="fig:databs"}](fig-4-update.pdf){width="0.6\columnwidth"} Likelihood Analysis {#sec:likelihood} ------------------- To quantitatively analyze the simulated data and compare to femtolensing predictions, we define a log-likelihood function[^8] $$\begin{aligned} -2 \log L_0(\vec{\mu}_s) \equiv \min_{\vec\mu_b} \bigg[ \sum_{j=1}^{\text{\# of bins}} \bigg( \frac{O_j - P_j(\vec{\mu}_b, \vec{\mu}_s)}{\sigma_j} \bigg)^2 \bigg] + \text{const} \,, \label{eq:likelihood-1}\end{aligned}$$ where $O_j$ (“number of observed events”) denotes the number of events in the $j$-th bin without lensing, and $P_j(\vec{\mu}_b, \vec{\mu}_s)$ (“number of predicted events”) denotes the number of events expected if there is a lens. $\sigma_j$ is the uncertainty associated with the $j$-th bin, which is given by adding in quadrature the statistical uncertainty $\sqrt{P_j(\vec{\mu}_b, \vec{\mu}_s)}$ and a systematic uncertainty. The latter accounts for quasi-random deviations of the source spectrum from the model, and we assume its magnitude to be 5%. (We will also show how our results change if the systematic uncertainty is set to 0% or 10% instead.) The vector $\vec\mu_b$ contains the relevant parameters of the background model (four for the BAND and BKN models, three for the power law model with exponential cutoff). To be conservative, we minimize over $\vec\mu_b$, i.e. we choose the background parameters that best fit the data. The vector $\vec\mu_s$ contains the parameters of the lens, namely its mass $M$, its redshift $z_L$, and its normalized impact parameter in the lens plane $y = \beta / \theta_E$. Thus, $L_0$ compares the unlensed spectrum to the lensed spectrum for fixed lens and source parameters. Note that we can use Gaussian rather than Poisson statistics here because the number of photon events per bin is large. We can define a lensing cross section $\sigma(D_L) = \pi (y_\text{max} \theta_E D_L)^2$, where $y_\text{max}$ is the maximal normalized distance from the lens to observer–source line of sight that still leads to a sizable lensing signal. (A sizable signal is defined as a signal that can be distinguished, at a given confidence level (CL) from an unlensed signal.) In other words, $y_\text{max}$ is obtained by solving $$\begin{aligned} -2 \log\bigg( \frac{L_0(\vec\mu_s | y = y_\text{max})} {L_0(\vec\mu_s | y = \infty)} \bigg) = \alpha \,, \label{eq:lensing-xsec}\end{aligned}$$ where $\alpha$ is the quantile one of the $\chi^2$ distribution with one degree of freedom corresponding to the chosen CL. For instance, $\alpha = 2.7$ for 90% CL and $\alpha = 9$ for $3\sigma$ CL. The optical depth $\tau$ of the source is obtained as $$\begin{aligned} \tau = \int_0^{z_S} \! \frac{dz_L}{H(z_L)} \, \sigma(D_L) \frac{\rho_\text{PBH}}{M} (1 + z_L)^2 \,, \label{eq:optical-depth}\end{aligned}$$ where $\rho_\text{PBH} / M$ is the number density of lenses at the present epoch. We integrate over the redshift of the lens, which is equivalent, up to the Hubble constant $H(z_L)$, to integrating over comoving distance. The factor $(1+z_L)^2$ appears as a combination of a factor $(1+z_L)^3$ that accounts for the increase in lens density with redshift and a factor $(1+z_L)^{-1}$ needed to convert between the physical and comoving longitudinal coordinate. To obtain limits on the density of lenses, we need to take into account the dependence of $L_0$ on $\vec\mu_s$. To this end, we define the overall likelihood for lensing of a single GRB source with size $a_S$ and redshift $z_S$ according to $$\begin{aligned} L^\text{1-GRB}(M, \rho_\text{PBH}, a_S, z_S) &\equiv L_0(0) + \int \! d^3x \, \frac{\rho_\text{PBH}}{M} (1+z_L)^3 \big[ L_0(\vec\mu_s) - L_0(0) \big] \,. \label{eq:llpbh}\end{aligned}$$ Here, the integral runs over physical (not comoving) coordinates, and the factor $(1+z_L)^3$ once again takes into account the increase of the lens density with redshift. The integration region extends from the observer to the source in the longitudinal direction, and out to infinity in the transverse directions. We here make the (realistic) assumption that the probability of a single GRB being lensed by multiple compact DM objects is $\ll 1$. Of course, the sensitivity can be significantly boosted by observing not a single GRB, but many of them. For $n$ observed GRBs, the total likelihood is $$\begin{aligned} L^\text{$n$-GRBs}(M, \rho_\text{PBH}, \{a_{S,k} \}, \{ z_{S,k} \}) &\equiv \prod_{k=1}^{n} L^\text{1-GRB}(M, \rho_\text{PBH}, h_{S,k}, z_{S,k}) \,, \label{eq:L-nGRBs}\end{aligned}$$ where $a_{S,k}$ and $z_{S,k}$ denote the size and redshift of the $k$-th GRB. We set 95% CL limits on the mass and density of the compact DM objects by equating the log-likelihood ratio with the 95% quantile of the $\chi^2$ distribution with two degrees of freedom ($M$ and $\rho_\text{PBH}$): $$\begin{aligned} -2 \log\bigg( \frac{L^\text{$n$-GRBs}(M, \rho_\text{PBH}, \{ a_{S,k} \}, \{ z_{S,k} \})} {L^\text{$n$-GRBs}(0, 0, \{ a_{S,k} \}, \{ z_{S,k} \})} \bigg) &= 5.99 & \text{(95\% CL)} \,. \label{eq:llr}\end{aligned}$$ Here, the denominator corresponds to the likelihood of the data in the absence of any lensing. Results {#sec:sensitivity-results} ------- ![Sensitivity of femtolensing searches to the primordial black hole contribution $\Omega_\text{PBH} / \Omega_\text{DM}$ to the overall DM density in the Universe. We show the projected sensitivity for different assumptions on the number of suitable GRBs with well-measured redshifts in the data sample. We also illustrate the dependence on the size $a_S$ of the emission region of a typical GRB, where $a_S = 10^8$ cm should be considered a highly optimistic value, and $a_S = 10^9$ cm an optimistic but possible value. The colored bands indicate the impact of systematic effects (uncorrelated random fluctuations in each bin). We use the BAND model for the GRB spectrum throughout. For other spectral models, limits would change by not more than a few tens of percent. We have assumed the redshift of all GRBs in the sample to be $z_S = 1$.[]{data-label="fig:contour1"}](PBHcontour){width="0.6\columnwidth"} With the above likelihood formalism in hand, we can now study the sensitivity of current and future data to compact DM objects. We focus on the mass range $10^{-17} M_\odot$ to $10^{-14} M_\odot$. This mass window is motivated by the requirement that a sizable phase difference between the different lensed images of the source should occur within the energy range of GRB spectra. Our results are shown in \[fig:contour1\]. We see that with 20 GRBs with well-measured redshifts (the number of GRBs used in [@Barnacka:2012bm]), a meaningful limit can only be set if the GRBs are assumed to be point-like, i.e. $a_S \lesssim 10^8$ cm. Here, by “meaningful limit” we mean a limit that constrains the cosmological PBH abundance, $\Omega_\text{PBH}$, to be less than the total DM abundance in the Universe, $\Omega_\text{DM}$. As argued in \[sec:extended-source\], the assumption $a_S \lesssim 10^8$ cm is almost certainly overly optimistic. If $a_S$ is only one order of magnitude larger (which is still very optimistic), no limit can be set. This may change in the future if the available sample of GRBs is significantly extended. With 100 GRBs, sensitivity to an $\mathcal{O}(30\%)$ fraction of PBHs can be achieved for $a_S \sim 10^9$ cm. The sensitivity would improve to $\Omega_\text{PBH} / \Omega_\text{DM} \sim 0.06$ if $a_S \sim 10^8$ cm. Our conclusions are essentially independent of the choice of GRB model (BAND vs. BKN vs. power law with exponential cutoff). They depend somewhat on the assumed systematic uncertainty, with the sensitivity deteriorating by at most a factor of two if the assumed systematic error is increased from 0% to 10%. Note that we have very conservatively assumed systematic errors to be completely uncorrelated between energy bins. ![image](PHBconstraint2){width="1.3\columnwidth"} In \[fig:pbh-all-constraints\], we put our projected constraints into a broader context by comparing to other limits on PBHs. We see that future femtolensing constraints, albeit weak, may cover a mass range that is otherwise inaccessible and where viable PBH DM could exist. While the projected limits shown in \[fig:contour1,fig:pbh-all-constraints\] apply only to PBH DM, we can use the arguments given at the end of \[sec:extended-lens\] as a starting point for estimating also the sensitivity to other compact DM structures. For UCMHs, we have argued above that the interference fringes are shifted to higher energies compared to the PBH case, with the magnitude of the shift, $m(\theta_0)/M_\text{cusp}$, given by \[eq:m-over-M-UCMH,eq:m-over-M-UCMH-numerical\] for UCMHs with $\rho(r) \propto (r_0/r)^{9/4}$. We therefore estimate that the projected femtolensing exclusion limits on such UCMHs will be similar in shape to the ones for PBHs if the PBH mass is understood as the “equivalent black hole mass” $m(\theta_0)$ from \[eq:m-over-M-UCMH,eq:m-over-M-UCMH-numerical\]. Expressed in terms of $M_\text{cusp}$, the limits on UCMHs are thus shifted to higher masses by a factor $M_\text{cusp} / m(\theta_0)$ compared to the limits shown in \[fig:contour1,fig:pbh-all-constraints\]. We moreover need to take into account the fact that the DM number density scales inversely with mass, hence we expect the projected exclusion curves to also move upwards by a factor $M_\text{cusp} / m(\theta_0)$, corresponding to a weakening of the limit by that factor. In view of this, and given that we have seen how difficult it will already be to constrain PBH DM using femtolensing, we refrain from a more detailed sensitivity study for UCMHs. Conclusions {#sec:conclusions} =========== To summarize, we have critically investigated the potential of gravitational femtolensing to constrain compact DM structures such as primordial black holes or ultracompact DM minihalos. Femtolensing exploits the tiny time delay between the multiple lensed images of a distant source. Interference between the images leads to characteristic fringes in the energy spectrum of observed photons, see \[fig:databs\]. As sources, we consider in particular gamma ray bursts at $\mathcal{O}(1)$ redshift. These sources are most easily observed at energies $\gtrsim 10$ keV, hence observable interference pattern require time delays $\lesssim 10^{-19}$ sec. This means that the best sensitivity is expected for compact DM objects in the mass range from $10^{-17}$ to $10^{-14}~M_\odot$. We have argued that the simple geometric picture of femtolensing based on point-like sources and lenses that is often used in the literature is not appropriate in reality. First, it is not true that photons travel from the source to the detector along one of just two discrete paths. In fact, when the time delay becomes comparable to the inverse photon frequency (which for point-like lenses is equivalent to the photon wave length becoming comparable to the Schwarzschild radius of the lens), wave optics effects become non-negligible. It is then necessary to integrate the photon amplitude over the whole lens plane. This leads to $\mathcal{O}(1)$ corrections to the interference pattern at the lower end of the photon energy spectrum. Second, while the approximation of a point-like lens works for primordial black holes, it is not satisfied for ultra-compact mini-halos, and even less so for NFW-like structures. We have therefore computed femtolensing effects for generic power-law density profiles, and have explicitly shown numerical results for the self-similar infall profile with $\rho(r) \propto r^{-9/4}$. The most important correction in femtolensing of GRBs is coming from the non-negligible size $a_S$ of the GRB source itself. In fact, we have argued that a GRB could only be treated as point-like for the purpose of femtolensing if the photon emission region was smaller than $a_S \sim 10^8$ cm. And while estimates for the size of the emission region can vary by a few orders of magnitude, $a_S \sim 10^8$ cm seems unrealistically small. For more realistic assumptions on the value of $a_S \gtrsim 10^{10}$ cm, the femtolensing effect is almost entirely washed out. This means that, contrary to previous claims, current GRB data is insufficient to constrain compact DM structures such as primordial black holes, even if their abundance $\Omega_\text{PBH}$ saturates the observed DM abundance $\Omega_\text{DM}$ in the Universe. We have, however, demonstrated that constraints down to $\Omega_\text{PBH} / \Omega_\text{DM} \sim 0.2$ would become possible with a sample of about 100 observed GRBs with well-measured redshifts and spectra, and with small $a_S$. Such GRBs are expected to be characterized by very fast intrinsic variability at sub-millisecond time scales. Since femtolensing constraints on compact DM objects are still out of reach we conclude that there are currently *no* firm bounds on PBH DM in the mass range of $10^{-17} \div 10^{-11}$ (with the exception of a small wedge of parameter space excluded by white dwarf observation). This is illustrated in \[fig:pbh-all-constraints\]. While several ideas has been put forward to constrain this mass range, for instance picolensing [@Nemiroff:1995ak] and the capture of PBHs by stars or neutron stars, none of these methods has yielded decisive bounds until now, so this region of parameter space still awaits exploration. We are grateful to Yacine Ali-Haïmoud, Juan Garcia Bellido, Kfir Blum, Lam Hui and Kathryn Zurek for useful discussions. JK has been supported by the German Research Foundation (DFG) under Grant Grant Nos. KO 4820/11, FOR 2239, EXC-1098 (PRISMA) and by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 637506, “$\nu$Directions”). JK would also like to thank CERN for hospitality and support. The size of the prompt emission region in GRBs {#app:GRBsize} ============================================== In this appendix we discuss the transverse size of the $\gamma$-ray emitting regions in GRBs, which, as we have seen in \[sec:extended-source\], plays a crucial role in the study of femtolensing. We will review estimates for the size of the emission region based on measurements of the variability time scale, and we will also discuss the lower bound following from the required transparency of the emission region to $\gamma$-rays. It is believed that the prompt $\gamma$-emission of GRBs is produced by electrons and positrons accelerated in relativistic shock waves. The non-thermal $\gamma$-ray spectrum implies that the emission region must be optically thin. To reconcile this requirement with the observed energetics of the bursts, the bulk Lorentz factor of the $\gamma$-ray emitting material must be large, $\varGamma\gtrsim 100$ [@Piran:2004ba]. We etsimate the transverse size of the emission region following the same approach as Ref. [@Barnacka:2014yja]. Consider a blob of material of size $a_S$ moving with velocity $v$ at an angle $\theta_\text{obs}$ to the observer’s line of sight. In the rest frame of the blob the minimal variability time scale of the emission is simply given by the light-crossing time, $$\begin{aligned} \hat t_\text{var}\sim a_S /c \,. \label{eq:tvarhat}\end{aligned}$$ The observed variability time $t_\text{var}$ is related to $\hat t_\text{var}$ by the relativistic Doppler formula, $$\begin{aligned} t_\text{var} = (1+z_S) \bigg(1 - \frac{v}{c} \cos\theta_\text{obs}\bigg) \varGamma \, \hat t_\text{var}\,, \label{eq:tvar}\end{aligned}$$ where $\varGamma \equiv \big(1 - (v/c)^2\big)^{-1/2} $ and we have taken into account the cosmological redshift of the source. Due to the relativistic beaming effect, we have $\theta_\text{obs}\sim 1/\varGamma$. Combining this with \[eq:tvarhat,eq:tvar\] one obtains the estimate $$\begin{aligned} a_S \sim \frac{c \varGamma \, t_\text{var}}{1+z_S} \simeq \frac{10^{11} \, \text{cm}}{1+z_S} \times \bigg( \frac{t_\text{var}}{0.03\,\text{sec}}\bigg) \bigg( \frac{\varGamma}{100}\bigg) \,. \label{eq:ahatestim}\end{aligned}$$ The minimal variability time scales for various GRBs have been determined in ref. [@Golkhou:2015lsa]. They lie within the range $t_\text{var}^\text{sGRB} \sim (0.01\div 0.1)\;\text{sec}$ for short GRBs and $t_\text{var}^\text{lGRB} \sim (0.1\div 1)\;\text{sec}$ for the long ones. These results are consistent with the earlier estimates of ref. [@Barnacka:2014yja] that give average variability time scales $t_\text{var}$ of $0.036$ sec and $1.2$ sec for short and long GRBs, respectively. We see that for a typical short GRB with $z\sim 1$, $t_\text{var}\sim 0.03$ sec, and $\varGamma \sim \mathcal{O}(100)$, the transverse size is, $$\begin{aligned} a_S \sim 10^{11} \; \text{cm} \,. \label{eq:ahatnum}\end{aligned}$$ For long GRBs this estimate becomes an order of magnitude larger.[^9] In the above estimates we adopted the standard picture of a GRB with a relativistic jet of $\gamma$-ray emitting material pointing towards the observer. One may wonder how the reasoning is modified in the case of off-axis observation, like in the recent GRB 170817A [@Goldstein:2017mmi; @Savchenko:2017ffs] accompanied by the gravitational wave event GW170817 [@TheLIGOScientific:2017qsa], which has confirmed the identification of neutron star mergers as progenitors of short GRBs. As discussed in [@Kasliwal:2017ngb], the observational data pertaining to this event across the electromagnetic spectrum are best described by a model in which a mildly relativistic wide-angle shock with $\varGamma\sim 2.5$ breaks off the ambient material and emits $\gamma$-rays at a distance $R_\text{em}\simeq 2.4\times 10^{11}$cm from the central engine. In this case relativistic beaming is practically absent, so $a_S$ is comparable to $R_\text{em}$. We again recover the estimate (\[eq:ahatnum\]). If all GRBs satisfy the estimate , observing any femtolensing in their spectra will be essentially hopeless. So let us ask how robust this estimate is. A hint that there may be GRBs with smaller sizes is already provided by the analysis of Ref. [@Golkhou:2015lsa], which finds that about 10% of the bursts (both short and long) exhibit faster variability, $t_\text{var} \lesssim 2\times 10^{-3}~\text{sec}$, which leads to an estimated size $a_S \lesssim 10^{10}$ cm. Thus, one can speculate that some rare GRBs at the tail of the distribution could have $t_\text{var}$ yet another order of magnitude shorter, bringing their sizes close to the values of $a_S$ required for efficient femtolensing. It is also worth pointing out that the determination of the minimal variability time scale is affected by instrumental systematics, such as the detector sensitivity and the light-curve sampling. Thus, it might happen, in principle, that some of the measured variability time scales overestimate the true intrinsic variability time scale of the source. An alternative method for constraining $a_S$ is based on the requirement that the emission region must be optically thin. Two processes can lead to absorption of $\gamma$-rays: production of $e^+e^-$ pairs in $\gamma\gamma$-scattering, and Compton scattering on electrons and positrons. Let us start with the first process. Consider a photon with energy $\hat E$ in the rest frame of the emitting blob. The optical depth for pair production is, $$\begin{aligned} \tau_{\gamma\gamma}(\hat E) = a_S \int_{\hat\epsilon_\text{th}} \! d\hat\epsilon \; \sigma_{\gamma\gamma}(\hat \epsilon) \, \hat n_\gamma(\hat\epsilon) \,, \label{eq:taugg}\end{aligned}$$ where $\sigma_{\gamma\gamma}(\hat\epsilon)$ is the (angular-averaged) cross section of collision with an ambient photon of energy $\hat\epsilon$, $\hat n_\gamma(\hat\epsilon)$ is the spectral density of such photons, and $$\begin{aligned} \hat\epsilon_\text{th} = \frac{m_e^2c^4}{\hat E}\end{aligned}$$ is the threshold energy for pair production. To estimate $\hat n_\gamma(\hat E)$ we notice that the number of photons with energy $\hat E$ emitted by the blob during its proper time $d\hat t$ is $$\begin{aligned} dN \simeq c \, a_S^2 \, \hat n_\gamma(\hat E) \, d\hat E \, d\hat t \,.\end{aligned}$$ This is related to the photon flux $f(E)$ seen by the observer at energy $E = \varGamma \hat E / (1+z_S)$ as $$\begin{aligned} dN \simeq \frac{d_S^2}{\varGamma^2(1+z_S)^2} \, f(E) \, dE \, dt \,,\end{aligned}$$ where $d_S$ is the luminosity distance of the GRB and the factor $1/\varGamma^2$ is due to the fact that in the observer’s frame the emission is beamed into a narrow cone with opening angle $\sim \varGamma^{-1}$. Next, due to the Doppler effect we have (cf. \[eq:tvar\]) $dt \simeq (1+z_S) \, d\hat t \, \varGamma$. Inserting all these relations into (\[eq:taugg\]) we obtain, $$\begin{aligned} \tau_{\gamma\gamma}(\hat E)\simeq\frac{d_S^2}{c \, a_S\varGamma^2(1+z_S)^2} \int_{\hat\epsilon_\text{th}}d\hat\epsilon\;\sigma_{\gamma\gamma}(\hat\epsilon)\, f\big(\varGamma\hat\epsilon/(1+z_S)\big)\;. \label{eq:taugg1}\end{aligned}$$ Following common practice, we now assume that the spectrum of the prompt emission is described by a power law, $$\begin{aligned} f(E) = A \, E^{\alpha} \,.\end{aligned}$$ with spectral index $\alpha$ close to $-2$. Then the integral in \[eq:taugg1\] can be evaluated with the result, $$\begin{aligned} \tau_{\gamma\gamma}(\hat E)\simeq\frac{d_S^2\sigma_T}{c a_S}\eta(\alpha)A\varGamma^{\alpha-2} (1+z_S)^{-\alpha-2}\bigg(\frac{m_e^2c^4}{\hat E}\bigg)^{\alpha+1}\;, \label{eq:taugg2}\end{aligned}$$ where $\sigma_T$ is the Thomson cross section and the numerical coefficient is [@Gould:1967zzb; @1987MNRAS.227..403S] $$\begin{aligned} \eta(\alpha) = \frac{3\big(\alpha(\alpha-5)(\alpha-1)-2\big)\sqrt{\pi} \, \Gamma(-\alpha)} {8\alpha(1-\alpha)(\alpha-2)(\alpha-1)\Gamma(3/2-\alpha)}\,.\end{aligned}$$ Expressing the optical depth in terms of the photon energy measured by the observer we arrive at, $$\begin{aligned} \tau_{\gamma\gamma}(E) \simeq \frac{d_S^2\sigma_T}{c a_S} \eta(\alpha) \varGamma^{2\alpha-1} (1+z_S)^{-2\alpha-3} \, \frac{m_e^2c^4}{E} f\bigg(\frac{m_e^2c^4}{E}\bigg) \,, \label{eq:taugg3}\end{aligned}$$ Up to a numerical coefficient of order one, this coincides with the expression obtained in [@2001ApJ...555..540L]. For $\alpha<-1$ the optical depth grows with the energy of the photon. Requiring that it is smaller than one for photons with the maximal observed energy $E_\text{max}$ we obtain, $$\begin{aligned} a_S > 2.5\times 10^6 \,\text{cm} \times \Big(\frac{d_S}{7\text{Gpc}}\Big)^2 \Big(\frac{f_{500}}{10^{-3}\text{sec}^{-1}\text{cm}^{-2}\text{keV}^{-1}}\Big) \Big(\frac{E_\text{max} }{1\text{MeV}}\Big) \Big(\frac{\varGamma}{1000}\Big)^{-5} \,, \label{agg}\end{aligned}$$ where $f_{500}$ is the $\gamma$-ray flux at $500$keV and we have assumed $\alpha=-2$, $z_S=1$ for the numerical estimate. Due to instrumental limitations the spectra of most GRBs are measured up to $E_\text{max}\sim 1$ MeV. This corresponds to the energy $\hat E\sim (1+z_S)\varGamma^{-1}\,\text{MeV}\ll 1\,\text{MeV}$ in the blob rest frame. If the spectrum were cut at these energies, all the photons would be below the threshold of pair production and we would not have to worry about absorption at all. However, it is believed that the spectrum of a typical GRB extends to much higher energies. This is supported by detection of high energy emission (up to $\sim 100\,\text{GeV}$) from several GRBs [@Nava:2018qkq]. When such observations are available, the corresponding maximal energy can be used in  to set a strong lower bound on the source size. We now turn to absorption due to Compton scattering of $\gamma$-photons on electrons and positrons. We will consider $e^{\pm}$ created by the process of $\gamma\gamma$ scattering discussed above. As will be seen shortly, the resulting constraint is stronger than (\[agg\]) in the case when a direct measurement of the high-energy component of $\gamma$-radiation is absent [@2001ApJ...555..540L]. Assuming, for simplicity, that electrons are non-relativistic in the rest frame of the emitting material, the optical depth for photons with $\hat E\lesssim m_ec^2$ is essentially independent of their energy, $$\begin{aligned} \tau_{\gamma e}\simeq a_S\sigma_T\hat n_e\,,\end{aligned}$$ where $\hat n_e$ is the total density of electrons. The latter is estimated as the density of photons for which the blob is optically thick with respect to pair production, $$\begin{aligned} \hat n_e = \int_{\hat\epsilon_\text{cr}}d\hat\epsilon\;\hat n_\gamma(\hat\epsilon)\,,\end{aligned}$$ where $\hat \epsilon_\text{cr}$ is determined from the equation $\tau_{\gamma\gamma}(\hat\epsilon_\text{cr})=1$. Using the expression we obtain, $$\begin{aligned} \tau_{\gamma e}\simeq\frac{\eta(\alpha)}{-\alpha-1}\bigg[ \frac{d_S^2\sigma_T}{c a_S}A\varGamma^{\alpha-2}(1+z_S)^{-\alpha-2}(m_ec^2)^{\alpha+1}\bigg]^2\;. \label{tauge1}\end{aligned}$$ Requiring that the emission region be optically thin, $\tau_{\gamma e}<1$, translates into $$\begin{aligned} a_S > \bigg(\frac{\eta(\alpha)}{-\alpha-1}\bigg)^{1/2}\frac{d_S^2\sigma_T}{c} \varGamma^{\alpha-2}(1+z_S)^{-\alpha-2}\,m_ec^2f(m_ec^2)\;. \label{age}\end{aligned}$$ For the numerical values $\alpha=-2$, $z_S=1$ this yields, $$\begin{aligned} a_S > 1.8\times 10^{9}\, \Big(\frac{d_S}{7\text{Gpc}}\Big)^2 \Big(\frac{f_{500}}{10^{-3}\text{sec}^{-1}\text{cm}^{-2}\,\text{keV}^{-1}}\Big) \Big(\frac{\varGamma}{1000}\Big)^{-4}\text{cm} \,. \label{age1}\end{aligned}$$ We see that for extreme values of the boost factor $\varGamma\sim 1200$ which may occur in some GRBs [@2011ApJ...738..138R] the lower bound on the source size is quite small. Of course, it would be too optimistic to interpret this lower bound as a plausible value of $a_S$. On the other hand, the above derivation is only an order-of-magnitude estimate. Moreover, it relies on a power-law extrapolation of the GRB spectrum. As such, it shows that the possibility for some GRBs to have sizes $a_S \lesssim 10^9$cm is not completely excluded. Further progress in our quantitative understanding of the physics of GRBs is required to conclude if this option is viable or not. Note that according to eq. (\[eq:ahatestim\]), if such GRBs exist, they are expected to have variablity at sub-millisecond time scale. [^1]: Note that the lensing potential is defined by \[eq:lensing-potential\] only up to an additive constant. We will ignore this constant here, which implies that the expression for the time delay, \[eq:timedelay\], is applicable only for calculating the time difference between different paths, but does nor necessary reflect the *absolute* time delay along the path. [^2]: The generic lens equation is $\vec \theta - \vec \beta = \vec \nabla \psi(\vec \theta) $. [^3]: We are grateful to Juan Garcia-Bellido for pointing out to us the importance of wave optics effects. [^4]: We leave aside the question whether such variability time scales can be measured with existing instruments, as the main focus of our study are future observations. [^5]: Note that some steep profiles tend to develop a core at the center of the distribution. In particular, the self-similar radial infall profile corresponding to $\delta=9/4$ cannot be a valid approximation all the way down to $r = 0$ (see e.g. [@Bringmann:2011ut] for an approach to estimate the size of the core). While we will for simplicity neglect the presence of the core in the sensitivity studies presented in this paper, it should be taken into account when analyzing real data. [^6]: The case isothermal sphere profile with $\delta=2$ lies at the boundary between the two regimes discussed here: there are two lensed images for $\beta<\theta_0$ and a single image for $\beta>\theta_0$. [^7]: The power law model is not strongly motivated physically, and is just used here as a cross check to verify the robustness of our predictions. In particular, if there is an exponential cutoff in the GRB spectrum, it is expected at an energy much higher than $150$ keV [@Nava:2018qkq]. [^8]: Note that our statistical procedure differs from the one in [@Barnacka:2012bm]. Our approach combines the likelihood of the lensed signal for every possible 3-dimensional position of the lens, weighted by the probability of finding a lens at a given position. Ref. [@Barnacka:2012bm], in contrast, investigates the observability of lensing as a function of the lens mass and the transverse position of the source, but without fully taking into account the dependence on the redshift of the lens $z_L$. [^9]: Note that Ref. [@Golkhou:2015lsa] derives also the distribution of the emission radii $R_\text{em}$, i.e. the distance from the GRB central engine at which $\gamma$-rays are emitted. The central value of this distribution is $R_\text{em}\sim 3\times 10^{13}$ cm ($10^{14}$ cm) for short (long) GRBs. Due to relativistic beaming, the transverse size of the patch visible by an observer on Earth is related to this distance as $a_S \sim R_\text{em}/\varGamma$. Assuming $\varGamma\sim {\cal O}(100)$, this again gives $ a_S\sim 10^{11}$cm ($10^{12}$cm) for short (long) GRBs.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We generalize several comparison results between algebraic, semi-topological and topological $K$-theories to the equivariant case with respect to a finite group.' address: 'Bergische Universität Wuppertal, Gau[ß]{}str. 20, D-42119 Wuppertal, Germany' author: - Jeremiah Heller - Jens Hornbostel title: 'Equivariant semi-topological $K$-homology and a theorem of Thomason' --- Introduction ============ In his landmark article [@Thomason:famous], Thomason establishes an étale Atiyah-Hirzebruch type spectral sequence relating étale cohomology and Bott-inverted algebraic $K$-theory with finite coefficients. When restricted to smooth complex varieties his results say, amongst other things, that there is an isomorphism $$K_{*}^{alg}(X;{\mathbb{Z}}/n)[\beta^{-1}]{\cong}KU^{-*}(X^{an};{\mathbb{Z}}/n)$$ between Bott-inverted algebraic $K$-theory with finite coefficients and topological $K$-theory with finite coefficients. In the last decade, Friedlander and Walker (see e.g., [@FW:compK], [@FW:ratisos]) have refined the comparison map ${\mathcal{K}}^{alg}(X)\to {\mathcal{K}}^{top}(X^{an})$ between the algebraic and the topological $K$-theory of complex varieties by introducing an intermediate theory ${\mathcal{K}}^{sst}(X)$, called [*semi-topological $K$-theory*]{}. One has natural morphisms of spectra $${\mathcal{K}}^{alg}(X) \to {\mathcal{K}}^{sst}(X) \to {\mathcal{K}}^{top}(X^{an})$$ where the left hand one induces a weak equivalence $${\mathcal{K}}^{alg}(X;{\mathbb{Z}}/n) \xrightarrow{{\simeq}} {\mathcal{K}}^{sst}(X;{\mathbb{Z}}/n)$$ for smooth quasi-projective $X$. Upon inverting (the unique lift of) the topological Bott element $\beta$, the right hand map induces a weak equivalence $${\mathcal{K}}^{sst}(X)[\beta^{-1}] \xrightarrow{{\simeq}} {\mathcal{K}}^{top}(X^{an})$$ for smooth quasi-projective $X$. There have been several proofs of the result that Bott-inverted semi-topological $K$-theory agrees with topological $K$-theory; the first ones relying on Thomason’s result itself. In [@Walker:Thomason] Walker introduces a [*bivariant*]{} semi-topological $K$-theory for quasi-projective complex varieties. One of the main results of that article is that the semi-topological $K$-homology of a smooth quasi-projective complex variety is isomorphic to the topological $K$-homology of its underlying complex manifold $X^{an}$. Using this result, Walker gives a new proof, in the case of smooth projective complex varieties, that Bott-inverted semi-topological $K$-theory agrees with complex $K$-theory. His proof does not rely on Thomason’s theorem and thus specializes to give a particularly elegant alternate proof of Thomason’s celebrated theorem comparing algebraic and complex $K$-theory with finite coefficients, in case of smooth projective complex varieties. In the present article, we generalize Walker’s results mentioned above to the equivariant setting with respect to an arbitrary finite group $G$. We begin by introducing a bivariant equivariant semi-topological $K$-theory ${\mathcal{K}}_{G}^{sst}(X,Y)$ for quasi-projective $G$-varieties $X$ and $Y$. To construct and study this bivariant theory, which is constructed as a $G$-spectrum, we rely on the machinery of equivariant $\Gamma$-spaces, established by Shimakawa. An important case is when $Y=\operatorname{\mathrm{Spec}}({\mathbb{C}})$, in which case ${\mathcal{K}}_{G}^{sst}(X,{\mathbb{C}})$ defines equivariant semi-topological $K$-theory. Similarly the equivariant semi-topological $K$-homology of $Y$ is ${\mathcal{K}}_{G}^{sst}({\mathbb{C}},Y)$. Our first main result is the following generalization of Walker’s comparison theorem, appearing as Theorem \[mainthm\] below. Let $Y$ be a smooth quasi-projective $G$-variety. Then there is a natural weak equivalence of $G$-spectra $${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},Y^{an})\xrightarrow{{\simeq}} {\mathcal{K}}^{sst}_{G}({\mathbb{C}},Y).$$ Here ${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},Y^{an})$ is the equivariant topological $K$-homology introduced in Section \[sec:top\] and is shown to be equivariantly weakly equivalent to $Y^{an}\wedge {\mathbf{bu}}_{G}$, where ${\mathbf{bu}}_{G}$ is the connective cover of the $G$-spectrum $KU_{G}$ representing equivariant complex $K$-theory. The topological Bott element lifts (uniquely) to a “semi-topological Bott element”, $\beta_{2}\in K_{2}^{G,\,sst}({\mathbb{C}},{\mathbb{C}})$ and our second main result is Theorem \[mainthmbottinv\], establishing that Bott-inverted equivariant semi-topological $K$-theory and equivariant topological $K$-theory agree. Let $X$ be a smooth complex projective $G$-variety. Then there are natural isomorphisms $$K^{G,sst}_{*}(X,{\mathbb{C}})[\beta^{-1}_{2}] \xrightarrow{{\cong}} KU_{G}^{-*}(X^{an}).$$ This result is proved following the outline of the argument in [@Walker:Thomason]. Namely, the proof relies on the equivariant version of Walker’s comparison theorem mentioned above, the pairings constructed in Section \[sec:pair\], and a good equivariant theory of fundamental and Thom classes. Regarding this last item, Walker uses the fact that nonequivariantly connective $K$-theory has Thom classes and satisfies Poincare duality. Here some nontrivial changes need to be made in the equivariant setting since the version of connective equivariant $K$-theory that appears in our work is not [*complex stable*]{}, see the discussion before Lemma \[kuisalmostKU\]. As a consequence of the rigidity property for equivariant algebraic $K$-theory established by Yagunov-[Ø]{}stv[æ]{}r [@YO:equirigid] and Friedlander-Walker’s recognition principle [@FW:ratisos], we establish in Theorem \[KalgKsemi\] an isomorphism $$K^{G,\,alg}_{*}(X;{\mathbb{Z}}/n) \xrightarrow{{\cong}} K_{*}^{G,\,sst}(X;{\mathbb{Z}}/n)$$ for smooth $X$. Here, in order to have a comparison map between our equivariant algebraic and semi-topological $K$-theories, it is important to have available an equivariant version of the Grayson-Walker theorem concerning geometric models for $K$-theory spectra. This result is proved by [Ø]{}stv[æ]{}r in [@Ostvaer]. As a consequence of the above isomorphism, Theorem \[mainthmbottinv\] specializes to give an alternate proof (in the case of smooth projective $G$-varieties) of the equivariant version of Thomason’s theorem [@Thomason:famousequi Theorem 5.9], comparing Bott-inverted equivariant algebraic $K$-theory and equivariant complex $K$-theory (with finite coefficients). Due to considerations of length, we have not discussed here several other generalizations and related results which are likely to be true. First, (some version of) Theorem \[mainthm\] should be true for Real semi-topological $K$-homology and real varieties. Second, the results of section \[sec:sstThom\] probably hold for quasi-projective varieties as well, by replacing the homology theories appearing there with a Borel-Moore type homology theory. Third, Theorem \[KalgKsemi\] probably holds for bivariant algebraic $K$-theory as well, as the base change, normalization and additivity property necessary to establish rigidity seem to extend to the corresponding categories of $G$-modules. Fourth, the ring structure on the equivariant algebraic $K$-theory introduced in section \[sec:pair\] presumably coincides with the previously considered ring structure and similarly for the topological theory. (Note that Proposition \[prop:conn\] implies the product on the cohomology theory is the same in positive degrees, see also [@Walker:Thomason Remark 6.11].) Finally, in light of the equivariant generalizations of [@Walker:Thomason] presented here, it would be interesting to know whether the more general results of [@Walker:Thomalg] admit an equivariant generalization as well. We conclude with an overview of the article. In section \[sec:pre\], we review some material on equivariant stable homotopy theory, in particular about equivariant $\Gamma$-spaces and equivariant group completion. In Section 3 we introduce and study various models for equivariant bivariant algebraic, semi-topological and topological $K$-theory we need to consider. In section \[sec:comp\] we establish the equivariant version of Walker’s comparison theorem between equivariant semi-topological and equivariant topological $K$-homology. Section \[sec:pair\] is devoted to a detailed study of pairings and operations (e.g. slant products) for the various equivariant $K$-theories appearing in this article. In section \[sec:sstThom\], we establish that Bott-inverted equivariant semi-topological $K$-theory and equivariant topological $K$-theory agree, for smooth projective complex $G$-varieties. In section \[sec:algThom\] we show how the semi-topological result implies Thomason’s result, thus giving a new proof in the the equivariant setting for smooth projective complex varieties. In a companion article [@Ostvaer], [Ø]{}stv[æ]{}r shows that the Grayson-Walker model of algebraic $K$-theory [@GW:Kmodels] allows for an equivariant generalization (which we use to write down the comparison map between the equivariant algebraic and semi-topological $K$-theories that is used in Section \[sec:algThom\]). We thank him for helpful discussions regarding this result. We are grateful to the referee for pointing out an error in our previous proof of Theorem \[mainthm\]. Additionally, the first author would like to thank M. Voineagu for several useful conversations on closely related topics. **Notation:** Unless stated otherwise, $G$ will be a finite group. We write ${\mathrm{Sch}}/{\mathbb{C}}$ for the category of quasi-projective complex varieties and ${\mathrm{Sm}}/{\mathbb{C}}$ for the full subcategory of smooth quasi-projective complex varieties. For a complex variety $X$, the set of complex points equipped with the Euclidean topology is denoted by $X^{an}$. Preliminaries {#sec:pre} ============= Stable equivariant homotopy theory, [$\Gamma_{G}$]{}-spaces, and [${\mathcal{W}}_{G}$]{}-spaces ----------------------------------------------------------------------------------------------- We write ${G\mathcal{T}}$ for the category whose objects are compactly generated Hausdorff spaces with $G$-action together with a $G$-invariant base-point and morphisms are based $G$-equivariant maps. Write ${\mathcal{T}}_{G}$ for the category with the same objects as ${G\mathcal{T}}$, but morphisms are all based continuous maps, hence $g(f(x)):=gf(g^{-1}x)$ defines a $G$-action on the morphism sets. Both these categories are enriched over topological spaces and ${\mathcal{T}}_{G}$ is enriched over ${G\mathcal{T}}$. In this paper, a $G$-spectrum means an orthogonal $G$-spectrum, unless otherwise specified; we usually omit the adjective orthogonal. If $A$ is an orthogonal $G$-spectrum we also write $A$ for its underlying (pre)-spectrum. We refer to [@May:equihomotopy] for background on and a good survey of equivariant stable homotopy theory, [@LLM] for further details concerning “classical” spectra in the equivariant setting, and [@MM:O] for equivariant orthogonal spectra. As is customary we write $[X,Y]_{G}$ for maps in the $G$-equivariant stable homotopy category. For a representation $V$ we write $$\pi_{V}^{G}X = [S^{V},X]_{G}.$$ In this paper our spectra arise primarily via equivariant $\Gamma$-spaces and ${\mathcal{W}}_{G}$-spaces, and we now recall some details on these. Let ${\mathcal{W}}_{G}$ denote the category of based $G$-spaces that are homeomorphic to finite $G$-$CW$-complexes and maps are all (base-point preserving) maps. A *${\mathcal{W}}_{G}$*-space is a based, equivariant functor $X:{\mathcal{W}}_{G}\to {\mathcal{T}}_{G}$ such that $$X:\operatorname{Map}(A,B) \to \operatorname{Map}(X(A),X(B))$$ is an equivariant continuous map of $G$-spaces. We have a map of $G$-spaces $X(A)\wedge B \to X(A\wedge B)$ obtained as the adjoint of the composition $B\to \operatorname{Map}(A,A\wedge B) \to \operatorname{Map}(X(A), X(A\wedge B))$. In particular, a ${\mathcal{W}}_{G}$-space $X$ functorially determines an orthogonal $G$-spectrum $\mathbb{U}X$ via $(\mathbb{U}X)(V) = X(S^{V})$ and hence it also determines a $G$-prespectrum. Moreover $\mathbb{U}$ is the right adjoint in a Quillen equivalence between ${\mathcal{W}}_{G}$-spaces and orthogonal $G$-spectra (indexed on a complete universe) and the category of ${\mathcal{W}}_{G}$-spaces has a smash-product such that $\mathbb{U}$ is lax symmetric monoidal, see [@AB] for details. There are two equivalent formulations of equivariant $\Gamma$-spaces. The first is as follows. Let $\Gamma$ denote the category whose objects are pointed sets $\underline{n}_{+}=\{0,1,\ldots,n\}$, pointed at $0$. Maps are base-point preserving set maps. An equivariant $\Gamma$-space is a functor $X:\Gamma \to {G\mathcal{T}}$ such that $X(0)= *$. Write $\Gamma[{G\mathcal{T}}]$ for the category whose objects are the equivariant $\Gamma$-spaces and morphisms are natural transformations. The second model is as follows. Let $\Gamma_{G}$ denote a skeletal category of finite $G$-sets with morphisms all pointed set maps. The category $\Gamma_{G}[{\mathcal{T}}_{G}]$ has as objects equivariant functors $X:\Gamma_{G}\to {\mathcal{T}}_{G}$ such that $X(0) = *$ and maps are equivariant natural transformations. A useful observation due to Shimakawa and May [@Shimakawa:note] is that there is an adjoint pair of functors $$\label{eqn:gammaeq} i:\Gamma_{G}[{\mathcal{T}}_{G}]\rightleftarrows \Gamma[{G\mathcal{T}}]:P$$ which are an equivalence of categories. Here $i$ is induced by the inclusion functor $i:\Gamma\to \Gamma_{G}$. In this paper we will generally work with the objects of $\Gamma[{G\mathcal{T}}]$ which we refer to simply as equivariant $\Gamma$-spaces. We refer to objects of $\Gamma_{G}[{\mathcal{T}}_{G}]$ as $\Gamma_{G}$-spaces. The equivalence $P$ is defined as follows. Let $X:\Gamma\to {G\mathcal{T}}$ be an equivariant $\Gamma$-space and $S$ a finite $G$-set. Write $S$ also for the contravariant functor $\operatorname{Map}(-,S)$ which it represents. The value of $PX:\Gamma_{G}\to {\mathcal{T}}_{G}$ on a $G$-set is defined via the left Kan extension $PX(S) = S\otimes_{\Gamma}X$. Alternatively, $PX(S)$ can be described as follows. A $G$-set $S$ corresponds to a group homomorphism $\rho:G\to \Sigma_{n}$ where $|S|=n$. Given a homomorphism $\rho:G\to \Sigma_{n}$ one defines a new $G$-action on $X(n)$ via the formula $g\cdot_{\rho}x = X(\rho(g))(gx)$ for $x\in X(n)$. Write $X(n)_{\rho}$ for this $G$-space. Then $PX(S) = X(n)_{\rho}$. 1. A $\Gamma_{G}$-space $X:\Gamma_{G}\to {\mathcal{T}}_{G}$ is said to be *special* provided $X(S) \to \operatorname{Map}_{cts*}(S,X(\underline{1}))$ is a $G$-weak equivalence for any $S$. 2. Say that an equivariant $\Gamma$-space is *special* if for every subgroup $H\subseteq G$ and homomorphism $\rho:H\to \Sigma_{n}$ the map $$X(n)_{\rho} \to (X(1)^{n})_{\rho}$$ is an $H$-weak equivalence, where $(X(1)^{n})_{\rho}$ is the $G$-space with action given by $g(x_{1},\ldots, x_{n}) = (gx_{\rho(g)(1)},\ldots, gx_{\rho(g)(n)})$. One easily checks that the two notions correspond to each other under the above equivalence. Segal introduced $\Gamma$-spaces in order to produce homotopy group completions. 1. A map $A\to B$ of homotopy associative, homotopy commutative $H$-spaces is said to be a *homotopy group completion* provided that 1. $\pi_{0}B$ is an abelian group and the map $\pi_{0}A\to \pi_{0}B$ is a group completion of the abelian monoid $\pi_{0}A$, and 2. $H_{*}(A,R)\to H_{*}(B,R)$ is the localization mapping $$H_{*}(A,R) \to {\mathbb{Z}}[\pi_{0}B]\otimes_{{\mathbb{Z}}[\pi_{0}A]}H_{*}(A,R),$$ for any commutative ring $R$. 2. Say that a $G$-space $A$ is an *equivariant homotopy commutative, associative $H$-space* if it is a homotopy commutative, associative $H$-space, the $H$-space structure map is equivariant, and the homotopies for associativity and commutativity can be taken to be equivariant. Say that an equivariant $H$-space map $A\to B$ is an *equivariant homotopy group completion* provided that $A^{K}\to B^{K}$ is a homotopy group completion for all subgroups $K\subseteq G$. Our basic example occurs when $X(-)$ is a special equivariant $\Gamma$-space (in $G$-$CW$-complexes). There is a functor from equivariant $\Gamma$-spaces to $G$-spectra generalizing the classical one for the trivial group $G$ as follows. Given an equivariant $\Gamma$-space $X$ we obtain a ${\mathcal{W}}_{G}$-functor, which we denote $\widehat{X}$ via $$\widehat{X}(M) = |B(M,\Gamma_{G}, PX)|$$ where $M$ is viewed as the functor $\operatorname{Map}(-,M):\Gamma_{G}^{op}\to {\mathcal{T}}_{G}$ which it represents and $B(-,-,-)$ denotes the two-sided bar construction. Write $$\mathbb{S}X = \{\widehat{X}(S^{V})\}$$ for the spectrum $\mathbb{U}\widehat{X}$ associated to the ${\mathcal{W}}_{G}$-space $\widehat{X}$. \[lem:Gspc\] Let $X$ be an equivariant $\Gamma$-space. 1. View $X(n)\to X(1)^{\times n}$ as a map of $G\times\Sigma_{n}$-spaces, where $(g,\sigma)$ acts on $X(n)$ via $(g,\sigma)\cdot x = X(\sigma)(gx)$ and on the $X(1)^{\times n}$ by $(g,\sigma)(x_{1},\ldots, x_{n}) = (gx_{\sigma(1)},\ldots, gx_{\sigma(n)})$. If $X(n)\to X(1)^{\times n}$ is a $G\times\Sigma_{n}$-weak equivariant equivalence for all $n$, then $X$ is special. 2. If $X$ is a special equivariant $\Gamma$-space, then $\mathbb{S}X$ is a positive $G-\Omega$-spectra and the map $$X(1){\simeq}\widehat{X}(S^{0}) \to \Omega\widehat{X}(S^{1})=\Omega \mathbb{S}X_{1}$$ is an equivariant group completion. See [@Shimakawa:Gloop]. If $X$ is an equivariant $\Gamma$-space or a $\Gamma_{G}$-space, $\underline{n}_{+}\mapsto X(n)^{H}$ defines an ordinary $\Gamma$-space. Given a $\Gamma$-space ${\mathcal{A}}(-)$ write $\mathbf{B}{\mathcal{A}} = ({\mathcal{A}}(\underline{1}_{+}), B{\mathcal{A}}(\underline{1}), B^{2}{\mathcal{A}}(\underline{1}),\ldots )$ for the associated spectrum as in [@Segal:gamma]. The condition in the following guarantees that the simplicial space $n\mapsto X(n)$ is good in the sense of [@Segal:gamma] \[lem:spcfix\] Let $X$ be a special equivariant $\Gamma$-space of the form $X(-) = |X^{\prime}(-)|$ where $X^{\prime}:\Gamma\to G{\mathrm{sSet}}$ and $H\subseteq G$ a subgroup. Then $\pi_{n}^{H}(\mathbb{S}X) {\cong}\pi_{n}(\mathbf{B}X^{H})$. Since $X$ is special, $\mathbb{S}X$ is a positive-$\Omega$-$G$-spectrum and so by [@MM:O Proposition V.3.2] we have that $\pi_{n}^{H}(\mathbb{S}X) = \pi_{n}(\mathbb{S}X^{H})$. Recall that if $A$ is an orthogonal $G$-spectrum and $V={\mathbb{R}}^{n}$ has trivial action then $A^{H}({\mathbb{R}}^{n}) = A({\mathbb{R}}^{n})^{H}$ (in terms of underlying pre-spectra: $(A^{H})_{n} = (A_{n})^{H}$). By [@Shimakawa:Gloop Proposition 1.2(c)] the map $|B(S^{n},\Gamma, X^{H})| \to |B(S^{n},\Gamma_{G}, X)^{H}|$ is a homotopy equivalence. By [@Segal:gamma Proposition 3.2], $B^{n}X^{H} = S^{n}\otimes_{\Gamma}X^{H}$ and because the overcategory $(\Gamma\downarrow S^{n})$ is filtered, the map $|B(S^{n},\Gamma, X^{H})|\to S^{n}\otimes_{\Gamma}X^{H}$ is a weak equivalence. Therefore $\pi_{n}(\mathbb{S}X)^{H} = \pi_{n}\mathbf{B}(X^{H})$. Pairings {#subsectionpairings} -------- The external product $X\barwedge Y$ of two ${\mathcal{W}}_{G}$-spaces $X$ and $Y$ is the ${\mathcal{W}}_{G}\times {\mathcal{W}}_{G}$-space given by $(X\barwedge Y)(A,B) = X(A)\wedge Y(B)$. The *smash product* $X\wedge Y$ of ${\mathcal{W}}_{G}$-spaces is defined as the left Kan extension of $X\barwedge Y$ along the functor $\wedge:{\mathcal{W}}_{G}\times {\mathcal{W}}_{G}\to {\mathcal{W}}_{G}$ given by $(A,B)\mapsto A\wedge B$, see [@AB]. By the universal property of Kan extension, giving a pairing $X\wedge Y \to Z$ of ${\mathcal{W}}_{G}$-spaces is equivalent to giving a map $X\barwedge Y \to Z\circ \wedge$ of $({\mathcal{W}}_{G}\times{\mathcal{W}}_{G})$-spaces. Since $\mathbb{U}$ is lax symmetric monoidal, the pairing $X\barwedge Y \to Z\circ \wedge$ defines a pairing of associated orthogonal $G$-spectra $\mathbb{U}X\wedge\mathbb{U}Y \to \mathbb{U}Z$. Let $X$, $Y$ be equivariant $\Gamma$-spaces and $X\barwedge Y$ is the equivariant $\Gamma\times\Gamma$-space $(\underline{p}_{+},\underline{q}_{+})\mapsto X(\underline{p}_{+})\wedge Y(\underline{q}_{+})$. Let $\wedge:\Gamma\times \Gamma\to \Gamma$ be defined by identifying $\underline{p}_{+}\wedge \underline{q}_{+}$ with $\underline{pq}_{+}$ via the lexicographical ordering. A map of equivariant $(\Gamma\times \Gamma)$-spaces $X\barwedge Y \to Z\circ \wedge$ determines a pairing of associated spectra, as we now explain. The functor $P$ which associates a $\Gamma_{G}$-space to an equivariant $\Gamma$-space, described in the previous section, can be extended in the evident way to a functor taking equivariant $(\Gamma\times\Gamma)$-spaces to $(\Gamma_{G}\times \Gamma_{G})$-spaces and we again denote this functor by $P$. A straightforward inspection shows that $P(X\barwedge Y) = PX\barwedge PY$. Together with the natural map $P(Z\circ \wedge) \to PZ\circ \wedge$, this implies that a map of equivariant $(\Gamma\times \Gamma)$-spaces $X\barwedge Y \to Z\circ\wedge$ gives rise to a map $PX\barwedge PY \to PZ\circ\wedge$ of $(\Gamma_{G}\times\Gamma_{G})$-spaces. This map in turn gives rise to a map $\widehat{X}\barwedge\widehat{Y}\to \widehat{Z}\circ\wedge$ of $({\mathcal{W}}_{G}\times{\mathcal{W}}_{G})$-spaces and therefore we obtain a pairing $\mathbb{S}X\wedge\mathbb{S}Y\to \mathbb{S}Z$. Homotopy colimits of $G$-spaces ------------------------------- Homotopy colimits can be viewed as the derived functors of the colimit functor. For our purposes it is important to use a functorial model for the homotopy colimit of a diagram of $G$-simplicial sets or spaces and we take the “standard model”. Explicitly, let $X:D\to {\mathcal{C}}$ be a functor, where ${\mathcal{C}}$ is the category of $G$-simplicial sets or spaces, then $$\operatorname*{\mathrm{hocolim}}_{D}X = |B(*,D,X)|$$ where $B(-,-,-)$ denotes the two-sided bar construction. Observe that this formula shows that $(\operatorname*{\mathrm{hocolim}}_{D}X)^{H} = \operatorname*{\mathrm{hocolim}}_{D}X^{H}$ for any subgroup $H\subseteq G$. [$G$]{}-modules --------------- If $X$ is a $G$-scheme then a coherent $G$-module on $X$ is a coherent ${\mathcal{O}}_{X}$-module ${\mathcal{M}}$ together with isomorphisms $\phi_{g}:{\mathcal{M}} \to g_{*}{\mathcal{M}}$ for each $g\in G$ such that $\phi_{e} = id$ and $\phi_{gh} = h_{*}\phi_{g}\phi_{h}$. If $X = \operatorname{\mathrm{Spec}}(R)$ then $R$ has a $G$-action which we write as a left-action. Specifying a coherent $G$-module ${\mathcal{M}}$ on $X$ is equivalent to specifying an $R$-module $M$ together with a $G$-action on $M$ which is compatible with the action on $R$ in the sense that $(g\cdot r)m = g\cdot (r\cdot (g^{-1}m))$ (i.e. $M$ is a module over the skew-group ring $R^{*}G$). Bivariant $K$-theories ====================== In this section we introduce the algebraic, semi-topological and topological bivariant $K$-theories with which we work in the paper. All of these are constructed as $G$-spectra. The construction of the algebraic bivariant $K$-theory spectrum as a $G$-spectrum makes use of the equivariant Grayson-Walker theorem proved by [Ø]{}stv[æ]{}r [@Ostvaer]. The bivariant semi-topological equivariant $K$-theory is constructed and studied in \[sub:sst\] and its topological counterparts are introduced and studied in \[sec:top\] and \[sec:tel\]. The comparison map between the semi-topological and topological $K$-theories is constructed and studied in the next section. The material in these sections corresponds mostly to material in Sections 3 and 4 in [@Walker:Thomason]. While the overall picture of the results presented here corresponds nicely to that in Walker’s paper, there are parts of the picture which differ. Before beginning, we point out some of the global differences of significance between our presentation of this material and the corresponding material there. First, Walker defines the semi-topological bivariant theory via topological spaces of algebraic maps while we use Friedlander-Walker’s simplicial model for this space, see Remark \[rem:mor\] below. Second, Walker makes use of $\Gamma$-spaces produced by taking nerves of certain topological categories while we prefer to simply describe our $\Gamma$-spaces as being obtained from a homotopy colimit of a certain diagram, see Remark \[rem:hocol\]. Algebraic $K$-theory {#sec:algK} -------------------- In this subsection we work with quasi-projective $G$-varieties over an arbitrary field $k$, where $G$ is a finite group whose order is coprime to $char(k)$ (though, only $k={\mathbb{C}}$ is used in later sections). Write ${\mathcal{P}}(G;X,Y)$ for the category of coherent $G$-modules on $X\times Y$ which are finite and flat over $X$ (this is the category ${\mathcal{P}}^{0}(G;X,Y)$ in the notation of [@Ostvaer]). This is an exact category, and we write ${\mathcal{K}}(G;X,Y)$ for the associated $K$-theory spectrum, as produced by Waldhausen’s construction. We explain how, when specialized to the case of finite groups, the material in [@Ostvaer] yields a $G$-spectrum ${\mathcal{K}}_{G}(X,Y)$ with the property that ${\mathcal{K}}_{G}(X,Y)^{H} {\simeq}{\mathcal{K}}(H;X\times\Delta^{\bullet}_{k},Y)$. Let $V$ be a finite dimensional $G$-representation over $k$. Then $V$ defines a $G$-bundle on $\operatorname{\mathrm{Spec}}(k)$ and we write ${\mathcal{V}}$ for this $G$-bundle. For any $G$-variety $Y$ over $k$, the pullback of ${\mathcal{V}}$ via the structure map $Y\to \operatorname{\mathrm{Spec}}(k)$ is a $G$-bundle on $Y$ which we denote as ${\mathcal{V}}_{Y}$. \[defn:G\] Let $X$,$Y$ be a quasi-projective $G$-schemes over $k$ and $V$ a representation of $G$. Write $\pi_{X}:X\times Y\to X$ for the projection map. Define the sheaf of sets ${\mathcal{G}}^{V}_{Y}(n):{\mathrm{Sch}}/k^{op} \to {\mathrm{Set}}$ to be the sheaf whose value on $X$ is the collection of quotient objects $[{\mathcal{V}}^{n}_{X\times Y}{\twoheadrightarrow}{\mathcal{M}}]$ which satisfy the following conditions. 1. The support of ${\mathcal{M}}$ is finite over $X$ and $(\pi_{X})_{*}{\mathcal{M}}$ is locally free. 2. The composition ${\mathcal{V}}^{n}_{X}\to (\pi_{X})_{*}{\mathcal{V}}^{n}_{X\times Y}\to (\pi_{X})_{*}{\mathcal{M}}$ is surjective. Forgetting the $G$-action, the underlying sheaf ${\mathcal{G}}^{V}_{Y}(n)$ is denoted by $G^{mn}_Y$ in [@Walker:Thomason], where $m=\dim V$. When condition (1) is satisfied we say that ${\mathcal{M}}$ is *finite and flat over $X$*. As remarked by Walker, condition (2) could be omitted and one would still obtain the same bivariant theory, see Remark \[rem:WGW\] and [@Ostvaer Theorem A.10]. Its inclusion has several advantages, one of which is that it allows the construction of functorial maps, e.g. Lemma \[lem:covariant\] below, which otherwise would merely be functorial up to homotopy. Let $Y$ be a $G$-variety over $k$. For each $g\in G$ we have isomorphisms $\phi_{g}:{\mathcal{V}}^{n}_{Y} \to g_{*}{\mathcal{V}}^{n}_{Y}$ such that $\phi_{e} = id$ and $\phi_{gh} = h_{*}\phi_{g}\phi_{h}$. For any $G$-variety $X$, we have a natural $G$-action on ${\mathcal{G}}^{V}_{Y}(n)(X)$ given by $$g\cdot[{\mathcal{V}}^{n}_{X\times Y}{\twoheadrightarrow}\mathcal{M}] = [{\mathcal{V}}^{n}_{X\times Y}\xrightarrow{\phi_{g}}g_{*}{\mathcal{V}}^{n}_{X\times Y}{\twoheadrightarrow}g_{*}\mathcal{M}].$$ This defines a $G$-action on the functor ${\mathcal{G}}^{V}_{Y}(n)$. By [@Walker:Thomason Lemma 2.2] (which is written for $k={\mathbb{C}}$, but it holds verbatim over any field $k$) the functor ${\mathcal{G}}^{V}_{Y}(n)$ is represented by a quasi-projective variety which we denote by the same symbol. In fact ${\mathcal{G}}^{V}_{Y}(n)$ is an open invariant subscheme of the Quot-scheme $\coprod_{r=0}^{n}Quot^{r}_{{\mathcal{V}}^{n}_{Y}/Y/\operatorname{\mathrm{Spec}}(k)}$, where $Quot^{r}_{{\mathcal{V}}^{n}_{Y}/Y/\operatorname{\mathrm{Spec}}(k)}$ is the functor which sends $U$ to the set of quotient objects $[{\mathcal{V}}^{n}_{U\times Y} {\twoheadrightarrow}{\mathcal{M}}]$ such that ${\mathcal{M}}$ is finite and flat over $U$ and $\pi_{*}{\mathcal{M}}$ is locally free of rank $r$ on $U$. The $G$-action on the functor ${\mathcal{G}}_{Y}^{V}(n):{\mathrm{Sch}}/k^{op}\to {\mathrm{Set}}$ defines an action on the representing scheme. Thus for a $G$-scheme $X$, we have an action on $\operatorname{Hom}_{{\mathrm{Sch}}/k}(X,\,{\mathcal{G}}^{V}_{Y}(n))$ defined by the usual formula, $g\cdot f$ is the function $x\mapsto gf(g^{-1}x)$. This action agrees with the previously described action on ${\mathcal{G}}^{V}_{Y}(n)(X)$. We now describe the equivariant $\Gamma$-space which will define the spectrum ${\mathcal{K}}_{G}(X,Y)$. Write $I$ for the category whose objects are $\underline{n}=\{1,2,\ldots, n\}$ for each $n\geq 0$ ( $\underline{0}$ is the empty set) and morphisms are injective, but not necessarily order-preserving, maps of sets. An injection $j:\underline{m}\hookrightarrow \underline{n}$ determines a map of coherent $G$-modules $j_{*}:{\mathcal{V}}^{m} \to {\mathcal{V}}^{n}$ where $j_{*}(e_{i}) = e_{j(i)}$. Write $j^{*}:{\mathcal{V}}^{n}{\twoheadrightarrow}{\mathcal{V}}^{m}$ for the transpose of this map. Explicitly we have $$j^{*}(e_{i})= \begin{cases} e_{k} & \textrm{if} \,\, j(k) = i,\,\,\textrm{and} \\ 0 & \textrm{if}\,\, i\notin \operatorname{\mathrm{im}}(j) . \end{cases}$$ Note that $j^{*}j_{*} = id$. We have induced morphisms $j^{*}:{\mathcal{V}}^{n}_{X\times Y}{\twoheadrightarrow}{\mathcal{V}}^{m}_{X\times Y}$ and precomposition with $j^{*}$ defines a natural transformation $j^{*}:{\mathcal{G}}_{Y}^{V}(m)(X)\to G_{Y}^{V}(n)(X)$. Note that $j^{*}[p:{\mathcal{V}}_{X\times Y}^{m}{\twoheadrightarrow}{\mathcal{M}}]=[q:{\mathcal{V}}_{X\times Y}^{n}{\twoheadrightarrow}{\mathcal{N}}]\in {\mathcal{G}}^{V}_{Y}(n)(X)$ exactly when there exists an isomorphism ${\mathcal{M}}\xrightarrow{{\cong}} {\mathcal{N}}$ such that the diagram commutes $$\xymatrix{ {\mathcal{V}}^{m}_{X\times Y} \ar@{->>}[r]^{p} & {\mathcal{M}} \ar@{-->}[d]^{{\cong}} \\ {\mathcal{V}}^{n}_{X\times Y} \ar@{->>}[u]^{j^{*}} \ar@{->>}[r]^{q} & {\mathcal{N}} . }$$ The equivariant maps $j^{*}$ make the assignment $$\underline{n}\mapsto \operatorname{Hom}_{{\mathrm{Sch}}/k}(X,{\mathcal{G}}_{Y}^{V}(n))$$ into a functor $I\to GSet$. \[lem:covariant\] Suppose that $f:Y\to Y'$ is an equivariant map of quasi-projective $G$-varieties over $k$. There is an equivariant morphism of $G$-varieties $f_{*}:{\mathcal{G}}^{V}_{Y}(n)\to {\mathcal{G}}^{V}_{Y'}(n)$, defined via the equivariant natural transformation of functors, $f_{*}:{\mathcal{G}}^{V}_{Y}(n)(-)\to {\mathcal{G}}^{V}_{Y'}(n)(-)$, which on $U$ sends the quotient $[q:{\mathcal{V}}^{n}_{U\times Y}{\twoheadrightarrow}{\mathcal{M}}]$ to $[{\mathcal{V}}^{n}_{U\times Y'}{\twoheadrightarrow}(id\times f)_{*}{\mathcal{M}}]$. Moreover $f_{*}$ is natural with respect to maps in $I$ and thus defines a transformation of $I$-functors. The map $f_{*}$ is well-defined as a consequence of condition (2) in Definition \[defn:G\]. The required naturality and equivariance statements are clear. Define the presheaf of $G$-simplicial sets ${\mathcal{A}}_{G}(-,\,Y)^{V}$ by $${\mathcal{A}}_{G}(X,\,Y)^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{{\mathrm{Sch}}/k}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(-)).$$ We extend this definition to pairs $(Y,y_{0})$ consisting of a quasi-projective $G$-variety together with an invariant $k$-rational basepoint $y_{0}\in Y$. This is useful later, when $k=\operatorname{\mathrm{Spec}}({\mathbb{C}})$, to discuss comparisons to the topological setting. Usually the pointed variety will be $Y_{+}= Y\coprod \operatorname{\mathrm{Spec}}(k)$, where we have adjoined a disjoint basepoint. Note that ${\mathcal{G}}_{y_{0}}^{V}(n)\subseteq {\mathcal{G}}_{Y}^{V}(n)$ and this inclusion is functorial in the variable $n$. Define $${\mathcal{G}}_{(Y,y_{0})}^{V}(n)(X) = {\mathcal{G}}_{Y}^{V}(n)(X)/{\mathcal{G}}_{y_{0}}^{V}(n)(X)$$ and $${\mathcal{A}}_{G}(X,\,(Y,y_{0}))^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{{\mathrm{Sch}}/k}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{(Y,y_{0})}^{V}(-)).$$ Note that ${\mathcal{A}}_{G}(X,\,(Y_{+},*))^{V} = {\mathcal{A}}_{G}(X,\,Y)^{V}$. Usually we omit the base-point from the notation when the context makes it clear what is meant. The $G$-simplicial sets ${\mathcal{A}}_{G}(X,\,Y)^{V}$ are clearly contravariantly natural in the first variable and are covariantly natural in the second variable by Lemma \[lem:covariant\]. Using the covariant naturality in the second variable we obtain equivariant $\Gamma$-spaces $$\underline{n}_{+}\mapsto |{\mathcal{A}}_{G}(X,\,\underline{n}_{+}\wedge Y_{+})^{V}|.$$ More generally the assignment $S\mapsto |{\mathcal{A}}_{G}(X,\, S\wedge Y_{+})^{V}|$ for a based finite $G$-set $S$, extends the above assignment to a $\Gamma_{G}$-space. By the equivalence of categories (\[eqn:gammaeq\]) it makes no difference whether we work with the equivariant $\Gamma$-space displayed above or with this $\Gamma_{G}$-space. We are most interested in the case of the regular representation $V=k[G]$. Recall that an equivariant $\Gamma$-space ${\mathcal{A}}$ naturally gives rise to a $G$-spectrum (see Lemma \[lem:Gspc\]) which we denote $\mathbb{S}{\mathcal{A}}$. \[defKalg\] Let $X$,$Y$ be quasi-projective $G$-schemes. Define the $G$-spectrum ${\mathcal{K}}_{G}(X,Y)$ by $${\mathcal{K}}_{G}(X,Y)={\mathbb{S}}|{\mathcal{A}}_{G}(X,\,-\wedge Y_+)^{k[G]}|,$$ the spectrum associated to the equivariant $\Gamma$-space $|{\mathcal{A}}_{G}(X,\,-\wedge Y_+)^{k[G]}|$. Write $K^{G,\,{\mathbb{A}}^{1}}_{*}(X,Y) = \pi^{G}_{*}{\mathcal{K}}_{G}(X,Y)$ for the homotopy groups of this spectrum. More generally if $H\subseteq G$ is a subgroup, write $K^{H,\,{\mathbb{A}}^{1}}_{*}(X,Y) = \pi^{H}_{*}{\mathcal{K}}_{G}(X,Y)$. The spectra ${\mathcal{K}}_{G}(X,Y)$ are contravariantly natural in the first variable and are covariantly natural in the second variable by Lemma \[lem:covariant\]. Later in this section we will see that $K^{G,\,{\mathbb{A}}^{1}}_{*}(X,Y) = \pi_{*}{\mathcal{K}}(G;X\times\Delta^{\bullet}_{k},Y)$. We now introduce a slightly different model which is often convenient to work with. Write $$\operatorname{Hom}_{{\mathrm{Sch}}/k}(X,\,{\mathcal{G}}_{Y}^{V}(\infty))=\operatorname*{\mathrm{colim}}_{n}\operatorname{Hom}_{{\mathrm{Sch}}/k}(X ,\,{\mathcal{G}}_{Y}^{V}(n))$$ where the colimit is over the standard inclusions $\underline{n}\subseteq \underline{n+1}$ given by $i\mapsto i$ and the transition maps are induced by precomposition with the canonical surjections ${\mathcal{V}}^{n+1} \to {\mathcal{V}}^{n}$. Thus an element is a quotient object $[{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{p} {\mathcal{M}}]$ where ${\mathcal{V}}^{\infty}_{X\times Y} \to {\mathcal{M}}$ factors as ${\mathcal{V}}^{\infty}_{X\times Y} \to {\mathcal{V}}^{N}_{X \times Y}\to {\mathcal{M}}$ for some $N$. If $j:\mathbb{N}\to \mathbb{N}$ is an injection we have an induced map given by $j^{*}[{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{p} {\mathcal{M}}] = [p\tilde{j}^{*}:{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{}{\mathcal{M}}]$, and similarly for pointed varieties $(Y,y_{0})$. In what follows it will be convenient to write $\infty$ for the set $\mathbb{N}$. Let $\tilde{I}$ be the category whose objects are the finite sets $\underline{n}$ together with the set $\infty$ and whose morphisms are injections. Let $M\subseteq \tilde{I}$ be the full subcategory containing the object $\infty$. The category $M$ consists of one object $\infty$ and $\operatorname{Hom}_{M}(\infty,\infty)$ is the monoid (under composition) of injective set maps $\mathbb{N}\to \mathbb{N}$. We will abuse notation and also write $M$ for this monoid. \[prop:alginf\] Let $X$ and $(Y,y_{0})$ be quasi-projective $G$-varieties over $k$. The natural maps $$\begin{gathered} \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(-)) \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}} \operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(-)) \\ \leftarrow \operatorname*{\mathrm{hocolim}}_{M}\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(\infty))\end{gathered}$$ are equivariant weak equivalences. The proposition could be obtained as a particular case of [@Shipley:THH Proposition 2.2.9] but for the convenience of the reader we sketch the full argument. Write $F(-) = \operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(-))$. Because the overcategory $(I\downarrow\infty)$ is filtered, we have the equivariant weak equivalence $L_{hK}F(\infty)\to L_{K}F(\infty)$ where $L_{K}F(-)$ and $L_{hK}(-)$ respectively denote the left Kan extension and the homotopy left Kan extension of $F|_{I}$ along $I\subseteq \tilde{I}$. Moreover, we have that the functors $F$,$L_{K}F:\tilde{I}\to G{\mathrm{sSet}}$ are equal. Since $\operatorname*{\mathrm{hocolim}}_{I} F \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}}L_{hK}F$ is an equivariant weak equivalence we have that $\operatorname*{\mathrm{hocolim}}_{I} F \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}}F$ is an equivariant weak equivalence as well. Since $M\subseteq \tilde{I}$ is right cofinal we have that $\operatorname*{\mathrm{hocolim}}_{M}F(\infty) \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}}F$ is an equivariant weak equivalence as well. \[prop:Malgwkeq\] The map $$\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(\infty)) \to \operatorname*{\mathrm{hocolim}}_{M}\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(\infty))$$ is an equivariant weak equivalence. Write $F(\infty)=\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{Y}^{V}(\infty))$. It suffices to show that $M$ acts on $F(\infty)$ by equivariant weak equivalences. Indeed, in this case it follows from [@Quillen:Ktheory Lemma p.90] that we have an equivariant homotopy fiber sequence $$F(\infty) \to \operatorname*{\mathrm{hocolim}}_{M}F(\infty) \xrightarrow{\pi} BM.$$ The map $\pi$ is surjective and as shown in [@GW:Kmodels proof of Lemma 3.1] $BM$ is contractible, from which the result follows. To see that $M$ acts by equivariant weak equivalences we proceed as follows. Let $\alpha\in M$ be an injection. Then $\alpha$ acts via $\alpha^{*}:F(\infty)\to F(\infty)$ where the quotient object $\alpha^{*}[q]$ is given by $$\alpha^{*}q(e_{i}) = \begin{cases} q(e_{j}) & \textrm{if} \,\, \alpha(j) = i,\,\,\textrm{and} \\ 0 & \textrm{if}\,\, i\notin \operatorname{\mathrm{im}}(\alpha) . \end{cases}$$ Write $t:\Delta^{1}_{k}{\cong}{\mathbb{A}}^{1}_{k}$ for any map sent to $0$ and $1$ under the face maps. An $n$-simplex of $F(\infty)\times \Delta^{1}$ is a pair $([q:{\mathcal{V}}^{\infty}_{X\times\Delta^{n}_{k}\times Y}\to {\mathcal{M}}], \delta:[n]\to [1])$. Associate to the above pair the quotient $[H]=[H([q],\delta)]$ defined by $$H([q],\delta) = \begin{cases} q(e_{j}) & \textrm{if} \,\, \alpha(j) = i,\,\,\textrm{and} \\ \delta^*(t) \cdot q(e_{i}) & \textrm{if}\,\, i\notin \operatorname{\mathrm{im}}(j) . \end{cases}$$ Here $t:\Delta_{k}^{1}\to {\mathbb{A}}_{k}^{1}$ is viewed as a global section of ${\mathcal{O}}_{\Delta_{k}^{1}}$ and $\delta^*(t)$ is thus viewed as a global section of ${\mathcal{O}}_{X\times \Delta^{n}_{k} \times Y}$ via pullback. This is easily seen to be a well-defined, equivariant map of simplicial sets and satisfies $H([q],0) = \alpha^{*}[q]$ and $H([q],1) = [q]$ as desired. \[rem:WGW\] The bivariant $K$-theories constructed above follow Walker’s constructions in [@Walker:Thomason] in the nonequivariant setting. In [@GW:Kmodels] Grayson-Walker use a slightly different construction and [@Ostvaer] is written following this version. These are related as follows. Let $V$ be any representation such that $V^{\infty}$ is a complete universe for $G$ (in particular the regular representation $k[G]$ satisfies this property). The $G$-ind-variety $K_{Y,0}^{(n)}$ used in [@Ostvaer] is the variety parameterizing $n$-tuples $[p_{1}:{\mathcal{V}}_{X\times Y}^{\infty}{\twoheadrightarrow}{\mathcal{M}}_{1}], \ldots, [p_{n}:{\mathcal{V}}_{X\times Y}^{\infty}{\twoheadrightarrow}{\mathcal{M}}_{n}]$ of finite and flat quotient objects which are in “general position” in the sense that $(p_{1},\ldots,p_{n})^{t}:{\mathcal{V}}_{X\times Y}^{\infty}\to \oplus{\mathcal{M}}_{i}$ is surjective. An element of $\operatorname{Hom}(X,\,{\mathcal{G}}_{\underline{n}_{+}\wedge Y_{+}}^{V}(\infty))$ is a quotient object $[{\mathcal{V}}^{\infty}_{X\times (\underline{n}\times Y)}{\twoheadrightarrow}{\mathcal{M}}]$ which is finite and flat over $X$ and satisfies the additional condition that ${\mathcal{V}}^{\infty}_{X}\to (\pi_{X})_{*}{\mathcal{M}}$ remains surjective. As remarked by Walker [@Walker:Thomason p. 219], giving such a quotient object naturally yields an $n$-tuple in general position. In other words we have a natural inclusion of $G$-ind-varieties $G^{V}_{\underline{n}\times Y}(\infty) \hookrightarrow K_{Y,0}^{(n)}$, identifying $\operatorname{Hom}(X,G^{V}_{\underline{n}\times Y}(\infty))$ with $n$-tuples of quotient objects $([p_{i}:{\mathcal{V}}^{\infty}_{X\times Y} {\twoheadrightarrow}{\mathcal{M}}_{i}])$ which are general position and satisfy the extra condition that ${\mathcal{V}}_{X}^{\infty}{\twoheadrightarrow}(\pi_{X})_{*}{\mathcal{M}}_{i}$ remains surjective. We thus have natural inclusions of equivariant $\Gamma$-spaces $$\operatorname{Hom}(X\times\Delta^{\bullet}_{k}, G^{V}_{-\times Y}(\infty)) \hookrightarrow \operatorname{Hom}(X\times\Delta^{\bullet}_{k},K_{Y,0}^{(-)})$$ which induces an equivalence of the associated $K$-theory spectra by [@Ostvaer Theorem A.10]. \[prop:algspecial\] For quasi-projective $G$-varieties $X$ and $Y$, the equivariant $\Gamma$-spaces $$\underline{n}_{+}\mapsto |{\mathcal{A}}_{G}(X,\,\underline{n}_{+}\wedge Y_{+})^{V}|$$ and $$\underline{n}_{+}\mapsto |\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{\underline{n}_{+}\wedge Y_{+}}^{ V}(\infty))|$$ are degreewise equivariantly weakly equivalent, and are both special. Propositions \[prop:alginf\] and \[prop:Malgwkeq\] imply that the $\Gamma_{G}$-space $|{\mathcal{A}}_{G}(X,\,-\wedge Y_{+})^{V}|$ is equivariantly weakly equivalent to the $\Gamma_{G}$-space $|\operatorname{Hom}(X\times{\Delta^{\bullet}_{k}},\,{\mathcal{G}}_{- \wedge Y_{+}}^{ V}(\infty))|$. Therefore it suffices to show that the latter is special. By Lemma \[lem:Gspc\] it suffices to show that $$\operatorname{Hom}(X\times{\Delta^{\bullet}_{k}},\,{\mathcal{G}}_{\underline{n}_{+} \wedge Y}^{ V}(\infty))\to \operatorname{Hom}(X\times{\Delta^{\bullet}_{k}},\,{\mathcal{G}}_{Y}^{ V}(\infty))^{\times n}$$ is a $G\times \Sigma_{n}$-equivalence. This follows by observing that the previous remark allows the argument given in [@GW:Kmodels Lemma 2.2] to carry over to our setting. That is, the maps defined there are equivariant, preserve the additional surjectivity condition and the explicit homotopy written there is $G\times\Sigma_{n}$-equivariant and preserves the additional surjectivity condition. We write ${\mathcal{P}}(G;X,Y)$ for the exact category of coherent $G$-modules on $X\times Y$ which are finite and flat over $X$ and write ${\mathcal{K}}(G;X,Y)$ for the associated the $K$-theory spectrum and ${\mathcal{K}}(G;X\times\Delta^{\bullet}_{k},Y)$ for the realization of the simplicial spectrum $d\mapsto {\mathcal{K}}(G;X\times\Delta^{d}_{k},Y)$. Note that when $X$ is smooth and $Y=\operatorname{\mathrm{Spec}}(k)$ then ${\mathcal{K}}(G;X)\xrightarrow{{\simeq}} {\mathcal{K}}(G;X\times\Delta^{\bullet}_{k},\operatorname{\mathrm{Spec}}(k))$ where ${\mathcal{K}}(G;X)$ is the equivariant algebraic $K$-theory spectrum introduced by Thomason [@Thomason:algKgroup]. \[prop:fix\] Let $X$, $Y$ be quasi-projective $G$-varieties and $H\subseteq G$ a subgroup. There are natural isomorphisms $$K_{n}^{H,{\mathbb{A}}^{1}}(X,Y)= \pi_{n}^{H}{\mathcal{K}}_{G}(X,Y) {\cong}\pi_{n}{\mathcal{K}}(H;X\times\Delta^{\bullet}_{k},Y).$$ By Proposition \[prop:algspecial\], we have that ${\mathcal{K}}_{G}(X,Y)$ is equivariantly weakly equivalent to the $G$-spectrum associated to $|\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{-\wedge Y_{+}}^{k[G]}(\infty))|$. Therefore by Lemma \[lem:spcfix\], we have $\pi_{n}^{H}{\mathcal{K}}_{G}(X,Y)=\pi_{n}\mathbf{B}|\operatorname{Hom}(X\times\Delta^{\bullet}_{k},\,{\mathcal{G}}_{-\wedge Y_{+}}^{k[G]}(\infty))^{H}|$. Now [@Ostvaer Theorems A.4 and A.10] show that this last spectrum is naturally weakly equivalent to the spectrum ${\mathcal{K}}(H;X\times\Delta^{\bullet}_{k},Y)$. In particular we have that $$K_{0}^{G,{\mathbb{A}}^{1}}(X,Y) = \operatorname{\mathrm{coker}}(i_{1}^{*}-i_{0}^{*}:K_{0}^{G,\,alg}(X\times{\mathbb{A}}^{1},Y)\to K_{0}^{G,\,alg}(X,Y)),$$ where $i_{k}$ is the inclusion at $k\in{\mathbb{A}}^{1}$ and $K_{0}^{G,\,alg}(X,Y)=\pi_{0}{\mathcal{K}}(G;X,Y)$. Semi-topological $K$-theory {#sub:sst} --------------------------- We now explain how to construct an equivariant version of the bivariant semi-topological $K$-theory introduced by Walker in [@Walker:Thomason section 2]. We begin by recalling a construction of Friedlander-Walker. If $F$ is a presheaf of sets on ${\mathrm{Sch}}/{\mathbb{C}}$ and $T$ is a topological space then $F(T)$ is defined as the left Kan extension of $F$ along the functor ${\mathrm{Sch}}/{\mathbb{C}}\to Top$ given by $X\mapsto X^{an}$. Explicitly $F(T)$ is the filtered colimit $$F(T) = \operatorname*{\mathrm{colim}}_{T\to U^{an}}F(U)$$ where the colimit is over continuous maps $T\to U^{an}$ with $U$ a (possibly singular) variety. Applying this in particular to the standard topological simplices $\Delta^{n}_{top}$ yields a presheaf of simplicial sets $n\mapsto F(\Delta^{n}_{top})$. More generally $F$ could be a presheaf of simplicial sets or spectra and we obtain a presheaf of bisimplicial sets or simplicial spectra and write $F(\Delta^{\bullet}_{top})$ for its realization. Using this construction we obtain the functor $I\to G{\mathrm{sSet}}$, $$\underline{n}\mapsto \operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(n)).$$ Define the presheaf of $G$-simplicial sets ${\mathcal{A}}^{{\mathrm{sst}}}_{G}(-,\,Y)^{V}$ by $${\mathcal{A}}^{{\mathrm{sst}}}_{G}(X,\,Y)^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(-)).$$ \[rem:mor\] We make use of the simplicial mapping spaces $\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}}, \, Y)$ rather than the topological spaces $\operatorname{Mor}(X,Y)$, which are used in [@Walker:Thomason]. Shortly after [@Walker:Thomason] was written, Friedlander-Walker developed techniques, especially in [@FW:ratisos], which make the spaces $\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}}, \, Y)$ easier to work with than the conceptually attractive space $\operatorname{Mor}(X,\,Y)$. When $X$ and $Y$ are projective, [@FW:sstfct Corollary 4.3] shows that we have a natural isomorphism of simplicial sets $\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,Y) {\cong}\operatorname{\mathrm{Sing}}_{\bullet}\operatorname{Mor}(X,\,Y)$. \[rem:hocol\] For $G$ trivial, Walker [@Walker:Thomason section 3.1] introduces ${\mathcal{A}}^{{\mathrm{sst}}}_{G}(X,\,Y)^{V}$ (or rather ${\mathcal{A}}^{{\mathrm{semi}}}_{G}(X,\,Y)^{V}$) as the nerve of a topological category. As noted in Lemma 3.2 of loc. cit., this agrees with the construction as above. It will also be convenient to extend this definition to pairs $(Y,y_{0})$ consisting of a quasi-projective $G$-variety together with an invariant basepoint $y_{0}\in Y$. Note that ${\mathcal{G}}_{y_{0}}^{V}(n)\subseteq {\mathcal{G}}_{Y}^{V}(n)$ and this inclusion is functorial as well in the variable $n$. Define $${\mathcal{G}}_{(Y,y_{0})}^{V}(n)(X) = {\mathcal{G}}_{Y}^{V}(n)(X)/{\mathcal{G}}_{y_{0}}^{V}(n)(X)$$ and $${\mathcal{A}}^{{\mathrm{sst}}}_{G}(X,\,(Y,y_{0}))^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{(Y,y_{0})}^{V}(-)).$$ Note that ${\mathcal{A}}^{sst}_{G}(X,\,(Y_{+},*))^{V} = {\mathcal{A}}^{sst}_{G}(X,\,Y)^{V}$. Usually we omit the base-point from the notation when the context makes it clear what is meant. The $G$-simplicial sets ${\mathcal{A}}^{sst}_{G}(X,\,Y)^{V}$ are clearly contravariantly natural in the first variable and are covariantly natural in the second variable by Lemma \[lem:covariant\]. By the covariant naturality in the second variable we obtain equivariant $\Gamma$-spaces $$\underline{n}_{+}\mapsto |{\mathcal{A}}^{sst}_{G}(X,\,\underline{n}_{+}\wedge Y_{+})^{V}|.$$ We are most interested in the case where $V={\mathbb{C}}[G]$. Recall that an equivariant $\Gamma$-space ${\mathcal{A}}$ gives rise to a $G$-spectrum (see Lemma \[lem:Gspc\]) which we write $\mathbb{S}{\mathcal{A}}$. \[defKsst\] Let $X$ and $Y$ be quasi-projective $G$-varieties. The *bivariant semi-topological $K$-theory spectrum* ${\mathcal{K}}^{sst}_{G}(X,Y)$ is the $G$-spectrum defined by $${\mathcal{K}}^{sst}_{G}(X,Y)={\mathbb{S}}|{\mathcal{A}}^{\mathrm{sst}}(X,\,-\wedge Y_+)^{{\mathbb{C}}[G]}|.$$ The *bivariant semi-topological $K$-theory groups* defined to be the homotopy groups $$K^{G,sst}_{*}(X,Y) = \pi^{G}_{*}{\mathcal{K}}^{sst}_{G}(X,Y)$$ of this spectrum. The spectra ${\mathcal{K}}^{sst}_{G}(X,Y)$ are evidently contravariantly natural in the first variable and are covariantly natural in the second variable by Lemma \[lem:covariant\]. \[lem:delsst\] For any quasi-projective $G$-varieties $X$,$Y$ we have a natural equivariant weak equivalence $${\mathcal{K}}_{G}^{sst}(X,Y)\xrightarrow{{\simeq}} {\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Y)$$ Let $V$ denote the regular representation. Note that if ${\mathcal{A}}= \operatorname*{\mathrm{colim}}_{i}{\mathcal{A}}_{i}$ is a filtered colimit of equivariant $\Gamma$-spaces then $\operatorname*{\mathrm{colim}}_{i}\mathbb{S}{\mathcal{A}}_{i} {\cong}\mathbb{S}{\mathcal{A}}$. Therefore we have $$\begin{aligned} {\mathcal{K}}_{G}(X\times \Delta^{n}_{top},Y) & {\cong}\mathbb{S}|{\mathcal{A}}_{G}(X\times\Delta^{n}_{top}, - \wedge Y_{+})^{V}|\\ & = \mathbb{S}|\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{{\mathbb{C}}}\times\Delta^{n}_{top},{\mathcal{G}}^{V}_{-\wedge Y_{+}}(-))|.\end{aligned}$$ The projection $\Delta^{\bullet}_{{\mathbb{C}}}\to \Delta^{0}_{{\mathbb{C}}}$ induces a natural map ${\mathcal{A}}^{sst}_{G}(X,Y)^{V}_{n} \to {\mathcal{A}}_{G}(X\times\Delta^{n}_{top},Y)^{V}$ and allowing $n$ to vary yields a natural map of bisimplicial sets. Taking realizations and associated spectra yields the map $$\mathbb{S}|{\mathcal{A}}^{sst}_{G}(X,Y)^{V}| \to \mathbb{S}|{\mathcal{A}}_{G}(X\times\Delta^{\bullet}_{top}, Y)^{V}|$$ and it remains to show that this is an equivariant weak equivalence. Note that ${\mathcal{A}}_{G}(X\times\Delta^{n}_{top},-\wedge Y_{+})^{V}$ is a special equivariant $\Gamma$-space, being a filtered colimit of such. By Lemma \[lem:spcfix\], it suffices to show that $$\mathbf{B}|\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{top},{\mathcal{G}}^{V}_{ Y})^{H}| \to \mathbf{B}|\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{{\mathbb{C}}}\times\Delta^{\bullet}_{top},{\mathcal{G}}^{V}_{ Y})^{H}|$$ is an equivalence of spectra for any subgroup $H\subseteq G$. The map of associated infinite loop spaces $$\Omega B|\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{top},{\mathcal{G}}^{V}_{ Y})^{H}| \to \Omega B|\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times\Delta^{\bullet}_{{\mathbb{C}}}\times\Delta^{\bullet}_{top},{\mathcal{G}}^{V}_{ Y})^{H}|$$ is a homology equivalence by [@FW:compK Lemma 1.2] and the result follows. The above together with the map induced by the projection $\Delta^{d}_{top}\to \Delta^{0}_{top}$ give us the natural transformations $$\label{eqn:algsst} {\mathcal{K}}_{G}(X, Y) \to {\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Y)\xleftarrow{{\simeq}}{\mathcal{K}}_{G}^{sst}(X,Y).$$ Write $$\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X,\,{\mathcal{G}}_{Y}^{V}(\infty))=\operatorname*{\mathrm{colim}}_{n}\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X ,\,{\mathcal{G}}_{Y}^{V}(n))$$ where the colimit is over the standard inclusions $\underline{n}\subseteq \underline{n+1}$ given by $i\mapsto i$ and the transition maps are induced by precomposition with the canonical surjections ${\mathcal{V}}^{n+1} \to {\mathcal{V}}^{n}$. Thus an element is a quotient object $[{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{p} {\mathcal{M}}]$ where ${\mathcal{V}}^{\infty}_{X\times Y} \to {\mathcal{M}}$ factors as ${\mathcal{V}}^{\infty}_{X\times U\times Y} \to {\mathcal{V}}^{N}_{X\times U \times Y}\to {\mathcal{M}}$ for some $m$. If $j:\mathbb{N}\to \mathbb{N}$ is an injection we have an induced map given by $j^{*}[{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{p} {\mathcal{M}}] = [p\tilde{j}^{*}:{\mathcal{V}}^{\infty}_{X \times Y} \xrightarrow{}{\mathcal{M}}]$. Similarly for pointed varieties $(Y,y_{0})$. \[rem:twomodels\] We will see in Corollary \[niceKsst\] below that $\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))$ gives rise to a $G$-spectrum that is weakly equivalent to ${\mathcal{K}}^{sst}_{G}(X,Y)$ which was obtained from $\operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(-))$. Because the first model is defined using a filtered colimit it is in many ways easier to work with and indeed we rely on this model to deduce many properties of our bivariant $K$-theory spectra. However the second model is better suited for the pairings and operations appearing in Section \[sec:pair\] which is crucial for this paper and so it is crucial to have both models available. In fact, there is a third model, namely the one provided by Lemma \[lem:delsst\], which allows for a convenient comparison map to equivariant algebraic $K$-theory as used in section \[sec:algThom\]. Recall that we write $\tilde{I}$ for the category whose objects are $\underline{n}$ and the set $\mathbb{N}$, which we denote by $\infty$, and whose morphisms are injections. Let $M\subseteq \tilde{I}$ be the full subcategory containing the object $\infty$. We also write $M$ for $\operatorname{Hom}_{M}(\infty,\infty)$, which is the monoid (under composition) of injective maps $\mathbb{N}\to \mathbb{N}$. \[prop:sstinf\] Let $X$ and $(Y,y_{0})$ be quasi-projective complex $G$-varieties. The natural maps $$\begin{gathered} \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(-)) \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}} \operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(-)) \\ \leftarrow \operatorname*{\mathrm{hocolim}}_{M}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))\leftarrow \operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))\end{gathered}$$ are equivariant weak equivalences. The proof that the first two arrows are weak equivalences is exactly as for Proposition \[prop:alginf\]. The proof of the last one is a variant of the argument in the proof of Proposition \[prop:Malgwkeq\]. That is, it suffices to show that $M$ acts by equivariant weak equivalences on $\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))$. We write $F(\infty)=\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))$. Let $\alpha\in M$ be an injection. Then $\alpha$ acts via $\alpha^{*}:F(\infty)\to F(\infty)$ where the quotient object $\alpha^{*}[q]$ is given by $$\alpha^{*}q(e_{i}) = \begin{cases} q(e_{j}) & \textrm{if} \,\, \alpha(j) = i,\,\,\textrm{and} \\ 0 & \textrm{if}\,\, i\notin \operatorname{\mathrm{im}}(\alpha) . \end{cases}$$ Let $g:\Delta^{1}_{top}\to ({\mathbb{A}}^{1})^{an}$ be a map which sends $0$ to $0$ and $1$ to $1$. We define a simplicial homotopy $H:F(\infty)\times \Delta^{1} \to F(\infty)$ between $\alpha^{*}$ and $id$ as follows. An $n$-simplex of $F(\infty)\times \Delta^{1}$ is represented by a triple $$(f:\Delta^{n}_{top}\to U^{an},[q:{\mathcal{O}}^{\infty}_{X\times U \times Y}\to {\mathcal{M}}], \delta:[n]\to [1]).$$ We denote by ${\mathcal{M}}'$ the pullback of ${\mathcal{M}}$ to $X\times U\times {\mathbb{A}}^{1} \times Y$ and associate to the above triple the element $(f\times \delta^{*}(g):\Delta^{n}_{top}\to (U\times {\mathbb{A}}^{1})^{an}, H)\in F(\infty)$ where $[H=H(f,[q],\delta):{\mathcal{O}}^{\infty}_{X\times U\times {\mathbb{A}}^{1}\times Y} \to {\mathcal{M}}']$ is $$H(f,[q],\delta) = \begin{cases} q(e_{j}) & \textrm{if} \,\, \alpha(j) = i,\,\,\textrm{and} \\ t \cdot q(e_{i}) & \textrm{if}\,\, i\notin \operatorname{\mathrm{im}}(j) . \end{cases}$$ Here $t=id:{\mathbb{A}}^{1}\to {\mathbb{A}}^{1}$ is viewed as a global section of ${\mathcal{O}}_{{\mathbb{A}}^{1}}$ and hence of ${\mathcal{O}}_{X\times U\times {\mathbb{A}}^{1} \times Y}$ via pullback. This is easily seen to be a well-defined, equivariant map of simplicial sets and satisfies $H([q],0) = \alpha^{*}[q]$ and $H([q],1) = [q]$ as desired. \[prop:sstspecial\] For quasi-projective $G$-varieties $X$ and $Y$ the equivariant $\Gamma$-spaces $$\underline{n}_{+}\mapsto |{\mathcal{A}}^{sst}_{G}(X,\,\underline{n}_{+}\wedge Y_{+})^{V}|$$ and $$\underline{n}_{+}\mapsto |\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{\underline{n}_{+}\wedge Y}^{ V}(\infty))|$$ are degreewise equivariantly weakly equivalent, and they are both special. Using Proposition \[prop:sstinf\], the proof is similar to the one of Proposition \[prop:algspecial\]. \[niceKsst\] For quasi-projective $G$-varieties $X$ and $Y$ there are equivariant weak equivalences of $G$-spectra $${\mathcal{K}}^{sst}_{G}(X,Y) \simeq \mathbb{S} |\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{-\wedge Y}^{ V}(\infty))|$$ and hence equivariant weak equivalences of associated infinite loop spaces $$\Omega^{\infty}{\mathcal{K}}^{sst}_{G}(X,Y) \simeq \Omega B|\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{ V}(\infty))|.$$ The first weak equivalence follows from Proposition \[prop:sstspecial\] and the second one from Lemma \[lem:Gspc\]. We finish this section by showing that the group $K_{0}^{G,\,sst}(X,Y)$ has the expected description in terms of certain coherent $G$-modules modulo algebraic equivalence. Let ${\mathcal{M}}_{1}$ and ${\mathcal{M}}_{2}$ be two coherent $G$-modules on $X\times Y$ which are finite and flat over $X$. We say that ${\mathcal{M}}_{1}$ and ${\mathcal{M}}_{2}$ are *algebraically equivalent* if there is a smooth, connected curve $C$ (without $G$-action), two closed points $c_{1},c_{2}\in C$, a coherent $G$-module ${\mathcal{N}}$ on $X\times C \times Y$ which is finite and flat over $X\times C$ such that $\iota_{k}^{*}{\mathcal{N}} = {\mathcal{M}}_{k}$, where $\iota_{k}$ is inclusion $X\times \{c_{k}\}\times Y \subseteq X\times C \times Y$. Write $\sim_{alg}$ for this equivalence relation. We write $K_{*}^{G,\,alg}(X,Y)$ for the algebraic $K$-theory of the exact category ${\mathcal{P}}_{G}(X,Y)$ of coherent $G$-modules on $X\times Y$ which are finite and flat over $X$. \[thm:algeq\] Let $X$ and $Y$ be quasi-projective $G$-varieties. We have an isomorphism $$K_{0}^{G,\,sst}(X,Y) {\cong}K_{0}^{G,\,alg}(X,Y)/\sim_{alg},$$ which is contravariantly natural in the first variable and covariantly in the second. In this proof we let $V={\mathbb{C}}[G]$. It follows from Corollary \[niceKsst\] that $$K_{0}^{G,\,sst}(X,Y)= [\pi_{0}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))^{G}]^{+}.$$ Using Lemma \[lem:pi0\] below together with the same argument as in [@FW:sstfct Proposition 2.10] shows that $\pi_{0}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))^{G}$ consists of equivalence classes of coherent $G$-modules ${\mathcal{M}}$ on $X\times Y$ which are finite and flat over $X$ and admit a surjection of the form ${\mathcal{V}}^{N}_{X\times Y}{\twoheadrightarrow}{\mathcal{M}}$ (the equivalence class of ${\mathcal{M}}$ is independent of the surjection). Here ${\mathcal{M}}_{1}$ and ${\mathcal{M}}_{2}$ are equivalent if there is a smooth, connected curve $C$, two closed points $c_{1},c_{2}\in C$, and a coherent $G$-module ${\mathcal{N}}$ on $X\times X\times Y$ such that ${\mathcal{M}}_{i} ={\mathcal{N}}|_{c_{i}}$ where ${\mathcal{N}}$ is finite and flat over $X\times C$ and it admits a surjection of the form ${\mathcal{V}}_{X\times C\times Y}^{K}{\twoheadrightarrow}{\mathcal{N}}$. The monoid structure on $\pi_{0}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))^{G}$ induced by the $H$-space structure is given by direct sum of modules. We thus have a natural map $$K_{0}^{G,\,sst}(X,Y) = [\pi_{0}\operatorname{Hom}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(\infty))^{G}]^{+} \to K_{0}^{G}(X,Y)/\sim_{alg}.$$ The argument given in [@FW:sstfct Proposition 2.12] applies here to show that this map is an isomorphism. The claim regarding the functorialities is easily verified. \[lem:pi0\] Let ${\mathcal{F}}$ be a presheaf of sets on ${\mathrm{Sch}}/{\mathbb{C}}$. Then $\pi_{0}{\mathcal{F}}(\Delta^{\bullet}_{top}) = {\mathcal{F}}({\mathbb{C}})/\sim$ where $\sim$ is the equivalence relation generated by $x_{1}\sim x_{2}$ if there is a smooth connected curve $C$, an element $z\in {\mathcal{F}}(C)$, two closed points $c_{0}, c_{1}\in C$ such that $x_{k}= \epsilon_{k}^{*}z$ where $\epsilon_{k}:c_{i}\to C$ is the inclusion. Write $\iota_{0},\iota_{1}:\Delta^{0}\to \Delta^{1}$ for the inclusions at $0$ and $1$. If $x_{1}\sim x_{2}$ then they are in the same path component of $\pi_{0}{\mathcal{F}}(\Delta^{\bullet}_{top})$. Suppose that $x_{1}$ and $x_{2}$ lie in the same path component. This means that there is an $(f:\Delta^{1}_{top}\to T^{an}, y\in {\mathcal{F}}(T))$ such that $\iota_{k}^{*}y= x_{i}$ for some $T$. Since we may assume that $T$ is connected, there is a sequence smooth connected curves $C_{1},\ldots C_{r}$, maps $g_{i}:C_{i}\to T$, and points $c_{i}, d_{i}\in C_{i}$ such that $g_{i}(d_{i}) = g_{i+1}(c_{i+1})$ with $g_{1}(c_{1})=f(0)$ and $g_{r}(d_{r})=f(1)$, from which the lemma follows. Topological $K$-theory {#sec:top} ---------------------- We now introduce the model for bivariant equivariant $K$-theory with which we work. We restrict attention to $G$-$CW$-complexes. For a $G$-$CW$ complex $T$ write ${\mathcal{C}}(T)$ for the nonunital topological ring of all continuous complex valued functions on $T$. When $(T,t_{0})$ is a based $G$-$CW$ complex $T$ write ${\mathcal{C}}_{0}(T)$ for the nonunital topological ring of continuous complex valued functions on $T$ which vanish at the base-point. Note when $T$ is unbased, that ${\mathcal{C}}(T) = {\mathcal{C}}_{0}(T_{+})$. Complex conjugation defines a natural involution on ${\mathcal{C}}_{0}(T)$. When $T$ is compact this makes ${\mathcal{C}}_{0}(T)$ into a $C^{*}$-algebra. Additionally the $G$-action on $T$ induces a $G$-action on ${\mathcal{C}}_{0}(T)$, where $G$ acts by continuous ${\mathbb{C}}$-algebra homomorphisms. If $V$ is a unitary complex $G$-representation then $\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})$ is also a $C^{*}$-algebra and $G$ acts on it via $C^{*}$-algebra homomorphisms. For a pointed $G$-$CW$-complex $T$ let $\underline{\operatorname{Hom}}_{*}({\mathcal{C}}_{0}(T),\,\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n}))$ be the space of involution-preserving, continuous, ${\mathbb{C}}$-algebra homomorphisms ($*$-map for short). We write $${\mathcal{F}}_{T}^{V}(n) = \underline{\operatorname{Hom}}_{*}({\mathcal{C}}_{0}(T),\,\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n}))$$ for this space. We have that $\operatorname{Map}(W,\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})) {\cong}{\mathcal{C}}_{0}(W)\otimes\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})$, which together with adjointness gives $$\operatorname{Map}(W,{\mathcal{F}}_{T}^{V}(n)) {\cong}\underline{\operatorname{Hom}_{*}}({\mathcal{C}}_{0}(T),{\mathcal{C}}_{0}(W)\otimes\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})).$$ The space ${\mathcal{F}}_{T}^{V}(n)=\underline{\operatorname{Hom}}_{*}({\mathcal{C}}_{0}(T),\,\operatorname{End}_{{\mathbb{C}}}(V^{\oplus n}))$ has a $G$-action given by the usual formula. A $*$-map $f:{\mathcal{C}}_{0}(T)\to \operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})$ factors as a composition of $*$-maps ${\mathcal{C}}_{0}(T)\to {\mathcal{C}}_{0}(\{t_{1},\ldots, t_{k}\})\to \operatorname{End}_{{\mathbb{C}}}(V^{n}))$ where $\{t_{1},\ldots, t_{k}\}\subseteq T$ is a finite set of points. Thus a point of ${\mathcal{F}}_{T}^{V}(n)$ is identified with a finite (unordered) list of points $t_{1},\ldots, t_{k}$ of $T$ together with a list of pairwise orthogonal subspaces $W_{1},\ldots, W_{k}$ of $V^{\oplus n}$. Then $g\in G$ acts by $g\cdot(t_{1},\ldots, t_{k}; W_{1},\ldots, W_{k}) = (g\cdot t_{1},\ldots g\cdot t_{k};g\cdot W_{1},\ldots, g\cdot W_{k})$. Thus for $H\subseteq G$, an $H$-invariant point of ${\mathcal{F}}_{T}^{V}(n)$ is specified by an $H$-set of points $\{t_{1},\ldots, t_{k}\}$ of $T$ a sub-$H$-space $W\subseteq V^{\oplus n}$ and a vector space decomposition $W = W_{1}\perp \cdots \perp W_{k}$ such that $h\cdot W_{ i} = W_{j}$ for some $j$. Given an injection $\alpha:\underline{m}\hookrightarrow \underline{n}$ write $\tilde{\alpha}:\operatorname{End}_{{\mathbb{C}}}(V^{\oplus m}) \to \operatorname{End}_{{\mathbb{C}}}(V^{\oplus n})$ for the map $$\tilde{\alpha}(f) = \alpha_{*}f\alpha^{*}.$$ Let $A$, $B$ be based $G$-$CW$ complexes. Using the maps $\tilde{\alpha}$, the assignment $$\underline{n}\mapsto \operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},{\mathcal{F}}_{B}^{V}(n)),$$ defines a functor $I\to G{\mathrm{sSet}}$. Now define for based $G$-$CW$-complexes $(A,a_{0})$ and $(B,b_{0})$ $${\mathcal{A}}^{top}_{G}(A, B)^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},{\mathcal{F}}_{B}^{V}(-)).$$ Note that we omit the base point from the notation, leaving it implicit. For each $V$ we now have $\Gamma$-spaces $$\underline{n}_{+}\mapsto |{\mathcal{A}}_{G}^{top}(A,\underline{n}_{+}\wedge B)^{V}|.$$ We will work mostly with unpointed spaces when comparing with the algebraic theories. In this context, if $A$ is unpointed it is convenient to still write $|{\mathcal{A}}_{G}^{top}(A, B)^{V}|$, which is to be interpreted as $|{\mathcal{A}}_{G}^{top}(A_{+}, B)^{V}|$, and similarly for $B$ unpointed. This abuse of notation should not lead to confusion. Again the main case of interest is $V={\mathbb{C}}[G]$. Let $A$, $B$ be based $G$-$CW$ complexes. Define $${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,\,B)$$ to be the $G$-spectrum associated to the equivariant $\Gamma$-space $|{\mathcal{A}}^{top}(A,-\wedge B)^{{\mathbb{C}}[G]}|$. Write $${\mathbf{bu}}^{G,{\mathfrak{c}}}_{*}(A,B) = \pi^{G}_{*}{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,B)$$ for the homotopy groups of this spectrum. More generally for a subgroup $H\subseteq G$, write ${\mathbf{bu}}^{H,{\mathfrak{c}}}_{*}(A,B) = \pi^{H}_{*}{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,B)$. Let $V$ be a real representation. Combining Corollary \[cor:cc\] and Proposition \[prop:conn\] we have that ${\mathbf{bu}}^{G,{\mathfrak{c}}}_{V}(A_{+},S^{0}) = KU^{-V}_{G}(A_{+})$. As before, it is convenient to introduce a variant of the construction above. We define $$\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(\infty)) = \operatorname*{\mathrm{colim}}_{n}\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(n))$$ where the colimit is over the standard inclusions $\underline{n}\subseteq \underline{n+1}$ given by $i\mapsto i$. As in the algebraic and semi-topological cases we have the following (see the paragraph preceding Proposition \[prop:alginf\] for a reminder on the indexing categories used below). \[prop:topinf\] For based $G$-$CW$-complexes $A$, $B$, the natural maps $$\begin{gathered} \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(-)) \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}} \operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(-)) \\ \leftarrow \operatorname*{\mathrm{holim}}_{M}\operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(\infty))\leftarrow \operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{B}^{V}(\infty)).\end{gathered}$$ are equivariant weak equivalences. This is similar to the proofs of Propositions \[prop:alginf\] and \[prop:sstinf\]. \[prop:topspecial\] For based $G$-$CW$-complexes, the equivariant $\Gamma$-space $$\underline{n}_{+}\mapsto |{\mathcal{A}}^{top}_{G}(A,\underline{n}_{+}\wedge B)^{V}|$$ and $$\underline{n}_{+}\mapsto |\operatorname{Hom}(A\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{\underline{n}_{+}\wedge B}^{ V}(\infty))|$$ are degreewise equivariantly weakly equivalent, and they are both special. Using Proposition \[prop:topinf\] and proceeding once more as in the proof of Proposition \[prop:algspecial\], it suffices to show the second equivariant $\Gamma$-space is special. The argument for this is again an adaptation of [@GW:Kmodels Lemma 2.2] to our present context. For any integer $M>0$, we will define a map $$\eta:{\mathcal{F}}_{B}^{V}(M)^{\times n}\to {\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(nM).$$ Composing $\eta$ and the inclusions $\epsilon:{\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(-) \to {\mathcal{F}}_{B}^{V}(-)^{\times n}$ yields the maps $${\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(M)\to {\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(nM) \,\,\textrm{and} \,\, {\mathcal{F}}_{B}^{V}(M)^{\times n}\to {\mathcal{F}}_{ B}^{V}(nM)^{\times n},$$ and the result follows by showing these are $G\times \Sigma_{n}$-equivariantly homotopic to the standard inclusions. We may identify ${\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(M)$ with the subspace of ${\mathcal{F}}_{B}^{V}(M)^{\times n}$ consisting of those $n$-tuples $(f_{i})$ of $*$-maps which satisfy $f_{i}(a)f_{j}(b) = 0$ for all $a$,$b\in {\mathcal{C}}_{0}(B)$. For an injection $\beta:\underline{p}\hookrightarrow \underline{q}$, recall that we write $\tilde{\beta}:\operatorname{End}_{{\mathbb{C}}}(V^{\oplus p})\to\operatorname{End}_{{\mathbb{C}}}(V^{\oplus q})$ for the induced map $\psi\mapsto \beta_{*}\psi\beta^{*}$. Let $\alpha_{i}:\underline{M} \hookrightarrow \underline{nM}$ be the injection $k\mapsto (i-1)M+k$. Define $\eta$ by sending the tuple $(f_{i})$ of $*$-maps to the tuple $(\tilde{\alpha_{i}}f_{i})$ which lies in ${\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(nM)$ as desired. Define the $G\times \Sigma_{n}$-equivariant homotopy ${\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(M)\times I\to {\mathcal{F}}_{\underline{n}_{+}\wedge B}^{V}(nM)$ between $\eta\epsilon$ and the standard inclusion by $((f_{i}), t) \mapsto (t\cdot\tilde{\alpha_{1}}f_{i} + (1-t)\cdot\tilde{\alpha_{i}}f_{i})$. Similarly one sees that $\epsilon\eta$ is equivariantly homotopic to the standard inclusion. As Walker observes in the nonequivariant case, if $B$ is not connected then for $A$ pointed and connected this is not a reasonable spectrum to work with. For example if $B = S^{0}$, $A$ is pointed and connected, and $V={\mathbb{C}}[G]$ then for any subgroup $K\subseteq G$ the space $|{\mathcal{A}}_{G}^{top}(A,\underline{n}_{+})^{V}|^{K}$ is equivalent to the space of $n$-tuples of $K$-bundles on $A$, each of which has rank zero at the base-point. Thus the equivariant $\Gamma$-space $|{\mathcal{A}}_{G}^{top}(A,\underline{n}_{+}\wedge S^{0})^{V}|$ is contractible. As in [@Walker:Thomason] we can remedy this by replacing $B$ with its suspensions. To do this we view $|{\mathcal{A}}_{G}^{top}(A,-\wedge B)^{V}|$ as a ${\mathcal{W}}_{G}$-space and consider the associated spectrum. Let $A$, $B$ be based $G$-$CW$ complexes. Define $${\mathbf{bu}}_{G}(A,\,B)$$ to be the $G$-spectrum associated to the ${\mathcal{W}}_{G}$-space $Y\mapsto |{\mathcal{A}}^{top}(A,Y\wedge B)^{{\mathbb{C}}[G]}|$. Write $${\mathbf{bu}}^{G}_{*}(A,B) = \pi^{G}_{*}{\mathbf{bu}}_{G}(A,B).$$ for the homotopy groups of this spectrum. More generally for a subgroup $H\subseteq G$, write ${\mathbf{bu}}^{H}_{*}(A,B) = \pi^{H}_{*}{\mathbf{bu}}_{G}(A,B)$ Recall from Section \[sec:pre\] that to an equivariant $\Gamma$-space $Y$ we naturally associate a ${\mathcal{W}}_{G}$-space $\widehat{Y}$, which may be described by the formula $$\widehat{Y}(X) = \operatorname*{\mathrm{hocolim}}_{S\to X}PY(S),$$ where $PY$ is the $\Gamma_{G}$-space associated to $Y$, and $S\to X$ is an object of the overcategory $(\Gamma_{G}\downarrow X)$. If $X$ is a ${\mathcal{W}}_{G}$-space and $Y$ is the equivariant $\Gamma$-space obtained from $X$ by restriction then we have a natural map $\widehat{Y}\to X$. We thus have a natural map of spectra $\mathbb{S}X \to \mathbb{U}X$. In particular we have a natural map of spectra $$\label{eqn:c2nc} {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,\,B) \to {\mathbf{bu}}_{G}(A,\,B).$$ \[prop:ev\] Let $B$ be a based $G$-$CW$-complex. For a based $G$-$CW$-complex $X$ there is a natural equivariant equivalence $$\operatorname*{\mathrm{hocolim}}_{S\to X}|{\mathcal{A}}_{G}^{top}(S^{0},S\wedge B)^{V}|\xrightarrow{{\simeq}}|{\mathcal{A}}_{G}^{top}(S^{0},X\wedge B)^{V}|.$$ In particular, ${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},\,B)$ and the spectrum $\{|{\mathcal{A}}_{G}^{top}(S^{0},S^{W}\wedge B)^{{\mathbb{C}}[G]}|\}$ are equivariantly equivalent. Since the indexing category $(\Gamma_{G}\downarrow X)$ is filtered, we have that $$\operatorname*{\mathrm{hocolim}}_{S\to X}|{\mathcal{A}}_{G}^{top}(S^{0},-\wedge B)^{V}| \xrightarrow{{\simeq}} \operatorname*{\mathrm{colim}}_{S\to X}|{\mathcal{A}}_{G}^{top}(S^{0},S\wedge B)^{V}|$$ is an equivariant weak equivalence. It thus suffices to show that $$\operatorname*{\mathrm{colim}}_{S\to X}|{\mathcal{A}}_{G}^{top}(S^{0},S\wedge B)^{V}| \xrightarrow{{\cong}} |{\mathcal{A}}_{G}^{top}(S^{0},X\wedge B)^{V}|$$ is an isomorphism. For this, it suffices to show that $\operatorname*{\mathrm{colim}}_{S\to X}{\mathcal{F}}_{S\wedge B}^{V}(n) \xrightarrow{{\cong}} {\mathcal{F}}^{V}_{X\wedge B}(n)$ is an isomorphism for any $n$. This follows by an argument similar to the argument given in [@Walker:Thomason Theorem 3.14], which we briefly sketch. That this map is bijective follows from the fact that for a based $G$-$CW$-complex $W$, a $*$-map ${\mathcal{C}}_{0}(W)\to \operatorname{End}_{{\mathbb{C}}}(V^{n})$ factors as ${\mathcal{C}}_{0}(W)\to {\mathcal{C}}_{0}(S) \hookrightarrow \operatorname{End}_{{\mathbb{C}}}(V^{n})$ where $S\subseteq W$ is a based finite $G$-set. The map $\operatorname*{\mathrm{colim}}_{S\to X}{\mathcal{F}}_{S\wedge B}^{V}(n) \to {\mathcal{F}}^{V}_{X\wedge B}(n)$ is the colimit of maps $\operatorname*{\mathrm{colim}}_{S\to W}{\mathcal{F}}^{V}_{S\wedge T}(n)\to {\mathcal{F}}^{V}_{W\wedge T}(n)$ where $T\subseteq B$ and $W\subseteq X$ range over all compact subspaces. When $T$, $W$ are compact we have that $\operatorname*{\mathrm{colim}}_{S\to W}{\mathcal{F}}^{V}_{S\wedge T}(n)\to {\mathcal{F}}^{V}_{W\wedge T}(n)$ is closed since ${\mathcal{F}}^{V}_{C}(n)$ is compact whenever $C$ is compact. It is therefore an equivariant homeomorphism and we are done. We have natural equivariant maps $$\label{eqn:adjmaps} |{\mathcal{A}}_{G}^{top}(A,\,B)^{V}|\xrightarrow{}\operatorname{Map}(A,\operatorname*{\mathrm{hocolim}}_{I}{\mathcal{F}}_{B}^{V}(-))\xleftarrow{}\operatorname{Map}(A,\,|{\mathcal{A}}_{G}^{top}(S^{0},\,B)^{V}|)$$ obtained as the composite $$\begin{aligned} |{\mathcal{A}}_{G}^{top}(A,B)^{V}| & = |\operatorname*{\mathrm{hocolim}}_{I}\operatorname{\mathrm{Sing}}_{\bullet}\operatorname{Map}(A,{\mathcal{F}}_{B}^{V}(-))| \\ & \xrightarrow{{\cong}} \operatorname*{\mathrm{hocolim}}_{I}|\operatorname{\mathrm{Sing}}_{\bullet}\operatorname{Map}(A,{\mathcal{F}}_{B}^{V}(-))| \to \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Map}(A,{\mathcal{F}}_{B}^{V}(-)) \\ & \to \operatorname{Map}(A,\operatorname*{\mathrm{hocolim}}_{I}{\mathcal{F}}_{B}^{V}(-)) \xleftarrow{{\simeq}} \operatorname{Map}(A,\operatorname*{\mathrm{hocolim}}_{I}|\operatorname{\mathrm{Sing}}_{\bullet}{\mathcal{F}}_{B}^{V}(-)|) \\ & \xleftarrow{{\cong}}\operatorname{Map}(A,\,|{\mathcal{A}}_{G}^{top}(S^{0},\,B)^{V}|),\end{aligned}$$ where the two displayed isomorphisms follow from [@Hirschhorn Theorem 18.9.10]. \[prop:adjmaps\] Let $A$, $B$ be based $G$-$CW$ complexes, with $A$ compact. Then the maps (\[eqn:adjmaps\]) are equivariant weak equivalences. The right-hand map in (\[eqn:adjmaps\]) is always a $G$-equivalence. Proposition \[prop:topinf\] together with the equivariant homeomorphism $$\operatorname*{\mathrm{colim}}_{n}\operatorname{Map}(A,{\mathcal{F}}^{V}_{B}(n)) \xrightarrow{{\simeq}} \operatorname{Map}(A,\operatorname*{\mathrm{colim}}_{n}{\mathcal{F}}^{V}_{B}(n)).$$ imply that the left-hand map is an equivariant weak equivalence. \[cor:3.17\] Let $A$, $B$ be based $G$-$CW$ complexes, with $A$ compact. Then the maps (\[eqn:adjmaps\]) induce an equivariant equivalence of $G$-spectra $${\mathbf{bu}}_{G}(A,\, B){\simeq}\operatorname{Map}(A,{\mathbf{bu}}_{G}^{{\mathfrak{c}}}(S^{0},B)) {\simeq}\operatorname{Map}(A,B\wedge {\mathbf{bu}}_{G}),$$ where ${\mathbf{bu}}_{G} = {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},S^{0})$. By the previous proposition the maps in (\[eqn:adjmaps\]) are equivariant weak equivalences. It follows that ${\mathbf{bu}}_{G}(A,\,B)$ is equivariantly weakly equivalent to the spectrum $\{\operatorname{Map}(A, |{\mathcal{A}}_{G}^{top}(S^{0},\,S^{W}\wedge B)^{{\mathbb{C}}[G]}|)\}$. By Proposition \[prop:ev\], we have that ${\mathbf{bu}}_{G}^{{\mathfrak{c}}}(S^{0},B)$ is equivariantly weakly equivalent to the spectrum $\{|{\mathcal{A}}_{G}^{top}(S^{0},\,S^{W}\wedge B)^{{\mathbb{C}}[G]}|\}$ and by [@AB Proposition 3.6] this spectrum is equivariantly weakly equivalent to $B\wedge {\mathbf{bu}}_{G}$. \[onemore\] It follows from Proposition \[prop:conn\] that there is a map of $G$-spectra ${\mathbf{bu}}_{G}\to KU_{G}$, where $KU_{G}$ is the spectrum representing equivariant complex $K$-theory. Furthermore ${\mathbf{bu}}_{G}$ is its connective cover, in the sense that this map induces isomorphisms $\pi^{H}_{n}{\mathbf{bu}}_{G} \xrightarrow{{\cong}} \pi_{n}^{H}KU_{G}$ for all subgroups $H\subseteq G$ and $n\geq 0$ and $\pi^{H}_{n}{\mathbf{bu}}_{G} = 0$ for $n<0$. Group completions via mapping telescopes {#sec:tel} ---------------------------------------- We finish this section by giving an alternate description of the equivariant homotopy group completion of the equivariant homotopy monoid $|{\mathcal{A}}_{G}^{top}(S^{0},\,S^{0})^{V}|$. This is the equivariant analogue of only a part of the results in [@Walker:Thomason Section 4]. It seems somewhat more complicated to establish the analogous result for the monoids $|{\mathcal{A}}_{G}^{top}(S^{0},\,B)^{V}|$ when $B$ has nontrivial action and we do not need it. Similarly we do not need the semi-topological analogues. An important consequence of the description of the equivariant homotopy group completion as a mapping telescope appears below in Corollary \[cor:cc\]. This is used in Section \[sec:pair\] in order to define a natural map of rings $${\mathbf{bu}}^{G}_{*}(W,S^{0}) \to KU_{G}^{-*}(W).$$ This natural transformation is crucial for our main results in Sections \[sec:sstThom\] and \[sec:algThom\]. Recall that ${\mathcal{F}}^{V}_{S^{0}}(n)$ is isomorphic to the space of linear subspaces in $V^{n}$. Below, we write ${\mathcal{F}}_{B}^{V}(\infty)=\operatorname*{\mathrm{colim}}_{n}{\mathcal{F}}_{B}^{V}(n)$ where the colimit is over the standard inclusions $\underline{n}\subseteq \underline{n+1}$ given by $i\mapsto i$. Consider the equivariant maps $\eta:{\mathcal{F}}^{V}_{S^{0}}(n)\to {\mathcal{F}}^{V}_{S^{0}}(n+1)$ given by $(W\subseteq V^{n}) \mapsto (V\oplus W \subseteq V\oplus V^{n})$. Taking colimits defines an equivariant map $\eta:{\mathcal{F}}^{V}_{S^{0}}(\infty)\to {\mathcal{F}}^{V}_{S^{0}}(\infty)$. Write $X_{i} = {\mathcal{F}}^{V}_{S^{0}}(\infty)$ considered as a based $G$-space with basepoint $x_{i}$, where $x_{0} = 0$ and $x_{i}= \eta^{i}x_{0}$. Then $\eta:X_{i}\to X_{i+1}$ is a based map, Write $$Tel(X_{i},\eta) = \operatorname*{\mathrm{hocolim}}(X_{0}\xrightarrow{\eta} X_{1}\xrightarrow{\eta} X_{2}\xrightarrow{\eta}\cdots )$$ \[thm:tel\] The map $${\mathcal{F}}^{V}_{S^{0}}(\infty) \to Tel(X_{i},\eta)$$ is an equivariant homotopy group completion. The proof is modeled on that of [@FW:sstfct Proposition 3.3], where $H$-space structures arising from operad actions rather than through $\Gamma$-spaces are used. We need to show that ${\mathcal{F}}^{V}_{S^{0}}(\infty) \to Tel(X_{i},\eta)$ is a map of equivariant homotopy commutative, associative $H$-spaces such that for any subgroup $K\subseteq G$ the map ${\mathcal{F}}^{V}_{S^{0}}(\infty)^{K} \to Tel(X_{i},\eta)^{K}$ is a homotopy group completion. Recall that this means that for each subgroup $K\subseteq G$, 1. the map $\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K}) \to \pi_{0}(Tel(X_{i},\eta)^{K})$ is a group completion of the monoid $\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K})$, and 2. $H_{*}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K};A) \to H_{*}(Tel(X_{i},\eta)^{K};A)$ is localization of the action of ${\mathbb{Z}}[\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K})]$ for any commutative ring $A$. First we have to show that $Tel(X_{i},\eta)$ has the structure of an equivariant homotopy commutative and associative $H$-space. We have equivariant homotopy equivalences $Tel(X_{i}\times X_{i}, \eta\times\eta) {\simeq}Tel(X_{i},\eta)\times Tel(X_{i},\eta)$ and $Tel(X_{i},\eta){\simeq}Tel(X_{2i}, \eta^{2})$. Thus to define the pairing it suffices to give a map $\mu:X_{n}\times X_{n} \to X_{2n}$ such that $\eta^{2}$ and $\mu\circ(\eta\times\eta)$ are equivariantly homotopic. We take $\mu$ to be the $H$-space product map $\mu:{\mathcal{F}}^{V}_{S^{0}}(\infty)\times {\mathcal{F}}^{V}_{S^{0}}(\infty)\to {\mathcal{F}}^{V}_{S^{0}}(\infty)$. Recall that the $H$-space structure on ${\mathcal{F}}^{V}_{S^{0}}(\infty)$ arises as follows. The equivariant $\Gamma$-space $\underline{n}_{+}\mapsto {\mathcal{F}}^{V}_{\underline{n}_{+}}(\infty)$ is special and $\mu$ arises by choosing a homotopy inverse to ${\mathcal{F}}_{\underline{2}_{+}}^{V}(\infty)\subseteq {\mathcal{F}}^{V}_{S^{0}}(\infty)\times {\mathcal{F}}^{V}_{S^{0}}(\infty)$ together with the multiplication map induced by $\underline{2}_{+}\to \underline{1}_{+}$ given by sending both $1$ and $2$ to $1$. Consider the commutative square $$\xymatrix{ {\mathcal{F}}^{V}_{S^{0}}(n)\times{\mathcal{F}}^{V}_{S^{0}}(n) \ar[r]^-{\eta\times \eta} & {\mathcal{F}}^{V}_{S^{0}}(n+1)\times{\mathcal{F}}^{V}_{S^{0}}(n+1) \ar[r]^-{\epsilon\gamma\times\gamma} & {\mathcal{F}}^{V}_{S^{0}}(n+2)\times{\mathcal{F}}^{V}_{S^{0}}(n+2) \\ {\mathcal{F}}^{V}_{\underline{2}_{+}}(n) \ar[r]^{\beta}\ar[d]^{m}\ar[r]\ar[u]^{\subseteq}_{{\simeq}} & {\mathcal{F}}^{V}_{\underline{2}_{+}}(n+2) \ar[d]^{m}\ar[ur]^{\subseteq}_{{\simeq}} & \\ {\mathcal{F}}^{V}_{S^{0}}(n) \ar[r]^{\eta^{2}} & {\mathcal{F}}^{V}_{S^{0}}(n+2) , & }$$ where $\beta$ sends the pair $(W_{1}, W_{2})$ to $(V\oplus 0 \oplus W_{1}, 0\oplus V \oplus W_{2})$ The map $\gamma$ is induced by $\gamma:\underline{k}_{+}\to \underline{k+1}_{+})$ where $\gamma(i) = i+1$ and $\epsilon$ is induced by $\epsilon:\underline{k}_{+}\to \underline{k}_{+}$ which interchanges $1$ and $2$ and is the identity on the other elements. As shown in the proof of Proposition \[prop:sstinf\], injections $\mathbb{N}\to \mathbb{N}$ induce equivariant homotopy equivalences ${\mathcal{F}}^{V}(\infty) \to {\mathcal{F}}^{V}(\infty)$. In particular $\gamma$,$\epsilon:{\mathcal{F}}^{V}(\infty) \to {\mathcal{F}}^{V}(\infty)$ are both equivariant homotopy equivalences. We thus have that $\eta^{2} {\simeq}\mu\circ(\eta\times \eta)$ and thus we obtain the required pairing giving a multiplication on $Tel(X_{i},\eta)$. Now we need to show that $\mu$ gives $Tel(X_{i},\eta)$ the structure of a homotopy commutative and homotopy associative $H$-space. First we show that the basepoint $x\in Tel(X_{i},\eta)$ is a right identity up to homotopy. The maps $\eta^{n}:X_{n}\to X_{2n}$ induce the homotopy equivalence $Tel(X_{n},\eta){\simeq}Tel(X_{2n},\eta^{2})$ and so to show that $x$ is a right homotopy identity it suffices to show that the maps $\alpha^{n}$, $\mu(-,x_{n}):X_{n}\to X_{2n}$ are homotopic. To show homotopy commutativity it suffices to show that the two maps $X_{i}\times X_{i}\to X_{2i}$ given by $\mu$ and $\mu\tau$ are homotopic, where $\tau$ is the map switching the factors. This follows from the fact that the maps $m$,$m\tau:{\mathcal{F}}^{V}_{\underline{2}_{+}}(n)\to {\mathcal{F}}^{V}_{S^{0}}(n)$ are equal, where $\tau$ is the map interchanging $1$ and $2$. Homotopy associativity follows in a similar fashion. The map ${\mathcal{F}}^{V}_{S^{0}}(\infty) \to Tel(X_{i},\eta)$ is an $H$-space map. The monoid $\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K})$ is the monoid (under direct sum) of isomorphism classes of $K$-modules which embed in $V^{\infty}$. Observe that $\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K})^{+}$ is obtained obtained by inverting the class of $V$ in $\pi_{0}({\mathcal{F}}^{V}_{S^{0}}(\infty)^{K})$. But this is exactly $\pi_{0}(Tel(X_{i},\eta)^{K})$. The condition on homology follows immediately since $H_{*}(Tel(X_{i},\eta)^{K};A)= \operatorname*{\mathrm{colim}}_{i}H_{*}(X_{i};A)$. The following corollaries are used later in Proposition \[prop:conn\] to define natural transformations to $KU_{G}^{*}(-)$. \[cor:gc\] Let $W$ be a compact, unbased $G$-$CW$ complex. Write ${\mathcal{A}}(-)$ for the equivariant $\Gamma$-space $\underline{n}_{+}\mapsto |{\mathcal{A}}_{G}^{top}(S^{0},\,\underline{n}_{+}\wedge S^{0})^{V}|$. The natural map $$\big[W_{+}, |{\mathcal{A}}_{G}^{top}(S^{0},\,S^{0})^{V}|\big]_{G}^{+} \to \big[W_{+}, \Omega \widehat{{\mathcal{A}}}(S^{1})\big]_{G}$$ is an isomorphism. The equivariant $\Gamma$-space ${\mathcal{A}}$ is special by Proposition \[prop:topspecial\]. By Lemma \[lem:Gspc\] the map $|{\mathcal{A}}_{G}^{top}(S^{0},\,S^{0})^{V}|\to \Omega \widehat{{\mathcal{A}}}(S^{1})$ is an equivariant homotopy group completion. By Proposition \[prop:topinf\] we have an equivariant equivalence $|{\mathcal{A}}_{G}^{top}(S^{0},\,S^{0})^{V}|{\simeq}{\mathcal{F}}^{V}_{S^{0}}(\infty)$ and so the previous theorem implies that we have an equivariant weak equivalence $\Omega\widehat{{\mathcal{A}}}(S^{1}){\simeq}Tel(X_{i},\eta)$. Therefore we have natural isomorphisms $\operatorname*{\mathrm{colim}}_{i}[W_{+}, X_{i}]_{G}= [W_{+},\Omega\widehat{{\mathcal{A}}}(S^{1})]_{G}$. As unbased spaces we have that $X_{i} = X_{0}$. Thus we have the natural identification $[W_{+}, X_{i}]_{G}{\cong}[W_{+}, X_{0}]_{G}$ and under this identification the transition maps in $\operatorname*{\mathrm{colim}}_{i}[W_{+}, X_{i}]_{G} = \operatorname*{\mathrm{colim}}_{i}[W_{+}, X_{0}]_{G}$ are addition by the same element and so $\operatorname*{\mathrm{colim}}_{i}[W_{+}, X_{i}]_{G} =[W_{+}, \Omega {\mathcal{A}}(S^{1})]_{G}$ is obtained by inverting this class in $[W_{+}, X_{0}]_{G}$. But since $[W_{+}, \Omega \widehat{{\mathcal{A}}}(S^{1})]_{G}$ is a group, this must be the group completion of $[W_{+}, X_{0}]_{G}$. \[cor:cc\] Let $X$ be a compact, unbased $G$-$CW$ complex. The natural map $${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(X_{+}, S^{0}) \to {\mathbf{bu}}_{G}(X_{+},S^{0})$$ induces an equivalence $$\Omega^{\infty}{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(X_{+}, S^{0}) \to \Omega^{\infty}{\mathbf{bu}}_{G}(X_{+},S^{0})$$ of associated equivariant infinite loop spaces. In particular, $$\pi^{G}_{0}{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(X_{+}, S^{0}) = KU_{G}^{0}(X) = \pi^{G}_{0}{\mathbf{bu}}_{G}(X_{+},S^{0}).$$ By Proposition \[prop:adjmaps\] we have that the equivariant $\Gamma$-space (resp.  ${\mathcal{W}}_{G}$-space) $|{\mathcal{A}}^{top}_{G}(X_{+}, -\wedge S^{0})^{{\mathbb{C}}[G]}|$ is equivariantly weakly equivalence to the equivariant $\Gamma$-space (resp.  ${\mathcal{W}}_{G}$-space) $\operatorname{Map}(X_{+}, |{\mathcal{A}}^{top}_{G}(S^{0}, -\wedge S^{0})^{{\mathbb{C}}[G]}|)$. For convenience write ${\mathcal{M}}(-) = \operatorname{Map}(X_{+}, |{\mathcal{A}}^{top}_{G}(S^{0}, -\wedge S^{0})^{{\mathbb{C}}[G]}|)$ and ${\mathcal{A}}(-) = |{\mathcal{A}}^{top}_{G}(S^{0}, -)^{{\mathbb{C}}[G]}|$. We show that $$\Omega^{W}\widehat{{\mathcal{M}}}(S^{W}) \to \Omega^{W}\operatorname{Map}(X_{+}, {\mathcal{A}}(S^{W}))$$ is an equivariant equivalence for any representation $W$ with $W^{G}\neq 0$. It follows from Proposition \[prop:topspecial\] that the equivariant $\Gamma$-space ${\mathcal{M}}$ is special and so by Lemma \[lem:Gspc\] the associated spectrum $\mathbb{S}{\mathcal{M}}$ is a positive $\Omega$-$G$-spectrum and ${\mathcal{M}}(S^{0})\to \Omega\widehat{{\mathcal{M}}}(S^{1})$ is an equivariant homotopy group completion. It therefore suffices to show that $$\operatorname{Map}(X_{+}, {\mathcal{A}}(S^{0})) \to \Omega^{W}\operatorname{Map}(X_{+}, {\mathcal{A}}(S^{W}) = \operatorname{Map}(X_{+}, \Omega^{W}{\mathcal{A}}(S^{W}))$$ is an equivariant homotopy group completion whenever $W^{G}\neq 0$. Using Proposition \[prop:ev\] and that ${\mathcal{A}}(-)$ is special, we have an equivariant weak equivalence $\Omega^{W}{\mathcal{A}}(S^{W}){\simeq}\Omega^{1}{\mathcal{A}}(S^{1}) {\simeq}\Omega^{1}\widehat{A}(S^{1})$. It follows from Corollary \[cor:gc\] that $$\pi_{0}\operatorname{Map}(X_{+}, {\mathcal{A}}(S^{0}))^{K}\to \pi_{0}\operatorname{Map}(X_{+}, \Omega^{W}{\mathcal{A}}(S^{W}))^{K}$$ is a group completion for any subgroup $K\subseteq G$. Furthermore, from Theorem \[thm:tel\] it follows that we have an equivariant weak homotopy equivalence $$\operatorname{Map}(X_{+}, \Omega^{W}{\mathcal{A}}(S^{W})) {\simeq}Tel(\operatorname{Map}(X_{+}, {\mathcal{A}}(S^{0})),\eta_{*})$$ from which the condition on homology with coefficients follows. For the last statement, we have that $$\pi^{G}_{0}{\mathbf{bu}}_{G}(X_{+},S^{0}) = \big[X_{+},\Omega{\mathcal{A}}(S^{1})\big]_{G} = \big[X_{+}, {\mathcal{F}}^{{\mathbb{C}}[G]}_{S^{0}}(\infty)\big]_{G}^{+}.$$ Since ${\mathcal{F}}^{{\mathbb{C}}[G]}_{S^{0}}$ is the Grassmannian of linear subspaces inside of ${\mathbb{C}}[G]^{\infty}$ it follows that this group is naturally identified with $KU_{G}^{0}(X)$. Walker’s comparison theorem {#sec:comp} =========================== In this section we define a comparison map between our semi-topological and topological bivariant equivariant $K$-theories. We present a proof of the equivariant version of Walker’s Fundamental Comparison Theorem [@Walker:Thomason Theorem 5.1]. Namely, we have the following result, which is obtained as a special case of Corollary \[cor:qtopeq\] below. \[mainthm\] Let $Y$ be a smooth quasi-projective $G$-variety. Then the natural map (\[eqn:comp\]) induces a weak equivalence of $G$-spectra, $${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},Y^{an})\xrightarrow{{\simeq}} {\mathcal{K}}^{sst}_{G}({\mathbb{C}},Y).$$ By Corollary \[cor:3.17\] we have an equivariant weak equivalence of $G$-spectra, ${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0}, Y^{an}) {\simeq}Y^{an}_{+}\wedge {\mathbf{bu}}_{G}$, where ${\mathbf{bu}}_{G} = {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},S^{0})$ is the connective cover of $KU_{G}$ (see Remark \[onemore\]). Thus this theorem says that semi-topological equivariant $K$-homology agrees with ${\mathbf{bu}}_{G}$-homology. The strategy of proof follows the nonequivariant one in [@Walker:Thomason]. The key is to show that we have a natural equivariant homotopy equivalence ${\mathcal{F}}_{Y^{an}}\to ({\mathcal{G}}_{Y})^{an}$. A map $f:X\to Y$ of varieties defines also a continuous map $f:X^{an}\to Y^{an}$ of associated analytic spaces. This gives rise to a natural map of $G$-simplicial sets $$\label{eqn:qtop} \operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(n)) \to \operatorname{Hom}_{cts}(X^{an}\times{\Delta^{\bullet}_{top}},\,{\mathcal{G}}_{Y}^{V}(n)^{an}).$$ To aid the comparison between the topological and semi-topological $K$-theories we introduce a bivariant theory based on the spaces ${\mathcal{G}}^{V}_{Y}(n)^{an}$. Let $A$ be a based $G$-$CW$-complex and $Y$ a quasi-projective $G$-variety and define $${\mathcal{A}}_{G}^{qtop}(A,Y)^{V} = \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},{\mathcal{G}}_{Y}^{V}(-)^{an}).$$ \[defKqtop\] We define ${\mathcal{K}}^{qtop}_{G}(A,Y)$ to be the $G$-spectrum associated to the equivariant $\Gamma$-space $$\underline{n}_{+}\mapsto |{\mathcal{A}}_{G}^{qtop}(A, \underline{n}_{+}\wedge Y_{+})^{{\mathbb{C}}[G]}|.$$ The map (\[eqn:qtop\]) above induces a natural transformation ${\mathcal{K}}^{sst}_{G} \to {\mathcal{K}}^{qtop}_{G}$. Write $$\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(\infty)^{an})=\operatorname*{\mathrm{colim}}_{n}\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(n)^{an})$$ where the colimit is over the standard inclusions $\underline{n}\subseteq \underline{n+1}$ given by $i\mapsto i$. As before we have the following (see the paragraph preceding Proposition \[prop:alginf\] for a reminder on the indexing categories used below). \[prop:qtopinf\] For based $G$-$CW$-complexes $A$ and a quasi-projective $G$-variety $Y$, the natural maps $$\begin{gathered} \operatorname*{\mathrm{hocolim}}_{I}\operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(-)) \to \operatorname*{\mathrm{hocolim}}_{\tilde{I}} \operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(-)) \\ \leftarrow \operatorname*{\mathrm{hocolim}}_{M}\operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(\infty))\leftarrow \operatorname{Hom}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(\infty)).\end{gathered}$$ are equivariant weak equivalences. Let $V$ be a unitary complex representation. By [@Walker:Thomason Section 5] there is a natural map of spaces $$\label{eqn:comp} {\mathcal{F}}_{Y^{an}}^{V}(n) \to {\mathcal{G}}^{V}_{Y}(n)^{an},$$ which will be shown in Theorem \[thm:ew\] to be an equivariant weak equivalence. This map can be described as follows. A point of ${\mathcal{F}}_{Y^{an}}^{V}(n)$ is specified by a list of distinct points $y_{1},\ldots, y_{k}$ of $Y^{an}$ and a list of pairwise orthogonal subspaces $W_{1},\ldots, W_{k}$ of $V^{n}$. To this point we associate the surjection $${\mathcal{V}}^{n}_{Y} \xrightarrow{(eval_{y_{1},\ldots,y_{k}})} \bigoplus_{i=1}^{k}V^{n} \longrightarrow \bigoplus_{i=1}^{k}W_{i}^{*}.$$ By [@Walker:Thomason Lemma 5.3] this defines a map ${\mathcal{F}}^{V}_{Y^{an}}(n)\subseteq {\mathcal{G}}^{V}_{Y}(n)^{an}$ which is a subspace inclusion and the image of ${\mathcal{F}}_{Y^{an}}^{V}(n)$ can be characterized as follows. When $Y=\operatorname{\mathrm{Spec}}(R)$ is affine, a point of ${\mathcal{G}}^{V}_{Y}(n)$ is represented by a pair $(V^{n}\to P, \rho:R\to \operatorname{End}_{{\mathbb{C}}}(P))$ where $P$ is a quotient vector space and $\rho$ is a map of ${\mathbb{C}}$-algebras. Two pairs $(V^{n}\to P, \rho:R\to \operatorname{End}_{{\mathbb{C}}}(P))$ and $(V^{n}\to Q, \rho:R\to \operatorname{End}_{{\mathbb{C}}}(Q))$ represent the same point when there is a vector-space isomorphism $P{\cong}Q$ which make the evident triangles commute. A pair $( V^{n}\to P, \rho:R\to \operatorname{End}_{{\mathbb{C}}}(P))$ is in the image of ${\mathcal{F}}^{V}_{Y^{an}}(n)$ exactly when $\rho$ is a normal map of ${\mathbb{C}}^{*}$-algebras (i.e. $\rho(r)$ and $\rho(r)^{*}$ commute with each other for all $r\in R$). The subspace ${\mathcal{F}}^{V}_{Y^{an}}(n)\subseteq {\mathcal{G}}^{V}_{Y}(n)^{an}$ is $G$-invariant and is compatible with the maps in $I$. We therefore have maps of equivariant $\Gamma$-spaces $${\mathcal{A}}^{sst}(X,-\wedge Y_{+})^{V} \to {\mathcal{A}}^{qtop}(X^{an},-\wedge Y_{+})^{V}\leftarrow{\mathcal{A}}^{top}(X^{an}, -\wedge Y^{an}_{+})^{V}$$ and thus natural maps of $G$-spectra $$\label{eqn:compbiv} {\mathcal{K}}^{sst}_{G}(X, Y)\to {\mathcal{K}}^{qtop}_{G}(X^{an}, Y) \xleftarrow{{\simeq}} {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(X^{an},Y^{an}).$$ We will see in Corollary \[cor:qtopeq\] that the right-hand map is a weak equivalence (note that the left map is an equivalence for $X=\operatorname{\mathrm{Spec}}({\mathbb{C}})$). As shown in [@Walker:Thomason Section 5], sending a quotient ${\mathcal{V}}^{n}_Y\to {\mathcal{M}}$ to its support defines a map of varieties $\theta:{\mathcal{G}}^{V}_{Y}(n)\to \coprod_{k=0}^{n|V|} {\mathrm{Sym}}^{k}Y$ and this is equivariant. Write $\theta_{top}$ for the restriction of $\theta$ to ${\mathcal{F}}_{Y}^{V}(n)$. The following lemma shows that in order to establish that (\[eqn:comp\]) is an equivariant weak equivalence, it suffices to show that for each subgroup $H\subseteq G$, the map ${\mathcal{F}}_{Y}^{V}(n)^{H} \to ({\mathcal{G}}^{V}_{Y}(n)^{an})^{H}$ is a weak equivalence locally on $(\coprod_{k=0}^{n|V|} {\mathrm{Sym}}^{k}Y^{an})^{H}$, in a suitable sense. This lemma and its proof are a slight modification of [@Walker:Thomason Lemma 5.4] in order to conclude a weak equivalence rather than a homology equivalence. \[lem:triangle\] Let $$\xymatrix{ W\ar[dr]_-{p}\ar[rr]^{f} & & X \ar[dl]^-{q}\\ & Z & }$$ be a commutative triangle of topological spaces with the property that for every $z\in U\subseteq Z$ with $U$ open, there is an open $V$ with $z\in V\subseteq U$ such that the induced map $f|_{p^{-1}(V)}:p^{-1}(V) \to q^{-1}(V)$ is a weak equivalence. Then $f:W\to X$ is a weak equivalence. The hypothesis implies that we can construct a hypercover ${\mathcal{V}}_\bullet\to Z$ with the property that each ${\mathcal{V}}_{n}$ is a disjoint union of open subsets $V\subseteq Z $ with the property that $p^{-1}(V) \to q^{-1}(V)$ is a weak equivalence. Write $({\mathcal{V}}_{n})_{X}$ and $({\mathcal{V}}_{n})_{W}$ for the pullback of ${\mathcal{V}}_{n}$ to $X$ and $W$ respectively. We thus have weak equivalences $({\mathcal{V}}_{n})_{W}\to ({\mathcal{V}}_{n})_{X}$ for each $n$, which induces a weak equivalence upon taking homotopy colimits. The result follows from the commutative diagram $$\xymatrix{ \operatorname*{\mathrm{hocolim}}_{\Delta}({\mathcal{V}}_{n})_{W} \ar[r]^{\simeq}\ar[d]_{\simeq} & \operatorname*{\mathrm{hocolim}}_{\Delta}({\mathcal{V}}_n)_{X} \ar[d]^{\simeq} \\ W \ar[r] & X. }$$ where the vertical maps are weak equivalences by [@DI:tophyp Theorem 1.3]. \[lem:locet\] Let $f:X\to Y$ be an equivariant étale map between quasi-projective complex $G$-varieties and $[q:{\mathcal{V}}^{n}_{X}\to {\mathcal{M}}]\in {\mathcal{G}}_{X}^{V}(n)({\mathbb{C}})$ a point such that $\theta([q])=m_{1}x_{1}+\cdots + m_{v}x_{v}$ (with $x_{i}\neq x_{j}$ for $i\neq j$) has the property that $f(x_{i})\neq f(x_{j})$ for $i\neq j$. Then the equivariant map $f_{*}:{\mathcal{G}}_{X}^{V}(n)\to {\mathcal{G}}_{Y}^{V}(n)$ is étale at each point in $G\smash\cdot[q]$. This follows from [@Walker:Thomason Lemma 5.5], as each $g\smash\cdot[q]$ satisfies the hypothesis of the lemma. For a subspace $D\subseteq Y^{an}$ write $({\mathcal{G}}^{V}_{Y}(n))^{an}_{D}$ (resp.  ${\mathcal{F}}^{V}_{Y}(n)_{D}$) for the inverse image of $\coprod {\mathrm{Sym}}^{k}D \subseteq \coprod {\mathrm{Sym}}^{k}Y^{an}$ under $\theta$ (resp.  under $\theta_{top}$). Observe that this is an open subspace (resp.  closed) when $D$ is open (resp.  closed) and it is invariant when $D$ is invariant. \[lem:lociso\] Suppose that $f:X\to Y$ is an equivariant map of quasi-projective $G$-varieties. Let $D\subseteq X^{an}$ be an invariant analytic open subset such that $f|_{D}$ is one-to-one and $f$ is étale at every point of $D$. Then the equivariant map $f_{*}:{\mathcal{G}}^{V}_{X}(n)\to{\mathcal{G}}^{V}_{Y}(n)$ induces an equivariant homeomorphism $$f_{*}:({\mathcal{G}}_{X}^{V}(n))^{an}_{D} \xrightarrow{{\cong}} ({\mathcal{G}}_{Y}^{V}(n))^{an}_{f(D)}.$$ Moreover this restricts to an equivariant homeomorphism $$f_{*}:({\mathcal{F}}_{X}^{V}(n))^{an}_{D} \xrightarrow{{\cong}} ({\mathcal{F}}_{Y}^{V}(n))^{an}_{f(D)}.$$ These maps are shown to be homeomorphisms in [@Walker:Thomason Lemma 5.7], in our situation they are additionally equivariant. \[thm:ew\] Let $Y$ be a smooth quasi-projective complex $G$-variety. Then ${\mathcal{F}}^{V}_{Y}(n) \to {\mathcal{G}}^{V}_{Y}(n)^{an}$ is an equivariant weak equivalence. We show that for each subgroup $H\subset G$ the triangle $$\label{eqn:tritemp} \xymatrix{ ({\mathcal{F}}_{Y}^{V}(n))^{H} \ar[rr]\ar[rd]_-{\theta_{top}} & & ({\mathcal{G}}_{Y}^{V}(n)^{an})^{H} \ar[ld]^-{\theta} \\ &(\coprod _{k=0}^{n|V|}{\mathrm{Sym}}^{k}Y^{an})^{H} & }$$ satisfies the hypothesis for Lemma \[lem:triangle\]. For a point $y\in Y^{an}$ with stabilizer $H_{y}$ write $[H/H_{y}\smash\cdot y]$ for the sum over the points of the $H$-orbit of $y$. Any point in $(\coprod {\mathrm{Sym}}^{k}Y^{an})^{H}$ can be written in the form $m_{1}[H/H_{1}\smash\cdot y_{1}] + \cdots + m_{r}[H/H_{r}\smash\cdot y_{r}]$ with $H/H_{i}\smash\cdot y_{i}\neq H/H_{j}\cdot y_{j}$ for $i\neq j$. Choose a point $P=m_{1}[H/H_{1}\smash\cdot y_{1}] + \cdots + m_{r}[H/H_{r}\smash\cdot y_{r}]$ in $(\coprod {\mathrm{Sym}}^{k}Y^{an})^{H}$ and an analytic open neighborhood $U$ of this point. As $Y$ is quasi-projective, we can find an invariant affine open subscheme of $Y$ which contains every point of the orbits $H/H_{i}\smash\cdot y_{i}$ and so we may assume that $Y$ is affine. We have equivariant étale maps $$H\times^{H_{i}} T_{y_{i}}Y \xleftarrow{q_{i}} H\times^{H_{i}} Y \xrightarrow{\pi_{i}} Y.$$ (the left-hand map is defined by using an $H_{i}$-equivariant splitting to the projection $d:m\to m/m^{2}$, where $m$ is the maximal ideal of $y_{i}$ in ${\mathcal{O}}_{Y,Y_{i}}$). These maps satisfy - $\pi_{i}$, $q_{i}$ are $H$-equivariant and $q_{i}$ maps $(h, y_{i})$ to $(h,0)$, - $\pi_{i}$, $q_{i}$ are étale at every point of $H/H_{i}\times\{y_{i}\}$, - the images of $\pi_{i}(H/H_{i}\times y_{i})$ and $\pi_{j}(H/H_{j}\times y_{j})$ are disjoint if $i\neq j$, and - the images of $q_{i}(H/H_{i}\times y_{i})$ and $q_{j}(H/H_{j}\times y_{j})$ are disjoint if $i\neq j$. Let $D_{i}\subseteq Y$ be $H_{i}$-invariant open analytic neighborhoods of $y_{i}\in Y$ such that $hD_{i}\cap h'D_{i} = \emptyset$ if $h$, $h'$ belong to different cosets of $H_{i}$, $H\smash\cdot D_{i}\cap H\smash\cdot D_{j}= \emptyset$ if $i\neq j$, and $q_{i}:D_{i}\to q_iD_{i}$ is a homeomorphism with $q_{i}D_{i}\subseteq T_{y_i}Y$ an open polydisk. We may take the $D_{i}$ small enough so that ${\mathrm{Sym}}^d (\cup H\smash\cdot D_{i})$ is contained in $U$ (where $d$ is such that $P\in{\mathrm{Sym}}^{d} Y$). We may further assume that $\pi_i$ and $q_i$ are étale on every $H\times^{H_i}D_{i}$. We show that $({\mathcal{F}}_{Y}^{V}(n))^{H}_{\cup H\smash\cdot D_{i}} \to ({\mathcal{G}}_{Y}^{V}(n)^{an})^{H}_{\cup H\smash\cdot D_{i}}$ is a weak equivalence. This in turn implies that $\theta_{top}^{-1}({\mathrm{Sym}}^{d}(\cup H\smash\cdot D_{i}))\to \theta^{-1}({\mathrm{Sym}}^{d}(\cup H\smash\cdot D_{i}))$ is a weak equivalence, which establishes that (\[eqn:tritemp\]) satisfies the hypothesis of Lemma \[lem:triangle\] as required. Applying Lemma \[lem:lociso\] we have the following commutative diagram of homeomorphisms (where for notational convenience we surpress both $n$ and the superscript $an$ and write $\tilde{D}_{i}= H\times^{H_{y_i}}D_{i}$) $$\xymatrix{ \big({\mathcal{F}}^{V}_{\coprod (H\times^{H_{i}}T_{y_{i}}Y)}\big)_{\coprod q_{i}(\tilde{D}_{i})}^{H} \ar[d] & \big({\mathcal{F}}^{V}_{(\coprod H\times^{H_{i}}Y)}\big)_{\coprod \tilde{D}_{i}}^{H}\ar[l]_-{{\cong}} \ar[d] \ar[r]^-{{\cong}} & \big({\mathcal{F}}^{V}_{Y}\big)_{\cup \pi_{i}(\tilde{D}_{i})}^{H} \ar[d] \\ \big({\mathcal{G}}^{V}_{\coprod (H\times^{H_{i}}T_{y_{i}}Y)}\big)_{\coprod q_{i}(\tilde{D}_{i})}^{H} & \big({\mathcal{G}}^{V}_{\coprod H\times^{H_{i}}Y}\big)_{\coprod \tilde{D}_{i}}^{H}\ar[l]_-{{\cong}} \ar[r]^-{{\cong}} & \big({\mathcal{G}}^{V}_{Y}\big)_{\cup \pi_{i}(\tilde{D}_{i})}^{H} . }$$ It therefore suffices to show that the left-hand vertical map is a weak equivalence. For each $i$ there is an $H_{i}$-equivariant linear map $T_{y_{i}}Y \to T_{y_{i}}Y$ mapping $q_i(D_{i})$ homeomorphically to the open unit polydisk $C_{i}\subseteq T_{y_{i}}Y$. Therefore it is enough to show that $$\big({\mathcal{F}}^{V}_{\coprod H\times^{H_{i}} T_{y_{i}}Y}\big)^{H}_{\coprod H\times^{H_{i}}C_{i}} \to \big({\mathcal{G}}^{V}_{\coprod H\times^{H_{i}} T_{y_{i}}Y}\big)^{H}_{\coprod H\times^{H_{i}}C_{i}}$$ is a weak equivalence. By Lemma \[lem:ehpt\] below, the inclusion $H/H_{i} \subseteq H\times^{H_{i}}T_{y_{i}}Y$, at the point $0\in T_{y_i}Y$, induce weak homotopy equivalences $$\big({\mathcal{G}}^{V}_{\coprod H/H_{i}}\big)^{H} \xrightarrow{{\simeq}} \big({\mathcal{G}}^{V}_{\coprod H\times^{H_{i}} T_{y_{i}}Y}\big)^{H}_{\coprod C_{i}^{h}} \;\;\textrm{and} \;\; \big({\mathcal{F}}^{V}_{\coprod H/H_{i}}\big)^{H} \xrightarrow{{\simeq}} \big({\mathcal{F}}^{V}_{\coprod H\times^{H_{i}} T_{y_{i}}Y}\big)^{H}_{\coprod C_{i}^{h}}.$$ It thus suffices to show that $\big({\mathcal{F}}^{V}_{\coprod H/H_{i}}(n)\big)^{H} \to \big({\mathcal{G}}^{V}_{\coprod H/H_{i}}(n)^{an}\big)^{H}$ is a weak equivalence. (These are equivariant analogues of the spaces denoted ${}^{0}({\mathcal{G}}^{n}_{T})^{an}_{C}$ and ${}^{0}({\mathcal{F}}^{n}_{T})_{C}$ in [@Walker:Thomason]). A point of the space $\big({\mathcal{G}}^{V}_{\coprod H/H_{i}}(n)^{an}\big)^{H}$ consists of an $H$-module quotient $[V^{n}{\twoheadrightarrow}W]$ and an equivariant map ${\mathbb{C}}^{H/H_{1}}\times \cdots \times{\mathbb{C}}^{H/H_{r}}\to \operatorname{End}_{{\mathbb{C}}}(W)$ of ${\mathbb{C}}$-algebras (here ${\mathbb{C}}^{H/H_i}{\cong}\oplus_{H/H_{i}}{\mathbb{C}}$ denotes the coordinate ring of $H/H_i$). This amounts to giving an $H$-module quotient $[V^{n}{\twoheadrightarrow}W]$, a direct sum decomposition $W= W_{1}\oplus \cdots \oplus W_{r}$ of $H$-modules, and $H_{i}$-module decomposition $W_{i}= \oplus_{H/H_{i}}W_{i}'$ inducing an $H$-module isomorphism $\mathrm{Ind}_{H_i}(W_i'){\cong}W_i$. Similarly a point of $\big({\mathcal{F}}^{V}_{\coprod H/H_{i}}(n)\big)^{H}$ is given by an $H$-module quotient $[V^{n}{\twoheadrightarrow}W]$ and an equivariant normal map of $C^{*}$-algebras ${\mathbb{C}}^{H/H_{1}}\times \cdots \times{\mathbb{C}}^{H/H_{r}}\to \operatorname{End}_{{\mathbb{C}}}(W)$. This amounts to giving an $H$-module quotient $[V^{n}{\twoheadrightarrow}W]$, an $H$-module orthogonal sum decomposition $W= W_{1}\perp \cdots \perp W_{r}$, and an orthogonal $H_{i}$-module decomposition $W_{i}= \perp_{H/H_{i}}W_{i}'$ inducing an $H$-module isomorphism $\mathrm{Ind}_{H_i}(W_i'){\cong}W_i$. Let $W$ be a finite dimensional Hermitian inner product $H$-module. For $H$-modules $W_1,\ldots, W_r$ such that $W{\cong}W_1\oplus \cdots \oplus W_r$, write $X(W_1,\ldots, W_r)$ for the space of decompositions $W=W_1\oplus \cdots \oplus W_r$ together with an $H_{i}$-module decomposition of each $W_{i}$, $W_{i}= \oplus_{H/H_{i}}W_{i}'$ which induces an $H$-module isomorphism $W_i{\cong}\mathrm{Ind}_{H_i}(W')$. Similarly, write $X^{{\mathrm{orth}}}(W_1,\ldots, W_r)$ for the space of orthogonal decompositions $W=W_1\perp \cdots \perp W_r$ with an orthogonal $H_{i}$-module decomposition of each $W_{i}$, $W_{i}= \perp_{H/H_{i}}W_{i}'$ which induces an $H$-module isomorphism $W_i{\cong}\mathrm{Ind}_{H_i}(W')$. First we consider the space of all decompositions of $W$ into an $r$-fold direct sum (resp.  orthogonal sum) of $H$-modules. These spaces break up into a disjoint union of connected components ${\mathrm{Grass}}_{W}(W_1,\ldots, W_{r})$ (resp.  ${\mathrm{Grass}}^{{\mathrm{orth}}}_{W}(W_1,\ldots, W_{r})$) of decompositions where the $i$th summand is $H$-isomorphic to $W_i$. As in [@Walker:Thomason], we proceed by induction on $r$ to show that ${\mathrm{Grass}}_{W}^{{\mathrm{orth}}}(W_1,\ldots, W_{r})\subseteq {\mathrm{Grass}}_{W}(W_1,\ldots, W_{r})$ is a weak equivalence for any $W$ and a tuple $(W_1,\ldots, W_r)$ of $G$-modules, such that $W = \oplus_iW_i$. The case $r=1$ is clear. The map ${\mathrm{Grass}}_W(W_1,\ldots, W_{r}) \to {\mathrm{Grass}}_W(W_1,\ldots, W_{r-1}\oplus W_{r})$ is a fibration with fiber ${\mathrm{Grass}}_{W'}(W_{r-1},W_r)$ where $W'= W_{r-1}\oplus W_{r}$. Similarly in the orthogonal case we have a fibration ${\mathrm{Grass}}^{{\mathrm{orth}}}_W(W_1,\ldots, W_{r}) \to {\mathrm{Grass}}^{{\mathrm{orth}}}_W(W_1,\ldots, W_{r-1}\oplus W_{r})$ with fiber ${\mathrm{Grass}}^{{\mathrm{orth}}}_{W'}(W_{r-1},W_r)$. Now the forgetful map from ${\mathrm{Grass}}^{{\mathrm{orth}}}_{W'}(W_{r-1},W_r)$ to the space ${\mathrm{Grass}}^{H}(W_{r-1}\subseteq W')$ of sub-$H$-planes isomorphic to $W_{r-1}$ is a homeomorphism. On the other hand ${\mathrm{Grass}}_{W'}(W_{r-1},W_r)\to {\mathrm{Grass}}^{H}(W_{r-1}\subseteq W')$ is a fibration with contractible fibers. We conclude that ${\mathrm{Grass}}_{W}^{{\mathrm{orth}}}(W_1,\ldots, W_{r})\subseteq {\mathrm{Grass}}_{W}(W_1,\ldots, W_{r})$ is a weak equivalence. The forgetful map $X(W_1,\ldots,W_r)\to {\mathrm{Grass}}_W(W_1,\ldots, W_{r})$ is also a fibration with fiber $F$ which is the space of $H_{i}$-module decompositions $W_{i} = \oplus_{H/H_{i}}W_i'$ inducing an $H$-module isomorphism $\mathrm{Ind}_{H_i}(W_{i}'){\cong}W_{i}$, $i=1,\ldots, r$. Such a decomposition is determined by an $H_{i}$-equivariant embedding $W_{i}'\subseteq W_{i}$, such that the induced map $\mathrm{Ind}_{H_i}(W_{i}'){\cong}W_{i}$ is an isomorphism, together with a choice of $H_{i}$-equivariant splitting of the projection $W_i\to W_{i}/W_i'$. This implies that the fibers of $F\to Y_{H_1}(W_1'\subseteq W_1)\times\cdots\times Y_{H_r}(W_r'\subseteq W_r)$, where $Y_{H_{i}}(W_{i}'\subseteq W_{i})$ is the space of $H_i$-equivariant embeddings $W_{i}'\subseteq W_{i}$ which induce an isomorphism $\mathrm{Ind}_{H_i}W_i'{\cong}W_{i}$, are contractible. Similarly $X^{{\mathrm{orth}}}(W_1,\ldots,W_r)\to {\mathrm{Grass}}_W^{{\mathrm{orth}}}(W_1,\ldots, W_{r})$ is a fibration with fiber $F^{{\mathrm{orth}}}$, which is the space of decompositions of $H_{i}$-module decompositions $W_{i} = \perp_{H/H_{i}}W_i'$ inducing an $H$-module isomorphism $\mathrm{Ind}_{H_i}(W_{i}'){\cong}W_{i}$. This space is homeomorphic to ${\mathrm{Grass}}^{H_1}(W_1'\subseteq W_1)\times\cdots\times {\mathrm{Grass}}^{H_r}(W_r'\subseteq W_r)$. \[lem:ehpt\] Let $G$ be a finite group and $V$ a representation. Let $\{H_{j}\}$ be finite set of subgroups of $G$, $W_{j}$ representations of $H_{j}$, $C_{j}$ the unit open polydisk in $W_j$, and $\tilde{C}_{j}= G\times^{H_j}C_j$. The inclusions $\iota:\coprod G/H_{j}\to \coprod G\times^{H_{j}}W_{j}$ induce equivariant weak equivalences of topological spaces $\iota_{*}:{\mathcal{G}}^{V}_{\coprod G/H_{i}}(n)^{an}\to ({\mathcal{G}}^{V}_{\coprod G\times^{H_{j}}W_{j}}(n)^{an})_{\coprod\tilde{C}_{j}}$ and $\iota_{*}:{\mathcal{F}}^{V}_{\coprod G/H_{i}}(n)\to ({\mathcal{F}}^{V}_{\coprod G\times^{H_{j}}W_{j}}(n))_{\coprod\tilde{C}_{j}}$. The arguments in both cases are similiar, so we treat only the first case. Let $U=\operatorname{\mathrm{Spec}}(A)$ be an affine $G$-variety. We have an equivariant natural transformation of functors $${\mathcal{G}}^{V}_{U}(n)\times {\mathbb{A}}^{1}\to {\mathcal{G}}^{V}_{U\times {\mathbb{A}}^1}(n)$$ defined on affine varieties as follows. An $R$-valued point of the left-hand side is a triple $(V^n\otimes_{{\mathbb{C}}} R{\twoheadrightarrow}Q, \rho:A\to \operatorname{End}_{R}(Q), f:{\mathbb{C}}[t]\to R)$, where $Q$ is a projective $R$-module, $\rho$ and $f$ are ${\mathbb{C}}$-algebra maps. Write $\phi$ for the ${\mathbb{C}}$-algebra map defined to be the composition $\eta f:{\mathbb{C}}[t]\to R\to \operatorname{End}_{R}(Q)$, where $\eta(r)$ is multiplication by $r$. Now define $F$ by sending this triple to $(V^{n}\otimes R{\twoheadrightarrow}Q, \rho\otimes\phi:A\otimes {\mathbb{C}}[t]\to \operatorname{End}_{R}(Q))$. Moreover, the resulting square $$\xymatrix{ {\mathcal{G}}^{V}_{U}(n)\times {\mathbb{A}}^{1}\ar[r]\ar[d]^{\theta_U\times\mathrm{id}} & {\mathcal{G}}^{V}_{U\times {\mathbb{A}}^1}(n) \ar[d]^{\theta_{U\times{\mathbb{A}}^1}}\\ (\coprod {\mathrm{Sym}}^k(U))\times {\mathbb{A}}^{1}\ar[r] & \coprod {\mathrm{Sym}}^{k}(U\times{\mathbb{A}}^{1}) }$$ is commutative. Write $\pi: \coprod G\times^{H_{j}}W_{j} \to \coprod G/H_{j}$ for the projection. Then $\pi_{*}\iota_{*} = id_{*}$. We obtain an equivariant homotopy between $\iota_{*}\pi_{*}$ and $id_{*}$ by using the restriction of the natural transformation $F$ to $I=[0,1]$ and the maps $(\coprod G\times^{H_{j}}W_{j})\times{\mathbb{A}}^{1}\to \coprod G\times^{H_{j}}W_{j}$ given by $(x,t)\mapsto t\smash\cdot x$. Note that the commutativity of the above square implies that the resulting homotopy restricts to an equivariant homotopy $({\mathcal{G}}^{V}_{\coprod G\times^{H_{j}}W_{j}}(n)^{an})_{\coprod\tilde{C}_{j}}\times I \to ({\mathcal{G}}^{V}_{\coprod G\times^{H_{j}}W_{j}}(n)^{an})_{\coprod\tilde{C}_{j}}$. Since we have that ${\mathcal{K}}^{sst}_{G}({\mathbb{C}},\, Y) = {\mathcal{K}}^{qtop}_{G}(S^0,\, Y)$, Theorem \[mainthm\] is a special case of the following equivariant generalization of [@Walker:Thomason Corollary 5.9]. \[cor:qtopeq\] Let $Y$ be a smooth quasi-projective $G$-variety and $A$ a based $G$-$CW$-complex. Then $${\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,\,Y^{an})\to {\mathcal{K}}^{qtop}_{G}(A,\, Y)$$ is an equivariant weak equivalence. By the previous theorem the map $$|\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{F}}_{Y^{an}}^{V}(n))| \to |\operatorname{Hom}_{cts*}(A\wedge\Delta^{\bullet}_{top\,+},\,{\mathcal{G}}_{Y}^{V}(n))^{an}|$$ is an equivariant weak equivalence for all $n$ and any $V$. Taking a homotopy colimit over $I$ gives an equivariant weak equivalence of equivariant $\Gamma$-spaces $$|{\mathcal{A}}^{top}_{G}(A,(-)\wedge Y_{+}^{an})^{V}|\to |{\mathcal{A}}^{qtop}_{G}(A,(-)\wedge Y_{+})^{V}|$$ and thus of associated $G$-spectra. The result is the particular case $V={\mathbb{C}}[G]$. Pairings and operations {#sec:pair} ======================= In this section we establish two basic pairings on our bivariant theories generalizing those of [@Walker:Thomason section 6] to the equivariant setting. These pairings are compatible with the natural comparison maps of $G$-spectra $$\begin{aligned} \label{eqn:manymaps} {\mathcal{K}}_{G}(X,Y) & \to {\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Y) \xleftarrow{{\simeq}} {\mathcal{K}}^{sst}_{G}(X,\, Y) \\ \nonumber & \to {\mathcal{K}}^{qtop}_{G}(X^{an},\, Y) \stackrel{\simeq}{\leftarrow} {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(X^{an},\,Y^{an}) \to {\mathbf{bu}}_{G}(X^{an},\,Y^{an}),\end{aligned}$$ obtained from (\[eqn:algsst\]), (\[eqn:compbiv\]), and (\[eqn:c2nc\]). The existence and compatibility of these pairings plays a crucial role in the applications in Sections \[sec:sstThom\] and \[sec:algThom\]. An important special case occurs when $Y=\operatorname{\mathrm{Spec}}({\mathbb{C}})$. Then the pairings specialize to give these spectra the structure of commutative ring spectra and these maps are maps of ring spectra. In particular, combined with Proposition \[prop:conn\] below, we have that $$K_{*}^{G}(X) \to K_{*}^{G,\,sst}(X)\to KU^{-*}_{G}(X^{an})$$ are graded ring homomorphisms. The construction of these bivariant theories all begin with the consideration of a functor $F:I\to G{\mathrm{sSet}}$, where $I$ is the category whose objects are the sets $\underline{n}=\{1,2,\ldots, n\}$ for each $n\geq 0$ (so $\underline{0}$ is the empty set) together with injective set maps. Let $\diamond:I\times I \to I$ denote the functor which sends the pair $(\underline{m},\underline{n})$ to $\underline{mn}$. Given injections $\alpha:\underline{m}\to \underline{n}$ and $\beta:\underline{p}\to \underline{q}$ in $I$ we define $\alpha\diamond\beta:\underline{mp}\to \underline{nq}$ by $$\alpha\diamond\beta((i-1)p+j) = (\alpha(i)-1)q+\beta(j)$$ for $i\in\underline{m}$, $j\in \underline{p}$, which is the map obtained by identifying $\underline{m}\times\underline{p}$ and $\underline{n}\times\underline{q}$ with $\underline{mp}$ and $\underline{nq}$ via the lexicographical ordering. Let $F$, $G$, $H:I\to G{\mathrm{sSet}}$ be $I$-diagrams of $G$-simplicial sets. A *pairing of $I$-diagrams* $F\times G\to H$, is a natural transformation $F\times G \to H\diamond$ of functors $I\times I\to G{\mathrm{sSet}}$. Such a pairing induces a pairing of $G$-simplicial sets $$\operatorname*{\mathrm{hocolim}}_{I}F\times\operatorname*{\mathrm{hocolim}}_{I}G\xrightarrow{{\cong}} \operatorname*{\mathrm{hocolim}}_{I\times I}F\times G \to \operatorname*{\mathrm{hocolim}}_{I\times I}H\diamond \to \operatorname*{\mathrm{hocolim}}_{I} H .$$ The *external product* is defined as follows. First we define a pairing $$\boxtimes:\operatorname{Hom}(X,\,{\mathcal{G}}_{Y}^{V}(m))\times \operatorname{Hom}(W,\,{\mathcal{G}}_{Z}^{V}(n)) \to \operatorname{Hom}(X\times W,\,{\mathcal{G}}_{Y\times Z}^{V}(mn))$$ of $I$-diagrams as follows. Define $$[p:{\mathcal{V}}^{m}_{X\times Y}{\twoheadrightarrow}{\mathcal{M}}]\boxtimes [q:{\mathcal{V}}^{n}_{W\times Z}{\twoheadrightarrow}{\mathcal{N}}]$$ to be the composition $${\mathcal{V}}^{mn}_{X\times W\times Y\times Z} {\cong}\pi_{X\times Y}^{*}{\mathcal{V}}^{m}_{X\times Y}\otimes \pi^{*}_{W\times Z}{\mathcal{V}}^{n}_{X\times Y} \to \pi^{*}_{X\times Y}{\mathcal{M}}\otimes\pi^{*}_{W\times Z}{\mathcal{N}},$$ where the first isomorphism is given using the lexicographical ordering, that is $e_{i}\otimes e_{j}$ is sent to $e_{(i-1)n+j}$. The maps $\pi_{X\times Y}$ and $\pi_{W\times Z}$ are the evident projections. It is clear that $\alpha^{*}\boxtimes \beta^{*} = (\alpha\diamond\beta)^{*}$ and thus we have a natural pairing of $I$-diagrams of $G$-sets. We thus obtain the natural pairing of $I$-diagrams of $G$-simplicial sets $$\begin{aligned} \boxtimes:\operatorname{Hom}(X\times\Delta^{\bullet}_{{\mathbb{C}}},&\,{\mathcal{G}}_{Y}^{V}(m))\times \operatorname{Hom}(W\times\Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Z}^{V}(n)) \\ & \to \operatorname{Hom}(X\times W\times\Delta^{\bullet}_{{\mathbb{C}}}\times\Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Y\times Z}^{V}(mn)) \\ & \to \operatorname{Hom}(X\times W\times\Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Y\times Z}^{V}(mn)) \end{aligned}$$ where the second map is induced by the diagonal $\Delta^{d}_{{\mathbb{C}}}\to \Delta^{d}_{{\mathbb{C}}}\times\Delta^{d}_{{\mathbb{C}}}$. Similarly making use of the diagonal $\Delta^{d}_{top}\to \Delta^{d}_{top}\times\Delta^{d}_{top}$ we obtain $$\begin{aligned} \boxtimes:\operatorname{Hom}(X\times\Delta^{\bullet}_{top},&\,{\mathcal{G}}_{Y}^{V}(m))\times \operatorname{Hom}(W\times\Delta^{\bullet}_{top},\,{\mathcal{G}}_{Z}^{V}(n)) \\ & \to \operatorname{Hom}(X\times W\times\Delta^{\bullet}_{top},\,{\mathcal{G}}_{Y\times Z}^{V}(mn)).\end{aligned}$$ Taking homotopy colimits we obtain the external pairing of $G$-simplicial sets $$\begin{aligned} \boxtimes&:{\mathcal{A}}_{G}(X,Y)^{V} \times {\mathcal{A}}_{G}(W,Z)^{V}\to {\mathcal{A}}_{G}(X\times W,Y\times Z)^{V},\\ \boxtimes&:{\mathcal{A}}_{G}(X\times\Delta^{\bullet}_{top},Y)^{V} \times {\mathcal{A}}_{G}(W\times\Delta^{\bullet}_{top},Z)^{V}\to {\mathcal{A}}_{G}(X\times W\times\Delta^{\bullet}_{top},Y\times Z)^{V},\,\,\textrm{and} \\ \boxtimes&:{\mathcal{A}}^{sst}_{G}(X,Y)^{V} \times {\mathcal{A}}^{sst}_{G}(W,Z)^{V}\to {\mathcal{A}}^{sst}_{G}(X\times W,Y\times Z)^{V}.\end{aligned}$$ A straightforward verification shows that this pairing is associative. To see that the three pairings are compatible with the first two natural transformations of \[eqn:manymaps\], use the naturality of $\operatorname{Hom}$ in the first variable applied to the diagonal $\Delta^{\bullet} \to \Delta^{\bullet} \times \Delta^{\bullet}$ and the projection $\Delta^{\bullet} \to \Delta^{0}$, both for ${\Delta^{\bullet}_{{\mathbb{C}}}}$ and $\Delta^{\bullet}_{top}$. The pairing $\boxtimes$ extends to give an associative pairing $$\boxtimes:{\mathcal{A}}^{qtop}_{G}(X^{an}, Y)^{V} \times {\mathcal{A}}^{qtop}_{G}(W^{an},Z)^{V} \to {\mathcal{A}}^{qtop}_{G}((X\times W)^{an},Y\times Z)^{V}.$$ compatible with the one on ${\mathcal{A}}^{sst}_{G}$. These pairings give rise to an external pairing of equivariant $\Gamma$-spaces $$\begin{gathered} (\underline{m}_{+}\mapsto {\mathcal{A}}_{G}(X,m_{+}\wedge Y_{+})^{V})\times (\underline{n}_{+}\mapsto {\mathcal{A}}_{G}(W,n_{+}\wedge Z_{+})^{V}) \\ \xrightarrow{\boxtimes} ((\underline{m}_{+},\underline{n}_{+})\mapsto {\mathcal{A}}_{G}(X\times W, m_{+}\wedge Y_{+} \wedge n_{+}\wedge Z_{+})\end{gathered}$$ and similarly for ${\mathcal{A}}^{sst}_{G}$ and ${\mathcal{A}}^{qtop}_{G}$. We therefore obtain by the discussion in Section \[subsectionpairings\] pairings of natural and associative pairings of spectra. $$\begin{aligned} \boxtimes&:{\mathcal{K}}_{G}(X,Y)\wedge{\mathcal{K}}_{G}(W, Z) \to {\mathcal{K}}_{G}(X\times W, Y\times Z), \\ \boxtimes&:{\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Y)\wedge{\mathcal{K}}_{G}(W\times\Delta^{\bullet}_{top}, Z) \to {\mathcal{K}}_{G}(X\times W\times\Delta^{\bullet}_{top}, Y\times Z), \\ \boxtimes&:{\mathcal{K}}^{sst}_{G}(X,Y)\wedge{\mathcal{K}}^{sst}_{G}(W, Z) \to {\mathcal{K}}^{sst}_{G}(X\times W, Y\times Z)\textrm{, and} \\ \boxtimes&:{\mathcal{K}}^{qtop}_{G}(X^{an},Y)\wedge{\mathcal{K}}^{qtop}_{G}(W^{an}, Z) \to {\mathcal{K}}^{qtop}_{G}(X^{an}\times W^{an}, Y\times Z).\end{aligned}$$ Now we define a pairing $$\begin{gathered} \boxtimes:{\mathcal{F}}^{V}_{S}(m)\times {\mathcal{F}}^{V}_{T}(n) = \underline{\operatorname{Hom}}_{*}({\mathcal{C}}_{0}(S), \operatorname{End}_{{\mathbb{C}}}(V^{m}))\times \underline{\operatorname{Hom}}_{*}({\mathcal{C}}_{0}(T),\operatorname{End}_{{\mathbb{C}}}(V^{n})) \\ \to \underline{\operatorname{Hom}}_{*}({\mathcal{C}}(S\wedge T), \operatorname{End}_{{\mathbb{C}}}(V^{mn})) = {\mathcal{F}}^{V}_{S\wedge T}(mn).\end{gathered}$$ Given $f:{\mathcal{C}}_{0}(S)\to \operatorname{End}_{{\mathbb{C}}}(V^{m})$ and $g:{\mathcal{C}}_{0}(T)\to \operatorname{End}_{{\mathbb{C}}}(V^{n})$ we define $f\boxtimes g$ to be the composition $$f\boxtimes g:{\mathcal{C}}_{0}(S\wedge T){\cong}{\mathcal{C}}_{0}(S)\otimes{\mathcal{C}}_{0}(T)\xrightarrow{f\otimes g} \operatorname{End}_{{\mathbb{C}}}(V^{m})\otimes \operatorname{End}_{C}(V^{n})\to \operatorname{End}_{{\mathbb{C}}}(V^{mn})$$ where the last map uses the lexicographical indexing $\ell:V^{m}\otimes V^{n}{\cong}V^{mn}$ via $e_{i}\otimes e_{j}\mapsto e_{n(i-1)+j}$. Given injections $\alpha:\underline{m}\to \underline{p}$ and $\beta:\underline{n}\to \underline{q}$ then under the above isomorphism we have that $\tilde{\alpha}f\otimes \tilde{\beta}g$ agrees with $\widetilde{\alpha\diamond\beta}\ell(f\otimes g)$. We obtain a natural pairing $$\begin{aligned} \boxtimes:\operatorname{Hom}_{cts*}(A,{\mathcal{F}}^{V}_{S}(m)) & \times \operatorname{Hom}_{cts*}(B,{\mathcal{F}}^{V}_{S}(n)) \\ & \to \operatorname{Hom}_{cts*}(A\wedge B,{\mathcal{F}}^{V}_{S}(m)\times{\mathcal{F}}^{V}_{T}(n))\\ & \to \operatorname{Hom}_{cts*}(A\wedge B,{\mathcal{F}}^{V}_{S\wedge T}(mn))\end{aligned}$$ of $I$-diagrams. This pairing is associative in the evident sense and induces the external pairing of equivariant $\Gamma$-spaces $$\boxtimes:{\mathcal{A}}^{top}_{G}(A,-\wedge S)^{V}\wedge {\mathcal{A}}^{top}_{G}(B,-\wedge T)^{V}\to {\mathcal{A}}^{top}_{G}(A\wedge B, -\wedge S\wedge - \wedge T)^{V}$$ and thus a pairing of $G$-spectra $$\boxtimes:{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A,S)\wedge {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(B,T) \to {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(A\wedge B, S\wedge T).$$ This pairing extends to an external pairing of ${\mathcal{W}}_{G}$-spaces and thus we obtain the pairing of $G$-spectra $$\boxtimes:{\mathbf{bu}}_{G}(A,S)\wedge {\mathbf{bu}}_{G}(B,T) \to {\mathbf{bu}}_{G}(A\wedge B, S\wedge T).$$ It is clear that the last of the natural transformations of \[eqn:manymaps\] is compatible with the pairings. To see this for the second last one, one uses the naturality of the maps discussed in section \[sec:comp\]. Write ${\mathcal{K}}^{?}_{G}(X,Y)$ for any one of the six bivariant theories appearing in (\[eqn:manymaps\]). The following proposition summarizes the preceding discussion. Let $X$, $Y$, $W$, $Z$, $S$, and $T$, be quasi-projective $G$-varieties. We have pairings $$\boxtimes:{\mathcal{K}}_{G}^{?}(X,Y) \wedge {\mathcal{K}}^{?}_{G}(W,Z)\to {\mathcal{K}}^{?}_{G}(X\times W, Y\times Z).$$ This pairing is associative in the sense that the two evident maps $${\mathcal{K}}^{?}_{G}(X,Y) \times {\mathcal{K}}^{?}_{G}(W,Z)\times {\mathcal{K}}^{?}_{G}(S,T)\to {\mathcal{K}}^{?}_{G}(X\times W\times S,Y\times Z\times T)$$ agree. Moreover, these pairings are compatible with the each of the natural transformations (\[eqn:manymaps\]). Taking $X=W=\operatorname{\mathrm{Spec}}({\mathbb{C}})$, here and below this is to be interpreted as $S^{0}$ in the topological case, in the external product defines the *external product for homology* $$\underline{\wedge}:{\mathcal{K}}^{?}_{G}({\mathbb{C}},Y)\wedge{\mathcal{K}}^{?}_{G}({\mathbb{C}},Z)\to {\mathcal{K}}^{?}_{G}({\mathbb{C}},Y\times Z).$$ Specializing to $Y=Z=\operatorname{\mathrm{Spec}}({\mathbb{C}})$ in the external pairing defines the *external product for cohomology*, $$\barwedge:{\mathcal{K}}^{?}_{G}(X,{\mathbb{C}})\wedge {\mathcal{K}}^{?}_{G}(W,{\mathbb{C}})\to {\mathcal{K}}^{?}_{G}(X\times W,{\mathbb{C}}).$$ We define the cup product by specializing further to $X=W$ and composing with the pullback along the diagonal $\Delta:X\to X\times X$, $$\label{eqn:cup} \cup:{\mathcal{K}}^{?}_{G}(X,{\mathbb{C}})\wedge {\mathcal{K}}^{?}_{G}(X,{\mathbb{C}}) \to {\mathcal{K}}^{?}_{G}(X,{\mathbb{C}}).$$ The cup-product turns $\pi_{*}{\mathcal{K}}^{?}_{G}(X,{\mathbb{C}})$ into a graded ring, and even into a graded $\pi_{*}{\mathcal{K}}^{?}_{G}({\mathbb{C}},{\mathbb{C}})$-algebra. Immediate from the definitions we have the following. \[prop:cup\] Let $X$ be a quasi-projective $G$-variety. The natural maps $$\begin{aligned} K^{G,\,{\mathbb{A}}^{1}}_{*}(X,{\mathbb{C}}) & \to K^{G}_{*}(X\times\Delta^{\bullet}_{top},{\mathbb{C}})\xleftarrow{{\cong}} K^{G,\, sst}_{*}(X,{\mathbb{C}}) \\ & \to K^{G,\,qtop}_{*}(X^{an},{\mathbb{C}}) \stackrel{{\cong}} {\leftarrow} {\mathbf{bu}}^{G,{\mathfrak{c}}}_{*}(X^{an},S^{0}) \to {\mathbf{bu}}^{G}_{*}(X^{an},S^{0})\end{aligned}$$ induced by (\[eqn:manymaps\]) are graded ring homomorphisms. Our second basic pairing is the composition pairing. There is a composition pairing $$\theta_{X,Y,Z}:\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X,\,{\mathcal{G}}_{Y}^{V}(m)) \times \operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(Y,\,{\mathcal{G}}_{Z}^{V}(n)) \to \operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X,\,{\mathcal{G}}_{Z}^{V}(mn))$$ defined as follows. Given a pair $$([p:{\mathcal{V}}^{m}_{X\times Y}{\twoheadrightarrow}{\mathcal{M}}],[q:{\mathcal{V}}^{n}_{Y\times Z}{\twoheadrightarrow}{\mathcal{N}}])$$ we have the quotient object $$[p\otimes q:{\mathcal{V}}^{mn}_{X\times Y\times Z}{\cong}{\mathcal{V}}^{m}_{X\times Y\times Z}\otimes {\mathcal{V}}^{n}_{X\times Y\times Z} \to \pi^{*}_{X\times Y}{\mathcal{M}}\otimes\pi_{Y\times Z}^{*}{\mathcal{N}}]$$ where the isomorphism is via the lexicographical ordering, as in the definition of the external product pairing above, and $\pi_{X\times Y}$ and $\pi_{Y\times Z}$ are the evident projections. Now pushforward along the projection $\pi_{X\times Z}$ define $\theta(p,q)$ $$[\theta(p,q):{\mathcal{V}}^{mn}_{X\times Z} \to (\pi_{X\times Z})_{*}(\pi^{*}_{X\times Y}{\mathcal{M}}\otimes\pi_{Y\times Z}^{*}{\mathcal{N}})].$$ It is easily verified that $\theta_{X,Y,Z}(p,q)\in \operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(X,\,{\mathcal{G}}_{Z}^{V}(mn))$ as needed. Moreover if $\alpha:\underline{m}\to\underline{m'}$ and $\beta:\underline{n}\to \underline{n'}$ are injections then $\theta(\alpha^{*}p,\beta^{*}q) = (\alpha\diamond\beta)^{*}\theta(p,q)$ and thus $\theta$ defines a pairing of $I$-diagrams. Abusing notation slightly, we also write $$\theta_{X,Y,Z}:\operatorname{Hom}(X\times U,\,{\mathcal{G}}_{Y}^{V}(m)) \times \operatorname{Hom}(Y\times U,\,{\mathcal{G}}_{Z}^{V}(n)) \to \operatorname{Hom}(X\times U,\,{\mathcal{G}}_{Z}^{V}(mn))$$ for the pairing obtained by composing with the map ${\mathcal{G}}_{Y}^{V}(n)\to {\mathcal{G}}_{Y\times U}^{V}(n)$ defined by $[q:{\mathcal{V}}^{n}_{X\times Y}{\twoheadrightarrow}{\mathcal{M}}]\mapsto [{\mathcal{V}}^{n}_{X\times Y\times U}{\twoheadrightarrow}\pi^{*}_{X\times Y}{\mathcal{M}}]$. We thus obtain pairings $$\operatorname{Hom}(X\times \Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Y}^{V}(m)) \times \operatorname{Hom}(Y\times \Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Z}^{V}(n)) \to \operatorname{Hom}(X\times \Delta^{\bullet}_{{\mathbb{C}}},\,{\mathcal{G}}_{Z}^{V}(mn))$$ and $$\operatorname{Hom}(X\times \Delta^{\bullet}_{top},\,{\mathcal{G}}_{Y}^{V}(n)) \times \operatorname{Hom}(Y\times \Delta^{\bullet}_{top},\,{\mathcal{G}}_{Z}^{V}(m)) \to \operatorname{Hom}(X\times \Delta^{\bullet}_{top},\,{\mathcal{G}}_{Z}^{V}(mn)).$$ After taking homotopy colimits we obtain the pairing of equivariant $\Gamma$-spaces $${\mathcal{A}}_{G}(X,\underline{m}_{+}\wedge Y_{+})\times {\mathcal{A}}_{G}(Y, \underline{n}_{+}\wedge Z_{+})\to {\mathcal{A}}_{G}(X,\underline{mn}_{+}\wedge Z_{+})$$ where as usual $\underline{m}_{+}\wedge\underline{n}_{+}$ is identified with $\underline{mn}_{+}$ via $(i,j)\mapsto (i-1)n+j$, and similarly for ${\mathcal{A}}^{sst}_{G}$. For a space $T$, $\operatorname{Hom}_{cts}(T,G_{Y}^{an}) = \operatorname*{\mathrm{colim}}_{T\to U^{an}}\operatorname{Hom}_{{\mathrm{Sch}}/{\mathbb{C}}}(U,G_{Y}^{an})$. With this observation, the definition of $\theta$ readily extends to a pairing for ${\mathcal{A}}^{qtop}_{top}$. We thus obtain natural pairings of $G$-spectra $$\begin{aligned} \theta&:{\mathcal{K}}_{G}(X,Y)\wedge {\mathcal{K}}_{G}(Y,Z) \to {\mathcal{K}}_{G}(X,Z), \\ \theta&:{\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Y)\wedge {\mathcal{K}}_{G}(Y\times\Delta^{\bullet}_{top},Z) \to {\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},Z), \\ \theta&:{\mathcal{K}}^{sst}_{G}(X,Y)\wedge {\mathcal{K}}^{sst}_{G}(Y,Z) \to {\mathcal{K}}^{sst}_{G}(X,Z), \,\,\textrm{and}\\ \theta&:{\mathcal{K}}^{qtop}_{G}(X,Y)\wedge {\mathcal{K}}^{qtop}_{G}(Y,Z) \to {\mathcal{K}}^{qtop}_{G}(X,Z)\end{aligned}$$ Now for based $G$-$CW$-complexes $S$,$T$, and $U$ we define the pairing $$\theta_{S,T,U}:\operatorname{Hom}_{cts*}(S,\,{\mathcal{F}}_{T}^{V}(m)) \times \operatorname{Hom}_{cts*}(T,\,{\mathcal{F}}_{U}^{V}(n)) \to \operatorname{Hom}_{cts*}(S,\,{\mathcal{F}}_{U}^{V}(mn))$$ defined by sending a pair of $*$-maps $$(p:{\mathcal{C}}_{0}(T)\to {\mathcal{C}}_{0}(S)\otimes\operatorname{End}_{{\mathbb{C}}}(V^{m}),\,\, q:{\mathcal{C}}_{0}(U)\to {\mathcal{C}}_{0}(T)\otimes\operatorname{End}_{{\mathbb{C}}}(V^{n}))$$ to the composite $$\begin{aligned} {\mathcal{C}}_{0}(U)\xrightarrow{q} {\mathcal{C}}_{0}(T)&\otimes\operatorname{End}_{{\mathbb{C}}}(V^{n})\xrightarrow{p\otimes 1} {\mathcal{C}}_{0}(S)\otimes\operatorname{End}_{{\mathbb{C}}}(V^{m})\otimes\operatorname{End}_{{\mathbb{C}}}(V^{n}) \\ &\to {\mathcal{C}}_{0}(S)\otimes \operatorname{End}_{{\mathbb{C}}}(V^{m}\otimes V^{n})\xrightarrow{{\cong}} {\mathcal{C}}_{0}(S)\otimes \operatorname{End}_{{\mathbb{C}}}(V^{mn}),\end{aligned}$$ where in the last map we have identified $V^{m}\otimes V^{n}$ with $V^{mn}$ via the lexicographical ordering as above. It is straightforward to check that $\theta(\tilde{\alpha}p,\tilde{\beta}q) = \widetilde{\alpha\diamond\beta}\theta(p,q)$. We thus obtain a pairing of $I$-diagrams which gives a pairing of $\Gamma$-spaces and therefore a pairing of $G$-spectra $$\theta:{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S,T)\wedge {\mathbf{bu}}_{G}^{{\mathfrak{c}}}(T,U) \to {\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S,U).$$ Similarly we have a pairing of ${\mathcal{W}}_{G}$-spaces leading to a pairing of $G$-spectra $$\theta:{\mathbf{bu}}_{G}(S,T)\wedge {\mathbf{bu}}_{G}(T,U)\to {\mathbf{bu}}_{G}(S,U).$$ The pairing $\theta$ enjoys the same properties as in the non-equivariant case, namely naturality, associativity and compatibility both with the pairing $\boxtimes$ and with the natural transformations (\[eqn:manymaps\]). That is, the equivariant analogues of [@Walker:Thomason Propositions 6.4, 6.5 and 6.6.] all hold. We define slant products and the cap product in the usual fashion. In the topological case, $\operatorname{\mathrm{Spec}}({\mathbb{C}})$ is interpreted as $S^{0}$. As above, we write ${\mathcal{K}}^{?}_{G}$ for any one of the bivariant theories appearing in (\[eqn:manymaps\]). Let $X$ and $Y$ be quasiprojective $G$-varieties resp.  $G-CW$-complexes. 1. We define the slant product pairing $$/: {\mathcal{K}}^{?}_{G}(X \times Y,{\mathbb{C}}) \wedge {\mathcal{K}}^{?}_{G}({\mathbb{C}},Y) \to {\mathcal{K}}^{?}_{G}(X,{\mathbb{C}})$$ by $a/b:=\theta_{X,X \times Y, {\mathbb{C}}}(\tau(a \wedge (1_X \boxtimes b)))$, where $\tau$ is the obvious involution. 2. We define the slant product pairing $$\backslash: {\mathcal{K}}^{?}_{G}(X,{\mathbb{C}}) \wedge {\mathcal{K}}^{?}_{G}({\mathbb{C}},X \times Y) \to {\mathcal{K}}^{?}_{G}({\mathbb{C}},X)$$ by $a \backslash b:= \theta_{{\mathbb{C}},X \times Y,Y}(\tau(a\boxtimes 1_Y)\wedge b))$. 3. Finally, we define the cap product $$\cap: {\mathcal{K}}^{?}_{G}(X,{\mathbb{C}}) \wedge {\mathcal{K}}^{?}_{G}({\mathbb{C}},X) \to {\mathcal{K}}^{?}_{G}({\mathbb{C}},X)$$ by $a \cap b := a \backslash \Delta_*(b)$, where $\Delta: X \to X \times X$ is the diagonal embedding. Again by definition, these products are compatible with the natural transformations (\[eqn:manymaps\]). As in [@Walker:Thomason Proposition 6.10] we observe that the operations given here coincide with the “classical” ones. (See [@May:equihomotopy section XIII.5] for a discussion of the “classical operations” in the equivariant setting.) The results [@Walker:Thomason Lemma 6.12, Proposition 6.13] also hold equivariantly and are needed later. We can also use these pairings to define transfer maps for finite (but not necessarily dominant) equivariant morphisms $f:X\to Y$ between smooth projective complex $G$-varieties, for the bivariant equivariant $K$-theories we consider. Recall that we write $K_{*}^{G,\,alg}(X,Y)$ for the algebraic $K$-theory of the exact category ${\mathcal{P}}(G;X,Y)$ of coherent $G$-modules on $X\times Y$ which are finite and flat over $X$. Write $K_{*}^{\prime}(G;X,Y)$ for the $K$-theory groups of the abelian category ${\mathcal{M}}(G;X,Y)$ of coherent $G$-modules on $X\times Y$ which are finite over $X$. To define these transfer maps, we make use of the following equivariant analog of [@Walker:Thomason Lemma 6.14]. Let $X$ and $Y$ be smooth quasi-projective $G$-varieties, with $Y$ projective. The natural mapping $${\mathcal{K}}_{n}^{G,\,alg}(X,Y) \to {\mathcal{K}}_{n}^{\prime}(G;X,Y)$$ induced by the inclusion ${\mathcal{P}}(G;X,Y)\subseteq {\mathcal{M}}(G;X,Y)$ is an isomorphism for all $n\geq 0$. The proof of [@Walker:Adamsop Lemma 2.2] concerning the nonequivariant bivariant $K$-theory with $Y=({\mathbb{P}}^1)^{\times s}$ generalizes to equivariant $K$-theory and $X$ an arbitrary smooth quasi-projective $G$-variety if one replaces the subscheme $D$ in the proof by the union $GD:=\cup_{g \in G}gD$. The map $GD \to X$ is quasi-finite as $D \to X$ is and because $G$ is finite. To see that it is proper (and hence finite), one uses again that $G$ is finite. The same argument as in [@Walker:Adamsop Lemma 2.3] then yields the equivalence for $Y={\mathbb{P}}^n$. For an arbitrary $Y$ with non-trivial $G$-action we have a finite, surjective equivariant map $Y\to Y/G$. We may assume that $Y/G$ is connected and applying Noether normalization to $Y/G$ we obtain a finite, surjective equivariant map $Y\to {\mathbb{P}}^{n}$. Arguing as in [@Walker:Thomason Lemma 6.14] we see that the result for $Y$ follows from the result for ${\mathbb{P}}^{n}$. When $X$, $Y$ are smooth with $Y$ projective, the previous lemma, Proposition \[prop:fix\], and the sequence of natural transformations (\[eqn:manymaps\]) imply that a class $[{\mathcal{M}}]\in K^{\prime}_{0}(G;X,Y)$ naturally defines a class in each of the bivariant theories in (\[eqn:manymaps\]). We will write $[{\mathcal{M}}]$ as well for the class induced in any one of these bivariant theories. We continue to use the notation ${\mathcal{K}}^{?}_{G}$ for any one of the bivariant theories appearing in (\[eqn:manymaps\]) and write $K^{G,\,?}_{*}$ for its homotopy groups. \[transfer\] Let $f: Y \to X$ be a finite morphism of smooth projective complex $G$-varieties, and let $[\Gamma_f^{t}] \in K^{G,\,alg}_0(X,Y)$ be the element represented by the transpose of the graph of $f$. We define transfer maps $$f_*:K^{G,\,?}_*(Y,Z) \to K^{G,\,?}_*(X,Z)$$ and $$f^*:K^{G,\,?}_*(Z,X) \to K^{G,\,?}_*(Z,Y)$$ by $f_*(a):=\theta([\Gamma_f^{t}],a)$ and $f^*(b):=\theta(b,[\Gamma_f^{t}])$. Note that the notation $f^*$ is used both for the transfer map in the second variable as well as the usual contravariance in the first variable (and a similar overlap for the meaning of $f_{*}$). We adopt this notation to conform to [@Walker:Thomason]. The compatibility of the pairings $\theta$ with the natural transformations (\[eqn:manymaps\]) between the various equivariant $K$-theories implies that these natural transformations are also compatible with the transfer maps. The cup product pairing (\[eqn:cup\]) gives ${\mathbf{bu}}^{G}_{*}(W,S^{0})$ the structure of a graded commutative ring. We now relate this ring to $KU^{-*}_{G}(W)$, the periodic equivariant complex $K$-theory. For details on equivariant $K$-theory we refer the reader to [@May:equihomotopy chapter XIV] or [@Segal]. Recall that $KU_{G}^{0}(S^{0}) = R(G)$, the complex representation ring. For real representations $\alpha$, $\beta\in \operatorname{Rep}_{{\mathbb{R}}}(G)$, the tensor product of bundles defines a product $$\begin{gathered} KU_{G}^{-\alpha}(W)\otimes KU_{G}^{-\gamma}(W) = KU_{G}^{0}(S^{\alpha}\wedge W)\otimes KU_{G}^{0}(S^{\gamma}\wedge W) \\ \to KU_{G}^{0}(S^{\alpha}\wedge W \wedge S^{\gamma}\wedge W) \to KU_{G}^{0}(S^{\alpha}\wedge S^{\gamma}\wedge W) = KU_{G}^{-\alpha-\gamma}(W),\end{gathered}$$ making $\oplus_{\alpha\in \operatorname{Rep}_{{\mathbb{R}}}(G)}KU_{G}^{-\alpha}(W)$ into a ring. \[prop:conn\] For any based, compact $G$-$CW$-complex $W$, there is a natural isomorphism of graded rings $$\oplus_{\alpha\in \operatorname{Rep}_{{\mathbb{R}}}(G)} {\mathbf{bu}}^{G}_{\alpha}(W,S^{0}) \xrightarrow{{\cong}} \oplus_{\alpha\in \operatorname{Rep}_{{\mathbb{R}}}(G)}KU_{G}^{-\alpha}(W).$$ The argument is similar to that of [@Walker:Thomason Proposition 6.18]. Briefly, we have that ${\mathbf{bu}}^{G}_{0}(W,S^{0}) = KU^{0}_{G}(W)$ since by Corollary \[cor:cc\] both groups are the quotient of $KU^{0}(W_{+})$ by the subgroup $KU^{0}(S^{0})$. From Corollary \[cor:3.17\] it follows that ${\mathbf{bu}}^{G}_{\alpha}(W,S^{0}) = {\mathbf{bu}}_{0}^{G}(S^{\alpha}\wedge W, S^{0})$ therefore it suffices to show that the diagram $$\xymatrix{ {\mathbf{bu}}_{\alpha}^{G}(W,S^{0}) \otimes {\mathbf{bu}}_{\gamma}^{G}(W,S^{0}) \ar[r]\ar[d] & {\mathbf{bu}}_{\alpha+\gamma}(W,S^{0}) \ar[d] \\ {\mathbf{bu}}_{0}^{G}(S^{\alpha}, S^{0}) \otimes {\mathbf{bu}}_{0}^{G}(S^{\gamma},S^{0}) \ar[r] & {\mathbf{bu}}_{0}^{G}(S^{\alpha}\wedge S^{\gamma}\wedge W,S^{0}) }$$ commutes. This is easily seen to hold by definition of the cup product. For any complex representation $V$ there is a *Bott element* $\beta_{V}\in KU_{G}^{0}(S^{V})$ such that for any $X$, multiplication by $\beta_{V}$ is an isomorphism $$-\cup\beta_{V}:KU_{G}^{0}(X) \xrightarrow{{\cong}} KU_{G}^{0}(X\wedge S^{V}).$$ In particular there is a Bott element $\beta_{2}\in KU^{-2}_{G}(S^{0}) = {\mathbf{bu}}^{G}_{2}(S^{0},S^{0})$ corresponding to the trivial one-dimensional complex representation. We refer the reader to [@May:equihomotopy Section XIV.3] and of course to [@Segal] for details on equivariant Bott periodicity. \[cor:bottinv\] Let $W$ be a based, compact $G$-$CW$-complex. There is a natural map of graded rings $$\oplus_{\alpha\in RO(G)}{\mathbf{bu}}^{G}_{\alpha}(W,S^{0}) \to \oplus_{\alpha\in RO(G)}KU_{G}^{-\alpha}(W).$$ Let $\Lambda$ be a complete set of irreducible complex representations. Inverting the Bott elements corresponding to $\Lambda$ yields an isomorphism of $RO(G)$-graded rings $$\oplus_{\alpha\in \operatorname{Rep}_{{\mathbb{R}}}(G)}{\mathbf{bu}}^{G}_{\alpha}(W,S^{0})[\beta_{V}^{-1},\,V\in\Lambda] \xrightarrow{{\cong}} \oplus_{\alpha\in RO(G)}KU_{G}^{-\alpha}(W).$$ For any representation $\alpha\in \operatorname{Rep}_{{\mathbb{R}}}(G)$ and any complex representation $V$ we have isomorphisms $$-\cup \beta_{V} :{\mathbf{bu}}_{\alpha}^{G}(W, S^{0}) \xrightarrow{{\cong}} {\mathbf{bu}}_{\alpha+V}(W, S^{0}).$$ If $\gamma$ is a real representation then $\gamma\oplus \gamma$ can be given the structure of a complex representation. For real representations $\alpha$, $\gamma \in \operatorname{Rep}(G)$ the composition $${\mathbf{bu}}^{G}_{\alpha-\gamma}(W,S^{0}) \xrightarrow{\beta_{2\gamma}} {\mathbf{bu}}^{G}_{\alpha+\gamma}(W,S^{0}) \xrightarrow{}KU_{G}^{-\alpha-\gamma}(W) \xrightarrow{\beta_{2\gamma}^{-1}} KU_{G}^{-\alpha + \gamma}(W)$$ defines the desired map $\oplus_{\alpha\in RO(G)}{\mathbf{bu}}^{G}_{\alpha}(W,S^{0}) \to \oplus_{\alpha\in RO(G)}KU_{G}^{-\alpha}(W)$. The second statement is immediate. Comparing semi-topological and topological equivariant $K$-theory {#sec:sstThom} ================================================================= The main result of this section is Theorem \[mainthmbottinv\] below, where we show that Bott-inverted equivariant semi-topological $K$-theory and equivariant topological $K$-theory agree for projective $G$-varieties. In the next section, we will see that this yields a new proof of the equivariant version of Thomason’s theorem. Similar to [@Walker:Thomason Theorem 7.11], Theorem \[mainthmbottinv\] follows by combining three ingredients: Theorem \[mainthm\], the compatibility of operations established in the previous section, and Theorem \[walker7.10equi\] below comparing the action of certain operations with multiplication by the Bott element. The most significant difference is that unlike in the nonequivariant case $ku^{-*}_{G}(-)$ need not satisfy Poincare duality. Consequently, we have to modify several arguments. We will write $ku^{*}_{G}(-) = {\mathbf{bu}}^{G}_{-*}(-,S^{0})$ and $ku_{*}^{G}(-)={\mathbf{bu}}^{G}_{*}(S^{0},-)$ in this section. By Corollary \[cor:3.17\] these agree with the cohomology and homology theories associated to the spectrum ${\mathbf{bu}}_{G}(S^{0},S^{0})$. From Proposition \[prop:cup\] and Corollary \[cor:bottinv\] we have natural maps of graded rings $$\label{eqn:maps} K^{G,\,sst}_{*}(X,{\mathbb{C}})\to ku_{G}^{-*}(X^{an}) \to KU_{G}^{-*}(X^{an}).$$ In this and the next section $\ast$ will always denote ${\bf Z}$-grading. Write $\beta_{2}\in K^{G,\,sst}_{2}({\mathbb{C}},{\mathbb{C}})$ for the element corresponding to the Bott element $\beta_{2}\in ku^{-2}_{G}(S^{0})$ (see the discussion preceding Corollary \[cor:bottinv\]) under the isomorphism $K^{G,\,sst}_{2}({\mathbb{C}},{\mathbb{C}}){\cong}ku^{-2}_{G}(S^{0})$ obtained from Theorem \[mainthm\]. The element $\beta_{2}\in K^{G,\,sst}_{2}({\mathbb{C}},{\mathbb{C}})$ is referred to as the *semi-topological Bott element*. By Corollary \[cor:bottinv\] the right map of (\[eqn:maps\]) induces an isomorphism of graded rings $ku_{G}^{-*}(X^{an})[\beta^{-1}_{2}] \stackrel{\simeq}{\to} KU_{G}^{-*}(X^{an})$. In Theorem \[mainthmbottinv\] below we show that that when $X$ is a smooth and projective complex $G$-variety, the maps (\[eqn:maps\]) induce isomorphisms $$K^{G,sst}_{*}(X,{\mathbb{C}})[\beta^{-1}_{2}] \stackrel{\cong}{\to} ku_{G}^{-*}(X^{an})[\beta^{-1}_{2}]\xrightarrow{{\cong}} KU_{G}^{-*}(X^{an}).$$ We define $\delta_X \in K_0^G(X \times X)$ to be the class of the coherent $G$-module ${\mathcal{O}}_{\Delta}$. We also write $\delta_X \in K_0^{G,sst}(X \times X)$ for its image in semi-topological $K$-theory, and $\delta_{X^{an}}$ for its image in either $ku^{0}_G(X^{an} \times X^{an})$ or $KU^{0}_G(X^{an} \times X^{an})$. Note that $\delta_{X}=\Delta_{*}(1)$, where $\Delta_{*}$ is the transfer map, defined in Definition \[transfer\]. The remaining ingredient for the proof of Theorem \[mainthmbottinv\] is the following. \[walker7.10equi\] Let $X$ be a smooth complex projective $G$-variety of dimension $d$. There are classes $[X]\in KU^{G,\, sst}_{2d}({\mathbb{C}},X)$ and $\delta_{X}\in K_{0}^{G,\,sst}(X\times X,{\mathbb{C}})$ such that the composition $$K^{G,sst}_{*}(X,{\mathbb{C}}) \xrightarrow{-\cap[X]} K^{G,sst}_{*+2d}({\mathbb{C}},X) \xrightarrow{\delta_{X}/-} K^{G,sst}_{*+2d}(X,{\mathbb{C}})$$ coincides with multiplication by $\beta_2^{d}\cup u$ for some unit $u\in K_{0}^{G,\,sst}(X,{\mathbb{C}})$. Similarly there are classes $[X^{an}]\in ku^{G}_{2d}(X^{an})$ and $\delta_{X^{an}}\in ku^{0}_{G}(X^{an})$ such that $$ku_{G}^{*}(X^{an}) \xrightarrow{-\cap[X^{an}]} ku^{G}_{2d-*}(X^{an}) \xrightarrow{\delta_{X^{an}}/-} ku^{G}_{*-2d}(X^{an})$$ coincides with multiplication by $\beta^{d}\cup v$ for some unit $v\in ku_{G}^{0}(X^{an})$. The proof of this theorem will occupy the remainder of this section but first, we prove the main result of this section. \[mainthmbottinv\] Let $G$ be a finite group, and let $X$ be a smooth complex projective $G$-variety of dimension $d$. Then the map of (\[eqn:maps\]) induces an isomorphism $$K^{G,sst}_{*}(X,{\mathbb{C}})[\beta^{-1}_{2}] \stackrel{\cong}{\to} ku_{G}^{-*}(X^{an})[\beta^{-1}_{2}] \stackrel{\cong}{\to} KU_{G}^{-*}(X^{an}).$$ That the right map is isomorphism is the second part of Corollary \[cor:bottinv\]. The argument that the left map is an isomorphism is the same as [@Walker:Thomason Theorem 7.11], in the nonequivariant case. Namely, we consider the diagram $$\xymatrix{ K^{G,sst}_{*}(X,{\mathbb{C}}) \ar[r]^{-\cap[X]}\ar[d] & K^{G,sst}_{*+2d}({\mathbb{C}},X) \ar[r]^{\delta_{X}/-}\ar[d]^{{\cong}} & K^{G,sst}_{*+2d}(X,{\mathbb{C}}) \ar[d] \\ ku_{G}^{-*}(X^{an}) \ar[r]^{-\cap[X^{an}]} & ku^{G}_{*+2d}(X^{an}) \ar[r]^{\delta_{X^{an}}/-} & ku^{G}_{-*-2d}(X^{an}) }$$ which commutes by the compatibility of operations established in the previous section. Using properties of the operations, one can see that the horizontal maps are multiplication by $\delta_{X}/[X]$. Moreover, by Theorem \[walker7.10equi\] they are both multiplication by $\beta_2^{d}$, up to a unit in $K^{G,sst}_{0}(X,{\mathbb{C}})$ (resp.  in $ku_{G}^{0}(X^{an})$). The result follows easily by a simple diagram chase. The remainder of this section is devoted to the proof of Theorem \[walker7.10equi\]. After some important modifications, its proof is similar to the nonequivariant case and we focus our attention on the necessary modifications. First, we recall some facts about equivariant complex orientation and Poincaré duality. These are significantly more complicated in the equivariant setting, but we can simplify things by restricting our attention to those theories which are complex stable. See [@May:equihomotopy Chapter XVI.9] and [@LLM Chapter III.6] for a general and comprehensive treatment of these topics. A useful summary of Poincaré duality for complex stable theories may be found in [@GreenleesWilliams]. Recall that an equivariant cohomology theory $E^{*}_{G}(-)$ is said to be *complex stable* if for each complex representation $V$, there is a class $\sigma_{V}\in \widetilde{E}_{G}^{|V|}(S^{V})$ which gives isomorphisms $$\widetilde{E}^{*}_{G}(S^{|V|}\wedge X) \xrightarrow{{\cong}} \widetilde{E}^{*}_{G}(S^{V}\wedge X),$$ for any $G$-space $X$. As equivariant complex topological $K$-theory satisfies Bott periodicity, it is complex stable. Let $E$ be a commutative ring $G$-spectrum representing a complex stable cohomology theory and $M$ a smooth $G$-manifold. For any $x\in M$ the slice theorem implies that there are isomorphisms $E_{*}^{G}(M,M-G\{x\}) {\cong}\widetilde{E}_{*}^{G}(G_{+}\wedge_{G_{x}}S^{V_{x}})$ and $E^{*}_{G}(M,M-G\{x\}) {\cong}\widetilde{E}^{*}_{G}(G_{+}\wedge_{G_{x}}S^{V_{x}})$ where $G_{x}\subseteq G$ is the isotropy subgroup of $x$ and $V_{x}$ is the tangent space to $M$ at $x$ (see e.g. [@GreenleesWilliams Lemma 3.1]). Let $E$ be as above and $M$ a smooth $G$-manifold of dimension $n$. For an $x\in M$ let $\phi_{G\{x\}}$ denote the composition $$E_{*}^{G}(M)\xrightarrow{} E_{*}^{G}(M,M-G\{x\}) {\cong}\widetilde{E}_{*}^{G}(G_{+}\wedge_{G_{x}} S^{V_{x}}) {\cong}\widetilde{E}^{G_{x}}_{*}(S^{V_{x}}),$$ where the first map is induced by the inclusion of pairs $(M,\emptyset)\subseteq (M,M-G\{x\})$, the second is the isomorphism from the paragraph above and the third is the change of groups isomorphism. An element $[M]\in \widetilde{E}^{G}_{n}(M)$ is called a *fundamental class* for $M$ if $\phi_{G\{x\}}([M])$ is an $\widetilde{E}^{G_{x}}_{*}$-module generator of $\widetilde{E}^{G_{x}}_{*}(S^{V_{x}})$ for all $x\in M$. \[Thomclass\] Let $E$ be as above and $M$ a smooth compact complex $G$-manifold of complex dimension $d$. For any orbit $i_{G\{x\}}:G\{x\}\to M$ let $\psi_{G\{x\}}$ denote the composition $$\begin{gathered} E^{*}_G(M \times M, M \times M - \Delta) \to E^{*}_G(M \times G\{x\},M \times G\{x\} - \Delta(G\{x\}) ) \\ \cong E^{*}_{G}(G_{+}\wedge_{G_{x}}S^{V_{x}}){\cong}E^{*}_{G_{x}}(S^{V_{x}})\end{gathered}$$ where the first map is obtained from the map of pairs induced by $id\times i_{G\{x\}}$. An element $t_M \in E^{2d}_G(M \times M, M \times M - \Delta)$ is called a [*Thom class*]{} for $M$ if $\psi_{G\{x\}}(t_{M})$ is an $E^*_{G_{x}}$-module generator of $E^*_{G_{x}}(S^{V_{x}})$ for all $x\in M$. Let $M$ be a smooth complex compact $G$-manifold of complex dimension $d$ and $E$ a commutative ring $G$-spectrum representing a complex stable cohomology theory. There is a bijection between $E$-Thom classes for $M$ and $E$-fundamental classes for $M$. This is [@LLM Proposition III.6.7]. One needs to observe that the definitions used there agree with the ones used here, as one can see using Remark \[allthomagree\] below. As a result of Bott periodicity and the Thom isomorphism for $KU_{G}^{0}$, any smooth complex compact $G$-manifold $M$ of complex dimension $d$ has a Thom class in $KU^{2d}_{G}(M)$ in the sense above and therefore it has a fundamental class. The map $$\cap[M]:KU^*_G(M) \stackrel{\cong}{\to} KU^G_{2d-*}(M)$$ is an isomorphism by [@LLM III.6.4] (or see [@GreenleesWilliams Theorem 3.6]). This is the *equivariant Poincaré duality* isomorphism. We know that $ku_{G}^{*}(-)$ is in general not complex stable (see e.g. [@Greenlees:equivforms Section 4]). Therefore, unlike in [@Walker:Thomason], we work with periodic rather than with connective equivariant $K$-theory from now on. This is not a problem by the following lemma. \[kuisalmostKU\] Let $W$ be a finite dimensional $G$-$CW$ complex. Then $$ku^{G}_{i}(W) \xrightarrow{{\cong}} KU^{G}_{i}(W)$$ is an isomorphism for $i\geq \dim(W)$. Let $W^{(n)}$ denote the $n$-skeleton of $W$. For each $n$, $W^{(n)}/W^{(n-1)}$ is wedge of spheres of the form $S^{n}\wedge G/H_{+}$. We show that $ku^{G}_{i}(W^{(n)}) \xrightarrow{{\cong}} KU^{G}_{i}(W^{(n)})$ is an isomorphism for $i\geq n$ and an injection of $i = n-1$. The map of $G$-spectra $ku^{G} \to KU^{G}$ induces a comparison of long-exact sequences $$\xymatrix{ \cdots \ar[r] & ku^{G}_{i}(X^{(n)}) \ar[r]\ar[d] & ku^{G}_{i}(X^{(n+1)}) \ar[r]\ar[d] & ku^{G}_{i}(X^{(n+1)}/X^{(n)}) \ar[d]\ar[r] & \cdots \\ \cdots \ar[r] & KU^{G}_{i}(X^{(n)}) \ar[r] & KU^{G}_{i}(X^{(n+1)}) \ar[r] & KU^{G}_{i}(X^{(n+1)}/X^{(n)}) \ar[r] & \cdots . }$$ The right-hand map is a sum of maps of the form $$ku^{G}_{i}(S^{n+1}\wedge G/H_{+}) \to KU^{G}_{i}(S^{n+1}\wedge G/H_{+}).$$ Via the change of groups isomorphism it is identified with $$ku^{H}_{i-n-1}(S^{0}) \to KU^{H}_{i-n-1}(S^{0})$$ which is an isomorphism for $i\geq n+1$ and $ku^{H}_{i-n-1}(S^{0})=0$ otherwise. The lemma allows us to lift fundamental classes to the equivariant homology theory $ku^{G}_{*}$ and thus to the semi-topological equivariant $K$-homology as well. Let $X$ be a smooth projective complex $G$-variety of complex dimension $d$ and $[X^{an}]\in KU^G_{2d}(X^{an})$ a fundamental class. Define classes $[X^{an}]\in ku_{2d}^{G}(X^{an})$ and $[X]\in K^{G,\,sst}_{2d}({\mathbb{C}},X)$ to be lifts of $[X^{an}]\in KU_{2d}(X^{an})$ under the isomorphisms $K^{G,\,sst}_{2d}({\mathbb{C}},X){\cong}ku_{2d}^{G}(X^{an}){\cong}KU^{G}_{2d}(X)$ provided by Theorem \[mainthm\] and Lemma \[kuisalmostKU\]. \[allthomagree\] There are several equivalent descriptions of the Thom space and thus of the cohomology groups in which Thom classes live. Let $p:X \to B$ be a complex vector bundle of rank $n$ with zero section $s$ and let $E^*$ a complex orientable cohomology theory. The Thom space $Th(p)$ is homotopy equivalent to the homotopy cofiber of $X-s(B) \to X$. If $B$ is compact, then the one point compactification $X^+$ is homeomorphic to $Th(p)$ and the Thom isomorphism for topological $K$-theory is often stated using $\widetilde{KU}{}^{*}(X^+)$. In [@Walker:Thomason], Thom classes for $M$ live in the cohomology group $E^*_M(M \times M)$, defined using the homotopy cofiber of $M\times M-\Delta \to M\times M$. This compares to the Thom space of its tangent bundle as the normal bundle of $\Delta:M \to M \times M$ is isomorphic to the tangent bundle of $M$ (see e.g. [@MilnorStasheff Lemma 11.5]). All of these weak equivalences remain valid in the equivariant case, as well. We can to establish an equivariant generalization of [@Walker:Thomason Lemma 7.8], with $ku$ replaced by $KU$. \[deltathom\] Let $X$ be a smooth projective complex $G$-variety of dimension $d$. Let $j^*:KU^{2d}_G(X^{an} \times X^{an}, X^{an} \times X^{an} - \Delta) \to KU^{2d}_G(X^{an} \times X^{an})$ be the map forgetting the support. Then $\delta_{X^{an}}=(\beta_2)^d \cup j^*(t)$ for some Thom class $t \in KU^{2d}_G(X^{an} \times X^{an}, X^{an} \times X^{an} - \Delta)$. We roughly follow Walker’s proof. We claim that the element $\delta_{X^{an}}$ lifts to an element $\tilde{\delta}_{X^{an}}\in KU^{0}_G(X^{an} \times X^{an}, X^{an} \times X^{an} - \Delta)$ such that the restriction of $\tilde{\delta}_{X^{an}}$ to $KU^{0}_G(X^{an}\times G\{x\}, X^{an} - \Delta G\{x\}) =KU^{0}_{G,\Delta G\{x\}}(X\times G\{x\})$ is a generator for any orbit $i_{G\{x\}}: G\{x\} \to X$. Using Bott periodicity, we therefore have that the element $\beta_{2}^{-d}\cup \tilde{\delta}_{X^{an}}\in KU_{G,\Delta}^{2d}(X^{an}\times X^{an})$ is a Thom class in the sense of Definition \[Thomclass\], from which the result follows. To begin we consider algebraic $K$-theory. Consider the following diagram $$\xymatrix{ K_0^{G}(X) \ar[r]^-{\Delta_*} \ar[d]_{(i_x)^*} & K_0^{G}(X \times X) \ar[r] \ar[d]^{(id_X\times i_{G\{x\}})^*} & K_0^{G}(X \times X - \Delta) \ar[d]^{(id_X\times i_{G\{x\}})^*} \\ K_0^{G}(G \{x\}) \ar[r]^-{\Delta'_*} & K_0^{G}(X\times G\{x\}) \ar[r] & K^{G}_0(X\times G\{x\}-\Delta G\{x\}). }$$ The right-hand square is evidently commutative. The commutativity of the left-hand square follows from the commutativity of $$\xymatrix{ K_0^{G_{x}}(X) \ar[r]^-{\Delta_*} \ar[d]_{(i_x)^*} & K_0^{G_{x}}(X \times X) \ar[d]^{(id_X\times i_{x})^*} \\ K_0^{G_{x}}(\{x\}) \ar[r]^-{(i_{x})_*} & K_0^{G_{x}}(X ) , }$$ which can be seen by using that $\pi:X\to \{x\}$ gives a $G_{x}$-equivariant section of $i_{x}$. The rows of the diagram are exact by [@Thomason:algKgroup Corollary 5.8] (that is the equivariant generalization of Quillen’s resolution theorem) and by [@Thomason:algKgroup Theorem 2.7]. The left hand side square induces a commutative square $$\xymatrix{ K_0^{G}(X) \ar[r]^-{\Delta_*}_-{{\cong}} \ar[d]_{(i_x)^*} & K_{0,\Delta}^{G}(X \times X) \ar[d]^{(id_X\times i_{G\{x\}})^*} \\ K_0^{G}(G \{x\}) \ar[r]^-{\Delta'_*}_-{{\cong}} & K_{0,\Delta G \{x\}}^{G}(X\times G\{x\}). }$$ Write $\tilde{\delta}_X=\Delta_*(1) \in K_{0,\Delta}^{G}(X \times X)$, which evidently maps to the element $\delta_X=\Delta_*(1) \in K_{0}^{G}(X \times X)$. Moreover, we have $(i_x)^*(1)=1 \in K_0^{G}(G \{x\})$ as $(i_x)^*$ is a ring homomorphism. We conclude that $(id_{X}\times i_{G\{x\}})^*(\tilde{\delta}_{X})=\Delta'_*(1)$. Now consider the natural transformation $\epsilon:K_{0}^{G}(-)\to KU_{G}^{0}(-)$ obtained as the composite $$K_{0}^{G}(X)\to K_{0}^{G,\,sst}(X)\to KU_{G}^{0}(X^{an})$$ of the natural transformations from Theorem \[thm:algeq\] and Corollary \[cor:qtopeq\]. The transformation $\epsilon$ is compatible with pullbacks, and it is a ring homomorphism. It is also compatible with pushforwards in both theories (where the push-forward in $K_{0}^{G}$ is the classical push-forward in (equivariant) algebraic $K$-theory and in $KU_{G}^{0}$ it is the one of Definition \[transfer\]) as a direct consequence of Definition \[transfer\]. We therefore conclude that we have $(id_{X^{an}}\times i_{G\{x\}})^*(\tilde{\delta}_{X^{an}})=\Delta'_*(1)$ in $KU^{0}_{G,\Delta G \{x\}}(X\times G\{x\})$. Now since $KU^{0}_{G,\Delta G \{x\}}(X\times G\{x\})$ is a free $KU^{G}_{G}(G\{x\})$-module of rank 1 and $\Delta'_*$ is a $KU^{G}_{G}(G\{x\})$-module homomorphism we conclude that $(id_{X^{an}}\times i_{G\{x\}})^*(\tilde{\delta}_{X^{an}})$ generates $KU^{0}_{G,\Delta G \{x\}}(X\times G\{x\})$ as required. *Proof of Theorem \[walker7.10equi\].* The proof proceeds as in [@Walker:Thomason Theorem 7.10]. Namely the argument there shows that it suffices to show that $\delta_{X}/[X] = \beta^{d}\cup u$ and $\delta_{X^{an}}/[X^{an}] = \beta^{d}\cup v$ for units $u$ and $v$. The second equation follows from Lemma \[deltathom\]. Using Lemma \[deltathom\] and the results about equivariant pairings from the previous section [@Walker:Thomason Proposition 7.9] generalizes to the equivariant setting, showing that $\pi^{*}(\delta_{Y}/a) =\delta_{X}/\pi^{*}(a)$ for a finite flat equivariant morphism $\pi:X\to Y$ of smooth projective $G$-varieties. It thus suffices to establish the result for $Y={\mathbb{P}}^{d}_{{\mathbb{C}}}$, because we can always find a finite, surjective (and hence flat) equivariant map $\pi:X\to {\mathbb{P}}^{d}_{{\mathbb{C}}}$. To see this, note that the quotient $X\to X/G$ is finite, surjective and equivariant and we may assume $X/G$ is connected so that Noether normalization yields a finite, surjective equivariant map $X/G\to {\mathbb{P}}^{d}_{{\mathbb{C}}}$. It thus remains to see that $\delta_{{\mathbb{P}}^{d}}/[{\mathbb{P}}^{d}] = \beta^{d}\cup u'$, for a unit $u'\in K_{0}^{G,\,sst}({\mathbb{P}}^{d}_{{\mathbb{C}}},{\mathbb{C}})$. This follows from seeing that we have an isomorphism $K_{*}^{G,\,sst}({\mathbb{P}}^{d}_{{\mathbb{C}}},{\mathbb{C}}) {\cong}{\mathbf{bu}}_{* }^{G,\,{\mathfrak{c}}}({\mathbb{P}}^{d}_{{\mathbb{C}}},S^{0})$. To see this we can argue as in [@FW:real Proposition 2.7] to see that we have an equivariant equivalence ${\mathcal{K}}_{G}(\Delta^{\bullet}_{top},{\mathbb{C}})^{n}{\simeq}{\mathcal{K}}_{G}({\mathbb{P}}^{1}_{{\mathbb{C}}}\times\Delta^{\bullet}_{top},{\mathbb{C}})$ and a similar compatible equivalence for ${\mathbf{bu}}^{{\mathfrak{c}}}_{G}$. The required isomorphism follows since ${\mathcal{K}}_{G}(\Delta^{\bullet}_{top},{\mathbb{C}}) {\simeq}{\mathbf{bu}}^{{\mathfrak{c}}}_{G}(S^{0},S^{0})$. Equivariant Thomason’s theorem {#sec:algThom} ============================== In this section, we explain how the work in the previous section gives an alternate proof of [@Thomason:famousequi Theorem 5.9]. This requires proving the expected comparison theorem between algebraic and semi-topological equivariant $K$-theory with finite coefficients. Working with the ${\Delta^{\bullet}_{top}}$-construction rather than with the topological mapping spaces $Mor(-,-)$ makes this particularly straightforward. Let $\mathcal{K}_G(-,-)$ be the bivariant presheaf on quasi-projective complex $G$-varieties with values in positive $\Omega$-$G$-spectra produced in Section \[sec:algK\]. By Proposition \[prop:fix\], we have that $$\pi_{n}^{H}\mathcal{K}_G(X,Y) {\cong}\pi_{n}\mathcal{K}(H;X\times\Delta^{\bullet}_{{\mathbb{C}}},Y)$$ for any subgroup $H < G$ and any $n$, where $\mathcal{K}(H;X,Y)$ is the $K$-theory spectrum of the category of coherent $H$-modules on $X\times Y$ which are finite and flat over $X$. The following result, whose proof is as in [@FW:ratisos Theorem 2.6], allows us to apply the work of the previous section to equivariant algebraic $K$-theory. Unlike Theorem \[mainthm\], the proof of the following theorem is rather formal and applies to equivariant theories other than ${\mathcal{K}}_{G}$, provided they satisfy an appropriate equivariant rigidity theorem. \[KalgKsemi\] For a smooth quasi-projective complex $G$-variety $X$ and any integer $n>0$, the maps (\[eqn:algsst\]) induce equivariant weak equivalences $${\mathcal{K}}_G(X,{\mathbb{Z}}/n) \xrightarrow{{\simeq}}{\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top};{\mathbb{Z}}/n) \xleftarrow{{\simeq}} {\mathcal{K}}^{sst}_G(X,{\mathbb{Z}}/n)$$ of $G$-spectra. The right hand map is an equivariant weak equivalence by Lemma \[lem:delsst\]. To show the left hand map is an equivariant weak equivalence, we must show that for all subgroups $H \subseteq G$, the map $\pi_*^{H}{\mathcal{K}}_{G}(X,{\mathbb{Z}}/n)\to \pi_{*}^{H}{\mathcal{K}}_{G}(X\times\Delta^{\bullet}_{top},{\mathbb{Z}}/n)$ is an isomorphism. Using Proposition \[prop:fix\], the proof follows along the lines of the argument given in [@FW:ratisos Theorem 2.8]. That is, we consider the map of presheaves ${{\mathcal{F}}}:={\mathcal{K}}(H;X,{\mathbb{Z}}/n) \to {\mathcal{G}}:={\mathcal{K}}(H;X \times - ,{\mathbb{Z}}/n)$ where the first presheaf is globally constant. By [@YO:equirigid] the map $$\pi_{n}{\mathcal{K}}(H;X,{\mathbb{Z}}/n) \xrightarrow{{\cong}} \pi_{n}{\mathcal{K}}(H;X \times {\mathcal{O}}_{T,t}^{h} ,{\mathbb{Z}}/n)$$ is an isomorphism, where $T$ is any smooth variety, $t\in T({\mathbb{C}})$, and ${\mathcal{O}}_{T,t}^{h}$ is the corresponding Henselian local ring. This rigidity isomorphism allows us to conclude the result as in the proof of [@FW:ratisos Theorem 2.8]. That is, the map of presheaves $\pi_{n}{\mathcal{F}}(-) \to \pi_{n}{\mathcal{G}}(-)$ becomes an isomorphism, upon sheafification, of étale sheaves on $Sm/{\mathbb{C}}$. It therefore becomes an isomorphism upon further sheafification, of sheaves in the $uad$-topology on $Sch/{\mathbb{C}}$ because resolutions of singularities are $uad$-covers. Thus one may apply [@FW:ratisos Theorem 2.6] in order to conclude the result. Using the preceding theorem, we may lift the element $\beta_2\in K_{2}^{G,\,sst}({\mathbb{C}},{\mathbb{Z}}/n)$ to an element $\beta$ in algebraic $K$-theory with finite coefficients. Proposition \[prop:cup\] implies that the isomorphism $K_{*}^{G}(X,{\mathbb{Z}}/n) {\cong}K_{*}^{G,\,sst}(X,{\mathbb{Z}}/n)$ from the previous theorem is a graded ring isomorphism and therefore $$K_*^G(X,{\mathbb{Z}}/n)[\beta^{-1}] {\cong}K_*^{G,sst}(X,{\mathbb{Z}}/n)[\beta_{2}^{-1}].$$ keeping in mind the usual warning (see e.g. [@Thomason:famous A.6]) concerning very small values of $n$. Also note that by [@Thomason:famous p. 503], the algebraic Bott element $\beta$ Thomason considers really is a lift of the topological one. Combining this theorem with the one of the previous section, we obtain a new proof of [@Thomason:famousequi Theorem 5.9] for finite groups: \[thm:lastapp\] For any smooth projective complex $G$-variety $X$ and any integer $n>0$, we have a natural isomorphism of graded rings $$K_*^G(X,{\mathbb{Z}}/n)[\beta^{-1}] \stackrel{{\cong}}{\to}KU_G^{-*}(X^{an},{\mathbb{Z}}/n).$$ This follows immediately from Theorem \[mainthmbottinv\] and Theorem \[KalgKsemi\]. [LMSM86]{} Andrew J. Blumberg, *Continuous functors as a model for the equivariant stable homotopy category*, Algebr. Geom. Topol. **6** (2006), 2257–2295. [MR ]{}[2286026 (2008a:55006)]{} Daniel Dugger and Daniel C. Isaksen, *Topological hypercovers and [$\mathbb A\sp 1$]{}-realizations*, Math. Z. **246** (2004), no. 4, 667–689. [MR ]{}[MR2045835 (2005d:55026)]{} Eric M. Friedlander and Mark E. Walker, *Comparing [$K$]{}-theories for complex varieties*, Amer. J. Math. **123** (2001), no. 5, 779–810. [MR ]{}[MR1854111 (2002i:19004)]{} [to3em]{}, *Semi-topological [$K$]{}-theory of real varieties*, Algebra, arithmetic and geometry, Part I, II (Mumbai, 2000), Tata Inst. Fund. Res. Stud. Math., vol. 16, Tata Inst. Fund. Res., Bombay, 2002, pp. 219–326. [MR ]{}[MR1940670 (2003h:19005)]{} [to3em]{}, *Semi-topological [$K$]{}-theory using function complexes*, Topology **41** (2002), no. 3, 591–644. [MR ]{}[MR1910042 (2003g:19005)]{} [to3em]{}, *Rational isomorphisms between [$K$]{}-theories and cohomology theories*, Invent. Math. **154** (2003), no. 1, 1–61. [MR ]{}[MR2004456 (2004j:19002)]{} J. P. C. Greenlees, *Equivariant forms of connective [$K$]{}-theory*, Topology **38** (1999), no. 5, 1075–1092. [MR ]{}[1688426 (2000i:55025)]{} Daniel R. Grayson and Mark E. Walker, *Geometric models for algebraic [$K$]{}-theory*, $K$-Theory **20** (2000), no. 4, 311–330, Special issues dedicated to Daniel Quillen on the occasion of his sixtieth birthday, Part IV. [MR ]{}[1803641 (2001m:19006)]{} J. P. C. Greenlees and G. R. Williams, *Poincaré duality for [$K$]{}-theory of equivariant complex projective spaces*, Glasg. Math. J. **50** (2008), no. 1, 111–127. [MR ]{}[2381737 (2008m:55007)]{} Philip S. Hirschhorn, *Model categories and their localizations*, Mathematical Surveys and Monographs, vol. 99, American Mathematical Society, Providence, RI, 2003. [MR ]{}[1944041 (2003j:18018)]{} L. G. Lewis, Jr., J. P. May, M. Steinberger, and J. E. McClure, *Equivariant stable homotopy theory*, Lecture Notes in Mathematics, vol. 1213, Springer-Verlag, Berlin, 1986, With contributions by J. E. McClure. [MR ]{}[866482 (88e:55002)]{} J. P. May, *Equivariant homotopy and cohomology theory*, CBMS Regional Conference Series in Mathematics, vol. 91, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 1996, With contributions by M. Cole, G. Comeza[ñ]{}a, S. Costenoble, A. D. Elmendorf, J. P. C. Greenlees, L. G. Lewis, Jr., R. J. Piacenza, G. Triantafillou, and S. Waner. [MR ]{}[1413302 (97k:55016)]{} M. A. Mandell and J. P. May, *Equivariant orthogonal spectra and [$S$]{}-modules*, Mem. Amer. Math. Soc. **159** (2002), no. 755, x+108. [MR ]{}[1922205 (2003i:55012)]{} John W. Milnor and James D. Stasheff, *Characteristic classes*, Princeton University Press, Princeton, N. J., 1974, Annals of Mathematics Studies, No. 76. [MR ]{}[0440554 (55 \#13428)]{} Paul Arne [Ø]{}stv[æ]{}r, *Appendix to “[E]{}quivariant [T]{}homason’s theorem and semi-topological [$K$]{}-homology”*, Preprint, 2012. Daniel Quillen, *Higher algebraic [$K$]{}-theory. [I]{}*, Algebraic [$K$]{}-theory, [I]{}: [H]{}igher [$K$]{}-theories ([P]{}roc. [C]{}onf., [B]{}attelle [M]{}emorial [I]{}nst., [S]{}eattle, [W]{}ash., 1972), Springer, Berlin, 1973, pp. 85–147. Lecture Notes in Math., Vol. 341. [MR ]{}[0338129 (49 \#2895)]{} Graeme Segal, *Equivariant [$K$]{}-theory*, Inst. Hautes Études Sci. Publ. Math. (1968), no. 34, 129–151. [MR ]{}[0234452 (38 \#2769)]{} [to3em]{}, *Categories and cohomology theories*, Topology **13** (1974), 293–312. [MR ]{}[MR0353298 (50 \#5782)]{} Kazuhisa Shimakawa, *Infinite loop [$G$]{}-spaces associated to monoidal [$G$]{}-graded categories*, Publ. Res. Inst. Math. Sci. **25** (1989), no. 2, 239–262. [MR ]{}[1003787 (90k:55017)]{} [to3em]{}, *A note on [$\Gamma_G$]{}-spaces*, Osaka J. Math. **28** (1991), no. 2, 223–228. [MR ]{}[1132161 (92j:55015)]{} Brooke Shipley, *Symmetric spectra and topological [H]{}ochschild homology*, $K$-Theory **19** (2000), no. 2, 155–183. [MR ]{}[1740756 (2001h:55010)]{} R. W. Thomason, *Algebraic [$K$]{}-theory and étale cohomology*, Ann. Sci. École Norm. Sup. (4) **18** (1985), no. 3, 437–552. [MR ]{}[826102 (87k:14016)]{} [to3em]{}, *Algebraic [$K$]{}-theory of group scheme actions*, Algebraic topology and algebraic [$K$]{}-theory ([P]{}rinceton, [N]{}.[J]{}., 1983), Ann. of Math. Stud., vol. 113, Princeton Univ. Press, Princeton, NJ, 1987, pp. 539–563. [MR ]{}[921490 (89c:18016)]{} [to3em]{}, *Equivariant algebraic vs. topological [$K$]{}-homology [A]{}tiyah-[S]{}egal-style*, Duke Math. J. **56** (1988), no. 3, 589–636. [MR ]{}[948534 (89f:14015)]{} Mark E. Walker, *Adams operations for bivariant [$K$]{}-theory and a filtration using projective lines*, $K$-Theory **21** (2000), no. 2, 101–140. [MR ]{}[1804538 (2002i:19005)]{} [to3em]{}, *Semi-topological [$K$]{}-homology and [T]{}homason’s theorem*, $K$-Theory **26** (2002), no. 3, 207–286. [MR ]{}[MR1936355 (2003k:19004)]{} [to3em]{}, *Thomason’s theorem for varieties over algebraically closed fields*, Trans. Amer. Math. Soc. **356** (2004), no. 7, 2569–2648. [MR ]{}[2052190 (2005d:19003)]{} Serge Yagunov and Paul Arne [Ø]{}stv[æ]{}r, *Rigidity for equivariant [$K$]{}-theory*, C. R. Math. Acad. Sci. Paris **347** (2009), no. 23-24, 1403–1407. [MR ]{}[2588790 (2010k:19010)]{}
{ "pile_set_name": "ArXiv" }
--- author: - 'I. KLICH' subtitle: 'An elementary derivation of Levitov’s formula' title: FULL COUNTING STATISTICS --- \#1[(\[\#1\])]{} The field of quantum noise has been rapidly developing in recent years, with the growing possibilities in precision measurements [@BlanterReview], and interest in mesoscopic systems as well as in technological applications of physical effects at the micrometer and nanometer scales. Of particular interest is the study of the statistics of charge transport between materials coupled through a contact or through a time dependent scatterer. The full statistics of charge transport was studied in a series of works by Levitov et al. [@Levitov; @Ivanov]. This approach yielded interesting results, and in particular, they where able to express the full counting statistics in terms of a determinant of a single particle operator. Several aspects of Levitov’s formula where discussed in following papers [@Andreev; @Nazarov]. Our aim in this paper is to present a novel derivation of the original Levitov formula. This is done by proving a trace formula , which relates certain traces in Fock space to single particle determinants. Using the present approach we find in addition several generalizations, such as a corresponding formula for Bosons. The full counting statistics ============================ The typical setting is the following: consider particle reservoirs, with given temperatures and chemical potentials, which are separated at time zero, and are evolving by a second quantized hamiltonian $H_0$. At some time the reservoirs are coupled, through a scattering region, and evolve by a new, time dependent hamiltonian. After a time $T$ they are decoupled again and one is interested in the statistics of charge transported from side to side, i.e. to compute $<Q>,<Q^2>$ and higher moments. The role of the third moment was recently discussed in [@LevitovReznikov]. The full statistics of charge transfer may be conveniently represented by the characteristic function of the (charge transport) probability distribution function, defined by: $$\chi(\lambda_1,...,;T)=\sum_{{\bf \alpha},{\bf \beta}}P({\bf \alpha}(t=0),{\bf \beta}(t=T))e^{iq\sum_{i}\lambda_i(\beta_i-\alpha_i)}$$ Here the summation is over all states ${\bf \alpha}=(\alpha_1,...),{\bf \beta}=(\beta_1,...)$ labelling the Fock space in the occupation number representation: for fermions these are vectors of zeros and ones and for bosons vectors with integer coefficients, where $\alpha_i$ is the number of particles occupying the single particle state $i$ in ${\bf\alpha}$. $P({\bf \alpha}(t=0),{\bf \beta}(t=T))$ is the probability that we started in state ${\bf \alpha}$ at time $0$ and finished in a state ${\bf \beta}$ at time $T$. Thus terms $(\alpha_i-\beta_i)$ appearing in the exponent are just the change in the number of particles occupying the single particle state $i$. And the parameters $\lambda_i$ are introduced in the standard manner to calculate different moments. By taking derivatives of $\chi$ with respect to $\lambda_i$, one can calculate arbitrary moments of the charge accumulation in state $i$. For example, $$<Q_i>=-i\partial_{\lambda_i}\log\chi|_{\lambda_1,...=0}$$ And $$(\Delta Q_i)^2=<Q_i^2-<Q_i>^2>=-\partial_{\lambda_i}^2\log\chi|_{\lambda_1,...=0}.$$ It is also possible to compute the moments of the charge which is transferred to a particular reservoir by taking the derivative with respect to $\lambda$ after setting $\lambda_i=\lambda$ for all states $i$ belonging to the desired reservoir. $\chi(\lambda)$ may be interpreted as the reaction of the system to coupling with a classical field $\lambda$ which measures the number of electrons on each side. For adiabatic change, and short scattering time Levitov et al [@Levitov], obtained the following expression for $\chi$: $$\begin{aligned} \label{LevitovForm} \chi({\lambda})=\det(1+n(S^{\dag}e^{iq\lambda}Se^{-iq\lambda}-1))\end{aligned}$$ Where $n$ is the occupation number operator and $S$ is the scattering matrix. As was remarked [@Levitov], this expression requires careful understanding and regularization. In the following we derive this formula in a new manner which, we hope will allow a convenient way to address these issues. In order to proceed we first write $\chi$ as a trace in Fock space: $$\begin{aligned} \label{chitr} & &\chi({\lambda},T) =\\ \nonumber & &\sum_{{\bf \alpha},{\bf \beta}}<{\bf \alpha}|\rho_0|{\bf \alpha}>|<{\bf \alpha}|{\mathbb U}^{\dag}|{\bf \beta}>|^2e^{iq\sum_i\lambda_i(\beta_i-\alpha_i)}\\ \nonumber & & ={\rm Tr}(\rho_0 {\mathbb U}^{\dag}e^{iq\sum\lambda_i a^{\dag}_i a_i}{\mathbb U}e^{-iq\sum\lambda_i a^{\dag}_i a_i})\end{aligned}$$ Where $\rho_0$ is the density matrix at the initial time ($t=0$), ${\mathbb U}$ is the evolution (in Fock space) from time $t=0$ to time $t=T$ and $a^{\dag}_i ,a_i$ are the creation and annihilation operators for a given one particle state $i$. Here it is assumed that the occupation number basis is chosen such that the initial time density matrix $\rho_0$ is diagonal in it, which implies that the states $\bf \alpha$ are eigenstates of the initial Hamiltonian, and measurement of charge in a specific state is meaningful. Next, we define the second quantized version of a single particle operator $A$ (i.e. an operator on the single particle Hilbert space) to be the Fock space operator: $$\label{secondquantized} \Gamma(A)=\sum <i|A|j>a^{\dag}_i a_j.$$ Then can be written as: $$\begin{aligned} \chi({\lambda},T)={\rm Tr}(\rho_0 e^{iq{\mathbb U}^{\dag}\Gamma ({\bf \lambda}){\mathbb U}}e^{-iq\Gamma({\bf \lambda})})\end{aligned}$$ Here ${\bf \lambda}$ is the matrix ${\rm diag}(\lambda_1,\lambda_2,...)$. To handle this kind of expressions (and to obtain Levitov’s formula) we prove in the following section a trace formula. A trace formula =============== In this section we prove the following: $$\label{formula} {\rm Tr}( e^{\Gamma(A)}e^{\Gamma(B)})={\rm det}(1-\xi e^{A}e^{B})^{-\xi}$$ Where $\xi=1$ for bosons and $\xi=-1$ for fermions (i.e. the creation and annihilation operators satisfy $a_j a^{\dag}_i-\xi a^{\dag}_i a_j=\delta_{ij}$). We prove this result for the finite dimensional Hilbert space case, and avoid at this point questions regarding the limit of infinite number of states, to be addressed elsewhere [@AK].\ [*Proof*]{}:\ For an $N$ dimensional single particle Hilbert space $\Gamma$ is a representation of the usual Lie algebra of matrices $gl(N)$. Indeed, substituting the definition , together with the relations obeyed by the creation and annihilation operators it is straightforward to check that $$\begin{aligned} \label{representation} & [\Gamma(A),\Gamma(B)]=\Gamma([A,B])\end{aligned}$$ is true for bosons and for fermions. By Baker Campbell Hausdorf there exists a matrix $C$ such that $e^A e^B=e^C$. $C$ is an element of $gl(N)$ and is given by a series of commutators, since $\Gamma$ is a representation, it holds that $$e^A e^B=e^C\rightarrow e^{\Gamma(A)}e^{\Gamma(B)}=e^{\Gamma(C)}.$$ Now let us evaluate ${\rm Tr}( e^{\Gamma(C)})$. Any matrix $C$ can be written in a basis in which it is of the form $\rm diag(\mu_1,..\mu_n)+K$ where $K$ is an upper triangular, thus we have $$\begin{aligned} & {\rm Tr}( e^{\Gamma(C)})= {\rm Tr}( e^{\Gamma({\rm diag}(\mu_1,..,\mu_n))+\Gamma(K)})={\rm Tr}( e^{\Gamma({\rm diag}(\mu_1,..,\mu_n))})=\\ \nonumber & {\rm Tr}( \prod_i e^{\mu_i a^{\dag}_ia_i})=\prod_i (1-\xi e^{\mu_i})^{-\xi}={\rm det}(1-\xi e^{C})^{-\xi}\end{aligned}$$ (One may also think of ${\rm Tr}( e^{\Gamma(C)})$ as the partition function of a system with Hamiltonian $-C$ at temperature $k_B T=1$). From this equation follows: $${\rm Tr}( e^{\Gamma(A)}e^{\Gamma(B)})={\rm Tr}( e^{\Gamma(C)})={\rm det}(1-\xi e^{C})^{-\xi}={\rm det}(1-\xi e^{A}e^{B})^{-\xi}$$ We remark at this point that this relation can immediately be generalized in the same way to products of more then two operators.\ $\bullet$ Let us illuminate our identity with a trivial example: Let ${\cal H}$ be an $N$ - dimensional Hilbert space, and choose $A=B=0$. Then the dimension of the appropriate Fock space is given by\ ${{\rm Tr}}({\mathbb{I}})={{\rm Tr}}( e^{\Gamma(0)}e^{\Gamma(0)})={\rm det}(1-\xi)^{-\xi}=\Big\{\begin{array}{lll} & 2^N & {\rm Femions} \\ \nonumber & \infty & {\rm Bosons}\end{array}$\ as it should be. Levitov’s formula ================= We now turn to give a novel derivation of Levitov’s result for the full counting statistics. In the framework of non interacting fermions the evolution ${\mathbb U}$ in the expression for $\chi$ is just the Fock space implementation of the single particle evolution $U$. That means that ${\mathbb U}^{\dag}\Gamma (\lambda){\mathbb U}=\Gamma ( U^{\dag}\lambda U)$ so that by the trace formula for 3 operators, we immediately have $$\begin{aligned} \label{chieq} & & \chi({\lambda},T)={\rm Tr}({e^{-\beta \Gamma(H_0)}\over Z} e^{iq\Gamma ( U^{\dag}\lambda U)}e^{-iq\Gamma({\bf \lambda})})\\ \nonumber & & {1\over Z}\det(1+e^{-\beta H_0} (U^{\dag}e^{iq\lambda}Ue^{-iq\lambda}))=\\ \nonumber & & \det(1+n (U^{\dag}e^{iq\lambda}Ue^{-iq\lambda}-1))\end{aligned}$$ Where $Z=\det(1+e^{-\beta H_0})$ and $n$ is the occupation number operator ${e^{-\beta H_0}\over 1+e^{-\beta H_0}}$ at the initial time. We note that the result should be viewed as the general expression for the counting statistics of noninteracting fermions, at any given time, and without any approximation. And may be a good start for studying different limits of the problem, as well as regularization difficulties. Finally, if the scattering time is very small compared with the entire evolution, then one may describe the problem in terms of dynamical scattering operators, $S=\lim_{t\rightarrow\infty} e^{iH_0 t} U(t,-t)e^{iH_0 t}$ where $H_0$ is the initial free evolution. Using the fact that $\lambda$ commutes with $H_0$, one obtains in the limit of $T\rightarrow\infty$: $$\begin{aligned} \chi({\lambda})=\det(1+n(S^{\dag}e^{iq\lambda}Se^{-iq\lambda}-1))\end{aligned}$$ Which is Levitov’s result , as promised.\ We now add a few remarks: 1\. [*Convergence and regularization:*]{} First we note that as long as we assume that ${\mathbb \rho}$ exists and has trace 1, then trace of $\rho$ times a bounded operator is also finite, so that $\chi$ is well defined. However, problems might arise when taking the thermodynamic limit. Taking the infinite volume limit may cause the Fock space density matrix to be ill defined (i.e it cannot be normalized to trace 1), however, expectation values obtained using it may still have meaning. If one uses the scattering matrix as in , regularization of the determinant is needed, since in this case one uses the static scattering matrix as an approximation for the true evolution. Indeed, in the limit $T\rightarrow\infty$ arbitrarily large charges can pass from side to side [^1], so that the information about the length of the time interval has to be put in by hand [@Levitov]. The equation , however, can be shown to be well defined even in the thermodynamic limit, for a finite time interval [@AK]. 2\. [*Bosons:*]{} It is now straightforward to derive an analogous formula for the full counting statistics of bosons. The result is simply: $$\begin{aligned} \chi_{B}({\lambda},T)={1\over\det(1-n_B(U^{\dag}e^{iq\lambda}Ue^{-iq\lambda}-1))}.\end{aligned}$$ Where $n_B$ is the occupation number operator for bosons. 3\. [*Rate of charge accumulation:*]{} Here we give an example of how one may compute the rate of charge accumulation in a box $A$. We choose a box in space, which can be described by a projection $P_A$ in the single particle Hilbert space (with matrix elements $<x|P_A|x'>=\delta(x-x')$ if $x$ is in the box, and zero otherwise). By setting $\lambda=0$ for all states that are outside the box $A$, one finds $$\begin{aligned} & &\dot{Q_A}=-i\partial_t\partial_{\lambda}\log\chi({\lambda},t)|_{\lambda=0}=\\ \nonumber & & q\partial_t{{\rm Tr}}(n(U^{\dag}P_A U-P_A))=q{{\rm Tr}}(n(\dot{U}^{\dag}P_A U+U^{\dag}P_A \dot{U}))=\\ \nonumber & & q\,{\rm Re}<U\dot{U}^{\dag}P_A>_t\end{aligned}$$ The angular brackets describe averaging over the distribution at the time of measurement $t$. Charge accumulation is equivalent to current if the box $A$ is connected via just one contact to the other reservoirs. This equation should be compared to the formula for the current in terms of scattering matrices [@BPT; @Brouwer; @Avron; @Makhlin] which is of fundamental interest in the field of quantum pumps. summary ======= To conclude, we presented a novel derivation of Levitov’s determinant formula for the full counting statistics of charge transfer. This was done by introducing a trace formula which is suitable for translating problems of non-interacting particles from Fock space to the single particle Hilbert space. The derivation is general enough to allow consideration of new problems of counting statistics, in particular, further problems involving bosons, or measurement of other operators then charge. We hope that some properties of the determinant under various limits, such as adiabatic and thermodynamic limits will now be easier to address.\ \ I am grateful to J. E. Avron for many discussions and remarks, and especially to O. Kenneth for his help in the proof of . I would also like to thank J. Feinberg, L. S. Levitov and M. Reznikov for remarks.\ \ Ya. M. Blanter and M. Büttiker, Phys. Rep. [**336**]{}, 1 (2000). L. S. Levitov and G. B. Lesovik, JETP Lett. [**58**]{}, 230 (1993); L. S. Levitov, H.-W. Lee, and G. B. Lesovik, Journal of Mathematical Physics, [**37**]{} (1996) 10. D. A. Ivanov, H. W. Lee and L. S. Levitov, Phys. Rev. B [**56**]{},6839 (1997). A. V. Andreev and A. Kamenev, Phys. Rev. Lett. [**85**]{}, 1294 (2000). Yu. V. Nazarov and M. Kindermann, cond-mat/0107133. L. S. Levitov and M. Reznikov, cond-mat/0111057. J. E. Avron and I. Klich, unpublished. M. Büttiker, H. Thomas, A. Prêtre, Z. Phys. B [**94**]{}, 133 (1994). P.W. Brouwer, Phys. Rev. B [**58**]{}, 10135 (1998). J.E. Avron, A. Elgart, G.M. Graf and L. Sadun, Phys. Rev. Lett. [**87**]{}, 236601 (2001) Y. Makhlin and A. D. Mirlin Phys. Rev. Lett. [**87**]{}, 276803 (2001). [^1]: To see this we note that the first moment of , which is the transported charge, diverges as time goes to infinity if there is a bias between the reservoirs
{ "pile_set_name": "ArXiv" }
--- abstract: 'Data confidentiality is an important requirement for clients when outsourcing databases to the cloud. Trusted execution environments, such as Intel SGX, offer an efficient, hardware-based solution to this cryptographic problem. Existing solutions are not optimized for column-oriented, in-memory databases and pose impractical memory requirements on the enclave. We present EncDBDB, a novel approach for client-controlled encryption of a column-oriented, in-memory databases allowing range searches using an enclave. EncDBDB offers nine encrypted dictionaries, which provide different security, performance and storage efficiency tradeoffs for the data. It is especially suited for complex, read-oriented, analytic queries, e.g., as present in data warehouses. The computational overhead compared to plaintext processing is within a millisecond even for databases with millions of entries and the leakage is limited. Compressed encrypted data requires less space than a corresponding plaintext column. Furthermore, the resulting code — and data — in the enclave is very small reducing the potential for security-relevant implementation errors and side-channel leakages.' author: - | Benny Fuhry\ \ \ Jayanth Jain H A\ \ \ Florian Kerschbaum\ \ \ bibliography: - 'ms.bib' title: 'EncDBDB: Searchable Encrypted, Fast, Compressed, In-Memory Database using Enclaves' --- Introduction {#sec:02:Introducton} ============ Data warehouses are used by companies for business intelligence and decision support. Such warehouses contain large datasets and the underlying database management systems (DBMS) are optimized for complex, read-oriented, analytic queries. Outsourcing the data and query processing to the cloud, more specifically to a Database-as-a-Service (DBaaS) provider, can reduce costs, minimize maintenance efforts and increase availability. However, companies are reluctant to outsource their sensitive data to an untrusted DBaaS provider due to possible data leakage, government intrusion, and legal liability. Cryptographic solutions can be a building block for an encrypted cloud database. For instance, fully homomorphic encryption (FHE) [@gentry_fully_2009] supports arbitrary computations on encrypted data, but is still too slow for practical deployability [@coron_fully_2011; @gentry_homomorphic_2012]. CryptDB [@popa_cryptdb:_2011] and Monomi [@Monomi] use multiple encryption schemes, e.g., probabilistic encryption, deterministic encryption, and order-preserving encryption [@agrawal_order_2004; @boldyreva_ope_2009; @boldyreva_ope_2011; @kerschbaum_optimal_2014] to perform different database functionalities. The encryption schemes are layered and/or stored in parallel, introducing a storage overhead, and careful query rewriting is necessary to receive a result securely and efficiently. An alternative approach is to build an encrypted database based on a trusted execution environment (TEE). TEEs provide an isolated, trusted environment for application code and data known as an *enclave*. Intel SGX [@Intel_SGX3; @costanintel; @Intel_SGX2; @SGX_Ref; @iscaTut; @Intel_SGX1], a TEE that is integrated into (most) modern Intel CPUs, sparked a new wave of research in the direction of TEE-based encrypted databases [@baumann_shielding_2014; @fuhry2017hardidx; @gribov2017stealthdb; @EnclaveDB; @ObliDB18]. SGX enclaves provide isolation against any other code, e.g., application code, other enclaves, and the OS. However, current TEE approaches assume an unrealistic size of enclaves [@baumann_shielding_2014; @EnclaveDB], do not provide DBMS functionality [@fuhry2017hardidx], do not support persistency [@ObliDB18], or leak the result of every primitive operation [@gribov2017stealthdb]. Also, these solutions do not consider data compression to reduce the size of large databases. We propose and implement EncDBDB, a high-performance, encrypted cloud database supporting analytic queries on large datasets. We focus on a complex, required query type: range queries. However, it is straightforward to also support e.g., count, aggregation, and average calculations. EncDBDB is based on a column-oriented, dictionary encoding based, in-memory database. Column-oriented data storage optimizes the processing of analytic workloads [@DBS-024; @boncz1999mil; @copeland1985decomposition; @stonebraker2005c], in-memory processing boosts the overall performance [@dewitt1984implementation; @garcia1992main; @larson2016modern], and dictionary encoding reduces the storage space overhead of large (encrypted) datasets [@abadi2006integrating; @Willhalm]. Table \[tab:02:teeDatabases\] compares our approach to the most relevant related work. \[tab:02:teeDatabases\] ---------------------------------- ------------------------- ----------------------------------------------------- ------------- ----------------------- --------------------------- -------- \[3\][\*]{}[Approach]{} \[3\][\*]{}[Workload]{} \[3\][\*]{}[Protection Object]{} \[3\][\*]{} \[3\][\*]{}[LOC]{} (lr)[5-6]{} Storage Performance EnclaveDB [@EnclaveDB] OLTP in-memory storage and query engine N/A $>\SI{20}{\percent}$ $\sim$ ObliDB [@ObliDB18] OLTP & OLAP data structure (array or $B^+$-tree) $>\SI{100}{\percent}$ $>\SI{200}{\percent}$ $\sim$ StealthDB [@gribov2017stealthdb] OLTP primitive operators (e.g., $\leq$, $\geq$, $+$, $$) $>\SI{300}{\percent}$ $>\SI{20}{\percent}$ $\sim$ EncDBDB OLAP data structure (dictionaries) $<\SI{100}{\percent}$ $\sim \SI{8.9}{\percent}$ ---------------------------------- ------------------------- ----------------------------------------------------- ------------- ----------------------- --------------------------- -------- The main contributions of EncDBDB are: - New architecture for search over encrypted data suitable for column-oriented, in-memory databases. - Nine different encrypted dictionaries from which the data owner can freely select on column granularity. They provide different security (order and frequency leakage), performance and storage efficiency tradeoffs. The security ranges from the equivalent of deterministic order-revealing encryption [@boneh2015semantically] to range predicate encryption [@lu_privacy-preserving_2012]. - Integration into MonetDB [@boncz_breaking_2008; @boncz2005monetdb; @idreos_monetdb:_2012], an open source DBMS. The enclave has only lines of code, reducing the potential for security-relevant implementation errors and side-channel leakages. Query optimization and auxiliary database functionalities, e.g., storage, transaction, and database recovery management still operate without changes to the original code. - Sub-millisecond overhead for encrypted range queries compared to plaintext range queries, on a real-world customer database containing millions of entries. - Less storage space required for a compressed, encrypted column with the appropriate encrypted dictionary than for a plaintext column with the same data. Background {#sec:02:background} ========== First, EncDBDB is a column-oriented, dictionary encoding based, in-memory database. Second, a TEE, more specifically Intel SGX, is used to protect and process data stored in the database. Third, the database data are encrypted with probabilistic authenticated encryption. We review these three concepts in this section. Column-oriented, Dictionary Encoding\ based, In-memory Databases {#subsec:02:background:database} ------------------------------------- **In-memory database.** Many commercial and open source DBMS vendors offer in-memory databases for analytical data processing, e.g., SAP HANA [@saphana], Oracle RDBMS [@oracledatabase], and MonetDB [@monetdbWebsite]. In-memory databases permanently store the primary data in main memory and use the disk as secondary storage. The major benefit of in-memory databases is the lower access time of main memory compared to disk storage. This speeds up every data access for which disk access would be necessary. Additionally, it leads to shorter locking times in concurrency control, thus fewer cache flushes and a better CPU utilization. See [@dewitt1984implementation; @garcia1992main; @larson2016modern] for more details. **Column-oriented, In-memory Database.** One possible database storage concept is to store the data column-oriented, i.e., successive values of each column are stored consecutively, and surrogate identifiers are (implicitly) introduced to connect the rows [@DBS-024; @boncz1999mil; @copeland1985decomposition; @stonebraker2005c]. The combination of in-memory databases and column-oriented storage reduces the number of cache misses, which strongly influences the in-memory performance. All in-memory databases mentioned above support column-oriented storage. The main drawbacks of column-oriented storage are: (1) so-called tuple-reconstruction is necessary to re-assemble a projection involving multiple attributes and (2) inserts and updates of a tuple are written to non-contiguous storage locations. These problems are not severe in the context of analytical applications, e.g., data warehousing and business intelligence, because analytical queries often involve a scan on a significant amount of all tuples, but only a small subset of all columns [@boncz1999mil; @krueger2010enterprise]. Additionally, bulk loading of data is often used in this context and complex, long, read-only queries are executed afterwards [@harizopoulos2006performance; @stonebraker2005c]. An example query is a report on total sales per country for products in a certain price range. Only the few columns that are involved in the query have to be loaded and they can be processed sequentially, which is beneficial as it decreases cache misses of CPUs. **Column-oriented, Dictionary Encoding based, In-memory Databases.** The three commercial DBMSes mentioned above and many other databases use data compression mechanisms to exploit redundancy within data [@abadi2006integrating; @Willhalm]. Abadi et al. [@abadi2006integrating] study multiple database compression schemes, e.g., null suppression, run-length encoding and dictionary encoding, and show how they can be applied to column-oriented databases. According to the authors, column-oriented databases particularly profit from compression. In this paper, we only consider dictionary encoding, because it is the most prevalent compression used in column-oriented databases [@abadi2006integrating]. Throughout this paper, we say a tuple $T{}$ contains $\abs{T{}}$ values, i.e., $\ensuremath{T{}}\xspace{} = (\allowbreak{}\ensuremath{T{}_{0}}\xspace,\allowbreak{}\ldots{},\allowbreak{}\ensuremath{T{}_{\ensuremath{\abs{T{}}}\xspace{} - 1}}\xspace )$. We use $\ensuremath{v}\xspace{} \in \ensuremath{T{}}\xspace{}$ as a shorthand for a value $v$ that is contained in the tuple $T{}$. The idea of dictionary encoding is to split a column $\ensuremath{C{}}\xspace{} = (\allowbreak{}\ensuremath{C{}_{0}}\xspace,\allowbreak{}\ldots{},\allowbreak{}\ensuremath{C{}_{\ensuremath{\abs{C{}}}\xspace{} - 1}}\xspace )$ into two structures: a *dictionary* $D{}$ and an *attribute vector* $AV{}$. The dictionary $\ensuremath{D{}}\xspace{} = ( \ensuremath{D{}_{0}}\xspace, \allowbreak \ldots{}, \allowbreak \ensuremath{D{}_{\ensuremath{\abs{D{}}}\xspace - 1}}\xspace )$ is filled with all values $\ensuremath{v}\xspace{} \in \ensuremath{C{}}\xspace{}$ and every $v$ has to be present in $D{}$ at least once. The index $i$ of a dictionary entry $D{}_{i}$is called the ValueID ($vid{}$) that corresponds to this value. The attribute vector $\ensuremath{AV{}}\xspace{} = (\allowbreak{} \ensuremath{AV{}_{0}}\xspace, \allowbreak \ldots{}, \allowbreak{} \ensuremath{AV{}_{\ensuremath{\abs{AV{}}}\xspace{} - 1}}\xspace )$ is constructed by replacing all values $\ensuremath{v}\xspace{} \in \ensuremath{C{}}\xspace{}$ with one $vid{}$ that corresponds to $v$. As a result, $AV{}$ contains $\ensuremath{\abs{AV{}}}\xspace{} = \ensuremath{\abs{C{}}}\xspace{}$ ValueIDs. The index $j$ of an entry $AV{}_{j}$is called its RecordID ($rid{}$). $un(C{})$ denotes the set of unique values in $C{}$, $\abs{un(C{})}$ the amount of unique values, $oc(C{},v{}$) the occurrence indices of a unique value $v$ in $C{}$, and $\abs{oc(C{},v{})}$ the number of occurrences of $v$. We define the correctness of a column split as follows: \[def:2:split\_cor\] Given a column $C{}$, we say that a split of $C{}$ into a dictionary $D{}$ and an attribute vector $AV{}$ is correct if $i$ is the ValueID stored in the attribute vector at position $j$ and $D{}_{i}$equals $C{}_{j}$, i.e., $\forall{} j \in [0, \ensuremath{\abs{AV{}}}\xspace{} - 1] \colon{} \allowbreak i = \ensuremath{AV{}_{j}}\xspace \wedge{} \ensuremath{D{}_{i}}\xspace = \ensuremath{C{}_{j}}\xspace$. In Figure \[fig:02:dictionaryEx\], we present a split example based on a small first name column (`FName`). For instance, Jessica was inserted in the dictionary at the ValueID $1$ and all positions from the original column that contained Jessica are replaced by this ValueID in the attribute vector (see RecordIDs $0$, $2$ and $3$). The set of unique values is $\ensuremath{un(C{})}\xspace{} = \{\text{Hans}, \text{Jessica},\allowbreak{}\text{Archie}\}$ and Archie occurs at the positions $\ensuremath{oc(C{},\text{Archie}}\xspace) = \{1, 5\}$. ![Dictionary encoding example[]{data-label="fig:02:dictionaryEx"}](figures2/dictionaryEx){width="\linewidth"} Note that a split column requires less space than the original column in many cases, because a ValueID of $i$ Bits is sufficient to represent $2^i$ different values in the attribute vector and the (variable-length) values only have to be stored once in the dictionary. For instance, a column that contains strings of 10 characters each, but only $256$ unique values, requires $256 \cdot \SI{10}{\byte}$ for the dictionary and $\num{10000} \cdot \SI{1}{\byte}$ for the attribute vector. In total, dictionary encoding reduces the required storage from to . Dictionary encoding has the best compression rate if columns contain few unique but many frequent values, because every value has to be stored only once. The real-world data used for our evaluation (see Section \[subsec:02:evaluation:perfEval\]) and other studies [@muller2014adaptive; @lemke_speeding_2010] show that this is a characteristic of many columns in data warehouses. High compression rates achieved by dictionary encoding sparingly use the scarce resource of in-memory databases — main memory. A search for all entries falling in a range $\ensuremath{R}\xspace{}$ is performed in two steps if dictionary encoding is used: a dictionary search followed by an attribute vector search. The dictionary search checks for every $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ if it falls into $R$ and returns the matching ValueIDs ($\mathbf{vid{}}$). The attribute vector search linearly scans the attribute vector searching for every value $\ensuremath{v}\xspace{} \in \ensuremath{\mathbf{vid{}}}\xspace{}$ and returns a list of matching RecordIDs ($\mathbf{rid{}}$). This operation is parallelizable with a speedup expected to be linear in the number of threads. In the example of Figure \[fig:02:dictionaryEx\], a search for $\ensuremath{R}\xspace{} = [\text{Archie}, \text{Hans}] $ in the dictionary returns $\ensuremath{\mathbf{vid{}}}\xspace{} = \{0,2\}$. Searching these ValueIDs in the attribute vector returns $\ensuremath{\mathbf{rid{}}}\xspace{} = \{1, 4, 5\}$. Intel Software Guard Extensions (SGX) {#subsec:02:background:sgx} ------------------------------------- Intel SGX is an instruction set extension that is available in Intel Core processors since the Skylake generation and in Intel Xeon processors since the Kaby Lake generation, making it a widely available TEE. It provides a secure, isolated processing area, called enclave, which guarantees confidentiality and integrity protection to code and data in it, even in an untrusted environment. We present SGX’s features used by EncDBDB. See [@Intel_SGX3; @costanintel; @Intel_SGX2; @SGX_Ref; @iscaTut; @Intel_SGX1] for more details. **Memory Isolation.** SGX v2 dedicates of the system’s main memory (RAM) for the so-called Processor Reserved Memory (PRM). All code and data in the PRM is encrypted while residing outside of the CPU, and decrypted and integrity checked when the data is loaded into the CPU. All other software on the system, including privileged software such as OS, hypervisor, and firmware, cannot access the PRM. Only about of the PRM can be used for enclave code and data, even if multiple enclaves are present. The OS can swap out enclave pages and SGX ensures integrity, confidentiality and freshness of swapped-out pages, but paging comes with a major performance overhead. Every program using SGX consists of an enclave and an untrusted part. The untrusted part is executed as an ordinary process within the virtual memory address space and the enclave memory is mapped into the virtual memory of the untrusted host process. This mapping allows the enclave to access the entire virtual memory of its host process, while the host process can invoke the enclave only through a well-defined interface. **Attestation.** SGX has a remote attestation feature, which allows verification of code integrity and authenticity on a remote system. This is done by hashing (called *measuring* in SGX terminology) the initial code and data loaded into the enclave. The authenticity of the measurement, as well as the fact that the measurement originates from a benign enclave, is ensured by SGX’s attestation feature (refer to [@Intel_SGX3] for details). The measurement can be provided to an external party to prove the correct creation of an enclave. Furthermore, the remote attestation feature allows establishment of a secure channel between an external party and an enclave. This secure channel can be used to deploy sensitive data, e.g., cryptographic keys, directly into the enclave. Probabilistic Authenticated Encryption {#subsec:02:background:authEnc} -------------------------------------- A probabilistic authenticated encryption ($\mathtt{PAE{}}$) scheme provides confidentiality, integrity, and authenticity of encrypted data. $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ takes a secret key $SK$, a random initialization vector $IV$ and a plaintext value $v$ as input and returns a ciphertext $c{}$. $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}$ takes $SK$ and $c{}$ as input and returns $v$ iff $v$ was encrypted with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under the initialization vector $IV$ and the secret key $SK$. AES-128 in GCM mode [@dworkin2007recommendation] can be used as a $\mathtt{PAE{}}$ implementation. High Level Design of EncDBDB {#sec:02:designOverview} ============================ In this section, we give an overview of EncDBDB’s setup and query phase, followed by the considered attacker model. EncDBDB Overview {#subsec:02:designOverview:overview} ---------------- EncDBDB provides nine encrypted dictionaries and also supports plaintext dictionaries. In the setup phase, one of these is selected per column of the data owner’s dataset. The selection determines how each column is split into a dictionary and an attribute vector. All values in encrypted dictionaries are encrypted with $\mathtt{PAE{}}$ under a key determined by the data owner. The encrypted dictionaries provide different tradeoffs regarding security, performance, and storage efficiency. EncDBDB is able to process all dictionary types together, even if they are mixed in one table. For brevity, plaintext dictionaries are not discussed any further, but performance measurements are shown in the evaluation section. The data owner’s dataset is deployed at a DBaaS provider that supports Intel SGX (see Figure \[fig:02:overview\]). SGX can be replaced by any other TEE that provides the required capabilities such as integrity and confidentiality protection of code and data, remote attestation, and secure data provisioning. Additionally, the data owner’s secret key ($SK_{DB}$) is deployed into the SGX enclave that is part of the DBMS and to a trusted proxy. ![High level design of EncDBDB[]{data-label="fig:02:overview"}](figures2/overviewSmall){width="\linewidth"} After this setup, the query phase starts, and an application can send queries to the DBaaS offering. These queries are routed through the proxy, where all values are encrypted with $\mathtt{PAE{}}$ and forwarded to the DBMS, which processes the queries in a pipeline. The pipeline outputs encrypted range queries on individual columns. The enclave is used for protected dictionary searches. Attribute vector searches and all other DBMS functionality are performed outside of the enclave. The DBMS combines the individual results into a total result and returns it to the proxy. At the proxy, the total result is decrypted and forwarded to the application for which the whole process is transparent. We introduce the encrypted dictionaries in detail in Section \[subsec:02:main:encDTs\]. In Section \[subsec:02:main:designDetail\], we provide an in-depth description of how the encrypted dictionaries are used as the main building block for implementing a protected DBMS that supports auxiliary DBMS functions. In Section \[subsec:02:main:dynData\], we explain how EncDBDB can handle data insertions, deletions, and updates. Attacker Model {#subsec:02:designOverview:attackerModel} -------------- Our attacker model considers the data owner, application and proxy as trusted. On the server side, we assume an honest-but-curious attacker, i.e., a passive attacker who follows the protocol, but tries to gain as much information as possible. The enclave code and data is protected by the TEE. The code is assumed not to have intentional data leakage. However, the attacker can observe all other software running at the DBaaS provider, e.g., the OS, the firmware and the DBMS. As a result, the attacker has full access to data stored on disk and main memory, and she is able to observe the access pattern to them. Additionally, she can track all communication between the enclave and resources outside of it, and all network communication between the proxy and the DBMS. Note that this includes the incoming queries in which only the data values are encrypted. Various research studies have shown that SGX is vulnerable to various side-channel attacks, e.g., cache attacks [@brasser2017software], timing attacks [@weichbrodt2016asyncshock] or page faults [@xu2015controlled]. Other researchers have presented solutions to these problems. For instance, how to mitigate the page fault side-channel [@shinde2016preventing], how to detect side-channels [@chen2017detecting] and how to protect against cache-based side-channels [@gruss2017strong]. We consider the side-channel exploitation and protection as an orthogonal problem and thus do not consider them in this work. However, we design our system to have minimal enclave code and therefore, the protections should be straightforward to integrate. Hardware attacks and Denial of Service (DoS) are out of scope. We assume that the attacker targets each database column independently, i.e., she does not use correlation information to target columns. It remains future work to evaluate how decorrelation of columns protects the database in practice. EncDBDB Design {#sec:02:main} ============== In this section, we continue the description of EncDBDB, which was already introduced in Section \[subsec:02:designOverview:overview\]. We first explain the nine encrypted dictionaries that EncDBDB supports and then elaborate how they can be used to build an encrypted DBMS. {#subsec:02:main:encDTs} The encrypted dictionaries differ from each other in two dimensions — repetition and order of values in $D{}$ — with three options each (see Table \[tab:02:encDTsCharacteristics\]). The repetition options are: frequency revealing, frequency smoothing, and frequency hiding. The order options are: sorted lexicographically, sorted and rotated around a random offset, and unsorted. An encrypted dictionary is defined by one option from each dimension, which leads to nine data structures with different security, search time and storage efficiency features. ------------- --------------------- -------- --------- ---------- (lr)[3-5]{} sorted rotated unsorted frequency revealing ED1 ED2 ED3 \[4pt\] frequency smoothing ED4 ED5 ED6 \[4pt\] frequency hiding ED7 ED8 ED9 ------------- --------------------- -------- --------- ---------- : Characteristics of encrypted dictionaries[]{data-label="tab:02:encDTsCharacteristics"} The idea of the repetition options is to increase the number of repetitions of dictionary values from frequency revealing to frequency hiding. This directly influences two features of the resulting encrypted dictionaries (see Table \[tab:02:encDTsRepetition\]): the security feature frequency leakage and dictionary size ($\abs{D{}}$). Note that $\abs{D{}}$ is fixed for frequency revealing and frequency hiding. For frequency smoothing, the worst-case size is $\abs{AV{}}$, but we give the average size, which depends on a configurable parameter $bs{}_{\max}$. repetition options frequency leak. dictionary size $\abs{D{}}$ --------------------- ----------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------- frequency revealing full $\abs{un(C{})}$ frequency smoothing bounded $\sim \sum_{\ensuremath{v}\xspace{} \in \ensuremath{C{}}\xspace{}}{\frac{2 \cdot \ensuremath{\abs{oc(C{},v{})}}\xspace{}}{1 + \ensuremath{bs{}_{\max}}\xspace{}}}$ frequency hiding none $\abs{AV{}}$ : Security feature frequency leakage and dictionary size of repetition options[]{data-label="tab:02:encDTsRepetition"} [0.14]{} ![image](figures2/eds_column) [0.28]{} ![image](figures2/eds_ed1) [0.28]{} ![image](figures2/eds_ed2) [0.28]{} ![image](figures2/eds_ed3) The order options also determine two features of the encrypted dictionaries (see Table \[tab:02:encDTsOrder\]). First, they determine the security feature order leakage, i.e., the information an attacker with memory access can learn about the plaintext order of the encrypted values in $D{}$. Second, they determine the search time combining the dictionary and attribute vector search time. The dictionary search time depends on $\abs{D{}}$ and the search algorithm, which differs for the order options. The attribute vector search time depends on the amount of ValueIDs returned by the dictionary search, because $AV{}$ has to be scanned for them. order options order leakage search time --------------- --------------- --------------------------------------------------------------------------------------------------------------------------- sorted full $O(\log \ensuremath{\abs{D{}}}\xspace{}) + O(\ensuremath{\abs{AV{}}}\xspace{})$ rotated bounded $O(\log \ensuremath{\abs{D{}}}\xspace{}) + O(\ensuremath{\abs{AV{}}}\xspace{})$ unsorted none $O(\ensuremath{\abs{D{}}}\xspace{}) + O(\ensuremath{\abs{AV{}}}\xspace{}\cdot\ensuremath{\mathbf{\abs{vid{}}}}\xspace{})$ : Security feature order leakage and search time of order options[]{data-label="tab:02:encDTsOrder"} Three operations differ for the nine encrypted dictionaries: (1) creation of the encrypted dictionaries, (2) dictionary search inside the enclave at the DBaaS provider, and (3) attribute vector search in the untrusted realm at the DBaaS provider. In the next sections, we denote the corresponding operations as (1) $\mathtt{EncDB}$, (2) $\mathtt{EnclDict}\-\mathtt{Search}$, and (3) $\mathtt{AttrVect}\-\mathtt{Search}$, and describe these operations in detail. As mentioned before, an encrypted dictionary is defined by an order and a repetition option. We start by describing the frequency revealing algorithm and then explain how it is combined with the three order options to instantiate ED1–ED3. Then, we do the same for the frequency smoothing algorithm and its combinations (ED4–ED6) followed by the frequency hiding algorithm and its combinations (ED7–ED9). We always assume a closed search range. Open or half-open ranges can be handled trivially, however we omit the details to provide a concise description. ### Frequency revealing {#frequency-revealing .unnumbered} In the frequency revealing algorithm, the split of a column $C{}$ is performed by inserting each unique value $\ensuremath{v}\xspace{} \in \ensuremath{un(C{})}\xspace{}$ into $D{}$ exactly once at an arbitrary position, i.e., $\ensuremath{\abs{D{}}}\xspace{} = \ensuremath{\abs{un(C{})}}\xspace{} \wedge \forall \ensuremath{v}\xspace{} \in \ensuremath{un(C{})}\xspace{} \colon \ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$. The ValueIDs in $AV{}$ are set such that the split is correct according to Definition \[def:2:split\_cor\]. The frequency revealing algorithm provides the best compression rate that is possible with dictionary encoding and thus is the most storage efficient repetition option. However, an attacker can learn the frequency of each value $\ensuremath{D{}_{j}}\xspace \in \ensuremath{D{}}\xspace{}$ by counting the occurrences of $j$ in $AV{}$. This is still true if each $\ensuremath{v}\xspace{} \in{} \ensuremath{D{}}\xspace{}$ is encrypted with probabilistic authenticated encryption. Therefore, the three frequency revealing encrypted dictionaries presented next have full frequency leakage. **ED1.** For each column $C{}$ that is protected with ED1, $\mathtt{EncDB\_}\allowbreak{}\mathtt{1}$performs the split operation according to the frequency revealing algorithm, sorts the values $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ lexicographically and adjusted the ValueIDs in $AV{}$ such that the split is correct. Afterwards, $\mathtt{EncDB\_}\allowbreak{}\mathtt{1}$derives $SK_{D}$ from the data owner’s secret key $SK_{DB}$, the table name, and the column name. It then encrypts all values $v$ individually with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under $SK_{D}$ and a random initialization vector $IV$. The resulting dictionary containing encrypted values is denoted as $eD{}$. Figure \[fig:02:exDT1toDT3\] (a) presents an example column $C{}$ and Figure \[fig:02:exDT1toDT3\] (b) the result of ED1 before $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ is performed. ED1 has full order leakage because an attacker knows the plaintext order of the encrypted values $\ensuremath{c{}}\xspace{} \in{} \ensuremath{eD{}}\xspace{}$. ED1’s dictionary search ($\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$), which is executed in the enclave at the DBaaS provider, is presented in Algorithm \[algo:02:dictSearch\_1\]. The function gets an encrypted range $\tau$ and an encrypted dictionary $eD{}$ as input. First it derives $SK_{D}$ and decrypts the start and end of the range individually. Then, one leftmost and one rightmost binary search is performed to find the dictionary indices where the searched range starts ($vid{}_{min}$) and ends ($vid{}_{max}$). All dictionary values are encrypted and stored in untrusted memory. Thus, the binary searches load the values into the enclave individually, decrypt them there, and compare them with the search value. The number of load, decrypt and compare operations is logarithmic in $\abs{D{}}$. In our implementation we use the result of the searches, and whether or not a value was found, to handle cases in which a value is not present. This detail is omitted for brevity in Algorithm \[algo:02:dictSearch\_1\]. $\ensuremath{SK_{D}}\xspace{} = \Call{DeriveKey}{\ensuremath{SK_{DB}}\xspace{},colName,tabName}$ $R$ = ($R_s$, $R_e$) = $\big(\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_s}\xspace{})}\xspace,\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_e}\xspace{})}\xspace\big)$ $vid{}_{min}$= $vid{}_{max}$= **return** $\mathbf{vid{}}$ = ($vid{}_{min}$, $vid{}_{max}$) Note that only very small, constant enclave memory is required for $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$as well as for the $\mathtt{EnclDict}\-\mathtt{Search}$ operations of all other encrypted dictionaries. Especially, the required enclave memory is independent of $\abs{D{}}$. Afterwards, $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$is executed in the untrusted realm at the DBaaS provider. It linearly scans the corresponding $AV{}$, checks if the ValueIDs fall between $vid{}_{min}$and $vid{}_{max}$, and returns the matching RecordIDs ($\mathbf{rid{}}$), i.e., $\ensuremath{\mathbf{rid{}}}\xspace{} = \{ i \, | \, \ensuremath{AV{}_{i}}\xspace \in{} \ensuremath{AV{}}\xspace{} \wedge \ensuremath{AV{}_{i}}\xspace \in [\ensuremath{vid{}_{min}}\xspace, \ensuremath{vid{}_{max}}\xspace]\} $. This operation is parallelizable with a speedup expected to be linear in the number of threads. **ED2.** The idea of the rotated algorithm, which is used in ED2, is to sort and randomly rotate $D{}$. $\mathtt{EncDB\_}\allowbreak{}\mathtt{2}$executes the frequency revealing algorithm, sorts the values in $D{}$ lexicographically, performs a random rotation of $D{}$ as explained in the following paragraph, adjusts the ValueIDs in $AV{}$ such that the split is correct, and encrypts all $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ with $\mathtt{PAE{}}$ under $SK_{D}$ and a random $IV$, resulting in $eD{}$. $\mathtt{EncDB\_}\allowbreak{}\mathtt{2}$generates a random offset (**rndOffset**) and rotates $D{}$ by this value. More formally, let $D'{}$ be the sorted dictionary, then $\ensuremath{D{}}\xspace{} = (\ensuremath{D{}_{i}}\xspace \, | \, \ensuremath{D{}_{i}}\xspace = \ensuremath{D'{}_{j}}\xspace \wedge{} i = (j + \text{\textit{rndOffset}{}}) \bmod{} \ensuremath{\abs{D'{}}}\xspace)$. $\mathtt{EncDB\_}\allowbreak{}\mathtt{2}$encrypts *rndOffset* with $\mathtt{PAE{}}$ under $SK_{D}$ and a random $IV$, and attaches the resulting **encRndOffset** to $eD{}$. Figure \[fig:02:exDT1toDT3\] (c) illustrates an example with $\textit{rndOffset}{} = 3$ (before encryption). For instance, has the ValueID $2$ in a sorted dictionary $D'{}$. After the rotation, the ValueID is $1 =(2 + 3) \bmod{} 4$. The order leakage is bounded, because an attacker who can observe no or a limited number of queries, does not know where the smallest and largest values are stored in $eD{}$. The idea of modular order-preserving encryption in the context of probabilistic encryption was introduced in [@Kerschbaum17]. The processing inside the enclave ($\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{2}$) is illustrated in Algorithm \[algo:02:dictSearch\_2\]. First, $SK_{D}$ is derived, and the encrypted range $\tau$ and *encRndOffset* are decrypted with it. Then, a special variant of binary search, which is explained in the next paragraph, is called to search the start and the end of the range — $vid{}_{min}$and $vid{}_{max}$. These indices have to be processed further inside of the enclave, because the positions of the indices relative to *rndOffset* define the final result of the dictionary search and *rndOffset* is sensitive. There are three possibilities: both indices are lower than *rndOffset*; both are greater than or equal to *rndOffset*; or $vid{}_{min}$is above and $vid{}_{max}$is below *rndOffset*. In the first and second case, the results are in the range $\big(\ensuremath{vid{}_{min}}\xspace, \ensuremath{vid{}_{max}}\xspace\big)$. In the third case, there are again two possibilities: $vid{}_{min}$does or does not equal $\abs{eD{}}$. In the first case, the range start was not found in $eD{}$, but it is higher than the last value in it. Accordingly, all results are in the range $\big(0, \ensuremath{vid{}_{max}}\xspace\big)$. Otherwise, the results are split in a lower range $\big(0, \ensuremath{vid{}_{max}}\xspace\big)$ and an upper range $\big(\ensuremath{vid{}_{min}}\xspace, \ensuremath{\abs{eD{}}}\xspace{} - 1\big)$. We always return a dummy range if the result is only one range to simplify attribute vector search. $\ensuremath{SK_{D}}\xspace{} = \Call{DeriveKey}{\ensuremath{SK_{DB}}\xspace{},colName,tabName}$ $\ensuremath{R}\xspace{} = (\ensuremath{R_s}\xspace{}, \ensuremath{R_e}\xspace{}) = \ensuremath{\big(\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_s}\xspace{})}\xspace,\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_e}\xspace{})}\xspace\big)}$ $\textit{rndOffset}{} = \ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\textit{encRndOffset}{})}\xspace$ $vid{}_{min}$= $vid{}_{max}$= $\mathbf{vid{}}$ = $\emptyset{}$ $\mathbf{vid{}}$ = {$\big(\ensuremath{vid{}_{min}}\xspace, \ensuremath{vid{}_{max}}\xspace\big)$, $\big(-1, -1\big)$} $\mathbf{vid{}}$ = {$\big(0, \ensuremath{vid{}_{max}}\xspace\big)$, $\big(\ensuremath{vid{}_{min}}\xspace, \ensuremath{\abs{eD{}}}\xspace{} - 1\big)$} $\mathbf{vid{}}$ = {$\big(0, \ensuremath{vid{}_{max}}\xspace\big)$, $\big(-1, -1\big)$ } **return** $\mathbf{vid{}}$ Algorithm \[algo:02:BinarySearchSpecial\] presents the details of the special binary search with slightly different handling of the range start and end. The goal is to perform a binary search that has an access pattern that is independent of *rndOffset*. A binary search that simply considers *rndOffset* during the data access would leak *rndOffset* in the first round, which would completely thwart the additional protection. The algorithm uses a string encoding operation (`ENCODE`), which converts string values of a fixed maximal length to an integer representation preserving the lexicographical data order. Each character is converted individually to an integer of fixed length and the integers are concatenated to one resulting integer. For instance, the encoding of would be $3334$ and would lead to $3433$. The lexicographical order is preserved by right padding the resulting integer to a fixed maximal length. In many DBMSes, the values in each column of a database have a fixed maximal length, which is fixed either implicitly by the datatype, e.g., $32$ bit for `INTEGER` columns (in MySQL), or fixed explicitly with the datatype, e.g., $30$ characters for `VARCHAR(30)` columns. For instance, `ENCODE` converts to the decimal $3334000000$ for a `VARCHAR(5)` column. $l = 0$, $h = \ensuremath{\abs{eD{}}}\xspace{}$ $r = \Call{encode}{\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{eD{}_{0}}\xspace)}\xspace}$ $N = \Call{encode}{\text{column maximum}}$ $\ensuremath{sVal{}}\xspace{} = (\Call{encode}{\ensuremath{sVal{}}\xspace{}} - r) \% N$ $j = \left \lceil{(l + h) / 2}\right \rceil$ $m = \Call{encode}{\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{eD{}_{j}}\xspace)}\xspace}$ $\ensuremath{cVal{}}\xspace{} = (m - r) \% N$ $l = j + 1$ $h = j$ **return** Algorithm \[algo:02:BinarySearchSpecial\] first initializes the low and high value of the search. A value $r$ is determined by decrypting $eD{}_{0}$and executing `ENCODE` on it. Then, `ENCODE` is performed on the maximum value that fits the column, which is implicitly defined by the fixed maximal length of the column. `ENCODE` is also executed on the search value ($\ensuremath{sVal{}}\xspace{}$), $r$ is subtracted from it and the result is taken modulo $N$. All values $m$ accessed during the search are loaded into the enclave, decrypted and handled as $\ensuremath{sVal{}}\xspace{}$. Note that is a possible value for *rndOffset*, because *rndOffset* is chosen uniformly at random between and $\ensuremath{\abs{D{}}}\xspace{} - 1$. We omit the special handling for brevity. Overall, the runtime of $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{2}$is logarithmic in $\abs{D{}}$ and the encoding introduces only a constant factor compared to $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$. $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{2}$linearly scans $AV{}$ outside of the enclave and checks if the values $\ensuremath{v}\xspace{} \in \ensuremath{AV{}}\xspace{}$ fall in either range that was returned by $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{2}$. The RecordIDs ($\mathbf{rid{}}$) of the matching values are returned by this operation. **ED3.** This encrypted dictionary combines the repetition option frequency revealing and the order option unsorted. Accordingly, $\mathtt{EncDB\_}\allowbreak{}\mathtt{3}$performs the frequency revealing algorithm and then shuffles the unique values $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ randomly resulting in an unsorted dictionary. Afterwards, the ValueIDs in $AV{}$ are set such that the split is correct and all values $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ are encrypted with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under $SK_{D}$ and a random $IV$. Figure \[fig:02:exDT1toDT3\] (d) shows an example for $\mathtt{EncDB\_}\allowbreak{}\mathtt{3}$before $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ is performed. $\mathtt{EncDB\_}\allowbreak{}\mathtt{3}$trivially has no order leakage. ED3’s unsorted dictionary prevents the use of any search with logarithmic runtime during $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$. Instead, a linear scan over all values $\ensuremath{c{}}\xspace{} \in \ensuremath{eD{}}\xspace{}$ has to be performed (see Algorithm \[algo:02:dictSearch\_7\]). First, $SK_{D}$ is derived and used to decrypt the encrypted search range $\tau$. Then, the algorithm loads each $\ensuremath{c{}}\xspace{} \in \ensuremath{eD{}}\xspace{}$ into the enclave, decrypts $c{}$ and checks if $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{c{}}\xspace{})$falls into $R$. The result is a list of all matching ValueIDs $\mathbf{vid{}}$. $\ensuremath{SK_{D}}\xspace{} = \Call{DeriveKey}{\ensuremath{SK_{DB}}\xspace{},colName,tabName}$ $R$ = ($R_s$, $R_e$) = $\big(\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_s}\xspace{})}\xspace,\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{\tau_e}\xspace{})}\xspace\big)$ $\mathbf{vid{}}$ = $\emptyset{}$ $v$ = $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Dec}(\allowbreak{}\ensuremath{SK_{D}}\xspace{},\allowbreak{}\ensuremath{eD{}_{i}}\xspace)$ $\mathbf{vid{}}$. **return** $\mathbf{vid{}}$ $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$has to compare every $\ensuremath{v}\xspace{} \in \ensuremath{AV{}}\xspace{}$ with every $\ensuremath{u}\xspace{} \in \ensuremath{\mathbf{vid{}}}\xspace{}$ returned by $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$. Thus, the runtime complexity is $O(\ensuremath{\abs{AV{}}}\xspace{}\cdot\ensuremath{\mathbf{\abs{vid{}}}}\xspace{})$. However, integers are compared in this case, which is a highly optimized operation in most CPUs. Additionally, $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$is easily parallelizable. ### Frequency smoothing {#frequency-smoothing .unnumbered} The main problem of the frequency revealing algorithm is that an attacker can learn the frequency of each value $\ensuremath{D{}_{j}}\xspace \in \ensuremath{D{}}\xspace{}$ even if the values are encrypted. The reason is that the underlying plaintext values are present only once with a unique ValueID. As a countermeasure, the frequency smoothing algorithm bounds the frequency leakage by inserting plaintext duplicates into $D{}$ during the column split. The foundation of this repetition option is the Uniform Random Salt Frequencies method [@Pouliot17]. In more detail, the frequency smoothing algorithm executes a parameterizable and probabilistic random experiment for each unique value $\ensuremath{v}\xspace{} \in \ensuremath{un(C{})}\xspace{}$ to determine how often $v$ should be inserted into $D{}$ (see Algorithm \[algo:02:bucketSizes\]). We say that a plaintext value $v$ is split into multiple buckets and every bucket has a specific size. The number of occurrences of $v$ in $C{}$ ($\abs{oc(C{},v{})}$) and a maximal bucket size ($bs{}_{\max}$) is passed to this experiment. The random size for an additional bucket is picked from the discrete uniform distribution $\mathcal{U}\{1, \ensuremath{bs{}_{\max}}\xspace{}\}$until the total size is above $\abs{oc(C{},v{})}$. The size of the last bucket is then set such that the total size matches $\abs{oc(C{},v{})}$. The experiment returns the bucket sizes ($bs_{\text{sizes}}$) and how many buckets were chosen ($\#bs{}$). The frequency smoothing algorithm inserts $\#bs{}$ repetitions of $v$ into $D{}$. For each $\ensuremath{C{}_{i}}\xspace \in{} \ensuremath{oc(C{},v{}}\xspace){}$, it randomly inserts one of the $\#bs{}$ possible ValueIDs into $AV{}_{i}$. Each ValueID is used exactly as often as defined by $bs_{\text{sizes}}$. As a result, the frequency leakage has a bound, because the number of occurrences of each $ValueID{} \in \ensuremath{AV{}}\xspace{}$ is guaranteed to be between 1 and $bs{}_{\max}$. $ prevTotal = total = \ensuremath{\#bs{}}\xspace{} = 0 $ $ \ensuremath{bs_{\text{sizes}}}\xspace{} = \emptyset{}$ $ \ensuremath{\#bs{}}\xspace{} \mathrel{+}= 1 $ $ rnd \xleftarrow{\$} [1, \ensuremath{bs{}_{\max}}\xspace{} ] $ $ \ensuremath{bs_{\text{sizes}}}\xspace{}.\Call{append}{rnd} $ $ prevTotal = total $ $ total \mathrel{+}= rnd $ $ \ensuremath{bs_{\text{sizes}}}\xspace{}.\Call{last}{} = \ensuremath{\abs{oc(C{},v{})}}\xspace{} - prevTotal $ **return** $ \ensuremath{bs_{\text{sizes}}}\xspace{}, \ensuremath{\#bs{}}\xspace{} $ $bs{}_{\max}$ can be chosen independently for each column. The selection influences $\abs{D{}}$, which impacts storage efficiency, search time and frequency leakage. For instance, a large $bs{}_{\max}$ leads to few repeating entries in $D{}$, which slightly increases $\abs{D{}}$ compared to the frequency revealing algorithm. This decreases the $\mathtt{EnclDict}\-\mathtt{Search}$ performance, because more data needs to be loaded into the enclave, more decryptions are performed, and more comparisons are necessary. The performance of $\mathtt{AttrVect}\-\mathtt{Search}$ also decreases, because more values have to be compared. A small $bs{}_{\max}$ leads to many repetitions in $D{}$, which further increases $\abs{D{}}$ and the search time. Yet, it leads to a low frequency leakage bound, as each ValueID in $AV{}$ is present at most $bs{}_{\max}$ times. Next, we explain how the frequency smoothing algorithm impacts the three order options, which were introduced in detail before. We omit the discussion of order leakage as it is independent of the repetition option. **ED4.** $\mathtt{EncDB\_}\allowbreak{}\mathtt{4}$performs the split of $C{}$ according to the frequency smoothing algorithm and sorts all values in $D{}$ lexicographically determining the order of repetitions randomly. Then, it adjusts the ValueIDs in $AV{}$ such that the split is correct while considering how often each ValueID can be used, which is defined by $bs_{\text{sizes}}$. Finally, $\mathtt{EncDB\_}\allowbreak{}\mathtt{4}$encrypts all $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under $SK_{D}$ and a random $IV$. Note that this only leads to the same ciphertexts with negligible probability, even if the plaintexts are equal. $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{4}$is equal to $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$, because leftmost and rightmost binary searches inherently handle repetitions. The performance penalty compared to ED1 is small, because the binary search only slows down logarithmically with a growing $\abs{D{}}$. $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{4}$equals $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{1}$. **ED5.** For this encrypted dictionary, $\mathtt{EncDB\_}\allowbreak{}\mathtt{5}$performs the split of $C{}$ according to the frequency smoothing algorithm, rotates the ValueIDs as described in $\mathtt{EncDB\_}\allowbreak{}\mathtt{2}$, sets the ValueIDs in $AV{}$ such that the split is correct (considering $bs_{\text{sizes}}$), and encrypts all $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under $SK_{D}$ and a random $IV$. Figure \[fig:02:exDT5\] shows an example for ED5 with $\ensuremath{bs{}_{\max}}\xspace{} = 3$ and $\textit{rndOffset}{} = 1$ not considering the encryption. ![Example for ED5 with $\ensuremath{bs{}_{\max}}\xspace{} = 3$ and $\textit{rndOffset}{} = 1$ without encryption[]{data-label="fig:02:exDT5"}](figures2/ed5){width="\linewidth"} The special binary searches are more complex for ED5 than for ED2, because they have to handle a corner case: the plaintext value of the last and first entry in $D{}$ might be equal and present more than two times (as in the example in Figure \[fig:02:exDT5\]). For the same reason, $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{5}$has to perform a more complicated postprocessing of $vid{}_{min}$and $vid{}_{max}$compared to $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{2}$. The performance penalty compared to ED2 is small, because the binary search slows down logarithmically in $\abs{D{}}$. **ED6.** For columns that are protected with ED6, $\mathtt{EncDB\_}\allowbreak{}\mathtt{6}$splits $C{}$ according to the frequency smoothing algorithm, shuffles the values in $D{}$, sets the ValueIDs in $AV{}$ such that the split is correct (considering $bs_{\text{sizes}}$), and encrypts all $\ensuremath{v}\xspace{} \in \ensuremath{D{}}\xspace{}$ with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ under $SK_{D}$ and a random $IV$. $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{6}$is equal to $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$and $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{6}$is equal to $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{3}$, but frequency smoothing severely impacts the performance of these operations. The reason is that the linear scan of $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{6}$needs to load, decrypt, and compare more values. Additionally, the number of comparisons in $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{6}$increases as $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{6}$potentially returns more values. ![image](figures2/overview){width="\textwidth"} ### Frequency hiding {#frequency-hiding .unnumbered} Now we discuss the frequency hiding algorithm, which prevents frequency leakage. The idea is to add a separate entry into $D{}$ for every value in $C{}$, i.e., $\forall{} i \in [0, \ensuremath{\abs{C{}}}\xspace{} - 1] \colon{} \ensuremath{D{}_{i}}\xspace = \ensuremath{C{}_{i}}\xspace $. As a result, every $\ensuremath{v}\xspace{} \in \ensuremath{C{}}\xspace{}$ has $\abs{oc(C{},v{})}$ possible ValueIDs at which $v$ is stored. In turn, $\ensuremath{v}\xspace{} \in \ensuremath{AV{}}\xspace{}$ is set to one of those ValueIDs at random and every ValueID is used once. The resulting dictionary encoding does not provide compression anymore ($\ensuremath{\abs{D{}}}\xspace{} = \ensuremath{\abs{AV{}}}\xspace{}$), but the frequency of every ValueID is perfectly equal, i.e., there is no frequency leakage. **ED7, ED8 and ED9.** $\mathtt{EncDB\_}\allowbreak{}\mathtt{7}$, $\mathtt{EncDB\_}\allowbreak{}\mathtt{8}$and $\mathtt{EncDB\_}\allowbreak{}\mathtt{9}$execute a split of $C{}$ according to the frequency hiding algorithm; sort, rotate, and shuffle $D{}$, respectively; adjusts the ValueIDs in $AV{}$ such that the split is correct and every index in $D{}$ is only used once in $AV{}$; and encrypt all values in $D{}$ as described before. Frequency hiding can be interpreted as a special case of frequency smoothing with a $bs{}_{\max}$ of . Therefore, the $\mathtt{EnclDict}\-\mathtt{Search}$ and $\mathtt{AttrVect}\-\mathtt{Search}$ operations are equal as described for ED4, ED5, and ED6, and the advantage and disadvantages are equivalent to the ones described for a small $bs{}_{\max}$. EncDBDB in detail {#subsec:02:main:designDetail} ----------------- In this section, we present how encrypted dictionaries are used as foundation for an encrypted DBMS. In one possible EncDBDB variant, the DBaaS provider is assumed trusted for the initial setup. The data owner can upload plaintext columns and the DBaaS provider can support the data owner in choosing a proper encrypted dictionary for each column. Afterwards, the DBaaS performs the appropriate column splits and encryptions. In the following, however, we discuss another variant in which the DBaaS provider is untrusted also during the setup (see Figure \[fig:02:detailDesign\]) and plaintext data never leaves the realm of the data owner. We first describe the system setup, followed by the processing during runtime. ### Setup {#setup .unnumbered} The data owner defines a security parameter $\lambda$ and generates a secret key $\ensuremath{SK_{DB}}\xspace{}=\ensuremath{\mathtt{PAE{}\_}\allowbreak{}\mathtt{Gen}(\allowbreak{}1^\lambda{})}\xspace{}$. The data owner uses SGX’s attestation feature to authenticate the DBaaS server’s enclave and to establish a secure connection to it (see Section \[subsec:02:background:sgx\] for details), which is used to deploy $SK_{DB}$ to the enclave. Additionally, $SK_{DB}$ is deployed at the proxy via a secure out-of-band mechanism. The data owner takes its plaintext database PDB and selects an encrypted dictionary for each column $\ensuremath{C{}}\xspace{} \in PDB{}$. He performs $\mathtt{EncDB}$ as explained in the previous section for each $C{}$ according to the selected encrypted dictionary. Each encrypted dictionary is encrypted with an individual key $SK_{D}$, which is derived from $SK_{DB}$, the table name, and the column name. The result is an encrypted database EDB. As a last step of the setup, the data owner uses the import functionality of the DBaaS provider to deploy EDB. The storage management of the in-memory database stores all data on disk for persistency and additionally loads it into main memory. ### Runtime {#runtime .unnumbered} From this point on, the application can send an arbitrary number of queries, which are processed as follows. The application issues an SQL query $Q{}$ to the proxy. W.l.o.g. we assume that $Q{}$ selects and filters only one column. The filter can be an equality select, an inequality select, a greater than select (inclusive or exclusive), a less than select (inclusive or exclusive) and a range select (inclusive or exclusive). The proxy converts all filters to a range select with range $\ensuremath{R}\xspace{} = \ensuremath{\big(\ensuremath{R_s}\xspace{}, \ensuremath{R_e}\xspace{}\big)}$. For instance, the SQL query `SELECT FName FROM t1 WHERE FName < Ella` is converted to `SELECT FName FROM t1 WHERE FName >=` `-\ensuremath{\infty}\xspace and FName < Ella` where $-\ensuremath{\infty}\xspace$ is a placeholder for the smallest domain value. Next, the proxy derives $SK_{D}$ using $SK_{DB}$, the table name, and the column name. Then, it encrypts the range start and end ($R_s$ and $R_e$) with $\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}$ using random initialization vectors. The resulting encrypted query $eQ{}$ of our SQL example is `SELECT FName FROM t1 WHERE FName >= \mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}(\allowbreak{}\ensuremath{SK_{D}}\xspace,\allowbreak{}\ensuremath{IV_{1}}\xspace,\allowbreak{}\ensuremath{-\ensuremath{\infty}\xspace}\xspace{})and FName <` `\mathtt{PAE{}\_}\allowbreak{}\mathtt{Enc}(\allowbreak{}\ensuremath{SK_{D}}\xspace,\allowbreak{}\ensuremath{IV_{2}}\xspace,\allowbreak{}\texttt{\textquotesingle{}Ella\textquotesingle{}})`. Because of the query conversion, the untrusted DBaaS provider cannot differentiate query types, and due the utilization of probabilistic authenticated encryption, it also cannot learn if the values were queried before. EncDBDB could also handle other query functionalities, e.g., counts, aggregations, and average calculations, but we do not consider these in this paper, because they are easier to support than range searches. Other researchers already presented encrypted joins [@arasu2013oblivious; @hahn2019; @li2008privacy] and it is an interesting future work to support joins while using encrypted dictionaries. $eQ{}$ is passed to the query pipeline of the DBaaS provider that is specific to the underlying DBMS. For instance, the query is processed by a *query parser*, a *query decomposer* and a *query optimizer*. The query optimizer selects a query plan and shares it with a *query evaluation engine*. It contains one $\big(\ensuremath{eD{}}\xspace{}, \ensuremath{AV{}}\xspace{}, \ensuremath{\tau}\xspace{}\big)$ tuple that is derived from $eQ{}$, i.e., an encrypted dictionary, a plaintext attribute vector and an encrypted range filter that has to be executed. The query evaluation engine enriches $eD{}$ with metadata: the table name, the column name, and the column size. Then, it passes $\tau$ and a reference to $eD{}$ to the enclave. The enclave performs the $\mathtt{EnclDict}\-\mathtt{Search}$ operation corresponding to the encrypted dictionary of the filtered column. During this search the necessary dictionary entries are loaded from the untrusted realm. Finally, it returns a list of ValueIDs ($\mathbf{vid{}}$) for which the corresponding values fall into $R$. The query evaluation engine performs the $\mathtt{AttrVect}\-\mathtt{Search}$ operation corresponding to the encrypted dictionary of the filtered column. These steps result in a list of RecordIDs ($\mathbf{rid{}}$). $\mathbf{rid{}}$ is passed to a *result renderer*, which would use $\mathbf{rid{}}$ to prefilter other columns in the same table if a filter query should be executed on them. Additionally, $\mathbf{rid{}}$ would be used if an unconditional select is performed on another column. One encrypted result column $eC{}$ is created by undoing the split in $eD{}$ and $AV{}$ on all entries in $\mathbf{rid{}}$, i.e., $\ensuremath{eC{}}\xspace{} = (\ensuremath{eD{}_{j}}\xspace \, | \, j = \ensuremath{AV{}_{i}}\xspace \wedge{} i \in \ensuremath{\mathbf{rid{}}}\xspace{})$. The result renderer enriches $eC{}$ with column metadata — table and column name — and passed $eC{}$ back to the proxy. The proxy receives one encrypted column $eC{}$ from the DBaaS provider and uses the attached column metadata to derive the column specific key $SK_{D}$. Every entry in $eC{}$ is decrypted individually with $SK_{D}$ resulting in one plaintext column $C{}$, which is passed back to the application. Notably, only a very small part of the query processing is done inside the trusted enclave and the required enclave memory is very limited. There is no need to modify auxiliary database functionalities such as persistency management, multiversion concurrency control or access management. Still, the complete processing is protected. Dynamic Data {#subsec:02:main:dynData} ------------ So far, we only discussed static data, which are prepared by the data owner before being uploaded to an EncDBDB-enabled DBaaS provider. This is sufficient for most analytical scenarios, because bulk loading of data is often used in this context and complex, read-only queries are executed afterwards [@harizopoulos2006performance; @stonebraker2005c]. For other usage scenarios, we present an approach on how EncDBDB can support dynamic data, i.e., data insertions, deletions, and updates. We propose to utilize a concept called *delta store* (or differential buffer): the database — more specifically each column — is split into a read optimized *main store* and a write optimized delta store (see [@farber2012sap; @hubner2011cost; @stonebraker2005c] for more details). Updates in a column do not change existing rows. Instead, all data changes are performed in the delta store. New values are simply appended. Updated values are handled using a validity vector for the two storage concepts. This vector stores stores a flag for each entry indicating whether or not it is valid. Deletions are realizable by an update on the validity bit. The overall state of the column is the combination of both stores. Thus, a read query becomes more complex: it is executed on both stores normally and then the results are merged while checking the validity of the entries. The delta store should be kept orders of magnitude smaller than the main store to efficiently handle read queries. This is done by periodically merging the data of the delta store into the main store. Hübner et al. describe different merging strategies [@hubner2011cost]. For EncDBDB, any encrypted dictionary can be used for the main store and ED9 should be employed for the delta store. New entries can simply be appended to a column of type ED9 by reencrypting the incoming value inside the enclave with a random $IV$. A search in this delta store is done by performing the linear scan as defined by $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{9}$. As a result, neither the data order nor the frequency is leaked during the insertion and search. A drawback of ED9 is that it has a high memory space overhead and low performance. However, the periodic merges mitigate this problem. The enclave handles the merging process as follows: First, it reencrypts every value in $D{}$. Then, the columns with the rotated order option are randomly rerotated and columns with the unsorted order option are reshuffle. The process has to be implemented in a way that does not leak the relationship between values in the old and new main store, e.g., with oblivious memory primitives [@zeroTrace; @zheng2017opaque]. Implementation {#sec:02:implementation} ============== For our experiments we implemented a prototype based on MonetDB, an open-source, column-oriented in-memory DBMS [@boncz_breaking_2008; @idreos_monetdb:_2012; @boncz2005monetdb]. MonetDB focuses on read-dominated, analytical workloads and thus perfectly fits our use case. It is a commercial relational DBMS, which exploits the large main memory of modern computer systems for processing and it uses disk storage for persistency. MonetDB uses a variant of dictionary encoding for all string columns. The attribute vector contains offsets to the dictionary, but the dictionary contains data in the order it is inserted (for non-duplicates). The dictionary does not contain duplicates if it is small (below ) and a hash table and collision lists are used to locate entries. The collision list is only used as long as the dictionary does not exceed a certain size. As a result, the dictionary might store values multiple times. The front-end query language of MonetDB is SQL. We implemented the nine encrypted dictionaries as SQL data types in the frontend and new internal data types in the backend. The encrypted dictionaries can be used in SQL create table statements like any other data type, e.g., `CREATE TABLE t1 (c1 ED7, c2 ED5, …)`. We further split each dictionary into a dictionary head and dictionary tail. The dictionary tail contains variable length values that are encrypted with AES-128 in GCM mode. The values are stored sequentially in a random order. The dictionary head contains fixed size offsets to the dictionary tail and the values are ordered according to the selected encrypted dictionary. This split is done to support variable length data while enabling an efficient binary search. For dictionary search, we pass a pointer to the encrypted dictionary into the enclave and it directly loads the data from the untrusted host process. Thus, only one context switch is necessary for each query. Furthermore, all operations mentioned as easily parallelizable run parallel in our implementation. Evaluation {#sec:02:evaluation} ========== In this section, we first provide security, storage and performance evaluations of our nine encrypted dictionaries. Based on those, we conclude the section with a usage guideline regarding the different encrypted dictionaries. Security Evaluation {#subsec:02:evaluation:secEval} ------------------- We start this section with a short discussion about enclave code size. In general, small enclave code size improves the security, as it reduces the probability of security-relevant implementation errors, unintended leakages, and hidden malware. Our enclave is written in C and besides the Intel SGX SDK (version 2.5), has only 1129 lines of code (LOC). Only 412 of those LOC are written by us, the remainder are taken up by a big integer library [@bigInteger] used for the dictionary search in ED2, ED5 and ED8. An enclave of this size can be efficiently verified by a user of EncDBDB. Now, we discuss the security of the nine encrypted dictionaries under the attacker model defined in Section \[subsec:02:designOverview:attackerModel\], i.e., an honest-but-curious attacker that targets each column independently. The attacker passively examines the processing of an encrypted dictionary $eD{}$ and an attribute vector $AV{}$ in multiple rounds and she knows which encrypted dictionary is used. First, we describe the security of ED1–ED3 and ED7–ED9 by comparing them with security schemes known in literature (see Table \[tab:02:securityClassification\]). A detailed analysis of the different security definitions is beyond the scope of this paper as it is highly data-dependent. However, we reference known attacks in Table \[tab:02:securityClassification\]. Afterwards, we describe the security of ED4–ED6 relative to the other encrypted dictionaries. The relation between the security provided by the different encrypted dictionaries is summarized in Figure \[fig:02:encDTsSecurity\]. ----- ------ --------- --------------------------------------------- ----------------------------------------------------------------------------- ED1 full full ideal, determ. ORE [@boneh2015semantically] [@TaoInference; @grubbs2016leakage] ED2 full bounded MOPE [@boldyreva_ope_2011] [@TaoInference; @grubbs2016leakage; @MOPE_revisited] ED3 full none DET [@bellare2007deterministic] [@TaoInference; @naveed_inference_2015] ED7 none full IND-FAOCPA [@kerschbaum2015frequency] [@GrubbsVolume; @GuiRangeQueryAttacks; @KellarisAttacksOutsourcedDatabases] ED8 none bounded IND-CPA-DS [@Kerschbaum17] [@GrubbsVolume; @GuiRangeQueryAttacks; @KellarisAttacksOutsourcedDatabases] ED9 none none RPE [@lu_privacy-preserving_2012] [@GrubbsVolume; @GuiRangeQueryAttacks; @KellarisAttacksOutsourcedDatabases] ----- ------ --------- --------------------------------------------- ----------------------------------------------------------------------------- : Security of ED1–ED3 and ED7–ED9[]{data-label="tab:02:securityClassification"} ![image](figures2/encDTsSecurity){width="\linewidth"} \[fig:02:encDTsSecurity\] ED1 provides protection that is comparable to an ideal, deterministic variant of order-revealing encryption (ORE) [@boneh2015semantically]. It is an ORE as the order of the values is revealed by a public — the dictionary. It is ideal as neither $eD{}$ itself nor the encrypted values leak anything but the order. It is deterministic as equal plaintexts have the same ciphertext. The security of modular OPE (MOPE) [@boldyreva_ope_2011] is a lower bound for ED2. As MOPE, a column protected with ED2 only leaks the order of the values. MOPE uses deterministic OPE and ED2 uses deterministic ORE, which is more secure. ED3 has no order leakage, but fully leaks the frequency of all values. These security features are equivalent to a column protected with deterministic encryption (DET) [@bellare2007deterministic]. A column protected with ED7 is IND-FAOCPA secure [@kerschbaum2015frequency]. Each ciphertext is present exactly once in $eD{}$ and the assignment of each $AV{}$ entry to a ValueID is done with the help of a if a plaintext is encrypted multiple times. Thus, the ValueIDs in $eD{}$ form a randomized order (see definition [@kerschbaum2015frequency]) of the plaintext values. ED8 protects a column with IND-CPA-DS security [@Kerschbaum17]. $\mathtt{EncDB\_}\allowbreak{}\mathtt{8}$and the `Enc` algorithm in [@Kerschbaum17] are different, but the results of the algorithms are equal. Furthermore, $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{8}$matches the `Search` algorithm in [@Kerschbaum17]. Therefore, the attacker learns the same information during processing. The security of a column protected with ED9 is comparable to the security of Range Predicate Encryption (RPE) [@lu_privacy-preserving_2012]. As defined by RPE’s plaintext privacy, $\mathtt{EnclDict}\-\mathtt{Search\_}\allowbreak{}\mathtt{9}$and $\mathtt{AttrVect}\-\mathtt{Search\_}\allowbreak{}\mathtt{9}$only leak the information that an entry falls into the search range. The of ED9 are plaintexts encrypted with $\mathtt{PAE{}}$ using a random initialization vector, which provides RPE’s predicate privacy. The frequency smoothing algorithm used by ED4 makes the ciphertext frequencies close to uniform by randomly selecting a frequency between $1$ and $bs{}_{\max}$, independent of the plaintext frequency. As ED1 fully leaks the ciphertext frequency and ED7 hides it completely, the security of ED4 lies between the security of ED1 and ED7. ED5 is more secure than ED2 and is less secure than ED8 for the same reason. The same is true for the triple ED6, ED3 and ED9. The frequency smoothing algorithm is based on an algorithm described in [@Pouliot17] and the authors only state that the last frequency is not selected from the same distribution, which might give an advantage to an attacker. An in-depth security evaluation is an open research question. Storage Evaluation {#subsec:02:evaluation:storageEval} ------------------ For our storage evaluation, we use a snapshot of a real-world SAP customer’s business warehouse (BW) system. The largest columns contain $168.7$ million data values. To evaluate the influence of the number of unique values to our algorithms, we search for columns having the same size, but different distributions. The dataset contains 30 large columns with $10.9$ million values. We present the results for two extreme cases: C1 with 6.96 million unique values and C2 with . Table \[tab:02:fileSizes\] presents the storage space requirements of different variants. The plaintext file contains all plaintext values present in the column without any compression. This file is comparable to a plaintext column for which dictionary encoding is not used. The encrypted file contains every value from the plaintext file, but individually encrypted with $\mathtt{PAE{}}$, which has the same storage requirements as an encrypted column without dictionary encoding. MonetDB’s storage requirements are presented as a baseline. size C1 size C2 -------------------------------------------------------- --------- --------- -- Plaintext file Encrypted file MonetDB ED1/ED2/ED3 ED4/ED5/ED6, $\ensuremath{bs{}_{\max}}\xspace{} = 100$ ED4/ED5/ED6, $\ensuremath{bs{}_{\max}}\xspace{} = 10$ ED4/ED5/ED6, $\ensuremath{bs{}_{\max}}\xspace{} = 2$ ED7/ED8/ED9 : Storage size of various variants[]{data-label="tab:02:fileSizes"} The size of the plaintext files decreases from C1 to C2, because the strings in these columns are 12 and 10 characters long. As expected, we see that EncDBDB requires less space if fewer unique values are present. We see that for C2 protected with ED1, ED2, or ED3, EncDBDB requires less storage space than the plaintext file, i.e., less space than a plaintext column without dictionary encoding. We also see a further expected behavior: a smaller $bs{}_{\max}$ increases the required storage space as more duplicates are stored. Note that the encrypted dictionaries are stored outside of the enclave and individual values are loaded and decrypted. Hence, the restricted enclave space does not constitute a limitation for EncDBDB. Performance Evaluation {#subsec:02:evaluation:perfEval} ---------------------- For the performance evaluation, we use the same columns introduced in the storage evaluation. Besides the original columns, which we call *full datasets*, we sample datasets from $1$ to $10$ million records using the distribution and values of the original columns. MonetDB is used as one baseline measurement in our experiments to compare ourselves against a commercial plaintext DBMS. Additionally, we implement *PlainDBDB* — a plaintext variant of EncDBDB. PlainDBDB uses the same algorithms as EncDBDB, but the dictionaries are plaintext and the algorithms are processed without an enclave. We use PlainDBDB as a second baseline to evaluate the performance overhead of encryption and SGX. All experiments are performed with the confidential computing offering of Microsoft Azure. We use a DC4s machine with RAM and 4 vCPU cores of an Intel Xeon E-2176G CPU @ 3.70GHz. All presented latencies measure the processing time spent at the server excluding any network delay or processing at the proxy or client. Our protocol runs in one round and only encrypts the values in the query. Thus, the communication and latency overhead compared to any database in the cloud is negligible. We use the term *range size* ($RS$) to describe how many consecutive unique values from the dataset are searched in a range query, i.e., if $sorted(\ensuremath{un(C{})}\xspace{}) = (\allowbreak{}\ensuremath{v_{0}}\xspace,\allowbreak{}\ldots{},\allowbreak{}\ensuremath{v_{\ensuremath{\abs{un(C{})}}\xspace - 1}}\xspace ) $ is a sorted list of all unique values in $C{}$, then $RS$ defines the search range $\ensuremath{R}\xspace{} = [\ensuremath{v_{i}}\xspace, \ensuremath{v_{i + \ensuremath{RS}\xspace{} - 1}}\xspace]$ for $i \in [0, \ensuremath{\abs{un(C{})}}\xspace{} - \ensuremath{RS}\xspace{}]$. For every dataset and encrypted dictionary, we perform 500 random range queries with range sizes 2 and 100. The same random range queries are executed for MonetDB, PlainDBDB, and EncDBDB. Note that the number of result rows returned by the server is greater than $RS$ if a value in the search range is present multiple times in the column (see Figure \[fig:02:returnedTuples\]). For instance, values are returned on average for the full dataset of C2 and $\ensuremath{RS}\xspace{} = 100$. ![Average number results returned by 500 random range queries for columns C1 and C2 (95% confidence interval; note that logarithmic y-axis distorts error bars)[]{data-label="fig:02:returnedTuples"}](figures2/C1,2_tuples) ![image](figures2/C1,2_ED1,2,3,4,5,6,7,8,9_times_used_2){width="\linewidth"} **ED1.** The first and fourth column in Figure \[fig:02:perfEval\] (a) present the latencies of ED1 for C1 and C2 and the range sizes 2 and 100. We highlight three observations from these plots. First, EncDBDB and PlainDBDB outperform MonetDB for both range sizes at both columns. The main reason is that MonetDB’s attribute vector search performs a linear number of string comparisons. In contrast, EncDBDB and PlainDBDB require only a logarithmic number of string comparisons in the dictionary search and a linear number of integer comparisons in the attribute vector search. Second, EncDBDB slows down if a column with equal size has less unique values: the average latencies increase from at C1 to at C2 for the full dataset and $\ensuremath{RS}\xspace{} = 100$. This seems counterintuitive, because fewer unique values result in a smaller dictionary size ($\abs{D{}}$) improving the dictionary search performance. However, only logarithmically fewer decryptions and string comparisons are necessary in the dictionary search, but many results are returned by the attribute vector search (see Figure \[fig:02:returnedTuples\]). As a result, the DBMS has to spend more time for tuple-reconstruction, i.e., to build the result set based on the found RecordIDs and the dictionary. Third, encryption is cheap: the average latency overhead of EncDBDB compared to PlainDBDB is (). The overhead is minor for two reasons: (1) as explained in the implementation section, we only requires one context switch per column, which is negligible in the overall latency and (2) we only use hardware-supported AES-GCM encryption. **ED2.** The second and fifth column in Figure \[fig:02:perfEval\] (a) present the latencies of ED2. The main observation is that the latency of EncDBDB and PlainDBDB is almost equal to the latency of ED1 for the two columns. The only difference between ED1 and ED2 is that ED2 uses a special binary search and post-processing of the resulting ValueIDs to handle the random rotation, which introduces only a minor overhead. In fact, the average latency overhead from ED1 to ED2 is for EncDBDB. **ED3.** The third and sixth column in Figure \[fig:02:perfEval\] (a) show the latencies of ED3. We observe that the average latencies of PlainDBDB and EncDBDB, and their relative latency differences, severely depend on the number of unique values and the range size ($RS$). C2 has a smaller $\abs{D{}}$ than C1, which decreases the latency of the linear dictionary search and therefore the average latency of the query execution. Additionally, a smaller $\abs{D{}}$ decreases the number of necessary decryptions for EncDBDB and therefore the relative latency difference between PlainDBDB and EncDBDB. **ED4, ED5, ED6.** Figure \[fig:02:perfEval\] (b) presents the latency plots for ED4–ED6. The latencies of MonetDB obviously do not change. In the following, we focus on EncDBDB discussing the latencies for ED4–ED6 compared to ED1–ED3. $\abs{D{}}$ is larger for ED4–ED6, because the frequency smoothing algorithm adds duplicates to $D{}$ ($\ensuremath{bs{}_{\max}}\xspace{} = 10$ in our experiments). For ED4 and ED5, $\abs{D{}}$ influences the latency only logarithmically. Compared to ED1 and ED2, the average overheads are only and , respectively. At ED6, the dictionary search might return more than $x$ ValueIDs for the range size $x$ as $eD{}$ contains duplicate plaintexts. Every returned value has to be compared to each attribute vector entry. This increases the average latencies for the full dataset at $\ensuremath{RS}\xspace{} = 100$ to and for C1 and C2. **ED7, ED8, ED9.** Figure \[fig:02:perfEval\] (c) presents the latency plots for ED7–ED9. We again focus on EncDBDB’s latency in ED7–ED9 compared to ED1–ED3. Compared to ED1 and ED2, the average overheads of ED7 and ED8 are and , respectively. For the full dataset at $\ensuremath{RS}\xspace{} = 100$, the average latencies of ED9 increase to and for C1 and C2, respectively. Usage Guideline {#subsec:02:evaluation:usageGuidelines} --------------- According to the security sensitivity of the data owner, an encrypted dictionary can be select per column. If plaintext is not an option, but the weakest security level is acceptable, ED1 can be used. It has a small storage size and it is almost as fast as PlainDBDB, even with different range sizes and unique value amounts. If order leakage should be reduced and a minor performance overhead is acceptable, ED2 is preferable over ED1. If order leakage is not acceptable, a column contains few unique values, and $RS$ is small, ED3 has a practical overhead. For instance, EncDBDB’s average latency overhead from ED1 to ED3 for C2 and $\ensuremath{RS}\xspace{} = 2$ is . If the frequency leakage should be bounded, ED5 can be used with a minor performance and storage overhead compared to ED2. In many cases, ED5 is the best security, latency and storage tradeoff among our encrypted dictionaries. If security and latency are critical, but not storage size, ED8 is the most favorable encrypted dictionary. If security is the main objective of a column, ED9 should be used. Related Work {#sec:02:relatedWork} ============ In this section, we compare EncDBDB to TEE-based encrypted databases, software-only encrypted databases, and searchable encryption. TEE-based Encrypted Databases {#subsec:02:relatedWork:teeBased} ----------------------------- In the following, we outline TEE-based approaches ranging from large to small enclave sizes, and classify EncDBDB accordingly. Haven [@baumann_shielding_2014] and SCONE [@scone] are approaches to shield complete applications on an untrusted system using SGX. Unmodified applications should be executable inside an SGX enclave, which could also be an off-the-shelf DBMS. However, a complete DBMS with millions of lines of code is prone to security-relevant implementation errors or side-channel leakages that could leak arbitrary data from the enclave. Furthermore, no TEE on the market does support the huge enclaves that are necessary for this concept. Priebe et al. proposed EnclaveDB [@EnclaveDB], a protected database engine that uses a TEE to provide confidentiality, integrity, and freshness for OLTP workloads. EnclaveDB has a large TCB, as the tables, indices, metadata, query engine, transaction manager and stored procedures are loaded into the TEE. The problems described for Haven and SCONE are only slightly less severe. Especially, it still does not fit into an existing TEE. Furthermore, all possible queries have to be known in advance. ObliDB [@ObliDB18] is an SGX based encrypted database that hides the access pattern using oblivious query processing algorithms on a $B^+$-tree index or a linear array. The additional protection introduces a latency overhead of 240% compared to a plaintext database. Additionally, ObliDB lacks transaction management and disk persistency. HardIDX [@fuhry2017hardidx] uses SGX to protect one specific data structure, a $B^+$-tree. Equality and range searches are done inside the enclave and either the whole dataset at once or parts on demand are loaded into enclave memory. Only a few megabytes of enclave memory are necessary, and the enclave has only a few lines of code. However, a $B^+$-tree is only presented as a building block of an encrypted database. TrustedDB [@bajaj2014trusteddb], Cipherbase [@transactioCipherbase], and StealthDB [@gribov2017stealthdb] use a secure co-processor, an FPGA, and SGX as TEE, respectively. They have the smallest enclave size by putting the execution of individual operators, e.g., $<$, $>$, and $=$ into a TEE. The operations are executed on encrypted data and the results are passed back. Only minor changes to an application (e.g., a database) are necessary as plaintext operations are just replaced by protected operators. However, much information is leaked as an attacker learns the result of each operation. Only the processing of data structures, e.g., dictionaries or $B^+$-trees, is protected, but the authors do not consider the inherently leaked information about the relation of individual data values. EncDBDB follows the same design philosophy as HardIDX: keep the enclave code and the required enclave memory as small as possible without leaking every individual decision by processing a data structure inside an enclave. As a main difference, we integrate EncDBDB into a DBMS. Software-Only Encrypted Databases {#subsec:02:relatedWork:encDB} --------------------------------- Software only encrypted databases, such as CryptDB [@popa_cryptdb:_2011] and Monomi [@Monomi], use property-preserving encryption for efficient search. Every database functionality requires its own encryption scheme with additional storage overhead. For instance, deterministic encryption [@bellare2007deterministic] is used to support equality selects, and OPE [@agrawal_order_2004; @boldyreva_ope_2009; @boldyreva_ope_2011; @kerschbaum_optimal_2014] allows range queries. Naveed et al. [@naveed_inference_2015] presented practical ciphertext-only attacks on property-preserving encryption and further attacks followed [@durak2016else; @grubbs2016leakage]. In EncDBDB, equality and range queries are handled by one encryption scheme with a small performance and storage overhead. Some encrypted dictionaries of EncDBDB are affected by these attacks, but the data owner can freely choose a security level that fits his requirements and all functionality is still supported. Other approaches for a secure DBMS allowing range query evaluation have been published: Cash et al. [@cash2013highly] introduce a protocol that allows evaluation of boolean queries on encrypted data. Faber et al. [@faber2015rich] extend this protocol to support range queries but either leak additional information on the queried range or the result set contains false positives. Pappas et al. [@pappas2014blind] evaluate encrypted bloom filters using secure multiparty computation. However, in order to achieve practical efficiency, they propose to split the server into two non-colluding parties. Egorov et al. [@egorov_zerodb_2016] presented ZeroDB, a database that enables a client to perform equality and range searches with the help of $B^+$-trees. It uses an interactive protocol requiring many rounds and thus is not usable for network-sensitive cloud computing. EncDBDB does neither require an additional party nor multiple rounds. Searchable Encryption {#subsec:02:relatedWork:searchEnc} --------------------- Song et al. introduced the first searchable encryption schemes for single plaintexts [@song_practical_2000]. In order to improve performance, Goh [@goh_secure_2003] and Curtmola et al. [@curtmola_searchable_2006] introduced encrypted (inverted) indices. However, these encryption schemes can only search for keyword equality and not ranges. The first range-searchable scheme by Boneh and Waters encrypts every entry linear in the size of the plaintext domain [@BW07]. The first scheme with logarithmic storage size per entry in the domain was proposed by Shi et al. in [@shi_multi-dimensional_2007]. Their security model (match-revealing) is somewhat weaker than standard searchable encryption. The construction is based on inner-product predicate encryption which has been made fully secure by Shen et al. in [@shen_predicate_2009]. All of these schemes have linear search time. Lu built the range-searchable encryption from [@shen_predicate_2009] into an index in [@lu_privacy-preserving_2012], thereby enabling polylogarithmic search time. Hahn and Kerschbaum proposed a construction building the search index incrementally [@hahn2016poly]. By doing so, they still support amortized polylogarithmic search time but increase security properties for non-queried values. Demertzis et al. presented multiple constructions that improve the constant factor of a range search [@demertzis_2016]. However, their construction without prohibitive storage cost and false positives (Logarithmic-URC) requires already more than a second to perform a range search within values [@fuhry2017hardidx]. EncDBDB operates on millions of entries in milliseconds. Conclusion {#sec:02:conclusion} ========== In this paper, we introduced EncDBDB — a high-performance, encrypted cloud database supporting analytic queries on large datasets. EncDBDB provides nine different encrypted dictionaries with distinct security, performance and storage efficiency tradeoffs. Even with no frequency leakage and bounded order leakage, range queries on datasets with millions of encrypted entries are executed within milliseconds. If some frequency leakage is acceptable, the compressed encrypted data requires less space than a plaintext column. Moreover, the TCB of EncDBDB consists only of 1129 lines of code exposing only a small attack surface. With those features, EncDBDB is ideally suited for an entity that wants to outsource its sensitive data to an untrusted cloud environment.
{ "pile_set_name": "ArXiv" }
--- abstract: 'It is shown that the power spectrum defined in the Synchronous Gauge can not be [*directly*]{} used to calculate the predictions of cosmological models on the large-scale structure of universe, which should be calculated [*directly*]{} by a suitable gauge-invariant power spectrum or the power spectrum defined in the Newtonian Gauge.' author: - | Wenbin Lin\ Institue of High Energy Physics, Chinese Academy of Science, P.O. Box 918-4, Beijing 100039, P.R. China title: Remarks on the Theory of Cosmological Perturbation ---  \  \ :   98.80.-k; 98.80.Hw; 98.80.Cq  \  \ It is indispensable to adopt the linearized Einstein equation to solve the problem of the evolution of density perturbation in the expanding universe, and this involves the freedom of gauge. Currently, the density perturbation is widely calculated in the Synchronous Gauge ($SG$) frame and then [*directly*]{} compared with the observations of the large-scale structure of universe, e.g., the abundance and distribution of the galaxies or the clusters (e.g., Refs. [@lss; @bardeen1; @white; @dodelson; @gawiser]). Because the power spectrums of density perturbation differ dramatically for small wavenumbers in different gauges, this raises an important problem: Does the density perturbation in the $SG$ frame characterize [*directly*]{} the large-scale structure of universe which is relative to the unperturbed expanding background of universe? I do not think so. Because the density perturbations defined in different gauges are generally different from each other, we need to construct some suitable gauge-invariant variables that can be directly confronted with the large-scale structure of universe. The gauge-invariant approach to the gravitational perturbation, which was pioneered by Bardeen[@bardeen2], has been applied to the cosmological perturbation (e.g., Refs.[@kodama1; @mukhanov]). However, there exist some errors in the comparison between the theoretical calculations and the observations of the large-scale structure of universe. The most general form of the line element for a spatially flat background and scalar metric perturbations can be written as [@mukhanov] $$ds^2=a^2\{(1+2\Psi)d\tau^2-2B_{\mid i}dx^id\tau-[\delta_{ij} -2(\Phi\delta_{ij}-E_{\mid_{ij}})]dx^idx^j~\},$$ where $a$ and $\tau$ are the conformal cosmic expansion scale factor and the conformal cosmic time; “$_\mid$” denotes the background three-dimensional covariant derivative. The corresponding perturbed energy-momentum tensor $T^{\mu}_{\nu}$ has the form $$\begin{aligned} \nonumber T^0_{~0}&=&\rho(1+\epsilon)~,\\ \nonumber T_0^{~i}&=&(\rho+p)U_{\mid i}~,\\ \nonumber T^0_{~i}&=&-(\rho+p)(U-B)_{\mid i}~,\\ T^i_{~j}&=&-p(1+\varepsilon)\delta^i_{~j}-(\rho+p) \Sigma_{\mid ij}~,\end{aligned}$$ here $\rho$ and $p$ are unperturbed density and pressure; $U$ and $\Sigma$ determine velocity perturbation and anisotropic shear perturbation, respectively; $\epsilon$ and $\varepsilon$ denote the fluctuation for density and pressure respectively. In many gauge-invariant approaches, the gauge-invariant density perturbation was constructed by [@bardeen2; @kodama1] $$\label{epsilonm} \epsilon_m\equiv \epsilon+3(1+w) \frac{\dot{a}}{a}(U-B)~,$$ where a dot denotes the derivative with respect to the conformal time $\tau$; $w=p/\rho$ is the ratio of the pressure to the density of background. $\epsilon_m$ coincides with the density perturbation $\epsilon_{(CTG)}$ in the Comoving Time-orthogonal Gauge ($CTG$, in which $U=B=0$), which denotes the density perturbation relative to the spacelike hypersurface which represents the matter local rest frame everywhere[@bardeen2]. This quantity also coincides with the density perturbation $\epsilon_{(SG)}$ in the Synchronous Gauge (in which $\Psi=B=0$) for the pressureless matter system. In other words, $\epsilon_m$ denotes the density perturbation relative to the observers everywhere comoving with the matter ([*not*]{} with the unperturbed expanding background of universe because the matter have local velocity other than the Hubble flow). So, $\epsilon_m$ has physical significance only for the small-scale perturbation everywhere, and this can be also derived from following facts: The relation between $\epsilon_m$ and the general gravitational potential $\phi$ (which was correctly constructed in literatures, see Eq.(\[potential\]) below) obeys Poisson equation[@bardeen2; @kodama1]: $$\label{poisson} \bigtriangledown^2\phi=-k^2\phi =4\pi G\rho a^2\epsilon_m~.$$ where $k$ is the (comoving) wavenumber of Fourier mode. Poisson equation is valid only in the case that the investigated scale is small compared to the Hubble radius $1/H$ with $H=\dot{a}/a^2$ being Hubble constant [^1][@lss]. Hence, $\epsilon_m$ can not be regarded as the physical density perturbation directly for the large-scale (comparable to or larger than the Hubble radius) perturbation mode in the expanding universe because the gravitational potential $\phi$ has been constructed correctly. Another fact is that the linearized Einstein equation shows that $\epsilon_m$ keeps on growing even for the perturbation modes outside both the Hubble radius and the horizon [@bardeen2; @kodama1], which also means that $\epsilon_m$ can not directly characterize the physical large-scale perturbation modes. Therefore, the physical density perturbation which comprises both the small-scale and the large-scale modes can not be described [*directly*]{} by $\epsilon_m$, though it is a gauge-invariant quantity. In fact, the large-scale density perturbation has physical significance only in the case that the corresponding density perturbation everywhere is relative to the unperturbed expanding background of universe. So we must construct a proper gauge-invariant variable to describe the physical density perturbation [*directly*]{}, and then confront it with the large-scale structure of universe. It is my viewpoint that the gauge-invariant quantity constructed by [@bardeen2; @kodama1; @mukhanov], $$\label{gidp} \epsilon_g\equiv \epsilon+\frac{\dot \rho}{\rho}(B-\dot{E})= \epsilon-3(1+w)\frac{\dot{a}}{a}(B-\dot{E})~.$$ is just what we sought for. $\epsilon_g$ coincides with the density perturbation $\epsilon_{(NG)}$ in the Newtonian Gauge ($NG$, in which $B=E=0$). Correspondingly, two gauge-invariant scalar potentials $\phi$ and $\psi$, both of which become the same as the gravitational potential in the Newtonian limit, are constructed from metric perturbations [@mukhanov]: $$\begin{aligned} \label{potential} \nonumber && \phi \equiv \Phi-\frac{\dot{a}}{a}(B-\dot{E})~,\\ && \psi \equiv \Psi+\frac{1}{a}\frac{d}{d\tau}[(B-\dot{E})a]~.\end{aligned}$$ The gauge-invariant pressure perturbation $\delta p/p$ and velocity perturbation $v_i$ are constructed as $$\label{p} \delta p/p\equiv \varepsilon+\frac{\dot p}{p} (B-\dot {E})~,$$ $$\label{v} v_i\equiv (U-\dot{E})_{|i}~.$$ The time-time part of the linearized Einstein equation gives [@mukhanov; @ma] $$\label{zero} \bigtriangledown^2\phi-3\frac{\dot{a}}{a}(\frac{\dot{a}}{a}\psi+\dot{\phi}) =-k^2\phi-3\frac{\dot{a}}{a}(\frac{\dot{a}}{a}\psi+\dot{\phi}) =4\pi G\rho a^2\epsilon_g~.$$ From this equation we can see that the relation between the potential $\phi$ and the physical density perturbation $\epsilon_g$ does not obey Poisson equation any more for the large-scale perturbation modes. At the same time, it can be deduced from the linearized Einstein equation that $\epsilon_g$ hardly grows when the perturbation mode is outside the Hubble radius ($k<aH$). Hence, though any gauge can be employed to work for the density perturbation, We should adopt the gauge-invariant variable which characterizes the density perturbation relative to the unperturbed expanding background of universe to compare with the large-scale structure of universe [ *directly*]{} (here I do not consider the biasing issue). For the multi-component system, $\epsilon_g$, $\delta p/p$ and $v_i$ denote the gauge-invariant variables of total density perturbation, total pressure perturbation and total velocity perturbation respectively, and the corresponding gauge-invariant variables for $\alpha$-component are constructed in similar ways [@kodama1]: $$\epsilon_{\alpha g}\equiv \epsilon_{\alpha}+\frac{\dot \rho_{\alpha}}{\rho_{\alpha}}(B-\dot{E})~,$$ $$\label{pa} (\delta p/p)_{\alpha} \equiv \varepsilon_{\alpha}+ \frac{\dot p_{\alpha}}{p_{\alpha}} (B-\dot {E})~,$$ $$\label{va} v_{\alpha i}\equiv (U_{\alpha}-\dot{E})_{|i}~.$$ From above equations we can see that all these gauge-invariant perturbation variables conincide with the corresponding ones in the Newtonian Gauge respectively. We can also adopt the density perturbation defined in this gauge to compare with the large-scale structure of universe [*directly*]{}. Any a real-space fluctuation $\delta(\bf{x},\tau)$ can be decomposed into Fourier modes $$\label{delta} \delta({\bf{x}},\tau)=\frac{1}{(2 \pi)^{\frac{3}{2}}} \int \delta_{\bf k}(\tau) e^{i\bf{k\cdot x}}d^3k~,$$ with the reality condition $\delta_{\bf k}^{\ast}=\delta_{-\bf{k}}$ . The power spectrum is defined as the mean square of the corresponding Fourier-mode amplitude: $$P_{\delta}(k,\tau)=<|\delta_{\bf k}(\tau)|^2>\equiv |\delta(k,\tau)|^2~.$$ All the statistical properties of the Gaussian random field can be determined by the power spectrum completely, and for simplicity we only consider this case here. At the same time, we are only interested in the fluctuation relative to the unpertured background everywhere which can be confronted with the large-scale structure of universe directly. The root mean square (rms) fluctuation and two-point correlation function of density on the scale $R$ can be calculated by the physical power spectrum $P_{\epsilon_g}(k,\tau)\equiv |\epsilon_g(k,\tau)|^2$ directly $$\begin{aligned} \label{sigma} \nonumber \sigma_{\epsilon_g}(R,\tau)&=&\left[\frac{1}{2\pi^2} \int_0^{\infty}P_{\epsilon_g}(k,\tau) W^2(kR)k^3\frac{dk}{k}\right]^{\frac{1}{2}}~,\\ \xi_{\epsilon_g}(R,\tau)&=&\frac{1}{2\pi^2} \int_0^{\infty}P_{\epsilon_g}(k,\tau) \frac{\sin kR}{kR}k^3\frac{dk}{k}~,\end{aligned}$$ where, $W(x)=\frac{3}{x}j_1(x)$ is the window function which truncates the contribution of the power spectrum of large wavenumbers to the rms fluctuation (here $j_1$ is the first-order spherical Bessel function). When calculating $\sigma_{\epsilon_g}(R,\tau)$ or $\xi_{\epsilon_g}(R,\tau)$, we need to integrate the power spectrum over all wavenumbers, not merely the modes within the Hubble radius (see Eq.(\[sigma\])). Because the spectrum $P_{\epsilon_m}(k,\tau)\equiv |\epsilon_m(k,\tau)|^2$ differs dramatically from $P_{\epsilon_g}(k,\tau)$ for small wavenumbers (e.g., $k\le aH$), generally we can not replace $P_{\epsilon_g}(k,\tau)$ with $P_{\epsilon_m}(k,\tau)$ in Eq.(\[sigma\]) directly to obtain the rms fluctuation and correlation function of the density. The temperature fluctuation of the cosmic microwave background (CMB) can be characterized by the angular power spectrum $C_l$. Based on the line-of-sight-integral method developed by Seljak and Zaldarriaga[@seljak], we have: $$\begin{aligned} \label{cl} C_l=\frac{1}{2\pi}\int_0^{\infty}\frac{dk}{k} k^3\left[\int_0^{\tau_0}\Theta(k,\tau) j_l(k\tau_0-k\tau)d\tau\right]^2~,\end{aligned}$$ with $$\begin{aligned} \label{effects} \nonumber \Theta(k,\tau) &=& g\left(\frac{1}{4}\epsilon_{\gamma}+\phi+ \frac{\dot{v}_{b}}{k}+ \frac{\dot{g}}{g}\frac{v_{b}}{k}\right)+e^{-\kappa}(\dot{\phi}+\dot{\psi})\\ && +g\left(\frac{\prod}{4}+\frac{3\ddot{\prod}}{4k^2}\right) +\dot{g} \frac{3\dot{\prod}}{4k^2} +\ddot{g}\frac{3\prod}{4k^2}~,\end{aligned}$$ where $j_l$ is the spherical Bessel function; $\epsilon_{\gamma}$ is the photon density perturbation; $\prod$ is a gauge-invariant quantity denoting the effect of polarization and anisotropic stress of photons; $v_{b}$ denotes the velocity perturbation of baryons; $\tau_0$ denotes present time; $g$ is the visibility function and $\kappa$ is the total optical depth at time $\tau$. Now, let us take the standard cold dark matter (SCDM) model as an example to illustrate the prediction differences between the conventional treatment and my viewpoint. The SCDM model (the cold dark matter density parameter $\Omega_c=0.95$, the baryon density parameter $\Omega_b=0.05$, $H_0=100h ~km~s^{-1}~Mpc^{-1}$ with $h=0.5$) assumes that the primordial density spectrum in the $SG$ frame $P_{\epsilon(SG)}(k,\tau_i)$ is the Harrizon-Zel’dovich (HZ) spectrum: $$\label{hz} P_{\epsilon(SG)}(k,\tau_i)\equiv \mid \epsilon_{(SG)}(k,\tau_i)\mid^2=Ak~,$$ here $A$ is a constant and $\tau_i$ is the primordial time. In the conventional treatment, the density perturbation $\epsilon_{(SG)}$ in the $SG$ frame was directly regarded as the physical density perturbation, so the rms fluctuation of the density on the scale of $r_8=8h^{-1}~Mpc$ was directly corresponded to $\sigma_{\epsilon (SG)}(r_8,\tau_0)$, which is related to $P_{\epsilon(SG)}(k,\tau_0)\equiv \mid \epsilon_{(SG)}(k,\tau_0)\mid^2$ by $$\begin{aligned} \label{rms8sg} \sigma_{\epsilon (SG)}(r_8,\tau_0)=\left[\frac{1}{2\pi^2}\int_0^{\infty} P_{\epsilon(SG)}(k,\tau_0)W^2(kr_8) k^3\frac{dk}{k}\right]^{\frac{1}{2}}~.\end{aligned}$$ In the foregoing discussions we have shown that this picture is wrong because $\epsilon_{(SG)}(k,\tau)$ (or its spectrum $P_{\epsilon(SG)}(k,\tau)\equiv \mid \epsilon_{(SG)}(k,\tau) \mid^2$) can not characterize the physical density fluctuation [*directly*]{}. The physical density spectrum corresponding to the assumption of the SCDM model on the primordial density spectrum should be $$\label{scdmp} P_{\epsilon_g}(k,\tau)= P_{\epsilon_g}(k,\tau_i)T^2_{\epsilon_g}(k,\tau)\simeq \left(\frac{3a^2_iH^2_i}{k^2}\right)^2 Ak T^2_{\epsilon_g}(k,\tau)~,$$ where, the subscript “$_i$” denotes the primordial time; $P_{\epsilon_g}(k,\tau_i)\simeq (3a^2_iH^2_i/k^2)^2 Ak $ is the physical primordial spectrum corresponding to the HZ spectrum in the $SG$ frame (notice that all the interesting modes are far outside the Hubble radius at time $\tau_i$, i.e., $k<<a_iH_i$), and $T_{\epsilon_g}(k,\tau)\equiv \frac{\epsilon_g (k,\tau)}{\epsilon_g (k,\tau_i)}$ is the transfer function of physical density spectrum due to the evolution of density perturbation from the primordial time $\tau_i$ to the time $\tau$. Fig.1 shows the physical density spectrum $P_{\epsilon_g}(k,\tau)$ of the SCDM model at the redshift $z=0$ and $z=10$, with the corresponding $P_{\epsilon(SG)}(k,\tau)$. It can be seen that $P_{\epsilon(SG)}(k,\tau)$ can approximate the physical spectrum only for the large wavenumbers (small-scale modes), and differs significantly from the latter for the wavenumbers outside the Hubble radius. By the way, $P_{\epsilon(SG)}(k,\tau)$ differs hardly from the gauge-invariant power spectrum $P_{\epsilon_m}(k,\tau)$ for the SCDM model in which the universe was dominated by the pressureless matter at very high redshift. In the conventional treatment, only the power spectrum of the wavenumbers inside the Hubble radius is taken seriously, and the part outside the Hubble radius is thought to have no physical significance. This viewpoint is not correct: When a real-space fluctuation is decomposed into Fourier modes, all these modes will have physical significance whether or not they are inside the Hubble radius (see Eq.(\[delta\])). Only the [*whole*]{} power spectrum can determine a real-space Gaussian fluctuation completely. In fact, the power spectrum of the wavenumbers outside the Hubble radius has also been included in the current calculation of the density fluctuations (e.g., see Eq.(\[rms8sg\])) and the angular power spectrum $C_l$ for CMB (see Eq.(\[cl\])). So, the conventional treatment is inconsistent upon this point. More important, the temperature fluctuation detected by the COBE satellite (e.g., $C_2$) is dominated by the power spectrum of the wavenumbers outside the Hubble radius at the recombination era! On the other hand, inflationary scenario has provided a well-known mechanism of the non-vanishing power spectrum for the wavenumbers outside the Hubble radius. Fig.1 also reveals another important difference between the physical spectrum and $P_{\epsilon(SG)}(k,\tau)$: the shape of the physical spectrum changes even for the pressureless matter system, because the physical density perturbation modes far outside the Hubble radius can not grow due to the causality; while the shape of $P_{\epsilon(SG)}(k,\tau)$ for the SCDM model hardly changes, because the growth rate of $\epsilon_{(SG)}(k,\tau)$ is independent of wavenumber $k$ for the presureless system. Now we re-check some predictions of the SCDM model. From Eqs.(\[sigma\]) and (\[scdmp\]) we find that the rms fluctuation $\sigma_{\epsilon_g}(R,\tau)$ and the correlation $\xi_{\epsilon_g}(R,\tau)$ of density are both divergent on any scale $R$ at any time $\tau$, so do those of the potential $\phi$, [*if*]{} the physical density spectrum has the HZ shape exactly in the limit of $k\rightarrow 0$ in the $SG$ frame. In fact, these unreasonable predictions caused by the divergent spectrum have also been reflected in the calculations of the temperature fluctuation of CMB, though it is very implicit. From the combination of Eqs.(\[cl\]), (\[effects\]) and (\[scdmp\]) we can obtain that $C_0$ is divergent too. This means that the temperature fluctuation of CMB is infinite in any direction. Because people are only interested in the angular distribution of the temperature anisotropy of CMB (e.g., $C_l$ for $l\ge 1$), this unreasonable prediction of the SCDM model has been ignored so far. All these unreasonable predictions are due to the fact that the power spectrum was not cut off effectively. Any physical power spectrum should become negligible as wavenumber $k$ tends to zero due to the causality. So, the primordial spectrum should be cut off effectively at a minimum wavenumber $k_{min}$, e.g., $a/k_{min}$ is at most the same order as the horizon (notice that the horizon may be much larger than the Hubble radius). After doing so, we can see that all calculation results are convergent and dependent on $k_{min}$. The key point lies in how to determine the effective minimum wavenumber $k_{min}$ for the power spectrum. If $k_{min}$ is inside the Hubble radius at any time, there will not be so many problems to discuss here. However, $k_{min}$ can be far outside the Hubble radius (for example, the quantum fluctuation in the early universe can be stretched far outside the Hubble radius by inflation process), otherwise we can not explain the temperature fluctuation detected by the COBE satellite, which showed that these fluctuation modes are outside the Hubble radius at the recombination era. In conclusion, the power spectrums differ dramatically for the modes outside the Hubble radius in different gauges, however, the calculation of the rms density fluctuation needs to integrate the power spectrum for all perturbation modes (see Eq.(\[sigma\])), not merely the modes within the Hubble radius, so do the calculations of the two-point correlation function, the abundance of the galaxies or clusters by Press-Schechter formula, and the angular power spectrum $C_l$ for CMB, etc. In the above demonstrations we have shown that these quantities should be calculated directly by the gauge-invariant spectrum $P_{\epsilon_g}(k,\tau)$, or the spectrum $P_{\epsilon(NG)}(k,\tau) \equiv \mid \epsilon_{(NG)}(k,\tau)\mid^2$ in the Newtonian Gauge which coincides with $P_{\epsilon_g}(k,\tau)$, but not directly by the spectrum $P_{\epsilon(SG)}(k,\tau)$ in the Synchronous Gauge. On the other hand, the density fluctuations and the CMB anisotropies are dependent on the cut-off wavenumber $k_{min}$ of the physical power spectrum. $k_{min}$ is not set by the Hubble radius simply, but by the physical process in early universe, such as inflation. In the inflationary picture there generally exists a non-vanishing power spectrum for the modes outside the Hubble radius, i.e., $k_{min}$ can be far outside the Hubble radius! These conclusions will change dramatically the scenarios of the current cosmological models on the formation of the large-scale structure of universe.  \ I am thankful to R. Brandenberger for reading manuscript and giving valuable comments. I also thank E. Bertschinger for using the COSMICS, and U. Seljak and M. Zaldarriaga for using the CMBFAST. This work was supported in part by the National Natural Science Foundation of China under Grant No.10047004. Peebles P J E 1980 The Large Scale Structure of the Universe (Princeton Univ. Press, Princeton) Bardeen J M, Bond J R, Kaiser N and Szalay A S 1986 ApJ 304 15 White M, Scott D and Silk J 1994 Ann. Rev. A&A 32 319 Dodelson S, Gates E I and Turner M S 1996 Science 274 69 \[gaw\] Gawiser E and Silk J 1998 Science 280 1405 Bardeen J M 1980 PRD 22 1882 Kodama H and Sasaki M 1984 Prog. Theo. Phys. Suppl. 78 1 \[muk11\]Mukhanov V F, Feldman H A and Brandenberger R H 1992 Phys. Rep. 215 203 Ma C P and Bertschinger E 1995 ApJ 455 7 \[seljak\]Seljak U and Zaldarriaga M 1996 ApJ 469 437 [**Captions for Figure**]{} Fig.1 The physical density spectrum for the SCDM model (denoted as the solid lines), with the corresponding spectrum in the Synchronous Gauge (denoted as the dashed lines). Since the physical rms fluctuation $\sigma_{\epsilon_g}(r_8,\tau_0)$ (see Eq.(\[sigma\])) is divergent for the SCDM model, for simplicity the normalization for the density spectrum is chosen so that $\sigma_{\epsilon(SG)}(r_8,\tau_0)=1$ (see Eq.(\[rms8sg\])) instead; The upper and the lower lines correspond to the cases of $z=0$ and $z=10$, respectively. =6 in [^1]: In the standard big-bang model, the Hubble radius is the same order as the horizon, but in the inflationary scenario, the horizon can be much large than the Hubble radius. The largest scale in causality is set by the horizon, while the physical perturbation modes outside the Hubble radius will hardly grow as long as the equation of state for the background of universe is unchanged. Notice that the perturbation amplitude for all modes will change if the equation of state of the background changes, e.g., $\phi$ changes during the transition from radiation-domonated era to matter-dominated era[@mukhanov].
{ "pile_set_name": "ArXiv" }
--- abstract: | We show that the bilinear complexity of multiplication in a non-split quaternion algebra over a field of characteristic distinct from $2$ is $8$. This question is motivated by the problem of characterising algebras of almost minimal rank studied in \[1\] This paper is a translation of a report submitted by the author to the XI international seminar “Discrete mathematics and its applications”. author: - | Vladimir Lysikov[^1]\ [Moscow State University]{} title: On bilinear algorithms for multiplication in quaternion algebras --- A sequence $(f_1,g_1,z_1;\dots;f_r,g_r,z_r)$ with $f_k\in U^{*}, g_k\in V^{*}, z_k\in W$ is called a *bilinear algorithm* of length $r$ for a bilinear mapping $\varphi\colon U\times V\to W$ if $$\varphi(u, v) = \sum_{k = 1}^r f_k(u)g_k(v)z_k\quad \forall u\in U, v\in V.$$ The minimal length of a bilinear algorithm for $\varphi$ is called the *bilinear complexity*, or the *rank*, of $\varphi$ and is denoted by $R(\varphi)$. Algorithms of minimal length are called optimal. Rank of bilinear map has a useful algebraic interpretation. A bilinear map can be thought of as a tensor in $U^{*}\otimes V^{*}\otimes W$. Bilinear algorithms are decompositions of this tensor and the bilinear complexity coincides with the tensor rank as defined in multilinear algebra. A quaternion algebra over field $F$ of characteristic $\mathop{\mathrm{char}} F\neq 2$ is a $4$-dimensional algebra generated by two elements $i, j$ with identities $$i^2 = p,\ j^2 = q,\ ij = -ji$$ for some $p,q\in F^{\times}$. Any quaternion algebra over $F$ is either isomorphic to the matrix algebra $F^{2\times 2}$ (in this case it is called split) or a noncommutative division algebra. It is known that $R(F^{2\times 2}) = 7$ and $R(H)\geq 8$ for non-split quaternion algebra $H$ \[3\]. Let $m = \dim U$, $n = \dim V$. We will prove some general results on algorithms of rank $m + n$ for some class of bilinear maps. We say that an element $u_0\in U$ is (left) $\varphi$-*regular* if the linear map $\varphi(u_0,\cdot)$ is an injection, i. e., if $$\varphi(u_0, v) = 0 \Leftrightarrow v = 0.$$ A bilinear algorithm $\varphi = \sum_{k = 1}^r f_k\otimes g_k\otimes z_k$ is called *two-component* if the set $\{1,\dots,r\}$ of indices can be partitioned into two sets $I$ and $J$ such that $\{f_i | i\in I\}$ and $\{g_j | j\in J\}$ are bases of $U^{*}$ and $V^{*}$ respectively. If $R(\varphi) = m + n$, $\mathop{\mathrm{lker}} \varphi = \mathbf{0}$ and every basis of $U$ contains a $\varphi$-regular element then all optimal bilinear algorithms for $\varphi$ are two-component. Let $\sum\limits_{k = 1}^{m + n} f_k\otimes g_k\otimes z_k$ be an optimal bilinear algorithm for $\varphi$. Since $\mathop{\mathrm{lker}} \varphi = \mathbf{0}$, the functionals $f_1,\dots,f_r$ span $U^{*}$. W.l.o.g. let $f_1,\dots,f_m$ be a basis and $u_1$ be a $\varphi$-regular element of the dual basis $u_1,\dots,u_n$. Since $u_1$ is regular, functionals $g_1,g_{m + 1},\dots,g_{m + n}$ span $V^{*}$. Case 1. $g_{m + 1},\dots, g_{m + n}$ are linearly independent. Then $I = \{1,\dots, m\}$ and $J = \{m + 1,\dots,m+n\}$ form a partition of $\{1,\dots,m+n\}$ required by definition of two-component algorithm. Case 2. $g_{m + 1},\dots, g_{m + n}$ are linearly dependent. In this case $\dim \mathrm{lin}(\{g_{m + 1},\dots, g_{m + n}\}) = n - 1$ and there is a unique (up to a constant) linear dependence of $g_{m + 1},\dots, g_{m + n}$. W.l.o.g. assume that $\sum_{k = 1}^s c_k g_{n + k} = 0$ where all the coefficients $c_k$ are nonzero. If there is an index $p$ such that $m + 1\leq p\leq m + s$ and $f_p$ has a nonzero first coordinate in basis $f_1,\dots,f_m$ then $I = \{2,\dots,m,p\}$ and $J = \{1,m + 1,\dots, m + n\}\setminus \{p\}$ form a required partition. If all $f_{m + 1},\dots,f_{m + s}$ have zero first coordinates, then we have a contradiction with $\varphi$-regularity of $u_1$, since $\varphi(u_1,v) = 0$ for a nonzero $v\in \ker g_1\cap \bigcap_{k = n + s + 1}^{m + n} \ker g_k$. A two-component bilinear algorithm for $\varphi$ exists if and only if there are bases $(u_1,\dots,u_m)$ and $(v_1,\dots,v_m)$ of spaces $U$ and $V$ resp. and collections $(z'_1,\dots,z'_m)$, $(z''_1,\dots,z''_n)$ with $z'_i, z''_j\in W$ such that the following condition holds: $$\varphi(u_i, v_j)\in \mathrm{lin}(\{z'_i, z''_j\}).$$ Any two-component bilinear algorithm for $\varphi$ can be written as $$\label{eqn5} \varphi = \sum_{i = 1}^m f_i\otimes (\sum_{j = 1}^n \lambda_{ij} g_j)\otimes z'_i + \sum_{j = 1}^n (\sum_{i = 1}^m \mu_{ij} f_i)\otimes g_j\otimes z''_j,$$ where $(f_i)$ and $(g_j)$ are bases of $U^{*}$ and $V^{*}$ respectively. Let $(u_i)$ and $(v_j)$ be the bases dual to $(f_i)$ and $(g_j)$. Then it follows that $$\label{eqn6} \varphi(u_i, v_j) = \lambda_{ij} z'_i + \mu_{ij} z''_j.$$ Conversely, if holds then is a two-component algorithm for $\varphi$. Let $A$ be a local algebra, $\dim A = n$, and $R(A) \geq 2n$. Equality $R(A) = 2n$ holds iff there are bases $(u_1 = 1, u_2,\dots u_n)$, $(v_1 = 1, v_2,\dots v_n)$ and collections $(z'_i, \dots, z'_n)$, $(z''_1,\dots,z''_n)$ such that $$u_i v_j \in \mathrm{lin}(\{z'_i, z''_j\}).$$ Let $F$ be a field of characteristic $\mathop{\mathrm{char}} F\neq 2$ and $H$ be a non-split quaternion algebra over $F$. Then $R(H) = 8$. Apply the preceding corollary with the following bases and collections: $$(u_1, u_2, u_3, u_4) = (v_1,v_2,v_3,v_4) = (1, i, j, k)\tag{\theequation a}$$ $$(z'_1, z'_2, z'_3, z'_4){=}(1 + \alpha i + \beta j + \gamma k, 1 + \alpha i - \beta j - \gamma k, 1 - \alpha i + \beta j - \gamma k, 1 - \alpha i - \beta j + \gamma k)\tag{\theequation b}$$ $$(z''_1, z''_2, z''_3, z''_4){=}(1 - \alpha i - \beta j - \gamma k, 1 - \alpha i + \beta j + \gamma k, 1 + \alpha i - \beta j + \gamma k, 1 + \alpha i + \beta j - \gamma k)\tag{\theequation c}$$ where $\alpha,\beta,\gamma\in F$, $\alpha\neq 0$, $\beta \neq 0$, $\gamma \neq 0$. By considering bilinear algorithms arising from different constants $\alpha,\beta,\gamma$ we also managed to prove the following fact about equivalence of optimal algorithms in the sense of de Groote \[2\]. Let $F$ be a field of characteristic $\mathop{\mathrm{char}} F\neq 2$ and $H$ be a non-split quaternion algebra over $F$. There are infinitely many de Groote equvalence classes of optimal bilinear algorithms for the multiplication in $H$. The author is grateful to Prof. V. B. Alekseyev for his attention to this research. This research was supported by RFBR grant 12-01-91331-DFG-a. 1. M. Bläser, A. M. de Voltaire. Semisimple algebras of almost minimal rank over the reals // Theoretical Computer Science. — 2009. — vol. 410, no. 50. — pp. 5202–5214. 2. H. F. de Groote. On varieties of optimal algorithms for the computation of bilinear mappings I // Theoretical Computer Science. — vol. 7, no. 1. — pp. 1–24. 3. P. Bürgisser,M. Clausen,M. A. Shokrollahi. Algebraic complexity theory. —Springer Verlag, 1997. [^1]: E-mail: `[email protected]`
{ "pile_set_name": "ArXiv" }
--- abstract: | We study the evolution of atomic and molecular gas in galaxies in semi-analytic models of galaxy formation that include new modeling of the partitioning of cold gas in galactic discs into atomic, molecular, and ionised phases. We adopt two scenarios for the formation of molecules: one pressure-based and one metallicity-based. We find that both recipes successfully reproduce the gas fractions and gas-to-stellar mass ratios of and 2 in local galaxies, as well as the and 2 disc sizes up to $z \leq 2$. We reach good agreement with the locally observed and 2 mass function, although both recipes slightly overpredict the low-mass end of the mass function. Both of our models predict that the high-mass end of the mass function remains nearly constant at redshifts $z < 2.0$. The metallicity-based recipe yields a higher cosmic density of cold gas and much lower cosmic 2 fraction over the entire redshift range probed than the pressure based recipe. These strong differences in mass function and cosmic density between the two recipes are driven by low mass galaxies ($\log{(M_*/M_\odot)} \leq 7$) residing in low mass halos ($\log{(M_{\rm{vir}}/M_\odot)} \leq 10$). Both recipes predict that galaxy gas fractions remain high from $z \sim 6 - 3$ and drop rapidly at lower redshift. The galaxy 2 fractions show a similar trend, but drop even more rapidly. We provide predictions for the CO $J= 1-0$ luminosity of galaxies, which will be directly comparable with observations with sub-mm and radio instruments. author: - | Gergö Popping$^{1}$[^1], Rachel S. Somerville$^{2}$ and Scott C. Trager$^{1}$\ $^{1}$Kapteyn Astronomical Institute, University of Groningen, Postbus 800, NL-9700 AV Groningen, the Netherlands\ $^{2}$Department of Physics and Astronomy, Rutgers University, 136 Frelinghuysen Road, Piscataway, NJ 08854, USA bibliography: - 'references.bib' title: Evolution of the atomic and molecular gas content of galaxies --- galaxies: formation - galaxies: evolution - galaxies: ISM - ISM: atoms - ISM: molecules Introduction ============ Attaining an understanding of when, how, and at what rate stars form out of interstellar gas, and of the mechanisms that regulate this process, is of key importance in building up a complete picture of galaxy formation and evolution. Observations across a range of scales have shown that star-formation (SF) is tightly linked to galaxy gas content. Observations in our Milky Way have shown that star formation takes place in dense giant molecular clouds [GMC; e.g., @Solomon1987; @McKee2007; @Bolatto2008]. Early observational work found a correlation between the surface density of the star formation rate (SFR) and the surface density of the total cold gas in galaxies [e.g. @Schmidt1959; @Kennicutt1998KS], while more recent work has emphasized that there is a strong correlation between the SFR density and the density of *molecular* hydrogen (2), while the correlation with atomic hydrogen () is weak or absent [@wong2002; @Bigiel2008; @Bigiel2011; @Schruba2011]. This work has stimulated a desire to understand and track the  and 2 content of galaxies separately in theoretical models. Observational studies of the  and 2 content of nearby galaxies have made great advances in recent years. The local  mass function down to masses of $\log{(M_{\rm{HI}}/M_\odot)} = 7$, and global density $\Omega_{\rm HI}$, has been quantified by blind surveys such as ALFALFA [@Giovanelli2005; @Martin2010]. The content and its relationship with other galaxy properties (such as stellar mass, stellar surface density, color, and concentration) have been quantified for a fairly large, homogeneously selected sample of nearby galaxies by GASS [Galex Arecibo SDSS survey: @Catinella2010; @Catinella2012; @Catinella2013]. The THINGS [The nearby galaxy survey: @Walter2008] and LITTLE THINGS [@Hunter2012] surveys mapped the atomic hydrogen density distribution in small samples of nearby galaxies. The molecular hydrogen content of galaxies has most commonly been studied through emission from $^{12}$CO (from here on CO) as a tracer. The CO mass-function of nearby galaxies was presented by @Keres2003, along with an estimate of the 2 mass-function resulting from the application of a constant conversion factor between CO luminosity and 2 mass. An updated estimate of the 2 mass function from the @Keres2003 sample, based on an empirical, and variable, CO-2 conversion factor, was presented by @Obreschkow2009. The BIMA SONG [BIMA survey of nearby galaxies: @Helfer2003], HERACLES [HERA CO-Line Extragalactic Survey: @Leroy2009_HERACLES] and COLD GASS [CO legacy database for GASS: @Saintonge2011] surveys mapped the CO-emitting gas in galaxies of the THINGS and GASS surveys, constraining the surface densities and gas-to-star ratios of molecular gas. Observations of atomic hydrogen in emission have up until now been restricted to galaxies at redshifts of $z\lesssim 0.2$ [@Verheijen2007; @Catinella2008]. Damped Lyman-$\alpha$ absorbers (DLA) have provided estimates of the global atomic gas content of the Universe ($\Omega_{\rm gas}$) at much higher redshifts [$z < 4.5$; e.g., @Rao2006; @Prochaska2009; @Noterdaeme2012], but the exact nature of these systems, and their connection to galaxies detected in emission, is still unclear, making the interpretation of these observations somewhat complicated [@Berry2013]. Direct observations of the molecular content of distant galaxies through the CO line have recently become available for small samples of objects, although these samples are usually biased towards the most gas-rich, actively star-forming galaxies [e.g. @Genzel2010; @Tacconi2010; @Riechers2011; @Bothwell2013; @Tacconi2013]. Although results are still inconclusive because of the small and potentially biased nature of the samples, and uncertainties in the 2-CO conversion factor, these studies suggest that galaxies at high redshift may have been considerably more rich in 2 than nearby galaxies. Moreover, a tight relationship between 2 surface density and SFR density seems to persist out to at least $z\sim 2$ [@Genzel2010; @Daddi2010]. The gas content of galaxies at high redshift has also been estimated using more indirect methods, such as by using Far-Infrared (FIR) observations and an assumed relationship between dust and 2 mass [@Magdis2012], or by using an empirical relationship between SFR density and total gas or 2 content along with SF tracers such as H-$\alpha$ or UV [@Erb2006; @Mannucci2009; @Popping2012]. All of the above efforts have already led to extremely valuable insights and constraints on galaxy formation models. However, our ability to measure and CO in emission, in unbiased samples of galaxies out to high redshift, is expected to undergo a revolution in the next decade, with new and upcoming facilities such as the Atacama Large Millimeter Array [ALMA; @Wootten2009] and the Square Kilometer Array [SKA; @Carilli2004] and its pathfinders the Karoo Array Telescope [MeerKAT; @Booth2009] and the Australian SKA Pathfinder [ASKAP; @Johnston2008] coming online. The observations expected from these facilities present a new and stringent challenge to theoretical models of galaxy formation. Until recently, most cosmological models and simulations of galaxy formation did not attempt to ‘partition’ gas into different phases, and used a total-gas based (+2) Kennicutt-Schmidt (KS) law to model star formation. However, aided by the insights gained from studies of the relationship between star formation and gas properties on $\sim$ kpc scales [e.g. @Bigiel2008; @Leroy2008] in external galaxies, theorists have also made considerable progress on developing physical models linking the efficiency of star formation on GMC scales with that on galactic scales. Several groups have implemented explicit modeling of detailed chemistry and simplified radiative transfer into galaxy-scale and cosmological numerical hydrodynamic simulations, tracking the multi-phase gas content and implementing 2-based star formation prescriptions [e.g. @Pelupessy2006; @Robertson2008; @Gnedin2011; @Christensen2012; @Kuhlen2012]. @Gnedin2011 [hereafter GK] presented fitting functions for the SFR in their simulations as a function of total cold gas density ($\Sigma_{\rm HI}+\Sigma_{\rm H_2}$), gas phase metallicity, and the intensity of the UV ionizing background. @Krumholz2009 presented analytic models for the formation of 2 as a function of total gas density and metallicity, supported by numerical simulations with simplified geometries [@Krumholz2008; @Krumholz2009], emphasizing the importance of metallicity as a controlling parameter in 2 formation. A somewhat different view is presented by @Ostriker2010, who propose that heating of the Interstellar Medium (ISM) by the stellar UV background plays a key role in regulating star formation. In their model, the thermal pressure in the diffuse ISM, which is proportional to the UV heating rate, adjusts until it balances the midplane pressure set by the vertical gravitational potential. This could provide an explanation for the strong empirical correlation between 2 fraction and disc midplane pressure found by @Blitz2006. These analytic models and fitting formulae can be implemented within semi-analytic models of galaxy formation. The modern semi-analytic approach applies simple, physically motivated recipes for physical processes that drive the formation and evolution of galaxies within the framework of a $\Lambda$ cold dark matter ($\Lambda$CDM) cosmology. These models can provide predictions of global galaxy properties (such as SFR, size, stellar mass and luminosity, gas content, metal enrichment) for large numbers of galaxies. Furthermore, they can efficiently explore the parameter space associated with the large number of “sub-grid” recipes that are used to model processes such as star formation, stellar feedback, black hole accretion and feedback from Active Galactic Nuclei (AGN). Semi-analytic models have been successful in reproducing many observed galaxy properties [e.g., @Kauffmann1993; @Cole1994; @Kauffmann1999; @Somerville1999; @Cole2000; @Somerville2001], in particular recent models that include ‘radio mode’ AGN feedback [e.g., @Bower2006; @Croton2006; @Kang2006; @Menci2006; @Monaco2007; @Somerville2008], although some puzzles remain. For example, SAMs from several different groups do not correctly reproduce the observed properties of low-mass galaxies [$\log{(M_*/M_\odot)} \sim 9 - 10.5$ @Fontanot2009; @Guo2010; @Weinmann2012]. These low-mass galaxies form too early in the models, and are too passive at late times. On the other hand, SAMs have also had difficulty reproducing enough very rapidly star forming, extreme starbursts (Ultra-luminous Infrared Galaxies; ULIRGS) at high redshift [@Somerville2012; @Niemi2012 and references therein]. However, numerical hydrodynamic simulations suffer from the same problems [@Weinmann2012; @Dave2010], and in fact produce very similar predictions to the SAMs, leading most theorists to conclude that it is likely to be limitations in our understanding of the sub-grid processes of star formation and stellar feedback, rather than inaccuracies of the semi-analytic approach, that are the root cause of the problems. Several groups have now used semi-analytic models to make predictions for the multi-phase gas content of galaxies. @Obreschkow2009_sam applied an empirical pressure-based recipe based on the results of @Blitz2006 [hereafter BR] in post-processing to compute the and 2 content of galaxies in the Millennium simulations [@DeLucia2007]. @Power2010 carried out a similar project based on post-processing. @Fu2010 [@Fu2012] modeled the partitioning of gas into and 2 in radial bins in each galaxy, using both the metallicity-dependent recipes of @Krumholz2009 [hereafter KMT] and the pressure-based recipe of BR, and self-consistently implemented a 2-based star formation recipe, within the established semi-analytic modeling framework of @Guo2011. @Lagos2011cosmic_evol [@Lagos2011sflaw] also estimated gas partitioning into an atomic and molecular component, and implemented a 2-based star formation recipe, within the GALFORM semi-analytic model [@Baugh2005; @Bower2006]. Somewhat simpler models in a similar spirit have also been presented by @Dutton2010 and @Krumholz2011. In this paper we explore how different models for 2 formation affect the evolution of the atomic and molecular gas content of galaxies with time. We implement an empirical, pressure-based recipe (BR) and a recipe based on numerical hydrodynamic simulations, dependent on metallicity and the local UV radiation field (GK) into the @Somerville2012 model, thus allowing a link to be made between the stellar and dust emission and the multi-phase gas content of galaxies. We anticipate that these predictions will be useful for planning upcoming observations of cold gas in galaxies at high redshift, and as these observations become available, this will provide insights into the physics that drives the formation of molecules in galaxies. Furthermore, we aim to give insight to what improvements need to be incorporated in cosmological galaxy evolution models to correctly model the gas content of galaxies. In Somerville, Popping & Trager (in prep.; SPT14) we implement a wider suit of star-formation and 2 recipes including the KMT recipe. We will present predictions for the observable properties of the *stellar* (and dust) emission over a broad range of redshifts, and discuss the sensitivity of these properties to the adopted SF-recipes. The structure of the paper is as follows. In Section \[sec:model\] we briefly present the semi-analytic model and its ingredients, focussing on the new recipes for gas partitioning and star formation. In Section \[sec:results\] we present our predictions for the scaling relations between stellar mass or surface density and and 2 content, relationship between mass and radius, and and 2 mass functions at $z\sim 0$. We further present predictions for the evolution in the SFR half-light radius vs. stellar mass, and 2 mass functions, global mass density of and 2, and and 2 fractions vs. stellar mass. We compare our predictions of 2 fractions and mass functions with observational estimates of these quantities obtained by applying a CO-2 conversion factor to the observations; we also adopt an alternate approach in which we use our knowledge of the physical properties of our model galaxies to estimate the CO content, and compare directly with the CO observations. In Section \[sec:discussion\] we discuss our findings and we summarize those in Section \[sec:conclusion\]. Model {#sec:model} ===== This section describes the semi-analytic model used in this paper. The model is based on the models presented in @Somerville1999, @Somerville2008, and @Somerville2012 and we refer the reader to those papers for details. In this section we provide a brief summary of the model framework and the ingredients relevant to this paper. Throughout this paper we adopt a flat $\Lambda$CDM cosmology with $\Omega_0 = 0.28,\,\Omega_\Lambda = 0.72, h = H_0/(100\, \rm{km}\,\rm{s}^{-1})=0.70, \sigma_8 = 0.812$ and a cosmic baryon fraction of $f_b = 0.1658$ [@Komatsu2009]. Unless stated otherwise we leave the free parameters associated with the galaxy-formation model fixed to the values given in @Somerville2012. Semi-analytic model framework ----------------------------- The merging histories of dark matter halos (merger trees) are constructed based on the Extended Press-Schechter formalism following the method described in @Somerville99tree and @Somerville2008. Each branch in the tree represents a merger event and is followed back in time to a minimum progenitor mass of $M_{\rm{res}}$, which we refer to as the mass resolution of our simulations. Whenever dark matter halos merge, the central galaxy of the largest progenitor halo becomes the new central galaxy, whereas all the other galaxies become ‘satellites’. Satellite galaxies may eventually merge with the central galaxy due to dynamical friction. Merger timescale are estimated using a variant of the Chandrasekhar formula from @Boylan2008. Tidal stripping and destruction of the satellites is included as described in @Somerville2008. Before reionisation of the Universe, each halo contains a mass of hot gas equal to the universal baryon fraction times the virial mass of the halo. After reionisation, the collapse of gas into low-mass halos is suppressed by the photoionising background. We model the fraction of baryons that can collapse into halos of a given mass after reionisation using the fitting functions provided by @Gnedin2000 and @Kravtsov2004. When a dark matter halo collapses or experiences a merger with a larger halo, the hot gas shock-heats to the virial temperature of the new halo. The radiating gas then gradually cools and collapses. To calculate the cooling rate of the hot gas we use the metallicity-dependent radiative cooling curves of @sutherland1993. The rate at which gas can cool is given by $$\dot{m}_{\mathrm{cool}}=\frac{1}{2}m_{\mathrm{hot}}\frac{r_{\mathrm{cool}}}{r_{\mathrm{vir}}}\frac{1}{t_{\mathrm{cool}}},$$ where $m_{\mathrm{hot}}$ is the mass of the hot halo gas, $r_{\mathrm{vir}}$ is the virial radius of the dark matter halo, and $r_{\mathrm{cool}}$ is the radius within which all of the gas can cool in a time $t_{\mathrm{cool}}$, which itself depends on density, metallicity and temperature. This cooling radius limited regime is associated with “hot flows”. In some cases the cooling radius can be larger than the virial radius. In this case the cooling rate is limited by the infall rate $$\dot{m}_{\mathrm{cool}}=\frac{1}{2}m_{\mathrm{hot}}\frac{1}{t_{\mathrm{cool}}}.$$ This infall limited cooling regime is associated with “cold flows” [@Birnboim2003; @Keres2005; @Dekel2006]. Although in reality satellite galaxies should continue to accrete some cold gas, we assume that the cold gas is only accreted by the central galaxy of the halo. When the gas cools we assume it initially collapses to form a rotationally supported disc. The scale radius of the disc is computed based on the initial angular momentum of the gas and the halo profile, assuming that angular momentum is conserved and that the self-gravity of the collapsing baryons causes contraction of the matter in the inner part of the halo [@Blumenthal1986; @Flores1993; @Mo1998]. Assuming that the halo initially has a density profile described by the Navarro-Frank-White [NFW; @Navarro1996] form, the size of the gas disc of a galaxy is given by $$r_{\rm gas} = \frac{1}{\sqrt{2}}f_j\lambda R_{\rm vir}f_c^{-1/2}f_{\rm R}(\lambda,c,f_d),$$ where $f_j \equiv (J_d/m_d)/(J_h/M_{\rm{vir}})$ is the ratio of the specific angular momentum of the disc and the halo, $c$ is the NFW concentration of the halo, and $f_d$ is the disc mass to the halo mass ratio. The functions $f_c^{-1/2}$ correct for the difference in energy of the NFW profile relative to that of a singular isothermal profile, and $f_{\rm R}$ accounts for the adiabatic contraction [see @Mo1998 for expressions governing $f_{\rm R}$ and $f_c$]. @Somerville2008size showed that this approach produced good agreement with the evolution of the size-stellar mass relation for disc-dominated galaxies from $z\sim 2$ to the present. Stars are formed through two modes, a “normal” mode in isolated discs, and a merger-driven “starburst” mode. We discuss star formation in the “normal” mode in below Section \[sec:h2\_sf\]. The efficiency and timescale of the “starburst” mode is set by the merger mass ratio and the gas fractions of the merger progenitors, based on the results of hydrodynamical simulations of binary galaxies [@Robertson2006; @Hopkins2009a]. When supernovae occur, they deposit some of their energy into the ISM, driving a large-scale outflow of cold gas from the galaxy. The mass outflow rate is given by $$\dot{m}_{\rm out} = \epsilon_{\rm SN} \left(\frac{V_0}{V_c} \right)^{\alpha_{\rm rh}} \dot{m}_*$$ where $V_c$ is the maximum circular velocity of the galaxy (here approximated by $V_{\rm max}$ of the uncontracted dark matter halo), $\dot{m}_*$ is the star formation rate, and $\epsilon_{\rm SN}$ and $\alpha_{\rm SN}$ are free parameters ($V_0=200$ km/s is an arbitrary normalization constant). Some fraction of the ejected gas escapes from the potential of the dark matter halo, whereas some is deposited in the hot gas reservoir within the halo and can cool again. The fraction of gas ejected from the disc and halo versus ejected from the disc but retained in the halo is a function of the halo circular velocity, such that low-mass halos lose a larger fraction of gas (see @Somerville2008 for details). We choose $\epsilon_{\rm SN} = 1.5$ and $\alpha_{\rm SN} = 2.2$ (similar to previous works) to obtain a good match with the observed $z\sim 0.0$ stellar mass function. Each generation of stars produces heavy elements that can enhance the metal content of a galaxy. Here, chemical enrichment is modelled in a simple manner using the instantaneous recycling approximation. For each parcel of new stars ${\rm d}m_*$, we also create a mass of metals ${\rm d}M_Z = y \, {\rm d}m_*$, which we assume to be instantaneously mixed with the cold gas in the disc. We assume the yield $y$ to be constant, and treat it as a free parameter. When supernova driven winds eject gas from the disc, a corresponding proportion of metals is also removed and deposited either in the hot gas or outside the halo, following the same proportions as the ejected gas. Mergers can remove angular momentum from the disc stars and build up a spheroid. The efficiency of disc destruction and build up of spheroids is a function of progenitor merger mass ratio and gas fractions, parameterised based on the simulations of binary galaxy systems [@Hopkins2009a]. These simulations indicate that more “major” and more gas-poor mergers are more efficient in removing angular momentum, destroying discs, and building spheroids. When implemented within the SAM, these recipes correctly predict the relative fractions of early vs. late type galaxies as a function of stellar mass [@Hopkins2009b]. The model tracks the growth of supermassive black holes and the energy they release [@Croton2006; @Somerville2008]. Each top-level DM halo is seeded with a $\sim 100 M_\odot$ black hole, and these black holes are able to grow via two different accretion modes. The first accretion mode is fueled by cold gas that is driven into the nucleus of the galaxy by mergers. This mode is radiatively efficient, and the accretion rates are close to the Eddington limit. The accretion continues until the energy being deposited into the ISM in the central region of the galaxy is sufficient to significantly offset and halt accretion via a pressure-drive outflow. Because this accretion mode is associated with optically bright classical quasars and AGN, it is sometimes referred to as “bright mode” or “quasar mode” accretion. The second mode of black hole growth, the “radio mode”, is thought to be associated with powerful jets observed at radio frequencies. Hot halo gas is assumed to be accreted according to the Bondi-Hoyle approximation [@Bondi1952]. We adjust the efficiency of “radio mode” heating to fit the observed number density of massive galaxies, and obtain $\kappa_{\rm radio} = 3.8 \times 10^{-3}$). Accretion rates in this mode are significantly sub-Eddington ($10^{-4}$ to $10^{-3}$ times the Eddington rate), so that most of the BH’s mass is acquired during “bright mode” accretion. However, the radio jets are assumed to couple very efficiently with the hot gas, and provide a heating term that can partially or completely offset cooling during the “hot flow” mode. Multi-phase Gas Partitioning and Star Formation Recipes {#sec:h2_sf} ------------------------------------------------------- In this section we describe the new ingredients of our model that we use to calculate the fraction of ionised, atomic, and molecular gas in each galaxy, and how we compute the SFR based on the molecular gas content. At each time step we compute the scale radius of the cold gas disc using the angular momentum argument described in the previous subsection. We assume that the cold gas is distributed in an exponential disc with scale radius $r_{\mathrm{gas}}$ and a central gas surface density of $m_{\rm cold}/(2\pi\,r_{\mathrm{gas}}^2)$, where $m_{\rm cold}$ is the mass of all cold gas in the disc. We assume that the stellar scale length is defined as $r_{\mathrm{star}} = r_{\mathrm{gas}}/\chi_{\mathrm{gas}}$, with $\chi_{\mathrm{gas}}=1.7$ fixed to match stellar scale lengths at $z = 0$. We divide the gas disc into radial annuli and compute the fraction of molecular gas in each annulus as described below. The integrated mass of and 2 in the disc at each time step is calculated using a fifth order Runga-Kutta integration scheme. ### Ionised gas We assume that the cold gas consists of an ionised, atomic and molecular component. The ionised component may be due to either an external background or by the radiation field from stars within the galaxy. We assume that some fraction of the cold gas in the galaxy, $f_{\rm ion, int}$, is ionised by the stars in the galaxy. The external background field ionises of a slab of gas on each side of the disc. Following @Gnedin2012, and assuming that all the gas with a surface density below some critical value $\Sigma_{\rm HII}$ is ionised, we use $$f_{\rm ion} = \frac{\Sigma_{\rm HII}}{\Sigma_0} \left[1 + \ln \left(\frac{\Sigma_0}{\Sigma_{\rm HII}} \right) + 0.5 \left(\ln \left(\frac{\Sigma_0}{\Sigma_{\rm HII}}\right) \right)^2 \right].$$ Throughout this paper we assume $f_{\rm ion, int} = 0.2$ (as in the Milky Way) and $\Sigma_{\rm HII} = 0.4 \, M_\odot \rm{pc}^{-2}$, supported by the results of @Gnedin2012. Although observations do not support a sharp transition to ionized gas at this surface density, we found that our model reproduced the results of the hydro simulations well with this choice of parameters. ### Molecular gas: pressure based partitioning In this work we consider two approaches for calculating the molecular fraction of the cold neutral gas in a galaxy. The first is based on the empirical pressure-based recipe presented by @Blitz2006 [BR]. They found a power-law relation between the disc mid-plane pressure and the ratio between molecular and atomic hydrogen, i.e., $$R_{\mathrm{H}_2} = \bigl(\frac{\Sigma_{\mathrm{H}_2}}{\Sigma_{\mathrm{HI}}}\bigr) = \bigl(\frac{P_m}{P_0}\bigr)^\alpha \label{eq:blitz2006}$$ where $\Sigma_{\mathrm{H}_2}$ and $\Sigma_{\mathrm{HI}}$ are the 2 and surface density, $P_0$ and $\alpha_{\rm BR}$ are free parameters that are obtained from a fit to the observational data, and $P_m$ the mid-plane pressure acting on the galactic disc. We adopted $\log P_0/k_B = 4.23$ cm$^3$ K and $\alpha_{\rm BR}=0.8$ from @Leroy2008. The hydrostatic pressure acting on the disc at a radius $r$ is estimated as [@Elmegreen1989; @Elmegreen1993; @Fu2010] $$P_m(r) = \frac{\pi}{2}\,G\,\Sigma_{\mathrm{gas}}(r)\left[\Sigma_{\mathrm{gas}}(r) + f_{\sigma}(r)\Sigma_*(r)\right] \label{eq:pressure}$$ where G is the gravitational constant, $f_\sigma(r)$ is the ratio between $\sigma_{\mathrm{gas}}(r)$ and $\sigma_*(r)$, the gas and stellar vertical velocity dispersion, respectively. The stellar surface density profile $\Sigma_*(r)$ is modeled as an exponential with scale radius $r_{\mathrm{star}}$ and central density $\Sigma_{*, 0} \equiv m_*/(2 \pi r_*^2)$. Following @Fu2010, we adopt $f_{\sigma}(r) = 0.1 \sqrt{\Sigma_{*,0}/\Sigma_*}$, based on empirical scalings for nearby disc galaxies. The fraction of non-ionized gas in a molecular state at each radial annulus can be calculated as $f_{\rm{H}_2} = R_{\rm{H}_2}/(1 + R_{\rm{H}_2})$. ### Molecular gas: metallicity based partitioning The second approach for computing molecular gas fractions in galaxies is based on the simulation by @Gnedin2011 [GK], who performed high-resolution “zoom-in” cosmological simulations with the Adaptive Refinement Tree (ART) code of @Kravtsov99, including gravity, hydrodynamics, non-equilibrium chemistry, and 3D on the fly radiative transfer. Based on their simulations, the authors find a fitting function for the 2 fraction which effectively parameterizes $f_{\rm H2}$ as a function of dust-to-gas ratio relative to the Milky Way, $D_{\rm MW}$, the UV ionizing background relative to the Milky Way, $U_{\rm MW}$, and the neutral gas surface density $\Sigma_{HI+H_2}$. The fraction of molecular hydrogen at each radial annulus is given by $$f_{H_2}(r) = \left[1+\frac{\tilde{\Sigma}}{\Sigma_{HI+H_2}(r)}\right]^{-2}$$ where $$\begin{aligned} \tilde{\Sigma} & = & 20\, {\rm M_\odot pc^{-2}} \frac{\Lambda^{4/7}}{D_{\rm MW}} \frac{1}{\sqrt{1+U_{\rm MW} D_{\rm MW}^2}} \\ \Lambda & = & \ln(1+g D_{\rm MW}^{3/7}(U_{\rm MW}/15)^{4/7})\\ g & = & \frac{1+\alpha s + s^2}{1+s}\\ s & = & \frac{0.04}{D_*+D_{\rm MW}}\\ \alpha & = & 5 \frac{U_{\rm MW}/2}{1+(U_{\rm MW}/2)^2}\\ D_* & = & 1.5 \times 10^{-3} \, \ln(1+(3U_{\rm MW})^{1.7})\end{aligned}$$ We take the dust-to-gas ratio to be proportional to the metallicity in solar units $D_{\rm MW} = Z/Z_{\odot}$. The local UV background relative to the MW is set by relating the SFR of the galaxy in the previous time step to the MW SFR as $U_{\rm MW}= \frac{SFR}{SFR_{\rm MW}}$, where we choose $SFR_{\rm MW} = 1.0\,\rm{M}_\odot\,\rm{yr}^{-1}$ [@Murray2010; @Robitaille2010]. The GK fitting functions are intended to characterize the formation of molecular hydrogen on dust grains, the dominant mechanism for forming 2 once gas is enriched to more than a few hundredths of Solar metallicity. Other channels for the formation of 2 in primordial gas must be responsible for producing the molecular hydrogen out of which the first stars were formed. Hydrodynamic simulations containing detailed chemical networks and analytic calculations have shown that 2 can form through other channels in dark matter halos above a critical mass $M_{\rm crit} \sim 10^5 \rm{M}_\odot$ [e.g., @Nakamura2001; @Glover2013]. This gas can then form “Pop III” stars which can enrich the surrounding ISM to $\rm{Z}_{\rm III} \sim 10^{-3}\,\rm{Z}_\odot$ [@Schneider2002; @Greif2010; @Wise2012]. These processes take place in halos much smaller than our resolution limit. We represent them by setting a “floor” to the molecular hydrogen fraction in our halos, $f_{\rm H2,floor}$. In addition, we “pre-enrich” the initial hot gas in halos, and the gas accreted onto halos due to cosmological infall, to a metallicity of $\rm{Z}_{\rm pre-enrich}$. We adopt typical values of $f_{\rm H2,floor} = 10^{-4}$ and $\rm{Z}_{\rm pre-enrich}=10^{-3}\rm{Z}_\odot$ [@Haiman1996; @Bromm2004]. We find that our results are not sensitive to the adopted values of these parameters within reasonable limits. Note that observations of resolved stars in the MW halo and local dwarfs have revealed stars with metallicities below $\rm{Z}\sim10^{-3}\, \rm{Z}_\odot$ [@Tolstoy2009; @Starkenburg2010], precluding much higher values for $\rm{Z}_{\rm pre-enrich}$. ### Molecular based star formation Star formation is modeled following empirical relationships from recent observations. @Bigiel2008 suggest, based on observations of spiral galaxies from the THINGS survey, that the star-formation rate surface density can be directly related to the surface density of molecular gas, i.e. $$\label{eqn:bigiel1} \Sigma_{\rm SFR} = A_{\rm SF} \, {\Sigma_{\rm H_2}}^{N}$$ with $N\simeq 1$. Observations of higher density environments suggest that above some critical 2 surface density, the slope of the relation described in equation \[eqn:bigiel1\] steepens. We therefore adopt a two-part scaling law given by: $$\label{eqn:bigiel2} \Sigma_{\rm SFR} = A_{\rm SF} \, (\Sigma_{\rm H_2}/10 M_\odot {\rm pc}^{-2}) \left(1+ \frac{\Sigma_{H_2}}{\Sigma_{\rm H_2, crit}}\right)^{N_{\rm SF}}$$ We adopt the “two-slope” star formation recipe in all of the models presented in this work. In addition, we adopt $A_{\rm SF}=5.98 \times 10^{-3}\, M_\odot {\rm yr}^{-1} {\rm kpc}^{-2}$, $\Sigma_{\rm H_2, crit} = 70 M_\odot$ pc$^{-2}$, and $N_{\rm SF}=1.0$. The value of $A_{\rm SF}$ is taken from the observations of @Bigiel2008, corrected to our system in which Helium is not included in the gas masses and densities. The values for $\Sigma_{\rm H_2, crit}$ and $N_{\rm SF}=1.0$ are motivated by the observational compilation presented in @Narayanan2012. Results {#sec:results} ======= In this section we show our predictions for the evolution of the and 2 content of galaxies over a range of redshifts from $z=0.0$ to $z = 6.0$. The simulations were run on a grid of halos with virial masses ranging from $5\times 10^8 M_\odot$ to $5\times 10^{14} M_\odot$ with a mass resolution of $5\times 10^6 M_\odot$. We first perform a comparison of our model predictions with observations of local galaxy properties, in order to validate our models. All presented gas masses are pure hydrogen masses and do not include a correction for Helium. Local galaxy properties ----------------------- In Figure \[fig:gas\_scaling\] we present the ratios of and 2 relative to stellar mass, and the ratio of 2 to , as functions of stellar mass and stellar surface density in disc-dominated galaxies ($M_{*_{\rm{bulge}}}/M_{*_{\rm{total}}} \leq 0.4$). We compare our results to a compilation of observations presented in @Leroy2008, @Saintonge2011, and @Catinella2013 based on the THINGS+HERACLES and GASS+COLDGASS surveys and in @Boselli2014 based on the Herschel reference survey. Both the pressure-based and metallicity-based recipes show very good agreement with the observed trends between stellar mass or stellar surface density and and 2 fractions. The fraction of relative to stars decreases with increasing stellar mass and surface density, whereas the fraction of 2 relative to stars remains roughly constant. Consequently, the fraction of cold gas in the form of 2 increases with stellar mass and surface density. The 2-to-ratio as a function of stellar mass is on average slightly too high in our models, although still within the scatter of the observations (particularly at low stellar masses). Here we focus on the gas fractions of disc-dominated galaxies. A similar exercise for a “blind” survey of galaxies would yield lower -to-stellar mass, 2-to-stellar mass, and 2-to-mass ratios. Spheroidal objects have much lower relative gas content than disc-dominated galaxies and most of the cold gas is atomic. We present the ratio of ionised hydrogen to galaxy stellar mass as a function of stellar mass in Figure \[fig:HII\]. We find a monotonic decrease in the ratio between ionised hydrogen and stellar mass, without any significant difference between the two 2 formation recipes. The ionised hydrogen mass ranges from about a tenth of the stellar mass in large galaxies, to about equal to the stellar mass in low mass objects ($\log{(M_*/M_\odot)} < 8 \,-\,9$), and up to 10-100 times the stellar mass in very low mass galaxies ($\log{(M_*/M_\odot)} < 6 \,-\,7$). These ratios are comparable to the ratio between galaxy and stellar mass, indicating that a significant amount of the cold gas in galaxies may be in an ionised component, as suggested by recent observations [@Tumlinson2011], but not accounted for in previous semi-analytic models of galaxy formation. We will further explore the predictions for ionised hydrogen in galaxies in future work. Both 2 recipes presented in this work rely on the estimated size of the galaxy disc, as this sets the surface density of the gas, one of the key parameters in calculating the 2 content of the gas. It is therefore of great importance to correctly predict the sizes of the gas disc in galaxies. Figure \[fig:HI\_size\] shows the disc size of a galaxy as a function of its mass. Following @Verheijen2001, we define the size of the disc as the radius at which the surface density of the gas equals $\Sigma_{\rm{HI}} = 1 M_\odot\,\rm{pc}^{-2}$. We calculate the location of $\Sigma_{\rm{HI}} = 1 M_\odot\,\rm{pc}^{-2}$ in post-processing, assuming an exponential distribution of the cold gas in the disc. Besides the fit presented in @Verheijen2001, we also present, where possible, the size of the discs of the THINGS sample of galaxies, which we computed from the radial profiles presented in [@Leroy2008]. There is good agreement between the modeled and observed disc sizes, spanning over three orders of magnitude in mass and two orders of magnitude in disc size. The good agreement between model and data is independent of the 2 formation prescription. We have shown that our models match the observed 2 fractions for nearby galaxies (Fig. \[fig:gas\_scaling\]), and @Somerville2008size has shown previously that the models also reproduce the size-stellar mass relation for disc-dominated galaxies from $z\sim 2$ to the present. Although this does not necessarily guarantee a match between the disc size and gas content of a galaxy, the agreement is an encouraging sanity check. Figure \[fig:HI\_mass\_func\_z0.0\] shows our predictions for the mass-functions at $z = 0.0$. Both star formation recipes show decent agreement with the observed mass functions at masses of $\log{(\rm{M}_{\rm{HI}}/\rm{M}_\odot)} \sim 10$ and higher. The pressure-based recipe slightly underpredicts the observed mass function in the mass range $\log{(\rm{M}_{\rm{HI}}/\rm{M}_\odot)} \sim 9 -10$, and slightly overpredicts the observations at lower masses. The metallicity-based recipe overpredicts the observed number of galaxies below $\log{(\rm{M}_{\rm{HI}}/\rm{M}_\odot)} \sim 8.5$. Figure \[fig:HI\_mass\_func\_z0.0\] shows that the galaxies responsible for the excess of low-mass objects are low-mass galaxies ($\log{(M_*/M_\odot)} \leq 7$) residing in low mass halos ($\log{(M_{\rm{vir}}/M_\odot)} \leq 10$). This underlines the importance of sufficiently high mass resolution in simulations that attempt to predict the properties of galaxies observed in . The predicted 2 mass function at $z=0.0$ is presented in Figure \[fig:H2\_mass\_func\_z0.0\], and compared with two observational estimates. Both estimates are based on the CO survey of @Keres2003. The estimated 2 mass function given by @Keres2003 was obtained by applying a fixed conversion factor to convert between CO and 2. @Obreschkow2009 estimated a variable 2-CO conversion factor based on the galaxy properties. Based on recent observations and theoretical work, a variable conversion factor that depends on galaxy properties (such as metallicity) is probably more reasonable (we discuss this further below). The predictions of both recipes are very similar, and we obtain good agreement with the observational estimates of @Keres2003, but significantly overproduce galaxies with large 2 masses relative to the @Obreschkow2009 results. It is possible that a process not included in our model, such as AGN feedback, could destroy or expel 2 in massive galaxies [@Saintonge2012] and possibly lower the number of 2 massive galaxies. Evolution of gas in galaxies ---------------------------- In this section we present our predictions for the evolution of the gas content in galaxies and make predictions for upcoming surveys of gas at high redshifts. ### Galaxy sizes Figure \[fig:CO\_size\] shows the SFR half-light radius of our modelled galaxies as a function of their stellar mass (i.e., the radius that encompasses half of the total SFR of the galaxy). We compare these results with radii presented in the literature for high-redshift galaxies [@foerster2006; @Genzel2010; @Tacconi2013] and, where possible, the CO half-light radius of the discs of the THINGS galaxies, which we computed from the radial profiles presented in [@Leroy2008 assuming a fixed conversion between the 2 and CO radial profiles]. Our results are in excellent agreement with the observations at high-redshift and in the local Universe, indicating that in spite of the simplicity of our model for computing disc sizes and surface density profiles, we appear to be able to correctly model the sizes and the location of star formation and the evolution of these quantities since $z=2$. For a fixed stellar mass, the SFR half-light radius increases with decreasing redshift. Consequently, the molecular gas is more compact in high redshift galaxies. This behavior is driven by the overall growth of galaxy discs with time, as they accrete gas with higher angular momentum. ### mass function Figure \[fig:HI\_mass\_func\_evol\] shows the predicted mass function at redshifts between $z=0$ and $z=6$. We overplot observations from @Zwaan2005 and @Martin2010 at $z=0.0$. For masses ($\log{(M_{\rm{HI}}/M_\odot)} \gtrsim 8$), the figure shows a clear monotonic increase in the number of galaxies at a given mass from $z=6$ to $z=2.0$. There is very weak evolution at $z\lesssim 2$, and almost none at all from $z\sim 1$–0. The weak evolution in the number of low mass galaxies shows that in our current model framework, the excess of these objects is already present at redshifts $\sim$ 2. We find little difference in the predicted evolution of the mass function between the metallicity- and pressure-based recipes. Although little evolution is seen in the mass function since $z\sim 2$, this of course does not mean that galaxies are static, or that is not being created or destroyed. It rather means that there is a kind of self-regulated equilibrium that arises naturally in these models. ### 2 mass function In Figure \[fig:H2\_mass\_func\_evol\] we show the predicted 2 mass function at redshifts between $z=0$ and $z=6$. The left panel contains mass functions obtained using the metallicity-based recipe, whereas the right pannel shows results obtained using a pressure-based recipe. These predictions are compared with the observational estimates of the 2 mass function at $z=0$, as shown in Figure \[fig:H2\_mass\_func\_z0.0\]. Both 2 recipes predict a gentle evolution in the 2 mass function at all 2 masses. In both recipes, the number of galaxies with large 2 masses increases from $z\sim 6$–2, then declines slightly to $z=0$. At lower masses, $\log{(M_{\rm{H2}/M_\odot})} \lesssim 9$, both models predict a slight increase in the number of low-2 mass galaxies from $z\sim 6$–4, then a more or less monotonic decline from $z\sim 4$ to $z\sim 0$. In both recipes, it is more difficult to form 2 in low-surface density gas. In our models, low-mass halos host galaxies with a larger fraction of their gas at low surface density (this is in accord with observational size-mass scaling relations), and therefore low-mass galaxies are less efficient at forming 2, as we saw in Figure \[fig:gas\_scaling\]. In the BR model, we would say this is because their disc midplane pressure is lower due to their smaller gravitational potential wells. In the GK model, we would say it is due to the lower availability of dust grains on which 2 can form. Thus the build-up of large 2-mass galaxies from $z\sim 6$–2 reflects the growth of structure and the formation of massive dark matter halos, while the decrease in the number of low-2 mass galaxies from $z\sim 4$ to $z\sim 0$ reflects the growth of galaxy discs resulting in lower cold gas surface densities, combined with low potential wells and/or low availability of dust grains. ### Evolution in galaxy gas-fractions In the following figures we present the gas fraction and relative 2 content of galaxies as a function of their stellar mass for different redshifts ($0<z<6$). In each case, we plot the conditional probability $P(f_{\rm gas}|m_{\rm star})$, and the reader should keep in mind that the most massive galaxies will be extremely rare at high redshift, and probably would not be included in any observed samples. Figure \[fig:gas\_frac\_evol\] shows the cold gas fraction of the modeled galaxies as a function of stellar mass, divided into redshift bins. We also included the indirectly derived gas fraction from @Popping2012. They calculated cold gas and 2 masses in galaxies from the [*COSMOS*]{} survey by inverting the @Bigiel2008 star-formation law in combination with the @Blitz2006 method to calculate the 2 fraction of cold gas. Including a recipe to calculate the 2 fraction of cold gas allowed Popping et al. to indirectly estimate both the molecular and the atomic hydrogen masses of thse galaxies. Our models predict that gas fractions decrease only mildly from $z\sim 6$–3. At lower redshifts the gas fractions decrease rapidly, such that galaxies with large stellar masses run out of gas first. This evolution is similar for both 2 recipes. Only in low mass galaxies ($\log{(M_*/M_\odot)} \leq 9$) do the two applied recipes give different predictions, with the metallicity-based recipe predicting slightly larger gas fractions. We find that our model predictions are in good agreement with the indirect estimates of @Popping2012 for $z\leq1.0$ in the mass range $\log{(M_*/M_\odot)} > 10$. At higher redshifts we find good agreement for objects with $\log{(M_*/M_\odot)} \geq 10.5$. We overpredict the indirect estimates from the literature at lower stellar mass, however, we did not take the selection criteria applied by Popping et al. into account here. Figure \[fig:H2\_star\_frac\_evol\] shows $f_{H2} \equiv \frac{M_{\rm{H2}}}{M_{\rm{H2}} + M_*}$ as a function of stellar mass at different redshifts. We included a compilation of observations presented in @Narayanan2012_gasfrac [taken from @Genzel2010 [@Daddi2010; @Tacconi2010; @Casey2011; @Bothwell2013]] and in @Tacconi2013. Besides the 2 masses quoted in the original literature, @Narayanan2012_gasfrac uses a novel approach to calculate the conversion between CO observation and 2 masses and their resulting gas fractions (see section \[sec:CO\] for a detailed description). We included the original values for $f_{H2}$ as well as the recalibrated values. Similar to the previous figure, there is no significant difference between the two studied recipes. The evolution in $f_{H2}$, however, is much stronger. At $z =0.0$ we find $f_{H2} \sim0.1$ at all probed stellar masses, whereas at $z=6.0$ we find values of $f_{H2} \sim 0.8$ over a large range of stellar masses. There is large scatter in $f_{H2}$ at redshifts $z=3.0 - 0.5$ over all probed stellar masses. This scatter is indicative of a transitional phase during which the relative 2 content of galaxies rapidly drops, however, this does not necessarily take place at the same time/rate in galaxies with similar stellar mass. This strong evolutionary trend, compared to the trends seen for the total cold gas fraction, indicates that the amount of 2 decreases not only due to the availability of less cold gas, but that the 2 fraction itself also drops [@Popping2012]. The rate at which this happens is independent of adopted recipe in our models. We find good agreement with the observations and their re-analysis by @Narayanan2012_gasfrac. Our model does not strongly favor either choice for the CO-2 conversion factor. Similar to the total cold gas fractions, we find that our model predicts a lower relative 2 content of galaxies than the indirect estimates by @Popping2012 suggest (especially at stellar masses $\log{(M_*/M_\odot)} \leq 10.5$). We again emphasize that so far, we did not take the selection bias inherent to the observations that went into @Popping2012 analysis into account. We will now discuss how selection criteria affect our results. Current samples of high-redshift galaxies are highly sensitive to their selection criteria and direct observation of the molecular content of the galaxies are usually biased towards the most gas rich galaxies. To study how this bias might affect the comparison of our model predictions with observations in the literature, we apply the selection criteria from the relevant surveys to our model galaxies assuming a metallicity-based 2 recipe and show the results in Figure. \[fig:frac\_evol\_literature\_GK\]. We compare our results to the gas fraction estimates for galaxies taken from the [*COSMOS*]{} sample with $I_{\rm{AB}} < 24$ [@Popping2012]. We also show predictions for a sample with H$_{\rm{AB}}<$ 25 mag, representative of galaxies in the CANDELS survey [@Grogin2011; @Koekemoer2011] for which reliable measurements of galaxy size are expected to be able to be obtained. When we account for the selection effects, we find good agreement with the indirect 2 fraction estimates from the [*COSMOS*]{} sample. At $z>1.5$ our model predicts slightly lower gas fractions than those suggested by the indirect estimates. The rough agreement is a very encouraging result for our model, but also emphasizes how important it is to properly take selection criteria into account when comparing models to observed galaxy samples. Our results also suggest that repeating the analysis on the deeper, $H$-band selected CANDELS sample will greatly expand the range of stellar mass and gas fraction that can be probed by the indirect method at $z>1.5$. We intend to repeat the @Popping2012 analysis on the CANDELS sample in the near future. These results will provide an interesting complement to direct measures of high redshift gas fractions that will become available from ALMA. ### Gas density evolution of the Universe Figure \[fig:density\_evolution\] shows the predicted global , 2, and total cold gas density (including ionised hydrogen) of the Universe as a function of time (in units of the critical density). We compare our results to local and 2 densities [@Keres2003; @Zwaan2005; @Obreschkow2009; @Martin2010; @Braun2012] and high-z estimates of the HI density obtained from Damped Lyman-$\alpha$ (DLA) absorption systems [e.g., @Peroux2005; @Rao2006; @Guimaraes2009; @Prochaska2009; @Noterdaeme2012; @Zafar2013]. We see that the two 2 formation recipes differ significantly in terms of both the total cold gas content of the Universe and the ratio between  and 2. The metallicity-based recipe predicts more cold gas overall at all redshifts, and also more . The pressure-based recipe produces more 2 overall, in spite of the lower amount of total cold gas. Both models underpredict $\Omega_{\rm HI}$ inferred from DLAS at $z\geq3$, the pressure-based model more dramatically. On the other hand, predictions by the metallicity-based model is in decent agreement with DLA observations at $z<2.5$. Overall the metallicity-based recipe is better in reproducing the observed values for $\Omega_{\rm HI}$ and $\Omega_{\rm H2}$. The two 2 formation recipes show a very different evolution in the global ratio of  to 2 with redshift (see Figure \[fig:density\_ratio\_evolution\]). The pressure-based recipe predicts a monotonic increase in $\Omega_{H2}/\Omega_{\rm gas}$ and $\Omega_{H2}/\Omega_{HI}$ with increasing redshift. The metallicity-based recipe predicts a very mild increase with increasing redshift up to $z>3.0$, then a flattening at higher redshifts. Especially worthwhile to note is that $\Omega_{H2}$ never exceeds $\Omega_{HI}$ for the metallicity-based recipe, whereas it does by up to a factor of three for the pressure-based recipe. We will give a detailed discussion about the origin of these differences, and how they can help to constrain the physics driving the partitioning of hydrogen into atomic and molecular hydrogen, in Section \[sec:discussion\]. As a comparison we also show predictions from @Obreschkow2009_letter and @Lagos2011cosmic_evol in Figure \[fig:density\_ratio\_evolution\]. Both authors use a pressure-based recipe similar to ours. Although the predictions differ in detail — unsurprising as many other aspects of the models differ — we find that our predictions for the pressure-based recipe are in qualitative agreement with other predictions from the literature, indicating a strong decline in $\Omega_{H2}/\Omega_{HI}$ with time. The slope of the decline differs significantly between the compared models. Only at $z>4.0$ do the Lagos et al. models predict an increase in $\Omega_{H2}/\Omega_{HI}$ with time. The authors claim this is due to a Monte-Carlo extension of the merger trees to very low mass halos dominated by , although our halo mass resolution is actually higher than theirs, so this seems unlikely to account for the difference with our results. Our predicted evolution in $\Omega_{H2}/\Omega_{HI}$ for the metallicity-based 2 recipe is much flatter compared with the predictions from pressure-based recipes. Predictions in Observation Space {#sec:CO} -------------------------------- Our model gives predictions for the 2 mass and surface density of galaxies, but these are difficult to observe directly. Observations typically use the CO luminosity as a tracer for the 2 content of a galaxy, assuming a CO-to-2 conversion factor. A proper prediction of the CO luminosity of galaxies requires the inclusion of detailed chemistry and radiative transfer calculations [@Lagos2012; @Popping2013RT]. In the present work we use a CO-to-2 conversion relation to convert our predicted 2 masses to more directly observable CO luminosities. The advantage of working in “Observation Space” is that the CO-to-2 conversion factor is thought to depend on galaxy properties such as internal density and metallicity, which are predicted by our models. Thus instead of attempting to convert CO luminosities to 2 masses for the observations, we can instead make use of our knowledge of our model galaxy properties to make a more physically motivated galaxy-by-galaxy conversion from 2 to CO. Recently, @Narayanan2012 and @Feldmann2012 coupled sub-grid models of the ISM with cosmological simulations of galaxy formation to calculate the CO-2 conversion factor for galaxies with different properties. Using a coupling of an 2-formation model and radiative-transfer calculations to simulated isolated and starburst galaxies, @Narayanan2012 found that the average CO-2 conversion factor in galaxies can be represented by $$\rm{X}_{\rm{CO}} = \frac{1.3\times10^{21}}{\rm{Z'}\times\Sigma_{H2}}$$ with $X$ in units of $\rm{cm}^{-2}\,(\rm{K}\,\rm{km}\,\rm{s}^{-1})^{-1}$, $\Sigma_{\rm{H2}}$ is the 2 surface density in units of $M_\odot\rm{pc}^{-2}$ and $\rm{Z}'$ is the gas metallicity in solar units. @Feldmann2012 use a coupling of sub-grid ISM models by @Glover2011 with cosmological simulations by @Gnedin2011. They find that, when averaged on kiloparsec scales, the CO-2 conversion factor is weakly dependent on column density and radiation field and can be described as a function of metallicity: $$\log{(\rm{X}_{\rm{CO}})} = a_1\log{(Z') }+ a_2$$ with $a_1=-0.66$ and $a_2=20.5$ (see the Feldmann, Gnedin & Kravtsov 2010 results averaged to 4 kpc). We estimate the CO luminosities of our model galaxies by applying three different assumptions for the CO-2 conversion factor; a fixed conversion of $\rm{X}_{\rm{CO}} = 2 \times 10^{20}\rm{cm}^{-2}\,(\rm{K}\,\rm{km}\,\rm{s}^{-1})^{-1}$, the approach presented by @Narayanan2012 and that of @Feldmann2012. Note that all CO luminosities presented here correspond to the CO J$=$1-0 transition. Figure \[fig:scaling\_CO\] shows the CO luminosity of our model galaxies as a function of stellar mass at $z=0.0$ for the three CO-2 conversion methods. Overplotted are CO luminosities observed by @Leroy2008 and @Saintonge2011. These CO luminosities have been obtained by converting the published 2 masses back to CO luminosities, using the CO-2 conversion factor assumed in the respective papers. The @Narayanan2012 and @Feldmann2012 methods produce very similar results, and when applied to our models both provide very good agreement with the observations. Both clearly produce better agreement with the observations than the fixed CO-2 conversion factor. The slope of the relation between CO luminosity and stellar mass varies slightly between the applied CO-2 conversion method, however, this is not very well constrained by the data. Figure \[fig:cofunc\_0.0\] shows the CO luminosity function at $z=0.0$ obtained using the three different CO-2 conversion methods. The @Feldmann2012 method gives the best overall agreement with the observed CO luminosity function. The @Narayanan2012 approach produces similar predictions, but with a slightly shallower low-luminosity end slope and more high CO luminosity galaxies. A fixed conversion factor of $\rm{X}_{\rm{CO}} = 2 \times 10^{20}\,\rm{cm}^{-2}\,(\rm{K}\,\rm{km}\,\rm{s}^{-1})^{-1}$ overpredicts the observations at all luminosities. All three methods slightly overpredict the number of high-CO luminosity objects. Figure \[fig:cofunc\_evol\] shows the predicted CO luminosity-functions at redshifts $z=0.0$–6 for all three applied CO-2 conversion methods. We overplot the $z=0$ CO luminosity function obtained by @Keres2003 to guide the eye. All three CO-2 conversion methods yield qualitatively the same evolutionary trends, but differ more in the details of the predicted evolution. All models predict a relatively mild flattening of the low-luminosity end of the CO luminosity function from $z\sim 6$–0, with a more rapid evolution on the bright end. The Feldmann et al. and Narayanan et al. approaches give almost identical results for the low-luminosity end, and differ more at high luminosities. The Feldmann et al. approach predicts fewer high-CO luminosity galaxies at high redshift. We show the evolution with redshift of the relation between stellar mass and CO luminosity in Figure \[fig:mstarCO\] (assuming the Feldmann et al. and Narayanan et al. approach for the CO-2 conversion factor). This diagram will, in the near future, be easily filled with observations of the CO luminosity of galaxies from surveys like [*GOODS*]{}, [*COSMOS*]{}, [*CANDELS*]{} using ALMA. As a comparison, we plot observational results presented in @Genzel2010 and @Tacconi2013, color-coded by redshift. We find that our model reproduces the observations very well. There is a clear linear relation between stellar mass and CO luminosity. The slope of this relation does not change with time and only slightly with CO-2 conversion method. The normalization of the relation does change with time, indicating that the relative amount of CO decreases at the same rate in galaxies spanning a wide range in stellar mass (and type). The @Narayanan2012 CO-2 conversion method predicts a stronger evolution in the CO luminosities with time, driven by the dependency of the CO-2 conversion on the 2 surface density. The high surface densities in high-redshifts galaxies decrease the CO-2 conversion factor, increasing the CO luminosity at a given stellar mass. We find only minor differences between the results obtained assuming a pressure- and metallicity-based recipe. Discussion {#sec:discussion} ========== In this paper we have presented new predictions for the evolution of the multiphase gas content and CO luminosity of galaxies from $z\sim 6$–0. We apply pressure- [@Blitz2006] and metallicity-based [@Gnedin2010] 2 formation recipes as two different approaches to calculating the molecular fraction of cold gas. Stars are formed following a power-law relation between the surface density of molecular gas and the SFR surface-density [@Bigiel2008]. Our goal is to assess the degree to which observations of the gas content of galaxies at high redshift can constrain the physics of the transformation of gas from one phase to another, and the conversion of cold dense gas into stars. In this section we discuss the results of this modeling effort and discuss our findings in comparison with previous studies using similar techniques. We will discuss the agreement and differences between the GK and BR model, draw general conclusions about the evolution of the gas content in galaxies, provide predictions that can help to guide future observational efforts, and discuss our results in the context of the physics driving galaxy formation in general. We find that both the pressure-based and metallicity-based 2 formation recipes do well at reproducing the gas fractions and gas-to-stellar-mass ratios of local galaxies and the trends with stellar mass and internal galaxy density. There are only very small differences in the scaling relations predicted by the pressure- and metallicity-based recipe over the entire stellar mass range probed. The predicted sizes of atomic hydrogen discs are in good agreement with observations at $z=0$, and the sizes of the modeled 2 discs are in good agreement with observations in the redshift range $z=0-2$. We note that these recipes were taken from empirical results calibrated to observations, or from numerical simulations, and were not tuned to cause our semi-analytic model to match these observations. This is an indication that, despite the simplicity of our model for gas partitioning, SF, and disc internal structure, we reproduce the distribution of gas in galaxies with reasonable accuracy. Both the pressure- and metallicity-based recipe do a fairly good job of reproducing the mass function over the whole range probed by observations, with a small excess of high-mass galaxies. Both models predict an excess of low--mass galaxies at $\log{(M_{\rm HI}/M_\odot)} < 8$ compared to observations. The galaxies responsible for the excess at low-masses in this model have low stellar masses ($\log{(M_*/M_\odot)} \leq 7$) and reside in low-mass-halos ($\log{(M_{\rm{vir}}/M_\odot)} < 9-10$). This shows that to properly model the smallest galaxies observed in , it is of key importance to resolve halos down to masses of $\log{(M_{\rm{vir}}/M_\odot)} \sim 8$, which frequently has not been possible in previous studies. For example, @Somerville2008 presented a predicted mass function that was apparently in much better agreement with the observed one, but this was merely an artifact of the relatively coarse halo mass resolution ($10^{10} M_\odot$) adopted in their simulations. Both recipes successfully predict the 2 mass function over the entire mass range probed. In both models, the number density of -massive galaxies shows an increase of about an order of magnitude from $z\sim 6$ to 4, then remains nearly constant to $z\sim 0$. This result indicates that there is a kind of self-regulated equilibrium that arises naturally in these models. To first order, the constant high-mass end of the mass function in our models is a consequence of the balance between accretion and the transformation of into 2. Observations have shown that saturates at surface densities of $\Sigma_{\rm{HI}} = 6-10 \, M_\odot\,\rm{pc}^{-2}$ and that higher cold gas densities are dominated by 2 [@Blitz2006; @Leroy2008]. In our models, as new gas is accreted, the amount of gas that is above surface densities where 2 formation is efficient increases, leading to conversion of into 2. The constant high-mass end of the mass function is a strong prediction that can be tested by the VLA up to $z \leq 0.4$ [@Fernandez2013], and SKA and its pathfinders ASKAP and MeerKat in the near future. It will not only probe the 2 formation recipes, but also the physics that drives the accretion, consumption, and heating and/or ejection of cold gas from galaxies. The number density of low 2-mass galaxies shows a strikingly different evolution, decreasing almost monotonically from $z\sim 4$ to $z\sim 0$. This behavior is qualitatively very similar in the two 2 formation models. It is intriguing that this behavior — weak evolution of massive objects, with a decrease in the number of low-mass objects — is qualitatively similar to the evolution of the observed stellar mass function [@Cimatti2006; @Marchesini2009], sometimes referred to as “mass assembly downsizing”. This suggests that “mass assembly downsizing” may be linked to the evolution of the molecular gas content of galaxies and the ability to form stars out of this molecular gas. We also find only minor differences in the evolution of galaxy gas fractions between the pressure- and metallicity-based recipes. Gas fractions are quite high ($\gtrsim 0.7$) over a broad range of stellar masses ($10^7 \lesssim M_* \lesssim 10^{12} M_\odot$) from $z\sim 6$–3, then drop fairly rapidly at lower redshifts. This drop in gas fraction occurs at higher redshift for galaxies with higher stellar mass — massive galaxies appear to consume or expel their gas earlier than less massive galaxies. A similar trend holds for the 2 fraction of galaxies, but the rate at which the 2 fraction drops is even faster than the rate of decline of the overall cold gas fractions. These trends are a different manifestation of mass assembly downsizing, and are in qualitative agreement with the observed evolution in galaxy gas fractions [@Tacconi2010; @Magdis2012; @Narayanan2012; @Popping2012; @Tacconi2013; @Sargent2013]. Future surveys of the molecular gas content of galaxies, as well as future efforts to indirectly estimate the gas content of galaxies, will be able to probe the gas content in galaxies over a much wider range in galaxy properties and environment, improving the constraints that can be obtained on models of galaxy formation. In a picture where galaxy gas fractions represent the competition between gas inflow, outflow and consumption through star formation [@dave2011], the decreasing gas fractions below redshifts of $z=3$ indicate that outflows and gas consumption largely dominate this competition. Galaxies run out of cold gas and of molecular gas, but not necessarily at the same rate [@Popping2012]. Taking into account that galaxies form their stars out of molecular gas, this means that declining SFRs are not only due to a decline in the cold gas available, but also due to an even more rapid decline of the 2 fraction of gas. The relative 2 content of galaxies with stellar masses below $10^{10} M_\odot$ predicted by our models appears to be slightly too low in the redshift regime $1.0 < z < 2.0$ compared to the predictions by @Popping2012. This effect is still present after taking selection criteria into account. It is probably related to the low-mass galaxy problem in models of galaxy formation, where galaxies in this mass regime are too passive at these redshifts with respect to the observations. We find that the low 2 content in these galaxies might be driving this problem, leading to inefficient star formation. A successful solution to the low-mass galaxy problem must also produce higher gas fractions in low-mass galaxies at intermediate redshift. The predictions of the cosmic-density evolution of , 2 and the total cold gas budget show the largest differences between pressure- and metallicity-based 2 recipes. The metallicity-based recipe yields a much higher cosmic density of cold gas and the density peaks at lower redshift. More striking is the difference in the evolution of the global 2 fraction, $\Omega_{H2}/\Omega_{\rm gas}$. The global 2 fraction assuming a metallicity-based 2 formation recipe shows only a mild decrease of a factor of $\sim 2$ from $z\sim 6$–0, whereas a pressure-based recipe predicts a strong decrease of a factor $\sim 6$ over this redshift range. Our predicted cosmic density of $\Omega_{\rm HI}$ is, at face value, in poor agreement with estimates from observations of DLAs for both 2 prescriptions. However, we have presented $\Omega_{\rm HI}$ for all galaxies without taking into account the selection criteria for DLAs. It is important to note that here we have computed the global density of gas by adding up all the gas in galaxies. However, DLAs may not provide an unbiased estimate of the total  content of the Universe. @Berry2013 present a detailed analysis of predicted DLA properties using the same semi-analytic models presented here, and show that $\Omega_{\rm HI}$ derived from DLAs as in the observational estimates shown here can differ substantially from the “true” underlying $\Omega_{\rm HI}$. They argue that a greater fraction of DLAs may arise from intergalactic or circumgalactic gas at $z\gtrsim 3$, while at lower redshifts, a large amount of may be in galaxies that have column densities too low for them to be selected as DLAs, leading to very weak evolution in $\Omega_{\rm DLA}$, as observed. The significant differences between the metallicity- and pressure-based recipes for 2 formation all find their origin in low mass galaxies ($\log{(M_*/M_\odot)} < 9$) within low mass halos ($\log{(M_{\rm{halo}}/M_\odot)} < 10$). A significant fraction of the cold gas and that leads to the higher cosmic densities of these quantities in the model with the metallicity-based recipe is within virtually “pristine” halos that contain less than $10^{6} M_\odot$ of stars [see also the discussion in @Berry2013]. These differences are driven by a lack of metals at high redshift, necessary for the metallicity-based recipe to form molecular gas. As a result fewer stars form, less gas is consumed and the cold gas content of galaxies piles up. Furthermore, the lack of formed stars slows down the production of metals necessary to form 2. Meanwhile, the high internal densities of high-redshift galaxies are highly conducive to the formation of molecules through a pressure-based recipe. It is important to note that both the pressure- and metallicity-based recipes predict a small excess of low--mass galaxies. None of the various 2 formation recipes that we have explored are able to remove this excess, suggesting that it may arise from other physical processes. We present predictions for the CO luminosities of our modelled galaxies using different methods to estimate the conversion between CO and 2. Although the general trends in CO are similar, different approaches to estimating the CO-2 conversion factor yield different predictions in detail, especially for lower-2-mass galaxies. The use of a fixed conversion factor between CO and 2 in our models overpredicts the observed luminosity function over a wide range of CO luminosities, although a different value for $X_{\rm CO}$ can change the normalization of the luminosity function. Using either the @Narayanan2012 or @Feldmann2012 CO-2 conversion prescriptions, which depend on galaxy properties, we obtain good agreement with the observed 2 luminosity function below the knee, but overpredict the number of high-2 mass galaxies by a significant amount, more so with the @Narayanan2012 prescription. The predicted evolution of the CO luminosity function is qualitatively similar to that of the 2 luminosity function, described above, although the detailed predictions depend somewhat on the adopted conversion prescription. Future surveys with sub-mm and radio telescopes such as the ALMA, PdBI, LMT, VLA, ATCA, and SKA, will be able to probe the CO $J=1-0$ luminosity function at $z \geq 2.0$ and provide valuable constraints for our models. Comparison with previous work ----------------------------- We now discuss our results with respect to other recent theoretical predictions of the evolution of atomic and molecular gas in semi-analytic galaxy formation models. We will attempt to not only point out the differences between the various modeling efforts, but also the common results that can shed more light on the physics at play in galaxy formation. We used the fitting functions provided by @Gnedin2011 [GK] in our metallicity-based recipe for the formation of 2, whereas most previous modeling efforts have used the analytic model of @Krumholz2009. These two approaches have been compared and were found to be very similar except at the lowest metallicities [@KrumholzGnedin]. We will show an explicit comparison of the two approaches in SPT14, and also find that they produce similar results. The GK fitting functions appear to be somewhat more robust and produce better agreement with observations, which is why we adopt them. Another difference in our approach is that we have separated the recipes for partitioning gas into an atomic and molecular component, and those for converting molecular gas into stars, while in some previous works both recipes were varied, making it more difficult to identify which aspects of the recipes may be causing differences in the results. In SPT14, we will present a systematic study of the effects of varying both the gas partitioning and star formation recipes separately. Here, we leave the star formation recipe fixed and vary only the gas partitioning recipes. A first attempt to study the atomic and molecular hydrogen content of galaxies was presented in @Obreschkow2009_sam and @Obreschkow2009_letter. The authors use the semi-analytic predictions from @DeLucia2007 and calculate the 2 and content of galaxies in post-processing using the @Blitz2006 pressure-based formalism. This model does not include an 2 based star-formation recipe, but rather assumes a traditional “total gas” based Kennicutt star formation relation, where stars form above some critical cold gas surface density. @Obreschkow2009_sam and @Obreschkow2009_letter find and 2 mass functions, 2 disc sizes and an evolution in universal density of 2 (Fig. \[fig:density\_ratio\_evolution\]) very similar to our findings when we assume a pressure based 2 formation recipe, and an 2-based star-formation recipe. @Obreschkow2009CO estimate the CO luminosity (ranging from CO J$=$1-0 to CO J$=$10-9) of a galaxy from its gas temperature based on the SFR surface density or AGN bolometric luminosity under local thermodynamic equilibrium (i.e. a single gas phase). The authors find that the low-luminosity end of the CO J$=$1-0 luminosity function is already in place at $z=2$, contrary to our predictions. The evolution of the bright end of the CO J$=$1-0 luminosity function is in much better agreement with our results. @Obreschkow2009CO point out that above $z>1$ the CMB starts to act as a bright background reducing the observed CO J$=$1-0 luminosity. At the same time, the higher excitation temperatures of the warm CMB in the early universe will ease the observability of CO emission [@Combes1999; @Gnedin2001], although the negative effect of the CMB dominates. These are effects we did not include in our model but which can play a significant role when observing young galaxies in the early universe. In particular, sources at $z>5$ without a strong heating by a starburst or AGN will not be detectable through low CO transitions. @Lagos2011cosmic_evol [@Lagos2011sflaw] study the evolution of the atomic and molecular gas content of galaxies using a pressure- and metallicity based 2 recipe in a semi-analytic model of galaxy formation. Their pressure-based model uses an 2-prescription from @Blitz2006 and a star-formation model from @Leroy2008, very similar to our pressure-based model. Their metallicity-based model follows the 2 prescription and star-formation model presented in @Krumholz2009. Although the authors vary the star-formation relation in their models, the models are not calibrated to necessarily fit the $z = 0$ luminosity functions, stellar and gas mass fractions and mass functions. The authors find that the metallicity-based recipes fail to reproduce the observed -mass functions and select the pressure-based recipes in combination with the @Bower2006 semi-analytic model as their preferred model. Taking into account that the @Lagos2011sflaw models are not calibrated to match local observations, we argue that a metallicity-based 2 and star-formation recipe should not be considered ruled out, although we also find (to a much lesser extent) that the metallicity-based 2 recipe tends to produce too many low-mass galaxies at $z=0$. The KMT model is known to break down at the lowest metallicities, due to a failure of the assumption of chemical equilibrium in the analytic model [@KrumholzGnedin]. This problem yields a rapid accumulation of large reservoirs in poor agreement with observations. We point out that the metallicity-based recipes require the cold gas in the initial time steps to be assigned a non-zero metallicity, otherwise no star formation will ever take place. The results can also be somewhat sensitive to the treatment of this “seed” metallicity, which may be provided by Pop III stars. Using their preferred model, @Lagos2011cosmic_evol find an evolution in the and 2 mass functions, gas fractions and 2 density of the universe very similar to our results. The authors find a bump in the mass function at $\log{(M_{HI}/\rm{h}^{-2} M_\odot)} \sim 7.5 - 8.0$, similar to (although much larger than) the excess number of low--mass galaxies we find. They ascribe this excess to a mismatch between the observed and modeled radii of the galaxy discs. We, however, have shown the sizes of the gas discs in our models (including the sizes of the and 2 components separately) are in good agreement with observations, so we do not think this is the main cause of the remaining excess of intermediate -mass galaxies in our models, though it may partially explain the better agreement of our metallicity-based model with observations. @Lagos2011cosmic_evol finds a good match between their preferred model and the observed CO luminosity function by @Keres2003. To obtain this match the authors need to assume a fixed CO-to-2 conversion factor of $X = 3.5\times10^{-20} \rm{cm}^{-2}/\rm{K}\,\rm{km}\,\rm{s}^{-1}$. @Fu2012 also studied the redshift evolution of atomic and molecular gas in galaxies, although the emphasis of their work lies more on the evolution of the mass-metallicity relation. In their work the authors use a variety of star-formation models (including the Bigiel et al. 2008 recipe) and apply both a metallicity- [based on @Krumholz2009] and a pressure-based 2 recipe [based on @Blitz2006]. Their model is calibrated to the local , 2 and stellar mass functions. The authors find that the evolution of the atomic and molecular gas fraction of galaxies is very similar for both applied 2 prescriptions, and is more dependent on the star-formation model. Similar to our findings, their results suggest that $\Omega_{\rm{H2}}/\Omega_{\rm{HI}}$ increases monotonically with increasing redshift for the pressure-based 2 recipe, whereas it decreases at redshifts $z>3$ for the metallicity-based recipe. The resolution of the models in @Fu2012 is not sufficient to study the differences in behavior of the low mass end of the 2 and mass function for the pressure- and metallicity based 2 recipes. This makes it difficult to compare our excess in low--mass galaxies and our constraints on the different 2 formation recipes with the results of @Fu2012. Despite the different implementations of the physical recipes, the three discussed models and ours all agree that a pressure-based recipe for 2 formation predicts a monotonic increase for $\Omega_{\rm{H2}}/\Omega_{\rm{HI}}$ with redshift [Fig. \[fig:density\_ratio\_evolution\], although note the decline in $\Omega_{\rm{H2}}/\Omega_{\rm{HI}}$ for the @Lagos2011cosmic_evol model], whereas it flattens out for a metallicity based recipe. We therefore conclude that the applied 2 recipe is likely to be responsible for these trends. As discussed extensively in the previous subsection, in the metallicity-based models, the low metallicities at early times make 2 formation and hence star formation very inefficient, in spite of the higher gas densities. Thus star formation, 2 formation, and enrichment are delayed in these models relative to the pressure-based models. Conclusions {#sec:conclusion} =========== We have presented predictions for the evolution of the atomic and molecular hydrogen content of galaxies from $z\sim 6 - 0$, based on a semi-analytic model of galaxy formation, including new modeling of the partitioning of cold gas in galactic discs into atomic, molecular, and ionised phases. We present results for two different 2 formation recipes: one a pressure-based recipe motivated by the empirical relation between molecular fraction and gas midplane-pressure from @Blitz2006, and one based on numerical hydrodynamic simulations in which the molecular fraction is highly dependent on the cold gas metallicity as well as the local UV background [@Gnedin2011]. We compared our predictions to local and high-redshift observations and adopted an alternate approach in which we estimate the CO content of galaxies and compare directly with CO observations. We summarize our main findings below. - Without any tuning, our models correctly predict the trends between gas fractions and gas-to-stellar-mass ratios of and 2 in local galaxies with mass and internal density. We furthermore reproduce the and 2 disc sizes of local and high redshift galaxies. - Both 2 formation recipes reproduce the observed $z=0$ mass function fairly well over the whole range probed by observations. Both models predict a small excess of low--mass galaxies. The high-mass end of the mass function remains remarkably constant at redshifts of $z \lesssim 2.0$ for both 2 formation recipes. - Both recipes correctly predict the 2 mass function over the entire mass range probed. The number density of 2-massive galaxies increases from $z\sim6$ to $z\sim4.0$ after which it remains fairly constant, whereas the number density of low-2 mass galaxies decreases almost monotonically from $z\sim4$ to $z\sim0$. - Galaxy gas fractions remain relatively high ($\gtrsim 0.7$) from $z\sim 6-3$, then drop fairly rapidly. A similar trend holds for the 2 fraction of galaxies, but the drop occurs at an even higher rate. - The metallicity-based recipe yields a much higher cosmic density of cold gas over the entire redshift range probed. The cosmic 2 fraction as predicted by the metallicity-based recipe is much lower than the 2 fraction predicted by the pressure-based recipe. - The galaxies responsible for the high cosmic gas density and low cosmic 2 fraction all reside in low-mass halos ($\log{(M_{\rm{halo}}/M_\odot)} < 10$), and contain negligible amounts of stellar material. The build-up of atomic gas in these low-mass halos is driven by a lack of metals at high redshift, necessary to form molecular gas, stars, and produce more metals. - The conversion of 2 masses to CO luminosities provides valuable direct predictions for future surveys with ALMA at low redshifts or radio interferometers such as the VLA at higher redshifts. None of the presented methods for the CO-to-2 conversion predicts perfect agreement with observations from the literature, although the physically motivated nature of the @Narayanan2012 and @Feldmann2012 approaches are favoured over a constant CO-to-2 conversion factor. The results presented in this paper can serve as predictions for future surveys of the atomic and molecular content of galaxies. We look forward to observations from new and upcoming facilities, that will be able to confront our predictions, further constraining the physics that drives the formation of molecules and the evolution of gas in galaxies. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Michael Berry, Ann Martin, Desika Narayanan, Danail Obreschkow, Linda Tacconi, and Martin Zwaan for providing observational data and Marc Verheijen for stimulating discussions. We thank the anonymous referee for many suggestions that have improved the paper. GP acknowledges NOVA (Nederlandse Onderzoekschool voor Astronomie) and LKBF (Leids Kerkhoven-Bosscha Fonds) for funding and the Physics and Astronomy department of Rutgers University for its hospitality. [^1]: E-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'Enfant terrible of charmonium spectroscopy, the resonance $X(3872)$, generated a stream of interpretations and ushered in a new exotic $XYZ$ spectroscopy. In the meantime, many (if not all) characteristics of $X(3872)$ are rather ambiguous. We construct spectra of decays of the resonance $X(3872)$ with good analytical and unitary properties which allows to define the branching ratio of the $X(3872) \to D^{*0}\bar D^0 + c.c.$ decay studying only one more decay, for example, the $X(3872)\to\pi^+\pi^- J/\psi(1S)$ decay. We next define the range of values of the coupling constant of the $X(3872)$ resonance with the $D^{*0}\bar D^0$ system. Finally, we show that our spectra are effective means of selection of models for the resonance $X(3872)$.' author: - 'N.N. Achasov$^{\,a}$ and E.V. Rogozina$^{\,a,b}$' title: | How learn the branching ratio $X(3872) \to D^{*0}\bar D^0 + c.c.$ --- Introduction ============ Discovery of the $X(3872)$ resonance became the first in discovery of the resonant structures $XYZ$ ($X(3872)$, $Y(4260)$, $Z_b^+(10610)$, $Z_b^+(10650)$, $Z_c^+(3900)$), the resonant interpretations of which assumes existence in them at least pair of heavy and pair of light quarks in this or that form. Thousand articles on this subject already were published in spite of the fact that many properties of new resonant structures are not defined yet and not all possible mechanisms of dynamic generation of these structures are studied, in particular, the role of the anomalous Landau thresholds is not studied. Below we suggest an approach which allows to define the branching ratio of the $X(3872) \to D^{*0}\bar D^0 + c.c.$ decay studying only one more decay of $X(3872)$ into a non-$D^{*0}\bar D^0$ channel and to select models the $X(3872)$ resonance. The paper is organized as follows. In Sec. 2 we construct spectra of decays of the resonance $X(3872)$ with good analytical and unitary properties and define the range of values of the coupling constant of the $X(3872)$ resonance with the $D^{*0}\bar D^0$ system. In Sec. 3 we show that the constructed spectra can effectively select the model proposed for the $X(3872)$ resonance. Spectra and coupling constant ============================= The $X(3872) \to D^{*0}\bar D^0 + c.c.$ and others spectra ---------------------------------------------------------- The mass spectrum $\pi^+\pi^-J/\psi(1S)$ in the $X(3872)\to\pi^+\pi^-J/\psi(1S) $ decay [@Belle11] looks as the ideal Breit-Wigner one, see Fig. \[fig1\]a. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ![a) The Belle data [@Belle11] on the invariant $\pi^+\pi^- J/\psi(1S)$ mass ($m$) distribution. The solid line is our theoretical one with taking into account the Belle energy resolution. The dotted line is second-order polynomial for the incoherent background. b) Our undressed theoretical line. []{data-label="fig1"}](Oth.eps "fig:"){width="8.5cm" height="5.9cm"} ![a) The Belle data [@Belle11] on the invariant $\pi^+\pi^- J/\psi(1S)$ mass ($m$) distribution. The solid line is our theoretical one with taking into account the Belle energy resolution. The dotted line is second-order polynomial for the incoherent background. b) Our undressed theoretical line. []{data-label="fig1"}](OthB.eps "fig:"){width="8.5cm" height="5.9cm"} (a) (b) -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- The mass spectrum $\pi^+\pi^-\pi^0 J/\psi(1S)$ in the $X(3872)\to\pi^+\pi^-\pi^0 J/\psi(1S)$ decay looks in a similar way [@Belle05; @BABAR10]. The mass spectrum $D^{*0}\bar D^0 + c.c.$ in the $X(3872)\to D^{*0}\bar D^0 + c.c.$ decay [@Belle10] looks as the typical resonance threshold enhancement, see Fig. \[fig2\]. [^1] ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ![The Belle data [@Belle10] on the invariant $D^{*0}\bar D^0 + c.c.$ mass ($m$) distribution. The solid line is our theoretical one with taking into account the Belle energy resolution. The dotted line is a square root function for the incoherent background. a) $D^{*0}\to D^0\pi^0$. b) $D^{*0}\to ![The Belle data [@Belle10] on the invariant $D^{*0}\bar D^0 + c.c.$ mass ($m$) distribution. The solid line is our theoretical one with taking into account the Belle energy resolution. The dotted line is a square root function for the incoherent background. a) $D^{*0}\to D^0\pi^0$. b) $D^{*0}\to D^0\gamma$. []{data-label="fig2"}](DD1.eps "fig:"){width="8.5cm" height="6cm"} D^0\gamma$. []{data-label="fig2"}](DD2.eps "fig:"){width="8.5cm" height="6cm"} (a) (b) ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- If structures in the above channels are manifestation of the same resonance, it is possible to define the branching ratio $X(3872) \to D^{*0}\bar D^0 + c.c.$, $BR(X(3872) \to D^{*0}\bar D^0 + c.c.)$ treating data only these (two) decay channels. We believe that the $X(3872)$ is the axial vector, $1^{++}$ [@LHCb; @PDG12]. In this case the S wave dominates in the $X(3872) \to D^{*0}\bar D^0 + c.c.$ decay and hence is described by the effective Lagrangian $$\label{L} L_{XD^{*0}D^0}(x)=g_AX^\mu\Bigl(D^0_\mu(x)\bar D^0(x)+ \bar D^0_\mu(x)D^0(x)\Bigr ).$$ The width of the $X\to D^{*0}\bar D^0 + c.c.$ decay $$\label{width} \Gamma(X \to D^{*0}\bar D^0 + c.c.\,,\, m )=\frac{g_A^2}{8\pi}\frac{\rho(m)}{m}\Biggl(1+\frac{{\bf k}^2}{3m_{D^{*0}}^2}\Biggr)\,,$$ where $\bf k$ is momenta of $D^{*0}$ (or $\bar D^0$) in the $D^{*0}\,\bar D^0$ center mass system, $m$ is the invariant mass of the $D^{*0}\,\bar D^0$ pair, $$\label{rho} \rho(m)=\frac{2|{\bf k}|}{m}= \frac{\sqrt{(m^2-m_+^2)(m^2-m_-^2)}}{m^2}\,,\ \ \ \ m_\pm =m_{D^{*0}}\pm m_{D^0}\,.$$ The second term in the right side of Eq. (\[width\]) is very small in our energy region and can be neglected. This gives us the opportunity to construct the mass spectra for the $X(3872)$ decays with the good analytical and unitary properties as in the scalar meson case [@ads; @nna-avk]. The mass spectrum in the $D^{*0}\bar D^0 + c.c.$ channel $$\label{SpectrumDD*} \frac{ dBR(X\to D^{*0}\bar D^0+c.c.\,,\, m )}{dm} =4\frac{1}{\pi}\frac{m^2\Gamma(X\to D^{*0}\bar D^0,\, m)}{|D_X(m)|^2}.$$ The branching ratio of $X(3872) \to D^{*0}\bar D^0 + c.c.$ $$\label{BRDD*} BR(X\to D^{*0}\bar D^0+c.c.) = 4\frac{1}{\pi}\int_{m_+}^\infty \frac{m^2\Gamma(X\to D^{*0}\bar D^0,\, m)}{|D_X(m)|^2} dm\,.$$ In others $\{i\}$ (non-$D^{*0}\bar D^0$) channels the $X(3872)$ state is seen as a narrow resonance that is why we write the mass spectrum in the $i$ channel in the form $$\label{Spectrumi} \frac{dBR(X\to i\,,\, m )}{dm} =2\frac{1}{\pi}\frac{m_X^2\,\Gamma_i}{|D_X(m)|^2}\,,$$ where $\Gamma_i$ is the width of the $X(3872)\to i$ decay. The branching ratio of $X(3872)\to i$ $$\label{BRi} BR(X\to i) = 2\frac{1}{\pi}\int_{m_0}^\infty \frac{m_X^2\Gamma_i}{|D_X(m)|^2} dm\,,$$ where $m_0$ is the threshold of the $i$ state. $$\label{inverseprop} D_X(m)= m_X^2-m^2 + Re(\Pi^{D^{*0}\bar D^0}_X(m_X))- \Pi^{D^{*0}\bar D^0}_X(m)-\imath m_X\Gamma\,,$$ where $\Gamma=\Sigma\Gamma_i$ is the total width of the $X(3872)$ decay into all non-$D^{*0}\bar D^0$ channels. When $m_+ \leq m$, $$\label{mm+} \hspace*{-21pt} \Pi^{D^{*0}\bar D^0}_X(m)=\frac{g_A^2}{8\pi^2}\left\{ \frac{(m^2-m_+^2)}{m^2}\frac{m_-}{m_+}\ln\frac{m_{D^{*0}}}{m_{D^0}} +\rho(m)\left [\imath\pi + \ln\frac{\sqrt{m^2-m_-^2}-\sqrt{m^2-m_+^2}}{\sqrt{m^2-m_-^2}+\sqrt{m^2-m_+^2}} \right ]\right\}.$$ When $m_- \leq m\leq m_+$, $$\label{m-mm+} \Pi^{D^{*0}\bar D^0}_X(m)=\frac{g_A^2}{8\pi^2}\left\{ \frac{(m^2-m_+^2)}{m^2}\frac{m_-}{m_+}\ln\frac{m_{D^{*0}}}{m_{D^0}} -2|\rho(m)|\arctan\frac{\sqrt{m^2-m_-^2}}{\sqrt{m_+^2-m^2}}\right\},$$ where $|\rho(m)|= \sqrt{(m_+^2-m^2)(m^2-m_-^2)}/m^2$. When $m\leq m_-$ and $m^2\leq 0$, $$\label{mm-} \Pi^{D^{*0}\bar D^0}_X(m)=\frac{g_A^2}{8\pi^2}\left\{ \frac{(m^2-m_+^2)}{m^2}\frac{m_-}{m_+}\ln\frac{m_{D^{*0}}}{m_{D^0}} -\rho(m)\ln\frac{\sqrt{m_+^2-m^2}-\sqrt{m_-^2-m^2}}{\sqrt{m_+^2-m^2}+\sqrt{m_-^2-m^2}} \right\}.$$ Our branching ratios satisfy unitarity $$\label{unitarity} 1= BR(X\to D^{*0}\bar D^0+c.c.)+\sum_iBR(X\to i)\,.$$ Fitting the Belle data [@Belle11; @Belle10], we take into account the Belle [@Belle11] results that $m_X= 3871.84\,\mbox{MeV}= m_{D^{*0}}+ m_{D^0}= m_+$ and $\Gamma_{X(3872)}<1.2$ MeV 90%CL that corresponds to $\Gamma <1.2$ MeV, which controls the width of the $X(2872)$ signal in the $\pi^+\pi^- J/\psi(1S)$ channel and in every non-$D^{*0}\bar D^0$ channel, see Fig. \[fig1\]b. The results of our fit are in the Table I. The current statistics is not sufficient for serious conclusions.\ \[12pt\] TABLE I. Results of the analysis of the Belle data [@Belle11; @Belle10].\ $BR_{seen}=BR(X\to D^{*0}\bar D^0+c.c.\,; m\leq 3891.84\,\mbox{MeV})$, $BR=BR(X\to D^{*0}\bar D^0+c.c.)$,\ $BR(Oth)_{seen}=\sum_iBR(X\to i\,;\, 3851.84\leq m\leq 3891.84\,\mbox{MeV})$. $\Gamma$ in MeV, $g_A$ in GeV.\ $\Gamma$ $g_A^2/{8\pi}$ $\chi^2/Ndf$ $BR_{seen}$ $BR$ $BR(Oth)_{seen}$ ---------------- --------------------------- -------------- -------------------------- -------------------------- --------------------------- $1.2_{-0.467}$ $0.857_{-0.481}^{+3.614}$ 43.74/42 $0.486_{-0.29}^{+0.061}$ $0.795_{-0.224}^{+0.19}$ $0.191_{-0.179}^{+0.223}$ \ Nevertheless, one can state that our results are consist with experiment. Really, in view of\ $BR(B\to X(3872)K)\times BR(X(3872)\to D^{*0}\bar D^0) = (0.80\pm 0.20\pm 0.1)\times 10^{-4}$ [@Belle10],\ $BR(B^+\to X(3872)K^+)\times BR(X(3872)\to\pi^+\pi^- J/\psi(1S)) = (8.61\pm 0.82\pm 0.52)\times 10^{-6}$ [@Belle11],\ $BR(B^+\to X(3872)K^+)\times BR(X(3872)\to\pi^+\pi^-\pi^0 J/\psi(1S)) = (0.6\pm 0.2\pm 0.1)\times 10^{-5}$ [@BABAR10],\ and $BR(B^+\to X(3872)K^+)\times BR(X(3872)\to\gamma J/\psi(1S)) = (1.78^{+0.48}_{-0.44}\pm 0.12)\times 10^{-6}$ [@Belle11g]\ it follows that $BR(X\to D^{*0}\bar D^0+c.c.\,; m\leq 3892\,\mbox{MeV})$ is a few times as large as the sum of all non-$D^{*0}\bar D^0$ known branching ratios. So, when fitting the $X(3872)\to D^{*0}\bar D^0$ data and data for any $X(3872)$ decay into non-$D^{*0}\bar D^0$ state, $X(3872)\to i $, we find $\Gamma$ and $g_A^2/8\pi$, which define $BR(X(3872)\to D^{*0}\bar D^0+c.c.)$. Generally speaking, we don’t need to know $BR(X(3872)\to i)$. Influence of the $X(3872) \to D^{*+}\bar D^- + c.c.$ channel ------------------------------------------------------------ As seen from Table I the sizeable part (near 40%) of $BR=BR(X\to D^{*0}\bar D^0+c.c.)$ accounts for the tail of the $X(3872)$ resonance ($m\geq 3891.84$ MeV). This gives an idea to take into account the $X(3872)\to D^{*+}D^- +c.c. $ decays [^2] on the $X(3872)$ tail. Since $X(3872)$ is an isoscalar, the effective Lagrangian has the form $$\label{L2} L(x)=g_AX^\mu\Bigl(D^0_\mu(x)\bar D^0(x)+ \bar D^0_\mu(x)D^0(x)+ D^+_\mu(x)D^-(x)+ D^+_\mu(x)D^-(x)\Bigr ).$$ Eq. \[inverseprop\] is replaced by $$\label{inverseprop2} D_X(m)= m_X^2-m^2 + Re(\Pi_X(m_X))- \Pi_X(m)-\imath m_X\Gamma\,,$$ where $$\label{Pi(m)} \Pi_X(m)= \Pi^{D^{*0}\bar D^0}_X(m)+ \Pi^{D^{*+}D^-}_X(m)\,.$$ $\Pi^{D^{*+}D^-}_X(m)$ is obtained from $\Pi^{D^{*0}\bar D^0}_X(m)$, see Eqs. (\[mm+\]), (\[m-mm+\]) and (\[mm-\]), by replacement of $m_{ D^{*0}}$ and $m_{ D^0}$ by $m_{ D^{*+}}$ and $m_{ D^+}$, respectively. The unitarity condition, Eq. (\[unitarity\]), takes the form $$\label{unitarity2} 1= BR(X\to D^{*0}\bar D^0+c.c.)+ BR(X\to D^{*+}\bar D^-+c.c.) + \sum_iBR(X\to i)\,.$$ The results of our fit are in the Table II.\ TABLE II. $\Gamma$ in MeV, $g_A$ in GeV.\ ------------------------------------------------------------------------------------------------------------------------------------------- $\Gamma$ $1.2_{-0.42}$ mode $X\to D^{*0}\bar $X\to D^{*+}D^{-}+c.c.$ $X\to Others$ D^0+c.c.$ ---------------- ------------------------ ------------- --------------------------- --------------------------- --------------------------- $g_A^2/{8\pi}$ $1.36_{-0.95}^{+4.85}$ $BR$ $0.586_{-0.101}^{+0.025}$ $0.315_{-0.16}^{+0.132}$ $0.098_{-0.096}^{+0.261}$ $\chi^2/Ndf$ 45.49/42 $BR_{seen}$ $0.285_{-0.188}^{+0.121}$ $0.028_{-0.019}^{+0.004}$ $0.091_{-0.084}^{+0.255}$ ------------------------------------------------------------------------------------------------------------------------------------------- \ The results in Tables I and II are compatible within the errors. The corresponding curves are similar to ones in Figs. \[fig1\] and \[fig2\]. Of course, one should take into account the $X(3872)\to D^{*+}D^- +c.c. $ channel in the case of the good statistics. conclusion ========== Our approach can serve as the guide in selection of theoretical models for the $X(3872)$ resonance. Indeed, if $3871.68$ MeV $<m_X<3871.95$ MeV [@PDG12] and $\Gamma_{X(3872)}=\Gamma <1.2$ MeV [@PDG12] then for $g_A^2/8\pi<0.2$ GeV$^2$ (that does not contradict current experiment, see Tables I and II) $BR(X\to D^{*0}\bar D^0+c.c.)=BR<0.3$. That is, unknown decays of $X(3872)$ into non-$D^{*0}\bar D^0$ states are considerable or dominant. For example, in Ref. [@Maiani] the authors considered $m_X=3871.68$ MeV, $\Gamma=1.2$ MeV and $g_{XDD^*}=g_A\sqrt{2}=2.5$ GeV, that is, $g_A^2/8\pi=0.1$ GeV$^2$. In this case $BR(X\to D^{*0}\bar D^0+c.c.)\approx 0.2$, that is, unknown decays $X(3872)$ into non-$D^{*0}\bar D^0$ states are dominant. For details see Table III.\ TABLE III. Branching ratios for the model from Ref. [@Maiani] without $D^{*+}D^{-}+c.c.$ channel. $m_X$ in MeV, $\Gamma$ in MeV, $g_A$ in GeV.\ $m_X$ $\Gamma$ $g_A^2/{8\pi}$ $BR_{seen}$ $BR$ $BR(Oth)_{seen}$ ----------- ---------- ---------------- ------------- --------- ------------------ $3871.68$ $1.2$ $0.1$ $0.152$ $0.189$ $0.792$ \ Account of influence of the $X(3872) \to D^{*+}\bar D^- + c.c.$ channel has little effect on the results for the model from Ref. [@Maiani] since $g_A^2/{8\pi}$ is small , see Table IV.\ TABLE IV. Branching ratios for the model from Ref. [@Maiani]. $m_X$ in MeV, $\Gamma$ in MeV, $g_A$ in GeV.\ ----------------------------------------------------------------------------------------------------------------- $m $3871.68$ mode $X\to D^{*0}\bar D^0+c.c.$ $X\to D^{*+}D^{-}+c.c.$ $X\to Others$ _X$ ---------------- ----------- ------------- ---------------------------- ------------------------- --------------- $\Gamma$ $1.2$ $BR$ $0.176$ $0.045$ $0.779$ $g_A^2/{8\pi}$ $0.1$ $BR_{seen}$ $0.14$ $0.011$ $0.761$ ----------------------------------------------------------------------------------------------------------------- [**Acknowledgments.**]{} We are grateful to T.A.-Kh. Aushev and G.V. Pakhlova for the consultation. This work was supported in part by RFBR, Grant No 13-02-00039, and Interdisciplinary project No 102 of Siberian division of RAS.\ [99]{} S.K. Choi, S.L. Olsen, K. Trabelsi et al. (Belle Callaboration), Phys. Rev. D [**84**]{}, 052004 (2011). K. Abe, K. Abe, I. Adachi et al. (Belle Callaboration), arXiv:hep-ex/0505037. P. del Amo Sanchez, J.P. Lees, V. Poireau et al. (BABAR Callaboration), Phys. Rev. D(R) [**82**]{}, 011101 (2010). T. Aushev, N. Zwahlen, I. Adachi et al. (Belle Callaboration), Phys. Rev. D [**81**]{}, 031103 (2010). R. Aaij, C. Abellan Beteta, B. Adeva et al. (LHCb Callaboration), Phys. Rev. Lett. [**110**]{}, 222001 (2013). J. Beringer, J.-F. Arguin, R.M. Barnett et al. (Particle Data Group), Phys. Rev. D [**86**]{}, 010001 (2012) and 2013 partial update for the 2014 edition. N.N. Achasov, S.A. Devyanin, and G.N. Shestakov, Yad. Fiz. [**32**]{}, 1098 (1980) \[Sov. J. Nucl. Phys. [**32**]{} 566 (1980)\]. N.N. Achasov and A.V. Kiselev, Phys. Rev. D [**70**]{}, 111901(R) (2004). V. Bhardwaj, K. Trabelsi, J.B. Singh et al. (Belle Callaboration), Phys. Rev. Lett. [**107**]{}, 9 (2011). L. Maiani, V. Riquer, R. Faccini, F. Piccinini, A. Pilloni, and A.D. Polosa, Phys. Rev. D [**87**]{}, 111102(R) (2013). [^1]: An interference between the $D^0\bar D^{*0}$ and $\bar D^0 D^{*0}$ channels is negligible for the narrowness of the $D^{*0}$ and $\bar D^{*0}$ states. Using the isotopical invariance of the $DD^*\pi$ interacion and the experimental information about the $D^{*+}\to D^0\pi^+$ and $D^{*+}\to D^+\pi^0$ decays, one can find $\Gamma_{D^{*0}}= 70$ KeV. [^2]: An interference between the $D^+ D^{*-}$ and $D^- D^{*+}$ channels is negligible for the narrowness of the $D^{*\pm}$ states, $\Gamma_{D^{*\pm}}= 96$ KeV.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Traditional approaches to extractive summarization rely heavily on human-engineered features. In this work we propose a data-driven approach based on neural networks and continuous sentence features. We develop a general framework for single-document summarization composed of a hierarchical document encoder and an attention-based extractor. This architecture allows us to develop different classes of summarization models which can extract sentences or words. We train our models on large scale corpora containing hundreds of thousands of document-summary pairs[^1]. Experimental results on two summarization datasets demonstrate that our models obtain results comparable to the state of the art without any access to linguistic annotation.' author: - | Jianpeng Cheng Mirella Lapata\ ILCC, School of Informatics, University of Edinburgh\ 10 Crichton Street, Edinburgh EH8 9AB\ [[email protected] [email protected]]{} bibliography: - 'acl2016.bib' title: Neural Summarization by Extracting Sentences and Words --- Introduction {#sec:introduction} ============ The need to access and digest large amounts of textual data has provided strong impetus to develop automatic summarization systems aiming to create shorter versions of one or more documents, whilst preserving their information content. Much effort in automatic summarization has been devoted to sentence extraction, where a summary is created by identifying and subsequently concatenating the most salient text units in a document. Most extractive methods to date identify sentences based on human-engineered features. These include surface features such as sentence position and length [@radev2004mead], the words in the title, the presence of proper nouns, content features such as word frequency [@nenkova2006compositional], and event features such as action nouns [@filatova2004event]. Sentences are typically assigned a score indicating the strength of presence of these features. Several methods have been used in order to select the summary sentences ranging from binary classifiers [@kupiec:ea:95], to hidden Markov models [@conroy:oleary:2011], graph-based algorithms [@erkan-radev:2004:EMNLP; @mihalcea:2005:PosterDemo], and integer linear programming [@woodsend2010automatic]. In this work we propose a data-driven approach to summarization based on neural networks and continuous sentence features. There has been a surge of interest recently in repurposing sequence transduction neural network architectures for NLP tasks such as machine translation [@sutskever2014sequence], question answering [@hermann2015teaching], and sentence compression [@rush2015neural]. Central to these approaches is an encoder-decoder architecture modeled by recurrent neural networks. The encoder reads the source sequence into a list of continuous-space representations from which the decoder generates the target sequence. An attention mechanism [@bahdanau2014neural] is often used to locate the region of focus during decoding. We develop a general framework for single-document summarization which can be used to extract sentences or words. Our model includes a neural network-based hierarchical document reader or encoder and an attention-based content extractor. The role of the reader is to derive the meaning representation of a document based on its sentences and their constituent words. Our models adopt a variant of neural attention to extract sentences or words. Contrary to previous work where attention is an *intermediate* step used to blend hidden units of an encoder to a vector propagating additional information to the decoder, our model applies attention *directly* to select sentences or words of the input document as the output summary. Similar neural attention architectures have been previously used for geometry reasoning [@vinyals2015pointer], under the name *Pointer Networks*. One stumbling block to applying neural network models to extractive summarization is the lack of training data, i.e., documents with sentences (and words) labeled as summary-worthy. Inspired by previous work on summarization [@woodsend2010automatic; @svore-vanderwende-burges:2007:EMNLP-CoNLL2007] and reading comprehension [@hermann2015teaching] we retrieve hundreds of thousands of news articles and corresponding highlights from the DailyMail website. Highlights usually appear as bullet points giving a brief overview of the information contained in the article (see Figure \[example\] for an example). Using a number of transformation and scoring algorithms, we are able to match highlights to document content and construct two large scale training datasets, one for sentence extraction and the other for word extraction. Previous approaches have used small scale training data in the range of a few hundred examples. Our work touches on several strands of research within summarization and neural sequence modeling. The idea of creating a summary by extracting words from the source document was pioneered in who view summarization as a problem analogous to statistical machine translation and generate headlines using statistical models for selecting and ordering the summary words. Our word-based model is similar in spirit, however, it operates over continuous representations, produces multi-sentence output, and jointly selects summary words and organizes them into sentences. A few recent studies [@kobayashisummarization; @yogatamaextractive] perform sentence extraction based on pre-trained sentence embeddings following an unsupervised optimization paradigm. Our work also uses continuous representations to express the meaning of sentences and documents, but importantly employs neural networks more directly to perform the actual summarization task. propose a neural attention model for abstractive sentence compression which is trained on pairs of headlines and first sentences in an article. In contrast, our model summarizes documents rather than individual sentences, producing multi-sentential discourse. A major architectural difference is that our decoder selects output symbols from the document of interest rather than the entire vocabulary. This effectively helps us sidestep the difficulty of searching for the next output symbol under *a large vocabulary*, with *low-frequency words* and *named entities* whose representations can be challenging to learn. and propose a similar “copy” mechanism in sentence compression and other tasks; their model can accommodate both generation and extraction by selecting which sub-sequences in the input sequence to copy in the output. We evaluate our models both automatically (in terms of <span style="font-variant:small-caps;">Rouge</span>) and by humans on two datasets: the benchmark DUC 2002 document summarization corpus and our own DailyMail news highlights corpus. Experimental results show that our summarizers achieve performance comparable to state-of-the-art systems employing hand-engineered features and sophisticated linguistic constraints. [|@[ ]{}p[15.5cm]{}@[ ]{}|]{} **AFL star blames vomiting cat for speeding**\ \ The 22-year-old AFL star, who drove 96km/h in a 60km/h road works zone on the South Eastern expressway in February, said he didn’t see the reduced speed sign because he was so distracted by his cat vomiting violently in the back seat of his car.\ \ He lost four demerit points, instead of seven, because of his significant training commitments.\ [=1em =-.8ex]{} *Adelaide Crows defender Daniel Talia admits to speeding but says he didn’t see road signs because his cat was vomiting in his car.* *22-year-old Talia was fined \$824 and four demerit points, instead of seven, because of his ’significant’ training commitments.* \ Problem Formulation {#sec:problem-formulation} =================== In this section we formally define the summarization tasks considered in this paper. Given a document $D$ consisting of a sequence of sentences $\{s_1, \cdots, s_m\}$ and a word set $\{w_1, \cdots, w_n\}$, we are interested in obtaining summaries at two levels of granularity, namely sentences and words. **Sentence extraction** aims to create a summary from $D$ by selecting a subset of $j$ sentences (where ). We do this by scoring each sentence within $D$ and predicting a label  indicating whether the sentence should be included in the summary. As we apply supervised training, the objective is to maximize the likelihood of all sentence labels given the input document $D$ and model parameters $\theta$: $$\log p(\mathbf{y}_L |D; \theta) = \sum\limits_{i=1}^{m} \log p(y_L^i |D; \theta) \label{sentence objective}$$ Although extractive methods yield naturally grammatical summaries and require relatively little linguistic analysis, the selected sentences make for long summaries containing much redundant information. For this reason, we also develop a model based on **word extraction** which seeks to find a subset of words[^2] in $D$ and their optimal ordering so as to form a summary . Compared to sentence extraction which is a sequence labeling problem, this task occupies the middle ground between full abstractive summarization which can exhibit a wide range of rewrite operations and extractive summarization which exhibits none. We formulate word extraction as a language generation task with an output vocabulary restricted to the original document. In our supervised setting, the training goal is to maximize the likelihood of the generated sentences, which can be further decomposed by enforcing conditional dependencies among their constituent words: $$\hspace*{-.2cm}\log p(\mathbf{y}_s |D; \theta)\hspace*{-.1cm}=\hspace*{-.1cm}\sum\limits_{i=1}^{k}\hspace*{-.1cm}\log p(w'_i | D, w'_1,\hspace*{-.1cm}\cdots\hspace*{-.1cm}, w'_{i-1}; \theta) \label{word objective}$$ In the following section, we discuss the data elicitation methods which allow us to train neural networks based on the above defined objectives. Training Data for Summarization {#sec:train-data-summ} =============================== Data-driven neural summarization models require a large training corpus of documents with labels indicating which sentences (or words) should be in the summary. Until now such corpora have been limited to hundreds of examples (e.g., the DUC 2002 single document summarization corpus) and thus used mostly for testing [@woodsend2010automatic]. To overcome the paucity of annotated data for training, we adopt a methodology similar to and create two large-scale datasets, one for sentence extraction and another one for word extraction. In a nutshell, we retrieved[^3] hundreds of thousands of news articles and their corresponding highlights from DailyMail (see Figure \[example\] for an example). The highlights (created by news editors) are genuinely abstractive summaries and therefore not readily suited to supervised training. To create the training data for **sentence extraction**, we reverse approximated the gold standard label of each document sentence given the summary based on their semantic correspondence [@woodsend2010automatic]. Specifically, we designed a rule-based system that determines whether a document sentence matches a highlight and should be labeled with 1 (must be in the summary), and 0 otherwise. The rules take into account the position of the sentence in the document, the unigram and bigram overlap between document sentences and highlights, the number of entities appearing in the highlight and in the document sentence. We adjusted the weights of the rules on 9,000 documents with manual sentence labels created by . The method obtained an accuracy of 85% when evaluated on a held-out set of 216 documents coming from the same dataset and was subsequently used to label 200K documents. Approximately 30% of the sentences in each document were deemed summary-worthy. For the creation of the **word extraction** dataset, we examine the lexical overlap between the highlights and the news article. In cases where all highlight words (after stemming) come from the original document, the document-highlight pair constitutes a valid training example and is added to the word extraction dataset. For out-of-vocabulary (OOV) words, we try to find a semantically equivalent replacement present in the news article. Specifically, we check if a neighbor, represented by pre-trained[^4] embeddings, is in the original document and therefore constitutes a valid substitution. If we cannot find any substitutes, we discard the document-highlight pair. Following this procedure, we obtained a word extraction dataset containing 170K articles, again from the DailyMail. Neural Summarization Model {#sec:model} ========================== The key components of our summarization model include a neural network-based hierarchical document reader and an attention-based hierarchical content extractor. The hierarchical nature of our model reflects the intuition that documents are generated compositionally from words, sentences, paragraphs, or even larger units. We therefore employ a representation framework which reflects the same architecture, with global information being discovered and local information being preserved. Such a representation yields minimum information loss and is flexible allowing us to apply neural attention for selecting salient sentences and words within a larger context. In the following, we first describe the document reader, and then present the details of our sentence and word extractors. Document Reader --------------- The role of the reader is to derive the meaning representation of the document from its constituent sentences, each of which is treated as a sequence of words. We first obtain representation vectors at the sentence level using a single-layer convolutional neural network (CNN) with a max-over-time pooling operation [@blunsom2013recurrent; @zhang2014chinese; @kim2015character]. Next, we build representations for documents using a standard recurrent neural network (RNN) that recursively composes sentences. The CNN operates at the word level, leading to the acquisition of sentence-level representations that are then used as inputs to the RNN that acquires document-level representations, in a hierarchical fashion. We describe these two sub-components of the text reader below. #### Convolutional Sentence Encoder We opted for a convolutional neural network model for representing sentences for two reasons. Firstly, single-layer CNNs can be trained effectively (without any long-term dependencies in the model) and secondly, they have been successfully used for sentence-level classification tasks such as sentiment analysis [@kim2014convolutional]. Let $d$ denote the dimension of word embeddings, and $s$ a document sentence consisting of a sequence of $n$ words $(w_1, \cdots, w_n)$ which can be represented by a dense column matrix $\mathbf{W} \in \mathbb{R}^{n \times d}$. We apply a temporal narrow convolution between $\mathbf{W}$ and a kernel $\mathbf{K} \in \mathbb{R}^{c \times d}$ of width $c$ as follows: $$\mathbf{f}^{i}_{j} = \tanh (\mathbf{W}_{j : j+c-1} \otimes \mathbf{K} + b)$$ where $\otimes$ equates to the Hadamard Product followed by a sum over all elements. $\mathbf{f}^i_j $ denotes the $j$-th element of the $i$-th feature map $\mathbf{f}^i$ and $b$ is the bias. We perform max pooling over time to obtain a *single* feature (the $i$th feature) representing the sentence under the kernel $\mathbf{K}$ with width $c$: $$\mathbf{s}_{i, \mathbf{K}}= \max_j \mathbf{f}_j^i$$ In practice, we use multiple feature maps to compute a list of features that match the dimensionality of a sentence under each kernel width. In addition, we apply multiple kernels with different widths to obtain a set of different sentence vectors. Finally, we sum these sentence vectors to obtain the final sentence representation. The CNN model is schematically illustrated in Figure \[sl\] (bottom). In the example, the sentence embeddings have six dimensions, so six feature maps are used under each kernel width. The blue feature maps have width two and the red feature maps have width three. The sentence embeddings obtained under each kernel width are summed to get the final sentence representation (denoted by green). #### Recurrent Document Encoder At the document level, a recurrent neural network composes a sequence of sentence vectors into a document vector. Note that this is a somewhat simplistic attempt at capturing document organization at the level of sentence to sentence transitions. One might view the hidden states of the recurrent neural network as a list of partial representations with each focusing mostly on the corresponding input sentence given the previous context. These representations altogether constitute the document representation, which captures local and global sentential information with minimum compression. The RNN we used has a Long Short-Term Memory (LSTM) activation unit for ameliorating the vanishing gradient problem when training long sequences [@hochreiter1997long]. Given a document , the hidden state at time step $t$, denoted by $\mathbf{h_t}$, is updated as: $$\begin{bmatrix} \mathbf{i}_t\\ \mathbf{f}_t\\ \mathbf{o}_t\\ \mathbf{\hat{c}}_t \end{bmatrix} = \begin{bmatrix} \sigma\\ \sigma\\ \sigma\\ \tanh \end{bmatrix} \mathbf{W}\cdot \begin{bmatrix} \mathbf{h}_{t-1}\\ \mathbf{s}_t \end{bmatrix}$$ $$\mathbf{c}_t = \mathbf{f}_t \odot \mathbf{c}_{t-1} + \mathbf{i}_t \odot \mathbf{\hat{c}}_t$$ $$\mathbf{h}_t = \mathbf{o}_t \odot \tanh(\mathbf{c}_t)$$ where $\mathbf{W}$ is a learnable weight matrix. Next, we discuss a special attention mechanism for extracting sentences and words given the recurrent document encoder just described, starting from the sentence extractor. ![A recurrent convolutional document reader with a neural sentence extractor. []{data-label="sl"}]("sentence".jpg){width="48.00000%"} Sentence Extractor ------------------ In the standard neural sequence-to-sequence modeling paradigm [@bahdanau2014neural], an attention mechanism is used as an intermediate step to decide which input region to focus on in order to generate the next output. In contrast, our sentence extractor applies attention to directly extract salient sentences after reading them. The extractor is another recurrent neural network that labels sentences sequentially, taking into account not only whether they are individually relevant but also mutually redundant. The complete architecture for the document encoder and the sentence extractor is shown in Figure \[sl\]. As can be seen, the next labeling decision is made with both the encoded document and the previously labeled sentences in mind. Given encoder hidden states $(h_1, \cdots, h_m)$ and extractor hidden states $(\bar{h}_1, \cdots, \bar{h}_m)$ at time step $t$, the decoder attends the $t$-th sentence by relating its current decoding state to the corresponding encoding state: $$\bar{\mathbf{h}}_{t} = \text{LSTM} ( p_{t-1} \mathbf{s}_{t-1}, \mathbf{\bar{h}}_{t-1}) \label{update}$$ $$p(y_L(t)=1 | D ) = \sigma(\text{MLP} (\mathbf{\bar{h}}_t : \mathbf{h}_t) )$$ where MLP is a multi-layer neural network with as input the concatenation of $\mathbf{\bar{h}}_t$ and $\mathbf{h}_t$. $p_{t-1}$ represents the degree to which the extractor believes the previous sentence should be extracted and memorized ($p_{t-1}$=1 if the system is certain; 0 otherwise). In practice, there is a discrepancy between training and testing such a model. During training we know the true label $p_{t-1}$ of the previous sentence, whereas at test time $p_{t-1}$ is unknown and has to be predicted by the model. The discrepancy can lead to quickly accumulating prediction errors, especially when mistakes are made early in the sequence labeling process. To mitigate this, we adopt a curriculum learning strategy [@bengio2015scheduled]: at the beginning of training when $p_{t-1}$ cannot be predicted accurately, we set it to the true label of the previous sentence; as training goes on, we gradually shift its value to the predicted label $p(y_L(t-1)=1 | d )$. ![Neural attention mechanism for word extraction.[]{data-label="we"}]("phrase".jpg){width="50.00000%"} Word Extractor -------------- Compared to sentence extraction which is a purely sequence labeling task, word extraction is closer to a generation task where relevant content must be selected and then rendered fluently and grammatically. A small extension to the structure of the sequential labeling model makes it suitable for generation: instead of predicting a label for the next sentence at each time step, the model directly outputs the next word in the summary. The model uses a *hierarchical* attention architecture: at time step $t$, the decoder softly[^5] attends each document sentence and subsequently attends each word in the document and computes the probability of the next word to be included in the summary $p(w'_t = w_i| d, w'_1, \cdots, w'_{t-1})$ with a softmax classifier: $$\bar{\mathbf{h}}_{t} = \text{LSTM} ( \mathbf{w'}_{t-1}, \mathbf{\bar{h}}_{t-1})\footnote{We empirically found that feeding the previous sentence-level attention vector as additional input to the LSTM would lead to small performance improvements. This is not shown in the equation.}$$ $$a_j^t = \mathbf{z}^\mathtt{T} \tanh(\mathbf{W}_e \mathbf{\bar{h}}_t + \mathbf{W}_r \mathbf{h}_j), h_j \in D$$ $$b_j^t = \text{softmax}(a_j^t)$$ $$\mathbf{\tilde{h}}_t = \sum\limits_{j=1}^{m} b_j^t \mathbf{h}_j$$ $$u_i^t = \mathbf{v}^\mathtt{T} \tanh(\mathbf{W}_{e'} \mathbf{\tilde{h}}_t + \mathbf{W}_{r'} \mathbf{w}_i), w_i \in D$$ $$p(w'_t = w_i| D, w'_1, \cdots, w'_{t-1}) = \text{softmax}(u_i^t)$$ In the above equations, $\mathbf{w}_i$ corresponds to the vector of the $i$-th word in the input document, whereas $\mathbf{z}$, $\mathbf{W}_e$, $\mathbf{W}_r$, $\mathbf{v}$, $\mathbf{W}_{e'}$, and $\mathbf{W}_{r'}$ are model weights. The model architecture is shown in Figure \[we\]. The word extractor can be viewed as a conditional language model with a vocabulary constraint. In practice, it is not powerful enough to enforce grammaticality due to the lexical diversity and sparsity of the document highlights. A possible enhancement would be to pair the extractor with a neural language model, which can be pre-trained on a large amount of unlabeled documents and then jointly tuned with the extractor during decoding [@gulcehre2015using]. A simpler alternative which we adopt is to use $n$-gram features collected from the document to rerank candidate summaries obtained via beam decoding. We incorporate the features in a log-linear reranker whose feature weights are optimized with minimum error rate training [@och2003minimum]. Experimental Setup ================== In this section we present our experimental setup for assessing the performance of our summarization models. We discuss the datasets used for training and evaluation, give implementation details, briefly introduce comparison models, and explain how system output was evaluated. #### Datasets We trained our sentence- and word-based summarization models on the two datasets created from DailyMail news. Each dataset was split into approximately 90% for training, 5% for validation, and 5% for testing. We evaluated the models on the DUC-2002 single document summarization task. In total, there are 567 documents belonging to 59 different clusters of various news topics. Each document is associated with two versions of 100-word[^6] manual summaries produced by human annotators. We also evaluated our models on 500 articles from the DailyMail test set (with the human authored highlights as goldstandard). We sampled article-highlight pairs so that the highlights include a minimum of 3 sentences. The average byte count for each document is 278. As there is no established evaluation standard for this task, we also report ROUGE evaluation on the entire DailyMail test set with varying limits. Please refer to the appendix for more information. #### Implementation Details We trained our models with Adam [@kingma2014adam] with initial learning rate 0.001. The two momentum parameters were set to 0.99 and 0.999 respectively. We performed mini-batch training with a batch size of 20 documents. All input documents were padded to the same length with an additional mask variable storing the real length for each document. The size of word, sentence, and document embeddings were set to 150, 300, and 750, respectively. For the convolutional sentence model, we followed [^7] and used a list of kernel sizes {1, 2, 3, 4, 5, 6, 7}. For the recurrent document model and the sentence extractor, we used as regularization dropout with probability 0.5 on the LSTM input-to-hidden layers and the scoring layer. The depth of each LSTM module was 1. All LSTM parameters were randomly initialized over a uniform distribution within . The word vectors were initialized with 150 dimensional pre-trained embeddings.[^8] Proper nouns pose a problem for [embedding-based]{} approaches, especially when these are rare or unknown (e.g., at test time). address this issue by adding a new set of features and a log-linear model component to their system. As our model enjoys the advantage of generation by extraction, we can force the model to inspect the context surrounding an entity and its relative position in the sentence in order to discover extractive patterns, placing less emphasis on the meaning representation of the entity itself. Specifically, we perform named entity recognition with the package provided by and maintain a set of randomly initialized entity embeddings. During training, the index of the entities is permuted to introduce some noise but also robustness in the data. A similar data augmentation approach has been used for reading comprehension [@hermann2015teaching]. A common problem with extractive methods based on sentence labeling is that there is no constraint on the number of sentences being selected at test time. We address this by reranking the positively labeled sentences with the probability scores obtained from the softmax layer (rather than the label itself). In other words, we are more interested in is the relative ranking of each sentence rather than their exact scores. This suggests that an alternative to training the network would be to employ a ranking-based objective or a *learning to rank* algorithm. However, we leave this to future work. We use the three sentences with the highest scores as the summary (also subject to the word or byte limit of the evaluation protocol). Another issue relates to the word extraction model which is challenging to batch since each document possesses a distinct vocabulary. We sidestep this during training by performing negative sampling [@mikolov2013distributed] which trims the vocabulary of different documents to the same length. At each decoding step the model is trained to differentiate the true target word from 20 noise samples. At test time we still loop through the words in the input document (and a stop-word list) to decide which word to output next. #### System Comparisons We compared the output of our models to various summarization methods. These included the standard baseline of simply selecting the “leading” three sentences from each document as the summary. We also built a sentence extraction baseline classifier using logistic regression and human engineered features. The classifier was trained on the same datasets as our neural network models with the following features: sentence length, sentence position, number of entities in the sentence, sentence-to-sentence cohesion, and sentence-to-document relevance. Sentence-to-sentence cohesion was computed by calculating for every document sentence its embedding similarity with every other sentence in the same document. The feature was the normalized sum of these similarity scores. Sentence embeddings were obtained by averaging the constituent word embeddings. Sentence-to-document relevance was computed similarly. We calculated for each sentence its embedding similarity with the document (represented as bag-of-words), and normalized the score. The word embeddings used in this baseline are the same as the pre-trained ones used for our neural models. In addition, we included a neural abstractive summarization baseline. This system has a similar architecture to our word extraction model except that it uses an open vocabulary during decoding. It can also be viewed as a hierarchical document-level extension of the abstractive sentence summarizer proposed by . We trained this model with negative sampling to avoid the excessive computation of the normalization constant. Finally, we compared our models to three previously published systems which have shown competitive performance on the DUC2002 single document summarization task. The first approach is the phrase-based extraction model of . Their system learns to produce highlights from parsed input (phrase structure trees and dependency graphs); it selects salient phrases and recombines them subject to length, coverage, and grammar constraints enforced via integer linear programming (ILP). Like ours, this model is trained on document-highlight pairs, and produces telegraphic-style bullet points rather than full-blown summaries. The other two systems, <span style="font-variant:small-caps;">tgraph</span> [@parveen-ramsl-strube] and <span style="font-variant:small-caps;">urank</span> [@wan2010towards], produce more typical summaries and represent the state of the art. <span style="font-variant:small-caps;">tgraph</span> is a graph-based sentence extraction model, where the graph is constructed from topic models and the optimization is performed by constrained ILP. <span style="font-variant:small-caps;">urank</span> adopts a unified ranking system for both single- and multi-document summarization. [|@[ ]{}l@|c@c@c@[ ]{}|]{} & <span style="font-variant:small-caps;">Rouge-1</span> & <span style="font-variant:small-caps;">Rouge-2</span> & <span style="font-variant:small-caps;">Rouge-l</span>\ <span style="font-variant:small-caps;">lead</span> & 43.6& 21.0 & 40.2\ <span style="font-variant:small-caps;">lreg</span> & 43.8 & 20.7 & 40.3\ <span style="font-variant:small-caps;">ilp</span> & 45.4 & 21.3 & 42.8\ <span style="font-variant:small-caps;">nn-abs</span> & 15.8 & 5.2 & 13.8\ <span style="font-variant:small-caps;">tgraph</span> & 48.1 & **24.3** & —\ <span style="font-variant:small-caps;">urank</span> & **48.5** & 21.5 & —\ <span style="font-variant:small-caps;">nn-se</span> & 47.4 & 23.0 & **43.5**\ <span style="font-variant:small-caps;">nn-we</span> & 27.0 & 7.9 & 22.8\ \ & <span style="font-variant:small-caps;">Rouge-1</span> & <span style="font-variant:small-caps;">Rouge-2</span> & <span style="font-variant:small-caps;">Rouge-l</span>\ <span style="font-variant:small-caps;">Lead</span> & 20.4 & 7.7 & 11.4\ <span style="font-variant:small-caps;">lreg</span> & 18.5 & 6.9 & 10.2\ <span style="font-variant:small-caps;">nn-abs</span> & 7.8 & 1.7 & 7.1\ <span style="font-variant:small-caps;">nn-se</span> & **21.2** & **8.3** & **12.0**\ <span style="font-variant:small-caps;">nn-we</span> & 15.7 & 6.4& 9.8\ #### Evaluation We evaluated the quality of the summaries automatically using <span style="font-variant:small-caps;">Rouge</span> [@lin:03]. We report unigram and bigram overlap () as a means of assessing informativeness and the longest common subsequence () as a means of assessing fluency. In addition, we evaluated the generated summaries by eliciting human judgments for 20 randomly sampled DUC 2002 test documents. Participants were presented with a news article and summaries generated by a list of systems. These include two neural network systems (sentence- and word-based extraction), the neural abstractive system described earlier, the lead baseline, the phrase-based ILP model[^9] of , and the human authored summary. Subjects were asked to rank the summaries from best to worst (with ties allowed) in order of informativeness (does the summary capture important information in the article?) and fluency (is the summary written in well-formed English?). We elicited human judgments using Amazon’s Mechanical Turk crowdsourcing platform. Participants (self-reported native English speakers) saw 2 random articles per session. We collected 5 responses per document. [|@[ ]{}p[16cm]{}@[ ]{}|]{} **sentence extraction**:\ \ \ \ at least three people poured gasoline on a car and lit it on fire at a *entity14* gas station explosive situation\ the passengers and the driver were not hurt during the incident but the car was completely ruined\ \ she said:’ he said he was pumping gas and some guys came up and asked for the car\ ’ they pulled out a gun and he took off running\ ’ they took the gas tank and started spraying\ ’ no one was injured during the fire , but the car ’s entire front end was torched , according to *entity52*\ \ surveillance video of the incident is being used in the investigation\ before the fire , which occurred at 12:15am on Saturday , the suspects tried to carjack the man hot case\ the *entity53* is investigating the incident at the *entity67* station as an arson\ **word extraction**:\ gang poured gasoline in the car, *entity5* Saturday morning. the driver argued with the suspects. his grandmother said the fire was lit by the suspects attempted to carjack her grandson.\ **entities**:\ *entity5*:California *entity13*:76-Station *entity14*: South LA *entity16*:Dodge Charger *entity33*:ABC *entity52*:NBC *entity53*:LACFD *entity67*:LA76\ Results {#sec:results} ======= Table \[duc\] (upper half) summarizes our results on the DUC 2002 test dataset using <span style="font-variant:small-caps;">Rouge</span>. represents our neural sentence extraction model, our word extraction model, and <span style="font-variant:small-caps;">nn-abs</span> the neural abstractive baseline. The table also includes results for the <span style="font-variant:small-caps;">lead</span> baseline, the logistic regression classifier (<span style="font-variant:small-caps;">lreg</span>), and three previously published systems (<span style="font-variant:small-caps;">ilp</span>, <span style="font-variant:small-caps;">tgraph</span>, and <span style="font-variant:small-caps;">urank</span>). Models 1$^{st}$ 2$^{nd}$ 3$^{rd}$ 4$^{th}$ 5$^{th}$ 6$^{th}$ MeanR ------------------------------------------------------ ---------- ---------- ---------- ---------- ---------- ---------- ------- <span style="font-variant:small-caps;">lead</span> 0.10 0.17 0.37 0.15 0.16 0.05 3.27 <span style="font-variant:small-caps;">ilp</span> 0.19 0.38 0.13 0.13 0.11 0.06 2.77 <span style="font-variant:small-caps;">nn-se</span> 0.22 0.28 0.21 0.14 0.12 0.03 2.74 <span style="font-variant:small-caps;">nn-we</span> 0.00 0.04 0.03 0.21 0.51 0.20 4.79 <span style="font-variant:small-caps;">nn-abs</span> 0.00 0.01 0.05 0.16 0.23 0.54 5.24 Human 0.27 0.23 0.29 0.17 0.03 0.01 2.51 : \[tab:humans\] Rankings (shown as proportions) and mean ranks given to systems by human participants (lower is better). The outperforms the <span style="font-variant:small-caps;">lead</span> and <span style="font-variant:small-caps;">lreg</span> baselines with a significant margin, while performing slightly better than the <span style="font-variant:small-caps;">ilp</span> model. This is an encouraging result since our model has only access to embedding features obtained from raw text. In comparison, <span style="font-variant:small-caps;">lreg</span> uses a set of manually selected features, while the <span style="font-variant:small-caps;">ilp</span> system takes advantage of syntactic information and extracts summaries subject to well-engineered linguistic constraints, which are not available to our models. Overall, our sentence extraction model achieves performance comparable to the state of the art without sophisticated constraint optimization (<span style="font-variant:small-caps;">ilp</span>, <span style="font-variant:small-caps;">tgraph</span>) or sentence ranking mechanisms (<span style="font-variant:small-caps;">urank</span>). We visualize the sentence weights of the model in the top half of Figure \[vis\]. As can be seen, the model is able to locate text portions which contribute most to the overall meaning of the document. <span style="font-variant:small-caps;">Rouge</span> scores for the word extraction model are less promising. This is somewhat expected given that <span style="font-variant:small-caps;">Rouge</span> is $n$-gram based and not very well suited to measuring summaries which contain a significant amount of paraphrasing and may deviate from the reference even though they express similar meaning. However, a meaningful comparison can be carried out between <span style="font-variant:small-caps;">nn-we</span> and <span style="font-variant:small-caps;">nn-abs</span> which are similar in spirit. We observe that <span style="font-variant:small-caps;">nn-we</span> consistently outperforms the purely abstractive model. As generates summaries by picking words from the original document, decoding is easier for this model compared to which deals with an open vocabulary. The extraction-based generation approach is more robust for proper nouns and rare words, which pose a serious problem to open vocabulary models. An example of the generated summaries for <span style="font-variant:small-caps;">nn-we</span> is shown at the lower half of Figure \[vis\]. Table \[duc\] (lower half) shows system results on the 500 DailyMail news articles (test set). In general, we observe similar trends to DUC 2002, with performing the best in terms of all <span style="font-variant:small-caps;">rouge</span> metrics. Note that scores here are generally lower compared to DUC 2002. This is due to the fact that the gold standard summaries (aka highlights) tend to be more laconic and as a result involve a substantial amount of paraphrasing. More experimental results on this dataset are provided in the appendix. The results of our human evaluation study are shown in Table \[tab:humans\]. Specifically, we show, proportionally, how often our participants ranked each system 1st, 2nd, and so on. Perhaps unsurprisingly, the human-written descriptions were considered best and ranked 1st 27% of the time, however closely followed by our <span style="font-variant:small-caps;">nn-se</span> model which was ranked 1st 22% of the time. The <span style="font-variant:small-caps;">ilp</span> system was mostly ranked in 2nd place (38% of the time). The rest of the systems occupied lower ranks. We further converted the ranks to ratings on a scale of 1 to 6 (assigning ratings 6$\dots$1 to rank placements 1$\dots$6). This allowed us to perform Analysis of Variance (ANOVA) which revealed a reliable effect of system type. Specifically, post-hoc Tukey tests showed that <span style="font-variant:small-caps;">nn-se</span> and <span style="font-variant:small-caps;">ilp</span> are significantly () better than <span style="font-variant:small-caps;">lead</span>, <span style="font-variant:small-caps;">nn-we</span>, and <span style="font-variant:small-caps;">nn-abs</span> but do not differ significantly from each other or the human goldstandard. Conclusions {#sec:conclusions} =========== In this work we presented a data-driven summarization framework based on an encoder-extractor architecture. We developed two classes of models based on sentence and word extraction. Our models can be trained on large scale datasets and learn informativeness features based on continuous representations without recourse to linguistic annotations. Two important ideas behind our work are the creation of hierarchical neural structures that reflect the nature of the summarization task and generation by extraction. The later effectively enables us to sidestep the difficulties of generating under a large vocabulary, essentially covering the entire dataset, with many low-frequency words and named entities. Directions for future work are many and varied. One way to improve the word-based model would be to take structural information into account during generation, e.g., by combining it with a tree-based algorithm [@cohn2009sentence]. It would also be interesting to apply the neural models presented here in a phrase-based setting similar to . A third direction would be to adopt an information theoretic perspective and devise a purely unsupervised approach that selects summary sentences and words so as to minimize information loss, a task possibly achievable with the dataset created in this work. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank three anonymous reviewers and members of the ILCC at the School of Informatics for their valuable feedback. The support of the European Research Council under award number 681760 “Translating Multiple Modalities into Text” is gratefully acknowledged. Appendix ======== In addition to the DUC 2002 and 500 DailyMail samples, we additionally report results on the entire DailyMail test set (Table \[ducap\]). Since there is no established evaluation standard for this task, we experimented with three different ROUGE limits: 75 bytes, 275 bytes and full length. [|@[ ]{}l@|c@c@c@[ ]{}|]{} & <span style="font-variant:small-caps;">Rouge-1</span> & <span style="font-variant:small-caps;">Rouge-2</span> & <span style="font-variant:small-caps;">Rouge-l</span>\ <span style="font-variant:small-caps;">Lead</span> & 21.9 & 7.2 & 11.6\ <span style="font-variant:small-caps;">nn-se</span> & 22.7 & 8.5 & 12.5\ <span style="font-variant:small-caps;">nn-we</span> & 16.0 & 6.4 & 10.2\ \ & <span style="font-variant:small-caps;">Rouge-1</span> & <span style="font-variant:small-caps;">Rouge-2</span> & <span style="font-variant:small-caps;">Rouge-l</span>\ <span style="font-variant:small-caps;">Lead</span> & 40.5 & 14.9 & 32.6\ <span style="font-variant:small-caps;">nn-se</span> & 42.2 & 17.3 & 34.8\ <span style="font-variant:small-caps;">nn-we</span> & 33.9 & 10.2 & 23.5\ \ & <span style="font-variant:small-caps;">Rouge-1</span> & <span style="font-variant:small-caps;">Rouge-2</span> & <span style="font-variant:small-caps;">Rouge-l</span>\ <span style="font-variant:small-caps;">Lead</span> & 53.5 & 21.7 & 48.5\ <span style="font-variant:small-caps;">nn-se</span> & 56.0 & 24.9 & 50.2\ <span style="font-variant:small-caps;">nn-we</span> & - & - & -\ [^1]: Resources are available for download at <http://homepages.inf.ed.ac.uk/s1537177/resources.html> [^2]: The vocabulary can also be extended to include a small set of commonly-used (high-frequency) words. [^3]: The script for constructing our datasets is modified from the one released in . [^4]: We used the Python Gensim library and the *GoogleNews* vectors. [^5]: A simpler model would use hard attention to select a sentence first and then a few words from it as a summary, but this would render the system non-differentiable for training. Although hard attention can be trained with the REINFORCE algorithm [@williams1992simple], it requires sampling of discrete actions and could lead to high variance. [^6]: According to the DUC2002 guidelines <http://www-nlpir.nist.gov/projects/duc/guidelines/2002.html>, the generated summary should be within 100 words. [^7]: The CNN-LSTM architecture is publicly available at <https://github.com/yoonkim/lstm-char-cnn>. [^8]: We used the *word2vec* [@mikolov2013distributed] skip-gram model with context window size 6, negative sampling size 10 and hierarchical softmax 1. The model was trained on the *Google 1-billion* word benchmark [@Chelba:ea:2014]. [^9]: We are grateful to Kristian Woodsend for giving us access to the output of his system. Unfortunately, we do not have access to the output of <span style="font-variant:small-caps;">tgraph</span> or <span style="font-variant:small-caps;">urank</span> for inclusion in the human evaluation.
{ "pile_set_name": "ArXiv" }
--- abstract: | We propose new construction of the polynomial integrals of motion related to the addition theorems. As an example we reconstruct Drach systems and get some new two-dimensional superintegrable Stäckel systems with third, fifth and seventh order integrals of motion.\ \ PACS numbers: 02.30.Jr, 02.30.Ik, 03.65.Fd\ Mathematics Subject Classification: 70H06, 70H20, 35Q72 author: - | A V Tsiganov\ *St.Petersburg State University, St.Petersburg, Russia\ *e–mail: [email protected]** title: Addition theorems and the Drach superintegrable systems --- Introduction ============ The Liouville classical theorem on completely integrable Hamiltonian systems implies that almost all points of the manifold $M$ are covered by a system of open toroidal domains with the action-angle coordinates $I=(I_1,\ldots,I_k)$ and $\omega=(\omega_1,\ldots,\omega_n)$: \[aa-br\] {I\_j,I\_k}={\_i,\_k}=0,{I\_j,\_k}=\_[ij]{}. The independent integrals of motion $H_1,\ldots,H_n$ are functions of the independent action variables $I_1,\ldots,I_n$ and the corresponding Jacobian does not equal to zero \[h-mat\] J0, J\_[ij]{}= . Let us introduce $n$ functions \[phi\_k\] \_j=\_[k]{} (J\^[-1]{})\_[kj]{}\_k,such that \[tr-alg\] {H\_i,\_j}=\_[k=1]{}\^[n]{} J\_[ik]{}(J\^[-1]{})\_[kj]{} =\_[ij]{}. If Hamiltonian $H=H_1$ then the ($n-1$) functions $\phi_2,\ldots,\phi_n$ are integrals of motion $$\dfrac{d\phi_j}{dt}=\{H_1,\phi_j\}=0,\qquad j=2,\ldots,n,$$ which are functionally independent on $n$ functions $H_1(I),\ldots,H_n(I)$. So, in classical mechanics any completely integrable system is superintegrable system in a neighborhood of any regular point of $M$ [@ts07f]. It means that the Hamiltonian $H=H_1$ has $2(n-1)$ integrals of motion $H_2,\ldots,H_n$ and $\phi_2,\ldots,\phi_n$ on any open toroidal domain. If the action-angle variables are global variables on the whole phase space $M$ and, therefore, we have superintegrable systems on $M$. For instance, the global action-angle variables for the open and periodic Toda lattices are discussed in [@hk08]. However, in generic case the angle variables $\omega_k$ are multi-valued functions on the whole phase space $M$. If we have $k$ additional single-valued algebraic integrals of motion $K$ the trajectories are closed (more generally, they are constrained to an $n-k$ dimensional manifold in phase space). Any additional integral is a function on the action-angle variables. Since we have to understand how to get single-valued additional integrals of motion from the multi-valued action-angle variables. In this paper we discuss a possibility to get polynomial integrals of motion from the multi-valued angle variables by using simplest addition theorem. The Stäckel systems. ==================== The system associated with the name of Stäckel [@st95; @ts99] is a holonomic system on the phase space $M=\mathbb R^{2n}$, with the canonical variables $q=(q_1,\ldots,q_n)$ and $p=(p_1,\ldots,p_n)$: =\_[j=1]{}\^n dp\_jdq\_j,{p\_j,q\_k}=\_[jk]{}.\[stw\] The nondegenerate $n\times n$ Stäckel matrix $S$, whose $j$ column depends on the coordinate $q_j$ only, defines $n$ functionally independent integrals of motion H\_k=\_[j=1]{}\^n ( S\^[-1]{})\_[jk]{}(p\_j\^2+U\_j(q\_j)). \[fint\] From this definition one immediately gets the separated relations p\_j\^2= \_[k=1]{}\^n H\_k S\_[kj]{}-U\_j(q\_j)\[stc\] and the angle variables $$\omega_i=\sum_{j=1}\int\,\dfrac{S_{ij}\,\mathrm dq_j}{p_j}= \sum_{j=1}\int\,\frac{S_{ij}\,\mathrm dq_j}{\sqrt{\sum_{k=1}^n H_k S_{kj}-U_j(q_j) } }\,.$$ It allows reducing solution of the equations of motion to a problem in algebraic geometry [@ts99]. Namely, let us suppose that there are functions $\mu_j$ and $\lambda_j$ on the canonical separated variables \[fr-2\] \_j=u\_j(q\_j)p\_j,\_j=v\_j(q\_j),{q\_i,p\_j}=\_[ij]{}, which allows us to rewrite separated equations (\[stc\]) as equations defining the hyperelliptic curves \[sthc\] C\_j:\_j\^2=P\_j(\_j)u\_j\^2(\_j)(\_[k=1]{}\^n H\_k S\_[kj]{}(\_j)-U\_j(\_j)), where $P_j(\lambda_j)$ are polynomials on $\lambda_j$. In this case the action variables $I_k=H_k$ (\[fint\]) have the canonical Poisson brackets (\[aa-br\]) with the angle variables \[w-st\] \_i=\_[j=1]{}\^n\_[A\_j]{} d\_j=\_[j=1]{}\^n \_[ij]{}(p\_j,q\_j), which are the sums of integrals $\vartheta_{ij}$ of the first kind Abelian differentials on the hyperelliptic curves $\mathcal C_j$ (\[sthc\]) [@ts99; @ts07f], i.e. they are sums of the multi-valued functions on the whole phase space. Addition theorems and algebraically superintegrable systems ----------------------------------------------------------- In generic case the action variables (\[w-st\]) are the sum of the multi-valued functions $\vartheta_{ij}$. However, if we are able to apply some addition theorem to the calculation of $\omega_i$ (\[w-st\]) \[add-th\] \_i=\_[j=1]{}\^n \_[ij]{}(p\_j,q\_j)=\_i (K\_i)+const, where $\Theta_i$ is a multi-valued function on the algebraic argument $K_i(p,q)$ then one will get algebraic integrals of motion $K_i(p,q)$ because $$\{H_1,\omega_i\}=\{H_1,\Theta_i \bigl(K_i\bigr)\}=\Theta_i'\cdot\{H_1,K_i\}=0.$$ So, the addition theorems (\[add-th\]) could help us to classify algebraically superintegrable systems and vice versa. Plane curves with the genus $g\geq 1$ are related to elliptic and Abelian integrals. Addition theorems of these functions are the content of Abel’s theorem [@we08]. The main result of this paper is that almost all the known examples of algebraically superintegrable systems relate with the one of the simplest addition theorem \[add-ln\] e\^x\^y=\^[x+y]{},(x\_1)+(x\_2)=(x\_1x\_2) associated with the zero-genus hyperelliptic curves $\mathcal C_j$ \[fr-1\] C\_j:\_j\^2=P\_j(\_j)=f\_j\_j\^2+g\_j\_j+h\_j,j=1,…,n, where $f_j,g_j,h_j$ are linear functions on $n$ integrals of motion $H_1,\ldots,H_n$. In fact, if $S_{ij}(\lambda)=1 $ then after substitution (\[fr-1\]) into (\[w-st\]) one gets the sum of the rational functions $$\vartheta_j=\int\dfrac{1 }{\sqrt{g_j\lambda_j+h_j}}\,\mathrm d \lambda_j=\frac{\mu_j}{g_j}$$ or logarithmic functions $$\vartheta_j=\int \dfrac{1}{\sqrt{f_j\lambda_j^2+g_j\lambda_j+h_j\,}}\,\mathrm d\lambda= f_j^{-1/2}\,\ln\left(\mu_j+\dfrac{2f_j\lambda_j+g_j}{2\sqrt{f_j}}\right)\,.$$ For the hyperelliptic curves of higher genus one gets elliptic functions, which have more complicated addition law [@we08]. In order to use addition low (\[add-ln\]) we have to make the following steps: - [We have to apply canonical transformation of the time that reduce $n$-th row of the Stäckel matrix to the canonical Brill-Noether form [@ts99a; @ts01] $$S_{nj}=1,\qquad j=1,\ldots n,$$ such that \[wk-n\] \_n=\_[j=1]{}\^n \^[v\_j(q\_j)]{} d , where $N$ is normalization, which is restored from $\{\omega_n,H_n\}=1.$ ]{} - [Then we have to use addition law (\[add-ln\]) for construction of the polynomial in momenta integrals of motion]{} - [We have to make inverse transformation of the time, which preserves the polynomial form of integrals such us it depends on $q$ variables only [@ts99a; @ts01]]{}. Let us consider construction of polynomial in the momenta integrals of motion at $n=2$. ### Case $f_1=f_2=0$ If $f_{1,2}=0$ we have \[cal-ex\] \_2=14\_[j=1]{}\^2 \^[v\_j(q\_j)]{} d = +=, where $$K=g_2\,p_1{u_1}+g_1\,p_2{u_2}$$ is the polynomial in the momenta integral of motion of the first or the third order. Remind that $g_{1,2}$ are linear functions on the Stäckel integrals $H_{1,2}$, which are the second order polynomials on $p_{1,2}$. Let us consider the two-dimensional Stäckel system defined by two Riemann surfaces $$\mathcal C_{1,2}:\qquad \mu^2=P_{1,2}(\lambda)= (H_1\pm H_2)\lambda+\alpha_{1,2},$$ and substitutions (\[fr-2\]) $$\mu_j=q_jp_j,\qquad\lambda_j=q_j^2\,,$$ which give rise to the following separated equations $$p_{1,2}^2 = H_1\pm H_2+\frac{\alpha_{1,2}}{q_{1,2}^2}\,.$$ Integrals of motion $H_1$ and $H_2$ are solutions of these separated equations \[cal-int\] H\_[1,2]{}=2-, which coincide with integrals of motion for the two-particle Calogero system after an obvious point transformation \[tr-c\] x=,p\_x=p\_1-p\_2,y=,p\_y=p\_1+p\_2. The angle variables (\[cal-ex\]) read as \_[1,2]{}&=& \^[q\_1\^2]{} d \^[q\_2\^2]{} d\ \ &=&.The corresponding cubic integral of motion $K$ (\[cal-ex\]) is equal to $$K=2(H_1+H_2)(H_1-H_2)\omega_2=-p_1q_1(H_1-H_2)+p_2q_2(H_1+H_2).$$ The bracket $$\{K,H_2\}=2(H_2^2-H_1^2)\,$$ is easily restored from the canonical brackets (\[aa-br\]). ### Case $f_1=k_1^2 f$ and $f_2=k_2^2f$ with integer $k_{1,2}$ In this case, we have \_2&=&\_[j=1]{}\^2 \^[v\_j(q\_j)]{} d = \_[j=1]{}\^2 ( +p\_j[u\_j]{}),\ &=& ,where $$P^{\,\prime}_{j}=\left.\frac{d P_{j}(\lambda)}{d\lambda}\right|_{\lambda=v_j(q_j)}=2f_jv_j(q_j)+g_j\,.$$ So function \_2&=& (2k\_1k\_2\_2)=( +p\_1[u\_1]{})\^[k\_2]{} ( +p\_2[u\_2]{})\^[k\_1]{}=\ &=&(K\_+K\_m)\[k5-gen\] is the generating function of polynomial integrals of motion $K_m$ and $K_\ell$ of the $m$-th and $\ell$-th order, respectively. The values of $m$ and $\ell$ depend on values of $k_{1,2}$. As above, the algebra of integrals $H_{1,2}$ and $K_{m}$ may be restored from the canonical brackets (\[aa-br\]). Let us consider the two-dimensional Stäckel system defined by the two Riemann surfaces $$\mathcal C_{1,2}:\qquad \mu_{1,2}^2=P_{1,2}(\lambda_{1,2})=k_{1,2}^2\lambda^2_{1,2}+\beta_{1,2}\lambda_{1,2}+(H_1\pm H_2)$$ and substitutions (\[fr-2\]) $$\mu_j=\,p_j,\qquad\lambda_j=q_j\,,$$ which give rise to the following separated equations $$p_{1,2}^2 = k_{1,2}^2 q_{1,2}^2+ \beta_{1,2}q_{1,2}+(H_1\pm H_2)\,.$$ Integrals of motion $H_1$ and $H_2$ are solutions of these separated equations \[os-int\] H\_[1,2]{}=2--, which coincide with integrals of motion for the harmonic oscillator. The same Stäckel system coincides with the Kepler problem after well-known canonical transformation of the time, which changes the row of the Stäckel matrix [@ts99a]. The angle variable (\[cal-ex\]) reads as \_[2]{}&=&-( \^[q\_j]{} d - \^[q\_2]{} d )=\ &=&.Using suitable branches of $\sqrt{k_{1,}^2}\,$ one gets function (\[k5-gen\]) $$\Phi=\exp(2k_1k_2\omega_2)=\left(p_1-\dfrac{2k_1^2 q_1+\beta_1}{2k_1}\right)^{k_2}\left(p_2+\dfrac{2k_2^2q_2+\beta_2}{2k_2}\right)^{k_1},$$ which generates polynomial integrals of motion $K_m$ and $K_\ell$. As an example, if $k_1=1$ and $k_2=3$ one gets third and fourth order integrals of motion $K_m$ and $K_\ell$ respectively. Of course, oscillator is one of the well-studied superintegrable systems, which was earmarked to illustrate generic construction only. Additional theorem (\[add-ln\]) allows us to get a huge family of the $n$-dimensional superintegrable systems, which have to be classified and studied. Classification -------------- In order to classify superintegrable systems associated with the addition theorem (\[add-ln\]) we have to start with a pair of the Riemann surfaces \[dr-eq\] C\_[j]{}:\^2=P\_j()= f\^2+g\_j+h\_j,j=1,2, where $$f=\alpha H_1+\beta H_2+\gamma,\quad g_j=\alpha^g_jH_1+\beta^g_j H_2+\gamma^g_j,\quad h_j=\alpha^h_jH_1+\beta^h_j H_2+\gamma^h_j,$$ and $\alpha$, $\beta$ and $\gamma$ are real or complex numbers. In order to use addition theorem (\[add-ln\]) we fix last row of the Stäckel matrix. Namely, substituting \[m-rest\] S\_[2j]{}()=\_j into the (\[stc\]-\[w-st\]) one gets \[th-eq\] \_[2j]{}== = \_jf\^[-1/2]{}(\_j+) , so the angle variable $$\omega_2=\dfrac{1}{\sqrt{f}}\,\ln \left[\left(p_1{u_1}+ \dfrac{P_1'}{2\sqrt{f}}\right)^{\kappa_1} \left(p_2{u_2}+\dfrac{P_2'}{2\sqrt{f}}\right)^{\kappa_2}\right], \qquad P^{\,\prime}_{j}=\left.\frac{d P_{j}(\lambda)}{d\lambda}\right|_{\lambda=v_j(q_j)},$$ is the multi-valued function on the desired algebraic argument $$K=\left(p_1{u_1}+ \dfrac{P_1'}{2\sqrt{f}}\right)^{\kappa_1} \left(p_2{u_2}+\dfrac{P_2'}{2\sqrt{f}}\right)^{\kappa_2}.$$ If $\kappa_{1,2}$ are positive integer, then \[k-kappa\] K=()\^[\_1+\_2]{}(K\_+K\_m) is the generating function of polynomial integrals of motion $K_m$ and $K_\ell$ of the $m$-th and $m\pm 1$-th order in the momenta, respectively. As an example we have \[k-m\] [ll]{} K\_m=2(p\_1[u\_1]{}P’\_2+p\_2[u\_2]{}P’\_1),&\_1=1, \_2=1,\ K\_m=2P’\_2(2p\_2u\_2P’\_1+p\_1u\_1P’\_2)+8p\_1u\_1p\_2\^2u\_2\^2f, &\_1=1,\_2=2,\ K\_m=2[P’\_2]{}\^2(3p\_2u\_2P’\_1+p\_1u\_1P’\_2)+8p\_2\^2u\_2\^2(p\_2u\_2P’\_1+3p\_1u\_1P’\_2)f, &\_1=1,\_2=3, where $m=1,3$, $m=3,5$ ¨ $m=3,7$, because $P'_{1,2}$ and $f$ are linear functions on $H_{1,2}$, which are the second order polynomials on momenta. The corresponding expressions for the $K_\ell$ look like \[k-ell\] [ll]{} K\_=P’\_1P’\_2+4p\_1p\_2[u\_1u\_2]{}f,&\_1=1, \_2=1,\ K\_=P’\_1[P’\_2]{}\^2+4fp\_2u\_2(p\_2u\_2P’\_1+2p\_1u\_1P’\_2), &\_1=1,\_2=2,\ K\_=P’\_1[P’\_2]{}\^3+12P’\_2p\_2u\_2(P’\_2p\_1u\_1+P’\_1p\_2u\_2)f+16p\_1u\_1p\_2\^3u\_2\^3f\^2. &\_1=1,\_2=3. The imposed condition (\[m-rest\]) leads to some restrictions on the functions $v_j(q_j)$ and $u_j(q_j)$. In fact, substituting canonical variables (\[fr-2\]) into the equations (\[dr-eq\]) we obtain the following expression for the Stäckel matrix \[s-fin\] S=( [cc]{} &\ \ & ),S0. So, for a given $\kappa_{1,2}$ expressions for $\vartheta_{2j}$ (\[th-eq\]) yield two differential equations on functions $u,v$ and parameters $\beta$: \[uv-eq\] S\_[2j]{}(q\_j)=\_j u\_jv’\_j=v\_j+\^g\_jv\_j+\^h\_j,j=1,2. For the Stäckel systems with rational or trigonometric metrics, we have to solve these equations in the space of the truncated Laurent or Fourier polynomials, respectively. If $\kappa_{j}\neq 0$ equations (\[uv-eq\]) have the following three monomial solutions \[uv-sol\] [llll]{} I&=0,\_j\^h=0,& u\_j=q\_j,&v\_j=q\_j\^,\ \ &\_j\^g=0,\_j\^h=0,&u\_j=1,& v\_j=-\_j( q\_j)\^[-1]{},\ \ &=0,\_j\^g=0,&u\_j=1,&v\_j=\_j\^[-1]{}\_j\^h q\_j, up to canonical transformations. The fourth solution $\mathrm{(\,IV)}$ is the combination of the first and third solutions for the different $j$’s. In order to prove this fact we can substitute $u=aq^m$ and $v=bq^k$ into the (\[uv-eq\]) and divide resulting equation on $q^{m+k-1}$ $$abk\kappa = b^2\beta q^{k+1-m}+q^{1-m}b\beta^g+q^{1-k-m}\beta^h.$$ Finally, we differentiate it by $q$ and multiply on $q^{m}$ $$0=-(m-k-1)b^2\beta q^{k}-(m-1)b\beta^g -(m+k-1)\beta^h\,q^{-k}.$$ Such as $b\neq 0$ and $k\neq 0$ one gets three solutions (\[uv-sol\]) only. Then we suppose that after some point transformation \[z-dr\] [ll]{} x=z\_1(q), & y=z\_2(q),\ p\_x=w\_[11]{}(q)p\_1+w\_[12]{}(q)p\_2,&p\_y=w\_[21]{}(q)p\_1+w\_[22]{}(q)p\_2, where $\mathrm w_{ij}\neq 0$, kinetic part of the Hamilton function $H_1=T+V$ has a special form $$T=\sum\left(S^{-1}\right)_{1j}p_j^2=\mathrm{g}_{11}(x,y)p_x^2 +\mathrm{g}_{12}(x,y)p_xp_y+\mathrm{g}_{22}(x,y)p_y^2,$$ where $\mathrm{g}$ is a metric on a configurational manifold. For instance, if we suppose that $$T=\sum\left(S^{-1}\right)_{1j}p_j^2=p_xp_y,$$ then one gets the following algebraic equations \[alg-dr\] w\_[11]{}w\_[21]{}=(S\^[-1]{})\_[11]{},w\_[12]{}w\_[21]{}+ w\_[11]{}w\_[22]{}=0,w\_[12]{}w\_[22]{}=(S\^[-1]{})\_[12]{} and the partial differential equations \[pde-dr\] {x,p\_x}={y,p\_y}=1,{p\_x,y}={p\_y,x}={p\_x,p\_y}=0. on parameters $\alpha$ and functions $z_{1,2}(q_1,q_2)$, $\mathrm w_{kj}(q_1,q_2)$. The remaining free parameters $\gamma,\gamma_j^h,\gamma_j^g$ determine the corresponding potential part of the Hamiltonian $V(x,y)$. In fact, since integrals $H_{1,2}$ is defined up to the trivial shifts $H_k\to H_k+c_k$, our potential $V(x,y)$ depends on three arbitrary parameters only. Summing up, in order to get all the superintegrable systems on a complex Euclidean space $E_{2}(\mathbb{C})$ associated with the addition theorem (\[add-ln\]) we have to solve equations (\[uv-eq\],\[alg-dr\],\[pde-dr\]) with respect to functions $u_{j}(q_j)$, $v_{j}(q_j)$, $z_{1,2}(q_1,q_2)$, $\mathrm w_{kj}(q_1,q_2)$ and parameters $\alpha$ and $\beta$. Let us consider second solution from the list (\[uv-sol\]) at $\kappa_1=1$ and $\kappa_2=2$. In this case equations (\[alg-dr\],\[pde-dr\]) have the following partial solution $$S=\left(\begin{array}{cc}1\quad& 0\\ \frac{1}{q_1^2}\quad& \frac{4}{q_2^2}\end{array}\right),\qquad \begin{array}{ll} x={\sqrt{q_1}q_2},\quad & p_x= \frac{\sqrt{q_1}}{q_2}\,p_1+\frac{1}{2\sqrt{q_1}}\,p_2\,,\\ y=\frac{\sqrt{q_1}}{q_2},\quad& p_y= \sqrt{q_1}q_2\,p_1-\frac{q_2^2}{2\sqrt{q_1}}\,p_2\,.\end{array}$$ Adding potential terms one gets two Riemann surfaces $$\mathcal C_{1}:\quad \mu^2={H_2}{\lambda^2}-{H_1}{\lambda}-{\gamma_1},\quad\mathrm{and}\quad \mathcal C_{2}:\quad\mu^2={H_2}{\lambda^2}-{2\gamma_2}{\lambda}+{4\gamma_3}$$ where $\mu_j=\,p_j$, $\lambda_j=q_j^{-1}$. Solutions of the corresponding separated variables are the second order Stäkel integrals of motion, which in physical variables look like \[h1-5\] H\_1=p\_xp\_y+\_1 xy ++,H\_2=4-\_2-\_3. It is new integrable systems, which is missed in the known lists of superintegrable systems [@cd06; @mw07; @ran97; @ran01]. In this case integrals of motion $K_5$ (\[k-m\]) and $K_6$ (\[k-ell\]) are the fifth and sixth order polynomials in the momenta, respectively. Of course, we can try to get quartic, cubic and quadratic integrals of motion $K_{4}$, $K_3$ and $K_{2}$ from the recurrence relations \[rec-H\] K\_5={K\_4,H\_2},K\_4={K\_3,H\_2},K\_3={K\_2,H\_2} and the equation $\{H_1,K_{j}\}=0$, $j=4,3,2$. Solving first recurrence equation one gets quartic integral of motion $$K_4=-4(xp_x-yp_y)^2(p_x^2+\gamma_1 y^2)-\frac{4\gamma_2^2}{xy} +\frac{8p_x(xp_x-yp_y)\gamma_2}{\sqrt{xy}} +\frac{16x(p_x^2+\gamma_1y^2)\gamma_3}{y}.$$ The other recurrence relations (\[rec-H\]) have not polynomial solutions. Of course, we can try modify recurrence relations $$\{K_3,H_2\}=K_4+F_4(H_1,H_2),\qquad \{K_2,H_2\}=K_3+F_3(H_1,H_2)\,$$ in order to get cubic and quadratic integrals of motion. However, using ansatz $$K_2=h_1(x,y)p_x^2+h_2(x,y)p_xp_y+h_3(x,y)p_y^2+h_4(x,y)$$ we can directly prove that there is not the additional quadratic integral of motion, which commute with $H_1$ (\[h1-5\]). The Drach systems ================= In 1935 Jules Drach published two articles on the Hamiltonian systems with the third order integrals of motion on a complex Euclidean space $E_{2}(\mathbb{C})$ with the following Hamilton function [@dr35] \[ham-dr\] H\_1=p\_xp\_y+U(x,y). Up to canonical transformations $x\to a x$ and $y\to b y$ the corresponding potentials look like [@ran97; @ts00]: (a)U&=&+x\^[r\_1]{}y\^[r\_2]{}+x\^[r\_2]{}y\^[r\_1]{},r\_j\^2+3r\_j+3=0,\ (b)U&=&++,\ (c)U&=&xy++,\ (d)U&=&+ +,\ (e)U&=&+ +,\ (f)U&=&xy +y+,\ (g)U&=&+(y-x)+3,\ (h)U&=&(y+3)\^[-2/3]{},\ (k)U&=&y\^[-1/2]{} +x y\^[-1/2]{} +x,\ (l)U&=&(y-3)+x\^[-1/2]{} +x\^[-1/2]{}(y-x). Such as Drach made some assumptions on the form of the third order integrals of motion $K_3$ (\[k3-dr\]) in the calculation it is not immediately clear whether the obtained list is complete. Non-separable systems. ---------------------- Let us discuss the Drach systems, which can not be reduced to the Stäckel systems by any point transformation of variables. The first system (a) is non-Stäckel system related to the three-particle periodic Toda lattice in the center-of-mass frame [@ts00] and there are global action-angle variables [@hk08]. The (h) system is reduced to the Stäckel system by non-point canonical transformation and, therefore, existence of the third order integral of motion is related with this non-point transformation [@ts00]. Later this system has been rediscovered by Holt [@ho82]. For the (k) case in the Drach papers [@dr35] we can find Hamiltonian \[hkd\] H\_1\^[(k)]{}=p\_xp\_y+y\^[-3/2]{} +x y\^[-3/2]{} +x and the following cubic integral of motion \[kkd\] K\_3\^[(k)]{}=6w(x,y)(p\_y -p\_x) -P(p\_x,p\_y,x,y),where $$P=3p_x^2p_y\,,\qquad w=-y.$$ It is easy to prove that $\{H_1^{(k)},K_3^{(k)}\}\neq 0$ and, therefore, we have to suggest the possibility of a small mistake in the Drach papers [@dr35]. Following to [@ts00], if we solve equation $\{p_xp_y+U(x,y),K_3^{(k)}\}=0$ with respect to $U(x,y)$, then one gets our case (k) $$H_1=p_xp_y+\alpha y^{-1/2} +\beta x y^{-1/2} +\gamma x.$$ On the other hand, we have proven directly that the Hamiltonian $H_1^{(k)}$ (\[hkd\]) has one second order integral of motion \[h2k\] H\_2\^[(k)]{}=p\_x\^2-4y\^[1/2]{}+2y and has not cubic integral of motion. Moreover, it is easy to see that quadratic integrals of motion $H_{1,2}^{(k)}$ (\[hkd\],\[h2k\]) can not be reduced to the Stäckel integrals by any point transformation of variables. Below we do not consider (a) and (h) systems and consider (k) case in our notation only. Classification. --------------- For the Drach systems $\kappa_1=\kappa_2=\pm1,1/2$ and \_2&=& (2\_2)=( +p\_1[u\_1]{})( +p\_2u\_2)=\ &=& (K\_+K\_m)\[phi-ln\] may be considered as the generating function of the polynomial integrals of motion (\[k-m\]-\[k-ell\]) K\_m&=&2(p\_1[u\_1]{}P’\_2+p\_2[u\_2]{} P’\_1),m=1,3, \[k3-dr\]\ K\_&=&P’\_1P’\_2+4p\_1p\_2[u\_1u\_2]{}f,=2,4, of the $m$-th and $\ell$-th order, respectively. It’s clear that $m=1,3$ and $\ell=2,4$, because $P'_{1,2}$ and $f$ are linear functions on $H_{1,2}$, which are second order polynomials on momenta. We have to underline, that we use different $\kappa_{1}=\kappa_2{2}=\pm1,1/2$ for the agreement of $K_m$ (\[k3-dr\]) with the initial Drach integrals of motion [@dr35] only. One gets third order polynomial integral of motion $K_3$ (\[k3-dr\]) if and only if $P'_1(\lambda)$ or $P'_2(\lambda)$ depends on $H_1$ or $H_2$. It leads to the additional restrictions on $\alpha$’s and $\beta$’s \[cub-dr\] \_[k=1]{}\^20 j=1 j=2. In order to get all the superintegrable systems on a complex Euclidean space $E_{2}(\mathbb{C})$ associated with addition theorem (\[add-ln\]) we have to solve equations (\[uv-eq\],\[alg-dr\],\[pde-dr\]) and (\[cub-dr\]) at $\kappa_{1,2}=1$ with respect to the functions $u_{j}(q_j)$, $v_{j}(q_j)$, $z_{1,2}(q_1,q_2)$, $\mathrm w_{kj}(q_1,q_2)$ and parameters $\alpha$ and $\beta$. The Drach list of the Stäckel systems with the cubic integral of motion (\[k3-dr\]) associated with the addition theorem (\[add-ln\]) is complete up to canonical transformations of the extended phase space. The results of corresponding calculations may be joined into the table: 0.5truecm -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $\mathcal C_{1,2}$ (\[dr-eq\]) subs. (\[fr-2\]) $z_{1,2}$ (\[z-dr\]) S --- ----------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------- -------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------- b $\begin{array}{c}\\ $\begin{array}{c} \\ \\ \mu_j=p_jq_j\end{array}$ $z_{1,2}=\frac{(q_1\pm q_2)^2}4$ $\left(\begin{smallmatrix} q_1^2&q_2^2\\1&1\end{smallmatrix}\right)$ \mu^2=H_1\lambda^2+(H_2+2\alpha)\lambda-\beta+2\gamma\\ \\ \mu^2=H_1\lambda^2+(H_2-2\alpha)\lambda-\beta-2\gamma\\ \end{array}$ c $\begin{array}{c}\\ $ \begin{array}{c}\lambda_j=q_j^2\\ \\ \kappa_j=\frac{1}{2}\end{array}$ $z_{1,2}=\frac{q_1\pm q_2}{2}$ $\left(\begin{smallmatrix} \frac12&-\frac12\\1&1\end{smallmatrix}\right)$ \mu^2=\dfrac{\alpha}4\lambda^2+\left(H_2+\dfrac{H_1}2\right)\lambda+\gamma\\ \\ \mu^2=\dfrac{\alpha}4\lambda^2+\left(H_2-\dfrac{H_1}2\right)\lambda-\beta\\ \end{array}$ d $\begin{array}{c}\\ $\begin{array}{c} \\ \\ \mu_j=p_j \end{array}$ $\begin{array}{l}z_1=\frac{q_1^2+q_2^2}{2q_1q_2}\\ \\ z_2= q_1q_2\end{array}$ $\left(\begin{smallmatrix} 1&1\\ \frac1{q_1^2}&\frac{1}{q_2^2}\end{smallmatrix}\right)$ \mu^2=H_2\lambda^2-\sqrt{8}(\alpha+\beta)\lambda+H_1-2\gamma\\ \\ \mu^2=H_2\lambda^2-\sqrt{8}(\alpha-\beta)\lambda+H_1+2\gamma\\ \end{array}$ f $\begin{array}{c}\\ $\begin{array}{c} \lambda_j=q_j^{-1}\\ \\ \kappa_j=-1\end{array}$ $\begin{array}{l}z_1=\frac{q_1-q_2^2}{2\sqrt{q_1q_2}}\\ \\ z_2= \sqrt{q_1q_2}\end{array}$ $\left(\begin{smallmatrix} \frac1{q_1}&\frac{-1}{q_2}\\ \frac1{q_1^2}&\frac{1}{q_2^2}\end{smallmatrix}\right)$ \mu^2=H_2\lambda^2-\left(\dfrac{\gamma}2-H_1\right)\lambda-\dfrac{\alpha}4-\dfrac{\beta}2\\ \\ \mu^2=H_2\lambda^2-\left(\dfrac{\gamma}2+H_1\right)\lambda-\dfrac{\alpha}4+\dfrac{\beta}2\\ \end{array}$ e $\begin{array}{c}\\ $\begin{array}{c} \\ \\ \mu_j=p_j\end{array}$ $z_{1,2}=\frac{(q_1\pm q_2)^2}4$ $\left(\begin{smallmatrix} q_1^2&q_2^2\\1&1\end{smallmatrix}\right)$ \mu^2=H_1\lambda^2+2(\beta+\gamma)\lambda+H_2+2\alpha\\ \\ \mu^2=H_1\lambda^2-2(\beta-\gamma)\lambda+H_2-2\alpha\\ \end{array}$ k $\begin{array}{c}\\ $\begin{array}{c} \lambda_j=q_j \\ \\ \kappa_j=1\end{array}$ $\begin{array}{l} z_1=\frac{q_1-q_2}{2}\\ \\ z_2=\frac{(q_1+q_2)^2}{4}\end{array}$ $\left(\begin{smallmatrix} q_1&-q_2\\1&1\end{smallmatrix}\right)$ \mu^2=\dfrac{\gamma}{2}\lambda^2+(\beta+H_1)\lambda+H_2+\alpha\\ \\ \mu^2=\dfrac{\gamma}{2}\lambda^2+(\beta-H_1)\lambda+H_2-\alpha\\ \end{array}$ g $\begin{array}{c}\\ $\begin{array}{c} $z_{1,2}=\frac{q_1\pm q_2}{2}$ $\left(\begin{smallmatrix} \frac12&-\frac12\\1&1\end{smallmatrix}\right)$ \mu^2=-\dfrac{\gamma}{3}\lambda^2+\left(\dfrac{H_1}2+H_2\right)\lambda+\alpha\\ \scriptstyle \mu_1=p_1q_1,\,\,\lambda_1=q_1^2\\ \\ \\ \mu^2=-\dfrac{\gamma}{3}\lambda^2-\dfrac{\beta}4\lambda+\dfrac{H_2}{4}-\dfrac{H_1}8\\ \scriptstyle \mu_2=2p_1,\,\,\lambda_2=q_2\\ \\ \\ \end{array}$ \scriptstyle \kappa_j=\frac{1}{2} \end{array}$ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- 0.5truecm Similar to the oscillator and the Kepler problem, the Kepler change of the time $t\to \widetilde{t}$, where \[t-change\] d=v(q)dt,v(q)=, relates the Drach systems (b),(d) and (e) with the systems (c),(f) and (k), respectively. Here $S$ are the Stäckel matrices for (b),(d) and (e) systems and $\widetilde{S}$ are the Stäckel matrices for (c),(f) and (k) systems. Remind, that these matrices $S$ and $\widetilde{S}$ have different first row only, see [@ts99a; @ts01]. Integrals of motion ------------------- Using definitions (\[k3-dr\]) we can prove that integral of motion $K_4$ (\[k3-dr\]) is the function on $H_1,H_2$ and $K_3$ K\_4\^2=16h\_1h\_2f\^2+(K\_3\^2-4h\_1g\_2\^2-4h\_2g\_1\^2)f+g\_1\^2g\_2\^2. Substituting this expression into the definition of $\Phi_2$ (\[phi-ln\]) we can get integral $K_3$ as function on the action-angle variables $I_{1,2}=H_{1,2}$ and $\omega_2$. As usual, polynomial algebra of integrals of motion $H_{1,2}$ and $K_3$ follows from the canonical brackets (\[aa-br\]) {H\_1,H\_2}&=&{H\_1,K\_3}={H\_1,K\_4}=0,\ {H\_2,K\_3}&=&K\_4,{H\_2,K\_4}=fK\_3,\[alg-b\]\ {K\_3,K\_4}&=&F\_Z(H\_1,H\_2,K\_3),where - $\delta=\phantom{-}4$,  $F_{I}=16f(g_1h_2+g_2h_1)-4g_1g_2(g_1+g_2)$ for b,c cases; - $\delta=-2$,  $F_{II}=K_3+32fh_1h_2-4(g_1^2h_2+g_2^2h_1)$ for d,f cases; - $\delta=\phantom{-}2$,  $F_{III}=4f(g_1^2+g_2^2)-16f^2(h_1+h_2)$ for e,k cases; - $\delta=\phantom{-}4$,  $F_{IV}=2f(g_1^2+8g_1h_2)-8f^2h_1-4g_1g_2^2$ for g case. The difference in the values of $\delta$ is related with the difference in $\kappa$’s which has been defined by the Drach integrals of motion [@dr35]. Of course, we can put $\kappa_1=\kappa_2=1$ in all the cases such that $\delta=2$ and polynomials $F$ look as in (\[alg-comm\]). It reduces polynomials $P_{1,2}$ in the table only. As above (\[rec-H\]), we can try to find another second order polynomial integral of motion $K_2$ from the equations $$K_3=\{H_2,K_2\},\qquad \{H_1,K_2\}=0.$$ Solutions of these equations K\_2=(2p\_1p\_2u\_1u\_2+2v\_1v\_2f+v\_1g\_2+v\_2g\_1)=\[int-k2\] have been found in [@ran97] in framework of the Lagrangian formalism. The algebras of quadratic integrals of motion $H_{1,2}$ and $K_2$ have been considered in [@cd06]. There is some opinion that all superintegrable systems with quadratic (linear) integrals of motion are multiseparable, i.e allows the separation of variables in the Hamilton-Jacobi equation in at least two different coordinate systems on the configuration space [@mw07]. The some of the Drach systems may be considered as counterexamples associated with the Lie surfaces [@cd06]. Namely, three of the superintegrable Drach systems with quadratic integrals $H_1,H_2,K_2$ are separable in the one coordinate system on the configuration space only. For the (b) and (c) Drach systems integrals of motion $H_1,K_2$ are separable in the coordinates $$x=\frac{q_2}{2q_1},\qquad y=q_1q_2,$$ and the corresponding separated relations do not allows us to get cubic integrals of motion. For the (e) and (g) cases integrals of motion $H_1,K_2$ are separable in the coordinates $$x=\frac{q_1-q_2}{2},\quad y=-\frac{(q_1+q_2)^2}{4}$$ and we can use the corresponding separated relations to the construction of the cubic integrals of motion. For the (d), (f) and (k) cases quadratic integrals of motion $H_1$ and $K_2$ do not separable by the point transformations. We can prove this Proposition by using computer program from [@ts05]. The (l) system. --------------- Without lost of generality we can put $\rho=-3$ in the (l) case. Substituting this Hamiltonian into the computer program from [@ts05] one gets the separated variables $$x=\dfrac{(q_1-q_2)^2}2,\qquad y= \dfrac{(q_1+q_2)^2}{2}$$ and the corresponding separated relations \[sep-rel-l\] p\_j\^2=P\_j(q\_j)= -4q\_j\^48q\_j\^3+4H\_1q\_1\^24q\_j+H\_2,j=1,2. which give rise to one hyperelliptic curve $\mu^2=P(\lambda)$ at $\mu=p_j$ and $\lambda=\pm q_j$. The angle variable \_2&=&12\^[q\_1]{} +12 \^[q\_2]{}\ &=&12\^[q\_1]{} -12 \^[-q\_2]{} is a sum of the incomplete elliptic integrals of the first kind on the common hyperelliptic curve. According to [@we08; @dr35] there is addition theorem and additional cubic integral of motion $${K}_3=2(\widetilde{P}'_1\,p_2+\widetilde{P}'_2\,p_1),$$ which looks like as the Drach integral (\[k3-dr\]), but in this case functions $$\widetilde{P}'_{1,2}=(q_1+q_2)^2\dfrac{\partial}{\partial q_{1,2}}\, \dfrac{P(\pm q_{1,2})}{(q_1+q_2)^4}\,$$ have completely another algebro-geometric explanation. All the details will be published in the forthcoming publications. As sequence, the algebra of integrals of motion $H_{1,2}$ and $K_3$ differs from the corresponding algebras for other Drach systems related with another addition theorem. As an example, the recurrence chain $ {K}_{j+1}=\{H_2, {K}_j\}$ terminates on the fourths step only $${K}_7=\{H_2, {K}_6\}=-480 {K}_4 {K}_3 +256H_1^2 {K}_3 -768\alpha H_2 {K}_3-3072\beta\gamma {K}_3\,.$$ The solution of the inverse recurrence chain looks like $${K}_2= p_y^2+2\alpha x-4 \gamma \sqrt{x}.$$ It is interesting, that the algebra of quadratic integrals of motion $H_{1,2}$ and $K_2$ is one of the standard cubic algebras [@cd06] and we do not explanation of this fact. New superintegrable systems on zero-genus hyperelliptic curves at $\kappa_1=1$ and $\kappa_2=2,3$ ================================================================================================= Let us put $\kappa_1=1$ and $\kappa_2=2$ in (\[uv-eq\]) and try to solve equations (\[alg-dr\])-(\[pde-dr\]). Here is one superintegrable system with cubic additional integral $K_m$ (\[k-m\]) and quadratic integral $K_\ell$ (\[k-ell\]) $$V_{III}= \gamma_1(3x+y)(x+3y)+{\gamma_2}{(x+y)}+{\gamma_3}{(x-y)}\,,\qquad S=\left( \begin{smallmatrix} a & b \\ 1 & 1 \end{smallmatrix} \right),$$ and seven systems with the real potentials $$\begin{array}{ll} V_{I}= \gamma_1(3x+y)(x+3y)+\dfrac{\gamma_2}{(x+y)^2}+\dfrac{\gamma_3}{(x-y)^2}, \qquad &S=\left( \begin{smallmatrix} a/q_1 & b/q_2 \\ 1/q_1 & 2/q_2 \end{smallmatrix} \right),\nn\\ V_{II}^{(1)}= \gamma_1 xy +\dfrac{\gamma_2}{\sqrt{x^3y}}+\dfrac{\gamma_3}{x^2}\,, \qquad &S=\left( \begin{smallmatrix} 1/q_1 & 1 \\ 1/q_1^2 & 4/q_2^2 \end{smallmatrix} \right),\nn\\ V_{II}^{(2)}= \dfrac{\gamma_1}{\sqrt{xy}}+\dfrac{\gamma_2}{x^2} +\dfrac{\gamma_3 y}{x^3}\,, \qquad &S=\left( \begin{smallmatrix} 0 & 1 \\ 1/q_1^2 & 4/q_2^2 \end{smallmatrix}\right),\nn\\ V_{II}^{(3)}= \dfrac{\gamma_1}{\sqrt{xy}}+\dfrac{\gamma_2}{\sqrt{x^3y}} +\dfrac{\gamma_3}{x^{5/4}y^{3/4}}\,, \qquad &S=\left( \begin{smallmatrix} 1 & 0 \\ 1/q_1^2 & 4/q_2^2 \end{smallmatrix}\right),\nn\\ V_{II}^{(4)}= \gamma_1 xy+\dfrac{\gamma_2 y}{x^3}+\dfrac{\gamma_3 y^3}{x^5}\,, \qquad &S=\left( \begin{smallmatrix} 0 & 4/q_2 \\ 1/q_1^2 & 4/q_2^2 \end{smallmatrix}\right),\nn\\ V_{IV}^{(1)}= \dfrac{\gamma_1}{\sqrt{xy}} +\dfrac{\gamma_2(\sqrt{x}-\sqrt{y})}{\sqrt{xy}}+\dfrac{\gamma_3}{\sqrt{xy}(\sqrt{x}+\sqrt{y})^2}\,, \qquad &S=\left( \begin{smallmatrix} 1 & q_2^2/4 \\ 1/q_1 & 1 \end{smallmatrix}\right),\nn\\ V_{IV}^{(2)}= \gamma_1 xy+\gamma_2(x-y)+\dfrac{\gamma_3}{(x+y)^2}\,, \qquad &S=\left( \begin{smallmatrix} a/q_1 & b \\ 1/q_1 & 1 \end{smallmatrix}\right)\nn \end{array}$$ for which integrals of motion $K_m$ and $K_\ell$ (\[k-m\]-\[k-ell\]) are fifth and sixth order polynomials in the momenta. Solution of the equations $K_m=\pm\{H_2,K_{m-1}\}$ and $\{H_1,K_{m-1}\}=0$ looks like K\_[m-1]{}&=&4p\_1p\_2u\_1u\_2(2fv\_2+g\_2) +4v\_2(2fv\_1+g\_1)(f v\_2+g\_2)+(4fh\_2+g\_2\^2)v\_1\ &=& 4\_2(\_1P’\_2+P’\_1)-(4fh\_2-g\_2\^2)\_1-4h\_2g\_1.It is additional integral of motion, which is second order polynomial for the system with potential $V_{III}$ and fourth order polynomial in the momenta for the other systems. Now we present some superintegrable Stäckel systems at $\kappa_1=1$ and $\kappa_2=3$. Here is one system with cubic additional integral $K_m$ (\[k-m\]) $$V_{III}^{(1)}= \gamma_1(x+2y)(2x+y)+\gamma_2(x+2y)+ \gamma_3(2x+y)\,,\qquad S=\left( \begin{smallmatrix} a & b \\ 1 & 1 \end{smallmatrix} \right),$$ and seven systems with the real potentials $$\begin{array}{ll} V_{I}= \gamma_1(x+2y)(2x+y)+\dfrac{\gamma_2}{(x+y)^2}+\dfrac{\gamma_3}{(x-y)^2}\,, \qquad &S=\left( \begin{smallmatrix} a/q_1 & b/q_2 \\ 1/q_1 & 3/q_2 \end{smallmatrix} \right), \nn\\ V_{II}^{(1)}= \gamma_1 xy+\dfrac{\gamma_2}{x^{2/3}y^{4/3}}+\dfrac{\gamma_3}{x^{1/3} y^{5/3}}, \qquad &S=\left( \begin{smallmatrix} 1/q_1 & 0 \\ 1/q_1^2 & 9/q_2^2 \end{smallmatrix} \right),\nn\\ V_{II}^{(2)}= \dfrac{\gamma_1}{\sqrt{xy}}+\dfrac{\gamma_2\sqrt{y}}{x^{5/2}}+\dfrac{\gamma_3 y^2} {x^{4}}, \qquad &S=\left( \begin{smallmatrix} 0 & 1 \\ 1/q_1^2 & 9/q_2^2 \end{smallmatrix}\right)\nn\\ V_{II}^{(3)}= \dfrac{\gamma_1}{\sqrt{xy}}+\dfrac{\gamma_2}{x^{4/3}y^{2/3}}+\dfrac{\gamma_3} {x^{7/6}y^{5/6}}, \qquad &S=\left( \begin{smallmatrix} 1 & 0 \\ 1/q_1^2 & 9/q_2^2 \end{smallmatrix}\right),\nn\\ V_{II}^{(4)}= \gamma_1 xy+\dfrac{\gamma_2 y^2}{x^4}+\dfrac{\gamma_3 y^5}{x^7}, \qquad &S=\left( \begin{smallmatrix} 0 & 1/q_2 \\ 1/q_1^2 & 9/q_2^2 \end{smallmatrix}\right),\nn\\ V_{III}^{(2)}= \gamma_1(x^2-5x\sqrt{y\,}+4y)+\dfrac{\gamma_2 x}{\sqrt{y\,}}+\dfrac{\gamma_3}{\sqrt{y\,}}\,,\qquad &S=\left( \begin{smallmatrix} 2q_1 & 2q_2 \\ 1 & 1 \end{smallmatrix}\right),\nn\\ V_{IV}=\gamma_1(x+5y)(5x+y)+\gamma_2(x-y)+\dfrac{\gamma_3}{(x-y)^2}\,, \qquad &S=\left( \begin{smallmatrix} a/q_1 & b \\ 1/q_1 & 1 \end{smallmatrix}\right)\nn \end{array}$$ for which integrals of motion $K_m$ and $K_\ell$ (\[k-m\]-\[k-ell\]) are seventh and eights order polynomials in the momenta. The algebra of integrals of motion $H_{1,2}$ and $K_m$ (\[k-m\]) is the fifth or seventh order polynomial algebra in terms of the coefficients of the hyperelliptic curves {H\_2,K\_m}=2K\_,{H\_2,K\_}=2fK\_m,{K\_m,K\_}=F\_Z,where polynomial $F_Z$ depends on the type of solution (\[uv-sol\]) only: F\_[I]{}&=&2(4fh\_2-g\_2\^2)\^[\_2-\_1]{}(4f(\_1\^2h\_2g\_1+\_2\^2h\_1g\_2)-g\_1g\_2(\_2\^2g\_1+\_1\^2g\_2)),\ F\_[II]{}&=&4(4fh\_2-g\_2\^2)\^[\_2-\_1]{}(4f(\_2+\_1)h\_2h\_1-\_1h\_1g\_2\^2-\_2h\_2g\_1\^2)K\_m\^2,\ F\_[III]{}&=&4(4fh\_2-g\_2\^2)\^[\_2-\_1]{}(4f(\_1h\_2+\_2h\_1)-\_1g\_2\^2-\_2g\_1\^2)f, \[alg-comm\]\ F\_[IV]{}&=&2(4fh\_2-g\_2\^2)\^[\_2-\_1]{}(4f(2\_2fh\_1-\_1h\_2g\_1)-2\_2fg\_1\^2+\_1g\_1g\_2\^2). Here choice of sign $+$ or $-$ depends on $\kappa$’s. As above, the Stäckel transformations (\[t-change\]) relate systems associated with one type of the solutions (\[uv-sol\]), whereas algebra of integrals of motion is invariant with respect to such transformations. The complete classification of such superintegrable systems requires further investigations. Conclusion ========== We discuss an application of the addition theorem to construction of algebraic integrals of motion from the multi-valued action-angle variables. We propose new algorithm to construction of the superintegrable Stäckel systems associated with zero-genus hyperelliptic curves. It allows us to prove that the Drach classification of the Stäckel systems with cubic integral of motion (\[k3-dr\]) associated with the addition theorem (\[add-ln\]) is complete. Moreover, we present some new two-dimensional superintegrable systems with third, fifth and seventh order integrals of motion. The proposed method may be applied to construction of the higher order additional polynomial integrals of motion for the $n$-dimensional superintegrable Stäckel systems on the different manifolds. On the other hand, we prove that there are some superintegrable systems, which miss out of this construction. It will be interesting to study a mathematical mechanism of the appearance such superintegrable systems. The research was partially supported by the RFBR grant 06-01-00140. [10]{} C. Daskaloyannis, K. Ypsilantis, , J. Math. Phys., v.47, 2006, 042904, 38 pages, math-ph/0412055. J. Drach. , v.200, p.22-26, 1935.\ J. Drach. , , v. 200, p.599-602, 1935. A. Henrici, T. Kappeler, , Preprint: arXiv:0802.4032, 2008. C.R. Holt, , J.Math.Phys., v.23, p.1037–1046, 1982. Yu.A. Grigoryev, A.V. Tsiganov, , Regular and Chaotic Dynamics, v.10(4), p.413-422, 2005. I. Marquette, P. Winternitz, , J.Math.Phys. v.48, p.012902, 2007.\ , Preprint: arXiv:0711.4783, 2007. M.F.Rañada, , J.Math.Phys., v.38, p.4165-4178, 1997. M.F. Ranada, M. Santander, , Phys. Lett. A, v.278, p.271–279, 2001. P. Stäckel, [*Uber die Integration der Hamilton–Jacobischen Differential Gleichung Mittelst Separation der Variabel*]{}, Habilitationsschrift, Halle, 1891. A.V. Tsiganov, , J. Math. Phys., v.40, p.279-298, 1999. A.V. Tsiganov, , J. Phys.A: Math. Gen., v.32, p.7965–7982, 1999. A.V. Tsiganov, , J. Phys.A: Math. Gen., v.33, p.7407-7423, 2000. A.V. Tsiganov, , J. Nonlinear Math.Phys, v.8(1), p.157-182, 2001. A.V. Tsiganov, , Regular and Chaotic Dynamics, v.13, p. 178-190, 2008. H. Weber, , Vieweg, Braunschweig, 1908.
{ "pile_set_name": "ArXiv" }
--- author: - 'D. Barducci' - 'G. Bélanger' - 'C. Hugonie' - 'A. Pukhov' bibliography: - 'nMSSM.bib' title: 'Status and prospects of the nMSSM after LHC Run-1' --- LAPTH-053/15\ LUPM:15-015 Introduction ============ The discovery of a Higgs boson with a mass of 125 GeV at the Large Hadron Collider (LHC) [@Aad:2012tfa; @Chatrchyan:2012ufa] can be viewed as an argument in favour of supersymmetry (SUSY) since a light Higgs boson is a landmark of this theory. However the mass of the new particle is only within a few GeV of the maximum value predicted in the minimal supersymmetric standard model (MSSM) and requires large contributions from the stop sector, thus raising the issue of fine-tuning [@Barbieri:1987fn; @Hall:2011aa]. In the next-to-minimal supersymmetric extension of the standard model, the NMSSM [^1], the fine-tuning issue is not as severe because of additional contributions to the lightest Higgs doublet mass, derived from the extra singlet superfield [@Ellwanger:2011mu; @Cao:2012fz; @Ellwanger:2012ke; @Perelstein:2012qg; @Agashe:2012zq; @Gherghetta:2012gb; @Cheng:2013fma; @Kim:2013uxa; @Fowlie:2014faa]. The NMSSM has the nice additional feature that the $\mu$ term is generated from the vacuum expectation value (VEV) of the new singlet field and is thus naturally at the SUSY scale, therefore solving the so-called $\mu$-problem [@Kim:1983dt]. For these reasons the discovery of the Higgs at the LHC has triggered a renewed interest in the NMSSM and phenomenological studies abound [@Munir:2013wka; @Belanger:2014roa; @Ellwanger:2014hia; @Jeong:2014xaa; @King:2014xwa; @Ellwanger:2014hca; @Bomark:2015fga; @Chakraborty:2015xia; @Potter:2015wsa]. The main focus has been on the Higgs sector, since the extra singlet can lead to new collider signatures, in particular when light, as the Standard Model (SM) like Higgs state $h$ with $m_h \sim 125$ GeV can decay into light singlet like scalars or pseudoscalars. Moreover its spin 1/2 SUSY partner, the singlino, can be at a mass well below the electroweak (EW) scale, giving also rise to peculiar SUSY signatures, especially when it is the lightest SUSY particle (LSP) and very light [@Ellwanger:2014hia; @Ellwanger:2014hca; @Potter:2015wsa; @Das:2012rr]. The NMSSM also provides a dark matter (DM) candidate, the LSP neutralino, and its properties have been analysed, both in the NMSSM with arbitrary parameters at the SUSY scale [@Belanger:2005kh; @Vasquez:2012hn] and in the grand unified theory (GUT) scale constrained models [@Hugonie:2007vd; @Belanger:2008nt; @Ellwanger:2014dfa]. In general the predictions are similar to those of the MSSM, but special features including the possibility of a light neutralino LSP with an important singlino component, can increase the annihilation into Higgs final states or resonant annihilation through a singlet Higgs. Light neutralinos LSP can also escape astrophysical constraints [@Vasquez:2010ru; @AlbornozVasquez:2011js; @AlbornozVasquez:2012px]. These studies were conducted within the framework of the $Z_3$ invariant NMSSM. However the superpotential of the general NMSSM does not necessarily possess this accidental $Z_3$ symmetry. The new minimal supersymmetric model (nMSSM), sometimes also called minimal next-to-minimal supersymmetric model (MNSSM) [@Panagiotakopoulos:1999ah; @Panagiotakopoulos:2000wp; @Dedes:2000jp], features instead a global discrete R-symmetry which forbids the singlet cubic self interaction in the superpotential, while avoiding problems with domain walls due to the $Z_3$ symmetry. Although the field content is the same as that of the ($Z_3$ invariant) NMSSM, the different superpotential and soft SUSY breaking terms lead to a peculiar phenomenology [@Panagiotakopoulos:2001zy; @Menon:2004wv; @Barger:2005hb; @Barger:2006dh; @Huber:2006wf; @Barger:2006kt; @Barger:2006sk; @Barger:2007nv; @Balazs:2007pf; @Ham:2007wu; @Huber:2007vva; @Chun:2008pg; @Ham:2008cg; @Cao:2009ad; @Cao:2010fi; @Ishikawa:2014owa]. The most striking feature of the model is that there is no mass term for the pure singlino. Only mixing effects with higgsinos can raise the singlino mass to the EW scale (up to $\sim75$ GeV [@Hesselbach:2007te]). The singlino is therefore naturally light, and the LSP generally contains a large singlino component, thus guaranteeing a phenomenology rather different from that of the MSSM both at colliders and in DM searches. Yet, all the phenomenological studies of the nMSSM have overlooked the possibility of a very light singlino LSP (below 5 GeV) in agreement with DM constraints. In addition, the results from LHC Run-1 and the prospects for Run-2 are still unexplored in this model. This is the gap we intend to fill here. In this paper we explore the parameter space of the nMSSM with unified conditions at the GUT scale, that is compatible with the latest Higgs results, with LHC searches for SUSY particles in Run-1 and with DM constraints. For this we rely on  [@Ellwanger:2005dv; @Ellwanger:2006rn] for the calculation of the spectrum and constraints on the Higgs sector, on  [@Kraml:2014sna; @Kraml:2013mwa] for comparisons with limits on simplified models, on 5 [@Conte:2012fm; @Conte:2014zja; @Dumont:2014tja] for a more complete implementation of LHC searches for sparticles and on  [@Belanger:2006is; @Belanger:2008sj; @Belanger:2014vza] for the computation of the DM observables including relic density, direct detection (DD) and indirect detection (ID). We show that this combined set of requirements lead to strong constraints on the model. We found that the allowed regions have very specific characteristics: they contain either a very light singlino LSP below 5 GeV, a mixed singlino-higgsino LSP with a mass around 45 GeV or 65 GeV, or a bino LSP with a mass around 65 GeV. We have also checked that the allowed regions in parameter space of the general model with arbitrary parameters at the SUSY scale ([*i.e.*]{} without unified conditions at the GUT scale) share the same general characteristics. After having discussed the main constraints on the model, we analyse for each region some distinctive signatures that can arise at the LHC Run-2. Here we consider both searches for new Higgs states and sparticles. The complementarity between collider and DD searches is also highlighted. Moreover, we investigate the potential of ID to probe this model, which turns out to be quite limited, except for very peculiar kinematics which could lead to an enhanced gamma-ray line [@Chalons:2011ia; @Das:2012ys; @Chalons:2012xf]. The paper is organized as follows. Section \[sec:sec2\] describes the model. Section \[sec:sec3\] contains the results of the parameter scan and the description of the different allowed regions. The LHC phenomenology of each region and the complementarity with DM searches is explored in detail in Section \[sec:sec4\]. Finally benchmark points corresponding to the channels with interesting signatures at the LHC are provided. Section \[sec:sec5\] contains our conclusions. The nMSSM {#sec:sec2} ========= The MSSM is defined by promoting each SM field $\Phi$ into a superfield $\widehat{\Phi}$, doubling the Higgs fields with two $SU(2)_L$ doublets $H_u$, $H_d$ and imposing R-parity conservation. SUSY breaking is assumed to occur in an invisible sector and to be mediated through gravitational interactions to the visible sector. The resulting theory contains a number of soft SUSY breaking terms proportional to powers of the SUSY breaking scale $M_{\rm susy}$. Unfortunately, a realistic realisation of EW symmetry breaking in the MSSM requires the presence of the so called $\mu$-term in the superpotential, coupling directly the two Higgs fields $H_u$ and $H_d$: W\_[MSSM]{} = \_u \_d + h\_u\^c\_u + h\_d\^c\_d + h\_e\^c\_d , with values of the arbitrary $\mu$ parameter close to $M_{\rm susy}$. There exist explanations for such a value of the $\mu$-term, alas, all in extended settings [@Giudice:1988yz]. The easiest solution to the $\mu$-problem is to introduce an extra gauge singlet S, coupled to the Higgs doublets and whose VEV is naturally of the order of $M_{\rm susy}$. This leads to the simplest extension of the MSSM, the NMSSM with a cubic (renormalisable) superpotential \[eq:WNMSSM\] W\_[NMSSM]{} = ł\_u\_d + \^3+ h\_u\^c\_u + h\_d\^c\_d + h\_e\^c\_d . The presence of the singlet cubic self interaction is necessary in order to avoid a global $U(1)$ Peccei-Quinn (PQ) symmetry which would lead to massless singlet states, experimentally excluded [@Ellwanger:2009dp]. The corresponding soft SUSY breaking potential is given by \[eq:VNMSSM\] V\_ &=& m\_[H\_u]{}\^2 | H\_u |\^2 + m\_[H\_d]{}\^2 | H\_d |\^2 + m\_[S]{}\^2 | S |\^2 + ( łA\_łH\_u H\_d S + A\_S\^3 + )\ && + m\_Q\^2|Q\^2| + m\_U\^2|U\_R\^2| + m\_D\^2|D\_R\^2| + m\_L\^2|L\^2| + m\_E\^2|E\_R\^2|\ && + ( h\_u A\_u Q H\_u U\_R\^c - h\_d A\_d Q H\_d D\_R\^c - h\_e A\_e L H\_d E\_R\^c + )\ && + M\_1 + M\_2 + M\_3 . The standard NMMSM given by Eqs. (\[eq:WNMSSM\]-\[eq:VNMSSM\]) possesses a global $Z_3$ symmetry under which all superfields are multiplied by $e^{2i\pi/3}$. Once this discrete symmetry is spontaneously broken during the phase transition associated with the EW symmetry breaking in the early universe, cosmologically dangerous domain walls are generated  [@Vilenkin:1984ib]. It has been argued in  [@Abel:1995wk; @Abel:1995uc] that these walls become harmless if they disappear before nucleosynthesis, which requires the presence in the effective potential of $Z_3$ breaking terms of magnitude V \~O(1 [MeV]{})\^4 . The magnitude of the $Z_3$ breaking terms corresponds to the presence in the superpotential or in the Kähler potential of $Z_3$ breaking operators suppressed by one inverse power of the Planck mass, $M_{\rm Planck}$. However, these $Z_3$ breaking (non-renormalisable) terms involving the singlet S induce divergent tadpoles [@Nilles:1982mp; @Lahanas:1982bk; @Ellwanger:1983mg; @Nilles:1997me; @Bagger:1993ji; @Jain:1994tk; @Bagger:1995ay] of the form W = \_F M\_[susy]{} M\_[Planck]{} S , V = \_S M\_[susy]{}\^2 M\_[Planck]{} (S + S\^\*) , thus reintroducing a hierarchy problem. The values of $\Xi_F$ and $\Xi_S$ depend on the loop order at which the tadpoles are generated, which in turn depends on the particular non-renormalisable terms that give rise to the tadpoles. A solution to both the domain wall and the stability problem is to impose a discrete $R$-symmetry on the complete theory (including non-renormalisable operators) such that the tadpole terms are generated at high loop order [@Panagiotakopoulos:1998yw]. One then obtains effective tadpole terms W = \_F S , V = \_S (S + S\^\*) , \_F M\_[susy]{}\^2 \_S M\_[susy]{}\^3 . In the case where $\xi_F \sim M_{\rm susy}^2$ and $\xi_S \sim M_{\rm susy}^3$ the singlet cubic self interaction in the superpotential (\[eq:WNMSSM\]) is not even phenomenologically required and can be omitted [@Panagiotakopoulos:1999ah]. The resulting model has been denoted as the new MSSM or nMSSM as, in the limit where SUSY is unbroken, the MSSM $\mu$ term is only traded for the dimensionless $\l$ coupling. Once SUSY is softly broken, the generated tadpole terms $\xi_F$ and $\xi_S$ break both the $Z_3$ and the PQ symmetry. The superpotential of the nMSSM then reads \[eq:WnMSSM\] W\_[nMSSM]{} = ł\_u\_d + \_F+ h\_u\^c\_u + h\_d\^c\_d + h\_e\^c\_d and the corresponding soft SUSY breaking potential is given by \[eq:VnMSSM\] V\_ &=& m\_[H\_u]{}\^2 | H\_u |\^2 + m\_[H\_d]{}\^2 | H\_d |\^2 + m\_[S]{}\^2 | S |\^2 + ( łA\_łH\_u H\_d S + \_S S + )\ && + m\_Q\^2|Q\^2| + m\_U\^2|U\_R\^2| + m\_D\^2|D\_R\^2| + m\_L\^2|L\^2| + m\_E\^2|E\_R\^2|\ && + ( h\_u A\_u Q H\_u U\_R\^c - h\_d A\_d Q H\_d D\_R\^c - h\_e A\_e L H\_d E\_R\^c + )\ && + M\_1 + M\_2 + M\_3 . In this paper we study the general nMSSM with arbitrary soft terms at the SUSY scale as well as the semi-universal nMSSM for which one imposes the following constraints on the soft terms at the GUT scale { m\_Q = m\_U = m\_D = m\_L = m\_E m\_0\ A\_u = A\_d = A\_e A\_0\ M\_1 = M\_2 = M\_3 M\_[1/2]{} . . In both cases one can trade the (free) parameters $m_{H_u}, m_{H_d}, m_{S}$ for the Higgs VEVs $v_u, v_d, s$, or equivalently for $\mu \equiv \l s$, $\tb \equiv \ds\frac{v_u}{v_d}$ and the known value of $M_Z^2 = g^2 v^2$, where $g^2 = (g_1^2 + g_2^2)/2$, $v = \sqrt{v_u^2 + v_d^2} \approx 174$ GeV, and $g_1$, $g_2$ denote the $U(1)_Y$ and $SU(2)_L$ gauge couplings, respectively. From the SUSY $F$- and $D$-terms and the soft SUSY breaking terms one obtains the potential for the neutral Higgs fields: V\_ & = & | - łH\_u\^0 H\_d\^0 + \_F |\^2 + (|H\_u\^0|\^2 - |H\_d\^0|\^2)\^2 + m\_[S]{}\^2 |S|\^2\ && + (m\_[H\_u]{}\^2 + |łS|\^2) |H\_u\^0|\^2 + (m\_[H\_d]{}\^2 + |łS|\^2) |H\_d\^0|\^2\ && + (- łA\_łH\_u\^0 H\_d\^0 S + \_S S + ) , which at the minimum is V\_0 & = & (-łv\_u v\_d + \_F)\^2 + (v\_u\^2 - v\_d\^2)\^2 + m\_[S]{}\^2 s\^2\ && + (m\_[H\_u]{}\^2 + \^2) v\_u\^2 + (m\_[H\_d]{}\^2 + \^2) v\_d\^2 - 2 łA\_łv\_u v\_d s + 2\_S s . The minimisation equations are given by \[eq:min\] { v\_u (m\_[H\_u]{}\^2 + \^2 + ł\^2v\_d\^2 +(v\_u\^2-v\_d\^2)) - v\_d (A\_ł+ł\_F)= 0 ,\ v\_d (m\_[H\_d]{}\^2 + \^2 + ł\^2v\_u\^2 +(v\_d\^2-v\_u\^2)) - v\_u (A\_ł+ł\_F)= 0 ,\ s(m\_[S]{}\^2 + ł\^2(v\_u\^2+v\_d\^2)) + \_S - łA\_łv\_u v\_d = 0 . . From the first two of these equations one can derive 2= , and from the third, one obtains in the limit where $s \gg v$ (or equivalently $\l \ll 1$) - . In the basis $(H_{dR}, H_{uR}, S_R)$ and after the elimination of $m_{H_d}^2$, $m_{H_u}^2$ and $m_{S}^2$ using the minimisation equations (\[eq:min\]), the elements of the $3 \times 3$ CP-even mass matrix ${\cal M}_S^2$ read \_[S,11]{}\^2 & = & g\^2 v\_d\^2 + (A\_ł+ł\_F) ,\ [M]{}\_[S,22]{}\^2 & = & g\^2 v\_u\^2 + (A\_ł+ł\_F)/ ,\ [M]{}\_[S,33]{}\^2 & = & ,\ [M]{}\_[S,12]{}\^2 & = & (2ł\^2 - g\^2) v\_u v\_d - (A\_ł+ ł\_F) ,\ [M]{}\_[S,13]{}\^2 & = & ł(2 v\_d - A\_łv\_u) ,\ [M]{}\_[S,13]{}\^2 & = & ł(2 v\_u - A\_łv\_d) . Dropping the Goldstone mode, one can express the $2 \times 2$ CP-odd mass matrix ${\cal P}_S^2$ in the basis (${A}, S_I$), where ${A} = \cos\b\, H_{uI}+ \sin\b\, H_{dI}$ \_[P,11]{}\^2 = , \_[P,22]{}\^2 = , \_[P,12]{}\^2 = łA\_łv . One can notice that ${\cal M}_{S,33}^2 = {\cal M}_{P,22}^2$, [*i.e.*]{} in the limit of small mixing between the singlet and doublet sectors, the CP-even and CP-odd singlet states have the same mass (up to radiative corrections, see below). In addition, this common (tree level) mass depends on the tadpole parameter $\xi_S$ which is a free parameter. Hence singlet like Higgs masses are arbitrary. In particular, they can be lighter than 125 GeV and still not excluded if their reduced couplings to SM particles (especially gauge bosons) are sufficiently suppressed. Finally, in the basis $\psi^0 = (-i\l_1, -i\l_2^3, \psi_d^0, \psi_u^0, \psi_S)$, the neutralino mass matrix reads \_0 = ( M\_1 & 0 & - & & 0\ & M\_2 & & - & 0\ & & 0 & -& -łv\_u\ & & & 0 & -łv\_d\ & & & & 0 )  . Therefore, in the limit of small singlino-higgsino mixing ($\l \ll 1 $), the singlino mass is $m_{\widetilde S} = 0$. On the other hand, if either $\mu$ or $M_1$ and $M_2$ are much larger than $M_Z$, one gets [@Hesselbach:2007te] \[eq:singlino\] m\_[S]{} 2 . The experimental lower bound on $\mu$ (from the non observation of a light chargino) and the theoretical upper bound on $\l$ (assuming perturbativity up to the GUT scale) therefore yield an upper bound on the singlino mass $m_{\widetilde S} \lesssim 75$ GeV. The physical CP-even Higgs states will be denoted as $h_i, i = 1,2,3$ (ordered in mass), and the physical CP-odd Higgs states as $a_i, i = 1, 2$. The neutralinos are denoted as $\tilde\chi^0_i, i = 1\dots5$ and their mixing angles $N_{i,j}$ such that $N_{1,5}$ indicates the singlino component of the lightest neutralino $\tilde\chi^0_1$ (assumed to be the LSP). All the above expressions are for tree level mass matrices. Loop corrections play an important role, especially in the Higgs sector where they account for a large part of the SM like Higgs mass at 125 GeV. To compute the SUSY and Higgs spectrum, we have used the  package, setting the precision for radiative corrections to the minimum (precision flag for Higgs calculations = 0 in the  input files). This includes the full one loop and the leading log two loop contributions from (s)top/(s)bottom, as well as the leading log one loop EW corrections. We have not used the most precise computation (precision flag for Higgs calculations = 2) of ref. [@Degrassi:2009yq], as it is valid only for the $Z_3$ invariant NMSSM. In this ($Z_3$ invariant) limit however, we have checked that the difference between the two computations is usually $\lesssim 3$ GeV for the SM like Higgs state near 125 GeV. In addition, a slight change of input parameters can always reproduce the same Higgs spectrum with both flags [^2]. The minimal precision for radiative corrections presents the extra advantage of using less CPU time, which is crucial for scans on large parameter space. In addition it allows to compute easily the complete effective Lagrangian in the Higgs sector with the same level of approximation. This complete Lagrangian can then be fed into  so as to compute the relic density of the LSP DM candidate as well as its DD and ID rates. Note that the higher-order corrections to the Higgs self-couplings encoded in the effective Lagrangian can in some cases have a significant effect on the DM relic density. Parameter scan {#sec:sec3} ============== The parameter exploration of the semi-universal nMSSM (as defined in Sec. \[sec:sec2\]) has been carried out using  `v4.6.0`, scanning over the following parameters [^3]: m\_0, M\_[1/2]{}, A\_0, , , ł, \_F, \_S, A\_ł, which are all defined at the GUT scale except $\tb$ (at $M_Z$) and $\l$, $\mu$ (at the SUSY scale). To efficiently scan over the nMSSM parameter space we have employed the Markov Chain Monte Carlo (MCMC) routines implemented in the  package, which we have tuned in order to cover in details regions of parameter space corresponding to lighter sparticles, [*i.e.*]{} with higher experimental prospects. Scenarios with very heavy sparticles (out of the LHC Run-2 reach) have been discarded. We have applied all the default constraints implemented into  (except for the constraint on the anomalous magnetic moment of the muon g$_{\mu}-2$), which include in particular [^4]: - No unphysical minimum of the Higgs potential, - No Landau pole below the GUT scale, - Invisible Z width $\Gamma_Z < 0.5$ MeV, - $B$-physics constraints, - LEP and Tevatron searches for sparticles and Higgs bosons, - Tevatron and LHC searches on charged Higgs via top decays, - At least one Higgs boson in the $125.1\pm3$ GeV mass range, - $\chi^2$ fit to the Higgs signal strengths [@Belanger:2013xza]. The latter indirectly takes into account the limit on non-standard decays of the SM like Higgs, such as the decay into light Higgs states or the invisible decays into the LSP which are somewhat dependent on shifts of other Higgs couplings. Moreover we have checked a posteriori that the direct limits on heavy Higgs states in the $WW$ channel were satisfied [@Aad:2015agg]. In performing our scan we have also required the DM relic density $\Omega h^2$ to be compatible with the relic abundance measured by Planck [@Adam:2015rua], $\Omega h^2_{\rm Planck}= 0.1186\pm0.0020$ at 68% CL. We have chosen to impose just an upper bound of the relic density, $\Omega h^2<0.131$, which takes into account $\sim$ 10% theoretical uncertainties that could arise from loop corrections into the DM annihilation cross section, see [*e.g.*]{} [@Baro:2007em]. We have also required the spin independent cross section for DD, rescaled for the local DM abundance ($\sigma^{\rm SI}_{\rm rescaled}=\sigma^{\rm SI} \Omega h^2/\Omega h^2_{\rm Planck}$), to be compatible with the latest LUX results [@Akerib:2013tjd]. We illustrate in Fig. \[fig:scan\_res\] the results of the scan mapped in the $m_{\tilde\chi^0_1}$ - $\Omega h^2$ plane, showing in blue the points with a DM relic density compatible with Planck (0.107$<\Omega h^2<$0.131) and in red the points for which it is below ($\Omega h^2<$0.107). We see that three different regimes for the LSP mass exist: a region with a very light LSP below 5 GeV (region 1), a region with a $\sim$ 45 GeV LSP (region 2) and a region with a $\sim 65$ GeV LSP (region 3). In the first region the DM annihilation proceeds through a light pseudoscalar (singlet like) Higgs resonance, the second corresponds to the $Z$ resonance and the third to the exchange of a SM like Higgs or $Z$ boson. Note that the gap for neutralino masses between 5 and 40 GeV is mostly due to constraints from the invisible width of the Higgs. The light LSP is a nearly pure singlino, so that the Higgs invisible width is very small. As the singlino mass increases, the same does its higgsino component, hence increasing the contribution to the Higgs invisible width. ![DM relic density $\Omega h^2$ in function of the LSP mass. Blue (red) points correspond to a DM relic density $0.107\le\Omega h^2\le 0.131$ ($\Omega h^2<0.107$).[]{data-label="fig:scan_res"}](figures/DM_LSP.pdf){width="46.00000%"} As we will discuss in the following, the 5 GeV and 65 GeV LSP region present two and three sub-regions respectively, mapped in different areas of the $m_0$-$ M_{1/2}$ parameter space. We report in Tabs. \[tab:scan-res-par1\]-\[tab:scan-res-par1\_bis\] the maximum and minimum values for all of the nMSSM input parameters, in the three regions and sub-regions that we have identified. Moreover in this Tables we also specify the weak scale gaugino masses and give a quick overview of the sparticle spectrum in the different sub-regions. All regions have small $\mu$ and several sub-regions feature sfermions and/or gauginos at the multi TeV scale. We distinguish region 1 for which the 125 GeV (SM like) Higgs state is the second lightest CP even ($h_2$) while the lightest ($h_1$) is mainly singlet, from regions 2 and 3 where the lightest CP even Higgs state ($h_1$) is SM like with a mass of 125 GeV. In the former case (with intermediate values of $\tb$) the SM like Higgs state takes its (relatively heavy) mass from $h_1 / h_2$ mixing effects. In the latter, $\tb \approx 2$ and $\l$ is near its maximal theoretically allowed value (from perturbativity up to the GUT scale). Hence the pure nMSSM contribution to the lightest Higgs mass is maximal. Region 1A 1B 2 ------------------ ---------------------------------- ----------------------------------------- ----------------------------------------- $\tb$ 6.6     10 6     8 1.5     2.1 $\l$ 0.33     0.53 0.49     0.52 0.68     0.80 $\mu$ 240     400 350     430 115     180 $m_0$ 0     1080 4040     4800 2$\cdot 10^4$     6$\cdot10^5$ $ M_{1/2}$ 630     1200 280     440 180     470 $A_0$ $-$1700     50 6700     7900 $-$3.7$\cdot10^4$     $-$2.5$\cdot10^3$ $A_\l$ 1400     6000 7000     7900 $-$1.3$\cdot10^4$     3.3$\cdot10^4$ $\xi_F$ 10     100 $-$1.5$\cdot10^4$     $-$1.4$\cdot10^4$ 3.7$\cdot10^4$     $5.1\cdot10^6$ $\xi_S$ $-$6$\cdot10^4$     2$\cdot10^4$ $-$1.9$\cdot10^7$     $-$1.6$\cdot10^7$ $-$5.2$\cdot10^{10}$     9.7$\cdot10^8$ $M_1$ 270     520 110     190 95     225 $M_2$ 500     950 200     340 160     400 $m_{\tilde q}$ 1300    2400 $>$ 3000 $>$ 20000 $m_{\tilde t_1}$ 350    1300 1050    1900 $>$ 3000 $m_{\tilde l}$ 180    1100 $>$ 3000 $>$20000 $m_{\tilde g}$ 1450    2600 780    1250 800    1500 : Maximum and minimum parameter values surviving the constraints implemented in  for regions 1 and 2. Also indicated are the $M_1$ and $M_2$ values at the SUSY scale and the mass ranges for sparticles. Dimensionfull parameters are expressed in GeV except $\xi_F$ and $\xi_S$ which are in GeV$^2$ and GeV$^3$ respectively.[]{data-label="tab:scan-res-par1"} Region 3A 3B 3C ------------------ -------------------------------------- ----------------------------------- ---------------------------------------- $\tb$ 1.8     2.5 1.8     2.8 1.3     1.8 $\l$ 0.64     0.77 0.66     0.74 0.65     0.74 $\mu$ $-$140     $-90$ $-$110     $-$90 110     150 $m_0$ 3.9$\cdot 10^3$     6.6$\cdot10^4$ 170     2500 0     3150 $ M_{1/2}$ 130     210 200     560 5.6$\cdot10^3$     2.3$\cdot10^4$ $A_0$ 11     3.3$\cdot10^4$ 440     3600 $-$1.9$\cdot10^4$     4.7$\cdot10^3$ $A_\l$ 6.2$\cdot 10^3$     2.7$\cdot10^4$ 450   3500 4     8200 $\xi_F$ 4.8$\cdot10^5$     3.4$\cdot10^7$ 4.3$\cdot10^5$     7.5$\cdot10^6$ 4.6$\cdot10^5$     $1.8\cdot10^7$ $\xi_S$ 2.0$\cdot10^9$     1.3$\cdot10^{11}$ 1.0$\cdot10^8$     4.9$\cdot10^9$ $-$1.9$\cdot10^{9}$     3.0$\cdot10^9$ $M_1$ 52     65 85     230 2.6$\cdot10^3$     1.1$\cdot10^4$ $M_2$ 83     108 160     430 4.6$\cdot10^3$     2.0$\cdot10^4$ $m_{\tilde q}$ $>$ 3000 780  2500 $>$ 10000 $m_{\tilde t_1}$ 500  $>$20000 150  550 $>$ 5000 $m_{\tilde l}$ $>$3000 100  2500 100  3000 $m_{\tilde g}$ 450  650 590  1300 $>$ 10000 : Maximum and minimum parameter values surviving the constraints implemented in  for region 3. See caption of Tab. \[tab:scan-res-par1\].[]{data-label="tab:scan-res-par1_bis"} In the next Section we will analyse these three regions separately, first checking what are the constraints set by LHC Run-1 and ID experiments, then describing the prospects of this scenarios for the LHC Run-2, together with the prospects for future DD experiments. To check the compatibility with the LHC constraints we have used both the packages  `v1.0.2` and 5. The former is a tool designed to decompose the signal of a given SUSY model into simplified topologies that are searched for by ATLAS and CMS, taking into account that in a generic SUSY spectrum the assumptions on the sparticles decays can be (and usually are) different from the ones assumed in the experimental searches. The input of  consists of an `SLHA` file [@Allanach:2008qq], containing particle spectrum and decay tables, together with SUSY production cross sections, which we have computed with  `v4.1.7`. For strong production of sparticles we have also applied next-to-leading-order and next-to-next-to-leading-log (NLO+NNLL) QCD $\kappa$-factors, which have been computed with the `nllfast` package [@Beenakker:1996ch; @Beenakker:1997ut; @Kulesza:2008jb; @Kulesza:2009kq; @Beenakker:2009ha; @Beenakker:2010nq; @Beenakker:2011fu]. Note that due to the presence of many states below the TeV scale, and in particular of a very light LSP, for some regions of parameter space the coloured sparticles possess a large number of possible decay chains, including long ones. Hence building the list of existing topologies can be much more computer-time consuming than in the MSSM. Conversely 5 is a multi purpose package designed for the analysis of events generated at parton level and/or reconstructed level with the inclusion of parton showering and detector effects. In this work we have exploited the 5 Public Analysis Database, [*i.e.*]{} a list of implemented and validated experimental searches, with which one can recast the experimental limits set by ATLAS and CMS in a generic model or in a particular model configuration. Clearly the approach of 5 is more general, since it does not rely on any simplified model assumption, but it has the drawback of being somewhat more time consuming, given the need to generate events with a Monte Carlo generator for each model point that one wants to test. In this paper we will use a combination of both tools to check the limits set by LHC Run-1 in the nMSSM, using 5 especially to set bounds on the gluino mass. LHC and DM phenomenology {#sec:sec4} ======================== After having defined the different regions of the GUT scale parameter space obeying the basic set of collider, astrophysical, cosmological and theoretical constraints, we now examine for each scenario the constraints from searches for sparticles at LHC Run-1 as well as from indirect searches for DM. For each scenario we then discuss what are the most relevant searches at LHC Run-2, both for sparticles and Higgs states, and we provide benchmark points with specific nMSSM signatures. The complementarity with direct DM searches is also highlighted. Region 1, $m_{\tilde\chi^0_1}<$ 5 GeV ------------------------------------- In this region the LSP is a quasi pure singlino state and it can appear in two different configurations of the $m_0$-$ M_{1/2}$ parameter space. One is characterised by small $m_0$ (${\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}$ 1 TeV) and $ M_{1/2}\sim$ 1 TeV (region 1A) while the second one has small $ M_{1/2}$ (${\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}$ 500 GeV) and large $m_0$ $\sim$ 4 TeV (region 1B). These two regions give rise to different sparticle spectra that we will analyse separately. We however stress three common features in these sub-regions. The first is that the lightest pseudoscalar and the singlino masses are linked, with $|m_{a_1}-2m_{\tilde\chi^0_1}|{\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}1.5$ GeV, in order to ensure efficient annihilation of the singlino through $a_1$ resonance, so as to satisfy the relic density constraint. The second follows from the singlino nature of the LSP. Decays of sparticles to a final state containing the LSP will often be suppressed thus, phase space allowing, longer decay chains involving heavier neutralinos, such as $\tilde t_1\to t \tilde\chi_2^0\to t \tilde\chi^0_1 Z$, will be preferred. This has important consequences for simplified models limits. Finally, the measured DM relic abundance can be accounted for just by the LSP itself. However, as explained in the previous Section, we will work in the less constraining assumption of a relic abundance $\Omega h^2<$ 0.131, thus allowing for the possibility of another DM component. ### Current constraints {#sec:1AB_constraints} In the region with small $m_0$ (1A) the spectrum is characterised by the presence of light stops ($m_{\tilde t_1}< 1300$ GeV) and light sleptons ($m_{\tilde l}\sim 180$–$1100$ GeV), with first and second generation squarks heavier than 1.3 TeV and gluinos heavier than 1.45 TeV. The most stringent constraints on coloured sparticles are thus easily evaded. Given the values of $\mu, M_1$ and $M_2$ reported in Tab. \[tab:scan-res-par1\], the EWino spectrum is characterised, besides the light singlino, by four neutral and charged states with a mass between 240 and 520 GeV corresponding to the three higgsinos and the bino. In the Higgs sector the role of the SM Higgs boson is played by $h_2$, while the $h_1$ mass is between 35 and 70 GeV and the lightest pseudoscalar $a_1$ has a mass $\sim$ 2 $m_{\tilde\chi^0_1}$. The masses of the three states of the heavy Higgs doublet are above 1.5 TeV. Note that pure scalar and pseudoscalar singlets are expected to have the same mass, however the mixing with the light scalar doublet is sufficient to increase the mass of the scalar singlet by a few tens of GeV. We show in Fig. \[fig:1A\_smodels\] the allowed and excluded points after the application of the constraints implemented in , projected in the $m_{\tilde e_L}$-$m_{\tilde t_1}$ plane. For each excluded point we indicate the most constraining analysis. This is defined in  as the analysis that has the larger ratio between the theoretical prediction and the experimental measurement for a given channel. However more than one channel can exclude the same point. From Fig. \[fig:1A\_smodels\] it is clear that the most constraining searches are the ATLAS search for slepton production [@Aad:2014vma] ($\tilde l\to l \tilde\chi^0_1$) and the CMS search for stop production [@Chatrchyan:2013xna] ($\tilde t\to b \tilde\chi^+_1$ and $\tilde t\to t \tilde\chi^0_1$). In particular the former is able to exclude sleptons lighter than $\sim 300$ GeV while the latter sets a lower bound of $\sim 550$ GeV on the lightest stop mass. These are stringent bounds, in the sense that no sparticles lighter than these limits are allowed. It is important to notice that, while the reach on the slepton mass is close to the official result of the ATLAS analysis for a 5 GeV LSP ($m_{\tilde l}{\stackrel{\displaystyle >}{\,_{\! \,_{\displaystyle \sim}}}}330$ GeV), this is not the case for the stop search, where the ATLAS limit is around 650 GeV. However this can be explained by the fact that the simplified model result assumes a 100% branching ratio either for $\tilde t_1\to t \tilde\chi^0_1$ or $\tilde t_1\to b \tilde\chi^+_1$, an assumption which is not satisfied here. First, the mass spectrum is such that there is always at least one decay channel into a heavier neutralino which is allowed. Moreover, the decays into heavier neutralinos typically have larger branching ratios than the decay into the singlino LSP. This causes therefore a small reduction of the LHC exclusion reach. ![Allowed and excluded points for region 1A in the $m_{\tilde e_L}$-$m_{\tilde t_1}$ plane. For each excluded point we indicate the search with the maximum sensitivity.[]{data-label="fig:1A_smodels"}](figures/1A_smodels.pdf){width="46.00000%"} The second region with a light LSP (1B) corresponds to large $m_0$ ($\sim 4$ TeV) and small $ M_{1/2}$. The large value of $m_0$ yields heavy sfermions with slepton and squark masses above 3 TeV, except for the lightest stop which is between 1–2 TeV. Therefore the only sub-TeV sparticles are EWinos and gluinos. In particular the neutralino spectrum has the following hierarchy: a light singlino ($m_{\tilde\chi^0_1}<$ 5 GeV), a bino ($m_{\tilde \chi_2^0}\sim 100$–$200$ GeV), a wino ($m_{\tilde\chi^0_3}\sim$ 200–300 GeV) which is degenerate with the lightest chargino, and heavier higgsino states since typically $\mu$ is larger than $M_1, M_2$ in region 1A (see Tab. \[tab:scan-res-par1\]). The gluinos lie in the 800–1200 GeV range, while $h_1$ is heavier than in region 1A, between 70 and 90 GeV. Contrary to what we found in region 1A,  does not set any constraint on the parameter space of this region. Clearly sleptons and squarks, including stops, escape the limits set by the experimental searches due to their high masses. Conversely, the reasons for EWinos escaping the LHC limits are less straightforward. The chargino mass in region 1B is typically above the bound set by ATLAS from the search for $\tilde\chi_1^+\tilde\chi^-_1$ production, with a subsequent $\tilde\chi^\pm_1\to W^\pm \tilde\chi^0_1$ decay, which, for a light LSP, is $m_{\tilde\chi^+} > 180$ GeV [@Aad:2014vma]. The other simplified topology analysed by ATLAS which is relevant for region 1B is $\tilde\chi^\pm_1\tilde\chi^0_2\to W Z \tilde\chi^0_1 \tilde\chi^0_1$. It assumes 100% EWinos decays into the LSP plus a SM gauge boson, as well as pure wino states ($m_{\tilde\chi^+_1} = m_{\tilde\chi^0_2}$). Considering the neutralino compositions, the latter criteria applies in our case only to $\tilde\chi^0_3$ while the decay assumption applies only to $\tilde\chi^\pm_1$. Indeed, the branching ratio for $\tilde\chi^0_3 \to Z\tilde\chi^0_1$ can be suppressed when decay rates into $h_{1,2}\tilde\chi^0_1$ and $a_1 \tilde\chi^0_1$ become important. The same comment can be made about the $\tilde\chi^0_2$ decays. Moreover the $\tilde\chi^0_2$ composition, which is mainly bino, causes a significant decrease of the $\tilde\chi^\pm_1\chi^0_2$ cross sections with respect to the wino case assumed in simplified models. Finally, the lack of sensitivity of the ATLAS search in a $3l+E_T^{\rm miss}$ final state [@Aad:2014nua] can be explained by the small branching ratios of the neutralinos in lepton pairs and LSP, which is typical of scenarios with heavy sleptons. The case of the gluino deserves special consideration, since a priori its mass can be well within the range excluded by simplified models analyses. As before, the issue is that we find reduced branching ratios in the standard search channels ($\tilde g\to t\bar t\tilde\chi^0_1,t\bar b\tilde\chi^0_1$, $q\bar q\tilde\chi^0_1$) because of competing decays into heavier neutral and charged EWinos, that can once again be linked to the singlino nature of the LSP. However, gluino searches generally rely on a large number of jets in the final state. These signal regions can be easily fed also from gluino decays different from the standard assumptions, in particular decays into heavier neutralinos. We have then used the recast [@MA5:CMS-SUS-13-016] CMS gluino search [@CMS:2013ija], implemented in the package 5, which relies on an opposite sign dilepton final state, more than 4 jets and more than 2 b-tag jets, together with a large $E_T^{\rm miss}$. This search sets a 95% CL exclusion when more than 4 events fall in its unique signal region. We have simulated gluino pair production with `MadGraph5 v1.5.11` [@Alwall:2014hca]. Parton showering, hadronization and decay of unstable particles have been performed with the package `Pythia v6.4` [@Sjostrand:2006za], while the 5 tuned version of `Delphes v3.2.0` [@deFavereau:2013fsa] has been used to simulate detector effects. Jets are reconstructed with `FastJet` [@Cacciari:2011ma], via an anti-$k_T$ [@Cacciari:2008gp] algorithm. We have then computed, through the efficiency map function implemented in 5, acceptances times efficiencies for various values of the gluino mass. From these, we have then calculated the final number of events for our nMSSM scenarios, using the production cross sections for gluino pair production computed via  and `nllfast`. We have found a lower limit on the gluino mass of $\sim 1.1$ TeV, close to the official CMS result [@CMS:2013ija]. This result strongly constrains region 1B, leaving just a few points with a heavy enough gluino. Finally, we have also computed the ID cross section for LSP pair annihilation into $\tau^+\tau^-$ final state (the most relevant at these masses) rescaled by the square of the local DM density, $[\Omega h^2/ (\Omega h^2)_{\rm Planck}]^2$, and compared it with the exclusion limits set by Fermi-LAT [@Ackermann:2015zua]. The results are shown in Fig. \[fig:1A\_ID\]. For a $\sim 5$ GeV LSP the Fermi-LAT limit is approximately constant and equal to $10^{-27}{\rm cm^3/s}$. In the left panel we show all the points surviving the LHC Run-1 constraints (implemented as discussed above), while in the right panel we zoom on to the region with small $m_{a_1}-2 m_{\tilde\chi^0_1}$. Blue (red) points correspond to a DM relic density $0.107\le\Omega h^2\le 0.131$ ($\Omega h^2<0.107$). From the plots it is clear that the current results from Fermi-LAT are able to completely exclude the portion of parameter space which corresponds to $m_{a_1}<2 m_{\tilde\chi^0_1}$, while above this threshold ID detection rates are well below the experimental limits. The reason for this is simply that when $2m_{\tilde\chi^0_1}$ is just below $m_{a_1}$, its annihilation cross section at small velocities can be significantly enhanced by the resonance effect, while in the early universe the thermal velocity of the neutralino is enough to overshoot the resonance [@AlbornozVasquez:2011js; @Bi:2009uj]. The increase in the rescaled cross section with $m_{a_1}-2 m_{\tilde\chi_1^0}>0.1~\rm{GeV}$ in Fig. \[fig:1A\_ID\] (left panel) is an effect of the rescaling. Indeed $\sigma v$ in the galaxy as well as in the early universe decreases when moving away from the narrow pseudoscalar resonance until one reaches a region where $m_{a_1}-2 m_{\tilde\chi_1^0}$ is too large to benefit from a strong resonance enhancement and the relic density is in agreement with PLANCK data. However since the rescaling factor is inversely proportional to the square of the annihilation cross section in the early universe, the net effect is an increase of the rescaled cross section for DD until $m_{a_1}-2 m_{\tilde\chi_1^0} \approx 1 $ GeV. Note that limits from AMS antiproton are not expected to be important for such low masses [@Cirelli:2013hv]. ![ID detection rates into $\tau^+\tau^-$ final state in function of $m_{a_1}-2 m_{\tilde\chi^0_1}$. Blue (red) points correspond to a DM relic density $0.107\le\Omega h^2\le 0.131$ ($\Omega h^2<0.107$). The horizontal lines represent the current limit from Fermi-LAT. Regions 1A and 1B are shown together.[]{data-label="fig:1A_ID"}](figures/1AB_ID.pdf "fig:"){width="46.00000%"}![ID detection rates into $\tau^+\tau^-$ final state in function of $m_{a_1}-2 m_{\tilde\chi^0_1}$. Blue (red) points correspond to a DM relic density $0.107\le\Omega h^2\le 0.131$ ($\Omega h^2<0.107$). The horizontal lines represent the current limit from Fermi-LAT. Regions 1A and 1B are shown together.[]{data-label="fig:1A_ID"}](figures/1AB_ID_zoom.pdf "fig:"){width="46.00000%"} ### LHC Run-2 and direct detection experiments prospects We now discuss the prospects for the LHC Run-2, including sparticles and Higgs searches, as well as DD experiments for the points surviving the constraints imposed by the LHC Run-1 and ID measurements. As shown in Fig. \[fig:1A\_smodels\] standard searches for stops quarks are already quite effective in constraining light stops in region 1A, for which the dominant decay mode is $\tilde t_1\to b\tilde\chi^+_1 $, and Run-2 is therefore expected to significantly extend the reach in this channel. The standard search channel $\tilde t_1\to t\tilde\chi^0_1$ does not offer as good prospects, since the branching ratio into the singlino LSP is typically suppressed. For example we found a 600 GeV stop with a $\sim$ 60% branching ratio into a chargino and only $\sim$ 5% in $t \tilde\chi^0_1$, with the rest of the decays being saturated by $t\tilde\chi^0_{2,3}$. Since $\tilde\chi^0_{2,3}$ mainly decay into a $Z$ or a SM like Higgs boson and a LSP, final states with $tZ+E_T^{\rm miss}$ or $th_2+E_T^{\rm miss}$, arising from one or both of the pair produced stops, provide a characteristic signature in this region. Also slepton searches will clearly act as a powerful probe of this region. In fact, only a factor two improvement in the LHC Run-2 mass reach, relative to the one of Run-1, would cover most of region 1A, leaving only a few points with sleptons heavier than 500 GeV. Upgrades of standard EWinos searches, as in ref. [@Aad:2014nua], will also be able to further extend the LHC reach. In fact, even if in Fig. \[fig:1A\_smodels\] the strongest bound represented is the one arising from slepton pair production, the decay pattern of $\tilde\chi^+_1$ and $\tilde\chi^0_2$ can be consistent with the assumption in the experimental searches ([*e.g.*]{} $\tilde\chi^0_2\to \tilde\chi^0_1 Z$ and $\tilde\chi^\pm_1\to \tilde\chi^0_1 W$), the latter especially if sleptons are heavier than EWinos. Conversely, for region 1B, gluino searches will clearly represent the major probe for LHC Run-2, since $m_{\tilde g}<1.2$ TeV and exclusion limits on the gluino mass are expected to greatly increase already with early 13 TeV data. The Higgs sector of these regions is characterised by a light pseudoscalar $a_1$, accompanied by a light scalar $h_1$, both with a dominant singlet component. The mass of $h_1$ lies between 30–70 GeV in region 1A and 70–90 GeV in region 1B. This leads, in region 1A, to the interesting possibility of exotic decays of the 125 GeV (SM like) Higgs state $h_2$ into a pair of light scalars ($h_2\to h_1 h_1$) as well as pseudoscalars ($h_2\to a_1 a_1$). Given the mass of the $a_1$ and of the $h_1$, possible interesting signatures to be explored are $h_2\to a_1 a_1\to 4 \tau/ 2\tau 2 \mu$ and $h_2\to h_1 h_1\to 4b/2b 2\tau$. The leptonic channels have already been investigated at LHC Run-1. The ATLAS collaboration has performed a search for a CP even Higgs boson produced through gluon fusion and decaying into a pair of pseudoscalars, and has set a limit on the production cross section times branching ratio into a $4\tau$ final state normalized to the SM Higgs boson cross section [@Aad:2015oqa]. This limit can be easily related to the $2\mu2\tau$ final state under the assumption that $Br(a_1\to\mu\mu)/Br(a_1\to\tau\tau)=m_\mu^2/(m_\tau^2\sqrt{1-(2m_\tau/m_{a_1}})^2)$. We show in Fig. \[fig:1AB\_ttmumu\] (left panel) the quantity $\sigma(gg\to h_2)/\sigma(gg\to h^{\rm SM}){\textrm{Br}}(h_2\to a_1 a_1){\textrm{Br}}^2(a_1\to \tau\tau)$ for region 1A and 1B, together with the limit set by the ATLAS collaboration. The rescaled production cross section has been computed using the reduced $ggh_2$ coupling provided by `NMSSMTools`, which is given with respect to a SM Higgs boson of the same mass. Given the small size of the deviations of the $h_2$ couplings with respect to the SM Higgs ($g_{ggh_2}/g^{\rm SM}_{ggh_2}\sim 1)$, and assuming $\sigma(gg\to h_2)_{\textrm SM}\sim 20$ pb for $m_h^{\rm SM}\sim 125$ GeV, this analysis sets a limit on the inclusive cross section times branching ratio of $\sim 2$ pb for $m_{a_1}=$ 5–10 GeV. While this analysis does not constrain at the moment region 1 of the nMSSM, prospects for LHC Run-2 are quite exciting, given an inclusive cross section that can reach the $\sim$ pb level in the $a_1$ considered mass range (right panel). A recently published CMS analysis [@Khachatryan:2015wka] also sets limits on the same production process, decaying however into a $4\mu$ final state. In the case where the only source of signal events is given by $h_2\to a_1 a_1\to 4\mu$, the limit is set to 1 fb for $m_{a_1} = 3.55$ GeV, which is well above the maximum rate obtained in the nMSSM ($\sim 0.01$ fb). ![$\sigma(gg\to h_2)/\sigma(gg\to h^{\rm SM}){\textrm{Br}}(h_2\to a_1 a_1){\textrm{Br}}^2(a_1\to \tau\tau)$ for the LHC Run-1 (left) and $\sigma(gg\to h_2){\textrm{Br}}(h_2\to a_1 a_1){\textrm{Br}}^2(a_1\to \tau\tau)$ for the LHC Run-2 (right) for region 1A and 1B. The horizontal line in the left plot is the limit set by ATLAS [@Aad:2015oqa].[]{data-label="fig:1AB_ttmumu"}](figures/1AB_tatamumu.pdf "fig:"){width="46.00000%"}![$\sigma(gg\to h_2)/\sigma(gg\to h^{\rm SM}){\textrm{Br}}(h_2\to a_1 a_1){\textrm{Br}}^2(a_1\to \tau\tau)$ for the LHC Run-1 (left) and $\sigma(gg\to h_2){\textrm{Br}}(h_2\to a_1 a_1){\textrm{Br}}^2(a_1\to \tau\tau)$ for the LHC Run-2 (right) for region 1A and 1B. The horizontal line in the left plot is the limit set by ATLAS [@Aad:2015oqa].[]{data-label="fig:1AB_ttmumu"}](figures/1AB_tatamumu_13.pdf "fig:"){width="46.00000%"} As mentioned, other interesting possibilities are the decays of the SM like Higgs boson into a pair of lighter CP even Higgs states ($h_2 \to h_1h_1$), which then decay into a $2b2\tau$ final state for which rates could reach $\mathcal{O}(250~{\rm fb})$ at the 13 TeV LHC. The decay into a 4$b$ final state has a larger rate, $\mathcal{O}(1~{\rm pb})$, but suffers from a higher QCD background [@Bomark:2014gya]. While our discussion has been carried out assuming dominant gluon fusion production of the SM like Higgs boson, the contribution of different production modes (vector boson fusion or Higgs Strahlung) can improve the signal over background ratio, despite smaller production cross sections. Finally, direct production of a light pseudoscalar $a_1$, either via gluon fusion or via $b\bar b$ associate production, can provide other possible signatures for the search of a light pseudoscalar [@Bomark:2014gya]. In region 1 the prospects for DM DD are quite pessimistic, due to the lightness of the LSP which has a mass below the lower limit that can be tested by the XENON1T experiment [@Aprile:2012zx]. The value of the (rescaled) spin independent DD cross section, being at most $\sim 10^{-10}$ pb, is also problematic. This is about 1 order of magnitude below the neutrino coherent scattering background for a WIMP mass of $\sim 5$ GeV [@Ruppin:2014bra]. We conclude this Section proposing in Tab. \[tab:1AB\_BMP\] three benchmark points in region 1A, with the Higgs sector signatures relevant for the 13 TeV run of the LHC discussed above. Tab. \[tab:1AB\_BMP\] reports also the relevant masses, cross sections (obtained with  `v1.5.0` [@Harlander:2012pb; @Liebler:2015bka] at NNLO) and branching ratios for the various benchmark points. These scenarios are amenable for a deeper phenomenological investigation, both from the theoretical and experimental side. The benchmark points BMP1A-I and BMP1A-II are in a configuration where $m_{h_2}>2 m_{h_1}$, and maximise Br$(h_2\to a_1 a_1)$ and Br$(h_2\to h_1 h_1)$ respectively, with the possibility of giving rise to the aforementioned multi-$\tau$ or multi-$b$ final states, although the branching ratios are only $\sim$4% for 4$b$ and $\sim$1% for $2b2\tau$. We also propose a third scenario, BMP1A-III, where the $h_2\to h_1 h_1$ channel is kinematically closed, and the decay rate of the SM like Higgs in two light pseudoscalars is below $10^{-3}$. The decay pattern of $h_1$ and $a_1$ in this configuration are similar to the ones of BMP1A-I and BMP1A-II, but clearly these states now need to be produced directly. We do not propose benchmark points for region 1B for two reasons. First, the characteristics of the Higgs sector are quite similar to the ones of region 1A when the decay pattern $h_2\to h_1 h_1$ is closed (recall in fact that in region 1B the $h_1$ mass is between 70 and 90 GeV). Secondly, and more importantly, the lighter gluino in this region is likely to be tested with early data from LHC Run-2 through conventional search channels. In summary, the Run-2 of the LHC at higher energy and luminosity will further probe region 1 through stop, slepton and gluino searches, while new Higgs decay channels involving light pseudoscalar and/or scalar bosons can provide characteristic signatures of an extended Higgs sector. Moreover, peculiar signatures from stop decays could also characterise the light singlino scenario. BMP1A-I BMP1A-II BMP1A-III ------------------------------ ----------------------------------------------- ----------------------------------------------- ----------------------------------------------- $\tb$ 7.8 8.32 9.48 $\l$ 0.372 0.4 0.509 $\mu$ 265 290 375 $m_0$ 0 0 505 $ M_{1/2}$ 766 790 888 $A_0$ -1146 -1050 9 $A_\l$ 2286 2700 5715 $\xi_F$ 28 48 79 $\xi_S$ -5.3$\cdot10^4$ -4.6$\cdot10^4$ 279 Masses $m_{h_1}=37.0$ $m_{a_1}=6.8$ $m_{h_1}=43.6$ $m_{a_1}=6.8$ $m_{h_1}=64.7$ $m_{a_1}=8.0$ $\sigma^{13~\rm TeV}$ \[pb\] $\sigma(gg\to h_2)=41.5$ $\sigma(gg\to h_2)=42.2$ $\sigma(gg\to h_2)=42.4$ $\sigma(gg\to h_1)=13.0$ $\sigma(gg\to h_1)=1.8$ $\sigma(gg\to h_1)=1.5$ $\sigma(gg\to a_1)=242.8$ $\sigma(gg\to a_1)=236.5$ $\sigma(gg\to a_1)=244.5$ Br($h_2$) Br($h_2\to a_1a_1$)=8% Br($h_2\to h_1h_1$)=10% Br($h_1$) Br($h_1\to b\bar b$)=85% Br($h_1\to b\bar b$)=65% Br($h_1\to b\bar b$)=58% Br($h_1\to \tau\tau)=$7% Br($h_1\to a_1a_1$)=28% Br($h_1\to a_1a_1)=$33% Br($h_1\to \tilde\chi^0_1\tilde\chi^0_1$)=7% Br($h_1\to \tau\tau$)=6% Br($h_1\to \tau\tau$)=5% Br($a_1$) Br($a_1\to \tilde\chi^0_1\tilde\chi^0_1$)=73% Br($a_1\to \tilde\chi^0_1\tilde\chi^0_1$)=73% Br($a_1\to \tilde\chi^0_1\tilde\chi^0_1$)=78% Br($a_1\to \tau\tau$)=25% Br($a_1\to \tau\tau$)=25% Br($a_1\to \tau\tau$)=20% : Benchmark points choices for region 1A. Dimensionful parameter are expressed in GeV, except $\xi_F$ and $\xi_S$ which are in GeV$^2$ and GeV$^3$ respectively and $\sigma$, expressed in pb. We indicate in the table the relevant mass spectrum, cross sections and branching ratios.[]{data-label="tab:1AB_BMP"} Region 2, $45 <m_{\tilde\chi^0_1}< 55$ GeV ------------------------------------------ This region is characterised by a large $m_0 \sim 10^5$ GeV and a small $M_{1/2} {\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}500$ GeV. The LSP is almost an equal admixture of higgsino and singlino, with a bino component $\lesssim$ 20%. The DM relic density is always below $\Omega h^2<$ 0.01, see Fig. \[fig:scan\_res\]. Due to the high value of $m_0$, all sfermions are heavy and decoupled and the only sub-TeV sparticles are EWinos and gluinos. The EWinos are rather mixed states, with a mass ${\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}200$ GeV, while the gluino lies in the 800–1500 GeV mass range. The lightest CP even Higgs state, $h_1$, is SM like, and all the other Higgs states are heavier than 300 GeV, with $h_2$ and $a_1$ carrying a dominant singlet component if lighter than 600 GeV (otherwise $h_3$ and $a_2$ are the dominant singlet states). To check the LHC Run-1 reach on the gluino sector, we have exploited the same technique used in Sec. \[sec:1AB\_constraints\], based on the recast of the CMS gluino search [@CMS:2013ija] implemented in 5, and obtained a similar limit of $m_{\tilde g}{\stackrel{\displaystyle >}{\,_{\! \,_{\displaystyle \sim}}}}1100$ GeV. Using  we did not find any additional constraint arising from the EWino sector. The most sensitive analysis, among the ones implemented in , is the CMS search for $\tilde\chi_2^0\tilde\chi^\pm_1$ production in $W^\pm Z + E_T^{\rm miss}$ final state with either on-shell or off-shell SM gauge bosons [@Khachatryan:2014qwa]. However, in this region of the nMSSM parameter space, the masses of the EWinos are approximately $m_{\tilde\chi^\pm_1}\sim$103–120 GeV and $m_{\tilde\chi^0_2}\sim$ 80–120 GeV. Their decays therefore occur via an off-shell $W$ or $Z$, falling close to the region where the CMS search looses sensitivity, $m_{\tilde\chi^0_2}=m_{\tilde\chi^0_1}+m_Z$. Another reason for this region not to be constrained by EWino searches is that there are other important decay channels for the neutralinos, which lead to a reduction of the branching into a Z boson. Notably, $\tilde\chi^0_2$ can decay into a $\gamma \tilde\chi^0_1$ final state, with a branching ratio up to $50\%$ for a mass splitting between $\tilde\chi^0_2$ and $\tilde\chi^0_1$ of $\sim 50$ GeV, as shown in Fig. \[fig:2\_n2decay\] (left panel). This opens up the interesting possibility of looking at $\tilde\chi^\pm_1\chi_2^0$ production, for which the cross section can be close to $\mathcal{O}(1~{\rm pb})$ already at LHC Run-1, in either the $2j+\gamma+E_T^{\rm miss}$ or $l+\gamma+E_T^{\rm miss}$ final states. Interestingly, the photon might have a sizeable $p_T$, as the $\tilde\chi^0_2-\tilde\chi^0_1$ mass difference can be $\mathcal{O}$(50 GeV). However such a mass splitting does not allow for large $E_T^{\rm miss}$, therefore it might be necessary to trigger on an ISR jet. We leave this point for further investigation and we provide one benchmark point for this scenario in Tab. \[tab:2\_BM\] (BMP2-I). ![Br$(\tilde\chi^0_2\to \gamma \tilde\chi^0_1)$ in function of $m_{\tilde\chi^0_2}-m_{\tilde\chi^0_1}$ (left) and $\sigma(gg\to h_2){\rm Br}(h_2\to h_1 h_1)$ in function of $m_{h_2}$ (right) for the 8 and 13 GeV LHC (cyan and orange points respectively). Both panels include only points in region 2 for which $m_{\tilde g}{\stackrel{\displaystyle >}{\,_{\! \,_{\displaystyle \sim}}}}1100$ GeV. The LSP mass is $\sim 45$–$55$ GeV.[]{data-label="fig:2_n2decay"}](figures/2_n2decay.pdf "fig:"){width="46.00000%"} ![Br$(\tilde\chi^0_2\to \gamma \tilde\chi^0_1)$ in function of $m_{\tilde\chi^0_2}-m_{\tilde\chi^0_1}$ (left) and $\sigma(gg\to h_2){\rm Br}(h_2\to h_1 h_1)$ in function of $m_{h_2}$ (right) for the 8 and 13 GeV LHC (cyan and orange points respectively). Both panels include only points in region 2 for which $m_{\tilde g}{\stackrel{\displaystyle >}{\,_{\! \,_{\displaystyle \sim}}}}1100$ GeV. The LSP mass is $\sim 45$–$55$ GeV.[]{data-label="fig:2_n2decay"}](figures/2_sigmah2h1h1.pdf "fig:"){width="47.80000%"} In this region the extra Higgs states $h_2$ and $a_1$ are either doublet or singlet like, depending on their masses, and almost degenerate. Only the states heavier than $\sim 450$ GeV survive LHC-Run 1 constraints, set by gluino searches. Possible decay patterns that have been searched for at the LHC are $a_1\to Z h_1$ and $h_2\to h_1h_1$ [@Khachatryan:2015lba; @Aad:2015wra; @Khachatryan:2014jya], which do not however impose any constraint on this region of the parameter space. This is due to the small Br$(a_1\to Z h_1)<0.5$% and to the fact that the mass of $h_2$ lies beyond the reach of the experimental analysis which is around 360 GeV. We show in Fig. \[fig:2\_n2decay\] (right panel) the rates $\sigma(gg\to h_2) {\rm Br}(h_2\to h_1 h_1)$ for the 8 and 13 TeV LHC (cyan and orange respectively), where the gluon fusion production cross sections have been obtained by rescaling the one provided by the LHC Higgs cross section working group [@Heinemeyer:2013tqa] with the reduced $ggh_2$ coupling provided by `NMSSMTools`. Recall that the current limit is $\sigma(gg\to h_2){\rm Br}(h_2\to h_1 h_1) {\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}4$ pb for $m_{h_2}=360$ GeV, thus roughly two orders of magnitude above the predicted value. ![$\sigma(gg\to h_2){\rm Br}(h_2\to t\bar t)$ (left) and $\sigma(gg\to h_2){\rm Br}(h_2 \to \tilde\chi^+_1\tilde\chi^-_1)$ (right) for the 8 and 13 TeV LHC (cyan and orange points respectively) as function of $m_{h_2}$.[]{data-label="fig:2_sigmah2ttchichi"}](figures/2_sigmah2tt.pdf "fig:"){width="46.00000%"}![$\sigma(gg\to h_2){\rm Br}(h_2\to t\bar t)$ (left) and $\sigma(gg\to h_2){\rm Br}(h_2 \to \tilde\chi^+_1\tilde\chi^-_1)$ (right) for the 8 and 13 TeV LHC (cyan and orange points respectively) as function of $m_{h_2}$.[]{data-label="fig:2_sigmah2ttchichi"}](figures/2_sigmah2x1x1.pdf "fig:"){width="48.00000%"} Given that the mass of the $h_2$ is above the $t\bar t$ threshold, of particular interest is its decays into a top pair. The cross section times branching ratio can reach $\mathcal{O}(100~{\rm fb})$ for the 13 TeV LHC, as illustrated in Fig. \[fig:2\_sigmah2ttchichi\] (left panel), and a similar value is predicted for $a_1$. This channel however deserves a deeper investigation, due to the well known possibility of dominant interference with the $t \bar t$ QCD background, for which dedicated cuts might need to be applied, see [*e.g.* ]{}[@Barger:2006hm]. In the MSSM, an analysis has shown that the LHC could find evidence of a Higgs boson in this channel for masses up to 1 TeV and low values of $\tb$ [@Djouadi:2015jea]. The possibility of $h_2$ decaying into a pair of EWinos is also interesting, although it weakens the discovery potential in the $t\bar{t}$ mode. We show as an illustrative example the rates for the process $gg\to h_2\to \tilde\chi^+_1\tilde\chi^-_1$ in Fig. \[fig:2\_sigmah2ttchichi\] (right panel) for the 8 and 13 TeV LHC. For the latter the cross section times branching can reach $\mathcal{O}(100~{\rm fb})$ for a $\sim 500$ GeV $h_2$, giving the possibility of a different search channel for $h_2$ and/or $a_1$. Clearly, for heavier Higgs states the rate drops rapidly. We provide in Tab. \[tab:2\_BM\] a benchmark point (BMP2-II) that simultaneously maximises the two aforementioned rates. Note that for this point the branching ratio into neutralinos is similar to the one into charginos, while the pure invisible decay has a 6% branching ratio. Rates for the lightest pseudoscalar $a_1$ are somewhat similar. Note that such large branching ratios in EWinos can also be found in the MSSM [@Belanger:2015vwa]. ![ID detection rate into $b \bar b$ final state (left) and rescaled spin independent cross section for DD (right) in function of $m_{\tilde\chi^0_1}$. The horizontal lines represent the current limit from FermiLAT (left) and LUX (right). Also shown in the right plot is the projection for XENON1T experiment. Only the points surviving the LHC Run-1 constraints are shown.[]{data-label="fig:2_DD_ID"}](figures/2_ID.pdf "fig:"){width="49.00000%"}![ID detection rate into $b \bar b$ final state (left) and rescaled spin independent cross section for DD (right) in function of $m_{\tilde\chi^0_1}$. The horizontal lines represent the current limit from FermiLAT (left) and LUX (right). Also shown in the right plot is the projection for XENON1T experiment. Only the points surviving the LHC Run-1 constraints are shown.[]{data-label="fig:2_DD_ID"}](figures/2_DD.pdf "fig:"){width="46.50000%"} Finally, unlike the first region, ID experiments do not set any additional constraints, since the rescaled rate into $b\bar b$ final state (the most relevant for the mass range of interest) is much below the current constraint from Fermi-LAT, as shown in Fig. \[fig:2\_DD\_ID\] (left panel). Prospects for DD measurements are once again quite pessimistic, as illustrated in the right panel of Fig. \[fig:2\_DD\_ID\]. The rescaled spin independent cross section for DD is more than one order of magnitude below the current limit set by the LUX experiment, and XENON1T will be able to marginally cover only a small region with $m_{\tilde\chi^0_1}> 50$ GeV. Such bleak DD prospects are typical of models where one relies on a resonance to achieve the correct relic density. BMP2-I BMP2-II ----------------- ------------------------------------------------------------------------ -------------------------------------------------------------- $\tb$ 1.7 1.82 $\l$ 0.73 0.72 $\mu$ 130 126.0 $m_0$ 2.34$\cdot10^5$ 5.3$\cdot10^4$ $ M_{1/2}$ 269 326 $A_0$ -5879 -3180 $A_\l$ 0 0 $\xi_F$ 5.46$\cdot10^5$ 1.52$\cdot10^5$ $\xi_S$ -301 -2971 Masses $m_{\tilde\chi^\pm_1}=105$, $m_{\tilde\chi^0_2}=102$ $m_{h_2}=478$, $m_{\tilde\chi^\pm_1}=104$ $\sigma$ \[fb\] $\sigma^{13 (8)~\rm TeV}_{\tilde\chi^\pm_1\tilde\chi^0_2}=$ 1729 (860) $\sigma^{13~\rm TeV}_{gg\to h_2}=$524 Br Br($\tilde\chi^0_2\to \tilde\chi^0_1\gamma$)=50% Br($h_2\to t\bar t$)=15% Br($h_2\to \tilde\chi^+_1\tilde\chi^-_1$)=29% Br($h_2\to \tilde\chi^0_{i\ne0} \tilde\chi^0_{j\ne 0})=$ 29% Br($h_2\to \tilde\chi^0_1 \tilde\chi^0_1)=$ 6% Br($\tilde\chi^\pm_1\to\tilde\chi^0_1 W^*$)=100% : Benchmark points choices for region 2. Dimensionful parameter are expressed in GeV, except $\xi_F$ and $\xi_S$ which are in GeV$^2$ and GeV$^3$ respectively and $\sigma$, expressed in fb. We indicate in the table the relevant mass spectrum, cross sections and branching ratios.[]{data-label="tab:2_BM"} Region 3, $m_{\tilde\chi^0_1}\sim 65$ GeV ----------------------------------------- The $\sim 65$ GeV LSP can appear in different regions of the $m_0$-$ M_{1/2}$ parameter space, in which it will have a different composition. We have identified the following three regions: - Region 3A, large $m_0$ ($\sim~10^4$ GeV) and small $ M_{1/2}$ (${\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}200$ GeV). In this region the LSP is mainly bino and the DM relic density can be compatible with the value measured by Planck, see Fig. \[fig:scan\_res\]. - Region 3B, $m_0\sim 1$ TeV and $ M_{1/2} {\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}500$ GeV. In this region the LSP is a mixed higgsino and singlino state. - Region 3C, small $m_0$ (${\stackrel{\displaystyle <}{\,_{\! \,_{\displaystyle \sim}}}}1$ TeV) and large $ M_{1/2}$ ($\sim$ 10$^4$ GeV). Here the LSP is also a higgsino/singlino admixture. ### Region 3A and 3B: bino or higgsino/singlino LSP with low $ M_{1/2}$ In region 3A and 3B the relic density is kept to a small value by annihilation of the LSP through the Z boson or the SM like Higgs boson $h_1$. The very small higgsino component of the LSP in region 3A (around 2%) is just sufficient to provide a DM relic density $\Omega h^2\approx 0.1$. The heavier CP even state, $h_2$, has a mass between 250 and 4000 GeV, almost degenerate with $a_1$, in both regions. Considering the values of $m_0$ and $ M_{1/2}$, the SUSY spectrum is characterised by light EWinos and light gluinos. Other sfermions are above the TeV scale and decoupled in region 3A while squarks, especially stops, can be below the TeV scale in region 3B. In region 3A all the EWinos are lighter than 200 GeV, and the gluino is always lighter than 600 GeV. This region is therefore already completely ruled out by the LHC Run-1 searches for gluinos. Indeed, with the same procedure as in Sec. \[sec:1AB\_constraints\], we have checked that gluino masses $\lesssim 1100$ GeV are excluded. Recall that in a model with gaugino mass unification at the GUT scale, the gluino and bino masses are respectively $m_{\tilde g}\approx 3 M_{1/2}$ and $m_{\tilde B}\approx 1/6 M_{1/2}$, thus the upper limit on the (bino) LSP mass set by the upper limit on the singlino mass from Eq. (\[eq:singlino\]), entails an upper limit on the gluino mass. This has the important consequence of leaving just region 1, the one with the very light singlino LSP, as an nMSSM explanation for the entire relic abundance of the universe, together with a heavy enough gluino to pass LHC Run-1 constraints. ![Spin independent cross section for DD as function of $m_{h_1}-2m_{\tilde\chi^0_1}$, for all the points $m_{\tilde\chi^0_1}\approx 60 {\rm GeV}$ The horizontal lines represent the current limit from LUX and the projected limits for the XENON1T experiment.[]{data-label="fig:3A_DD"}](figures/3AB_DD.pdf){width="46.00000%"} BMP3B-I ----------------- -------------------------------------------------- $\tb$ 1.96 $\l$ 0.688 $\mu$ -105 $m_0$ 260 $ M_{1/2}$ 529 $A_0$ 3042 $A_\l$ 1489 $\xi_F$ 1.66$\cdot10^6$ $\xi_S$ 3.82$\cdot10^8$ Masses $m_{h_2}=276$, $m_{\tilde\chi^\pm_1}=115$ $\sigma$ \[fb\] $\sigma^{13~\rm TeV}_{gg\to h_2}=$ 180 Br Br($h_2\to\tilde\chi^+_1\tilde\chi^-_1$)=51% Br($\tilde\chi^\pm_1\to\tilde\chi^0_1 W^*$)=100% : Benchmark points choices for region 3B. Dimensionful parameter are expressed in GeV, except $\xi_F$ and $\xi_S$ which are in GeV$^2$ and GeV$^3$ respectively and $\sigma$, expressed in fb. We indicate in the table the relevant mass spectrum, cross sections and branching ratios.[]{data-label="tab:3_BM"} The same bound on the gluino mass has also a strong impact in region 3B. Here the value of $ M_{1/2}$ is slightly higher than in region 3A, thus the gluino is somewhat heavier, $600 <m_{\tilde g}< 1200$ GeV. Nevertheless a large part of this region is already excluded by LHC Run-1. In addition, all EWinos except the heavier neutral and charged states, have a mass below 200 GeV, and are slightly constrained by direct EWinos searches, though the main constraint remains the one on the gluino mass. Moreover the lightest stop lies in the 200–500 GeV range because of the smaller value of $m_0$. In principle, for a 60 GeV LSP such a light stops could be excluded. However this is not the case here for two reasons. The first is that stops with a mass $\sim 250$ GeV are close to the kinematic edge $m_{\tilde t}-m_{\tilde\chi^0_1}=m_Z$ (see for example refs. [@Chatrchyan:2013xna; @Aad:2014kra]) where standard stop searches loose sensitivity, causing therefore a poor exclusion limit. The second is that for higher values of the stop mass, around 400 GeV, the same analyses loose sensitivity since the stop decays do not fulfill the simplified model assumptions. In particular, the branching ratios Br($\tilde t\to\tilde\chi^\pm_1 b)$ and Br($\tilde t\to\tilde\chi^0_1 t)$ are both suppressed. The reason is that the coupling of the stop to the singlino is small, hence other decay channels into heavier neutralinos are favoured despite the reduced phase space, causing therefore a reduction of the exclusion power of the experimental analyses. The two previous statements about the non exclusion of such light stops have been checked via the CMS search [@Chatrchyan:2013xna] available in the 5 Public Analysis Database [@MA5:CMS-SUS-13-011]. The reinterpretation of the LHC Run-1 results on gluino searches leave therefore just a small window of parameter space available, namely for $m_{\tilde g}{\stackrel{\displaystyle >}{\,_{\! \,_{\displaystyle \sim}}}}1100$ GeV. While extra Higgs searches such as $a_1\to Z h_1$ and $h_2\to h_1h_1$ do not set any constraint on this region of parameter space [^5], exotic decays of heavy Higgs states in a pair of EWinos can have substantial rates. This leads again to the interesting possibility of searching for these states through their EWino decay channels. Cross sections of $\mathcal{O}$(100 fb) are expected for $gg\to h_2, h_2\to \tilde\chi^+_1\tilde\chi^-_1$ at the 13 TeV LHC, with a $h_2$ $\sim 300$ GeV. We provide one such benchmark in Tab. \[tab:3\_BM\] (BMP3-I). Note that despite of a $h_2$ mass below 300 GeV, this point is safely below the recent limits of CMS in the $WW$ channel [@Aad:2015agg]. As concerns ID measurements, the relevant rates are well below the limits set by the Fermi-LAT experiment. Prospects for future underground experiments for DD are quite interesting. The SI cross section lies just below the limits set by the LUX experiment and is roughly constant in this region. It is determined essentially by the $h_1$ coupling to the LSP which does not vary much in this region. The rescaled cross section is however suppressed by up to almost two orders of magnitude, since the thermally averaged DM annihilation cross section is strongly enhanced when $m_{h_1}-2 m_{\tilde\chi_1^0}$ is $\sim 1 $ GeV, see Fig. \[fig:3A\_DD\]. Nevertheless, as shown in this figure, XENON1T will be able to probe all this region of parameter space, offering a valid complement to collider based analyses. ### Region 3C: higgsino/singlino LSP with high $ M_{1/2}$ This region at high $ M_{1/2}$ and low $m_0$ is characterised by light sleptons and light higgsinos and singlino, while the bino and wino states are heavy and decoupled. Squarks and gluinos are also heavy and decoupled, and the Higgs states $h_2$ and $a_1$ have a mass greater than 600 GeV. Note that even though $m_0$ is small at the GUT scale, the squarks are heavy because the renormalisation group equations that are used to derive their mass at the SUSY scale receive important contributions from $ M_{1/2}$ which is large. Despite the light EWinos, $m_{\tilde\chi^0_2,\tilde\chi^\pm_1}\sim 150$ GeV, sleptons, $m_{e_R/\mu_R}\sim 200$ GeV, and $\tilde \tau_1$ as light 100 GeV, no constraints are obtained by the LHC Run-1 SUSY searches. First and second generation right handed charged sleptons with a mass 100–300 GeV for a $\sim 65$ GeV LSP lie on the edge of the exclusion set by the ATLAS slepton search with $2l+E_T^{\rm miss}$ final state [@Aad:2014vma]. The limit set by ATLAS on the $\tilde \tau^+\tilde \tau^-$ production cross section varies between $0.76$ pb for $m_{\tilde \tau} = 100$ GeV and $0.02$ pb for $m_{\tilde \tau}= 300$ GeV for a 60 GeV LSP [@Aad:2014yka]. This limit is roughly one order of magnitude larger than the predicted values, therefore making this search not sensitive to this scenario. Since the EWinos are mostly higgsino or singlino, their production cross sections are small. Moreover, the $\tilde\chi^0_{1,2}$ and $\tilde\chi^+_1$ masses lie close to the kinematic edge for off-shell decay where standard searches loose sensitivity, making it therefore easy to escape the LHC limits. We expect however that upgrade of the standard simplified model searches will be able to cover this region of parameter space with early LHC Run-2 data given the standard decay modes of these particles, and the fact that in this region  gives a ratio between the theoretical prediction and the experimental measurement which reaches the value of 0.7, [*i.e.*]{} not too far from the value of 1, for which we claim the point to be excluded. Finally, as above, ton-scale underground experiments for DD, such as XENON1T, will be able to fully cover this region of parameter space, see Fig. \[fig:stau\_xs\], while no constraints are actually enforced by ID measurements. ![ Spin independent cross section for DD as function of $m_{h_1}-2m_{\tilde\chi^0_1}$. The horizontal lines represent the current limit from LUX and the projected limits for the XENON1T experiment.[]{data-label="fig:stau_xs"}](figures/3C_DD.pdf){width="46.00000%"} Conclusions {#sec:sec5} =========== In this paper we have explored the parameter space of the nMSSM with input parameters defined at the GUT scale, and identified three broad regions for the LSP mass which satisfy theoretical, astrophysical, cosmological and collider constraints. In particular we have considered collider constraints on the Higgs boson and sparticles, as well as flavour constraints available in , upper limit on DM DD from LUX and relic abundance of DM as measured by Planck. The nMSSM is characterised by the presence of a singlino with a mass below 75 GeV, since only mixing with higgsinos contribute to its, otherwise vanishing, mass. In only two of the three regions the lightest neutralino LSP can account completely for the DM abundance. The first features a light singlino LSP ($m_{\tilde\chi^0_1}\sim 5$ GeV) the second a heavier one ($m_{\tilde\chi^0_1}\sim 65$ GeV), mostly bino. In all the other allowed regions the computed relic density lies below the value measured by Planck. A closer scrutiny at the constraints arising from SUSY searches at the LHC, implemented via two public tools for analyses recast,  and 5, showed that some of these regions (or sub-regions) were actually completely exclude by results of LHC Run-1, notably when they featured a gluino below the TeV scale. In particular the only valid region not requiring an additional DM component is the one with a light singlino. We have investigated how the model could be probed at the LHC Run-2 or through non collider DM searches. In particular, the region with a light singlino is characterised by the presence of a light pseudoscalar with a mass which is twice the one of the singlino, and a scalar Higgs with a mass below that of the recently discovered scalar. Hence, searches for these extra light states could lead to a distinctive signal of the nMSSM. Relevant signatures include the decay of the 125 GeV Higgs into light scalars or pseudoscalars, leading for example to $4b,2b/2\tau$ or $4\tau$ final states, or the direct production of one of the light scalar or pseudoscalar. Some of these signatures are under investigation at LHC, since the possibility of a light CP odd Higgs is one of the hallmark signature of generic NMSSM, and searches for light Higgs bosons are a priority in the LHC Run-2 program. In other regions where the lightest Higgs is the already discovered boson at 125 GeV, the model can be probed via searches for the heavy Higgs states. One characteristic of the model, as opposed to the MSSM but shared with the generic NMSSM, is that values of $\tb\approx 1$–$2$ can be compatible with all the present constraints. Thus, searches for heavy Higgs states decaying into top pairs will provide further probes of the model. Another characteristic feature is the possibility of a heavy Higgs state decaying with a high rate into a pair of charginos or neutralinos. In particular we have found large rates even with a Higgs mass below 300 GeV. While a large branching ratio into charginos for the heavy Higgs state can occur in the MSSM, the Higgs mass is typically much higher [@Belanger:2015vwa], hence this process suffers from suppression due to the somewhat smaller cross section. In the nMSSM the decays of sparticles have also peculiar features, thus searches for SUSY in Run-2 of the LHC could provide not only a validation of the theoretical SUSY framework, but also a sign of its non-minimality. The most distinctive feature occurs once again in the case of a light singlino LSP. Since the singlino is very weakly coupled to the sfermions, the decay channels involving the singlino in the final state typically have small partial widths. Therefore decays of sfermions, and in particular squarks, proceed preferably through the heavy neutralinos rather than directly in the LSP. Moreover the heavier neutralinos and charginos can decay into the LSP and a gauge boson (always permitted by phase space) and into a LSP and a light scalar or pseudoscalar. When the singlino LSP is not so light, we have found another unusual signature corresponding to $\tilde\chi_2^0$, which can decay with a substantial rate into a $\tilde\chi^0_1\gamma$ final state even when there is a considerable mass difference between these two states. Of course conventional SUSY searches at LHC Run-2 will probe further the allowed parameter space of the model, notably through searches for squarks and gluinos but also EWinos and sleptons. We have also pointed out that DD searches are complementary to collider searches, and that in particular a ton-scale detector could completely probe the remaining allowed region with a LSP around 65 GeV, while large signals in ID are not expected, except in very specific kinematic configurations. A thorough investigation of these new signatures at the LHC, which lies beyond the scope of this work, is required to assess the potential of the LHC to discover and/or probe the nMSSM. Acknowledgments {#acknowledgments .unnumbered} =============== DB thanks Ursula Laa for useful discussions regarding the use of . This work was supported in part by the LIA-TCAP of CNRS, by the French ANR, Project DMAstro-LHC, ANR-12-BS05-0006, by the [*Investissements d’avenir*]{}, Labex ENIGMASS. The work of AP was also supported by the Russian foundation for Basic Research, grant RFBR-15-52-16021-CNRS-a. The authors acknowledge the support of France Grilles for providing cloud computing resources on the French National Grid Infrastructure. [^1]: For a review, see refs. [@Maniatis:2009re; @Ellwanger:2009dp]. [^2]: For a review of NMSSM Higgs mass calculations in public codes (including ), see ref. [@Staub:2015aea]. [^3]: The value of the top quark pole mass has been fixed to $m_{\rm top} = 173.1$ GeV. [^4]: See <http://www.th.u-psud.fr/NMHDECAY/nmssmtools.html> for a detailed list of the implemented constraints. [^5]: Note that the main search channel at low values of $\tb$ is in gauge bosons and that the constraints on the SM like Higgs imply a suppressed coupling of the heavy Higgs to gauge bosons, as in the MSSM. The suppression is even stronger for the singlet Higgs.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Space-filling designs are popular choices for computer experiments. A sliced design is a design that can be partitioned into several subdesigns. We propose a new type of sliced space-filling design called sliced rotated sphere packing designs. Their full designs and subdesigns are rotated sphere packing designs. They are constructed by rescaling, rotating, translating and extracting the points from a sliced lattice. We provide two fast algorithms to generate such designs. Furthermore, we propose a strategy to use sliced rotated sphere packing designs adaptively. Under this strategy, initial runs are uniformly distributed in the design space, follow-up runs are added by incorporating information gained from initial runs, and the combined design is space-filling for any local region. Examples are given to illustrate its potential application.' author: - | Xu He[^1]\ Academy of Mathematics and System Sciences,\ Chinese Academy of Sciences bibliography: - 'SlicedDesigns.bib' title: '**Sliced rotated sphere packing designs**' --- \#1 0 [0]{} 1 [0]{} [**Sliced rotated sphere packing designs**]{} [*Keywords:*]{} Design of experiment; Expected improvement; Maximin distance; Nested design; Sequential design. Introduction {#sec:intro} ============ Space-filling designs whose points are uniformly scattered in the design space are popular choices for computer experiments [@Santner:book; @Sacks:1989]. In this work, we consider designs which contain points in $[0,1]^p$. The separation distance of a design $\mathbf{D}$ is $$\label{eqn:dp} \min_{\mathbf{x}_1,\mathbf{x}_2\in \mathbf{D}, \mathbf{x}_1\neq \mathbf{x}_2} \left( \|\mathbf{x}_1-\mathbf{x}_2\|_2 \right),$$ and the fill distance of a design $\mathbf{D}$ is $$\label{eqn:df} \sup_{\mathbf{z}\in [0,1]^p}\left\{ \min_{\mathbf{x}\in \mathbf{D}}(\|\mathbf{z}-\mathbf{x}\|_2) \right\}.$$ As discussed in @Johnson:1990 and @Haaland:2017, designs with high separation distance or low fill distance have some optimal or asymptotically optimal characteristics that are desirable for computer experiments. Many space-filling designs are Latin hypercube designs, which achieve optimal one-dimensional projective uniformity [@McKay:1979]. Maximin distance Latin hypercube designs [@Morris:1995], which are generated by numerically maximizing the separation distance within the class of Latin hypercube design, are popular among space-filling designs. Lattice-based designs are another type of space-filling designs [@Heitmann:2016; @RSPD; @ILmMD]. A lattice is the collection of infinitely many orthogonal or nonorthogonal grid points, and a lattice-based design consists of the lattice points that are located in the design space. Lattice-based designs have identical structure at any local area and are therefore space-filling globally and locally. In particular, rotated sphere packing designs are constructed by rescaling, rotating and translating the lattice that has asymptotically lowest fill distance [@RSPD]. If a space-filling design can be partitioned into several space-filling subdesigns, we call the full design, together with the slicing rule, a sliced space-filling design [@Qian:2009:SSFD]. @Qian:2012 proposed sliced Latin hypercube designs whose full designs and subdesigns are Latin hypercube designs. Later, @Ba:2015 proposed optimal sliced Latin hypercube designs whose full designs and subdesigns are Latin hypercube designs with high separation distances. For illustration, an optimal sliced Latin hypercube design in two dimensions is presented in Figure \[fig:intro\](a). Other types of sliced space-filling designs are proposed by @Qian:2009:SSFD [@Yang:2013; @Ai:2014; @Huang:2014; @Sun:2014; @Xie:2014; @Deng:2015; @Liu:2015; @SOALH], among others. These designs are useful for computer experiments with quantitative and qualitative variables [@Qian:Wu:Wu:2008; @Deng:2016], computer experiments with multiple levels of accuracy [@Qian:Wu:2008], and model validation [@Zhang:2013]. ![Two sliced space-filling designs in two dimensions. Different slices are depicted by different symbols. \[fig:intro\]](introfigS2.eps){width="13.6cm"} In this paper, we propose a new class of sliced space-filling design called sliced rotated sphere packing designs. Their full designs and subdesigns are rotated sphere packing designs. Sliced rotated sphere packing designs are constructed based on sliced lattices, which are lattices that can be partitioned into several sublattices. An example of sliced rotated sphere packing design in two dimensions is displayed in Fig. \[fig:intro\](b). We provide two algorithms to construct sliced rotated sphere packing designs. The first algorithm partitions an ordinary rotated sphere packing design and the second algorithm enlarges an ordinary rotated sphere packing design. Both algorithms are simple without any numerical steps. The full design and subdesigns of a sliced rotated sphere packing design achieve the same degree of uniformity as ordinary rotated sphere packing designs that are based on the same type of lattice. While any type of lattice can be sliced, we provide a space-filling type of sliced lattice, based on which sliced rotated sphere packing designs achieve better separation distances by (\[eqn:dp\]) than optimal sliced Latin hypercube designs for low-dimensional cases. Sliced rotated sphere packing designs are also useful as sequential or adaptive designs. Usually, computer experiments are carried out sequentially. Many sequential space-filling designs distribute points uniformly in the design space for both initial and follow-up runs  [@Qian:2009; @Qian:Ai:2010; @NOALH; @Xu:2015; @Kong:2016]. However, as we gain more knowledge on the input-output system, we may find some regions more interesting for further investigation than others. This requires adaptive designs that can incorporate information gained from completed computer runs. One such example is sequential minimum energy designs [@Joseph:energy] whose points are representative of a probability density. By assigning higher density for more interesting areas, we can obtain nonuniform designs that focus on critical areas. Such designs are adaptive if the density function is set based on information gained from completed computer runs. Similar to the idea of sequential minimum energy designs, we propose a strategy to use sliced rotated sphere packing designs adaptively. Under this strategy, design points are generated with two densities, a higher density for more interesting regions and a lower density for the remaining space, with the high density region chosen adaptively based on information gained from initial runs that are uniformly distributed in the design space. Unlike most adaptive design methods that search for optimal next-point-to-add from the whole design space, we propose to search over a short list of candidate follow-up runs. Because of separation distance properties among all initial and candidate follow-up runs, the generated points are space-filling globally for low density regions and locally for high density regions. Furthermore, adjacent points of sliced rotated sphere packing designs are connected with special rules which may simplify the definition of high density region. As a result, this strategy is useful for emulation of nonstationary computer experiments and optimization of computer experiments. The rest of the paper is organized as follows: Section \[sec:lattice\] gives preliminary mathematical results on lattices and sliced lattices. In Section \[sec:const\], we give the algorithms to construct sliced rotated sphere packing designs. In Section \[sec:simu\], we compare sliced rotated sphere packing designs with other classes of sliced designs numerically. In Section \[sec:adaptive\], we give the strategy to use sliced rotated sphere packing designs adaptively and show its applications. Conclusions and discussion are provided in Section \[sec:conclusion\]. Proofs are given in the appendix. Lattices and sliced lattices {#sec:lattice} ============================ In this section, we give necessary definitions and results of lattices and sliced lattices. A set of points in $\mathbb{R}^p$ is called a lattice if it forms a group. The lattice points consist of linear combinations of $p$ basis vectors with integer coefficients. We call a $p\times p$ matrix a generator matrix of the lattice if its rows are the basis vectors. As an example, the set of integer vectors, $\mathbb{Z}^p$, is called the $p$-dimensional integer lattice, which can be generated from the $p$-dimensional identity matrix. Two important properties of lattices are their densities and thicknesses. If we place identical balls in $\mathbb{R}^p$ centered at the lattice points, then the maximum radius of the balls such that no two balls overlap is called the packing radius of the lattice, and the minimum radius of the balls such that the union of overlapped balls cover $\mathbb{R}^p$ is called the covering radius of the lattice. The Voronoi cell of a point $\mathbf{x}_0$ in a lattice $\mathbf{L}$ is the region $$\text{Vor}(\mathbf{x}_0) = \{ \mathbf{z}: |\mathbf{z}-\mathbf{x}_0| \leq |\mathbf{z}-\mathbf{x}|, \mbox{ for any } \mathbf{x} \in \mathbf{L} \}.$$ The density and thickness of a lattice is the volume of one ball with packing and covering radius, respectively, divided by the volume of one Voronoi cell. Lattice-based designs with highest possible density and lowest possible thickness have asymptotically optimal separation distance and fill distance, respectively. In this paper, we focus on two types of lattices, $A_p$ and $A_p^*$. The $A_p$ is called the $p$-dimensional zero-sum root lattice, with one possible generator matrix $$\begin{aligned} \label{eqn:A_p} \mathbf{M}_p &=& \frac{\sqrt{2}}{2} \mathbf{I}_p - \frac{\sqrt{p+1}+1}{\sqrt{2}p} \mathbf{J}_p, \end{aligned}$$ where $\mathbf{I}_p$ is the $p\times p$ identity matrix and $\mathbf{J}_p$ is the $p\times p$ matrix with all entries being one. The $A_p^*$ is called the dual of the $p$-dimensional zero-sum root lattice, with one generator matrix $$\begin{aligned} \label{eqn:A_p^*} \mathbf{M}_p^* &=&\frac{\sqrt{p+1}}{\sqrt{p}} \mathbf{I}_p - \frac{1}{\sqrt{p}(\sqrt{p+1}-1)} \mathbf{J}_p. \end{aligned}$$ The $A_p$ and $A_p^*$ are equivalent when $p=2$. Their densities and thicknesses for $2\leq p\leq 10$ are given in Tables \[tab:density\] and \[tab:thickness\], respectively. The $A_p^*$ has the best known thickness for $2\leq p\leq 22$ and the $A_p$ has the best known density for $p=2$ and $3$. @RSPD recommended to use $A_p^*$ for constructing rotated sphere packing designs, but as can be seen from the table, both lattices are substantially more space-filling than $\mathbb{Z}^p$. For a comprehensive review of lattices, see @Conway:1998 or @Zong:1999. $p$ 2 3 4 5 6 7 8 9 10 ---------------- ------- ------- ------- ------- ------- ------- ------- ------- ------- $A_p$ 0.907 0.740 0.552 0.380 0.244 0.148 0.085 0.046 0.024 $A_p^*$ 0.907 0.680 0.441 0.255 0.135 0.065 0.030 0.013 0.005 $\mathbb{Z}^p$ 0.785 0.524 0.308 0.164 0.081 0.037 0.016 0.006 0.002 : Density of three lattices, $A_p$, $A_p^*$ and $\mathbb{Z}^p$ \[tab:density\] $p$ 2 3 4 5 6 7 8 9 10 ---------------- ------ ------ ------ ------ ------ ------ ------ ------- ------- $A_p$ 1.21 2.09 3.18 5.92 9.84 18.9 33.0 64.4 116.0 $A_p^*$ 1.21 1.46 1.77 2.12 2.55 3.06 3.67 4.39 5.25 $\mathbb{Z}^p$ 1.57 2.72 4.93 9.20 17.4 33.5 64.9 126.8 249.0 : Thickness of three lattices, $A_p$, $A_p^*$ and $\mathbb{Z}^p$ \[tab:thickness\] Next, we give our definition and some theoretical results for sliced lattices. Suppose $\mathbf{K}$ is a proper subgroup of a lattice $\mathbf{L}$. For any $\mathbf{q} \in \mathbf{L}$, the set $\mathbf{K} \oplus \mathbf{q} = \{ \mathbf{k}+\mathbf{q} : \mathbf{k}\in \mathbf{K}\}$ is called a coset of $\mathbf{K}$. If $\mathbf{Q}$ is a finite subset of $\mathbf{L}$ and any coset of $\mathbf{K}$ can be uniquely expressed by $\mathbf{K} \oplus \mathbf{q}$ with a $\mathbf{q} \in \mathbf{Q}$, then the cosets of $\mathbf{K}$ partition $\mathbf{L}$ and we call $(\mathbf{L},\mathbf{K},\mathbf{Q})$ a sliced lattice. It is not hard to see that any lattice $\mathbf{L}$ can be partitioned into sublattices. Suppose $\mathbf{G}$ is a generator matrix of $\mathbf{L}$ and $z$ is an integer greater than one. Let $\mathbf{K}$ denote the lattice generated from $z \mathbf{G}$, then $(\mathbf{L},\mathbf{K},\{0,\ldots,z-1\}^p)$ is a sliced lattice with $z^p$ slices. Despite its simplicity, $(\mathbf{L},\mathbf{K},\{0,\ldots,z-1\}^p)$ is not useful for medium to large $p$ due to its large number of slices. A more practical type of sliced lattice is $(A_p^*,A_p,\mathbf{B})$, where $\mathbf{B}=\{\mathbf{u}_0,\ldots,\mathbf{u}_p\}$ and $\mathbf{u}_j$ is the $p$-vector with the first $j$ elements being one and other elements being zero. Here $\mathbf{u}_0$ is the zero vector. Proposition \[prp:slice\] below shows that $(A_p^*,A_p,\mathbf{B})$ is a sliced lattice with $p+1$ slices. \[prp:slice\] Let $\mathbf{L}$ and $\mathbf{K}$ be the lattices generated by $\mathbf{M}_p^*$ in (\[eqn:A\_p\^\*\]) and $\{2(p+1)\}^{1/2}\mathbf{M}_p$ in (\[eqn:A\_p\]), respectively. Suppose $\mathbf{a}^T \mathbf{M}_p^* \in \mathbf{L}$ where $\mathbf{a}=(a_1,\ldots,a_p)^T$ is an integer vector with $\sum a_i \mod (p+1) =z$, then $\mathbf{a}^T \mathbf{M}_p^* \in \mathbf{K} \oplus \mathbf{u}_z$. Furthermore, $\{ \mathbf{K} \oplus \mathbf{u}_0, \ldots, \mathbf{K} \oplus \mathbf{u}_p \}$ is a partition of $\mathbf{L}$. For a sliced lattice $(\mathbf{L},\mathbf{K},\mathbf{Q})$, think of $\mathbf{L}$ as being enlarged from $\mathbf{K}$. We call points in $\mathbf{K}$ “adult” points and the remaining points of $\mathbf{L}$ “baby” points. For a given baby point, the adult points nearest to it are called the “parents” of the baby point and the baby point is called a “child” of its parents. As shall be shown in Section \[sec:adaptive\], the parent-child relation is useful for adaptive designs. Proposition \[prp:parent\] below gives the parent-child relation of the $(A_p^*,A_p,\mathbf{B})$ sliced lattice. \[prp:parent\] (i) Suppose $(\mathbf{L},\mathbf{K},\mathbf{B})$ is a sliced lattice where $\mathbf{L}$ and $\mathbf{K}$ are generated by $\mathbf{M}_p^*$ in (\[eqn:A\_p\^\*\]) and $\{2(p+1)\}^{1/2}\mathbf{M}_p$ in (\[eqn:A\_p\]), respectively. For any baby point $(b_1,\ldots,b_p) \mathbf{M}_p^* \in \mathbf{L}\setminus \mathbf{K}$ with $\sum b_i \mod (p+1) =z$, its parents are $\{(b_1,\ldots,b_p) - (c_1,\ldots,c_p)\} \mathbf{M}_p^* $, of which either $( c_i\in\{0,1\}, i=1,\ldots p, \sum c_i = z )$ or $( c_i \in \{0,-1\}, i=1,\ldots p, \sum c_i = z-(p+1) )$. \(ii) Conversely, any adult point $(a_1,\ldots,a_p) \mathbf{M}_p^* \in \mathbf{K}$ has $2^{p+1}-2$ children, which can be written as $\{(a_1,\ldots,a_p) - (c_1,\ldots,c_p)\} \mathbf{M}_p^*$ with either $( c_i \in \{0,1\}, i=1,\ldots p, (c_1,\ldots,c_p)\neq 0 )$ or $( c_i \in \{0,-1\}, i=1,\ldots p, (c_1,\ldots,c_p)\neq 0 )$. Construction {#sec:const} ============ Construction of rotated sphere packing designs {#sec:const:RSPD} ---------------------------------------------- Before proposing our algorithms to construct sliced rotated sphere packing designs, we first give a brief review of the construction of rotated sphere packing designs proposed in @RSPD. A rotated sphere packing design is a finite set of points generated from rescaling, rotating, translating and extracting the points from a lattice. With $p$, $n$ and the generator matrix $\mathbf{G}$ given, the algorithm has five major steps: 1. Obtain a rotation matrix $\mathbf{R}$, which is a $p\times p$ orthogonal matrix. 2. Obtain a large design given by $\mathbf{E} = \mathbf{F} \mathbf{G} \mathbf{R}$, where $\mathbf{F}$ is an integer matrix sufficiently large such that $\{ \mathbf{f}^T\mathbf{G}\mathbf{R} : \mathbf{f} \in \mathbb{Z}^p, \mathbf{f}^T\mathbf{G}\mathbf{R} \in [-l/2-\rho_c,l/2+\rho_c]^p \}$ is a subset of rows of $\mathbf{E}$, where $ l = \left(n \Omega_p/\Theta\right)^{1/p} \rho_c$, $\Omega_p$ is the volume of one unit sphere in $\mathbb{R}^p$, $\Theta$ is the thickness of the lattice and $\rho_c$ is the covering radius of the lattice. 3. Search for a perturbation vector $\boldsymbol{\delta} =(\delta_1,\ldots,\delta_p)^T \in \text{Vor}(0)$ such that there are exactly $n$ points of $\mathbf{E}$ contained in the region $\otimes_{k=1}^p [-l/2-\delta_k,l/2-\delta_k]$, where $\text{Vor}(0)$ is the Voronoi cell of $(0,\ldots,0) \in \mathbf{E}$. A theorem in @RSPD guarantees the existence of such $\boldsymbol{\delta}$. 4. Obtain the design $\mathbf{D}$ by extracting points of $ \mathbf{\tilde E}/l+1/2 $ that lie in $[0,1]^p$, where $\mathbf{\tilde E}$ is the matrix obtained by adding $\boldsymbol{\delta}^T$ to rows of $\mathbf{E}$. 5. Repeat Steps 1-4 for $w$ times and select the $(\mathbf{R},\boldsymbol{\delta})$ combination that maximizes the empirical projected uniformity measured by the criterion [@Roshan:2015] $$\label{eqn:criterion:MaxPro} \psi(\mathbf{D}) = \left\{ \{n(n-1)\}^{-1} \sum_{1\leq i<j\leq n} \frac{1}{\prod_{k=1}^p (x_{i,k}-x_{j,k})^2} \right\}^{1/p}.$$ In @RSPD, $\mathbf{G}$ is recommended to be $\mathbf{M}_p^*$ in (\[eqn:A\_p\^\*\]). A Givens rotation $\mathbf{R}_p(i,j,\alpha)$ is the $p\times p$ identity matrix with the $(i,i)$th, $(i,j)$th, $(j,i)$th and $(j,j)$th elements being replaced by $\cos(\alpha)$, $-\sin(\alpha)$, $\sin(\alpha)$ and $\cos(\alpha)$, respectively. For $p=2$, $\mathbf{R}=\mathbf{I}_2$ and $w=1$ was recommended. For $p>2$, it was recommended to use $w=100$ and generate $\mathbf{R}$s randomly by multiplying $p(p-1)/2$ sequential Givens rotations with $\alpha$ sampled independently and uniformly from $[0,2\pi]$. Construction of sliced rotated sphere packing designs {#sec:const:SRSPD} ----------------------------------------------------- We now give two algorithms to construct sliced rotated sphere packing designs based on a sliced lattice $(\mathbf{L},\mathbf{K},\mathbf{Q})$ and an ordinary rotated sphere packing design. The algorithms are general for any type of sliced lattice. Let $\mathbf{G}$ and $\mathbf{H}$ be the generator matrices of $\mathbf{L}$ and $\mathbf{K}$, respectively, and assume $\mathbf{Q}=\{\mathbf{q}_1,\ldots,\mathbf{q}_s\}$ with $\mathbf{q}_1=0$. For the $(A_p^*,A_p,\mathbf{B})$, we can use $\mathbf{G}=\mathbf{M}_p^*$ in (\[eqn:A\_p\^\*\]), $\mathbf{K}=\{2(p+1)\}^{1/2}\mathbf{M}_p$ in (\[eqn:A\_p\]) and $s=p+1$. The first algorithm partitions an $\mathbf{L}$-based rotated sphere packing design with the following three steps: 1. Obtain $\mathbf{D}$, an ordinary $\mathbf{G}$-based rotated sphere packing design with $n$ points as in Section \[sec:const:RSPD\], and express its points as $\mathbf{x}_i=(\mathbf{a}_i^T \mathbf{G} \mathbf{R} + \boldsymbol{\delta}^T)/l + 1/2$, $\mathbf{a}_i \in \mathbb{Z}^p$, $i=1,\ldots,n$. 2. Determine the coset $\mathbf{a}_i^T \mathbf{G}$ belongs to, $i=1,\ldots,n$. 3. Obtain $\mathbf{D}_k = \{\mathbf{x}_i: \mathbf{a}_i^T \mathbf{G} \in \mathbf{K} \oplus \mathbf{q}_k\}$, $k=1,\ldots,s$. The sliced rotated sphere packing design is given by $(\mathbf{D}_1,\ldots,\mathbf{D}_s)$. This algorithm is suitable for simultaneous construction of sliced rotated sphere packing designs. The second algorithm enlarges a $\mathbf{K}$-based rotated sphere packing design with the following five steps: 1. Obtain $\mathbf{D}_1$, an ordinary $\mathbf{H}$-based rotated sphere packing design with $n_1$ points as in Section \[sec:const:RSPD\], and express its points as $\mathbf{x}_i=(\mathbf{a}_i^T \mathbf{H} \mathbf{R} + \boldsymbol{\delta}^T)/l + 1/2$, $\mathbf{a}_i \in \mathbb{Z}^p$, $i=1,\ldots,n_1$. 2. Obtain a large design given by $\mathbf{E} = \mathbf{F} \mathbf{G} \mathbf{R}$, where $\mathbf{F}$ is an integer matrix sufficiently large such that $\{ \mathbf{f}^T\mathbf{G}\mathbf{R} : \mathbf{f} \in \mathbb{Z}^p, \mathbf{f}^T\mathbf{G}\mathbf{R} \in [-l/2-\rho_c,l/2+\rho_c]^p \}$ is a subset of rows of $\mathbf{E}$. 3. Obtain the design $\mathbf{D}$ by extracting points of $ \mathbf{\tilde E}/l+1/2 $ that lie in $[0,1]^p$, where $\mathbf{\tilde E}$ is the matrix obtained by adding $\boldsymbol{\delta}^T$ to rows of $\mathbf{E}$. 4. Determine the coset $\mathbf{a}_i^T \mathbf{G}$ belongs to, $i=1,\ldots,n$. 5. Obtain $\mathbf{D}_k = \{\mathbf{x}_i: \mathbf{a}_i^T \mathbf{G} \in \mathbf{K} \oplus \mathbf{q}_k\}$, $k=2,\ldots,s$. The sliced rotated sphere packing design is given by $(\mathbf{D}_1,\ldots,\mathbf{D}_s)$. This algorithm is appealing for sequential experiments in which the adult points are given by $\mathbf{D}_1$ and the baby points are given by $\cup_{k=2}^s \mathbf{D}_k$. From both algorithms, the desired sliceable structure comes with no lose of uniformity and little extra computation. The full design achieves the same degree of uniformity as an ordinary $\mathbf{L}$-based rotated sphere packing design while the subdesigns achieve the same degree of uniformity as ordinary $\mathbf{K}$-based rotated sphere packing designs. Although the algorithms are applicable to arbitrary sliced lattices, the resulted sliced design is space-filling only if its underlying sliced lattice is space-filling. As a result, in this paper we focus on $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing designs. We can define the parent-child relation of sliced rotated sphere packing designs similarly to that for sliced lattices. We illustrate the parent-child relation of an $(A_2^*,A_2,\mathbf{B})$-based sliced rotated sphere packing design in Figure \[fig:child\]. ![The parent-child relation of an $(A_2^*,A_2,\mathbf{B})$-based sliced rotated sphere packing design. The panel on the left depicts the six children (pluses) of an adult point (triangle) and other adult points (circles). The panel on the right depicts the three parents (triangles) of a baby point (plus) and other adult points (circles). \[fig:child\]](introfigS2children.eps){width="13.6cm"} Let $n_j$ denote the number of points for $\mathbf{D}_j$ and $n=\sum n_j$, the first algorithm allows $n$ to be pre-specified and the second algorithm allows $n_1$ to be pre-specified. However, neither algorithm allows us to simultaneously set the values of $(n_1,\ldots,n_s)$. This is a major limitation of sliced rotated sphere packing designs. In some applications the balance property that $n_1=\cdots=n_s$ may be important. The balance property of a sliced design $\mathbf{D}$ can be measured by the criterion $$\label{eqn:sn} \varphi(\mathbf{D}) = \sum_{j=1}^{s} (n_j-n/s)^2.$$ Recall that in the construction algorithm of rotated sphere packing designs, we generate a number of $(\mathbf{R},\boldsymbol{\delta})$ randomly. Figure \[fig:dnpsi\] plots $\varphi(\mathbf{D})$ and $\psi(\mathbf{D})$ in (\[eqn:criterion:MaxPro\]) of 100 random designs when $p=4$ and $n=50$. It is observed that the $n_1,\ldots,n_s$ are likely to be roughly equal but not exactly the same. Ten of these designs achieve minimum $\varphi(\mathbf{D})$, i.e., 2. If balance is the primary concern, we can choose the design with minimum $\psi(\mathbf{D})$ among those 10 designs. This design is almost balanced and has $n_1=9$, $n_2=11$ and $n_3=n_4=n_5=10$. Meanwhile, it also achieves good projection uniformity. We can further reduce $\varphi(\mathbf{D})$ by searching for $\boldsymbol{\delta}$ around $(\mathbf{R},\boldsymbol{\delta})$ combinations that yield low $\varphi(\mathbf{D})$, but we omit the details here. From our experience, it is not hard to find a balanced $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing design for $p\leq 5$ with $n\leq 200$. However, as $p$ and $n$ grow, it becomes much harder to find strictly balanced designs. ![The $\varphi(\mathbf{D})$ and $\psi(\mathbf{D})$ from 100 randomly generated designs, $p=4$, $n=50$. \[fig:dnpsi\]](dnpsi4.eps){width="10cm"} Numerical comparison on separation distance {#sec:simu} =========================================== In this section, we compare $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing designs with optimal sliced Latin hypercube designs and sliced Latin hypercube designs using the separation distance criterion by (\[eqn:dp\]). As discussed, this criterion reflects the uniformity of designs and is what optimal sliced Latin hypercube designs aim to maximize. For $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing designs generated from the first algorithm, the separation distance for the full design is $p^{1/2} (p+1)^{(1-p)/(2p)} n^{-1/p}$, the same to that of an ordinary $A_p^*$-based rotated sphere packing design, and the separation distance among points in the same slice is $\sqrt{2} (p+1)^{1/(2p)} n^{-1/p}$, the same to that of an ordinary $A_p$-based rotated sphere packing design with $n/(p+1)$ points. We obtain separation distance of the other two types of sliced designs numerically. The comparison for $2\leq p\leq 10$, $n=10(p+1)$ and $n=40(p+1)$ are shown in Figure \[fig:maximin\]. Seen from the results, for $n=10(p+1)$, sliced rotated sphere packing design is the best for $2\leq p\leq 6$; for $n=40(p+1)$, sliced rotated sphere packing design is the best for $2\leq p\leq 7$. These results were expected by us, since ordinary rotated sphere packing designs have better separation distance than maximin distance Latin hypercube designs for $2\leq p\leq 6$ [@RSPD], sliced rotated sphere packing designs retain the same separation distances as ordinary rotated sphere packing designs, and optimal sliced Latin hypercube designs have inferior separation distances to maximin distance Latin hypercube designs. From our experience, sliced rotated sphere packing design is more competitive as $n$ grows. Besides, the construction of sliced rotated sphere packing designs is fast for $2\leq p\leq 6$. For instance, it takes 106 seconds to generate a sliced rotated sphere packing design with $p=6$ and $n=600$ on a laptop. To sum it, sliced rotated sphere packing designs have good distance-based properties and should be useful for applications such as computer experiments with quantitative and qualitative variables [@Qian:Wu:Wu:2008; @Deng:2016], computer experiments with multiple levels of accuracy [@Qian:Wu:2008], and model validation [@Zhang:2013]. ![Separation distance of all points (left) and points in the same slice (right) for three types of sliced designs: sliced rotated sphere packing designs (circles), optimal sliced Latin hypercube designs (pluses) and sliced Latin hypercube designs (triangles). \[fig:maximin\]](S1s.eps){width="13.6cm"} Adaptive sliced rotated sphere packing designs {#sec:adaptive} ============================================== In this section, we propose a strategy to use sliced rotated sphere packing designs adaptively. Adaptive designs are widely used in many computer experiment problems including emulation of nonstationary computer experiments [@Jin:2002], global optimization of black-box functions [@EI], finding several promising points for response surface optimum [@Joseph:energy], finding an excursion set whose output is above a target value [@Chevalier:2014] and estimating a percentile of the output distribution [@Oakley:2004]. Here we focus on the emulation and optimization objectives. ![Counter plot of the Franke’s function. \[fig:resShow\]](simu3FrankeresFunc.eps){width="10cm"} Many computer experiments are nonstationary. For example, Figure \[fig:resShow\] gives the contour plot of the Franke’s function [@Joseph:energy]. It can be seen that the output has larger volatility in the bottom-left corner than in other places. Thus, in order to obtain an overall accurate emulator, a space-filling design with denser points in the bottom-left corner is desired. The major challenge here is to identify the high volatility region based on limited computer runs. Although Gaussian process emulators can give variance estimates for any position in the design space, a stationary Gaussian process model will not yield high variance estimate for high volatility regions. One notable nonstationary Gaissuan process model is the treed Gaussian process model which partitions the design space based on volatility and fit different Gaussian process models separately [@Gramacy:2008]. An adaptive design approach using treed Gaussian process model was proposed in @Gramacy:2009. However, from our experience, this approach does not work well for small sample sizes. Another cross-validation based approach was proposed in @Jin:2002 which has three major steps below: 1. Carry out initial runs that come from a maximin distance Latin hypercube design. 2. Fit Gaussian process emulators using completed runs and add follow-up runs one-by-one. Let the cross-validation error be defined by $$~\label{eqn:e} e(\mathbf{x}_{n+1}) = \left[ \left\{ \sum_{i=1}^n \left( \hat f_{-i}(\mathbf{x}_{n+1}) - \hat f(\mathbf{x}_{n+1}) \right)^2 \right\}/n \right]^{1/2},$$ where $n$ is the number of completed runs, $\hat f(\mathbf{x}_{n+1})$ is the predicted outcome from emulating all completed runs and $\hat f_{-i}(\mathbf{x}_{n+1})$ is the predicted outcome without using the $i$th run. Let $$~\label{eqn:g} g(\mathbf{x}_{n+1}) = e(\mathbf{x}_{n+1}) \text{min}_{i=1}^n \|\mathbf{x}_{n+1}-\mathbf{x}_i\|.$$ The new point $\mathbf{x}_{n+1}$ shall maximize $g(\mathbf{x}_{n+1})$ where $\mathbf{x}_{n+1}\in [0,1]^p$. 3. Stop when a certain number of points are added or $\sup_{\mathbf{x}} g(\mathbf{x})$ goes below a value. In this algorithm, high $e(\mathbf{x})$ implies high volatility around $\mathbf{x}$ and high $\text{min}_{i=1}^n \|\mathbf{x}-\mathbf{x}_i\|$ implies good interpoint distance. As a result, the added runs tend to locate in high volatility regions and not too close to any completed run. Finding the response surface minimum is another important objective for computer experiments. A related objective is to find several promising points for response surface minimum. The promising points can be further investigated by extra experiments based on same or different responses. For this objective, sequential minimum energy designs [@Joseph:energy] are suitable which has three major steps below: 1. Carry out initial runs that come from a maximin distance Latin hypercube design. 2. Fit a Gaussian process emulator using completed runs and add follow-up runs one-by-one. Let the density function be defined by $$~\label{eqn:d} d(\mathbf{x}) = \hat f_{\text{max}} - \hat f(\mathbf{x}),$$ where $\hat f_{\text{max}}$ is the estimated global maximum output value and $\hat f(\mathbf{x})$ is the predicted outcome value at $\mathbf{x}$. Let the energy function be $$\label{eqn:energy} r(\mathbf{x}) = \sum_{i=1}^n \left\{ d(\mathbf{x}_i)^{-2}d(\mathbf{x})^{-2} \|\mathbf{x}_i-\mathbf{x}\|^{-4p} \right\}.$$ The new point $\mathbf{x}_{n+1}$ shall minimize $r(\mathbf{x}_{n+1})$ where $\mathbf{x}_{n+1}\in [0,1]^p$. 3. Stop when a certain number of points are added or $\inf_{\mathbf{x}} r(\mathbf{x})$ goes above a value. In this algorithm, high $d(\mathbf{x})$ implies relatively low output values and high $\|\mathbf{x}_i-\mathbf{x}\|$ implies good interpoint distance. As a result, the added runs tend to locate in low outcome regions and not too close to any completed run. It can be seen that the two algorithms are very similar to each other. The primary difference between them, as well as many other adaptive design methods, lies in the criterion to choose follow-up runs (e.g., $g(\mathbf{x})$ in (\[eqn:g\]) and $r(\mathbf{x})$ in (\[eqn:energy\])). The criterion is the key to the success of adaptive designs. It needs to generate denser points in more interesting regions while scattering points uniformly in local regions. Most adaptive design methods are greedy in assuming that the next run to be added is the last run. As a result, if many follow-up runs are added in a local area, these points have no space-filling property. Sliced rotated sphere packing designs provide a non-greedy approach for adaptive designs. For the objective of emulating nonstationary computer experiments, our first strategy has the following three steps: 1. Generate $(D_1,\ldots,D_s)$, an $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing design using the second algorithm in Section \[sec:const:SRSPD\]. Run experiments using $D_1$ and obtain the outputs. 2. Fit Gaussian process emulators using completed runs and add follow-up runs one-by-one. Treat points in $\cup_{j=2}^s D_j$ as candidates for follow-up runs. The new point $\mathbf{x}_{n+1}$ shall maximize $g(\mathbf{x}_{n+1})$ in (\[eqn:g\]) where $\mathbf{x}_{n+1}\in \cup_{j=2}^s D_j$. 3. Stop when a certain number of points are added or the $\sup_{\mathbf{x}} g(\mathbf{x})$ goes below a value. Instead of searching for the entire $[0,1]^p$ to find an $\mathbf{x}$ that maximizes $g(\mathbf{x})$, we propose to search over a short list of candidate points. Apart from the apparent advantage of reduced computation, the new strategy ensures space-filling properties when multiple follow-up runs are added in a local area. For a local area with $m$ initial runs, up to roughly $pm$ follow-up runs can be added while preserving the $p^{1/2} (p+1)^{(-1-p)/(2p)} {n_1}^{-1/p}$ separation distance among all initial and follow-up runs, where $n_1$ is the number of initial runs. In the above strategy, we use the same criterion, namely $g(\mathbf{x})$, and the same stopping rules. Because baby points have the same interpoint distance, using $g(\mathbf{x})$ is equivalent to using $e(\mathbf{x})$ in (\[eqn:e\]) alone. We now propose a simpler criterion. For $i=1,\ldots,n_1$, let $$\label{eqn:te} \tilde e(\mathbf{x}) = \left[ \left\{ \sum_{i=1}^{n_1} \left( \hat f_{-i}(\mathbf{x}) - \hat f(\mathbf{x}) \right)^2 \right\}/n_1 \right]^{1/2},$$ where $\hat f(\mathbf{x})$ is the predicted outcome from emulating all initial runs and $\hat f_{-i}(\mathbf{x})$ is the predicted outcome without using the $i$th run. For any baby point, let $\tilde e(\mathbf{x})$ be defined as the average $\tilde e$ value of its parents. In some rare cases, none of the parents of a baby point is located in $[0,1]^p$. Such baby points are assigned with highest $\tilde e(\mathbf{x})$. The $\tilde e(\mathbf{x})$ criterion is a further simplification from $e(\mathbf{x})$; using this criterion, we do not need to refit Gaussian process models after new runs completed. Below we summarize design strategies introduced for the emulation objective: MmLH : Use a non-adaptive maximin distance Latin hypercube design. MmLH-CV : Use a maximin distance Latin hypercube design for $n_1$ initial runs; use the cross-validation based criterion $g(\mathbf{x})$ in (\[eqn:g\]) to add follow-up runs. SRSPD-CV : Use sliced rotated sphere packing design with $n_1$ adult points; use the cross-validation based criterion $g(\mathbf{x})$ in (\[eqn:g\]) to add follow-up runs. SRSPD-CV2 : Use sliced rotated sphere packing design with $n_1$ adult points; use a modified error function $\tilde e(\mathbf{x})$ in (\[eqn:te\]) to add follow-up runs. We compare these methods numerically on average prediction error from Gaussian process emulation over 10000 independently and uniformly sampled testing locations, assuming $n_1=13$ and altogether $n\geq 13$ runs are used. For each method, the results are averaged based on 100 randomly generated designs. To add randomness into SRSPD-CV and SRSPD-CV2, we use $w=100$ and randomly generated $R$s for sliced rotated sphere packing designs, which is different from our general recommendation for $p=2$. The results are shown in Figure \[fig:resPEns2\]. ![Mean prediction error for emulating the Franke’s function. \[fig:resPEns2\]](simu3CFrankeresPEns2.eps){width="10cm"} Seen from the results, adaptive methods perform better than MmLH for $n\leq 20$. Clearly, adding no more than seven points in the bottom-left corner is better than scattering points uniformly in the design space. However, it is not beneficial to add more than seven follow-up runs since the bottom-left corner cannot contain too many points. The SRSPD-CV performs uniformly better than MmLH-CV. There are three possibly reasons for this. Firstly, follow-up runs of MmLH-CV may not be located in space-filling locations because the greedy one-at-a-time strategy cannot simultaneously control locations of multiply follow-up runs. Secondly, for MmLH-CV the balance between volatility and interpoint distance may not be ideal. This may result in too much focus on either volatility or interpoint distance. Indeed, it is unknown if $g(\mathbf{x})$ in (\[eqn:g\]) is inferior to $e(\mathbf{x}) \text{min}_{i=1}^n \|\mathbf{x}-\mathbf{x}_i\|^2$ or $e(\mathbf{x}) \text{min}_{i=1}^n \|\mathbf{x}-\mathbf{x}_i\|^{1/2}$. In contrast, for SRSPD-CV separation distance properties are ensured by the sliced lattice structure and $g(\mathbf{x})$ is used solely to measure volatility. Lastly, because it is computationally infeasible to compute $g(\mathbf{x})$ for every $\mathbf{x} \in [0,1]^p$, for MmLH-CV we only compute $g(\mathbf{x})$ on 5000 randomly generated positions as recommended by @Jin:2002. Thus, the added runs may be suboptimal in $g(\mathbf{x})$. The SRSPD-CV2 performs better than SRSPD-CV for $17\leq n\leq 22$. This might because of the deficiency of $e(\mathbf{x})$. Clearly, $e(\mathbf{x})$ decreases as more points are added near $\mathbf{x}$. This may hinder adding more points in the left-bottom corner. In contrast, adult points from a sliced rotated sphere packing design have the same interpoint distance to other points, making $\tilde e(\mathbf{x})$ a fair measure on volatility. To sum it, when using sliced rotated sphere packing designs, much simpler criterion can be used; the parent-child relation may help in defining the criterion. Besides having better performance than MmLH and MmLH-CV, SRSPD-CV and SRSPD-CV2 take much less time. The SRSPD-CV2 also allows follow-up runs to be added in parallel. We now return to the global minimization problem. The most popular method to the minimization problem is the EI algorithm below [@EI]: 1. Carry out $n_2$ initial runs that is uniformly distributed in the design space. 2. Fit a Gaussian process emulator using completed runs and add follow-up runs one-by-one. Let the expected improvement of a new point $\mathbf{x}_{n+1}$ be $$\label{eqn:EI} \text{EI}(\mathbf{x}_{n+1}) = \text{E} \left( \left\{ \min_{i=1,\ldots,n} f(\mathbf{x}_i) - f(\mathbf{x}_{n+1}) \right\}^+ \mid \mathbf{x}_1,\ldots,\mathbf{x}_n,f(\mathbf{x}_1),\ldots,f(\mathbf{x}_n) \right),$$ where $\mathbf{x}_1,\ldots,\mathbf{x}_n$ are the evaluated runs, $z^+=z$ if $z\geq 0$ and $z^+=0$ if $z<0$. In the formula, $f(\mathbf{x}_{n+1})$ is random because the new run has not been carried out yet. @EI gave a deterministic formula to compute $\text{EI}(\mathbf{x}_{n+1})$ for any given $\mathbf{x}_{n+1}$. The new point $\mathbf{x}_{n+1}$ shall maximize $\text{EI}(\mathbf{x}_{n+1})$ where $\mathbf{x}_{n+1}\in [0,1]^p$. 3. Stop when a certain number of points are added or $\sup_{\mathbf{x}} \text{EI}(\mathbf{x})$ goes below a value. The minimization problem is different from the emulation problem. In order to ensure optimality, many points need to cluster around the minimum; these points cannot be space-filling. The EI criterion does exactly this. It was reported in @Joseph:energy that sequential minimum energy designs do not work well for finding the single minimum. This is presumably because the energy function in (\[eqn:energy\]) tends to spreads points away from each other. However, because sequential minimum energy designs give promising points for the response surface minimum, we develop an algorithm that combines the energy function with the EI criterion: 1. Carry out $n_1$ initial runs that come from a maximin distance Latin hypercube design. 2. Fit a Gaussian process emulator using completed runs and add $n_2-n_1$ follow-up runs one-by-one. The new point $\mathbf{x}_{n+1}$ shall minimize the energy function $r(\mathbf{x}_{n+1})$ in (\[eqn:energy\]) where $\mathbf{x}_{n+1}\in [0,1]^p$. 3. Fit a Gaussian process emulator using completed runs and add more follow-up runs one-by-one. The new point $\mathbf{x}_{n+1}$ shall maximize $\text{EI}(\mathbf{x}_{n+1})$ in (\[eqn:EI\]) where $\mathbf{x}_{n+1}\in [0,1]^p$. 4. Stop when a certain number of points are added or $\sup_{\mathbf{x}} \text{EI}(\mathbf{x})$ goes below a value. The four-step algorithm above replaces the first step of the EI algorithm by sequential minimum energy designs with the same number of total runs. This adds adaptiveness to the $n_2$ initial runs. From our experience, the new algorithm generally outperforms the original EI algorithm. We further modify the algorithm using sliced rotated sphere packing designs. Because baby points of sliced rotated sphere packing designs have the same interpoint distance, it suffices to use the density function $d(\mathbf{x})$ in (\[eqn:d\]) to replace the energy function $r(\mathbf{x}_{n+1})$. Also because baby points are located in the center of their parents, it suffices to use the average output value of parents as the predicted outcome of baby points. Clearly, this criterion is much simpler and model-free. Our proposed algorithm has the following three steps: 1. Generate $(D_1,\ldots,D_s)$, an $(A_p^*,A_p,\mathbf{B})$-based sliced rotated sphere packing design with $n_1$ adult points using the second algorithm in Section \[sec:const:SRSPD\]. Run experiments using $D_1$ and obtain the outputs. 2. Fit a Gaussian process emulator using completed runs and add $n_2-n_1$ follow-up runs one-by-one. The new point $\mathbf{x}_{n+1}$ shall minimize the average output value among parents of $\mathbf{x}_{n+1}$, where $\mathbf{x}_{n+1} \in \cup_{j=2}^s D_j$. However, baby points with no parent are carried out with highest priority. 3. Fit a Gaussian process emulator using completed runs and add more follow-up runs one-by-one. The new point $\mathbf{x}_{n+1}$ shall maximize $\text{EI}(\mathbf{x}_{n+1})$ in (\[eqn:EI\]) where $\mathbf{x}_{n+1}\in [0,1]^p$. 4. Stop when a certain number of points are added or $\sup_{\mathbf{x}} \text{EI}(\mathbf{x})$ goes below a value. We compare the above-mentioned methods numerically: MmLH : The original EI algorithm using a maximin distance Latin hypercube design in the first step. SMED : The four-step algorithm using the energy function. SRSPD : The four-step algorithm using the average-parent-output criterion. As recommend in @EI, we use $n_2=10p$ for all methods. Remark that for all three methods, the same EI criterion is used for the $(n_2+1)$th and subsequent runs. The difference lies in how the first $n_2$ runs are generated. For SMED and SRSPD, we set $n_1=13$ for $p=2$ and $n_1=5p$ for $p>2$. For a fair comparison, the stopping rule is set on the number of runs. We consider the four test functions that were used in @EI, namely the Branin function, the Goldstein-Price function, the Hartmann 3 function and the Hartmann 6 function [@test]. Their dimensions are 2, 2, 3 and 6, respectively. For each function with each strategy, we repeat the procedure for 100 times and compute the response value, namely the minimum output value among completed runs. To add randomness into the SRSPD strategy, we use $w=100$ and randomly generated $R$s for sliced rotated sphere packing designs. We depict the response value as a function of the number of completed runs in Figure \[fig:EI\]. ![Average response value as a function of the number of completed runs for algorithms using three types of initial designs. \[fig:EI\]](simu4TechR.eps){width="13.6cm"} Seen from the results, both SRSPD and SMED perform well in finding promising points using the first $10p$ runs. In most cases, they continue to find good input sites earlier than MmLH. In particular, SRSPD is the best method for the Branin and Goldstein Price functions and one of the best methods for the Hartmann 6 function. Although not as good as SMED for Hartmann 3, SRSPD has the best overall performance. This clearly suggests that adaptive sliced rotated sphere packing designs are useful for the minimization problem. Similar to the emulation problem, the benefit may come from the distance properties of follow-up runs, the robustness of the simple average-parent-output criterion and the fact that we can obtain the exact optimum of the criterion. Besides, SRSPD takes less time and allows parallel computation in the second step. Although the most important component of adaptive designs is their criteria for choosing follow-up runs, our main focus here is not to provide new adaptive designs with new powerful criteria. Instead, our goal is to show the advantage of using a short list of candidate points for follow-up runs and that sliced rotated sphere packing designs are suitable under this strategy. We have shown that adaptive sliced rotated sphere packing designs can be used in combination with exact or simplified criterion that has been proposed before. For complex problems that no adaptive design criterion has been proposed, it should be easier to invent a criterion for our strategy than for usual adaptive designs. As discussed, a criterion for sliced rotated sphere packing designs only needs to measure how interesting positions are. In contrast, a criterion for usual adaptive designs need to strike a proper balance between more points in interesting regions and better distance properties. Furthermore, the lattice structure and the parent-child relation may help in developing fair criteria. The down side is that our strategy only allows points coming from two densities; it does not allow very dense points in a small region. Conclusions and discussion {#sec:conclusion} ========================== In this paper, we propose a new class of sliced space-filling design called sliced rotated sphere packing designs. We also propose a space-filling type of sliced lattice, based on which sliced rotated sphere packing designs achieve good distance properties. Because of their delicate local structure, sliced rotated sphere packing designs are suitable as adaptive designs. The construction algorithms proposed in Section \[sec:const:SRSPD\] apply to any types of sliced lattices. Sometimes we should consider sliced lattices other than $(A_p^*,A_p,\mathbf{B})$. For example, to design computer experiments with one qualitative variable of $\tilde s$ levels and several quantitative variables, sliced lattices with exactly $\tilde s$ slices are desired. A future research problem is to construct sliced rotated sphere packing designs with flexible number of slices. We also propose a strategy to use sliced rotated sphere packing designs adaptively. The strategy requires a criterion for choosing follow-up runs that are suitable to the specific scientific goal. The criteria we have proposed for emulation and optimization problems may not be optimal. Further improvement by using more complex criteria are possible. Our main focus is to corroborate the usefulness of the new strategy. Separate studies are needed to find the best algorithm for various applications such as finding an excursion set whose output is above a target value [@Chevalier:2014] and estimating a percentile of the output distribution [@Oakley:2004]. We plan to work on these problems in the future. Similar to ordinary rotated sphere packing designs, a major restriction of sliced rotated sphere packing designs is on the number of dimensions. Although sliced rotated sphere packing designs are useful for $2\leq p\leq 6$, they are not suitable for high-dimensional problems. Appendix {#appendix .unnumbered} ======== From (\[eqn:A\_p\^\*\]) and (\[eqn:A\_p\]), we have $$\left\{2(p+1)\right\}^{1/2}\mathbf{M}_p = (\mathbf{I}_p+\mathbf{J}_p) \mathbf{M}_p^*; \quad \mathbf{M}_p^* = \left\{\mathbf{I}_p-\mathbf{J}_p/(p+1)\right\} \{2(p+1)\}^{1/2}\mathbf{M}_p.$$ Therefore, $(\mathbf{a}-\mathbf{u}_z)^T \mathbf{M}_p^* = (\mathbf{a}-\mathbf{u}_z)^T \left\{\mathbf{I}_p-\mathbf{J}_p/(p+1)\right\} \{2(p+1)\}^{1/2}\mathbf{M}_p$. Because $(\mathbf{a}-\mathbf{u}_z)^T \left\{\mathbf{I}_p-\mathbf{J}_p/(p+1)\right\}$ is an integer vector if and only if $\sum a_i \mod (p+1) = z$, $\mathbf{a}^T \mathbf{M}_p^* \in \mathbf{K} \oplus \mathbf{u}_z$ and $\cup_{k=0}^p ( \mathbf{K} \oplus \mathbf{u}_k ) =\mathbf{L}$. Since $\mathbf{u}_y$ and $\mathbf{u}_z$ do not belong to the same coset for $y,z\in\{0,\ldots,p\}$ and $y\neq z$, $\{ \mathbf{K} \oplus \mathbf{u}_0, \ldots, \mathbf{K} \oplus \mathbf{u}_p \}$ is a partition of $\mathbf{L}$. \(i) Consider an arbitrary adult point $\mathbf{a}^T \mathbf{M}_p^* \in \mathbf{K}$. Let $\mathbf{a}=(a_1,\ldots,a_p)^T$, $\mathbf{b}=(b_1,\ldots,b_p)^T$ and $\mathbf{c}=(c_1,\ldots,c_p)^T=\mathbf{b}-\mathbf{a}$. Consider three cases for $\mathbf{c}$. Firstly, assume there exist $i,j \in \{1,\ldots,p\}$ such that $c_j-c_i\geq 2$. Let $ \mathbf{\tilde c} = \mathbf{c} +\mathbf{e}_i -\mathbf{e}_j$ where $\mathbf{e}_z$ is the $p$-vector with the $z$th element being one and other elements being zero. Then $$\| \mathbf{c}^T \mathbf{M}_p^* \|^2 = \mathbf{c}^T \mathbf{M}_p^* (\mathbf{M}_p^*)^T \mathbf{c} = \mathbf{\tilde c}^T \mathbf{M}_p^* (\mathbf{M}_p^*)^T \mathbf{\tilde c} + 2 (\mathbf{c}-\mathbf{\tilde c})^T \mathbf{M}_p^* (\mathbf{M}_p^*)^T \mathbf{\tilde c} + (\mathbf{c}-\mathbf{\tilde c})^T \mathbf{M}_p^* (\mathbf{M}_p^*)^T (\mathbf{c}-\mathbf{\tilde c})$$ $$> \| \mathbf{\tilde c}^T \mathbf{M}_p^* \|^2 + 2 (\mathbf{c}-\mathbf{\tilde c})^T \left\{ \frac{p+1}{p} \mathbf{I}_p + \frac{-p-2+2\sqrt{p+1}}{p(\sqrt{p+1}-1)^2} \mathbf{J}_p \right\} \mathbf{\tilde c}$$ $$= \| \mathbf{\tilde c}^T \mathbf{M}_p^* \|^2 + 2(p+1) (\mathbf{e}_j - \mathbf{e}_i)^T \mathbf{\tilde c} /p \geq \| \mathbf{\tilde c}^T \mathbf{M}_p^* \|^2.$$ Because $(\mathbf{b}- \mathbf{\tilde c})^T \mathbf{M}_p^*$ is an adult point closer to $\mathbf{b}^T \mathbf{M}_p^*$ than $\mathbf{a}^T \mathbf{M}_p^*$, $\mathbf{a}^T \mathbf{M}_p^*$ is not a parent of $\mathbf{b}^T \mathbf{M}_p^*$. Secondly, assume $\min c_i \geq 1$ and there is a $j$ such that $c_j= 2$. Let $\mathbf{\tilde c} = \mathbf{c} -\mathbf{u}_p -\mathbf{e}_j$. Then $$\| \mathbf{c}^T \mathbf{M}_p^* \|^2 > \| \mathbf{\tilde c}^T \mathbf{M}_p^* \|^2 + 2(p+1) \mathbf{e}_j^T \mathbf{\tilde c} /p \geq \| \mathbf{\tilde c}^T \mathbf{M}_p^* \|^2.$$ Because $(\mathbf{b}- \mathbf{\tilde c})^T \mathbf{M}_p^*$ is an adult point closer to $\mathbf{b}^T \mathbf{M}_p^*$ than $\mathbf{a}^T \mathbf{M}_p^*$, $\mathbf{a}^T \mathbf{M}_p^*$ is not a parent of $\mathbf{b}^T \mathbf{M}_p^*$. Similarly, the $\mathbf{a}^T \mathbf{M}_p^*$ with $\min c_i \leq -1$ and a $j$ such that $c_j= -2$ is not a parent of $\mathbf{b}^T \mathbf{M}_p^*$, either. Combining the three cases, a necessary condition for $\mathbf{a}^T \mathbf{M}_p^*$ being a parent of $\mathbf{b}^T \mathbf{M}_p^*$ is either $( c_i\in\{0,1\}, i=1,\ldots p, \sum c_i = z )$ or $( c_i \in \{0,-1\}, i=1,\ldots p, \sum c_i = z-(p+1) )$. Because all $\mathbf{a}^T \mathbf{M}_p^*$ that satisfy the above conditions have the same distance to $\mathbf{b}^T \mathbf{M}_p^*$, we conclude that all of them are parents of $\mathbf{b}^T \mathbf{M}_p^*$. It is not hard to derive (ii) from (i). [^1]: He’s work is partial supported by Special National Key Research and Development Plan under Grant No. 2016YFD0400206, National Natural Science Foundation of China (NSFC 11501550 and NSFC 11671386) and funding from Chinese Ministry of Science and Technology (Grant No. 2016YFF0203801).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Interconnected ensembles of biological entities are perhaps some of the most complex systems that modern science has encountered so far. In particular, scientists have concentrated on understanding how the complexity of the interacting structure between different neurons, proteins or species influences the functioning of their respective systems. It is well-established that many biological networks are constructed in a highly hierarchical way with two main properties: short average paths that join two apparently distant nodes (neuronal, species, or protein patches) and a high proportion of nodes in modular aggregations. Although several hypotheses have been proposed so far, still little is known about the relation of the modules with the dynamical activity in such biological systems. Here we show that network modularity is a key ingredient for the formation of self-organising patterns of functional activity, independently of the topological peculiarities of the structure of the modules. In particular, we show that macroscopic spatial patterns at the modular scale can develop in this case, which may explain how spontaneous order in biological networks follows their modular structural organisation. Our results also show that Turing patterns on biological complex networks can be a signature of the presence of modular structure and consequently [a possible]{} protocol for community detection. We test our results on real-world networks to confirm the important role of modularity in creating macro-scale patterns.' author: - 'Bram A. Siebert$^{1}$, Cameron L. Hall$^{1,2}$, James P. Gleeson$^{1}$, Malbor Asllani$^{1}$' bibliography: - 'mybib.bib' title: 'The role of modularity in self-organisation dynamics in biological networks' --- Introduction {#sec:intro} ============ Patterns are macroscopic structures that are the distinctive mark of the self-organisation in a system of microscopic interacting entities [@Nicolis1977]. They are ubiquitous in nature and can be seen in the spots of a leopard’s fur or the coloured scales of a butterfly’s wing [@Murray2001]. In 1952, Alan Turing published his seminal work on pattern formation, *The Chemical Basis of Morphogenesis* where he laid down [an elegant and plausible theory that can be used to explain]{} the formation of patterns [@Turing1952]. Turing developed a simple model of pattern formation that established the minimal requirements for a biochemical system to self-organise. Turing’s minimal system is composed of two “competing” chemicals, an activator and an inhibitor, which share the same spatial domain where they react and diffuse. Based on a diffusion-driven instability mechanism, today known as Turing instability, Turing showed that it is possible to explain [and predict the growth of spatially inhomogeneous perturbations away from a spatially homogeneous steady state. These perturbations in concentration]{} are later stabilised by nonlinearities in the system, yielding the celebrated Turing patterns. The emergence of stable heterogeneous patterns, the result of Turing instabilities, was initially at odds with the general understanding that diffusion by itself is a smoothing process. Nevertheless, it has been shown that the right combination of short-range activation and long-range inhibition[, caused by slowly diffusing activators and rapidly diffusing inhibitors,]{} enables the pattern forming phenomenon [@Gierer1972]. [Conventionally, an activator-inhibitor system is modelled using a set of reaction-diffusion equations that describe the evolution of the concentrations of activator and inhibitor throughout a continuous medium. These equations can readily be adapted to describe activator-inhibitor systems in discrete systems such as regular lattices, and they have been used in this way to describe pattern formations in cellular tissues [@Othmer1971; @Gierer1972].]{} However, biological tissue often takes more complex forms, and the spatial support cannot always be adequately formulated via regular lattices. Inspired by the network structures of early stages of embryogenesis [@Schnabel2006], ecological meta-populations [@Holland2008] or coupled chemical reactors [@Horsthemke2004], researchers have extended the reaction-diffusion formalism to complex biological networks [@Othmer1971; @Nakao2010; @Asllani2014Multiplex; @Asllani2014Directed]. These discrete structures [consist of]{} graphs where the nodes usually represent the cells inside which reactions occur, and the edges usually represent the routes through which cells communicate by exchanging chemicals. Abundant studies in recent years have concentrated on the effect that different topologies of interactions (a peculiarity of the given cellular tissue) have on the formation or destruction of patterns [@Nakao2010; @Asllani2014Multiplex; @Asllani2015; @Asllani2016] H[ü]{}tt et al. [@Hutt] recently argued that the formalism of activator-inhibitor systems is relevant to the dynamical processes evolving in the brain [@Hutt] . The brain consists of billions of cells, the neurons, connected together into an extremely complex system [@seung_2013; @Bullmore2009]. The implementation of network tools for analysing the brain’s structure has been used since the first years of network science [@Watts1998]. In their seminal work, Watts and Strogatz [@Watts1998] studied the topology of the neuronal network of the nematode *C. elegans* and discovered that these networks possess small average shortest paths, as well as a high clustering coefficient They described networks with these two properties as being “small-world” networks. Later, it was verified that many brain networks are small world networks [@Meunier2010; @HarrigerLogan2012; @Hahn2019]. It has been argued and widely accepted that the small-world property of brain connectomes should help the communication between neurons inside the brain by integrating multiple segregated sources of information [@SpornsBook]. A further property of brain networks is that they are often modular [@Meunier2010] so that the neurons can be segregated into communities (referred to as modules) where two neurons chosen at random from the same module are much more likely to be connected than two neurons chosen at random from different modules. [The effect of network modularity on pattern formation will be the main focus of this paper.]{} The functional role that the topology of brain connections has been discussed from several perspectives. For example, due to the increased structural stability [@Simon1962; @SpornsBook], the modularity might have been crucial in the evolution and development of the brain. According to [@Meunier2010; @Simon1962] modular topology can also optimise the wiring cost in the case of spatial networks. A small number of long range (and thus costly) connections reduces the diameter of the network, and allows the remaining nodes, now grouped into communities or modules, to form dense small world networks. Also, more compact segregation of neurons may contribute to the specialisation of the neurons in their functional duties [@SpornsBook]. To ensure both a low shortest path length, and a high clustering coefficient, brain networks are organised in a strict hierarchical manner [@Sporns2007; @Bullmore2009; @Meunier2010] where at the first level of the hierarchy sets of nodes (the modules) are connected to mimic a small-world topology and the same happens at the second level of hierarchy and so on, until the single node level. For a more detailed discussion of the role of the hierarchy in the pattern formation process see the Appendix. More generally, modularity is a common topological property that naturally emerges in biological, ecological, and social scenarios where the different communities are associated with different functions of the system represented by the network as a whole [@NewmanPNAS]. There are many examples of this: in protein interaction networks, the proteins that share similar functions are grouped together in modules [@protein]; in metabolic networks, there are structural/functional communities corresponding to cycles or pathways [@metabolic]; and in citation networks, scientific papers are clustered according to their research topic [@Redner]. In addition to these properties, in this paper, we propose a new mathematical mechanism that highlights the role that modularity [takes]{} in self-organising processes in biological networks. Using the Turing theory of pattern formation, we show that spatially extended patterns can be triggered by the segregation of the nodes (neurons) in distinguishable communities. To formally analyse the chances of such networks self-organising, we use a linear stability approach known in the literature as the dispersion relation [@Murray2001]. We focus on modular networks, which (in contrast to other networks, e.g., small-world ones), are characterised by a small spectral gap, i.e., a small distance of the second largest eigenvalue [^1] of the Laplacian from the origin. To anticipate some of the technical details, we discuss the key features of modular networks in the following paragraphs and outline how these affect pattern formation. For modular networks, the Laplacian eigenvalues that may be responsible for the Turing instability can be split into two sets. In one set, we have the eigenvalues emerging due to the global modularity of the network, which we denote as “modular eigenvalues”. In Sec. \[sec:whymodular\], we will show that when only this part of the spectrum is [responsible for the]{} instability, then the shape of the associated pattern follows that of the network in the sense that nodes belonging to the same modules have very similar concentrations of the species among themselves but these concentrations are distinctly different from the concentrations in other modules. In contrast, if the instability is caused by the remaining set of eigenvalues, which correspond to the local connectivity of nodes, here denoted as “non-modular eigenvalues”, then all the nodes have (in principle) different concentrations making the pattern globally heterogeneous. In this latter case, if the eigenvalues responsible for the instability are limited to the eigenvalues belonging to a single module, then the pattern will first emerge in that module. We aim to create a bridge between the role of the structure in many biological networks with the dynamical activity therein. In particular, in our model, we explain how communities of biological entities (cells, individuals, etc.) can act as functional units in their corresponding biological systems. As a consequence, we argue that this approach can potentially be used in community detection methods [@Fortunato; @Newman2004; @NewmanBook] for networked biological systems where Turing patterns are known to exist. In this paper we begin in Sec. \[sec1\] with a description of the mathematical background of Turing patterns. This will lead us into a discussion as to why modularity is critical to the formation of patterns in Sec. \[subsec:DvDu\]. We describe the different types of patterns which form in Sec. \[sec:whymodular\], and show how increasing the modularity helps in the formation of patterns. Finally in Sec. \[sec:realNetworks\] we look for Turing patterns in some real world networks. Pattern formation on a networked system {#sec1} ======================================= The process of the formation of Turing patterns was originally put forward by Turing [@Turing1952] on continuous domains to explain the emergence of spontaneous order in biological contexts, but it was formulated on networked systems only in the 1970s by Othmer & Scriven [@Othmer1971; @Nakao2010]. In a continuous domain, [the most simple Turing mechanism is given]{} in terms of reaction-diffusion equations that describe the evolution [through time and space]{} of the concentrations of two competing chemical species, called the activator (with concentration denoted $u(x,t)$) and the inhibitor (with concentration denoted $v(x,t)$) [@Turing1952; @Murray2001]. In general, an activator increases production of both itself and the inhibitor. The inhibitor, in turn, slows down the growth in activator. When the spatial support is instead discrete, constituted by spatial patches (nodes) connected through communicating routes (links) the reaction-diffusion mechanism can be formulated using ODEs, instead of PDEs [@Othmer1971]. In general, a two-species reaction-diffusion model on a network of $N$ nodes will take the form, $$\begin{aligned} \frac{d u_i}{d t} &=& f(u_i,v_i) + D_u \sum_j {L}_{ij}u_j,\, \forall i=1,\dots, N \\ \frac{d v_i}{d t} &= & g(u_i,v_i) + D_v \sum_j {L}_{ij}v_j,\, \forall i=1,\dots, N, \end{aligned} \label{eq:turingPatternNetworks}$$ where $u_i$ and $v_i$ represent the concentrations of activator and inhibitor respectively at node $i$, $f$ and $g$ are nonlinear functions that describe the net production rates of activator and inhibitor respectively, $D_u$ and $D_v$ are the diffusion coefficients of activator and inhibitor respectively, and $\mathbf{L}$ is the graph Laplacian operator. The entries $L_{ij}$ of the graph Laplacian are defined by $L_{ij} = A_{ij} - k_{i} \delta_{ij}$, where $\mathbf{A}$ is the adjacency matrix, $k_i$ is the degree of node $i$, $\delta$ is the Kronecker delta, and where we do not sum over repeated indices. In order to understand the development of spatial patterns, we analyse the linear stability of the system starting from a homogeneous steady state $(u^*,v^*)$ that is stable in the absence of diffusion. If the diffusion coefficients are nonzero and the ratio $\rho=D_v/D_u$ is large enough, the steady state $(u^*,v^*)$ becomes unstable and small random perturbations of the previous steady state will grow. This growth is exponential in the initial linear regime, and may then be stabilised by the nonlinear terms of the functions $f$ and $g$ so that the system reaches a stable but spatially inhomogeneous steady state. Such a mechanism is responsible for the emergence of Turing patterns. The linearised system in matrix form reads: [$$\frac{d (\delta \textbf{x})}{d t} = \left(\hat{\textbf{J}} + \textbf{D} \hat{\textbf{L}}\right)\delta \textbf{x}, \label{eq:turingPatternsLinearisedNetwork}$$ where $\delta\mathbf{x}=(\textbf{u}-u^*\mathds{1}_N,\textbf{v}-v^*\mathds{1}_N)$ is the perturbations vector of the activator $\textbf{u}$ and inhibitor $\textbf{v}$ species, $\mathds{1}_N$ is the all-ones $N-$dimensional vector, and $$\textbf{D} = \begin{bmatrix} D_u\textbf{I}_N & 0 \\ 0 & D_v\mathbf{I}_N \end{bmatrix}$$ is the diffusion constant matrix. Note that $\mathbf{I}_N$ represents the $N$ by $N$ identity matrix, so that $\mathbf{D}$ is $2N$ by $2N$. The Jacobian matrix and the extended Laplacian are correspondingly $$\hat{\textbf{J}}= \begin{bmatrix} f_u\textbf{I}_N & f_v\textbf{I}_N \\ g_u\textbf{I}_N & g_v\textbf{I}_N \end{bmatrix}, \;\;\; \hat{\textbf{L}}= \begin{bmatrix} \textbf{L} & 0 \\ 0 & \textbf{L} \end{bmatrix}.$$ Note here that the notation $\textbf{J}$ will be reserved to identify the Jacobian of the $2 \times 2$ reactions matrix: $$\textbf{J}= \begin{bmatrix} f_u & f_v \\ g_u & g_v \end{bmatrix}.$$ ]{} We then look for solutions to Eq.  of the form $$\begin{aligned} \delta \textbf{u} = \sum_{\alpha = 1}^N b_\alpha e^{\sigma(\Lambda_\alpha) t}\boldsymbol{\Phi}^\alpha,\\ \delta \textbf{v} = \sum_{\alpha = 1}^N c_\alpha e^{\sigma(\Lambda_\alpha) t}\boldsymbol{\Phi}^\alpha, \end{aligned}$$ where $\Lambda_\alpha$, $\boldsymbol{\Phi}^\alpha$ are respectively the eigenvalues and eigenvectors of the Laplacian $\textbf{L}$ matrix, $\sigma(\Lambda_\alpha)$ are the eigenvalues of the extended Jacobian $(\hat{\textbf{J}}+ \textbf{D}\hat{\textbf{L}})$, and $\alpha$ is the index term. Following the standard approach described by [@Murray2001; @Othmer1971; @Nakao2010], we substitute the expansion of the perturbations into Eq. , which decomposes the extended Jacobian to a $2\times 2$ eigenvalue problem for each index $\alpha$: $$\textbf{J}_\alpha = \begin{bmatrix} f_u + D_u\Lambda_\alpha & f_v \\ g_u & g_v + D_v\Lambda_\alpha \end{bmatrix},$$ where subscripts on the activation function $f(u,v)$ and the inhibition function $g(u,v)$ represent partial derivatives evaluated at $(u^*,v^*)$ . To study the stability of the linear system we look for positive real parts of the eigenvalues of $\textbf{J}_\alpha$. Turing instability occurs when the real part of the larger of the two eigenvalues $\sigma(\Lambda_\alpha) = \left(\text{tr}\textbf{J}_\alpha + \sqrt{(\text{tr}\textbf{J}_\alpha)^2 - 4 \text{det}\textbf{J}_\alpha}\right)\big/2$ is positive. The function $\sigma(\Lambda_\alpha)$, is known in the literature as the dispersion relation [@Murray2001]. For an activator-inhibitor system the necessary conditions for stability are $\text{tr}\textbf{J}_\alpha <0$ and $\text{det}\textbf{J}_\alpha > 0$. The first condition is always true, since $\text{tr}\textbf{J}_\alpha = \text{tr}\textbf{J} + (D_u+D_v)\Lambda_\alpha$, and this is negative since the stability of the fixed point in the absence of diffusion implies that $\text{tr}\textbf{J} <0$, while the non-positivity of the Laplacian spectrum implies $\Lambda_\alpha < 0$. We therefore turn our attention to the second condition for stability, which concerns $\text{det}\textbf{J}_\alpha=\text{det}\textbf{J} + \left(f_uD_v + g_vD_u\right) \Lambda_\alpha + D_uD_v\Lambda^{2}_\alpha$. In order for a Turing instability to occur, we require $\text{det}\textbf{J}_\alpha < 0$. Noting that the stability of the fixed point in the absence of diffusion implies that $\text{det}(\textbf{J}) > 0$ and noting that $\Lambda_\alpha < 0$, it is straightforward to conclude that the only way for $\text{det}\textbf{J}_\alpha $ to be negative is for $(f_u D_v + g_v D_u)$ to be positive. Without loss of generality we define $u$ to be the activator and $v$ to be the inhibitor. Since $u$ corresponds to the activator, its increased presence increases the production of $u$; since $v$ corresponds to the inhibitor, its increased presence decreases the production of $v$; as a result of this, the signs of the respective partial derivatives are $f_u > 0$, and $g_v < 0$. Therefore, we require $\rho=D_v/D_u > 1$ for instability, implying that the inhibitor should diffuse faster than the activator in order for Turing patterns to arise. In many practical cases, this difference needs to be very large in order to achieve $\det(\textbf{J}_\alpha)>0$ ![image](mod_vs_NW.pdf){width=".85\textwidth"} The case for $D_v \gtrsim D_u$ {#subsec:DvDu} ------------------------------ From experimental observations [@Horsthemke1987; @Pearson; @exp_Turing1; @exp_Turing2] it is rarely true that the inhibitor diffuses much faster than the activator, but instead the chemicals diffuse with similar rates. In the case where $D_v \gtrsim D_u$, it can be shown that the dispersion relation is positive only for values of the spectrum of the Laplacian very near to the origin. To prove this we analyse the behaviour of $\text{det}(\textbf{J}_\alpha)$ when considered as a function of $\Lambda_\alpha$; more precisely, we focus on the value of $\Lambda_\alpha$ corresponding to a minimum of $\text{det}(\textbf{J}_\alpha)$. [It is known in literature [@Murray2001] that for the continuous case, it will always exist a non-positive value of $\Lambda_\alpha$ such that the $\text{det}(\textbf{J}_\alpha)<0$ or, in other words, that Turing instability can occur. In order to proceed with our analysis, in the following, we will consider that $\Lambda_\alpha$ takes continuous values and will see that the spectrum of a (strongly) modular networks fits in the domain of the continuous dispersion relation for which the instability occurs for the particular case, $D_v \gtrsim D_u$. We start by]{} differentiating with respect to $\Lambda_\alpha$ and after some algebraic manipulation, we find that the minimum of $\text{det}\textbf{J}_\alpha$ is found at $\Lambda_\alpha = \Lambda_\text{min}$ where $$\Lambda_\text{min} = -\frac{f_u\rho + g_v}{2D_v}. \label{eq:rho=1}$$ From relation  we note that if $D_v$ is kept fixed while $\rho \rightarrow 1$ then $\Lambda_\text{min}\rightarrow 0$. To show this we set $\rho = 1 + \epsilon$. Under the conditions of the Turing instability, $\Lambda_\text{min}$ is non-positive, so $(1 + \epsilon)f_u + g_v >0$. Rearranging, we can write $(1 + \epsilon)f_u + g_v = \text{tr}\textbf{J}+\epsilon f_u$ and, noting that $\text{tr}\textbf{J}$ is necessarily negative, we conclude that the positive quantity $\text{tr}\textbf{J}+\epsilon f_u$ can be at most of order $\epsilon$, since $\epsilon f_u > |\text{tr}\textbf{J}|$. This shows that $\Lambda_\text{min}$ is of order $\epsilon$. Therefore, as $\epsilon$ decreases, the value of $\Lambda_\alpha$ for which $\text{det}\textbf{J}_\alpha$ is at its minimum tends towards the origin. Hence, the possible values of $\Lambda_\alpha$ [that may permit Turing instabilities tend towards zero as the ratio $\rho$ of diffusivities tends to $1$. In practice, this implies that the range of values of $\Lambda_\alpha$ for which instabilities can occur decreases in size and is restricted to small values of $\Lambda_\alpha$]{}. Therefore, a small spectral gap is needed to allow patterns to form. [This is significant for the analysis of modular networks that follows since, as shown in the following section, modular networks are characterised by a small spectral gap $\lvert \Lambda_2 - \Lambda_1 \rvert$. Hence the Laplacian of a modular network will have eigenvalues close to the origin. Because of this, we are able to find modular networks where Turing instabilities, and thus pattern formation, may occur where otherwise (i.e., in non-modular networked systems) they would not. This modular pattern formation may even occur for values of $\rho$ that are close to those observed in real systems.]{} ![image](change_modular.pdf){width="102.00000%"} ![image](class_patt.pdf){width="102.00000%"} Turing patterns on modular networks {#sec:whymodular} =================================== It has been argued that that the existence of a small-world topology in many types of networks, including brain networks, is of a crucial importance in several important processes from neuronal communication [@Sporns_Zwi] to structural robustness [@albert_barabsi]. Such functional properties are based on the short average path length that characterises this family of networks. Nevertheless these advantages also present a counter effect, that of a large spectral gap. We emphasised in the preceding subsection that the spectral gap is an important ingredient for the Turing instability. In this section, we illustrate the importance of a small spectral gap to Turing instabilities by comparing pattern formation on a Newman–Watts (NW) network (as an example of a small-world network) with pattern formation on a modular network generated using the Stochastic Block Model (SBM). [As described in [@SpornsBook; @Sporns2004; @Meunier2010], modular structure has been identified in many brain networks. Since the FitzHugh–Nagumo model [@FitzHugh1961; @Nagumo1962] is both useful for modelling neuronal dynamics  [@Murray2001], and since it can exhibit spatial pattern formation [@Murray2001; @Asllani2014Directed], we will use this model throughout this paper. In dimensionless form, FitzHugh–Nagumo dynamics correspond to using the functions $f(u,v)= u - u^3 - v$ and $g(u,v) = c(u - a + bv)$ to describe the net production of activator and inhibitor in Eq. (\[eq:turingPatternNetworks\]) where $a$, $b$, and $c$ are constants.]{} In Fig. \[fig:Mod\_vs\_NW\] we [ compare the pattern on a]{} single-module NW network (of $125$ nodes and $660$ edges) and a modular network with $5$ communities, each with $25$ nodes and a [local]{} Erdős–Rényi (ER) topology. As can be observed from the dispersion relation in Fig. \[fig:Mod\_vs\_NW\] b), the distribution of the eigenvalues of the Laplacian matrix for the NW network shows a large spectral gap. This makes the Turing instability impossible for the given choice of parameters (including $\rho = 5.5$), since the instability (i.e., values of $\Lambda$ corresponding to positive values of the continuous curve) is concentrated near the origin. We could potentially create an instability by significantly increasing $\rho$. As $t \to \infty$, the Fitzhugh–Nagumo models considered in this paper will tend to an equilibrium. One way to depict these equilibria is to plot the concentration of the activator species at long times. For the NW network described above, this is shown in Fig. \[fig:Mod\_vs\_NW\] a) and we see that the activator concentration is homogeneous across all nodes as expected. In contrast to this, for a strongly modular topology the spectrum is divided into two distinct sets of eigenvalues. The first set is those nonzero eigenvalues near the origin (of which there are $M-1$ where $M$ is the number of the modules) and the second set is composed of all the remaining eigenvalues that are far from the origin [@Peixoto]. [We note that both the NW network and the modular network have the same number of nodes and edges, so the difference between the networks’ spectra cannot be attributed to a difference in the number of nodes or in the average degree of these nodes.]{} As already anticipated, we will refer to the first set of [nonzero]{} eigenvalues as the modular eigenvalues (for example in Fig. \[fig:Mod\_vs\_NW\] d) the first four non-zero eigenvalues). [In Fig. \[fig:Mod\_vs\_NW\] d) we observe that the modular eigenvalues are sufficiently close to 0 that one of them corresponds to an instability; in Fig. \[fig:Mod\_vs\_NW\] c) we see that this leads to a pattern in the activator concentrations at equilibrium.]{} To understand the reason why the spectrum of a modular network can be divided into two subsets we should first explain the reason behind the spectral gap in small-world networks. As mentioned earlier, the denomination “small-world” [refers to a certain class of networks, one feature of which is the small average distance between nodes.]{} In [@Bojan1991], Bojan shows that the absolute value of the second largest Laplacian eigenvalue $\Lambda_2$ is bounded below by $\dfrac{4}{Nd}$, where $N$ is the number of nodes in the network and $d$ is the diameter. This means that for a fixed value of the size $N$ of the network, the spectral gap (equivalently, $|\Lambda_2|$) is larger when the diameter $d$ is smaller; thus, a NW network will have a larger spectral gap than a modular network. To further investigate how the spectral gap changes for different network topologies, we look at three different networks in Fig. \[fig:changingModularity\]. We create these networks in a simple way. [First we divide our $125$ nodes into five modules of nodes, and define the total number of intra-edges (connections within modules) and inter-edges (connections between modules). [Then we allocate each module an equal number of intra-edges and inter-edges and randomly connect nodes within and between the modules, while avoiding double entries in both cases.]{} If we define the number of intra-edges to be much larger than the number of inter-edges, then this process will yield a network with as strong modular structure.]{} We describe three examples of these networks with increasing “modularity”, where modularity is defined by the $Q$ function described in [@NewmanFiedler; @Reichardt2006]. We first look at an ER graph, as shown in Fig. \[fig:changingModularity\] $b)$. Notice that there is a large spectral gap in the corresponding dispersion relation, as shown in Fig. \[fig:changingModularity\] $a)$. [By simply modifying the ratio of inter-edges to intra-edges, we can then generate a new network which begins to close the spectral gap, as in Fig. \[fig:changingModularity\] $c)$ and $d)$.]{} Finally in Fig. \[fig:changingModularity\] $f$) we have reduced the number of inter-edges such that patterns form, and the spectral gap is greatly decreased, as in Fig. \[fig:changingModularity\] $g)$. This leads us to ask why a highly modular network closes the spectral gap so well. [Note that in the Appendix, we consider the hierarchical case where each module is arranged in a small-world fashion.]{} To understand the small spectral gap of modular networks, we first imagine a scenario in which the modules are disconnected from each other. Individually, these modules are denser and smaller than the Newman–Watts network, therefore each of them is expected to have a relatively large spectral gap. From the Perron–Frobenius theorem [@Golub] we know that the number $M$ of the connected components (the modules in this case) corresponds with the number of zero eigenvalues of the Laplacian operator. However, once the modules are connected with a small number of links then $M-1$ of these eigenvalues will move away from zero. From a spectrum perturbative analysis, we find that these become very small nonzero eigenvalues, with only one zero eigenvalue still remaining to signify that the whole modular network is connected. This explains the small size of the spectral gap in modular networks and consequently the emergence or not of Turing patterns respectively in small-world and modular networks[ [@Donetti; @Andreotti].]{} ![image](eigen_patt.pdf){width=".85\textwidth"} We notice from Figure \[fig:Mod\_vs\_NW\] $c)$ that although the pattern is highly heterogeneous at a global level, the patterns on nodes within each single module are quite homogeneous, having almost the same concentration of the species for each node in the module. Such macroscopic spatially extended patterns where densely connected entities (e.g., of biological nature) show the same amount of activity have been observed in different biological contexts [@NewmanPNAS; @metabolic; @protein] and in particular in dynamics of the brain [@Smith2018; @Oldguys]. To the best of our knowledge, we here propose the first self-organising mechanism that explains the uniformity at the module level of Turing patterns in biological networks. We can obtain insight into the patterns of $u$ and $v$ observed at equilibrium by constructing and analysing the eigenvectors associated with the Turing instabilities. From an initial condition close to the unstable homogeneous equilibrium, the rate of change in the concentrations $u$ and $v$ will initially be dominated by the eigenvector associated with the largest positive eigenvalue of the Jacobian. This initial growth will ultimately be stabilised by nonlinear terms, and we expect that the state equilibrium pattern of concentrations will be reminiscent of the eigenvectors associated with the instability [@Turing1952; @Murray2001]. To begin our analysis of the resultant patterns, we select parameters which lead to a single modular eigenvalue being positive, and observe the final “homogeneous by module” pattern as in Fig. \[fig:class\_patt\] $a)$, $b)$. The situation changes when the instability is exclusively induced from the non modular eigenvalues. In this case the concentration is no longer uniform for each module as shown in Fig. \[fig:class\_patt\] $c)$, $d)$. A hybrid state is obtained instead when both sets of eigenvalues contribute to the Turing instability as in Fig. \[fig:class\_patt\] $e)$, $f)$. These hybrid states can lead to patterns that are similar to either the modular patterns or the heterogeneous patterns. This is because the Turing instability in this case involves a competition between the eigenvectors associated with the unstable modular eigenvalues and the eigenvectors associated with the non-modular eigenvalues. The dominant instability (and therefore the eigenvector that we expect to be most similar to the equilibrium pattern) will be the eigenvector associated with the largest eigenvalue of the Jacobian. In Fig. \[fig:class\_patt\] *f)* for example, we observe that the largest eigenvalue of the Jacobian is associated with one of the modular eigenvalues of the Laplacian, and this is associated with a pattern in Fig. \[fig:class\_patt\] *e)* that could be described as almost being modular. In the Supplementary Material (SM) we discuss several criteria to establish which eigenvalue is dominating over the others. ![image](param_space.pdf){width="102.00000%"} In order to understand why the final shape of the pattern can be modular we focus on the study of the eigenvectors as plotted in Fig. \[fig:eigen\_patt\]. From the stability analysis we know that initially the pattern is shaped according to the unstable eigenvectors and this form is largely retained in the final nonlinear regime. Nevertheless, what surprises is the particular form of the eigenvectors associated with the modular eigenvalues as in Fig. \[fig:eigen\_patt\] $a)$; [in particular, the fact that the components of the modular eigenvectors are very small in all but one module.]{} To shed light on this peculiarity we will resort again to spectral perturbation theory. As anticipated earlier, the smallest non zero eigenvalue of the Laplacian $\Lambda_2$ defines the spectral gap known also in the literature as the Fiedler eigenvalue and defines the algebraic connectivity [@Fiedler; @Chung]. Its corresponding eigenvector is known as the Fiedler eigenvector and has the property that the entries of the nodes corresponding to the same modules take very similar values. Because of this property, the Fiedler eigenvector has been extensively used as the basis of several community detection methods [@NewmanFiedler; @Donetti; @Andreotti]. The other modular eigenvectors also behave in a similar manner to the Fiedler eigenvector; their entries are segregated by module [@Donetti; @Andreotti]. [Since the modular eigenvectors are often the fastest growing modes in the Turing instability, this means that the modular]{} [shape]{} of the global pattern is a consequence of the modularity of the structure of the network itself. On the other hand, when the instability is caused strictly by the non modular eigenvalues, another behaviour occurs during the pattern forming phenomenon. [This is best considered by again considering a modular network to be a perturbation of a network with initially $M$ disconnected components. In such a case, each nonzero eigenvalue of the Laplacian will correspond to an eigenvector whose components are all zero outside a single component. A modular network will be a small perturbation to this, and so the non-modular eigenvectors will also be close to zero except within a single component. If only one non-modular eigenvalue corresponds to a Turing instability, then only one module of the network will show pattern formation, as illustrated in Fig. \[fig:eigen\_patt\] b). Thus, we can predict the module on which pattern formation will occur by looking at the components of the eigenvector whose eigenvalue corresponds to the fastest growing mode of the Turing instability.]{} So far we have considered the contribution in the formation of patterns of both modular and non modular eigenvalues, however when we deal with Turing patterns in real scenarios the ratio $\rho=D_v/D_u$ is quite close to one [@Horsthemke1987; @Pearson; @exp_Turing1; @exp_Turing2]. To evaluate the conditions under which different patterns form in real conditions we now explore the parameter space of the Fitzhugh–Nagumo dynamics in more detail. In Fig. \[fig:param\_space\] it can be observed that although different types of patterns can be found in the space of the parameters $a$ and $c$, as the ratio of diffusivities gets closer to 1 the region where patterns can occur shrinks and, more importantly, the only possible Turing patterns are modular ones (indicated in green colour). [The result that brain networks have optimised their spatial interaction matrix in order to allow pattern formation has been already claimed by experimental observers [@Smith2018; @Oldguys]; to the best of our knowledge we present the first mechanism that explains the role of modularity in achieving this pattern formation.]{} ![image](zebraherd.pdf){width=".9\textwidth"} Self-organisation in real modular networks {#sec:realNetworks} ========================================== Heretofore we have discussed the role of modularity in the formation of patterns only for synthetic networks. In this part we will illustrate our findings in real examples of biological or ecological networks. The neuronal networks of several primitive animals such as nematodes have been well characterised. Indeed, it was the study of nematode neuronal networks that first inspired the development of small-world network models [@Watts1998]. In Fig. \[fig:C\_elegans\]$a)$ we show the final modular pattern of the nematode *P. pacificus* [@ppacificus]. This follows from the theoretical prediction of the unstable Fiedler eigenvector, shown in Fig. \[fig:C\_elegans\]$b)$ Here we have used the Fiedler eigenvector to identify the communities of neurons [@NewmanFiedler]. In this particular case two modules are clearly distinguishable and the level of activity of the nodes inside the modules are quite homogeneous. Other examples of Turing patterns in neuronal networks are presented in the Supplementary Material. Although the modularity of brain networks has been well-studied [@SpornsBook; @Meunier2010; @Sporns2004; @Sporns_Zwi] other types of natural networks manifest this property also. For instance, this is the case for ecological networks where the individuals are connected to each other through trophic relations [@Murray2001; @zebra]. Such modular contact networks have also been shown to be crucial for the pattern of disease spreading [@Sun; @Sun_rev]. In Fig. \[fig:C\_elegans\] $c)$ and $d)$ we present respectively the [equilibrium]{} pattern [of the FitzHugh–Nagumo equations]{} and its comparison to the unstable eigenvector of [the contact network of]{} a zebra herd [@zebra] where a community of $11$ individuals out of a total of $23$ is clearly visible. However, the formation of patterns of spreading are not limited only to contact networks, which in general can be small in size. Modularity is a common property in other types of networks which, although they are not directly related to biological systems, are still essential for biological phenomena occurring on them. This is for instance, the case for networks of human mobility, such the roads networks in the city of Chicago presented in the SM [@konectLink; @konectArticle1; @konectArticle2; @konectProceedings], which are decisive for the spreading of an epidemics in the entire urban area [@Sun; @Sun_rev]. These examples all show agreement with the mathematical analysis we have shown so far. Discussion and conclusions ========================== In this paper we have analytically and numerically explored pattern formation on modular networks. We have shown that modularity, a ubiquitous topological feature of many biological networks, is crucial for the self-organisation of the global dynamics on a network. To study this behaviour we have considered here the Turing instability as a paradigmatic mechanism for pattern formation in biology, ecology or neuroscience. [The possibility of pattern formation via the Turing mechanism]{} on non-modular networks is limited to unrealistically extreme ratios of the diffusion constants of the activator and inhibitor species making the small spectral gap of the Laplacian matrix a fundamental requirement for the Turing instability. This feature is a structural advantage of modular networks which follows from spectral perturbation theory. A strongly modular network can be considered as a set of connected components weakly attached with [a small number of intermodule links. From spectral perturbation theory this yields a number – equal to one fewer than the number of modules – of non zero eigenvalues very near to the origin.]{} This characterisation at the linear stability level influences the shape of the spatially extended patterns. Due to the segregation of the entries of the eigenvectors corresponding to the set of modular eigenvalues, we are able to explain why Turing patterns are homogeneous per module on these networks. This result opens to an important aspect regarding the functional resolution of the brain modes which was hypothesised [@Sporns2016; @Hutt; @Meunier2010] in several experimental observations [@Smith2018; @Oldguys]. To the best of our knowledge, the model we present here constitutes the first self-organising mechanism where the modules are presented as functional blocks of biological networks. In this sense, we argue that the module is the smallest spatial unit to be taken into account from the functional point of view [i.e. if we “zoom” out far enough from a modular network, the individual modules behave like individual supernodes.]{} For the particular example of the brain the modules might be the super-nodes of the functional connectomes [@Sporns2004; @Sporns2016; @Smith2018]. Indeed, the (self-)segregation of the network structure in modules [@NewmanPNAS] influences also the shape of the dynamical pattern on it. Based on the fact (see [@Meunier2010; @Hutt; @Sporns2004] and Fig.\[fig:C\_elegans\]) that in real scenarios Turing patterns should be exclusively modular, we believe that the results we have shown here can be potentially used to formulate a community detection protocol [@NewmanPNAS; @NewmanFiedler; @NewmanBook] in the case where patterns of self-organised activity are known to exist. In the case when we relax Turing conditions to allow the instability for the non-modular part of the spectrum, then we can [use the eigenvector corresponding to the largest eigenvalue to indicate]{} the module in which the Turing pattern is first seeded before finally spreading to the rest of the network. This behaviour can potentially make the pattern formation process a powerful diagnostic tool for studying and eventually controlling the emergence of abnormal dynamics which characterise many neurological diseases [@AsllaniPLOS] or the spread of an epidemic in a group of individuals [@Sun; @Sun_rev]. [We test our theoretical results on several real connection data sets of neuronal, ecological and infrastructure networks verifying the correctness of our findings, that modularity is crucial for the development of patterns, and that when the instability is derived from the first set of modular eigenvalues, that the resultant self-organisation follows the modular structure of the network.]{} The results we have presented here can extend also to more complicated scenarios. This is, for example, the case when the hierarchy of a network is considered as a complement to its modularity. In the Appendix we show that in a hierarchical modular network the modular eigenvalues are even more relevant for the Turing pattern forming process. Further extensions of our approach are also possible; for example to consider the effect of directed edges in a modular network. In this case we expect a richer dynamics where travelling Turing waves should emerge in a directed modular networks [@Asllani2014Directed]. Acknowledgement {#acknowledgement .unnumbered} =============== B. A. Siebert. acknowledges funding from the Irish Research Council under grant GOIPG/2018/3026. This work is partly funded by Science Foundation Ireland (grant numbers 16/IA/4470, 16/RC/3918, 12/RC/2289 P2, 18/CRT/6049) and co-funded under the European Regional Development Fund. Appendix {#appendix .unnumbered} ======== The FitzHugh–Nagumo model ------------------------- [We have used the Fitzhugh–Nagumo model throughout this paper[@FitzHugh1961; @Nagumo1962],]{} which is one of the first and most well known mathematical models used to describe the spiking dynamics of neurons. In terms of mathematical equations the behaviour of a single neuron is described by $$\begin{aligned} \frac{d u}{d t} &=& u - u^3 - v\nonumber\\ \frac{d v}{d t} &=& c(u - av - b)\end{aligned}$$ where $u$ is the membrane potential and $v$ the recovery variable. The model itself was first introduced by FitzHugh [@FitzHugh1961] to explain the generation of spikes in excitable systems, i.e., neurons. A spike is a short-lasting elevation of the membrane voltage $u$ diminished over time by a slower and linear recovery variable $v$ once the system is periodically excited by an external current. The following year Nagumo *et al.* [@Nagumo1962] developed the electric circuit which mimics such behaviour. However, although the model itself is mainly used to describe the oscillatory behavior of neurons, it also admits a stable fixed point, which is a necessary requirement [for Turing instabilities]{}. Once this model is equipped with a diffusion term, it turns out in a perfect candidate for pattern formation [@Murray2001]. In recent years, with the rapid development of network science, the Fitzhugh–Nagumo model has been extended to diffusively coupled networks [@Perc2005; @Asllani2014Directed]. The role of hierarchy of the brain networks in the pattern formation {#subsec:appendixHierarchy} -------------------------------------------------------------------- We have discussed the role that modularity has on pattern formation, isolating it from other important features such as the small-world property, which is in fact an integral aspect of many networks, including brain networks. So then a question that arises naturally is, how does the brain cope with maintaining both features and their functional properties at the same time? We now are able to answer this question by recalling an important empirical results that characterises most real networks, their hierarchical structure [@hierarchical; @NewmanBook; @Meunier2010]. In fact, most of the connectomes studied are organised in a modular structure, however each module is further organised in a small-world fashion. In a hierarchical modular network the entire network is organised in modules which are attached to each other so as to have a small diameter and at the same time the nodes in the modules are connected in such way to form sub-modules again minimising their diameter and this process goes on this way up to smallest building unity, the single nodes. A hierarchical structure stresses once more the necessity of modularity for the self-organising phenomena in the networks. In Fig. \[fig:hierar\_mod\] we show that the difference of the smallest non modular eigenvalue from the origin is larger when the modules have a small-world topology compared to when they are organised at random (e.g. ER network) for the same number of nodes, edges and modules. The reason for this can be found once more by taking a perturbative approach. The spectral gap of an individual module (disconnected from the rest of the network) is larger when its diameter is smaller, as it is in the Newman–Watts network used in Fig. \[fig:hierar\_mod\]. ![image](hierar_mod.pdf){width=".9\textwidth"} Thus, in the presence of hierarchy, the cyan and the yellow regions in Fig. \[fig:param\_space\] would be even smaller making the modularity region shown in green larger compared to the previous two. We notice, however, that the instability invariance is still valid for values of the diffusivites ratio $\rho$ near to $1$, that is when only the green region in the parameter space is available. In conclusion, a hierarchical arrangement where each module is arranged in a small world fashion, and these modules are again connected in a small world fashion, are even better candidates for forming modular patterns, than then modular networks studied in the main part of this paper. [^1]: The definition of the spectral gap depends on the way one defines the Laplacian matrix [@NewmanBook]. In our case the spectrum of the Laplacian is non positive.
{ "pile_set_name": "ArXiv" }
--- abstract: 'This article considers the relation between the spanning properties of lattice orbits of discrete series representations and the associated lattice co-volume. The focus is on the density theorem, which provides a trichotomy characterizing the existence of cyclic vectors and separating vectors, and frames and Riesz sequences. We provide an elementary exposition of the density theorem, that is based solely on basic tools from harmonic analysis, representation theory, and frame theory, and put the results into context by means of examples.' address: - | Faculty of Mathematics, University of Vienna, Oskar-Morgenstern-Platz 1, A-1090 Vienna, Austria\ and Acoustics Research Institute, Austrian Academy of Sciences, Wohllebengasse 12-14 A-1040, Vienna, Austria - 'Faculty of Mathematics, University of Vienna, Oskar-Morgenstern-Platz 1, A-1090 Vienna, Austria' author: - José Luis Romero - Jordy Timo van Velthoven date: - - title: The density theorem for discrete series representations restricted to lattices --- [ ]{} Introduction ============ Let $G$ be a second countable locally compact group and let $(\pi, {\mathcal{H}_{\pi}})$ be an irreducible, square-integrable unitary representation of $G$, a so-called *discrete series representation*. For a lattice $\Gamma \subset G$, we consider the relation between certain spanning properties of lattice orbits of $\pi$ under a vector $g \in {\mathcal{H}_{\pi}}$, $$\begin{aligned} \label{eq:orbit_intro} \pi (\Gamma) g = \big\{ \pi (\gamma) g : \gamma \in \Gamma \big\},\end{aligned}$$ and the *lattice co-volume* $\operatorname*{vol}(G/\Gamma)$ of $\Gamma$, i.e., the volume of a fundamental domain of $\Gamma$. The spanning properties that we consider are the existence of cyclic, separating, frame and Riesz vectors; see Section \[sec:vectors\_frames\] for the precise definitions. The notions of cyclic and separating vectors occur primarily in the theory of operator algebras, in particular, von Neumann algebras, and they provide (if they exist) a powerful tool in studying the structure of these algebras. The stronger notions of frames and Riesz sequences, on the other hand, form the core of Gabor and wavelet theory, and are important in applications as they guarantee unconditionally convergent and stable Hilbert space expansions. The central theorem relating the spanning properties of systems and the corresponding lattice co-volume is referred to as the *density theorem*. Under the assumption that the lattice $\Gamma$ is an infinite conjugacy class (ICC) group, i.e., any conjugacy class $ \{\gamma \gamma_0 \gamma^{-1} \; | \; \gamma \in \Gamma \}$ for $\gamma_0 \in \Gamma \setminus \{e\}$ has infinite cardinality, the density theorem provides the following trichotomy: \[thm:intro\_ICC\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series representation of a second countable unimodular group $G$ of formal dimension $d_{\pi} > 0$. Suppose $\Gamma \subset G$ is an ICC lattice. Then the following assertions hold: (i) If $\operatorname*{vol}(G / \Gamma) d_{\pi} < 1$, then $\pi|_{\Gamma}$ admits a Parseval frame, but neither a separating vector, nor a Riesz sequence; (ii) If $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$, then $\pi|_{\Gamma}$ admits an orthonormal basis; (iii) If $\operatorname*{vol}(G / \Gamma) d_{\pi} > 1$, then $\pi|_{\Gamma}$ admits an orthonormal system, but not a cyclic vector. (While $d_{\pi}$ and $\operatorname*{vol}(G/\Gamma)$ depend on the normalization of the Haar measure on $G$, their product $\operatorname*{vol}(G/\Gamma) d_{\pi}$ does not.) The density theorem characterizes the spanning properties of the lattice orbits in terms of the lattice co-volume or its reciprocal, often called the *density* of the lattice. In the setting of a general unimodular group, the assumption that the lattice is ICC is essential and cannot be omitted —see Example \[sec:SL\_ICC\] below— although a more general version of Theorem \[thm:intro\_ICC\] for possibly non-ICC lattices was obtained by Bekka [@bekka2004square]. The existence claims in Theorem \[thm:intro\_ICC\] are not accompanied by constructions of explicit vectors. The criteria for the existence of cyclic and separating vectors in Theorem \[thm:intro\_ICC\] are well-known to be consequences of the general theory underlying the so-called *Atiyah-Schmid formula* [@atiyah1976elliptic; @atiyah1977geometric; @goodman1989coxeter], and, for certain classes of representations, also a consequence of Rieffel’s work [@rieffel1981von; @rieffel1988projective]. The stronger statements on the existence of Parseval frames (part (i)) and orthonormal bases (part (ii)) can also be obtained by similar techniques as shown by Bekka [@bekka2000square; @bekka2004square]. The statement on orthonormal systems (part (iii)) does not seem to have explicitly occurred in the literature before. While the interest in the density theorem is broad and manifold, as it is encompasses operator algebras, representation theory, mathematical physics, and Gabor and wavelet analysis, the available proofs rely on advanced theory of von Neumann algebras, and may only be accessible to a smaller community of experts. This expository article provides an elementary and self-contained presentation of the density theorem, that is based solely on basic tools from harmonic analysis, representation theory, and frame theory, and should be accessible to an interested non-expert. While almost all methods employed exist in some antecedent form in the different specialized literatures, their particular combination here makes the basic structure underlying the density theorem transparent; see Section \[sec:technical\_overview\]. The elementary arguments in this article fall, however, short of deriving the more general version of Theorem \[thm:intro\_ICC\] by Bekka [@bekka2004square]. We hope that this article motivates the non-specialist to delve deeper into operator-algebraic methods. We also expect that the concrete exposition contributes to the study of quantitative aspects of Theorem \[thm:intro\_ICC\], such as the relation between the distance between $\operatorname*{vol}(G / \Gamma) d_{\pi}$ and the critical value $1$, and special qualities of the corresponding cyclic or separating vectors, such as smoothness in the case of Lie groups. Context and related work ------------------------ In the setting of Theorem \[thm:intro\_ICC\], for any non-zero $g \in {\mathcal{H}_{\pi}}$, the system $\{ \pi (x) g : x \in G \}$ is *overcomplete*, i.e., it contains proper subsystems that are still complete. The fundamental question as to whether subsystems corresponding to lattices remain complete was posed by Perelomov in his group-theoretical approach towards the construction of coherent states [@perelomov1972coherent; @perelomov1986generalized]. In fact, a criterion for the completeness of subsystems of coherent states similar to Theorem \[thm:intro\_ICC\] was posed as a question in [@perelomov1972coherent p.226] [^1]. These criteria have been considered for specific systems and vectors in, e.g., [@bargmann1971on; @perelomov1971remark; @perelomov1973coherent; @klauder1994wavelets; @neretin2006perelomov; @ramanathan1995incompletness; @monastyrsky1974coherent; @groechenig2016completeness]. The related question as to whether a system is a (discrete) frame is at the core of modern frame theory [@daubechies1986painless] and has, in particular, a long history in Gabor theory [@heil2007history]. The existence of a frame vector is also studied in representation theory, in whose jargon such a vector is called *admissible* [@fuehr2005abstract; @fuehr2002admissible]. While the mere existence of a frame or Riesz vector for a given lattice is quite different from the validity of these properties for one specific vector, there is an interesting interplay between the two problems. In Section \[sec:examples\] we discuss a selection of examples, including one where Theorem \[thm:intro\_ICC\] yields seemingly unnoticed consequences. Projective versions ------------------- The density theorem can also be formulated for *projective* unitary representations [@bekka2004square; @radulescu1998berezin; @han2017note; @gabardo2003frame], and allows for applications to representations that are square-integrable only modulo a central subgroup (as in the case of nilpotent or reductive Lie groups). The proofs that we present work transparently for projective representations and we formulate the main results in that generality in Theorem \[th\_main\]. In the projective setting, the lattice is not assumed to be ICC, but is assumed to satisfy the weaker *Kleppner condition* [@kleppner1962structure], a compatibility condition between the lattice and the cocycle of the projective representation. The projective formulation greatly simplifies the treatment of concrete examples such as weighted Bergman spaces and Gabor systems in Section \[sec:examples\]. Technical comments {#sec:technical_overview} ------------------ The common approach to the density theorem is through the coupling theory of von Neumann algebras, and a self-contained presentation in this spirit can be found in [@goodman1989coxeter; @bekka2000square]. Although we make no explicit reference to the coupling theory, some of the arguments we give are simplifications of standard results, as we point out throughout the text. Most significantly, we circumvent certain technicalities associated with the so-called trace of a group von Neumann algebra. In finding elementary arguments, we benefited particularly from reading [@cowling1991irreducibility; @alaoglu1940ergodic; @kuhn1992restrictions; @fuehr2005abstract; @rieffel2004integrable]. An important simplification in the proof of Theorem \[thm:intro\_ICC\] occurs in the derivation of the necessity of the volume or density conditions for cyclicity and separateness, which also play an essential role in deriving the existence of frame and Riesz vectors. Our argument is inspired by Janssen’s “classroom proof” of the density theorem for Gabor frames [@janssen_density], and underscores the power of frame-theoretic methods. In this article such argument is pushed further to yield consequences for cyclicity and separateness. While the necessity of the density conditions for frames and Riesz sequences is an active field of research [@balan2006density; @fuehr2017density; @mitkovski1], most abstract results are not applicable to groups of non-polynomial growth. It is therefore remarkable that the particular lattice structure of the systems in question leads to simple and conclusive results. Preliminaries {#sec:prelim} ============= Throughout the article, the locally compact group $G$ is assumed to be second countable and unimodular. We fix a Haar measure $\mu_G$ on $G$. Some of the notions below depend on this normalization, but the main results do not. Cocycles and projective representations --------------------------------------- A *cocycle* or *multiplier* on $G$ is a Borel measurable function $\sigma : G \times G \to \mathbb{T}$ such that (i) For all $x,y,z \in G$, $\sigma(x, yz) \sigma(y,z) = \sigma(xy,z) \sigma(x,y);$ (ii) For the identity $e \in G$ and all $x \in G$, $\sigma(x,e) = \sigma(e,x) = 1$. A *projective unitary representation* $(\pi, {\mathcal{H}_{\pi}})$ of $G$ on a Hilbert space ${\mathcal{H}_{\pi}}$ is a mapping $\pi : G \to \mathcal{U} ({\mathcal{H}_{\pi}})$ satisfying the following conditions: (i) The map $x \mapsto \pi(x)$ is weakly measurable, i.e., the map $G \ni x \mapsto \langle \pi(x) f, g \rangle \in \mathbb{C}$ is Borel for all $f, g \in {\mathcal{H}_{\pi}}$; (ii) There exists a function $\sigma : G \times G \to \mathbb{T}$ such that $\pi(x) \pi(y) = \sigma(x,y) \pi (xy)$ for all $x, y \in G$; (iii) $\pi (e) = I$. In this case, the map $\sigma$ in (ii) is uniquely determined and it is a cocycle. A projective unitary representation with cocycle $\sigma$ is called a *$\sigma$-representation*. Given two $\sigma$-representations $(\pi_1, \mathcal{H}_{\pi_1})$ and $(\pi_2, \mathcal{H}_{\pi_2})$, a linear operator $T : \mathcal{H}_{\pi_1} \to \mathcal{H}_{\pi_2}$ is said to *intertwine* $\pi_1$ and $\pi_2$ if $$T \pi_1 (x) = \pi_2 (x) T, \qquad \mbox{for all }x \in G.$$ See [@mackey1958unitary] and [@varadarajan1985geometry Chapter VII] for more on cocycles and projective representations. Square-integrable $\sigma$-representations ------------------------------------------ Let $(\pi, {\mathcal{H}_{\pi}})$ be a $\sigma$-representation of $G$. For $f, g \in {\mathcal{H}_{\pi}}$, the associated *matrix coefficient* is defined by $ C_{g} f (x) = \langle f, \pi(x) g \rangle$ for $x \in G. $ The $\sigma$-representation $(\pi, {\mathcal{H}_{\pi}})$ is called *square-integrable* if there exists a norm dense subspace $\mathcal{D} \subset {\mathcal{H}_{\pi}}$ such that $$\begin{aligned} \label{eq:squareintegrable_dense} C_g f = \langle f, \pi (x) g \rangle \in L^2 (G), \quad f \in {\mathcal{H}_{\pi}}, \; g \in \mathcal{D}. \end{aligned}$$ The $\sigma$-representation $(\lambda_G^{\sigma}, L^2 (G))$ given by $$\begin{aligned} (\lambda_G^{\sigma} (y) F)(x) = \sigma(y, y^{-1} x) F(y^{-1} x), \qquad F \in L^2 (G),\, x,y \in G,\end{aligned}$$ is called the *$\sigma$-regular representation* and satisfies the *covariance property* or *intertwining property*: $$\begin{aligned} \label{eq:covariance} C_g (\pi(y) f)(x) = \sigma(y, y^{-1} x) C_g f (y^{-1} x) = \big( \lambda_G^{\sigma} (y) C_g f\big) (x), \quad x,y \in G, \end{aligned}$$ for all $f \in {\mathcal{H}_{\pi}}$, $g \in \mathcal{D}$. A $\sigma$-representation $(\pi, {\mathcal{H}_{\pi}})$ is called *irreducible* if the only closed $\pi(G)$-invariant subspaces of ${\mathcal{H}_{\pi}}$ are $\{0\}$ and ${\mathcal{H}_{\pi}}$ and is said to be a *discrete series $\sigma$-representation* if it is both square-integrable and irreducible. Given a discrete series $\sigma$-representation $(\pi, {\mathcal{H}_{\pi}})$, there exists a unique number $d_{\pi} > 0$, called the *formal dimension* of $\pi$, such that the *orthogonality relations* $$\begin{aligned} \label{eq:ortho_rel} \int_G \langle \pi (x) f_1, g_1 \rangle \overline{\langle \pi(x) f_2, g_2 \rangle} \; d\mu_G (x) = d_{\pi}^{-1} \langle f_1, f_2 \rangle \overline{\langle g_1, g_2 \rangle}\end{aligned}$$ hold for all $f_1, f_2, g_1, g_2 \in {\mathcal{H}_{\pi}}$. See [@radulescu1998berezin; @robert1983introduction] and [@neeb2000holomorphy Appendix VII]. Fundamental domains and lattices -------------------------------- Let $\Gamma \subseteq G$ be a discrete subgroup. A left (resp. right) fundamental domain of $\Gamma$ in $G$ is a Borel set $\Omega \subseteq G$ satisfying $G = \Gamma \cdot \Omega$ and $\gamma \Omega \cap \gamma' \Omega = \emptyset$ (resp. $G = \Omega \cdot \Gamma$ and $\Omega \gamma \cap \Omega \gamma' = \emptyset$) for all $\gamma, \gamma' \in \Gamma$ with $\gamma \neq \gamma'$. If $\Omega$ is a left (resp. right) fundamental domain, then $\Omega^{-1}$ is a right (resp. left) fundamental domain. The discrete subgroup $\Gamma \subseteq G$ is called a *lattice* if it admits a left (or right) fundamental domain of finite measure. Equivalently, a discrete subgroup $\Gamma$ is a lattice if and only if the quotient $G / \Gamma$ admits a finite $G$-invariant regular Borel measure. Any two fundamental domains have the same measure, and thus, we may define the *co-volume* of $\Gamma$ as $\operatorname*{vol}(G/\Gamma) := \mu_G (\Omega)$. This depends of course on the choice of the Haar measure for $G$. See [@raghunathan1972discrete] and [@bekka2008kazhdan Appendix B] for more on lattices and fundamental domains. ICC groups and Kleppner’s condition ----------------------------------- Let $\Gamma$ be a discrete countable group and let $\sigma : \Gamma \times \Gamma \to \mathbb{T}$ be a cocycle. An element $\gamma_0 \in \Gamma$ satisfying $\sigma(\gamma_0, \gamma) = \sigma(\gamma, \gamma_0)$ for all elements $\gamma \in \Gamma$ commuting with $\gamma_0$ is called *$\sigma$-regular*. The pair $(\Gamma, \sigma)$ is said to satisfy *Kleppner’s condition* if the conjugacy class $C_{\gamma_0} := \{ \gamma \gamma_0 \gamma^{-1} \; | \; \gamma \in \Gamma\}$ of any $\sigma$-regular element $\gamma_0 \in \Gamma \setminus \{e\}$ is infinite. The group $\Gamma$ is called an *infinite conjugacy class* (ICC) group if any conjugacy class $C_{\gamma_0}$ for $\gamma_0 \in \Gamma \setminus \{e\}$ is infinite. Any ICC group $\Gamma$ satisfies Kleppner’s condition for any cocycle $\sigma : G \times G \to \mathbb{T}$. Von Neumann algebras {#sec_vn} -------------------- Let $\mathcal{H}$ be a separable complex Hilbert space. A net $(T_{\alpha} )_{\alpha \in \Lambda}$ of bounded linear operators $T_\alpha \in \mathcal{B}(\mathcal{H})$ converges in the strong operator topology (SOT) to an operator $T \in \mathcal{B}(\mathcal{H})$ if $T_\alpha f \longrightarrow Tf$ in the norm of $\mathcal{H}$ for all $f \in \mathcal{H}$, and it converges in the weak operator topology (WOT) if ${\ensuremath{\left<T_\alpha f,g\right>}} \longrightarrow {\ensuremath{\left<Tf,g\right>}}$ for all $f,g \in \mathcal{H}$. A subalgebra $\mathcal{A} \subseteq \mathcal{B}(\mathcal{H})$ is called a *von Neumann algebra* if $\mathcal{A}$ is self-adjoint, i.e., $\mathcal{A}=\mathcal{A}^*$, contains the identity $I$ and is weakly closed in $\mathcal{B}(\mathcal{H})$. The *commutant* $M'$ of a set $M \subseteq \mathcal{B}(\mathcal{H})$ is the class of all bounded linear operators that commute with each operator of $M$, i.e., $$M' := \{ T \in \mathcal{B}(\mathcal{H}) \; : \; TS = ST, \; \forall S \in M \}.$$ By von Neumann’s density theorem, it follows that if $\mathcal{A} \subseteq \mathcal{B}(\mathcal{H})$ is a self-adjoint algebra containing the identity, then $\mathcal{A}'' := (\mathcal{A}')'$ is contained in the strong closure of $\mathcal{A}$ in $\mathcal{B}(\mathcal{H})$. In particular, the double commutant $\mathcal{A}''$ is the smallest von Neumann algebra containing $\mathcal{A}$ and equals the strong and weak closure of $\mathcal{A}$. Thus, for every operator $T \in \mathcal{A}''$, there exist a net of operators of $\mathcal{A}$ converging to $T$ in the SOT topology. Moreover, by Kaplansky’s density theorem, the net may be assumed to be uniformly bounded in operator norm. For a family of operators $\mathcal{A} \subset \mathcal{H}$ and a vector $g \in \mathcal{H}$, the closed linear span of $\mathcal{A} g = \{ A g \; : \; A \in \mathcal{A}\}$ in $\mathcal{H}$ is denoted by $[\mathcal{A}g] := \overline{\operatorname*{span}\mathcal{A} g}$. Given a von Neumann algebra $\mathcal{A} \subseteq \mathcal{B}(\mathcal{H})$ and an orthogonal projection $P_{\mathcal{K}}$ onto a closed subspace $\mathcal{K} \subseteq \mathcal{H}$, the space $\mathcal{K}$ is invariant under $\mathcal{A}$, i.e., $\mathcal{A}(\mathcal{K}) \subset \mathcal{K}$, if and only if $P_{\mathcal{K}} \in \mathcal{A}'$. This observation is known as the *projection lemma*. For more background on von Neumann algebras, see [@kadison1983fundamentals1; @dixmier1981vonneumann]. Partial isometries and the polar decomposition ---------------------------------------------- Let $\mathcal{H}$ and $\mathcal{K}$ be complex Hilbert spaces. A bounded linear operator $U : \mathcal{K} \to \mathcal{H}$ is called a *partial isometry* if $U$ is an isometry when restricted to the orthogonal complement $\mathcal{N}(U)^{\perp}$ of its null space $\mathcal{N}(U)$. The subspace $\mathcal{N}(U)^{\perp}$ is called the *initial space* of $U$ and the range $\mathcal{R} (U)$ of $U$ is the *final space* of $U$, i.e., the image of $\mathcal{N}(U)^{\perp}$ under the isometry $U|_{\mathcal{N} (U)^{\perp}}$ A linear operator $T : \operatorname*{dom}(T) \subset \mathcal{H} \to \mathcal{K}$ is *densely defined* if its domain $\operatorname*{dom}(T)$ is a norm dense subspace in $\mathcal{H}$ and is called *closed* if its graph $\mathcal{G}(T) := \{ (f, Tf) \; | \; f \in \mathcal{H} \}$ is closed in $\mathcal{H} \oplus \mathcal{K}$. For a closed, densely defined linear operator $T : \operatorname*{dom}(T) \subset \mathcal{H} \to \mathcal{K}$, its *adjoint* is denoted by $T^*$ and its *modulus* by $|T| := (T^* T)^{1/2}$. The operator $|T|$ is defined by Borel functional calculus and has domain $\operatorname*{dom}(|T|)=\operatorname*{dom}(T)$. The *polar decomposition* of $T$ is uniquely given by $$T = U_T |T| = |T^*| U_T,$$ where $U_T : \mathcal{H} \to \mathcal{K}$ is a partial isometry with initial space $\mathcal{N}(T)^{\perp} = \overline{\mathcal{R}(|T|)}$ and final space $\overline{\mathcal{R}(T)}$. For more details and background, see e.g. [@fell1988representations1 VI, Section 13]. Orbits of square-integrable representations {#sec:vectors_frames} =========================================== Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$ on a separable (complex) Hilbert space ${\mathcal{H}_{\pi}}$. For a vector $g \in {\mathcal{H}_{\pi}}$, we consider the orbit $\pi(\Gamma) g$ of $g$ under $(\pi, {\mathcal{H}_{\pi}})$, i.e., $${\pi (\Gamma) g}:= \big\{ \pi (\gamma) g : \gamma \in \Gamma \big\}.$$ We treat the system ${\pi (\Gamma) g}$ as a family indexed by $\Gamma$ and allow for repetitions. Cyclic and separating vectors {#sec_comp_sep} ----------------------------- A vector $g \in {\mathcal{H}_{\pi}}$ is called *cyclic* or *complete* if $[\pi(\Gamma) g] = {\mathcal{H}_{\pi}}$. By von Neumann’s density theorem, the vector $g \in {\mathcal{H}_{\pi}}$ is cyclic if and only if $[\pi (\Gamma)'' g ] = {\mathcal{H}_{\pi}}$. A vector $g \in {\mathcal{H}_{\pi}}$ is called *separating* for $\pi (\Gamma)''$ if $T \in \pi (\Gamma)''$ and $Tg = 0$ imply $T = 0$, that is, if the map $\pi(\Gamma)'' \ni T \mapsto Tg \in {\mathcal{H}_{\pi}}$ is injective. A vector $g \in {\mathcal{H}_{\pi}}$ is separating for $\pi(\Gamma)''$ if and only if $[\pi(\Gamma)' g]= {\mathcal{H}_{\pi}}$. Indeed, if $[\pi (\Gamma)'g] \neq {\mathcal{H}_{\pi}}$, then the projection $P_{\mathcal{K}}$ onto $\mathcal{K} := [\pi(\Gamma)' g]$ is in $\pi(\Gamma)''$ and $P_{\mathcal{K}} \neq I$. Thus $I-P_{\mathcal{K}} \neq 0$ and $(I - P_{\mathcal{K}} )g = 0$, showing that $g$ is not separating for $\pi(\Gamma)''$. Conversely, if $[\pi(\Gamma)' g] = {\mathcal{H}_{\pi}}$ and $T \in \pi(\Gamma)''$ is such that $T g = 0$, then $0 = S Tg = T S g$ for all $S \in \pi(\Gamma)'$, and hence $T = 0$ since $\pi(\Gamma)' g$ is norm dense in ${\mathcal{H}_{\pi}}$. Intuitively, a vector $g \in {\mathcal{H}_{\pi}}$ is cyclic if the corresponding orbit $\pi(\Gamma) g$ is rich enough so as to provide approximations for every vector in ${\mathcal{H}_{\pi}}$. On the other hand, if $g$ is separating for $\pi(\Gamma)''$, then $\pi (\Gamma)''$ cannot be too rich, because $\pi(\Gamma)'' \ni T \mapsto Tg \in {\mathcal{H}_{\pi}}$ is injective. The central question of this article is the relation between the existence of cyclic and separating vectors on the one hand, and the co-volume of $\Gamma$ within a larger group $G$. As a key tool, we consider certain strengthened notions of cyclicity and separation. Frames and Riesz sequences -------------------------- A system ${\pi (\Gamma) g}$ is called a *frame* for ${\mathcal{H}_{\pi}}$ if there exist constants $A, B > 0$, called *frame bounds*, such that the following *frame inequalities* hold: $$\begin{aligned} \label{eq_frame_bounds} A \| f \|_{{\mathcal{H}_{\pi}}}^2 \leq \sum_{\gamma \in \Gamma} |\langle f, \pi (\gamma) g \rangle |^2 \leq B \| f \|_{{\mathcal{H}_{\pi}}}^2, \qquad f \in {\mathcal{H}_{\pi}}.\end{aligned}$$ A vector $g$ is a *frame vector* if $\pi(\Gamma) g$ is a frame. A system ${\pi (\Gamma) g}$ forming a frame is complete by the first (lower) bound in . The second of the frame inequalities (upper bound), $$\begin{aligned} \label{eq_Bessel} \sum_{\gamma \in \Gamma} |\langle f, \pi (\gamma) g \rangle |^2 \leq B \| f \|_{{\mathcal{H}_{\pi}}}^2 , \qquad f \in {\mathcal{H}_{\pi}},\end{aligned}$$ is known as a *Bessel bound*. A vector $g$ satisfying is a *Bessel vector*. Note that the definition concerns $\pi(\Gamma) g$ as an indexed family. Two indexations of the same underlying set can have, for example, different frame bounds. The Bessel condition is equivalent to the *frame operator* $$S_{g, \Gamma} : {\mathcal{H}_{\pi}}\to {\mathcal{H}_{\pi}}, \; S_{g, \Gamma} f = \sum_{\gamma \in \Gamma} \langle f, \pi(\gamma) g \rangle \pi(\gamma) g$$ being well-defined and bounded. The full two-sided frame inequality is equivalent to the frame operator being a positive-definite (bounded, invertible) operator on ${\mathcal{H}_{\pi}}$. A frame ${\pi (\Gamma) g}$ for which the frame bounds can be chosen as $A=B=1$ is called a *Parseval frame*, because it gives the identity $$\begin{aligned} \| f \|_{{\mathcal{H}_{\pi}}}^2 = \sum_{\gamma \in \Gamma} |\langle f, \pi (\gamma) g \rangle |^2, \qquad f \in {\mathcal{H}_{\pi}}.\end{aligned}$$ Equivalently, ${\pi (\Gamma) g}$ is a Parseval frame for ${\mathcal{H}_{\pi}}$ if and only if its frame operator $S_{g,\Gamma}$ is the identity on ${\mathcal{H}_{\pi}}$. Whenever well-defined and bounded, the frame operator $S_{g, \Gamma}$ commutes with $\pi(\gamma)$ for all $\gamma \in \Gamma$. \[rem\_parsevalization\] An arbitrary frame ${\pi (\Gamma) g}$ can be turned into a Parseval frame by considering $\tilde g := S_{g, \Gamma}^{-1/2} g$. Indeed, if ${\pi (\Gamma) g}$ is a frame, then $S_{g, \Gamma}$ is a positive operator, and, therefore, $\tilde g$ is well-defined. Moreover, since $S_{g, \Gamma}^{-1/2}$ also commutes with each $\pi(\gamma)$, for $f \in {\mathcal{H}_{\pi}}$, $$\begin{aligned} S_{\tilde g, \Gamma} f = \sum_{\gamma \in \Gamma} \langle f, \pi(\gamma) S_{g, \Gamma}^{-1/2} g \rangle \pi(\gamma) S_{g, \Gamma}^{-1/2}g = S_{g, \Gamma}^{-1/2} S_{g, \Gamma} S_{g, \Gamma}^{-1/2} f = f,\end{aligned}$$ showing that $\pi(\Gamma) \tilde{g}$ is a Parseval frame for ${\mathcal{H}_{\pi}}$. A system ${\pi (\Gamma) g}$ is called a *Riesz sequence* in ${\mathcal{H}_{\pi}}$ if there exist constants $A, B > 0$, called *Riesz bounds*, such that $$A \| c \|_{\ell^2}^2 \leq \bigg\| \sum_{\gamma \in \Gamma} c_{\gamma} \pi (\gamma) g \bigg\|_{{\mathcal{H}_{\pi}}}^2 \leq B \| c \|_{\ell^2}^2, \quad c = (c_{\gamma})_{\gamma \in \Gamma} \in \ell^2 (\Gamma).$$ A duality argument, shows that a Riesz sequence satisfies the Bessel bound . Moreover, a Riesz sequence is linearly independent and $\omega$-independent, and hence cannot admit repetitions. A vector $g$ yielding a Riesz sequence $\pi(\Gamma) g$ is a *Riesz vector*. A complete Riesz sequence ${\pi (\Gamma) g}$ is called a *Riesz basis* for ${\mathcal{H}_{\pi}}$. Equivalently, a system ${\pi (\Gamma) g}$ is a Riesz basis for ${\mathcal{H}_{\pi}}$ if it is the image of an orthonormal basis under a bounded, invertible operator on ${\mathcal{H}_{\pi}}$. If $\pi(\Gamma) g$ is a Riesz basis for ${\mathcal{H}_{\pi}}$, then $\pi(\Gamma) g$ and $ \pi(\Gamma) S^{-1}_{g,\Gamma} g = S^{-1}_{g,\Gamma} \pi(\Gamma) g$ are biorthogonal sequences in ${\mathcal{H}_{\pi}}$, i.e., $\langle \pi(\gamma') g, S_{g,\Gamma}^{-1} \pi(\gamma) g \rangle = \delta_{\gamma', \gamma}$ for $\gamma, \gamma' \in \Gamma. $ It will be shown in Proposition \[prop\_riesz\_sep\] that, under Kleppner’s condition, if $\pi(\Gamma)g$ is a Riesz sequence, then $g$ is separating for $\pi(\Gamma)''$. \[rem\_Riesz\_on\] If $\pi(\Gamma)g$ is a Riesz sequence in ${\mathcal{H}_{\pi}}$, then it is a Riesz basis for $[\pi(\Gamma) g]$ and hence the frame operator $S_{g, \Gamma} : [\pi(\Gamma) g ] \to [\pi(\Gamma) g]$ is well-defined and bounded. The biortogonality of the systems $\pi(\Gamma)g$ and $\pi (\Gamma) S^{-1}_{g,\Gamma} g$ yields that $$\big\langle S^{-1/2}_{g, \Gamma} \pi(\gamma') g, S^{-1/2}_{g, \Gamma} \pi(\gamma) g \big\rangle = \big \langle \pi(\gamma') g, S^{-1}_{g, \Gamma} \pi(\gamma) g \big\rangle = \delta_{\gamma', \gamma}, \quad \gamma, \gamma' \in \Gamma,$$ showing that $\pi(\Gamma) S^{-1/2}_{g,\Gamma} g = S^{-1/2}_{g, \Gamma} \pi(\Gamma) g$ is an orthonormal sequence in ${\mathcal{H}_{\pi}}$. For more on frames and Riesz bases, see e.g. the books [@christensen2016introduction; @young2001introduction]. Bounded operators and Bessel vectors ------------------------------------ The *coefficient operator* and *reconstruction operator* associated with ${\pi (\Gamma) g}$ are given respectively by $$\begin{aligned} \label{eq:coefficient} C_{g,\Gamma} f = \big( \langle f, \pi(\gamma) g \rangle \big)_{\gamma \in \Gamma}, \quad f \in {\mathcal{H}_{\pi}},\end{aligned}$$ and $$\begin{aligned} \label{eq:reconstruction} D_{g, \Gamma} c = \sum_{\gamma \in \Gamma} c_{\gamma} \pi(\gamma) g, \quad c = (c_{\gamma})_{\gamma \in \Gamma} \in c_{00} (\Gamma), \end{aligned}$$ where $c_{00} (\Gamma) \subseteq \mathbb{C}^{\Gamma}$ denotes the space of finite sequences on $\Gamma$. Recall that ${\pi (\Gamma) g}$ is called a *Bessel sequence* if there exists $B > 0$ such that holds. In this case, the coefficient operator is well-defined and bounded as a map from ${\mathcal{H}_{\pi}}$ into $\ell^2 (\Gamma)$, and its adjoint $D_{g, \Gamma}$ is well-defined and bounded from $\ell^2 (\Gamma)$ into ${\mathcal{H}_{\pi}}$. The space of Bessel vectors is denoted by $\mathcal{B}_{\pi}$. The assumption that $(\pi, {\mathcal{H}_{\pi}})$ is square-integrable in the sense of , together with the uniform boundedness principle, yields that the space ${\mathcal{B}_{\pi}}$ is norm dense in ${\mathcal{H}_{\pi}}$. Coefficient and reconstruction as unbounded operators ----------------------------------------------------- In the sequel, we treat the coefficient mapping and reconstruction mapping as operators from domains and on images in which they do not necessarily act as bounded operators. The coefficient operator $C_{g, \Gamma}$, with domain $$\operatorname*{dom}(C_{g, \Gamma}) := \big\{f \in \mathcal{H}_{\pi} \; : \; C_{g, \Gamma} f \in \ell^2 (\Gamma) \big\}$$ is given by $f \mapsto ( \langle f, \pi(\gamma) g \rangle )_{\gamma \in \Gamma}$ and well-defined from $\operatorname*{dom}(C_{g,\Gamma})$ into $\ell^2 (\Gamma)$. The reconstruction operator $D_{g, \Gamma}$, with domain $$\begin{aligned} \label{eq:dom_reconstruction} \operatorname*{dom}(D_{g, \Gamma} ) := \bigg\{ c \in \ell^2 (\Gamma) \; \bigg | \; \exists f \in \mathcal{H}_{\pi} \; : \; \sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi (\gamma) g, h \rangle \ = \langle f, h \rangle, \; \forall h \in \mathcal{B}_{\pi} \bigg\}\end{aligned}$$ is given by $D_{g, \Gamma} c = f$ and well-defined from $\operatorname*{dom}(D_{g, \Gamma})$ into ${\mathcal{H}_{\pi}}$, where $f \in {\mathcal{H}_{\pi}}$ is the vector occurring in the domain definition . Note that $f$ is uniquely determined since ${\mathcal{B}_{\pi}}$ is a dense subspace in ${\mathcal{H}_{\pi}}$. For simplicity, we also sometimes write $$\begin{aligned} D_{g, \Gamma} c = \sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) g;\end{aligned}$$ the series is however a formal expression for the vector $f$ in . The following result provides basic properties of the (possibly) unbounded coefficient and reconstruction operators. \[prop:denselydefined\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Let $g \in {\mathcal{H}_{\pi}}$ be an arbitrary vector. (i) The coefficient operator $C_{g, \Gamma} : \operatorname*{dom}(C_{g, \Gamma}) \to \ell^2 (\Gamma)$, $C_{g, \Gamma} f = ( \langle f, \pi(\gamma) g \rangle )_{\gamma \in \Gamma}$ is a closed, densely defined operator. (ii) The reconstruction operator $D_{g, \Gamma} : \operatorname*{dom}(D_{g, \Gamma}) \to \mathcal{H}_{\pi}$, $D_{g, \Gamma} c = f$, is a closed, densely defined operator. \(i) The map $C_{g, \Gamma} : \operatorname*{dom}(C_{g, \Gamma}) \to \ell^2 (\Gamma)$ is densely defined since the dense space of Bessel vector ${\mathcal{B}_{\pi}}\subset \operatorname*{dom}(C_{g, \Gamma})$. To show that $C_{g, \Gamma}$ is closed, let $f_n \to f$ in ${\mathcal{H}_{\pi}}$ with $f_n \in \operatorname*{dom}(C_{g, \Gamma})$ and assume that $C_{g, \Gamma} f_n \to c$ in $\ell^2 (\Gamma)$ as $n \to \infty$. By Cauchy-Schwarz, $$|C_{g, \Gamma} f_n (\gamma) - C_{g, \Gamma} f (\gamma)| = |\langle f_n - f, \pi(\gamma)g \rangle | \leq \| f_n - f \|_{{\mathcal{H}_{\pi}}} \| g \|_{{\mathcal{H}_{\pi}}} \to 0$$ as $n \to \infty$, yielding that $c = C_{g, \Gamma} f$. This shows that $C_{g,\Gamma} f \in \ell^2 (\Gamma)$, and hence $f \in \operatorname*{dom}(C_{g,\Gamma})$. \(ii) Note that the map $D_{g, \Gamma}$ is densely defined since the space of finite sequences $c_{00} (\Gamma) \subseteq \operatorname*{dom}(D_{g, \Gamma})$. To show that $D_{g, \Gamma}$ is closed, let $(c^{(k)})_{k \in \mathbb{N}} \subset \operatorname*{dom}(D_{g, \Gamma})$ be such that $c^{(k)} \to c$ in $\ell^2 (\Gamma)$ and $f_k := D_{g, \Gamma} c^{(k)} \to f$ for some $f \in \mathcal{H}_{\pi}$ as $k \to \infty$. Let $h \in \mathcal{B}_{\pi}$ be arbitrary. Then, $$\big\langle c^{(k)}, C_{g, \Gamma} h \rangle_{\ell^2 (\Gamma)} = \sum_{\gamma \in \Gamma} c_{\gamma}^{(k)} \langle \pi (\gamma) g, h \rangle = \langle f_k , h \rangle.$$ Since $C_{g,\Gamma} h \in \ell^2 (\Gamma)$ as $h \in \mathcal{B}_{\pi}$, it follows that $\langle c^{(k)}, C_{g,\Gamma} h \rangle_{\ell^2 (\Gamma)} \to \langle c, C_{g, \Gamma} h \rangle$ as $k \to \infty$, and hence $$\sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi (\gamma) g, h \rangle = \lim_{k \to \infty} \big\langle c^{(k)}, C_{g,\Gamma} h \big\rangle_{\ell^2 (\Gamma)} = \lim_{k \to \infty} \langle f_k , h \rangle = \langle f, h \rangle.$$ Thus $c \in \operatorname*{dom}(D_{g, \Gamma})$ and $D_{g, \Gamma} c = f$, which shows that $D_{g, \Gamma}$ is a closed operator. For a general frame $\{f_i\}_{i \in I}$ in an abstract Hilbert space $\mathcal{H}$, the coefficient operator $f \mapsto ( \langle f, f_i \rangle )_{i \in I}$ is always closed, but not necessarily densely defined, on its canonical domain. The reconstruction operator $(c_i)_{i \in I} \mapsto \sum_{i \in I} c_i f_i$ may fail to be closed on the domain $$\bigg\{ c = (c_i)_{i \in I} \in \ell^2 (I) \; : \; \sum_{i \in I} c_i f_i \mbox{ converges in the norm of }\mathcal{H} \bigg\},$$ see [@christensen1995frames]. Crucially, in and part (ii) of Proposition \[prop:denselydefined\], we define the series in a suitably weak form. Uniqueness for the extended representation {#sec:uniqueness} ------------------------------------------ Given $c = (c_{\gamma})_{\gamma \in \Gamma} \in \ell^2 (\Gamma)$, define the operator $$\begin{aligned} \label{eq_pic} \pi(c) : {\mathcal{B}_{\pi}}\to {\mathcal{H}_{\pi}}, \quad \pi(c) g :=\sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) g.\end{aligned}$$ Note that $\pi(c) = \sum_{\gamma \in \Gamma} c_{\gamma} \pi(\gamma)$ is well-defined since the series representing $\pi(c)g$ converges unconditionally in ${\mathcal{H}_{\pi}}$ by the Bessel property. In the notation of , conjugating the operator $\pi(c)$ simply corresponds to (twisted) conjugation of the corresponding sequence $c$. \[lem:conjugate\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Let $c \in \ell^2 (\Gamma)$. Then, for all $\gamma \in \Gamma$, $$\pi(\gamma) \pi (c) \pi(\gamma)^* = \pi(\vartheta_{\Gamma}^{\sigma} (\gamma) c),$$ where $$\begin{aligned} \label{eq:twosided_translation} (\vartheta_\Gamma^{\sigma} (\gamma) c)_{\gamma'} := \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1}) c_{\gamma^{-1} \gamma' \gamma}, \qquad \gamma, \gamma' \in \Gamma.\end{aligned}$$ Let $\gamma \in \Gamma$ be fixed. The identity $ \pi(\gamma) \pi( \gamma') \pi(\gamma)^* = \sigma(\gamma, \gamma') \overline{\sigma(\gamma \gamma' \gamma^{-1}, \gamma)} \pi(\gamma \gamma' \gamma^{-1}) $ holds for any $\gamma' \in \Gamma$. Therefore, $$\begin{aligned} \pi(\gamma) \pi (c) \pi(\gamma)^* &= \sum_{\gamma' \in \Gamma} c_{\gamma'} \sigma(\gamma, \gamma') \overline{\sigma(\gamma \gamma' \gamma^{-1}, \gamma)} \pi(\gamma \gamma' \gamma^{-1}) \\ &= \sum_{\gamma' \in \Gamma} c_{\gamma^{-1} \gamma' \gamma} \sigma(\gamma, \gamma^{-1} \gamma' \gamma) \overline{\sigma(\gamma', \gamma)} \pi(\gamma'), {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:conjugate1}\end{aligned}$$ where the second equality follows from the change of variable $\gamma' \mapsto \gamma \gamma' \gamma^{-1}$. Combining the identity $$\sigma(\gamma, \gamma^{-1} \gamma' \gamma) \sigma(\gamma' \gamma, \gamma^{-1}) = \sigma(\gamma, \gamma^{-1} \gamma') \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1})$$ with $$\sigma(\gamma, \gamma^{-1} \gamma') \sigma(\gamma^{-1}, \gamma') = \sigma(\gamma, \gamma^{-1}) = \sigma(\gamma', \gamma) \sigma(\gamma' \gamma, \gamma^{-1}),$$ yields that $\sigma(\gamma, \gamma^{-1} \gamma' \gamma) \overline{\sigma(\gamma', \gamma)} = \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1})$ for all $\gamma' \in \Gamma$. Inserting this in gives $$\begin{aligned} \pi(\gamma) \pi (c) \pi(\gamma)^* &= \sum_{\gamma' \in \Gamma} c_{\gamma^{-1} \gamma' \gamma} \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1}) \pi(\gamma') = \sum_{\gamma' \in \Gamma} \big(\vartheta_{\Gamma}^{\sigma} (\gamma) c\big)_{\gamma'} \pi (\gamma'),\end{aligned}$$ as desired. Under Kleppner’s condition, we have the following important uniqueness result. \[prop\_uniq\_op\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. Suppose $c \in \ell^2(\Gamma)$ is such that $\pi(c) \equiv 0$ on $\mathcal{B}_{\pi}$. Then $c = 0$. The proof is divided into four steps. **Step 1.** (Invariance of kernel). Let $\vartheta = \vartheta_{\Gamma}^{\sigma}$ be the unitary action of $\Gamma$ on $\ell^2 (\Gamma)$ given by . Define the closed subspace $$\mathcal{K} := \bigg\{ c \in \ell^2 (\Gamma) \; : \; \pi(c) g = 0, \quad \forall g \in B_{\pi} \bigg\} = \bigcap_{g \in \mathcal{B}_{\pi}} \mathcal{N} (D_{g, \Gamma})$$ of $\ell^2 (\Gamma)$. The space $\mathcal{K}$ is $\vartheta (\Gamma)$-invariant. Indeed, for $c \in \mathcal{K}$, by Lemma \[lem:conjugate\], $$\begin{aligned} \bigg\langle \pi\big( \vartheta(\gamma)c\big) g, h \bigg \rangle = \bigg\langle \sum_{\gamma' \in \Gamma} c_{\gamma'} \pi (\gamma') \pi(\gamma)^* g, \pi(\gamma)^* h \bigg\rangle = 0\end{aligned}$$ for all $\gamma \in \Gamma$ and $g, h \in {\mathcal{B}_{\pi}}$. Moreover, the space $\mathcal{K}$ is $\lambda_{\Gamma}^{\sigma} (\Gamma)$-invariant: For $\gamma \in \Gamma$ and $g,h \in \mathcal{B}_{\pi}$, $$\begin{aligned} \bigg\langle \pi \big(\lambda_{\Gamma}^{\sigma} (\gamma) c \big) g, h \bigg\rangle &= \sum_{\gamma' \in \Gamma} \sigma(\gamma, \gamma^{-1} \gamma') c_{\gamma^{-1} \gamma'} \langle \pi(\gamma') g,h \rangle \\ &= \sum_{\gamma' \in \Gamma} \sigma(\gamma, \gamma') \overline{\sigma(\gamma, \gamma')} c_{\gamma'} \langle \pi(\gamma) \pi(\gamma') g, h \rangle \\ &= \bigg \langle \sum_{\gamma' \in \Gamma} c_{\gamma'} \pi(\gamma') g, \pi(\gamma)^* h \bigg\rangle = 0,\end{aligned}$$ where the second equality follows from the change of variable $\gamma' \mapsto \gamma \gamma'$. **Step 2.** (Minimal fixed point). Let $c \in \mathcal{K}$ be arbitrary and consider the norm-closed convex hull $\overline{\operatorname*{co}(\vartheta (\Gamma) c)}$ in the Hilbert space $\mathcal{K}$. Then there exists a unique $d \in \overline{\operatorname*{co}(\vartheta (\Gamma) c)}$ of minimal norm. By uniqueness, the vector $d$ must be $\vartheta (\Gamma)$-invariant, that is, $$\begin{aligned} \label{eq:d_constant} d_{\gamma'} = \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1}) d_{\gamma^{-1} \gamma' \gamma}, \qquad \mbox{for all }\gamma, \gamma' \in \Gamma.\end{aligned}$$ Therefore, $|d|$ is constant on conjugacy classes. **Step 3.** ($\sigma$-regularity of non-zero entries). Let $\gamma' \in \Gamma$ be such that $d_{\gamma'} \neq 0$. Suppose $\gamma \in \Gamma$ commutes with $\gamma'$. Then, by , $$\begin{aligned} \label{eq:sigma-regular} 0 \neq d_{\gamma'} = \overline{\sigma(\gamma, \gamma')} \sigma(\gamma \gamma' \gamma^{-1}, \gamma ) d_{\gamma \gamma' \gamma^{-1}} = \overline{\sigma(\gamma, \gamma')} \sigma(\gamma', \gamma) d_{\gamma'},\end{aligned}$$ and, therefore, $\sigma(\gamma, \gamma') = \sigma(\gamma', \gamma)$. Thus $\gamma'$ is $\sigma$-regular. **Step 4.** (Vanishing coefficients on regular classes). By Step 3, if $\gamma \in \Gamma$ is such that $d_\gamma \not= 0$, then $\gamma$ is $\sigma$-regular, and by Kleppner’s condition, the conjugacy class $C_\gamma$ is infinite, unless $\gamma=e$. On the other hand, by , $|d|$ is constant on $C_\gamma$, while $|d| \in \ell^2 (\Gamma)$, and therefore $C_\gamma$ must be finite. We conclude that $d_{\gamma} = 0$ for all $\gamma \in \Gamma \setminus \{e\}$. Moreover, since $d \in \mathcal{K}$, also $d_e = 0$, and hence $d=0$. **Step 5.** (Conclusion). The above shows that for an arbitrary $c \in \mathcal{K}$, we have $0 \in \overline{\operatorname*{co}(\vartheta (\Gamma) c)}$. Since $(\vartheta(\gamma) c)_e = c_e$ for all $\gamma \in \Gamma$, it follows that $c_e = 0_e = 0$. The $\lambda^{\sigma}_{\Gamma} (\Gamma)$-invariance of $\mathcal{K}$ now yields that $c_{\gamma} = \overline{\sigma(\gamma^{-1}, \gamma)} \big(\lambda_{\Gamma}^{\sigma} (\gamma^{-1}) c\big)_e = 0$ for all $\gamma \in \Gamma$. This completes the proof. Step 2 in the proof of Proposition \[prop\_uniq\_op\] is an application of the *minimal method* for ergodic theorems [@alaoglu1940ergodic Section 10]. Improving spanning properties {#sec:improving} ============================= Mackey-type version of Schur’s lemma ------------------------------------ We will repeatedly use the following folklore result. \[prop:intertwiner\] For $i \in \{1,2\}$, let $(\pi_i, \mathcal{H}_{\pi_i})$ be $\sigma$-representations of a locally compact group $G$. Suppose that $T : \mathcal{H}_{\pi_1} \to \mathcal{H}_{\pi_2}$ is a closed, densely defined operator intertwining $(\pi_1, \mathcal{H}_{\pi_1})$ and $(\pi_2, \mathcal{H}_{\pi_2})$; that is, the domain and range of $T$ are respectively $\pi_1(G)$ and $\pi_2(G)$-invariant, and $$\begin{aligned} T \pi_1(x) = \pi_2(x) T, \quad x \in G.\end{aligned}$$ If $$T = U |T|$$ is the polar decomposition of $T$, then $|T| : \operatorname*{dom}(T) \to \mathcal{H}_{\pi_2}$ commutes with $(\pi_1, \mathcal{H}_{\pi_1})$ and the isometry $U : \mathcal{N}(T)^{\perp} \to \overline{\mathcal{R}(T)}$ isometrically intertwines $(\pi_1, \mathcal{H}_{\pi_1})$ and $(\pi_2, \mathcal{H}_{\pi_2})$. Note that $\pi_i (x)^* = \overline{\sigma(x, x^{-1})} \pi_i (x^{-1})$ and let $\tau(x) := \overline{\sigma(x, x^{-1})} \in \mathbb{T}$ for $x \in G$. Using that $\pi_1 (x)^* T^* = T^* \pi_2 (x)^*$ for all $x \in G$, a direct calculation entails $$T^* T \pi_1 (x) = T^* \pi_2 (x) T = T^* \overline{\tau (x)} \pi_2 (x^{-1})^* T = \overline{\tau(x)} \pi_1 (x^{-1})^* T^* T = \pi_1 (x) T^* T,$$ showing that $T^*T$ intertwines $(\pi_1, \mathcal{H}_{\pi_1})$. The operator $|T|$ is obtained from $T^*T$ by Borel functional calculus, and thus also commutes with $(\pi_1, \mathcal{H}_{\pi_1})$, e.g., see [@fell1988representations1 Theorem 12.14]. Using this, it follows directly that $$U \pi_1 (x) |T| = U |T| \pi_1 (x) = \pi_2 (x) U |T|,$$ whence $(U \pi_1 (x) - \pi_2 (x) U)|T| = 0$ for $x \in G$. Hence $(U \pi_1 (x) - \pi_2 (x) U) \equiv 0$ on $\mathcal{R}(|T|)$. Since $\mathcal{R}(|T|)$ is dense in $\mathcal{N}(T)^{\perp} = \overline{\mathcal{R}(|T|)}$, the desired conclusion follows. Mackey-type versions of Schur’s lemma for representations of $*$-algebras can be found in [@fell1988representations1]. From cyclic vectors to Parseval frames -------------------------------------- We show the existence of Parseval frames ${\pi (\Gamma) g}$ whenever $\pi$ admits a complete vector. \[prop:cyclic\_tight-frame\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Let $h \in {\mathcal{H}_{\pi}}$ be arbitrary. Then there exists $g \in \mathcal{H}_{\pi}$ such that ${\pi (\Gamma) g}$ is a Parseval frame for $[\pi(\Gamma)h]$. In particular, if $\pi$ is cyclic, then there exists a Parseval frame ${\pi (\Gamma) g}$ for ${\mathcal{H}_{\pi}}$. We split the proof into two steps. **Step 1.** (*Unitary intertwiner*). For $h \in {\mathcal{H}_{\pi}}$, the map $C_{h,\Gamma} : \operatorname*{dom}(C_{h, \Gamma}) \subseteq \mathcal{H}_{\pi} \to \ell^2 (\Gamma)$ is closed and densely defined by Proposition \[prop:denselydefined\]. Moreover, $C_{h,\Gamma}$ intertwines $\pi$ and $\lambda^{\sigma}_{\Gamma}$ by the covariance property . Thus the partial isometry $U : \mathcal{N} (C_{h,\Gamma})^{\perp} \to \overline{\mathcal{R} (C_{h,\Gamma})}$ of the decomposition $C_{h, \Gamma} = U |C_{h, \Gamma}|$ intertwines $\pi$ and $\lambda^{\sigma}_{\Gamma}$ by Proposition \[prop:intertwiner\]. Since $\mathcal{N} (C_{h,\Gamma})^{\perp} = [\pi(\Gamma) h]$, it follows that $U :[\pi(\Gamma) h] \to \overline{\mathcal{R} (C_{h, \Gamma})}$ is a unitary intertwiner. **Step 2.** (*Parseval frame*). Let $P_{\mathcal{K}} : \ell^2 (\Gamma) \to \ell^2 (\Gamma)$ be the orthogonal projection onto $\mathcal{K} := \overline{\mathcal{R} (C_{\eta,\Gamma})}$. Then $P_{\mathcal{K}} \in \lambda^{\sigma}_{\Gamma} (\Gamma)'$ by the projection lemma, and $\lambda^{\sigma}_{\Gamma} (\Gamma) P_{\mathcal{K}} \delta_{e} = P_{\mathcal{K}} \lambda_{\Gamma}^{\sigma} (\Gamma) \delta_e$ satisfies $$\| f \|_{{\mathcal{H}_{\pi}}}^2 = \|P_{\mathcal{K}} f \|_{{\mathcal{H}_{\pi}}}^2 = \sum_{\gamma \in \Gamma} |\langle P_{\mathcal{K}} f, \lambda_{\Gamma}^{\sigma} (\gamma) \delta_e \rangle|^2 = \sum_{\gamma \in \Gamma} |\langle f, \lambda_{\Gamma}^{\sigma} (\gamma) P_{\mathcal{K}} \delta_e \rangle|^2, \quad f \in \mathcal{K} ,$$ showing that $\lambda^{\sigma}_{\Gamma} (\Gamma) P \delta_{e}$ is a Parseval frame for $\mathcal{K}$. Since $U$ is unitary, the system $ \pi (\Gamma) U^{*} P \delta_e = U^* \lambda^{\sigma}_{\Gamma} (\Gamma) P \delta_{e}$ is a Parseval frame for the span $[\pi(\Gamma) h]$. The construction of the unitary operator in Step 1 above is standard, e.g., see [@rieffel1969square; @rieffel2004integrable]. It is also used, for example, in [@gabardo2003frame; @barbieri2015riesz; @fuehr2005abstract]. From separating vectors to orthonormal sequences ------------------------------------------------ The following result complements Proposition \[prop:cyclic\_tight-frame\] with a similar result for separating vectors and orthonormal sequences. In contrast to Proposition \[prop:cyclic\_tight-frame\], the result requires the assumption that Kleppner’s condition is satisfied. \[prop:separating\_riesz\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition and that $\pi (\Gamma)''$ admits a separating vector. Then there exists $g \in \mathcal{H}_{\pi}$ such that ${\pi (\Gamma) g}$ forms an orthonormal sequence in $\mathcal{H}_{\pi}$. For an arbitrary $\eta \in {\mathcal{H}_{\pi}}$, the map $D_{\eta, \Gamma} : \operatorname*{dom}(D_{\eta, \Gamma}) \to \mathcal{H}_{\pi}$ is a closed, densely defined operator by Proposition \[prop:denselydefined\]. The proof will be split into three steps: **Step 1.** (Auxiliary operator $\widetilde{\pi}(c)$). For a fixed $c \in \ell^2 (\Gamma)$, consider the auxiliary operator $\widetilde{\pi}(c) : \operatorname*{dom}(\widetilde{\pi}(c)) \to \mathcal{H}_{\pi}$, with domain $$\operatorname*{dom}(\widetilde{\pi} (c) ) := \bigg\{ g \in \mathcal{H}_{\pi} \; | \; \exists f \in \mathcal{H}_{\pi} \; : \; \sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi (\gamma) g, h \rangle \ = \langle f, h \rangle, \; \forall h \in \mathcal{B}_{\pi} \bigg\},$$ defined by $\widetilde{\pi} (c) g = D_{g, \Gamma} c$. Note that $\mathcal{B}_{\pi} \subseteq \operatorname*{dom}(\widetilde{\pi} (c))$ and hence $\widetilde{\pi} (c)$ is densely defined. A similar argument as in part (ii) of Proposition \[prop:denselydefined\] shows that $\widetilde{\pi} (c)$ is a closed operator. **Step 2.** ($D_{\eta, \Gamma}$ is injective for separating $\eta$). We show that $D_{\eta, \Gamma}$ is injective if $\eta \in {\mathcal{H}_{\pi}}$ is separating for $\pi(\Gamma)''$. For this, let $c \in \operatorname*{dom}(D_{\eta, \Gamma})$ be such that $D_{\eta, \Gamma} c = 0$. Then, $\eta \in \operatorname*{dom}(\widetilde{\pi}(c))$. Let $T \in \pi(\Gamma)'$, $g \in \operatorname*{dom}(\widetilde{\pi}(c))$ and $h \in {\mathcal{B}_{\pi}}$ be arbitrary. Then $$\langle T \widetilde{\pi}(c) g, h \rangle = \sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi(\gamma) g, T^* h \rangle = \sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi(\gamma) T g, h \rangle = \big\langle \widetilde{\pi}(c) T g, h \big\rangle,$$ and $Tg \in \operatorname*{dom}(\widetilde{\pi}(c))$. Hence, $\widetilde{\pi}(c) Tg = T \widetilde{\pi}(c)g$ by density of $\mathcal{B}_{\pi}$. Let $\widetilde{\pi} (c) = J |\widetilde{\pi} (c)| $ be the polar decomposition of $\widetilde{\pi} (c)$. Then $T |\widetilde{\pi} (c)| = |\widetilde{\pi} (c)| T$ by the Borel functional calculus, e.g., see [@fell1988representations1 Theorem 12.14]. Hence $T J |\widetilde{\pi} (c)| = J |\widetilde{\pi} (c)| T = J T |\widetilde{\pi}(c)|$, which yields that $(TJ - JT) \equiv 0$ on the dense subspace $\mathcal{R} (|\widetilde{\pi} (c)|)$. This shows that $J \in \pi (\Gamma)''$. Since $0 = D_{\eta, \Gamma} c = \widetilde{\pi} (c) \eta = |(\widetilde{\pi}(c))^*|J \eta$ and $\mathcal{N}(|(\widetilde{\pi} (c))^*|) = \mathcal{R}(\widetilde{\pi}(c))^{\perp}$, it follows that $J\eta = 0$. The separateness of $\eta \in \mathcal{H}_{\pi}$ yields $J = 0$, hence $\widetilde{\pi} (c) = 0$. In particular, $$\sum_{\gamma \in \Gamma} c_{\gamma} \langle \pi(\gamma) g, h \rangle = 0$$ for all $g, h \in {\mathcal{B}_{\pi}}$. By Proposition \[prop\_uniq\_op\], it follows that $c = 0$. Thus $D_{\eta, \Gamma}$ is injective. **Step 3.** (Isometric intertwiner). Since $D_{\eta, \Gamma} : \operatorname*{dom}(D_{\eta, \Gamma}) \to {\mathcal{H}_{\pi}}$ intertwines $\lambda_{\Gamma}^{\sigma}$ and $\pi$, it follows by Proposition \[prop:intertwiner\] that $U : \ell^2 (\Gamma) = \mathcal{N} (D_{\eta, \Gamma})^{\perp} \to \overline{\mathcal{R} (D_{\eta, \Gamma})}$ in the polar decomposition $D_{\eta, \Gamma} = U |D_{\eta, \Gamma}|$ intertwines $\lambda^{\sigma}_{\Gamma}$ and $\pi$. Therefore, the system $ \pi(\Gamma) U \delta_e = U \lambda^{\sigma}_{\Gamma} (\Gamma) \delta_e $ is an orthonormal basis for $\overline{\mathcal{R} (D_{\eta, \Gamma})}$, thus an orthonormal sequence in $\mathcal{H}_{\pi}$. Expansions in the von Neumann algebra {#sec_exps} ===================================== Expansions {#sec:expansions} ---------- The following theorem provides, under Kleppner’s condition, a Fourier-type series expansion for every operator in $\pi(\Gamma)''$. \[th\_series\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. Then, for every operator $T \in \pi(\Gamma)''$, there exists a unique $c \in \ell^2(\Gamma)$ such that $T = \pi(c)$ on $\mathcal{B}_{\pi}$, i.e., $$\begin{aligned} \label{eq_Tf} T f = \sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) f, \quad f \in {\mathcal{B}_{\pi}}. \end{aligned}$$ The uniqueness claim follows from Proposition \[prop\_uniq\_op\]. For the existence claim, consider the space $$\begin{aligned} \mathcal{A} := \left\{ \pi(c) = \sum_{\gamma \in \Gamma} c_{\gamma} \pi(\gamma) \; \bigg| \; c \in c_{00} (\Gamma) \right\} \subset \pi(\Gamma)'',\end{aligned}$$ where $c_{00}(\Gamma) \subset \mathbb{C}^{\Gamma}$ are finite sequences on $\Gamma$. The space $\mathcal{A}$ is a self-adjoint algebra containing $\pi(\Gamma) \subset \mathcal{U}({\mathcal{H}_{\pi}})$. By von Neumann’s density theorem, the von Neumann algebra $\pi(\Gamma)''$ is the SOT closure of $\mathcal{A}$. To provide for arbitrary $T \in \pi(\Gamma)''$, we first construct a vector-valued orthonormal sequence. **Step 1.** (Existence of vector-valued orthonormal sequence) Let $\eta = (\eta_k)_{ k \in \mathbb{N}}$ be a sequence of vectors $\eta_k \in {\mathcal{H}_{\pi}}$ such that $\{\eta_k : k \in \mathbb{N} \}$ is norm dense in ${\mathcal{H}_{\pi}}$. Consider the direct sum ${\mathcal{H}}:= \bigoplus_{n \in \mathbb{N}} {\mathcal{H}_{\pi}}= \ell^2(\mathbb{N}, {\mathcal{H}_{\pi}})$ and the associated direct sum $\sigma$-representation $({\bigoplus_{n \in \mathbb{N}} \pi}, {\mathcal{H}})$ of $\Gamma$, given by $$\begin{aligned} {\bigoplus_{n \in \mathbb{N}} \pi}(\gamma) f = (\pi(\gamma) f_k)_{k \in \mathbb{N}}, \quad f = (f_k)_{k \in \mathbb{N}} \in {\mathcal{H}}.\end{aligned}$$ The associated von Neumann algebra $({\bigoplus_{n \in \mathbb{N}} \pi}(\Gamma))''$ consists of operators $T \in \mathcal{B}({\mathcal{H}})$ acting as $$\begin{aligned} \label{eq_aaaaa} T (f_k)_{k \in \mathbb{N}} := (A f_k)_{k \in \mathbb{N}},\end{aligned}$$ for some $A \in \pi(\Gamma)''$. We claim that $\eta = (\eta_k)_{k \in \mathbb{N}}$ is a separating vector for $({\bigoplus_{n \in \mathbb{N}} \pi}(\Gamma))''$. Indeed, if $T \in ({\bigoplus_{n \in \mathbb{N}} \pi}(\Gamma))''$ annihilates $\eta$, then, for $A$ as in , $A \eta_k = 0$ for all $k \in \mathbb{N}$, and, by density, $A=0$, which implies $T=0$. The space of Bessel vectors $\mathcal{B}_{\oplus \pi}$ of the direct sum $({\bigoplus_{n \in \mathbb{N}} \pi}, {\mathcal{H}})$ is norm dense in ${\mathcal{H}}$ since it contains $\{ (\delta_{j,k} h)_{k \in \mathbb{N}} \; : \; h \in {\mathcal{B}_{\pi}},\; j \in \mathbb{N} \}$. Therefore, Proposition \[prop:separating\_riesz\] is applicable to obtain a vector $g = (g_k)_{k \in \mathbb{N}} \in {\mathcal{H}}$ such that $({\bigoplus_{n \in \mathbb{N}} \pi}(\Gamma) )g$ is orthonormal in ${\mathcal{H}}$. Hence, $$\begin{aligned} \label{eq:vector-valued_orthonormal} {\lVertc\rVert}_2^2 = {\Big\lVert\sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) g\Big\rVert}_{{\mathcal{H}}}^2 = \sum_{k \in \mathbb{N}} {\Big\lVert\sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) g_k\Big\rVert}_{{\mathcal{H}_{\pi}}}^2\end{aligned}$$ for all $c \in \ell^2(\Gamma)$, and, in particular, $$\begin{aligned} \label{eq:vector-valued_norm1} {\lVertg\rVert}^2_{{\mathcal{H}}}= \sum_{k \in \mathbb{N}} \| g_k \|_{{\mathcal{H}_{\pi}}}^2 = 1. \end{aligned}$$ **Step 2.** (Strong closure of $\mathcal{A}$). Let $T \in \pi(\Gamma)''$. By von Neumann’s and Kaplansky’s density theorem, there exists a bounded net $(T_{\alpha} )_{\alpha \in \Lambda}$ of operators $T_{\alpha} \in \mathcal{A}$ such that $T_{\alpha} {\xrightarrow{\textit{SOT}}}T$. Let $g = (g_k)_{ k \in \mathbb{N}} \in {\mathcal{H}}$ be as in Step 1 satisfying and . Select sequences $c^{(\alpha)} \in c_{00} (\Gamma) \subset \ell^2(\Gamma)$ such that $T_{\alpha} = \pi(c^{(\alpha)})$. Then, for each $\alpha \in \Lambda$, $$\begin{aligned} {\lVertc^{(\alpha)}\rVert}_2 &= \sum_{k \in \mathbb{N}} {\Big\lVert\sum_{\gamma \in \Gamma} c^{(\alpha)}_\gamma \pi(\gamma) g_k\Big\rVert}_{{\mathcal{H}_{\pi}}}^2 = \sum_{k \in \mathbb{N}} {\big\lVertT_{\alpha} g_k\big\rVert}_{{\mathcal{H}_{\pi}}}^2 \\ &\leq {\lVertT_{\alpha}\rVert}_{op}^2 \sum_{k \in \mathbb{N}} {\lVertg_k\rVert}_{{\mathcal{H}_{\pi}}}^2 \leq \sup_{\alpha' \in \Lambda} {\lVertT_{\alpha'}\rVert}_{op}^2 < \infty.\end{aligned}$$ By the Banach-Alaoglu theorem, we may pass to a subnet and assume that $c^{(\alpha)} {\xrightarrow{w}}c$ for some $c \in \ell^2(\Gamma)$. Let $f \in \mathcal{B}_\pi$ and $h \in {\mathcal{H}_{\pi}}$ be arbitrary. Then $({\ensuremath{\left<\pi(\gamma) f,h\right>}})_{\gamma \in \Gamma} \in \ell^2(\Gamma)$, and, thus, $$\begin{aligned} {\ensuremath{\left<T_{\alpha} f,h\right>}} = \sum_{\gamma \in \Gamma} c^{(\alpha)}_\gamma {\ensuremath{\left<\pi(\gamma) f,h\right>}} \longrightarrow \sum_{\gamma \in \Gamma} c_\gamma {\ensuremath{\left<\pi(\gamma) f,h\right>}}.\end{aligned}$$ On the other hand ${\ensuremath{\left<T_{\alpha} f,h\right>}} \longrightarrow {\ensuremath{\left<Tf,h\right>}}$. Hence, $\pi(c) = T$ on $\mathcal{B}_\pi$, as desired. Coherent Riesz sequences are generated by separating vectors ------------------------------------------------------------ As a first application of Theorem \[th\_series\], we show the following. \[prop\_riesz\_sep\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. If $\pi(\Gamma)g$ is a Riesz sequence in ${\mathcal{H}_{\pi}}$, then $g$ is separating for $\pi(\Gamma)''$. Suppose that $\pi(\Gamma)g$ is a Riesz sequence in ${\mathcal{H}_{\pi}}$ and assume that $T \in \pi(\Gamma)''$ annihilates $g$. By Theorem \[th\_series\], there exists a sequence $c \in \ell^2(\Gamma)$ such that $T=\pi(c)$. Since $\pi(\Gamma)g$ is a Riesz sequence, we have $g \in \mathcal{B}_\pi$, and, therefore, $$\begin{aligned} 0= {\lVertT g\rVert}^2_{{\mathcal{H}_{\pi}}} = {\Big\lVert\sum_{\gamma \in \Gamma} c_\gamma \pi(\gamma) g\Big\rVert}_{{\mathcal{H}_{\pi}}}^2 \asymp {\lVertc\rVert}_{\ell^2(\Gamma)}^2.\end{aligned}$$ Thus $c = 0$, and, therefore, $T = 0$, as desired. Doubly invariant subspaces -------------------------- As a second application of Theorem \[th\_series\], we show that ${\mathcal{H}_{\pi}}$ does not admit so-called doubly invariant subspaces. \[prop\_no\_double\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. Let $\mathcal{K} \leq {\mathcal{H}_{\pi}}$ be a closed subspace that is invariant under $\pi(\Gamma)$ and $\pi(\Gamma)'$. Then $\mathcal{K}=\{0\}$ or $\mathcal{K}={\mathcal{H}_{\pi}}$. Consider the orthogonal projection $P_{\mathcal{K}} : {\mathcal{H}_{\pi}}\to {\mathcal{H}_{\pi}}$ onto $\mathcal{K}$. Since $\mathcal{K}$ is $\pi(\Gamma)'$-invariant, it follows that $P_{\mathcal{K}} \in \pi(\Gamma)''$ by the projection lemma. Theorem \[th\_series\] then yields a unique sequence $c \in \ell^2(\Gamma)$ such that $$\begin{aligned} \label{eq:expansion_doubly} P_{\mathcal{K}} = \pi (c) = \sum_{\gamma' \in \Gamma} c_{\gamma'} \pi(\gamma') \end{aligned}$$ as an operator on ${\mathcal{B}_{\pi}}$. Since $\mathcal{K}$ is also $\pi(\Gamma)$-invariant, it follows also that $P_{\mathcal{K}} \in \pi(\Gamma)'$. Therefore $P_{\mathcal{K}} = \pi(\gamma) P_{\mathcal{K}} \pi(\gamma)^*$ for all $\gamma \in \Gamma$. By Lemma \[lem:conjugate\], $$\begin{aligned} \pi(\gamma) P_{\mathcal{K}} \pi(\gamma)^{*} = \sum_{\gamma' \in \Gamma} (\vartheta_{\Gamma}^{\sigma} (\gamma) c)_{\gamma'} \pi(\gamma')\end{aligned}$$ The uniqueness of the expansion shows that $$\begin{aligned} \label{eq:sequence_constantconjugacy2} c_{\gamma'} = (\vartheta_{\Gamma}^{\sigma} (\gamma) c)_{\gamma'} = \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1}) c_{\gamma^{-1} \gamma' \gamma}\end{aligned}$$ for all $\gamma, \gamma' \in \Gamma$. Thus $|c|$ is constant on conjugacy classes. We now use Kleppner’s condition together with the fact that $c \in \ell^2 (\Gamma)$, as in Steps 3 and 4 of the proof of Proposition \[prop\_uniq\_op\], to conclude that $c_\gamma=0$, for $\gamma \in \Gamma \setminus \{e\}$. This shows that either $P_{\mathcal{K}} = 0$ or $P_{\mathcal{K}} = I_{{\mathcal{H}_{\pi}}}$, as claimed. Proposition \[prop\_no\_double\] shows that, under Kleppner’s condition, the center $\pi (\Gamma)'' \cap \pi(\Gamma)'$ of the algebra $\pi(\Gamma)''$ does not contain non-trivial projections, and thus equals $\mathbb{C} I_{{\mathcal{H}_{\pi}}}$. In technical terms: The von Neumann algebra $\pi(\Gamma)''$ is a *factor*. Kleppner’s condition is also necessary for $\pi(\Gamma)''$ to be a factor. Indeed, if $C_{\gamma_0}$ is a finite non-trivial $\sigma$-regular conjugacy class, then the sequence $c \in \ell^2 (\Gamma)$ defined by $$c_{\gamma'} = \begin{cases} \sigma(\gamma, \gamma_0) \overline{\sigma(\gamma \gamma_0 \gamma^{-1}, \gamma)}, \quad & \text{if} \; \gamma' \in C_{\gamma_0}, \; \; \gamma' = \gamma \gamma_0 \gamma^{-1} \\ 0, & \text{if} \; \gamma' \notin C_{\gamma_0} \end{cases}$$ is well-defined and satisfies $c_{\gamma'} = \overline{\sigma(\gamma^{-1}, \gamma')} \sigma(\gamma^{-1} \gamma' \gamma, \gamma^{-1}) c_{\gamma^{-1} \gamma' \gamma}$ for all $\gamma' \in C_{\gamma_0}$ and $\gamma \in \Gamma$, and by Lemma \[lem:conjugate\], one can see that $$\begin{aligned} T := \sum_{\gamma' \in C_{\gamma_0}} c_{\gamma'} \pi(\gamma') \in \pi(\Gamma)''\cap \pi(\Gamma)'\end{aligned}$$ and that $T \notin \mathbb{C} I_{{\mathcal{H}_{\pi}}}$. Therefore $\pi(\Gamma)''$ is not a factor. See [@kleppner1962structure; @omland2014primeness] for similar arguments. Existence of cyclic or separating vectors {#sec:existence} ========================================= In this section we investigate how to produce large cyclic subspaces for $\pi(\Gamma)$. As a first step, we investigate when the sum of two orthogonal cyclic subspaces, $[\pi(\Gamma)g_1]$ and $[\pi(\Gamma)g_2]$ is again cyclic. The following key lemma shows that this is the case, provided that the corresponding cyclic subspaces generated by the commutant algebra $\pi (\Gamma)'$, i.e., $[\pi(\Gamma)'g_1]$ and $[\pi(\Gamma)' g_2]$, are also orthogonal. \[lemma\_sum\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose $(g_k)_{k \in I}$ is a countable family of unit-norm vectors $g_k \in {\mathcal{H}_{\pi}}$ satisfying the following *simultaneous orthogonality conditions* $$\begin{aligned} \label{eq_d1} \pi(\Gamma) g_k \perp \pi(\Gamma) g_j, \qquad k \not= j, \\ \label{eq_d2} \pi(\Gamma)' g_k \perp \pi(\Gamma)' g_j, \qquad k \not= j.\end{aligned}$$ Let $a \in \ell^1(I)$ with $a_k \not =0$ for all $k \in I$, and set $g := \sum_{k \in I} a_k g_k$. Then $$\begin{aligned} \label{eq_d1a} [\pi(\Gamma) g] = \bigoplus_{k \in I} [\pi(\Gamma) g_k], \\ \label{eq_d1b} [\pi(\Gamma)' g] = \bigoplus_{k \in I} [\pi(\Gamma)' g_k].\end{aligned}$$ Clearly, $[\pi(\Gamma) g] \subseteq \bigoplus_{k \in I} [\pi(\Gamma) g_k]$. For the other inclusion, let $k \in I$, and note that the projection $P_{\mathcal{K}_k}$ onto $\mathcal{K}_k = [\pi(\Gamma)' g_k]$ is in $\pi(\Gamma)''$ as $[\pi(\Gamma)' g_k]$ is $\pi(\Gamma)'$-invariant. Therefore $g_k = {a_k}^{-1} P_{\mathcal{K}_k} g \in [\pi(\Gamma) g]$ for all $k \in I$. This gives . The identity follows similarly, interchanging the roles of $\pi(\Gamma)'$ and $\pi(\Gamma)''$. \[prop\_either\] Let $(\pi, {\mathcal{H}_{\pi}})$ be a square-integrable $\sigma$-representation of a countable discrete group $\Gamma$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. Then $\pi$ admits a cyclic vector or $\pi(\Gamma)''$ admits a separating vector (possibly both). By Zorn’s Lemma, we can select a family $(g_k)_{k \in I}$ of unit-norm vectors $g_k \in {\mathcal{H}_{\pi}}$ satisfying the simultaneous orthogonality conditions and , and maximal with respect to that property. The set $I$ is countable because ${\mathcal{H}_{\pi}}$ is assumed to be separable. Let $g := \sum_{k \in I} a_k g_k$ be as in Lemma \[lemma\_sum\], so that and hold. The maximality of $(g_k)_{ k \in I}$ implies that $$\begin{aligned} \label{eq_aaa} [\pi(\Gamma) g]^\perp \cap [\pi(\Gamma)' g]^\perp = \{ 0 \};\end{aligned}$$ otherwise, we could choose a unit-norm vector $h \in [\pi(\Gamma) g]^\perp \cap [\pi(\Gamma)' g]^\perp$, and extend the family $(g_k)_{k \in I}$. We claim that, in addition, $$\begin{aligned} \label{eq_bbb} [\pi(\Gamma) g]^\perp \perp [\pi(\Gamma)' g]^\perp.\end{aligned}$$ To see this, let $P_1$ and $P_2$ be the orthogonal projections onto $[\pi(\Gamma) g]$ and $[\pi(\Gamma)' g]$ respectively. Then $P_1 \in \pi(\Gamma)''$ and $P_2 \in \pi(\Gamma)'$, and therefore $P_1$ and $P_2$ commute. Hence, by , $$\begin{aligned} \mathcal{R}( (I-P_1)(I-P_2) ) = \mathcal{R}( (I-P_2) (I-P_1) ) \subseteq [\pi(\Gamma) g]^\perp \cap [\pi(\Gamma)' g]^\perp=\{0\}.\end{aligned}$$ Therefore $(I-P_1)(I-P_2)=0$, which implies . Note that $[\pi(\Gamma) g]^\perp$ is $\pi(\Gamma)''$ invariant, while $[\pi(\Gamma)' g]^\perp$ is $\pi(\Gamma)'$ invariant. As a consequence, the subspaces $$\begin{aligned} \mathcal{K}_1 := [\pi(\Gamma)' \big([\pi(\Gamma)'' g]^\perp\big)], \qquad \mathcal{K}_2 := [\pi(\Gamma)''\big([\pi(\Gamma)' g]^\perp\big)]\end{aligned}$$ are also orthogonal. Indeed, for $T' \in \pi(\Gamma)'$, $f_1 \in [\pi(\Gamma)'' g]^\perp$, $T \in \pi(\Gamma)''$, $f_2 \in [\pi(\Gamma)' g]^\perp$, the commutativity of $T$ and $T'$ implies that ${\ensuremath{\left<T' f_1,T f_2\right>}} = {\ensuremath{\left<T^* f_1,(T')^* f_2\right>}} = 0$. On the other hand, the subspaces $\mathcal{K}_1$ and $\mathcal{K}_2$ are doubly-invariant: $\pi(\Gamma) \mathcal{K}_i = \pi(\Gamma)' \mathcal{K}_i=\mathcal{K}_i$ for $i=1,2$. Lemma \[prop\_no\_double\] therefore implies that $\mathcal{K}_i = \{0\}$ or $\mathcal{K}_i = {\mathcal{H}_{\pi}}$ for $i=1,2$. The possibility $\mathcal{K}_1=\mathcal{K}_2={\mathcal{H}_{\pi}}$ is excluded (unless ${\mathcal{H}_{\pi}}=\{0\}$) because $\mathcal{K}_1 \perp \mathcal{K}_2$. Thus, either $\mathcal{K}_1 = \{0\}$, or $\mathcal{K}_2 = \{0\}$. If $\mathcal{K}_1=\{0\}$, then $[\pi(\Gamma)'' g]^\perp = \{0\}$, yielding a cyclic vector: ${\mathcal{H}_{\pi}}= [\pi(\Gamma)'' g]$. If $\mathcal{K}_2=\{0\}$, then $[\pi(\Gamma)' g]={\mathcal{H}_{\pi}}$, which implies that $g$ is a separating vector for $\pi(\Gamma)''$ by the discussion in Section \[sec\_comp\_sep\]. Lemma \[lemma\_sum\] and Proposition \[prop\_either\] are simplifications and adaptions of standard results on central projections in reduced von Neumann algebras [@dixmier1981vonneumann I.2, Proposition 3]. Discrete series representations restricted to lattices {#sec:discreteseries} ====================================================== Let $G$ be a second countable unimodular group and let $\Gamma \subset G$ be a lattice subgroup. Let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$, i.e., irreducible and square-integrable. This section is devoted to orbits of the restriction $\pi|_{\Gamma}$ of $(\pi, {\mathcal{H}_{\pi}})$ to $\Gamma$, i.e., $$\pi(\Gamma)g = \big\{ \pi (\gamma) g \; : \; \gamma \in \Gamma \big\}$$ for some $g \in {\mathcal{H}_{\pi}}$. In order to apply the results obtained in the previous sections, it is essential that the restriction $\pi|_{\Gamma}$ be square-integrable in the sense of . The following observation guarantees this. Let $\Gamma \subseteq G$ be a lattice and let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$. The Bessel vectors ${\mathcal{B}_{\pi}}$ of the restriction $\pi|_{\Gamma}$ are norm dense in ${\mathcal{H}_{\pi}}$. The space of Bessel vectors of $\pi|_{\Gamma}$ is $\pi(G)$-invariant, and, therefore, if non-zero, it is norm dense in ${\mathcal{H}_{\pi}}$ by the irreducibility of $\pi$. Hence, it suffices to show that $\pi|_{\Gamma}$ admits a non-zero Bessel vector. Using the orthogonality relations , choose $\eta \in {\mathcal{H}_{\pi}}$ such that the map $C_{\eta} : {\mathcal{H}_{\pi}}\to L^2 (G)$ is an isometry. Let $P_{\mathcal{K}} : L^2 (G) \to L^2 (G)$ be the orthogonal projection onto the closed subspace $\mathcal{K} := C_{\eta} ({\mathcal{H}_{\pi}})$. Then $P_{\mathcal{K}} \in \lambda_{G}^{\sigma} (G)'$. Let $\Omega \subset G$ be a left fundamental domain for $\Gamma$ and consider the collection of indicator functions $\mathcal{S}_{\Omega} := \left\{ \chi_E \,:\, E \subseteq \Omega \mbox{ measurable } \right\}$. Then $ \left[\lambda_{G}^{\sigma} (\Gamma) \mathcal{S}_{\Omega} \right] = L^2 (G). $ Hence, there exists a measurable subset $E_0 \subseteq \Omega$ such that $P_{\mathcal{K}} (\chi_{E_0}) \not= 0$. The family $$\begin{aligned} \lambda_{G}^{\sigma} (\Gamma) \chi_{E_0} = \{\lambda_G^{\sigma}(\gamma) \chi_{E_0} : \gamma \in \Gamma\}\end{aligned}$$ is orthogonal in $L^2 (G)$, and consequently Bessel. Hence, also $\lambda_{G}^{\sigma} (\Gamma) P_{\mathcal{K}} \chi_{E_0} = P_{\mathcal{K}} \lambda_{G}^{\sigma} (\Gamma) \chi_{E_0}$ is Bessel in $\mathcal{K}$. The map $D_{\eta} := C_{\eta}^{*}|_{\mathcal{K}} : \mathcal{K} \to {\mathcal{H}_{\pi}}$ is unitary, and, by the covariance , $\pi(\Gamma) D_{\eta} P_{\mathcal{K}}(\chi_{E_0}) = D_{\eta} \lambda_{G}^{\sigma}(\Gamma) P_{\mathcal{K}}(\chi_{E_0})$, yielding that the non-zero vector $D_\eta P_{\mathcal{K}}(\chi_{E_0})$ is a Bessel vector for $\pi|_{\Gamma}$. Frame bounds and density {#sec_density_1} ------------------------ The following proposition relates frame bounds , formal dimension and co-volume. \[prop:density\_quasi\] Let $\Gamma \subseteq G$ be a lattice and let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. If ${\pi (\Gamma) g}$ admits a Bessel bound $B > 0$, then $ d_{\pi}^{-1} \| g \|_{\mathcal{H}_{\pi}}^2 \leq B \operatorname*{vol}(G / \Gamma)$. Moreover, if ${\pi (\Gamma) g}$ also admits a lower frame bound $A > 0$, then $$\begin{aligned} \label{eq:framebounds_formal} A \operatorname*{vol}(G /\Gamma) \leq d_{\pi}^{-1} \| g \|_{\mathcal{H}_{\pi}}^2 \leq B \operatorname*{vol}(G / \Gamma). \end{aligned}$$ Let $\Omega \subseteq G$ be a right fundamental domain of $\Gamma \subseteq G$. Then $$\begin{aligned} \int_G |\langle f, \pi (x) g \rangle |^2 \; d\mu_G (x) = \sum_{\gamma \in \Gamma} \int_{\Omega} |\langle f, \pi(x \gamma ) g \rangle|^2 \; d\mu_{G} (x) = \int_{\Omega} \sum_{\gamma \in \Gamma} |\langle \pi(x)^* f, \pi(\gamma) g \rangle|^2 \; d\mu_G (x)\end{aligned}$$ for $f \in \mathcal{H}_{\pi}$. This, together with the orthogonality relations , yields $$\begin{aligned} d_{\pi}^{-1} \|f \|_{\mathcal{H}_{\pi}}^2 \|g\|_{\mathcal{H}_{\pi}}^2 = \int_{\Omega} \sum_{\gamma \in \Gamma} |\langle \pi(x)^* f, \pi(\gamma) g \rangle|^2 \; d\mu_G (x). \end{aligned}$$ Thus, if ${\pi (\Gamma) g}$ is Bessel with bound $B$, then $ d_{\pi}^{-1} \|f \|_{\mathcal{H}_{\pi}}^2 \|g\|_{\mathcal{H}_{\pi}}^2 \leq B \int_{\Omega} \| f \|_{\mathcal{H}_{\pi}}^2 d\mu_G (x), $ which shows the upper bound in . The desired lower bound is proven similarly. The proof of Proposition \[prop:density\_quasi\] also works for discrete subgroups $\Gamma \subset G$ having possibly infinite co-volume. However, the lower bound in shows that the restriction $\pi|_{\Gamma}$ admits a frame only if $\Gamma \subset G$ has finite co-volume. The lattice assumption is in fact even necessary for $\pi|_{\Gamma}$ to admit a cyclic vector [@bekka2004square Corollary 2]. The idea of periodizing the orthogonality relations by means of Weil’s integral formula can also be found in [@cowling1991irreducibility; @kuhn1992restrictions]. Proposition \[prop:density\_quasi\] will be subsequently substantially sharpened by eliminating the frame bounds in the conclusion. Necessary density conditions {#sec_density_2} ---------------------------- The following result provides necessary density conditions for several spanning properties. Note that Kleppner’s condition is not assumed in parts (i) and (ii). \[thm:necessary\_density\] Let $\Gamma \subseteq G$ be a lattice and let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. (i) If $\pi|_{\Gamma}$ admits a cyclic vector, then $\operatorname*{vol}(G / \Gamma) d_{\pi} \leq 1$. In particular, if $\pi|_{\Gamma}$ admits a frame vector, then $\operatorname*{vol}(G / \Gamma) d_{\pi} \leq 1$. (ii) If $\pi|_{\Gamma}$ admits a Riesz vector, then $\operatorname*{vol}(G / \Gamma) d_{\pi} \geq 1$. (iii) Suppose $(\Gamma, \sigma)$ satisfies Kleppner’s condition. If $\pi(\Gamma)''$ admits a separating vector, then $\operatorname*{vol}(G / \Gamma)d_{\pi} \geq 1$. \(i) Suppose first that there exists a vector $g \in \mathcal{H}_{\pi}$ such that ${\pi (\Gamma) g}$ is a Parseval frame for $\mathcal{H}_{\pi}$. Then, $\operatorname*{vol}(G/\Gamma) d_{\pi} = \|g \|_{\mathcal{H}_{\pi}}^2$, by Proposition \[prop:density\_quasi\]. Since ${\pi (\Gamma) g}$ is a Bessel sequence with bound $1$, necessarily $\| g \|_{\mathcal{H}_{\pi}}^2 \leq 1$. Hence $\operatorname*{vol}(G/\Gamma) d_{\pi} \leq 1$, as claimed. Second, if $\pi|_{\Gamma}$ admits a cyclic vector, then it also admits a Parseval frame by Proposition \[prop:cyclic\_tight-frame\]. \(ii) Suppose that $\pi|_{\Gamma}$ admits a Riesz vector. Then, by Remark \[rem\_Riesz\_on\], there also exists $g \in \mathcal{H}_{\pi}$ such that ${\pi (\Gamma) g}$ is orthonormal. Hence, ${\pi (\Gamma) g}$ has Bessel bound $1$, and, therefore, by Proposition \[prop:density\_quasi\], $d_{\pi}^{-1} \leq \operatorname*{vol}(G/ \Gamma)$. \(iii) Finally, under Kleppner’s condition, if $\pi(\Gamma)''$ admits a separating vector, then it also admits an orthonormal sequence by Proposition \[prop:separating\_riesz\], and we can apply part (ii). The idea of relating the orthogonality relations and the frame inequalities for proving a density theorem as Theorem \[thm:necessary\_density\] was used in Janssen’s “classroom proof” of the density theorem for Gabor frames [@janssen_density]. The use of an auxiliary tight frame to deduce the density condition can be be found in [@christensen2016introduction Theorem 11.3.1]. A similar combination of these ideas have been used in [@jakobsen2016density]. In this article, these ideas are further refined, implying necessary conditions for completeness. The arguments for Riesz sequences seem to be new. Critical density ---------------- This section is devoted to the spanning properties of $\pi|_{\Gamma}$ for lattices possessing the *critical density* $\operatorname*{vol}(G/\Gamma) d_{\pi} = 1$. \[lem:critical\_complete\] Let $\Gamma \subseteq G$ be a lattice and let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. Suppose $g \in \mathcal{H}_{\pi}$ is a unit vector such that ${\pi (\Gamma) g}$ is an orthonormal system in ${\mathcal{H}_{\pi}}$. Then the following are equivalent: (i) The system ${\pi (\Gamma) g}$ is complete in $\mathcal{H}_{\pi}$. (ii) $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$. That (i) implies (ii) follows from Proposition \[prop:density\_quasi\]. Conversely, suppose that $\operatorname*{vol}(G/ \Gamma) d_{\pi} = 1$. Let $\Omega \subseteq G$ be a right fundamental domain of $\Gamma \subseteq G$, and $\{f_n: n \in \mathbb{N}\}$ a norm dense subset of ${\mathcal{H}_{\pi}}$. Fix $n \in \mathbb{N}$. By the orthogonality relations and the assumption $\| g \|_{\mathcal{H}_{\pi}} = 1$, $$\begin{aligned} d_{\pi}^{-1} \| f_n \|_{\mathcal{H}_{\pi}}^2 = \int_G | \langle f_n, \pi (x) g \rangle |^2 \; d\mu_G (x) = \sum_{\gamma \in \Gamma} \int_{\Omega} | \langle f_n, \pi (x \gamma) g \rangle |^2 \; d\mu_G (x) . \end{aligned}$$ Since $\operatorname*{vol}(G/ \Gamma)d_{\pi} = 1$, $$\begin{aligned} 0 &= d_{\pi} \int_{\Omega} \| f_n \|_{\mathcal{H}_{\pi}}^2 \; d\mu_G(x) - d_{\pi} \sum_{\gamma \in \Gamma} \int_{\Omega} | \langle f_n, \pi (x \gamma) g \rangle |^2 \; d\mu_G (x) \\ &= d_{\pi} \int_{\Omega} \bigg( \|f_n\|_{\mathcal{H}_{\pi}}^2 - \sum_{\gamma \in \Gamma} | \langle f_n, \pi (x \gamma) g \rangle |^2 \bigg)\; d\mu_G (x). {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:nonzero_integrand}\end{aligned}$$ But $\sum_{\gamma \in \Gamma} | \langle f_n, \pi (x \gamma) g \rangle |^2 \leq \| f \|^2_{\mathcal{H}_{\pi}}$ for any $x \in \Omega$ by Bessel’s inequality. Thus the integrand in is $0$ for $x \in \Omega \setminus E_n$, where $E_n \subseteq \Omega$ is a null measure set. Since $\bigcup_{n \in \mathbb{N}} E_n$ has null measure, we can choose $x_0 \in \Omega \setminus \bigcup_{n \in \mathbb{N}} E_n$. Therefore, $$\begin{aligned} \sum_{\gamma \in \Gamma} | \langle f, \pi (x_0 \gamma) g \rangle |^2 = \| f \|^2_{\mathcal{H}_{\pi}},\end{aligned}$$ holds for all $f \in \{f_n : n \in \mathbb{N}\}$, and extends by density to all $f \in {\mathcal{H}_{\pi}}$. Replacing $f$ by $\pi(x_0)f$ gives $ \sum_{\gamma \in \Gamma} | \langle f, \pi (\gamma) g \rangle |^2 = \| f \|^2_{\mathcal{H}_{\pi}}$, for all $f \in {\mathcal{H}_{\pi}}$. This shows that ${\pi (\Gamma) g}$ is complete. \[prop:critical\_density\] Let $\Gamma \subseteq G$ be a lattice and let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. The following assertions are equivalent: (i) The system ${\pi (\Gamma) g}$ is a Riesz (resp. orthonormal) basis for $\mathcal{H}_{\pi}$. (ii) The system ${\pi (\Gamma) g}$ is a frame (resp. Parseval frame) for $\mathcal{H}_{\pi}$ with $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$. (iii) The system ${\pi (\Gamma) g}$ is a Riesz (resp. orthonormal) sequence in $\mathcal{H}_{\pi}$ with $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$. The implications (i) $\Rightarrow$ (ii) and (i) $\Rightarrow$ (iii) follow directly from Theorem \[thm:necessary\_density\]. \(ii) $\Rightarrow$ (i) First, we show that a Parseval frame $\pi(\Gamma) g$ with $\operatorname*{vol}(G/\Gamma) d_{\pi} = 1$ is an orthonormal basis for ${\mathcal{H}_{\pi}}$. Indeed, we have $\| g \|_{{\mathcal{H}_{\pi}}}^2 = 1$ by Proposition \[prop:density\_quasi\], and hence $$\begin{aligned} \label{eq:parseval_unit} 1 = \| \pi(\gamma') g \|_{{\mathcal{H}_{\pi}}}^2 = \sum_{\gamma \in \Gamma} | \langle \pi(\gamma') g, \pi(\gamma) g \rangle|^2 = 1 + \sum_{\gamma \in \Gamma \setminus \{\gamma'\}} |\langle \pi(\gamma') g, \pi(\gamma) g \rangle|^2,\end{aligned}$$ which shows that $\langle \pi(\gamma') g, \pi(\gamma) g \rangle = \delta_{\gamma, \gamma'}$ for all $\gamma, \gamma' \in \Gamma$. Thus $\pi(\Gamma) g $ is an orthonormal basis for $\mathcal{H}_{\pi}$. Second, if $\pi(\Gamma) g$ is an arbitrary frame with $\operatorname*{vol}(G/\Gamma) d_{\pi} = 1$, then $\pi (\Gamma) S_{g, \Gamma}^{-1/2} g$ is a Parseval frame for ${\mathcal{H}_{\pi}}$, and hence an orthonormal basis for ${\mathcal{H}_{\pi}}$ by the above. But ${\pi (\Gamma) g}= S_{g,\Gamma}^{1/2} \pi(\Gamma) S_{g,\Gamma}^{-1/2} g$, and thus ${\pi (\Gamma) g}$ is a Riesz basis for ${\mathcal{H}_{\pi}}$. \(iii) $\Rightarrow$ (i) Suppose ${\pi (\Gamma) g}$ is a Riesz sequence in ${\mathcal{H}_{\pi}}$. Then $S^{-1/2}_{g,\Gamma} : [\pi(\Gamma) g] \to [\pi(\Gamma)g]$ is well-defined and bounded. Hence, the system $\pi(\Gamma) S_{g,\Gamma}^{-1/2} g $ is orthonormal in $\mathcal{H}_{\pi}$ by Remark \[rem\_Riesz\_on\], thus complete by Lemma \[lem:critical\_complete\]. As above, $\pi (\Gamma) g = S_{g,\Gamma}^{1/2} \pi(\Gamma) S_{g,\Gamma}^{-1/2} g$, showing that ${\pi (\Gamma) g}$ is a Riesz basis. Moreover, if ${\pi (\Gamma) g}$ itself is orthonormal, then its completeness follows directly by Lemma \[lem:critical\_complete\]. Proof of the density theorem {#sec:density_thm} ============================ We finally can prove the main result of the article. \[th\_main\] Let $\Gamma \subseteq G$ be a lattice in a second countable unimodular group $G$. Let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. Suppose that $(\Gamma, \sigma)$ satisfies Kleppner’s condition. Then the following assertions hold: - If $\operatorname*{vol}(G / \Gamma) d_{\pi} < 1$, then $\pi|_{\Gamma}$ admits a Parseval frame, but $\pi(\Gamma)''$ does not admit a separating vector. (In particular, $\pi|_{\Gamma}$ does not admit a Riesz vector.) - If $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$, then $\pi|_{\Gamma}$ admits an orthonormal basis. - If $\operatorname*{vol}(G / \Gamma) d_{\pi} > 1$, then $\pi|_{\Gamma}$ admits an orthonormal sequence, but not a cyclic vector. (In particular, $\pi|_{\Gamma}$ does not admit a frame vector.) \(i) Assume that $\operatorname*{vol}(G / \Gamma) d_{\pi} < 1$. Then, by Theorem \[thm:necessary\_density\], $\pi(\Gamma)''$ does not admit a separating vector. Combining this information with Proposition \[prop\_either\], it follows that $\pi|_{\Gamma}$ admits a cyclic vector, and by Proposition \[prop:cyclic\_tight-frame\] also a Parseval frame. The “in particular" part also follows from Theorem \[thm:necessary\_density\]. \(ii) Assume that $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$. By Proposition \[prop\_either\], $\pi|_{\Gamma}$ admits either a cyclic or separating vector. In the first case, by Proposition \[prop:cyclic\_tight-frame\], $\pi|_{\Gamma}$ also admits a a Parseval frame ${\pi (\Gamma) g}$, and hence an orthonormal basis by Proposition \[prop:critical\_density\]. In the second case, by Proposition \[prop:separating\_riesz\], $\pi|_{\Gamma}$ admits a a orthonormal sequence ${\pi (\Gamma) g}$, which forms an orthonormal basis by Proposition \[prop:critical\_density\]. \(iii) Assume that $\operatorname*{vol}(G / \Gamma) d_{\pi} > 1$. Then, by Theorem \[thm:necessary\_density\], $\pi|_{\Gamma}$ does not admit a cyclic vector. Combining this information with Proposition \[prop\_either\], it follows that $\pi(\Gamma)''$ admits a separating vector, and by Proposition \[prop:separating\_riesz\], also an orthonormal sequence. Proof of Theorem \[thm:intro\_ICC\] ----------------------------------- The hypotheses of Theorem \[thm:intro\_ICC\] are a particular case of Theorem \[th\_main\]. Indeed, an ICC lattice $\Gamma$ satisfies Kleppner’s condition for any cocycle $\sigma$, in particular, for $\sigma \equiv 1$. [^2] [$\square$]{} A far reaching generalization of Theorem \[thm:intro\_ICC\] without the ICC condition is due to Bekka [@bekka2004square]; see Section \[sec:examples\]. Examples and applications {#sec:examples} ========================= The density theorem for semisimple Lie groups --------------------------------------------- For certain center-free semisimple Lie groups, a lattice is automatically ICC, and hence Kleppner’s condition is satisfied. For reference purposes, we state Theorem \[th\_main\] in this setting. \[th\_semi\] Let $G$ be a center-free connected semisimple real Lie group all of whose connected, normal, compact subgroups are trivial.[^3] Let $(\pi, {\mathcal{H}_{\pi}})$ be a discrete series $\sigma$-representation of $G$ of formal dimension $d_{\pi} > 0$. Let $\Gamma \subseteq G$ be a lattice. Then - If $\operatorname*{vol}(G / \Gamma) d_{\pi} < 1$, then $\pi|_{\Gamma}$ admits a Parseval frame, but $\pi(\Gamma)''$ does not admit a separating vector. (In particular, $\pi|_{\Gamma}$ does not admit a Riesz vector.) - If $\operatorname*{vol}(G / \Gamma) d_{\pi} = 1$, then $\pi|_{\Gamma}$ admits an orthonormal basis. - If $\operatorname*{vol}(G / \Gamma) d_{\pi} > 1$, then $\pi|_{\Gamma}$ admits an orthonormal sequence, but not a cyclic vector. (In particular, $\pi|_{\Gamma}$ does not admit a frame vector.) Under the hypothesis, the lattice $\Gamma$ is an ICC group, see, e.g., [@goodman1989coxeter Lemma 3.3.1] or [@bekka2004square Theorem 2]. Therefore, $(\Gamma, \sigma)$ satisfies Kleppner’s condition, and the conclusion follows from Theorem \[th\_main\]. As we show in Example \[sec:SL\_ICC\], the conclusion of Theorem \[th\_semi\] may fail when the center of the group is non-trivial. A more general version of Theorem \[th\_semi\], that does not require the ICC condition, was derived by Bekka [@bekka2004square], and applies to semisimple Lie groups with a possibly non-trivial center [@bekka2004square Theorem 2], and to a class of algebraic groups over more general fields. Theorem \[th\_semi\] can also be phrased more generally for such algebraic groups, provided they have a trivial center. We now illustrate an important instance of Theorem \[th\_semi\]. \[sec\_bergman\] The group $G = \mathrm{PSL}(2, \mathbb{R}) = \mathrm{SL}(2, \mathbb{R}) / \{-I, I\}$ is a connected simple Lie group with trivial center [@fulton1991representation; @varadarajan1984lie], and acts on the upper half plane $$\mathbb{C}^+ = \{z \in \mathbb{C} : \Im (z) > 0\}$$ through Moebius transforms as $$G \times \mathbb{C}^+ \ni \big( \begin{pmatrix} a & b \\ c & d \end{pmatrix}, z \big) \mapsto \frac{a z + b}{cz + d} \in \mathbb{C}^+.$$ The measure $d\mu (z) = (\Im (z))^{-2} dx dy$, where $z = x+iy$ and $dxdy$ is the Lebesgue measure on $\mathbb{C}^+$, is $G$-invariant. Let $\mathrm{PSO}(2,\mathbb{R}) := \mathrm{SO}(2,\mathbb{R}) / \{-I, I\}$ be the compact subgroup of rotations. We use the diffeomorphism, $$\begin{aligned} \label{eq_diff_i} G / \mathrm{PSO}(2, \mathbb{R}) &\to \mathbb{C}^+, \\ [ {\mathrm{m}}] &\mapsto m \cdot i,\end{aligned}$$ to fix a Haar measure on $G / \mathrm{PSO}(2, \mathbb{R})$, and equip $\mathrm{PSO}(2, \mathbb{R})$ with a normalized Haar measure $\mu_{T}$ of total measure $1$. This fixes the Haar measure $\mu_G$ on $G$ as $d\mu_G \simeq d\mu d\mu_T$. With this normalization, for measurable $E \subseteq \mathbb{C}^+$, $$\begin{aligned} \label{eq_norm_haar} \mu_G\left( \{m \in G: m \cdot i \in E \}\right) = \mu (E).\end{aligned}$$ In the remainder of this article, the Haar measure on $G = \mathrm{PSL}(2,\mathbb{R})$ is always assumed to have this normalization. For $\alpha > 1$, define the measure $d\mu_{\alpha} (z) = (\Im (z))^{\alpha -2} dxdy$ and the weighted Bergman space of holomorphic functions $ A^2_{\alpha} (\mathbb{C}^+) := \mathcal{O} (\mathbb{C}^+) \cap L^2 (\mathbb{C}^+, d\mu_{\alpha})$, equipped with norm $$\begin{aligned} \label{eq:bergman_norm} \| f \|^2_{A^2_{\alpha}} = \int_{\mathbb{C}^+} |f(z)|^2 \; d\mu_{\alpha} (z).\end{aligned}$$ Define $j : \mathrm{SL}(2, \mathbb{R}) \times \mathbb{C}^+ \to \mathbb{C} \setminus \{0\}$ by $$\begin{aligned} \label{eq_pialpha} j({\mathrm{m}}, z) = (cz + d)^{-1}, \quad {\mathrm{m}}= \begin{pmatrix} a & b \\ c & d \end{pmatrix}, z \in \mathbb{C}^+.\end{aligned}$$ Then $j$ satisfies $ j({\mathrm{m}}_1 {\mathrm{m}}_2, z) = j({\mathrm{m}}_1, {\mathrm{m}}_2 z) j ({\mathrm{m}}_2, z) $ for all ${\mathrm{m}}_1, {\mathrm{m}}_2 \in \mathrm{SL}(2, \mathbb{R})$ and $z \in \mathbb{C}^+$. Let $z^\alpha$ be defined with respect to the principal branch of the argument: $\arg(z) \in (-\pi, \pi]$. Since $j({\mathrm{m}}, z) \in \mathbb{C}\setminus\mathbb{R}$, we can form $j({\mathrm{m}}, z)^\alpha$, and $$j({\mathrm{m}}_1 {\mathrm{m}}_2, z)^\alpha = \lambda({\mathrm{m}}_1, {\mathrm{m}}_2, \alpha, z) j({\mathrm{m}}_1, {\mathrm{m}}_2 z)^\alpha j ({\mathrm{m}}_2, z)^\alpha$$ for a unimodular function $\lambda({\mathrm{m}}_1, {\mathrm{m}}_2, \alpha, z)$. The analyticity of $j({\mathrm{m}}, z)$ on $z$, implies that $\lambda({\mathrm{m}}_1, {\mathrm{m}}_2, \alpha, z)=\lambda({\mathrm{m}}_1, {\mathrm{m}}_2,\alpha)$ is independent of $z$. A projective unitary representation $(\pi'_{\alpha}, A^2_{\alpha} (\mathbb{C}^+))$ of $\mathrm{SL}(2, \mathbb{R})$ is therefore given by $$\begin{aligned} \label{eq:holomorphicdiscreteseries} (\pi'_{\alpha}({\mathrm{m}}) f )(z) = j({\mathrm{m}}^{-1}, z)^{\alpha} f ({\mathrm{m}}^{-1} \cdot z), \quad {\mathrm{m}}\in \mathrm{SL}(2, \mathbb{R}), \; z \in \mathbb{C}^+.\end{aligned}$$ Let $\tau: G \to \mathrm{SL}(2, \mathbb{R})$ be a Borel cross-section of the quotient map, i.e., a Borel measurable function that chooses a representative; see [@mackey1952induced Lemma 1.1.] or [@varadarajan1985geometry Chapter 5]. Since $j(-{\mathrm{m}}, z)=-j({\mathrm{m}}, z)$, $\pi_\alpha := \pi'_\alpha \circ \tau$ defines a projective unitary representation of $G$ on $A^2_{\alpha} (\mathbb{C}^+)$, the so-called *holomorphic discrete series $\sigma$-representation*. For any $\alpha > 1$, $(\pi_{\alpha}, A^2_{\alpha} (\mathbb{C}^+))$ is irreducible and square-integrable of formal dimension $$d_{\pi_{\alpha}} = \frac{\alpha - 1}{4\pi}.$$ See [@radulescu1998berezin; @robert1983introduction] for the details. Lattices $\Gamma \subseteq G$ are known as *Fuchsian groups*. By the normalization , we have $\operatorname*{vol}(G/\Gamma)=\mu(D)$, where $D \subseteq \mathbb{C}^+$ is a so-called *Dirichlet fundamental domain* for $\Gamma$, that provides the tessellation $\mathbb{C}^+ = \bigcup_{\gamma \in \Gamma} \gamma D$, up to sets of null measure. According to Theorem \[th\_main\], the existence of a function $g \in A^2_{\alpha} (\mathbb{C}^+)$ such that $\pi_{\alpha}(\Gamma) g$ is complete in (resp. frame for, resp. Parserval frame for) $A^2_{\alpha} (\mathbb{C}^+)$ is equivalent to the condition $$\label{eq_compl_sl2r} \operatorname*{vol}(G/\Gamma) \leq \frac{4\pi}{\alpha - 1},$$ while the existence of a Riesz sequence $\pi_{\alpha}(\Gamma) g$ (resp. orthonormal sequence, resp. $g$ separating vector) in $A^2_{\alpha} (\mathbb{C}^+)$ is equivalent to the condition $$\label{eq_sep_sl2r} \operatorname*{vol}(G/\Gamma) \geq \frac{4\pi}{\alpha - 1}.$$ For examples of Fuchsian groups, and formulae for their co-volume, see [@beardon1983geometry]. The following example demonstrates that Kleppner’s condition (or the ICC condition) cannot be removed as an assumption in Theorem \[th\_main\]. \[sec:SL\_ICC\] Let $G = \mathrm{SL}(2, \mathbb{R})$, with center $Z(G) = \{-I, I\}$. For $\alpha > 1$, the group $G$ acts on the Bergman space $A^2_{\alpha} (\mathbb{C}^+)$ by the representation $\pi'_{\alpha}$ whose action is given by . Equip $Z(G)$ with the counting measure and $G / Z(G) = \mathrm{PSL}(2, \mathbb{R}) $ with the Haar measure $\mu_{G/Z}$ normalized as in Example \[sec\_bergman\]. The Haar measure on $G$ is then fixed by Weil’s formula: $d\mu_G \simeq d\mu_{G/Z} d\mu_Z$. By the orthogonality relations of the holomorphic discrete series of $\mathrm{PSL}(2, \mathbb{R})$, it follows then that, for $f \in A^2_{\alpha} (\mathbb{C}^+)$, $$\int_G | \langle f, \pi'_{\alpha}(x) f \rangle|^2 \; d\mu_G (x) = \int_{G/Z(G)} \sum_{ Z(G)} |\langle f, \pi'_{\alpha} (\dot{x}) f \rangle |^2 \; d\mu_{G/Z} (xZ) = (2 \cdot d^{-1}_{\pi_{\alpha}} )\| f \|_{{\mathcal{H}_{\pi}}}^4,$$ where $d_{\pi_{\alpha}} := (\alpha - 1)/(4\pi)$ as in Example \[sec\_bergman\]. Thus $\pi'_{\alpha}$ is a discrete series representation of $G$ of formal dimension $d'_{\pi_{\alpha}} = (\alpha - 1)/(8\pi)$. Let $\Gamma \subseteq G$ be a lattice such that $Z(G) \subset \Gamma$, and $\Omega \subset G$ a right fundamental domain. Denote by $p : G \to G/Z(G)$ the canonical projection, and $\widetilde{\Gamma} = p(\Gamma)$. As $Z(G) \subset \Gamma$, $\chi_{\Omega}(x)+\chi_{\Omega}(-x)=\chi_{p(\Omega)}(xZ)$, and, therefore, Weil’s formula gives $$\begin{aligned} \operatorname*{vol}(G/\Gamma) &= \int_G \chi_{\Omega} (x) \; d\mu_G (x) = \int_{G / Z(G)} \chi_{p(\Omega)}(xZ) \; d\mu_{G/Z} (xZ)\\ &= \mu_{G/Z} (p(\Omega)) = \operatorname*{vol}(\mathrm{PSL}(2, \mathbb{R}) / \widetilde{\Gamma}),\end{aligned}$$ as $p(\Omega)$ is a fundamental domain for $\widetilde{\Gamma}$ in $\mathrm{PSL}(2, \mathbb{R})$. Consider the representation $\pi_\alpha$ from Example \[sec\_bergman\]. Since $\pi'_\alpha(-I)=\pm I$, for any $g \in A^2_{\alpha} (\mathbb{C}^+)$, $$\begin{aligned} [\pi'_{\alpha} (\Gamma) g] = [\pi_{\alpha} (\widetilde{\Gamma}) g].\end{aligned}$$ We conclude that there exists $g \in A^2_{\alpha} (\mathbb{C}^+)$ such that $\pi'_{\alpha} (\Gamma) g$ is complete if and only if $$\begin{aligned} \label{completeness_SL} \operatorname*{vol}(G/\Gamma) = \operatorname*{vol}(\mathrm{PSL}(2,\mathbb{R}) / \widetilde{\Gamma}) \leq \frac{4\pi}{\alpha - 1} = \frac{1}{2} (d'_{\pi_{\alpha}})^{-1},\end{aligned}$$ or, equivalently, $\operatorname*{vol}(G/\Gamma) d'_{\pi_{\alpha}} \leq 1/2$. (This conclusion follows also from Bekka’s result [@bekka2004square Example 1], where a different normalizations of the Haar measure is used.) Therefore, the completeness part of Theorem \[th\_main\] fails for $G$ and $\Gamma$. Of course, $(\Gamma,\sigma)$ does not satisfy Kleppner’s condition, as the central element $-I \in G$ has a finite conjugacy class. Second, note that there does not exist a Riesz sequence in $A^2_{\alpha} (\mathbb{C}^+)$ of the form $\pi'_{\alpha} (\Gamma) g$, regardless of the value of $\operatorname*{vol}(G/\Gamma)$, as the (indexed) system $\pi'_{\alpha} (\Gamma) g$ is always linearly dependent: $\pi'_{\alpha} (I) g = g = \pm \pi'_{\alpha} (-I)g$. Hence, also in that respect, the conclusion of Theorem \[th\_main\] fails for $G$ and $\Gamma$. ### Perelomov’s uniqueness problem {#sec:perelomov_uniqueness} A set of points $\Lambda \subseteq \mathbb{C}^+$ is called a *set of uniqueness* for the Bergman space $A^2_{\alpha} (\mathbb{C}^+)$ if the only function $f \in A^2_{\alpha} (\mathbb{C}^+)$ that vanishes identically on $\Lambda$ is the zero function. Perelomov [@perelomov1973coherent] studied this question when $\Lambda$ is the orbit of a point $w \in \mathbb{C}^+$ through a Fuchsian group $\Gamma$ in $G = \mathrm{PSL}(2,\mathbb{R})$.[^4] The link with lattice orbits of $\pi_{\alpha}$ is provided by the special choice of vector ${k^{(\alpha)}_w}(z)=2^{\alpha-2}\pi^{-1}(\alpha-1)i^\alpha (z-\overline{w})^{-\alpha}$, which has the *reproducing property*: $$\begin{aligned} \label{eq_rep_for} f({\mathrm{m}}\cdot w) = c_\alpha (cw+d)^{-\alpha} \langle f, \pi_{\alpha}({\mathrm{m}}) {k^{(\alpha)}_w}\rangle_{A^2_\alpha}, \qquad f \in A^2_{\alpha} (\mathbb{C}^+), {\mathrm{m}}\in \mathrm{PSL}(2, \mathbb{R}),\end{aligned}$$ where $c_\alpha \in \mathbb{T}$ is a unimodular constant and the notation of is used. Hence, $\Lambda=\Gamma w$ is a set of uniqueness for $A^2_{\alpha} (\mathbb{C}^+)$ if and only if $\pi_{\alpha} (\Gamma) {k^{(\alpha)}_w}$ is complete in $A^2_{\alpha} (\mathbb{C}^+)$. Perelomov [@perelomov1973coherent] showed that this is the case if $$\begin{aligned} \label{eq_perelomov} \#F_w \operatorname*{vol}(G/\Gamma) < \frac{4\pi}{\alpha - 1},\end{aligned}$$ where $ F_w = \left\{\gamma \in \Gamma : \gamma \cdot w = w \right\} $ is the stabilizer subgroup of $w$.[^5] When $\#F_w=1$, the sufficient condition for the completeness of $\pi_{\alpha}(\Gamma) {k^{(\alpha)}_w}$ in $A^2_{\alpha} (\mathbb{C}^+)$ almost matches , which is necessary for the completeness of *any* orbit $\pi_{\alpha}(\Gamma) g$. For $\#F_w>1$, a necessary condition for the completeness of $\pi_{\alpha}(\Gamma) {k^{(\alpha)}_w}$ in $A^2_{\alpha} (\mathbb{C}^+)$ almost matching was proved by Kelly-Lyth [@kelly Theorem 5.4]: if $\Lambda$ is a uniqueness set for $A^2_{\alpha} (\mathbb{C}^+)$, then $$\begin{aligned} \label{eq_kelly_lyth} \#F_w \operatorname*{vol}(G/\Gamma) \leq \frac{4\pi}{\alpha - 1}.\end{aligned}$$ Thus, while $(\pi_{\alpha}, A^2_{\alpha} (\mathbb{C}^+))$ admits a cyclic vector $g$ if and only if $\operatorname*{vol}(G/\Gamma) \leq \frac{4\pi}{\alpha - 1}$, in the smaller range $\operatorname*{vol}(G/\Gamma) < \frac{4\pi}{\#F_w(\alpha - 1)}$ the specific choice $g={k^{(\alpha)}_w}$ is possible, and in the range $\frac{4\pi}{\#F_w(\alpha - 1)} < \operatorname*{vol}(G/\Gamma) \leq \frac{4\pi}{\alpha - 1}$ it is not. The completeness of $\pi_{\alpha} (\Gamma){k^{(\alpha)}_w}$ when $\#F_w \operatorname*{vol}(G/\Gamma) = \frac{4\pi}{\alpha - 1}$ does not seem to have been treated in the literature. Perelomov’s original work also contains a necessary condition for the completeness of $\pi_{\alpha}(\Gamma) {k^{(\alpha)}_w}$ in $A^2_{\alpha} (\mathbb{C}^+)$, formulated in terms of the smallest weight $m_0^+$ for which the space of parabolic $\Gamma$-modular forms on $\mathbb{C}^+$ is at least two-dimensional [@perelomov1973coherent Theorem 3]: if $\Lambda$ is a uniqueness set for $A^2_{\alpha} (\mathbb{C}^+)$, then $$\begin{aligned} \label{eq_nec_per} \frac{2 \pi}{m_0^+} \leq \frac{4\pi}{\alpha - 1}.\end{aligned}$$ As shown in [@kelly Lemma 5.3], $$\begin{aligned} \frac{2 \pi}{m_0^+} \leq \frac{\operatorname*{vol}(G/\Gamma)}{1+ \#P} \leq \operatorname*{vol}(G/\Gamma),\end{aligned}$$ where $\#P$ denotes the number of inequivalent cusps for $\Gamma$. Thus the necessity of for cyclicity is stronger than Perelomov’s automorphic weight bound for the cyclicity of one specific vector , but weaker than Kelly-Lyth’s . Under the assumption that fails, Perelomov uses certain $\Gamma$-modular forms to construct a non-zero function in $A^2_{\alpha} (\mathbb{C}^+)$ that vanishes on $\Gamma w$. Under the assumption that fails, Kelly-Lyth also provides such function, by calculating the so-called upper Beurling-Seip density of $\Gamma w$ in terms of the co-volume of $\Gamma$, and by resorting to Seip’s interpolation theorem [@seip1993beurling]. While this article gives a very elementary argument for the necessity of for the completeness of $\pi_\alpha(\Gamma)g_w$, we do not have a similarly simple argument for . ### Frames and Riesz sequences of reproducing kernels {#sec:frame_kernel} By Theorem \[th\_semi\], under , there exists $g \in A^2_{\alpha} (\mathbb{C}^+)$ such that the orbit $\pi_{\alpha}(\Gamma) g$ is a (Parseval) frame for $A^2_{\alpha} (\mathbb{C}^+)$. In light of Section \[sec:perelomov\_uniqueness\], it is natural to ask whether the specific choice $g={k^{(\alpha)}_w}$ also provides a frame. Here the answer depends on whether or not $\Gamma$ is co-compact (that is, $G/\Gamma$ is compact). Using , the frame property reads $$\begin{aligned} \label{eq_samp_1} A{\lVertf\rVert}^2_{A^2_{\alpha}} \leq \Im(w)^{-\alpha} \sum_{\gamma \in \Gamma} \Im(\gamma \cdot w)^\alpha {\ensuremath{\left| f(\gamma \cdot w) \right| }}^2 \leq B {\lVertf\rVert}^2_{A^2_{\alpha}}, \qquad \qquad f \in A^2_{\alpha} (\mathbb{C}^+),\end{aligned}$$ for some constants $A,B>0$. The stabilizer subgroup $F_w$ is finite because it is simultaneously contained in the discrete set $\Gamma$ and in the compact subgroup ${\mathrm{m}}_0 \mathrm{PSO}(2,\mathbb{R}) {\mathrm{m}}_0^{-1}$, where ${\mathrm{m}}_0 i = w$. Hence, we can rewrite as a *sampling inequality*: $$\begin{aligned} \label{eq_samp_2} A'{\lVertf\rVert}^2_{A^2_{\alpha}} \leq \Im(w)^{-\alpha} \# F_w \sum_{z \in \Gamma w} \Im(z)^\alpha {\ensuremath{\left| f(z) \right| }}^2 \leq B' {\lVertf\rVert}^2_{A^2_{\alpha}}, \qquad \qquad f \in A^2_{\alpha} (\mathbb{C}^+).\end{aligned}$$ Based on the characterization of sampling inequalities by Seip [@seip1993beurling], Kelly-Lyth showed that if $\Gamma$ is not co-compact, then $\Gamma w$ never satisfies , because its so-called lower Beurling-Seip density is zero [@kelly p.44]. Thus, in this case, $\pi_{\alpha} (\Gamma) {k^{(\alpha)}_w}$ fails to be a frame for $A^2_{\alpha} (\mathbb{C}^+)$. On the other hand, if $\Gamma$ is co-compact, the lower Beurling-Seip density of $\Gamma w$ can be computed in term of the co-volume of $\Gamma$ and yields that $\pi_{\alpha} (\Gamma) {k^{(\alpha)}_w}$ is a frame for $A^2_{\alpha} (\mathbb{C}^+)$ if and only if holds, see [@kelly p.44]. Similarly, under , Theorem \[th\_semi\] provides $g \in A^2_{\alpha} (\mathbb{C}^+)$ such that $\pi_{\alpha} (\Gamma) g$ forms a Riesz sequence in $A^2_{\alpha} (\mathbb{C}^+)$, and one may wonder if, under the corresponding strict inequality, the particular choice $g={k^{(\alpha)}_w}$ is also possible. This is indeed the case if the stabilizer subgroup $F_w$ is trivial: as shown by Kelly-Lyth [@kelly Theorem 5.8] by invoking Seip’s interpolation theorem [@seip1993beurling], the system $\pi_{\alpha} (\Gamma) {k^{(\alpha)}_w}$ is a Riesz sequence if and only if $$\operatorname*{vol}(G/\Gamma) > \frac{4\pi}{\alpha - 1}.$$ If the stabilizer subgroup $F_w$ is non-trivial, then $\pi_{\alpha}(\Gamma) {k^{(\alpha)}_w}$ is not a Riesz sequence, because it is linearly dependent (as an indexed set). Indeed, shows that $\pi_{\alpha}(\gamma) {k^{(\alpha)}_w}$ is a multiple of ${k^{(\alpha)}_w}$ when $\gamma \in F_w$. To make the problem meaningful, we can eliminate repetitions by considering the reduced orbit $$\begin{aligned} \tilde\pi_{\alpha} (\Gamma){k^{(\alpha)}_w}=\left\{\pi_{\alpha} (\gamma){k^{(\alpha)}_w}: \gamma \in \Gamma_w \right\},\end{aligned}$$ where $\Gamma_w$ is a set of representatives of $\Gamma/F_w$. With this correction, [@kelly Theorem 5.8] implies that if $\Gamma \subset \mathrm{PSL}(2, \mathbb{R})$ is a Fuchsian group satisfying $$\#F_w \operatorname*{vol}(G/\Gamma) > \frac{4\pi}{\alpha - 1},$$ then $\tilde\pi_{\alpha} (\Gamma) {k^{(\alpha)}_w}$ is a Riesz sequence in $A^2_{\alpha} (\mathbb{C}^+)$. Thus, in contrast to the frame property, a Riesz sequence can exist even for lattices that are not co-compact. ### Perelomov’s problem with respect to other special vectors The functions $${h^{(\alpha)}_n}(z) = \Big(\frac{z-i}{z+i} \Big)^n (z+i)^{-\alpha}, \qquad n \in \mathbb{N}_0,$$ form a distinguished orthogonal basis for $A^2_{\alpha} (\mathbb{C}^+)$. Note that $h_0^{(\alpha)}$ is a multiple of the reproducing kernel $k^{(\alpha)}_i \in A^2_{\alpha} (\mathbb{C}^+)$ at $i$ discussed in Section \[sec:perelomov\_uniqueness\]. In the language of Perelomov [@perelomov1972coherent; @perelomov1973coherent], each ${h^{(\alpha)}_n}$ is a *stationary vector* of the subgroup of rotations $\mathrm{PSO}(2, \mathbb{R})$ in $G = \mathrm{PSL}(2, \mathbb{R})$: for each $r \in \mathrm{PSO}(2, \mathbb{R})$, there exists $\phi_r \in \mathbb{R}$ such that: $$\begin{aligned} \pi_{\alpha} (r) {h^{(\alpha)}_n}= e^{i \phi_r} {h^{(\alpha)}_n},\end{aligned}$$ as a direct calculation shows. Because of stationarity, given a Fuchsian group $\Gamma \subset G$, the orbit $\pi_{\alpha}(\Gamma) {h^{(\alpha)}_n}$ can be reduced by selecting for each $\gamma \in \Gamma$ just one representative modulo $\mathrm{PSO}(2, \mathbb{R})$, the specific choice being immaterial. The resulting set is a *subsystem of coherent states* in the sense of Perelomov[@perelomov1972coherent; @perelomov1973coherent], and it is complete in $A^2_{\alpha} (\mathbb{C}^+)$ if and only if the orbit $\pi_\alpha(\Gamma) {h^{(\alpha)}_n}$ is. The coherent state subsystems associated with ${h^{(\alpha)}_n}$ can be more concretely described as follows [@bertrand2002characterization; @klauder1994wavelets]. The subgroup of affine transformations $$\begin{aligned} \label{eq:P} P := \bigg \{ {\mathrm{m}}_{x,y} = \begin{pmatrix} \sqrt{y} & x / \sqrt{y} \\ 0 & 1/ \sqrt{y} \end{pmatrix} : (x, y) \in \mathbb{R} \times \mathbb{R}^+ \bigg\} \subset \mathrm{PSL}(2,\mathbb{R})\end{aligned}$$ provides representatives for the quotient $G / \mathrm{PSO}(2, \mathbb{R})$, since $G = P \cdot \mathrm{PSO}(2, \mathbb{R})$ and $P \cap \mathrm{PSO}(2, \mathbb{R}) = \{I\}$. In particular, every ${\mathrm{m}}\in G$ can be written as ${\mathrm{m}}= {\mathrm{m}}_{x,y} r$ for a unique ${\mathrm{m}}_{x,y} \in P$ and some $r \in \mathrm{PSO}(2, \mathbb{R})$. Recall that $i \in \mathbb{C}^+$ is a fixed point of $\mathrm{PSO}(2, \mathbb{R})$, and, hence, $(x,y)$ is $x+iy = {\mathrm{m}}_{x,y} \cdot i ={\mathrm{m}}\cdot i$. Therefore, the coherent state associated with ${h^{(\alpha)}_n}$ can be realized as an *affine system*: $$\begin{aligned} \label{eq_per} {\mathcal{A}}_\alpha({h^{(\alpha)}_n},\Gamma) = \left\{ \pi_\alpha({\mathrm{m}}_{x,y}) {h^{(\alpha)}_n}: x+iy \in \Gamma i \right\} = \left\{ y^{-\alpha/2} {h^{(\alpha)}_n}\big(\tfrac{\cdot-x}{y}\big) : x+iy \in \Gamma i \right\}.\end{aligned}$$ Perelomov’s problem concerns the completeness of ${\mathcal{A}}_\alpha({h^{(\alpha)}_n},\Gamma)$ in $A^2_{\alpha} (\mathbb{C}^+)$. While Theorem \[th\_semi\] shows that is necessary for completeness, we are unaware of literature on corresponding sufficient conditions. The completeness problem can be alternatively reformulated on the real half-line. The connection is provided by the *Paley-Wiener theorem for Bergman spaces* [@duren2007paley; @sally1967analytic]: the Fourier-Laplace transform $$\begin{aligned} \mathcal{F}f (z) = \int_0^\infty f(t) e^{i z t} \, dt, \qquad z \in \mathbb{C}^+,\end{aligned}$$ is a multiple of an isometric isomorphism between the weighted $L^2$-space $$\begin{aligned} L^2(\mathbb{R}^+, t^{-(\alpha-1)}\,dt) = \left\{ f:\mathbb{R}^+ \to \mathbb{C} \mbox{ measurable} : \int_{\mathbb{R}^+} |f(t)|^2 t^{-(\alpha-1)} \, dt< \infty \right\}\end{aligned}$$ and the Bergman space $A^2_{\alpha} (\mathbb{C}^+)$. In addition, the special vectors ${h^{(\alpha)}_n}$ correspond under the isomorphism to multiples of $$\begin{aligned} \label{eq:Fourier} H_n^{(\alpha)} (t) := t^{\alpha - 1} e^{-t} L^{(\alpha - 1)}_n (2t), \qquad t >0,\end{aligned}$$ where $L_n^{\alpha-1}$ is the Laguerre polynomial of degree $n \in \mathbb{N}$ and index $\alpha-1$; see [@duren2007paley]. The inverse Fourier-Laplace transform thus maps the affine system into the system $$\begin{aligned} \label{eq_abc} \mathcal{F}^{-1} {\mathcal{A}}_\alpha({h^{(\alpha)}_n},\Gamma) =\left\{ d^\alpha_n \,y^{-\alpha/2+1} e^{-i \pi x \,\cdot} H^{(\alpha)}_n (y \cdot) \right\},\end{aligned}$$ in $L^2 (\mathbb{R}^+, t^{-(\alpha -1)} dt)$ for a suitable $d^\alpha_n \in \mathbb{C}$. This yields another equivalent formulation of Perelomov’s completeness problem. See also [@combescure2012coherent Section 8.6]. With a certain physical motivation, part of Perelomov’s work [@perelomov1973coherent] has been adapted to the special vectors $H_n^{(\alpha)}$ by Abreu, Balazs, de Gosson and Mouayn [@ABDM]. Conditionally to the existence of modular forms having certain special properties, and under certain restrictions on $\alpha > 1$, [@ABDM Corollary 1] asserts that if is complete in $L^2(\mathbb{R}^+, t^{-(\alpha-1)}\,dt)$, then $$\begin{aligned} \label{eq_abdm} \operatorname*{vol}(G/\Gamma) \leq (n+1) \frac{8\pi}{(\alpha-1)}.\end{aligned}$$ On the other hand, Theorem \[th\_semi\] provides the sharper bound $$\begin{aligned} \operatorname*{vol}(G/\Gamma) \leq \frac{4\pi}{(\alpha-1)},\end{aligned}$$ which is valid without assumptions on the existence of adequate modular forms, for all $\alpha>1$, and is also applicable to any vector in lieu of ${h^{(\alpha)}_n}$. [^6] Heisenberg projective representation and Gabor systems {#sec:gabor} ------------------------------------------------------ Let $G = \mathbb{R}^{2d}$. Define the projective representation $(\pi, L^2 (\mathbb{R}^d))$ through the action $$\begin{aligned} \label{eq:timefreq} \pi (z) f (t) = e^{2\pi i y \cdot t} f(t - x), \quad z = (x,\xi) \in \mathbb{R}^{2d}, \; t \in \mathbb{R}^d. \end{aligned}$$ Then $ \pi(z + z') = e^{2\pi i \xi' \cdot x } \pi(z) \pi(z')$ for $z = (x, \xi) \in \mathbb{R}^{2d}$ and $z' = (x', \xi') \in \mathbb{R}^{2d}$. Thus the cocycle of $(\pi, L^2 (\mathbb{R}^d))$ is $\sigma (z, z') = e^{2\pi i \xi' \cdot x } \in \mathbb{T}$. Moreover, $\pi$ is is irreducible and square-integrable of formal dimension $d_{\pi} = 1$. For background, and the appearance of the Heisenberg group, see [@folland1989harmonic; @groechenig2001foundations]. Systems of the form $\pi(\Gamma) g$, with $g \in L^2 (\mathbb{R}^d)$ and $\Gamma \subset \mathbb{R}^{2d}$ a lattice, are known as *Gabor systems* or *Weyl-Heisenberg systems*, and are important in several branches of pure and applied mathematics. Gabor systems are sometimes also called *canonical* coherent state subsystems in mathematical physics. The literature on Gabor systems focuses mainly on frames, Riesz sequences, and completeness. Kleppner’s condition for a lattice $\Gamma \subseteq \mathbb{R}^{2d}$ and the cocycle $\sigma$ reads: for all $\gamma \in \Gamma \setminus \{0\}$ there exists $\gamma' \in \Gamma$ such that $$\sigma(\gamma, \gamma') \overline{\sigma(\gamma, \gamma')} = e^{2\pi i (\gamma_2' \cdot \gamma_1 - \gamma_2 \cdot \gamma_1')} \neq 1.$$ While for separable lattices $\Gamma = \alpha \mathbb{Z}^d \times \beta \mathbb{Z}^d$, with $\alpha,\beta \in \mathbb{R}$, Kleppner’s condition reduces to $\alpha \beta \not\in \mathbb{Q}$, an explicit characterization of Kleppner’s condition for more general lattices is subtle, e.g., see [@omland2014primeness; @packer1989twisted; @han2017note]. Provided that $(\Gamma, \sigma)$ satisfies Kleppner’s condition, Theorem \[th\_main\] shows that $\pi(\Gamma)$ admits a frame vector if and only if it admits a complete vector, if and only if $$\begin{aligned} \label{eq:density_gabor} \operatorname*{vol}(G/\Gamma) \leq 1;\end{aligned}$$ while the condition for the existence a Riesz vector is $$\begin{aligned} \label{eq:density_gabor_2} \operatorname*{vol}(G/\Gamma) \geq 1.\end{aligned}$$ In fact, Theorem \[thm:necessary\_density\] shows that the necessity of the density conditions for completeness, frames, and Riesz sequences holds without assuming Kleppner’s condition. Direct proofs of this necessity go back to Baggett [@baggett1990processing], Daubechies, Landau and Landau [@daubechies1995gabor], and Ramanathan and Steger [@ramanathan1995incompletness], and are also implicitly contained in Rieffel’s work [@rieffel1981von; @rieffel1988projective]. Our proof of Theorem \[thm:necessary\_density\] is partially inspired by Janssen’s “classroom proof” [@janssen_density]. Much deeper is the fact that the sufficiency of and for the existence of frames and Riesz vectors also holds without assuming Kleppner’s condition. The stronger claim, shown by Rieffel [@rieffel1981von; @rieffel1988projective], and also a consequence of Bekka’s work [@bekka2004square Theorem 4], lies beyond the elementary approach presented in this article. Indeed, Rieffel’s and Bekka’s work require considering not only the operator algebras $\pi(\Gamma)'$ and $\pi(\Gamma)''$, but also certain so-called induced algebras, and in this way fully exploit the coupling theory of von Neumman algebras. We hope that our elementary introduction motivates the reader to delve deeper into operator-algebraic methods. For lattices of the form $\Gamma = A \mathbb{Z}^d \times B \mathbb{Z}^d$, with $A, B \in \mathrm{GL}(d, \mathbb{R})$, Han and Wang gave a constructive proof of the sufficiency of for the existence of frame vectors [@han2001lattice]. ### Gaussians and Bargmann-Fock spaces The question of choosing specific cyclic or frame vectors has been intensively studied for $d=1$ and lattices in $\mathbb{R}^2$ of the form $\Gamma = \alpha \mathbb{Z} \times \beta \mathbb{Z}$. In his work on foundations of quantum mechanics, von Neumann [@neumann1968mathematische] claimed without proof that the Gabor system $\pi(\Gamma)g$ generated by the Gaussian function $$\begin{aligned} \label{eq_gauss} g(t) = 2^{-1/4} e^{-\pi |t|^2}, \qquad t \in \mathbb{R},\end{aligned}$$ is complete in $L^2(\mathbb{R})$ if and only if holds. Proofs of the claim were given by Perelomov [@perelomov1971remark], Bargmann [@bargmann1971on], and Neretin [@neretin2006perelomov]. For rational lattices (i.e., $\alpha \beta \in \mathbb{Q}$), the same claim holds when the Gaussian function is multiplied by a rational function with no real poles [@groechenig2016completeness]. The related question, under which conditions the Gabor system generated by the Gaussian is a frame for $L^2 (\mathbb{R})$ or a Riesz sequence was first considered by Daubechies and Grossmann [@daubechies1988frames], and fully answered independently by Lyubarskiĭ [@lyubarski1992frames], and Seip and Wallstén [@seip1992density; @seip1992density2]: $$\begin{aligned} \operatorname*{vol}(G/\Gamma) < 1,\end{aligned}$$ is necessary and sufficient for the frame property, while $$\begin{aligned} \operatorname*{vol}(G/\Gamma) > 1,\end{aligned}$$ is necessary and sufficient for the Riesz property. The proofs of Lyubarskiĭ [@lyubarski1992frames] and Seip-Wallstén [@seip1992density] work with a $\sigma$-representation unitarily equivalent to $(\pi, L^2 (\mathbb{R}))$ on the Bargmann-Fock space $\mathcal{F}^2(\mathbb{C})$ of entire functions $F:\mathbb{C} \to \mathbb{C}$ having finite norm $$\| F \|^2_{\mathcal{F}^2} = \int_{\mathbb{C}} | F(z) |^2 e^{- \pi |z|^2} \; dxdy.$$ As in Example \[sec\_bergman\], the distinguished vector $g$ corresponds under the new representation to the reproducing kernel, that is, the vector representing the evaluation functional $F \mapsto F(0)$. A simple proof of the density results was derived by Janssen [@janssen1994signal]. The characterization of the frame and Riesz property for other vectors $g$ is a topic of intense study [@groechenig2014mystery]. [10]{} L. D. Abreu, P. Balazs, M. de Gosson, and Z. Mouayn. Discrete coherent states for higher [L]{}andau levels. , 363:337–353, 2015. L. Alaoglu and G. Birkhoff. General ergodic theorems. , 41:293–309, 1940. M. Atiyah and W. Schmid. A geometric construction of the discrete series for semisimple [L]{}ie groups. , 42:1–62, 1977. M. F. Atiyah. Elliptic operators, discrete groups and von [N]{}eumann algebras. In [*Colloque “[A]{}nalyse et [T]{}opologie” en l’[H]{}onneur de [H]{}enri [C]{}artan ([O]{}rsay, 1974)*]{}, pages 43–72. Astérisque, No. 32–33. 1976. L. W. Baggett. Processing a radar signal and representations of the discrete [H]{}eisenberg group. , 60/61(1):195–203, 1990. R. Balan, P. G. Casazza, C. Heil, and Z. Landau. Density, overcompleteness, and localization of frames. [I]{}. [T]{}heory. , 12(2):105–143, 2006. D. Barbieri, E. Hernández, and J. Parcet. Riesz and frame systems generated by unitary actions of discrete groups. , 39(3):369–399, 2015. V. Bargmann, P. Butera, L. Girardello, and J. R. Klauder. On the completeness of the coherent states. , 2(4):221–228, 1971. A. F. Beardon. , volume 91 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 1983. B. Bekka. Square integrable representations, von [N]{}eumann algebras and an application to [G]{}abor analysis. , 10(4):325–349, 2004. B. Bekka, P. de la Harpe, and A. Valette. , volume 11 of [*New Mathematical Monographs*]{}. Cambridge University Press, Cambridge, 2008. M. B. Bekka. Square integrable representations, lattices and von [N]{}eumann algebras. In [*Lie theory and its applications in physics, [III]{} ([C]{}lausthal, 1999)*]{}, pages 27–40. World Sci. Publ., River Edge, NJ, 2000. J. Bertrand and M. Irac-Astaud. Characterization of [${\rm SU}(1,1)$]{} coherent states in terms of affine group wavelets. , 35(34):7347–7357, 2002. O. Christensen. Frames and pseudo-inverses. , 195(2):401–414, 1995. O. Christensen. . Applied and Numerical Harmonic Analysis. Birkhäuser/Springer, \[Cham\], second edition, 2016. M. Combescure and D. Robert. . Theoretical and Mathematical Physics. Springer, Dordrecht, 2012. M. Cowling and T. Steger. The irreducibility of restrictions of unitary representations to lattices. , 420:85–98, 1991. I. Daubechies and A. Grossmann. Frames in the [B]{}argmann space of entire functions. , 41(2):151–164, 1988. I. Daubechies, A. Grossmann, and Y. Meyer. Painless nonorthogonal expansions. , 27(5):1271–1283, 1986. I. Daubechies, H. J. Landau, and Z. Landau. Gabor time-frequency lattices and the [W]{}exler-[R]{}az identity. , 1(4):437–478, 1995. J. Dixmier. , volume 27 of [*North-Holland Mathematical Library*]{}. North-Holland Publishing Co., Amsterdam-New York, 1981. With a preface by E. C. Lance, Translated from the second French edition by F. Jellett. P. Duren, E. A. Gallardo-Gutiérrez, and A. Montes-Rodríguez. A [P]{}aley-[W]{}iener theorem for [B]{}ergman spaces with application to invariant subspaces. , 39(3):459–466, 2007. J. M. G. Fell and R. S. Doran. , volume 125 of [*Pure and Applied Mathematics*]{}. Academic Press, Inc., Boston, MA, 1988. Basic representation theory of groups and algebras. G. B. Folland. , volume 122 of [*Annals of Mathematics Studies*]{}. Princeton University Press, Princeton, NJ, 1989. H. Führ. Admissible vectors for the regular representation. , 130(10):2959–2970, 2002. H. Führ. , volume 1863 of [*Lecture Notes in Mathematics*]{}. Springer-Verlag, Berlin, 2005. H. Führ, K. Gröchenig, A. Haimi, A. Klotz, and J. L. Romero. Density of sampling and interpolation in reproducing kernel [H]{}ilbert spaces. , 96(3):663–686, 2017. W. Fulton and J. Harris. , volume 129 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 1991. A first course, Readings in Mathematics. J.-P. Gabardo and D. Han. Frame representations for group-like unitary operator systems. , 49(2):223–244, 2003. F. M. Goodman, P. de la Harpe, and V. F. R. Jones. , volume 14 of [ *Mathematical Sciences Research Institute Publications*]{}. Springer-Verlag, New York, 1989. K. Gröchenig. . Applied and Numerical Harmonic Analysis. Birkhäuser Boston, Inc., Boston, MA, 2001. K. Gröchenig. The mystery of [G]{}abor frames. , 20(4):865–895, 2014. K. Gröchenig, A. Haimi, and J. L. Romero. Completeness of [G]{}abor systems. , 207:283–300, 2016. D. Han. A note on the density theorem for projective unitary representations. , 145(4):1739–1745, 2017. D. Han and Y. Wang. Lattice tiling and the [W]{}eyl-[H]{}eisenberg frames. , 11(4):742–758, 2001. C. Heil. History and evolution of the density theorem for [G]{}abor frames. , 13(2):113–166, 2007. M. S. Jakobsen and J. Lemvig. Density and duality theorems for regular [G]{}abor frames. , 270(1):229–263, 2016. A. J. E. M. Janssen. Signal analytic proofs of two basic results on lattice expansions. , 1(4):350–354, 1994. A. J. E. M. Janssen. Classroom proof of the density theorem for [G]{}abor systems. , 2005. R. V. Kadison and J. R. Ringrose. , volume 100 of [*Pure and Applied Mathematics*]{}. Academic Press, Inc. \[Harcourt Brace Jovanovich, Publishers\], New York, 1983. Elementary theory. D. Kelly-Lyth. Uniform lattice point estimates for co-finite [F]{}uchsian groups. , 78(1):29–51, 1999. J. R. Klauder and R. F. Streater. Wavelets and the [P]{}oincaré half-plane. , 35(1):471–478, 1994. A. Kleppner. The structure of some induced representations. , 29:555–572, 1962. G. Kuhn and T. Steger. Restrictions of the special representation of [${\rm Aut}({\rm tree}_3)$]{} to two cocompact subgroups. , 22(4):1349–1363, 1992. Y. I. Lyubarskiĭ. Frames in the [B]{}argmann space of entire functions. In [*Entire and subharmonic functions*]{}, volume 11 of [*Adv. Soviet Math.*]{}, pages 167–180. Amer. Math. Soc., Providence, RI, 1992. G. W. Mackey. Induced representations of locally compact groups. [I]{}. , 55:101–139, 1952. G. W. Mackey. Unitary representations of group extensions. [I]{}. , 99:265–311, 1958. M. Mitkovsi and A. Ramirez. Density results for continuous frames. . M. I. Monastyrsky and A. M. Perelomov. Coherent states and bounded homogeneous domains. , 6:1–14, 1974. H. Moscovici. Coherent state representations of nilpotent [L]{}ie groups. , 54(1):63–68, 1977. H. Moscovici and A. Verona. Coherent states and square integrable representations. , 29(2):139–156, 1978. K.-H. Neeb. , volume 28 of [*De Gruyter Expositions in Mathematics*]{}. Walter de Gruyter & Co., Berlin, 2000. Y. A. Neretin. The [P]{}erelomov problem of the inversion of the [B]{}argmann-[S]{}egal transform. , 40(4):104–107, 2006. T. A. n. Omland. Primeness and primitivity conditions for twisted group [$C^*$]{}-algebras. , 114(2):299–319, 2014. J. A. Packer. Twisted group [$C^*$]{}-algebras corresponding to nilpotent discrete groups. , 64(1):109–122, 1989. A. Perelomov. . Texts and Monographs in Physics. Springer-Verlag, Berlin, 1986. A. M. Perelomov. Remark on the completeness of the coherent state system. , 6(2):213–224, 1971. A. M. Perelomov. Coherent states for arbitrary [L]{}ie groups. , 26:222–236, 1972. A. M. Perelomov. Coherent states for the [L]{}obačevskiĭ plane. , 7(3):57–66, 1973. M. S. Raghunathan. . Springer-Verlag, New York-Heidelberg, 1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 68. J. Ramanathan and T. Steger. Incompleteness of sparse coherent states. , 2(2):148–153, 1995. M. A. Rieffel. Square-integrable representations of [H]{}ilbert algebras. , 3:265–300, 1969. M. A. Rieffel. von [N]{}eumann algebras associated with pairs of lattices in [L]{}ie groups. , 257(4):403–418, 1981. M. A. Rieffel. Projective modules over higher-dimensional noncommutative tori. , 40(2):257–338, 1988. M. A. Rieffel. Integrable and proper actions on [$C^*$]{}-algebras, and square-integrable representations of groups. , 22(1):1–53, 2004. A. Robert. , volume 80 of [*London Mathematical Society Lecture Note Series*]{}. Cambridge University Press, Cambridge-New York, 1983. F. Rădulescu. The [$\Gamma$]{}-equivariant form of the [B]{}erezin quantization of the upper half plane. , 133(630):viii+70, 1998. P. J. Sally, Jr. . Memoirs of the American Mathematical Society, No. 69. American Mathematical Society, Providence, R. I., 1967. K. Seip. Density theorems for sampling and interpolation in the [B]{}argmann-[F]{}ock space. [I]{}. , 429:91–106, 1992. K. Seip. Beurling type density theorems in the unit disk. , 113(1):21–39, 1993. K. Seip and R. Wallstén. Density theorems for sampling and interpolation in the [B]{}argmann-[F]{}ock space. [II]{}. , 429:107–113, 1992. V. S. Varadarajan. , volume 102 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 1984. Reprint of the 1974 edition. V. S. Varadarajan. . Springer-Verlag, New York, second edition, 1985. J. von Neumann. . Unveränderter Nachdruck der ersten Auflage von 1932. Die Grundlehren der mathematischen Wissenschaften, Band 38. Springer-Verlag, Berlin-New York, 1968. R. M. Young. . Academic Press, Inc., San Diego, CA, first edition, 2001. [^1]: Perelomov uses the term coherent state with a slightly different meaning, as systems are not the full orbit of a group representation, but parametrized by a homogeneous space to eliminate redundancies. See [@moscovivi1977coherent; @moscovici1978coherent] for the relation between the two notions. [^2]: In part (i) of Theorem \[thm:intro\_ICC\], the assertion that $\pi|_{\Gamma}$ does not admit a separating vector means that $\pi(\Gamma)''$ does not admit such a vector. [^3]: In the jargon of semisimple Lie groups, a group all of whose connected, normal, compact subgroups are trivial is sometimes referred to as *a group without compact factors*. [^4]: Perelomov formulates his results on the unit disk. [^5]: In [@perelomov1973coherent Theorems 3 and 4] Perelomov implicitly assumes that $\#F_w=1$, the general case follows after some minor adaptations, as explained in [@kelly Theorem 5.1]. The case $\Gamma=\mathrm{PSL}(2,\mathbb{Z})$ is proved independently in [@klauder1994wavelets], after observing that the physically-motivated restrictions the authors impose on $\alpha$ play no role in the argument. [^6]: The bound stated in [@ABDM Corollary 1] is with $\alpha$ instead of $\alpha-1$. We understand this as a miscalculation caused by inconsistent normalization of the Bergman space on [@ABDM page 352]. The result in [@ABDM] is (equivalently) formulated in terms of the completeness of the system of functions $(yt)^{-\alpha/2+1} e^{\pi x i \,t/2} H^{(\alpha)}_n (yt/2)$ within $L^2(\mathbb{R}^+, t^{-1}\,dt)$.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that if $K$ is a knot in $S^3$ and $\Sigma$ is a bridge sphere for $K$ with high distance and $2n$ punctures, the number of perturbations of $K$ required to interchange the two balls bounded by $\Sigma$ via an isotopy is $n$. We also construct a knot with two different bridge spheres with $2n$ and $2n-1$ bridges respectively for which any common perturbation has at least $3n-1$ bridges. We generalize both of these results to bridge surfaces for knots in any 3-manifold.' address: - | -Jesse Johnson\ Mathematics Department\ Oklahoma State University. - | - Maggy Tomova\ Mathematics Department\ University of Iowa. author: - Jesse Johnson - Maggy Tomova title: Flipping bridge surfaces and bounds on the stable bridge number --- [^1] Introduction ============ Reidemeister [@Re] and Singer [@Sin] showed that any two Heegaard splittings for a 3-manifold $M$ have a common stabilization, i.e., if ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ are two Heegaard surfaces for $M$ there exists a Heegaard surface ${\mbox{$\Sigma$}}''$ that is isotopic to a stabilization of ${\mbox{$\Sigma$}}$ as well as to a stabilization of ${\mbox{$\Sigma$}}'$. A long standing question in Heegaard splittings asks what is the minimal genus of ${\mbox{$\Sigma$}}''$ in terms of the genera of ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$. Examples of Heegaard splittings that required many stabilizations were presented in [@DB], [@JJ2] and [@HTT]. Bridge splittings are the natural extension of Heegaard splittings in the context of a compact orientable manifold $M$ containing a properly embedded tangle $T$. A *bridge splitting* for $(M,T)$ is a triple $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ where ${\mbox{$\Sigma$}}$ is a connected surface that decomposes $M$ into compression bodies $H^+$ and $H^-$ and decomposes $T$ into collections of arcs $\tau^+$ and $\tau^-$ that are embedded in the corresponding compression bodies in specific ways. The surface ${\mbox{$\Sigma$}}$ is called a *bridge surface* for $(M, T)$. Note that if $T=\emptyset$, then $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ is a Heegaard splitting for $M$. Given a bridge surface ${\mbox{$\Sigma$}}$ of $(M,T)$ one can always obtain another bridge surface ${\mbox{$\Sigma$}}''$ by performing stabilizations and perturbations. These operations are discussed in detail in [@STo3] and they behave in a manner similar to stabilizations of Heegaard splittings. In this paper we consider pairs of bridge splittings ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ for $(M,T)$ and study bridge splittings ${\mbox{$\Sigma$}}''$ that can be obtained from both ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ via stabilizations and perturbations. The results we obtain are similar but somewhat weaker than the results obtained by Johnson for Heegaard splittings in [@JJ] and [@JJ2] due to the additional difficulties introduced by the presence of the knot. At first we will distinguish a bridge splitting $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ from the bridge splitting $({\mbox{$\Sigma$}}, (H^-, \tau^-), (H^+, \tau^+))$ in which the order of the compression bodies is reversed. We ask what is the minimum value of $2-\chi({\mbox{$\Sigma$}}'')$ such that $({\mbox{$\Sigma$}}'', (H''^+, \tau''^+), (H''^-, \tau''^-))$ is isotopic to stabilizations and perturbations of both bridge splittings $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ and $({\mbox{$\Sigma$}}, (H^-, \tau^-), (H^+, \tau^+))$. This value is called the [*flip Euler characteristic of*]{} ${\mbox{$\Sigma$}}$ and it is analogous to the flip genus of a Heegaard splitting defined in [@JJ]. We give a bound on this quantity in terms of the Euler characteristic of ${\mbox{$\Sigma$}}$ and the distance of $T$ with respect to ${\mbox{$\Sigma$}}$ (Definition \[def:distance\]). \[thm:flip\] Let $T$ be a prime tangle properly embedded in a compact orientable irreducible 3-manifold $M$ and let $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ be a bridge splitting for $(M,T)$ such that $\chi({\mbox{$\Sigma$}})\leq -4$. Then the flip Euler characteristic of ${\mbox{$\Sigma$}}$ is at least $max\{2-2\chi({\mbox{$\Sigma$}}), d({\mbox{$\Sigma$}},T)\}$. \[cor:flip\] Let $T$ be a prime knot in $S^3$ and let ${\mbox{$\Sigma$}}$ be a bridge sphere for $T$ with $n\geq 3$ bridges such that $d(T, {\mbox{$\Sigma$}}) \geq 4n$. If $({\mbox{$\Sigma$}}'', (H''^+, \tau''^+), (H''^-, \tau''^-))$ is a minimal bridge number perturbation of both bridge splittings $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ and $({\mbox{$\Sigma$}}, (H^-, \tau^-), (H^+, \tau^+))$, then $T$ has exactly $2n$ bridges with respect to ${\mbox{$\Sigma$}}''$. We next consider the problem of distinguishing bridge surfaces without keeping track of the order of compression bodies. To make this clear, we will consider only the bridge surface rather than the bridge splitting. In this case we obtain the following result. \[thm:knot\] There exist infinitely many manifolds $M_{\alpha}$ each containing a knot $K_{\alpha}$ so that each pair $(M_{\alpha}, K_{\alpha})$ has two bridge surfaces ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ with $\chi({\mbox{$\Sigma$}})=2s$ and $\chi({\mbox{$\Sigma$}}')=2s-2$ so that for every bridge surface ${\mbox{$\Sigma$}}''$ that is isotopic to stabilizations and perturbations of both ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$, $\chi({\mbox{$\Sigma$}}')\leq 3s+2$. As a corollary of the above we obtained the following result: \[cor:secondmain\] For every $n \geq 2$ there exists a knot $\tilde K$ in $S^3$ with bridge spheres $\Sigma$ and $\Sigma'$ with bridge numbers $2n-1$ and $2n$ respectively such that every bridge surface $\Sigma''$ which is isotopic to a perturbation of both has at least $3n-1$ bridges. In Section \[sec:prel\] we give the definition of a bridge splitting for a pair $(M,T)$ and explain how a sweep-out is associated to any bridge splitting. Furthermore we define two conditions on a pair of sweep-outs: A sweep-out $g$ can split a second sweep-out $f$ for the same manifold or can span it. Generically these are the only two options for how $g$ behaves with respect to $f$. In Sections \[sec:span\] and \[sec:split\] we consider two bridge splittings ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ for $(M,T)$ with associated sweep-outs $f$ and $g$. We show that if $g$ spans $f$, then the Euler characteristic of the punctured bridge surface ${\mbox{$\Sigma$}}$ is bounded below by the Euler characteristic of the punctured bridge surface ${\mbox{$\Sigma$}}'$. Next we define the distance of a bridge splitting and we show that if $g$ splits $f$ then the distance of ${\mbox{$\Sigma$}}$ is bounded above by the Euler characteristic of ${\mbox{$\Sigma$}}'$. Finally we consider the case where $g$ neither spans not splits $f$ and we show that this can only occur if $\chi({\mbox{$\Sigma$}})\geq -3$. Using these results in Section \[sec:flipping\] we prove Theorem \[thm:flip\] and in Section \[sec:example\] we prove Theorem \[thm:knot\]. Preliminaries {#sec:prel} ============= Compression bodies containing trivial arcs ------------------------------------------ Let $H$ be a compression body. Recall that a spine of $H$ is a complex ${\mbox{$\partial$}}_- H \cup \Gamma$ where $\Gamma \subset H$ is a properly embedded finite graph with no valence $1$ vertices in the interior of $H$ and such that $H$ is isotopic to a regular neighborhood of ${\mbox{$\partial$}}_- H \cup \Gamma$. A set of properly embedded arcs $\tau=\{t_1,..., t_n\}$ in $H$ is [*trivial*]{} if each $t_i$ is either parallel to ${\mbox{$\partial$}}_+H$ or is a vertical arc with one endpoint in ${\mbox{$\partial$}}_+H$ and the other endpoint in ${\mbox{$\partial$}}_-H$. If an arc is parallel to ${\mbox{$\partial$}}_+H$ the disk of parallelism is called a [*bridge disk*]{}. We will denote the pair of a compression body $H$ containing properly embedded trivial arcs $\tau$ by $(H, \tau)$. The arcs $\tau$ can be isotoped in $H$ so that the projection $H - spine(H) \cong {\mbox{$\partial$}}H \times [0, 1) \to [0, 1)$ has no critical points in the vertical arcs and a single critical point, say a maximum, in each boundary parallel arc. Let $s_i$ be a collection of vertical arcs each connecting a single maximum of $\tau$ to a spine of $H$. Let $spine((H, \tau))=spine(H) \cup \{s_i\}$ and note that there is a map $({\mbox{$\partial$}}H, {\mbox{$\partial$}}H \cap \tau) \times I \to (H, \tau)$ which is a homeomorphism except over the spine, and the map gives a neighborhood of the spine a mapping cylinder structure. Bridge splittings ----------------- Let $T$ be a properly embedded tangle in a compact oriented $3$-manifold $M$ and let ${\mbox{$\Sigma$}}$ be a properly embedded surface transverse to $T$ such that ${\mbox{$\Sigma$}}$ splits $M$ into two compression bodies $H^+$ and $H^-$ and such that $\tau^+=H^+\cap T$ and $\tau^-=H^-\cap T$ are trivial arcs in the corresponding compression body. In this case we say that $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ is a bridge splitting for $(M,T)$ and ${\mbox{$\Sigma$}}$ is a bridge surface. As every compact orientable 3-manifold has a Heegaard splitting it is easy to see that every properly embedded tangle in any 3-manifold has a bridge splitting. Surfaces in $(M,T)$ ------------------- Suppose $M$ is a compact, irreducible, orientable 3-manifold containing a properly embedded tangle $T$ and let $F$ be a surface in $M$ transverse to $T$. The surface $F$ gives rise to a punctured surface in the complement of a regular neighborhood $\eta(T)$ of $T$. We will refer to this punctured surface as $F$ also and we will specify if we are referring to the punctured or the closed surface whenever it is not clear from context. Two surfaces in $(M,T)$ will be considered isotopic only if there is an isotopy between them transverse to the tangle. A simple closed curve in $F-\eta(T)$ is [*essential*]{} if it does not bound a disk in $F$ and it is not parallel to the boundary of a puncture. A properly embedded arc in $F$ with endpoints in $F \cap {\mbox{$\partial$}}M$ is essential if it does not cobound a disk with an arc in $F \cap {\mbox{$\partial$}}M$. An embedded disk $D$ in $M$ is a [*compressing disk*]{} for $F$ if $D \cap T =\emptyset$, $D \cap F = {\mbox{$\partial$}}D$ and ${\mbox{$\partial$}}D$ is an essential curve in $F-\eta(T)$. A properly embedded disk $D^c$ in $M$ is a [*cut-disk*]{} for $F$ if $D^c \cap T$ is a single point in the interior of $D^c$, $D^c \cap F = {\mbox{$\partial$}}D^c$ and ${\mbox{$\partial$}}D^c$ is an essential curve in $F-\eta(T)$. A [*c-disk*]{} is either a cut or a compressing disk. Obtaining new bridge splittings from known ones ----------------------------------------------- We will consider two geometric operations which allow us to produce new bridge surfaces from existing ones. These are generalizations of stabilizations for Heegaard splittings. Following [@HS1], the bridge surface $\Sigma$ will be called [*stabilized*]{} if there is a pair of compressing disks on opposite sides of $\Sigma$ that intersect in a single point. The bridge surface is called [*perturbed*]{} if there is a pair of bridge disks $D_i$ on opposite sides of $\Sigma$ such that $ \emptyset \neq ({\mbox{$\partial$}}D_1 \cap {\mbox{$\partial$}}D_2) \subset (\Sigma \cap T)$ and $|{\mbox{$\partial$}}D_1 \cap {\mbox{$\partial$}}D_2|=1$. These operations are discussed in detail in [@STo3]. Sweep-outs ---------- Suppose $(M,T)=(H^+, \tau^+)\cup_{{\mbox{$\Sigma$}}} (H^-, \tau^-)$. From the definition of a spine one can construct a map $f:M \to [-1,1]$ such that $f^{-1}(1)$ is isotopic to a spine of $(H^+,\tau^+)$, $f^{-1}(-1)$ is isotopic to a spine of $(H^-, \tau^-)$ and $f^{-1}(t)$ is a surface isotopic to the punctured surface ${\mbox{$\Sigma$}}$ for every $t \in (-1,1)$. This function is called a *sweep-out* representing $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$. We give a brief overview of how sweep-outs can be applied to study bridge surfaces for tangles in a 3-manifold. Further details can be found in [@T2]. Consider a tangle properly embedded in a 3-manifold with two bridge splittings. Let $f$ be a sweep-out representing the bridge splitting $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$ and let $g$ be another sweep-out representing a second bridge splitting for $(M,T)$ which we denote $({\mbox{$\Sigma$}}',(H'^-, \tau'^-), (H'^+, \tau'^+))$. Consider the two parameter sweep-out $f \times g$ mapping $(M,T)$ into the square $[-1,1] \times [-1,1]$. Each point $(s,t)$ in the square represents a pair of surfaces ${\mbox{$\Sigma$}}_t = f^{-1}(t)-\eta(T)$ isotopic to the punctured surface ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'_s = g^{-1}(s)-\eta(T)$ isotopic to ${\mbox{$\Sigma$}}'$. The [*graphic*]{} is the subset ${\mbox{$\Gamma$}}$ of the square consisting of all points $(s,t)$ where either $\Sigma_t$ is tangent to $\Sigma'_s$ or $\Sigma_t \cap \Sigma'_s$ contains a point of $T$. We say that $f \times g$ is [*generic*]{} if it is stable on the complement of the spines and each arc $\{t\} \times [-1,1]$ and $[-1,1] \times \{s\}$ contains at most one vertex of the graphic. If $f\times g$ is generic then at each (valence four) vertex of ${\mbox{$\Gamma$}}$ there are two points of tangency, two points of $T$ in the intersection, or one of each. By general position of the spines $f^{-1}(\pm 1)$ with the surface ${\mbox{$\Sigma$}}'$, the graphic ${\mbox{$\Gamma$}}$ is incident to ${\mbox{$\partial$}}I \times I$ in only a finite number of points corresponding to tangencies between $f^{-1}(\pm 1)$ and ${\mbox{$\Sigma$}}'$. Splitting and spanning sweep-outs --------------------------------- Suppose $f$ and $g$ are sweep-outs for $(M,T)$ and $f \times g$ is generic. Generalizing [@JJ], for some fixed values of $s$ and $t$ we will say that ${\mbox{$\Sigma$}}_t$ is *mostly above* ${\mbox{$\Sigma$}}'_s$ if each component of ${\mbox{$\Sigma$}}_t \cap H'^-_s$ (if there are any) is contained in a disk or a once-punctured disk in ${\mbox{$\Sigma$}}_t$. Similarly we will say that ${\mbox{$\Sigma$}}_t$ is *mostly below* ${\mbox{$\Sigma$}}'_s$ if each component of ${\mbox{$\Sigma$}}_t \cap H'^+_s$ is contained in a disk or once-punctured disk in ${\mbox{$\Sigma$}}_t$. We will say that $g$ *spans* $f$ if there are values $t_+$, $t_-$ and $s$ for which ${\mbox{$\Sigma$}}_{t_+}$ is mostly above ${\mbox{$\Sigma$}}'_s$ and ${\mbox{$\Sigma$}}_{t_-}$ is mostly below ${\mbox{$\Sigma$}}'_s$. We will say that $g$ spans $f$ *positively* if $t_- <t_+$ and *negatively* otherwise. These conditions are shown at the top of Figure \[fig:splitspan\]. Note that $g$ may span $f$ both positively and negatively. We will say that $g$ *splits* $f$ if there is a value of $s$ such that the horizontal line $[-1,1] \times \{s\} \subset [-1,1]\times[-1,1]$ does not intersect any vertices of $\Gamma$ and for every $t$ the surface $\Sigma_t$ is neither mostly above nor mostly below $\Sigma'_s$. This is shown at the bottom left of Figure \[fig:splitspan\]. Note that this condition is equivalent to the condition that there exists an $s$ such that for every $t$, $\Sigma'_s \cap \Sigma_t$ contains at least one curve that is essential in ${\mbox{$\Sigma$}}_t$. ![The blue regions are those where $\Sigma_t$ is mostly below $\Sigma'_s$ and the red regions are where $\Sigma_t$ is mostly above $\Sigma'_s$. The figures show $g$ spanning $f$ positively, negatively, both positively and negatively and $g$ splitting $f$.[]{data-label="fig:splitspan"}](spanning) Spanning sweep-outs and bounds on Euler characteristic {#sec:span} ====================================================== As in the last section, we will let $f$ and $g$ be sweep-outs for the pair $(M,T)$ associated to the two bridge splittings $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$ and $({\mbox{$\Sigma$}}',(H'^-, \tau'^-), (H'^+, \tau'^+))$ respectively. Define $\Sigma_t = f^{-1}(t)-\eta(T)$, and $\Sigma'_s = g^{-1}(s)-\eta(T)$ for $s,t \in(-1,1)$. We will also name the compression bodies $H'^-_s = g^{-1}(-1,s]$ containing the trivial arcs $\tau'^-_s$ and $H'^+_s = g^{-1}[s,1)$ containing the trivial arcs $\tau'^+_s$. \[thm:span\] Let $f$ and $g$ be sweep-outs associated to bridge surfaces $\Sigma$ and $\Sigma'$ for a prime, unsplit tangle $T$ in an irreducible orientable 3-manifold $M$. Suppose that $f \times g$ is generic and suppose that there are values $s$ and $t_+>t_0$ such that ${\mbox{$\Sigma$}}_{t_+}$ is mostly above ${\mbox{$\Sigma$}}'_s$ and ${\mbox{$\Sigma$}}_{t_0}$ is mostly below ${\mbox{$\Sigma$}}'_s$. Then there is a sequence of compressions and cut compressions of $\Sigma'_s$ after which there is a component of the compressed surface which is parallel to $\Sigma$. If there are values $s$ and $t_+>t_0>t_-$ such that ${\mbox{$\Sigma$}}_{t_+}$ and ${\mbox{$\Sigma$}}_{t_-}$ are mostly above ${\mbox{$\Sigma$}}'_s$ and ${\mbox{$\Sigma$}}_{t_0}$ is mostly below ${\mbox{$\Sigma$}}'_s$ then there is a sequence of compressions and cut compressions of $\Sigma'$ after which there are two components of the compressed surface which are parallel to $\Sigma$. Because the Euler characteristic is non-decreasing under c-compression, this theorem implies the following corollary: \[cor:span\] Let $f$ and $g$ be sweep-outs associated to bridge surfaces $\Sigma$ and $\Sigma'$ for a prime, unsplit tangle $T$ in an irreducible orientable 3-manifold $M$. If $f \times g$ is generic and $g$ spans $f$ positively (or negatively) then $\chi(\Sigma')\leq \chi(\Sigma)$. If $g$ spans $f$ both positively and negatively then $\chi(\Sigma')\leq 2\chi(\Sigma)$. [*Proof of Theorem \[thm:span\]*]{} We will only prove the second statement as the proof of the first statement is similar but simpler. Suppose there are values $s$ and $t_+>t_0>t_-$ such that ${\mbox{$\Sigma$}}_{t_+}$ and ${\mbox{$\Sigma$}}_{t_-}$ are mostly above ${\mbox{$\Sigma$}}'_s$ and ${\mbox{$\Sigma$}}_{t_0}$ is mostly below ${\mbox{$\Sigma$}}'_s$. By definition, each loop in the intersection ${\mbox{$\Sigma$}}'_s \cap ({\mbox{$\Sigma_{t_+}$}}\cup {\mbox{$\Sigma$}}_{t_0} \cup {\mbox{$\Sigma$}}_{t_-})$ bounds a disk or a once-punctured disk in ${\mbox{$\Sigma_{t_+}$}}\cup {\mbox{$\Sigma$}}_{t_0} \cup {\mbox{$\Sigma$}}_{t_-}$. To facilitate this discussion color ${H'_s}^-$ blue and ${H'_s}^+$ red. This induces a coloring on ${\mbox{$\Sigma$}}_{t_+}$, ${\mbox{$\Sigma$}}_{t_-}$ and ${\mbox{$\Sigma$}}_{t_0}$. As ${\mbox{$\Sigma$}}_{t_+}$ and ${\mbox{$\Sigma$}}_{t_-}$ are mostly above ${\mbox{$\Sigma$}}'_s$ every component of $({\mbox{$\Sigma$}}_{t_+} \cup {\mbox{$\Sigma$}}_{t_-})-{\mbox{$\Sigma$}}'_s$ that is not contained in a possibly once punctured sub-disk of ${\mbox{$\Sigma$}}_{t_+}$ or ${\mbox{$\Sigma$}}_{t_-}$ is red. Every component of ${\mbox{$\Sigma$}}_{t_0}-{\mbox{$\Sigma$}}'_s$ that is not contained in a possibly once punctured sub-disk of ${\mbox{$\Sigma$}}_{t_0}$ is blue. Let $\ell$ be an innermost in ${\mbox{$\Sigma_{t_+}$}}$ curve of ${\mbox{$\Sigma$}}'_s \cap {\mbox{$\Sigma_{t_+}$}}$. By hypothesis $\ell$ is necessarily inessential in ${\mbox{$\Sigma_{t_+}$}}$. If $\ell$ is also inessential in ${\mbox{$\Sigma$}}'_s$ then it can be removed by an isotopy of ${\mbox{$\Sigma$}}'_s$ as $M$ is irreducible and $T$ is prime. If $\ell$ is essential in $\Sigma'_s$ then the possibly punctured disk it bounds in ${\mbox{$\Sigma_{t_+}$}}$ is a c-disk for $\Sigma'_s$. Replace $\Sigma'_s$ with the surface $F_0$ that results from c-compressing $\Sigma'_s$ along this c-disk. Note that neither of these two moves affects the coloring of any region of ${\mbox{$\Sigma_{t_+}$}}-{\mbox{$\Sigma$}}'_s$ that is not contained in a possibly punctured subdisk of ${\mbox{$\Sigma_{t_+}$}}$. We can then repeat this construction with an innermost loop of $F_0 \cap {\mbox{$\Sigma_{t_+}$}}$, producing a surface $F_1$ and so on until we find a surface $F_k$ disjoint from ${\mbox{$\Sigma_{t_+}$}}$. At the end of this sequence of isotopies and c-compressions, ${\mbox{$\Sigma_{t_+}$}}$ will be entirely red. Repeat the above process with ${\mbox{$\Sigma$}}_{t_-}$ and $F_k$ playing the roles of ${\mbox{$\Sigma_{t_+}$}}$ and ${\mbox{$\Sigma$}}'_s$ respectively to obtain a surface $F_\ell$ disjoint from both ${\mbox{$\Sigma_{t_+}$}}$ and ${\mbox{$\Sigma$}}_{t_-}$ and leaving ${\mbox{$\Sigma$}}_{t_-}$ entirely red. Finally repeat the process beginning with ${\mbox{$\Sigma$}}_{t_0}$ and $F_\ell$ to obtain a surface $F_m$ disjoint from all of ${\mbox{$\Sigma_{t_+}$}}\cup {\mbox{$\Sigma$}}_{t_0} \cup {\mbox{$\Sigma$}}_{t_-}$ and leaving ${\mbox{$\Sigma$}}_{t_0}$ entirely blue. Maximally c-compress the surface $F_n=F_m \cap f^{-1}(t_-, t_+)$ in the complement of ${\mbox{$\Sigma_{t_+}$}}\cup {\mbox{$\Sigma_{t_0}$}}\cup {\mbox{$\Sigma_{t_-}$}}$ to get a surface $\tilde F$. Each component of $\tilde F$ is contained in a 3-manifold homeomorphic to ${\mbox{$\Sigma$}}\times I$ and is c-incompressible in this manifold. By [@T1 Corollary 3.7] each component of $\tilde F$ is either a sphere disjoint from $T$, a sphere bounding a ball containing a trivial subarc of $T$ or a component parallel to ${\mbox{$\Sigma_{t_0}$}}$. Note that $\tilde F$ was obtained from ${\mbox{$\Sigma$}}'_s$ by c-compressions and therefore it cannot have sphere components disjoint from $T$ as ${\mbox{$\Sigma$}}'_s$ does not have any such component, i.e., all component of $\tilde F$ have non-positive Euler characteristic. In addition $\tilde F$ separates ${\mbox{$\Sigma_{t_0}$}}$ from ${\mbox{$\Sigma_{t_+}$}}$ and ${\mbox{$\Sigma_{t_0}$}}$ from ${\mbox{$\Sigma_{t_-}$}}$ as ${\mbox{$\Sigma_{t_0}$}}$ is entirely blue and ${\mbox{$\Sigma_{t_+}$}}$ and ${\mbox{$\Sigma_{t_-}$}}$ are red. Therefore $\tilde F$ must have at least two components parallel to ${\mbox{$\Sigma_{t_0}$}}$, one lying in the product region between ${\mbox{$\Sigma_{t_-}$}}$ and ${\mbox{$\Sigma_{t_0}$}}$ and one lying in the product region between ${\mbox{$\Sigma_{t_0}$}}$ and ${\mbox{$\Sigma_{t_+}$}}$. Splitting sweep-outs and bounds on distance {#sec:split} =========================================== We briefly review the definition of distance of a bridge surface. For more details see [@T2]. \[def:distance\] Suppose $M$ is a compact, orientable, irreducible $3$–manifold containing a properly embedded tangle $T$ and suppose $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$ is a bridge splitting for $(M,T)$. The curve complex $\mathcal{C}(\Sigma,T)$ is a graph with vertices corresponding to isotopy classes of essential simple closed curves in $\Sigma-\eta(T)$. Two vertices are adjacent in $\mathcal{C}(\Sigma,T)$ if their corresponding classes of curves have disjoint representatives. Let $\mathcal{V}^+$ (resp $\mathcal{V}^-$) be the set of all essential simple closed curves in $\Sigma-\eta(T)$ that bound disks in $H^+ -\eta(T)$ (resp $H^- -\eta(T)$). Then the [*distance*]{} of the bridge splitting, $d(\Sigma,T)$, is the minimum distance between a vertex in $\mathcal{V}^+$ and a vertex in $\mathcal{V}^-$ measured in $\mathcal{C}(\Sigma,T)$ with the path metric. \[thm:split\] Let $f$ and $g$ be sweep-outs associated to bridge surfaces $\Sigma$ and $\Sigma'$ for a prime tangle $T$ in an irreducible 3-manifold $M$ and suppose that $\chi(\Sigma)\leq -1$. If $f \times g$ is generic and $g$ splits $f$ then $d({\mbox{$\Sigma$}},T) \leq 2-\chi({\mbox{$\Sigma$}}')$. Let $s$ be such that for every $t\in(-1,1)$ the intersection $\Sigma'_s \cap {\mbox{$\Sigma$}}_t$ contains a curve that is essential in ${\mbox{$\Sigma$}}_t$ and $[-1,1]\times \{s\}$ is disjoint from all vertices of $\Gamma$ so in particular $g|_{{\mbox{$\Sigma$}}_t}$ is Morse. Let $H_t^-=f^{-1}[-1,t)$ and $H_t^+=f^{-1}(t,1]$ be the two components of $M-\Sigma_t$. For values of $t$ close to $-1$, all curves of ${\mbox{$\Sigma$}}'_s \cap {\mbox{$\Sigma$}}_t$ bound disks in $H_t^-$ because ${\mbox{$\Sigma$}}'_s$ is transverse to the spine $f^{-1}(-1)$. Similarly for values of $t$ close to $1$, all curves of ${\mbox{$\Sigma$}}'_s \cap {\mbox{$\Sigma$}}_t$ bound disks in $H_t^+$. If for some value $t$ there is a curve of ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_t$ that is essential in ${\mbox{$\Sigma$}}_t$ and bounds a disk in $H_t^+$ and simultaneously there is a curve of ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_t$ that is essential in ${\mbox{$\Sigma$}}_t$ and bounds a disk in $H_t^-$, then $d({\mbox{$\Sigma$}},T) \leq 1$. We will say that two values $t_-$, $t_+$ are *adjacent* if there is a single critical value for $f$ between them. In this case, the projections of the curves ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_{t_-}$ to ${\mbox{$\Sigma$}}_{t_+}$ can be isotoped disjoint from the curves ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_{t_+}$. Thus, if for some adjacent values $t_-$, $t_+$, there is a curve of ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_{t_-}$ that is essential in ${\mbox{$\Sigma$}}_{t_-}$ and bounds a disk in $H_{t_-}^-$ and a curve of ${\mbox{$\Sigma$}}' \cap {\mbox{$\Sigma$}}_{t_+}$ that is essential in ${\mbox{$\Sigma$}}_{t_+}$ and bounds a disk in $H_{t_+}^+$, we can again conclude that $d({\mbox{$\Sigma$}},T) \leq 1$. The above discussion shows that either $d({\mbox{$\Sigma$}},T) \leq 1\leq 2-\chi(\Sigma')$ or there is an interval $[\alpha,\beta]$, where $\alpha \neq \beta$ are critical values for $f|_{{\mbox{$\Sigma$}}'_s}$ such that for every $t \in (\alpha,\beta)$, no curve of ${\mbox{$\Sigma$}}'_s \cap {\mbox{$\Sigma$}}_t$ is both essential in ${\mbox{$\Sigma$}}_t$ and bounds a disk in ${\mbox{$\Sigma$}}'_s$. Moreover for a very small $\epsilon$, ${\mbox{$\Sigma$}}'_s \cap \Sigma_{\alpha-\epsilon}$ contains curve that is essential in $\Sigma_{\alpha-\epsilon}$ and bounds a disk in $H^-_{\alpha-\epsilon}$ and ${\mbox{$\Sigma$}}'_s \cap \Sigma_{\beta+\epsilon}$ contains a curve that is essential in $\Sigma_{\beta+\epsilon}$ and bounds a disk in $H^+_{{\beta+\epsilon}}$. Let $\alpha'$ be just above $\alpha$ and $\beta'$ be just below $\beta$. Suppose some component of ${\mbox{$\Sigma$}}'_s \cap {\mbox{$\Sigma$}}_{\alpha'}$ bounds a disk in ${\mbox{$\Sigma$}}'_s$. Then this component must also bound a disk in ${\mbox{$\Sigma$}}_{\alpha'}$ and therefore ${\mbox{$\Sigma$}}'_s$ can be isotoped to remove this component. After some number of isotopies we obtains a surface ${\mbox{$\Sigma$}}''$ so that no curve of ${\mbox{$\Sigma$}}'' \cap {\mbox{$\Sigma$}}_{\alpha'}$ or ${\mbox{$\Sigma$}}'' \cap {\mbox{$\Sigma$}}_{\beta'}$ bounds a disk in ${\mbox{$\Sigma$}}''$. Define $S = {\mbox{$\Sigma$}}'' \cap f^{-1}[\alpha',\beta']$. Because the boundary curves of $S$ do not bound disks in ${\mbox{$\Sigma$}}''$, it follows that $\chi(S) \geq \chi({\mbox{$\Sigma$}}'_s)$. Let $\pi$ be the projection map from $f^{-1}[\alpha', \beta']$ to ${\mbox{$\Sigma$}}_0$. By [@JJ Lemma 22], isotopy classes of loops in $S$ project to isotopy classes in ${\mbox{$\Sigma$}}_0$. Although we are now dealing with punctured surfaces the proof of this result is the same so we will not repeat it here. As in [@JJ] we let $L$ be the set of isotopy classes of loops of $f|_S$ and let $\pi^*$ be the natural map from $L$ to $\mathcal{C}(\Sigma_0,T)$, together with $\{0\}$ where each curve in $L$ maps to the vertex that corresponds to its projection in ${\mbox{$\Sigma$}}_0$ unless it is inessential in ${\mbox{$\Sigma$}}_0$ in which case it is mapped to $\{0\}$. Note that $L$ determines a decomposition of $S$ into pairs of pants and punctured annuli. Lemma 23 in [@JJ] shows that if $\ell$ and $\ell'$ are cuffs of the same pair of pants, then their images under $\pi^*$ are adjacent vertices in $\mathcal{C}({\mbox{$\Sigma$}}_0,T)$. The same is true if $\ell$ and $\ell'$ are the two boundary components of a punctured annulus. For if that is the case, then $f|_{\Sigma'}$ passes through a puncture so it contains a level component which is an arc with both of its endpoints lying in a boundary component of $\Sigma'$. The projection of this component to $\Sigma_0$ is also an arc with both endpoints on some boundary component. The boundary curves of a regular neighborhood of the arc together with the boundary component are isotopic to the projections of $\ell$ and $\ell'$ and thus $\ell$ and $\ell'$ are disjoint. Let $L'=\pi_*(L)\cap \mathcal{C}({\mbox{$\Sigma$}}_0,T)$. By [@JJ Lemma 24] this set is connected and has diameter equal to at most the number of components of $S-L$. Each component of $S-L$ is a punctured annulus or a pair of pants and therefore contributes $-1$ to $\chi(S)$. It follows that $diam(L')\leq -\chi(S)$. Recall that for a very small $\epsilon$, ${\mbox{$\Sigma$}}'\cap \Sigma_{\alpha-\epsilon}$ contains a curve that bounds a compressing disk for $H_{\alpha-\epsilon}^-$ and ${\mbox{$\Sigma$}}'\cap \Sigma_{\beta+\epsilon}$ contains a curve that bounds a compressing disk for $H_{\beta+\epsilon}^+$. As the intervals $(\alpha-\epsilon,\alpha')$ and $(\beta', \beta+\epsilon)$ contain exactly one critical point each, every curve in the set $\pi( {\mbox{$\Sigma$}}'\cap \Sigma_{\alpha-\epsilon})$ is distance at most one from every curve in the set $\pi( {\mbox{$\Sigma$}}'\cap \Sigma_{\alpha'})$ and similarly every curve in the set $\pi( {\mbox{$\Sigma$}}'\cap \Sigma_{\beta+\epsilon})$ is distance at most one from every curve in the set $\pi( {\mbox{$\Sigma$}}'\cap \Sigma_{\beta'})$. Adding these distances we obtain the inequality $d(\Sigma)\leq diam(L')+2\leq 2-\chi(S)\leq 2-\chi(\Sigma')$ as desired. In this and in the previous section we saw that if $f$ and $g$ are two sweep-outs associated to bridge surfaces $\Sigma$ and $\Sigma'$ for the pair $(M,T)$ and $g$ spans $f$, then we can relate $\chi(\Sigma)$ and $\chi(\Sigma')$ and if $g$ splits $f$ then we can relate $d({\mbox{$\Sigma$}},T)$ and $\chi({\mbox{$\Sigma$}}')$. It is clear that if $g$ and $f$ are sweep-outs such that $f \times g$ is generic, then either $g$ spans $f$, $g$ splits $f$ or there is are values of $s$ and $t$ such that for a small $\epsilon$, ${\mbox{$\Sigma$}}_{t}$ is mostly above ${\mbox{$\Sigma$}}'_{s-\epsilon}$ and ${\mbox{$\Sigma$}}_{t}$ is mostly below ${\mbox{$\Sigma$}}'_{s+\epsilon}$. We now consider a slight generalization of this last case. \[lem:other\] Suppose $f$ and $g$ are sweep-outs for a tangle in a manifold such that $f \times g$ is generic except possibly for a single vertex of order 6 or two vertices of order 4 with the same $s$ coordinate. Suppose the graphic of $f \times g$ has a vertex at coordinates $(s,t)$ such that for a small $\epsilon$, ${\mbox{$\Sigma$}}_{t}$ is mostly above ${\mbox{$\Sigma$}}'_{s-\epsilon}$ and ${\mbox{$\Sigma$}}_{t}$ is mostly below ${\mbox{$\Sigma$}}'_{s+\epsilon}$. If this vertex has valence 4, then $\chi({\mbox{$\Sigma$}})\geq -2$. If the vertex has valence 6, then $\chi({\mbox{$\Sigma$}})\geq -3$. By the definition of $f \times g$ it follows that, $g|_{{\mbox{$\Sigma$}}_{t'}}$ is Morse where $t'=t+\epsilon'$ for a small $\epsilon'$. Furthermore there are two critical values for $g|_{{\mbox{$\Sigma$}}_{t'}}$, $a < b$ with at most one other critical value between them (if the valence of $(s,t)$ is 6) such that if $a'$ is a regular value directly below $a$ and $b'$ is a regular value directly above $b$, then ${\mbox{$\Sigma$}}_{t'}$ is mostly above ${\mbox{$\Sigma$}}'_{a'}$ and mostly below ${\mbox{$\Sigma$}}'_{b'}$. Consider first ${\mbox{$\Sigma$}}_{t'} \cap {\mbox{$\Sigma$}}'_{a'}$. By definition each component of ${\mbox{$\Sigma$}}_{t'} \cap H'^-_{a'}$ is contained in a possibly punctured disk subset of ${\mbox{$\Sigma$}}_{t'}$. Let $\Lambda$ be the set of all curves of ${\mbox{$\Sigma$}}_{t'} \cap {\mbox{$\Sigma$}}'_{a'}$ that are not contained in the interior of a disk or punctured disk component of ${\mbox{$\Sigma$}}_{t'} - {\mbox{$\Sigma$}}'_{a'}$, see Figure \[fig:curves\]. Then ${\mbox{$\Sigma$}}_{t'}-\Lambda$ is a collection of components all but one of which are possibly punctured disks. Note that the Euler characteristics of each of these possibly punctured disk components is at least 0. Passing through each critical point between $a'$ and $b'$ is equivalent to adding a band between two components of ${\mbox{$\Sigma$}}_{t'}-{\mbox{$\Sigma$}}'_{a'}$ or banding a component to itself. In either case the sum of the Euler characteristics of all components is decreased by one. As these bands correspond to a sweep-out they all lie on the same side of ${\mbox{$\Sigma$}}'$. As ${\mbox{$\Sigma$}}_{t'}$ is mostly below ${\mbox{$\Sigma$}}'_{b'}$, it follows that after attaching at most three bands to a collection of at most once punctured disks, the result is a surface isotopic to ${\mbox{$\Sigma$}}_{t'}$ with possibly some disks and once punctured disks missing, i.e. ${\mbox{$\Sigma$}}_{t'}\cap {\mbox{$\Sigma$}}'_{b'}$ is also as in Figure \[fig:curves\] but now the subsurface which is not contained in a punctured disk is below ${\mbox{$\Sigma$}}'_{b'}$. As at most three bands were added, it follows that $\chi({\mbox{$\Sigma$}}_{t'})\geq -3$. If the vertex $(s,t)$ has valence four, then only two bands need to be added so $\chi({\mbox{$\Sigma$}}_{t'})\geq -2$. ![[]{data-label="fig:curves"}](curves) Using the results in this and the previous section we can obtain the following generalization of the main result in [@T2]. \[thm:boundbridge\] Suppose $N$ is a manifold containing a tangle $K$ and let $M$ be submanifold such that $T=K \cap M$ is a properly embedded tangle. Let $\Sigma$ be a bridge surface of $(M,T)$ and let $\Sigma'$ be a bridge surface of $(N,K)$. Then one of the following holds: - There is an isotopy to ${\mbox{$\Sigma$}}'$ followed by some number of compressions and cut-compressions of $\Sigma'\cap M$ in $M$ giving a compressed surface $\Sigma''$, such that at least one component of $\Sigma'' \cap M$ is parallel to $\Sigma$, - $d(\Sigma,T) \leq 2-\chi(\Sigma')$, - $\chi(\Sigma)\geq -2 $. Because $(M,T) \subset (N,K)$ for values of $s$ close to $-1$, $\Sigma_t$ is mostly above $\Sigma'_s$ and for values of $s$ close to $1$, $\Sigma_t$ is mostly below $\Sigma'_s$. Therefore there are three possibilities. Either $g$ spans $f$, $g$ splits $f$ or the graphic of $f \times g$ has a vertex of valence 4 at coordinates $(s,t)$ such that for a small $\epsilon$, ${\mbox{$\Sigma$}}_{t}$ is mostly above ${\mbox{$\Sigma$}}'_{s-\epsilon}$ and ${\mbox{$\Sigma$}}_{t}$ is mostly below ${\mbox{$\Sigma$}}'_{s+\epsilon}$. Thus there are three cases to consider. **Case 1:** If $g$ spans $f$ then there are values $s$ and $t_+$ and $t_-$ such that ${\mbox{$\Sigma$}}_{t_+}$ is mostly above ${\mbox{$\Sigma$}}'_s$ and ${\mbox{$\Sigma$}}_{t_-}$ is mostly below ${\mbox{$\Sigma$}}'_s$. By the arguments in Theorem \[thm:span\] it follows that after some number of compressions and cut-compressions of $\Sigma_s'\cap M$ in $M$ we obtain an incompressible surface $\Sigma''$ that separates ${\mbox{$\Sigma$}}_{t_+}$ and ${\mbox{$\Sigma$}}_{t_-}$ and therefore it is parallel to $\Sigma$ as desired. **Case 2:** If $g$ splits $f$, then the arguments in the proof of Theorem \[thm:split\] show that $d(\Sigma,T)\leq 2-\chi(\Sigma')$ as desired. **Case 3:** Finally suppose that the graphic of $f \times g$ has a vertex at coordinates $(s,t)$ such that for a small $\epsilon$, ${\mbox{$\Sigma$}}_{t}$ is mostly above ${\mbox{$\Sigma$}}'_{s-\epsilon}$ and ${\mbox{$\Sigma$}}_{t}$ is mostly below ${\mbox{$\Sigma$}}'_{s+\epsilon}$. Then by Lemma \[lem:other\] it follows that $\chi(\Sigma)\geq -2$. Flipping bridge surfaces {#sec:flipping} ======================== In this section we want to restrict our attention to oriented isotopies, i.e., if ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ are bridge splittings for $(M,T)$ splitting the manifold into compression bodies $H^+, H^-$ and $H'^+, H'^-$ respectively, the bridge splittings $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ and $({\mbox{$\Sigma$}}', (H'^+, \tau'^+), (H'^-, \tau'^-))$ will be called [*orientation isotopic*]{} if there is an isotopy mapping ${\mbox{$\Sigma$}}$ to ${\mbox{$\Sigma$}}'$, $(H^+, \tau^+)$ to $(H'^+, \tau'^+)$ and $(H^-, \tau^-)$ to $(H'^-, \tau'^-)$. Following [@JJ] we will say that a bridge surface ${\mbox{$\Sigma$}}$ is [*flippable*]{} if $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ is orientation isotopic to $({\mbox{$\Sigma$}}, (H^-, \tau^-), (H^+, \tau^+))$. Suppose $({\mbox{$\Sigma$}}', (H'^+, \tau'^+), (H'^-, \tau'^-))$ is a bridge splitting for $(M,T)$ isotopic to stabilizations and perturbations of both bridge splittings $({\mbox{$\Sigma$}}, (H^+, \tau^+), (H^-, \tau^-))$ and $({\mbox{$\Sigma$}}, (H^-, \tau^-), (H^+, \tau^+))$. The minimal value of $2-\chi({\mbox{$\Sigma$}}')$ is called the [*flip Euler characteristic of ${\mbox{$\Sigma$}}$*]{} and it is analogous to the flip genus of a Heegaard splitting defined in [@JJ]. We will take advantage of several results previously proven for sweep-outs of Heegaard splittings. The proofs carry over with only minor modifications. \[lem:span\] (See Lemma 9 and Lemma 11 in [@JJ]) If $(\Sigma,(H^+,\tau^+), (H^-,\tau^-))$ is a bridge decomposition for some knot $K \subset M$ then $(\Sigma,(H^+,\tau^+), (H^-,\tau^-))$ spans itself positively. If $(\Sigma',(H'^+,\tau'^+),(H'^-,\tau'^-))$ is a perturbation or stabilization of $(\Sigma,(H^+,\tau^+),(H^-,\tau^-))$, then it spans the bridge splittings $(\Sigma, (H^+,\tau^+),(H^-,\tau^-))$ positively and spans $(\Sigma, (H^-,\tau^-), (H^+,\tau^+))$ negatively. Let $f$ be a sweep-out for the decomposition $(\Sigma,(H^+,\tau^+), (H^-,\tau^-))$. Then there is a second sweep-out $g$ for $(\Sigma, (H^+,\tau^+),(H^-,\tau^-))$ such that $g^{-1}(0)=\Sigma'_0$ is disjoint from and separates the spines $f^{-1}(-1)$, $f^{-1}(1)$. Thus for $t_-$ near $-1$ and $t_+$ near $1$, $\Sigma_{t_-}$ will be mostly below $\Sigma'_0$ and $\Sigma_{t_+}$ will be mostly above. This implies that $(\Sigma,(H^+,\tau^+), (H^-,\tau^-))$ spans itself positively. Similarly, for any perturbation or stabilization of the bridge decomposition, we can perturb or stabilize $\Sigma'_0$ while keeping it disjoint from $f^{-1}(-1)$, $f^{-1}(1)$, then extend this surface to a sweep-out for the perturbed or stabilized bridge decomposition. Thus if $(\Sigma',(H'^+,\tau'^+),(H'^-,\tau'^-))$ is a perturbation or stabilization of $(\Sigma,(H^+,\tau^+),(H^-,\tau^-))$, then it spans the bridge splitting $(\Sigma, (H^+,\tau^+),(H^-,\tau^-))$ positively and spans $(\Sigma, (H^-,\tau^-), (H^+,\tau^+))$ negatively. Let ${\mbox{$\Sigma$}}'$ be a common stabilization or perturbation of the two bridge splittings $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$ and $({\mbox{$\Sigma$}},(H^+, \tau^+), (H^-, \tau^-))$. In particular there are sweep-outs $g$ and $g'$ representing $(\Sigma',(H'^-, \tau'^-), (H'^+, \tau'^+))$, and a sweep out $f$ representing $({\mbox{$\Sigma$}},(H^-, \tau^-), (H^+, \tau^+))$ such that $g$ spans $f$ positively and $g'$ spans $f$ negatively. As $g$ and $g'$ represent the same bridge decomposition and are therefore orientation isotopic it follows that there is a family of sweep-outs $\{g_r|r \in [0,1]\}$ such that $g_0=g$, $g_1=g'$. ![[]{data-label="fig:span"}](spanningchange) [@JJ Lemma 26]\[lem:iso\] Let $g$ and $g'$ be sweep-outs such that $ f\times g$ and $f \times g'$ are generic and $g$ is isotopic to $g'$. There is a family of sweep-outs $\{g_r|r \in [0,1]\}$ such that $g_0=g$, $g_1=g'$ and for all but finitely many $r$, $f \times g_r$ is generic. At the finitely many non-generic points there are at most two valence two or four vertices at the same level or there is a single valence 6 vertex. We can now prove our first main result. Consider the family of sweep-outs $\{g_r|r \in [0,1]\}$ described in Lemma \[lem:iso\]. As $g_0$ spans $f$ positively and $g_1$ spans $f$ negatively there must be some $r$ such that either $g_r$ splits $f$, $g_r$ spans $f$ both positively and negatively, or the hypothesis of Lemma \[lem:other\] are satisfied. This is illustrated in Figure \[fig:span\]. Case 1: $g_r$ splits $f$. In this case by Theorem \[thm:split\] it follows that $d({\mbox{$\Sigma$}},T)\leq 2-\chi(\Sigma')$ so $\chi({\mbox{$\Sigma$}}')\leq 2-d({\mbox{$\Sigma$}},T)$. Case 2: $g_r$ spans $f$ both positively and negatively. In this case by Theorem \[thm:span\] it follows that $\chi(\Sigma)\leq 2\chi(\Sigma')$. Case 3: There are at most two valence two or four vertices at the same level or there is a valence 6 vertex. By an argument identical to the one in the proof of [@JJ Lemma 26] it follows that either we are in one of cases 1 or 2 or there is a vertex of valence 4 or valence 6 corresponding to coordinates $(s,t)$ such that for a very small $\epsilon$ the surface $f^{-1}_{t+\epsilon}$ is mostly above $g^{-1}(s)$ and $f^{-1}_{t-\epsilon}$ is mostly below $g^{-1}(s)$. Lemma \[lem:other\] shows that in this case $\chi({\mbox{$\Sigma$}})\geq -3$ contradicting the hypothesis. ![[]{data-label="fig:flip"}](flip) [*Proof of Corollary \[cor:flip\]*]{} Figure \[fig:flip\] shows that if $T$ has $n$ bridges with respect to a bridge sphere ${\mbox{$\Sigma$}}$, then there is a flippable bridge sphere $\Sigma'$ obtained from ${\mbox{$\Sigma$}}$ by perturbations with respect to which $T$ has $2n$ bridges. The fact that there is no such bridge sphere with fewer punctures follows by Theorem \[thm:flip\]. Bridge surfaces that require a large number of stabilization and perturbations to become equivalent {#sec:example} =================================================================================================== [*Proof of Theorem \[thm:knot\]*]{}. The proof of this theorem consists of a construction for a pair $(M,K)$ where $K$ is a knot with two distinct bridge surfaces, ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$ with $\chi({\mbox{$\Sigma$}})=2s$ and $\chi({\mbox{$\Sigma$}}')=2s-2$ so that for every common stabilization/perturbation ${\mbox{$\Sigma$}}''$ of ${\mbox{$\Sigma$}}$ and ${\mbox{$\Sigma$}}'$, $\chi({\mbox{$\Sigma$}}')\leq 3s+2$. In particular this construction gives examples of knots in $S^3$ with distinct bridge spheres with bridge number $2n$ and $2n-1$, respectively, for which any common perturbation has at least $3n-1$ bridges. Let $K$ be a knot in a manifold $M$ and let $(\Sigma,(H^+,\tau^+),(H^-,\tau^-))$ be a bridge splitting for $(M,K)$ so that, $\chi(\Sigma)\leq -4$ and $d(\Sigma,K)\geq-3\chi(\Sigma)$. Suppose $f$ is a sweep-out for $M$ associated to $\Sigma$. Let $\Gamma_+$ be a spine of $(H^+,\tau^+)$ and let $\Gamma_-$ be a spine of $(H^-,\tau^-)$ so that $\Gamma_-=f^{-1}(-1)$ and $\Gamma_+=f^{-1}(1)$. Choose an edge of $\Gamma_-$ that has a valence 1 vertex; i.e., an edge that has one endpoint in $K$. Let $B$ be a ball that is a regular neighborhood of this edge and let $M^-$ be the closure of $M \setminus B$ containing the one strand tangle $K^-=K \cap M^-$. The sweep-out $f$ on $M$ can be modified to be a sweep-out of $M^-$ by perturbing $f$ to be constant in $B$. We will use $f$ to refer to either sweep-out when the manifold is clear from context. Let $P$ be a manifold homeomorphic to $S^2 \times I$ containing two vertical arcs $\tau_1$ and $\tau_2$. Construct a new manifold $M\#M$ by gluing a copy of $M^-$ to each of the boundary sphere of $P$ so that the endpoints of each copy of $K^-$ are identified with one endpoint of $\tau_1$ and one endpoint of $\tau_2$ to obtain a new knot $K\#K$. Then $(M\#M,K\#K)$ is the connect sum of two copies of $(M,K)$. The pair $(M\#M, K\#K)$ has two natural generalized Heegaard splittings $\mathcal{H}_1$ and $\mathcal{H}_2$ induced by the bridge splittings for $M$ and $P$, shown in Figures \[fig:generalized\] and \[fig:composite\]. In both cases we will take $\Sigma$ to be the bridge surface for each copy of $(M^-,K^-)$. However in the first generalized Heegaard splitting we will take the surface $S^2 \times \{1/2\}$ to be the bridge surface for $(P, \tau_1 \cup \tau_2)$ and for the second one we will take the bridge surface for $P$ to be the surface obtained by tubing together the two spheres which are boundaries of small collars of $S^2 \times \{0\}$ and $S^2 \times \{1\}$ respectively along a vertical tube, see Figure \[fig:generalized\]. Let $\Sigma_1$ and $\Sigma_2$ be the two bridge surfaces for $(M\#M,K\#K)$ obtained by amalgamating $\mathcal{H}_1$ and $\mathcal{H}_2$ respectively. Note that $\chi(\Sigma_1)=2\chi(\Sigma)$ and $\chi(\Sigma_2)=2\chi(\Sigma)-2$. ![Schematic depiction of $\mathcal{H}_1$ and $\mathcal{H}_2$.[]{data-label="fig:generalized"}](generalized) ![The figure depicts $\Sigma_1$ and $\Sigma_2$ if $M=S^3$ and $\Sigma=S^2$.[]{data-label="fig:composite"}](composite) \[lem:example\] If $\Sigma'$ is isotopic to a surface obtained via a sequence of stabilizations and perturbations of $\Sigma_1$ and also to a surface obtained via sequence of stabilizations and perturbations of $\Sigma_2$, then $\chi(\Sigma') \leq 3\chi(\Sigma)+2$. Let $\mathcal {H}^-_i$ and $\mathcal {H}^+_i$ be the thin and thick surfaces for $\mathcal{H}_i$. Then $M\#M-\mathcal{H}^-_i$ has two components that are homeomorphic to $M^-$ and each of these components has a Heegaard surface $\Sigma$. Let $M^-_1$ and $M^-_2$ be these components and let $f_1$ and $f_2$ be their sweep-outs associated to $\Sigma$. The following remark is clear in the case when $M=S^3$ and $\Sigma=S^2$, as shown in Figure \[fig:composite\]. In the general case the proof is very similar to the proof of [@JJ Lemma 14] so we leave the details to the reader. \[rmk:amalgamation\] Let ${\mbox{$(f_1\uparrow \uparrow f_2)$}}$ be the sweep-out associated to $\Sigma_1$ and ${\mbox{$(f_1\uparrow \downarrow f_2)$}}$ be the sweep-out associated to $\Sigma_2$. Then ${\mbox{$(f_1\uparrow \uparrow f_2)$}}$ spans $f_1$ and $f_2$ positively and ${\mbox{$(f_1\uparrow \downarrow f_2)$}}$ spans $f_1$ positively and $f_2$ negatively. Let $g$ and $g'$ be sweep-outs for $(M\#M,K\#K)$ defined by perturbing and stabilizing sweep-outs ${\mbox{$(f_1\uparrow \uparrow f_2)$}}$ and ${\mbox{$(f_1\uparrow \downarrow f_2)$}}$ enough times so that $g$ and $g'$ represent isotopic bridge decompositions. By Lemma \[lem:span\] it follows that $g$ spans ${\mbox{$(f_1\uparrow \uparrow f_2)$}}$ positively. By Remark \[rmk:amalgamation\] it follows that ${\mbox{$(f_1\uparrow \uparrow f_2)$}}$ spans both $f_1$ and $f_2$ positively. Therefore we conclude that $g$ spans both $f_1$ and $f_2$ positively. Similarly, $g'$ spans $f_1$ positively and $f_2$ negatively. As $g$ and $g'$ represent isotopic bridge decompositions, the sweep-out $g$ is isotopic to either $g'$ or $-g'$. In other words, there is a family of sweep-outs $\{g_r|r \in [0,1]\}$ such that $g_0=g$, $g_1=\pm g'$. Consider the family of sweep-outs $\{g_r|r \in [0,1]\}$ described in Lemma \[lem:iso\]. Because $g'$ spans $f_1$ positively and $f_2$ negatively, the sweep-out $g_1 = \pm g'$ spans one of $f_1$ or $f_2$ positively and the other negatively. Without loss of generality, assume $g_1$ spans $f_1$ negatively. As $g_0$ spans $f_1$ positively and $g_1$ spans $f_1$ negatively, Lemma \[lem:iso\] implies that there is an $r$ satisfying one of the following: Case 1: $g_r$ splits $f_1$ or $g_r$ splits $f_2$. The argument is the same so suppose $g_r$ splits $f_1$. In this case by Theorem \[thm:split\] it follows that $d(\Sigma_1,K) \leq 2-\chi(\Sigma')$. As $d(\Sigma_1,K) \geq -3\chi(\Sigma)$ by construction, it follows that $\chi(\Sigma') \leq 2+4\chi(\Sigma)\leq 3\chi(\Sigma)+2$ as required. Case 2: $g_r$ spans $f_1$ both positively and negatively and $g_r$ spans $f_2$, say positively. By the definition of spanning there exist $s$ and $t_+>t_0>t_-$ such that $(\Sigma_1)_{t_+}$ and $(\Sigma_1)_{t_-}$ are mostly above ${\mbox{$\Sigma$}}'_s\cap M^-_1$ and $(\Sigma_1)_{t_0}$ is mostly below ${\mbox{$\Sigma$}}'_s\cap M^-_1$ and there exist $u$ and $t'_0 < t'_+$ so that $(\Sigma_2)_{t'_0}$ is mostly below ${\mbox{$\Sigma$}}'_u \cap M^-_2$ and $(\Sigma_2)_{t'_+}$ is mostly above ${\mbox{$\Sigma$}}'_u\cap M^-_2$. If we can choose $s$ and $u$ to be equal, then by Theorem \[thm:span\] it follows that $\chi({\mbox{$\Sigma$}}'_u)=\chi({\mbox{$\Sigma$}}'_u \cap M^-_1)+\chi({\mbox{$\Sigma$}}'_u\cap M^-_2)\leq 2\chi(\Sigma_1)+\chi(\Sigma_2)=3\chi(\Sigma)$ as desired. Suppose that no such value exists. Without loss of generality suppose that $s < u$ and choose $s$ and $u$ to be such that $u-s$ is minimal. By the choice of $s$ and Theorem \[thm:span\] it follows that $\chi({\mbox{$\Sigma$}}'_s\cap M^-_1 )\leq 2\chi(\Sigma_1)$. Let $S$ be the decomposing sphere for $(M\#M, K\#K)$ (we may take $S=S^2\times \{1/2\}$ in $P$). The surface ${\mbox{$\Sigma$}}'_u$ can be obtained from ${\mbox{$\Sigma$}}'_s$ by a series of boundary compressions of $\Sigma'_u \cap M^-_1$ and $\Sigma'_u \cap M^-_2$ together with isotopies of the surface that are the identity in a neighborhood of $S$. Let $\Sigma'_0, \Sigma'_1,...,\Sigma'_n$ be the sequence of surfaces so that $\Sigma'_0$ is isotopic relative to $S$ to $\Sigma'_s$, $\Sigma'_n$ is isotopic relative to $S$ to $\Sigma'_u$ and $\Sigma'_k$ is obtained from $\Sigma'_{k-1}$ by performing a boundary compression of $\Sigma'_{k-1} \cap M^-_1$ or a boundary compression of $\Sigma'_{k-1} \cap M^-_2$ along a disk $D_k$. Following [@J] we will call these boundary compressions $\alpha$-isotopies of $\Sigma'$ and we will call the boundary compressing disks, $\alpha$-disks. As the isotopy between $\Sigma'_s$ and $\Sigma'_u$ represents a sweep-out, all $\alpha$-disks are on the same side of ${\mbox{$\Sigma$}}'$, in this case, the positive side as $s <u$. Every $\alpha$-disk $D_k$ has a dual $\alpha$-disk $E_k$ contained in the negative side of $\Sigma'_k$ that can be used to perform an $\alpha$-isotopy on $\Sigma'_{k}$ to recover $\Sigma'_{k-1}$, as in Figure \[fig:dual\]. ![[]{data-label="fig:dual"}](dual) **Claim:** There is a collection of disks $E_1,...,E_n$ such that for every $k$ and $j$ the following hold - ${\mbox{$\partial$}}E_k$ is the endpoint union of an arc in $S$ and an arc in ${\mbox{$\Sigma$}}'_u$, - $E_k \cap E_j$ is either empty or it is equal to some disk $E_l$, - one component of the boundary of an $\epsilon$-neighborhood of ${\mbox{$\Sigma$}}'_u \cup (E_n\cup...\cup E_{k+1})$ is isotopic to ${\mbox{$\Sigma$}}'_u$ relative to $S$ and the other is isotopic to ${\mbox{$\Sigma$}}'_k$ relative to $S$, - $E_k$ is an $\alpha$-disk for $\Sigma'_k$. In particular if $\mathcal{E}=\cup_1^n E_i$ then one component of the boundary of an $\epsilon$-neighborhood of ${\mbox{$\Sigma$}}'_u \cup \mathcal{E}$ is isotopic to ${\mbox{$\Sigma$}}'_u$ relative to $S$ and the other is isotopic to ${\mbox{$\Sigma$}}'_s$ relative to $S$. [*Proof of claim:*]{} Let $\mathcal{E}_j=\cup_j^n E_j$. We will prove the claim by induction on $n-j$. The result is clear for $\mathcal{E}_n$ as this collection contains a single disk. Suppose the result holds for $\mathcal{E}_{k+1}$. Therefore $\Sigma'_k$ is isotopic relative $S$ to one of the two boundary components of a regular neighborhood of ${\mbox{$\Sigma$}}'_u \cup \mathcal{E}_{k+1}$. Consider a disk $D_k$ that realizes the $\alpha$-isotopy between ${\mbox{$\Sigma$}}'_{k-1}$ and ${\mbox{$\Sigma$}}'_k$. Then the boundary of a regular neighborhood of ${\mbox{$\Sigma$}}'_{k-1}\cup D_k$ has two components, one is ${\mbox{$\Sigma$}}'_{k-1}$ and the other is ${\mbox{$\Sigma$}}'_k$. If $D_k \cap \mathcal{E}_{k+1}=\emptyset$, let $E_k$ be the disk dual to $D_k$. If $D_k$ has a nonempty intersection with each of $E_{i_1},..., E_{i_l}$, let $\gamma_{i_r}$ be a small neighborhood of $D \cap E_{i-r}$ in ${\mbox{$\partial$}}E_{i+r}$. Identify all arcs $\gamma_{i_r}$ for $r=1,...,l$ to a single arc $\gamma$ and let $E_k$ be the disk dual to $D_k$ so that $E_k \cap \Sigma'=\gamma$. Consider the surface $\Sigma'_u \cap M^-_2$. By Theorem \[thm:span\] there exists a collection ${\bf D}$ of compressing and cut-compressing disks for $\Sigma'_u \cap M^-_2$ after which the resulting union of surfaces $(\Sigma'_u)^{\bf D} \cap M^-_2$ contains a component parallel to $\Sigma_2$. We can choose these disks so that a subcollection ${\mbox{$\partial$}}\mathcal{D}_u$ is contained in $S$ and so that compressing $\Sigma'_u$ along the collection ${\mbox{$\partial$}}\mathcal{D}_u$ produces a surface disjoint from $S$. Note that $\chi((\Sigma'_u)^{\bf D}\cap M^-_2) \leq \chi(\Sigma)$ The surface $\Sigma'_u$ is isotopic to the union of $(\Sigma'_u)^{\bf D}$ together with some tubes between the components. Some of these tubes may run along the knot. Let $\mathcal{D}_u\subset {\bf D}$ be the collecting of possibly nested disks and punctured disks these tubes bound in $S$. Then $\mathcal{E}\cap S$ is a collection of disjoint arcs $\Lambda$ with endpoints in ${\mbox{$\partial$}}\mathcal{D}_u$. Label the regions of $S-\Sigma'_u$ positive or negative depending on which side of $\Sigma'_u$ they lie in and recall that the collection of arcs $\Lambda$ lies in the negative regions. The surface $S$ intersects the knot $K$ in two points, which we will label $p_1$, $p_2$. The curves in $\Sigma_u' \cap S$ can be classified into three categories depending on whether the corresponding c-disk in $\mathcal{D}_u$ is a disk, a punctured disk containing $p_1$, or a punctured disk containing $p_2$. Let $\Gamma_1$, $\Gamma_2$ and $\Gamma_3$ be these collection of curves and let $\Gamma=\Gamma_1\cup\Gamma_2 \cup \Gamma_3$. Note that the only arcs in $\mathcal{E}\cap S$ that have endpoints in both $\Gamma_2$ and $\Gamma_3$ are between the outermost curve $\gamma_2$ in $\Gamma_2$ and the outermost curve $\gamma_3$ in $\Gamma_3$. Also note that if an arc has both of its endpoints in a curve $\gamma$ so that $\gamma \in \Gamma_2 \cup \Gamma_3$, then the arc is parallel to a subarc of $\gamma$. Let $\Delta$ be the union of all these, possibly nested, disks of parallelism. Let $p$ be a point in $S-(\mathcal{D}\cup \Delta)$. Let $F$ be the twice punctured disk obtained by removing a neighborhood of $p$ from $S$. Then $\mathcal{D}_u \subset F$ and no curve in $\Sigma'_u\cap F$ is parallel to ${\mbox{$\partial$}}F$. Furthermore the boundary of a regular neighborhood of $(\mathcal{E}\cap S)\cup\Gamma$ contains at most one curve that bounds a twice punctured disk in $F$. This curve is obtained by taking a regular neighborhood of the component of $\Gamma\cup (\mathcal{E}\cap S)$ containing $\gamma_2 $, $\gamma_3$ and an arc of $\mathcal{E}\cap S$ connecting the two. Let $\mathcal{D}_s$ be the collection of possibly punctured disks that the boundary of a regular neighborhood of $(\mathcal{E}\cap S)\cup\Gamma$ bounds in $F$. By the claim, the boundary of a regular neighborhood of $\Sigma'_u \cup \mathcal{E}$ contains two surfaces. One is isotopic to $\Sigma'_u$ relative $S$, and the other one is isotopic to $\Sigma'_s$ relative to $S$. Then $\Sigma'_s\cap S = {\mbox{$\partial$}}\mathcal{D}_s$. Let $(\Sigma'_s)^{\mathcal{D}_s}$ be the surface obtained from $\Sigma'_s$ by compressing it along $\mathcal{D}_s$. Note that some of these compressions may be along cut-disks and at most one may be along a disk with two punctures. Therefore $\chi(\Sigma'_s \cap M^-_2)\leq \chi((\Sigma'_s)^{\mathcal{D}_s}\cap M^-_2)+2$. Note that $(\Sigma'_s)^{\mathcal{D}_s} \cap M^-_2$ is isotopic to $(\Sigma'_u)^{\mathcal{D}_u} \cap M^-_2$ so has Euler Characteristic at most $\chi(\Sigma)$ and therefore $\chi(\Sigma'_s \cap M^-_2)\leq \chi(\Sigma)+2$. On the other hand, by our choice of $s$, $\chi(\Sigma'_s \cap M^-_1)\leq 2\chi(\Sigma)$. Therefore $\chi(\Sigma')=\chi(\Sigma'_s \cap M^-_1)+\chi(\Sigma'_s \cap M^-_2)\leq 3\chi(\Sigma)+2$ as desired. Case 3: There are at most two valence two or valence four vertices at the same level or there is a valence 6 vertex. As in Theorem \[thm:flip\] this implies that $\chi(\Sigma)\geq -3$ contrary to our hypothesis. [10]{} D. Bachman. Heegaard splittings of sufficiently complicated 3-manifolds I: Stabilization. . Jesse Johnson Flipping and stabilizing Heegaard splittings. Jesse Johnson Bounding the stable genera of Heegaard splittings from below. David Bachman and Saul Schleimer. Distance and bridge position. , 219(2):221–235, 2005. A. J. Casson and C. McA. Gordon. Reducing [H]{}eegaard splittings. , 27(3):275–283, 1987. Wolfgang Haken. Some results on surfaces in [$3$]{}-manifolds. , 39–98, 1968. J. Hass, A. Thompson, and W. Thurston. Common stabilizations of Heegaard splittings. . Chuichiro Hayashi and Koya Shimokawa. Heegaard splittings of the pair of the solid torus and the core loop. , 14(2):479–501, 2001. Chuichiro Hayashi and Koya Shimokawa. Heegaard splittings of trivial arcs in compression bodies. , 10(1):71–87, 2001. Chuichiro Hayashi and Koya Shimokawa. Thin position of a pair (3-manifold, 1-submanifold). , 197(2):301–324, 2001. William Jaco. , volume 43 of [*CBMS Regional Conference Series in Mathematics*]{}. American Mathematical Society, Providence, R.I., 1980. K. Reidemesiter Zur dreidimensionaen Topologie. , 11, 189–194, 1933. Martin Scharlemann and Abigail Thompson. Heegaard splittings of [$({\rm surface})\times I$]{} are standard. , 295(3):549–564, 1993. Martin Scharlemann and Abigail Thompson. Thin position for [$3$]{}-manifolds. In [*Geometric topology (Haifa, 1992)*]{}, volume 164 of [ *Contemp. Math.*]{}, pages 231–238. Amer. Math. Soc., Providence, RI, 1994. Martin Scharlemann and Maggy Tomova. . , . J. Singer Three dimensional manifolds and their Heegaard diagrams. , 35, 88-111, 1933. Maggy Tomova. . , 80:85–98, 2009. Maggy Tomova. . , 7:957–1006, 2007. F. Waldhausen. , , 7:195–203, 1968. [^1]: Research partially supported by an NSF grant.
{ "pile_set_name": "ArXiv" }
--- abstract: | Edge-on spiral galaxies offer a unique perspective on the vertical structure of spiral disks, both stars and the iconic dark dustlanes. The thickness of these dustlanes can now be resolved for the first time with [*Herschel*]{} in far-infrared and sub-mm emission. We present NHEMESES, an ongoing project that targets 12 edge-on spiral galaxies with the PACS and SPIRE instruments on [*Herschel*]{}. These vertically resolved observations of edge-on spirals will impact on several current topics. First and foremost, these [*Herschel*]{} observations will settle whether or not there is a phase change in the vertical structure of the ISM with disk mass. Previously, a dramatic change in dustlane morphology was observed as in massive disks the dust collapses into a thin lane. If this is the case, the vertical balance between turbulence and gravity dictates the ISM structure and consequently star-formation and related phenomena (spiral arms, bars etc.). We specifically target lower mass nearby edge-ons to complement existing [*Herschel*]{} observations of high-mass edge-on spirals (the HEROES project). Secondly, the combined data-set, together with existing [*Spitzer*]{} observations, will drive a new generation of spiral disk Spectral Energy Distribution models. These model how dust reprocesses starlight to thermal emission but the dust geometry remains the critical unknown. And thirdly, the observations will provide an accurate and unbiased census of the cold dusty structures occasionally seen extending out of the plane of the disk, when backlit by the stellar disk. To illustrate the NHEMESES project, we present early results on NGC 4244 and NGC 891, two well studies examples of a low and high-mass edge-on spiral. title: 'New HErschel Multi-wavelength Extragalactic Survey of Edge-on Spirals (NHEMESES)' --- Edge-on spiral galaxies offer a unique perspective on spiral disks. An observer can explore the vertical structure of both stars and ISM using different wavelengths. Dust is mechanically linked to the cold ISM [@Allen86; @Weingartner01b]. [@Dalcanton04] used the appearance of dust lanes as a probe of vertical stability of spiral disks. They found that in massive spiral disks, the ISM collapses into a thin dustlane, while the less massive disks show more flocculant dust morphology. In addition, observations of edge-ons often reveal dark structures extending out of the plane of massive disks [e.g., @Howk99a], but the quantity of extra-planar dust has been impossible to constrain from extinction measures. A comprehensive approach is to model multi-wavelength images of edge-on spirals with a bulge, a stellar and a dust disk to constrain stellar and ISM structure [e.g., @Xilouris99; @Bianchi08; @Baes10a; @Popescu11]. However, until now, these models are often degenerate in the vertical distributions of stellar light and ISM due to lack of resolution and wavelength coverage. With the advent of [*Herschel*]{}, the vertical structure of some nearby edge-on disks can be now resolved. Our [*Herschel*]{} program NHEMESES[^1] (PI B.W. Holwerda) aims to (1) measure the vertical scale of dust in spiral disks (an unknown in the above models), (2) the extra-planer dust component, and (3) whether or not the strong break seen in dust lanes by [@Dalcanton04] is indeed a real phase-change in the cold ISM with disk mass. ![\[f:bwh:f2\] The vertical profile of NGC4244, same legend as Fig. \[f:bwh:f1\]. The SPIRE profile stays within the 3.6 $\mu$m height.](./ngc4244_SPIRE_radial.pdf){width="\textwidth"} ![\[f:bwh:f2\] The vertical profile of NGC4244, same legend as Fig. \[f:bwh:f1\]. The SPIRE profile stays within the 3.6 $\mu$m height.](./ngc4244_SPIRE_vertical.pdf){width="\textwidth"} We show some first NHEMESES results with the SPIRE morphology of NGC 4244, the prototypical disk-dominated low-mass spiral galaxy ($\rm v_{rot} = 95 ~ km s^{-1}$). The radial profile (Fig. \[f:bwh:f1\]) appears to truncate beyond a peak in flux from a small star-formation region on both sides of the disks (see also Fig. \[f:bwh:f3\]). The radius of truncation is similar to those initially found by [@vdKruit81a] and more recently by [@de-Jong07] for different stellar populations. The vertical sub-mm profiles (Fig. \[f:bwh:f2\]) are similar in width as the [*Spitzer*]{} 3.6 $\mu$m emission, a good tracer of the stellar mass (e.g., Meidt et al., [*this volume*]{}). [@Comeron11a] find evidence for a thick and thin stellar disk in this galaxy and the dusty ISM appears associated with the inner (thin) disk. In contrast, a second, thicker vertical component in NGC 891 have been reported [@Kamphuis07 Seon et al., this volume]. Fig. \[f:bwh:f3\] shows the color image based on the 250, 350 and 500 $\mu$m SPIRE images with the  contours from [@Zschaechner11a; @Heald11]. The dust emission is restricted to the highest  contour; it is a single, concentrated disk. This is in contrast to the more massive NGC 891, where [@Popescu03; @Bianchi11] find evidence for dust throughout the  disk and extended envelope [@Oosterloo07]. ![ \[f:bwh:f4\] The Spectral Energy Distribution model from [@MacLachlan11] (see also MacLachlan et al.([*this volume*]{}) and the SPIRE fluxes for NGC 4244. The SED under-predicts the sub-mm fluxes by only a factor 2.](./n4244_spire_hi.pdf){width="\textwidth"} ![ \[f:bwh:f4\] The Spectral Energy Distribution model from [@MacLachlan11] (see also MacLachlan et al.([*this volume*]{}) and the SPIRE fluxes for NGC 4244. The SED under-predicts the sub-mm fluxes by only a factor 2.](./n4244_simone_sed.pdf){width="\textwidth"} The Spectral Energy Distribution (SED) model by [@MacLachlan11], in Fig \[f:bwh:f4\], finds an optically thin disk with s similar scale-height for both dust and stars [as did @Seth05a based on stellar populations]. They report a dust mass of $ 2.38 \times 10^6 M_\odot $ and a dust scale-length 1.8 times its stellar scale length. We find that their predicted sub-mm fluxes are lower than what we observe, possibly because the dusty ISM is much clumpier than the smooth diffuse ISM they assume for the SED model. However, we find that their high value of the scale-length of the dusty disk strokes with the radial profile of sub-mm emission we observe (Fig. \[f:bwh:f1\]). We find that dust in NGC 4244 is distributed vertically throughout the (thin) stellar disk, in accordance with the prediction from [@Dalcanton04], but that, unlike the much more massive NGC 891, there is little evidence for a second dust component associated with the outer  envelope. SED models for this galaxy are converging to a solution which includes a large scales (-height and -length) for the dust compared to the stellar ones but distributed almost exclusively in clumps. This different distribution of the dusty geometry for lower-mass spiral disks may help explain the colder dust temperature and higher relative dust mass observed by H-ATLAS [@Dunne11 Dunne et al. [*this volume*]{}]. The NHEMESES project aims to observe 12 low-mass spirals with PACS and SPIRE on board [*Herschel*]{}. Combined with existing [*Herschel*]{} observations of massive edge-on spiral galaxies, we aim to (1) ascertain if the ISM indeed goes through a phase-change at 120 km/s, (2) take a census of dusty outflows in nearby, quiescent, spiral disks and (3) provide a suite of multi-wavelength data (in combination with SDSS and [*Spitzer*]{} observations) to serve as a benchmark for SED models of edge-on spirals galaxies. [2]{} [20]{} natexlab\#1[\#1]{} , R. J. [et al.]{} 1986, , 319, 296 , M. [et al.]{} 2010, , 518, L39+ , S. 2008, , 490, 461 , S., & [Xilouris]{}, E. M. 2011, , 531, L11+ , S. [et al.]{} 2011, , 729, 18 , J. J. [et al.]{} 2004, , 608, 189 , R. S. [et al.]{} 2007, , 667, L49 , L. [et al.]{} 2011, , 1395 , G. [et al.]{} 2011, ArXiv e-prints , J. C. 1999, , 269, 293 , P. [et al.]{} 2007, , 471, L1 , J. M. [et al.]{} 2011, ArXiv e-prints , T. [et al.]{} 2007, , 134, 1019 , C. C., & [Tuffs]{}, R. J. 2003, , 410, L21 , C. C. [et al.]{} 2011, , 527, A109+ , A. C. [et al.]{} 2005, , 129, 1331 , P. C., & [Searle]{}, L. 1981, , 95, 105 , J. C., & [Draine]{}, B. T. 2001, , 553, 581 , E. M. [et al.]{} 1999, , 344, 868 , L. K. [et al.]{} 2011, ArXiv e-prints [^1]: [**N**]{}ew [**HE**]{}rshel [**M**]{}ulti-wavelength [**E**]{}xtragalactic [**S**]{}urvey of [**E**]{}dge-on [**S**]{}pirals, a 10.3 hour, OT1, priority 2 program to supplement the GT program on large disks, [**HER**]{}shel [**O**]{}bservations [**E**]{}dge-on [**S**]{}pirals, PI M. Baes.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper firstly, we generalize the concept of codismantlable graphs to hypergraphs and show that some special vertex decomposable hypergraphs are codismantlable. Then we introduce and generalize some combinatorial invariants of graphs such as matching number and induced matching number and present some upper bounds for the regularity of Stanley-Reisner ring of $\Delta_{\mathcal{H}}$ for certain hypergraphs $\mathcal{H}$ in terms of the introduced matching numbers. At last, we generalize the concept of bouquet in graphs to hypergraphs to extend some combinatorial invariants of graphs about disjointness of a set of bouquets. We use these invariants to characterize the projective dimension of Stanley-Reisner ring of special hypergraphs in some sense.' address: - 'Fahimeh Khosh-Ahang, Department of Mathematics, Ilam University, P.O.Box 69315-516, Ilam, Iran.' - 'Somayeh Moradi, Department of Mathematics, Ilam University, P.O.Box 69315-516, Ilam, Iran and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O.Box: 19395-5746, Tehran, Iran.' author: - 'Fahimeh Khosh-Ahang$^*$ and Somayeh Moradi' title: 'Codismantlable hypergraphs, projective dimension and regularity of edge ideal of special hypergraphs' --- Introduction {#introduction .unnumbered} ============ There is a natural correspondence between simplicial complexes and hypergraphs in the way that for a hypergraph $\mathcal{H}$, the faces of the simplicial complex associated to it are the independent sets of vertices of $\mathcal{H}$, i.e. the sets which do not contain any edge of $\mathcal{H}$. This simplicial complex is called the **independence complex** of $\mathcal{H}$ and is denoted by $\Delta_{\mathcal{H}}$. Square-free monomial ideals can be studied using these combinatorial stuffs. Recently, edge ideals of graphs, as the easiest class of squarefree monomial ideals, has been studied by many researchers and some nice characterizations of the algebraic invariants, in terms of data from graphs, have been proved (cf. [@HD], [@Kimura], [@KM], [@Mor], [@VT] and [@Zheng]). Extending the concepts in graphs to hypergraphs and finding more general results in hypergraphs, which will cover all squarefree monomial ideals, is of great interest and in some senses there are generalizations, see for example [@Emt], [@HT1], [@HW], [@MVi] and [@Wood]. In [@BC], the authors introduced a new class of graphs in terms of a codominated vertex, which they call codismantlable and studied some algebraic and combinatorial properties of these graphs. They proved that a $(C_4,C_5)$-free vertex decomposable graph is codismantlable. In this paper, we extend the concept of codismantlability to hypergraphs and gain some generalizations of the results for graphs in this context. The **Castelnuovo-Mumford regularity** (or simply regularity) and the projective dimension of an $R$-module $M$ are respectively defined as $${\mbox{reg}\,}(M) := \max\{j-i |\ \beta_{i,j}(M)\neq 0\},$$ and $${\mbox{pd}\,}(M) := \max\{i |\ \beta_{i,j}(M)\neq 0 \ {\mathrm}{for \ some \ } j\},$$ where $\beta_{i,j}$ is the $(i,j)$-th Betti number of $M$. Explaining the Castelnuovo-Mumford regularity and the projective dimension of $R/I_{\Delta_{\mathcal{H}}}$ in terms of invariants of $\mathcal{H}$ has been subject of many works. In [@HT1], [@KM] and [@VT], respectively for chordal graphs, $C_5$-free vertex decomposable graphs and sequentially Cohen-Macaulay bipartite graphs $G$, it was shown that ${\mbox{reg}\,}(R/I(G))=c_G$, where $I(G)$ is the edge ideal of $G$ and $c_G$ is the induced matching number of $G$. There has been much activity surrounding combinatorial characterizations of the projective dimension and the Castelnuovo-Mumford regularity of the edge ideal of hypergraphs. In [@HW], the authors defined the concept of a 2-collage in a simple hypergraph as a generalization of the matching number in graphs and proved that the Castelnuovo-Mumford regularity of the edge ideal of a simple hypergraph is bounded above by a multiple of the minimum size of a 2-collage. Morey and Villarreal [@MVi], gave a lower bound for the regularity of the edge ideal of any simple hypergraph in terms of an induced matching of the hypergraph. Moreover, in [@HT1], for $d$-uniform properly-connected hypergraphs a lower bound for the regularity is given. For more results see [@C; @DS; @DHS; @FV; @LM; @N; @WW]. In this paper, we also study the regularity and projective dimension of the Stanley-Reisner ring of $\Delta_{\mathcal{H}}$ for some families of hypergraphs and relate them to some combinatorial concepts and generalize or improve some results, which had been gained, for graphs such as [@HT1 Theorem 6.7] and Theorem 1.9, Propositions 2.6 and 2.7 and Corollary 2.9 in [@KM]. The paper proceeds as follows. In Section 1, we study codismantlable hypergraphs. In fact, in Definition \[codis\], we generalize codominated vertex concept to hypergraphs. We show that under some circumstances a vertex $x$ of a hypergraph is a codominated vertex if and only if it is a shedding vertex (see Theorem \[main\]). By using this fact, in Corollary \[coro\], we conclude that a $C_5$-free vertex decomposable hypergraph, which all of its $3$-cycles have edges of cardinality just two, is codismantlable. Hence, we generalize Lemma 2.6 in [@BC] and Lemma 2.3 in [@KM]. In the light of [@MVi Corollary 3.9(a)], $c_\mathcal{H}$ is a lower bound for ${\mbox{reg}\,}(R/I_{\Delta_{\mathcal{H}}})$, when $\mathcal{H}$ is a hypergraph. In Section 2, we are going to obtain some upper bounds for ${\mbox{reg}\,}(R/I_{\Delta_{\mathcal{H}}})$ for a hypergraph $\mathcal{H}$. We define an induced matching, a semi induced matching and matching number for a hypergraph $\mathcal{H}$, which we denote by $c_{\mathcal{H}}$, $c'_{\mathcal{H}}$ and $m_{\mathcal{H}}$, respectively and compare them together under different conditions. Mainly, we show that in graphs $c_{\mathcal{H}}$ and $c'_{\mathcal{H}}$ coincide, but in general it is not true. Moreover, we show that under the special conditions on $\mathcal{H}$, ${\mbox{reg}\,}(R/I_{\Delta_{\mathcal{H}}})\leq m_{\mathcal{H}}$. As another class of hypergraphs, vertex decomposable hypergraphs has been studied and in Theorem \[reg\], it is proved that if a vertex decomposable hypergraph $\mathcal{H}$ is $(C_2,C_5)$-free, then ${\mbox{reg}\,}(R/I_{\Delta_{\mathcal{H}}})\leq c'_{\mathcal{H}}\leq {\mbox{dim}\,}\Delta_{\mathcal{H}}+1$. This improves a result on graphs proved in [@KM], which states that for a $C_5$-free vertex decomposable graph $G$, ${\mbox{reg}\,}(R/I(G))=c_G$. Finally, in Section 3, we study the projective dimension of Stanley-Reisner ring of certain hypergraphs. To this aim, by generalizing the concept of bouquet from graphs to hypergraphs, we introduce some invariants such as $d_\mathcal{H}$ and $d'_\mathcal{H}$ for a hypergraph $\mathcal{H}$. We show that $c_\mathcal{H}\leq d_\mathcal{H}\leq d'_\mathcal{H}$ and for vertex decomposable hypergraph $\mathcal{H}$, $d'_\mathcal{H}$ is an upper bound for ${\mbox{pd}\,}(R/I_{\Delta_{\mathcal{H}}})$. Also, we find special circumstances in Theorem \[final\] under which, $d'_\mathcal{H}$ precisely characterizes ${\mathrm}{bigheight}(I(\mathcal{H}))$ and ${\mbox{pd}\,}(R/I_{\Delta_{\mathcal{H}}})$. In fact in this section we generalize Propositions 2.6, 2.7 and Corollary 2.9 in [@KM] in some sense. Codominated vertex, shedding vertex and codismantlable hypergraphs ================================================================== Throughout this paper, we assume that $\mathcal{H}=(V(\mathcal{H}), \mathcal{E}(\mathcal{H}))$ is a simple hypergraph with vertex set $V(\mathcal{H})$ and edge set $\mathcal{E}(\mathcal{H})$ whose elements are a family of subsets of $V(\mathcal{H})$ such that no element of $\mathcal{E}(\mathcal{H})$ contains another. Also, for any vertex $x\in V(\mathcal{H})$, $\mathcal{H}\setminus x$ is a hypergraph with vertex set $V(\mathcal{H})\setminus\{x\}$ and edge set $\{E\in \mathcal{E}(\mathcal{H}):\ x\notin E\}$. Moreover $\mathcal{H}/x$ is a hypergraph with vertex set $V(\mathcal{H})\setminus\{x\}$ whose edges are the minimal elements (with respect to inclusion) of the set $\{E\setminus \{x\} \ : \ E\in \mathcal{E}(\mathcal{H})\}$. It is clear that $\mathcal{H}\setminus x$ and $\mathcal{H}/x$ are two simple hypergraphs. They are called **deletion** and **contraction** of $\mathcal{H}$ by $x$, respectively. Note that for a vertex $x\in V(\mathcal{H})$, ${\mbox{del}\,}_{\Delta_\mathcal{H}}(x)=\Delta_{\mathcal{H}\setminus x}$ and ${\mbox{lk}\,}_{\Delta_{\mathcal{H}}}(x)=\Delta_{\mathcal{H}/x}$. *Let $\Delta$ be a simplicial complex on the vertex set $V = \{x_1,\ldots, x_n\}$. Then $\Delta$ is **vertex decomposable** if either:* 1\) The only facet of $\Delta$ is $\{x_1,\ldots, x_n\}$, or $\Delta=\emptyset$. 2\) There exists a vertex $x\in V$ such that ${\mbox{del}\,}_{\Delta}(x)$ and ${\mbox{lk}\,}_{\Delta}(x)$ are vertex decomposable, and such that every facet of ${\mbox{del}\,}_{\Delta}(x)$ is a facet of $\Delta$. A vertex $x\in V$ for which every facet of ${\mbox{del}\,}_{\Delta}(x)$ is a facet of $\Delta$ is called a **shedding vertex** of $\Delta$. Note that this is equivalent to say that no facet of ${\mbox{lk}\,}_{\Delta}(x)$ is a facet of ${\mbox{del}\,}_{\Delta}(x)$. A hypergraph $\mathcal{H}$ is called vertex decomposable, if the independence complex $\Delta_{\mathcal{H}}$ is vertex decomposable and a vertex of $\mathcal{H}$ is called a shedding vertex if it is a shedding vertex of $\Delta_{\mathcal{H}}$. It is easily seen that if $x$ is a shedding vertex of $\mathcal{H}$ and $\{E_1, \dots, E_k\}$ is the set of all edges of $\mathcal{H}$ containing $x$, then every facet of $\mathcal{H}\setminus x$ contains $E_i\setminus \{x\}$ for some $1\leq i\leq k$. Also, recall that a vertex $x$ of a graph $G$ is called codominated if there is a vertex $y$ in $G$ such that $N_G[y]\subseteq N_G[x]$, where $N_G[x]$ is the set of vertices of $G$ consisting $x$ and all neighbors of $x$. Now, we extend the notion of codominated vertex to hypergraphs. \[codis\] Let $x,y$ be two distinct vertices of the hypergraph $\mathcal{H}$. We set $$N_{\mathcal{H}}(x\setminus y)=\{E\setminus \{x\} \ : \ E\in \mathcal{E}(\mathcal{H}), x\in E, y\not\in E\}.$$ Also, assume that $x\in V(\mathcal{H})$ and $\{E_1, \dots, E_k\}$ is the set of all edges of $\mathcal{H}$ containing $x$. Then the vertex $x$ is called a **codominated** vertex if there is an integer $1\leq i \leq k$ such that for each vertex $y\in E_i\setminus \{x\}$ we have $$N_{\mathcal{H}}(y\setminus x)\subseteq N_{\mathcal{H}}(x\setminus y).$$ Note that this definition is a natural generalization of one in graph theory. Now, we are going to find relations between two concepts of shedding vertex and codominated vertex. The following lemma shows that in general every codominated vertex is a shedding vertex. (Compare [@BC Lemma 2.6].)\[codominated\] In a hypergraph $\mathcal{H}$, every codominated vertex is a shedding vertex. Assume that $x$ is a codominated vertex and $\{E_1, \dots, E_k\}$ is the set of all edges of $\mathcal{H}$ containing $x$. Then there is an integer $1\leq i \leq k$ such that for each vertex $y\in E_i\setminus \{x\}$ we have $$N_{\mathcal{H}}(y\setminus x)\subseteq N_{\mathcal{H}}(x\setminus y).$$ Assume that $S$ is a facet of $\Delta_{\mathcal{H}/x}$. Then there is a vertex $y\in (E_i\setminus\{x\})\setminus S$. Now, if we show that $S\cup \{y\}$ is a face of $\Delta_{\mathcal{H}\setminus x}$, the proof will be done. To this end, suppose in contrary that there is an edge $E'$ of $\mathcal{H}$ such that $x\not\in E'$ and $E'$ is contained in $S\cup \{y\}$. Since $S$ is a facet of $\Delta_{\mathcal{H}/x}$, we should have $E'\setminus\{y\}\in N_{\mathcal{H}}(y\setminus x)$. So $E'\setminus\{y\}\in N_{\mathcal{H}}(x\setminus y)$ which means that $(E'\setminus\{y\})\cup\{x\}\in \mathcal{E}(\mathcal{H})$. Therefore $E'\setminus\{y\}\in \mathcal{E}(\mathcal{H}/x)$. But $E'\setminus\{y\}\subseteq S$ which is a contradiction. Hence $S\cup \{y\}$ is a face of $\Delta_{\mathcal{H}\setminus x}$ as required. In the following definition we recall the concept of a cycle hypergraph from [@Berge]. A closed chain is called a **cycle**. More precisely, assume that $\mathcal{H}$ is a hypergraph, $x_1, \dots, x_n$ are distinct vertices of $\mathcal{H}$ and $E_1, \dots, E_n$ are distinct edges of $\mathcal{H}$ such that $x_1, x_n\in E_n$ and $x_i,x_{i+1}\in E_i$, for each $1\leq i\leq n-1$. Then we call $x_1-E_1-x_2-E_2-\dots -x_n-E_n-x_1$ or briefly $E_1-\dots -E_n$ a cycle of length $n$ or an $n$-cycle and we denote it by $C_n$. We say that $\mathcal{H}$ is $C_n$-free if it doesn’t contain any cycle $C_n$ as a subhypergraph. In [@KM Lemma 2.3], it is shown that in a $C_5$-free graph, any shedding vertex is codominated. In the following we generalize this result to hypergraphs. \[shedding\] Assume that $\mathcal{H}$ is a $C_5$-free hypergraph which all of its $3$-cycles have edges of cardinality just two. Then every shedding vertex is a codominated vertex. Assume that $x$ is a shedding vertex and $\{E_1, \dots, E_k\}$ is the set of all edges containing $x$. Moreover, assume, in contrary, that for each $1\leq i \leq k$, there is $y_{i}\in E_i\setminus\{x\}$ and an edge $E_i'$ of $\mathcal{H}$ such that $y_i\in E_i'$, $x\not\in E_i'$ and $E_i'\setminus \{y_i\}\in N_{\mathcal{H}}(y_i\setminus x)\setminus N_{\mathcal{H}}(x\setminus y_i)$. So, $(E_i'\setminus \{y_i\})\cup \{x\}$ is not an edge of $\mathcal{H}$. Note that for two distinct integers $1\leq i,j\leq k$, if $y_i=y_j$, we can choose $E_i'=E_j'$. Also, if $y_i\neq y_j$, then we should have $\{y_i,y_j\}\not\subseteq E_i'$ and $\{y_i,y_j\}\not\subseteq E_j'$. Because otherwise, if for instance $\{y_i,y_j\}\subseteq E_i'$, then $x - E_i - y_i -E_i' - y_j-E_j - x$ is a cycle which $|E_i'|>2$ or $|E_j|>2$ which contradicts with our assumption (note that if $|E_i'|=|E_j|=2$, then we should have $E_i'=\{y_i, y_j\}$ and $E_j=\{x,y_j\}$ and hence, ($E_i'\setminus \{y_i\})\cup \{x\}=E_j$, which is an edge of $\mathcal{H}$ and contradicts with the choice of $E_i'$). Set $S=\bigcup_{i=1}^k (E_i'\setminus \{y_i\})$. Since $\mathcal{H}$ is $C_5$-free, $S$ is an independent set of vertices in $\mathcal{H}/x$. To be more precise, if $S$ is not an independent set of vertices in $\mathcal{H}/x$, there is an edge $E\setminus \{x\}$ of $\mathcal{H}/x$ contained in $S$. So, there are distinct edges $E_i'$ and $E_j'$ with $1\leq i,j \leq k$ and distinct vertices $z,w\in E\setminus \{x\}$ such that $z\in E\cap (E_i'\setminus\{y_i\})$ and $w\in E\cap (E_j'\setminus\{y_i\})$ (otherwise $E\setminus \{x\}\subseteq E_i'\setminus\{y_i\}$ for some $1\leq i\leq k$. If $x\notin E$, then $E=E\setminus \{x\}\subseteq E_i'\setminus\{y_i\} \subset E'_i$, which is impossible. If $x\in E$, then $E=E_j$ for some $1\leq j\leq k$ and $y_j\in E_j\setminus \{x\}=E\setminus \{x\}\subseteq E'_i\setminus \{y_i\}$. Thus $i\neq j$ and $\{y_i,y_j\}\subseteq E'_i$, which is a contradiction). Therefore, one can easily check that $w - E - z - E_i' - y_i - E_i - x - E_j - y_j - E_j' - w$ forms a cycle of length five in $\mathcal{H}$ which is a contradiction. Now, we extend $S$ to a facet $F$ of $\Delta _{\mathcal{H}/x}$. $F$ is a face of $\Delta _{\mathcal{H}\setminus x}$ and so it is contained in a facet $G$ of $\Delta _{\mathcal{H}\setminus x}$. Now, since $x$ is a shedding vertex, $G$ should contain $E_i\setminus \{x\}$ for some $1\leq i\leq k$. Hence, $E'_i\subseteq G$ which contradicts to the fact that $G$ is a facet of $\Delta _{\mathcal{H}\setminus x}$. Now, by means of Lemmas \[codominated\] and \[shedding\], we have the following result, which is one of our main results of this paper. \[main\] Assume that $\mathcal{H}$ is a $C_5$-free hypergraph which all of its $3$-cycles have edges of cardinality just two. Then a vertex $x$ of $\mathcal{H}$ is a shedding vertex if and only if it is a codominated vertex. (Compare [@BC Theorem 2.7] and [@KM Lemma 2.3].) In a $C_5$-free graph, a vertex is a shedding vertex if and only if it is a codominated vertex. In the following, we generalize the concept of codismantlability to hypergraphs. (Compare [@BC Definition 2.2].) Given two hypergraphs $\mathcal{G}$ and $\mathcal{H}$, we say that $\mathcal{G}$ is codismantlable to $\mathcal{H}$ if there exist hypergraphs $\mathcal{H}_0, \mathcal{H}_1, \dots, \mathcal{H}_{k+1}$ satisfying $\mathcal{G}\cong \mathcal{H}_0$, $\mathcal{H}\cong \mathcal{H}_{k+1}$ and $\mathcal{H}_{i+1}=\mathcal{H}_i\setminus x_i$, for each $0\leq i \leq k$, where $x_i$ is a codominated vertex of $\mathcal{H}_i$. A hypergraph $\mathcal{H}$ is called **codismantlable** if either it is an edgeless hypergraph or it is codismantlable to an edgeless hypergraph. The next corollary, which is a generalization of Corollary 2.9 in [@BC], illustrates that certain vertex decomposable hypergraphs are codismantlable. \[coro\] Assume that $\mathcal{H}$ is a $C_5$-free vertex decomposable hypergraph which all of its $3$-cycles have edges of cardinality just two. Then $\mathcal{H}$ is codismantlable. The result can be easily gained by an induction and Theorem \[main\]. So, in graph theory, we have the following remarkable result. (Compare [@BC Corollary 2.9].) Every $C_5$-free vertex decomposable graph is codismantlable. Regularity of edge ideal of certain hypergraphs =============================================== In this section, first inspiring the definition of an induced matching in [@MVi], we introduce the concepts of induced matching number and semi induced matching number of a hypergraph. Then we give some equality and inequalities between these invariants and finally we show that for a hypergraph $\mathcal{H}$, these invariants give bounds for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$ and for some families of hypergraphs we give the precise amount of ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$ in terms of these numbers. A set $\{E_1, \dots, E_k\}$ of edges of a hypergraph $\mathcal{H}$ is called a **semi induced matching** if the only edges contained in $\bigcup_{\ell=1}^kE_\ell$ are $E_1, \dots, E_k$. A semi induced matching which all of its elements are mutually disjoint is called an **induced matching**. Also, we set $$c_{\mathcal{H}}:=\max\{|\bigcup_{\ell=1}^kE_\ell|-k \ : \ \{E_1, \dots, E_k\} \ {\mathrm}{is \ an \ induced \ matching \ in \ }\mathcal{H}\},$$ $$c'_{\mathcal{H}}:=\max\{|\bigcup_{\ell=1}^kE_\ell|-k \ : \ \{E_1, \dots, E_k\} \ {\mathrm}{is \ a \ semi \ induced \ matching \ in \ }\mathcal{H}\},$$ and we call them **induced matching number** and **semi induced matching number** of $\mathcal{H}$, respectively. The following lemma compares the invariants $c_{\mathcal{H}}$, $c'_{\mathcal{H}}$ and ${\mathrm}{dim}(\Delta_{\mathcal{H}})$ for an arbitrary hypergraph $\mathcal{H}$. \[dim\] For any hypergraph $\mathcal{H}$, we have the following inequalities. $$c_{\mathcal{H}}\leq c'_{\mathcal{H}}\leq {\mathrm}{dim}(\Delta_{\mathcal{H}})+1$$ It is clear that every induced matching of $\mathcal{H}$ is a semi induced matching. So, we have $c_{\mathcal{H}}\leq c'_{\mathcal{H}}$. To prove the last inequality, suppose that $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}$ such that $c'_{\mathcal{H}}=|\bigcup_{\ell=1}^kE_\ell|-k$. Set $S_0=\emptyset$ and for each $1\leq i\leq k$, if $E_i\cap S_{i-1}\neq \emptyset$, then set $S_i=S_{i-1}$; else, choose a vertex $x_i\in E_i$ and set $S_i=S_{i-1}\cup \{x_i\}$. Now, consider the set $G=(\bigcup_{\ell=1}^kE_\ell)\setminus S_k$. We claim that $G$ is an independent set of vertices in $\mathcal{H}$. By contrary, assume that $E\subseteq G$ for some $E\in \mathcal{E}(\mathcal{H})$. Then $E\cap S_k=\emptyset$ and $E\subseteq \bigcup_{\ell=1}^kE_\ell$. So $E=E_i$ for some $1\leq i\leq k$, since $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}$. From the choice of $x_i$’s, it is clear that $x_j\in E_i\cap S_k$ for some $1\leq j\leq i$, which is a contradiction. Therefore, $G$ is contained in a facet $F$ of $\Delta_{\mathcal{H}}$. Since $|S_k|\leq k$, we have $c'_{\mathcal{H}}\leq |G|\leq |F|\leq {\mathrm}{dim}(\Delta_{\mathcal{H}})+1$, which completes the proof. The following example illustrate that the equalities in Lemma \[dim\] can be strict. \[ex\] Let $\mathcal{H}$ be a hypergraph with vertex set $V=\{x_1, \dots, x_6\}$ and edges $E_1=\{x_1, x_2, x_3\}, E_2=\{x_2, x_3, x_4\}$ and $E_3=\{x_4, x_5, x_6\}$. Then one can see that $c_{\mathcal{H}}=2$ and $c'_{\mathcal{H}}=3$. So $c_{\mathcal{H}}<c'_{\mathcal{H}}$. Assume that $G$ is a star graph with vertex set $V=\{x_1, \dots, x_4\}$ and edges $\{x_1, x_2\}, \{x_1, x_3\}, \{x_1, x_4\}$. Then one can easily see that $c'_G=1$, but ${\mathrm}{dim}(\Delta_G)=2$. So, even when $\mathcal{H}$ is a graph, the second equality in Lemma \[dim\] can be strict. It is easily seen that when $\mathcal{H}$ is a graph, $c_{\mathcal{H}}$ is the well-known induced matching number of $\mathcal{H}$; i.e. the maximum number of $3$-disjoint edges in $\mathcal{H}$. But the following theorem shows more than this in graphs. \[graph\] If $G$ is a simple graph, then we have $c_G=c'_G$. It is obvious that if $\{E_1, \dots, E_k\}$ is a set of pairwise 3-disjoint edges in $G$ such that $c_G=k$, then it is also a semi induced matching in $G$ and $|\bigcup_{j=1}^kE_j|-k=c_G$. So, $c'_G\geq c_G$. Therefore, it is enough to prove $c'_G\leq c_G$. To this end, it is sufficient to show that for each semi induced matching $\{E_1, \dots, E_k\}$ in $G$, there is a subset $S$ of $\{1, \dots, k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is an induced matching in $G$ and we have $$|\bigcup_{\ell=1}^kE_\ell|-k \leq |\bigcup_{\ell\in S}E_\ell|-|S|.$$ We proceed by induction on $k$. The case $k=1$ is obvious. Inductively suppose the result has been proved for smaller values of $k$. The following three cases should be considered. **Case I.** Suppose that there is an integer $1\leq i\leq k$ such that $E_i\cap (\bigcup_{\ell =1,\ell\neq i}^k E_\ell)=\emptyset$. Then $\{E_\ell \ : \ 1\leq \ell \leq k, \ell \neq i\}$ is a semi induced matching in $G$. So, by inductive hypothesis, there is a subset $S$ of $\{1, \dots, i-1, i+1, \dots , k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is an induced matching in $G$ and we have $$|\bigcup_{\ell=1, \ell\neq i}^kE_\ell|-(k-1) \leq |\bigcup_{\ell\in S}E_\ell|-|S|.$$ Now, set $S'=S\cup \{i\}$. We claim that $\{E_\ell \ : \ \ell\in S'\}$ is an induced matching in $G$. Clearly, the edges are mutually disjoint. Let $E$ be an edge of $G$ such that $E\subseteq \bigcup_{\ell\in S'}E_\ell$. Then $E\subseteq \bigcup_{\ell=1}^k E_\ell$ and thus $E=E_j$ for some $1\leq j\leq k$. By contrary if $j\notin S'$, then $E_j\nsubseteq \bigcup_{\ell\in S}E_\ell$ and $E_j\neq E_i$. Therefore $E_j\cap E_i\neq \emptyset$. So $E_i\cap (\bigcup_{\ell=1, \ell\neq i}^kE_\ell)\neq \emptyset$, which contradicts to our assumption. So the claim is proved. Thus we have $$\begin{aligned} |\bigcup_{\ell=1}^kE_\ell|-k&=|\bigcup_{\ell=1, \ell\neq i}^kE_\ell|-(k-1)+ |E_i|-1 \\ &\leq |\bigcup_{\ell\in S}E_\ell|-|S|+ |E_i|-1 \\ &=|\bigcup_{\ell\in S'}E_\ell|-|S'|.\end{aligned}$$ **Case II.** Suppose that there is an integer $1\leq i\leq k$ such that $|E_i\cap (\bigcup_{\ell =1,\ell\neq i}^k E_\ell)|=1$. Then $\{E_\ell \ : \ 1\leq \ell \leq k, \ell \neq i\}$ is a semi induced matching in $G$. So, by inductive hypothesis, there is a subset $S$ of $\{1, \dots, i-1, i+1, \dots , k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is an induced matching in $G$ and we have $$|\bigcup_{\ell=1, \ell\neq i}^kE_\ell|-(k-1) \leq |\bigcup_{\ell\in S}E_\ell|-|S|.$$ So, we have $$\begin{aligned} |\bigcup_{\ell=1}^kE_\ell|-k&=|\bigcup_{\ell=1, \ell\neq i}^kE_\ell|-(k-1)\\ &\leq |\bigcup_{\ell\in S}E_\ell|-|S|.\end{aligned}$$ **Case III.** Suppose that for each $1\leq i\leq k$, $E_i\subseteq \bigcup_{\ell=1, \ell\neq i}^kE_\ell$. Let $\bigcup_{\ell=1}^kE_\ell=\{x_1,\ldots,x_t\}$ and $T$ be a subgraph of $G$ with vertex set $\{x_1,\ldots,x_t\}$ and edge set $\{E_1,\ldots,E_k\}$. From the assumption it is easy to see that $\deg_T(x_i)\geq 2$ for any $1\leq i\leq t$. Thus $2t\leq \sum_{i=1}^t \deg_T(x_i)=2|E(T)|=2k$. So $t\leq k$. This means that $|\bigcup_{\ell=1}^kE_\ell|\leq k$ in this case. So, if we set $S=\{1\}$, then clearly, $$|\bigcup_{\ell=1}^k E_\ell|-k \leq|\bigcup_{\ell\in S}E_\ell|-|S|.$$ The proof is complete. The concept of matching number of a hypergraph is known as a generalization of one in graph theory (see [@Berge]). H$\grave{a}$ and Van Tuyl in [@HT1] showed that when $\mathcal{H}$ is a graph, its matching number is an upper bound for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$. Here, by benefitting their work, we improve the definition of matching number of a hypergraph so that we can generalize this result to special class of hypergraphs. So, we present a new definition for matching number of a hypergraph as follows. (Compare [@HT1 Definition 6.6].) A set of edges of a hypergraph $\mathcal{H}$ is called a **matching** if they are pairwise disjoint. Also, we set $$m_{\mathcal{H}}:=\max\{|\bigcup_{\ell=1}^kE_\ell|-k \ : \ \{E_1, \dots, E_k\} \ {\mathrm}{is \ a \ matching \ in \ }\mathcal{H}\},$$ and we call it the **matching number** of $\mathcal{H}$. One can see that this definition is a natural generalization of one in graph theory, i.e. when $\mathcal{H}$ is a graph, $m_{\mathcal{H}}$ is the largest size of a maximal matching in $\mathcal{H}$. Furthermore, it is obvious that $c_{\mathcal{H}}\leq m_{\mathcal{H}}$ for any hypergraph $\mathcal{H}$. Although, at one look, no relation can be seen between $c'_{\mathcal{H}}$ and $m_{\mathcal{H}}$, but Theorem \[graph\] shows that $c'_G\leq m_G$, for any simple graph $G$. Now, we generalize this inequality for special class of hypergraphs consisting simple graphs. Recall that a hypergraph is called **$d$-uniform** if every edge of $\mathcal{H}$ has the same cardinality $d$. So, every simple graph is a $2$-uniform hypergraph. Note that the mentioned condition in the following proposition is different from the property of *strongly connected* for hypergraphs. Recall that a $d$-uniform hypergraph $\mathcal{H}$ is called strongly connected if for each distinct edges $E$ and $E'$, there is a chain $E=E_0, E_1, \dots, E_{k-1}, E_k=E'$ of edges of $\mathcal{H}$ such that for each $i:=0, 1, \dots, k-1$, $|E_i\cap E_{i+1}|=d-1$. \[MH\] Assume that $\mathcal{H}$ is a $d$-uniform hypergraph such that for each distinct edges $E$ and $E'$, $E\cap E'\neq \emptyset$ implies that $|E\cap E'|=d-1$. Then $$c_{\mathcal{H}}\leq c'_{\mathcal{H}}\leq m_{\mathcal{H}}.$$ It is enough to prove the second inequality. To this end, it is sufficient to show that for each set $\{E_1, \dots, E_k\}$ of edges of $\mathcal{H}$, there is a subset $S$ of $\{1, \dots, k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is a matching of $\mathcal{H}$ and $$|\bigcup_{\ell=1}^kE_\ell|-k \leq|\bigcup_{\ell\in S}E_\ell|-|S|.$$ We use induction on $k$. The result is clear when $k=1$. So assume inductively that $k>1$ and the result is true for smaller values of $k$. We may consider the following cases. **Case I.** Suppose that there is an integer $1\leq i\leq k$ such that $E_i\cap (\bigcup_{\ell =1,\ell\neq i}^k E_\ell)=\emptyset$. Then by inductive hypothesis, there is a subset $S$ of $\{1, \dots, i-1, i+1, \dots , k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is a matching in $G$ and we have $$|\bigcup_{\ell=1, \ell\neq i}^kE_\ell|-(k-1) \leq |\bigcup_{\ell\in S}E_\ell|-|S|.$$ Now, set $S'=S\cup \{i\}$. It is obvious that $\{E_\ell \ : \ \ell\in S'\}$ is a matching of $\mathcal{H}$ and we have $$\begin{aligned} |\bigcup_{\ell=1}^kE_\ell|-k&=|\bigcup_{\ell=1,\ell\neq i}^k E_\ell|-(k-1)+|E_i|-1\\ &\leq|\bigcup_{\ell\in S}E_\ell|-|S|+|E_i|-1\\ &=|\bigcup_{\ell\in S'}E_\ell|-|S'|\end{aligned}$$ as desired. **Case II.** Suppose that there is an integer $1\leq i\leq k$ such that $0<|E_i\cap (\bigcup_{\ell=1, \ell\neq i}^kE_\ell)|< |E_i|.$ Then inductive hypothesis implies that there is a subset $S$ of $\{1, \dots, i-1, i+1, \dots , k\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is a matching of $\mathcal{H}$ and $$|\bigcup_{\ell=1,\ell\neq i}^kE_\ell|-(k-1) \leq|\bigcup_{\ell\in S}E_\ell|-|S|.$$ On the other hand, by our assumption on $\mathcal{H}$, we should have $|E_i\cap (\bigcup_{\ell=1,\ell\neq i}^kE_\ell)|=d-1$. Now, we have $$\begin{aligned} |\bigcup_{\ell=1}^kE_\ell|-k&=|\bigcup_{\ell=1,\ell\neq i}^kE_\ell|-(k-1)+|E_i|-|E_i\cap (\bigcup_{\ell=1, \ell\neq i}^k E_\ell)|-1\\ &\leq|\bigcup_{\ell\in S}E_\ell|-|S|+d-(d-1)-1\\ &=|\bigcup_{\ell\in S}E_\ell|-|S|\end{aligned}$$ as desired. **Case III.** Suppose that for each $1\leq i\leq k$, $E_i\subseteq \bigcup_{\ell=1, \ell\neq i}^kE_\ell$. Then by inductive hypothesis, there is a subset $S$ of $\{1, \dots, k-1\}$ such that $\{E_\ell \ : \ \ell\in S\}$ is a matching of $\mathcal{H}$ and $$|\bigcup_{\ell=1}^{k-1}E_\ell|-(k-1) \leq|\bigcup_{\ell\in S}E_\ell|-|S|.$$ So, we have $$\begin{aligned} |\bigcup_{\ell=1}^kE_\ell|-k&=|\bigcup_{\ell=1}^{k-1}E_\ell|-(k-1)-1\\ &\leq|\bigcup_{\ell\in S}E_\ell|-|S|-1\\ &\leq|\bigcup_{\ell\in S}E_\ell|-|S|\end{aligned}$$ as desired. \[mH\] Morey and Villarreal in [@MVi] showed that $c_{\mathcal{H}}$ is a lower bound for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$ for a simple hypergraph $\mathcal{H}$. Hereafter, we are trying to find circumstances under which $c'_{\mathcal{H}}$ or $m_{\mathcal{H}}$ is an upper bound for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$. Note that in the light of [@HT1 Theorem 6.7], $m_{\mathcal{H}}$ is an upper bound for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$, where $\mathcal{H}$ is a simple graph. But we may have this result for more hypergraphs. In this regard, recall that a subset $C$ of the edges of a hypergraph $\mathcal{H}$ is called a 2-collage for $\mathcal{H}$ if for each edge $E$ of $\mathcal{H}$ we can delete a vertex $v$ so that $E\setminus \{v\}$ is contained in some edge of $C$. Hence if $\mathcal{H}$ is a $d$-uniform hypergraph such that for each distinct edges $E$ and $E'$, $E\cap E'\neq \emptyset$ implies that $|E\cap E'|=d-1$, one can easily see that any maximal matching in $\mathcal{H}$ is a 2-collage. So, in view of [@MVi Corollary 3.9(a)] and [@HW Theorem 1.2], one can have $$c_{\mathcal{H}}\leq {\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})\leq m_{\mathcal{H}}.$$ As another main result of this paper, we are going to show that $c'_{\mathcal{H}}$ is an upper bound for ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})$ for a certain class of hypergraphs. To this end, we need to illustrate the relations between $c'_{\mathcal{H}}$, $c'_{\mathcal{H}\setminus x}$ and $c'_{\mathcal{H}/x}$ for a vertex $x$ of $\mathcal{H}$. Note that it is obvious that $c_{\mathcal{H}\setminus x}\leq c_{\mathcal{H}}$ and $c'_{\mathcal{H}\setminus x}\leq c'_{\mathcal{H}}$. Now, suppose that $\{E_1\setminus\{x\}, \dots, E_k\setminus\{x\}\}$ is a semi induced matching in $\mathcal{H}/x$ such that $c'_{\mathcal{H}/x}=|\bigcup_{\ell=1}^k (E_\ell\setminus\{x\})|-k$. The following example shows that it is not necessarily true that $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}$. Let $\mathcal{H}$ be a hypergraph with $V(\mathcal{H})=\{x_1, \dots, x_5\}$ and $\mathcal{E}(\mathcal{H})=\{E_1=\{x_1, x_2, x_3\}, E_2=\{x_2, x_3, x_4\}, E_3= \{x_4, x_5\}\}$. Then $\mathcal{E}(\mathcal{H}/x_1)=\{E_1\setminus\{x_1\}, E_3\setminus \{x_1\}\}$. It is clear that $\{E_1\setminus\{x_1\}, E_3\setminus \{x_1\}\}$ is a semi induced matching in $\mathcal{H}/x_1$ but $\{E_1, E_3\}$ is not a semi induced matching in $\mathcal{H}$. Now, the following two lemmas provide conditions under which we can get to a semi induced matching in $\mathcal{H}$ from one in $\mathcal{H}/x$, for a vertex $x$ of $\mathcal{H}$. \[1\] Assume that $\mathcal{H}$ is a $C_2$-free hypergraph, $x$ is a vertex of $\mathcal{H}$ and $k$ is the smallest integer such that there exists a semi induced matching $\{E_1\setminus\{x\}, \dots, E_k\setminus\{x\}\}$ in $\mathcal{H}/x$ so that $c'_{\mathcal{H}/x}=|\bigcup_{\ell=1}^k (E_\ell\setminus\{x\})|-k$. Then $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}$ and so if $x\in E_i$ for some $1\leq i\leq k$, we have $c'_{\mathcal{H}/x}+1\leq c'_{\mathcal{H}}$. Suppose that there is an edge $E$ of $\mathcal{H}$ such that $E\subseteq\bigcup_{\ell=1}^k E_\ell$. Then $E\setminus \{x\}\subseteq\bigcup_{\ell=1}^k (E_\ell\setminus \{x\})$. Now, we have three cases: **Case I.** If $x\in E$, then $E\setminus \{x\}=E_i\setminus \{x\}$, for some $1\leq i\leq k$. If $x\not\in E_i$, then $E$ strictly contains $E_i$ which is a contradiction. So, $x\in E_i$ and hence $E=E_i$ as desired. **Case II.** If $x\not\in E$ and $E$ is an edge of $\mathcal{H}/x$, then $E=E_i\setminus \{x\}$, for some $1\leq i\leq k$. If $x\in E_i$, then $E_i$ strictly contains $E$ which is a contradiction. So, $x\not\in E_i$ which implies that $E=E_i$ as desired. **Case III.** If $x\not\in E$ and $E$ is not an edge of $\mathcal{H}/x$, then there is an edge $E'$ of $\mathcal{H}$ containing $x$ such that $E'\setminus \{x\}\subset E$ and $E'\setminus \{x\}$ is an edge of $\mathcal{H}/x$. So, $E\cap E'=E'\setminus\{x\}$. Since $\mathcal{H}$ is $C_2$-free, $|E'\setminus\{x\}|=1$. Since $E'\setminus \{x\}\subseteq \bigcup_{\ell=1}^k (E_\ell\setminus \{x\})$, then $E'\setminus \{x\}=E_i\setminus \{x\}$ for some $1\leq i\leq k$. Thus $|E_i\setminus\{x\}|=1$. Moreover, $E_i\setminus\{x\}\nsubseteq \bigcup_{\ell=1,\ell\neq i}^k (E_\ell\setminus \{x\})$, since otherwise $E_i\setminus\{x\}\subseteq E_j\setminus\{x\}$ for some $j\neq i$, which is impossible. Therefore, $\{E_\ell\setminus \{x\}, 1\leq \ell\leq k, \ell\neq i\}$ is a semi inducing matching in $\mathcal{H}/x$ and $|\bigcup_{\ell=1,\ell\neq i}^k (E_\ell\setminus\{x\})|-(k-1)=|\bigcup_{\ell=1}^k (E_\ell\setminus\{x\})|-1-(k-1)=c'_{\mathcal{H}/x}$, which contradicts to our assumption on $k$. So this case can’t occur. Hence, $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}$. Now, if $x\in E_i$ for some $1\leq i\leq k$, we have $$c'_{\mathcal{H}/x}=|\bigcup_{\ell=1}^k (E_\ell\setminus\{x\})|-k=|\bigcup_{\ell=1}^k E_\ell|-k-1\leq c'_{\mathcal{H}}-1,$$ which completes the proof. \[2\] Assume that $\mathcal{H}$ is a $(C_2, C_5)$-free hypergraph, $x$ is a shedding vertex of $\mathcal{H}$ and $\{E_1\setminus\{x\}, \dots, E_k\setminus\{x\}\}$ is a semi induced matching in $\mathcal{H}/x$ such that $x\not\in E_\ell$ for all $1\leq \ell \leq k$. Then there is an edge $F$ of $\mathcal{H}$ containing $x$ such that $\{E_1, \dots, E_k, F\}$ is a semi induced matching in $\mathcal{H}$. Moreover, $c'_{\mathcal{H}/x}+1\leq c'_{\mathcal{H}}$. Let $\{F_1, \dots, F_s\}$ be the set of all edges containing $x$ and suppose, in contrary, that for each $F_i$, there is an edge $F_i'$ of $\mathcal{H}$ such that $F_i'\not\in \{E_1, \dots, E_k, F_i\}$, $F_i'\cap F_i\neq\emptyset$ and $F_i'\setminus F_i\subseteq \bigcup_{\ell=1}^kE_\ell$. Note that if $F_i'=F_j$ for some $1\leq j\leq s$, then $F_j\setminus F_i\subseteq\bigcup_{\ell=1}^kE_\ell$. Since $\mathcal{H}$ is $C_2$-free, then $F_i\cap F_j=\{x\}$ and $F_j\setminus F_i=F_j\setminus \{x\}$. So $F_j\setminus \{x\}\subseteq\bigcup_{\ell=1}^kE_\ell$. This is a contradiction, since $F_j\setminus \{x\}\in \mathcal{E}(\mathcal{H}/x)$. Also, note that for each distinct integers $1\leq i,j\leq s$, $F_i'\neq F_j'$. Because, if we have $F_i'=F_j'$ for some distinct integers $1\leq i,j\leq s$, then we should have $F_i'\setminus (F_i\cap F_j)\subseteq \bigcup_{\ell=1}^kE_\ell$. On the other hand, we know that $x\not\in F_i'$ and since $\mathcal{H}$ is $C_2$-free, $F_i\cap F_j=\{x\}$. Hence, $F_i'\subseteq \bigcup_{\ell=1}^kE_\ell$. So there exists an edge $E\setminus\{x\}\in \mathcal{E}(\mathcal{H}/x)$ such that $E\setminus\{x\}\subseteq F'_i\subseteq \bigcup_{\ell=1}^kE_\ell$, which is a contradiction. Moreover, note that for each distinct integers $1\leq i,j\leq s$, $F_j\cap F'_j\not\subseteq F'_i\setminus F_i$, because otherwise since $F'_i\setminus F_i$ and $F'_j\setminus F_j$ are contained in $\bigcup_{\ell=1}^k E_\ell$, we should have $F'_j\subseteq \bigcup_{\ell=1}^k E_\ell$, which is a contradiction. Hence, $E_\ell\cap F'_i\neq F_j\cap F'_j$ for all $1\leq \ell\leq k$. Now, set $S=\bigcup_{i=1}^s(F_i'\setminus F_i)$. At first, we are going to show that $S$ is an independent set of vertices in $\mathcal{H}/x$. Suppose, in contrary, that $S$ is not independent. Then, since $S\subseteq \bigcup_{\ell=1}^kE_\ell$ and $\{E_1, \dots, E_k\}$ is a semi induced matching in $\mathcal{H}/x$, there should exist an $E_\ell$ which intersects with two distinct edges $F_i'$ and $F_j'$. So, since $\mathcal{H}$ is $C_2$-free, $E_\ell - F_i' - F_i - F_j - F_j' - E_\ell$ forms a subhypergraph $C_5$ in $\mathcal{H}$ which is a contradiction. Thus, $S$ is an independent set of vertices in $\mathcal{H}/x$. We extend $S$ to a facet $G$ of $\Delta_{\mathcal{H}/x}$. $G$ is also a facet of $\Delta_{\mathcal{H}\setminus x}$; because otherwise $G$ is contained in a facet $K$ of $\Delta_{\mathcal{H}\setminus x}$. Now, since $x$ is a shedding vertex, $K$ contains $F_i\setminus \{x\}$ for some $1\leq i\leq s$. Hence, $F'_i\subseteq K$, because $F'_i\setminus F_i\subseteq S\subseteq G\subseteq K$ and $x\not\in F'_i$. This is a contradiction, since $F'_i\in \mathcal{E}(\mathcal{H} \setminus x)$. So we found a facet of $\Delta_{\mathcal{H}/x}$, which is a facet of $\Delta_{\mathcal{H}\setminus x}$. But this contradicts to the fact that $x$ is a shedding vertex. So, we proved that $\{E_1, \dots, E_k, F_i\}$ is a semi induced matching in $\mathcal{H}$, for some edge $F_i$ containing $x$. Now, let $c'_{\mathcal{H}/x}=|\bigcup_{\ell=1}^k(E_\ell\setminus\{x\})|-k$. Since $F_i\setminus \{x\}\nsubseteq \bigcup_{\ell=1}^kE_\ell$, $c'_{\mathcal{H}/x}+1\leq |(\bigcup_{\ell=1}^kE_\ell)\cup F_i|-(k+1)\leq c'_{\mathcal{H}}$ as required. Now, we are ready to state one of our main results of this section. \[reg\] Let $\mathcal{H}$ be a $(C_2,C_5)$-free vertex decomposable hypergraph.Then $${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})\leq c'_{\mathcal{H}}\leq {\mathrm}{dim}(\Delta_{\mathcal{H}})+1.$$ In the light of Lemma \[dim\] it is enough to prove ${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})\leq c'_{\mathcal{H}}$. In this regard, we use induction on $|V(\mathcal{H})|$. If $|V(\mathcal{H})|=2$, the result is clear. Suppose, inductively, that the result has been proved for smaller values of $|V(\mathcal{H})|$. Assume that $x$ is a shedding vertex of $\mathcal{H}$. Let $\Delta=\Delta_{\mathcal{H}}$, $\Delta_1=\Delta_{\mathcal{H}\setminus x}$ and $\Delta_2=\Delta_{\mathcal{H}/x}$. Then $\mathcal{H}\setminus x$ and $\mathcal{H}/x$ are $(C_2,C_5)$-free vertex decomposable hypergraphs and no facet of $\Delta_2$ is a facet of $\Delta_1$. By inductive hypothesis we have $${\mathrm}{reg}(R/I_{\Delta_1})\leq c'_{\mathcal{H}\setminus x} \ {\mathrm}{and} \ {\mathrm}{reg}(R/I_{\Delta_2})\leq c'_{\mathcal{H}/x}.$$ On the other hand, we have the inequality $${\mathrm}{reg}(R/I_{\Delta})\leq\max\{{\mathrm}{reg}(R/I_{\Delta_1}),{\mathrm}{reg}(R/I_{\Delta_2})+1\}.$$ Hence $${\mathrm}{reg}(R/I_{\Delta})\leq \max\{c'_{\mathcal{H}\setminus x},c'_{\mathcal{H}/x}+1\}.$$ Now, the result immediately follows from Lemmas \[1\] and \[2\]. It is obvious that $c_{\mathcal{H}}=c'_{\mathcal{H}}$ doesn’t necessarily imply that $\mathcal{H}$ is a graph. Now since this equality in conjunction with Theorem \[reg\] can characterize the regularity of the Stanley-Reisner ring $R/I_{\Delta_{\mathcal{H}}}$ precisely, this question arises that when the equality $c_{\mathcal{H}}=c'_{\mathcal{H}}$ holds? In the light of Remark \[mH\], the similar question can be asked about the equality $c_{\mathcal{H}}=m_{\mathcal{H}}$. With this point of view, [@MVi Corollary 3.9(a)], Lemma \[dim\], Proposition \[MH\], Remark \[mH\] and Theorems \[graph\] and \[reg\] imply the next corollary. (Compare [@KM Theorem 1.9].) - Assume that $\mathcal{H}$ is a $(C_2,C_5)$-free vertex decomposable hypergraph such that $c_{\mathcal{H}}={\mathrm}{dim}(\Delta_{\mathcal{H}})+1$. Then $${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})= c_{\mathcal{H}}=c'_{\mathcal{H}}={\mathrm}{dim}(\Delta_{\mathcal{H}})+1.$$ - If $\mathcal{H}$ is a $(C_2,C_5)$-free vertex decomposable hypergraph such that $c_{\mathcal{H}}=c'_{\mathcal{H}}$, then $${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})= c_{\mathcal{H}}.$$ - In particular, if $G$ is a simple $C_5$-free vertex decomposable graph, then $${\mathrm}{reg}(R/I_{\Delta_G})= c_G.$$ - Assume that $\mathcal{H}$ is a $d$-uniform hypergraph such that $c_{\mathcal{H}}=m_{\mathcal{H}}$ and for each distinct edges $E$ and $E'$, $E\cap E'\neq \emptyset$ implies that $|E\cap E'|=d-1$. Then $${\mathrm}{reg}(R/I_{\Delta_{\mathcal{H}}})= c_{\mathcal{H}}=c'_{\mathcal{H}}=m_{\mathcal{H}}.$$ Projective dimension of edge ideal of certain hypergraphs ========================================================= In this section, we are going to characterize or even find some bounds for the projective dimension of edge ideal of special hypergraphs up to our ability. To this end, firstly we generalize some concepts from graphs to hypergraphs so that we generalize some combinatorial invariants of hypergraphs as follows. A hypergraph $\mathcal{H}$ is called a **bouquet** if $\bigcap_{E\in \mathcal{E}(\mathcal{H})}E\neq \emptyset$. In this case, if $\mathcal{H}$ has at least two edges, then all elements in $\bigcap_{E\in \mathcal{E}(\mathcal{H})}E$ are called the **roots** of $\mathcal{H}$, all of its edges are called the **stems** of $\mathcal{H}$ and the elements of $\bigcup_{E\in \mathcal{E}(\mathcal{H})}(E\setminus \bigcap_{E\in \mathcal{E}(\mathcal{H})}E)$ are called the **flowers** of $\mathcal{H}$. When $\mathcal{H}$ has only one edge $E$, any proper subset of $E$ can be considered as roots and its complement as flowers of the bouquet. A subhypergraph of a simple hypergraph $\mathcal{T}$ which is a bouquet is called a bouquet of $\mathcal{T}$. Let $\mathcal{B}=\{B_1,\ldots,B_n\}$ be a set of bouquets of $\mathcal{H}$. We use the following notations. $$\ \ \ F(\mathcal{B})=\{w\in V(\mathcal{H})\ | \ w \text{ is a flower of some bouquet in } \mathcal{B}\}$$ $$R(\mathcal{B})=\{z\in V(\mathcal{H}) \ | \ z \text{ is a root of some bouquet in } \mathcal{B}\}$$ $$S(\mathcal{B})=\{E\in E(\mathcal{H}) \ | \ E \text{ is a stem of some bouquet in } \mathcal{B}\}$$ Kimura in [@Kimura] introduced two notions of disjointness of a set of bouquets in graphs. In the following, we generalize these notions to hypergraphs. (See [@Kimura Definitions 2.1 and 5.1].) A set of bouquets $\mathcal{B}=\{B_1,\ldots,B_n\}$ is called **strongly disjoint** in $\mathcal{H}$ if we can choose a stem $E_i$ from each bouquet $B_i\in \mathcal{B}$ such that $\{E_1,\ldots,E_n\}$ is an induced matching in $\mathcal{H}$. A set of bouquets $\mathcal{B}=\{B_1,\ldots,B_n\}$ is called **semi-strongly disjoint** in $\mathcal{H}$ if $R(\mathcal{B})$ is an independent set of vertices in $\mathcal{H}$. Now, set $$d_\mathcal{H}:=\max\{|F(\mathcal{B})| \ | \ \mathcal{B}\ \text{is a strongly disjoint set of bouquets of } \mathcal{H} \}$$ and $$\ \ \ \ \ \ d'_\mathcal{H}:=\max\{|F(\mathcal{B})| \ | \ \mathcal{B}\ \text{is a semi-strongly disjoint set of bouquets of } \mathcal{H} \}.$$ Note that when $G$ is a graph, to define $d_G$ and $d'_G$, the condition $(i)$ in Definitions 2.1 and 5.1 in [@Kimura] is redundant. Therefore, the above definitions are suitable generalizations to hypergraphs and when $\mathcal{H}$ is a graph, they are coincide to those in graphs defined by Kimura in [@Kimura]. The following result shows that $d_\mathcal{H}$ and $d'_\mathcal{H}$ are comparable with $c_\mathcal{H}$ and $c'_\mathcal{H}$. \[cd\] - For each hypergraph $\mathcal{H}$, we have $$c_\mathcal{H}\leq d_\mathcal{H}\leq d'_\mathcal{H}.$$ Specially when $G$ is a graph, we have $$c_G=c'_G\leq d_G\leq d'_G.$$ - If $\mathcal{H}$ is a $C_2$-free hypergraph, then $$c_\mathcal{H}\leq c'_\mathcal{H}\leq d'_\mathcal{H}.$$ <!-- --> - It is easy to see that if $\{E_1, \dots, E_k\}$ is an induced matching in $\mathcal{H}$, then one can consider it as a set of bouquets with only one stem and one vertex as their roots. So, the number of its flowers equals to $c_\mathcal{H}$. Also, one may see that every strongly disjoint set of bouquets in $\mathcal{H}$ is semi-strongly disjoint. Hence, we have $c_\mathcal{H}\leq d_\mathcal{H}\leq d'_\mathcal{H}$ as desired. The last assertion immediately follows from \[graph\]. - In view of Lemma \[dim\] it is enough to prove the second inequality. In this regard, assume that $\{E_1, \dots, E_k\}$ is a semi-induced matching in $\mathcal{H}$ such that $c'_\mathcal{H}=|\bigcup_{\ell=1}^kE_\ell |-k$. At first, remove all edges $E_j$ which is contained in $\bigcup_{\ell=1,\ell\neq j}^kE_\ell$. Then set $s=0$ and assume that $B_0$ is a bouquet in $\mathcal{H}$ with $\mathcal{E}(B_0)=\emptyset$. Then for $i =1, \dots, k$, if $E_i \cap (\bigcap_{E_\ell\in \mathcal{E}(B_j)}E_\ell)\neq\emptyset$ for some $j<s+1$, then set $\mathcal{E}(B_j):=\mathcal{E}(B_j)\cup \{E_i\}$ (note that if there exist more than one $j<s+1$ such that $E_i \cap (\bigcap_{E_\ell\in \mathcal{E}(B_j)}E_\ell)\neq\emptyset$, then we add $E_i$ to edges of just one of these $B_j$s); else consider $B_{s+1}$ as a bouquet with $\mathcal{E}(B_{s+1})=\{E_i\}$ and set $s:=s+1$. Now, since $\{E_1, \dots, E_k\}$ is a semi-induced matching in $\mathcal{H}$, $\mathcal{B}=\{B_1, \dots, B_s\}$ is a set of semi-strongly disjoint set of bouquets in $\mathcal{H}$. Also, we have $$|F(\mathcal{B})|\geq |(\bigcup_{\ell=1}^kE_\ell)\setminus R(\mathcal{B})|=|\bigcup_{\ell=1}^kE_\ell|-|R(\mathcal{B})|.$$ Now, since $\mathcal{H}$ is $C_2$-free, by considering only one vertex as the root of bouquets with one stem, we have $|R(\mathcal{B})|\leq k$. Hence, $$|F(\mathcal{B})|\geq |\bigcup_{\ell=1}^kE_\ell|- k=c'_\mathcal{H},$$ which implies that $c'_\mathcal{H}\leq d'_\mathcal{H}$ as required. The following corollary is an immediate consequence of Theorem \[reg\] and part (ii) of Proposition \[cd\]. Assume that $\mathcal{H}$ is a $(C_2,C_5)$-free vertex decomposable hypergraph. Then $$c_\mathcal{H}\leq {\mathrm}{reg}(R/I_{\Delta_\mathcal{H}})\leq c'_\mathcal{H}\leq d'_\mathcal{H}.$$ So, if moreover $c_\mathcal{H}=d'_\mathcal{H}$, then $${\mathrm}{reg}(R/I_{\Delta_\mathcal{H}})=c_\mathcal{H}=d'_\mathcal{H}.$$ For our main result of this section, we need the following two lemmas. \[1\] Assume that $\mathcal {H}$ is a hypergraph and $x$ is a shedding vertex of $\mathcal{H}$. Then $d'_{\mathcal{H}/x}\leq d'_\mathcal{H}$. Suppose that $\mathcal{B}=\{B_1,\ldots,B_n\}$ is a set of semi-strongly disjoint set of bouquets of $\mathcal{H}/x$ such that $d'_{\mathcal{H}/x}=|F(\mathcal{B})|$ and for each $1\leq i \leq n$, $S(B_i)=\{E_{i,1}\setminus \{x\}, \dots, E_{i,m_i}\setminus \{x\}\}$. Then if for each $1\leq i \leq n$ we set $B'_i$ as a bouquet in $\mathcal{H}$ with $S(B'_i)=\{E_{i,1}, \dots, E_{i,m_i}\}$, then clearly $\mathcal{B'}=\{B'_1,\ldots,B'_n\}$ is a set of semi-strongly disjoint set of bouquets of $\mathcal{H}$ with $|F(\mathcal{B'})|\geq d'_{\mathcal{H}/x}$. Hence, $d'_{\mathcal{H}/x}\leq d'_\mathcal{H}$ as required. \[2\] Assume that $\mathcal {H}$ is a hypergraph and $x$ is a shedding vertex of $\mathcal{H}$. Then $d'_{\mathcal{H}\setminus x}+1\leq d'_\mathcal{H}$. Suppose that $\mathcal{B}=\{B_1,\ldots,B_n\}$ is a set of semi-strongly disjoint set of bouquets of $\mathcal{H}\setminus x$ such that $d'_{\mathcal{H}\setminus x}=|F(\mathcal{B})|$. Also, suppose that $\{E_1,\ldots,E_k\}$ is the set of all edges of $\mathcal{H}$ containing $x$. Now, we are going to prove that there is an integer $1\leq i\leq k$ such that for each $y_i\in E_i\setminus \{x\}$, $R(\mathcal{B})\cup \{y_i\}$ is an independent set of vertices in $\mathcal{H}$. If we prove this claim, then by choosing this $E_i$ as a bouquet of $\mathcal{H}$ with one stem $E_i$ and some $y_i\in E_i\setminus \{x\}$ as its root, $\mathcal{B'}=\mathcal{B}\cup \{E_i\}$ forms a set of semi-strongly disjoint set of bouquets of $\mathcal{H}$ with $|F(\mathcal{B'})|\geq d'_{H\setminus x}+1$, since $x\in F(E_i)\setminus V(\mathcal{B})$. This completes the proof. To prove the claim suppose in contrary that for each $1\leq i\leq k$, there is a vertex $y_i\in E_i\setminus \{x\}$ and an edge $E'_i$ of $\mathcal{H}$ containing $y_i$ such that $E'_i\setminus \{y_i\}\subseteq R(\mathcal{B})$. Note that if for two distinct integers $1\leq i,j\leq k$, $y_i=y_j$, then one can choose $E'_i=E'_j$. Also, if $y_i\neq y_j$, then we should have $\{y_i,y_j\}\not\subseteq E_i'$ and $\{y_i,y_j\}\not\subseteq E_j'$. Because otherwise, if for instance $\{y_i,y_j\}\subseteq E_i'$, then $y_j\in E_i'\setminus \{y_i\}\subseteq R(\mathcal{B})$. Hence $E'_j\setminus \{y_j\}\subseteq R(\mathcal{B})$ insures that $E'_j\subseteq R(\mathcal{B})$ which contradicts to independence of $R(\mathcal{B})$ in $\mathcal{H}\setminus x$. Now, if we set $S=\bigcup_{i=1}^k (E_i'\setminus \{y_i\})$, then $S$ is an independent set of vertices in $\mathcal{H}/x$. To be more precise, if $S$ is not an independent set of vertices in $\mathcal{H}/x$, there is an edge $E\setminus \{x\}$ of $\mathcal{H}/x$ contained in $S$. If $x\in E$, then $E=E_j$ for some $1\leq j\leq k$. Hence, $y_j\in E\setminus \{x\}\subseteq S$, which implies that $\{y_i, y_j\}\subseteq E'_i$ for some integer $1\leq i\leq k$ with $i\neq j$ which is a contradiction. So, $E$ is an edge of $\mathcal{H}\setminus x$ which is impossible, because $E\subseteq S\subseteq R(\mathcal{B})$ and $R(\mathcal{B})$ is independent in $\mathcal{H}\setminus x$. Now, we extend $S$ to a facet $F$ of $\Delta _{\mathcal{H}/x}$. $F$ is a face of $\Delta _{\mathcal{H}\setminus x}$ and so it is contained in a facet $G$ of $\Delta _{\mathcal{H}\setminus x}$. Now, since $x$ is a shedding vertex, $G$ should contain $E_i\setminus \{x\}$ for some $1\leq i\leq k$. Hence, $E'_i\subseteq G$ which contradicts to the fact that $G$ is a facet of $\Delta _{\mathcal{H}\setminus x}$. This proves our claim and so completes the proof. Now, we are ready to state our main result of this section which is a generalization of Proposition 2.6 in [@KM]. (Compare [@KM Proposition 2.6].)\[pd\] Assume that $\mathcal{H}$ is a vertex decomposable hypergraph. Then $${\mathrm}{pd}(R/I_{\Delta_\mathcal{H}})\leq d'_\mathcal{H}.$$ We proceed by induction on $|V(H)|$. If $|V(\mathcal{H})|=2$, the result is clear. Suppose, inductively, that the result has been proved for smaller values of $|V(\mathcal{H})|$. Assume that $x$ is a shedding vertex of $\mathcal{H}$. Let $\Delta=\Delta_{\mathcal{H}}$, $\Delta_1=\Delta_{\mathcal{H}\setminus x}$ and $\Delta_2=\Delta_{\mathcal{H}/x}$. Then $\mathcal{H}\setminus x$ and $\mathcal{H}/x$ are vertex decomposable hypergraphs and no facet of $\Delta_2$ is a facet of $\Delta_1$. By inductive hypothesis we have $${\mathrm}{pd}(R/I_{\Delta_1})\leq d'_{\mathcal{H}\setminus x} \ {\mathrm}{and} \ {\mathrm}{pd}(R/I_{\Delta_2})\leq d'_{\mathcal{H}/x}.$$ On the other hand, by Corollary 2.10 in [@MK], we have the equality $${\mathrm}{pd}(R/I_{\Delta})=\max\{{\mathrm}{pd}(R/I_{\Delta_1})+1,{\mathrm}{pd}(R/I_{\Delta_2})\}.$$ Hence $${\mathrm}{pd}(R/I_{\Delta})\leq \max\{d'_{\mathcal{H}\setminus x}+1,d'_{\mathcal{H}/x}\}.$$ Now, the result immediately follows from Lemmas \[1\] and \[2\]. We end this paper by the following result, which is a generalization of Proposition 2.7 and Corollary 2.9 in [@KM] and can characterize the projective dimension of certain hypergraphs in special circumstances. (Compare [@KM Proposition 2.7 and Corollary 2.9].)\[final\] Assume that $\mathcal{B}=\{B_1,\ldots,B_n\}$ is a set of semi-strongly disjoint set of bouquets of $\mathcal{H}$ such that $d'_{\mathcal{H}}=|F(\mathcal{B})|$. Then - there exists a minimal vertex cover of $\mathcal{H}$ contained in $F(\mathcal{B})$. - If moreover all edges of $S(\mathcal{B})$ has cardinality two (specially if $\mathcal{H}$ is a graph), then $F(\mathcal{B})$ is a minimal vertex cover of $\mathcal{H}$ and so $$c_\mathcal{H}\leq d_\mathcal{H}\leq d'_\mathcal{H}\leq{\mathrm}{bigheight}(I(\mathcal{H}))\leq {\mathrm}{pd}(R/I(\mathcal{H})).$$ - If $H$ is a vertex decomposable hypergraph and all edges of $S(\mathcal{B})$ has cardinality two (specially if $\mathcal{H}$ is a graph), then $${\mathrm}{bigheight}(I(\mathcal{H}))={\mathrm}{pd}(R/I(\mathcal{H}))=d'_\mathcal{H}.$$ <!-- --> - Suppose that $E(\mathcal{H})\setminus S(\mathcal{B})=\{E_1, \dots, E_k\}$. Then for each $1\leq i\leq k$, $E_i\cap F(\mathcal{B})\neq\emptyset$. Because otherwise, if $E_i\cap F(\mathcal{B})=\emptyset$, then by adding the edge $E_i$ to $\mathcal{B}$ (as an stem or a bouquet with one stem), one can get to a semi-strongly disjoint set of bouquets of $\mathcal{H}$ with more flowers than $d'_{\mathcal{H}}$ which is impossible. Now, let $S(\mathcal{B})=\{E_{k+1}, \dots, E_t\}$ and set $S_0=\emptyset$. For each $1\leq i \leq t$, if $E_i\cap S_{i-1}\neq \emptyset$, then set $S_i:=S_{i-1}$, else choose a vertex $x_i\in E_i\cap F(\mathcal{B})$ and set $S_i:=S_{i-1}\cup \{x_i\}$. It can be easily seen that $S_t$ is a minimal vertex cover of $\mathcal{H}$ contained in $F(\mathcal{B})$ as desired. - The first statement can be easily seen by (i). The inequalities can be gained by Proposition \[cd\], the first statement and [@MVi Corollary 3.33]. - follows from (ii) and Theorem \[pd\]. [10]{} Hypergraphs. Combinatorics of finite sets. Translated from the French. North-Holland Mathematical Library, 45. North-Holland Publishing Co., Amsterdam, 1989. arXiv:1205.5631. arXiv:1308.1299. J. Combin. Theory Ser. A 120 (2013), no. 2, 453–469. J. Algebraic Combin. 38 (2013), no. 1, 37–55. Math. Scand. 106 (2010), no. 1, 50–66. Proc. Amer. Math. Soc. 137 (2009), no. 10, 3271–3282. J. Algebraic Combin. 27 (2008), no. 2, 215–245. arXiv:1301.6779. European J. Combin. 25 (2004), no. 7, 949–960. Harmony of Gröbner bases and the modern industrial society, 153–168, World Sci. Publ., Hackensack, NJ, 2012. To appear in Proc. Amer. Math. Soc. arXiv:1309.7948. Bull. Iranian Math. Soc. 36 (2010), no. 2, 267–277. To appear in Math. Scand. Progress in Commutative Algebra, Combinatorics and Homology, Vol. 1 (C. Francisco, L. C. Klingler, S. Sather-Wagstaff and J. C. Vassilev, Eds.), De Gruyter, Berlin, 2012, pp. 85–126. J. Combin. Theory Ser. A 118 (2011), no. 2, 491–501. Arch. Math. (Basel) 93 (2009), no. 5, 451–459. Proc. Amer. Math. Soc. [**137**]{} (2009), no. 10, 3235–3246. The electronic journal of combinatorics, [**18**]{} (2011), no. 1. To appear in J. Commut. Algebra. arXiv:1009.2756. Comm. Algebra [**32**]{} (2004), no. 6, 2301–2324.
{ "pile_set_name": "ArXiv" }
--- abstract: | [**Abstract**]{} In certain scenarios of deformed relativistic symmetries relevant for non-commutative field theories particles exhibit a momentum space described by a non-abelian group manifold. Starting with a formulation of phase space for such particles which allows for a generalization to include group valued momenta we discuss quantization of the corresponding field theory. Focusing on the particular case of $\kappa$-deformed phase space we construct the one-particle Hilbert space and show how curvature in momentum space leads to an ambiguity in the quantization procedure reminiscent of the ambiguities one finds when quantizing fields in curved space-times. The tools gathered in the discussion on quantization allow for a clear definition of the basic deformed field mode operators and two-point function for $\kappa$-quantum fields. author: - Michele Arzano title: 'Anatomy of a deformed symmetry: field quantization on curved momentum space' --- Introduction ============ A characteristic feature of quantum field theory in curved space-time is that different observers, in general, do not agree on the particle content of [*the same*]{} quantum state of the field i.e. a “natural" definition of particle does not exist [@Davies:1984rk]. This is ultimately due to the lack of a unique choice of notion of time which guides the distinction between positive and negative frequency/energy modes and thus of particle and antiparticle at the quantum level. Such ambiguity is already present when the space is a maximally symmetric one even though in these cases the symmetries available offer some criteria to pick up a particular choice of vacuum state. The familiar Poincaré invariant vacuum in Minkowski space and the Bunch-Davies vacuum in de Sitter space are well known examples of such states.\ The main motivation for the present work was the observation that, perhaps not surprisingly, similar issues regarding ambiguities in the definition of frequency/energy arise in a quite different setting namely for certain classes of “non-commutative" field theories in which usual commuting space-time coordinates are replaced by generators of a Lie algebra. In such theories, as discussed in detail in the rest of the paper, momentum space will turn into a non-abelian Lie group and thus into a curved manifold. A natural question is why should one be interested in studying field theories defined on a curved momentum space. One motivation comes from lower dimensional physics. As first pointed out by ’t Hooft [@'tHooft:1996uc] the momentum of a particle coupled to three-dimensional gravity as a conical defect is given by an angle leading to Lie algebra valued particle coordinates (see [@Matschull:1997du] and references therein for an extended discussion). In higher dimension we encounter two more contexts in which field theories with curved momentum space play a major role. On one side field theories defined on group manifolds are very useful tools in non-perturbative quantum gravity where they provide a way of generating amplitudes for spin-foam models (see e.g. [@Oriti:2009wn]). On the other hand certain models of non-commutative field theories are associated with momentum spaces given by homogeneous spaces other than the usual $\mathbb{R}^{3,1}$. In these cases the curvature in momentum space introduces an energy scale which is [*invariant*]{} under the action of deformed relativistic symmetry generators [@Amelino-Camelia:2001fd; @Arzano:2007ef; @Arzano:2007gr; @KowalskiGlikman:2003we; @Freidel:2007yu; @Joung:2008mr; @Meljanac:2010ps; @Battisti:2010sr; @Girelli:2009ii; @Girelli:2010wi].\ Since the operational interpretation of non-commuting space-time coordinates is not immediate the starting point of our discussion will be a “symmetry based" description of the phase space of a relativistic particle alternative to the usual formulation in terms of cotangent bundle of a configuration space. We will describe how this picture of a classical phase space naturally leads to the definition of a quantum one-particle Hilbert space. However the crucial step that permits the distinction between particle and antiparticle states i.e. positive and negative energy states requires the introduction of a complex structure “by hand". This will be discussed in detail in Section III where we also recall how the arbitrariness of this choice is at the root of the ambiguity one encounters in the choice of vacuum state in curved space-times. In Section IV we introduce the notion of “curved" momentum space at the level of phase space focusing on a four dimensional model based on the $\kappa$-deformed Poincaré algebra where momentum space is embedded in a Lie group described by a sub-manifold of de Sitter space. The structure of the momentum space group manifold is described in more in detail in the beginning of Section V as a preparation for the following discussion on the one-particle quantization from the deformed phase space and the related ambiguities. In Section VI we provide a practical construction of the one-particle Hilbert space and field operators obtaining an explicit form of the two-point function and discussing the behaviour of quantum fluctuations of deformed field modes. We conclude, in Section VII, with a summary of the results and a brief discussion. From particles to fields ======================== Classical relativistic particle: phase space and symmetries ----------------------------------------------------------- In classical mechanics one has two equivalent ways of describing the phase space of a free relativistic particle. The usual approach is simply to take as the configuration space the ÒrangeÓ space of the coordinates of a particle (Minkowski space, $\mathbb{R}^{3,1}$) and define the (unreduced) phase space as the cotangent bundle of such configuration space. The physical phase space will be given by a six-dimensional sub-manifold of the unreduced phase space whose coordinates parametrize geodesics in Minkowski space. From an abstract mathematical point of view such phase space consists of a [*symplectic manifold*]{} $(\mathcal{M},\,\Omega)$, with $\mathcal{M}$ the cotangent bundle of the configuration space equipped with a closed non-degenerate two-form $\Omega$ (for more details see e.g. [@AbrMars]).\ For a classical mechanical system which admits a continuous group of symmetries $G$ the phase space can be alternatively described by a group theoretic construction known [*co-adjoint orbit method*]{} [@Kir] which emphasizes the deep relation between $\mathcal{M}$ and $G$. In this case the phase space can be constructed starting from the algebra $\mathfrak{g}^*$ dual to the Lie algebra $\mathfrak{g}$ of the symmetry group $G$. Since the symmetry group $G$ has a natural [*co-adjoint*]{} action on $\mathfrak{g}^*$ the phase space manifold $\mathcal{M}$ will be given by the orbit $\mathcal{O}_Y$ of the co-adjoint action of $G$ on an element $Y\in \mathfrak{g}^*$. The symplectic structure on $\mathcal{O}_Y$ will be induced by the natural symplectic structure on the dual algebra $\mathfrak{g}^*$. The latter is defined as follows. Take an element $Y\in \mathfrak{g}^*$, since $\mathfrak{g}^*$ is a vector space the tangent space $T_Y\mathfrak{g}^*\simeq \mathfrak{g}^*$. If we take a smooth function on the dual algebra $f\in C^{\infty}(\mathfrak{g}^*)$ then the differential $(df)_Y: T_Y\mathfrak{g}^* \rightarrow \mathbb{R}$ i.e. $(df)_Y$ can be seen as an element of the Lie algebra $\mathfrak{g}$ since $(df)_Y\in ( \mathfrak{g}^*)^*\simeq \mathfrak{g}$. The Poisson bracket on $C^{\infty}(\mathfrak{g}^*)$ is then given in terms of the commutators of $\mathfrak{g}$ by $$\label{LiePoiss} \{f,g\}(Y)\equiv \langle Y, [(df)_Y, (dg)_Y]\rangle \,,$$ where we used the natural pairing $\langle Y, \xi\rangle$ of $\mathfrak{g}$ and $ \mathfrak{g}^*$ as vector spaces. The orbits $\mathcal{O}_Y$ of the co-adjoint action of $G$ on an element $Y\in \mathfrak{g}^*$ equipped with the symplectic structure above become symplectic manifolds which describe the phase spaces of $G$-symmetric mechanical systems.\ In our specific context we are interested in the phase space of a relativistic point particle and thus we take the symmetry group $G$ to be the Poincaré group $ISO(3,1)= SO(3,1)\ltimes \mathbb{R}^{3,1}$. In this case $\mathfrak{g}^*=\mathfrak{iso}^*(3,1)\equiv \mathfrak{so}^*(3,1) \oplus (\mathbb{R}^{3,1})^*$ and the co-adjoint orbits $\mathcal{O}_{m,s}$ are given by level hyper-surfaces of the two Casimir functions $\mathcal{C}_1(p)$ and $\mathcal{C}_2(w)$ on $\mathfrak{iso}^*(3,1)$. More specifically if we fix a set of co-ordinates $(p^0,\, p^i,\, j^i, k^i)$ on $\mathfrak{iso}^*(3,1)$ then we take $p=(p^0, p^i)$ and define the Pauli-Lubanski four vector $w=(w^0, w^i)$ by $$w^0={\bf p}\cdot{\bf j}\,,\,\,\,\,\,\,\,\, \vec{w}={\bf p}\times{\bf k}+p^0 {\bf j}\,.$$ The mass and spin labels of the co-adjoint orbit $m$ and $s$ will be related to the fixed values of the functions $\mathcal{C}_1=p\cdot p$ and $\mathcal{C}_2=w\cdot w$. Writing explicitly the Poisson structure on $\mathcal{O}_{m,s}$ for a specific choice of coordinate functions it can be seen [@Carinena:1989uw] that $\mathcal{O}_{m,s}\simeq \mathbb{R}^6\times S^2$ as a Poisson manifold i.e. a symplectic manifold describing the phase space of a relativistic spinning particle. Notice here that the main advantage of the co-adjoint method approach is that it offers the most general formulation of a relativistic particleÕs phase space since encompasses the case of spinning particle which is normally not straightforward to describe in terms of the cotangent bundle on a configuration space [@Balachandran:1991zj].\ From here on we will focus on the phase space of a spinless relativistic particle. In this case the Pauli-Lubanski vector vanishes identically and we denote the co-adjoint orbit by $\mathcal{O}_{m,0}$. As mentioned above the dual algebra $\mathfrak{g}^*=\mathfrak{iso}^*(3,1)$ carries a natural Poincaré invariant Poisson structure directly related to the commutators of the Lie algebra $\mathfrak{g}=\mathfrak{iso}(3,1)$. Indeed every $\xi\in \mathfrak{g}$ defines a linear co-ordinate function on $\mathfrak{g}^*$ given by $f_{\xi}$ such that $f_{\xi}(Y)=\langle Y, \xi \rangle$. As we pointed out above for any function $f$ on $\mathfrak{g}^*$ the one-form $df$ can be seen as an element of the Lie algebra $\mathfrak{g}$. In particular if we consider co-ordinate functions on $\mathfrak{g}^*$ associated with the generators of the Lie algebra $\xi_i$ then $df_{\xi_i}\equiv \xi_i$. Denoting $h_i\equiv f_{\xi_i}$ it is easy to see that the Poisson brackets induced by the commutators of the Lie algebra $\mathfrak{g}$ will be given by $$\{h_i, h_j\}=c^k_{ij} h_k\, ,$$ where $c^k_{ij}$ are the structure constants of $\mathfrak{g}$. Starting from the coordinate functions $(p^0,\, p^i,\, j^i, k^i)$ on $\mathfrak{iso}^*(3,1)$ one can define a set of canonical co-ordinates on $\mathcal{O}_{m,0}$ using the spatial momentum coordinates $p^i$ and defining the position coordinates $$q^i=\frac{k^i}{p^0}\, ,$$ with the coordinates satisfying the constraints $w^i =w^0 =0$ and $(p^0)^2-{\bf p}^{\,2}=m^2$. Using the general formulae above it is easy to check that the canonical “phase space" coordinates $\{q^i,p^i\}$ close the usual Poisson brackets $$\{q_i, q_j\}=\{p_i, p_j\}=0\,,\,\,\,\,\,\,\,\,\,\{q_i, p_j\}=\delta_{ij}\, ,$$ and thus $\mathcal{O}_{m,0}\simeq \mathbb{R}^6$ as expected. Describing the phase space in terms of the co-adjoint orbit is in some way equivalent to consider a symplectic manifold whose natural coordinates are “Poincaré momenta”. Using co-adjoint orbits to describe phase space we have a straightforward connection with the irreducible representations of $\mathfrak{iso}(3,1)$ since the latter are also labelled by the eigenvalues of the two invariant functions $\mathcal{C}_1$ and $\mathcal{C}_2$. We devote the rest of this section to such connection.\ Phase space of a classical field -------------------------------- As a preparation for the discussion below it will be useful to make a short digression on the meaning of “positions" and “momenta" when describing the phase space and symmetries of a relativistic particle. Let us denote with $T$ the group of space-time translation. For ordinary relativistic symmetries this is just $\mathbb{R}^{3,1}$ seen as a group under addition. The Lie algebra $\mathfrak{t}$ of translation generators, as a tangent space to the identity element, can be identified with $\mathbb{R}^{3,1}$ as vector spaces. The (trivial) Lie bracket on $\mathfrak{t}$ is induced by the addition law of the group $T\equiv\mathbb{R}^{3,1}$. The dual group $T^*$ is, by definition, given by equivalence classes of unitary irreducible representations of $T$ and in the case $T\equiv\mathbb{R}^{3,1}$ elements of $T^*\equiv(\mathbb{R}^{3,1})^*$ are given by one-dimensional characters or in physics language ‘plane waves’. When we write a plane wave like $e_{p}\in(\mathbb{R}^{3,1})^*$ we are simply saying that such element of the dual group $(\mathbb{R}^{3,1})^*$ has coordinates given by the four-vector $p$.\ One usually refers to ’positions’ (as elements of the ambient space on which we build the (unreduced) configuration space) as given by coordinates on Minkowski space i.e. the translation group $T\equiv\mathbb{R}^{3,1}$. Indeed, in the usual description, the unreduced phase space of a non-spinning relativistic particle is given by the cotangent bundle of the group of translations $T$ which is isomorphic [@AbrMars] to $T \times \mathfrak{t}^*$. From this point of view ’momenta’ are just coordinates on the dual Lie algebra $\mathfrak{t}^*$. Let us point out that one also speaks of ’momenta’ when referring the space-time translation generators i.e. a basis of the Lie algebra $\mathfrak{t}$. In this case space-time ‘coordinates’ correspond to the basis of generators the dual algebra $\mathfrak{t}^*$. In ordinary relativistic theories we can refer to coordinates and momenta without specifying the objects we are referring to because $T$ and $\mathfrak{t}$ can be identified and so can their duals $T^*$ and $\mathfrak{t}^*$. Notice how, instead, from the more general point of view of the co-adjoint orbit description of phase space it is only correct to say that the dual algebra $\mathfrak{iso}^*(3,1)$ provides the ambient space on which both position and momenta are defined. As we will see in Section IV the distinction between $T$, $T^*$ and their respective Lie algebras will be crucial when momentum space becomes “curved". In that context a description of phase space in terms of co-adjoint orbits will provide a very clear characterization of the structures that lie at the basis of symmetry deformation.\ Going back to our spinless relativistic particle, in the language of co-adjoint orbits its “momentum space" will be given by the subspace $M_{m}\subset \mathcal{O}_{m,0}$ (the “mass-shell") obtained by considering the restriction to the co-adjoint orbits of the abelian subalgebra $\mathfrak{t}^*\equiv(\mathbb{R}^{3,1})^*$ of $\mathfrak{g}^*=\mathfrak{iso}^*(3,1)$ dual to the algebra of translation generators. Since for a ordinary relativistic particle in Minkowski space we can identify $\mathfrak{t}^*$ with $T^*$ the momentum space $M_{m}$ can be characterized in a [*coordinate independent*]{} way as a orbit of a character (“plane wave") under the action of the group $SO(3,1)$ (see [@Barut:1986dd]), i.e. $$M_m \equiv\{\gamma e_{p}: e_{p} \in (\mathbb{R}^{3,1})^*,\, \gamma\in SO(3,1) \}\, ,$$ which, keeping in mind the discussion above, can be described in terms of the co-ordinate functions on the dual algebra $\mathfrak{t}^*\equiv(\mathbb{R}^{3,1})^*$ by the two-sheeted hyperboloid $(p^0)^2-{\bf p}^{\,2}=m^2$. From its definition as a orbit of a symmetry group $M_m$ has a natural structure of a homogeneous space, indeed $$M_m \simeq SO(3,1)/SO(3)\, ,$$ with $SO(3)$ the “isotropy" subgroup of $SO(3,1)$ which leaves invariant the point $(m,0,0,0)$. Like any homogeneous space (under some additional assumptions see Barut pag 130) $M_m$ admits an invariant measure on its space of functions. On the space of complex valued functions on the mass-shell $C^{\infty}(M_m )$ we can define the invariant measure $d\mu_m$ using the following trick [@Sternberg:1994tw]: one looks for the volume 3-form which satisfies $$dV= d (\mathcal{C}_1(p))\wedge d\mu_m$$ where $dV$ is the ordinary volume 4-form on $\mathbb{R}^{3,1}$. The invariant measure on $C^{\infty}(M_m )$ can be usefully written as a “$\delta$-measure" $$d\mu_m = d V \delta(\mathcal{C}_1(p))\, .$$ In the same spirit we can think of elements of $C^{\infty}( M_{m} )$ as distributions on $(\mathbb{R}^{3,1})^*$ given by $$\tilde{\phi}(p)= \delta(\mathcal{C}_1(p)) \tilde{f}(p)$$ with $\tilde{f}(p)\in C^{\infty}((\mathbb{R}^{3,1})^*)$. A necessary and sufficient condition for a distribution to be of the form above is that $$(\mathcal{C}_1(p)-m^2)\tilde{\phi}(p)=0\, .$$ On the space of functions $C^{\infty}((\mathbb{R}^{3,1})^*)$ we can introduce a notion of Fourier transform which is just a special (trivial) case of the general Fourier transform of functions on a group (which will be useful later on) $$f(\Lambda) = (d_{\Lambda})^{-1} \int_{G}d\mu(g)\, \tilde{f}(g)\, \hat{n}_{g}(\Lambda)$$ where $\Lambda$ is an index of an irreducible representation of $G$, $d_{\Lambda}$ its dimension and $\hat{n}_{g}(\Lambda)$ the character of such representation. In our particular case for $\tilde{f}(p)\in C^{\infty}((\mathbb{R}^{3,1})^*)$ and $\tilde{\phi}(p)\in C^{\infty}( M_m )$ one has the familiar expressions $$f(x) = \int_{(\mathbb{R}^{3,1})^*}d\mu(p)\, \tilde{f}(p)\, e_p(x)\,,\,\,\,\,\,\,\,\,\,\,\,\,\phi(x) = \int_{(\mathbb{R}^{3,1})^*}d\mu(p)\, \delta(\mathcal{C}_1(p))\, \tilde{f}(p)\, e_p(x)\, .$$ where $d\mu(p)=\frac{d^4 p}{(2\pi)^{3/2}}$ and $e_p(x)=\exp(-i p x )$. Finally noting that under Fourier transform $\partial_i \phi(x)\rightarrow i p_i \tilde{\phi}(p)$ we have that $$(\mathcal{C}_1(p)-m^2)\tilde{\phi}(p)=0\,\Longleftrightarrow (\Box+m^2)\phi(x)=0\,,$$ the Fourier transform maps functions on the mass shell hyperboloid into the space of solutions of the Klein-Gordon equation $\mathcal{S}$. Notice that we also have $-\partial_i \phi^*(x)\rightarrow -i p_i \tilde{\phi}^*(p)$ and due to the quadratic nature of the equations above we make the identification $\tilde{\phi}^*(p)=\tilde{\phi}(-p)$ and $\phi^*(x)=\phi(x)$ i.e. the solutions of the Klein-Gordon equation are [*real valued*]{} functions. The phase space of a classical field is then given by the symplectic manifold $(\mathcal{S},\omega)$ with symplectic structure provided by the antisymmetric bilinear form $\omega$ given by the Wronskian[^1] associated to the Klein-Gordon equation $$\label{omega} \omega(\phi_1,\phi_2)=\int_{\Sigma}(\phi_2\nabla_{\mu}\phi_1-\phi_1\nabla_{\mu}\phi_2)d\Sigma^{\mu}\, .$$ This exhibits nicely the connection between phase space of a relativistic spinless point particle and the phase space of a classical scalar field. Let us remark here that in Minkowski space (and in general on any globally hyperbolic space) the field’s phase space is given by an equivalent description in terms of the space of initial data $\{\varphi, \pi\}$ on a given Cauchy surface $\Gamma_{\Sigma}$ with the symplectic form given by the restriction of $\omega$ above to such space. In the next section we will discuss how a natural structure of inner product can be defined on the field’s phase space and how this can be used to construct the “one-particle" Hilbert space of the corresponding quantum field theory. Complex numbers and field quantization ====================================== As we discussed above classical fields are [*real*]{} fields. In classical field theory complex variables are often used as a computational tool with no physical meaning. When we turn to the quantum setting however complex numbers become fundamental. From the point of view of quantum observables the imaginary unit $i$ is introduced in order to turn differential operators into self-adjoint operators (e.g. momenta as generators of translations). From the point of view of quantum states these are now rays of a [*complex*]{} Hilbert space. Indeed, from a modern perspective, the very concept of quantization of a classical field amounts to the introduction of an appropriate [*complex structure*]{} $J$ on the classical phase space of the theory [@Bongaarts:1971cu; @Ashtekar:1975zn; @Panangaden:1979mi; @Gibbons:1993iv].\ In the section above we discussed how the phase space of a classical field can be described by the space of solutions of the classical equations of motions $\mathcal{S}$. This characterization of phase space will give an intuitive physical interpretation of the role of the complex structure since as we will see in more detail below, $J$ provides a direct sum decomposition of the [*complexification*]{} of $\mathcal{S}$, $\mathcal{S}^{\mathbb{C}}$ into “positive and negative energy" subspaces which will represent, respectively, the “one-particle" Hilbert space of the theory $\mathcal{H}$ and its complex conjugate $\bar{\mathcal{H}}$ once they are equipped with an appropriate inner product. Of course the choice of $J$ is not unique but in certain specific cases it will be dictated by further physical inputs. For example for a real scalar field in Minkowski space there exists a unique Poincaré invariant complex structure and it corresponds to the familiar textbook decomposition of the field in positive and negative frequency modes. In more general space-times there will be no unique choice of $J$ and this is at the basis of the well known phenomenon of particle creation. In this case different observers will decompose the field according to different notion of positive and negative energy and will define different vacuum states for their quantum field. From a more fundamental point of view such observers are just choosing different complex structures in representing the Hilbert space of their quantum field theory.\ Let’s try to be more concrete. To introduce a [*complex structure*]{} on $\mathcal{S}$ amounts to define an automorphism $J:\mathcal{S}\rightarrow\mathcal{S}$ such that $J^2=-1$. As we mentioned above the introduction of $J$ corresponds to a choice of decomposition of $\mathcal{S}^{\mathbb{C}}$ in positive and negative energy subspaces. Recall that the complexification $\mathcal{S}^{\mathbb{C}}$ of $\mathcal{S}$ is defined by $$\mathcal{S}^{\mathbb{C}}\equiv\mathcal{S}\otimes\mathbb{C}\,.$$ The complex linear extension of $J$ to $\mathcal{S}^{\mathbb{C}}$ is given by $$J(\phi\otimes z)\equiv J(\phi)\otimes z\, .$$ The introduction of $J$ gives rise to a natural decomposition of $\mathcal{S}^{\mathbb{C}}$ into two subspaces, $\mathcal{S}^{\mathbb{C}+}$ and $\mathcal{S}^{\mathbb{C}-}$ spanned, respectively, by the eigenvectors of $J$ with eigenvalues $\pm i$ i.e. $J(\phi^{\pm})=\pm i(\phi^{\pm})$. We can define projectors $P^{\pm}:\mathcal{S}\rightarrow\mathcal{S}^{\mathbb{C}\pm}$ $$P^{\pm}\equiv \frac{1}{2}(1\mp iJ)\,,$$ with $$\mathcal{S}^{\mathbb{C}}=\mathcal{S}^{\mathbb{C}+}\oplus \mathcal{S}^{\mathbb{C}-}\,.$$ The connection with positive and negative energy decomposition is now easily seen. If the background space-time admits a timelike and hypersurface orthogonal Killing vector field $\mathcal{L}_t$, i.e. it is [*static*]{}, one can decompose any real solution $\phi\in\mathcal{S}$ in normal modes (e.g. plane waves) of positive and negative energy components with respect to $\mathcal{L}_t$ $$\phi=\phi^+ +\phi^-\,.$$ Then the map $J=-(-\mathcal{L}_t\mathcal{L}_t)^{-1/2}\mathcal{L}_t$ is such that $$J\phi=i\phi^+ +(-i)\phi^-\,,\,\,\,\,\,\,\,\,\,\,\,\,P^{\pm}\phi= \phi^{\pm}$$ i.e. $J$ is a complex structure on $\mathcal{S}$ and it provides a decomposition of $\mathcal{S}^{\mathbb{C}}$ in positive and negative energy subspaces. Put the other way around [*a decomposition of $\mathcal{S}^{\mathbb{C}}$ in positive and negative energy subspaces singles out a preferred complex structure $J$*]{}. Of course in order to obtain the “one-particle" Hilbert space $\mathcal{H}$ from $\mathcal{S}^{\mathbb{C}+}$ we need to equip the latter with a positive definite inner product. This can be constructed using $J$ itself and the natural symplectic structure (\[omega\]) of the classical phase space under the further requirement that the complex structure be [*compatible*]{} with the symplectic structure $\omega$, namely $$\omega(J\phi_1,J\phi_2)=\omega(\phi_1,\phi_2)\, .$$ The positive definite inner product on the positive energy subspace $\mathcal{S}^{\mathbb{C}+}$ will be given by $$\label{innp} (\phi^+_1,\phi^+_2)\,=-i \omega(\overline{P^+\phi_1},P^+\phi_2)= \frac{1}{2}\left(\omega(J\phi_1,\phi_2)-i \omega(\phi_1,\phi_2)\right) \, .$$ It is easily checked that such product is positive definite on $\mathcal{S}^{\mathbb{C}+}$ and thus the one particle Hilbert space $\mathcal{H}$ of the theory is obtained by taking the completion of $\mathcal{S}^{\mathbb{C}+}$ with respect to the above inner product. The complex conjugate space $\bar{\mathcal{H}}$ can be thus identified with the subspace $\mathcal{S}^{\mathbb{C}-}$ and corresponds to the “one-antiparticle" space. The point that should be stressed (for a detailed discussion see [@Wald:1995yp]) is that to each choice of complex structure will correspond a inner product (and a corresponding Hilbert space construction) and vice versa.\ It would be good at this point to make contact with the usual textbook formalism to see concrete realizations of these rather abstract constructions. The Fourier transform of the an element $\phi\in\mathcal{S}$ can be recast as a normal mode expansion $$\label{realsol2} \phi({\bf x}, t)=\int d\mu({\bf k})\,\left [\phi^+({\bf k}) e_{{\bf k}} +\phi^-({\bf k}) \bar{e}_{{\bf k}} \right]$$ where $e_{{\bf k}}$ is a positive energy plane wave solution $$e_{\bf k}\equiv \frac{1}{(2\pi)^{3/2}} \exp(i {\bf kx }-i\omega_{{\bf k}}t)$$ with $\omega_{{\bf k}}=\sqrt{{\bf k}^2+m^2}$, $d\mu({\bf k})=\frac{d{\bf k}}{2\omega_{{\bf k}}}$ and the following relation between the modes (\[realsol2\]) and the Fourier coefficients: $\phi^+({\bf k})=\tilde{\phi}(-\omega_{{\bf k}}, - {\bf k})\,,\,\, \phi^-({\bf k})= \tilde{\phi}(\omega_{{\bf k}}, {\bf k})$. Positive and negative energy modes are defined w.r.t. the inertial time translation Killing vector $\partial_t$ and thus according to the discussion above $J=\frac{-\partial_t}{(-\partial_t\partial_t)^{1/2}}$ and in terms of the time translation generator $P_0=i\partial_t$ $$iJ=-\frac{P_0}{|P_0|}\,,\,\,\,\,\,\,\,\,\,\,\,P^{\pm}=\frac{1}{2}\left(1\pm \frac{P_0}{|P_0|}\right)\, .$$ Using the expression for the projector above we have $$\label{posolcov} \phi^+(x)=\frac{1}{(2\pi)^{3/2}}\int dk\, \delta(k^2-m^2)\, \theta(k_0)\, \tilde{\phi}(k)\exp(-ikx)=\frac{1}{(2\pi)^{3/2}}\int \frac{d{\bf k}}{2\omega_{{\bf k}}}\, \phi^+({\bf k})\exp(i {\bf kx }-i\omega_{{\bf k}}t)\, ,$$ with $\phi^-(x)\equiv\overline{\phi^+}(x)$ and from (\[omega\]) $$\label{caninner} (\phi_1^+,\phi_2^+)\equiv-i\omega(\overline{P^+\phi}_1,P^+\phi_2)=\int \frac{d{\bf k}}{2\omega_{{\bf k}}}\,\, \phi_1^-({\bf k}) \phi^+_2({\bf k})\, .$$ This shows how an equivalent description of the one particle Hilbert space is given by $\mathcal{H}=(M_m^+; d\mu({\bf k}))$, the space of functions on the positive mass-shell square integrable with respect to the Lorentz invariant measure $d\mu({\bf k})=\frac{d{\bf k}}{2\omega_{{\bf k}}}$. The inner product defined above extends to a natural inner product on the whole mass-shell $M_m=M_m^+\cup M_m^-$ given by $$\omega(\overline{\phi}_1,\phi_2)=i\int d^4 k\,\delta(k^2-m^2)\, \overline{\tilde{\phi}}_1(k) \tilde{\phi}_2(k)\, ,$$ from which is is easy to write the covariant version of (\[caninner\]) $$\label{caninnerco} (\phi^+_1,\phi^+_2)=\int d^4 k\,\delta(k^2-m^2)\,\theta(k^0)\, \overline{\tilde{\phi}}_1(k) \tilde{\phi}_2(k)\, .$$ Notice how the $\delta$-measure $d^4 k\,\delta(k^2-m^2)$ is exactly the invariant measure on the the space of functions on the homogenous space $M_m\simeq SO(3,1)/SO(3)$ we introduced in the previous section and that the complex structure, through the projection operator $P^+$, singles out a subspace of it, that of functions on the ‘positive energy’ mass-shell. A basis of one-particle states will be given by monochromatic plane wave solutions $e_{\bf k}$ which we denote by kets $|\bf{k}\rangle\in\mathcal{H}$. From (\[posolcov\]) we see that the modes associated with such solutions are $$e^+_{{\bf k}}({\bf p})\equiv 2\omega_{\bf k} \,\, \delta^3({\bf p}-{\bf k})\, .$$ It is easily checked that the normalized plane wave solutions above provide an orthogonal basis for $\mathcal{H}$ indeed $$\langle {\bf k}_1|{\bf k}_2\rangle\equiv \,(e^+_{\bf k_1},e^+_{\bf k_2})=\int \frac{d{\bf k}}{2\omega_{{\bf k}}}\,\, e_{\bf k_1}^-({\bf k}) e^+_{\bf k_2}({\bf k})=2 \omega_{\bf k_1}\delta^3({\bf k_1}-{\bf k_2})\, ,$$ as expected. In the rest of the paper we will show how the construction above can be extended to the quantization of a classical relativistic particle with a deformed phase space and group-valued momenta. Bending phase space =================== The main point of this and the following section will be to show that when the space $M_m$ is embedded in a group there will be quite dramatic consequences for field quantization. In particular the introduction of curvature in momentum space leads to an ambiguity in the definition of the energy of one-particle states in terms of field modes. This is somewhat analogous to what happens for quantum fields in curved space where one does not have a preferred notion of vacuum due to the lack of a unique way of measuring time and energy for different observers. In our case to each choice of co-ordinates on (curved) momentum space will correspond a choice of field modes or “linear momentum" of one-particle states.\ To start off let us make more clear the notion of “momentum becoming group valued". In Section II we saw how the ambient space on which the momentum sector of the phase space of a classical relativistic particle is built is the Lie algebra $\mathfrak{t}^*$ dual to the algebra of translation generators $\mathfrak{t}$. When we say that the momentum becomes “group valued" we mean that the Lie algebra $\mathfrak{t}^*$ acquires non-trivial Lie brackets i.e. it becomes non-abelian (unlike the case of a particle in ordinary Minkowski space). This is to say that the dual group $T^*$ is now a non-abelian group and thus momenta, as labels of plane waves, will obey a non-abelian composition rule. Let’s first see what consequences this has in general and then discuss a particular four-dimensional example.\ First of all according to the discussion in Section II and eq. (\[LiePoiss\]) a non-trivial Lie bracket on $\mathfrak{t}^*$ will correspond to a non-trivial Poisson-Lie structure on its dual algebra i.e. coordinate functions $x^{\mu}$ on $\mathfrak{t}$ will now have non-trivial Poisson brackets $$[\cdot,\cdot]_{\mathfrak{t}^*}\neq 0\longrightarrow \{\cdot,\cdot\}_{\mathfrak{t}}\neq 0\,\, .$$ The second consequence is that a non-trivial Lie bracket on $\mathfrak{t}^*$ induces a new structure on $\mathfrak{t}$, a “non-trivial co-commutator" i.e. a function $\delta: \mathfrak{t}\rightarrow \mathfrak{t}\otimes \mathfrak{t}$ (which, as we will see in the next Section, will give the leading order deviation from the Leibniz rule (co-product) for a basis of the algebra of polynomials of the translation generators) defined by $$\delta (Y) (\xi_1,\xi_2)\equiv \langle Y, [\xi_1,\xi_2] \rangle\,,\,\,\,\,\,\,\,\,\,\,[\cdot,\cdot]_{\mathfrak{t}^*}\neq 0\longrightarrow \delta(\cdot)_{\mathfrak{t}}\neq 0\,\, .$$ For more details about the interplay between Poisson-Lie structures and Lie-bialgebra structures we refer the reader to [@Chari:1994pz]. Notice how even when the new structures are introduced the algebra of translation generators $\mathfrak{t}$ is still abelian and thus at the [*Lie algebra level*]{} the Poincaré algebra is unchanged. This means that the adjoint orbits of the Poincaré group on its Lie algebra are the same as in the classical case and consequently, under the dual pairing (which at the Lie algebra level does not involve any product or co-product structures) the [*co-adjoint orbits are the same*]{}. This means that the classical phase space is unaffected by the introduction of a non-trivial Lie bracket on $\mathfrak{t}^*$.\ For the case of interest to us, the $\kappa$-Poincaré algebra [@Lukierski:1992dt], the most important new ingredient is that the dual algebra of translations gets equipped with the following bracket $$\label{kdual} [P^*_{\mu},P^*_{\nu}]=-\frac{1}{\kappa} (P^*_{\mu}\delta^0_{\nu}-P^*_{\nu}\delta^0_{\mu})\, .$$ The algebra generated by $P^*_{\mu}$ is isomorphic to the quotient Lie algebra $\mathfrak{b}\equiv\mathfrak{so}(4,1)/\mathfrak{so}(3,1)$ (see e.g. [@KowalskiGlikman:2004tz]). The non-trivial co-commutators on $\mathfrak{t}$ are then given by $$\label{cocomm} \delta(P^0)=0\,,\,\,\,\,\,\,\delta(P^i)=\frac{1}{\kappa} P^i \wedge P^0\, .$$ The Lie algebra structure of $\mathfrak{t}^*=\mathfrak{b}$ will correspond to a Poisson structure on $\mathfrak{t}$ given by $$\{x_{i},x_{j}\}=0\,,\,\,\,\,\{x_{0},x_{j}\}=\frac{1}{\kappa}x_j\, .$$ Such Poisson brackets bear the same structure of the commutation relations of the so-called $\kappa$-Minkowski non-commutative space-time [@Majid:1994cy] but we should be careful in identifying such co-ordinates with positions of a classical relativistic particle. Indeed as discussed in detail in Section II when building phase space from the co-adjoint orbit position variables should be constructed from the dual algebra. As in the undeformed case we have here a choice of canonical co-ordinates on the co-adjoint orbit given by $\{p_i,x_i\}$ as discussed in section II. In other words the classical phase space of a $\kappa$-particle is built from orbits of the undeformed Poincaré algebra on its dual. Even if the latter has non trivial Lie brackets the orbits are still orbits on a linear (flat) space and thus there is no ambiguity in the choice of canonical coordinates (for more details on this conclusion drawn from an alternative approach see [@Arzano:2010kz]). Quantum fields and vacuum structure: a new quantization ambiguity ================================================================= As in the undeformed case plane waves will be the key ingredient in the construction the “one-particle" Hilbert space of the theory. In the deformed phase space setting, as remarked in the previous section, the translation group $T$ is still an abelian group and thus we can define the dual group $T^*$ as the set of plane waves (characters). As unitary irreducible representations of $T$ we can denote plane waves as $e_{x}=\exp(i x_{\mu}P^{\mu})$ and as elements of the non-abelian group $T^*=B$, obtained by exponentiating the Lie algebra $\mathfrak{b}$ above we write $e_{p}=\exp(i p^{\mu}P^*_{\mu})$. What is important to notice is that, unlike the undeformed case, such plane waves will have composition law w.r.t. $T$ and $T^*$ which are respectively abelian and non-abelian $$e_{p}e_{q}\equiv e_{p\oplus q}\neq e_{q\oplus p}\equiv e_{q}e_{p}\, ,$$ and $$e_{x}e_{y}\equiv e_{x + y}= e_{y + x}\equiv e_{y}e_{x}\, .$$ Likewise we will have different behaviours under group inversion $$(e_{p})^{-1}\equiv e_{\ominus p}\, ,\,\,\,\,\,\,\, (e_{x})^{-1}\equiv e_{- x}\, .$$ The non-abelian composition rule for the $T^*$ labels can be derived in terms of the Baker-Campbell-Hausdorff formula using the Lie brackets of $\mathfrak{b}$ (see e.g. [@Kosinski:1999dw]). Notice however that the explicit form of such composition rule will depend on the choice of co-ordinates on the group manifold $T^*=B$. Some of this coordinate systems will correspond to group decompositions of $B$ which reflect in a splitting of the plane wave $e_{p}$ in purely spatial and purely temporal components. As an example we will consider the following one-parameter family of decompositions of $B$ parametrized by $0\leq|\beta|\leq 1$ $$e_{p}\equiv e^{-i\frac{1-\beta}{2}p^0 P^*_{0}}e^{ip^jP^*_{j}}e^{-i\frac{1+\beta}{2}p^0P^*_{0}}\ .$$ Such parametrization will correspond to the different momentum composition rules $$\label{copr} p\oplus_{\beta}q = ( p^0+ q^0;\, p^j\ e^{\frac{1-\beta}{2\kappa}q^0}+q^j\ e^{-\frac{1+\beta}{2\kappa}p^0})\,$$ and “antipodes" $$\label{antip} \ominus_{\beta} p = (-p^0;\, - e^{\frac{-\beta}{\kappa} p^0} p^i )\, .$$ The non trivial behaviours of the “deformed" plane waves above can be understood in terms of coordinate choices on the group manifold $B$. In order to see that let us first note that as a group manifold $B$ is represented by a submanifold of de Sitter space. If we describe the latter as a four-dimensional hyper-surface embedded in five dimensional Minkowski space $$\label{4} -z_0^2 + z_1^2 + z_2^2 + z_3^2 + z_4^2 =\kappa^2\, ,$$ it can be shown [@KowalskiGlikman:2004tz] that the “momentum space" $B$ is given by the submanifold[^2] defined by the inequality $z_0-z_4>0$. Each choice of group splitting will correspond to a particular choice of co-ordinates on $B$ (these are obtained from acting with a matrix representation of the group element on the stability point $(0,...,\kappa)\in \mathbb{R}^{4,1} $ seen as a column vector). For example to the ordering $\beta=1$ will correspond “flat slicing" coordinates $p_\mu$ given by $$\begin{aligned} \label{bicrossp} {z_0}(p_0, \mathbf{p}) &=& \kappa\sinh {{p_0}/\kappa} + \frac{\mathbf{p}^2}{2\kappa}\, e^{ {p_0}/\kappa}, \nonumber\\ z_i(p_0, \mathbf{p}) &=& - p_i \, e^{{p_0}/\kappa}, \nonumber\\ {z_4}(p_0, \mathbf{p}) &=& -\kappa \cosh {{p_0}/\kappa} + \frac{\mathbf{p}^2}{2\kappa}\, e^{ {p_0}/\kappa}.\end{aligned}$$ With a straightforward but tedious calculation one can easily obtain a general expression for co-ordinate systems associated to each value of the parameter $\beta$ $$\begin{aligned} \label{bicrosspbeta} {z_0}(p_0, \mathbf{p}) &= &\kappa \left(\sinh_{+}{[p_0]} \cosh_{-}{[p_0]}+\cosh_{+}{[p_0]} \sinh_{-}{[p_0]}\right)+\nonumber\\ & &+ \left(\frac{\mathbf{p}^2}{2\kappa}\right) \left(\sinh_{+}{[p_0]} \cosh_{-}{[p_0]} + \cosh_{+}{[p_0]} \cosh_{-}{[p_0]} - \sinh_{+}{[p_0]} \sinh_{-}{[p_0]}-\cosh_{+}{[p_0]} \cosh_{-}{[p_0]}\right) \nonumber\\ z_i(p_0, \mathbf{p}) &=& - p_i \, \exp_{+}{[p_0]}, \nonumber\\ {z_4}(p_0, \mathbf{p}) &=& -\kappa \left(\sinh_{+}{[p_0]} \cosh_{-}{[p_0]}+\cosh_{+}{[p_0]} \sinh_{-}{[p_0]}\right)+\nonumber\\ & & +\left(\frac{\mathbf{p}^2}{2\kappa}\right) \left(\sinh_{+}{[p_0]} \cosh_{-}{[p_0]} + \cosh_{+}{[p_0]} \cosh_{-}{[p_0]} - \sinh_{+}{[p_0]} \sinh_{-}{[p_0]}-\cosh_{+}{[p_0]} \cosh_{-}{[p_0]}\right)\nonumber\\\end{aligned}$$ where we used the compact notation $h_{\pm}[p_0]\equiv h(\frac{1\pm\beta}{2\kappa}p_0)$ for the exponential and hyperbolic functions appearing above.\ From a mathematical point of view the different composition laws and choices of coordinates reflect the different choices of bases of the universal enveloping algebra (UEA) $U(\mathfrak{b})$ which we use to label the elements of $B$. Recall here that roughly speaking the UEA of $\mathfrak{t}$, $U(\mathfrak{t})$, is the associative algebra of polynomials of the translation generators (see [@Tjin:1991me] for a pedagogical introduction). The very important aspect of UEA of Lie algebras is that they can be endowed with an additional “co-algebra" structure which encodes the way their representations extend to tensor product spaces. In particular such rule of extending representations to tensor product spaces is defined by a map $\Delta: U(\mathfrak{t})\rightarrow U(\mathfrak{t})\otimes U(\mathfrak{t})$ called the “co-product" which for ordinary UEA is nothing but the analogous of the familiar [*Leibniz rule*]{} for derivatives acting on products of two elements. In mathematical language a UEA equipped with the additional co-algebra structure (and appropriate compatibility axioms) becomes a [*Hopf algebra*]{}. The important thing to note is that the algebra of functions on $C^{\infty}(T^*)$ also has a natural Hopf algebra structure. Indeed it turns out that $U(\mathfrak{t})$ is dual as a Hopf algebra to $C^{\infty}(T^*)$ and a choice of basis in $U(\mathfrak{t})$ will correspond to a choice of basis of coordinate functions on $C^{\infty}(T^*)$. To each composition rule related to the different group splittings described above one can associate a specific co-product for the basis elements given by $$\Delta(P_0)=P_0\otimes 1+1\otimes P_0\,\,\,\,\,\,\Delta(P_i)=P_i\otimes e^{\frac{1-\beta}{2\kappa}P_0}+e^{-\frac{1+\beta}{2\kappa}P_0}\otimes P_i\ ,$$ and the corresponding antipodes, which reflect the group inversion law of $B$ on $U(\mathfrak{b})$, given by $$\label{antipode} S(P_0)=-P_0\,\,\,\,\,\, S(P_i)= -e^{\frac{\beta}{\kappa}P_0} P_i\ .$$ From these basic ingredients, under certain compatibility requirements for the action of the Lorentz group on the deformed momentum space, one can reconstruct the structure of the whole deformed $\kappa$-Poincaré algebra (see [@Majid:1994cy] for details of the construction and [@Arzano:2007qp] for a condensed review of the $\kappa$-Poincaré algebra). Notice that the $1/\kappa$ term of the antisymmetric part of the different co-products which reproduces the co-commutator (\[cocomm\]) [*does not*]{} depend on the choice of co-ordinates and thus all the structures at the level of Lie algebra are [*uniquely*]{} defined which means that there is no ambiguity in describing the phase space of a classical relativistic particle even when the deformations are introduced.\ After this digression on the structure of the dual group $T^*=B$ we turn back to our main task which is the definition of a one-particle Hilbert space from the classical phase space described in the previous section. Now that we have identified the (deformed) space of characters, in analogy with the undeformed case, we will consider the orbits under the action of the Lorentz group. Indeed on elements of $T^*=B$ one can define a natural action[^3] which is induced from the action of the Lorentz group on the five dimensional Minkowski space, in which the de Sitter hyperboloid is embedded, keeping the $z_4$ co-ordinate fixed. This will lead to an action of the usual Lorentz group $SO(3,1)$ leaving invariant the hyperboloid [@Arzano:2009ci] $$\label{4} -z_0^2 + z_1^2 + z_2^2 + z_3^2 =\kappa^2-\tilde{m}^2\, ,$$ which describes the “deformed" mass-shell given by $$M^{\kappa}_m \equiv\{\gamma e_{p}: e_{p}\in B,\, \gamma\in SO(3,1) \}\, .$$ As for the undeformed mass shell described in Section II, the space $M^{\kappa}_m $ as the orbit of a symmetry group will have a natural geometrical interpretation as a homogenous space. The deformed one-particle Hilbert space will be built from the space of functions on such homogenous space $C^{\infty}(M^{\kappa}_m)$. As discussed above a choice of co-ordinates on $B$ is associated to a choice of basis of $U(\mathfrak{b})$ and to the hyperboloid above will correspond with an invariant mass Casimir operator $C_1(P)\in U(\mathfrak{b})$. Functions on the mass-shell $\phi \in C^{\infty}(M^{\kappa}_m)$ will thus satisfy the “wave equation" $$\label{waveq} C_1(P)\,\phi =m^2 \phi \,,$$ where $m^2=\tilde{m}^2-\kappa^2$. In particular (\[waveq\]) will hold for plane waves themselves. Notice that for any Lie group $G$ the space of complex valued functions square integrable w.r.t. the inner product defined using the Haar measure $d\mu (g)$ $$(f_1,f_2)=\int_G d\mu (g)\, \, \bar{f}_1(g)\, f_2(g)\, .$$ defines a Hilbert space. In our case, as functions on a homogeneous space we can define a natural invariant measure and a inner product on $C^{\infty}(M^{\kappa}_m)$ (see discussion in Section II) with the latter given by $$\label{kinner} (\phi_1,\phi_2)_{\kappa}=\int_B d\mu (p)\, \, \delta(\mathcal{C}_1(p))\,\, \bar{\phi}_1(p)\, \phi_2(p)\, .$$ Here $d\mu (p)$ is the left-invariant Haar measure on $T^*=B$ [@Freidel:2007hk] which in cartesian and flat slicing co-ordinates reads respectively $$\label{19a} d\mu \equiv \frac{1}{(2\pi)^4\,z_4}\, d z_0\, d^3\mathbf{z} = \frac{e^{3p_0/\kappa}}{(2\pi)^4}\, dp_0\, d^3\mathbf{p}\,.$$ To define a Hilbert space from $C^{\infty}(M_m)$ we need to find a criterion which ensures that the inner product (\[kinner\]) is positive definite. As discussed at length in Section III this entails the introduction of a complex structure on $C^{\infty}(M_m)$. Roughly speaking this corresponds to a choice of a “time-like" element of $P_0 \in U(\mathfrak{b})$ such that $$P_0 \, \phi(p)^{\pm}= \omega^{\pm}(p)\, \phi(p)^{\pm}\,,$$ i.e. the equivalent of an energy co-ordinate function on the homogenous space $M^{\kappa}_m$. The complex structure will be, as usual, given by $$\label{complexs} J=i \frac{P_0}{|P_0|}\, ,$$ (properly speaking such element is not in the UEA but in the “enveloping field" [@Barut:1986dd]) and, as in the undeformed case, can be used to define positive and negative energy projection operators.\ Now we come to our main point. In order to choose the energy operator $P_0$ from which we define the complex structure we need to make an explicit choice of basis in the commutative UEA $U(\mathfrak{t})$ with which we decompose the element $C_1(P)$. In ordinary local QFT the requirement of “local action" of a symmetry generator singles out a [*unique choice*]{} of basis of translation generators $P_0,\,P_i$ for which $C_1(P)=P_0^2-{\bf P}^2_i$. Indeed in this case a choice of cartesian co-ordinates on $C^{\infty}(\mathbb{R}^{3,1})$ will correspond to the set of basis elements $P_0,\,P_i$ of $U(\mathbb{R}^{3,1})$ for which $$\Delta P_{\mu} = P_{\mu} \otimes 1 + 1 \otimes P_{\mu}\, .$$ Elements of a UEA for which the co-product has such form are called “primitive". In everyday language the trivial form of the co-product above is telling us that primitive elements act according to the Leibnitz rule i.e. additively and thus are “local symmetry generators" (see [@Arzano:2007nx] for a detailed discussion).\ In our deformed setting the peculiarity of $U(\mathfrak{b})$ is that now [*there is no choice*]{} of a commuting set of primitive elements with which we decompose the Casimir. Indeed since the dual Hopf algebra of $U(\mathfrak{b})$ is, loosely speaking, the algebra of functions on the [*non-abelian*]{} group $B$ the co-product of $U(\mathfrak{b})$ [*no matter which basis we choose*]{} will be non co-commutative, namely $\sigma \circ \Delta \neq \Delta$ (where $\sigma (a\otimes b) = b \otimes a)$. In other words the action of translation generators will be non-Leibniz and non-symmetric for ANY choice of basis of $U(\mathfrak{b})$. This is the most profound and truly “basis independent" statement in the context of deformed relativistic symmetries. We thus conclude that there is no preferred choice of translation symmetry generators from which we can define an energy coordinate function on $M^{\kappa}_m$ and thus [*no preferred choice of complex structure*]{} in constructing the one-particle Hilbert space of a relativistic particle with curved momentum space.\ Note that in QFT in curved space one faces an analogous situation: in this case the ambiguity in the the definition of the complex structure $J$ comes from the fact that there is no global time-like Killing vector that can be used to define such object. In the most optimistic cases one has a preferred notion of the time-translation only in certain regions of space-time and this ultimately leads to particle production when one evolves from a region to another. Notice that while for us to allow for a generalization of the quantization formalism to curved momentum space we had to start with a phase space described in terms of co-adjoint orbits in QFT in curved space the starting point is the phase space of the field described as solutions of the equation of motion which is well defined on any global hyperbolic manifold (which can have no global symmetries at all). Field modes and vacuum fluctuations =================================== We now give a concrete realization of the deformed one-particle Hilbert space and introduce tools to describe the behaviour of deformed field modes. Let us focus on the choice of basis $P_0, P_i$ in $U(\mathfrak{b})$ related to the “flat slicing" co-ordinates (\[bicrossp\]) i.e. to the group splitting parameter $\beta=1$. The wave equation defining the mass-shell is given by the mass Casimir which for such choice of basis reads $$\mathcal{C}_{\kappa}(P)=\left(2\kappa \sinh\left(\frac{P_0}{2\kappa}\right)\right)^2-{\bf P}^2 e^{P_0/\kappa}\,.$$ For simplicity we focus on the massless case. For on-shell plane waves $e_{\bf p}\equiv \{e_p: \mathcal{C}_{\kappa}(P)\, e_p=0\}$, and in general of any function on $M^{\kappa}_{m}$, the generator $P_0$ will read off the energy coordinate $$P_0\,\, e_{\bf p} = \omega^{\pm}_{\kappa}({\bf p})\,\, e_{\bf p}\, ,$$ with $$\omega^{\pm}_{\kappa}({\bf p})=-\kappa \log\left(1\mp \frac{|{\bf p}|}{\kappa}\right) \, .$$ We can now use $P_0$ to define the complex structure (\[complexs\]) and the operator $P^+= 1/2 1- iJ$ to project a generic element of $C^{\infty}(M^{\kappa}_m)$ on the positive energy subspace $C^{\infty}(M^{\kappa +}_m)$. The inner product on such space given by $$(\phi_1,\phi_2)_{\kappa}=\int_{M^{\kappa +}_m} \frac{d\mu({\bf p})}{2\omega_{\kappa}({\bf p})}\,\, \bar{\phi}_1({\bf p})\, \phi_2({\bf p})\, ,$$ (we omitted for the $+$ superscripts for notational clarity), which can be written in covariant form as [@Arzano:2007ef] $$\label{sinner} (\phi_1,\phi_2)_{\kappa}=\int_{B} d\mu(p)\,\,\delta(\mathcal{C}_1(p))\, \theta(p_0)\, \overline{\tilde{\phi}}_1(p)\, \tilde{\phi}_2(p)\, ,$$ is indeed positive definite and thus turns $C^{\infty}(M^{\kappa +}_m)$ into our deformed one-particle Hilbert space $\mathcal{H}_{\kappa}$. Using the group Fourier transform discussed in Section II we can write the “space-time" counterpart of $ \phi({\bf p})\in C^{\infty}(M^{\kappa +}_m)$ $$\phi(x) = \int_{B} d\mu(p)\,\,\delta(\mathcal{C}_1(p))\, \theta(p_0)\, \tilde{\phi}(p)\, e_p(x) = \int_{M^{\kappa +}_m} \frac{d\mu({\bf p})}{2\omega_{\kappa}({\bf p})}\,\,\, \phi({\bf p})\, e_{{\bf p}}(x)$$ which shows how, due to the group nature of the plane waves $e_p(x)$, the fields $\phi(x)$ form a non-commutative algebra and thus the Fourier transformed version of elements of $\mathcal{H}_{\kappa}$ describe the one-particle Hilbert space of a [*non-commutative quantum field theory*]{}.\ For a practical description of the states $\mathcal{H}_{\kappa}$ we can introduce a normalized basis of delta functions[^4] which correspond to the “modes” of the on-shell plane waves $e_{\bf p}$ $$e_{{\bf p}}({\bf k})\equiv 2\omega_{\kappa}({\bf k}) \,\, \delta^3({\bf p}\oplus (\ominus {\bf k}))\, ,$$ where $\oplus$ and $\ominus$ denote respectively the (non-abelian) composition and antipode for spatial momenta which can be read off (\[copr\]) and (\[antip\]) and are explicitely given by $${\bf p}\oplus {\bf q} = {\bf p} + e^{\frac{-p^0}{\kappa}} {\bf q} \,,\,\,\,\,\,\,\,\,\,\,\,\,\,\, \ominus {\bf p} = - e^{\frac{p^0}{\kappa} } {\bf p}\, .$$ Introducing a bra-ket notation $e_{\bf p} \equiv |{\bf p} \rangle$ we have for the inner product of one-particle states [@Arzano:2007ef] $$\langle{\bf k_1}|{\bf k_2} \rangle \equiv (e_{\bf k_1},e_{\bf k_2})_{\kappa}=2\omega_{\kappa}({\bf k}_1) \,\, \delta^3({\bf k}_1\oplus (\ominus {\bf k}_2))\,.$$ Of course the same construction above can be repeated for any other choice of the group splitting parameter $\beta$. In this case the Hilbert space $\mathcal{H}^{\kappa}_{\beta}$ will be spanned by basis vectors $|{\bf k} \rangle_{\beta}$ bearing a different relation between energy and linear momentum through $\omega^{\beta}_{\kappa}({\bf k})$ and a different composition rule for the eigenvalues of the deformed translation generators $P^{\beta}_{\mu}$. Notice also that unlike the case of quantum fields in curved space the different Hilbert space constructions share the [*same*]{} vacuum state.\ Within the context of one-particle quantization we can proceed a step further and study the basic observables of the theory in order to get some insight on the vacuum structure and quantum fluctuations of the theory. One-particle observables will be given by the quantized counterpart of classical observables i.e. functions on phase space. The latter can be written in terms of the symplectic structure as $\mathcal{O}_{\phi}\equiv \omega(\phi, \cdot)$ with $\phi\in C^{\infty}(M^{\kappa}_m)$. Quantization of such observable gives the most general expression of the field operator $\hat{\mathcal{O}}_{\phi}\equiv\Psi(\phi)$ which for specific choices of $\phi$ reduces to the familiar field operator (see [@Bongaarts:1971cu] for a nice discussion). The one-particle creation and annihilation operator will be obtained upon quantization of the following functions on phase space $$\begin{aligned} a(\phi)(\cdot)&\equiv\frac{1}{2}\left(\omega(J\phi,\cdot)-i\omega(\phi,\cdot)\right)&=\,\langle\phi,\cdot \rangle\\\, , a^*(\phi)(\cdot)&\equiv\frac{1}{2}\left(\omega(J\phi,\cdot)+i\omega(\phi,\cdot)\right)&=\,\langle\cdot,\phi \rangle\, .\end{aligned}$$ In terms of the delta function basis written above we denote the quantized counterparts of such functions by $$a(\bar{e}_{\bf k})\equiv a({\bf k}) \, ,\,\,\,\,\,\,\,\, a^{\dagger}(e_{\bf k})\equiv a^{\dagger}({\bf k})\, ,$$ so that $$\label{aphi} a(\phi)= \int \frac{d\mu({\bf k})}{2\omega_{\kappa}({\bf k})}\, \phi({\bf k})\, a({\bf k})\,.$$ and $$\label{aphi+} a^{\dagger}(\phi)= \int \frac{d\mu({\bf k})}{2\omega_{\kappa}({\bf k})}\, \phi(\ominus{\bf k})\, a^{\dagger}({\bf k})\, ,$$ where the antipode in the last expression comes from the reality condition on the classical phase space element $\phi\in C^{\infty}(M^{\kappa}_m)$. The “generalized" filed operator can be written in terms of such creation and annihilation operators[^5] as $$\Psi(\phi)=i(a(\phi)-a^{\dagger}(\phi))\,,$$ and from $\Psi(\phi)$ we can write down the field mode operator or the quantum equivalent of the classical oscillator coordinate. Indeed using the expansions (\[aphi\]) and (\[aphi+\]) $$\Psi(\phi)=i \int \frac{d\mu({\bf k})}{2\omega_{\kappa}({\bf k})} \tilde{\phi}({\bf k}) \left(a({\bf k}) + \mathcal{J}_{\ominus}({\bf k})\, a^{\dagger}(\ominus{\bf k})\right)\, ,$$ with $\mathcal{J}_{\ominus}({\bf k})$ defined by $d\mu(\ominus {\bf k})= \mathcal{J}_{\ominus}({\bf k})d\mu({\bf k})$. We have for the Schroedinger picture field mode operator $$\hat{\varphi}_{\kappa}({\bf k})\equiv\frac{1}{2\omega_{\kappa}({\bf k})} (a({\bf k}) + \mathcal{J}_{\ominus}({\bf k})\, a^{\dagger}(\ominus{\bf k}))\,.$$ We can evolve $\hat{\varphi}_{\kappa}({\bf k})$ in time using the translation generator $P_0$ obtaining the field mode operator in the Heisenberg representation $$\hat{\varphi}_{\kappa}({\bf k},t)\equiv\frac{1}{2\omega_{\kappa}({\bf k})} (a({\bf k}) \exp(-i \omega_{\kappa}({\bf k})t)+ \mathcal{J}_{\ominus}({\bf k})\, a^{\dagger}(\ominus{\bf k})\exp(i \omega_{\kappa}({\bf k})t))\,.$$ We can now take the expectation value of the product of two mode-field operators above in the vacuum state $|0\rangle$ such that $a^{\dagger}({\bf k})|0\rangle\equiv |{\bf k}\rangle$ and $a({\bf k})|0\rangle\equiv 0\,\,\, \forall\,{\bf k}$. Thus we obtain the deformed equivalent of the spatial Fourier transform of the two-point function $$G_+({\bf k}_1,t;{\bf k}_2,s)\equiv\langle 0| \hat{\varphi}_{\kappa}({\bf k}_1,t) \hat{\varphi}_{\kappa}({\bf k}_2, s) |0\rangle =\frac{\delta^3({\bf k}_1\oplus{\bf k}_2)}{2\omega_{\kappa}({\bf k}_1)} \mathcal{J}_{\ominus}({\bf k}_1)\exp(-i\omega_{\kappa}({\bf k}_1)(t-s)) \, .$$ This provides us with the fundamental building block for $\kappa$-deformed field theory and for all the applications in which the two-mode point function plays a fundamental role.\ As an immediate application of the formalism introduced we can calculate the [*vacuum fluctuations*]{} of the field modes $\hat{\varphi}_{\kappa}({\bf k})$ which will be given by $$\delta\hat{\varphi}_{\kappa}({\bf k})=(\langle 0| \hat{\varphi}_{\kappa}({\bf k}) \hat{\varphi}^{\dagger}_{\kappa}({\bf k})|0\rangle)^{1/2}\sim \frac{\mathcal{J}_{\ominus}({\bf k})}{2\omega_{\kappa}({\bf k})} \, .$$ For the illustrative case of $\beta=1$ we have that $\mathcal{J}_{\ominus}({\bf k})=\exp(-3\omega_{\kappa}({\bf k})/\kappa)$ and thus $$\delta\hat{\varphi}_{\kappa}({\bf k})\rightarrow 0\, ,\,\,\,\,\,\,\,\,\ |{\bf k}|\rightarrow \kappa\, ,$$ i.e. quantum fluctuations [*freeze*]{} when the modulus of the linear momentum of the field mode approaches the value of the deformation parameter $\kappa$. Notice how such result heavily relies on the definition of linear modes for the field one is choosing. From this point of view the study of mode fluctuations seem to be a good candidate to establish, via some physical requirement, whether or not a “preferred" notion of field mode exist in the quantization procedure we outlined. Such question will be addressed in future work. Summary ======= We presented a detailed account of the quantization of a relativistic particle with momentum space given by a group manifold. This was done starting from a description of the phase space of the particle as a co-adjoint orbit of the relativistic symmetry group. The reason for adopting such formulation was twofold: on one side it is naturally connected with the description of the corresponding classical and quantum field theory spaces of states on the other hand allows for generalizations to models of relativistic particles with group valued momenta for which a notion of configuration space is less straightforward. We discussed how, in general, at the phase space level “curving" momentum space boils down to the introduction of a non-trivial Lie bracket on the dual Lie algebra of translations. In particular we considered the “group" momentum space associated with $\kappa$-deformations of the Poincaré algebra which is obtained by exponentiating the $\kappa$-Minkowski Lie brackets and which, as a manifold, is given by a sub-manifold of de Sitter space. Our analysis shows that, at least at the kinematical level, there is no effect of such deformations on the classical phase space of a single relativistic particle, a result which confirms what suggested in [@Arzano:2010kz].\ Effects of the deformation do indeed appear, and quite dramatically, at the quantum level. We recalled how a necessary step in the construction of a quantum Hilbert space from a classical field’s phase space is the introduction of a [*complex structure*]{} which defines the notion of positive and negative energy states. We showed that, for a deformed field theory related to a relativistic particle with curved momentum space, this step is non-trivial since it involves a choice of basis in the algebra of polynomials of the generators of deformed translations. As for field quantization in curved space-time, in a deformed setting one does not have a criterion to pick a preferred notion of energy (and linear momentum). This is to contrast with ordinary local quantum field theory in which such criterion exists and consists in picking a basis of translation generators which act according to the Leibniz rule on tensor product states i.e. whose momenta combine according to usual addition. Even though our discussion was limited to the example of $\kappa$-deformed momentum space, the conclusion we reach applies to [*any*]{} field theory with group valued momenta and in particular to the “quantum double" of the Lorentz group, a deformation of the Poincaré algebra relevant for relativistic particles coupled to three dimensional gravity [@Bais:2002ye].\ The tools introduced in the discussion of the quantization of the $\kappa$-deformed field theory were used in the last section to provide a concrete realization of a $\kappa$-one-particle Hilbert space. We defined the basic field observable of the theory and were able to explicitly derive the quantized mode operators which were used to write down the deformed two-point function in the linear momentum representation and the vacuum fluctuations of the modes, which, as expected, exhibit a non-trivial behaviour when their modulus gets closer to the (UV) deformation scale $\kappa$. This further step in the understanding the quantum properties of $\kappa$-deformed field theories finally opens the window to what we think are most promising applications of these models namely their use for investigating trans-planckian issues [@Corley:1996ar; @Martin:2000xs; @Parentani:2007dw] in semi-classical gravity from cosmology to black hole radiance. The author would like to thank Jerzy Kowalski-Glikman for many valuable discussions and encouragement and Tomasz Trzasniewski for comments on the last version of the paper. This research is supported by the EC’s Marie Curie Actions. I would like to thank the Perimeter Institute for Theoretical Physics for hospitality while part of this project was being carried out. Research at Perimeter Institute for Theoretical Physics is supported in part by the Government of Canada through NSERC and by the Province of Ontario through MRI. [99]{} P. C. W. Davies, “Particles Do Not Exist,” [*In \*Christensen, S.M. ( Ed.): Quantum Theory Of Gravity\*, 66-77, Adam Hilger Ltd. Bristol, 1984*]{} G. ’t Hooft, Class. Quant. Grav.  [**13**]{}, 1023 (1996) \[arXiv:gr-qc/9601014\]. H. J. Matschull and M. Welling, Class. Quant. Grav.  [**15**]{}, 2981 (1998) \[arXiv:gr-qc/9708054\]. D. Oriti, arXiv:0912.2441 \[hep-th\]. G. Amelino-Camelia and M. Arzano, Phys. Rev. D [**65**]{} (2002) 084044 \[arXiv:hep-th/0105120\]. M. Arzano and A. Marciano, Phys. Rev.  D [**76**]{}, 125005 (2007) \[arXiv:0707.1329 \[hep-th\]\]. M. Arzano and A. Marciano, Phys. Rev.  D [**75**]{}, 081701 (2007) \[arXiv:hep-th/0701268\]. J. Kowalski-Glikman and S. Nowak, Class. Quant. Grav.  [**20**]{}, 4799 (2003) \[arXiv:hep-th/0304101\]. L. Freidel and J. Kowalski-Glikman, PoS (QG-Ph) 023 \[arXiv:0710.2886 \[hep-th\]\]. E. Joung, J. Mourad and K. Noui, J. Math. Phys.  [**50**]{}, 052503 (2009) \[arXiv:0806.4121 \[hep-th\]\]. S. Meljanac and A. Samsarov, arXiv:1007.3943 \[hep-th\]. M. V. Battisti and S. Meljanac, Phys. Rev.  D [**82**]{}, 024028 (2010) \[arXiv:1003.2108 \[hep-th\]\]. F. Girelli and E. R. Livine, arXiv:0910.3107 \[hep-th\]. F. Girelli and E. R. Livine, arXiv:1004.0621 \[hep-th\]. Abraham, R., and J.E. Marsden, [*Foundation of Mechanics*]{}, Benjamin/Cummings Publishing Company, 1978. Kirillov, A.A., [*Elements of the Theory of Representations*]{}, Springer-Verlag, 1976. J. F. Carinena, J. M. Gracia-Bondia and J. C. Varilly, J. Phys. A [**23**]{}, 901 (1990). A. P. Balachandran, G. Marmo, B. S. Skagerstam and A. Stern, [*Singapore: World Scientific (1991) 358 p*]{} A. O. Barut and R. Raczka, “Theory Of Group Representations And Applications,” [*Singapore: World Scientific (1986)*]{} S. Sternberg, “Group theory and physics,” [*Cambridge, UK: Univ. Pr. (1994) 429 p*]{} P. J. M. Bongaarts, “Linear Fields According To I. E. Segal,” [*In \*London 1971, Mathematics Of Contemporary Physics\*, Academic Publishing, London 1972, 187-208*]{} A. Ashtekar and A. Magnon, Proc. Roy. Soc. Lond.  A [**346**]{}, 375 (1975). P. Panangaden, J. Math. Phys.  [**20**]{}, 2506 (1979). G. W. Gibbons and H. J. Pohle, Nucl. Phys.  B [**410**]{}, 117 (1993) \[arXiv:gr-qc/9302002\]. R. M. Wald, “Quantum field theory in curved space-time and black hole thermodynamics,” [*Chicago, USA: Univ. Pr. (1994) 205 p*]{} H. Ruegg and V. N. Tolstoi, Lett. Math. Phys.  [**32**]{}, 85 (1994) \[arXiv:hep-th/9406146\]. V. Chari and A. Pressley, “A Guide To Quantum Groups,” [*Cambridge, UK: Univ. Pr. (1994) 651 p*]{} J. Lukierski, A. Nowicki and H. Ruegg, Phys. Lett.  B [**293**]{}, 344 (1992). J. Kowalski-Glikman and S. Nowak, arXiv:hep-th/0411154. M. Arzano and J. Kowalski-Glikman, arXiv:1008.2962 \[hep-th\]. S. Majid and H. Ruegg, Phys. Lett.  B [**334**]{}, 348 (1994) \[arXiv:hep-th/9405107\]. P. Kosinski, J. Lukierski and P. Maslanka, Czech. J. Phys.  [**50**]{}, 1283 (2000) \[arXiv:hep-th/0009120\]. L. Freidel, J. Kowalski-Glikman and S. Nowak, Int. J. Mod. Phys.  A [**23**]{} (2008) 2687 \[arXiv:0706.3658 \[hep-th\]\]. M. Arzano, J. Kowalski-Glikman and A. Walkus, Class. Quant. Grav.  [**27**]{}, 025012 (2010) \[arXiv:0908.1974 \[hep-th\]\]. T. Tjin, Int. J. Mod. Phys.  A [**7**]{}, 6175 (1992) \[arXiv:hep-th/9111043\]. M. Arzano, PoS [**QG-PH**]{}, 005 (2007) \[arXiv:0711.3222 \[hep-th\]\]. M. Arzano, Phys. Rev.  D [**77**]{}, 025013 (2008) \[arXiv:0710.1083 \[hep-th\]\]. M. Arzano and D. Benedetti, Int. J. Mod. Phys.  A [**24**]{}, 4623 (2009) \[arXiv:0809.0889 \[hep-th\]\]. J. G. Bu, H. C. Kim, Y. Lee, C. H. Vac and J. H. Yee, Phys. Lett.  B [**665**]{}, 95 (2008) \[arXiv:hep-th/0611175\]. T. R. Govindarajan, K. S. Gupta, E. Harikumar, S. Meljanac and D. Meljanac, Phys. Rev.  D [**77**]{}, 105010 (2008) \[arXiv:0802.1576 \[hep-th\]\]. C. A. S. Young and R. Zegers, Nucl. Phys.  B [**809**]{}, 439 (2009) \[arXiv:0807.2745 \[hep-th\]\]; arXiv:0812.3257 \[math.QA\]. M. Arzano, arXiv:0909.4873 \[hep-th\]. F. A. Bais, N. M. Muller and B. J. Schroers, Nucl. Phys.  B [**640**]{}, 3 (2002) \[arXiv:hep-th/0205021\]. S. Corley and T. Jacobson, Phys. Rev.  D [**54**]{}, 1568 (1996) \[arXiv:hep-th/9601073\]. J. Martin and R. H. Brandenberger, Phys. Rev.  D [**63**]{}, 123501 (2001) \[arXiv:hep-th/0005209\]. R. Parentani, Class. Quant. Grav.  [**25**]{}, 154015 (2008) \[arXiv:0710.4664 \[hep-th\]\]. [^1]: In Minkowski space the integral is taken over a Cauchy surface $\Sigma_t$ at fixed time $t$ $$\label{qftinnprod} \omega(\phi_1,\phi_2)= \int_{\Sigma_t}(\phi_2\dot{\phi}_1-\phi_1\dot{\phi}_2)\, d^3\vec{x}\, .$$ [^2]: In [@Freidel:2007hk] it was argued that the action of Lorentz boosts on negative frequency plane waves could take their momentum out of the submanifold describing the Lie group $B$ thus breaking Lorentz symmetry. It was later observed by one of the authors of [@Freidel:2007hk], myself and a collaborator [@Arzano:2009ci] that the correct way of handling the action of Lorentz generators on such antiparticle states is via their “antipode" (see (\[antipode\]) below). In this way the particle/antiparticle structures and Lorentz symmetry are fully consistent. [^3]: Recall even if the action of the Lorentz group on $B$ is not a representation the action on the space of functions on $B$ does provide a representation. [^4]: Recall that the Dirac delta for functions on a group $G$ is such that $$\int_G d\mu(g) \delta(g) f(g) = f(e)\,,\,\,\,\,\,\,\,\, \int_G d\mu(g) \delta(gh^{-1}) f(g)=f(h)\, ,$$ where $g,h \in G$ and $e$ is the unit element which in the notation used in the preceding Sections this reads $$\int_B d\mu(p) \delta(p) f(p) = f(0)\,,\,\,\,\,\,\,\,\, \int_B d\mu(p) \delta(p\oplus(\ominus q)) f(p)=f(q)\, .$$ [^5]: Let us remark here that, as widely discussed in the literature [@Arzano:2007ef; @Arzano:2008bt; @Bu:2006dm; @Govindarajan:2008qa; @Young:2008zm], the extension of the creation and annihilation operators defined above to the multi-particle sector of the theory is highly non-trivial. In fact in the construction of a deformed Fock space the non-symmetric nature of the co-product requires a “momentum-shifting" symmetrization [@Arzano:2007ef; @Arzano:2009wp]. The existence of a covariant deformed symmetrization procedure depends on the availability of an operator known as quantum $R$-matrix (see [@Arzano:2008bt; @Young:2008zm] for an extended discussion) whose explicit construction for the $\kappa$-Poincaré algebra has been a topic of various studies without a commonly agreed outcome. We should notice however that our analysis goes beyond the illustrative example of $\kappa$-deformation and, for example, would also apply to the case of deformed relativistic symmetries described by the so-called Lorentz double [@Bais:2002ye]. For such models one has a rather straightforward definition of $R$-matrix and thus, in principle, no obstacles in the construction of a consistent Fock space.
{ "pile_set_name": "ArXiv" }
--- abstract: | Attribute-Based Encryption (ABE) is a powerful cryptographic tool that allows fine-grained access control over data. Due to its features, ABE has been adopted in several applications, such as encrypted storage or access control systems. Recently, researchers argued about the non acceptable performance of ABE when implemented on mobile devices. Indeed, the non feasibility of ABE on mobile devices would hinder the deployment of novel protocols and services–that could instead exploit the full potential of such devices. However, we believe the conclusion of non usability was driven by a not-very efficient implementation. In this paper, we want to shine a light on this concern by studying the feasibility of applying ABE on smartphone devices. In particular, we implemented [<span style="font-variant:small-caps;">AndrABEn</span>]{}, an ABE library for Android operating system. Our library is written in the C language and implements two main ABE schemes: Ciphertext-Policy Attribute-Based Encryption, and Key- Policy Attribute-Based Encryption. We also run a thorough set of experimental evaluation for [<span style="font-variant:small-caps;">AndrABEn</span>]{}, and compare it with the current state-of-the-art (considering the same experimental setting). The results confirm the possibility to effectively use ABE on smartphone devices, requiring an acceptable amount of resources in terms of computations and energy consumption. Since the current state-of-the-art claims the non feasibility of ABE on mobile devices, we believe that our study (together with the [<span style="font-variant:small-caps;">AndrABEn</span>]{} library that we made available online) is a key result that will pave the way for researchers and developers to design and implement novel protocols and applications for mobile devices. author: - | Moreno Ambrosin\ \ Mauro Conti [^1]\ \ Tooska Dargahi\ \ \ bibliography: - 'ref.bib' title: | On the Feasibility of Attribute-Based Encryption\ on Smartphone Devices --- Introduction {#sec:intro} ============ Attribute-Based Encryption (ABE) is a public key encryption scheme first introduced in 2005 by Sahai and Waters [@sahai2005fuzzy]. In this scheme, both encryption and decryption are based on [*attributes*]{} (e.g., age, gender, or job position), that can be either related to the private keys of the users, or to the ciphertext. A user can restrict access to a specific piece of data by defining an [*access policy*]{}. As an example, an access policy can be expressed as a boolean expression such as $(A \wedge B)\vee C$, where $A$, $B$ and $C$ are attributes and the possible values for attributes are implicitly $true$ or $false$. Researchers proposed two main types of ABE schemes, namely Key-Policy Attribute-Based Encryption (KP-ABE) [@goyal2006attribute] and Ciphertext-Policy Attribute-Based Encryption (CP-ABE) [@bethencourt2007ciphertext]. Compared to the other encryption approaches, ABE presents several advantages [@lee2013survey]. First, it allows the data owner to apply a fine grained access control over data, based on attributes and policies. Second, ABE schemes are scalable and independent of the number of authorized users. Moreover, ABE by construction is resilient against collusion attacks. Finally, while traditional public key infrastructures impose a noticeable communication and storage overhead due to the exchange of cryptographical material, using ABE the data owner can encrypt data by using a set of attributes, without exchanging any certificates or identifying the client [@lee2013survey]. Several factors influence the performance of ABE in real applications, such as the number of attributes used for defining an access policy, the desired security level, and the capabilities of the underlying device, in terms of available memory and CPU speed. Some researchers have already studied the feasibility of using ABE on mobile devices [@ABE_icc_2014]; however, most of the existing studies do not consider all the factors, or actually do not implement ABE on smartphone. In [@ABE_icc_2014], Wang et al. evaluated the performance of CP-ABE and KP-ABE on laptop and smartphone devices. They implemented both schemes by Java language, and evaluated different metrics. The authors concluded that the ABE performance is unacceptable on Android smartphone. However, the usefulness of ABE in mobile applications is evident, and its non feasibility on such devices would be a big obstacle to deployment of new services and to benefit from its advantages. Therefore, obtaining acceptable performance is the biggest challenge for guaranteeing the use of ABE on resource constraint devices. #### Contribution The contribution of this paper is a comprehensive and careful study of the feasibility of ABE operations on Android smartphone devices. In particular, differently from what claimed in [@ABE_icc_2014], we show that it is possible to achieve reasonable performance for the main ABE operations, even on smartphone devices. We provide [<span style="font-variant:small-caps;">AndrABEn</span>]{}, an implementation of CP-ABE [@bethencourt2007ciphertext] and KP-ABE [@goyal2006attribute] as a C library for Android smartphones. We integrated such library into Android using the Android Native Development Kit (NDK) tool. We evaluated our implementation, and compared its performance with the Java-based study and implementation proposed by Wang et al. in their work presented at ICC 2014 [@ABE_icc_2014], considering the same experimental settings[^2]. The results of our thorough evaluation show that the performance of our solution is an order of magnitude higher than the one in [@ABE_icc_2014]. Accordingly, our results prove that applying ABE is indeed feasible on current mobile devices, such as Samsung Galaxy Nexus smartphone—on which we ran our experiments. Finally, we made the [<span style="font-variant:small-caps;">AndrABEn</span>]{} library freely available [@andraben_impl] for researchers and developers. #### Organization The rest of the paper is organized as follows. In Section \[sec:related\] we present some related work. In Section \[sec:background\] we introduce the preliminaries and background on ABE. In Section \[sec:our-implementation\], we present [<span style="font-variant:small-caps;">AndrABEn</span>]{}, our proposed implementation for CP-ABE and KP-ABE, and provide an analysis of its performance. We also discuss and compare our results with the ones in [@ABE_icc_2014]. Finally, in Section \[sec:conclusion\] we draw our conclusions. Related work {#sec:related} ============ Due to the increasing use of mobile devices, evaluating the performance of cryptographic algorithms on mobile devices is an important issue that has been considered in several research studies [@tillich2004survey; @kawahara2006efficient; @voyiatzis2011increasing; @braga2012portability]. As an example, Braga and Nascimento [@braga2012portability] evaluated the feasibility of cryptographic algorithms on Android smartphone devices. They assessed the portability of cryptographic libraries on a Samsung i9100 and measured the performance of applying these libraries on the Android devices; however, in their analysis they did not consider ABE. Recently, the concept of ABE has been used in various schemes to deal with data confidentiality, privacy and access control issues. Unfortunately, most of these research studies did not evaluate the actual feasibility of adopting ABE in their proposed approaches. However, few researchers focused on such assessment; we discuss some of them in the following. In [@Baden:2009:POS:1594977.1592585], Baden et al. presented Persona, an Online Social Network, where users are able to hide their personal information. They provided privacy by encrypting data with ABE. They implemented their solution on a first generation iPhone device, by cross-compiling the [cpabe]{} library [@cpabe] and its dependencies for the iPhone SDK 2.2.1. They obtained an average decryption and encryption time of 254 ms and 926 ms respectively, considering access structures containing one to five attributes. However, the authors did not provide an implementation of ABE on the Android smartphone. Along the same line of studies, we also proposed a system for efficient software updates distribution, over untrusted distribution networks in [@Moreno_ESORICS]. We used CP-ABE to guarantee flexible access control. In this work, we only provided evaluations measured on a laptop device with two 2.4 GHz Intel Core 2 Duo CPUs. Based on our experimental results, the CP-ABE encryption and decryption time for five attributes are 77.47 ms and 32.62 ms, respectively. We realized that not providing a proper feasibility assessment is a big limitation for this type of proposals. In this paper, we aim to fill this gap and pave the way for developing further solutions assuming the feasibility of ABE for mobile devices. A study similar to the one we are going to discuss in this paper was carried out in [@ABE_icc_2014]. The authors evaluated the performance of CP-ABE and KP-ABE in terms of execution time, data overhead, energy consumption, CPU and memory usage. They implemented these two ABE schemes using Java on a laptop with a 1.60 GHz Intel Quad-Core i7 2677M CPU and a smartphone runs Android 4.04 with a 1.60 GHz Intel Atom Z2460. The authors stated that applying ABE on Android smartphone devices is not practical with acceptable performance. In this paper, we show that this conclusion mostly depends on the specific implementation provided in [@ABE_icc_2014], and it does not hold in general. Finally, in [@green2011outsourcing] Green et al. proposed an alternative approach for efficient ABE decryption. In their solution, a part of the ABE decryption is outsourced to a third party cloud, highly reducing the load on the client device. However, while representing a good option to facilitate the ABE operations on devices with limited resources, this solution requires additional resources compared to in-device decryption, such as a third-party cloud entity, as well as Internet connectivity. To the best of our knowledge, we are the first that show the reasonable performance of ABE on Android devices, and provide a publicly available implementation for Android smartphone devices [@andraben_impl]. Indeed, existing ABE implementations for Android showed a high computation overhead on mobile platforms [@6654173; @ABE_icc_2014]. We prepared and cross-compiled two ABE C libraries, to be used on the Android mobile devices, thus proving the feasibility of such schemes on this platform. While we leave this as a future work, we expect that [<span style="font-variant:small-caps;">AndrABEn</span>]{} can be easily extended for other mobile devices and Internet of Things (IoT) devices. Background on ABE {#sec:background} ================= This section provides the fundamentals of Key-Policy ABE (KP-ABE) [@goyal2006attribute], and Ciphertext-Policy ABE (CP-ABE) [@bethencourt2007ciphertext]. Both CP-ABE and KP-ABE, are public key schemes. In a KP-ABE scheme, the data owner encrypts the data specifying a set of attributes. Each user owns a private key $D$ that reflects a specific policy. She will be able to decrypt a ciphertext if and only if the attributes embedded into the ciphertext satisfy the policy in $D$. It consists of four functions: - [**Setup.**]{} It takes as input an implicit security parameter and outputs the public parameter ${pk^{ABE}}$, and a master key ${mk^{ABE}}$. - [**Encryption.**]{} It takes as input a message $M$, a set of attributes $\gamma$, and the public parameter ${pk^{ABE}}$, and outputs the ciphertext $E$. - [**KeyGen.**]{} It takes as input an access policy $A$, the master key ${mk^{ABE}}$ and the public parameter ${pk^{ABE}}$. It outputs a decryption key $D$ reflecting the given policy. - [**Decryption.**]{} It takes as input the ciphertext $E$ that is encrypted under the set of attributes $\gamma$; the decryption key $D$, that represents the access policy $A$; and the public parameter ${pk^{ABE}}$. It outputs the message $M$ if and only if $\gamma$ “satisfies” the access policy $A$. Different from KP-ABE, in the CP-ABE scheme, the data owner encrypts her data enforcing an access policy. Users are provided with private keys representing a set of attributes. Only users having attributes that satisfy an access policy will be able to decrypt the ciphertext. A CP-ABE scheme provides the following functions: - [**Setup.**]{} It takes as input an implicit security parameter and outputs the public parameter ${pk^{ABE}}$, and a master key ${mk^{ABE}}$. - [**Encryption.**]{} It takes as input a message $M$, an access policy $A$, and the public parameter ${pk^{ABE}}$, and outputs the ciphertext $E$. - [**KeyGen.**]{} It takes as input a set of attributes $\gamma$, the master key ${mk^{ABE}}$ and the public parameter ${pk^{ABE}}$. It outputs a decryption key $D$ reflecting the given attributes. - [**Decryption.**]{} It takes as input the ciphertext $E$ that is encrypted under the access policy $A$; the decryption key $D$ representing a set of attributes $\gamma$; and the public parameter ${pk^{ABE}}$. It outputs the message $M$ if and only if $\gamma$ “satisfies” the access policy $A$. Similar to other pairing-based schemes, the complexity of CP-ABE and KP-ABE depends on the number of exponentiations and pairing operations performed by each of their algorithms [@li2010data]. In the CP-ABE scheme [@bethencourt2007ciphertext], the efficiency of the [**KeyGen**]{} algorithm depends on the number of attributes to be applied to the newly generated key. Herein, the algorithm performs two exponentiations for each attribute. Similarly, the [**Encryption**]{} operation requires two exponentiations for each attribute in the specified policy. The same complexity is required also by the [**KeyGen**]{} and [**Encryption**]{} operations in the KP-ABE scheme [@goyal2006attribute]. However, the efficiency of the [**Decryption**]{} function for CP-ABE mainly depends on how the policy enforced on the ciphertext and on the private key used for its decryption. This makes an estimation of the complexity of such operation a non trivial task [@bethencourt2007ciphertext]. The same holds for the KP-ABE [**Decryption**]{} operation, which strongly depends on the attributes set and the access policy specified in the ciphertext and the private key, respectively. ANDRABEN: Implementation and Analysis {#sec:our-implementation} ===================================== In this section, we provide an in-depth analysis of the performance of [<span style="font-variant:small-caps;">AndrABEn</span>]{} (Section \[sec:performance\]), and a comparison with the implementation in [@ABE_icc_2014] (Section \[sec:discussion\]). Performance Evaluation {#sec:performance} ---------------------- Our ABE implementation [@andraben_impl] comprises two libraries: the [cpabe]{} library [@cpabe], which implements the scheme proposed by Bethencourt et al. in [@bethencourt2007ciphertext], and a publicly available custom implementation [@kpabe_impl] of the KP-ABE scheme proposed by Goyal et al. in [@goyal2006attribute]. The original code has been slightly modified, in order to be integrated into Android mobile devices. Both libraries employ Type $A$ pairings provided by the PBC library [@libpbc]. Type $A$ pairings are built on top of an elliptic curve: $y^2 = x^3 + x$ over a finite field $F_q$, for some prime $q=3~mod~4$, and have a fixed embedding degree $k = 2$ [@pbc-thesis]. Therefore, the security strength of the scheme can be tuned by modifying two parameters: the size of the field $q$, and the prime order $r$ of the base point $P\in E(F_q)$ [@ABE_icc_2014]. Table \[tbl:table\_strength\] shows the security level of both CP-ABE and KP-ABE schemes, according to [@brown2001software]. ----------------------------- ---------------------------------------- -- -- [**Security level bits**]{} **80 & **112 & **128\ bit length of $r$ & 160 & 224 & 256\ bit length of $q$ & 512 & 1024 & 1536\ ****** ----------------------------- ---------------------------------------- -- -- : Security strength of CP-ABE and KP-ABE.[]{data-label="tbl:table_strength"} [<span style="font-variant:small-caps;">AndrABEn</span>]{} is implemented on Android 4.3, “Jelly Beam”. We carried out our experimental evaluation on a Samsung Galaxy Nexus device (1.2 GHz dual-core ARM Cortex-A9 CPU, 1 GB RAM). For completeness and comparison, we also tested our libraries on a laptop device (Ubuntu 14.4 LTS, 1.8 GHz 4x Intel Core i7-4500U CPU, 8 GB RAM). We evaluated [**KeyGen**]{}, [**Encryption**]{} and [**Decryption**]{} operations, varying the number of attributes adopted from one to 30. We consider this range to be representative enough for a wide range of real world applications of ABE [@ABE_icc_2014]. We tested both ABE schemes with security levels of 80, 112 and 128 bits, and measured average execution time, CPU and memory utilization, and energy consumption on mobile devices. Note that, as in other public key schemes, in ABE the actual encryption of the ciphertext is performed by means of a symmetric key, which is in turn encrypted with the public key. Therefore, we evaluate both encryption and decryption operations performed on a symmetric key. This makes our analysis independent from the size of the ciphertext. #### Execution Time Figure \[fig:time\_android\] presents the average time overhead for [**Encryption**]{} and [**Decryption**]{} operations for both CP-ABE and KP-ABE schemes. The results are presented for both Android and Laptop devices. They have been obtained as an average of 100 executions for each operation, varying the number of employed attributes and adopting different levels of security. [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](cpabe_keygen "fig:"){width=".95\columnwidth"} [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](kpabe_keygen "fig:"){width=".95\columnwidth"} \ [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](cpabe_enc "fig:"){width=".95\columnwidth"} [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](kpabe_enc "fig:"){width=".95\columnwidth"} \ [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](cpabe_dec "fig:"){width=".95\columnwidth"} [[0.49]{}]{} ![Average execution time (and std. deviation in errorbar) for main CP-ABE and KP-ABE algorithms.[]{data-label="fig:time_android"}](kpabe_dec "fig:"){width=".95\columnwidth"} As we can see in Figure \[fig:time\_android\], in general, the time required to perform each operation depends directly on the number of attributes that are used. Adopting a security value of 80 bits (which is reasonable for several medium-level security applications [@cryptoeprint:2009:389]) the CP-ABE [**Encryption**]{} operation remains under 4 s on the Android smartphone. Similarly, the CP-ABE [**KeyGen**]{} operation requires less than 2 s to be executed. However, adopting a security level of 112 or 128 bits, the time overhead imposed by CP-ABE on Android smartphone is much higher, while we argue being still usable for non-interactive applications (e.g., encrypted data to be uploaded to a cloud storage service). Indeed, in such applications the encryption could be carried out in a background process. For KP-ABE, however, the time required to perform the various operations is lower, and even the adoption of a security level of 128 bits is feasible on Android smartphone. On laptop, the evaluation results confirm the practicality of both CP-ABE and KP-ABE schemes, requiring a reasonable time for both [**KeyGen**]{} ($<\!2$ s), and [**Encryption**]{} ($<\!2$ s) operations. #### CPU utilization We measured CPU utilization on Android smartphone by collecting the required information from the system files [/proc/stat]{}, and [/proc/\[pid\]/stat]{}, where [pid]{} is the id of the application’s process. The CPU utilization remains under $50\%$ for each of the three operations, for both CP-ABE and KP-ABE, i.e., the operations fully utilize one of the two CPUs provided by the underlying platform. #### Memory utilization We measured the average memory space required by [<span style="font-variant:small-caps;">AndrABEn</span>]{}, adopting a range between one to 30 attributes. We realized that our implementations utilize between $13.5$ and $14.5$ MBytes of RAM space. We argue that such amount is acceptable for modern smartphones such as the Samsung Galaxy Nexus—used in our experiments. [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](cpabe_keygen_RAM "fig:"){width=".95\columnwidth"} [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](kpabe_keygen_RAM "fig:"){width=".95\columnwidth"} \ [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](cpabe_enc_RAM "fig:"){width=".95\columnwidth"} [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](kpabe_enc_RAM "fig:"){width=".95\columnwidth"} \ [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](cpabe_dec_RAM "fig:"){width=".95\columnwidth"} [0.49]{} ![Average RAM usage for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_RAM"}](kpabe_dec_RAM "fig:"){width=".95\columnwidth"} To better understand the behavior of [<span style="font-variant:small-caps;">AndrABEn</span>]{} in large scale scenarios, we also measured the average memory consumption employing 10, 100 and 1000 attributes. Figure \[fig:android\_RAM\] shows the obtained results. As expected, the amount of required RAM grows with the number of employed attributes. Here, one of the main advantages of running CP-ABE and KP-ABE on native code is the possibility to manage the heap usage directly, without relying on the Dalvik VM. Indeed, in such case the use of the expensive garbage collector significantly slows down the overall execution [@6654173]. #### Energy Consumption Energy consumption is a major concern in mobile devices. Therefore, a desirable implementation of a cryptographic tool for mobile devices, should consume as less energy as possible. We measured the average energy consumption for each of the CP-ABE and KP-ABE operations, by using the well-known PowerTutor Android application [@powerTutor]. Figure \[fig:android\_battery\] shows the obtained results. As we can see, with a security level of 80 and 112 bits, the energy required by both schemes remains low, making their use suitable on smartphones. [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](cpabe_keygen_battery "fig:"){width=".95\columnwidth"} [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](kpabe_keygen_battery "fig:"){width=".95\columnwidth"} \ [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](cpabe_enc_battery "fig:"){width=".95\columnwidth"} [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](kpabe_enc_battery "fig:"){width=".95\columnwidth"} \ [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](cpabe_dec_battery "fig:"){width=".95\columnwidth"} [0.49]{} ![Average energy consumption for CP-ABE and KP-ABE on Android smartphone.[]{data-label="fig:android_battery"}](kpabe_dec_battery "fig:"){width=".95\columnwidth"} Discussion {#sec:discussion} ---------- In what follows, we provide a brief comparison of our proposal, [<span style="font-variant:small-caps;">AndrABEn</span>]{}, against the implementation proposed in [@ABE_icc_2014], on the Android platform. Unfortunately, the code used in [@ABE_icc_2014] does not seem to be publicly available. Moreover, in [@ABE_icc_2014], the authors performed their evaluation on a device (and a specific processor in smartphone) which is not easily available anymore (an Android smartphone with a 1.60 GHz Intel Atom Z2460 processor and 1 GB RAM). For these reasons, to compare with the implementation proposed in [@ABE_icc_2014], we can only rely on the numbers reported in [@ABE_icc_2014] itself. Most of the discussion that follows is based on this approach. Moreover, for a further validation we also re-implemented the solution proposed in [@ABE_icc_2014] and performed some additional comparison. As a side note we observe that, since our implementation is single threaded, it does not take advantage from the dual-core CPU of the device we used for our measurements. Furthermore, both the device we adopted and the one used in [@ABE_icc_2014] are equipped with the same RAM memory (i.e., 1 GB). As shown in Figure \[fig:time\_android\], the execution time for both CP-ABE and KP-ABE with [<span style="font-variant:small-caps;">AndrABEn</span>]{} is significantly lower compared to the results reported in [@ABE_icc_2014]. In particular, the CP-ABE key generation with [<span style="font-variant:small-caps;">AndrABEn</span>]{} requires less than 30 s, while with the implementation proposed in [@ABE_icc_2014], key generation requires around 200 s. Similarly, [<span style="font-variant:small-caps;">AndrABEn</span>]{} performs CP-ABE encryption and decryption in less than 30 s and 20 s, respectively, while in the implementation proposed in [@ABE_icc_2014], encryption and decryption operations take on average 70 s and 80 s, respectively. Moreover, the average execution time reported in [@ABE_icc_2014] for all the three main KP-ABE operations, considering 26 attributes and a security level of 128 bits, is $\approx45$ s for encryption, while decryption and key generation operations require between $\approx90$ s and $\approx110$ s. Instead, our [<span style="font-variant:small-caps;">AndrABEn</span>]{} implementation of KP-ABE requires a considerably lower execution time for each of the main operations, i.e., 12 s for decryption (Figure \[kpabe\_dec\_android\]), and 3 s for encryption (Figure \[kpabe\_enc\_android\]) and key generation (Figure \[kpabe\_keygen\_android\]). In order to compare the energy consumption of [<span style="font-variant:small-caps;">AndrABEn</span>]{} (which is illustrated in Figure \[fig:android\_battery\]) with the implementation proposed in [@ABE_icc_2014] for CP-ABE, let us consider 10 attributes and a security level of 128 bits. With [<span style="font-variant:small-caps;">AndrABEn</span>]{}, each of the main CP-ABE operations, i.e., key generation, encryption and decryption, consumes almost 5 J, while with the implementation proposed in [@ABE_icc_2014], key generation consumes between 70 J and 100 J, and encryption and decryption operations consume 30 J and 40 J, respectively. Furthermore, compared to [<span style="font-variant:small-caps;">AndrABEn</span>]{}, also the KP-ABE implementation of key generation, encryption and decryption provided in [@ABE_icc_2014] require a considerably higher amount of energy. Indeed, while our implementation of the three main KP-ABE operations requires less than 5 J, the implementation in [@ABE_icc_2014] requires between 15 J and 40 J. [0.49]{} ![Comparison between the CP-ABE implementation of [<span style="font-variant:small-caps;">AndrABEn</span>]{} and in [@ABE_icc_2014].[]{data-label="fig:comparison"}](jcpabe_keygen "fig:"){width=".95\columnwidth"} [0.49]{} ![Comparison between the CP-ABE implementation of [<span style="font-variant:small-caps;">AndrABEn</span>]{} and in [@ABE_icc_2014].[]{data-label="fig:comparison"}](jcpabe_enc "fig:"){width=".95\columnwidth"} [0.49]{} ![Comparison between the CP-ABE implementation of [<span style="font-variant:small-caps;">AndrABEn</span>]{} and in [@ABE_icc_2014].[]{data-label="fig:comparison"}](jcpabe_dec "fig:"){width=".95\columnwidth"} The comparison reported up to this point considered directly the performance results reported in [@ABE_icc_2014]. However, for a further validation we also developed our own Java-based implementation of the CP-ABE scheme in [@bethencourt2007ciphertext], following the specifications reported in [@ABE_icc_2014]. We evaluated our Java implementation (which we believe being similar to the one used in [@ABE_icc_2014], for which code is not available) on the same device we used to evaluate [<span style="font-variant:small-caps;">AndrABEn</span>]{}. Due to space limitations, we only provide a comparison of the different execution time for the two solutions, which is presented in Figure \[fig:comparison\]. As we can see, the execution time for CP-ABE with [<span style="font-variant:small-caps;">AndrABEn</span>]{} is significantly lower compared to the results obtained with our Java based implementation, that in turn presents results that are consistent with the ones provided in [@ABE_icc_2014]. As an example, let us consider the execution time obtained with 25 attributes and a security level of 128 bits. Our CP-ABE [<span style="font-variant:small-caps;">AndrABEn</span>]{} implementation requires $\approx$25 s to perform the key generation, while the Java-based implementation requires $\approx$360 s to perform the same task. Similarly, the CP-ABE encryption and decryption operations with [<span style="font-variant:small-caps;">AndrABEn</span>]{} require on average $\approx$26 s and $\approx$19 s, while our Java-based implementation performs encryption and decryption in $\approx$340 s and $\approx$172 s, respectively. Overall, we can conclude that, compared with the approach discussed in [@ABE_icc_2014], [<span style="font-variant:small-caps;">AndrABEn</span>]{} provides significantly better performance, in terms of execution time, memory and CPU usage, and energy consumption. Conclusion {#sec:conclusion} ========== With the increasing use of cloud environment and smart devices connected to the Internet of Things, exchanged data confidentiality and access control to the stored data become a challenging issue. Attribute-Based Encryption is one of the best solutions that can be used to satisfy users privacy concerns [@ABE_icc_2014]. However, its performance on resource constraint devices is a challenging issue, and still represents a big concern for researchers willing to use ABE to develop novel privacy-preserving and access control solutions for such devices. In this paper, we studied the feasibility of applying ABE on smartphone devices and presented [<span style="font-variant:small-caps;">AndrABEn</span>]{}, an implementation of ABE in C language. We also provided a comparative analysis with a similar research study [@ABE_icc_2014] in which the authors proposed a Java-based implementation of ABE for Android smartphone. Based on the results of our thorough experiments, we conclude that using ABE on Android smartphones and similar devices is feasible. The evidence that we bring in this paper will be a reference for applicability of ABE in resource-constrained devices. [^1]: Mauro Conti is supported by a European Marie Curie Fellowship (N. PCIG11-GA-2012-321980). This work is also partially supported by the Italian MIUR PRIN Project TENACE (N. 20103P34XC), and the University of Padua PRAT 2014 Project on Mobile Malware. [^2]: We re-implemented the proposal in [@ABE_icc_2014] due to the unavailability of the original source code.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present results obtained from spectroscopic observations of red giants located in the fields of the Large Magellanic Cloud (LMC) globular clusters (GCs) NGC1928 and NGC1939. We used the GMOS and AAOmega+2dF spectrographs to obtain spectra centred on the CaII triplet, from which we derived individual radial velocities (RVs) and metallicities. From cluster members we derived mean RVs of RV$_{\rm NGC\,1928}=249.58\pm$4.65 km/s and RV$_{\rm NGC\,1939}=258.85\pm$2.08 km/s, and mean metallicities of \[Fe/H\]$_{\rm NGC\,1928}=-1.30\pm$0.15 dex and \[Fe/H\]$_{\rm NGC\,1939}=-2.00\pm$0.15 dex. We found that both GCs have RVs and positions consistent with being part of the LMC disc, so that we rule out any possible origin but that in the same galaxy. By computing the best solution of a disc that fully contains each GC, we obtained circular velocities for the 15 known LMC GCs. We found that 11/15 of the GCs share the LMC rotation derived from $HST$ and $Gaia$ DR2 proper motions. This outcome reveals that the LMC disc existed since the very early epoch of the galaxy formation and experienced the steep relatively fast chemical enrichment shown by its GC metallicities. The four remaining GCs turned out to have circular velocities not compatible with an [*in situ*]{} cluster formation, but rather with being stripped from the SMC.' author: - | A.E. Piatti$^{1,2}$[^1], N. Hwang$^{3}$, A.A. Cole$^{4}$, M.S. Angelo$^{5}$ and B. Emptage$^{4}$\ $^{1}$Consejo Nacional de Investigaciones Científicas y Técnicas, Av. Rivadavia 1917, C1033AAJ, Buenos Aires, Argentina\ $^{2}$Observatorio Astronómico de Córdoba, Laprida 854, 5000, Córdoba, Argentina\ $^{3}$Korea Astronomy and Space Science Institute, 776 Daedeokdae-Ro Yuseong-Gu, Daejeon 305-348, Korea\ $^{4}$School of Natural Sciences, University of Tasmania, Private Bag 37, Hobart, 7001 TAS, Australia\ $^{5}$Laboratório Nacional de Astrofísica, R. Estados Unidos 154, 37530-000 Itajubá, MG, Brazil\ date: 'Accepted XXX. Received YYY; in original form ZZZ' title: Accurate radial velocity and metallicity of the Large Magellanic Cloud old globular clusters NGC1928 and NGC1939 --- \[firstpage\] galaxies: individual: LMC – galaxies: star clusters: general Introduction ============ Only fifteen old GCs (GCs, ages $\ga$ 12 Gyr) are known to survive in the Large Magellanic Cloud (LMC) [@pg13], of which NGC1928 and NGC1939 have only recently been added by @detal99 [hereafter D99]. Their first colour-magnitude diagrams come from $HST$ photometry [@mg04], confirming their old ages. As far as we are aware, neither NGC1928 nor NGC1939 have published accurate metallicity or radial velocity (RV) measurements. The orbital motions of LMC ancient GCs are satisfactorily described by a disc-like rotation with no GC appearing to have halo kinematics [@shetal10]. @s92 found that these clusters form a disc that agrees with the parameters of the optical isophotes and inner HI rotation curve. There are some other galaxies that appear to have GC systems with kinematic properties related to the HI discs [e.g. @olsenetal2004], which might suggest a benign evolutionary history, such as might be expected if the LMC has evolved in a low density environment. However, the destruction of a GC system that is on a coplanar orbit about a larger galaxy could also produce such a disc-like rotation geometry [@leamanetal2013]. Furthermore, @vdbergh2004 showed that the possibility that the LMC old GCs formed in a pressure-supported halo, rather than in a rotating disc, should not be discarded. In this sense, @carreraetal2008 argued that the lack of evidence of such a hot stellar halo in the LMC is related to a low contrast of the halo population with respect to that of the disc, particularly at the innermost galactocentric radii where NGC1928 and NGC1939 are located. On the other hand, @carpinteroetal2013 modelled the dynamical interaction between the Small Magellanic Cloud (SMC) and the LMC, and found that at least some of the oldest clusters observed in the LMC could have originated in the SMC. The LMC old GCs have also been compared to those of the Milky Way (MW). @brocatoetal1996, @mucciarellietal2010 and @wagnerkaiseretal2017, among othes, showed that the old LMC GCs resemble the MW ones in age and in many chemical abundance patterns. In contrast, @johnsonetal2006 found that many of the abundances in the LMC old GCs are distinct from those observed in the MW, while @pg13 suggested that the most likely explanation for the difference between the old GC and field star age-metallcity relationships is a very rapid early chemical enrichment traced by the very visible old GCs. Indeed, the integrated spectroscopic metallicities obtained by @detal99 suggest that NGC1928 is one of the most metal-rich (\[Fe/H\] $\sim$ -1.2 dex) old GCs, whereas NGC1939 one of the most metal-poor (\[Fe/H\] $\sim$ -2.0 dex) old GCs. In Section 2 we describe the spectroscopic observations performed with the aim of deriving for the first time accurate mean cluster RVs (Section 3) and metallicities (Section 4). These quantities are considered in Section 5 to investigate whether NGC1928 and NGC1939 have been born in the LMC disc, or have other origins. Finally, a summary of the results is presented in Section 6. Observational data sets ======================= We carried out spectroscopic observations centred on the CaII infrared triplet ($\sim$ 8500 Å) of red giant stars located in the fields of NGC1928 and 1939. Most of the targets were selected from the $HST$ photometric data set of @mg04, bearing in mind their loci in the cluster colour-magnitude diagrams (CMDs). Because of the relatively small cluster angular sizes ($\la$ 1 arcmin) and their high crowding, many cluster red giants were discarded. For this reason, we considered some few other relatively bright red giant stars (4 in NGC1928 and 1 in NGC1939) without $HST$ photometry. Fig. \[fig:fig1\] illustrates the positions of the selected targets in the cluster fields and CMDs, respectively. In the case of NGC1939, we have also available $CT_1$ Washington photometry [@p17e], from which we built the cluster CMD of Fig. \[fig:fig2\]. ![image](fig1_ngc1928){width="\columnwidth"} ![image](ngc1928_cmd){width="\columnwidth"} ![image](fig1_ngc1939){width="\columnwidth"} ![image](ngc1939_cmd){width="\columnwidth"} ![NGC1939’s $CT_1$ CMD for stars located in the same area as in Fig. \[fig:fig1\], taken from @p17e. Selected stars are highlighted with filled red circles.[]{data-label="fig:fig2"}](ngc1939_ct1){width="\columnwidth"} Gemini South Observatory: GMOS spectra -------------------------------------- We carried out spectroscopic observations of stars in the field of NGC1928 and NGC1939 using the Gemini Multi-Object Spectrograph (GMOS) of Gemini South observatory during the nights of October 21 and 25, 2017, through programmes GS-2017B-Q-23 and GS-2017B-Q-71 (PI: Piatti), respectively. For each star cluster, we took four consecutive exposures of 900 sec for a single mask, as well as CuAr arcs and flats before and after the individual science exposures in order to secure a stable wavelength calibration. The total integration time for the science targets was 3600 sec. We used the R831 grating and the OG515 (&gt; 520 nm) filter, combined with a mask of 1.0 arcsec wide slits placed on the target stars, which gave a spectral sampling of $\sim$ 0.75Åper pixel with the 2$\times$2 CCD binning configuration. We observed 11 and 9 science target stars in the field of NGC1928 and NGC1939, respectively. We reduced the spectra following the standard GMOS data reduction procedure using the IRAF.[gemini.gmos]{} package. The wavelength calibration was derived using the [gswavelength]{} task, which compares the observed spectra with GCAL arc lamp data, and a wavelength solution was derived with a rms less than 0.20Å. We also used sky OH emission lines to further constrain the wavelength calibration and applied small offsets of about 0.3$-$0.5Å to the science spectra. The final dispersion of our data turned out to be 26.47 km/sec per pixel and the S/N ratio of the resulting spectra ranges from 30 up to 100, measured using the local continuum of the CaII triplet. Fig. \[fig:fig3\] illustrates spectra of some science targets. ![Normalised spectra of some observed stars.[]{data-label="fig:fig3"}](fig3){width="\columnwidth"} Anglo-Australian Telescope: AAOmega+2dF spectra ----------------------------------------------- We observed the region around ($\alpha$, $\delta$) = (5:24, $-$68:48) with the AAOmega spectrograph and 2dF fibre positioner at the 3.9m AAT on 2017 December 10–11, as part of a followup program intended to identify the most metal-poor red giants in the LMC (Emptage et al., in preparation). The fibre positions were chosen to optimise overlap with the targets in @coleetal2005 and @vanderswaelmenetal2013 in order to provide metallicity cross-calibration. NGC1939 is not far from the field centre, so 6 fibres were assigned to red giants within 3$^{\prime}$ of the cluster, over two configurations of the fibre plate. No fibers were assigned to stars in the vicinity of NGC 1928, a sparser cluster farther from the 2dF field centre. On 10 December, the field was observed for 3$\times$1800 s in 1$\farcs$6 seeing, and the following night a second fibre configuration was observed for 3$\times$1200 s in 1$\farcs$4 seeing. The red arm of the spectrograph was employed with the 1700D grating, centred on $\lambda_c$ = 8600 Å, for a dispersion of $\approx$0.24 Å per pixel, and a resolution R $\approx$11,000, depending on the position of the fibre image on the CCD. Arc and fibre flat exposures were taken immediately prior to each set of three science exposures. The data were reduced using the standard [*2dfdr*]{} data reduction package, which tunes the extraction parameters to optimise the signal to noise, producing wavelength-calibrated, sky-subtracted spectra. We obtained typical SNR values in the continuum of $\approx$15–50 depending on target I magnitude and fibre centring accuracy. Continuum normalisation was performed using the IRAF task [*continuum*]{}, with a sixth-order cubic spline fit and rejection of unusually low points, which are assumed to be photospheric lines. The spectra were not flux-calibrated, as we intend only to measure equivalent widths and radial velocities. Radial velocity measurements ============================ GMOS spectra ------------ We measured RVs by cross-correlating the observed spectra and synthetic ones taken from the PHOENIX library [@husser2013]. The synthetic spectra library covers the wavelength range 500 $-$ 55000Åand provides a wide coverage in effective temperatures ($T_{\textrm{eff}}$ $\sim 2300 - 12000\,^{\rm o}$K), surface gravities (log($g$) $\sim 0.0 - 6.0$dex) and metallicities (\[Fe/H\] $\sim -4.0 - +1.0$dex). We selected templates with $T_{\textrm{eff}}$ in the range $4000-5500\,^{\rm o}$K and log($g$) between $1.5-3.0$dex, which correspond to giant stars with MK types $\sim$ G0$-$K4. In the case of NGC1928, we restricted the templates to those with \[Fe/H\]$=-1.0$dex, while for NGC1939 we employed those with \[Fe/H\]$=-2.0$dex (see Section 4). In both cases, we selected 224 templates and checked that the restriction in metallicity already has a negligible impact on the RV estimates, since variations of 1.0 dex in \[Fe/H\] resulted in a change of $\sim$ 1 km/s in the derived RV (see also Fig. \[fig:fig4\]). The observed spectra were continuum normalised before the cross-correlation procedure and the synthetic templates had their spectral resolution degraded to match the resolution of our science spectra. We employed the transformation equations of @ciddor1996 to convert the wavelength grids from vacuum ($\lambda_{\textrm{vac}}$) to air wavelengths ($\lambda_{\textrm{air}}$; see also Section 3.2 of @angelo2017 for more details). Spectral fluxes ($F_{\lambda}^{\textrm{vac}}=\frac{dE_{\lambda}}{dt\,d\lambda_{\textrm{vac}}\,d\textrm{Area}}$) were also converted from vacuum to air values through the expression: $$F_{\lambda}^{\textrm{air}}=\frac{dE_{\lambda}}{dt\,d\lambda_{\textrm{air}}\,d\textrm{Area}}=F_{\lambda}^{\textrm{vac}}\left( \frac{d\lambda_{\textrm{vac}}}{d\lambda_{\textrm{air}}} \right)$$ Each observed spectrum was cross-correlated against the whole selected synthetic template sample by making use of the IRAF.[fxcor]{} task, which implements the algorithm described in @tonry1979 for the construction of the cross-correlation function (CCF) of each object - template pair of spectra. Besides the RV estimates, [fxcor]{} returns the CCF normalised peak ($h$) $-$ an indicator of the degree of similarity between the correlated spectra $-$ and the Tonry & Davis ratio (TDR) defined as TDR$=h/(\sqrt{2}\sigma_{a})$, where $\sigma_{a}$ is root mean square of the CCF antisymmetric component. For each object spectrum we assigned the RV value resulting from the cross-correlation with the highest $h$ value, which was in all cases greater than 0.8. We finally carried out the respective heliocentric corrections. Table \[tab:table1\] lists the resulting RVs with their respective uncertainties, while Fig. \[fig:fig4\] illustrates the cross-correlation procedure. ![Continuum normalised spectrum of star \#7 of NGC1928 (black lines) and the best fitted synthetic template spectra for \[Fe/H\]$=-1.0$dex (red line) and $-2.0$dex (green line), respectively.[]{data-label="fig:fig4"}](fig4.pdf){width="\columnwidth"} AAOmega+2dF spectra ------------------- We measured radial velocities by Fourier cross-correlation between the target star spectra and a set of templates obtained at the same resolution and signal-to-noise ratio. The templates were LMC field red giants observed by @coleetal2005, with metallicities from $-1 \lesssim$ \[M/H\] $\lesssim$ $-2$, and radial velocities between 200–300 km/s. Relative velocities for each star compared to each template in turn were calculated using [*fxcor*]{} in IRAF, and converted to a heliocentric frame. The radial velocities based on each template were averaged together, weighted by the cross-correlation peak height. The uncertainty in the resulting average velocities is on the order of $\pm$3 km/s, and is dominated by scatter between the various templates. Therefore the line of sight velocity dispersion of the field stars is highly resolved by these measurements, but cluster is expected to be unresolved. ------------ ------------ ----------- ------------ ------- ------------------ ----------------- ----------------- ----------------- ---------------- ID Instrument R.A. Dec. S/N RV W8498 W8542 W8662 \[Fe/H\] (deg) (deg) (km/s) (Å) (Å) (Å) (dex) NGC1928-1 GMOS 80.217297 -69.475896 71.3 250.29$\pm$3.03 1.024$\pm$0.010 2.516$\pm$0.086 1.724$\pm$0.023 -1.35$\pm$0.13 NGC1928-2 GMOS 80.230389 -69.481931 102.5 245.02$\pm$2.77 1.259$\pm$0.090 2.606$\pm$0.035 1.960$\pm$0.063 — NGC1928-3 GMOS 80.234166 -69.476247 33.1 238.61$\pm$6.92 0.983$\pm$0.040 1.963$\pm$0.070 1.847$\pm$0.093 -1.32$\pm$0.17 NGC1928-4 GMOS 80.240765 -69.472867 66.7 268.80$\pm$2.52 1.832$\pm$0.112 3.567$\pm$0.106 2.207$\pm$0.107 — NGC1928-5 GMOS 80.241292 -69.480756 64.7 286.96$\pm$3.37 1.863$\pm$0.105 3.510$\pm$0.160 2.247$\pm$0.048 — NGC1928-6 GMOS 80.245312 -69.471593 49.2 261.24$\pm$5.92 0.816$\pm$0.036 2.284$\pm$0.052 1.55$\pm$0.056 -1.33$\pm$0.15 NGC1928-7 GMOS 80.251698 -69.479367 86.3 251.72$\pm$3.08 1.089$\pm$0.033 2.678$\pm$0.071 2.039$\pm$0.048 — NGC1928-8 GMOS 80.255200 -69.472097 46.3 216.78$\pm$6.25 1.021$\pm$0.061 2.875$\pm$0.139 2.329$\pm$0.140 -0.71$\pm$0.24 NGC1928-9 GMOS 80.263638 -69.472226 53.6 214.11$\pm$3.81 0.970$\pm$0.032 2.780$\pm$0.120 1.875$\pm$0.094 -0.93$\pm$0.20 NGC1928-10 GMOS 80.216320 -69.480298 47.7 280.37$\pm$5.58 1.286$\pm$0.080 3.065$\pm$0.114 2.316$\pm$0.084 -0.61$\pm$0.23 NGC1928-11 GMOS 80.227749 -69.486631 29.0 231.37$\pm$8.27 0.950$\pm$0.100 1.958$\pm$0.157 1.631$\pm$0.040 -1.33$\pm$0.21 NGC1939-1 GMOS 80.339326 -69.944931 56.6 279.94$\pm$6.76 0.630$\pm$0.095 1.351$\pm$0.026 1.274$\pm$0.035 -1.95$\pm$0.13 NGC1939-2 GMOS 80.350991 -69.942040 39.6 258.18$\pm$4.92 1.599$\pm$0.133 3.203$\pm$0.220 2.486$\pm$0.123 -0.41$\pm$0.30 NGC1939-3 GMOS 80.352639 -69.954109 83.1 261.36$\pm$3.69 0.688$\pm$0.046 1.584$\pm$0.030 1.351$\pm$0.042 -2.05$\pm$0.10 NGC1939-4 GMOS 80.370202 -69.954369 22.1 260.40$\pm$2.93 0.932$\pm$0.013 2.139$\pm$0.012 1.577$\pm$0.068 -2.00$\pm$0.09 NGC1939-5 GMOS 80.371903 -69.956597 40.5 241.33$\pm$13.40 0.711$\pm$0.060 1.576$\pm$0.044 0.899$\pm$0.030 -2.02$\pm$0.12 NGC1939-6 GMOS 80.375680 -69.958244 49.9 250.52$\pm$3.92 1.589$\pm$0.050 3.528$\pm$0.095 2.634$\pm$0.051 -0.41$\pm$0.19 NGC1939-7 AAO+2dF 80.342251 -69.952236 25.3 259.90$\pm$2.50 0.900$\pm$0.100 1.176$\pm$0.118 1.384$\pm$0.138 -2.14$\pm$0.15 NGC1939-8 AAO+2dF 80.351829 -69.943060 26.6 282.50$\pm$3.40 0.800$\pm$0.100 2.348$\pm$0.235 2.502$\pm$0.250 -1.57$\pm$0.24 NGC1939-9 AAO+2dF 80.358841 -69.955288 24.9 259.90$\pm$2.80 0.809$\pm$0.081 2.048$\pm$0.205 1.131$\pm$0.113 -1.94$\pm$0.21 NGC1939-10 AAO+2dF 80.359225 -69.943083 19.2 259.60$\pm$2.80 0.569$\pm$0.057 2.067$\pm$0.207 1.224$\pm$0.122 -1.98$\pm$0.20 NGC1939-11 AAO+2dF 80.362350 -69.957974 13.8 257.90$\pm$3.50 1.400$\pm$0.100 4.016$\pm$0.402 2.416$\pm$0.242 -0.60$\pm$0.35 NGC1939-12 AAO+2dF 80.385450 -69.951819 25.0 270.20$\pm$2.80 1.495$\pm$0.150 3.880$\pm$0.388 2.950$\pm$0.295 -0.44$\pm$0.42 NGC1939-13 GMOS 80.339554 -69.953529 76.8 268.59$\pm$3.61 0.681$\pm$0.040 1.747$\pm$0.046 1.445$\pm$0.030 -1.95$\pm$0.11 NGC1939-14 GMOS 80.343735 -69.951279 40.3 253.30$\pm$11.01 0.502$\pm$0.078 1.573$\pm$0.055 1.093$\pm$0.030 -1.85$\pm$0.14 NGC1939-15 GMOS 80.384301 -69.945343 64.0 264.30$\pm$3.19 0.450$\pm$0.111 1.743$\pm$0.041 1.401$\pm$0.033 -1.93$\pm$0.14 ------------ ------------ ----------- ------------ ------- ------------------ ----------------- ----------------- ----------------- ---------------- Overall metallicity estimates ============================= GMOS spectra ------------ Equivalent widths of the CaII infrared triplet lines were measured from the normalised spectra using the [splot]{} package within IRAF. Their resulting average values and the respective uncertainties are listed in Table \[tab:table1\]. The latter were estimated by computing equivalent widths using different continua, bearing in mind the presence of TiO bands and the spectra S/N ratio. We then overplotted the sum of the equivalent widths of the three CaII lines ($\Sigma$W(CaII)) in the $\Sigma$W(CaII) versus $V-V_{\rm HB}$ plane, that has been calibrated in terms of metallicity [see, e.g., @coleetal2004]. In that diagram $V_{\rm HB}$ refers to the mean magnitude of the cluster horizontal branch. For NGC1928 and NGC1939 we adopted the individual $V$ magnitudes of the selected stars and $V_{\rm HB}$ $=19.3$mag, taken from @mg04 (see also Fig. \[fig:fig1\]). We also took advantage of the Washington photometry of @p17e (see also Fig. \[fig:fig2\]) to convert $T_1$ magnitudes of the selected stars into $V$ magnitudes $-$ for those stars without $HST$ $V$ mags $-$ using the theoretical red giant branches computed by @betal12, and the cluster reddening and distance moduli derived by @mg04. Fig. \[fig:fig5\] shows the resulting plots, where we included iso-abundance lines according to eq. (5) of @coleetal2004 for $\beta = 0.64$Å/mag [@rutledgeetal97], while the last column of Table \[tab:table1\] lists the interpolated \[Fe/H\] values. The errors were calculated by propagating those of the coefficients in eq. (5) [@coleetal2004], $\sigma$($\beta$) [@rutledgeetal97], the $HST$ [@mg04] and Washington [@p17e] photometric errors and $\sigma(\Sigma$W(CaII)), respectively. AAOmega+2dF spectra ------------------- Equivalent widths of the Ca II triplet lines were measured using the program [*EW*]{}, originally written by G.S. da Costa and used by @coleetal2005 and many others [e.g. @dacosta2016]. The lines were fit by a sum of Gaussian plus Lorentzian profiles, constrained to have a common centroid. The metallicities were measured as for the GMOS stars, described above. Because of the lower SNR, we tested the results against the method of @starkenburgetal2010, using only the two strongest lines of the Ca triplet; no significant differences were found within the errorbars. The total error on metallicity is dominated by systematic effects (e.g., possible differences in detailed abundance ratios between the target stars and those used to form the calibration sample) rather than random error from photon noise. For the field stars in the vicinity of NGC1939 in common with @coleetal2005, comparing the equivalent widths measured in the 2017 AAOmega spectra shows an average difference of $\sum$W$_{\mathrm{AAO-VLT}}$ = 0.06 $\pm$0.38 Å, highly consistent with no systematic offset. ![image](fig5a){width="\columnwidth"} ![image](fig5b){width="\columnwidth"} Analysis and discussion ======================= We first assigned to the observed stars cluster membership probabilities according to three different criteria, namely: the position of the stars in the cluster CMDs, the dispersion of their RV values and that for their \[Fe/H\] values, respectively. For NGC1928, we previously discarded stars \#8 and 9, which fall outside the cluster radius recently estimated by @pm2018 [31.6$\pm$7.3 arcsec] from a radial profile that reaches out to $\sim$ 4 times the cluster’s tidal radius. By looking at the cluster CMDs (Figs. \[fig:fig1\] and \[fig:fig2\]) we considered possible members any star located along the cluster red giant branches, within the observed spread of those sequences. We included the results of our assessment in column 2 of Table \[tab:table2\]. Note that this criterion could lead us to conclude on the cluster membership of any star that belongs to the LMC star field, because of the superposition with LMC field features. This is the case, for instance, of the LMC field red clump. We then built RV distribution functions by summing all the individual RV values, each of them represented by a Gaussian with centre and $\sigma$ equal to the mean RV value and the associated error, respectively (see Table \[tab:table1\]). Every Gaussian was assigned the same amplitude. The resulting RV distributions are shown in Fig. \[fig:fig6\], where the cluster RV ranges can be clearly identified from the FWHM of the primary peak (shadowed regions). In NGC1928’s panel, we intentionally included stars \#8 and 9 (red curve), thus confirming that they are probably non-members. In NGC1939, we also plotted the RV distributions obtained from using only stars observed with GMOS (green curve) and with AAOmega+2dF (magenta curve), respectively. As can be seen, there is a negligible shift between both RV scales, so that we summed them to produce the overall RV distribution (black curve). Table \[tab:table2\] lists the RV membership status assigned to each star on the basis of whether its RV falls within the shadowed regions. As for the metallicity membership probability, we visually inspected Fig. \[fig:fig5\], in which star sequences along a constant \[Fe/H\] value can be recognised, with some dispersion. For instance, the $\Sigma$W(CaII) versus $V-V_{\rm HB}$ diagram for NGC1928 (left panel) shows stars \#8 and 9 $-$ initially discarded because they fall outside the cluster radius $-$ and \#10 at a very distinguishable higher metallicity level. These stars have also RVs quite different from those of the observed cluster members. For the remaining stars, we do not have any argument as to deny them cluster membership. In the case of NGC1939 (right panel), the observed more metal-poor sequence contains more than three times the number of stars in the more metal-rich sequence (\[Fe/H\] $\sim -0.4$dex), so that we concluded that the former corresponds to that of the cluster. Note that the separation between both sequences is similar for $\Sigma$W(CaII) obtained from GMOS and AAOmega+2dF spectra, respectively. According to @coleetal2005 [see their figure 6], the derived \[Fe/H\] values for the observed stars meant to be LMC field stars (red symbols) are in excellent agreement with the bulk of metallicity values of LMC bar field giants. The final membership status of each star is listed in the last column of Table \[tab:table2\]. Only stars \#1 and 6 observed in the field of NGC1939 have RV memberships different from those adopted using separately their positions in the cluster CMDs and their metallicities, respectively. Nevertheless, we rely on the possibility that LMC field stars can have either RVs or metal-contents similar to that of the cluster. This is not the case of the field giant \#2 observed also along the line-of-sight of NGC1939, whose $V$ magnitude and $V-I$ colour place it superimposed on the cluster red giant branch (see Fig. \[fig:fig1\]). For the remaining stars observed in both cluster fields, the three membership criteria totally agree. We finally used the RV and \[Fe/H\] values of all cluster members to derive the mean cluster RVs and metallicities by employing a maximum likelihood approach. The relevance lies in accounting for individual star measurements, which could artificially inflate the dispersion if ignored. We optimized the probability $\mathcal{L}$ that a given ensemble of stars with velocities RV$_i$ and errors $\sigma_i$ are drawn from a population with mean RV $<$$RV$$>$ and dispersion W [e.g., @pm1993; @walker2006], as follows: $$\small \mathcal{L}\,=\,\prod_{i=1}^N\,\left( \, 2\pi\,(\sigma_i^2 + W^2 \, ) \right)^{-\frac{1}{2}}\,\exp \left(-\frac{(RV_i \,- <RV>)^2}{\sigma_i^2 + W^2} \right) .$$ where the errors on the mean and dispersion were computed from the respective covariance matrices[^2]. We obtained for NGC1928, $<$RV$_{\rm NGC\,1928}> = 249.58\pm$4.65 km/s and $<$\[Fe/H\]$_{\rm NGC\,1928}> = -1.30\pm$0.15 dex, while for NGC1939 the mean values turned out to be $<$RV$_{\rm NGC\,1939}> = 258.85\pm$2.08 km/s and $<$\[Fe/H\]$_{\rm NGC\,1939}> = -2.00\pm$0.15 dex. We compared our mean cluster RVs with those previously obtained by D99, who mentioned that their integrated spectra were not particularly suitable for accurate velocity measurements. Fig. \[fig:fig7\] shows the results, where other LMC GCs with RV estimates available in the literature were added. One of the diagnostic diagrams most frequently used to assess whether a cluster belongs to the LMC disc is that which shows the relationship between position angles (PAs) and RVs [@s92; @getal06; @shetal10; @vdmareletal2002; @vdmk14] for a disc-like rotation geometry. We here followed the recipe used by @s92, who converted the observed heliocentric cluster RVs to Galactocentric RVs through eq.(4) in @fw79. We computed cluster PAs by adopting the LMC disc central coordinates and their uncertainties obtained by @vdmk14 from $HST$ average proper motion measurements for stars in 22 fields. Fig. \[fig:fig8\] shows the disc solution derived for those $HST$ proper motions [Table 1 in @vdmk14] represented with a solid line, as well as those considering the uncertainties in the LMC disc line-of-sight systemic velocity, circular velocity and PA of the line-of-nodes and the derived velocity dispersion (dotted lines). As can be seen, NGC1928 and 1939 are placed within the fringes of the LMC disc at 1$\sigma$ confidence, similarly to many of the remaining 13 GCs included in the figure for comparison purposes. Therefore, assuming that both GCs belong to the LMC disc, we then sought for the best disc solutions for their respective RVs and position in the galaxy, i.e., we looked for the circular velocity ($v_{\rm rot}$) and PA of the line-of-nodes (PA$_{\rm LOS}$) of the discs that fully contain them. To do that, we used a grid of $v_{\rm rot}$ and PA$_{\rm LOS}$ values to evaluate eq.(1) of @s92 for the cluster PAs and their uncertainties, and then to find the most likely pair ($v_{\rm rot}$,PA$_{\rm LOS}$) that minimizes by $\chi^2$ the difference between the cluster RVs with their errors and those calculated above. We used a grid of $v_{\rm rot}$ from 0.0 up to 200.0 km/s in steps of 1.0 km/s, and a range of PA$_{\rm LOS}$ from 0.0 up to 360.0 degrees in steps of 1.0 degree. For NGC1928, the most suitable disc turned out to be that with $v_{\rm rot} = 45.0\pm$10.0 km/s and PA$_{\rm LOS} = 85.0\pm$10.0 degrees, while the resulting one for NGC1939 is that with $v_{\rm rot} = 35.0\pm$10.0 km/s and PA$_{\rm LOS} = 130.0\pm$10.0 degrees. For comparison purposes, we also computed $v_{\rm rot}$ and PA$_{\rm LOS}$ values for the remaining LMC GCs (see Table \[tab:table3\]). Fig. \[fig:fig9\] depicts the resulting $v_{\rm rot}$ values as a function of the deprojected distances ($r$, see Table \[tab:table3\]). The latter were computed using the LMC disc fitted by @vdmk14 from $HST$ proper motions in 22 fields, whose rotation curve is represented in the figure by a solid black line. The rotation curves obtained from line-of-sight (LOS) velocities of young and old stars [@vdmk14] are drawn with red and blue solid lines, respectively, and that from $Gaia$ DR2 proper motions [@vasiliev2018] with a magenta line. The figure reveals that NGC1928 and 1939 very well match the $HST$ proper motion rotation curve, as also do many other GCs. Reticulum ($r=$ 10.2 kpc, $v_{\rm rot}=$35 km/s) seems to rotate slower than the old stellar population LOS rotation curve, while NGC1835, 1898 and 2210 ($v_{\rm rot}>$ 100 km/s) are high circular velocity objects. Because the disc-like rotation geometry is shared by most the GCs (age $\ga$ 12 Gyr), we infer that the LMC disc had to exist since the early epoch of the galaxy formation, not only as an structure in itself but also from a dynamical point of view with a non-negligible angular momentum. The GCs that follow such rotation pattern span the entire metallicity range of all the GCs in the galaxy (-2.0 $\la$ \[Fe/H\] (dex) $\la$ -1.3, see also Table \[tab:table3\]), so that the LMC disc had also to experience a similar chemical enrichment within $\sim$ 3 Gyr of its GC formation [12 $\la$ age (Gyr) $\la$ 14, @piattietal2009; @wagnerkaiseretal2018]. Furthermore, because of the lack of a clear metallicity gradient among the disc GCs, we conclude that the whole disc - except possibly its very outskirts ($r >$ 15 kpc) - has been chemically evolved similarly. The four GCs mentioned above that significantly depart from the LMC rotation curve have ages and metallicities in the same ranges as those disc GCs. However, it is hard to figure out an [*in situ*]{} GC scenario for them, because of their very different $v_{\rm rot}$ values. Note that the velocity dispersion for young and old stellar population derived by @vdmk14 is 11.6 and 22.8 km/s, respectively [see also @s92; @vdmareletal2002], so that their velocities differ by more than three times the LOS velocity dispersion of the LMC old population. One alternative is to conclude that these four objects were stripped from the SMC, whose oldest stellar population has ages and metallicities compatible with them [@pg13]. Indeed, such a possibility has been suggested by @carpinteroetal2013, who modelled the dynamical interaction between both galaxies. Consequently, our results become in the first observational evidence that the LMC have accreted not only populations of SMC field stars [@olsenetal2011] but also some of its present GCs. ![image](fig6a){width="\columnwidth"} ![image](fig6b){width="\columnwidth"} ![Comparison of LMC GC RVs derived by D99 with those available in the literature (see Table \[tab:table3\]).[]{data-label="fig:fig7"}](fig7){width="\columnwidth"} ![Galactocentric RVs versus PAs diagram for LMC GCs. RVs were taken from the literature (see Table \[tab:table3\]). We included the curves derived by @vdmk14 from $HST$ proper motions of 22 LMC fields (see text for details).[]{data-label="fig:fig8"}](fig8){width="\columnwidth"} ![LMC rotation curve as a function of the deprojected distance ($r$) derived from $HST$ proper motions of 22 fields, and from LOS velocities of young and old stellar populations drawn with black, red and blue solid lines, respectively [taken from figure 7 of @vdmk14]. The rotation curve derived by @vasiliev2018 is drawn with a magenta line. All the 15 GCs have been superimposed with filled circles; NGC1928 and 1939 in red.[]{data-label="fig:fig9"}](fig9){width="\columnwidth"} ------------ ------------------ ----- ---- ---------- --------- ID Distance to CMD RV \[Fe/H\] Adopted cluster’s centre NGC1928-1 m m m m m NGC1928-2 m – m – m NGC1928-3 m m m m m NGC1928-4 m – m – m NGC1928-5 m – nm – nm NGC1928-6 m m m m m NGC1928-7 m – m – m NGC1928-8 nm – nm nm nm NGC1928-9 nm – nm nm nm NGC1928-10 m m? nm nm nm NGC1928-11 m m m m m NGC1939-1 m m nm m nm NGC1939-2 m m nm nm nm NGC1939-3 m m m m m NGC1939-4 m m m m m NGC1939-5 m m m m m NGC1939-6 m nm m nm nm NGC1939-7 m m m m m NGC1939-8 m nm nm nm nm NGC1939-9 m m m m m NGC1939-10 m m m m m NGC1939-11 m nm nm nm nm NGC1939-12 m nm nm nm nm NGC1939-13 m m m m m NGC1939-14 m m m m m NGC1939-15 m m m m m ------------ ------------------ ----- ---- ---------- --------- : Cluster membership of the observed stars.[]{data-label="tab:table2"} ----------- ---------------- -------------- ---------------- ------ ---------------- ------- ---------------- ---------------- ID PA $r$ RV Ref. \[Fe/H\] Ref. PA$_{\rm LOS}$ $v_{rot.}$ (deg) (kpc) (km/s) (dex) (deg) (km/s) NGC1466 250.8$\pm$2.0 8.9$\pm$0.9 200.0$\pm$5.0 1 -1.90$\pm$0.10 10 190.0$\pm$10.0 90.0$\pm$10.0 NGC1754 234.2$\pm$6.7 2.4$\pm$0.9 234.1$\pm$5.4 3 -1.50$\pm$0.10 5,6 100.0$\pm$10.0 75.0$\pm$10.0 NGC1786 313.7$\pm$7.7 1.8$\pm$0.9 279.9$\pm$4.9 3 -1.75$\pm$0.10 5,6,7 350.0$\pm$10.0 75.0$\pm$10.0 NGC1835 256.2$\pm$16.8 1.0$\pm$0.9 188.0$\pm$5.0 3 -1.72$\pm$0.10 5,6 130.0$\pm$10.0 170.0$\pm$10.0 NGC1841 183.0$\pm$2.0 14.2$\pm$0.8 210.3$\pm$0.9 2 -2.02$\pm$0.10 5 130.0$\pm$10.0 85.0$\pm$10.0 NGC1898 163.0$\pm$18.5 0.4$\pm$0.9 210.0$\pm$5.0 1 -1.32$\pm$0.10 5,6,8 110.0$\pm$10.0 130.0$\pm$10.0 NGC1916 124.4$\pm$26.0 0.3$\pm$0.9 278.0$\pm$5.0 1 -1.54$\pm$0.10 9 160.0$\pm$10.0 65.0$\pm$10.0 NGC2005 113.3$\pm$10.0 1.4$\pm$0.9 270.0$\pm$5.0 1 -1.74$\pm$0.10 5,6,8 280.0$\pm$10.0 35.0$\pm$10.0 NGC2019 120.1$\pm$8.5 1.6$\pm$0.9 280.6$\pm$2.3 2 -1.56$\pm$0.10 5,6,8 150.0$\pm$10.0 65.0$\pm$10.0 NGC2210 89.2$\pm$3.0 5.2$\pm$0.9 343.0$\pm$5.0 1 -1.55$\pm$0.10 7,9 140.0$\pm$10.0 160.0$\pm$10.0 NGC2257 59.1$\pm$1.5 9.8$\pm$0.9 301.6$\pm$0.8 2 -1.77$\pm$0.10 5,7,9 100.0$\pm$10.0 70.0$\pm$10.0 Hodge11 97.3$\pm$2.8 5.2$\pm$0.9 245.1$\pm$1.0 2 -2.00$\pm$0.10 11 335.0$\pm$10.0 70.0$\pm$10.0 Reticulum 334.1$\pm$2.0 10.2$\pm$0.9 247.5$\pm$1.5 2 -1.57$\pm$0.10 2 170.0$\pm$10.0 35.0$\pm$10.0 NGC1928 119.0$\pm$20.0 0.6$\pm$0.9 249.6$\pm$12.8 4 -1.30$\pm$0.15 4 85.0$\pm$10.0 45.0$\pm$10.0 NGC1939 144.2$\pm$16.0 0.8$\pm$0.9 258.8$\pm$7.4 4 -2.00$\pm$0.15 4 130.0$\pm$10.0 35.0$\pm$10.0 ----------- ---------------- -------------- ---------------- ------ ---------------- ------- ---------------- ---------------- Ref.: (1)@s92; (2)@getal06; (3)@shetal10; (4) this work; (5)@setal92; (6)@beetal2002; (7)@mucciarellietal2010; (8)@johnsonetal2006; (9)@wagnerkaiseretal2018; (10)@walker1992; (11)@matelunaetal2012. Conclusions =========== With the aim of investigating the origin of the LMC GCs NGC1928 and 1939, we carried out spectroscopic observations of giant stars located in their fields with the GMOS and the AAOmega+2dF spectrographs of the Gemini South and the Australian Astronomical Observatories, respectivey. The targets were selected bearing in mind their positions along the red giant branch or red clump in $HST$ cluster CMDs, the only available photometric data set at the moment of preparing the observations. Some few candidates without $HST$ photometry were also selected. The resulting high S/N spectra centred on the CaII infrared triplet allowed us to measure accurate individual RVs for 11 and 15 stars in the fields of NGC1928 and 1939, respectively. The RVs were obtained through cross-correlation of the observed spectra with template spectra. We also measured equivalent widths of the three CaII lines and derived individual metallicities (\[Fe/H\]) for those stars with available photometry using a previous well-established calibration. The accuracy in the individual \[Fe/H\] values ranges 0.1-0.3 dex. By considering as membership probability criteria the position of the observed stars in the cluster CMDs, and their position in the RV and metallicity distribution functions, we concluded that 7 and 9 observed stars are probable cluster members of NGC1928 and 1939, respectively. The combined three criteria resulted to be a robust approach to assess the cluster membership of the observed stars. From the adopted cluster members we estimated for the first time accurate mean cluster RVs and metallicities. We found that NGC1928 is one of the most-metal rich GCs (\[Fe/H\]=-1.3 dex), and NGC1939 is one of the most metal-poor ones (\[Fe/H\]=-2.0 dex). Both GCs are located in the innermost region of the LMC (deprojected distance &lt; 1 kpc) and have RVs consistent with being part of the LMC disc. Therefore, we rule out any possible origin but that in the same galaxy. Indeed, we computed the best solution for a rotation disc that fully contains each GC, separately, and found that the resulting circular velocities at the deprojected cluster distances very well match the rotation curves fitted from $HST$ and $Gaia$ DR2 proper motions, respectively. We extended our kinematics analysis to all the 15 LMC GCs by obtaining also circular velocities. The outcomes show that most of the GCs share the LMC rotation curve. Since they span the whole LMC GC metallicity range with no evidence of a metallicity gradient, we concluded that the LMC disc has existed since the early epoch of the galaxy formation and has also experienced the abrupt chemical enrichment seen in its GC populations in an interval of time of $\sim$ 3 Gyr. Four objects out of the fifteen GCs (NGC1835, 1898, 2210 and Reticulum) have estimated circular velocities which notably depart from the LMC rotation curve. We think that they are witnesses of having been stripped by the LMC from the SMC, an scenario predicted from numerical simulations of the galaxy dynamical interactions and confirmed from observation of field star populations. Acknowledgements {#acknowledgements .unnumbered} ================ Based on observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the National Research Council (Canada), CONICYT (Chile), Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina), and Ministério da Ciência, Tecnologia e Inovação (Brazil). We thank Dougal Mackey for providing us with the $HST$ photometric data base. We thank the referee for the thorough reading of the manuscript and timely suggestions to improve it. \[lastpage\] [^1]: E-mail: [email protected] [^2]: @pm1993 noted that this approach underestimates the true velocity dispersion for small sample sizes.
{ "pile_set_name": "ArXiv" }
![Pulsar sky map in Galactic coordinates. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. Black dots: Pulsars for which gamma-ray pulsation searches were conducted using rotational ephemerides. Gray dots: Known pulsars which were not searched for pulsations. \[SkyMap\]](Figures/glast_900psr_atnf.ps){width="100.00000%"} ![ $P-\dot P$ diagram. Dashed lines: characteristic age $\tau_{\rm c}$. Dot-dashed lines: rotational energy loss rate $\dot E$. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. Black dots: Pulsars for which gamma-ray pulsation searches were conducted using rotational ephemerides. Gray dots: Known pulsars which were not searched for pulsations. \[PPdot\]](Figures/ppdot_distrib.eps){width="80.00000%"} ![Galactic plane pulsar distribution (polar view). The star represents the Galactic center. The two circles centered at the Earth’s position have radii of 3 kpc and 5 kpc. For pulsars with different possible distances, the nearer values from Table \[tab:dist\] are used. Note that the millisecond pulsars (MSPs), while having a significantly lower $\dot E$ than the other pulsars (see Figure \[IndexvsEdot\]), are detectable due to their close proximity. The one exception (PSR J0218+4232) also exhibits a significantly higher $\dot E$ than the other MSPs. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. Black dots: Pulsars for which gamma-ray pulsation searches were conducted using rotational ephemerides. Gray dots: Known pulsars which were not searched for pulsations. \[SkyProjection\]](Figures/CmyPolar.eps){width="100.00000%"} ![Phase difference $\Delta$ between the gamma-ray peaks, versus the phase lag $\delta$ between the main radio peak and the nearest gamma-ray peak. Pulsars without a radio detection are plotted with $\delta=0$. With present light curves we cannot generally measure $\Delta < 0.15$; objects classified as single-peaked are plotted with $\Delta$=0. Two such objects, both MSPs, are off the plot at $\delta>0.8$. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. The plot has been corrected for the updated value of $\delta$ for PSR J1124-5916, as per an Erratum sent to the ApJ (December 2010). \[RadioSep\]](Figures/radiolag_psep.eps){width="60.00000%"} ![Magnetic field strength at the light cylinder $B_{\rm LC}$ versus pulsar characteristic age $\tau_{\rm c}$. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. Black dots: Pulsars for which gamma-ray pulsation searches were conducted using rotational ephemerides. Gray dots: Known pulsars which were not searched for pulsations. \[BlcvsAge\]](Figures/AgeBlc.eps){width="60.00000%"} ![Gamma-ray luminosity $L_\gamma$ versus the rotational energy loss rate $\dot E$. Dashed line: $L_\gamma$ equal to $\dot E$. Dot-dashed line: $L_\gamma$ proportional to the square root of $\dot E$. $L_\gamma$ is calculated using a beam correction factor $f_\Omega=1$ for all pulsars and the integral energy flux $G_{100}$ from the on-pulse spectral analysis (see Section 2.2), except for PSRs J1836+5925 and J2021+4026 which use the total background-corrected phase-averaged flux, including a relatively bright unpulsed component (see Section 2.2). For the Crab we also plot the total high energy luminosity, $L_{\rm tot} = L_X + L_\gamma$, indicated by \*. Several notable pulsars have been labeled. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. Unfilled markers indicate pulsars for which only a DM-based distance estimate is available (see Table \[tab:dist\]). Pulsars with two distance estimates have two markers connected with dashed error bars. \[LEdot\]](Figures/LvsEdot.eps){width="100.00000%"} ![Value of the exponential cutoff $E_{\rm cutoff}$ versus the magnetic field at the light cylinder, $B_{\rm LC}$. The statistical uncertainties on $E_{\rm cutoff}$ are shown. An additional systematic bias of ($+20\%,\,-10\%$) may affect $E_{\rm cutoff}$ (see text). The histogram of $E_{\rm cutoff}$ values is projected along the right-hand axis. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. \[EcvsBlc\]](Figures/CutOffvsBlc.eps){width="50.00000%"} ![Photon index $\Gamma$ versus the rotational energy loss rate, $\dot E$. For $\Gamma$, the statistical uncertainties combined with the systematic uncertainties due to the diffuse emission model are shown. An additional systematic bias of ($+0.3,\,-0.1$) affects $\Gamma$ (see text). The histogram of the photon indices is projected along the right-hand axis. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. \[IndexvsEdot\]](Figures/IndexvsEdot_H.eps){width="50.00000%"} ![Aitoff projection sky map of the $5\sigma$ sensitivity in units of logarithmic [**photon**]{} flux (Log($L_\gamma$) ph cm$^{-2}$s$^{-1}$) for six months of *Fermi* LAT sky-survey data. The sensitivity analysis uses the model of the diffuse gamma-ray background described in the text (Section 4), and pulsar spectra with differential photon indices of $\Gamma = 1.4$ with an exponential cutoff energy of $E_{\rm cutoff} = 2.2$ GeV. \[SensitivitySkyMap\]](Figures/sensitivPSR_aitoff_6month_2p2GeVcutoff.eps){width="50.00000%"} ![Measured integral photon flux above 100 MeV, $F_{100}$, versus the $5\sigma$ flux sensitivity described in Figure \[SensitivitySkyMap\]. For $F_{100}$, the statistical uncertainties combined with the systematic uncertainties due to the diffuse emission model are shown. An additional systematic bias of ($+30\%,\,-10\%$) affects $F_{100}$ (see text). The effective blind search sensitivity is comparable to the $2\times 5\sigma$ line, although a few pulsars are discovered at lower flux, presumably due to favorable pulse profiles, spectra or local backgrounds. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. \[FluxVsSensitivity\]](Figures/Senslim.eps){width="50.00000%"} ![LogN–LogS distribution as described in Section 4 for all the detected pulsars (black dashed line), the radio-selected gamma-ray pulsars including MSPs (grey histogram), and the gamma-ray-selected pulsars (blue hatched histogram). \[LogN-LogS\]](Figures/LogN-LogS_2.eps){width="60.00000%"} ![Pulsar “detectability” metric (${\dot E}^{1/2}/d^2$, normalized to Vela) vs. spin period. Detected MSPs (red triangles) and young pulsars (radio-selected, green circles; gamma-ray-selected, blue squares) all have high values of this metric. For objects with a distance range in Table \[tab:dist\], we use here the geometric mean of the maximum and minimum values. Searched, but presently undetected objects (gray dots) are plotted using DM-derived distances. For the possible causes of non-detection see Section 5.1. \[DetectMetric\]](Figures/edotfigscaledv3.eps){width="70.00000%"} ![Separations $\Delta$ between the gamma-ray peaks, for those pulsars with two identified peaks, versus the spin-down power $\dot E$. The histogram of peak separations is projected along the right-hand axis. Blue squares: gamma-ray-selected pulsars. Red triangles: millisecond gamma-ray pulsars. Green circles: all other radio loud gamma-ray pulsars. \[DeltavsEdot\]](Figures/PeakSeparationEdot.eps){width="70.00000%"}
{ "pile_set_name": "ArXiv" }
--- abstract: | The goal of this paper is to study uniqueness of a one-dimensional Hamilton-Jacobi equation $$\begin{cases} u_t=|u_x|^2+R(x,I(t)) &\text{in }{\mathbb{R}}\times (0,\infty), \\ \max_{{\mathbb{R}}} u(\cdot,t)=0 &\text{on }[0,\infty), \end{cases}$$ with an initial condition $u_0(x,0)=u_0(x)$ on ${\mathbb{R}}$. A reaction term $R(x,I(t))$ is given while $I(t)$ is an unknown constraint (Lagrange multiplier) that forces maximum of $u$ to be always zero. In the paper, we prove uniqueness of a pair of unknowns (u,I) using dynamic programming principle in one dimensional space for some particular class of nonseparable reaction $R(x,I(t))$. address: ' Department of Mathematics, University of Wisconsin Madison, 480 Lincoln Drive, Madison, WI 53706, USA' author: - Yeoneung Kim title: 'On the Uniqueness for One-Dimensional Constrained Hamilton-Jacobi Equations' --- [^1] Introduction ============ The non-local parabolic equations arising in adaptive dynamics(see [@Darwin1; @Darwin2; @Darwin3; @Darwin4]) have an interesting feature so called Dirac concentration of density as a diffusion coefficient vanishes. To illustrate this, we consider the following evolution equation $$\begin{cases} n^{\varepsilon}_t - {\varepsilon}\Delta n^{\varepsilon}= \frac{n^{\varepsilon}}{{\varepsilon}} R(x,I^{\varepsilon}(t)) &\text{in } {\mathbb{R}}^n \times (0,\infty),\\ n^{{\varepsilon}}(x,0)=n^{\varepsilon}_0 \in L^1({\mathbb{R}}^n) &\text {on }{\mathbb{R}}^n,\\ I^{\varepsilon}(t)=\int_{{\mathbb{R}}^n} \psi(x) n^{\varepsilon}(t,x)dx, \end{cases}$$ where the spatial variable $x$ denotes ‘traits’ in the environment. Furthermore, $n^{\varepsilon}$, $R(x,I^{\varepsilon}(t))$, ${\varepsilon}$ and $\psi(x)$ describe density of the population, reproduction rate, mutation rate and consumption rate by a trait $x$. Here $\psi$ assumed to be a nonnegative compactly supported function. We then take Hopf-Cole transformation $n^{\varepsilon}(x,t)=e^{u^{\varepsilon}(x,t) /{\varepsilon}}$. It was shown in many literatures that as mutation rate ${\varepsilon}$ vanishes, $u^{\varepsilon}$ converges locally uniformly to $u$ which is a viscosity solution to $$\label{origin} \begin{cases} u_t=|Du|^2+R(x,I(t)) &\text{in } {\mathbb{R}}^n \times (0,\infty),\\ \max_{ {\mathbb{R}}^n} u(\cdot,t)=0 &\text{on } [0,\infty),\\ u(x,0)=u_0(x) &\text{on } {\mathbb{R}}^n. \end{cases}$$ The constraint of $u$ is obtained from the property that $I^{\varepsilon}$ is positive and uniformly bounded. It was also shown that $$n^{\varepsilon}(x,t) n(x,t) \rightharpoonup {{\overline}\rho}(x)(x(t)-{\overline}x(t))\text{ weakly in the sense of measure}$$ where $$u({\overline}x(t),t)=\max_{{\mathbb{R}}} u(\cdot,t)=0 \text{ and } \rho(t)=\frac{I(t)}{\psi(x)}$$ for the solution $n^{\varepsilon}(x,t)$ to (\[origin\]) (see [@Dirac.C; @Convergence]). Despite the existence of solutions to (\[origin\]) is quite well understood, the uniqueness is relatively less known. In the recent work by S. Mirrahimi, J. -M. Roquejoffre [@Uniqueness], the uniqueness of the solution is shown when the reaction and initial condition $u_0(x)$ are strictly concave so that regularity of maximum point is obtained. However, the uniqueness for general initial data and a nonconave reaction is still open. In this paper, the uniqueness property for constrained Hamilton-Jacobi equations in 1-D with some nonseparable reaction terms is obtained using dynamic programming principle. Setting and main result ----------------------- We need following assumptions on $$R(x,I):{\mathbb{R}}\times [0,\infty) \rightarrow {\mathbb{R}}\text{ and } u_0(x): {\mathbb{R}}\rightarrow {\mathbb{R}}.$$ where the reaction term is defined as $$R(x,I)= \begin{cases} b(x)-Q(I) & \text{ for } x \geq 0,\\ R'(x,I) & \text{ for } x <0. \end{cases}$$ - $R$ is smooth and $R'(\cdot, I)<0$ on $(-\infty,0)$ for any positive $I$ ; - $\sup_{0\leq I \leq I_M} \|R(\cdot,I)\|_{W^{2,\infty}} <\infty$ and $R$ is strictly decreasing in $I$; - $Q(I)\geq 0$ is strictly increasing in $I$ and $Q(0)=0$ - $\sup_{{\mathbb{R}}} R(\cdot,I_M)=0$; - $\min_{{\mathbb{R}}} R(\cdot,0)=0$; - $b(x)$ is strictly increasing on $[0,\infty)$ with b(0)=0; - $b'(x)$ is Lipschitz continuous, hence, nonnegative ; - $u_0(x) \in C^2({\mathbb{R}})$ with $\|u_0\|_{C^2({\mathbb{R}})}<\infty$, $\max_{x\in {\mathbb{R}}} u_0(\cdot)=u_0(0)=0$ and $u_0(x)<0$ elsewhere. Additionally, $f \in {W^{1,\infty}}({\mathbb{R}}^n)$, that is; $\|f\|_{L^\infty({\mathbb{R}}^n)}+\|Df\|_{L^\infty({\mathbb{R}}^n)} <\infty$. Now we are ready to state our main theorem. Under the assumptions above, we consider the following equation. $$\label{HJ} \begin{cases} u_t={u_x}^2+R(x,I(t)) &\text{in } {\mathbb{R}}\times (0,\infty),\\ \max_{ {\mathbb{R}}} u(\cdot,t)=0 &\text{on } [0,\infty),\\ u(x,0)=u_0(x) &\text{on } {\mathbb{R}}. \end{cases}$$ \[main\] There exists at most one pair $(u,I)$ such that $u(x,t) \in C({\mathbb{R}}\times (0,\infty))$ solves (\[HJ\]) in viscosity sense and $I(t) \in C([0,\infty))$ is strictly increasing. Preliminary =========== Throughout the section, let us assume $(u,I)\in C({\mathbb{R}}\times (0,\infty)) \times C([0,\infty))$ is a pair of solution to (\[HJ\]) in viscosity sense. By a Lipschitz estimate provided by the author in [@yeon], one can assume further that $u$ is Lipschitz continuous in ${\mathbb{R}}\times [0,T]$ for any positive $T$. Now we follow dynamic programming principle arguments presented in [@Uniqueness], which yields $$u(x,t)=\sup_{{\gamma}(t)=x} \{F({\gamma}):{\gamma}\in AC([0,t];{\mathbb{R}}) \}$$ where $$F({\gamma}):=u_0({\gamma}(0))+\int_0 ^t \left( -\frac{ \dot {\gamma}^2}{4} +R({\gamma}(s),I(s) \right) ds.$$ Furthermore, one can actually show that there exists a path ${\gamma}(s) \in C^1([0,t);{\mathbb{R}}))$ such that $$u(x,t)=u_0({\gamma}(0))+\int_0 ^t \left( -\frac{ \dot {\gamma}^2}{4} +R({\gamma}(s),I(s) \right) ds$$ with ${\gamma}(t)=0$ and it satisfies Euler-Lagrange equation $$\label{EL} \begin{cases} \ddot {\gamma}(s)+2R_x({\gamma}(s),I(s))=0,\\ \dot {\gamma}(0)+2\dot {u_0}({\gamma}(0))=0,\\ {\gamma}(t)=x. \end{cases}$$ For the details, see [@Uniqueness] and references therein. There could be more than one solution to the equation above. However, the Euler-Lagrange equation reduces to a simpler equation that results in the existence of a unique solution in our setting. We start with some generic properties. \[prop1\] Assume that $\max_{{\mathbb{R}}} u(\cdot, t)=u(x',t)=0$. Then $R(x',I(t))$=0. By viscosity subsolution test, one can easily obtain $R(x',I(t) \geq 0$. Now we assume that $R(x',I(t))>0$. Then there exists $t_0>0$ such that $R(x',I(s))>0$ on $[t,t+t_0]$ by the continuity of $I$ and $R$. Integrating (\[HJ\]) both sides over $\{x'\} \times [t,t+t_0]$ yields $$\begin{aligned} u(x',t+t_0)-u(x',t) & \geq \int_t ^{t+t_0} R(x',I(s))ds >0\end{aligned}$$ Hence, we get $$u(x',t+t_0) >0,$$ which violates the maximum constraint. We define $x(t) \in {\mathbb{R}}$ to satisfy $$R(x(t),I(t))=0$$ for $t>0$ and a strictly increasing $I(t)$. Then, together with Propositition \[prop1\], we have $$\label{p} max_{{\mathbb{R}}} u(\cdot,t)=u(x(t),t))=0.$$ for a solution pair $(u,I)$. \[bd\] $I(0)=0$ and $I(s) \leq I_M$ on $[0,\infty)$. Let us first prove $I(0)=0$ when $(u,I)$ is a pair of solution. We may assume $I(0)>0$. From the property (\[p\]), we deduce $$0=\lim_{t\rightarrow0+} u(x(t),t)=u(x(0+),0)<0$$ where $x(0+)$ is a right limit of $x(t)$, which yields contradiction. Therefore, $I(0)=0$. The second part of the proposition, $I(s) \leq I_M$, is a straight consequence of Proposition \[prop1\] due to the assumption on $R$. We also need some regularity properties of the solution $u(x,t)$, which play crucial roles in analyzing the trajectory ${\gamma}(s)$. For a real valued function $u(x)$ define for $x\in {\mathbb{R}}^n$, we define super differential and sub differential at $x$ as $$\begin{aligned} D^{+}u(x)&=\{p \in {\mathbb{R}}^n : \liminf_{y \rightarrow x} \frac{u(y)-u(x)- p \cdot (y-x)} {|y-x|} \geq 0\} \\ D^{-}u(x)&=\{p \in {\mathbb{R}}^n : \limsup_{y \rightarrow x} \frac{u(y)-u(x)-p\cdot (y-x)}{|y-x|} \leq 0\}\end{aligned}$$ \[reg\] A solution $u(x,t)$ is semiconvex in $x\in {\mathbb{R}}$ for any fixed positive $T$. Let us define $v(x,t)=-u(x,t)$ and prove $v(x,t)$ is semiconcave in ${\mathbb{R}}\times [0,T]$. Cleary, $v$ satisfies $$\begin{cases} v_t+v_x^2 +R(x,I(t))=0 &\text{ in } {\mathbb{R}}^n \times (0,T],\\ v(x,0)=-u(x,0) &\text{ on }{\mathbb{R}}\end{cases}$$ in viscosity sense. To prove semiconcavity of $v$, we first provide a priori estimate for $v^{\varepsilon}$ where $v^{\varepsilon}$ is a unique solution to $$\begin{aligned} \label{ap} \begin{cases} v_{t}^{\varepsilon}+ (v_{x}^{{\varepsilon}})^2 +R(x,I(t))={\varepsilon}v_{xx}^{\varepsilon}&\text{ in } {\mathbb{R}}\times (0,T],\\ v^{\varepsilon}(x,0)=-u_0(x):=v_0(x) &\text{ on } {\mathbb{R}}. \end{cases} \end{aligned}$$ Differentiating (\[ap\]) twice with respect to $x$ and substituting $w$ for $v_{xx}^{\varepsilon}$ yields $$w_t+2w^2+2v_x w_x+R_{xx}={\varepsilon}w_{xx}.$$ It is known that $w$ is bounded but the bound depends on ${\varepsilon}$. However, one can actually show that the bound is uniform in ${\varepsilon}$. To justify this, we first notice that $w$ is a subsolution to the following parabolic equation $$\begin{aligned} \label{para} \begin{cases} w_t+v_x w_x +R_{xx}={\varepsilon}w_{xx} &\text{ in } {\mathbb{R}}\times (0,T],\\ w(x,0)=v''_0(x) &\text{ on } {\mathbb{R}}. \end{cases}\end{aligned}$$ On the other hand, $v_0+Ct$ and $v_0-Ct$ are supersolution and subsolution to (\[para\]) respectively where $C$ depends only on the bound for $R_{xx}$. Therefore, by comparison principle, one can obtaiin $|w|<C$ where $C$ does not depend on ${\varepsilon}$. As a last step, we need the following estimate. *There exists positive $C$ that depends only on $T$ such that* $$\|v_{t}^{\varepsilon}\|_{L^\infty({\mathbb{R}}\times [0,T])} +\|v_{x}^{\varepsilon}\| _{L^\infty({\mathbb{R}}\times [0,T])} < C$$ Since $0\leq I(t) \leq I_M$, $R(x,I)$ is bounded, for $C>0$ large enough, we have $$v_0(x)-Ct \leq v^{\varepsilon}(x,t) \leq v_0(x)+Ct$$ by the comparison principle. Using the comparison principle one more time yields $$v^{\varepsilon}(s+t) \geq v^{\varepsilon}(t)-Cs$$ for $s,t \geq 0$. Therefore, $v_{t}^{\varepsilon}> -C$ in ${\mathbb{R}}\times [0,T]$. On the other hand, observing the original equation (\[ap\]), we can derive $\|v_{x} ^{\varepsilon}\|_{L^\infty({\mathbb{R}}\times [0,T])}< C$ as $v_{xx} ^{\varepsilon}$ is bounded above. Finally, an upper bound for $v_{t}^{\varepsilon}$ is obtained, and such bounds depend only on $T$. As a consequence, $v_{\varepsilon}$ converges locally uniformly to $v$ as ${\varepsilon}$ goes to $0$ by Arzela-Ascoli and by the uniqueness and stability of a viscosity solution. Moreover, the semiconcaivity of $v^{\varepsilon}$ in $x$ implies that $$v^{\varepsilon}(x,t)-K|x|^2$$ is concave in $x$ for some positive $K$. Combining it with locally uniform convergence of $v^{\varepsilon}$, we get semiconcavity of $v$ in $x$. Thereroe, $u$ is locally semiconvex in $x$. \[lemma\] For each $t \in (0,\infty)$, $u(x,t)$ is differentiable at $(x(t),t)$ with respect to the space variable $x$ and it satisfies $$\label{re} 0=u_x(x(t),t)=-\frac{\dot {\gamma}_x(t)}{2}.$$ In addition to that, by the maximum constraint, we have $\dot {\gamma}_x(t)=0$. By Lemma \[reg\], $v(x,t)=-u(x,t)$ is semiconvcave in $x$. Hence, supper differential at $(x(t),t)$ is nonempty. On the other hand $p=0$ is a subdifferential of for $v$ at $(x(t),t)$. Therefore, $u$ is differentiable with respect to the space variable at $(x(t),t)$. Moreover, the derivative is $0$. A classical result in [@Concave] suggests that $$\eta(t) \in {\nabla}^+ v(x(t),t)$$ where $\dot {\gamma}_x(s)= 2\eta(s)$ for $s\in[0,t]$ and $v$ is defined as above. Combining these two, we get the result using the differentiability of $v$ at $(x(t),t)$. \[imp\] Let ${\gamma}(s)\in C^1([0,t];{\mathbb{R}})$ an optimizing path whose terminal point is $x(t)$ and $x(s)\in {\mathbb{R}}$ satisfy $R(x(s),I(s))=0$ for $s>0$. Then we have ${\gamma}(s)>x(s)$ for $s\in(0,t)$. We may assume first that ${\gamma}(s) \geq 0$ since $F({\gamma}^+) \geq F({\gamma})$ where $$\begin{aligned} {\gamma}^+(s)= \begin{cases} {\gamma}(s) &\text{if }{\gamma}(s)>0\\ 0 &\text{if } {\gamma}(s) \leq 0 \end{cases}\end{aligned}$$ Now we assume ${\gamma}(s)<x(s)$ on $(0,t)$. Then $R({\gamma}(s),I(s))<R(x(s),I(s))=0$ on $(0,t)$, which yields $$0=u(x(t),t)= \int_{t_0} ^{t} \left( -\frac{ \dot {\gamma}^2}{4} +R({\gamma}(s),I(s) \right) +u_0({\gamma}(0))<0.$$ Hence, There exists $t' \in (t_0,t)$ such that ${\gamma}(t')=x(t')$. On the other hand, ${\gamma}(s)$ satisfies the Euler-Lagrange equation, which is, $$\label{el} \ddot {\gamma}(s) + R_x({\gamma}(s),I(s))=\ddot {\gamma}(s) + b'({\gamma}(s))=0.$$ Integrating the equation from $t'$ to $t$ gives $$0=\dot {\gamma}(t)-\dot {\gamma}(t_0) = \int_{t_0}^{t} b'({\gamma}(s)) >0,$$ by the lemma above. Therefore, ${\gamma}(s)>x(s)$ on $(0,t)$. Proof of the theorem \[main\] ============================= We assume that we have two pairs of solutions $(u_1,I_1)$ and $(u_2,I_2)$ to (\[HJ\]) for $n=1$ and consider two cases. Let us fix the time $T$. [*[Case 1]{}*]{} : $I_1(s)$ and $I_2(s)$ intersect only at the origin for $s\in [0,T]$.\ Without loss of generality, let us assume $I_1<I_2$ except for the terminal point. Then $u_1$ is a viscosity supersolution to $$\label{comp} \begin{cases} (u_2)_t={(u_2)_x}^2+R(x,I_2(t)) &\text{in } {\mathbb{R}}\times (0,t],\\ u(x,0)=u_0(x) &\text{on } {\mathbb{R}}. \end{cases}$$ By the comparison principle and the maximum constraint, we have $x_1(s)=x_2(s)$ for all $s$, where $x_1,x_2$ are defined as above, which is a contradiction. [*[Case 2]{}*]{} : $I_1(s)$ and $I_2(s)$ intersect at more than one point including the terminal point $t$. Let $t_0<t_1 \in [0,t]$ be points such that $$I_1(t_i)=I_2(t_i) \text{ for } i=1,2.$$ Hence, we have $x_1(t_0)=x_2(t_0):=\alpha$ and $x_1(t_1)=x_2(t_1):=\beta$. In addition to that, we may assume that $$I_1>I_2 \text{ for } i \in (t_0,t_1).$$ For the $t_i$’s above, we define ${\gamma}_1(s)$ and $\eta_1(s)$ as optimizing trajectories corresponding to $I_1$ whose terminal points are $\alpha$ and $\beta$ respectively. Similarly, one can define ${\gamma}_2(s)$ and $\eta_2(s)$ as optimizing trajectories corresponding to $I_2$ whose terminal points are $\alpha$ and $\beta$ respectively. By Proposition \[imp\] and Lemma \[lemma\], for each $i=1,2$, ${\gamma}_i$ satisfies $$\begin{cases} \ddot {\gamma}_i +2b'({\gamma}_i) =0,\\ \dot {\gamma}_i(t)=0,\\ {\gamma}(t)=\alpha. \end{cases}$$ Similarly, for each $i=1,2$, $\eta_i$ is a solution to $$\label{comp} \begin{cases} \ddot \eta_i +2b'(\eta_i) =0,\\ \dot \eta_i(t)=0,\\ {\gamma}(t)=\beta. \end{cases}$$ Therefore, ${\gamma}_1={\gamma}_2:={\gamma}$ and $\eta_1=\eta_2=\eta$. Applying this property to the relations $$\begin{aligned} 0=u_1(\beta,t_1)=\int_0^{t_1} \left(-\frac{\dot {\gamma}^2 }{4} + b({\gamma})-Q(I_1)\right)ds+u_0({\gamma}(0),0),\\ 0=u_2(\beta,t_0)=\int_0^{t_1} \left(-\frac{\dot {\gamma}^2 }{4} + b({\gamma})-Q(I_2)\right)ds+u_0({\gamma}(0),0),\\ 0=u_1(\alpha,t_1)=\int_0^{t_1} \left(-\frac{\dot \eta^2 }{4} + b({\gamma})-Q(I_1)\right)ds+u_0(\eta(0),0),\\ 0=u_2(\alpha,t_0)=\int_0^{t_1} \left(-\frac{\dot \eta^2 }{4} + b({\gamma})-Q(I_2)\right)ds+u_0(\eta(0),0),\end{aligned}$$ we end up getting $$0=\int_{t_0}^{t_1} \left(Q(I_1)-Q(I_2)\right) ds,$$ which contradicts $I_1>I_2$ on $(t_0,t_1)$. [30]{} S. Armstrong, H. V. Tran, *Viscosity solutions of general viscous Hamilton–Jacobi equations*, Mathematische Annalen. 361 (2014), 647-687. G. Barles, *Discontinuous viscosity solutions of first-order Hamilton-Jacobi equations: a guided visit*, Nonlinear Analysis: Theory, Methods & Appl. 20 (1999), no. 9, 1123-1134. P. Cannarsa, C. Sinestrari, *Semiconcave Functions, Hamilton-Jacobi Equations, and Optimal Control*, Progress in Nonlinear Differential Equations and Their Applications G. Barles, S. Mirrahimi, B. Perthame, *Concentration in Lotka-Volterra parabolic or integral equations: a general convergence result*, Methods Appl. Anal. 16 (2009), no. 3, pp.321-340. G. Barles, B. Perthame, *Concentrations and constrained Hamilton-Jacobi equations arising in adaptive dynamics*, Contemporary Math. 439 (2007), 57-68. O. Diekmann, P.-E. Jabin, S. Mischler, B. Perthame, *The dynamics of adaptation : an illuminating example and a Hamilton-Jacobi approach*, Th. Pop. Biol. 67 (2005), no. 4, 257-271. M. G. Crandall, L. C. Evans, P.-L. Lions, *Some properties of viscosity solutions of Hamilton-Jacobi equations*, Transaction of American Mathematical Society, 282 (1984), no. 2, 487-502. O. Diekmann, *Beginner’s guide to adaptive dynamics*, Banach Center Publications 63 (2004), 47-86. S. A. H. Geritz, E. Kisdi, G. Mészena, J. A. J. Metz, *Dynamics of adaptation and evolutionary branching*, Phy. Rev. Letters 78 (1997), 2024-2027. S. A. H. Geritz, E. Kisdi, G. Mészena, J. A. J. Metz, *Evolutionary singular strategies and the adaptive growth and branching of the evolutionary tree*, Evolutionary Ecology 12 (1998), 35-57. S. A. H. Geritz, E. Kisdi, , M. Gyllenberg, F. J. Jacobs, J. A. J. Metz *Link between population dynamics and dynamics of Darwinian evolution*, Phy. Rev. Letters 95 (2005), no. 7. N. Q. Le, H. Mitake, H. V. Tran, *Dynamical and Geometric Aspects of Hamilton-Jacobi and Linearized Monge-Ampere Equations*, Lecture notes in Mathematics 2183 (2016). S. Mirahimi, J.-M. Roquejoffre, *A class of Hamilton-Jacobi equations with constraint: Uniqueness and constructive approach*, J. of Differential Equations 250.5 (2016), 4717-4738. B. Perthame, G. Barles, *Dirac concentrations in Lotka-Volterra parabolic PDEs*, Indiana Univ. Math., J. 57 (2008), no. 7, 3275-3301. Y. Kim, *Wellposedness for constrained Hamilton-Jacobi equations*, preprint [^1]: Supported in part by NSF grant DMS-1664424
{ "pile_set_name": "ArXiv" }
--- abstract: 'We simulate black hole binary interactions to examine the probability of mergers and black hole growth and gravitational radiation signals using a specific initial distribution of masses for black holes in globular clusters and a simple semi-analytic formalism for dynamical interactions. We include 3-body recoil and the latest results in numerical relativity for gravitational radiation recoil. It is found that while 99% of binaries are ejected from low metallicity, low mass clusters; metal rich massive clusters retain 5% of their binaries. An interesting fraction of the ejected binaries, especially those from high mass, high metallicity systems, merge on timescales short enough to be gravitational radiation sources during their mergers with rates approaching those expected for galactic field black hole binaries. While the merger rates are comparable, the much larger mass of these binaries and their localization will make them appealing targets for advanced LIGO. We single out two possible Milky Way clusters (NGC 6441 and NGC 6388) as having the properties for a good probability of retention.' author: - 'Kenneth Moody, Steinn Sigurdsson' title: 'Modeling the Retention Probability of Black Holes in Globular Clusters: Kicks and Rates' --- Introduction ============ Observed black hole masses occupy two regimes, $M_{BH}\lesssim100M_{\sun}$ for black holes formed from core collapse supernova, and supermassive black holes with $M_{BH}\gtrsim10^6M_{\sun}$ which reside in the centers of galaxies. Observations of some objects, however have suggested that a middle regime of intermediate mass black holes (IMBH, review by @Miller04) could exist with masses between stellar and supermassive black holes. Ultraluminous X-ray sources ($L_X>10^{39}$ ergs sec$^{-1}$) have been found in intense star forming regions outside the nuclei of some galaxies [@Kaaret01; @Matsumoto01; @Fabbiano01], and recently even in one globular cluster [@Maccarone07]. The lower limit on the mass for these objects, assuming isotropic emission at the Eddington limit, is a few hundred solar masses. A stellar mass black hole would require special geometry of its accretion disk for sufficient beaming to occur and the accretion to be sub-Eddington [@King01], or need special conditions on the gas to provide a super-Eddington accretion rate [@Begelman01]. Conversely, a supermassive black hole ($M\gtrsim10^6 M_{\sun}$) would experience dynamical friction and sink to the center of its galaxy in too short a time to be plausibly observed at locations within the host galaxies where the IMBH candidates are projected to be seen today [@Kaaret01]. An intermediate mass for these objects would seem to be indicated. In the nearly bulgeless galaxy NGC4395, there has been found [@Fillipenko03] an AGN the black hole mass for which seems to be $\lesssim 10^5M_{\sun}$, which is in the upper range of IMBH masses. An object with a similar mass may also exist in the galaxy POX 52 [@Barth04]. Another possible place to look for IMBHs besides in starburst regions in galaxies is in the centers of globular clusters. Observations have seen an increase in the mass-to-light ratio towards the centers of two globular clusters that might be consistent with a massive object, a $M>10^4M_{\sun}$ object in the Andromeda Galaxy cluster G1 [@GRH02; @GRH05], a $4\times10^4M_{\sun}$ object found by @Noyola08 in the cluster $\omega$ Cen, and a few thousand solar mass object in M15 @vanderMarel02 [@Gerssen02], although in the case of M15, @Baumgardt03 are able to simulate the observations without an IMBH using smaller compact objects. The velocity dispersion of the central stars in the cores of these globular clusters as compared to the conjectured mass of the IMBH put these clusters on the same M-$\sigma$ relation as the bulges of galaxies with supermassive black holes [@GRH02; @vanderMarel02; @Gerssen02]. While the theory of the origin of the M-$\sigma$ relation for supermassive black holes would probably not apply to globular clusters, it is intriguing that, at least in these two cases, the IMBHs in these clusters are consistent with it. Two possible formation scenarios for the formation of IMBHs within a globular cluster have been proposed recently. The first involves the process of core collapse, by which the heavier stars in a cluster first sink to the middle through mass segregation. The stellar density goes very high, and might be sufficient that several stars collide forming a very large star (mass a few hundred solar masses), which collapses directly to an IMBH [@Begelman78; @PortZ02; @Freitag07]. Otherwise, stellar evolution causes the high mass stars to form black holes, which can become binaries through exchange into existing binaries of lower mass main sequence or neutron stars [@Sigurdsson93b]. We assume that formation of an IMBH by runaway merger does not occur in this case. Three-body interactions, which in the cores of clusters are dominated by interactions with all three objects being black holes, can then begin to work to harden binaries to the point where they merge. It is this scenario we intend to investigate. Previous studies of dynamical formation of a IMBH from stellar mass black holes have been performed. Black holes have been shown to be dynamically important in such aspects as the radius-age relation of Magellanic Cloud clusters [@Mackey07; @Mackey08], and even in galactic nuclei [@Lee95]. @PortZ00 include a study of how important black hole binaries in clusters are to gravitational wave research. Recently, @HB07 did a study of mergers of black holes in a system already containing a few hundred solar mass black hole, which represents the next step in IMBH formation after our work. @Kulkarni93 and @Sigurdsson93a used only 10 M$_{\sun}$ black holes, and determined that the formation of 10$^3$ M$_{\sun}$ objects is possible. On the other hand, @Miller02 showed that these binaries tend to be ejected before reaching a size at which recoil becomes unimportant, precluding further growth. Since black holes are produced from progenitors with a wide range of masses (20-100 M$_{\sun}$) and have varied evolution histories (wind losses and mass transfer) just prior to becoming black holes, a distribution in masses may better reflect the actual situation in globular clusters. @OLeary06 did a study using the distribution of black hole masses and binary periods as given in @Belczynski04 and a more complicated method of computing interactions than the semi-analytic model we use. They use the old prescription for gravitational radiation recoil similar to that found in @Favata04. Numerically simulating black hole mergers through ringdown has now been done [@Gonzalez06] and definitive recoil velocities determined, so that the sole remaining uncertainty in determining the circumstances of black hole mergers is the initial distributions of their masses. Our semi-analytic method can more quickly respond to updates in stellar population synthesis models than direct many-body integration. We describe the conditions under which the simulations were done, including initial conditions of the binary and the analytical form of the 3-body interactions and relevant time scales, then report results from several simulation runs, with a few parameters (e.g. metallicity) adjusted after each one. Finally, we discuss what the results imply for observed systems and suggest two systems that fall in the higher probability category for harboring an IMBH. Simulation Conditions ===================== We simulate the history for a total of 100,000 binaries for each set of initial conditions. The ensemble of initial conditions includes two values each for metallicity and escape velocity. An examination of @Belczynski04 shows that there exist only two distinct shapes of the period and mass distributions based on metallicity, therefore we only include a qualitative distinction with the changeover coming at $\log [Fe/H]=-1.3$ for observed clusters. The escape velocities were chosen as a proxy for several properties to represent moderately sized clusters and heavy clusters; smaller clusters which might have had a lower value for the escape velocity are not expected to retain any binaries that interact due to their extremely shallow potentials. The models’ initial conditions are described in Table \[modeltab\]. Each binary history is run until one of three fates is determined: ejection as a binary through 3-body superelastic recoil (Fate 1), ejection as a single object upon merger from gravitational radiation recoil (Fate 2), or a retained merged single object (Fate 3). For those systems which come under fate 1, we also calculate the gravitational radiation merger timescale for the ejected binary and determine the fraction of those in the run that coalesce within a Hubble time. Results for the runs are in Table \[restab\]. Whereas previously (c.f. @Sigurdsson93a) the distribution of black hole masses has been assumed to be single-valued at 10M$_{\sun}$, we start with a multi-valued initial mass function (IMF) for black holes. Based on figure 3 of @Fryer01 for f=1 (fully efficient supernova) and stellar IMF power-law index $\gamma$=2.0, we propose using for our low metallicity distribution a smooth power law, which gives the probability of a black hole having mass $M$ proportional to $10^{-.05M/M_{\sun}}$ for masses in the range 3M$_{\sun}$ to 80M$_{\sun}$. This is the same analytic form as the low metallicity (Z=0.001) IMF found by @Belczynski04 for his standard model parameters. @Fryer01 calculate a black hole IMF for single progenitor stars, and are used by @Belczynski04 to predict black hole masses from a range of progenitor masses, while also considering binary evolution effects such as common envelope phases. The most uncertainty in determining mass functions from binary evolution models comes during the common envelope phase and in calculating mass loss. These are common to any attempt to use a distribution of masses. The IMFs in @Belczynski04 are presented as histograms, but we convert these to analytical probability distributions for ease of use with computer-based simulations. Binaries are constructed using the IMF to pick both masses, a distribution in periods taken from @Belczynski04 as appropriate for the metallicity studied and also converted to an analytical form, and an eccentricity from a thermal distribution (P(e)=2e). We chose the masses independently as we expect the stars to have, for the most part, developed independently in well separated binaries. The simulations take place in a regime, being the center of a dense cluster, where all of the stars not heavy enough to have become evolved have been ejected to the outskirts of the cluster, except in the first  100Myr when massive stars were evolving rapidly and drastically changing core conditions. We assume that we start after this time, unlike other studies such as @PortZ02, and have a quasi-static cluster environment. The forms for both the IMF and periods are as follows [@Belczynski04]: $$P_{{\rm low Z}}(M)=0.152\cdot10^{-0.05M/M_{\sun}}{,\,\,\rm all\,\,M}.$$ $$P_{{\rm high Z}}(M)=\cases{0.028, &$3M_{\sun}\geq M\geq 15M_{\sun}$, \cr 10^{0.6M/M_{\sun}}, &$15M_{\sun}\geq M\geq 55M_{\sun}$, \cr 0, &$55M_{\sun}\geq M\geq 80M_{\sun}$.}$$ $$P(P)=\cases{\sqrt{1-(P-1)^2}, &$0\geq \log P({\rm days})\geq 2$ low Z only, \cr (P-2), &$2\geq \log P({\rm days})\geq 6$ all Z.}$$ Binaries are subjected to encounters with a third black hole whose mass is randomly drawn from the IMF. The time scale for the encounter $t_{enc}$ $$t_{enc}=1.5\times10^9 {\rm years} \frac{m_3 v_{10}}{\mu_{1\,2}m_T a_{AU}n_4}$$ is calculated from equation (2.9) of @Sigurdsson93b with $m_3$ the mass of the third black hole, $\mu_{1\,2}$ the reduced mass of the binary, $m_T$ the total mass of all three objects, the masses being expressed in terms of $M_{\sun}$, $a_{AU}$ the binary’s semimajor axis, $n_4$ the density of stars $n/10^4$ pc$^{-3}$, and $v_{10}$ the relative velocity of the third object $v/10$ km/s, which is taken to be 1 for typical globular clusters, which have velocity dispersions on the order of 10 km/sec. For this work we use a value of the dimensionless cross section $\tilde{\sigma}=10$ as defined in equation (2.7), as this value is broadly consistent with interacting systems having mass ratios in the range of those in our simulations as given in tables 3A and 3B of @Sigurdsson93b. This time scale is compared to the time for merger by emission of gravitational radiation [@Peters64] $$t_{GW}=3.151\times10^{17}{\rm years}\ g(e) \left(\frac{a}{AU} \right)^4 \left(\frac{M_{\sun}}{m_1}\right) \left(\frac{M_{\sun}} {m_2}\right) \left(\frac{M_{\sun}}{m_1+m_2}\right)$$ $$g(e)=\left(1-e^2\right)^{7/2} \left(1 + \frac{73}{24}e^2 + \frac{37}{96}e^4 \right)$$ For $t_{enc}<t_{GW}$, the encounter takes place. This entails choosing a new eccentricity from the thermal distribution and a change in the semimajor axis $a$ such that the binding energy of the binary is changed by $$\Delta = 1-\frac{a_{in}}{a_{out}}\frac{m_am_b}{m_1m_2}$$ where $a_{in}$ and $a_{out}$ are the starting and ending semimajor axes of the binary for the encounter, $m_1$ and $m_2$ are the masses of the two original objects in the binary, and $m_a$ and $m_b$ are the masses of the two new objects. Based on the results in @Sigurdsson93b, we assume that the two most massive of the three interacting objects form the new binary, leading to the possibility of membership change. If there is not a change, $a$ is simply reduced by a factor $(1-\Delta)$. With an exchange in membership, it is possible for $a$ to increase dramatically. For simplicity, we choose a fixed value of $\Delta = 0.4$, characteristic of the mean energy transferred in the same encounters from @Sigurdsson93b that gave us our value for $\tilde{\sigma}$. This is warranted, if there are multiple encounters per system before ejection. For a check on the adequacy of a single value of $\Delta$, we ran one set of simulations allowing $\Delta$ for each interaction to vary in a normal distribution around 0.4. Even with a variance of 0.2, the effect was negligible. We do not track stellar interactions, implicitly we assume that we are in a regime where there are multiple black holes which have formed a dense sub-core in the cluster, and that the interactions of these black holes dominates the fate of any binary. In general, interactions of the sub-core with stars are not important. They become very important in late stages, particularly for the ”last” black hole binary. Discussion of stellar interactions with the binary is beynd the scope of this paper. Besides changes in the internal dynamics of the binary, the conservation of momentum among the two systems requires the binary to recoil. The magnitude of the recoil is $$v_{rec}=\frac{m_e}{m_T}\sqrt{\frac{m_3(m_1+m_2)} {m_e(m_a+m_b)}\sigma_{GC} + \frac{2\Delta m_TGm_1m_2}{m_e(m_a+m_b)a_{in}}}$$ If the recoil is smaller than the assumed escape velocity of the globular cluster, the time is incremented by $t_{enc}$, and the run continues by choosing a black hole mass independently from the mass distribution. The dynamical friction timescale for the binary is approximately $\langle m\rangle/M_{BH} t_r$ [@OLeary06], where $\langle m\rangle$ is the average stellar mass, and $t_r$ is the relaxation timescale. Very few binaries are kicked with the narrow range of velocity required to have a turning point of several half-mass radii, and these are highly radial orbits. Therefore, the core relaxation time applies. The core relaxation times for most clusters are $10^7-10^8$ years [@Harris96], so most binaries will return to the core in less than a million years. If the binary is ejected, the run is stopped and $t_{GW}$ is calculated for the binary. Those for which $t_{GW}$ is less than 10$^{10}$ years may be field gravitational radiation sources. The run will also be stopped once $t_{enc}>t_{GW}$, at which time the recoil velocity from asymmetric emission of gravitational radiation is calculated. We used the zero spin expression for the gravitational radiation recoil from @Gonzalez06 $$v_{GW}=1.20\times 10^4\eta^2\sqrt{1-4\eta}(1-0.93\eta)\ {\rm km/sec}$$ where $\eta$ is the symmetric mass ratio defined using $\eta=q/(1+q)^2$, $q$ being the mass ratio of the two objects in the binary ($0\geq q\geq1,0\geq \eta \geq \ 0.25$). Stellar mass black holes, unlike supermassive black holes in galaxy centers which have accreted most of their mass from a thin disk where the spin goes to 0.98, are not likely to have a large spin parameter. @OShaun05b find by analogy with neutron star birth spins [@Lorimer05; @Kramer03; @Migliazzo02] that expected spins should be less than 0.1, unless otherwise spun up by fallback from the supernova explosion. @Burrows07 find that only rapidly spinning cores may produce the phenomena called hypernovae. Since these types of objects are rare, we may infer that most supernovae that produce black holes make slowly spinning ones. Any accretion that does occur while the black hole is in a binary with a mass donor is expected by @Belczynski08 to increase the spin parameter $a=J/M^2$ beyond 0.5. From the fully spin dependent form of the recoil velocity [@Campanelli07] and using the approximation of $1/\sqrt{2}$ for the values of the sine and cosine of the angles, we find that $v_{GW}(a)/v_{GW}(a=0)$ goes above 2 for values of $a>\sim 0.4$ except for extreme mass ratios which are more sensitive to spin. The merged object is then ejected or retained in the globular cluster depending on the magnitude of $v_{GW}$. Results ======= As seen in Table \[restab\], the most likely conditions for a black hole to be retained are in massive, metal-rich clusters. The change in mass distribution with metallicity is the main driver of whether or not a binary may be retained. What changes most between the two metallicities is the distribution of mass ratios, seen in Figure \[ratio\]. The initial distribution of mass ratios for metal poor binaries is nearly constant above 0.2, which is the condition @OLeary06 place on their binaries [*a priori*]{}. For the distribution of initial mass ratios for high metallicity binaries, there is a peak at $q=0.25$ due to systems with one member from each of the two parts of the distribution (centered at 10 and 40M$_{\sun}$). The mass ratio distribution of ejected binaries is shifted toward higher values of $q$ for both high and low metallicity distributions, meaning more equal mass binaries are more likely to be ejected. This follows from previous attempts at this problem [@Kulkarni93; @Sigurdsson93b] with the failure of equal mass binaries to produce a retained object. We compare our results of ejected binaries to those of @OLeary06 by collecting enough of our runs within a single model to create an N=512 cluster. The models of theirs most similar to our were the e5e5king11, v2e5k11, and v3e5k11 models for low escape velocity, and the e5e5king7, v2e5k7, and v3e5k7 models for high escape velocity. All of these fall into what we consider low metallicity models, as that corresponds to the models of @Belczynski04 the authors used. In the models indicated, @OLeary06 find the ejection fraction of black holes in binaries in low $v_{esc}$ clusters is 0.14, and a fraction of 0.1 in high $v_{esc}$ clusters. We find similar values for the fraction of ejected black holes in binaries. The large numbers of ejected binaries produced in each cluster make them an interesting target of investigation. Some fraction of these we find have a $t_{GW}$ less than 10Gyr, and so would make a background source for gravitational radiation detectors [@PortZ00]. Which property of the cluster is more important in determining the efficiency of producing binaries that will merge in less than a Hubble time is complex. High mass metal rich clusters produce the most binaries that merge in less than 10Gyr. The next most come from high mass metal poor clusters. Low mass metal poor clusters produce more binaries that merge in less than 10Gyr than low mass metal rich clusters. The change in Hubble time mergers with mass is expected as a heavier cluster would allow the binary to become harder before ejecting it. The binaries for the most part stay within their host galaxy. Figure \[vrec\] shows the distribution of velocities of the ejected binaries. For the two low metallicity models, which have binaries that start with tight orbits having $\log P(days)<2$, there are a few systems ( 1%) that are ejected with a velocity higher than 300 km/sec, but most binaries (and all of the high metallicity ones) have $v_{rec}< 200$km/sec. This means that while they leave their parent cluster, they are still confined to their parent galaxy unless it is a dwarf galaxy. The distributions of masses for the retained merged objects are in Figure \[mass\]. For low metallicity systems, the distribution is flat up to 60M$_{\sun}$, after which it drops. High metallicity systems show one peak of 20-30M$_{\sun}$, and another between 80 and 100M$_{\sun}$, which reflects the underlying initial mass distribution. For ejected objects, the low metallicity systems show a monotonic decline from 15M$_{\sun}$ to the maximum mass seen at  120M$_{\sun}$, with a slight break downward at 60M$_{\sun}$. For the high metallicity systems, there is a peak around 40-60M$_{\sun}$ where the distribution of masses for retained objects has a deficit. The currently used recoil velocity function has its peak at a mass ratio of about a third, so that if a binary in the high metallicity model consists of one member from each of the two regions, it will have a total mass of about 40-60M$_{\sun}$ and a mass ratio of 0.3-0.4 which will most likely be ejected, whereas a binary with both members from the same region will have a mass of either 20 or 80-100M$_{\sun}$ and $q$ close to one and be retained (if it survives 3-body interactions of course). While the binaries are in the globular cluster, they may go through short-lived phases with large semi-major axes due to exchanges of membership. These stages may be important in transferring angular momentum from the binary to the cluster as a whole through interactions with stars as shown in @Mapelli05. For a fraction of the binaries, the histories of $a$ are recorded and examined to determine how much time they spend with $a>10^2$, $10^3$, and $10^4$ AU. The distributions of the two high metallicity samples are the same, but the low mass metal poor model shows fewer binaries that get to high separations. The low metallicity low and high mass models have 34.2% and 16.2% that never have $a>10$AU respectively, while for high metallicity this percentage is 41.6%. Other than this difference, the distributions of time spent at high separations is similar. The plots for the number of binaries that exist at high separations for the time indicated are in figure \[hia\]. Discussion ========== The event rate from merging black hole binaries can be calculated from the fraction of systems that merge within a Hubble time and the relative contributions from low and high metallicity systems and light or massive clusters. To conservatively estimate the event rate, we assume 100 globular clusters per galaxy (e.g. the Milky Way is currently thought to have about 150) and 100 BH per globular cluster ($N_{BH}\sim10^{-4}N_{\star}$). We assume that the break between high versus low metallicity is at an \[Fe/H\] of -1.3 and that light globulars have $v_{esc}$ of less than 30 km sec$^{-1}$ and heavy globulars above this value. The escape velocity for a globular cluster is given by $v_{esc}=\sqrt{2\Phi_0}$, where $\Phi_0$ is the central potential of the cluster, $W=\Phi_0/\sigma^2$ is the King parameter and is correlated with the cluster concentration, and $\sigma$ is approximately equal to the velocity dispersion except in the case for shallow globulars. We determine the concentration and thus $W$ from the catalog of @Harris96, while 1D velocity dispersion data were obtained from @Pryor93. For the 56 Milky Way clusters for which we could determine the escape velocity, we find that the percentage of clusters in each of our models is as follows: A 45% (25), B 21% (12), C 20% (11), and D 14% (8). Including data from table \[restab\] on the number of mergers within a Hubble time, we find that ejected binaries account for $\sim$640 mergers per galaxy in a Hubble time, with another 335 coming from those binaries that merge while still in the cluster. Over a Hubble time this gives a rate of 10$^{-7}$ per year per galaxy. These are very conservative estimates for the rate, as the Milky Way is assumed to have about 150 globular clusters, and giant ellipticals can have on the order of $10^3$. Assuming a value of 300 globular clusters per galaxy and 300 black hole binaries per cluster, there would be an order of magnitude jump in the rate to 10$^{-6}$ per year per galaxy. We also expect a further increase in rate from the additional mergers produced by black holes that are retained after their first merger. Galactic binary BH merger rates are estimated at $~10^{-6}$ per year [@OShaun05a]. The mergers are expected to be delayed from the formation of the clusters, which in the case of globulars is close to the beginning of the universe. The last interaction before the binary is ejected typically happens when the semimajor axis is 0.1-1 AU, giving a $t_{enc}$ of $~10^8-10^9$ years. The timescales for the gravitational merger of the ejected binaries spans a wide range of values ($5<\log t_{GW}<20$). Figure \[tGW\] shows the distribution of merger timescales for ejected binaries for each of the models. We find that the percentage of binaries which merge between 1 and 10 Gyr is 2.6% for model A, 2.3% for model B, 4.3% for model C, and 8.2% for model D. While we have used a thermal distribution ($P(e)=2e$, $\langle e\rangle=0.67$) for the eccentricity after an exchange, the 3-body study by @Sigurdsson93b found that this works for equal mass exchanges, but for non-equal masses, the eccentricities may be higher ($\langle e\rangle\approx 1-1.3(m_3/m_2)$). This does not affect recoil velocities, but the $t_{GW}$ would be shortened, and the expected rates of black hole mergers would increase by a factor of a few. If we choose black hole binaries as we do for the simulation runs and determine their $t_{GW}$ without any interactions, we find that for metal rich systems, only 0.1% merge in less that a Hubble time, whereas 7% of low metallicity binaries do so. This we explain as a model dependent result, the low metallicity period distribution includes systems which have periods shorter than 100 days while the metal rich distribution does not. Interactions are of great importance in metal rich systems for producing observable mergers, while they are ambivalent in metal poor systems. The chirp masses for cluster binary mergers are much higher due to exchanges undergone while the binary was in the cluster. We find that, while galactic mergers have chirp masses of 3-8M$_{\sun}$ [@Belczynski07], the chirp masses for the ejected binaries are 15-25M$_{\sun}$ in the metal rich case and 20-40M$_{\sun}$ in metal poor clusters. The higher chirp masses, while dependent on the models used for the initial mass function of the black holes, is a distinct prediction characteristic of the globular cluster binaries, and easily observable by gravitational radiation instruments. Since the strain due to gravitational radiation scales as $M/r$, the factor of 4-6 increase in mass of the cluster binaries makes them visible over a factor of 60-200 larger volume, which makes them almost as important source as galactic binaries for the conservative values of GC/galaxy and binaries/GC. If we assume the less conservative numbers, the cluster binary inspirals would dominate the signal. LIGO will have an abundance of targets from the ejected binaries. Our work in this paper provides a first step from population synthesis to the possible formation of an IMBH in the center of a globular cluster. Examining a second merger once the merged object has exchanged into a new binary is beyond the scope of this work, but has been studies by @HB07. To connect our theoretical models to observed clusters, a plot of metallicity versus $v_{esc}$ is given in figure \[metvesc\] using metallicity data from the catalog by @Harris96 and the escape velocity as described above. Two clusters that have both high metallicity and $v_{esc}>50$ km sec$^{-1}$ are NGC 6388 and NGC 6441. Both of these clusters lie within 4 kpc of the galactic center. These clusters are most well known for their contribution to the “second parameter” problem in that they have more extended blue horizontal branches than their metallicity would indicate. It is speculated that this might be due to dynamical interactions in the clusters [@Rich97; @Miocchi07]. They note that the M31 cluster G1, a cluster suspected of having an IMBH by @GRH02, also shows an extended blue horizontal branch. These clusters may have been at one point the nuclei of dwarf galaxies, as a couple of other suspected nuclei appear in interesting regions of the metallicity-$v_{esc}$ plot. Other clusters suspected of being dwarf galaxy nuclei are M54 (due to its association with the Sagittarius dwarf galaxy) by @Ibata94 and $\omega$ Cen [@Norris96; @Norris97]. The presence of extended blue horizontal branch stars in the metal-rich clusters NGC 6388 and NGC 6441 is thought to give a similar argument for their being formed in a similar manner [@Piotto97]. A couple others which are less outstanding but still in the upper right part of the diagram are NGC 2808 and M62. Observations of variability in the recently discovered ULX in a globular cluster of NGC 4472 by @Maccarone07 lead to estimates of a 300M$_{\sun}$ IMBH, though their other solution gives a mass of 30M$_{\sun}$. They find a metallicity of the cluster of -1.7 from color-metallicity relations, and the luminosity gives it a absolute magnitude of -9.2. When compared to analogous clusters in the Milky Way (e.g. NGC 6273), this cluster fits into category C, low metallicity high mass. Figure \[metvesc\] places NGC 6388 and NGC 6441 in context with other massive, well studied globular clusters. In conclusion, we find that within our simplified model assumptions, most black hole binaries are ejected through gravitational 3-body interaction from the cluster into the general potential of the galaxy. Of those binaries that survive to merge by gravitational radiation, about 2/3 to half are ejected through gravitational radiation recoil. Between 0.5% and 3.5%, depending on metallicity and cluster escape velocity, of all black hole binaries in clusters are predicted to be retained upon merger of the binary, with typical final masses of 20-50 M$_\sun$, but in some instances over 100M$_{\sun}$. Of course if other formation channels dominate, or there is significant gas accretion after the dynamical interaction phase, then the final black hole masses may be very different (higher if there is significant accretion). We find that the rate per galaxy of black hole binary mergers from the globular cluster population, through gravitational radiation is competitive with the total merger rate from the parent galaxy, but biased towards higher masses. While most globular clusters in massive galaxies probably form at high redshift, this suggests that black hole binary coalescence from clusters in low mass, nearby star forming galaxies may be a significant contributor to the total high frequency gravitational radiation signal in the local universe. The current results are dependent on the exact form of the initial conditions of mass distributions and period distributions obtained from population synthesis. As the formation mechanisms for black holes become more well understood, it would be appropriate and easy to refine the results in this paper. We thank the Center for Gravitational Wave Physics for support, and Ben Owen and Richard O’Shaughnessy for helpful discussions in making this paper. We would also like to thank the referee for his helpful suggestions. This research was funded by NSF grant PHY 0114375. Barth, A. J., Ho, L. C., Rutledge, R. E., Sargent, W. L. W.,  2004, , 607, 90 Baumgardt, H., Hut, P., Makino, J., McMillan, S., Portegies Zwart, S.,  2003, , 582, 21 Begelman, M. C.,  2001, , 551, 897 Begelman, M. C., Rees, M. J.   1978, , 185, 847 Belczynski, K., Sadowski, A., Rasio, F. A., 2004 , , 611, 1068 Belczynski, K., Taam, R. E., Kalogera, V., Rasio, F. A., Bulik, T.,  2007, , 662, 504 Belczynski, K., Taam, R. E., Rantsiou, E., van der Sluys, M.,  2008, , 682, 474 Burrows, A., Dessart, L., Livne, E., Ott, C. D., & Murphy, J.,  2007, , 664, 416 Campanelli, M., Lousto, C. O., Zlochower, Y., & Merritt, D.,  2007, , 659, 5 Fabbiano, G., Zezas, A. L., Murray, S. S.  2001, , 554, 1035 Favata, M., Hughes, S. A., Holz, D. E.,  2004, , 607, 5 Fillipenko, A. V., & Ho L. C.,  2003, , 588, 13 Freitag, M., Guerkan, M. A., Rasio, F. A.,  2007, ASPC, 367, 707 Fryer, C. L., Kalogera, V., 2001, , 554, 548 Gebhardt, K., Rich, R. M., Ho, L. C.  2002, , 578, 41 Gebhardt, K., Rich, R. M., Ho, L. C.  2005, , 634, 1093 Gerssen, J., van der Marel, R. P., Gebhardt, K., Guhathakurta, P., Peterson, R. C., & Pryor, C.  2002, , 124, 3270 Gonzalez, J. A., Sperhake, U., Brügmann, B., Hannam, M., Husa, S., 2007, Physical Review Letters, 98, 091101 Harris, W. E.  1996, , 112, 1487 Holley-Bockelmann, K., Gultekin, K., Shoemaker, D., Yunes, N., astro-ph 0707.1334 Ibata, R. A., Gilmore, G., Irwin, M. J.,  1994, Nature, 370, 194 Kaaret, P., Prestwich, A. H., Zezas, A., Murray, S. S., Kim D.-W., Kilgard, R. E., Schlegel, E. M., Ward, M. J.  2001, , 321, 29 King, A. R., Davies, M. B., Ward M. J., Fabbiano, G., Elvis, M.  2001, , 522, 109 Kramer, M., et al.,  2003, , 342, 1299 Kulkarni, S. R., Hut, P., McMillan, S. L. W.  1993, Nature, 364, 421 Lee, M. H.,  1995, , 272, 605 Lorimer, D. R., et al.,  2005, ASP Conf. Ser. 328, Binary Radio Pulsars, ed. F. A. Rasio & I. H. Stairs (San Francisco: ASP), 113 Mackey, A. D., Wilkinson, M. I., Davies, M. B., Gilmore, G. F.,  2007, , 379, 40 Mackey, A. D., Wilkinson, M. I., Davies, M. B., Gilmore, G. F.,  2008, , 386, 65 Mapelli, M., Colpi, M., Possenti, A., Sigurdsson, S.  2005, , 364, 1315 Maccarone, T. J., Kundu, A., Zepf, S. E., Rhode, K. L.  2007, Nature, 445, 183 Matsumoto, H., Tsuru, K. G., Koyama, K., Awaki, H., Canizares, C. R., Kawai, N., Matsushita, S., Kawabe, R.   2001, , 320, 30 Migliazzo, J. M., Gaensler, B. M., Backer, D. C., Stappers, B. W., Strom, R. G., & van der Swaluw, E.,  2002, ASP Conf. Ser. 271, Neutron Stars in Supernova Remnants, ed. P. O. Slane & B. M. Gaensler (San Francisco: ASP), 57 Miller, M. C., Hamilton D. P.   2002, , 330, 232 Miller, M. C. & Colbert, E. J. M. 2004, International Journal of Modern Physics D, 13, 1 Miocchi, P.,  2007, , 381, 103 Norris, J. E., Freeman, K. C., Mighell, K. J.,  1996, , 462, 241 Norris, J. E., Freeman, K. C., Mayor, M., Seitzer, P.,  1997, , 487, 187 Noyola, E., Gebhardt, K., & Bergmann, M.,  2008, , 676, 1008 O’Leary, R. M., Rasio, F. A., Fregeau, J. M., Ivanova, N., & O’Shaughnessy, R.,  2006, , 637, 937 O’Shaughnessy, R., Kalogera, V., & Belczynski, K.,  2005, , 620, 385 O’Shaughnessy, R., Kaplan, J., Kalogera, V., & Belczynski, K.,  2005, , 632, 1035 Peters, P. C.  1964, Phys. Rev. B, 136, 1224 Piotto, G., et al. 1997, Advances in Stellar Evolution, 84 Portegies Zwart, S. F., & McMillan, S. L. W.  2000, , 528, 17 Portegies Zwart, S. F., & McMillan, S. L. W.  2002, , 576, 899 Pryor, C. & Meylan, G., 1993, ASP Conf. Ser.  50: Structure and Dynamics of Globular Clusters, 357 Rich, R. M., et al.  1997, , 484, 25 Sigurdsson, S., Hernquist, L.  1993, Nature, 364, 423 Sigurdsson, S., Phinney, E. S.,  1993, , 415, 631 van der Marel, R. P., Gerssen, J., Guhathakurta, P., Peterson, R. C., & Gebhardt, K.  2002, , 124, 3255 [cccc]{} A & low & 30\ B & high & 30\ C & low & 50\ D & high & 50\ [ccccc]{} A & 95.7 (95744) & 3.8 (3790) & 0.5 (466) & 5.3 (5071)\ B & 98.2 (98242) & 1.0 (997) & 0.8 (761) & 4.1 (4023)\ C & 94.1 (94069) & 4.5 (4530) & 1.4 (1401) & 8.3 (7799)\ D & 92.7 (92666) & 3.8 (3842) & 3.5 (3492) & 13.3 (12362)\
{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the techniques of chiral effective field theories we evaluate the so called “generalized polarizabilities", which characterize the structure dependent components in virtual Compton scattering (VCS) off the nucleon as probed in the electron scattering reaction $eN\rightarrow e''N\gamma$. Results are given for both spin-dependent and spin-independent structure effects to ${\cal O}(p^3)$ in SU(2) Heavy Baryon Chiral Perturbation Theory and to ${\cal O}(\epsilon^3)$ in the SU(2) “Small Scale Expansion”.' address: - '$^a$ Forschungszentrum Jülich, Institut für Kernphysik (Th), D-52425 Jülich, Germany ' - '$^b$ Department of Physics and Astronomy, University of Massachusetts, Amherst, MA 01003, USA' - '$^c$ Institut f[" u]{}r Kernphysik, Johannes Gutenberg-Universit[" a]{}t, D-55099 Mainz, Germany' author: - 'Thomas R. Hemmert$^a$[^1], Barry R. Holstein$^{a,b}$[^2] Germar Kn[" o]{}chlein$^c$[^3], and Dieter Drechsel$^c$[^4]' title: | \ \ Generalized Polarizabilities of the Nucleon in Chiral Effective Theories --- Introduction ============ One of the primary goals of contemporary particle/nuclear physics is to understand the structure of the nucleon. Indeed this is being pursued at the very highest energy machines such as HERA and SLAC, at which one probes the quark/parton substructure, as well as at lower energy accelerators such as BATES, ELSA and MAMI, wherein one studies the low energy structure of the nucleon via electron scattering. In recent years another important low energy probe has been (real) Compton scattering, by which one can study the deformation of the nucleon under the influence of quasi-static electric and/or magnetic fields [@LV]. For example, in the presence of an external electric field $\vec{E}$ the quark distribution of the nucleon becomes distorted, leading to an induced electric dipole moment $$\vec{p}=4\pi\alpha_E\vec{E}$$ in the direction of the applied field, where $\alpha_E$ is the electric polarizability. The interaction of this dipole moment with the field leads to a corresponding interaction energy $$U=-{1\over 2}4\pi\alpha_E\vec{E}^2.$$ Similarly in the presence of an applied magnetizing field $\vec{H}$ there will be an induced magnetic dipole moment $$\vec{\mu}=4\pi\beta_M\vec{H}$$ and an interaction energy $$U=-{1\over 2}4\pi\beta_M\vec{H}^2.$$ For wavelengths large compared to the size of the system, the effective Hamiltonian for the interaction of a system of charge $e$ and mass $m$ with an electromagnetic field is, of course, given by the simple form $$H^{(0)}={(\vec{p}-e\vec{A})^2\over 2m}+e\phi,$$ and the Compton scattering cross section has simply the familiar Thomson form $${d\sigma\over d\Omega}=\left({\alpha_{em}\over m}\right)^2\left({\omega'\over \omega}\right)^2[{1\over 2} (1+\cos^2\theta)],$$ where $\alpha_{em}$ is the fine structure constant and $\omega,\omega'$ are the initial, final photon energies respectively. As the energy increases, however, so does the resolution and one must take into account also polarizability effects, whereby the effective Hamiltonian becomes $$H_{\rm eff}=H^{(0)}-{1\over 2}4\pi(\alpha_E\vec{E}^2+\beta_M\vec{H}^2).$$ The Compton scattering cross section from such a system (taken, for simplicity, to be spinless) is given then by $$\begin{aligned} {d\sigma\over d\Omega}&=&\left({\alpha_{em}\over m}\right)^2\left({\omega'\over \omega}\right)^2[{1\over 2} (1+\cos^2\theta)\nonumber\\ &-&{m\omega\omega'\over \alpha_{em}}[{1\over 2}(\alpha_E+\beta_M)(1+\cos\theta)^2 +{1\over 2}(\alpha_E-\beta_M)(1-\cos\theta)^2+\ldots].\label{eq:sss}\end{aligned}$$ It is clear from Eq.(\[eq:sss\]) that from careful measurement of the differential scattering cross section, extraction of these structure dependent polarizability terms is possible provided that i) the energy is large enough that these terms are significant compared to the leading Thomson piece and ii) that the energy is not so large that higher order corrections become important. In this way the measurement of electric and magnetic polarizabilities for the proton has recently been accomplished using photons in the energy range 50 MeV $<\omega <$ 100 MeV, yielding[@PPol] [^5] $$\begin{aligned} \alpha_E^p&=&(12.1\pm 0.8\pm 0.5)\times 10^{-4}\; {\rm fm}^3\nonumber\\ \beta_M^p&=&(2.1\mp 0.8\mp 0.5)\times 10^{-4}\; {\rm fm}^3. \label{abexp}\end{aligned}$$ Note that in practice one generally exploits the strictures of causality and unitarity as manifested in the validity of the forward scattering dispersion relation, which yields the Baldin sum rule[@bgm] $$\alpha_E^{p,n}+\beta_M^{p,n}={1\over 2\pi^2}\int_0^\infty{d\omega\over \omega^2} \sigma_{\rm tot}^{p,n}=\left\{ \begin{array}{ll}(13.69\pm 0.14)\times 10^{-4}{\rm fm}^3& {\rm proton}\\ (14.40\pm 0.66)\times 10^{-4}{\rm fm}^3& {\rm neutron} \end{array}\right.$$ as a rather precise constraint because of the small uncertainty associated with the photoabsorption cross section $\sigma_{\rm tot}^p$. [From]{} these results, which imply that the polarizabilities of the proton are nearly a factor of a thousand smaller than its volume, we learn that the nucleon is a relatively rigid object when compared to the hydrogen atom, for example, for which the electric polarizability and volume are comparable. Additional probes of proton structure are possible if one exploits its spin $\vec{S}$. Thus, for example, the presence of a time varying electric field in the plane of a rotating system of charges will lead to a charge separation with induced electric dipole moment $$\vec{p}=-\gamma_1\vec{S}\times{\partial \vec{E}\over \partial t}$$ and corresponding interaction energy $$U_1=-\vec{p}\cdot\vec{E}=\gamma_1\vec{E}\cdot\vec{S}\times (\vec{\nabla}\times\vec{B}),\label{spinpola1}$$ where we have used the Maxwell equations in writing this form. This is a quantum mechanical analog of the familiar Faraday rotation. (Note that the “extra” time or spatial derivative is required by time reversal invariance since $\vec{S}$ is T-odd.) Similarly other possible structures are[@Ragusa; @NS] $$\begin{aligned} \label{spinpolas} U_2&=&\gamma_2\vec{B}\cdot\vec{\nabla}\vec{S}\cdot\vec{E}\nonumber\\ U_3&=&\gamma_3\vec{E}\cdot\vec{\nabla}\vec{S}\cdot\vec{B}\nonumber\\ U_4&=&\gamma_4\vec{B}\cdot\vec{S}\times(\vec{\nabla}\times\vec{E}),\end{aligned}$$ and the measurement of these various “spin-polarizabilities” $\gamma_i$ via polarized Compton scattering provides a rather different probe for nucleon structure. Because of the requirement for polarization not much is known at present about such spin-polarizabilities, although from dispersion relations the combination $$\gamma_0^p\equiv \gamma_1^p-\gamma_2^p-2\gamma_4^p\approx\left\{ \begin{array}{cc} -1.34\times 10^{-4}\; {\rm fm}^4 & {\rm SAID}\cite{DHS}\\ -0.80\times 10^{-4}\; {\rm fm}^4 & {\rm Mainz}\cite{HDT} \end{array}\right.$$ has been evaluated and from a global analysis of unpolarized Compton data, to which it contributes at ${\cal O}(\omega^4)$, Tonnison et al.[@ton] have determined the so-called backward spin-polarizability to be $$\gamma_\pi=\gamma_1+\gamma_2+2\gamma_4=(27.7\pm 2.3\pm 2.5)\times 10^{-4}\; {\rm fm}^4.$$ Clearly such measurements represent an important goal for the future. At the same time it has come to be realized that a high resolution probe of nucleon structure is available, in principle, via the use of [*virtual*]{} Compton scattering—VCS—wherein virtual photons produced from scattered electrons are scattered off a nucleon into real final state photons, transferring a three-momentum $\bar{q}$ to the target. The outcome of such measurements is, in principle, $\bar{q}$-dependent values of the polarizabilities (usually termed “generalized polarizabilities” and denoted by GPs in the following) which can be thought of as the Fourier transforms of [*local*]{} polarization densities in the nucleon. At the present time a VCS experiment has already taken place at MAMI, and there exist approved experiments at BATES and TJNAF. Preliminary results have been reported from MAMI and will be discussed in the conclusion [@nstar]. It is therefore appropriate to have a base of solid theoretical predictions with which such data can be confronted. The here presented approach, which utilizes the techniques of chiral effective theories in the heavy fermion formulation, has already yielded several results[@HHKS1; @HHKS2]. In the first chiral calculation of generalized polarizabilities utilizing SU(2) Heavy Baryon Chiral Perturbation Theory (HBChPT) [@HHKS1], the leading momentum-dependent modification of the (generalized) electric ($\bar{\alpha}_E(\bar{q})$) and magnetic ($\bar{\beta}_M(\bar{q})$) polarizabilities was analyzed. Later, in a short communication[@HHKS2], numerical studies for the full $\bar{q}$-dependence of all 10 generalized (Guichon) polarizabilities were presented—again using the framework of SU(2) HBChPT. In this work we present the details behind the numerical study of ref.[@HHKS2] and, for the first time in the field of VCS, are able to present simple analytical expressions for all GPs in a momentum range from $0<\bar{q}^2<0.5$ GeV$^2$ utilizing SU(2) HBChPT. These new expressions greatly facilitate the study of the influence of the chiral “pion cloud” on the GPs and the comparison with model calculations. Furthermore, we also investigate the leading modifications of the GPs’ $\bar{q}$-dependence due to $\Delta$(1232) resonance contributions utilizing a different effective chiral lagrangian approach—the so called “small scale expansion” (SSE) [@HBDel]. SSE results have already been reported for real Compton scattering [@delta; @HHKK], and in the present work we generalize the analysis to the VCS case. In the next section we shall discuss the definition of the generalized polarizabilities, while in Section III we present an introduction to the way in which our heavy baryon calculations—valid to one loop—are carried out. In section IV we show how to connect our predictions to the general formulation of VCS and how to extract the desired generalized polarizabilities. In section V, we present the results of our calculations. Finally, we summarize our findings in a concluding section VI. Generalized Polarizabilities ============================ Recently a new frontier in Compton scattering has been opened (see, [*e.g.*]{}, [@NF]) and is now in the beginning of being explored: the study of the electron scattering process $e p \rightarrow e' p' \gamma$ ([*cf. Fig. 1*]{}) in order to obtain information concerning the virtual Compton scattering[^6] (VCS) process $\gamma^* N \rightarrow \gamma N$. As will be discussed below, in addition to the two kinematical variables of real Compton scattering— the scattering angle $\theta$ and the energy $\omega'$ of the outgoing photon—the invariant structure functions for VCS [@BAS; @Guichon] depend on a [*third*]{} kinematical variable, [*e.g.*]{} the magnitude of the three–momentum transfer to the nucleon in the hadronic c.m. frame, $\bar{q}\equiv|\vec{q}|$. As shown in ref. [@Guichon], the VCS amplitude can then be characterized in terms of $\bar{q}$-dependent GPs, in analogy to the well-known polarizability coefficients in real Compton scattering. However, due to the specific kinematic approximation chosen in [@Guichon] there does not exist a one–to–one correspondence between the real Compton polarizabilities and the GPs of Guichon et al. in VCS[@Guichon; @fm1; @fm2]. The advantage of VCS lies in the virtual nature of the initial state photon and the associated possibility of an [*independent*]{} variation of photon energy and momentum, thus rendering access to a much greater variety of structure information than in the case of real Compton scattering. For example, one can hope to identify individual signatures of specific nucleon resonances which cannot be obtained in other processes[@NF]. In this regard, it should be noted that a great deal of theoretical work already exists, such as predictions within a non–relativistic constituent quark model[@Guichon], a one–loop calculation in the linear sigma model[@Metz], a Born term model including nucleon resonance effects[@Vanderhaeghen], a HBChPT calculation of the leading $\bar{q}$-dependence of the generalized electric and magnetic polarizability [@HHKS1], a calculation of $\bar{\alpha}_E(\bar{q}^2)$ in the Skyrme model[@KM97] and the numerical study of all 10 GPs again utilizing HBChPT [@HHKS2]. For an overview of the status at higher energies and in the deep inelastic regime we refer to [@NF]. The GPs of the nucleon have been defined by Guichon et al. in terms of electromagnetic multipoles as functions of the initial photon momentum $\bar{q}$[@Guichon], $$\begin{aligned} P^{(\rho' L' , \rho L)S} (\bar{q}^2) & = & \left[ \frac{1}{\omega'^{L} \bar{q}^{L}} H^{(\rho' L' , \rho L)S} (\omega' , \bar{q}) \right]_{\omega' = 0} \, , \nonumber\\ \hat{P}^{(\rho' L' , L)S} (\bar{q}^2) & = & \left[ \frac{1}{\omega'^{L} \bar{q}^{L+1}} \hat{H}^{(\rho' L' , L)S} (\omega' , \bar{q}) \right]_{\omega' = 0} \, ,\end{aligned}$$ where $L$ ($L'$) denotes the initial (final) photon angular momentum, $\rho$ ($\rho'$) the type of multipole transition ($0 = C$ (scalar, Coulomb), $1 = M$ (magnetic), $2 = E$ (electric)), and $S$ distinguishes between non–spin–flip ($S=0$) and spin–flip ($S=1$) transitions. In addition, mixed–type polarizabilities, ${\hat{P}}^{(\rho' L' , L)S} (\bar{q}^2)$, have been introduced, which are neither purely electric nor purely Coulomb type. It is important to note that the above definitions are based on the kinematical approximation that the multipoles are expanded around $\omega' = 0$ and [*[only terms linear in $\omega'$ are retained]{}*]{}, which, together with current conservation, yields selection rules for the possible combinations of quantum numbers of the GPs. In this approximation, 10 GPs have been introduced in [@Guichon] as functions of $\bar{q}^2$: $P^{(01,01)0}$, $P^{(11,11)0}\,$, $P^{(01,01)1}\,$, $P^{(11,11)1}\,$, $P^{(01,12)1}\,$, $P^{(11,02)1}\,$, $P^{(11,00)1}\,$, ${\hat{P}}^{(01,1)0}\,$, ${\hat{P}}^{(01,1)1}\,$, ${\hat{P}}^{(11,2)1}\,$. However, recently it has been proven[@fm1; @fm2], using crossing symmetry and charge conjugation invariance, that only [*six*]{} of the above ten GPs are independent. With $$\omega_0 \equiv M_N - \sqrt{M_N^2+\bar{q}^2}=-\;\frac{\bar{q}^2}{2M_N}+ {\cal O}(1/M_N^3) \label{om0}$$ and $M_N$ being the nucleon mass, the four constraints implied by $C$ invariance and crossing can be written as $$\begin{aligned} 0 & = & \sqrt{\frac{3}{2}} P^{(01,01)0}(\bar{q}^2) + \sqrt{\frac{3}{8}} P^{(11,11)0}(\bar{q}^2) + \frac{3 \bar{q}^2}{2 \omega_0} \hat{P}^{(01,1)0}(\bar{q}^2) \, , \nonumber\\ 0 & = & P^{(11,11)1}(\bar{q}^2) + \sqrt{\frac{3}{2}} \omega_0 P^{(11,02)1}(\bar{q}^2) + \sqrt{\frac{5}{2}} \bar{q}^2 \hat{P}^{(11,2)1}(\bar{q}^2) \, , \nonumber\\ 0 & = & 2 \omega_0 P^{(01,01)1}(\bar{q}^2) + 2 \frac{\bar{q}^2}{\omega_0} P^{(11,11)1}(\bar{q}^2) - \sqrt{2} \bar{q}^2 P^{(01,12)1}(\bar{q}^2) + \sqrt{6} \bar{q}^2 \hat{P}^{(01,1)1}(\bar{q}^2) \, , \nonumber\\ 0 & = & 3 \frac{\bar{q}^2}{\omega_0} P^{(01,01)1}(\bar{q}^2) - \sqrt{3} P^{(11,00)1}(\bar{q}^2) - \sqrt{\frac{3}{2}} \bar{q}^2 P^{(11,02)1}(\bar{q}^2) \, .\label{cinv}\end{aligned}$$ In the scalar (i.e. spin–independent) sector the first of Eqs.(\[cinv\]) allows us to eliminate the mixed polarizability ${\hat{P}}^{(01,1)0}$ in favor of $P^{(01,01)0}$ and $P^{(11,11)0}$, which are simply generalizations of the familiar electric and magnetic polarizabilities in real Compton scattering $$\begin{aligned} \bar{\alpha}_E (\bar{q}^2) & = & - \frac{e^{2}}{4 \pi} \sqrt{\frac{3}{2}} P^{(01,01)0} (\bar{q}^2) \,, \nonumber\\ \bar{\beta}_M (\bar{q}^2) & = & - \frac{e^{2}}{4 \pi} \sqrt{\frac{3}{8}} P^{(11,11)0} (\bar{q}^2) \,.\end{aligned}$$ In the limit $\bar{q}\rightarrow 0$ they reduce to the real Compton polarizabilities $\bar{\alpha}_E,\,\bar{\beta}_M$ of Eq.(\[abexp\]). In the spin-dependent sector it is not a priori clear which three of the seven GPs $P^{(01,01)1}\,$, $P^{(11,11)1}\,$, $P^{(01,12)1}\,$, $P^{(11,02)1}\,$, $P^{(11,00)1}\,$, ${\hat{P}}^{(01,1)1}\,$, ${\hat{P}}^{(11,2)1}\,$. should be eliminated by use of Eq.(\[cinv\]). However, the chiral analysis performed here shows that to leading order only 4 of the 7 spin GPs can be calculated—$P^{(01,12)1},\,P^{(11,02)1},\,P^{(11,00)1},\,{\hat{P}}^{(01,1)1}$. Naturally we focus on these four spin GPs, as $P^{(01,01)1},\,P^{(11,11)1},\,{\hat{P}}^{(11,2)1}$ possess an extra suppression factor of $1/M_N$ (see section \[HBChPTspin\]) which pushes them outside the validitiy of our analysis. Still, one can reconstruct the whole set of spin GPs via Eq.(\[cinv\]) if one wishes to do so. Finally, we note that in the spin-sector one can also establish a (partial) connection between the GPs defined in the context of VCS by Guichon et al. [@Guichon] and the 4 real Compton spin polarizabilities $\gamma_i,\;i=1...4$ of Ragusa [@Ragusa] given in Eqs.(\[spinpola1\]) and (\[spinpolas\]): $$\begin{aligned} \label{spinconnection} \gamma_3&=&-\frac{e^2}{4\pi}\,\frac{3}{\sqrt{2}}\,P^{(01,12)1}(\bar{q}\rightarrow 0) \nonumber \\ \gamma_2+\gamma_4&=&-\frac{e^2}{4\pi}\,\frac{3\sqrt{3}}{2\sqrt{2}}\,P^{(11,02)1} (\bar{q}\rightarrow 0)\end{aligned}$$ These model-independent relations might provide an interesting possibility to determine some of the elusive (Ragusa) spin-polarizabilities by the upcoming experiments. The chiral framework ==================== Pion-Nucleon ChPT ----------------- We want to perform the VCS calculation to ${\cal O}(p^3)$ in Heavy Baryon Chiral Perturbation Theory (HBChPT) ([*e.g.*]{} see [@review]). We therefore need the lagrangians $$\begin{aligned} {\cal L}_{VCS}^{(3)}={\cal L}_{N}^{(3)}+{\cal L}_{\pi}^{(4)}.\end{aligned}$$ We begin our discussion in the nucleon sector. For VCS to ${\cal O}(p^3)$ we need the lagrangians $${\cal{L}}_{N}^{(3)} = {\cal{L}}_{\pi N}^{(1)} + {\cal{L}}_{\pi N}^{(2)} + {\cal{L}}_{\pi N}^{(3)} \, , \label{eq:chiL}$$ with $$\begin{aligned} {\cal{L}}_{\pi N}^{(1)} & = & \bar N_v(iv \cdot D + \dot{g}_A S \cdot u) N_v \, , \nonumber\\ {\cal{L}}_{\pi N}^{(2)} & = & \frac{1}{2M_0} \bar N_v \left\{(v\cdot D)^2-D^2- \frac{i}{2}\left[S^\mu,S^\nu\right]\left[\left( 1+\dot{\kappa}_v\right)f_{\mu\nu}^+ +2\left(1+\dot{\kappa}_s\right)v_{\mu\nu}^{(s)} \right]+\dots\right\}N_v\; , \nonumber\\ {\cal{L}}_{\pi N}^{(3)} & = & \frac{-1}{8 M_0^2} \bar N_v \left\{\left(1+2\dot{\kappa}_v\right) \left[S_\mu,S_\nu\right]f_+^{\mu\sigma}v_\sigma D^\nu+2\left(\dot{\kappa}_s-\dot{\kappa}_v\right) \left[S_\mu,S_\nu\right]v_{(s)}^{\mu\sigma}v_\sigma D^\nu + \mathrm{h.c.}+\dots\right\} N_v\; , \label{piN}\end{aligned}$$ where we have only kept those terms[^7] which contribute to our VCS calculation. Furthermore, all terms which vanish in the “Coulomb gauge” $v \cdot A=0$, with $v_\mu$ being the velocity vector $(v^2=1)$ of the nucleon and $A_\mu$ denoting a photon field, have been omitted. The velocity-dependent nucleon field $N_v$ is projected from the relativistic nucleon Dirac field $\Psi_N$ via $$N_v = {\rm{exp}} \left[ i M_0 v \cdot x \right] P_v^+ \Psi_N \, ,$$ where the velocity projection operator is given by $$P_v^+ = \frac{1}{2} \left( 1 + \not\!{v} \right) \, .$$ $S_\mu$ denotes the usual Pauli-Lubanski vector ([*e.g.*]{} [@review]) and $D_\mu$ corresponds to the covariant derivative of the nucleon $$D_{\mu}\,N_v =\left[\partial_{\mu} + \Gamma_{\mu} - i v_{\mu}^{(s)}\right]N_v.$$ One also encounters the following chiral tensors in the VCS calculation: $$\begin{aligned} U & = & u^2={\mathrm{exp}} \left(i \vec \tau \cdot \vec \pi /F_\pi \right)\nonumber \\ \Gamma_{\mu} & = & \frac{1}{2} \left\{ u^{\dagger} \left( \partial_{\mu} - i \,e\,\frac{\tau^3}{2}\,A_\mu\right) u + u \left( \partial_{\mu} - i\,e\,\frac{\tau^3}{2}\,A_\mu \right) u^{\dagger} \right\} \; , \nonumber\\ u_{\mu} & = & i \left\{ u^{\dagger} \left( \partial_{\mu} - i\,e\,\frac{\tau^3}{2}\,A_\mu \right) u - u \left( \partial_{\mu} - i\,e\,\frac{\tau^3}{2}\,A_\mu \right) u^{\dagger} \right\} \; . \label{eq:aa}\end{aligned}$$ In Eq.(\[eq:aa\]) $\vec \tau$ are the conventional Pauli isospin matrices, while $\vec \pi$ represents the interpolating pion field. Furthermore, $v_\mu^{(s)}=e\,\frac{1}{2}\,A_\mu$ denotes an isoscalar photon field and the corresponding field strength tensors in Eq.(\[piN\]) are defined as $$\begin{aligned} v_{\mu \nu}^{(s)} & = & \partial_{\mu} v_{\nu}^{(s)} - \partial_{\nu} v_{\mu}^{(s)} \, , \nonumber\\ f_+^{\mu \nu} & = & u\,e\frac{\tau^3}{2}\,\left(\partial_\mu A_\nu-\partial_\nu A_\mu\right) u^{\dagger} + u^{\dagger}\,e\frac{\tau^3}{2}\,\left(\partial_\mu A_\nu-\partial_\nu A_\mu\right) u \, .\end{aligned}$$ [From]{} the pion sector we require information up to ${\cal O}(p^4)$ for a ${\cal O}(p^3)$ VCS calculation. Utilizing “standard ChPT” [@GL] ([*i.e.*]{} the assumption of a “large” quark condensate parameter $B$) one finds $$\begin{aligned} {\cal L}_{\pi}^{(4)}={\cal L}_{\pi\pi}^{(2)}+{\cal L}_{\pi\pi}^{(4)}\end{aligned}$$ with $$\begin{aligned} \label{Lpi} {\cal{L}}_{\pi \pi}^{(2)} &=& \frac{F_0^2}{4} {\mathrm{tr}} \left[ \left(\nabla_{\mu} U\right)^{\dagger} \nabla^{\mu} U +\chi^\dagger U+\chi U^\dagger \right] \; , \nonumber \\ {\cal{L}}_{\pi \pi}^{(4)} &=&{e^2\over 32\pi^2 F_0}\epsilon^{\mu\nu\alpha\beta} F_{\mu\nu}F_{\alpha\beta}\pi^0+\ldots \; ,\end{aligned}$$ where again we have omitted all terms not required for the VCS calculation. Note that the only piece shown from the chiral ${\cal O}(p^4)$ meson lagrangian is the so called “anomalous” or “Wess-Zumino” term [@WZ], which one needs for the ${\cal O}(p^3)$ pion-pole diagram of VCS shown in Fig.\[figgborn\](f). In the lagrangians of Eq.(\[Lpi\]) one also encounters the chiral tensors $$\begin{aligned} \nabla_{\mu} U &=& \partial_{\mu} U -i\,\frac{e}{2}\,A_{\mu} \left[ \tau_3, U \right] \; ,\nonumber \\ \chi&=&2\,B\,{\it M}\; ,\end{aligned}$$ where ${\it M}$ denotes the SU(2) quark mass matrix in the isospin limit $m_u=m_d$. Finally, we emphasize that we do not require any additional diagrams compared to the ${\cal O}(p^3)$ calculation for real Compton scattering [@BKKM1]. The complete set of non-zero diagrams we have to calculate is given in Fig.\[figgborn\] ((a) $s$-channel, (b) $u$-channel, (c) contact diagram and (f) $t$-channel pole term) and Fig.\[figgnpi\] ($N\pi$-loop diagrams). In the following we will treat the tree and loop parts of the amplitudes separately, $$A_i = A_i^{tree} + A_i^{loop} \, ,$$ since the generalized polarizabilities are contained only in the latter. $\Delta$(1232) and the Small Scale Expansion {#deltatheory} -------------------------------------------- In standard SU(2) HBChPT, nucleon resonances like the $\Delta$(1232) are considered to be much heavier than the nucleon and therefore only contribute via local counterterms. This approach is particularly well-suited for near-threshold processes ([*e.g.*]{} the multipole $E_{0+}$ in threshold pion photoproduction) where the resonance contributions are small and their contribution to counterterms can be estimated by a simple Born diagram analysis. However, if one wants to move away from threshold, nucleon resonances, in particular the lowest lying SU(2) resonance $\Delta$(1232), contribute as dynamical degrees of freedom and the theoretical treatment in terms of local counterterms generates a slowly converging perturbative series. In this kinematical regime it is therefore advantageous to formulate an effective field theory which keeps the resonance as an explicit degree of freedom. In addition to this dynamical consideration there is also another practical concern regarding the inclusion of resonance effects via counterterms. Even if simple Born exchange might be the dominant contribution of a particular resonance, the local counterterm in the chiral Lagrangian that subsumes this effect might be of higher order in the calculation, so that the leading and even the subleading result can misrepresent the perturbative series. A well-known example of this type are the so-called spin-polarizabilities of the nucleon, wherein one encounters very large contributions due to $\Delta$(1232) Born graphs that only start contributing via counterterms at ${\cal O}(p^5)$ in the chiral calculation ([*e.g.*]{} [@HHKK]). Situations of this type require a “resummation” of the standard chiral expansion in order to push resonance effects into lower orders to restore meaningful perturbative expansions for quantities of interest in low energy baryon physics. In order to address these two different but related issues in the field of resonance physics in baryon CHPT, the so called “small scale expansion” of SU(2) baryon ChPT has recently been formulated [@HBDel; @HHK97]. In this chiral effective theory one treats the nucleon and the first nucleon resonance—$\Delta$(1232)—as explicit degrees of freedom, and, to address the second problem, the chiral power counting is modified to bring $\Delta$(1232) related effects into lower orders of the calculation. In the “small scale expansion” one organizes the Lagrangian and the calculation in powers of the scale “$\epsilon$”, which, in addition to the chiral expansion parameters of small momenta $q$ and the pion mass $m_\pi$, also includes the $\Delta(1232)-N(940)$ mass splitting $\Delta=M_\Delta-M_N$. Of course, this modification of the chiral counting implies that one has to repeat the whole procedure of construction of the Lagrangian and the determination of counterterms and coupling constants, even for processes which only involve nucleons in the initial and final states. For first results regarding the modified renormalization of nucleon parameters we refer to [@bfhm]. For our calculation below, which (as far as the GPs are concerned) is done only to leading order—${\cal O}(\epsilon^3)$—in the small scale expansion of the (generalized) polarizabilities, we shall require only the propagator involving the $\Delta(1232)$ as well as the couplings $NN\gamma$, $NN\gamma\gamma$, $N\Delta\pi$ and $N\Delta\gamma$. Details of the “small scale expansion” formalism are given in ref. [@HHK97]. Here we only list the minimal structures necessary for the present calculation. The systematic 1/M-expansion of the coupled $N\Delta$-system starts with the most general relativistic chiral invariant lagrangian involving spin 1/2 ($\psi_N$) and spin 3/2 ($\psi^i_\mu$) baryon fields[^8]. The “light” spin 3/2 field $T^i_\mu$ in the effective low-energy theory is projected from its relativistic Rarita-Schwinger counterpart $\psi^i_\mu$ via $$T_{\mu}^i (x) \equiv P_{v}^{+} \; P^{3/2}_{(33)\mu\nu} \; \psi^{\nu}_i (x) \; \mbox{exp}(i M_0 v \cdot x), \label{eq:T}$$ where we have introduced a spin 3/2 projection operator for fields with [*fixed velocity*]{} $v_\mu$ $$P^{3/2}_{(33)\mu \nu} = g_{\mu \nu} - \frac{1}{3} \gamma_{\mu} \gamma_{ \nu} - \frac{1}{3} \left( \not\!{v} \gamma_{\mu} v_{\nu} + v_{\mu} \gamma_{\nu}\not\!{v}\right) . \label{eq:proj}$$ The remaining components, $$G_{\mu}^i (x) = \left( g_{\mu\nu}-P_{v}^{+} \; P^{3/2}_{(33)\mu\nu}\right) \psi^{\nu}_i (x) \; \mbox{exp}(i M_0 v \cdot x) , \label{eq:G}$$ can be shown to be “heavy” [@HHK97] and are integrated out. Resulting from this procedure one finds the (non-relativistic) chiral lagrangians of the “small scale expansion” (SSE): $${\cal L}^{SSE}={\cal L}_{N}^{SSE}+{\cal L}_{\Delta}^{SSE}+\left({\cal L}_{ N\Delta}^{SSE} + h.c. \right). \label{eq:4L}$$ To the order we are working here ${\cal L}_{N}^{SSE}$ agrees with the chiral lagrangian ${\cal L}_{N}^{(3)}$ (Eq.(\[eq:chiL\])) needed for VCS. From the chiral SSE lagrangians explicitly involving the $\Delta$ field we need the structures [@HHK97] $$\begin{aligned} {\cal L}^{(1)}_{\Delta}&=&-\bar{T}^\mu_i g_{\mu\nu}\left[iv\cdot D^{ij}-\Delta_0\; \delta^{ij}+\ldots\right] T^\nu_j\nonumber\\ {\cal L}^{(1)}_{N\Delta}&=&g_{\pi N\Delta}\bar{T}^\mu_i\;w_\mu^i\;N+{\rm h.c.}\nonumber\\ {\cal L}^{(2)}_{N\Delta}&=&\bar{T}^\mu_i\left[{ib_1\over M_0}\;S^\nu\;f_{+\mu\nu}^i+ \ldots \right] N +{\rm h.c.}, \label{eq:xxx}\end{aligned}$$ where $\Delta_0=M_\Delta-M_0$ can be identified with the [*physical*]{} delta-nucleon mass difference to the order we are working, [*i.e.*]{} $M_0\equiv M_N$. The corresponding chiral tensors needed for VCS read $$\begin{aligned} D_{\mu}^{ij}&=&\partial_{\mu}\delta^{ij}-i\frac{e}{2}\left(1+\tau_3\right)A_\mu \delta^{ij}+e\epsilon^{i3j}A_\mu + \ldots \nonumber\\ w_\mu^i &=&-\frac{1}{F_\pi}\partial_\mu\pi^i-\frac{e}{F_\pi}A_\mu\epsilon^{i3j} \pi^j+\ldots \nonumber\\ f_{+\mu\nu}^i&=&e\delta^{i3}\left(\partial_\mu A_\nu-\partial_\nu A_\mu\right)+\dots\end{aligned}$$ The coupling constants defined in Eq.(\[eq:xxx\]) are determined from fits to the strong and electromagnetic decay widths of the Delta resonance within the “small scale expansion”. To the order we are working one requires[^9] [@HHKK; @trh] $g_{\pi N\Delta}=1.05\pm 0.02$ and $b_1=3.85\pm 0.15$. The leading propagator for a $\Delta$(1232) field with small momentum $k_\mu$ is then given by $$S^{3/2}_{\mu\nu}={-iP^{3/2}_{\mu\nu}\over v\cdot k-\Delta+i\eta}\xi^{ij}_{I=3/2},$$ where $P^{3/2}_{\mu\nu}$ is the spin-${3\over 2}$ heavy baryon projector in d-dimensions [@HHK97] $$P^{3/2}_{\mu\nu}=g_{\mu\nu}-v_\mu v_\nu+{4\over d-1}S_\mu S_\nu ,$$ and $$\xi^{ij}_{I=3/2}=\delta^{ij}-{1\over 3}\tau^i\tau^j$$ is the corresponding isospin projector. The vertices relevant for our calculation can be read off directly from Eq.(\[eq:xxx\]). As in the nucleon case, the resulting diagrams can be separated into two classes—one-loop graphs and Born graphs. The systematics of the “small scale expansion” uniquely fixes the number and type of diagrams for VCS to be calculated to ${\cal O}(\epsilon^3)$. It turns out that to the order we are working there are two Born diagrams involving the $\Delta$(1232) (Fig.\[figgborn\](d,e)) and nine $\Delta\pi$-loop diagrams (Fig.\[figgdpi\]), which turn out to have exactly the same structure as their chiral $N\pi$ analogues (cf. Fig.\[figgnpi\]). However, before undertaking any such calculation, it is necessary to work out the formalism for VCS. Virtual Compton Scattering ========================== General Structure {#svcs} ----------------- We begin by specifying our notation for the virtual Compton process $$\gamma^*(\epsilon^\mu,q^\mu)+N(p_i^\mu)\rightarrow\gamma(\epsilon^{\prime *\mu},q^{ \prime\mu})+N(p_f^\mu).$$ Here the nucleon four-momenta in the initial and final states are denoted by $p_i^\mu=(E_i,\vec{p}_i)$ and $p_f^\mu=(E_f,\vec{p}_f)$ respectively. The virtual initial \[real final\] state photon is characterized by its four-momentum $q^\mu=(\omega,\vec{q}),\; q^2<0$ \[$q^{\prime\mu}=(\omega^\prime,\vec{q}^{\;\prime}),\;q^{\prime 2}=0$\] and polarization vector $\epsilon^\mu=(\epsilon_0,\vec{\epsilon})$ \[$\epsilon^{\prime\mu}=(\epsilon^{\prime}_0,\vec{\epsilon}^{\;\prime})$\]. Since our discussion refers to an electron scattering experiment, wherein the virtual photon is exchanged between the electron and hadron currents, the polarization vector of the incoming photon is given by $$\epsilon_\mu=e\;\bar{u}_{e'}(k_1)\;\gamma_\mu \;u_e(k_2)\; \frac{1}{q^2} \; ,$$ where $u_e(k_1),\bar{u}_{e'}(k_2)$ are electron Dirac spinors with four-momenta $k_1^\mu \;(k_2^{\mu})$ before (after) emission of the virtual photon. The unit charge $e$ is taken as $e=\sqrt{4\pi/137}>0$. In addition to the proper VCS process displayed in Fig.2a there are also Bethe-Heitler processes taking place (Fig.2b,c), $${\cal M}_{eN\rightarrow e'N\gamma}={\cal M}^{VCS}+{\cal M}^{Bethe-Heitler},$$ and such Bethe-Heitler contributions must be carefully evaluated before one can infer any information about the VCS matrix element from the electron scattering cross section.[^10] In the following, however, we will focus on the evaluation of the VCS matrix element ${\cal M}^{VCS}$ (Fig.2a). For details on ${\cal M}^{Bethe-Heitler}$ and the calculation of the cross section we refer to [@nstar] and references therein. [From]{} now on we will work in the center of mass system of the final state photon-nucleon subsystem, $$\begin{aligned} \vec{p}_f=-\vec{q}^{\;\prime},& & \vec{p}_i=-\vec{q}=-\bar{q}\;\hat{e}_z\; , \nonumber \\ \omega^\prime+\sqrt{M_{N}^2+\omega^{\prime 2}}&=&\omega +\sqrt{M_{N}^2+\bar{q}^2} \; , \label{eq:omega}\end{aligned}$$ where the $z$-axis is defined by the three-momentum vector $\vec{q}$ of the incoming virtual photon. Utilizing the Lorentz gauge[^11], $$\epsilon \cdot q=0,\qquad \epsilon_0=\frac{\bar{q}}{\omega}\; \epsilon_z \; ,$$ with $\vec{\epsilon}=\vec{\epsilon}_T+\epsilon_z \hat{e}_z$, one can express the VCS matrix element in terms of twelve[^12] independent kinematic forms $$\begin{aligned} {\cal M}^{VCS}&=&i\;e^2\left\{\vec{\epsilon}^{\;\prime *}\cdot\vec{\epsilon}_T\;A_1 +\vec{\epsilon}^{\;\prime *}\cdot\hat{q}\;\vec{\epsilon}_T\cdot\hat{q}^\prime\;A_2+i \vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime *}\times\vec{\epsilon}_T\right)A_3+i\;\vec{ \sigma}\cdot\left(\hat{q}^\prime\times\hat{q}\right)\vec{\epsilon}^{\;\prime *}\cdot \vec{\epsilon}_T\;A_4 \right. \nonumber \\ & &\phantom{i\;e^2 }+i\;\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime *}\times\hat{q} \right)\vec{\epsilon}_T\cdot\hat{q}^\prime\;A_5+i\;\vec{\sigma}\cdot\left(\vec{ \epsilon}^{\;\prime *}\times\hat{q}^\prime\right)\vec{\epsilon}_T\cdot\hat{q}^\prime\; A_6 \nonumber \\ & &\phantom{i\;e^2}-i\;\vec{\sigma}\cdot\left(\vec{\epsilon}_T\times\hat{q}^\prime \right)\vec{\epsilon}^{\;\prime *}\cdot\hat{q}\;A_7-i\;\vec{\sigma}\cdot\left(\vec{ \epsilon}_T\times\hat{q}\right)\vec{\epsilon}^{\;\prime *}\cdot\hat{q}\;A_8 \nonumber \\ & &\phantom{i\;e^2 }\left.+\frac{q^2}{\omega^2}\;\epsilon_z\left[\vec{\epsilon}^{\; \prime *}\cdot\hat{q}\;A_9+i\;\vec{\sigma}\cdot\left(\hat{q}^\prime\times\hat{q} \right)\vec{\epsilon}^{\;\prime *}\cdot\hat{q}\;A_{10}+i\;\vec{\sigma}\cdot\left( \vec{\epsilon}^{\;\prime *}\times\hat{q}\right)A_{11}+i\;\vec{\sigma}\cdot\left(\vec{ \epsilon}^{\;\prime *}\times\hat{q}^\prime\right)A_{12}\right]\right\}, \label{eq:vcs12}\end{aligned}$$ where $\sigma_i,\;i=x,y,z$ are Pauli spin matrices. Utilizing Eq.(\[eq:omega\]), each amplitude $A_i$, i=1,12 is then a function of three independent kinematic quantities—$\omega',\bar{q}$ and $\theta$. Separation of Born- and Structure-Part {#separation} -------------------------------------- The twelve VCS amplitudes $A_i(\omega^\prime,\theta,\bar{q})$ can be decomposed into a (nucleon) Born part $A^{Born}_i(\omega^\prime,\theta,\bar{q})$ and a structure dependent part $\bar{A}_i(\omega^\prime,\theta,\bar{q})$, $$A_i(\omega^\prime,\theta,\bar{q})=A^{Born}_i(\omega^\prime,\theta,\bar{q})+ \bar{A}_i(\omega^\prime,\theta,\bar{q}).$$ To third order in both the chiral and small scale expansions, the Born part contains the nucleon pole diagrams (Fig.\[figgborn\](a,b)), the Thomson seagull graph (Fig.\[figgborn\](c)) and the (anomalous) pion-pole graph (Fig.\[figgborn\](f)). In the case of a proton target one finds $$\begin{aligned} A_1^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&- \frac{1}{M_N}+{\cal O}(1/(M_{N}^3, \Lambda_{\chi}^2M_N)) \nonumber\\ A_2^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&\frac{\bar{q}}{M_{N}^2}+{\cal O}( 1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_3^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&\frac{\left(1+2\kappa_p\right) \omega^\prime-\left(1+\kappa_p\right)^2\cos\theta\;\bar{q}}{2M_{N}^2} -{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^\prime\left(\omega^{\prime 2}+ \bar{q}^2-2\omega^{\prime}\bar{q}\cos\theta\right)}{m_\pi^2+\omega^{ \prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos\theta} \nonumber\\ & &\phantom{\frac{\left(1+2\kappa_p\right) \omega^\prime-\left(1+\kappa_p\right)^2\cos\theta\;\bar{q}}{2M_{N}^2}} +{\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_4^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-\frac{\bar{q}\left(1+\kappa_p\right)^2} {2 M_{N}^2}+{\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_5^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&\frac{\bar{q}\left(1+\kappa_p\right)^2} {2 M_{N}^2}-{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^{\prime 2}\bar{q}} {m_\pi^2+\omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos\theta}+ {\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_6^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-\frac{\omega^\prime\left(1+\kappa_p \right)}{2 M_{N}^2}+{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^{\prime 3}} {m_\pi^2+\omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos\theta}+ {\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_7^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&\frac{\bar{q}\left(1+\kappa_p \right)^2}{2 M_{N}^2}-{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^{\prime 2} \bar{q}}{m_\pi^2+\omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos \theta}+{\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_8^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-\frac{1+\kappa_p}{2 M_{N}^2}\;\frac{ \bar{q}^2}{\omega^\prime}+{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^\prime \bar{q}^2}{m_\pi^2+\omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos \theta}+{\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \nonumber\\ A_9^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-\frac{1}{M_N}+\frac{2\;\omega^\prime \bar{q}\cos\theta+\bar{q}^2}{2 M_{N}^2\omega^\prime}+{\cal O}(1/(M_{N}^3, \Lambda_{\chi}^2M_N)) \nonumber\\ A_{10}^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-{g_A\over 8\pi^2F_\pi^2}\;\frac{ \omega^{\prime 2}\bar{q}}{m_\pi^2+\omega^{\prime 2}+\bar{q}^2-2 \omega^\prime\bar{q}\cos\theta}+{\cal O}(1/(M_{N}^3,\Lambda_{\chi}^2 M_N)) \nonumber\\ A_{11}^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&\frac{\left(1+2\kappa_p\right) \omega^\prime}{2 M_{N}^2}-{g_A\over 8\pi^2F_\pi^2}\;\frac{\omega^{ \prime 2}\left(\omega^\prime-\bar{q}\cos\theta\right)}{m_\pi^2+ \omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos\theta}+{\cal O} (1/(M_{N}^3,\Lambda_{\chi}^2M_N))\nonumber\\ A_{12}^{Born\;(3)}(\omega^\prime,\theta,\bar{q})&=&-\frac{\left(1+\kappa_p\right) \omega^\prime\cos\theta}{2 M_{N}^2}-{g_A\over 8\pi^2F_\pi^2}\;\frac{ \omega^{\prime 2}\left(\bar{q}-\omega^\prime\cos\theta\right)}{m_\pi^2 +\omega^{\prime 2}+\bar{q}^2-2\omega^\prime\bar{q}\cos\theta}+{\cal O} (1/(M_{N}^3,\Lambda_{\chi}^2M_N)) \,,\label{eq:bg}\end{aligned}$$ where $\Lambda_\chi=4\pi F_\pi$ denotes the scale of chiral symmetry breaking[@sca]. One can easily verify that the low energy forms of these structure functions are in agreement with the constraints implied by the Low theorem in the case of real Compton scattering[@low]—$\bar{q}=0$—and with the generalized low energy theorem in the case of VCS[@Guichon; @SK]. [From]{} the above expressions it can also be seen that the pion-pole contributions—Fig.\[figgborn\](f)—which scale linearly with $g_A$, affect only the spin-dependent structure amplitudes, as expected from the pion-nucleon coupling structure. All additional contributions are contained in the structure-dependent parts $\bar{A}_i(\omega^\prime,\theta,\bar{q})$ of the amplitudes, from which one can extract the (generalized) polarizabilities. Connection with the GPs ----------------------- In this section we present the formulae by which the GPs are related to the twelve structure-dependent amplitudes $\bar{A}_i(\omega^\prime,\theta,\bar{q}),\;i=1\dots 12$ [*to ${\cal O}(p^3)$ in HBChPT and to ${\cal O}(\epsilon^3)$ in SSE*]{}. First, we focus on the spin-independent GPs. To leading order in both the chiral and small scale expansions the spin-independent GPs $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$ can be found from the structure functions $\bar{A}_9(\omega^\prime,\theta,\bar{q}),\;\bar{A}_2(\omega^\prime,\theta,\bar{q})$ via [@HHKS1] $$\begin{aligned} \bar{\alpha}^{(3)}_E(\bar{q})&=&\frac{e^2}{8\pi} \; \frac{\partial^2}{\partial \omega^{\prime 2}}\bar{A}^{(3)}_9(\omega^\prime,\theta,\bar{q})\biggl|_{\omega^\prime =0} \; , \nonumber\\ \bar{\beta}^{(3)}_M(\bar{q})&=&-\frac{e^2}{4\pi} \;\frac{1}{\bar{q}}\;\frac{\partial}{ \partial\omega^\prime}\bar{A}^{(3)}_2(\omega^\prime,\theta,\bar{q})\biggl|_{\omega^\prime =0} \; . \label{defab}\end{aligned}$$ Note that the structure amplitudes in general have a dependence on the scattering angle $\theta$, whereas the GPs are only functions of $\bar{q}$. The independence of the GPs on $\theta$ therefore serves as a non-trivial check on the calculation. Likewise, the four independent spin-dependent GPs can be found from the relations [@Germar] $$\begin{aligned} \hat{P}_{(01,1)1}^{(3)}(\bar{q})&=&-\frac{\sqrt{2}}{3\sqrt{3}}\;\frac{1}{\bar{q}^2}\; \frac{\partial}{\partial\omega^\prime}\left[2\;\bar{A}_{3}^{(3)}(\omega^\prime,\theta, \bar{q})+\bar{A}_{8}^{(3)}(\omega^\prime,\theta,\bar{q})\right]_{\omega^\prime=0} \nonumber\\ P_{(01,12)1}^{(3)}(\bar{q})&=&-\frac{\sqrt{2}}{3}\;\frac{1}{\bar{q}^2}\;\frac{\partial}{ \partial\omega^\prime}\bar{A}_{8}^{(3)}(\omega^\prime,\theta,\bar{q})\biggl|_{\omega^\prime=0} \nonumber\\ P_{(11,02)1}^{(3)}(\bar{q})&=&-\frac{\sqrt{2}}{3\sqrt{3}}\;\frac{1}{\bar{q}}\;\frac{ \partial^2}{\partial\omega^{\prime2}}\bar{A}_{10}^{(3)}(\omega^\prime,\theta,\bar{q})\biggl|_{ \omega^\prime=0} \nonumber\\ P_{(11,00)1}^{(3)}(\bar{q})&=&\frac{\bar{q}}{\sqrt{3}}\;\frac{\partial^2}{\partial \omega^{\prime 2}}\left[\bar{A}_{12}^{(3)}(\omega^\prime,\theta,\bar{q})-\frac{2}{3}\; \bar{A}_{10}^{(3)}(\omega^\prime,\theta,\bar{q})\right]_{\omega^\prime=0}.\label{defspin}\end{aligned}$$ We note that these relations are only exact to third order in the chiral and in the small scale expansion. The analysis of ref.[@Germar] must be generalized before one can perform any fourth order calculations. Thus, to the order we are working, the remaining three spin-dependent GPs $P_{(01,01)1}^{(3)},P_{(11,11)1}^{(3)},\hat{P}_{(11,2)1}^{(3)}$ and the additional scalar GP $\hat{P}_{(01,1)0}^{(3)}$ can only be reconstructed[^13] with the help of the charge-conjugation constraint of Eqs.(\[cinv\]), yielding $$\begin{aligned} \hat{P}_{(01,1)0}^{(3)}(\bar{q})&=&\frac{\omega_0}{3\;\bar{q}^2}\left[\frac{\partial^2}{ \partial\omega^{\prime 2}}\bar{A}^{(3)}_9(\omega^\prime,\theta,\bar{q})-\frac{2}{ \bar{q}}\;\frac{\partial}{\partial\omega^\prime}\bar{A}^{(3)}_2(\omega^\prime,\theta, \bar{q})\right]_{\omega^\prime =0} \nonumber\\ P_{(01,01)1}^{(3)}(\bar{q})&=&\frac{\omega_0}{3\;\bar{q}}\;\frac{\partial^2}{\partial \omega^{\prime 2}}\left[\bar{A}_{12}^{(3)}(\omega^\prime,\theta,\bar{q})-\bar{A}_{10}^{ (3)}(\omega^\prime,\theta,\bar{q})\right]_{\omega^\prime =0} \nonumber\\ P_{(11,11)1}^{(3)}(\bar{q})&=&\frac{\omega_0}{3\;\bar{q}^2}\left\{2\;\frac{\partial}{ \partial\omega^\prime}\bar{A}_{3}^{(3)}(\omega^\prime,\theta,\bar{q})\biggl|_{\omega^\prime=0} -\frac{\omega_{0}^2}{\bar{q}}\;\frac{\partial^2}{\partial\omega^{\prime 2}}\left[ \bar{A}_{12}^{(3)}(\omega^\prime,\theta,\bar{q})-\bar{A}_{10}^{(3)}(\omega^\prime, \theta,\bar{q})\right]_{\omega^\prime=0}\right\} \nonumber\\ \hat{P}_{(11,2)1}^{(3)}(\bar{q})&=&\frac{\sqrt{2}\;\omega_0}{3\sqrt{5}\;\bar{q}^3} \left\{\left[\frac{\omega_{0}^2}{\bar{q}^2}\;\frac{\partial^2}{\partial\omega^{ \prime 2}}\bar{A}_{12}^{(3)}(\omega^\prime,\theta,\bar{q})+\left(1-\frac{\omega_{0}^2}{ \bar{q}^2}\right)\frac{\partial^2}{\partial\omega^{\prime 2}}\bar{A}_{10}^{(3)}( \omega^\prime,\theta,\bar{q})\right]_{\omega^\prime=0}\right. \nonumber \\ & &\phantom{\frac{\sqrt{2}\;\omega_0}{3\sqrt{5}\;\bar{q}^3} }\left. -\frac{2}{\bar{q}}\frac{\partial}{\partial\omega^\prime}\bar{A}_{3}^{(3)}( \omega^\prime,\theta,\bar{q})\biggl|_{\omega^\prime=0}\right\} , \label{defcharge}\end{aligned}$$ with $\omega_0=M_N-\sqrt{M_{N}^2+\bar{q}^2}$. Note that the spin-dependent GPs are just functions of the three-momentum transfer $\bar{q}$, whereas their generating structure amplitudes in Eqs.(\[defspin\]-\[defcharge\]) also depend on the scattering angle $\theta$—leading again to a non-trivial check on the calculation as in the case of the spin-independent GPs. With these definitions of the GPs we now turn to the results of the chiral and small scale expansions. Results ======= In this section we present the results for the generalized polarizabilities calculated in [*two different chiral effective theories*]{}—${\cal O}(p^3)$ HBChPT and ${\cal O}(\epsilon^3)$ SSE. ${\cal O}(p^3)$ Heavy Baryon ChPT --------------------------------- ### Structure Amplitudes {#chiamp} The only diagrams left at ${\cal O}(p^3)$ for the structure dependent part are the nine $N\pi$-continuum diagrams (Fig.\[figgnpi\]), which correspond to the pion-cloud of the nucleon in the formalism of baryon chiral perturbation theory. All other diagrams have already been accounted for in the Born part of section \[separation\]. We can now calculate the ${\cal O}(p^3)$ contributions to the 12 VCS structure amplitudes defined in Eq.(\[eq:vcs12\]), with all our results given in the CMS of the the final state photon-nucleon subsystem. From appendix \[appnpi\] one can read off the spin-independent structure amplitudes to ${\cal O}(p^3)$, yielding $$\begin{aligned} \bar{A}^{(3)}_1(\omega^\prime,\theta,\bar{q})&=& -\frac{g_{A}^2}{16\pi F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{\frac{m_{\pi}^2-4 m_{f}^2}{\sqrt{m_{f}^2}}+2\sqrt{m_{\pi}^2-\omega^{\prime 2}}-2\sqrt{m_{\pi}^2-\omega^{ \prime 2}x^2}-2\sqrt{\tilde{m}^2-\omega^{\prime 2}x^2} \right. \nonumber \\ & & \phantom{-\frac{g_{A}^2}{16\pi F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy }\left.+2\left(1-y\right)\frac{K^2 T+\left(6\hat{m}^2-m_{\pi}^2-6 T^2\right)\omega^\prime}{\omega^\prime\sqrt{\hat{m}^2-T^2}}\right\} \nonumber\\ \bar{A}^{(3)}_2(\omega^\prime,\theta,\bar{q})&=& +\frac{g_{A}^2}{8\pi F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy \frac{\bar{q} \omega^\prime\left(1-y\right) }{\sqrt{\hat{m}^2-T^2}}\left\{-1+x-8x y+7\left(y-y^2+x y^2\right) \right. \nonumber \\ & & \phantom{-\frac{g_{A}^2}{8\pi F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy \frac{\bar{q} \omega^\prime\left(1-y\right) }{\sqrt{\hat{m}^2-T^2}} } \left. +\left(1-x\right)y\left(1-y\right)\frac{\left(m_{\pi}^2-\hat{m}^2+T^2\right) \omega^\prime-K^2 T}{\omega^\prime\left(\hat{m}^2-T^2\right)}\right\} \nonumber\\ \bar{A}^{(3)}_9(\omega^\prime,\theta,\bar{q})&=& \bar{A}^{(3)}_1(\omega^\prime,\theta,\bar{q})+\cos\theta\;\bar{A}^{(3)}_2( \omega^\prime,\theta,\bar{q})+ \frac{g_{A}^2}{16\pi F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{ 1}dy\;\bar{q}^2\left[\frac{x\left(1-2x\right)}{\sqrt{\tilde{m}^2-\omega^{\prime 2}}} \right. \nonumber \\ & &\left. -\left(1-2y\right)y\left(1-y\right)\frac{\left(m_{\pi}^2-\hat{m}^2+T^2\right) \omega^\prime-K^2 T}{\omega^\prime\left(\hat{m}^2-T^2\right)^{3/2}}+ \frac{\left(1-y\right)\left(1-9y+14y^2\right)}{\sqrt{\hat{m}^2-T^2}}\right] , \label{eq:a1a2}\end{aligned}$$ with the “energy” and “mass” variables $$\begin{aligned} \label{definitions} T&=&\omega^\prime x\left(1-y\right) \nonumber\\ K^2&=&\omega^{\prime 2}-\omega^\prime\bar{q}\cos\theta \nonumber\\ \tilde{m}^2&=&m_{\pi}^2-q^2 x\left(1-x\right) \nonumber\\ \hat{m}^2&=&m_{\pi}^2-q^2 y\left(1-y\right)+2 q\cdot q^\prime\left(1-x\right)y \left(1-y\right) \nonumber\\ m_{f}^2&=&m_{\pi}^2-\left(q-q^\prime\right)^2 x\left(1-x\right) .\end{aligned}$$ The spin-dependent structure amplitudes to ${\cal O}(p^3)$ in the chiral expansion can also be found from the expressions in appendix \[appnpi\]— $$\begin{aligned} \bar{A}^{(3)}_3(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{-\sqrt{m_{\pi}^2-\omega^{\prime 2}}\arcsin\left[\frac{\omega^\prime}{m_\pi}\right]+\sqrt{m_{\pi}^2-\omega^{\prime 2}x^2}\arcsin\left[\frac{\omega^\prime x}{m_\pi}\right] \right. \nonumber \\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy } +\sqrt{\tilde{m}^2-\omega^{\prime 2} x^2}\arcsin\left[\frac{\omega^\prime x}{\tilde{m}}\right] +\omega^\prime x\log\left[\frac{\tilde{m}}{m_{\pi}}\right] \nonumber\\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy } \left. +\sin^2\theta\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \nonumber\\ \bar{A}^{(3)}_4(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left(1-y\right)\frac{\bar{q}\; T}{\sqrt{\hat{m}^2-T^2}}\arcsin\left[\frac{T}{\hat{m}}\right] \nonumber\\ \bar{A}^{(3)}_5(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{\left(x-1\right)\left(1-y\right)^2\frac{ \bar{q}\;\omega^\prime\arcsin\left[\frac{T}{\hat{m}}\right]}{\sqrt{\hat{m}^2-T^2}} \right. \nonumber \\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy}\left. +\cos\theta\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \nonumber\\ \bar{A}^{(3)}_6(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{\left(1-x\right)\left(1-y\right)^2\frac{ \omega^{\prime 2}\arcsin\left[\frac{T}{\hat{m}}\right]}{\sqrt{\hat{m}^2-T^2}} \right. \nonumber \\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy}\left. -\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \nonumber\\ \bar{A}^{(3)}_7(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{y\left(y-1\right)\frac{\omega^\prime \bar{q}\arcsin\left[\frac{T}{\hat{m}}\right]}{\sqrt{\hat{m}^2-T^2}} \right. \nonumber \\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy}\left. +\cos\theta\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \nonumber\\ \bar{A}^{(3)}_8(\omega^\prime,\theta,\bar{q})&=& \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left\{y\left(1-y\right)\frac{\bar{q}^{2} \arcsin\left[\frac{T}{\hat{m}}\right]}{\sqrt{\hat{m}^2-T^2}} \right. \nonumber \\ & &\phantom{ \frac{g_{A}^2}{4\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy}\left. -\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \nonumber\\ \bar{A}^{(3)}_{10}(\omega^\prime,\theta,\bar{q})&=&\bar{A}^{(3)}_{4}(\omega^\prime, \theta,\bar{q})+\bar{A}^{(3)}_{7}(\omega^\prime,\theta,\bar{q})\nonumber \\ & &\phantom{\bar{A}^{(3)}_{4}(\omega^\prime,\theta,\bar{q})} +\frac{g_{A}^2}{8\pi^2 F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1}dy\left(1-y\right)^2 x\left(2y-1\right)y\;\frac{ \bar{q}^3\omega^{\prime}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{ \hat{m}}\right]\right)}{\hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \nonumber\\ \bar{A}^{(3)}_{11}(\omega^\prime,\theta,\bar{q})&=& A_3(\omega',\theta,\bar{q}) +A_5(\omega',\theta,\bar{q})\nonumber\\ &+&{g_A^2\over 8\pi^2F_\pi^2}\int_0^1 dx \int_0^1 dy\left\{ x(1-2x)\bar{q}^2{\arcsin[{\omega'x\over \tilde{m}}]\over \sqrt{\tilde{m}^2-{\omega'}^2x^2}}+(1-y)(1-2y)\bar{q}^2 {\arcsin[{T\over \hat{m}}]\over \sqrt{\hat{m}^2-T^2}}\right.\nonumber\\ & &\phantom{sin}\left.-x(1-x)y(1-y)^3\bar{q}^3\omega'{T\sqrt{\hat{m}^2-T^2} +\hat{m}^2\arcsin[{T\over \hat{m}}]\over \hat{m}^2(\hat{m}^2-T^2)^{3\over 2}}\right\} \nonumber\\ \bar{A}^{(3)}_{12}(\omega^\prime,\theta,\bar{q})&=&\cos\theta\;\bar{A}^{(3)}_{6}( \omega^\prime,\theta,\bar{q})+\frac{g_{A}^2}{8\pi^2F_{\pi}^2}\int_{0}^{1}dx\int_{0}^{1} dy\left\{\left(1-y\right)\left(2y-1\right)\frac{\omega^\prime\bar{q}\arcsin\left[ \frac{T}{\hat{m}}\right]}{\sqrt{\hat{m}^2-T^2}}\right. \nonumber \\ & &\phantom{\cos\theta\;\bar{A}^{(3)}_{4}(\omega^\prime,\theta,\bar{q}) }\left. +2\cos\theta\left(1-x\right)x\left(1-y\right)^3y\;\frac{\bar{q}^2\omega^{\prime 2}\left(T\sqrt{\hat{m}^2-T^2}+\hat{m}^2\arcsin\left[\frac{T}{\hat{m}}\right]\right)}{ \hat{m}^2\left(\hat{m}^2-T^2\right)^{3/2}} \right\} \label{eq:a3a12}\end{aligned}$$ Eqs.(\[eq:a1a2\],\[eq:a3a12\]) constitute the [*full*]{} ${\cal O}(p^3)$ HBChPT results for the structure part in virtual Compton scattering off the nucleon. As such, they are independent of the particular formalism of Guichon and could also be used to extract alternative descriptions of generalized polarizabilities, [*e.g.*]{} see the recent paper by Unkmeir et al[@Christine]. ### Spin-independent Polarizabilities {#HBChPTscalar} [From]{} the ${\cal O}(p^3)$ HBChPT results for the 12 structure amplitudes given in the previous section one can now extract the GPs as defined by Guichon, following the general formulae given in Eqs.(\[defab\]-\[defcharge\]). In this subsection we first focus on the spin-independent GPs $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$. The leading $\bar{q}$-dependent modification of $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$ has already been analyzed in ref. [@HHKS1] and one finds $$\begin{aligned} \bar{\alpha}^{(3)}_E(\bar{q}) &=& \frac{5 e^2 g_{A}^2}{384\pi^2 F_{\pi}^2 m_\pi}\left[1-\frac{7}{50}\frac{\bar{q}^2}{m_{\pi}^2}+\frac{81}{2800}\frac{ \bar{q}^4}{m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] , \nonumber\\ \bar{\beta}^{(3)}_M(\bar{q}) &=& \frac{e^2 g_{A}^2}{768\pi^2 F_{\pi}^2 m_\pi}\left[1+\frac{1}{5}\frac{\bar{q}^2}{m_{\pi}^2}-\frac{39}{560}\frac{\bar{q}^4}{ m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] . \label{abseries}\end{aligned}$$ First, we note that in the limit $\bar{q}\rightarrow 0$ one recovers the well-known real Compton results at $q^2=0$ [@BKKM1] $$\begin{aligned} \bar{\alpha}^{(3)}_E &=& \bar{\alpha}^{(3)}_E(\bar{q}=0)= \frac{5 e^2 g_{A}^2}{384\pi^2 F_{\pi}^2m_\pi}=12.5\times 10^{-4}\;\mbox{fm}^3 \nonumber \\ \bar{\beta}^{(3)}_M &=& \bar{\beta}^{(3)}_M(\bar{q}=0)= \frac{e^2 g_{A}^2}{768\pi^2 F_{\pi}^2m_\pi}=1.25\times 10^{-4}\;\mbox{fm}^3 \; ,\end{aligned}$$ which work extremely well when compared with the existing experimental information given in Eq.(\[abexp\]). As already pointed out in ref. [@HHKS1], the slope of $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$ with respect to $\bar{q}$ shows the [*opposite sign*]{} for the two spin-independent polarizabilities. These respective slopes are uniquely determined by the chiral structure of the nucleon, i.e. the “pion-cloud”, as given by the $\pi N$-loop diagrams of Fig.(\[figgnpi\]). At ${\cal O}(p^3)$ ChPT therefore leads to the remarkable prediction that the (generalized) magnetic polarizability $\bar{\beta}_M(\bar{q})$ [*rises*]{} with increasing three-momentum transfer in a small window near $\bar{q}=0$. The subleading, i.e. ${\cal O}(p^4)$, correction to this result is not known at this point, but in section \[sres\] we discuss the leading modification of the slopes due to the $\Delta$(1232) resonance. Starting from the expression for the individual Feynman diagrams given in appendix \[appnpi\], the $\pi N$-loop contributions to $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$ can be shown to possess analytic expressions for their $\bar{q}$-dependence. To ${\cal O}(p^3)$ we find the remarkably simple closed form expressions $$\begin{aligned} \bar{\alpha}^{(3)}_E(\bar{q})&=& \frac{e^2 g_{A}^2 m_\pi}{64\pi^2 F_{\pi}^2}\;\frac{4+2\frac{\bar{q}^2}{m_{\pi}^2}-\left(8-2\frac{\bar{q}^2}{m_{\pi}^2} -\frac{\bar{q}^4}{m_{\pi}^4}\right)\frac{m_\pi}{\bar{q}}\arctan\frac{\bar{q}}{2 m_{\pi}}}{\bar{q}^2\left(4+\frac{\bar{q}^2}{m_{\pi}^2}\right)} \; , \nonumber\\ \bar{\beta}^{(3)}_M(\bar{q})&=& \frac{e^2 g_{A}^2 m_\pi}{128\pi^2 F_{\pi}^2}\;\frac{-\left(4+2\frac{\bar{q}^2}{m_{\pi}^2}\right)+\left(8+6\frac{ \bar{q}^2}{m_{\pi}^2}+\frac{\bar{q}^4}{m_{\pi}^4}\right)\frac{m_\pi}{\bar{q}}\arctan \frac{\bar{q}}{2 m_{\pi}}}{\bar{q}^2\left(4+\frac{\bar{q}^2}{m_{\pi}^2}\right)} \; . \label{eq:bq}\end{aligned}$$ These HBChPT predictions for $\bar{\alpha}_E(\bar{q}), \; \bar{\beta}_M(\bar{q})$ are also shown in Fig.\[figab\]. One observes a relatively sharp fall-off in the electric GP, whereas the magnetic GP shows the [*rising*]{} behaviour for low values of $\bar{q}$ as described above. This remarkable effect has its origin in the chiral structure of the pion cloud surrounding the nucleon and poses a formidable challenge to form-factor-supplemented Born-models of the GPs ([*e.g.*]{} see [@Vanderhaeghen]). From Eq.(\[eq:bq\]) the maximum of the magnetic GP can be determined to be $$\bar{\beta}_{M}^{max.}\left(\bar{q}=2.38 m_\pi\right)=1.29 \times \bar{\beta}_M(0) \; ,$$ indicating a 30% enhancement of this GP relative to its value at the real photon point. Using the $C$-invariance relations Eqs.(\[cinv\]), we can also read off the remaining spin-independent GP $$\begin{aligned} \hat{P}^{(01,1)0}(\bar{q})&=&-{g_A^2m_\pi\over 16\pi F_\pi^2}{\omega_0\over \bar{q}^4} {4+2{\bar{q}^2\over m_\pi^2}+(-8+10{\bar{q}^2\over m_\pi^2}+3{\bar{q}^4\over m_\pi^4}) {m_\pi\over \bar{q}}\arctan{\bar{q}\over 2m_\pi}\over (4+{\bar{q}^2\over m_\pi^2})}\nonumber\\ &=& -{11g_A^2\over 576\pi F_\pi^2 M_Nm_\pi}\left[1-{6\over 55} {\bar{q}^2\over m_\pi^2}+{123\over 560}{\bar{q}^4\over m_\pi^4}+{\cal O}(\bar{q}^6)\right] \label{extra}\end{aligned}$$ Once more we note that $\hat{P}^{(01,1)0}(\bar{q})$ is not an independent GP, but can be found as a linear combination of $\bar{\alpha}^{(3)}_E(\bar{q}),\;\bar{\beta}^{(3)}_M(\bar{q})$ via the charge-conjugation constraint Eq.(\[cinv\]). The extra suppression by $1/M_N$ compared to Eq.(\[abseries\]) arises from the expansion of the $\omega_0$ factor defined in Eq.(\[om0\]). Having discussed the scalar (spin-independent) structure of the nucleon, we now move on to the spin-dependent analysis. ### Spin-dependent Generalized Polarizabilities {#HBChPTspin} Following the identification of the GPs from the 12 structure amplitudes via Eqs.(\[defspin\]-\[defcharge\]) we can also analyze the behaviour of the spin-dependent GPs near $\bar{q}=0$. For the four independent spin GPs we find $$\begin{aligned} P_{(01,12)1}^{(3)}(\bar{q})&=&-\frac{\sqrt{2}\;g_{A}^2}{288\pi^2F_{\pi}^2 m_{\pi}^2}\left[1-\frac{\bar{q}^2}{5\;m_{\pi}^2}+\frac{ 3\;\bar{q}^4}{70\;m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] \nonumber\\ P_{(11,02)1}^{(3)}(\bar{q})&=&-\frac{\sqrt{2}\;g_{A}^2}{144\sqrt{3}\;\pi^2F_{\pi}^2 m_{\pi}^2}\left[1-\frac{\bar{q}^2}{5\;m_{\pi}^2}+\frac{3\; \bar{q}^4}{70\;m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] \nonumber\\ P_{(11,00)1}^{(3)}(\bar{q})&=&-\frac{5\;g_{A}^2}{144\sqrt{3}\pi^2F_{\pi}^2}\left[0+ \frac{\bar{q}^2}{m_{\pi}^2}-\frac{7\;\bar{q}^4}{50\; m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] \nonumber\\ \hat{P}_{(01,1)1}^{(3)}(\bar{q})&=&-\frac{g_{A}^2}{48\sqrt{6}\;\pi^2F_{\pi}^2m_{\pi}^2} \left[1-\frac{2\;\bar{q}^2}{15\;m_{\pi}^2}+\frac{ \bar{q}^4}{42\;m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] , \label{spintaylor}\end{aligned}$$ whereas the remaining three spin-dependent GPscan be determined via Eq.(\[defcharge\]) as a consequence of the $C$-invariance relations Eq.(\[cinv\]): $$\begin{aligned} P_{(01,01)1}^{\;l.o.\;\chi}(\bar{q})&=&\frac{g_{A}^2}{144\pi^2F_{\pi}^2}\;\frac{1}{M_N} \left[0 +\frac{\bar{q}^2}{m_{\pi}^2}-\left(\frac{3}{20}+\frac{ \mu^2}{4}\right)\frac{\bar{q}^4}{m_{\pi}^4}+{\cal O}( \bar{q}^6)\right] \nonumber\\ P_{(11,11)1}^{\;l.o.\;\chi}(\bar{q})&=&-\frac{g_{A}^2}{288\pi^2F_{\pi}^2}\;\frac{1}{M_N} \left[0 +\frac{\bar{q}^2}{m_{\pi}^2}-\left(\frac{1}{10}-\frac{ \mu^2}{4}\right)\frac{\bar{q}^4}{m_{\pi}^4}+{\cal O}( \bar{q}^6)\right] \nonumber\\ \hat{P}_{(11,2)1}^{\;l.o.\;\chi}(\bar{q})&=&\frac{\sqrt{2}\;g_{A}^2}{2880\sqrt{5}\;\pi^2 F_{\pi}^2m_{\pi}^2}\;\frac{1}{M_N}\left[0+\left(1+5\; \mu^2\right)\frac{\bar{q}^2}{m_{\pi}^2}-\left(\frac{ 2}{7}+\mu^2+\frac{15\;\mu^4}{4}\right)\frac{ \bar{q}^4}{m_{\pi}^4}+{\cal O}(\bar{q}^6)\right] ,\end{aligned}$$ with $\mu=m_\pi /M_N$. As in the case of $\hat{P}^{(01,1)0}(\bar{q})$ of Eq.(\[extra\]), one can clearly see that these three GPs are formally suppressed by an additional factor of $1/M_N$ relative to the four independent spin GPs of Eq.(\[spintaylor\]) and therefore ordinarily would not be accessible in a ${\cal O}(p^3)$ calculation. It is only the charge-conjugation constraint that allows us to extract them from the $\bar{A}_i^{(3)}$ VCS amplitudes. It is also interesting to note that four of the generalized spin-polarizabilities [*vanish*]{} in the real Compton limit—$\bar{q}\rightarrow 0$. In the case of $P^{(3)}_{(11,00)1}(\bar{q}), P^{l.o.\chi}_{(01,01)1}(\bar{q}), P^{l.o.\chi}_{(11,11)1}(\bar{q})$ this follows from charge conjugation invariance and crossing symmetry, as pointed out by Drechsel et al.[@fm2]. On the other hand, for $\hat{P}^{l.o.\chi}_{(11,2)1}(\bar{q})$ the zero appears to be a numerical accident which is only true at this order, since the linear sigma model calculation of ref. [@Metz] violates this condition. Nevertheless the zero in the first three cases is a powerful confirmation of the internal consistency of the ChPT approach to generalized polarizabilities. As in the case of the spin-independent sector it is possible to give analytic expressions for the 7 spin-dependent GPs. Defining the auxiliary function $$\label{auxfunction} g(x)={{\rm sinh}^{-1}(x)\over x\sqrt{1+x^2}}\; ,$$ the four independent generalized spin-polarizabilities to third order in the chiral expansion read $$\begin{aligned} P_{(01,12)1}^{(3)}(\bar{q})&=& -\frac{g_{A}^2}{24\sqrt{2}\;\pi^2 F_{\pi}^2\bar{q}^2}\left[1-g({\bar{q}\over 2m_\pi}) \right]\nonumber\\ P_{(11,02)1}^{(3)}(\bar{q})&=&-\frac{g_{A}^2}{12\sqrt{6}\;\pi^2F_{\pi}^2\bar{q}^2} \left[1-g({\bar{q}\over 2m_\pi})\right]\nonumber\\ P_{(11,00)1}^{(3)}(\bar{q})&=&\frac{g_{A}^2}{12\sqrt{3}\;\pi^2F_{\pi}^2} \left[2-\left(2+{3\bar{q}^2\over 4m_\pi^2}\right) g({\bar{q}\over 2m_\pi})\right] \nonumber\\ \hat{P}_{(01,1)1}^{(3)}(\bar{q})&=&\frac{g_{A}^2}{24\sqrt{6}\;\pi^2F_{\pi}^2\bar{q}^2} \left[3-\left(3+{\bar{q}^2\over m_\pi^2}\right) g({\bar{q}\over 2m_\pi})\right] .\label{4spin}\end{aligned}$$ The ${\cal O}(p^3)$ HBChPT results for these four spin-dependent GPs are shown in Fig.(\[figspin1\]). All are found to be negative in the low energy regime and three of them show a steep rise with $\bar{q}$ at low three-momentum transfer—except for $P_{(11,00)1}^{(3)}(\bar{q})$, which vanishes for $\bar{q}\rightarrow 0$ and is strongly falling off for small finite values of $\bar{q}$. The remaining three C-constrained GPs are found to be $$\begin{aligned} P_{(01,01)1}^{\;l.o.\;\chi}(\bar{q})&=&\frac{g_{A}^2\omega_0}{24\pi^2F_{\pi}^2\bar{q}^2} \left[1-(1+{\bar{q}^2\over 2 m_\pi^2})g({\bar{q}\over 2m_\pi})\right] ,\nonumber\\ P_{(11,11)1}^{\;l.o.\;\chi}(\bar{q})&=&-\frac{g_{A}^2\omega_0M_{N}^2}{24\pi^2 F_{\pi}^2\bar{q}^4}\left[\left(\frac{2\; \omega_0}{M_N}+\frac{3\;\bar{q}^2}{M_{N}^2}\right) -\left(\frac{3\;m_{\pi}^2+ \bar{q}^2}{M_{N}^2}\;\frac{\bar{q}^2}{m_{\pi}^2}+ \frac{\omega_0}{M_N}\left(2+\frac{ \bar{q}^2}{m_{\pi}^2}\right)\right) g({\bar{q}\over 2m\pi})\right], \nonumber\\ \hat{P}_{(11,2)1}^{\;l.o.\;\chi}(\bar{q})&=&\frac{g_{A}^2\omega_0M_{N}^2}{6\sqrt{10}\; \pi^2 F_{\pi}^2\bar{q}^6}\;\left[ \left(\frac{\omega_0}{M_N}+\frac{2\; \bar{q}^2}{M_{N}^2}\right)-{1\over 2} \left(\frac{\omega_0}{M_N}\left(2+\frac{ \bar{q}^2}{m_{\pi}^2}\right)+\left(4+\frac{ \bar{q}^2}{m_{\pi}^2}\right)\frac{\bar{q}^2}{ M_{N}^2}\right)g({\bar{q}\over 2m\pi})\right], \label{3spin}\end{aligned}$$ with $\omega_0$ defined in Eq.(\[om0\]). Their resulting $\bar{q}$-dependence is shown in Fig.(\[figspin2\]). $P_{(01,01)1}(\bar{q}),\; P_{(11,11)1}(\bar{q})$ vanish for $\bar{q}\rightarrow 0$ as required by C-invariance [@fm2], whereas the unconstrained spin-dependent GP $\hat{P}_{(11,2)1}(\bar{q})$ rises at low $\bar{q}$ and shows an unusual turnover point near $\bar{q}^2\sim 0.2$ GeV$^2$. Once more we note that these three particular GPs, strictly speaking, lie beyond a ${\cal O}(p^3)$ calculation and could only be deduced via the C-invariance constraints of Eq.(\[cinv\]). [From]{} an analysis of the corresponding spin-polarizabilities in real Compton scattering [@HHKK] one knows that in some cases there exist large corrections at $q^2=0$ to these chiral ${\cal O}(p^3)$ results of the spin-polarizabilities due to the Delta resonance. On the other hand, the spin-independent polarizabilities $\bar{\alpha}_E,\;\bar{\beta}_M$ are known to be well described within ${\cal O}(p^3)$ $\pi N$ HBChPT (see Eq.(\[abseries\]) in the limit $\bar{q}\rightarrow 0$ and [@BKKM1]). In the next section we will therefore analyze the leading effects of the $\Delta$(1232) on the GPs in a [*different chiral effective framework*]{}, which contains the $\Delta$(1232) as an explicit degree of freedom. ${\cal O}(\epsilon^3)$ Small Scale Expansion {#sres} -------------------------------------------- ### General comments regarding SSE and Compton scattering In HBChPT the effects of $\Delta$(1232) are incorporated via higher order contact interactions, i.e. the effects of this particular resonance are not directly tractable in the calculation. If one is interested in such kind of questions, one needs a chiral effective framework which includes $\Delta$(1232) as an explicit degree of freedom [*in a consistent power counting framework*]{}—one approach of this kind is SSE as laid out in section \[deltatheory\]. First, we would like to stress again that any SSE calculation to ${\cal O}(\epsilon^3)$ does not just equal the corresponding ${\cal O}(p^3)$ HBChPT calculation plus some additional diagrams with explicit Delta degrees of freedom. SSE constitutes a chiral effective theory separate from HBChPT—for example, even single nucleon coupling structures which look the same in the (bare) lagrangians of the two theories can undergo quite a different coupling constant renormalization or acquire different beta-functions, for details we refer the interested reader to ref.[@bfhm]. For the particular case of (real) Compton scattering we would like to remind the reader that HBChPT and SSE show quite a different convergence behaviour for the (real) Compton polarizabilities [@HHKK], which is expected to also hold true for the here discussed generalized polarizabilities of VCS. In principle there are two kinds of [*additional*]{} contributions to the ${\cal O}(p^3)$ HBChPT results presented in the previous section—$\Delta$(1232) pole graphs (Fig.\[figgborn\](d,e)) and $\Delta\pi$-continuum effects (Fig.\[figgdpi\]). The latter are straightforwardly obtained from the results given in appendix \[appdpi\], whereas the Delta pole effects to be discussed here are identical to their (real) Compton contributions discussed in [@HHKK]. ### Spin-independent results {#sir} First we discuss the ${\cal O}(\epsilon^3)$ SSE results for the spin-independent GPs $\bar{\alpha}_E^{(III)}(\bar{q}),\;\bar{\beta}_M^{(III)}(\bar{q})$ near $\bar{q}=0$ to facilitate the comparison between HBChPT and SSE. One finds $$\begin{aligned} \bar{\alpha}^{(III)}_E(\bar{q}) &=& \frac{5 e^2 g_{A}^2}{384\pi^2F_{\pi}^2m_\pi} +{e^2g_{\pi N\Delta}^2\over216\pi^3F_\pi^2}\left( {9\Delta\over \Delta^2-m_\pi^2}+{\Delta^2-10m_\pi^2\over (\Delta^2-m_\pi^2)^{3\over 2}}\ln R\right) \nonumber\\ &+&\frac{\bar{q}^2}{m_\pi^2}\left[-\frac{7e^2g_{A}^2}{3840\pi^2 F_{\pi}^2m_\pi}-{e^2g_{\pi N\Delta}^2\over1080\pi^3F_\pi^2} \left({2\Delta^3-17\Delta m_\pi^2\over (\Delta^2-m_\pi^2)^2}+{8\Delta^2m_\pi^2+7m_\pi^4\over (\Delta^2- m_\pi^2)^{5\over 2}}\ln R\right)\right]+{\cal O}(\bar{q}^4) \nonumber \\ &=&\left\{12.5+4.22+\frac{\bar{q}^2}{m_\pi^2}\left[-1.75-0.240\right] +\frac{\bar{q}^4}{m_\pi^4}\left[0.362+0.018\right]+\dots\right\} \times 10^{-4} \;\mbox{fm}^3 \nonumber \\ \bar{\beta}^{(III)}_M(\bar{q}) &=& \frac{e^2 g_{A}^2}{768\pi^2 F_{\pi}^2m_\pi} +\frac{e^22\,b_1^2}{9\pi M_N^2\Delta} +{e^2g_{\pi N\Delta}^2\over 216\pi^3F_\pi^2} {1\over \sqrt{\Delta^2-m_\pi^2}}\ln R \nonumber\\ &+&\frac{\bar{q}^2}{m_\pi^2}\left[\frac{e^2g_{A}^2}{3840\pi^2 F_{\pi}^2m_\pi}+{e^2g_{\pi N\Delta}^2\over 1080\pi^3F_\pi^2} \left({\Delta\over (\Delta^2-m_\pi^2)}-{m_\pi^2\over (\Delta^2 -m_\pi^2)^{3\over2}}\ln R\right)\right]+{\cal O}(\bar{q}^4) \nonumber \\ &= &\left\{1.25+7.20+0.725+\frac{\bar{q}^2}{m_\pi^2}\left[0.250+0.078\right] +\frac{\bar{q}^4}{m_\pi^4}\left[-0.087-0.020\right]+\dots\right\} \times 10^{-4}\;\mbox{fm}^3 \label{SSEtaylorab}\end{aligned}$$ with $$R={\Delta\over m_\pi}+\sqrt{{\Delta^2\over m_\pi^2}-1}\; .$$ The important point to note in Eq.(\[SSEtaylorab\]) is the fact that the $\bar{q}$-dependence is only modified in a very weak fashion by the inclusion of explicit delta degrees of freedom. In that respect SSE to ${\cal O}(\epsilon^3)$ and HBChPT to ${\cal O}(p^3)$ are quite compatible. However, the same problems known from real Compton scattering [@delta; @HHKK] appear in the limit $\bar{q}\rightarrow 0$, which in the Guichon definition of the GPs corresponds to the real photon point. In $\bar{\alpha}_E(0)\rightarrow \bar{\alpha}_E$ the $\Delta\pi$-continuum of Fig.(\[figgdpi\]) produces a shift of $4.2\times 10^{-4}\;\mbox{fm}^3$, which when added to the $12.5\times 10^{-4}\;\mbox{fm}^3$ from the $N\pi$-continuum of Fig.(\[figgnpi\]), leads to a much larger number then the current values of $\bar{\alpha}_E$ [@PPol]. In $\bar{\beta}_M(0)\rightarrow\bar{\beta}_M$ the effect is even more dramatic. Here it is the large magnetic contribution[^14] of $7.2\times 10^{-4}\;\mbox{fm}^3$ coming from the [*$\bar{q}$-independent*]{} delta pole graphs of Fig.(\[figgborn\]d,e) which spoil any agreement with the currently accepted number for $\bar{\beta}_M$ of the proton [@PPol]. On the other hand, the sum of the contributions from the $N\pi$- and from the $\Delta\pi$-continuum has the right magnitude of $\sim\,2\times 10^{-4}\;\mbox{fm}^3$ for the magnetic polarizability, constituting the “chiral version” of the unwanted presence of a large $\Delta$(1232)-induced paramagnetism, which is well-known in the literature [@MNZ]. A large source of diamagnetism due to the pion-cloud has been identified in refs.[@BKSM] in the case of (real) Compton scattering, but this mechanism, which leads to a sensible (central) value of $\bar{\beta}_M\sim 3.5\;10^{-4}\,\mbox{fm}^3$ for the proton, can only be implemented in a ${\cal O}(p^4)$ HBChPT (respectively ${\cal O}(\epsilon^4)$ SSE ?) calculation and is therefore beyond the scope of this analysis. Keeping these problems in mind, we nevertheless are convinced that the $\bar{q}$-dependence is described reasonably well by the ${\cal O}(\epsilon^3)$ calculation and that the problems described above only refer to the correct normalization of the theory at the real photon point $\bar{q}\rightarrow 0$. We base this expectation on the observation that the relevant scale of the $\bar{q}$-evolution in Eq.(\[SSEtaylorab\]) at small momentum transfer is given by the quantity $\bar{q}^2/m_\pi^2$, i.e. the momentum dependence arises from the “pion-cloud” of the nucleon. At the next order—${\cal O}(\epsilon^4)$—new diagrams are expected to correct the normalization at the photon point. The $\bar{q}$-dependence of these diagrams however is then expected to scale with $\bar{q}^2/(M_N m_\pi)$, i.e. it should be much weaker due to the appearance of the extra suppression factor $m_\pi/M_N$ of the next order. Whether this expectation will hold true can, of course, only be decided once $\bar{\alpha}_E(\bar{q}),\;\bar{\beta}_M(\bar{q})$ have been explictly calculated to ${\cal O}(\epsilon^4)$. An analysis of the renormalization of $\bar{\alpha}_E,\; \bar{\beta}_M$ in real Compton scattering to ${\cal O}(\epsilon^4)$ is under way [@GHKM] and will later be extended to the case of VCS at ${\cal O}(\epsilon^4)$. For completeness we also give formal expressions for the two spin-independent GPs. Unlike the case of ${\cal O}(p^3)$ HBChPT in SSE to ${\cal O}(\epsilon^3)$ we were not able to obtain closed form expressions— $$\begin{aligned} \label{eq:bqSSE} \bar{\alpha}^{(III)}_E(\bar{q})&=& \frac{e^2 g_{A}^2 m_\pi}{64\pi^2 F_{\pi}^2}\;\frac{4+2\frac{\bar{q}^2}{m_{\pi}^2}- \left(8-2\frac{\bar{q}^2}{m_{\pi}^2} -\frac{\bar{q}^4}{m_{\pi}^4}\right)\frac{m_\pi}{\bar{q}} \arctan\frac{\bar{q}}{2 m_{\pi}}}{\bar{q}^2\left(4+\frac{\bar{q}^2}{m_{\pi}^2}\right)} \nonumber \\ &+&\frac{e^2}{8\pi}\,\frac{8\,g_{\pi N\Delta}^2}{9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial^2}{\partial w\,^2}\; \mbox{\Large \{}\; J_0\left(\omega^\prime-\Delta,m_\pi^2\right)+J_0\left(-\omega^\prime -\Delta,m_\pi^2\right)\nonumber \\ & &-2\left[J_2^\prime\left(\omega^\prime x-\Delta,m_\pi^2\right)+J_2^\prime \left(-\omega^\prime x-\Delta,m_\pi^2\right)+J_2^\prime \left(\omega^\prime x-\Delta,\tilde{m}^2\right)+J_2^\prime \left(-\omega^\prime x-\Delta,\tilde{m}^2\right)\right] \nonumber \\ & &+4\left(1-y\right)\left[5\left(J_6^{\prime\prime}\left(T-\Delta, \hat{m}^2\right)+J_6^{\prime\prime}\left(-T-\Delta, \hat{m}^2\right)\right)-\left(T^2+m_\pi^2-\hat{m}^2- T\,\omega^\prime\right)\right.\nonumber \\ & &\phantom{+4\left(1-y\right)}\left. \left(J_2^{\prime\prime}\left(T-\Delta, \hat{m}^2\right)+J_2^{\prime\prime}\left(-T-\Delta, \hat{m}^2\right)\right)\right] \nonumber \\ & &-2\left[3\,J_2^\prime\left(-\Delta,m_f^2\right)+\left(m_f^2-m_\pi^2\right) J_0^\prime\left(-\Delta,m_f^2\right)\right] \nonumber \\ & &-x\left(1-2x\right)\bar{q}^2\left(J_0^\prime \left(\omega^\prime x-\Delta,\tilde{m}^2\right)+J_0^\prime \left(-\omega^\prime x-\Delta,\tilde{m}^2\right)\right) \nonumber \\ & &-2\left[\left(1-y\right)\left(14y^2-9y+1\right)\bar{q}^2\left( J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)+ J_2^{\prime\prime}\left(-T-\Delta,\hat{m}^2\right)\right)\right. \nonumber \\ & &\phantom{-2}\left.+ y\left(1-y\right)\left(1-2y\right)\bar{q}^2\left(T^2+m_\pi^2 -\hat{m}^2-T\,w\right)\left( J_0^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)+ J_0^{\prime\prime}\left(-T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0, \omega^\prime\rightarrow 0}\; , \nonumber\\ \bar{\beta}^{(III)}_M(\bar{q})&=& \frac{e^2 g_{A}^2 m_\pi}{128\pi^2 F_{\pi}^2}\;\frac{-\left(4+2\frac{\bar{q}^2}{m_{\pi}^2}\right) +\left(8+6\frac{\bar{q}^2}{m_{\pi}^2}+\frac{\bar{q}^4}{m_{\pi}^4} \right)\frac{m_\pi}{\bar{q}}\arctan \frac{\bar{q}}{2 m_{\pi}}}{\bar{q}^2\left(4+\frac{\bar{q}^2}{ m_{\pi}^2}\right)}+\frac{e^22\,b_1^2}{9\pi M_N^2\Delta}\nonumber\\ &+&\frac{e^2}{4\pi}\,\frac{1}{\bar{q}}\,\frac{32\,g_{\pi N\Delta}^2}{9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial}{\partial w}\; \mbox{\Large \{}\;\nonumber \\ & &\left[\left(1-y\right)\left(-1+x-8xy+7(y-y^2+xy^2)\right)\bar{q}\,w \left(J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)+ J_2^{\prime\prime}\left(-T-\Delta,\hat{m}^2\right)\right)\right. \nonumber \\ & &\left.-y\left(1-y\right)^2\left(1-x\right)\left(T^2+m_\pi^2 -\hat{m}^2-T\,w\right)\bar{q}\,w\left(J_0^{\prime\prime} \left(T-\Delta,\hat{m}^2\right)+ J_0^{\prime\prime}\left(-T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0, \omega^\prime\rightarrow 0} \; . \nonumber \\\end{aligned}$$ We note that the relevant J-functions are defined in appendix \[Jfunctions\] and the mass-/energy variables occuring in Eq.(\[eq:bqSSE\]) have been given in Eq.(\[definitions\]). The results of Eq.(\[eq:bqSSE\]) are also shown in Fig.\[figabSSE\]. Once more, we do not advocate the use of these ${\cal O}(\epsilon^3)$ SSE curves in a realistic analysis of VCS at this point as there are large known cancellations which are not included yet to this order. A realistic use of these curves could be the prescription $$\begin{aligned} \bar{\alpha}_E^{ren.}(\bar{q})&=&\bar{\alpha}_E^{(III)}(\bar{q})- \bar{\alpha}_E^{(III)}(0)+\bar{\alpha}_E^{exp.} \nonumber \\ \bar{\beta}_M^{ren.}(\bar{q})&=&\bar{\beta}_M^{(III)}(\bar{q}) -\bar{\beta}_M^{(III)}(0)+\bar{\beta}_M^{exp.}\;, \label{prescript}\end{aligned}$$ where the index $exp.$ refers to the current experimental numbers for $\bar{\alpha}_E,\;\bar{\beta}_M$ of ref.[@PPol]. The results of this operation are shown in Fig.\[figabren\]. There one can clearly see that the $\Delta$(1232) related effects at ${\cal O}(\epsilon^3)$ SSE [*enhance*]{} the $\bar{q}$-trend already seen at ${\cal O}(p^3)$ HBChPT. Of course we want to emphasize that the prescription of Eq.(\[prescript\]) leaves the strict realm of chiral effective theories and just constitutes an ad hoc fix to include [*some*]{} effects that are of higher order in the (slowly converging) SSE expansion for the spin-independent GPs. Now we move to the generalized spin polarizabilities in SSE. ### Spin-dependent results Once more we start from a discussion of the GPs near $\bar{q}=0$. It should be noted that there are no $\Delta$ pole contributions[^15] to any of the generalized spin polarizabilities at ${\cal O}(\epsilon^3)$, quite in contrast to the real Compton (Ragusa) spin polarizabilities $\gamma_2,\;\gamma_4$ [@HHKK]. The ${\cal O}(\epsilon^3)$ results for the four independent spin GPs therefore exclusively arise from the $N\pi$- and $\Delta\pi$-continuum graphs of Figs.(\[figgnpi\],\[figgdpi\]) and can be found from the expressions given in appendices \[appnpi\] and \[appdpi\]. One obtains $$\begin{aligned} \label{SSEspinreihe} P_{(01,12)1}^{(III)}(\bar{q})&=&-\frac{\sqrt{2}\;g_{A}^2}{288\pi^2F_{\pi}^2m_{\pi}^2} -{\sqrt{2}g_{\pi N\Delta}^2\over 324\pi^2F_\pi^2}\left({1\over \Delta^2-m_\pi^2}-{\Delta\over (\Delta^2-m_\pi^2)^{3\over 2}} \ln R\right)\nonumber\\ & &+\frac{\bar{q}^2}{m_\pi^2}\left[\frac{\sqrt{2}\;g_{A}^2}{1440 \pi^2F_{\pi}^2m_{\pi}^2}-{\sqrt{2}g_{\pi N\Delta}^2\over 3240\pi^2F_\pi^2} \left({\Delta^2+2m_\pi^2\over (\Delta^2-m_\pi^2)^2} -{3\Delta m_\pi^2\over (\Delta^2-m_\pi^2)^{5\over 2}}\ln R\right) \right] +{\cal O}(\bar{q}^4) \nonumber \\ &=&\left\{-7.28+0.735+\frac{\bar{q}^2}{m_\pi^2}\left[1.46-0.067\right] +\frac{\bar{q}^4}{m_\pi^4}\left[-0.312+0.009\right]+\dots\right\} \times 10^{-3}\;\mbox{fm}^4\nonumber \\ P_{(11,02)1}^{(III)}(\bar{q})&=&-\frac{\sqrt{2}\;g_{A}^2}{144\sqrt{3}\;\pi^2F_{\pi}^2 m_{\pi}^2} -\sqrt{2\over 3}{g_{\pi N\Delta}^2\over 162\pi^2 F_\pi^2}\left({1\over \Delta^2-m_\pi^2}-{\Delta\over (\Delta^2-m_\pi^2)^{3\over 2}}\ln R\right) \nonumber\\ & &+\frac{\bar{q}^2}{m_\pi^2}\left[\frac{\sqrt{2}\;g_{A}^2}{720\sqrt{3} \;\pi^2F_{\pi}^2m_{\pi}^2}-\sqrt{2\over 3}{g_{\pi N\Delta}^2\over 1620\pi^2 F_\pi^2}\left({\Delta^2+2m_\pi^2\over (\Delta^2-m_\pi^2)^2} -{3\Delta m_\pi^2\over (\Delta^2-m_\pi^2)^{5\over 2}}\ln R\right) \right] +{\cal O}(\bar{q}^4) \nonumber \\ &=&\left\{-8.41+0.848+\frac{\bar{q}^2}{m_\pi^2}\left[1.68-0.077\right] +\frac{\bar{q}^4}{m_\pi^4}\left[-0.360+0.010\right]\dots\right\} \times 10^{-3}\;\mbox{fm}^4\nonumber \\ P_{(11,00)1}^{(III)}(\bar{q})&=&0+\frac{\bar{q}^2}{m_{\pi}^2} \left[-\frac{5\;g_{A}^2}{144\sqrt{3}\pi^2F_{\pi}^2} -\sqrt{1\over 3}{5g_{\pi N\Delta}^2\over 162\pi^2 F_\pi^2}\left({m_\pi^2\over \Delta^2-m_\pi^2}-{\Delta m_\pi^2\over (\Delta^2-m_\pi^2)^{3\over 2}}\ln R\right)\right]+{\cal O}(\bar{q}^4) \nonumber\\ &=&\left\{0+0+\frac{\bar{q}^2}{m_\pi^2}\left[-1.49+0.15\right] +\frac{\bar{q}^4}{m_\pi^4}\left[0.208-0.002\right]+\dots\right\} \times 10^{-2}\;\mbox{fm}^2 \nonumber\\ \hat{P}_{(01,1)1}^{(III)}(\bar{q})&=&-\frac{g_{A}^2}{48\sqrt{6}\;\pi^2 F_{\pi}^2m_{\pi}^2}-\sqrt{1\over 6}{g_{\pi N\Delta}^2\over 54\pi^2F_\pi^2}\left({1\over \Delta^2-m_\pi^2}-{\Delta\over (\Delta^2-m_\pi^2)^{3\over 2}}\ln R\right)\nonumber\\ & &+\frac{\bar{q}^2}{m_\pi^2}\left[\frac{g_{A}^2}{360\sqrt{6}\;\pi^2 F_{\pi}^2m_{\pi}^2}-\sqrt{1\over 6}{g_{\pi N\Delta}^2\over 810\pi^2F_\pi^2}\left({\Delta^2+2m_\pi^2\over (\Delta^2-m_\pi^2)^2} -{3\Delta m_\pi^2\over (\Delta^2-m_\pi^2)^{5\over 2}}\ln R\right) \right] +{\cal O}(\bar{q}^4) \nonumber \\ &=&\left\{-12.6+1.272+\frac{\bar{q}^2}{m_\pi^2}\left[1.68-0.077\right] +\frac{\bar{q}^4}{m_\pi^4}\left[-0.300+0.009\right]+\dots\right\} \times 10^{-3}\;\mbox{fm}^4.\end{aligned}$$ First, we observe that SSE to ${\cal O}(\epsilon^3)$ obeys the C-invariance constraint [@fm2] $\lim_{\bar{q}\rightarrow 0}P_{(11,00)1}^{(III)}(\bar{q})=0$, as does the ${\cal O}(p^3)$ HBChPT calculation in Eq.(\[spintaylor\]). Second, we note that there is no strong renormalization of the above[^16] spin-dependent GPs at the real photon point due to $\Delta$(1232) related effects. We observe that in general the effects from the $\Delta\pi$-continuum are small and [*always*]{} interfere [*destructively*]{} with the corresponding contribution from the $N\pi$-continuum, in contrast to the [*constructive*]{} interference in the spin-independent sector of section \[sir\]. As in the previous section, we were not able to give the full spin-dependent ${\cal O}(\epsilon^3)$ results in a closed form expression but utilize a Feynman-parameter representation and the J-functions defined in appendix \[Jfunctions\]: $$\begin{aligned} \label{sdspin} P_{(01,12)1}^{(III)}(\bar{q})&=& -\frac{g_{A}^2}{24\sqrt{2}\;\pi^2 F_{\pi}^2\bar{q}^2}\left[1-g({\bar{q}\over 2m_\pi}) \right]\nonumber \\ & &-\frac{\sqrt{2}}{3}\,\frac{1}{\bar{q}}\,\frac{16\,g_{\pi N\Delta}^2}{9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial}{\partial w}\;\mbox{\Large \{} \left[y\left(1-y\right)\bar{q}^2\left(J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2 \right)-J_2^{\prime\prime}\left(-T-\Delta,\hat{m}^2\right)\right)\right. \nonumber \\ & &\left. +x\left(1-x\right)y\left(1-y\right)^3\bar{q}^2\omega^{\prime\,2}\left( J_0^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_0^{\prime\prime} \left(-T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0,\omega^\prime\rightarrow 0} \nonumber\\ P_{(11,02)1}^{(III)}(\bar{q})&=&-\frac{g_{A}^2}{12\sqrt{6}\;\pi^2F_{\pi}^2\bar{q}^2} \left[1-g({\bar{q}\over 2m_\pi})\right]\nonumber \\ &-&\frac{\sqrt{2}}{3\sqrt{3}}\,\frac{1}{\bar{q}}\,\frac{8\,g_{\pi N\Delta}^2} {9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial^2}{\partial w^{\prime\;2}} \;\mbox{\Large \{} \left[x\,y\left(1-y\right)^2\left(1-2y\right)\bar{q}^3\omega^\prime\left( J_0^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_0^{\prime\prime}\left( -T-\Delta,\hat{m}^2\right)\right)\right] \nonumber\\ & &+2\left[x\left(1-y\right)^2\bar{q}\,\omega^\prime\left( J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_2^{\prime\prime}\left( -T-\Delta,\hat{m}^2\right)\right)\right] \nonumber\\ & &-2\left[y\left(1-y\right)\bar{q}\,\omega^\prime\left( J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_2^{\prime\prime}\left( -T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0,\omega^\prime\rightarrow 0} \nonumber\\ P_{(11,00)1}^{(III)}(\bar{q})&=&\frac{g_{A}^2}{12\sqrt{3}\;\pi^2F_{\pi}^2} \left[2-\left(2+{3\bar{q}^2\over 4m_\pi^2}\right) g({\bar{q}\over 2m_\pi})\right] +\sqrt{2}\,\bar{q}^2\,P_{(11,02)1}^{(III)}(\bar{q}) -\frac{\bar{q}}{\sqrt{3}}\,\frac{8\,g_{\pi N\Delta}^2} {9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial^2}{\partial w^{\prime\;2}} \;\mbox{\Large \{}\nonumber\\ & &\left[\left(1-y\right)\left(1-2y\right)\bar{q}\,\omega^\prime \left(J_2^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_2^{\prime\prime} \left(-T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0,\omega^\prime\rightarrow 0} \nonumber\\ \hat{P}_{(01,1)1}^{(III)}(\bar{q})&=&\frac{g_{A}^2}{24\sqrt{6}\;\pi^2F_{\pi}^2\bar{q}^2} \left[3-\left(3+{\bar{q}^2\over m_\pi^2}\right) g({\bar{q}\over 2m_\pi})\right] +\sqrt{1\over 3}P_{(01,12)1}^{(III)}(\bar{q})\nonumber\\ &-&\frac{2\sqrt{2}}{3\sqrt{3}}\,\frac{1}{\bar{q}^2}\,\frac{4\,g_{\pi N\Delta}^2} {9\,F_\pi^2} \int_0^1 dx\int_0^1 dy\;\frac{\partial}{\partial w^{\prime}} \;\mbox{\Large \{}\left[-\left(J_0^\prime\left(\omega^\prime-\Delta,m_\pi^2\right) -J_0^\prime\left(-\omega^\prime-\Delta,m_\pi^2\right)\right)\right]\nonumber\\ & &+2\,\frac{3}{d-1} \left[J_2^\prime\left(\omega^\prime x-\Delta,m_\pi^2\right)-J_2^\prime \left(-\omega^\prime x-\Delta,m_\pi^2\right)+J_2^\prime\left(\omega^\prime x- \Delta,\tilde{m}^2\right)-J_2^\prime\left(-\omega^\prime x-\Delta,\tilde{m}^2 \right)\right]\nonumber\\ & &-4\left[x\left(1-x\right)y\left(1-y\right)^3\bar{q}^2\omega^{\prime\;2}\left( J_0^{\prime\prime}\left(T-\Delta,\hat{m}^2\right)-J_0^{\prime\prime}\left( -T-\Delta,\hat{m}^2\right)\right)\right] \mbox{\Large \} }\biggl|_{\,\cos\theta\rightarrow 0,\omega^\prime\rightarrow 0}\end{aligned}$$ Note that the auxiliary function $g(x)$ has already been defined in Eq.(\[auxfunction\]) and the mass-/energy-variables again correspond to the structures introduced in Eq.(\[definitions\]). We present the [*absolute*]{} ${\cal O}(\epsilon^3)$ SSE predictions for the four independent spin GPs in Fig.\[figSSEspin1\]. It clearly shows that the ${\cal O}(\epsilon^3)$ curves are always lying higher than the corresponding ${\cal O}(p^3)$ HBChPT ones. In all cases the two curves share a similar behaviour in their $\bar{q}$-dependence—leading to the conclusion that there is “no dramatic” signal of the $\Delta$(1232) resonance in the spin-dependent GPs to ${\cal O}(\epsilon^3)$ compared to the dominant contributions from the $N\pi$-continuum. Finally we note that the remaining (linearly dependent) generalized spin polarizabilities $P_{(01,01)1}^{(III)},P_{(11,11)1}^{(III)},\hat{P}_{(11,2)1}^{(III)}$ may be found via the charge-conjugation constraint Eq.(\[cinv\]). The Mainz Experiment ==================== As mentioned above, the pioneering VCS experiment[^17] has taken place at Mainz, and preliminary results of the analysis are now available [@nstar]. The measurement was performed at $\bar{q}^2=0.36$ GeV$^2$ and used parallel kinematics although relativistic forward-focussing allowed access to events as much as $\pm$ 26 degrees out of plane. Nevertheless the desired generalized polarizabilities were hidden behind a very large Bethe-Heitler background and their extraction was a real experimental tour de force. Consulting Fig.\[figab\], we note that at $\bar{q}^2=0.36$ GeV$^2$ the ${\cal O}(p^3)$ HBChPT calculation predicts that $\bar{\alpha}_E(\bar{q})$ should have decreased by as much as 50% from its real photon value, whereas the much smaller GP $\bar{\beta}_M(\bar{q})$ is predicted to have slightly increased. As can be seen from the HBChPT predictions in Figs.\[figspin1\],\[figspin2\], the spin-dependent GPs will dramatically change with regard to the real photon point. Thus the confrontation of theoretical predictions with the MAMI results offers a chance to realistically test theoretical pictures of nucleon structure. Essentially two quantities were determined experimentally—the combination $P_{LL}-P_{TT}/\epsilon$ of longitudinal and transverse response functions, which is primarily sensitive to the generalized electric polarizability $\alpha_E(\bar{q})$ (plus linear combinations of spin GPs) [@nstar; @Guichon], as well as the interference term $P_{LT}$, depending on the generalized magnetic polarizability $\beta_M(\bar{q})$ and the spin GP $P^{(01,01)1}(\bar{q})$ [@nstar; @Guichon] (which itself can be expressed as a linear combination of the 2 spin GPs $P^{(11,00)1}(\bar{q})$ and $P^{(11,02)1}(\bar{q})$ via the C-invariance constraint of Eq.(\[cinv\])). Results of the experiment together with predictions from ${\cal O}(p^3)$ HBChPT[^18] and other theoretical models are given in Table 1. Obviously only the chiral picture (refs.[@HHKS1; @HHKS2]; sections \[HBChPTscalar\], \[HBChPTspin\] of this work; Figs. \[figab\], \[figspin1\], \[figspin2\]) is able to explain the experimental features at this point. Of course this is only a single experiment at a single momentum transfer and results from other laboratories and other values of $\bar{q}^2$ are needed in order to confirm our predictions. Nevertheless the agreement is certainly encouraging. Summary ======= Virtual Compton scattering—$eN\rightarrow e'N\gamma$—opens the way to high resolution study of nucleon structure by measuring generalized polarizabilities (GPs), which are momentum-dependent analogues of the familiar polarizabilities determined in real Compton scattering. In this work, we have calculated these quantities within the framework of conventional heavy baryon chiral perturbation theory to third order in the momentum expansion as well as to third order in the “small-scale expansion”, which contains the $\Delta(1232)$ as an explicit degree of freedom. As originally defined by Guichon et al., there exist ten such GPs, three being associated with spin-independent correlations and seven connected with spin-flip structures. At third order both in HBChPT and in SSE only six of these—two spin-independent and four spin-dependent—survive, and we have calculated these directly. In the case of the $\Delta$-pole and $\pi N$ loop contributions, we were able to obtain results for the GPs which are simple analytic forms, while in the case of the corrresponding $\pi\Delta(1232)$-continuum contributions only numerical results could be given. We briefly discussed the results from the first VCS experiment on the proton from Mainz at $Q^2=0.33$ GeV$^2$. The success in predicting the measured response functions resulted from a combination of a sharp falloff of $\bar{\alpha}_E^{(3)}(\bar{q})$, a slight rise of $\bar{\beta}_M^{(3)}(\bar{q})$ and a strong incarease in the contributing spin GPs with momentum-transfer $\bar{q}$. All these effects are intimately related to the chiral dynamics of the pion cloud, which can be calculated very precisely in chiral effective theories like HBChPT and SSE—with HBChPT at least in the spin-independent sector having the better convergence behaviour as far as we can tell at this point. In particular, for the case of the generalized magnetic polarizability both HBChPT and SSE predict a [*rising*]{} behaviour as one goes away from the real photon point—$\bar{q}^2=0$—up to a momentum $\bar{q}^2\sim 0.1$ GeV$^2$. This is a distinctive feature of the chiral calculations and generally not found in simple quark model evaluations. It expresses the feature that chiral invariance requires local regions [*both*]{} of paramagnetic (at small distances) and diamagnetic (at larger distances) magnetic polarizability densities in the nucleon. Aside from the widely discussed $\bar{q}$-dependences of the generalized electric and magnetic polarizabilities, the strong variation of the GPs in the spin-sector is likely to be of interest for further study, both on the experimental and on the theoretical side. Considering the results of the chiral calculations for the spin polarizabilities in real Compton scattering we believe that the ${\cal O}(\epsilon^3)$ SSE calculation should be quite competitive with the ${\cal O}(p^3)$ HBChPT analysis at least as far as the generalized spin-polarizabilities are concerned. Future measurements at Bates, MAMI and TJNAF will clarify this issue. It goes without saying that our calculation is preliminary in that it does not include important corrections arising at ${\cal O}(p^4)/{\cal O}(\epsilon^4)$—see, [*e.g.*]{} the discussed normalization problems in $\bar{\alpha}_E^{(III)}(0),\,\bar{\beta}_M^{(III)}(0)$. An ${\cal O}(p^4)$ HBChPT analysis has been carried out in the case of the real Compton electric and magnetic polarizabilities in ref.[@BKSM] and important corrections and uncertainties were found which, while not drastically modifying the basic numerical predictions obtained at ${\cal O}(p^3)$, did introduce sizable uncertainties into the predictions due to unknown counterterms which had to be estimated via resonance exchange. We may then anticipate a similar behavior here—that such higher order corrections will not change the basic pattern of the chiral ${\cal O}(p^3)/{\cal O}(\epsilon^3)$ predictions, but mainly only correct the (photon-point) normalization. However, verification of this assumption awaits detailed future calculations. Lastly, we stress once more the motivation for performing electron scattering experiments on the nucleon: Different theoretical approaches may yield comparable results at the real photon point, but the details of the underlying dynamics can be analyzed in a much more powerful way by studying the $Q^2$-dependence. In conclusion, VCS on the nucleon has matured to become a precise testing ground for our notions of nucleon structure at low energies. [**Acknowledgments**]{} The authors acknowledge many helpful discussions with N. d’Hose, U.-G. Mei[ß]{}ner, A. Metz, R. Miskimen, S. Scherer, J. Shaw and M. Vanderhaeghen. BRH would like to acknowledge the support of the Alexander von Humboldt Foundation and the National Science Foundation, as well as the hospitality of the IKP at Forschungszentrum J[" u]{}lich. This work is also supported in part by the Deutsche Forschungsgemeinschaft (SFB443). Loop Functions {#Jfunctions} ============== The formalism to calculate the loop diagrams for Compton scattering both in ChPT and in the small scale expansion has been described in detail in the appendices of ref.[@delta]. Therefore we shall only give some definitions of the basic building blocks. We express the invariant amplitudes of Feynman diagrams containing pion-nucleon loops in terms of $d$-dimensional J-functions, defined via $$\begin{aligned} {1\over i}\int{d^d\ell\over (2\pi)^d} {\{1,\ell_\mu\ell_\nu,\ell_\mu\ell_\nu\ell_\alpha\ell_\beta\}\over (v\cdot\ell-W-i\eta)(M^2-\ell^2-i\eta)}&=& \left\{J_0(W,M), g_{\mu\nu}J_2(W,M)+v_\mu v_\nu J_3(W,M), \right.\nonumber \\ & &\left. (g_{\mu\nu} g_{\alpha\beta}+{\rm perm.})J_6(W,M)+\dots \right\} , \label{eq:props}\end{aligned}$$ with the small imaginary part $\eta$ denoting the location of the pole. In the case of Compton scattering at ${\cal O}(p^3)$ or ${\cal O}(\epsilon^3)$, all loop-integrals can be expressed in terms of the four functions $\Delta_M,J_0(W,M),J_2(W,M),J_6(W,M)$, which are related via $$\begin{aligned} J_2 \left(W,M\right) & = & \frac{1}{d-1} \left[ \left( M^2 - W^2 \right) J_0 \left(W,M\right) - W \; \Delta_{M} \right] \label{J_2}\nonumber\\ J_6 \left(W,M\right) & = & \frac{1}{d+1} \left[ \left( M^2 - W^2 \right) J_2 \left(W,M\right) - \frac{M^2 W}{d} \; \Delta_M \right] \label{J_6} \, ,\end{aligned}$$ with $\Delta_M$ denoting the meson integral $$\Delta_M=\frac{1}{i}\int\frac{d^d \ell}{(2\pi)^d}\;\frac{1}{M^2-\ell^2-i\eta} \; ,$$ and $J_0\left(W,M\right)$ being the basic meson-baryon integral with arbitrary energy $W$ and mass variable $M$. Explicit representations for these building blocks can be found in appendix A of ref.[@delta]. Finally, we remind the reader that all propagator structures encountered in the calculation can be reduced to the basic forms of Eq.(\[eq:props\]) by taking derivatives of the J-functions with respect to the square of the mass— $$\begin{aligned} J_i'\left(W,M\right) & = & \frac{\partial}{\partial \left( M^2 \right)} J_i \left(W,M\right) \, , \nonumber\\ J_i''\left(W,M\right) & = & \frac{\partial^2}{\partial \left( M^2 \right)^2} J_i \left(W,M\right) .\end{aligned}$$ For a more detailed discussion we refer to ref.[@trh]. $N\pi$ Loop Amplitudes in VCS {#appnpi} ============================= Using the J-function formalism defined in Appendix A, one can get exact solutions for the nine $N\pi$-loop diagrams of Fig.\[figgnpi\]. By $\tilde{\epsilon}_\mu \; (q_\mu)$ we denote the polarization-vector (four-momentum) of the incoming virtual photon, and by $\epsilon_{\mu}^{\prime} \; (q^{\prime}_\mu)$ the corresponding quantities in the outgoing real photon with energy $\omega^\prime$. In order to make contact with the VCS amplitudes defined in Eq.(\[eq:vcs12\]), we use the Coulomb gauge $$\tilde{\epsilon}^\mu=(0,\vec{\epsilon}_T +{q^2\over \omega^2}\vec{\epsilon}\cdot\hat{q}\hat{q})$$ The amplitudes can then be cast in the form $$\begin{aligned} Amp_{1+2}^{N\pi} &=& i \frac{g_{A}^2}{F_{\pi}^2} \; \bar{u}_2(r^\prime) \left\{ -\frac{1}{2} \tilde{\epsilon} \cdot \epsilon^\prime \left[ J_0 ( \omega^\prime , m_{\pi}^2 ) +J_0(-\omega^\prime , m_{\pi}^2 ) \right] + \; [ S \cdot \epsilon^\prime,S\cdot \tilde{\epsilon} ] \left[ J_0(\omega^\prime ,m_{\pi}^2)- J_0(-\omega^\prime ,m_{\pi}^2) \right] \right\} u_1(r) \nonumber\\ Amp_{3+6}^{N\pi} &=& i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) \left\{ \tilde{\epsilon} \cdot\epsilon^\prime \left[ J^{\prime}_2 ( \omega^\prime x , m_{\pi}^2 ) + J^{\prime}_2 ( -\omega^\prime x, m_{\pi}^2 ) \right] \right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } \left.-2 \; [S\cdot\epsilon^\prime, S\cdot \tilde{\epsilon} ] \left[ J^{\prime}_2 ( \omega^\prime x,m_{\pi}^2 ) - J^{\prime}_2 ( -\omega^\prime x, m_{\pi}^2 ) \right] \right\} u_1(r) \nonumber\\ Amp_{4+5}^{N\pi} &=& i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) \left\{ \tilde{\epsilon} \cdot \epsilon^\prime \left[ J_{2}^\prime ( \omega^\prime x , \tilde{m}^2 ) + J_{2}^\prime ( - \omega^\prime x ,\tilde{m}^2)\right]\right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } -2 \; [ S\cdot \epsilon^\prime , S\cdot\tilde{\epsilon} ] \left[ J_{2}^\prime ( \omega^\prime x , \tilde{m}^2 ) - J_{2}^\prime ( - \omega^\prime x , \tilde{m}^2 ) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } -\frac{1}{2}x\left(1-2x\right) \tilde{\epsilon} \cdot q \; \epsilon^\prime \cdot q \left[ J_{0}^\prime ( \omega^\prime x , \tilde{m}^2 ) + J_{0}^\prime(-\omega^\prime x, \tilde{m}^2 ) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } \left. + x\left(1-2x\right) [ S\cdot\epsilon^\prime,S\cdot q ] \; \tilde{\epsilon} \cdot q \left[ J_{0}^\prime ( \omega^\prime x, \tilde{m}^2 ) - J_{0}^\prime ( - \omega^\prime x , \tilde{m}^2)\right]\right\} u_1(r) \nonumber\\ Amp_{7+8}^{N\pi} &=& i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \int_{0}^{1}dy \left(1-y\right) \bar{u}_2(r^\prime) \times \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } \left\{ \tilde{\epsilon} \cdot \epsilon^\prime \left[-2\left(d+1\right)\left(J_{6}^{\prime\prime}(T,\hat{m}^2) + J_{6}^{\prime\prime} (-T,\hat{m}^2 )\right)\right.\right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \{ \tilde{\epsilon} \cdot \epsilon^\prime } \left. + 2\left(T^2-(\hat{m}^2-m_{\pi}^2+\frac{q\cdot q^\prime}{ \omega^\prime}T)\right)\left(J_{2}^{\prime\prime}(T,\hat{m}^2)+ J_{2}^{\prime\prime}(-T,\hat{m}^2)\right) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +\left[\left(1-d\right)\epsilon^\prime\cdot a\; \tilde{\epsilon}\cdot b -\epsilon^\prime\cdot\left(c+d\right) \tilde{\epsilon}\cdot b- 2\;\epsilon^\prime\cdot a\; \tilde{\epsilon}\cdot\left(c+d\right)\right] \left( J_{2}^{\prime\prime} (T, \hat{m}^2)+J_{2}^{\prime\prime}(-T,\hat{m}^2)\right)\nonumber\\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +\left(T^2-(\hat{m}^2-m_{\pi}^2+\frac{q\cdot q^\prime}{ \omega^\prime} T)\right) \epsilon^\prime\cdot a\; \tilde{\epsilon}\cdot b \left(J_{0}^{\prime\prime} (T,\hat{m}^2)+ J_{0}^{\prime\prime}(-T,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +2\;\tilde{\epsilon}\cdot b\; [S\cdot\epsilon^\prime,S\cdot(q-q^\prime)] \left(J_{2}^{\prime\prime}(T,\hat{m}^2)-J_{2}^{\prime\prime}( -T,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +4\;\epsilon^\prime\cdot a\; [S\cdot\tilde{\epsilon},S\cdot(q-q^\prime)] \left(J_{2}^{\prime\prime}(T,\hat{m}^2)-J_{2}^{\prime\prime}( -T,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +4\;\tilde{\epsilon}\cdot\epsilon^\prime\;[S\cdot c,S\cdot d]\left( J_{2}^{\prime\prime} (T,\hat{m}^2)- J_{2}^{\prime\prime}( -T,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} }\left. +2\; \epsilon^\prime\cdot a\;\tilde{\epsilon}\cdot b\; [S\cdot c,S\cdot d] \left(J_{0}^{\prime\prime}(T,\hat{m}^2)-J_{0}^{\prime\prime}( -T,\hat{m}^2)\right) \right\} u_1(r) \nonumber\\ Amp_{9}^{N\pi} &=& i \frac{g_{A}^2}{F_{\pi}^2} \; \bar{u}_2(r^\prime) \; \tilde{\epsilon} \cdot \epsilon^\prime \; u_1(r) \int_{0}^{1}dx \left\{ \left(d-1\right) J_{2}^\prime ( 0 , m_{f}^2 ) + \left(m_{f}^2-m_{\pi}^2\right) J_{0}^\prime ( 0 ,m_{f}^2) \right\} ,\end{aligned}$$ with $$\begin{aligned} a_\mu&=&-q_\mu \; y \nonumber\\ b_\mu&=&q^{\prime}_\mu\left(2y+2x-2y x-2\right)-q_\mu\left(2y-1\right) \nonumber\\ c_\mu&=&q^{\prime}_\mu\left(y+x-y x\right)-q_\mu y \nonumber\\ d_\mu&=&q^{\prime}_\mu\left(y+x-y x-1\right)-q_\mu\left(y-1\right),\end{aligned}$$ and the energy and mass variables $T,\hat{m},\tilde{m},m_f$ as defined in section \[chiamp\]. $\Delta\pi$ Loop Amplitudes in VCS {#appdpi} ================================== The 9 $\pi\Delta$ continuum diagrams are shown in Fig.\[figgdpi\]. We find $$\begin{aligned} Amp_{1+2}^{\Delta\pi}&=& i \frac{8 g_{\pi N\Delta}^2}{3 F_{\pi}^2} \; \bar{u}_2(r^\prime) \left\{ -{1\over 2}\frac{d-2}{d-1} \tilde{\epsilon} \cdot \epsilon^\prime \left[J_0(\omega^\prime -\Delta ,m_{\pi}^2)+J_0(-\omega^\prime -\Delta, m_{\pi}^2)\right]\right.\nonumber\\ &-&\left.{1\over 2}\frac{2}{d-1}\;[S\cdot\epsilon^\prime, S\cdot\tilde{\epsilon}]\left[J_0(\omega^\prime -\Delta ,m_{\pi}^2)- J_0(-\omega^\prime -\Delta,m_{\pi}^2)\right] \right\} u_1(r) \nonumber\\ Amp_{3+6}^{\Delta\pi} &=& i {8g_{\pi N\Delta}^2\over 3F_\pi^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) \left\{ {d-2\over d-1} \tilde{\epsilon} \cdot\epsilon^\prime \left[ J^{\prime}_2 ( \omega^\prime x -\Delta, m_{\pi}^2 ) + J^{\prime}_2 ( -\omega^\prime x-\Delta, m_{\pi}^2 ) \right] \right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } \left.+{2\over d-1} \; [S\cdot\epsilon^\prime, S\cdot \tilde{\epsilon} ] \left[ J^{\prime}_2 ( \omega^\prime x-\Delta,m_{\pi}^2 ) - J^{\prime}_2 ( -\omega^\prime x-\Delta, m_{\pi}^2 ) \right] \right\} u_1(r) \nonumber\\ Amp_{4+5}^{\Delta\pi} &=& i {8g_{\pi N\Delta}^2\over 3F_\pi^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) \left\{{d-2\over d-1} \tilde{\epsilon} \cdot \epsilon^\prime \left[ J_{2}^\prime ( \omega^\prime x -\Delta, \tilde{m}^2 ) + J_{2}^\prime ( - \omega^\prime x -\Delta,\tilde{m}^2 )\right]\right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } +{2\over d-1} \; [ S\cdot \epsilon^\prime , S\cdot\tilde{\epsilon} ] \left[ J_{2}^\prime ( \omega^\prime x-\Delta , \tilde{m}^2 ) - J_{2}^\prime ( - \omega^\prime x-\Delta , \tilde{m}^2 ) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } -\frac{1}{2}{d-2\over d-1} x\left(1-2x\right) \tilde{\epsilon} \cdot q^\prime \; \epsilon^\prime \cdot q \left[ J_{0}^\prime ( \omega^\prime x -\Delta, \tilde{m}^2 ) + J_{0}^\prime (-\omega^\prime x-\Delta, \tilde{m}^2 ) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \int_{0}^{1}dx \; \bar{u}_2(r^\prime) } \left. -{1\over d-1} x\left(1-2x\right) [ S\cdot\epsilon^\prime ,S\cdot q ] \; \tilde{\epsilon} \cdot q \left[ J_{0}^\prime ( \omega^\prime x-\Delta, \tilde{m}^2 ) - J_{0}^\prime ( - \omega^\prime x -\Delta, \tilde{m}^2 )\right]\right\} u_1(r) \nonumber\\ Amp_{7+8}^{\Delta\pi} &=& i {8g_{\pi N\Delta}^2\over 3F_\pi^2} \int_{0}^{1}dx \int_{0}^{1}dy \left(1-y\right) \bar{u}_2(r^\prime) \times \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } \left\{ \tilde{\epsilon} \cdot \epsilon^\prime \left[-2{(d+1)(d-2)\over d-1} \left(J_{6}^{\prime\prime}(T-\Delta,\hat{m}^2) + J_{6}^{\prime\prime} (-T-\Delta,\hat{m}^2 ) \right)\right.\right. \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} \{ \tilde{\epsilon} \cdot \epsilon^\prime } \left. + 2{d-2\over d-1}\left(T^2-(\hat{m}^2-m_{\pi}^2+\frac{q\cdot q^\prime}{ \omega^\prime} T)\right)\left(J_{2}^{\prime\prime}(T-\Delta,\hat{m}^2)+ J_{2}^{\prime\prime}(-T-\Delta,\hat{m}^2)\right) \right] \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +{d-2\over d-1}\left[\left(1-d\right)\epsilon^\prime\cdot a\;\tilde{\epsilon}\cdot b -\epsilon^\prime\cdot\left(c+d\right) \tilde{\epsilon}\cdot b- 2\;\epsilon^\prime\cdot a\;\tilde{\epsilon}\cdot\left(c+d\right)\right] \left( J_{2}^{\prime\prime} (T-\Delta, \hat{m}^2)+J_{2}^{\prime\prime} (-T-\Delta,\hat{m}^2)\right)\nonumber\\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } +{d-2\over d-1} \left(T^2-(\hat{m}^2-m_{\pi}^2+\frac{q\cdot q^\prime}{ \omega^\prime} T)\right) \epsilon^\prime\cdot a\;\tilde{\epsilon}\cdot b \left(J_{0}^{\prime\prime} (T-\Delta,\hat{m}^2)+ J_{0}^{\prime\prime}(-T-\Delta,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } -{2\over d-1}\;\tilde{\epsilon}\cdot b\;[S\cdot\epsilon^\prime,S\cdot(q-q^\prime)] \left( J_{2}^{\prime\prime}(T-\Delta,\hat{m}^2)-J_{2}^{\prime\prime}( -T-\Delta,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } -{4\over d-1}\;\epsilon^\prime\cdot a\;[S\cdot\tilde{\epsilon},S\cdot(q-q^\prime)] \left( J_{2}^{\prime\prime}(T-\Delta,\hat{m}^2)-J_{2}^{\prime\prime}( -T-\Delta,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} } -{4\over d-1}\;\tilde{\epsilon}\cdot\epsilon^\prime\; [S\cdot c,S\cdot d]\left( J_{2}^{\prime\prime} (T-\Delta,\hat{m}^2)- J_{2}^{\prime\prime}( -T-\Delta,\hat{m}^2)\right) \nonumber \\ & & \phantom{i \frac{g_{A}^2}{F_{\pi}^2} }\left. -{2\over d-1}\; \epsilon^\prime\cdot a\; \tilde{\epsilon}\cdot b\;[S\cdot c,S\cdot d] \left(J_{0}^{\prime\prime}(T-\Delta,\hat{m}^2) -J_{0}^{\prime\prime}( -T-\Delta,\hat{m}^2)\right) \right\} u_1(r) \nonumber\\ Amp_{9}^{\Delta\pi} &=& i {8g_{\pi N\Delta}^2\over 3F_\pi^2}\; \bar{u}_2(r^\prime) \; \tilde{\epsilon} \cdot \epsilon^\prime \; u_1(r) \int_{0}^{1}dx \left\{ \left(d-2\right) J_{2}^\prime ( -\Delta , m_{f}^2 ) + {d-2\over d-1}\left(m_{f}^2-m_{\pi}^2\right) J_{0}^\prime ( -\Delta ,m_{f}^2) \right\} ,\end{aligned}$$ $\pi^0$-Pole Contributions {#anomalypolas} ========================== In this section we explicitly give the ${\cal O}(p^3)\equiv {\cal O}(\epsilon^3)$ contribution of $\pi^0$-exchange in the t-channel—Fig.\[figgborn\]f—to the generalized spin-polarizabilities of Eqs.(\[4spin\],\[3spin\]). In the main part of this work we had included this particular effect in the Born part of the structure amplitudes $A_i^{Born}(\omega^\prime,\theta,\bar{q})$ (c.f. Eq.(\[eq:bg\])). However, in the existing literature of VCS many authors prefer to consider $\pi^0$-exchange as a genuine contribution to the spin-polarizabilities. For easier comparison we list our results below and show the resulting GPs in Fig.\[figanom1\]. $$\begin{aligned} \alpha_E(\bar{q})&=&\beta_M(\bar{q})=0\nonumber\\ \hat{P}^{\rm anom}_{(01,1)1}&=&\sqrt{2\over 3}{g_A\over 24\pi^2F_\pi^2} {1\over m_\pi^2+\bar{q}^2}\nonumber\\ P_{(01,12)1}^{\rm anom}&=&-\sqrt{2}{g_A\over 24\pi^2F_\pi^2} {1\over m_\pi^2+\bar{q}^2}\nonumber\\ P_{(11,02)1}^{\rm anom}&=&\sqrt{2\over 3}{g_A\over 12\pi^2F_\pi^2} {1\over m_\pi^2+\bar{q}^2}\nonumber\\ P_{(11,00)1}^{\rm anom}&=&-\sqrt{1\over 3}{g_A\over 12\pi^2F_\pi^2} {\bar{q}^2\over m_\pi^2+\bar{q}^2}\nonumber\\ \hat{P}^{\rm anom}_{(11,2)1}&=&\hat{P}^{\rm anom}_{(01,1)0}= P^{\rm anom}_{(01,01)1}=0\nonumber\\ P^{\rm anom}_{(11,11)1}&=&-\omega_0{g_A\over 12\pi^2F_\pi^2}{1\over m_\pi^2+\bar{q}^2}\end{aligned}$$ A useful identity {#ident} ================= It should be noted that, while making the transition from the chiral loop amplitudes in appendices \[appnpi\], \[appdpi\] to the twelve VCS structure amplitudes $A_i,\;i=1\dots 12$ of Eq.(\[eq:vcs12\]), one also encounters the matrix element $$\vec{\epsilon}^{\;\prime}\cdot\hat{q}\;\vec{\epsilon}\cdot\hat{q}^\prime\;\vec{\sigma} \cdot\left(\hat{q}^\prime\times\hat{q}\right)=-\left(\vec{\epsilon}^{\;\prime}\times \vec{\epsilon}\right)\cdot\left(\hat{q}^\prime\times\hat{q}\right)\;\vec{\sigma}\cdot \left(\hat{q}^\prime\times\hat{q}\right), \label{eq:cross}$$ which has to be brought into a form which accompanies one of the twelve structure amplitudes. To achieve this we start from the identity $$\begin{aligned} \vec{a}\cdot\hat{e}_x\;\vec{b}\cdot\hat{e}_x+\vec{a}\cdot\hat{e}_y\;\vec{b}\cdot \hat{e}_y+\vec{a}\cdot\hat{e}_z\;\vec{b}\cdot\hat{e}_z=\vec{a}\cdot\vec{b} \nonumber\end{aligned}$$ and then construct the 3 orthonormal unit vectors $\hat{e}_a ,\;a=x,y,z$ from the direction vectors $\hat{q},\hat{q}^\prime$ via $$\begin{aligned} \hat{e}_x=\frac{1}{\sin\theta}\left(\hat{q}^\prime\times\hat{q}\right)\times\hat{q} \; ,\quad \hat{e}_y=\frac{1}{\sin\theta}\left(\hat{q}^\prime\times\hat{q}\right) , \quad \hat{e}_z=\hat{q} . \nonumber\end{aligned}$$ Identifying $\vec{a}=(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon})\,{\rm and}\, \vec{b}=\vec{\sigma}$ one finds a relation for the structure of interest, Eq.(\[eq:cross\]), $$\begin{aligned} \left(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon}\right)\cdot\left(\hat{q}^\prime\times \hat{q}\right)\;\vec{\sigma}\cdot\left(\hat{q}^\prime\times\hat{q}\right)= \sin^2\theta\;\sigma\cdot\left(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon}\right)- \vec{a}\cdot\vec{c}\;\vec{b}\cdot\vec{c}-\sin^2\theta\left(\vec{\epsilon}^{\;\prime} \times\vec{\epsilon}\right)\cdot\hat{q}\;\vec{\sigma}\cdot\hat{q} \; , \nonumber\end{aligned}$$ with $\vec{c}=\hat{q}\;\cos\theta-\hat{q}^\prime$. Noting that $$\begin{aligned} \vec{a}\cdot\vec{c}\;\vec{b}\cdot\vec{c}&=&\sin^2\theta\;\vec{a}\cdot\vec{b}+\left( \vec{a}\times\vec{c}\right)\cdot\left(\vec{c}\times\vec{b}\right), \nonumber \\ \left(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon}\right)\cdot\hat{q}\;\vec{\sigma}\cdot \hat{q}&=&-\vec{\epsilon}\cdot\hat{q}\;\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime} \times\hat{q}\right)+\vec{\epsilon}^{\;\prime}\cdot\hat{q}\;\vec{\sigma}\cdot\left( \vec{\epsilon}\times\hat{q}\right)+\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime}\times \vec{\epsilon}\right), \nonumber\end{aligned}$$ one obtains $$\begin{aligned} \left(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon}\right)\cdot\left(\hat{q}^\prime\times \hat{q}\right)\;\vec{\sigma}\cdot\left(\hat{q}^\prime\times\hat{q}\right)&=& \vec{\epsilon}\cdot\hat{q}\;\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime}\times\hat{q} \right)-\vec{\epsilon}^{\;\prime}\cdot\hat{q}\;\vec{\sigma}\cdot\left(\vec{\epsilon} \times\hat{q}\right)-\cos\theta\;\vec{\epsilon}\cdot\hat{q}\;\vec{\sigma}\cdot\left(\vec{ \epsilon}^\prime\times\hat{q}^\prime\right)-\cos\theta\;\vec{\epsilon}\cdot \hat{q}^\prime\;\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime}\times\hat{q}\right) \nonumber \\ & &+\;\vec{\epsilon}\cdot\hat{q}^\prime\;\vec{\sigma}\cdot\left(\vec{\epsilon}^{\;\prime} \times\hat{q}^\prime\right)+\cos\theta\;\vec{\epsilon}^{\;\prime}\cdot\hat{q}\;\vec{ \sigma}\cdot\left(\vec{\epsilon}\times\hat{q}^\prime\right)-\sin^2\theta\;\vec{\sigma} \cdot\left(\vec{\epsilon}^{\;\prime}\times\vec{\epsilon}\right) .\end{aligned}$$ [99]{} See, [*e.g.*]{}, A.I. L’vov, Int. J. Mod. Phys. [**A8**]{}, 5267 (1993); B.R. Holstein, Comm. Nucl. Part. Phys. [**20**]{}, 301 (1992). F.J. Federspiel et al., Phys. Rev. Lett. [**67**]{}, 1511 (1991); A.L. Hallin et al., Phys. Rev. [**C48**]{}, 1497 (1993); A. Zieger et al., Phys. Lett. [**B278**]{}, 34 (1992); B.E. MacGibbon et al., Phys. Rev. [**C52**]{}, 2097 (1995). J. Schmiedmayer et al., Phys. Rev. Lett. [**66**]{}, 1015 (1991). L. Koester, Phys. Rev. [**C51**]{}, 3363 (1995). S. Beane, M. Malheiro, D.R. Phillips, and U. van Kolck, nucl-th/9905023; to appear in Nucl. Phys. [**A**]{}. D. Babusci, G. Giordano, and G. Matone, Phys. Rev. [**C57**]{}, 291 (1998). S. Ragusa, Phys. Rev. [**D47**]{}, 3757 (1993); ibid. [**D49**]{}, 3157 (1994). See, [*e.g.*]{} D. Babusci, G. Giordano, A. L’vov, and A. Nathan, Phys. Rev. [**C58**]{}, 1013 (1998). A.M. Sandorfi et al., Phys. Rev. [**D50**]{}, R6681 (1994). J. Tonnison et al., Phys. Rev. Lett. [**80**]{}, 4382 (1998). D. Drechsel, G. Krein, and O. Hanstein, Phys. Lett. [**B420**]{}, 248 (1998); D. Drechsel, M. Gorchtein, B. Pasquini, and M. Vanderhaeghen, hep-ph/9904290. J. Roche et al., “The first dedicated Virtual Compton Scattering experiment at MAMI”; forthcoming. T.R. Hemmert, B.R. Holstein, G. Kn[" o]{}chlein and S. Scherer, Phys. Rev. [**D55**]{}, 2630 (1997). T.R. Hemmert, B.R. Holstein, G. Kn[" o]{}chlein and S. Scherer, Phys. Rev. Lett. [**79**]{}, 22 (1997). T.R. Hemmert, B.R. Holstein, and J. Kambor, Phys. Lett. [**B395**]{}, 89 (1997). T.R. Hemmert, B.R. Holstein and J. Kambor, Phys. Rev. [**D55**]{}, 5598 (1997) T.R. Hemmert, B.R. Holstein, J. Kambor and G. Kn[" o]{}chlein, Phys. Rev. [**D57**]{} 5746, (1998). See, [*e.g.*]{}, Proc. Workshop on Virtual Compton Scattering VCS96, Clermont-Ferrand, 1996, Ed. V. Breton; P.A.M. Guichon and M. Vanderhaeghen, Prog. in Part. and Nucl. Phys. [**41**]{}, 125 (1998). J. Edelmann, N. Kaiser, G. Piller and W. Weise, Nucl. Phys. [**A641**]{}, 119 (1998); D. Drechsel, S.S. Kamalov, G. Krein, B. Pasquini, and L. Tiator, nucl-th/9907056, to be published in Nucl. Phys. [**A**]{}; X. Ji and J. Osborne, hep-ph/9905410. R.A. Berg and C.N. Lindner, Nucl. Phys. [**26**]{}, 259 (1961); H. Arenh[" o]{}vel and D. Drechsel, Nucl. Phys. [**A233**]{}, 153 (1974). P.A.M. Guichon, G.Q. Liu, and A.W. Thomas, Nucl. Phys. [**A591**]{}, 606 (1995) and Aust. J. Phys. [**49**]{}, 905 (1996). D. Drechsel, G. Kn[" o]{}chlein, A. Metz, and S. Scherer, Phys. Rev. [**C55**]{}, 424 (1997). D. Drechsel, G. Kn[" o]{}chlein, A. Yu. Korchin, A. Metz, and S. Scherer, Phys. Rev. [**C57**]{}, 941 (1998). A. Metz and D. Drechsel, Z. Phys. [**A356**]{}, 351 (1996) and [**A359**]{}, 165 (1997). M. Vanderhaeghen, Phys. Lett. [**B368**]{}, 13 (1996). M. Kim and D.-P. Min, hep-ph/9704381. V. Bernard, N. Kaiser, and U.-G. Mei[ß]{}ner, Int. J. Mod. Phys. [**E4**]{}, 193 (1995). V. Bernard, H.W. Fearing, T.R. Hemmert and U.-G. Mei[ß]{}ner, Nucl. Phys. [**A635**]{}, 121 (1998); (E) [**A642**]{}, 536 (1998). J. Gasser and H. Leutwyler, Ann. Phys. (NY) [**158**]{}, 142 (1984); Nucl. Phys. [**B250**]{}, 465 (1985). J. Wess and B. Zumino, Phys. Lett. [**B37**]{}, 95 (1971); E. Witten Nucl. Phys. [**B223**]{}, 422 (1983). V. Bernard, J. Kambor, N. Kaiser, and U.-G. Mei[ß]{}ner, Nucl. Phys. [**B388**]{}, 315 (1992). T.R. Hemmert, B.R. Holstein, and J. Kambor, J. Phys. [**G24**]{}, 1831 (1998). T.R. Hemmert, University of Massachusetts Ph.D. Thesis, Amherst 1997, UMI-98-09346-mc (microfiche). A. Manohar and H. Georgi, Nucl. Phys. [**B234**]{}, 189 (1984); J.F. Donoghue, E. Golowich, and B.R. Holstein, Phys. Rev. [**D30**]{}, 587 (1984). F.E. Low, Phys. Rev. [**96**]{}, 1428 (1954); M. Gell-Mann and M.L. Goldberger, Phys. Rev. [**96**]{}, 1433 (1954). S. Scherer, A.Yu. Korchin, and J.H. Koch, Phys. Rev. [**C54**]{}, 904 (1996). G. Kn[" o]{}chlein, Universität Mainz Ph.D. Thesis, Shaker Verlag (Aachen) 1997. C. Unkmeir, S. Scherer, A.I. L’vov, and D. Drechsel, hep-ph/9904442. See, [*e.g.*]{}, N.C. Mukhopadhyay, A.M. Nathan and L. Zhang, Phys. Rev. [**D47**]{}, R7 (1993). V. Bernard, N. Kaiser, A. Schmidt, and U.-G. Mei[ß]{}ner, Phys. Lett. [**B319**]{}, 269 (1993); Z. Phys. [**A348**]{}, 317 (1994). G. Gellas, T.R. Hemmert, C. Ktorides and U.-G. Mei[ß]{}ner, forthcoming. ---------------------------------- --------------- ------ ------ ------ ------ -- Quantity Expt. ChPT LSM ELM NRQM $P_{LL}-{1\over \epsilon}P_{TT}$ $30.5\pm 6.2$ 26.3 10.9 5.9 17.0 $P_{LT}$ $-8.6\pm 3.9$ -5.7 0 -1.9 -1.7 ---------------------------------- --------------- ------ ------ ------ ------ -- : Experimental values of the response functions measured at MAMI at $Q^2=0.33$ GeV$^2$ compared with predictions from chiral perturbation theory at ${\cal O}(p^3)$, the linear sigma model (LSM) of Metz and Drechsel [@Metz], the effective lagrangian model (ELM) of Vanderhaeghen [@Vanderhaeghen], and the nonrelativistic quark model (NRQM) of Guichon et al. [@Guichon]. This table is taken from ref.[@nstar]. [^1]: email: [email protected] [^2]: email: [email protected] [^3]: email: [email protected] [^4]: email: [email protected] [^5]: Results for the neutron extracted from $n-Pb$ scattering cross section measurements have been reported[@npol] but have been questioned[@ques]. Extraction via studies using a deuterium target may be possible in the future[@bean]. [^6]: Chiral analyses of double virtual Compton scattering $\gamma^\ast p\rightarrow \gamma^\ast p$ in the forward direction and its connection with the spin structure of the nucleon have recently been published [@dvcs]. [^7]: We note that to the order we are working in the VCS calculation the nucleon mass parameter $M_0$ can be replaced by the physical nucleon mass $M_N$, the axial-vector coupling in the chiral limit $\dot{g}_A$ can be replaced with the physical axial-vector coupling constant $g_A=1.267$ and the isoscalar \[isovector\] anomalous magnetic moment of the nucleon in the chiral limit $\dot{\kappa}_s\;[\dot{\kappa}_v]$ can be replaced with the physical isoscalar \[isovector\] anomalous magnetic moment $\kappa_s=\kappa_p+\kappa_n=-0.120\, \mbox{n.m.}\;[\kappa_v=\kappa_p-\kappa_n=3.71\,\mbox{n.m.}]$. Details of the renormalization of these parameters in the chiral lagrangian by loop effects and higher order counter terms can be found in ref.[@bfhm], both for HBChPT and SSE. [^8]: In order to take into account the isospin 3/2 property of the $\Delta$(1232) we supply the Rarita-Schwinger spinor with an additional isospin index $i$, subject to the subsidiary condition $\tau_i \; \psi_{\mu}^i (x) = 0$. [^9]: Note that these values are determined from the width expressions within the “small scale expansion" and therefore differ from those obtained in a relativistic analysis, [*e.g.*]{} see ref. [@delta]. [^10]: In fact, the primary source of information about the structure of the nucleon in the process $eN\rightarrow e'N\gamma$ comes from the interference between ${\cal M}^{VCS}$ and ${\cal M}^{Bethe-Heitler}$. [^11]: Our calculations are actually performed in the Coulomb gauge, see the discussion in appendix \[appnpi\]. [^12]: It is helpful to employ the identity given in appendix \[ident\] when reducing Pauli structures to the 12 structure amplitudes employed here. [^13]: The origin of this impediment lies in the fact that the quantity $\omega_0=M_N-\sqrt{M_{N}^2+\bar{q}^2}$ strictly speaking is suppressed by a factor of $1/M_N$ in both the chiral and small scale expansions. Full sensitivity to $\omega_0$ dependent quantities can therefore only be achieved in ${\cal O}(p^4)$, respectively ${\cal O}(\epsilon^4)$ calculations. [^14]: As expected, $\bar{\alpha}_E$ is completely free of delta pole contributions to this order, quite analogous to the case of real Compton scattering [@delta; @HHKK]. [^15]: We observe that there does exist a $\Delta$-pole contribution to the spin GP $\hat{P}_{(01,1)0}(\bar{q})$, $$\hat{P}_{(01,1)0}^{\Delta-{\rm pole}}=-{4\omega_0\over 27\bar{q}^2}{b_1^2\over M^2\Delta}\;,$$ which, however, is suppressed by an additional factor of $1/M_N$ originating in $\omega_0$ of Eq.(\[om0\]) and therefore is counted as a ${\cal O}(\epsilon^4)$ effect. [^16]: This is to be contrasted with the individual (real) Compton spin polarizabilities $\gamma_2$ and $\gamma_4$ defined by Ragusa, see ref.[@HHKK]. To be more specific about the connection between VCS and real Compton scattering we utilize Eq.(\[spinconnection\]) and find $$\begin{aligned} \gamma_3^{(III)}&=&1.0\times 10^{-4}\;\mbox{fm}^4 \nonumber \\ \gamma_2^{(III)}+\gamma_4^{(III)}&=&1.0\times 10^{-4}\;\mbox{fm}^4\; ,\end{aligned}$$ with the input from Eq.(\[SSEspinreihe\]). This is in complete agreement with the results of ref.[@HHKK]. It turns out that the large $\Delta$(1232) pole contribution cancels in this particular linear combination of $\gamma_2$ and $\gamma_4$. [^17]: We note that the theoretical predictions of the $\bar{q}$-dependence in the GPs given in refs.[@HHKS1; @HHKS2] [*preceded*]{} the analysis of the experiment at Mainz. [^18]: We do not give predictions for the response functions of ${\cal O}(\epsilon^3)$ SSE due to the discussed normalization problem in $\bar{\alpha}_E^{(III)}(\bar{q}=0)$. However, we believe that the ${\cal O}(\epsilon^3)$ SSE predictions for the spin GPs will be helpful for ongoing studies on double polarization VCS experiments, which might provide the possibility to study the connection between Ragusa and Guichon spin-polarizabilities as indicated by Eq.(\[spinconnection\]).
{ "pile_set_name": "ArXiv" }
--- abstract: 'A graph is a data structure composed of dots (i.e. vertices) and lines (i.e. edges). The dots and lines of a graph can be organized into intricate arrangements. The ability for a graph to denote objects and their relationships to one another allow for a surprisingly large number of things to be modeled as a graph. From the dependencies that link software packages to the wood beams that provide the framing to a house, most anything has a corresponding graph representation. However, just because it is possible to represent something as a graph does not necessarily mean that its graph representation will be useful. If a modeler can leverage the plethora of tools and algorithms that store and process graphs, then such a mapping is worthwhile. This article explores the world of graphs in computing and exposes situations in which graphical models are beneficial.' author: - 'Marko A. Rodriguez$^1$' - Peter Neubauer$^2$ bibliography: - '../marko.bib' title: Constructions from Dots and Lines --- =1 The Bits and Pieces of the Dots and Lines ========================================= A model is a representation of some aspect of reality. Many models can be thought of as a collection of objects (e.g. people, concepts) and the relationships that exist between them (e.g. friendships, subclasses). Such objects and relations form a network. Graphically, an object in a network can be denoted by a dot and a relationship can be denoted by a line. A structure formed by dots and lines is known as a graph—the mathematical term for a network [@intrograph:trudeau1976]. The most common type of graph is the simple graph. An example instance is diagrammed in Figure \[fig:graph-example\]. In a simple graph there are a set of vertices (i.e. dots) and a set of edges (i.e. lines), where edges are undirected, connect two unique vertices (i.e. no loops), and no two edges exist between the same pair of vertices. ![image](images/graph-example){width="62.50000%"} Contrary to the title of this article, dots and lines are not the only components in a graph modeler’s toolkit. There are many more bits and pieces in the world of graphs. In practice, rarely are vertices and edges the only data contained within a graph. For instance, sometimes its useful to have a name associated with a vertex, a weight and direction associated with an edge, etc. From primitive dots and lines various bits and pieces can be added to yield a more flexible, more expressive graph. Figure \[fig:graph-types\] diagrams a collection of different graph types. A short summary of each graph type is provided in the itemization below. Note that in many cases, these bits and pieces can be used in combination with one another (i.e. they are not necessarily mutually exclusive). ![image](images/graph-types){width="95.00000%"} - **half-edge graph**: a unary edge (i.e. an edge that “connects" one vertex) has limited practical application and is primarily discussed in mathematics. - **multi-graph**: there are many situations in which it is desirable to have multiple edges between the same two vertices. - **simple graph**: the prototypical graph, where an edge connects two vertices and no loops are allowed. - **weighted graph**: used to represent strength of ties or transition probabilities. - **vertex-labeled graph**: most every graph makes use of labeled vertices (e.g. an identifier). - **semantic graph**: used to model cognitive structures such as the relationship between concepts and the instances of those concepts [@sowa:semantic1991]. - **vertex-attributed**: used in applications where it is desirable to append non-relational metadata to a vertex. - **edge-labeled graph**: used to denote the way in which two vertices are related (e.g. friendships, kinships, etc.). - **directed graph**: orders the vertices of an edge to denote edge orientation. - **hypergraph**: generalizes a binary edge whereby an edge connects an arbitrary number of vertices [@hyper:gallo1993]. - **undirected graph**: the typical graph that is used when the relationship is symmetric (e.g. friendship). - **resource description framework graph**: a graph standard developed by the the World Wide Web consortium that denotes vertices and edges by Uniform Resource Identifiers [@rdfintro:miller1998]. - **edge-attributed graph**: used in applications where it is desirable to append non-relational metadata to an edge. - **pseudo graph**: used to denote a reflexive relationship. The list presented is not the complete space of all graph types, nor are the terms generally accepted in all domains. Many of these structures have been rediscovered in different domains and under different names. The important point is that there are numerous graph types and, consequently, there are systems and algorithms that exist to store and process them. A common graph type supported by most graph systems is the directed, labeled, attributed, multi-graph—also known as a “property graph." Graphs of this form allow for the representation of labeled vertices, labeled edges, and attribute metadata (i.e. properties) for both vertices and edges. The property graph is common because by simply abandoning or adding particular bits and pieces, other graph types can be expressed. For example, by not allowing loops or multiple edges between two vertices, a simple graph is generated. By not allowing vertex/edge attributes, a standard semantic graph is generated. By restricting the vertex/edge labels to Uniform Resource Identifiers (URIs), a Resource Description Framework (RDF) graph is generated.[^1] By adding a weight attribute to an edge, a weighted graph is generated. The various graph types and the morphisms that yield one graph type from another are diagrammed in Figure \[fig:graph-types-morphisms\]. Note the location of the property graph within this diagram. Finally, while it is possible to model a hypergraph in a property graph, it comes at the expense of using vertices in the property graph to denote both vertices and edges in the hypergraph. For this reason, there exist specialized hypergraph systems.[^2] For the remainder of this article, the more common property graph and its supporting technologies are discussed. ![image](images/graph-types-morphisms){width="85.00000%"} Preserving Dots and Lines ========================= The computer science community has recently seen an explosion of database technologies. For decades, the relational database of Codd’s relational algebra has been the primary storage and query mechanism for large data sets [@rdbms:codd1970]. However, with the continued growth of data and an increasingly variegated application landscape, new databases have emerged. In this space, no database is seen as the single solution to all problems. Instead, each database attempts to solve a particular data management issue. Itemized below is a short description of recent database types. - **document database**: These databases have the “document" as their atomic entity. Such objects are semi-structured and usually represented in XML or JSON. A document can be retrieved by means of pattern matching a query document (i.e. a semi-populated document) against all the documents contained in the database. The benefit of this model is that these databases scale horizontally with relative ease. This is due to the fact that documents lack references between one another. The drawback is that data is not interrelated and thus, cross database analyses are costly. For many web applications the document databases is a well-suited solution that supports data scale and a convenient symmetry between the document structure and the processing language (e.g. languages that natively support XML and/or JSON). Examples of such databases include MongoDB[^3] and CouchDB[^4]. - **key/value store**: This family of databases is focused on the scaling of large amounts of data over a large number of machines and, in turn, supporting heavy read/write loads. Most of the databases in this class were inspired by Amazon’s Dynamo [@dynamo:decandia2007]. A popular open-source key/value store is Tokyo Cabinet.[^5] - **triple/quad store**: Triple/quad-stores were developed to support the demands of the Semantic Web/Web of Data/Linked Data community. These databases are optimized for storing and querying data represented according to the Resource Description Framework (RDF) [@rdfintro:miller1998]. Typical use cases include description logic reasoning [@baader:dl2003] and SPARQL-based graph pattern matching [@sparql:prud2004]. AllegroGraph is a high-performance quad-store with a large suite of extensions and features.[^6] - **column store**: Most column stores are modeled after Google’s BigTable database [@bigtable:chang2006]. A big table is a sparse, distributed, persistent multi-dimensional sorted map. The map is indexed by a row key, column key, and a time-stamp. Real-world services implemented with BigTable include GoogleAnalytics and GoogleEarth. Cassandra is a popular open-source column store.[^7] - **graph database**: Graph databases are optimized for the efficient processing of dense, interrelated datasets. In these databases, the atomic entity is the graph as a whole. The typical data model is the property graph. By supporting the interrelation of data, graph databases allow for fast traversals along the edges between vertices [@traversal:rodriguez2010]. A popular graph database of this form is Neo4j.[^8] There are numerous databases in this growing space that were not mentioned. Moreover, there are other databases types not mentioned. It is out of the scope of this article to explore this space in depth. The interested reader is directed to related discussions, blog posts, and presentations that are made freely available on the Internet. Of particular relevance to this article is the graph database and the property graph data model. Figure \[fig:property-graph-example\] diagrams a property graph containing people, their articles, and a university. In this particular domain model, each vertex has a [name]{} property and a [type]{} property. Edges denote both a directionality and a relationship type (i.e. an edge label). Moreover, its possible to also include properties on an edge to further refine the way in which two vertices are related (e.g. Josh started attending RPI in 2007). ![image](images/property-graph-example){width="82.00000%"} A consequence of the flexibility of a graph is that other graph structures can be represented along with the domain model. A typical use case of such graph extensions include endogenous indices. An index is usually a tree-structure that allows for the fast look-up of elements within a collection. If there were no indices into a collection, then to determine if a particular element had a particular property, each element in the collection would have to be examined. The cost of a linear scan of this kind is ${\mathcal}{O}(n)$, where $n$ is the number of elements. What an index provides is the ability to partition the elements into increasingly fine-grained bins. Most indices have a lookup cost of ${\mathcal}{O}(\text{log}_2 \, n)$. While an index creates more data (the tree structure), it makes up for this cost by greatly increasing the speed of element retrieval. Figure \[fig:index-example\] demonstrates a [name]{}-property index over the example graph diagrammed in Figure \[fig:property-graph-example\]. Together, the domain model and the index of the domain model are seen as a single atomic entity. Searching for an element and moving between elements are accomplished by a unified framework: the graph traversal. ![image](images/index-example){width="100.00000%"} Jumping from Dot to Dot ======================= The first aspect of using a graph is creating a graph. Once a graph has been created, it can be subjected to algorithms that quantify aspects of its structure, alter its structure, or solve-problems that are a function of its structure. At the root of any of these algorithms is the graph traversal [@traversal:rodriguez2010]. A graph traversal is a “walk" along the elements of a graph—from vertex, to edge, to vertex, etc. As this walk proceeds, aspects of the graph can be saved or manipulated and in general, an algorithm can be computed. In principle, any of the data models and databases presented in the previous section (and including typical relational databases) can be used to represent and process a graph. However, when traversing a graph is the ultimate use case for a graph data set, then a graph databases is the optimal solution.[^9] To get a better understanding of how graph traversals work, the examples in this section will be expressed in terms of a graph programming language called Gremlin.[^10] In Gremlin, moving over vertices and edges is analogous, in many ways, to moving through the directory structure of a local filesystem. To demonstrate, a naïve friend-of-a-friend query is represented as follows: ./outE[@label=`friend']/inV/outE[@label=`friend']/inV Reading from left to right, this expression states: - Start at the root vertex ([.]{}, i.e. the vertex to evaluate the expression on). - Traverse to all the outgoing edges of the root vertex ([/outE]{}). - Filter out all edges that are not labeled “friend" ([\[@label=‘friend’\]]{}). - For all those friend-labeled edges, go to their incoming/head vertices ([/inV]{}). - For all the friends of the root vertex, get their outgoing edges ([/outE]{}). - Filter out all edges that are not labeled “friend" ([\[@label=‘friend’\]]{}). - For all those friend-labeled edges, go to their incoming/head vertices ([/inV]{}). At the end of this expression, the resultant vertices are the friends of the friends of the root vertex. Figure \[fig:foaf-example\]a diagrams the traversal, where the grey vertices are the returned vertices. This example is “naïve" because in many cases, its important to retrieve the root vertex’s friends of friends that are not also its friends. In such situations, the traverser must remember if a located friend-of-friend was not already a friend. In order to calculate the friend-of-a-friend, the friends must be determined first. Therefore, its possible to save this information for later use. This idea is diagrammed in Figure \[fig:foaf-example\]b and the Gremlin expression is presented below, where the variable [\$x]{} references the friends of the root vertex. ./outE[@label=`friend']/inV[g:assign(`$x')]/ outE[@label=`friend']/inV[g:except($x)] ![image](images/foaf-example){width="95.00000%"} An important aspect of working with property graphs is that the edges are typed/labeled. The standard suite of graph algorithms found in most graph/network-theory textbooks are not immediately useful for property graphs [@netanal:brandes2005]. This is because, most graph algorithms have been developed for unlabeled graphs. When vertices can be related by many different ways and vertices can represent various types of objects, the meaning of the rankings, paths, etc. returned by standard graph algorithms are ambiguous. However, by interpreting a path through a graph as an edge, its possible to express standard graph algorithms on property graphs [@pathalg:rodriguez2009]. The previously presented Gremlin expression followed a path from the root vertex to its friends’ friends. This path can be considered a “virtual" (i.e. inferred, derived) edge. From the perspective of this expression, a new implicit graph is created over the graph’s vertices that only contains edges labeled “friend-of-a-friend." This idea is diagrammed in Figure \[fig:foaf-virtual-example\]. As such, this “virtual" graph is equivalent to an unlabeled graph because all edges having the same meaning. Therefore, all the standard graph algorithms can be meaningfully applied to this derived graph—e.g. the shortest path between person $A$ and person $B$ through their friends of friends. The benefit of edge-labeled graphs (e.g. property graphs) is that there are as many types of rankings, scorings, etc. as there are types of paths that exist between the elements of the graph. ![image](images/foaf-virtual-example){width="45.00000%"} Conclusion ========== The concept of a graph was introduced in the late 19$^\text{th}$ century. During the many decades that followed, the world of graphs was primarily left to the toiling of mathematicians. In the last few decades, the sociology, physics, and computer science communities introduced a suite of algorithms and insightful realizations about the nature of graphs found in the real-world. Moreover, the increasingly large volume of data made available by the Internet has yielded datasets that reflect the graphs found in our technological and social systems. To satiate the need to handle and process these large-scale graphs, graph databases have come to the forefront. To make use of the graphs beyond simply representing their explicit structure, graph traversal frameworks and algorithms have been developed in order to shape graphs by driving the evolution of the entities that they model—e.g. humans and their relationships to one another and the objects of their world [@faith2:rodriguez2009]. [^1]: This is not completely true as an RDF graph makes use of URIs, literals, and blank/anonymous nodes. The distinction between these concepts are outside the scope of this article. [^2]: HyperGraphDB is an example hypergraph database that is available at: [http://www.kobrix.com/hgdb.jsp]{}. [^3]: MongoDB is available at [http://www.mongodb.org/]{}. [^4]: CouchDB is available at [http://couchdb.apache.org/]{}. [^5]: Tokyo Cabinet is available at [http://1978th.net/tokyocabinet/]{}. [^6]: AllegroGraph is available at [http://www.franz.com/agraph/allegrograph/]{}. [^7]: Cassandra is available at [http://cassandra.apache.org/]{}. [^8]: Neo4j is available at [http://neo4j.org/]{}. [^9]: This is an import point. A graph database is optimized for graph traversals because elements (i.e. vertices and edges) maintain direct references to their adjacent elements. It is this design choice that makes traversing a graph structure within a graph database fast and efficient. [^10]: Gremlin is available at [http://gremlin.tinkerpop.com/]{}.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We have studied the scaling properties of the electromagnetic response functions of $^4$He and $^{12}$C nuclei computed by the Green’s Function Monte Carlo approach, retaining only the one-body current contribution. Longitudinal and transverse scaling functions have been obtained in the relativistic and non relativistic cases and compared to experiment for various kinematics. The characteristic asymmetric shape of the scaling function exhibited by data emerges in the calculations in spite of the non relativistic nature of the model. The results are consistent with scaling of zeroth, first and second kinds. Our analysis reveals a direct correspondence between the scaling and the nucleon-density response functions.' author: - 'N. Rocco' - 'L. Alvarez-Ruso' - 'A. Lovato' - 'J. Nieves' bibliography: - 'biblio.bib' title: 'Electromagnetic scaling functions within the Green’s Function Monte Carlo approach' --- Introduction ============ A realistic description of the electromagnetic response of atomic nuclei is a challenging many-body problem as it requires an accurate understanding of both the nuclear dynamics and of the interaction vertex. In this regard a valuable strategy consists in analyzing the scaling properties of nuclear response functions in a variety of kinematic setups [@West:1974ua; @Day:1987az; @Donnelly:1998xg]. Scaling of the first kind is said to occur when the electron-nucleus cross section or longitudinal/transverse response functions, divided by an appropriate function describing the single-nucleon physics, do no longer depend on two variables (for example energy transfer $\omega$ and absolute value of the 3-momentum transfer ${|\bf q|}$ in the Laboratory frame), but only upon a specific function of them, which defines the scaling variable. Scaling of the second kind takes place when there is no dependence on the nuclear species. Finally, the simultaneous occurrence of both kinds of scaling is denoted as superscaling [@Alberico:1988bv]. Superscaling is exactly fulfilled by the Global Relativistic Fermi gas (GRFG) model, for which a simple and symmetric scaling function can be derived in terms of the dimensionless scaling variable $\psi$ [@Barbaro:1998gu] (explicit expressions are provided in Sec. \[GRFG\] below). However, contrary to the GRFG model predictions, the results extracted from experimental data reveal an asymmetric shape of the scaling function, with a tail that extends to high values of $\psi$ (and $\omega$) [@Amaro:2004bs]. These results represent a strong constraint for theoretical models of electron scattering reactions. Extensive studies with a large variety of models reveal the importance of a proper description of the interaction of knocked-out nucleons with the residual nucleus—final state interactions (FSI)—to obtain the tail of the scaling function [@Caballero:2005sj; @Caballero:2006wi; @Caballero:2007tz; @Meucci:2009nm; @Antonov:2011bi]. The authors of Refs. [@Caballero:2005sj; @Caballero:2006wi] argue that, while this asymmetry in the scaling function is largely absent in non-relativistic mean-field models, it can be recovered within the relativistic impulse approximation, given that FSI are described using a strong relativistic mean field (RMF) potential. Asymmetric scaling functions also emerge in semi-relativistic models when FSI are described by local potentials derived from the RMF one [@Caballero:2007tz]. On the other hand, the comparison between semi-relativistic and relativistic results shows a breakdown of the zeroth-kind scaling, [i.e.]{} different scaling functions in the longitudinal and transverse channel, only when the fully relativistic mean field approach is employed. According to Ref. [@Caballero:2007tz] this effect has been ascribed to the dynamical enhancement of the lower component of the Dirac spinors, which are not present in the semi-relativistic approach. In this work we analyze the scaling properties exhibited by Green’s Function Monte Carlo (GFMC). GFMC is an [*ab initio*]{} method allowing for a very accurate description of the properties of $A\leq 12$ nuclei, in which the dynamics of constituent nucleons are fully considered [@Lovato:2013cua; @Lovato:2014eva; @Lovato:2015qka]. The longitudinal and transverse electromagnetic response functions of $^{12}$C, recently computed within GFMC turn out to be in very good agreement with experiment, when two-body currents are accounted for [@Lovato:2016gkq]. Despite this remarkable result, GFMC is currently limited to $^{12}$C because of the exponentially growing cost of the calculation with the number of nucleons. In addition to that, the inclusion of relativistic kinematic and baryon resonance production would involve non trivial difficulties. The study of the behavior of the scaling functions obtained from the GFMC calculations, while being interesting in its own right, is aimed at elucidating the role of initial and final state correlations in the asymmetric shape of the scaling function. In Section \[scaling\] we review the derivation of the electron-nucleus cross section, as well as its expression in terms of longitudinal and transverse response functions, which are necessary to introduce the concept of scaling. In Section \[GFMC\], the main elements of the Green’s Function Monte Carlo approach are briefly outlined, while in Section \[GRFG\] we explicitly derive the expression of the longitudinal and transverse scaling functions in the context of the GRFG model, both in the relativistic and non relativistic cases. In Section \[results\] we report the results of our analysis of the scaling features of the GFMC response functions for $^4$He and $^{12}$C nuclei and in different kinematics. We then discuss a novel interpretation of the longitudinal and transverse scaling function in terms of the nucleon-density function. Finally, in Section \[conclusion\] we summarize our findings and state the conclusions. Scaling of the nuclear electromagnetic response within the Green’s Function Monte Carlo approach {#scaling} ================================================================================================ In the one-photon-exchange approximation, the double differential electron-nucleus cross section can be written in the form \[xsec\] =L\_W\^  , where $k_e=(E_e,{\bf k}_e)$ and $k_{e^\prime}=(E_{e^\prime},{\bf k}_{e^\prime})$ are the laboratory four-momenta of the incoming and outgoing electrons, respectively; $\alpha \simeq 1/137$ is the fine structure constant, $d\Omega_{e^\prime}$, the differential solid angle in the direction of ${\bf k}_{e^\prime}$, and $q=k_e - k_{e^\prime} =(\omega,{\bf q})$ the four momentum transfer. The leptonic tensor is given by $$\begin{aligned} L^{\mu\nu}=2 \left( k_{e^\prime}^\mu k_e^\nu+ k_e^\mu k_{e^\prime}^\nu- g^{\mu\nu}k_{e^\prime}\cdot k_e \right)\,. \end{aligned}$$ The hadronic tensor encompasses the electromagnetic transitions from the target nucleus to all possible final states. It is thus given by $$\begin{aligned} \label{response:tensor} W^{\mu \nu} =\sum_f \langle 0| {J^\mu}^\dagger(q) | f \rangle \langle f | J^\nu(q) | 0 \rangle \, \delta^{(4)}(P_0+q-P_f) \ ,\end{aligned}$$ where $| 0 \rangle$ and $| f \rangle$ denote the initial and final hadronic states with four-momenta $P_0 = ( E_0,{\bf p}_0 )$ and $P_f = (E_f,{\bf p}_f) $, while $J(q)$ is the electromagnetic nuclear current operator. Equation can be rewritten in terms of two response functions, denoted by $R_L({\bf q}, \omega)$ and $R_T({\bf q},\omega)$, describing interactions with longitudinally (L) and transversely (T) polarized photons, respectively. The resulting expression reads $$\begin{aligned} \frac{d^2\sigma}{d E_{e^\prime} d\Omega_{e^\prime}} & =\left( \frac{d \sigma}{d\Omega_{e^\prime}} \right)_{\rm{M}} \Big[ A_L(|{\bf q}|,\omega,\theta_{e^\prime}) R_L(|{\bf q}|,\omega) \nonumber \\ & + A_T(|{\bf q}|,\omega,\theta_{e^\prime}) R_T(|{\bf q}|,\omega) \Big] \ ,\end{aligned}$$ where $$\begin{aligned} A_L = \Big( \frac{q^2}{{\bf q}^2}\Big)^2 \ \ \ , \ \ \ A_T = -\frac{1}{2}\frac{q^2}{{\bf q}^2}+\tan^2\frac{\theta_e}{2} \ , \end{aligned}$$ and $$\begin{aligned} \label{Mott} \left( \frac{d \sigma}{d \Omega_{e^\prime}} \right)_{\rm{M}}= \left[ \frac{\alpha \cos(\theta_{e^\prime}/2)}{2 E_{e^\prime}\sin^2(\theta_{e^\prime}/2) }\right]^2\end{aligned}$$ is the Mott cross section. The L and T response functions can be readily expressed in terms of specific components of the hadronic tensor. Choosing the $z$-axis along the direction of the momentum transfer one finds $$\begin{aligned} \label{RL} R_L & = W^{00}\ ,\\ \label{RT} R_T &= \sum_{ij=1}^3\Big(\delta_{ij}-\frac{q_iq_j}{{\bf q}^2}\Big)W^{ij} \ . \end{aligned}$$ The Green’s function Monte Carlo approach {#GFMC} ----------------------------------------- GFMC provides a suitable framework to carry out accurate calculations of a variety of nuclear properties in the non relativistic regime, typically corresponding to $|{\bf q}| \lsim 500 \ {\rm MeV}$ (for a recent review of Quantum Monte Carlo methods for nuclear physics see, e.g., Ref. [@Carlson:2014vla]). The longitudinal and transverse response function are given by $$\begin{aligned} R_L({\bf q}, \omega)&= \sum_f \la 0|\rho^\dagger({\bf q})|f\ra\la f|\rho({\bf q})|0\ra\delta(\omega+E_0-E_f)\,, \nonumber\\ R_T({\bf q}, \omega)&= \sum_f \la 0|{\bf j}_T^\dagger({\bf q})|f\ra\la f|{\bf j}_T({\bf q})|0\ra\delta(\omega+E_0-E_f) \,, \label{resp:GFMC} \end{aligned}$$ where $\rho({\bf q})$ and ${\bf j}_T({\bf q})$ denote non-relativistic reductions of the nuclear-charge and transverse-current operators, respectively [@Carlson:2001mp]. Valuable information on the L and T responses can be obtained from their Laplace transforms, also referred to as Euclidean responses \_[T,L]{}([**q**]{}, )= \_[\_[[el]{}]{}]{}\^[d]{} e\^[-]{}R\_[T,L]{}([**q**]{}, ) . The lower integration limit $\omega_{\rm{el}}= {\bf q}^2/2M_A$, $M_A$ being the mass of the target nucleus, is the elastic scattering threshold—corresponding to the $|f \rangle = |0 \rangle$ term in the sum of Eq. —whose contribution is excluded. Within GFMC, the Euclidean responses are evaluated from $$\begin{aligned} \nonumber \widetilde{E}_L({\bf q},\tau) & = \langle 0| \rho^\dagger({\bf q}) e^{-(H-E_0)\tau} \rho({\bf q})|0\rangle \\ & - |\langle 0 | \rho({\bf q}) | 0 \rangle|^2 e^{-\omega_{\rm el} \tau} \ , \label{eq:eucL_mat_el}\end{aligned}$$ and $$\begin{aligned} \nonumber \widetilde{E}_T({\bf q},\tau) & = \langle 0| {\bf j}_T^\dagger({\bf q}) e^{-(H-E_0)\tau} {\bf j}_T({\bf q})|0\rangle \\ & - |\langle 0 | {\bf j}_T({\bf q}) | 0 \rangle|^2 e^{-\omega_{\rm el} \tau} \ . \label{eq:eucT_mat_el}\end{aligned}$$ Note that, although the states $|f \rangle \neq | 0 \rangle$ do not appear explicitly in Eqs. and , the Euclidean responses include the FSI effects of the particles involved in the electromagnetic interaction, both among themselves and with the spectator nucleons. The inversion of the Laplace transform, needed to retrieve the energy dependence of the responses, is long known to involve severe difficulties. However, maximum-entropy techniques, based on Bayesian inference arguments, have been successfully exploited to perform accurate inversions, supplemented by reliable estimates of the theoretical uncertainty. In the case of $^{12}$C, particular care has to be devoted to the subtraction of contributions arising from elastic scattering and the transitions to the low-lying $2^+$, $0^+_2$, and $4^+$ states [@Lovato:2015qka]. Scaling within the relativistic Fermi gas model {#GRFG} ----------------------------------------------- The easiest, albeit quite crude approximation, to describe the hadron tensor consists on using the GRFG model. Within this approach the scattering process is assumed to take place on a single nucleon with four-momentum $p=(E({\bf p}),{\bf p})$, where $E({\bf p})=\sqrt{|{\bf p}|^2+m^2}$, $m$ being the nucleon mass. The requirement that the struck nucleon is in the target nucleus implies that $|{\bf p}|$ is smaller than the Fermi momentum $p_F$. Furthermore, the outgoing nucleon with four-momentum ${p^\prime}^\mu=(p+q)^\mu$ should lay above the Fermi surface. The expression of the hadron tensor describing the response of the target nucleus then reads $$\begin{aligned} W^{\mu\nu}=& \frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\frac{m^2}{E({\bf p})E({\bf p+q})}\ w^{\mu\nu}(p+q,p)\nonumber\\ &\times \theta(p_F-|{\bf p}|)\theta(|{\bf p+q}|-p_F)\nonumber\\ &\times \delta(\omega+E({\bf p})-E({\bf p+q}))\ . \label{had:tens:FG}\end{aligned}$$ Once we only discuss symmetric nuclei, $\mathcal{N}$ denotes both the number of protons and neutrons in the nucleus. The single-nucleon response tensor $w^{\mu\nu}(p+q,p)$ encodes the response of a system in which a nucleon with 4-momentum $p$ in the initial state is scattered by a (virtual) photon, leading to a final state with a nucleon carrying a 4-momentum $(p+q)$. The following general expression $$\begin{aligned} \label{w12} w^{\mu\nu}(p+q,p)=& - W_1(\tau)\Big(g^{\mu\nu}-\frac{q^\mu q^\nu}{q^2}\Big)\nonumber\\ &+ W_2(\tau)\frac{1}{m^2}\Big(p^\mu-\frac{p\cdot q}{q^2}q^\mu\Big)\nonumber\\ &\times \Big( p^\nu -\frac{p\cdot q}{q^2}q^\nu\Big)\ ,\end{aligned}$$ where $\tau= -q^2/4m^2= Q^2/4m^2\geq0$, holds. It is well known that the nucleon structure functions $W_{1,2}$ can be written in terms of the proton and neutron electric and magnetic form factors as $$\begin{aligned} W_1(\tau)=& \tau G^2_M(\tau)\ ,\nonumber\\ W_2(\tau)=& \frac{G_E^2(\tau)+\tau G_M^2(\tau)}{(1+\tau)}\ ,\end{aligned}$$ and $$\begin{aligned} G_E(\tau)&= G^p_{E}(\tau)\frac{1}{2}(1+\tau_{z,i})+ G_E^n(\tau)\frac{1}{2}(1-\tau_{z,i})\ ,\nonumber\\ G_M(\tau) &= G_M^p(\tau)\frac{1}{2}(1+\tau_{z,i})+ G_M^n(\tau)\frac{1}{2}(1-\tau_{z,i})\ ,\end{aligned}$$ where $\tau_{z,p/n} = \pm 1$. Using the GRFG model to parametrize the nuclear amplitudes, the integral entering Eq.  can be analytically solved. We start by evaluating the function $$\begin{aligned} F(p_F,q)= & \frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}(p_F,q,{\bf p})\ , \label{f_scal}\end{aligned}$$ with $$\begin{aligned} \mathcal{F}(p_F,q,{\bf p})=&\frac{m^2}{E({\bf p})E({\bf p+q})}\nonumber\\ &\times \theta(p_F-|{\bf p}|)\theta(|{\bf p+q}|-p_F)\nonumber\\ &\times \delta(\omega+E({\bf p})-E({\bf p+q}))\, \label{eq:scale_integrand} \end{aligned}$$ resulting in [@Alberico:1988bv; @Donnelly:1991qy] $$\begin{aligned} \mathcal{F}(p_F,q,{\bf p}) =\frac{3\mathcal{N}m^2}{2p_F^3|{\bf q}|}\theta(E_F-\Gamma)(E_F-\Gamma)\ ,\end{aligned}$$ where we have introduced $E_F=\sqrt{p_F^2+m^2}$ and $$\begin{aligned} \Gamma&=\rm{Max}\{\Gamma_1,\Gamma_2,\Gamma_3\}\nonumber\\ &=\rm{Max}\Big\{m, E_F-\omega,\frac{-\omega+|{\bf q}|\sqrt{1+1/\tau}}{2}\Big\}\ . \end{aligned}$$ It is convenient to introduce the widespread set of dimensionless variables [@Alberico:1988bv] $$\begin{aligned} \lambda=\omega/2m\ ,\nonumber\\ \kappa=|{\bf q}|/2m\ ,\nonumber\\ \eta_F=p_F/m\ .\end{aligned}$$ The minimum $\Gamma_3/m=1$ at $$\begin{aligned} \lambda=\lambda^0=\frac{1}{2}\Big[ \sqrt{(1+4 \kappa^2)}-1\Big]\ ,\end{aligned}$$ corresponds to the quasi elastic peak $\tau = \lambda$ [@Alberico:1988bv]. In the limit of large $|{\bf q}|$, the relation $\Gamma=\Gamma_3$ is satisfied for each value of $\omega$. Hence, a dimensionless scaling variable can be defined in terms of this quantity as [@Alberico:1988bv] $$\begin{aligned} \psi=sign(\lambda-\lambda^0)\Big[\frac{1}{\xi_F}\Big(\frac{\Gamma_3}{m}-1\Big)\Big]^{1/2}\ ,\end{aligned}$$ with $\xi_F= E_F/m -1$ and such that $\psi=0$ at the quasi elastic peak. Note that this definition of the scaling variable is equivalent to the more common expression $$\begin{aligned} \psi= \frac{1}{\sqrt{\xi_F}}\frac{\lambda-\tau}{\sqrt{(1+\lambda)\tau+\kappa\sqrt{\tau(1+\tau)}}} \,.\end{aligned}$$ Collecting previous results one obtains $$\begin{aligned} F(p_F,q)=\frac{3\mathcal{N}\xi_F}{4 \eta_F^3 m \kappa}\big(1-\psi^2)\theta(1-\psi^2)\ .\end{aligned}$$ Substituting Eq. and into Eqs. , leads to the following expressions for the L and T response functions $$\begin{aligned} R_L=&\frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}(p_F,q,{\bf p})\Big\{ -W_1(\tau)\Big(1-\frac{{\omega}^2}{q^2}\Big)\nonumber\\ &+\frac{W_2}{m^2}\Big[ E_p - \frac{p\cdot q}{q^2}\omega\Big]^2\Big\}\ ,\nonumber\\ R_T=&\frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}(p_F,q,{\bf p})\Big\{2 W_1(\tau)+\frac{W_2(\tau)}{m^2}{\bf p}_T^2\Big\}\ .\end{aligned}$$ After performing the integrations, the responses can be cast in the form $$\begin{aligned} R_L=& \frac{3\mathcal{N}\xi_F}{4 \eta_F^3 m \kappa}\big(1-\psi^2)\theta(1-\psi^2)\nonumber\\ &\times \Big\{\frac{\kappa^2}{\tau}[G_E^2(\tau)+W_2(\tau)\Delta ]\Big\}\ ,\nonumber\\ R_T=& \frac{3\mathcal{N}\xi_F}{4 \eta_F^3 m \kappa}\big(1-\psi^2)\theta(1-\psi^2)\nonumber\\ &\times \Big\{ 2\tau G^2_M(\tau)+ W_2(\tau)\Delta\Big\}\ ,\end{aligned}$$ where $$\begin{aligned} \Delta= \xi_F(1-\psi^2)\Big[\frac{\sqrt{\tau(1+\tau)}}{\kappa}+\xi_F(1-\psi^2)\frac{\tau}{3\kappa^2}\Big]\ .\end{aligned}$$ The next step consists in the definition of the longitudinal and transverse scaling functions [@Maieron:2001it] $$\begin{aligned} f_L(\psi)= p_F\times \frac{R_L}{G_L}\ ,\nonumber\\ f_T(\psi)= p_F\times \frac{R_T}{G_T}\ ,\end{aligned}$$ where $$\begin{aligned} G_L=\frac{\mathcal{N}}{2\kappa} \Big\{\frac{\kappa^2}{\tau}[G_E^2(\tau)+W_2(\tau)\Delta ]\Big\}\nonumber\\ G_T=\frac{\mathcal{N}}{2\kappa}\Big\{ 2\tau G^2_M(\tau)+ W_2(\tau)\Delta\Big\}\ . \label{FG:pre:fact}\end{aligned}$$ Within the GRFG the same scaling function for the the longitudinal and transverse channel arises. This is a symmetric function centered in $\psi=0$ $$\begin{aligned} f(\psi)=f_L(\psi)= f_T(\psi)=\frac{3\xi_F}{2 \eta_F^2}\big(1-\psi^2)\theta(1-\psi^2)\ .\end{aligned}$$ In the non relativistic limit the L and T responses can be expressed as $$\begin{aligned} R_L=& \frac{3\mathcal{N}}{4\pi p_F^3}\int d^3p \frac{1}{2}\sum_{s,s^\prime}\Big\{ \chi^\dagger_s \rho^\dagger({\bf q})\chi_{s^\prime}\chi^\dagger_{s^\prime}\rho({\bf q})\chi_s\Big\}\nonumber\\ &\times \theta(p_F-|{\bf p}|)\theta(|{\bf p+q}|-p_F)\nonumber\\ &\times \delta\Big(\omega +\frac{{\bf p}^2}{2m}-\frac{|{\bf p+q}|^2}{2m}\Big)\ ,\nonumber\\ R_T=& \frac{3\mathcal{N}}{4\pi p_F^3}\int d^3p \frac{1}{2}\sum_{s,s^\prime}\Big\{ \chi^\dagger_s {\bf j}_T^\dagger({\bf q})\chi_{s^\prime}\chi^\dagger_{s^\prime}{\bf j}_T({\bf q})\chi_s\Big\}\nonumber\\ &\times \theta(p_F-|{\bf p}|)\theta(|{\bf p+q}|-p_F)\nonumber\\ &\times \delta\Big(\omega +\frac{{\bf p}^2}{2m}-\frac{|{\bf p+q}|^2}{2m}\Big)\ , \end{aligned}$$ where $s$ and $s^\prime$ are the spin quantum numbers of the nucleon in the initial and final state, respectively. In the following, non relativistic scaling variable and functions are introduced with the same non relativistic reduction of the current operator and relativistic corrections as in the GFMC calculations [@Carlson:2001mp]. Neglecting the very small spin-orbit relativistic correction in the definition of charge operator, the charge and current operators read $$\begin{aligned} \rho({\bf q})=&\frac{G_E(\tau)}{\sqrt{1+\tau}}\ ,\nonumber\\ {\bf j}_T({\bf q})=& \Big[ \frac{G_E(\tau)}{m}{\bf p}_T-i \frac{G_M(\tau)}{2m}{\bf q}\times {\bm \sigma}\Big]\ .\end{aligned}$$ As opposed to the semi relativistic model of Ref. [@Amaro:2006if], in the GFMC relativistic corrections enter only in the current definition, while the kinematics is fully non relativistic.\ In the non relativistic limit, Eq.  reduces to $$\begin{aligned} F^{nr}(p_F,q)= & \frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}^{nr}(p_F,q,{\bf p})\nonumber\\ = &\frac{3\mathcal{N}m^2}{2p_F^3|{\bf q}|}\theta(E^{nr}_F-\Gamma)(E^{nr}_F-\Gamma^{nr})\ , \label{f_scal_nr} \end{aligned}$$ with $$\begin{aligned} \mathcal{F}^{nr}(p_F,q,{\bf p})&= \theta(p_F-|{\bf p}|)\theta(|{\bf p+q}|-p_F)\nonumber\\ &\times \delta\Big(\omega +\frac{{\bf p}^2}{2m}-\frac{|{\bf p+q}|^2}{2m}\Big)\ , \end{aligned}$$ and $$\begin{aligned} \Gamma^{nr}&={\rm Max}\{\Gamma^{nr}_1,\Gamma^{nr}_2\}\nonumber\\ &={\rm Max}\Big\{ E^{nr}_F-\omega,m+\frac{1}{2m}\Big( \frac{\omega m}{|{\bf q}|}-\frac{|{\bf q}|}{2}\Big)^2\Big\}\ . \label{eps_nr}\end{aligned}$$ The non relativistic Fermi energy reads $E^{nr}_F=m+ {p_F^2}/{2m}$. We can then introduce a non relativistic scaling variable given by $$\begin{aligned} \psi^{nr}= &\Big[\frac{1}{\xi^{nr}_F}\Big(\frac{\Gamma^{nr}}{m}-1\Big)\Big]^{1/2}= \frac{1}{\sqrt{2\xi^{nr}_F}}\Big(\frac{\lambda}{\kappa}-{\kappa}\Big)\ .\end{aligned}$$ In the limit of large $|{\bf q}|$, Eq.  can be written in terms of $\psi^{nr}$ as $$\begin{aligned} F^{nr}(p_F,q)=\frac{3\mathcal{N}\xi^{nr}_F}{4 \eta_F^3 m \kappa}\big(1-{\psi^{nr}}^2)\theta(1-{\psi^{nr}}^2)\ .\end{aligned}$$ In analogy with the relativistic case, the longitudinal and transverse responses are expressed as $$\begin{aligned} R^{nr}_L&=\frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}^{nr}(p_F,q,{\bf p})\Big\{\frac{G^2_E(\tau)}{1+\tau}\Big\}\ \nonumber\\ &=\frac{3\mathcal{N}\xi_F}{4 \eta_F^3 m \kappa}\big(1-{\psi^{nr}}^2)\theta(1-{\psi^{nr}}^2) \Big\{\frac{G^2_E(\tau)}{1+\tau}\Big\}\ ,\\ R^{nr}_T&=\frac{3 \mathcal{N} }{4\pi p_F^3}\int d^3p\ \mathcal{F}^{nr}(p_F,q,{\bf p})\Big\{ \frac{G^2_E(\tau)}{m^2}p_T^2\nonumber\\ &+ \frac{G^2_M(\tau)}{2m^2}|{\bf q}|^2\Big\}\nonumber\\ &= \frac{3\mathcal{N}\xi_F}{4 \eta_F^3 m \kappa}\big(1-{\psi^{nr}}^2)\theta(1-{\psi^{nr}}^2)\nonumber\\ &\times\Big\{ G_E^2(\tau)\Delta^{nr}+ 2 G_M^2(\tau)\kappa^2\Big\}\ ,\end{aligned}$$ where $$\begin{aligned} \Delta^{nr}= \xi^{nr}_F (1-{\psi^{nr}}^2)\ .\end{aligned}$$ We then define the non relativistic longitudinal and transverse scaling functions as $$\begin{aligned} f^{nr}_L(\psi^{nr})&= p_F\times \frac{R^{nr}_L}{G^{nr}_L}\ ,\nonumber\\ f^{nr}_T(\psi^{nr})&= p_F\times \frac{R^{nr}_T}{G^{nr}_T}\ ,\end{aligned}$$ where $$\begin{aligned} G^{nr}_L&= \frac{\mathcal{N}}{2\kappa}\Big\{\frac{G^2_E(\tau)}{1+\tau}\Big\}\ ,\nonumber\\ G^{nr}_T&= \frac{\mathcal{N}}{2\kappa}\Big\{ G_E^2(\tau)\Delta^{nr}+ 2 G_M^2(\tau)\kappa^2\Big\} \ . \label{g_nr}\end{aligned}$$ In order to compare our results with the data, we introduce the experimental scaling functions obtained from the extracted longitudinal and transverse responses for $^4$He and $^{12}$C $$\begin{aligned} f^{exp}_L&= p_F\times\frac{R^{exp}_L}{G_L}\ ,\nonumber\\ f^{exp}_T&= p_F\times \frac{R^{exp}_T}{G_T}\ .\end{aligned}$$ It is long known that $f^{exp}_L$ clearly shows a scaling behavior in the limit of large momentum transfer. On the other hand, sizable scaling violations occur in the transverse channel, due to significant contributions given by two-body currents, resonance excitations and inelastic scattering. Hence, the comparison with the experimental data will be performed considering only the longitudinal contribution, $f^{exp}_L$. Results ======= ![(color online) Ratio of the non relativistic and relativistic expressions of the prefactors entering the definition of the scaling function plotted as a function of $\psi$, for $|{\bf q}|$= 300, 380, 570 MeV. The blue solid and red dashed lines correspond to the longitudinal and transverse channels, respectively. []{data-label="prefact"}](gl_gt_300_rat "fig:") ![(color online) Ratio of the non relativistic and relativistic expressions of the prefactors entering the definition of the scaling function plotted as a function of $\psi$, for $|{\bf q}|$= 300, 380, 570 MeV. The blue solid and red dashed lines correspond to the longitudinal and transverse channels, respectively. []{data-label="prefact"}](gl_gt_380_rat "fig:") ![(color online) Ratio of the non relativistic and relativistic expressions of the prefactors entering the definition of the scaling function plotted as a function of $\psi$, for $|{\bf q}|$= 300, 380, 570 MeV. The blue solid and red dashed lines correspond to the longitudinal and transverse channels, respectively. []{data-label="prefact"}](gl_gt_570_rat "fig:") Here we analyze the scaling features of the GFMC responses. In order to highlight the underlying nuclear dynamics we first divide them by the non relativistic prefactors $G^{nr}_{L,T}$. These have been obtained expanding the relativistic-current matrix elements in powers of $1/m$ retaining terms up to $\mathcal{O}[1/m^2]$ [@Lovato:2016gkq]. Relativistic corrections appear as terms of $\mathcal{O}[1/m^2]$ in the longitudinal channel while they are $\mathcal{O}[1/m^3]$ in the transverse one, and are therefore neglected in this case. This difference plays a relevant role in the interpretation of the results presented below. For a meaningful comparison with the scaling functions extracted from experimental data, we also present the results obtained using the relativistic prefactors $G_{L,T}$. Figure \[prefact\] clearly shows the different behavior of $G^{nr}_{L,T}$ and $G_{L,T}$ for three values of the momentum transfer. Relativistic effects are particularly relevant in the transverse case; at $|{\bf q}|$= 570 MeV the ratio $G^{nr}_T/G_T$ significantly differs from 1 for $\psi\geq 0$. ![(color online) Longitudinal (solid blue) and transverse (dashed red) scaling functions obtained from the GFMC calculation of the longitudinal and transverse responses of $^{12}$C at $|{\bf q}|= 300$ MeV. [**Upper panel**]{}: the responses have been divided by the non relativistic prefactors and the resulting curves are plotted as a function of $\psi^{nr}$. [**Lower panel**]{}: the standard definition of the prefactors given in Eq.  has been used to get both the theoretical curves and the experimental points obtained from the data of Ref. [@Barreau:1983ht] . []{data-label="300_12C"}](12C300_rlrt_gfmc_nr2 "fig:") ![(color online) Longitudinal (solid blue) and transverse (dashed red) scaling functions obtained from the GFMC calculation of the longitudinal and transverse responses of $^{12}$C at $|{\bf q}|= 300$ MeV. [**Upper panel**]{}: the responses have been divided by the non relativistic prefactors and the resulting curves are plotted as a function of $\psi^{nr}$. [**Lower panel**]{}: the standard definition of the prefactors given in Eq.  has been used to get both the theoretical curves and the experimental points obtained from the data of Ref. [@Barreau:1983ht] . []{data-label="300_12C"}](12C300_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_12C\] but for $|{\bf q}|= 380$ MeV.[]{data-label="380_12C"}](12C380_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_12C\] but for $|{\bf q}|= 380$ MeV.[]{data-label="380_12C"}](12C380_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_12C\] but for $|{\bf q}|= 570$ MeV.[]{data-label="570_12C"}](12C570_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_12C\] but for $|{\bf q}|= 570$ MeV.[]{data-label="570_12C"}](12C570_rlrt_gfmc_rel "fig:") In Figs. \[300\_12C\], \[380\_12C\] and \[570\_12C\] we show the longitudinal (blue solid lines) and transverse (red dashed lines) scaling functions extracted from the GFMC calculations of the $^{12}$C response functions. The results in the upper panels, obtained dividing the GFMC calculations by $G^{nr}_{L(T)}$, are plotted as a function of the non relativistic scaling variable $\psi^{nr}$. In the lower panels a comparison between the theoretical curves and the experimental points, in which the relativistic form of the prefactors has been adopted, is presented. It is important to point out that the longitudinal response of $^{12}$C is known to be affected by the elastic and the low lying excited states$J^{\pi}=2^+,\ 0_2^+,$ and $4^+$contributions. In order to compare experiments which refer only to the inclusive quasi-elastic responsewith GFMC calculations, these contributions have been explicitly subtracted by using the experimental values of excitations energies and form factors. Because of the fast drop of the form factors with increasing momentum transfer, in Ref. [@Lovato:2016gkq] it is argued that these corrections are expected to be significant in the longitudinal channel at $|{\bf q}|= 300$ MeV, but almost negligible at $|{\bf q}|= 570$ MeV. On the other hand, in the transverse channel such contributions are expected to be always negligible. The scaling functions displayed in the upper panels exhibit a clearly asymmetric shape, with a tail extending in the region $\psi^{nr} > 0$, as opposed to the GRFG model predictions. The difference in magnitude between the longitudinal and transverse GFMC scaling functions, which become less evident for larger values of $|{\bf q}|$, is likely to be ascribed to small residual effects of the low lying excited state contributions. For the aforementioned reason, in the lower panels the agreement between the longitudinal GFMC scaling function and the experimental data improves with increasing momentum transfer. The different behavior of the transverse scaling functions displayed in the upper and lower panels deserves some comments. In the lower panels, the red curves present a large non vanishing tail for $\psi >1$, although those are expected to approach zero, as shown in the upper panels. This discrepancy can be best understood considering the results of Fig.\[prefact\]. The relativistic and non relativistic expressions of the transverse prefactors used to extract the scaling functions are sizably different in the kinematic setups considered. In particular, for $|{\bf q}|=570$ MeV, these are very similar for $-1.5\leq \psi \leq 0$ where their ratio is almost 1, while in the region $\psi\geq 0$ their trend is significantly different and $G^{nr}_T/G_T$ increases for larger values of $\psi$. ![(color online) Experimental scaling functions of $^{12}$C obtained from the longitudinal responses for $|{\bf q}|=300,\ 380,\ 570$ MeV [@Barreau:1983ht]. []{data-label="fl_exp_12C"}](12Cfl_exp_all) ![(color online) Longitudinal scaling functions of $^{12}$C obtained from GFMC calculations for $|{\bf q}|=300,\ 380,\ 570$ MeV as a function of $\psi^{nr}$. []{data-label="fl_all_12C"}](12Cfl_th_nr_all) Figure \[fl\_exp\_12C\] shows the experimental scaling functions of $^{12}$C extracted from the experimental data of Ref. [@Barreau:1983ht] for $|{\bf q}|=300,\ 380,$ and 570 MeV. Although scaling is expected to occur in the limit of large momentum transfer, within the error bars of the different data points, the longitudinal response functions scale to a universal curve over the entire quasi-elastic peak, even in the region of moderate $|{\bf q}|$. In Fig. \[fl\_all\_12C\] the longitudinal GFMC scaling functions are shown as a function of $\psi^{nr}$ for $|{\bf q}|=300,\ 380,$ and 570 MeV. The theoretical results seem to indicate that first-kind scaling occurs. However, the interpretation of the differences between the three curves is obscured by the residual effect of the low-lying transitions discussed above. A more meaningful comparison can be carried out in the transverse channel, where the response functions are not affected by this effect. ![(color online) Transverse scaling functions of $^{12}$C obtained from GFMC calculations for $|{\bf q}|=300,\ 380,\ 570$ MeV as a function of $\psi^{nr}$. []{data-label="ft_all_12C"}](12Cft_th_nr_all) Figure \[ft\_all\_12C\] shows the GFMC results for the transverse scaling functions. The difference between the three curves in the ${\psi}^{nr}<0$ region suggests that, for $|{\bf q}|=300,\ 380$ MeV, the requirement $\Gamma=\Gamma_2$ \[see Eq. \]which is necessary to introduce the scaling variableis not satisfied for all the values of $\omega$. Indeed, the scaling violation in the low-energy transfer region is clearly visible. ![(color online) Experimental scaling functions obtained from the longitudinal responses of $^4$He for $|{\bf q}|=$300, 400, 500, 600 and 700 MeV [@Carlson:2001mp]. The value of the Fermi momentum of $^4$He has been set to $180$ MeV. The black dots correspond to the scaling function obtained from the experimental longitudinal response of $^{12}$C at $|{\bf q}|=570$ MeV [@Barreau:1983ht]. []{data-label="fl_exp_4He"}](4hefl_exp) To better elucidate the scaling properties of the GFMC calculations, it is worth to analyze the $^4$He nucleus, whose longitudinal response functions are not affected by low-lying transitions. In Fig. \[fl\_exp\_4He\], the scaling functions obtained from the experimental data of the longitudinal responses of $^4$He at $|{\bf q}|= 300,\ 400,\ 500\ ,600,$ and $700$ MeV are shown. Choosing the Fermi momentum equal to $180$ MeV, we observe that the points corresponding to different values of the momentum transfer tend to lay on top of each other, and the agreement with the $^{12}$C data at $|{\bf q}|= 570$ MeV is also remarkable. ![(color online) Longitudinal (solid blue) and transverse (dashed red) scaling functions obtained from the GFMC calculation of the longitudinal and transverse responses of $^{4}$He at $|{\bf q}|= 300$ MeV. [**Upper panel**]{}: the responses have been divided by the non relativistic prefactors and the resulting curves are plotted as a function of $\psi^{nr}$. [**Lower panel**]{}: the standard definition of the prefactors given in Eq. has been used to get both the theoretical curves and the experimental points obtained from the data of Ref. [@Carlson:2001mp] .[]{data-label="300_4He"}](4he300_rlrt_gfmc_nr2 "fig:") ![(color online) Longitudinal (solid blue) and transverse (dashed red) scaling functions obtained from the GFMC calculation of the longitudinal and transverse responses of $^{4}$He at $|{\bf q}|= 300$ MeV. [**Upper panel**]{}: the responses have been divided by the non relativistic prefactors and the resulting curves are plotted as a function of $\psi^{nr}$. [**Lower panel**]{}: the standard definition of the prefactors given in Eq. has been used to get both the theoretical curves and the experimental points obtained from the data of Ref. [@Carlson:2001mp] .[]{data-label="300_4He"}](4he300_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 400$ MeV.[]{data-label="400_4He"}](4he400_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 400$ MeV.[]{data-label="400_4He"}](4he400_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 500$ MeV.[]{data-label="500_4He"}](4he500_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 500$ MeV.[]{data-label="500_4He"}](4he500_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 600$ MeV.[]{data-label="600_4He"}](4he600_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 600$ MeV.[]{data-label="600_4He"}](4he600_rlrt_gfmc_rel "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 700$ MeV.[]{data-label="700_4He"}](4he700_rlrt_gfmc_nr2 "fig:") ![ Same as in Fig. \[300\_4He\] but for $|{\bf q}|= 700$ MeV.[]{data-label="700_4He"}](4he700_rlrt_gfmc_rel "fig:") ![(color online) Longitudinal scaling functions obtained from GFMC calculations of the longitudinal response of $^{4}$He for $|{\bf q}|=400,\ 500,\ 600,\ 700$ MeV and of $^{12}$C at $|{\bf q}|=570$ MeV. []{data-label="fl_all_4He"}](4hefl_all_th) ![(color online) Transverse scaling functions obtained from GFMC calculations of the transverse response of $^{4}$He for $|{\bf q}|=400,\ 500,\ 600,\ 700$ MeV and of $^{12}$C at $|{\bf q}|=570$ MeV.[]{data-label="ft_all_4He"}](4heft_all_th) In Figs. \[300\_4He\]-\[700\_4He\] we show the longitudinal (solid blue) and transverse (dashed red) scaling functions extracted from the GFMC calculations of $^{4}$He at $|{\bf q}|=300,\ 400\ ,500\ ,600,$ and 700 MeV. In the upper and lower panels the same scheme followed to present the $^{12}$C scaling functions has been adopted. In the longitudinal channel, theoretical calculations and experimental data reported in the lower panels present are in very nice agreement in all the kinematic setups. Finding this agreement up to $|{\bf q}|= 700$ MeV may appear surprising since the GFMC is a non relativistic approach. This can be understood because all the relativistic corrections coming from both the Dirac-spinors and the currents are kept up to $\mathcal{O}[1/m^2]$. However, this is not the case in the transverse channel where relativistic corrections are subleading and have been neglected. Moreover, the differences in magnitude of the transverse scaling functions, following the discussion carried out for $^{12}$C, are likely to be ascribed to relativistic effects in the prefactors. The upper panels of Figs. \[300\_4He\]-\[700\_4He\] clearly show that in the $^4$He case the scaling of the zeroth-kind is manifest when the effects of nuclear dynamics are singled out by using the non relativistic expressions for the prefactors. The absence of low-lying transition contributions makes the scaling of the first kind apparent. The curves of Figs \[fl\_all\_4He\] and \[ft\_all\_4He\], where we compare the longitudinal and transverse scaling functions of $^4$He for different values of the momentum transfer, present a remarkably good scaling behavior. The $^4$He results for $|{\bf q}|= 600\ ,700$ MeV are almost coincident and in good agreement with the longitudinal scaling function of $^{12}$C computed at $|{\bf q}|= 570$ MeV. Figures \[fl\_all\_4He\] and \[ft\_all\_4He\] prove that the asymmetric shape of the scaling function does not depend upon the momentum transfer. Consequently, it is not likely to be ascribed to collective excitation modes, that can be accounted for within the random phase approximation. This analysis, carried out for a variety of kinematics suggests that scaling occurs in the GFMC calculations of the longitudinal and transverse response functions of both $^4$He and $^{12}$C nuclei. Comparing the definition of the longitudinal response function and the one of the corresponding prefactor, see Eq. and , while neglecting the spin-orbit contribution, one is lead to conclude that the scaling function corresponds to $$\begin{aligned} f_L= \frac{2\kappa\ R_\varrho}{\mathcal{N}} \label{n:dens}\end{aligned}$$ where $R$ is the nucleon-density response function defined as $$\begin{aligned} R_\varrho\equiv &\sum_f \langle 0| \varrho^\dagger(\mathbf{q}) | f \rangle \langle f | \varrho(\mathbf{q}) | 0 \rangle\, \delta (E_0+\omega-E_f) \ ,\end{aligned}$$ in terms of the nucleon-density operator $$\varrho\equiv \sum_i e^{i \mathbf{q}\cdot \mathbf{r}_i} \frac{(1\pm\tau_{i,z})}{2}\, ,$$ where the $\pm$ applies to protons and neutrons, respectively. Note that Eq. holds also in the relativistic case, provided that relativistic expressions for the energies are used and spinors are normalized as $\bar{u}u= \sqrt{m/E}$ to absorb the factor $m^2/(E({\bf p})E({\bf p+q}))$ of Eq. (\[eq:scale\_integrand\]). On the other hand, the transverse scaling function corresponds to the spin-response, which reduces to the nucleon-density response defined above in the limit of high momentum-transfer, where the impulse approximation is expected to be accurate and where $|\mathbf{q}| \gg |{\bf p}_T|$. We found that the longitudinal and transverse response functions obtained retaining only the one body current operator scale to the same universal scaling function: the nucleon-density response function. The results presented in Ref. [@Lovato:2016gkq] show that two-body currents lead to a significant enhancement of the transverse response of $^{12}$C in the region of the quasi elastic peak. We expect that the inclusion of this contribution in the scaling analysis, while leaving the longitudinal scaling function unchanged, would contribute to the observed scaling violation of the experimental scaling function in the transverse channel for $\psi\geq0$. Conclusions {#conclusion} =========== We have performed a scaling analysis of the GFMC electromagnetic response functions of $^4$He and $^{12}$C for a variety of kinematic setups. Despite the non relativistic nature of the calculation, all the GFMC scaling functions analyzed are strongly asymmetric, with a tail extending to the large $\psi$ region. Within the present picture, this is a consequence of nuclear correlations in both the initial and final states. This is at variance with the findings of Ref. [@Caballero:2007tz], where the asymmetric shape was ascribed to relativistic effects in the treatment of the final state interactions. In this regard, it is interesting to point out that the symmetry of the scaling function is not recovered even for momentum transfer as low as $|{\bf q}| =300$ MeV in both $^4$He and $^{12}$C. When the nuclear dynamics is properly singled out, the $^{12}$C response function shows a fairly good scaling behavior. However, the presence of the low lying transitions which is known to affect the longitudinal channel, introduces non trivial difficulties in drawing definitive conclusions. A better understanding is given by the analysis of $^4$He responses which are free from the uncertainties coming from these contributions. Our results for this nucleus indicate that both the zeroth- and first-kind of scaling occur. Moreover, the $^4$He and $^{12}$C scaling functions fulfill scaling of the second kind once the Fermi momentum of $^4$He is appropriately tuned. From our analysis, a novel interpretation of the scaling function emerges. If the spin-orbit contribution to the density-current operator is neglected, it can be easily noted that the longitudinal scaling function corresponds to the nucleon-density response. In the transverse channel, for sufficiently large momentum transfer the term proportional to the transverse momentum of the incoming nucleon can be safely neglected and the scaling function is proportional to the spin-response. In nuclei characterized by total spin $S=0$, such as $^4$He, $^{12}$C, $^{16}$O and $^{40}$Ca, in the impulse approximation the spin-response reduces to the nucleon-density response. Our findings on the occurrence of zeroth-kind scaling are consistent with this interpretation. In fact, within GFMC the scaling violation of the transverse response in the quasi-elastic region is likely to come from two-body currents. This was first noted by the authors of Ref. [@Carlson:2001mp] in which a better agreement between the experimental data and the theoretical calculation of the Euclidean responses of $^3$He and $^4$He was found, once that this term was accounted for. The role played by two-body current contributions in the electromagnetic responses of $^{12}$C have been recently investigated in Ref. [@Lovato:2016gkq], where a significant enhancement of the transverse response is observed at all momentum transfers: not only in the [*dip*]{} region, but in the whole quasi-elastic peak region, extending below the pion-production threshold. In the pioneering work of Ref. [@Fabrocini:1996bu], it has been shown that such enhancement is mainly due to the interference between one- and two-body currents leading to single knock out final state. In this case the kinematics would be very similar to the those analyzed in this paper, where only one-body current contributes. Hence, we expect that it would be possible to define an appropriate scaling function for these processes. The consequences of the two-body current contribution in the GFMC scaling functions as well as the study of the scaling properties of the total nuclear response including both one- and two-body termswill be the subject of a future work. Acknowledgements {#acknowledgements .unnumbered} ================ Research partially supported by the Spanish Ministerio de Economía y Competitividad and the European Regional Development Fund, under contracts FIS2014-51948-C2-1-P and SEV-2014-0398, by Generalitat Valenciana under contract PROMETEOII/2014/0068, and by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, under contract DE-AC02-06CH11357 (A.L.). Under an award of computer time provided by the INCITE program, this research used resources of the Argonne Leadership Computing Facility at Argonne National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under contract DE-AC02-06CH11357.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In these lecture notes, we present the equations presently used in stellar interior models in order to compute the effects of axial rotation. We discuss the hypotheses made. We suggest that the effects of rotation might play a key role at low metallicity.' author: - Georges Meynet title: Physics of rotation in stellar models --- Physics of rotation =================== Axial rotation modifies the hydrostatic equilibrium configuration by adding a centrifugal acceleration term in the hydrostatic equation, induces many instabilities driving the transport of angular momentum and of chemical species in radiative zones and changes the mass loss rates. In the present lecture notes we shall consider the case of models without magnetic fields. Hydrostatic effects of rotation ------------------------------- ### The equations of stellar structure. In a rotating star, the equations of stellar structure need to be modified [@KipTho70]. The usual spherical coordinates must be replaced by new coordinates characterizing the equipotentials. The classical method applies when the effective gravity can be derived from a potential $\Psi = \Phi - \frac{1}{2} \Omega^2 r^2 \sin^2 \theta$, i.e. when the problem is conservative. There, $\Phi$ is the gravitational potential, $\Omega$ the angular velocity, $r$ the radius at the colatitude $\theta$. If the rotation law is shellular ([*i.e.*]{} such that $\Omega$ is constant on isobaric surfaces see below), the problem is non–conservative. Most existing models of rotating stars apply, rather inconsistently, the classical scheme by [@KipTho70]. However, as shown by [@MM97], the equations of stellar structure can still be written consistently, in term of a coordinate referring to the mass inside the isobaric surfaces[^1] . Thus, the problem of the stellar structure of a differentially rotating star in a shellular rotation state can be kept one–dimensional. ### The Roche model. In all the derivations, we shall use the Roche model, [*i.e.*]{} we approximate the gravitational potential by $GM_{\overline{r}}/\overline{r}$ where $M_{\overline{r}}$ is the mass inside the isobaric surface with a mean radius $\overline{r}$. The radius $\overline{r}$ which labels each isobaric surface is defined by $\overline{r}=\left( V_{\overline{r}}/(4/3 \pi)\right )^{1/3}$ where $V_{\overline{r}}$ is the volume (deformed by rotation) inside the isobaric surface considered. Apart from the case of extreme rotational velocities, the parameter $\overline{r}$ is close to the average radius of an isobar, which is the radius at $P_2 (\cos \vartheta) =0$, namely for $\vartheta = 54.7$ degrees. In the frame of the Roche model, the shape of a meridian at the surface of the star (which is an isobaric surface) is given by couples of $R$ and $\theta$ values satisfying the following equation: $${GM \over R}+{1 \over 2}\Omega^2 R^2 \sin^2 \theta={GM \over R_{\rm p}}, \label{eq1}$$ where $R$ is the radius at colatitude $\theta$, $\Omega$ the angular velocity, $M$ the mass inside the surface and $R_{\rm p}$, the polar radius. Thus the shape of the surface (as well as of any isobaric surface inside the star) is determined by three parameters $M$, $\Omega$ and $R_{\rm p}$. The first two $M$ and $\Omega$ are independent variables. The third one is derived from the first two and the equations of stellar structure. Setting $x=\left({GM \over \Omega^2} \right)^{-1/3} R$, one can write Eq. \[eq1\] (see [@KipTho70]) $${1 \over x}+{1 \over 2}x^2 \sin^2 \theta={1 \over x_{\rm p}}. \label{eq1b}$$ With this change of variable, the shape of an equipotential is uniquely determined by only one parameter $x_{\rm p}$. Setting $$f={R_{\rm e} \over R_{\rm p}}, \label{eq2}$$ where $R_{\rm e}$ is the equatorial radius, one easily obtains from Eq. \[eq1\] that $$R_{\rm p}=\left({GM \over \Omega^2} \right)^{1/3} \left({2 (f-1)\over f^3} \right)^{1/3}=\left({GM \over \Omega^2} \right)^{1/3} x_{\rm p}. \label{eq3}$$ The above equation relates the inverse of the oblateness $f$ to $R_{\rm p}$. ### The von Zeipel theorem and its consequences. The von Zeipel theorem [@vZ24] expresses that the radiative flux $\vec{F}$ at some colatitude $\vartheta$ in a rotating star is proportional to the local effective gravity $\vec{g_\mathrm{eff}}$. [@Ma99] has generalized this theorem to the case of shellular rotation and the expression of the flux $\vec{F}$ for a star with angular velocity $\Omega$ on the isobaric stellar surface is $$\vec{F} = - \frac{L(P)}{4 \pi GM_{\star}} \vec{g_{\rm{eff}}} [1 + \zeta(\vartheta)] \quad {\mathrm{with}} \quad \label{Ma23}$$ $$M_{\star} = M \left( 1 - \frac{\Omega^2} {2 \pi G \rho_{\rm{m}}} \right) \quad {\mathrm{and}} \quad \label{Ma24}$$ $$\zeta(\vartheta) = \left[\left(1 - \frac{\chi_T}{\delta}\right) \Theta + \frac{H_T}{\delta} \frac{d\Theta}{dr}\right] P_{2}(\cos \vartheta). \label{Ma25}$$ There, $\rho_{\rm{m}}$ is the internal average density, $\chi = 4acT^3/(3 \kappa \rho)$ and $\chi_{T}$ is the partial derivative with respect to T. The quantity $\Theta$ is defined by $\Theta = \frac{\tilde{\rho}}{\bar{\rho}}$, i.e. the ratio of the horizontal density fluctuation to the average density on the isobar [@Z92]. One has the thermodynamic coefficients $\delta = - (\partial \ln\rho / \partial \ln T)_{P, \mu}$, $H_{T}$ is the temperature scale height. The term $\zeta(\vartheta)$, which expresses the deviations of the von Zeipel theorem due to the baroclinicity of the star, is generally very small, (cf. [@Ma99]). Let us emphasize that the flux is proportional to $\vec{g_\mathrm{eff}}$ and not to $\vec{g_\mathrm{tot}}$. This results from the fact that the equation of hydrostatic equilibrium is $\frac{\vec{\nabla} P}{\rho} = - \vec{g_\mathrm{eff}}$. The effect of radiation pressure is already counted in the expression of $P$, which is the total pressure. We may call $M_{\star}$ the effective mass, i.e. the mass reduced by the centrifugal force. This is the complete form of the von Zeipel theorem in a differentially rotating star with shellular rotation, whether or not one is close to the Eddington limit. This theorem has numerous consequences. Some of them are discussed below. A fast rotating star has stronger radiative fluxes at the pole than at the equator. Therefore the position of such a star in the HR diagram will depend of the angle between the line of sight and the rotational axis (inclination angle). If for instance that angle is 90 degrees, a great part of the light will come from the equatorial belt characterized by lower radiative fluxes and cooler effective temperatures, while when the star is observed pole-on most of the light will come from the hot polar region characterized by stronger fluxes and higher effective temperatures. Thus the perceived luminosity and effective temperature (and also effective gravity) of a star depend on the inclination angle. This has to be kept in mind when comparisons are made with observed quantities. Computations of the effect of the inclination angle on the emergent luminosity, colors and spectrum have been performed by [@MP70]. The effect of the inclination angle on the determination of the effective gravity is discussed in [@HG06]. In general, a theoretical evolutionary track is given in term of total luminosity and of an average effective temperature defined by $T_{\rm eff}^4= L/(\sigma S(\Omega))$, where $\sigma$ is Stefan’s constant and $S(\Omega)$ the total actual stellar surface. The total luminosity (corresponding to the integrated flux over the surface) does not depend on the angle of view, but cannot be directly compared to the “observed” luminosity deduced from the apparent luminosity coming from the hemisphere directed toward us. Let us note however that for surface velocities inferior to about 70% of the critical velocity these effects remain quite modest. As a numerical example, the ratio ($T_{\rm eff}$(pole)-$T_{\rm eff}$(equator))/$T_{\rm eff}$(equator) becomes superior to 10% only for $\omega > 0.7$. At break-up, the effective temperature of the polar region is about a factor two higher than that of the equatorial one. Let us express the total gravity at some colatitude $\vartheta$, taking into account the radiative acceleration (cf. [@Ma99]) $$\vec{g_\mathrm{rad}} = \frac{1}{\rho} \vec{\nabla} P_\mathrm{rad} = \frac{\kappa(\vartheta)\vec{F}}{c} \; ,$$ thus one has $$\begin{aligned} \vec{g_\mathrm{tot}} =\vec{g_\mathrm{eff}}+\vec{g_\mathrm{rad}}= \vec{g_\mathrm{eff}}+\frac{\kappa(\vartheta)\vec{F}}{c} $$ The rotation effects appear both in $\vec{g_\mathrm{eff}}$ and in $\vec{F}$. We may also consider the local limiting flux. The condition $\vec{g_\mathrm{tot}}= \vec{0}$ allows us to define a limiting flux, $$\vec{F_{\mathrm{lim}}}(\vartheta) = - \frac{c}{\kappa(\vartheta)} \vec{g_{\mathrm{eff}}}(\vartheta) \; .$$ From that we may define the ratio $\Gamma_{\Omega}(\vartheta)$ of the actual flux (see Eq. \[Ma23\]) $F(\vartheta)$ to the limiting local flux in a rotating star, $$\begin{aligned} \Gamma_{\Omega}(\vartheta) = \frac{\vec{F}(\vartheta)}{\vec{F_{\mathrm{lim}}}(\vartheta)}= \frac{ \kappa (\vartheta) \; L(P)[1+\zeta (\vartheta)]}{4 \pi cGM \left( 1 - \frac{\Omega^2}{2 \pi G \rho_{\rm{m}}} \right) }\; . \label{Ma29}\end{aligned}$$ As a matter of fact, $\Gamma_{\Omega}(\vartheta)$ is the local Eddington ratio and $$\begin{aligned} L_{\rm Edd} = \frac{4 \pi cGM \left( 1 - \frac{\Omega^2}{2 \pi G \rho_{\rm{m}}} \right) }{\kappa (\vartheta) \; [1+\zeta (\vartheta)] } \label{Lmax}\end{aligned}$$is the local Eddington luminosity. For a certain angular velocity $\Omega$ on the isobaric surface, the maximum permitted luminosity of a star is reduced by rotation, with respect to the usual Eddington limit. In the above relation, $\kappa(\vartheta)$ is the largest value of the opacity on the surface of the rotating star. For O–type stars with photospheric opacities dominated by electron scattering, the opacity $\kappa$ is the same everywhere on the star. For zero rotation, the usual expressions are found: $\Gamma_{\Omega}(\vartheta) = \Gamma=\frac{\kappa L}{4 \pi c GM}$ and $L_{\rm Edd} =\frac{4 \pi cGM }{\kappa}$ . Critical limits correspond to values of respectively the luminosity and/or the velocity which impose that the total gravity become equal to zero at least at some peculiar places at the surface. We may identify different limits [@mm6]: – We speak of the Eddington or $\Gamma$–limit, when rotation effects can be neglected and $\vec{g_\mathrm{rad}} + \vec{g_\mathrm{grav}} = \vec{0}$[^2], which implies that $$\Gamma = \frac{\kappa L}{4 \pi c GM} \; \rightarrow \; 1.$$ In that case $L = L_{\rm Edd}=4 \pi c GM/\kappa$. The opacity $\kappa$ considered here is the total opacity. –The critical velocity or $\Omega$–limit is reached for a star with an angular velocity $\Omega$ at the surface, when the effective gravity $\vec{g_\mathrm{eff}} = \vec{g_\mathrm{grav}} + \vec{g_\mathrm{rot}} = \vec{0}$ and in addition when radiation pressure effects can be neglected. –*The $\Omega \Gamma$–limit is reached when the total gravity $\vec{g_\mathrm{tot}} = \vec{0}$, with significant effects of both rotation and radiation.* This is the general case. It should lead to the two above cases in their respective limits. Using relation \[Ma29\], we may write the expression for the total gravity as $$\vec{g_\mathrm{tot}} = \vec{g_\mathrm{eff}} \left[ 1 - \Gamma_{\Omega}(\vartheta) \right] \; .$$ This shows that the expression for the total acceleration in a rotating star is similar to the usual one, except that $\Gamma$ is replaced by the local value $\Gamma_{\Omega}(\vartheta)$. Indeed, contrarily to expressions such as $\vec{g_\mathrm{tot}} = \vec{g_\mathrm{eff}} \left( 1 - \Gamma \right)$ often found in literature, we see that the appropriate Eddington factor given by Eq. \[Ma29\] also depends on the angular velocity $\Omega$ on the isobaric surface. From Eq. \[Ma29\], we note that over the surface of a rotating star, which has a varying gravity and T$_\mathrm{eff}$, $\Gamma_{\Omega}(\vartheta)$ is the highest at the latitude where $\kappa(\vartheta)$ is the largest, (if we neglect the effects of $\zeta (\vartheta)$, which is justified in general). If the opacity increases with decreasing T as in hot stars, the opacity is the highest at the equator and there the limit $\Gamma_{\Omega}(\vartheta) = 1 $ may by reached first. Thus, it is to be stressed that if the limit $\Gamma_{\Omega}(\vartheta) = 1 $ happens to be met at the equator, it is not because $\vec{g_\mathrm{eff}}$ is the lowest there, but because the opacity is the highest ! The reason for no direct dependence on $\vec{g_\mathrm{eff}}$ is because both terms $\vec{g_\mathrm{eff}}$ cancel each other in the expression \[Ma29\] of the flux ratio. The critical velocity is reached when somewhere on the star one has $\vec{g_\mathrm{tot}}=\vec{0}$, i.e. $$\vec{g_\mathrm{eff}} \;\left[1 - \Gamma_{\Omega}(\vartheta) \right] = \vec{0} . \label{Ma313}$$ This equation has two roots. The first one $v_\mathrm{crit,1}$ is given by the usual condition $\vec{g_\mathrm{eff}}= \vec{0}$, which implies the equality $\Omega^2 R^{3}_\mathrm{eb}/(GM) =1$ at the equator. This corresponds to an equatorial critical velocity $$v_\mathrm{crit, 1} = \Omega \; R_\mathrm{eb} = \left( \frac{2}{3} \frac{GM}{R_\mathrm{pb}} \right)^{\frac{1}{2}} \; . \label{Ma314}$$ $R_\mathrm{eb}$ and $R_\mathrm{pb}$ are respectively the equatorial and polar radius at the critical velocity. We notice that the critical velocity $v_\mathrm{crit, 1}$ is independent on the Eddington factor. To this extent, this is in agreement with [@Gla98]. The basic physical reason for this independence is quite clear: the radiative flux tends toward zero when the effective gravity is zero, thus there is no effect of the radiative acceleration when this occurs. Equation \[Ma313\] has a second root, which is given by the condition $\Gamma_{\Omega}(\vartheta)$ = 1. If we call the Eddington ratio $\Gamma_\mathrm{max}$ the maximum value of $\kappa(\vartheta) L(P)/(4 \pi c GM)$ over the surface (in general at equator), [@mm6] has shown that for $\Gamma_\mathrm{max} < 0.639$, no value of $\Omega$ can lead to $\Gamma_{\Omega}(\vartheta)$ = 1. Physically this means that when the star is sufficiently far from the Eddington limit, the effects of rotation on the radiative equilibrium are not sufficient for lowering the Eddington luminosity such that it may have an impact on the value of the critical velocity. In that case, Eq. \[Ma313\] has only one root given by the classical expression. For a given large enough $\Gamma_\mathrm{max}$ (i.e. larger than 0.639), a second root is obtained given by $$\begin{aligned} v_\mathrm{crit, 2}^2 = \frac{9}{4} \;v_\mathrm{crit, 1}^2 \; \frac{1-\Gamma}{V^{\prime}(\omega)} \; \frac{ R^{2}_\mathrm{e}(\omega)}{R^2_{\mathrm{pb}}}\end{aligned}$$ The quantity $\omega$ is the fraction $\Omega/\Omega_\mathrm{c}$ of the angular velocity at break–up. The quantity $V^{\prime}(\omega)$ is the ratio of the actual volume of a star with rotation $\omega$ to the volume of a sphere of radius $R_\mathrm{pb}$. $V^{\prime}(\omega)$ is obtained by the integration of the solutions of the surface equation for a given value of the parameter $\omega$. For $\Gamma_\mathrm{max} > 0.639$, this second root is inferior to the first one. Thus it is encountered first and is therefore the expression of the critical velocity that has to be used. Due to the von-Zeipel theorem, the radiative flux, which is the driving force for the stellar winds of massive hot stars, varies as a function of the colatitude. This effect, when accounted for in the theory of the line driven wind theory, lead to an enhancement of the quantities of mass lost and to wind anisotropies. We shall describe in more details these effects in the Section 1.3 below. 1) The classical critical angular velocity or the $\Omega$-limit (to distinguish it from the $\Omega\Gamma$-limit as defined by [@mm6]) in the frame of the Roche model is given by $$\Omega_{\rm crit}=\left({2 \over 3}\right)^{3 \over 2}\left({GM \over R^3_{\rm pb}}\right)^{1 \over 2}, \label{eq4}$$ where $R_{\rm pb}$ is the polar radius when the surface rotates with the critical velocity. The critical velocity is given by $$\upsilon_{\rm crit}=\left({2 \over 3}{GM \over R_{\rm pb}}\right)^{1 \over 2}. \label{eq4b}$$ Replacing $\Omega$ in Eq. \[eq3\] by $(\Omega/\Omega_{\rm crit}) \cdot \Omega_{\rm crit}$ and using Eq. \[eq4\], one obtains a relation between $R_{\rm p}$, $f$ and $R_{\rm pb}$, $$R_{\rm p}={3 \over 2} R_{\rm pb} \left({\Omega_{\rm crit} \over \Omega}\right)^{2/3} \left({2 (f-1)\over f^3} \right)^{1/3}, \label{eq5}$$ Thus one has $${\Omega \over \Omega_{\rm crit}}=\left({3 \over 2}\right)^{3/2} \left({R_{\rm pb} \over R_{\rm p}}\right)^{3/2} \left({2 (f-1)\over f^3} \right)^{1/2}. \label{eq6}$$ With a good approximation (see below) one has that $R_{\rm pb}/ R_{\rm p} \simeq 1$ and therefore $${\Omega \over \Omega_{\rm crit}} \simeq \left({3 \over 2}\right)^{3/2} \left({2 (f-1)\over f^3} \right)^{1/2}. \label{eq7}$$ This equation is quite useful since it allows the determination of $\Omega_{\rm crit}$ from quantities obtained with a model computed for $\Omega$. This is not the case of Eq. \[eq5\] which involves $R_{\rm pb}$ whose knowledge can only be obtained by computing models at the critical limit. In general Eq. \[eq7\] gives a very good approximation of $\Omega_{\rm crit}$ (see [@EMM] for a discussion of this point). Setting $\upsilon$ the velocity at the equator, one has that $${\upsilon \over \upsilon_{\rm crit}}={\Omega R_{\rm e}\over\Omega_{\rm c}R_{\rm eb}} ={\Omega \over \Omega_{\rm crit}}{R_{\rm e}\over R_{\rm p}}{R_{\rm p}\over R_{\rm pb}}{R_{\rm pb}\over R_{\rm eb}}, \label{eq7bis}$$ where $R_{\rm eb}$ is the equatorial radius when the surface rotates with the critical velocity. Using Eq. \[eq5\] above, and the fact that in the Roche model $R_{\rm pb}/ R_{\rm eb}=2/3$, one obtains $${\upsilon \over \upsilon_{\rm crit}}= \left({\Omega \over \Omega_{\rm crit}}2(f-1)\right)^{1/3} \label{eq8}$$ The relations between $\upsilon/\upsilon_{\rm crit}$ and $\Omega/\Omega_{\rm crit}$ obtained in the frame of the Roche model (see Eq. \[eq8\]) for the 1 and 60 M$_\odot$ stellar models at $Z=0.02$ are shown in Fig. \[ocvc\]. In case we suppose that the polar radius remains constant $R_{\rm pb}/R_{\rm p}=1$, then Eq. \[eq7\] can be used and one obtains a unique relation between $\Omega/\Omega_{\rm crit}$ and $\upsilon/\upsilon_{\rm crit}$, independent of the mass, metallicity and evolutionary stage considered. One sees that the values of $\upsilon/\upsilon_{\rm crit}$ is smaller than that of $\Omega/\Omega_{\rm crit}$ by at most $\sim$25%. At the two extremes the ratios are of course equal. An interesting quantity is the ratio of the centrifugal acceleration, $a_{\rm cen}$, to the gravity, $g_{\rm e}$, at the equator $${a_{\rm cen} \over g_{\rm e}}={\Omega^2 R^3_{\rm e}\over G M}= \left({\Omega \over \Omega_{\rm crit}}\right)^2 \left({2 \over 3}\right)^3 f^3 \left({R_{\rm p}\over R_{\rm pb}}\right)^3, \label{eq9}$$ where we have used Eq. \[eq4\] and divided/multiplied by $R_{\rm pb}^3$. Replacing $R_{\rm pb}/R_{\rm p}$ by its expression deduced from Eq. \[eq5\], we obtain $${a_{\rm cen} \over g_{\rm e}}=2(f-1). \label{eq10}$$ We can check that at the critical limit, when $f=3/2$, then $a_{\rm cen}=g_{\rm e}$. 2) It is interesting to note that the fact that the effective temperature varies as a function of the colatitude on an isobaric surface does not necessarily imply that the temperature varies as a function of the colatitude on an isobaric surface. For instance in the conservative case, the temperature is constant on equipotentials which are also isobaric surfaces. This simply illustrates the difference between the effective temperature whose definition is related to the radiative flux ($F=\sigma T_{\rm eff}^4$) and hence to the [*temperature gradient*]{} and the temperature itself. Interestingly, one has that in a conservative case, $\Gamma_\Omega(\theta)$ is constant on isobaric surfaces (neglecting the term $\zeta(\vartheta)$). This means that when the $\Omega\Gamma$-limit is reached, it is reached over the whole stellar surface at the same time. This is in contrast with the $\Omega$-limit which is reached first at the equator. 3) Recently we have examined the effects of rotation on the thermal gradient and on the Solberg–Hoiland term by analytical developments and by numerical models [@MCM08]. Writing the criterion for convection in rotating envelopes, we show that the effects of rotation on the thermal gradient are much larger and of opposite sign to the effect of the Solberg–Hoiland criterion. On the whole, rotation favors convection in stellar envelopes at the equator and to a smaller extent at the poles. In a rotating 20 M$_{\odot}$ star at 94% of the critical angular velocity, there are two convective envelopes, the biggest one has a thickness of 13.2% of the equatorial radius. The convective layers are shown in Fig. \[P94\]. They are more extended than without rotation. In the non-rotating model, the corresponding convective zone has a thickness of only 4.6% of the radius. The occurrence of outer convection in massive stars has many consequences (see [@MCM08]). ![2–D representation of the external convective zones (in red) and of the convective core (blue) in a model of 20 M$_{\odot}$ with $X=0.70$ and $Z=0.020$ at the end of MS evolution with fast rotation ( $\Omega/\Omega_{\mathrm{crit}}=0.94$). The axis are in units of cm. Figure taken from [@MCM08].[]{data-label="P94"}](P020z20S7ZCsansSH.eps){width="8.8cm"} Transport mechanisms of angular momentum and of chemical species ---------------------------------------------------------------- In a solid body rotation state, the first instability to set up is a thermal instability called the meridional circulation (see below). It consists in large meridional currents which transport angular momentum either inside-out or conversely from the outer regions toward the inner ones. Such meridional currents rapidly build ut gradients of the angular momentum both in the “horizontal direction” ([*i.e.*]{} along isobaric surface) and in the vertical one. Along isobaric surface, any gradient of $\Omega$ triggers a strong horizontal turbulence. Indeed in that direction the instability can develop without having to overcome any stable density gradient. As a consequence any gradient of $\Omega$ along isobaric surfaces is rapidly erased and the star settles into a “shellular” rotation state [@Z92]. This means that $\Omega$ can be considered as nearly constant along isobars. In the vertical direction, where in a radiative zone, a stable density gradient counteracts any instability, the gradients of $\Omega$ are eroded on much longer timescales (see below). The equations below describe the interactions of meridional currents and of shear instabilities in a state of shellular rotation [@Z92]. ### Meridional circulation. Meridional circulation is an essential mixing mechanism in rotating stars and there is a considerable literature on the subject (see ref. in [@Tass90]). The velocity of the meridional circulation in the case of shellular rotation was derived by [@Z92]. The velocity of meridional circulation is derived from the equation of energy conservation [@Mes53] $$\begin{aligned} \rho T\left[{\partial S \over \partial t}+({\bf e}_r \dot r +{\bf U})\cdot{\bf \nabla}S\right]={\rm div}(\chi{\bf \nabla}T)+\rho\epsilon-{\rm div}{\bf F}_h \label{eqn9}\end{aligned}$$ where $S$ is the entropy per unit mass, $\chi$ the thermal conductivity, $\epsilon$ the rate of nuclear energy per unit mass and ${\bf F}_h$ the flux of thermal energy due to horizontal turbulence. All the quantities are expanded linearly around their average on a level surface or isobar, using Legendre Polynomials $P_2(\cos\theta)$. For instance $$T(P,\theta)=\bar T(P)+\tilde{T}P_2(\cos\theta).$$ Then Eq. \[eqn9\] is linearized and an expression for $U_2$ can be deduced [@Z92]. Using the same method [@MZ98] revised the expression for $U_2$ to account for expansion and contraction in non–stationary models. They also studied the effects of the $\mu$–gradients (mean molecular weight gradients), of the horizontal turbulence and considered a general equation of state. They obtained $$\begin{aligned} U_2(r)={P \over \bar \rho \bar g C_P\bar T[\nabla_{\rm ad}-\nabla+(\varphi/\delta)\nabla_\mu ]} \times\left[{L\over M_*}(E_\Omega+E_\mu)+{C_P \over \delta}{\partial \Theta \over \partial t}\right], \label{Umer}\end{aligned}$$ where $M_{\star}=M \left( 1 - \frac{\Omega^2}{2 \pi G \rho_{\rm{m}}} \right)$ is the reduced mass and the other symbols have the same meaning as in [@Z92] and [@MZ98] [^3]. The driving term in the square brackets in the second member is $E_{\Omega}$. It behaves mainly like $E_{\Omega} \simeq \frac{8}{3} \left[ 1 - \frac{{\Omega^2}} {2\pi G\overline{\rho}}\right] \left( \frac{\Omega^2r^3}{GM}\right)$ The term $\overline{\rho}$ means the average on the considered equipotential. The term with the minus sign in the square bracket is the Gratton–Öpik term, which becomes important in the outer layers when the local density is small. This term produces negative values of $U_2(r)$ (noted $U(r)$ from now), meaning that the circulation is going down along the polar axis and up in the equatorial plane. This makes an outward transport of angular momentum, while a positive $U(r)$ gives an inward transport. At lower $Z$, the Gratton–Öpik term is negligible, which contributes to make larger $\Omega$–gradients in lower $Z$ stars. Recently [@Mat04] rederived the system of partial differential equations, which govern the transport of angular momentum, heat and chemical elements. They expand the departure from spherical symmetry to higher order and include explicitly the differential rotation in latitude, to first order. Boundary conditions for the surface and at the frontiers between radiative and convective zones are also explicitly given in this paper. ### Shellular rotation. The differential rotation which results from the evolution and transport of the angular momentum makes the stellar interior highly turbulent. As explained above, the turbulence is very anisotropic, with a much stronger geostrophic–like transport in the horizontal direction than in the vertical one [@Z92], where stabilisation is favoured by the stable density gradient. This strong horizontal transport is characterized by a large diffusion coefficient $D_{\rm{h}}$. Various expressions have been proposed: - The usual expression for the coefficient $\nu_{\mathrm{h}}$ of viscosity due to horizontal turbulence and for the coefficient $D_{\mathrm{h}}$ of horizontal diffusion, which is of the same order, is, according to [@Z92], $$D_{\mathrm{h}} \simeq \nu_{\mathrm{h}} = \frac{1}{c_{\mathrm{h}}} r \;|2V(r) - \alpha U(r)| \,\, , \label{Zahn92}$$where $r$ is the appropriately defined eulerian coordinate of the isobar [@MM97]. $V(r)$ is defined by $u_\theta(r,\theta)=V(r) {{\rm d}P_2(\cos\theta) \over {\rm d}r}$ where $u_\theta$ is the horizontal component of the velocity of the meridional currents[^4], $\alpha = \frac{1}{2} \frac{d \ln r^{2} \Omega} {d \ln r}$ and $c_{\mathrm{h}}$ is a constant of order of unity or smaller. This equation was derived assuming that the differential rotation on an isobaric surface is small [@MZ98]. - [@Ma99] has derived an expression for the coefficient $D_{\mathrm{h}}$ of diffusion by horizontal turbulence in rotating stars. He has obtained $$\begin{aligned} D_{\mathrm{h}} \propto \; r \; \left(r \overline{\Omega}(r) \; V \; \left[ 2 V - \alpha U \right]\right)^\frac{1}{3} \;. \label{nuh}\end{aligned}$$This expression can be written in the usual form $\nu_{\mathrm{h}}= \frac{1}{3} \; l \cdot v$ for a viscosity, where the appropriate velocity $v$ is a geometric mean of 3 relevant velocities: a velocity $(2 V - \alpha U)$ as in Eq \[Zahn92\] by [@Z92], the horizontal component $V$ of the meridional circulation, the average local rotational velocity $ r \overline{\Omega}(r)$. This rotational velocity is usually much larger than either $U(r)$ or $V(r)$, typically by 6 to 8 orders of a magnitude in an upper Main Sequence star rotating with the average velocity. - From torque measurements in the classical Couette-Tayler experiment [@RZ99], [@Mati04] have found the following expression $$\begin{aligned} \nu_{\mathrm{h}} = \left({\beta \over 10}\right)^{1/2} \; \left(r^2 \overline{\Omega}(r) \; \left[r \left|2 V - \alpha U \right|\right]\right)^\frac{1}{2} \; , \label{mat}\end{aligned}$$with $\beta\approx 1.5 \times 10^{-5}$ [@RZ99]. The horizontal turbulent coupling favours an essentially constant angular velocity $\Omega$ on the isobars. This rotation law, constant on shells, applies to fast as well as to slow rotators. As an approximation, it is often represented by a law of the form $\Omega = \Omega (r)$ ([@Z92]; see also [@ES76]). Let us note here that the exact value of the diffusion coefficient $D_h$ is not well known. Indeed the values of the numerical factors intervening in the various expressions shown above may vary to some extent. Since the expression of $D_h$ intervenes in the formulas for $U_r$, for $D_{\rm shear}$ and $D_{\rm eff}$ (see below), these uncertainties have some impact on the amplitudes of the transport mechanisms. ### Shear turbulence and mixing. In a radiative zone, shear due to differential rotation is likely to be a most efficient mixing process. Indeed shear instability grows on a dynamical timescale that is of the order of the rotation period [@Z92]. The usual criterion for shear instability is the Richardson criterion, which compares the balance between the restoring force of the density gradient and the excess energy present in the differentially rotating layers, $$Ri = \frac{N^{2}_{\mathrm{ad}}}{(0.8836\ \Omega\frac {d\ln\Omega}{d\ln r})^2} < \frac {1}{4},$$ where we have taken the average over an isobar, $r$ is the radius and $N_{\rm{ad}}$ the Brunt-Väisälä frequency given by $$\begin{aligned} N^2_{\rm{ad}} = \frac{g \delta}{H_{P}} \left[ \frac{\varphi} {\delta} \nabla_{\mu} + \nabla_{ad} - \nabla_{\rm{rad}} \right].\end{aligned}$$ When thermal dissipation is significant, the restoring force of buoyancy is reduced and the instability occurs more easily. Its timescale is however longer, being the thermal timescale. This case is referred to as “secular shear instability”. The criterion for low Peclet numbers $Pe$ (i.e. of large thermal dissipation, see below) has been considered by [@Za74], while the cases of general Peclet numbers $Pe$ have been considered by [@Mae95], [@MM96], who give $$Ri = \frac{g \delta}{(0.8836\ \Omega\frac {d\ln\Omega}{d\ln r})^{2} H_{P}} \left[ \frac{\Gamma}{\Gamma +1} (\nabla_{ad} -\nabla) + \frac{\varphi}{\delta} \nabla_{\mu} \right] < \frac{1}{4}$$ The quantity $\Gamma = Pe/6 $, where the Peclet number $Pe$ is the ratio of the thermal cooling time to the dynamical time, i.e. $Pe = \frac{v \ell}{K}$ where $v$ and $\ell$ are the characteristic velocity and length scales, and $K = (4acT^3)/ (3 C_P \kappa \rho^2 )$ is the thermal diffusivity. A discussion of shear–driven turbulence by [@Ca98] suggests that the limiting $Ri$ number may be larger than $\frac{1}{4}$. To account for shear transport and diffusion, we need a diffusion coefficient. Amazingly, a great variety of coefficients $D_{\mathrm{shear}} = \frac{1}{3} v \ell$ have been derived and applied (see a more extended discussion in [@MMV]): 1. [@Z92] defines the diffusion coefficient corresponding to the eddies which have the largest $Pe$ number so that the Richardson criterion is just marginally satisfied. However, the effects of the vertical $\mu$–gradient are not accounted for and the expression only applies to low Peclet numbers. The same has been done by [@MM96], who considers also the effect of the vertical $\mu$–gradient, the case of general Peclet numbers and, in addition they account for the coupling due to the fact that the shear also modifies the local thermal gradient. This coefficient has been used by [@MM97] and by [@Den99]. The comparisons of model results and observations of surface abundances have led many authors to conclude that the $\mu$–gradients appear to inhibit the shear mixing too much with respect to what is required by the observations ([@Cha95a],[@MM97],[@He00]). 2. Instead of using a gradient $\nabla_{\mu}$ in the criterion for shear mixing, [@Cha95a] and [@He00] write $f_{\mu} \nabla_{\mu}$ with a factor $f_{\mu} = 0.05$ or even smaller. This procedure is not satisfactory since it only accounts for a small fraction of the existing $\mu$–gradients in stars. The problem is that the models depend at least as much (if not more) on $f_{\mu}$ than on rotation, [*i.e.*]{} a change of $f_{\mu}$ in the allowed range (between 0 and 1) produces as important effects as a change of the initial rotational velocity. This situation has led to two other more physical approaches discussed below. Also [@He00] introduces another factor $f_c$ to adjust the ratio of the transport of the angular momentum and of the chemical elements like [@Pin89]. 3. Around the convective core in the region where the $\mu$–gradient inhibits mixing, there is anyway some turbulence due to both the horizontal turbulence and to the semiconvective instability, which is generally present in massive stars. This situation has led to the hypothesis [@Mae97] that the excess energy in the shear, or a fraction $\alpha$ of it of the order of unity, is degraded by turbulence on the local thermal timescale. This progressively changes the entropy gradient and consequently the $\mu$–gradient. This hypothesis leads to a diffusion coefficient $D_{\rm shear}$ given by $$\begin{aligned} D_{\rm shear} = 4 \frac{K}{N^{2}_{\rm{ad}}} \left[ \frac{1}{4} \alpha \left(0.8836\ \Omega \frac{d\ln\Omega}{d\ln r} \right)^2 - (\nabla^{\prime} -\nabla) \right]. \label{Mae97}\end{aligned}$$ The term $\nabla^{\prime} -\nabla$ in Eq. \[Mae97\] expresses either the stabilizing effect of the thermal gradients in radiative zones or its destabilizing effect in semiconvective zones (if any). When the shear is negligible, $D_{\rm shear}$ tends toward the diffusion coefficient for semiconvection by [@La83] in semiconvective zones. When the thermal losses are large ($\nabla^{\prime} =\nabla$), it tends toward the value $$D_{\rm shear} = \alpha (K/N^2_{\rm{ad}}) \left(0.8836\ \Omega \frac{d\ln\Omega}{d\ln r}\right)^2 , \label{orig}$$ given by [@Z92]. Eq. \[Mae97\] is completed by the three following equations expressing the thermal effects [@Mae97] $$\begin{aligned} D_{\rm shear} = 2 K \Gamma \;\;\;\;\;\;\;\; \nabla=\frac{\nabla_{rad}+ (\frac{6 \Gamma^2}{1+\Gamma}) \nabla_{ad}}{1+(\frac{6 \Gamma^2}{1+\Gamma})}, \label{Mae972}\end{aligned}$$ $$\begin{aligned} \nabla^{\prime} -\nabla = \frac{\Gamma}{\Gamma +1} (\nabla_ {\mathrm{ad}} - \nabla). \label{Mae973}\end{aligned}$$ The system of 4 equations given by Eqs. \[Mae97\], \[Mae972\] and \[Mae973\] form a coupled system with 4 unknown quantities $D_{\rm shear}$, $\Gamma$, $\nabla$ and $\nabla^{\prime}$. The system is of the third degree in $\Gamma$. When it is solved numerically, we find that as a matter of fact the thermal losses in the shears are rather large in massive stars and thus that the Peclet number $Pe$ is very small (of the order of 10$^{-3}$ to 10$^{-4}$). For very low Peclet number $Pe =6 \Gamma$, the differences $(\nabla^{\prime} -\nabla)$ are also very small as shown by Eq. \[Mae973\]. Thus, we conclude that Eq. \[Mae97\] is essentially equivalent, at least in massive stars, to the original Eq. \[orig\] above, as given by [@Z92]. We may suspect that this is not necessarily true in low and intermediate mass stars since there the $Pe$ number may be larger. 4. [@TZ97] found that the diffusion coefficient for the shears is modified by the horizontal turbulence. The change can be an increase or a decrease of the diffusion coefficient depending on the various parameters, as discussed below. Thus, we have $$\begin{aligned} D = \frac{ (K + D_{\mathrm{h}})} {\left[\frac{\varphi}{\delta} \nabla_{\mu}(1+\frac{K}{D_{\mathrm{h}}})+ (\nabla_{\mathrm{ad}} -\nabla_{\mathrm{rad}}) \right] }\; \times \label{dsh} \\[2mm] \nonumber \frac{H_{\mathrm{p}}}{g \delta} \; \left [ \alpha\left( 0.8836\Omega{d\ln \Omega \over d\ln r} \right)^2 -4 (\nabla^{\prime} -\nabla) \right]\end{aligned}$$ where $D_{\mathrm{h}}$ is the coefficient of horizontal diffusion (cf. [@Z92]). We ignore here the thermal coupling effects discussed by Maeder ([@Mae97]) because they were found to be relatively small and they increase the numerical complexity. Interestingly, we see that in regions where $\nabla_{\mu} \simeq 0$, Eq. \[dsh\] leads us to replace $K$ by $(K+D_{\mathrm{h}})$ in the usual expression (cf. [@TZ97]), i.e. it reinforces slightly the diffusion in regions which are close to chemical homogeneity. On the contrary, in regions where $\nabla_{\mu}$ dominates with respect to $(\nabla_{\mathrm{ad}} -\nabla_{\mathrm{rad}})$, the transport is proportional to $D_{\mathrm{h}}$ rather than to $K$, which is quite logical since the diffusion is then determined by $D_{\mathrm{h}}$ rather than by thermal effects. The above result shows the importance of the treatment for the meridional circulation, since in turn it determines the size of $D_{\mathrm{h}}$ and to some extent the diffusion by shears. Of course, the Reynolds condition $D_{\rm shear} \geq \frac{1}{3} \nu Re_c$ must be satisfied in order that the medium is turbulent. The quantity $\nu$ is the total viscosity (radiative + molecular) and $Re_c$ the critical Reynolds number estimated to be around 10 (cf. [@Den99]; [@Z92]). The numerical results indicate that the conditions for the occurrence of turbulence are satisfied. ### Transport of the angular momentum Let us express the rate of change of the angular momentum, ${{\rm d}{\mathcal L}\over{\rm d}t}$, of the element of mass in the volume ABCD represented in Fig. \[schema\]: $${{\rm d}{\mathcal L}\over{\rm d}t}={\bf M},$$ where ${\bf M}$ is the momentum of the forces acting on the volume element. We assume that angular momentum is transported only through advection (by a velocity field [**U**]{}) and through turbulent diffusion, which may be different in the radial (vertical) and tangential (horizontal) direction. The component of the angular momentum aligned with the rotational axis is equal to [^5] $$\underbrace{\rho r^2 \sin\theta {\rm d}\theta {\rm d}\varphi {\rm d} r}_{\rm Mass\ of\ ABCD}\ \ \underbrace{r\sin\theta \Omega}_{\rm velocity}\underbrace{r\sin\theta ,}_{\rm distance\ to\ axis}$$ where $ \Omega=\dot\varphi$. Since the mass of the volume element ABCD does not change, the rate of change of the angular momentum can be written $$\begin{aligned} \rho r^2 \sin\theta {\rm d}\theta {\rm d}\varphi {\rm d} r {{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r}. \label{eqn1}\end{aligned}$$ Due to shear, forces apply on the surfaces of the volume element. The force on the surface AB is equal to $$\underbrace{\eta_v}_{\rm vertical\ viscosity}\ \ \underbrace{r \sin\theta {\partial\Omega\over\partial r}}_{\rm vertical\ shear} \underbrace{r^2\sin\theta{\rm d}\theta{\rm d}\varphi}_{\rm surface\ AB}.$$ The component of the momentum of this force along the rotational axis is $$\underbrace{\eta_v r^3\sin^2\theta {\partial\Omega\over\partial r}{\rm d}\theta{\rm d}\varphi}_{\rm force\ on\ AB} \ \underbrace{r\sin\theta .}_{\rm distance\ to\ axis}$$ The component along the rotational axis of the resultant momentum of the forces acting on AB and CD is equal to $$\begin{aligned} {\partial \over \partial r}(\eta_v r^4 \sin^3\theta {\rm d}\theta{\rm d}{\varphi}{\partial\Omega\over\partial r}){\rm d}r. \label{eqn2}\end{aligned}$$ The force on the surface AC due to the tangential shear is equal to $$\eta_h \underbrace{r\sin\theta {\partial \Omega\over r\partial\theta}}_{\rm tangential\ shear} \underbrace{r\sin\theta{\rm d}\varphi{\rm d}r }_{\rm surface\ AC},$$ where $\eta_h$ is the horizontal viscosity. The component along the rotational axis of the resultant momentum of the forces acting on AC and BD is equal to $$\begin{aligned} {\partial \over r \partial \theta}(\eta_h r^2\sin^3\theta{\rm d}r{\rm d}\varphi {\partial \Omega \over \partial \theta})r{\rm d}\theta. \label{eqn3}\end{aligned}$$ Using Eqs. \[eqn1\], \[eqn2\] and \[eqn3\], simplifying by ${\rm d}r{\rm d}\theta{\rm d}\varphi$, one obtains the equation for the transport of the angular momentum $$\begin{aligned} \rho r^2 \sin\theta {{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r}= {\partial \over \partial r}(\eta_v r^4 \sin^3\theta {\partial\Omega\over\partial r})+ {\partial \over \partial \theta}(\eta_h r^2\sin^3\theta {\partial \Omega \over \partial \theta}). \label{eqn4}\end{aligned}$$ Setting $\eta_v=\rho D_v$ and $\eta_h=\rho D_h$ and dividing the left and right member by $r^2\sin\theta$, one obtains $$\begin{aligned} \rho{{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r}={\sin^2\theta \over r^2} {\partial \over \partial r}(\rho D_v r^4 {\partial\Omega\over\partial r})+ {1\over \sin\theta} {\partial \over \partial \theta}(\rho D_h \sin^3\theta {\partial \Omega \over \partial \theta}). \label{eqn43}\end{aligned}$$ Now, the left–handside term can be written $$\rho{{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r}={{\rm d}\over {\rm d}t} (\rho r^2 \sin^2\theta \Omega)_{M_r}-r^2 \sin^2\theta \Omega{{\rm d}\rho\over {\rm d}t}|_{M_r}.$$ Using the relation between the Lagrangian and Eulerian derivatives, one has $$\begin{aligned} \rho{{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r}=\nonumber\end{aligned}$$ $$\begin{aligned} {\partial\over \partial t} (\rho r^2 \sin^2\theta \Omega)_{r} +{\bf U}\cdot{\bf \nabla}(\rho r^2 \sin^2\theta \Omega)-r^2 \sin^2\theta \Omega{{\rm d}\rho\over {\rm d}t}|_{M_r}. \label{eqn42}\end{aligned}$$ Using $${{\rm d} \rho \over {\rm d}t}|_{M_r}={\partial \rho \over \partial t}|_{r}+{\bf U} \cdot {\bf \nabla} \rho,$$ and the continuity equation $${\partial \rho \over \partial t}|_{r}=-{\rm div}(\rho {\bf U}),$$ one obtains ${\rm d}\rho/{\rm d}t|_{M_r}+\rho {\rm div}{\bf U}=0$, which incorporated in Eq. \[eqn42\] gives $$\rho{{\rm d}\over {\rm d}t} (r^2 \sin^2\theta \Omega)_{M_r} ={\partial \over \partial t} (\rho r^2 \sin^2\theta \Omega)_{r}+{\bf \nabla}({\bf U}\rho r^2 \sin^2\theta \Omega).$$ Developing the divergence in spherical coordinates and using Eq. \[eqn43\], one finally obtains the equation describing the transport of the angular momentum ([@MZ98]; [@Mat04]) $$\begin{aligned} {\partial\over \partial t} (\rho r^2 \sin^2\theta \Omega)_{r}+{1 \over r^2}{\partial \over \partial r}(\rho r^4\sin^2\theta w_r\Omega)+{1 \over r\sin\theta} {\partial \over \partial \theta}(\rho r^2\sin^3\theta w_{\theta} \Omega)= \nonumber \\ {\sin^2\theta \over r^2} {\partial \over \partial r}(\rho D_v r^4 {\partial\Omega\over\partial r})+ {1\over \sin\theta} {\partial \over \partial \theta}(\rho D_h \sin^3\theta {\partial \Omega \over \partial \theta}), \label{eqn5}\end{aligned}$$ where $w_r=U_r+\dot r$ is the sum of the radial component of the meridional circulation velocity and the velocity of expansion/contraction, and $w_\theta=U_\theta$, where $U_\theta$ is the horizontal component of the meridional circulation velocity. Assuming, as in [@Z92], that the rotation depends little on latitude due to strong horizontal diffusion, we write $$\Omega(r,\theta)=\bar\Omega(r)+\hat \Omega(r,\theta),$$ with $\hat\Omega \ll \bar\Omega$. The horizontal average $\bar \Omega$ is defined as being the angular velocity of a shell rotating like a solid body and having the same angular momentum as the considered actual shell. Thus $$\bar \Omega={\int \Omega \sin^3 \theta{\rm d} \theta \over \int \sin^3 \theta {\rm d}\theta}.$$ Any vector field whose Laplacian is nul can be decomposed in spherical harmonics. Thus, the meridional circulation velocity can be written [@Mat04] $${\bf U}=\underbrace{\sum_{l > 0} U_l (r) P_l (\cos\theta)}_{u_r} {\bf e}_r+ \underbrace{ \sum_{l > 0} V_l(r) {{\rm d} P_l (\cos\theta) \over {\rm d} \theta}}_{u_\theta}{\bf e}_\theta,$$ where ${\bf e_r}$ and ${\bf e_\theta}$ are unit vectors along the radial and colatitude directions respectively. Multiplying Eq. \[eqn5\] by $\sin\theta{\rm d}\theta$ and integrating it over $\theta$ from 0 to $\pi$, one obtains [@MZ98] $$\begin{aligned} {\partial \over \partial t}(\rho r^2 \bar \Omega)_r={1 \over 5 r^2}{\partial \over \partial r}(\rho r^4 \bar \Omega [U_2(r)-5\dot r]) +{1 \over r^2}{\partial \over \partial r}\left(\rho D_v r^4 {\partial \bar \Omega \over \partial r} \right). \label{eqn6}\end{aligned}$$ It is interesting to note that only the $l=2$ component of the circulation is able to advect a net amount of angular momentum. As explained in [@Spie92] the higher order components do not contribute to the vertical transport of angular momentum. Note also that the change in radius $\dot r$ of the given mass shell is included in Eq. \[eqn6\], which is the Eulerian formulation of the angular momentum transport equation. In its Lagrangian formulation, the variable $r$ is linked to $M_r$ through ${\rm d}M_r=4\pi r^2 \rho {\rm d}r$, and the equation for the transport of the angular momentum can be written $$\begin{aligned} \rho{\partial \over \partial t}(r^2 \bar \Omega)_{M_r}={1 \over 5 r^2}{\partial \over \partial r}(\rho r^4 \bar \Omega U_2(r)) +{1 \over r^2}{\partial \over \partial r}\left(\rho D_v r^4 {\partial \bar \Omega \over \partial r} \right). \label{eqn7}\end{aligned}$$ The characteristic time associated to the transport of $\Omega$ by the circulation is [@Z92] $$\begin{aligned} t_\Omega\approx t_{KH} \left({\Omega^2 R \over g_s}\right)^{-1}, \label{eqn8}\end{aligned}$$ where $g_s$ is the gravity at the surface and $t_{KH}$ the Kelvin–Helmholtz timescale, which is the characteristic timescale for the change of $r$ in hydrostatic models. From Eq. \[eqn8\], one sees that $t_\Omega$ is a few times $t_{KH}$, which itself is much shorter that the Main Sequence lifetime. For shellular rotation, the equation of transport of angular momentum in the vertical direction is in lagrangian coordinates (cf. [@Z92]; [@MZ98]) $$\begin{aligned} \lefteqn{\rho \frac{d}{d t} \left( r^2 \Omega\right)_{M_r} = } \nonumber \\[2mm] && \frac{1}{5 r^2} \frac{\partial}{\partial r} \left(\rho r^4 \Omega U(r) \right) + \frac{1}{r^2} \frac{\partial}{\partial r} \left(\rho D r^4 \frac{\partial \Omega}{\partial r} \right) . \label{full}\end{aligned}$$ $\Omega(r)$ is the mean angular velocity at level $r$. The vertical component $u(r,\theta)$ of the velocity of the meridional circulation at a distance $r$ to the center and at a colatitude $\theta$ can be written $$\begin{aligned} u(r,\theta)=U(r)P_2(\cos \theta), \label{urt}\end{aligned}$$ where $P_2(\cos \theta)$ is the second Legendre polynomial. Only the radial term $U(r)$ appears in Eq. \[full\]. The quantity $D$ is the total diffusion coefficient representing the various instabilities considered and which transport the angular momentum, namely convection, semiconvection and shear turbulence. As a matter of fact, a very large diffusion coefficient as in convective regions implies a rotation law which is not far from solid body rotation. In this work, we take $D = D_{\rm{shear}}$ in radiative zones, since as extra–convective mixing we consider shear mixing and meridional circulation. In case the outward transport of the angular momentum by the shear is compensated by an inward transport due to the meridional circulation, we obtain the local conservation of the angular momentum. We call this solution the *stationary solution*. In this case, $U(r)$ is given by (cf. [@Z92]) $$U(r)= - \frac{5 D}{\Omega} \frac{\partial \Omega}{\partial r} \; . \label{stati}$$ The full solution of Eq. \[full\] taking into account $U(r)$ and $D$ gives the *non–stationary solution* of the problem. In this case, $\Omega (r)$ evolves as a result of the various transport processes, according to their appropriate timescales, and in turn differential rotation influences the various above processes. This produces a feedback and, thus, a self–consistent solution for the evolution of $\Omega (r)$ has to be found. Fig. \[Ur\] shows the evolution of $U(r)$ in a model of a 20 M$_{\odot}$ star with $Z$ = 0.004 and an initial rotation velocity $v_{\mathrm{ini}}$ = 300 km s$^{-1}$ [@MMVII]. $U(r)$ is initially positive in the interior, but progressively the fraction of the star where $U(r)$ is negative is growing. This is due to the Gratton–Öpik term in Eq. (2), which favors a negative $U(r)$ in the outer layers, when the density decreases. This negative velocity causes an outward transport of the angular momentum, as well as the shears[^6]. The transport of angular momentum by circulation has often been treated as a diffusion process ([@ES76]; [@Pin89]; [@He00]). From Eq. \[full\], we see that the term with $U$ (advection) is functionally not the same as the term with $D$ (diffusion). Physically advection and diffusion are quite different: diffusion brings a quantity from where there is a lot to other places where there is little. This is not necessarily the case for advection. A circulation with a positive value of $U(r)$, i.e. rising along the polar axis and descending at the equator, is as a matter of fact making an inward transport of angular momentum. Thus, we see that when this process is treated as a diffusion, like a function of $\frac{\partial \Omega}{\partial r}$, even the sign of the effect may be wrong. The expression of $U(r)$ given above (Eq. \[Umer\]) involves derivatives up to the third order, thus Eq. \[full\] is of the fourth order, which makes the system very difficult to solve numerically. In practice, we have applied a Henyey scheme to make the calculations. Eq. \[full\] also implies four boundary conditions. At the stellar surface, we take (cf. [@Ta97]) $$\begin{aligned} \frac{\partial \Omega}{\partial r} = 0 \;\;\;\; \mathrm{and} \;\;\; \; U(r) =0 \label{BE}\end{aligned}$$ and at the edge of the core we have $$\begin{aligned} \frac{\partial \Omega}{\partial r} = 0 \; \; \; \; \mathrm{and} \; \; \; \; \Omega(r) = \Omega_{\mathrm{core}.}\end{aligned}$$ We assume that the mass lost by stellar winds is just embarking its own angular momentum. This means that we ignore any possible magnetic coupling, as it occurs in low mass stars. It is interesting to mention here, that in case of no viscous, nor magnetic coupling at the stellar surface, [*i.e.*]{} with the boundary conditions \[BE\], the integration of Eq. \[full\] gives for an external shell of mass $\Delta M$ [@Ma99] $$\begin{aligned} \Delta M {d \over dt} (\Omega r^2)=-{4\pi \over 5} \rho r^4 \Omega U(r).\end{aligned}$$ This equation is valid provided the stellar winds are spherically symmetric. When the surface velocity approches the critical velocity, it is likely that there are anisotropies of the mass loss rates (polar ejection or formation of an equatorial ring) and thus the surface condition should be modified according to the prescriptions of [@Ma99]. ### Mixing and transport of the chemical elements. A diffusion–advection equation like Eq. \[full\] should normally be used to express the transport of chemical elements. However, if the horizontal component of the turbulent diffusion $D_{\rm{h}}$ is large, the vertical advection of the elements can be treated as a simple diffusion [@Cha92] with a diffusion coefficient $D_{\rm eff}$. As emphasized by [@Cha92] , this does not apply to the transport of the angular momentum. $D_{\rm eff}$ is given by $$D_{\rm eff} = \frac{\mid rU(r) \mid^2}{30 D_h} \; , \label{deff}$$ where $D_{\rm{h}}$ is the coefficient of horizontal turbulence. Eq. \[deff\] expresses that the vertical advection of chemical elements is severely inhibited by the strong horizontal turbulence characterized by $D_{\rm{h}}$. Thus, the change of the mass fraction $X_i$ of the chemical species $i$ is simply $$\begin{aligned} \left( \frac{dX_i}{dt} \right)_{M_r} = \left(\frac{\partial }{\partial M_r} \right)_t \left[ (4\pi r^2 \rho)^2 D_{\rm mix} \left( \frac{\partial X_i} {\partial M_r}\right)_t \right] + \left(\frac{d X_i}{dt} \right)_{\rm nucl} . \label{fullxi}\end{aligned}$$ The second term on the right accounts for composition changes due to nuclear reactions. The coefficient $D_{\rm mix}$ is the sum $D_{\rm mix} = D_{\rm shear}+D_{\rm eff}$ and $D_{\rm eff}$ is given by Eq. \[deff\]. The characteristic time for the mixing of chemical elements is therefore $t_{\rm mix} \simeq \frac{R^2}{D_{\rm mix}} $ and is not given by $t_{\rm circ} \simeq \frac{R}{U}$, as has been generally considered [@Sch58]. This makes the mixing of the chemical elements much slower, since $D_{\rm eff}$ is very much reduced. In this context, we recall that several authors have reduced by large factors, up to 30 or 100, the coefficient for the transport of the chemical elements, with respect to the transport of the angular momentum, in order to better fit the observed surface compositions (cf. [@He00]). This reduction of the diffusion of the chemical elements is no longer necessary with the more appropriate expression of $D_{\rm eff}$ given here. Surface enrichments due to rotation are illustrated in Fig \[ncp\]. The tracks are plotted in the plane ${\rm (N/C)/(N/C)_{ini}}$ versus $P$ where $P$ is the rotational period in hours. During the evolution the surface is progressively enriched in CNO burning products, [*i.e.*]{} is enriched in nitrogen and depleted in carbon. At the same time, the rotational period increases. ![Evolutionary tracks in the plane surface N/C ratio, normalized to its initial value, versus the rotational period in hours for different initial mass stars, various initial velocities and for the metallicities $Z$=0.02 and 0.002. Positions of some periods in days are indicated at the bottom of the figure. The dotted tracks never reach the critical limit during the MS phase. The short dashed tracks reach the critical limit during the MS phase. The dividing line between the shaded and non-shaded areas corresponds to the entrance into the phase when the star is at the critical limit during the MS phase. If Be stars are stars rotating at or very near the critical limit, present models would predict that they would lie in the vicinity of this dividing line or above it. Note the different vertical scales used when comparing similar masses at different metallicities. Figure taken from [@EMM].[]{data-label="ncp"}](ncplow.eps){height="0.755\textheight"} When the effects of the shear and of the meridional circulation compensate each other for the transport of the angular momentum (*stationary solution*), the value of $U$ entering the expression for $D_{\rm eff}$ is given by Eq. \[stati\]. Rotation and mass loss ---------------------- We can classify the effects of rotation on mass loss in three categories. 1. The structural effects of rotation. 2. The changes brought by rotation on the radiation driven stellar winds. 3. The mass loss induced by rotation at the critical limit. Let us now consider in turn these various processes. ### Structural effects of rotation on mass loss. Rotation, by changing the chemical structure of the star, modifies its evolution. For instance, moderate rotation at metallicities of the Small Magellanic Cloud (SMC) favors redward evolution in the Hertzsprung-Russel diagram. This behavior can account for the high number of red supergiants observed in the SMC [@MMVII], an observational fact which is not at all reproduced by non-rotating stellar models. Now it is well known that the mass loss rates are greater when the star evolves into the red part of the HR diagram, thus in this case, rotation modifies the mass loss indirectly, by changing the evolutionary tracks. The $\upsilon_{\rm ini}=0$, 200, 300 and 400 km s$^{-1}$ models lose respectively 0.14, 1.40, 1.71 and 1.93 M$_\odot$ during the core He-burning phase (see Table 1 in [@MMVII]). The enhancement of the mass lost reflects the longer lifetimes of the red supergiant phase when velocity increases. Note that these numbers were obtained assuming that the same scaling law between mass loss and metallicity as in the MS phase applies during the red supergiant phase. If, during this phase, mass loss comes from continuum-opacity driven wind then the mass-loss rate will not depend on metallicity (see the review by [@vL06]). In that case, the redward evolution favored by rotation would have a greater impact on mass loss than that shown by the computations shown above. Of course, such a trend cannot continue forever. For instance, at very high rotation, the star will have a homogeneous evolution and will never become a red supergiant [@M87]. In this case, the mass loss will be reduced, although this effect will be somewhat compensated by other processes: first by the fact that the Main-Sequence lifetime will last longer, second, by the fact that the star will enter the Wolf-Rayet phase (a phase with high mass loss rates) at an earlier stage of its evolution, and third by the fact that the star may encounter the $\Omega$-limit. ### Radiation driven stellar winds with rotation. The effects of rotation on the radiation driven stellar winds result from the changes brought by rotation to the stellar surface. They induce changes of the morphologies of the stellar winds and increase their intensities. Naively we would first guess that a rotating star would lose mass preferentially from the equator, where the effective gravity (gravity decreased by the effect of the centrifugal force) is lower. This is probably true when the star reaches the $\Omega$-limit (i.e. when the equatorial surface velocity is such that the centrifugal acceleration exactly compensates the gravity), but this is not correct when the star is not at the critical limit. Indeed as recalled above, a rotating star has a non uniform surface brightness, and the polar regions are those which have the most powerful radiative flux. Thus one expects in case the opacity does not vary at the surface, that the star will lose mass preferentially along the rotational axis. This is correct for hot stars, for which the dominant source of opacity is electron scattering. In that case the opacity only depends on the mass fraction of hydrogen and does not depends on other physical quantities such as temperature. In that way, rotation induces anisotropies of the winds ([@MD01];[@DO02]). This is illustrated in the left panel of Fig. \[ani\]. Wind anisotropies have consequences for the angular momentum that a star retains in its interior. Indeed, when mass is lost preferentially along the polar axis, little angular momentum is lost. This process allows loss of mass without too much loss of angular momentum a process which might be important in the context of the evolutionary scenarios leading to Gamma Ray Bursts. Indeed in the framework of the collapsar scenario ([@W93]), one has to accommodate two contradictory requirements: on one side, the progenitor needs to lose mass in order to have its H and He-rich envelope removed at the time of its explosion, and on the other hand it must have retained sufficient angular momentum in its central region to give birth to a fast rotating black-hole. The quantity of mass lost through radiatively driven stellar winds is enhanced by rotation. This enhancement can occur through two channels: by reducing the effective gravity at the surface of the star, by increasing the opacity of the outer layers through surface metallicity enhancements due to rotational mixing. - [*reduction of the effective gravity:* ]{} The ratio of the mass loss rate of a star with a surface angular velocity $\Omega$ to that of a non-rotating star, of the same initial mass, metallicity and lying at the same position in the HR diagram is given by [@mm6] $$\frac{\dot{M} (\Omega)} {\dot{M} (0)} \simeq \frac{\left( 1 -\Gamma\right) ^{\frac{1}{\alpha} - 1}} {\left[ 1 - \frac{4}{9} (\frac{v}{v_{\mathrm{crit, 1}}})^2-\Gamma \right] ^{\frac{1}{\alpha} - 1}} \; ,$$ where $\Gamma$ is the electron scattering opacity for a non–rotating star with the same mass and luminosity, $\alpha$ is a force multiplier [@La95]. The enhancement factor remains modest for stars with luminosity sufficiently far away from the Eddington limit [@mm6]. Typically, $\frac{\dot{M} (\Omega)} {\dot{M} (0)} \simeq 1.5$ for main-sequence B–stars. In that case, when the surface velocity approaches the critical limit, the effective gravity decreases and the radiative flux also decreases. Thus the matter becomes less bound when, at the same time, the radiative forces become also weaker. When the stellar luminosity approaches the Eddington limit, the mass loss increases can be much greater, reaching orders of magnitude. This comes from the fact that rotation lowers the maximum luminosity or the Eddington luminosity of a star. Thus it may happen that for a velocity still far from the classical critical limit, the rotationally decreased maximum luminosity becomes equal to the actual luminosity of the star. In that case, strong mass loss ensues and the star is said to have reached the $\Omega\Gamma$ limit [@mm6]. - [*Effects due to rotational mixing:* ]{} During the core helium burning phase, at low metallicity, the surface may be strongly enriched in both H-burning and He-burning products, [*i.e.*]{} mainly in nitrogen, carbon and oxygen. Nitrogen is produced by transformation of the carbon and oxygen produced in the He-burning core and which have diffused by rotational mixing in the H-burning shell [@MMVIII]. Part of the carbon and oxygen produced in the He-core also diffuses up to the surface. Thus at the surface, one obtains very high value of the CNO elements. For instance a 60 M$_\odot$ with Z=$10^{-8}$ and $\upsilon_{\rm ini}=800$ km s$^{-1}$ has, at the end of its evolution, a CNO content at the surface equivalent to 1 million times its initial metallicity! In case the usual scaling laws linking the surface metallicity to the mass loss rates is applied, such a the star would lose due to this process more than half of its initial mass. ### Mass loss induced by rotation As recalled above, during the Main-Sequence phase the core contracts and the envelope expands. In case of local conservation of the angular momentum, the core would thus spin faster and faster while the envelope would slow down. In that case, it can be easily shown that the surface velocity would evolve away from the critical velocity (see e.g. [@Vl06]). In models with shellular rotation however an important coupling between the core and the envelope is established through the action of the meridional currents. As a net result, angular momentum is brought from the inner regions to the outer ones. Thus, would the star lose no mass by radiation driven stellar winds (as is the case at low Z), one expects that the surface velocity would increase with time and would approach the critical limit. In contrast, when radiation driven stellar winds are important, the timescale for removing mass and angular momentum at the surface is shorter than the timescale for accelerating the outer layers by the above process and the surface velocity decreases as a function of time. It evolves away from the critical limit. Thus, an interesting situation occurs: when the star loses little mass by radiation driven stellar winds, it has more chance to lose mass by reaching the critical limit. On the other hand, when the star loses mass at a high rate by radiation driven mass loss then it has no chance to reach the critical limit and thus to undergo a mechanical mass loss. This is illustrated in the right panel of Fig. \[ani\]. ![[*Left panel*]{}: Iso-mass loss distribution for a 120 M$_\odot$ star with Log L/L$_\odot$=6.0 and T$_{\rm eff}$ = 30000 K rotating at a fraction 0.8 of critical velocity (figure from [@MD01]). [*Right panel*]{}: Evolution of the surface velocities for a 60 M$_ {\odot}$ star with 3 different initial metallicities. []{data-label="ani"}](V60Z.eps){width="2.5in" height="2.5in"} ### Discussion. At this point it is interesting to discuss three aspects of the various effects described above. First what are the main uncertainties affecting them? Second, what are their relative importance? And finally what are their consequences for the interstellar medium enrichment? In addition to the usual uncertainties affecting the radiation driven mass loss rates, the above processes poses three additional problems: 1. [*What does happen when the CNO content of the surface increases by six orders of magnitude as was obtained in the 60 M$_\odot$ model described above?*]{} Can we apply the usual scaling law between Z and the mass losses? This is what we have done in our models, but of course this should be studied in more details by stellar winds models. For instance, for WR stars, [@V05] have shown that at $Z=Z_\odot/30$, 60% of the driving is due to CNO elements and only 10% to Fe. Here the high CNO surface enhancements result from rotational mixing which enrich the radiative outer region of the star in these elements, but also from the fact that the star evolves to the red part of the HR diagram, making an outer convective zone to appear. This convective zone plays an essential role in dredging up the CNO elements at the surface. Thus what is needed here is the effects on the stellar winds of CNO enhancements in a somewhat red part of the HR diagram (typical effective temperatures of the order of Log T$_{\rm eff}\sim$3.8). 2. [*Do stars can reach the critical limit?*]{} For instance, [@BO70] obtain that during pre-main sequence evolution of rapidly rotating massive stars, “equatorial mass loss” or “rotational mass ejection” never occur (see also [@BO73]). In these models the condition of zero effective gravity is never reached. However, these authors studied pre-main sequence evolution and made different hypotheses on the transport mechanisms than in the present work. Since they were interested in the radiative contraction phase, they correctly supposed that “the various instabilities and currents which transport angular momentum have characteristic times much longer than the radiative-contraction time”. This is no longer the case for the Main-Sequence phase. In our models, we consistently accounted for the transport of the angular momentum by the meridional currents and the shear instabilities. A detailed account of the transport mechanisms shows that they are never able to prevent the star from reaching the critical velocity. Another difference between the approach in the work of [@BO70] and ours is that [@BO70] consider another distribution of the angular velocity than in our models. They supposed constant $\Omega$ on cylindrical surface, while here we adopted, as imposed by the theory of [@Z92], a “shellular rotation law”. They resolved the Poisson equation for the gravitational potential, while here we adopted the Roche model. Let us note that the Roche approximation appears justified in the present case, since only the outer layers, containing little mass, are approaching the critical limit. The majority of the stellar mass has a rotation rate much below the critical limit and is thus not strongly deformed by rotation. Thus these differences probably explain why in our models we reach situations where the effective gravity becomes zero. 3. [*What does happen when the surface velocity reaches the critical limit?*]{} Let us first note that when the surface reaches the critical velocity, the energy which is still needed to make equatorial matter to escape from the potential well of the star is still important. This is because the gravity of the system continues of course to be effective all along the path from the surface to the infinity and needs to be overcome. If one estimates the escape velocity from the usual equation energy for a piece of material of mass $m$ at the equator of a body of mass $M$, radius $R$ and rotating at the critical velocity, $${1 \over 2}m \upsilon_{\rm crit}^2+{1 \over 2}m \upsilon_{\rm esc}^2-{GMm \over R}=0,$$ one obtains, using $\upsilon_{\rm crit}^2={GM/R}$ that the escape velocity is simply reduced by a factor $1/\sqrt{2}=0.71$ with respect to the escape velocity from a non-rotating body [^7]. Thus the reduction is rather limited and one can wonder if matter will be really lost. A way to overcome this difficulty is to consider the fact that, at the critical limit, the matter will be launched into a keplerian orbit around the star. Thus, probably, when the star reaches the critical limit an equatorial disk is formed like for instance around Be stars. Here we suppose that this disk will eventually dissipate by radiative effects and thus that the material will be lost by the star. Practically, in the present models, we remove the supercritical layers. This removal of material allows the outer layers to become again subcritical at least until secular evolution will bring again the surface near the critical limit (see [@MEM06] for more details in this process). Secular evolution during the Main-Sequence phase triggers two counteracting effects: on one side, the stellar surface expands. Local conservation of the angular momentum makes the surface to slow down and the surface velocity to evolve away from the critical limit. On the other hand, meridional circulation continuously brings angular momentum to the surface and accelerates the outer layers. This last effect in general overcomes the first one and the star rapidly reach again the critical limit. How much mass is lost by this process? As seen above, the two above processes will maintain the star near the critical limit for most of the time. In the models, we adopt the mass loss rate required to maintain the star at about 95-98% of the critical limit. Such a mass loss rate is imposed as long as the secular evolution brings back the star near the critical limit. In general, during the Main-Sequence phase, once the critical limit is reached, the star remains near this limit for the rest of the Main-Sequence phase. At the end of the Main-Sequence phase, evolution speeds up and the local conservation of the angular momentum overcomes the effects due to meridional currents, the star evolves away from the critical limit and the imposed “critical” mass loss is turned off. “Spinstars” at very low metallicities? ====================================== Let us call “spinstars” those stars with a sufficiently high initial rotation in order to have their evolution significantly affected by rotation. In this section, we present some arguments supporting the view according to which spinstars might have been more common in the first generations of stars in the Universe. A direct way to test this hypothesis would be to obtain measures of surface velocity of very metal poor massive stars and to see whether their rotation is superior to those measured at solar metallicity. At the moment, such measures can be performed only for a narrow range of metallicities for $Z$ between 0.002 and 0.020. Interestingly already some effects can be seen. For instance [@Keller04] presents measurements of the projected rotational velocities of a sample of 100 early B-type main-sequence stars in the Large Magellanic Cloud (LMC). He obtains that the stars of the LMC are more rapid rotators than their Galactic counterparts and that, in both galaxies, the cluster population exhibits significantly more rapid rotation than that seen in the field (a point also recently obtained by [@HG06]). More recently [@Martayan07] obtain that the angular velocities of B (and Be stars) are higher in the SMC than in the LMC and MW. For B-type stars, the higher values obtained at lower $Z$ can be the result of two processes: 1) the process of star formation produces more rapid rotators at low metallicity; 2) the mass loss being weaker at low $Z$, less angular momentum is removed from the surface and thus starting from the same initial velocity, the low $Z$ star would be less slowed down by the winds. In the case of B-type stars, the mass loss rates are however quite modest and we incline to favor the first hypothesis, [*i.e.*]{} a greater fraction of fast rotators at birth at low metallicity. Another piece of argument supporting this view is the following: in case the mass loss rates are weak (which is the case on the MS phase for B-type stars), then the surface velocity is mainly determined by two processes, the initial value on the ZAMS and the efficiency of the angular momentum transport from the core to the envelope. In case of very efficient transport, the surface will receive significant amount of angular momentum transported from the core to the envelope. The main mechanism responsible for the transport of the angular momentum is meridional circulation. The velocities of the meridional currents in the outer layers are smaller when the density is higher thus in more metal poor stars. Therefore, starting from the same initial velocity on the ZAMS, one would expect that B-type stars at solar metallicity (with weak mass loss) would have higher surface velocities than the corresponding stars at low $Z$. The opposite trend is observed. Thus, in order to account for the higher velocities of B-type stars in the SMC and LMC, in the frame of the present rotating stellar models, one has to suppose that stars on the ZAMS have higher velocities at low $Z$. Very interestingly, the fraction of Be stars (stars rotating near the critical velocity) with respect to the total number of B stars is higher at low metallicity ([@MaederGrebel99]; [@Wisniewski06]). This confirms the trend discussed above favoring a higher fraction of fast rotators at low $Z$. There are at least four other striking observational facts which might receive an explanation based on massive fast rotating models. [*First*]{}, indirect observations indicate the presence of very helium-rich stars in the globular cluster $\omega$Cen [@Piotto05]. Stars with a mass fraction of helium, $Y$, equal to 0.4 seem to exist, together with a population of normal helium stars with $Y=0.25$. Other globular clusters appear to host helium-rich stars [@Caloi07], thus the case of $\omega$Cen is the most spectacular but not the only one. There is no way for these very low mass stars to enrich their surface in such large amounts of helium and thus they must have formed from protostellar cloud having such a high amount of helium. Where does this helium come from? We proposed that it was shed away by the winds of metal poor fast rotating stars [@MMocen]. [*Second*]{}, in globular clusters, stars made of material only enriched in H-burning products have been observed (see the review by [@Gratton04]). Probably these stars are also enriched in helium and thus this observation is related to the one reported just above. The difference is that proper abundance studies can be performed for carbon, nitrogen, oxygen, sodium, magnesium, lithium, fluorine …, while for helium only indirect inferences based on the photometry can be made. [@DecressinI] propose that the matter from which the stars rich in H-burning products are formed, has been released by slow winds of fast rotating massive stars. Of course, part of the needed material can also be released by AGB stars. The massive star origin presents however some advantages: first a massive star can induce star formation in its surrounding, thus two effects, the enrichment and the star formation can be triggered by the same cause. Second, the massive star scenario allows to use a less flat IMF than the scenario invoking AGB stars [@PC06]. The slope of the IMF might be even a Salpeter’s one in case the globular cluster lost a great part of its first generation stars by tidal stripping (see [@DCM07]). [*Third*]{}, the recent observations of the surface abundances of very metal poor halo stars[^8] show the need of a very efficient mechanism for the production of primary nitrogen [@Chiappinial05]. As explained in [@Chiappinial06]), a very nice way to explain this very efficient primary nitrogen production is to invoke fast rotating massive stars. Very interestingly, fast rotating massive stars help not only in explaining the behavior of the N/O ratio at low metallicity but also those of the C/O. Predictions for the behaviour of the $^{12}$C/$^{13}$C ratios at the surface of very metal poor non-evolved stars have also been obtained [@Chiappinial08]. [*Fourth*]{}, below about \[Fe/H\] $<$ -2.5, a significant fraction of very iron-poor stars are C-rich (see the review by [@BC05]). Some of these stars show no evidence of $s$-process enrichments by AGB stars and are thus likely formed from the ejecta of massive stars. The problem is how to explain the very high abundances with respect to iron of CNO elements. [@MEM06] and [@Hi07] proposed that these stars might be formed from the winds of very metal poor fast rotating stars. It is likely that rotation also affects the composition of the ejecta of intermediate mass stars. [@MEM06] predict the chemical composition of the envelope of a 7 M$_\odot$ E-AGB star which have been enriched by rotational mixing. The composition presents striking similarities with the abundance patterns observed at the surface of CRUMPS. The presence of overabundances of fluorine and of $s$-process elements might be used to discriminate between massive and intermediate mass stars. All the above observations seem to point toward the same direction, an important population of spinstars at low Z. How many? What is the origin of the fast rotation? What are the consequences for the Gamma ray Burst progenitors? All these questions have still to be addressed in a quantitative way and offer nice perspective for future works. Acknowledgment {#acknowledgment .unnumbered} ============== My warm thanks to André Maeder whose enlightened theoretical developments allowed to explore the effects of rotation in stellar models. [8.]{} Beers, T.C., Christlieb, N. 2005, ARAA, 43, 531 P.Bodenheimer, P., Ostriker, J.P. *Rapidly Rotating Stars.VI. Pre-Main - Evolution of Massive Stars*, *ApJ*, 161, 1101 (1970) P.Bodenheimer, P., Ostriker, J.P. *Rapidly Rotating Stars. VIII. Zero-Viscosity Polytropic Sequences*, *ApJ*, 180, 159 (1973) Caloi, V., D’Antona, F. 2007, A&A, 463, 949 Canuto V.M., 1998, ApJ 508, 767 Chaboyer, B., Zahn, J.–P. 1992, A&A 253, 173 Chaboyer B., Demarque P., Pinsonneault M.H., 1995a, ApJ 441, 865 Chiappini, C., Matteucci, F. & Ballero, S.K. 2005, A&A, 437, 429 Chiappini, C., Hirschi, R., Meynet, G., Ekstroem, S., Maeder, A., Matteucci, F. 2006, A&A Letters, 449, 27 Chiappini, C., Ekstroem, S.,Hirschi, R., Meynet, G., Maeder, A., Charbonnel, C. 2008, A&A Letters, in press Decressin, T., Meynet, G., Charbonnel, C., Prantzos, N., Ekstr[ö]{}m, S. 2007, A&A, 464, 1029 Decressin, T., Charbonnel, C., Meynet, G. 2007, A&A, 475, 859 Denissenkov P.A., Ivanova N.S., Weiss A., 1999, A&A 341, 181 Dwarkadas, V.V., Owocki, S.P. 2002, ApJ, 581, 1337 Endal A.S., Sofia S., 1976, ApJ 210, 184 Ekström S., Meynet, G., Maeder A. 2008, A&A, in press, (astro-ph/0711.1735) Glatzel W., 1998, A&A 339, L5 Gratton, R., Sneden, C., Carretta, E. 2004, ARAA, 42, 385 Heger A., Langer N., Woosley S.E., 2000, ApJ 528, 368 Hirschi, R. 2007, A&A, 461, 571 Huang, W., Gies, D.R. 2006, ApJ, 648, 591 Keller, S. C. 2004, PASP, 21, 310 Kippenhahn R., Thomas H.C., 1970, A Simple Method for the Solution of the Stellar Structure Equations Including Rotation and Tidal Forces. In: Slettebak A. (ed.) Proc. IAU Coll. 4, Stellar Rotation. Gordon and Breach Science Publishers, p. 20 Lamers, H.J.G.L.M., Snow, T.P., Lindholm, D.M. 1995, ApJ, 455, 269 Langer N., Fricke K.J., Sugimoto D., 1983, A&A 126, 207 Maeder, A. 1987, A&A, 158, 179 Maeder A., 1995, A&A 299, 84 Maeder A., 1997, A&A 321, 134 (Paper II) Maeder A., 1999, A&A 347, 185 (Paper IV) Maeder, A. 2003, A&A, 399, 263 Maeder, A., Desjacques, V. 2001, A&A, 372, L9 Maeder A., Meynet G., 1996, A&A 313, 140 Maeder, A., & Meynet, G. 2001, A&A, 373, 555, (Paper VII) Maeder, A., Meynet, G. 2006, A&A, 448, L37 Maeder A., Peytremann E., 1970, A&A 7, 120 Maeder A., Zahn J.P., 1998, A&A 334, 1000 (Paper III) Maeder, A.,Grebel, E. K., Mermilliod, J.-C. 1999, A&A, 346, 459 Maeder A., Georgy C., Meynet G. 2008, A&A, in press, (astro-ph/0801.1018) Maeder, A., & Meynet, G. 2000, A&A, 361, 159, (Paper VI) Martayan, C., Fr[é]{}mat, Y., Hubert, A.-M., Floquet, M., Zorec, J., Neiner, C. 2007, A&A, 462, 683 Mathis, S., Zahn, J.–P. 2004, A&A, 425, 229 Mathis, S., Palacios, A., Zahn, J.–P. 2004, A&A, 425, 243 Mestel, L. 1953, MNRAS, 113, 716 Meynet G., Maeder A., 1997, A&A 321, 465 (Paper I) Meynet, G., & Maeder, A. 2000, A&A 361, 101, (Paper V) Meynet, G., & Maeder, A. 2002, A&A, 390, 561, (Paper VIII) G. Meynet, A. Maeder, in proceedings of *Stars with the B\[e\] phenomenon*, M. Kraus & A.S. Miroshnichenko (eds), ASP Conf. Series, in press (astro-ph/0511269) G. Meynet, S. Ekström, A. Maeder, *The early star generations: the dominant effect of rotation on the CNO yields*, *A&A* [**447**]{}, 623 (2006) Pinsonneault M.H., Kawaler S.D., Sofia S., Demarque P., 1989, ApJ 338, 424 Piotto, G., Villanova, S., Bedin, L. R., Gratton, R., Cassisi, S., Momany, Y., Recio-Blanco, A., Lucatello, S., Anderson, J., King, I. R., Pietrinferni, A., Carraro, G. 2005, ApJ, 621, 777 Prantzos, N., Charbonnel, C. 2006, A&A, in press (2006), (astro-ph/0606112) Richard, D., Zahn, J.–P. 1999, A&A, 347, 734 Schwarzschild, M. 1958, Structure and evolution of the stars, Princeton, Princeton University Press Spiegel, E., Zahn, J.–P. 1992, A&A, 265, 106 Talon S., Zahn J.P., 1997, A&A 317, 749 Talon S., Zahn J.P., Maeder A., Meynet G., 1997, A&A, 322, 209 Tassoul J.L., 1990, The Effects of Rotation on Stellar Structure and Evolution. In: Willson L.A., Stalio R. (eds.) Angular Momentum and Mass Loss for Hot Stars. Kluwer Acad. Publ., p. 7 van Loon, J.Th., 2006, in *Stellar Evolution at Low Metallicity: Mass Loss, Explosions, Cosmology*, H.J.G.L.M. Lamers, N. Langer, T. Nugis, K. Annuk, ASP Conf. in press, (astro-ph/0512326) J.S. Vink, A. de Koter, *On the metallicity dependence of Wolf-Rayet winds*, *A&A*, [**442**]{}, 587 (2005) von Zeipel H. 1924, MNRAS 84, 665 Wisniewski, J. P., Bjorkman, K. S. 2006, ApJ, 652, 458 Woosley, S.E. 1993, ApJ, 405, 273 Zahn J.P., 1974, Rotational instabilities and stellar evolution. In: Ledoux P. (ed.) Proc. IAU Symp. 59, Stellar instability and evolution. Reidel, Dordrecht, p. 185 Zahn, J.-P. 1992, A&A, 265, 115 [^1]: For shellular rotation, the shape of the isobaric surfaces are given by the same expression as the one giving the shape of the equipotentials in conservative cases provided some changes of variables are performed. [^2]: $\vec{g_\mathrm{grav}}$ is the gravitational acceleration. [^3]: $U_2$ is the same as $U$ in Eq. \[urt\] [^4]: $V(r)$ can be obtained from $U(r)$ see Eq. 2.10 in [@Z92] [^5]: The components perpendicular to the rotational axis cancel each other when the integration is performed over $\varphi$. [^6]: When U is negative, the meridional currents turn anticlockwise, [*i.e.*]{} go inwards along directions parallel to the rotational axis and go outwards in directions parallel to the equatorial plane. [^7]: We suppose here that the vector $\upsilon_{\rm esc}$ is normal to the direction of the vector $\upsilon_{\rm crit}$. [^8]: These stars are in the field and present \[Fe/H\] as low as -4, thus well below the metallicities of the globular clusters.
{ "pile_set_name": "ArXiv" }
--- abstract: | We define and study fractional versions of the well-known Gamma subordinator $\Gamma :=\{\Gamma (t),$ $t\geq 0\},$ which are obtained by time-changing $\Gamma $ by means of an independent stable subordinator or its inverse. Their densities are proved to satisfy differential equations expressed in terms of fractional versions of the shift operator (with fractional parameter greater or less than one, in the two cases). As a consequence, the fractional generalization of some Gamma subordinated processes (i.e. the Variance Gamma, the Geometric Stable and the Negative Binomial) are introduced and the corresponding fractional differential equations are obtained. **Keywords**: Gamma subordinator; Variance Gamma process; Geometric Stable subordinator; Negative Binomial process; Fractional shift operator. *AMS Mathematical Subject Classification (2010).* 60G52, 34A08, 33E12, 26A33. author: - 'Luisa Beghin[^1]' title: 'Fractional Gamma process and fractional Gamma-subordinated processes' --- Introduction and preliminaries ============================== The Gamma subordinator $\Gamma (t),t>0$ is a very well-known process, applied to many different fields such as, for example, engineering reliability, maintenance theory, risk theory, option pricing and so on. It can be considered as a particular case of the tempered stable subordinators and thus used in financial modelling (see e.g. [@CON]). We define here two fractional versions of $\Gamma $, obtained by a random time-change by means of an independent stable subordinator $\mathcal{A}_{1/\nu }$, or, alternatively, the inverse stable subordinator $\mathcal{L}_{\nu }$ (see section 2.2 for their exact definitions). Thus we define, for any $t\geq 0,$$$\left\{ \begin{array}{l} \Gamma _{\nu }(t):=\Gamma (\mathcal{L}_{\nu }(t)),\quad 0<\nu <1 \\ \overline{\Gamma }_{\nu }(t):=\Gamma (\mathcal{A}_{1/\nu }(t)),\quad \nu >1\end{array}\right. \label{pr6}$$(the case $\nu =1$ corresponds to the standard Gamma process $\Gamma (t)$). Only for $\nu >1$, $\overline{\Gamma }_{\nu }$ represents itself a Lévy process, since it is obtained by subordinating a Lévy process to a stable subordinator. The processes defined in (\[pr6\]) can be considered as fractional versions of $\Gamma $, since we prove that, in both cases, their distributions satisfy differential equations expressed in terms of a new operator, that we call “fractional shift operator”. We recall the definition of the (integer order) shift operator: let $D_{x}^{n}:=d^{n}/dx^{n}$, for any $n\in \mathbb{N}$, then$$e^{cD_{x}}f(x):=\sum_{n=0}^{\infty }\frac{c^{n}D_{x}^{n}}{n!}f(x)=f(x+c), \label{shi}$$for any analytic function $f:\mathbb{R}\rightarrow \mathbb{R}$ and $c\in \mathbb{R}.$ The fractional counterpart of (\[shi\]) is obtained by replacing the $n$-th order derivative by the $n$-fold iterated fractional derivative. We will adopt the Caputo definition of fractional derivative, for $0<\nu \leq 1$, i.e.$$D_{x}^{\nu }u(x):=\frac{1}{\Gamma (1-\nu )}\int_{0}^{x}\frac{1}{(x-s)^{\nu }}\frac{d}{ds}u(s)ds, \label{ca}$$while, for $\nu \geq 1$, we use the right sided fractional Riemann-Liouville derivative on $\mathbb{R}^{+},$ i.e.$$\mathcal{D}_{-,x}^{\nu }u(t):=\frac{1}{\Gamma (m-\nu )}\left( -\frac{d}{dx}\right) ^{m}\int_{x}^{+\infty }\frac{1}{(s-x)^{1+\nu -m}}u(s)ds\text{,\qquad }m-1<\nu <m, \label{rl}$$for $x>0$ (see (2.2.4) of [@KIL], p.80). Correspondingly we define the two operators $\mathcal{O}_{c,x}^{\nu },$ for $\nu \leq 1,$ and $\overline{\mathcal{O}}_{c,x}^{\nu },$ for $\nu \geq 1,$ as fractional variants of (\[shi\]) (see Definitions 1 and 2 below). A generalized exponential operator of fractional order has been presented in [@DATT] (see also [@BABU]): when applied to power functions, i.e. $f(x)=x^{k},$ $k\in \mathbb{N}$, it is proved to produce the so-called Hermite-Kampé de Feriét polynomials. More recently, another fractional version of (\[shi\]) has been proposed in [@MIS] (where the exponential is replaced by the Mittag-Leffler function). By means of the operators $\mathcal{O}_{c,x}^{\nu }$ and $\overline{\mathcal{O}}_{c,x}^{\nu }$, we obtain, in the next section, the fractional differential equations satisfied by the one-dimensional distributions of $\Gamma _{\nu }$ and $\overline{\Gamma }_{\nu }$, in the two ranges of the fractional parameter $\nu $. The expedience of using (\[ca\]) and (\[rl\]) in the definitions of $\mathcal{O}_{c,x}^{\nu }$ and $\overline{\mathcal{O}}_{c,x}^{\nu }$ for $\nu <1$ and $\nu >1$, respectively, has been suggested by the results on stable subordinators and their inverse: indeed it is known that the law of $\mathcal{A}_{1/\nu }$ (that we will denote as $h_{1/\nu }(x,t)$) satisfies the following fractional equation of order $\nu >1$$$\mathcal{D}_{-,\,t}^{\nu }h_{1/\nu }=\frac{\partial }{\partial x}h_{1/\nu },\quad \;x,\text{ }t\geq 0,\;h_{1/\nu }(x,0)=\delta (x)$$(with other appropriate initial conditions, see [@DOV] and [@BEG1] for details). On the other hand the density of the process $\mathcal{L}_{\nu }(t):=\inf \{z:\mathcal{A}_{\nu }(z)>t\}$ (denoted hereafter as $l_{\nu }(x,t)$) satisfies the fractional equation of order $\nu <1$ $$D_{\,t}^{\nu }l_{\nu }=-\frac{\partial }{\partial x}l_{\nu },\quad \;x,\text{ }t\geq 0,\text{ }l_{\nu }(x,0)=\delta (x),$$(see [@HAH]). Moreover, we prove that the one-dimensional distributions of the processes in (\[pr6\]) satisfy, alternatively, a differential equation expressed in terms of the fractional version of the operator $\mathcal{P}_{c,x}$, defined as$$\mathcal{P}_{c,x}f(x):=\sum_{j=1}^{\infty }\frac{(-1)^{j+1}}{jc^{j}}D_{x}^{j}f(x),\text{ \qquad }c,x\in \mathbb{R}\text{,} \label{a}$$for any infinitely differentiable function $f.$ As a consequence of all the previous results, we derive in Section 3 the differential equations satisfied by the fractional versions of the following Gamma subordinated processes: the Variance Gamma process, the Geometric Stable subordinator and the Negative Binomial process. The Variance Gamma (hereafter VG) process (alternatively defined as Laplace motion) is obtained by subordinating a Brownian motion to an independent Gamma subordinator:$$X(t):=B(\Gamma (t)),\qquad t\geq 0,$$where $B$ is a standard Brownian motion. The VG process is a particular case of a symmetric geometric $\nu $-stable process, for $\nu =2,$ and it is widely used in the financial theory, in order to model the logarithm of stock prices (see, e.g. [@MAD], [@KOT]). It has been already proposed a fractional version of the VG process in [@KOZ], defined by subordinating a fractional Brownian motion $B_{H}$ to an independent Gamma subordinator, i.e. as $X_{H}(t):=B_{H}(\Gamma (t)),$ where $H\in (0,1)$ is the Hurst exponent. This process is useful to model hydraulic conductivity fields in geophysics, as well as financial time series. We propose here different fractional versions of $X$, defined as $X_{\nu }(t):=B(\Gamma _{\nu }(t)),$ $t\geq 0,$ for $\nu <1$, and $\overline{X}_{\nu }(t):=B(\overline{\Gamma }_{\nu }(t)),$ $t\geq 0$, for $\nu >1$. Again, in the last case, we get a Lévy process and the corresponding Lévy symbol is obtained. Moreover, by definition, it is clear that the marginal distributions of the fractional VG processes are scales mixtures of normal laws: indeed it is$$X_{\nu }(t)\overset{d}{=}\Gamma _{\nu }(t)Z,\text{ for }\nu <1\text{ and }\overline{X}_{\nu }(t)\overset{d}{=}\overline{\Gamma }_{\nu }(t)Z,\text{ for }\nu >1\text{,} \label{re}$$where $Z$ is a standard Gaussian variable and $\overset{d}{=}$ denotes the equality of one-dimensional distributions. By comparing (\[re\]) with formula (1.5) of [@KOZ], we can note that here $\Gamma _{\nu }$ and $\overline{\Gamma }_{\nu }$ play the same role of the generalized Gamma (or Amoroso) random process $G_{t}^{2H}$ process (whose law is reported there in (2.1)). Thus they can represent the stochastic variance or volatility, in financial terms. Also the Geometric Stable (hereafter GS) subordinator is widely studied and applied, especially in financial contexts (see [@KOZ3]); it is one of the special subordinators for which the potential measure has a decreasing density, thus a wide potential theory has been established for it (see [BOG]{}, [@SIK]). The GS process is defined as a stable subordinator time-changed by means of a Gamma process (see (\[gs\]) below). The differential equation satisfied by its density has been obtained in [BEG]{} and we generalize it to the fractional case. The Negative Binomial (hereafter NB) process is a discrete valued process, which can be defined, alternatively, as a compound Poisson process with logarithmic jumps or as a mixed Poisson process (i.e. a Poisson process subordinated to an independent Gamma subordinator, see, e.g., [@KOZ2]). Through the first definition, a fractional version of the NB process has been introduced in [@BEG2] and the corresponding densities are proved to solve fractional recursive differential equations, which generalize the Kolmogorov ones. By exploiting the mixing representation, we obtain here alternative differential equations, involving the fractional shift operators. Fractional Gamma processes ========================== We first recall the following preliminary result, proved in [@BEG]: the one-dimensional distribution of the Gamma subordinator $\Gamma (t),t\geq 0,$ of parameter $b>0,$ i.e.$$f_{\Gamma }(x,t):=\Pr \left\{ \Gamma (t)\in dx\right\} =\left\{ \begin{array}{l} \frac{b^{t}}{\Gamma (t)}x^{t-1}e^{-bx},\qquad x\geq 0 \\ 0,\qquad x<0\end{array}\right. , \label{gam}$$satisfies the following Cauchy problem, for $x,t\geq 0,$$$\left\{ \begin{array}{l} \frac{\partial }{\partial x}f_{\Gamma }=-b(1-e^{-\partial _{t}})f_{\Gamma } \\ f_{\Gamma }(x,0)=\delta (x) \\ \lim_{|x|\rightarrow +\infty }f_{\Gamma }(x,t)=0\end{array}\right. , \label{res}$$where $e^{-\partial _{t}}$ is the partial derivative version of the shift operator defined in (\[shi\]) and $\delta (x)$ is the Dirac delta function. Then we need to introduce the definition of the *fractional shift operators*, for the two cases $\nu \leq 1$ and $\nu \geq 1.$ Let $f:\mathbb{R}^{+}\rightarrow \mathbb{R}$ be a continuous function with fractional derivative $D_{x}^{\nu }$ defined in (\[ca\]), for $\nu \in (0,1],$ then$$\mathcal{O}_{c,x}^{\nu }f(x):=\sum_{n=0}^{\infty }\frac{c^{n}}{n!}\underbrace{D_{x}^{\nu }...D_{x}^{\nu }}_{n-times}f(x), \label{fr1}$$provided that the series converges. Let $f:\mathbb{R}^{+}\rightarrow \mathbb{R}$ be a continuous function with fractional derivative $\mathcal{D}_{-,x}^{\nu }$ defined in (\[rl\]), for $\nu \geq 1,$ then $$\overline{\mathcal{O}}_{c,x}^{\nu }f(x):=\sum_{n=0}^{\infty }\frac{(-c)^{n}}{n!}\underbrace{\mathcal{D}_{-,x}^{\nu }...\mathcal{D}_{-,x}^{\nu }}_{n-times}f(x), \label{fr2}$$provided that the series converges. The semigroup property does not hold for the fractional derivatives $D_{x}^{\nu }$ and $\mathcal{D}_{-,x}^{\nu }$ and thus for the operators ([fr1]{}) and (\[fr2\]) cannot be used the formalism $e^{cD_{x}^{\nu }}$ adopted in [@DATT]. It is easy to check that, for $\nu =1$, the fractional shift operator defined in (\[fr1\]) coincides with the standard shift operator in ([shi]{}): indeed we get$$\mathcal{O}_{c,x}^{1}f(x)=\sum_{n=0}^{\infty }\frac{c^{n}}{n!}D_{x}^{n}f(x)=e^{cD_{x}}f(x).$$On the other hand, for $\nu =1$, formula (\[fr2\]) reduces to$$\overline{\mathcal{O}}_{c,x}^{1}f(x):=\sum_{n=0}^{\infty }\frac{(-c)^{n}}{n!}(-1)^{n}D_{x}^{n}f(x)=e^{cD_{x}}f(x),$$since $\mathcal{D}_{-,x}^{n}=(-1)^{n}D_{x}^{n}$ (see (2.2.5) of [@KIL]). We note that (\[fr1\]) and (\[fr2\]) do not coincide with the fractional analogue of the Taylor’s series expansion introduced in [@OSL] and its generalizations presented in [@TRUJ], [@JUM]. The case $\protect\nu <1$: the fractional Gamma process ------------------------------------------------------- We consider now the first fractional Gamma process, defined as$$\Gamma _{\nu }(t):=\Gamma (\mathcal{L}_{\nu }(t))\text{,\qquad }\nu \in (0,1],\text{ }t\geq 0, \label{pr}$$where $\Gamma $ is a Gamma process independent of $\mathcal{L}_{\nu }$ and by $\mathcal{L}_{\nu }$ we denote the inverse of a stable subordinator $\mathcal{A}_{\nu }$ of index $\nu $ (with parameters $\mu =0,$ $\beta =1,$ $\sigma =(t\cos \pi \nu /2)^{1/\nu }$, in the notation of [@SAMO]). Thus, by definition, $\mathcal{L}_{\nu }(t):=\inf \left\{ z\geq 0:\mathcal{A}_{\nu }(z)>t\right\} \ $and we recall that $$\mathbb{E}e^{-k\mathcal{L}_{\nu }(t)}=E_{\nu ,1}(-kt^{\nu }),\qquad k>0, \label{elle2}$$where $$E_{\nu ,\beta }(x)=\sum_{j=0}^{\infty }\frac{x^{j}}{\Gamma (\nu j+\beta )},\quad \mathcal{R}(\nu )>0,\text{ }\beta ,x\in \mathbb{C}$$is the Mittag-Leffler function. We start by deriving the fractional equation satisfied by the one-dimensional distribution of the process $\Gamma _{\nu }$ defined in (\[pr\]). Let $f_{\Gamma _{\nu }}(x,t):=\Pr \{\Gamma _{\nu }(t)\in dx\}$, for $x,t\geq 0$, and $l_{\nu }(x,t):=\Pr \{\mathcal{L}_{\nu }(t)\in dx\},$ then the density$$f_{\Gamma _{\nu }}(x,t)=\int_{0}^{\infty }f_{\Gamma }(x,z)l_{\nu }(z,t)dz \label{pr3}$$satisfies, for $\nu \in (0,1)$ and $t>1$, the following equation$$\frac{\partial }{\partial x}f_{\Gamma _{\nu }}=-b(1-\mathcal{O}_{-1,t}^{\nu })f_{\Gamma _{\nu }},\qquad x\geq 0, \label{pr4}$$with initial condition$$f_{\Gamma _{\nu }}(0,t)=0. \label{pr5}$$ The initial condition is immediately satisfied by (\[pr3\]), since $f_{\Gamma }(0,t)=0,$ for $t>1.$ In order to verify equation (\[pr4\]) we evaluate the Laplace transform of (\[pr3\]), with respect to $x$, by denoting $\widetilde{f}_{\Gamma _{\nu }}(\theta ,t):=\int_{0}^{\infty }e^{-\theta x}f_{\Gamma _{\nu }}(x,t)dx,$$$\begin{aligned} \widetilde{f}_{\Gamma _{\nu }}(\theta ,t) &=&\int_{0}^{\infty }\widetilde{f}_{\Gamma }(\theta ,z)l_{\nu }(z,t)dz \label{sub} \\ &=&\int_{0}^{\infty }\exp \left\{ -z\log \left( 1+\frac{\theta }{b}\right) \right\} l_{\nu }(z,t)dz \notag \\ &=&[\text{by (\ref{elle2}) with }k=\log [1+\theta /b)]>0] \notag \\ &=&E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) . \notag\end{aligned}$$By taking the Laplace transform of the r.h.s. of (\[pr4\]) we get$$\begin{aligned} &&-b(1-\mathcal{O}_{-1,t}^{\nu })\widetilde{f}_{\Gamma _{\nu }}(\theta ,t) \label{rs} \\ &=&-bE_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) +b\sum_{n=0}^{\infty }\frac{(-1)^{n}}{n!}\underbrace{D_{t}^{\nu }...D_{t}^{\nu }}_{n-times}E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) \notag \\ &=&[\text{by n applications of formula (2.4.58) of \cite{KIL}]} \notag \\ &=&-bE_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) +bE_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) \sum_{n=0}^{\infty }\frac{\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{n}}{n!} \notag \\ &=&-b\left[ 1-1-\frac{\theta }{b}\right] E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) =\theta E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) , \notag\end{aligned}$$which coincides with the Laplace transform of the l.h.s. of (\[pr4\]), i.e. $$\begin{aligned} \theta \widetilde{f}_{\Gamma _{\nu }}(\theta ,t) &=&\theta E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) -f_{\Gamma _{\nu }}(0,t) \label{ls} \\ &=&[\text{by (\ref{pr5}) }] \notag \\ &=&\theta E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) . \notag\end{aligned}$$ We can alternatively prove that the density (\[pr3\]) satisfies the following time-fractional equation, for any $\nu \in (0,1)$ and $t>1$:$$D_{t}^{\nu }f_{\Gamma _{\nu }}=-\mathcal{P}_{b,x}f_{\Gamma _{\nu }},\qquad x\geq 0, \label{re5}$$with initial conditions$$\left. \frac{\partial ^{j}}{\partial x^{j}}f_{\Gamma _{\nu }}(x,t)\right\vert _{x=0}=0,\qquad j=0,1,... \label{re6}$$where $D_{t}^{\nu }$ denotes, as usual, the Caputo fractional derivative. Indeed by taking the Laplace transform of (\[re5\]) we get $$\begin{aligned} \mathcal{L}\left\{ D_{t}^{\nu }f_{\Gamma _{\nu }}(\cdot ,t);\theta \right\} &=&D_{t}^{\nu }E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) \label{eq4} \\ &=&-\log \left( 1+\frac{\theta }{b}\right) E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) \notag \\ &=&-\mathcal{L}\left\{ \mathcal{P}_{b,x}f_{\Gamma _{\nu }}(\cdot ,t);\theta \right\} , \notag\end{aligned}$$where $\mathcal{L}\left\{ f(\cdot );\theta \right\} :=\int_{0}^{\infty }e^{-\theta x}f_{\Gamma }(x,t)dx.$ The last equality in (\[eq4\]) is obtained by considering the well-known formula$$\mathcal{L}\left\{ D_{x}^{l}f(\cdot );\theta \right\} =\theta ^{l}\widetilde{f}(\theta )-\left. \sum_{j=0}^{l-1}\theta ^{j}D_{x}^{l-1-j}f(x)\right\vert _{x=0}. \label{aa}$$Indeed, by the definition of $\mathcal{P}_{c,x}$ given in (\[a\]), for $c=b $, we get$$\begin{aligned} \mathcal{L}\left\{ \mathcal{P}_{b,x}f_{\Gamma _{\nu }}(\cdot ,t);\theta \right\} &=&\sum_{l=1}^{\infty }\frac{(-1)^{l+1}}{lb^{l}}\int_{0}^{\infty }e^{-\theta x}D_{x}^{l}f_{\Gamma _{\nu }}(x,t)dx \\ &=&[\text{by (\ref{re6})}] \\ &=&\sum_{l=1}^{\infty }\frac{(-1)^{l+1}\theta ^{l}}{lb^{l}}E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) \\ &=&\log \left( 1+\frac{\theta }{b}\right) E_{\nu ,1}\left( -\log \left( 1+\frac{\theta }{b}\right) t^{\nu }\right) .\end{aligned}$$ We note that the process $\Gamma _{\nu }$ is no longer a subordinator, since it is clear from (\[sub\]) that its density is not infinitely divisible. To avoid this problem we present in the next section a fractional version of the Gamma process which is still infinitely divisible, and, being increasing, is also a subordinator. We analyze here some properties of the process $\Gamma _{\nu }$, such as its moments. The expected value is finite and can be obtained by taking its Laplace transform and considering the well-known result$$\int_{0}^{\infty }e^{-st}l_{\nu }(x,t)dt=s^{\nu -1}e^{-s^{\nu }x}.$$Indeed we get$$\begin{aligned} \mathcal{L}\left\{ \mathbb{E}\Gamma _{\nu }(t);s\right\} &=&\int_{0}^{\infty }e^{-st}\int_{0}^{\infty }\mathbb{E}\Gamma (z)l_{\nu }(z,t)dzdt \\ &=&\frac{1}{b}\int_{0}^{\infty }z\int_{0}^{\infty }e^{-st}l_{\nu }(z,t)dtdz \\ &=&\frac{s^{\nu -1}}{b}\int_{0}^{\infty }ze^{-xs^{\nu }}dz=\frac{1}{bs^{\nu +1}}\end{aligned}$$and thus $$\mathbb{E}\Gamma _{\nu }(t)=\frac{t^{\nu }}{b\Gamma (\nu +1)},$$which, for $\nu =1$, reduces to the well-known expected value of the Gamma process. The $r$-th absolute moments of $\Gamma _{\nu }$ are given by$$\mathbb{E}\Gamma _{\nu }(t)^{r}=\frac{1}{b^{r}}\sum_{k=0}^{r}\QATOPD[ ] {r}{k}\frac{k!t^{\nu k}}{\Gamma (\nu k+1)},\qquad r\in \mathbb{Z}, \label{mom}$$where $\QATOPD[ ] {r}{k}$ denotes the (unsigned) Stirling numbers of the first kind. Recall that, for the Gamma subordinator,$$\mathbb{E}\Gamma (t)^{r}=\frac{1}{b^{r}}\frac{\Gamma (r+t)}{\Gamma (t)}, \label{lem}$$so that we get$$\begin{aligned} \mathcal{L}\left\{ \mathbb{E}\Gamma _{\nu }(\cdot )^{r};s\right\} &=&\int_{0}^{\infty }e^{-st}\int_{0}^{\infty }\mathbb{E}\Gamma (z)^{r}l_{\nu }(z,t)dzdt \label{mom2} \\ &=&\frac{s^{\nu -1}}{b^{r}}\int_{0}^{\infty }\frac{\Gamma (r+z)}{\Gamma (z)}e^{-s^{\nu }z}dt \notag \\ &=&\frac{s^{\nu -1}}{b^{r}}\int_{0}^{\infty }z^{(r)}e^{-s^{\nu }z}dt, \notag\end{aligned}$$where $z^{(r)}$ denotes the rising factorial defined as $z^{(r)}:=\Gamma (r+z)/\Gamma (r).$ We recall the following expansion for the rising factorials:$$z^{(r)}=\sum_{k=0}^{r}\QATOPD[ ] {r}{k}z^{k}, \label{ris}$$therefore (\[mom2\]) can be rewritten as$$\begin{aligned} \mathcal{L}\left\{ \mathbb{E}\Gamma _{\nu }(\cdot )^{r};s\right\} &=&\frac{s^{\nu -1}}{b^{r}}\sum_{k=0}^{r}\QATOPD[ ] {r}{k}\int_{0}^{\infty }z^{k}e^{-s^{\nu }z}dt \\ &=&\frac{s^{\nu -1}}{b^{r}}\sum_{k=0}^{r}\QATOPD[ ] {r}{k}\frac{k!}{s^{\nu k+\nu }}=\frac{1}{b^{r}}\sum_{k=0}^{r}\QATOPD[ ] {r}{k}\frac{k!}{s^{\nu k+1}}.\end{aligned}$$By inverting the Laplace transform we get (\[mom\]). In order to obtain the variance of $\Gamma _{\nu }$ we choose $r=2$ in ([mom]{}), so that we get $$var\left( \Gamma _{\nu }(t)\right) =\frac{2t^{2\nu }}{b^{2}\Gamma (2\nu +1)}+\frac{t^{\nu }}{b^{2}\Gamma (\nu +1)}-\frac{t^{2\nu }}{b^{2}\Gamma ^{2}(\nu +1)}.$$Again, for $\nu =1$, we obtain the variance of the Gamma process. The case $\protect\nu >1$: the fractional Gamma subordinator ------------------------------------------------------------ The second fractional Gamma process we present here is defined as$$\overline{\Gamma }_{\nu }(t):=\Gamma (\mathcal{A}_{1/\nu }(t))\text{,\qquad }\nu >1,\text{ }t\geq 0, \label{sec}$$where $\Gamma $ is a Gamma process and $\mathcal{A}_{1/\nu }$ is the (independent) stable subordinator of index $1/\nu $ (with parameters $\mu =0, $ $\theta =1,$ $\sigma =(t\cos \pi /2\nu )^{\nu })$. It is well-known that $$\mathbb{E}e^{-k\mathcal{A}_{1/\nu }(t)}=e^{-k^{1/\nu }t},\qquad k>0. \label{sec2}$$ Let $f_{\overline{\Gamma }_{\nu }}(x,t):=\Pr \{\overline{\Gamma }_{\nu }(t)\in dx\}$, $x,t\geq 0$, and $h_{1/\nu }(x,t):=\Pr \{\mathcal{A}_{1/\nu }(t)\in dx\},$ then the density$$f_{\overline{\Gamma }_{\nu }}(x,t)=\int_{0}^{\infty }f_{\Gamma }(x,z)h_{1/\nu }(z,t)dz \label{sec3}$$satisfies, for $\nu >1$ and $t>1$, the following equation$$\frac{\partial }{\partial x}f_{\overline{\Gamma }_{\nu }}=-b(1-\overline{\mathcal{O}}_{-1,t}^{\nu })f_{\overline{\Gamma }_{\nu }},\qquad x\geq 0, \label{sec4}$$with initial condition$$f_{\overline{\Gamma }_{\nu }}(0,t)=0. \label{sec5}$$ The condition (\[sec5\]) is immediately satisfied by (\[sec3\]). As far as equation (\[sec4\]) is concerned, as in the proofs of Theorem 3, we take the Laplace transform of the r.h.s. of (\[sec4\]) which reads$$\begin{aligned} &&-b(1-\overline{\mathcal{O}}_{-1,t}^{\nu })\int_{0}^{\infty }\widetilde{f}_{\Gamma }(\theta ,z)h_{1/\nu }(z,t)dz \\ &=&-b(1-\overline{\mathcal{O}}_{-1,t}^{\nu })\int_{0}^{\infty }\exp \left\{ -z\log \left( 1+\frac{\theta }{b}\right) \right\} h_{1/\nu }(z,t)dz \\ &=&\text{[by (\ref{sec2}) with }k=\log [1+\theta /b)]\text{]} \\ &=&-b(1-\overline{\mathcal{O}}_{-1,t}^{\nu })\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \\ &=&-b\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} +b\sum_{n=0}^{\infty }\frac{1}{n!}\underbrace{\mathcal{D}_{-,t}^{\nu }...\mathcal{D}_{-,t}^{\nu }}_{n-times}\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \\ &=&[\text{by n applications of (2.2.15) in \cite{KIL}}] \\ &=&-b\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} +b\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \sum_{n=0}^{\infty }\frac{1}{n!}\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{n} \\ &=&-b\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} +b\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \frac{b+\theta }{b},\end{aligned}$$which coincides with the Laplace transform of the l.h.s. of (\[sec4\]), i.e.$$\begin{aligned} \mathcal{L}\{\frac{\partial }{\partial x}f_{\overline{\Gamma }_{\nu }}(\cdot ,t);\theta \} &=&\theta \widetilde{f}_{\overline{\Gamma }_{\nu }}(\theta ,t)-f_{\overline{\Gamma }_{\nu }}(0,t) \\ &=&[\text{by (\ref{sec5})]} \\ &=&\theta \exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} .\end{aligned}$$ We can alternatively prove that the density (\[sec3\]) satisfies the following time-fractional equation, for any $\nu \in (1,+\infty )$ and $t>1$:$$\mathcal{D}_{-,t}^{\nu }f_{\overline{\Gamma }_{\nu }}=-\mathcal{P}_{b,x}f_{\overline{\Gamma }_{\nu }},\qquad x\geq 0, \label{re8}$$with initial conditions$$\left. \frac{\partial ^{j}}{\partial x^{j}}f_{\overline{\Gamma }_{\nu }}(x,t)\right\vert _{x=0}=0,\qquad j=0,1,... \label{re9}$$where $\mathcal{D}_{-,t}^{\nu }$ denotes the Riemann-Liouville fractional derivative defined in (\[rl\]). Indeed by taking the Laplace transform of (\[re8\]) we get $$\begin{aligned} \mathcal{L}\left\{ \mathcal{D}_{-,t}^{\nu }f_{\overline{\Gamma }_{\nu }}(\cdot ,t);\theta \right\} &=&\mathcal{D}_{-,t}^{\nu }\exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \\ &=&-\log \left( 1+\frac{\theta }{b}\right) \exp \left\{ -t\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }\right\} \\ &=&[\text{by (\ref{aa}) and (\ref{re9})}] \\ &=&-\mathcal{L}\left\{ \mathcal{P}_{b,x}f_{\overline{\Gamma }_{\nu }}(\cdot ,t);\theta \right\} .\end{aligned}$$ We remark that the process $\overline{\Gamma }_{\nu }$ defined in (\[sec\]) is obtained by subordinating a Lévy process to the independent stable subordinator $\mathcal{A}_{1/\nu }$. Thus it is itself a Lévy process (see, e.g. [@APPL], Theorem 1.3.25) and, being also real valued and increasing, it is a subordinator. Its Laplace exponent can be evaluated directly: $$\psi _{\overline{\Gamma }_{\nu }}(\theta ):=-\frac{1}{t}\log \mathbb{E}e^{-\theta \overline{\Gamma }_{\nu }(t)}=-\frac{1}{t}\log \widetilde{f}_{\overline{\Gamma }_{\nu }}(\theta ,t)=\left( \log \left( 1+\frac{\theta }{b}\right) \right) ^{1/\nu }, \label{lp}$$which reduce, for $\nu =1,$ to the Laplace exponent of $\Gamma .$ It can be checked that (\[lp\]) is a Bernstein function, by verifying that $(-1)^{n}d^{(n)}\psi _{\overline{\Gamma }_{\nu }}(\theta )/d\theta ^{(n)}\geq 0.$ Moreover we have $lim_{\theta \rightarrow 0}\psi _{\overline{\Gamma }_{\nu }}(\theta )=0$. Thus, by Theorem 1.3.4, p.45 in [@APPL], for $\psi _{\overline{\Gamma }_{\nu }}(\theta )$ there exists the following representation$$\psi _{\overline{\Gamma }_{\nu }}(\theta )=a+b\theta +\int_{0}^{\infty }(1-e^{-y\theta })\lambda (dy),$$for a measure $\lambda $ s.t. $\int_{0}^{\infty }(y\wedge 1)\lambda (dy)<\infty $, for all $\theta >0,$ $a,b\geq 0.$ Loosely speaking formula (\[lp\]) implies that $\overline{\Gamma }_{\nu }$ grows more slowly than the standard Gamma subordinator, as $x\rightarrow +\infty .$ Indeed $\Gamma $ increases at a logarithmic rate (see, for example, [@LIN]). The Lévy symbol $\eta _{\overline{\Gamma }_{\nu }}(u):=\log \mathbb{E}e^{iu\overline{\Gamma }_{\nu }(t)}/t$ of $\overline{\Gamma }_{\nu }$ can be obtained by applying Proposition 1.3.27 in [@APPL]: for any $u\in \mathbb{R},$$$\eta _{\overline{\Gamma }_{\nu }}(u)=-\psi _{\mathcal{A}_{1/\nu }}(-\eta _{\Gamma }(u)),$$where $\psi _{\mathcal{A}_{1/\nu }}(\theta )=\theta ^{1/\nu }$ is the Laplace exponent of the stable subordinator and $\eta _{\Gamma }(u)=-\log (1-iu/b)$ the Lévy symbol of $\Gamma $. Thus we get$$\eta _{\overline{\Gamma }_{\nu }}(u)=-\left[ \log (1-iu/b)\right] ^{1/\nu },$$which, for $\nu =1$ reduces to $\eta _{\Gamma }.$ Finally we note that the process $\overline{\Gamma }_{\nu }$ does not possess any finite moment, since the same is true for the subordinator $\mathcal{A}_{1/\nu }.$ Applications to Gamma-subordinated processes ============================================ We will define and analyze the fractional versions of well-known processes, such as the VG, the GS and the NB processes, which are all expressed through a random time change by the Gamma subordinator. By applying the previous results, we will be able to derive the fractional equations satisfied by their distributions, expressed by means of the fractional shift operators $\mathcal{O}_{-1,t}^{\nu }$ and $\overline{\mathcal{O}}_{-1,t}^{\nu }$. Fractional Variance Gamma processes ----------------------------------- We consider now the process obtained by the composition of a Brownian motion with one of the two fractional versions of the Gamma process studied so far. Thus we define$$\left\{ \begin{array}{c} X_{\nu }(t):=B(\Gamma _{\nu }(t)),\qquad t\geq 0,\text{ }\nu \in (0,1) \\ \overline{X}_{\nu }(t):=B(\overline{\Gamma }_{\nu }(t)),\qquad t\geq 0,\text{ }\nu \in (1,\infty )\end{array}\right. \label{mee}$$where $B$ is a standard Brownian motion and $\Gamma _{\nu }$ and $\overline{\Gamma }_{\nu }$ are independent of $B$. For $\nu =1$, the processes in ([mee]{}) coincides with the Variance Gamma process. Let, for simplicity, $b=1$ from now onwards. Let us denote the one-dimensional distributions of $X_{\nu }$ and $\overline{X}_{\nu }$ as $$f_{X_{\nu }}(x,t)=\int_{0}^{\infty }f_{B}(x,z)f_{\Gamma _{\nu }}(z,t)dz,\qquad x,t\geq 0,\text{ }\nu \in (0,1), \label{vg}$$and$$f_{\overline{X}_{\nu }}(x,t)=\int_{0}^{\infty }f_{B}(x,z)f_{\overline{\Gamma }_{\nu }}(z,t)dz,\qquad x,t\geq 0,\text{ }\nu >1, \label{vg2}$$where $f_{B}$ is the transition density of the standard Brownian motion $B.$ The density (\[vg\]) of $X_{\nu }$ satisfies the following fractional differential equation$$\frac{1}{2}\frac{\partial ^{2}}{\partial x^{2}}f_{X_{\nu }}=(1-\mathcal{O}_{-1,t}^{\nu })f_{X_{\nu }},\qquad x\geq 0,\text{ }t>1,\qquad \nu \in (0,1), \label{vg1}$$while the density (\[vg2\]) of $\overline{X}_{\nu }$ satisfies the equation$$\frac{1}{2}\frac{\partial ^{2}}{\partial x^{2}}f_{\overline{X}_{\nu }}=(1-\overline{\mathcal{O}}_{-1,t}^{\nu })f_{\overline{X}_{\nu }},\qquad x\geq 0,\text{ }t>1,\qquad \nu >1. \label{vg3}$$The boundary conditions for both equations are$$\left\{ \begin{array}{l} \lim_{|x|\rightarrow \infty }f(x,t)=0 \\ \lim_{|x|\rightarrow \infty }\frac{\partial }{\partial x}f(x,t)=0\end{array}\right. , \label{con}$$where $f:=f_{X_{\nu }},$ for $\nu \in (0,1)$ and $f:=f_{\overline{X}_{\nu }},$ for $\nu \in (1,\infty ).$ The conditions (\[con\]) can be easily checked by considering that their analogues are satisfied by $f_{B}$. For $\nu \in (0,1),$ in order to prove (\[vg1\]), we take the Fourier transform with respect to $x$ of its r.h.s.: let$$\widehat{f}(u):=\mathcal{F}\left\{ f(\cdot );u\right\} =\int_{-\infty }^{+\infty }e^{iux}f(x)dx,$$then$$\begin{aligned} \mathcal{F}\left\{ (1-\mathcal{O}_{-1,t}^{\nu })f_{X_{\nu }}(\cdot ,t);u\right\} &=&(1-\mathcal{O}_{-1,t}^{\nu })\widehat{f}_{X_{\nu }}(u,t) \label{cl1} \\ &=&(1-\mathcal{O}_{-1,t}^{\nu })E_{\nu ,1}\left( -t^{\nu }\log \left( 1+\frac{u^{2}}{2}\right) \right) \notag \\ &=&-\frac{u^{2}}{2}E_{\nu ,1}\left( -t^{\nu }\log \left( 1+\frac{u^{2}}{2}\right) \right) , \notag\end{aligned}$$where for the last step we have performed some calculations similar to ([rs]{}). For the l.h.s. of (\[vg1\]), by considering the conditions ([con]{}), we get instead$$\mathcal{F}\left\{ \frac{1}{2}\frac{\partial ^{2}}{\partial x^{2}}f_{X_{\nu }}(\cdot ,t);u\right\} =-\frac{u^{2}}{2}\widehat{f}_{X_{\nu }}(u,t), \label{cl2}$$which coincides with (\[cl1\]). Analogously, for $\nu >1,$ we get$$\begin{aligned} \mathcal{F}\left\{ (1-\overline{\mathcal{O}}_{-1,t}^{\nu })f_{\overline{X}_{\nu }}(\cdot ,t);u\right\} &=&(1-\overline{\mathcal{O}}_{-1,t}^{\nu })\widehat{f}_{\overline{X}_{\nu }}(u,t) \\ &=&(1-\overline{\mathcal{O}}_{-1,t}^{\nu })\exp \left\{ -t\left( \log \left( 1+\frac{u^{2}}{2}\right) \right) ^{1/\nu }\right\} \notag \\ &=&-\frac{u^{2}}{2}\exp \left\{ -t\left( \log \left( 1+\frac{u^{2}}{2}\right) \right) ^{1/\nu }\right\} , \notag\end{aligned}$$which is equal to the analogue of (\[cl2\]), with $f_{X_{\nu }}$ replaced by $f_{\overline{X}_{\nu }}.$ Alternatively we can prove the previous result directly, without resorting to the Fourier transform, by considering the heat equation together with (\[pr4\]), for $\nu \in (0,1)$, and (\[sec4\]), for $\nu >1$: indeed, in the first case, we get$$\begin{aligned} \frac{1}{2}\frac{\partial ^{2}}{\partial x^{2}}f_{X_{\nu }}(x,t) &=&\frac{1}{2}\frac{\partial ^{2}}{\partial x^{2}}\int_{0}^{\infty }f_{B}(x,z)f_{\Gamma _{\nu }}(z,t)dz \\ &=&\int_{0}^{\infty }\frac{\partial }{\partial z}f_{B}(x,z)f_{\Gamma _{\nu }}(z,t)dz \\ &=&\left[ f_{B}(x,z)f_{\Gamma _{\nu }}(z,t)\right] _{z=0}^{z=\infty }-\int_{0}^{\infty }f_{B}(x,z)\frac{\partial }{\partial z}f_{\Gamma _{\nu }}(z,t)dz \\ &=&[\text{by (\ref{pr5})}] \\ &=&-\int_{0}^{\infty }f_{B}(x,z)\frac{\partial }{\partial z}f_{\Gamma _{\nu }}(z,t)dz,\end{aligned}$$which coincides with the right-hand side of (\[vg1\]). We can obtain ([vg3]{}) analogously, for $\nu >1.$ We analyze now the properties of the two versions of fractional VG process, starting with their absolute moments. The absolute $q$-moments of $X_{\nu }$ are given by$$\mathbb{E}|X_{\nu }(t)|^{q}=\left\{ \begin{array}{l} \frac{2^{r}}{\sqrt{\pi }}\Gamma \left( r+\frac{1}{2}\right) \sum_{k=0}^{r}\QATOPD[ ] {r}{k}\frac{k!t^{\nu k}}{\Gamma (\nu k+1)},\qquad \text{for }q=2r \\ \frac{2^{r+\frac{1}{2}}}{\sqrt{\pi }}\Gamma \left( r+1\right) \int_{0}^{\infty }\frac{\Gamma \left( r+z+\frac{1}{2}\right) }{\Gamma (z)}l_{\nu }(z,t)dz\qquad \text{for }q=2r+1\end{array}\right. , \label{mom3}$$for $\nu <1$, while $\mathbb{E}|\overline{X}_{\nu }(t)|^{q}=\infty $, for any $q\geq 1$, for $\nu >1.$ For $\nu <1,$ by considering (\[vg\]) and taking into account the well-known form of the absolute moments of the Brownian motion $B$, we can write$$\begin{aligned} \mathbb{E}|X_{\nu }(t)|^{q} &=&\int_{0}^{+\infty }|x|^{q}\int_{0}^{\infty }f_{B}(x,z)f_{\Gamma _{\nu }}(z,t)dzdx \label{las} \\ &=&\int_{0}^{\infty }\mathbb{E}|B(z)|^{q}f_{\Gamma _{\nu }}(z,t)dz \notag \\ &=&\frac{\sqrt{2}^{q}}{\sqrt{\pi }}\Gamma \left( \frac{q+1}{2}\right) \int_{0}^{\infty }z^{q/2}f_{\Gamma _{\nu }}(z,t)dz. \notag\end{aligned}$$For $q=2r$, by considering (\[mom\]), with $b=1$, we immediately get the first line in (\[mom3\]). For $q=2r+1$ we rewrite (\[las\]) as$$\begin{aligned} &&\mathbb{E}|X_{\nu }(t)|^{q} \notag \\ &=&\frac{\sqrt{2}^{q}}{\sqrt{\pi }}\Gamma \left( \frac{q+1}{2}\right) \int_{0}^{\infty }\mathbb{E}|\Gamma (z)|^{q/2}l_{\nu }(z,t)dz, \notag\end{aligned}$$which, by (\[lem\]), coincides with the second line in (\[mom3\]). It is easy to check that, for $\nu =1$, formula (\[mom3\]) reduces to $$\mathbb{E}|X(t)|^{q}=\frac{2^{q/2}}{\sqrt{\pi }}\Gamma \left( \frac{q}{2}+\frac{1}{2}\right) \frac{\Gamma \left( \frac{q}{2}+t\right) }{\Gamma (t)}\text{ \qquad }q=1,2....,$$which is the well-known formula for the absolute moments of the VG process (see [@KOZ], formula (2.2) for $H=1/2$). Indeed, for $q=2r,$ we get from (\[mom3\])$$\begin{aligned} \mathbb{E}|X_{\nu }(t)|^{2r} &=&\frac{2^{r}}{\sqrt{\pi }}\Gamma \left( r+\frac{1}{2}\right) \sum_{k=0}^{r}\QATOPD[ ] {r}{k}t^{k} \\ &=&[\text{by (\ref{ris})}] \\ &=&\frac{2^{r}}{\sqrt{\pi }}\Gamma \left( r+\frac{1}{2}\right) \frac{\Gamma \left( r+t\right) }{\Gamma (t)},\end{aligned}$$while, for $q=2r+1$, it is$$\mathbb{E}|X_{\nu }(t)|^{2r+1}=\frac{2^{r+\frac{1}{2}}}{\sqrt{\pi }}\Gamma \left( r+1\right) \frac{\Gamma \left( r+t+\frac{1}{2}\right) }{\Gamma (t)},$$since $l_{\nu }(z,t)=\delta (z-t)$, for $\nu =1.$ It is evident from ([mom3]{}) that the even-order absolute moments are never linear in $t$: in particular the variance is given by$$var(X_{\nu }(t))=\frac{t^{\nu }}{\Gamma (\nu +1)}.$$ The most important feature of the fractional VG process is that, for $\nu >1, $ it is a Lévy process and thus infinitely divisible for any $t$, since it is obtained by subordinating the Brownian motion (which is a Lévy process) to the subordinator $\overline{\Gamma }_{\nu }$. The Lévy symbol $\eta _{\overline{X}_{\nu }}(u)$ of $\overline{X}_{\nu }$ can be obtained by applying again Proposition 1.3.27 in [@APPL]: for any $u\in \mathbb{R},$$$\eta _{\overline{X}_{\nu }}(u)=-\psi _{\overline{\Gamma }_{\nu }}(-\eta _{B}(u)),$$where $\psi _{\overline{\Gamma }_{\nu }}(\theta )=\left( \log \left( 1+\theta \right) \right) ^{1/\nu }$ is the Laplace exponent of $\overline{\Gamma }_{\nu }$ (see (\[lp\]), for $b=1$) and $\eta _{B}(u)=-u^{2}/2$ the Lévy symbol of $B.$ Thus we get$$\eta _{\overline{X}_{\nu }}(u)=-\left( \log \left( 1+\frac{u^{2}}{2}\right) \right) ^{1/\nu } \label{dic}$$which, for $\nu =1,$ reduces to $\eta _{X}$ (see [@APPL], p.57)$.$ From (\[dic\]) it is evident that $\overline{X}_{\nu }$ cannot be represented as difference of two independent Gamma suborinators, as happens for $X.$ Let $\Gamma ^{\prime }$ and $\Gamma ^{\prime \prime }$ be two independent Gamma random processes with parameters $a=1$, $b=\sqrt{2}$. Then it is well-known that $X(t)\overset{d}{=}\Gamma ^{\prime }(t)-\Gamma ^{\prime \prime }(t)$, $t\geq 0.$ But, in the fractional case, for $\nu >1$, the characteristic functions are respectively$$\mathbb{E}e^{iu\overline{X}_{\nu }(t)}=\exp \left\{ -t\left( \log \left( 1+\frac{u^{2}}{2}\right) \right) ^{1/\nu }\right\} ,$$and$$\mathbb{E}e^{iu\overline{\Gamma }_{\nu }^{\prime }(t)-iu\overline{\Gamma }_{\nu }^{\prime \prime }(t)}=\exp \left\{ -t\left[ \log (1-iu/\sqrt{2})\right] ^{1/\nu }-t\left[ \log (1+iu/\sqrt{2})\right] ^{1/\nu }\right\} ,$$for $\overline{\Gamma }_{\nu }^{\prime }$ and $\overline{\Gamma }_{\nu }^{^{\prime \prime }}$ independent and defined in (\[sec\]) with $b=\sqrt{2}.$ The previous expressions coincide only in the special case $\nu =1.$ Fractional Geometric Stable subordinator ---------------------------------------- The Geometric Stable (GS) subordinator of index $\alpha $ is defined by the following subordinating relationship$$G_{\alpha }(t):=\mathcal{A}_{\alpha }(\Gamma (t)),\qquad t\geq 0,\text{ }0<\alpha \leq 1, \label{gs}$$where $\mathcal{A}_{\alpha }$ denotes a stable subordinator of index $\alpha $ and $\Gamma $ an independent Gamma subordinator with $b=1$, for simplicity (see [@SIK] and [@KOZ3]). Its Laplace exponent is $$\psi _{G_{\alpha }}(\theta )=\log \left( 1+\theta ^{\alpha }\right) \notag$$so that its Lévy measure is equal to$$\lambda (dx)=\alpha x^{-1}E_{\alpha ,1}(x)dx.$$For $\alpha =1,$ the process $G_{\alpha }$ reduces to the Gamma subordinator. The fractional equation satisfied by the density of $G_{\alpha }$ has been already obtained in [@BEG], in the general case of the GS process (i.e. not necessarily totally skewed to the right). It is expressed in terms of the fractional Riesz-Feller derivative. We derive here an analogous equation, in terms of the Caputo fractional derivative $D_{x}^{\alpha }$ of order $\alpha \in (0,1]$ and then we extend it to the fractional version of the GS subordinator. We define the latter as $$\left\{ \begin{array}{l} G_{\alpha }^{\nu }(t):=\mathcal{A}_{\alpha }(\Gamma _{\nu }(t)),\qquad t\geq 0,\text{ }0<\nu <1, \\ \overline{G}_{\alpha }^{\nu }(t):=\mathcal{A}_{\alpha }(\overline{\Gamma }_{\nu }(t)),\qquad t\geq 0,\text{ }\nu >1,\end{array}\right. \label{ai4}$$ where $\alpha \in (0,1]$, $\Gamma _{\nu }$ and $\overline{\Gamma }_{\nu }$ are independent of $B$. For $\nu =1$, the processes in (\[ai4\]) reduce to the GS subordinator. Let us denote the one-dimensional distributions of $G_{\alpha }^{\nu }$ and $\overline{G}_{\alpha }^{\nu }$ as $$f_{G_{\alpha }^{\nu }}(x,t)=\int_{0}^{\infty }h_{\alpha }(x,z)f_{\Gamma _{\nu }}(z,t)dz,\qquad x,t\geq 0,\text{ }\nu \in (0,1), \label{mee2}$$and$$f_{\overline{G}_{\alpha }^{\nu }}(x,t)=\int_{0}^{\infty }h_{\alpha }(x,z)f_{\overline{\Gamma }_{\nu }}(z,t)dz,\qquad x,t\geq 0,\text{ }\nu >1. \label{mee3}$$ The density (\[mee2\]) of $G_{\alpha }^{\nu }$ satisfies the following doubly fractional differential equation$$D_{x}^{\alpha }f_{G_{\alpha }^{\nu }}=(1-\mathcal{O}_{-1,t}^{\nu })f_{G_{\alpha }^{\nu }},\qquad x\geq 0,\text{ }t>1,\qquad \alpha \in (0,1],\text{ }\nu \in (0,1) \label{mee4}$$while the density (\[mee3\]) of $\overline{G}_{\alpha }^{\nu }$ satisfies the equation$$D_{x}^{\alpha }f_{\overline{G}_{\alpha }^{\nu }}=(1-\overline{\mathcal{O}}_{-1,t}^{\nu })f_{\overline{G}_{\alpha }^{\nu }},\qquad x\geq 0,\text{ }t>1,\qquad \alpha \in (0,1],\text{ }\nu >1, \label{mee5}$$The initial condition for both equations is $f(0,t)=0$ (where $f:=f_{G_{\alpha }^{\nu }},$ for $\nu \in (0,1)$ and $f:=f_{\overline{G}_{\alpha }^{\nu }},$ for $\nu >1$). We can apply the result given in (\[res\]) to obtain the differential equation satisfied by the density of $G_{\alpha }$. By (\[gs\]), the density of the GS subordinator can be written as$$f_{G_{\alpha }}(x,t)=\int_{0}^{\infty }h_{\alpha }(x,z)f_{\Gamma }(z,t)dz, \label{ai}$$where $h_{\alpha }(x,t)$ is the density of $\mathcal{A}_{\alpha }(t),$ $t\geq 0.$ Then $f_{G_{\alpha }}$ satisfies the following space-fractional differential equation$$D_{x}^{\alpha }f_{G_{\alpha }}=-(1-e^{-\partial _{t}})f_{G_{\alpha /2}},\qquad x,t\geq 0, \label{ai2}$$with initial condition$$f_{G_{\alpha }}(0,t)=0. \label{ai3}$$In order to get (\[ai2\]) we apply a well-known result on stable subordinators: the density $h_{\alpha }$ satisfies, for $\alpha \in (0,1],$ the following equation $$D_{x}^{\alpha }h_{\alpha }=-\frac{\partial }{\partial t}h_{\alpha },\quad h_{\alpha }(x,0)=\delta (x),\qquad x,t\geq 0.$$Thus$$\begin{aligned} D_{x}^{\alpha }f_{G_{\alpha }}(x,t) &=&\int_{0}^{\infty }D_{x}^{\alpha }h_{\alpha }(x,z)f_{\Gamma }(z,t)dz \\ &=&-\int_{0}^{\infty }\frac{\partial }{\partial z}h_{\alpha }(x,z)f_{\Gamma }(z,t)dz \\ &=&[h_{\alpha }(x,z)f_{\Gamma }(z,t)]_{z=0}^{\infty }+\int_{0}^{\infty }h_{\alpha }(x,z)\frac{\partial }{\partial z}f_{\Gamma }(z,t)dz,\end{aligned}$$which, by considering (\[res\]), with $b=1,$ gives (\[ai2\]). The initial condition is trivially satisfied since $h_{\alpha }(0,z)=0$. Equations (\[mee4\]) and (\[mee5\]) can be easily obtained by considering theorems 3 and 6. Also in this case, only for $\nu >1$, the fractional GS process is still a subordinator and its Lévy symbol can be obtain as follows:$$\begin{aligned} \eta _{\overline{G}_{\alpha }^{\nu }}(u) &=&-\psi _{\overline{\Gamma }_{\nu }}(-\eta _{\mathcal{A}_{\alpha }}(u)) \\ &=&-\left\{ \log [1+(-iu)^{\alpha }]\right\} ^{1/\nu },\end{aligned}$$since $\eta _{\mathcal{A}_{\alpha }}(u)=-(-iu)^{\alpha }.$ Fractional Negative Binomial process ------------------------------------ The NB process $M(t),t>0,$ is a jump Lévy process with the following distribution $$q_{k}(t):=\Pr \left\{ M(t)=k\right\} =\binom{t+k-1}{k}p^{t}(1-p)^{k},\qquad t>0,\text{ }k\in \mathbb{N},\text{ }p\in \left( 0,1\right) , \label{uno}$$where $\binom{x}{k}$ is the generalized binomial coefficient defined, for any $x\in \mathbb{R}$, as $$\binom{x}{k}:=\frac{(x)_{k}}{k!}=\frac{x(x-1)...(x-k+1)}{k!}$$and $(x)_{k}$ is the falling factorial. The NB process has two alternative representations (see [@KOZ2]). The first one is in terms of compound Poisson process: let $X_{j}$ be i.i.d. random variables with discrete logarithmic distribution$$\Pr \left\{ X_{j}=k\right\} =-\frac{(1-p)^{k}}{k\ln p},\qquad k=1,2,... \label{log}$$for any $j=1,2,...$, then the process defined as$$M(t):=\sum_{j=1}^{N_{\lambda }(t)}X_{j},\qquad t\geq 0, \label{due}$$where $N_{\lambda }$ denotes an homogeneous Poisson process of parameter $\lambda =\ln (1/p)$ (independent from $X_{j},$ for any $j$), is proved to be equivalent to the process defined by (\[uno\]). Thus $M$ can be considered as a particular case of the so called continuous-time random walks. Another possible representation of the NB process is in terms of Cox process with Gamma distributed directing measure, i.e. $N_{\Lambda (0,t]},$ $t\geq 0, $ where $\Lambda (0,t]$ is a Gamma r.v. with shape parameter $t$ and scale parameter $(1-p)/p$. Thus we immediately get the following equality in distribution $$M(t)\overset{d}{=}N_{1}(\Gamma (t)),\qquad t\geq 0, \label{coxx}$$where $\Gamma $ has density (\[gam\]) with $a=1,$ $b=p/(1-p).$ \[lm:eqNB\] The distribution of the NB process of parameter $p,$ given in (\[uno\]), satisfies the following differential equation:$$\left\{ \begin{array}{l} \frac{d}{dt}q_{0}(t)=\ln p\,q_{0}(t) \\ \frac{d}{dt}q_{k}(t)=\left[ \ln p+\sum_{j=0}^{k-1}\frac{1}{t+j}\right] q_{k}(t),\qquad k\geq 1\end{array}\right. , \label{lem1}$$for $t\geq 0,$ with initial conditions $$q_{k}(0)=\left\{ \begin{array}{c} 1,\qquad k=0 \\ 0,\qquad k>0\end{array}\right. .$$**Proof** We take the derivative of (\[uno\]), by considering that$$\binom{t+k-1}{k}=\frac{(t)^{(k)}}{k!}, \label{a1}$$where $(x)^{(k)}$ is the rising factorial defined as $(x)^{(k)}:=\Gamma (x+k)/\Gamma (x)$:$$\begin{aligned} \frac{d}{dt}q_{k}(t) &=&\frac{(1-p)^{k}}{k!}\frac{d}{dt}\left[ p^{t}(t)^{(k)}\right] \\ &=&\frac{(1-p)^{k}}{k!}\left[ p^{t}\ln p\,(t)^{(k)}+p^{t}\frac{d}{dt}(t)^{(k)}\right] \\ &=&\ln p\,q_{k}(t)+\frac{(1-p)^{k}}{k!}p^{t}(t)^{(k)}\left[ \Psi ^{(0)}(t+k)-\Psi ^{(0)}(t)\right] ,\end{aligned}$$where, in the last step, we have used the well-known result on the derivative of the rising factorial and $\Psi ^{(0)}$ denotes the Digamma function defined as $\Psi ^{(0)}(x):=d\ln \Gamma (x)/dx$ (see, for example, [@CHOI]). By the properties of the Digamma function it is easy to see that $\Psi ^{(0)}(t+k)-\Psi ^{(0)}(t)=\sum_{j=0}^{k-1}\frac{1}{t+j}$, so that we get (\[lem1\]). The initial conditions are trivially satisfied, by considering that $(0)^{(k)}=0$ for any $k>0$ and $(0)^{(k)}=1$ for $k=0.$$\square $ We now define a fractional version of the NB process by substituting in ([coxx]{}) the standard Gamma subordinator with its fractional counterpart defined in (\[pr6\]). Thus we set$$\left\{ \begin{array}{l} M_{\nu }(t):=N_{1}(\Gamma _{\nu }(t)),\text{\qquad }\nu \in (0,1) \\ \overline{M}_{\nu }(t):=N_{1}(\overline{\Gamma }_{\nu }(t)),\qquad \nu \in (1,\infty )\end{array}\right. , \label{af}$$for any $t\geq 0.$ For $\nu =1$, $M_{\nu }(t):=M(t).$ From (\[pr\]) and (\[sec\]), it is clear that $$\left\{ \begin{array}{l} M_{\nu }(t)\overset{d}{=}M(\mathcal{L}_{\nu }(t))\text{,\qquad }\nu \in (0,1) \\ \overline{M}_{\nu }(t)\overset{d}{=}M(\mathcal{A}_{1/\nu }(t))\text{,\qquad }\nu \in (1,\infty )\end{array}\right. . \label{coxx2}$$Unfortunately, due to the presence of non-constant coefficients in ([lem1]{}), we are not able to derive its fractional analogue. By resorting to the compound Poisson representation, it is proved in [@BEG2] that, for $\nu <1$, the distribution $q_{k}^{\nu }(t):=\Pr \{M_{\nu }(t)=k\}$ satisfies the following birth-type (or Kolmogorov forward) fractional equations$$\left\{ \begin{array}{l} D_{t}^{\nu }q_{0}^{\nu }=\ln pq_{0}^{\nu } \\ D_{t}^{\nu }q_{k}^{\nu }=\ln pq_{k}^{\nu }-\ln p\sum_{i=1}^{k}\frac{(1-p)^{i}}{i\ln p}q_{k-i}^{\nu },\qquad k>0\end{array}\right. \label{vec}$$with initial conditions$$q_{0}^{\nu }(0)=1,\quad q_{k}^{\nu }(0)=0\text{, for all integer }k>0,$$Analogously, for $\nu >1$, the distribution $\overline{q}_{k}^{\nu }(t):=\Pr \{\overline{M}_{\nu }(t)=k\}$ satisfies $$\left\{ \begin{array}{l} \mathcal{D}_{-,t}^{\nu }q_{0}^{\nu }=-\ln pq_{0}^{\nu } \\ \mathcal{D}_{-,t}^{\nu }q_{k}^{\nu }=-\ln pq_{k}^{\nu }+\ln p\sum_{i=1}^{k}\frac{(1-p)^{i}}{i\ln p}q_{k-i}^{\nu },\qquad k>0\end{array}\right. \label{vec2}$$with initial conditions$$\overline{q}_{0}^{\nu }(0)=1,\quad \overline{q}_{k}^{\nu }(0)=0\text{, for all integer }k>0.$$By applying the results of section 2, we can obtain alternative equations satisfied by the distribution of the NB process and its fractional versions defined in (\[af\]), in terms of the shift and fractional shift operators. The distribution of the NB process $M$ solves the following fractional differential equations, for $t>1,$$$e^{-D_{t}}q_{k}=\frac{1}{p}q_{k}-\frac{1-p}{p}q_{k-1},\qquad k\geq 0, \label{gam3}$$with initial condition$$q_{0}(0)=1,\quad q_{k}(0)=0\text{, for any integer }k>0$$and $q_{-1}(t)=0$, for any $t.$ From (\[coxx\]) we can write that$$q_{k}(t)=\int_{0}^{+\infty }p_{k}(z)f_{\Gamma }(z,t)dz, \label{gam4}$$where $p_{k}(t):=\Pr \{N_{1}(t)=k\}$ is the distribution of a Poisson process with intensity $1$. By applying to (\[gam4\]) the shift operator (\[shi\]) with $c=-1$, we get, for any $k\geq 0,$$$\begin{aligned} e^{-D_{t}}q_{k}(t) &=&\int_{0}^{+\infty }p_{k}(z)e^{-\partial _{t}}f_{\Gamma }(z,t)dz \label{dim} \\ &=&[\text{by (\ref{res}) with }b=p/(1-p)] \notag \\ &=&\frac{1-p}{p}\int_{0}^{+\infty }p_{k}(z)\frac{\partial }{\partial z}f_{\Gamma }(z,t)dz+\int_{0}^{+\infty }p_{k}(z)f_{\Gamma }(z,t)dz \notag \\ &=&[\text{integrating by parts}] \notag \\ &=&\frac{1-p}{p}\left[ p_{k}(z)f_{\Gamma }(z,t)\right] _{z=0}^{z=+\infty }-\frac{1-p}{p}\int_{0}^{+\infty }\frac{\partial }{\partial z}p_{k}(z)f_{\Gamma }(z,t)dz+q_{k}(t) \notag \\ &=&[\text{by (\ref{gam})-(\ref{res})}] \notag \\ &=&-\frac{1-p}{p}\int_{0}^{+\infty }D_{z}p_{k}(z)f_{\Gamma }(z,t)dz+q_{k}(t), \notag\end{aligned}$$which coincides with (\[gam3\]), by considering the well-known equation satisfied by the Poisson distribution. By applying theorems 3 and 6, we obtain the fractional equations satisfied by the distribution of the fractional NB processes defined in (\[coxx2\]). The distribution $q_{k}^{\nu }(t)$ satisfies the following fractional differential equations:$$\mathcal{O}_{-1,t}^{\nu }q_{k}^{\nu }=\frac{1}{p}q_{k}^{\nu }-\frac{1-p}{p}q_{k-1}^{\nu },\qquad k\geq 0,\text{ }t>1,\qquad \nu \in (0,1) \label{gam5}$$(where $\mathcal{O}_{-1,t}^{\nu }$ is defined in (\[fr1\])). The distribution $\overline{q}_{k}^{\nu }(t)$ satisfies$$\overline{\mathcal{O}}_{-1,t}^{\nu }\overline{q}_{k}^{\nu }=\frac{1}{p}\overline{q}_{k}^{\nu }-\frac{1-p}{p}\overline{q}_{k-1}^{\nu },\qquad k\geq 0,\text{ }t>1,\qquad \nu >1 \label{gam6}$$(where $\overline{\mathcal{O}}_{-1,t}^{\nu }$ is defined in (\[fr2\])) The initial conditions for both equations are $$q_{0}(0)=1,\quad q_{k}(0)=0\text{, for any integer }k>0$$and $q_{-1}(t)=0$, for any $t,$ where we set $q:=q^{\nu }$, for $\nu <1$ and $q:=\overline{q}^{\nu }$, for $\nu >1.$ By (\[af\]) we get, for $\nu <1$,.$$q_{k}^{\nu }(t)=\int_{0}^{+\infty }p_{k}(z)f_{\Gamma _{\nu }}(z,t)dz,$$thus we can write that$$\begin{aligned} \mathcal{O}_{-1,t}^{\nu }q_{k}^{\nu } &=&\int_{0}^{+\infty }p_{k}(z)\mathcal{O}_{-1,t}^{\nu }f_{\Gamma _{\nu }}(z,t)dz \\ &=&\left[ \text{by (\ref{pr4}), with }b=p/(1-p)\right] \\ &=&\frac{1-p}{p}\int_{0}^{+\infty }p_{k}(z)\frac{\partial }{\partial z}f_{\Gamma _{\nu }}(z,t)dz+\int_{0}^{+\infty }p_{k}(z)f_{\Gamma _{\nu }}(z,t)dz.\end{aligned}$$We obtain equation (\[gam5\]), by some steps similar to (\[dim\]). The case $\nu >1$ can be treated analogously. As happens with the fractional VG process, by definition (\[coxx2\]), it is clear that, in the case $\nu >1$, also $\overline{M}_{\nu }$ is a Lévy process, since the same is true for $M$ and $\mathcal{A}_{1/\nu }$ is a subordinator. The Lévy symbol can be obtain by evaluating the characteristic function of the process:$$\begin{aligned} \mathbb{E}e^{iu\overline{M}_{\nu }(t)} &=&\sum_{k=0}^{\infty }e^{iuk}q_{k}^{\nu }(t) \\ &=&\sum_{k=0}^{\infty }e^{iuk}\int_{0}^{+\infty }q_{k}(z)h_{1/\nu }(z,t)dz \\ &=&\int_{0}^{+\infty }\mathbb{E}e^{iuM(z)}h_{1/\nu }(z,t)dz \\ &=&\int_{0}^{+\infty }\left( \frac{p}{1-(1-p)e^{iu}}\right) ^{z}h_{1/\nu }(z,t)dz \\ &=&\int_{0}^{+\infty }\exp \left\{ -z\log \frac{1-(1-p)e^{iu}}{p}\right\} h_{1/\nu }(z,t)dz \\ &=&\exp \left\{ -t\left( \log \frac{1-(1-p)e^{iu}}{p}\right) ^{1/\nu }\right\} .\end{aligned}$$Thus we get$$\eta _{\overline{M}_{\nu }}(u)=-\left( \log \frac{1-(1-p)e^{iu}}{p}\right) ^{1/\nu },$$which coincides with$$-\psi _{\mathcal{A}_{1/\nu }}(-\eta _{M}(u))=-\left( -\eta _{M}(u)\right) ^{1/\nu },$$where $\eta _{M}(u)=\log \frac{p}{1-(1-p)e^{iu}}$ (see formula (1.1) in [KOZ2]{}). For the reader’s convenience we sum up the results on the Lévy symbols obtained so far in the following table: recall that in all the cases below it is $\nu >1.$ For $\nu =1$ we obtain the well-known Lévy symbols of the corresponding non-fractional processes. $\begin{array}{ll} \text{\textbf{Process}} & \text{\textbf{L\'{e}vy symbol}} \\ \text{Fractional Gamma }\overline{\Gamma }_{\nu } & \eta _{\overline{\Gamma }_{\nu }}(u)=-\left[ \log (1-iu/b)\right] ^{1/\nu } \\ \text{Fractional VG }\overline{X}_{\nu } & \eta _{\overline{X}_{\nu }}(u)=-\left[ \log \left( 1+\frac{u^{2}}{2}\right) \right] ^{1/\nu } \\ \text{Fractional GS }\overline{G}_{\alpha }^{\nu } & \eta _{\overline{G}_{\alpha }^{\nu }}(u)=-\left[ \log \left( 1+(-iu)^{\alpha }\right) \right] ^{1/\nu } \\ \text{Fractional NB }\overline{M}_{\nu } & \eta _{\overline{M}_{\nu }}(u)=-\left[ \log \frac{1-(1-p)e^{iu}}{p}\right] ^{1/\nu }\end{array}$ [99]{} **Applebaum D. (2009)**,** ***Lévy Processes and Stochastic Calculus*, Cambridge Studies in Advanced Mathematics, Cambridge. **Babusci D., Dattoli G., Sacchetti D**. **(2010)**, Integral equations, fractional calculus and shift operator, *arXiv*: 1007.5211v1 \[math-ph\]. **Beghin L. (2013)**, Geometric stable processes and fractional differential equation related to them, *arXiv* 1304.7915v1 \[math-ph\], submitted. **Beghin L., Macci C. (2012)**, Alternative forms of compound fractional Poisson processes, *Abstract Appl. Anal.*, 2012, 1-30. **Beghin L., Macci C. (2013)**, Fractional discrete processes: compound and mixed Poisson representations, *Journ. Appl. Probab.,* in press. **Bogdan K., Byczkowski T., Kulczycki T., Ryznar M., Song R., Vondracek Z. (2009)**, *Potential Analysis of Stable Processes and its Extensions***,** Lecture Notes in Math., Editors P.Graczyk, A.Stos, Elsevier. **Choi, J., Srivastava, H.M., Kim, Y. (2001)**, Applications of a certain family of hypergeometric summation formulas associated with Psi and Zeta functions, *Comm. Korean Math. Soc.,* 16, 319-332. **Cont R., Tankov P. (2008)** *Financial Modelling with Jump Processes*, 2nd Ed., Chapman & Hall / CRC Press. **Dattoli G., Ricci P.E., Sacchetti D**. **(2003)**, Generalized shift operators and pseudo-polynomials of fractional order, *Appl. Math. Comp.,* 141, 215-224. **D’Ovidio M.** **(2011),** On the fractional counterpart of the higher-order equations, *Stat. Prob. Lett.*, 81, (12), 1929-1939. **Hahn M. G., Kobayashi K., Umarov, S.** **(2011),** Fokker-Plank-Kolmogorov equations associated with time-changed fractional Brownian motion. *Proceed. of the Amer. Math. Soc.*, 139, (2), 691–705. **Jumarie, G. (2009)**, Probability calculus of fractional order and fractional Taylor’s series application to Fokker-Plank equation and information of non-random functions, *Chaos, Solitons and Fractals,* 40, 1428-1448. **Kilbas A.A., Srivastava H.M., Trujillo J.J. (2006),** *Theory and Applications of Fractional Differential Equations*, vol. 204 of North-Holland Mathematics Studies, Elsevier Science B.V., Amsterdam. **Kotz S., Kozubowski T.J., Podgòrski K. (2001)**, *The Laplace Distribution and Generalizations: A Revisit with Applications to Communications, Economics, Engineering and Finance*, Birkhauser, Boston, MA. **Linde W., Shi Z. (2004)**, Evaluating the small deviations probabilities for subordinated Lévy processes, *Stochastic Processes and their Applications,* 113, 273-287. **Madan D.B., Seneta E. (1990)**, The variance gamma (V.G.) model for share market returns, *Journ. Business*, 63, 511–524. **Miškinis P. (2012),** On Integral Representation of the Translation Operator, *Mathematical Modelling and Analysis*, 17, (1), 100-112. **Kozubowski T.J.(1999),** Univariate Geometric Stable Laws, *Journ. Comp. Anal. Appl.*, 1, (2), 177-217. **Kozubowski T.J., Meerschaert M.M., Podgòrski K. (2006)**, Fractional Laplace motion, *Adv. Appl. Probab.,* 38, 451-464. **Kozubowski, T.J., Podgórski, K. (2009),** Distributional properties of the Negative Binomial Lévy process. *Probab. Math. Statist.* 29, 43–71. **Osler, T J. (1971)** Taylor’s series generalized for fractional derivatives and applications. *SIAM J. Math. Anal.*, 2, 37–48. **Samorodnitsky G., Taqqu M.S. (1994)** *Stable Non-Gaussian Random Processes*, Chapman and Hall, New York. **Sikic H., Song R., Vondracek Z. (2006)**, Potential theory of geometric stable processes, *Probab. Theory Relat. Fields,* 135, 547–575. **Trujillo J. J., Rivero M., Bonilla, B. (1999)** On a Riemann–Liouville generalized Taylor’s formula, *J. Math. Anal*. *Appl*., 231, 255–265. [^1]: Address: Department of Statistical Sciences, Sapienza University of Rome, P.le A. Moro 5, I-00185 Roma, Italy. e-mail: `[email protected]`
{ "pile_set_name": "ArXiv" }
--- abstract: 'The theory of finitely supported algebraic structures is related to Pitts theory of nominal sets (by equipping finitely supported sets with finitely supported internal algebraic laws). It represents a reformulation of Zermelo Fraenkel set theory obtained by requiring every set theoretical construction to be finitely supported according to a certain action of a group of permutations of some basic elements named atoms. Its main purpose is to let us characterize infinite algebraic structures, defined involving atoms, only by analyzing their finite supports. The first goal of this paper is to define and study different kinds of infinities and the notion of ‘cardinality’ in the framework of finitely supported structures. We present several properties of infinite cardinalities. Some of these properties are extended from the non-atomic Zermelo Fraenkel set theory into the world of atomic objects with finite support, while other properties are specific to finitely supported structures. We also compare alternative definitions of ‘infinite finitely supported set’, and we finally provide a characterization of finitely supported countable sets.' author: - | Andrei Alexandru\ Romanian Academy, Institute of Computer Science\ Iaşi, Romania\ `[email protected]`\ Gabriel Ciobanu\ Romanian Academy, Institute of Computer Science\ and A.I.Cuza University of Iaşi, Romania\ `[email protected]` title: Infinities within Finitely Supported Structures --- Introduction ============ The theory of finitely supported algebraic structures which is known under the name of ‘nominal sets’ (when dealing with computer science applications) or ‘Finitely Supported Mathematics’ (in some pure set theoretical papers related to the foundations of mathematics) represents an alternative framework for working with infinite structures hierarchically constructed by involving some basic elements (called atoms) by dealing only with a finite number of entities that form their supports. The theory of nominal sets is presented in a categorical manner as a Zermelo-Fraenkel (ZF) alternative to Fraenkel and Mostowski 1930s permutation models of set theory with atoms [@pitts-2]. A nominal set is defined as a usual ZF set endowed with a group action of the group of (finitary) permutations over a certain fixed countable ZF set $A$ (also called the set of atoms by analogy with the Fraenkel and Mostowski framework) formed by elements whose internal structure is not taken into consideration (i.e. by elements that can be checked only for equality), satisfying a finite support requirement. This requirement states that for any element in a nominal set there should exist a finite set of atoms such that any permutation fixing pointwise this set of atoms also leaves the element invariant under the related group action. Nominal sets represents a categorical mathematical theory of names studying scope, binding, freshness and renaming in formal languages based upon symmetry. Inductively defined finitely supported sets (that are finitely supported elements in the powerset of a nominal set) involving the name-abstraction together with Cartesian product and disjoint union can encode syntax modulo renaming of bound variables. In this way, the standard theory of algebraic data types can be extended to include signatures involving binding operators. In particular, there is an associated notion of structural recursion for defining syntax-manipulating functions and a notion of proof by structural induction. Various generalizations of nominal were used in order to study automata, languages or Turing machines that operate over infinite alphabets; for this a relaxed notion of finiteness, called ‘orbit finiteness’, was defined and means ‘having a finite number of orbits under a certain group action’ [@boj]. Finitely Supported Mathematics (FSM) is an alternative name for nominal algebraic structures, used in theoretical papers focused on the foundations of set theory (rather than on applications in computer science). In order to describe FSM as a theory of finitely supported algebraic structures (that is finitely supported sets *together with finitely supported internal algebraic laws*), we use nominal sets (without the requirement that the set $A$ of atoms is countable) which by now on will be called invariant sets motivated by Tarski’s approach regarding logicality (i.e. a logical notion is defined by Tarski as one that is invariant under the permutations of the universe of discourse). The cardinality of the set of atoms *cannot* be internally compared with any other ZF cardinality, and so we just say that atoms form an infinite set without any specifications regarding its cardinality. In FSM we actually study the finitely supported subsets of invariant sets together with finitely supported relations (order relations, functions, algebraic laws etc), and so FSM becomes a theory of atomic algebraic structures constructed/defined according to the finite support requirement. The requirement of being finitely supported under a canonical action of the group of permutation of atoms (constructed under the rules in Proposition \[p1\]) is actually an axiom adjoined to ZF, and so non-finitely supported structures are not allowed (they do not exist) in FSM. FSM contains the family of ‘non-atomic’ (ordinary) ZF sets (which are proved to be trivial FSM sets) and the family of ‘atomic’ sets with finite supports (hierarchically constructed from the empty set and the fixed ZF set $A$). The main question now is whether a classical ZF result (obtained in ZF framework for non-atomic sets) can be adequately reformulated by replacing ‘non-atomic element/set’ with ‘atomic finitely supported element/set’ (according to the canonical actions of the group of one-to-one transformations of $A$ onto itself) in order to be valid also for atomic sets with finite supports. The (non-atomic) ZF results cannot be directly translated into the framework of atomic finitely supported sets, unless we are able to reprove their new formulations internally in FSM, i.e. by involving only *finitely supported structures* even in the intermediate steps of the proof. This is because the family of finitely supported sets is not closed under subset constructions, and we cannot use something outside FSM in order to prove something in FSM. The meta-theoretical techniques for the translation of a result from non-atomic structures to atomic structures are fully described in [@book] (or in [@pitts-2], with the mention that, working on foundations of mathematics, and so we use a slightly different terminology for the same concept). They are based on a refinement of the finite support principle form [@pitts-2] called “$S$-finite supports principle" claiming that for any finite set $S$ of atoms, anything that is definable in higher order logic from $S$-supported structures using $S$-supported constructions is also $S$-supported. The formal involvement of the $S$-finite support principles implies a constructive method for defining the support of a structure by employing the supports of the sub-structures of a related structure. In this paper we introduce the notion of ‘cardinality’ of a finitely supported set, and we prove several properties of this concept. Some properties are naturally extended from the non-atomic ZF into the world of atomic structures. In this sense we prove that Cantor-Schr[ö]{}der-Bernstein theorem for cardinalities is still valid in FSM. Several other cardinality properties are preserved from ZF. However, although Cantor-Schr[ö]{}der-Bernstein theorem can be successfully translated into FSM, its ZF dual is no longer valid in FSM. Other specific FSM properties of cardinalities (that do not have related ZF correspondents) are also emphasized. We introduce various definition for infinity and we compare them, providing relevant examples of atomic sets verifying the conditions of each such definition. Finally, we introduce and study the concept of countability in FSM. Finitely Supported Sets {#FMset} ======================= A ZF finite set is referred to a set for which there is a bijection with a finite ordinal; a ZF infinite set is a set that is not finite. Adjoin to ZF a special infinite set $A$ (called ‘the set of atoms’; despite classical set theory with atoms we do not need to modify the axiom of extensionality). Actually, atoms are entities whose internal structure is considered to be irrelevant which are considered as basic for a higher-order construction, i.e. their internal structure is not taken into consideration. A *transposition* is a function $(a\, b):A\to A$ given by $(a\, b)(a)=b$, $(a\, b)(b)=a$ and $(a\, b)(n)=n$ for $n\neq a,b$. A *(finitary) permutation* of $A$ in FSM is a one-to-one transformation of $A$ onto itself (a bijection of $A$) generated by composing finitely many transpositions. We denote by $S_{A}$ the set of all finitary permutations of $A$. According to Proposition 2.6 from [@book], a function $f:A \to A$ is a bijection on $A$ in FSM if and only if it leaves unchanged all but finitely many elements of $A$. Thus, in FSM a function is a one-to-one transformation of $A$ onto itself if and only if it is a (finitary) permutation of $A$. Thus, the notions ‘permutation (bijection) of $A$’ and ‘finitary permutation of $A$’ coincide in FSM. \[2.4\] Let $X$ be a ZF set. 1. An *$S_{A}$-action* on $X$ is a function $\cdot:S_{A}\times X\rightarrow X$ having the properties that $Id\cdot x=x$ and $\pi\cdot(\pi'\cdot x)=(\pi\circ\pi')\cdot x$ for all $\pi,\pi'\in S_{A}$ and $x\in X$, where $Id$ is the identity mapping on $A$. An *$S_{A}$-set* is a pair $(X,\cdot)$ where $X$ is a ZF set, and $\cdot:S_{A}\times X\to X$ is an $S_{A}$-action on $X$. 2. Let $(X,\cdot)$ be an $S_{A}$-set. We say that *$S\subset A$ supports $x$* whenever for each $\pi\in Fix(S)$ we have $\pi\cdot x=x$, where $Fix(S)=\{\pi\,|\,\pi(a)=a,\forall a\in S\}$. The least finite set supporting $x$ (which exists according to Proposition \[p11\]) is called *the support of $x$* and is denoted by $supp(x)$. An empty supported element is called *equivariant*; this means that $x \in X$ is equivariant if and only if $\pi \cdot x=x$, $\forall \pi \in S_{A}$. 3. Let $(X,\cdot)$ be an $S_{A}$-set. We say that $X$ is an *invariant set* if for each $x\in X$ there exists a finite set *$S_{x}\subset A$* which supports $x$. \[p11\] [@book] Let $X$ be an $S_{A}$-set and let $x\in X$. If there exists a finite set supporting $x$ (particularly, if $X$ is an invariant set), then there exists a least finite set supporting $x$ which is constructed as the intersection of all finite sets supporting $x$. \[2.15\] [@book] Let $(X,\cdot)$ be an $S_{A}$-set, and $\pi\in S_{A}$. If $x\in X$ is finitely supported, then $\pi\cdot x$ is finitely supported and $supp(\pi\cdot x)=\pi(supp(x))$. \[2.7\]    1. The set $A$ of atoms is an $S_{A}$-set with the $S_{A}$-action $\cdot:S_{A}\times A\rightarrow A$ defined by $\pi\cdot a:=\pi(a)$ for all $\pi\in S_{A}$ and $a\in A$. $(A,\cdot)$ is an invariant set because for each $a\in A$ we have that $\{a\}$ supports $a$. Furthermore, $supp(a)=\{a\}$ for each $a\in A$. 2. The set $S_{A}$ is an $S_{A}$-set with the $S_{A}$-action $\cdot:S_{A}\times S_{A}\rightarrow S_{A}$ defined by $\pi\cdot\sigma:=\pi\circ\sigma\circ\pi^{-1}$ for all $\pi,\sigma\in S_{A}$. $(S_{A},\cdot)$ is an invariant set because for each $\sigma\in S_{A}$ we have that the finite set $\{a\in A\,|\,\sigma(a)\neq a\}$ supports $\sigma$. Furthermore, $supp(\sigma)=\{a\in A\,|\,\sigma(a)\neq a\}$ for each $\sigma\in S_{A}$. 3. Any ordinary (non-atomic) ZF-set $X$ (such as $\mathbb{N},\mathbb{Z},\mathbb{Q}$ or $\mathbb{R}$ for example) is an invariant set with the single possible $S_{A}$-action $\cdot:S_{A}\times X\rightarrow X$ defined by $\pi\cdot x:=x$ for all $\pi \in S_{A}$ and $x\in X$. \[p1\] Let $(X,\cdot)$ and $(Y,\diamond)$ be $S_{A}$-sets. 1. The Cartesian product $X\times Y$ is also an $S_{A}$-set with the $S_{A}$-action $\otimes:S_{A}\times(X\times Y)\rightarrow(X\times Y)$ defined by $\pi\otimes(x,y)=(\pi\cdot x,\pi\diamond y)$ for all $\pi\in S_{A}$ and all $x\in X$, $y\in Y$. If $(X,\cdot)$ and $(Y,\diamond)$ are invariant sets, then $(X\times Y,\otimes)$ is also an invariant set. 2. The powerset $\wp(X)=\{Z\,|\, Z\subseteq X\}$ is also an $S_{A}$-set with the $S_{A}$-action $\star: S_{A}\times\wp(X) \rightarrow \wp(X)$ defined by $\pi\star Z:=\{\pi\cdot z\,|\, z\in Z\}$ for all $\pi \in S_{A}$, and all $Z \subseteq X$. For each invariant set $(X,\cdot)$, we denote by $\wp_{fs}(X)$ the set formed from those subsets of $X$ which are finitely supported according to the action $\star$ . $(\wp_{fs}(X),\star|_{\wp_{fs}(X)})$ is an invariant set, where $\star|_{\wp_{fs}(X)}$ represents the action $\star$ restricted to $\wp_{fs}(X)$. 3. The finite powerset of $X$ $\wp_{fin}(X)=\{Y \subseteq X\,|\, Y \text{finite}\}$ and the cofinite powerset of $X$ $\wp_{cofin}(X)=\{Y \subseteq X\,|\, X\setminus Y \text{finite}\}$ are $S_{A}$-sets with the $S_{A}$-action $\star$ defined as in item 2. If $X$ is an invariant set, then both $\wp_{fin}(X)$ and $\wp_{cofin}(X)$ are invariant sets. 4. Let $(X,\cdot)$ and $(Y,\diamond)$ be $S_{A}$-sets. We define the disjoint union of $X$ and $Y$ by $X+Y=\{(0,x)\,|\, x\in X\}\cup\{(1,y)\,|\, y\in Y\}$. $X+Y$ is an $S_{A}$-set with the $S_{A}$-action $\star:S_{A}\times(X+Y)\rightarrow(X+Y)$ defined by $\pi\star z=(0,\pi\cdot x)$ if $z=(0,x)$ and $\pi\star z=(1,\pi\diamond y)$ if $z=(1,y)$. If $(X,\cdot)$ and $(Y,\diamond)$ are invariant sets, then $(X+Y,\star)$ is also an invariant set: each $z\in X+Y$ is either of the form $(0,x)$ and supported by the finite set supporting $x$ in $X$, or of the form $(1,y)$ and supported by the finite set supporting $y$ in $Y$. \[2.14\] 1. Let $(X,\cdot)$ be an $S_{A}$-set. A subset $Z$ of $X$ is called *finitely supported* if and only if $Z\in\wp_{fs}(X)$ with the notations from Proposition \[p1\]. A subset $Z$ of $X$ is *uniformly supported* if all the elements of $Z$ are supported by the same set $S$ (and so $Z$ is itself supported by $S$ as an element of $\wp_{fs}(X)$). Generally, an FSM set is a finitely supported subset (possibly equivariant) of an invariant set. 2. Let $(X,\cdot)$ be a finitely supported subset of an $S_{A}$- set $(Y, \cdot)$. A subset $Z$ of $Y$ is called *finitely supported subset of $X$* (and we denote this by $Z \in \wp_{fs}(X)$) if and only if $Z\in\wp_{fs}(Y)$ and $Z \subseteq X$. Similarly, we say that a uniformly supported subset of $Y$ contained in $X$ is a *uniformly supported subset of $X$*. From Definition \[2.4\], a subset $Z$ of an invariant set $(X, \cdot)$ is finitely supported by a set $S \subseteq A$ if and only if $\pi \star Z \subseteq Z$ for all $\pi \in Fix(S)$. This is because any permutation of atoms should have finite order. \[4.4-9\] 1. Let $X$ be a finite subset of an invariant set $(U, \cdot)$. Then $X$ is finitely supported and $supp(X)=\cup\{supp(x)\,|\, x\in X\}$. 2. Let $X$ be a uniformly supported subset of an invariant set $(U, \cdot)$. Then $X$ is finitely supported and $supp(X)=\cup\{supp(x)\,|\, x\in X\}$. 1\. Let $X=\left\{ x_{1},\ldots, x_{k}\right\}$, and $S=supp(x_{1})\cup\ldots\cup supp(x_{k})$. Obviously, $S$ supports $X$. Indeed, let us consider $\pi\in Fix(S)$. We have that $\pi\in Fix(supp(x_{i}))$ for each $i\in\{1,\ldots ,k\}$. Therefore, $\pi\cdot x_{i}=x_{i}$ for each $i\in\{1,\ldots ,k\}$ because $supp(x_{i})$ supports $x_{i}$ for each $i\in\{1,\ldots ,k\}$, and so $supp(X) \subseteq S$. It remains to prove that $S \subseteq supp(X)$. Consider $a \in S$. This means there exists $j\in\{1,\ldots ,k\}$ such that $a \in supp(x_{j})$. Let $b$ be an atom such that $b \notin supp(X)$ and $b \notin supp(x_{i})$, $\forall i\in\{1,\ldots ,k\}$. Such an atom exists because $A$ is infinite, while $supp(X)$ and $supp(x_{i})$, $ i\in\{1,\ldots ,k\}$, are all finite. We prove by contradiction that $(b\; a) \cdot x_{j} \notin X$. Indeed, suppose that $(b\; a) \cdot x_{j} \in X$. In this case there is $y \in X$ with $(b\; a) \cdot x_{j}=y$. Since $a \in supp(x_{j})$, we have $b \in (b\; a)(supp(x_{j}))$. However, according to Proposition \[2.15\], we have $supp(y)=(b\; a)(supp(x_{j}))$. We obtain that $b \in supp(y)$ for some $y \in X$, which is a contradiction with the choice of $b$. Therefore, $(b\; a) \star X \neq X$, where $\star$ is the standard $S_{A}$-action on $\wp(U)$ is defined in Proposition \[p1\](2). Since $b \notin supp(X)$, we prove by contradiction that $a \in supp(X)$. Indeed, suppose that $a \notin supp(X)$. It follows that the transposition $(b\; a)$ fixes each element from $supp(X)$, i.e. $(b\; a) \in Fix(supp(X))$. Since $supp(X)$ supports $X$, by Definition \[2.4\], it follows that $(b\; a) \star X=X$, which is a contradiction. Thus, $a \in supp(X)$, and so $S \subseteq supp(X)$. 2\. Since $X$ is uniformly supported, there exists a finite subset of atoms $T$ such that $T$ supports every $x \in X$, i.e. $supp(x) \subseteq T$ for all $x \in X$. Thus, $\cup\{supp(x)\,|\, x\in X\} \subseteq T$. Clearly, $supp(X) \subseteq \cup\{supp(x)\,|\, x\in X\}$. Conversely, let $a \in \cup\{supp(x)\,|\, x\in X\}$. Thus, there exists $x_{0} \in X$ such that $a \in supp(x_{0})$. Let $b$ be an atom such that $b \notin supp(X)$ and $b \notin T$. Such an atom exists because $A$ is infinite, while $supp(X)$ and $T$ are both finite. We prove by contradiction that $(b\; a) \cdot x_{0} \notin X$. Indeed, suppose that $(b\; a) \cdot x_{0}=y \in X$. Since $a \in supp(x_{0})$, we have $b =(b\;a)(a) \in (b\; a)(supp(x_{0}))=supp((b\; a) \cdot x_{0})=supp(y)$. Since $supp(y) \subseteq T$, we get $b \in T$, a contradiction. Therefore, $(b\; a) \star X \neq X$. Since $b \notin supp(X)$, we have that $a \in supp(X)$ as in the above item. Let $X$ be a uniformly supported subset of an invariant set. Then $X$ is uniformly supported by $supp(X)$. Since $supp(X)=\cup\{supp(x)\,|\, x\in X\}$, we have $supp(x) \subseteq supp(X)$ for all $x \in X$ which means $supp(X)$ supports every $x \in X$. \[p111\] We have $\wp_{fs}(A)=\wp_{fin}(A) \cup \wp_{cofin}(A)$. We know that $B$ is finitely supported with $supp(B)=B$ whenever $B \subset A$ and $B$ is finite. If $C \subseteq A$ and $C$ is cofinite, then $C$ is finitely supported by $A \setminus C$ with $supp(C)=A \setminus C$. However, if $D \subsetneq A$ is neither finite nor cofinite, then $D$ is not finitely supported. Indeed, assume by contradiction that there exists a finite set of atoms $S$ supporting $D$. Since $S$ is finite and both $D$ and its complementary $C_{D}$ are infinite, we can take $a \in D \setminus S$ and $b \in C_{D} \setminus S$. Then the transposition $(a\,b)$ fixes $S$ pointwise, but $(a\,b) \star D \neq D$ because $(a\,b)(a)=b \notin D$; this contradicts the assertion that $S$ supports $D$. Therefore, $\wp_{fs}(A)=\wp_{fin}(A) \cup \wp_{cofin}(A)$. \[2.10-1\] Let $X$ and $Y$ be invariant sets. 1. A function $f:X\rightarrow Y$ is *finitely supported* if $f\in\wp_{fs}(X\times Y)$. The set of all finitely supported functions from $X$ to $Y$ is denoted by $Y^{X}_{fs}$. 2. Let $Z$ be a finitely supported subset of $X$ and $T$ a finitely supported subset of $Y$. A function $f:Z\rightarrow T$ is *finitely supported* if $f\in\wp_{fs}(X\times Y)$. The set of all finitely supported functions from $Z$ to $T$ is denoted by $T^{Z}_{fs}$. \[2.18’\] [@book] Let $(X,\cdot)$ and $(Y,\diamond)$ be two invariant sets. 1. $Y^{X}$ (i.e. the set of all functions from $X$ to $Y$) is an $S_{A}$-set with the $S_{A}$-action $\widetilde{\star}:S_{A}\times Y^{X}\rightarrow Y^{X}$ defined by $(\pi \widetilde{\star}f)(x) = \pi\diamond(f(\pi^{-1}\cdot x))$ for all $\pi\in S_{A}$, $f\in Y^{X}$ and $x\in X$. A function $f:X\rightarrow Y$ is finitely supported in the sense of Definition \[2.10-1\] if and only if it is finitely supported with respect the permutation action $\widetilde{\star}$. 2. Let $Z$ be a finitely supported subset of $X$ and $T$ a finitely supported subset of $Y$. A function $f:Z\rightarrow T$ is supported by a finite set $S \subseteq A$ if and only if for all $x \in Z$ and all $\pi \in Fix(S)$ we have $\pi \cdot x \in Z$, $\pi \diamond f(x) \in T$ and $f(\pi\cdot x)=\pi\diamond f(x)$. Particularly, a function $f:X\rightarrow Y$ is supported by a finite set $S \subseteq A$ if and only if for all $x \in X$ and all $\pi \in Fix(S)$ we have $f(\pi\cdot x)=\pi\diamond f(x)$. Cardinalities and Order Properties ================================== \[FM-event struc\]   - An *invariant partially ordered set (invariant poset)* is an invariant set $(P,\cdot)$ together with an equivariant partial order relation $\sqsubseteq$ on $P$. An invariant poset is denoted by $(P,\sqsubseteq,\cdot)$ or simply $P$. - A *finitely supported partially ordered set (finitely supported poset)* is a finitely supported subset $X$ of an invariant set $(P,\cdot)$ together with a partial order relation $\sqsubseteq$ on $X$ that is finitely supported as a subset of $P\times P$. Two FSM sets $X$ and $Y$ are called equipollent if there exists a finitely supported bijection $f:X \to Y$. The FSM cardinality of $X$ is defined as the equivalence class of all FSM sets equipollent to $X$, and is denoted by $|X|$. This means that for two FSM sets $X$ and $Y$ we have $|X|=|Y|$ if and only if there exists a finitely supported bijection $f:X \to Y$. On the family of cardinalities we can define the relations: - $\leq$ by:  $|X| \leq |Y|$ if and only if there is a finitely supported injective mapping $f:X \to Y$; - $\leq^{*}$ by: $|X| \leq^{*} |Y|$ if and only if there is a finitely supported surjective mapping $f:Y \to X$. \[cardord\] 1. The relation $\leq$ is equivariant, reflexive, anti-symmetric and transitive, but it is not total. 2. The relation $\leq^{*}$ is equivariant, reflexive and transitive, but it is not anti-symmetric, nor total. - $\leq$ and $\leq^{*}$ are equivariant because for any FSM sets $X$ and $Y$, whenever there is a finitely supported injection/ surjection $f:X \to Y$, according to Proposition \[2.15\], we have that $\pi \star f:\pi \star X \to \pi \star Y$, defined by $(\pi \star f)(\pi \cdot x)=\pi \cdot f(x)$ for all $x \in X$, is a finitely supported injective/surjective mapping, and so $\pi \star X$ is comparable with $\pi \star Y$ (under $\leq$ or $\leq^{*}$, after case). - $\leq$ and $\leq^{*}$ are obviously reflexive because for each FSM set $X$, the identity of $X$ is an equivariant bijection from $X$ to $X$. - $\leq$ and $\leq^{*}$ are transitive because for any FSM sets $X$, $Y$ and $Z$, whenever there are two finitely supported injections/surjections $f:X \to Y$ and $g:Y \to Z$, there exists an injection/surjection $g \circ f:X \to Z$ which is finitely supported by $supp(f) \cup supp(g)$. - The anti-symmetry of $\leq$. \[lem1\] Let $(B, \cdot)$ and $(C, \diamond)$ be two invariant sets. If there exist a finitely supported injective mapping $f: B \to C$ and a finitely supported injective mapping $g: C \to B$, then there exists a finitely supported bijective mapping $h:B \to C$. Furthermore, $supp(h) \subseteq supp(f) \cup supp(g)$. *Proof of Lemma \[lem1\].* Let us define $F:\wp_{fs}(B) \to \wp_{fs}(B)$ by $F(X)=B-g(C-f(X))$ for all finitely supported subsets $X$ of $B$. **Claim 1:** $F$ is correctly defined, i.e. $Im(F) \subseteq \wp_{fs}(B)$.\ For every finitely supported subset $X$ of $B$, we have that $f(X)$ is supported by $supp(f) \cup supp(X)$. Indeed, let $\pi \in Fix(supp(f) \cup supp(X))$. Let $y$ be an arbitrary element from $f(X)$; then $y=f(x)$ for some $x \in X$. However, because $\pi \in Fix (supp(X))$, it follows that $\pi \cdot x \in X$ and so, because $supp(f)$ supports $f$ and $\pi$ fixes $supp(f)$ pointwise, from Proposition \[2.18’\] we get $\pi \diamond y= \pi \diamond f(x)= f(\pi \cdot x) \in f(X)$. Thus $\pi \widetilde{\star} f(X)=f(X)$, where $\widetilde{\star}$ is the $S_{A}$-action on $\wp_{fs}(C)$ defined as in Proposition \[p1\]. Analogously, $g(Y)$ is finitely supported by $supp(g) \cup supp(Y)$ for all $Y \in \wp_{fs}(C)$. It is easy to remark that for every finitely supported subset $X$ of $B$ we have that $C-f(X)$ is also supported by $supp(f) \cup supp(X)$, $g(C-f(X))$ is supported by $supp(g) \cup supp(f) \cup supp(X)$, and $B-g(C-f(X))$ is supported by $supp(g) \cup supp(f) \cup supp(X)$. Thus, $F$ is well-defined. **Claim 2:** $F$ is a finitely supported function.\ We prove that $F$ is finitely supported by $supp(f) \cup supp(g)$. Let us consider $\pi \in Fix(supp(f) \cup supp(g))$. Since $\pi \in Fix(supp(f))$ and $supp(f)$ supports $f$, according to Proposition \[2.18’\] we have that $f(\pi \cdot x)=\pi \diamond f(x)$ for all $x \in B$. Thus, for every finitely supported subset $X$ of $B$ we have $f(\pi \star X)=\{f(\pi \cdot x)\;|\;x \in X\}=\{\pi \diamond f(x)\;|\;x \in X\}=\pi \widetilde{\star} f(X)$, where $\star$ is the $S_{A}$-action on $\wp_{fs}(B)$ and $\widetilde{\star}$ is the $S_{A}$-action on $\wp_{fs}(C)$. Similarly, $g(\pi \widetilde{\star} Y)=\pi \star g(Y)$ for any finitely supported subset $Y$ of $C$. Therefore, $F(\pi \star X)=B-g(C-f(\pi \star X))=B-g(C-\pi \widetilde{\star} f(X)) \overset{\pi \widetilde{\star}C=C}{=}B-g(\pi \widetilde{\star}(C-f(X)))=B-(\pi \star g(C-f(X))) \overset{\pi \star B=B}{=} \pi \star (B- g(C-f(X)))=\pi \star F(X)$. From Proposition \[2.18’\] it follows that $F$ is finitely supported. Moreover, because $supp(F)$ is the least set of atoms supporting $F$, we have $supp(F) \subseteq supp(f) \cup supp(g)$. **Claim 3:** For any $X,Y \in \wp_{fs}(B)$ with $X \subseteq Y$, we have $F(X) \subseteq F(Y)$. This remark follows by direct calculation. **Claim 4:** The set $S:=\{X\;|\; X \in \wp_{fs}(B), X \subseteq F(X)\}$ is a non-empty finitely supported subset of $\wp_{fs}(B)$. Obviously, $\emptyset \in S$. We claim that $S$ is supported by $supp(F)$. Let $\pi \in Fix(supp(F))$, and $X \in S$. Then $X \subseteq F(X)$. From the definition of $\star$ (see Proposition \[p1\]) we have $\pi \star X \subseteq \pi \star F(X)$. According to Proposition \[2.18’\], because $supp(F)$ supports $F$, we have $\pi \star X \subseteq \pi \star F(X)=F(\pi \star X)$, and so $\pi \star X \in S$. It follows that $S$ is finitely supported, and $supp(S) \subseteq supp(F)$. **Claim 5:** $T:=\underset{X \in S}{\cup}X$ is finitely supported by $supp(S)$.\ Let $\pi \in Fix(supp(S))$, and $t \in T$. Since $T=\underset{X \in S}{\cup}X$, we have that there exists $Z \in S$ such that $t \in Z$. Therefore, $\pi \cdot t \in \pi \star Z$. However, since $\pi$ fixes $supp(S)$ pointwise and $supp(S)$ supports $S$, we have that $\pi \star Z \in S$. Thus, there exists $Y \in S$ such that $\pi \star Z=Y$. Therefore $\pi \cdot t \in Y$, and so $\pi \cdot t \in \underset{X \in S}{\cup}X$. It follows that $\underset{X \in S}{\cup}X$ is finitely supported, and so $T=\underset{X \in S}{\cup}X \in \wp_{fs}(B)$. Furthermore, $supp(T) \subseteq supp(S)$. **Claim 6:** We prove that $F(T)=T$.\ Let $X \in S$ arbitrary. We have $X \subseteq F(X) \subseteq F(T)$. By taking the supremum on $S$, this leads to $T \subseteq F(T)$. However, because $T \subseteq F(T)$, from Claim 3 we also have $F(T) \subseteq F(F(T))$. Furthermore, $F(T)$ is supported by $supp(F) \cup supp(T)$ (i.e. by $supp(f) \cup supp(g)$), and so $F(T) \in S$. According to the definition of $T$, we get $F(T) \subseteq T$. We get $T=B-g(C-f(T))$, or equivalently, $B-T=g(C-f(T))$. Since $g$ is injective, we obtain that for each $x \in B-T$, $g^{-1}(x)$ is a set containing exactly one element. Let us define $h:B \to C$ by $$h(x)=\left\{ \begin{array}{ll} f(x), & \text{for}\: x \in T;\\ g^{-1}(x), & \text{for}\: x \in B-T.\end{array}\right.$$ **Claim 7:** We claim that $h$ is supported by the set $supp(f) \cup supp(g) \cup supp(T)$ (more exactly, by $supp(f) \cup supp(g)$, according to the previous claims). Let $\pi \in Fix(supp(f) \cup supp(g) \cup supp(T))$, and $x$ an arbitrary element of $B$. If $x \in T$, because $\pi \in Fix(supp(T))$ and $supp(T)$ supports $T$, we have $\pi \cdot x \in T$. Thus, from Proposition \[2.18’\] we get $h(\pi \cdot x)=f(\pi \cdot x)=\pi \diamond f(x)=\pi \diamond h(x)$. If $x \in B-T$, we have $\pi \cdot x \in B-T$. Otherwise, we would obtain the contradiction $x=\pi^{-1} \cdot (\pi \cdot x) \in T$ because $\pi^{-1}$ also fixes $supp(T)$ pointwise. Thus, because $g$ is finitely supported, according to Proposition \[2.18’\] we have $h(\pi \cdot x)=g^{-1}(\pi \cdot x)=\{y \in C\;|\; g(y)$=$\pi \cdot x\}=\{y \in C\;|\; \pi^{-1} \cdot g(y)$=$x\}=\{y\in C\;|\; g(\pi^{-1} \diamond y)$= $x\}\overset{\pi^{-1} \diamond y:= z}{=}\{\pi \diamond z \in C\;|\; g(z)$=$x\}=\pi \diamond \{ z \in C\;|\; g(z)$=$x\}=\pi \diamond g^{-1}(x)=\pi \diamond h(x)$. We obtained $h(\pi \cdot x)=\pi \diamond h(x)$ for all $\pi \in Fix(supp(f) \cup supp(g) \cup supp(T))$ and all $x \in B$. According to Proposition \[2.18’\], we get that $h$ is finitely supported. Furthermore, we also have that $supp(h) \subseteq supp(f) \cup supp(g) \cup supp(T) \overset{Claim\; 5}{\subseteq} supp(f) \cup supp(g) \cup supp(S)$$ \overset{Claim \; 4}{\subseteq} supp(f) \cup supp(g) \cup supp(F) \overset{Claim \; 2}{\subseteq} supp(f)$ $ \cup supp(g)$. **Claim 8:** $h$ is a bijective function.\ First we prove that $h$ is injective. Let us suppose that $h(x)=h(y)$. We claim that either $x,y \in T$ or $x,y \in B-T$. Indeed, let us suppose that $x\in T$ and $y \notin T$ (the case $x \notin T$, $y \in T$ is similar). We have $h(x)=f(x)$ and $h(y)=g^{-1}(y)$. If we denote $g^{-1}(y)=z$, we have $g(z)=y$. However, we supposed that $y \in B-T$, and so there exists $u \in C-f(T)$ such that $y=g(u)$. Since $y=g(z)$, from the injectivity of $g$ we get $u=z$. This is a contradiction because $u \notin f(T)$, while $z=f(x) \in f(T)$. Since we proved that both $x,y$ are contained either in $T$ or in $B-T$, the injectivity of $h$ follows from the injectivity of $f$ or $g$, respectively. Now we prove that $h$ is surjective. Let $y \in C$ be arbitrarily chosen. If $y \in f(T)$, then there exists $z \in T$ such that $y=f(z)$, and so $y=h(z)$. If $y \in C-f(T)$, and because $g(C-f(T))=B-T$, there exists $x \in B-T$ such that $g(y)=x$. Thus, $y \in g^{-1}(x)$. Since $g$ is injective, and so $g^{-1}(x)$ is a one-element set, we can say that $g^{-1}(x)=y$ with $x \in B-T$. Thus we have $y=h(x)$. \[lem2\] Let $(B, \cdot)$ and $(C, \diamond)$ be two invariant sets (in particular, $B$ and $C$ could coincide), $B_{1}$ a finitely supported subset of $B$ and $C_{1}$ a finitely supported subset of $C$. If there exist a finitely supported injective mapping $f: B_{1} \to C_{1}$ and a finitely supported injective mapping $g: C_{1} \to B_{1}$, then there exists a finitely supported bijective mapping $h:B_{1} \to C_{1}$. Furthermore, $supp(h) \subseteq supp(f) \cup supp(g) \cup supp(B_{1}) \cup supp(C_{1})$. *Proof of Lemma \[lem2\].* We follow the proof of Lemma \[lem1\]. We define $F:\wp_{fs'}(B_{1}) \to \wp_{fs'}(B_{1})$ by $F(X)=B_{1}-g(C_{1}-f(X))$ for all $X \in \wp_{fs'}(B_{1})$, where $\wp_{fs'}(B_{1})$ is a finitely supported subset of the invariant set $\wp_{fs}(B)$ (supported by $supp(B_{1})$) defined by $\wp_{fs'}(B_{1})=\{X \in \wp_{fs}(B)\;|\;X \subseteq B_{1}\}$. As in the previous lemma, but using Proposition \[2.18’\], we get that $F$ is well-defined, i.e. for every $X \in \wp_{fs'}(B_{1})$ we have that $F(X)$ is supported by $supp(f) \cup supp(g) \cup supp(B_{1}) \cup supp(C_{1}) \cup supp(X)$ which means $F(X) \in \wp_{fs'}(B_{1})$. Moreover, $F$ is itself finitely supported (in the sense of Definition \[2.10-1\]) by $supp(f) \cup supp(g) \cup supp(B_{1}) \cup supp(C_{1})$. The set $S:=\{X\;|\; X \in \wp_{fs'}(B_{1}), \ X \subseteq F(X)\}$ is contained in $\wp_{fs'}(B_{1})$ and it is supported by $supp(F)$ as a subset of $\wp_{fs}(B)$. The set $T:=\underset{X \in S}{\cup}X \in \wp_{fs'}(B_{1})$ is finitely supported by $supp(S)$, and it is a fixed point of $F$. As in the proof of Lemma \[lem1\], we define the bijection $h:B_{1} \to C_{1}$ by $$h(x)=\left\{ \begin{array}{ll} f(x), & \text{for}\: x \in T;\\ g^{-1}(x), & \text{for}\: x \in B_{1}-T.\end{array}\right.$$ According to Proposition \[2.18’\], we obtain that $h$ is finitely supported by $ supp(f) \cup supp(g) \cup supp(B_{1}) \cup supp(C_{1}) \cup supp(T)$, and $supp(h) \subseteq supp(f) \cup supp(g) \cup supp(B_{1}) \cup supp(C_{1})$. Thus, $h$ is the required finitely supported bijection between $B_{1}$ and $C_{1}$. The anti-symmetry of $\leq$ follows from Lemma \[lem1\] and Lemma \[lem2\] because FSM sets are actually finitely supported subsets of invariant sets. It is worth noting that $\leq^{*}$ is not anti-symmetric. \[lem3\] There are two invariant sets $B$ and $C$ such that there exist both a finitely supported surjective mapping $f: C \to B$ and a finitely supported surjective mapping $g: B \to C$, but it does not exist a finitely supported bijective mapping $h:B \to C$. *Proof of Lemma \[lem3\].* Let us consider the invariant set $(A, \cdot)$ of atoms. The family $T_{fin}(A)=\{(x_{1}, \ldots, x_{m}) \subseteq (A \times \ldots \times A)\,|\,m \geq 0\}$ of all finite injective tuples from $A$ (including the empty tuple denoted by $\bar{\emptyset}$) is an $S_{A}$-set with the $S_{A}$-action $\star:S_{A}\times T_{fin}(A) \rightarrow T_{fin}(A)$ defined by $\pi \star (x_{1}, \ldots, x_{m})=(\pi \cdot x_{1}, \ldots, \pi \cdot x_{m})$ for all $ (x_{1}, \ldots, x_{m}) \in T_{fin}(A)$ and all $\pi \in S_{A}$. Since $A$ is an invariant set, we have that $T_{fin}(A)$ is an invariant set. Whenever $X$ is an invariant set, we have that each injective tuple $(x_{1}, \ldots, x_{m})$ of elements belonging to $X$ is finitely supported, and, furthermore, $supp(x_{1}, \ldots, x_{m})=supp(x_{1}) \cup \ldots \cup supp(x_{m})$. Particularly, we obtain that $supp(a_{1}, \ldots, a_{m})=\{a_{1}, \ldots, a_{m}\}$, for any injective tuple of atoms $(a_{1}, \ldots, a_{m})$ (similarly as in Proposition 2.2 from [@book]). Since $supp(\bar{\emptyset})=\emptyset$, it follows that $T^{*}_{fin}(A)=T_{fin}(A)\setminus \bar{\emptyset}$ is an equivariant subset of $T_{fin}(A)$, and is itself an invariant set. Let us fix an atom $a \in A$. We define $f:T_{fin}(A) \to T_{fin}(A)\setminus \bar{\emptyset}$ by $$f(y)=\left\{ \begin{array}{ll} y, & \text{if}\: \text{$y$ is an injective non-empty tuple};\\ (a), & \text{if}\: \text{$y=\bar{\emptyset}$}\: .\end{array}\right.$$ Clearly, $f$ is surjective. We claim that $f$ is supported by $supp(a)$. Let $\pi \in Fix(supp(a))$, i.e. $a=\pi(a)=\pi \star(a)$. If $y$ is a non-empty tuple of atoms, we obviously have $f(\pi \star y)=\pi \star y=\pi \star f(y)$. If $y=\bar{\emptyset}$, we have $\pi \star y=\bar{\emptyset}$, and so $f(\pi \star y)=(a)=\pi(a)=\pi \star f(y)$. Thus, $f(\pi \star y)=\pi \star f(y)$ for all $y \in T_{fin}(A)$. According to Proposition \[2.18’\], we have that $f$ is finitely supported. We define an equivariant surjective function $g:T_{fin}(A)\setminus \bar{\emptyset} \to T_{fin}(A)$ by $$g(y)=\left\{ \begin{array}{ll} \bar{\emptyset}, & \text{if}\: \text{$y$ is a tuple with exactly one element};\\ y', & \text{otherwise}\: ; \end{array}\right.$$ where $y'$ is a new tuple formed by deleting the first element in tuple $y$ (the first position in a finite injective tuple exists without requiring any form of choice). Clearly, $g$ is surjective. Indeed, $\bar{\emptyset}=g((a))$ for some one-element tuple $(a)$ ($A$ is non-empty, and so it has at least one atom). For a fixed finite injective non-empty $m$-tuple $y$, we have that $y$ can be seen as being “contained" in an injective $(m+1)$-tuple $z$ of form $(b,y)$ (whose first element is a certain atom $b$, and the following elements are precisely the elements of $y$). The related atom $b$ exists because $y$ is finite, while $A$ is infinite (generally, we can always find an atom $b \notin supp(y)=\{y\}$ according to the finite support requirement in FSM - more details in Section 2.9 of [@book]). We get $y=g (z)$. For proving the surjectivity of $g$ we do not need to ‘choose’ a precise such an element $b$ (we do not need to define an inverse function for $g$); it is sufficient to ascertain that $g(b,y)=y$ for every $b \in A\setminus\{y\}$ and $A\setminus\{y\}$ is non-empty (the axiom of choice is not required because for proving only the surjectivity of $g$ we do not involve the construction of a system of representatives for the family $(g^{-1}(y))_{y \in T_{fin}(A)}$). We claim now that $g$ is equivariant. Let $(x)$ be a one-element tuple from $A$ and $\pi$ an arbitrary permutation from $S_{A}$. We have that $\pi \star (x)=(\pi(x))$ is a one-element tuple from $A$, and so $g(\pi \star (x))=\bar{\emptyset} = \pi \star \bar{\emptyset} = \pi \star g((x))$. Now, let us consider $(x_{1}, \ldots, x_{m}) \in T_{fin}(A), m \geq 2$ and $\pi \in S_{A}$. We have $ g(\pi \star (x_{1}, \ldots, x_{m}))=g((\pi \cdot x_{1}, \ldots, \pi \cdot x_{m}))=g((\pi (x_{1}), \ldots, \pi (x_{m})))=(\pi (x_{2}), \ldots, \pi (x_{m}))=\pi \star (x_{2}, \ldots, x_{m})=\pi \star g(x_{1}, \ldots, x_{m})$. According to Proposition \[2.18’\], we have that $g$ is empty-supported (equivariant). We prove by contradiction that there could not exist a finitely supported injective $h: T_{fin}(A) \to T_{fin}(A)\setminus \bar{\emptyset}$. Let us suppose there is a finitely supported injection $h:T_{fin}(A) \rightarrow T_{fin}(A)\setminus \bar{\emptyset}$. We have $\bar{\emptyset} \notin Im(h)$ because $Im(h) \subseteq T_{fin}(A) \setminus \bar{\emptyset}$. We can form an infinite sequence $\mathcal{F}$ which has the first term $y_{0}=\bar{\emptyset}$, and the general term $y_{n+1}=h(y_{n})$ for all $n \in \mathbb{N}$. Since $\bar{\emptyset}\notin Im(h)$, it follows that $\bar{\emptyset} \neq h(\bar{\emptyset})$. Since $h$ is injective and $\bar{\emptyset} \notin Im(h)$, we obtain by induction that $h^{n}(\bar{\emptyset}) \neq h^{m}(\bar{\emptyset})$ for all $n,m \in \mathbb{N}$ with $n \neq m$. We prove now that for each $n \in \mathbb{N}$ we have that $y_{n+1}$ is supported by $supp(h)\cup supp(y_{n})$. Let $\pi \in Fix(supp(h)\cup supp(y_{n}))$. According to Proposition \[2.18’\], because $\pi \in Fix(supp(h))$ we have $h(\pi \star y_{n})=\pi \star h(y_{n})$. Since $\pi \in Fix(supp(y_{n}))$ we have $\pi \star y_{n}=y_{n}$, and so $h(y_{n})=\pi \star h(y_{n})$. Thus, $\pi \star y_{n+1}= \pi \star h(y_{n}) = h(y_{n}) = y_{n+1}$. Furthermore, because $supp(y_{n+1})$ is the least set supporting $y_{n+1}$, we have $supp(y_{n+1}) \subseteq supp(h)\cup supp(y_{n})$ for all $n \in \mathbb{N}$. Since each $y_{n}$ is a finite injective tuple of atoms, it follows that $supp(y_{n})=\{y_{n}\}$ for all $n \in \mathbb{N}$ (where by $\{y_{n}\}$ we denoted the set of atoms forming $y_{n}$). We get $\{y_{n+1}\} = supp(y_{n+1}) \subseteq supp(h)\cup supp(y_{n}) = supp(h) \cup \{y_{n}\}$. By repeatedly applying this result, we get $\{y_{n}\} \subseteq supp(h) \cup \{y_{0}\}= supp(h) \cup \emptyset =supp(h)$ for all $n\in\mathbb{N}$. Since $supp(h)$ has only a finite number of subsets, we contradict the statement that the infinite sequence $(y_{n})_{n}$ never repeats. Thus, there does not exist a finitely supported bijection between $T_{fin}(A)\setminus \bar{\emptyset}$ and $T_{fin}(A)$. - $\leq$ and $\leq^{\star}$ are not total. We prove that whenever $X$ is an infinite ordinary (non-atomic) ZF-set, for any finitely supported function $f : A \to X$ and any finitely supported function $g : X \to A$, $Im(f)$ and $Im(g)$ are finite. As a direct consequence there are no finitely supported injective mappings and no finitely supported surjective mappings between $A$ and $X$. Let us consider a finitely supported mapping $f:A \to X$. Let let us fix an element $b\in A$ with $b\notin supp(f)$. Let $c$ be an arbitrary element from $A\setminus supp(f)$. Since $b\notin supp(f)$, we have that $(b\, c)$ fixes every element from $supp(f)$, i.e. $(b\, c)\in Fix(supp(f))$. However, $supp(f)$ supports $f$, and so, by Proposition \[2.18’\], we have $f((b\, c)(a))=(b\,c) \diamond f(a)=f(a)$ for all $a\in A$. In particular, $f(c)=f((b\,c)(b))=f(b)$. Since $c$ has been chosen arbitrarily from $A \setminus supp(f)$, it follows that $f(c)=f(b)$, for all $c \in A \setminus supp(f)$. If $supp(f)=\{a_{1}, \ldots, a_{n}\}$, then $Im(f) = \{f(a_{1})\}\cup \ldots \cup \{f(a_{n})\} \cup \{f(b)\}$. Thus, $Im(f)$ is finite (because it is a finite union of singletons). Let $g : X \to A$ be a finitely supported function. Assume by contradiction that $Im(g)$ is infinite. Pick any atom $a \in Im(g) \setminus supp(g)$ (such an atom exists because $supp(g)$ is finite). There exists an $x \in X$ such that $g(x) = a$. Now pick any atom $b \in Im(g) \setminus (supp(g) \cup \{a\})$, The transposition $(a\,b)$ fixes $supp(g)$ pointwise, and so $g(x)=g((a\,b) \diamond x)=(a\,b) \cdot g(x) =(a\, b)(a)= b$, contradicting the fact that $g$ is a function. Thus, $Im(g)$ is finite. \[corcor\] There exist two invariant sets $B$ and $C$ such that there is a finitely supported bijection between $\wp_{fs}(B)$ and $\wp_{fs}(C)$, but there is no finitely supported bijection between $B$ and $C$. Firstly we prove the following lemma. \[lemlem\] Let $X$ and $Y$ be two FSM sets and $f:X \to Y$ a finitely supported surjective function. Then the mapping $g:\wp_{fs}(Y) \to \wp_{fs}(X)$ defined by $g(V)=f^{-1}(V)$ for all $V \in \wp_{fs}(Y)$ is well defined, injective and finitely supported by $supp(f) \cup supp(X) \cup supp(Y)$. *Proof of Lemma \[lemlem\].* Let $V$ be an arbitrary element from $\wp_{fs}(Y)$. We claim that $f^{-1}(V) \in \wp_{fs}(X)$. Indeed we prove that the set $f^{-1}(V)$ is supported by $supp(f) \cup supp(V) \cup supp(X) \cup supp(Y)$. Let $\pi \in Fix(supp(f) \cup supp(V) \cup supp(X) \cup supp(Y))$, and $x \in f^{-1}(V) $. This means $f(x) \in V$. According to Proposition \[2.18’\], and because $\pi$ fixes $supp(f)$ pointwise and $supp(f)$ supports $f$, we have $f(\pi \cdot x)= \pi \cdot f(x) \in \pi \star V = V$, and so $\pi \cdot x \in f^{-1}(V)$ (we denoted the actions on $X$ and $Y$ generically by $\cdot$, and the actions on their powersets by $\star$). Therefore, $f^{-1}(V)$ is finitely supported, and so the function $g$ is well defined. We claim that $g$ is supported by $supp(f) \cup supp(X) \cup supp(Y)$. Let $\pi \in Fix(supp(f) \cup supp(X) \cup supp(Y))$. For any arbitrary $V \in \wp_{fs}(Y)$ we get $\pi \star V \in \wp_{fs}(Y)$ and $\pi \star g(V) \in \wp_{fs}(X)$, and by Proposition \[2.18’\] we have that $\pi^{-1} \in Fix(supp(f))$, and so $f(\pi^{-1} \cdot x)=\pi^{-1} \cdot f(x)$ for all $x\in X$. For any arbitrary $V \in \wp_{fs}(Y)$, we have that $z \in g(\pi \star V) = f^{-1}(\pi \star V) \Leftrightarrow f(z) \in \pi \star V \Leftrightarrow \pi^{-1} \cdot f(z) \in V \Leftrightarrow f(\pi^{-1} \cdot z) \in V \Leftrightarrow \pi^{-1} \cdot z \in f^{-1}(V) \Leftrightarrow z \in \pi \star f^{-1}(V)=\pi \star g(V)$. If follows that $g(\pi \star V)=\pi \star g(V)$ for all $V \in \wp_{fs}(Y)$, and so $g$ is finitely supported. Moreover, because $f$ is surjective, a simple calculation shows us that $g$ is injective. Indeed, let us suppose that $g(U)=g(V)$ for some $U,V \in \wp_{fs}(Y)$. We have $f^{-1}(U) = f^{-1}(V)$, and so $f(f^{-1}(U)) = f(f^{-1}(V))$. Since $f$ is surjective, we get $U = f(f^{-1}(U)) = f(f^{-1}(V)) = V$. We start the proof of Corollary \[corcor\]. As in Lemma \[lem3\], we consider the sets $B=T_{fin}(A)\setminus \bar{\emptyset}$ and $C=T_{fin}(A)$. According to Lemma \[lem3\] there exists a finitely supported surjective function $f:C \to B$ and a finitely supported (equivariant) surjection $g:B \to C$. Thus, according to Lemma \[lemlem\], there exist a finitely supported injective function $f':\wp_{fs}(B) \to \wp_{fs}(C)$ and a finitely supported injective function $g':\wp_{fs}(C) \to \wp_{fs}(B)$. According to Lemma \[lem1\], there is a finitely supported bijection between $\wp_{fs}(B)$ and $\wp_{fs}(C)$. However, we proved in Lemma \[lem3\] that there is no finitely supported bijection between $B=T_{fin}(A)\setminus \bar{\emptyset}$ and $C=T_{fin}(A)$. The following result communicated by Levy in 1965 for non-atomic ZF sets can be reformulated in the world of finitely supported atomic structures. Let $X$ and $Y$ be two invariant sets with the property that whenever $|2^{X}_{fs}|=|2^{Y}_{fs}|$ we have $|X|=|Y|$. If $|X|\leq^{\star}|Y|$ and $|Y|\leq^{\star}|X|$, then $|X|=|Y|$. According to the hypothesis and to Lemma \[lemlem\] there exist two finitely supported injective functions $f:\wp_{fs}(Y) \to \wp_{fs}(X)$ and $g:\wp_{fs}(X) \to \wp_{fs}(Y)$. According to Lemma \[lem1\], there is a bijective mapping $h:\wp_{fs}(X) \to \wp_{fs}(Y)$ . According to Theorem \[comp\], we get $|2^{X}_{fs}|=|2^{Y}_{fs}|$, and so we get $|X|=|Y|$. \[Cantor\] Let $X$ be a finitely supported subset of an invariant set $(Y, \cdot)$. Then $|X| \lneq |\wp_{fs}(X)|$ and $|X| \lneq^{*} |\wp_{fs}(X)|$ First we prove that there is no finitely supported bijection between $X$ and $\wp_{fs}(X)$, and so their cardinalities cannot be equal. Assume, by contradiction, that there is a finitely supported surjective mapping $f: X \to \wp_{fs}(X)$. Let us consider $Z=\{x \in X\,|\,x \notin f(x)\}$. We claim that $supp(X) \cup supp(f)$ supports $Z$. Let $\pi \in Fix(supp(X) \cup supp(f))$. Let $x \in Z$. Then $\pi \cdot x \in X$ and $\pi \cdot x \notin \pi \star f(x)=f(\pi \cdot x)$. Thus, $\pi \cdot x \in Z$, and so $Z \in \wp_{fs}(X)$. Therefore, since $f$ is surjective there is $x_{0} \in X$ such that $f(x_{0})=Z$. However, from the definition of $Z$ we have $x_{0} \in Z$ if and only if $x_{0}\notin f(x_{0})=Z$, which is a contradiction. Now, it is clear that the mapping $i: X \to \wp_{fs}(X)$ defined by $i(x)=\{x\}$ is injective and supported by $supp(X)$. Thus, $|X| \lneq |\wp_{fs}(X)|$. Let us fix an atom $y \in X$. We define $s:\wp_{fs}(X) \to X$ by $$s(U)=\left\{ \begin{array}{ll} u, & \text{if}\: \text{$U$ is an one-element set \{u\} };\\ y, & \text{if}\: \text{$U$ has more than one element}\: .\end{array}\right.$$ Clearly, $s$ is surjective. We claim that $s$ is supported by $supp(y) \cup supp(X)$. Let $\pi \in Fix(supp(y) \cup supp(X))$. Thus, $y=\pi \cdot y$. If $U$ is of form $U=\{u\}$, we obviously have $s(\pi \star U)=s(\{\pi \cdot u\})=\pi \cdot u=\pi \cdot s(U)$. If $U$ has more than one element, then $\pi \star U$ has more than one element, and we have $s(\pi \star U)=y=\pi \cdot y=\pi \cdot s(U)$. Thus, $\pi \star U \in \wp_{fs}(X)$, $\pi \cdot s(U) \in X$, and $s(\pi \star U)=\pi \cdot s(U)$ for all $U \in \wp_{fs}(X)$ . According to Proposition \[2.18’\], we have that $s$ is finitely supported. Therefore, $|X| \lneq^{*} |\wp_{fs}(X)|$. In Proposition \[Cantor\] we used a technique for constructing a surjection starting from an injection defined in the opposite way, that can be generalized as follows. \[pco2\]Let $X$ and $Y$ be finitely supported subsets of an invariant set $U$. If If $|X| \leq |Y|$, then $|X| \leq ^{\star} |Y|$. The converse is not valid. However, if $|X| \leq ^{\star} |Y|$, then $|X| \leq |\wp_{fs}(Y)|$. Suppose there exists a finitely supported injective mapping $f: X \to Y$. We consider the case $Y \neq \emptyset$ (otherwise, the result follows trivially). Fix $x_{0} \in X$. Define the mapping $f':Y \to X$ by $$f'(y)=\left\{ \begin{array}{ll} f^{-1}(y), & \text{if}\: \text{$y \in Im(f)$ };\\ x_{0}, & \text{if}\: \text{$y \notin Im(f)$}\: .\end{array}\right.$$ Since $f$ is injective, it follows that $f^{-1}(y)$ is an one-element set for each $y \in Im(f)$, and so $f'$ is a function. Clearly, $f'$ is surjective. We claim that $f'$ is supported by the set $supp(f) \cup supp(x_{0}) \cup supp(X) \cup supp(Y)$. Indeed, let us consider $\pi \in Fix(supp(f) \cup supp(x_{0}) \cup supp(X) \cup supp(Y))$. Whenever $y \in Im(f)$ we have $y=f(z)$ for some $z \in X$ and $\pi \cdot y=\pi \cdot f(z)=f(\pi \cdot z) \in Im(f)$, which means $Im(f)$ is finitely supported by $supp(f)$. Consider an arbitrary $y_{0} \in Im(f)$, and thus $\pi \cdot y_{0} \in Im(f)$. Then $f'(y_{0})= f^{-1}(y_{0})=z_{0}$ with $f(z_{0})=y_{0}$, and so $f(\pi \cdot z_{0})= \pi \cdot f(z_{0})=\pi \cdot y_{0}$, which means $f'(\pi \cdot y_{0})=f^{-1}(\pi \cdot y_{0})=\pi \cdot z_{0}=\pi \cdot f^{-1}(y_{0})=\pi \cdot f'(y_{0})$. Now, for $y \notin Im(f)$ we have $\pi \cdot y \notin Im(f)$, which means $f'(\pi \cdot y)=x_{0}=\pi \cdot x_{0}=\pi \cdot f(y)$ since $\pi$ fixes $x_{0}$ pointwise. Thus, $|X| \leq ^{\star} |Y|$. Conversely, from the proof of Lemma \[lem3\], we know that there is a finitely supported surjection $g:T_{fin}(A)\setminus \bar{\emptyset} \to T_{fin}(A)$, but there does not exist a finitely supported injection $h: T_{fin}(A) \to T_{fin}(A)\setminus \bar{\emptyset}$. Assume now there is a finitely supported surjective mapping $f:Y \to X$. We proceed similarly as in the proof of Lemma \[lemlem\]. Fix $x \in X$. Then $f^{-1}(\{x\})$ is supported by $supp(f) \cup supp(x) \cup supp(X)$. Indeed, let $\pi \in Fix(supp(f) \cup supp(x) \cup supp(X))$, and $y \in f^{-1}(\{x\})$. This means $f(y)=x$. According to Proposition \[2.18’\], we have $f(\pi \cdot y)= \pi \cdot f(y)= \pi \cdot x=x$, and so $\pi \cdot y \in f^{-1}(\{x\})$. Define $g:X \to \wp_{fs}(Y)$ by $g(x)=f^{-1}(\{x\})$. We claim that $g$ is supported by $supp(f) \cup supp(X)$. Let $\pi \in Fix(supp(f) \cup supp(X))$. For any arbitrary $x \in X$, we have that $z \in g(\pi \cdot x) = f^{-1}(\{\pi \cdot x\}) \Leftrightarrow f(z)=\pi \cdot x \Leftrightarrow \pi^{-1} \cdot f(z)=x \Leftrightarrow f(\pi^{-1} \cdot z) =x \Leftrightarrow \pi^{-1} \cdot z \in f^{-1}(\{x\}) \Leftrightarrow z \in \pi \star f^{-1}(\{x\})=\pi \star g(x)$. From Proposition \[2.18’\] it follows that $g$ is finitely supported. Since $g$ is also injective, we get $|X| \leq |\wp_{fs}(Y)|$. \[pco1\] Let $X,Y,Z$ be finitely supported subsets of an invariant set $U$. The following properties hold. 1. If $|X| \leq |Y|$, then $|X|+|Z| \leq |Y|+|Z|$; 2. If $|X| \leq |Y|$, then $|X| \cdot |Z| \leq |Y| \cdot |Z|$; 3. If $|X| \leq |Y|$, then $|X^{Z}_{fs}| \leq |Y^{Z}_{fs}|$; 4. If $|X| \leq |Y|$ and $Z\neq \emptyset$, then $|Z^{X}_{fs}| \leq |Z^{Y}_{fs}|$; 5. $|X|+|Y| \leq |X|\cdot|Y|$ whenever both $X$ and $Y$ have more than two elements. 1\. Suppose there is a finitely supported injective $f: X \to Y$, and define the injection $g: X+Z\to Y+Z$ by $$g(u)=\left\{ \begin{array}{ll} (0,f(x)), & \text{if}\: u=(0,x)\: \text{with}\: x \in X;\\ (1,z), & \text{if}\: u=(1,z)\: \text{with}\: z \in Z .\end{array}\right.$$ Since $f$ is finitely supported we have that $f(\pi \cdot x)=\pi \cdot f(x)$ for all $x \in X$ and $\pi \in Fix(supp(f))$. By using Proposition \[2.18’\], i.e verifying that $g(\pi \star u)=\pi \star g(u)$ for all $u \in X+Z$ and all $\pi \in Fix(supp(f) \cup supp(X) \cup supp(Y) \cup supp(Z))$, we have that $g$ is also finitely supported. 2\. Suppose there exists a finitely supported injective mapping $f: X \to Y$. Define the injection $g: X\times Z\to Y \times Z$ by $g((x,z))=(f(x),z)$ for all $(x,z) \in X \times Z$. Clearly $g$ is injective. Since $f$ is finitely supported we have that $f(\pi \cdot x)=\pi \cdot f(x)$ for all $x \in X$ and $\pi \in Fix(supp(f))$, and so $g(\pi \otimes(x,z))=g((\pi \cdot x, \pi \cdot z))=(f(\pi \cdot x),\pi \cdot z)=(\pi \cdot f(x),\pi \cdot z)=\pi \otimes g((x,z))$ for all $(x,z) \in X \times Z$ and $\pi \in Fix(supp(f) \cup supp(X) \cup supp(Y) \cup supp(Z))$, which means $g$ is supported by $supp(f) \cup supp(X) \cup supp(Y) \cup supp(Z)$. 3\. Suppose there exists a finitely supported injective mapping $f: X \to Y$. Define $g:X^{Z}_{fs} \to Y^{Z}_{fs}$ by $g(h)=f \circ h$. We have that $g$ is injective and for any $\pi \in Fix(supp(f))$ we have $\pi \widetilde{\star} f=f$, and so $g(\pi \widetilde{\star} h)=f \circ (\pi \widetilde{\star} h)= (\pi \widetilde{\star} f) \circ (\pi \widetilde{\star} h)=\pi \widetilde{\star} (f \circ h)=\pi \widetilde{\star} g(h)$ for all $h \in X^{Z}_{fs}$. We used the relation $(\pi \widetilde{\star} f) \circ (\pi \widetilde{\star} h)=\pi \widetilde{\star} (f \circ h)$ for all $\pi \in S_{A}$. This can be proved as follows. Fix $x\in Z$, we have $(\pi\widetilde{\star}(f\circ h))(x)=\pi\cdot(f(h(\pi^{-1}\cdot x)))$. Also, if we denote $(\pi\widetilde{\star} h)(x)=y$ we have $y=\pi\cdot(h(\pi^{-1}\cdot x))$ and $((\pi\widetilde{\star} f)\circ(\pi\widetilde{\star} h))(x)=(\pi\widetilde{\star} f)(y)=\pi\cdot(f(\pi^{-1}\cdot y))=\pi\cdot(f((\pi^{-1}\circ\pi)\cdot h(\pi^{-1}\cdot x)))=\pi\cdot(f(h(\pi^{-1}\cdot x)))$. We finally obtain that $g$ is supported by $supp(f) \cup supp(X) \cup supp(Y) \cup supp(Z)$. 4\. Suppose there exists a finitely supported injective mapping $f: X \to Y$. According to Proposition \[pco2\], there is a finitely supported surjective mapping $f':Y \to X$. Define the injective mapping $g:Z^{X}_{fs} \to Z^{Y}_{fs}$ by $g(h)=h \circ f'$. As in item 3 one can prove that $g$ is finitely supported by $supp(f') \cup supp(X) \cup supp(Y) \cup supp(Z)$. 5\. Fix $x_{0}, x_{1} \in X$ with $x_{0} \neq x_{1}$ and $y_{0}, y_{1} \in Y$ with $y_{0}\neq y_{1}$. Define the injection $g: X+Y\to X \times Y$ by $$g(u)=\left\{ \begin{array}{ll} (x, y_{0}), & \text{if}\: u=(0,x)\: \text{with}\: x \in X, x \neq x_{0};\\ (x_{0},y), & \text{if}\: u=(1,y)\: \text{with}\: y \in Y;\\ (x_{1}, y_{1}), & \text{if}\: u=(0,x_{0})\end{array}\right.$$ It follows that $g$ is supported by $supp(x_{0}) \cup supp(y_{0}) \cup supp(x_{1}) \cup supp(y_{1}) \cup supp(X) \cup supp(Y)$, and $g$ is injective. \[comp\] Let $(X, \cdot)$ be a finitely supported subset of an invariant set $(Z, \cdot)$. There exists a one-to-one mapping from $\wp_{fs}(X)$ onto $\{0,1\}^{X}_{fs}$ which is finitely supported by $supp(X)$, where $\wp_{fs}(X)$ is considered the family of those finitely supported subsets of $Z$ contained in $X$. Let $Y$ be a finitely supported subset of $Z$ contained in $X$, and $\varphi_{Y}$ be the characteristic function on $Y$, i.e. $\varphi_{Y}:X \to \{0,1\}$ is defined by $\varphi_{Y}(x)\overset{def}{=}\left\{ \begin{array}{ll} 1 & \text{for}\: x \in Y\\ 0 & \text{for}\: x\in X \setminus Y \end{array}\right.$.\ We prove that $\varphi_{Y}$ is a finitely supported function from $X$ to $\{0,1\}$ (according to Proposition \[p1\], $\{0,1\}$ is a trivial invariant set), and the mapping $Y \mapsto \varphi_{Y}$ defined on $\wp_{fs}(X)$ is also finitely supported in the sense of Definition \[2.10-1\]. First we prove that $\varphi_{Y}$ is supported by $supp(Y) \cup supp(X)$. Let $\pi \in Fix(supp(Y) \cup supp(X))$. Thus $\pi \star Y=Y$ (where $\star$ represents the canonical permutation action on $\wp(Z)$), and so $\pi \cdot x \in Y$ if and only if $x \in Y$. Since we additionally have $\pi \star X=X$, we obtain $\pi \cdot x \in X \setminus Y$ if and only if $x \in X \setminus Y$. Thus, $\varphi_{Y}(\pi\cdot x)=\varphi_{Y}(x)$ for all $x \in X$. Furthermore, because $\pi$ fixes $supp(X)$ pointwise we have $\pi \cdot x \in X$ for all $x \in X$, and from Proposition \[2.18’\] we get that $\varphi_{Y}$ is supported by $supp(Y) \cup supp(X)$. We remark that $\{0,1\}^{X}_{fs}$ is a finitely supported subset of the set $(\wp_{fs}(Z \times \{0,1\}), \widetilde{\star})$. Let $\pi \in Fix(supp(X))$ and $f:X \to \{0,1\}$ finitely supported. We have $\pi\widetilde{\star} f=\{(\pi\cdot x,$ $\pi\diamond y)\,|\,(x,y)\in f\}=\{(\pi\cdot x,$ $y)\,|\,(x,y)\in f\}$ because $\diamond$ is the trivial action on $\{0,1\}$. Thus, $\pi\widetilde{\star}f $ is a function with the domain $\pi\star X=X$ which is finitely supported as an element of $(\wp (Z \times \{0,1\}), \widetilde{\star})$ according to Proposition \[2.15\]. Moreover, $(\pi \widetilde{\star} f)(\pi\cdot x)= f(x)$ for all $x \in X$ (1). According to Proposition \[2.18’\], to prove that the function $g:=Y \mapsto \varphi_{Y}$ defined on $\wp_{fs}(X)$ (with the codomain contained in $\{0,1\}^{X}_{fs}$) is supported by $supp(X)$, we have to prove that $\pi \widetilde{\star}g(Y)=g(\pi \star Y)$ for all $\pi \in Fix(supp(X))$ and all $Y \in \wp_{fs}(X)$ (where $\widetilde{\star}$ symbolizes the induced $S_{A}$-action on $\{0,1\}^{X}_{fs}$). This means that we need to verify the relation $\pi \widetilde{\star} \varphi_{Y} = \varphi_{\pi \star Y}$ for all $\pi \in Fix(supp(X))$ and all $Y \in \wp_{fs}(X)$. Let us consider $\pi \in Fix(supp(X))$ (which means $\pi \cdot x \in X$ for all $x \in X$) and $Y \in \wp_{fs}(X)$. For any $x \in X$, we know that $x \in \pi \star Y$ if and only if $\pi^{-1} \cdot x \in Y$. Thus, $\varphi_{Y}(\pi^{-1} \cdot x )=\varphi_{\pi \star Y}(x)$ for all $x \in X$, and so $(\pi \widetilde{\star} \varphi_{Y})(x) \overset{(1)}{=}\varphi_{Y}(\pi^{-1} \cdot x )=\varphi_{\pi \star Y}(x)$ for all $x\in X$. Moreover, from Proposition \[2.15\], $\pi \star Y$ is a finitely supported subset of $Z$ contained in $\pi \star X=X$, and $\{0,1\}^{X}_{fs}$ can be represented as a finitely supported subset of $\wp_{fs}(Z \times \{0,1\})$ (supported by $supp(X)$). According to Proposition \[2.18’\] we have that $g$ is a finitely supported function from $\wp_{fs}(X)$ to $\{0,1\}^{X}_{fs}$. Obviously, $g$ is one-to-one. Now we prove that $g$ is onto. Let us consider an arbitrary finitely supported function $f: X \to \{0,1\}$. Let $Y_{f}\overset{def}{=}\{x \in X\;|\; f(x)=1\}$. We claim that $Y_{f} \in \wp_{fs}(X)$. Let $\pi \in Fix(supp(f))$. According to Proposition \[2.18’\] we have $\pi \cdot x \in X$ and $f(\pi \cdot x)=f(x)$ for all $x \in X$. Thus, for each $x \in Y_{f}$, we have $\pi \cdot x \in Y_{f}$. Therefore $\pi \star Y_{f}=Y_{f}$, and so $Y_{f}$ is finitely supported by $supp(f)$ as a subset of $Z$, and it is contained in $X$. A simple calculation show us that $g(Y_{f})=f$, and so $g$ is onto. One can easy verify that the properties of $\leq$ presented in Proposition \[pco1\] (1), (2) and (4) also hold for $\leq^{\star}$. We left the details to the reader. \[cardord1\] There exists an invariant set $X$ (particularly the set $A$ of atoms) having the following properties. 1. $|X \times X| \nleq^{*} |\wp_{fs}(X)|$; 2. $|X \times X| \nleq |\wp_{fs}(X)|$; 3. $|X \times X| \nleq^{*} |X|$; 4. $|X \times X| \nleq |X|$; 5. For each $n \in \mathbb{N}, n \geq 2$ we have $|X| \lneq |\wp_{n}(X)| \lneq |\wp_{fs}(X)|$, where $\wp_{n}(X)$ is the family of all $n$-sized subsets of $X$; 6. For each $n \in \mathbb{N}$ we have $|X| \lneq^{*} |\wp_{n}(X)| \lneq^{*} |\wp_{fs}(X)|$; 7. $|X| \lneq |\wp_{fin}(X)| \lneq |\wp_{fs}(X)|$; 8. $|X| \lneq^{*} |\wp_{fin}(X)| \lneq^{*} |\wp_{fs}(X)|$; 9. $|\wp_{fs}(X) \times \wp_{fs}(X)| \nleq^{*} |\wp_{fs}(X)|$; 10. $|\wp_{fs}(X) \times \wp_{fs}(X)| \nleq |\wp_{fs}(X)|$; 11. $|X+X| \lneq^{*} |X\times X|$; 12. $|X+X| \lneq |X\times X|$. 1\. We prove that that there does not exist a finitely supported surjective mapping $f: \wp_{fs}(A) \to A \times A$. Suppose, by contradiction, that there is a finitely supported surjective mapping $f: \wp_{fs}(A) \to A \times A$. Let us consider two atoms $a,b\notin supp(f)$ with $a \neq b$. These atoms exist because $A$ is infinite, while $supp(f) \subseteq A$ is finite. It follows that the transposition $(a\, b)$ fixes each element from $supp(f)$, i.e. $(a\, b) \in Fix(supp(f))$. Since $f$ is surjective, it follows that there exists an element $X \in \wp_{fs}(A)$ such that $f(X)=(a,b)$. Since $supp(f)$ supports $f$ and $(a\, b) \in Fix(supp(f))$, from Proposition \[2.18’\] we have $f((a\,b) \star X)=(a\,b) \otimes f(X)=(a\,b) \otimes (a,b)=((a\,b)(a), (a\,b)(b))=(b,a)$. Due to the functionality of $f$ we should have $(a\,b) \star X \neq X$. Otherwise, we would obtain $(a,b)=(b,a)$. We claim that if both $a,b \in supp(X)$, then $(a\,b)\star X=X$. Indeed, suppose $a,b \in supp(X)$. Since $X$ is a finitely supported subset of $A$, then $X$ is either finite or cofinite. If $X$ is finite, then $supp(X)=X$, and so $a,b \in X$. Moreover, $(a\, b)(a)=b$, $(a\, b)(b)=a$, and $(a\, b)(c)=c$ for all $c \in X$ with $c\neq a,b$. Therefore, $(a\,b) \star X=\{(a\,b)(x)\,|\,x \in X\}=\{(a\,b)(a)\} \cup \{(a\,b)(b)\} \cup \{(a\,b)(c)\,|\,c \in X \setminus\{a,b\}\}=\{b\} \cup \{a\} \cup (X \setminus \{a,b\})=X$. Now, if $X$ is cofinite, then $supp(X)=A \setminus X$, and so $a,b \in A \setminus X$. Since $a,b \notin X$, we have $a,b \neq x$ for all $x \in X$, and so $(a\,b)(x)=x$ for all $x \in X$. Thus, in this case we also have $(a\,b) \star X=X$. Since when both $a,b \in supp(X)$ we have $(a\,b)\star X=X$, it follows that one of $a$ or $b$ does not belong to $supp(X)$. Suppose $b \notin supp(X)$ (the other case is analogue). Let us consider $c\neq a,b$, $c \notin supp(f)$, $c \notin supp(X)$. Then $(b\, c) \in Fix(supp(X))$, and, because $supp(X)$ supports $X$, we have $(b\,c)\star X=X$. Furthermore, $(b\, c) \in Fix(supp(f))$, and by Proposition \[2.18’\] we have $(a,b)=f(X)=f((b\,c) \star X)=(b\,c) \otimes f(X)=(b\,c) \otimes (a,b)=((b\,c)(a), (b\,c)(b))=(a,c)$ which is a contradiction because $b\neq c$. Thus, $|A \times A| \nleq^{*} |\wp_{fs}(A)|$. 2\. We prove that there does not exist a finitely supported injective mapping $f: A \times A \to \wp_{fs}(A)$. Suppose, by contradiction, that there is a finitely supported injective mapping $f: A \times A \to \wp_{fs}(A)$. According to Proposition \[pco2\], one can define a finitely supported surjection $g: \wp_{fs}(A) \to A \times A$. This contradicts the above item. Thus, $|A \times A| \nleq |\wp_{fs}(A)|$. 3\. We prove that there does not exist a finitely supported surjection $f: A \to A \times A$. Since there exists a surjection $s$ from $\wp_{fs}(A)$ onto $A$ defined by $$s(X)=\left\{ \begin{array}{ll} x, & \text{if}\: \text{$X$ is an one-element set $\{x\}$ };\\ a, & \text{if}\: \text{$X$ is not an one-element set,}\: \end{array}\right.$$ where $a$ is a fixed atom, and $s$ is finitely supported (by$\{a\}$), the result follows from item 1. Thus, $|A \times A| \nleq^{*} |A|$. 4\. We prove that there does not exist a finitely supported injection $f: A \times A \to A$. Since there exists an equivariant injection from $A$ into $\wp_{fs}(A)$ defined as $x \mapsto \{x\}$, the result follows from item 2. Thus, $|A \times A| \nleq |A|$; Alternatively, one can prove that there does not exist a one-to-one mapping from $A \times A$ to $A$ (and so neither a finitely supported one). Suppose, by contradiction, that there is a an injective mapping $i: A \times A \to A$. Let us fix two atoms $x$ and $y$ with $x \neq y$. The sets $\{i(a,x)\,|\,a \in A\}$ and $\{i(a,y)\,|\,a \in A\}$ are disjoint and infinite. Thus, $\{i(a,x)\,|\,a \in A\}$ is a infinite and coinfinite subset of $A$, which contradicts the fact that any subset of $A$ is either finite or cofinite. 5\. We prove that $|A| \lneq |\wp_{n}(A)| \lneq |\wp_{fs}(A)|$ for all $n \in \mathbb{N}, n \geq 2$. Consider $a_{1},a_{2}, \ldots, a_{n-1}$ $, a^{1}_{1}, \ldots, a^{n}_{1}, \ldots, a^{1}_{n-1}, \ldots, a^{n}_{n-1} \in A$ a family of pairwise different elements. Then $i: A \to \wp_{n}(A)$ defined by $$i(x)=\left\{ \begin{array}{ll} \{x,a_{1},a_{2}, \ldots, a_{n-1}\}, & \text{if}\: \text{$x \neq a_{1}, \ldots, a_{n-1}$ };\\ \{ a^{1}_{1}, \ldots, a^{n}_{1}\}, & \text{if}\: \text{$x=a_{1}$ }\\ \vdots \\ \{ a^{1}_{n-1}, \ldots, a^{n}_{n-1}\}, & \text{if}\: \text{$x=a_{n-1}$ } \: \end{array}\right.$$ is obviously an injective mapping from $(A, \cdot)$ to $(\wp_{n}(A), \star)$. Furthermore, we can easy check that $i$ is supported by the finite set {$a_{1},a_{2}, \ldots, a_{n-1} $ $, a^{1}_{1}, \ldots, a^{n}_{1}, \ldots, a^{1}_{n-1}, \ldots, a^{n}_{n-1}\}$, and so $|A| \leq |\wp_{n}(A)|$ in FSM. We claim that there does not exist a finitely supported injection from $\wp_{n}(A)$ into $A$. Assume on the contrary that there exists an finitely supported injection $f:\wp_{n}(A) \to A$. First, we claim that, for any $Y \in \wp_{n}(A)$ which is disjoint from $supp(f)$, we have $f(Y) \notin Y$. Assume by contradiction that $f(Y) \in Y$ for a fixed $Y$ with $Y \cap supp(f)=\emptyset$. Let $\pi$ be a permutation of atoms which fixes $supp(f)$ pointwise, and interchanges all the elements of $Y$ (e.g. $\pi$ is a cyclic permutation of $Y$). Since $\pi$ permutes all the elements of $Y$, we have $\pi \cdot f(Y)=\pi(f(Y)) \neq f(Y)$. However, $\pi \star Y =\{\pi(a_{1}), \ldots, \pi(a_{n})\}=\{a_{1}, \ldots, a_{n}\}=Y$. Since $\pi$ fixes $supp(f)$ pointwise and $supp(f)$ supports $f$, we have $\pi (f(Y))=\pi \cdot f(Y)= f(\pi \star Y)=f(Y)$, a contradiction. Since $supp(f)$ is finite, there are infinitely many such $Y$ with the property that $Y \cap supp(f)=\emptyset$. Thus, because it is injective, $f$ takes infinitely many values on those $Y$. Since $supp(f)$ is finite, there should exist at least one element in $\wp_{n}(A)$, denoted by $Z$ such that $Z \cap supp(f)=\emptyset$ and $f(Z) \notin supp(f)$. Thus, $f(Z) = a$ for some $a \in A \setminus (Z \cup supp(f))$. Let $b \in A \setminus (supp(f) \cup Z \cup \{a\})$ and also let $\pi = (a\, b)$. Then $\pi \in Fix(supp(f) \cup Z)$, and hence $f(Z) =f((a\,b) \star Z)=(a\,b)(f(Z))= b$, a contradiction. We obtained that $|A| \neq |\wp_{n}(A)|$ in FSM, and so $|A|<|\wp_{n}(A)|$. We obviously have $|\wp_{n}(A)| \leq |\wp_{fs}(A)|$. We prove below that there does not exist a finitely supported injective mapping from $\wp_{fs}(A)$ onto one of its finitely supported proper subsets, i.e. any finitely supported injection $f:\wp_{fs}(A) \rightarrow \wp_{fs}(A)$ is also surjective. Let us consider a finitely supported injection $f:\wp_{fs}(A) \rightarrow \wp_{fs}(A)$. Suppose, by contradiction, $Im(f) \subsetneq \wp_{fs}(A)$. This means that there exists $X_{0}\in\wp_{fs}(A)$ such that $X_{0}\notin Im(f)$. Since $f$ is injective, we can define an infinite sequence $\mathcal{F}=(X_{n})_{n}$ starting from $X_{0}$, with distinct terms of form $X_{n+1}=f(X_{n})$ for all $n \in \mathbb{N}$. Furthermore, according to Proposition \[2.18’\], for a fixed $k \in \mathbb{N}$ and $\pi \in Fix(supp(f) \cup supp(X_{k}))$, we have $\pi \star X_{k+1}=\pi \star f(X_{k})=f(\pi \star X_{k})=f(X_{k})=X_{k+1}$. Then, $supp(X_{n+1}) \subseteq supp(f)\cup supp(X_{n})$ for all $n \in \mathbb{N}$, and by induction on $n$ we have that $supp(X_{n}) \subseteq supp(f)\cup supp(X_{0})$ for all $n \in \mathbb{N}$. We obtained that each element $X_{n}\in \mathcal{F}$ is supported by the same finite set $S:=supp(f)\cup supp(X_{0})$. However, there could exist only finitely many subsets of $A$ (i.e. only finitely many elements in $\wp_{fs}(A)$) supported by $S$, namely the subsets of $S$ and the supersets of $A\setminus S$ (where a superset of $A \setminus S$ is of form $A\setminus X$ with $X \subseteq S$). We contradict the statement that the infinite sequence $(X_{n})_{n}$ never repeats. Thus, $f$ is surjective, and so there could not exist a bijection between $\wp_{fin}(A)$ and $\wp_{fs}(A)$, which means $|\wp_{n}(A)| \neq |\wp_{fs}(A)|$. 6\. Fix $n \in \mathbb{N}$. As in the above item there does not exist neither a finitely supported bijection between $\wp_{n}(A)$ and $\wp_{fs}(A)$, nor a finitely supported bijection between $A$ and $\wp_{n}(A)$. However, there exists a finitely supported injection $i: A \to \wp_{n}(A)$. Fix an atom $a \in A$. The mapping $s: \wp_{n}(A) \to A $ defined by $$s(X)=\left\{ \begin{array}{ll} i^{-1}(X), & \text{if}\: \text{$X\in Im(i)$ };\\ a, & \text{if}\: \text{$X \notin Im(i)$}\: \end{array}\right.$$ is supported by $supp(i) \cup \{a\}$ and is surjective. Now, fix $n$ atoms $x_{1}, \ldots, x_{n}$. The mapping $g:\wp_{fs}(A) \to \wp_{n}(A)$ defined by $$g(X)=\left\{ \begin{array}{ll} X, & \text{if}\: \text{$X\in \wp_{n}(A)$ };\\ \{x_{1}, \ldots x_{n}\}, & \text{if}\: \text{$X \notin \wp_{n}(A)$}\: \end{array}\right.$$ is supported by $\{x_{1}, \ldots x_{n}\}$ and is surjective. 7\. We prove that $|A| \lneq |\wp_{fin}(A)| \lneq |\wp_{fs}(A)|$. We obviously have that $|A| \leq |\wp_{fin}(A)|$ by taking the equivariant injective mapping $f:A \to \wp_{fin}(A)$ defined by $f(a)=\{a\}$ for all $a \in A$. We prove, by contradiction, that there is no finitely supported surjection from $A$ onto $\wp_{fin}(A)$. Assume that $g:A \to \wp_{fin}(A)$ is a finitely supported surjection. Let us fix two atoms $x$ and $y$. We define the function $h: \wp_{fin}(A) \to \wp_{2}(A)$ by $ h(X)=\left\{ \begin{array}{ll} X, & \text{if}\: \text{$|X|=2$ };\\ \{x,y\}, & \text{if}\: \text{$|X| \neq 2$}\: .\end{array}\right. $. Since for every $\pi \in S_{A}$ and $X \in \wp_{fin}(A)$ we have $|\pi \star X|=|X|$, we conclude that $h$ is finitely supported by $\{x,y\}$. Thus, $h \circ g$ is a surjection from $A$ onto $\wp_{2}(A)$ supported by $supp(g)\cup\{x,y\}$, which contradicts the previous item. Therefore, $|A|<|\wp_{fin}(A)|$. Since every element in $\wp_{fin}(A)$ belongs to $\wp_{fs}(A)$, but there does not exist a finitely supported injective mapping from $\wp_{fs}(A)$ onto one of its finitely supported proper subsets, we also have $|\wp_{fin}(A)|<|\wp_{fs}(A)|$. 8\. As in the above item there does not exist neither a finitely supported bijection between $\wp_{fin}(A)$ and $\wp_{fs}(A)$, nor a finitely supported bijection between $A$ and $\wp_{fin}(A)$. Fix an atom $a \in A$. The mapping $s: \wp_{fin}(A) \to A$ defined by $$s(X)=\left\{ \begin{array}{ll} x, & \text{if}\: \text{$X$ is an one-element set $\{x\}$ };\\ a, & \text{if}\: \text{$X$ is not an one-element set}\: \end{array}\right.$$ is supported by $\{a\}$ and is surjective. Now, fix an atom $b$. The mapping $g:\wp_{fs}(A) \to \wp_{fin}(A)$ defined by $$g(X)=\left\{ \begin{array}{ll} X, & \text{if}\: \text{$X\in \wp_{fin}(A)$ };\\ \{b\}, & \text{if}\: \text{$X \notin \wp_{fin}(A)$}\: \end{array}\right.$$ is supported by $\{b\}$ and is surjective. 9\. According to Theorem \[cardord1\](1) there is no finitely supported surjection from $\wp_{fs}(A)$ onto $A \times A$. Suppose there is a finitely supported surjective mapping $f: \wp_{fs}(A) \to \wp_{fs}(A) \times \wp_{fs}(A)$. Obviously, there exists a supported surjection $s:\wp_{fs}(A) \to A$ defined by $$s(X)=\left\{ \begin{array}{ll} a, & \text{if}\: \text{$X$ is an one-element set \{a\} };\\ x, & \text{if}\: \text{$X$ has more than one element}\: .\end{array}\right.$$ where $x$ is a fixed atoms of $A$. The surjection $s$ is supported by $supp(x)=x$. Thus, we can define a surjection $g:\wp_{fs}(A) \times \wp_{fs}(A) \to A \times A$ by $g(X,Y)=(s(X),s(Y))$ for all $X,Y \in \wp_{fs}(A)$. Let $\pi \in Fix(supp(s))$. Since $supp(s)$ supports $s$, by Proposition \[2.18’\] we have $g(\pi \otimes_{\star} (X,Y))=g(\pi \star X,\pi \star Y)=(s(\pi \star X),s(\pi \star Y))=(\pi \cdot s(X),\pi \cdot s(Y))=\pi \otimes (s(X),s(Y))$ for all $X,Y \in \wp_{fs}(A)$, where $\otimes_{\star}$ and $\otimes$ represent the $S_{A}$-actions on $\wp_{fs}(A) \times \wp_{fs}(A)$ and $A \times A$, respectively. Thus, $supp(s)$ supports $g$, and so $supp(g) \subseteq supp(s)$. Furthermore, the function $h=g \circ f: \wp_{fs}(A) \to A \times A$ is surjective and finitely supported by $supp(s) \cup supp(f)$. This is a contradiction, and so $|\wp_{fs}(A) \times \wp_{fs}(A)| \nleq^{*} |\wp_{fs}(A)|$. 10\. Suppose, by contradiction, that there is a finitely supported injective mapping $f: \wp_{fs}(A) \times \wp_{fs}(A) \to \wp_{fs}(A)$. In the view of Proposition \[pco2\], let us fix two finitely supported subsets of $A$, namely $U$ and $V$. We define the function $g: \wp_{fs}(A) \to \wp_{fs}(A) \times \wp_{fs}(A)$ by $$g(X)=\left\{ \begin{array}{ll} f^{-1}(X), & \text{if}\: \text{$X\in Im(f)$ };\\ (U,V), & \text{if}\: \text{$X \notin Im(f)$}\: .\end{array}\right.$$ Clearly, $g$ is surjective. Furthermore, $g$ is supported by $supp(f) \cup supp(U) \cup supp(V)$ (the proof uses the fact that $Im(f)$ is a subset of $\wp_{fs}(A)$ supported by $supp(f)$). This contradicts the above item, and so $|\wp_{fs}(A) \times \wp_{fs}(A)| \nleq |\wp_{fs}(A)|$. 11\. In the view of Proposition \[pco1\](5) there is a finitely supported injection from $A+A$ into $A \times A$, and a finitely supported surjection from $A \times A$ onto $A+A$ according to Proposition \[pco2\]. Thus $|A+A|\leq |A \times A|$ and $|A+A|\leq^{*}|A \times A|$ Fix three different atoms $a,b,c \in A$. Define the mapping $f:A+A \to \wp_{fs}(A)$ by $$f(u)=\left\{ \begin{array}{ll} \{x\}, & \text{if}\: u=(0,x)\: \text{with}\: x \in A;\\ \{a,y\}, & \text{if}\: u=(1,y)\: \text{with}\: y \in A, y \neq a;\\ \{b,c\}, & \text{if}\: u=(1,a)\end{array}\right.$$ One can directly prove that $f$ is injective and supported by $\{a,b,c\}$. According to Proposition \[pco2\], we have $|A+A|\leq^{*}|\wp_{fs}(A)|$. If we had $|A \times A|=|A+A|$, we would obtain $|A\times A| \leq^{*} |\wp_{fs}(A)|$ which contradicts item 1. 12\. According to the above item $|A+A|\leq |\wp_{fs}(A)|$. If we had $|A \times A|=|A+A|$, we would obtain $|A\times A| \leq |\wp_{fs}(A)|$ which contradicts item 2. \[cardord2’\] There exists an invariant set $X$ having the following properties: 1. $|X| \lneq |X|+|X|$; 2. $|X| \lneq^{*} |X|+|X|$. 1\. First we prove that in FSM we have $|\wp_{fs}(A)|=2|\wp_{fin}(A)|$. Let us consider the function $f:\wp_{fin}(A) \to \wp_{cofin}(A)$ defined by $f(U)=A \setminus U$ for all $U \in \wp_{fin}(A)$. Clearly, $f$ is bijective. We claim that $f$ is equivariant. Indeed, let $\pi \in S_{A}$. To prove that $f(\pi \star U)$=$\pi \star f(U)$ for all $U \in \wp_{fin}(A)$, we have to prove that $A \setminus (\pi \star U)=\pi \star (A \setminus U)$ for all $U \in \wp_{fin}(A)$. Let $y \in A \setminus (\pi \star U)$. We can express $y$ as $y=\pi \cdot (\pi^{-1} \cdot y)$. If $\pi^{-1} \cdot y \in U$, then $y \in \pi \star U$, which is a contradiction. Thus, $\pi^{-1} \cdot y \in (A \setminus U)$, and so $y \in \pi \star (A \setminus U)$. Conversely, if $y \in \pi \star (A \setminus U)$, then $y=\pi\cdot x$ with $x \in A \setminus U$. Suppose $y \in \pi \star U$. Then $y=\pi\cdot z$ with $z \in U$. Thus, $x=z$ which is a contradiction, and so $y \in A \setminus (\pi \star U)$. Since $f$ is equivariant and bijective, it follows that $|\wp_{fin}(A)|=|\wp_{cofin}(A)|$. However, every finitely supported subset of $A$ is either finite or cofinite, and so $\wp_{fs}(A)$ is the union of the disjoint subsets $\wp_{fin}(A)$ and $\wp_{cofin}(A)$. Thus, $|\wp_{fs}(A)|=2|\wp_{fin}(A)|$. Moreover, there exists an equivariant injection $i: \wp_{fin}(A) \to \wp_{fs}(A)$ defined by $i(U)=U$ for all $U \in \wp_{fin}(A)$. However, there does not exist a finitely supported one-to-one mapping from $\wp_{fs}(A)$ onto one of its finitely supported proper subsets. Thus, there could not exist a bijection $f: \wp_{fs}(A) \to \wp_{fin}(A)$. Therefore, $|\wp_{fin}(A)| \neq |\wp_{fs}(A)|=2|\wp_{fin}(A)|$. We can consider $X=\wp_{fin}(A)$ or $X=\wp_{cofin}(A)$. 2\. It remains to prove that there is a finitely supported surjection from $\wp_{fs}(A)$ onto $\wp_{fin}(A)$. We either use Proposition \[pco2\] or effectively construct the surjection as below. Fix $a \in A$. We define $g:\wp_{fs}(A) \to \wp_{fin}(A)$ by $$g(U)=\left\{ \begin{array}{ll} U, & \text{if}\: \text{$U\in \wp_{fin}(A)$ };\\ \{a\}, & \text{if}\: \text{$U \notin \wp_{fin}(A)$}\: .\end{array}\right.$$ Clearly, $g$ is supported by $\{a\}$ and surjective. We can consider $X=\wp_{fin}(A)$ or $X=\wp_{cofin}(A)$. Forms of Infinite in Finitely Supported Structures {#chap9} ================================================== The equivalence of various definitions for infinity is provable in ZF under the consideration of the axiom of choice. Since in FSM the axiom of choice fails, our goal is to study various FSM forms of infinite and to provide several relations between them. Let $X$ be a finitely supported subset of an invariant set. 1. $X$ is called *FSM usual infinite* if $X$ does not correspond one-to-one and onto to a finite ordinal. We simply call *infinite* an FSM usual infinite set. 2. $X$ is *FSM covering infinite* if there is a finitely supported directed family $\mathcal{F}$ of finitely supported sets with the property that $X$ is contained in the union of the members of $\mathcal{F}$, but there does not exist $Z\in \mathcal{F}$ such that $X\subseteq Z$. 3. $X$ is called *FSM Tarski I infinite* if there exists a finitely supported one-to-one mapping of $X$ onto $X \times X$. 4. $X$ is called *FSM Tarski II infinite* if there exists a finitely supported family of finitely supported subsets of $X$, totally ordered by inclusion, having no maximal element. 5. $X$ is called *FSM Tarski III infinite* if $|X|=2|X|$. 6. $X$ is called *FSM Mostowski infinite* if there exists an infinite finitely supported totally ordered subset of $X$. 7. $X$ is called *FSM Dedekind infinite* if there exist a finitely supported one-to-one mapping of $X$ onto a finitely supported proper subset of $X$. 8. $X$ is *FSM ascending infinite* if there is a finitely supported increasing countable chain of finitely supported sets $X_{0}\subseteq X_{1}\subseteq\ldots\subseteq X_{n}\subseteq\ldots$ with $X\subseteq\cup X_{n}$, but there does not exist $n\in\mathbb{N}$ such that $X\subseteq X_{n}$; Note that in the definition of FSM Tarski II infinity for a certain $X$, the existence of a finitely supported family of finitely supported subsets of $X$ is required, while in the definition of FSM ascending infinity for $X$, the related family of finitely supported subsets of $X$ has to be FSM countable (i.e. the mapping $n \mapsto X_{n}$ should be finitely supported). It is immediate that if $X$ is FSM ascending infinite, then it is also FSM Tarski II infinite. Let $X$ be a finitely supported subset of an invariant set. Then $X$ is FSM usual infinite if and only if $X$ is FSM covering infinite. Let us suppose that $X$ is FSM usual infinite. Let $\mathcal{F}$ be the family of all FSM usual non-infinite (FSM usual finite) subsets of $X$ ordered by inclusion. Since $X$ is finitely supported, it follows that $\mathcal{F}$ is supported by $supp(X)$. Moreover, since all the elements of $\mathcal{F}$ are finite sets, it follows that all the elements of $\mathcal{F}$ are finitely supported. Clearly, $\mathcal{F}$ is directed and $X$ is the union of the members of $\mathcal{F}$. Suppose by contradiction, that $X$ is not FSM covering infinite. Then there exists $Z\in\mathcal{F}$ such that $X\subseteq Z$. Therefore, $X$ should by FSM usual finite which is a contradiction with our original assumption. Conversely, assume that $X$ is FSM covering infinite. Suppose, by contradiction that $X$ is FSM usual finite, i.e. $X=\{x_{1}, \ldots x_{n}\}$. Let $\mathcal{F}$ be a directed family such that $X$ is contained in the union of the members of $\mathcal{F}$ (at least one such a family exists, for example $\wp_{fs}(X)$). Then for each $i \in \{1, \ldots, n\}$ there exists $F_{i} \in \mathcal{F}$ such that $x_{i} \in F_{i}$. Since $\mathcal{F}$ is directed, there is $Z \in \mathcal {F}$ such that $F_{i} \subseteq Z$ for all $i \in \{1, \ldots, n\}$, and so $X \subseteq Z$ with $Z\in \mathcal{F}$, which is a contradiction. \[ti1\] The following properties of FSM Dedekind infinite sets hold. 1. Let $X$ be a finitely supported subset of an invariant set $Y$. Then $X$ is FSM Dedekind infinite if and only if there exists a finitely supported one-to-one mapping $f: \mathbb{N} \to X$. As a consequence, an FSM superset of an FSM Dedekind infinite set is FSM Dedekind infinite, and an FSM subset of an FSM set that is not Dedekind infinite is also not FSM Dedekind infinite. 2. Let $X$ be an infinite finitely supported subset of an invariant set $Y$. Then the sets $\wp_{fs}(\wp_{fin}(X))$ and $\wp_{fs}(T_{fin}(X))$ are FSM Dedekind infinite. 3. Let $X$ be an infinite finitely supported subset of an invariant set $Y$. Then the set $\wp_{fs}(\wp_{fs}(X))$ is FSM Dedekind infinite. 4. Let $X$ be a finitely supported subset of an invariant set $Y$ such that $X$ does not contain an infinite subset $Z$ with the property that all the elements of $Z$ are supported by the same set of atoms. Then $X$ is not FSM Dedekind infinite. 5. Let $X$ be a finitely supported subset of an invariant set $Y$ such that $X$ does not contain an infinite subset $Z$ with the property that all the elements of $Z$ are supported by the same set of atoms. Then $\wp_{fin}(X)$ is not FSM Dedekind infinite. 6. Let $X$ and $Y$ be two finitely supported subsets of an invariant set $Z$. If neither $X$ nor $Y$ is FSM Dedekind infinite, then $X \times Y$ is not FSM Dedekind infinite. 7. Let $X$ and $Y$ be two finitely supported subsets of an invariant set $Z$. If neither $X$ nor $Y$ is FSM Dedekind infinite, then $X + Y$ is not FSM Dedekind infinite. 8. Let $X$ be a finitely supported subset of an invariant set $Y$. Then $\wp_{fs}(X)$ is FSM Dedekind infinite if and only if $X$ is FSM ascending infinite. 9. Let $X$ be a finitely supported subset of an invariant set $Y$. If $X$ is FSM Dedekind infinite, then $X$ is FSM ascending infinite. The reverse implication is not valid. <!-- --> 1. Let us suppose that $(X, \cdot)$ is FSM Dedekind infinite, and $g: X \rightarrow X$ is an injection supported by the finite set $S \subsetneq A$ with the property that $Im(g) \subsetneq X$. This means that there exists $supp(g) \subseteq S$ and there exists $x_{0}\in X$ such that $x_{0}\notin Im(g)$. We can form a sequence of elements from $X$ which has the first term $x_{0}$ and the general term $x_{n+1}=g(x_{n})$ for all $n \in \mathbb{N}$. Since $x_{0}\notin Im(g)$ it follows that $x_{0} \neq g(x_{0})$. Since $g$ is injective and $x_{0} \notin Im(g)$, by induction we obtain that $g^{n}(x_{0}) \neq g^{m}(x_{0})$ for all $n,m \in \mathbb{N}$ with $n \neq m$. Furthermore, $x_{n+1}$ is supported by $supp(g)\cup supp(x_{n})$ for all $n \in \mathbb{N}$. Indeed, let $\pi \in Fix(supp(g)\cup supp(x_{n}))$. According to Proposition \[2.18’\], $\pi \cdot x_{n+1}= \pi \cdot g(x_{n})=g(\pi \cdot x_{n})=g(x_{n})=x_{n+1}$. Since $supp(x_{n+1})$ is the least set supporting $x_{n+1}$, we obtain $supp(x_{n+1}) \subseteq supp(g)\cup supp(x_{n})$ for all $n \in \mathbb{N}$. By finite recursion, we have $supp(x_{n}) \subseteq supp(g)\cup supp(x_{0})$ for all $n\in\mathbb{N}$. Since all $x_{n}$ are supported by the same set of atoms $supp(g)\cup supp(x_{0})$, we have that the function $f:\mathbb{N} \to X$, defined by $f(n)=x_{n}$, is also finitely supported (by the set $supp(g)\cup supp(x_{0}) \cup supp(X)$ not depending on $n$). Indeed, for any $\pi \in Fix(supp(g)\cup supp(x_{0}) \cup supp(X))$ we have $f(\pi \diamond n)=f(n)=x_{n}=\pi \cdot x_{n}=\pi\cdot f(n)$, $\forall n \in \mathbb{N}$, where by $\diamond$ we denoted the trivial $S_{A}$-action on $\mathbb{N}$. Furthermore, because $\pi$ fixes $supp(X)$ pointwise we have $\pi \cdot f(n) \in X$ for all $n \in \mathbb{N}$. From Proposition \[2.18’\] we have that $f$ is finitely supported. Obviously, $f$ is also injective. Conversely, suppose there exists a finitely supported injective mapping $f: \mathbb{N} \to X$. According to Proposition \[2.18’\], it follows that for any $\pi \in Fix(supp(f))$ we have $\pi \cdot f(n)=f(\pi \diamond n)=f(n)$ and $\pi \cdot f(n) \in X$ for all $n \in \mathbb{N}$. Let us define $g:X \to X$ by $$g(x)=\left\{ \begin{array}{ll} f(n+1), & \text{if}\: \text{there exists $n \in \mathbb{N}$ with $x=f(n)$};\\ x, & \text{if}\: \text{$x \notin Im(f)$}\: .\end{array}\right.$$ We claim that $g$ is supported by $supp(f) \cup supp(X)$. Indeed, let us consider $\pi \in Fix(supp(f) \cup supp(X))$ and $x \in X$. If there is some $n$ such that $x=f(n)$, we have that $\pi \cdot x=\pi \cdot f(n)=f(n)$, and so $g(\pi \cdot x)=g(f(n))=f(n+1)=\pi \cdot f(n+1)=\pi \cdot g(x)$. If $x \notin Im(f)$, we prove by contradiction that $\pi \cdot x \notin Im(f)$. Indeed, suppose that $\pi \cdot x \in Im(f)$. Then there is $y \in \mathbb{N}$ such that $\pi \cdot x=f(y)$ or, equivalently, $x = \pi^{-1} \cdot f(y)$. However, since $\pi \in Fix(supp(f))$, from Proposition \[2.18’\] we have $\pi^{-1} \cdot f(y)=f(\pi^{-1} \diamond y)$, and so we get $x=f(\pi^{-1} \diamond y)=f(y) \in Im(f)$ which contradicts the assumption that $x \notin Im(f)$. Thus, $\pi \cdot x \notin Im(f)$, and so $g(\pi \cdot x)=\pi \cdot x=\pi \cdot g(x)$. We obtained that $g(\pi \cdot x)=\pi \cdot x=\pi \cdot g(x)$ for all $x \in X$ and all $\pi \in Fix(supp(f) \cup supp(X))$. Furthermore, $\pi \cdot g(x) \in \pi \star X=X$ (where by $\star$ we denoted the $S_{A}$-action on $\wp_{fs}(Y)$), and so $g$ is finitely supported. Since $f$ is injective, it follows immediately that $g$ is injective. Furthermore, $Im(g)=X \setminus \{f(0)\}$ which is a proper subset of $X$, finitely supported by $supp(f(0)) \cup supp(X)=supp(f) \cup supp(X)$. 2. The family $\wp_{fin}(X)$ represents the family of those finite subsets of $X$ (these subsets of $X$ are finitely supported as subsets of the invariant set $Y$ in the sense of Definition \[2.14\]). Obviously, $\wp_{fin}(X)$ is a finitely supported subset of the invariant set $\wp_{fs}(Y)$, supported by $supp(X)$. This is because whenever $Z$ is an element of $\wp_{fin}(X)$ (i.e. whenever $Z$ is a finite subset of $X$) and $\pi$ fixes $supp(X)$ pointwise, we have that $\pi \star Z$ is also a finite subset of $X$. The family $\wp_{fs}(\wp_{fin}(X))$ represents the family of those subsets of $\wp_{fin}(X)$ which are finitely supported as subsets of the invariant set $\wp_{fs}(Y)$ in the sense of Definition \[2.14\]. As above, according to Proposition \[2.15\], we have that $\wp_{fs}(\wp_{fin}(X))$ is a finitely supported subset of the invariant set $\wp_{fs}(\wp_{fs}(Y))$, supported by $supp(\wp_{fin}(X)) \subseteq supp(X)$. Let $X_{i}$ be the set of all $i$-sized subsets from $X$, i.e. $X_{i}=\{Z \subseteq X\,|\,|Z|=i\}$. Since $X$ is infinite, it follows that each $X_{i}, i \geq 1$ is non-empty. Obviously, we have that any $i$-sized subset $\{x_{1}, \ldots, x_{i}\}$ of $X$ is finitely supported (as a subset of $Y$) by $supp(x_{1}) \cup \ldots \cup supp(x_{i})$. Therefore, $X_{i} \subseteq \wp_{fin}(X)$ and $X_{i} \subseteq \wp_{fs}(Y)$ for all $i \in \mathbb{N}$. Since $\cdot$ is a group action, the image of an $i$-sized subset of $X$ under an arbitrary permutation is an $i$-sized subset of $Y$. However, any permutation of atoms that fixes $supp(X)$ pointwise also leaves $X$ invariant, and so for any permutation $\pi \in Fix(supp(X))$ we have that $\pi \star Z$ is an $i$-sized subset of $X$ whenever $Z$ is an $i$-sized subset of $X$. Thus, each $X_{i}$ is a subset of $\wp_{fin}(X)$ finitely supported by $supp(X)$, and so $X_{i} \in \wp_{fs}(\wp_{fin}(X))$. We define $f: \mathbb{N} \to \wp_{fs}(\wp_{fin}(X))$ by $f(n)=X_{n}$. We claim that $supp(X)$ supports $f$. Indeed, let $\pi \in Fix(supp(X))$. Since $supp(X)$ supports $X_{n}$ for all $n \in \mathbb{N}$, we have $\pi \star f(n)=\pi \star X_{n}=X_{n}=f(n)= f(\pi \diamond n)$ (where $\diamond$ is the trivial $S_{A}$-action on $\mathbb{N}$) and $\pi \star f(n)=\pi \star X_{n}=X_{n} \in \wp_{fs}(\wp_{fin}(X))$ for all $n \in \mathbb{N}$. According to Proposition \[2.18’\], we have that $f$ is finitely supported. Furthermore, $f$ is injective and, by item 1, we have that $\wp_{fs}(\wp_{fin}(X))$ is FSM Dedekind infinite. If we consider $Y_{i}$ the set of all $i$-sized injective tuples formed by elements of $X$, we have that each $Y_{i}$ is a subset of $T_{fin}(X)$ supported by $supp(X)$, and the family $(Y_{i})_{i \in \mathbb{N}}$ is a countably infinite, uniformly supported, subset of $\wp_{fs}(T_{fin}(X))$. From item 1 we get that $\wp_{fs}(T_{fin}(X))$ is FSM Dedekind infinite. 3. The proof is actually the same as in the above item because every $X_{i} \in \wp_{fs}(\wp_{fs}(A))$. 4. If there does not exist a uniformly supported subset of $X$, then there does not exist a finitely supported injective mapping $f:\mathbb{N} \to X$, and so $f$ cannot be FSM Dedekind infinite. 5. We prove the following lemma: \[lem4\] Let $X$ be a finitely supported subset of an invariant set $Y$ such that $X$ does not contain an infinite uniformly supported subset. Then the set $\wp_{fin}(X)=\{Z\!\subseteq\! X\,|\, Z\, \text{finite}\}$ does not contain an infinite uniformly supported subset. *Proof of Lemma \[lem4\].* Suppose, by contradiction, that the set $\wp_{fin}(X)$ contains an infinite subset $\mathcal{F}$ such that all the elements of $\mathcal{F}$ are different and supported by the same finite set $S$. Therefore, we can express $\mathcal{F}$ as $\mathcal{F}=(X_{i})_{i \in I} \subseteq \wp_{fin}(X)$ with the properties that $X_{i} \neq X_{j}$ whenever $i \neq j$ and $supp(X_{i}) \subseteq S$ for all $i \in I$. Fix an arbitrary $j \in I$. However, from Proposition \[4.4-9\], because $supp(X_{j})=\underset{x \in X_{j}}{\cup}supp(x)$, we have that $X_{j}$ has the property that $supp(x) $ $\subseteq S$ for all $x \in X_{j}$. Since $j$ has been arbitrarily chosen from $I$, it follows that every element from every set of form $X_{i}$ is supported by $S$, and so $\underset{i}{\cup}X_{i}$ is an uniformly supported subset of $X$ (all its elements being supported by $S$). Furthermore, $\underset{i \in I}{\cup}X_{i}$ is infinite because the family $(X_{i})_{i \in I}$ is infinite and $X_{i} \neq X_{j}$ whenever $i \neq j$. Otherwise, if $\underset{i}{\cup}X_{i}$ was finite, the family $(X_{i})_{i \in I}$ would be contained in the finite set $\wp (\underset{i}{\cup}X_{i})$, and so it couldn’t be infinite with the property that $X_{i} \neq X_{j}$ whenever $i \neq j$. We were able to construct an infinite uniformly supported subset of $X$, namely $\underset{i}{\cup}X_{i}$, and this contradicts the hypothesis that $X$ does not contain an infinite uniformly supported subset. *Proof of this item* According to the above lemma, if $X$ does not contain an infinite uniformly supported subset, then $\wp_{fin}(X)$ does not contain an infinite uniformly supported subset. Suppose, by contradiction, that $\wp_{fin}(X)$ is FSM Dedekind infinite. According to item 1, there exists a finitely supported injective mapping $f: \mathbb{N} \to \wp_{fin}(X)$. Thus, because $\mathbb{N}$ is a trivial invariant set, according to Proposition \[2.18’\], there exists an infinite injective (countable) sequence $f(\mathbb{N})=(X_{i})_{i \in \mathbb{N}} \subseteq \wp_{fin}(X)$ having the property $supp(X_{i}) \subseteq supp(f)$ for all $i \in \mathbb{N}$. We obtained that $\wp_{fin}(X)$ contains an infinite uniformly supported subset $(X_{i})_{i \in \mathbb{N}}$, which is a contradiction. 6. Suppose, by contradiction, that $X \times Y$ is FSM Dedekind infinite. According to item 1, there exists a finitely supported injective mapping $f: \mathbb{N} \to X \times Y$ Thus, according to Proposition \[2.18’\], there exists an infinite injective sequence $f(\mathbb{N})=((x_{i},y_{i}))_{i \in \mathbb{N}} \subseteq X \times Y$ with the property that $supp((x_{i}, y_{i})) \subseteq supp(f)$ for all $i \in \mathbb{N}$ (1). Fix some $j \in \mathbb{N}$. We claim that $supp((x_{j}, y_{j})) =supp(x_{j}) \cup supp(y_{j})$. Let $U=(x_{j}, y_{j})$, and $S=supp(x_{j})\cup supp(y_{j})$. Obviously, $S$ supports $U$. Indeed, let us consider $\pi\in Fix(S)$. We have that $\pi\in Fix(supp(x_{j}))$ and also $\pi\in Fix(supp(y_{j}))$ Therefore, $\pi\cdot x_{j}=x_{j}$ and $\pi\cdot y_{j}=y_{j}$, and so $\pi \otimes (x_{j}, y_{j})=(\pi \cdot x_{j}, \pi \cdot y_{j})=(x_{j}, y_{j})$, where $\otimes$ represent the $S_{A}$ action on $X \times Y$ described in Proposition \[p1\]. Thus, $supp(U) \subseteq S$. It remains to prove that $S \subseteq supp(U)$. Fix $\pi \in Fix(supp(U))$. Since $supp(U)$ supports $U$, we have $\pi \otimes (x_{j}, y_{j})=(x_{j}, y_{j})$, and so $(\pi \cdot x_{j}, \pi \cdot y_{j})=(x_{j}, y_{j})$, from which we get $\pi \cdot x_{j}=x_{j}$ and $\pi \cdot y_{j}= y_{j}$. Thus, $supp(x_{j}) \subseteq supp(U)$ and $supp(y_{j}) \subseteq supp(U)$. Hence $S=supp(x_{j})\cup supp(y_{j}) \subseteq supp(U)$. According to relation (1) we obtain, $supp(x_{i})\cup supp(y_{i}) \subseteq supp(f)$ for all $i \in \mathbb{N}$. Thus, $supp(x_{i}) \subseteq supp(f)$ for all $i \in \mathbb{N}$ and $supp(y_{i}) \subseteq supp(f)$ for all $i \in \mathbb{N}$ (2). Since the sequence $((x_{i},y_{i}))_{i \in \mathbb{N}}$ is infinite and injective, then at least one of the sequences $(x_{i})_{i \in \mathbb{N}}$ and $(y_{i})_{i \in \mathbb{N}}$ is infinite. Assume that $(x_{i})_{i \in \mathbb{N}}$ is infinite. Then there exists an infinite subset $B$ of $\mathbb{N}$ such that $(x_{i})_{i \in B}$ is injective, and so there exists an injection $u: B \to X$ defined by $u(i)=x_{i}$ for all $i \in B$ which is supported by $supp(f)$ (according to relation (2) and Proposition \[2.18’\]). However, since $B$ is an infinite subset of $\mathbb{N}$, there exists a ZF bijection $h: \mathbb{N} \to B$. The construction of $h$ requires only the fact that $\mathbb{N}$ is well-ordered which is obtained from the Peano construction of $\mathbb{N}$ and does not involve a form of the axiom of choice. Since both $B$ and $\mathbb{N}$ are trivial invariant sets, it follows that $h$ is equivariant. Thus, $u \circ h$ is an injection from $\mathbb{N}$ to $X$ which is finitely supported by $supp(u) \subseteq supp(f)$. This contradicts the assumption that $X$ is not FSM Dedekind infinite. \[rrr\] Analogously, using the relation $supp(x) \cup supp(y)=supp((x,y))$ for all $x \in X$ and $y \in Y$ derived from Proposition \[4.4-9\], it can be proved that $X \times Y$ does not contain an infinite uniformly supported subset if neither $X$ nor $Y$ contain an infinite uniformly supported subset. 7. Suppose, by contradiction, that $X + Y$ is FSM Dedekind infinite. According to item 1, there exists a finitely supported injective mapping $f: \mathbb{N} \to X + Y$. Thus, there exists an infinite injective sequence $(z_{i})_{i \in \mathbb{N}} \subseteq X+ Y$ such that $supp(z_{i}) \subseteq supp(f)$ for all $i \in \mathbb{N}$. According to the construction of the disjoint union of two $S_{A}$-sets (see Proposition \[p1\]), as in the proof of item 6, there should exist an infinite subsequence of $(z_{i})_{i}$ of form $((0, x_{j}))_{x_{j} \in X}$ which is uniformly supported by $supp(f)$, or an infinite sequence of form $((1, y_{k}))_{y_{k} \in Y}$ which is uniformly supported by $supp(f)$. Since $0$ and $1$ are constants, this means there should exist at least an infinite uniformly supported sequence of elements from $X$, or an infinite uniformly supported sequence of elements from $Y$. This contradicts the hypothesis neither $X$ nor $Y$ is FSM Dedekind infinite. Analogously, it can be proved that $X + Y$ does not contain an infinite uniformly supported subset if neither $X$ nor $Y$ contain an infinite uniformly supported subset. 8. Assume, by contradiction, that $(X_{n})_{n \in \mathbb{N}}$ is an infinite countable family of different subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported. Thus, each $X_{n}$ is supported by the same set $S=supp(n \mapsto X_{n})$. We define a countable family $(Y_{n})_{n \in \mathbb{N}}$ of subsets of $X$ that are non-empty and pairwise disjoint. A ZF construction of such a family belongs to Kuratowski and can also be found in Lemma 4.11 from [@herrlich]. This approach works also in FSM in the view of the $S$-finite support principle because every $Y_{k}$ is defined only involving elements in the family $(X_{n})_{n \in \mathbb{N}}$, and so whenever $(X_{n})_{n \in \mathbb{N}}$ is uniformly supported (meaning that all $X_{n}$ are supported by the same set of atoms), we get that $(Y_{n})_{n \in \mathbb{N}}$ is uniformly supported. Formally the sequence $(Y_{n})_{n \in \mathbb{N}}$ is recursively constructed as below. For $n \in \mathbb{N}$, assume that $Y_{m}$ is defined for any $m<n$ such that the set $\{ X_{k} \setminus \underset{m<n}\cup Y_{m}\,|\,k \geq n\}$ is infinite. Define $n'=min\{k\,|\,k \geq n \:\text{and}\: X_{k} \setminus \underset{m<n}\cup Y_{m} \neq \emptyset \:\text{and}\: (X \setminus X_{k}) \setminus \underset{m<n}\cup Y_{m} \neq \emptyset\}$. We define $$Y_{n}=\left\{ \begin{array}{ll} X_{n'} \setminus \underset{m<n}\cup Y_{m}, & \text{if}\: \{X_{k}\setminus (X_{n'} \cup \underset{m<n}\cup Y_{m}) \,|\, k>n'\}\:\text{is infinite};\\ (X \setminus X_{n'})\setminus \underset{m<n}\cup Y_{m}, & \text{otherwise} .\end{array}\right.$$ Obviously, $Y_{1}$ is supported by $S \cup supp(X)$. By induction, assume that $Y_{m}$ is supported by $S \cup supp(X)$ for each $m<n$. Since $Y_{n}$ is defined as a set combination of $X_{i}$’s (which are all $S$-supported) and $Y_{m}$’s with $m<n$, we get that $Y_{n}$ is supported by $S \cup supp(X)$ according to the $S$-finite support principle. Therefore the family $(Y_{i})_{i \in \mathbb{N}}$ is uniformly supported by $S \cup supp(X)$. Let $U_{i}=Y_{0} \cup \ldots \cup Y_{i}$ for all $i \in \mathbb{N}$. Clearly all $U_{i}$ are supported by $S \cup supp(X)$, and $U_{0} \subsetneq U_{1} \subsetneq U_{2} \subsetneq \ldots \subsetneq X$. Let $V_{n}=(X\setminus\underset{i\in\mathbb{N}}{\cup}U_{i})\cup U_{n}$. Clearly, $X=\underset{n\in \mathbb{N}}{\cup} V_{n}$. Moreover, $V_{n}$ is supported by $S \cup supp(X)$ for all $n \in \mathbb{N}$. Therefore, the mapping $n\mapsto V_{n}$ is finitely supported. Obviously, $V_{0} \subsetneq V_{1} \subsetneq V_{2} \subsetneq \ldots \subsetneq X$. However, there does not exist $n\in\mathbb{N}$ such that $X=V_{n}$, and so $X$ is FSM ascending infinite. The converse holds since if $X$ is FSM ascending infinite, there is a finitely supported increasing countable chain of finitely supported sets $X_{0}\subseteq X_{1}\subseteq\ldots\subseteq X_{n}\subseteq\ldots$ with $X\subseteq\cup X_{n}$, but there does not exist $n\in\mathbb{N}$ such that $X\subseteq X_{n}$. In this sequence there should exist infinitely many different elements of form $X_{i}$ (otherwise their union will be a term of the sequence), and the result follows from Proposition \[cou\]. 9. Suppose $X$ is FSM Dedekind infinite. Therefore, $\wp_{fs}(X)$ is FSM Dedekind infinite. According to item 8, we have that $X$ is FSM ascending infinite. The reverse implication is not valid because, as it is proved in Proposition \[tari\], $\wp_{fin}(A)$ is FSM ascending infinite, but not FSM Dedekind infinite. \[ti2\] The following sets and all of their FSM usual infinite subsets are FSM usual infinite, but they are not FSM Dedekind infinite. 1. The invariant set $A$ of atoms. 2. The powerset $\wp_{fs}(A)$ of the set of atoms. 3. The set $T_{fin}(A)$ of all finite injective tuples of atoms. 4. The invariant set $A^{A}_{fs}$ of all finitely supported functions from $A$ to $A$. 5. The invariant set of all finitely supported functions $f:A \to A^{n}$, where $n \in \mathbb{N}$. 6. The invariant set of all finitely supported functions $f:A \to T_{fin}(A)$. 7. The invariant set of all finitely supported functions $f:A \to \wp_{fs}(A)$. 8. The sets $\wp_{fin}(A)$, $\wp_{cofin}(A)$, $\wp_{fin}(\wp_{fs}(A))$, $\wp_{fin}(\wp_{cofin}(A))$, $\wp_{fin}(\wp_{fin}(A))$, $\wp_{fin}(A^{A}_{fs})$. 9. Any construction of finite powersets of form $\wp_{fin}(\ldots \wp_{fin}(A))$, $\wp_{fin}(\ldots \wp_{fin}(P(A)))$, or $\wp_{fin}(\ldots \wp_{fin}(\wp_{fs}(A)))$. 10. Every finite Cartesian combination between the set $A$, $\wp_{fin}(A)$, $\wp_{cofin}(A)$, $\wp_{fs}(A)$ and $A^{A}_{fs}$ . 11. The disjoint unions $A+A^{A}_{fs}$, $A+\wp_{fs}(A)$, $\wp_{fs}(A)+A^{A}_{fs}$ and $A+\wp_{fs}(A)+A^{A}_{fs}$ and all finite disjoint unions between $A$, $A^{A}_{fs}$ and $\wp_{fs}(A)$. <!-- --> 1. $A$ does not contain an infinite uniformly supported subset, and so it is not FSM Dedekind infinite (according to Theorem \[ti1\](4)). 2. $\wp_{fs}(A)$ does not contain an infinite uniformly supported subset because for any finite set $S$ of atoms there exist only finitely many elements of $\wp_{fs}(A)$ supported by $S$, namely the subsets of $S$ and the supersets of $A\setminus S$. Thus, $\wp_{fs}(A)$ it is not FSM Dedekind infinite (Theorem \[ti1\](4)). 3. $T_{fin}(A)$ does not contain an infinite uniformly supported subset because the finite injective tuples of atoms supported by a finite set $S$ are only those injective tuples formed by elements of $S$, being at most $1+A_{|S|}^{1}+A_{|S|}^{2}+\ldots+A_{|S|}^{|S|}$ such tuples, where $A_{n}^{k}=n(n-1)\ldots (n-k+1)$. 4. We prove the following lemmas. \[lem”\] Let $S=\{s_{1},\ldots,s_{n}\}$ be a finite subset of an invariant set $(U, \cdot)$ and $X$ a finitely supported subset of an invariant set $(V, \diamond)$. Then if $X$ is does not contain an infinite uniformly supported subset, we have that $X^{S}_{fs}$ does not contain an infinite uniformly supported subset. *Proof of Lemma \[lem”\]*. First we prove that there is an FSM injection $g$ from $X^{S}_{fs}$ into $X^{|S|}$. For $f \in X^{S}_{fs}$ define $g(f)=(f(s_{1}),\ldots, f(s_{n}))$. Clearly $g$ is injective (and it is also surjective). Let $\pi \in Fix(supp(s_{1}) \cup \ldots \cup supp(s_{n}) \cup supp(X))$. Thus, $g(\pi \widetilde{\star} f)=(\pi \diamond f(\pi^{-1} \cdot s_{1}),\ldots, \pi \diamond f(\pi^{-1} \cdot s_{n}))=(\pi \diamond f( s_{1}),\ldots, \pi \diamond f(s_{n}))$ $=\pi \otimes g(f)$ for all $f \in X^{S}_{fs}$, where $\otimes$ is the $S_{A}$-action on $X^{|S|}$ defined as in Proposition \[p1\]. Hence $g$ is finitely supported, and the conclusion follows from Theorem \[ti1\](1) and by repeatedly applying similar arguments as in Theorem \[ti1\](6) (if we slightly modify the proof of the theorem, using the fact that $supp(x) \cup supp(y)=supp((x,y))$ for all $x,y \in X$, we show that the $|S|$-time Cartesian product of $X$, i.e. $X^{|S|}$ does not contain an infinite uniformly supported subset; otherwise $X$ should contain itself an infinite uniformly supported subset, which contradicts the hypothesis). \[lemyy\] Let $S=\{s_{1},\ldots,s_{n}\}$ be a finite subset of an invariant set $(U, \cdot)$ and $X$ a finitely supported subset of an invariant set $(V, \diamond)$. Then if $X$ is not FSM Dedekind infinite, we have that $X^{S}_{fs}$ is not FSM Dedekind-infinite. *Proof of Lemma \[lemyy\]* First we proved that there is an FSM injection $g$ from $X^{S}_{fs}$ into $X^{|S|}$. The conclusion follows from Theorem \[ti1\](1) and by repeatedly applying Theorem \[ti1\](6) (from which we know that the $|S|$-time Cartesian product of $X$, i.e. $X^{|S|}$, is not FSM Dedekind infinite). \[lem”’\]Let $f:A \to A$ be a function that is finitely supported by a certain finite set of atoms $S$. Then either $f|_{A \setminus S}=Id$ or $f|_{A \setminus S}$ is an one-element subset of $S$. *Proof of Lemma \[lem”’\]* Let $f:A \to A$ be a function that is finitely supported by the finite set of atoms $S$. We distinguish two cases: I. There is $a \notin S$ with $f(a)=a$. Then for each $b\notin S$ we have that $(a\, b) \in Fix(S)$, and so $f(b)=f((a\, b)(a))=(a\, b)(f(a))=(a\, b)(a)=b$. Thus, $f|_{A \setminus S}=Id$. II\. For all $a \notin S$ we have $f(a) \neq a$. We claim that $f(a) \in S$ for all $a \notin S$. Suppose, by contradiction, that $f(a)=b \in A \setminus S$ for a certain $a \notin S$. Thus, $(a\, b) \in Fix(S)$, and so $f(b)=f((a\, b)(a))=(a\, b)(f(a))=(a\, b)(b)=a$. Let us consider $c \in A \setminus S$, $c \neq a,b$. Thus, $(a\, c) \in Fix(S)$, and so $f(c)=f((a\, c)(a))=(a\, c)(f(a))=(a\, c)(b)=b$. Furthermore, $(b\, c) \in Fix(S)$, and so $f(b)=f((b\, c)(c))=(b\, c)(f(c))=(b\, c)(b)=c$. However, $f(b)=a$ which contradicts the functionality of $f$. Thus $f(a) \in S$ for any $a \notin S$. If $x,y \notin S$, then we should have $f(x),f(y) \in S$, and so, because $(x\,y) \in Fix(S)$, we get $f(x)=f((x\, y)(y))=(x\, y)(f(y))=f(y)$ since both $x$ and $y$ belong to $A\setminus S$ which means they are different from $f(y)$ belonging to $S$. Therefore there is $x_{0} \in S$ such that $f|_{A \setminus S}$ $=\{x_{0}\}$. *Proof of this item*. Assume, by contradiction, that $A^{A}_{fs}$ contains an infinite, uniformly supported subset, meaning that there are infinitely many functions from $A$ to $A$ supported by the same finite set $S$. According to Lemma \[lem”’\], any $S$-supported function $f:A \to A$ should have the property that either $f|_{A \setminus S}=Id$ or $f|_{A \setminus S}$ is an one-element subset of $S$. A function from $A$ to $A$ is precisely characterized by the set of values it takes on the elements of $S$ and on the elements of $A\setminus S$, respectively. For each possible definition of such an $f$ on $S$ we have at most $|S|+1$ possible ways to define $f$ on $A \setminus S$. Since we assumed that there exist infinitely many finitely supported functions from $A$ to $A$ supported by the same set $S$, there should exist infinitely many finitely supported functions from $S$ to $A$ supported by the set $S$. But this is a contradiction according to Lemma \[lem”\] which states that $A^{S}_{fs}$ is does not contain an infinite uniformly supported subset (because $A$ does not contain an infinite uniformly supported subset). 5. There is an equivariant bijective mapping between $(A^{n})^{A}_{fs}$ and $(A^{A}_{fs})^{n}$ defined as follows. If $f:A \to A^{n}$ is a finitely supported function with $f(a)=(a_{1},\ldots, a_{n})$, we associate to $f$ the Cartesian pair $(f_{1},\ldots, f_{n})$ where for each $i \in \mathbb{N}$, $f_{i}:A \to A$ is defined by $f_{i}(a)=a_{i}$ for all $a \in A$. We omit technical details since they are based only on the application of Proposition \[2.18’\]. We proved above that $A^{A}_{fs}$ does not contain an infinite uniformly supported subset, and so neither $(A^{A}_{fs})^{n}$ contains an infinite uniformly supported subset by involving a similar proof as of Theorem \[ti1\](6) (see the proof of Lemma \[lem”\]). 6. Assume by contradiction that $T_{fin}(A)^{A}$ contains an infinite $S$-uniformly supported subset. If $f:A \to T_{fin}(A)$ is a function supported by $S$, then consider $f(a)=x$ for some $a \notin S$. For $b \notin S$ we have $(a\,b) \in Fix(S)$, and so $f(b)=f((a\,b)(a))=(a\,b)\otimes f(a)=(a\,b)\otimes x$ which means $|f(a)|=|f(b)|$ for all $a,b \notin S$. Each $S$-supported function $f:A \to T_{fin}(A)$ is fully described the values it takes on the elements of $S$ and on the elements of $A\setminus S$, respectively, i.e., by the elements of $f(S)$ and of $f(A\setminus S)$. More precisely, each $S$-supported function $f:A \to T_{fin}(A)$ can be uniquely decomposed into two $S$-supported functions $f|_{S}$ and $f|_{A \setminus S}$ (this follows from Proposition \[2.18’\] and because both $S$ and $A \setminus S$ are supported by $S$). However, $f(A\setminus S) \subseteq A'^{n}$ for some $n \in \mathbb{N}$, where $A'^{n}$ is the set of all injective $n$-tuples of $A$. According to Lemma \[lem”\] we have at most finitely many $S$-supported functions from $S$ to $T_{fin}(A)$. According to item 5, we have at most finitely many $S$-supported functions from $A\setminus S$ to $A'^{n}$ for each fixed $n \in \mathbb{N}$. This is because $A'^{n}$ is a subset of $A^{n}$ and $A \setminus S$ is a subset of $A$, and so by involving Proposition \[pco1\](3) and (4) we find a finitely supported injection $\varphi$ from $(A'^{n})^{A \setminus S}$ and $(A^{n})^{A}$; if $\mathcal{K}$ was an infinite subset in $(A'^{n})^{A \setminus S}$ uniformly supported by $T$, then $\varphi(\mathcal{K})$ would be an infinite subset of $(A^{n})^{A}$ uniformly supported by $T \cup supp(\varphi)$. Therefore, there should exist an infinite subset $M \subseteq \mathbb{N}$ such that we have at least one $S$-supported function $g:A\setminus S \to A'^{k}$ for any $k \in M$. We do not need to find a set of representatives for such $g$’s; we consider all of them. Fix $a \in A \setminus S$. For each of the above $g$’s (that form an $S$-supported family $\mathcal{F}$) we have that $g(a)$’s form an uniformly supported family (by $S \cup \{a\}$) of $T_{fin}(A)$, which is also infinite because tuples having different cardinalities are different and $M$ is infinite. However, we contradict the proof of item 3 stating that $T_{fin}(A)$ does not contain an infinite uniformly supported subset. Alternatively, one can remark that if $|S \cup \{a\}|=l$ with $l$ fixed, then there is $m\in M$ fixed with $m>l$. Moreover, $g(a)$ for some $g:A\setminus S \to A'^{m}$ in $\mathcal{F}$ (we need to select only a function from those functions $g:A\setminus S \to A'^{m}$ with $m$ fixed depending only on the fixed $l$, *and not* a set of representatives for the entire family $(\{g:A\setminus S \to A'^{k}\})_{k \in M}$), which is an injective $m$-tuple of atoms, cannot be supported by $S \cup \{a\}$; thus, the set of all $g(a)$’s cannot be infinite and uniformly supported. 7. We can use a similar approach as in item 6, to prove that there exist at most finitely many $S$-supported functions from $A$ to $\wp_{fin}(A)$. For this we just replace $A'^{n}$ with the set of all $n$-sized subsets of $A$, $\wp_{n}(A)$. All it remains is to prove that, for each $n \in \mathbb{N}$, there cannot exist infinitely many functions $g:A \to \wp_{n}(A)$ supported by the same set $S'$. Fix $n \in \mathbb{N}$. Assume, by contradiction that there exist infinitely many functions $g:A \to \wp_{n}(A)$ supported by the same set $S'$. According to Lemma \[lem”\] there are only finitely many functions from $S'$ to $\wp_{n}(A)$ supported by the same set of atoms, and so there should exist infinitely many functions $g:(A\setminus S') \to \wp_{n}(A)$ supported by $S'$. For such a $g$, let us fix an element $a\in A$ with $a\notin S'$. There exist $x_{1}, \ldots, x_{n} \in A$ fixed (depending only on the fixed $a$) and different such that $g(a)=\{x_{1}, \ldots, x_{n}\}$. Let $b$ be an arbitrary element from $A\setminus S'$, and so $(a\, b)\in Fix(S')$ which means $g(b)=g((a\,b)(a))=(a\,b) \star g(a)=(a\,b) \star \{x_{1}, \ldots, x_{n}\}=\{(a\,b)(x_{1}),\ldots, (a\,b)(x_{n})\}$. We analyze the two possibilities: Case 1: One of $x_{1}, \ldots, x_{n}$ coincides to $a$. Suppose $x_{1}=a$. We claim that $x_{2}, \ldots, x_{n} \in S'$. Assume the contrary, that is, there exists $i \in \{2,\ldots,n\}$ such that $x_{i} \notin S'$. Without losing the generality suppose $x_{2} \notin S'$, which means $(a\,x_{2}) \in Fix(S')$, and so $g(x_{2})=g((a\,x_{2})(a))=(a\,x_{2}) \star g(a)=(a\,x_{2}) \star \{a,x_{2}, \ldots, x_{n}\}=\{a,x_{2}, \ldots, x_{n}\}$. Let $c \in A\setminus S'$ with $c$ different from $a,x_{2}, \ldots, x_{n}$. We have $g(c)=g((a\,c)(a))=(a\,c) \star g(a)=(a\,c) \star \{a,x_{2}, \ldots, x_{n}\}=\{c,x_{2}, \ldots, x_{n}\}$, and hence $g(x_{2})=g((c\,x_{2})(c))=(c\,x_{2}) \star g(c)=(c\,x_{2}) \star \{c,x_{2}, \ldots, x_{n}\}=\{c,x_{2}, \ldots, x_{n}\}$ which contradicts the functionality of $g$. Therefore, $g(b)=(b,x_{2},\ldots, x_{n})$ for all $b \in A \setminus S'$, and so only the selection of $x_{2}, \ldots x_{n}$ provides the distinction between $g$’s. Since $S'$ is finite, $\{x_{2}, \ldots, x_{n}\}$ can be selected in $C_{|S'|}^{n-1}$ ways if $|S'|\geq n-1$, or in $0$ ways otherwise. Case 2: Consider now that all $x_{1}, \ldots, x_{n}$ are different from $a$. Then $g(b)= \left\{ \begin{array}{ll} \{x_{1}, \ldots, x_{n}\}, & \text{if}\: \text{$b \neq x_{1}, \ldots, x_{n}$ };\\ \{a,x_{2}, \ldots, x_{n}\}, & \text{if}\: \text{$x_{1} \notin S'$ and $b=x_{1}$ };\\ \ldots\\ \{x_{1},\ldots, x_{n-1}, a\}, & \text{if}\: \text{$x_{n} \notin S'$ and $b=x_{n}$} \: .\end{array}\right.$ Since $x_{1}, \ldots, x_{n},a$ are fixed atoms, then $g(A \setminus S')$ is finite. However, $Im(g)$ should be supported by $S'$. According to Proposition \[4.4-9\], since $Im(g)$ is finite, it should be uniformly supported by $S'$. We obtain that $x_{1}, \ldots, x_{n} \in S'$, and so $g(A \setminus S') =\{x_{1}, \ldots, x_{n}\}$. Otherwise, if some $x_{i} \notin S'$, we would get $\{x_{1}, \ldots, a, \ldots, x_{n}\} \in Im(g)$ (where $a$ replaces $x_{i}$) and so $\{x_{1}, \ldots, a, \ldots, x_{n}\}$ is supported by $S'$. Again by Proposition \[4.4-9\] we would have that $a$ is supported by $S'$ which means $\{a\}=supp(a)\subseteq S'$ contradicting the choice of $a$. Alternatively, for proving that all $x_{1}, \ldots, x_{n} \in S'$, assume by contradiction that one of them (say $x_{1}$) does not belong to $S'$. Let $c$ be an atom from $A \setminus S'$ with $c$ different from $a, x_{1},x_{2}, \ldots, x_{n}$. We have $g(c)=g((a\,c)(a))=(a\,c) \star g(a)=(a\,c) \star \{x_{1}, \ldots, x_{n}\}=\{x_{1}, \ldots, x_{n}\}$, and hence $g(x_{1})=g((c\,x_{1})(c))=(c\,x_{1}) \star g(c)=(c\,x_{1}) \star \{x_{1}, x_{2}, \ldots, x_{n}\}=\{c,x_{2}, \ldots, x_{n}\}$. However $g(x_{1})=\{a,x_{2}, \ldots, x_{n}\}$ which contradicts the functionality of $g$. Since $S'$ is finite, $\{x_{1}, \ldots, x_{n}\}$ can be selected in $C_{|S'|}^{n}$ ways $|S'|\geq n$ or in $0$ ways otherwise. In either case, there couldn’t exist infinitely many $g$’s supported by $S'$, and so for each $n \in \mathbb{N}$, there exist at most finitely many functions from $A$ to $\wp_{n}(A)$ supported by the same set of atoms. Assume by contradiction that $\wp_{fin}(A)^{A}$ contains an infinite $S$-uniformly supported subset. If $f:A \to \wp_{fin}(A)$ is a function supported by $S$, then we have $|f(a)|=|(a\,b)\star f(a)|=|f((a\,b)(a))|=|f(b)|$ for all $a,b \notin S$. According to Proposition \[2.18’\] (since both $S$ and $A \setminus S$ are supported by $S$), each $S$-supported function $f:A \to \wp_{fin}(A)$ is uniquely decomposed into two $S$-supported functions $f|_{S}$ and $f|_{A \setminus S}$. However $f(A\setminus S) \subseteq \wp_{n}(A)$ for some $n \in \mathbb{N}$. According to Lemma \[lem”\] there are at most finitely many $S$-supported functions from $S$ to $\wp_{fin}(A)$. Furthermore, there exist at most finitely many $S$-supported functions from $A\setminus S$ to $\wp_{n}(A)$ for each fixed $n \in \mathbb{N}$. Therefore, there should exist an infinite subset $M \subseteq \mathbb{N}$ such that we have at least one $S$-supported function $g:A\setminus S \to \wp_{k}(A)$ for any $k \in M$. Fix $a \in A \setminus S$. For each of the above $g$’s (that form an $S$-supported family $\mathcal{F}$) we have that $g(a)$’s form an uniformly supported family (by $S \cup \{a\}$) of $\wp_{fin}(A)$. If $|S \cup \{a\}|=l$ with $l$ fixed, then there is $m\in M$ fixed with $m>l$. Moreover, $g(a)$ for $g:A\setminus S \to \wp_{m}(A) \in \mathcal{F}$, which is an $m$-sized subset of atoms, cannot be supported by $S \cup \{a\}$ (according to Proposition \[4.4-9\]); thus, the set of all $g(a)$’s cannot be infinite and uniformly supported. Analogously, there there exist at most finitely many $S$-supported functions from $A$ to $\wp_{cofin}(A)$ (using eventually the fact that there is an equivariant bijection $X \mapsto A\setminus X$ between $\wp_{fin}(A)$ and $\wp_{cofin}(A)$). Assume by contradiction that $\wp_{fs}(A)^{A}$ contains an infinite $S$-uniformly supported subset. If $f:A \to \wp_{fs}(A)$ is a function supported by $S$, then consider $f(a)=X$ for some $a \notin S$. For $b \notin S$ we have $f(b)=(a\,b)\star X$ which means $f(A \setminus S)$ is formed only by finite subsets of atoms if $X$ is finite, and $f(A \setminus S)$ is formed only by cofinite subsets of atoms if $X$ is cofinite. Thus, whenever $f:A \to \wp_{fs}(A)$ is a function supported by $S$, we have either $f(A \setminus S) \subseteq \wp_{fin}(A)$ or $f(A \setminus S) \subseteq \wp_{cofin}(A)$. Each $S$-supported function $f:A \to \wp_{fs}(A)$ is fully described by $f(S)$ and $f(A\setminus S)$. According to Lemma \[lem”\] we have at most finitely many $S$-supported functions from $S$ to $\wp_{fs}(A)$. Furthermore, we have at most finitely many $S$-supported functions from $A\setminus S$ to $\wp_{fin}(A)$, and at most finitely many $S$-supported functions from $A\setminus S$ to $\wp_{cofin}(A)$. Thus, $\wp_{fs}(A)^{A}$ does not contain an infinite uniformly supported subset. 8. The sets $\wp_{fin}(A)$, $\wp_{cofin}(A)$, $\wp_{fin}(\wp_{fs}(A))$, $\wp_{fin}(\wp_{cofin}(A))$, $\wp_{fin}(\wp_{fin}(A))$, $\wp_{fin}(A^{A}_{fs})$ do not contain infinite uniformly supported subsets, and so they are not FSM Dedekind infinite (Theorem \[ti1\](5)). 9. Directly from Theorem \[ti1\](5). 10. According to Theorem \[ti1\](6). 11. According to Theorem \[ti1\](7). There exist two FSM sets that whose cardinalities incomparable via the relation $\leq$ on cardinalities, and none of them is FSM Dedekind infinite. According to Corollary \[ti2\], none of the sets $A \times A$ and $\wp_{fs}(A)$ is FSM Dedekind infinite. According to Theorem \[cardord1\], there does not exist a finitely supported injective mapping $f: A \times A \to \wp_{fs}(A)$. According to Lemma 11.10 from [@jech] that is preserved in FSM (proof omitted) there does not exist a finitely supported injective mapping $f: \wp_{fs}(A) \to A \times A$. \[ti3\] The following sets and all of their supersets, their powersets and the families of their finite subsets, are both FSM usual infinite and FSM Dedekind infinite. 1. The invariant sets $\wp_{fs}(\wp_{fs}(A))$, $\wp_{fs}(\wp_{fin}(A))$ and $\mathbb{N}$. 2. The set of all finitely supported mappings from $X$ to $Y$, and the set of all finitely supported mappings from $Y$ to $X$, where $X$ is a finitely supported subset of an invariant set with at least two elements, and $Y$ is an FSM Dedekind infinite set. 3. The set of all finitely supported functions $f:\wp_{fin}(Y) \to X$ and the set of all finitely supported functions $f:\wp_{fs}(Y) \to X$, where $Y$ is an infinite finitely supported subset of an invariant set, and $X$ is a finitely supported subset of an invariant set with at least two elements. 4. The set $T^{\delta}_{fin}(A)=\underset{n\in \mathbb{N}}{\cup}A^{n}$ of all finite tuples of atoms (not necessarily injective). <!-- --> 1. This follows from Theorem \[ti1\](3) and Theorem \[ti1\] (2). 2. Let $(y_{n})_{n \in \mathbb{N}}$ be an injective, uniformly supported, countable sequence in $Y$ (that exists from Theorem \[ti1\](1)). Thus, each $y_{n}$ is supported by the same set $S$ of atoms. In $Y^{X}$ we consider the injective family $(f_{n})_{n \in \mathbb{N}}$ of functions from $X$ to $Y$ where for each $i \in \mathbb{N}$ we define $f_{i}(x)=y_{i}$ for all $x \in X$. According to Proposition \[2.18’\], each $f_{i}$ is supported by $S$, and so is the infinite family $(f_{n})_{n \in \mathbb{N}}$, meaning that there is an $S$-supported injective mapping from $\mathbb{N}$ to $Y^{X}$. In this case it is necessary to require only that $X$ is non-empty. Fix two different elements $x_{1}, x_{2} \in X$. Take $\mathcal{F}=(y_{n})_{n \in \mathbb{N}}$ an injective, uniformly supported, countable sequence in $Y$. In $X^{Y}$ we consider the injective family $(g_{n})_{n \in \mathbb{N}}$ of functions from $Y$ to $X$ where for each $i \in \mathbb{N}$ we define $g_{i}(y)=\left\{ \begin{array}{ll} x_{1} & \text{if}\: y=y_{i}\\ x_{2} & \text{if}\: y=y_{j} \; \text{with}\; j \neq i, \; \text{or}\; y\notin \mathcal{F}\end{array} \right.$. According to Proposition \[2.18’\], each $g_{i}$ is supported by the finite set $supp(x_{1}) \cup supp(x_{2}) \cup supp(\mathcal{F})$, and so the infinite family $(g_{n})_{n \in \mathbb{N}}$ is uniformly supported meaning that there is an injective mapping from $\mathbb{N}$ to $X^{Y}$ supported by $supp(x_{1}) \cup supp(x_{2}) \cup supp(\mathcal{F})$. 3. From Theorem \[comp\], there exists a one-to-one mapping from $\wp_{fs}(U)$ onto $\{0,1\}^{U}_{fs}$ for an arbitrary finitely supported subset of an invariant set $U$. Fix two distinct elements $x_{1},x_{2} \in X$. There exists a finitely supported (by $supp(x_{1}) \cup supp(x_{2})$) bijective mapping from $\{0,1\}^{U}_{fs}$ to $\{x_{1},x_{2}\}^{U}_{fs}$ which associates to each $f \in \{0,1\}^{U}_{fs}$ an element $g \in \{x_{1},x_{2}\}^{U}_{fs}$ defined by $g(x)=\left\{ \begin{array}{ll} x_{1} & \text{for}\: f(x)=0\\ x_{2} & \text{for}\: f(x)=1 \end{array}\right.$ for all $x \in U$ and supported by $supp(x_{1}) \cup supp(x_{2}) \cup supp(f)$. Obviously, there is a finitely supported injection between $\{x_{1},x_{2}\}^{U}_{fs}$ and $X^{U}_{fs}$. Thus, there is a finitely supported injection from $\wp_{fs}(U)$ into $X^{U}_{fs}$. If we take $U=\wp_{fin}(Y)$ or $U=\wp_{fs}(Y)$, the result follows from Theorem \[ti1\](1), Theorem \[ti1\](2) and Theorem \[ti1\](3). 4. Fix $a \in A$ and $i \in \mathbb{N}$. We consider the tuple $x_{i}=(a,\ldots,a) \in A^{i}$. Clearly $x_{i}$ is supported by $\{a\}$ for each $i \in \mathbb{N}$, and so $(x_{n})_{n \in \mathbb{N}}$ is a uniformly supported subset of $T^{\delta}_{fin}(A)$. \[propro\] Let $X$ be a finitely supported subset of an invariant set such that $\wp_{fs}(X)$ is not FSM Dedekind infinite. Then each finitely supported surjective mapping $f:X \to X$ should be injective. Let $f: X \to X$ be a finitely supported surjection. Since $f$ is surjective, we can define the function $g:\wp_{fs}(X) \rightarrow \wp_{fs}(X)$ by $g(Y) = f^{-1}(Y)$ for all $Y\in \wp_{fs}(X)$ which is finitely supported and injective according to Lemma \[lemlem\]. Since $\wp_{fs}(X)$ is not FSM Dedekind infinite, it follow that $g$ is surjective. Now let us consider two elements $a,b \in X$ such that $f(a)=f(b)$. We prove by contradiction that $a=b$. Suppose that $a \neq b$. Let us consider $Y=\{a\}$ and $Z=\{b\}$. Obviously, $Y,Z \in \wp_{fs}(X)$. Since $g$ is surjective, for $Y$ and $Z$ there exist $Y_{1}, Z_{1} \in \wp_{fs}(X)$ such that $f^{-1}(Y_{1})=g(Y_{1})=Y$ and $f^{-1}(Z_{1})=g(Z_{1})=Z$. We know that $f(Y) \cap f(Z)= \{f(a)\}$. Thus, $f(a) \in f(Y)=f(f^{-1}(Y_{1})) \subseteq Y_{1}$. Similarly, $f(a) =f(b) \in f(Z)=f(f^{-1}(Z_{1})) \subseteq Z_{1}$, and so $f(a) \in Y_{1} \cap Z_{1}$. Thus, $a \in f^{-1}(Y_{1} \cap Z_{1})=f^{-1}(Y_{1}) \cap f^{-1}(Z_{1})=Y \cap Z$. However, since we assumed that $a \neq b$, we have that $Y \cap Z = \emptyset$, which represents a contradiction. It follows that $a=b$, and so $f$ is injective. \[propro”\] 1. Let $X$ be a finitely supported subset of an invariant set. If $\wp_{fin}(X)$ is FSM Dedekind infinite, then $X$ should be FSM non-uniformly amorphous, meaning that $X$ should contain two disjoint, infinite, uniformly supported subsets. 2. Let $X$ be a finitely supported subset of an invariant set. If $\wp_{fs}(X)$ is FSM Dedekind infinite, then $X$ should be FSM non-amorphous, meaning that $X$ should contain two disjoint, infinite, finitely supported supported subsets. The reverse implication is not valid. 1\. Assume that $(X_{n})_{n \in \mathbb{N}}$ is a countable family of different finite subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported. Thus, each $X_{n}$ is supported by the same set $S=supp(n \mapsto X_{n})$. Since each $X_{n}$ is finite (and the support of a finite set coincides with the union of the supports of its elements), as in the proof of Lemma \[lem4\], we have that $\underset{n \in \mathbb{N}}\cup X_{n}$ is uniformly supported by $S$. Furthermore, $\underset{n \in \mathbb{N}}\cup X_{n}$ is infinite since all $X_{i}$ are pairwise different. Moreover, the countable sequence $(Y_{n})_{n \in \mathbb{N}}$ defined by $Y_{n}=X_{n} \setminus \underset{m<n}\cup X_{m}$ is a uniformly supported (by $S$) sequence of pairwise disjoint uniformly supported sets with $\underset{n \in \mathbb{N}}\cup X_{n}=\underset{n \in \mathbb{N}}\cup Y_{n}$. Again since each $Y_{n}$ is finite (and the support of a finite set coincides with the union of the supports of its elements), any element belonging to a set from the sequence $(Y_{n})_{n \in \mathbb{N}}$ is $S$-supported. Since the union of all $Y_{n}$ is infinite, and each $Y_{n}$ is finite, there should exist infinitely many terms from the sequence $(Y_{n})_{n \in \mathbb{N}}$ that are non-empty. Assume that $(Y_{n})_{n \in M \subseteq \mathbb{N}}$ with $M$ infinite is a subset of $(Y_{n})_{n \in \mathbb{N}}$ formed by non-empty terms. Let $U_{1}=\{\cup Y_{k}\,|\, k \in M, k \;\text{is odd}\}$ and $U_{2}=\{\cup Y_{k}\,|\, k \in M, k\; \text{is even}\}$. Then $U_{1}$ and $U_{2}$ are disjoint, uniformly $S$-supported and infinite subsets of $X$. 2\. Assume that $\wp_{fs}(X)$ is FSM Dedekind infinite. As in the proof of Theorem \[ti1\](8), we can define a uniformly supported, countable family $(Y_{n})_{n \in \mathbb{N}}$ of subsets of $X$ that are non-empty and pairwise disjoint. Let $V_{1}=\{\cup Y_{k}\,|\, k \;\text{is odd}\}$ and $V_{2}=\{\cup Y_{k}\,|\, k\; \text{is even}\}$. Then $V_{1}$ and $V_{2}$ are disjoint, infinite subsets of $X$. Since each $Y_{i}$ is supported by $S'=supp(n \mapsto Y_{n})$ we have $\pi \star Y_{i}=Y_{i}$ for all $i \in \mathbb{N}$ and $\pi \in Fix(S')$. Fix $\pi \in Fix(S')$ and $x \in V_{1}$. Thus, there is $l \in \mathbb{N}$ such that $x \in Y_{2l+1}$. We obtain $\pi \cdot x \in \pi \star Y_{2l+1}=Y_{2l+1}$, and so $\pi \cdot x \in V_{1}$. Thus, $V_{1}$ is supported by $S'$. Analogously, $V_{2}$ is supported by $S'$, and so $X$ is FSM non-amorphous. Conversely, the set $A+A=\{0,1\}\times A$ (the disjoint union of $A$ and $A$) is obviously non-amorphous because $\{(0,a)\,|\,a \in A\}$ is equivariant, infinite and coinfinite. One can define the equivariant bijection $f: \wp_{fs}(A) \times \wp_{fs}(A) \to \wp_{fs}(\{0,1\}\times A)$ by $f(U,V)=\{(0,x)\,|\,x \in U\} \cup \{(1,y)\,|\,y \in V\}$ for all $U,V \in \wp_{fs}(A)$. Clearly $f$ is equivariant because for each $\pi \in S_{A}$ we have $f(\pi \star U,\pi \star V)=\pi \star f(U,V)$. However, $\wp_{fs}(A) \times \wp_{fs}(A)$ is not FSM Dedekind infinite according to Corollary \[ti2\](2) and Theorem \[ti1\](6). It is also worth noting that non-uniformly amorphous FSM sets are non-amorphous FSM sets since uniformly supported sets are obviously finitely supported. The converse however is not valid since $\wp_{fin}(A)$ is non-amorphous but it has no infinite uniformly supported subset (the only finite subsets of atoms supported by a finite set $S$ of atoms being the subsets of $S$), and so it cannot be non-uniformly amorphous. \[propro’\] Let $X$ be a finitely supported amorphous subset of an invariant set (i.e. any finitely supported subset of $X$ is either finite or cofinite). Then each finitely supported surjective mapping $f:X \to X$ should be injective. Since any finitely supported subset of $X$ is either finite or cofinite, then any uniformly supported subset of $X$ is either finite or cofinite. From Proposition \[propro”\], $\wp_{fin}(X)$ is not FSM Dedekind infinite. For the rest of the proof we follow step-by-step the proof of Proposition \[propro\] (and of Lemma \[lemlem\]). If $X$ is finite, we are done, so assume $X$ is infinite. If $Y \in \wp_{fin}(X)$, then $f^{-1}(Y) \in \wp_{fs}(X)$ (supported by $supp(f) \cup supp(X) \cup supp(Y)$). Since $X$ is amorphous, it follows that $f^{-1}(Y)$ is either finite or cofinite. If $f^{-1}(Y)$ is cofinite, then its complementary $\{x \in X \;|\; f(x) \notin Y\}$ is finite. This means that all but finitely many elements in $X$ would have their image under $f$ belonging to the finite set $Y$. Therefore, $Im(f)$ would be a finite subset of $X$, which contradicts the surjectivity of $f$. Thus, $f^{-1}(Y)$ is a finite subset of $X$. In this sense we can well-define define the function $g:\wp_{fin}(X) \rightarrow \wp_{fin}(X)$ by $g(Y) = f^{-1}(Y)$ which is supported by $supp(f) \cup supp(X)$ and injective. Since $\wp_{fin}(X)$ is not FSM Dedekind infinite, it follow that $g$ is surjective, and so $f$ is injective exactly as in the last paragraph of the proof of Proposition \[propro\]. 1. Let $X$ be an FSM Dedekind infinite set. Then there exists a finitely supported surjection $j:X \to \mathbb{N}$. The reverse implication is not valid. 2. If $X$ is a finitely supported subset of an invariant set such that there exists a finitely supported surjection $j:X \to \mathbb{N}$, then $\wp_{fs}(X)$ is FSM Dedekind infinite. The reverse implication is also valid. 1\. Let $X$ be an FSM Dedekind infinite set. According to Theorem \[ti1\](1), there is a finitely supported injection $i:\mathbb{N} \to X$. Let us fix $n_{0} \in \mathbb{N}$. We define the function $j: X \to \mathbb{N}$ by $$j(x)=\left\{ \begin{array}{ll} i^{-1}(x), & \text{if}\: \text{$x\in Im(i)$ };\\ n_{0}, & \text{if}\: \text{$x\notin Im(i)$}\: .\end{array}\right.$$ Since $Im(i)$ is supported by $supp(i)$ and $n_{0}$ is empty supported, by verifying the condition in Proposition \[2.18’\] we have that $j$ is supported by $supp(i) \cup supp(X)$. Indeed, when $\pi \in Fix(supp(i) \cup supp(X))$, then $x \in Im(i) \Leftrightarrow \pi \cdot x \in Im(i)$, and $n=i^{-1}(\pi \cdot x) \Leftrightarrow i(n)=\pi \cdot x \Leftrightarrow \pi^{-1} \cdot i(n)=x \Leftrightarrow i(\pi^{-1} \diamond n)=x \Leftrightarrow i(n)=x \Leftrightarrow n=i^{-1}(x)$, where $\diamond$ is the trivial action on $\mathbb{N}$; similarly $y \notin Im(i) \Leftrightarrow \pi \cdot y \notin Im(i)$ and $j(\pi \cdot y)=n_{0}=\pi \diamond n_{0}=\pi \diamond j(y)$. Clearly, $j$ is surjective. However, the reverse implication is not valid because the mapping $f:\wp_{fin}(A) \to \mathbb{N}$ defined by $f(X)=|X|$ for all $X \in \wp_{fin}(A)$ is equivariant and surjective, but $\wp_{fin}(A)$ is not FSM Dedekind infinite. 2\. Suppose now there exists a finitely supported surjection $j:X \to \mathbb{N}$. Clearly, for any $n \in \mathbb{N}$, the set $j^{-1}(\{n\})$ is non-empty and supported by $supp(j)$. Define $f: \mathbb{N} \to \wp_{fs}(X)$ by $f(n)=j^{-1}(\{n\})$. For $\pi \in Fix(supp(j))$ and an arbitrary $n \in \mathbb{N}$ we have $j(x)=n \Leftrightarrow j(\pi^{-1} \cdot x)=n$, and so $x \in j^{-1}(\{n\}) \Leftrightarrow \pi^{-1} \cdot x \in j^{-1}(\{n\})$, which means $f(n) = \pi \star f(n)$ for all $n \in \mathbb{N}$, and so $f$ is supported by $supp(j)$. Since $f$ is also injective, by Theorem \[ti1\](1) we have that $\wp_{fs}(X)$ is FSM Dedekind infinite. Conversely, assume that $\wp_{fs}(X)$ is FSM Dedekind infinite. As in the proof of Theorem \[ti1\](8), we can define a uniformly supported, countable family $(Y_{n})_{n \in \mathbb{N}}$ of subsets of $X$ that are non-empty and pairwise disjoint. The mapping $f$ can be defined by $ f(x)=\left\{ \begin{array}{ll} n, & \text{if}\; \exists n. x\in Y_{n} ;\\ 0, & \text{otherwise} \end{array}\right.$, and, obviously, $f$ is supported by $supp(n \mapsto Y_{n})$. Let $X$ be an infinite finitely supported subset of an invariant set. Then there exists a finitely supported surjection $f:\wp_{fs}(X) \to \mathbb{N}$. Let $X_{i}$ be the set of all $i$-sized subsets from $X$, i.e. $X_{i}=\{Z \subseteq X\,|\,|Z|=i\}$. The family $(X_{i})_{i \in \mathbb{N}}$ is uniformly supported by $supp(X)$ and all $X_{i}$ are non-empty and pairwise disjoint. Define the mapping $f$ by $f(Y)=\left\{ \begin{array}{ll} n, & \text{if}\: Y\in X_{n};\\ 0, & \text{if}\; Y\; \text{is infinite} .\end{array}\right.$ According to Proposition \[2.18’\], $f$ is supported by $supp(X)$ (since any $X_{n}$ is supported by $supp(X)$) and it is surjective. We actually proved the existence of a finitely supported surjection from $\wp_{fin}(X)$ onto $\mathbb{N}$. The sets $A$ and $\wp_{fin}(A)$ are both FSM usual infinite and none of them is FSM Dedekind infinite. We prove below that $A$ is not FSM ascending infinite, while $\wp_{fin}(A)$ is FSM ascending infinite. \[tari\] - The set $A$ is not FSM ascending infinite. - Let $X$ be a finitely supported subset of an invariant set $U$. If $X$ is FSM usual infinite, then the set $\wp_{fin}(X)$ is FSM ascending infinite. In order to prove that $A$ is not FSM ascending infinite, we prove firstly that each finitely supported increasing countable chain of finitely supported subsets of $A$ must be stationary. Indeed, if there exists an increasing countable chain $X_{0}\subseteq X_{1}\subseteq\ldots\subseteq A$ such that $n \mapsto X_{n}$ is finitely supported, then, according to Proposition \[2.18’\] and because $\mathbb{N}$ is a trivial invariant set, each element $X_{i}$ of the chain must be supported by the same $S=supp(n \mapsto X_{n})$. However, there are only finitely many such subsets of $A$ namely the subsets of $S$ and the supersets of $A\setminus S$. Therefore the chain is finite, and, because it is ascending, there exists $n_{0}\in\mathbb{N}$ such that $X_{n}=X_{n_{0}},\forall n\geq n_{0}$. Now, let $Y_{0}\subseteq Y_{1}\subseteq\ldots\subseteq Y_{n}\subseteq\ldots$ be a finitely supported countable chain with $A\subseteq \underset{n \in \mathbb{N}}{\cup} Y_{n}$. Then $A \cap Y_{0}\subseteq A \cap Y_{1}\subseteq\ldots\subseteq A \cap Y_{n}\subseteq\ldots \subseteq A$ is a finitely supported countable chain of subsets of $A$ (supported by $supp(n \mapsto Y_{n})$) which should be stationary (finite). Furthermore, since $\underset{i \in \mathbb{N}}{\cup}(A\cap Y_{i})=A \cap(\underset{i \in \mathbb{N}}{\cup} Y_{i})=A$, there is some $k_{0}$ such that $A \cap Y_{k_{0}}=A$, and so $A \subseteq Y_{k_{0}}$. Thus, $A$ is not FSM ascending infinite. We know that $\wp_{fin}(X)$ is a subset of the invariant set $\wp_{fin}(U)$ supported by $supp(X)$. Let us consider $X_{n}\!=\!\{Z\!\in\!\wp_{fin}(X)\,|\,|Z|\!\leq\! n\}$. Clearly, $X_{0}\subseteq X_{1}\subseteq\ldots\subseteq X_{n}\subseteq\ldots$. Furthermore, because permutations of atoms are bijective, we have that for an arbitrary $k \in \mathbb{N}$, $|\pi \star Y| =|Y|$ for all $\pi \in S_{A}$ and all $Y \in X_{k}$, and so $\pi \star Y \in X_{k}$ for all $\pi \in Fix(supp(X))$ and all $Y \in X_{k}$. Thus, each $X_{k}$ is a subset of $\wp_{fin}(X)$ finitely supported by $supp(X)$, and so $(X_{n})_{n \in \mathbb{N}}$ is finitely (uniformly) supported by $supp(X)$. Obviously, $\wp_{fin}(X)=\underset{n \in \mathbb{N}}{\cup} X_{n}$. However, there exists no $n\in\mathbb{N}$ such that $\wp_{fin}(X)=X_{n}$. Thus, $(\wp_{fin}(X), \star)$ is FSM ascending infinite. \[Tt\] Let $X$ be a finitely supported subset of an invariant set $(Z, \cdot)$. 1. If $X$ is FSM Dedekind infinite, then $X$ is FSM Mostowski infinite. 2. If $X$ is FSM Mostowski infinite, then $X$ is FSM Tarski II infinite. The reverse implication is not valid. 1\. Suppose $X$ is FSM Dedekind infinite. According to Theorem \[ti1\](1) there exists an uniformly supported infinite injective sequence $T=(x_{n})_{n \in \mathbb{N}}$ of elements from $X$. Thus, each element of $T$ is supported by $supp(T)$ and there is a bijective correspondence between $\mathbb{N}$ and $T$ defined as $n \mapsto x_{n}$ which is supported by $supp(T)$. If we define the relation $\sqsubset$ on $T$ by: $x_{i} \sqsubset x_{j}$ if and only if $i<j$, we have that $\sqsubset$ is a (strict) total order relation supported by $supp(T)$. Thus $T$ is an infinite finitely supported (strictly) totally ordered subset of $X$, and so $X$ is FSM Mostowski infinite since any strict total order can be extended to a total order. 2\. Suppose that $X$ is not FSM Tarski II infinite. Then every non-empty finitely supported family of finitely supported subsets of $X$ which is totally ordered by inclusion has a maximal element under inclusion. Let $(U, <)$ be a finitely supported strictly totally ordered subset of $X$ (any total order relation induces a strict total order relation). We prove that $U$ is finite, and so $X$ is not FSM Mostowski infinite. In this sense it is sufficient to prove that $<$ and $>$ are well-orderings. Since both of them are (strict) total orderings, we need to prove that any finitely supported subset of $U$ has a least and a greatest element wrt $<$, i.e. a minimal and a maximal element (because $<$ is total). Let $Y$ be a finitely supported subset of $U$. The set $\downarrow z=\{y \in Y\,|\,y<z\}$ is supported by $supp(z) \cup supp(Y) \cup supp(<)$ for all $z \in Y$. The family $T=\{\downarrow z\,|\,z \in Y\}$ is itself finitely supported by $supp(Y) \cup supp(<)$ because for all $\pi \in Fix(supp(Y) \cup supp(<))$ we have $\pi \cdot \downarrow z=\downarrow \pi \cdot z$. Since $<$ is transitive, we have that $T$ is (strictly) totally ordered by inclusion, and so it has a maximal element, which means $Y$ has a maximal element. Analogously, the set $\uparrow z=\{y \in Y\,|\,z<y\}$ is supported by $supp(z) \cup supp(Y) \cup supp(<)$ for all $z \in Y$ and the family $T'=\{\uparrow z\,|\,z \in Y\}$ is itself finitely supported by $supp(Y) \cup supp(<)$ because for all $\pi \in Fix(supp(Y) \cup supp(<))$ we have $\pi \cdot \uparrow z=\uparrow \pi \cdot z$. The family $T'$ is (strictly) totally ordered by inclusion, and so it has a maximal element, from which $Y$ has a minimal element. We used the obvious properties $z<t$ if and only if $\downarrow z \subset \downarrow t$, and $z<t$ if and only if $\uparrow t \subset \uparrow z$. Conversely, according to Proposition \[tari\], $\wp_{fin}(A)$ is FSM ascending infinite, and so it is FSM Tarski II infinite. However, $\wp_{fin}(A)$ is not FSM Mostowski infinite, according to Corollary \[pTII\]. \[propamo\] Let $X$ be a finitely supported subset of an invariant set $(Z, \cdot)$. If $X$ is FSM Mostowski infinite, then $X$ is non-amorphous meaning that $X$ can be expressed as a disjoint union of two infinite finitely supported subsets. The reverse implication is not valid. Suppose that there is an infinite finitely supported totally ordered subset $(Y, \leq)$ of $X$. Assume, by contradiction, that $Y$ is amorphous, meaning that any finitely supported subset of $Y$ is either finite or cofinite. As in the proof of Theorem \[Tt\] (without making the requirement that $\leq$ is strict, which anyway would not essentially change the proof), for $z \in Y$ we define the finitely supported subsets $\downarrow z=\{y \in Y\,|\,y \leq z\}$ and $\uparrow z=\{y \in Y\,|\,z \leq y\}$ for all $z \in Y$. We have that the mapping $z \mapsto \downarrow z$ from $Y$ to $T=\{\downarrow z\,|\,z \in Y\}$ is itself finitely supported by $supp(Y) \cup supp(\leq)$. Furthermore it is bijective, and so $T$ is amorphous. Thus, any subset $Z$ of $T$ is either finite or cofinite, and obviously any subset $Z$ of $T$ is finitely supported. Analogously, the mapping $z \mapsto \uparrow z$ from $Y$ to $T'=\{\uparrow z\,|\,z \in Y\}$ is finitely supported and bijective, which means that any subset of $T'$ is either finite or cofinite, and clearly any subset of $T'$ is finitely supported. We distinguish the following two cases: 1\. There are only finitely many elements $x_{1}, \ldots, x_{n} \in Y$ such that $\downarrow x_{1}, \ldots, \downarrow x_{n}$ are finite. Thus, for $y \in U= Y \setminus \{x_{1}, \ldots, x_{n}\}$ we have $\downarrow y$ infinite. Since $\downarrow y$ is a subset of $Y$, it should be cofinite, and so $\uparrow y$ is finite (because $\leq$ is a total order relation). Let $M=\{\uparrow y\,|\,y \in U\}$. As in Theorem \[Tt\] we have that $M$ is totally ordered with respect to sets inclusion. Furthermore, for an arbitrary $y \in U$ we cannot have $y \leq x_{k}$ for some $k \in \{1,\ldots,n\}$ because $\downarrow y$ is infinite, while $\downarrow x_{k}$ is finite, and so $\uparrow y$ is a subset of $U$. Thus, $M$ is an infinite, finitely supported (by $supp(U) \cup supp(\leq)$), totally ordered family formed by finite subsets of $U$. Since $M$ is finitely supported, for each $y \in U$ and each $\pi \in Fix(supp(M))$ we have $\pi \cdot \uparrow y \in M$. Since $\uparrow y$ is finite, we have that $\pi \cdot \uparrow y$ is finite having the same number of elements as $\uparrow y$. Since $\pi \cdot \uparrow y$ and $\uparrow y$ are comparable via inclusion, they should be equal. Thus, $M$ is uniformly supported. Since $\leq$ is a total order, for $\pi \in Fix(supp(\uparrow y))$ we have $\uparrow \pi \cdot y=\pi \cdot \uparrow y=\uparrow y$, and so $\pi \cdot y=y$, from which $supp(y) \subseteq supp(\uparrow y)$. Thus, $U$ is uniformly supported. Since any element of $U$ has only a finite number of successors (leading to the conclusion that $\geq$ is an well-ordering on $U$ uniformly supported by $supp(U)$) and $U$ is *uniformly supported*, we can define an order monomorphism between $\mathbb{N}$ and $U$ which is supported by $supp(U)$. For example, choose $u_{0}\neq u_{1} \in U$, then let $u_{2}$ be *the greatest element* (w.r.t. $\leq$) in $U\setminus \{u_{0}, u_{1}\}$, $u_{3}$ be *the greatest element* in $U\setminus \{u_{0}, u_{1}, u_{2}\}$ (no choice principle is used since $\geq$ is an well-ordering, and so such a *greatest* element is precisely defined), and so on, and find an infinite, uniformly supported countable sequence $u_{0}, u_{1}, u_{2},\ldots$. Since $\mathbb{N}$ is non-amorphous (being expressed as the union between the even elements and the odd elements), we conclude that $U$ is non-(uniformly) amorphous containing two infinite uniformly supported disjoint subsets. 2\. We have cofinitely many elements $z$ such that $\downarrow z$ is finite. Thus, there are only finitely many elements $y_{1}, \ldots, y_{m} \in Y$ such that $\downarrow y_{1}, \ldots, \downarrow y_{m}$ are infinite. Since every infinite subset of $Y$ is cofinite, only $\uparrow y_{1}, \ldots, \uparrow y_{m}$ are finite. Let $z \in Y \setminus \{y_{1}, \ldots, y_{m}\}$ which means $\uparrow z$ infinite. Since $\uparrow z$ is a subset of $Y$ it should be cofinite, and so $\downarrow z$ is finite. As in the above item, the set $M'=\{\downarrow z\,|\,z \in Y \setminus \{y_{1}, \ldots, y_{m}\} \}$ is an infinite, finitely supported, totally ordered (by inclusion) family of finite sets, and so it has to be uniformly supported, from which $Y \setminus \{y_{1}, \ldots, y_{m}\}$ is uniformly supported, and so $\leq$ is an FSM well ordering on $Y \setminus \{y_{1}, \ldots, y_{m}\}$. Therefore, $Y \setminus \{y_{1}, \ldots, y_{m}\}$ has an infinite, uniformly supported, countable subset, and so $Y \setminus \{y_{1}, \ldots, y_{m}\}$ is non-(uniformly) amorphous containing two infinite uniformly supported disjoint subsets. Thus $Y$ is non-amorphous, and so $X$ is non-amorphous. Conversely, the set $A+A$ (the disjoint union of $A$ and $A$) is obviously non-amorhpous because because $\{(0,a)\,|\,a \in A\}$ is equivariant, infinite and coinfinite. However, if we assume there exists a finitely supported total order relation on an infinite subset of $A+A$, then there should exist an infinite, finitely supported, total order on at least one of the sets $\{(0,a)\,|\,a \in A\}$ or $\{(1,a)\,|\,a \in A\}$, which leads to an infinite finitely supported total order relation on $A$. However $A$ is not FSM Mostowski infinite by Corollary \[pTII\]. \[Tti\] Let $X$ be a finitely supported subset of an invariant set $(Z, \cdot)$. If $X$ contains no infinite uniformly supported subset, then $X$ is not FSM Mostowski infinite. Assume, by contradiction, that $X$ is FSM Mostowski infinite, meaning that $X$ contains an infinite, finitely supported, totally ordered subset $(Y, \leq)$. We claim that $Y$ is uniformly supported by $supp(\leq) \cup supp(Y)$. Let $\pi \in Fix (supp(\leq) \cup supp(Y))$ and let $y \in Y$ an arbitrary element. Since $\pi$ fixes $supp(Y)$ pointwise and $supp(Y)$ supports $Y$, we obtain that $\pi \cdot y \in Y$, and so we should have either $y<\pi \cdot y$, or $y=\pi \cdot y$, or $\pi \cdot y<y$. If $y<\pi \cdot y$, then, because $\pi$ fixes $supp(\leq)$ pointwise and because the mapping $z\mapsto \pi \cdot z$ is bijective from $Y$ to $\pi \star Y$, we get $y<\pi \cdot y<\pi^{2} \cdot y<\ldots < \pi^{n} \cdot y$ for all $n \in \mathbb{N}$. However, since any permutation of atoms interchanges only finitely many atoms, it has a finite order in the group $S_{A}$, and so there is $m \in \mathbb{N}$ such that $\pi^{m}=Id$. This means $\pi^{m} \cdot y=y$, and so we get $y<y$ which is a contradiction. Analogously, the assumption $\pi \cdot y<y$, leads to the relation $\pi^{n} \cdot y<\ldots<\pi \cdot y<y$ for all $n \in \mathbb{N}$ which is also a contradiction since $\pi$ has finite order. Therefore, $\pi \cdot y=y$, and because $y$ was arbitrary chosen form $Y$, $Y$ should be a uniformly supported infinite subset of $X$. Looking to the proof of Proposition \[propamo\], the following result follows directly. Let $X$ be a finitely supported subset of an invariant set $(Z, \cdot)$. If $X$ is FSM Mostowski infinite, then $X$ is non-uniformly amorphous meaning that $X$ has two disjoint, infinite, uniformly supported subsets. \[remarema\] In a permutation model of set theory with atoms, a set can be well-ordered if and only if there is a one-to-one mapping of the related set into the kernel of the model. Also it is noted that axiom of choice is valid in the kernel of the model [@jech]. Although FSM/nominal is somehow related to (has connections with) permutation models of set theory with atoms, it is independently developed over ZF without being necessary to relax the axioms of extensionality or foundation. FSM sets are ZF sets together with group actions, and such a theory makes sense over ZF without being necessary to require the validity of the axiom of choice on ZF sets. Thus, FSM is the entire ZF together with atomic sets with finite support (where the set of atoms is a fixed ZF formed by element whose internal structure is ignored and which are basic in the higher order construction). There may exist infinite ZF sets that do not contain infinite countable subsets, and as well there may exist infinite uniformly supported FSM sets (particularly such ZF sets) that do not contain infinite countable, uniformly supported, subsets. \[pTII\] 1. The sets $A$, $A+A$ and $A \times A$ are FSM usually infinite, but there are not FSM Mostowski infinite, nor FSM Tarski II infinite. 2. None of the sets $\wp_{fin}(A)$, $\wp_{cofin}(A)$, $\wp_{fs}(A)$ and $\wp_{fin}(\wp_{fs}(A))$ is Mostowski infinite in FSM. 3. None of the sets $A^{A}_{fs}$, $T_{fin}(A)^{A}_{fs}$ and $\wp_{fs}(A)^{A}_{fs}$ is FSM Mostowksi infinite. In the view of Theorem \[Tti\] it is sufficient to prove that none of the sets $A$, $\wp_{fin}(A)$, $\wp_{cofin}(A)$, $\wp_{fs}(A)$, $A+A$, $A \times A$, $A^{A}_{fs}$,$T_{fin}(A)^{A}_{fs}$ and $\wp_{fs}(A)^{A}_{fs}$ contain infinite uniformly supported subsets. For $A$, $\wp_{fin}(A)$, $\wp_{cofin}(A)$ and $\wp_{fs}(A)$ this is obvious since for any finite set $S$ of atoms there are at most finitely many subsets of $A$ supported by $S$, namely the subsets of $S$ and the supersets of $A \setminus S$. Moreover, $\wp_{fin}(\wp_{fs}(A))$ does not contain an infinite uniformly supported subset according to Lemma \[lem4\] since $\wp_{fs}(A)$ does not contain an infinite uniformly supported subset. Regarding $A^{A}_{fs}$, the things are also similar with Corollary \[ti2\](4). According to Lemma \[lem”’\], any $S$-supported function $f:A \to A$ should have the property that either $f|_{A \setminus S}=Id$ or $f|_{A \setminus S}$ is an one-element subset of $S$. For each possible definition of such an $f$ on $S$ we have at most $|S|+1$ possible ways to define $f$ on $A \setminus S$, and so at most $|S|+1$ possible ways to completely define $f$ on $A$. If there was an infinite uniformly $S$-supported sequence of finitely supported functions from $A$ to $A$, there should exist infinitely many finitely supported functions from $S$ to $A$ supported by the same finite set $S$. But this contradicts the fact that $A^{|S|}$ does not contain an infinite uniformly supported subset (this follows by applying finitely many times the result that $X \times X$ does not contain an infinite uniformly supported subset whenever $X$ does not contain an infinite uniformly supported subset). Analyzing the proofs of Corollary \[ti2\](6) and (7), we also conclude that $T_{fin}(A)^{A}_{fs}$ and $\wp_{fs}(A)^{A}_{fs}$ do not contain infinite uniformly supported subsets. We also have that $A$ is not FSM Tarski II infinite because $\wp_{fs}(A)$ contains no infinite uniformly supported subsets, and so every totally ordered subset (particularly via inclusion) of $\wp_{fs}(A)$ should be finite meaning that it should have a maximal element. Furthermore, we have that there is an equivariant bijection between $\wp_{fs}(A+A)$ and $\wp_{fs}(A)\times \wp_{fs}(A)$. Since $\wp_{fs}(A)$ does not contain an infinite uniformly supported subset, we have that $\wp_{fs}(A) \times \wp_{fs}(A)$ does not contain an infinite uniformly supported subset (the proof is quasi-identical to the one of Theorem \[ti1\](6) without taking count on the countability of the related infinite uniformly supported family). Therefore, any infinite totally ordered (via inclusion) uniformly supported family of $\wp_{fs}(A+A)$ should be finite containing a maximal element. There is an equivariant bijection between $\wp_{fs}(A)^{A}_{fs}$ and $\wp_{fs}(A \times A)$. Therefore any uniformly supported totally ordered subset of $\wp_{fs}(A \times A)$ should be finite containing a maximal element. Let $X$ be a finitely supported subset of an invariant set $Y$ such that $X$ does not contain an infinite uniformly supported subset. Then the set $\wp_{fin}(X)$ is not FSM Mostowski infinite. According to Lemma \[lem4\], $\wp_{fin}(X)$ does not contain an infinite uniformly supported subset. Thus, by Theorem \[Tti\], $\wp_{fin}(X)$ is not FSM Mostowski infinite. \[TTRR\] Let $X$ be a finitely supported subset of an invariant set $(Y, \cdot)$. 1. If $X$ is FSM Tarski I infinite, then $X$ is FSM Tarski III infinite. The converse does not hold. However if $X$ is FSM Tarski III infinite, then $\wp_{fs}(X)$ is FSM Tarski I infinite. 2. If $X$ is FSM Tarski III infinite, then $X$ is FSM Dedekind infinite. The converse does not hold. However if $X$ is FSM Dedekind infinite, then $\wp_{fs}(X)$ is FSM Tarski III infinite. 1\. We consider the case when $X$ has at least two elements (otherwise the theorem is trivial). Let $X$ be FSM Tarski I infinite. Then $|X \times X|= |X|$. Fix two elements $x_{1}, x_{2} \in X$ with $x_{1}\neq x_{2}$. We can define an injection $f:X \times \{0,1\} \to X \times X$ by $f(u)=\left\{ \begin{array}{ll} (x,x_{1}) & \text{for}\: u=(x,0)\\ (x,x_{2}) & \text{for}\: u=(x,1) \end{array}\right.$. Clearly, by checking the condition in Proposition \[2.18’\] and using Proposition \[p1\], we have that $f$ is supported by $supp(X) \cup supp(x_{1}) \cup supp(x_{2})$ (since $\{0,1\}$ is necessarily a trivial invariant set), and so $|X\times\{0,1\}|\leq |X \times X|$. Thus, $|X\times\{0,1\}| \leq |X|$. Obviously, there is an injection $i: X \to X\times\{0,1\}$ defined by $i(x)=(x,0)$ for all $x \in X$ which is supported by $supp(X)$. According to Lemma \[lem2\], we get $2|X|=|X \times \{0,1\}|=|X|$. Let us consider $X=\mathbb{N} \times A$. We make the remark that $|\mathbb{N}\times \mathbb{N}|=|\mathbb{N}|$ by considering the equivariant injection $h:\mathbb{N} \times \mathbb{N} \to \mathbb{N}$ defined by $h(m,n)=2^{m}3^{n}$ and using Lemma \[lem2\]. Similarly, $|\{0,1\}\times \mathbb{N}|=|\mathbb{N}|$ by considering the equivariant injection $h':\mathbb{N} \times \{0,1\}\to \mathbb{N}$ defined by $h'(n,0)=2^{n}$ and $h'(n,1)=3^{n}$ and using Lemma \[lem2\]. We have $2|X|=2|\mathbb{N}||A|=|\mathbb{N}||A|=|X|$. However, we prove that $|X \times X|\neq |X|$. Assume the contrary, and so we have $|\mathbb{N} \times (A \times A)|=|\mathbb{N} \times A \times \mathbb{N} \times A|=|\mathbb{N} \times A|$. Thus, there is a finitely supported injection $g: A \times A \to \mathbb{N} \times A$, and by Proposition \[pco2\] there is a finitely supported surjection $f:\mathbb{N} \times A \to A \times A$. Let us consider three different atoms $a,b,c \notin supp(f)$. There exists $(i,x) \in \mathbb{N} \times A$ such that $f(i,x)=(a,b)$. Since $(a\,b) \in Fix(supp(f))$ and $\mathbb{N}$ is trivial invariant set, we have $f(i,(a\,b)(x))=(a\,b)f(i,x)=(a\,b)(a,b)=((a\,b)(a),(a\,b)(b))=(b,a)$. We should have $x=a$ or $x=b$, otherwise $f$ is not a function. Assume without losing the generality that $x=a$, which means $f(i,a)=(a,b)$. Therefore $f(i,b)=f(i,(a\,b)(a))=(a\,b)f(i,a)=(a\,b)(a,b)=(b,a)$. Similarly, since $(a\,c),(b\,c) \in Fix(supp(f))$, we have $f(i,c)=f(i,(a\,c)(a))=(a\,c)f(i,a)=(a\,c)(a,b)=(c,b)$ and $f(i,b)=f(i,(b\,c)(c))=(b\,c)f(i,c)=(b\,c)(c,b)=(b,c)$. But $f(i,b)=(b,a)$ contradicting the functionality of $f$. Therefore, $X$ is FSM Tarski III infinite, but it is not FSM Tarski I infinite. Now, suppose that $X$ is FSM Tarski III infinite, which means $|\{0,1\}\times X|=|X|$. We define the mapping $\psi:\wp_{fs}(X) \times \wp_{fs}(X) \to \wp_{fs}(\{0,1\}\times X)$ by $f(U,V)=\{(0,x)\,|\,x \in U\} \cup \{(1,y)\,|\,y \in V\}$ for all $U,V \in \wp_{fs}(X)$. Clearly $\psi$ is well defined and bijective, and for each $\pi \in Fix(supp(X))$ we have $\psi(\pi \star U,\pi \star V)=\pi \star \psi(U,V)$ which means $\psi$ is finitely supported. Therefore, $|\wp_{fs}(X) \times \wp_{fs}(X)|=|\wp_{fs}(\{0,1\}\times X)|=|\wp_{fs}(X)|$. The last equality follows by applying twice Lemma \[lemlem\] (using the fact that there is a finitely supported surjection from $X$ onto $X\times\{0,1\}$ and a finitely supported surjection from $X\times\{0,1\}$ onto $X$, we obtain there is a finitely supported injection from $\wp_{fs}(X\times\{0,1\})$ into $\wp_{fs}(X)$, and a finitely supported injection from $\wp_{fs}(X)$ into $\wp_{fs}(X\times\{0,1\})$) and Lemma \[lem3\]. 2\. Let us assume that $X$ is FSM Tarski III infinite. Let us consider an element $y_{1}$ belonging to an invariant set (whose action is also denoted by $\cdot$) with $y_{1}\notin X$ (such an element can be, for example, a non-empty element in $\wp_{fs}(X) \setminus X$). Fix $y_{2} \in X$. One can define a mapping $f:X \cup \{y_{1}\} \to X \times \{0,1\}$ by $f(x)=\left\{ \begin{array}{ll} (x,0) & \text{for}\: x \in X\\ (y_{2}, 1) & \text{for}\: x=y_{1} \end{array}\right.$. Clearly $f$ is injective and it is supported by $S=supp(X) \cup supp(y_{1}) \cup supp(y_{2})$ because for all $\pi$ fixing $S$ pointwise we have $f(\pi \cdot x)=\pi \cdot f(x)$ for all $x \in X \cup \{y_{1}\}$. Therefore, $|X \cup \{y_{1}\}| \leq |X \times \{0,1\}|=|X|$, and so there is a finitely supported injection $g:X \cup \{y_{1}\} \to X$. The mapping $h:X \to X$ defined by $h(x)=g(x)$ is injective, supported by $supp(g) \cup supp(X)$, and $g(y_{1}) \in X \setminus h(X)$, which means $h$ is not surjective. It follows that $X$ is FSM Dedekind infinite. Let us consider $X=A \cup \mathbb{N}$. Since $A$ and $\mathbb{N}$ are disjoint, we have that $X$ is an invariant set (similarly as in Proposition \[p1\]). Clearly $X$ is FSM Dedekind infinite. Assume, by contradiction, that $|X|=2|X|$, that is $|A \cup \mathbb{N}|=|A+A+\mathbb{N}|=|(\{0,1\}\times A) \cup \mathbb{N}|$. Thus, there is a finitely supported injection $f:(\{0,1\}\times A) \cup \mathbb{N}\to A \cup \mathbb{N}$, and so there exists a finitely supported injection $f:(\{0,1\}\times A) \to A \cup \mathbb{N}$. We prove that whenever $\varphi:A \to A \cup \mathbb{N}$ is finitely supported and injective, for $a \notin supp(\varphi)$ we have $\varphi(a) \in A$. Assume, by contradiction, that there is $a \notin supp(\varphi)$ such that $\varphi(a)\in \mathbb{N}$. Since $supp(\varphi)$ is finite, there exists $b \notin supp(\varphi)$, $b \neq a$. Thus, $(a\,b) \in Fix(supp(\varphi))$, and so $\varphi(b)=\varphi((a\,b)(a))=(a\,b)\diamond \varphi(a)=\varphi(a)$ since $(\mathbb{N}, \diamond)$ is a trivial invariant set. This contradicts the injectivity of $\varphi$. We can consider the mappings $\varphi_{1},\varphi_{2}: A \to A \cup \mathbb{N}$ defined by $\varphi_{1}(a)=f(0,a)$ for all $a \in A$ and $\varphi_{2}(a)=f(1,a)$ for all $a \in A$, that are injective and supported by $supp(f)$. Therefore, $f(\{0\} \times A)=\varphi_{1}(A)$ contains at most finitely many element from $\mathbb{N}$, and $f(\{1\} \times A)=\varphi_{2}(A)$ also contains at most finitely many element from $\mathbb{N}$. Thus, $f$ is an injection from $(\{0,1\}\times A)$ to $A \cup Z$ where $Z$ is a finite subset of $\mathbb{N}$. It follows that $f(\{0\} \times A)$ contains an infinite subset of atoms $U$, and $f(\{1\} \times A)$ contains an infinite subset of atoms $V$. Since $f$ is injective, it follows that $U$ and $V$ are infinite disjoint subsets of $A$, which contradicts Proposition \[p111\] stating that $A$ is amorphous. Now, if $X$ is FSM Dedekind infinite, we have that there is a finitely supported injection $h$ from $X$ onto a finitely supported proper subset $Z$ of $X$. Consider an element $y_{1}$ belonging to an invariant set with $y_{1}\notin X$. We can define an injection $h': X \cup \{y_{1}\} \to X$ by taking $h'(x)=h(x)$ for all $x \in X$ and $h'(y_{1})=b$ with $b \in X \setminus Z$. Clearly $h'$ is supported by $supp(h) \cup supp(y_{1}) \cup supp(b)$. Since there also exists an $supp(X)$-supported injection from $X$ to $X \cup \{y_{1}\}$, according to Lemma \[lem3\], one can define a finitely supported bijection $\psi$ from $X$ to $X \cup \{y_{1}\}$. According to Lemma \[lemlem\] the mapping $g:\wp_{fs}(X \cup \{y_{1}\}) \to \wp_{fs}(X)$ defined by $g(V)=f^{-1}(V)$ for all $V \in \wp_{fs}(X \cup \{y_{1}\})$ is finitely supported and injective. Therefore, $2^{|X|}\geq 2^{|X|+1} =2\cdot 2^{|X|}$ which in the view of Lemma \[lem3\] leads to the conclusion that $\wp_{fs}(X)$ is FSM Tarski III infinite. \[ti4\] The following sets are FSM usual infinite, but they are not FSM Tarski I infinite, nor FSM Tarski III infinite. 1. The invariant set $A$. 2. The invariant set $\wp_{fs}(A)$. 3. The invariant sets $\wp_{fin}(A)$ and $\wp_{cofin}(A)$. 4. The set $\wp_{fin}(X)$ where $X$ is a finitely supported subset of an invariant set containing no infinite uniformly supported subset. The result follows directly because the related sets are not FSM Dedekind infinite, according to Theorem \[ti1\] and Corollary \[ti2\]. Let $X$ be an infinite finitely supported subset of an invariant set. Then $\wp_{fs}(\wp_{fs}(\wp_{fs}(X)))$ is FSM Tarski III infinite and, consequently, $\wp_{fs}(\wp_{fs}(\wp_{fs}(\wp_{fs}(X))))$ is FSM Tarski I infinite. Since $\wp_{fs}(\wp_{fs}(X))$ is FSM Dedekind infinite, as in the proof of Theorem \[TTRR\](2) one can prove $|\wp_{fs}(\wp_{fs}(X))|+1=|\wp_{fs}(\wp_{fs}(X))|$. The result now follows directly using arithmetic properties of FSM cardinalities proved above. In a future work we intend to prove an even stronger result claiming that $\wp_{fs}(\wp_{fs}(X))$ is FSM Tarski III infinite and, consequently, $\wp_{fs}(\wp_{fs}(\wp_{fs}(X)))$ is FSM Tarski I infinite, whenever $X$ is an infinite finitely supported subset of an invariant set. \[CAN1\] The sets $A^{\mathbb{N}}_{fs}$ and $\mathbb{N}^{A}_{fs}$ are FSM Tarski I infinite, and so they are also Tarski III infinite. There is an equivariant bijection $\psi$ between $(A^{\mathbb{N}})^{2}_{fs}$ and $A^{\mathbb{N} \times \{0,1\}}_{fs}$ that associates to each Cartesian pair $(f,g)$ of mappings from $\mathbb{N}$ to $A$ a mapping $h:\mathbb{N} \times\{0,1\} \to A$ defined as follows. $$h(u)=\left\{ \begin{array}{ll} f(n) & \text{if}\: u=(n,0) \\ \\ g(n) & \text{if}\: u=(n,1) \end{array}\right.,$$ The equivariance of $\psi$ follows from Proposition \[2.18’\] because if $\pi \in S_{A}$ we have $\psi(\pi \widetilde {\star} f, \pi \widetilde {\star} g)=h'$ where $h'(n,0)=(\pi \widetilde {\star} f)(n)=\pi(f(n))$ and $h'(n,1)=(\pi \widetilde {\star} g)(n)=\pi(g(n))$. Thus, $h'(u)=\pi(h(u))$ for all $u \in \mathbb{N} \times \{0,1\}$ which means $h'=\pi \widetilde {\star} h=\pi \widetilde {\star} \psi (f,g)$. There also exists an equivariant bijection $\varphi$ between $(\mathbb{N}^{A})^{2}_{fs}$ and $(\mathbb{N} \times \mathbb{N})^{A}_{fs}$ that associates to each Cartesian pair $(f,g)$ of mappings from $A$ to $\mathbb{N}$ a mapping $h: A \to \mathbb{N} \times \mathbb{N}$ defined by $h(a)=(f(a),g(a))$ for all $a \in A$. The equivariance of $\varphi$ follows from Proposition \[2.18’\] because if $\pi \in S_{A}$ we have $\varphi(\pi \widetilde {\star} f, \pi \widetilde {\star} g)=h'$ where $h'(a)=((\pi \widetilde {\star} f)(a), (\pi \widetilde {\star} g)(a))=(f(\pi^{-1}(a)), g(\pi^{-1}(a)))=h(\pi^{-1}(a))=(\pi \widetilde {\star}h)(a)$ for all $a \in A$, and so $h'=\pi \widetilde {\star} h=\pi \widetilde {\star} \varphi(f,g)$. Therefore $|(A^{\mathbb{N}})^{2}_{fs}|=|A^{\mathbb{N} \times \{0,1\}}_{fs}|= |A^{\mathbb{N}}_{fs}|$. Therefore, $|(\mathbb{N}^{A})^{2}_{fs}|=|(\mathbb{N} \times \mathbb{N})^{A}_{fs}|=|\mathbb{N}^{A}_{fs}|$ according to Proposition \[pco1\](3) and Lemma \[lem2\] (we used $|\mathbb{N} \times \mathbb{N}|=|\mathbb{N}|$). Let $X$ be a finitely supported subset of an invariant set $(Y, \cdot)$. If $\wp_{fs}(X)$ is FSM Tarski I infinite, then $\wp_{fs}(X)$ is FSM Tarski III infinite. The converse does not hold. The direct implication is a consequence of Theorem \[TTRR\](1). Thus, we focus on the proof of the invalidity of the reverse implication. Firstly we make the remark that whenever $U,V$ are finitely supported subsets of an invariant set with $U \cap V=\emptyset$, we have that there is a finitely supported (by $supp(U) \cup supp(V)$) bijection from $\wp_{fs}(U\cup V)$ into $\wp_{fs}(U) \times \wp_{fs}(V)$ that maps each $X \in \wp_{fs}(U\cup V)$ into the pair $(X \cap U, X \cap V)$. Analogously, whenever $B,C$ are invariant sets there is an equivariant bijection from $\wp_{fs}(B) \times \wp_{fs}(C)$ into $\wp_{fs}(B+C)$ that maps each pair $(B_{1},C_{1}) \in \wp_{fs}(B) \times \wp_{fs}(C)$ into the set $\{(0,b)\,|\,b \in B_{1}\} \cup \{(1,c)\,|\,c \in C_{1}\}$. This follows directly by verifying the conditions in Proposition \[2.18’\]. Let us consider the set $A \cup \mathbb{N}$ which is FSM Dedekind infinite. According to Theorem \[TTRR\](2), we have that $\wp_{fs}(A \cup \mathbb{N})$ is FSM Tarski III infinite. We prove that it is not FSM Tarski I infinite. Assume, by contradiction that $|\wp_{fs}(A \cup \mathbb{N}) \times \wp_{fs}(A \cup \mathbb{N})|=|\wp_{fs}(A \cup \mathbb{N})|$ which means $|\wp_{fs}(A + \mathbb{N}+A + \mathbb{N})|=|\wp_{fs}(A \cup \mathbb{N})|$, and so $|\wp_{fs}(A +A + \mathbb{N})|=|\wp_{fs}(A \cup \mathbb{N})|$. Thus, according to Proposition \[pco1\](4), there is a finitely supported injection from $\wp_{fs}(A + A)$ to $\wp_{fs}(A \cup \mathbb{N})$, which means there is a finitely supported injection from $\wp_{fs}(A) \times \wp_{fs}(A)$ to $ \wp_{fs}(A) \times \wp_{fs}(\mathbb{N})$, and so there is a finitely supported injection from $A \times A$ to $ \wp_{fs}(A) \times \wp_{fs}(\mathbb{N})$. According to Proposition \[pco2\], there should exist a finitely supported surjection $f: \wp_{fs}(A) \times \wp_{fs}(\mathbb{N}) \to A \times A$. Let us consider two atoms $a,b\notin supp(f)$ with $a \neq b$. It follows that $(a\, b) \in Fix(supp(f))$. Since $f$ is surjective, there exists $(X,M) \in \wp_{fs}(A) \times \wp_{fs}(\mathbb{N})$ such that $f(X,M)=(a,b)$. According to Proposition \[2.18’\] and because $\mathbb{N}$ is a trivial invariant set meaning that $(a\,b) \star M = M$, we have $f((a\,b) \star X,M)=f((a\,b) \otimes (X,M))=(a\,b) \otimes f(X,M)=(a\,b) \otimes (a,b)=((a\,b)(a), (a\,b)(b))=(b,a)$. Due to the functionality of $f$ we should have $((a\,b) \star X,M) \neq (X,M)$, which means $(a\,b) \star X \neq X$. We prove that if both $a,b \in supp(X)$, then $(a\,b)\star X=X$. Indeed, suppose $a,b \in supp(X)$. Since $X \in \wp_{fs}(A)$, from Proposition \[p111\] we have that $X$ is either finite or cofinite. If $X$ is finite, then $supp(X)=X$, and so $a,b \in X$. Therefore, $(a\,b) \star X=\{(a\,b)(x)\,|\,x \in X\}=\{(a\,b)(a)\} \cup \{(a\,b)(b)\} \cup \{(a\,b)(c)\,|\,c \in X \setminus\{a,b\}\}=\{b\} \cup \{a\} \cup (X \setminus \{a,b\})=X$. Now, if $X$ is cofinite, then $supp(X)=A \setminus X$, and so $a,b \in A \setminus X$. Since $a,b \notin X$, we have $a,b \neq x$ for all $x \in X$, which means $(a\,b)(x)=x$ for all $x \in X$, and again $(a\,b)\star X=X$. Thus, one of $a$ or $b$ does not belong to $supp(X)$. Assume $b \notin supp(X)$. Let us consider $c\neq a,b$, $c \notin supp(f)$, $c \notin supp(X)$. Then $(b\, c) \in Fix(supp(X))$, and so $(b\,c)\star X=X$. Moreover, $(b\, c) \in Fix(supp(f))$, and by Proposition \[2.18’\] we have $(a,b)=f(X,M)=f((b\,c) \star X,M)=f((b\,c)\otimes (X,M))=(b\,c) \otimes f(X,M)=(b\,c) \otimes (a,b)=(a,c)$ which is a contradiction because $b\neq c$. Let $X$ be a finitely supported subset of an invariant set $(Y, \cdot)$. If $X$ is FSM Tarski III infinite, then there exists a finitely supported bijection $g:\mathbb{N} \times X \to X$. The reverse implication is also valid. By hypothesis, there is a finitely supported bijection $\varphi:\{0,1\} \times X \to X$. Let us consider the mappings $f_{1},f_{2}: X \to X$ defined by $f_{1}(x)=\varphi(0,x)$ for all $x \in X$ and $f_{2}(x)=\varphi(1,x)$ for all $x \in X$, that are injective and supported by $supp(\varphi)$ according to Proposition \[2.18’\]. Since $\varphi$ is injective we also have $Im(f_{1}) \cap Im(f_{2})=\emptyset$, and because $\varphi$ is surjective we get $Im(f_{1}) \cup Im(f_{2})=X$. We prove by induction that the $n$-times auto-composition of $f_{2}$, denoted by $f_{2}^{n}$, is supported by $supp(f_{2})$ for all $n\in \mathbb{N}$. For $n=1$ this is obvious. So assume that $f_{2}^{n-1}$ is supported by $supp(f_{2})$. By Proposition \[2.18’\] we must have $f_{2}^{n-1}(\sigma \cdot x)=\sigma \cdot f_{2}^{n-1}(x)$ for all $\sigma \in Fix(supp(f_{2}))$ and $x \in X$. Let us fix $\pi \in Fix(supp(f_{2}))$. According to Proposition \[2.18’\], we have $f_{2}^{n}(\pi \cdot x)=f_{2}(f_{2}^{n-1}(\pi \cdot x))=f_{2}(\pi\cdot f_{2}^{n-1}(x))=\pi \cdot f_{2}(f_{2}^{n-1}(x))=\pi \cdot f_{2}^{n}(x)$ for all $x \in X$, and so $f_{2}^{n}$ is finitely supported from Proposition \[2.18’\]. Define $f:\mathbb{N} \times X \to X$ by $f((n,x))=f_{2}^{n}(f_{1}(x))$. Let $\pi \in Fix(supp(f_{1}) \cup supp(f_{2}))$. According to Proposition \[2.18’\] and because $(\mathbb{N}, \diamond)$ is a trivial invariant set we get $f(\pi \otimes (n,x))=f((n, \pi \cdot x))=f_{2}^{n}(f_{1}(\pi \cdot x))=f_{2}^{n}(\pi \cdot f_{1}(x))=\pi \cdot f_{2}^{n}(f_{1}(x))=\pi \cdot f((n,x))$ for all $(n,x) \in \mathbb{N} \times X$, which means $f$ is supported by $supp(f_{1}) \cup supp(f_{2})$. We prove the injectivity of $f$. Assume $f((n,x))=f((m,y))$ which means $f_{2}^{n}(f_{1}(x))=f_{2}^{m}(f_{1}(y))$. If $n>m$ this leads to $f_{2}^{n-m}(f_{1}(x))=f_{1}(y)$ (since $f_{2}$ is injective) which is in contradiction with the relation $Im(f_{1}) \cap Im(f_{2})=\emptyset$. Analogously we cannot have $n<m$. Thus, $n=m$ which leads to $f_{1}(x)=f_{1}(y)$, and so $x=y$ due to the injectivity of $f_{1}$. Therefore, $f$ is injective. Since we obviously have a finitely supported injection from $X$ into $\mathbb{N} \times X$ (e.g $x \mapsto (0,x)$ which is supported by $supp(X)$), in the view of Lemma \[lem2\] we can find a finitely supported bijection between $X$ and $\mathbb{N} \times X$. The reverse implication is almost trivial. There is a finitely supported injection from $\{0,1\} \times X$ into $\mathbb{N} \times X$. If there is a finitely supported injection from $\mathbb{N} \times X$ into $X$, then there is a finitely supported injection from $\{0,1\} \times X$ into $X$. The desired result follows from Lemma \[lem2\]. Countability {#countable} ============ Let $Y$ be a finitely supported subset of an invariant set $X$. Then $Y$ is *countable in FSM (or FSM countable)* if there exists a finitely supported onto mapping $f: \mathbb{N} \to Y$. Let $Y$ be a finitely supported countable subset of an invariant set $(X, \cdot)$. Then $Y$ is uniformly supported. There exists a finitely supported onto mapping $f: \mathbb{N} \to Y$. Thus, for each arbitrary $y\in Y$, there exists $n \in \mathbb{N}$ such that $f(n)=y$. According to Proposition \[2.18’\], for each $\pi \in Fix(supp(f))$ we have $\pi \cdot y=\pi \cdot f(n)=f(\pi \diamond n)=f(n)=y$, where $\diamond$ is the necessarily trivial action on $\mathbb{N}$. Thus, $Y$ is uniformly supported by $supp(f)$. Let $Y$ be a finitely supported subset of an invariant set $X$. Then $Y$ is countable in FSM if and only if there exists a finitely supported one-to-one mapping $g: Y \to \mathbb{N}$. Suppose that $Y$ is countable in FSM. Then there exists a finitely supported onto mapping $f: \mathbb{N} \to Y$. We define $g: Y \to \mathbb{N}$ by $g(y)=min[f^{-1}(\{y\})]$, for all $y \in Y$. According to Proposition \[2.18’\], $g$ is supported by $supp(f) \cup supp(Y)$. Obviously, $g$ is one-to-one. Conversely, if there exists a finitely supported one-to-one mapping $g: Y \to \mathbb{N}$, then $g(Y)$ is supported is equivariant as a subset of the trivial invariant set $\mathbb{N}$. Thus, there exists a finitely supported bijection $g: Y \to g(Y)$, where $g(Y) \subseteq \mathbb{N}$. We define $f: \mathbb{N} \to Y$ by $$f(n)=\left\{ \begin{array}{ll} g^{-1}(n) & \text{if}\: n \in g(Y) \\ \\ t & \text{if}\: n \in \mathbb{N} \setminus g(Y) \end{array}\right.,$$ where $t$ is a fixed element of $Y$. According to Proposition \[2.18’\], we have that $f$ is supported by $supp(g) \cup supp(Y) \cup supp(t)$. Moreover, $f$ is onto. \[cou\] Let $Y$ be an infinite, finitely supported, countable subset of an invariant set $X$. Then there exists a finitely supported bijective mapping $g: Y \to \mathbb{N}$. First we prove that for any infinite subset $B$ of $\mathbb{N}$, there is an injection from $\mathbb{N}$ into $B$. Fix such a $B$. It follows that $B$ is well ordered. Define $f:\mathbb{N} \to B$ by: $f(1)=min(B)$, $f(2)=min(B \setminus f(1))$, and recursively $f(m)= min(B \setminus \{f(1),f(2),...,f(m-1)\})$ for all $m \in \mathbb{N}$ (since $B$ is infinite). Since $\mathbb{N}$ is well ordered, choice is not involved. Obviously since both $B$ and $\mathbb{N}$ are trivial invariant sets, we have that $f$ is equivariant. Since $B$ is a subset of $\mathbb{N}$ we also have an equivariant injective mapping $h:B \to \mathbb{N}$. According to Lemma \[lem1\], there is an equivariant bijection between $B$ and $\mathbb{N}$ (we can even prove that $f$ is bijective). Since $Y$ is countable, there exists a finitely supported one-to-one mapping $u: Y \to \mathbb{N}$. Thus, the mapping $u: Y \to u(Y)$ is finitely supported and bijective. Since $u(Y) \subseteq N$, we have that there is an equivariant bijection $v$ between $u(Y)$ and $\mathbb{N}$, and so there exists a finitely supported bijective mapping $g: Y \to \mathbb{N}$ defined by $g=v \circ u$. From [@book] we know that the (in)consistency of the choice principle $\textbf{CC(fin)}$ in FSM is an open problem, meaning that we do not know whether this principle is consistent or not in respect of the FSM axioms. A relationship between countable union principles and countable choice principles is presented in ZF in [@herrlich]. Below we prove that such a relationship is preserved in FSM. 1. The Countable Choice Principle for finite sets in FSM **CC(fin)** has the form “Given any invariant set $X$, and any countable family $\mathcal{F}=(X_{n})_{n}$ of finite subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported, there exists a finitely supported choice function on $\mathcal{F}$." 2. The Countable Union Theorem for finite sets in FSM, **CUT(fin)**, has the form “Given any invariant set $X$ and any countable family $\mathcal{F}=(X_{n})_{n}$ of finite subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported, then there exists a finitely supported onto mapping $f: \mathbb{N} \to \underset{n}{\cup}X_{n}$" 3. The Countable Union Theorem for $k$-element sets in FSM, **CUT(k)**, has the form “Given any invariant set $X$ and any countable family $\mathcal{F}=(X_{n})_{n}$ of $k$-element subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported, then there exists a finitely supported onto mapping $f: \mathbb{N} \to \underset{n}{\cup}X_{n}$" 4. The Countable Choice Principle for sets of $k$-element sets in FSM, **CC(k)** has the form “Given any invariant set $X$ and any countable family $\mathcal{F}=(X_{n})_{n}$ of $k$-element subsets of $X$ in FSM such that the mapping $n\mapsto X_{n}$ is finitely supported, there exists a finitely supported choice function on $\mathcal{F}$." In FSM, the following equivalences hold. 1. **CUT(fin)** $\Leftrightarrow$ **CC(fin)**; 2. **CUT(2)** $\Leftrightarrow$ **CC(2)**; 3. **CUT(n)** $\Leftrightarrow$ **CC(i)** for all $i \leq n$. 1\. Let us assume that **CUT(fin)** is valid in FSM. We consider the finitely supported countable family $\mathcal{F}=(X_{n})_{n}$ in FSM, where each $X_{n}$ is a non-empty finite subset of an invariant set $X$ in FSM. From **CUT(fin)**, there exists a finitely supported onto mapping $f: \mathbb{N} \to \underset{n}{\cup}X_{n}$. Since $f$ is onto and each $X_{n}$ is non-empty, we have that $f^{-1}(X_{n})$ is a non-empty subset of $\mathbb{N}$ for each $n \in \mathbb{N}$. Consider the function $g: \mathcal{F} \to \cup \mathcal{F}$, defined by $g(X_{n})=f(min[f^{-1}(X_{n})])$. We claim that $supp(f) \cup supp(n \mapsto X_{n})$ supports $g$. Let $\pi \in Fix(supp(f) \cup supp(n \mapsto X_{n}))$. According to Proposition \[2.18’\], and because $\mathbb{N}$ is a trivial invariant set and each element $X_{n}$ is supported by $supp(n \mapsto X_{n})$, we have $\pi \cdot g(X_{n})=\pi \cdot f(min[f^{-1}(X_{n})])= f(\pi \diamond min[f^{-1}(X_{n})])=f(min[f^{-1}(X_{n})])=g(X_{n})=g(\pi \star X_{n})$, where by $\star$ we denoted the $S_{A}$-action on $\mathcal{F}$, by $\cdot$ we denoted the $S_{A}$-action on $\cup \mathcal{F}$ and by $\diamond$ we denoted the trivial action on $\mathbb{N}$. Therefore, $g$ is finitely supported. Moreover, $g(X_{n}) \in X_{n}$, and so $g$ is a choice function on $\mathcal{F}$. Conversely, let $\mathcal{F}=(X_{n})_{n}$ be a countable family of finite subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported. Thus, each $X_{n}$ is supported by the same set $S=supp(n \mapsto X_{n})$. Since each $X_{n}$ is finite (and the support of a finite set coincides with the union of the supports of its elements), as in the proof of Lemma \[lem4\], we have that $Y=\underset{n \in \mathbb{N}}\cup X_{n}$ is uniformly supported by $S$. Moreover, the countable sequence $(Y_{n})_{n \in \mathbb{N}}$ defined by $Y_{n}=X_{n} \setminus \underset{m<n}\cup X_{m}$ is a uniformly supported (by $S$) sequence of pairwise disjoint uniformly supported sets with $Y=\underset{n \in \mathbb{N}}\cup Y_{n}$. Consider the infinite family $M \subseteq \mathbb{N}$ such that all the terms of $(Y_{n})_{n \in M}$ are non-empty. For each $n \in M$, the set $T_{n}$ of total orders on $Y_{n}$ is finite, non-empty, and uniformly supported by $S$. Thus, by applying **CC(fin)** to $(T_{n})_{n \in M}$, there is a choice function $f$ on $(T_{n})_{n \in M}$ which is also supported by $S$. Furthermore, $f(T_{n})$ is supported by $supp(f) \cup supp(T_{n})=S$ for all $n \in M$. One can define a uniformly supported (by $S$) total order relation on $Y$ (which is also a well order relation on $Y$) as follows $x \leq y$ if and only if $\left\{ \begin{array}{ll} x \in Y_{n} \; \text{and}\; y \in Y_{m} \;\text{with}\; n<m \\ \text{or}\\ x,y \in Y_{n}\; \text{and}\; xf(T_{n})y \end{array}\right.$. Clearly, if $Y$ is infinite, then there is an $S$-supported order isomorphism between $(Y, \leq)$ and $M$ with the natural order, which means, in the view of Proposition \[cou\], that $Y$ is countable. 2\. As in the above item $\textbf{CUT(2)} \Rightarrow \textbf{CC(2)}$. For proving $\textbf{CC(2)} \Rightarrow \textbf{CUT(2)}$, let $\mathcal{F}=(X_{n})_{n}$ be a countable family of 2-element subsets of $X$ such that the mapping $n\mapsto X_{n}$ is finitely supported. According to **CC(2)** we have that there exists a finitely supported choice function $g$ on $(X_{n})_{n}$. Let $x_{n}=g(X_{n}) \in X_{n}$. As in the above item, we have that $supp(n \mapsto X_{n})$ supports $x_{n}$ for all $n \in \mathbb{N}$. For each $n$, let $y_{n}$ be the unique element of $X_{n}\setminus \{x_{n}\}$. Since for any $n$ both $x_{n}$ and $X_{n}$ are supported by the same set $supp(n \mapsto X_{n})$, it follows that $y_{n}$ is also supported by $supp(n \mapsto X_{n})$ for all $n \in \mathbb{N}$. Define $f: \mathbb{N} \to \underset{n}{\cup}X_{n}$ by $ f(n)=\left\{ \begin{array}{ll} x_{\frac{n}{2}} & \text{if}\: n \;\text{is even}\\ \\ y_{\frac{n-1}{2}} & \text{if}\: n \;\text {is odd} \end{array}\right.$. We can equivalently describe $f$ as being defined by $f(2k)=x_{k}$ and $f(2k+1)=y_{k}$. Clearly, $f$ is onto. Furthermore, because all $x_{n}$ and all $y_{n}$ are uniformly supported by $supp(n \mapsto X_{n})$, we have that $f(n)=\pi \cdot f(n)$, for all $\pi \in Fix(supp(n \mapsto X_{n}))$ and all $n \in \mathbb{N}$. Thus, according to Proposition \[2.18’\], we obtain that $f$ is also supported by $supp(n \mapsto X_{n})$, and so $\underset{n}{\cup}X_{n}$ is FSM countable. 3\. As in the proof of the first item. We can easily remark that under $\textbf{CC(fin)}$ a finitely supported subset $X$ of an invariant set is FSM Dedekind infinite if and only if $\wp_{fin}(X)$ is FSM Dedekind infinite. Let $Y$ be a finitely supported countable subset of an invariant set $X$. Then the set $\underset{n \in \mathbb{N}}{\cup}Y^{n}$ is countable, where $Y^{n}$ is defined as the n-time Cartesian product of $Y$. Since $Y$ is countable, we can order it as a sequence $Y=\{x_{1}, \ldots, x_{n}, \ldots\}$. The other sets of form $Y^{k}$ are *uniquely* represented in respect of the previous enumeration of the elements of $Y$. Since $Y$ is finitely supported and countable, all the elements of $Y$ are supported by the same set $S$ of atoms. Thus, in the view of Proposition \[p1\], for each $k \in \mathbb{N}$, all the elements of $Y^{k}$ are supported by $S$. Fix $n \in \mathbb{N}$. On $Y^{n}$ define the S-supported strict well order relation $\sqsubset$ by: $(x_{i_{1}}, x_{i_{2}}, \ldots, x_{i_{n}}) \sqsubset (x_{j_{1}}, x_{j_{2}}, \ldots, x_{j_{n}})$ if and only if $\left\{ \begin{array}{ll} i_{1} < j_{1} \\ \text{or}\\ i_{1}=j_{1}\; \text{and}\; i_{2}<j_{2}\\ \text{or}\\ \ldots \\ \text{or}\\ i_{1}=j_{1}, \ldots, i_{n-1}=j_{n-1}\; \text{and}\; i_{n}<j_{n} \end{array}\right.$. Now, define an $S$-supported strict well order relation $\prec$ on $\underset{n \in \mathbb{N}}{\cup}Y^{n}$ by $u \prec v$ if and only if $\left\{ \begin{array}{ll} u \in Y^{n} \; \text{and}\; v \in Y^{m} \;\text{with}\; n<m \\ \text{or}\\ u,v \in Y_{n}\; \text{and}\; u \sqsubset v \end{array}\right.$. Therefore, there exists an $S$-supported order isomorphism between $(\underset{n \in \mathbb{N}}{\cup}Y^{n}, \prec)$ and $(\mathbb{N}, <)$. Conclusion ========== It is known that, when an infinite family of elements having no internal structure is considered by weakening some axioms of the ZF set theory, the results in ZF may lose their validity. According to Theorem 5.4 in [@hal], multiple choice principle and Kurepa’s antichain principle are both equivalent to the axiom of choice in ZF. However, in Theorem 9.2 of [@jech] it is proved that multiple choice principle is valid in the Second Fraenkel Model, while the axiom of choice fails in this model. Furthermore, Kurepa’s maximal antichain principle is valid in the Basic Fraenkel Model, while multiple choice principle fails in this model. This means that the following two statements (that are valid in ZF) *‘Kurepa’s principle implies axiom of choice’* and *‘Multiple choice principle implies axiom of choice’* fail in Zermelo Fraenkel set theory with atoms. FSM is related to set theory with atoms, however in our approach $A$ is considered as a ZF set (without being necessary to modify the axioms of foundation or of extensionality), and invariant sets are defined as sets with group actions. Additionally, FSM involves an axiom of finite support which states that only atomic finitely supported structures (under a canonical hierarchical set-theoretical construction) are allowed in the theory. Therefore, there is indeed a similarity between the development of permutation models of set theory with atoms and FSM, but this framework is developed over the standard ZF in the form ‘usual sets together with actions of permutation groups’ without being necessary to consider an alternative set theory. The goal of this paper is to answer to a natural question whether the theorems involving the usual/non-atomic ZF sets remain valid in the framework of atomic sets with finite supports modulo canonical permutation actions. It is already known that there exist results that are consistent with ZF, but the are invalid when replacing ‘non-atomic structure’ with ‘atomic finitely supported structure’. The ZF results are not valid in FSM unless we are able to reformulate them with respect to the finite support requirement. The proofs of the FSM results should not brake the principle that any structure has to be finitely supported, which means that the related proofs should be *internally consistent in FSM* and not retrieved from ZF. The methodology for moving from ZF into FSM is based on the formalization of FSM into higher order logic (and this is not a simple task due to some important limitations) or on the hierarchical construction of supports using the $S$-finite support reasoning that actually represents an hierarchical method for defining the support of a structure using the supports of the sub-structures of the related structure. Since any structure has to be finitely supported in FSM, specific results (that are not derived from ZF) can also be obtained. In this paper we study infinite cardinalities of finitely supported structures. The preorder relation $\leq$ on FSM cardinalities defined by involving finitely supported injective mappings is antisymmetric, but not total. The preorder relation $\leq^{*}$ on FSM cardinalities defined by involving finitely supported surjective mappings is not antisymmetric, nor total. Thus, Cantor-Schr[ö]{}der-Bernstein theorem (in which cardinalities are ordered by involving finitely supported injective mappings) is consistent with the finite support requirement of FSM. However, the dual of Cantor-Schr[ö]{}der-Bernstein theorem (in which cardinalities are ordered by involving finitely supported surjective mappings) is not valid for finitely supported structures. Several other specific properties of cardinalities are presented in Theorem \[cardord1\]. The idea of presenting various approaches regarding ‘infinite’ belongs to Tarski who formulates several definitions of infinite in [@tarski24]. The independence of these definitions was later proved in set theory with atoms in [@levy1]. Such independence results can be transferred into classical ZF set theory by employing Jech-Sochor’s embedding theorem stating that permutation models of set theory with atoms can be embedded into symmetric models of ZF, and so a statement which holds in a given permutation model of set theory with atoms and whose validity depend only on a certain fragment of that model, also holds in some well-founded model of ZF. In this paper we reformulate the definitions of (in)finiteness from [@tarski24] internally into FSM, in terms of finitely supported structures. The related definitions for ‘FSM infinite’ are introduced in Section \[chap9\]. We particularly mention FSM usual infinite, FSM Tarski (of three types) infinite, FSM Dedekind infinite, FSM Mostowski infinite, FSM Kuratowski infinite, or FSM ascending infinite. We were able to establish comparison results between them and to present relevant examples of FSM sets that satisfy certain specific infinity properties. These comparison results are proved internally in FSM, by employing only finitely supported constructions. Some of the results are obtained by using the classical translation technique from ZF into FSM involving the $S$-finite support principle, while many other properties (especially those revealing uniform supports) are specific to FSM. We also provide connections with FSM (uniformly) amorphous sets. We particularly have focused on the notion of FSM Dedekind infinity, and we proved a full characterization of FSM Dedekind infinite sets. For example, we were able to prove that $T_{fin}(A)$, $\wp_{fin}(\wp_{fs}(A))$, $A^{A}_{fs}$, $\wp_{fin}(A^{A}_{fs})$, $(A^{n})^{A}_{fs}$ (for a fixed $n \in \mathbb{N}$), $T_{fin}(A)^{A}_{fs}$ and $\wp_{fs}(A)^{A}_{fs}$ are not FSM Dedekind infinite (nor FSM Mostowski infinite), while $\wp_{fs}(\wp_{fin}(A))$ and $T^{\delta}_{fin}(A)$ are FSM Dedekind infinite. The notion of ‘countability’ is described in FSM in Section \[countable\], where we present connections between countable choice principles and countable union theorems within finitely supported sets. In Figure \[fig:1\] we point out some of the relationships between the FSM definitions of infinite. The ‘red arrows’ symbolize *strict* implications (of from $p$ implies $q$, but $q$ does not imply $p$), while ‘black arrows’ symbolize implications for which we have not proved yet if they may be strict or not (analyze this in respect of Remark \[remarema\]). Blue arrows represent equivalences. \[node distance = 1.1cm, auto,font=, every node/.style=[node distance=2.1cm]{}, comment/.style=[rectangle, inner sep= 4pt, text width=3cm, node distance=0.25cm, font=]{}, force/.style=[rectangle, draw, fill=black!10, inner sep=4pt, text width=3cm, text badly centered, minimum height=1.0cm, font=****]{}\] (TI) [X is FSM Tarski I infinite]{}; (TIII) [X is FSM Tarski III infinite]{}; (PTI) [The FSM powerset of X is FSM Tarski I infinite]{}; (D inf) [X is FSM Dedekind infinite]{}; (PTIII) [The FSM powerset of X is FSM Tarski III infinite]{}; (NTIII) [There is a finitely supported bijection between X and the Cartesian product of the set of positive integers and X]{}; (M inf) [X is FSM Mostowski infinite]{}; (PD inf) [The finite powerset of X is FSM Dedekind infinite]{}; (PDD inf) [The finite powerset of X contains an infinite uniformly supported subset]{}; (A inf) [X is FSM ascending infinite]{}; (cont) [X contains an infinite uniformly supported subset]{}; (TII) [X is FSM Tarski II infinite]{}; (nua) [X is FSM non uniformly amorphous]{}; (usual) [X is FSM usual infinite]{}; (n-a) [X is FSM non amorphous]{}; (usual) [X is FSM usual infinite]{}; (ci) [X is FSM covering infinite]{}; (PPI) [The FSM powerset of the finite powerset of X is FSM Dedekind infinite]{}; (PAI) [The finite powerset of X is FSM ascending infinite]{}; (PPPP) [The powerset of X is FSM Dedekind infinite]{}; (PPPP’) [There exists a finitely supported surjection from X onto the set of positive integers]{}; (TI) edge (TIII) (PTI) edge (PTIII) (TIII) edge (D inf) (D inf) edge (A inf) (M inf) edge (TII) (TII) edge (usual) (cont) edge (usual) (n-a) edge (usual) (nua) edge (n-a) (D inf) edge (PPPP) (PPPP) edge (n-a) (D inf) edge (PTIII) (PD inf) edge (PPPP); (TIII) edge (PTI) (D inf) edge (M inf) (D inf) edge (PD inf) (PD inf) edge (PDD inf) (A inf) edge (TII) (M inf) edge (cont) (PD inf) edge (nua) (M inf) edge (nua) (PTIII) edge (A inf); (TIII) edge (NTIII) (NTIII) edge (TIII) (PDD inf) edge (cont) (cont) edge (PDD inf) (ci) edge (usual) (usual) edge (ci) (PPPP) edge (A inf) (A inf) edge (PPPP) (usual) edge (PPI) (PPI) edge (usual) (PAI) edge (ci) (ci) edge (PAI) (PPPP) edge (PPPP’) (PPPP’) edge (PPPP); [|l|l|l|l|l|l|l|l|p[0.5cm]{}|]{} Set & Tarski I inf & Tarski III inf & Ded. inf & Most. inf & Asc. inf & Tarski II inf & Non-amorph.\ $A$ & No & No & No & No & No & No & No\ $A+A$ & No & No & No & No & No & No & Yes\ $A \times A$ & No & No & No & No & No & No & Yes\ $\wp_{fin}(A)$ & No & No & No & No & Yes & Yes & Yes\ $T_{fin}(A)$ & No & No & No & No & Yes & Yes & Yes\ $\wp_{fs}(A)$ & No & No & No & No & Yes & Yes & Yes\ $\wp_{fin}(\wp_{fs}(A))$ & No & No & No & No & Yes & Yes & Yes\ $A^{A}_{fs}$ & No & No & No & No & Yes & Yes & Yes\ $T_{fin}(A)^{A}_{fs}$ & No & No & No & No & Yes & Yes & Yes\ $\wp_{fs}(A)^{A}_{fs}$ & No & No & No & No & Yes & Yes & Yes\ $A \cup \mathbb{N}$ & No & No & Yes & Yes & Yes & Yes & Yes\ $A \times \mathbb{N}$ & No & Yes & Yes & Yes & Yes & Yes & Yes\ $\wp_{fs}(A \cup \mathbb{N})$ & No & Yes & Yes & Yes & Yes & Yes & Yes\ $\wp_{fs}(\wp_{fs}(A))$ & ? & Yes & Yes & Yes & Yes & Yes & Yes\ $A^{\mathbb{N}}_{fs}$ & Yes & Yes & Yes & Yes & Yes & Yes & Yes\ $\mathbb{N}^{A}_{fs}$ & Yes & Yes & Yes & Yes & Yes & Yes & Yes\ In this final table we present the forms of infinity satisfied by the classical FSM sets. [99]{} Alexandru, A., Ciobanu, G.: Finitely Supported Mathematics: An Introduction, Springer (2016). Bojanczyk, M., Klin, B., Lasota, S.: Automata with group actions. In 26th Symposium on Logic in Computer Science, pp. 355–364. IEEE Computer Society Press (2011). Halbeisen, L.: Combinatorial Set Theory: With a Gentle Introduction to Forcing, 2nd edition, Springer (2017). Herrlich, H.: Axiom of Choice, Lecture Notes in Mathematics, Springer (2006). Jech, T.J.: The Axiom of Choice, Studies in Logic and the Foundations of Mathematics, North-Holland (1973). Levy, A.: The independence of various definitions of finiteness, Fundamenta Mathematicae, 46, 1-13 (1958). Pitts, A.M.: Nominal Sets Names and Symmetry in Computer Science, Cambridge University Press (2013). Tarski, A.: Sur les ensembles finis, Fundamenta Mathematicae, 6, 45-95 (1924).
{ "pile_set_name": "ArXiv" }
--- abstract: | A basic problem in information theory is the following: Let ${{\bf P}}= ({{\bf X}}, {{\bf Y}})$ be an arbitrary distribution where the marginals ${{\bf X}}$ and ${{\bf Y}}$ are (potentially) correlated. Let Alice and Bob be two players where Alice gets samples $\{x_i\}_{i \ge 1}$ and Bob gets samples $\{y_i\}_{i \ge 1}$ and for all $i$, $(x_i, y_i) \sim {{\bf P}}$. What joint distributions ${{\bf Q}}$ can be simulated by Alice and Bob without any interaction? Classical works in information theory by G[á]{}cs-K[ö]{}rner and Wyner answer this question when at least one of ${{\bf P}}$ or ${{\bf Q}}$ is the distribution on $\{0,1\} \times \{0,1\}$ where each marginal is unbiased and identical. However, other than this special case, the answer to this question is understood in very few cases. Recently, Ghazi, Kamath and Sudan showed that this problem is decidable for ${{\bf Q}}$ supported on $\{0,1\} \times \{0,1\}$. We extend their result to ${{\bf Q}}$ supported on any finite alphabet. We rely on recent results in Gaussian geometry (by the authors) as well as a new *smoothing argument* inspired by the method of *boosting* from learning theory and potential function arguments from complexity theory and additive combinatorics. author: - | Anindya De[^1]\ Northwestern University\ [[email protected]]{} - | Elchanan Mossel[^2]\ MIT\ [[email protected]]{} - | Joe Neeman\ UT Austin\ [[email protected]]{} bibliography: - 'allrefs.bib' title: Non interactive simulation of correlated distributions is decidable --- Introduction ============ The starting point of this paper is a rather basic problem in information theory and communication complexity, known as the problem of *non-interactive simulation of joint distributions*: Consider two non-communicating players Alice and Bob. Suppose that we give Alice and Bob the sequences $\{{{\bf X}}_1\}_{i=1}^\infty$ and $\{{{\bf Y}}_i\}_{i=1}^\infty$ respectively, where the pairs $({{\bf X}}_i, {{\bf Y}}_i)$ are independently drawn from some joint distribution ${{\bf P}}$. Without communicating with each other, which joint distributions ${{\bf Q}}$ can Alice and Bob jointly simulate? To state the problem more precisely, suppose that ${{\bf P}}$ is a distribution on $\mathcal{Z} \times \mathcal{Z}$ and that ${{\bf Q}}$ is a distribution on $\mathcal{W} \times \mathcal{W}$. A *non-interactive strategy* for Alice and Bob simply denotes a triple $(n, f, g)$ such that $f, g: \mathcal{Z}^n \to \mathcal{W}$, and for which $(f({{\bf X}}^n), g({{\bf Y}}^n))$ has distribution ${{\bf Q}}$ whenever $({{\bf X}}_i, {{\bf Y}}_i)$ are drawn independently from ${{\bf P}}$ (here, ${{\bf X}}^n$ denotes ${{\bf X}}_1, \dots, {{\bf X}}_n$). The main question that we consider here is whether a non-interactive strategy exists for a given input distribution ${{\bf P}}$ and a given target distribution ${{\bf Q}}$. Note that not every pair of input and target distributions admits a non-interactive strategy. The most obvious example of this is the case where the two coordinates of ${{\bf P}}$ are independent; in this case, one can obviously only simulate distributions ${{\bf Q}}$ whose coordinates are also independent. Witsenhausen [@witsenhausen] introduced the problem of non-interactive simulation, and he studied the case where ${{\bf Q}}$ is a Gaussian measure on $\mathbb{R}^2$. In this case, he showed that ${{\bf Q}}$ can be approximately simulated by ${{\bf P}}$ if and only if the absolute value of the correlation between the components of ${{\bf Q}}$ is at most the so-called “maximal correlation coefficient” (which we will define later) of ${{\bf P}}$. In this case, Witsenhausen showed that for any $\delta>0$, Alice and Bob can simulate ${{\bf Q}}$ up to error $\delta$ with $n = {\mathrm{poly}}(|\mathcal{Z}|, \log (1/\delta))$. Further, he gave an explicit algorithm to compute $f$ and $g$ in time $\mathsf{poly}(n)$. Various other questions of this flavor have been explored in information theory. We discuss two examples here. Let us use $\mathsf{Eq}$ to denote the distribution supported on $\{0,1\} \times \{0,1\}$ where (i) both marginals are unbiased Bernoullis and (ii) both marginals are identical with probability $1$. 1. In their seminal paper, G[á]{}cs and K[ö]{}rner [@gacs1973common] studied non-interactive simulation in the case ${{\bf Q}}= \mathsf{Eq}$. In this case, they obtained a simple and complete characterization of all ${{\bf P}}$ such that it is possible to non-interactively simulate ${{\bf Q}}$ from ${{\bf P}}$. They also studied the *simulation capacity*: roughly, how many samples from ${{\bf P}}$ are needed to produce each sample from ${{\bf Q}}$? They showed that the simulation capacity is equal to another quantity, which is now known as the *G[á]{}cs-K[ö]{}rner common information of ${{\bf P}}$*. 2. Around the same time, Wyner [@wyner1975] considered the complementary problem where ${{\bf P}}= \mathsf{Eq}$ and ${{\bf Q}}$ is arbitrary. In other words, Alice and Bob have access to shared random bits and they want to simulate ${{\bf Q}}$. In this case it is always possible to approximately simulate ${{\bf Q}}$; Wyner studied the simulation capacity, and showed that it is equal to what is now known as the *Wyner common information of ${{\bf Q}}$*. When the target distribution ${{\bf Q}}$ is not Gaussian or $\mathsf{Eq}$, and the input distribution ${{\bf P}}$ is not $\mathsf{Eq}$, the problem becomes much more complicated (see, for example, [@KA15] and the references therein). Nevertheless, the preceding examples suggest that the answer should depend on how much common information there is between the coordinates of ${{\bf P}}$ and between the coordinates of ${{\bf Q}}$. To explore this notion more closely, let $\mathbf{G}_{\rho,2}$ be the centered Gaussian measure on $\mathbb{R}^2$, where each coordinate has unit variance and the correlation between the coordinates is $\rho>0$. Consider the setting where ${{\bf P}}= \mathbf{G}_{\rho,2}$. If Alice and Bob want to produce unbiased bits with maximal correlation, Borell’s noise stability inequality [@Borell:85] can be interpreted as saying that the best they can do is to output the sign of their first input. In other words, a pair of unbiased, positively correlated bits can be simulated from ${{\bf P}}$ if and only if their correlation is at most $\frac{2}{\pi} \sin^{-1}(\rho)$. The problem becomes much more difficult if Alice and Bob want to produce random variables with three equally likely outcomes each. To begin with, the analogue of Borell’s result is not known: we don’t know what Alice and Bob should to in order to maximize the probability that they agree. This issue was partially addressed in a recent work of the authors [@DMN16a]: while [@DMN16a] does not solve the simulation problem, [@DMN16a] shows that they can approximately compute a strategy that maximizes the agreement probability, to an arbitrarily small error. Note that this still does not settle the simulation problem, since joint distributions with three outcomes each are (unlike the case of two outcomes) not determined by the marginal probabilities and the agreement probability. In this work, we extend to framework of [@DMN16a] to answer the non-interactive simulation problem. Specifically, we show that if ${{\bf Q}}$ can be non-interactively simulated from ${{\bf P}}$ then one can compute, for every $\delta > 0$, a $\delta$-approximate simulation protocol. Here is an equivalent formulation, in which $|{{\bf P}}|$ denotes the size of some standard encoding of ${{\bf P}}$:  \[thm:main\] Let $(\mathcal{Z} \times \mathcal{Z}, {{\bf P}})$ and $([k] \times [k], {{\bf Q}})$ be probability spaces, and let ${{\bf X}}^n = ({{\bf X}}_1, \dots, {{\bf X}}_n)$ and ${{\bf Y}}^n = ({{\bf Y}}_1, \dots, {{\bf Y}}_n)$, where $({{\bf X}}_i, {{\bf Y}}_i)$ are independently drawn from ${{\bf P}}$. For every $\delta > 0$, there is an algorithm running in time $O_{|{{\bf P}}|, \delta}(1)$ which distinguishes between the following two cases: 1. There exist $n \in \mathbb{N}$ and $f,g: \mathcal{Z}^n \rightarrow [k]$ such that ${{\bf Q}}$ and the distribution of $(f({{\bf X}}^n), g({{\bf Y}}^n))$ are $\delta$-close in total variation distance. In this case, there is an explicit $n_0 = n_0(|{{\bf P}}|, \delta)$ such that we may choose $n \le n_0$. Further, the functions $f$ and $g$ can be explicitly computed. 2. For any $n \in \mathbb{N}$ and $f,g: \mathcal{Z}^n \rightarrow [k]$, ${{\bf Q}}$ and the distribution of $(f({{\bf X}}^n), g({{\bf Y}}^n))$ are $8\delta$-far in total variation distance. We remark here that the bound $n_0$, while computable, is not primitive recursive and has an Ackermann type growth, which is introduced by our application of a regularity lemma from [@DS14]. It is easy to see that to prove Theorem \[thm:main\], it suffices to prove the following theorem.  \[thm:junta\] With the notation of Theorem \[thm:main\], suppose there exist $f, g: \mathcal{Z}^n \to [k]$ such that $(f({{\bf X}}^n), g({{\bf Y}}^n)) \sim {{\bf Q}}$. Then, there exist $n_0 = n_0 (|\mathbf{P}|, \delta)$ and $f_{\delta}, g_\delta : \mathcal{Z}^{n_0} \rightarrow [k]$ such that ${{\bf Q}}$ and the distribution of $(f_{\delta}({{\bf X}}^{n_0}), g_{\delta}({{\bf Y}}^{n_0}))$ are $\delta$-close in total variation distance. Moreover, $n_0$ is computable. Further, the functions $f_\delta$ and $g_\delta$ can be explicitly computed. The gist of the above theorem is that if a distribution can be simulated then it can be approximately simulated with a bounded number of samples. (The crucial point in the previous sentence is that the bound is explicit, and that it depends only on ${{\bf P}}$ and the desired accuracy.) Recent work, and the difficulty of going from two to three ---------------------------------------------------------- In a recent paper, Ghazi, Kamath, and Sudan [@GKS16] proved Theorems \[thm:main\] and \[thm:junta\] in the case $k=2$. Moreover, they gave an explicit doubly exponential bound on $n_0$ and the running time of the algorithm. Borell’s noise stability theorem (which is not available for $k > 2$) played an important role in their analysis. To explain the bottleneck in extending their result for any $k$, we will elaborate on the case where $\mathcal{Z} = \mathbb{R}$ and ${{\bf P}}= \mathbf{G}_{\rho,2}$. We begin by recalling Borell’s inequality [@Borell:85] on Gaussian noise stability. \[thm:Borell\] [@Borell:85] Let ${{\bf P}}= \mathbf{G}_{\rho,2}$. For any $\mu_1, \mu_2 \in [0,1]$, let $f, g : \mathbb{R}^n \rightarrow \{0,1\}$ such that $\mathbf{E}[f] = \mu_1$ and $\mathbf{E}[g] = \mu_2$. Let us choose $\kappa_1, \kappa_2$ such that for $f_{\mathsf{LTF}}, g_{\mathsf{LTF}} : \mathbb{R} \rightarrow \{0,1\}$ defined as $f_{\mathsf{LTF}}(x) = \mathsf{sign}(x- \kappa_1)$ and $g_{\mathsf{LTF}}(x) = \mathsf{sign}(x- \kappa_2)$, we have $\mathbf{E}[f_{\mathsf{LTF}}] = \mu_1$ and $\mathbf{E}[g_{\mathsf{LTF}}] = \mu_2$. Then, $$\Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f_{\mathsf{LTF}}({{\bf X}}) = g_{\mathsf{LTF}}({{\bf Y}})] \geq \Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f({{\bf X}}) = g({{\bf Y}})].$$ Likewise, if we define $h_{\mathsf{LTF}} = \mathsf{sign} (-x - \kappa_2)$, then $\mathbf{E}[h_{\mathsf{LTF}}] = \mu_2$ and $$\Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f_{\mathsf{LTF}}({{\bf X}}) = h_{\mathsf{LTF}}({{\bf Y}})] \leq \Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f({{\bf X}}) = g({{\bf Y}})].$$ To explain the intuitive meaning of these theorems, let us define $\mathsf{Corr}_{\max}(\rho, \mu_1, \mu_2)$ and $\mathsf{Corr}_{\min}(\rho, \mu_1, \mu_2)$ as $$\mathsf{Corr}_{\max}(\rho, \mu_1, \mu_2)= \Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f_{\mathsf{LTF}}({{\bf X}}) = g_{\mathsf{LTF}}({{\bf Y}})],$$ $$\mathsf{Corr}_{\min}(\rho, \mu_1, \mu_2)= \Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f_{\mathsf{LTF}}({{\bf X}}) = h_{\mathsf{LTF}}({{\bf Y}})]$$ where $f_{\mathsf{LTF}}$, $g_{\mathsf{LTF}}$ and $h_{\mathsf{LTF}}$ are halfspaces defined in Theorem \[thm:Borell\]. Then, Borell’s result implies that for any given measures $\mu_1, \mu_2$ and functions $f, g$ with these measures, the probability that $f({{\bf X}})$ and $g({{\bf Y}})$ are identical lies between $\mathsf{Corr}_{\max}(\rho, \mu_1, \mu_2)$ and $\mathsf{Corr}_{\min}(\rho, \mu_1, \mu_2)$. Further, now, it easily follows that for any $\eta$ such that $\mathsf{Corr}_{\min }(\rho, \mu_1, \mu_2) \le \eta \le \mathsf{Corr}_{\max}(\rho, \mu_1, \mu_2)$, there is a function $g_{\eta}: \mathbb{R} \rightarrow \{0,1\}$ such that $\mathbf{E}[g_{\eta}] = \mu_2$ and $ \eta = \Pr_{({{\bf X}}, {{\bf Y}}) \sim P} [f({{\bf X}}) = g_{\eta}({{\bf Y}})]$. In fact, it is also easy to see that $g_\eta$ can be assumed to be the indicator function of an interval. Now, consider any distribution ${{\bf Q}}$ on $\{0,1\} \times \{0,1\}$, and take $({{\bf U}}, {{\bf V}}) \sim {{\bf Q}}$. Assume that there exist $f, g: \mathbb{R}^n \rightarrow \{0,1\}$ such that $(f({{\bf X}}^n) , g({{\bf Y}}^n)) \sim {{\bf Q}}$. Defining $\mu_{1, {{\bf Q}}} = \mathbf{E}[{{\bf U}}]$, $\mu_{2, {{\bf Q}}} = \mathbf{E}[{{\bf V}}]$ and $\eta_{{{\bf Q}}} = \Pr[{{\bf U}}= {{\bf V}}]$ and applying Theorem \[thm:Borell\], we obtain that there are functions $f_{{\bf Q}},g_{{\bf Q}}: \mathbb{R} \rightarrow \{0,1\}$ which satisfy $$\mathbf{E}[f_{{\bf Q}}({{\bf X}}) ] = \mu_{1, {{\bf Q}}}, \quad \mathbf{E}[g_{{\bf Q}}({{\bf Y}}) ] = \mu_{2, {{\bf Q}}},$$ and $$\Pr_{({{\bf X}}, {{\bf Y}}) \sim {{\bf P}}} [f_{{\bf Q}}({{\bf X}}) = g_{{{\bf Q}}}({{\bf Y}})] = \eta_{{\bf Q}}.$$ Further, the functions $f_{{{\bf Q}}}$ and $g_{{{\bf Q}}}$ are in fact indicators of intervals and given $\mu_{1, {{\bf Q}}}$, $\mu_{2, {{\bf Q}}}$ and $\eta_{{{\bf Q}}}$, the functions $f_{{{\bf Q}}}$ and $g_{{{\bf Q}}}$ can be explicitly computed. Observe that any distribution ${{\bf Q}}$ over $\{0,1\} \times \{0,1\}$ is characterized by the quantities $\mu_{1, {{\bf Q}}}$, $\mu_{2, {{\bf Q}}}$ and $\eta_{{{\bf Q}}}$. Thus, it implies that $(f_{{\bf Q}}({{\bf X}}), g_{{{\bf Q}}}({{\bf Y}})) \sim {{\bf Q}}$. This completely settles the non-interactive simulation problem in the case $k=2$, when ${{\bf P}}$ is the Gaussian measure $\mathbf{G}_{\rho,2}$ on $\mathbb{R}^2$. In particular, we see that when ${{\bf P}}$ is Gaussian, the result of [@GKS16] is a straightforward consequence of Theorem \[thm:Borell\]. Indeed, their main contribution was to show that the general case reduces to the Gaussian case. Moreover, that part of their argument turns out to generalize to $k > 2$ (as we will discuss later). Therefore, let us continue examining the case where ${{\bf P}}$ is Gaussian, and see why $k > 2$ causes trouble. There are two problems: 1. The analogue of Borell’s result for $k>2$ is not known. In particular, the following simple question is still open: let $\boldsymbol{\mu} \in \Delta_k$ where $\Delta_k$ is the convex hull of the standard unit vectors $\{ \mathbf{e}_1, \ldots, \mathbf{e}_k\}$. Let $A_{\boldsymbol{\mu}} = \{f : \mathbb{R}^n \rightarrow [k] : \mathbf{E}[f] =\boldsymbol{\mu} \}$. Among all $f \in A_{\boldsymbol{\mu}}$, what $f$ maximizes the probability $\Pr_{({{\bf X}},{{\bf Y}}) \sim {{\bf P}}} [f({{\bf X}}) = f({{\bf Y}})]$? If $k=2$, then Theorem \[thm:Borell\] asserts that $f$ is the indicator of some halfspace; for $k > 3$, the answer is almost completely unknown. Of particular relevance to us, it is not even known whether the optimal value can be achieved in any finite dimension (whereas in the case $k=2$, it is achieved in one dimension). 2. For $k=2$, any distribution ${{\bf R}}= ({{\bf R}}_1, {{\bf R}}_2)$ supported on $[k] \times [k]$ is completely defined by $\mathbf{E}[{{\bf R}}_1]$, $\mathbf{E}[{{\bf R}}_2]$ and $\Pr[{{\bf R}}_1 = {{\bf R}}_2]$. However, this is no longer true when $k>2$. In [@DMN16a], the authors partially circumvented the first issue. To explain the result of [@DMN16a], we will need to introduce two notions. The first is that of the (standard) Ornstein-Uhlenbeck noise operator. Namely, for any $t \ge 0$ and $f: \mathbb{R}^n \rightarrow \mathbb{R}$, we define $P_t f : \mathbb{R}^n \rightarrow \mathbb{R}$ as $$\label{eq:noise-operator-def} P_tf(x) = \mathop{\mathbf{E}}_{y \sim \gamma_n} [f(e^{-t} x +\sqrt{1-e^{-2t}} y)].$$ To see the connection between $P_t$ and our $\rho$-correlated Gaussian distribution $\mathbf{P}=\mathbf{G}_{\rho,2}$, choose $t$ so that $e^{-t} = \rho$. Then $$\mathbf{E}_{({{\bf X}}, {{\bf Y}})^n \sim {{\bf P}}^n} [f({{\bf X}}^n) \cdot f({{\bf Y}}^n)] = \mathbf{E}_{{{\bf X}}^n \sim \gamma_n} [f({{\bf X}}^n) \cdot P_t f ({{\bf X}}^n)].$$ The above quantity is often referred to as the noise stability of $f$ at noise rate $t>0$. Note that the operator $P_t$ is a linear operator on the space of functions mapping $\mathbb{R}^n$ to $\mathbb{R}$. In fact, the noise operator can be syntactically extended to functions $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$ with the same definition as in (\[eq:noise-operator-def\]). Embedding $\Delta_k$ in $\mathbb{R}^k$ and identifying $[k]$ with the vertices of $\Delta_k$, we obtain that $$\mathbf{E}_{({{\bf X}}, {{\bf Y}})^n \sim {{\bf P}}^n} [\langle f({{\bf X}}^n) , f({{\bf Y}}^n) \rangle] = \mathbf{E}_{{{\bf X}}^n \sim \gamma_n} [\langle f({{\bf X}}^n) , P_t f ({{\bf X}}^n) \rangle].$$ Let us now recall the notion of a multivariate polynomial threshold function (PTF) from [@DMN16a]. Given polynomials, $p_1, \ldots, p_k: \mathbb{R}^n \rightarrow \mathbb{R}$, define $f= \mathsf{PTF}(p_1, \ldots, p_k)$ as $$f(x) = \begin{cases} j &\textrm{if }p_j(x)>0\textrm{ and }p_i(x) \le 0 \textrm{ for all }j\not =i, \\ 1 &\textrm{ otherwise}. \\ \end{cases}$$ In [@DMN16a], the authors proved the following theorem. A notation we will adopt for the rest of the paper is that unless explicitly mentioned otherwise, the expectation is always w. r. t. the variable being a standard Gaussian where the ambient dimension will be clear from the context.  \[thm:DMN1\] Let $f: \mathbb{R}^n \rightarrow [k]$ such that $\mathbf{E}[f] = \boldsymbol{\mu} \in \mathbb{R}^k$. Then, given any $t>0, \epsilon>0$, there exists an explicitly computable $n_0 = n_0(t, k, \epsilon)$ and $d= d(t, k, \epsilon)$ such that there is a degree-$d$ PTF $g: \mathbb{R}^{n_0} \rightarrow [k]$ with 1. $\Vert \mathbf{E}[g] - \boldsymbol{\mu} \Vert_1 \le \epsilon$. 2. $\mathbf{E}[\langle g, P_t g \rangle] \ge \mathbf{E}[\langle f, P_t f \rangle] - \epsilon$. In other words, Theorem \[thm:DMN1\] shows that for any given $\boldsymbol{\mu}$ and error parameter $\epsilon>0$, there is a low-degree, low-dimensional PTF $g$ which approximately maximizes the noise stability and whose expectation is close to $\boldsymbol{\mu}$. [[[[We remark here that the issue of matching the expectation exactly versus approximately is insignificant since expectations can always be made to match exactly by suffering a tiny change in the correlation.]{}]{}]{}]{} The proof of Theorem \[thm:DMN1\] has two separate steps: 1. (**Smooth**) The first step is to show that given any $f: \mathbb{R}^n \rightarrow [k]$ with $\mathbf{E}[f] = \boldsymbol{\mu}$, there is a degree $d = d(t, k, \epsilon)$ PTF $h$ on $n$ variables such that $\Vert \mathbf{E}[h] -\boldsymbol{\mu}\Vert_1 \le \epsilon$ and $\mathbf{E}[\langle h, P_t h \rangle] \ge \mathbf{E}[\langle f, P_tf \rangle]-\epsilon$. In other words, reduce the degree but not the dimension. The main idea here is to modify the function $f$ by first smoothing it and then rounding it back to the discrete set $[k]$. It is fairly easy to show that this procedure doesn’t decrease the noise stability of $f$ (as long as the amount of smoothing is chosen to match the noise parameter $t$). The more difficult part is to show that the result of this procedure is close to a low-degree PTF. This is done using a randomized rounding argument: we show that by rounding the smoothed function at a random threshold, the expected Gaussian surface area of the resulting partition is bounded; in particular, there exists a good way to round. A well-known link between Gaussian surface area and Hermite expansions then implies that the rounded, smoothed function is almost a low-degree PTF. [[[[This argument uses the co-area formula, gradient bounds and is inspired by ideas from [@KNOW14; @Neeman14].]{}]{}]{}]{} 2. (**Reduce**) The second step is to show that given a multivariate PTF $h$, there is a multivariate PTF $g$ on $n_0 = n_0(t,k, \epsilon)$ variables such that the noise stability of $g$ is the same as that of the noise stability of $h$ up to an additive error $\epsilon$. This step uses several ideas and results from [@DS14]. To give a brief overview of this part, we start with the notion of an *eigenregular* polynomial which was introduced in [@DS14]. A polynomial is said to be $\delta$-eigenregular if for the canonical tensor $\mathcal{A}_p$ associated with the polynomial, the ratio of the maximum singular value to its Frobenius norm is at most $\delta$. Let us assume that $h = \mathsf{PTF}(p_1, \ldots, p_k)$. The *regularity lemma* from [@DS14], roughly speaking, shows that each of the polynomials $p_1, \ldots, p_k$ can be written as a low-degree “outer” polynomial composed with a bounded number of $\delta$-eigenregular, low-degree “inner” polynomials. Using the central limit theorem from [@DS14] and several other new technical ingredients, one can replace the whole collection of inner polynomials by a new collection of inner polynomials on a bounded number of variables. Moreover, one can do this replacement while hardly affecting the distribution of the outer polynomial. In particular, this whole procedure constructs a new PTF on a bounded number of inputs, and with approximately the same noise stability as the original PTF. **How to prove Theorem \[thm:junta\]:** We will first outline the proof of Theorem \[thm:junta\] in the case that ${{\bf P}}= \mathbf{G}_{\rho,2}$ (the $\rho$-correlated Gaussian measure on $\mathbb{R}^2$). As we observed earlier, any function with codomain $[k]$ naturally maps to $\mathbb{R}^k$ by identifying $i \in [k]$ with the standard unit vector $\mathbf{e}_i \in \mathbb{R}^k$. Also, for any function $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$ and $1 \le j \le k$, we let $f_j: \mathbb{R}^n \rightarrow \mathbb{R}$ denote the $j^{th}$ coordinate of $f$. Then, observe that for all $1 \le i, j \le k$, $$\Pr_{({{\bf X}}^n, {{\bf Y}}^n) \sim {{\bf P}}^n} [f({{\bf X}}^n) = i \wedge g({{\bf Y}}^n) = j ] = \mathbf{E}[f_{i} P_t g_{j}].$$ In particular, to prove Theorem \[thm:junta\] in the case ${{\bf P}}= \mathbf{G}_{\rho,2}$ it suffices to prove an improvement of Theorem \[thm:DMN1\], where the inequality $\mathbf{E}[\langle g, P_t g \rangle] \ge \mathbf{E}[\langle f, P_t f \rangle] - \epsilon$ is replaced by an almost-equality: $|\mathbf{E} [g_i P_t g_j] - \mathbf{E} [f_i P_t f_j]| \le \epsilon$ for all $i, j$. In fact, we will prove something slightly stronger, by starting with two functions instead of one. The proof of Theorem \[thm:junta\] will follow the same smooth/reduce outline as the proof of Theorem \[thm:DMN1\]. Moreover, the “reduce” step will be essentially the same as the one in [@DMN16a]. Therefore, we will outline only the “smooth” step. Define the set $\Delta_{k,\epsilon}$ as $$\Delta_{k,\epsilon} = \{x \in \mathbb{R}^k : \exists y \in \Delta_k, \ \ \Vert x - y \Vert_1 \le \epsilon\}.$$ Thus, if $\epsilon = 0$, then $\Delta_{k,\epsilon} = \Delta_k$. In the “smooth” step for the proof of Theorem \[thm:junta\], we will show that for any pair $f$, $g$ of functions ${{\bf R}}^n \to [k]$, there exist functions $\tilde f, \tilde g: {{\bf R}}^n \to {{\bf R}}^k$ such that For every $\epsilon>0$, we will show that there are functions $f_1, g_1: \mathbb{R}^n \rightarrow \mathbb{R}^k$ satisfying the following conditions: - $ \Vert \mathbf{E}[f] - \mathbf{E}[\tilde f] \Vert_1 \le \epsilon$, $ \Vert \mathbf{E}[g] - \mathbf{E}[\tilde g] \Vert_1 \le \epsilon$; - the functions $f_1, g_1$ are linear combinations of $O_{k,t,\epsilon}(1)$ low-degree PTFs (with some special structure that we will describe later); - $\Pr[\tilde f({{\bf X}}^n) \in \Delta_{k,\epsilon}] \ge 1 - \epsilon$ and $\Pr[\tilde g({{\bf Y}}^n) \in \Delta_{k,\epsilon}] \ge 1-\epsilon$; and - for any $1 \le i, j \le k$, $\big| \mathbf{E}[\langle f_{i} P_t g_{j} \rangle] - \mathbf{E}[\langle \tilde f_{i} P_t \tilde g_{j} \rangle ] \big| \le \epsilon$. The precise statement corresponding to this step is given in Lemma \[lem:smoothing\], which contains most of the technically new ideas in the paper. In particular, we employ a new “boosting” based idea to obtain the functions $\tilde f$ and $\tilde g$. The proof of Lemma \[lem:smoothing\] comes in two main steps. We start with arbitrary functions $f$ and $g$. First, we show that there are projections of polynomial threshold functions ${f_{\mathsf{sm}}}$ and ${g_{\mathsf{sm}}}$ which have the same low-level Hermite spectrum as $f$ and $g$. This is carried out in an iterative argument using a potential function, and is inspired by similar iterative algorithms appearing in boosting [@Schapire:90; @Fre95] from learning theory, the hardcore lemma in complexity theory [@Imp95] and dense model theorems in graph theory [@Frieze1999] and additive combinatorics [@Tao:07; @TTV09:conf]. While these iterative algorithms have recently been used to prove structural results in complexity theory [@DDFS14; @LRS15; @TTV09:conf], since our algorithm is in the multidimensional setting, it is somewhat more delicate than these applications. The main argument here is carried out in Lemma \[lem:Boosting\], and we bound the degree of the resulting polynomials in Corollary \[corr:fsm\]. The next step is to show that we can replace the projected polynomial threshold functions by polynomials that with high probability take values very close to the simplex (call them ${f_{\mathsf{sm}}}'$ and ${g_{\mathsf{sm}}}'$). This is carried out in Lemma \[lem:smoothing-1\], using Bernstein approximations for Lipschitz functions. Finally, we use some probabilistic tricks to replace ${f_{\mathsf{sm}}}'$ and ${g_{\mathsf{sm}}}'$ by functions $\tilde f$ and $\tilde g$ which are linear combinations of low-degree PTFs. This finishes the proof of Lemma \[lem:smoothing\]. What happens when ${{\bf P}}$ is not Gaussian? ----------------------------------------------  \[sec:non-gauss\] So far, the discussion has pertained to the case when ${{\bf P}}= \mathbf{G}_{\rho,2}$. What happens if ${{\bf P}}$ is a different probability distribution? As we have remarked earlier, the main result of [@GKS16] is that the $k=2$ case of Theorem \[thm:junta\] essentially reduces to the special case ${{\bf P}}= \mathbf{G}_{\rho,2}$. Their argument uses quite general tools from Boolean function analysis such as the invariance principle [@MOO10; @Mossel2010] and regularity lemmas for low-degree polynomials [@DSTW:10; @DDS14]. A similar argument can be used to prove Theorem \[thm:junta\] by reducing to the Gaussian case; however, we will actually need a slightly stronger Gaussian version of Theorem \[thm:junta\]: \[thm:junta-strong\] Let ${{\bf P}}= \mathbf{G}_{\rho,2}$ and let $f^{(1)}, \ldots, f^{(\ell)}: \mathbb{R}^n \rightarrow [k]$ and $g^{(1)}, \ldots, g^{(\ell)}: \mathbb{R}^n \rightarrow [k]$ where we define ${{\bf Q}}_{i,j}$ as ${{\bf Q}}_{i,j} = (f^{(i)}({{\bf X}}^n), g^{(j)}({{\bf Y}}^n))$. Then, for every $\delta>0$, there is an explicitly defined constant $n_0 = n_0(\ell, k, \delta)$ and explicitly defined functions $f^{(1)}_{\mathsf{junta}}, \ldots, f^{(\ell)}_{\mathsf{junta}}: \mathbb{R}^{n_0} \rightarrow [k]$ and $g^{(1)}_{\mathsf{junta}}, \ldots, g^{(\ell)}_{\mathsf{junta}}: \mathbb{R}^{n_0} \rightarrow [k]$ such that for every $1 \le i, j \le \ell$, ${d_{\mathrm TV}}((f^{(i)}_{\mathsf{junta}}({{\bf X}}^{n_0}), g^{(j)}_{\mathsf{junta}}({{\bf Y}}^{n_0})), {{\bf Q}}_{i,j}) \le \delta$. Note that the $\ell=1$ case of Theorem \[thm:junta-strong\] is exactly the ${{\bf P}}= \mathbf{G}_{\rho,2}$ case of Theorem \[thm:junta\], the proof of which we outlined above. Then $\ell > 1$ case has essentially the same proof, but with more notation. In order to prove Theorem \[thm:junta\] from Theorem \[thm:junta-strong\], Alice and Bob both execute a “decision tree.” By standard arguments from Boolean function analysis (see [@ODonnell:book] for definitions of the terminology that follows), Alice and Bob can represent $f$ and $g$ by small decision trees, such that most of the “leaf” functions (call them $\{f^{(i)}\}_{1 \le i \le \ell}$ and $\{g^{(i)}\}_{1 \le i \le \ell}$) are *low-influence* functions. The invariance principle of Mossel *et al.* [@MOO10; @Mossel2010] allows us to replace $\{f^{(i)}\}_{1 \le i \le \ell}$ and $\{g^{(i)}\}_{1 \le i \le \ell}$ by functions of Gaussian variables; essentially, we can pretend that Alice and Bob have access to independent copies of $\mathbf{G}_{\rho,2}$ where $\rho$ is the so-called maximal correlation coefficient of $({{\bf X}}, {{\bf Y}})$. Finally, we apply Theorem \[thm:junta-strong\] to this collection of Gaussian “leaf” functions. In the end, we have replaced Alice and Bob’s initial functions by a pair of decision trees of bounded size, where every leaf function is a function of a bounded number of Gaussian variables. We give a more detailed overview of this reduction in Section \[section:GKS\]. Acknowledgements ---------------- We thank Pritish Kamath, Badih Ghazi and Madhu Sudan for pointing out that the $\ell=1$ case of Theorem \[thm:junta-strong\] is not sufficient to derive Theorem \[thm:junta\]. (An earlier version of this paper incorrectly claimed that it was.) We also thank the anonymous reviewers who pointed out the same gap. Technical preliminaries ======================= We will start by defining some technical preliminaries which will be useful for the rest of the paper. For $k \in \mathbb{N}$ and $1\le i \le k$, let $\mathbf{e}_i$ be the unit vector along coordinate $i$ and let $\Delta_k$ be the convex hull formed by $\{ \mathbf{e}_i \}_{1\le i \le k}$. In this paper, we will be working on the space of functions $f: \mathbb{R}^n \rightarrow \mathbb{R}$ where the domain is equipped with the standard $n$ dimensional normal measure (denoted by $\gamma_n(\cdot)$). Unless explicitly mentioned otherwise, all the functions considered in this paper will be in $L^2(\gamma_n)$. A key property of such functions is that they admit the so-called Hermite expansion. Let us define a family of polynomials $H_q: \mathbb{R} \rightarrow \mathbb{R}$ (for $q \ge 0$) as $$H_0(x) = 1 ; \ H_1(x) = x ; \ H_q(x) = \frac{(-1)^q}{\sqrt{q!}} \cdot e^{x^2/2} \cdot \frac{d^q}{dx^q} e^{-x^2/2}.$$ Let $\mathbb{Z}^{\ast}$ denote the subset of non-negative integers and $S \in \mathbb{Z}^{\ast n}$. Define $H_S: \mathbb{R}^n \rightarrow \mathbb{R}$ as $$H_S(z) = \prod_{i=1}^n H_{S_i}(z_i).$$ It is well known that the set $\{H_S\}_{S \in \mathbb{Z}^{\ast n}}$ forms an orthonormal basis for $L^2(\gamma_n)$. In other words, every $f \in L^2(\gamma_n)$ may be written as $$f = \sum_{S \in \mathbb{Z}^{\ast n}} \widehat{f}(S) \cdot H_S,$$ where $\widehat{f}(S)$ are typically referred to as the *Hermite coefficients* and expansion is referred to as the *Hermite expansion*. The notion of Hermite expansion can be easily extended to $f: R^n \rightarrow \mathbb{R}^k$ as follows: Let $f = (f_1, \ldots, f_k)$ and let $$f_i = \sum_{S \in \mathbb{Z}^{\ast n}} \widehat{f_i}(S) \cdot H_S.$$ Then, the Hermite expansion of $f$ is given by $\sum_{S \in \mathbb{Z}^{\ast n}} \widehat{f}(S) \cdot H_S$ where $\widehat{f}(S) = (\widehat{f_1}(S), \ldots, \widehat{f_k}(S))$. In this setting, we also have Parseval’s identity: $$\label{eq:parseval} \int \Vert f (x) \Vert_2^2 \ \gamma_n(x) dx = \sum_{S \in \mathbb{Z}^{\ast n}} \Vert \widehat{f}(S) \Vert_2^2$$ For $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$ and $d \in \mathbb{N}$, define $f_{\le d} : \mathbb{R}^n \rightarrow \mathbb{R}^k$ by $$f_{\le d} (x) = \sum_{S: |S| \le d} \widehat{f}(S) \cdot H_S(x).$$ Here $|S|$ denotes the $\ell_1$ norm of the vector $S$. We will define $\mathsf{W}^{\le d} [f] = \Vert f_{\le d} \Vert_2^2$ and $\mathsf{W}^{> d} [f] = \sum_{|S|>d} \Vert \widehat{f}(S) \Vert_2^2$. ### Ornstein-Uhlenbeck operator {#ornstein-uhlenbeck-operator .unnumbered} The Ornstein-Uhlenbeck operator $P_t$ is defined for $t \in [0, \infty)$ such that for any $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$, $$(P_t f)(x) = \int_{y \in \mathbb{R}^n} f(e^{-t} \cdot x + \sqrt{1- e^{-2t}} \cdot y) d \gamma_n(y).$$ Note that if $f : \mathbb{R}^n \rightarrow \Delta_k$, then so is $P_t f$ for every $t>0$. A basic fact about the Ornstein-Uhlenbeck operator is that the functions $\{H_S\}$ are eigenfunctions of this operator. We leave the proof of the next proposition to the reader. For $S \in \mathbb{Z}^{\ast n}$, $P_t H_S = e^{-t \cdot |S|} \cdot H_S$. ### Probabilistic inequalities \[thm:hyper\] Let $p: \mathbb{R}^n \rightarrow \mathbb{R}$ be a degree-$d$ polynomial. Then, for any $t>0$, $$\Pr_{x} \big[|p(x) - \mathbf{E}[p(x)]| \ge t \cdot \sqrt{\mathsf{Var}[p]}\big] \leq d \cdot e^{-t^{2/d}}.$$ \[thm:combine-hyper\] Let $a, b: \mathbb{R}^n \rightarrow \mathbb{R}$ be degree $d$ polynomials satisfying $\mathbf{E}_x [a(x) - b(x)]=0$ and $\mathsf{Var}[a-b] \le (\tau/d)^{3d} \cdot \mathsf{Var}[a]$. Then, $\Pr_{x}[\mathsf{sign}(a(x)) \not = \mathsf{sign}(b(x))] = O(\tau)$. ### Producing non-integral functions Instead of producing functions $\{f^{(j)}_{\mathsf{junta}}\}_{1 \le i \le \ell}$ and $\{g^{(j)}_{\mathsf{junta}}\}_{1 \le i \le \ell}$ (in Theorem \[thm:junta-strong\]) with range $[k]$, we will actually produce functions $\{\tilde{f}^{(j)}_{\mathsf{junta}}\}_{1 \le i \le \ell}$ and $\{\tilde{g}^{(j)}_{\mathsf{junta}}\}_{1 \le i \le \ell}$ whose range will be close to $\Delta_{k,\epsilon}$. The next two lemmas show that functions with range $\Delta_{k,\epsilon}$ can be converted to non-interactive simulation strategies with range $[k]$ with nearly the same guarantee. More precisely, we show that given $f', g' : \mathbb{R}^n \rightarrow \Delta_{k,\epsilon}$, there are functions $f,g: \mathbb{R}^n \rightarrow [k]$ such that $\mathbf{E}[f] \approx \mathbf{E}[f']$, $\mathbf{E}[g] \approx \mathbf{E}[g']$ and for any $1 \le j_1, j_2 \le k$, $\mathbf{E}[f_{j_1} P_t g_{j_2}] \approx \mathbf{E}[f'_{j_1} P_t g'_{j_2}]$. To define this, let us adopt the notation that given a point $x \in \mathbb{R}^k$, $\mathsf{Proj}(x)$ denotes the closest point to $x$ in $\Delta_k$ in Euclidean distance.  \[lem:round2\] Let $f : \mathbb{R}^n \rightarrow \mathbb{R}^k$ which satisfies the following two conditions: 1. $\Pr_{x} [f(x) \not \in \Delta_{k,\delta}] \le \delta$. 2. For all $x$, $\Vert f(x) \Vert_\infty \le k$. Then, there is a function $f_1: \mathbb{R}^n \rightarrow \Delta_k$ such that $\Vert f-f_1 \Vert_1 = O(k \cdot \delta)$. Define $f_1 = \mathsf{Proj}(f)$. Note that if $x$ is such that $f(x) \in \Delta_{k,\delta}$, then by definition, $\Vert f_1(x) -f(x) \Vert_1 \le \delta$. On the other hand, for any $x$, $\Vert f(x) -f_1(x) \Vert_1 \le k$. This proves the claim.  \[lem:round3\] Let $f_1, g_1 : \mathbb{R}^n \rightarrow \Delta_k$. Then, there exist (explicitly defined) $f_2, g_2: \mathbb{R}^{n+2} \rightarrow [k]$ such that 1. $\mathbf{E}[f_2 ] = \mathbf{E}[f_1]$ and $\mathbf{E}[g_2 ] = \mathbf{E}[g_1]$. 2. For any $1 \le j, \ell \le k$, $$\mathbf{E}[ f_{1,j} P_t g_{1,\ell} ]=\mathbf{E}[ f_{2,j} P_t g_{2,\ell}].$$ [Further, the function $f_2$ (resp. $g_2$) is dependent only on $f_1$ (resp. $g_1$).]{} Let $z = (x, z_1, z_2)$ where $x \in \mathbb{R}^n$ and $z_1, z_2 \in \mathbb{R}$. For any $y \in \Delta_k$, let us divide $\mathbb{R}$ into $k$ intervals $S_1, \ldots, S_k$ such that for $z \sim \gamma$, $\Pr[z \in S_i] = y_i$. For $y \in \Delta_k$ and $z' \in \mathbb{R}$, $\mathsf{Part}(y,z) = i$ if $z' \in S_i$. Define $f_2: \mathbb{R}^{n+2} \rightarrow [k]$ as $$f_2(z) =f_2(x,z_1,z_2) = \mathsf{Part}(f_1(x), z_1).$$ $$g_2(z) =g_2(x,z_1,z_2) = \mathsf{Part}(g_1(x), z_2).$$ We will now verify the claimed properties. First of all, observe that the codomain of $f_2$ and $g_2$ is indeed $k$. Second, by definition, it is easy to follow that $\mathbf{E}[f_1]= \mathbf{E}[f_2]$ and $\mathbf{E}[g_1]= \mathbf{E}[g_2]$. Finally, note that $$\mathbf{E}[ f_{1,j} P_t g_{1,\ell} ]= \mathbf{E}_{({{\bf X}}^n, {{\bf Y}}^n) \sim {{\bf P}}^n}[ f_{1,j}({{\bf X}}^n) g_{1,\ell}({{\bf Y}}^n) ].$$ On the other hand, suppose $z_1, z_2 \sim \gamma$. Then, $$\Pr_{z_1, z_2 \sim \gamma} [f_2(x,z_1, z_2) = j \ \wedge \ g_2(y,z_1, z_2) = \ell] = f_{1,j}(x) g_{1,\ell}(y).$$ Thus, we obtain that $$\mathbf{E}[ f_{2,j} P_t g_{2,\ell} ] = \mathbf{E}_{({{\bf X}}^n, {{\bf Y}}^n) \sim {{\bf P}}^n} [f_{1,j}({{\bf X}}^n) g_{1,\ell} ({{\bf Y}}^n)] = \mathbf{E}[ f_{1,j} P_t g_{1,\ell} ].$$ Proof strategy for the main theorem ----------------------------------- To describe the proof strategy for the main section, we first define a class of $k$-ary functions called *polynomial plurality functions* (PPFs) which are closely related to the multivariate PTFs defined in the introduction but are somewhat different. For this, let us first define the function $\arg \max$ as follows $\arg \max : \mathbb{R}^k \rightarrow \mathbb{R}^k$ is defined as $$\arg \max (x_1, \ldots, x_k) = \begin{cases} \mathbf{e}_i \ \ &\text{if } x_i > x_j \ \textrm{ for all } j\not =i \\ 0 &\textrm{otherwise} \\ \end{cases}$$ A function $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$ is said to be a PPF of degree-$d$ if there exists a polynomial $p: \mathbb{R}^n \rightarrow \mathbb{R}$ of degree $d$ and an index $1 \le j \le x$ such that $f= \arg \max(z)$ where $z_i = \delta_{i=j} \cdot p(x)$. Given polynomial $p: \mathbb{R}^n \rightarrow \mathbb{R}$ and $ 1 \le j \le k$, we define the function $\mathsf{PPF}_{p, j}$ as $$\mathsf{PPF}_{p, j}(x) = \arg \max \big(\underbrace{0, \ldots, 0}_{(j-1) \textrm{ times}}, p(x) , \underbrace{0, \ldots, 0}_{(n-j) \textrm{ times}} \big).$$  \ The following is a basic fact about PPFs.  \[fact:balanced\] For any PPF $f$ of degree $d$, if $f=\mathsf{PPF}_{p, j}$, we can assume without loss of generality that $\mathsf{Var}(p)=1$. Further, by changing $f$ in at most $\delta$ fraction of places, we can assume that $|\mathbf{E}[p(x) ]| \le d \cdot \log^{d/2}(1/\delta)$. Such a PPF is said to be a $(d,\delta)$-balanced PPF. The fact about variance follows simply by scaling. To bound $|\mathbf{E}[p(x) ]|$, note that if $|\mathbf{E}[p(x)]| > d \cdot \log^{d/2}(1/\delta)$, then $\Pr_x [\mathsf{sign}(p(x)) = \mathsf{sign}(\mathbf{E}[p(x)])] \ge 1-\delta$ (using Theorem \[thm:combine-hyper\]). Thus, if we set $q(x) = p(x) - \mathbf{E}[p(x)] + d \cdot \log^{d/2}(1/\delta) \cdot \mathsf{sign}(\mathbf{E}[p(x)])$, then $\Pr_x [p(x) \not = q(x)] \le \delta$. The PPF defined as $\mathsf{PPF}_{q,j}$ satisfies all the desired properties. To prove our main theorem (Theorem \[thm:junta-strong\]), we will prove the following two intermediate results.  \[lem:smoothing\] For $1 \le i \le \ell$, let $f^{(i)}, g^{(i)}: \mathbb{R}^n \rightarrow [k]$ such that $\mathbf{E}[f^{(i)}] = \boldsymbol{\mu}^{(i)}_f$ and $\mathbf{E}[g^{(i)}]= \boldsymbol{\mu}^{(i)}_g$. Then, for any $t>0$, $\delta>0$, $d_0 = d_0(t,k,\delta) = (2/t) \cdot \log( k^2/\delta)$ and $1\le i \le \ell$, there are functions $f^{(i)}_1, g^{(i)}_1: \mathbb{R}^n \rightarrow \mathbb{R}^k$ which satisfy the following conditions: 1. For any $x \in \mathbb{R}^n$ and $1 \le i \le \ell$, $f^{(i)}_1(x), g^{(i)}_1(x)$ always lies in the positive orthant. 2. For any $x \in \mathbb{R}^n$ and $1 \le i \le \ell$, $\Vert f^{(i)}_1(x) \Vert_\infty , \Vert g^{(i)}_1(x) \Vert_\infty\le 1$. 3. For $ 1 \le i \le \ell$, $\Pr_{x} [f^{(i)}_1(x) \not \in \Delta_{k,k\delta/2}]\le\delta/2$ and $\Pr_{x} [g^{(i)}_1(x) \not \in \Delta_{k,k\delta/2}] \le \delta/2$. 4. For $ 1 \le i \le \ell$, $|\mathbf{E}[f^{(i)}_1] - \boldsymbol{\mu}^{(i)}_f|,\ |\mathbf{E}[g^{(i)}_1] - \boldsymbol{\mu}^{(i)}_g| =O(k \delta)$. 5. For $ 1 \le i,j \le \ell$ and for any $1 \le s_1, s_2 \le k$, $|\mathbf{E}[ f^{(i)}_{1,s_1} P_t g^{(j)}_{1,s_2} ] -\mathbf{E}[ f^{(i)}_{s_1} P_t g^{(j)}_{s_2} ]| =O(k \cdot \delta)$. 6. For $1 \le i \le \ell$, $f^{(i)}_1$ and $g^{(i)}_1$ are of the following form. There are degree-$d_0$ polynomials $\{p^{(i)}_{s,j,1}\}_{1\le i \le \ell, 1\le s \le k, 1 \le j \le m}$ and $\{p^{(i)}_{s,j,2}\}_{1\le i \le \ell, 1\le s \le k, 1 \le j \le m}$ $$f^{(i)}_1 = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{p^{(i)}_{s,j,1},j}(x) \ , \ g^{(i)}_1 = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{p^{(i)}_{s,j,2},j}(x),$$ such that the resulting PPFs $\mathsf{PPF}_{p^{(i)}_{s,j,1},j}(x)$ and $\mathsf{PPF}_{p^{(i)}_{s,j,2},j}(x)$ are $(d_0,\delta)$-balanced PPFs. Here $m= O(1/\delta)$. [[[[Further, the function $f^{(i)}_1$ (resp. $g^{(i)}_1$) is dependent only on $f^{(i)}$ (resp. $g^{(i)}$), $t$, $k$ and $\delta$. ]{}]{}]{}]{}  \[lem:junta-construction\] Let $\{p^{(i)}_{s,j,1}\}_{1 \le i \le \ell, 1\le s \le k, 1 \le j \le m}$ and $\{p^{(i)}_{s,j,2}\}_{1 \le i \le \ell,1\le s \le k, 1 \le j \le m}$ be degree-$d_0$ polynomials. For $1 \le i \le \ell$, let $f^{(i)}_1, g^{(i)}_1: \mathbb{R}^n \rightarrow \mathbb{R}^k$ be defined as in Lemma \[lem:smoothing\] and satisfy the following two conditions: 1. For $1\le i \le \ell$, $1 \le s \le k$ and $1 \le j \le m$, all the PPFs $\mathsf{PPF}_{p^{(i)}_{s,j,1},j}$ and $\mathsf{PPF}_{p^{(i)}_{s,j,2},j}$ are $(d_0,\delta)$-balanced PPFs. 2. For $1\le i \le \ell$, $\Pr_{x} [f^{(i)}_1(x) \not \in \Delta_{k,\delta}] \le \delta$ and $\Pr_{x} [g^{(i)}_1(x) \not \in \Delta_{k,\delta}] \le \delta$. Then, there exists an explicit constant $n_0 = n_0 (d_0,k,\delta,\ell)$ such that there are polynomials $\{r_{s,j,1}^{(i)}\}_{1 \le i \le \ell, 1 \le s \le k, 1 \le j \le m}$ and $\{r_{s,j,2}^{(i)}\}_{1 \le i \le \ell,1 \le s \le k, 1 \le j \le m}$ satisfying the following conditions: For $1 \le i \le \ell$, let us define the functions $f^{(i)}_{\mathsf{junta}}, g^{(i)}_{\mathsf{junta}}: \mathbb{R}^{n_0} \rightarrow \mathbb{R}^k$ defined as $$f^{(i)}_{\mathsf{junta}} = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{r^{(i)}_{s,j,1},s}(x) \ , \ g^{(i)}_{\mathsf{junta}} = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{r^{(i)}_{s,j,2},s}(x),$$ Then, they satisfy the following three conditions: For all $1 \le i \le \ell$, 1. $ \Vert \mathbf{E}[f^{(i)}_{\mathsf{junta}}] - \mathbf{E}[f^{(i)}_1] \Vert_1 \le \delta$ and $ \Vert \mathbf{E}[g^{(i)}_{\mathsf{junta}}] - \mathbf{E}[g^{(i)}_1] \Vert_1 \le \delta$. 2. $\Pr_{x} [f^{(i)}_{\mathsf{junta}}(x) \not \in \Delta_{k,\sqrt{\delta}}] \le \sqrt{\delta}$ and $\Pr_{x} [g^{(i)}_{\mathsf{junta}}(x) \not \in \Delta_{k,\sqrt{\delta}}] \le \sqrt{\delta}$. 3. For any $1 \le i, j \le \ell$, $1 \le s_1, s_2 \le k$, $|\mathbf{E}[ f^{(i)}_{1,s_1} P_t g^{(j)}_{1,s_2} ] -\mathbf{E}[ f^{(i)}_{\mathsf{junta},s_1} P_t g^{(j)}_{\mathsf{junta},s_2} ]| \le \delta$. **Proof of Theorem \[thm:junta-strong\]:**The proof of Theorem \[thm:junta-strong\] follows by applying Lemma \[lem:smoothing\] on the set $\{f^{(i)} \cup g^{(i)}\}_{1 \le i \le \ell}$ and subsequently applying Lemma \[lem:junta-construction\]. While the range of functions produced by $\{f_{\mathsf{junta}}^{(i)} \cup g_{\mathsf{junta}}^{(i)}\}_{1 \le i \le \ell}$ is not $\Delta_k$, by applying Lemma \[lem:round2\] and Lemma \[lem:round3\], we can rectify this issue. We note here that the functions obtained in this process, namely $\{f_{\mathsf{junta}}^{(i)} \cup g_{\mathsf{junta}}^{(i)}\}_{1 \le i \le \ell}$ are explicit. Namely, the functions obtained before applying Lemma \[lem:round2\] and Lemma \[lem:round3\] are low-degree PPFs. Lemma \[lem:round2\] applies a projection on to the standard simplex $\Delta_k$. Likewise, Lemma \[lem:round3\] also produces an explicit function as its output. We now explain why $\{f_{\mathsf{junta}}^{(i)} \cup g_{\mathsf{junta}}^{(i)}\}_{1 \le i \le \ell}$ satisfy the stated guarantees. In particular, overloading notation, let us denote the functions obtained by application of Lemma \[lem:round2\] and Lemma \[lem:round3\] as $f_{\mathsf{junta}}^{(i)}$ and $ g_{\mathsf{junta}}^{(i)}$. Then, we see that $$\Vert \mathbf{E}[f^{(i)}_{\mathsf{junta}}] - \mathbf{E}[f^{(i)}_1] \Vert_1 \le O (k \cdot \sqrt{\delta}), \ \Vert \mathbf{E}[g^{(i)}_{\mathsf{junta}}] - \mathbf{E}[g^{(i)}_1] \Vert_1 \le O(k \cdot \sqrt{\delta}),$$ $$~\label{eq:tv-dist} \textrm{For any }1 \le i, j \le \ell, \ 1 \le s_1, s_2 \le k, \ |\mathbf{E}[ f^{(i)}_{1,s_1} P_t g^{(j)}_{1,s_2} ] -\mathbf{E}[ f^{(i)}_{\mathsf{junta},s_1} P_t g^{(j)}_{\mathsf{junta},s_2} ]| \le \delta$$ Note that the functions $\{f^{(i)} \cup g^{(i)}\}_{1 \le i \le \ell}$ have arity $n_0$. Further, observe that for $1 \le s_1, s_2 \le k$ and $1 \le i, j \le \ell$, $$\begin{aligned} \Pr [f_{\mathsf{junta}}^{(i)}({{\bf X}}^{n_0}) = s_1 \ \wedge \ g_{\mathsf{junta}}^{(j)}({{\bf Y}}^{n_0}) = s_2] &=& \mathbf{E}[f^{(i)}_{\mathsf{junta},s_1} P_t g^{(j)}_{\mathsf{junta},s_2} ] \ \textrm{and} \\ \Pr [f^{(i)}({{\bf X}}^{n}) = s_1 \ \wedge \ g^{(j)}({{\bf Y}}^{n}) = s_2] &=& \mathbf{E}[f^{(i)}_{s_1} P_t g^{(j)}_{s_2} ]. \end{aligned}$$ Thus, for $1\le s_1, s_2 \le k$, $$\big| \Pr [f_{\mathsf{junta}}^{(i)}({{\bf X}}^{n_0}) = s_1 \ \wedge \ g_{\mathsf{junta}}^{(j)}({{\bf Y}}^{n_0}) = s_2]- \Pr [f^{(i)}({{\bf X}}^{n}) = s_1 \ \wedge \ g^{(j)}({{\bf Y}}^{n}) = s_2]\big| \le \delta.$$ This immediately implies that $${d_{\mathrm TV}}\big(\big(f_{\mathsf{junta}}^{(i)}({{\bf X}}^{n_0}) , g_{\mathsf{junta}}^{(j)}({{\bf Y}}^{n_0})\big),\big(f_{}^{(i)}({{\bf X}}) , g_{}^{(j)}({{\bf Y}})\big)\big) = O(k^2 \delta),$$ which finishes the proof. Proof of Lemma \[lem:smoothing\] ================================ The proof of Lemma \[lem:smoothing\] shall proceed in several steps. Note that Lemma \[lem:smoothing\] claims existence of $\{f_1^{(i)}\}$ and $\{g_1^{(i)}\}$ which satisfies six different properties. The functions $\{f^{(i)}\}$ and $\{g^{(i)}\}$ themselves satisfy the first five properties and thus, the only non-trivial task that remains is to achieve the sixth property. The sixth property will be achieved by gradual modification of $\{f^{(i)}\}$ and $\{g^{(i)}\}$ in a sequence of steps which are explained below. 1. Corollary \[corr:Boosting\] allows us to replace $f^{(i)}$ (resp. $g^{(i)}$) with ${f_{\mathsf{sm}}}^{(i)}$ (resp. ${g_{\mathsf{sm}}}^{(i)}$), which is the projection onto $\Delta_k$ of a polynomial, and which shares the same low-degree Hermite expansion as $f^{(i)}$ (resp. $g^{(i)}$). Coupled with Claim \[clm:noise-degree\], this shows that if $f^{(i)}$ is replaced by ${f_{\mathsf{sm}}}^{(i)}$ and $g^{(i)}$ is replaced by ${g_{\mathsf{sm}}}^{(i)}$, then the first five properties in Lemma \[lem:smoothing\] hold. On the other hand, note that while ${f_{\mathsf{sm}}}^{(i)}$ and ${g_{\mathsf{sm}}}^{(i)}$ do not have the full structure claim in Property 6, they do have some resemblance to PPFs. Corollary \[corr:Boosting\] is the technically most innovative part of the proof and in turn relies on Lemma \[lem:Boosting\]. [[[[A crucial point for the application to non-interactive simulation is that the construction of ${f_{\mathsf{sm}}}^{(i)}$ (resp. ${g_{\mathsf{sm}}}^{(i)}$) is dependent only on $f^{(i)}$ (resp. $g^{(i)}$) and the error parameters. ]{}]{}]{}]{} 2. Applying Bernstein-type approximations for Lipschitz functions in terms of low-degree polynomials, Lemma \[lem:smoothing-1\] shows that ${f_{\mathsf{sm}}}^{(i)}$ and ${g_{\mathsf{sm}}}^{(i)}$ can be replaced by ${f_{\mathsf{sm}}}^{'(i)}$ and ${g_{\mathsf{sm}}}^{'(i)}$ where each coordinate of ${f_{\mathsf{sm}}}^{'(i)}$ and ${g_{\mathsf{sm}}}^{'(i)}$ is a low-degree multivariate polynomial. [[[[Again, crucially for the application to non-interactive simulation, the function ${f_{\mathsf{sm}}}^{'(i)}$ (resp. ${g_{\mathsf{sm}}}^{'(i)}$) is dependent only on ${f_{\mathsf{sm}}}^{(i)}$ (resp. ${g_{\mathsf{sm}}}^{(i)}$) and the error parameters.]{}]{}]{}]{} 3. Finally, the functions ${f_{\mathsf{sm}}}^{'(i)}$ and ${g_{\mathsf{sm}}}^{'(i)}$ are changed to $f_1^{(i)}$ and $g_1^{(i)}$ which are linear combinations of PPFs (as promised in Lemma \[lem:smoothing\]) using some simple probabilistic observations. [[[[Again, the conversion of ${f_{\mathsf{sm}}}^{'(i)}$ to $f_1^{(i)}$ is only dependent on ${f_{\mathsf{sm}}}^{'(i)}$ and desired error parameters. ]{}]{}]{}]{} Projections of polynomials -------------------------- We begin with the first step described above. The first lemma relates the (by now, well-known) connection between the low-degree Hermite expansion of a function and its noise stability. In particular, it shows that if a pair of functions $(f^{(1)}, g^{(1)})$ (whose range is $\Delta_k$) is replaced by another pair $(\underline{f}^{(1)}, \underline{g}^{(1)})$ such that low-degree Hermite spectrum of $f^{(1)}$ (resp. $g^{(1)}$) is close to that of $\underline{f}^{(1)}$ (resp. $\underline{g}^{(1)}$) are close to each other, then for any $1 \le s_1, s_2 \le k$, $\mathbf{E}[f^{(1)}_{s_1} P_t g^{(1)}_{s_2}] \approx \mathbf{E}[\underline{f}^{(1)}_{s_1} P_t \underline{g}^{(1)}_{s_2}]$. \[clm:noise-degree\] Let $f^{(1)}, g^{(1)} , \underline{f}^{(1)}, \underline{g}^{(1)} : \mathbb{R}^n \rightarrow \Delta_k$ such that for $d_1= d_1(\delta, t) = \frac{1}{t} \log (k^2/\delta)$ we have $$\mathsf{W}^{\le d_1} [ (f^{(1)}-\underline{f}^{(1)}) ] , \ \mathsf{W}^{\le d_1} [g^{(1)}-\underline{g}^{(1)}]\le \delta^2/k^4.$$ Then, $\sum_{1 \le s_1, s_2 \le k} | \mathbf{E}[ f^{(1)}_{s_1} P_{t} g^{(1)}_{s_2} ] - \mathbf{E}[ \underline{f}^{(1)}_{s_1} P_t \underline{g}^{(1)}_{s_2} ] | \le \delta$. For any $1 \le s_1, s_2 \le k$, $$\begin{aligned} \big| \mathbf{E}[f^{(1)}_{s_1} P_t g^{(1)}_{s_2}]-\mathbf{E}[\underline{f}^{(1)}_{s_1} P_t \underline{g}^{(1)}_{s_2}] \big| &\le& \big| \mathbf{E}[ (f^{(1)}_{s_1} - \underline{f}^{(1)}_{s_1}) P_t g^{(1)}_{s_2} ] \big| + \big| \mathbf{E}[ \underline{f}^{(1)}_{s_1} P_t (g^{(1)}_{s_2} - \underline{g}^{(1)}_{s_2}) ] \big| \end{aligned}$$ By using the self-adjointness of the noise operator and applying the Jensen’s inequality, the first term can be bounded as $$\big| \mathbf{E}[ (f^{(1)}_{s_1} - \underline{f}^{(1)}_{s_1}) P_t g^{(1)}_{s_2} ] \big| \le \sqrt{\mathbf{E}[P_t(f^{(1)} - \underline{f}^{(1)})_{s_1}^2]}\sqrt{\mathbf{E}[(g^{(1)})_{s_2}^2]} \le \sqrt{\mathbf{E}[P_t(f^{(1)}- \underline{f}^{(1)})_{s_1}^2]}.$$ Similarly bounding $ \big| \mathbf{E}[ \underline{f}^{(1)}_{s_1} P_t (g^{(1)}_{s_2} - \underline{g}^{(1)}_{s_2}) ] \big| $, we obtain $$\big| \mathbf{E}[ (f^{(1)}_{s_1} - \underline{f}^{(1)}_{s_1}) P_t g^{(1)}_{s_2} ] \big| + \big| \mathbf{E}[ \underline{f}^{(1)}_{s_1} P_t (g^{(1)}_{s_2} - \tilde{g}^{(1)}_{s_2}) ] \big| \le \sqrt{\mathbf{E}[P_t(f^{(1)}_{s_1} - \underline{f}^{(1)}_{s_1})^2]} + \sqrt{\mathbf{E}[P_t(g^{(1)}_{s_2} - \underline{g}^{(1)}_{s_2})^2]}.$$ Now, applying the condition that $\mathsf{W}^{\le d_1} [ (f^{(1)}-f^{(2)}) ] \le \delta^2/k^4$, we get that $$\mathbf{E}[\Vert P_t(f^{(1)} - \underline{f}^{(1)}) \Vert_2^2] \le \frac{\delta^2}{k^4} + e^{-2td_1} \cdot \mathbf{E}[\Vert (f^{(1)} - \underline{f}^{(1)}) \Vert_2^2] \le \frac{2 \delta^2}{k^4}.$$ The last inequality uses the fact that for all $x$, $\Vert f^{(1)}(x) - \underline{f}^{(1)}(x) \Vert_1 \le 1$. Likewise, we also get $\mathbf{E}[\Vert P_t(g^{(1)} - \underline{g}^{(1)}) \Vert_2^2 \leq 2 \delta^2/k^4$. Combining this, we obtain that for all $1 \le s_1, s_2 \le k$, $$\big| \mathbf{E}[ f^{(1)}_{s_1} P_{t} g^{(1)}_{s_2} ] - \mathbf{E}[ \underline{f}^{(1)}_{s_1} P_t \underline{g}^{(1)}_{s_2} ] \big| \le \frac{2 \delta}{k^2}.$$ Summing over all $1 \le s_1, s_2 \le k$, we get the stated bound. Next, we state the main technical lemma of this section. To state the lemma, we define the function $\mathsf{Proj}: \mathbb{R}^k \rightarrow \Delta_k$ such that $\mathsf{Proj}(x) = y$ if $y$ is the closest point (in Euclidean distance) to $x$ in $\Delta_k$. While the authors are aware that technically, we require $\mathsf{Proj}$ to be quantified by the parameter $k$, the relevant $k$ shall always be clear from the context.  \[lem:Boosting\] Let $F: \mathbb{R}^n \rightarrow \Delta_k$ and let $g_1, \ldots, g_m : \mathbb{R}^n \rightarrow \mathbb{R}^k$ be an orthonormal sequence of functions under the standard $n$-dimensional Gaussian measure $\gamma_n$. Here the function $g_1 : x \mapsto (1,\ldots, 1)$. Then, for any $\delta>0$, there exists a function $F_{\mathsf{proj}} : \mathbb{R}^n \rightarrow \Delta_k$ of the form $F_{\mathsf{proj}} = \mathsf{Proj} (\sum_{i=1}^m \kappa_i g_i)$ satisfying $$\sum_{i=1}^m ({{\bf E}}[g_i F] - {{\bf E}}[g_i F_{\mathsf{proj}}])^2 \le \delta.$$ Further, $\sum_{i=1}^m \Vert \kappa_i \Vert_2^2 \le \delta^{-2}$. Before proving Lemma \[lem:Boosting\], we first see why this lemma is useful. In particular, we have the following corollary. Essentially, the corollary says that given $f, g: \mathbb{R}^n \rightarrow \Delta_k$, there are functions $f_{\mathsf{sm}}$ and $g_{\mathsf{sm}}$ such that (i) the low-level Hermite spectrum of $f$ (resp. $g$) is close to $f_{\mathsf{sm}}$ (resp. $g_{\mathsf{sm}}$) (ii) Both ${f_{\mathsf{sm}}}$ and ${g_{\mathsf{sm}}}$ are obtained by applying the function $\mathsf{Proj}$ on a low-degree polynomial. In essence, we are obtaining *simple* functions ${f_{\mathsf{sm}}}$ and ${g_{\mathsf{sm}}}$ which simultaneously (i) have the same low-level Hermite spectrum as $f$ and $g$ (ii) and have range $\Delta_k$. \[corr:Boosting\] Given function $f: \mathbb{R}^n \rightarrow [k]$, $d \in \mathbb{N}$ and error parameter $\delta>0$, there is a function ${f_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$ which has the following properties: 1. The function ${f_{\mathsf{sm}}}$ has the following form: $${f_{\mathsf{sm}}}(x) = \mathsf{Proj}\bigg(\sum_{|S| \le d} \alpha_{f,s} H_S(x)\bigg) ,$$ where $H_S(x)$ is the Hermite polynomial corresponding to the multiset $S$. 2. $\sum_{|S| \le d} \Vert \alpha_{f,S} \Vert_2^2 \le \delta^{-2}$. 3. Define $\beta_{f,S} = \mathbf{E}[{f_{\mathsf{sm}}}(x) \cdot H_S(x)]$. Then, $\sum_{|S| \le d} \Vert \beta_{f,S} - \alpha_{f,S} \Vert_2^2 \le \delta$. We note that for a scalar-valued function $H_S$ and a vector-valued function ${f_{\mathsf{sm}}}$, we compute $\mathbf{E}[{f_{\mathsf{sm}}}\cdot H_S]$ pointwise for each coordinate of the vector valued function ${f_{\mathsf{sm}}}$. The proof of this corollary follows straightaway by instantiating Lemma \[lem:Boosting\] with $\{g_1, \ldots, g_m\} = \{H_{S}\}_{|S| \le d}$ with $F=f$ and $F=g$. [Lemma \[lem:Boosting\]]{} We will prove this lemma via an iterative argument. We will define a sequence of functions $\{F_t\}_{t \ge 0}$ iteratively such that for all $t \ge 0$, $F_t : \mathbb{R}^n \rightarrow \Delta_k$. Define the vector $\beta \in \mathbb{R}^m$ by $\beta_j = \langle F, g_j \rangle$. Also, for every $t \ge 0$, we will define $\beta_t\in \mathbb{R}^m$ by $\beta_{t,j} = \langle F_t, g_j \rangle$. The iterative process has the following property: If for any $t$, $\Vert \beta_t - \beta \Vert_2^2 \le \delta$, then we terminate the process. Else, we modify $F_t$ to obtain the function $F_{t+1}$. We now define the initial function $F_0$ as well as the modification to obtain $F_{t+1}$ from $F_t$ (when $t \ge 0$). The function $F_0 : \mathbb{R}^n \rightarrow \Delta_k$ is defined as $F_0 : x \rightarrow (1/k, \ldots, 1/k)$. Next, given $F_t$, we define $F_{t+1}$. To do this, we will also need to define an auxiliary sequence of functions $\{G_t \}_{t \ge 0}$ where $G_0 = F_0$. The iterative process is defined in Figure \[fig:DS\]. ------------------------------------------------------------------------ 1. Define $\rho_t = \Vert \beta_t - \beta \Vert_2$. 2. If $\rho_t^2 \le \delta$, then stop the process. Else, we define $J_t = \sum_{j=1}^m (\beta -\beta_t)_j \cdot g_j$. 3. Define $G_{t+1} = G_t+ J_t/2$. Define $F_{t+1} = \mathsf{Proj}(G_{t+1})$ and $t \leftarrow t+1$. Go to Step 1. ------------------------------------------------------------------------  \ It is clear that if this process terminates at step $t=t_0$, then the function $F_{\mathsf{proj}}= F_{t_0}$ satisfies the required properties. Thus, we now need to bound the convergence rate of the process. To do this, we introduce a potential function $\Psi(t)$ defined as follows: $$\Psi(t) = \mathbf{E}[ \langle F - F_t , F - 2 G_t + F_t \rangle].$$ The basic observation here is that $\Psi(0) = O(1)$. We will prove two main lemmas. The first will prove that in every iteration of the process in Figure \[fig:DS\], $\Psi(t)$ decreases by a fixed amount. The second is that $\Psi(t)$ is always non-negative. These two facts, in conjunction, automatically imply an upper bound on the maximum number of steps in the algorithm.  \[clm:potential-decrease\] $$\mathbf{E}[\langle F - F_t, J_t \rangle] = \rho_t^2.$$ By orthogonality of the functions $\{g_j \}_{j=1}^m$, $$\mathbf{E}[\langle F - F_t, J_t \rangle] = \sum_{j=1}^m (\beta -\beta_t)_j \mathbf{E}[\langle g_j , F- F_t \rangle] = \sum_{j=1}^m (\beta -\beta_t)_j \cdot (\beta-\beta_t)_j = \Vert \beta - \beta_t \Vert_2^2.$$ We now recall a basic fact about projective maps (see, e.g. [@CheneyGoldstein:59 Theorem 3]).  \[fact:convex\] Let $C$ be a closed, convex set and let $\mathsf{Proj}_C : \mathbb{R}^n \rightarrow C$ be defined as $x \mapsto \arg \min_{y \in C} \Vert x - y \Vert_2$. Then the map $\mathsf{Proj}_C$ is uniquely defined, and always contractive i.e. for any $z, z' \in \mathbb{R}^n$, $\Vert \mathsf{Proj}_C (z) - \mathsf{Proj}_{C}(z') \Vert_2 \le \Vert z- z' \Vert_2$. Moreover, for any $x \in C$ and any $z \in {{\bf R}}^n$, $\langle z - \mathsf{Proj}_C(z), x - \mathsf{Proj}_C(z)\rangle \le 0$.  \[clm:non-negative\] For all $t$, $\Psi(t) \geq 0$. $$\begin{aligned} \Psi(t) &=& \mathbf{E}[ \langle F - F_t , F - 2 G_t + F_t \rangle] \\ &=& \mathbf{E}[ \langle F - F_t , F - F_t \rangle] + 2 \cdot \mathbf{E}[ \langle F - F_t , F_t - G_t \rangle] . \end{aligned}$$ The first term is clearly non-negative. The second is non-negative by Fact \[fact:convex\], taking $z = G_t$ and $x = F$. The next lemma shows that the potential function always decreases by a fixed quantity.  \[lem:descent\] $$\Psi(t+1) - \Psi(t) \le -\frac{\rho_t^2}{4}.$$ $$\begin{aligned} \Psi(t+1) - \Psi(t) &=& \mathbf{E}[ \langle F - F_{t+1} , F - 2 G_{t+1} + F_{t+1} \rangle]- \mathbf{E}[ \langle F - F_t , F - 2 G_t + F_t \rangle] \\ &=& \mathbf{E}[\langle F - F_t, 2 (G_t - G_{t+1}) \rangle] + \mathbf{E}[\langle F_{t+1} - F_t , 2 G_{t+1} - F_t - F_{t+1} \rangle] \\ &=& \mathbf{E}[\langle F - F_t, - J_t \rangle] + \mathbf{E}[\langle F_{t+1} - F_t, 2 G_{t+1} - F_t - F_{t+1} \rangle] \\ &=& - \rho_t^2 + \mathbf{E}[\langle F_{t+1} - F_t , 2 G_{t+1} - F_t - F_{t+1} \rangle] ~\textrm{(applying Claim~\ref{clm:potential-decrease})} \\ &=& -\rho_t^2 + 2 \cdot \mathbf{E}[\langle F_{t+1} - F_t , G_{t+1} - F_{t+1} \rangle] + \mathbf{E}[\langle F_{t+1} - F_t , F_{t+1} - F_{t} \rangle] \\ &=& -\rho_t^2 + \mathbf{E}[\Vert F_{t+1} - F_{t} \Vert_2^2] + 2 \cdot \mathbf{E}[\langle F_{t+1} - F_t , G_{t+1} - F_{t+1} \rangle] \\ &\le& -\rho_t^2 + \mathbf{E}[ \Vert G_{t+1} - G_{t} \Vert_2^2 ] + 2 \cdot \mathbf{E}[\langle F_{t+1} - F_t , G_{t+1} - F_{t+1} \rangle]~ \textrm{(applying Fact~\ref{fact:convex})} \\ &=& - \frac{3 \rho_t^2}{4} + 2 \cdot \mathbf{E}[\langle F_{t+1} - F_t , G_{t+1} - F_{t+1} \rangle]\end{aligned}$$ It remains to show that $\mathbf{E}[\langle F_{t+1} - F_t , G_{t+1} - F_{t+1} \rangle] \le \frac{\rho_t^2}{4}$. Indeed, the Cauchy-Schwarz inequality yields $$\begin{aligned} \|F_{t+1} - F_t\|_2 \|G_{t+1} - G_t\|_2 &\ge \langle G_{t+1} - G_t, F_{t+1} - F_t \rangle \\ &= \langle G_{t+1} - F_{t+1}, F_{t+1} - F_t \rangle + \langle F_{t+1} - F_t, F_{t+1} - F_t\rangle + \langle F_t - G_t, F_{t+1} - F_t\rangle\end{aligned}$$ In the last line above, the second term is obviously non-negative. Moreover, the third term is non-negative by Fact \[fact:convex\] (take $z = G_t$ and $x = F_{t+1}$). Hence, $$\langle G_{t+1} - F_{t+1}, F_{t+1} - F_t \rangle \le \|F_{t+1} - F_t\|_2 \|G_{t+1} - G_t\|_2 \le \|G_{t+1} - G_t\|_2^2 = \frac{\rho_t^2}{4},$$ where the second inequality follows from Fact \[fact:convex\]. Combining Claim \[clm:non-negative\] and Lemma \[lem:descent\], we obtain that the iterative process described in Figure \[fig:DS\] stops in at most $4/\delta$ steps. If the above iteration stops after $t=t_0$ steps, we let $F_{\mathsf{proj}}= F_{t_0}$. Note that $ F_{\mathsf{proj}} = \mathsf{Proj}(\sum_{0 \le t < t_0} J_t/2)$. Thus, it is clear that $F_{\mathsf{proj}} = \mathsf{Proj}(\sum_{i=1}^m \kappa_i g_i)$. To bound $\sum_{i=1}^m \Vert \kappa_i \Vert_2^2$, note that $$\sum_{i=1}^m \Vert \kappa_i \Vert_2^2 = \Vert \sum_{0 \le t < t_0} J_t/2 \Vert_2^2 \le t_0 \cdot \sum_{0 \le t < t_0} \Vert J_t/2\Vert_2^2 \le t_0^2 \cdot \max_t \Vert J_t/2 \Vert_2^2 \le t_0^2.$$ The very last inequality uses the fact that $\Vert J_t \Vert_2 \le \Vert (F_t - F) \Vert_2 \le 1$. Plugging the upper bound of $O(1/\delta^2)$ on $t_0^2$, we obtain that $\sum_{i=1}^m \Vert \kappa_i \Vert_2^2 \le O(1/\delta^2)$. This concludes the proof.  \[corr:fsm\] For $t>0$, error parameter $\delta>0$ and any function $f: \mathbb{R}^n \rightarrow [k]$, there is a function ${f_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$ such that for $d =(2/t) \cdot \log(k^2/\delta)$, we have the following: 1. $\Vert \mathbf{E}[{f_{\mathsf{sm}}}] -\mathbf{E}[f] \Vert_1 \le \delta$. 2. The function ${f_{\mathsf{sm}}}=\mathsf{Proj}(p_{f,1}(x), \ldots, p_{f,k}(x))$ where for all $1 \le s \le k$, $p_{f,s} : \mathbb{R}^n \rightarrow \mathbb{R}$ are polynomials of degree $d$ and $\mathsf{Var}(p_{f,s}) \le k^8/\delta^4$. 3. For any $g: \mathbb{R}^n \rightarrow [k]$ and the corresponding function ${g_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$, we have $\sum_{1 \le s_1, s_2 \le k} |\mathbf{E}[f_{\mathsf{sm},s_1} P_t g_{\mathsf{sm},s_2}]-\mathbf{E}[f_{s_1} P_t g_{s_2}]| \leq \delta$. Given the function $f: \mathbb{R}^n \rightarrow [k]$, we apply Corollary \[corr:Boosting\] to obtainthe function ${f_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$ where $${f_{\mathsf{sm}}}=\mathsf{Proj}(p_{f,1}(x), \ldots, p_{f,k}(x)),$$ where for all $1 \le s \le k$, $p_{f,s}: \mathbb{R}^n \rightarrow \mathbb{R}$ are polynomials of degree $d = (1/t) \cdot \log(k^2/\delta)$ such that $\mathsf{W}^{\le d} [({f_{\mathsf{sm}}}- f)] \le \delta^2/k^4$. Further, for each $1 \le s \le k$, $\mathsf{Var}(p_{f,s}) \le (k^8/\delta^4)$. This immediately implies both items 1 and 2. To prove Item 3, note that we also have $\mathsf{W}^{\le d} [({g_{\mathsf{sm}}}- g)] \le \delta^2/k^4$. Applying Claim \[clm:noise-degree\], we obtain that $\sum_{1 \le s_1, s_2 \le k} |\mathbf{E}[f_{\mathsf{sm},s_1} P_t g_{\mathsf{sm},s_2}]-\mathbf{E}[f_{s_1} P_t g_{s_2}]| \leq \delta$. This proves Item 3. This completes the first step in the outline of Lemma 5: we have replaced arbitrary functions by projections of polynomials. Bernstein approximation ----------------------- The next step in the proof of Lemma 5 is the removal of the projection. The basic idea is just to approximate the projection map by a polynomial. Then, the projection of a polynomial becomes the composition of two polynomials, which is still a polynomial. For $0 \le k \le d$, efine $p_{k,d}(x) = \binom{d}{k} x^k (1-x)^{d-k}$. For a function $f: [0, 1]^\ell \to \mathbb{R}$, define the polynomial $\mathsf{BP}_{f, d_1, \dots, d_\ell}$ by $$\mathsf{BP}_{f, d_1, \dots, d_\ell}(x) = \sum_{k_1, \dots, k_\ell} f\left(\frac{k_1}{d_1}, \dots, \frac{k_\ell}{d_\ell}\right) p_{k_1,d_1}(x_1) \cdots p_{k_\ell,d_\ell}(x_\ell).$$ We call $\mathsf{BP}_{f,d_1, \dots, d_\ell}$ the multivariate Bernstein approximation for $f$ with degrees $(d_1, \dots, d_\ell)$. \[thm:Bernstein\]**Multivariate Bernstein approximations** Let $f: [0,1]^\ell \rightarrow \mathbb{R}$ be a $L$-Lipschitz function in $[0,1]^{\ell}$. In other words, $\Vert f(x) - f(y) \Vert_2 \le L \cdot \Vert x - y \Vert_2$. Then $\mathsf{BP}_{f, d_1, \ldots, d_\ell}$ satisfies the inequality $$\sup_{z \in [0,1]^{\ell}} \big| f(z) - \mathsf{BP}_{f, d_1, \ldots, d_\ell}(z) \big| \le \frac{L}{2} \cdot \bigg( \sum_{j=1}^\ell \frac{1}{d_j}\bigg)^{1/2}$$ The proof of Theorem \[thm:Bernstein\] is folklore; we provide one for completeness. Fix $z \in [0, 1]^\ell$. Note that each $p_{k_i,d_i}(z_i)$ is non-negative, and that $\sum_{k_i=0}^{d_i} p_{k_i,d_i}(z_i) = 1$. Hence, $$\begin{aligned} f(z) - \mathsf{BP}_{f,d_1, \dots, d_\ell}(z) & = \sum_{k_1, \dots, k_\ell} \left[f(z) - f\left(\frac{k_1}{d_1}, \dots, \frac{k_\ell}{d_\ell}\right)\right] p_{k_1,d_1}(z_1) \cdots p_{k_\ell,d_\ell}(z_\ell) \\ & \le L \sum_{k_1, \dots, k_\ell} \left\|z - \left(\frac{k_1}{d_1}, \dots, \frac{k_\ell}{d_\ell}\right)\right\|_2 p_{k_1,d_1}(z_1) \cdots p_{k_\ell,d_\ell}(z_\ell) \\ & \le L \left[\sum_{k_1, \dots, k_\ell} \left\|z - \left(\frac{k_1}{d_1}, \dots, \frac{k_\ell}{d_\ell}\right)\right\|_2^2 p_{k_1,d_1}(z_1) \cdots p_{k_\ell,d_\ell}(z_\ell)\right]^{1/2} \\ & = L \left[\sum_{i=1}^\ell \sum_{k_i=0}^{d_i} \Big(z_i - \frac{k_i}{d_i}\Big)^2 p_{k_i,d_i}(z_i)\right]^{1/2}. \end{aligned}$$ Finally, note that $\sum_{k=0}^d (x - k/d)^2 p_{k,d}(x)$ is just the variance of a binomial random variable with $d$ trials and success probability $x$. This is bounded by $\frac{1}{4d}$. Plugging in this bound for each $i$ separately completes the proof. Rescaling the function, we have the following corollary. To state this corollary, we let $B(x, r) = \{z : \Vert z - x \Vert_2 \le r\}$ i.e. the $\ell_2$ of radius $r$ at $x$. \[corr:Bernstein\] Let $f: B(x,r) \rightarrow \mathbb{R}$ be a $1$-Lipschitz function (where $B(x,r) \subseteq \mathbb{R}^\ell$). Then, given any error parameter $\eta>0$, there is a polynomial $p_{f, r, \eta}$ whose degree in every variable is at most $d_{B}(\eta, r,\ell) = \ell \cdot 4r^2 \cdot (1/\eta^2) $ such that $$\sup_{z \in B(x,r)} \big|p_{f, r, \eta}(z) - f(z) \big| \le \eta.$$ To prove this, we will rely on Theorem \[thm:Bernstein\]. First, define $B_{\infty}(x,r) =\{z: \Vert z-x\Vert_\infty \le r\}$. We extend $f$ to $B_{\infty}(x,r)$ as follows: $ f(z) = f(\mathsf{Proj}_{B(x,r)}(z)). $ Note that the extension is $1$-Lipschitz (using Fact \[fact:convex\]). Define the function $g: [0,1]^\ell \rightarrow \mathbb{R}$ as $$g(z) = f\bigg(x + \bigg(z-\mathbf{\frac{1}{2}}\bigg) \cdot 2r \bigg).$$ Here $\mathbf{\frac{1}{2}}$ is the point in $\mathbb{R}^{\ell}$ which is $1/2$ in every coordinate. It is easy to see that the function $g$ is $2r$-Lipschitz. Thus, if we choose the function $\mathsf{BP}_{g,d_1, \ldots, d_{\ell}}$, then we have $$\sup_{z \in [0,1]^{\ell}} \big|\mathsf{BP}_{g,d_1, \ldots, d_{\ell}} - g(z) \big| \le 2r \cdot \bigg(\sum_{j=1}^{\ell} \frac{1}{d_j} \bigg)^{1/2}.$$ In particular, we set all the degrees $d_1 = \ldots =d_{\ell} = \ell \cdot 4r^2 \cdot (1/\eta^2)$, then $\sup_{z \in [0,1]^{\ell}} \big|\mathsf{BP}_{g,d_1, \ldots, d_{\ell}} - g(z) \big| \le \eta$. Thus, if we set $p_{f,r,\eta}(z)$ as $$p_{f,r,\eta}(z) = \mathsf{BP}_{g,d_1, \ldots, d_{\ell}} \bigg( \frac{z-x}{2r} + \mathbf{\frac{1}{2}}\bigg).$$ It is clear that the polynomial $p_{f,r,z}$ satisfies $\sup_{z \in B(x,r)} \big|p_{f, r, \eta}(z) - f(z) \big| \le \eta$. We next modify the function ${f_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$ obtained in Corollary \[corr:fsm\] to obtain the function ${f_{\mathsf{sm}}}' : \mathbb{R}^n \rightarrow \mathbb{R}^k$ which is a (i) low-degree polynomial and (ii) ${f_{\mathsf{sm}}}$ is close to ${f_{\mathsf{sm}}}'$ with high probability on the Gaussian measure $\gamma_n$. \[lem:smoothing-1\] Given the function ${f_{\mathsf{sm}}}: \mathbb{R}^n \rightarrow \Delta_k$ from Corollary \[corr:fsm\], there is a function ${f_{\mathsf{sm}}}': \mathbb{R}^n \rightarrow \mathbb{R}^k$ such that ${f_{\mathsf{sm}}}' = (p'_{f,1}(x), \ldots, p'_{f,k}(x))$ where for all $1 \le s \le k$, $p'_{f,s}: \mathbb{R}^n \rightarrow \mathbb{R}$ are polynomials satisfying the following conditions: 1. For $1 \le s \le k$, the polynomials $\{p'_{f,s}\}$ have degree $d' = \log^{d}(dk/\delta) \cdot \mathsf{poly}(k/\delta) \cdot d$ where $d$ is the degree appearing in Corollary \[corr:fsm\]. 2. $\Pr_{x \sim \gamma_n} [\Vert {f_{\mathsf{sm}}}(x) - {f_{\mathsf{sm}}}'(x) \Vert_\infty \le \delta/4] \le \delta/2$. Let the function ${f_{\mathsf{sm}}}(x) = \mathsf{Proj}(p_{f,1}(x), \ldots, p_{f,k}(x))$. Since all the polynomials are degree $d$ and have variance at most $\sigma_{\mathsf{sm}}^2 = k^{8}/\delta^4$, using Theorem \[thm:hyper\], we obtain the following: $$\label{eq:inf-ball-1} \Pr_{x \sim \gamma_n} \sup_{1 \le s \le k} [|p_{f,s} - \mathbf{E}[p_{f,s}]| \le \log^{d/2} (2dk/\delta) \cdot \sigma_{\mathsf{sm}}] \le \frac{\delta}{2}.$$ Define the point $\boldsymbol{\mu}_{sm,f} = (\mathbf{E}[p_{f,1}], \ldots, \mathbf{E}[p_{f,s}])$. Also, let $r_{sm} = \log^{d/2} (2dk/\delta) \cdot \sigma_{\mathsf{sm}}$. Since the projection from ${{\bf R}}^k$ to $\Delta_k$ is Lipschitz, Corollary \[corr:Bernstein\] implies that there exist polynomials $p_{\mathsf{sm},s}: \mathbb{R}^{k} \rightarrow \mathbb{R}$ (for $1 \le s\le k$) whose degree in every variable is at most $k \cdot 4 r_{sm}^2 \cdot 16/\delta^2 = \log^d(dk/\delta) \cdot \mathsf{poly}(k/\delta)$, and which satisfy $$~\label{eq:inf-ball} \textrm{for all } z \in B(\boldsymbol{\mu}_{sm,f}, r_{sm}) \textrm{, we have } \ |p_{\mathsf{sm},s}(z) - \mathsf{Proj}_s(z)| \le \frac{\delta}{4}$$ Let $p_{\mathsf{sm}}: \mathbb{R}^k \rightarrow \mathbb{R}^k$ be defined as the map $p_{\mathsf{sm}}(x) = (p_{\mathsf{sm},1}(x), \ldots, p_{\mathsf{sm},k}(x))$. Recall that ${f_{\mathsf{sm}}}= \mathsf{Proj} (p_{f,1}(x), \ldots, p_{f,k}(x))$. We define $p'_f= p_{\mathsf{sm}} \circ (p_{f,1}, \ldots, p_{f,k})$. We now define ${f_{\mathsf{sm}}}' = (p'_{f,1}(x), \ldots, p'_{f,k}(x))$. It is clear that for $1 \le s \le k$, $p'_{f,s}$ is a polynomials of degree $\log^{d}(dk/\delta) \cdot \mathsf{poly}(k/\delta) \cdot d$. Likewise, combining (\[eq:inf-ball\]) and (\[eq:inf-ball-1\]), we obtain that $\Pr_{x \sim \gamma_n} [\Vert {f_{\mathsf{sm}}}(x) - {f_{\mathsf{sm}}}'(x) \Vert_\infty \le \delta/2] \le \delta/2$. Converting to PPFs ------------------ Before we finish the proof of Lemma \[lem:smoothing\], we will need to make a couple of elementary observations. First of all, observe that if $\alpha$ is uniformly random in $[0,1]$, then for any $x \in [0,1]$, $\mathbf{E}[\mathbf{1}_{x - \alpha \ge 0}] = x$. Here $\mathbf{1}_{x-\alpha \ge 0}$ denotes the function which is $1$ if $x-\alpha \ge 0$ and $0$ otherwise. Now, for any parameter $\eta>0$, define the distribution $\mathsf{Int}_\eta$ to be uniformly random over the set $\{i \cdot \eta\}_{i \ge 0} \cap [0,1]$. Then, we have the following simple claim.  \[clm:expectation-delta\] Let $\zeta>0$ and $y \in \Delta_{k,\zeta}$. Then, $$\bigg\Vert \mathop{\mathbf{E}}_{(\alpha_1,\ldots, \alpha_k) \sim \mathsf{Int}_\eta^k} \bigg[\sum_{s=1}^k \arg \max (\underbrace{0, \ldots, 0}_{s-1 \textrm{ times}}, y_s-\alpha_s,\underbrace{0, \ldots, 0}_{k-s \textrm{ times}} )\bigg] - y \bigg\Vert_1 \le 2( \zeta + k \cdot \eta).$$ Let the point closest to $y$ in $\Delta_k$ be $x$. Then, we have $\Vert x-y \Vert_1=\zeta$. We have the following: $$\bigg\Vert \mathop{\mathbf{E}}_{(\alpha_1,\ldots, \alpha_k) \sim \mathsf{Int}_\eta^k} \bigg[\sum_{s=1}^k \arg \max (\underbrace{0, \ldots, 0}_{s-1 \textrm{ times}}, x_s-\alpha_s,\underbrace{0, \ldots, 0}_{k-s \textrm{ times}} )\bigg] - x \bigg\Vert_1 \le k \cdot \eta.$$ Combining this with $\Vert x-y \Vert_1\le \zeta$, we obtain $$~\label{eq:inter} \bigg\Vert \mathop{\mathbf{E}}_{(\alpha_1,\ldots, \alpha_k) \sim \mathsf{Int}_\eta^k} \bigg[\sum_{s=1}^k \arg \max (\underbrace{0, \ldots, 0}_{s-1 \textrm{ times}}, x_s-\alpha_s,\underbrace{0, \ldots, 0}_{k-s \textrm{ times}} )\bigg] - y \bigg\Vert_1 \le k \cdot \eta +\zeta.$$ Next, for any $1 \le s\le k$, $$\Vert \mathop{\mathbf{E}}_{(\alpha_1,\ldots, \alpha_k) \sim \mathsf{Int}_\eta^k} \arg \max (\underbrace{0, \ldots, 0}_{s-1 \textrm{ times}}, x_s-\alpha_s,\underbrace{0, \ldots, 0}_{k-s \textrm{ times}} ) - \arg \max (\underbrace{0, \ldots, 0}_{s-1 \textrm{ times}}, y_s-\alpha_s,\underbrace{0, \ldots, 0}_{k-s \textrm{ times}} ) \Vert_1 \le |x_s-y_s| + \eta.$$ Summing over all $1 \le s \le k$ and combining with (\[eq:inter\]), we obtain the claim. [Lemma \[lem:smoothing\]]{} For $1 \le i \le \ell$, let $\{{f_{\mathsf{sm}}}^{'(i)}\}$ and $\{{g_{\mathsf{sm}}}^{'(i)}\}$ be the functions obtained by applying Corollary \[corr:fsm\] and Lemma \[lem:smoothing-1\] to the family of functions $\{f^{(i)}\}$ and $\{g^{(i)}\}$. In particular, let ${f_{\mathsf{sm}}}^{'(i)} = (p^{'(i)}_{f,1}, \ldots, p^{'(i)}_{f,k})$ and ${g_{\mathsf{sm}}}^{'(i)} = (p^{'(i)}_{g,1}, \ldots, p^{'(i)}_{g,k})$. For $ \eta>0$ (to be fixed later), let us define $f^{(i)}_1$ and $g^{(i)}_1$ as follows: $$f^{(i)}_1 = \sum_{s=1}^k \mathop{\mathbf{E}}_{(\alpha_1, \ldots, \alpha_k) \in \mathsf{Int}_{\eta}^k} \arg \max \big(\underbrace{0, \ldots, 0}_{s-1 \ \textrm{times}} , p^{'(i)}_{f,s} - \alpha_s, \underbrace{0, \ldots, 0}_{k-s \ \textrm{times}} \big)$$ $$g^{(i)}_1 = \sum_{s=1}^k \mathop{\mathbf{E}}_{(\alpha_1, \ldots, \alpha_k) \in \mathsf{Int}_{\eta}^k} \arg \max \big(\underbrace{0, \ldots, 0}_{s-1 \ \textrm{times}} , p^{'(i)}_{g,s} - \alpha_s, \underbrace{0, \ldots, 0}_{k-s \ \textrm{times}} \big)$$ We will now verify the properties of the construction.  \ **Proof of Items 1 and 2:** Both these items are straight forward from the construction. \ **Proof of Item 3:** By the second item of Lemma \[lem:smoothing-1\], we have $\Pr_{x \sim \gamma_n} [{f_{\mathsf{sm}}}^{'(i)}(x) \in \Delta_{k,k\delta/4}] \ge 1-\delta/2$. By applying Claim \[clm:expectation-delta\], we obtain that whenever ${f_{\mathsf{sm}}}^{'(i)}(x) \in \Delta_{k,k\delta/4}$, $f^{(i)}_1(x) \in \Delta_{k, O(k \delta + k \eta)}$. Thus, as long as $\eta \le \delta/k$, this proves Item 3 for $f^{(i)}_1$. The proof for $g^{(i)}_1$ is similar.  \ **Proof of Items 4 and 5:** We first observe that $\Pr_{x \sim \gamma_n}[\Vert {f_{\mathsf{sm}}}^{'(i)}(x) - {f_{\mathsf{sm}}}^{(i)}(x) \Vert_1 \le k \cdot \delta/4 ] \ge 1- \delta/2$. By applying Claim \[clm:expectation-delta\], we obtain that $\Pr_{x \sim \gamma_n} [\Vert f^{(i)}_1(x) - {f_{\mathsf{sm}}}^{(i)}(x) \Vert_1 \le O(k \delta + k \eta)] \ge 1- \delta/2$. However, note that by definition, $\Vert f^{(i)}_1(x) - {f_{\mathsf{sm}}}^{(i)}(x) \Vert_\infty \le k$. This implies that $\mathbf{E}[\Vert {f_{\mathsf{sm}}}^{'(i)} (x) - f^{(i)}_1(x) \Vert_1] = O(k\delta + k \eta)$. As long as $\eta \le \delta/k$, we have $\mathbf{E}[\Vert {f_{\mathsf{sm}}}^{(i)} (x) - f^{(i)}_1(x) \Vert_1] = O(k\delta)$. Combining with the guarantees of Corollary \[corr:Boosting\] yields Items 4 and 5.  \ **Proof of Item 6:** To prove Item 6, note that for any $1 \le s \le k$ and $\alpha_s \in [0,1]$, $$\arg \max \big(\underbrace{0, \ldots, 0}_{s-1 \ \textrm{times}} , p^{'(i)}_{f,s} - \alpha_s, \underbrace{0, \ldots, 0}_{k-s \ \textrm{times}} \big) = \mathsf{PPF}_{p^{'(i)}_{f,s} - \alpha_s,s}.$$ Thus, if we define $p^{(i)}_{s,j,1} = p^{'(i)}_{f,s}- \eta \cdot j$ and $p^{(i)}_{s,j,2} = p^{'(i)}_{g,s} - \eta \cdot j$, then $$f_1^{(i)} = \sum_{s=1}^k \sum_{j=0}^{m} \frac{1}{m} \mathsf{PPF}_{p^{(i)}_{s,j,1} , s} \ \textrm{and} \ g_1^{(i)} = \sum_{s=1}^k \sum_{j=0}^{m} \frac{1}{m} \mathsf{PPF}_{p^{(i)}_{s,j,2} , s},$$ where $m = \lceil 1/\eta \rceil$. As $\eta \leq \delta /k$, $m = O(k/\delta)$. By Lemma \[lem:smoothing-1\], $\mathsf{deg}(p^{'(i)}_{f,s})$ and $\mathsf{deg}(p^{'(i)}_{g,s})$ is at most $d' = d \cdot \mathsf{poly}(k/\delta) \cdot \log^{d} (d k /\delta)$ where $d =2/t \cdot \log(dk/\delta)$ (coming from Corollary \[corr:Boosting\]). If we set $d_0(t,k,\delta) = d'$, then $\mathsf{deg}(p^{'(i)}_{f,s})$ and $\mathsf{deg}(p^{'(i)}_{g,s})$ is at most $d_0(t,k,\delta)$. As $\mathsf{deg}(p^{(i)}_{s,j,1}) = \mathsf{deg}(p^{'(i)}_{f,s})$ and $\mathsf{deg}(p^{(i)}_{s,j,2}) = \mathsf{deg}(p^{'(i)}_{g,s})$, this proves Item 6. (We can make the PPFs balanced by applying Fact \[fact:balanced\]). Construction of junta polynomials ================================= This section is dedicated to the proof of Lemma \[lem:junta-construction\]. To prove this lemma, we will first recall the following important result from [@DMN16a] (Theorem 41 in that paper).  \[thm:junta-construct\] Let $p_1, \ldots, p_{\ell}: \mathbb{R}^n \rightarrow \mathbb{R}$ be degree-$d$ polynomials and for $\delta>0$, the following two conditions: (i) For all $1 \le s \le \ell$, $\mathsf{Var}(p_s)=1$ and (ii) For all $1 \le s \le \ell $, $|\mathbf{E}[p_s]| \le \log^{d/2} (k \cdot d /\delta)$. For $1 \le s \le \ell$ and $t>0$, define $u_s: \mathbb{R}^{2n} \rightarrow \mathbb{R}$ as follows: $u_s(x,y) = p_s(e^{-t} x + \sqrt{1-e^{-2t}} y)$. Then, there is an explicitly computable $n_0 = n_0(\ell, d, \xi)$ and polynomials $r_1, \ldots, r_{\ell}: \mathbb{R}^{n_0} \rightarrow \mathbb{R}$ with the following properties: For $1 \le s \leq \ell$, define $v_s: \mathbb{R}^{2n_0} \rightarrow \mathbb{R}$ as $v_s(x,y) = r_s(e^{-t} x + \sqrt{1-e^{-2t}} y)$. Then, for $ 1 \le s, s' \le \ell$, 1. $\big|\Pr_{x \sim \gamma_n} [{p}_s \ge 0] - \Pr_{x \sim \gamma_n} [{r}_s \ge 0] \big|\le \xi$. 2. $\big|\Pr_{x,y \sim \gamma_{n}} [{u}_s \ge 0] - \Pr_{x,y \sim \gamma_{n_0}} [{v}_s \ge 0] \big|\le \xi$. 3. $\big|\Pr_{x \sim \gamma_n} [{p}_s \cdot {p}_{s'} \ge 0] - \Pr_{x \sim \gamma_{n_0}} [{r}_s \cdot {r}_{s'} \ge 0] \big|\le \xi$. 4. $\big|\Pr_{x,y \sim \gamma_n} [{u}_s\cdot {u}_{s'} \ge 0] - \Pr_{x,y \sim \gamma_{n_0}} [{v}_s \cdot {v}_{s'} \ge 0] \big|\le \xi$. 5. $\big|\Pr_{x,y \sim \gamma_n} [{p}_s\cdot {u}_{s'} \ge 0] - \Pr_{x,y \sim \gamma_{n_0}} [{v}_s \cdot {v}_{s'} \ge 0] \big|\le \xi$. We now derive an additional property of the polynomials $\{p_s\}_{1 \le s \le \ell}$ and $\{r_s \}_{1 \le s \le \ell}$ defined in Theorem \[thm:junta-construct\] which will be useful later.  \[corr:const-and-1\] Let $p_1, \ldots, p_{\ell}: \mathbb{R}^n \rightarrow \mathbb{R}$ and $u_1, \ldots, u_{\ell}: \mathbb{R}^n \rightarrow \mathbb{R}$ be as defined in Theorem \[thm:junta-construct\]. Then, for any $1 \le s, s' \le k$, $$\big|\Pr_{x \sim \gamma_n} [({p}_s (x) \ge 0 ) \wedge ( {p}_{s'}(x) \ge 0)] - \Pr_{x \sim \gamma_{n_0}} [({r}_s (x) \ge 0 ) \wedge ( {r}_{s'}(x) \ge 0)] \big| \le 2\xi.$$ The main observation here is that if $A , B \not =0$, then $$\mathbf{1}[A \ge 0 ] \cdot \mathbf{1}[ B \ge 0] = \frac{1}{2} \big(\mathbf{1}[A \cdot B \ge 0] +\mathbf{1}[A \ge 0]+\mathbf{1}[ B \ge 0]-1\big) .$$ Now, note that because $p_s$, $p_{s'}$, $r_{s}$ and $r_{s'}$ are degree-$d$ polynomials, any of these functions vanish over the Gaussian measure with probability $0$. Thus, $$\begin{aligned} \Pr_{x \sim \gamma_n} [({p}_s (x) \ge 0 ) \wedge ( {p}_{s'}(x) \ge 0)] &=& \frac{1}{2} \big(\Pr_{x \sim \gamma_n} [{p}_s (x) \ge 0 ] + \Pr_{x \sim \gamma_n} [ {p}_{s'}(x) \ge 0] + \Pr_{x \sim \gamma_n} [ p_{s} \cdot {p}_{s'}(x) \ge 0]-1\big) \nonumber \\ \Pr_{x \sim \gamma_{n_0}} [({r}_s (x) \ge 0 ) \wedge ( {r}_{s'}(x) \ge 0)] &=& \frac{1}{2} \big(\Pr_{x \sim \gamma_{n_0}} [{r}_s (x) \ge 0 ] + \Pr_{x \sim \gamma_{n_0}} [ {r}_{s'}(x) \ge 0] + \Pr_{x \sim \gamma_{n_0}} [ r_{s} \cdot {r}_{s'}(x) \ge 0]-1\big) \nonumber \end{aligned}$$ Combining the above equations with items 1 and 3 in Theorem \[thm:junta-construct\] yields the corollary.  \ We now describe the proof of Lemma \[lem:junta-construction\]. [Lemma \[lem:junta-construction\]]{} Let us consider the collection of degree-$d_0$ polynomials $\{p_{s,j,1}^{(i)}\}_{1 \le i \le \ell, 1 \le s \le k, 1 \le j \le m} \cup \{p_{s,j,2}^{(i)}\}_{1 \le i \le \ell,1 \le s \le k, 1 \le j \le m}$. We now apply Theorem \[thm:junta-construct\] to obtain polynomials $\{r_{s,j,1}^{(i)}\}_{1 \le i \le \ell,1 \le s \le k, 1 \le j \le m} \cup \{r_{s,j,2}^{(i)}\}_{1 \le i \le \ell,1 \le s \le k, 1 \le j \le m}$ with $\xi= \delta/(40 k^2)$. We now define $$f^{(i)}_{\mathsf{junta}} = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{r_{s,j,1}^{(i)},s}(x) \ , \ g^{(i)}_{\mathsf{junta}} = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathsf{PPF}_{r_{s,j,2}^{(i)},s}(x)$$ We now verify the properties of the construction.  \ **Proof of Item 1:** Observe that for $1 \le s\le k$, we have the following $$\mathbf{E}[(f_{1}^{(i)}(x))_s] = \sum_{j=1}^m \frac{1}{m} \cdot \mathbf{E}_{x} [\mathsf{PPF}_{p_{s,j,1}^{(i)},s}(x)] = \sum_{j=1}^m \frac{1}{m} \cdot \Pr_{x} [p_{s,j,1}^{(i)}(x) \ge 0]$$ $$\mathbf{E}[(f_{\mathsf{junta}}^{(i)}(x))_s] = \sum_{j=1}^m \frac{1}{m} \cdot \mathbf{E}_{x} [\mathsf{PPF}_{r_{s,j,1}^{(i)},s}(x)] = \sum_{j=1}^m \frac{1}{m} \cdot \Pr_{x} [r_{s,j,1}^{(i)}(x) \ge 0]$$ Thus, we obtain $$\big| \mathbf{E}[(f^{(i)}_{1}(x))_s] - \mathbf{E}[(f^{(i)}_{\mathsf{junta}}(x))_s] \big| \le \sup_{1 \le j \le m} \big| \Pr_{x} [p_{s,j,1}^{(i)}(x) \ge 0]-\Pr_{x} [r_{s,j,1}^{(i)}(x) \ge 0] \big| \le \xi .$$ The penultimate inequality follows by applying Theorem \[thm:junta-construct\] to $p^{(i)}_{s,j,1}$ and $r^{(i)}_{s,j,1}$. This immediately implies that $\Vert \mathbf{E}[f^{(i)}_{1}(x)] - \mathbf{E}[f^{(i)}_{\mathsf{junta}}(x)] \Vert_1 \le k \cdot \xi \le \delta$. The proof for $ \big| \mathbf{E}[(g^{(i)}_{1}(x))_s] - \mathbf{E}[(g^{(i)}_{\mathsf{junta}}(x))_s] \big| \le \delta. $ is exactly identical.  \ **Proof of Item 2:** Like Item 1, we will only prove that $\Pr_{x} [f^{(i)}_{\mathsf{junta}}(x) \in \Delta_{k,\sqrt{\delta}}] \le \sqrt{\delta}$. The proof for $\Pr_{x} [g^{(i)}_{\mathsf{junta}}(x) \in \Delta_{k,\sqrt{\delta}}] \le \sqrt{\delta}$. To prove this, we first observe that for all $x$ both $f^{(i)}_1(x)$ and $f^{(i)}_{\mathsf{junta}}(x)$ always lie in the positive orthant and secondly, $\Vert f^{(i)}_1(x) \Vert_\infty, \Vert f^{(i)}_{\mathsf{junta}}(x) \Vert_\infty \le 1$. Next, $$\begin{aligned} \mathbf{E}[(\Vert f^{(i)}_1(x) \Vert_1 - 1)^2 ] &\le& \Pr_{x} [f^{(i)}_1(x) \in \Delta_{k,\delta}] \cdot \delta^2 + \Pr_{x} [f^{(i)}_1(x) \not \in \Delta_{k,\delta}] \cdot k^2 \nonumber \\ &\le& \delta^2 + k^2 \cdot \delta. \label{eq:square-diff}\end{aligned}$$ The first inequality uses $\sup_x \Vert f^{(i)}_1(x) \Vert_1 \le k$ and the second inequality uses $\Pr_{x} [f^{(i)}_1(x) \not \in \Delta_{k,\delta}] \le \delta$. Next, observe that $$\Vert f^{(i)}_1(x) \Vert_1 = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathbf{1}[p^{(i)}_{s,j,1}(x) \ge 0] \ \ , \ \ \Vert f^{(i)}_{\mathsf{junta}}(x) \Vert_1 = \sum_{s=1}^k \sum_{j=1}^m \frac{1}{m} \cdot \mathbf{1}[r^{(1)}_{s,j,1}(x) \ge 0]$$ This implies $$\begin{aligned} ( \Vert f^{(i)}_1(x) \Vert_1 - 1)^2 = \sum_{s=1}^k \sum_{s'=1}^k \sum_{j=1}^m \sum_{j'=1}^m \frac{1}{m^2} \mathbf{1}[p^{(i)}_{s,j,1}(x) \ge 0] \cdot \mathbf{1}[p^{(i)}_{s',j',1}(x) \ge 0] + 1 - \frac{2}{m}\sum_{s=1}^k \sum_{j=1}^m \mathbf{1}[p^{(i)}_{s,j,1}(x) \ge 0] .\label{eq:f-1}\end{aligned}$$ $$\begin{aligned} (\Vert f^{(i)}_{\mathsf{junta}}(x) \Vert_1 - 1)^2 = \sum_{s=1}^k \sum_{s'=1}^k \sum_{j=1}^m \sum_{j'=1}^m \frac{1}{m^2} \mathbf{1}[r^{(i)}_{s,j,1}(x) \ge 0] \cdot \mathbf{1}[r^{(i)}_{s',j',1}(x) \ge 0] + 1 - \frac{2}{m}\sum_{s=1}^k \sum_{j=1}^m \mathbf{1}[r^{(i)}_{s,j,1}(x) \ge 0] .\label{eq:f-junta-1}\end{aligned}$$ Recall that by construction, we have $$\sup_{1 \le s \le k, \ 1 \le j \le m} \big| \Pr_{x} [p_{s,j,1}^{(i)}(x) \ge 0]-\Pr_{x} [r_{s,j,1}^{(i)}(x) \ge 0] \big| \le \xi \label{eq:diff-p-r}$$ Applying Corollary \[corr:const-and-1\], we also obtain $$\sup_{1 \le s,s' \le k, \ 1 \le j,j' \le m} \big| \Pr_{x} [(p_{s,j,1}^{(i)}(x) \ge 0) \wedge (p_{s',j',1}^{(i)}(x) \ge 0)]-\Pr_{x} [(r_{s,j,1}^{(i)}(x) \ge 0) \wedge (r_{s',j',1}^{(i)}(x) \ge 0)] \big| \le 2\xi. \label{eq:diff-int-p-r}$$ Applying (\[eq:diff-p-r\]) and (\[eq:diff-int-p-r\]) to (\[eq:f-1\]) and (\[eq:f-junta-1\]), we obtain $$\big| \mathbf{E}[(\Vert f_{\mathsf{junta}}^{(i)}(x) \Vert_1 - 1)^2 ]- \mathbf{E}[(\Vert f^{(i)}_1(x) \Vert_1 - 1)^2 ]\big| \le 2 k^2 \cdot \xi + 2 k \cdot \xi \le \delta.$$ Combining this with (\[eq:square-diff\]), we obtain $ \mathbf{E}[(\Vert f^{(i)}_{\mathsf{junta}}(x) \Vert_1 - 1)^2 ] \le 2k^2 \cdot \delta. $ Applying Markov’s inequality, we obtain that $\Pr[| \ \Vert f^{(i)}_{\mathsf{junta}}(x) \Vert_1 - 1| > k\sqrt{\delta}] \le 2 k \sqrt{\delta}$. Since $f^{(i)}_{\mathsf{junta}}(x)$ lies in the positive orthant for any $x$, this proves Item 2. \ **Proof of Item 3:** To prove Item 3, we observe that for any $1 \le s_1, s_2 \le k$, $$\begin{aligned} \mathbf{E}[f_{1,s_1} P_t g_{1,s_2}] &=& \frac{1}{m^2}\sum_{j=1}^m \sum_{j'=1}^m \mathbf{E}\big[\mathsf{PPF}_{p^{(1)}_{s_1,j_1}}(x) P_t \ \mathsf{PPF}_{p^{(2)}_{s_2,j_2}}(x)\big] \nonumber \\ &=& \frac{1}{m^2}\sum_{j=1}^m \sum_{j'=1}^m \mathbf{E}_{x,y}\big[\mathsf{PPF}_{p^{(1)}_{s_1,j_1}}(x) \mathsf{PPF}_{p^{(2)}_{s_2,j_2}} (e^{-t} x + \sqrt{1-e^{-2t}} y)\big] \nonumber\\ &=&\frac{1}{m^2}\sum_{j=1}^m \sum_{j'=1}^m \Pr_{x, y} [ (p^{(1)}_{s_1,j_1}(x) \ge 0) \wedge (p^{(2)}_{s_2,j_2}(e^{-t} x + \sqrt{1-e^{-2t}} y) \ge 0)] \nonumber \\ &=& \frac{1}{m^2}\sum_{j=1}^m \sum_{j'=1}^m \Pr_{x, y} [ (p^{(1)}_{s_1,j_1}(x) \ge 0) \wedge (u^{(2)}_{s_2,j_2}(e^{-t} x + \sqrt{1-e^{-2t}} y) \ge 0)] \label{eq:fpg-1}.\end{aligned}$$ Likewise, we can obtain $$\label{eq:fpg-2} \mathbf{E}[f_{\mathsf{junta},s_1} P_t g_{\mathsf{junta},s_2}] = \frac{1}{m^2} \sum_{j=1}^m \sum_{j'=1}^m \Pr_{x, y} [ (r^{(1)}_{s_1,j_1}(x) \ge 0) \wedge (v^{(2)}_{s_2,j_2}(e^{-t} x + \sqrt{1-e^{-2t}} y) \ge 0)].$$ Combining (\[eq:fpg-1\]) and (\[eq:fpg-2\]) with Item 5 in Theorem \[thm:junta-construct\] yields $$\big| \mathbf{E}[f_{1,s_1} P_t g_{1,s_2}] -\mathbf{E}[f_{\mathsf{junta},s_1} P_t g_{\mathsf{junta},s_2}] \big | \le \xi.$$ This finishes the proof. Acknowledgments {#acknowledgments .unnumbered} --------------- We thank Pritish Kamath, Badih Ghazi and Madhu Sudan for pointing out that the $\ell=1$ case of Theorem \[thm:junta-strong\] is not sufficient to derive Theorem \[thm:junta\]. (An earlier version of this paper incorrectly claimed that it was.) We also thank the anonymous reviewers who pointed out the same gap. Reduction from arbitrary ${{\bf P}}$ to the Gaussian case =========================================================  \[section:GKS\] We first restate Theorem \[thm:junta\] below. [**2**]{} Suppose there exist $f, g: \mathcal{Z}^n \to [k]$ such that $(f({{\bf X}}^n), g({{\bf Y}}^n)) \sim {{\bf Q}}$. Then, there exist $n_0 = n_0 (|\mathbf{P}|, \delta)$ and $f_{\mathsf{junta}}, g_{\mathsf{junta}} : \mathcal{Z}^{n_0} \rightarrow [k]$ such that ${{\bf Q}}$ and the distribution of $(f_{\mathsf{junta}}({{\bf X}}^{n_0}), g_{\mathsf{junta}}({{\bf Y}}^{n_0}))$ are $\delta$-close in total variation distance. Moreover, $n_0$ is computable. Further, the functions $f_{\mathsf{junta}}$ and $g_{\mathsf{junta}}$ can be explicitly computed. Next, we restate Theorem \[thm:junta-strong\]. [**5**]{} Let ${{\bf P}}= ({{\bf X}},{{\bf Y}}) = \mathbf{G}_{\rho,2}$ and let $f^{(1)}, \ldots, f^{(\ell)}: \mathbb{R}^n \rightarrow [k]$ and $g^{(1)}, \ldots, g^{(\ell)}: \mathbb{R}^n \rightarrow [k]$ where we define ${{\bf Q}}_{i,j}$ as ${{\bf Q}}_{i,j} = (f^{(i)}({{\bf X}}^n), g^{(j)}({{\bf Y}}^n))$. Then, for every $\delta>0$, there is an explicitly defined constant $n_0 = n_0(\ell, k, \delta)$ and explicitly defined functions $f^{(1)}_{\mathsf{junta}}, \ldots, f^{(\ell)}_{\mathsf{junta}}: \mathbb{R}^{n_0} \rightarrow [k]$ and $g^{(1)}_{\mathsf{junta}}, \ldots, g^{(\ell)}_{\mathsf{junta}}: \mathbb{R}^{n_0} \rightarrow [k]$ such that for every $1 \le i, j \le \ell$, ${d_{\mathrm TV}}((f^{(i)}_{\mathsf{junta}}({{\bf X}}^{n_0}), g^{(j)}_{\mathsf{junta}}({{\bf Y}}^{n_0})), {{\bf Q}}_{i,j}) \le \delta$. The main purpose of this section is to show how proving Theorem \[thm:junta\] reduces to proving Theorem \[thm:junta-strong\]. While the reduction essentially follows just going over the steps in [@GKS16] *mutatis mutandis* (which in turn relies on standard tools from Boolean function analysis), for the purposes of clarity, we give a brief overview of the reduction here. First, let us fix some notation. 1. We recall the notion of maximal correlation coefficient: Namely, given a probability space $({{\bf X}}, {{\bf Y}})$, we let $\rho({{\bf X}}, {{\bf Y}})$ be defined as $$\rho({{\bf X}}, {{\bf Y}}) = \sup \mathbf{E}[\Psi_1({{\bf X}}) \cdot \Psi_2 ({{\bf Y}})],$$ where the supremum is taken over all functions which satisfy $\mathbf{E}[\Psi_1({{\bf X}})]= \mathbf{E}[\Psi_2({{\bf Y}})]=0$ and $\mathsf{Var}[\Psi_1({{\bf X}})] = \mathsf{Var}[\Psi_2({{\bf Y}})]=1$. 2. For a given set $H \subseteq [n]$, $x_H \in {{\bf X}}^{|H|}$ and function $f: {{\bf X}}^n \rightarrow \mathbb{R}^k$, we let $f(x_H, .) : {{\bf X}}^{[n] \setminus H} \rightarrow \mathbb{R}^k$ denote the function obtained by fixing the coordinates of $f$ in $H$ to $x_H$. As we have stated before, for the case $k=2$, Ghazi, Kamath and Sudan [@GKS16] reduce Theorem \[thm:junta\] for the general ${{\bf P}}$ case to the case when ${{\bf P}}= \mathbf{G}_{\rho,2}$. In other words, for $k=2$, [@GKS16] reduces Theorem \[thm:junta\] for the general ${{\bf P}}$ case to Theorem \[thm:junta-strong\] with $\ell=1$. We now give a sketch of why Theorem \[thm:junta\] reduces to Theorem \[thm:junta-strong\] for $k>1$. **Overview of the reduction:** Using the regularity lemma for low-degree polynomials [@DSTW:10; @DDS14] and other ideas from Boolean function analysis (along the lines of [@GKS16]), one can easily show the following: Let $\tau>0$ be any error parameter. Then, there exists a set $H \subseteq [n]$ such that $|H| = O_{\tau,|{{\bf P}}|,k}(1)$ and for $(x_H, y_H) \sim ({{\bf X}}, {{\bf Y}})^H$, with probability $1-\tau$, the following holds: The functions $f(x_H, \cdot)$ and $g(y_H, \cdot)$ are *low-influence* functions namely, $$\max_{i \in [n] \setminus H} \mathsf{Inf}_i(f(x_H, \cdot)) \leq \tau, \ \ \max_{i \in [n] \setminus H} \mathsf{Inf}_i(g(y_H, \cdot)) \leq \tau.$$ In the above definition, for $f: \mathbb{R}^n \rightarrow \mathbb{R}^k$, we let $\mathsf{Inf}_i(f)$ denotes the quantity $$\mathsf{Inf}_i(f) = \sum_{i \in S: S \in \mathbb{Z}^{\ast n}} \Vert \widehat{f}(S) \Vert_2^2,$$ where $\widehat{f}(S)$ denotes the Hermite coefficient of $f$ corresponding to $S$. Note that this is the standard definition of “influence" from Boolean function analysis (see [@ODonnell:book; @Mossel2010]). In fact, one can also additionally assume that every coordinate of $f$ and $g$ is essentially a low-degree polynomial. To understand why the low-influence condition is useful, let ${{\bf P}}_{G} = \mathbf{G}_{\rho,2}$ where $\rho = \rho({{\bf X}},{{\bf Y}})$. Further, let $({{\bf X}}_G, {{\bf Y}}_G) = {{\bf P}}_G$. Likewise, let $\tilde{f}(x_H, \cdot)$ (resp. $\tilde{g}(y_H, \cdot)$) be the multilinear extension of $f(x_H, \cdot)$ (resp. ${g}(y_H, \cdot)$) to the Gaussian space. Then, the invariance principle of Mossel *et al.* [@MOO10; @Mossel2010] shows that as long as $\tau$ is chosen to be sufficiently small in $\delta$, for any pair $(x_H, y_H)$ where $f(x_H, \cdot)$ and $g(y_H, \cdot)$ are low-influence functions, the following holds: $${d_{\mathrm TV}}( (\tilde{f}(x_H, {{\bf X}}_G^{[n] \setminus H}), \tilde{g}(y_H, {{\bf Y}}_G^{[n] \setminus H})) , (f(x_H, {{\bf X}}^{[n] \setminus H}), f(y_H, {{\bf Y}}^{[n] \setminus H})) \le \delta/4.$$ Note that the total number of $(x_H, y_H)$ pairs is bounded by $|\mathsf{supp}({{\bf P}})|^{2|H|}$. Let us denote this number by $\mathbf{N}_{sup}$. By applying Theorem \[thm:junta-strong\],we obtain that for any $\delta>0$, there is $n_0 = n_0(\mathbf{N}_{sup},k,\delta)$ such that corresponding to every function $\tilde{f}(x_H, \cdot)$ (resp. $\tilde{g}(y_H, \cdot)$ ), there is a function $\underline{f_{x_H}} : \mathbb{R}^{n_0} \rightarrow [k]$ (resp. $\underline{g_{y_H}}: \mathbb{R}^{n_0} \rightarrow [k]$ ) such that $${d_{\mathrm TV}}\big(\big(\tilde{f}(x_H, {{\bf X}}_G^{[n] \setminus H}), \tilde{g}(y_H, {{\bf Y}}_G^{[n] \setminus H}\big), \big(\underline{f}_{x_H}( {{\bf X}}_G^{n_0}), \underline{g}_{y_H}( {{\bf Y}}_G^{n_0})\big) \big) \le \delta/4.$$ Note that here we are crucially using the fact that Theorem \[thm:junta-strong\] is valid for an arbitrary $\ell \ge 1$ and not just $\ell=1$. Let us define $m_0 = n_0 \cdot (1/\kappa^2)$. We next define $\underline{f}_{\mathsf{low}, \ x_H} : \mathbb{R}^{m_0} \rightarrow [k]$ as $$\underline{f}_{\mathsf{low}, \ x_H} \big(x_{1,1}, \ldots, x_{n_0, \kappa^{-2}} \big) = \underline{f}_{x_H} \big( \kappa \cdot (x_{1,1} + \ldots + x_{1,\kappa^{-2}}) , \ldots, \kappa \cdot (x_{n_0,1} + \ldots + x_{n_0,\kappa^{-2}})\big).$$ $$\underline{g}_{\mathsf{low}, \ y_H} \big(y_{1,1}, \ldots, y_{n_0, \kappa^{-2}} \big) = \underline{g}_{y_H} \big( \kappa \cdot (y_{1,1} + \ldots + y_{1,\kappa^{-2}}) , \ldots, \kappa \cdot (y_{n_0,1} + \ldots + y_{n_0,\kappa^{-2}})\big).$$ From the definition of $\underline{f}_{x_H}$ and $\underline{g}_{y_H}$, it easily follows that, $$\big(\underline{f}_{x_H}( {{\bf X}}_G^{n_0}), \underline{g}_{y_H}( {{\bf Y}}_G^{n_0})\big) = \big(\underline{f}_{\mathsf{low}, x_H}( {{\bf X}}_G^{m_0}), \underline{g}_{\mathsf{low}, y_H}( {{\bf Y}}_G^{m_0})\big)$$ Let ${f}_{\mathsf{low}, x_H}$ and ${g}_{\mathsf{low}, y_H}$ denote the multilinear extensions of $\underline{f}_{\mathsf{low}, x_H}$ and $\underline{g}_{\mathsf{low}, y_H}$ to the space $({{\bf X}}^{m_0}, {{\bf Y}}^{m_0})$. Observe that the functions $\underline{f}_{\mathsf{low}, x_H}$ and $\underline{g}_{\mathsf{low}, y_H}$ have influence bounded by $\kappa$. Thus, as long as $\kappa$ is chosen to be a sufficiently small function of $\delta$, the invariance principle [@Mossel2010] implies that $${d_{\mathrm TV}}\big( ({f}_{\mathsf{low}, x_H} ( {{\bf X}}^{m_0}),{g}_{\mathsf{low}, y_H} ({{\bf Y}}^{m_0})), (\underline{f}_{\mathsf{low}, x_H}( {{\bf X}}_G^{m_0}), \underline{g}_{\mathsf{low}, y_H}( {{\bf Y}}_G^{m_0}))\big) \le \delta/4.$$ Combining the above three equations, we get that $${d_{\mathrm TV}}\big( ({f}_{\mathsf{low}, x_H} ( {{\bf X}}^{m_0}),{g}_{\mathsf{low}, y_H} ({{\bf Y}}^{m_0})), ({f}(x_H, {{\bf X}}^{[n] \setminus H}), {g}(y_H, {{\bf Y}}^{[n] \setminus H}))\big) \le \frac{3 \delta}{4}.$$ With this, we define functions $f_{\mathsf{junta}}: \mathbb{R}^{m_0 + |H|} \rightarrow [k]$ and $g_{\mathsf{junta}}: \mathbb{R}^{m_0 + |H|} \rightarrow [k]$ as follows. Split $x \in \mathbb{R}^{m_0 +H}$ as $(x_H, x_{m_0})$ and $y \in \mathbb{R}^{m_0 +H}$ as $(y_H, y_{m_0})$. $$f_{\mathsf{junta}}(x_H, x_{m_0}) = {f}_{\mathsf{low}, x_H} (x_{m_0}) ; \ g_{\mathsf{junta}}(y_H, y_{m_0}) = {g}_{\mathsf{low}, y_H} (y_{m_0}).$$ This immediately implies $${d_{\mathrm TV}}((f({{\bf X}}^n), g({{\bf Y}}^n)),(f_{\mathsf{junta}}({{\bf X}}^{m_0+|H|}), g_{\mathsf{junta}}({{\bf Y}}^{m_0 +|H|})) \le \frac{3\delta}{4} + \tau.$$ Once we choose $\tau \le \delta/4$, the reduction is complete. [^1]: Supported by a start-up grant from Northwestern University [^2]: Supported by ONR grant N00014-16-1-2227 and NSF grant CCF 1320105.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Protein-nucleic acid complexes are important for many cellular processes including the most essential function such as transcription and translation. For many protein-nucleic acid complexes, flexibility of both macromolecules has been shown to be critical for specificity and/or function. Flexibility-rigidity index (FRI) has been proposed as an accurate and efficient approach for protein flexibility analysis. In this work, we introduce FRI for the flexibility analysis of protein-nucleic acid complexes. We demonstrate that a multiscale strategy, which incorporates multiple kernels to capture various length scales in biomolecular collective motions, is able to significantly improve the state of art in the flexibility analysis of protein-nucleic acid complexes. We take the advantage of the high accuracy and ${\cal O}(N)$ computational complexity of our multiscale FRI method to investigate the flexibility of large ribosomal subunits, which is difficult to analyze by alternative approaches. An anisotropic FRI approach, which involves localized Hessian matrices, is utilized to study the translocation dynamics in an RNA polymerase.' author: - | Kristopher Opron$^1$, Kelin Xia$^2$, Zach Burton$^1$ and Guo-Wei Wei$^{3}$ [^1] [^2]\ $^1$ Department of Biochemistry and Molecular Biology\ Michigan State University, MI 48824, USA\ $^2$ Department of Mathematics\ Michigan State University, MI 48824, USA\ $^3$ Mathematical Biosciences Institute\ The Ohio State University, Columbus, Ohio 43210, USA title: 'Flexibility-Rigidity Index for Protein-Nucleic Acid Flexibility and Fluctuation Analysis ' --- Key words: Thermal fluctuation, Atomic flexibility, protein-nucleic acid complex, Multiscale. Introduction {#sec:Intro} ============ Proteins and the nucleic acids, which include deoxyribonucleic acid (DNA) and ribonucleic acid (RNA), are among the most essential biomolecules for all known forms of life. In cells, proteins have a wide variety of important functions, including supporting organism structure, catalyzing reactions involved in transcription and the cell cycle, participating in signal transduction, and working as immune agents. Nucleic acids typically function in association with proteins and play a crucial role in encoding, transmitting and expressing genetic information. Genetic information is stored through the nucleic acid sequence, i.e., the order of nucleotides within a DNA or RNA molecule and transmitted via transcription and translation processes. [Protein rigidity, flexibility and electrostatics strongly correlate to protein structure and function[@Anfinsen:1973]. ]{} The impact of [biomolecular]{} electrostatics to their structure, function and dynamics has been a subject of intensive study. However the importance of biomolecular flexibility and rigidity to their structure and function has been overlooked. In general, protein rigidity is responsible for protein three-dimensional (3D) equilibrium geometric shapes and structural function in forms of tubulin, collagen, elastin, and keratin, while protein flexibility is an important factor in all other protein functions [@Frauenfelder:1991]. DNA flexibility is an important effect in DNA packing. [Although the flexibility of biomolecules is often associated with their motion and dynamics]{}, which are their response to the external stimuli and die out at the absolute zero temperature, [ flexibility is an intrinsic property.]{} Biomolecular flexibility and rigidity can be measured directly or indirectly by many experimental approaches, such as X-ray crystallography, nuclear magnetic resonance (NMR) and single-molecule force experiments [@Dudko:2006]. In single-molecule force experiments, including optical tweezers and nanopore force spectroscopy, the intrinsic rupture rate can be a direct measure of the flexibility and rigidity. In the X-ray structure, Debye-Waller factors, also known as B-factors or temperature factors, are computed as the uncertainty for each atom in the least square fitting of between the X-ray diffraction data and the theoretical model. Debye-Waller factors are interpreted as atomic mean-square-fluctuations at the given experimental temperature, and are associated with biomolecular flexibility and rigidity. NMR is known for its ability to analyze biomolecular flexibility and rigidity under physiological conditions, and at various timescales. The availability of experimental data makes the theoretical study of biomolecular flexibility and rigidity an interesting and important topic, in which quantitative models can be calibrated and validated. Molecular dynamics (MD) [@McCammon:1977] can be used to elucidate biomolecular collective motion and fluctuation. MD is a powerful technique for the understanding of the conformational landscapes of biomolecules. However, biomolecular flexibility and rigidity are intrinsic properties that are better measured at the motionless and fluctuation free state. Therefore, MD is not efficient for biomolecular flexibility and rigidity analysis. Alternative approaches including normal mode analysis (NMA) [@Go:1983; @Tasumi:1982; @Brooks:1983; @Levitt:1985], graph theory [@Jacobs:2001] and elastic network model (ENM) [@Bahar:1997; @Bahar:1998; @Atilgan:2001; @Hinsen:1998; @Tama:2001; @LiGH:2002] become the main workhorses for biomolecular flexibility and rigidity analysis during the past two decades. In analogy to the time-dependent and time-independent Schrödinger’s equations, these approaches are designed as time-independent counterparts of the corresponding MD methods [@JKPark:2013]. Consequently, a diagonalization of the interaction matrix or Hamiltonian of a biomolecule is a required procedure to obtain biomolecular eigenmodes and associated eigenvalues, which are further organized to predict the biomolecular temperature factors. The low order eigenmodes computed from diagonalizing the Kirchhoff matrix or the Hessian matrix can be interpreted as the slow motions of the biomolecule around the equilibrium state and thus shed light on the long-time behavior of the biomolecular dynamics beyond the reach of MD simulations [@Tasumi:1982; @Brooks:1983; @Levitt:1985]. Tirion argued that the potential in the NMA can be simplified to retain only the harmonic potential for elasticity, which is the dominant term in the MD Hamiltonian [@Tirion:1996]. Network theory [@Flory:1976] has had considerable impact in flexibility analysis. The combination of elasticity and coarse-grained network gives rise to elastic network model (ENM) [@Hinsen:1998]. Many other network based approaches, including Gaussian network model (GNM) [@Bahar:1997; @Bahar:1998] and anisotropic network model (ANM) [@Atilgan:2001], have been developed for biomolecular flexibility analysis. It has been demonstrated by Yang et al. [@LWYang:2008] that GNM is about one order more efficient than most other flexibility approaches. GNM is also typically more accurate than ANM in B-factor prediction [@JKPark:2013; @Opron:2014]. Applications have been demonstrated in stability [@Livesay:2004] analysis, docking simulation [@Gerek:2010], viral capsids [@Rader:2005; @Tama:2005] and domain motions of hemoglobin [@CXu:2003], F1 ATPase [@WZheng:2003; @QCui:2004], chaperonin GroEL [@Keskin:2002; @WZheng:2007] and the ribosome [@Tama:2003; @YWang:2004]. More details can be found in a few recent reviews [@JMa:2005; @LWYang:2008; @Skjaven:2009; @QCui:2010]. A common feature of the above mentioned time-independent methods is that they resort to the matrix diagonalization procedure. The computational complexity of the matrix diagonalization is typically of the order of ${\cal O}(N^3)$, where $N$ is the number of elements in the matrix. Such a computational complexity calls for new efficient strategies for the flexibility analysis of large biomolecules. It is well known that NMA and GNM offer poor flexibility analysis for many macromolecules [@Kundu:2002; @Kondrashov:2007; @Hinsen:2008; @GSong:2007]. Park et al. had studied the performance of NMA and GNM methods for three sets of structures [@JKPark:2013]. They found that both methods fail to work and deliver negative correlation coefficients (CCs) for many structures [@JKPark:2013]. They have shown that mean correlation coefficients (MCCs) for the NMA B-factor prediction of small-sized, medium-sized and large-sized sets of structures are about 0.480, 0.482 and 0.494, respectively [@JKPark:2013; @Opron:2014]. The GNM is considerably more accurate and delivers MCCs of 0.541, 0.550 and 0.529 for the above test sets [@JKPark:2013; @Opron:2014]. Indeed, various improvements, including crystal environment, solvent type and co-factors, are proposed [@Kundu:2002; @Kondrashov:2007; @Hinsen:2008; @GSong:2007]. Additionally, density - cluster rotational - translational blocking has been considered [@Tama:2000; @Demerdash:2012]. Alternative approaches have been proposed for the flexibility analysis of hinges in proteins using bioinformatics [@hingeatlas], graph theory [@hingeprot; @stonehinge; @flexprot] and energetics [@flexoracle]. Moreover, low quality experimental data due to collection conditions and structural refinement procedures may also contribute to poor flexibility predictions. From observation of the relationship between flexibility and local packing density, Halle [@Halle:2002] proposed a much simplified model called local density model (LDM), and bypassed the whole eigenmode analysis for protein B-factor prediction. In this method, the inverse of contact density, defined as the number of noncovalent neighbor atoms within a local region, is found to be proportional to atomic mean-square displacements, thus can be directly used to predict the experimental B-factors. Another interesting method is the local contact model (LCM) proposed by Zhang et al [@FLZhang:2002]. In this approach, the generalized order parameter of the atom is approximated by the summation of a series of exponential functions of atomic distances. Both LDM and LCM demonstrate great potential for protein flexibility prediction. Based on these approaches, many modifications have been proposed in the literature [@CPLin:2008; @SWHuang:2008; @DWLi:2009]. Among them, the weighted contact number (WCN) is able to deliver a better accuracy than GNM using an inverse square distance function [@CPLin:2008]. Recently, we have proposed a few matrix-decomposition-free methods for flexibility analysis, including molecular nonlinear dynamics [@KLXia:2014b], stochastic dynamics [@KLXia:2013f] and flexibility-rigidity index (FRI) [@KLXia:2013d; @Opron:2014]. Among them, flexibility-rigidity index (FRI) has been introduced to evaluate protein flexibility and rigidity, that are further required in a multiscale formalism called continuum elasticity with atomic rigidity (CEWAR) for macromolecular the elasticity analysis [@KLXia:2013d]. The FRI method appears to be akin to the “flexibility index” proposed independently by von der Lieth et al. [@vonderLieth:1996] and Jacobs et al. [@Jacobs:2001] to describe bond strengths. However, these flexibility indices have little in common with FRI, which does not resort to any protein interaction Hamiltonian for predicting protein flexibility and rigidity. Instead, the FRI is a structure based approach. The fundamental assumptions of the FRI method are as follows. Protein functions, such as flexibility, rigidity, and energy, are fully determined by the structure of the protein and its environment, and the protein structure is in turn determined by the relevant interactions. Therefore, whenever the protein structure is available, there is no need to analyze protein flexibility and rigidity by tracing back to the protein interaction Hamiltonian. Consequently, the FRI bypasses the ${\cal O}(N^3)$ matrix diagonalization. In fact, FRI does not even require the 3D geometric information of the protein structure. It assesses topological connectivity of the protein distance geometry and analyzes the geometric compactness of the protein structure. It can be regarded as a kernel generalization of the local density model [@Halle:2002]. Our initial FRI [@KLXia:2013d] has the computational complexity of of ${\cal O}(N^2)$ and our fast FRI (fFRI) [@Opron:2014] based on a cell lists algorithm [@Allen:1987] is of ${\cal O}(N)$. The FRI and the fFRI have been extensively validated by a set of 365 proteins for parametrization, accuracy and reliability. The parameter free fFRI is about ten percent more accurate than the GNM on the 365 protein test set and is orders of magnitude faster than GNM on a set of 44 proteins. FRI is able to predict the B-factors of an HIV virus capsid (313 236 residues) in less than 30 seconds on a single-core processor, which would require GNM more than 120 years to accomplish if the computer memory is not a problem [@Opron:2014]. However, earlier FRI methods do not work for many protein structures that NMA and GNM are unable to deliver good predictions. In addition to problems mentioned above, the neglecting of multiple characteristic length scales in protein structures is another drawback of [present]{} flexibility analysis. Indeed, biomolecules have many characteristic length scales, ranging from covalent bond scale, hydrogen bond scale, wan der Waals bond scale, intraresidue scale, interresidue scale, alpha helix and beta sheet scale, domain scale and protein interaction scale. When GNM or FRI is parametrized at a given cutoff or scale parameter, it captures only a subset of the characteristic length scales but inevitably misses other characteristic length scales of the protein. Consequently, these methods fail to offer accurate B-factor prediction for many multiscale biomolecules. A multiscale strategy has been proposed to resolve this problem by introducing two or three kernels that are parametrized at relatively small, medium and/or large length scales in the FRI formulation [@Opron:2015a]. We demonstrate that the resulting multiscale FRI (mFRI) works extremely well for many proteins that the GNM method fails to offer accurate flexibility analysis [@Opron:2015a]. Based on a set of 364 proteins, mFRI is 20% more accurate than GNM. It is interesting to note that there is no obvious way to incorporate multiple length scales in the aforementioned matrix diagonalization based approaches. Although being developed independently, our FRI methods are akin to LDM, LCM and WCN in terms of matrix diagonalization free. However, our FRI methods differ from LDM, LCM and WCN in the following aspects. First, our original FRI methods were motivated from continuum mechanics, the CEWAR [@KLXia:2013d; @Opron:2014]. As a result, our FRI methods offer not only discrete flexibility index and discrete rigidity index, but also continuous flexibility function and continuous rigidity function. The latter is equivalent to volumetric biomolecular density distribution. Consequently, one can use the FRI rigidity function to fit electron microscope (EM) density maps [@MTopf:2008; @Wriggers:1999; @KLXia:2015b]. A byproduct of our FRI rigidity function is the smooth biomolecular surface extracted by setting an isosurface value, including the Gaussian surface as a special case [@Krone:2012; @KLXia:2015d]. In contrast, LDM, LCD and WCN do not admit any continuum representation. Additionally, the discrete FRI formulations differ from those of LDM, LCD and WCN by admitting the diagonal term in the summation. Moreover, we have considered the multiscale effects in biomolecules. Our mFRI captures biomolcular thermal fluctuations at various length scales and thus substantially improves the accuracy of the original FRI method. Finally, we have proposed an anisotropic FRI (aFRI) method to describe biomolecular collective motions. A unique feature of our aFRI method is that it allows adaptive Hessian matrices, from a completely global $3N\times 3N$ matrix to completely local $3 \times 3 $ matrices. Therefore, one can use aFRI to pinpoint one’s flexibility analysis to a given domain or region. The objective of the present work is to develop FRI methods for the flexibility analysis of protein-nucleic acid complexes. Protein and nucleic acid are dramatically different biomolecules. Amino acid residues and [nucleotides]{} have different length scales and interaction characteristics. Therefore, a good model should not only allow residues and/or [nucleotides]{} to be treated with different length scales, but also adapt a multiscale description of each residue and/or [nucleotide]{}. Unlike elastic network models that are parametrized in only one length scale for each particle, the mFRI provides a simultaneous multiscale description. Therefore, the present mFRI is able to better capture multiscale collective motions of protein-nucleic acid complexes. Additionally, many protein-nucleic acid complexes are very large biomolecules and pose difficulty to conventional mode decomposition based methods. The ${\cal O}(N)$ scaling FRI methods provide an efficient approach to the flexibility analysis of large protein-nucleic acid complexes. The rest of this paper is organized as follows. Section \[sec:methods\] is devoted to methods and algorithms. To establish notation and facilitate further discussion, the basic FRI approach is briefly discussed. We then present the multikernel based mFRI method to improve the accuracy of biomolecular flexibility analysis. The basic formulation of the aFRI is discussed. In Section \[sec:validation\], we first analyze the benefit of adding an additional kernel with an appropriate length scale by comparing the performance of the B-factor prediction for a set of 64 protein-nucleic acid complexes introduced by Yang et al. [@Yang:2006] and a larger database of 203 high resolution protein-nucleic acid structures. Three different coarse-grain representations of protein-nucleic acid complexes introduced by Yang et al [@Yang:2006] are examined. Section \[Sec:CaseStudy\] is devoted to the application of the mFRI and aFRI methods. We consider a large ribosomeal structure to explore the utility and demonstrate the performance of the proposed mFRI. Further, we explore the use of aFRI for the prediction of collective motions of bridge helix, trigger loop and nucleic acids in an RNA polymerase. Methods and algorithms {#sec:methods} ====================== Flexibility-rigidity index {#sec:Flexibility} -------------------------- In FRI, the topological connectivity of a biomolecule is measured by rigidity index and flexibility index. In particular, the rigidity index represents the protein density profile. Consider an $N$-atom representation of a biomolecule. The coordinates of these atoms are given as $\{ {\bf r}_{j}| {\bf r}_{j}\in \mathbb{R}^{3}, j=1,2,\cdots, N\}$. We denote $ \|{\bf r}_i-{\bf r}_j\|$ the Euclidean space distance between the $i$th atom and the $j$th atom. A general correlation kernel, $ \Phi( \|{\bf r} - {\bf r}_j \|;\eta_{j})$, is a real-valued monotonically decreasing radial basis function satisfying $$\begin{aligned} \label{eq:couple_matrix1-1} \Phi( \|{\bf r} - {\bf r}_j \|;\eta_{j})&=&1 \quad {\rm as }\quad \|{\bf r} - {\bf r}_j \| \rightarrow 0\\\ \Phi( \|{\bf r} - {\bf r}_j \|;\eta_{j})&=&0 \quad {\rm as }\quad \|{\bf r} - {\bf r}_j \| \rightarrow\infty,\end{aligned}$$ where $\eta_{j}$ is an atomic type dependent scale parameter. The correlation between the $i$th and $j$th particles is given by $$\begin{aligned} \label{eq:couple_matrix0} {C}_{ij} = \Phi( \|{\bf r}_i - {\bf r}_j \|;\eta_{j}).\end{aligned}$$ The correlation matrix $\{ C_{ij}\}$ can be computed to visualize the connectivity among protein particles. We define a position (${\bf r}$) dependent (continuous) rigidity function or density function [@KLXia:2013d; @Opron:2014] $$\begin{aligned} \label{eq:rigidity3} \mu({\bf r}) & = & \sum_{j=1}^N w_{j} \Phi( \|{\bf r} - {\bf r}_j \|;\eta_{j} ), \end{aligned}$$ where $w_{j}$ is an atom type dependent weight. For example, carbon, nitrogen and phosphorus atoms can have different weights. Equation (\[eq:rigidity3\]) can be understood as a discrete to continuum mapping. It maps a set of discrete values $\{w_j\}$ at $\{ {\bf r}_j\}$ to the continuum domain. Although Delta sequences of the positive type discussed in an earlier work [@GWei:2000] are all good choices, generalized exponential functions $$\begin{aligned} \label{eq:couple_matrix1} \Phi(\|{\bf r} - {\bf r}_j \|;\eta_{j}) = e^{-\left(\|{\bf r} - {\bf r}_j \|/\eta_{j}\right)^\kappa}, \quad \kappa >0\end{aligned}$$ and generalized Lorentz functions $$\begin{aligned} \label{eq:couple_matrix2} \Phi(\|{\bf r} - {\bf r}_j \|;\eta_{j}) = \frac{1}{1+ \left( \|{\bf r} - {\bf r}_j \|/\eta_{j}\right)^{\upsilon}}, \quad \upsilon >0 \end{aligned}$$ have been commonly used in our recent work [@KLXia:2013d; @Opron:2014; @Opron:2015a]. Since the rigidity function can be directly interpreted as a density distribution, it can been used to define the rigidity surface of a biomolecule by taking an isovalue. By taking $\kappa =2$ in Eq. (\[eq:couple\_matrix1\]), we result in a formula for a Gaussian surface from Eq. (\[eq:rigidity3\]). Similarly, we define a position (${\bf r}$) dependent (continuous) flexibility function [@KLXia:2013d; @Opron:2014] $$\begin{aligned} \label{eq:flexibility1} F({\bf r}) & = & \frac{1}{\sum_{j=1}^N w_{j} \Phi( \|{\bf r} - {\bf r}_j \|;\eta_{j} )}. \end{aligned}$$ This function is well defined in the computational domain containing the biomolecule. The flexibility function can be visualized by its projection on a given surface, such as the solvent excluded surface of a biomolecule. The (discrete) rigidity index for the $i$th particle is obtained by restricting ${\bf r}$ to a given atomic position ${\bf r}_i$ $$\begin{aligned} \label{eq:rigidity4} \mu_i & = & \sum_{j=1}^N w_{j} \Phi( \|{\bf r}_i - {\bf r}_j \|;\eta_{j} ). \end{aligned}$$ Here $ \mu_i$ measures the total density or rigidity at the $i$th particle. In a similar manner, we define a set of (discrete) flexibility indices by $$\begin{aligned} \label{eq:flexibility2} f_i & = & \frac{1}{\sum_{j=1}^N w_{j} \Phi( \|{\bf r}_i - {\bf r}_j \|;\eta_{j} )}. \end{aligned}$$ The flexibility index $ f_i$ is directly associated with the B-factor of $i$th particle $$\begin{aligned} \label{eq:regression} B_i^t = a f_i + b, \quad \forall i =1,2,\cdots,N\end{aligned}$$ where $ \{B_i^t\}$ are theoretically predicted B-factors, and $a$ and $b$ are two constants to be determined by a simple linear regression. This allows us to use experimental data to validate the FRI method. In our earlier work [@KLXia:2013d; @Opron:2014; @Opron:2015a], we set $w_j=1$ for the coarse-grained C$_\alpha$ representation of proteins. We have also developed parameter free FRI (pfFRI), such as $(\kappa=1, \eta=3)$ and $(\upsilon=3, \eta=3)$, to make our FRI robust for protein C$_\alpha$ B-factor prediction. Multiscale Flexibility-rigidity index {#sec:MFRI} ------------------------------------- The basic idea of multiscale FRI or multi-kernel FRI (mFRI) is quite simple. Since macromolecules are inherently multiscale in nature, we utilize multiple kernels that are parametrized at multiple length scales to characterize the multiscale thermal fluctuations of macromolecules $$\begin{aligned} \label{eq:flexibility3} f^{n}_i & = & \frac{1}{\sum_{j=1}^N w^{n}_{j} \Phi^{n}( \|{\bf r}_i - {\bf r}_j \|;\eta^{n}_{j} )}, \end{aligned}$$ where $w^{n}_{j}$, $\Phi^{n}( \|{\bf r}_i - {\bf r}_j \|;\eta^{n}_{j}) $ and $\eta^{n}_{j}$ are the corresponding quantities associated with the $n$th kernel. We seek the minimization of the form $$\begin{aligned} \label{eq:regression2} {\rm Min}_{a^{n},b} \left\{ \sum_i \left| \sum_{n}a^n f^{n}_i + b-B^e_i\right|^2\right\}\end{aligned}$$ where $\{B^e_i\}$ are the experimental B-factors. In principle, all parameters can be optimized. For simplicity and computational efficiency, we only determine $\{a^n\}$ and $b$ in the above minimization process. For each kernel $\Phi^n$, $w^n_j$ and $\eta^n_j$ will be selected according to the type of particles. Specifically, for a simple C$_\alpha$ network, we can set $w^n_j=1$, $\eta_j^{n}=\eta^{n}$ and choose a single kernel function parametrized at different scales. The predicted B-factors can be expressed as $$\begin{aligned} \label{eq:flexibility4} B^{\rm mFRI}_i = b+ \sum_{n=1}\frac{a^n}{\sum_{j=1}^N \Phi( \|{\bf r}_i - {\bf r}_j \|;\eta^{n} )}. \end{aligned}$$ The difference between Eqs. (\[eq:flexibility3\]) and (\[eq:flexibility4\]) is that, in Eqs. (\[eq:flexibility3\]), both the kernel and the scale can be changed for different $n$. In contrast, in Eq. (\[eq:flexibility4\]), only the scale is changed. One can use a given kernel, such as $$\begin{aligned} \label{eq:couple_matrixn} \Phi(\|{\bf r} - {\bf r}_j \|;\eta^n) = \frac{1}{1+ \left( \|{\bf r} - {\bf r}_j \|/\eta^n\right)^{3}}, \end{aligned}$$ to achieve good multiscale predictions [@Opron:2015a]. Anisotropic flexibility-rigidity index {#sec:AnisoFlexibility} -------------------------------------- The anisotropic flexibility-rigidity index (aFRI) model was built in a very unique manner. Different from the existing normal mode analysis or anisotropic elastic network models, in which the Hessian matrix is always global, our aFRI model delivers a local and adaptive Hessian matrix. This means that for a molecule with $N$ particles, the Hessian matrix is always $3N\times 3N$ for ANM, whereas, our Hessian matrix may vary from a set of $N$ $3\times 3$ matrices for a completely local aFRI to $3N\times 3N$ for a complete global aFRI, depending on the need of a physical problem. For instance, if one is particularly interested in certain structures like alpha helices, domains, or binding sites of a protein, or certain subunits of a biomolecular complex, one design and an aFRI for these portions of the molecule. We partition all the $N$ particles in a molecule or a biomolecular complex into a total of $M$ clusters $\{c_1, c_2,\cdots, c_k, \cdots, c_M \}$. Cluster $c_k$ has $N_k$ particles or atoms so that $N=\sum_{k=1}^M N_{k}$. We choose clusters based on our physical interest as mentioned above. In this way, two very special situations can be found. The first one corresponds to the completely local situation, i.e., $N$ clusters and each cluster has only one atom. The other situation contains only one cluster, which is then completely global. It is straightforward to construct a $3N\times 3N$ Hessian matrix and analyze the collective motion. The problem arises when we consider the global motion of a selected cluster, at the same time include the influence from the rest clusters. The essential idea is to construct a cluster Hessian matrix for each cluster individually and then incorporate the information from nearby clusters into its diagonal terms. For example, if we want to know the thermal fluctuation of a particular cluster $c_k$ with $N_{k}$ particles or atoms, we need to find $3N_k$ eigenvectors for the cluster. Let us keep in mind that each position vector in $\mathbb{R}^3$ has three components, i.e., ${\bf r}=(x,y,z)$. For each given pair of particles $i$ and $j$, we can define a local anisotropic matrix $\Phi^{ij}=\left( \Phi^{ij}_{uv} \right)$ as $$\Phi^{ij}=\left( \begin{array}{ccc} \Phi^{ij}_{xx} & \Phi^{ij}_{xy}& \Phi^{ij}_{xz}\\ \Phi^{ij}_{yx} & \Phi^{ij}_{yy}& \Phi^{ij}_{yz}\\ \Phi^{ij}_{zx} & \Phi^{ij}_{zy}& \Phi^{ij}_{zz} \end{array} \right),$$ where $\Phi^{ij}_{uv}$ are defined as $$\begin{aligned} \label{eq:Anisorigidity1} \Phi^{ij}_{uv} = \frac{\partial}{\partial u_i} \frac{\partial}{\partial v_j} \Phi( \|{\bf r}_i - {\bf r}_j \|; \eta_{j} ), \quad u,v= x, y, z; i,j =1,2,\cdots,N.\end{aligned}$$ Due to the inner connection between rigidity and flexibility, we have two different aFRI algorithms. The specially designed cluster Hessian matrix with a smaller size can incorporate nonlocal geometric impact and predict collective thermal motions of the cluster. The details are presented below. ### Anisotropic rigidity {#sec:arFRI} In anisotropic rigidity based aFRI, a rigidity Hessian matrix is needed. For a cluster $c_k$, if we denote its rigidity Hessian matrix as $ \left(\mu_{uv}^{ij}(c_k)\right)$ with elements, $$\begin{aligned} \label{eq:Anisorigidity2} \mu^{ij}_{uv}(c_k) =& - w_{j}\Phi^{ij}_{uv}, &\quad i,j \in c_k; i\neq j; u,v= x, y, z \\ \label{eq:Anisorigidity3} \mu^{ii}_{uv}(c_k)=& \sum_{j=1}^N w_{j} \Phi^{ij}_{uv}, &\quad i \in c_k; u,v= x, y, z \\ \label{eq:Anisorigidity4} \mu^{ij}_{uv}(c_k)=& 0, &\quad i,j \notin c_k; u,v= x, y, z.\end{aligned}$$ In this way, the rigidity Hessian matrix is of $3N_k\times 3N_k$ dimensions. More importantly, the information from all other clusters are built into diagonal terms, even if the cluster itself is completely localized, i.e., $N_k=1,~ \forall k$. For B-factor prediction, we define a set of anisotropic rigidity (AR) based flexibility indices by $$\begin{aligned} \label{eq:Anisorigidity34} f_{i}^{\rm AR} = \frac{1}{\mu^{i}_{\rm diag}},\end{aligned}$$ where the $i$th diagonal term $\mu^{i}_{\rm diag}$ is of the form, $$\begin{aligned} \label{eq:Anisorigidity33} \mu^{i}_{\rm diag} &=& {\rm Tr}\left(\mu_{uv}^{i}\right)\\ &=& \sum_{j=1}^N w_{j} \left[\Phi^{ij}_{xx}+\Phi^{ij}_{yy}+ \Phi^{ij}_{zz}\right].\end{aligned}$$ Here, $f_{i}^{\rm AR}$ is employed in the linear regression to determine B-factors. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ![ Illustration highlighting atoms used for coarse-grained representations in protein-nucleic acid complexes for FRI and GNM. In addition to protein C$\alpha$ atoms, Model M1 considers the backbone P atoms for nucleotides. Model M2 includes M1 atoms and adds the sugar O4’ atoms for nucleotides. Model M3 includes M1 atoms and adds the sugar C4’ atoms and the base C2 atoms for nucleotides. []{data-label="CoarseG"}](ATP.pdf "fig:"){width="40.00000%"} ![ Illustration highlighting atoms used for coarse-grained representations in protein-nucleic acid complexes for FRI and GNM. In addition to protein C$\alpha$ atoms, Model M1 considers the backbone P atoms for nucleotides. Model M2 includes M1 atoms and adds the sugar O4’ atoms for nucleotides. Model M3 includes M1 atoms and adds the sugar C4’ atoms and the base C2 atoms for nucleotides. []{data-label="CoarseG"}](TTP.pdf "fig:"){width="40.00000%"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ### Anisotropic flexibility The other way to construct aFRI is to construct a flexibility Hessian matrix, which is denoted as ${\bf F}(c_k)$ for cluster $c_k$ with elements, $$\begin{aligned} \label{eq:Anisoflexibility} {\bf F}^{ij}(c_k) =& - \frac{1}{w_{j}} (\Phi^{ij})^{-1}, &\quad i,j \in c_k; i\neq j; u,v= x, y, z \\ \label{eq:Anisoflexibilityy3} {\bf F}^{ii}(c_k)=& \sum_{j=1}^N \frac{1}{w_{j}} (\Phi^{ij})^{-1}, &\quad i \in c_k; u,v= x, y, z \\ \label{eq:Anisoflexibility4} {\bf F}^{ij}(c_k)=& 0, &\quad i,j \notin c_k; u,v= x, y, z.\end{aligned}$$ Note that $(\Phi^{ij})^{-1}$ represents the unscaled inverse of matrix $\Phi^{ij}$ such that $\Phi^{ij}(\Phi^{ij})^{-1}=| \Phi^{ij}|$. The diagonalization of ${\bf F}(c_k)$ gives rise to eigenmodes, which represent the cluster motions. Additionally, the diagonal part ${\bf F}^{ii}(c_k)$ has built in information from all particles in the system. In this way, we deliver a cluster Hessian matrix. By diagonalizing ${\bf F}(c_k)$, we obtain $3N_k$ eigenvectors for the $N_k$ particles in the cluster $c_k$ of interest. Furthermore, instead of predicting the B-factors via the eigenvalues and eigenmodes, we directly predict the B-factors by using our anisotropic flexibility (AF) based flexibility indices defined as, $$\begin{aligned} \label{eq:Anisoflexibility2} f_i^{\rm AF} &=&{\rm Tr} \left({\bf F}(c_k)\right)^{ii}, \\ &=& \left({\bf F}(c_k)\right)^{ii}_{xx}+ \left({\bf F}(c_k)\right)^{ii}_{yy}+ \left({\bf F}(c_k)\right)^{ii}_{zz}.\end{aligned}$$ Finally, we employ $f_i^{\rm AF}$ to predict B-factors. Implementation and validation {#sec:validation} ============================== In this section, we parametrize and test the previously described mFRI on protein-nucleic acid structures. A immediate concern is whether the proposed mFRI is as efficient on protein-nucleic structures as it is on protein-only structures as shown in a previous study [@KLXia:2015a]. The accuracy of the mFRI method is tested by the B-factor prediction of two sets of protein-nucleic acid structures, including a set of 64 molecules used in a recent GNM study [@Yang:2006] and a set of 203 molecules for more accurate parametrization of mFRI. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![ MCCs for single kernel parameter test using the M1 (squares), M2 (circles) and M3 (triangles) representations. Lorentz kernel with $\upsilon=3$ is used. The parameter $\eta$ is varied to find the maximum MCC on the test set of structures. The results for a set of 64 protein-nucleic structures ( PDB IDs listed in Table \[table:64set\]) are shown on the left, while results for a separate set of 203 structures (PDB IDs listed in Table \[table:203set\]) is shown on the right for more general selections.[]{data-label="single_etasigma"}](test64.pdf "fig:"){width="45.00000%"} ![ MCCs for single kernel parameter test using the M1 (squares), M2 (circles) and M3 (triangles) representations. Lorentz kernel with $\upsilon=3$ is used. The parameter $\eta$ is varied to find the maximum MCC on the test set of structures. The results for a set of 64 protein-nucleic structures ( PDB IDs listed in Table \[table:64set\]) are shown on the left, while results for a separate set of 203 structures (PDB IDs listed in Table \[table:203set\]) is shown on the right for more general selections.[]{data-label="single_etasigma"}](test203.pdf "fig:"){width="45.00000%"} -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Coarse-grained representations of protein-nucleic acid complexes {#sec:nucleic} ---------------------------------------------------------------- In this section, we consider flexibility analysis of protein-nucleic acid complexes. To this end, we need coarse-grained representations. We consider three coarse-grained representation of nucleic acids to be used in conjugation with the C$\alpha$-only representation used for proteins. These three models are identical to those used by Yang et al. [@Yang:2006] and are named M1, M2 and M3. Model M1 consists of the backbone P atoms and protein C$\alpha$atoms. Model M2 contains the same atoms as M1 but also includes sugar O4’ atoms. Model M3 includes atoms from M1 and adds the sugar C4’ atoms and base C2 atoms, see Fig. \[CoarseG\]. Model M1 is similar to protein C$\alpha$ representations because they are both backbone-only representations. The atoms in M1 are 6 bonds apart while C$\alpha$ atoms are 3 bonds apart. Model M2 includes P atoms and adds the O4’ atoms located on the ribose portion of the nucleotide. Finally, model M3 includes atoms of P, C4’ and C2, a carbon from the base portion of the nucleotide, see Fig. \[CoarseG\]. As point out by Yang et al. [@Yang:2006], nucleotides are approximately three times more massive than amino acids and so model M3 with three nodes per nucleotide is consistent in this sense with using C$\alpha$ atoms for the protein representation. Multiscale/Multikernel FRI {#sec:multiFRI} --------------------------- To parametrize and test the accuracy of multikernel fFRI on protein-nucleic acid structures, we use a dataset from Yang et al. [@Yang:2006] containing 64 structures. In addition, we construct a larger database of 203 high resolution structures. This expanded protein-nucleic structure set was obtained by searching the Protein Data Bank (PDB) for structures that contain both Protein and DNA and structure which have an X-ray resolution between 0.0 and 1.75 Å. All PDB files are processed by removing low occupancy atomic coordinates for structures having residues with multiple possible coordinates. The PDB IDs of the 64 and 203 structures can be found in Table \[table:64set\] and Table \[table:203set\], respectively. To quantitatively assess the performance of the proposed multikernel FRI method, we consider the correlation coefficient (CC) $$\begin{aligned} \label{correlation} {\rm CC}=\frac{\sum^N_{i=1}\left(B^e_i-\bar{B}^e \right)\left( B^t_i-\bar{B}^t \right)} { \left[\sum^N_{i=1}(B^e_i- \bar{B}^e)^2\sum^N_{i=1}(B^t_i-\bar{B}^t)^2\right]^{1/2}},\end{aligned}$$ where $\{B^t_i, i=1,2,\cdots,N\}$ are a set of predicted B-factors by using the proposed method and $\{B^e_i, i=1,2,\cdots, N\}$ are a set of experimental B-factors read from the PDB file. Here $\bar{B}^t$ and $\bar{B}^e$ the statistical averages of theoretical and experimental B-factors, respectively. ### Multikernel FRI testing on protein-nucleic structures {#sec:multisearch} Previous tests of single kernel FRI indicate that the Lorentz type and exponential type correlation kernels are the two most accurate kernel types. This leads us to try the combination of these two types of kernels. The resulting multikernel FRI method requires four parameters, namely, $\kappa$ and $\eta$ for the exponential kernel and $\upsilon$ and $\eta$ for the Lorentz kernel. ### Single kernel FRI testing {#sec:singleFRI} In order to compare FRI and GNM methods for protein-nucleic acid structures, we test our single kernel FRI at a range of $\eta$ values. For this test we use the Lorentz kernel with $\upsilon=3$ for B-factor prediction on both structures sets and all three representations (M1, M2 and M3). The results are shown in Figure \[single\_etasigma\]. For the 64 structure set, single kernel FRI has a maximum mean correlation coefficient (MCC) to experimental B-factors for M1, M2 and M3 representations of 0.620, 0.612 and 0.555. Comparatively, GNM had a MCC of approximately 0.59, 0.58 and 0.55 for M1, M2 and M3 for the same data set [@Yang:2006]. The maximum MCCs for FRI on the larger data set for M1, M2 and M3 are 0.613, 0.625 and 0.586, respectively. The M1 and M2 representations perform better than the M3 representation. ### Parameter-free multikernel FRI {#sec:pfFRI} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- ![ Mean correlation coefficients (MCCs) for two-kernel FRI models on a set of 203 protein-nucleic structures. From left to right, MCC values are shown for M1, M2 and M3 representations. We use one Lorentz kernel with $\upsilon=3.0$ and one exponential kernel with $\kappa=1.0$. The values of parameter $\eta$ for both kernels are varied from 2 to 20 Å. []{data-label="multi_etasigma"}](M1_matrix.pdf "fig:"){width="33.00000%"} ![ Mean correlation coefficients (MCCs) for two-kernel FRI models on a set of 203 protein-nucleic structures. From left to right, MCC values are shown for M1, M2 and M3 representations. We use one Lorentz kernel with $\upsilon=3.0$ and one exponential kernel with $\kappa=1.0$. The values of parameter $\eta$ for both kernels are varied from 2 to 20 Å. []{data-label="multi_etasigma"}](M2_matrix.pdf "fig:"){width="33.00000%"} ![ Mean correlation coefficients (MCCs) for two-kernel FRI models on a set of 203 protein-nucleic structures. From left to right, MCC values are shown for M1, M2 and M3 representations. We use one Lorentz kernel with $\upsilon=3.0$ and one exponential kernel with $\kappa=1.0$. The values of parameter $\eta$ for both kernels are varied from 2 to 20 Å. []{data-label="multi_etasigma"}](M3_matrix.pdf "fig:"){width="33.00000%"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- As with protein-only structures, we develop multikernel FRIs with multiple kernels to improve accuracy of prediction on protein-nucleic acid structures. In order to simplify the FRI method, we try to develop an accurate parameter-free version for a two-kernel mFRI. We use a combination of one Lorentz and one exponential kernel. Values for parameters $\upsilon$ and $\kappa$ are set to 3.0 and 1.0 respectively based on the results of previous FRI studies [@Opron:2014]. The optimal values for $\eta$ in both kernels are determined by testing a range of possible values from 2 to 20 Å. All three representations (M1, M2 and M3) described previously are considered. The results of these tests on the set of 203 protein-nucleic acid structures are shown in Figure \[multi\_etasigma\]. [width=,totalheight=0.93,keepaspectratio]{} -------- ------- ------ ------- ------ ------- ------- PDB ID CC CC CC 1asy 0.647 1114 0.645 1248 0.631 1382 1b23 0.751 471 0.774 537 0.714 603 1c0a 0.763 653 0.704 721 0.598 789 1CX0 0.821 162 0.763 234 0.627 306 1drz 0.846 162 0.754 234 0.585 306 1efw 0.537 1286 0.647 1412 0.660 1538 1egk\* 0.273 104 0.298 212 0.267 320 1ehz\* 0.623 62 0.706 124 0.722 186 1evv\* 0.710 62 0.769 124 0.770 186 1f7u 0.577 670 0.588 734 0.603 798 1ffk 0.759 6482 0.793 9310 0.809 12138 1ffy 0.520 991 0.549 1066 0.568 1141 1fg0\* 0.720 498 0.723 996 0.721 1494 1fir\* 0.687 61 0.576 122 0.439 183 1fjg 0.461 3915 0.585 5428 0.600 6941 1gid\* 0.649 316 0.643 632 0.583 948 1gtr 0.724 603 0.747 677 0.645 751 1h3e 0.717 507 0.724 586 0.645 663 1h4s 0.671 1011 0.704 1076 0.626 1141 1hr2\* 0.599 313 0.589 628 0.585 943 1i94 0.489 3923 0.615 5437 0.652 6951 1i9v\* 0.615 73 0.631 147 0.642 220 1j1u 0.730 372 0.671 446 0.456 520 1j2b 0.686 1300 0.712 1448 0.672 1596 1j5a 0.532 3158 0.548 5932 0.510 8706 1j5e 0.427 3909 0.546 5422 0.553 6935 1jj2 0.799 6567 0.839 9443 0.836 12319 1jzx 0.586 3158 0.600 5932 0.561 8706 1l8v\* 0.700 312 0.688 626 0.672 940 1l9a 0.849 211 0.789 336 0.675 461 1lng 0.780 183 0.595 280 0.405 377 1m5k 0.904 402 0.841 622 0.760 842 1m5o 0.921 405 0.872 629 0.810 853 1mfq 0.773 341 0.688 468 0.543 595 1mms 0.507 317 0.548 433 0.646 549 1n32 0.388 3916 0.494 5447 0.517 6978 1nbs\* 0.547 270 0.566 540 0.573 810 1o0c 0.766 602 0.758 676 0.636 750 1qf6 0.608 710 0.578 779 0.540 848 1qrs 0.671 603 0.672 677 0.586 751 1qtq 0.620 602 0.640 676 0.596 750 1qu2 0.520 991 0.549 1066 0.568 1141 1qu3 0.579 954 0.599 1029 0.613 1104 1rc7 0.599 256 0.566 296 0.470 336 1s72 0.823 6636 0.839 9507 0.831 12378 1ser 0.748 855 0.743 917 0.657 978 1sj3 0.880 167 0.805 240 0.614 313 1tn2\* 0.686 62 0.712 124 0.676 186 1tra\* 0.624 62 0.670 124 0.660 186 1ttt 0.578 1401 0.564 1587 0.515 1773 1u0b 0.757 535 0.754 609 0.621 683 1u6b 0.476 312 0.490 531 0.506 750 1u9s\* 0.446 155 0.432 310 0.419 465 1vby 0.877 167 0.792 240 0.587 313 1vc0 0.878 167 0.804 240 0.611 313 1vc5 0.861 164 0.840 234 0.685 304 1y0q\* 0.491 230 0.484 463 0.472 696 1y26\* 0.677 70 0.697 141 0.709 212 1yfg\* 0.565 64 0.600 128 0.623 192 1yhq 0.835 6636 0.840 9507 0.831 12378 1yij 0.836 6636 0.851 9507 0.842 12378 2tra\* 0.614 65 0.614 130 0.613 195 3tra\* 0.645 64 0.615 128 0.620 192 4tra\* 0.679 62 0.715 124 0.694 186 -------- ------- ------ ------- ------ ------- ------- : Correlation coefficients (CCs) between predicted and experimental B-factors for the set of 64 protein-nucleic structures [@Yang:2006]. Here N1, N2 and N3 values represent the number of atoms used for the M1, M2 or M3 representations for each structure. We use the parameter-free two-kernel mFRI model, i.e., one exponential kernel ($\kappa=1$ and $\eta=18$ Å) and one Lorentz kernels ($\upsilon=3$, $\eta=18$ Å. [PDB IDs marked with an asterisk (\*) indicate structure containing only nucleic-acid residues. ]{} \[table:64set\] PDB ID PDB ID PDB ID PDB ID PDB ID PDB ID PDB ID PDB ID PDB ID PDB ID -------- -------- -------- -------- -------- -------- -------- -------- -------- -------- 1A1H 1A1I 1AAY 1AZP 1BF4 1C8C 1D02 1D2I 1DC1 1DFM 1DP7 1DSZ 1EGW 1EON 1F0V 1FIU 1H6F 1I3W 1JK2 1JX4 1K3W 1K3X 1L1Z 1L3L 1L3S 1L3T 1L3V 1LLM 1MNN 1NJX 1NK0 1NK4 1OJ8 1ORN 1PFE 1QUM 1R2Z 1RFF 1RH6 1SX5 1T9I 1U4B 1VTG 1WTO 1WTQ 1WTV 1XJV 1XVK 1XVN 1XVR 1XYI 1ZS4 2ADW 2AXY 2BCQ 2BCR 2BOP 2C62 2C7P 2EA0 2ETW 2EUW 2EUX 2EUZ 2EVF 2EVG 2FMP 2GB7 2HAX 2HEO 2HHV 2IBT 2IH2 2ITL 2NQ9 2O4A 2OAA 2ODI 2P2R 2PY5 2Q10 2R1J 2VLA 2VOA 2WBS 2XHI 2Z70 2ZKD 3BIE 3BKZ 3BM3 3BS1 3D2W 3EY1 3EYI 3FC3 3FDE 3FDQ 3FSI 3FYL 3G00 3G9M 3G9O 3G9P 3GO3 3GOX 3GPU 3GQ4 3HPO 3HT3 3HTS 3I0W 3I2O 3I3M 3I49 3I8D 3IGK 3JR5 3JX7 3JXB 3JXY 3JXZ 3KDE 3KXT 3M4A 3MR3 3MXM 3NDH 3O1M 3O1P 3O1S 3O1T 3O1U 3OQG 3PV8 3PVI 3PX0 3PX4 3PX6 3PY8 3QEX 3RKQ 3RZG 3S57 3S5A 3SAU 3SJM 3TAN 3TAP 3TAQ 3TAR 3THV 3TI0 3U6E 3U6P 3V9W 3ZDA 3ZDB 3ZDC 3ZDD 4A75 4B21 4B9S 4DFK 4DQI 4DQP 4DQQ 4DS4 4DS5 4DSE 4DSF 4E0D 4ECQ 4ECV 4ECX 4ED0 4ED2 4ED7 4ED8 4EZ6 4F1H 4F2R 4F2S 4F3O 4F4K 4F8R 4FPV 4GZ1 4GZN 4HC9 4HIK 4HIM 4HLY 4HTU 4HUE 4HUF 4HUG 4IBU 4IX7 4KLG 4KLI 4KLM 4KMF : The PDB IDs of the 203 high resolution protein-nucleic structures used in our single-kernel FRI parameter test. IDs marked with an asterisk indicate those containing only nucleic acids residues. \[table:203set\] As expected, the addition of another kernels results in an overall increase in accuracy for the 203 complex set. For two-kernel mFRI, the MCCs increase up to 0.68 for M1, 0.67 for M2 and 0.63 for M3. The choice of $\eta$ turns out to be very robust based on results shown in Figure \[multi\_etasigma\]. We have also carried out a similar test of two-kernel mFRI ($\upsilon=3.0$ and $\kappa=1.0$) for the set of 64 protein-nucleic acid structures. Note that this has many large complexes. The MCCs for M1, M2 and M3 models are 0.668, 0.666 and 0.620, respectively, which are similar to what we have found for the set of 203 structures. [The set of 64 structures includes 19 structures composed of nucleic acids and no amino acids. The MCCs for this nucleic acid-only subset 0.608, 0.617 and 0.603 for M1, M2 and M3 models. The correlation coefficients for all 64 individual molecular complexes are listed in Table \[table:64set\].]{} To summarize the performance of Gaussian network model, single kernel FRI, and two-kernel mFRI, we list their MCCs for the 64 protein-nucleic acid structures in Table \[table:compare\]. It can be seen that, the FRI outperforms GNM in all three representations, and two-kernel mFRI further significantly improves the accuracy of our method and achieves up to 15% improvement compared with GNM [@Yang:2006]. Based on our earlier test [@Opron:2015a], we believe that our three-kernel mFRI can deliver a better prediction. [**GNM**]{}[@Yang:2006] [**FRI**]{} [**Two-kernel mFRI**]{} ---- ------------------------- ------------- ------------------------- M1 0.59 0.620 0.666 M2 0.58 0.612 0.668 M3 0.55 0.555 0.620 : MCCs of Gaussian network model (GNM) [@Yang:2006], single kernel flexibility-rigidity index (FRI) and two-kernel mFRI for three coarse-grained representations (M1, M2,and M3). A set of 64 protein-nucleic acid structures [@Yang:2006] is used. \[table:compare\] Applications {#Sec:CaseStudy} ============= In this section we briefly explore the applications of the mFRI and aFRI methods to large protein-nucleic acid complexes. We highlight a few particular examples where mFRI improves upon previous FRI methods, in particular, for the flexibility prediction of ribosomes. Further, we show how aFRI is well suited for the study of the dynamics of large macromolecular complexes using the bacterial RNA polymerase active site as an example. Multikernel FRI flexibility prediction for protein-nucleic acid structures - ribosomes {#sec:ribo} -------------------------------------------------------------------------------------- Some of the largest and most biologically important structures that contain both protein and nucleic acids are ribosomes. Ribosomes are the protein synthesizers of the cell and connect amino acid into polymer chains. In ribosomes, proteins and RNA interact through intermolecular effects, such as electrostatic interactions, hydrogen bonding, hydrophobic interactions, base stacking and base pairing. RNA tertiary structures can significantly influence protein-RNA interactions. Ribosomes are primarily composed of RNA with many smaller associated proteins as shown in Fig. \[ribosome3\]. The top of Fig. \[ribosome3\] shows the 50S subunit of the ribosome (PDB ID: 1YIJ) with the nucleic acids in a smooth surface representation with the protein subunits bound and shown in a secondary structure representation. The set of 64 structures used in our tests contains a number of ribosomal subunits. Due to their multiscale nature, these structures also happen to be among those that benefit the most from using multikernel FRI over single kernel FRI or GNM. For example, in the case of ribosome 50S subunit structure (PDB ID:1YIJ), B-factor prediction with three-kernel FRI yields a CC value of 0.85, while that of single kernel FRI is only around 0.3. GNM does not provide a good B-factor prediction for this structure either. The three-kernel mFRI model we used is one exponential kernel ($\kappa=1$ and $\eta=15$ Å) and two Lorentz kernels ($\upsilon=3$, $\eta=3$ Å  and $\upsilon=3$, $\eta=7$Å). The comparison between mFRI-predicted and experimental B-factors for ribosome 50S subunit structure is demonstrated in Fig. \[ribosome3\]. By using the fitting coefficients from the above 50S subunit (1YIJ) flexibility analysis, we have obtained flexibility predictions for the entire ribosome (PDB ID:4V4J) as well as many protein subunits and other RNAs that associate with it, see Fig. \[ribosome3\]. To avoid confusion, the B-factors for 4V4J are uniquely determined by using not only the same three-kernel mFRI model from the case 1YIJ, and also its fitting parameters, i.e., $a^1=, a^2=, a^3,$ and $b$. Again, the FRI values are mapped by color to the smooth surface of the nucleic acids, however, in these bottom figures the protein subunits are omitted to draw attention instead to the various types of RNA involved in this structure. Anisotropic FRI for conformational motion prediction of an RNA polymerase {#sec:RNAP} ------------------------------------------------------------------------- RNA polymerase is one of the essential enzymes for all life on Earth as we know it today and possibly from the very beginning of life [@Iyer2003; @Burton2014old]. Despite this importance, the mechanisms for many of the polymerase’s functions are still not well understood on the atomic level. Considerable effort has been spent both experimentally and computationally to understand RNAP polymerase function in more detail but many questions remain. The study of RNA polymerase experimentally or computationally is difficult and often expensive due to the size of the system and variety of molecules involved. The minimal required elements for a bacterial or eukaryotic RNA polymerase include multiple protein subunits, a double stranded DNA molecule, a single stranded RNA molecule, free nucleotides, various ions (Mg$^{2+}$, Zn$^{2+}$, Na$^+$ etc.) and solvent. A typical setup for this system in all-atom molecular dynamics includes  300,000 atoms when solvated. With this number of atoms and current computer power, it is often not feasible to simulate these molecules on biologically relevant timescales using MD. Perhaps the most popular tool for studying long time dynamics of biomolecules is normal mode analysis (NMA) and its related methods such as the anisotropic network model (ANM). These methods have been successfully used to study protein dynamics for many proteins, however, at their maximum accuracy, their computational complexity is of ${\cal O}(N^3)$, where $N$ is the number of atoms. This is a problem because many cellular functions involve a large number of macromolecules with many thousands to millions of residues to consider. Therefore, future computational studies of biomolecules beyond the protein scale will require methods with better scaling properties such as FRI and aFRI. In this example, we use completely local anisotropic FRI to examine correlated motions in regions near the active site of bacterial RNA polymerase, including the bridge helix, trigger loop and nucleic acid chains. We examine the relationship between these components’ motions and their contributions to critical functions such as catalysis and translocation. We use the anisotropic rigidity form in Sec. \[sec:arFRI\] with the Lorentz kernel ($\upsilon=2$ and $\eta=3$ Å). Figure \[RNAP\_flex\]a is a simplified representation of RNA polymerase (PDB ID 2PPB) that shows these important features which are buried in the core of the largest protein subunits, $\beta$ and $\beta'$. The bridge helix and trigger loop, shown in green and blue respectively, are parts of the protein that have been implicated in most of the essential functions of the polymerase. Mutational studies of these regions result in modulation of the polymerase speed and accuracy, both positively and negatively, indicating the regions are important for normal functioning of the enzyme. How these regions aid these functions and how they interact remains an open question. With this demonstration of local aFRI analysis we hope to shed some light on how these essential parts of RNA polymerase work together. Local aFRI, as described in earlier work, is much less computationally costly than global aFRI or NMA and has been shown to have qualitatively similar results for small to large size single proteins. To further validate the local aFRI method we compare the conclusions from a local aFRI study of RNAP to those of NMA based studies. The RNA polymerase elongation complex is a relatively large system but it is still tenable for NMA methods. NMA has been applied to both bacterial and eukaryotic RNA polymerase in the past [@Van2004; @Feig2010rna] which provides us with a point of comparison for our results. Local aFRI produces three modes of motion sorted from lowest to highest frequency vibration according to eigenvalue as in NMA. In Figure \[RNAP\_flex\] we present findings from the lowest frequency mode effectively focusing on the most dominant motion of each conformation. Two major conformations of RNA polymerase are considered, those with open and closed trigger loop regions (Figures \[RNAP\_flex\]c and \[RNAP\_flex\]d.) A closed trigger loop is one that is completely folded into two parallel alpha helices while an open trigger loop has a region of disordered loop between two shorter helices and is slightly bent away from the bridge helix . The closing or folding of the trigger loop into the closed conformation is assumed to follow binding of an NTP in the active site and to precede catalysis. After catalysis, it is suspected that the trigger loop opens or unfolds to facilitate translocation and permit new NTPs to enter the active site. The results of aFRI analysis on the effect of trigger loop closing reveal a distinct change in correlated motions in open and closed trigger loop conformations. These changes involve interactions between the bridge helix, the trigger loop and the nucleic acid regions. In Figure \[RNAP\_flex\]b, regions of high correlation are color coded which reveals that the bridge helix is composed of two highly self correlated portions suggesting the presence of a hinge in the bridge helix. In fact, the central portion of the bridge helix has been observed as a kinked or bent helix in a yeast RNAP structure[@Wang2006]. Additionally, it is observed that a portion of the bridge helix and the N-terminal helix of the trigger loop are highly correlated in the closed trigger loop structure only. This set of two helices is situated directly next to the active site and could provide stability to aid catalysis after trigger loop closing. Additionally, correlation between nucleic acids and protein shows marked differences from the open trigger loop to closed trigger loop structures. The motions indicated in Figures \[RNAP\_flex\]c and \[RNAP\_flex\]d show that the open trigger loop structure is primed to translocate based on the direction of highly correlated motions of the upstream and downstream nucleic acids. By contrast, the closed trigger loop nucleic acid motions are considerably less correlated and not in the direction of translocation. This is the expected relationship as it matches the results from previous biological and NMA studies of RNA polymerase. [@Feig2010rna] These differences between a closed trigger loop and open trigger loop structure reveal potentially important structural changes that arise as the RNA polymerase switches between open and closed trigger loop conformations during the transition between translocation and catalysis. Specifically, the results for the closed trigger loop conformation suggest the presence of a stabilized catalytic area which is made of the N-terminal helix of the trigger loop and the bridge helix. The results for the open trigger loop conformation show no such coordination of the active site helices and instead indicates a less defined hinge and coordinated motion in the direction of translocation. Taken together these results provide a potential explanation for how trigger loop opening and closing is correlated with translocation and catalysis respectively. Concluding remarks {#sec:Conclusion} ================== Protein-nucleic acid complexes are essential to all living organisms. The function of these complexes depends crucially on their flexibility, an intrinsic property of a macromolecule. However, for many large protein-nucleic acid complexes, such as ribosomes and RNA polymerases, the present flexibility analysis approaches can be problematic due to their computational complexity scaling of ${\cal O}(N^3)$ and neglecting multiscale effects. This work introduces the flexibility-rigidity index (FRI) methods [@KLXia:2013d; @Opron:2014; @Opron:2015a] for the flexibility analysis of protein-nucleic acid structures. We show that a multiscale FRI (mFRI) realized by multiple kernels parametrized at multiple length scales is able to significantly outperform the Gaussian network model (GNM) for the B-factor prediction of a set of 64 protein-nucleic acid complexes [@Yang:2006]. The FRI methods are not only accurate, but also efficient, as their computational complexity scales as ${\cal O}(N)$. Additionally, anisotropic FRI (aFRI), which has cluster Hessian matrices, offers collective motion analysis for any given cluster, i.e, subunit or domain in a biomolecular complex. We apply FRI methods to a large ribosomal subunit (1YIJ) with multiple subunits. We note that both original single-scale FRI and GNM do not work well for this structure. It is found that that the multiscale strategy is crucial for the flexibility analysis of multi-subunit structures. The correlation coefficients between FRI predictions and experimental B-factors for 1YIJ improve from 0.3 for single-scale FRI to 0.85 for multiscale FRI. We further use the fitting coefficients obtained from 1YIJ to predict the flexibility of a entire ribosome, 4V4J. We found that mFRI has an advantage for analyzing large biomolecular complexes due to both higher speeds and accuracy. We have also demonstrated the utility of the anisotropic FRI (aFRI) for analyzing the translocation of an RNA polymerase, which involves protein, DNA, RNA, nucleotide substrates and various ions. Both experimental and computational studies of RNA polymerases are difficult and expensive due to the size and complexity of the biomolecular complex. The molecular mechanism of RNA polymerase translocation is an interesting, open research topic. The present work makes use of localized aFRI to elucidate the synergistic local motions of a bacterial RNA polymerase. Our findings are consistent with those from much more expensive molecular dynamics simulations and normal mode analysis [@Feig2010rna; @Feig2010rna2]. [The study of hinges has been an important topic and much research has been done in the past [@hingeprot; @flexoracle; @hingeatlas; @stonehinge; @flexprot]. Identification of hinge residues is useful for inferring motion and function when molecules are too large for MD simulation on relevant timescales. Other methods, such as GNM and NMA have been utilized. FRI-based methods could place a significant role in hinge analysis. This aspect will be carefully analyzed in our future work. ]{} Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported in part by NSF grants, IIS-1302285, and DMS-1160352, NIH Grant R01GM-090208 and Michigan State University Center for Mathematical Molecular Biosciences Initiative. The authors acknowledge the Mathematical Biosciences Institute for hosting valuable workshops. [10]{} M. P. Allen and D. J. Tildesley. . Oxford: Clarendon Press, 1987. C. B. Anfinsen. Einfluss der configuration auf die wirkung den. , 181:223 – 230, 1973. A. R. Atilgan, S. R. Durrell, R. L. Jernigan, M. C. Demirel, O. Keskin, and I. Bahar. Anisotropy of fluctuation dynamics of proteins with an elastic network model. , 80:505 – 515, 2001. I. Bahar, A. R. Atilgan, M. C. Demirel, and B. Erman. Vibrational dynamics of proteins: Significance of slow and fast modes in relation to function and stability. , 80:2733 – 2736, 1998. I. Bahar, A. R. Atilgan, and B. Erman. Direct evaluation of thermal fluctuations in proteins using a single-parameter harmonic potential. , 2:173 – 181, 1997. B. R. Brooks, R. E. Bruccoleri, B. D. Olafson, D. States, S. Swaminathan, and M. Karplus. Charmm: A program for macromolecular energy, minimization, and dynamics calculations. , 4:187–217, 1983. Z. F. Burton. The old and new testaments of gene regulation: Evolution of multi-subunit rna polymerases and co-evolution of eukaryote complexity with the rnap ii ctd. , 5(3), 2014. Q. Cui and I. Bahar. . Chapman and Hall/CRC, 2010. Q. Cui, G. J. Li, J. Ma, and M. Karplus. A normal mode analysis of structural plasticity in the biomolecular motor f(1)-atpase. , 340(2):345 – 372, 2004. O. N. A. Demerdash and J. C. Mitchell. . , [80]{}([7]{}):[1766–1779]{}, [JUL]{} [2012]{}. O. K. Dudko, G. Hummer, and A. Szabo. Intrinsic rates and activation free energies from single-molecule pulling experiments. , 96:108101, 2006. U. Emekli, S. Dina, H. Wolfson, R. Nussinov, and T. Haliloglu. automated prediction of hinges in protein structures. , 70(4):1219–1227, 2008. M. Feig and Z. F. Burton. Rna polymerase ii flexibility during translocation from normal mode analysis. , 78(2):434–446, 2010. M. Feig and Z. F. Burton. Rna polymerase ii with open and closed trigger loops: active site dynamics and nucleic acid translocation. , 99(8):2577–2586, 2010. S. Flores and M. Gerstein. predicting flexible hinges by identification of stable domains. , 8(1), 2007. S. Flores, L. Lu, J. Yang, N. Carriero, and M. Gerstein. Hinge atlas: relating protein sequence to sites of structural flexibility. , 8, 2007. P. J. Flory. Statistical thermodynamics of random networks. , 351:351 – 378, 1976. H. Frauenfelder, S. G. Slihar, and P. G. Wolynes. . , [254]{}([5038]{}):[1598–1603]{}, [DEC 13]{} [1991]{}. Z. N. Gerek and S. B. Ozkan. A flexible docking scheme to explore the binding selectivity of pdz domains. , 19:914–928, 2010. N. Go, T. Noguti, and T. Nishikawa. Dynamics of a small globular protein in terms of low-frequency vibrational modes. , 80:3696 – 3700, 1983. B. Halle. Flexibility and packing in proteins. , 99:1274–1279, 2002. K. Hinsen. Analysis of domain motions by approximate normal mode calculations. , 33:417 – 429, 1998. K. Hinsen. Structural flexibility in proteins: impact of the crystal environment. , 24:521 – 528, 2008. S. W. Huang, C. H. Shih, C. P. Lin, and J. K. Hwang. Prediction of nmr order parameters in proteins using weighted protein contact-number model. , 121(3-4):197–200, 2008. L. M. Iyer, E. V. Koonin, and L. Aravind. Evolutionary connection between the catalytic subunits of dna-dependent rna polymerases and eukaryotic rna-dependent rna polymerases and the origin of rna polymerases. , 3(1):1, 2003. D. J. Jacobs, A. J. Rader, L. A. Kuhn, and M. F. Thorpe. . , [44]{}([2]{}):[150–165]{}, [AUG 1]{} [2001]{}. K. S. Keating, S. C. Flores, M. B. Gerstein, and L. A. Kuhn. hinge prediction by network analysis of individual protein structures. , 18(2):359–371, 2009. O. Keskin, I. Bahar, D. Flatow, D. G. Covell, and R. L. Jernigan. Molecular mechanisms of chaperonin groel-groes function. , 41:491 – 501, 2002. D. A. Kondrashov, A. W. Van Wynsberghe, R. M. Bannen, Q. Cui, and J. G. N. Phillips. Protein structural variation in computational models and crystallographic data. , 15:169 – 177, 2007. M. Krone, J. E. Stone, T. Ertl, and K. Schulten. Fast visualization of [Gaussian]{} density surfaces for molecular dynamics and particle system trajectories. , 2012:67–71, 2012. S. Kundu, J. S. Melton, D. C. Sorensen, and J. G. N. Phillips. Dynamics of proteins in crystals: comparison of experiment with simple models. , 83:723 – 732, 2002. M. Levitt, C. Sander, and P. S. Stern. Protein normal-mode dynamics: Trypsin inhibitor, crambin, ribonuclease and lysozyme. , 181(3):423 – 447, 1985. D. W. Li and R. Br[ü]{}schweiler. All-atom contact model for understanding protein dynamics from crystallographic b-factors. , 96(8):3074–3081, 2009. G. H. Li and Q. Cui. A coarse-grained normal mode approach for macromolecules: an efficient implementation and application to [Ca(2+)-ATPase]{}. , 83:2457 – 2474, 2002. C. P. Lin, S. W. Huang, Y. L. Lai, S. C. Yen, C. H. Shih, C. H. Lu, C. C. Huang, and J. K. Hwang. Deriving protein dynamical properties from weighted protein contact number. , 72(3):929–935, 2008. D. R. Livesay, S. Dallakyan, G. G. Wood, and D. J. Jacobs. . , [576]{}:[468–476]{}, [2004]{}. J. P. Ma. Usefulness and limitations of normal mode analysis in modeling dynamics of biomolecular complexes. , 13:373 – 180, 2005. J. A. McCammon, B. R. Gelin, and M. Karplus. Dynamics of folded proteins. , 267:585–590, 1977. K. Opron, K. L. Xia, and G. W. Wei. Fast and anisotropic flexibility-rigidity index for protein flexibility and fluctuation analysis. , 140:234105, 2014. K. Opron, K. L. Xia, and G. W. Wei. Communication: Capturing protein multiscale thermal fluctuations. , 142(211101), 2015. J. K. Park, R. Jernigan, and Z. Wu. Coarse grained normal mode analysis vs. refined gaussian network model for protein residue-level structural fluctuations. , 75:124 –160, 2013. A. J. Rader, D. H. Vlad, and I. Bahar. Maturation dynamics of bacteriophage hk97 capsid. , 13:413 – 421, 2005. M. Shatsky, R. Nussinov, and H. J. Wolfson. alignment of flexible protein structures without a predefinition of hinge regions. , 11(1):83–8106, 2004. L. Skjaerven, S. M. Hollup, and N. Reuter. Normal mode analysis for proteins. , 898:42 – 48, 2009. G. Song and R. L. Jernigan. vgnm: a better model for understanding the dynamics of proteins in crystals. , 369(3):880 – 893, 2007. F. Tama and C. K. Brooks III. Diversity and identity of mechanical properties of icosahedral viral capsids studied with elastic network normal mode analysis. , 345:299 – 314, 2005. F. Tama, F. X. Gadea, O. Marques, and Y. H. Sanejouand. Building-block approach for determining low-frequency normal modes of macromolecules. , 41(1):1–7, 2000. F. Tama and Y. H. Sanejouand. Conformational change of proteins arising from normal mode calculations. , 14:1 – 6, 2001. F. Tama, M. Valle, J. Frank, and C. K. Brooks III. Dynamic reorganization of the functionally active ribosome explored by normal mode analysis and cryo-electron microscopy. , 100(16):9319 – 9323, 2003. M. Tasumi, H. Takenchi, S. Ataka, A. M. Dwidedi, and S. Krimm. Normal vibrations of proteins: Glucagon. , 21:711 – 714, 1982. M. M. Tirion. Large amplitude elastic motions in proteins from a single-parameter, atomic analysis. , 77:1905 – 1908, 1996. M. Topf, K. Lasker, B. Webb, H. Wolfson, and W. C. A. Sali. Protein structure fitting and refinement guided by cryo-em density. , 16:295–307, 2008. A. Van Wynsberghe, G. Li, and Q. Cui. Normal-mode analysis suggests protein flexibility modulation throughout rna polymerase’s functional cycle. , 43(41):13083–13096, 2004. C. W. von der Lieth, K. Stumpf-Nothof, and U. Prior. A bond flexibility index derived from the constitution of molecules. , 36:711–716, 1996. D. Wang, D. A. Bushnell, K. D. Westover, C. D. Kaplan, and R. D. Kornberg. Structural basis of transcription: role of the trigger loop in substrate specificity and catalysis. , 127(5):941–954, 2006. Y. Wang, A. J. Rader, I. Bahar, and R. L. Jernigan. Global ribosome motions revealed with elastic network model. , 147:302 – 314, 2004. G. W. Wei. Wavelets generated by using discrete singular convolution kernels. , 33:8577 – 8596, 2000. W. Wriggers, R. A. Milligan, and J. A. McCammon. Situs: [A]{} package for docking crystal structures into low-resolution maps from electron microscopy. , 125:185–195, 1999. K. L. Xia, X. Feng, Y. Y. Tong, and G. W. Wei. Persistent homology for the quantitative prediction of fullerene stability. , 36:408–422, 2015. K. L. Xia, K. Opron, and G. W. Wei. Multiscale multiphysics and multidomain models — [ Flexibility]{} and rigidity. , 139:194109, 2013. K. L. Xia and G. W. Wei. A stochastic model for protein flexibility analysis. , 88:062709, 2013. K. L. Xia and G. W. Wei. Molecular nonlinear dynamics and protein thermal uncertainty quantification. , 24:013103, 2014. K. L. Xia and G. W. Wei. Persistent topology for [cryo-EM]{} data analysis. , 31:e02719, 2015. K. L. Xia, Z. X. Zhao, and G. W. Wei. Multiresolution topological simplification. , 22:1–5, 2015. C. Xu, D. Tobi, and I. Bahar. Allosteric changes in protein structure computed by a simple mechanical model: hemoglobin t &lt;–&gt; r2 transition. , 333:153 – 168, 2003. L. Yang, A. Rader, X. Liu, C. Jursa, S. Chen, H. Karimi, and I. Bahar. online computation of structural dynamics using the gaussian network model. , 34(Web Server issue):W24–W31, 2006. L. W. Yang and C. P. Chng. Coarse-grained models reveal functional dynamics–[I]{}. elastic network models–theories, comparisons and perspectives. , 2:25 – 45, 2008. F. L. Zhang and R. Br[ü]{}schweiler. Contact model for the prediction of nmr nh order parameters in globular proteins. , 124(43):12654–12655, 2002. W. Zheng, B. R. Brooks, and D. Thirumalai. Allosteric transitions in the chaperonin groel are captured by a dominant normal mode that is most robust to sequence variations. , 93:2289 – 2299, 2007. W. J. Zheng and S. Doniach. A comparative study of motor-protein motions by using a simple elastic-network model. , 100(23):13253 – 13258, 2003. [^1]: On leave from the Department of Mathematics, Michigan State University [^2]: Address correspondences to Guo-Wei Wei. E-mail:[email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'In Alzheimer’s disease (AD), the progressive atrophy leads to aberrant network reconfigurations both at structural and functional levels. In such network reorganization, the core and peripheral nodes appear to be crucial for the prediction of clinical outcome due to their ability to influence large-scale functional integration. However, the role of the different types of brain connectivity in such prediction still remains unclear. Using a multiplex network approach we integrated information from DWI, fMRI and MEG brain connectivity to extract an enriched description of the core-periphery structure in a group of AD patients and age- matched controls. Globally, the regional coreness - i.e., the probability of a region to be in the multiplex core - significantly decreased in AD patients as a result of the randomization process initiated by the neurodegeneration. Locally, the most impacted areas were in the core of the network - including temporal, parietal and occipital areas - while we reported compensatory increments for the peripheral regions in the sensorimotor system. Furthermore, these network changes significantly predicted the cognitive and memory impairment of patients. Taken together these results indicate that a more accurate description of neurodegenerative diseases can be obtained from the multimodal integration of neuroimaging-derived network data.' author: - Jeremy Guillon - Mario Chavez - Federico Battiston - Yohan Attal - Valentina La Corte - Michel Thiebaut de Schotten - Bruno Dubois - Denis Schwartz - Olivier Colliot - Fabrizio De Vico Fallani bibliography: - 'ms.bib' title: 'Disrupted core-periphery structure of multimodal brain networks in Alzheimer’s Disease' --- Introduction {#introduction .unnumbered} ============ The brain is a complex network where differently specialized areas are anatomically and functionally connected. Because of such interconnected structure, focal damages can affect the rest of the network through the interruption of communication pathways. Indeed, many neurological disorders affecting language, motor and sensory abilities are often due to a disconnection syndrome caused by the anatomical connectivity breakdown between the relevant brain areas [@SCHMAHMANN_2008; @geschwind_disconnexion_1965]. In the case of neurodegenerative diseases, the disconnection hypothesis is supported by a progressive death of neurons and synapses that induce gross atrophy. Empirical evidence has shown that Alzheimer’s disease (AD) patients with severe motor and cognitive impairments exhibited anatomical disconnections among regions between cerebral hemispheres that resemble those observed in split-brain subjects [@Delbeuck_2007; @Lakmache_1998]. In Parkinson’s disease (PD) intrahemispheric dissociations between subcortical and cortical structures have been linked to disturbances in cognition, perception, emotion, and sleep [@Cronin_Golomb_2010]. In addition, functional connectivity alterations within and between hemispheres have been reported in both AD [@Blinowska_2017; @Sankari_2010; @Adler_2003; @Babiloni_2009] and PD [@Luo_2015; @Olde_Dubbelink_2013] suggesting their potential role in the early diagnosis. Altogether, these findings suggest that neurodegenerative diseases must be considered as a network problem. Recent approaches based on network theory have greatly advanced our understanding of the connection mechanisms characterizing brain diseases [@Stam_2014]. Among others, decreased efficiency, modularity and hub centrality have been largely reported in neurodegeneration and associated with the stage of disease. Increasing evidence suggests that the core-periphery structure of the human connectome - supporting global integration of information among distant areas - is highly affected by the AD process and that resulting changes might be effective predictors of cognitive declines. On one hand, brain areas forming the core of the network - i.e. central and mutually connected nodes - have been reported to be preferentially attacked by AD [@Yan_2018]. On the other hand, brain regions forming the periphery of the network - i.e. nodes that are only weakly connected to the other units in the network - appear to be crucial for the degeneration [@Daianu_2015]. While these results refer to structural brain connectivity, the relative contribution of functional brain connectivity into the network core-periphery changes remain poorly understood. Based on the aforementioned empirical and theoretical grounds, we hypotesize that neurodegeneration would affect the core-periphery structure of the brain network at both anatomical and functional levels. More specifically, we expected that the extraction of the core-periphery organization by integrating information from multimodal brain networks would give more accurate predictor of AD and cognitive impairment. Finally, based on the evidence that hubs are the most attacked nodes in many neurological diseases and psychiatric disorders [@van_den_Heuvel_2013a], we hypothesize that the core brain regions would be mostly impacted by the AD atrophy process. To test these predictions, we considered multiple brain networks derived from DWI, fMRI and MEG data recorded in a group of AD patients and age-matched healthy controls (HC). Cognitive impairments in AD patients were described using multidomain behavioral measurements. We extracted the multimodal core-periphery structure of the brain networks through a multiplex network approach, where all the available information is kept at different connectivity layers. We evaluated how AD impacted the multiplex core-periphery organization and we tested the correlation of the regional coreness with the cognitive and memory impairment of patients. See Material and methods for more details on the experimental design and methods of analysis. ![image](figure1.png){width="\textwidth"} Results {#results .unnumbered} ======= Multimodal core of brain networks {#multimodal-core-of-brain-networks .unnumbered} --------------------------------- We integrated multimodal information by constructing nine-layer multiplex brain networks containing DWI, fMRI and MEG connectivity between 68 cortical regions of interest (ROIs) (Material and methods). To estimate the likelihood of each ROI $i$ to be in the multiplex core we computed its *coreness* $\mathcal{C}_i$ by counting how many times it was in the multiplex core across different density thresholds [@Battiston_2018]. At each threshold, the multiplex core-periphery structure was obtained by linearly combining the node strength of all the layers through a vector parameter $c$ (Material and methods). Because we do not know a-priori the best combination, we derived the optimal $c^{*}$ by using a data-driven approach that efficiently explores the parameter space to maximize the difference between AD and HC regional coreness. Specifically, we used the particles swarm optimization algorithm (PSO) to maximize the Fisher’s criterion $F(c)$ (Material and methods). Results show that the optimal $c^{*}$ components are found to be highly heterogenous and that the DWI layer, as well as MEG-alpha1 and fMRI layers, are the main contributors to separate the AD and HC group (Table \[corenesscoefftable\], Figure \[661774\]a). Layer $m$ $c^{*[m]}$ ------------------ ------------ MEG$_{\delta}$ 0.000 MEG$_{\theta}$ 0.001 MEG$_{\alpha_1}$ 0.258 MEG$_{\alpha_2}$ 0.000 MEG$_{\beta_1}$ 0.000 MEG$_{\beta_2}$ 0.002 MEG$_{\gamma}$ 0.000 fMRI 0.104 DWI 0.961 : [Vector of the optimal layers weight for the coreness computation.]{}[]{data-label="corenesscoefftable"} In the HC group, the multiplex core tended to include large portions of temporal, superior parietal and occipital cortices, and to a minor extent central and superior frontal regions (Figure \[661774\]b). On average AD patients exhibited a loss of coreness with respect to HC particularly in the temporal, superior parietal and occipital cortices. These regions were already known to form the core of multiplex brain networks derived from DTI and fMRI data [@Battiston_2018]. ![image](figure2.png){width="\textwidth"} Reorganization of core-periphery structure in AD {#reorganization-of-core-periphery-structure-in-ad .unnumbered} ------------------------------------------------ To quantify the observed network changes we defined the coreness disruption index $\kappa$ as the slope of the line obtained by regressing the difference between the average coreness (at each ROI and across subjects) of the two groups with the average coreness of the healthy one [@Termenon_2016] (Material and methods). We found a significant negative $\kappa$ value indicating that AD preferentially attack ROIs with a high coreness ($\kappa=-0.20,p=2.45\mathrm{e}{-10}$). This result was also consistent at the individual level when we extracted the coreness disruption index in each patient (Supplementary Table 1). In particular, by statistically comparing the average coreness of the two groups, we reported a significant decrease of coreness in core regions, such as temporal, parietal and occipital cortices as well as a significant increase of coreness in the right paracentral area which are instead more peripheral, ($p<0.025$, Figure \[482763\]a,b, Supplementary Table 2). ![image](figure3.png){width="\textwidth"} Based on the hypothesis that AD is a disconnection syndrome [@geschwind_disconnexion_1965; @delbeuck_alzheimers_2003] leading to disorganized network configurations [@sanz-arigita_loss_2010], we next generated a series of synthetic multiplex networks starting from the ones observed in the HC group and then progressively randomizing the same amount of links in each layer (Materials and methods). As expected the coreness disruption index decreased with the percentage of links that was randomly rewired. Notably, we could obtain the same significantly impacted $\kappa$ values observed in the multiplex brain networks of the AD group ($p=2.77\mathrm{e}{-4}$) by rewiring between 45% and 60% of the links in the HC multiplex brain networks (Figure \[722240\]). Altogether, these findings indicate that the AD is associated with a pervasive random reconfiguration of structural and functional connectivity that primarly affects the nodes of the multiplex core. ![[**Coreness disruption index as a function of network randomization**. Boxplots show the values of coreness disruption index ($\kappa$) obtained by randomly rewiring an increasing percentage of links in the multiplex brain networks of the HC group (see Materials and methods for more details). For example, $\text{RA}_{15}$ means that 15% of the links were reshuffled randomly in each layer. The blue and red boxplots illustrate respectively the $\kappa$ values for the HC and AD group. The circles in the boxes show the median; the bottom and top edges of the boxes denote the 25th and 75th percentile, respectively. Whiskers connect the most extreme points not considered outliers, and outliers are plotted individually as circles. [\[722240\]]{}]{}](figure4.png){width="250pt"} Coreness disruption predicts cognitive and memory deficits {#coreness-disruption-predicts-cognitive-and-memory-deficits .unnumbered} ---------------------------------------------------------- We finally conducted a correlation analysis to better understand how the observed multiplex brain network changes were associated with the behavioral performance of AD patients. Results show that both cognitive and memory deficits could be predicted by the individual loss of regional coreness. At the global scale, the coreness disruption index significantly correlated with the MMSE ($R=0.46,p=0.028$) as well as with the Immediate ($R=0.47,p=0.024$) and Free Recall ($R=0.59,p=0.005$) scores. The higher the $\kappa$ values, the better was the performance of the patients (Figure \[609234\]a, Supplementary material). At the local scale, temporal, parietal and cortices were highly positively correlated with the behavior of patients. Notably, these ROIs overlapped with those exhibiting significant decreases of regional coreness with respect to healthy controls (Figure \[482763\]b). We found similar positive correlations for bilateral middle frontal ROIs ($R=0.36,p=0.092$ for left, $R=0.35,p=0.100$ for right), while areas in the motor-system appeared not to be involved except for the paracentral lobule that tended to negatively correlate with the MMSE ($R=-0.55,p=0.007$) and Immediate Recall scores ($R=-0.36,p=0.089$) scores (Figure \[609234\]b). ![image](figure5.png){width="\textwidth"} Discussion {#discussion .unnumbered} ========== Multiplex brain networks {#multiplex-brain-networks .unnumbered} ------------------------ The increasing availability of multimodal neuroimaging data holds a great potential to enrich our knowledge about fundemental neural mechanisms and to improve the precision of predictive biomarkers of brain diseases [@Calhoun_2016]. However, how to integrate information from different neuroimaging modalities is still an open issue. Existing approaches have mainly focused on merging information at the level of the native data structure (e.g., signal or images) [@Uluda__2014; @Biessmann_2011]. Only recently, investigators have started to propose fusion algorithms in an effort to infer brain connectivity [@Ng_2012] or to detect mental states [@Lei_2011]. Here, we adopted a complementary solution - based on the nascent field of multilayer network theory - which preserves the original nature of the different connectivity types. Similar approaches have been already used in the case of temporal [@Betzel_2017; @De_Domenico_2017], multifrequency [@Guillon_2017; @de_domenico_mapping_2016] and DTI-fMRI brain networks [@Battiston_2017]. This study considers for the first time brain networks obtained from three different neuroimaging modalities - DWI, fMRI and MEG - to construct multiplex brain networks consisting of nine connectivity layers and to derive an augmented description of their core-periphery structure in healthy and Alzheimer’s diseased subjects. A crucial step in the multiplex construction is how to weight the different layers, which typically contain connectivity measured in units (e.g, number of fiber tracks and signal correlation) of different scales. While this is in general an arbitrary choice, here we established an objective way to associate a weight to each layer by maximizing the difference of regional coreness - i.e. the likelihood of each region to be in the core - between the groups. Results showed that all the three modalities are necessary to the group separation. In particular, for MEG only alpha1 was determinant while the other frequency layers had very low, or null, weights. This is in line with current evidence showing that alpha1 frequency band contains the most discriminant power and connectivity changes in AD [@Babiloni_2004; @Blinowska_2017]. Notably, DWI had a very high contribution coefficient as compared to the other layers. Core-periphery structure of diffusion-based networks is known to be very robust [@van_den_Heuvel_2011; @Hagmann_2008] with respect to functional layers and this might possibly depend on the heterogeneity of the node degree distribution. Further resarch is needed to better understand how to normalize connectivity weights between layers when construction multiplex network construction. Network reorganization in Alzheimer’s disease {#network-reorganization-in-alzheimers-disease .unnumbered} --------------------------------------------- AD is associated with network changes affecting the structure and function of the brain at multiple spatial and temporal scales [@Stam_2014]. It has been hypothesized that these network reconfigurations could result from dysconnection patterns initiated by the gross atrophy of the brain. While several studies have found significant changes in terms of network efficiency, modularity and node centrality, the direction of these alterations - in terms of increments or decrements with respect to healyhy controls - is often unclear and modality dependent [@Tijms_2013]. Here, we focused on the core-periphery structure of the human brain which has been shown to have a significant impact on cognition ensuring global integration across remote cortical areas [@van_den_Heuvel_2011]. Structural connectome studies have reported that AD patients, from the preclinical to dementia stages, have significant hub-concentrated lesion distributions [@Crossley_2014; @Buckner_2009; @Dai_2014; @Brier_2014; @Shu_2018]. However, recent evidence is suggesting that network disruption is prevalent in the peripheral network components in both AD [@Daianu_2015] and mild cognitive impairement (MCI) patients [@Zhao_2017]. These inconsistent findings suggest that the network disruption mechanisms remain unclear. By integrating information from structural and functional brain networks we aimed to overcome this controversy and provide a more comprehensive insight. Our multiplex network approach shows that core regions were globally affected in AD patients as compared to HC subjects and that this result could be modeled by a global random rewiring process. Specifically, we reported significant decrements of coreness in temporal and parietal cortices, which are heavily affected by atrophy processes and beta-amyloid deposition [@Buckner_2008]. However, this change was paralled by a significant increase of coreness in the paracentral lobules, which originally belonged to the multiplex periphery. Because regions of the sensorimotor system - such as paracentral lobule - are not directly affected by the atrophy process [@Agosta2010], we speculate that possible compensatory mechanisms could have therefore taken place. In line with this hypothesis, recent findings suggest that more efficient motor commands in mild cognitive impaired patients could trigger the later functional decline [@Kubicki_2016]. Longitudinal studies involving healthy subjects converting into AD will be fundamental to confirm or reject this prediction [@Dubois_2016]. Connectivity-based biomarkers of clinical behavior {#connectivity-based-biomarkers-of-clinical-behavior .unnumbered} -------------------------------------------------- [\[490891\]]{} Brain wiring organization is critically associated with human cognition, behavior as well as with several neurological and psychiatric disorders [@Stam_2014]. Network indices describing core-periphery  and rich-club organization in structural brain networks have been shown to predict cognitive and motor deficits in multiple sclerosis [@Stellmann_2017], and Huntington disease [@Harrington_2015], as well as communication impairment in schizophrenia [@van_den_Heuvel_2013]. More pertinent to this work, rich-club biomarkers extracted from DTI networks have been shown to correlate with cognitive and memory deficits in Alzheimer’s disease [@Daianu_2015; @tijms_alzheimers_2013; @stam_graph_2009]. Here, we showed that the coreness disruption index - quantifying the global tendency to weaken core-periphery structure in multimodal brain networks - determined the cognitive and memory performance of our AD patients. Patients with a stronger core-periphery organization had better MMSE and FSCRT scores. At the local scale, temporal, parietal as well as frontal areas tended to positively correlate with patient’s behavior. These association regions have been shown to be implicated in the prediction of AD cognitive performance [@Khachiyants_2012] and more in general in memory and language [@Squire_1992; @Gordon_1995; @Pochon_2002]. We also found negative correlations with the paracentral lobule (especially right), a region that is typically involved in motor-related tasks but not in integrative functions. From a network perspective, the coreness of regions that tended to be in the multiplex core - such as temporal, parietal and occipital cortices - were positively correlated with patients’ performance, while among the peripheral areas the paracentral lobule was negatively correlated with the behavior. This means that in presence of more severe cognitive and memory deficits the relative decrease of connectivity in core regions tended to be replaced by periphery components of the brain system. This result would confirm the existence of an adjusting mechanism, where the sensorimotor system might be involved in the compensation of connectivity loss in systems that are directly impacted by Amyloid-beta plaques and tau neurofibrillary tangles accumulation [@Iaccarino2018]. Methodological considerations {#methodological-considerations .unnumbered} ----------------------------- [\[624728\]]{} The basic algorithm behind the detection of the core-periphery structure in multiplex networks is purely deterministic [@Ma_2015]. This means that in principle we could not evaluate the statistical relevance of the identified structure. To overcome this limitation, we adopted a procedure that consisted in extracting the core-periphery structure from a series of multiplex networks obtained by filtering the actual brain multiplex network with increasing density thresholds [@Battiston_2018]. This way we could derive a probabilistic measure of coreness by counting how many times each ROI was assigned to the core across all the possible thresholds. For the sake of simplicity we filtered each brain multiplex by retaining the strongest links so that the average node degree of each layer ranged from $k=1$ to $k=N-1$ (Materials and methods). After filtering we did not binarize the surviving links so that we applied the core-periphery algorithm to sparse weighted multiplex networks. This approach allows us to exploit all the available information in the multiplex brain networks. At the same time, we remark that additional care is needed, as it introduces issues related to the different nature and distributions of the link weights [@Buld__2018]. Here, we mitigated this problem by using in the core-periphery algorithm the vector $c$ of parameters that can weight the contribution of each layer (Material and methods). Alternative solutions have been recently proposed taking into account the normalization of the weight across the layers by means of singular value decomposition [@Mandke_2018]. We used an optimization algorithm - namely the particle swarm optimization - to find the best combination of $c$ components that maximized the difference of the coreness between AD and HC subjects (Material and methods). This method presents two limitations that are important to mention here. First, the time complexity increases exponentially with the number of layers $$$M$ in order to find a stable solution. We verified that for $M>10$ the research complexity becomes rapidly intractable due to the large space of parameter combination to explore. Second, the cost function optimized by the algorithm and used to evaluate how segregated the two groups are (i.e. two sets of coreness vectors) should be carefully chosen as its accuracy is highly impacted by the size of the feature space (here $N$, the number of ROIs) and the size of the samples (here the size of the cohorts). More advanced techniques taking into account the possible nonlinear and/or non-Euclidean nature of the feature space should be considered for very large networks  (e.g., support vector machines, Riemannian geometry). Conclusion {#conclusion .unnumbered} ---------- [\[786908\]]{} Consistent with our hypothesis, we have shown that AD atrophy process generates multimodal connectivity changes that can be quantified by a multilayer network approach. Specifically we have identified that both core and - to a minor extent - peripheral cortical areas are affected in AD, and that the direction of the effect was opposite. Decrease of coreness in temporal, parietal and occipital areas - forming the rich core of the human brain - is paralleled by a possible compensatory increment in cortical regions that are in the sensorimotor system and that are more peripheral. These cortical network signatures varied over individuals and were significant predictors of cognitive and memory deficits.  Furthermore, we reported a general framework for the statistical comparison of core-periphery organization in arbitrary multiplex networks. Taken together, our results offer new insights into the crucial role of core-periphery organization in neurodegenerative diseases. Material and methods {#material-and-methods .unnumbered} ==================== Cohort inclusion {#cohort-inclusion .unnumbered} ---------------- [\[904153\]]{} The study involved 23 Alzheimer’s diseased (AD) patients (13 women) and 26 healthy age-matched control (HC) subjects (19 women). All participants underwent the Mini-Mental State Examination (MMSE) for global cognition and the Free and Cued Selective Reminding Test (FCSRT) for verbal episodic memory. Inclusion criteria for all participants were: *i)* age between 50 and 90; *ii)* absence of general evolutive pathology; *iii)* no previous history of psychiatric diseases; *iv)* no contraindication to MRI examination; *v)* French as a mother tongue. Specific criteria for AD patients were: *i)* clinical diagnosis of Alzheimer’s disease; *ii)* Mini-Mental State Examination (MMSE) score greater or equal to 18. All subjects gave written informed consent for participation in the study, which was approved by the local ethics committee of the Pitie-Salpetriere Hospital. All experiments were performed in accordance with relevant guidelines and regulation. Data acquisition and pre-processing {#data-acquisition-and-pre-processing .unnumbered} ----------------------------------- [\[577935\]]{} Magnetic resonance imaging (MRI) acquisitions were obtained using a 3T system (Siemens Trio, 32-channel system, with a 12-channel head coil). The MRI examination included: *(i)* 3D T1-weighted volumetric magnetization-prepared rapid gradient echo (MPRAGE) sequence with the following parameters: thickness = 1 mm isotropic, repetition time (TR) = 2300 ms, echo time (TE) = 4.18 ms, inversion time (TI) = 900 ms, acquisition matrix = 256 × 256; (*ii) *echo planar imaging (EPI) sequence with the following parameters: one image with no diffusion sensitization (b0 image) and 50 diffusion-weighted images (DWI) at b = 1500 s/mm^2^ , thickness = 2 mm isotropic, TR = 13000 ms, TE = 92 ms, flip angle = 90°, acquisition matrix = 128 × 116; *(iii) *functional MRI (fMRI) resting-state sequence [sensitive to blood oxygenation level-dependent (BOLD) contrast with the following parameters: ]{}200 images, thickness = 3 mm isotropic, TR = 2400 ms, TE = 30 ms, flip angle = 90°, acquisition matrix = 64 × 64. All MR All MR images were processed using the [Clinica](http://clinica.run) software (<http://www.clinica.run>). We first used the `t1-freesurfer-cross-sectional` pipeline to process T1-weighted images. to process T1-weighted images. This pipeline is a wrapper of different tools of the FreeSurfer software (<http://surfer.nmr.mgh.harvard.edu/>)  [@Fischl_2012]. It includes segmentation of subcortical structures, extraction of cortical surfaces, cortical thickness estimation, spatial normalization onto the FreeSurfer surface template (FsAverage), and parcellation of cortical regions. Functional MRI images pre-processing have been conducted using the `fmri-preprocessing` pipeline. Slice timing correction, head motion correction and unwarping have been applied using SPM12 tools ([www.fil.ion.ucl.ac.uk/spm](http://www.fil.ion.ucl.ac.uk/spm)). Separately the brain mask has been extracted from the T1 image of each subject using FreeSurfer. The resulting fMRI images have then been registered to the brain-masked T1 image of each subject using SPM’s registration tool. Finally, diffusion-weighted images have been processed using the `dwi-preprocessing` pipeline of Clinica. For each subject, all raw DWI volumes were rigidly registered (6 degrees of freedom (dof)) to the reference b0 image (DWI volume with no diffusion sensitization) to correct for head motion. The diffusion weighting directions were appropriately updated [@Leemans_2009]. An affine registration (12 dof) was then performed between each DWI volume and the reference b0 to correct for eddy current distortions. These registrations were done using the FSL flirt tool ([www.fmrib.ox.ac.uk/fsl](http://www.fmrib.ox.ac.uk/fsl)). To correct for echo-planar imaging (EPI) induced susceptibility artifacts, the field map image was used as proposed by [@Jezzard_1995] with the FSL prelude/fugue tools. Finally, the DWI volumes were corrected for nonuniform intensity using the ANTs N4 bias correction algorithm [@Tustison_2010]. A single multiplicative bias field from the reference b0 image was estimated, as suggested in [@Jeurissen_2014]. The magnetoencephalography (MEG) experimental protocol consisted in a resting-state with eyes-closed (EC). Subjects seated comfortably in a dimly lit electromagnetically and acoustically shielded room and were asked to relax. MEG signals were collected using a whole-head MEG system with 102 magnetometers and 204 planar gradiometers (Elekta Neuromag TRIUX MEG system) at a sampling rate of 1000 Hz and on-line low-pass filtered at 330 Hz. The ground electrode was located on the right shoulder blade. An electrocardiogram (EKG, Ag/AgCl electrodes) was placed on the left abdomen for artifacts correction and a vertical electrooculogram (EOG) was simultaneously recorded. Four small coils were attached to the participant in order to monitor head position and to provide co-registration with the anatomical MRI. The physical landmarks (the nasion, the left and right pre-auricular points) were digitized using a Polhemus Fastrak digitizer (Polhemus, Colchester, VT). We extracted three consecutive clean epochs of approximately 2 minutes each.  Signal space separation was performed using [MaxFilter](http://imaging.mrc-cbu.cam.ac.uk/meg/Maxfilter) to remove external noise. We used in-house software to remove cardiac and ocular blink artifacts from MEG signals by means of principal component analysis. We visually inspected the preprocessed MEG signals in order to remove epochs that still presented spurious contamination. At the end of the process, we obtained a coherent dataset consisting of three clean preprocessed epochs per subject. We reconstructed the MEG activity on the cortical surface by using a source imaging technique [@He1999; @Baillet2001]: *i) *We used the previously segmented T1-weighted images of each single subject [@Fischl2002; @Fischl2004]  to import cortical surfaces in the Brainstorm software [@Tadel2011] where they were modeled with approximately 20000 equivalent current dipoles (i.e., the vertices of the cortical meshes). *ii) *We applied the wMNE (weighted Minimum Norm Estimate) algorithm with overlapping spheres [@Lin2006] to solve the linear inverse problem. Both magnetometer and gradiometer, whose position has been registered on the T1 image using the digitized head points, were used to localize the activity over the cortical surface. Construction of brain networks {#construction-of-brain-networks .unnumbered} ------------------------------ [\[621327\]]{} We built, for each modality, one or multiple brain connectivity networks whose nodes are regions of interests (ROIs) defined by the Desikan cortical atlas parcellation [@Desikan_2006] ($N=68$ regions); and links are weighted by a given connectivity measure estimated between each pair of nodes resulting in $68\times68$ fully symmetric adjacency matrices. [In the case of MEG, we used the spectral coherence as a connectivity estimator with the following parameters: ]{} window length = 2 s, window type = sliding Hanning, overlap = 25% number of FFT points (NFFT) = 2000 for a frequency resolution of 0.5 Hz between 2 Hz and 45 Hz included[. ]{} We then averaged the connectivity matrices within the following characteristic frequency bands [@Stam2002; @Babiloni2004]: *delta* (2–4 Hz), *theta* (4.5–7.5 Hz), *alpha1* (8–10.5 Hz), *alpha2* (11–13 Hz), *beta1* (13.5–20 Hz), *beta2* (20.5–29.5 Hz) and *gamma* (30–45 Hz). We finally averaged the connectivity matrices across the three available epochs to obtain a robust estimate of the individual brain networks. For fMRI data, we focused our analysis on the scale 2 wavelet correlation matrices that represented - with a TR = 2400ms - the functional connectivity in the frequency interval 0.05–0.10Hz [@Biswal_1995; @Cordes2001; @Achard_2006]. For DWI data, we used the [Clinica](http://clinica.run) software to estimate the fiber orientation distributions (FODs) using constrained spherical deconvolution (CSD) algorithm from MRtrix3 `dwi2fod` tool and tractography based on iFOD2 algorithm from [MRtrix3 `tckgen` tool. The connectome is finally estimated by counting the number of tracts connecting each pair of nodes according to the given parcellation file using MRtrix3 ]{}`tck2connectome` tool.  Network methods and models {#network-methods-and-models .unnumbered} -------------------------- [\[303802\]]{} We constructed multiplex brain networks in each subject by spatially aligning DWI, fMRI and MEG source reconstructed connectivity networks. This led to the following multiplex network with $M=9$ layers: $$\mathcal{M} = \big\{ W^{[m]}, \forall m \in \{\textrm{MEG}_{\delta},...,\textrm{MEG}_{\gamma}, \textrm{fMRI}, \textrm{DWI}\} \big\},$$ where $W^{[m]} = \big\{w_{ij}^{[m]}\big\}$ is the connectivity matrix containing the weights of the connections between the ROIs $i$ and $j$ in the modality $m$. To extract the coreness of the nodes from the resulting multiplex networks, we followed the procedure described by @Battiston_2018. First, we filtered each layer by preserving the strongest weights for a broad range of increasing thresholds. Specifically, we considered density-based thresholds so that each layer had the same average node degree from $k=1$ to $k=N-1$. Then, for each threshold we computed the core-periphery of the filtered multiplex network by evaluating *(i)* the multiplex richness $\mu_i$ of node $i$, defined as follows: $$\mu_i = \sum_{m=1}^{M}{c^{[m]}s_i^{[m]}},$$ with $s_i^{[m]}$ the strength of the node in the $m$-th layer, and $c^{[m]}$ the components of the vector $c$ that modulate the contribution of each modality-specific layer. And *(ii)*, similarly to the original paper, we decomposed the richness function into two components based on the links of node $i$ that are going towards nodes with lower richness and those towards nodes with higher richness $s^{[m]}=s^{[m]-}+s^{[m]+}$. Thus, the multiplex richness of a node towards richer nodes is defined as follows: $$\mu_i^{+} = \sum_{m=1}^{M}{c^{[m]}s_i^{[m]+}}.$$ We finally counted the number of times that each node was in the core across all the explored thresholds and we normalized by the maximum theoretical value. As a result, we obtained the coreness $\mathcal{C}_i$ that can be written as follow: $$\mathcal{C}_i = \frac{1}{N-1} \sum_{k=1}^{N-1}{\delta^{[k]}_i},$$ where $$\delta^{[k]}_i = \begin{cases} 1, & \text{if node $i$ is in the core for the average node degree $k$}.\\ 0, & \text{otherwise}. \end{cases}$$ We generated random multiplex networks by shuffling a given percentage of links separately in each layer starting from the actual multiplex brain network of the HC group. This way the weight distribution was unchanged and only the topology of the network is impacted. Specifically, for each HC individual, we generated $n_{\textrm{rand}}$ new randomized multiplexes. We chose the minimum number of randomizations necessary to obtain a variance approximatively equal to the one observed in the HC and AD groups. This number was $n_{\textrm{rand}}=3$ and gave in total $N_{\textrm{RA}}=n_{\textrm{rand}} \times N_{\textrm{HC}}=78$ samples. Particles swarm optimization and statistical analysis {#particles-swarm-optimization-and-statistical-analysis .unnumbered} ----------------------------------------------------- We used the PSO algorithm [@Kennedy] under the MATLAB(R) software with the default parameters. The Fisher’s criterion $F(c)$ was defined as follow: $$F(c)= \frac{ \left(\bar I_{\textrm{AD}}(c) - \bar I_{\textrm{HC}}(c)\right)^{2} } { s_{\textrm{AD}}^{2} + s_{\textrm{HC}}^{2} },$$ with $\bar I_{\textrm{Pop}}(c)$, the average local (i.e. node level) index, here the coreness $\mathcal{C}$, over a population $\textrm{Pop}$, which in our case belongs to $\{\textrm{AD}, \textrm{HC}\}$, and, $$s_{\textrm{Pop}}^{2} = \sum_{s \in \textrm{Pop}}{ (I_{s}(c) - \bar I_{\textrm{Pop}}(c))^{2} },$$ with $s$ a subject belonging the population $\textrm{Pop}$. Since, in our case, $F(c) = F(ac), \forall a \in \mathbb{R}^{+}$, and in order to save one dimension in the searching space, we expressed the coefficient $c$ as a point on the positive section of the unitary hypersphere of dimension $M=9$ such that: $$c = \begin{pmatrix} \sin \phi_1 ... \sin \phi_{8} \\ \sin \phi_1 ... \sin \phi_{7} \cos \phi_{8} \\ \sin \phi_1 ... \sin \phi_{6} \cos \phi_{7} \\ \sin \phi_1 ... \sin \phi_{5} \cos \phi_{6} \\ \sin \phi_1 ... \sin \phi_{4} \cos \phi_{5} \\ \sin \phi_1 ... \sin \phi_{3} \cos \phi_{4} \\ \sin \phi_1 \sin \phi_{2} \cos \phi_{3} \\ \sin \phi_1 \cos \phi_{2} \\ \cos \phi_1 \end{pmatrix}, \phi_k \in \Big[0, \frac{\pi}{2}\Big], \forall k \in [1,M-1].$$ To consider the non-gaussian nature of the data we considered non-parametric statistics when assessing differences between populations and prediction of behavioral scores. To these ends, we used respectively permutation t-tests and Spearman correlation coefficients. The statistical threholds were set to $\alpha=0.05$ and we applied a rough false discovery rate (FDR) correction to account for the $N=68$ post-hoc tests at the level of brain regions ($\alpha_{FDR}=0.025$). Acknowledgements {#acknowledgements .unnumbered} ================ [\[427711\]]{} The research leading to these results has received funding from the program “Investissements d’avenir” ANR-10-IAIHU-06 (Agence Nationale de la Recherche-10-IA Institut Hospitalo-Universitaire-6), ANR-11-IDEX-004 (Agence Nationale de la Recherche-11- Initiative d’Excellence-004, project LearnPETMR number SU-16-R-EMR-16)  and from Agence Nationale de la Recherche (project HM-TC, grant number ANR-09-EMER-006). FD acknowledges support from the “Agence Nationale de la Recherche” through contract number ANR-15-NEUC-0006-02. The content is solely the responsibility of the authors and does not necessarily represent the official views of any of the funding agencies.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study bound states of two fermions with opposite spins in an extended Hubbard chain. The particles interact when located both on a site or on adjacent sites. We find three different types of bound states. Type U is predominantly formed of basis states with both fermions on the same site, while two states of type V originate from both fermions occupying neighbouring sites. Type U, and one of the states from type V, are symmetric with respect to spin flips. The remaining one from type V is antisymmetric. V-states are characterized by a diverging localization length below some critical wave number. All bound states become compact for wave numbers at the edge of the Brilloin zone.' author: - 'Jean-Pierre Nguenang$^{1,2}$' - 'Sergej Flach$^{1}$' title: 'Fermionic bound states on a one-dimensional lattice' --- [*Introduction.*]{} Advances in experimental techniques of manipulation of ultracold atoms in optical lattices make it feasible to explore the physics of few-body interactions. Systems with few quantum particles on lattices have new unexpected features as compared to the condensed matter case of many-body interactions, where excitation energies are typically small compared to the Fermi energy. In particular, a recent experiment explored the repulsive binding of bosonic atom pairs in an optical lattice [@Winkler2006Nature441], as predicted theoretically decades earlier [@aao69; @EilbeckPhysicaD78] (see also [@fg08] for a review). In this work, we study binding properties of fermionic pairs with total spin zero. We use the extended Hubbard model, which contains two interaction scales - the on site interaction $U$ and the nearest neighbour intersite interaction $V$. The nonlocal interaction $V$ is added in condensed matter physics to emulate remnants of the Coulomb interaction due to non-perfect screening of electronic charges. For fermionic ultracold atoms or molecules with magnetic or electric dipole-dipole interactions, it can be tuned with respect to the local interaction $U$ by modifying the trap geometry of a condensate, additional external dc electric fields, combinations with fast rotating external fields, etc (for a review and relevant references see [@mab08]). The paper is organized as follows. We first describe the model and introduce the basis we use to write down the Hamiltonian matrix to be diagonalized. Then we derive the quantum states of the lattice containing one and two fermions, with opposite spins in the latter case. We study the obtained bound states for two fermions. We obtain analytical expressions for the energy spectrum of the bound states. Some of the bound states are characterized by a critical momentum below which they dissolve with the two-particle continuum. [*Model and basis choice.*]{} We consider a one-dimensional lattice with $f$ sites and periodic boundary conditions described by the extended Hubbard model with the following Hamiltonian: $$\label{eq:hamiltonian} \hat{H} = \hat{H}_0 + \hat{H}_U + \hat{H}_V,$$ where $$\hat{H}_0=-\sum_{j,\sigma} \hat{a}^+_{j,\sigma} (\hat{a}_{j-1,\sigma} + \hat{a}_{j+1,\sigma} )\;,$$ $$\label{eq:haminteraction1} \hat{H}_U = -U\sum_{j} \hat{n}_{j,\uparrow} \hat{n}_{j,\downarrow}\;,\;\hat{n}_{j,\sigma}= \hat{a}^+_{j,\sigma} \hat{a}_{j,\sigma}\;,$$ $$\label{eq:haminteraction2} \hat{H}_V = -V\sum_{j} \hat{n}_j \hat{n}_{j+1}\;,\; \hat{n}_j = \hat{n}_{j,\uparrow}+ \hat{n}_{j,\downarrow}\;.$$ $\hat{H}_0$ describes the nearest-neighbor hopping of fermions along the lattice. Here the symbols $\sigma = \uparrow,\downarrow$ stand for a fermion with spin up or down. $\hat{H}_U$ describes the onsite interaction between the particles, and $\hat{H}_V$ the intersite interaction of fermions located at adjacent sites. $ \hat{a}^+_{j,\sigma}$ and $\hat{a}_{j,\sigma}$ are the fermionic creation and annihilation operators satisfying the corresponding anticommutation relations: $[\hat{a}^+_{j,\sigma},\hat{a}_{l,\sigma'}]=\delta_{j,l}\delta_{\sigma,\sigma'}$, $[\hat{a}^+_{j,\sigma},\hat{a}^+_{l,\sigma'}]=[\hat{a}_{j,\sigma},\hat{a}_{l,\sigma'}]=0$. Note that throughout this work we consider $U$ and $V$ positive, which leads to bound states located below the two-particle continuum. A change of the sign of $U,V$ will simply swap the energies. The Hamiltonian (\[eq:hamiltonian\]) commutes with the number operator $\hat{N}=\sum_{j}\hat{n}_j$ whose eigenvalues are $n=n_{\uparrow}+n_{\downarrow}$, i.e. the total number of fermions in the lattice. We consider $n=2$, with $n_{\uparrow}=1$ and $n_{\downarrow}=1$. Therefore we construct a basis starting with the eigenstates of $\hat{N}$. We use a number state basis $|\Phi_n\rangle=|n_1;n_2\cdots n_f\rangle$ [@EilbeckPhysicaD78], where $n_i=n_{i,\uparrow}+ n_{i,\downarrow}$ represents the number of fermions at the i-th site of the lattice. $|\Phi_n\rangle$ is an eigenstate of the number operator $\hat{N}$ with eigenvalue $n=\sum_{j=1}^f n_j$. To observe the fermionic character of the considered states, any two-particle number state is generated from the vacuum $|O\rangle$ by first creating a particle with spin down, and then a particle with spin up: e.g. $\hat{a}^+_{2,\uparrow} \hat{a}^+_{1,\downarrow} |O\rangle$ creates a particle with spin down on site 1 and one with spin up on site 2, while $\hat{a}^+_{2,\uparrow} \hat{a}^+_{2,\downarrow} |O\rangle$ creates both particles with spin down and up on site 2. Due to periodic boundary conditions the Hamiltonian (\[eq:hamiltonian\]) commutes also with the translation operator $\hat{T}$, which shifts all lattice indices by one. It has eigenvalues $\tau=exp(ik)$, with Bloch wave number $k=\frac{2\pi\nu}{f}$ and $\nu=0,1,2,...,f-1$. [*Single particle spectrum.*]{} For the case of having only one fermion (either spin up or spin down) in the lattice ($n=1$), a number state has the form $|j\rangle = \hat{a}^+_{j,\sigma} |O\rangle$. The interaction terms $\hat{H}_U$ and $\hat{H}_V$ do not contribute. For a given wave number $k$, the eigenstate to (\[eq:hamiltonian\]) is therefore given by: $$|\Psi_1\rangle=\frac{1}{\sqrt{f}} \sum_{s=1}^f\Big(\frac{\hat{T}}{\tau}\Big)^{s-1}|1\rangle\;.$$ The corresponding eigenenergy $$\label{eq:1fermionenergy} \varepsilon_k=-2\cos(k).$$ [*Two fermions with opposite spins.*]{} For two particles, the number state method involves $N_{s}= f^2$ basis states, which is the number of ways one can distribute two fermions with opposite spins over the $f$ sites including possible double occupamcy of a site. Below we consider only cases of odd $f$ for simplicity. Extension to even values of $f$ is straightforward. We define basis states to a given value of the wave number $k$: $$\begin{aligned} |\Phi_1\rangle = \frac{1}{\sqrt{f}} \sum_{s=1}^f\Big(\frac{\hat{T}}{\tau}\Big)^{s-1} \hat{a}^+_{1,\uparrow} \hat{a}^+_{1,\downarrow} |O\rangle\;, \\ % |\Phi_{j,+}\rangle = \frac{1}{\sqrt{f}} \sum_{s=1}^f\Big(\frac{\hat{T}}{\tau}\Big)^{s-1} \hat{a}^+_{j,\uparrow} \hat{a}^+_{1,\downarrow} |O\rangle\;,\\ % |\Phi_{j,-}\rangle = \frac{1}{\sqrt{f}} \sum_{s=1}^f\Big(\frac{\hat{T}}{\tau}\Big)^{s-1} \hat{a}^+_{1,\uparrow} \hat{a}^+_{j,\downarrow} |O\rangle\;, \\ \; 1 < j \leq (f+1)/2\;.\end{aligned}$$ Note, that the sign $\pm$ discriminates between states, where the distance from the spin up particle to the spin down particle is smaller (larger) than vice versa. Note that distances are measured by scanning the (periodic) chain in the direction of increasing lattice site number, passing a given particle (say with spin up) and then counting the distance to the spin down particle. In the limit of an infinite system the two signs discriminate between states where the spin up particle is to the left or right of the spin down one. Therefore, a complete wavefunction is given by $$\label{eq:basis1} |\Psi_2\rangle= c_1|\Phi_1\rangle+\sum_{j=2}^{\frac{f+1}{2}}c_{j+}|\Phi_{j,+}\rangle +\sum_{j=2}^{\frac{f+1}{2}}c_{j-}|\Phi_{j,-}\rangle\;.$$ Any vector in our given Hilbert space is then spanned by the numbers $|c_1,c_{2+},c_{2-},c_{3+},...\rangle$. Next we calculate the matrix elements of the Hamiltonian (\[eq:hamiltonian\]) in the framework of the basis (\[eq:basis1\]). We arrive at a $f \times f$ matrix with elements $H(i,j)$ ($i,j=1,\ldots,d$): $$\begin{aligned} \label{eq:matel2} H(i,j) = -\left( \begin{array}{cccccccc} U & q &q & 0 & \ldots & ~& ~& \\ q^* & V&0& q & 0 & \ldots & & \\ q^* &0&V&0&q & 0 & \ldots & \\ 0 &q^* & 0 & 0 & 0 & q & 0 & \\ 0 &0 &\ddots&\ddots &\ddots &\ddots &\ddots & 0 \\ &\ldots&0&q^*&0&0&0&q\\ & & \ldots & 0 & q^* & 0 & 0 & p \\ & & &\ldots & 0 &q^*& p & 0 \end{array} \right)\;.\end{aligned}$$ Here $q=1+\tau$ and $p=\tau^{-(f+1)/2} + \tau^{-(f-1)/2}$. ![\[spectrum\]Energy spectrum of the two fermion states. The eigenvalues are plotted as a function of the wave number $k$. Here $U=1\;,\;V=1,\;f=101$. Symbols are from analytical derivation, lines are the result of numerical diagonalization. The arrows indicate the location of the critical wave numbers (see text). ](fig1.eps){width="0.95\columnwidth"} In Fig. \[spectrum\] we show the energy spectrum of the Hamiltonian matrix (\[eq:matel2\]) obtained by numerical diagonalization for the interaction parameters $U=2$ and $V=2$ and $f=101$. At $U=0$ and $V=0$, the spectrum is given by the two fermion continuum, whose eigenstates are characterized by the two fermions independently moving along the lattice. In this case the eigenenergies are the sum of the two single-particle energies: $$\label{eq:energyunperturbed} E_{k_1,k_2}^0=-2[\cos(k_1)+\cos(k_2)],$$ with $k_{1,2}={\pi}\nu_{1,2}/(f+1)$ , $\nu_{1,2}= 1,\ldots,f$. The Bloch wave number $k = k_1+k_2 \mod 2\pi$. Therefore, if $k=\pm \pi$, the continuum degenerates into points. The continuum is bounded by the hull curves $h_{\pm}(k)=\pm 4 \cos \frac{k}{2}$. The same two-particle continuum is still observed in Figure.\[spectrum\] for nonzero interaction. However, in addition to the continuum, we observe one, two or three bound states dropping out of the continuum, which depends on the wave number. For any nonzero $U$ and $V$, all three bound states drop out of the continuum at $k=\pm \pi$. One of them stays bounded for all values of $k$. The two other ones merge with the continuum at some critical value of $|k|$ upon approaching $k=0$ as observed in Fig.\[spectrum\]. Note that for $k=\pm \pi$ and $U=V$, all three bound states are degenerate. Upon increasing $U$ and $V$, we observe that a second bound state band separates from the continuum for all $k$ (Fig.\[pbcspectrum\]). At the same time, when $U \neq V$, the degeneracy at $k=\pm \pi$ is reduced to two. ![\[pbcspectrum\]Energy spectrum of the two fermion states. The eigenvalues are plotted as a function of the wave number k. Here $U=4\;,\;V=3\;,\;f=101$. The symbols are from analytical derivation, lines are the results of numerical diagonalization. The arrow indicates the location of the critical wave number (see text). ](fig2.eps){width="0.95\columnwidth"} Finally, for even larger values of $U$ and $V$, all three bound state bands completely separate from the continuum (Fig.\[symspect\]). [*Symmetric and antisymmetric state representation.*]{} In order to obtain analytical estimates on the properties of the observed bound states, we use the fact that the Hamiltonian for a two fermion state is invariant under flipping the spins of both particles. We define symmetric basis states $$\label{eq:basisym} |\Phi_{j,s}\rangle=\frac{1}{\sqrt{2}}(|\Phi_{j,+}\rangle + |\Phi_{j,-}\rangle)$$ and antisymmetric states $$\label{eq:basisas} |\Phi_{j,a}\rangle=\frac{1}{\sqrt{2}}(|\Phi_{j,+}\rangle - |\Phi_{j,-}\rangle)\;.$$ Note that $|\Phi_1\rangle$ is a symmetric state as well. Then the matrix (\[eq:matel2\]) can be decomposed into irreducible symmetric and antisymmetric ones: $$\begin{aligned} \label{eq:matsym} H^s(i,j) = -\left( \begin{array}{cccccc} U & q\sqrt{2} & & & & \\ q^*\sqrt{2} &V & q & & & \\ & q^* & 0 & q & & \\ & & \ddots & \ddots &\ddots & \\ & & & q^* & 0 & q \\ & & & & q^* & p \end{array} \right)\;,\end{aligned}$$ $$\begin{aligned} \label{eq:matasym} H^a(i,j) = -\left( \begin{array}{cccccc} V & q& & & & \\ q^* & 0 & q & & & \\ & q^* & 0 & q & & \\ & & \ddots & \ddots &\ddots & \\ & & & q^* & 0 & q \\ & & & & q^* &- p \end{array} \right)\;.\end{aligned}$$ The rank of the matrix $H^s$ is $(f+1)/2$, while the rank of $H^a$ is $(f-1)/2$. ![\[symspect\] Energy spectrum for the two fermion states. The eigenvalues are plotted as a function of the wave number k. Here $U=8\;,\;V=8\;,\;f=101$. The symbols are from analytical derivation, lines are the results of numerical diagonalization. ](fig3.eps){width="0.95\columnwidth"} [*Antisymmetric bound states.*]{} The antisymmetric states exclude double occupation. Therefore the spectrum is identical with the one of two spinless fermions [@EilbeckPhysicaD78]. Following the derivations in [@EilbeckPhysicaD78] we find that the antisymmetric bound state, if it exists, has an energy $$\label{eanti} E^a_2(k)=-(V+\frac{4}{V}\cos^2(\frac{k}{2}))\;.$$ This result is valid as long as the the bound state energy stays outside of the continuum. The critical value of $k$ at which validity is lost, is obtained by requesting $|E^a_2(k)|=|h_{\pm}(k)|$. It follows $ V=2\cos(\frac{k}{2})$. Therefore the antisymmetric bound state merges with the continuum at a critical wave number $$\label{kac} k^a_{c}=2\arccos(\frac{V}{2})\;,$$ setting a critical length scale $\lambda^a_{c}=\frac{2\pi}{ k^a_{c}}$. For $V=1$ it follows $k^a_{c}/\pi\approx 0.667$ (see Fig.\[spectrum\]). The equation (\[eanti\]) is in excellent agreement with the numerical data in Figs. \[spectrum\],\[pbcspectrum\],\[symspect\] (cf. open triangles). We also note, that the antisymmetric bound state is located between the two symmetric bound states, which we discuss next. [*Symmetric bound states.*]{} A bound state can be searched for by assuming an unnormalized eigenvector to (\[eq:matsym\]) of the form $|c,1,\mu,\mu^2,\mu^3,...\rangle$ with $|\mu| \equiv \rho \leq 1$. We obtain $$\begin{aligned} E c = U c + \sqrt{2} q^*\;, \nonumber \\ E = \sqrt{2} q c + V + q^* \mu \;, \\ E = \frac{1}{\mu} + q^* \mu \;. \nonumber\end{aligned}$$ It follows that $\mu = \rho \exp(ik/2)$ and $${\label{eq:ensym}} E^s_2(k) = -2({\rho+\frac{1}{\rho})\cos{k/2}}\;.$$ The parameter $\rho$ satisfies a cubic equation $$\label{eq:cubiceq} a{\rho}^3+b{\rho}^2+c{\rho}+d=0$$ with the real coefficients a, b. c and d given by $a=2Vcos(\frac{k}{2})$, $b= 4cos^2(\frac{k}{2})-UV$, $c=2(U+V)cos(\frac{k}{2})$, $d=-4cos^2({\frac{k}{2}})$. An analytic solution to (\[eq:cubiceq\]) can be obtained, but is cumbersome to be presented here. We plot the results in Figs. \[spectrum\],\[pbcspectrum\],\[symspect\] (cf. open circles and squares). We obtain excellent agreement. At the Brilloin zone edge $k=\pm \pi$ the cubic equation (\[eq:cubiceq\]) is reduced to a quadratic one, and can be solved to obtain finally $\rho \rightarrow 0$ and $${\label{eq:enrhosmal}} E^{s1}_{2}(k\to\pm\pi)=-U\;,\;E^{s2}_{2}(k\to\pm\pi) =-V \;.$$ In particular we find for $k=\pm \pi$ that $E^{s2}_{2} = E^a_2$. In addition, if $U=V$, all three bound states degenerate at the zone edge. If $V=0$, the cubic equation (\[eq:cubiceq\]) is reduced to a quadratic one in the whole range of $k$ and yields [@EilbeckPhysicaD78] $$E^{s1}_{2}(k) = -\sqrt{U^2 + 16 \cos^2 ( k/2)}\;.$$ Next we determine the critical value of $k$ for which the bound state with energy $E^{s2}_{2}$ is joining the continuum. Since at this point $\rho=1$, we solve (\[eq:cubiceq\]) with respect to $k_c$ and find $$\label{ks2c} k^{s2}_{c}=2\arccos\Big( \frac{UV}{2(U+2V)}\Big)$$ setting another critical length scale $\lambda^s_{c}=\frac{2\pi}{ k^s_{c}}$. E.g. for $U=V=1$ $k^{s2}_{c}/\pi \approx 0.89$, in excellent agreement with Fig.\[spectrum\]. For $U=4$ and $V=3$ we find $k^{s2}_{c}/\pi \approx 0.59$ confirming numerical results in Fig.\[pbcspectrum\]. [*Conclusions.*]{} Two fermionic particles with opposite spin allow for three different types of bound states on a one-dimensional lattice with onsite $U$ and nearest neighbour $V$ interaction. Two of them are symmetric with respect to spin flips, and one is antisymmetric. The antisymmetric bound state is characterized by a critical wave number separates wave numbers with bound states from wave numbers without. It follows from (\[kac\]) that this happens for $V < 2$. For larger values of $V$ the whole wave number space becomes available for antisymmetric bound states, similar to one of the symmetric bound states for any nonzero $U$. The second symmetric bound state also observes a critical wave number. It follows from (\[ks2c\]) that this happens for $U < 4V/(V-2)$, while the whole wave number space becomes available otherwise. It could be a challenging task to observe these different phases with one, two, or three bound states experimentally, by tuning $U$, $V$, and $k$. For higher lattice dimensions more nearest neighbours have to be taken into account, similar to an increase of the interaction range. In these cases, we expect consequently more bound states to appear.\ \ Acknowledgements\ We thank M. Haque, D, Krimer, A. Ponno and Ch. Skokos for useful discussions. J.-P. Nguenang acknowledges the warm hospitality of the Max Planck Institute for the Physics of Complex Systems in Dresden. [99]{} K. Winkler, G. Thalhammer, F. Lang, R. Grimm, J. Ecker Denshlag, A. J. Daley, A. Kantian, H. P. Büchler, and P. Zoller, Nature [**441**]{}, 853 (2006). A. A. Ovchinnikov, Zh. Eksp. Teor. Fiz./Soviet Phys. JETP [**57/30**]{}, 263/147 (1969/1970). A.C Scott, J.C. Eilbeck and H.Gilhøj, Physica D [**78**]{}, 194 (1994). S. Flach and A. V. Gorbach, Phys. Rep. [**467**]{}, 1 (2008). M. A. Baranov, Phys. Rep. [**464**]{}, 71 (2008).
{ "pile_set_name": "ArXiv" }
--- abstract: | Compactification of Type IIB superstring on an $AdS_5 \times S^5/\Gamma$ background leads to SU(N) gauge field theories with prescribed matter representations. In the ’t Hooft limit of large N such theories are conformally finite. For finite N and broken supersymmetry ($\cal N$ = 0) I derive the constraints to be two-loop conformal and examine the consequences for a wide choice of $\Gamma$ and its embedding $\Gamma \subset {\cal C}^3 (\supset S^5)$. address: - 'Department of Physics and Astronomy,' - 'University of North Carolina, Chapel Hill, NC 27599-3255' author: - '**Paul H. Frampton**' title: '**ADS/CFT String Duality and Conformal Gauge Theories.**' --- [*Introduction.*]{} Recently the relationship of string theory to gauge theory received stimulus from the conjecture by Maldacena[@mald] (related earlier papers are [@K; @GKT; @GK; @P]) stemming from string duality which makes in its strongest form the assertion that the information contained in superstring theory is encoded in a four-dimensional gauge field theory including its non-perturbative sector. This has been vigorously pursued by many authors, especially Witten[@W1; @W2; @W3]. A brief review is in [@schwarz]. This relationship appears ironic when one recalls that the earliest string theories, the dual resonance models for strong interactions, were abandoned in favor of an $SU(3)$ gauge theory 25 years ago. String theory has generally been regarded as much more general than gauge field theory because of its far richer structure; however, that perception was based on perturbative arguments, and the new developments of Maldacena [*et al.*]{} are essentially non-perturbative. The idea is to consider N coincident D3-branes with 4-dimensional world volume theories having superconformal symmetry. This is conjectured [@mald] to be dual (weak coupling related to strong coupling) to type IIB superstring theory in a spacetime with geometry $AdS_5 \times S^5$. The world volume theory is in this case an ${\cal N}= 4$ supersymmetric Yang-Mills theory with gauge group $SU(N)$. Originally it is $U(N)$ but this is broken to $SU(N)$. The radii of the $AdS_5$ and $S^5$ are equal and both given by $R = \lambda^{1/4} l_s$ where $\lambda$ is the\ ’t Hooft parameter[@tH] $\lambda = g_{YM}^2 N$ ($g_{YM}^2 = g_S$, the string coupling constant) and $l_S^2 = \alpha^{'}$ the universal Regge slope. The string tension is $T = (2 \pi \alpha^{'})^{-1}$. The ${\cal N} = 4$ $ SU(N)$ gauge theory has been known to be ultra-violet finite for many years[@mand]. This is true not only for $N \rightarrow \infty$, the conformal limit of Maldacena, but also for finite $N$. [*Breaking supersymetries.*]{} By factoring out a discrete group $\Gamma$ in $S^5/\Gamma$ it is possible to break some or all of the ${\cal N} = 4$ supersymmetries. The isometry of $S^5$ is $SO(6) \sim SU(4)$ which may be identified with the R-parity of the ${\cal N} = 4$ conformal gauge theory. The spinors are in the [**4**]{} and the scalars are in the [**6**]{} of this $SU(4)$. I shall here consider only abelian groups $\Gamma = Z_p$, although non-abelian $\Gamma$ are worth further study (see [*e.g.*]{} [@HH; @GLR]). I am considering only $AdS_5 \times S^5/\Gamma$, although the second 5-dimensional orbifold can be more general [*e.g.*]{} the $T^{p,q}$ spaces considered in [@KW]. The number of unbroken symmetries has been studied in [*e.g.*]{} [@moore; @morrison] with the result that if $\Gamma \subset SU(2)$ there remains ${\cal N} = 2$ supersymmetry; if that is not satisfied but $\Gamma \subset SU(3)$ there remains ${\cal N} = 1$ supersymmetry; finally if even that is not satisfied one is left with ${\cal N} = 0$ or no supersymmetry. This last case is of most interest here. It has been demonstrated that the large N limit of the resultant gauge theory coincides with that of the ${\cal N} = 4$ case. Such arguments have been made both using string theory [@vafa] and directly at the field theory level [@bershadsky]. In the latter case the proof involves a monodromy of the representation for the group $\Gamma$. For finite N, however, there is no argument that the resultant gauge theory is conformal, especially for ${\cal N} = 0$ where there are no non-renormalization theorems. Nevertheless, if there does exist a conformal gauge theory in four dimensions with ${\cal N} = 0$, it would be so tightly constrained as to be possibly unique and would be of interest especially if it could contain the standard $SU(3) \times SU(2) \times U(1)$ model with its peculiar representations for the quarks and leptons. The representations which occur in the resultant ${\cal N} \leq 2$ gauge theories from the orbifold construction have been studied using quiver diagrams[@moore]. I will find that these diagrams, while convenient for the cases ${\cal N} \geq 1$ need augmentation for the case ${\cal N} = 0$. To specify the potentially conformal gauge theory I need to state how the group $\Gamma$ is embedded in ${\cal C}^3$. Let the three complex coordinates of ${\cal C}^3$ be denoted by $\underline{X} = (X_1, X_2, X_3)$. The action of $Z_p$ is the specified by: $$\underline{X} \rightarrow (\alpha^{a_1} X_1, \alpha^{a_2} X_2, \alpha^{a_3} X_3) \label{Zp}$$ where $\alpha = exp(2 \pi i/ p)$ and the three integers $a_{\mu} = (a_1, a_2, a_3)$ specify the embedding. In order to ensure an ${\cal N} = 0$ result, I must insist that $\Gamma$ is not contained in $SU(3)$ by the requirement that $$a_1 + a_2 + a_3 \neq 0~~~ (mod~~~ p) \label{su4}$$ At the same time, for the correct behavior of the spinors we need in addition $$a_1 + a_2 + a_3 = 0~~~ (mod~~~ 2) \label{mod2}$$ For any given p, there is a finite $\nu(p)$ number of choices satisfying Eq.(\[su4\]) and Eq.(\[mod2\]). We shall indicate later how to enumerate these $\nu(p)$. [*Matter representations.*]{} Because the discrete group $Z_p$ leads to the identification of $p$ points in ${\cal C}^3$ and the N coinciding D3-branes converge on all $p$ copies, the gauge group becomes $SU(N)^p$. The surviving states are invariant under the product of a gauge transformation and a $Z_p$ transformation defined as in Eq.(\[Zp\]) above. For the scalars, it then follows that the scalars fall into the representations $$\sum_{\mu} (N_i, \bar{N}_{i \pm a_{\mu}}) \label{scalars}$$ For $a_{\mu} \neq 0$ these are bi-fundamentals and for $a_{\mu} = 0$ complex adjoints. If we focus on one $SU(N)$ the only non-singlet representations (the same will be true for the fermions) are fundamentals, anti-fundamentals and adjoints. These representations also follow from the Douglas-Moore quiver diagram. For the fermions we must consider the transformation of a 4-spinor by making four combinations $A_{\lambda} (1 \leq \lambda \leq 4)$ of the $a_{\mu}$ $$\begin{aligned} A_1 & = & (a_1 + a_2 + a_3) / 2 \\ A_2 & = & (a_1 - a_2 - a_3) / 2 \\ A_3 & = & (-a_1 + a_2 - a_3) / 2 \\ A_4 & = & (-a_1 - a_2 + a_3) / 2 \\\end{aligned}$$ Again the surviving states are invariant under a product of the $Z_p$ and gauge transformations. This leads to the fermion representation: $$\sum_{\lambda} (N_i, \bar{N}_{i+A_{\lambda}}) \label{fermions}$$ which can, if required, be deduced from a (different) quiver diagram. [*Two-loop $\beta$-functions.*]{} I may take the detailed formula for the gauge coupling $\beta$-function $\beta_g$ from [@MV]. The two leading orders are: $$\beta_g = \beta_g^{(1)} + \beta_g^{(2)}$$ with $$\beta_g^{(1)} = - \frac{g^3}{(4 \pi)^2} \left[ \frac{11}{3} C_2(G) - \frac{4}{3} \kappa S_2 (F) - \frac{1}{6} S_2 (S) \right] \label{beta1}$$ and $$\beta_g^{(2)} = - \frac{g^5}{(4 \pi)^4} \left[ \frac{34}{3} (C_2(G))^2 - \kappa \left[ 4C_2(F) + \frac{20}{3} C_2(G) \right] S_2(F) -\left[ 2 C_2(S) + \frac{1}{3} C_2(G) \right] S_2(S) + \frac {2 \kappa Y_4(F)}{g^2} \right] \label{beta2}$$ Here $C_2, S_2$ are the quadratic Casimir, Dynkin index respectively for the representations indicated, $\kappa$ is 1/2, 1 for Weyl, Dirac fermions respectively, products like $C_2(R)S_2(R)$ imply a sum over irreducible representations and finally the Yukawa term is included naturally in the two-loop term (unlike in [@MV]) because here the Yukawa couplings are proportional to the gauge coupling. The crucial quantity $Y_4(F)$ is defined in terms of the Yukawa matrix $Y_{ij}^a\psi_i\zeta\psi_j\phi^a$ by $$Y_4(F) = Tr \left( C_2(F) Y^a Y^{\dagger a} \right) \label{Y4}$$ Looking first at ${\cal N} = 4$, the values are easily seen to $C_2(G)=N, S_2(F)=4N, S_2(S)=6N$ while $C_2(F)S_2(F)=4N^2$ and $C_2(S)S_2(S)=6N^2$. Finally $Y_4(F) = 24g^2N^2$. It follows from Eq.(\[beta1\]) and Eq.(\[beta2\]) the $\beta_g = 0$ for ${\cal N} = 4$ at two loops, as is well known[@mand]. However, the situation for ${\cal N} = 0$ is much more complicated. At one-loop level for ${\cal N} = 0$ the evaluation of $\beta_g^{(1)}$ is the same term-by-term as for ${\cal N} = 4$. This is already in [@KS1; @KKS; @KS2] for the one-loop level and since the one-loop $\beta$-function is purely leading-order in $N$ it conforms to the general arguments of [@vafa; @bershadsky]. At two-loop order I must examine the non-leading terms in $1/N$ in Eq.(\[beta2\]). The first, third and fifth terms are always the same for ${\cal N} = 0$ as for ${\cal N} = 4$, respectively $34N^2/3 - 40N^2/3 - 2N^2 = -4N^2$. To evaluate the second, fourth and sixth terms I find it necessary to distinguish four cases which are designated ($\alpha,\beta,\gamma,\delta$) as follows: $$a_1 = a_2 ; ~~~a_3 = 0. ~~~A_1 = - A_4 \neq 0; ~~~A_2 = A_3 = 0. ~~~~(Case~~~ \alpha).$$ $$a_1 \neq a_2; ~~~a_3 = 0. ~~~A_1 = - A_4 \neq 0; ~~~A_2 = - A_3 \neq 0. ~~~~(Case ~~~\beta).$$ $$All ~~~~a_{\mu} \neq 0. ~~~~One ~~A_{\lambda} = 0; ~~~~three ~~A_{\lambda} \neq 0. ~~~~(Case ~~~\gamma).$$ $$All ~~~~a_{\mu} \neq 0. ~~~~All ~~A_{\lambda} = 0. ~~~~(Case ~~~\delta).$$ These possibilities lead to fermion and scalar representations of $SU(N)$ which are different for the four cases. They exhaust the choices which leave ${\cal N} = 0$ which requires that Eqs.(\[su4\]),(\[mod2\]) are fulfilled. (Note that at least two $a_{\mu}$ must be non-zero). The evaluation of the remaining terms in Eq.(\[beta2\]) can now be done case by case. In Case $\alpha$, where both fermions and scalar appear in both fundamentals and adjoints, we find that $C_2(F)S_2(F) = 4N^2(1 - 1/(2N^2))$, $C_2(S)S_2(S) = 6N^2(1 - 2/(3N^2))$ and $Y_4(F) = (24N^2 - 16)g^2$. Substituting in Eq.(\[beta2\]) leads, as generally expected to an non-vanishing $\beta_g$ for finite N and a non-conformal gauge theory. For the other cases, I find for the three group theory expressions $C_2(F)S_2(F)$, $C_2(S)S_2(S)$ and $Y_4(F)$ respectively the following: $$4N^2(1 - 1/(N^2)), ~~~6N^2(1 - 2/(3N^2)) ~~~and~~~(24N^2 - 24). ~~~(Case~~~\beta)$$ $$4N^2(1 - 3/(4N^2)), ~~~6N^2(1 - 1/(N^2)) ~~~and~~~(24N^2 - 18). ~~~(Case~~~\gamma)$$ $$4N^2(1 - 1/(N^2)), ~~~6N^2(1 - 1/(N^2)) ~~~and~~~(24N^2 - 24). ~~~(Case~~~\delta)$$ Substituting in Eq. (\[beta2\]), I find that $\beta_g^{(2)}$ is non-vanishing except in the Case $\gamma$. For this surviving theory, the fermions are in both fundamentals and adjoints, while all scalars are in fundamentals. This is therefore the only combination of matter representations of further interest. [*Directions.*]{} A subsequent question to be addressed is what happens at three-loop and even higher orders. Also one must consider running of the Yukawa and quartic Higgs self-couplings due to possible non-vanishing of their $\beta$-functions $\beta_Y$ and $\beta_H$. It is planned to publish a more complete analysis elsewhere; I conclude this proposal with comments and possible future directions. Often low-energy supersymmetry is adopted in order to solve the hierarchy problem of the Planck or GUT scale to the weak scale. This hierarchy is theory-generated and one may instead be agnostic about physics at $\gtrsim 1000$TeV scale where there is no real information. For example, recent ideas about extra Kaluza-Klein dimensions at reduced scales [*e.g.*]{} [@DDG1; @DDG2; @DDG3; @ADD] avoid the hierarchy altogether and hence remove the main motivation for low-energy supersymmetry. The possible role of an ${\cal N} = 0$, $d=4$ conformal gauge theory may be put in context by imagining the level of skepticism to infinite renormalization of QED in 1948 (and later of the standard model) if the example of [@mand] had been found four decades earlier. The exciting possibility is that the standard model is part of such an ${\cal N} = 0$ conformal gauge theory. The mass scales $\Lambda_{QCD}$ and $M_W$ would arise from necessarily non-perturbative effects, and gravity would be accommodated through the holographic principle[@tH2; @W1]. Using AdS/CFT duality could help identify the relevant conformal theory. If so, this could shed light on the outstanding questions (families, CP violation, etc.) posed by the standard model. I wish to acknowledge useful discussions with J. Lykken, D.R. Morrison and S. Trivedi. This work was supported in part by the US Department of Energy under Grant No. DE-FG05-85ER-40219. [99]{} J. Maldacena, Adv. Theor. Math. Phys. [**2,**]{} 231 (1998). [*hep-th/9711200.*]{} I.R. Klebanov, Nucl. Phys. [**B496,**]{} 231 (1997). [*hep-th/9702076.*]{} S.S. Gubser, I.R. Klebanov and A.A. Tsyetlin, Nucl. Phys. [**B499,**]{} 217 (1997). [*hep-th/9703040.*]{} S.S. Gubser and I.R. Klebanov, Phys. Lett. [**B413,**]{} 41 (1997). [*hep-th/9708005.*]{} A.M. Polyakov, Nucl. Phys. Suppl. [**68,**]{} 1 (1998). [*hep-th/9711002.*]{} E. Witten, Adv. Theor. Math. Phys. [**2,**]{} 253 (1998). [*hep-th/9802150.*]{} E. Witten, Adv. Theor. Math. Phys. [**2,**]{} 505 (1998). [*hep-th/9803131.*]{} E. Witten, J. High Energy Phys. 9807:006 (1998). J.H. Schwarz, CALT-68-2204 (1998). [*hep-th/9812037.*]{} G. ’t Hooft, Nucl. Phys. [**B72,**]{} 461 (1974). S. Mandelstam, Nucl. Phys. [**B213,**]{} 149 (1983). A. Hanary and Y.-H. He. [*hep-th/9811183.*]{} B.R. Greene, C.I. Lazariou and M. Raugas. [*hep-th/9811201.*]{} I.R. Klebanov and E. Witten. [*hep-th/9807080.*]{} M.R. Douglas and G. Moore. [*hep-th/9603167.*]{} M.R. Douglas, B.R. Greene and D.R. Morrison, Nucl. Phys. [**B506,**]{} 84 (1997). M. Bershadsky, Z. Kakushadze and C. Vafa, Nucl. Phys. [**B523,**]{} 59 (1998). [*hep-th/9803076.*]{} M. Bershadsky and A. Johansen. [*hep-th/9803249.*]{} M.E. Machacek and M.T. Vaughn, Nucl. Phys. [**B222,**]{} 83 (1983). S. Kachru and E. Silverstein, Phys. Rev. Lett. [**80,**]{} 4855 (1998). [*hep-th/9802183.*]{} S. Kachru, J. Kumar and E. Silverstein. [*hep-th/9807076.*]{} S. Kachru and E. Silverstein, J. High Energy Phys. 9811:001 (1998). [*hep-th/9808056.*]{} K.R. Dienes, E. Dudas and T. Gherghetta, Phys. Lett. [**436,**]{} 55 (1998). [*hep-ph/9803466.*]{} K.R. Dienes, E. Dudas and T. Gherghetta. [*hep-ph/9806292.*]{} K.R. Dienes, E. Dudas and T. Gherghetta. [*hep-ph/9807522.*]{} N. Arkani-Hamed, S. Dimopoulos and G. Dvali, Phys. Lett. [**B429,**]{} 263 (1998). [*hep-ph/9803315.*]{} G. ’t Hooft, in Salamfestschrift. Editors: A. Ali, J. Ellis and S. Ranjbar-Daemi.\ World Scientific Publishing Company (1994) page 284. [*gr-qc/9310026.*]{}
{ "pile_set_name": "ArXiv" }
--- abstract: 'By ascribing a complex anticommuting variable $\zeta$ to each basic [ *property*]{} of a field it is possible to describe all the fundamental particles as combinations of only five $\zeta$ and understand the occurrence of particle generations. An extension of space-time $x$ to include property then specifies the ‘where, when and what’ of an event and allows for a generalized relativity where the gauge fields lie in the $x-\zeta$ sector and the Higgs fields in the $\zeta-\zeta$ sector.' --- [*R. Delbourgo*]{} [School of Mathematics and Physics, University of Tasmania\ Private Bag 37 GPO, Hobart, Australia 7001 ]{} Preamble {#preamble .unnumbered} ======== First of all let me say why I am so pleased to be present at this FestSchrift in honour of Girish’ and Bruce’s retirements. When I arrived in Tasmania one of my first tasks was to make contact with physicists in other Australian institutions. Amongst these Melbourne University had high priority and I was very glad to welcome Bruce as one of the earliest visitors to Hobart. Since then we have interacted many times and I have relished my own visits to Melbourne to give occasional seminars and find out what Joshi and McKellar were up to. Let me wish them both a long and happy retirement and say how much I have appreciated their support and friendship throughout the last 30 years! Retirement has much to recommend it and they might care to view their future condition as changing from ‘battery hen’ to ‘free-range chicken’, since they are no longer obliged to feed on grants in order to lay eggs. Introduction ============ As most of you will have already surmised, the title of my talk has nothing to do with real estate and perhaps sounds all the more mysterious for it. In fact property values have everything do with quantum fields and accurately reflect the contents of what follows. The motivation for this work is to be able to describe the ‘what’ as well as the ‘where-when’ of an event. To make an analogy with personality traits in humans, psychologists may characterise a person as optimistic/pessimistic, happy/sad, aggressive/submissive, etc. (A person will possess some combination or superposition over these trait states.) The same sort of characterisation applies to a quantum field and is usually underlined by attaching a label to each distinct field. Thus we speak of an electron/positron field, a neutron/antineutron field, a quark/antiquark field, etc. and draw them in Feynman diagrams say with solid or dashed lines and arrows, plus legends if necessary, to distinguish them from one another. An event is some confluence between these labels with a possible interchange of traits, as specified by an interaction Lagrangian. Often we assume a symmetry group is operational which ‘rotates’ labels; this constrains the interchange of property and, if the symmetry is local, it can be gauged. The basic idea I wish to put to you is that traits/labels are normally discrete: thus a field is either an electron or it is not; similarly for a proton, and so on. It makes sense to attach a separate coordinate to each such property and choose the coordinate to be anticommuting since its occupation number is either one or zero. (So a person can be pessimistic & sad & aggressive with a product of the three traits.) Furthermore if we make the coordinate $\zeta$ complex we can describe the converse property by simple conjugation. When applying this idea to quantum fields the question arises: how many coordinates are needed? The fewer the better of course. From my investigations thus far [@RD1; @RD2] I have concluded that one can get a reasonable description of the known fundamental particle spectrum by using just five complex cordinates $\zeta^\mu; \, \mu = 0,1,2,3,4$. I have found [@RD1] that four $\zeta$ are insufficient to account for the three known light generations and their features. The way generations arise in this context comes by multiplying traits by neutral products of other traits. For instance a person can be pessimistic or (pessimistic $\times$ sad-happy) or (pessimistic $\times$ aggressive-submissive) and so on. With $N$ $\,\,\zeta$ one potentially encounters $2^{N-1}$ generations of a particular property by forming such neutral products; this is interesting because it suggests that any anticommuting property scheme produces an even number of generations. This is not a cause for panic; all we know at present is that there exist 3 light neutrinos, so there may be other (possibly sterile) heavy neutrinos accompanied by other quarks and charged leptons. The proposal therefore is to append a set of five complex anticommuting coordinates $\zeta^\mu$ to space-time $x^m$ which are to be associated with property. This is in contrast to traditional brane-string schemes which append yet unseen bosonic extra degrees of freedom to space-time. Now one of the most interesting aspects of fermionic degrees of freedom is that they act oppositely to bosonic ones in several respects: we are familiar with sign changes in commutation relations, statistical formulae and, most importantly, quantum loop contributions; but less well-known is that in certain group theoretical representations [@Neg] the $SO(2N)$ Casimirs are continuations to negative $N$ of $Sp(2N)$ Casimirs so that anticommuting coordinates [*effectively subtract*]{} dimensions. The opposite (to bosons) loop sign is extra confirmation of this statement and this is put to great use in standard supersymmetry (SUSY) in order to resolve the fine-tuning problem. So it is not inconceivable that with a correct set of anticommuting property coordinates $\zeta$ appended to $x$ we might end up with zero net dimensions as it was presumably before the BIG BANG; at the very least that we may arrive at a scheme where fermions cancel quantum effects from bosons as in SUSY. For this it is not necessary to embrace all the tenets of standard SUSY; indeed we shall contemplate an edifice where property coordinates are Lorentz scalar. In that respect they are similar to the variables occurring in BRST transformations for quantized vector gauge-fixed models — without implying any violation of the spin-statistics theorem for normal physical fields. Property superfields ==================== Now let us get down to the nuts and bolts of the construction and find out what the edifice will look like. We might be tempted to match the four space-time $x^m$ by four $\zeta^\mu$, but as we have shown elsewhere that is not enough – lepton generations do not ensue. However, all is not lost: we can add 5 $\zeta^\mu$ (which are sufficient), but only take half the states to get correct statistics. Associate each Lorentz scalar anticommuting numbers with a ‘property’ or ‘trait’; this invites us to consider symmetry groups[@Groups] like $SU(5)$ or $Sp(10)$ or $SO(10)$ to reshuffle the properties. (That these popular groups pop up is probably no accident.) The only way I have found to obtain the well-established quantum numbers of the fundamental particle spectrum as superpositions over traits is to postulate the following charge $Q$ and fermion number $F$ assignments, $$Q(\zeta^{0,1,2,3,4}) = (0,1/3,1/3,1/3,-1);\quad F(\zeta^{0,1,2,3,4}) = (1,-1/3,-1/3,-1/3,1).$$ Other properties are to built up as composites of these. For example $\zeta^4$ may be identified as a negatively charged lepton, but then so can its product with neutral combinations like $\bar{\zeta}_0\zeta^0$, $\bar{\zeta}_i\zeta^i$,... strongly suggesting how generations can arise in this framework. We’ll return to this shortly. Since the product of an even number of $\zeta$ is a (nilpotent) commuting property a Bose superfield $\Phi$ should be a Taylor series in even powers of $\zeta,\bar{\zeta}$. Similarly a superfield which encompasses fermions $\Psi_\alpha$ will be a series in odd powers of $\zeta,\bar{\zeta}$ — up to the 5th: $$\Phi(x,\zeta,\bar{\zeta})=\sum_{even\,r+\bar{r}} (\bar{\zeta})^{\bar{r}}\phi_{(\bar{r}),(r)}(\zeta)^r;$$ $$\Psi_\alpha(x,\zeta,\bar{\zeta})= \sum_{odd\, r+\bar{r}} (\bar{\zeta})^{\bar{r}}\psi_{\alpha(\bar{r}),(r)}(\zeta)^r.$$ So far as labels $\mu$ on $\zeta^\mu$ go, we can characterise - label 0 as [neutrinicity]{} - labels [1 - 3]{} as (down) [chromicity]{} - label 4 as [charged leptonicity]{} You will notice that these expansions produce too many states for $\psi_\alpha$ and $\phi$, viz. 512 properties in all, so they demand cutting down. An obvious tactic is to associate conjugation ($^c$) with the operation $\zeta \leftrightarrow \bar{\zeta}$ $$\psi_{(r),(\bar{r})} = \psi^{(c)}_{(\bar{r}),(r)},$$ corresponding to reflection along the main diagonal when we expand $\Psi$ as per the table below. Indeed if we suppose that all fermion field components are left-handed the conjugation/reflection operation then automatically includes right-handed particle states. Even so there remain too many components and we may wish to prune more. One strategy is to notice that under reflection about the [*cross-diagonal*]{} the $F$ and $Q$ quantum numbers are not altered. We shall call this cross-diagonal reflection a [*duality*]{} ($^\times$) transformation. For example, $$(\bar{\zeta}_\alpha\zeta^\mu\zeta^\nu)^{\times} = \frac{1}{3!} \epsilon^{\rho\sigma\tau\mu\nu}\bar{\zeta}_\rho\bar{\zeta}_\sigma \bar{\zeta}_\tau.\frac{1}{4!}\epsilon_{\alpha\beta\gamma\delta\epsilon} \zeta^\beta\zeta^\gamma\zeta^\delta\zeta^\epsilon.$$ By imposing the antidual reflection symmetry: $\psi_{(r),(\bar{r})} = - \psi_{(5-\bar{r}),(5-r)}$ we roughly halve the remaining number of components. So whereas previously we had the separate set of neutrino states, for instance, $$\zeta^0,\zeta^0(\bar{\zeta}_4\zeta^4),\zeta^0(\bar{\zeta}_i\zeta^i), \zeta^0(\bar{\zeta}_4\zeta^4)(\bar{\zeta}_i\zeta^i), \zeta^0(\bar{\zeta}_i\zeta^i)^2,\,(\bar{\zeta}_i\zeta^i)^3, \zeta^0(\bar{\zeta}_4\zeta^4)(\bar{\zeta}_i\zeta^i)^2, \zeta^0(\bar{\zeta}_4\zeta^4)(\bar{\zeta}_i\zeta^i)^3$$ antiduality sifts out half the combinations, namely: $$\zeta^0[1-(\bar{\zeta}_4\zeta^4)(\bar{\zeta}_i\zeta^i)^3/6],\quad \zeta^0[\bar{\zeta}_4\zeta^4)-(\bar{\zeta}_i\zeta^i)^2/2],$$ $$\zeta^0[(\bar{\zeta}_i\zeta^i)-(\bar{\zeta}_4\zeta^4)(\bar{\zeta}_j\zeta^j)^2/2], \quad\zeta^0[(\bar{\zeta}_i\zeta^i)(\bar{\zeta}_4\zeta^4)-(\bar{\zeta}_j\zeta^j)^2/2]$$ In particular as $\bar{\zeta}_0\bar{\zeta}_4\zeta^1\zeta^2\zeta^3$ and $\bar{\zeta}_4\zeta^0\zeta^1\zeta^2\zeta^3$ are self-dual, imposing antiduality eliminates these unwanted states, a good thing since they respectively have $F=3$ and $Q=-2$. Applying antiduality and focussing on left-$\Psi$, the resulting square contains the following varieties of up ($U$), down ($D$), charged lepton ($L$) and neutrinos ($N$), where the subscript distinguishes between repetitions. In the $\zeta\bar{\zeta}$ expansion table, $\times$ are duals, \* are conjugates: $r\backslash\bar{r}$ 0 1 2 3 4 5 ---------------------- ---- --------------------- ----------------------------------------- ------------------------------- --------------------- ---------- 0 $\!\!L_1,N_1,D_5^c$ $L_5^c,D_1,U_1$ 1 \* $\!\!L_{2,3},N_{2,3},D^c_{3,6,7},U_3^c$ $\!\!L_6^c,D_2,U_2$ 2 \* $\!\!L_4,N_4,D^c_{4,8},U_4^c$ $\times$ 3 \* \* $\times$ 4 \* \* $\times$ 5 \* \* \* Observe that colour singlet and triplet fermions come in 4s, 6s and 8s which comfortably contain the known three generations. However you will see that whereas $U,D$ in the first and second generation are bona fide weak isospin doublets, the third and fourth family $U,D$ are accompanied by another exotic colour triplet quark, call it $X$ say, having charge $Q =-4/3$, and make up a weak isospin triplet. So this is a departure from the standard model! Specifically the weak isospin generators are: $$T_+ = \zeta^0\partial_4 - \bar{\zeta}_4\bar{\partial}^0, \quad T_- = \zeta^4\partial_0 - \bar{\zeta}_0\bar{\partial}^4;$$ $$2T_3 = [T_+,T_-] = \zeta^0\partial_0-\zeta^4\partial_4+ \bar{\zeta}_4\bar{\partial}^4-\bar{\zeta}_0\bar{\partial}^0,\, {\rm so~we~meet}$$ $${\rm doublets~like~} (N_1,L_1), (U_1,D_1)\sim (\zeta^0,\zeta^4), \quad (-\bar{\zeta}_4,\bar{\zeta}_0)$$ $${\rm singlets~like~}L_5\,\,D_5 \sim (\bar{\zeta}_0\zeta^0 + \bar{\zeta}_4\zeta^4), \quad (\bar{\zeta}_0\zeta^0\bar{\zeta}_4\zeta^4),$$ $${\rm triplets~like~} (U_3,D_3,X_3)\sim (-\bar{\zeta}_4\zeta^0, [\bar{\zeta}_0\zeta^0-\bar{\zeta}_4\zeta^4]/\sqrt{2},\bar{\zeta}_0\zeta^4).$$ Exotic Particles, Generations & the Mass Matrix =============================================== The $U$-states arise from combinations like $\zeta^i\zeta^j\zeta^0$ lying in the $\overline{10}$-fold SU(5) combination $\zeta^\lambda\zeta^\mu\zeta^\nu$ and as $\bar{\zeta}_k\bar{\zeta}_4\zeta^0$. (Note that $\zeta^0\zeta^4\zeta^k$ has the exotic value $F=5/3$ and cannot be identified with $U^c$.) The $D$-states occur similarly, as do the $N$’s and $L$’s. However, observe that $L_{5,6}\sim \bar{\zeta}_3\bar{\zeta}_2\bar{\zeta}_1$ and $D_{5,6,7,8} \sim \bar{\zeta}_k$ are nominally weak isosinglets, which again [*differs*]{} from the standard model. These mysterious states are definitely charged but do not possess any weak interactions, so their behaviour is very curious. One might even regard $L_{5,6}$ like colourless charged baryons (antiproton-like) but really until one sees how these states mix with the usual weak isodoublet leptons it is dangerous to label them one thing rather than another without further research. ‘Pentaquarks’ such as $\Theta^+ \sim uudd\bar{s}$ & $\Xi^{--} \sim ddss\bar{u}$ were recently discovered(?) with quite narrow widths and many people have advanced models to describe these new resonances as well as tetraquark mesons. But who is to say unequivocally that they are not composites of ordinary quarks and another $U$-quark or other $D$-family quarks which my scheme indicates? Further, I get a fourth neutrino, which could be essentially sterile and might help explain the mystery of $\nu$ masses. This is clearly fertile ground for investigation and I have only scratched the surface here. Possibly conflicts with experiment may arise that will ultimately invalidate the entire property values scheme. One of the first matters to be cleared up is the mass matrix and flavour mixing which affects quarks as well as leptons. If we assume it is due to a Higgs $\Phi$ field’s expectation values, there are nine colourless possibilities having $F=Q=0$ lying in an antidual boson superfield: - one $\phi_{(0)(0)} = \langle\phi\rangle = {\it M}$ - one $\phi_{(0)(4)} = \langle\phi_{1234}\rangle = {\it H}{\rm ~complex}$ - three $\phi_{(1)(1)}=\langle\phi^0_0,\phi_4^4,\phi_i^i\rangle = {\it A,B,C}$ - four $\phi_{(2)(2)}=\langle\phi_{04}^{04},\phi_{0k}^{0k}, \phi_{4k}^{4k},\phi_{ij}^{ij}\rangle = {\it D,E,F,G},$ others being related by duality. With nine $\langle\phi\rangle$ the mass matrix calculation already becomes a difficult task! To ascertain what happens, consider the subset of quarks involving $U_{1-4}$ and $D_{1-4,7}$ interacting with anti-selfdual superHiggs. The Lagrangian $\int d^5\zeta d^5\bar{\zeta}\,\,\bar{\Psi}\langle\Phi\rangle\Psi$ produces $U$ and simplistic $D$ mixing matrices: $$2M(U)\!\rightarrow\!\left( \begin{array}{cccc} 2{\it M}+{\it F}/\sqrt{3}&{\it B}+{\it C}/\sqrt{3}&-{\it H}^* & 0\\ {\it B}+{\it C}/\sqrt{3} & 2{\it M} & 0 & 0 \\ -{\it H} & 0 & 2{\it M}+{\it G}/\sqrt{3} & 2{\it C}/\sqrt{3} \\ 0 & 0 & 2{\it C}/\sqrt{3} & 2{\it M} \end{array} \right)$$ $$2M(D)\!\!\rightarrow\!\!\left( \begin{array}{ccccc} \!\!2{\it M}+{\it E}/\sqrt{3}&\!{\it A}+{\it C}/\sqrt{3}& -{\it H}^*&0&0\\ \!\!{\it A}+{\it C}/\sqrt{3} & 2{\it M}& 0 & 0 & 0 \\ -{\it H}&0& 2{\it M}\!+{\it G}/\sqrt{3}&2{\it C}/\sqrt{3}&\! ({\it E}\!-\!{\it F})/\sqrt{3}\\ 0&0& 2{\it C}/\sqrt{3} & 2{\it M} & {\it A}-{\it B} \\ 0 & 0 & -({\it E}\!+\!{\it F})/{\sqrt{3}} & {\it A}-{\it B} & 2{\it M}-{\it D} \\ \end{array} \right).$$ Similar expressions can be found for the leptons and quarks. Such matrices have to be diagonalised via unitary transformations $V(U)$ & $V(D)$. However the coupling of the weak bosons is [*different*]{} for isodoublets $U_{1,2}~D_{1,2}$ and isotriplets $U_{3,4}~D_{3,4}$ (and zero for the weak isosinglet $D_7$). So we can’t just evaluate $V^{-1}(U)V(D)$ for the unitary CKM matrix now. Rather the weak interactions connecting $U$ & $D$ mass-diagonalised quarks will not be quite unitary (because of the different coupling factors): $$L_{\rm weak}/g_w = W^+(\bar{U}_1D_1 +\bar{U}_2D_2 ) +\sqrt{2}W^+(\bar{U}_3D_3+\bar{U}_4D_4+\bar{D_3}X_3+\bar{D}_4X_4)$$ $$\quad\quad + W^-(\bar{D}_1U_1 +\bar{D}_2U_2 ) +\sqrt{2}W^-(\bar{D}_3U_3+\bar{D}_4U_4+\bar{X_3}D_3+\bar{X}_4D_4)+ W^3{\rm ~terms}$$ So this is another departure from the standard picture: nonunitarity of the $3\times 3$ CKM matrix is a test of the scheme. Also CP violation is an intrinsic feature of the property formalism because $\it H$ is naturally complex, unlike the other expectation values ${\it A,B...M}$. A more realistic attempt for getting the quark and lepton masses would be to abandon antiduality in the Higgs sector and use all 18 expectation values; otherwise it may prove impossible to cover the 12 or more orders of magnitude all the way from the electron neutrino to the top quark (and higher). Generalized relativity ====================== We know that gauge fields can transport/communicate property from one place to another so where are they? Maybe one can mimic the SUSY procedure and supergauge the massless free action for $\Psi$, without added complication of spin. But there is a more compelling way, which has the benefit of incorporating gravity. Construct a fermionic version of Kaluza-Klein (KK) theory [@KKF], this time without worrying about infinite modes which arise from squeezing normal bosonic coordinates. These are the significant points of such an approach: - One must introduce a fundamental length $\Lambda$ in the extended $X$, as property $\zeta$ has no dimensions; maybe this is the gravity scale $\kappa= \sqrt{8\pi G_N}$? - Gravity (plus gauge field products) fall within the $x-x$ sector, gauge fields in $x-\zeta$ and the Higgs scalars must form a matrix in $\zeta-\zeta$, - Gauge invariance is connected with the number of $\zeta$ so $SU(5)$ or $Sp(10)$ [*or perhaps a subgroup*]{} are indicated, - There is no place for a gravitino as spin is absent ($\zeta$ are Lorentz scalar), - There are necessarily a small finite number of modes, - Weak left-handed SU(2) is associated with rotations of $\zeta$ [**not**]{} $\bar{\zeta}$ so may have something to with $\zeta$-analyticity. The real metric specifies the separation in location as well as property: it tells us how ‘far apart’ and ‘different in type’ two events are. Setting $\bar{\zeta}^{\bar{\mu}}\equiv \bar{\zeta}_\mu$, $$ds^2 = dx^mdx^n{G_{nm}} + dx^md\zeta^\nu{G_{\nu m}} + dx^n d\bar{\zeta}^{\bar{\mu}}{G_{\bar{\mu} n}} + d\bar{\zeta}^{\bar{\mu}}d\zeta^\nu{G_{\bar{\mu}\nu}}$$ where the tangent space limit corresponds to Minkowskian $$G_{ab}\rightarrow I_{ab}= \eta_{ab},\, {G_{\bar{\alpha}\beta}}\rightarrow I_{\bar{\alpha}\beta} = \Lambda^2{\delta_{\alpha\beta}},$$ multiplied at least by $(\bar{\zeta}\zeta)^5$ — to arrange correct property integration. Proceeding to curved space the components should contain the force fields, leading one to a ‘superbein’ $${\bar{E}_M}^A = \left( \begin{array}{cc}{{e_m}^a}& {i\Lambda(A_m)_\mu^\alpha\zeta^\mu} \\ 0 & {\Lambda\delta_\mu^\alpha} \end{array} \right),$$ and the metric “tensor” arising from $$ds^2 = dx^m dx^n g_{nm}+2\Lambda^2 [d\bar{\zeta}^{\bar{\mu}}-idx^m\bar{\zeta}^{\bar{\kappa}} (A_m)_{\bar{\kappa}}^{\bar{\mu}}]\delta_{{\bar{\mu}}\nu} [d\zeta^\nu + idx^n (A_n)^\nu_\lambda\zeta^\lambda];$$ $$g_{mn} = e_m^ae_n^b\eta_{ab}.$$ Gauge symmetry corresponds to the special change $\zeta^\mu\rightarrow \zeta'^\mu=[\exp(i\Theta(x))]^\mu_\nu\zeta^\nu $ with $x'=x$. Given the standard transformation law $${G_{\zeta m}}(X)=\frac{\partial X^{\prime R}}{\partial x^m} \frac{\partial X^{\prime S}}{\partial\zeta} G'_{SR}(X')(-1)^{[R]} ={\frac{\partial \zeta'}{\partial\zeta}}{G'_{\zeta m}}- {\frac{\partial \bar{\zeta}'}{\partial x^m}}{ \frac{\partial \zeta'}{\partial\zeta}G'_{\zeta\bar{\zeta}}}.$$ this translates into the usual gauge variation (a matrix in property space), $$A_m(x) =\exp(-i\Theta(x))[A'_m(x) - i\partial_m]\exp(i\Theta(x)).$$ The result is consistent with other components of the metric tensor but does not fix what (sub)group is to be gauged in property space although one most certainly expects to take in the nonabelian colour group and the abelian electromagnetic group, so as to agree with physics. Some notational niceties ======================== When dealing with commuting and anticommuting numbers within a single coordinate framework $X^M = (x^m,\zeta^\mu)$ one has to be [**exceedingly**]{} careful with the order of quantities and of labels. I cannot stress this enough. (It took me six months and much heartache to get the formulae below correct.) For derivatives the rule is $dF(X) = dX^M (\partial F/\partial X^M) \equiv dX^M \partial_M F$, [**not**]{} with the $dX$ on the right, and for products of functions: $d(FG..)=dF\,G+F\,dG+..$. Coordinate transformations read $dX^{\prime M}=dX^N(\partial X^{\prime M}/\partial X^N)$, and [*in that particular order*]{}. Since $ds^2 = dX^N dX^M \,G_{MN}$, the symmetry property of the metric is $G_{MN}= (-1)^{[M][N]}G_{NM}$. Let $G^{LM}G_{MN} = \delta_N^L$ for the inverse metric, so $G^{MN}=(-1)^{[M]+[N]+[M][N]}\,G^{NM}$. The notation here is $[M]=0$ for bosons and 1 for fermions. Changing coordinate system from $X$ to $X'$, we have to be punctilious with signs and orders of products, things we normally never care about; the correct transformation law is $$G_{NM}(X) = \left(\frac{\partial X^{\prime R}}{\partial X^M}\right)\left(\frac {\partial X^S}{\partial X^N}\right)G'_{SR}(X')\,(-1)^{[N]([R]+[M])}.$$ Transformation laws for contravariant and covariant vectors read: $$V^{\prime M}(X') = V^R(X)\left(\frac{\partial X^{\prime M}}{\partial X^R}\right) \quad {\rm and}\quad A'_M(X')=\left(\frac{\partial X^R} {\partial X^{\prime M}}\right)A_R(X),$$ in the order stated. Thus the invariant contraction is $$V^{\prime M}(X')A'_M(X') = V^R(X)A_R(X) = (-1)^{[R]}A_R(X)V^R(X).$$ The inverse metric $G^{MN}$ can be used to raise and lower indices as well as forming invariants, so for instance $V_R\equiv V^SG_{SR}$ and $V^{\prime R}V^{\prime S} G'_{SR} = V^MV^NG_{NM}$. The next issue is covariant differentiation; we insist that $A_{M;N}$ should transform like $T_{MN}$, viz. $$T'_{MN}(X') =(-1)^{[S]+[N])[R]}\left(\frac{\partial X^R}{\partial X^{\prime M}} \right)\left(\frac{\partial X^S}{\partial X^{\prime N}}\right) T_{RS}(X).$$ After some work we find that $$A_{M;N} = (-1)^{[M][N]}A_{M,N} - A^L\Gamma_{\{MN,L\}},$$ where the connection is given by $$\begin{aligned} \Gamma_{\{MN,L\}}&\equiv& [(-1)^{([L]+[M])[N]}G_{LM,N} +(-1)^{[M][L]}G_{LN,M} -G_{MN,L}]/2 \\ &=& (-1)^{[M][N]}\Gamma_{\{NM,L\}}.\nonumber\end{aligned}$$ Another useful formula is the raised connection $${\Gamma_{MN}}^K \equiv (-1)^{[L]([M]+[N])}\Gamma_{\{MN,L\}}G^{LK} = (-1)^{[M][N]}{\Gamma_{NM}}^K,$$ whereupon one may write $$A_{M;N} = (-1)^{[M][N]}A_{M,N} - {\Gamma_{MN}}^LA_L.$$ Similarly one can show that for double index tensors the true differentiation rule is $$T_{LM;N}\!\equiv\!(-1)^{[N]([L]\!+\![M])}T_{LM,N}- (-1)^{[M][N]}{\Gamma_{LN}}^KT_{KM} -(-1)^{[L]([M]\!+\![N]\!+\![K])}{\Gamma_{MN}}^KT_{LK}.$$ As a nice check, the covariant derivative of the metric properly vanishes: $$G_{LM;N} \equiv (-1)^{[N]([L]+[M])}G_{LM,N} -(-1)^{[L][M]}{\Gamma}_{\{LN,M\}}- \Gamma_{\{MN,L\}} \equiv 0.$$ Moving on to the Riemann curvature we form doubly covariant derivatives: $$A_{K;L;M} - (-1)^{[L][M]}A_{K;M;L} \equiv (-1)^{[K]([L]+[M])}{R^J}_{KLM}A_J$$ where one discovers that $$\begin{aligned} {R^J}_{KLM}&\equiv& (-1)^{[K][M]}({\Gamma_{KM}}^J)_{,L} - (-1)^{[L]([K]+[M])}({\Gamma_{KL}}^J)_{,M} \nonumber \\ & & + (-1)^{[M]([K]+[L])+[K][L]}{\Gamma_{KM}}^N{\Gamma_{NL}}^J - (-1)^{[K]([M]+[L])}{\Gamma_{KL}}^N{\Gamma_{NM}}^J.\end{aligned}$$ Evidently, ${R^J}_{KLM} = -(-1)^{[L][M]}{R^J}_{KML}$ and, less obviously, the cyclical relation takes the form $$(-1)^{[K][L]}{R^J}_{KLM} + (-1)^{[L][M]}{R^J}_{LMK}+ (-1)^{[M][K]}{R^J}_{MKL}=0.$$ The fully covariant Riemann tensor is $R_{JKLM}\equiv (-1)^{([J]+[K])[L]}{R^N}_{KLM}G_{NJ}$ with pleasing features: $$\begin{aligned} R_{JKLM} & = & -(-1)^{[L][M]}R_{JKML} = -(-1)^{[J][K]}R_{KJLM},\\ 0 & = & (-1)^{[J][L]}R_{JKLM}+(-1)^{[J][M]}R_{JLMK}+(-1)^{[J][K]}R_{JMKL} \\ R_{JKLM}& = & (-1)^{([J]+[K])([L]+[M])} R_{LMJK}.\nonumber\end{aligned}$$ Finally proceed to the Ricci tensor and scalar curvature: $$\begin{aligned} R_{KM} &\equiv& (-1)^{[J]+[K][L]+[J]([K]+[M])}G^{LJ}R_{JKLM}\nonumber \\ & = &(-1)^{[L]([K]+[L]+[M])}{R^L}_{KLM}\!=\!(-1)^{[K][M]}R_{MK},\end{aligned}$$ $$R\equiv G^{MK}R_{KM}.$$ Curvatures of space-time-property ================================= When Einstein produced his general theory of relativity with Grossmann he wrote all his expressions in terms of real variables. In order to avoid any confusion with complex variables, we shall copy him by writing everything in terms of real coordinates $\xi,\eta$ rather than complex $\zeta=(\xi+i\eta)/\sqrt{2}$. (Note that the real invariant is $\bar{\zeta}\zeta = i\xi\eta$ and that a phase transformation of $\zeta$ corresponds to a real rotation in $(\xi,\eta)$ space.) For simplicity consider just one extra pair (rather than five pairs) and the following two examples, which are complicated enough as it is. \(1) Decoupled property and space-time, but both curved: $$\begin{aligned} ds^2&=&dx^mdx^nG_{nm}(x,\xi,\eta)+2id\xi d\eta G_{\eta\xi}(x,\xi,\eta)\nonumber \\ &\equiv& dx^mdx^ng_{nm}(x)(1+if\xi\eta)+2i\Lambda^2d\xi d\eta(1+ig\xi\eta)\end{aligned}$$ from which we can read off the metric components ($G^{LM}G_{MN}\equiv\delta^L_N$) $$\left( \begin{array}{ccc} G_{mn} & G_{m\xi} & G_{m\eta}\\ G_{\xi n} & {G_{\xi\xi}} & {G_{\xi\eta}}\\ G_{\eta n} & {G_{\eta\xi}} & {G_{\eta\eta}} \end{array} \right) = \left( \begin{array}{ccc} g_{mn}(1+if\xi\eta) & 0 & 0 \\ 0 & 0 & -i\Lambda^2(1+ig\xi\eta)\\ 0 & i\Lambda^2(1+ig\xi\eta) & 0 \end{array} \right),$$ $$\left( \begin{array}{ccc} G^{lm} & G^{l\xi} & G^{l\eta}\\ G^{\xi m} & {G^{\xi\xi}} & G^{\xi\eta}\\ G^{\eta m} & G^{\eta\xi} & G^{\eta\eta} \end{array} \right) = \left( \begin{array}{ccc} g^{lm}(1-if\xi\eta) & 0 & 0 \\ 0 & 0 & -i(1-ig\xi\eta)/\Lambda^2\\ 0 & i(1-ig\xi\eta)/\Lambda^2 & 0 \end{array} \right).$$ The non-zero connections in the property sector are $$\Gamma_{\xi\eta}^\xi = -{\Gamma_{\eta\xi}}^\xi = ig\xi,\quad {\Gamma_{\xi\eta}}^\eta = -{\Gamma_{\eta\xi}}^\eta = ig\eta$$ so $$\begin{aligned} { {R^\eta}_{\xi\eta\eta}}&=& -2ig(1+ig\xi\eta) = -{{R^\xi}_{\eta\xi\xi}}\\ {{R^\xi}_{\xi\xi\eta}}&=& -ig(1+ig\xi\eta) =- {{R^\eta}_{\eta\eta\xi}}\\ {\rm so~}{R_{\xi\eta}}&=&-{R_{\eta\xi}}=3ig(1+ig\xi\eta).\end{aligned}$$ Consequently the total curvature is given by $$R = {G^{mn}R_{nm}}+2{G^{\eta\xi}R_{\xi\eta}} = R^{(g)}(1-if\xi\eta)-6g/\Lambda^2.$$ Since $\sqrt{-G..}= -i\Lambda^2\sqrt{-g..}(1+2if\xi\eta)(1+ig\xi\eta)$, we obtain an action $$I \equiv \frac{1}{2\Lambda^4}\int\,R\,\sqrt{G..}\,d^4x d\eta d\xi = \frac{1}{2\kappa^2}\int\, d^4 x\sqrt{-g..} \left[R^{(g)} +\lambda\right]$$ where $\kappa^2 \equiv 8\pi G_N = \Lambda^2/(f+g)$, $R^{(g)}$ is the standard gravitational curvature and $\lambda = -6g(2f+g)/\Lambda^2(f+g)$ corresponds to a cosmological term. \(2) Our second example leaves property space flat (in the $\eta,\xi$ sector) but introduces a U(1) gauge field $A$, governed by the metric, $$\left( \begin{array}{ccc} G_{mn} & G_{m\xi} & G_{m\eta}\\ G_{\xi n} & G_{\xi\xi} & G_{\xi\eta}\\ G_{\eta n} & G_{\eta\xi} & G_{\eta\eta} \end{array} \right)\!\! =\!\! \left( \begin{array}{ccc} g_{mn}(1\!+\!if\xi\eta)\!+\!2i\Lambda^2\xi A_mA_n\eta & i\Lambda^2A_m\xi & i\Lambda^2A_m\eta \\ i\Lambda^2 A_n\xi & 0 & -i\Lambda^2\\ i\Lambda^2A_n\eta & i\Lambda^2 & 0 \end{array} \right).$$ Simplify the analysis somewhat by going to flat (Minkowski) space first as there are then fewer connections. After some work ($F_{mn}\equiv A_{m,n} -A_{n,m}$) one obtains, $$\begin{aligned} {\Gamma_{\xi\eta}}^\xi&=& {\Gamma_{\xi\eta}}^\eta ={\Gamma_{\xi\eta}}^k = 0,\\ {{\Gamma_{m\xi}}^\xi} &=& {{\Gamma_{m\eta}}^\eta} =i\Lambda^2A^lF_{lm}\xi\eta/2, \quad {{\Gamma_{m\xi}}^\eta}=-{{\Gamma_{m\eta}}^\xi} = A_m,\\ {{\Gamma_{m\xi}}^l}&=& i\Lambda^2{F^l}_m\xi/2,\quad {{\Gamma_{m\eta}}^l} = i\Lambda^2{F^l}_m\eta/2,\\ {{\Gamma_{mn}}^\xi} &=& -A_mA_n\xi -(A_{m,n}+A_{n,m})\eta/2,\\ {{\Gamma_{mn}}^\eta} &=& -A_mA_n\eta +(A_{m,n}+A_{n,m})\xi/2,\\ {{\Gamma_{mn}}^k} &=& i\Lambda^2(A_m{F^k}_n+A_n{F^k}_m)\xi\eta.\end{aligned}$$ Other Christoffel symbols can be deduced through symmetry of indices. Hence $${R_{km}}={{R^l}_{klm}-{R^\xi}_{k\xi m}-{R^\eta}_{k\eta m}}= -i\Lambda^2(A_{k,l}+A_{l,k}){F^l}_m\xi\eta/2 + {\rm total~der.}$$ $${R_{k\xi}}={R^l}_{kl\xi}+{R^\xi}_{k\xi\xi}+{R^\eta}_{k\eta\xi}= i\Lambda^2[{F^l}_{k,l}\xi/2+A^lF_{k,l}\eta]+{\rm total~der.}$$ $${R_{k\eta}}={R^l}_{kl\eta}+{R^\xi}_{k\xi\eta} +{R^\eta}_{k\eta\eta} = i\Lambda^2[{F^l}_{k,l}\eta/2-A^lF_{k,l}\xi]+{\rm total~der.}$$ $${R_{\xi\eta}}=-\Lambda^4F_{kl}F^{lk}\xi\eta.$$ Then covariantize by including the gravitational component $g_{mn}(1+if\xi\eta)$ to end up with the total curvature: $$\begin{aligned} R &=& {G^{mn}R_{nm}} + 2{G^{m\xi}R_{\xi m}} + 2{G^{m\eta}R_{\eta m}}+2{G^{\eta\xi}R_{\xi\eta}}\\ &\rightarrow& R^{(g)} -3i\Lambda^2 g^{km}g^{ln}F_{kl}F_{nm}\xi\eta/2\end{aligned}$$ Finally rescale $A$ to identify the answer as electromagnetism + gravitation: $$\int R\sqrt{-G..}d^4x d\eta d\xi /4\Lambda^4 = \int d^4x \sqrt{-g..} \left[R^{(g)}/2\kappa^2 - F^{kl}F_{kl}/4\right],$$ where $\kappa^2=\Lambda^2/f =8\pi G_N$. It is a nice feature of the formalism that the gauge field Lagrangian arises from space-property terms — like the standard K-K model (from the tie-up between ordinary space-time and the fifth dimension). Where do we go from these two examples? Well some generalizations come to mind: - replace property couplings $f$ and $g$ by two fields (dilaton and Higgs), - combine the two models; this should lead to gravity + em + cosmic const., - extend fully to five $\zeta$; it is easy enough to incorporate the standard gauge model and one can even entertain a GUT $SU(5)$ of some ilk, - work out the particle mass spectrum from all the $\langle\phi\rangle$. Acknowledgements {#acknowledgements .unnumbered} ================ A considerable amount of the material has been drawn from previous papers written with several collaborators including Peter Jarvis, Ruibin Zhang, Roland Warner and Martin White. I would like to take the opportunity to thank them all for their valuable insights and help. [9]{} R. Delbourgo, [*J. Phys. A*]{} [**39**]{}, 5175 (2006). R. Delbourgo, [*J. Phys. A*]{} [**39**]{}, 14735 (2006). A. McKane, Phys. Lett. [*A76*]{}, 22 (1980);\ P. Cvitanovic and A.D. Kennedy, Phys. Scripta [**26**]{}, 5 (1982);\ I.G. Halliday and R.M. Ricotta, Phys. Lett. [**B193**]{}, 241 (1987);\ G.V. Dunne, J. Phys. [**A22**]{}, 1719 (1989). P.D. Jarvis and M. White, Phys. Rev. [**D43**]{}, 4121 (1991);\ R. Delbourgo, P.D. Jarvis, R.C. Warner, Aust. J. Phys. [**44**]{}, 135 (1991). R. Delbourgo, S. Twisk and R. Zhang, Mod. Phys. Lett. [**A3**]{}, 1073 (1988);\ P. Ellicott and D.J. Toms, Class. Quant. Grav. [**6**]{}, 1033 (1989);\ R. Delbourgo and M. White, Mod. Phys. Lett. [**A5**]{}, 355 (1990).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Tunneling spectroscopy is applied to tunnel junctions with only one or no ferromagnetic electrode to study the excitation of quasi particles in magnetic tunnel junctions. The bias dependence is investigated with high accuracy by inelastic electron tunneling spectroscopy. Both types of junctions show a zero bias anomaly that is different in size and sign compared to magnetic tunnel junctions, i.e. junctions with two ferromagnetic electrodes. A pronounced difference is also found depending on the material that is probed by the tunneling electrons, which might be attributed to the excitation of magnons.' author: - Volker Drewello - Zoë Kugler - Günter Reiss - Andy Thomas bibliography: - 'paper-database.bib' title: Tunneling spectroscopy probing magnetic and nonmagnetic electrodes in tunnel junctions --- Magnetic tunnel junctions (MTJs) with MgO as a crystalline barrier show large tunnel magnetoresistance (TMR) ratios of up to 1000% at low temperature [@Ikeda2008; @Tezuka2009]. Nevertheless, the TMR ratio at room temperature is still by a factor of 2 to 3 smaller. Decreasing this temperature dependence is one way to gain higher TMR ratios at room temperature and increase the applicability of MTJ based spintronic devices. The main reason for the decreasing TMR values are intrinsic excitations within the junctions [@Bratkovsky1998; @Han2001; @Drewello2008] which can be studied by means of inelastic electron tunneling spectroscopy (IETS). In this paper, the different excitations in magnetic and nonmagnetic electrodes are investigated. Tungsten is chosen as the nonmagnetic metal electrode because of its high melting point. It also has a lower lattice mismatch with the MgO barrier than e.g. Tantalum (6% vs. 11%). The samples are prepared in a self-made sputter deposition tool with a base pressure below $10^{-9}$mbar. The layers are deposited on top of thermally oxidized silicon wafers at an Ar pressure of $5\times10^{-3}$mbar (MgO at $10^{-2}$mbar). The first sample is a metal/ insulator/ metal tunnel junction (M-I-M) with the layer stacking W 20/ MgO 1.8/ W 20 (all values in nm). The second sample has a ferromagnet as the lower electrode (FM-I-M). The layer stack is W 15/ Co$_{40}$Fe$_{40}$B$_{20}$ 6/ MgO 1.8/ W 20. Both stacks are annealed at 723K for 1 hour in a high vacuum furnace. A capping of Ta and Au is added for protection and to form contact pads. The samples are structured by optical lithography and Ar ion beam etching. The low temperature measurements are done at 13K in a closed cycle Helium cryostat by a standard two-probe technique. The bias voltage is defined with respect to the upper electrode. Thus, negative bias results in electrons tunneling into the upper electrode. A Lock-In technique is used to obtain the d$I$/d$V$ curves which are differentiated numerically to get the IET spectra (d$^2I/$d$V^2$). Details of the measurement setup and procedure can be found elsewhere [@Drewello2009]. ![The tunnel spectrum (d$I$/d$V$) normalized to its value at $V=0$ and the IET spectrum (d$^2I$/d$V^2$) of the W/ MgO/ W sample.[]{data-label="fig1"}](MIM_IETS) The spectra for the M-I-M junction are shown in Figure \[fig1\]. The conductance has the predicted parabolic shape [@Simmons1964] with variations at bias voltages up to 100mV. In the IET spectrum these features can be seen much clearer. First, at approximately 9mV a negative zero bias anomaly is found, i.e. the conductance decreases at low bias compared to zero bias. In typical full MTJs (i.e. with two magnetic electrodes) the zero bias anomaly shows the opposite sign. Often the magnitude of the effect is larger and dominates the spectra [@Moodera1998; @Miao2006; @Drewello2009]. Second, several broad peaks are found up to 100mV. Figure \[fig2\] shows these features in more detail. As no magnetic materials are used and no magnetic impurities are expected, phonon excitation is an explanation for these peaks. Peaks of the electrode phonons typically have energies around 30meV [@Klein1973] while the phonons of barrier oxides have higher energies. Here, the first peaks correspond to an excitation of tungsten phonons with an energy of 26meV [@Olejniczak1998]. The MgO tunneling barrier leads to phonon peaks at 66 and 81mV [@Klein1973]. A strong peak is indeed found at 66mV but only a shoulder is found at 81mV. In MgO based full MTJs this peak is typically more pronounced [@Miao2006; @Drewello2009]. ![The IET spectrum of the W/ MgO/ W junction. The low bias region is shown for both polarities. The arrows mark energies of known excitations.[]{data-label="fig2"}](MIM_IETS_mag) ![The normalized tunnel spectrum and the IET spectrum of the Co-Fe-B/ MgO/ W sample. []{data-label="fig3"}](FMIM_IETS) ![Comparison of the IET spectra for positive (electrons tunnel into Co-Fe-B) and negative (into W) bias voltage for the Co-Fe-B/ MgO/ W sample. []{data-label="fig4"}](FMIM_IETS_mag) The spectra of the FM-I-M junction show a strong asymmetry in the slope of the dI/dV-curves (Figure \[fig3\]). This leads to a different height of the IET spectra for positive compared to negative bias, nevertheless the peak structure is visible. Magnified spectra for negative (electrons tunnel into W) and positive bias (electrons tunnel into Co-Fe-B) can be compared in Figure \[fig4\]. Most prominent is the zero bias feature which is still negative for both bias polarities. For negative bias it is found at approximately 10mV, while it is closer to zero for positive bias. This is presumably caused by smearing with the following peaks, which have a much higher intensity (X). For both polarities several features are found. The spectrum for negative bias looks similar to the spectrum of the first sample, with the tungsten phonon peak at 26mV and a MgO phonon peak at 66mV. The 81mV shoulder is more pronounced. For positive bias, i.e. tunneling into the ferromagnet, the peaks are higher and less sharp. After the neagtive ZBA the spectrum rises strongly to the first peak. This must be the excitation of magnons (X). Also, the MgO phonon peak at 81mV peak is more pronounced than the one at 66mV. ![Comparison of the IET spectra of the Co-Fe-B/ MgO/ W sample and those of a typical MTJ (scaled). Arrows mark the ZBA.[]{data-label="fig5"}](FMIM_IETS_compare_with_PSV430) In Figure \[fig5\] the IET spectra are compared to those of a typical MTJ [^1].The FM-I-M spectrum for tunneling into W (negative bias) is compared to the spectrum for the parallel (P) magnetic state. For tunneling into the FM (pos. bias) the spectrum is compared to the MTJ in antiparallel (AP) magnetic state. In both cases the spectra have a similar shape. The striking difference is the zero bias anomaly, which is positive for the MTJ and leads to a huge peak in the P state. This effect is also visible in the AP state, but the zero bias peak is only a shoulder. As a result the spectrum immediately rises after zero bias, while for the FM-I-M junction the flank seems shifted to higher bias. The MgO phonon peaks are also similar in both cases. The spectra of the full MTJ in the P state are negative over a wide bias range. In the AP state, the spectra are much lower compared to the FM-I-M-spectra in the same bias region. Now the results will be discussed. In the presented spectra most peaks can be identified as belonging to the target electrode and the barrier with the exception of the zero bias anomaly (ZBA). Its size and sign is obviously not simply a matter of the material of the target electrode. This can be seen from the results of the FM-I-M junction, where the ZBA is roughly the same for the two different electrode materials. Furthermore, the same ferromagnet shows large positive ZBA peaks in the MTJ. The two different results suggest that it is not a excitation of surface magnons which is causing the ZBA. This leaves tunneling through impurities as an explanation. It is known that the zero bias effect depends on the impurity material [@Cooper1973]. A model that would explain impurities in the barrier is implantation of upper electrode atoms during preparation. In this case the material would be tungsten in both samples (in difference to the MTJ case). ![Comparison of IET spectra for positive/negative bias and even and odd spectra of the Co-Fe-B/ MgO/ W sample.[]{data-label="fig6"}](FMIM_IETS_EO) The M-I-M sample shows near ideal symmetry and the anticipated peaks. The asymmetry in the spectra of the FM-I-M sample, however, rises the question of magnon excitation in the ferromagnet. If even/odd spectra [^2] are calculated, peak structures are indeed found in the odd spectrum (Figure \[fig6\]). The first peak must be the excitation of magnons, as observed before. However, the origin of the second, much broader peak is not clear. It’s maximum is at high bias voltage around 150mV and leads up to some hundred mV. This result is different compared to the peak at 10mV that Paluskar et al. find for incoherent Alumina based junctions [@Paluskar2007]. The high energies we find would be equivalent to temperatures higher than 1000 K and, therefore, $T_C$. The total magnon density of states in a ferromagnet is large at these energies [@Halilov1997] and bulk magnons are suspected to contribute to the tunneling process at some voltage [@Bratkovsky1998]. However, it is not clear if the interaction potential allows these modes to be excited [@Balashov2008]. Nevertheless, the similarity of the FM-I-M spectra to the state-specific spectra of the MTJ also fits in this model. In the case of electrons tunneling into Co-Fe-B magnons should be the dominant excitation. This is also the case in the AP state of a MTJ where the direct tunneling contribution is smaller due to the inverse spinpolarization of the electrodes. As the major difference is caused by the anomaly around zero bias, it can be suspected that the broad contribution, which is not seen in the P state, is also the excitation of magnons. Lastly, the broad dip (or gap) in the tunnel spectrum (d$I$/d$V$) of the FM-I-M junction resembles the P state spectrum of MTJs that incorporate one half-metallic Heusler compound electrode [@Sakuraba2006; @DrewelloAPL2009]. In both cases magnon excitation is prohibited for the corresponding bias polarity, while it is allowed for the other one. This is another hint that the magnon excitation is the origin of the broader feature (gap in d$I$/d$V$, background in IETS). A minor note regards the MTJ spectrum in the P state. It is shifted to negative values, which might indicate a large coherent tunneling contribution [@Tsunegi2008]. This is not seen in the M-I-M sample, which might be tentatively ascribed to different growth of the MgO barrier depending on the lower electrode material. The different strength of the MgO phonons peaks could then also be a effect of different barrier and interface properties. In summary, tunnel junctions with only one or no ferromagnetic electrode have been investigated by inelastic electron tunneling spectroscopy. The excitations of electrode and barrier phonons are observed in all junctions. For the junction with one ferromagnetic electrode the excitation spectra show a strong asymmetry, which is attributed to magnon excitation. In contrast to full magnetic tunnel junction, the presented junctions show a negative zero bias anomaly. We gratefully acknowledge Jan Schmalhorst for helpful discussions, Patryk Krzysteczko and Markus Schäfers for technical assistance, and the DFG (Grant RE 1052/13-1) for financial support. [^1]: The full MTJ is a pseudo spin valve with a similar layer stack: Ta 20 / Co-Fe-B 5.4/ MgO 2.4/ Co-Fe-B 2.4/ Ta 20. It is annealed at 723K for 1 hour. [^2]: The even (odd) spectrum is the average (difference) of the spectra for positive and negative bias.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Stein’s method is a popular method used to derive upper-bounds of distances between probability distributions. It can be viewed, in certain of its formulations, as an avatar of the semi-group or of the smart-path method used commonly in Gaussian analysis. We show how this procedure can be enriched by Malliavin calculus leading to a functional approach valid in infinite dimensional spaces.' address: 'Institut Mines-Telecom, Telecom ParisTech, CNRS LTCI, Paris, France' author: - 'L. Decreusefond' title: 'The Stein-Dirichlet-Malliavin method' --- Introduction {#sec:introduction} ============ Distances between probability or probability metrics is a very old topic since it is rich of a wide range of applications. As mathematical objects, it is natural to define a metric topology on spaces of probability measures. As modeling objects, it is natural to compare probability measures which appear in the mathematical representations of random phenomena. This topic has at least three facets: The diverse definitions of probability metrics which are tailored for each applications; the computations and comparisons of these different distances for the widest possible range of situations and at last, the applications which go from mathematical considerations like functional inequalities to more practical results of rate of convergence of stochastic algorithms. The Figure \[fig\_decreusefond\_esaim:mindmap\] shows a partial view of the different aspects of this subject. node\[concept \] [Probability\ metrics]{} \[clockwise from=0\] child\[concept color=macouleur2\] [node\[concept\] [Computations]{} \[clockwise from=22\] child\[concept color=blue!60\] [node\[concept\] [Girsanov\ $\Q\ll \P$]{}]{} child\[concept color=red\] [node\[concept\] [Stein\ $\Q=T^*m$]{}]{} ]{} child\[concept color=macouleur2\] [node\[concept\] [Types]{} child\[concept color=macouleur3\] [node\[concept\] [Optimal\ Transport]{} \[clockwise from=-10\] child\[concept color=red\] [node\[concept\] [Rubinstein]{}]{} child\[concept color=blue!60\] [node\[concept\] [Wasserstein]{}]{} ]{} child\[concept color=macouleur3\] [node\[concept\] [Prohorov]{}]{} child\[concept color=blue!60\] [node\[concept\] [Entropy]{}]{} ]{} child\[concept color=macouleur2\] [node\[concept\] [Applications]{} \[clockwise from=-45\] child\[concept color=blue!60\] [node\[concept\] [Functional\ inequality]{}]{} child\[concept color=red\] [node\[concept\] [Convergence\ rate]{}]{} child\[concept color=macouleur3\] [node\[concept\] [Ergodicity]{}]{} ]{} ; A few words are in order to explain the blue and red colors. For the computations of distances between measures $\mu$ and $\nu$, we need to impose some relationships between these two measures. Absolute continuity is one very frequent type of relationships between two measures. The Radon-Nykodim theorem gives a precious tool to estimate divergence-like and Wasserstein distances (see for instance [@MR2036490] for such an application). One may also reverse the point of view: Given a positive function $F$, compare the $\mu$ and $\nu=F\d\mu$ to obtain some precious functional inequalities on $F$ (see [@MR3155209]). These results thus belong to the same *spirit* and are colored in blue. Another natural way to put a structure between two measures is to have a map which transforms a known measure into another one and to compare this transformed measure to a reference probability. This is exactly the framework in which the Stein’s method performs well if we consider Kantorovitch-Rubinstein type distances (defined below). Typical applications of these form of distances are to give the convergence rates of celebrated theorem like CLT or Berry-Esseen Theorem or of random algorithms [@MR99k:28007]. The links between these different points justify that they are all colored in red. This paper is a rather informal introduction to the Stein-Dirichlet-Malliavin method (SDM for short henceforth). This is an extension of the classical Stein’s method, enriched by the structure given by Dirichlet forms and Malliavin calculus. We hope that this new point of view will lead to more systematic proofs of convergence, extending their applicability. The price to pay is to master some new concepts from Malliavin calculus like the gradient and its associated adjoint. That is why we tried to maintain the technicalities at the lowest possible level, insisting more on the ideas at play. We first show the different kinds of probability metrics that exist in the literature. We do not pretend to be exhaustive but aim to point out to the wide diversity of possible definitions. In Section \[sec:probability-metrics\], we establish the principles of the SDM method and show how it can be applied to the Poisson-Gaussian convergence. We then explain how to construct the necessary structures to extend this procedure to infinite dimensional spaces. In Section \[sec:edgeworth-expansion\], Edgeworth expansions are obtained by iterating the previous procedure as often as desired. Taxonomy of probability metrics {#sec:probability-metrics} =============================== In what follows, all the probability measures are defined on Polish spaces denoted either by $\EE$ or $\FF$, whose borelian $\sigma$-fields is $\mathfrak B(\EE)$, respectively $\mathfrak B(\FF)$. There are several notions of metrics between probability measures. An interesting survey of the main variants and their mutual relationships can be found in [@INSR:INSR419]. Each of one is often adapted to a particular purpose. They can roughly and partly be classified in three types. The first one is the so-called Prokhorov distance. $$\text{Dist}_{\text{Pro}}(\P,\Q)=\inf\Bigl\{\epsilon >0, \P(A)\le \Q(A^\epsilon)+\epsilon \text{ for all } A\in \mathfrak B(\EE)\Bigr\},$$ where $A^\epsilon$ is the $\epsilon$-neighborhood of $A$ defined by $ A^\epsilon=\{y\in \EE,\ \exists x\in A, d(x,y)\le \epsilon\}.$ This distance is crucial as its associated topology is precisely the topology of the convergence in distribution, i.e. we have the following theorem which can be found in [@MR982264]. A sequence $(\P_n,\, n\ge 1)$ of probability measures converges weakly to $\P$ if and only if $ \text{Dist}_{\text{Pro}}(\P_n,\P)$ tends to $0$ as $n$ goes to $\infty$. Unfortunately, this distance is hardly computable and that justifies the search for alternative and more tractable definitions. A vast category of probability metrics is represented by the $f$-divergence defined as follows. Let $f$ be a convex function such that $f(1)=0$. Then, for two probability measures $\P$ and $\Q$ on a Polish space $\EE$, $$D_f(\Q \| \P)= \begin{cases} \displaystyle{\int_{\EE}} f\left(\frac{\xdif\Q}{\xdif\P}\right)\xdif \P & \text{ if } \Q\ll \P,\\ \infty & \text{ otherwise.} \end{cases}$$ For instance, if we choose $f=t\ln t$, we obtain the Kullblack-Leibler distance. The Hellinger distance corresponds to the case where $f(t)=(\sqrt{t}-1)^2$. Total variation between absolutely continuous measures boils down to take $f(t)=|t-1|$. Another class of distances between measures can be obtained via optimal transportation theory. For general results about this theory, we refer to the books [@MR99k:28006; @MR99k:28007; @Villani:2007fk; @MR1964483]. Let $(\EE, \, \P)$ and $(\FF,\, \Q)$ two Polish spaces equipped with a probability measure and $c$ a semi-continuous function from $\EE\times \FF$ to $\R^+\cup \{\infty\}$. The optimal-transportation problem or Monge-Kantorovitch problem $\text{MKP}(\P,\Q)$ is to find $$\min_{\gamma \in \Sigma (\P,\Q)}\int_{\EE\times \FF} c(x,y)\xdif \gamma(x,y)$$ where $ \Sigma (\P,\Q)$ denoted the space of probability measures on $\EE\times \FF$ with first marginal $\P$ and second marginal $\Q$. Said otherwise in a more probabilistic way, it amounts to find the coupling between $\P$ and $\Q$ which minimizes the cost, i.e. to construct on the same probability space, two random variables $X$ and $Y$ of respective distribution $\P$ and $\Q$ which minimizes $\esp{}{c(X,Y)}$ among all the possible constructions. The usual cost functions are of the type $c(x,y)=\text{dist}(x,y)^p$ where $\text{dist}$ is a distance and $p$ a positive real number. For the Euclidean distance and $p=2$, we can construct the so-called Wasserstein distance by considering $$W(\P,\Q)=\sqrt{\min_{\gamma \in \Sigma (\P,\Q)}\int_{\R^d\times \R^d} \|x-y\|^2\xdif \gamma(x,y)}.$$ All the distances viewed so far are not unrelated as many functional inequalities do exist between all of them. Just to mention two examples, the Pinsker inequality states that the total variation distance is controlled by the Kullblack-Leibler distance. $$D_{|t-1|}(\P,\Q)\le \sqrt{\frac12 D_{t\ln t}(\P,\Q) }.$$ On the other hand, the so-called HWI identity (see [@MR1964483]) relates the relative entropy (H), the Wasserstein distance (W) and the Fischer information (I) as follows. \[sec:scope-steins-method-1\] Let $\P$ and $\Q$ two probability measures on $\R^n$ such that $\P=\exp(-V)\xdif x$ with $\nabla^2V\ge K\Id_n$. Then, $$D_{t\ln t}(\P,\Q)\le W(\P,\Q)\sqrt{D_{\nabla |\ln t|^2}(\P,\Q)}-\frac{K}{2} W(\P,\Q)^2.$$ These examples are here only to give a glimpse of the vast subject of the relationship between all these notions of distances. However, this is not the true subject of the present paper. The theorem which justifies the sequel is known as Kantorovitch-Rubinstein theorem (see [@MR982264; @MR622552]) and says the following. For $\P$ and $\Q$ two probability measures on a Polish space $\EE$, consider the Monge-Kantorovitch problem for a cost function $c$ which is a distance on $\EE$. Then, we have the following representation $$\min_{\gamma \in \Sigma (\P,\Q)}\int_{\EE\times \FF} c(x,\, y)\xdif \gamma(x,y)=\sup_{F \in {\operatorname{Lip}}_c(1)}\left(\esp{\P}{F}-\esp{\Q}{F}\right),$$ where $F \in {\operatorname{Lip}}_c(1)$ means that $F$ is $c$-Lipschitz continuous: $|F(x)-F(y)|\le c(x,y)$ for all $x,\, y\in E$. The resulting distance between $\P$ and $\Q$, will be called henceforth the Kantorovitch-Rubinstein distance as in [@MR1964483]. This formulation of a distance motivates alternative definitions by changing the set of test functions. For instance, for $\FFF=\{\car_{(-\infty;\, x]},\, x\in R\}$, $$\sup_{F \in \FFF}\left|\esp{\P}{F}-\esp{\Q}{F}\right|$$ is the total-variation distance. It turns out that Stein’s method is particularly well suited to estimate such kind of distances as we shall see now. Stein’s method {#sec:steins-method} ============== Historically, the Stein’s method for Gaussian distribution dates back to the seminal paper of Stein [@stein1972]. It was soon extended to the Poisson distribution in the paper of Chen [@MR0370693]. It is then impossible to track all the extensions of this approach, made mainly by A. Barbour and his collaborators, to several other distributions like compound Poisson [@MR1920275], Poisson point processes[@MR2001a:60058], stationary measure of birth-death process, even Brownian motion [@MR1035659]. For a whole account of all this period, one may refer to the books [@MR1708412; @MR93g:60043] and references therein. The main breakthrough came with the paper of Nourdin and Peccati [@MR2520122], in which it is shown that combining Malliavin calculus and Stein’s approach, one can obtain a rather simple proof of the striking *fourth moment theorem*, established earlier in [@MR2118863]. This was the starting point of a bunch of articles with with a wide area of applications: rate of convergence in the central limit theorem, Berry-Esseen theorem, iterated-logarithm theorem, limit theorems on manifolds, etc. Dirichlet-Malliavin structure {#sec:going-functional} ----------------------------- The procedure of the Stein’s method can be abstracted within the setting of Dirichlet structures (for details, we refer to [@bouleau-hirsch; @MR569058; @MR1214375]). The subsequent explanations are at a very formal level since the hard part for this machinery to work is to find the convenient functional spaces for each case of applications. The first idea underlying the Stein’s method is to characterize the target measure by an algebraic equation: Find a functional operator $L$ on $\mathcal F$ such that $\esp{\Q}{LF}=0$ for any $F$ in $\mathcal F$ if and only if $\Q=\P$. It turns out that this functional operator $L$ can be viewed as the (infinitesimal) generator of a Markovian semi-group, which we denote by $P=(P_t,\, t\ge 0)$ whose stationary measure is $\P$: The image measure of $\P$ by $P_t$ is still $\P$ for any $t\ge 0$. Under some technical hypothesis, there exists a strong ergodic Markov process $X=(X(t),\, t\ge 0)$ of invariant measure $\P$ and of generator $L$. It must be noted that the knowledge of one of $L$, $P$ or $X$ is equivalent to the knowledge of the other two. Formally speaking, for any $x\in E$, $$P_tf(x)=e^{tL}f(x), \ Lf(x)=\left.\frac{dP_tf(x)}{dt}\right|_{t=0}, \ P_tf(x)=\esp{}{f(X(t))\, |\, X(0)=x}.$$ One can also associate to $X$, the so-called Dirichlet form defined formally by $$\EEE(F,G)=\esp{\P}{LF\, G},$$ for any $F$ and $G$ sufficiently regular. As before, if we are given such a bilinear form $\EEE$, one can retrieve $L$ by the following relationship: For any $F$, $LF$ is the unique element $H$ such that for any $G$, $\EEE(F,\, G)=\esp{\P}{HG}$. This means that whichever of $L$, $X$, $P$ or $\EEE$ we are given, the others are uniquely determined (the reader is referred to the particularly illuminating Diagram 2, page 36 of [@MR1214375]). Within this framework, it is easy to see that the Stein-Dirichlet representation formula holds: For any bounded $F$, $$\label{eq_decreusefond_esaim:1} \esp{\Q}{F}-\esp{\P}{F}=\esp{\Q}{\int_0^\infty LP_tF \d t}.$$ This formula is also known as *the semi-group method* or *the smart-path formula* in the Stein’s method literature. This means that we can write $${\operatorname{dist}}_{\mathcal F}(\P,\Q)=\sup_{F\in \mathcal F}\left| \esp{\P}{F}-\esp{\Q}{F}\right|=\sup_{F\in \mathcal F}\left|\esp{\Q}{\int_0^\infty LP_tF \d t} \right|.$$ Instead of using coupling arguments to estimate this right-hand-side as usually done in the Stein’s method, we use another functional operator which is the *gradient* in the sense of Malliavin. It is usually denoted by $D$ and satisfies the identity $L=D^*D$ where $D^*$ is the adjoint of $D$. This a square root of the symmetric operator $L$, but not all square-roots are interesting as we also need a nice commutation relationship between $D$ and $P$. A few examples are the best way to illustrate what we mean. One dimensional examples {#sec:gaussian-measures} ------------------------ If $\P$ denote the standard Gaussian measure on $\R$, then $X$ is the Ornstein-Uhlenbeck process defined by $$dX(t)=\sqrt{2}\d B(t)-X(t)\d t, \ X(0)=x,$$ where $B$ is a standard one-dimensional Brownian motion. A straightforward application of the Itô formula gives the following expression of $X$: $$X(t)=e^{-t}x+\sqrt{2}\int_0^te^{-(t-s)}\d B(s).$$ It is then easy to see that $X(t)\sim\mathcal N(e^{-t}x,\, 1-e^{-2t})$, which, in turn, entails the Mehler representation formula: $$P_tF(x)=\int_\R F(e^{-t}x+\sqrt{1-e^{-2t}}y)\d\P(y).$$ It follows by differentiation and integration by parts that for $F\in \mathcal C^2_b$, $$LF(x)=xF^\prime(x)-F^{\prime\prime}(x), \text{ for all } x\in \R.$$ The *Malliavin* gradient is the usual derivative operator and standard computations show that $$\int_\R DF(x) \, G(x)\d\P(x)=\int_\R F(x)(xG(x)-DG(x))\d\P(x),$$ hence that $D^*G(x)=xG(x)-DG(x)$ and $L=D^*D$. Moreover, we have $DP_tF(x)=e^{-t}P_tDF(x)$ which is the commutation relationship alluded above. If $\P$ represents the Poisson measure on $\N$ of parameter $\lambda$, the process $X$ can be viewed as the number of occupied servers in an M/M/$\infty$ queue (see [@Decreusefond:2012sys]), $L$ is the corresponding generator: $$LF(x)=\lambda(F(x+1)-F(x))+x(F(x-1)-F(x)), \text{ for all } x\in \N,$$ with the convention that $0.F(-1)=0$. The *gradient* is defined by $$DF(x)=F(x+1)-F(x),$$ and we have $DP_tF=e^{-t}P_tDF$ (see [@Decreusefond:2012sys Theorem 11.16] or [@DST:functional]). For the scalar product in $L^2(\P)$, we have $$\label{eq_decreusefond_esaim:3} \int_\N DF(x)\, G(x)\d\P(x)=\int_\N F(x)(\frac{x}{\lambda}G(x-1)-G(x))\d \P(x).$$ Hence, $$D^*F(x)=\frac{x}{\lambda}G(x-1)-G(x) \text{ and } L=D^*D.$$ We now show how these constructions do articulate to give a new approach to the Stein’s method. It is well known that for $Z_\lambda$ a Poisson random variable of parameter $\lambda$, $$\hat{Z}_\lambda =\frac{Z_\lambda-\lambda}{\sqrt{\lambda}}\xrightarrow{\lambda \to \infty} \mathcal N(0,\ 1) \text{ in distribution.}$$ We are going to use the Stein-Dirichlet-Malliavin method to evaluate the rate of convergence. We are in a situation where the target measure in defined $\R$ whereas the initial randomness comes from a probability measure on $\N$. The map $T$ defined by $$\begin{aligned} T\, :\, \EE=\N&\longrightarrow \FF=\R\\ n&\longmapsto \frac{n-\lambda}{\sqrt{\lambda}},\end{aligned}$$ maps one space to the other and we are to evaluate the distance between $T^*\Q_\lambda$, the image measure of $\Q_\lambda$, the Poisson($\lambda$) probability, by the map $T$ and $\P$ the standard normal distribution on $\R$. This is a particular case of the general situation illustrated in Figure \[fig\_decreusefond\_esaim:comparaison\]. at (0,2) [Initial space]{}; at (5,2) [Target space]{}; (a) at (0,0) [$(\EE,\Q)$]{}; (b) at (5,0) [$(\FF,\P)$]{}; [(c) at (5,-3) [$(\mathfrak F,T^*\Q)$]{}; (a) edge node\[above\] [$T$]{} (c);]{} [at (0,-3) [$\text{dist}_{\mathcal F} (T^*\Q,\, \P)$ ?]{};]{} In view of , we have to estimate $$\sup_{F\in \mathcal F} \int_0^\infty \int_\R x.(P_tF)^\prime(x)-(P_tF)^{\prime\prime}(x) \d T^*\Q_\lambda(x) \d t,$$ where is the Ornstein-Uhlenbeck semi-group given by the Mehler formula above and $\mathcal F$ is a functional space to be conveniently chosen. According to the definition of $T$, the quantity to maximize is equal to $$\esp{}{\int_0^\infty \hat{Z}_\lambda.(P_tF)^\prime(\hat{Z}_\lambda)-(P_tF)^{\prime\prime}(\hat{Z}_\lambda)\d t}.$$ Applying to $G=1$ and $ F\circ T$, we get $$\sqrt{\lambda} \ \esp{}{F(\hat{Z}_\lambda+\frac{1}{\sqrt{\lambda}})-F(\hat{Z}_\lambda)}=\esp{}{\hat{Z}_\lambda\,F(\hat{Z}_\lambda)}.$$ Hence, $$\label{eq_decreusefond_esaim:5} \esp{}{\hat{Z}_\lambda.(P_tF)^\prime(\hat{Z}_\lambda)}=\sqrt{\lambda} \esp{}{(P_tF)^\prime(\hat{Z}_\lambda+\frac{1}{\sqrt{\lambda}})-(P_tF)^\prime(\hat{Z}_\lambda)}.$$ For any $t>0$, the regularizing properties of $P_t$ entails that $P_tF$ is thrice differentiable. Hence, $$\label{eq_decreusefond_esaim:6} (P_tF)^\prime(\hat{Z}_\lambda+\frac{1}{\sqrt{\lambda}})-(P_tF)^\prime(\hat{Z}_\lambda)= \frac{1}{\sqrt{\lambda}} (P_tF)^{\prime\prime}(\hat{Z}_\lambda)+\frac{1}{{\lambda}} \int_0^1 (1-r) (P_tF)^{(3)}(\hat{Z}_\lambda+\frac{r}{\sqrt{\lambda}})\d r.$$ And then, a miracle occurs: The term involving the second order derivative vanishes and we are lead to maximize $$\label{eq_decreusefond_esaim:4} \frac{1}{\sqrt{\lambda}}\ \esp{}{\int_0^\infty \int_0^1 (1-r) (P_tF)^{(3)}(\hat{Z}_\lambda+\frac{r}{\sqrt{\lambda}})\d r \d t}$$ for $F$ over $\mathcal F$. There is now a delicate point. If $F$ is in $\mathcal C^1_b$, we already mentioned that $$(P_tF)^\prime(x)=e^{-t}P_t(F^\prime)(x).$$ Furthermore, by integration by parts with respect to the Gaussian measure, it is easy to see that $$(P_tF)^{(k)}(x)=\left(\frac{e^{-t}}{\sqrt{1-e^{-2t}}}\right)^k\int_\R F(e^{-t}x+\sqrt{1-e^{-2t}}y)\, y^k\d\P(y),$$ whenever $F$ is bounded, for any $k\ge 1$. At first glance, it seems easy to bound by using the previous formula for $k=3$. Unfortunately, the term $\exp(-kt)(1-\exp(-2t))^{-k/2}$ is integrable over $[0,+\infty)$ only for $k=1$. Hence, we must choose $\mathcal F=\{F\in \mathcal C^2_b, \Vert F\Vert_{\mathcal C^2_b}\le 1\}$ and then we have $$\begin{gathered} \left|(P_tF)^{(3)}(x)\right|=\left|\frac{e^{-3t}}{\sqrt{1-e^{-2t}}} \int_\R F^{(2)}(e^{-t}x+\sqrt{1-e^{-2t}}y)\, y\d\P(y)\right|\\ \le \frac{e^{-3t}}{\sqrt{1-e^{-2t}}}\Vert F^{(2)}\Vert _\infty\ \int_\R |y|\d\P(y).\end{gathered}$$ Plugging this inequality into , we get $$\begin{gathered} \label{eq_decreusefond_esaim:8} \sup_{\Vert F\Vert_{\mathcal C^2_b}\le 1}\left| \esp{}{F(\hat{Z}_\lambda)}-\int F\d\P\right| \\ \le \frac{1}{\sqrt{\lambda}}\int_0^1 (1-r)\d r\ \int_0^\infty\frac{e^{-3t}}{\sqrt{1-e^{-2t}}}\ dt \ \int_\R |y|\d\P(y) =\frac{\sqrt{\pi}}{4\sqrt{2}}\ \frac{1}{\sqrt{\lambda}}\cdotp\end{gathered}$$ Hence we have established the rate of convergence for the Kantorovitch-Rubinstein distance associated to $\mathcal F=\{F\in \mathcal C^2_b, \Vert F\Vert_{\mathcal C^2_b}\le 1\}$. In dimension $1$, for Gaussian approximation, we could have used $LF(x)=xF(x)-F^\prime(x)$ as a characterizing operator and thus used only $1$-Lipschitz functions with a slightly different constant in front of the $\lambda^{-1}$ factor, namely $$\sup_{F\in {\operatorname{Lip}}(1)}\left| \esp{}{F(\hat{Z}_\lambda)}-\int F\d\P\right|\le \frac{1}{\sqrt{2\pi}}\ \frac{1}{\lambda}\cdotp$$ Note that this upper-bound is better than the bound obtained by the classical Stein’s method where $(2\pi)^{-1/2}$ is replaced by $1$. However, this line of thought is not applicable to higher dimensions. More generally, the recipe of the Stein-Dirichlet-Malliavin method is the following. - Characterize the target measure as the stationary distribution of an ergodic Markov process, - Construct the two Dirichlet-Malliavin structure on both initial and target spaces, - Perform an integration by parts on the initial space (see ), - Replace the gradient on the initial space by a function of the gradient on the target space (this is done here by the Taylor formula ), at the price of additional terms to be controlled, - Finish the computations in the target space using the commuting relationship : $DP_t=e^{-t}P_t D$. Higher dimensions {#sec:higher-dimensions} ----------------- This procedure can be generalized to any dimension provided that we have Dirichlet-Malliavin structures on both the initial and the target spaces. For the Gaussian measure in dimension $d$, the generator is given by $$\label{eq_decreusefond_esaim:2} LF(x)=x.D F(x)-\Delta F(x), \text{ for all } x\in \R^d,$$ where $D$ is the usual gradient in $\R^d$ and $\Delta$ is the Laplacian operator. The Mehler formula stays formally the same with an integral over $\R^d$ instead of $\R$ and $X$ is the $\R^d$-valued process composed of $d$ independent copies of the one dimensional Ornstein-Uhlenbeck process. The Malliavin gradient is still the usual gradient and the commutation relationship between $D$ and $P_t$ is easily seen to hold again. We can then retrieve the results of [@MR2727319]. Real difficulties arise when we try to generalize this approach to infinite dimensional spaces like the Wiener space. It is tempting to define $L$ formally as in , replacing the Laplacian by the trace of $D\circ D$. Unfortunately, for this trace term to exist, we need to restrict the space $\mathcal F$ of test functions and to choose conveniently the space $\FF$. There are actually two papers which address this problem. In both of them [@CD:2012; @Shih20111236], despite apparent dissimilarities, we end by considering $\FF$ a Hilbert space with a Gaussian measure. Let us show how it works on an example. For $N_\lambda$ a Poisson process on $\R^+$ of intensity $\lambda$, it is known that $$\hat{N}_\lambda(t)=\frac{N_\lambda(t)-\lambda t}{\sqrt{\lambda}}\xrightarrow{\lambda\to \infty} B(t) \text{ in distribution},$$ where $B$ is a standard Brownian motion and the convergence is understood to hold in $\mathbb D$, the Skorohod space of rcll functions. To compare the two distributions implies to find a common Hilbert space which supports both the distribution of $B$ and $\hat{N}_\lambda$. In principle, any Sobolev-like space should do. In [@CD:2012], we chose the so-called Besov-Liouville space $I^{\beta, 2}$ for $\beta<1/2$ defined by $$I^{\beta, 2}=\{f, \, \exists \dot f\in L^2([0,1]) \text{ such that } f(x)=\frac{1}{\Gamma(\beta)}\, \int_0^x (x-y)^{\beta-1}\dot f(y)\d y\}.$$ It is a Hilbert space when equipped with the scalar-product $\langle f,\, g\rangle_{\beta,2}=\langle \dot f,\, \dot g\rangle_{L^2}$. The Wiener measure on this space, denoted by $\mu_\beta$, is defined by $$\esp{\mu_\beta}{\exp(i\langle \eta,\, \omega\rangle_{{\beta,\, 2}})}=\exp(-\frac 12 \langle V_\beta\eta,\, \eta\rangle_{{\beta,\, 2}}).$$ where $$\begin{gathered} I_{0^+}^\beta f(x)=\frac{1}{\Gamma(\beta)}\, \int_0^x (x-y)^{\beta-1}\dot f(y)\d y,\ I_{1^-}^\beta f(x)=\frac{1}{\Gamma(\beta)}\, \int_x^1 (y-x)^{\beta-1}\dot f(y)\d y\\ \text{ and } V_\beta=I_{0^+}^\beta\circ I_{0^+}^{1-\beta}\circ I_{1^-}^{1-\beta}\circ I_{0^+}^{-\beta}.\end{gathered}$$ The Ornstein-Uhlenbeck semi-group on $( I^{\beta, 2},\, \mu_\beta)$ is defined for any $F\in L^2( I^{\beta, 2},\, \mu_\beta)$ by $$\begin{aligned} P_t^\beta F(u)&:=\int_{I^{\beta, 2}} F(e^{-t}u+\sqrt{1-e^{-2t}}\, v)\d \mu_\beta(v). \end{aligned}$$ The gradient is the Fréchet gradient on $I^{\beta, 2}$ and all the other properties still holds formally as in finite dimension. As initial space, we consider $\EE=\mathfrak N$, the space of locally finite configurations on $\R^+$ equipped with the vague topology. The measure $\mathbf Q_\lambda$ is such that the canonical process, denoted by $N_\lambda$, is a Poisson process of intensity $\lambda$, for details we refer to [@CD:2012]. On the initial space, we actually only need to know the gradient and an integration by parts formula. Here, we take $$D_xF(N_\lambda)=F(N_\lambda+\delta_x)-F(N_\lambda),$$ where $N_\lambda+\delta_x$ is the configuration $N_\lambda$ with an additional atom at location $x$. The well-known Campbell-Mecke formula ([@kallenberg83; @MR1113698]) is equivalent to say that $$\esp{\Q_\lambda}{F\ \int_0^1 G_\tau (\d N_\lambda(\tau)-\lambda \d \tau))}=\lambda\ \esp{\Q_\lambda}{\int_0^1 D_\tau F\ G_\tau\d \tau},$$ for $G$ a deterministic process. The map $T$ is defined by $$\begin{aligned} T\, :\, \, \mathfrak N& \longrightarrow I^{\beta,2}\\ N & \longmapsto (t\mapsto \frac{N(t)-\lambda t}{\sqrt{\lambda}})\cdotp\end{aligned}$$ Proceeding exactly along the same lines as before, one can show that there exists $c_\beta>0$ such that $$\label{eq_decreusefond_esaim:7} \sup_{\Vert F\Vert _{ \mathcal C^2_b(I^{\beta,2};\, \R)}\le 1}\left| \esp{\Q_\lambda}{F}-\esp{\mu_\beta}{F}\right| \le \frac{c_\beta}{\sqrt{\lambda}},$$ where $ \mathcal C^2_b(I^{\beta,2};\, \R)$ is the set of twice Fréchet differentiable functionals on $I^{\beta,2}$, with bounded differentials. This is the generalization we could expect of . Other examples of the application of this procedure, involving other functional spaces, can be found in the papers [@CD:2012; @DST:functional]. A similar approach with Malliavin calculus replaced by a coupling argument appears in [@schumacher14]. Edgeworth expansion {#sec:edgeworth-expansion} =================== The Stein’s method as developed here can be iterated to obtain Edgeworth expansions. We now want to precise the expansion obtained in . For, we go one step further in the Taylor formula : $$\psi(\hat{Z}_\lambda+1/\sqrt{\lambda})-\psi(\hat{Z}_\lambda)=\frac{1}{\sqrt{\lambda}}\psi^\prime(\hat{Z}_\lambda)+\frac{1}{2\lambda}\psi^{\prime\prime}(\hat{Z}_\lambda)+\frac{1}{6\lambda^{3/2}}\psi^{(3)}(\hat{Z}+\theta/\sqrt{\lambda}).$$ Hence, $$\begin{gathered} \label{eq_gaussapp20:4} \esp{}{ \hat{Z}_\lambda DP_tF(\hat{Z}_\lambda)-D^{(2)}P_tF(\hat{Z}_\lambda)}\\ = \frac{1}{2\sqrt{\lambda}}\esp{}{D^{(3)}P_tF(\hat{Z}_\lambda)}+\frac{1}{6\lambda}\esp{}{D^{(4)}P_tF(\hat{Z}+\theta/\sqrt{\lambda})}.\end{gathered}$$ If $F$ is thrice differentiable with bounded derivatives then $P_tF$ is four times differentiable, hence the last term of is bounded by $\lambda^{-1}\frac{e^{-4t}}{\sqrt{1-e^{-2t}}}\Vert F^{(3)}\Vert_\infty/6$. Moreover, applying to $DP_tF$ shows that $$\esp{}{D^{(3)}P_tF(\hat{Z}_\lambda)}= \esp{\P}{D^{(3)}P_tF}+O(\lambda^{-1/2}).$$ Combining the last two results, we obtain that for $F$ thrice differentiable $$\esp{}{F(\hat{Z}_\lambda)}-\esp{\P}{F}=\frac{1}{2\sqrt{\lambda}}\esp{\P}{\int_0^\infty D^{(2)}P_tF\d t} +O(\lambda^{-1}).$$ This line of thought can be pursued at any order provided that $F$ is assumed to have sufficient regularity and we get an Edgeworth expansion up to any power of $\lambda^{-1/2}$. Using the properties of Hermite polynomials, this leads to the expansion: $$\esp{}{F(\hat{Z}_\lambda)}-\esp{\P}{F} =\frac{1}{6\sqrt{\lambda}}\esp{\P}{F {{\mathcal H}}_3}+O(\lambda^{-1}),$$ where ${{\mathcal H}}_n$ is the $n$-th Hermite polynomials. In [@CD:2014], we generalized this approach to the Poisson process-Brownian motion convergence established in . Conclusion {#sec:conclusion} ========== We showed how the Stein’s method can be abstracted in the framework of Dirichlet forms and Malliavin calculus. This gives raise to a new method of proof which can be applied to infinite dimensional spaces and iterated to get Edgeworth expansions. One open question is to apply this approach to other limiting processes like stable or max-stable processes, Brownian bridges, etc. [10]{} D. Bakry, I. Gentil, and M. Ledoux. , volume 348 of [*Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\]*]{}. Springer, Cham, 2014. A. D. Barbour. Stein’s method for diffusion approximations. , 84(3):297–322, 1990. A. D. Barbour, T. C. Brown, and A. Xia. Point processes in time and [S]{}tein’s method. , 65(1-2):127–151, 1998. A. D. Barbour, L. Holst, and S. Janson. , volume 2 of [*Oxford Studies in Probability*]{}. The Clarendon Press Oxford University Press, 1992. Oxford Science Publications. A. D. Barbour and M. M[aa]{}nsson. Compound [P]{}oisson process approximation. , 30(3):1492–1537, 2002. N. Bouleau and F. Hirsch. . De Gruyter Studies in Mathematics, 1992. L. H. Y. Chen. On the convergence of [P]{}oisson binomial to [P]{}oisson distributions. , 2(1):178–180, 1974. L. [Coutin]{} and L. [Decreusefond]{}. tein’s method for [B]{}rownian approximations. , 7(5):349–372, 2013. L. [Coutin]{} and L. [Decreusefond]{}. Higher order expansions via [S]{}tein’s method. , May 2014. Schumacher D. and Stucki K. Gibbs point process approximation: total variation bounds using [S]{}tein’s method. , 42(5):1911–1951, 2014. L. Decreusefond and P. Moyal. . ISTE Ltd and John Wiley & Sons Inc, 2012. L. [Decreusefond]{}, M. [Schulte]{}, and C. [Thäle]{}. Functional [P]{}oisson approximation in [R]{}ubinstein distance. , to appear, 2015. R. M. Dudley. . The Wadsworth & Brooks/Cole Mathematics Series. Wadsworth & Brooks/Cole Advanced Books & Software, Pacific Grove, CA, 1989. X. Fernique. Sur le th[é]{}or[è]{}me de [K]{}antorovitch-[R]{}ubinstein dans les espaces polonais. In [*Seminar on Probability, XV (Univ. Strasbourg, Strasbourg, 1979/1980) (French)*]{}, volume 850 of [*Lecture Notes in Math.*]{}, pages 6–10. Springer, Berlin, 1981. D. Feyel and A. S. [Ü]{}st[ü]{}nel. Monge-[K]{}antorovitch measure transportation and [M]{}onge-[A]{}mpère equation on [W]{}iener space. , 128(3):347–385, 2004. M. Fukushima. , volume 23 of [ *North-Holland Mathematical Library*]{}. North-Holland Publishing Co., Amsterdam, 1980. A. L. Gibbs and F. E. Su. On choosing and bounding probability metrics. , 70(3):419–435, 2002. O. Kallenberg. . Academic Press, 3rd edition, 1983. A. F. Karr. , volume 7 of [*Probability: Pure and Applied*]{}. Marcel Dekker Inc., New York, second edition, 1991. Z. M. Ma and M. R[ö]{}ckner. . Universitext. Springer-Verlag, Berlin, 1992. I. Nourdin and G; Peccati. Stein’s method on [W]{}iener chaos. , 145(1-2):75–118, 2009. David Nualart and Giovanni Peccati. Central limit theorems for sequences of multiple stochastic integrals. , 33(1):177–193, 2005. Giovanni Peccati and Cengbo Zheng. Multi-dimensional [G]{}aussian fluctuations on the [P]{}oisson space. , 15:no. 48, 1487–1527, 2010. S. T. Rachev and L. R[ü]{}schendorf. . Probability and its Applications (New York). Springer-Verlag, New York, 1998. Theory. S. T. Rachev and L. R[ü]{}schendorf. . Probability and its Applications (New York). Springer-Verlag, New York, 1998. Applications. H.-H. Shih. On [S]{}tein’s method for infinite-dimensional [G]{}aussian approximation in abstract [W]{}iener spaces. , 261(5):1236 – 1283, 2011. Charles Stein. A bound for the error in the normal approximation to the distribution of a sum of dependent random variables, 1972. C. Villani. , volume 58 of [*Graduate Studies in Mathematics*]{}. American Mathematical Society, Providence, RI, 2003. C. Villani. . Lectures Notes in Mathematics. Springer Verlag, New York, 2007. A. Xia. Poisson approximation, compensators and coupling. , 18(1):159–177, 2000.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We construct high-precision models of the Universe that contain radiation, a cosmological constant, and periodically distributed inhomogeneous matter. The density contrasts in these models are allowed to be highly non-linear, and the cosmological expansion is treated as an emergent phenomenon. This is achieved by employing a generalised version of the post-Newtonian formalism, and by joining together inhomogeneous regions of space-time at reflection symmetric junctions. Using these models, we find general expressions that precisely and unambiguously quantify the effect of small-scale inhomogeneity on the large-scale expansion of space (an effect referred to as “back-reaction", in the literature). We then proceed to specialize our models to the case where the matter fields are given by a regular array of point-like particles. This allows us to derive extremely simple expressions for the emergent Friedmann-like equations that govern the large-scale expansion of space. It is found that the presence of radiation tends to reduce the magnitude of back-reaction effects, while the existence of a cosmological constant has only a negligible effect.' author: - 'Viraj A. A. Sanghai' - Timothy Clifton title: Cosmological backreaction in the presence of radiation and a cosmological constant --- Introduction ============ In previous work we developed a new formalism for constructing cosmological models with a periodic lattice structure [@vaas; @Tim1]. This was done by taking regions of space-time that we described using the post-Newtonian perturbative expansion, and patching them together at reflection symmetric boundaries to form a global solution to Einstein’s equations. The advantages of this approach are (i) that it allows extremely large density contrasts to be consistently included in cosmology, at higher orders in perturbation theory, without the imposition of any continuous symmetries (i.e. Killing vectors), and (ii) that it allows the cosmological expansion to be viewed as an emergent phenomenon, resulting from the junction conditions between patches [@Is1], rather than being specified from the outset. These two features, taken together, make our lattice models ideally suited for studying the effects that non-linear structure has on the large-scale expansion of space. Such effects, usually referred to as “back-reaction" in the cosmology literature [@br1; @br2; @br3], are important to understand if we are to have faith that the homogeneous and isotropic Friedmann-Lemaître-Robertson-Walker (FLRW) models are suitable for interpreting observations in a lumpy universe (such as the one within which we live). They may also be important for the much heralded era of “precision cosmology" [@euclid; @ska], especially if observations become good enough to isolate higher-order relativistic effects. However, while they may constitute interesting devices for studying back-reaction, and while they can help to illustrate the complementary nature of cosmology and weak-field gravity, the lattice models constructed in [@vaas; @Tim1] are not fully realistic. One way in which this situation can be improved upon, and on which we focus in this paper, is by adding other types of matter fields, beyond the non-relativistic matter that is usually included in studies of post-Newtonian gravity. In this regard, particular matter fields that are of interest in cosmology are radiation, and the cosmological constant, $\Lambda$. The former of these becomes increasingly important at early times, while the latter (if it is non-zero) comes to dominate the expansion at late times. In this paper we extend the post-Newtonian formalism by including the contribution of barotropic fluids with non-vanishing pressure, $p=p(\rho)$, to the energy-momentum tensor. Such an approach can be used to include a fluid of radiation, with $p=\frac{1}{3} \rho$, or a cosmological constant, with $p=-\rho$. It could also be used to include a variety of other matter fields that are commonly considered in cosmology. We then use this extended formalism to model the gravitational fields that exist within each of our lattice cells, and proceed to determine (lengthy) general expressions for the effect that such fluids have on the large-scale expansion of space. This is done in full generality, without assuming anything about the distribution of matter within each cell. In order to develop these ideas further we then specialize the distribution of matter to a particularly simple example: a single point-like mass at the centre of each cell, in the presence of radiation and $\Lambda$. Globally, this corresponds to a regular array of massive particles sitting in a sea of radiation. The result of considering this specific set-up is an expression for cosmological back-reaction that takes an extremely simple form. Its effect on the Friedmann equation is to add an extra term, so that we have $$\begin{aligned} \left( \frac{\dot{a}}{a} \right)^2 =& \frac{8\pi G}{3}( \rho_{M} + \rho_{r}) -\frac{k}{a^2} +\frac{\Lambda}{3} +\mathcal{B} \ , \nonumber\end{aligned}$$ where the back-reaction term, $\mathcal{B}$, is given by $$\begin{aligned} \mathcal{B} &\simeq -\left( 2\pi G L \rho_{M} a \right)^2 \left(1.50 - 0.80 \frac{\Omega_{r}}{\Omega_{M}} + 1.76 \frac{\Omega_{k}}{\Omega_{M}} \right) \, , \nonumber\end{aligned}$$ where $\Omega_{M}$, $\Omega_{r}$ and $\Omega_{k}$ are the standard cosmological density parameters for matter, radiation and spatial curvature, respectively, and where $L$ is the length of the edge of a cell (see Section \[results\] for details). It can be seen that the discretely distributed matter contributes a term that looks like radiation to the effective Friedmann-like equation that governs the large-scale expansion of space, just as was found in [@vaas]. We find that presence of radiation, however, reduces the magnitude of this back-reaction term, while the presence of $\Lambda$ has no noticeable effect on it at all. As shown in [@vaas], a negative value of spatial curvature increases the amplitude of the back-reaction, and a positive value decreases it. The physical set-up that we consider in the latter parts of this paper, consisting of a universe full of point sources, has received considerable attention over the past few years. This includes studies of the initial data of such models [@crt; @yoo1; @ben1; @kor1; @kor2], as well as their evolution [@ben2; @ben3; @yoo2; @numlam; @cgr; @cgr2; @hinder]. Studies of back-reaction in the presence of radiation and $\Lambda$ have also been performed using both perturbative methods [@lam; @radlam1; @radlam2], and by solving the full Einstein equations [@numlam; @numrad]. Our work is complementary to these previous studies. It builds on them by developing and applying a versatile perturbative framework that incorporates non-linear density contrasts, while avoiding the ambiguities that can arise when averaging in general relativity. The plan for the rest of this paper is as follows: In section \[sec2\] we set out the equations that describe the geometry and dynamics of our lattice cells. In section \[sec3\] we use these equations to determine the cosmological expansion of our lattice, in the presence of an arbitrary barotropic fluid, and for any general distribution of matter. In section \[sec4\] we then look at the specific case of regularly arranged point masses in cubic cells in the presence of radiation, spatial curvature and a cosmological constant. Throughout the paper we use latin letters ($a$, $b$, $c$, ...) to denote space-time indices, and greek letters ($\mu$, $\nu$, $\rho$, ...) to denote spatial indices. We reserve the first half of the capital latin alphabet ($A$, $B$, $C$, ...) to denote the spatial components of tensors in the boundary of a cell, and the latter half ($I$, $J$, $K$, ...) as labels to denote quantities associated with our various different matter fields. The geometry of a lattice cell {#sec2} ============================== In this section we present the equations that describe the geometry within each of our lattice cells, and the dynamics of their boundaries. We begin by briefly recapping the set-up of our bottom-up approach to cosmology, before moving on to discuss how we extend the post-Newtonian formalism to include a barotropic fluid, as well as non-relativistic matter. After this, we make use of reflection symmetric junction conditions to find the evolution of the boundary of every cell. Altogether, this gives us just enough information to work out the expansion of each of our cells, and hence the lattice as a whole, to the first post-Newtonian level of accuracy. Lattice structure ----------------- We begin by splitting the Universe into a large number of identical cells, in order to construct a periodic lattice structure. When doing this, we allow the cell shapes to be chosen as any regular convex polyhedra that tessellates a three-dimensional space that is either flat or has constant positive or negative spatial curvature. There are six such tessellations for a space of positive curvature, one for a flat space, and four for a space of negative curvature (see Table I of [@vaas], and reference [@poly], for details). The analysis we present in this section is valid for any cell shape, and for any of these possible tessellations. Let us now consider any one cell, and take, as a first approximation, the space-time within this cell to be close to Minkowski space. We can then choose a Cartesian set of spatial coordinates ($x$, $y$, $z$), and rotate these coordinates until the vector $\partial/ \partial x$ is orthogonal to one of the faces (as all cell faces are identical, it does not matter which one we choose). This situation is illustrated in Fig. \[fig1\], for the case of a dodecahedral cell. The position of every point on this cell face is then given by $x = X(t,y,z)$. Now, as the evolution of this cell face is a $(2+1)$-dimensional time-like hypersurface, we can define a space-like unit vector, $n^{a}$, as its normal. The leading-order contributions to the covariant components of this vector are given by [@vaas] $$n_a = \left( -X_{,t}, 1, -X_{,y}, -X_{,z} \right) \, ,$$ where commas denote partial differentiation. Reflection symmetry implies that junction conditions at the cell face should be invariant under the change $n^a \rightarrow - n^a$, which means that the extrinsic curvature of every $(2+1)$-dimensional cell face must vanish. This requirement provides the information necessary for specifying the boundary conditions for the field equations within each cell, as well as for the motion of the cell face itself. Post-Newtonian expansion ------------------------ The matter content and geometry within each of our cells is described using the post-Newtonian perturbative expansion. This formalism, valid in the limit of weak gravitational fields and slow motions, assigns orders of smallness to quantities in the metric and the energy-momentum tensor using the parameter $$\label{2} \epsilon \equiv \frac{|\bm{v}|}{c} \ll 1 \, ,$$ where $\bm{v}$ is the three-velocity associated with the matter fields, and $c$ is the speed of light. The post-Newtonian expansion also requires that time derivatives add an order of smallness, compared to space derivatives, so that $X_{,t} /X_{,y} \sim O(\epsilon)$, [*etc*]{}. The explicit expansion of the metric is then given by the following line element: $$\begin{aligned} ds^2 = \left( -1+ h^{(2)}_{tt} + h^{(4)}_{tt} \right) dt^2 &+ 2 h^{(3)}_{t\mu} dt dx^{\mu} \nonumber \\ &\hspace{-10pt} + \left( \delta_{\mu\nu}+ h^{(2)}_{\mu\nu} \right) dx^{\mu} dx^{\nu} \label{metric} \, ,\end{aligned}$$ where $h^{(2)}_{tt}$, $h^{(2)}_{\mu\nu}$, $h^{(3)}_{t\mu}$ and $h^{(4)}_{tt}$ are perturbations to the Minkowski metric, and where superscripts in brackets represent the order of smallness of a quantity. The metric has been expanded to $O(\epsilon^4)$ in the time-time component, to $O(\epsilon^3)$ in the time-space components, and $O(\epsilon^2)$ in the space-space components. These are the orders of accuracy required in order to consistently write the equations of motion for time-like particles to first post-Newtonian accuracy. We can similarly expand the matter fields in powers of $\epsilon$. To do so, we define the energy density, $\rho$, and isotropic pressure, $p$, as $$\begin{aligned} \rho =& T_{ab} u^{a} u^{b} \, , \\ p =& \frac{1}{3}T_{ab} (g^{ab} + u^{a} u^{b}) \, , \end{aligned}$$ where $T_{ab}$ is the energy-momentum tensor, $g_{ab}$ is the metric of space-time, and $u^{a}$ is a reference four-velocity that satisfies $u^{a}u_{a} = -1$. We can expand the energy density and pressure as $$\begin{aligned} \rho =& \rho^{(2)} +\rho^{(4)}+ O(\epsilon^6) \, , \\ p =& p^{(2)} + p^{(4)} + O(\epsilon^6) \, , \label{denpress}\end{aligned}$$ and write the expanded four-velocity as $$\begin{aligned} u^{a} =& \bigg(1 +\frac{h^{(2)}_{tt}}{2}+ \frac{v^2}{2}\bigg)(1;v^{\mu}) + O(\epsilon^4) \, ,\end{aligned}$$ where $v$ is the three-velocity of the fluid we are considering, and $v^2= v^{\mu}v_{\mu}$. The reader may note that we have included a contribution to the pressure at $O(\epsilon^2)$, which is usually taken to vanish in post-Newtonian gravity. We have done this in order to include barotropic fluids, which generally have the leading-order contribution to pressure at the same order as energy density. For further details of post-Newtonian expansions, the reader is referred to [@will]. Matter content {#conserve} -------------- Let us now consider the matter content of our space-time. We wish to model a universe that contains both non-relativistic matter, with $p^{(2)}=0$, and a barotropic fluid, with equation of state $p=p(\rho)$. For simplicity, and as a first approximation, we will take the latter of these to be a perfect fluid that does not strongly interact with the non-relativistic matter. Such a fluid could be used to model radiation ($p=\frac{1}{3} \rho$), vacuum energy ($p=-\rho$), or a massless scalar field ($p=\rho$). The non-relativistic matter is intended to represent both baryonic matter and cold dark matter. We therefore write the total energy-momentum tensor for these two fluids as $$\begin{aligned} T^{ab} = T^{ab}_M + T^{ab}_I\, ,\end{aligned}$$ where subscripts $M$ and $I$ refer to quantities associated with the non-relativistic matter fields and the barotropic fluid, respectively. In what follows, the cosmological constant, $\Lambda$, is included directly in the field equations. If we now take the reference four-vector for each of the fluids to be given by $$\begin{aligned} u_{M}^{a} =& \bigg(1 +\frac{h^{(2)}_{tt}}{2}+ \frac{v_{M}^2}{2}\bigg)(1;v_{M}^{\mu}) + O(\epsilon^4) \, , \nonumber\\ u_{I}^{a} =& \bigg(1 +\frac{h^{(2)}_{tt}}{2}+ \frac{v_{I}^2}{2}\bigg)(1;v_{I}^{\mu}) + O(\epsilon^4) \, , \label{4-vel}\end{aligned}$$ where $v_M$ and $v_I$ are the three-velocities of our two fluids, then we can write the components of the perturbed energy-momentum tensor as $$\begin{aligned} T_{tt}=& \rho^{(2)} (1 - h^{(2)}_{tt}) + \rho^{(2)}_{I} v_{I}^2 + \rho^{(2)}_{M}v_{M}^2 \nonumber\\ & + \rho^{(2)}_{M} \Pi_{M} + \rho^{(4)}_{I} + p^{(2)}_{I} v_{I}^2 + O(\epsilon^6) \label{emtt} \, , \\ \nonumber\\ T_{t\mu} =& - \rho^{(2)}_{M} v_{M\mu} -( \rho^{(2)}_{I} + p^{(2)}_{I}) v_{I\mu} + O(\epsilon^5) \label{emtx} \, , \\ \nonumber\\ T_{\mu \nu} =& \rho^{(2)}_{M} v_{M\mu} v_{M\nu} + ( \rho^{(2)}_{I} + p^{(2)}_{I}) v_{I\mu} v_{I\nu} \nonumber \\ & + (p^{(4)}_{M} + p^{(2)}_{I} + p^{(4)}_{I}) g_{\mu\nu} + O(\epsilon^6) \, , \label{em}\end{aligned}$$ where $\rho^{(2)} = \rho^{(2)}_{M} + \rho^{(2)}_{I}$, and where $\rho^{(2)}_{M}$ is the rest-mass energy density of the non-relativistic matter fields, $\Pi_{M}$ is their specific energy density, and $p^{(4)}_{M}$ is their pressure. Similarly, ${\rho^{(2)}_{I}}$ and ${\rho^{(4)}_{I}}$ are the two lowest-order parts of the energy density of the barotropic fluid, and $p^{(2)}_{I}$ and $p^{(4)}_{I}$ are the two lowest-order contributions to its pressure. The reader may note the we have set $p^{(2)}_M=0$ for the non-relativistic matter fields, as we want this to represent dust-like sources such as galaxies and clusters. Before considering Einstein’s equations, we note that we can use the energy-momentum conservation equations for the non-interacting barotropic fluid to write $$\begin{aligned} \nabla p^{(2)}_{I} = 0 \, .\label{emcon1}\end{aligned}$$ This is the leading-order part of the Euler equation of the barotropic fluid, and it immediately implies that both $p^{(2)}_{I}$ and $\rho^{(2)}_{I}$ must be functions of time only \[as $p=p(\rho)$, for this fluid\]. It also means that the leading-order part of the continuity equation for the barotropic fluid, which also follows directly from energy-momentum conservation, is given by $$\begin{aligned} \rho^{(2)}_{I,t} + (\rho^{(2)}_{I} + p^{(2)}_{I}) \nabla \cdot {\bm v}_{I} = 0 \, . \label{emcon2}\end{aligned}$$ This is very similar to the conservation equation for a homogeneous fluid in FLRW models, and we later use it in the same way as that equation to determine the cosmological evolution. Einstein’s field equations {#sec2a} -------------------------- In order to find the geometry of the space-time within each cell, and to solve for the motion of its boundary, we need to use Einstein’s field equations, $$\begin{aligned} R_{ab} &= 8\pi G \left( T_{ab} - \frac{1}{2} T g_{ab} \right) + g_{ab}\Lambda \, , \label{Riccifield}\end{aligned}$$ where $R_{ab}$ is the Ricci tensor, $g_{ab}$ is the metric of space-time, $\Lambda$ is the cosmological constant, $G$ is Newton’s constant, $T_{ab}$ is the energy-momentum tensor, and $T= g^{ab} T_{ab}$ is its trace. Using the perturbed metric given in Eq. , and the energy-momentum tensor from Eq. , we can write the leading-order contributions to the $tt$-component of Einstein’s equations as $$\begin{aligned} \nabla^2 h^{(2)}_{tt} = - 8\pi G\rho^{(2)} -24\pi G p^{(2)}_{I} + 2\Lambda \label{nablaphi} \, ,\end{aligned}$$ where $\nabla^2 = \partial_{\alpha} \partial_{\alpha}$ is the three-dimensional Laplacian. Here we have taken the cosmological constant $\Lambda$ to contribute at $O(\epsilon^2)$, which means we are modelling a scenario where $\Lambda \sim \rho^{(2)} \sim h^{(2)}_{tt}$. This happens on scales of about $100$ Mpc, where the cosmological constant is comparable to the background gravitational potential. This is still well below the cosmological horizon scale, where our post-Newtonian formalism is satisfied. The solution to Eq. can be formally written as $$\begin{aligned} h^{(2)}_{tt} \equiv 2\Phi = 2\Phi_{M} + 2\Phi_{I} +6\Phi_{p}+ 2\Phi_{\Lambda} \, , \label{phi} $$ where the potentials $\Phi_{M}$, $\Phi_{I}$, $\Phi_{p}$ and $\Phi_{\Lambda}$ are given implicitly as the solutions to $$\begin{aligned} \nabla^2 \Phi_{M} \equiv& - 4 \pi G \rho^{(2)}_{M} \, , \label{phim} \\ \nabla^2 \Phi_{I} \equiv& - 4 \pi G {\rho^{(2)}_{I}} \, , \label{phii} \\ \nabla^2 \Phi_{p} \equiv& - 4 \pi G p^{(2)}_{I} \, , \\ \nabla^2 \Phi_{\Lambda} \equiv& \Lambda \, . \label{newpots}\end{aligned}$$ Using the symmetries of our lattice model, and the fact that $p^{(2)}_{I}$ is a function of time only, the potentials $\Phi_{p_{I}}$ and $\Phi_{\Lambda}$ can be written explicitly as $$\begin{aligned} \Phi_{p_{I}}=& -\frac{2\pi G p^{(2)}_{I}}{3}(x^2 + y^2 + z^2) \ , \\ \Phi_{\Lambda}=& \frac{\Lambda}{6}(x^2 + y^2 + z^2) \, . \label{plamsolns}\end{aligned}$$ Solutions to Eqs. (\[phim\]) and (\[phii\]) can be given in terms of Green’s functions, as shown in [@vaas]. Auxiliary functions of time can also be added in $h_{tt}^{(2)}$, and absorbed into the matter potential, $\Phi_{M}$. To go further, we now need to make a gauge choice. We make the following choice at $O(\epsilon^2)$, so that we remain as close as possible to the standard post-Newtonian gauge, $$\begin{aligned} & \frac{1}{2} h^{(2)}_{tt,\mu} + h^{(2)}_{\mu\nu, \nu} - \frac{1}{2} h^{(2)}_{\nu \nu,\mu} = 3 \Phi_{p , \mu} +\frac{3}{2} \Phi_{\Lambda, \mu} \ .\label{gauge1}$$ This ensures that the metric is diagonal at $O(\epsilon^2)$, and there are no $O(\epsilon)$ contributions to the $t\mu$-component of the metric. Using Eqs. and , the $\mu\nu$-component of Einstein’s equations can now be written as $$\begin{aligned} \nabla^2 h^{(2)}_{\mu\nu} = - (8\pi G \rho^{(2)} +\Lambda)\delta_{\mu\nu} \, . \label{nablapsi}\end{aligned}$$ The solution to this equation is given by $$\begin{aligned} h^{(2)}_{\mu\nu} \equiv 2 \Psi \delta_{\mu\nu} = (2\Phi_{M} + 2\Phi_{I} - \Phi_{\Lambda})\delta_{\mu\nu} \, .\label{psi} $$ The reader may note that in this formalism we have $\Psi \neq \Phi$ in the presence of either a cosmological constant or a barotropic fluid (or both). This differs from the case of cosmological perturbation theory, where $\Phi = \Psi$ in the absence of anisotropic stress. To solve for the ${t\mu}$-component of Einstein’s equations, we now need to make a gauge choice at $O(\epsilon^3)$, which we do as follows: $$\begin{aligned} & h^{(3)}_{\nu t, \nu} - \frac{1}{2} h^{(2)}_{\nu\nu,t} = 0 \, . \label{gauge2}\end{aligned}$$ Using both of our gauge conditions, Eqs. and , the $t\mu$-component of Einstein’s equations can be written as $$\begin{aligned} & \nabla^2 h^{(3)}_{t\mu} +\Psi_{,t\mu} = 16\pi G \left[ \rho^{(2)}_{M} v_{M\mu} + (\rho^{(2)}_{I} + p^{(2)}_{I}) v_{I\mu} \right]\, . \label{thirdorderpert}\end{aligned}$$ The solution to this equation is given by $$\begin{aligned} h^{(3)}_{t\mu} = -4V_{M\mu} - 4V_{I\mu} + \frac{1}{2}\chi_{,t\mu} \, , \label{thirdsoln}\end{aligned}$$ where we have used the two vector potentials $$\begin{aligned} \nabla^2 V_{M\mu} \equiv& -4\pi \rho^{(2)}_{M} v_{M\mu} \, , \\ \nabla^2 V_{I\mu} \equiv& - 4\pi G (\rho^{(2)}_{I} + p^{(2)}_{I}) v_{I\mu} \, ,\end{aligned}$$ and the superpotential $$\begin{aligned} \nabla^2 \chi \equiv -2\Psi \, .\end{aligned}$$ The gauge conditions imply that the divergence of these vector potentials must obey $V_{M\mu,\mu} + V_{I\mu, \mu} = - \Psi_{,t}$. Finally, we can write the $O(\epsilon^4)$ part of the tt-component of Einstein’s equations. Using the energy-momentum tensor from Eq. , both our gauge conditions, and the lower-order solutions for $h_{tt}^{(2)}$, $h_{\mu \nu}^{(2)}$ and $h_{t\mu}^{(3)}$, this equation becomes $$\begin{aligned} \nabla^2 h_{tt}^{(4)} = & -2\nabla(\Phi \nabla \Phi) - \nabla (\Psi \nabla \Phi + \Phi \nabla \Psi) +4\pi G\rho^{(2)}\Phi +24\pi G p^{(2)}_{I}\Phi - \frac{5}{2}\Lambda\Phi -20\pi G\rho^{(2)}\Psi -60\pi G p^{(2)}_{I}\Psi \label{nablasqhtt4} \\ &\quad + 5\Lambda\Psi - 16\pi G\rho^{(2)}_{M}v_{M}^2 - 16\pi G\rho^{(2)}_{I}v_{I}^2 -8\pi G\rho^{(2)}_{M}\Pi_{M} -8\pi G\rho^{(4)}_{I} -16 \pi G p^{(2)}_{I}v_{I}^2 -24\pi G p^{(4)}_{M} -24\pi G p^{(4)}_{I} \, . \nonumber\end{aligned}$$ These equations can also be solved using the Green’s functions from [@vaas]. Once this has been done, and the distribution of matter has been specified, this gives us sufficient information to find the geometry of each of our lattice cells, to post-Newtonian order of accuracy. Nowhere in this analysis have we assumed asymptotic flatness, as is conventionally done when applying the post-Newtonian formalism to the case of isolated systems. Instead, we have a system of equations that can be directly applied to solve for the gravitational fields of astrophysical bodies in a cosmological setting. Cosmological Expansion {#sec3} ====================== In this section, we derive the acceleration and constraint equations for the boundary of each of our cells, up to the first post-Newtonian level of accuracy. Due to the periodicity of our lattice models, these equations will also describe the large-scale expansion of the Universe as a whole. At Newtonian order, these equations take exactly the same form as the acceleration and constraint equations of a FLRW universe containing dust, a barotropic fluid, spatial curvature and a cosmological constant. At first post-Newtonian order, we obtain the leading-order corrections to these equations in a lattice universe. Using reflection symmetric boundary conditions, as we do in this study, implies that the extrinsic curvature of each of the $(2+1)$-dimensional boundaries of every cell must vanish (see [@vaas] for details). This condition leads directly to the equation of motion of the cell boundary, which to post-Newtonian accuracy can be written as follows: $$\begin{aligned} X_{,tt} =& \bigg[ \Phi_{,x} - 2\Psi\Phi_{,x}+ \frac{h_{tt,x}^{(4)}}{2} - h_{tx,t} -(2\Phi_{,x} + \Psi_{,x}) X_{,t}^{2} \nonumber \\ &\quad -( 2\Psi_{,t}+\Phi_{,t}) X_{,t} - X^{(2)}_{,A} \Phi_{,A}\bigg] \bigg|_{x=X} + O(\epsilon^6) \, , \label{X1} \end{aligned}$$ which can also be derived from the geodesic equation. Likewise, we obtain a set of equations that describes the spatial curvature of the cell boundaries, and their rate of change, as $$\begin{aligned} X_{,AB} =& \delta_{AB} ( \Psi_{,x})|_{x=X} + O(\epsilon^4) \, , \label{X3} $$ and $$\begin{aligned} X_{,tA} =& \frac{1}{2} \bigg[ h_{tA,x} - h_{tx,A} - 2(\Phi_{,A} +\Psi_{,A}) X_{,t}\bigg]\bigg|_{x=X} + O(\epsilon^5) \, . \label{X2} \end{aligned}$$ Each of the quantities in these equations must be evaluated on the boundary of the cell. Together, they give us enough information to relate the evolution of the boundaries of our cells to the matter content within them. We will now do this to Newtonian, and then post-Newtonian, levels of accuracy. Newtonian accuracy ------------------ For a regular polyhedron, at the Newtonian order of accuracy, the total surface area and volume of a cell are given by $A=\alpha_{\kappa}X^2$ and $V=\frac{1}{3} \alpha_{\kappa}X^3$, where $\alpha_{k}$ is a set of constants that depend on the cell shape in question (numerical values can be found in [@vaas]). By applying Gauss’ theorem, and using Eq. , we can re-write the evolution equation for $X$ as $$\begin{aligned} X_{,tt} = \frac{-4\pi G M - 4\pi G \int_{V}({\rho^{(2)}_{I}}+3p^{(2)}_{I}) \ dV^{(0)} }{\alpha_{\kappa}X^2} + \frac{\Lambda}{3} X \, , \label{acc1}\end{aligned}$$ where $M$ is the gravitational mass of the non-relativistic matter, defined by $M\equiv \int_V \rho^{(2)}_{M} \ dV^{(0)}$, the integrals are over the spatial volume interior to the cell, and $dV^{(0)}$ is the spatial volume element at zeroth order. This equation can be simplified, and integrated, by making use of Reynold’s transport theorem. This theorem states that for any function on space-time, $f$, we have $$\begin{aligned} \frac{d}{dt} \int f \ dV = \int f_{,t} \ dV + \int f \bm{v} \cdot d\bm{A} \, .\end{aligned}$$ Taking $f$ to be the energy density, $\rho^{(2)}_{I}$, and using the conservation equations and , then gives $$\begin{aligned} \frac{d \int \rho^{(2)}_{I} \ dV}{dt} = - \int p^{(2)}_{I} \bm{v_{I}} \cdot d \bm{A} = -p^{(2)}_{I} X_{,t} A \, . \label{halfcont}\end{aligned}$$ where we have required the barotropic fluid to be co-moving with the boundary of the cell, at all points on the boundary, and where we have made use of the fact that $\rho^{(2)}_I$ and $p^{(2)}_I$ are functions of time only. We then have the following conservation equation for the barotropic fluid $$\begin{aligned} \rho^{(2)}_{I,t} + 3 \frac{X_{,t}}{X} ({\rho^{(2)}_{I}} + p^{(2)}_{I} )=0 \, . \label{continuity1}\end{aligned}$$ This is strongly reminiscent of the corresponding equation in FLRW cosmology, as it should be. We can now simplify the evolution equation (\[acc1\]), and integrate it using the continuity equation (\[continuity1\]), to get $$\begin{aligned} \frac{X_{,tt}}{X}&= \frac{-4\pi G M}{\alpha_{\kappa}X^3} - \frac{4\pi G}{3} ({\rho^{(2)}_{I}}+3p^{(2)}_{I}) +\frac{\Lambda}{3}\, , \label{acc}\end{aligned}$$ and $$\begin{aligned} \left( \frac{X_{,t}}{X} \right)^2 &= \frac{8\pi G M}{\alpha_{\kappa}X^3} + \frac{8\pi G}{3} \rho^{(2)}_{I} - \frac{C}{X^2} + \frac{\Lambda}{3} \, , \label{constraint}\end{aligned}$$ where $C$ is an integration constant. These equations are identical to the acceleration and constraint equations of an FLRW universe filled with dust, a barotropic fluid, and a cosmological constant, with $C$ taking the role of the spatial curvature. Finally, using Eqs. and , we can read off that $\nabla . \bm{v}_{I} = 3 {X_{,t}}/{X}$. The three-velocity of the barotropic fluid is therefore given by $$\begin{aligned} \bm{v}_{I}^{\mu} = \frac{X_{,t}}{X} (x, y, z) \, . \label{vel} $$ This expression will be very useful for evaluating some of the more complicated post-Newtonian expressions that will follow. Post-Newtonian accuracy ----------------------- In this section we calculate the post-Newtonian contributions to the equations of motion of the boundary, following a similar approach to the one used in [@vaas]. The principal difference in the present case is the inclusion of the barotropic fluid, and of $\Lambda$. These lead directly to extra terms in the energy-momentum tensor, but also result in $\Phi \neq \Psi$. We must therefore keep track of each of these potentials separately. We begin by observing that the functional form of $X$, up to $O(\epsilon^2)$, is given by $$\begin{aligned} X = \zeta + \frac{1}{2} (y^2+ z^2) \bm{n} \cdot \nabla \Psi + O(\epsilon^4) \, , \label{xfunc}\end{aligned}$$ where $\zeta=\zeta(t)$ is a function of time only, and corresponds to the position of the centre of a cell face in the $x$-direction. This observation follows from the lowest order parts of Eqs. - , from the gauge conditions and , and from symmetry arguments imposed at the centre of the cell face. Taking time derivatives of Eq. , and substituting in from Eq. , then gives $$\begin{aligned} \zeta_{,tt} =& X_{,tt} - \frac{1}{2}(y^2+ z^2)(\bm{n} \cdot \nabla \Psi)^{\ddot{}} + O(\epsilon^6) \ \nonumber \\ =& \Phi_{,x} - 2\Psi\Phi_{,x} + \frac{h^{(4)}_{tt,x}}{2} - h_{tx,t} -(2\Phi_{,x} + \Psi_{,x}) X_{,t}^{2} \nonumber \\ & -( 2\Psi_{,t}+\Phi_{,t}) X_{,t} - X^{(2)}_{,A} \Phi_{,A} \nonumber \\ &- \frac{1}{2}(y^2+ z^2)(\bm{n} \cdot \nabla \Psi)^{\ddot{}} + O(\epsilon^6) \ , \label{zetatt}\end{aligned}$$ where $^{.}$ represents a time derivative along the boundary and where all quantities in this equation should be evaluated on the boundary of the cell. Several of the terms in Eq. can be related to the matter content within the cell by an application of Gauss’ theorem. For example, we can use Eq. to obtain $$\begin{aligned} \bm{n} \cdot \nabla \Psi &= -\frac{4\pi G M}{\alpha_{\kappa}X^2} -\frac{4\pi G {\rho^{(2)}_{I}} X}{3} -\frac{\Lambda X}{6} \, . \label{normpsi} \end{aligned}$$ We can also replace a number of terms in Eq. using either the gauge condition, given in Eq. , or the lower-order solutions given in Eqs. and . As an example of this, we can replace the $h_{tx,t}$ term in Eq. by using Eq. and Gauss’ theorem. This gives $$\begin{aligned} \kappa \int_{S} n_{\alpha}h_{ t\alpha, t} \ dS &= \int_{\Omega} 3\Psi_{,tt} \ dV \, . \label{gauge3}\end{aligned}$$ Finally, using the lower-order solutions for $\Phi$ and $\Psi$, from Eqs. and , we can write the generalized form of the acceleration equation in terms of the potentials defined in Eqs. -. This gives $$\begin{aligned} X_{,tt} =&-\frac{4\pi G M}{A} + (-4\pi G {\rho^{(2)}_{I}} -12\pi G p^{(2)}_{I} + \Lambda )\frac{V}{A} - \frac{3\kappa}{\alpha_{\kappa}X^2} \int_{S} \bigg( (\Phi_{M} + \Phi_{I} + \Phi_{p_{I}})_{,t} X_{,t} \bigg) \ dS \nonumber \\ & +\frac{\kappa}{\alpha_{\kappa}X^2} \int_{S} \bigg(2\Phi_{M} + 2\Phi_{I} + 3\Phi_{p_{I}} + \frac{1}{2}\Phi_{\Lambda}\bigg)\bigg(\frac{4\pi G M}{\alpha_{\kappa}X^2} + \frac{4\pi G}{3} ({\rho^{(2)}_{I}}+3p^{(2)}_{I} ) X -\frac{\Lambda X}{3} \bigg) \ dS \nonumber \\ &+ \frac{1}{\alpha_{\kappa}X^2}\bigg[4\pi G {\langle {\rho^{(2)}_{I}}(\Phi_{M}+ \Phi_{I} + 3\Phi_{p_{I}}+ \Phi_{\Lambda})\rangle} +4\pi G {\langle \rho^{(2)}_{M} (-2\Phi_{M}-2\Phi_{I} + 3\Phi_{p_{I}}+ \frac{5}{2} \Phi_{\Lambda})\rangle} \nonumber \\ &\qquad \qquad + 12\pi G {\langle p^{(2)}_{I} (\Phi_{M}+ \Phi_{I} + 3\Phi_{p_{I}}+ \Phi_{\Lambda})\rangle} - {\langle \Lambda (\Phi_{M}+ \Phi_{I} + 3\Phi_{p_{I}}+ \Phi_{\Lambda})\rangle} -12\pi G {\langle p^{(4)}_{M}\rangle} \nonumber \\ & \qquad \qquad - 8\pi G{\langle \rho^{(2)}_{M}v_{M}^2\rangle} - 8\pi G{\langle \rho^{(2)}_{I}v_{I}^2\rangle} -4\pi G{\langle \rho^{(2)}_{M}\Pi_{M}\rangle} -4\pi G{\langle \rho^{(4)}_{I}\rangle} -8 \pi G {\langle p^{(2)}_{I}v_{I}^2\rangle} -12\pi G{\langle p^{(4)}_{I}\rangle} \bigg]\nonumber \\ & + \frac{96\pi^2 G^2 M^2}{\alpha_{\kappa}^2 X^3} + \frac{64\pi^2 G^2 M {\rho^{(2)}_{I}}}{\alpha_{\kappa}} - \frac{12\pi G M C}{\alpha_{\kappa}X^2} + \frac{32\pi^2 G^2 {\rho^{(2)}_{I}}^2 X^3}{3} - 4\pi G {\rho^{(2)}_{I}} C X \nonumber \\ &+ \frac{64\pi^2 G^2 M p^{(2)}_{I} }{\alpha_{\kappa}} + \frac{8}{3}\pi G p^{(2)}_{I} \Lambda X^3 + \frac{64\pi^2 G^2 {\rho^{(2)}_{I}} p^{(2)}_{I} X^3}{3} - 8\pi G p^{(2)}_{I} C X -\frac{\Lambda^2 X^3}{6} + \frac{\Lambda C X }{2} \nonumber \\ & - \frac{3}{\alpha_{\kappa}X^2}\int_{V} (\Phi_{M} + \Phi_{I} - \frac{1}{2}\Phi_{\Lambda})_{,tt} \ dV - \frac{1}{2} (\bm{n} \cdot \nabla \Psi)^{\ddot{}} \bigg[\frac{\kappa}{\alpha_{\kappa}X^2} \int_{S} (y^2+ z^2) \ dS - (y^2 + z^2)\bigg] + O(\epsilon^6) \ , \label{final_pots}\end{aligned}$$ where $V$ is the volume of the cell, $A$ is the total surface area of the cell, and $\kappa$ is the number of faces of the cell. The notation ${\langle \varphi\rangle} = \int_V \varphi \ dV$ is used to denote quantities integrated over the volume interior to the cell, where $\varphi$ is some scalar function on the space-time. The quantity $( \bm{n} \cdot \nabla \Psi)^{\ddot{}}$, in this equation, can be found to be given by $$\begin{aligned} ( \bm{n} \cdot \nabla \Psi)^{\ddot{}}&= -\frac{224\pi^2 G^2 M^2}{\alpha_{\kappa}^2 X^5} -\frac{14\pi G M \Lambda}{3\alpha_{\kappa} X^2} -\frac{448\pi^2 G^2 M {\rho^{(2)}_{I}}}{3\alpha_{\kappa}X^2}+\frac{24\pi G MC}{\alpha_{\kappa}X^4} -\frac{112\pi^2 G^2 M p^{(2)}_{I} }{\alpha_{\kappa} X^2} \nonumber \\ &\quad - \frac{224\pi^2 G^2 {\rho^{(2)}_{I}}^2 X}{9} - \frac{112\pi^2 G^2 {\rho^{(2)}_{I}} p^{(2)}_{I} X}{3} - \frac{14 \pi G {\rho^{(2)}_{I}} \Lambda X}{9} - 16\pi^2 G^2 {p^{(2)}_{I}}^2 X \nonumber \\ &\quad -\frac{2\pi G p^{(2)}_{I} \Lambda X }{3} - \frac{\Lambda^2 X}{18} + \frac{8 \pi G {\rho^{(2)}_{I}} C}{X} + \frac{8 \pi G p^{(2)}_{I} C}{X} + 4\pi G p^{(2)}_{I, t} X_{,t} \ . \label{psiddot}\end{aligned}$$ The acceleration equation is fully general, being valid for any cell shape and any distribution of matter in the presence of a barotropic fluid and a cosmological constant. This complicated equation reduces to the one derived in [@vaas], in the absence of the barotropic fluid and the cosmological constant. In addition, however, the present equation contains several cross terms between the different types of matter. These arise due to the non-linearity of Einstein’s equations, and should be expected to alter the effects of back-reaction. Before moving on to consider simple matter distributions, we can simplify Eq. a little by looking at the specific case of cubic cells. In this case the total volume of a cell is given by $$\begin{aligned} V = 8 \zeta^3 + 8 (\bm{n} \cdot \nabla \Psi) \zeta^4 + 3\int_{V} \Psi \ dV + O(\epsilon^4) \, ,\end{aligned}$$ and the total surface area is given by $$\begin{aligned} A&= 24\zeta^2\bigg(1 + \frac{4}{3} (\bm{n} \cdot \nabla \Psi) \zeta + \frac{1}{2\zeta^2} \int_{S} \Psi \ dS\bigg) + O(\epsilon^4) \, .\end{aligned}$$ We can also use $\kappa = 6$ and $\alpha_{k}=24$, for the specific case of cubic cells, and rewrite the acceleration equation as $$\begin{aligned} X_{,tt} =&\frac{- \pi G M}{6 \zeta^2} -\frac{4\pi G}{3}( {\rho^{(2)}_{I}} + 3 p^{(2)}_{I}) \zeta + \frac{\Lambda\zeta}{3} \nonumber \\ &+ \frac{7\pi^2 G^2 M^2}{27 X^3} +\frac{118\pi^2 G^2 M{\rho^{(2)}_{I}} }{27} +\frac{5\pi G M\Lambda}{108} + 4\pi^2 G^2 M p^{(2)}_{I} + \frac{496\pi^2 G^2 {\rho^{(2)}_{I}}^2 X^3}{27} + 32\pi^2 G^2 {\rho^{(2)}_{I}}P X^3\nonumber \\ &+\frac{16\pi G {\rho^{(2)}_{I}} \Lambda X^3}{27} + \frac{8\pi G p^{(2)}_{I} \Lambda X^3}{3} - \frac{7\Lambda^2 X^3}{54} - \frac{5\pi G M C}{6X^2} - \frac{20 \pi G {\rho^{(2)}_{I}} C X}{3} - \frac{32 \pi G p^{(2)}_{I} C X}{3} +\frac{\Lambda C X }{2} \nonumber \\ & +\frac{1}{4X^2} \int_{S} \bigg(4\Phi_{M} + 4\Phi_{I} + 3\Phi_{p_{I}} - \frac{1}{2}\Phi_{\Lambda}\bigg)\bigg(\frac{\pi G M}{6X^2} + \frac{4\pi G}{3} ({\rho^{(2)}_{I}}+3p^{(2)}_{I} ) X - \frac{\Lambda X}{3} \bigg) \ dS \nonumber \\ &-\frac{3}{4X^2} \int_{S} \bigg((\Phi_{M} + \Phi_{I} + \Phi_{p_{I}})_{,t} X_{,t} \bigg) \ dS +\frac{16\pi^2 G^2 {p^{(2)}_{I}}^2 X}{3} - \frac{ 4\pi G p^{(2)}_{I, t} X_{,t} X^2}{3} \nonumber \\ &+ \frac{1}{24X^2}\bigg[{\langle (4 \pi G (\rho^{(2)} + 3p^{(2)}_{I} ) - \Lambda)(- 2\Phi_{M} - 2\Phi_{I} + 3\Phi_{p_{I}}+ \frac{5}{2}\Phi_{\Lambda})\rangle} - 8\pi G{\langle \rho^{(2)}_{M}v_{M}^2\rangle} - 8\pi G{\langle \rho^{(2)}_{I}v_{I}^2\rangle} \nonumber \\ & \qquad \qquad -4\pi G{\langle \rho^{(2)}_{M}\Pi_{M}\rangle} -4\pi G{\langle \rho^{(4)}_{I}\rangle} -8 \pi G {\langle p^{(2)}_{I}v_{I}^2\rangle}-12\pi G {\langle p^{(4)}_{M}\rangle} -12\pi G{\langle p^{(4)}_{I}\rangle} \bigg]\nonumber \\ & - \frac{1}{8X^2} \int_{V} (\Phi_{M} + \Phi_{I} - \frac{1}{2}\Phi_{\Lambda})_{,tt} \ dV + \frac{1}{2} (\bm{n} \cdot \nabla \Psi)^{\ddot{}} (y^2 + z^2) + O(\epsilon^6) \ . \label{cubicacc}\end{aligned}$$ Every term in this equation can be solved for in complete generality using the Green’s function formalism set out in [@vaas], but it still remains a very complicated expression. Instead, and in order to show the effects of back-reaction in a simple illustrative example, we look at the case of regularly arranged point-like particles in a sea of radiation, and in the presence of a cosmological constant. Point sources with radiation, spatial curvature and $\Lambda$ {#sec4} ============================================================= To find an explicit solution to the acceleration equation, let us consider the case of a point source located at the centre of each cell, in the presence of radiation and a cosmological constant. To simplify matters further, let us evaluate the acceleration equation at the centre of a cell face ([*i.e.*]{} at $y=z=0$). Solutions --------- In the case of point sources we have $v_{M}^{\alpha} = p^{(4)}_{M} = \Pi_{M} ={\langle \rho^{(2)}_{M} \Phi_{M}\rangle} ={\langle \rho^{(2)}_{M} \Phi_{I}\rangle}= {\langle \rho^{(2)}_{M} \Phi_{p_{I}}\rangle} = {\langle \rho^{(2)}_{M} \Phi_{\Lambda}\rangle} = 0$. Hence, in this case, the potentials defined in Eq. simplify to $$\begin{aligned} \label{110} \nabla^2 \Phi_{M} = -4\pi G M \delta(\mathbf{x}) \, , \quad \nabla^2 \Phi_{I} = -4\pi G {\rho^{(2)}_{r}} \, , \quad \nabla^2 \Phi_{p_{I}} = -\frac{4\pi G}{3} \rho^{(2)}_{r} \, , \quad {\rm and} \quad \nabla^2 \Phi_{\Lambda} = \Lambda \, ,\end{aligned}$$ where $M$ is the gravitational mass of the point source at the centre of the cell, and ${\rho^{(2)}_{r}}$ is the energy density of the radiation. The first of these potentials can be solved for, using the method of images, and can be used to absorb all auxiliary functions of time (see [@vaas] for details). This gives $$\begin{aligned} \Phi_{M} &= \lim _{\mathcal{N} \to \infty} \sum_{\bm{\beta} = - \mathcal{N}}^{\mathcal{N}}\frac{G M}{\sqrt{(x -2\beta_{1} X)^2 + (\hat{y} -2 \beta_{2} X)^2 + (\hat{z}- 2\beta_{3} X)^2 }} - \lim _{\mathcal{N} \to \infty} \sum_{\bm{\beta^{*}} = - \mathcal{N}}^{\mathcal{N}}\frac{G M}{2 |\bm{\beta}| X} \, , \label{intphi} $$ where $\bm{\beta^{*}}$ indicates that the null triplet has been removed. The remaining potentials are given by $$\begin{aligned} \Phi_{I}= -\frac{2\pi G {\rho^{(2)}_{r}}}{3}(x^2 + y^2 + z^2) \, , \quad \Phi_{p_{I}}= -\frac{2\pi G \rho^{(2)}_{r}}{9}(x^2 + y^2 + z^2) \, , \quad {\rm and} \quad \Phi_{\Lambda}= \frac{\Lambda}{6}(x^2 + y^2 + z^2) \, . \label{pot_solns}\end{aligned}$$ If we now assume that the radiation does not interact with the point sources, then we have $p^{(4)}_{r} = \frac{1}{3}\rho^{(4)}_{r}$. Using the energy-momentum conservation equation at $O(\epsilon^4)$, the velocity of the barotropic fluid given in Eq. , and the lower-order acceleration and constraint equations, the energy density of radiation at $O(\epsilon^4)$ can then be seen to be given by $$\begin{aligned} \rho^{(4)}_{r} =& \bigg[\frac{\pi G M \rho^{(2)}_{r}}{X^3} + \frac{16}{3}\pi G {\rho^{(2)}_{r}}^2 - 2 \frac{\rho^{(2)}_{r}C}{X^2} + \frac{2\rho^{(2)}_{r}\Lambda}{3} \bigg] r^2 + 4\rho^{(2)}_{r}\Phi_{M} \, , \label{rho4}\end{aligned}$$ where $r^2=x^2 + y^2 + z^2$. Using all of this information, the acceleration equation can then be found to reduce to $$\begin{aligned} X_{,tt}=&-\frac{\pi G M}{6X^2} - \frac{8\pi G}{3} \rho^{(2)}_{r} X +\frac{\Lambda X}{3} +\frac{\pi G^2 M^2}{X^3} \mathcal{A}_{1} + \pi G^2 M\rho^{(2)}_{r} \mathcal{A}_{2} + G M \Lambda \mathcal{A}_{3} \nonumber \\ & + \frac{G M C}{X^2} \mathcal{A}_{4} - \frac{64}{9}\pi^2 G^2 {\rho^{(2)}_{r}}^2 X^3 - \frac{8}{9}\pi G \rho^{(2)}_{r} \Lambda X^3 -\frac{2}{9} \Lambda^2 X^3 +\frac{4}{3} \pi G \rho^{(2)}_{r} C X + \frac{1}{2} \Lambda C X + O(\epsilon^6) \ , \label{accmink}\end{aligned}$$ where $\mathcal{A}_{1}$, $\mathcal{A}_{2}$, $\mathcal{A}_{3}$ and $\mathcal{A}_{4}$ are constants whose values are given in Table \[tab1\], and whose relationship to the variables used in [@vaas] are given in the appendix. **Constant** **Numerical value** ------------------- --------------------------------- $\mathcal{A}_{1}$ $\qquad \phantom{-}1.27 \ldots$ $\mathcal{A}_{2}$ $\qquad -9.29 \dots$ $\mathcal{A}_{3}$ $\qquad -0.219 \dots$ $\mathcal{A}_{4}$ $\qquad \phantom{-}0.809 \dots$ : \[tab1\] The numerical values of $\mathcal{A}_{1}$, $\mathcal{A}_{2}$, $\mathcal{A}_{3}$ and $\mathcal{A}_{4}$, from Eq. . These are the numbers approached as the number of reflections in the method of images diverges to infinity. Although already dramatically simplified, we can reduce this equation further by transforming into a FLRW background. This can be achieved using the following coordinate transformations [@vaas]: $$\begin{aligned} t &= \hat{t} + \frac{a_{,\hat{t}} a}{2} (\hat{x}^2 + \hat{y}^2 + \hat{z}^2) + O(\epsilon^3) \label{timetrans} \\ \nonumber \\ x &= a \hat{x} \bigg[1 + \frac{(a_{,\hat{t}})^2}{4} (\hat{x}^2 + \hat{y}^2 + \hat{z}^2)\bigg] +O(\epsilon^4) \\ \nonumber \\ y &= a \hat{y} \bigg[1 + \frac{(a_{,\hat{t}})^2}{4} (\hat{x}^2 + \hat{y}^2 + \hat{z}^2)\bigg] +O(\epsilon^4) \\ \nonumber \\ z &= a \hat{z} \bigg[1 + \frac{(a_{,\hat{t}})^2}{4} (\hat{x}^2 + \hat{y}^2 + \hat{z}^2)\bigg] +O(\epsilon^4) \, , \end{aligned}$$ where the new coordinates $\hat{t}, \hat{x}, \hat{y}, \hat{z}$ are the standard set in an FLRW background, and where $a(\hat{t})$ is the scale factor of that background. The energy density in these new coordinates is given by $$\begin{aligned} \rho^{(2)}_{r}(t) &= \rho^{(2)}_{r}(\hat{t}) - 2a_{,\hat{t}}^2 \rho^{(2)}_{r}(\hat{t}) (\hat{x}^2 + \hat{y}^2 + \hat{z}^2) +O(\epsilon^4)\, .\end{aligned}$$ Evaluating this expression at the centre of a cell face, and using the the lower-order constraint equation , gives $$\begin{aligned} \rho^{(2)}_{r}(t) =& \hat{\rho}^{(2)}_{r} - \bigg(\frac{2\pi G (\hat{\rho}^{(2)}_{r}) M }{3 a \hat{X}_{0}^3} + \frac{16\pi G (\hat{\rho}^{(2)}_{r})^2 a^2}{3} \label{rhotrans} \\ & \qquad \qquad \qquad \qquad + \frac{2\hat{\rho}^{(2)}_{r}\Lambda a^2}{3} - 2\hat{\rho}^{(2)}_{r}k\bigg) \hat{X}_{0}^2 \, . \nonumber\end{aligned}$$ In this last equation we have introduced the abbreviated notation $\hat{\rho}^{(2)}_{r} = \rho^{(2)}_{r}(\hat{t})$, and used $k$ to denote Gaussian curvature in the background FLRW geometry. Similarly, at the centre of a cell face the position of the boundary transforms as $$\begin{aligned} X=& a \hat{X}_{0}\bigg[1+ \frac{a_{,\hat{t}}^2}{4} \hat{X}_{0}^2\bigg] \label{XtoXhat} \\ =& a \hat{X}_{0} \bigg[1+ \bigg( \frac{\pi G M }{12 a \hat{X}_{0}^3} + \frac{2\pi G \hat{\rho}^{(2)}_{r} a^2}{3} + \frac{\Lambda}{12} a^2 - \frac{k}{4} \bigg) \hat{X}_{0}^2 \bigg] , \nonumber\end{aligned}$$ where $a=a(\hat{t})$ in this expression. Results ------- Finally, using Eqs. - , the acceleration equation simplifies down to $$\begin{aligned} \frac{\ddot{a}}{a} = &-\frac{4\pi G}{3} (\hat{\rho}^{(2)}_{M} +2 \hat{\rho}^{(2)}_{r}) + \frac{\Lambda}{3} +\mathcal{B}_{1} + O(\epsilon^6) \, , \label{backacc}\end{aligned}$$ where overdots in this equation denote derivatives with respect to $\hat{t}$, and where the back-reaction term, $\mathcal{B}_{1}$, is given by $$\begin{aligned} \mathcal{B}_{1} &\simeq \left(4 \pi G \hat{\rho}^{(2)}_{M} a \hat{X}_{0}\right)^2 \left(1.50 - 1.20 \frac{\Omega_{r}}{\Omega_{M}} + 0.88 \frac{\Omega_{k}}{\Omega_{M}}\right) \, . \label{b1}\end{aligned}$$ In writing these equations we have used the expression $\hat{\rho}^{(2)}_{M} \equiv {M}/{8 a^3 \hat{X}_{0}^3}$ for the average mass density in a cell, and have introduced the usual cosmological parameters $$\begin{aligned} \Omega_{M} \equiv& \frac{8\pi G \hat{\rho}^{(2)}_{M}}{3 H^2} \ , \quad \Omega_{r} \equiv \frac{8 \pi G \hat{\rho}^{(2)}_{r}}{3 H^2} \ , \quad \Omega_{k} \equiv& -\frac{k}{a^2 H^2} \, , \nonumber\end{aligned}$$ where $H \equiv \dot{a}/a$. The numerical values inside the brackets in Eq. are calculated from the constants in Table \[tab1\], and are quoted to the second decimal place only. The reader will note that $\Lambda$ does not appear in this expression, and so does not contribute to this back-reaction term at this level of accuracy. It can also be seen that, in the absence of the point-like particles, the acceleration equation reduces to the standard Friedmann equation for a universe with radiation, spatial curvature and a cosmological constant, as expected. The back-reaction term, $\mathcal{B}_{1}$, is strongly influenced by the presence of radiation and spatial curvature, but not $\Lambda$. As can be seen from Fig. \[fig2\], the magnitude of $\mathcal{B}_{1}$ decreases as the amount of radiation in the Universe increases. This is independent of the expected suppression in the growth of structure that radiation is known to cause, as the discrete nature of the non-relativistic matter in this example exists for all time. Fig. \[fig2\] also shows us that the back-reaction effect reduces for a closed universe, and increases for an open universe. In Fig. \[fig3\] we plot the consequences of having non-zero amounts of both radiation and positive spatial curvature, while in Fig. \[fig4\] we show the corresponding plot for negative spatial curvature. In this latter case the spatial curvature and radiation can have compensating effects as they are simultaneously increased. As well as an acceleration equation, we can integrate Eq. to obtain a constraint equation. This is given by $$\begin{aligned} \left( \frac{\dot{a}}{a} \right)^2 =& \frac{8\pi G}{3}( \hat{\rho}^{(2)}_{M} + \hat{\rho}^{(2)}_{r}) -\frac{k}{a^2} +\frac{\Lambda}{3} +\mathcal{B}_{2} + O(\epsilon^6) \, , \label{backcon}\end{aligned}$$ where we have introduced $\mathcal{B}_{2}$ to denote the leading-order contribution to the back-reaction in this equation, and written $C=k\hat{X}_{0}^2 + O(\epsilon^4)$. The back-reaction term can be written explicitly as $$\begin{aligned} \label{b2} \mathcal{B}_{2} &\simeq -\left( 4\pi G \hat{\rho}^{(2)}_{M} a \hat{X}_{0}\right)^2 \left(1.50 - 0.80 \frac{\Omega_{r}}{\Omega_{M}} + 1.76 \frac{\Omega_{k}}{\Omega_{M}} \right) \, .\end{aligned}$$ Let us now consider how different forms of matter affect the back-reaction in the Hubble rate. From Fig. \[fig5\] it can be seen that the effect of radiation is to decrease the back-reaction term in this equation. In the Hubble rate, the back-reaction effect from the non-relativistic matter itself is negative. This means that radiation increases the value of the Hubble rate. The cosmological constant again makes a negligible contribution to the back-reaction. Finally, at $O(\epsilon^4)$, the Hubble rate is greater for a universe with positive spatial curvature, and smaller for a universe with negative spatial curvature. In Figs. \[fig6\] and \[fig7\] we plot the results of simultaneously adding radiation and spatial curvature. Once again, if spatial curvature is negative, then the effect it has on the back-reaction term can compensate that of radiation. If spatial curvature is positive, however, the effect it has on back-reaction is complementary to that of radiation. Let us now consider the functional form of the different terms in the back-reaction equations. Recall that the lowest-order parts of the matter density and radiation density both scale in exactly the same way as in a FLRW model. This means that the leading-order correction arising from the non-relativistic matter itself is a radiation-like term, as identified in [@vaas]. The non-linear effect from radiation, on the other hand, scales as a fluid with equation of state $p= \frac{2}{3} \rho$. This is somewhere between the behaviour expected from a free scalar field, and that of normal radiation. The leading-order correction from the spatial curvature scales in the same way as non-relativistic matter, and effectively renormalises the value of the gravitational mass in the Universe. Let us now consider the deceleration parameter, $q_0$. Using Eqs. and , we find this parameter to be given by $$\begin{aligned} q_{0} &\equiv - \frac{\ddot{a} a}{\dot{a}^2} = \frac{(\Omega_{M} +2 \Omega_{r} - 2 \Omega_{\Lambda} )}{ 2( \Omega_{M} + \Omega_{r} + \Omega_{\Lambda} + \Omega_{k})} + \mathcal{B}_{3} + O(\epsilon^4) \label{backdec} \, ,\end{aligned}$$ where the back-reaction term in this equation is $$\begin{aligned} \label{b3} \mathcal{B}_{3} =& -\frac{3\mathcal{B}_{1}}{8\pi G \hat{\rho}^{(2)}_{M} \left( 1 + \frac{\Omega_{r}}{\Omega_{M}} + \frac{\Omega_{\Lambda}}{\Omega_{M}} + \frac{\Omega_{k}}{\Omega_{M}} \right) } \nonumber \\ \nonumber \\ &- \frac{3\mathcal{B}_{2} (1 +2 \frac{\Omega_{r}}{\Omega_{M}} - 2 \frac{\Omega_{\Lambda}}{\Omega_{M}} )}{16\pi G \hat{\rho}^{(2)}_{M} \left( 1 + \frac{\Omega_{r}}{\Omega_{M} } + \frac{\Omega_{\Lambda}}{\Omega_{M} } + \frac{\Omega_{k}}{\Omega_{M}} \right)^2 } \, , $$ where $\Omega_{\Lambda} \equiv \Lambda/3 H^2$, and where the values of $\mathcal{B}_{1}$ and $\mathcal{B}_{2}$ are given in Eqs. (\[b1\]) and (\[b2\]). The effect that radiation, spatial curvature and a cosmological constant have on the back-reaction term $\mathcal{B}_{3}$ is displayed graphically in Fig. \[fig8\]. Unlike the cases of $\mathcal{B}_{1}$ and $\mathcal{B}_{2}$, it can be seen that $\mathcal{B}_{3}$ is only of order $\epsilon^2$. This is because the deceleration parameter, $q_0$, is itself an order $1$ quantity. The back-reaction in this quantity is therefore still small compared to the corresponding FLRW value, even though its absolute magnitude has increased from the terms that enter into the Friedmann equations. At scales of about $100$ Mpc, we estimate that these corrections amount to changes at the level of about $1$ part in $10^{4}$ in the deceleration parameter. The value of $\mathcal{B}_{3}$ in the absence of radiation and a cosmological constant is negative, meaning that discretizing the matter in this way leads to a small increase in acceleration. This is no surprise, as back-reaction has already been shown to increase $\ddot{a}/a$ and decrease $\dot{a}^2/a^2$. As the value of $q_0$ is simply given by the ratio of these two quantities (with a minus sign), we have that both types of back-reaction contribute cumulatively to the acceleration measured by this dimensionless parameter. It can be seen from Fig. \[fig8\] that radiation increases the back-reaction that occurs in the deceleration parameter. Positive values of $\Lambda$ have a small effect on $\mathcal{B}_{3}$, even though it does not have a noticeable effect on $\mathcal{B}_{1}$ or $\mathcal{B}_{2}$. This is because, in Eq. (\[b3\]), we find that $\Lambda$ enters into the background terms that multiply $\mathcal{B}_{1}$ and $\mathcal{B}_{2}$. Negative values of $\Lambda$ can make a more sizeable contribution to the back-reaction of $q_0$, and can even cause the back-reaction term to contribute to deceleration, if its magnitude is large enough. The effect of positive spatial curvature on $\mathcal{B}_{3}$ can also be large, but in this case causes extra acceleration. One should keep in mind, however, that for both of these last two cases the background value of the deceleration also diverges as $\Omega_{\Lambda} \rightarrow - \Omega_{M}$ and $\Omega_{k} \rightarrow -\Omega_{M}$. Finally, and unlike in the acceleration and constraint equations, a negative value for the spatial curvature provides only a small correction to the value of $\mathcal{B}_{3}$. The effects on $\mathcal{B}_{3}$ of simultaneously adding negative spatial curvature, positive cosmological constant, and non-zero radiation are displayed in Figs. \[fig10\]-\[fig13\]. It can be seen from Fig. \[fig10\] that, in the presence of radiation, negative spatial curvature has only a small effect on the back-reaction. Similarly, in Fig. \[fig11\], it can be seen that positive values of $\Lambda$ have a small effect on the back-reaction term, when radiation is present. On the other hand, in Fig. \[fig13\], it can be seen that although positive $\Lambda$ and negative spatial curvature have only a small effect on the back-reaction in the absence of radiation, these effects are comparable to each other when radiation is absent. In this case, for small values of $\Lambda$, we have a small correction to the absolute value of $\mathcal{B}_{3}$, with a maximum at $\Omega_{\Lambda} = 0.5\Omega_{M}$. Negative spatial curvature does not affect $\mathcal{B}_{3}$ for small values of $\Lambda$, but does become increasingly significant as the value of $\Lambda$ increases. CONCLUSIONS =========== We have constructed an original framework that can be used to quantify the effects that radiation, spatial curvature and $\Lambda$ have on the cosmological back-reaction that results from the existence of non-linear inhomogeneities. Our approach is based on modelling the universe as a regular lattice, in which all structure is periodic. The geometry of space-time within each individual cell is then taken to be close to Minkowski space, and a post-Newtonian perturbative expansion is used to model all gravitational fields and matter content. By patching these cells together, using Israel’s junction conditions at reflective symmetric boundaries, we finally construct a global and dynamical space-time. We derived an acceleration equation that describes the expansion of this emergent cosmology, and which is valid for any arbitrary distribution of matter within each cell (as long as it is distributed periodically). This equation is valid in the presence of both a barotropic fluid, with unspecified equation of state, and a cosmological constant. Having derived the equations that govern the general case, we then simplified our equations by considering the specific example of a point-like mass at the centre of each lattice cell, in a sea of radiation and in the presence of a cosmological constant. The back-reaction terms generated by the matter fields alone behave like radiation in the Friedmann equation, as found in [@vaas]. The presence of actual radiation, however, reduces the magnitude of the back-reaction in both the acceleration and constraint equations. In contrast, we find that the cosmological constant has a negligible effect on back-reaction, and that spatial curvature can have a significant effect depending on whether the Universe is open or closed. These results explain why the leading-order effects of back-reaction occur at the level of linear-order perturbations in cosmological perturbation theory [@lam; @radlam1; @radlam2], even though they require second-order gravity in order to be calculated. In future work we will calculate observables in these models, by solving the equations that govern the expansion of a beam of light [@toapp2]. We also aim to further improve their realism by reducing the symmetries required at the junctions between cells. ACKNOWLEDGEMENTS {#acknowledgements .unnumbered} ================ We are grateful to I. Brown, P. Carrilho, D. Dold, P. Fleury and S. Imrith for helpful discussions and comments. VAAS and TC both acknowledge support from the STFC. **Constant** **Asymptotic value** -------------- --------------------------------- $D$ $\qquad \phantom{-}1.44 \ldots$ $E$ $\qquad \phantom{-}0.643 \dots$ $F$ $\qquad -1.62 \dots$ $P$ $\qquad \phantom{-}0.304 \dots$ $V_{1}$ $\qquad \phantom{-}2.31 \dots$ : \[tab3\] The numerical values of $D$, $E$, $F$, $P$, and $V_{1}$ that are approached as the number of reflections used in the method of images diverges to infinity. Appendix: Numerical coefficients {#appA .unnumbered} ================================ The numerical constants that appear in the acceleration equation are given, in terms of the variables used in [@vaas], by $$\begin{aligned} \mathcal{A}_{1} =& \frac{D}{3} - \frac{E}{2} + \frac{7\pi}{27} - \frac{F}{6} + \frac{P}{12} \, , \nonumber \\ \nonumber \\ \mathcal{A}_{2} =& \frac{13\pi}{27} + \frac{16D}{3} - 4E - 8V_{1} - \frac{4F}{3}+ \frac{4P}{3} \, , \nonumber \\ \nonumber \\ \mathcal{A}_{3} =&\frac{5\pi}{216} -\frac{2D}{3} - \frac{E}{2} + \frac{V_{1}}{3} - \frac{F}{6}-\frac{P}{6} \, , \nonumber \\ \nonumber \\ \mathcal{A}_{4} =& -\frac{5 \pi}{6} +\frac{F}{2} +\frac{3E}{2} \, .\end{aligned}$$ The numerical values of $\mathcal{A}_{1}$, $\mathcal{A}_{2}$, $\mathcal{A}_{3}$ and $\mathcal{A}_{4}$ are given in Table \[tab1\], and the numerical values of $D$, $E$, $F$, $P$ and $V_{1}$ are given in Table \[tab3\]. The quantity $V_1$, which is defined by $$V_1 \equiv \frac{\int_{-X}^{X} \Phi_M dx dy dz}{4 G M X^2} \, ,$$ converges to its limiting value quickly as the number of image masses is increased, as illustrated in Fig. \[fig14\]. The convergence of $D$, $E$, $F$, $P$ and $V_{1}$ is given in [@vaas]. [99]{} V. A. A. Sanghai and T. Clifton, [*Phys. Rev. D*]{} [**91**]{}, 103532 (2015); [*Phys. Rev. D*]{} [**93**]{}, 089903(E) (2016). T. Clifton, [*Classical Quantum Gravity*]{} [**28**]{}, 164011 (2011). W. Israel, [*Nuovo Cimento B*]{} **44**, 1 (1966); [*Nuovo Cimento B*]{} **48**, 463 (1967). C. Clarkson, G. Ellis, J. Larena and O. Umeh, [*Rep. Prog. Phys.*]{} [**74**]{}, 112901 (2011). T. Buchert and S. Räsänen, [*Annu. Rev. Nucl. Part. Sci.*]{} [**62**]{}, 57 (2012). T. Clifton, [*Int. J. Mod. Phys. D*]{} [**22**]{}, 133004 (2013). http://sci.esa.int/euclid. http://www.skatelescope.org/. T. Clifton, K. Rosquist and R. Tavakol, [*Phys. Rev. D*]{} [**86**]{}, 043506 (2012). C.-M. Yoo, H. Abe, Y. Takamori and K.-i. Nakao, [*Phys. Rev. D*]{} [**86**]{}, 044027 (2012). E. Bentivegna, [*Classical Quantum Gravity*]{} [**31**]{}, 035004 (2014). M. Korzyński, [*Classical Quantum Gravity*]{} [**31**]{}, 085002 (2014). M. Korzyński, [*Classical Quantum Gravity*]{} [**32**]{}, 215013 (2015). E. Bentivegna and M. Korzyński, [*Classical Quantum Gravity*]{} [**29**]{}, 165007 (2012). E. Bentivegna and M. Korzyński, [*Classical Quantum Gravity*]{} [**30**]{}, 235008 (2013). C.-M. Yoo, H. Okawa and K.-i. Nakao, [*Phys. Rev. Lett.*]{} [**111**]{}, 161102 (2013). C.-M. Yoo and H. Okawa, [*Phys. Rev. D*]{} [**89**]{}, 123502 (2014). T. Clifton, D. Gregoris and K. Rosquist, [*Classical Quantum Gravity*]{} [**31**]{}, 105012 (2014). T. Clifton, D. Gregoris and K. Rosquist, [*Classical Quantum Gravity*]{} [**31**]{}, 105012 (2014). M. Korzyński, I. Hinder and E. Bentivegna, [*J. Cosmol. Astropart. Phys.*]{} [**08**]{} (2015) 025. C. Clarkson, K. Ananda and J. Larena, [*Phys. Rev. D*]{} [**80**]{}, 083525 (2009). I. A. Brown, J. Behrend and K. A. Malik, [*J. Cosmol. Astropart. Phys.*]{} [**11**]{} (2009) 027. I. A. Brown, G. Robbers and J. Behrend, [*J. Cosmol. Astropart. Phys.*]{} [**04**]{} (2009) 016. W. C. Lim, M. Regis and C. Clarkson, [*J. Cosmol. Astropart. Phys.*]{} [**10**]{} (2013) 010. H. S. M. Coxeter, [*Regular Polytopes*]{} (Dover Publications, New York, 1973). C. M. Will, [*Theory and Experiment in Gravitational Physics*]{} (Cambridge University Press, Cambridge, 1993). V. A. A. Sanghai, P. Fleury and T. Clifton, ([*to be published*]{}).
{ "pile_set_name": "ArXiv" }
--- author: - Sergei Nayakshin bibliography: - 'references.bib' title: THE PHYSICS OF ACCRETION DISKS WITH MAGNETIC FLARES --- definition abstract.tex statement.tex acknowledgment.tex Introduction ============ intro.tex Magnetic Flares in Accretion Disks: Preliminaries ================================================= chapter2.tex Pressure Equilibrium and Containment ==================================== chapter3.tex Pressure-Ionization Instability in X-ray Reflection =================================================== chapter4.tex X-ray reflection in AGNs and The BBB ==================================== chapter5.tex Energy budget of the corona =========================== chapter6.tex Classification of Accretion Disk States and Comparison to Observations ====================================================================== chapter7.tex
{ "pile_set_name": "ArXiv" }
--- abstract: 'We investigate the Kerr black hole defined on canonically deformed space-time. Particulary, we find the corresponding event horizon, the ergosphere, the temperature and the entropy of such deformed object.' ---    [[${\rm {Marcin\;Daszkiewicz}}$]{}]{}\ [ ${\rm{ University\; of\; Wroclaw,\; pl.\; Maxa\; Borna\; 9,\; 50-206\; Wroclaw,\; Poland}}$]{} [ ${\rm{ e-mail:\; [email protected]}}$]{} [[${\rm {Cezary\;J.\;Walczyk}}$]{}]{}\ [ ${\rm{ University\; of\; Bialystok,\; ul.\; Lipowa\;41,\; 15-424\; Bialystok,\; Poland}}$]{} [ ${\rm{ e-mail:\; [email protected]}}$]{} [[Introduction]{}]{} ==================== The idea to use noncommutative coordinates is quite old - it goes back to Heisenberg and was firstly formalized by Snyder in [@snyder]. Recently, however, there were found new formal arguments based mainly on Quantum Gravity [@2], [@2a] and String Theory models [@recent], [@string1], indicating that space-time at Planck scale should be noncommutative, i.e. it should have a quantum nature. On the other side, the main reason for such considerations follows from many phenomenological considerations, which state that relativistic space-time symmetries should be modified (deformed) at Planck scale, while the classical Poincare invariance still remains valid at larger distances [@1a], [@1d]. In accordance with the Hopf-algebraic classification of all deformations of relativistic and nonrelativistic symmetries (see [@clas1], [@clas2]) the most general form of space-time noncommutativity looks as follows $$[\,x_{\mu},x_\nu\,] = \theta_{\mu\nu} (x)\;,\label{nonco}$$ where $$\theta_{\mu\nu} (x) = \theta_{\mu\nu}^{(0)} + \theta_{\mu\nu}^{(1)\,\rho}x_{\rho} + \theta_{\mu\nu}^{(2)\,\rho\tau}x_{\rho}x_{\tau}\;.\label{par}$$ For the simplest, canonical noncommutativity $(\theta_{\mu\nu} (x) = \theta_{\mu\nu}^{(0)})$, the corresponding Poincare Hopf algebra has been provided in [@oeckl] and [@chi] with the use of twist procedure [@drin]-[@twist1], while its nonrelativistic counterparts have been discovered by various contraction schemes in [@dasz]. The Lie-algebraic $(\theta_{\mu\nu} (x) =\theta_{\mu\nu}^{(1)\,\rho}x_{\rho})$ relativistic and nonrelativistic symmetries have been proposed in [@kappaP] and [@kappaG] respectively. In the literature they are known as $\kappa$-Poincare and $\kappa$-Galilei Hopf algebra with mass-like deformation parameter $\kappa$. Besides, there were proposed the twist deformations of a Lie-type at relativistic and nonrelativistic level in [@lie2], [@lie1] and [@dasz]. The quadratic deformation $(\theta_{\mu\nu} (x) =\theta_{\mu\nu}^{(2)\,\rho\tau}x_{\rho}x_{\tau})$ has been studied in [@paolo] and [@lie2]. Recently, there appeared a lot of papers dealing with classical ([@deri]-[@daszwal]) and quantum ([@qm1]-[@oscy]) mechanics, Doubly Special Relativity frameworks ([@dsr1a], [@dsr1b]), statistical physics ([@maggiore], [@rama]) and field theoretical models (see e.g. [@przeglad]), defined on the canonically and Lie-algebraically deformed space-times[^1]. It should be noted, however, that the especially interesting studies have been performed in articles [@treatment]-[@inne2] in the context of so-called black hole physics, i.e. there have been investigated (with use of different methods and techniques) the basic types of noncommutative black hole metrics. In this article we study the effect of canonical noncommutativity on the Kerr solution of General Relativity equation with use of treatment proposed in [@treatment]-[@treatment2]. Particulary, we investigate the impact of quantum space on the event horizon as well as on the shape of ergosphere. Besides, as in the case of basic nonrotating black holes described by Schwarzschild and Reisner-Nordstrom metric tensors respectively, we find the temperature and entropy of such canonically deformed object. In our investigation we proceed in accordance with the following algorithm (see [@treatment]-[@treatment2])[^2]. Firstly, we assume that particle-like gravitational source remains point-like. Next, we exchange the commutative variables in classical Kerr solution of field equation by noncommutative ones. Further, we rewrite the obtained in such a way metric tensor in terms of commutative phase space variables and deformation parameter $\theta$ (see formula (\[rep\])). Finally, we perform the basic analysis of such a prepared mathematical object. It should be noted that the classically noncommutative Kerr black hole has been already analyzed in article [@inne2]. However, the used therein techniques are completely different than the methods preferred by us. For example, there is assumed that due to the space-time noncommutativity the mass density particle-like gravitational source is smeared. Such an assumption leads to the new (Kerr) solution of General Relativity equations for which the corresponding event horizon can be found only numerically. The paper is organized as follows. In Sect. 2 we recall basic facts concerning the canonical deformed Poincare Hopf algebra and the corresponding quantum space-time provided in article [@chi]. The third section is devoted to the short review of Kerr black hole defined in commutative (classical) space. In Sect. 4 we analyze the effect of space-time noncommutativity (\[minkowski\]) on the Kerr metric. The final remarks are presented in the last section. Canonically deformed Minkowski space-time ========================================= In this section we recall basic facts associated with $\theta$-deformed Poincare Hopf algebra ${\cal U}_{\theta}({P})$ and with the corresponding canonically deformed quantum space-time [@chi]. Firstly, it should be noted that such objects can be get by so-called twist procedure, in which the algebraic sector of Hopf structure ${\cal U}_{\theta}({P})$ remains undeformed, i.e. it takes the form $$\begin{aligned} &&\left[\; M_{\mu \nu },M_{\rho \sigma }\;\right] =i\left( \eta _{\mu \sigma }\,M_{\nu \rho }-\eta _{\nu \sigma }\,M_{\mu \rho }+\eta _{\nu \rho }M_{\mu \sigma }-\eta _{\mu \rho }M_{\nu \sigma }\right) \;, \notag \\ %&~~& \cr &&\left[\; M_{\mu \nu },P_{\rho }\;\right] =i\left( \eta _{\nu \rho }\,P_{\mu }-\eta _{\mu \rho }\,P_{\nu }\right) \;\;\;,\;\;\; \left[\; P_{\mu },P_{\nu }\;\right] =0\;. \label{poincare}\end{aligned}$$ Besides, the coproduct of considered algebra is given by $$\Delta _{0}(a) \to \Delta _{\theta}(a) = \mathcal{F}_{\theta }\circ \,\Delta _{0}(a)\,\circ \mathcal{F}_{\theta }^{-1}\;,\label{twist}$$ where $${\cal F}_{\theta}=\exp \left[\;{\frac{{ i}}{2}\theta^{\mu\nu}P_\mu\otimes P_\nu}\;\right]\;,\label{factor}$$ denotes the canonical twist factor, while $\Delta _{0}(a) = a \otimes 1 + 1 \otimes a$. Consequently, using (\[poincare\])-(\[factor\]) we get $$\begin{aligned} \Delta_{\theta}(P_\mu) &=& P_\mu\otimes 1\ +\ 1\otimes P_\mu\;, \nonumber \\ \Delta_\theta (M_{\mu\nu}) &=& M_{\mu\nu}\otimes 1\ +\ 1\otimes M_{\mu\nu} - \frac{1}{2}({\theta_\mu}^\rho P_\nu-{\theta_\nu}^\rho P_\mu)\otimes P_\rho\ + \label{cop}\\ &+&\frac{1}{2}P_\rho\otimes({\theta_\mu}^\rho P_\nu-{\theta_\nu}^\rho P_\mu)\;. \nonumber\end{aligned}$$ The corresponding quantum Minkowski space-time is defined as the representation space (Hopf module) for Poincare Hopf algebra ${\cal U}_{\theta}({P})$. It is given by the following commutation relationes $$[\;{\hat x}_\mu,{\hat x}_\nu\;] =\ { i}\theta_{\mu\nu}\;,\label{minkowski}$$ and it can be extended to the whole algebra of momentum and position operators as follows $$[\;{\hat x}_\mu,{\hat x}_\nu\;] =\ { i}\theta_{\mu\nu}\;\;\;,\;\;\; [\;{\hat p}_\mu,{\hat p}_\nu\;] =0\;\;\;,\;\;\; [\;{\hat x}_\mu,{\hat p}_\nu\;]={ i}\eta_{\mu\nu}\;. \label{phase}$$ Of course, for deformation parameter $\theta$ running to zero the all above objects become classical. [Kerr black hole]{} =================== The Kerr metric describes the solution of field equation for rotating (with angular momentum $L$), uncharged and axially-symmetric massive (with mass $M$) object in empty space-time. In spherical coordinate system $(r,\varphi,\phi)$ it can be written as follows [@schw] $$\begin{aligned} c^{2} d\tau^{2} &=& \left( 1 - \frac{r_{s} r}{\rho^{2}} \right) c^{2} dt^{2} - \frac{\rho^{2}}{\Delta} dr^{2} - \rho^{2} d\varphi^{2} + \label{kerr}\\ &-& \left( r^{2} + \alpha^{2} + \frac{r_{s} r \alpha^{2}}{\rho^{2}} \sin^{2} \varphi \right) \sin^{2} \varphi \ d\phi^{2}+\frac{2r_{s} r\alpha \sin^{2} \varphi }{\rho^{2}} \, c \, dt \, d\phi\;,\nonumber\end{aligned}$$ with $r_s$ denoting the Schwarzschild radius $$r_{s} = \frac{2GM}{c^{2}}\;,\label{rsradius}$$ and with the length-scales $\alpha$, $\rho$ and $\Delta$ introduced for brevity $$\begin{aligned} \alpha &=& \frac{L}{Mc}\;,\label{alfa}\\ \rho^{2} &=& r^{2} + \alpha^{2} \cos^{2} \varphi\;,\label{ro}\\ \Delta &=& r^{2} - r_{s} r + \alpha^{2}\;.\label{delta}\end{aligned}$$ Of course, for parameter $\alpha$ approaching zero the above metric passes into well-known solution for Schwarzschild black hole [@schw]. It should be noted that the Kerr metric (\[kerr\]) has four (only two of them are physically relevant) surfaces on which it appears to be singular. First pair occurs when the purely radial component of metric goes to infinity. Then the solutions of corresponding quadratic equation $$\frac{1}{g_{rr}} = \frac{\Delta}{\rho^2} = 0\;,\label{equation1}$$ look as follows $$\begin{aligned} r_{\rm inner} &:=& r_{i+} = \frac{r_{s} + \sqrt{r_{s}^{2} - 4\alpha^{2}}}{2}\;,\label{sol1}\end{aligned}$$ and $$\begin{aligned} r_{i-} &=& \frac{r_{s} - \sqrt{r_{s}^{2} - 4\alpha^{2}}}{2}\;,\label{sol2}\end{aligned}$$ respectively. The first of them (physical one) describes the event horizon while the second solution (due to the fact that radius $r_{i+}$ is bigger than $r_{i-}$) has just unphysical properties.\ The second pair of singularities occurs when the purely temporal component of the Kerr metric changes sign. Again, solving a quadratic equation of the form $${g_{tt}} = \left( 1 - \frac{r_{s} r}{\rho^{2}} \right) = 0\;,\label{equation2}$$ we get $$\begin{aligned} r_{\rm outer} &:=& r_{o+} = \frac{r_{s} + \sqrt{r_{s}^{2} - 4\alpha^{2}\cos^2\varphi}}{2}\;,\label{sol3}\\ r_{o-} &=& \frac{r_{s} - \sqrt{r_{s}^{2} - 4\alpha^{2}\cos^2\theta}}{2}\;,\label{sol4}\end{aligned}$$ where only radius $r_{o+}$ remains physical $(r_{o+}>r_{i+}>r_{o-})$.\ One can observe that due to the $\cos^2 \varphi$ term in (\[sol3\]) the outer sphere touches the inner one (see (\[sol1\])) at the poles of rotation axis, where $\varphi$ equals 0 or $\pi$; the space between these two surfaces is called the ergosphere. Obviously, for parameter $\alpha$ running to zero we have $$\lim_{\alpha\to 0}r_{\rm inner} = \lim_{\alpha\to 0}r_{\rm outer} = r_s\;,\label{limit}$$ and the ergosphere disappears.\ Finally, it should be mentioned that in accordance with article [@bh1], [@bh2] the temperature and entropy of Kerr black hole are given by $$T = \frac{\hbar}{4\pi kc}\frac{\left(r_{i+}-r_{i-}\right)}{\left(r_{i+}^2+\alpha^2\right)}\;,\label{temp}$$ and $$S = \frac{kc}{4\hbar}{\left(r_{i+}^2+\alpha^2\right)}\;,\label{ent}$$ respectively, with symbol $k$ denoting the Boltzman’s constant. Noncommutative Kerr black hole ============================== Following the treatment proposed in papers [@treatment]-[@treatment2] we define the metric for Kerr black hole in $\theta$-deformed space-time with $\theta_{0i} =0$ by $$\begin{aligned} c^{2} d\tau^{2} &=& \left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{\rho}\hat{\rho}} \right) c^{2} dt^{2} - \frac{\hat{\rho}\hat{\rho}}{\hat{\Delta}} d\hat{r}d\hat{r} - \hat{\rho}\hat{\rho} d\varphi^{2} +\label{nonkerr}\\ &-&\left( \hat{r}\hat{r} + {\alpha}^2 + \frac{r_{s} \sqrt{\hat{r}\hat{r}} {\alpha}^2}{\hat{\rho}\hat{\rho}} \sin^{2} \varphi \right) \sin^{2} \varphi \ d\phi^{2} +\frac{2r_{s} \sqrt{\hat{r}\hat{r}}{\alpha} \sin^{2} \varphi }{\hat{\rho}\hat{\rho}} \, c \, dt \, d\phi\;, \nonumber\end{aligned}$$ where $$\begin{aligned} \hat{\rho}\hat{\rho} &=& \hat{r}\hat{r} + {\alpha}^2 \cos^{2} \varphi\;,\label{nonro}\\ \hat{\Delta} &=& \hat{r}\hat{r} - r_{s} \sqrt{\hat{r}\hat{r}} + {\alpha}^2\;,\label{nondelta}\end{aligned}$$ and where the components of $\hat{r} = \sqrt {\hat{x}_1^2 + \hat{x}_2^2 + \hat{x}_3^2}$ satisfy the commutation relations (\[minkowski\]). The solutions of $$\begin{aligned} \frac{\hat{\Delta}}{\hat{\rho}\hat{\rho}} &=& \frac{\hat{r}\hat{r} - r_{s} \sqrt{\hat{r}\hat{r}} + {\alpha}^2}{\hat{r}\hat{r} + {\alpha}^2 \cos^{2} \varphi} = 0\;, \label{nonequation1}\\ \left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{\rho}\hat{\rho}} \right) &=& \left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{r}\hat{r} + {\alpha}^2 \cos^{2} \varphi} \right) = 0\;,\label{nonequation2}\end{aligned}$$ are the singularities of the metric (\[nonkerr\]).\ In order to analyze the above system we represent the noncommutative varibles $({\hat x}_i, {\hat p}_i)$ in terms of classical phase space $({ x}_i, { p}_i)$ as (see e.g. [@kijanka], [@lukiluk2]) $${\hat x}_{i} = { x}_{i} - \frac{1}{2}\theta_{ij}p_j \;\;\;,\;\;\; {\hat p}_{i}= p_i\;, \label{rep}$$ where $$[\;x_i,x_j\;] = 0 =[\;p_i,p_j\;]\;\;\;,\;\;\; [\;x_i,p_j\;] ={i}\delta_{ij}\;. \label{classpoisson}$$ Then, the equations (\[nonequation1\]) and (\[nonequation2\]) take the form $$\frac{\hat{\Delta}}{\hat{\rho}\hat{\rho}} = \frac{\left(x_i - \frac{1}{2}\theta_{ij}p_j\right)\left(x_i - \frac{1}{2}\theta_{ik}p_k\right)-r_s\sqrt{\left(x_i - \frac{1}{2}\theta_{ij}p_j\right)\left(x_i - \frac{1}{2}\theta_{ik}p_k\right)}+\alpha^2} {\left(x_i - \frac{1}{2}\theta_{ij}p_j\right)\left(x_i - \frac{1}{2}\theta_{ik}p_k\right) +\alpha^2\cos^2 \varphi} =0\;, \label{classnonequation1}$$ and $$\left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{\rho}\hat{\rho}} \right) = \left( 1 - \frac{r_{s} \sqrt{\left(x_i - \frac{1}{2}\theta_{ij}p_j\right)\left(x_i - \frac{1}{2}\theta_{ik}p_k\right)}}{\left(x_i - \frac{1}{2}\theta_{ij}p_j\right)\left(x_i - \frac{1}{2}\theta_{ik}p_k\right) +\alpha^2\cos^2 \varphi} \right) =0 \;, \label{classnonequation2}$$ respectively. First of all, one should notice that for $\alpha$ running to zero the above conditions become the same, and we reproduce the noncommutative Schwarzschild black hole singularity equation given by [@treatment], [@treatment1] $$\frac{\hat{\Delta}}{\hat{\rho}\hat{\rho}} = \left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{\rho}\hat{\rho}} \right) = 1 - \frac{r_{s}}{ \sqrt{\hat{r}\hat{r}}} = 0\;. \label{toschwarz}$$ Further, we rewrite the formulas (\[classnonequation1\]) and (\[classnonequation2\]) as follows $$\begin{aligned} \frac{\hat{\Delta}}{\hat{\rho}\hat{\rho}} &=& \frac{\tilde{r}^2-r_s \tilde{r}+\alpha^2}{\tilde{r}^2+\alpha^2\cos^2\varphi}=0\;, \label{taylor1}\\ \left( 1 - \frac{r_{s} \sqrt{\hat{r}\hat{r}}}{\hat{\rho}\hat{\rho}} \right) &=& 1 - \frac{r_s\tilde{r}}{\tilde{r}^2+\alpha^2\cos^2\varphi}=0 \;,\label{taylor2}\end{aligned}$$ where $$\tilde{r}=\sqrt{r^2-\frac{1}{2}\vec{\theta}\vec{L}+\frac{1}{16}\left(\vec{p}\times\vec{\theta}\right)^2}\;\;\;,\;\;\;\vec{L}=\vec{x}\times\vec{p} \;\;\;\;\;{\rm and}\;\;\;\;\;\theta_{ij}=\frac{1}{2}\epsilon_{ijk} \theta_k\;,\label{rTilde}$$ as well as we find the corresponding solutions in the form $$\tilde{r}_\pm(a) = \frac{r_s\pm\sqrt{r_s^2-4a}}{2}\;,$$ with $a=\alpha^2$ and $a=\alpha^2\cos^2\varphi$ in the case of equations (\[taylor1\]) and (\[taylor2\]) respectively.\ Hence, in accordance with the formula (\[rTilde\]) we get $$r_\pm(a)=\frac{1}{2}\sqrt{\left(r_s\pm\sqrt{r_s^2-4 a}\right)^2+2\vec{\theta}\vec{L}-\frac{1}{4}\left(\vec{p}\times\vec{\theta}\right)^2}\;, \label{rTaylor12}$$ while in the limit of commutative space, one obtains $$\begin{aligned} r_{-}(\alpha^2) &=& r_{i-}\;,\label{limits2}\\ r_{+}(\alpha^2) &=& r_{i+} \;=\; r_{\rm inner}\;,\label{limits3}\\ r_{-}(\alpha^2\cos^2\varphi) &=& r_{o-}\;,\label{limits5}\\ r_{+}(\alpha^2\cos^2\varphi) &=& r_{o+} \;=\; r_{\rm outer}\;.\label{limits6}\end{aligned}$$ Consequently, due to the above limits we define the $\theta$-deformed ergosphere as the space occurring between radiuses $r_{+}(\alpha^2)$ and $r_{+}(\alpha^2\cos^2\varphi)$, i.e. as the region existing between deformed outer radius and “noncommutative” event horizon. Besides, one can observe that for $\alpha$ and $\theta_{ij}$ approaching zero, we reproduce the Schwarzschild event horizon $$\lim_{\alpha, \theta \to 0}r_{+}(\alpha^2) = \lim_{\alpha, \theta \to 0}r_{+}(\alpha^2\cos^2\varphi) = r_s \;,$$ with remaining radiuses vanishing.\ Finally, it should be mentioned that in accordance with formulas (\[temp\]) and (\[ent\]) as well as due to the limits (\[limits2\])-(\[limits6\]), the temperature and entropy of noncommutative Kerr black hole are given by $$\begin{aligned} T = \frac{\hbar}{4\pi kc}\frac{\left({r}_{+}(\alpha^2)-r_{-}(\alpha^2)\right)}{\left(r_{+}^2(\alpha^2)+\alpha^2\right)}\;,\label{ntemp}\end{aligned}$$ and $$\begin{aligned} S = \frac{kc}{4\hbar}{\left(r_{+}^2(\alpha^2)+\alpha^2\right)}\;,\label{nent}\end{aligned}$$ respectively. Final remarks ============= In this article we investigate the Kerr black hole defined on canonically deformed space-time. Particulary, we find the corresponding event horizon, the proper ergosphere, the temperature and the entropy of such deformed object. Besides, for parameters $\alpha$ and $\theta_{ij}$ approaching zero we reproduce the classical Schwarzschild black hole solution, while in the limit of commutative space we arrive to the case of undeformed Kerr black hole metric tensor; the presented studies has been performed with use of methods and techniques proposed in articles [@treatment], [@treatment1] and [@treatment2]. Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to thank M. Szczachor, P. Gusin and J. Lukierski for valuable discussions. This paper has been financially supported by Polish NCN grant No 2011/01/B/ST2/03354. [99]{} H.S. Snyder, Phys. Rev. 72, 68 (1947) S. Doplicher, K. Fredenhagen, J.E. Roberts, Phys. Lett. B 331, 39 (1994) A. Kempf and G. Mangano, Phys. Rev. D 55, 7909 (1997); hep-th/9612084 A. Connes, M.R. Douglas, A. Schwarz, JHEP 9802, 003 (1998); hep-th/9711162 N. Seiberg and E. Witten, JHEP 9909, 032 (1999); hep-th/9908142 S. Coleman, S.L. Glashow, Phys. Rev. D 59, 116008 (1999); hep-ph/9812418 R.J. Protheore, H. Meyer, Phys. Lett. B 493, 1 (2000); astro-ph/0005349 S. Zakrzewski, *“Poisson Structures on the Poincare group”*; q-alg/9602001 Y. Brihaye, E. Kowalczyk, P. Maslanka, *“Poisson-Lie structure on Galilei group”*; math/0006167 R. Oeckl, J. Math. Phys. 40 (1999) M. Chaichian, P.P. Kulish, K. Nashijima, A. Tureanu, Phys. Lett. B 604, 98 (2004) V.G. Drinfeld, Soviet Math. Dokl. 32, 254 (1985); Algebra i Analiz (in Russian), 1, Fasc. 6, p. 114 (1989) V. Chari, A. Pressley, *“A Guide to Quantum Groups”*, Cambridge University Press, Cambridge, 1994 S. Majid, *“Foundations of quantum group theory”*, Cambridge University Press, 2000 M. Daszkiewicz, Mod. Phys. Lett. A 23, 7 (2008), IFT-UWR-LV-420; arXiv: 0801.1206 \[hep-th\] J. Lukierski, A. Nowicki, H. Ruegg and V.N. Tolstoy, Phys. Lett. B 264, 331 (1991); J. Lukierski, A. Nowicki and H. Ruegg, Phys. Lett. B 293, 344 (1992) S. Giller, P. Kosinski, M. Majewski, P. Maslanka and J. Kunz, Phys. Lett. B 286, 57 (1992) J. Lukierski and M. Woronowicz, Phys. Lett. B 633, 116 (2006); hep-th/0508083 J. Lukierski, A. Nowicki, H. Ruegg and V.N. Tolstoy, J. Phys. A 27, 2389 (1994) P. Aschieri, L. Castellani, A.M. Scarfone, Eur. Phys. J. C 7, 159 (1999); q-alg/9709032 A. Deriglazov, JHEP 0303, 021 (2003); hep-th/0211105 A.E.F. Djemaï, *“On noncommutative classical mechanics”*; hep-th/0309034 J.M. Romero and J.D. Vergara, Mod. Phys. Lett. A 18, 1673 (2003); hep-th/0303064 J.M. Romero, J.A. Santiago, J.D. Vergara, Phys. Lett. A 310, 9 (2003); hep-th/0211165 Y. Miao, X. Wang, S. Yu, *“Classical mechanics on noncommutative space with Lie-algebraic structure”*; arXiv: 0911.5227 \[math-ph\] E. Harikumar, A.K. Kapoor, *“Newton equation on the kappa space-time and the Kepler problem”*; arXiv: 1003.4603 \[hep-th\] M. Daszkiewicz, C.J. Walczyk, Phys. Rev. D 77, 105008 (2008); 0802.3575 \[mat-ph\], \[hep-th\] M. Chaichian, M.M. Sheikh-Jabbari, A. Tureanu, Phys. Rev. Lett. 86, 2716 (2001); hep-th/0010175 M. Chaichian, M.M. Sheikh-Jabbari, A. Tureanu, Euro Phys. J. C 36, 251 (2004); hep-th/0212259 M. Chaichian, A. Demichev, P. Presnajder, M.M. Sheikh-Jabbari, A. Tureanu, Nucl. Phys. B 11, 383 (2001) A. Kijanka, P. Kosinski, Phys. Rev. D 70, 127702 (2004); hep-th/0407246 M. Daszkiewicz, C.J. Walczyk, Acta Phys. Pol. B 40, 293 (2009); arXiv: 0812.1264 \[hep-th\] G. Amelino-Camelia, Phys. Lett. B 510, 255 (2001); hep-th/0012238 G. Amelino-Camelia, Mod. Phys. Lett. A 17, 899 (2002); gr-qc/0204051 M. Maggiore, Phys. Rev. D 49, 5182 (1994); hep-th/9305163 S.M. Rama, Phys. Lett. B 519, 103 (2001); hep-th/0107255 R.J. Szabo, Phys. Rept. 378, 207 (2003); hep-th/0109162 J. Lukierski, H. Ruegg, W.J. Zakrzewski, Ann. Phys. 243, 90 (1995); hep-th/9312153 J. Lukierski, P. Stichel, W.J. Zakrzewski, Ann. Phys. 260, 224 (1997); hep-th/9612017 F. Nasseri, Gen. Rel. Grav. [37]{}, 2223 (2005); hep-th/0508051 F. Nasseri, Int. J. Mod. Phys. D [15]{}, 1113 (2006); hep-th/0508122 F. Nasseri, S.A. Alavi, *“Reissner Nordstrom black holes in noncommutative spaces”*; private notes P. Nicolini, A. Smailagic, E. Spallucci, Phys. Lett. B 632, 547 (2006); gr-qc/0510112 Y.G. Miao, Z. Xue and S.J. Zhang, Int. J. Mod. Phys. D [21]{}, 1250018 (2012); arXiv: 1102.0074 \[hep-th\] S.A. Hughes, *“Trust but verify: The Case for astrophysical black holes”*, e-Conf C [0507252]{}, L006 (2005); hep-ph/0511217 J.D. Beckenstein, Phys. Rev. D 7, 2333 (1973) S.W. Hawking, Commun. Math. Phys. 43, 199 (1975) A. Kijanka, P. Kosinski, Phys. Rev. D 70, 127702 (2004); hep-th/0407246 [^1]: For earlier studies see [@lukiluk1] and [@lukiluk2]. [^2]: It should be noted, however, that the results obtained by us make full solutions in deformation parameter $\theta$ of singularity equations, contrary to the results of articles [@treatment]-[@treatment2].
{ "pile_set_name": "ArXiv" }
--- abstract: 'The role of uncertainty quantification (UQ) in deep learning has become crucial with growing use of predictive models in high-risk applications. Though a large class of methods exists for measuring deep uncertainties, in practice, the resulting estimates are found to be poorly calibrated, thus making it challenging to translate them into actionable insights. A common workaround is to utilize a separate recalibration step, which adjusts the estimates to compensate for the miscalibration. Instead, we propose to repurpose the heteroscedastic regression objective as a surrogate for calibration and enable any existing uncertainty estimator to be inherently calibrated. In addition to eliminating the need for recalibration, this also regularizes the training process. Using regression experiments, we demonstrate the effectiveness of the proposed heteroscedastic calibration with two popular uncertainty estimators.' address: '$^{\dagger}$Arizona State University, $^{\ddagger}$ Lawrence Livermore National Labs' bibliography: - 'refs.bib' title: Heteroscedastic Calibration of Uncertainty Estimators in Deep Learning --- uncertainty quantification, deep uncertainties, calibration, heteroscedastic regression, dropout. Introduction {#sec:intro} ============ Background {#sec:background} ========== Proposed Approach {#sec:approach} ================= Empirical Results {#sec:experiments} ================= Conclusions {#sec:conclusions} ===========
{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the theory of functions of several complex variables, we prove that if an analytic function in several variables satisfies a system of $q$-partial differential equations, then, it can be expanded in terms of the product of the Rogers-Szegő polynomials. This expansion theorem allows us to develop a general method for proving $q$-identities. A general $q$-transformation formula is derived, which implies Watson’s $q$-analog of Whipple’s theorem as a special case. A multilinear generating function for the Rogers-Szegő polynomials is given. The theory of $q$-exponential operator is revisited.' address: 'East China Normal University, Department of Mathematics, 500 Dongchuan Road, Shanghai 200241, P. R. China' author: - 'Zhi-Guo Liu' title: 'On the $q$-partial differential equations and $q$-series' --- [^1] [^2] [^3] Introduction ============ Throughout the paper, we use the standard $q$-notations. For $0<q<1$, we define the $q$-shifted factorials as $$(a; q)_0=1,\quad (a; q)_n=\prod_{k=0}^{n-1}(1-aq^k), \quad (a; q)_\infty=\prod_{k=0}^\infty (1-aq^k);$$ and for convenience, we also adopt the following compact notation for the multiple $q$-shifted factorial: $$(a_1, a_2,...,a_m;q)_n=(a_1;q)_n(a_2;q)_n ... (a_m;q)_n,$$ where $n$ is an integer or $\infty$. The basic hypergeometric series ${_r\phi_s}$ is defined as $${_r\phi_s} \left({{a_1, a_2, ..., a_{r}} \atop {b_1, b_2, ..., b_s}} ; q, z \right) =\sum_{n=0}^\infty \frac{(a_1, a_2, ..., a_{r};q)_n} {(q, b_1, b_2, ..., b_s ;q)_n}\left((-1)^n q^{n(n-1)/2}\right)^{1+s-r} z^n.$$ For any function $f(x)$ of one variable, the $q$-derivative of $f(x)$ with respect to $x,$ is defined as $$\mathcal{D}_{q,x}\{f(x)\}=\frac{f(x)-f(qx)}{x},$$ and we further define $\mathcal{D}_{q,x}^{0} \{f\}=f,$ and for $n\ge 1$, $\mathcal{D}_{q, x}^n \{f\}=\mathcal{D}_{q, x}\{\mathcal{D}_{q, x}^{n-1}\{f\}\}.$ For any nonnegative integer $n,$ we have the higher-order $q$-derivative formula $$\mathcal{D}^n_{q, x} \left\{\frac{1}{(sx; q)_\infty}\right\}=\frac{s^n}{(sx; q)_\infty}, \label{qd:eqn1}$$ which is the case $t=0$ of the following general higher-order $q$-derivative formula: $$\mathcal{D}^n_{q, x} \left\{\frac{(tx; q)_\infty}{(sx; q)_\infty}\right\} =s^n(t/s; q)_n \frac{(q^ntx; q)_\infty}{(sx; q)_\infty}. \label{qd:eqn2}$$ \[qpde\] A $q$-partial derivative of a function of several variables is its $q$-derivative with respect to one of those variables, regarding other variables as constants. The $q$-partial derivative of a function $f$ with respect to the variable $x$ is denoted by $\partial_{q, x}\{f\}$. The Gaussian binomial coefficients also called the $q$-binomial coefficients are $q$-analogs of the binomial coefficients, which are given by $${n\brack k}_q=\frac{(q; q)_n}{(q; q)_k(q; q)_{n-k}}. \label{qd:eqn3}$$ Now we introduce the definition of the Rogers–Szegő polynomials which were first studied by Rogers [@Rogers1893] and then by Szegő [@Szeg]. \[rspolydefn\]With the $q$-binomial coefficients be defined as in (\[qd:eqn3\]), the Rogers-Szegő polynomials are defined by $$h_n(x, y|q)=\sum_{k=0}^n {n\brack k}_q x^k y^{n-k}.$$ If $q$ is replaced by $q^{-1}$ in the Rogers-Szegő polynomials, we can obtain the Stieltjes-Wigert polynomials (see, for example, [@Carlitz; @Szeg]). \[swpolydefn\] The Stieltjes-Wigert polynomials are defined by $$g_n(x, y|q)=h_n(x, y|q^{-1})=\sum_{k=0}^n {n\brack k}_q q^{k(k-n)}x^k y^{n-k}.$$ \[gefunpp\] $h_n(x, y|q)$ and $g_n(x, y|q)$ satisfy the identities $$\partial_{q, x} \{h_n(x, y|q)\}=\partial_{q, y} \{h_n(x, y|q)\}=(1-q^n)h_{n-1}(x, y|q), \label{rseqn3}$$ $$\partial_{q^{-1}, x} \{g_n(x, y|q)\}=\partial_{q^{-1}, y} \{g_n(x, y|q)\}=(1-q^{-n})g_{n-1}(x, y|q). \label{rseqn4}$$ Using the identity, $\partial_{q, x} \{ x^k\}=(1-q^k) x^{k-1},$ we immediately find that $$\partial_{q, x}\left\{h_n(x, y|q)\right\} =\sum_{k=1}^n {n\brack k}_q (1-q^k) x^{k-1} y^{n-k}.$$ In the same way, using the identity, $\partial_{q, y} \{ y^{n-k}\}=(1-q^{n-k}) y^{n-k-1},$ we deduce that $$\partial_{q, y}\{h_n(x, y|q)\} =\sum_{k=0}^{n-1} {n\brack k}_q (1-q^{n-k}) x^{k} y^{n-k-1}.$$ If we make the variable change $k+1 \to k$ in the right-hand side of the above equation, we can find that $$\partial_{q, y}\left\{h_n(x, y|q)\right\} =\sum_{k=1}^{n} {n\brack {k-1}}_q (\alpha; q)_{k} (1-q^{n-k+1}) x^{k-1} y^{n-k}.$$ From the definition of the $q$-binomial coefficients, it is easy to verify that $${n\brack k}_q (1-q^k)={n\brack {k-1}}_q (1-q^{n-k+1}).$$ Thus, the identity in (\[rseqn3\]) holds. In this same way, we can prove (\[rseqn4\]). This completes the proof of Proposition \[gefunpp\]. To explain our motivation of this paper, we begin with the following proposition. \[ppmotivation\] If $f(x, y)$ is a two variables analytic function in a neighbourhood of $(0, 0) \in \mathbb{C}^2$, satisfying the partial differential equation $f_x(x,y)=f_y(x, y),$ then, we have $f(x, y)=f(x+y, 0).$ Now we begin to solve the partial differential equation in the above proposition. From the theory of two complex variables, we may assume that near $(x, y)=(0, 0),$ $$f(x, y)=\sum_{n=0}^\infty A_n(x)y^n.$$ If this is substituted into $ f_x(x,y)=f_y(x, y)$, we immediately conclude that $$\sum_{n=0}^\infty A_n'(x)y^n=\sum_{n=0}^\infty (n+1)A_{n+1}(x)y^n.$$ Equating the coefficients of $y^n$ on both sides of the above equation, we find that for each integer $n\ge 1, A_n(x)={A_{n-1}'(x)}/{n}.$ By iteration, we deduce that $A_n(x)={A_0^{(n)}(x)}/{n!}.$ It is obvious that $A_0(x)=f(x, 0).$ Using the Taylor expansion, we deduce that $$f(x, y)=\sum_{n=0}^\infty \frac{f^{(n)}(x, 0)}{n!}y^n=f(x+y, 0),$$ which completes the proof of Proposition \[ppmotivation\]. In order to find the $q$-extension of Proposition \[ppmotivation\], we are led to the following proposition. \[qppmotivation\] If $f(x,y)$ is a two-variable analytic function at $(0,0)\in \mathbb{C}^2$, then, we have - $f$ can be expanded in terms of $h_n(x, y|q)$ if and only if $f$ satisfies the $q$-partial differential equation $ \partial_{q, x}\{f\}=\partial_{q, y}\{f\}. $ - $f$ can be expanded in terms of $g_n(x, y|q)$ if and only if $f$ satisfies the $q$-partial differential equation $ \partial_{q^{-1}, x}\{f\}=\partial_{q^{-1}, y}\{f\}. $ This proposition can be extended to the following more general expansion theorem for the analytic functions in several variables, which is the main result of this paper. \[mainthmliu\] If $f(x_1,y_1, \ldots, x_k, y_k)$ is a $2k$-variable analytic function at $(0,0, \cdots, 0)\in \mathbb{C}^{2k}$, then, we have - $f$ can be expanded in terms of $h_{n_1}(x_1, y_1|q)\cdots h_{n_k}(x_k, y_k|q)$ if and only if $f$ satisfies the $q$-partial differential equations $ \partial_{q, x_j}\{f\}=\partial_{q, y_j}\{f\}~\text{for}~j=1, 2, \ldots, k. $ - $f$ can be expanded in terms of $g_{n_1}(x_1, y_1|q)\cdots g_{n_k}(x_k, y_k|q)$ if and only if $f$ satisfies the $q$-partial differential equation $ \partial_{ q^{-1}, x_j}\{f\}=\partial_{q^{-1}, y_j}\{f\}~\text{for}~j=1, 2, \ldots, k. $ Proposition \[qppmotivation\] is the special case $k=1$ of Theorem \[mainthmliu\]. This theorem is useful in $q$-series, which allows us to develop a general method for proving $q$-identities. Many applications of this expansion theorem to $q$-series are discussed in this paper. To determine if a given function is an analytic functions in several complex variables, we often use the following theorem (see, for example, [@Taylor p. 28]). \[hartogthm\] [(Hartog’s theorem).]{} If a complex valued function $f(z_1, z_2, \ldots, z_n)$ is holomorphic (analytic) in each variable separately in a domain $U\in\mathbb{C}^n,$ then, it is holomorphic (analytic) in $U.$ Proof of the expansion theorem ============================== In order to prove Theorem \[mainthmliu\], we need the following fundamental property of several complex variables (see, for example, [@Malgrange p. 5, Proposition  1], [@Range p. 90]). \[mcomplexpp\] If $f(x_1, x_2, \ldots, x_k)$ is analytic at the origin $(0, 0, \ldots, 0)\in \mathbb{C}^k$, then, $f$ can be expanded in an absolutely convergent power series, $$f(x_1, x_2, \ldots, x_k)=\sum_{n_1, n_2, \ldots, n_k=0}^\infty \alpha_{n_1, n_2, \ldots, n_k} x_1^{n_1} x_2^{n_2}\cdots x_k^{n_k}.$$ Now we begin to prove Proposition \[qppmotivation\] using Proposition \[mcomplexpp\]. The proofs of (i) and (ii) are similar, so we only prove (i). Since $f$ is analytic at $(0, 0),$ from Proposition \[mcomplexpp\], we know that $f$ can be expanded in an absolutely convergent power series in a neighborhood of $(0, 0)$. Thus there exists a sequence $\{\alpha_{m, n}\}$ independent of $x$ and $y$ such that $$f(x, y)=\sum_{m, n}^\infty \alpha_{m, n} x^my^n=\sum_{n=0}^\infty y^n \left\{\sum_{m=0}^\infty \alpha_{m, n} x^m\right\}. \label{pthm:eqn1}$$ Substituting this into the $q$-partial differential equation $\partial_{q, x}\{f(x, y)\}=\partial_{q, y}\{f(x, y)\}$ and using the fact $\partial_{q, y}\{y^n\}=(1-q^n)y^{n-1},$ we find that $$\sum_{n=0}^\infty y^n \partial_{q, x}\left\{\sum_{m=0}^\infty \alpha_{m, n} x^m\right\} =\sum_{n=1}^\infty (1-q^n)y^{n-1}\left\{\sum_{m=0}^\infty \alpha_{m, n} x^m\right\}.$$ Equating the coefficients of $y^{n-1}$ on both sides of the above equation, we easily deduce that $$\sum_{m=0}^\infty \alpha_{m, n} x^m=\frac{1}{1-q^n} \partial_{q, x}\left\{\sum_{m=0}^\infty \alpha_{m, n-1} x^m\right\}.$$ Iterating the above equation $(n-1)$ times, we conclude that $$\sum_{m=0}^\infty \alpha_{m, n} x^m=\frac{1}{(q; q)_n} \partial^n_{q, x}\left\{\sum_{m=0}^\infty \alpha_{m, 0} x^m\right\}.$$ With the help of the identity, ${\partial}^n_{q, x}\{x^m\}=(q; q)_m x^{m-n}/(q; q)_{m-n}$, we obtain $$\sum_{m=0}^\infty \alpha_{m, n} x^m=\sum_{m=n}^\infty \alpha_{m, 0}{m\brack n}_q x^{m-n}.$$ Noting that the series in (\[pthm:eqn1\]) is absolutely convergent, substituting the above equation into (\[pthm:eqn1\]) and interchanging the order of the summation, we deduce that $$f(x, y)=\sum_{m=0}^\infty \alpha_{m, 0} \sum_{n=0}^m {m \brack n}_qy^n x^{m-n} =\sum_{m=0}^\infty \alpha_{m, 0}h_m(x, y|q),$$ since $h_n(x, y|q)$ is symmetric in $x$ and $y$. Conversely, if $f(x, y)$ can be expanded in terms of $h_n(x, y|q), $ then using (\[rseqn3\]), we find that $\partial_{q, x}\{f(x, y)\}=\partial_{q, y}\{f(x, y)\}.$ This completes the proof of Proposition \[qppmotivation\]. [It can be shown that Proposition \[qppmotivation\] is equivalent to [@Liu2010 Theorem 4], but the proof in [@Liu2010] is less rigorous. This paper may be viewed as an improved of improved version of [@Liu2010].]{} Now we begin to prove Theorem \[mainthmliu\] by using Proposition \[qppmotivation\] and mathematical induction. The proof of (ii) is similar to that of (i), so we only prove (i). From Proposition \[qppmotivation\] we conclude that the theorem holds when $k=1.$ Now, we assume that the theorem is true for the case $k-1$ and consider the case $k$. If we regard $f(x_1, y_1, \ldots, x_k, y_k)$ as a function of $x_1$ and $y_1, $ then $f$ is analytic at $(0, 0)$ and satisfies $\partial_{q, x_1}\{f\}=\partial_{q, y_1}\{f\}.$ Thus from (i) in Proposition \[qppmotivation\], there exists a sequence $\{c_{n_1}(x_2, y_2, \ldots, x_k, y_k)\}$ independent of $x_1$ and $y_1$ such that $$f(x_1, y_1, \ldots, x_k, y_k)=\sum_{n_1=0}^\infty c_{n_1}(x_2, y_2, \ldots, x_k, y_k)h_{n_1}(x_1, y_1|q). \label{pthm:eqn2}$$ Setting $y_1=0$ in the above equation and using $h_{n_1}(x_1, 0|q)=x_1^{n_1},$ we obtain $$f(x_1, 0, x_2, y_2, \ldots, x_k, y_k)=\sum_{n_1=0}^\infty c_{n_1}(x_2, y_2, \ldots, x_k, y_k)x_1^{n_1}.$$ Using the Maclaurin expansion theorem, we immediately deduce that $$c_{n_1}(x_2, y_2, \ldots, x_k, y_k)=\frac{\partial^{n_1} f(x_1, 0, x_2, y_2, \ldots, x_k, y_k)}{{n_1}!\partial {x_1}^{n_1}}\Big|_{x_1=0}$$ Since $f(x_1, y_1, \ldots, x_k, y_k)$ is analytic near $(x_1, y_1, \ldots, x_k, y_k)=(0, \ldots, 0)\in \mathbb{C}^{2k},$ from the above equation, we know that $c_{n_1}(x_2, y_2, \ldots, x_k, y_k)$ is analytic near $(x_2, y_2, \ldots, x_k, y_k)=(0, \ldots, 0)\in \mathbb{C}^{2k-2}.$ Combining (\[pthm:eqn2\]) with (i) in Theorem \[mainthmliu\], we find, for $j=2, \ldots k$, that $$\begin{aligned} &\sum_{n_1=0}^\infty \partial_{q, x_j}\{c_{n_1}(x_2, y_2, \ldots, x_k, y_k)\} h_{n_1}(x_1, y_1|q)\\ &=\sum_{n_1=0}^\infty \partial_{q, y_j}\{c_{n_1}(x_2, y_2, \ldots, x_k, y_k)\} h_{n_1}(x_1, y_1|q).\end{aligned}$$ By equating the coefficients of $h_{n_1}(x_1, y_1|q)$ in the above equation, we find that for $j=2, \ldots, k,$ $$\partial_{q, x_j}\{c_{n_1}(x_2, y_2, \ldots, x_k, y_k)\} =\partial_{q, y_j}\{c_{n_1}(x_2, y_2, \ldots, x_k, y_k)\}.$$ Thus by the inductive hypothesis, there exists a sequence $\{\alpha_{n_1, n_2, \ldots, n_k}\}$ independent of $x_2, y_2, \ldots, x_k, y_k$ (of course independent of $x_1$ and $y_1$) such that $$c_{n_1}(x_2, y_2, \ldots, x_k, y_k)=\sum_{n_2, \ldots, n_k=0}^\infty \alpha_{n_1, n_2, \ldots, n_k} h_{n_2}(x_2, y_2|q)\ldots h_{n_k}(x_k, y_k|q).$$ Substituting this equation into (\[pthm:eqn2\]), we find that $f$ can be expanded in terms of $h_{n_1}(x_1, y_1|q)\cdots h_{n_k}(x_k, y_k|q).$ Conversely, if $f$ can be expanded in terms of $h_{n_1}(x_1, y_1|q)\cdots h_{n_k}(x_k, y_k|q),$ then using (\[rseqn3\]), we find that $\partial_{q, x_j}\{f\}=\partial_{q, y_j}\{f\}$ for $j=1, 2, \ldots, k.$ This completes the proof of Theorem \[mainthmliu\]. The generating functions of the Rogers-Szegő polynomials and the Stieltjes-Wigert polynomials ============================================================================================= \[gefunthm\] If $h_n(x, y|q)$ and $g_n(x, y|q)$ are given by Definitions \[rspolydefn\] and \[swpolydefn\], then, we have $$\sum_{n=0}^\infty h_n (x, y | q) \frac{t^n}{(q; q)_n}=\frac{1}{(xt, yt; q)_\infty},~\max\{|xt|, |yt|\}<1, \label{rseqn1}$$ $$\sum_{n=0}^\infty (-1)^n q^{n(n-1)/2} g_n(x, y|q) \frac{t^n}{(q; q)_n} =(xt, yt; q)_\infty. \label{rseqn2}$$ We only prove (\[rseqn1\]). The proof of (\[rseqn2\]) is similar, so is omitted. It is well-known that $1/(xt; q)_\infty$ is an analytic function of $x$ for $|xt|<1,$ and $1/(yt; q)_\infty$ is an analytic function of $y$ for $|yt|<1.$ Thus, $1/(xt, yt; q)_\infty$ is an analytic function of $x$ and $y$ for $\max\{|xt|, |yt|\}<1$. If we use $f(x, y)$ to denote the right-hand side of (\[rseqn1\]), then $f(x, y)$ is analytic near $(0, 0)\in \mathbb{C}^2.$ A direct computation shows that $$\partial_{q, x}\{f(x, y)\}= \partial_{q, y}\{f(x, y)\}=tf(x, y).$$ Thus by (i) in Proposition \[qppmotivation\], there exists a sequence $\{\alpha_n\}$ independent of $x$ and $y$ such that $$f(x, y)=\frac{1}{(xt, yt; q)_\infty}=\sum_{n=0}^\infty \alpha_n h_n(x, y|q).$$ Taking $y=0$ in the above equation, using $h_n(x, 0|q)=x^n$ and the $q$-binomial theorem, we obtain $$\frac{1}{(xt; q)_\infty}=\sum_{n=0}^\infty \frac{(xt)^n}{(q; q)_n}=\sum_{n=0}^\infty \alpha_n x^n.$$ Equating the coefficients of $x^n$ on both sides of the above equation, we deduce that $\alpha_n=t^n/(q; q)_n.$ Thus, we arrive at (\[rseqn1\]). We can also prove Theorem \[gefunthm\] by multiplying two copies of $q$-binomial theorem together. $q$-Mehler formulas for the Rogers-Szegő polynomials and the Stieltjes-Wigert polynomials ========================================================================================= Using Theorem \[mainthmliu\], we can derive easily the $q$-Mehler formulas for the Rogers-Szegő polynomials and the Stieltjes-Wigert polynomials. \[aqmehler\] If $\max\{|xut|, |xvt|, |yut|, |yvt|\}<1$, then, we have $$\sum_{n=0}^\infty h_n (x, y|q) h_n (u, v|q) \frac {t^n}{(q; q)_n} =\frac{(xyuvt^2; q)_\infty} {(xut, xvt, yut, yvt; q)_\infty}.$$ \[bqmehler\] If $|xyuvt^2/q|<1$, then, we have the identity $$\sum_{n=0}^\infty(-1)^n g_n (x, y|q) g_n (u, v|q) \frac {q^{n(n-1)/2}t^n}{(q; q)_n} =\frac {(xut, xvt, yut, yvt; q)_\infty}{(xyuvt^2/q; q)_\infty}.$$ The $q$-Mehler formula for the Rogers-Szegő polynomials was first given by Rogers [@Rogers1893] in 1893 and later reproved by Carlitz [@Carlitz]. The $q$-Mehler formula for $g_n(x, y|q)$ was first proved by L. Carlitz [@Carlitz]. We first prove Theorem \[aqmehler\]. If we use $f(x, y)$ to denote the right-hand side of the equation in Theorem \[aqmehler\], it is obvious that $f(x, y)$ is analytic in $x$ and $y$ separately, so by Hartog’s theorem, we know that $f(x, y)$ is analytic at $(0, 0)$. Using the identity $(z; q)_\infty=(1-z)(qz; q)_\infty$ and a direct computation, we find that $$\partial_{q, x}\{f(x, y)\}=\partial_{q, y}\{f(x, y)\} =\frac{t(u+v)-xuvt^2-yuvt^2}{1-xyuvt^2} f(x, y).$$ Thus by (i) in Proposition \[qppmotivation\], there exists a sequence $\{\alpha_n\}$ independent of $x$ and $y$ such that $$\frac{(xyuvt^2; q)_\infty} {(xut, xvt, yut, yvt; q)_\infty}=\sum_{n=0}^\infty \alpha_n h_n(x, y|q). \label{meqn1}$$ Putting $y=0$ in the above equation, using the fact $h_n(x, 0|q)=x^n,$ and the generating function for $h_n$ in (\[rseqn1\]), we find that $$\sum_{n=0}^\infty \alpha_n x^n=\frac{1}{(xut, xvt; q)_\infty} =\sum_{n=0}^\infty h_n(u, v|q)\frac{(xt)^n}{(q; q)_n}.$$ Equating the coefficients of $x^n$ on both sides of the above equation, we deduce that $\alpha_n=h_n(u, v) t^n/(q; q)_n.$ Substituting this into (\[meqn1\]), we complete the proof of Theorem \[aqmehler\]. Now we turn to prove Theorem \[bqmehler\]. Denote the right-hand side of the equation in Theorem \[bqmehler\] by $g(x, y)$, then, by the identity, $(z; q)_\infty=(1-z)(qz; q)_\infty$ and a direct computation, we deduce that $$\partial_{q^{-1}, x}\{g(x, y)\}=\partial_{q^{-1}, y}\{g(x, y)\} =\frac{qt(u+v)-(x+y)uvt^2}{q^2-xyuvt^2} g(x, y).$$ Thus by (ii) in Proposition \[qppmotivation\], there exists a sequence $\{\beta_n\}$ independent of $x$ and $y$ such that $$\frac {(xut, xvt, yut, yvt; q)_\infty}{(xyuvt^2/q; q)_\infty}=\sum_{n=0}^\infty \beta_n g_n(x, y|q). \label{meqn2}$$ Setting $y=0$ in the above equation, using the fact $g_n(x, 0|q)=x^n,$ and the generating function for $g_n$ in (\[rseqn2\]), we find that $$\sum_{n=0}^\infty \beta_n x^n =(xut, xvt; q)_\infty =\sum_{n=0}^\infty (-1)^n q^{n(n-1)/2} g_n(u, v|q)\frac{(xt)^n}{(q; q)_n}.$$ It follows that $\beta_n=(-1)^n q^{n(n-1)/2}g_n(u, v|q) t^n/(q; q)_n.$ Substituting this into (\[meqn2\]), we complete the proof of Theorem \[bqmehler\]. Carlitz’s extension of the $q$-Mehler formula for the Rogers-Szegő polynomials =============================================================================== Carlitz’s extension of the $q$-Mehler formula for the Rogers-Szegő polynomials [@Carlitz1972 p. 96, Eq. (4.1)] is equivalent to the following theorem. In this section we will prove it using Proposition \[qppmotivation\]. \[Carlitzthm\] For $\max\{|aut|, |but|, |avt|, |bvt|\}<1,$ we have that $$\begin{aligned} \sum_{n=0}^\infty h_{n+k}(a, b|q) h_n(u, v|q) \frac{t^n}{(q; q)_n} &=\frac{(abuvt^2; q)_\infty}{(aut, but, avt, bvt; q)_\infty}\\ &\times \sum_{j=0}^k {k \brack j}_q \frac{b^j a^{k-j}(aut, avt; q)_j}{(abuvt^2; q)_j}.\end{aligned}$$ Differentiating $k$ times the generating function for $h_n(a, b|q)$ with respect to $t$, we have that $$\sum_{n=0}^\infty h_{n+k}(a, b|q) \frac{t^n}{(q; q)_n} =\frac{1}{(at, bt; q)_\infty} \sum_{j=0}^k {k \brack j}_q b^j a^{k-j}(at; q)_j. \label{careqn1}$$ If $t$ is replaced by $tu,$ we arrive at $$\sum_{n=0}^\infty h_{n+k}(a, b|q) \frac{u^nt^n}{(q; q)_n} =\frac{1}{(aut, but; q)_\infty} \sum_{j=0}^k {k \brack j}_q b^j a^{k-j}(aut; q)_j. \label{careqn2}$$ Denote the right hand side of the equation in Theorem \[Carlitzthm\] by $f(u, v).$ Then by a direct computation, we find that $$\begin{aligned} &\partial_{q, u}\{f(u, v)\}=\partial_{q, v}\{f(u, v)\}\\ &=\frac{(abuvt^2; q)_\infty}{(aut, but, avt, bvt; q)_\infty} \sum_{j=0}^k {k \brack j}_q \frac{b^j a^{k-j}(aut, avt; q)_j}{(abuvt^2; q)_j} {\left( }\frac{atq^j+bt-ab(u+v)t^2q^j}{1-abuvt^2q^j}{\right) }.\end{aligned}$$ Thus, by (i) in Proposition \[qppmotivation\], there exists a sequence $\{\alpha_n\}$ independent of $u$ and $v$ such that $$f(u, v)=\sum_{n=0}^\infty \alpha_n h_n(u, v|q). \label{careqn6}$$ Taking $v=0$ in the above equation and noting the definition of $f(u, v),$ we find that $$\frac{1}{(aut, but; q)_\infty} \sum_{j=0}^k {k \brack j}_q b^j a^{k-j}(aut; q)_j =\sum_{n=0}^\infty \alpha_n u^n.$$ Comparing this equation with (\[careqn2\]), we find that $\alpha_n=t^n h_{n+k}(a, b|q)/(q; q)_n.$ Thus we complete the proof of the theorem. An extension of Rogers’s summation =================================== The Rogers summation formula [@Gas+Rah p. 44] is one of the most important results for $q$-series, which can be stated in the following proposition. \[rogerspp\] For $|\alpha abc/q^2|<1,$ we have the summation $$\begin{aligned} &{_6 \phi_5} \left({{\alpha, q\sqrt{\alpha}, -q\sqrt{\alpha}, q/a, q/b, q/c} \atop{\sqrt{\alpha}, -\sqrt{\alpha},\alpha a, \alpha b, \alpha c}}; q, \frac{\alpha abc}{q^2}\right)\\ &=\frac{(\alpha q, \alpha ab/q, \alpha ac/q, \alpha bc/q; q)_\infty} {(\alpha a, \alpha b, \alpha c, \alpha abc/q^2; q)_\infty}. \nonumber\end{aligned}$$ In [@Liu2011], we give the following extension of the Rogers $_6\phi_5$ summation formula by using the operator method. \[rogersliuthm\] For $\max\{|\alpha \beta abc/q^2|, |\alpha \gamma abc/q^2|\}<1$, we have the identity $$\begin{aligned} &\sum_{n=0}^\infty \frac{(1-\alpha q^{2n})(\alpha, q/a, q/b, q/c; q)_n}{(q, \alpha a, \alpha b, \alpha c; q)_n} {\left( }\frac{\alpha abc}{q^2}{\right) }^n {_4 \phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop{q/a, q/b,\alpha \beta \gamma ab/q}}; q, q\right)\\ &=\frac{(\alpha, \alpha ac/q, \alpha bc/q, \alpha \beta ab/q, \alpha \gamma ab/q, \alpha \beta \gamma abc/q^2; q)_\infty} {(\alpha a, \alpha b, \alpha c, \alpha \beta abc/q^2, \alpha \gamma abc/q^2, \alpha \beta \gamma ab/q; q )_\infty}.\end{aligned}$$ Now we will use Proposition \[qppmotivation\] to give a simple proof of the above theorem. Using Watson’s $q$-analogue of Whipple’s theorem, the Rogers $_6\phi_5$ summation formula and Tannery’s theorem, we can obtain the asymptotic formula (see [@Liu2011 p. 1406 ] for the details) for $n\to \infty$, $${_4 \phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop{q/a, q/b,\alpha \beta \gamma abc/q}}; q, q\right) \sim \frac{(\gamma, q/a\beta, q/b\beta, ab\alpha \gamma/q; q)_\infty \beta^n} {(q/a, q/b, \gamma/\beta, ab\alpha \beta \gamma/q; q)_\infty}.$$ Thus, by the ratio test, we know that the left-hand side of the equation in Theorem \[rogersliuthm\] converges to an analytic function of $\beta$ for $|\alpha \beta abc/q^2|<1$. It is obvious that the left-hand side of the equation in Theorem \[rogersliuthm\] is symmetric in $\beta$ and $\gamma,$ so the left-hand side of the equation in Theorem \[rogersliuthm\] also converges to an analytic function of $\gamma$ for $|\alpha \gamma abc/q^2|<1$. It follows that the left-hand side of the equation in Theorem \[rogersliuthm\] is analytic near $(\beta, \gamma)=(0, 0).$ For simplicity, we temporarily introduce $A_n$ and $f_k(\alpha, \beta)$ as follows $$\begin{aligned} &A_n:=\frac{(1-\alpha q^{2n})(\alpha, q/a, q/b, q/c; q)_n}{(q, \alpha a, \alpha b, \alpha c; q)_n} {\left( }\frac{\alpha abc}{q^2}{\right) }^n, \\ &f_k(\beta, \gamma):=\frac{(\alpha \beta \gamma ab q^k/q; q)_\infty}{(\beta q^k, \gamma q^k, \alpha \beta ab/q, \alpha \gamma ab/q; q)_\infty}.\end{aligned}$$ Using $f_0(\beta, \gamma)$ to multiply the left-hand side of the equation in Theorem \[rogersliuthm\], we find that $$\begin{aligned} f_0(\beta, \gamma) \sum_{n=0}^\infty {_4 \phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop{q/a, q/b,\alpha \beta \gamma abc/q}}; q, q\right)A_n\\ =\sum_{n=0}^\infty A_n \sum_{k=0}^n \frac{(q^{-n}, \alpha q^n; q)_k q^k}{(q, q/a, q/b; q)_k}f_k(\beta, \gamma).\end{aligned}$$ It is obvious that $f_0(\beta, \gamma)$ is analytic near $(\beta, \gamma)=(0, 0)$. Thus, the right-hand side of the above equation is also analytic near $(\beta, \gamma)=(0, 0)$. By a direct computation, we easily find that $$\partial_{q, \beta}\{f_k(\beta, \gamma)\}=\partial_{q, \gamma}\{f_k(\beta, \gamma)\} =f_k(\beta, \gamma)\left(\frac{\alpha ab+q^{k+1}-(\beta+\gamma)\alpha abq^k}{q-\alpha \beta \gamma abq^k}\right).$$ Thus, there exists a sequence $\{B_n\}$ independent of $\beta$ and $\gamma$ such that $$f_0(\beta, \gamma) \sum_{n=0}^\infty {_4 \phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop{q/a, q/b,\alpha \beta \gamma abc/q}}; q, q\right)A_n =\sum_{n=0}^\infty B_n h_n(\beta, \gamma|q).\label{rleqn1}$$ Setting $\gamma=0$ in the above equation and using $h_n(\beta, 0|q)=\beta^n$, we obtain $$\frac{1}{(\beta, \alpha \beta ab/q; q)_\infty}\sum_{n=0}^\infty {_3 \phi_2} \left({{q^{-n}, \alpha q^n, \beta} \atop{q/a, q/b}}; q, q\right)A_n =\sum_{n=0}^\infty B_n \beta^n.$$ Ismail, Rahman and Suslov [@IsmailRS p. 559, Theorem 5.1] have proved that $$\begin{aligned} \sum_{n=0}^\infty {_3 \phi_2} \left({{q^{-n}, \alpha q^n, \beta} \atop{q/a, q/b}}; q, q\right)A_n =\frac{(\alpha, \alpha ac/q, \alpha bc/q, \alpha \beta ab/q; q)_\infty} {(\alpha a, \alpha b, \alpha c, \alpha \beta abc/q^2; q)_\infty}.\end{aligned}$$ Comparing the above two equations, we are led to the following identity: $$\sum_{n=0}^\infty B_n \beta^n=\frac{(\alpha, \alpha ac/q, \alpha bc/q; q)_\infty} {(\alpha a, \alpha b, \alpha c, \beta, \alpha \beta abc/q^2; q)_\infty}.$$ Using the generating function for $h_n$ in Theorem \[gefunthm\], we immediately obtain $$\frac{1}{(\beta, \alpha \beta abc/q^2; q)_\infty}=\sum_{n=0}^\infty \frac{\beta^n}{(q; q)_n} h_n(1, \alpha abc/q^2|q).$$ Hence we have $$B_n=\frac{(\alpha, \alpha ac/q, \alpha bc/q; q)_\infty} {(\alpha a, \alpha b, \alpha c; q)_\infty}\frac{h_n(1, \alpha abc/q^2|q)}{(q; q)_n}.$$ Substituting the above equation into (\[rleqn1\]), we find that the right-hand side of (\[rleqn1\]) becomes $$\frac{(\alpha, \alpha ac/q, \alpha bc/q; q)_\infty} {(\alpha a, \alpha b, \alpha c; q)_\infty} \sum_{n=0}^\infty \frac{1}{(q; q)_n}h_n(\beta, \gamma|q)h_n(1, \alpha abc/q^2|q).$$ Using the $q$-Mehler formula for $h_n$ in Theorem \[aqmehler\], we easily find that $$\sum_{n=0}^\infty \frac{1}{(q; q)_n}h_n(\beta, \gamma|q)h_n(1, \alpha abc/q^2|q)=\frac{(\alpha\beta\gamma abc/q^2; q)_\infty} {(\beta, \gamma, \alpha\beta abc/q^2, \alpha\gamma abc/q^2; q)_\infty}.$$ Combining the above two equations, we find that the right-hand side of (\[rleqn1\]) equals $$\frac{(\alpha, \alpha ac/q, \alpha bc/q, \alpha\beta\gamma abc/q^2; q)_\infty} {(\alpha a, \alpha b, \alpha c, \beta, \gamma, \alpha\beta abc/q^2, \alpha\gamma abc/q^2; q)_\infty}.$$ It follows that $$\begin{aligned} f_0(\beta, \gamma) \sum_{n=0}^\infty {_4 \phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop{q/a, q/b,\alpha \beta \gamma abc/q}}; q, q\right)A_n\\ =\frac{(\alpha, \alpha ac/q, \alpha bc/q, \alpha\beta\gamma abc/q^2; q)_\infty} {(\alpha a, \alpha b, \alpha c, \beta, \gamma, \alpha\beta abc/q^2, \alpha\gamma abc/q^2; q)_\infty},\end{aligned}$$ which is equivalent to the equation in Theorem \[rogersliuthm\]. This completes the proof of Theorem \[rogersliuthm\]. An extension of the Andrews-Askey integral =========================================== It is known that the Jackson $q$-integral can be defined as $$\int_{a}^b f(x)d_q x=(1-q)\sum_{n=0}^\infty [bf(bq^n)-af(aq^n)]q^n.$$ Using Ramanujan’s ${_1}\psi{_1}$ summation, Andrews and Askey [@Andrews+Askey] proved the following $q$-integral formula. \[AAintegralpp\] If there are no zero factors in the denominator of the integral, then, we have $$\int_{u}^v \frac{(qx/u, qx/v; q)_\infty}{(ax, bx; q)_\infty}d_q x =\frac{(1-q)v(q, u/v, qv/u, abuv; q)_\infty}{(au, bu, av, bv; q)_\infty}.$$ To extend the above $q$-integral, we introduce $H_k(a, b, u, v)$ defined by $$H_k(a, b, u, v)=\sum_{r=0}^k {k \brack r}_q\frac{(au, av; q)_r}{(abuv; q)_r}b^ra^{k-r}. \label{aaeqn1}$$ We extend the Andrews and Askey integral to the following more general integral. \[aaliuthm\] If there are no zero factors in the denominator of the integral and $\max\{|au|, |bu|, |cu|, |du|, |av|, |bv|, |cv|, |dv|\}<1,$ then, we have $$\begin{aligned} &\int_{u}^v \frac{(qx/u, qx/v; q)_\infty}{(ax, bx, cx, dx; q)_\infty}d_q x =\frac{(1-q)v(q, u/v, qv/u, abuv, cduv; q)_\infty}{(au, bu, cu, du, av, bv, cv, dv; q)_\infty}\\ &\qquad \qquad \qquad \qquad \times \sum_{k=0}^\infty \frac{q^{k(k-1)/2}(-uv)^k}{(q; q)_k}H_k(a, b, u, v)H_k(c, d, u, v).\end{aligned}$$ We temporarily use $I(a, b, c, d)$ to denote the $q$-integral in the left-hand side of the above equation. Writing $I(a, b, c, d)$ in the series form, we easily find that it is an analytic function of $a, b, c, d$ for $$\max\{|au|, |bu|, |cu|, |du|, |av|, |bv|, |cv|, |dv|\}<1.$$ It is easy to check that $I(a, b, c, d)$ satisfies the two partial differential equations $$\begin{aligned} \partial_{q, a}\{I\}=\partial_{q, b}\{I\}=\int_{u}^v \frac{x(qx/u, qx/v; q)_\infty}{(ax, bx, cx, dx; q)_\infty}d_q x,\\ \partial_{q, c}\{I\}=\partial_{q, d}\{I\}=\int_{u}^v \frac{x(qx/u, qx/v; q)_\infty}{(ax, bx, cx, dx; q)_\infty}d_q x.\end{aligned}$$ Thus, by (i) in Theorem \[mainthmliu\], there exists a sequence $\{\alpha_{m, n}\}$ such that $$\int_{u}^v \frac{(qx/u, qx/v; q)_\infty}{(ax, bx, cx, dx; q)_\infty}d_q x =\sum_{m, n=0}^\infty \alpha_{m, n} h_m(a, b|q)h_n(c, d|q). \label{aaeqn2}$$ Setting $b=d=0$ in the above equation and noting that $h_m(a, 0|q)=a^m$ and $h_n(c, 0|q)=c^n,$ we obtain $$\int_{u}^v \frac{(qx/u, qx/v; q)_\infty}{(ax, cx; q)_\infty}d_q x =\sum_{m, n=-\infty}^\infty \alpha_{m, n} a^mc^n.$$ Using the Andrews-Askey integral, we find that the above equation can be written as $$\sum_{m, n=0}^\infty \alpha_{m, n} a^mc^n=\frac{(1-q)v(q, u/v, qv/u, acuv; q)_\infty}{(au, cu, av, cv; q)_\infty}.$$ Using the generating function for $h_n$ in Theorem \[gefunthm\], we find that the above equation can be written as $$\begin{aligned} &\sum_{m, n=0}^\infty \alpha_{m, n} a^mc^n\\ &=(1-q)v(q, u/v, qv/u, acuv; q)_\infty \sum_{m, n=0}^\infty \frac{h_m(u, v|q)h_n(u, v|q)a^m c^n} {(q; q)_m(q; q)_n}.\end{aligned}$$ Equating the coefficients of $a^mc^n$ in the both sides of the above equation, we deduce that $$\begin{aligned} \alpha_{m, n}&=(1-q)v(q, u/v, qv/u; q)_\infty\\ &\times \sum_{k \ge 0}\frac{q^{k(k-1)/2}(-uv)^k h_{m-k}(u, v|q)h_{n-k}(u, v|q)}{(q; q)_k (q; q)_{m-k} (q; q)_{n-k}}.\end{aligned}$$ Substituting the above equation into (\[aaeqn2\]) and simplifying, we conclude that $$\begin{aligned} &\int_{u}^v \frac{(qx/u, qx/v; q)_\infty}{(ax, bx, cx, dx; q)_\infty}d_q x \label{aaeqn3}\\ &=(1-q)v(q, u/v, qv/u; q)_\infty \sum_{k=0}^\infty \frac{q^{k(k-1)/2}(-uv)^k}{(q; q)_k}A_kB_k, \nonumber\end{aligned}$$ where $A_k$ and $B_k$ are given by $$A_k=\sum_{m=0}^\infty \frac{h_{m+k}(a, b|q)h_m(u, v|q)}{(q; q)_m} \ \text{and} \ B_k=\sum_{n=0}^\infty \frac{h_{n+k}(c, d|q)h_n(u, v|q)}{(q; q)_n}.$$ Thus using the Carlitz formula in Theorem \[Carlitzthm\], we immediately find that $$\begin{aligned} A_k=\frac{(abuv; q)_\infty}{(au, bu, av, bv; q)_\infty}H_k(a, b, u, v) \\ B_k=\frac{(cduv; q)_\infty}{(cu, du, cv, dv; q)_\infty}H_k(c, d, u, v).\end{aligned}$$ Substituting the above two equations into (\[aaeqn3\]), we complete the proof of the theorem. The following $q$-integral formula is equivalent to Sears’s identity for the sum of two nonterminating balanced ${}_3\phi_2$ series, which was first noticed by Al-Salam and Verma [@SalamVerma] \[ssvthm\] If there are no zero factors in the denominator of the integral and $\max\{|au|, |av|, |bu|, |bv|, |cu|, |cv|\}<1$, then, we have $$\int_{u}^v \frac{(qx/u, qx/v, abcuvx; q)_\infty}{(ax, bx, cx; q)_\infty}d_q x =\frac{(1-q)v(q, u/v, qv/u, abuv, acuv, bcuv; q)_\infty}{(au, bu, cu, av, bv, cv; q)_\infty}.$$ Denote $f(a, c)$ as $$f(a, c)=\frac{1}{(abuv, bcuv; q)_\infty}\int_{u}^v \frac{(qx/u, qx/v, abcuvx; q)_\infty}{(ax, bx, cx; q)_\infty}d_q x.$$ By a direct computation, we find that $\partial_{q, a}\{f(a, c)\}=\partial_{q, c}\{f(a, c)\}$ equals $$\frac{1}{(abuv, bcuv; q)_\infty}\int_{u}^v (x+buv-(a+c)buvx)\frac{(qx/u, qx/v, qabcuvx; q)_\infty}{(ax, bx, cx; q)_\infty}d_q x.$$ Hence, by (i) in Proposition \[qppmotivation\], there exists a sequence $\{\alpha_n\}$ independent of $a$ and $c$ such that $$\frac{1}{(abuv, bcuv; q)_\infty}\int_{u}^v \frac{(qx/u, qx/v, abcuvx; q)_\infty}{(ax, bx, cx; q)_\infty}d_q x =\sum_{n=0}^\infty \alpha_n h_n(a, c|q). \label{aaeqn4}$$ Putting $c=0$ in the above equation, using $h_n(a, 0|q)=a^n$ and the Andrews-Askey integral, we find that $$\sum_{n=0}^\infty \alpha_n a^n= =\frac{(1-q)v(q, u/v, qv/u; q)_\infty }{(au, av, bu, bv; q)_\infty}.$$ It follows that $$\alpha_n =\frac{(1-q)v(q, u/v, qv/u; q)_\infty h_n(u, v|q)}{(q; q)_n( bu, bv; q)_\infty}.$$ Substituting the above equation into (\[aaeqn4\]) and then using the $q$-Mehler formula for $h_n$, we find that $$\begin{aligned} &\frac{1}{(abuv, bcuv; q)_\infty}\int_{u}^v \frac{(qx/u, qx/v, abcuvx; q)_\infty}{(ax, bx, cx; q)_\infty}d_q x\\ &=\frac{(1-q)v(q, u/v, qv/u; q)_\infty }{( bu, bv; q)_\infty}\sum_{n=0}^\infty \frac{h_n(u, v|q) h_n(a, c|q)}{(q; q)_n}\\ &=\frac{(1-q)v(q, u/v, qv/u, acuv; q)_\infty }{( au, av, bu, bv, cu, cv; q)_\infty}. \end{aligned}$$ Multiplying both sides of the above equation by $(abuv, bcuv; q)_\infty,$ we complete the proof of Theorem \[ssvthm\]. Generalizations of Ramanujan’s reciprocity formula =================================================== Ramanujan’s reciprocity formula (see, for example, [@BCYY]) can be stated as in the following proposition. \[ramppreciprocity\] [(Ramanujan’s reciprocity theorem )]{}. $$\begin{aligned} &v\sum_{n=0}^\infty (-1)^n \frac{q^{n(n+1)/2}(u/v)^n}{(cv; q)_n} -u\sum_{n=0}^\infty (-1)^n \frac{q^{n(n+1)/2}(v/u)^n}{(cu; q)_n}\label{rreqn1}\\ &\quad=\frac{(v-u)(q, qv/u, qu/v; q)_\infty}{(cu, cv; q)_\infty}.\nonumber\end{aligned}$$ By using the $q$-exponential operator to Ramanujan’s $_1\psi_1$ summation, we [@Liu2003 Theorem 6] proved the following reciprocity formula, which is equivalent to an identity of Andrews. This formula was used to give a simple evaluation of the Askey-Wilson integral [@Liu2009]. \[liuppreciprocity\] [(Liu [@Liu2003 Theorem 6])]{}. For $\max\{|bu|, |bv|\}<1,$ we have the reciprocity formula $$\begin{aligned} &v\sum_{n=0}^\infty \frac{(q/bu, cduv; q)_n(bv)^n}{(cv, dv; q)_{n+1}} -u\sum_{n=0}^\infty \frac{(q/bv, cduv; q)_{n}(bu)^n}{(cu, du; q)_{n+1}}\\ &=\frac{(v-u)(q, qv/u, qu/v, bcuv, bduv, cduv; q)_\infty}{(bu, bv, cu, cv, du, dv; q)_\infty}.\end{aligned}$$ Ramanujan’s reciprocity formula is the special case $b=d=0$ of Proposition \[liuppreciprocity\]. Using a limiting case of Watson’s $q$-analog of Whipple’s theorem, Kang [@Kang] found the following equivalent form of Proposition \[liuppreciprocity\]. \[kangppreciprocity\] [(Kang [@Kang Theorem 1.2])]{}. We have $$\begin{aligned} &v\sum_{n=0}^\infty (1-q^{2n+1}v/u)\frac{(q/bu, q/cu, q/du; q)_n}{(bv, cv, dv; q)_{n+1}} q^{n(n-1)/2} (-bcduv^2)^n\\ &-u\sum_{n=0}^\infty (1-q^{2n+1}u/v)\frac{(q/bv, q/cv, q/dv; q)_n}{(bu, cu, du; q)_{n+1}} q^{n(n-1)/2} (-bcdvu^2)^n\\ &=(v-u)\frac{(q, qv/u, qu/v, bcuv, bduv, cduv; q)_\infty}{(bu, bv, cu, cv, du, dv; q)_\infty}.\end{aligned}$$ For completeness, we will reproduce Kang’s proof here. Watson’s $q$-analog of Whipple’s theorem [@Gas+Rah Eq. (2.5.1)] can be stated as follows $$\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} {_4\phi_3}\left({{q^{-m}, q/a, q/b, \alpha cd/q }\atop{\alpha c, \alpha d, q^2/{\alpha ab q^m}}}; q, q\right) \label{wweqn}$$ $$={_8\phi_7}{\left( }{{q^{-m}, q\sqrt{\alpha}, -q\sqrt{\alpha}, \alpha, q/a, q/b, q/c, q/d} \atop{\sqrt{\alpha}, -\sqrt{\alpha}, \alpha a, \alpha b, \alpha c, \alpha d, \alpha q^{m+1}}}; q, \frac{\alpha^2 abcdq^m}{q^2}{\right) }.$$ Setting $a=1$ and then letting $m\to \infty $ in the above equation and simplifying, we find for, $|\alpha b/q^2|<1$, that $$\begin{aligned} &(1-\alpha b/q) \sum_{n=0}^\infty \frac{(q/b, \alpha cd/q; q)_n}{(\alpha c, \alpha d; q)_n} (\alpha b/q)^n \\ &=\sum_{n=0}^\infty \frac{(1-\alpha q^{2n}) ( q/b, q/c, q/d; q)_n} {( \alpha b, \alpha c, \alpha d; q)_n} (-1)^n q^{n(n-1)/2} {\left( }\frac{\alpha^2 bcd}{q^2}{\right) }^n.\end{aligned}$$ Replacing $(\alpha, b, c, d)$ by $(qv/u, bu, cu, du)$ in the above equation, then dividing both sides of resulting equation by $(1-bv)(1-cv)(1-dv)$, and finally multiplying both sides by $v$, we find that the first summation in the left-hand side of the equation in Proposition \[liuppreciprocity\] equlas $$v\sum_{n=0}^\infty (1-q^{2n+1}v/u)\frac{(q/bu, q/cu, q/du; q)_n}{(bv, cv, dv; q)_{n+1}} q^{n(n-1)/2} (-bcduv^2)^n.$$ Interchanging $u$ and $v$, it is found that the second summation in the left-hand side of the equation in Proposition \[liuppreciprocity\] is equal to $$u\sum_{n=0}^\infty (1-q^{2n+1}u/v)\frac{(q/bv, q/cv, q/dv; q)_n}{(bu, cu, du; q)_{n+1}} q^{n(n-1)/2} (-bcdu^2v)^n.$$ Thus, we complete the proof of Proposition \[kangppreciprocity\]. The following general reciprocity formula was derived by Chu and Zhang from Bailey’s $_6\psi_6$ summation, which is in fact a variant form of Bailey’s $_6\psi_6$ summation. We will show that this reciprocity formula can be derived from Kang’s reciprocity formula by using Proposition \[qppmotivation\]. Thus, we give a new proof of Bailey’s $_6\psi_6$ summation. \[Chuzhangpp\] [(Chu and Zhang [@ChuZhang Theorem 5])]{}. We have $$\begin{aligned} &v\sum_{n=0}^\infty (1-q^{2n+1}v/u)\frac{(q/au, q/bu, q/cu, q/du; q)_n}{(av, bv, cv, dv; q)_{n+1}} (abcdu^2v^2/q)^n\\ &-u\sum_{n=0}^\infty (1-q^{2n+1}u/v)\frac{(q/av, q/bv, q/cv, q/dv; q)_n}{(au, bu, cu, du; q)_{n+1}} (abcdu^2v^2/q)^n\\ &=(v-u)\frac{(q, qv/u, qu/v, abuv, acuv, aduv, bcuv, bduv, cduv; q)_\infty}{(au, av, bu, bv, cu, cv, du, dv, abcdu^2v^2/q; q)_\infty}.\end{aligned}$$ For the sake of brevity, we first introduce the compact notation $A_n(a, b, u, v), B_n(u, v)$ and $f(a, b)$ as follows $$\begin{aligned} &A_n(a, b, u, v)=\frac{(auq^{-n}, buq^{-n}, avq^{n+1}, bvq^{n+1}; q)_\infty} {(abuv; q)_\infty},\\ &B_n(u, v)=v (1-q^{2n+1}v/u)\frac{(q, q/cu, q/du; q)_n}{(cv, dv; q)_{n+1}} (cdv^2/q)^n,\\ &f(a, b)=\sum_{n=0}^\infty A_n(a, b, u, v)B_n(u, v)q^{n^2+n} -\sum_{n=0}^\infty A_n(a, b, v, u)B_n(v, u) q^{n^2+n}.\end{aligned}$$ Using the ratio test, we can show that $f(a, b)$ is analytic near $(a, b)=(0, 0).$ By a direct computation, we find that $\partial_{q^{-1}, a}\{f(a, b)\}=\partial_{q^{-1}, b}\{f(a, b)\}$ equals $$\begin{aligned} &\frac{1}{(q-abuv)}\sum_{n=0}^\infty A_n(a, b, u, v)B_n(u, v)q^{n^2+n}{{\left( }u/q^{n}+vq^{n+1}-(a+b)uv{\right) }}\\ &-\frac{1}{(q-abuv)}\sum_{n=0}^\infty A_n(a, b, v, u)B_n(v, u) q^{n^2+n}{{\left( }v/q^{n}+uq^{n+1}-(a+b)uv{\right) }}.\end{aligned}$$ Thus, by (ii) in proposition \[qppmotivation\], there exists a sequence $\{\alpha_n\}$ independent of $a$ and $b$ such that $$f(a, b)=\sum_{n=0}^\infty \alpha_n g_n(a, b|q).$$ Setting $a=0$ in the above equation, and using the fact that $g_n(0, b|q)=b^n$, we find that $$\begin{aligned} \sum_{n=0}^\infty \alpha_n b^n &=\sum_{n=0}^\infty B_n(u, v) q^{n^2+n} (buq^{-n}, bvq^{n+1}; q)_\infty\\ &\quad-\sum_{n=0}^\infty B_n(v, u) q^{n^2+n} (bvq^{-n}, buq^{n+1}; q)_\infty.\end{aligned}$$ By a direct computation, we find, for any complex $z$ and any integer $n$, that $$z^n (q/z; q)_n=(-1)^n q^{n(n+1)/2} {(zq^{-n}; q)_\infty}/{(z; q)_\infty}. \label{rrameqn1}$$ It follows that $$\begin{aligned} \sum_{n=0}^\infty \alpha_n b^n &=(bu, bv; q)_\infty\sum_{n=0}^\infty B_n(u, v) (-1)^n q^{n(n+1)/2}\frac{(q/bu; q)_n (bu)^n}{(bv; q)_{n+1}}\\ &\quad-(bu, bv; q)_\infty\sum_{n=0}^\infty B_n(v, u) q^{n(n+1)/2}\frac{(q/bv; q)_n (bv)^n}{(bu; q)_{n+1}}.\end{aligned}$$ By Proposition \[kangppreciprocity\], we find that the right-hand side of the above equation is equal to $$\sum_{n=0}^\infty \alpha_n b^n =(v-u)\frac{(q, qv/u, qu/v, bcuv, bduv, cduv; q)_\infty}{(cu, cv, du, dv; q)_\infty}.$$ Using the generating function for $g_n$ in Theorem \[gefunthm\], we find that $$\sum_{n=0}^\infty (-1)^n q^{n(n-1)/2}g_n(cuv, duv|q) \frac{b^n}{(q; q)_n} =(bcuv, bduv; q)_\infty.$$ Equating the coefficients in the above two equations, we deduce that $$\alpha_n=(-1)^n(v-u) q^{n(n-1)/2}\frac{g_n(cuv, duv|q)(q, qv/u, qu/v, cduv; q)_\infty} {(q; q)_n(cu, cv, du, dv; q)_\infty}.$$ Thus we have $$\begin{aligned} f(a, b)&=(v-u)\frac{(q, qv/u, qu/v, cduv; q)_\infty}{(cu, cv, du, dv; q)_\infty}\\ &\qquad \times \sum_{n=0}^\infty (-1)^n q^{n(n-1)/2} \frac{g_n(a, b|q)g_n(cuv, duv|q)}{(q; q)_n}.\end{aligned}$$ Using the $q$-Mehler formula for $g_n$ in Theorem \[bqmehler\], we conclude that $$\sum_{n=0}^\infty (-1)^n q^{n(n-1)/2} \frac{g_n(a, b|q)g_n(cuv, duv|q)}{(q; q)_n} =\frac{(acuv, aduv, bcdu, bcdv; q)_\infty}{(abcdu^2v^2/q; q)_\infty}.$$ Combining the above two equations, we immediately find that $$f(a, b)=(v-u)\frac{(q, qv/u, qu/v, acuv, aduv, bcdu, bcdv, cduv; q )_\infty} {(cu, cv, du, dv, abcdu^2v^2/q ; q)_\infty},$$ which is the same as Proposition \[Chuzhangpp\] after applying (\[rrameqn1\]) to the left-hand side of the above equation. A general $q$-transformation formula ====================================== We proved the following general expansion formula for $q$-series [@Liu2013 Theorem 1.1]. \[newliuthm\] If $f(x)$ is an analytic function near $x=0,$ then, under suitable convergence conditions, we have $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_\infty} {(\alpha a, \alpha b; q)_\infty} f(\alpha a)\\ =\sum_{n=0}^\infty & \frac{(1-\alpha q^{2n}) (\alpha, q/a; q)_n (a/q)^n} {(1-\alpha )(q, \alpha a; q)_n} \sum_{k=0}^n \frac{(q^{-n}, \alpha q^n; q)_k q^k} {(q, \alpha b; q)_k}f(\alpha q^{k+1}).\nonumber\end{aligned}$$ The main aim of this section is using the above theorem and Proposition \[qppmotivation\] to prove the following transformation formula for terminating $q$-series. \[newliuthma\] If $m$ is an nonnegative integer and $\{A_n\}$ is a arbitrary complex sequence, then, we have $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} \sum_{n=0}^m \frac{(q^{-m}, q/a, q/b; q)_nq^n}{(q^2/{\alpha ab q^m}; q)_n}A_n\\ &=\sum_{n=0}^m \frac{(1-\alpha q^{2n})(q^{-m}, \alpha, q/a, q/b; q)_n(\alpha ab q^{m-1})^n} {(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_n} \sum_{k=0}^n (q^{-n}, \alpha q^n; q)_k q^k A_k.\end{aligned}$$ To prove Theorem \[newliuthma\], we first prove the following lemma by using Theorem \[newliuthm\]. \[liulem\] For any nonnegative integer $m$ and arbitrary complex sequence $\{A_n\}$, we have $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_\infty} {(\alpha a, \alpha b; q)_\infty} \sum_{n=0}^m A_n (q^{-m}, q/a; q)_n (\alpha a q^m)^n\\ &=\sum_{n=0}^\infty \frac{(1-\alpha q^{2n})(\alpha, q/a, q/b; q)_n (-\alpha ab/q)^n q^{n(n-1)/2}} {(1-\alpha) (q, \alpha a, \alpha b; q)_n}\\ &\quad \times \sum_{l=0}^n \frac{(q^{-m}, q^{-n}, \alpha q^n; q)_l}{(q/b; q)_l} \left(\frac{q^{m+2}}{b}\right)^l A_l. \end{aligned}$$ Suppose that $m$ is an nonnegative integer and $\{A_n\}$ is a arbitrary complex sequence. In Theorem \[newliuthm\], we can choose $f(x)$ as follows $$f(x)=\sum_{l=0}^m A_l (q^{-m}, q\alpha/x; q )_l (q^m x)^l.$$ Letting $x=\alpha a$ in the above equation, we immediately find that $$f(\alpha a)=\sum_{l=0}^m A_l (q^{-m}, q/a; q )_l (\alpha a q^m)^l. \label{teqn1}$$ It is easy to check that $(q^{-k}; q)_l=0$ for $l>k.$ Thus we at once deduce that $$\begin{aligned} f(\alpha q^{k+1})=\sum_{l=0}^m A_l (q^{-m}, q^{-k}; q)_l (\alpha q^{m+k+1})^l =\sum_{l=0}^k A_l (q^{-m}, q^{-k}; q)_l (\alpha q^{m+k+1})^l.\end{aligned}$$ It follows that $$\sum_{k=0}^n \frac{(q^{-n}, \alpha q^n; q)_k}{(q, \alpha b; q)_k} f(\alpha q^{k+1}) =\sum_{k=0}^n \frac{(q^{-n}, \alpha q^n; q)_k}{(q, \alpha b; q)_k}\sum_{l=0}^k A_l (q^{-m}, q^{-k}; q)_l (\alpha q^{m+k+1})^l.$$ Interchanging the order of the summation on the right-hand side of the above equation, we find that the right-hand side of the above equation becomes $$\sum_{l=0}^n A_l (q^{-m}; q)_l (\alpha q^{m+1})^l \sum_{k=l}^n \frac{(q^{-n}, \alpha q^n; q)_k (q^{-k}; q)_l} {(q, \alpha b; q)_k} q^{k(l+1)}.$$ Using $(q^{-k}; q)_l (q; q)_{k-l}=(-1)^l (q; q)_k q^{l(l-1)/2-kl}$, we find that the above equation becomes $$\sum_{l=0}^n A_l (q^{-m}; q)_l (-\alpha q^{m+1})^l q^{l(l-1)/2}\sum_{k=l}^n \frac{(q^{-n}, \alpha q^n; q)_k q^k } {(\alpha b; q)_k (q; q)_{k-l}}.$$ Making the variable change $k-l=j$ in the above equation, we deduce that $$\sum_{l=0}^n A_l \frac{(q^{-n}, q^{-m}; q)_l}{(\alpha b; q)_l} (-\alpha q^{m+1})^l q^{l(l+1)/2} \sum_{j=0}^{n-l} \frac{(q^{-n+l}, \alpha q^{n+l}; q)_j q^j } {(\alpha b q^l; q)_j (q; q)_{j}}. \label{teqn2}$$ Using the $q$-Chu-Vandermonde summation formula, we find that the inner summation of the above equation equals $$\frac{(bq^{-n}; q)_{n-l} (\alpha q^{n+l})^{n-l}}{(\alpha b q^l; q)_{n-l}} =(-\alpha b)^{n-l} q^{n(n-1)/2-l(l-1)/2} \frac{(q/b; q)_n (\alpha b; q)_l}{(q/b; q)_l (\alpha b; q)_n}.$$ Substituting this equation into (\[teqn2\]), we conclude that $$\begin{aligned} &\sum_{k=0}^n \frac{(q^{-n}, \alpha q^n; q)_k}{(q, \alpha b; q)_k} f(\alpha q^{k+1}) \label{teqn3}\\ &= \frac{(q/b; q)_n}{(\alpha b; q)_n} (-\alpha b; q)^n q^{n(n-1)/2} \sum_{l=0}^n \frac{(q^{-n}, q^{-m}, \alpha b q^l; q)_l}{(q/b; q)_l} \left(\frac{q^{m+2}}{b}\right)^l A_l.\nonumber \end{aligned}$$ Substituting the above equation and equation (\[teqn1\]) into Theorem \[newliuthm\], we complete the proof of Lemma \[liulem\]. Now we begin to prove Theorem \[newliuthma\] by using the above lemma and Proposition \[qppmotivation\]. Setting $b=q^{m+1}$ in Lemma \[liulem\] and then replacing $A_n$ by $A_n (q/b; q)_n$, we immediately deduce that $$\begin{aligned} &\frac{(\alpha q; q)_m}{(\alpha a; q)_m} \sum_{n=0}^{m} A_n (q^{-m}, q/a, q/b; q)_n (\alpha a q^m)^n \label{teqn4}\\ =\sum_{n=0}^m & \frac{(1-\alpha q^{2n}) (q^{-m}, \alpha, q/a; q)_n (-\alpha aq^{m})^n q^{n(n-1)/2}} {(1-\alpha)(q, \alpha a, \alpha q^{m+1}; q)_n}B_n,\nonumber\end{aligned}$$ where $B_n$ is defined as $$B_n=\sum_{k=0}^n (q^{-n}, \alpha q^n; q)_k q^k A_k.$$ For simplicity, we temporarily introduce $f_{m,n}(a, b), f(a, b)$ and $g(a, b)$ as follows $$\begin{aligned} &f_{m, n} (a, b)=\frac{(aq^{-n}, bq^{-n}, \alpha aq^m, \alpha b q^m; q)_\infty} {(\alpha ab q^{m-n-1}; q)_\infty},\\ &f(a, b)=\sum_{n=0}^m \frac{(1-\alpha q^{2n}) (q^{-m}, \alpha; q)_n \alpha^n q^{n^2+mn}} {(1-\alpha) (q, \alpha q^{m+1}; q)_n}f_{n, n} (a, b)B_n,\\ &g(a, b)=(\alpha q; q)_m \sum_{n=0}^{m} (q^{-m}; q)_n (-\alpha q^m)^n q^{n(n+1)/2}f_{m, n}(a, b)A_n.\end{aligned}$$ It is easily seen that $$\begin{aligned} f(a, 0)&=(a, \alpha a; q)_\infty \sum_{n=0}^m \frac{(1-\alpha q^{2n}) (q^{-m}, \alpha, q/a; q)_n (-\alpha aq^{m})^n q^{n(n-1)/2}} {(1-\alpha)(q, \alpha a, \alpha q^{m+1}; q)_n}B_n,\\ g(a, 0)&=(a, \alpha a; q)_\infty \frac{(\alpha q; q)_m}{(\alpha a; q)_m} \sum_{n=0}^{m} A_n (q^{-m}, q/a; q)_n (\alpha a q^m)^n.\end{aligned}$$ Combining the above two equations and (\[teqn4\]), we find the identity, $f(a, 0)=g(a, 0).$ By a direct computation, we find that $$\partial_{q^{-1}, a} \{f_{m, n}(a, b)\}=\partial_{q^{-1}, b} \{f_{m, n}(a, b)\}=\frac{(1+\alpha q^{m+n}-\alpha(a+b)q^{m-1})f_{m, n} (a, b)}{q^{n+1}(1-\alpha ab q^{m-n-2})}.$$ It follows that $$\partial_{q^{-1}, a} \{f(a, b)\}=\partial_{q^{-1}, b} \{f(a, b)\}, ~\text{and}~ \partial_{q^{-1}, a} \{g(a, b)\}=\partial_{q^{-1}, b} \{g(a, b)\}.$$ Thus, by (ii) in Proposition \[qppmotivation\], there exist two sequences $\{\beta_l\}$ and $\{\gamma_l\}$ independent of $a$ and $b$ such that $$\begin{aligned} f(a, b)=\sum_{l=0}^\infty \beta_l h_l(a, b|q),\quad g(a, b)=\sum_{l=0}^\infty \gamma_l h_l(a, b|q).\end{aligned}$$ Setting $b=0$ in the above equations and using $f(a, 0)=g(a, 0),$ we find that $\beta_l=\gamma_l$ for any nonnegative integer $l$. It follows that $f(a, b)=g(a, b),$ which gives $$\begin{aligned} &\frac{(\alpha q; q)_m}{(\alpha a, \alpha b; q)_m}\sum_{n=0}^m (q^{-m}, q/a, q/b; q)_n q^{-n(n+1)/2}(-\alpha ab q^m)^n (\alpha ab/q; q)_{m-n}A_n\\ &=\sum_{n=0}^m \frac{(1-\alpha)(q^{-m}, \alpha, q/a, q/b; q)_n(\alpha ab q^{m-1})^n}{(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_n}B_n.\end{aligned}$$ From the definition of the $q$-shifted factorial and by a direct computation, we find that $$(\alpha ab/q; q)_{m-n}= \frac{(-1)^nq^{n(n+1)/2}(\alpha ab/q; q)_m}{(q^2/\alpha ab q^m; q)_n (\alpha ab q^{m-1})^n}.$$ Combining the above two equations, we finish the proof of Theorem \[newliuthma\]. An extension of Watson’s $q$-analog of Whipple’s theorem ========================================================= Taking $A_k=(\beta, \gamma; q)_kz^k/(q, c, d, h; q)_k$ in Theorem \[newliuthma\], we immediately obtain the following theorem. \[newliuthmb\] For any nonnegative integer $m$, we have the $q$-formula $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} {_5\phi_4}\left({{q^{-m}, q/a, q/b, \beta, \gamma }\atop{q^2/{\alpha ab q^m, c, d, h}}}; q, qz\right)\\ =&\sum_{n=0}^m \frac{(1-\alpha q^{2n})(q^{-m}, \alpha, q/a, q/b; q)_n(\alpha ab q^{m-1})^n} {(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_n} {_4\phi_3}{\left( }{{q^{-n}, \alpha q^n, \beta, \gamma}\atop{c, d, h}}; q, qz{\right) }.\end{aligned}$$ If we take $h=\gamma=0$ and $z=1$ in the above equation, we immediately deduce the following proposition. \[Exwwpp\] For any nonnegative integer $m$, then, we have $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} {_4\phi_3}\left({{q^{-m}, q/a, q/b, \beta}\atop{q^2/{\alpha ab q^m, c, d}}}; q, q\right)\\ =&\sum_{n=0}^m \frac{(1-\alpha q^{2n})(q^{-m}, \alpha, q/a, q/b; q)_n(\alpha ab q^{m-1})^n} {(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_n} {_3\phi_2}{\left( }{{q^{-n}, \alpha q^n, \beta}\atop{c, d}}; q, q{\right) }.\end{aligned}$$ This proposition contains Watson’s $q$-analog of Whipple’s theorem as a special case. Thus we may regard it an extension of Watson’s $q$-analog of Whipple’s theorem. \[wwpp\] [(Watson’s $q$-analog of Whipple’s theorem)]{}. $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} {_4\phi_3}\left({{q^{-m}, q/a, q/b, \alpha cd/q }\atop{\alpha c, \alpha d, q^2/{\alpha ab q^m}}}; q, q\right)\\ &={_8\phi_7}{\left( }{{q^{-m}, q\sqrt{\alpha}, -q\sqrt{\alpha}, \alpha, q/a, q/b, q/c, q/d} \atop{\sqrt{\alpha}, -\sqrt{\alpha}, \alpha a, \alpha b, \alpha c, \alpha d, \alpha q^{m+1}}}; q, \frac{\alpha^2 abcdq^m}{q^2}{\right) }.\end{aligned}$$ If $(c, d, \beta)$ is replaced by $(\alpha c, \alpha d, \alpha cd/q)$, we find that the left-hand side of the equation in Proposition \[Exwwpp\] becomes the left-hand side of the equation in Proposition \[wwpp\], and the right-hand side becomes $$\sum_{n=0}^m \frac{(1-\alpha q^{2n})(q^{-m}, \alpha, q/a, q/b; q)_n(\alpha ab q^{m-1})^n} {(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_n} {_3\phi_2}{\left( }{{q^{-n}, \alpha q^n, \alpha cd/q}\atop{\alpha c, \alpha d}}; q, q{\right) }.$$ Using the $q$-Pfaff-Saalschütz formula (see, for example [@Gas+Rah p. 13, Eq. (1.7.2)]), we find that $${_3 \phi_2} \left({{q^{-n}, \alpha q^n, \alpha cd/q}\atop{\alpha c, \alpha d}}; q, q\right)=\frac{(q/c, q/d; q)_n}{(\alpha c, \alpha d; q)_n}\left(\frac{\alpha cd}{q}\right)^n.$$ Combining the above two equations we arrive at the right-hand side of the equation in Proposition \[wwpp\]. Thus we complete the proof of the proposition. Letting $m\to \infty$ in Theorem \[newliuthma\], we immediately obtain the following theorem. Many important applications of this theorem to mock-theta function identities have been discussed in the paper [@Liumocktheta]. \[newliuthmc\] For $|\alpha abz/q|<1$, we have the $q$-transformation formula $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_\infty} {(\alpha a, \alpha b; q)_\infty} {_4\phi_3} \left({{q/a, q/b, \beta, \gamma} \atop { c, d, h}} ; q, \frac{\alpha ab z}{q} \right) \\ &=\sum_{n=0}^\infty \frac{(1-\alpha q^{2n}) (\alpha, q/a, q/b; q)_n (-\alpha ab/q)^n q^{n(n-1)/2}} {(1-\alpha)(q, \alpha a, \alpha b; q)_n} {_4\phi_3} \left({{q^{-n}, \alpha q^n, \beta, \gamma} \atop {c, d, h}} ; q, qz \right).\end{aligned}$$ Some $q$-series transformation formulas ======================================= In this section we will use Theorem \[newliuthmb\] to derive some $q$-transformation formula. \[liutfthma\] If $m$ is an nonnegative integer, then, we have $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_m}{(\alpha^2 a, \alpha^2 b; q^2)_m} {_5\phi_4}\left({{q^{-2m}, q^2/a, q^2/b, \lambda, q\lambda }\atop{\alpha, q\alpha, q^2\lambda^2, q^4/{\alpha^2 ab q^{2m}}}}; q^2, q^2 \right)\\ =&\sum_{n=0}^m \frac{(1-\alpha^2 q^{4n})(q^{-2m}, \alpha^2, q^2/a, q^2/b; q^2)_n (-q, \alpha /\lambda; q)_n (\alpha^2 \lambda ab q^{2m-2})^n} {(1-\alpha^2)(q^2, \alpha^2 q^{2m+2}, \alpha^2 a, \alpha^2 b; q^2)_n (\alpha, -q\lambda; q)_n}.\end{aligned}$$ We first replace $q$ by $q^2$ and $\alpha$ by $\alpha^2$ in Theorem \[newliuthmb\] and then set $(\beta, \gamma, c, d, h, z) =(\lambda, q\lambda, \alpha, q\alpha, \lambda^2 q^2, 1)$ in the resulting equation, we deduce that $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_m}{(\alpha^2 a, \alpha^2 b; q^2)_m} {_5\phi_4}\left({{q^{-2m}, q^2/a, q^2/b, \lambda, q\lambda }\atop{\alpha, q\alpha, q^2\lambda^2, q^4/{\alpha^2 ab q^{2m}}}}; q^2, q^2 \right)\\ =&\sum_{n=0}^m \frac{(1-\alpha^2 q^{4n})(q^{-2m}, \alpha^2, q^2/a, q^2/b; q^2)_n(\alpha^2 ab q^{2m-2})^n} {(1-\alpha^2)(q^2, \alpha^2 q^{2m+2}, \alpha^2 a, \alpha^2 b; q^2)_n}\\ &\qquad\times{_4\phi_3}{\left( }{{q^{-2n}, \alpha^2 q^{2n}, \lambda, q\lambda}\atop{\alpha, q\alpha, q^2\lambda^2}}; q^2, q^2 {\right) }.\end{aligned}$$ Verma and Jain [@VermaJain Eq. (5.3)] (see also [@Gas+Rah p. 110, Ex. (3.34)]) proved that $${_4\phi_3} \left({{q^{-2n}, \alpha^2 q^{2n}, \lambda, q\lambda} \atop { \alpha, q \alpha, q^2\lambda^2}} ; q^2, q^2 \right) =\frac{\lambda^n (-q, \alpha /\lambda; q)_n}{(\alpha, -q\lambda; q)_n}. \label{rogers:eqn1}$$ Combining the above two equations, we complete the proof of Theorem \[liutfppa\]. Letting $m\to \infty$ in Theorem \[liutfthma\], we obtain the following proposition. \[liutfppa\] $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_\infty}{(\alpha^2 a, \alpha^2 b; q^2)_\infty} {_4\phi_3}\left({{q^2/a, q^2/b, \lambda, q\lambda }\atop{\alpha, q\alpha, q^2\lambda^2 }}; q^2, \frac{\alpha^2 ab}{q^2} \right)\\ =&\sum_{n=0}^\infty \frac{(1-\alpha^2 q^{4n})( \alpha^2, q^2/a, q^2/b; q^2)_n (-q, \alpha /\lambda; q)_n (-\alpha^2 \lambda ab)^n q^{n^2-3n}} {(1-\alpha^2)(q^2, \alpha^2 a, \alpha^2 b; q^2)_n (\alpha, -q\lambda; q)_n}.\end{aligned}$$ Let $(a, b, \alpha, \gamma)=(0, 0, - q, 0)$ in Proposition \[liutfppa\], we have the evaluation $$\begin{aligned} \sum_{n=0}^\infty \frac{q^{2n^2+2n}}{(-q; q)_{2n}(q^2; q^2)_n} &=\frac{1}{(q^2; q^2)_\infty} \sum_{n=-\infty}^\infty (-1)^n q^{\frac{7n^2+3n}{2}} \label{rrideqn1}\\ =\frac{(q^2, q^5, q^7; q^7)_\infty}{(q^2; q^2)_\infty}.\nonumber\end{aligned}$$ \[liutfthmb\] For any nonnegative integer $m,$ we have $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_m}{(\alpha^2 a, \alpha^2 b; q^2)_m} {_5\phi_4}\left({{q^{-2m}, q^2/a, q^2/b, \lambda, q\lambda }\atop{q\alpha, q^2\alpha, \lambda^2, q^4/{\alpha^2 ab q^{2m}}}}; q^2, q^2 \right)\\ =&\sum_{n=0}^m \frac{(1+\alpha q^{2n})(q^{-2m}, \alpha^2, q^2/a, q^2/b; q^2)_n (-q, q\alpha /\lambda; q)_n (\alpha^2 \lambda ab q^{2m-2})^n} {(1+\alpha)(q^2, \alpha^2 q^{2m+2}, \alpha^2 a, \alpha^2 b; q^2)_n (\alpha, -\lambda; q)_n}.\end{aligned}$$ Replacing $q$ by $q^2$ and $\alpha$ by $\alpha^2$ in Theorem \[newliuthmb\] and then set $(\beta, \gamma, c, d, h, z) =(\lambda, q\lambda, q\alpha, q^2\alpha, \lambda^2 , 1)$ in the resulting equation, we deduce that $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_m}{(\alpha^2 a, \alpha^2 b; q^2)_m} {_5\phi_4}\left({{q^{-2m}, q^2/a, q^2/b, \lambda, q\lambda }\atop{q\alpha, q^2\alpha, \lambda^2, q^4/{\alpha^2 ab q^{2m}}}}; q^2, q^2 \right)\\ =&\sum_{n=0}^m \frac{(1-\alpha^2 q^{4n})(q^{-2m}, \alpha^2, q^2/a, q^2/b; q^2)_n(\alpha^2 ab q^{2m-2})^n} {(1-\alpha^2)(q^2, \alpha^2 q^{2m+2}, \alpha^2 a, \alpha^2 b; q^2)_n}\\ &\qquad\times{_4\phi_3}{\left( }{{q^{-2n}, \alpha^2 q^{2n}, \lambda, q\lambda}\atop{q\alpha, q^2\alpha, \lambda^2}}; q^2, q^2 {\right) }.\end{aligned}$$ Verma and Jain [@VermaJain Eq. (5.4)] derived the following series summation formula: $${_4\phi_3} \left({{q^{-2n}, \alpha^2 q^{2n}, \lambda, q\lambda} \atop { q\alpha, q^2 \alpha, \lambda^2}} ; q^2, q^2 \right) =\frac{\lambda^n (-q, q\alpha /\lambda; q)_n (1-\alpha)}{(\alpha, -\lambda; q)_n (1-\alpha q^{2n})}. \label{rogers:eqn2}$$ Combining the above two equations, we finish the proof of Theorem \[liutfthmb\]. Letting $m\to \infty$ in Theorem \[liutfthmb\], we obtain the following proposition. \[liutfppb\] $$\begin{aligned} &\frac{(\alpha^2 q^2, \alpha^2 ab/q^2; q^2)_\infty}{(\alpha^2 a, \alpha^2 b; q^2)_\infty} {_4\phi_3}\left({{q^2/a, q^2/b, \lambda, q\lambda }\atop{q\alpha, q^2\alpha, \lambda^2}}; q^2, \frac{\alpha^2 ab}{q^2} \right)\\ =&\sum_{n=0}^\infty \frac{(1+\alpha q^{2n})(\alpha^2, q^2/a, q^2/b; q^2)_n (-q, q\alpha /\lambda; q)_n (-\alpha^2 \lambda ab )^nq^{n^2-3n}} {(1+\alpha)(q^2, \alpha^2 a, \alpha^2 b; q^2)_n (\alpha, -\lambda; q)_n}.\end{aligned}$$ Setting $(a, b, \alpha, \gamma)=(0, 0, -q, 0)$ and $(0, 0, -1, 0)$ in Proposition \[liutfppb\], respectively, we find that $$\begin{aligned} \sum_{n=0}^\infty \frac{q^{2n^2+2n}}{(-q; q)_{2n+1}(q^2; q^2)_n} &=\frac{(q, q^6, q^7; q^7)_\infty}{(q^2; q^2)_\infty},\label{rrideqn2}\\ \sum_{n=0}^\infty \frac{q^{2n^2}}{(-q; q)_{2n}(q^2; q^2)_n} &=\frac{(q^3, q^4, q^7; q^7)_\infty}{(q^2; q^2)_\infty}. \label{rrideqn3}\end{aligned}$$ Identities (\[rrideqn1\]), (\[rrideqn2\]) and (\[rrideqn3\]) are called the Rogers-Selberg identities (see, for example, [@McLaughlinS Eqs. (2.7.1), (2.7.2), (2.7. 3)]). It is easily seen that the identity of Verma and Jain [@VermaJain Eq. (2.28)] is equivalent to the summation $${_3\phi_2} \left({{q^{-n}, \alpha q^n, 0} \atop {\sqrt{q\alpha}, -\sqrt{q\alpha}}} ; q, q \right) =\begin{cases} 0 &\text {if $n$ is odd}\\ (-1)^l q^{l^2} \frac{(q; q^2)_l \alpha^l }{(q\alpha; q^2)_l}, & \text{if $n=2l$} \end{cases} \label{rogers:eqn3}$$ Setting $(\beta, \gamma, h, c, d, z)=(0, 0, 0, \sqrt{q\alpha}, -\sqrt{q\alpha}, 1)$ in Theorem \[newliuthmb\] and then using (\[rogers:eqn3\]), we obtain the following theorem. \[liutfthmc\] For any nonnegative integer $m,$ we have $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_m}{(\alpha a, \alpha b; q)_m} {_4\phi_3}\left({{q^{-m}, q/a, q/b, 0 }\atop{\sqrt{q\alpha}, -\sqrt{q\alpha}, q^2/{\alpha ab q^m}}}; q, q\right)\\ &=\sum_{n=0}^{[m/2]} \frac{(1-\alpha q^{4n})(q^{-m}, \alpha, q/a, q/b; q)_{2n} (q; q^2)_{n}(-\alpha^3 a^2b^2 q^{2m-2})^n q^{n^2}} {(1-\alpha)(q, \alpha q^{m+1}, \alpha a, \alpha b; q)_{2n} (q\alpha; q^2)_n}.\end{aligned}$$ Letting $m\to \infty$ in Theorem \[liutfthmc\], we obtain the following proposition. \[liutfppc\] $$\begin{aligned} &\frac{(\alpha q, \alpha ab/q; q)_\infty}{(\alpha a, \alpha b; q)_\infty} {_3\phi_2}\left({{q/a, q/b, 0 }\atop{\sqrt{q\alpha}, -\sqrt{q\alpha}}}; q, \frac{\alpha ab}{q}\right)\\ &=\sum_{n=0}^{\infty} \frac{(1-\alpha q^{4n})(\alpha, q/a, q/b; q)_{2n} (q; q^2)_{n}(-\alpha^3 a^2b^2 )^n q^{3n^2-3n}} {(1-\alpha)(q, \alpha a, \alpha b; q)_{2n} (q\alpha; q^2)_n}.\end{aligned}$$ Putting $(a, b, \alpha)=(0, 0, 1)$ and $(0, 0, q^2)$ in Proposition \[liutfppc\], respectively, we obtain the Rogers identities (see, for example, [@McLaughlinS Eqs. (2.14.1), (2.14.3)]) $$\begin{aligned} \sum_{n=0}^\infty \frac{q^{n^2}} {(q; q)_n (q; q^2)_n}&=\frac{(q^6, q^8, q^{14}; q^{14})_\infty} {(q; q)_\infty},\label{rrideqn4}\\ \sum_{n=0}^\infty \frac{q^{n^2+2n}}{(q; q)_n (q; q^2)_{n+1}}&=\frac{(q^2, q^{12}, q^{14}; q^{14})_\infty}{(q; q)_\infty}. \label{rrideqn5}\end{aligned}$$ A multilinear generating function for the Rogers-Szegő polynomials =================================================================== \[mgrogers\] If $\max\{ |a|, |c|, |x_1|, |y_1|, \ldots, |x_k|, |y_k|\}<1,$ then, we have the following multilinear generating function for the Rogers-Szegő polynomials: $$\begin{aligned} \sum_{n_1, n_2, \ldots, n_k=0}^\infty \frac{(a; q)_{n_1+n_2+\cdots+n_k}h_{n_1}(x_1, y_1|q) h_{n_2}(x_2, y_2|q)\cdots h_{n_k}(x_k, y_k|q)} {(c; q)_{n_1+n_2+\cdots+n_k}(q; q)_{n_1}(q; q)_{n_2}\cdots (q; q)_{n_k}}\\ =\frac{(a; q)_\infty}{(c, x_1, y_1, x_2, y_2, \ldots, x_k, y_k)_\infty} {_{2k+1}\phi_{2k}}\left({{c/a, x_1, y_1, x_2, y_2, \ldots, x_k, y_k} \atop{0, 0, \ldots, 0}}; q, a\right).\end{aligned}$$ If we use $f(x_1, y_1, \ldots, x_k, y_k)$ to denote the right-hand side of the above equation, then, using the ratio test, we find that $f$ is an analytic function of $x_1, y_1, x_2, y_2, \ldots, x_k, y_k$ for $\max\{ |a|,|c|, |x_1|, |y_1|, \ldots, |x_k|, |y_k|\}<1.$ By a direct computation, we deduce that for $j=1, 2\ldots, k$, $\partial_{q, x_j}\{f\}=\partial_{q, y_j}\{f\}$ equals $$\frac{(a; q)_\infty}{(c, x_1, y_1, x_2, y_2, \ldots, x_k, y_k)_\infty} {_{2k+1}\phi_{2k}}\left({{c/a, x_1, y_1, x_2, y_2, \ldots, x_k, y_k} \atop{0, 0, \ldots, 0}}; q, qa\right).$$ Thus, by (i) in Theorem \[mainthmliu\], there exists a sequence $\{\alpha_{n_1, n_2, \ldots, n_k}\}$ independent of $x_1, y_1, \ldots, x_k, y_k$ such that $$\begin{aligned} &f(x_1, y_1, x_2, y_2, \ldots, x_k, y_k)\label{mgeneqn}\\ &=\sum_{n_1, n_2, \ldots, n_k=0}^\infty \alpha_{n_1, n_2, \ldots, n_k} h_{n_1}(x_1, y_1|q) h_{n_2}(x_2, y_2|q)\cdots h_{n_k}(x_k, y_k|q).\nonumber\end{aligned}$$ Setting $y_1=y_2\cdots=y_k=0$ in the above equation, we immediately obtain $$\begin{aligned} &\sum_{n_1, n_2, \ldots, n_k=0}^\infty \alpha_{n_1, n_2, \ldots, n_k} x_{1}^{n_1}x_2^{n_2}\cdots x_{k}^{n_k}\label{mgeneqn1}\\ &=\frac{(a; q)_\infty}{(c, x_1, x_2, \ldots, x_k)_\infty} {_{k+1}\phi_{k}}\left({{c/a, x_1, x_2, \ldots, x_k} \atop{0, 0, \ldots, 0}}; q, a\right).\nonumber\end{aligned}$$ Andrews [@Andrews1972] has proved the following formula for the $q$-Lauricella function: $$\begin{aligned} \sum_{n_1, n_2, \ldots, n_k=0}^\infty \frac{(a; q)_{n_1+n_2+\cdots+n_k}(b_1; q)_{n_1}(b_2; q)_{n_2}\cdots (b_k; q)_{n_k} x_1^{n_1} x_2^{n_2}\cdots x_k^{n_k}} {(c; q)_{n_1+n_2+\cdots+n_k}(q; q)_{n_1}(q; q)_{n_2}\cdots (q; q)_{n_k}}\\ =\frac{(a, b_1x_1, b_2x_2, \ldots, b_kx_k; q)_\infty}{(c, x_1, x_2, \ldots, x_k)_\infty} {_{k+1}\phi_{k}}\left({{c/a, x_1, x_2, \ldots, x_k} \atop{bx_1, bx_2, \ldots, bx_k}}; q, a\right).\end{aligned}$$ Taking $b_1=b_2=\cdots=b_k=0$ in the above equation, we conclude that $$\begin{aligned} \sum_{n_1, n_2, \ldots, n_k=0}^\infty \frac{(a; q)_{n_1+n_2+\cdots+n_k} x_1^{n_1} x_2^{n_2}\cdots x_k^{n_k}} {(c; q)_{n_1+n_2+\cdots+n_k}(q; q)_{n_1}(q; q)_{n_2}\cdots (q; q)_{n_k}}\\ =\frac{(a; q)_\infty}{(c, x_1, x_2, \ldots, x_k)_\infty} {_{k+1}\phi_{k}}\left({{c/a, x_1, x_2, \ldots, x_k} \atop{0, 0, \ldots, 0}}; q, a\right).\end{aligned}$$ Equating the above equation and (\[mgeneqn1\]), we find that $$\alpha_{n_1, n_2, \ldots, n_k}=\frac{(a; q)_{n_1+n_2+\cdots+n_k}} {(c; q)_{n_1+n_2+\cdots+n_k}(q; q)_{n_1}(q; q)_{n_2}\cdots (q; q)_{n_k}}.$$ Substituting this into (\[mgeneqn\]), we complete the proof of Theorem \[mgrogers\]. Some notes on the $q$-exponential operator identities ====================================================== In this section, we will revisit the theory of $q$-exponential operator using Proposition \[qppmotivation\]. Using the $q$-derivative operator, we can construct the $q$-exponential operator $T(y \mathcal{D}_{q, x})$ by $$T(y \mathcal{D}_{q, x})=\sum_{n=0}^\infty \frac{(y \mathcal{D}_{q, x})^n }{(q; q)_n}. \label{liueqn1}$$ If we replace $q$ by $q^{-1}$ in the above equation, we obtain the dual operator of $T(y \mathcal{D}_{q, x})$ as follows $$T(y \mathcal{D}_{q^{-1}, x})=\sum_{n=0}^\infty \frac{(-y \mathcal{D}_{q^{-1}, x})^n q^{n(n+1)/2}}{(q; q)_n}. \label{liueqn2}$$ \[opedefn\] If $f(x)$ is analytic near $x=0,$ then, we define $$\begin{aligned} T(y \mathcal{D}_{q, x})\{f(x)\}&=\sum_{n=0}^\infty \frac{y^n }{(q; q)_n}\mathcal{D}_{q, x}^n \{f(x)\},\\ T(y \mathcal{D}_{q^{-1}, x})\{f(x)\}&=\sum_{n=0}^\infty \frac{(-y)^n q^{n(n+1)/2} }{(q; q)_n}\mathcal{D}_{q^{-1}, x}^n \{f(x)\}.\end{aligned}$$ The following lemma is a modified version of [@Liu2010 Theorems 1 and 2], which is the $q$-exponential operator form of Proposition \[qppmotivation\]. This lemma indicates that the $q$-exponential operator method has a solid mathematical foundation. \[liulemope\] Suppose that $f(x, y)$ is a two-variable analytic function near $(x, y)=(0, 0).$ Then - $f(x, y)=T(y \mathcal{D}_{q, x})\{f(x, 0)\},$ if and only if $\partial_{q, x}\{f\}=\partial_{q, y}\{f\}$. - $f(x, y)=T(y \mathcal{D}_{q^{-1}, x})\{f(x, 0)\},$ if and only if $\partial_{q^{-1}, x}\{f\}=\partial_{q^{-1}, y}\{f\}$. We only prove (i). The proof of (ii) is similar to that of (i), so is omitted. If $f(x, y)=T(y \mathcal{D}_{q, x})\{f(x, 0)\},$ then, by a direct computation, we find that $\partial_{q, x}\{f\}=\partial_{q, y}\{f\}=T(y \mathcal{D}_{q, x})\mathcal{D}_{q, x}\{f(x, 0)\}.$ Conversely, if $f$ is a two-variable analytic function near $(x, y)=(0, 0),$ which satisfying $\partial_{q, x}\{f\}=\partial_{q, y}\{f\},$ then, we may assume that $$f(x, y)=\sum_{n=0}^\infty A_n(x) y^n.$$ Substituting the equation into the $q$-partial differential equation, $\partial_{q, x}\{f\}=\partial_{q, y}\{f\}$, we easily find that $$A_n(x)=\frac{\mathcal{D}_{q, x}\{A_n(x)\}}{1-q^n}=\cdots=\frac{\mathcal{D}^n_{q, x}\{A_0(x)\}}{(q; q)_n}.$$ It is easily seen that $A_0(x)=f(x, 0).$ Thus, we find that $$f(x, y)=\sum_{n=0}^\infty \frac{\mathcal{D}^n_{q, x}\{f(x, 0)\}}{(q; q)_n}y^n =T(y \mathcal{D}_{q, x})\{f(x, 0)\}.$$ \[cliupp\]Suppose that $T(y \mathcal{D}_{q, x})$ and $T(y \mathcal{D}_{q^{-1}, x})$ are defined by $(\ref{liueqn1})$. Then - For $\max\{|xs|, |xt|, |ys|, |yt|\}<1$, we have the operator identity $$T(y \mathcal{D}_{q, x})\left\{\frac{1}{(xs, xt; q)_\infty}\right\} =\frac{(xyst; q)_\infty}{(xs, xt, ys, yt; q)_\infty}.$$ - If $|xyst/q|<1,$ then, we have the operator identity $$T(y \mathcal{D}_{q^{-1}, x})\left\{(xs, xt; q)_\infty\right\} =\frac{(xs, xt, ys, yt; q)_\infty}{(xyst/q; q)_\infty}.$$ Many applications of these two $q$-operator identities and their extensions have appeared in the literature, see, for example, [@Cao2009; @Cao2010; @CL1; @CL2; @Liu2003; @Zhang; @ZhangWang]. The original proofs of these two operator identities in [@CL1; @CL2] depend on the method of formal computation. Next we use Lemma \[liulemope\] to give a very simple proof of Proposition \[cliupp\]. The proof of (ii) is similar to that of (ii), so we only prove (i). If we use $f(x, y)$ to denote the right-hand side of the first equation in Proposition \[cliupp\]. Then it is easily seen that $f(x, y)$ is analytic near $(0, 0)$ and satisfies $$\partial_{q, x}\{f\}=\partial_{q, y}\{f\} =\frac{s+t-(x+y)st}{1-xyst}f(x, y).$$ Thus, by (i) in Proposition \[liulemope\], we have $f(x, y)=T(y \mathcal{D}_{q, x})\{f(x, 0)\}.$ This completes the proof of Proposition \[cliupp\]. From Definition \[opedefn\], by a direct computation, we find the $q$-exponential operator representations of $h_n$ and $g_n$ as follows $$T(y \mathcal{D}_{q, x})\{x^n\}=h_n(x, y|q), ~ T(y \mathcal{D}_{q^{-1}, x})\{x^n\}=g_n(x, y|q). \label{liueqn3}$$ \[liuopeppa\] If $f(x)$ is analytic near $x=0,$ then $T(y \mathcal{D}_{q, x})\{f(x)\}$ is analytic near $(x, y)=(0, 0).$ Since $f(x)$ is analytic near $x=0,$ there exists a positive number $r<1$ (without loss of generality), such that $$f(x)=\sum_{k=0}^\infty a_k x^k,\quad |x|<r<1. \label{liueqn4}$$ From Definition \[opedefn\] and by a direct computation, we conclude that $$\begin{aligned} T(y \mathcal{D}_{q, x})\{f(x)\}&=\sum_{n=0}^\infty \frac{y^n}{(q; q)_n} \sum_{k=n}^\infty \frac{(q; q)_k}{(q; q)_{k-n}}a_k x^{k-n}\\ &=\sum_{n=0}^\infty \frac{y^n}{(q; q)_n} \sum_{m=0}^\infty \frac{(q; q)_{n+m}a_{n+m}}{(q; q)_{m}} x^{m}.\end{aligned}$$ Since $f(x)$ is analytic in $|x|<r,$ we have, by the root test, that $ \lim_{n\to \infty} \text{sup}|a_n|^{1/n}\le 1. $ It is obvious that $\lim_{n\to \infty}|(q; q)_{n+m}|=|(q; q)_\infty|<\infty. $ It follows that $$\lim_{n\to \infty} \text{sup}|a_{n+m}(q; q)_{n+m}|^{1/(n+m)}\le 1.$$ Thus, for $r$ in (\[liueqn4\]), there exists a large integer $N$ such that $|a_{n+m}(q; q)_{n+m}|\le (1+r)^{m+n}$ for all $n\ge N$. It follows that $$\begin{aligned} \Big|\sum_{n=N}^\infty \frac{y^n}{(q; q)_n} \sum_{m=0}^\infty \frac{(q; q)_{n+m}a_{n+m}}{(q; q)_{m}} x^{m}\Big| \le \sum_{n=N}^\infty \frac{|(1+r)y|^n}{(q; q)_n}\sum_{m=0}^\infty \frac{(1+r)^{m}|x|^m}{(q; q)_{m}}\end{aligned}$$ We further assume that $|x|\le r/2$, then from the above equation, we deduce that $$\begin{aligned} \Big|\sum_{n=N}^\infty \frac{y^n}{(q; q)_n} \sum_{m=0}^\infty \frac{(q; q)_{n+m}a_{n+m}}{(q; q)_{m}} x^{m}\Big| \le \sum_{n=N}^\infty \frac{|(1+r)y|^n}{(q; q)_n}\sum_{m=0}^\infty \frac{(r(r+1)/2)^{m}}{(q; q)_{m}},\end{aligned}$$ which is a uniformly and absolutely convergent series for $|y|<1/(1+r).$ Thus $T(y \mathcal{D}_{q, x})\{f(x)\}$ is analytic near $y=0.$ Thus, by Hartog’s theorem, we know that $T(y \mathcal{D}_{q, x})\{f(x)\}$ is analytic near $(x, y)=(0, 0).$ A deep result of the $q$-exponential operator is the following operator identity [@Liu2010 Eq.(3.1)]. \[liuopeppb\]For $\max\{|as|, |at|, |au|, |bs|, |bt|, |bu|, |abstu/v|\}<1, $ we have $$\begin{aligned} T(b \mathcal{D}_{q, a})\left\{\frac{(av; q)_\infty}{(as, at, au; q)_\infty}\right\} &=\frac{(av, bv, abstu/v; q)_\infty}{(as, at, au, bs, bt, bu; q)_\infty}\\ &\quad \times{_3\phi_2}\left({{v/s, v/t, v/u}\atop{av, bv}}; q, \frac{abstu}{v}\right).\end{aligned}$$ Now we will give a new proof of this identity. If we use $f(a, b)$ to denote the left-hand side of the equation in Proposition \[liuopeppb\], then from Proposition \[liuopeppa\], we know that $f(a, b)$ is analytic near $(a, b)=(0, 0).$ Using Lemma \[liulemope\], we find that $\partial_{q, a}\{f\}=\partial_{q, b}\{f\}.$ Thus, there exists a sequence $\{\alpha_n\}$ independent of $a$ and $b$ such that $$f(a, b)=\sum_{n=0}^\infty \alpha_n h_n(a, b|q). \label{liueqn5}$$ Taking $b=0$ in the above equation and using $h_n(a, 0|q)=a^n$, we immediately deduce that $$\sum_{n=0}^\infty \alpha_n a^n=\frac{(av; q)_\infty}{(as, at, au; q)_\infty}. \label{liueqn6}$$ Using the $q$-binomial theorem and the generating function for $h_n,$ we obtain $$\frac{(av; q)_\infty}{(as; q)_\infty}=\sum_{n=0}^\infty \frac{(v/s; q)_n (as)^n}{(q; q)_n}, \quad \frac{1}{(at, au; q)_\infty}=\sum_{n=0}^\infty h_n(t, u|q) \frac{a^n}{(q; q)_n}.$$ Substituting the above two equations into (\[liueqn6\]) and equating the coefficients of $a^n,$ we deduce that $$\alpha_n=\frac{1}{(q; q)_n}\sum_{k=0}^n {n\brack k}_q (v/s; q)_k s^k h_{n-k}(t, u|q).$$ Substituting the above equation into (\[liueqn6\]), we arrive at $$f(a, b)=\sum_{n=0}^\infty \frac{h_n(a, b|q)}{(q; q)_n} \sum_{k=0}^n {n\brack k}_q (v/s; q)_k s^k h_{n-k}(t, u|q).$$ Interchanging the order of the summation of the above equation and simplifying, we find that $$f(a, b)=\sum_{k=0}^\infty \frac{s^k(v/s; q)_k}{(q; q)_k} \sum_{m=0}^\infty \frac{h_{m+k}(a, b|q)h_m(t, u|q)}{(q; q)_m}. \label{liueqn7}$$ Using the Carlitz formula in Theorem \[Carlitzthm\], we find that the inner summation of the above equation equals $$\frac{(abtu; q)_\infty}{(at, au, bt, bu; q)_\infty} \sum_{j=0}^k {k\brack j}_q \frac{b^j a^{k-j}(at, au; q)_j}{(abtu; q)_j}.$$ Substituting the above equation into (\[liueqn7\]) and then interchanging the order of the summation, we conclude that $$f(a, b)=\frac{(av, abtu; q)_\infty}{(as, at, au, bt, bu; q)_\infty} {_3\phi_2}\left({{v/s, at, au}\atop{av, abtu}}; q, bs\right).$$ Using Sears’ $_3\phi_2$ transformation formula (see, for example, [@Liu2003 Theorem 3]), we have $${_3\phi_2}\left({{v/s, at, au}\atop{av, abtu}}; q, bs\right) =\frac{(abstu/v, bv; q)_\infty}{(abtu, bs; q)_\infty} {_3\phi_2}\left({{v/s, v/t, v/u}\atop{av, bv}}; q, \frac{abstu}{v}\right).$$ Combining the above two equations, we complete the proof of Proposition \[liuopeppb\]. In the same way, we can also obtain the following $q$-exponential operator identity [@Liu2010 Eq.(3.2)]. \[liuopeppc\]For $\max\{|av|, |bv|, |abstu/qv|\}<1, $ we have $$\begin{aligned} T(b \mathcal{D}_{q^{-1}, a})\left\{\frac{(as, at, au; q)_\infty}{(av; q)_\infty}\right\} &=\frac{(as, at, au, bs, bt, bu; q)_\infty}{(av, bv, abstu/qv; q)_\infty}\\ &\quad \times{_3\phi_2}\left({{s/v, t/v, u/v}\atop{q/av, q/bv}}; q, q\right).\end{aligned}$$ The following proposition reveals a relationship of $q$-integral and $q$-exponential operator. \[liuopeppd\] If $T(y \mathcal{D}_{q, x})$ is defined by $(\ref{liueqn1})$, then, we have $$\begin{aligned} &T(b \mathcal{D}_{q, a})\left\{\frac{(acst; q)_\infty}{(as, at, au; q)_\infty}\right\}\\ &=\frac{(cs, ct; q)_\infty}{(1-q)t (q, s/t, qt/s, au, bu; q)_\infty}\int_{s}^t \frac{(qx/s, qx/t, abux; q)_\infty}{(ax, bx, cx; q)_\infty} d_q x.\end{aligned}$$ If we use $f(a, b)$ to denote the right-hand side of the above equation, then, it is to verify that $f(a, b)$ is analytic near $(a, b)=(0, 0).$ A direct computation shows that $\partial_{q, a}\{f\}=\partial_{q, b}\{f\}$ equals $$\frac{(cs, ct; q)_\infty}{(1-q)t (q, s/t, qt/s, au, bu; q)_\infty}\int_{s}^t \frac{(qx/s, qx/t, qabux; q)_\infty(x+u-(a+b)ux)} {(ax, bx, cx; q)_\infty} d_q x.$$ Using the Andrews-Askey integral in Proposition \[AAintegralpp\], we easily find that $$\begin{aligned} f(a, 0)&= \frac{(cs, ct; q)_\infty}{(1-q)t (q, s/t, qt/s, au; q)_\infty}\int_{s}^t \frac{(qx/s, qx/t; q)_\infty}{(ax, cx; q)_\infty} d_q x\\ &=\frac{(acst, q)_\infty}{(au, as, at; q)_\infty}.\end{aligned}$$ Thus, using (i) in Lemma \[liulemope\], we immediately conclude that $$f(a, b)=T(b\mathcal{D}_{q, a})\{f(a, 0)\} =T(b \mathcal{D}_{q, a})\left\{\frac{(acst; q)_\infty}{(as, at, au; q)_\infty}\right\}.$$ Combining Proposition \[liuopeppb\] and Proposition \[liuopeppd\], we can obtain the following proposition, which is equivalent to [@Liu2010 Theorem 9]. \[liuopeppe\] For $\max\{|as|, |bs|, |cs|, |at|, |bt|, |ct|, |abu/c|\}<1,$ we have $$\begin{aligned} &\int_{s}^t \frac{(qx/s, qx/t, abux; q)_\infty}{(ax, bx, cx; q)_\infty} d_q x\\ &=\frac{(1-q)t(q, s/t, qst, acst, bcst, abu/c; q)_\infty} {(as, bs, cs, at, bt, ct; q)_\infty} {_3\phi_2}\left({{cs, ct, cst/u}\atop{acst, bcst}}; q, \frac{abu}{c}\right).\end{aligned}$$ \[liuopeppf\]Let $\{f_n(x)\}$ be a sequence of analytic functions near $x=0$, such that the series $$\sum_{n=0}^\infty f_n(x)$$ converges uniformly to an analytic function $f(x)$ near $x=0$, and the series $$\sum_{n=0}^\infty T(yD_{q, x})\{f_n(x)\}$$ converges uniformly to an analytic function $f(x, y)$ near $(x, y)=(0, 0)$. Then we have $f(x, y)=T(yD_{q, x})\{f(x)\}$, or $$T(yD_{q, x})\left\{\sum_{n=0}^\infty f_n(x)\right\}=\sum_{n=0}^\infty T(yD_{q, x})\{f_n(x)\}.$$ If we use $f_n(x, y)$ to denote $T(yD_{q, x})\{f_n(x)\}, $ then, by Lemma \[liulemope\], we find that $\partial_{q, x}\{f_n(x, y)\}=\partial_{q, y}\{f_n(x, y)\}$. It follows that $\partial_{q, x}\{f(x, y)\}=\partial_{q, y}\{f(x, y)\}.$ Thus, using Lemma \[liulemope\] again, we have $$f(x, y)=T(yD_{q, x})\{f(x, 0)\}=T(yD_{q, x})\{f(x)\}.$$ If we replace $T(yD_{q, x})$ by $T(yD_{q^{-1}, x})$ in proposition \[liuopeppf\], we obtain the following proposition. \[liuopeppg\]Let $\{f_n(x)\}$ be a sequence of analytic functions near $x=0$, such that the series $$\sum_{n=0}^\infty f_n(x)$$ converges uniformly to an analytic function $f(x)$ near $x=0$, and the series $$\sum_{n=0}^\infty T(yD_{q^{-1}, x})\{f_n(x)\}$$ converges uniformly to an analytic function $f(x, y)$ near $(x, y)=(0, 0)$. Then we have $f(x, y)=T(yD_{q^{-1}, x})\{f(x)\}$, or $$T(yD_{q^{-1}, x})\left\{\sum_{n=0}^\infty f_n(x)\right\}=\sum_{n=0}^\infty T(yD_{q^{-1}, x})\{f_n(x)\}.$$ Next we will use the above proposition to give an extension of $q$-Gauss summation. For this purpose, we also need Tannery’s Theorem (see, for example, [@Boas]). \[tanthm\] For each nonnegative integer $n$, let $s(n)=\sum_{k>o} f_k(n)$ is a finite sum or a convergent series. If $\lim_{n\to \infty} f_k(n)=f_k$, $|f_k(n)|\le M_k$, and $\sum_{k> 0}M_k<\infty,$ then $$\lim_{n\to \infty} s(n)=\sum_{k=0}^\infty f_k.$$ \[liuopeppgauss\] For $\max\{|x|, |abxy|\}<1,$ we have the summation $$\begin{aligned} \sum_{n=0}^\infty \frac{(a, b; q)_n x^n}{(q, abxy; q)_n} {_3\phi_2}{\left( }{{q^{-n}, 1/x, 1/y}\atop{a, b}}; q, abxyq^n{\right) }=\frac{(ax, bx; q)_\infty}{(x, abxy; q)_\infty}.\end{aligned}$$ When $y=1,$ the above proposition immediately reduces to the $q$-Gauss summation. If we denote the $_3\phi_2$ series in Proposition \[liuopeppgauss\] by $f_k(n), $ then, by a simple computation, we have $$\begin{aligned} &f_k(n) =(-1)^k {n\brack k}_q \frac{(1/x, 1/y ;q)_k}{(q, a, b; q)_k }q^{k(k-1)/2}(abxy)^k,\\ &\lim_{n\to \infty}f_k(n)=(-1)^k \frac{(1/x, 1/y ;q)_k}{(q, a, b; q)_k }q^{k(k-1)/2}(abxy)^k,\\ &\quad|f_k(n)| \le \frac{|(1/x, 1/y ;q)_k|}{(q;q)_k |(a, b; q)_k|} |abxy|^k=M_k.\end{aligned}$$ Using the ratio test, we find, for $|abxy|<1,$ that $\sum_{k \ge 0} M_k<\infty.$ Thus, by Tannery’s Theorem, we deduce that for $|abxy|<1,$ $$\lim_{n\to \infty} {_3\phi_2}\left({{q^{-n}, 1/x, 1/y}\atop{a, b}}; q, abxyq^n \right) =\sum_{k=0}^\infty (-1)^k \frac{(1/x, 1/y ;q)_k}{(q, a, b; q)_k }q^{k(k-1)/2}(abxy)^k.$$ Hence, using the ratio test, we find the series in Proposition \[liuopeppgauss\] converges uniformly to an analytic function $f(a, b)$ for $\max\{|x|, |abxy|\}<1.$ Rewrite the $q$-binomial theorem in the form $$\sum_{n=0}^\infty \frac{x^n}{(q; q)_n} \frac{(a; q)_\infty}{(aq^n, ax; q)_\infty} =\frac{1}{(x; q)_\infty}.$$ Multiplying both sides of the above equation by $1/(ay; q)_\infty,$ we deduce that $$\sum_{n=0}^\infty \frac{x^n}{(q; q)_n} \frac{(a; q)_\infty}{(aq^n, ax, ay; q)_\infty} =\frac{1}{(x; q)_\infty(ay; q)_\infty}. \label{qgauss1}$$ Using Proposition \[liuopeppb\], we have, for $\max\{|ax|, |ay|, |a|, |abxy|\}<1,$ that $$\begin{aligned} &\sum_{n=0}^\infty \frac{x^n}{(q; q)_n} T(b \mathcal{D}_{q, a})\left\{\frac{(a; q)_\infty}{(aq^n, ax, ay; q)_\infty}\right\}\\ &=\frac{( abxy; q)_\infty}{( ax, ay, bx, by; q)_\infty} \sum_{n=0}^\infty \frac{(a, b; q)_n x^n}{(q, abxy; q)_n} {_3\phi_2}{\left( }{{q^{-n}, 1/x, 1/y}\atop{a, b}}; q, abxyq^n{\right) }.\end{aligned}$$ We have proved that the right-hand side of the above equation is analytic near $(a, b)=(0, 0).$ Thus, we find that the right-hand side of the above equation equals $$\begin{aligned} T(b \mathcal{D}_{q, a})\left\{\sum_{n=0}^\infty \frac{x^n}{(q; q)_n} \frac{(a; q)_\infty}{(aq^n, ax, ay; q)_\infty}\right\} &=\frac{1}{(x; q)_\infty} T(b \mathcal{D}_{q, a})\left\{\frac{1}{(ay; q)_\infty}\right\}\\ &=\frac{1}{(x, ay, by; q)_\infty}.\end{aligned}$$ Combining the above two equations, we complete the proof of Proposition \[liuopeppgauss\]. In the same way, we can prove the following transformation for $q$-series, which reduces to the Jackson q-analogue of the Euler transformation when $b=y.$ \[qeulerpp\] For $\max\{|b|, |x|, |abxy|\}<1,$ we have $$\begin{aligned} \\ &\frac{(b; q)_\infty}{(x; q)_\infty} {_2\phi_1}\left({{ax, cx}\atop{acxy}}; q, b\right)\\ &=\sum_{n=0}^\infty \frac{(ab, bc; q)_n x^n}{(q, acxy; q)_n} {_3\phi_2}\left({{q^{-n}, b/x, b/y}\atop{ab, bc}}; q, abxyq^n\right).\end{aligned}$$ We end this section with an extension of the Sears $_4\phi_3$ transformation formula. \[liuopepph\] We have the $q$-transformation formula $$\begin{aligned} &\sum_{j=0}^n \frac{(q^{-n}, be^{i\theta}, be^{-i\theta}, abcduq^{n-1}/v; q)_j q^j} {(q, ab, bc, bd; q)_j}{_3\phi_2}\left({{bq^{j}/v, cq^{n}/v, u/v}\atop{q/av, q/dv}}; q, q\right)\\ &=\frac{(ac, cd; q)_n }{(ab, bd; q)_n }\left(\frac{b}{c}\right)^n\sum_{j=0}^n \frac{(q^{-n}, ce^{i\theta}, ce^{-i\theta}, abcduq^{n-1}/v; q)_j q^j} {(q, ac, bc, cd; q)_j}{_3\phi_2}\left({{cq^{j}/v, bq^{n}/v, u/v}\atop{q/av, q/dv}}; q, q\right).\end{aligned}$$ If we let $u=v$ in the above equation and use that obvious fact that $(1; q)_0=1$ and $(1; q)_k=0$ for $k\ge 1,$ we find that Proposition \[liuopepph\] reduces to the Sears $_4\phi_3$ transformation [@Liu2003 Theorem 4]. Recall the Sears $_3\phi_2$ transformation formula (see, for example, [@Liu2003 p. 123]), which states $$\begin{aligned} &{_3\phi_2}{\left( }{{a_1, a_2, a_3}\atop{b_1, b_2}}; q, \frac{b_1b_2}{a_1a_2a_3}{\right) }\\ &=\frac{(b_2/b_3, b_1b_2/a_1a_2; q)_\infty} {(b_2, b_1b_2/a_1a_2a_3; q)_\infty} {_3\phi_2}{\left( }{{b_1/a_1, b_1/a_2, a_3}\atop{b_1, b_1b_2/a_1a_2}}; q, \frac{b_2}{a_3}{\right) }.\end{aligned}$$ Taking $a_3=q^{-n}$ in the above equation, replacing $(a_1, a_2, b_1, b_2, q)$ by $$(1/a_1, 1/a_2, 1/b_1, 1/b_2, 1/q)$$ and performing some calculation, we obtain the transformation formula $$\begin{aligned} {_3\phi_2}{\left( }{{q^{-n}, a_1, a_2}\atop{b_1, b_2}}; q, q{\right) }=\frac{(b_1b_2/a_1a_2; q)_n}{(b_2; q)_n} {\left( }\frac{a_1a_2}{b_1}{\right) }^n {_3\phi_2}{\left( }{{q^{-n}, \frac{b_1}{a_1},\frac {b_1}{a_2}}\atop{b_1, \frac{b_1b_2}{a_1a_2}}}; q, q{\right) }.\end{aligned}$$ Replacing $(a_1, a_2, b_1, b_2)$ by $(be^{i\theta}, be^{-i\theta}, bc, ab)$ in the above equation and making some calculation, we find that $$\begin{aligned} &\sum_{j=0}^n \frac{(q^{-n}, be^{i\theta}, be^{-i\theta}; q )_j q^j}{(q, bc; q)_j}(abq^j, acq^n; q)_\infty\\ &={\left( }\frac{b}{c}{\right) }^n \sum_{j=0}^n \frac{(q^{-n}, ce^{i\theta}, ce^{-i\theta}; q )_j q^j}{(q, bc; q)_j}(acq^j, abq^n; q)_\infty.\end{aligned}$$ Multiplying both sides of the above equation by $(au; q)_\infty/(av; q)_\infty,$ and then taking the action of $T(d D_{q^{-1}, a})$ on both sides of the resulting equation, we deduce that $$\begin{aligned} &\sum_{j=0}^n \frac{(q^{-n}, be^{i\theta}, be^{-i\theta}; q )_j q^j}{(q, bc; q)_j}T(d D_{q^{-1}, a})\left\{\frac{(abq^j, acq^n, au; q)_\infty}{(av; q)_\infty}\right\}\\ &={\left( }\frac{b}{c}{\right) }^n \sum_{j=0}^n \frac{(q^{-n}, ce^{i\theta}, ce^{-i\theta}; q )_j q^j}{(q, bc; q)_j} T(d D_{q^{-1}, a})\left\{\frac{(acq^j, abq^n, au; q)_\infty}{(av; q)_\infty}\right\}.\end{aligned}$$ Using the $q$-exponential operator in Proposition \[liuopeppc\], we immediately find that $$\begin{aligned} T(d D_{q^{-1}, a})\left\{\frac{(abq^j, acq^n, au; q)_\infty}{(av; q)_\infty}\right\} &=\frac{(abq^j, acq^n, au, bdq^j, cdq^n, du; q)_\infty}{(av, dv, abcduv^{-1}q^{n+j-1}; q)_\infty}\\ &\quad\times {_3\phi_2}{\left( }{{q^{-n}, bq^j/v, cq^n/v, u/v}\atop{q/av, q/dv}}; q, q{\right) },\end{aligned}$$ $$\begin{aligned} T(d D_{q^{-1}, a})\left\{\frac{(acq^j, abq^n, au; q)_\infty}{(av; q)_\infty}\right\} &=\frac{(acq^j, abq^n, au, cdq^j, bdq^n, du; q)_\infty}{(av, dv, abcduv^{-1}q^{n+j-1}; q)_\infty}\\ &\quad\times {_3\phi_2}{\left( }{{q^{-n}, cq^j/v, bq^n/v, u/v}\atop{q/av, q/dv}}; q, q{\right) }.\end{aligned}$$ Combining the above three equations, we complete the proof of Proposition \[liuopepph\]. Acknowledgments =============== I am grateful to the referee for many very helpful comments. The author was supported in part by the National Science Foundation of China. [9]{} Al-Salam, W. A.; Verma, A.: Some remarks on $q$-beta integral, Proc. Amer. Math. Soc. 85 (1982) 360–362. Andrews, G. E.: Summations and transformations for basic Appell series, J. London Math. Soc. (2) 4 (1972) 618–622. Andrews, G. E.; Askey, R.: Another $q$-extension of the beta function, Proc. Amer. Math. Soc. 81 (1981) 97–100. Berndt, B. C.; Chan, S. H.; Yeap, B. P.; Yee, A.J.: A reciprocity theorem for certain $q$-series found in Ramanujan’s lost notebook, Ramanujan J. 13 (2007) 29–40. Boas, R. P.: Tannery¡¯s theorem, Math. Mag. 38 (1965) 66. L. Carlitz, Some polynomials related to theta functions, Ann. Mat. Pure Appl. 41 (1955) 359–373. Cao, J. : New proofs of generating functions for Rogers-Szegő polynomials, Appl. Math. Comput. 207 (2009) 486–492. Cao, J. : Notes on Carlitz’s $q$-operators, Taiwanese J. Math. 14 (2010) 2229–2244. Carlitz, L. : Generating functions for certain $q$-orthogonal polynomials, Collectanea Math. 23(1972) 91–104. Chen, W. Y. C.; Liu, Z.-G.: Parameter augmentation for basic hypergeometric series, I, Mathematical Essays in honor of Gian-Carlo Rota (edited by B. E. Sagan and R. P. Stanley) , Birkäuser, 1998, 111–129. Chen, W. Y. C.; Liu, Z.-G.: Parameter augmentation for basic hypergeometric series, II, J. Combin. Theory Ser. A 80 (1997) 175–195. Chu, W.; Zhang, W.: Bilateral $q$-series identities and reciprocal formulae. Funct. Approx. Comment. Math. 42 (2010) 153–162. Gasper, G.; Rahman, M.: Basic Hypergeometric Series, Cambridge Univ. Press, Cambridge, MA,2004. Ismail, M.; Rahman, M.; Suslov, K.: Some summation theorems and transformations for $q$-series, Can. J. Math. 49 (1997) 543–567. Kang, S. Y.: Generalizations of Ramanujan’s reciprocity theorem and their applications, J. London Math. Soc. (2) 75 (2007) 18–34. Liu, Z.-G.: Some operator identities and $q$-series transformation formulas, Disc. Math. 265 (2003) 119–139. Liu, Z.-G.: An identity of Andrews and the Askey–Wilson integral, Ramanujan J. 19 (2009) 115–119. Liu, Z.-G.: Two $q$-difference equations and $q$-operator identities, J. Difference Equ. Appl. 16 (2010) 1293–1307. Liu, Z.-G.: An extension of the non-terminating ${}_6\phi_5$ summation and the Askey-Wilson polynomials, J. Difference Equ. Appl. 17 (2011) 1401–1411. Liu, Z.-G.: A $q$-series expansion formula and the Askey-Wilson polynomials, Ramanujan J. 30 (2013) 193-210. Liu, Z.-G.: On the $q$-derivative and $q$-series expansion, Int. J. Number Theory, doi: 10.1142/S1793042113500759, in press. McLaughlin, J.; Sills, A. V.: Rogers-Ramanujan-Slater type identities, Electronic J. Combin. 15 (2008) 59 pp. Malgrange, B.: Lectures on functions of several complex variables, Springer-verlag, Berlin, 1984. Range, R. M. : Complex analysis: A brief tour into higher dimensions, Am. Math. Monthly 110 (2003) 89–108. Rogers, L. J.: On a three-fold symmetry in the elements of Heine’s series, Proc. London. Math. Soc. 24 (1893)171–179. Szegő, G.: Ein Betrag zur Theorie der Thetafunktionen, Sitz. Preuss. Akad. Wiss. Phys. Math. 19 (1926) 242–252. Taylor, J.: Several Complex Variables with Connections to Algebraic Geometry and Lie Groups, Graduate Studies in Mathematics, vol. 46. Am. Math. Soc., Providence, 2002. Verma, A.; Jain, V. K. : Transformations between basic hypergeometric series on different bases and identities of Rogers-Ramanujan type, J. Math Anal. Appl. 76 (1980) 230-269. Zhang, Z. ; Wang, J.: Two operator identities and their applications to terminating basic hypergeometric series and $q$-integrals, J. Math. Anal. Appl. 312 (2005) 653–665. Zhang, Z.: Operator identities and several $U(n+1)$ generalizations of the Kalnins-Miller transformations, J. Math. Anal. Appl. 324 (2006) 1152–1167. [^1]: Dedicated to Srinivasa Ramanujan on the occasion of his 125th birth anniversary [^2]: 2010 Mathematics Subject Classifications : 05A30, 33D15, 11E25, 11F27. [^3]: Keywords: $q$-series, $q$-derivative, $q$-partial differential equation, $q$-exponential operator, $q$-identities, analytic functions, Rogers-Szegő polynomials
{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent photometric analysis of the colour-magnitude diagrams (CMDs) of young massive clusters (YMCs) have found evidence for splitting in the main sequence and extended main sequence turn-offs, both of which have been suggested to be caused by stellar rotation. Comparison of the observed main sequence splitting with models has led various authors to suggest a rather extreme stellar rotation distribution, with a minority ($10-30$%) of stars with low rotational velocities and the remainder ($70-90$%) of stars rotating near the critical rotation (i.e., near break-up). We test this hypothesis by searching for Be stars within two YMCs in the LMC (NGC 1850 and NGC 1856), which are thought to be critically rotating stars with decretion disks that are (partially) ionised by their host stars. In both clusters we detect large populations of Be stars at the main sequence turn-off ($\sim30-60$% of stars), which supports previous suggestions of large populations of rapidly rotating stars within massive clusters.' author: - | N. Bastian$^{1}$, I. Cabrera-Ziri$^{1,2}$, F. Niederhofer$^{3}$, S. de Mink$^{4}$, C. Georgy$^{5}$, D. Baade$^{2}$, M. Correnti$^{3}$, C. Usher$^{1}$, M. Romaniello$^{2}$\ $^{1}$Astrophysics Research Institute, Liverpool John Moores University, 146 Brownlow Hill, Liverpool L3 5RF, UK\ $^{2}$European Southern Observatory, Karl-Schwarzschild-Straße 2, D-85748 Garching bei München, Germany\ $^{3}$Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA\ $^{4}$Astronomical Institute Anton Pannekoek, University of Amsterdam, PO Box 94249, NL-1090GE Amsterdam, the Netherlands\ $^{5}$Geneva Observatory, University of Geneva, Maillettes 51, 1290, Sauverny, Switzerland\ date: 'Accepted. Received; in original form' --- \[firstpage\] galaxies - star clusters Introduction {#sec:intro} ============ A number of recent studies have found split main-sequences (split-ms) within young massive clusters in the Large Magellanic Cloud (LMC - Milone et al. 2015, 2016, Bastian et al. 2016; Niederhofer et al. in prep., Correnti et al. in prep.). Attempts to model the split-ms using different aged populations, CNO variation, or Fe-spreads have not resulted in consistent fits to the observations (e.g., Milone et al. 2015). However, stellar rotation appears to provide satisfactory fits (e.g., D’Antona et al. 2015) which is in agreement with studies that have explained the observed extended main-sequence turnoffs (eMSTOs) within young and intermediate age clusters through stellar rotation effects (e.g., Bastian & de Mink 2009; Brandt & Huang 2015; Niederhofer et al. 2015b). One potential complication with this explanation, however, is that in order to have discrete main-sequences within the young clusters, a rather extreme distribution of stellar rotation rates is required, namely a bi-modal distribution with a peak near zero velocity and another near the critical rotation[^1] ($\omega = \Omega_{\rm rot} / \Omega_{\rm critical} \sim 0.9$ - i.e., D’Antona et al. 2015). In the massive clusters studied so far, the “red ms" (interpreted as being due to the high $\omega$ stars) is the dominant one, containing $\sim67$% or more of the stars within the magnitude interval where it is visible, although there is considerable scatter from cluster to cluster (e.g., Milone et al. 2015, 2016; Niederhofer et al. in prep.). D’Antona et al. (2015) have attempted to explain this bi-modal rotational distribution by assuming that all stars begin their lives rapidly rotating and stars in binary systems can be effectively braked due to tidal synchronization. However, it is currently unexplained why stars within massive clusters would all begin their lives as rapid rotators (at least in the stellar mass range explored so far) as stars in the field or in the lower mass open clusters do not appear to contain large fractions of rapid rotators ($<10\%$ rapid rotators, e.g., McSwain & Gies 2005). In order to confirm or refute the existence of a large population of rapidly rotating stars within young massive clusters, the ideal way would be to spectroscopically measure rotation rates for large samples of stars, sampling both the ’blue ms’ and ’red ms’. An alternative method is to search for Be stars within the cluster, based on the presence of H$\alpha$ emission (e.g., Keller et al. 2000). Be stars are rapidly rotating stars, with $\omega \ge 0.88-0.95$ (e.g., Townsend et al. 2004; Frémat et al. 2005; Delaa et al. 2011; Meilland et al. 2012; Rivinius et al. 2013), that develop decretion discs that become ionised by the host star, hence appear in H$\alpha$ emission. While this measurement is certainly more crude than the measurements of stellar rotation rates directly, it should be able to discern whether a large population of rapidly rotating stars exist within the clusters. In the present work, we search for Be stars in two young massive clusters in the LMC, NGC 1850 and 1856, both of which have been observed to have split main sequences (Bastian et al. 2016; Niederhofer et al. in prep; Correnti et al. in prep. and Milone et al. 2015, respectively). We use existing HST based photometric catalogues and archival data to search for stars with H$\alpha$ emission. In § \[sec:obs\] we present the catalogues used and in § \[sec:analysis\] we present an analysis techniques. Finally, in § \[sec:discussion\] we discuss our results and their implications and present our conclusions. [lcccccc]{} ID& log (Age/yr) & log (Mass/) & Be star fraction & M$_{\rm TO}$ ()$^{a}$.\ NGC 1850 & 7.9 & 4.86 & 0.19-0.62$^{b}$ & $4.6-5$\ NGC 1856 & 8.45 & 4.88 & 0.33 & $2.9-3.1$\ Observations, Catalogues and Methods {#sec:obs} ==================================== Catalogues ---------- For the present study we make use of previously published catalogues and archival data. For NGC 1850, we use HST/WFC3 F336W and F438W photometry presented in Bastian et al. (2016). Additionally, we use the HST/WFPC2 images in F656N and F675W which were part of the GO-6101 programme (PI Gilmozzi) in order to search for stars with excess H$\alpha$ emission. The WFPC2 images were reduced and the photometry carried out in the same way as described in Panagia et al. (2000) and Romaniello et al. (2002). We focus on the inner $5.3$ pc (22") region of the cluster in order to avoid any contaminating H$\alpha$ emission from the ionised nearby regions. For NGC 1856 we use the HST/WFC3 photometry presented in Correnti et al. (2015), namely the F438W, F555W, F814W, and the F656N measurements. As in Correnti et al. (2015) we find that there exists differential extinction across the field of view. We have corrected for this using the technique described in Milone et al. (2012), however we note that correcting for differential extinction or not does not change our main conclusions. Methods ------- In order to search for stars with H$\alpha$ excess we follow the method of Keller et al. (2000). First, we make a colour-magnitude diagram (CMD) with the continuum flux of a star (e.g., either the F675W magnitude or a linear combination of nearby filters such as the F555W and F814W - see Zeidler et al. 2015) minus the F656N magnitude (which includes both the continuum and any H$\alpha$ emission that is present) vs. F675W (or a nearby filter). Sources with H$\alpha$ excess scatter to the red in such a diagram. Next we fit a polynomial to the locus of points without H$\alpha$ excess (i.e. the nominal ridge line). Finally, we search for stars that are $3\sigma$ or more to the red from the ridge line, based on their estimated errors. The selection of stars with excess H$\alpha$ emission is shown in the left panels of Figs. \[fig:cmd\_n1850\] &  \[fig:cmd\_n1856\], for NGC 1850 and NGC 1856, respectively. In the right panels of Figs. \[fig:cmd\_n1850\] &  \[fig:cmd\_n1856\] we show the position of the H$\alpha$ excess stars in the broad-band CMDs of the clusters. The dashed horizontal lines indicate the magnitude at which we expect to be complete until, based on histograms of the H$\alpha$ magnitudes within the clusters. Correnti et al. (2015) used a similar approach (see also McSwain & Gies 2005) for NGC 1856 to identify H$\alpha$ emitting stars, but they focussed their attention on fainter magnitudes (and redder broad-band colours) in order to search for pre-main sequence stars. However, they also found a large population of H$\alpha$ emitting stars near the main sequence turn-off. Both clusters host a large population of Be stars within them. We note, however, that the method used here results in a lower limit of Be stars within the cluster, as the Be phase itself is transitory (e.g., Rivinius et al. 2013, and references therein) as Be stars can temporarily lose their line emission, i.e., their disk, completely (e.g., Baade et al. 1988). However, the relative fraction of time stars spend in the emission or non-emission state is currently unknown, so we will use the observed values as lower limits. We note that the H$\alpha$ excess observed in stars in NGC 1850 is significantly larger than for NGC 1856. This is likely due to the different ages of the clusters (NGC 1850 is $\sim80$ Myr while NGC 1856 is $\sim280$ Myr - e.g., Bastian & Silva-Villa 2013), resulting in significant temperature differences between stars on the respective turnoffs. For NGC 1850 we expect the turn-off stars to be $\sim5$ , while for NGC 1856 the turnoff stars have masses of $\sim3$ . The difference in temperature, and subsequently ionising flux between the stars on the turnoff, mean that the H$\alpha$ excess is expected to be larger in NGC 1850, as observed. In Table \[tab:objects\] we give the basic properties of the clusters and their stellar populations. As a point of reference, the splitting of the main sequence in NGC 1850 can be seen just above and below the dashed horizontal line the right panel of Fig. \[fig:cmd\_n1850\]. This will be studied in more detail in Niederhofer et al. (in prep) and Correnti et al. (in prep.). The split in the main sequence in NGC 1856 is not clearly visible in the right panel of Fig. \[fig:cmd\_n1856\] due to the choice of filters, as it becomes more evident in ultraviolet filters (e.g., Milone et al. 2015, D’Antona et al. 2015). ![image](n1850_ha_excess.eps){width="8cm"} ![image](n1850_cmd_ha.eps){width="8cm"} ![image](n1856_ha_excess.eps){width="8cm"} ![image](n1856_cmd_ha.eps){width="8cm"} Analysis {#sec:analysis} ======== NGC 1850 -------- As seen in Fig. \[fig:cmd\_n1850\] a significant fraction of stars near the MSTO in NGC 1850 display H$\alpha$ emission. We have quantified the fraction of stars with H$\alpha$ emission as a function of magnitude and find fractions of 0.04, 0.19, 0.40 and 0.62 for the magnitude ranges of $20 > F438W \ge 19$, $19 > F438W \ge 18$, $18 > F438W \ge 17$, $17 > F438W \ge 16.5$, respectively. A dependence of the fraction of H$\alpha$ emitting stars as a function of magnitude has been previously found (Keller et al. 2000; 2001). This may simply be due to the fact that brighter stars (at a fixed age) have higher temperatures, allowing for the discs to be more ionised and hence brighter in emission lines. Alternatively it may be that stars are more likely to possess discs as their approach the end of their MS lifetimes. Another possible cause is that at the brightest luminosities, stars that are non-rapid rotators will have already left the MS, leaving only rapid rotators behind (which stay on the MS for longer due to rotational mixing). We can compare the observed Be star fraction with that between the blue-MS and red-MS, which are the slow/non-rotating and rapidly rotating stars, respectively in the interpretation that their position is due to rotation (e.g., D’Antona et al. 2015). The splitting in the MS can be seen in the right panel of Fig. \[fig:cmd\_n1850\] just above the dashed horizontal line. Niederhofer et al. (in prep) show that the red-MS is the dominant population, making up $\sim92$% at a given luminosity (where the two sequences can be clearly separated). This is similar to that found in NGC 1755 (Milone et al. 2016), a similarly aged ($\sim80$ Myr), although slightly lower mass ($\sim3\times10^4$ ) cluster, where the red-MS makes up $\sim75$% of the population. For NGC 1850, the high fraction of Be stars is consistent with the high fraction of red-MS stars, given that the Be star fraction is a lower limit. Finally, we have created a synthetic cluster using the SYCLIST models (Georgy et al. 2014) made up of two populations at an age of 100 Myr, one that is made up of slow rotators ($\omega < 0.1$) and one where all stars are rotating at $\omega=0.9$, with the two populations having the same number of stars. The results are shown in Fig. \[fig:cmd\_example\], where the two separate MSs can be seen. The inclination angles have been distributed randomly for both populations. Note that the rapidly rotating stars are observed across the entire MSTO region and not just located on one side. This is consistent with the observations (i.e., Fig. \[fig:cmd\_n1850\]) and may make it difficult to directly associate Vsini measurements with the spread in the MSTO. We note that these simulations do not include extinction to the stars from the decretion disc itself. Hence, in the observations we would expect some of the rapidly rotating stars (those seen nearly edge on) to shift towards fainter magnitudes and redder colours, consistent with the observations (see Fig. \[fig:cmd\_n1850\]). The models also do not include include Balmer continuum emission (from the stars with strong Balmer line emission). NGC 1856 -------- Next we study the Be fraction in NGC 1856. As this cluster is significantly older than NGC 1850 ($\sim300$ Myr vs. $\sim100$ Myr) the stars on the MSTO have lower masses ($\sim3$ ), and hence have lower ionising fluxes. Hence, some stars with discs may not be observable as Be stars, i.e. only stars with relatively strong emission will be detectable. For stars brighter than $V=19.5$ we find a Be star fraction of 0.33. Unlike NGC 1850, however, we do not find any significant change in the Be star fraction as a function of magnitude. This may be due to the fact that at this mass the ionising flux from the host star changes less steeply with magnitude. We can compare the observed Be star fraction with the ratio of blue (non-rotating) and red (rapidly rotating) stars on the MS. In the magnitude range where the two sequences were readily distinguishable Milone et al. (2015) found that the red MS dominates by number, making up $\sim2/3$ of the stars. Note that in Fig. \[fig:cmd\_n1856\] the splitting of the MS is not seen, which is due to the filter choice (the split is larger in the ultraviolet - e.g., Milone et al. 2015; D’Antona et al. 2015). The range where the two sequences in the CMD can be differentiated is $\sim1.7 - 3$ . The observe Be star fraction of 0.33 is smaller than the fraction of the red MS ($\sim0.67$), however as discussed above, the Be star fraction is a lower limit to the number of rapid rotators in the cluster, and it is significantly higher than in typical open clusters or in the field (e.g., McSwain & Gies 2005). In NGC 1856, the fraction of stars on the red MS and the fraction of Be stars is lower than in NGC 1850. The conclusion reached here is similar as that for NGC 1850, that a significant fraction of stars within the cluster are rotating near the critical rotation rate. Fig. \[fig:cmd\_example\] (bottom panel) shows a synthetic cluster made with the SYCLIST models for an age similar to NGC 1856 ($\sim300$ Myr). As was found for NGC 1850 (top panel), the rapidly rotating stars are found across the MSTO, not just on a single side, consistent with the observed colour/magnitude distribution of the H$\alpha$ excess stars. ![[**Top:**]{} A synthetic cluster made with the SYCLIST models with two populations that differ only in their rotation rates ($\omega=0$ and $\omega=0.9$) at an age of 100 Myr (meant to broadly reproduce NGC 1850). Representative errors were taken from the observations of NGC 1850, and we have not included binaries. Additionally, we have not taken into account reddening of stars that have decretion discs seen nearly edge on, nor have we included bound-free Balmer continuum emission from the disc itself. Note that rapid rotators are expected to be present across the MSTO, and not just limited to one side. [**Bottom:**]{} The same but now for an age of 300 Myr, meant to broadly reproduce NGC 1856. In both panels note that the filters are similar, but not identical to, the HST filters used in the present work. []{data-label="fig:cmd_example"}](cmd_model_example.eps "fig:"){width="8cm"} ![[**Top:**]{} A synthetic cluster made with the SYCLIST models with two populations that differ only in their rotation rates ($\omega=0$ and $\omega=0.9$) at an age of 100 Myr (meant to broadly reproduce NGC 1850). Representative errors were taken from the observations of NGC 1850, and we have not included binaries. Additionally, we have not taken into account reddening of stars that have decretion discs seen nearly edge on, nor have we included bound-free Balmer continuum emission from the disc itself. Note that rapid rotators are expected to be present across the MSTO, and not just limited to one side. [**Bottom:**]{} The same but now for an age of 300 Myr, meant to broadly reproduce NGC 1856. In both panels note that the filters are similar, but not identical to, the HST filters used in the present work. []{data-label="fig:cmd_example"}](cmd_model_example_bi.eps "fig:"){width="8cm"} Discussion and Conclusions {#sec:discussion} ========================== Keller et al. (2000) have used HST/WFPC2 photometry to search for Be stars within three LMC (NGC 1818, 2004, 2100) and one SMC (NGC 330) clusters with ages between $\sim15$ and $\sim25$ Myr. They found Be star fractions of $10-15$% for stars brighter than $V=19$, similar to, although slightly lower than found in NGC 1850 and NGC 1856 in the current study. However, they also found that the Be star fraction increases substantially towards the MSTO, peaking at 0.35-0.5. The authors note that these values are substantially higher than that found in the field, which is suggestive that stellar rotation (at least this evolutionary state) is dependent on the surrounding environment. Martayan et al. (2010) used slitless spectroscopy to search for Be stars in a sample of SMC clusters. They also found that the Be star fraction depends on spectral type (luminosity), but that by spectral types of B5 or B6 (those at the MSTO in NGC 1850) the fraction of Be stars was below 5%. Hence, NGC 1850 and NGC 1856 have much higher Be star fractions than the field or lower mass clusters, when similar spectral types are compared. As discussed previously, the dual MSs observed in a growing number of young massive clusters in the LMC/SMC is likely due to a bi-modal rotational distribution, with one peak at lower rotation rates and another near the critical rotation rate. Such a large population of rapidly rotating stars would appear to be unique amongst stars within massive clusters. The observations of large fractions of Be stars within such clusters lends support for this interpretation (i.e. of the presence of a large population of rapidly rotating stars within clusters). It is worth noting that Dufton et al. (2013) found a bi-modal rotational distribution of early-type B stars in the young region 30 Doradus in the LMC. Hence, such distributions appear at all ages ($<10$ Myr in 30 Doradus to $\sim300$ Myr for NGC 1856) which implies that it is imprinted at birth (see also - see also Martayan et al. 2007) and is not an evolutionary feature (i.e. binary interactions slowing down rapid rotators to form the peak at low rotational periods - e.g., D’Antona et al. 2015). While a consensus on the minimum rotation rate required for Be stars to form has not been fully reached, the majority of estimates place this value at $\omega_{\rm Be} \approx 0.8$ or higher (c.f. Rivinius et al. 2013 and references therein). This is in agreement with the rapid rotation rates required by D’Antona et al. (2015) to explain the dual MSs observed in NGC 1856 ($\omega \approx 0.9$). However, we note that while Be stars are good tracers of rapidly rotating stars, the Be phenomenon is intermittent, so there is likely a population of rapidly rotating stars that we have missed with our method. Hence, the observed fraction of stars that display H$\alpha$ emission is a lower limit to the actual number of Be stars within the clusters. Huang & Gies (2008) and Huang et al. (2010) have argued that while the incidence of rapid stellar rotators amongst B stars is higher in clusters than in the field, this is due to the relative youth of the clusters in previous surveys, and that the field stars were on average older, meaning that they have had a longer time to spin down. The two clusters studied here have densities more than an order of magnitude higher than those used in the Huang et al. study, and are significantly older than the clusters used in that work. The fact that we find high fractions of rapidly rotating stars in these two relative old clusters leads us to conclude that stellar density does play a major role in setting the rotation rates of stars, at least near globular cluster type masses ($\sim10^5$ ). We conclude that high fraction of Be stars within these two YMCs, which implies a high fraction of rapidly rotating stars, argues in favour of the interpretation that the observed main sequence splits in YMCs is due to a bi-modal rotation distribution, with a majority of stars rotating near the critical velocity (e.g., D’Antona et al. 2015). Such high fractions of rapid rotators are not seen in the field (c.f. Keller et al. 2000; Rivinius et al. 2013), suggesting that the cluster environment may impact the angular momentum distribution of the stars within them, which in turn may affect their evolution. Such a high fraction of rapid rotators would manifest in extended main sequence turn-offs in clusters, as often observed in young and intermediate age clusters (e.g., Milone et al. 2009; Niederhofer et al. 2015b). A high fraction of rapid rotators would be unique in stellar clusters and shows that the environment in which a star forms may influence its further evolution. A number of anomalies (i.e. chemical and morphological CMD features) have been observed in the ancient globular clusters, i.e., multiple populations. If GCs are indeed the ancient analogues to YMCs forming locally (see Kruijssen 2015 for a recent summary) we may expect them to share the rotation distribution properties observed in clusters like NGC 1850 and 1856. Such potential stellar evolutionary means to form multiple populations (to explain why they are only found in GCs and not in the field in large numbers) may be an attractive avenue for future research as all scenarios put forward so far, mostly based on multiple generations of star-formation taking place within clusters, have significant problems and are all but ruled out (e.g., Bastian 2015). In order to quantify the exact rotation rate distribution (Vsini) within clusters dedicated medium/high resolution spectroscopic surveys of cluster members are required. The MSTO region of the CMDs of the two clusters presented here are well within reach of the capabilities of instruments like FLAMES/UVES on the VLT. Acknowledgments {#acknowledgments .unnumbered} =============== NB gratefully acknowledges financial support from the Royal Society (University Research Fellowship) and the European Research Council (ERC-CoG-646928, Multi-Pop). F.N. gratefully acknowledges financial support for this project provided by NASA through grant HST-GO-14069 from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS526555. SdM acknowledges support by a Marie Sklodowska-Curie Action (H2020 MSCA-IF-2014, project id 661502). [99]{} Baade, D., Dachs, J., van de Weygaert, R., & Steeman, F. 1988, A&A, 198, 211 Bastian, N., & de Mink, S. E. 2009, MNRAS, 398, L11 Bastian, N., & Silva-Villa, E. 2013, MNRAS, 431, L122 Bastian, N. 2015, IAUS 319 (arXiv:1510.01330) Bastian, N., Niederhofer, F., Kozhurina-Platais, V., et al. 2016, MNRAS, 460, L20 Brandt, T. D., & Huang, C. X. 2015, ApJ, 807, 25 Correnti, M., Goudfrooij, P., Puzia, T. H., de Mink, S. E. 2015, MNRAS, 450, 3054 D’Antona, F., Di Criscienzo, M., Decressin, T., et al. 2015, MNRAS, 453, 2637 Delaa, O., Stee, P., Meilland, A., et al. 2011, A&A, 529, A87 Dufton, P. L., Langer, N., Dunstall, P. R., et al. 2013, A&A, 550, A109 Fr[é]{}mat, Y., Zorec, J., Hubert, A.-M., & Floquet, M. 2005, A&A, 440, 305 Georgy, C., Granada, A., Ekström, S., et al. 2014, A&A, 566, 21 Gilmozzi, R., Kinney, E. K., Ewald, S. P., Panagia, N., & Romaniello, M. 1994, ApJL, 435, L43 Granada, A., Ekstr[ö]{}m, S., Georgy, C., et al. 2013, A&A, 553, A25 Huang, W., & Gies, D. R. 2008, ApJ, 683, 1045-1051 Huang, W., Gies, D. R., & McSwain, M. V. 2010, ApJ, 722, 605 Keller, S. C., Bessell, M. S., & Da Costa, G. S. 2000, AJ, 119, 1748 Keller, S. C., Grebel, E. K., Miller, G. J., & Yoss, K. M. 2001, AJ, 122, 248 Kruijssen, J. M. D. 2015, MNRAS, 454, 1658 Martayan, C., Fr[é]{}mat, Y., Hubert, A.-M., et al. 2007, A&A, 462, 683 Martayan, C., Baade, D., & Fabregat, J. 2010, A&A, 509, A11 McSwain, M. V., & Gies, D. R. 2005, ApJS, 161, 118 Meilland, A., Millour, F., Kanaan, S., et al. 2012, A&A, 538, A110 Milone, A. P., Bedin, L. R., Piotto, G., & Anderson, J. 2009, A&A, 497, 755 Milone, A. P., Piotto, G., Bedin, L. R., et al. 2012, ApJ, 744, 58 Milone, A. P., Bedin, L. R., Piotto, G, et al. 2015, MNRAS, 450, 3750 Milone, A. P., Marino, A. F., D’Antona, F., et al. 2016, MNRAS, 458, 4368 Niederhofer, F., Hilker, M., Bastian, N., & Silva-Villa, E. 2015a, A&A, 575, A62 Niederhofer, F., Georgy, C., Bastian, N., & Ekstr[ö]{}m, S. 2015b, MNRAS, 453, 2070 Panagia, N., Romaniello, M., Scuderi, S., & Kirshner, R. P. 2000, ApJ, 539, 197 Rivinius, T., Carciofi, A. C., & Martayan, C. 2013, A&ARv, 21, 69 Romaniello, M., Panagia, N., Scuderi, S., & Kirshner, R. P. 2002, AJj, 123, 915 Townsend, R. H. D., Owocki, S. P., & Howarth, I. D. 2004, MNRAS, 350, 189 Zeidler, P., Sabbi, E., Nota, A., et al. 2015, AJ, 150, 78 \[lastpage\] [^1]: As in previous works, i.e. Granada et al. (2013), we use critical rotation to mean the velocity required for the centrifugal force to counterbalance gravity at the equator.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the degeneracy of topological solitons in the gauged O(3) non-linear sigma model with Chern-Simons term may be removed by chosing a self-interaction potential with symmetry - breaking minima. The topological solitons in the model has energy,charge,flux and angular momentum quantised in each topological sector.' --- = 16true cm = 20true cm =-1true cm =-2true cm plus 1pt [**ON THE QUESTION OF DEGENERACY OF TOPOLOGICAL SOLITONS IN A GAUGED O(3) NON-LINEAR SIGMA MODEL WITH CHERN-SIMONS TERM**]{} [^1]\ [Department of Physics\ A.B.N.Seal College,Coochbehar\ West Bengal,India]{} The 2+1 dimensional O(3) nonlinear sigma model has been studied extensively over a long period of time\[1\] due to its own interest of providing topologically stable soliton solutions which are exactly integrable in the Bogomol’nyi limit \[2\] and also for its applications in condensed matter physics \[3\].Soliton solutions of the model modified by the addition of Hopf term characterising maps from $S^3$ to $S^2$ reveal the occurence of fractional spin and statistics \[4,5\]. The system can be cast in the form of a genuine gauge theory by the inclusion of the Chern - Simons (C-S) term which implements fractional spin and statistics in the context of local field theories .A gauge-independent analysis \[6\] of the O(3) sigma model with C-S term \[7\] shows conclusively that this fractional spin is a physical effect and not an artifact of the gauge.The fractional spin of the topologically stable solitons of the model were shown to scale as the square of the vortex number .It is also established that this property is not specific of the particular model \[8-11\] and rather shared by the Chern-Simons vortices in general \[12\]. A characteristic feature of the soliton solutions of the O(3) sigma model in (2+1) dimensions is its scale - invariance which prevents particle interpretation on quantisation \[13\]. An interesting method of breaking this scale-invariance is to gauge the U(1) subgroup as well as including a potential term \[14\].Recently it has been shown that gauging the U(1) subgroup by Chern - Simons term with a particlar potential gives rise to both topological and nontopological solitons\[15\]. However the topological solitons are infinitely degenerate in a given topo- logical sector with quantised energy but degenerate charge , flux and angular momentum .This is certainly very characteristic outcome in the light of the findings detailed above about the Chern - Simons vortices. The potential used in \[15\] has two discrete minima $\phi_3 = \pm 1$ and the U(1) symmetry is not spontaneously broken.The topologically stable soliton solutions of the model are classified according to the homotopy $\Pi_2(S_2)=Z$ just as the model without the gauge field coupling.The observed infinite degeneracy in each topological sector is thus physically undesirable .In the present letter we will show that inclusion of a different form of self-interaction with symmetry breaking minima leads to topologically stable soliton solutions which have all the desired features with quantised energy,charge,flux and angular momentum in each topological sector. The Lagrangian of our model is given by $${\cal L} ={1 \over 2}D_\mu {\bf \phi}\cdot D^\mu{\bf \phi} +{k \over 4}\epsilon^{\mu\nu\lambda}A_\mu\partial_\nu A_\lambda+ U({\bf \phi})$$ Here ${\bf \phi}$ is a triplet of scalar fields constituting a vector in the internal space with unit norm $$\begin{aligned} \phi_a = {\bf n_a}\cdot {\bf \phi},(a=1,2,3)\\ {\bf \phi}\cdot{\bf \phi} =\phi_a\phi_a= 1\end{aligned}$$ where ${\bf n}_a$ constitute a basis of unit orthogonal vectors in the internal space. We work in the Minkowskian space - time with the metric tensor diagonal, $g_{\mu\nu} = (1,-1,-1)$. $D_\mu {\bf \phi}$ is the covariant derivative given by $$D_\mu {\bf{\phi}}=\partial_\mu {\bf{\phi}} + A_\mu {\bf n}_3\times{\bf{\phi}}$$ The SO(2) (U(1)) subgroup is gauged by the vector potential $A_\mu$ whose dynamics is dictated by the CS term.The potential $$U({\bf{\phi}})=-{1 \over {2k^2}} \phi_3^2(1-\phi_3^2)$$ gives a self interaction of the fields $\phi_a$.Note that the minima of the potential arise when either, $$\begin{aligned} \phi_1 =& \phi_2 = 0\hspace{.2cm} and \hspace{.2cm}\phi_3 = \pm 1\\ or,\phi_3 =& 0\hspace{.2cm} and \hspace{.2cm}\phi_1^2+\phi_2^2 =1\end{aligned}$$ In (6) the U(1) symmetry is unbroken whereas (7) corresponds to the spontaneous breaking of the same.For obvious reasons we will refer to (6) as the symmetric minima and (7) as the symmetry breaking minima. The Euler - Lagrange equations of the system (1) is derived subject to the constraint (3) by the Lagrange multiplier technique $$\begin{aligned} D_\nu (D^\nu {\bf{ \phi}})& =& [D_\nu (D^\nu {\bf{ \phi}}) \cdot{\bf{ \phi}}]{\bf{ \phi}} -{1\over k^2}{\bf n}_3\phi_3(1-2\phi_3^2) +{1\over k^2}\phi_3^2(1-2\phi_3^2){\bf \phi}\\ {k \over 2}\epsilon^{\mu\nu\lambda}F_{\nu\lambda}& =& j^\mu\end{aligned}$$ where $$j^\mu = -{\bf n}_3\cdot{\bf J}^\mu\hspace{.2cm} and\hspace{.2cm} {\bf J}^\mu ={\bf{ \phi}}\times D^\mu {\bf{ \phi}}$$ Using (8) we get $$D_\mu {\bf J}^\mu = {1 \over k^2}({\bf n}_3 \times {\bf {\phi}})\phi_3(1-2\phi_3^2)$$ From (9) we find $$j^0 = k\epsilon_{ij}\partial^iA^j = - kB$$ where B = curl[**A**]{} is the magnetic field. Integrating (12) over the entire space we obtain $$\Phi = -{Q \over k}$$ where Q is the charge and $\Phi$ is the magnetic flux.The relation (13) is characteristic of the CS theories. The energy functional is now obtained from Schwinger’s energy - momentum tensor \[16\] which in the static limit becomes $$E = {1 \over 2}\int d^2x[(D_i{\bf {\phi}})\cdot(D_i{\bf {\phi}})+ {{k^2B^2}\over{1-\phi_3^2}}+{1 \over k^2}\phi_3^2(1-\phi_3^2)]$$ We have eliminated $A_0$ using (10) and (12).The energy functional (14) is subject to the constraint (3). We can construct a conserved current $$K_\mu = {1 \over {8\pi}}\epsilon_{\mu\nu\lambda}[ {\bf {\phi}}\cdot D^\nu {\bf {\phi}} \times D^\lambda{\bf {\phi}} - F^{\nu\lambda}\phi_3]$$ By a straightforward calculation it can be shown that $$\partial_\mu K^\mu = 0$$ The corresponding conserved charge is $$T = \int d^2x K_0$$ Using (15) and (17) we can write $$T = \int d^2x[{1 \over{8\pi}}\epsilon_{ij}{\bf {\phi}} \cdot(\partial^i{\bf {\phi}} \times \partial^j {\bf {\phi}})] +{ 1 \over {4\pi}}\int_{boundary}\phi_3 A_\theta r d\theta$$ where r,$\theta$ are polar coordinates in the physical space and $A_\theta = {\bf e}_\theta \cdot {\bf A}$. Let us now consider the symmetric minima (6).For finite value of the energy functional (14) we require the fields at the spatial infinity to be equal to either $\phi_3=1 or -1$.The physical infinity is thus one point compactified to either the north or the south pole of the internal sphere.The static field configurations are thus classified according to the degree of the mapping from $S_2$ to $S_2$.Note that the first term of (18) gives the winding number of the mapping \[3\].But $\phi_3 \to \pm 1$ on the boundary.As a result the value of the topological charge is not quantised.So the static finite energy solutions corresponding to the symmetric minima are nontopological.Unlike \[15\] topological solitons are not obtained in this limit.These observations may be compared with earlier findings about the Chern - Simons solitons \[17\]. The situation changes dramatically when we consider the symmetry breaking minima (7).Here the physical vaccua bear representation of the U(1) symmetry. $$\psi \approx e^{in\theta}$$ where $\psi = \phi_1 + i\phi_2$ and n gives the number of times the infinite circle of the physical space circuitting around the equatorial circle of the internal sphere.The topological solitons of the model are now classified according to this winding number.When the equatorial circle is traversed once the physical space is mapped on a hemisphere of the internal sphere.In general the topological charge (18) will be quantised by $$T ={ n \over 2}$$ allowing half integral values of T. Using the definition of $\phi$ we can write $$D_i {\bf {\phi}}\cdot D_i{\bf{\phi}} = |(\partial_i + iA_i)\psi|^2 +(\partial_i\phi_3)^2$$ From (14) and (21) we observe that for finite energy configurations we require the covariant derivative $(\partial_i + iA_i)\psi$ to vanish at the physical boundary.Using (19) we then get on the boundary $${\bf{A}}={\bf{e_\theta}}{n\over r}$$ It is really interesting to observe that the asymptotic form (22) is sufficient to find the magnetic flux $\Phi$,charge Q and spin S. Thus the magnetic flux $$\Phi = \int B d^2x = \int_{boundary}A_\theta r d\theta = 2\pi n$$ and spin \[6,12\] $$S = -{k\over 2} \int_{boundary} \partial^i[x_iA^2 - A_ix_jA^j]d^2x = -\pi kn^2$$ Using (13) and (23) we then find $$Q = -2\pi kn$$ Equations (23) to (25) show that the soliton solutions corresponding to the symmetry breaking vaccua have charge,flux and angular momentum quantised in each topological sector. We then turn to show that the model satisfies Bogomol’nyi conditions. Rearrenging (14) we can write \[18\] $$E = {1 \over 2} \int d^{2}x[{1 \over 2}(D_i{\bf {\phi}} \pm \epsilon_{ij}{\bf {\phi}}\times D_{j}{\bf {\phi}})^2 + {{k^2} \over {1-\phi_3^2}}(F_{12}\pm {1 \over k^2}\phi_3(1- \phi_3^2))^2] \pm 4\pi T$$ Equation (26) gives the Bogomol’nyi conditions $$\begin{aligned} D_i{\bf {\phi}}\pm \epsilon_{ij} {\bf {\phi}}\times D_j{\bf \phi} = 0\\ F_{12}\pm {1 \over k^2}\phi_3(1-\phi_3^2) =0 \end{aligned}$$ which minimises the energy functional in a particular topological sector, the upper sign corresponds to +ve and the lower sign corresponds to -ve value of the topological charge.The equations can be handled in the usual method\[3,15\] to show that the scale invariance is removed by the artifice of the gauge-field coupling. We will now show the consistency of (27) and (28) using the well-known Ansatz\[14,19\] $$\begin{aligned} \phi_1(r,\theta) = \sin F(r) \cos n\theta\nonumber\\ \phi_2(r,\theta) = \sin F(r) \sin n\theta\nonumber\\ \phi_3(r,\theta) = \cos F(r)\nonumber\\ {\bf A}(r,\theta)= -{\bf e}_\theta {{na(r)} \over r} \end{aligned}$$ From (7) we observe that we require the boundary condition $$F(r) \to \pm {\pi \over 2}\hspace{.2cm} as\hspace{.2cm} r \to \infty$$ and equation (22) dictates that $$a(r) \to -1 \hspace{.2cm}as\hspace{.2cm} r \to \infty$$ Remember that equation (22) was obtained so as the solutions have finite energy. Again ,for the fields to be well defined at the origin we require $$F(r) \to 0 or \pi \hspace{.2cm}and \hspace{.2cm} a(r) \to 0\hspace{.2cm} as \hspace{.2cm}r \to 0$$ Substituting the Ansatz(29) into (27) and (28) we find that $$\begin{aligned} F^\prime (r) = \pm {{n(a+1)}\over r} \sin F\\ a^\prime (r) = \mp {r \over {nk^2}}\sin^2F \cos F \end{aligned}$$ where the upper sign holds for +ve T and the lower sign corresponds to -ve T.Equations (33) and (34) are not exactly integrable.They may be solved numerically subject to the appropriate boundary conditions to get the exact profiles. Using the Ansatz (29) we can explicitly compute the topological charge T by performing the integration in (18).The result is $$T = -{n\over 2}[cosf(\infty)-cosf(0)]$$ So we find that according to (30) and (32) T =$\pm {n\over 2}$ which is in agreement with our observation (20).Note that f(0) corresponds to +ve T and f(0) = $\pi$ corresponds to -ve T. If we take +ve T we find F(r) bounded between 0 and $\pi\over 2$ is consistent with (30),(32) and (33).Again a(r) bounded between 0 and -1 is consistent with (31),(32) and (34).Thus for +ve topological charge the ansatz (29) with the following boundary conditions $$\begin{aligned} F(0) = 0\hspace{.2cm} a(0) = 0\nonumber\\ F(\infty)={\pi \over 2} \hspace{.2cm}a(\infty)= -1 \end{aligned}$$ is consistent with the Bogomol’nyi conditions. Similarly the consistency may be verified for -ve T. To conclude,we find that gauging the U(1) subgroup of the nonlinear O(3) sigma model along with the inclusion of a self-interaction potential with degenerate minima where the U(1) symmetry is spontaneously broken provides topologically stable soliton solutions which have the desirable feature of the removal of scale invariance with quantised energy,charge,flux and angular momentum pertaining to each topological sector in contrast to \[15\] where such solutions are infinitely degenerate.This breaking of the degeneracy is ascribed to the topology of the minima of the potential considered in our model.We have demonstrated that the theory satisfies Bogomol’nyi conditions and discussed the consistency of the solutions.Detailed calculations of the profiles are pending with other related issues.We propose to take up these works subsequently. I like to thank Dr.R.Banerjee for helpful discussions and Dr.S.Roy for his encouragements.I also thank Professor C.K.Majumdar,Director S.N. Bose National centre for Basic Sciences for allowing me to use some of his institute facilities.Finally my earnest thanks are due to the referee for his comments which largely enabled me to put the work in proper perspectives. [99]{} A.A. Belavin and A.M. Polyakov,JETP Lett.22(1975) 245. E.B. Bogomol’nyi,Sov. J. Nucl. Phys.24(1976)449. R.Rajaraman, Solitons and Instantons (North Holland Publishing Company). F.Wilczek and A.Zee, Phys. Rev. Lett.51(1983)2250. M.Bowick,D.Karabali and L.C.R.Wijewardhana, Nucl.Phys. B271(1986)417. R.Banerjee,Phys.Rev.Lett.69(1992)17;Phys.Rev.D48(1993)2905 R.Banerjee,Nucl.Phys.B419\[FS\](1994) 611. J.Hong,Y.Kim and P.Pac,Phys.Rev.Lett.64(1990)2230. R.Jackiw and E.J.Weinberg,Phys.Rev.Lett.64(1990)2234. J.Burzlaff,A.Chakrabarti and D.Tchrakian,Phys.Lett.B293 (1992)127. B.Chakrabarti,Ann. of Phys.(N.Y.)244(1995)312. R.Banerjee and P.Mukherjee,Nucl.Phys.B478(1996)235. R.A.Leese,M.Peyrard and W.J.Zakrzewski, Nonlinearity 3(1990)387. B.J.Schroers,Phys.Lett.B 356(1995)291. P.K.Ghosh and S.K.Ghosh,Phys.Lett.B366(1996)199. J.Schwinger,Phys.Rev.127(1962)324. R.Jackiw,K.Lee and E.J.Weinberg Phys.Rev.D42(1990)3488. Note that a factor 1/2 is missing in eqn(7) of \[15\]. Y.S.Yu and A.Zee,Phys.Lett. B147(1984) 325. [^1]: e-mail:[email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: | In this article we demonstrate how algorithmic probability theory is applied to situations that involve uncertainty. When people are unsure of their model of reality, then the outcome they observe will cause them to update their beliefs. We argue that classical probability cannot be applied in such cases, and that subjective probability must instead be used. In Experiment 1 we show that, when judging the probability of lottery number sequences, people apply subjective rather than classical probability. In Experiment 2 we examine the conjunction fallacy and demonstrate that the materials used by involve model uncertainty. We then provide a formal mathematical proof that, for every uncertain model, there exists a conjunction of outcomes which is more subjectively probable than either of its constituents in isolation. **Keywords:** Conjunction fallacy; algorithmic statistics; likelihood judgments; surprise; subjective probability. author: - | [**Phil Maguire ([email protected])**]{}\ Department of Computer Science\ NUI Maynooth, Ireland\ Department of Computer Science\ NUI Maynooth, Ireland\ School of Business, National College of Ireland\ Mayor Street, IFSC, Dublin 1, Ireland\ School of Computer Science and Informatics\ UCD, Dublin 4, Ireland bibliography: - 'References.bib' title: A Computational Theory of Subjective Probability --- Introduction ============ *Breaking news: Pandemonium erupted today at the National Lottery headquarters as the numbers 1, 2, 3, 4, 5 and 6 were drawn for the third week in a row. Lottery officials, stunned by a sense of déjà vu, scrambled to release a statement insisting that the lottery drum selection mechanism meets the highest standards for randomness. Meanwhile thousands are celebrating after ignoring the opinions of mathematicians who had viewed the two previous draws as a statistical fluke. Commentators in the media are demanding an immediate investigation, describing the incident as a fiasco.* The mathematical concept of probability, originally formulated to describe the highly constrained environment of games of chance, has now found its way into everyday parlance, with people using it to quantify the likelihood of everything from the possibility of economic recession to the risk of global warming. Such has been the unquestioned adoption of the probability concept into mainstream culture that it has become the default assumption that probability theory provides the only logical way for people to think about likelihood. For instance, applied probability theory to real-world situations involving personality decisions, medical judgments, criminal motives and political forecasts. On observing consistent deviations from the mathematical theory, they interpreted their findings as evidence of a serious flaw in human reasoning (see , for a review of the associated debate). In this article we adopt the alternative stance that consistent deviations between human reasoning and a simplified, artificial mathematical theory are far more likely to reflect deficiencies in the theory than they are to reflect sub-optimality in how people think about likelihood. Classical Probability ===================== Probability theory was formalised by Kolmogorov in the 1930s through the notion of probability space, whereby a set of possible outcomes is mapped to a number that represents its likelihood by a probability measure function. For example, a perfect dice outputs the numbers from 1 to 6 with equal frequency. However, in the real world it is rarely feasible to identify the theoretical probability measure function which underlies the events we observe. Because we have to work backwards, using the events to deduce the original function, we can never be sure if the model we are using is correct. For example, according to classical probability theory, no conceivable sequence of numbers produced by rolling a dice will ever lead us to revise our beliefs about the nature of the dice. Even if we rolled 1, 1, 1, 1... the hypothesis of the sequence being a statistical fluke would always remain infinitely more likely than the possibility that the dice is biased. In reality, nobody has beliefs which are strong enough to stand up to the requirements of classical probability theory. We strongly believe that the numbers drawn from the lottery are random, yet there are certain sequences which, as in the introductory lottery example, would cause us to question our assumptions and consider other possibilities. If the sequence 1, 2, 3, 4, 5, 6 was drawn three weeks in succession, it might suggest that the balls were not equally weighted, the drum mechanism was defective, or that one of the lottery officials was playing a practical joke. The point where we start to ask questions reveals how strongly we hold our beliefs. But no matter how confident we are about a particular model of reality, there will always be some sequence of events which will cause us to change our mind. This poses a crucial problem for probability theory. Let’s consider the probability of 1, 2, 3, 4, 5, 6 being drawn in a lottery for three weeks in a row. If the draw is unbiased then this sequence of events is just as likely as any other. In a lottery with 45 numbers, the exact probability is $C(45,6)^3$. But if this sequence of events actually unfolded, it would lead us to believe that the draw mechanism is biased. Given the new updated belief, then the probability of getting 1, 2, 3, 4, 5, 6 is actually far higher. So what is the true probability of this sequence of events? To apply classical probability theory a single model of reality must be selected. We must assume either that the lottery draw is biased or that it isn’t. But doing so would be a mistake because we don’t actually know which world is the case. The situation involves model-outcome dependence, insofar as the outcome affects our beliefs about the system that generated it. Stating that the probability of drawing 1, 2, 3, 4, 5, 6 is $C(45,6)^3$ is misleading because, if this sequence of events actually occurred, we would no longer trust the assumptions involved in computing that probability. Uncertainty in the Real World ============================= The issue here is that classical probability theory only applies to cases involving a definitive probability measure function, while models of reality always involve uncertainty. Though useful for reasoning about games deliberately engineered to generate pseudo-randomness, classical probability has less applicability to everyday life, where reducing uncertainty and optimising models of reality are the principal goals. In our previous work examining the difference between surprise and probability judgments [@Maguire2011] we presented a cognitive theory of uncertainty modeling which views the maintenance of an up-to-date representation of reality as the principal motivation guiding information seeking behaviour. People rely on observational data to continually refine their model of the environment, thus maintaining the optimality of their decision making. In particular, the signal that they rely on to diagnose discrepancies between their model and the real world is randomness deficiency. The best model of a set of observational data is the one which describes it most concisely, so that the description of the data relative to the model is ‘incompressible’ or random (see ; ). In the case that one’s model of reality is optimal, then new sensory data should still be random with respect to it. The experience of randomness deficiency (i.e. a pattern which could be described more concisely using an alternative model) causes alarm bells to go off, because it indicates that one’s model is likely to be suboptimal. This is known as surprise. When surprise occurs there are two potential resolutions. First, more observation data can be gathered, which might mitigate the randomness deficiency by revealing it to be a statistical fluke. If this does not resolve the discrepancy then the remaining alternative is to update one’s model to fit the data. Either way, the resolution process necessitates urgent sampling of information from the environment. During the surprise response, eye widening, opening of the mouth and enlargement of the nasal cavity serve to facilitate the intake of sensory information (see ). Consider for example looking at the floor and seeing some crumbs which spell out the words “YOU ARE BEING WATCHED”. When crumbs fall on the floor it is just as probable that they will arrange themselves into this pattern as any other. If we were certain that the crumbs had fallen randomly then it would not be interesting. However, where knowledge is uncertain then people respond to randomness deficiency. The pattern of crumbs is randomness deficient because there is another model which can explain it more concisely: Somebody might have deliberately arranged the crumbs in this way. The first strategy is to look at the rest of the floor. If the rest of the floor is covered in many crumbs which have no other patterns then the overall randomness deficiency is mitigated. If these are the only crumbs on the floor then finding a satisfactory explanation becomes critical. People are motivated to seek out randomness deficiency in the world [@Dessalles2006]. The experience of randomness deficiency with subsequent resolution through representational updating is what makes subjects interesting, films entertaining and jokes funny [@Schmidhuber2009]. Accordingly, when people speak intuitively about likelihood and probability, it is the concept of representational updating which is relevant to them. Subjective Probability ====================== Because it assumes a definitive probability measure function, classical probability theory cannot be applied to the concept of representational updating. This limitation means that the theory is, for the large part, irrelevant to everyday life and thus inappropriate for evaluating the nature of human reasoning. Developments in algorithmic statistics have allowed probability theory to be extended to situations involving an uncertain probability measure function (see, e.g., ; ). The optimal model which can be derived from a set of observations is the one which maximizes the compression of that dataset, yielding the Minimum Description Length (MDL), a concept which formalizes Occam’s razor. Whenever an observation is no longer typical with respect to an MDL model it should be adjusted to lower the randomness deficiency of the data (see , for details on how the updating process is carried out). We can quantify the extent of this representational adjustment in terms of the amount of information that, given the original model, would be required to obtain the updated model. The more the information required, the more significant (and less likely) the update. The model that people hold of reality represents the very best that they can do in representing their environment and provides the very best that they can achieve in terms of predictions. If we assume that our representation is a reliable predictor of events then the larger a potential update to that representation, the rarer it should be. Accordingly, we can apply probability theory to speak about the likelihood of an outcome requiring an update of a particular size. The uncertainty which precludes probability theory from being applied to real-world scenarios is circumvented by shifting the focus from an underdetermined probability measure function to the immutable mechanism of representational updating. Preliminaries ------------- A computable probability density function $p$ can be interpreted as a model for a string generating device. Given such a device, described by $p$, there are some “type of strings” we expect to be output, whereas some others are surprising. String $x$ is said $p$-typical if it is a random string relative to the model described by $p$, i.e. the model already describes all the regularities in $x$. Formally, let $\alpha >0$ be a constant, called the surprise threshold, which represents the level of randomness deficiency that necessitates representational updating. String $x$ is $p$-typical with surprise threshold $\alpha$ (or $(p,\alpha)$-typical) if the length of its shortest description given $p$ is at least the number of bits a Shannon-Fano code based on $p$ would require (an encoding where the more $p$-likely a string is, the shorter its encoding will be) after subtracting the surprise level $\alpha$, i.e., $$K(x|p^*) \geq -\log p(x) - \alpha.$$ The idea behind the minimal description length (MDL) of a string $x$ [@Gacs2001] is to take the shortest (in description length) among all models for which $x$ is typical. To avoid overfitting (i.e. the model is specifically built for $x$ instead of for all “strings of type $x$”) the description length of both the model and the string given the model, should be equal to the description of the string on its own. Formally, probability density function $p$ is optimal for string $x$ if the shortest description of $x$ has the same length (up to an additive constant) as the shortest description of $p$ plus the number of bits required for a Shannon-Fano encoding of $x$ based on $p$, i.e., $$K(x) = K(p) -\log p(x) \pm O(1)$$ where $O(1)$ means the equality holds up to an additive constant. The MDL of string $x$ is the shortest (description length) among all optimal probability density functions for $x$ for which $x$ is typical. Subjective information and probability -------------------------------------- Suppose an observer experiences observations $d_1,d_2,\ldots$ generated by some source with computable probability density $p_{\mathrm{source}}$. The observer tries to learn the probability density $p_{\mathrm{source}}$ by finding the shortest optimal model based on the observations made so far. Formally, after having observed strings $d_1,d_2,\ldots,d_n$, the observer seeks to construct a hypothetical model $p_n$ where $$\begin{aligned} p_n &= \arg\min\{|p^*|: p \text{ is optimal for } d_1,d_2,\ldots,d_n \text{ and }\\ &d_1,d_2,\ldots,d_n \text{ are } (p,\alpha)\text{-typical}\}.\end{aligned}$$ If the next observation $d_{n+1}$ is surprising, action may be required. Formally, observation $d_{n+1}$ is $\alpha$-surprising if the length of its shortest description given $p$ is less than the number of bits a Shannon-Fano code based on $p$ would require after subtracting the surprise level $\alpha$, i.e., $$K(d_{n+1}|p_n^*)<-\log p_n(d_{n+1}) - \alpha.$$ If an update is performed, then the subjective information of $d_{n+1}$ (the “cost” of the update) is the amount of information needed to update the model to the latest, that is the length of the shortest description of the new model, given the old model, i.e., $$\mathrm{subjective \ information}(d_{n+1}) = K(p_{n+1}^*|p_n^*).$$ Subjective probability (the probability of the update) can then be quantified based on the amount of information it contains, i.e., $$\mathrm{subjective\ probability}(d_{n+1}) =2^{-K(p_{n+1}^*|p_n^*)}.$$ Experiment 1 ============ In the following experiment we investigated the hypothesis that people use subjective probability rather than classical probability to judge the likelihood for real-world events. We used an example for which the use of classical probability theory seems particularly compelling, namely lottery sequences (see ). A naive application of classical probability suggests that all lottery sequences are just as likely. Method ------ In a lottery system where 6 numbers are drawn from 45, each ordered sequence has a classical probability of $C(45,6)$. According to the theory outlined in the previous section, the subjective probability of an outcome is related to its randomness deficiency. People expect the lottery numbers to be Kolmogorov-random. The more they deviate from a typical random string, the lower the subjective probability that they reflect the output of a random source. The randomness deficiency of a string is quantified precisely by its MDL. However, since this theoretical construct is not computable in practice, we are obliged to create a heuristic compressor which approximates it. We considered the patterns to which people are sensitive in discriminating predictable sequences from random ones. Overtly non-typical random patterns include ones in which the numbers are consecutive (e.g. 3, 4, 5, 6, 7, 8) or where they increase in a constant step size. To compress these patterns we created a simple compressor which takes in an ordered sequence of six numbers, and computes the six step sizes between them (with the first number counting as the first step). A Huffman encoding scheme is then applied, which relates bit size to step size. A breakdown of the structure of the associated Huffman tree is provided in Table \[huffman\]. Using this system the sequence 10, 32, 33, 35, 39, 45 is transformed to step sizes of +10, +22, +1, +2, +4, +6 which is then encoded using 8 + 8 + 2 + 3 + 4 + 6 = 31 bits. Analysing six years of bi-weekly Irish National Lottery draws revealed a mean compressed length of 30.9 bits, with a mode of 31 bits. The most randomness deficient of the 624 sequences was 2, 4, 32, 34, 36, 37 (description length of 20 bits), while the most random was 9, 20, 26, 27, 34, 45 (description length of 39 bits). The theoretical minimum description length of our system was 12 (e.g. 1, 2, 3, 4, 5, 6), while the theoretical maximum was 43 (e.g. 7, 13, 20, 29, 36, 45). The number of bits needed to perfectly encode an ordered random sequence of six numbers between 1 and 45 is 23.0 bits. Although our compressor cannot compute MDL, it delivers compression for randomness deficient outputs (i.e. it compresses below 23.0 bits for certain non-typical random sequences) and can therefore be used to evaluate the hypothesis that people use subjective rather than classical probability. 0.12in Level Depth Leaves \#Branches ------------- ------------------ ------------ 1 - 2 2 +1, repeat 2 3 +2, +3 2 4 +4 3 5 +5 5 6 +6 9 7 +7, +8 16 8 +9 up to +40     - : Structure of Huffman encoding scheme.[]{data-label="huffman"} ### Participants 130 undergraduate students from NUI Maynooth participated voluntarily in this study. ### Procedure As an initial step we purchased two quickpick (i.e. randomly selected) lottery tickets for the next week’s Irish National Lottery, with six ordered numbers ranging from 1 to 45. Participants were informed that we had purchased these tickets and that, for each of the two quickpick sequences, their goal was to identify it from among a group of five candidate sequences. No mention was made of how the other four sequences had been generated. Each quickpick sequence was presented on a screen along with four other sequences randomly generated using our compressor algorithm. The four distractor sequences met the constraints of having compressed bit-sizes of between 15 and 18 bits, 19 and 22 bits, 23 and 26 bits, and 27 and 29 bits respectively. As it happened, the first lottery ticket sequence had a compressed description length of 31 bits, and the second had a length of 30 bits. The ordering of the five sequences on the screen was randomized. Participants ranked each set of five sequences in order of likelihood of being the quickpick sequence, from highest probability to lowest probability. After the process was complete participants were shown the actual lottery tickets so that, as promised, they could see if they had made the correct judgment or not. Unfortunately for the experimenters, the lottery tickets did not turn out to be winning ones. Results and Discussion ---------------------- An individual applying classical probability would view all sequences as equally likely and would thus only have a 20% chance of correctly identifying one quickpick sequence mixed with four others. However, 64% of participants correctly identified the numbers on the first ticket, and 66% on the second ticket (i.e. ranked these sequences in first place). When participants were shown the lottery tickets at the end of the experiment they were surprised that their intuition had, in the majority of cases, led them to make the correct choice. Figure \[experiment1results\] shows the mean compressed bit size for sequences ranked from first to fifth place across the two presentations. The overall correlation between ranking and compressed description length was 0.965, $p < .001$. ![Mean compressed bit size according to rankings of likelihood.[]{data-label="experiment1results"}](experiment1results2.pdf){width="8cm"} These results demonstrate that, not only do people use subjective probability, they also enhance the accuracy of their judgments by using it. While the naive mathematician assumes all lottery sequences are equally likely, the savvy layperson realises there is an element of uncertainty involved in how those sequences were generated. The greater the randomness deficiency of a sequence, the greater the subjective probability that it was generated by a non-random generative mechanism. Our central argument in this article is that, because models of reality always involve uncertainty, people apply subjective probability rather than classical probability in everyday life. In the following experiment we investigated whether the application of subjective probability can explain experimental observations which have previously been interpreted as examples of fallacious reasoning. Experiment 2 ============ The conjunction effect is a situation in which people assert that a conjunction of two outcomes is more probable than either of those outcomes in isolation. According to classical probability theory this is a fallacy because requiring two outcomes to be validated is always a stricter criterion than requiring a single one to be validated (i.e. $P(x\wedge y) \leq P(y)$). The most celebrated example of the fallacy involves one of the materials used by , involving an individual named Linda. *Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations*. Which is more probable? a\) Linda is a bank teller b\) Linda is a bank teller and is active in the feminist movement. report that, when the two possible outcomes are listed together as above, 85% of people violate the conjunction rule by identifying b) as more probable. @TverskyKahneman1983’s explanation is that people get confused by what they call ‘representativeness’. They found that participants’ responses reflect the extent to which the descriptions match a stereotype, with a correlation of 0.98 between mean ranks of probability and representativeness. It is interesting to note that this correlation closely matches the observed correlation of 0.97 between mean ranks of probability and compressed description length in Experiment 1. This suggests the possibility that representativeness and randomness deficiency are closely related concepts. In Experiment 1 we found that, when there is uncertainty as to the generative mechanism which produced an outcome, people rely on randomness deficiency to make judgments. The uncertainty in Experiment 1 concerned the fact that participants were given no information as to how four of the five lottery sequences were generated. Rather than assuming that all the sequences were generated randomly, they correctly used randomness deficiency to make inferences that resolved the uncertainty. In the Linda example, some information about Linda is provided, but there is much about her that remains unknown (e.g. has she settled down since her student days?) In the case of uncertainty regarding the underlying probability measure function, then classical probability cannot be applied. For example, if we find out that Linda is a bank teller, then we might infer that she has settled down. In contrast, hearing that she is still active in the feminist movement suggests that she has not changed much since her student days. Because these two models of Linda are quite different, there is no definitive probability measure function relative to which classical probability can be expressed. Method ------ In the following experiment we investigated whether the outcomes for the Linda scenario cause participants to adjust their model of Linda. ### Materials For this experiment we altered the Linda scenario by including the outcomes as part of the description. We removed the information that she is single, outspoken and very bright and included at the end of the description either that “Linda is a bank teller” (Version 1) or “Linda is a bank teller and is active in the feminist movement” (Version 2). Participants were then asked to rate the probability of Linda having the attributes of being single, outspoken and very bright (from 0 to 100%). In order for classical probability to be applicable, then the probabilities provided for Versions 1 and 2 should not differ significantly. Linda should be just as independent, outspoken and bright regardless of whether she is active in the feminist movement or not. ### Participants 106 undergraduate students from NUI Maynooth participated voluntarily in this study. ### Procedure Participants were randomly assigned either Version 1 or Version 2 of Linda’s description and wrote down their probabilities for the three characteristics, which were randomly ordered along with three other filler characteristics (Linda plays golf, Linda is dyslexic, Linda suffers from anxiety). Results and Discussion ---------------------- The mean probabilities for the three characteristics are shown in Table \[lindaresults\]. When Linda was described as a bank teller and active in the feminist movement she was rated as significantly more likely to be single, demonstrating that the outcomes used in the Linda scenario cause participants to adjust their model of Linda. 0.12in Ver. 1 Ver. 2 t-test ------------- -------- -------- -------------------------- Single 47% 64% $t(104) = 4.11, p <.001$ Outspoken 77% 80% $p > .05$ Very Bright 59% 63% $p > .05$ : Mean probability ratings, t-test scores and significance for the two descriptions of Linda.[]{data-label="lindaresults"} The numbers generated by a perfect dice never lead us to update our beliefs about the nature of the dice, yet finding out about Linda’s current activities does lead people to update their beliefs about her. Because the model of Linda is uncertain, subjective probability must be applied. What people are quantifying when they identify the conjunction as more probable is that the conjunction contains more subjective information, and that, relative to the process of representational updating, the likelihood of an outcome diminishes with the amount of subjective information it carries. Basing decisions on subjective probability is mathematically the correct approach when dealing with uncertainty regarding the underlying probability measure function. In the following section we build on this result by proving that for every situation involving uncertainty (i.e. all real world scenarios) there is a conjunction of events which is more subjectively probable than either of its constituents in isolation. Proof that Conjunction Effect is not a Fallacy ============================================== In this section we prove that given any hypothetical model $p$, there are always two strings of events $x,y$ such that $x$ is a substring of $y$ but $y$ has higher subjective probability. The idea of the proof is that any long enough typical string of events can always be decomposed into a substring of events that carries greater subjective information. Let $E_1,E_2,\ldots E_m$ be $m$ independent events and let $p$ be the associated computable probability measure function. Let $\alpha>0$ be a surprise threshold. There exists a conjunction of events $A=A_1 \wedge A_2 \wedge\ldots\wedge A_n$ with a constituent $B$ (i.e. $p(A)<p(B)$) such that $B$ is $(p,\alpha)$-surprising (i.e. carries subjective information) and $A$ is $(p,\alpha)$-typical (i.e. has a subjective probability of 1). Let $E_1,E_2,\ldots E_m$, $p$ and $\alpha>0$ be as above. Without loss of generality $m=2^k$ and $p$ can be seen as a probability on strings of length $k$ (each coding one event $E_i$) extended multiplicatively i.e., $p:2^{k}\rightarrow [0,1]$ is extended multiplicatively by $p(xy):=p(x)p(y)$. Let $n$ be a large integer. Let $y\in 2^{kn}$ be a $(p,\alpha)$-typical string. $y$ can be viewed as the concatenation of $n$ strings of length $k$ (i.e. the conjunction of $n$ events). By the pigeon hole principle, there must be such a string that occurs at least $n/2^k$ times. Denote this string by $s$, and let $l$ be the number of occurences of $s$ in $y$, i.e. $l\geq n/2^k$. Because $y$ is $(p,\alpha)$-typical we have $p(s)>0$. Thus $p(s)=2^{-c}$ for some $c>0$. Let $x$ be $l$ concatenations of $s$. Because $p$ is extended multiplicatively we have $p(x)>p(y)$. Let us show that $x$ is $(p,\alpha)$-surprising. To describe $x$ it suffices to describe $l$ plus a few extra bits that say “print $s$ $l$ times”. Since $l$ can be described in less than $2\log l$ bits (by a prefix free program) we have $K(x) < 3 \log l$ for $n$ large enough. We have $$\begin{aligned} -\log p(x) - \alpha &= -\log p(s^l) - \alpha = -\log p(s)^l -\alpha\\ &= -l\log 2^{-c} - \alpha = cl -\alpha > 3 \log l > K(x) \\ &\geq K(x|p^*)\end{aligned}$$ for $n$ large enough. Thus $x$ is $(p,\alpha)$-surprising, but $y$ is not. Conclusion ========== Although identified an association between representativeness and the conjunction effect, they never provided an explanation for why such an association might exist, instead being satisfied to pass it off as an arbitrary reasoning fallacy. Had they questioned participants regarding their judgments, rather than dismissing them as fallacious, then the resultant findings may have facilitated the extension of classical probability theory. In sum, perhaps the most salient fallacy on display in @TverskyKahneman1983’s [-@TverskyKahneman1983] study is the misplaced belief that mathematical theories which have been developed for precision models in the exact sciences retain their validity when used to describe complex cognition in the real world. posed the following question: “Why do intelligent and reasonably well-educated people fail to recognize the applicability of the conjunction rule in transparent problems?” Here, we have presented the answer: Because often it’s not applicable.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We introduce a new test of how well language models capture meaning in children’s books. Unlike standard language modelling benchmarks, it distinguishes the task of predicting syntactic function words from that of predicting lower-frequency words, which carry greater semantic content. We compare a range of state-of-the-art models, each with a different way of encoding what has been previously read. We show that models which store explicit representations of long-term contexts outperform state-of-the-art neural language models at predicting semantic content words, although this advantage is not observed for syntactic function words. Interestingly, we find that the amount of text encoded in a single memory representation is highly influential to the performance: there is a sweet-spot, not too big and not too small, between single words and full sentences that allows the most meaningful information in a text to be effectively retained and recalled. Further, the attention over such window-based memories can be trained effectively through self-supervision. We then assess the generality of this principle by applying it to the CNN QA benchmark, which involves identifying named entities in paraphrased summaries of news articles, and achieve state-of-the-art performance.' author: - | Felix Hill[^1],  Antoine Bordes, Sumit Chopra & Jason Weston\ Facebook AI Research\ 770 Broadway\ New York, USA\ `[email protected],{abordes,spchopra,jase}@fb.com` bibliography: - 'iclr2016\_conference.bib' title: 'The Goldilocks Principle: Reading Children’s Books with Explicit Memory Representations' --- Introduction ============ The Children’s Book Test ======================== Studying Memory Representation with Memory Networks =================================================== Baseline and Comparison Models ============================== Results ======= Conclusion ========== ### Acknowledgments {#acknowledgments .unnumbered} The authors would like to thank Harsha Pentapelli and Manohar Paluri for helping to collect the human annotations and Gabriel Synnaeve for processing the QA CNN data. [^1]: The majority of this work was done while FH was at Facebook AI Research, and was completed at his current affiliation, University of Cambridge, Computer Laboratory, Cambridge, UK.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The [*leader trajectory*]{} function defined in this article is an approximate solution of a differential equation. It is defined by some independent one-dimensional differential equations. The generalized main result of this article asserts that if the [*leader trajectory*]{} exists then it is at finite distance from the solution of the system. However, the differential equation can be controlled by a reduced one-dimensional differential equations. The application of the generalized main result is to control the trajectory of the periodic systems. We prove that for any [*periodic*]{} system and any initial condition there exists a [*leader trajectory*]{} which is a linear function of the time variable. In other words, we find an exact [*Rotation vector formula*]{} which is the relation between the [*rotation vector*]{} and the initial condition. In addition, we find easily a necessary and sufficient condition for the existence of a locally constant rotation vector under perturbation of the system, known by the [*Arnold tongue*]{}. The main result answer to many problems in natural sciences.' author: - | W. Oukil$^{1}$, Ph. Thieullen$^{2}$ and A. Kessi$^{3}$\ \ \ \ \ \ nocite: '[@*]' title: Reduced dimension and Rotation vector formula of ordinary differential equation --- Differential equations, bounded solution, periodic system, rotation vector. Introduction and main results ============================= One important application of the result is to control the trajectory of the periodic systems as cited in the abstract; The applications of the periodic differential equations are various. The issue is to estimate the value of the [*rotation vector*]{}. The list of the related works is long; for some biological works see  [@bonilla; @PhysRevE.86.016211] [@choi; @coray] [@Cumin2007; @Ermentrouttheeffects] [@opac-b1102236; @Quinn] [@WinfreeModel; @Gerstner]. The main result answer to many problems because we find an explicit relation between the rotation vector and the initial condition of the orbit. We study in this article the following system $$\label{systemNotNormalized} \dot{x}= f(x),\quad t\in\mathbb{R},\quad x=(x_i)_{i=1}^n,\quad x(0) =x_0,\quad x_0\in\mathbb{R}^n.$$ where $f: \mathbb{R}^n\to \mathbb{R}^n$ is a $C^2$ function. The function $ x(t)=(x_i(t))_{i=1}^n$ is the state of the system. Notice that any non-autonomous system can be written as an autonomous system. We introduce in the present Section some notations, the definition of the leader trajectory and the main results. The Section \[Sec:RDE\] is the ingredients to study the linearized of system which has the form of the Riccati differential equation. In the rest of Sections we prove the main results Periodic systems {#secperiodicandhypo} ---------------- Many mathematical models are defined by the periodic systems  [@Cheban; @Corduneanu; @Fink1974]. For example, in biological sciences the components of the \[rotationvector\] are called the [*frequency*]{}  [@Cumin2007; @Ermentrouttheeffects; @OukilKessiThieullen]. When the periodic system is stable, in the sens of the [*Lyapunov stability*]{}, the \[rotationvector\] exists  [@Cheban; @Corduneanu; @Fink1974; @Saito1971]. Let be $q,p\in\mathbb{N}^*$ and let $g=(g_i)_1^p : \mathbb{R}^q\to\mathbb{R}^p$ be a function. We say that $g$ is [*periodic*]{} if it is periodic relative to each variable. In other words, $$g_i\Big(y_1,\ldots,y_j+1,\ldots,y_q\Big)=g_i(y),\ \forall i=\overline{1,q},\forall j=\overline{1,q},\ \forall y=(y_i)_{i=1}^q\in\mathbb{R}^q.$$ In this article, we use the usual norm on $\mathbb{R}^q$, $q\in\mathbb{N}^*$, defined by $\norm{z}=\max_{1\le i\le q}|z_i|$ for all $z:=(z_i)_{i=1}^q\in\mathbb{R}^q$. We also denote $\norm{.}$ the associated matrix norm. Let be $q\in\mathbb{N}^*$ and $\lambda\in\mathbb{R}^q$. Let $g:\mathbb{R}\to\mathbb{R}^{q}$ be a continuous function. We say that $\lambda$ is the [*[rotation vector]{}*]{} of the function ${g}$ if $$\sup_{t\in\mathbb{R}}\norm{g(t)-\lambda t}<+\infty.$$ Notice that if the rotation vector exists then it is unique. To formulate the main results let us introduce the following notations; Let $p,q\in\mathbb{N}^*$, for every $q\times p$ matrix $M=[m_{i,j}]_{i,j}$ we denote $$\label{SigmaDefinition} \sigma(M): =\sum_{1\le i\le q}\sum_{1\le j\le p}m_{i,j}.$$ Let be $u\in\mathbb{R}^q$, we denote $\operatorname{Diag}(u)$ the diagonal matrix of diagonal $u$. Let be $ 1\le i \le q$, define the diagonal matrix $I_{i,q}$ as $$\label{Def:desIiq} I_{i,q}=\operatorname{Diag}(v),\ v=(v_j)_{j=1}^q\ : \ v_i=-1 \ \text{and}\ v_j = 1,\ \forall 1\le j\neq i\le q.$$ Denote by $I_q$ the $q\times q$ identity matrix and put $I_{0,q}=I_q$. The particularity of the matrix $I_{i,q}$ is the relation $2y_i=\sigma(I_{0,q}y)-\sigma(I_{i,q}y)$ for any vector $(y_i)_{i=1}^q$. Let $g: \mathbb{R}^q\to \mathbb{R}^q$ be a $C^2$ function we denote $dg$ and $d^2g$ the first and the second differential of $g$ respectively. Let be $ x_0\in \mathbb{R}^{n}$ and let $x$ be the solution of the Equation . Suppose that $f$ is periodic, then $x$ admits a \[rotationvector\] $\rho\in\mathbb{R}^n$ which is the unique solution of the following [*Rotation vector formula*]{} $$\label{RotaionVFMR} \lim_{t\to+\infty}\frac{1}{t}\Big[\Psi_0[f_{x_0}](\rho, t)-\Psi_i[f_{x_0}](\rho, t)\Big]=0,\quad\forall i=\overline{1,n},$$ where the functions $\Big(\Psi_i[f_{x_0}]\Big)_{i=0}^n$ are defined by $$\begin{gathered} \Psi_i[{f}_{x_0} ](\rho,t):= \int_{0}^t A_i(\rho,s)\exp\Big( \frac{1}{n}\int_{s}^t\sigma\Big(I_{i,n} d f( \nu {\rho} +x_0 ) I_{i,n} \Big) d\nu\Big)ds,\\ A_i(\rho,t):=\sigma\Big(I_{i,n} (\rho - f(t {\rho} +x_0))\Big),\quad \forall t \in\mathbb{R},\quad\forall i=\overline{0,n}.\end{gathered}$$ Generalized Main result ----------------------- Define the leader trajectory function. \[Def:LeaderTrajec\]\[Leader trajectory\] Let be $q\in\mathbb{N}^*$, let $g:\mathbb{R}^q\to\mathbb{R}^q$ and $\mu :\mathbb{R}\to\mathbb{R}^q$ be a $C^1$ functions. We say that $\mu$ is a [*leader trajectory of the function $g$*]{} if the following hypotheses are satisfied - $\sup\Big\{\Big|\int_{0}^t\sigma\Big( dg( \mu( \nu) ) \Big) d\nu\Big|,\quad t\in\mathbb{R}\Big\}< +\infty$. - $\forall i=\overline{1,q}$, $\exists \tau_i\in\{-1,1\}$: $\sup_{t\geq 0} \int_{0}^t\tau_i \sigma\Big(I_{i,q} dg( \mu(\tau_i \nu) ) I_{i,q} \Big) d\nu < +\infty$. - The following one-dimensional differential equations $$\begin{gathered} \dot{\psi}_i=\sigma\Big(I_{i,q} \dot{\mu}( t)- I_{i,q}g( \mu(t))\Big) + \frac{1}{q} \sigma\Big(I_{i,q} dg(\mu( t)) I_{i,q} \Big){\psi}_i,\\ \quad {\psi}_i(0)=0,\quad i=\overline{0,q},\quad t\in \mathbb{R}.\end{gathered}$$ admit a solution ${\psi}_i:\mathbb{R}\to\mathbb{R}$ of initial condition ${\psi}_i(0)=0$ such that $$\begin{gathered} \sup_{t\in\mathbb{R}}|{\psi}_{0}( \tau_i t)-{\psi}_{i}( \tau_i t)|<+\infty,\quad\forall i=\overline{1,q}.\end{gathered}$$ Let be $ x_0\in \mathbb{R}^{n}$ and let $x$ be the solution of the Equation . Suppose that $$\norm{d f}_\infty+\norm{d^2 f}_\infty<+\infty.$$ Let $\mu :\mathbb{R}\to\mathbb{R}^n$ be a leader trajectory of $f$ such that $\mu(0)=x_0$. Then there exists $D>0$ such that $$\label{estimationMainResult1} \norm{\mu(t)-{x}(t)}\le D,\quad\forall t\in\mathbb{R}.$$ In order to prove the Generalized mains result, we use the following linearized system of : Let $\mu :\mathbb{R}\to\mathbb{R}^n$ be a $C^1$ function such that $\mu(0)=x_0$. Consider the change of variables $$\eta(t)= \mu(t)-\Phi^{ t}(x_0) ,\ t\in\mathbb{R},\ \eta:=(\eta_{i})_{i=1}^n,$$ where $\Phi^t$ is the flow of the system . Use Taylor formula to obtain the following linearized system $$\begin{gathered} \label{LinearizedPincipalSystem}\dot{ {\eta}} = \dot{\mu}(t)-f(\mu(t))+ df( \mu(t)) {\eta} + {\eta}^T \zeta_\mu( t ) {\eta},\quad\norm{\eta(0)}=0, \quad t\in\mathbb{R},\end{gathered}$$ where $\zeta_\mu : \mathbb{R} \times\mathbb{R}^{n\times n\times n}$ is a continuous function such that $\norm{\zeta_\mu}_\infty<\norm{d^2f}_\infty$ and where ${\eta}^T$ is the transpose of the function ${\eta}$. Equation take the form of Riccati differential equation. The aim is to prove that the solution ${\eta}:\mathbb{R}\to\mathbb{R}^n$ is uniformly bounded on $\mathbb{R}$. Riccati differential equation {#Sec:RDE} ============================= We denote $v^T$ the transpose of a vector $v$ and for any $q\in\mathbb{N}^*$ we denote $$\mathbb{1}_q:=(1,\ldots,1)^T\in\mathbb{R}^q\quad\text{and}\quad\mathbb{0}_q:=(0,\ldots,0)^T\in\mathbb{R}^q.$$ In this Section, we study a generalized Riccati differential equation. Let $ \mathpzc{A} :\mathbb{R}\to\mathbb{R}^q$, $\mathpzc{B} :\mathbb{R}\to\mathbb{R}^{q\times q}$ and $\mathpzc{H} :\mathbb{R}\to\mathbb{R}^{q \times q\times q}$ be a continuous functions such that $$\label{normegamma} \exists \gamma<1:\ \max\Big\{\norm{\mathpzc{H} }_\infty,\norm{\mathpzc{B}}_\infty\Big\}\le\gamma.$$ Consider the following Riccati differential equation $$\label{equ:systemRiccatiAA}\dot{ y} = \mathpzc{A}(t) +{\mathpzc{B}}(t)y +y^T \mathpzc{H}( t) y ,\ y(0)=\mathbb{0}_q,\ y=(y_i)_{i=1}^q,\ t\in\mathbb{R}.$$ Recall that the notation $\sigma(.)$ and the diagonal matrix $I_{i,q}$ are defined by Equations and respectively. Consider the following hypotheses\ $$\sup\Big\{\Big|\int_{0}^t\sigma({\mathpzc{B}}( \nu)) d\nu\Big| ,\quad t\in\mathbb{R}\Big\}< +\infty,$$ and for every $i=\overline{1,q}$ there exists $\tau_i \in\{-1,1\}$ such that $$\sup\Big\{\tau_i \int_{0}^t\sigma(I_{i,q}{\mathpzc{B}}( \tau_i \nu)I_{i,q}) d\nu,\quad t\geq 0\Big\}< +\infty,$$ The following one-dimensional differential equations $$\begin{gathered} \label{HypoPsiTheoRR}\dot{\psi}_i= \sigma(I_{i,q}\mathpzc{A}(t)) + \frac{1}{q} \sigma(I_{i,q}{\mathpzc{B}}(t)I_{i,q}) {\psi}_i,\quad t\geq0,\end{gathered}$$ admit a solution ${\psi}_i:\mathbb{R}\to\mathbb{R}$ of initial condition ${\psi}_i(0)=0$ such that $$\begin{gathered} \sup_{t\in\mathbb{R}}|{\psi}_{0}(\tau_i t)- {\psi}_i(\tau_i t)|<+\infty,\quad\forall i=\overline{1,q}.\end{gathered}$$ The following theorem is the main ingredient used to prove the main results. \[Prop:RiccatirotationvectorconditionTheo\] Let $y$ be the solution of the differential Equation and let be $\gamma\in(0,1)$ as defined by Equation . Suppose that Hypotheses and are satisfied. Then there exists $\gamma_*>0$ such that for all $\gamma\in(0,\gamma_*)$ there exists $D_\gamma>0$ such that the solution $y$ of Equation satisfies $\sup_{t\in\mathbb{R}}\norm{y(t)}< D_\gamma$. The strategy it to use an appropriate change of variable in and use the differential equation comparison. In order to simplify, we introduce the following notation. Define the cone $V_+^q$ of $\mathbb{R}^q$ by $$\label{ConeDefini} V_+^q :=\{z=(z_i)_{i=1}^q\in\mathbb{R}^q:\ z_i>0,\quad\forall i=\overline{1,q}\}.$$ We consider the partial order $\succeq$ on $\mathbb{R}^q$ defined by $$\forall z\in\mathbb{R}^q:\quad z\succ\mathbb{0}_q \iff z\in V_+^q\quad\text{and}\quad z\succeq\mathbb{0}_q \iff z\in {\overline{V_+^q}}.$$ Define the sets $\mathcal{K}_q^+$ and $\mathcal{K}_{q}$ of the $q\times q$ diagonal matrix by $$\label{defi:setK}\mathcal{K}_q^+ =\{I_{i,q},\ i=\overline{0,q}\}\quad \text{and}\quad \mathcal{K}_{q} =\{-I:\ \ I\in\mathcal{K}_q^+\}\cup\mathcal{K}_q^+.$$ We have the following lemma \[estimationzbyoverunder\] Let $y$ be the solution of the differential Equation . Let be $\gamma>0$ given by Equation . Let be $T>0$ and put $U=[-T,T]$. For every $I\in\mathcal{K}_{q}$ and $\tau\in\{-1,1\}$ denote $$\label{DefDeltDetTild} {\delta}_I(t):= {{\psi}}_I(\tau t)-\sigma(I{y}( \tau t))$$ where ${{\psi}}_I:\mathbb{R}\to\mathbb{R}$ satisfies $$\label{HypoPsiTheoRRLemma}\dot{\psi}_I= \sigma(I\mathpzc{A}( t)) +\frac{1}{q} \sigma(I{\mathpzc{B}}( t)I) {\psi}_I,\ t\in\mathbb{R}.$$ Let be $D>0$ and suppose that $\sup_{t\in U}\norm{ y(t)}<D/\gamma$, then $$\begin{aligned} \forall t\in U,\ \forall I\in\mathcal{K}_{q}:\ {\delta}_I(t) =\frac{D^2}{\gamma(1-\gamma)}\implies \frac{d}{dt}[\exp(qt){{\delta}}_I(t)]>0.\end{aligned}$$ Denote $\tilde{I}_q:=[c_{i,j}]_{1\le i,j\le q}$ be the $q\times q$ matrix such that $c_{i,j}=1$ for all $ i,j=\overline{1,q}$. Let be $I\in\mathcal{K}_{q}$ and $\tau\in\{-1,1\}$. Consider the change of variable $$\begin{gathered} {z}_I( \tau,t):= \exp(t\tilde{I}) I{y}( \tau t),\quad t\in U.\end{gathered}$$ Put $\tilde{y}_I(t):= I{y}( \tau t)$, since the inverse matrix of $I$ is itself the system becomes $$\begin{gathered} \notag \dot{\tilde{y}}_I= \tau I\mathpzc{A}(\tau t) +\tau I{\mathpzc{B}}(\tau t) I \tilde{y}_I(t)+ \tau I y^T(\tau t) {\mathpzc{H}}( \tau t)y(\tau t),\quad \forall t\in U.\end{gathered}$$ That follows $$\begin{aligned} \label{EquaKer} \dot{ {z}}_I(\tau,t)= \exp(t\tilde{I}) \Big[ \tau I\mathpzc{A}(\tau t)+[\tilde{I}+ \tau I{\mathpzc{B}}(\tau t) I ] \tilde{y}_I+\tau y^T(\tau t) {\mathpzc{H}}( \tau t)y(\tau t)\Big].\end{aligned}$$ Remark that $\tilde{I}_q^2 = q \tilde{I}_q$. We have the relation $$\begin{gathered} \label{expoitilde}\exp(t\tilde{I}_q)=I_q+\frac{1}{q}(\exp(qt)-1)\tilde{I}_q\ \text{and}\ \norm{\exp(t\tilde{I}_q)}\le \exp(qt), \ \forall t\geq0.\end{gathered}$$ By hypothesis $\sup_{t\in U}\norm{\gamma y(t)}<D$ and $\norm{\mathpzc{H}}_\infty\le{\gamma}$ then $$\norm{ \gamma \exp(t\tilde{I}) I y^T(\tau t) {\mathpzc{H}}(\tau t)y(\tau)}<\exp(qt) D^2,\quad\forall t\in U.$$ From , $$\begin{aligned} \dot{ {z}}_I(\tau,t)\prec \exp(t\tilde{I}) \Big[ \tau I\mathpzc{A}(\tau t)+[\tilde{I}+ \tau I{\mathpzc{B}}(t) I ] \tilde{y}_I\Big]+\frac{1}{\gamma}\exp(qt) D^2\mathbb{1}_q.\end{aligned}$$ Suppose that $\exp(-qt){\delta}_I(t) ={D^2}/(\gamma({1-\gamma}))$ then $\exp(-qt){\delta}_I(t)>0$. Use the fact that for every $c=(c_i)_{i=1}^q\in\mathbb{R}^q$ and $v=(v_i)_{i=1}^q\in\mathbb{R}^q$ such that $\sigma(v)\neq0$ there exists $\tilde{c}\in\mathbb{R}$ that satisfies $|\tilde{c}|\le \norm{c}$ and such that $$\sum_{j=1}^q c_j v_j=\tilde{c} \sigma(v).$$ The above vector $v$ play the following role $$v_i= \frac{1}{q}{\psi}_I(\tau t)- \tilde{y}_i(t),\quad i=\overline{1,q}.$$ Use the hypothesis hypothesis $\norm{\mathpzc{B}(t)}< \gamma$, to obtain $$\begin{gathered} - \Big[\tilde{I} +\tau I\mathpzc{B}(t) I\Big] (\frac{1}{q} {\psi}_I(\tau t)\mathbb{1}_q- \tilde{y} (t) ) \prec- (1-\gamma) {\delta}_I(t) \mathbb{1}_q,\quad\forall t\in U.\end{gathered}$$ Since $\exp(t\tilde{I}) V_q^+\subset V_q^+$, Equation implies that $$\begin{aligned} \dot{ {z}}_I(\tau,t)\ &\prec \tau \exp(t\tilde{I}) I\mathpzc{A}(\tau t) + \frac{1}{q}\exp(t\tilde{I})[ \tilde{I}+ \tau I{\mathpzc{B}}(\tau t ) I ] {\psi}_I(\tau t) \mathbb{1}_q\\ \notag&-(1-\gamma) \exp(qt){\delta}(t)\mathbb{1}_q+\frac{1}{\gamma}\exp(qt) D^2\mathbb{1}_q,\ \forall t\in U.\end{aligned}$$ By Equation , we have $$\sigma(I\mathpzc{B}(\tau t)I )=\exp(-qt)\sigma(\exp(t\tilde{I})I\mathpzc{B}(\tau t)I\mathbb{1}_q)\ \text{and}\ \exp(t\tilde{I}) \tilde{I}=\exp(qt)\mathbb{1}_q.$$ By addition, we get $$\begin{aligned} \frac{d}{dt}[ \exp(qt) {{\psi}}_I(\tau t) -\sigma( {z}_I(\tau,t) ]&<- q(1-\gamma) \exp(qt) {\delta}(t) + q\frac{1}{\gamma}\exp(qt) D^2,\end{aligned}$$ We have proved ${\delta}_I(t) ={D^2}/(\gamma({1-\gamma}))$ implies that $\frac{d}{dt}[\exp(qt){{\delta}}_I(t)] >0$. Recall that the diagonal matrix $I_{i,q}$ define the sets $\mathcal{K}_q^+$ and $\mathcal{K}_{q}$ in Equation and recall that $I_q$ is the $q\times q$ identity matrix. \[LemmeMAJOD\] Let $v:\mathbb{R}\to\mathbb{R}^q$ and ${\psi} :\mathcal{K}_{q}\times\{-1,1\}\times\mathbb{R}_+\to\mathbb{R}$ be a functions such that - For all $(I,\tau,t)\in\mathcal{K}_{q}\times\{-1,1\}\times\mathbb{R}_+$, $$\label{equu:symetry} {\psi}(-I,\tau,t)=-{\psi}(I,\tau,t).$$ - There exist $\theta,\alpha, T>0$ such that for all $I\in\mathcal{K}_{q}$ there exists $\tau_I \in\{-1,1\}$ such that $\tau_I=\tau_{-I}$ and such that $$\theta+{\psi}(I,\tau_I,t )-\sigma(Iv(\tau_I t))>\alpha>{\psi}(I,\tau_I,-t )-\sigma(Iv(-\tau_I t)),\ \forall t\in(0,T).$$ - For all $I_0\in\{I_q,-I_q\}$ and all $ \tau\in\{-1,1\}$ $$\theta+{\psi}(I_0,\tau,t )-\sigma(I_0v(\tau t))>\alpha>{\psi}(I_0,\tau,-t )-\sigma(I_0v(-\tau t)),\ \forall t\in(0,T).$$ Then, $$\norm{v(t)}\le \theta+\alpha+ \frac{1}{2}\beta_T ,\quad\forall t\in [-T, T],$$ where $${\beta}_T:=\sup\Big\{\Big| {\psi}(I_q,\tau_I, t)-{\psi}(I, \tau_I, t)\Big|:\quad t\in [-T, T],\quad I\in\mathcal{K}_q^+ \Big\}.$$ By addition and by Equation , we have $$\begin{gathered} 2(\theta-\alpha)+{\psi}(I_0,\tau_I,t )+{\psi}(I,\tau_I,t )> \sigma( I_0 v(\tau_I t))+\sigma(Iv(\tau_I t)) ,\\ 2\alpha-{\psi}(I,\tau_I,-t )-{\psi}(I_0,\tau_I,-t )>\sigma(I_0v(-\tau_I t))+\sigma(Iv(-\tau_I t)).\end{gathered}$$ Use the fact $\tau_I=\tau_{-I}$ and again Equation : - For $I_0=I_q$ and $I=-I_{i,q}$ we obtain $$\begin{gathered} 2(\theta-\alpha)+{\psi}(I_q,\tau_I,t )-{\psi}(I_{i,q},\tau_I,t )> \sigma( I_q v(\tau_I t))-\sigma(I_{i,q}v(\tau_I t)),\\ 2\alpha+{\psi}(I_{i,q},\tau_I,-t )-{\psi}(I_q,\tau_I,-t )>\sigma(I_qv(-\tau_I t))-\sigma(I_{i,q}v(-\tau_I t)).\end{gathered}$$ - For $I_0=-I_q$ and $I=I_{i,q}$ we obtain $$\begin{gathered} 2(\theta-\alpha)-{\psi}(I_q,\tau_I,t )+{\psi}(I_{i,q},\tau_I,t )>-[\sigma( I_q v(\tau_I t))-\sigma(I_{i,q}v(\tau_I t))],\\ 2\alpha-{\psi}(I,\tau_I,-t )+{\psi}(I_q,\tau_I,-t )>-[\sigma(I_qv(-\tau_I t))-\sigma(I_{i,q}v(-\tau_I t))].\end{gathered}$$ Now, use the relation $$\forall u=(u_i)_{i=1}^q\in\mathbb{R}^q:\quad u_i=\frac{1}{2}[\sigma(I_q u)-\sigma(I_{i,q} u)],\quad\forall i=\overline{1,q},$$ to deduce that for every $i=\overline{1,q}$ there exists $\tau_i\in\{-1,1\}$ such that $$\begin{aligned} \Big|v_i(\tau_i t)+\frac{1}{2}\Big[{\psi}(I_q,\tau_i,t )-{\psi}(I_{i,q},\tau_i,t ) \Big]\Big|&< \theta+\alpha,\ \forall t\in(0,T)\\ \Big|v_i(-\tau_i t)-\frac{1}{2}\Big[{\psi}(I_q,\tau_i,-t )-{\psi}(I_{i,q},\tau_i,-t )\Big]\Big|&<\theta+\alpha,\ \forall t\in(0,T).\end{aligned}$$ Finally, by definition of the constants ${\beta}_T$ of the present Lemma, we find $$\norm{v(t )}\le \theta+\alpha+\frac{1}{2}{\beta}_T,\quad\forall t\in [-T, T].$$   \ For every $\alpha\in\mathbb{R}$ and $i=\overline{0,q}$ let ${\psi}_{i}(t,\alpha)$ be the solutions of the Equation of initial condition ${\psi}_{i}(0,\alpha)=\alpha$. For every $i=\overline{1,q}$ le be $\tau_i\in\{-1,1\}$ defined by the hypothesis \[H1\] and put $\tau_0=1$. Let $y$ be the solution of the differential Equation and put $$\overline{\delta}_i(t,\alpha):= -{\psi}_{i}(\tau_i t,\alpha)+\sigma(I_{i,q}{y}( \tau_i t)) \ \text{and}\ \underline{\delta}_i( t,\alpha):= -\overline{\delta}_i(t,\alpha),\quad\forall i=\overline{0,q},$$ Let be $D>0$ and put $\alpha_*=D^2/(\gamma(1-\gamma))$. Since ${y}(0)=\mathbb{0}_q$ then $\overline{\delta}_i(0,\alpha_*)=\underline{\delta}_i(0,\alpha_*)=\alpha_*$. By Lemma \[estimationzbyoverunder\], $$\dot{\overline{\delta}}_i(0,\alpha_*)>0\quad \text{and}\quad\dot{\underline{\delta}}_i(0,\alpha_*)>0 \quad \forall i=\overline{0,q}.$$ Since $\overline{\delta}_i(0,0)=\underline{\delta}_i(0,0)=0$, then there exists $\epsilon>0$ such that $$\begin{aligned} \overline{\delta}_i(t,\alpha_*)> \alpha_*> \overline{\delta}_i(-t,0) \quad\forall t\in (0,\epsilon),\quad \forall i=\overline{0,q},\\ \underline{\delta}_i(t,\alpha_*)> \alpha_*> \underline{\delta}_i(-t,0) \quad\forall t\in (0,\epsilon),\quad \forall i=\overline{0,q}.\end{aligned}$$ Define $$\begin{aligned} \overline{T}=&\sup\{t\geq 0:\quad\overline{\delta}_i(s,\alpha_*)> \alpha_*> \overline{\delta}_i(-s,0),\quad\forall s\in(0,t),\quad\forall i=\overline{0,q}\},\\ \underline{T}=&\sup\{t\geq 0:\quad\underline{\delta}_i(s,\alpha_*)> \alpha_*> \underline{\delta}_i(-s,0),\quad\forall s\in(0,t),\quad\forall i=\overline{0,q}\}\end{aligned}$$ Our strategy is to prove that $ \overline{T}=\underline{T}=+\infty$. By contradiction, suppose that $\overline{T}<+\infty$ or $\underline{T}<+\infty$, then there exists $i\in\{0,q\}$ such that one of the following cases is true $$\begin{aligned} &\overline{\delta}_i(\overline{T},\alpha_*) = \alpha_*,\quad \quad &&\overline{\delta}_i(-\overline{T},0) = \alpha_*,\\ &\underline{\delta}_i(\underline{T},\alpha_*) = \alpha_*,\quad &&\underline{\delta}_i(-\underline{T},0) = \alpha_*.\end{aligned}$$ Then $\overline{T}$ and $\underline{T}$ satisfies $$\begin{aligned} \label{contarditetoile}\frac{d}{dt}\Big[\exp(qt)\overline{\delta}_i(\overline{T},\alpha_*)\Big]\le 0\quad \text{and}\quad \frac{d}{dt}\Big[\exp(qt)\overline{\delta}_i(-\overline{T},0)\Big]\le0,\\ \label{contarditetoileB}\frac{d}{dt}\Big[\exp(qt)\underline{\delta}_i(\underline{T},\alpha_*)\Big]\le 0\quad \text{and}\quad \frac{d}{dt}\Big[\exp(qt)\underline{\delta}_i(-\underline{T},0)\Big]\le0.\end{aligned}$$ Integrate Equation to obtain $${\psi}_i(t,\alpha)=\alpha \exp\Big(\frac{1}{q} \int_{0}^t\sigma(I_{i,q}{\mathpzc{B}}( \nu)I_{i,q}) d\nu\Big)+{\psi}_{i}(t,0),\quad\forall t\in\mathbb{R},\quad\forall i=\overline{0,q}$$ By hypothesis there exists $\omega>0$ such that for all $\tau\in\{-1,1\}$ we have $$\begin{gathered} \sup\Big\{ \frac{\tau}{q} \int_{0}^t\sigma({\mathpzc{B}}( \tau\nu)) d\nu ,\quad\forall t\in\mathbb{R}\Big\}< \omega,\end{gathered}$$ and for all $i=\overline{1,q}$ there exists $\tau_i\in\{-1,1\}$ such that $$\begin{gathered} \sup\Big\{ \frac{\tau_i}{q} \int_{0}^t\sigma(I_{i,q}{\mathpzc{B}}( \tau_i \nu)I_{i,q}) d\nu,\quad\forall t\geq 0\Big\}< \omega.\end{gathered}$$ Let be $ T_*=\min\{\underline{T},\overline{T}\}$. We obtain the following estimations $$\begin{aligned} {\psi}_0(\tau t,\alpha)&\le \exp(\omega)\alpha+{\psi}_{0}(\tau t,0),\quad\forall \tau\in\{-1,1\},\quad\forall t\in(0,T_*),\\ {\psi}_i(\tau_i t,\alpha)&\le \exp(\omega)\alpha+{\psi}_{i}(\tau_i t,0),\quad\forall i=\overline{1,q},\quad\forall t\in(0,T_*).\end{aligned}$$ Put $$\psi(I_{i,q},\tau ,t):={\psi}_{i}(\tau t,0),\quad\forall i=\overline{0,q},\quad \forall \tau\in\{-1,1\}.$$ By definition of the function ${\psi}_{i}(t,0)$ we have: $\psi(-I_{i,q},\tau,t)=-\psi(I_{i,q},\tau,t)$ for every $i=\overline{0,q}$ and $\tau\in\{-1,1\}$. By definition of $T_*$ we have: for all $I\in\mathcal{K}_q$ there exists $1\le i\le q$ and $\tau_I\in\{-1,1\}$ such that $\tau_{I}=\tau_{-I}=\tau_i$ and such that for all $t\in(0,T_*)$ we have $$\begin{aligned} \exp(\omega)\alpha_*+\psi(I,\tau_I,t)-\sigma(I{y}( \tau_I t))> \alpha_*> \psi(I,\tau_I,-t)-\sigma(I{y}(- \tau_I t)),\end{aligned}$$ further, for all $\tau\in\{-1,1\}$ and $I\in\{I_q,-I_q\}$ we have $$\begin{aligned} \exp(\omega)\alpha_*+\psi(I,\tau,t)-\sigma(I{y}( \tau t))> \alpha_*> \psi(I,\tau,-t)-\sigma(I{y}( -\tau t)),\end{aligned}$$ where we recall that $\alpha_*=D^2/(\gamma(1-\gamma))$. By Hypothesis there exists $\beta>0$ such that $$\begin{gathered} \sup_{t\in\mathbb{R}}|{\psi}_{0}(\tau_i t)- {\psi}_i(\tau_i t)|<\beta,\quad\forall i=\overline{1,q}.\end{gathered}$$ By Lemma \[LemmeMAJOD\] , $$\label{Majorating} \norm{\gamma y(t)}<\gamma (\exp(\omega)+1) \alpha_* +\frac{1}{2}\gamma \beta ,\quad\forall [-T_*,T_*].$$ There exists $\gamma_*>0$ such that for all $\gamma \in(0,\gamma_*)$ there exists $D_\gamma>0$ such that $$\label{Equ:ValeurD} \frac{\exp(\omega)+1}{1-\gamma}D_\gamma^2+\gamma\beta<D_\gamma,$$ Choose $\alpha_*=D_\gamma^2/(\gamma(1-\gamma))$ and use equation to get $$\norm{\gamma y(t)}< \frac{\exp(\omega)+1}{1-\gamma}D_\gamma^2+\gamma \beta<D_\gamma,\quad\forall [-T_*,T_*].$$ Then $\norm{\gamma y(t)}<D_\gamma$ for every $t\in[-T_*,T*]$. In particular $\norm{\gamma y(T_*)}<D_\gamma$ and $\norm{\gamma y(-T_*)}<D_\gamma$. By Lemma \[estimationzbyoverunder\] and Equation , $$\begin{aligned} \frac{d}{dt}\Big[\exp(qt)\overline{\delta}_i(T_*,\alpha_*)\Big]> 0\quad \text{and}\quad \frac{d}{dt}\Big[\exp(qt)\overline{\delta}_i(-T_*,0)\Big]>0,\\ \frac{d}{dt}\Big[\exp(qt)\underline{\delta}_i(T_*,\alpha_*)\Big]> 0\quad \text{and}\quad \frac{d}{dt}\Big[\exp(qt)\underline{\delta}_i(-T_*,0)\Big]>0.\end{aligned}$$ We obtain a contradiction with Equation . Then $T_*=+\infty$ and we have $$\begin{aligned} \overline{\delta}_i(t,\alpha_*)> \alpha_*> \overline{\delta}_i(-t,0) \quad\forall t\geq0,\quad \forall i=\overline{0,q},\\ \underline{\delta}_i(t,\alpha_*)> \alpha_*> \underline{\delta}_i(-t,0) \quad\forall t\geq0\quad \forall i=\overline{0,q}.\end{aligned}$$ Finally, use again Equation and Lemma \[LemmeMAJOD\] to deduce that $$\norm{\gamma y(t)}< \frac{\exp(\omega)+1}{1-\gamma}D_\gamma^2+\gamma\beta<D_\gamma,\quad\forall t\in\mathbb{R}.$$ Proof of the Generalized main result ====================================  \ Consider the Equation . By hypothesis of the Generalized main result there exist $L$ such that $$\max\{\norm{df}_\infty,\norm{d^2f}_\infty\}<L.$$ Let be $\epsilon\in(0,1)$ and consider the change of variable $$\eta(t)= \epsilon^2\Big[ \mu( {\epsilon} t)- \Phi^{ {\epsilon} t}(x_0)\Big] ,\ t\in\mathbb{R},\ \eta:=(\eta_{i})_{i=1}^n,$$ where $\Phi^t:=(\Phi_i^t)_{i=1}^n$ is the flow of the system . Use Taylor formula $$\begin{gathered} \dot{ {\eta}} = \epsilon^3 [\dot{\mu}(\epsilon t)-f(\mu( \epsilon t))]+ {\epsilon}^2 df( \mu( \epsilon t)) {\eta} + {\eta}^T \zeta_\mu( t ) {\eta},\quad\eta(0)=\mathbb{0}_n, \quad t\in\mathbb{R},\end{gathered}$$ where he function $\zeta_\mu : \mathbb{R} \times\mathbb{R}^{n\times n\times n}$ is a continuous function and satisfies $\norm{\zeta_\mu}_\infty\le \norm{d^2f}_\infty<\epsilon L$. By hypothesis the function $\mu:\mathbb{R}\to\mathbb{R}^n$ is a leader trajectory of $f$ such that $\mu(0)=x_0$. By Theorem \[Prop:RiccatirotationvectorconditionTheo\] there exists $\gamma_*>0$ such that for all $\epsilon L<\gamma_*$ there exists $D_{\epsilon, L}$ such that $$\norm{\mu( t)- \Phi^{ t}(x_0)}=\frac{1}{\epsilon^2}\norm{\eta(\frac{t}{\epsilon})}<\frac{1}{\epsilon^2}D_{\epsilon, L},\quad \forall t\in\mathbb{R}.$$ Leader trajectory and Rotation vector formula ============================================= We show in this Section that for any periodic function $g$ there exists a constant vector $\rho\in \mathbb{R}^q$ such that the function $t\mapsto t\mathbb{1}_q $ is a leader trajectory of the function $t\mapsto I_{\rho}^{-1}g(\rho t)$. In order to prove the result let us introduce the following notations; Let $g=(g_i)_{i=1}^q :\mathbb{R}^{q}\to \mathbb{R}^q$ be a $C^1$ periodic function. Recall that the cone $V_+^q$ is defined by Equation . Let $ \Psi_i[g]:V_+^q\times\mathbb{R}\to\mathbb{R}$ be the functions defined by $$\begin{gathered} \notag\Psi_i[g](z,t):= \int_{0}^t A_i(z,s)\exp\Big( \frac{1}{q} \int_{s}^t\sigma\Big(I_{i,q} I_z^{-1} dg( \nu z)I_z I_{i,q} \Big) d\nu\Big)ds,\\ \label{Equ:DefPsi}A_i(z,s):=\sigma\Big(I_{i,q}(\mathbb{1}_q - I_z^{-1} g( s z))\Big),\quad\forall i=\overline{0,q},\end{gathered}$$ where $I_z$ is the diagonal matrix of diagonal $z$. For every $i=\overline{0,q}$ the function $\Psi_i[g](z,t)$ is the solution of the following one-dimensional differential equation $$\begin{gathered} \label{EquaPsiZFin}\dot{\psi}:= \sigma\Big(I_{i,q} (\mathbb{1}_q - I_z^{-1} g( t z))\Big)+\frac{1}{q} \sigma\Big(I_{i,q} I_z^{-1}dg( t z) I_z I_{i,q} \Big)\psi, \\ \notag \psi(0)=0,\quad t\in\mathbb{R}.\end{gathered}$$ \[Prop:conditionsatisfied\] Let $g=(g_i)_{i=1}^q :\mathbb{R}^{q}\to \mathbb{R}^q$ be a $C^1$ periodic function such $$\exists \beta,\gamma>0:\ \inf \{g_{i}(z),\ z\in\mathbb{R}^q,\ i=\overline{1,q}\}>1+\beta\quad \text{and}\quad \norm{dg}_\infty<\gamma.$$ Then there exits $\gamma_*>0$ such that for all $\gamma\in(0,\gamma_*)$ there exists $\rho\in V_+^q$ such that the function $t\mapsto t\mathbb{1}_q $ is a leader trajectory of the function $t\mapsto I_{\rho}^{-1}g(\rho t)$. Further, $\rho$ is solution of the following Equation $$\lim_{t\to+\infty} \frac{1}{t}\Big[\Psi_0[g](\rho, t )-\Psi_i[g](\rho, t)\Big]=0,\quad\forall i=\overline{1,q}.$$ For the proof of the following Lemma we refer to  [@Cheban]. \[PropWithoutProof\] Let be $q\in\mathbb{N}^*$ and $g,h:\mathbb{R}^q\to\mathbb{R}$ be a periodic function. Let be $z\in\mathbb{R}^q$ and suppose that $$\lim_{t\to+\infty}\frac{1}{t}\int_{0}^{t}g(zs)ds\le 0,$$ then there exists $\lambda_z\in\mathbb{R}$ and a bounded function $p_z:\mathbb{R}\to\mathbb{R}$ such that the following differential equation $$\dot{\psi}=h(zt)+g(zt)\psi,\quad\psi(0)=0,$$ admits a solution $\psi:\mathbb{R}\to\mathbb{R}$ of the form $$\psi(t)=\lambda_z t+p_z(t), \quad\forall t\in\mathbb{R}.$$ In particular, the Lemma is true when $g(\nu)=0$ for all $\nu\in\mathbb{R}^q$; in this case we have $$\norm{p_z}_\infty<\norm{h}_\infty,\quad\forall z=(z_i)_{i=1}^q,\quad z_i\geq1,\quad\forall i=\overline{1,q}.$$ Thanks to Lemma \[PropWithoutProof\], for every $i=\overline{0,q}$ and $z\in V_+^q$ let $\tau_i:\mathbb{R}^q\to\{-1,1\}$ defined by $$\label{Equ:DerTauIz}\tau_{i}(z)=-\operatorname{Sign}(\lambda_{i,z}),\quad \lambda_{i,z}:=\lim_{t\to+\infty}\frac{1}{t}\int_{0}^t\sigma\Big(I_{i,q} I_z^{-1} dg( \nu z)I_z I_{i,q} \Big)d\nu,$$ where the function $\operatorname{Sign}:\mathbb{R}\to\{-1,1\}$ is defined by $\operatorname{Sign}(t)=1$ if $t\geq0$ and $\operatorname{Sign}(t)=-1$ if $t<0$. Remark that for all $\tau\in\{-1,1\}$ we have $$\begin{gathered} \label{gderiveiq}\sigma\Big( I_z^{-1} dg( \tau t z) I_z \Big)=\tau\frac{d}{dt}\Big[I_z^{-1} g( \tau t z)\Big],\quad\forall z\in V_+^q,\\ \label{gderiveiqb}\sigma\Big( I_{i,q} I_z^{-1} dg( \tau t z) I_z I_{i,q} \Big)={ \tau }\frac{d}{dt}\Big[I_z^{-1} g( \tau t z)\Big]-\frac{2}{z_i}\frac{\partial}{\partial z_i}g(\tau z t),\quad\forall z\in V_+^q,\end{gathered}$$ where $\frac{\partial}{\partial z_i}g$ is the $i^{th}$ partial derivative of $g$. In order to prove that the function $t\mapsto t\mathbb{1}_q$ is a leader trajectory of the function $t\mapsto I_\rho^{-1}g(\rho t)$, we must prove that $$\sup_{t\in\mathbb{R }}\Big|\Psi_0[g](\rho, \tau_i(\rho)t )-\Psi_i[g](\rho,\tau_i(\rho) t)\Big|<+\infty,\ \forall i=\overline{1,q}.$$ Again, by Lemma \[PropWithoutProof\] , it is sufficient to prove that $$\label{equasufficent} \lim_{t\to+\infty} \frac{1}{t}\Big[\Psi_0[g](\rho,t )-\Psi_i[g](\rho,t)\Big]=0,\ \forall i=\overline{1,q}.$$ In order to simplify, for every $i=\overline{0,q}$, $z\in\mathbb{R}^q$ and $t, s\geq0$ denote $$\begin{aligned} \theta_i(z,t,s)&:= \exp\Big( \gamma\frac{\tau_i(z)}{q} \int_{s}^t\sigma\Big( I_{i,q} I_z^{-1} dg( \tau_i(z) \nu z) I_z I_{i,q} \Big)d\nu\Big),\\ \theta_{0}^{i}(z,t,s)&:= \exp\Big( \gamma \frac{\tau_i(z)}{q}\int_{s}^t\sigma\Big( I_z^{-1} dg( \tau_i(z) \nu z) I_z \Big)d\nu\Big),\\ \underline{\theta}_i(z,t,s)&:= \theta_{0}^{i}(z,t,s)-\theta_i(z,t,s),\\ \Theta_i(z,t)&:= \int_{0}^t {\theta}_{0}^{i}(z,t,s) +\frac{1}{2} \underline{\theta}_i(z,t,s) \sigma\Big(I_{i,q}(\mathbb{1}_q - I_z^{-1}g( \tau_i(z) s z))\Big) ds.\end{aligned}$$ For every fixed $k\in\mathbb{N}$, define the function $\Gamma_k=(\Gamma_{i,k})_{i=1}^q:z\in V_+^q\to\mathbb{R}^q$ by $$\begin{aligned} \Gamma_{i,k}(z):=& \frac{1}{ \Theta_i(z,k) }{\int_{0}^{k} {\theta}_{0}^{i}(z,k,s) g_i( \tau_{i}(z) s z) ds },\quad z\in V_+^q.\end{aligned}$$ Equation is satisfied if there exists $\rho\in z\in V_+^q$ and a sequence $u_k\to+\infty$ such that $\lim_{k\to+\infty}\Gamma_{u_k}(\rho)=\rho$. We have $$\Theta_i(z,k) \Gamma_{i,k}(z)- {\int_{0}^{k} {\theta}_{0}^{i}(z,k,s) g_i( \tau_{i}(z) s z) ds }=0.$$ By the Mean value Theorem, for all $z\in V_+^q$ and all $k\in\mathbb{N}$ there exists $s_{i,k}^z\geq0$ such that $$\begin{gathered} {\theta}_{0}^{i}(z,k,s_{i,k}^z) +\frac{1}{2} \underline{\theta}_i(z,k,s_{i,k}^z) \sigma\Big(I_{i,q}(\mathbb{1}_q - I_z^{-1}g( s_{i,k}^z\tau_i(z) z))\Big) \Gamma_{i,k}(z)\\ = {\theta}_{0}^{i}(z,k,s_{i,k}^z) g_i( s_{i,k}^z\tau_{i}(z) z) .\end{gathered}$$ That follows $$\label{Equa:EstimGama}\Gamma_{i,k}(z):=\frac{ g_i( s_{i,k}^z \tau_{i}(z) z) }{ 1+\zeta_i(z,k) },$$ where $$\begin{aligned} \notag\zeta_i(z,k)&=\frac{1}{2}\frac{ \underline{\theta}_i(z,k,s_{i,k}^z) }{{\theta}_{0}^{i}(z,k,s_{i,k}^z)}\sigma\Big(I_{i,q}(\mathbb{1}_q - I_z^{-1}g( s_{i,k}^z\tau_i(z) z))\Big)\\ \label{Equ:DefZeta}&=\frac{1}{2}\Big[1-\frac{ \theta_i(z,k,s_{i,k}^z) }{{\theta}_{0}^{i}(z,k,s_{i,k}^z)}\Big]\sigma\Big(I_{i,q}(\mathbb{1}_q - I_z^{-1}g(s_{i,k}^z \tau_i(z) z))\Big).\end{aligned}$$ Let be $L>1$ and denote $$\mathcal{V}_L=\Big\{z=(z_i)_{i=1}^q\in z\in V_+^q:\ \norm{z}\le (L-1)\norm{g}_\infty\quad\text{and}\quad z_i>1,\ \forall i=\overline{1,q}\Big\}.$$ By definition of the coefficients $\tau_i(z)$ in Equation and by Lemma \[PropWithoutProof\], using Equations and we find the following relations $$\begin{gathered} \notag\Big|\frac{1}{q} \int_{0}^t\sigma\Big( I_z^{-1} dg( \nu z) I_z \Big)d\nu\Big|<\gamma ,\quad\forall z\in \mathcal{V}_L,\quad t\in\mathbb{R},\\ \frac{\tau_i(z)}{q}\int_{0}^t \sigma\Big( I_{i,q} I_z^{-1} dg( \tau_i(z) \nu z)I_{i,q} I_z \Big)d\nu<\gamma ,\quad\forall z\in \mathcal{V}_L,\quad\forall t\geq0.\end{gathered}$$ That implies $$\begin{gathered} \theta_i(z,t,s)<\exp(\gamma ),\quad\forall i=\overline{1,q},\quad\forall t\geq s\geq0,\quad\forall z\in\mathcal{V}_L,\\ \exp(-\gamma )<\theta_{0}^{i}(z,t,s)<\exp(\gamma ),\quad\forall i=\overline{1,q},\quad\forall t\geq s\geq0,\quad\forall z\in\mathcal{V}_L.\end{gathered}$$ For all $t\geq0$ and all $z\in\mathcal{V}_L$, we have the following estimation $$|\zeta_i(z,t)|\le \frac{1}{2} q[\exp(2\gamma)-1]\Big[\norm{z}+\norm{g}_\infty\Big]\le \frac{1}{2}qL[\exp(2\gamma )-1]\norm{g}_\infty.$$ Again, by the hypotheses of the present Proposition, there exists $\beta>0$ such that $$\inf \{g_{i}(z),\quad z\in\mathbb{R}^q,\ i=\overline{1,q}\}>1+\beta$$ There exists $\gamma_*>0$ such that for all $\gamma\in(0,\gamma_*)$ there exists ${L_{\gamma}}>0$ such that $$\begin{aligned} {L_{\gamma}}>\frac{ 2 }{ 2- q{L_{\gamma}}(\exp(2\gamma)-1) \norm{g}_\infty}\ \text{and}\ \frac{ 2(1+\beta) }{ 2+q{L_{\gamma}}(\exp(2\gamma )-1)\norm{g}_\infty}>1\end{aligned}$$ By Equations and , for a fixed $\gamma\in(0,\gamma_*)$ and $L_\gamma>1$, the compact and convex set $\mathcal{V}_{L_{\gamma}}$ satisfies $$\Gamma_k(\mathcal{V}_{L_{\gamma}}) \subset \mathcal{V}_{L_{\gamma}},\quad\forall k\in\mathbb{N}.$$ and the function $\Gamma_k$ is $C^1$ on $\mathcal{V}_{L_{\gamma}}$. By Brouwer fixed-point theorem, for all $k\in\mathbb{N}$ there exists $\rho_k\in\mathcal{V}_{L_{\gamma}}$ such that $\Gamma_k(\rho_k) =\rho_k$. There exists $\rho \in \mathcal{V}_{L_{\gamma}}$ and there exists a sequence $(u_k)_{k\in\mathbb{N}}\subset\mathbb{N}$ such that $$\lim_{k\to+\infty}\rho_{u_k}=\rho\ \text{and}\ \lim_{k\to+\infty}\frac{1}{u_k}\Gamma_{u_k}(\rho,u_k)=\rho.$$ In other words, $$\lim_{k\to+\infty}\frac{1}{u_k}[\Psi_0[g](\rho, u_k)-\Psi_i[g](\rho,u_k)]=0,\quad\forall i=\overline{1,q}.$$ Hence, Equation is satisfied. By Equation of this proof, we deduce that the the function $t\mapsto t\mathbb{1}_q $ is a leader trajectory of the function $t\mapsto I_{\rho}^{-1}g(\rho t)$. Proof of the Main result ======================== Let $x$ be the solution of the system \[systemNotNormalized\] and suppose that $f$ is periodic. By periodicity, there exists $>0$ such that $$\label{majofdff}\max\Big\{\norm{f}_\infty, \norm{df}_\infty,\norm{df}_\infty\Big\}<\omega.$$ Let be ${c},{\gamma}>0$, and consider the following change of variable $$\tilde{x}(t):={c}t\mathbb{1}_n+ x( {\gamma}t)-x_0,\quad\forall t\in\mathbb{R}.$$ System \[systemNotNormalized\], becomes $$\dot{\tilde{x}}= {c}\mathbb{1}_n+\gamma \tilde{f}_{x_{0}}(\tilde{x},t),\quad\tilde{x}(0)=\mathbb{0}_n,\quad t\in\mathbb{R}.$$ $$\tilde{f}_{x_{0}}(z,z_{n+1})= {c}\mathbb{1}_n+\gamma f( z+x_0-c z_{n+1}\mathbb{1}_n),\quad\forall z=(z_i)_{i=1}^{n+1}\in\mathbb{R}^{n+1}.$$ Put $\tilde{f}_{x_{0}}:= (\tilde{f}_{x_{0},i})^n_{i=0}$, thanks to Equation we get $$\begin{gathered} \forall (z,t)\in\mathbb{R}^n\times\mathbb{R}:\quad\max\Big\{ \norm{\frac{d}{dz}\tilde{f}_{x_{0}}(z, t) },\norm{\frac{d^2}{dz^2}\tilde{f}_{x_{0}}(z, t) }\Big\}<{\gamma}{\omega},\\ \inf\Big\{\tilde{f}_{x_{0},i}(z,z_{n+1}),\quad i=\overline{1,n},\quad z=(z_i)_{i=1}^{n+1}\in\mathbb{R}^{n+1}\Big\}>c-\gamma \omega.\end{gathered}$$ Le be $\beta>0$ fixed then fora all $c>1+\beta+\gamma \omega$ we get $$\label{equ:carbitra} \tilde{f}_{x_{0}}(z,z_{n+1})>1+\beta, \quad\forall z=(z_i)_{i=1}^{n+1}\in\mathbb{R}^{n+1}.$$ By Proposition \[Prop:conditionsatisfied\], there exist $\gamma>0$ and ${\rho}_*\in\mathbb{R}^n$ such that the function $\mu(t):=t\mathbb{1}_n $ is a leader trajectory of the function $I_{{\rho}_*}^{-1}\tilde{f}_{x_{0}}$. Since $\tilde{x}_0=\mathbb{0}_n$ and by the Generalized main result we deduce that $$\sup_{t\in\mathbb{R}}\norm{ t\mathbb{1}_n -I_{{\rho}_*}^{-1}\tilde{x}(t)}<+\infty.$$ By definition of $\tilde{x}$ we obtain $$\sup_{t\in\mathbb{R}}\norm{\frac{t}{\gamma}({\rho}_*-c\mathbb{1}_n) -x_0+ x( t)}<+\infty.$$ In other words, the solution $x$ of the Equation admits a \[rotationvector\] $\rho:=({\rho}_*-c\mathbb{1}_n)/\gamma$ . Again by Proposition \[Prop:conditionsatisfied\] and by the uniqueness of the rotation vector the vector we deduce that $\rho_*$ is the unique solution of the following equation $$\quad \lim_{t\to+\infty}\frac{1}{t}\Big[\Psi_0[\tilde{f}_{x_0}](\rho_*, t)-\Psi_i[\tilde{f}_{x_0}](\rho_*, t)\Big]=0,\quad\forall i=\overline{1,n}$$ where the functions $\Psi_i[\tilde{f}_{x_0}]$ are defined in Equation of the previous Section. Since ${\rho}_*=\gamma\rho+c\mathbb{1}_n$ and by using the change of variable $t\to \gamma t$ then it is equivalent $$\lim_{t\to+\infty}\frac{1}{t}\Big[\Psi_0[f_{x_0}](\rho_*,c, t)-\Psi_i[f_{x_0}](\rho_*,c, t)\Big]=0,\quad\forall i=\overline{1,n},$$ where $$\begin{gathered} \Psi_i[{f}_{x_0} ](\rho,c,t):= \int_{0}^t A_i(\rho,s)\exp\Big( \frac{1}{n} \int_{s}^t\sigma\Big(I_{i,n}I_{{\rho}_*}^{-1} d f( \nu {\rho} +x_0 )I_{{\rho}_*} I_{i,n} \Big) d\nu\Big)ds,\\ A_i(\rho,c,s):=\sigma\Big(I_{i,n} I_{{\rho}_*}^{-1} (\rho - f(s {\rho} +x_0))\Big),\quad\forall i=\overline{0,n}.\end{gathered}$$ By Equation the coefficient $c$ is arbitrary in $(1+\beta+\gamma \omega,+\infty)$, since $\rho$ is unique, we get $$\lim_{c\to\infty}\lim_{t\to+\infty}\frac{c}{t}\Big[\Psi_0[\tilde{f}_{x_0}](\rho_*,c, t)-\Psi_i[\tilde{f}_{x_0}](\rho_*,c,t)\Big]=0,\quad\forall i=\overline{1,n}.$$ Finally, by Lemma \[PropWithoutProof\] of the previous Section, we deduce that $\rho$ is the unique solution of the following Rotation vector formula $$\lim_{t\to+\infty}\frac{1}{t}\Big[\Psi_0[f_{x_0}](\rho, t)-\Psi_i[f_{x_0}](\rho, t)\Big]=0,\quad\forall i=\overline{1,n},$$ where $$\begin{gathered} \Psi_i[{f}_{x_0} ](\rho,t):= \int_{0}^t A_i(\rho,s)\exp\Big( \frac{1}{n} \int_{s}^t\sigma\Big(I_{i,n} d f( \nu {\rho} +x_0 ) I_{i,n} \Big) d\nu\Big)ds,\\ A_i(\rho,s):=\sigma\Big(I_{i,n} (\rho - f(s {\rho} +x_0))\Big),\quad\forall i=\overline{0,n}.\end{gathered}$$ Conclusion and open problem =========================== We have proved that for any periodic field $f$ and any initial condition the solution of the system admits a unique rotation vector which is the solution of the Rotation vector formula . We conjecture that the solution $x$ of initial condition $x_0$ and rotation vector $\rho$ is periodic on the torus if and only if there exists $T>0$ such that $T\rho\in\mathbb{Q}^n$ and $\Psi_0[f_{x_0}](\rho, T)-\Psi_i[f_{x_0}](\rho,T) =0$ for every $i=\overline{1,n}$. Consider the systems $$\dot{x}=f(x),\quad x(0)=x_0\in\mathbb{R}^n\quad\text{and}\quad\dot{y}=f(y)+\zeta(t),\quad y(0)=x_0,$$ where $f:\mathbb{R}^n\to\mathbb{R}^n$ is a $C^1$ periodic function and $\zeta:\mathbb{R}\to\mathbb{R}^n$ is an integrable function. Let $\rho_x$ be the rotation vector of the solution $x$. Thanks to the Rotation vector formula and the Generalized main result, we deduce that the rotation vector $\rho_y$ of the solution $y$ exists and satisfies $\rho_x=\rho_y$ if and only if $$\begin{gathered} \sup_{t\in\mathbb{R}}\Big|\Psi_0[\zeta](t)-\Psi_i[\zeta](t)\Big|<+\infty,\quad\forall i=\overline{1,n},\\ \Psi_i[\zeta](t):=\int_{0}^t \exp\Big(\frac{1}{n}\int_{s}^t \sigma\Big(I_{i,n} df(\rho \nu +x_0) I_{i,n} \Big)d\nu\Big) \sigma( I_{i,n}\zeta(s))ds,\ i=\overline{0,n},\end{gathered}$$ which gives information about the Arnold Tongue. The leader trajectory defined in this article is an approximate solution. [99]{} , [The [K]{}uramoto model: A simple paradigm for synchronization phenomena]{}, [Rev. Mod. Phys.]{}, [77]{},[137–185]{},[2005]{}. , [Phys. Rev. Lett]{}, [2001]{}, [86]{}, [4278]{}. , [Critical exponents of the transition from incoherence to partial oscillation death in the [W]{}infree model]{}, [Journal of Statistical Mechanics: Theory and Experiment]{}, [2009]{}, [10]{}, [P10014]{}, [2009]{}. , [Phase synchronization of bursting neurons in clustered small-world networks]{}, [Phys. Rev. E]{}, [86]{}, [1]{}, [016211]{}, = [12]{}, [2012]{}. D.N. Cheban, Asymptotically Almost Periodic Solutions of Differential Equations, Hindawi Publishing Corporation, New York - Cairo, 2009. , [Complete synchronization of [K]{}uramoto oscillators with finite inertia]{}, [Physica D: Nonlinear Phenomena]{}, [1]{}, [32–44]{}, [240]{}, [2011]{}. , [N. Chopra, M. W. Spong]{},[On Exponential Synchronization of [K]{}uramoto Oscillators]{}, [IEEE Transactions on Automatic Control]{}, [54]{}, [2]{}, [353–357]{} year=[2009]{}. , [Cooray, G.]{},[The [K]{}uramoto Model]{}, [U.U.D.M. project report. Uppsala University, Department of Mathematics.]{}, [2008]{}. Corduneanu, B., Almost periodic oscillations and waves., New York ; London : Springer, ISBN 9780387098180 (hbk.), 2008. , [Generalising the [K]{}uramoto model for the study of neuronal synchronisation in the brain]{}, [Physica D: Nonlinear Phenomena]{}, [2]{}, [181–196]{}, [226]{}, [2007]{}. , [Stable Periodic Solutions to Discrete and Continuum Arrays of Weakly Coupled Nonlinear Oscillators]{}, [SIAM Journal on Applied Mathematics]{}, [52]{}, [6]{}, [1665–1687]{}, [1992]{}. , [The Effects of Spike Frequency Adaptation and Negative Feedback on the Synchronization of Neural Oscillators]{}, [Neural Computation]{}, [13]{}, [6]{}, [1285–1310]{}, [2001]{}. Ermentrout, G. B. and Kopell, N., Multiple pulse interactions and averaging in systems of coupled neural oscillators, Journal of Mathematical Biology, 29, 3, 195–217, 1432–1416, 1991. , [Almost Periodic Functions. L.N.M. ]{}, [ Springer]{},[1974]{}. Françoise, J-P., Oscillations en biologie : analyse qualitative et modèles, Mathématiques et applications, Springer, Berlin, Heidelberg, New York, 3-540-25152-9, 2005. , [Global attractor and asymptotic dynamics in the [K]{}uramoto model for coupled noisy phase oscillators]{}, [Nonlinearity]{}, [25]{}, [5]{}, [1247]{}, [2012]{}. , [Phase diagram of a generalized [W]{}infree model]{}, [ Phys. Rev. E]{}, [75]{}, [ 051104]{},2007. , [Synchrony, stability, and firing patterns in pulse-coupled oscillators]{}, [Physica D]{},, [163]{}, [191–216]{}, 2002. , [ Emergence of phase-locked states for the [W]{}infree model in a large coupling regime]{}, [Discrete and Continuous Dynamical Systems, Series A, 35]{}, [8]{}, [3417–3436]{}, 2015. , [Emergent dynamics of [W]{}infree oscillators on locally coupled networks]{}, [Journal of Differential Equations, 260]{}, [5]{}, [4203 - 4236]{},2016. , [A quantitative description of membrane current and its application to conduction and excitation in nerve]{}, [The Journal of physiology]{}, [4]{}, [500–544]{}, [The Physiological Society]{}, [117]{}, [1952]{}. , [ On the stability of the [K]{}uramoto model of coupled nonlinear oscillators]{}, [ Proceedings of the American Control Conference]{}, [R4296–4301]{},2004. , [Cerveau [&]{} comportement]{}, [9782744501371]{}, [Neurosciences [&]{} cognition]{}, [De Boeck]{}, 2002. , [International Symposium on Mathematical Problems in Theoretical Physics]{}, [ Lecture Notes in Physics, Springer, New York]{}, [39]{}, [1975]{}. ,[Chemical Oscillations, Waves, and Turbulence]{}, [978-3-642-69691-6]{}, , [Springer Berlin Heidelberg]{},[19]{}, [1984]{}. , [ Coexistence of coherence and incoherence in nonlocally coupled phase oscillators]{}, [ Nonlinear Phenom. Complex Syst]{},[5]{}, [380]{},2002. S. Louca and F. M. Atay, Spatially structured networks of pulse-coupled phase oscillators on metric spaces, Discrete and Continuous Dynamical Systems, 34, 9, 3703–3745, 2014. Y. L. Maistrenko, O. V. Popovych and P. A. Tass, [Desynchronization and chaos in the [K]{}uramoto model]{}, [Springer Berlin Heidelberg]{}, [Berlin, Heidelberg]{}, [285–306]{}, [978-3-540-31520-9]{},2005. , [Synchronization of Pulse-Coupled Biological Oscillators]{}, [SIAM Journal on Applied Mathematics]{},[50]{}, [1645–1662]{},1990. R. E. Mirollo and S. H. Strogatz, The spectrum of the locked state for the [K]{}uramoto model of coupled oscillators , Physica D: Nonlinear Phenomena , 205, [1-4]{}, 249 - 266, 0167-2789, 2005. , [Phase shifts between synchronized oscillators in the [W]{}infree and [K]{}uramoto models]{}, [InterJournal Complex Systems]{}, [1408]{}, [2004]{}. , [Coherence-incoherence patterns in a ring of non-locally coupled phase oscillators]{}, [Nonlinearity]{}, [26]{}, [2469–2498]{},2013. Oukil, W, Kessi, A. and Thieullen, Ph , Synchronization hypothesis in [W]{}infree model, Dynamical Systems, Taylor [&]{} Francis, 2016. , [ Chimera states: coexistence of coherence and incoherence in networks of coupled oscillators]{},[Nonlinearity]{}, [28]{}, [R67]{},2015. , [Low-dimensional Dynamics of Populations of Pulse-Coupled Oscillators]{}, [Phys. Rev. X ]{}, [4]{}, [011009]{},2014. , [Phase chaos in coupled oscillators]{}, [Phys. Rev. E]{}, [71]{}, [6]{}, [065201]{}, [4]{}, [2005]{}. , [Physical Review E]{}, [75]{}, [036218]{},2007. , [On dynamical systems in n-dimensional torus]{}, [Proceedings of the Symposium on Differential Equations and Dynamical Systems. Lecture Notes in Mathematics]{}, [206]{}, [18–19]{}, [1971]{}. , [ Biological rhythms and the behavior of populations of coupled oscillators]{}, [J. Theor. Bio]{},[16]{}, [15–42]{}, [1967]{}. Winfree, A. T., The geometry of biological time, Biomathematics, Springer Verlag, New York, 0-387-09373-7, 1980. , [Spiking Neuron Models: An Introduction]{}, [0521890799]{}, [Cambridge University Press]{}, [New York, NY, USA]{},2002. Y. Yi, On almost automorphic oscillations. Differences and differential equations,, Fields Inst. Commun. Amer. Math. Soc., Providence, RI, 42, 2004.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Cardiac magnetic resonance imaging (MRI) is an essential tool for MRI-guided surgery and real-time intervention. The MRI videos are expected to be segmented on-the-fly in real practice. However, existing segmentation methods would suffer from drastic accuracy loss when modified for speedup. In this work, we propose Multiscale Statistical U-Net (MSU-Net) for real-time 3D MRI video segmentation in cardiac surgical guidance. Our idea is to model the input samples as multiscale canonical form distributions for speedup, while the spatio-temporal correlation is still fully utilized. A parallel statistical U-Net is then designed to efficiently process these distributions. The fast data sampling and efficient parallel structure of MSU-Net endorse the fast and accurate inference. Compared with vanilla U-Net and a modified state-of-the-art method GridNet, our method achieves up to 268% and 237% speedup with 1.6% and 3.6% increased Dice scores.' author: - Tianchen Wang - Jinjun Xiong - Xiaowei Xu - Meng Jiang - Haiyun Yuan - Meiping Huang - Jian Zhuang - Yiyu Shi bibliography: - 'bib.bib' title: 'MSU-Net: Multiscale Statistical U-Net for Real-time 3D Cardiac MRI Video Segmentation' --- Introduction ============ Real-time Magnetic Resonance Imaging (MRI) techniques have been providing fast and accurate visual guidance in multiple fields. The duration of cardiac surgery (e.g., prosthetic valve implantation in the correct location at the aortic annulus) has been significantly shortened since interactive real-time MRI was getting applied [@mcveigh2006real]. Interventional real-time MRI has also been adopted for congenital, ischemic, and structural heart disease for its capacity of visualizing 3D anatomy and assessing myocardial tissue as well as local hemodynamics [@campbell2017real]. To achieve real-time MRI guidance, the images need to be segmented on-the-fly, at a speed of at least **30**, preferably up to 100 frames per second (FPS) [@schaetz2017accelerated; @iltis2015high]. However, performing real-time segmentation on cardiac MRI images is a challenging task. In addition to the difficult effects such as anisotropic resolution, cardiac border ambiguity and large variations among targeting objects from patients [@zheng20183d], the requirement of real-time fast segmentation demands a lightweight and efficient processing framework. Existing approaches used complicated neural network architectures to achieve good accuracy and were not able to make inference in real time [@isensee2017automatic; @zotti2018convolutional]. Recently, Statistical Convolutional Neural Network (SCNN) was proposed to speed up conventional CNNs with little performance loss in video object detection [@wang2019scnn]. Instead of feeding the input samples as deterministic values, SCNN used Independent Component Analysis (ICA) to extract parameterized statistical distributions in canonical form to compactly model the temporally and contextually correlated information. Then the network model propagated the distributions in canonical form more efficiently than deterministic values. In this work, we propose Multiscale Statistical U-Net (MSU-Net) for real-time cardiac MRI segmentation. We incorporate SCNN and a new multiscale data sampling method with the U-Net to capture spatio-temporal correlation in the input data. Our model adopts a parallel architecture to efficiently propagate the multiscale distributions. Specifically, we apply ICA with multiple sets of temporal image patches to generate a cluster of canonical form distributions, each of which represents a different scale to model the input data. This multiscale sampling method can preserve the information of spatio-temporal correlations at different scales. Then we implement a number of parallel yet light-weight encoder-decoder style branches for efficient inference. Each branch propagates the specific scale of canonical form distributions. Experimental results show that our MSU-Net achieves up to 268% and 237% speedup with 1.6% and 3.6% increased Dice scores compared with vanilla U-Net and a modified state-of-the-art method GridNet [@zotti2018convolutional]. Background ========== SCNN [@wang2019scnn] was the first model that feeds CNNs with a reasonable number of statistical distributions that were decomposed from the input data. SCNN is lighter and thus of higher speed than conventional CNNs that conduct deterministic operations (such as sum and max). SCNN applied ICA to decompose video frames that exhibit spatio-temporal correlation into canonical form distributions as follows: $$D\!=\!a_{0}\!+\!a_{1}X_1\!+\!\dots\!+\!a_{m}X_m\!+\!a_{r}R, \label{eq:icar}$$ where (1) $D$ is a random multivariate signal, which in video object detection represents the same pixel across multiple frames in a snippet; (2) $a_{0}$ is the mean value of $D$; (3) $X_{i}~(i\in\{1 \dots m\})$ are additive independent subcomponents of $D$; (4) $a_{i}~(i\in\{1 \dots m\})$ are the corresponding weight act as mixing matrix; (5) $R$ denotes uncorrelated Gaussian noise and $a_{r}$ is the weight of $R$; (6) $m$ is the basis dimension of the canonical form distribution. With the help of predefined core operations (weighted [*sum*]{} and [*max*]{}) that keep their outputs still in canonical form distributions, SCNN needs little modification to the standard gradient descent based scheme. It can be trained using the same forward and back propagation procedures as conventional CNNs. At the output, the results are mixed to form a temporal feature map for each sample by plugging in the values of independent sources $X_i$ from the ICA process. By processing multiple frames at a time through distributions, SCNN significantly speedups object detection in videos over conventional CNNs with slight accuracy degradation. Method ====== In this section, we first present a multiscale sampling method to extract canonical form distributions from input 3D MRI videos. Then we introduce the architecture of MSU-Net and explain how it processes these distributions for real-time segmentation. Distribution Extraction with Multiscale Data Sampling {#sec:corr_model} ----------------------------------------------------- In order to build linear distributions in parameterized canonical form (Equation \[eq:icar\]) via ICA, we need to decide how to properly extract samples from 3D MRI video to feed into ICA, i.e., what information each $D$ should represent. In the approach of SCNN for video object detection [@wang2019scnn], the video clips are resized and split into small snippets, and each distribution $D$ models the same pixel across multiple frames in the same snippet. However, this cannot be applied to 3D MRI video directly since lots of semantic details important to segmentation would be lost. Thus, we propose to use $D$ to represent a patch within a small range (both spatially and temporally) where strong correlation exists. Specifically, we denote the dimension of an input 3D MRI video as \[$X$, $Y$, $Z$, $T$\], where $X-Y$ plane is the short axis plane, $Z$ is the short axis and $T$ is the temporal dimension. The common issues of slice shifting as well as large inter-slice gap in MRI cardiac images along short-axis ($Z$ axis) [@zotti2018convolutional] lead to minimum spatial correlation in $X-Z$ and $Y-Z$ planes. Therefore, we extract patches within the dimension \[$X$, $Y$, $T$\], independent of $Z$. ![(a) Illustration of multi-scale data sampling from cropped 3D MRI video. Each canonical form is extracted using the samples from the same position in the $X-Y$ plane and collected at different time steps. Different canonical forms can have different patch sizes. (b) Restored inputs with multi-scale data sampling method at different time steps (t) using different compression ratio (r). The restored inputs with r=1/40 have more noises than those with r=1/10. []{data-label="fig:cropping_restore"}](fig-cropping_restore.pdf){width="1\linewidth"} Before extracting the patches, the 3D MRI videos are normalized first to remove offsets among videos. Each patch is then extracted using a window of size $(n, n)$ on the $X-Y$ plane over $t$ time steps. We call $t$ as snippet span. We propose to allow different canonical forms to have different $n$ and $t$, as such an approach covers potential spatio-temporal correlations at different scales. We call this cluster of distributions with multiple patch sizes as multiscale distributions. An example of the extraction process of multiscale distributions with different patch sizes on one slice is shown in Fig. \[fig:cropping\_restore\] (a). The patches are collected at the same position over time and fed to ICA to extract canonical form distributions. ICA has to be used because the propagation of the canonical form distributions requires all the bases to be independent. Other approahces such as PCA cannot guarantee this unless the samples follow Gaussian distributions, which is not the case in our problem. As a result, the snippet of 3D MRI video is “collapsed” into a smaller 3D image, with each voxel representing a canonical form distribution (Equation \[eq:icar\]) that has both spatial and temporal correlations: with patch size $(n, n, t)$ and predetermined independent basis dimension $d$, a compression ratio of $r=d/(n^2t)$ is achieved. To show the feasibility of the proposed data sampling, we extract the multiscale canonical form distributions using our procedure with various compression ratio $r$ by changing the basis dimension $d$ with $n=7$ and $t=5$. The visual results along with the compression ratio are shown in Fig. \[fig:cropping\_restore\] (b). With a larger ratio ($\text{r}\!=\!1/40$, smaller basis dimension of canonical form distribution), the restored video gain more noise with vague contours, which would bring obstruction to the segmentation task. With a smaller ratio ($\text{r}\!=\!1/10$), the difference between the input video and the restored mixing video is negligible. Therefore, we adopt $\text{r}\!=\!1/10$ as the compression ratio in our following experiments. ![The architecture of MSU-Net. The number of Blocks in DownTube/UpTube varies to accommodate the various input dimensions.[]{data-label="fig:msunet"}](fig-arch.pdf){width="1\linewidth"} Real-time Segmentation with MSU-Net {#sec:msunet} ----------------------------------- The multiscale canonical form distributions provide compact data representation for efficient processing. In this subsection, we explore a parallel structure, namely MSU-Net, that can further speedup the segmentation. Fig. \[fig:msunet\] illustrates our MSU-Net which consists of multiple DownTubes (DTs), UpTubes (UTs), Center blocks, and a final evaluator (FE). The DTs and UTs act as the encoders and decoders in U-Net for feature propagation. Multiple DTs are built for a set of splitting patch sizes, each consisting of multiple blocks with downscaling convolution layers to perform feature downscaling and reuse. The ICA process and the corresponding mixing operations are done before and after the operations in DT, repsectively, and the operations in DT are performed in canonical form distributions similar to the work developed in SCNN [@wang2019scnn]. The features in UT are propagated and upscaled with the blocks made of convolutional layers, and transposed convolutional layers, respectively. The features after each upscaling are concatenated with the one skipped from DT for feature reuse. After the outputs are obtained from UTs with various patch sizes, all features would have the same dimensions, which are then concatenated and forwarded to the final evaluator to generate the final output. The number of blocks in DT/UT varies to accommodate the input dimensions of 3D images. Experiments =========== Experiment Setup ---------------- The evaluation task is to segment right ventricle (RV), myocardium (MYO), and left ventricle (LV) from MRI video clips in real time. We evaluate the proposed MSU-Net and competitive baselines on segmenting the RV, MYO and LV from the frames of End Diastolic (ED) and End Systolic (ES) instant. These frames were collected from the ACDC MICCAI 2017 challenge dataset [@bernard2018deep] with additional labeling done by experience radiologists. These frames have similar properties as 3D cardiac MRI videos. The dataset has 150 exams from different patients with 100 for training and 50 for testing. The images were collected following the common clinical SSFP cine-MRI sequence with a series of short-axis slices starting from the mitral valves down to the apex of the left ventricle. We perform 5-fold cross-validation and use the Dice score to evaluate the segmentation accuracy. We implement two versions of MSU-Net with specific snippet spans ($\text{t}=5$ and $10$, denoted as T5, T10, respectively) for evaluation. The ICA processing time is included when we evaluate the inference time of MSU-Net. Existing approaches have reported their FPS on the same dataset: $\sim1$ [@isensee2017automatic], and $5.56$[@zotti2018convolutional]. Clearly, none of them can perform real-time inference (i.e., at least 30 FPS). Therefore, we modify and rebuild these approaches to speed them up so that they can be compared with MSU-Net on a relatively fair basis. We implement a set of shallower/slimmer versions (i.e., with fewer layers/fewer channels) of the models. Specifically, we modify the 2D U-Net [@ronneberger2015u] to a shallow version with a depth of 3 and initial filter of 8 or 16. We denote them as D3+IF8 and D3+IF16, respectively. The 2D U-Net with vanilla configuration (D5+IF64) is also included. We also modify GridNet to shallower versions with a depth of 2 or 3 and initial filter of 32, denoted as D2+IF32 and D3+IF32, respectively. The vanilla version of GridNet is one of the best models in the ACDC 2017 challenge. All these methods are fully trained after modification. In our experiments, we do not include 3D U-Net for comparison due to its excessive memory consumption, unbalanced input dimensions, and slow inference speed [@ronneberger2015u; @isensee2017automatic]. Meanwhile we do not include lightweight networks such as ShuffleNet [@ma2018shufflenet] or MobileNet [@sandler2018mobilenetv2] which is designed with small memory footprint for mobile devices in image classification/object detection rather than medical segmentation, while inference speed is not their primary concern (which mainly depends on the network depth). We have tried ShuffleNet/MobileNet in our experiment settings and the speeds are only at 8.45/11.58 FPS, which are slower than the nets we reported. We implement MSU-Net and 2D U-Nets using PyTorch. The GridNet was implemented using TensorFlow [@zotti2018convolutional]. All experiments run on a machine with 16 cores of Intel Xeon E5-2620 v4 CPU, 256G memory, and an NVIDIA GeForce GTX 1080 GPU. ------------------- ---------- ------------------- ------------------- ------------------- ------------------- **RV** **MYO** **LV** **Average** GridNet (D3+IF32) 15.7 .842$\pm$.028 .804$\pm$.026 .901$\pm$.036 .849$\pm$.014 U-Net (D5+IF64) 16.1 **.865$\pm$.036** .761$\pm$.039 **.911$\pm$.026** .846$\pm$.025 GridNet (D2+IF32) 18.2 .815$\pm$.025 .812$\pm$.014 .851$\pm$.033 .826$\pm$.011 U-Net (D3+IF16) 33.2 .564$\pm$.071 .738$\pm$.045 .767$\pm$.026 .690$\pm$.036 U-Net (D3+IF8) 43.2 .552$\pm$0.079 .674$\pm$.060 .759$\pm$.059 .662$\pm$.058 MSU-Net (T5) **43.2** .855$\pm$.026 **.836$\pm$.022** .897$\pm$.017 **.862$\pm$.011** MSU-Net (T10) **70.2** .837$\pm$.034 .811$\pm$.049 .854$\pm$.040 .834$\pm$.020 ------------------- ---------- ------------------- ------------------- ------------------- ------------------- : Comparison between baseline methods and our MSU-Net on Dice score and FPS for 3D MRI video segmentation. “T5”/“T10” denotes the video snippet span in MSU-Net (t=5/t=10). “D” and “IF” denote the depth of the network and the initial filters number of the input layer, respectively. \[table:compare\] Results ------- Table \[table:compare\] presents the comparison among U-Net, GridNet, and the proposed MSU-Net on Dice score and FPS. Our MSU-Nets can achieve the fastest processing speed (highest FPS) and the best Dice score. Compared with the fastest baseline method U-Net (D3+IF8), our MSU-Net (T10) runs 1.63$\times$ faster and makes an improvement of 26% on segmentation accuracy. Compared with the most accurate baseline method (with the highest Dice score) GridNet (D3+IF32), our MSU-Net (T5) can achieve a slightly higher accuracy and 2.75$\times$ faster processing speed. From the table, it is clear that MSU-Nets are the only capable method to segment real-time 3D MRI videos. For MSU-Net, a bigger video snippet span (T10) can obtain a faster processing speed with a slight accuracy degradation (only 0.028). However, for U-Net, when it is modified into shallow/slim versions such as U-Net (D3+IF16) and (D3+IF8) for real-time processing ($\ge 30$ FPS), the accuracy degrades significantly: We observe that the accuracy drops from 0.846 to 0.690 and 0.662, respectively. We observe the same pattern for GridNet, and conclude that MSU-Net can achieve a stable accuracy when configured for segmentation in real time. Finally, Fig. \[fig:seg\_out\] shows the examples of MSU-Net segmentation results at various time steps. Note that our MSU-Net can accurately segment the target areas. The boundaries are clearly extracted on most of the slices. In the base and middle slices, the segmentation fits the contours of targets. In some of the apex slices, the segmentation of RV (labeled in blue) is not as accurate as MYO and LV, because of the unclear boundaries between the instances. ![The segmentation results of our method MSU-Net (T5) on the testing data. The rows indicate the slices at the base, the middle, and the apex of LV. The columns show the results at various time steps in series. RV, MYO, and LV are labeled in blue, green and red, respectively. []{data-label="fig:seg_out"}](fig-seg_outs.pdf){width="1\linewidth"} Conclusions =========== In this paper, we proposed Multiscale Statistical U-Net (MSU-Net) for real-time 3D cardiac MRI video segmentation. Based on the scheme of Statistical Convolutional Neural Network, we model the input samples as multiscale canonical form distributions for speedup, while the spatio-temporal correlationis still fully utilized. A parallel statistical U-Net is then proposed to process these multiscale distributions efficiently. On the 3D cardiac MRI videos from the ACDC MICCAI 2017 dataset, MSU-Net achieves up to 268% and 237% speedup with 1.6% and 3.6% increased Dice scores compared with vanilla U-Net and a modified state-of-the-art method GridNet, respectively. Acknowledgement =============== This work was approved by the Research Ethics Committee of Guangdong General Hospital, Gunagdong Academy of Medical Sciences with protocol No. 20140316. This work was supported by the National key Research and Development Program \[2018YFC1002600\], Science and Technology Planning Project of Guangdong Province, China \[No. 2017A070701013, 2017B090904034, 2017030314109, and 2019B020230003\], National Science Foundation grant \[CCF-1919167\], and Guangdong peak project \[DFJH201802\].
{ "pile_set_name": "ArXiv" }
--- abstract: 'Session types offer a type-based discipline for enforcing communication protocols in distributed programming. We have previously formalized simple session types in the setting of multi-threaded $\lambda$-calculus with linear types. In this work, we build upon our earlier work by presenting a form of dependent session types (of DML-style). The type system we formulate provides linearity and duality guarantees with no need for any runtime checks or special encodings. Our formulation of dependent session types is the first of its kind, and it is particularly suitable for practical implementation. As an example, we describe one implementation written in ATS that compiles to an Erlang/Elixir back-end.' author: - Hanwen Wu - Hongwei Xi bibliography: - './library.bib' title: Dependent Session Types --- Introduction {#sec:introduction} ============ A session is a sequence of interactions among concurrently running programs. We assign session types [@Honda:1993eh; @Honda:1998fm; @Takeuchi:1994bv; @Honda:2008hi] to communication channels to ensure session fidelity, which means each participant in the session communicates according to a chosen protocol. Recent works [@Wadler:2012ua; @Wadler:2014bta; @Caires:2010gi; @Carbone:2015hl; @Carbone:2016kd] have established a form of Curry-Howard correspondence where logical propositions are interpreted as session types for terms in variants of $\pi$-calculus [@Milner:1992fc; @Milner:1992gv]. Instead of $\pi$-calculus, it is also possible to formulate session types in the setting of $\lambda$-calculus [@Lindley:2015cr; @Xi:2016ue]. This paper formulates a form of dependent session types by extending our prior work [@Xi:2016ue]. More specifically, the formulation is based on Applied Type Systems ( [@Xi:2003kl; @Xi:2004vt]), a type system supporting dependent types (of DML-style [@Xi:1999bh]), linear types, and programming with theorem proving.  takes a layered approach to dependent types in which *statics*, where types are formed and reasoned about, are completely separate from *dynamics*, where programs are constructed and evaluated. Based on , session protocols are then captured by extending statics with session types (static terms of sort ${\text{\rm\it stype}}$), while communication channels are *linear* dynamic values whose types are *indexed* by such session protocols. When compared to other similar works (e.g. [@Toninho:2011in]), a very important difference of our formulation is that our session types describe the intended behavior *globally*, instead of using a polarized presentation where dual session types are used to describe dual endpoints of a channel *locally*. This is especially so when quantifiers are involved. Suppose that we want to provide an equality testing service, which receives two integers $m$ and $n$, and then sends out a boolean value indicating whether they are equal. Let us use *roles* 0 (server) and 1 (client), to refer to the two endpoints of a channel. We may use `S` for 0 and `C` for 1. We use `equal` for the following (static) term which describes the protocol for the equality testing service, $${\text{\rm\tt equal}}{\coloncolonequals}{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}})\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}})\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{\text{\rm\bf bool}})\coloncolon{{\text{\rm\tt end}}}({{\text{\rm\tt S}}})$$ We use ${{\text{\rm\tt msg}}}(r,\hat\tau)$ to mean that the endpoint $r$ (more precisely, the party holding endpoint $r$) is to send a (linear) value of type $\hat\tau$, and $\coloncolon$ for chaining, and ${{\text{\rm\tt end}}}(r)$ to mean that the endpoint $r$ is to initiate the termination of the session (while the other side waits for it). With dependent session types, `equal` can be given a more precise definition as follows, $$\begin{gathered} {\text{\rm\tt equal}}{\coloncolonequals}{{\text{\rm\tt quan}}}({{\text{\rm\tt C}}},\lambda m{:}{\text{\rm\it int}}.{{\text{\rm\tt quan}}}({{\text{\rm\tt C}}},\lambda n{:}{\text{\rm\it int}}.\\ {{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(m))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(n))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{\text{\rm\bf bool}}(m=n))\coloncolon{{\text{\rm\tt end}}}({{\text{\rm\tt S}}})))\end{gathered}$$ where ${{\text{\rm\tt quan}}}$ is a *global* encoding of quantifiers. For any role $r$, ${{\text{\rm\tt quan}}}(r,\cdot)$ means universal quantification at endpoint $r$, and dually, existential quantification at the other endpoint ($1-r$). In `equal`, `quan` means universal quantification at the client side, meaning the client process can send any integers onto the endpoint. Dually, `quan` refers to existential quantification at the server side, indicating that the server process can only send back a boolean value representing the equality of the two *received* integers. Note that ${\text{\rm\bf int}}$ and ${\text{\rm\bf bool}}$ are type constructors (static functions of c-sort ${\text{\rm\it int}}\Rightarrow{\text{\rm\it type}}$ and ${\text{\rm\it bool}}\Rightarrow{\text{\rm\it type}}$, respectively) while ${\text{\rm\it int}}$ and ${\text{\rm\it bool}}$ are *sorts* for static terms. Both ${\text{\rm\bf int}}(i)$ and ${\text{\rm\bf bool}}(b)$ are singleton types representing values that equal $i$ and $b$, respectively. In ATS, which uses ML-like syntax, an example program of the type ${{\text{\rm\bf chan}}}({{\text{\rm\tt S}}},{\text{\rm\tt equal}})\rightarrow{\text{\rm\bf 1}}$ that provides such service on the server side endpoint can be written as follows, [ eq\_test (ch:chan(S,equal)): void = ]{} [ () = exify ch ]{} [ () = exify ch ]{} [ m = recv ch ]{} [ n = recv ch ]{} [ () = send (ch, m = n)]{} [ close ch ]{} Let us use this code sample to introduce some key concepts. We use (linear) *channels* for communication. A channel consists of two *endpoints*. When one process sends a value onto one endpoint, the value is automatically transmitted to the other endpoint of the channel. `ch` is one such endpoint of the channel at party `S`, whose type is ${{\text{\rm\bf chan}}}({{\text{\rm\tt S}}},{\text{\rm\tt equal}})$. The *linear* type constructor, ${{\text{\rm\bf chan}}}$, will construct a linear type ${{\text{\rm\bf chan}}}(r,\pi)$ given a role $r$ and a global session type $\pi$. The combination of $r$ and $\pi$ is where a global session type gets “projected” locally. This can be used to type an endpoint of a channel at party $r$. As `equal` is globally quantified by session type constructor `quan`, we need to locally interpret it at party `S`, by calling a session API `exify` twice, which essentially turns ${{\text{\rm\bf chan}}}({{\text{\rm\tt S}}},{\text{\rm\tt equal}})$ into $$\exists m{:}{\text{\rm\it int}}.\exists n{:}{\text{\rm\it int}}. {{\text{\rm\bf chan}}}({{\text{\rm\tt S}}},{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(m))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(n))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{\text{\rm\bf bool}}(m=n))\coloncolon{{\text{\rm\tt end}}}({{\text{\rm\tt S}}}))$$ for use with other session API, e.g. `recv`. The *guard* in the signature of `exify` (see \[fig:sessionapi\]), $r\neq r_0$, specifies that, for any ${{\text{\rm\tt quan}}}(r_0,\cdot)$ at endpoint ${{\text{\rm\bf chan}}}(r,\cdot)$, only when $r\neq r_0$ is true that `exify` can be invoked to turn ${{\text{\rm\bf chan}}}(r,{{\text{\rm\tt quan}}}(r_0,\cdot))$ into $\exists a{:}\sigma.{{\text{\rm\bf chan}}}(r,\cdot)$. Dually, before the client can use the channel to send two integers, it has to locally interpret `quan` at party `C`, by calling `unify` (see \[fig:sessionapi\]) whose guard is $r=r_0$, which is the dual of `exify` since roles can only be 0 or 1 in a binary session. It will turn the endpoint at the client side into $$\forall m{:}{\text{\rm\it int}}.\forall n{:}{\text{\rm\it int}}. {{\text{\rm\bf chan}}}({{\text{\rm\tt C}}},{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(m))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(n))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{\text{\rm\bf bool}}(m=n))\coloncolon{{\text{\rm\tt end}}}({{\text{\rm\tt S}}}))$$ Essentially, a universally quantified endpoint *inputs* a static term from the user to eliminate the quantifier, while an existentially quantified endpoint *outputs* the witness to the user to eliminate the quantifier. Note that the user of an endpoint is the process holding such endpoint as mentioned above. So “inputs from the user” means the user writes a program to *send* a value using the endpoint. Such a twist is found in other works as well, e.g. [@Wadler:2012ua; @Wadler:2014bta]. The main contribution of this paper lies in the formulation of a form of dependent session types (of DML-style) in the setting of $\lambda$-calculus, which is the first of its kind. In particular, this formulation is based on unpolarized presentation. Our technical results include preservation and progress properties, which guarantee session fidelity and deadlock-freeness. We also mention at the end an implementation of our system that targets Erlang/Elixir. The rest of the paper is organised as follows. \[sec:mtlc-linear\] briefly sets up multi-threaded $\lambda$-calculus with linear types, denoted as ${\mathcal{L}_0}$. \[sec:mtlc-linear-dep\] introduces *predicatization* to extend ${\mathcal{L}_0}$ into multi-threaded $\lambda$-calculus with dependent types and linear types, denoted as ${\mathcal{L}_{\forall,\exists}}$. \[sec:session-types\] further extends ${\mathcal{L}_{\forall,\exists}}$ to formulate dependent session types as ${\mathcal{L}_{\forall,\exists}^{\pi}}$. \[sec:implementations\] describes technical details of our implementations. \[sec:examples\] demonstrates the benefits of dependent session types through examples. We then mention extensions (multi-party sessions, polymorphism, etc) in \[sec:ext\], related works in \[sec:related-works\] and finally conclude in \[sec:conclusion\]. Multi-threaded \\lambda-calculus with Linear Types {#sec:mtlc-linear} ================================================== The formulation of multi-threaded $\lambda$-calculus with linear types is largely standard and follows exactly from our previous work [@Xi:2016ue] except for some minor cosmetic changes. Therefore, we only present it very briefly and refer the readers to our prior work for details. Syntax ------ The syntax is shown in \[fig:mtlc0syntax\] which is mostly standard. $\delta$/$\hat\delta$ are non-linear/linear base types. “vtype” is just linear type. Note that a type $\tau$ is also a linear type $\hat\tau$, but it is not regarded as a *true* linear type. ${\textit{dcc}}$/${\textit{dcf}}$ are dynamic constant constructors/functions (pre-defined constructors/functions). ${\textit{dcr}}$ are dynamic constant resources that are treated *linearly*. ${\mathcal{S}}$ are dynamic signatures that assign types to dynamic constants, and these types are called *c-types*. Note that $\vv{\cdot}$ stands for a possibly empty sequence of $\cdot$, i.e. $\vv{e}$ is a possibly empty sequence of dynamic terms. ${\textit{dcx}}\,(\vv{e})$ is a term of type $\tau$ if *dcx* is a constant of c-type $(\tau_1,\dotsc,\tau_n)\Rightarrow\tau$ in ${\mathcal{S}}$ and for each $e_i~(1 \leq i \leq n)$ in $\vv{e}$, $e_i$ has type $\tau_i$. We use $[]$ for the empty mapping and $[a_1,\dotsc,a_n \mapsto b_1,\dotsc,b_n]$ for a mapping that maps $a_i$ to $b_i$ for $1\leq i\leq n$, in which case we write $m(a_i)$ to mean $b_i$. We use ${\text{\rm\bf dom}}(m)$ for the domain of a mapping $m$. If $a \notin {\text{\rm\bf dom}}(m)$, then $m[a \mapsto b]$ means to extend $m$ with a new link from $a$ to $b$. We also use $m\backslash a$ to mean the mapping obtained by removing $a$ from ${\text{\rm\bf dom}}(m)$, and $m[a\colonequals b]$ to mean $(m\backslash a)[a \mapsto b]$. Substitution $\theta$ is a mapping from variables to dynamic values. We write $e[\theta]$ for the result of applying $\theta$ to $e$. Pool $\Pi$ is a mapping from thread identifiers $t$ (represented as natural numbers) to closed dynamic expressions such that $0 \in {\text{\rm\bf dom}}(\Pi)$. We use $\Pi(t), t \in {\text{\rm\bf dom}}(\Pi)$ to refer to a thread in $\Pi$ whose thread identifier is $t$. We use $\Pi(0)$ for the main thread. Typing contexts are divided into a non-linear part $\Gamma$ and a linear part $\Delta$. They are intuitionistic meaning that it is required that each variable occurs at most once in a non-linear context $\Gamma$ or a linear context $\Delta$. Given $\Gamma_1, \Gamma_2$ s.t. ${\text{\rm\bf dom}}(\Gamma_1)\cap{\text{\rm\bf dom}}(\Gamma_2)=\varnothing$, we write $(\Gamma_1, \Gamma_2)$ for the union of the two. The same notion also applies to linear context $\Delta$. Given non-linear context $\Gamma$ and linear context $\Delta$, we can form a combined context $(\Gamma;\Delta)$ when ${\text{\rm\bf dom}}(\Gamma)\cap{\text{\rm\bf dom}}(\Delta)=\varnothing$. Given $(\Gamma;\Delta)$, we may write $(\Gamma;\Delta),x:\hat\tau$ for either $(\Gamma;\Delta,x:\hat\tau)$ or $(\Gamma,x:\hat\tau;\Delta)$ if $\hat\tau$ is indeed a non-linear type. Besides integers and booleans, we also assume a constant function `thread_create` in *dcx* whose c-type in ${\mathcal{S}}$ is $({\text{\rm\bf 1}}\multimap{\text{\rm\bf 1}})\Rightarrow{\text{\rm\bf 1}}$. A function of type ${\text{\rm\bf 1}}\multimap{\text{\rm\bf 1}}$ takes no argument and returns no result (if it terminates). Since it is a true linear function, it can be invoked exactly once. Intuitively, `thread_create` creates a thread that evaluates the linear function. Its semantic is to be formally introduced later. To manage resources, we follow [@Xi:2016ue] and define $\rho(\cdot)$ (\[fig:rho\]) to compute the multiset (bag) of constant resources in a given expression and $\res$ (**RES** in [@Xi:2016ue]) to range over such multisets of resources. We say $R$ is valid if $R\in\res$ holds. Intuitively, $\res$ can be thought as all the resources of all the programs and $R$ the resources of a single program. We need to make sure that resource allocation to different programs is consistent in $\res$. For precise definitions, please refer to our prior work. Sementics --------- Typing rules are the same as [@Xi:2016ue], and we push it to \[fig:mtlc0typing\] in the appendix. The c-type judgment based on the signature is of the form ${\mathcal{S}}\vDash e:\hat\tau$. A typing judgment is of the form $\Gamma;\Delta\vdash e:\hat\tau$ which is standard. By inspecting the rules in \[fig:mtlc0typing\], we can readily see that a closed value cannot contain resources if it can be assigned a non-linear type $\tau$. The *Lemma of Canonical Forms* and the *Lemma of Substitution* are the same as our previous work ([@Xi:2016ue] Lemma 2.2 and Lemma 2.3), we thus omit them completely. ${\mathcal{L}_0}$ has a call-by-value semantic, and the definition of evaluation context ($E$), redex, and reducts are completely standard and are the same as our previous work. We thus omit the details and present just reduction on pools and properties of ${\mathcal{L}_0}$. Given pools $\Pi_1,\Pi_2$, we define *reductions on pools* $\Pi_1\rightarrow\Pi_2$ as follows, $$\begin{gathered} \prftree[r]{\bf pr0} {e_1\rightarrow e_2} {\Pi[t\mapsto e_1]\rightarrow\Pi[t\mapsto e_2]} \quad \prftree[r]{\bf pr2} {t > 0} {\Pi[t\mapsto\langle\rangle]\rightarrow\Pi} \\ \prftree[r]{\bf pr1} {\Pi(t)=E[\texttt{thread\_create}({\text{\rm\bf lam}}~x.e)]} {\Pi\rightarrow\Pi[t\colonequals E[\langle\rangle]][t'\mapsto{\text{\rm\bf app}}({\text{\rm\bf lam}}~x.e, \langle\rangle)]}\end{gathered}$$ \[thm:subjectreduction\] Assume $\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable and $\Pi_1\rightarrow\Pi_2$ holds for some $\Pi_2$ satisfying $\rho(\Pi_2)\in\res$. Then $\varnothing;\varnothing\vdash\Pi_2:\hat\tau$ is also derivable. \[thm:progress\] Assume that $\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable. Then we have - $\Pi_1$ is a singleton mapping $[0\mapsto v]$ for some value $v$, or - $\Pi_1\rightarrow\Pi_2$ holds for some $\Pi_2$ s.t. $\rho(\Pi_2)\in\res$. \[thm:soundness\] Assume that $\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable. Then for any $\Pi_2$, $\Pi_1\rightarrow^* \Pi_2$ implies that either $\Pi_2$ is a singleton mapping $[0\mapsto v]$ for some value $v$ or $\Pi_2\rightarrow\Pi_3$ for some $\Pi_3$ satisfying $\rho(\Pi_3)\in\res$, where $\rightarrow^*$ is the transitive and reflective closure of $\rightarrow$. Follows directly from \[thm:subjectreduction\] and \[thm:progress\]. Predicatization {#sec:mtlc-linear-dep} =============== In this section, we extremely briefly describe an approach to extend ${\mathcal{L}_0}$ to support both universally and existentially quantified types. Such process is *predicatization* and is mostly standard in the framework of  [@Xi:2003kl]. Predicatization is extensively described in [@Xi:1998wa; @Xi:1999bh; @Xi:2007te], and has been employed in several other papers based on , e.g. [@Shi:2013hr; @Shi:2008vp]. We thus only summarize the process to prepare for the development of ${\mathcal{L}_{\forall,\exists}^{\pi}}$, and omit any technical details. As an applied type system, ${\mathcal{L}_{\forall,\exists}}$ is layered into *statics* and *dynamics*. The dynamics of ${\mathcal{L}_{\forall,\exists}}$ is based on ${\mathcal{L}_0}$, while the statics will be a newly introduced layer underlying ${\mathcal{L}_0}$. The predicatization process concerns mostly about formalizing the type index language while maintaining the dynamic semantics of ${\mathcal{L}_0}$, and reducing type equality problems into constraint solving problems w.r.t. some constraint domain, such as integer arithmetic. General steps of predicatization involve the followings: - Formalizing statics, the language of type index. This involves its syntax, sorting rules, and specifically, non-linear type/linear type formation rules, etc. - Formalizing type equality in terms of subtyping relations and regular constraint relations. - Extending dynamics. This involves extending the syntax, typings, evaluation context, and reduction relations to accommodate, for instance, the introduction and elimination of quantifiers. The language of statics can be regarded as a simply typed $\lambda$-calculus. The “types” for static terms are denoted as *sorts* to avoid confusion. The syntax for statics is shown in \[fig:statics\] which is mostly standard. We assume base sorts $b$ to include ${\text{\rm\it int}}$, ${\text{\rm\it bool}}$, ${\text{\rm\it type}}$ for types, and ${\text{\rm\it vtype}}$ for linear types. Non-linear/linear types in the ${\mathcal{L}_{\forall,\exists}}$ are now static terms of sorts ${\text{\rm\it type}}$/${\text{\rm\it vtype}}$, respectively. We reformulate types in the dynamics in \[fig:types\]. Given a proposition $P$ (a static term of sort ${\text{\rm\it bool}}$) and a type $\tau$, $P{\,{\supset}\,}\tau$ is a *guarded type*, and $P{{\wedge}}\tau$ is an *asserting type*. Formal definition of guarded types and asserting types can be found in [@Xi:2007te]. Intuitively, in order to turn a value of type $P{\,{\supset}\,}\tau$ into a value of type $\tau$, we must establish the proposition $P$, thus “guarded”; if a value of type $P{{\wedge}}\tau$ is generated, we can assume that the proposition $P$ holds, thus “asserting”. The extended syntax of ${\mathcal{L}_{\forall,\exists}}$ over that of ${\mathcal{L}_0}$ is given in \[fig:mtlcdepsyntax\]. Typing judgement in ${\mathcal{L}_{\forall,\exists}}$ is of the form $\Sigma;\vv{P};\Gamma;\Delta\vdash e:\hat\tau$ where $\Sigma$ is the sorting environment for static terms and $\vv{P}$ is a sequence of propositions keeping track of the constraints. We present only some additional typing rules in \[fig:mtlcdeptyping\]. We claim that \[thm:subjectreduction\], \[thm:progress\], and \[thm:soundness\] can be carrier over to ${\mathcal{L}_{\forall,\exists}}$ following the proof in [@Xi:2007te]. Dependent Session Types {#sec:session-types} ======================= Dependent types are types that depend on terms, and they offer much more expressive power for specifying intended behavior of a program through types. A restricted form of dependent types, we call dependent types of DML-style [@Xi:2007te], are types that depend on *static* terms. In this section, we will formally develop *dependent session types* (of DML-style), where session types can have quantification over static terms. Based on ${\mathcal{L}_{\forall,\exists}}$, we first extend the statics, then extend the dynamics, and finally discuss the soundness of ${\mathcal{L}_{\forall,\exists}^{\pi}}$. Extending Statics ----------------- The syntax of extended statics is given in \[fig:sessionsyntax\]. We add ${\text{\rm\it stype}}$ as a new base sort to represnet session types. Session types $\pi$ are now static terms of sort ${\text{\rm\it stype}}$. We use $i$ for static integers and $b$ for static booleans. ${{\text{\rm\tt end}}}(i)$ means party $i$ (the party holding endpoint $i$) will close the session while the other party will wait for closing. Given linear type $\hat\tau$ and a session type $\pi$, ${{\text{\rm\tt msg}}}(i,\hat\tau)\coloncolon\pi$ means party $i$ should send a message to the other party, and then continue as $\pi$. ${{\text{\rm\tt branch}}}(i,\pi_1,\pi_2)$ is for branching, where party $i$ should choose to continue as $\pi_1$ or $\pi_2$ while the other party simply follows the choice. Beyond these basic session type constructs, we have ${{\text{\rm\tt ite}}}$[^1] for conditional branch, ${{\text{\rm\tt quan}}}$ for universal/existential quantification, and ${{\text{\rm\tt fix}}}$ for recursions. Given a static boolean expression, ${{\text{\rm\tt ite}}}(b, \pi_1,\pi_2)$ represents $\pi_1$ when $b$ is $\top$ (true), or $\pi_2$ when $b$ is $\bot$ (false). Given a static function of sort $\sigma\rightarrow{\text{\rm\it stype}}$, ${{\text{\rm\tt quan}}}(i,\lambda a{:}\sigma.\pi)$ is interpreted *intuitively*[^2] as universally quantified $\forall a{:}\sigma.\pi$ by party $i$, or as existentially quantified $\exists a{:}\sigma.\pi$ by the other party. Note that this is actually a session type *scheme* and we assume the existance of such ${{\text{\rm\tt quan}}}$ for every sort $\sigma$. The need for a unified representation of quantifiers, `quan`, is a must since we essentially formulate all session types as *global*, as compared to polarized presentation where session types are all *local*. Given a static function of sort ${\text{\rm\it stype}}\rightarrow{\text{\rm\it stype}}$, ${{\text{\rm\tt fix}}}(\lambda a{:}{\text{\rm\it stype}}.\pi)$ is an encoding of the fixpoint operator that represents the fixpoint of the input function. In practice, we may write recursive definitions directly as a syntax sugar (as shown in \[ex:array\]). Besides, we also introduce *role* as a subset sort $\{r{:}{\text{\rm\it int}}\mid r=0\lor r=1\}$ to represent two parties, server (0) and client (1), involved in a binary session. Note that subset sorts are merely syntax sugars for a guarded/asserting type [@Xi:1999bh]. For instance, $\forall r{:}{\text{\rm\it role}}.{\text{\rm\bf int}}(r)$ is desugared into $\forall r{:}{\text{\rm\it int}}.(r=0\lor r=1){\,{\supset}\,}{\text{\rm\bf int}}(r)$. We also add the following *linear* type constructor as a static constant[^3], $${{\text{\rm\bf chan}}}: ({\text{\rm\it role}},{\text{\rm\it stype}})\Rightarrow{\text{\rm\it vtype}}$$ that represents a linear channel. Given role $r$ and session type $\pi$, ${{\text{\rm\bf chan}}}(r,\pi)$ is endpoint $r$ of a channel held by a party. The channel is governed by the session type $\pi$, and the endpoint interprets this session type *locally* as role $r$. Extending Dynamics ------------------ We add the following dynamic constant functions (pre-defined functions), shown in \[fig:sessionapi\], to create, use, and consume linear channels. We will refer to them as *session API* or just the API. We break up the figure and present them with explanations here. $$\texttt{create} : \forall r_1,r_2{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.(r_1\neq r_2){\,{\supset}\,}({\text{\rm\bf chan}}(r_2,\pi)\multimap{\text{\rm\bf 1}})\Rightarrow{\text{\rm\bf chan}}(r_1,\pi)$$ `create` is to create a session of two threads, connected via a channel of session type $\pi$, and each thread holds an endpoint of the channel. One party is holding endpoint $r_1$ of type ${{\text{\rm\bf chan}}}(r_1,\pi)$ as returned by `create` in the current thread, while the other party is holding endpoint $r_2(\neq r_1)$ of type ${{\text{\rm\bf chan}}}(r_2,\pi)$ in a newly spawned thread evaluating the given *linear* function of type ${{\text{\rm\bf chan}}}(r_2,\pi)\multimap{\text{\rm\bf 1}}$. As the (closure) function may contains resources, it must be linear to guarantee that it can be called *exactly once*. The channel endpoint will be consumed in this function as it is linear. $$\begin{aligned} \texttt{send} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.\forall \hat\tau{:}{\text{\rm\it vtype}}.(r=r_0){\,{\supset}\,}({{\text{\rm\bf chan}}}(r,{{\text{\rm\tt msg}}}(r_0,\hat\tau)\coloncolon\pi), \hat\tau)\Rightarrow{{\text{\rm\bf chan}}}(r,\pi) \\ \texttt{recv} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.\forall \hat\tau{:}{\text{\rm\it vtype}}.(r\neq r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt msg}}}(r_0,\hat\tau)\coloncolon\pi)\Rightarrow \hat\tau\otimes{{\text{\rm\bf chan}}}(r,\pi) \end{aligned}$$ `send` is for sending linear values. Given global session type ${{\text{\rm\tt msg}}}(r_0,\hat\tau)\coloncolon\pi$, its interpretation at $r$ where $r=r_0$ is to send a message of linear type $\hat\tau$ then to proceed as $\pi$. The `send` function *consumes* the channel, uses the capability of sending denoted by ${{\text{\rm\tt msg}}}(r_0,\hat\tau)$, and returns another channel of type ${{\text{\rm\bf chan}}}(r,\pi)$, where the sending capability is now removed. Dually, the interpretation of ${{\text{\rm\tt msg}}}(r_0,\hat\tau)\coloncolon\pi$ is to receive at party $r(\neq r_0)$, implemented by `recv`. Note that even though we encode it here in the style of continuation, our implementation directly *changes* the type of channel without consuming it. In ATS programming language, it is presented in the following style, $$\begin{gathered} \texttt{send}:\forall r,r_0{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.\forall \hat\tau{:}{\text{\rm\it vtype}}.\\ (r=r_0){\,{\supset}\,}(!{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt msg}}}(r_0,\hat\tau)\coloncolon\pi)\gg{{\text{\rm\bf chan}}}(r,\pi), \hat\tau)\Rightarrow{\text{\rm\bf 1}}\end{gathered}$$ Similarly, `close` is for terminating a session while `wait` is waiting for the other side to close. $$\begin{aligned} \texttt{close} &: \forall r,r_0{:}{\text{\rm\it role}}.(r=r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{\text{\rm\tt end}}(r_0))\Rightarrow{\text{\rm\bf 1}} \\ \texttt{wait} &: \forall r,r_0{:}{\text{\rm\it role}}.(r\neq r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{\text{\rm\tt end}}(r_0))\Rightarrow{\text{\rm\bf 1}} \end{aligned}$$ The interpretation of ${{\text{\rm\tt branch}}}(r_0,\pi_1,\pi_2)$ at party $r(\neq r_0)$ is to offer two choices, $\pi_1$ and $\pi_2$. Therefore, `offer` function will consume the endpoint and return a linear pair of the other party’s choice (as a singleton boolean) and the endpoint whose session type is a conditional branch between $\pi_1,\pi_2$ using the received tag $b$ as the condition. Dually, `choose` will choose $\pi_1$ and $\pi_2$ respectively according to the boolean tag provided by the user. Note that these two functions are completely unnecessary since they can be encoded using other functions/session types. We present them here just to stay inline with others where `offer/choose` are usually treated as standard constructs. $$\begin{aligned} \texttt{offer} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall \pi_1,\pi_2{:}{\text{\rm\it stype}}.(r\neq r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{\text{\rm\tt branch}}(r_0,\pi_1,\pi_2)) \\ &\qquad\qquad\Rightarrow\exists b{:}{\text{\rm\it bool}}.{\text{\rm\bf bool}}(b)\otimes{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt ite}}}(b,\pi_1,\pi_2)) \\ \texttt{choose} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall \pi_1,\pi_2{:}{\text{\rm\it stype}}.\forall b{:}{\text{\rm\it bool}}.(r=r_0){\,{\supset}\,}({{\text{\rm\bf chan}}}(r,{\text{\rm\tt branch}}(r_0,\pi_1,\pi_2)), {\text{\rm\bf bool}}(b)) \\ &\qquad\qquad\Rightarrow {{\text{\rm\bf chan}}}(r,{{\text{\rm\tt ite}}}(b,\pi_1,\pi_2)) \end{aligned}$$ `unify` is to interpret ${{\text{\rm\tt quan}}}(r_0,\cdot)$ at party $r(= r_0)$ as universal quantifier, while `exify` is to interpret it dually as existential quantifier at party $r(\neq r_0)$. $$\begin{aligned} \texttt{unify} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.\forall f{:}\sigma\rightarrow{\text{\rm\it stype}}. \\&\qquad\qquad (r= r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt quan}}}(r_0,f))\Rightarrow\forall s{:}\sigma.{{\text{\rm\bf chan}}}(r,f(s)) \\ \texttt{exify} &: \forall r,r_0{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.\forall f{:}\sigma\rightarrow{\text{\rm\it stype}}. \\&\qquad\qquad (r\neq r_0){\,{\supset}\,}{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt quan}}}(r_0,f))\Rightarrow\exists s{:}\sigma.{{\text{\rm\bf chan}}}(r,f(s)) \end{aligned}$$ `itet` and `itef` reduces the conditional branching session type ${{\text{\rm\tt ite}}}(b,\pi_1,\pi_2)$ according to static boolean expression $b$. `recurse` unrolls the fixpoint encoding. $$\begin{aligned} \texttt{itet} &: \forall r{:}{\text{\rm\it role}}.\forall \pi_1,\pi_2{:}{\text{\rm\it stype}}.{{\text{\rm\bf chan}}}(r,{\text{\rm\tt ite}}(\top,\pi_1,\pi_2))\Rightarrow{{\text{\rm\bf chan}}}(r,\pi_1)\\ \texttt{itef} &: \forall r{:}{\text{\rm\it role}}.\forall \pi_1,\pi_2{:}{\text{\rm\it stype}}.{{\text{\rm\bf chan}}}(r,{\text{\rm\tt ite}}(\bot,\pi_1,\pi_2))\Rightarrow{{\text{\rm\bf chan}}}(r,\pi_2)\\ \texttt{recurse} &: \forall r{:}{\text{\rm\it role}}.\forall f{:}{\text{\rm\it stype}}\rightarrow{\text{\rm\it stype}}.{{\text{\rm\bf chan}}}(r,{{\text{\rm\tt fix}}}(f))\Rightarrow{{\text{\rm\bf chan}}}(r,f({{\text{\rm\tt fix}}}(f))) \end{aligned}$$ Note that these functions (`unify`/`exify`/`itet`/`itef`/`recurse`) are *proof* functions that merely change the types of endpoints. They have no runtime counterparts and thus can be eliminated after type checking has passed. *Duality* is not explicitly encoded as is usually done in session types literature [@Lindley:2016du; @Pucella:2008dt; @Jespersen:2015ka]. Instead, we choose to make the duality as general as possible and use a *global* session type $\pi$ paired with a role $r$ to guide the local interpretation at endpoint $r$. Given that $r$ can only be $0$ or $1$, we can define that ${{\text{\rm\bf chan}}}(0,\pi)$ and ${{\text{\rm\bf chan}}}(1,\pi)$ are *dual* endpoints of a channel. Session API come in dual pairs, and the dual usage of dual endpoints are realized by the corresponding session API pairs with the help of guarded types. The typing rules for guarded types will force one endpoint to be only used with one API in the pair while the dual endpoint to be only used with the dual API in the same pair. A crucial indication of such formulation is that we essentially reduce the duality checking problem into a simple integer comparison problem, which greatly simplifies our formulation. Also, it reduces the number of the dynamic constants in \[fig:sessionapi\] in half by avoiding coercion between so-called input/output types [@Lindley:2016du]. In our previous work [@Xi:2016ue], we used a polarized presentation, e.g. ${\text{\rm\bf chanpos}}(p)$ and ${\text{\rm\bf channeg}}(p)$ where $p$ is a *local* type. This is similar to `In[]`/`Out[]` in [@Scalas:2016uh], $S_?$/$S_!$ in [@Lindley:2016du] Section 6, and *dual*/*notDual* in [@Sackman:2008ux]. We found this polarized presentation is not suitable for extending to multi-party sessions, whereas our “global+role+guard” formulation can be very easily adapted to multi-party sessions based on [@Xi:2017wv]. For example, in a three-party session, we can define ${{\text{\rm\bf chan}}}(0,\pi)$, ${{\text{\rm\bf chan}}}(1,\pi)$, and ${{\text{\rm\bf chan}}}(2,\pi)$ to be *compatible*, as a generalization to duality. We very briefly mention such extension in \[sec:ext\]. $$\texttt{cut} : \forall r_1,r_2{:}{\text{\rm\it role}}.\forall\pi{:}{\text{\rm\it stype}}.(r_1\neq r_2){\,{\supset}\,}({{\text{\rm\bf chan}}}(r_1,\pi),{{\text{\rm\bf chan}}}(r_2,\pi))\Rightarrow{\text{\rm\bf 1}}$$ Given *dual* endpoints, `cut` will link together the endpoints by performing *bi-directional forwarding*. In other words, it will send onto one endpoint each received value from the other endpoint. `cut` is often used to implement delegation of service. It can be proven that these two endpoints must belong to *different* channels since otherwise, it will obviously deadlock. We will explain more in \[sec:implementations\]. Dynamic Semantics ----------------- The dynamic semantics of ${\mathcal{L}_{\forall,\exists}^{\pi}}$ is indeed the same as our prior work except that we have added a branching construct and we use a more general unpolarized presentation. We thus push additional reduction ruls on pools in \[fig:sessionreduction\] and \[fig:sessionreductioncut\] to the appendix. Note that, as mentioned above, `unify`/`exify`/`itet`/`itef`/`recurse` do not have any dynamic semantics. The meaning of these rules should be intuitively clear. For instance, **pr-msg** states, if thread $t_1$ in pool $\Pi$ is of the form $E[\texttt{send}(ch_{i,r_1},v)]$, and thread $t_2$ in pool $\Pi$ is of the form $E[\texttt{recv}(ch_{i,r_2})]$, then $\Pi$ can be reduced to another pool where $t_1$ is replaced by $E[ch_{i,r_1}]$ and $t_2$ is replaced by $E[\langle v,ch_{i,r_2}\rangle]$. Soundness of the Type System ---------------------------- While \[thm:subjectreduction\] can be easily established for ${\mathcal{L}_{\forall,\exists}^{\pi}}$, \[thm:progress\] is more involved due to the addition of session API. However, based on [@Xi:2003kl; @Xi:1999bh], ${\mathcal{L}_{\forall,\exists}}$ and ${\mathcal{L}_{\forall,\exists}^{\pi}}$ are *conservative* extensions of ${\mathcal{L}_0}$, and the deadlock-freeness is proven for ${\mathcal{L}_0}$ with channels in [@Xi:2016ue] using a technique known as *DF-Reducibility*. Thus the same results can be proven for ${\mathcal{L}_{\forall,\exists}^{\pi}}$ using the exact same technique since the dynamic semantics are the same. We thus refer readers to [@Xi:2016ue; @Xi:1999bh] for detailed proofs. We can then establish the same deadlock-freeness guarantee as stated in Lemma 3.1 of [@Xi:2016ue] Assume that $\varnothing;\varnothing;\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable and $\Pi_1\rightarrow\Pi_2$ s.t. $\rho(\Pi_2)\in\res$. Then $\varnothing;\varnothing;\varnothing;\varnothing\vdash\Pi_2:\hat\tau$ is also derivable. Assume that $\varnothing;\varnothing;\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable and $\rho(v)$ contains no channel endpoins for every $v:\hat\tau$. Then - $\Pi_1$ is a singleton mapping $[0\mapsto v]$ for some $v$, or - $\Pi_1\rightarrow\Pi_2$ holds for some $\Pi_2$ s.t. $\rho(\Pi_2)\in\res$. Assume that $\varnothing;\varnothing;\varnothing;\varnothing\vdash\Pi_1:\hat\tau$ is derivable and $\rho(v)$ contains no channel endpoins for every $v:\hat\tau$. Then for any $\Pi_2$ satisfying $\rho(\Pi_2)\in\res$, $\Pi_1\rightarrow^*\Pi_2$ implies either $\Pi_2$ is a singleton mapping $[0\mapsto v]$ for some $v$, or $\Pi_2\rightarrow\Pi_3$ for some $\Pi_3$ s.t. $\rho(\Pi_3)\in\res$. Implementations {#sec:implementations} =============== Our implementations consist of two parts, a session API library in ATS, and a runtime implementation of the session API (referred to as a *back-end*) in a target language. ATS is a programming language based on , and it supports a style of *co-programming* with many target languages by compiling an ATS program into the target language. Its default compilation target is C. For the purpose of this paper, besides a native back-end in ATS/C itself, we also support back-ends in Erlang/Elixir and JavaScript. A session-typed program will be firstly type-checked based on the type system of ${\mathcal{L}_{\forall,\exists}^{\pi}}$, and then compiled into a target language (if passed type checking). The compiler/interpreter of the target language will then be invoked to compile/interpret the program together with the corresponding back-end. Although formalized as synchronous sessions (for the sake of simplicity), our implementations can fully support asynchronous communications. Our linear typing guarantees *no resources leaks*. For instance, in our Erlang/Elixir back-end, there are no process leaks related to channels. Our session API library in ATS is (almost) a direct translation of those listed in \[fig:sessionapi\], except for some slight syntax differences. For example, `send` is translated into the followings. [ send ]{} [ (!chan(r,msg(r0,v)::p) &gt;&gt; chan(r,p), v): void]{} where `{}` is universal quantification (and `[]` is existential quantification), `!` means call-by-value, which indicates *not* to consume a linear value, and `>>` means to *change* the linear type after the function returns. As mentioned before, whenever possible, the API will change the types of endpoints directly instead of relying on continuations. There are a couple other minor changes. First, with guarded recursive data types [@Xi:2003dp] and pattern matching, the API formulates `offer`/`choose` in a simpler way as follows, [ choice (stype, stype, stype) = ]{} [| {p,q:stype} Left (p, p, q) ]{}[ ()]{} [| {p,q:stype} Right (q, p, q) ]{}[ ()]{} [ offer ]{} [ (!chan(r,branch(r0,p,q)) &gt;&gt; chan(r,s)): ]{}[ choice (s,p,q)]{} [ choose ]{} [ (!chan(r,branch(r0,p,q)) &gt;&gt; chan(r,s), choice(s,p,q)): void]{} where `choice` is a guarded recursive data type that essentially captures the equality on session types. Also, since it is existentially quantified, the type-checker will enforce *exhuastive* case analysis on the received choice to instantiate `s`. Note that `s` as in `>> chan(r,s)` is in the scope of quantifier `[s:stype]` even though it appears before the quantifier. We briefly mention some technical details below and refer the readers to <http://multirolelogic.org> for pointers to all the source code. Due to space limitation, we assume that the readers are reasonably familiar with these target languages. Message-passing Back-end in Erlang/Elixir ----------------------------------------- Erlang offers functional distributed programming abilities through its powerful virtual machine. Elixir offers a more friendly syntax and better tooling on top of the same runtime. In Erlang/Elixir, every process has a unique `pid` (process identifier), and an associated mailbox. Communications are achieved via message-passing asynchronously and can be done across different nodes. In this particular implementation, `choose` and `offer` are implemented as `send` and `receive`, respectively. `close` and `wait` are implemented both to terminate the process directly. This back-end relies on order-preserving messages and is inherently asynchronous and distributed. In Erlang/Elixir back-end, a message is represented by a label, a `pid`, a `ref`, and a payload. A channel endpoint is identified through a combination of a `pid` and a `ref`. The message labels are used to identify the kind of messages, e.g. `:send`/`:receive`. The `pid` is used to locate the message’s origin, or an endpoint’s mailbox. The `ref`’s are globally unique references, generated through a built-in function `make_ref` for every endpoint. The need for `ref` is discussed in [@Mostrous:2011vu]. Intuitively speaking, the `ref` acts as a signature of the message and every out-going message is signed using the sending endpoint’s own `ref`. Thus it can be used both to distinguish in-session messages from out-of-session messages[^4], and to identify requests from the endpoint’s owning process and messages from the dual endpoint. An endpoint will run a loop in a dedicated process and talk to the owning process through messages-passing. The endpoint loop keeps track of two parameters: `self`, which is its own signature as a `ref`, and `dual`, which is the dual endpoint’s `pid` and `ref`. In every iteration, the loop will receive a request from the owning process by pattern matching against messages signed by `self`, and then process the request accordingly. For instance, when the owning process sends a message with label `:receive` signed with `self`, the endpoint will then pattern match against messages in the endpoint’s mailbox and block until it finds the first message whose label is `:send` and is signed by the dual endpoint’s `ref`, which is `dual.ref`. The found message will then be delivered to the owning process’s mailbox, fulfilling the request. `cut` is implemented as delegation, where `:send` requests are handled as before, but `:receive` requests are delegated to an endpoint involved in a `cut`. Suppose we have dual endpoints `A:chan(0,p)`/`A:chan(1,p)` and dual endpoints `B:chan(0,p)`/`B:chan(1,p)` of some session type `p`, and we are to perform `cut(A,B)`. The owning process $P_2$ of both `A` and `B`, will send a `:cut` request to `A` and `B`, with a payload of the `pid` and `ref` of `B` and `A`, respectively. The info about `B` will be forwarded to `A`, and `A` will delegate `:receive` requests to `B`. Similarly, the info about `A` will be forwarded to `B`. and `B` will delegate `:receive` requests to `A`. A delegated request will change its signature from the original requester’s `ref`, to the delegator’s `ref`, so that the delegator can still process the request as if the request comes from its owning process. An example is illustrated in \[fig:elixir\], where $\leftrightarrow$ is for endpoint ownership, $\Leftrightarrow$ connects dual endpoints, and dashed arrow denotes delegation. Now, if $P_1$ sends a message to $P_3$, it will be sent through endpoint `A`, and then delivered to the mailbox of `A`. When $P_3$ tries to receive the message, it will send a `:receive` request to `B`, and `B` delegates it to `A`, and `A` will fulfill the request since the message is in its mailbox. We also have a shared memory implementation in ATS/C which implements our own message queue guarded by locks, and a continuation-based implementation in JavaScript using WebWorker. Examples {#sec:examples} ======== We will show some example dependent session types or programs in the followings. We will assume that the server plays role 0 (`S`), and the client plays role 1 (`C`). We will use ATS’s ML-like syntax to present the program (after omitting some insignificant details), which can be easily mapped to ${\mathcal{L}_{\forall,\exists}^{\pi}}$. We also use syntax sugar and implementation optimizations described in \[sec:implementations\] and extensions from \[sec:ext\]. Again, the source code can be found online through <http://multirolelogic.org>, and all the code can be type-checked, compiled, and executed. One can easily define a counter as an integer stream. But more precisely, we can define dependently session typed constructor as $${\text{\rm\tt counter}} (n{:}{\text{\rm\it int}}) {\coloncolonequals}{{\text{\rm\tt branch}}}({{\text{\rm\tt C}}}, {{\text{\rm\tt msg}}}({{\text{\rm\tt S}}}, {\text{\rm\bf int}}(n)) \coloncolon {\text{\rm\tt counter}}(n+1),{{\text{\rm\tt end}}}({{\text{\rm\tt C}}}))$$ which says, in every iteration, the client can choose to receive an integer $n$ and let the session continue from $n+1$, or to end the session. `counter` makes use of higher-order fixpoint encoding, `fix`, which is better explained in \[ex:array\]. On top of `counter`, we can define a service `from` that given an integer $n$, returns an endpoint of session type `counter`$(n)$. $${\text{\rm\tt from}} {\coloncolonequals}{{\text{\rm\tt quan}}}({{\text{\rm\tt C}}}, \lambda n{:}int.{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{\text{\rm\bf int}}(n))\coloncolon{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{\text{\rm\bf chan}}({{\text{\rm\tt C}}},{\text{\rm\tt counter}}(n)))\coloncolon{{\text{\rm\tt end}}}({{\text{\rm\tt C}}}))$$ Since ${\text{\rm\bf chan}}$ is a linear type constructor, a channel can then be sent over another channel just as other linear values, and `send` will consume it. This forms a higher-order session type. We omit any testing code since it is similar to \[ex:array\]. Due to space limitation, we push other examples to \[app:ex\]. Extensions {#sec:ext} ========== We very briefly describe possible extensions of ${\mathcal{L}_{\forall,\exists}^{\pi}}$. First, it is straightforward to add *general recursion* to our language (not to the session type) as has been done in [@Xi:2016ue]. Second, one can always introduce a *higher-order* `fix` into session types, such as $${\text{\rm\tt fix}}(\lambda f{:}(\vv\sigma\rightarrow{\text{\rm\it stype}}).\lambda \vv a{:}\vv\sigma.\pi),\vv s)$$ where $f$ is a static function of sort $(\vv\sigma\rightarrow{\text{\rm\it stype}})\rightarrow\vv\sigma\rightarrow{\text{\rm\it stype}}$, and $\vv s$ are static terms of matching sorts $\vv\sigma$. Correspondingly, we need to introduce another `recurse` to unroll it. A higher-order `fix` will input static terms to form a new session type that dependents on these static terms. Thus these are also a form of dependent session types. Third, binary branching can be extended as well. For instance, we can introduce ${{\text{\rm\tt branch}}}(i,\pi_1,\pi_2,\pi_3), i\in\{0,1,2\}$ and cooresponding session API similar to `ite` to unroll it. More importantly, we can extend ${\mathcal{L}_{\forall,\exists}^{\pi}}$ to support *multi-party session types* based on [@Xi:2017wv]. Roles will be extended from $\{0,1\}$ to a larger set of natural numbers, ${{\text{\rm\bf chan}}}(r,\pi)$ will be extended to ${{\text{\rm\bf chan}}}(R,\pi)$ where $R$ is now a *set* of roles. This is essential because of the need to represent one party’s *complement* roles, which has to be a set. Guards in session API will change from $r=r_0$ to $r_0\in R$, and from $r\neq r_0$ to $r_0\notin R$. `cut` will be extended to another form based on [@Xi:2017wv]. Also, both predicative quantification (dependent types) and higher-order/impredicative quantification (polymorphism) are supported by , and our formulation naturally supports *polymorphic session types* in the sense of [@Caires:2013jb] since `quan` and higher-order `fix` can input session types to form a session type. We give such an example in \[ex:poly\]. However, we focus on dependent session types in this paper. Related Works {#sec:related-works} ============= To our best knowledge, [@Toninho:2011in] is the only other formalization of dependent session types in a similar sense as ours. It is based on intuitionistic linear type theory for a variant of $\pi$-calculus, which extends the work in [@Caires:2010gi] where a kind of Curry-Howard isomorphism is established. The work concerns with two layers, an unspecified dependently typed layer for functional terms that assign meanings to atomic propositions, and a session typed layer that composes sessions and interprets linear logic connectives. Quantifiers connect these two layers where universal quantifier inputs a functional term and existential quantifier outputs a functional term. Their line of works presents session types in a polarized style, corresponding to the left/right introduction/elimination rules of the logic. Our work is different in many ways. Our work is based on $\lambda$-calculus instead of $\pi$-calculus/linear logic, and we have shown our concrete implementations to support the argument that such formulation is practical. Quantifiers are handled slightly differently. We present unpolarized global quantifiers in the session type, then locally interpreted it as $\forall$/$\exists$ through our session API. However, the input/output action is not limited to follow the quantifiers immediately as they do. Our unpolarized style is easier to extend to multi-party sessions, while theirs is inherently binary due to the nature of duality in the logic. [@Caires:2013jb] and [@Pfenning:2011ce] are based on [@Toninho:2011in] which focus on polymorphic session types and proof-carrying code in session types, respectively. Our work supports polymorphic session types in the sense of [@Caires:2013jb] but we do not have space to formally address it. There are many attempts to integrate session types into practical programming languages. [@Pucella:2008dt; @Lindley:2016du; @Sackman:2008ux] embed session types into Haskell, [@Scalas:2016uh] in Scala, [@Jespersen:2015ka] in Rust, [@Ng:2012hu] in C, and [@Hu:2008hta; @Ng:2011go; @Hu:2010fj] in Java. The single sailent feature is that we support dependent session types while none of above supports. Our type system also guarantees linearity and duality natively and staticly without any special encoding. Due to the lack of linear types, [@Lindley:2016du] relies on an encoding of linear $\lambda$-calculus, [@Pucella:2008dt; @Sackman:2008ux] rely on indexed monads. [@Jespersen:2015ka] makes use of affine types in Rust that guarantees “at most once” usage which is still not enough. Other works did not capture linearity in the type system. Duality is encoded as a proof system using type classes in [@Pucella:2008dt; @Lindley:2016du], and using traits in [@Jespersen:2015ka]. [@Scalas:2016uh] uses Scala’s `In[-]`/`Out[-]` types where `-` is a *local* type, and similarly [@Sackman:2008ux] uses `dual/notDual`, and they are both similar to our prior work using **chanpos** and **channeg**. [@Hu:2008hta] ensures duality in the runtime and [@Ng:2011go; @Hu:2010fj] are its extensions. There are other works proposing new languages to support session types, such as [@Toninho:2013iu; @Gay:2010gt; @Wadler:2012ua] and `SILL`[^5] [@Caires:2010gi], but these are not as practical in their current states. There are other works that are loosely related to ours, such as those investigating links between logics and session types [@Wadler:2012ua; @Wadler:2014bta; @Caires:2010gi]. Please refer to [@Xi:2016ue] for more due to space limitations. Conclusion {#sec:conclusion} ========== We have presented a form of dependent session type system ${\mathcal{L}_{\forall,\exists}^{\pi}}$ based on $\lambda$-calculus using unpolarized presentation. Our type system handles quantification over static terms in session types, allowing more precise session protocols to be described elegantly. Linearity is guaranteed statically by the type system, and duality is guaranteed by a combination of global session types, roles at a local endpoint, and guards in the session API. ${\mathcal{L}_{\forall,\exists}^{\pi}}$ also supports delegations, higher-order sessions, polymorphic sessions, and recursively defined sessions. Our type system enjoys subject reduction and progress properties, which guarantees session fidelity and deadlock-freeness. We have shown the practicality of ${\mathcal{L}_{\forall,\exists}^{\pi}}$ by providing a back-end in Erlang/Elixir, which is asynchronous, distributed, and leak-free. Our formulation can also be adapted to multi-party sessions based on multirole logic and we leave this as a future work. Appendix - More Examples {#app:ex} ======================== One can safely send an array by sending a length $n$ first, then followed by $n$ messages for $n$ elements of the array. Such a channel can be encoded in the following dependent session types. $$\begin{aligned} {\text{\rm\tt repeat}} (\tau{:}{\text{\rm\it type}}, n{:}{\text{\rm\it int}}) &{\coloncolonequals}{\text{\rm\tt ite}}(n>0, {{\text{\rm\tt msg}}}({{\text{\rm\tt S}}}, \tau) \coloncolon {\text{\rm\tt repeat}}(\tau,n-1),{{\text{\rm\tt end}}}({{\text{\rm\tt S}}})) \\ {\text{\rm\tt array}} (\tau{:}{\text{\rm\it type}}) &{\coloncolonequals}{{\text{\rm\tt quan}}}({{\text{\rm\tt S}}}, \lambda n{:}{\text{\rm\it int}}.{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}}, {\text{\rm\bf int}}(n))::{\text{\rm\tt repeat}}(\tau, n))\end{aligned}$$ where `repeat` is a recursive session type constructor written in direct style, and its desugared version is as follows, $$\begin{gathered} {\text{\rm\tt repeat}} (\tau{:}{\text{\rm\it type}}, n{:}{\text{\rm\it int}}) {\coloncolonequals}\\ {{\text{\rm\tt fix}}}( \lambda p{:}{\text{\rm\it int}}\rightarrow{\text{\rm\it stype}}. \lambda n{:}{\text{\rm\it int}}. {{\text{\rm\tt ite}}}(n>0, {{\text{\rm\tt msg}}}({{\text{\rm\tt S}}}, \tau) \coloncolon p(n-1), {{\text{\rm\tt end}}}({{\text{\rm\tt S}}})), n) \end{gathered}$$ Note that `repeat` and `array` are session type constructors, which are just static functions returning static terms of sort ${\text{\rm\it stype}}$. Also, the `fix` is a higher-order fixpoint described in \[sec:ext\]. `repeat`$(\tau,n)$ then says, if $n>0$ is true, the session proceeds to allow sending of a value of type $\tau$ from party `S` (${{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},\tau)$), then proceeds as `repeat`$(\tau,n-1)$. If $n>0$ is false, the session can only be terminated by party `S` (${{\text{\rm\tt end}}}({{\text{\rm\tt S}}})$). Similarly, `array` says, party `S` is to send an integer $n$ followed by $n$ repeated messages described by `repeat`$(\tau,n)$. Therefore, the server side can be programmed as follows, [ server ]{} [ (ch:chan(S,array(a)), data:arrref(a,n), len:int(n)): void = ]{} [ () = unify ch ]{} [ () = send (ch, len) ]{} [ sendarr ]{} [ (ch:chan(S,repeat(a,n)), x:int(n), data:arrref(a,m), len:int(m)): void = ]{} [ x = ]{} [ () = recurse ch]{} [ () = itef ch]{} [ close ch ]{} [ () = recurse ch]{} [ () = itet ch]{} [ () = send (ch, data\[len-x\])]{} [ sendarr (ch, x-]{}[, data, len) ]{} [ sendarr (ch, len, data, len) ]{} And its type is $$\texttt{server} : \forall\tau{:}{\text{\rm\it type}}.\forall n{:}{\text{\rm\it nat}}.({\text{\rm\bf chan}}({{\text{\rm\tt S}}},{\text{\rm\tt array}}(\tau)), {\text{\rm\bf arrref}}(\tau,n), {\text{\rm\bf int}}(n))\rightarrow{\text{\rm\bf 1}}$$ where `data` is the array to be sent, whose type is indexed by the type of elements and the length of array. `len` is the length of array, whose type is a singleton integer that equals the length of `data`. `prval` denotes a proof value that has no runtime semantics. After type-checking has passed, these values will be eliminated. The example comes from [SILL]{}[^6], an implementation of binary session types based on [@Caires:2010gi]. As compared to a simple queue, we define a dependently typed queue indexed by its length as follows, with the higher-order ${{\text{\rm\tt fix}}}$ introduced in \[sec:ext\], $$\arraycolsep=1pt \begin{array}{rl} {\text{\rm\tt queue}}(\tau{:}{\text{\rm\it type}}, n{:}{\text{\rm\it int}}){\coloncolonequals}{{\text{\rm\tt branch}}}({{\text{\rm\tt C}}}, &{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},\tau)\coloncolon{\text{\rm\tt queue}}(\tau,n+1),\\ &{{\text{\rm\tt ite}}}(n>0, {{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},\tau)\coloncolon{\text{\rm\tt queue}}(\tau,n-1),{{\text{\rm\tt end}}}({{\text{\rm\tt S}}}))) \end{array}$$ where the client can choose to either enqueue or dequeue an element of type $\tau$. In the dequeue case, instead of encoding an optional value as a `branch` to deal with dequeuing from an empty queue, we use the length of the queue to decide the continuation of the session type. If the length $n$ is greater than 0, the endpoint allows dequeuing. Otherwise, the endpoint can only be closed. As mentioned before, `itet`/`itef` are proof functions that have no runtime cost, while a non-dependently session typed queue will require `choose`/`offer` that need to communicate a tag at runtime. We follow their example, and present the `elem` function as follows, which given a queue and an element `e`, constructs a new queue where `e` will be inserted into the queue as if it is the first element, and `e` will be the first to be dequeued. [ elem ]{} [ (q:chan(C,queue(a,n)), e:a): chan(C,queue(a,n+]{}[)): void = ]{} [ server ]{} [ (out:chan(S,queue(a,n+]{}[)), inp:chan(C,queue(a,n))): void = ]{} [ () = recurse out ]{} [ c = offer out ]{} [ case c ]{} [ | Right () =&gt; ]{} [ () = itet out]{} [ () = send (out, e)]{} [ cut (out, inp) ]{} [ | Left () =&gt; ]{} [ y = recv out]{} [ () = recurse inp]{} [ () = choose (inp, Left())]{} [ () = send (inp, y)]{} [ server (out, inp) ]{} [ create (]{}[ out =&gt; server (out, queue)) ]{} We define a polymorphic cloud service that, given any unlimited function, will provide replicated services of such function. The example is taken from [@Caires:2013jb] that makes use of higher-order quantification over session types, and high-order sessions. We define polymorphic session types as follows, $$\begin{aligned} {\text{\rm\tt service}}(\pi{:}{\text{\rm\it stype}})&{\coloncolonequals}{{\text{\rm\tt branch}}}({{\text{\rm\tt C}}},{{\text{\rm\tt msg}}}({{\text{\rm\tt S}}},{{\text{\rm\bf chan}}}({{\text{\rm\tt C}}},\pi))\coloncolon{\text{\rm\tt service}}(\pi), {{\text{\rm\tt end}}}({{\text{\rm\tt C}}})) \\ {\text{\rm\tt cloud}}&{\coloncolonequals}{{\text{\rm\tt quan}}}({{\text{\rm\tt C}}},\lambda\pi{:}{\text{\rm\it stype}}.{{\text{\rm\tt msg}}}({{\text{\rm\tt C}}},{{\text{\rm\bf chan}}}({{\text{\rm\tt S}}},\pi)\rightarrow{\text{\rm\bf 1}})\coloncolon{\text{\rm\tt service}}(\pi))\end{aligned}$$ Here, ${\text{\rm\tt service}}(\pi)$ is a polymorphic session type constructor that says a client can repeatedly choose to use a service through a newly created endpoint disciplined by session type $\pi$, or to close it. `cloud` is a polymorphic session type that says, as long as the client sends an *unlimited/non-linear* function that can provide the functionality described by $\pi$, the server will turn it into a replicated service. Corresponding server and client programs could be written like the followings. [ server (ch:chan(S,cloud)): void = ]{} [ () = exify ch ]{} [ f = recv ch ]{} [ srv ]{}[ (ch:chan(S,service(p)), f:chan(S,p)-&gt;void): void = ]{} [ () = recurse ch]{} [ c = offer ch]{} [ case c ]{} [ | Right () =&gt; wait ch ]{} [ | Left () =&gt; ]{} [ ep = create (]{}[ ch =&gt; f ch)]{} [ () = send (ch, ep)]{} [ srv (ch, f) ]{} [ srv (ch, f)]{} [ client (ch:chan(C,cloud)): void = ]{} [ echo (ch:chan(S,msg(C,string)::]{}[(C))): void = ]{} [ () = print (recv ch)]{} [ wait ch ]{} [ () = unify ch ]{} [ () = send (ch, echo) ]{} [ prt (ch:chan(C,service(msg(C,string)::]{}[(C))), n:int): void =]{} [ () = recurse ch ]{} [ n &lt;= ]{} [ (choose (ch, Right()); close ch) ]{} [ () = choose (ch, Left())]{} [ ep = recv ch ]{} [ () = send (ep, “hello world!”)]{} [ () = close ep ]{} [ prt (ch, n-]{}[) ]{} [ prt (ch, ]{}[)]{} Appendix - Figures ================== ------------------------------------------------------------------------ [^1]: Note that ${{\text{\rm\tt branch}}}$ is just a special case of ${{\text{\rm\tt ite}}}$ and we can indeed encode ${{\text{\rm\tt branch}}}$ using ${{\text{\rm\tt ite}}}$. [^2]: This is only intuitively interpreted. Its accurate interpretation should be considered together with an endpoint since $\pi$ is global. See later sections. [^3]: It is indeed ${{\text{\rm\bf chan}}}:({\text{\rm\it int}},{\text{\rm\it stype}})\Rightarrow{\text{\rm\it vtype}}$ since in , subset sort is not allowed in a c-sort. We use ${\text{\rm\it role}}$ here just to simplify our presentation. [^4]: This is because that knowing just the `pid` is enough for any process to randomly inject messages to its mailbox. [^5]: <https://github.com/ISANobody/sill> [^6]: <https://github.com/ISANobody/sill>
{ "pile_set_name": "ArXiv" }
--- abstract: | Skeleton-based action recognition has attracted increasing attention due to its potentially broad applications such as autonomous and anonymous surveillance. With the help of deep learning techniques, it has also witnessed substantial progress and achieved excellent accuracy in non-adversarial environments. However, in practice, potential adversaries might easily deceive an action recognition model by performing actions with imperceptible perturbations. Deploying such a model without understanding its adversarial vulnerability might lead to severe consequences, [*e.g.*]{}, recognizing a violent action as a normal one. Despite these security concerns, research on the vulnerability of skeleton-based action recognition remains scant, partly due to the challenges caused by the unique nature of human skeletons and actions. Specifically, we argue that for imperceptible and reproducible adversarial skeleton actions: 1) the bone lengths should be maintained roughly the same as the original bone lengths; 2) the changes of joint angles should be small; 3) the adversarial motion speeds should be restricted. These unique constraints hinder direct applications of existing attack methods to adversarial skeleton actions. In this paper, we conduct a thorough study towards understanding the adversarial vulnerability of skeleton-based action recognition. We first formulate the generation of adversarial skeleton actions as a constrained optimization problem by representing or approximating the constraints with mathematical equations. To deal with the intractable primal optimization problem with equality constraints, we propose to optimize its unconstrained dual problem using ADMM. We further design an efficient plug-in defense, inspired by recent theories and empirical observations, against adversarial skeleton actions. Extensive evaluations demonstrate the effectiveness of our attack and defense, and reveal the properties of adversarial skeleton actions. author: - 'Tianhang Zheng$^{1}$, Sheng Liu$^{2}$, Changyou Chen$^{2}$, Junsong Yuan$^{2}$, Baochun Li$^{1}$, Kui Ren$^{3}$' bibliography: - 'sample-base.bib' title: 'Towards Understanding the Adversarial Vulnerability of Skeleton-based Action Recognition' --- Introduction ============ Action recognition is an important task in multimedia and computer vision, motivated by many downstream applications such as video surveillance and indexing, and human-machine interaction [@cheng2015advances]. It is also a challenging task since it requires capturing long-term spatial-temporal motion patterns to understand the semantics of actions. Many recent works from the multimedia & computer vision community [@wang2016action; @ke2017a; @gao2019optimized; @shi2019two; @shi2019skeleton; @yang2020hierarchical] propose to learn action recognition on the human skeleton motion captured by cameras or depth sensors, where an action is represented by a time series of poses represented as 3D body skeletons. Compared with video streams, skeleton representation is more robust to the variance of background clutters, and also easier-to-handle for machine learning models due to its compact representation. Recent advances in deep learning techniques have been applied to skeleton-based action recognition, including convolutional neural networks [@ke2017a; @li2018co], recurrent neural networks [@li2018CVPR; @si2019attention], and graph neural networks [@yan2018spatial; @shi2019two; @gao2019optimized; @liu2020disentangling]. ![The targeted setting: misleading the model to recognize “kicking person" as “drinking water" (normal action) by perturbing the skeleton action. To launch the attack in a real-world scenario ([*e.g.,*]{} under a surveillance camera), the adversarial skeleton action should satisfy certain constraints. The figure is drawn based on [@shahroudy2016ntu].](diagram_v2.png){width="0.95\linewidth"} \[fig:target\_diagram\] On the other hand, existing work has demonstrated the vulnerability of deep learning techniques to adversarial examples in many application domains, such as face recognition and object detection. This phenomenon motivates us to suspect that, despite achieving high accuracy in non-adversarial environments, the deep neural networks (DNNs) for skeleton-based action recognition might also be vulnerable to adversarial skeleton actions. It is worth noting that a thorough study on the adversarial vulnerability of action-recognition models is indispensable before deploying them to real-world applications such as surveillance systems. Otherwise, the potential adversaries might easily deceive those systems by performing specific adversarial actions, leading to significant consequences, as shown in Fig. \[fig:target\_diagram\]. To our knowledge, the study on adversarial skeleton actions is scant and non-trivial[^1], due to the fundamental differences between the properties of adversarial skeleton actions and other adversarial examples. The differences are caused by the bones between joints and the joint angles, which impose unique spatial constraints on skeleton data [@shahroudy2016ntu]. Specifically, in the generated adversarial skeleton actions, lengths of bones must be maintained the same, and simultaneously, joint angles cannot violate certain physiological structures. In addition, considering the physical properties of human bodies, the speeds of motions in the adversarial actions should also be constrained. [*If any constraint is not satisfied, the adversarial skeleton actions might be easily perceived and detected or could not be performed by the actors.*]{} To understand the adversarial vulnerability of skeleton-based action recognition, we first study how to generate adversarial skeleton actions. Specifically, we formulate the generation of adversarial skeleton actions as a constrained optimization problem by representing the spatio-temporal constraints with mathematical equations. Since the primal constrained problem is intractable, we turn to solve its dual problem. Moreover, since all the constraints are represented by mathematical equations, both primal and dual variables become unconstrained in the dual problem. We further specify an efficient algorithm based on ADMM to solve the unconstrained dual problem, in which the internal minimization objective is optimized by an Adam optimizer, and the external maximization objective is optimized by one-step gradient ascent. We show that this algorithm can find an adversarial skeleton action within a couple of hundred internal steps. Other than the attack, we further propose an efficient defense against adversarial skeleton actions based on recent theories and empirical observations. Our defense consists of two core steps, [*i.e.,*]{} adding Gaussian noise and Gaussian filtering to action data. The first step, adding Gaussian noise, is inspired by the recent advance in certified defenses. Specifically, adding Gaussian noise to the input is proved to be a certified defense, which means additive Gaussian noise on the adversarial examples can guarantee the model to output a correct prediction (with high probability), as long as the adversarial perturbation is restricted within a certain radius in the neighbor of the original data sample. Note that there are several other methods to certify model robustness, such as dual approach, interval analysis, and abstract interpretations [@dvijotham2018dual; @wong2018provable; @mirman2018differentiable; @gowal2018effectiveness; @wang2018efficient]. We adopt the Gaussian noise method because it is simple, effective, and more importantly, scalable to complicated models. Note that skeleton-based action recognition models are always more complicated than the common ConvNets certified by [@dvijotham2018dual; @wong2018provable; @mirman2018differentiable; @gowal2018effectiveness; @wang2018efficient]. The second step is to smooth the skeleton frames along the temporal axis using a Gaussian filter. This step will not affect the robustness certified by the first step according to the post-processing property [@lecuyer2018certified; @li2018second; @cohen2019certified], but can always filter out a certain amount of adversarial perturbation and random noise in practice, thus making our defense applicable to normally trained models. Our proposed attack and defense are evaluated on two opensource models, [*i.e.*]{}, 2s-AGCN and HCN [@shi2019two; @li2018co] [^2]. Extensive evaluations show that our attack can achieve $100\%$ attack success rate with almost no violation of the constraints. Moreover, the visualization results, including images and videos, demonstrate that the difference between the original and adversarial skeleton actions is imperceptible. Extensive evaluations also show that our defense is effective and efficient. Specifically, our defense can improve the empirical accuracy of normally trained models to over $60\%$ against adversarial skeleton actions under different settings. To summarize, our main contribution is three-fold: 1. We identify the constraints needed to be considered in adversarial skeleton actions, and formulate the problem of generating adversarial skeleton actions as a constrained optimization problem by formulating those constraints as mathematical equations. We further propose to solve the primal constrained problem by optimizing its dual problem using ADMM, achieving 100% attack success rate. 2. We propose an efficient two-step defense against adversarial skeleton actions based on previous theories and empirical observations, and specify the defense in both inference and certification stages. The proposed defense achieves high robust accuracy under mild perturbations. 3. We conduct extensive evaluations on two opensource models and two datasets. We also provide several interesting observations regarding the properties of adversarial skeleton actions based on the experimental results. Preliminaries ============= Definitions and Notations ------------------------- Let $\xb$ and $l \in \{1, 2,...,L\}$ respectively denote a data sample and the label, where $L$ is the number of all possible classes. For an image, $\xb$ is a 2D matrix. For a skeleton action studied in this paper, $\xb \triangleq \{(x^\tau_i, y^\tau_i, z^\tau_i)_{i=1}^I\}_{\tau=1}^\Tau$, where $(x^\tau_i, y^\tau_i, z^\tau_i)$ denotes the position (coordinates) of the $i$-th joint of the $\tau$-th skeleton frame in an action sequence, with $I$ and $\Tau$ denoting the number of joints in a skeleton and the number of skeleton frames in an action sequence, respectively. The corresponding adversarial skeleton action is denoted by $\xb' \triangleq \{(x'^\tau_i, y'^\tau_i, z'^\tau_i)_{i=1}^I\}_{\tau=1}^\Tau$. We take the skeletons in the largest dataset, *i.e.*, NTU RGB+D dataset, as an example. As shown in Fig. \[fig:skeleton\], in a skeleton, there are totally 25 joints in a skeleton frame, and thus $I=25$. The number of frames $\Tau$ differs for each skeleton action, and usually, we subsample a constant number of frames from each sequence or pad zeros after each sequence to endow all the skeleton actions with the same $\Tau$. Let $\Fb_\Thetab(\cdot)$ denote a classification network, where $\Thetab$ represents the network weights. The logit output on $\xb$ is denoted by $\Fb_\Thetab(\xb)$ with $L$ elements ($\{\Fb_{\Thetab, k}(\xb)~|~k=1,...,L\}$). $\Fb_\Thetab(\cdot)$ can correctly classify $\xb$ iff $\argmax_{k} \Fb_{\Thetab, k}(\mathbf{x}) = l$. The goal of adversarial attacks is to find an adversarial sample $\xb'$, which satisfies several pre-defined constraints, such that $\argmax_{k} \Fb_{\Thetab, k}(\xb') \neq l$ or $\argmax_{k} \Fb_{\Thetab, k}(\mathbf{x}) = l_t$ ($l_t$ is the target label). A commonly-used constraint is that $\xb'$ should be close to the original sample $\xb$ according to some distance metric. ![Skeleton Representation](skeleton-v1.pdf "fig:"){width="0.36\linewidth"} ![Skeleton Representation](joint_angle.png "fig:"){width="0.34\linewidth"} \[fig:skeleton\] DNNs for Skeleton-based Action Recognition ------------------------------------------ In the following, we briefly introduce the two DNNs used for evaluation of our proposed attack method in this project. HCN is a CNN-based end-to-end hierarchical network for learning global co-occurrence features from skeleton data [@li2018co]. HCN is designed to learn different levels of features from both raw skeleton and skeleton motion. The joint-level features are learned by a multi-layer CNN, and the global co-occurrence features are learned from the fused joint-level features. At the end, the co-occurrence features are also fed to a fully-connected network for action classification. 2s-AGCN is one of the state-of-the-art GCN-based models for skeleton-based action recognition. In contrast to the earliest GCN-based model, ([*i.e.,*]{} ST-GCN), 2s-AGCN learns the appropriate graph topology of every skeleton action rather than prefine the graph topology. This enables 2s-AGCN to capture the implicit connections between joints in certain actions, such as the connection between hand and face in the “wiping face" action. Besides, 2s-AGCN also adopts the two-stream framework to learn from both static and motion information. Overall, 2s-AGCN significantly improves the accuracy of ST-GCN by nearly 7%. Adversarial Attacks {#subsec:adv_attack} ------------------- After the discovery of adversarial examples, the community has developed hundreds of attack methods to generate adversarial samples. In the following, we mainly introduce four attack methods plus a parallel work, with a discussion on the difference between our proposed method and these attacks. #### Fast Gradient Sign Method (FGSM) FGSM is a typical one-step adversarial attack algorithm proposed by [@goodfellow2014]. The algorithm updates a benign sample along the direction of the gradient of the loss w.r.t. the sample. Formally, FGSM follows the update rule as $$\xb^\prime = \clip_{[v_{min}, v_{max}]}\{\xb + \epsilon \cdot \sign(\nabla_{\xb} \mathcal{L}(\Thetab, \xb, l))\}~,$$\[eq:fgsm\] where $\epsilon$ controls the maximum $\ell_\infty$ perturbation of the adversarial samples; $[v_{min}, v_{max}]$ is the valid element-wise value range and $\clip_{[a,b]}(\cdot)$ function clips its input into the range of $[a, b]$. #### Projected Gradient Descent (PGD) PGD [@kurakin2016adversarial; @madry2017towards] is a strong iterative version of FGSM, which executes Eq. \[eq:fgsm\] for multiple steps with a smaller step size and then projects the updated adversarial examples into the pre-defined $\ell_p$-norm ball. Specifically, in each step, PGD updates the sample by $$\begin{aligned} \xb^\prime_{t+1} = Proj\{\xb^\prime_{t} + \alpha \cdot \sign(\nabla_{\xb^\prime_{t}} \mathcal{L}(\Thetab, \xb^\prime_{t}, l)\}~\end{aligned}$$ The $Proj$ function is a clip function for $\ell_\infty$-norm balls, and an $\ell_2$ normalizer for $\ell_\infty$-norm balls. #### Carlini and Wagner Attack [@carlini2017towards] proposes an attack called C&W attack, which generates $\ell_p$-norm adversarial samples by optimization over the C&W loss: $$\begin{aligned} \label{eq:cw_adv_optimization} \min_{\xb'}D(\xb, \xb') + c \cdot loss(\xb')~.\end{aligned}$$ In the C&W loss, $D(\xb, \xb')$ represents some distance metric between the benign sample $\xb$ and the adversarial sample $\xb'$, and the metrics used in [@carlini2017towards] include $\ell_\infty$, $\ell_0$, and $\ell_2$ distances. $loss(\cdot)$ is a customized loss. It is worth noting that our proposed attack is completely different from PGD or C&W attack. For PGD, C&W, or many other attacks, the simple constraints on the pixel value can be resolved by projection functions or naturally incorporated into the objective by $sigmoid$/$tanh$ function. However, in our scenario, the constrained optimization problem is much more complicated, and thus has to be solved by more advanced methods. #### ADMM-based Attack [@zhao2018admm] also proposes a framework based on ADMM to generate $\ell_p$ adversarial examples. However, we note that our proposed attack is completely different from theirs in two aspects: First, the constraints we consider in this paper are much more complicated than the $\ell_p$-norm constraints in [@zhao2018admm]. Second, we formulate the problem in a very different manner. Specifically, [@zhao2018admm] follows the ADMM framework to break the problem defined like Eq. \[eq:cw\_adv\_optimization\] into two sub-problems; while our attack formulates a different problem with indispensable equality constraints, where ADMM is a natural solution to this problem. #### Adversarial Attack on Skeleton Action Note that [@liu2019adversarial] is a parallel work that proposes an attack based on FGSM and BIM (PGD) to generate adversarial skeleton actions. Specifically, [@liu2019adversarial] adapts the FGSM and BIM to skeleton-based action recognition by using a clipping function and an alignment operation to impose the bone and joint constraints on the updated adversarial skeleton actions in each iteration. [*However, [@liu2019adversarial] is very different from our work.*]{} First, the joint constraints considered in [@liu2019adversarial] are not the constraints for joint angles mentioned before. Second, the alignment operation might corrupt the perturbation learned in each iteration. In contrast to [@liu2019adversarial], we attempt to formulate adversarial skeleton action generation as a constrained optimization problem with equality constraints. Releasing the equality constraints by Lagrangian multipliers yields an unconstrained dual optimization problem, which does not need any complicated additional operation in the optimization process. Third, we propose to solve the the dual optimization problem by ADMM, which is a more appropriate method to optimize complicated constrained problems. Therefore, the attack achieves better performance than [@liu2019adversarial], which will be detailed in section \[sec:attack\_performance\]. Finally, we specify a defense method against adversarial skeleton actions based on the state-of-the-art theories and our observations. Alternating Direction Method of Multipliers (ADMM) -------------------------------------------------- Alternating Direction Method of Multipliers (ADMM) is a powerful optimization algorithm to handle large-scale statistical tasks in diverse application domains. It blends the decomposability of dual ascent with the great convergence property of the method of multipliers. Currently, ADMM plays a significant role in solving statistical problems, such as support vector machines [@forero2010consensus], trace norm regularized least squares minimization [@yang2013fast], and constrained sparse regression [@bioucas2010alternating]. Except for convex problems, ADMM is also a widely used solution to some nonconvex problems, whose objective function could be nonconvex, nonsmooth, or both. [@wang2019global] shows that ADMM is able to converge as long as the objective has a smooth part, while the remaining part can be coupled or nonconvex, or include separable nonsmooth functions. Applications of ADMM to nonconvex problems include network reference [@miksik2014distributed], global conformal mapping [@lai2014splitting], noisy color image restoration [@lai2014splitting]. Adversarial Defenses -------------------- Both learning and security communities have developed many defensive methods against adversarial examples. Among them, adversarial training and several certified defenses attract the most attention due to their outstanding/guaranteed performance against strong attacks [@he2017adversarial; @uesato2018adversarial; @athalye2018obfuscated]. In the following, we briefly introduce adversarial training and several certified defenses, including the randomized smoothing method adopted in this paper. #### Adversarial Training Adversarial training is one of the most successful empirical defenses in the past few years [@goodfellow2014; @madry2017towards; @zhang2019theoretically]. The intuition of adversarial training is to improve model robustness by training the model with adversarial examples. Although adversarial training achieves tremendous success against many strong attacks [@zheng2019distributionally; @andriushchenko2019square; @tashiro2020output], its performance is not theoretically guaranteed and thus might be compromised in the future. Besides, adversarial training always requires much more computational resource than standard training, making it not scalable to complicated models. #### Certified Defenses A defense with a theoretical guarantee on its defensive performance is considered as a certified defense. In general, there are three main approaches to design certified defenses. The first approach is to formulate the certification problem as an optimization problem and bound it by dual approach and convex relaxations [@dvijotham2018dual; @raghunathan2018certified; @wong2018provable]. The second approach approximates a convex set that contains all the possible outputs of each layer to certify an upper bound on the range of the final output [@mirman2018differentiable; @gowal2018effectiveness; @wang2018efficient]. The third is the randomized smoothing method used in this paper. The only essential operation for this method is to add Gaussian/Laplace noise to the inputs, which is simple and applicable to any deep learning models. [@lecuyer2018certified] first proves that randomized smoothing is a certified defense by theories on differential privacy. [@li2018second] improves the certified bound using a lemma on Renyi divergence. Cohen et al. [@cohen2019certified] proves a tight bound on the $\ell_2$ robust radius certified by adding Gaussian noise using the Neyman-Pearson lemma. [@jia2019certified] further extends the approach of [@cohen2019certified] to the $top-k$ classification setting. Since the bound proved by [@cohen2019certified] is the tightest, the method in [@cohen2019certified] is used for certification. In this paper we adopt the approach in [@lecuyer2018certified] due to its ability for efficient inference in practice. Threat Model ============ Adversary Knowledge: White-box Setting -------------------------------------- In this paper, we follow the white-box setting, where the adversary has full access to the model architecture and parameters. We make this assumption because (i) it is always a safe, conservative, and realistic assumption since we might never know the knowledge of potential adversaries about the model [@carlini2017towards], which varies among different adversaries and also changes over time. (ii) For systems/devices equipped with an action recognition model, recognition is more likely to be done locally, or on a local cloud, making the adversary easily acquire the model parameters with his own system/device. Note that although most of the experiments on the proposed attack and defense are done under the white-box setting, we also have several experiments on evaluating the transferability of our attack. Adversary Goal: Targeted & Untargeted label Setting --------------------------------------------------- Under the targeted setting, the goal of an adversary is to mislead the recognition model to predict the adversarial skeleton action as a targeted label pre-defined by the adversary. For instance, suppose the adversary is “kicking“ someone under a surveillance camera equipped with an action recognition model. It may launch a targeted attack to mislead the model to recognize this violent action as a normal one such as ”drinking water". Under the untargeted label settings, an adversary only aims to disable the recognition and thus is considered successful as long as the model makes wrong predictions instead of a specific targeted prediction. In this paper, we propose two objectives suitable for the above two settings respectively, which will be detailed in section \[subsec:constrained\]. Imperceptibility & Reproducibility ---------------------------------- Except for the aforementioned adversary goals, the adversary also requires the adversarial perturbation to be both imperceptible and reproducible. Here “imperceptibility" means it should be difficult for human vision to figure out the difference between the original and adversarial skeleton actions. Imperceptibility is not only a common requirement in the previous attacks, but also a useful one in our scenario. Note that it is natural to schedule a periodical examination for an autonomous surveillance system by human labor to check if the system works well. If the system has been fooled by a seemingly “normal" adversarial skeleton action, the mistake might be considered due to the system itself rather than the adversary who performs the adversarial skeleton action in the examination process. Here “reproducibility” is an additional requirement specific to our scenario. As mentioned in the introduction, the adversarial skeleton action could be a real threat when it can be reproduced under a real-world system. Thus, to make our attack a real-world threat, the generated adversarial skeleton actions should satisfy three concrete constraints to be imperceptible and reproducible, which will be detailed in section \[sec:attack\]. Adversarial Skeleton Action {#sec:attack} =========================== In this section, we present our proposed attack, [*i.e.,*]{} ADMM attack. We first introduce how to formulate the three constraints into mathematical equations. Then we formulate the constrained optimization problem to generate adversarial skeleton actions under both targeted and untargeted settings. Finally, we elaborate on how to solve the optimization problem by ADMM. Bone Constraints ---------------- We again take the skeletons in the NTU RGB+D dataset as an example. As shown in Fig. \[fig:skeleton\], in a skeleton, there are totally 25 joints, forming a total of 24 bones. While the bones are not explicitly considered in modeling, they are strictly connecting to the 25 joints, thus imposing 24 bone-length constraints, [**i.e.**]{}, the distance between the joints at the two ends of a bone should remain the same in adversarial skeleton actions. To mathematically represent the 24 bones, we associate each joint with its preceding joint, forming the two ends of a bone. As a result, the 24 preceding-joints for joint-2$\sim$joint-25 are denoted as $\mathcal{P} \triangleq \{(x^\tau_{pi}, y^\tau_{pi}, z^\tau_{pi})_{i=2}^{25}\}$. The corresponding joint indices of the elements in $\mathcal{P}$ are {1, 21, 3, 21, 5, 6, 7, 21, 9, 10, 11, 1, 13, 14, 15, 1, 17, 18, 19, 2, 8, 8, 12, 12}. We define the $i$-th bone’s length as $B^\tau_i \triangleq \sqrt{(x^\tau_i - x^\tau_{pi})^2 + (y^\tau_i - y^\tau_{pi})^2 + (z^\tau_i - z^\tau_{pi})^2}$. In this regard, the bone constraints can be represented as $B^\tau_i = B'^\tau_i$. Due to the measurement errors in the NTU dataset itself, here we also tolerate very small difference between $B^\tau_i$ and $B'^\tau_i$. Therefore, we can finally formulate the bone constraints as $$\label{eq:con1} |B'^\tau_i - B^\tau_i|/B^\tau_i \leq \epsilon_L,$$ where $\epsilon_L$ is usually set as $0.01\sim0.03$. [*Note that inequality constraints in the primal problem will impose inequality constraints on the corresponding Lagrangian variables in the dual problem.*]{} In order to avoid this in the dual problem, we reformulate the above inequality constraints as mathematical equations, [*i.e.*]{}, is equivalent to $$\label{eq:bone_constraints} \max\{|B'^\tau_i - B^\tau_i|/B^\tau_i - \epsilon_L, 0\} = 0.$$ Joint Angle Constraints ----------------------- Except for the bone-length constraints, we also need to impose constraints on the rotations of the joint angles according to the physiological structures of human beings. Let us also use the NTU dataset as an example. Each joint angle corresponds to the angle between two bones, and thus can be represented by the three joint locations of those two corresponding bones as illustrated in the right of Fig. \[fig:skeleton\]. Note that a natural way to compute the joint angle as shown in Fig. \[fig:skeleton\] is to first compute the cosine value and then input the value into the arccos function. However, the gradient of arccos function is likely exploded, causing large numerical errors when the $\cos$ value of the joint angle is close to $1$ ($\frac{d}{dx} arccos x = -\frac{1}{\sqrt{1 - x^2}}$). To deal with this issue, we derive an approximate upper bound for the changes of joint angle value to avoid computing the arccos function and its gradient. Again, take the right of Fig. \[fig:skeleton\] as an example, the angle change $\Delta\theta_9$ caused by the displacement of joint-9 ([*i.e.,*]{} $x'^\tau_{9} - x^\tau_{9}$, $y'^\tau_{9} - y^\tau_{9}$, $z'^\tau_{9} - z^\tau_{9}$) can be approximated by $\sin\Delta\theta_9 \approx \frac{\sqrt{(x'^\tau_{9} - x^\tau_{9})^2 + (y'^\tau_{9} - y^\tau_{9})^2 + (z'^\tau_{9} - z^\tau_{9})^2}}{\sqrt{(x^\tau_{10} - x^\tau_{9})^2 + (y^\tau_{10} - y^\tau_{9})^2 + (z^\tau_{10} - z^\tau_{9})^2}}$. In particular, when the angle change $\Delta\theta$ is smaller than $0.1$ ([*i.e.,*]{} $5.73^{\circ}$), we can consider $sin\Delta\theta$ almost same as $\Delta\theta$. The total angle change $\Delta\theta$ is upper bounded by the sum of the changes caused by the displacements of joint-9, joint-10, and joint-11. Therefore the upper bound can be represented by $ J'^\tau = \frac{\sqrt{(x'^\tau_{9} - x^\tau_{9})^2 + (y'^\tau_{9} - y^\tau_{9})^2 + (z'^\tau_{9} - z^\tau_{9})^2} }{\sqrt{(x^\tau_{10} - x^\tau_{9})^2 + (y^\tau_{10} - y^\tau_{9})^2 + (z^\tau_{10} - z^\tau_{9})^2}} + \\ \frac{\sqrt{(x'^\tau_{10} - x^\tau_{10})^2 + (y'^\tau_{10} - y^\tau_{10})^2 + (z'^\tau_{10} - z^\tau_{10})^2} }{\sqrt{(x^\tau_{10} - x^\tau_{9})^2 + (y^\tau_{10} - y^\tau_{9})^2 + (z^\tau_{10} - z^\tau_{9})^2}} + \frac{\sqrt{(x'^\tau_{10} - x^\tau_{10})^2 + (y'^\tau_{10} - y^\tau_{10})^2 + (z'^\tau_{10} - z^\tau_{10})^2} }{\sqrt{(x^\tau_{11} - x^\tau_{10})^2 + (y^\tau_{11} - y^\tau_{10})^2 + (z^\tau_{11} - z^\tau_{10})^2}} \\ + \frac{\sqrt{(x'^\tau_{11} - x^\tau_{11})^2 + (y'^\tau_{11} - y^\tau_{11})^2 + (z'^\tau_{11} - z^\tau_{11})^2} }{\sqrt{(x^\tau_{11} - x^\tau_{10})^2 + (y^\tau_{11} - y^\tau_{10})^2 + (z^\tau_{11} - z^\tau_{10})^2}} $ Although this representation looks more complicated than the arccos function, its gradient can be computed efficiently and accurately. Given such an approximation, the joint angle constraints can be similarly represented as $$\label{eq:joint_angle_constraint} \max\{J_k'^{\tau} - \epsilon_J, 0\} = 0$$ where $\epsilon_J$ is set as $0.1\sim0.2$ ($6^\circ \sim 12^\circ$). [*Note that $J_k'^{\tau}$ represents the approximation of the change of a joint angle.*]{} Speed Constraints ----------------- According to the physical conditions of human beings, we should consider one more type of constraints, [*i.e.,*]{} temporal smoothness constraints. By those constraints, we attempt to restrict the speeds of the motions in the generated adversarial skeleton actions. Specifically, the speeds of the motions can be approximated by the displacements between two consecutive temporal frames, [*i.e.,*]{} $S^\tau_m \approx \sqrt{(x^{\tau+1}_m - x^\tau_m)^2 + (y^{\tau+1}_m - y^\tau_m)^2 + (z^{\tau+1}_m - z^\tau_m)^2}$. Then, similar to Eq. \[eq:bone\_constraints\], we bound the change of speeds by $$\label{eq:speed_constraints} \max\{|S'^\tau_m - S^\tau_m|/S^\tau_m -\epsilon_S, 0\} = 0,$$ where $\epsilon_L$ is usually set as (smaller than) $ 10\%$. Constrained Primal Problem Formulation {#subsec:constrained} -------------------------------------- In this subsection, we introduce the main objectives used under the untargeted setting and targeted setting. #### Untargeted Setting Under the untargeted setting, the adversary achieves its goal as long as the DNN makes a prediction other than the ground-truth label, [*i.e.,*]{} $\argmax_{k} \Fb_{\Thetab, k}(\xb') \neq l$. This will hold iff $F_{\Thetab, l}(\xb') < \max_{k, k\neq l}F_{\Thetab, k}(\xb')$. Therefore, we define the objective as minimizing $\max\{F_{\Thetab, l}(\xb') - max_{k, k\neq l}F_{\Thetab, k}(\xb') + conf, 0\}$, where $conf > 0$ is the desired confidence value of the DNN on the wrong prediction. Note that if the objective is equal to $0$, we have $max_{k, k\neq l}F_{\Thetab, k}(\xb') \geq F_{\Thetab, l}(\xb') + conf$. #### Targeted Setting The goal of the adversary is to render the prediction result to be the attack target $l_t$, [*i.e.,*]{} $\argmax_{k \in \mathcal{K}} \Fb_{\Thetab, k}(\xb') = l_t$. Therefore, the primal objective is defined as minimizing the cross entropy between $\Fb_{\Thetab, k}(\xb')$ and $l_t$, or $\max\{\max_{k, k\neq l_t}F_{\Thetab, k}(\xb') - F_{\Thetab, l_t}(\xb') + conf, 0\}$ following the logic of the untargeted setting. We can also adopt other objectives for our purpose. However, it turns out the above two main objectives are the most commonly-used ones in previous work [@kurakin2016adversarial; @madry2017towards; @carlini2017towards]. For simplicity, we denote the main loss by $\mathcal{L}(\xb, l)$. The constrained primal problem can then be formulated as $$\begin{aligned} \min\limits_{\xb'}~ & \mathcal{L}(\xb', l) \\ \mbox{subject to}~&~\mbox{Eq.~(\ref{eq:bone_constraints}), (\ref{eq:joint_angle_constraint}), (\ref{eq:speed_constraints})} ~\nonumber\end{aligned}$$ ![Evolution of the averaged loss items and the constraints ($\beta=1.0$)[]{data-label="fig:analysis"}](loss.png "fig:"){width="0.48\linewidth"} ![Evolution of the averaged loss items and the constraints ($\beta=1.0$)[]{data-label="fig:analysis"}](constraints.png "fig:"){width="0.48\linewidth"} Dual Optimization by ADMM ------------------------- Note that our constrained primal problems are in general intractable. Instead of searching for a solution to the constrained primal problem, we propose to formulate and optimize its unconstrained dual problem via ADMM. The algorithm is illustrated in Alg. \[alg:lagrangian\]. Specifically, we first define the augmented Lagrangian of the constrained primal as shown in Alg. \[alg:lagrangian\]. The additional term $\frac{\beta}{2}(\|\mathcal{\Bb}'\|_2^2 + \|\mathcal{\Jb}'\|_2^2 + \|\mathcal{\mathbf{S}}'\|_2^2)$, which is commonly used in ADMM (for nonconvex problems), aims to further penalize any violation of the equality constraints. [*We note that larger $\beta$ usually leads to smaller violation but larger final main objective (decreases the attack success rate)*]{}. Specifically, given the Lagrangian $\mathcal{G}(\xb, l; \bm{\lambda}, \bm{\nu}, \bm{\omega})$ (defined in Alg. \[alg:lagrangian\]), the dual problem is $\max\limits_{\bm{\lambda}, \bm{\nu}, \bm{\omega}}\min\limits_{\xb'}\mathcal{G}(\xb, l; \bm{\lambda}, \bm{\nu}, \bm{\omega})$. Note that since the internal function $\min\limits_{\xb'}\mathcal{G}(\xb, l; \bm{\lambda}, \bm{\nu}, \bm{\omega})$ is an affine function w.r.t. the variables $\bm{\lambda}, \bm{\nu}, \bm{\omega}$, we can simply use single-step gradient ascent with a large step size (usually set as $\beta$ in ADMM) to update those dual variables. However, $\mathcal{G}(\xb, l; \bm{\lambda}, \bm{\nu}, \bm{\omega})$ is an extremely complicated nonconvex function w.r.t. the adversarial sample $\xb'$. Therefore, in most cases, we could only guarantee local optima for the internal minimization problem. Fortunately, it turns out that even the local optima can always fool the DNNs. To find a local optimum efficiently, we adopt the Adam optimizer instead of the vanilla stochastic gradient descent (SGD) since Adam optimizer always converges faster than vanilla SGD. Theoretically, a local minimum is guaranteed because the Adam optimizer stops updating the variables when the gradients are (close to) $0$. Next, we further look into the evolution of the loss during the optimization process. As shown in Fig. \[fig:analysis\], at the very beginning ([*i.e.,*]{} the first stage), the internal minimization problem finds adversarial samples with large violation of the constraints. The large violation will cause the Lagrangian multipliers $\bm{\lambda}, \bm{\nu}, \bm{\omega}$ to increase rapidly, and thus significantly increase the loss terms $\langle \bm{\lambda}, \mathcal{\Bb}' \rangle$ (bone loss), $\langle \bm{\nu}, \mathcal{\Jb}' \rangle$ (joint loss), and $\langle \bm{\omega}, \mathcal{\mathbf{S}}' \rangle$ (speed loss). As a result, the algorithm proceeds into the second stage, where the Adam optimizer focuses more on diminishing the constraint violation $\Bb'$, $\Jb'$, and $\mathcal{\mathbf{S}}'$ when optimizing $\xb'$. Finally, the algorithm proceeds into a relatively stable stage where we can stop the algorithm. According to Fig. 2, our algorithm is very efficient in the sense that it only needs 200 (internal) iterations to enter the final stable stage. Loss function $\mathcal{L}(\xb, l)$, hyper-parameter $\beta$, adam optimizer for the adversarial skeleton action $\xb'$, maximum number of iterations $T$. **Define Constraints:** $\mathcal{B}'^\tau_j \triangleq \max\{|B'^\tau_j - B^\tau_j|/B^\tau_j - \epsilon_L, 0\}$, $\mathcal{J}'^\tau_k \triangleq \max\{J'^{\tau}_k - \epsilon_J, 0\}$, and $\mathcal{S}'^{\tau}_m \triangleq \max\{|S'^\tau_m - S^\tau_m|/S^\tau_m -\epsilon_S, 0\}$. (Vector Representations: $\mathcal{\Bb}'$, $\mathcal{\Jb}'$, and $\mathcal{\mathbf{S}}'$) **Define Lagrangian Variables:** $\bm{\lambda}$, $\bm{\nu}$, and $\bm{\omega}$ (Corresponding to $\mathcal{\Bb}'$, $\mathcal{\Jb}'$, and $\mathcal{\mathbf{S}}'$) **Define Augmented Lagrangian:** $\mathcal{G}(\xb, l; \bm{\lambda}, \bm{\nu}, \bm{\omega}) \triangleq \mathcal{L}(\xb, l) + \langle \bm{\lambda},\mathcal{\Bb}' \rangle + \langle \bm{\nu}, \mathcal{\Jb}' \rangle + \langle \bm{\omega}, \mathcal{\mathbf{S}}' \rangle + \frac{\beta}{2}(\|\mathcal{\Bb}'\|_2^2 + \|\mathcal{\Jb}'\|_2^2 + \|\mathcal{\mathbf{S}}'\|_2^2)$. **Update $\xb'$:** fix the multipliers $\bm{\lambda}(t), \bm{\nu}(t), \bm{\omega}(t)$ $\xb'(t+1) \in \argmin_{\xb'} \mathcal{G}(\xb', l; \bm{\lambda}(t), \bm{\nu}(t), \bm{\omega}(t))$ updated by the adam optimizer **Update Multipliers:** compute $\mathcal{\Bb'}(t+1)$, $\mathcal{\Jb'}(t+1)$, and $\mathcal{\mathbf{S}'}(t+1)$ based on $\xb'(t+1)$ $\bm{\lambda}(t+1) = \bm{\lambda}(t) + \beta \mathcal{\Bb'}(t+1)$; $\bm{\nu}(t+1) = \bm{\nu}(t) + \beta \mathcal{\Jb'}(t+1)$; $\bm{\omega}(t+1) = \bm{\omega}(t) + \beta \mathcal{\mathbf{S}'}(t+1)$ **Output** $\xb'(T)$ ![image](visualization_main.png){width="0.9\linewidth"} Defense against Adversarial Skeleton Actions ============================================ Note that although the method proposed in [@li2018second; @cohen2019certified] can certify larger robust radii than [@lecuyer2018certified]. However, the sample complexity to compute the confidence intervals in [@li2018second; @cohen2019certified] will lead to computational overhead in the inference stage. [*Therefore, we only use the method in [@cohen2019certified] in the certification process. In the inference stage, we modify the method in [@lecuyer2018certified] to build a relatively efficient defense against adversarial skeleton actions, as shown in Alg. \[alg:randomized\_defense\]*]{}. In general, our proposed defense consists of two steps: adding Gaussian noise and temporal filtering by Gaussian kernel. In the following, we will detail these two steps and explain why we include them in the defense. Additive Gaussian Noise ----------------------- Our first step is adding Gaussian noise to the skeleton actions. In the inference stage, we follow [@lecuyer2018certified] to make the prediction as $\argmax_k E(\Fb_{\Thetab, k}(\mathcal{M}(\xb')))$ given input $\xb'$, where $\mathcal{M}(\xb) = \Gb(\xb + \zb)$ is randomized mechanism with Gaussian noise $\zb$ and post-processing function $\Gb$. In order to estimate $E(\Fb_{\Thetab}(\mathcal{M}(\xb')))$, we sample N noisy samples $\tilde\xb'(n) = \xb'+\tilde\zb(n)$ from $\mathcal{N}(\xb', \sigma^2\Ib)$ and feed them into the post-processing function $\Gb$ and the neural network $\Fb_\Thetab$. $E(\Fb_{\Thetab}(\mathcal{M}(\xb')))$ is estimated by $\frac{1}{N}\sum_{n=1}^{N}\Fb_{\Theta}(\Gb(\tilde\xb'(n)))$, and according to the Chernoff bound [@boucheron2013concentration], the error of this estimation is bounded by $$Pr(|\frac{1}{N}\sum_{n=1}^{N}\Fb_{\Theta, l}(\Gb(\tilde\xb'(n))) - E(\Fb_{\Thetab, l}(\mathcal{M}(\xb')))| < \epsilon) \sim \mathcal{O}(e^{-Nt^2})$$. In the certification stage, we rely on the main theorem from [@cohen2019certified], which gives the currently tightest bound: \[lem:lemma2\] Denote an mechanism randomized by Gaussian noise by $\mathcal{M}(\xb) = \Gb(\xb + \zb)$, and the ground-truth label by $l$. Define $f(\xb) = \argmax_k\Fb_{\Thetab, k}(\mathcal{M}(\xb))$. Suppose $\underline{p_A}$ & $\overline{p_B}$ satisfy $$\begin{aligned} Pr(f(\xb) = l) \geq \underline{p_A} \geq \overline{p_B} \geq \max_{i\neq l} Pr(f(\xb) = i),\end{aligned}$$ the $\ell_2$ robust radius is $R = \frac{\sigma}{2}(\Phi^{-1}(\underline{p_A}) - \Phi^{-1}(\overline{p_B}))$. Lemma \[lem:lemma2\] indicates that as long as $\|\xb' - \xb\|_2 < R$, $\argmax_i Pr(f(\xb) = i) = l$, [*i.e.*]{}, the prediction is correct. The algorithm using the above lemma for certification is detailed in Algorithm \[alg:randomized\_certification\]. In the next subsection, we will detail the post-processing function mentioned before. Neural Network $\Fb_\Thetab(\cdot)$, standard deviation of the additive Gaussian noise $\sigma$, skeleton action $\xb'$ (probably adversarial), number of noisy samples for inference of $n$. Sample $N$ samples from $\mathcal{N}(\xb', \sigma^2\Ib)$ $\rightarrow$ $\{\tilde\xb'(n)|n=1, 2,..., N\}$ Smooth $\tilde\xb'(n)$ by a $1\times5$ or $1\times7$ Gaussian filter $\rightarrow$ $\Gb(\tilde\xb'(n))$ Feed $\Gb(\tilde\xb'(n))$ into the network $\rightarrow$ $\Fb_\Theta(\Gb(\tilde\xb'(n)))$ **Output** $\argmax_l\sum_{n=1}^{N}\Fb_{\Theta, l}(\Gb(\tilde\xb'(n)))$ Neural Network $\Fb_\Thetab(\cdot)$, standard deviation of the additive Gaussian noise $\sigma$, original and adversarial skeleton action $\xb$ & $\xb'$, number of noisy samples for inference of $n$, a predefined confidence value p for hypothesis test (usually $95\%$). **Recognition:** Sample $N$ samples from $\mathcal{N}(\xb, \sigma^2\Ib)$ $\rightarrow$ $\{\tilde\xb(n)|n=1, 2,..., N\}$ Smooth $\tilde\xb(n)$ by a $1\times5$ or $1\times7$ Gaussian filter $\rightarrow$ $\Gb(\tilde\xb(n))$ Feed $\tilde\xb(n)$ into the (normally trained) network $\rightarrow$ $\Fb_\Theta(\Gb(\tilde\xb(n)))$ **Confidence Interval:** Compute the number (counts) top two indices in $\{\argmax_k\Fb_{\Theta, k}(\Gb(\tilde\xb(n)))~|~n=1,2,...,N\}$ $\rightarrow c_A, c_B$ Compute the lower bound for $p_A$ and the upper bound for $p_B$ by the method in [@goodman1965simultaneous] with confidence $p$ $\rightarrow \underline{p_A}, \overline{p_B}$. **Certification:** Compute the certified $\ell_2$ radius by $R = \frac{\sigma}{2}(\Phi^{-1}(\underline{p_A}) - \Phi^{-1}(\overline{p_B}))$. Output $\max\{R, 0\}$ **if** [*$p_A$ corresponds to the ground-truth label $l$*]{} **else** $-1$ (certified robust radius) Compare R with $\|\xb'-\xb\|_2$, and if $R$ is larger, then output the index corresponding to $c_A$ (if inputs have $\xb'$) Temporal Filtering by Gaussian Kernel ------------------------------------- After adding Gaussian noise to the skeleton actions, we propose to further smooths the action along the temporal axis by a $1\times5$ or $1\times7$ Gaussian filter. The intuition is that the adjacent frames in a skeleton action sequence are very similar to each other, and thus can be used as references to rectify the adversarial perturbations. Although this additional operation does not improve the certification results, we observe that it can help our defense become more compatible with a normally trained model than the original randomized smoothing method in [@lecuyer2018certified; @cohen2019certified]. Also, we argue that this simple operation is not usually used in previous work because it is not very suitable in the image recognition domain, where no adjacency information (along the temporal axis) is available. Experiments =========== Attack Performance {#sec:attack_performance} ------------------ #### Main Results The main results of our attack are shown in Table \[tab:attack\_performance\]. As we can see, our proposed attack can achieve 100% success rates with very small violation of the constraints. The averaged normalized bone-length difference is approximately $1\% \sim 2\%$, and the violation of the joint angles is smaller than $10^\circ$. [*Considering the skeleton data is usually noisy, this subtle violation is considered “very common" in real world.*]{} We also note that adversarial-sample generation under the untargeted setting is usually easier than that under the targeted setting since a targeted adversarial sample is guaranteed to be an untargeted adversarial sample, but not vice versa. This fact is also reflected by the results in Table \[tab:attack\_performance\]. Furthermore, in Fig. \[fig:visual\], we show the visualization result of an adversarial skeleton action (recognized as a normal action “drinking water") generated by our attack, which is almost visually indistinguishable from its original skeleton action (“kicking"). \[tab:trans\] #### Transferability In order to shed light on the transferability of our attack, we feed the adversarial skeleton actions generated on a HCN model to another HCN model and 2s-AGCN, respectively. In order to boost the transferability performance, we set $\beta$ as $0.01$ or $0.1$ to generate adversarial skeleton actions with larger perturbation. The attack success rates are given in Table \[tab:trans\]. [*Similar to 3D adversarial point clouds [@xiang2018generating], the transferability of the adversarial skeleton actions is also a little limited compared with adversarial images.*]{} #### Comparison with C&W Attack We use C&W attack as an example to shed light on the difference between our attack and the existing attacks. C&W attack has been demonstrated as a successful optimization-based adversarial attack in many application domains. However, since C&W attack mainly considers minimizing the $\ell_2$ distance between original and adversarial skeletons, it might easily violate the constraints, as shown in our simple case study (Table \[tab:comparison\]). Defense Performance ------------------- #### Empirical Results We demonstrate the performance of the defense for inference in Table \[tab:empirical\_defense\]. We set $\beta=1.0$ to generate adversarial examples, and set $N=50$ (Alg. \[alg:randomized\_defense\]), which is more smaller than the number of samples required for certification but can achieve good empirical performance, as shown in Table \[tab:empirical\_defense\]. [*it is much easier to defend adversarial skeleton actions under the targeted setting than the untargeted setting.*]{} Note that the accuracy of HCN on NTU-CV and NTU-CS is respectively $91.1\%$ and $86.5\%$ [@li2018co], and the accuracy of 2s-AGCN is respectively $95.1\%$ and $88.5\%$ [@shi2019two]. #### Certified Results Due to the high computational cost of the certification method (N=1000), we mainly evaluate the certification algorithm on HCN. The certified accuracy achieved by different levels of noise is shown in Fig. \[fig:certify\]. Note that we use the same level of noise to train the model as the noise for certification. [*As we can see, with sacrificing $10\%\sim20\%$ accuracy on the clean samples, the method is able to achieve about $50\%$ certified accuracy ($\ell_2 = 0.5$)*]{}. ![Certification accuracy on HCN[]{data-label="fig:certify"}](CV_certify.png "fig:"){width="0.48\linewidth"} ![Certification accuracy on HCN[]{data-label="fig:certify"}](CS_certify.png "fig:"){width="0.48\linewidth"} Additional Experimental Results ------------------------------- #### Additional visualization results Here we provide more visualization results. We use “drinking water" as the attack target because “drinking water" is a normal action, which looks completely different from the some violent/abnormal actions like throwing, kicking, pushing, and punching. Despite the obvious visual differences between “drinking water” and those abnormal actions, our attack can still fool the state-of-the-art models to recognize those abnormal actions as “drinking water” by imperceptible and reproducible perturbation. In Fig. \[fig:visual\_hcn\], we show that our attack can fool the HCN model to recognize the “throwing" and “kicking a person” actions as a normal action “drinking water" by imperceptible adversarial perturbation. Similarly, in Fig. \[fig:visual\_agcn\], we show that our attack can fool the 2s-AGCN model to recognize the “throwing” and “punching a person" actions as a normal action “drinking water". [*We also attach the more videos to show the original and adversarial skeleton actions in the supplementary material*]{}. These visualization results along with the quantitative results in Table 1 (in the paper) demonstrate that the perturbations are indeed imperceptible and reproducible. #### Kinetics Dataset Except for the NTU dataset, we also evaluate our attack on another popular dataset, [*i.e.,*]{} Kinetics-400 dataset under both the untargeted and targeted settings. As shown in Table \[tab:attack\_performance\], under the untargeted setting, our attack can achieve 100% attack success rates with very small violation of the constraints, similar to its performance on the NTU dataset. However, under the targeted setting, it is much more difficult for our attack to find targeted adversarial skeleton actions with very small violations of the constraints. This is because Kinetics-400 has 400 classes of actions, and the original NTU dataset only has 60 classes of actions. Also, we argue that the results on Kinetics under the targeted setting do not devalue our attack since, even for most of the clean testing samples from Kinetics, it is difficult for the state-of-the-models to predict their ground-truth labels (targets). ![image](hcn_throw.png){width="0.9\linewidth"} ![image](hcn_kick.png){width="0.9\linewidth"} ![image](visualize_throwing.png){width="0.9\linewidth"} ![image](agcn_punch.png){width="0.9\linewidth"} Conclusion ========== We study the problem of adversarial vulnerability of skeleton-based action recognition. We first identify and formulate three main constraints that should be considered in adversarial skeleton actions. Since the corresponding constrained optimization problem is intractable, we propose to optimize its dual problem by ADMM, which is a generic method proposed in this paper to generate adversarial examples with complicated constraints. To defend against adversarial skeleton actions, we further specify an efficient defensive inference algorithm and a certification algorithm. The effectiveness of the attack and defense is demonstrated on two opensource models, and the results induce several interesting observations that can help us understand the properties of adversarial skeleton actions. [^1]: The only parallel work is detailed in section \[subsec:adv\_attack\]. [^2]: We select these two models because the authors have released the code and hyperparameters on Github so that we can correctly reproduce the results. Also, these two models achieve fairly good performance.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The posit number system is arguably the most promising and discussed topic in Arithmetic nowadays. The recent breakthroughs claimed by the format proposed by John L. Gustafson have put posits in the spotlight. In this work, we first describe an algorithm for multiplying two posit numbers, even when the number of exponent bits is zero. This configuration, scarcely tackled in literature, is particularly interesting because it allows the deployment of a fast sigmoid function. The proposed multiplication algorithm is then integrated as a template into the well-known FloPoCo framework. Synthesis results are shown to compare with the floating point multiplication offered by FloPoCo as well. Second, the performance of posits is studied in the scenario of Neural Networks in both training and inference stages. To the best of our knowledge, this is the first time that training is done with posit format, achieving promising results for a binary classification problem even with reduced posit configurations. In the inference stage, 8-bit posits are as good as floating point when dealing with the MNIST dataset, but lose some accuracy with CIFAR-10.' author: - | Raúl Murillo Montero\ Department of Computer Architecture and Automation\ Complutense University of Madrid\ Madrid, 28040\ `[email protected]`\ Alberto A. Del Barrio\ Department of Computer Architecture and Automation\ Complutense University of Madrid\ Madrid, 28040\ `[email protected]`\ Guillermo Botella\ Department of Computer Architecture and Automation\ Complutense University of Madrid\ Madrid, 28040\ `[email protected]`\ bibliography: - 'references.bib' title: 'Template-Based Posit Multiplication for Training and Inferring in Neural Networks' --- Introduction {#sec:introduction} ============ Multiple floating-point representations have been used in computers over the years, although the IEEE Standard for Floating-Point Arithmetic (IEEE 754) [@IEEE754_1985] is the most common implementation that modern computing systems have adopted. Since it was established in 1985, the standard has only been revisited in 2008 (IEEE 754-2008) [@IEEE_Std], but the main characteristics of the original remain to keep compatibility with existing implementations and it is not adopted by all computer systems. However, multiple shortcomings have been identified in the IEEE 754 standard, which are listed below [@gustafson2017end]: - Different computers using the same IEEE floating-point format are not required produce the same results. When a computation does not fit into the chosen number representation, the number will be rounded. Even in the last revision of the standard they introduce the *round-to-nearest, ties away from zero* rounding scheme and provide recommendations for computations reproducibility, hardware designers are not coerced to implement them. Therefore, identical computations can lead to multiple results across different computing platforms [@kahan1998java]. - Multiple bit patterns are used for handling exceptions such as the *Not a Number* (NaN) value, which indicates that a value is not representable or undefined – for example dividing by zero results in a NaN. The problem is that the amount of bit patterns that represent NaN may be more than necessary, making hardware design more complex and decreasing the available number of exactly representable values. - IEEE 754 makes use of overflow – accepting $\infty$ or $-\infty$ as a substitute for large-magnitude finite numbers – and underflow – accepting 0 as a substitute for small-magnitude nonzero numbers. Thus, major problems can be produced, as the above mentioned. - Rounding is performed on individual operands of every calculation, so associativity and distributivity properties are not always held in floating-point representation. The last revision of the standard tries to solve this issue including the Multiply Accumulation (FMA) operation. However, again this may not be supported by all computer systems. The above listed shortcomings led to the idea of developing a new number system that can serve as a replacement for the now ubiquitous IEEE 754 arithmetic. At the beginning of 2017, John L. Gustafson introduced the *posit* number representation system, a Type III unum format [@gustafson2017end; @Jaiswal_2018_adder] that has no underflow, overflow or wasted NaN values. Gustafson claims that posits are not only a suitable replace for the current IEEE Standard for Floating-Point Arithmetic, but also provide more accurate answers with an equal or smaller number of bits and simpler hardware [@gustafson2017beating]. As it is illustrated in [@gustafson2017beating; @gustafson2017posit; @gustafson2016radical] there are important benefits when using posits, as a better dynamic range, accuracy, closure and consistency between machines than with conventional floating point. However, posits are still in development and there is still some controversy about their improvement [@deDinechin2019posits; @ughen2019HAL]. ![Layout of an $\langle n,es \rangle$ posit number.[]{data-label="fig:posit_format"}](images/Posit_format.pdf){height="1.2in"} The numerical value of a posit number $X$, whose bits are distributed as shown in Figure \[fig:posit\_format\], is given by Equation \[eq:posit\_decimal\_value\]: $$\label{eq:posit_decimal_value} X = (-1)^{s}*(2^{2^{es}})^{k}*2^{e}*1.f~,$$ where $k$ is the *regime* value, $e$ is the unsigned exponent (if $es$ &gt; 0), and $f$ is the mantissa of the number without the implicit one. In terms of format layout, the main differences with floating point are the existence of the regime field and together with the unsigned and unbiased exponent, if there exists such exponent field. The regime is a sequence of bits with the same value (*r*) finished with the negation of such value ($\overline{r}$). Provided that $X=x_{n-1}x_{n-2}...x_{1}x_{0}$, this regime can be expressed as Equation \[eq:posit\_regime\] shows. $$\label{eq:posit_regime} k = - x_{n-2} + \sum_{i=n-2}^{x_{i}\neq~x_{n-2}}(-1)^{1 - x_{i}}~.$$ In other words, the regime basically counts the number of occurrences of the bit labelled as $r$ in Figure \[fig:posit\_format\]. If $r=1$ then the regime is the number of 1’s minus 1, while if $r=0$, the regime is the negative value of the number of 0’s. For instance, if the regime is 4-bits wide, the value `1110` would be interpreted as $k=2$, while the value `0001` would be $k=-3$. Hence, detecting the leading 1’s and 0’s is critical for performing this step [@Kim_2018; @Kim_2019; @tavares2019design]. However, the main difficulty to detect the regime, and consequently unpack the posit, is that its length varies dynamically. In numbers close to zero, there will be many fraction bits and few regime bits and, in large numbers there will be many regime bits and fewer fraction bits, but there is no fixed amount of bits. Finally, it must be noted that the scaling factor defined by the regime ($k$) and the amount of exponent bits ($es$) is typically named $useed$. This variability of configurations that the posit numbers provide is an excellent opportunity to research in the design of efficient units able to implement operations among posits. The recommended formats for obtaining similar accuracy results with respect to floating point are: $\langle 8,0 \rangle$ (w.r.t *minifloat* [@minifloat]), $\langle 16,1 \rangle$ (w.r.t. half precision) and $\langle 32,2 \rangle$ (w.r.t. single precision). Nevertheless, the aforementioned variability of posit configurations opens a gate to look for functional units providing the best trade-offs [@Barrio2019; @barrio2017slack]. It is therefore desirable to provide a generic architecture and a generic flow to implement posit functional units. For this purpose, in this paper we leverage the FloPoCo framework [@de_Dinechin_2011] capabilities. FloPoCo (Floating-Point Cores, but not only) is an open-source C++ framework for the generation of arithmetic datapaths which provides a command-line interface that inputs operator specifications and outputs synthesizable VHDL specially suited for Field-Programmable Gate Arrays (FPGAs). In this paper we have defined a generic posit multiplication algorithm and integrated it as a template within the FloPoCo framework. This way, it is possible to generate synthesizable multipliers with any posit configuration, including $es=0$. The code is publicly available at github [^1]. Moreover, we have evaluated the performance of this algorithm and other posit operators in one of the scenarios where they can have a greater impact because of their accuracy with reduced formats: the Neural Networks (NNs). To the best of our knowledge, this is the first time that posits are evaluated in the training stage, providing promising results. In the inference phase, short posit formats match floating point for the MNIST dataset and lose some accuracy with CIFAR-10. The rest of the paper is organized as follows: Section \[sec:related work\] describes the state of the art regarding posit multipliers and other functional units; Section \[sec:posit multiplier\] presents our algorithm for multiplying in the posit format as well as its integration with the FloPoCo framework; Section \[sec:posits and nns\] presents a study on posits and NNs; Section \[sec:experiments\] provides synthesis and simulation results to validate the approach and finally Section \[sec:conclusions\] draws our conclusions and future lines of work. Related Work {#sec:related work} ============ Since the posit number system was introduced, the interest on a hardware implementation for this format has increased rapidly. Despite the short life time of posits, several hardware implementations have been proposed since 2017. The posit arithmetic unit proposed in [@Jaiswal_2018; @Jaiswal_2018_adder], includes floating-point to posit conversion, posit to floating-point conversion, addition/subtraction and multiplication. The work in [@Thesis:2018] improves the decimal accuracy of the latters and defines a posit vectorized unit. Although these works seem totally parametrized, there is no support for the case of zero exponent bits. This is important because in the Deep Learning scenario, the configuration $\langle 8,0 \rangle$ provides an extremely fast approach for the sigmoid function [@gustafson2017posit]. Another general design for posit arithmetic unit that includes adders and multipliers is presented in [@Chaurasiya_2018]. In contrast to the implementations shown in [@Jaiswal_2018; @Thesis:2018], the posit decoder proposed in this architecture uses only a leading zero detector for decoding the regime, while [@Jaiswal_2018; @Jaiswal_2018_adder] were using a leading one detector too. The work in [@jaiswal2019pacogen] improves the prior ones by employing just one leading one detector and provides a tool for generating different units with different posit formats. Authors in [@podobas2018hardware] present a C++ template compliant with Intel OpenCL SDK. Nonetheless, the configuration with $es=0$ is not included either in any of the aforementioned articles. J. Johnson [@Johnson2018] employs posit addition to complement the logarithmic multiplication when performing inference in Convolutional Neural Networks (CNNs), studying the $\langle 8,0 \rangle$ configuration too, but just in simulation. Other studies tackling the inference stage of CNNs are performed in [@carmichael2019date; @carmichael2019performance; @Langroudi2018PositNNTP]. In all these works, the training stage is performed in floating point and the weights converted to posit format, while the inference stage is performed in posit format. Finally, for the sake of completeness it must be noted that since the posit standard [@posit_standard] includes fused operations such as the fused dot product, and due to the importance of this operation for matrix calculus, some research and development for this kind of implementations has been done. Different matrix-multiply units for posits are presented in [@Chen_2018; @Thesis:2018; @carmichael2019date; @carmichael2019performance]. They make use of the *quire* register [@posit_standard] to accumulate the partial additions that are involved in the dot product, so the result is rounded only after the whole computation. Nevertheless, the fused operations are out of the scope of this work, as the functional unit that will be described is a posit multiplier. In this paper we propose an algorithm to perform the multiplication of two posit numbers and integrate it into a well-known framework as FloPoCo [@de_Dinechin_2011], providing support for generating synthesizable multipliers for any posit configuration, including the case $es=0$. Furthermore, we present a study in both the training and inference stages of NNs. As it there have been mentioned, there have been several studies for the inference stage, but not on the training phase. The Posit Multiplier {#sec:posit multiplier} ==================== At hardware level, posits were designed to be easy to compute, i.e., to have a circuitry similar to the existing floating point. The main encoding difference between float and posit formats is the fact that the second one includes a run-time varying scaling component. This leads to a format that has no fixed fields at run-time, which is a hardware design challenge. Below we present a fully functional posit multiplier operator. Analogously to performing computations with IEEE floats, it is necessary to unpack/decode the operands fields before carrying out any computation. Therefore, we first present the posit decoding process in Algorithm \[alg:1\]. It must be noted that the decoder can also be used in other arithmetic modules, e.g. a posit adder. The explanation of such algorithm is as follows: - Sign and special cases are detected checking the Most Significant Bit (MSB) and ORing the remaining bits, respectively (lines 2–5). - Since posit arithmetic uses 2’s complement for representing negative numbers, dealing with the absolute value simplifies the data extraction process. Therefore, the 2’s complement of the inputs are obtained, only if it is necessary, by XORing the input with the replicated sign bit and adding the sign to the Least Significant Bit (LSB) (line 6). - The $twos[N-2]$ bit aids to determine the regime value. In order to use only a leading zero detector [@Chaurasiya_2018], we invert the bits of $twos$ if the regime consists on a sequence of ones (line 8). Then we count the sequence of `0` bits terminating in a `1` bit using a leading zero detector module (line 9). - For extracting the exponent and the fraction bits, the regime is shifted out from $twos$, so the exponent is aligned to the left (line 10). - The first $es$ bits of the shifted string (if $es=0$ this instruction is omitted) correspond to the exponent bits (line 11), while the remaining bits correspond to the fraction (line 12). It must be noted that here the hidden bit is appended as the MSB. - The regime depends on the sequence of identical bits that constitute this field. The regime value is $zc-1$ when the bits are `1` (positive regime) or $-zc$ when it consists on a sequence of `0` bits (negative regime). Note that an extra `0` is added to maintain sign bit of the operation (line 13), as $zc$ is a positive value. $nzero \gets \bigvee in[N-2:0]$ $sign \gets in[N-1]$ $z \gets \neg (sign \vee nzero)$ $inf \gets sign \wedge \neg (nzero)$ $twos \gets (\{N-1 \{sign\} \} \oplus in[N-2:0]) + sign$ $rc \gets twos[N-2]$ $inv \gets \{N-1\{rc\}\} \oplus twos$ $zc \gets {\textsc}{LZD}(inv)$ $tmp \gets twos[N-4:0] \ll (zc-1)$ $exp \gets tmp[N-4:N-es-3]$ $frac \gets {nzero \text{ \& } tmp[N-es-4:0]}$ $reg \gets {rc \text{ ? } {\texttt{`0'} \text{ \& } zc-1} : -({\texttt{`0'} \text{ \& } zc})}$ $sign, reg, exp, frac, z, inf$ The process of posit multiplication is almost the same as for floating-point multiplication, i.e. the scaling factors are added and the fractions are multiplied and rounded. There are few differences when multiplying posits due to the regime field. The pseudocode for posit multiplication is shown in Algorithm \[alg:2\] and the explanation of the flow is as follows: $sign_A, reg_A, exp_A, frac_A, z_A, inf_A \gets {\textsc}{Decode}(in_A)$ $sign_B, reg_B, exp_B, frac_B, z_B, inf_B \gets {\textsc}{Decode}(in_B)$ $sign \gets sign_A \oplus sign_B$ $z \gets z_A \vee z_B$ $inf \gets inf_A \vee inf_B$ $sf_A \gets {reg_A \text{ \& } exp_A}$ $sf_B \gets {reg_B \text{ \& } exp_B}$ $frac_{mult} \gets frac_A \times frac_B$ $ovf_m \gets frac_{mult}[MSB]$ $norm_{frac} \gets {ovf_m \text{ ? } {\texttt{`0'} \text{ \& } frac_{mult}} : {frac_{mult} \text{ \& } \texttt{`0'}}}$ $sf_{mult} \gets ({sf_A[MSB] \text{ \& } sf_A}) + ({sf_B[MSB] \text{ \& } sf_B}) + ovf_m$ $sf_{sign} \gets sf_{mult}[MSB]$ $nzero \gets \bigvee frac_{mult}$ $exp \gets sf_{mult}[es-1:0]$ $reg_{tmp} \gets sf_{mult}[MSB-2:es]$ $reg \gets {sf_{sign} \text{ ? } - reg_{tmp} : reg_{tmp}}$ $ovf_{reg} \gets reg[MSB]$ $reg_f \gets {ovf_{reg} \text{ ? } {\texttt{`0'} \text{ \& } \{\lceil\log_2(N)\rceil\{\texttt{`1'}\} \}} : reg}$ $ovf_{regf} \gets \bigwedge reg_f[MSB-2:0]$ $exp_f \gets {(ovf_{reg} \vee ovf_{regf} \vee \neg nzero) \text{ ? } \{es\{\texttt{`0'}\}\} : exp}$ $tmp1 \gets {nzero \text{ \& } {\texttt{`0'} \text{ \& } {exp_f \text{ \& } {norm_{frac}[MSB-3:0] \text{ \& } \{N-1\{\texttt{`0'}\} \}}}}}$ $tmp2 \gets {\texttt{`0'} \text{ \& } {nzero \text{ \& } {exp_f \text{ \& } {norm_{frac}[MSB-3:0] \text{ \& } \{N-1\{\texttt{`0'}\} \}}}}}$ $shift_{neg} \gets {ovf_{regf} \text{ ? } reg_f - 2 : reg_f - 1}$ $shift_{pos} \gets {ovf_{regf} \text{ ? } reg_f - 1 : reg_f}$ $tmp \gets {sf_{sign} \text{ ? } tmp2 \gg shift_{neg} : tmp1 \gg shift_{pos}}$ $LSB, G, R \gets tmp[MSB-(N-1):MSB-(N+1)]$ $S \gets \bigvee tmp[MSB-(N+2):0]$ $round \gets {(ovf_{reg} \vee ovf_{regf}) \text{ ? } \texttt{`0'} : G \wedge (LSB \vee R \vee S)}$ $result_{tmp} \gets {{\texttt{`0'} \text{ \& } (tmp[MSB:MSB-(N-1)] + round)}}$ $result \gets {inf \text{ ? } infinity : {z \text{ ? } zero : {sign \text{ ? } - result_{tmp} : result_{tmp}}}}$ $result$ - When the two operands are decoded (lines 2–3), the sign and special cases are handled easily (lines 4–6). - The Scaling Factor (SF) of each operand consists of the regime and the exponent values, one after the other (lines 7–8). This is due to how posit decimal values are computed using regime and exponent. - The resulting fraction field is the outcome after multiplying the two operands fractions as if they were integer values (line 9). Recall that multiplying two $n-bit$ integers results in an integer of $2n$ bits. In addition, the decoder module returns fractions with the hidden bit as MSB, so the first two bits of the fractions multiplication do not strictly belong to the fraction field of the result, since they correspond to the multiplication of the hidden bits plus the possible carry bit due to fraction overflow. Therefore, the MSB of the result aids to detect any overflow when multiplying the fractions(line 10). - If a fraction overflow occurs, the resulting fraction has to be normalized shifting one bit to the right. In order to avoid losing any bit for rounding, instead of shifting, we just append a `0` bit as MSB, or as LSB if there is no overflow (line 11). - The resulting scaling factor is obtained by adding both operand scales, plus the possible fraction overflow. The result of adding two bit strings of same size may overflow, and in this case that carry bit indicates the sign of resulting regime, so it is necessary to replicate the MSB of both scaling factors before adding them (lines 12–13). - Exponent and regime are extracted from the scaling factors addition. The obtained regime may be negative, but it is more suitable to handle absolute values (lines 15–17). A similar action is performed in the decoding stage to simplify the following steps. - Adding two high-magnitude regimes may result in overflow, so in that case the regime is truncated to the maximum possible value and the exponent is set to 0 (lines 18–21). - Once the resulting fields have been computed and adjusted, they have to be packed in the correct order. To construct the regime correctly, the packed fields have to be right-shifted as a signed integer according to the sign and value of the regime. It is important to avoid losing any fraction bit to round correctly, so an amount of `0` bits has to be appended on the right (lines 22–26). - Posits, same as IEEE 754 floats, follow a *round-to-nearest-even* scheme. To perform a correct unbiased rounding, the LSB, Guard (G), Round (R) and Sticky (S) bits are needed [@Koren:1993] (lines 27–29). The rounded result is finally adjusted according to the sign and exceptions. Integration with FloPoCo {#subsec:integration flopoco} ------------------------ In this subsection we can briefly comment how to integrate the aforementioned algorithms within the FloPoCo framework [@de_Dinechin_2011]. FloPoCo follows an object-oriented class hierarchy, where all operators inherit from a baseline `Operator` virtual class. Thus, by extending such class and incorporating Algorithms \[alg:1\] and \[alg:2\] it is possible to create a posit multiplier with $n$ and $es$ as input parameters. Then, using the command `flopoco <options> <operator specification list>`, FloPoCo will generate a single synthesizable VHDL file [@de_Dinechin_2011]. Figure \[fig:FloPoCo\] illustrates this process for the command `flopoco PositMult N=8 es=1`, with which we obtain the VHDL code for a $\langle 8,1 \rangle$ multiplier, and changing the values on `N` and `es` we can obtain a new multiplier for any other posit configuration. [0.46]{} ![Generation of synthesizable VHDL from C++ code with FloPoCo.[]{data-label="fig:FloPoCo"}](images/in.pdf "fig:"){width="\textwidth"} [0.46]{} ![Generation of synthesizable VHDL from C++ code with FloPoCo.[]{data-label="fig:FloPoCo"}](images/out.pdf "fig:"){width="\textwidth"} It is important to mention that, in contrast with the works presented in [@Jaiswal_2018; @Jaiswal_2018_adder; @Chaurasiya_2018] which only provide implementations with a non-zero value for $es$, we designed a generic template that can be used to automatically generate multipliers for any posit configuration, not only those with $es > 0$. What is more, it is possible to generate combinational and sequential and even FPGA-customized versions of the multiplier by just changing the options when invoking FloPoCo. This will be shown in Section \[sec:experiments\]. Case Study: Posits and Neural Networks {#sec:posits and nns} ====================================== The recent surge of interest in Artificial Intelligence, and in particular in Deep Learning (DL), together with the limitations this sector currently has in terms of power consumption and memory resources make us wonder if posits can be helpful in this field. As described in this paper, the posit number system has many interesting properties, such as lack of underflow or overflow or the fast approximation of sigmoid function that some configurations of posits can do ($es=0$). These, along with the so-called *tapered precision* [@Langroudi2018PositNNTP], suggest that posits may be suitable for performing DL tasks. In a format with tapered precision the values mass around 0 and sparse to higher or lower numbers in less frequency, so representation of small values is more accurate than using other formats. When we use a number system with tapered precision, such as posits, the values follow a normal distribution centered in 0. That is the same distribution that DNN weight parameters usually follow, but even more grouped around 0. Figure \[fig:tapered\] illustrates this concept, which suggests that using posits for DNN may provide more accurate results. [0.475]{} ![Distributions of posit values and NN weights.[]{data-label="fig:tapered"}](images/posit_dist.pdf "fig:"){width="\textwidth"} [0.475]{} ![Distributions of posit values and NN weights.[]{data-label="fig:tapered"}](images/NN_weights.pdf "fig:"){width="\textwidth"} Training {#subsec:training} -------- The NN training with the posit format has been done using the posit-arithmetic library PySigmoid [@pysigmoid]. The choice of this particular package is due to the fact that it allows working with specific posit configurations, not only the “common” [Posit$\langle 8, 0 \rangle$]{}, [Posit$\langle 16, 1 \rangle$]{} and [Posit$\langle 32, 2 \rangle$]{}, and that it has a function that simulates the hardware operation for fast sigmoid, which approximates the original function when posits have $es=0$, in particular when using the [Posit$\langle 8, 0 \rangle$]{} configuration. According to [@gustafson2017posit], this fast sigmoid can be achieved by flipping the first bit of the posit and shifting it right two places, so given an $n$-bits posit $X$, this behavior can be modeled by Equation \[eq:fast\_sigmoid\], describes how the fast sigmoid is implemented for an $n$-bit input. $$\label{eq:fast_sigmoid} \sigma_{fast}(X,n) = X \oplus 2^{n-1}>> 2~.$$ To measure how well posits perform at deep learning tasks, the binary classification problem depicted in Figure \[fig:dataset\] has been studied in detail. The samples consist only of two features and classes are obviously separated by a non-linear boundary. ![Classification problem for posit training.[]{data-label="fig:dataset"}](images/NN_dataset.pdf){width="60.00000%"} The case study consists in training a NN architecture with two hidden layers of 4 and 8 neurons, respectively, to solve the aforementioned binary classification problem and check if posits may be suitable for training. Although there are multiple libraries and frameworks for Machine Learning (ML) that accelerate and simplify these kind of tasks, our test requires that all the internal computations involving parameters of the network are done in the posit format. Hence, the only option is to implement the NN from scratch, casting the input into posit type and replacing all the internal operands by the ones from PySigmoid library [@pysigmoid]. In this way the fused dot product with the quire accumulator can be employed too. In order to perform operations with the quire, the original library was modified and can be found at github [^2]. The original version of this library allowed underflow for certain values, which was fixed in the modified version we present in this paper. Inference {#subsec:inference} --------- As some research papers show [@Chen2014; @Gupta2015; @Courbariaux2014], it is difficult to apply lower numerical precision to the training of NNs, especially when using less than 16 bits. However, many research papers have shown that it is possible to apply low-precision computing to the inference stage of NNs after training with exact arithmetic [@rodriguez2018lower; @Jacob2017; @Hubara2016]. These results lead to the idea of using a reduced posit format such as [Posit$\langle 8, 0 \rangle$]{} for carrying out the DL inference. Performing low-precision inference can be extremely helpful in embedded systems and applications that make use of DL techniques such as Autonomous Driving [@Cococcioni2018]. There are some characteristics of the posit format that can be an advantage when using this format in NNs: - The comparison of posits uses the same hardware as for comparing integers, which is much faster than floats comparison. Thus, the typical pooling layers of CNNs could be easily implemented. - If using a format with $es=0$, e.g. [Posit$\langle 8, 0 \rangle$]{}, the sigmoid function can be approximately calculated as described by Equation \[eq:fast\_sigmoid\]. - Input values for NNs are usually normalized between \[-1,1\]. Because of the aforementioned tapered precision, the addition of two posit numbers is pretty accurate near 0. Experiments {#sec:experiments} =========== In this section the experimental results are presented. First, the synthesis results of template-based multiplier are discussed and second, the performance of posits operators is evaluated on NNs in both training and inference stages. Synthesis Results {#subsec:synthesis} ----------------- Prior to synthesizing the posit multiplier described in Section \[sec:posit multiplier\], the verification has been done as follows: first the golden solution has been obtained with the help of the Mathematica environment [@mathematica2019]. Afterwards, the VHDL the test bench is run using Xilinx Vivado Design Suite [@vivado] and the outputs compared with those from Mathematica, producing no mismatch. Different posit configurations have been successfully tested. In particular, complete tests have been done for $\langle 8,0 \rangle$, $\langle 8,1 \rangle$ and $\langle 8,2 \rangle$, and also, but less exhaustive, for $\langle 16,1 \rangle$ and $\langle 32,2 \rangle$. Several posit multipliers have been synthesized using Synopsys Design Compiler with a 65 target-library and without placing any timing constraint. The delay, area, power and energy of different posit multipliers have been measured. Besides synthesizing multipliers for different posit formats, the capabilities of FloPoCo have been leveraged to generate these units with different styles, namely: pipelined, combinational and combinational without employing hard multipliers nor DSP blocks. Also, several floating-point multipliers have been generated using FloPoCo as well. In this case, the notatio $\langle exp,mant \rangle$ indicates the amount of bits for the exponent and mantissa bits, respectively. Given that there is also a sign bit, the three explored FloPoCo floating point configurations would correspond with minifloat ($\langle 4,3 \rangle$), IEEE Half Precision ($\langle 5,10 \rangle$) and IEEE Single Precision ($\langle 8,23 \rangle$). Nevertheless, it must be noted that FloPoCo floating point does not handle exceptional cases (infinity, NaN) and subnormal numbers. Table \[tab:synthesis\] presents the synthesis results. In case of the pipelined designs, the number of stages is indicated between parenthesis next to the delay value. A first conclusion that can be extracted is the fact that posit pipelined designs are not optimized in terms of pipeline depth, as they have many stages in comparison with equivalent floating point units. For example, 8-bit multipliers require at least 7 stages, which is a lot for this kind of components. Second, the FloPoCo units are more efficient but, as it has been mentioned before, they are not complete while our multiplier is. In order to compare results with a state-of-the-art multiplier [@Chaurasiya_2018], we have used Xilinx Vivado for synthesizing our proposed multiplier on a ZedBoard Zynq-7000 SoC, the same target as in [@Chaurasiya_2018]. This comparison in terms of LUTs and DSPs is shown in Table \[tab:FPGA\]. In the case of the pipelined design, it must be also considered that extra resources are necessary. For example, for the case of 32-bits, 65 LUTRAMs, 910 FFs and one BUFG are required. ------------------------ ----------- ----- ----------- ----- Slice LUT DSP Slice LUT DSP **[@Chaurasiya_2018]** 218 1 572 4 Pipelined 321 1 891 2 Combinational 266 1 927 2 Combinational, No hm 266 1 1640 0 ------------------------ ----------- ----- ----------- ----- : Comparison of posit multipliers synthesis area results.[]{data-label="tab:FPGA"} As can be observed in Table \[tab:FPGA\], the results are not as good as the ones presented in [@Chaurasiya_2018]. Nonetheless, it must be observed that for 32-bits our implementations employ less DSPs. Furthermore, it must be reminded that our flow also supports the case of $es=0$. Neural Networks Training {#subsec:neural networks training} ------------------------ As it has been described in Section \[subsec:training\], different [Posit$\langle n, es \rangle$]{} configurations has been studied for the binary classification problem. Moreover, the posit configurations employing $es=0$ have utilized the fast sigmoid function too, while those with $es>0$ or floating point formats have made use of the regular sigmoid. The weights and biases are randomly initialized and the Mean Square Error (MSE) has been used as loss function to compare the outcomes of the network. An amount of 2500 epochs have been set to compare the losses of the different formats throughout the whole training. In this manner it is possible to compare whether the network converges or not and also how fast. Table \[tab:nn\] and Figure \[fig:nn\] depict the obtained results. -- -------- -------- -------- -------- -------- -------- 0 250 500 750 1000 1250 0.3701 0.2346 0.1726 0.0839 0.0023 0.0010 0.3701 0.2346 0.1727 0.1124 0.0023 0.0010 0.3681 0.1882 0.1491 0.1530 0.1530 0.1530 0.3653 0.2129 0.1359 0.0938 0.1478 0.1264 0.3650 0.2467 0.1758 0.1684 0.0140 0.0081 0.3648 0.2817 0.1716 0.1622 0.0645 0.0035 0.3337 0.1772 0.1453 0.0440 0.0019 0.0011 0.3337 0.1758 0.1658 0.0328 0.0017 0.0009 -- -------- -------- -------- -------- -------- -------- : Loss function during the NN training.[]{data-label="tab:nn"} ![Loss function along the NN training.[]{data-label="fig:nn"}](images/NN_training.pdf){width="80.00000%"} As can be seen, there is almost no difference in using single or double precision floats. Both [Posit$\langle 32, 2 \rangle$]{} and [Posit$\langle 16, 1 \rangle$]{} present the same behavior as floats, even with less MSE during the fists epochs. Posit configurations with 16, 14 and 12 bits takes some extra epochs to converge, and only those configurations with 10 and 8 bits present an irregular convergence. In these cases, the lack of underflow is undermining the posits convergence. In fact, this type of behavior has already appeared in a Newton-Raphson study presented at [@bachelor2018]. Therefore, it can be concluded that posits have converged as well as floats, even with some short formats employing the fast sigmoid approach. Although the proposed NN is a reduced example, these facts point in a good direction to train more complex CNNs. Neural Networks Inference {#subsec:neural networks inference} ------------------------- The performance of [Posit$\langle 8, 0 \rangle$]{} format is evaluated on two datasets: MNIST and CIFAR-10 run on the LeNet-5 architecture [@Lecun1998]. In this case, the networks are firstly trained with floating point arithmetic and then the weights are converted to posit format prior to the inference stage. The networks have been trained using Keras [@chollet2015keras] and TensorFlow [@tensorflow2015-whitepaper] frameworks. The posit computations during inference have been simulated with the help of a NumPy library version which includes a posit data type [@numpy]. In this case, computations are much faster than using the PySigmoid library, but there is not a fast sigmoid implementation, so the simpler ReLU module is used as activation function instead. The obtained results are shown in Table \[tab:CNN\]. -- ----------- ------------ ----------- ----------- ----------- ----------- Top-1 Top-5 Top-1 Top-5 Top-1 Top-5 $99.22\%$ $100.00\%$ $99.32\%$ $99.94\%$ $99.40\%$ $100\%$ $68.04\%$ $96.47\%$ $56.11\%$ $92.42\%$ $58.92\%$ $95.62\%$ -- ----------- ------------ ----------- ----------- ----------- ----------- : Performance on CNNs inference.[]{data-label="tab:CNN"} As can be observed, posits get slightly higher Top-1 and Top-5 accuracies than single precision floats. On the other hand, when dealing with a more complex dataset as CIFAR-10, there is a loss of $12\%$ in Top-1 and $4\%$ in Top-5. In order to mitigate this, a hybrid posit-float architecture has been considered. Following the idea described in [@Johnson2018], only the additions are performed in posit format. Under this scenario the accuracies are higher, but still floating point is superior. Finally, it must be emphasized that the posits employed in these tests are [Posit$\langle 8, 0 \rangle$]{}, which in exchange for the accuracy loss can reduce the memory footprint to a quarter in comparison to single precision, the functional units complexity and so on. Conclusions {#sec:conclusions} =========== In this work an algorithm for performing multiplication between two posit numbers has been presented. The algorithm is generic for whatever [Posit$\langle n, es \rangle$]{} configuration and it has been integrated into the FloPoCo framework. Furthermore, this multiplication algorithm, together with other posit operations, has been employed in the neural networks scenario for performing training and inference, obtaining promising trade-off results. In the future, further studies must be made in order to integrate posits to train larger NNs, such as CNNs, and to perform inference with higher accuracies. A possible direction may be combining different formats, as [@Johnson2018] proposed, bracing the transprecision concept. [^1]: <https://github.com/RaulMurillo/Posit-Multiplier_FloPoCo> [^2]: <https://github.com/RaulMurillo/PySigmoid/tree/master/PySigmoid>
{ "pile_set_name": "ArXiv" }
--- author: - 'L. Nortmann' - 'E. Pallé' - 'F. Murgas' - 'S. Dreizler' - 'N. Iro' - 'A. Cabrera-Lavers' bibliography: - 'paper\_h32\_arXiv.bib' date: 'Received 8 September 2015; Accepted 19 April 2016' subtitle: 'Confirmation of the flat transmission spectrum of HAT-P-32b' title: The GTC exoplanet transit spectroscopy survey IV --- Introduction {#chap:int} ============ Two decades after the first detection of an exoplanet around a solar-type star by @1995Natur.378..355M the field of exoplanet science is fast-moving and has expanded into many sub-fields. A new main focus is the characterization of exoplanet atmospheres. The most successful approach to studying the planet’s atmospheric properties has been the measurement of their transmission and emission spectra from multi-color observations of the occultation events in transiting planetary systems. While the emission of the planet can be inferred from the drop in flux during the secondary eclipse the planet’s transmission spectrum can be obtained during the primary eclipse. This is possible since the planet’s atmosphere will be opaque at wavelengths where the atmospheric constituents absorb light causing a larger effective planet radius and, thus, a deeper transit. Many successful measurements of wavelength-dependent planet radii have been obtained from space using the Hubble Space Telescope (HST) [e.g. @2002ApJ...568..377C; @2008MNRAS.385..109P; @2012ApJ...747...35B; @2015MNRAS.446.2428S]. Moreover, in the last 4 years, ground-based observations have also yielded promising results . However, both space-based and ground-based data often are affected by systematic noise signals, which need to be addressed before a high quality transmission spectrum can be extracted. In the past, the correct treatment of these noise signals has been subject of scientific debate and has led to disagreements between the conclusions of several groups studying the same data sets [e.g. @2007Natur.448..169T; @2007ApJ...668L.179E; @2009ApJ...699..478D; @2011MNRAS.411.2199G for HD 189733b]. As a general consequence this has created doubts concerning the robustness of presented results. In this paper we aim to demonstrate on the case of HAT-P-32b that reliable results for a planet’s transmission spectrum can be obtained from the ground. The hot Jupiter with a mass of $M=0.860 \pm 0.16$ $M_\mathrm{Jup}$ and a radius of $R= 1.789 \pm 0.025$ $R_\mathrm{Jup}$ was discovered by @2011ApJ...742...59H around an late-type F dwarf star (Vmag=11.44) at an $2.15$ day orbit. The planet’s dayside temperature was measured to be $T_\mathrm{eq}=2042 \pm 50$ K by @2014ApJ...796..115Z from secondary eclipse observations in the $H$, $K_S$, $3.6$ and $4.5$ $\mathrm{\mu m}$ bands. An optically close companion was discovered in [[email protected]] by @2013AJ....146....9A The stellar companion was recently studied in more detail and concluded to be an M-dwarf bound to the HAT-P-32 system from proper motion and AO measurements [@2015ApJ...800..138N; @2014ApJ...796..115Z]. Both studies place the effective temperature of the companion at about $T_\mathrm{eff}\approx3500$ K. Following the notation used in these works in the following we will refer to this stellar companion as HAT-P-32B and to the planet host star and the planet as HAT-P-32A and HAT-P-32Ab, respectively. @2014ApJ...785..126K observed the HAT-P-32 system among several other planet host stars for radial velocity (RV) trends that could indicate additional companions. They found a long trend signal for HAT-P-32A pointing to the existence of yet another body in the system. A transit timing variation (TTV) study of 45 transit events by @2014MNRAS.441..304S looking for evidence of an additional body found no evidence for variations larger than 1.5 [email protected] obtained a ground-based optical transmission spectrum of HAT-P-32Ab using Gemini North/GMOS. Their results show a flat transmission spectrum. In the study presented in this paper we probed a very similar wavelength range using the long slit method at the OSIRIS instrument at the 10-meter class telescope GTC aiming to verify the nature of the transmission spectrum and further demonstrate the potential of GTC/OSIRIS as a reliable survey instrument for observations of this kind. The paper is organized as follows. We first will describe the observing set up and data reduction in Sect. \[sec:obs\]. This is followed by a description of the white light curve analysis in Sect. \[sec:wlfit\] and a discussion of the white light curve results in Sect. \[sec:resdiswhite\]. Here we also address systematic noise signals we found in both data sets. In Sect. \[sec:systematics\] we will describe the source of the largest noise signal and motivate its correction for the narrow band light curves followed by a description of the extraction of the transmission spectrum during the analysis of the narrow band light curves in Sect \[sec:colorfit\]. We present and discuss our results for the transmission spectrum in Sect. \[sec:results\] and draw our conclusions in Sect. \[sec:conclusion\]. The study of photometric and spectroscopic data of the companion HAT-P-32B in order to derive its stellar parameters and extract values necessary for the correction of its diluting effect on the transit depth of HAT-P-32Ab can be found in the Appendix \[sec:sec\]. Observations and data {#sec:obs} ===================== Star RA DEC Vmag -------------- ----------------------- --------------------------------------- ------- HAT-P-32 $02$h $04$m $10.278$s $+46\degr$ $41\arcmin$ $16.21\arcsec$ 11.44 Ref1 (Run 1) $02$h $04$m $15.060$s $+46\degr$ $40\arcmin$ $49.57\arcsec$ 13.59 Ref2 (Run 2) $02$h $03$m $51.771$s $+46\degr$ $41\arcmin$ $32.23\arcsec$ 10.97 : Coordinates of the planet host star HAT-P-32A and the reference stars ‘Ref1’ and ‘Ref2’ used in the first and second observing run.[]{data-label="tab:coord"} We observed HAT-P-32Ab twice during transit on 2012 September 15 (JD 2456185.5, hereafter referred to as *Run 1*) and on 2012 September 30 (JD 2456200.5, hereafter referred to as *Run 2*) with the OSIRIS instrument (Optical System for Imaging and low-Intermediate-Resolution Integrated Spectroscopy; @2012SPIE.8446E..4TS) mounted at the Spanish 10.4 m telescope GranTeCan (GTC). We chose the method of long slit spectroscopy, in which the planet host star and a suitable reference star are both placed inside one long slit. The grism R1000R was used to disperse the light over the range from 518 to 918 nm. With an exposure time of 10 seconds (Run 1) and 7 seconds (Run 2) continuous time series of 321 (Run 1) and 700 (Run 2) optical spectra were obtained in each night covering the duration of the whole transit event in both cases. We used slightly different observing set-ups in each night, working with different regions of the CCD detector. The OSIRIS detector consists of a mosaic of two $2048\times4096$ pixel Marconi CCD42-82 chips. In Run 1 the chosen reference star ‘Ref1’ was considerably fainter ($\Delta$ Vmag $=2.15$) than HAT-P-32A and located at a distance of $56.0\arcsec=0.93\arcmin$. A custom made $12\arcsec$ wide slit was used and both stars were placed on CCD1. In Run 2 we chose a brighter reference star ($\Delta$ Vmag $=-0.467$) located at a $191.0\arcsec=3.18\arcmin$ separation from HAT-P-32A. Due to the larger distance between the stars, they could not both be placed in CCD1. In order to still have both stars on the same CCD and, thus, avoid possible complications from variations in the detector properties we placed both stars on CCD2 which has a larger unvignetted field of view through the slit than CCD1. We also exchanged the custom made $12\arcsec$ wide slit for a $10\arcsec$ wide slit, since the latter is slightly longer (extending $0.567\arcmin$ further into CCD2) giving both stars more room in spatial direction. The set up for both observing runs is illustrated in Fig. \[fig:setup\] and the coordinates for both reference stars are given in Table \[tab:coord\]. The observing conditions during both nights were good, with an average seeing of $1.06\arcsec$ in Run 1 and $1.12\arcsec$ in Run 2. The seeing was not stable in either run, varying between $0.78\arcsec$ and $1.82\arcsec$ during the first, and between $0.69\arcsec$ and $2.42\arcsec$ during the second night. Due to complications during the observation almost no out-of-transit data was obtained during Run 1 and part of the data was rendered useless by a light reflection passing over the detector contaminating the red part of the spectrum of HAT-P-32A for approximately 20 minutes (35 frames) during the second half of the transit. An example of a contaminated frame is shown in Fig. \[fig:reflec\]. After each run about 50 bias frames and 100 sky flats were taken. For wavelength calibration, spectra of the HgAr, Xe and Ne lamps were taken with a $1\arcsec$ wide slit. Data reduction and spectral extraction {#sec:specextract} -------------------------------------- We employed standard data reduction procedures, subtracting the median averaged bias and dividing by the median averaged flat field. We then applied a wavelength calibration to every column of the image re-binning the data to a homogeneous wavelength grid using the IDL routine `rebinw` from the PINTofALE package [@2000HEAD....5.2705K], guaranteeing flux conservation. This step ensured that every pixel in an image row corresponded to the same wavelength. The extraction of the stellar spectra from the images was then performed using the optimal extraction algorithm [@1986PASP...98..609H]. The algorithm performed well on our data sets, yielding lower noise levels than other, more simple approaches. During the extraction a median averaged spatial profile was used to identify and mask cosmic ray strikes. Due to the close projected distance of HAT-P-32A and the fainter M-dwarf companion (which was not yet discovered at the time of the observations) it was not possible to reliably exclude HAT-P-32B by choosing a narrow extraction aperture. Instead we choose a very wide aperture (80 pixel = 20.32) ensuring that HAT-P-32B was fully within this aperture at all times. As a consequence the diluting effect of its additional flux on the transit depth had to be corrected during transit modeling (see Sect. \[sec:wlfit\]).During the observations the stars drifted slightly in spatial and also in dispersion direction. We monitored the drift in spatial direction by fitting a Gaussian function to the stellar profile tracing the position of the peaks. In this step we also retrieved the FWHM of the fitted Gaussian profile to monitor the seeing variations. The drift of the stars in dispersion direction caused small shifts in the wavelength solution with respect to the one obtained for a star perfectly centered within the slit. We monitored and corrected these wavelength shifts by calculating the cross correlation of each spectrum with the first spectrum of the respective run. Light curves {#sec:lightcurves} ------------ Since any telluric variations during the observing runs will have affected both planet host star and reference star in the same manner, these effects can be neglected if only the relative light curves of these two objects are considered. We created white light curves from the data by dividing the total summed flux over all wavelengths for HAT-P-32A by the total sum of the reference star spectra for each measurement. The resulting white light curves are shown in Fig. \[fig:rawwlboth\]. We further created twenty narrow band channel light curves from the data of Run 2 by dividing the wavelength range into intervals of 20 nm. The spectra of HAT-P-32A and reference star Ref2 are shown in Fig. \[fig:spectra\_run2\] together with the channel limits. Channel \#13 encompassed both the telluric oxygen bands and the potassium resonance lines ( at $766.5$ and $769.9$ nm) predicted for exoplanet atmospheres at moderate temperatures. Akin to what has been reported by we found that the noise level, estimated from the standard deviation of the out-of-transit light curve scatter in each individual wavelength point, rises significantly in the deeper of the two telluric oxygen bands, negatively affecting the signal-to-noise of the entire channel \#13 light curve. We therefore constructed an additional channel (channel \#13b) of only $15$ nm width, spanning from $763$ to $788$ nm. This channel was still encompassing the expected potassium lines but avoided the high noise region as sketched in Fig. \[fig:c13b\]. Analysis of the white light curves {#sec:wlfit} ================================== The relative white light curves for both runs are shown in Fig. \[fig:rawwlboth\]. We found both white light curves to be affected by red noise. Previous works dealing with data obtained with GTC/OSIRIS have reported on similar noise signals . The three works explored different systematics models which included terms depending on the seeing and air mass. We also found indicators for a possible correlation of the red noise with auxiliary parameters and explored these possible correlations by including a systematic noise model in our light curve fitting. We found the minimal $\chi^2$ using the IDL implementation of the Levenberg-Markward algorithm `mpfit` by @2009ASPC..411..251M while modeling the light curves with a model of the form: $$\mathcal{M}=\left(\mathcal{T}+c_w\right)\cdot\mathcal{S}$$ where $\mathcal{T}$ is the analytical transit model described by @2002ApJ...580L.171M, $\mathcal{S}$ is a systematic noise model and $c_w$ is the relative flux contribution of the stellar companion HAT-P-32B in white light (i.e. $c_w=f_{\text{HAT-P-32B}}/f_{\text{HAT-P-32A}}$, where $f$ is the total flux integrated over the white light wavelength range). In the following we will refer to $c_w$ and its equivalents for the narrow band channels as the dilution factor. The @2002ApJ...580L.171M transit model parameterizes the transit using the radius ratio between the planet and the star $R_p/R_{\star}$, the semi-major axis of the planet’s orbit in units of the stellar radius $a_p/R_{\star}$, the inclination of the planet orbit $i$, a reference time for the mid-transit $T_\mathrm{C_1}$ and two coefficients $u_1$ and $u_2$ describing the stellar limb darkening with a quadratic limb darkening law. Also needed is the period of the planet’s orbit $P$, which we kept fixed to the value given by @2011ApJ...742...59H ($P = 2.150008$ days) and the eccentricity of the orbit $e$, which we kept fixed at $e = 0$. The code implementation of @2013PASP..125...83E was used for the calculation of the transit model. The relative flux contribution of the stellar M-dwarf companion $c_w$ and its uncertainty were determined from our own data of Run 2 as described in detail in Appendix \[sec:sec\] and this value was kept fixed during this optimization step. We tested 135 different systematic noise models $\mathcal{S}$ where each model was a different combination of polynomial functions depending on the four auxiliary parameters: position in spatial direction $xpos$, position in dispersion direction $ypos$, seeing/FWHM of the stellar profile $fwhm$ and air mass $airm$. All combinations of different polynomial orders between 0 and 2 (between 0 and 4 for the FWHM) for all four auxiliary parameters were explored. The most complex form of $\mathcal{S}$ we tested, consequently, was of the form: $$\begin{split} \mathcal{S}=~&n_0+\\ & x_1\cdot xpos+x_2\cdot xpos^2+\\ & y_1\cdot ypos+y_2\cdot ypos^2+\\ & a_1\cdot airm+a_2\cdot airm^2+\\ & f_1\cdot fwhm+f_2\cdot fwhm^2+f_3\cdot fwhm^3+f_4\cdot fwhm^4 \end{split}$$ where $n_0$ is the normalization and $x_{1,2}$, $y_{1,2}$, $a_{1,2}$ and $f_{1,2,3,4}$ are model parameters scaling the influence of the respective auxiliary parameter.In order to avoid over fitting due to the fact that the minimal $\chi^2$ generally decreases with a rising number of free parameters, a compromise between the number of free parameters and minimization of the $\chi^2$ should be found by the means of an objective criterion. We used the Bayesian Information Criterion for the case of an unknown variance of the data $$BIC =N\ln(RSS/N) + k \cdot \ln(N) \label{eq:bic}$$ where $RSS$ is the sum of squared residuals, $k$ is the number of free parameters and $N$ is the number of data points, to determine the best choice of $\mathcal{S}$. We compared the $BIC$ of all tested models and found that using a combination of a first-order polynomial of the position drift in dispersion direction ($ypos$), a first-order polynomial of the air mass and a third-order polynomial function of the FWHM yields the lowest BIC for the data of Run 1, i.e. a model of the form: $$\begin{split} \mathcal{S}_{\mathrm{Run 1}}^*=~&n_0+y_1\cdot ypos + a_1\cdot airm+\\ & f_1\cdot fwhm+f_2\cdot fwhm^2+f_3\cdot fwhm^3 \end{split} \label{eq:syswr1}$$ An abbreviated compilation of the results of the $BIC$ comparison can be found in Table \[tab:bich321\], where, for practicality, only the models with an $\Delta BIC=BIC- BIC_{\text{min}} <10$ are listed. The best fit results for the planet-to-star radius ratio derived by using these model approaches differ only slightly, and all lie well within the $1\sigma$ uncertainty interval (derived in Sect. \[sec:wlerror\]) of the best fit result obtained with the lowest BIC model. In Table \[tab:bich321\] the models are identified by a short code, which is to be read as follows: ‘xpos’ is the keyword for the position drift in spatial direction, ‘ypos’ for the drift in dispersion direction, ‘airm’ for the air mass and ‘fwhm’ for the seeing/FWHM of the stellar profile. The number following each of these keywords indicates the highest polynomial order that was allowed to be non-zero and free in the model fit. Using the systematic noise model $\mathcal{S}_{\mathrm{Run 1}}^*$ given in Eq. \[eq:syswr1\] we achieved a good fit to the data of Run 1 with an almost Gaussian distribution of the residuals (with a normalized standard deviation of $463$ ppm). The remaining correlation of the residuals is explored in the Sect. \[sec:wlerror\], which focusses on the error estimates of the results. A plot of the white light curve with the best fit model can be found in Fig. \[fig:white\_fit\_r1\]. Model $\Delta {BIC}$ $\Delta R_p/R_\star$ ------------------------- ---------------- ---------------------- xpos0 ypos0 airm0 fwhm3 5.76  0.000586 xpos0 ypos0 airm1 fwhm3 4.28  0.000292 xpos0 ypos0 airm1 fwhm4 9.63  0.000332 xpos0 ypos1 airm0 fwhm3 7.96  0.000399 xpos0 ypos1 airm1 fwhm2 6.92  0.000050 xpos0 ypos1 airm1 fwhm3 **0.00**  0.000000 xpos0 ypos1 airm1 fwhm4 5.70 -0.000125 xpos0 ypos2 airm1 fwhm3 4.63  0.000019 xpos1 ypos0 airm0 fwhm3 7.23  0.000300 xpos1 ypos0 airm1 fwhm3 9.77  0.000443 xpos1 ypos0 airm2 fwhm3 7.46  0.000492 xpos1 ypos1 airm0 fwhm3 5.54 -0.000456 xpos1 ypos1 airm1 fwhm3 5.69  0.000046 xpos1 ypos2 airm0 fwhm3 9.77 -0.000346 : Model comparison for the white light curve of Run 1. Different model approaches for the systematic noise and the respective values for $\Delta {BIC}= {BIC}- {BIC}_{ \mathrm{min}}$ and the respective change in best fitting planet-to-star radius ratio $\Delta R_p/R_\star= R_p/R_\star- \left(R_p/R_\star\right)_{{BIC}_{ \mathrm{min}}}$.[]{data-label="tab:bich321"} After repeating this process for the data of Run 2 we found the minimal $BIC$ for a systematics model consisting of a first-order polynomial of the position drift in spatial direction ($xpos$), a first-order polynomial of the air mass and a third-order polynomial function of the FWHM. However, neither this model nor any other tested combination did yield a good fit to the whole curve leaving the residuals still strongly correlated (see Fig. \[fig:white\_fit\_r2\]). This lead us to conclude that the systematic noise, which is strongly distorting the first part of the light curve of Run 2, has a different origin. We will investigate this origin in Sect. \[sec:systematics\]. Errors {#sec:wlerror} ------ For the white light curve fit of Run 1 we probe the probability distributions of the model parameters with Markov Chain Monte Carlo (MCMC) sampling using the `emcee` package [@2013PASP..125..306F]. We used non-informative priors for all parameters except the dilution coefficient $c_w$, which we allowed to vary within its uncertainties using a Gaussian prior. The MCMC sampling was run with an ensemble of 600 walkers each starting with slightly different parameter populations as seeds for the chains. We let the chains run for 20000 accepted steps and defined a burn-in phase, in which the chains are not yet fully converged, to be over after 5000 steps omitting all chain-steps prior to this mark. The results can be considered robust if all the chains converge to the same probability distribution, which they did in our case. We determined a thinning factor for each chain as the largest autocorrelation length of any parameter within the chain. The average thinning factor was $\approx 200$ which resulted in a total of 45891 accepted sample points for the merged distribution of an accepted chain steps. The correlation plots of the posterior distributions for all parameters of the white light curve fit of Run 1 are shown in Fig. \[fig:confidence\_intervals\_h321\_wlight\].The $1\sigma$ uncertainties of each parameter were calculated as the limits encompassing $68.27\%$ of all sampled points. ### Red noise estimation {#sec:rednoise} In order to determine the level of red noise remaining in the light curve residuals and consequent underestimation of the parameter uncertainties, we calculated the red noise factor $\beta$. This factor was introduced by @2008ApJ...683.1076W and is based on a comparison of the progression of the standard deviation of the time-binned light curve residuals with the behavior expected for a light curve purely affected by Gaussian (i.e. white) noise. The latter is described by Eq. \[eq:sigmabeta\] if $\sigma_1$ is the amplitude of the Gaussian white noise, $N$ the number of adjacent points binned together and $M$ the total number of bins. $$\sigma_N^\mathrm{theory}=\frac{\sigma_1}{\sqrt{N}}\sqrt{\frac{M}{M-1}} \label{eq:sigmabeta}$$ The difference between theoretical and actual progression of the standard deviation of the binned residuals is caused by red noise. @2008ApJ...683.1076W define the ratio between the two curves as the ‘red noise factor’ $\beta=\sigma_N^\mathrm{actual}/\sigma_N^\mathrm{theory}$ which can be used to inflate the MCMC chain derived error bars of all parameters. We determined the maximal value of this ratio to be $\beta_{\mathrm{Run 1}}=1.345$ for the white light curve data from the first run when the systematics model with the lowest $BIC$ was used and $\beta_\mathrm{Run 2}\geq6.5$ for the data from Run 2, regardless of which systematics model $\mathcal{S}$ was used. The $\beta$ factor close to unity for Run 1 demonstrates that the systematics model $\mathcal{S}_{\mathrm{Run 1}}^*$ (Eq. \[eq:syswr1\]) used was appropriate to correct all noise sources. In contrast, the high value derived for Run 2 clearly indicates a remaining noise source (see Sect. \[sec:systematics\]), suggesting the model approach used for Run 1 to be insufficient for the analysis of the data of Run 2. Results and discussion of the white light curves {#sec:resdiswhite} ================================================ \#1\#2\#3 --------------------------------------------------------- -------------- ------------------------------------- ---------------------- ------------------------------------------------------- -- Parameter This work @2011ApJ...742...59H ($e\equiv0$) @2013MNRAS.436.2974G @2014MNRAS.441..304S Filter/Range $518-918$ nm $I$, $z$, $g$ $520.6-932.5$ nm $r_\mathrm{S}$, $R_\mathrm{B}$, $R_\mathrm{C}$, Clear $R_\mathrm{p}/R_{\star}$ $a_\mathrm{p}/R_{\star}$ $i$ $P$ \[d\] (fixed) (fixed) $T_\mathrm{C_1}-2456185 [\mathrm{BJD}_{\mathrm{UTC}}] $ $e$ (fixed) (fixed) (fixed) (fixed) $u_1$ $\tol{I~\mathrm{ band}=0.2045}{}{}$ $\tol{R \mathrm{~band}= 0.28}{}{}$ $u_2$ $\tol{I \mathrm{~band}=0.3593}{}{}$ $\tol{R \mathrm{~band}= 0.35}{}{}$ $T_\mathrm{C_2}-2456200 [\mathrm{BJD}_{\mathrm{UTC}}]$ --------------------------------------------------------- -------------- ------------------------------------- ---------------------- ------------------------------------------------------- -- The results for the white light curve fit to the data from Run 1 are listed in Table \[tab:wlresult\] together with the literature parameters of recent high precision studies of HAT-P-32Ab. The error bars of these results were inflated with the red noise factor of $\beta=1.345$ derived in Sect. \[sec:wlerror\]. We find our results for the planet parameters to be consistent with the literature values. The difference in radius ratio compared to the results of @2011ApJ...742...59H and @2014MNRAS.441..304S might be explained by the different wavelength region probed in this study but is more likely caused by the transit depth dilution from additional flux of the M-dwarf companion HAT-P-32B, which was not accounted for in these two studies.A rough dilution correction of the given radius ratios using $$\left(R_p/R_{\star}\right)_{\mathrm{corrected}}=\left(R_p/R_{\star}\right)_{\mathrm{uncorrected}}\cdot\sqrt{1+c_{\mathrm{filter}}} \label{eq:roughdi}$$ with the dilution factor values $c_i'=0.006 \pm 0.002$, $c_z'=0.012 \pm 0.004$, $c_g'<0.0018$ provided by @2014ApJ...796..115Z in similar broadband filters as the ones used in these two studies, yields comparable radius ratios to the one derived by us and @2013MNRAS.436.2974G. Transit timing comparison {#sec:ttv} ------------------------- In order to compare our best fit transit time, $T_{\mathrm{C}_1,2456185}$ with literature predictions we use the ephemeris given in each paper to calculate the expected transit time and its uncertainty using Eq. \[eq:ephem\]. $$T_{\mathrm{C}}=T_{0}+E\cdot P \label{eq:ephem}$$ Where $T_{0}$ is a reference transit time, $E$ the epoch and $P$ the orbital period given in the respective paper. We find that our result for the transit time best agrees with the prediction given by the discovery paper [@2011ApJ...742...59H]. Using their ephemeris information and our measured transit time to refine the period we yield $P_{\mathrm{new}}=2.15000806(24)$ days.We also calculated the predicted transit time for Run 2 $T_{\mathrm{C}_2,2456200}$, which we could not reliably measure from a white light curve transit fit. To calculate the prediction for $T_{\mathrm{C}_2,2456200}$ from our Run 1 results we used Eq. \[eq:ephem\] with $T_{0}=T_{\mathrm{C}_1,2456185}$, $E=7$ and $P=P_{\mathrm{new}}$. The results are given in Table \[tab:wlresult\]. Comparison with theoretical limb darkening {#sec:limbd} ------------------------------------------ As the limb darkening coefficients are wavelength-dependent and cannot be well compared with literature results obtained in different filters, we compared them to theoretical values. To derive these theoretical coefficients we calculated the wavelength-dependent theoretical limb darkening profiles for a star with the basic stellar properties of HAT-P-32A (\[Fe/H\] $=-0.04\pm0.08$, $\log g$ (cgs) $=4.33 \pm 0.01$, @2011ApJ...742...59H, and $T_{\mathrm{eff}} =6269\pm64$ K, @2014ApJ...796..115Z) by interpolation from the PHOENIX specific intensity spectra library by . We then weighted each wavelength with its actual contribution to the measured stellar flux during observation, taking into account the instrument response function and the telluric absorption and summed all contributing limb darkening profiles to derive the theoretical white light profile. We then renormalized the model information so that $\mu=0$ actually corresponds to the outer edge of the star i.e. the region where the mean optical depth corresponds to unity. This is not the case for the raw model data due to the spherical symmetry assumption used in the PHOENIX code. We repeated this process varying the adopted stellar parameters for HAT-P-32A within their reported errors to explore the uncertainties of the intensity profile. Finally the theoretical limb darkening coefficients and their errors were derived by fitting the intensity profiles with a quadratic limb darkening law. We found that our best fit limb darkening coefficients were lower than the values predicted by the PHOENIX stellar models ($u_{1,\mathrm{ theory}}=0.340\pm0.056$ and $u_{2,\mathrm{ theory}}=0.245\pm0.073$). This difference might be caused by insufficiencies of the PHOENIX models, errors in the assumed stellar parameters or undetected systematic noise in the white light curve of Run 1. Consequences for the retrieval of the planet’s transmission spectrum -------------------------------------------------------------------- Despite the good agreement of our Run 1 white light curve results with the literature data, the lack of extensive out-of-transit data lead us to deem the data set unsuitable for a transmission spectroscopy study where a reliable measurement of very small changes in the transit depth is essential.Further, we considered the results obtained for the white light curve fit of the Run 2 as unreliable since a clear and dominant systematic noise signal in the data remained uncorrected (see Sect. \[sec:rednoise\]). We will further investigate this noise signal in Sect. \[sec:systematics\] and motivate a correction for the narrow band channel data of Run 2 which then can be used to derive a transmission spectrum of HAT-P-32Ab. GTC/OSIRIS instrument specific systematic noise affecting the data of Run 2 {#sec:systematics} =========================================================================== In Sect. \[sec:wlfit\] we determined that the systematics affecting the data of Run 2 cannot be sufficiently corrected by solely using a similarly simple de-correlation function of auxiliary parameters as was sufficient for the data of Run 1. The largest non-astrophysical systematic present in the white light curve of Run 2 is the large distortion of the first half of the light curve. It is present in all narrow band channel light curves (see Fig. \[fig:raw\_and\_corrected\_curves\_blue\], left panel). When studying the raw white light curves of Run 2 we found that the distortion in the relative light curve coincides with a sinusoidal-like feature found in both the raw light curve of HAT-P-32A and the reference star (see Fig. \[fig:whitelight\_h322\_raw\_curves\]). This particular and slightly curious feature in the first half of the observation does not appear to be caused by telluric atmosphere variations. A closer investigation of the two ‘bumps’ showed that while they are present in the data of both stars, they are slightly shifted in phase (and possibly have a different amplitude). As a consequence they do not cancel out when the ratio of both curves is taken to correct any telluric effects. We found that the ‘bumps’ re-appear with a frequency of roughly $60\degr$ of the instrument rotation angle. To better visualize our findings we divided the raw light curve of HAT-P-32A by a theoretical transit light curve using our best fit transit parameters from Run 1, cleaned both raw light curves from an air mass trend and plotted them over the instrument rotation angle (see Fig. \[fig:whitelight\_h322\_raw\_curves\_vs\_rot\]). We fitted all four ‘bumps’ with a Gaussian function to derive an estimate for the values of rotation angle at which the peak flux occurs. Since the instrument rotation angle changed during the first half of the observing run and then stayed nearly constant (see Fig. \[fig:rotan\_run2\]) this lead to the ‘bumps’ only manifesting in the first half of the run. The $60\degr$ symmetry of the ‘bump’-feature prompted us to suspect these flux amplitude variations to be caused by vignetting in pupil space. A vignetting of the pupil would reduce the overall amount of captured light that reaches the detector. Due to the hexagonal shape of the GTC primary mirror the pupil also is not invariant under rotation but exhibits a $60\degr$ rotational symmetry. As a consequence, the amount of vignetted pupil area would not only depend on the distance of the projected pupil but also on its effective rotation angle to the source of the vignetting. Vignetting could for example occur at the secondary and/or tertiary mirror for off-axis rays. It is indicated in the GTC conceptual design document [@GTCCD Fig. 4.4] that this should only have a significantly impact on targets located at separations larger than $\approx4\arcmin$ from the telescope optical axis (which includes Ref2 but not HAT-P-32A in Run 2, see Fig. \[fig:setup\]). Potential vignetting caused by other objects located in the optical path in pupil space might, however, also affect targets located closer to the telescope pointing.We constructed a simple toy model, to simulate the expected footprint of pupil vignetting at a hypothetical source. The effect predicted by this toy model exhibited a $60\degr$-periodicity with rotation angle very similar to the signal found in our data of Run 2 (see Fig. \[fig:theo\_vignetting\]). The toy model further showed a dependency of the signal on the initial target position relative to the telescope optical axis, which was regulating the phase, amplitude and precise shape of the effect. The true signal affecting the data could be more complex in case the source of vignetting is located off-center from the telescope optical axis (divergent from the toy model assumption) or if the signal is a superposition of several vignetting signals.The difference in systematic noise signal amplitude between the two stars of Run 2, which are separated by more than $3\arcmin$ ($\equiv$ 1/3 of the OSIRIS field of view) appears to be small, suggesting that for any given observing set up no signals with significantly larger amplitudes (i.e. larger than $1\%$) are to be expected. Systematic noise signals of amplitudes slightly lower than $1\%$ might be negligible for many other science cases. In our case, however, noise signals of this amplitude are more than two orders larger than the astrophysical signals we are aiming to detect and, therefore, need to be addressed. Since the amount of vignetting is only dependent on the projected pupil position and effective pupil rotation angle it should be wavelength-independent as long as the vignetting occurs before the light has passed any dispersing optical elements. In this case the produced systematic noise signal will affect all narrow band channel light curves and the white light curve in exactly the same way and should divide out if the ratio between any color light curve and the white light curve is taken. Analysis of the color light curves of Run 2 {#sec:colorfit} =========================================== We proceeded by dividing all narrowband light curves of Run 2 by the white light curve. This served to cancel out all wavelength-independent systematic noise signals, i.e. signals which affected all curves in an identical manner. Such common-mode corrections are usual practice when dealing with spectrophotometric transit data [@2013MNRAS.436.2974G; @2015MNRAS.446.2428S]. In accordance with expectation the correction appears to fully dispose of the light curve distortion we suspect to be caused by the wavelength-independent pupil vignetting (see Sect. \[sec:systematics\]). All remaining (i.e. wavelength-dependent) noise signals in the narrow band light curves are found to be linked with auxiliary parameters of the observation, in a similar manner as it was the case for the white light curve of Run 1 (see Sect. \[sec:wlfit\]). We fitted all differential light curves with a model of the form: $$\mathcal{M}_{\mathrm{diff},n}=\left(\mathcal{T}_n+c_n\right)/\left(\mathcal{T}_w+c_w\right)\cdot\mathcal{S}_n \label{eq:mdiff}$$ where $\mathcal{T}_n$ is the analytical transit model for the narrow band light curve, $c_n$ the dilution factor (i.e. relative flux contribution of the stellar companion HAT-P-32B) in this particular narrow band, $\mathcal{T}_w$ the model for the white light transit and $c_w$ dilution factor in white light and $\mathcal{S}_n$ a systematic noise model for all remaining wavelength-dependent noise sources. The latter was of the same form as the systematic model used during the white light curve fit, i.e. a combination of polynomial functions of the auxiliary parameters detector position, air mass and FWHM. Alternatively the original relative light curves can be fitted by a model of the form: $$\mathcal{M}_n=\left(\mathcal{T}_n+c_n\right)\cdot \mathcal{S}_{\mathrm{CM}}\cdot\mathcal{S}_n$$ where $$\mathcal{S}_{\mathrm{CM}}=\frac{\mathcal{D}_w}{\left(\mathcal{T}_w+c_w\right)}$$ is the common mode systematic noise signal derived from the residuals of the white light curve data $\mathcal{D}_w$ of Run 2. This approach is mathematically identical to the one given in Eq. \[eq:mdiff\].In both cases all wavelength-independent transit model parameters (i.e. $a_p/R_{\star}$, $i$, $T_\mathrm{C_2}$, $P$, the white light planet-to-star radius ratio, the white light limb darkening coefficients and the dilution parameters $c_w$ and $c_n$), were kept fixed in this optimization. The values for $c_n$ were determined as explained in Appendix \[sec:sec\]. We chose to fix the white light and wavelength-independent transit parameters to the values we derived from Run 1 in Sect. \[sec:wlfit\], as they represent a measurement taken with the same instrument and at the same wavelength interval. Not fixing the values but letting them vary in their uncertainty intervals did not change our final results for the relative wavelength-dependent change in radius ratio, but did only offset the exact value around which the narrow band radius ratios and limb darkening coefficients varied. The same held true when we fixed the wavelength-independent and white light transit parameters to the literature values provided by @2013MNRAS.436.2974G, @2011ApJ...742...59H or @2014MNRAS.441..304S. Between these tested configurations adopting our best fit Run 1 white light parameters did yield the lowest overall $\chi^2$. We, therefore, adopted the values resulting from these white light curves values as our final results. To allow an independent use of our derived transmission spectrum, we also provide the relative change in transit depth with respect to the white light curve transit depth, i.e. $\left(R_p/R_{\star}\right)^2-\left(R_p/R_{\star}\right)^2_\mathrm{white~light}$, together with our results for the absolute values of the radius-ratios in Table \[tab:tspec\_results\]. These differential transit depths are independent of the white light curve parameters chosen during the analysis and, therefore, free of any possible systematic errors caused by the uncertainties inherent to these parameters.We again tested all different combinations of systematics models. This time we allowed the highest polynomial order for the air mass and spatial and dispersion position to be 2 and for the FWHM to be 3. The $BIC$ was again calculated using Eq. \[eq:bic\]. As we were aiming to determine the model which best explained all 20 channels simultaneously, the squared sum of all 20 channel’s joined light curve residuals (i.e. channel \#1 - \#20, excluding the sub channel \#13b) were used for the calculation. The overall lowest $BIC$ was reached by using the model: $$\begin{split} \mathcal{S}_n=~&n_0+x_1\cdot xpos+y_1\cdot ypos+ a_1\cdot airm\\ & f_1\cdot fwhm+f_2\cdot fwhm^2 \end{split}$$ where $n_0$, $x_1$, $y_1$, $a_1$ and $f_{1,2}$ are wavelength-dependent parameters different for every channel. An overview of the $BIC$ comparison can be found in Table \[tab:bich322\] where all models with a $\Delta BIC < 1000$ are listed. Model $\Delta {BIC}$ ------------------------- ---------------- xpos0 ypos1 airm1 fwhm2 114.82 xpos0 ypos1 airm1 fwhm3 211.20 xpos0 ypos2 airm1 fwhm2 282.08 xpos1 ypos1 airm1 fwhm2 **0.00** xpos1 ypos1 airm1 fwhm3 77.08 xpos1 ypos2 airm1 fwhm2 170.87 xpos1 ypos2 airm1 fwhm3 240.33 xpos2 ypos1 airm1 fwhm2 179.62 xpos2 ypos1 airm1 fwhm3 255.41 xpos2 ypos2 airm1 fwhm2 350.57 xpos2 ypos2 airm1 fwhm3 418.44 : Model comparison for the systematic noise in the narrow band channel curves of Run 2 listing the respective $\Delta {BIC}= {BIC}- {BIC}_{ \mathrm{min}}$ values.[]{data-label="tab:bich322"} When investigating the individual channels separately we found that most of the color light curves were satisfied with a less complex model, without a spatial position ($xpos$) dependent term. But the $\chi^2$ of the three bluest channels was significantly minimized by introducing this additional parameter, leading to an overall favored $BIC$ without affecting the final results for the other channels.The results from the different models listed in Table \[tab:bich322\] are plotted in Fig. \[fig:bicrprs\] as green lines. The values we adapted as our final results are shown as black diamonds. It can be seen that at the blue end of the spectrum a few model give notably different results. The models yielding such deviant results are, however, not the ones favored when determining the best model for each channel individually. The resulting radius-ratios for the individually determined best models (red squares) are in good agreement with the adopted results derived from the homogeneous analysis. Errors {#sec:errh22} ------ We repeated the exploration of the posterior parameter distributions with MCMC as described for the white light curve analysis in Sect. \[sec:wlerror\], now letting the values for the wavelength-dependent dilution factors $c_n$ vary within their uncertainty intervals using Gaussian priors. This time we used 300 walkers and let them run for 10000 accepted steps. We discarded all points of the burn-in period, which was over after 1000 steps. The typical thinning factors for the chains were $\approx 82$, leaving us with a final distribution of an average of 33300 points for each wavelength channel light curve fit. An example of the resulting posterior parameter distributions is show for channel \#9 in Fig. \[fig:confidence\_intervals\_h322\_bin8\]. We then calculated the red noise factor for every narrow band channel light curve as described in Sect. \[sec:rednoise\] and inflated the MCMC derived error bars accordingly. The resulting values for $\beta$ lie between $1.022$ and $1.950$ and the resulting uncertainties of the wavelength-dependent radius ratio after multiplication with $\beta$ lie between $337$ and $972$ ppm. The light curves before and after the full systematic noise correction are shown in Fig.\[fig:raw\_and\_corrected\_curves\_blue\]. The resulting radius ratios and limb darkening coefficients as well as $\beta$ factors are given in Table \[tab:tspec\_results\]. \#1\#2\#3 Channel $\lambda$ \[nm\] $R_p/R_{\star}$ $\Delta$ Transit depth $\left[\%\right]$ $u_1$ $u_2$ Dilution factor $c_n$ $\left[\%\right]$ Red noise factor $\beta$ --------- ------------------------------- ----------------- ------------------------------------------ ------- ------- ----------------------------------------- -------------------------- \#1 $ 518 - 538$ $ 0.1574\pm 0.0522$ $ 1.862$ \#2 $ 538 - 558$ $ 0.1689\pm 0.0565$ $ 1.950$ \#3 $ 558 - 578$ $ 0.1796\pm 0.0541$ $ 1.612$ \#4 $ 578 - 598$ $ 0.1710\pm 0.0553$ $ 1.022$ \#5 $ 598 - 618$ $ 0.1944\pm 0.0554$ $ 1.761$ \#6 $ 618 - 638$ $ 0.1855\pm 0.0525$ $ 1.950$ \#7 $ 638 - 658$ $ 0.3335\pm 0.0601$ $ 1.908$ \#8 $ 658 - 678$ $ 0.2665\pm 0.0587$ $ 1.575$ \#9 $ 678 - 698$ $ 0.2700\pm 0.0606$ $ 1.702$ \#10 $ 698 - 718$ $ 0.3621\pm 0.0583$ $ 1.738$ \#11 $ 718 - 738$ $ 0.5098\pm 0.0646$ $ 1.285$ \#12 $ 738 - 758$ $ 0.7963\pm 0.0613$ $ 1.626$ (\#13 $ 758 - 778$ $ 0.6573\pm 0.0588$ $ 1.731$) \#13b $ 763 - 778$ $ 0.6211\pm 0.0568$ $ 1.191$ \#14 $ 778 - 798$ $ 0.7601\pm 0.0644$ $ 1.886$ \#15 $ 798 - 818$ $ 1.0250\pm 0.0684$ $ 1.212$ \#16 $ 818 - 838$ $ 1.0954\pm 0.0670$ $ 1.994$ \#17 $ 838 - 858$ $ 1.0714\pm 0.0688$ $ 1.230$ \#18 $ 858 - 878$ $ 1.1793\pm 0.0703$ $ 1.380$ \#19 $ 878 - 898$ $ 1.2985\pm 0.0643$ $ 1.633$ \#20 $ 898 - 918$ $ 1.4459\pm 0.0701$ $ 1.789$ Results and discussion of the transmission spectrum {#sec:results} =================================================== The result for the wavelength-dependent planet-to-star radius ratio of HAT-P-32Ab is shown in Fig \[fig:results\_rprs\]. We find it to show little variation in the probed region between $518$ and $918$ nm. The variations are significantly smaller than two planetary atmospheric scale heights ($H_p\approx1400$ km). This estimate for $H_P$ is derived with Eq. \[eq:scaleh\] $$H_p=\frac{k_bT_p}{mg_p}\label{eq:scaleh}$$ using the planet equilibrium temperature $T_p=2042$ K given by @2014ApJ...796..115Z, the planetary gravitational acceleration of $g_p=6.6069$ $\mathrm{m/s^2}$ from @2011ApJ...742...59H and the approximation of the mean molecular weight of the atmosphere $m$ as that of a solar abundance hydrogen helium mixture. In Eq. \[eq:scaleh\] $k_B$ stands for the Boltzmann constant. The best fit results for the limb darkening coefficients are shown in Fig. \[fig:results\_ld2\]. They vary smoothly with wavelength except for the coefficients for channel \#13, which encompasses the telluric oxygen bands at $\approx761$ nm. We compare the results for the limb darkening coefficients with theoretical predictions. Our approach to the narrow band channel analysis entailed the division by the white light curve. Therefore, effectively only relative changes of the transit depth and the stellar limb darkening to the white light transit depth and white light limb darkening were measured. Consequently, the absolute values of our resulting color dependent limb darkening values are affected by the assumed white light limb darkening values. Since we adopted the best fit white light curve parameters from Run 1, where the limb darkening coefficients are divergent from theoretical expectations, the resulting color dependent coefficients are by default divergent as well. In order to still independently compare them to the theoretical expectations we calculate modified theoretical values which correspond to the theoretical predictions for the color dependent coefficients under assumption of the white light coefficients fixed to the best fit results of Run 1. These modified theoretical values were derived from PHOENIX specific intensity spectra as described in Sect. \[sec:limbd\], where instead of the whole white light wavelength region, now only the corresponding narrow band channels were summed to derive the limb darkening profile. The resulting modified theoretical prediction is shown together with the actual measured limb darkening coefficients in Fig. \[fig:results\_ld2\]. It can be seen that the strong change in limb darkening in respect to the neighboring wavelength regions that we see for channel \#13 is not expected by theory. Since the results for the sub-channel \#13b do not show the same divergence from theory, we conclude that the results of channel \#13 are affected by uncorrected noise due to the strong telluric absorption bands and are un-reliable. Channel \#13 does not carry any significant information that is not also represented by channel \#13b. Therefore, we decide to exclude the results for the planet-to-star radius ratio of channel \#13 in the following comparison to literature data and planet atmosphere model predictions. Comparison to theoretical models and literature data ---------------------------------------------------- We compared the results to theoretical atmosphere models with various degrees of alkali metal abundance depletion (see Fig. \[fig:compare\_rprs\_models\]). The atmospheric models were calculated with a line by line radiative code described in , with updated opacities described in and @2015ApJ...811...55M. We assumed a clear atmosphere (no clouds) in thermochemical equilibrium with a solar abundance of elements for the nominal model. Furthermore, other models with various changes in TiO and alkali abundances were explored. The stellar heating was provided using a stellar synthetic spectrum[^1] of a G-type star with an effective temperature of 6000 K from @2003IAUS..210P.A20C. We found no agreement between the measured transmission spectrum and the expected absorption features of the sodium and potassium resonance doublets ( at 589.0 and 589.6 nm, at 766.5 and 769.9 nm, with the doublets being unresolved at our resolution) predicted by the nominal model (see top panel of Fig. \[fig:compare\_rprs\_models\], green model). The radius ratio in channel \#4 encompassing the sodium doublet is lower by $\approx12.9\sigma$ than the predicted value and the radius ratio in channel \#13b encompassing the potassium doublet is lower by $\approx7.4\sigma$. Further, there is no indication of a detection of the predicted broad wings of the sodium feature in the channels neighboring channel \#4. Neither an enhancement of the titanium oxide (TiO) abundance in the probed upper layers of the atmosphere (see top panel of Fig. \[fig:compare\_rprs\_models\], grey dashed model) nor a depletion of either only sodium or only potassium (see middle panel of Fig. \[fig:compare\_rprs\_models\]) improves the fit between the model and data. Reducing the abundance of both these alkali metals by a factor of 1000 yields a model which agrees fairly well with the data with a reduced $\chi^2$ of $1.58$ (20 degrees of freedom). This is a similar but slightly worse match than the one to a straight line representative of a grey atmosphere (see bottom panel of Fig. \[fig:compare\_rprs\_models\]) which results in a reduced $\chi^2 $ of $1.50$ (20 degrees of freedom). A grey atmosphere signal could be caused by a high altitude cloud layer masking the fingerprint of the atmosphere below. Alternatively, a significantly lower atmospheric scale height of the terminator region than the measured equilibrium temperature of the planet would suggest could be responsible. The amplitude of all expected atmospheric features would shrink accordingly and in an extreme case they would be hidden in the uncertainties of the measurement. To be in very good agreement with the data the scale height would have to be lower than the current prediction by a factor of $6.8$. Such a decrease in scale height could for example be caused by either a lower than predicted terminator temperature ($\approx300$ K) or an increase of the assumed mean molecular weight by that factor (i.e. $m\approx 16$ kg/kmole; for comparison the mean molecular weight of water vapor is $m_{\rm H_2O}=18.02$ kg/kmole). A more plausible explanation, however, would be that both of these two regulating factors are divergent from the values assumed in the current scale height calculation.Our results for the planet-to-star radius ratio are in good agreement with the study of @2013MNRAS.436.2974G who observed two transit events of HAT-P-32Ab with GMOS at Gemini North. They analyzed both data sets separately and then combined the final results for the planet-to-star radius ratio of both data sets. In Fig. \[fig:compare\_rprs\_g13\] we compare our results with their combined results. Potential of GTC/OSIRIS as a tool for transmission spectroscopy --------------------------------------------------------------- We found the light curves to be heavily affected by systematics which can, however, be modeled. Only the channel containing the telluric oxygen absorption band at $\approx761$ nm was too heavily affected by noise to be sufficiently corrected. The data showed correlation with the position of the stars on the chip (and in the slit), with air mass (which is expected due to the difference in color between the planet host star and the reference star) and with seeing. The latter correlation suggests slit losses affecting the planet host star and reference star differently. This problem can be overcome by choosing a larger slit width. We adopted this for the following runs . While the rotation dependent distortion cannot be overcome, it will affect the data less when the changes in the rotation angle are small. It will also have a smaller impact when the two stars are close to each other on the chip since vignetting then should be similar for both stars. In such cases, i.e. when no obvious distortions can be seen in the data, disregarding their possible hidden existence could lead to systematically erroneous transit parameters, but should have negligible impact on the derived transmission spectrum, as long as only relative variations in the radius ratio are considered. Conclusion {#sec:conclusion} ========== We were able to derive a high precision transmission spectrum for the inflated hot Jupiter HAT-P-32Ab showing no prominent absorption features and, thus, supporting the results of the earlier work by @2013MNRAS.436.2974G. The study allowed us to detect and understand the low-level instrument systematics affecting GTC/OSIRIS and will help to improve future measurements. The independent confirmation of ground-based results from a different ground-based facility affected by different systematic noise signals is a step towards re-establishing faith in the reliability of (ground-based) transmission spectroscopy measurements. We have confidence in the potential of ground-based facilities and GTC/OSIRIS in particular as an excellent tool for larger surveys. This work is based on observations made with the Gran Telescopio Canarias (GTC), installed in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofisica de Canarias, in the island of La Palma. It is partly financed by the Spanish Ministry of Economics and Competitiveness through projects ESP2013-48391-C4-2-R and ESP2014-57495-C2-1-R. LN acknowledges support from the DFG Graduiertenkolleg 1351 *Extrasolar Planets and their Host Stars*. FM acknowledges the support of the French Agence Nationale de la Recherche (ANR), under the program ANR-12-BS05-0012 Exo-atmos. Many of the plots shown in this paper were made using Matplotlib [@Hunter2007]. We would like to note that during a late stage of this paper’s referee process another study of HAT-P-32b by @2016arXiv160309136M was uploaded to ArXiv. This study further confirms the measurement of a flat transmission spectrum for this planet’s atmosphere. Companion HAT-P-32B {#sec:sec} =================== In [[email protected]] @2013AJ....146....9A discovered an optical companion to HAT-P-32A using adaptive optics (AO) with Aries, a near infrared diffraction limited imager and spectrograph (PI: Don McCarthy), which is fed by the Multiple Mirror Telescope’s (MMT) AO beam. The optical companion was later confirmed to be bound to the HAT-P-32 system by @2015ApJ...800..138N from proper motion measurements and AO imaging. HAT-P-32B was further characterized as an M-dwarf by @2014ApJ...796..115Z and @2015ApJ...800..138N who used near infrared broad band AO imaging to constrain its stellar parameters.Due to the close proximity of the two stars HAT-P-32A and B the flux of the latter was contributing to our measurements of the former. In order to correctly include this effect in our models we needed to determine the wavelength-dependent flux ratio between the two stars as precisely as possible. In the following we describe how we extracted this information from our GTC/OSIRIS spectra. Subsequently, we use this data and additionally obtained near infrared observations to derive improved stellar parameters for HAT-P-32B. Optical spectrum - GTC/OSIRIS ----------------------------- Since HAT-P-32B was undiscovered prior to our observations the observing set up was not optimized to maximize the projected distance of the two spectra of HAT-P-32A and B on the chip. We still were able to detect HAT-P-32B as a separated object in Run 2 as a deformation of the spatial profile of HAT-P-32A. Using the out-of-transit data available in Run 2 we determined the flux peak of the HAT-P-32A spectrum in every frame at every wavelength and then added frames in 10-frame time-bins resulting in 25 images with increased signal-to-noise. We then fitted a double-peak model based on an empirical profile function to the spatial profiles of HAT-P-32A and B for every wavelength cut though the spectrum for all 25 images. The used empirical profile was based on a Moffat function and assumed symmetry of the stellar profile. It is described in detail in Sect. \[sec:emprofile\]. A sample fit to the spatial double profile can be found in Fig. \[fig:profile\]. ### Empirical profile {#sec:emprofile} We acquired a first approximation of the peak position and amplitude of the two stellar profiles by fitting each of them with a Moffat function enforcing identical values for the half widths and for the Moffat indices of both profiles. The result showed that the wings of the actual stellar profiles were underestimated by the Moffat approximation. Consequently, we moved on to an empirical approach. For this we assumed that the actual profile is symmetrical in respect to its central axis and that the point spread functions (PSFs) of both stars are the same. We constructed an empirical profile $\mathcal{P}_e$ using the side of the stellar profile of HAT-P-32A which is not distorted by the additional flux of HAT-P-32B and mirroring it at the central axis of the HAT-P-32A profile. We then used a scaled down, shifted in pixel position version of $\mathcal{P}_e$ to model the profile of HAT-P-32B. The final model for the superposition of both stellar profiles $\mathcal{W}$ was of the form: $$\mathcal{W}\left(x\right)= \mathcal{P}_e\left(x\right)+s\cdot\mathcal{P}_e\left(x-x_{A-B}\right) \label{eq:mp}$$ In this model the scaling factor ($s$), the center of the stellar profile of HAT-P-32A ($x$), and the distance of the centers of both stars ($x_{A-B}$) were free parameters. From the optimization of this model we obtained the flux for HAT-P-32A ($f_A$) and HAT-P-32B ($f_B$) for every wavelength element in every of the 25 images by summing over the respective theoretical profiles. ### Dilution of the optical transit depth measurements of HAT-P-32A b Since HAT-P-32B was within the aperture used in Sect. \[sec:specextract\] to determine the flux of HAT-P-32A, its additional flux had a diluting effect on the transit depth. This effect is color dependent as HAT-P-32B contributes different amounts of flux in the individual wavelength channels defined in Sect. \[sec:lightcurves\]. We determined the contribution in every wavelength channel by first summing the measured flux of both stars separately within the channel limits in each of the 25 frames. For each channel we then divided the resulting total flux of both stars and adopted the mean of the 25 flux ratios as the final result of the dilution factor $c_n=f_A/f_B$ and their standard deviation as its uncertainty. The resulting values for each channel are listed in Table \[tab:tspec\_results\] in Sect. \[sec:results\]. ### Resulting relative optical spectrum While for the correction of the multi-color channel light curves in this paper only low resolution information for HAT-P-32b was needed, higher resolution information was available. We used this data to determine the stellar properties of HAT-P-32B in Sect. \[sec:mdwarf2phoenix\]. In order to remove the OSIRIS instrument sensitivity function from the data and clean it of any telluric absorption we used the relative spectrum of HAT-P-32B and HAT-P-32A for this analysis. Consequently, all wavelength-dependent telluric and instrumental effects are divided out. Since the slit alignment was optimized to center HAT-P-32A and the reference star (Ref2, see Table \[tab:coord\]) within it, HAT-P-32B was not perfectly centered in the slit. As a consequence the wavelength solution for its spectrum is slightly shifted compared to the wavelength solution of HAT-P-32A. We first applied this shift in wavelength and then calculated the flux ratio for every wavelength element in each of the 25 frames. We again adopted the mean of the 25 results as the final value and their standard deviation as the uncertainty. The resulting relative spectrum covering the wavelength range $518-918$ nm is shown in Fig. \[fig:h32bvsa\_optical\]. Infrared photometry - WHT/LIRIS ------------------------------- In addition to the optical spectrum we obtained $J$, $H$, and $K_S$ band measurements with the William Herschel Telescope (WHT) in the night of the 6th of October 2012 using LIRIS (Long-slit Intermediate Resolution Infrared Spectrograph) in imaging mode. We took 18 images and obtained 15 flats and darks 15 for each filter. A sample image of these observations for each filter is shown in Fig. \[fig:wht\_phot\]. We used PSF fitting to obtain the absolute flux for HAT-P-32B and HAT-P-32A. We then calculated the flux ratio $f_{\text{B}}/f_{\text{A}}$ for every image and adopted the mean as the final value and the standard deviation as the uncertainty for each filter. The results are listed in the first column of Table \[tab:c\_ir\]. ----------------------------------- ------------------- -------------- -- ------- $f_{\text{B}}/f_{\text{A}}$ \[%\] This work in Filter $J$ $ 2.65\pm 0.45 $ ... ... $H$ $3.73\pm 0.65 $ $4.4\pm0.5$ ... $K'$ ... ... ... $K_S$ $4.64 \pm 0.21$ $4.7\pm 0.2$ 0.044 ----------------------------------- ------------------- -------------- -- ------- : Flux ratios between HAT-P-32B and HAT-P-32A obtained from infrared imaging in this work, [@2015ApJ...800..138N], [@2014ApJ...796..115Z] and [@2013AJ....146....9A].[]{data-label="tab:c_ir"} Comparison to literature {#sec:mdwarf2lit} ------------------------ Source $T_\mathrm{eff}$ (K) $\log g$ (cgs) \[Fe/H\] (fixed/varied within prior) ------------------------------------- ---------------------- ------------------- -------------------------------------- -- This work (2012-Sep-30) @2014ApJ...796..115Z (2013-Mar-02) $3565 \pm82$ ... $-0.04$ @2015ApJ...800..138N  (2012-Feb-02) $3516\pm12$ $4.8930\pm0.0098$ $0.00$ @2015ApJ...800..138N  (2013-Mar-02) $3551 \pm 10$ $4.8677\pm0.0070$ $0.00$ We compared our results of the flux ratio $c_{\mathrm{filter}}=f_A/f_B$ of HAT-P-32B and HAT-P-32A with the broad band measurements from the literature obtained in the optical ($g'$, $r'$, $i'$, $z'$) by @2014ApJ...796..115Z with the Robo-AO instrument [@2014ApJ...790L...8B] on the 60 inch telescope at the Palomar Observatory and in the near infrared ($J$, $H$, $K'$ and $Ks$) by @2014ApJ...796..115Z and @2015ApJ...800..138N with NIRC2 a near-infrared imager (PI: Keith Matthews) using the AO system of Keck-II [@2000PASP..112..315W] and @2013AJ....146....9A using MMT/Aries.The optical wavelength region probed with our GTC/OSIRIS transit observations fully overlaps with the $r'$ and $i'$ bands. We folded our data with the respective filter curves and found the results ($c_{r'}=0.0023(7)$, $c_{i'}=0.0064(8)$) to be consistent with the study of @2014ApJ...796..115Z ($c_{\textrm{Zhao},r'}=0.003(1)$, $c_{\textrm{Zhao},i'}=0.006(2)$).Since the WHT/LIRIS filter curves for the near infrared broadband filters $J$, $H$, and $K_S$ differ slightly from the ones used by the Keck-II/NIRC2 facility, an exact comparison between the respective measurements is not feasible. When neglecting these small differences in filter transmission, we, however, found that our results are consistent within $1\sigma$ with the near infrared values derived by all three studies @2014ApJ...796..115Z, @2015ApJ...800..138N and @2013AJ....146....9A. Both @2014ApJ...796..115Z and @2015ApJ...800..138N independently analyzed the Keck-II/NIRC2 data obtained in the $H$ and $K_S$ band passes (while the $J$ and $K'$ band data was only analyzed by @2015ApJ...800..138N) and arrived at different results, which are only consistent with each other within $2\sigma$. Comparison to theoretical models {#sec:mdwarf2phoenix} -------------------------------- Using their broadband measurements @2014ApJ...796..115Z and @2015ApJ...800..138N both determined physical properties of HAT-P-32B and arrived at similar results with effective temperatures around $T_\mathrm{eff}=3550$ K. The exact results are listed in Table \[tab:stellarh32B\].Both studies made use of the PHOENIX stellar atmosphere models but relied on slightly different approaches, stellar parameters for HAT-P-32A and different fixed values for the metallicity.We conducted our own analysis by fitting a theoretical model of the form: $$\mathcal{R}=k\cdot \frac{M_{\mathrm{HAT-P-32B}}\left(\lambda,T_\mathrm{eff},\log g, \left[\mathrm{Fe/H}\right]\right)}{M_{\mathrm{HAT-P-32A}}\left(\lambda,T_\mathrm{eff},\log g, \left[\mathrm{Fe/H}\right]\right)} \label{eq:pspecmodel}$$ to our optical and near infrared data as well as the literature broadband measurements (excluding the upper limit @2014ApJ...796..115Z give for the $g'$ band and the $K_S$ band data point of @2013AJ....146....9A for which no uncertainties were given). In Eq. \[eq:pspecmodel\] $k$ is a scaling factor (corresponding to the squared radius ratio of the two stars $\left(R_B/R_A\right)^2$) and $M_{\mathrm{HAT-P-32B}}$ and $M_{\mathrm{HAT-P-32A}}$ are PHOENIX stellar models, interpolated to specific stellar parameters from the model grid provided by with trilinear interpolation. During the fit the stellar parameters of HAT-P-32A were allowed to vary within the uncertainties of the given literature values (\[Fe/H\] $=-0.04\pm0.08$, $\log g$ (cgs) $=4.33 \pm 0.01$, @2011ApJ...742...59H, and $T_{\mathrm{eff}}=6269\pm64$ K, @2014ApJ...796..115Z) and an identical metallicity for both stars was enforced. For the comparison with the broadband points we folded the PHOENIX model with the respective filter curves. In this step we used the exact instrument specific filter curve (downloaded from the facility web-pages) for each broadband data point. For the optical spectrum we folded the PHOENIX model spectrum with a Gauss function to reduce the resolution to match the data. We found that if we only fit the broadband measurements we arrive at similar values for HAT-P-32B’s stellar parameters as the ones derived by @2014ApJ...796..115Z and @2015ApJ...800..138N. When including the optical data, however, the fit clearly favored cooler temperatures. Due to the inconsistencies in the literature regarding the results obtained for the $H$ and $K_S$ bands with Keck-II/NIRC2 we decided to only include our own data in the final optimization. The difference in the best fit effective temperature due to this exclusion of broadband points is insignificant (6 K). Our results for HAT-P-32B’s stellar parameters are given in Table \[tab:stellarh32B\]. The uncertainties of these results were inflated with the red noise factor $\beta$, which was calculated as described in Sect. \[sec:rednoise\]. The errors do not incorporate any intrinsic uncertainties of the PHOENIX stellar atmosphere models.The best fit model relative spectrum is plotted together with all data points (including those that were not regarded in the fit) in Fig. \[fig:h32bvsa\_optical\] and \[fig:h32bvsa\_ir\].It stands out that the $K'$ band result of @2015ApJ...800..138N diverges significantly from all $K_S$ band values including their own. This offset is too large to be explained by the difference between the $K'$ and $K_S$ band passes. Since the $K_S$ and $K'$ observations by @2015ApJ...800..138N were conducted at different dates (13 month apart), a more likely explanation for this inconsistency could be stellar variability of HAT-P-32B due to activity i.e. star spots or flares. We measure a prominent $H_{\alpha}$ emission line at $656$ nm in the optical spectrum (see Fig. \[fig:h32bvsa\_optical\]), indicative of such stellar activity. [^1]: (available at ftp://ftp.stsci.edu/cdbs/grid/ck04models)
{ "pile_set_name": "ArXiv" }
--- abstract: 'We provide the first mathematical proof that the connective constant of the hexagonal lattice is equal to $\sqrt{2+\sqrt{2}}$. This value has been derived non rigorously by B. Nienhuis in 1982, using Coulomb gas approach from theoretical physics. Our proof uses a parafermionic observable for the self avoiding walk, which satisfies a half of the discrete Cauchy-Riemann relations. Establishing the other half of the relations (which conjecturally holds in the scaling limit) would also imply convergence of the self-avoiding walk to SLE($8/3$).' author: - 'Hugo Duminil-Copin and Stanislav Smirnov' title: 'The connective constant of the honeycomb lattice equals $\sqrt{2+\sqrt 2}$' --- Introduction ============ A famous chemist P. Flory [@Flory] proposed to consider self-avoiding (*i.e.* visiting every vertex at most once) walks on a lattice as a model for polymer chains. Self-avoiding walks turned out to be a very interesting object, leading to rich mathematical theories and challenging questions, see [@MadrasSlade]. Denote by $c_n$ the number of $n$-step self-avoiding walks on the hexagonal lattice ${\mathbb{H}}$ started from some fixed vertex, *e.g.* the origin. Elementary bounds on $c_n$ (for instance $\sqrt{2}^n\leq c_n\leq 3\cdot 2^{n-1}$) guarantee that $c_n$ grows exponentially fast. Since a $(n+m)$-step self-avoiding walk can be uniquely cut into a $n$-step self-avoiding walk and a parallel translation of a $m$-step self-avoiding walk, we infer that $$c_{n+m}\leq c_nc_m,$$ from which it follows that there exists $\mu\in(0,+\infty)$ such that $$\mu:=\lim_{n\rightarrow \infty}c_n^{\ \frac 1n}.$$ The positive real number $\mu$ is called the *connective constant* of the hexagonal lattice. Using Coulomb gas formalism, B. Nienhuis [@Nienhuis; @Nienhuis-jsp] proposed physical arguments for $\mu$ to have the value $\sqrt{2+\sqrt 2}$. We rigorously prove this statement. While our methods are different from those applied by Nienhuis, they are similarly motivated by considerations of vertex operators in the $O(n)$ model. Our methods do not directly apply to the square lattice, for which the value of the connective constant is different and currently unknown. \[theorem\] For the hexagonal lattice, $$\mu=\sqrt{2+\sqrt{2}}.$$ It will be convenient to consider walks between *mid-edges* of ${\mathbb{H}}$, *i.e.* centers of edges of ${\mathbb{H}}$ (the set of mid-edges will be denoted by $H$). We will write ${\gamma}:a\rightarrow E$ if a walk ${\gamma}$ starts at $a$ and ends at some mid-edge of $E\subset H$. In the case $E=\{b\}$, we simply write ${\gamma}:a\rightarrow b$. The *length* $\ell({\gamma})$ of the walk is the number of vertices visited by ${\gamma}$. We will work with the partition function $$Z(x)=\sum_{{\gamma}\ :\ a\rightarrow H}x^{\ell({\gamma})}\quad\in(0,+\infty].$$ This sum does not depend on the choice of $a$, and is increasing in $x$. Establishing the identity $\mu=\sqrt{2+\sqrt 2}$ is equivalent to showing that $Z(x)=+\infty$ for $x>1/\sqrt{2+\sqrt{2}}$ and $Z(x)<+\infty$ for $x<1/\sqrt{2+\sqrt 2}$. To this end, we analyze walks restricted to bounded domains and weighted depending on their winding. The modified sum can be defined as a *parafermionic observable* arising from a disorder operator. Such observables exist for other models, see [@CardyIkhlef; @chelkak-smirnov-iso; @smirnov-icm2010]. The paper is organized as follows. In Section 2, the parafermionic observable is introduced and its key property is derived. Section 3 contains the proof of Theorem \[theorem\]. Section 4 discusses conformal invariance conjectures for self-avoiding walks. To simplify formulæ, below we set $x_c:=1/\sqrt{2+\sqrt 2}$ and $j={{\rm e}}^{{{\rm i}}2\pi/3}$. Parafermionic observable ======================== A (hexagonal lattice) *domain* $\Omega\subset H$ is a union of all mid-edges emanating from a given collection of vertices $V(\Omega)$ (see Fig. \[fig:SAWpicture\]): a mid-edge $z$ belongs to $\Omega$ if at least one end-point of its associated edge is in $\Omega$, it belongs to $\partial \Omega$ if only one of them is in $\Omega$. We further assume $\Omega$ to be simply connected, *i.e.* having a connected complement. ![**Left.** A domain $\Omega$ with boundary mid-edges labeled by small black squares, and vertices of $V(\Omega)$ labeled by circles. **Right.** Winding of a curve ${\gamma}$.[]{data-label="fig:SAWpicture"}](SAWpicture){width="0.6\hsize"} For a self-avoiding walk ${\gamma}$ between mid-edges $a$ and $b$ (not necessarily the start and the end), we define its *winding ${{\rm W}}_{\gamma}(a,b)$* as the total rotation of the direction in radians when ${\gamma}$ is traversed from $a$ to $b$, see Fig. \[fig:SAWpicture\]. Our main tool is given by the following The *parafermionic observable* for $a\in \partial \Omega$, $z\in \Omega$, is defined by $$F(z)=F(a,z,x,\sigma)=\sum_{{\gamma}\subset \Omega:\ a\rightarrow z}{{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{\gamma}(a,z)} x^{\ell({\gamma})}.$$ \[lem:relation\] If $x=x_c$ and $\sigma=\frac 58$, then $F$ satisfies the following relation for every vertex $v\in V(\Omega)$: $$\label{relation around vertex} (p-v)F(p)+(q-v)F(q)+(r-v)F(r)=0,$$ where $p,q,r$ are the mid-edges of the three edges adjacent to $v$. Note that with $\sigma=5/8$, the complex weight ${{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{\gamma}(a,z)}$ can be interpreted as a product of terms $\lambda$ or $\bar{\lambda}$ per left or right turn of ${\gamma}$ drawn from $a$ to $z$, with $$\lambda=\exp \left(-{{\rm i}}\frac{5}{8}\cdot \frac{\pi}{3}\right)= \exp \left(-{{\rm i}}\frac{5\pi}{24}\right).$$ We start by choosing notation so that $p$, $q$ and $r$ follow counter-clockwise around $v$. Note that the left-hand side of can be expanded into the sum of contributions $c({\gamma})$ of all possible walks ${\gamma}$ finishing at $p,q$ or $r$. For instance, if a walk ends at the mid-edge $p$, its contribution will be given by $$c({\gamma})=(p-v)\cdot {{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{{\gamma}}(a,p)}~x_c^{\ell({\gamma})}.$$ One can partition the set of walks ${\gamma}$ finishing at $p,q$ or $r$ into pairs and triplets of walks in the following way, see Fig \[fig:pairs\]: - If a walk ${\gamma}_1$ visits all three mid-edges $p,q,r$, it means that the edges belonging to ${\gamma}_1$ form a disjoint self-avoiding path plus (up to a half-edge) a self-avoiding loop from $v$ to $v$. One can associate to ${\gamma}_1$ the walk passing through the same edges, but exploring the loop from $v$ to $v$ in the other direction. Hence, walks visiting the three mid-edges can be grouped in pairs. - If a walk ${\gamma}_1$ visits only one mid-edge, it can be associated to two walks ${\gamma}_2$ and ${\gamma}_3$ that visit exactly two mid-edges by prolonging the walk one step further (there are two possible choices). The reverse is true: a walk visiting exactly two mid-edges is naturally associated to a walk visiting only one mid-edge by erasing the last step. Hence, walks visiting one or two mid-edges can be grouped in triplets. If one can prove that the sum of contributions to of each pair or triplet vanishes, then their total sum is zero, and holds. Let ${\gamma}_1$ and ${\gamma}_2$ be two associated walks as in the first case. Without loss of generality, we may assume that ${\gamma}_1$ ends at $q$ and ${\gamma}_2$ ends at $r$. Note that ${\gamma}_1$ and ${\gamma}_2$ coincide up to the mid-edge $p$, and then follow an almost complete loop in two opposite directions. It follows that $$\ell({\gamma}_1)=\ell({\gamma}_2)\quad\quad \text{and}\quad \quad \left\{\ \substack{ {{\rm W}}_{{\gamma}_1}(a,q)={{\rm W}}_{{\gamma}_1}(a,p)+{{\rm W}}_{{\gamma}_1}(p,q)={{\rm W}}_{{\gamma}_1}(a,p)-\frac{4\pi}{3}\\ \ \\ {{\rm W}}_{{\gamma}_2}(a,r)={{\rm W}}_{{\gamma}_2}(a,p)+{{\rm W}}_{{\gamma}_2}(p,r)={{\rm W}}_{{\gamma}_1}(a,p)+\frac{4\pi}{3}}\right.\quad.$$ In order to evaluate the winding of ${\gamma}_1$ between $p$ and $q$ above, we used the fact that $a$ is on the boundary and $\Omega$ is simply connected. We conclude that $$\begin{aligned} c({\gamma}_1)+c({\gamma}_2)&=(q-v){{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{{\gamma}_1}(a,q)} x_c^{\ell({\gamma}_1)}+(r-v){{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{{\gamma}_2}(a,r)} x_c^{\ell({\gamma}_2)}\\ &=(p-v){{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{{\gamma}_1}(a,p)}x_c^{\ell({\gamma}_1)}\left(j\bar{\lambda}^4+\bar{j}\lambda^4\right)=0\end{aligned}$$ where the last equality holds since $j\bar{\lambda}^4=-i$ by our choice of $\lambda=\exp (-{{\rm i}}5\pi/24)$. Let ${\gamma}_1,{\gamma}_2,{\gamma}_3$ be three walks matched as in the second case. Without loss of generality, we assume that ${\gamma}_1$ ends at $p$ and that ${\gamma}_2$ and ${\gamma}_3$ extend ${\gamma}_1$ to $q$ and $r$ respectively. As before, we easily find that $$\ell({\gamma}_2)=\ell({\gamma}_3)=\ell({\gamma}_1)+1\quad\quad\text{and}\quad\quad\left\{\ \substack{{{\rm W}}_{{\gamma}_2}(a,r)={{\rm W}}_{{\gamma}_2}(a,p)+{{\rm W}}_{{\gamma}_2}(p,q)={{\rm W}}_{{\gamma}_1}(a,p)-\frac{\pi}{3}\\ \ \\ {{\rm W}}_{{\gamma}_3}(a,r)={{\rm W}}_{{\gamma}_3}(a,p)+{{\rm W}}_{{\gamma}_3}(p,r)={{\rm W}}_{{\gamma}_1}(a,p)+\frac{\pi}{3}}\right.\quad.$$ Plugging these values into the respective contributions, we obtain $$\begin{aligned} c({\gamma}_1)+c({\gamma}_2)+c({\gamma}_3)&=(p-v){{\rm e}}^{-{{\rm i}}\sigma {{\rm W}}_{{\gamma}_1}(a,p)}x_c^{\ell({\gamma}_1)}\left(1+x_cj\bar{\lambda}+x_c\bar{j}\lambda\right)=0.\end{aligned}$$ Above is the *only* place where we use that $x$ takes its critical value, i.e. $x_c^{-1}=\sqrt{2+\sqrt2}=(2\cos \frac{\pi}{8})$. The claim of the lemma follows readily by summing over all pairs and triplets. ![**Left:** a pair of walks visiting all the three mid-edges emanating from $v$ and differing by rearranged connections at $v$. **Right:** a triplet of walks, one visiting one mid-edge, the two others visiting two mid-edges, and obtained by prolonging the first one through $v$.[]{data-label="fig:pairs"}](pairs){width="0.90\hsize"} \[rem:CR\] Coefficients in are three cube roots of unity multiplied by $p-v$, so its left-hand side can be seen as a discrete $dz$-integral along an elementary contour on the dual lattice. The fact that the integral of the parafermionic observable along discrete contours vanishes suggests that it is discrete holomorphic and that self-avoiding walks have a conformally invariant scaling limit, see Section 4. Proof of Theorem \[theorem\] ============================ #### Counting argument in a strip domain. We consider a vertical strip domain $S_T$ composed of $T$ strips of hexagons, and its finite version $S_{T,L}$ cut at heights $\pm L$ at angles $\pm \pi/3$, see Fig. \[fig:domain\]. Namely, position a hexagonal lattice ${\mathbb{H}}$ of meshsize 1 in ${\mathbb{C}}$ so that there exists a horizontal edge $e$ with mid-edge $a$ being 0. Then $$\begin{aligned} V(S_T)&=\{z\in V({\mathbb{H}}):0\leq {\rm Re}(z)\leq \frac{3T+1}{2} \},\\ V(S_{T,L})&=\{z\in V(S_T):|\sqrt 3~{\rm Im}(z)-{\rm Re}(z)|\leq 3L\}.\end{aligned}$$ Denote by $\alpha$ the left boundary of $S_{T}$, by $\beta$ the right one. Symbols ${\varepsilon}$ and $\bar{{\varepsilon}}$ denote the top and bottom boundaries of $S_{T,L}$. Introduce the following (positive) partition functions: $$A_{T,L}^x:=\sum_{\substack{{\gamma}\subset S_{T,L} :\ a\rightarrow\alpha\setminus \{a\}}}x^{\ell(\gamma)},~~ B_{T,L}^x:=\sum_{\substack{{\gamma}\subset S_{T,L}:\ a\rightarrow\beta}}x^{\ell(\gamma)},~~ E_{T,L}^x:=\sum_{\substack{{\gamma}\subset S_{T,L}:\ a\rightarrow{\varepsilon}\cup\bar{{\varepsilon}}}}x^{\ell(\gamma)}.$$ In the next lemma, we deduce from relation a global identity without the complex weights. ![Domain $S_{T,L}$ and boundary intervals $\alpha$, $\beta$, ${\varepsilon}$ and $\bar{{\varepsilon}}$.[]{data-label="fig:domain"}](domain){width="0.30\hsize"} For critical $x=x_c$, the following identity holds $$\label{equation box} 1=c_{\alpha}A^{x_c}_{T,L}+B^{x_c}_{T,L}+c_{{\varepsilon}}E^{x_c}_{T,L},$$ with positive coefficients $c_{\alpha}=\cos \left(\frac {3\pi}{8}\right)$ and $c_{{\varepsilon}}=\cos \left(\frac{\pi}{4}\right)$. Sum the relation over all vertices in $V(S_{T,L})$. Values at interior mid-edges disappear and we arrive at the identity $$\label{sum} 0=-\sum_{z\in \alpha}F(z)+\sum_{z\in \beta}F(z)+j\sum_{z\in {\varepsilon}}F(z)+\bar{j}\sum_{z\in \bar{{\varepsilon}}}F(z).$$ The symmetry of our domain implies that $F(\bar{z})=\bar{F}(z)$, where $\bar{x}$ denotes the complex conjugate of $x$. Observe that the winding of any self-avoiding walk from $a$ to the bottom part of $\alpha$ is $-\pi$ while the winding to the top part is $\pi$. Thus $$\begin{aligned} \sum_{z\in \alpha}F(z)&=F(a)+\sum_{z\in \alpha\setminus \{a\}}F(z) =F(a)+\frac12\sum_{z\in \alpha\setminus \{a\}}\left(F(z)+F(\bar z)\right)\\ &=1+\frac{{{\rm e}}^{-{{\rm i}}\sigma \pi}+{{\rm e}}^{{{\rm i}}\sigma \pi}}{2}A^x_{T,L}=1-\cos \left(\frac{3\pi}{8}\right)\ A^x_{T,L}=1-c_{\alpha}A^x_{T,L}.\end{aligned}$$ Above we have used the fact that the only walk from $a$ to $a$ is a trivial one of length $0$, and so $F(a)=1$. Similarly, the winding from $a$ to any half-edge in $\beta$ (resp. ${\varepsilon}$ and $\bar{{\varepsilon}}$) is 0 (resp. $\frac {2\pi}3$ and $-\frac{2\pi}3$), therefore $$\sum_{z\in \beta}F(z)=B^x_{T,L}\quad \text{and}\quad j\sum_{z\in {\varepsilon}}F(z)+\bar{j}\sum_{z\in \bar{{\varepsilon}}}F(z)=\cos \left(\frac\pi 4\right)\ E^x_{T,L}=c_{{\varepsilon}}E^x_{T,L}.$$ The lemma follows readily by plugging the last three formulæ into . Observe that sequences $(A^x_{T,L})_{L>0}$ and $(B^x_{T,L})_{L>0}$ are increasing in $L$ and are bounded for $x\leq x_c$ thanks to and their monotonicity in $x$. Thus they have limits $$A^x_T:=\lim_{L\rightarrow \infty}A^x_{T,L}=\sum_{{\gamma}\subset S_T:\ a\rightarrow\alpha\setminus \{a\}} x^{\ell({\gamma})},~~~~ B^x_T:=\lim_{L\rightarrow \infty}B^x_{T,L}=\sum_{{\gamma}\subset S_T:\ a\rightarrow \beta} x^{\ell({\gamma})}.$$ Identity then implies that $(E^{x_c}_{T,L})_{L>0}$ decreases and converges to a limit $E^{x_c}_T=\lim_{L\rightarrow \infty}E^{x_c}_{T,L}$. Passing to a limit in , we arrive at $$\label{equation strip} 1=c_{\alpha}A^{x_c}_T+B^{x_c}_T+c_{{\varepsilon}}E^{x_c}_T.$$ We start by proving that $Z(x_c)=+\infty$, and hence $\mu\geq \sqrt{2+\sqrt 2}$. Suppose that for some $T$, $E^{x_c}_T>0$. As noted before, $E^{x_c}_{T,L}$ decreases in $L$ and so $$Z(x_c)\geq \sum_{L>0}E^{x_c}_{T,L}\geq \sum_{L>0} E^{x_c}_T=+\infty,$$ which completes the proof. Assuming on the contrary that $E^{x_c}_T=0$ for all $T$, we simplify to $$\label{infinite strip} 1=c_{\alpha}A_T^{x_c}+B_T^{x_c}.$$ Observe that a walk $\gamma$ entering into the count of $A_{T+1}^{x_c}$ and not into $A_T^{x_c}$ has to visit some vertex adjacent to the right edge of $S_{T+1}$. Cutting $\gamma$ at the first such point (and adding half-edges to the two halves), we uniquely decompose it into two walks crossing $S_{T+1}$ (these walks are usually called bridges), which together are one step longer than $\gamma$. We conclude that $$\label{rec relation} A_{T+1}^{x_c}-A_T^{x_c}\leq x_c\left(B_{T+1}^{x_c}\right)^2.$$ Combining for two consecutive values of $T$ with , we can write $$\begin{aligned} 0&=1-1=(c_{\alpha}A_{T+1}^{x_c}+B_{T+1}^{x_c})-(c_{\alpha}A_{T}^{x_c}+B_{T}^{x_c})\\ &=c_{\alpha}(A_{T+1}^{x_c}-A_T^{x_c})+B_{T+1}^{x_c}-B_T^{x_c} \leq c_{\alpha}x_c\left(B_{T+1}^{x_c}\right)^2+B_{T+1}^{x_c}-B_T^{x_c},\end{aligned}$$ and so $$c_{\alpha}x_c\left(B_{T+1}^{x_c}\right)^2+B_{T+1}^{x_c}\geq B_T^{x_c}.$$ It follows easily by induction, that $$B_T^{x_c}\geq {\min [B_1^{x_c},1/(c_{\alpha}x_c)]}~/~{T}$$ for every $T\geq1$, and therefore $$Z(x_c)\geq \sum_{T>0}B_T^{x_c}=+\infty.$$ This completes the proof of the estimate $\mu\geq x_c^{-1}= \sqrt{2+\sqrt2}$. It remains to prove the opposite inequality $\mu\leq x_c^{-1}$. To estimate the partition function from above, we will decompose self-avoiding walks into bridges. A *bridge* of width $T$ is a self-avoiding walk in $S_T$ from one side to the opposite side, defined up to vertical translation. The partition function of bridges of width $T$ is $B_T^x$, which is at most $1$ by . Noting that a bridge of width $T$ has length at least $T$, we obtain for $x<x_c$ $$B_T^x\leq \left(\frac{x}{x_c}\right)^TB_T^{x_c}\leq \left(\frac{x}{x_c}\right)^T.$$ Thus, for $x<x_c$, the series $\sum_{T>0} B_T^x$ converges and so does the product $\prod_{T>0} (1+B_T^x)$. Let us assume for the moment the following fact: *any self-avoiding walk can be canonically decomposed into a sequence of bridges of widths $T_{-i}<\cdots<T_{-1}$ and $T_0>\cdots>T_j$, and, if one fixes the starting mid-edge and the first vertex visited, the decomposition uniquely determines the walk*. Such decomposition was first introduced by Hammersley and Welsh in [@HammersleyWelsh] (for a modern treatment, see Section 3.1 of [@MadrasSlade]). Applying the decomposition to walks starting at $a$ (the first visited vertex is 0 or -1), we can estimate $$Z(x)\leq 2\sum_{\substack{T_{-i}<\cdots<T_{-1}\\ T_j<\cdots<T_0}} \left(\prod_{k=-i}^jB_{T_k}^x\right)=2\prod_{T>0}(1+B_T^x)^2<\infty.$$ The factor 2 is due to the fact that there are two possibilities for the first vertex once we fix the starting mid-edge. Therefore, $Z(x)<+\infty$ whenever $x<x_c$ and $\mu\leq x_c^{-1}=\sqrt{2+\sqrt2}$. To complete the proof of the theorem it only remains to prove that such a decomposition into bridges does exist. Once again, this fact is well-known [@MadrasSlade; @HammersleyWelsh], but we include the proof for completeness. ![**Left:** Decomposition of a half-plane walk into four bridges with widths $8>3>1>0$. The first bridge corresponds to the maximal bridge containing the origin. Note that the decomposition contains one bridge of width 0. **Right:** The reverse procedure. If the starting mid-edge and the first vertex are fixed, the decomposition is unambiguous.[]{data-label="fig:decomposition"}](decomposition){width="0.60\hsize"} First assume that $\tilde{{\gamma}}$ is a half-plane self-avoiding walk, meaning that the start of $\tilde{{\gamma}}$ has extremal real part: we prove by induction on the width $T_0$ that the walk admits a canonical decomposition into bridges of widths $T_0>\cdots>T_j$. Without loss of generality, we assume that the start has minimal real part. Out of the vertices having the maximal real part, choose the one visited last, say after $n$ steps. The $n$ first vertices of the walk form a bridge $\tilde{{\gamma}}_1$ of width $T_0$, which is the first bridge of our decomposition when prolonged to the mid-edge on the right of the last vertex. We forget about the $(n+1)$-th vertex, since there is no ambiguity in its position. The consequent steps form a half-plane walk $\tilde{{\gamma}}_2$ of width $T_1<T_0$. Using the induction hypothesis, we know that $\tilde{{\gamma}}_2$ admits a decomposition into bridges of widths $T_1>\cdots>T_j$. The decomposition of $\tilde{{\gamma}}$ is created by adding $\tilde{{\gamma}}_1$ before the decomposition of $\tilde{{\gamma}}_2$. If the walk is a reverse half-plane self-avoiding walk, meaning that the end has extremal real part, we set the decomposition to be the decomposition of the reverse walk in the reverse order. If ${\gamma}$ is a self-avoiding walk in the plane, one can cut the trajectory into two pieces ${\gamma}_1$ and ${\gamma}_2$: the vertices of ${\gamma}$ up to the first vertex of maximal real part, and the remaining vertices. The decomposition of ${\gamma}$ is given by the decomposition of ${\gamma}_1$ (with widths $T_{-i}<\cdots <T_{-1}$) plus the decomposition of ${\gamma}_2$ (with widths $T_0>\cdots >T_j$). Once the starting mid-edge and the first vertex are given, it is easy to check that the decomposition uniquely determines the walk by exhibiting the reverse procedure, see Fig. \[fig:decomposition\] for the case of half-plane walks. \[remark\] The proof provides bounds for the number of bridges from $a$ to the right side of the strip of width $T$, namely, $$\frac{c}{T}\leq B_T^{x_c}\leq 1.$$ In paragraphs 3.3.3 and 3.4.3 of [@LawlerSchrammWerner], precise behaviors are conjectured for the number of self-avoiding walks between two points on the boundary of a domain, which yields the following (conjectured) estimate: $$\sum_{\gamma\subset S_T:0\rightarrow T+{\rm i}yT} x_c^{\ell(\gamma)} \approx T^{-5/4} H(0,1+{\rm i}y)^{5/4}$$ where $H$ is the boundary derivative of the Poisson kernel. Integrating with respect to $y$, we obtain that $B_T^{x_c}$ should decay as $T^{-1/4}$ when $T$ goes to infinity. Similar estimates are conjectured for walks in $S_T$ from $0$ to ${\rm i}yT$. Conjectures =========== In [@Nienhuis; @Nienhuis-jsp], Nienhuis proposed a more precise asymptotical behavior for the number of self-avoiding walks: $$\label{number}c_n~\sim~ A~n^{\gamma-1}~\sqrt{2+\sqrt 2}^{~n},$$ with $\gamma=43/32$. Here the symbol $\sim$ means that the ratio of two sides is of the order $n^{o(1)}$, or perhaps even tends to a constant. Moreover, Nienhuis gave arguments in support of Flory’s prediction that the mean-square displacement $\langle |\gamma(n)|^2\rangle$ satisfies $$\label{mean-square}\langle |\gamma(n)|^2\rangle~=~\frac1{c_n}~\sum_{\gamma~n-\text{step SAW}}~|\gamma(n)|^2~=~n^{2\nu+o(1)}~,$$ with $\nu=3/4$. Despite the precision of the predictions and , the best rigorously known bounds are very far apart and almost 50 years old (see [@MadrasSlade] for an exposition). The derivation of these exponents seems to be one of the most challenging problems in probability. It was shown by G. Lawler, O. Schramm and W. Werner in [@LawlerSchrammWerner] that $\gamma$ and $\nu$ could be computed if the self-avoiding walk would posses a conformally invariant scaling limit. More precisely, let $\Omega\neq \mathbb C$ be a simply connected domain in the complex plane $\mathbb C$ with two points $a$ and $b$ on the boundary. For $\delta>0$, we consider the discrete approximation given by the largest finite domain $\Omega_\delta$ of $\delta\mathbb H$ included in $\Omega$, and $a_\delta$ and $b_\delta$ to be the vertices of $\Omega_\delta$ closest to $a$ and $b$ respectively. A probability measure $\mathbb P_{x,\delta}$ is defined on the set of self-avoiding trajectories $\gamma$ between $a_\delta$ and $b_\delta$ that remain in $\Omega_\delta$ by assigning to $\gamma$ a weight proportional to $x^{\ell(\gamma)}$. We obtain a random curve denoted $\gamma_\delta$. Conjectured conformal invariance of self-avoiding walks can be stated as follows, see [@LawlerSchrammWerner]: \[SLE\] Let $\Omega$ be a simply connected domain (not equal to $\mathbb C$) with two distinct points $a$, $b$ on its boundary. For $x=x_c$, the law of $\gamma_\delta$ in $(\Omega_\delta,a_\delta,b_\delta)$ converges when $\delta\rightarrow 0$ to the (chordal) Schramm-Loewner Evolution with parameter $\kappa=8/3$ in $\Omega$ from $a$ to $b$. As discussed in [@lsw-ust; @Smirnov], to prove convergence of a random curve to SLE it is sufficient to find a discrete observable with a conformally covariant scaling limit. Thus it would suffice to show that a normalized version of $F_\delta$ has a conformally invariant scaling limit, which can be achieved by showing that it is holomorphic and has prescribed boundary values. As discussed in [@smirnov-icm2010], the winding of an interface leading to a boundary edge $z$ is uniquely determined, and coincides with the winding of the boundary itself. Thus one can say that $F_\delta$ satisfies a discrete version of the following *Riemann boundary value problem* (a homogeneous version of the Riemann-Hilbert-Privalov BVP): $${\mathrm{Im}}\,\left({F(z)\cdot\left({\mathrm{tangent~to~}\partial\Omega}\right)^{5/8}}\right)~=~0~,~~~z\in\partial\Omega~, \label{eq:rbvp}$$ with a singularity at $a$. Note that the problem above has conformally covariant solutions (as $(dz)^{5/8}$-forms), and so is well defined even in domains with fractal boundaries. As noted in Remark \[rem:CR\], relation amounts to saying that discrete contour integrals of $F_\delta$ vanish. So any (subsequential) scaling limit of $F_\delta$ would have to be holomorphic. Unfortunately, relation alone, is unsufficient to deduce the existence of such a limit, unlike in the Ising case [@chelkak-smirnov-iso]. The reason is that for a domain with $E$ edges, imposes $\approx \frac23 E$ relations (one per vertice) for $E$ values of $F_\delta$, making it impossible to reconstruct $F_\delta$ from its boundary values. So $F_\delta$ is not exactly holomorphic, it can be rather thought of as a divergence-free vector field, which seems to have non-trivial curl. However, we expect that in the limit the curl vanishes, which is equivalent to $F_\delta(z)$ having the same limit regardless of the orientation of the edge $z$. The Riemann BVP is easily solved, and we arrive at the following conjecture: Let $\Omega$ be a simply connected domain (not equal to $\mathbb C$), let $z\in \Omega$, and let $a$, $b$ be two distinct points on the boundary of $\Omega$. We assume that the boundary of $\Omega$ is smooth near $b$. For $\delta>0$, let $F_\delta$ be the holomorphic observable in the domain $(\Omega_\delta,a_\delta)$ approximating $(\Omega,a)$, and let $z_\delta$ be the closest point in $\Omega_\delta$ to $z$. Then $$\lim_{\delta\rightarrow 0}~\frac{F_\delta(z_\delta)}{F_\delta(b_\delta)} ~=~\left(\frac{\phi'(z)}{\phi'(b)}\right)^{5/8}\label{convergence result}$$ where $\Phi$ is a conformal map from $\Omega$ to the upper half-plane mapping $a$ to $\infty$ and $b$ to 0. The right-hand side of is well-defined, since the conformal map $\phi$ is unique up to multiplication by a real factor. Proving this conjecture would be a major step toward Conjecture \[SLE\] and the derivation of critical exponents. #### Acknowledgements. The authors would like to thank G. Slade for useful comments on the manuscript, and G. Lawler for suggesting Remark \[remark\]. This research was supported by the EU Marie-Curie RTN CODY, the ERC AG CONFRA, as well as by the Swiss [FNS]{}. The second author was partially supported by the Chebyshev Laboratory (Department of Mathematics and Mechanics, St.-Petersburg State University) under RF governement grant 11.G34.31.0026 [1]{} J. Cardy and Y. Ikhlef, Discretely holomorphic parafermions and integrable loop models, **42(10), 102001** 11 pages (2009). D. Chelkak and S. Smirnov, Universality in the 2[D]{} [I]{}sing model and conformal invariance of fermionic observables. Invent. Math., to appear. Preprint, arXiv:0910.2045 (2009). P. Flory, Principles of Polymer Chemistry, (1953). N. Madras and G. Slade, Self-avoiding walks, (1993). J. M. Hammersley and D. J. A. Welsh, Further results on the rate of convergence to the connective constant of the hypercubical lattice. **13** 108–110 (1962). G. Lawler, O. Schramm and W. Werner, On the scaling limit of planar self-avoiding walk. , 339–364. [*Proc. Sympos. Pure. Math.*]{} [**72**]{}, [*Part 2, Amer. Math. Soc.*]{} Providence, RI (2004). G. Lawler, O. Schramm and W. Werner, Conformal invariance of planar loop-erased random walks and uniform spanning trees. , [**32(1B)**]{} 939–995 (2004). B. Nienhuis, Exact critical point and critical exponents of O(n) models in two dimensions. **49** 1062–1065 (1982). B. Nienhuis, Critical behavior of two-dimensional spin models and charge asymmetry in the Coulomb gas. **34** 731–761 (1984). S. Smirnov, Towards conformal invariance of 2D lattice models. **Vol. II** 1421–1451 (2006). S. Smirnov, . <span style="font-variant:small-caps;">Département de Mathématiques</span> <span style="font-variant:small-caps;">Université de Genève</span> <span style="font-variant:small-caps;">Genève, Switzerland</span> <span style="font-variant:small-caps;">E-mail:</span> `[email protected] ; [email protected]`
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the best approximation to the difference operators on the cyclic groups of order $2^n$ by a dyadic convolution operator are the restrictions of a generalized dyadic derivative. This answers a question on the “intuitive” interpretation of the dyadic derivative posed by Butzer and Wagner more than 30 years ago.' address: | Faculty of Mathematics\ University of Vienna\ Nordbergstrasse 15\ A-1090 Vienna, AUSTRIA author: - Andreas Klotz bibliography: - 'walshbib.bib' title: The Optimal Dyadic Derivative --- [^1] Introduction {#sec:introduction} ============ Background and Notation {#sect:not} ======================= The dyadic expansion of $x \in {\ensuremath{[0;1)}}$ is $x=\sum_{k=1}^ \infty x_k 2^{-k}$ with $x_k \in \set{0,1}$; if $x= 2^{-m}k$ for an integer $k$ we choose the expansion ending in zeros. The dyadic expansion of $n \in \bn_0$ is $n=\sum_{k=0}^\infty n_k 2^k$ with $n_k \in \set{0,1}$. Denote by ${\ensuremath{\bz_2}}$ the cyclic group of order 2, i.e. ${\ensuremath{\bz_2}}=\set{0,1}$ with addition modulo 2. The Haar measure on ${\ensuremath{\bz_2}}$ assigns the value $1/2$ to each singleton. The [dyadic group]{} ${\bz_2^\bn}=\prod_{k=1}^\infty{\ensuremath{\bz_2}}$ consists of all sequences $(x_k)_{k \in \bn}$ with $x_k \in {\ensuremath{\bz_2}}$. It is generated by the elements $e_k= (\delta_{jk})_{k \in \bn}$. Addition $\operatorname{\dot{+}}$ in ${\bz_2^\bn}$ is defined componentwise modulo 2: $(x \operatorname{\dot{+}}y)_k= x_k + y_k \mod 2$. With the natural product topology ${\bz_2^\bn}$ is a totally disconnected, compact abelian group. The Haar measure ${{d}}\lambda$ on ${\bz_2^\bn}$ is the product measure of the Haar measure on ${\ensuremath{\bz_2}}$ with total mass one. For $x \in {\bz_2^\bn}$ the dyadic balls $$B_n(x)=x \operatorname{\dot{+}}\set{z \in {\bz_2^\bn}\colon z_k=0, 1\leq k \leq n} = \set{z \in {\bz_2^\bn}\colon z_k=x_k, 1\leq k \leq n}$$ form a finite partition of ${\bz_2^\bn}$ for each nonnegative integer $n$, and are a basis of the topology of ${\bz_2^\bn}$. Dyadic balls are either disjoint or included in each other. The balls $B_n(0)$ are subgroups of ${\bz_2^\bn}$. In particular ${\bz_2^\bn}$ can be partitioned uniquely into the balls $ \set{B_n(x)}_{x \in {\bz_2^n}} $ for each $n \in \bn$. The characters on ${\bz_2^\bn}$, i.e. the group homomorphisms to the complex unit circle $\bt$ are the *[Walsh function]{}s*, defined for $x \in{\bz_2^\bn}$ and for $n \in \bno$ by $$w_n(x)=(-1)^{\sum_{k=0}^\infty n_k x_{k+1}} =(-1)^{{[ n,x ]}}\,,$$ where ${{[ n,x ]}}=\sum_{k=0}^\infty n_k x_{k+1} \mod 2$. The enumeration is called the Paley enumeration, and $w_n$ are the Walsh-Paley functions. The multiplicative group ${\mathbb{D}}=\set{w_m \colon m \in \bno}$ is the Pontryagin dual of the [dyadic group]{}, and an orthonormal basis of $L^2({\bz_2^\bn})$. If $m, n \in \bn_0$ with dyadic expansions as above, then dyadic addition $ m \operatorname{\dot{+}}n= \sum_{k=0}^\infty (m_k \operatorname{\dot{+}}n_k) 2^k $ turns $\bno$ into a group which is isomorphic to ${\mathbb{D}}$. In particular, for $x, y \in {\bz_2^\bn}$, and $m, n \in \bno$ $$\begin{aligned} w_mw_n&=w_{m \operatorname{\dot{+}}n} \,, \\ w_m(x \operatorname{\dot{+}}y) &=w_m(x)w_m(y)\,.\end{aligned}$$ For $n \in \bno$ the sets ${\mathbb{D}}_n = \set{w_m \in {\mathbb{D}}\colon m < 2^n}$ are multiplicative subgroups of ${\mathbb{D}}$ of order $2^n$, and the functions in ${\mathbb{D}}_n$ are constant on the dyadic intervals $B_n(x)$. In particular, the functions in ${\mathbb{D}}_n$ are uniquely determined by their values on ${\bz_2^n}$. In other words, the span of ${\mathbb{D}}_n$ is $$\mF_n=\set{f \colon {\bz_2^\bn}\to \bc , \quad f \; \text{ constant on cosets of } \; B_{n}} \,.$$ Sequency ordering ----------------- Besides the Paley ordering, there is another ordering of the [Walsh function]{}s – the *sequency* ordering – that is more closely related to the frequency concept of Fourier analysis. To define the sequency ordering we need the interpretation of [Walsh function]{}s as being defined on the unit interval. For $x \in {\ensuremath{[0;1)}}$ we define *Fine’s map* $\rho: {\ensuremath{[0;1)}}\to {\bz_2^\bn}; \; x \mapsto (x_k)_{k \in \bn}$, where the $x_k$ are the components of the dyadic expansion of $x$. The absolute value of $x \in {\bz_2^\bn}$ is $\abs{x}=\sum_{k=1}^ \infty x_k2^{-k}$, a metric on ${\bz_2^\bn}$ is given by $d(x,y)= \abs{x-y}$. By abuse of notation we use the absolute value also for $k \in \bno$. On the interval ${\ensuremath{[0;1)}}$ the [Walsh function]{}s are defined by $$\widetilde w_n(x)=w_n(\rho(x)) \,,$$ and the set $\widetilde {\mathbb{D}}= \set{\widetilde w_n\colon n \in \bno}$ with pointwise multiplication is isomorphic to ${\mathbb{D}}$ as a group; the definition of the subgroups $\widetilde {\mathbb{D}}_n$ is obvious. For $1 \leq p < \infty$ the Banach spaces $L^p({\ensuremath{[0;1)}})$ and $L^p({\bz_2^\bn})$ are isometrically isomorphic via the mapping (see, e.g.,[@ScWaSi90]) $$L^p({\bz_2^\bn}) \to L^p({\ensuremath{[0;1)}}); \; f \mapsto f \circ \abs{\phantom{x}}$$ We want to order the [Walsh function]{}s in increasing *sequency*, that is the number of sign changes on $(0;1)$. We develop the needed concepts from scratch, as we shall use them later, see also [@ScWaSi90] for a similar derivation. \[lem-sign-change\] The function $\widetilde{w_k} \in \widetilde{{\mathbb{D}}_n}$ changes sign at $x \in (0;1)$ only if $\rho(x) \in {\bz_2^n}$. If $x=\sum_{j=1}^{M(x)-1}x_j2^{-j} +2^{-M(x)}$, where $1 \leq M(x) \leq n$, set $h_{M(x)}=\sum_{k=M(x)}^N e_{k} $. Then $\widetilde{w_k} \in {\mathbb{D}}_n$ changes sign at $x$ if and only if ${[ k, h_{M(x)} ]}=1$. As the values of $w_k \in {\mathbb{D}}_n$ are constant on the balls ${B_n(x)}$, the functions $\widetilde w_k$ can change sign only at the points of the form $\sum_{j=1}^nx_j2^{-j}$, $x_j \in {\ensuremath{\bz_2}}$. So $\widetilde w_k$ changes sign at $x= \sum_{j=1}^nx_j2^{-j}$, if $$\label{eq:2} \widetilde w_k(x)=-\widetilde w_k({x}-2^{-n}) \,.$$ Equivalently, for $x=\sum_{j=1}^nx_je_{j}$ $$\label{eq:5} w_k(x \operatorname{\dot{+}}\rho( \abs{x-2^{-n}})=-1 \,.$$ If $x=\sum_{j=1}^{M(x)-1}x_je_j +e_{M(x)}$ for $1 \leq M(x) \leq n$, then $$\rho( \abs{x-2^{-n}})=\sum_{j=1}^{M(x)-1}x_je_j+ \sum_{j=M(x)+1}^ne_j \,,$$ so $$\label{eq:3} x \operatorname{\dot{+}}\rho( \abs{x-2^{-n}}) = \sum_{j=M(x)}^n e_j =h_{M(x)}\,.$$ The statement of the Lemma follows by combining equations (\[eq:5\]) and (\[eq:3\]). We note for later use that, by setting $M(0)=1$ and interpreting the difference $x-2^{-n}$ modulo $1$, i.e. $-2^{-n}=\sum_{j=1}^n2^{-j}$, Equation (\[eq:3\]) is also true for $x=0$. Define the ${\ensuremath{\bz_2}}$-linear mapping $\mS$ for $n=\sum_{j=0}^\infty n_j 2^j \in \bno$ by $ \mS n = \sum_{j=0}^\infty n_{j+1} 2^j \,. $ Set $\mG= {\ensuremath{\mathcal{I}}}+ \mS$, where [$\mathcal{I}$]{} is the identity on . The matrix representation of $\mG k$ on the [$\bz_2$]{}-subspace $\set{0,1, \dotsc, 2^n-1}$ of  with respect to the basis $2^l, l=0,\dotsc,n-1$ is $$\mG k= \begin{pmatrix} 1 & 1 & 0 & \hdotsfor{2} & 0 \\ 0 & 1 & 1 & 0 &\dotsc & 0 \\ \hdotsfor{6} \\ 0 & \hdotsfor{2} & 0& 1 & 1 \\ 0 & \hdotsfor{3} & 0& 1 \end{pmatrix} \begin{pmatrix} k_{0} \\ k_{1} \\ \dotsc \\k_{n-2}\\ k_{n-1} \end{pmatrix} = \begin{pmatrix} k_{0} \operatorname{\dot{+}}k_1 \\ k_{1} \operatorname{\dot{+}}k_{2} \\ \dotsc \\k_{n-2} \operatorname{\dot{+}}k_{n-1}\\ k_{n-1} \end{pmatrix}$$ \[prop:sequency-ordering\] The function $\widetilde w_{\mG k}$ has exactly $k$ sign-changes in $(0;1)$. Equation implies that whenever $w_k$ changes sign at $x \in {\bz_2^n}$ it changes sign for all $z \in {\bz_2^n}$ with $M(z)=M(x)$. Define $M_r=\set{x \in {\bz_2^n}\setminus \set{0} \colon M(x)=r}$ for $1 \leq r \leq n$. By definition of $M(x)$ we obtain that $\abs{M_r}=2^{r-1}$, and the disjoint union of the $M_r$ is $\bigcup_{r=1}^n M_r ={\bz_2^n}\setminus \set{0}$. As $w_k$ changes sign at $x \in M_r$, if and only if $$\begin{aligned} 1= {[ k, h_{r} ]} ={[ k, \sum_{i=r}^ne_i ]} = \sum_{i=r-1}^{n-1} k_i \mod 2 \,,\end{aligned}$$ the total number of sign changes of $w_k$ in the unit interval is $$\begin{aligned} & \sum_{r=1}^n \bigl( \sum_{i=r-1}^{n-1} k_i \mod 2 \bigr)2^{r-1} =\sum_{r=0}^{n-1} \bigl( \sum_{i=r}^{n-1} k_i \mod 2 \bigr)2^{r}\\ &= \begin{vmatrix} \begin{pmatrix} k_{n-1} \operatorname{\dot{+}}k_{n-2} \operatorname{\dot{+}}\dotsc \operatorname{\dot{+}}k_0\\ \dotsc \\ k_{n-1} \operatorname{\dot{+}}k_{n-2} \\ k_{n-1} \end{pmatrix} \end{vmatrix} = \begin{vmatrix} \begin{pmatrix} 1 & 1 & \hdotsfor{2} & 1\\ 0 & 1 & \hdotsfor{2} & 1\\ \hdotsfor{5} \\ 0 & 0 & \dotsc & 0 & 1 \end{pmatrix} \begin{pmatrix} k_{0} \\ k_{1} \\ \dotsc \\ k_{n-1} \end{pmatrix} \end{vmatrix} \\ &= \bigabs{\sum_{i=0}^{n-1} \mS^i k} = \abs{\mG ^{-1}k} \,.\end{aligned}$$ So $w_{Gk}$ has $k$ sign changes in $(0;1)$. Best Approximation by Dyadic Convolution Operators {#sec:sequency-ordering} ================================================== [Hilbert-Schmidt]{} norm {#sec:hs-norm} ------------------------ Let ${\ensuremath{\mathcal{H}}}$ be a Hilbert space with orthonormal basis $\set{e_\lambda}_{\lambda \in \Lambda}$. The Hilbert-Schmidt-norm (or Frobenius norm) of an operator $A \in \mB({\ensuremath{\mathcal{H}}})$ is $\norm{A}_{{F}}= \bigl(\sum_{\lambda \in \Lambda}\norm{A e_\lambda}_{\ensuremath{\mathcal{H}}}^2 \bigr)^{1/2} $, and this norm is independent of the choice of the basis. It is well known that this norm stems from an inner product. With the trace operator $\operatorname{trace}{A}=\sum_{\lambda \in \Lambda} \inprod {A e_\lambda, e_\lambda}$ this inner product is $ \inprod{A,B}_{{F}}= \operatorname{trace}(B^* A) \,. $ Therefore the (unique) best approximation of a [Hilbert-Schmidt]{} operator $A$ on a (closed) subspace of the [Hilbert-Schmidt]{} operators in the [Hilbert-Schmidt]{}-norm is the orthogonal projection on this subspace with respect to the [Hilbert-Schmidt]{} inner product. We will need the following immediate consequence of the definition of a [Hilbert-Schmidt]{} operator. \[prop:hs–norm-equivalence\] If $A$ is a [Hilbert-Schmidt]{} operator on ${\ensuremath{\mathcal{H}}}$, and $T \colon {\ensuremath{\mathcal{H}}}\to {\ensuremath{\mathcal{K}}}$ a Hilbert space isomorphism, then $TA \inv T $ is a [Hilbert-Schmidt]{} operator on ${\ensuremath{\mathcal{K}}}$ with the same norm. [Walsh transform]{} and dyadic convolution operators {#sec:wt-dyad-conv} ---------------------------------------------------- Let $f,g \in L^1({\bz_2^\bn})$. The dyadic convolution of $f$ and $g$ is $$f \operatorname{\dot{\ast}}g (x) =\int_{t \in {\bz_2^\bn}} f(x \operatorname{\dot{+}}t) g(t) \dd \lambda (t) \,.$$ The dyadic convolution operator $C_f$ is $C_f g = f \operatorname{\dot{\ast}}g$.\ The *Walsh Transform* of $f$ is $$ {\ensuremath{\mathcal{W}}}f (k)= f^\sim(k) = \inprod{f,w_k}=\int_{t \in {\bz_2^\bn}} f(t) w_k(t) \dd \lambda(t) \, \text{ for } k \in \bno \,.$$ The [Walsh transform]{} is an isometry between $L^2({\bz_2^\bn})$ and $\ell^2(\bno)$. It is straightforward that for $f,g \in \mF_n$ $$f^\sim (k) = \begin{cases} 2^{-n}\sum_{t \in {\bz_2^n}} f(t) w_k(t), \quad & k < 2^n\\ 0, \quad & \text{else} \,, \end{cases}$$ and $$f(x)=\sum_{k=0}^{2^n-1} f^\sim(k)w_k(x) \,.$$ Following an approach of Pearl [@Pearl75] we want to characterize the best approximation of operators on $L^2({\bz_2^\bn})$ by dyadic convolution operators $C_f$, $f \in L^2({\bz_2^\bn})$. \[prop:bestappr-dyconvop\] Assume that $A$ is a [Hilbert-Schmidt]{}-operator on $L^2({\bz_2^\bn})$. Then the best approximation of $A$ by a dyadic convolution operator $C_f$ in the [Hilbert-Schmidt]{}-norm is given by $f^\sim(k)= \inprod{A w_k, w_k}$. Consider the commutative diagram $$\begin{CD} L^2({\bz_2^\bn}) @>A >> L^2({\bz_2^\bn})\\ @V \simeq V {\ensuremath{\mathcal{W}}}V @V \simeq V {\ensuremath{\mathcal{W}}}V\\ \ell^2(\bno) @> \tilde A >> \ell^2(\bno) \end{CD}$$ i.e. $\tilde A \tilde f = {\ensuremath{\mathcal{W}}}(A f)$, and $\norm{A}_{{F}}=\norm{\tilde A}_{{F}}$. For the entries of the matrix $\tilde A$ we obtain $$\begin{aligned} \tilde A(k,l)=&\inprod{\tilde A e_l, e_k}_{\ell^2(\bno)} =\inprod{\tilde A \,{\ensuremath{\mathcal{W}}}\, w_l, {\ensuremath{\mathcal{W}}}\, w_k}_{\ell^2(\bno)} \\ =&\inprod{{\ensuremath{\mathcal{W}}}^* \tilde A \, {\ensuremath{\mathcal{W}}}\, w_l, w_k}_{L^2({\bz_2^\bn})} = \inprod{ A w_l, w_k}_{L^2({\bz_2^\bn})} \,.\end{aligned}$$ As $\widetilde C_f =\diag (\tilde f)$, where $\diag$ denotes a diagonal matrix, we see that minimal norm $\norm{A-C_f}_{{F}}$ is obtained for $\tilde f(k) = \inprod{A w_k, w_k}$. We still have to show that $f \in L^2({\bz_2^\bn})$: $$\norm{f}_2^2=\norm{\tilde f}_2^2=\sum_{k=0}^\infty \abs{\inprod{A \, w_k, w_k}}^2 \leq \sum_{k=0}^\infty \norm{A\,w_k}_2^2 =\norm{A}_{{F}}^2 \,. \qedhere$$ Best Approximation of Cyclic Difference Operators {#sec:best-appr-cycl} ------------------------------------------------- We will need an appropriate definition of a dyadic derivative. In [@BuWa73; @BuWa75] Butzer and Wagner introduced the concept of a *dyadic derivative*, which was an extension of the *logical derivative* used by Gibbs [@gibbs1970] on ${\bz_2^n}$. This operator has the property that $$D w_k = \abs k w_k \text{ for all } k \in \bno \,,$$ thus mimicking the behavior of the classical differentiation operator on the exponentials. In [@On79], Onneweer mentioned the dependence of the definition on the ordering on the [Walsh function]{}s, and offered an alternative definition of a dyadic derivative with the property that $$D w_k = 2^{\floor{\log_2 \abs k}} w_k \,.$$ We follow the definition of He Zelin  [@Zelin93], who defined a generalized dyadic derivative. We adapt the definition for ${\bz_2^\bn}$. Assume that $\gamma= (\gamma(k))_{k \in \bno}$ is a sequence of complex numbers. The *generalized dyadic derivative* $f \in L^1({\bz_2^\bn})$ is $$D_\gamma f = \lim_{n \to \infty} f \operatorname{\dot{\ast}}{\sum_{k=0}^{2^n-1} \gamma (k) w_k} \,,\label{eq:6}$$ whenever the expression on the right side converges in $L^1({\bz_2^\bn})$. The set of all $f \in L^1({\bz_2^\bn})$ such that the limit exists is the *domain* $\mD(D_\gamma)$ of $D_\gamma$. The generalized dyadic derivative is a closed operator from $\mD(D_\gamma)$ to $L^1({\bz_2^\bn})$. This follows form [@Zelin93 Thm 1]. It is elementary that $$D_\gamma w_k = \gamma(k) w_k \,.$$ The case $\gamma(k)=\abs{k}$ corresponds to the dyadic derivative considered by Butzer and Wagner, and $\gamma(k)= 2^{\floor{\log_2\abs k}}$ is the dyadic derivative of Onneweer. For the formulation of the main theorem we need to introduce the translation operator on $[0;1)$ as $T_xf (t)= f(t-x)$, the difference is to be understood modulo 1, and the difference operator $\Delta_n = 2^n (T_{2^{-n}} -I)$. By abuse of notation these operators are also defined on ${\bz_2^n}$, if $x \in {\bz_2^n}$. \[thm\_main\] There is a unique generalized dyadic derivative $D_\gamma$ that approximates the ordinary derivative in the following sense: For any $n \in \bn$, the restriction of $D_\gamma$ is the best approximation to $\Delta_n$ in the class of dyadic convolution operator, measured in the [Hilbert-Schmidt]{} norm, $$\label{eq:7} \norm{D_\gamma\vert_{\mF_n} - \Delta_n}_{{F}} = \min \set{\norm{C_f-\Delta_n}_{{F}} \colon f \in \mF_n} \,.$$ The coefficients of $\gamma$ are $$\gamma(\mG k) = 2 (k_0 + \abs k)$$ In particular, $$D_\gamma f = 2 \lim_{n \to \infty} f \operatorname{\dot{\ast}}\sum_{k=0}^{2^n-1} (k_0 +\abs k) w_{\mG k} \,.$$ \(a) It is remarkable that *one* operator satisfies (\[eq:7\]) for *all* indices $n$.\ (b) The theorem answers in a way the question posed by Butzer and Wagner: There is a (generalized) dyadic derivative that can be uniquely described as the limit of optimal approximations to classical difference operators by dyadic convolution operators. So in this way the dyadic derivative is the best approximation to the classical differentiation operator. It should be no surprise that the description of $D_\gamma$ is simplest for the sequence ordering of the [Walsh function]{}s, as sequency mimics frequency. (Another example would be the position of the maximum of the Fourier transform of [Walsh function]{}s, which can be described most transparently for the sequency ordering.)\ (c) On an $n$-dimensional Hilbert space the [Hilbert-Schmidt]{}-norm admits a statistical interpretation [@boettcher03 Thm. 2.2]: If $X$ is a uniformly distributed random variable on the unit sphere $S_{n-1}$ then $$ \norm{A}_{{F}}^2= n \mathbb{E}(\norm{AX}_2^2) $$ where $\mathbb{E}$ is the expectation operator. (Remark: A simple proof of this statement can be based on the divergence theorem.) So the generalized dyadic derivative obtained in the theorem approximates the classical difference operators best in a statistical sense. This might be of interest in signal processing applications. The statements of theorem are a simple consequence of the following lemma. \[lem-main\] The best approximation of the cyclic translation operator $T_{2^{-n}}$ by a dyadic convolution operator $C_f$ on $\mF_n$ in the [Hilbert-Schmidt]{}-norm satisfies $$f^\sim (\mG m)= 1-2^{-n+1}(\abs m +m_0)$$ This follows from $$\begin{aligned} f^\sim (k) &= 2^{-n}\sum_{x \in {\bz_2^n}} T_{2^{-n}}w_k(x)w_k(x)= 2^{-n}\sum_{x \in {\bz_2^n}} w_k(\phi(\abs x-2^{-n}) \operatorname{\dot{+}}x) \\ &= 2^{-n}\sum_{x \in {\bz_2^n}} w_k(h_M(x))=2^{-n}(w_k(h_M(0))+\sum_{x \in {\bz_2^n}\setminus \set{0}} w_k(h_M(x))\end{aligned}$$ by . As $\abs {M_k} = 2^{k-1}$ for $k>0$ and $h_m(0)=h_1$ the sum above can be rewritten as $$\begin{aligned} f^\sim (k) &=2^{-n}\bigl(w_k(h_1)+\sum_{r=1}^n w_k(h_r)2^{r-1}\bigr) \\ & =2^{-n}\bigl((-1)^{{[ k,h_1 ]}}+\sum_{r=1}^n (-1)^{{[ k,h_r ]}}2^{r-1}\bigr)\end{aligned}$$ We now use that $(-1)^m = 1 - 2 m$ for $m \in {\ensuremath{\bz_2}}$ to obtain $$\begin{aligned} f^\sim (k) &= 2^{-n}\bigl(1-2{{[ k,h_1 ]}}+\sum_{r=1}^n (1- 2{{[ k,h_r ]}})2^{r-1}\bigr) \\ &= 2^{-n}\bigl(-2{{[ k,h_1 ]}}+ 2^n -2\sum_{r=1}^n {{[ k,h_r ]}}2^{r-1}\bigr) \\ &= 2^{-n}\bigl(-2{{[ k,h_1 ]}}+ 2^n -2\abs{\inv \mG k}\bigr) \,.\end{aligned}$$ Substituting $k=\mG m$ yields $$f^\sim (\mG m) =1 - 2^{-n+1}(m_0 + \abs m) \,,$$ and that is what we wanted to prove. We only have to observe that $$T_{2^{-n}}-C_f = (T_{2^{-n}}-I) -(C_f-I)=\Delta_n-C_{f-\delta} \qedhere$$ The obtained result is in a way rather peculiar, as is shown in the following two examples. The best approximation of the symmetric difference operator $2^{n-1}(T_{2^{-n}}-T_{2^{-n}})$ by a dyadic convolution operator is by the zero operator. This follows easily from the inspection of the diagonal elements (see Proposition \[prop:bestappr-dyconvop\]) $$\inprod{T_{2^{-n}}w_k,w_k} - \inprod{T_{-2^{-n}}w_k,w_k}=0 \,,$$ as $$\inprod{T_{-2^{-n}}w_k,w_k}=\inprod{w_k,T_{2^{-n}}w_k}=\inprod{T_{2^{-n}}w_k,w_k} \,.$$ The operator $\mJ$ of *anti-differentiation* on [$[0;1)$]{} is $$\mJ f(x) = \int_0^x f(t) \dd t \,.$$ Let us determine the best approximation of the [Hilbert-Schmidt]{}-operator $\mJ$ by a dyadic convolution operator. By Proposition \[prop:bestappr-dyconvop\] it is sufficient to compute $\inprod{\mJ w_k, w_k}$. Using the expansion  [@Fi49 Eq.(3.6)], valid for $k \geq 1$, $$\mJ w_k = 2^{-n-2} (w_{k'}-\sum_{r=1}^\infty 2^{-r}w_{2^{n+r}+k}) \,,$$ where $k\geq1$, $k=2^n+k'$, and $0 \leq k' <2^n$, we obtain that $\inprod{\mJ w_k,w_k} =0$ for all $k\geq 1$. On the other hand it is straightforward that $\inprod{\mJ w_0, w_0}=1/2$, so $\gamma_0=1/2$ and $\gamma_k=0$ for all $k \geq 1$. The resulting operator is $$D_\gamma f = \frac{1}{2} \int_0^1 f(\lambda) \dd \lambda \,.$$ We end this note with some questions, which might be tractable by the methods used above. 1. What are the best approximations of $\Delta_n^k$ , $k>1$ by dyadic convolution operators? 2. It is possible to adapt the approach given above and to consider classical differentiation operators on the space of trigonometric polynomials of degree $n$ and their approximation by dyadic convolution operators. Does this change the result of Theorem \[thm\_main\]? 3. Can a similar result be obtained if the [Hilbert-Schmidt]{} norm is replaced by, e.g., the operator norm? 4. What is the generalization to Vilenkin groups? [^1]: A. K.  was supported by the FWF project P22746N13
{ "pile_set_name": "ArXiv" }
--- abstract: 'We compute the Kodaira dimension of the universal Picard variety ${P_{d,g}}$ parameterizing line bundles of degree $d$ on curves of genus $g$ under the assumption that $(d-g+1,2g-2)=1$. We also give partial results for arbitrary degrees $d$ and we investigate for which degrees the universal Picard varieties are birational.' author: - 'Gilberto Bini, Claudio Fontanari, Filippo Viviani' title: | On the birational geometry\ of the universal Picard variety --- Introduction ============ The study of the birational geometry of the moduli spaces has become a very active research area after the unexpected result of Harris-Mumford-Eisenbud ([@HarMum], [@EH]) that the moduli space $M_g$ of curves of genus $g$ is a variety of general type for $g \ge 24$, contradicting a long-standing conjecture of Severi on the unirationality of moduli of curves. More recently, also the birational geometry of other moduli spaces has been widely investigated: the moduli space of pointed curves ([@Log], [@CF]), the moduli space of Prym varieties ([@FL]), the moduli space of spin curves ([@Lud2], [@Far], [@FV]), to mention at least some contributions in this area. The aim of the present paper is to investigate the birational geometry of the universal Picard variety $$P_{d,g} \to M_g,$$ parameterizing smooth curves of genus $g$ together with a line bundle of degree $d$. The following result is due to Verra (see [@Verra Thm. 1.2]). *([@Verra])*\[Ver-thm\] The variety $P_{d,g}$ is unirational for $g\leq 9$ and any $d$. Our main result is the computation of the Kodaira dimension of $P_{d,g}$ with $g\geq 10$ under a technical assumption on the degree $d$. Recall that, since $P_{d,g}$ is singular and not projective, the Kodaira dimension of $P_{d,g}$, which we denote by $\kappa(P_{d,g})$, is defined as the Kodaira dimension of any smooth projective model of it (see [@Laz Example 2.1.5]). The previous result of Verra implies that $\kappa(P_{d,g})=-\infty$ for $g\leq 9$ and any $d$. \[Kod-geom\] Assume that $(d-g+1,2g-2)=1$ and $g\geq 10$. The Kodaira dimension of ${P_{d,g}}$ is equal to $$\kappa({P_{d,g}})=\begin{cases} 0 & \text{ if } g=10,\\ 19 & \text{ if } g=11,\\ 3g-3 & \text{ if } g\geq 12. \end{cases}$$ In Propositions \[dim-Kod11\] and \[dim-Kod\], we also determine the Iitaka fibration (see [@Laz Def. 1.3.6]) of ${P_{d,g}}$ in the non-trivial cases, namely for $g\geq 11$. Without any assumption on the degree $d$, we obtain the following partial result: \[kod-nongeo\] The Kodaira dimension of ${P_{d,g}}$ (for $g\geq 10$) satisfies the following inequalities $$\kappa({P_{d,g}})\leq \begin{cases} 0 & \text{ if } g=10,\\ 19 & \text{ if } g=11,\\ 3g-3 & \text{ if } g\geq 12. \end{cases}$$ Moreover, $\kappa({P_{d,g}})=3g-3$ if $\kappa({M_g})\geq 0$ (and in particular for $g\geq 22$). Let us now explain the strategy that we use to prove the above results. The main tool we use is the GIT compactification constructed by Caporaso (see [@Cap]) $$\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$$ of $P_{d,g}$ over the Deligne-Mumford moduli space ${\overline{M}_g}$ of stable curves of genus $g$. The projective normal variety ${\overline{P}_{d,g}}$ is a good moduli space for the stack ${\overline{\mathcal Pic}_{d,g}}$ (see [@Capbis] and [@Melo]), whose section over a scheme $S$ is the groupoid ${\overline{\mathcal Pic}_{d,g}}(S)$ of families of quasistable curves of genus $g$ $$f: (\mathcal{C}, \mathcal{L}) \to S$$ endowed with a balanced line bundle ${\mathcal{L}}$ of degree $d$ (see \[def-stack-scheme\] for details). Furthermore, ${\overline{P}_{d,g}}$ is a coarse moduli scheme for ${\overline{\mathcal Pic}_{d,g}}$ if and only if $(d-g+1,2g-2)=1$, which is precisely the numerical hypothesis on the degree $d$ in Theorem \[Kod-geom\]. Albeit ${\overline{P}_{d,g}}$ is singular, we can prove (under the same assumption on the degree) that ${\overline{P}_{d,g}}$ has canonical singularities and therefore pluricanonical forms on the smooth locus lift to any desingularization: \[conj-sing\] Assume that $(d-g+1,2g-2)=1$ and that $g\geq 4$. Then ${\overline{P}_{d,g}}$ has canonical singularities. In particular, if ${\widetilde}{{\overline{P}_{d,g}}}$ is a resolution of singularities of ${\overline{P}_{d,g}}$, then every pluricanonical form defined on the smooth locus ${\overline{P}_{d,g}}^{\rm reg}$ of ${\overline{P}_{d,g}}$ extends holomorphically to ${\widetilde}{{\overline{P}_{d,g}}}$, that is, for all integers $m$ we have $$h^0({\overline{P}_{d,g}}^{\rm reg},m K_{{\overline{P}_{d,g}}^{\rm reg}})=h^0({\widetilde}{{\overline{P}_{d,g}}}, m K_{{\widetilde}{{P_{d,g}}}}).$$ The proof of this theorem is given in Section \[Sec-sing\]. The restriction on the degree $d$ comes from the fact that ${\overline{P}_{d,g}}$ has finite quotient singularities if and only if $(d-g+1,2g-2)=1$; hence only for such degrees $d$ we can apply the Reid–Tai criterion for the canonicity of finite quotient singularities (see e.g. [@HarMum p. 27] or [@Lud Thm. 4.1.11]). Indeed, we establish in Theorem \[sing-Pdst\] a similar statement without any restriction on $d$ for the open subset ${{\overline}{P}_{d,g}^{\rm st}}\subset {\overline{P}_{d,g}}$ of GIT-stable points of ${\overline{P}_{d,g}}$, which coincides with ${\overline{P}_{d,g}}$ if and only if $(d-g+1,2g-2)=1$. In proving Theorem \[sing-Pdst\] (from which Theorem \[conj-sing\] follows), we determine the non-smooth locus of ${{\overline}{P}_{d,g}^{\rm st}}$ in Proposition \[sing\]. Note that a proof of Theorem \[conj-sing\] for all degrees $d$ would imply the validity of Theorem \[Kod-geom\] without any assumptions on the degree $d$. The above Theorem \[conj-sing\] is crucial for our purposes because it allows us to compute the Kodaira dimension of ${\overline{P}_{d,g}}$ as the Iitaka dimension (see [@Laz Def. 2.1.3]) of the canonical divisor $K_{{\overline{P}_{d,g}}}$ on the modular variety ${\overline{P}_{d,g}}$, instead of working on some (a priori non modular) desingularization of ${\overline{P}_{d,g}}$. The class of $K_{{\overline{P}_{d,g}}}$ is given by the following \[K\] For any $g\geq 4$, we have $$K_{{\overline{P}_{d,g}}}= \phi_d^*(14\lambda-2\delta),$$ where $\lambda, \delta$ denote the Hodge and the boundary class on ${\overline{M}_g}$, respectively. The proof of this theorem is given in Section \[canonical\]. We first compute in Theorem \[can-stack\] the canonical class of ${\overline{\mathcal Pic}_{d,g}}$ through a careful application of the Grothendieck-Riemann-Roch theorem to the universal family over ${\overline{\mathcal Pic}_{d,g}}$. Then we show that the pull-back of $K_{{\overline{P}_{d,g}}}$ via the canonical map $p:{\overline{\mathcal Pic}_{d,g}}\to {\overline{P}_{d,g}}$ is equal to $K_{{\overline{\mathcal Pic}_{d,g}}}$. Note that this is in contrast with what happens for ${\overline{M}_g}$ (or for the moduli space of Prym or spin curves), where the pull-back of the canonical class of the coarse moduli space is equal to the canonical class of the moduli stack plus some (small) corrections at the boundary. Theorem \[K\] allows us to compute the Iitaka dimension of $K_{{\overline{P}_{d,g}}}$ as the Iitaka dimension of the divisor $14\lambda-2\delta$ on ${\overline{M}_g}$ (because $\phi_d$ is a regular fibration). By exploiting the rich available knowledge on the birational geometry of ${\overline{M}_g}$, we prove the following \[Iitaka-Mg\] The Iitaka dimension of $K_{{\overline{P}_{d,g}}}$ on ${\overline{P}_{d,g}}$ is equal to $$\kappa(K_{{\overline{P}_{d,g}}})=\kappa(14\lambda-2\delta)=\begin{cases} -\infty &\text{ if } g\leq 9,\\ 0 & \text{ if } g=10,\\ 19 & \text{ if } g=11,\\ 3g-3 & \text{ if } g\geq 12. \end{cases}$$ The proof of the above theorem is given in Section \[Iitaka\] by combining Propositions \[kod-infty\], \[dim-Kod\], \[dim-Kod10\], \[dim-Kod11\]. With the above results it is now easy to prove Theorems \[Kod-geom\] and \[kod-nongeo\]. Indeed, note that we have always the inequality $$\label{comp-Kod} \kappa({P_{d,g}})\leq \kappa({K_{{\overline{P}_{d,g}}}}),$$ with equality if $(d-g+1,2g-2)=1$ by Theorem \[conj-sing\]. From (\[comp-Kod\]) and Theorem \[Iitaka-Mg\], we deduce Theorem \[Kod-geom\] and the first part of Theorem \[kod-nongeo\]. The second part of Theorem \[kod-nongeo\] follows from Proposition \[kod-fib\], which is proved in Section \[Fibrat\] via a careful analysis of the regular fibration $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$. In the final Section \[bir-Pdg\], inspired by Lemma 8.1 in [@Cap], we investigate for which values of $d$ and $d'$ the varieties $P_{d,g}$ and $P_{d',g}$ are birational. We prove the following \[bira-thm\] Assume that $g\geq 22$ or $g\geq 12$ and $(d-g+1,2g-2)=1$. Then $P_{d,g}$ is birational to $P_{d',g}$ if and only if $d'\equiv \pm d \mod (2g-2)$. In this case, $P_{d,g}$ is isomorphic to $P_{d',g}$. This follows from Theorem \[birationa\], where we also determine the possible birational maps between the varieties $P_{d,g}$ for $g$ big enough. From the same result, we obtain a description the group of birational self-maps of $P_{d,g}$ (see Corollary \[Cor-bira1\]) and we deduce that the boundary of ${\overline}{P}_{d,g}$ is preserved by any automorphism of ${\overline}{P}_{d,g}$ (see Corollary \[Cor-bira2\]). While this work was being written down, Farkas and Verra posted on the arXiv the preprint [@FarVer], where they determine, among other things, the Kodaira dimension of $P_{g,g}$ (note that the degree $g$ satisfies the assumptions of our Theorem \[Kod-geom\], so that their result is a particular case of our main theorem). However, their strategy is different from ours and it seems to apply only in the special case $d=g$. Indeed, the authors of loc. cit. consider the global Abel-Jacobi map $$A_{g,d}: M_{g,d}/S_d\to P_{d,g},$$ obtained by sending a curve $C$ together with a collection of unordered points $\{p_1,\ldots, p_d\}$ into the pair $(C,{\mathcal{O}}_C(p_1+\cdots+p_d))$. It is well-known that the map $A_{g,d}$ is a birational isomorphism in degree $d=g$ (and only in this case). Using this fact, Farkas and Verra determine the Kodaira dimension of $P_{g,g}$ by studying the pluricanonical forms on the Deligne-Mumford-Knudsen compactification ${\overline}{M}_{g,g}/S_g$ (instead of the Caporaso compactification ${\overline}{P}_{g,g}$, as we do in this paper). Throughout this paper, we work over the complex field $\mathbb{C}$. Moreover, we fix two integers $g\geq 2$ and $d$. Preliminaries ============= [*The stack $\pdb$ and the scheme ${\overline{P}_{d,g}}$*]{}\[def-stack-scheme\] In this subsection, we recall the definition of the stack ${\overline{\mathcal Pic}_{d,g}}$ and its good moduli space ${\overline{P}_{d,g}}$, and collect some of their properties to be used later on. Let ${{\mathcal Pic}_{d,g}}$ be the universal Picard stack over the moduli stack ${\mathcal{M}_g}$ of smooth curves of genus $g$. The fiber ${{\mathcal Pic}_{d,g}}(S)$ of ${{\mathcal Pic}_{d,g}}$ over a scheme $S$ is the groupoid whose objects are families of smooth curves ${\mathcal{C}}\to S$ endowed with a line bundle ${\mathcal{L}}$ over ${\mathcal{C}}$ of relative degree $d$ over $S$ and whose arrows are the obvious isomorphisms. $\gd$ is a smooth irreducible (Artin) algebraic stack of dimension $4g-4$ endowed with a natural forgetful map $\Phi_d: {{\mathcal Pic}_{d,g}}\to {\mathcal{M}_g}$. The stack ${{\mathcal Pic}_{d,g}}$ admits a good moduli scheme ${P_{d,g}}$ of dimension $4g-3$ which has a natural forgetful map $\phi_d: {P_{d,g}}\to {M_g}$ onto the coarse moduli scheme of smooth curves of genus $g$. We have the following commutative diagram $$\label{diag0} \xymatrix{ {{\mathcal Pic}_{d,g}}\ar[r] \ar@{->>}_{\Phi_d}[d] &{P_{d,g}}\ar@{->>}^{\phi_d}[d]\\ {\mathcal{M}_g}\ar[r] & {M_g}. }$$ \[gerbe\] The fact that ${{\mathcal Pic}_{d,g}}$ has dimension $4g-4$ (and not $4g-3$ as ${P_{d,g}}$) is due to the fact that on each object $({\mathcal{C}}\to S,{\mathcal{L}})$ of ${{\mathcal Pic}_{d,g}}(S)$ there is an action of the multiplicative group ${\mathbb{G}_m}$ via scalar multiplication on ${\mathcal{L}}$. Therefore the map $\Phi_d$ factors as $$\Phi_d:{{\mathcal Pic}_{d,g}}\to {{\mathcal Pic}_{d,g}}^{{\mathbb{G}_m}}\to P_{d,g},$$ where ${{\mathcal Pic}_{d,g}}^{{\mathbb{G}_m}}$ (which is denoted by ${{\mathcal Pic}_{d,g}}\fatslash {\mathbb{G}_m}$ by some authors) is the ${\mathbb{G}_m}$-rigidification of ${{\mathcal Pic}_{d,g}}$ along the subgroup ${\mathbb{G}_m}$. Note that ${{\mathcal Pic}_{d,g}}^{{\mathbb{G}_m}}$ is a Deligne-Mumford stack of dimension $4g-3$ while ${{\mathcal Pic}_{d,g}}$ is an Artin stack of dimension $4g-4$ which is [not]{} Deligne-Mumford. However, we will never need the rigidified stack ${{\mathcal Pic}_{d,g}}^{{\mathbb{G}_m}}$ in this work so that we refer to [@Melo Sec. 4] for more details (note that in loc. cit. our stack ${{\mathcal Pic}_{d,g}}$ is denoted by $\mathcal{G}_{d,g}$ while its rigidification ${{\mathcal Pic}_{d,g}}^{{\mathbb{G}_m}}$ is denoted by $\mathcal{P}_{d,g}$). The stack ${{\mathcal Pic}_{d,g}}$ and the scheme ${P_{d,g}}$ have been compactified in a modular way in [@Cap], [@Capbis] and [@Melo]. To describe these compactifications, we need to recall some definitions. \[quasi-stable\] A connected, projective nodal curve $C$ is said to be *quasistable* if it is (Deligne-Mumford) semistable and the exceptional components of $C$ do not meet. Given a quasi-stable curve $C$, we will denote by $C_{\rm exc}$ the subcurve of $C$ (called exceptional subcurve) given by the union of all the exceptional components of $C$; by ${\widetilde}{C}:=\overline{C\setminus C_{\rm exc}}$ its complementary subcurve (called non-exceptional subcurve) and by $C^{\rm st}$ the stabilization of $C$. Moreover, we will denote by $\gamma({\widetilde}{C})$ the number of connected components of ${\widetilde}{C}$. \[balanced\] Let $C$ be a quasistable curve of genus $g\ge 2$ and $L$ a degree $d$ line bundle on $C$. (i) \[bala1\] We say that $L$ is *balanced* if - for every subcurve $Z$ of $C$ the following (“Basic Inequality”) holds $$\label{basic} \frac{d \deg_Z ({\omega_C}_{|Z})}{2g-2}-\frac {k_Z}2\le \deg_ZL\leq \frac{d \deg_Z ({\omega_C}_{|Z})}{2g-2}+\frac{k_Z}2,$$ where $k_Z$ is the number of intersection points of $Z$ with the complementary subcurve $Z^c:={\overline}{C\setminus Z}$. - $\deg_EL=1$ for every exceptional component $E$ of $C$. (ii) \[bala2\]We say that $L$ is *strictly balanced* if it is balanced and if for each proper subcurve $Z$ of $C$ for which one of the two inequalities in (\[basic\]) is not strict, then the intersection $Z\cap Z^c$ is contained $C_{\rm exc}$. (iii) \[bala3\] We say that $L$ is *stably balanced* if it is balanced and if for each proper subcurve $Z$ of $C$ for which one of the two inequalities in (\[basic\]) is not strict, then either $Z$ or $Z^c$ is entirely contained in $C_{\rm exc}$. The above Definitions \[balanced\] and \[balanced\] are taken from [@CCC Def. 5.1.1] (see also [@Capbis Def. 4.6]) and they are equivalent, respectively, to the definitions of semistable in [@Cap Sec. 5.5] and G-stable in [@Cap Sec. 6.2]. The Definition \[balanced\] is taken from [@Cap-Lis Sec. 4.1] and it is equivalent to the definition of [extremal]{} in [@Cap Sec. 5.2]. There is an equivalence relation of the set of balanced line bundles on a quasi-stable curve $C$. \[equiv-rel\] Given two balanced line bundles $L$ and $L'$ on a quasi-stable curve $C$, we say that $L$ and $L'$ are [equivalent]{}, and we write $(C,L)\equiv (C,L')$, if $L_{|{\widetilde}{C}}\cong L'_{|{\widetilde}{C}}$. The equivalence class of a pair $(C,L)$ is denoted by $[(C,L)]$. Note that the above equivalence relation $\equiv$ clearly preserves the multidegree of the line bundles, hence it preserves the condition of being strictly balanced or stably balanced. \[GIT-inter\] In the GIT construction of ${\overline{P}_{d,g}}$ given in [@Cap], the equivalence classes $[(C,L)]$ such that $C$ is quasi-stable and $L$ is balanced (resp. strictly balanced, resp. stably balanced) correspond to the GIT-semistable (resp. GIT-polystable, resp. GIT-stable) orbits (see [@Cap Prop. 6.1, Lemma 6.1] and also [@CCC Thm. 5.1.6]). The relationship between stably balanced and strictly balanced line bundles is given by the following \[compa-bal\] A line bundle $L$ on a quasi-stable curve $C$ is stably balanced if and only if it is strictly balanced and ${\widetilde}{C}$ is connected. Assume first that $L$ is strictly balanced and that ${\widetilde}{C}$ is connected. Let $Z$ be a proper subcurve of $C$ such that one of the two inequalities in is not strict. Then $Z\cap Z^c\subset C_{\rm exc}$ because $L$ is strictly balanced by hypothesis. Therefore the non-exceptional subcurve ${\widetilde}{C}$ can be written as a disjoint union of the two subcurves $Z\cap {\widetilde}{C}$ and $Z^c\cap {\widetilde}{C}$. Since ${\widetilde}{C}$ is connected by hypothesis, we must have that either $Z\cap {\widetilde}{C}=\emptyset$ or $Z^c\cap {\widetilde}{C}=\emptyset$, which implies that either $Z\subseteq C_{\rm exc}$ or $Z^c\subseteq C_{\rm exc}$, respectively. This shows that $L$ is stably balanced. Conversely, assume that $L$ is stably balanced. Clearly this implies that $L$ is strictly balanced. Assume, by contradiction, that ${\widetilde}{C}$ is not connected. Then we can find two proper disjoint subcurves $D_1$ and $D_2$ of $C$ that are not contained in $C_{\rm exc}$ and such that $E:=(D_1\cup D_2)^c$ is the union of $r\geq 1$ exceptional components of $C$. It is easily checked that $$\left\{\begin{aligned} & \deg_{D_1\cup E}(\omega_C)=\deg_{D_1}(\omega_C), \\ & k_{D_1\cup E}=k_{D_1}=r, \\ & \deg_{D_1\cup E} L=\deg_{D_1} L +r. \end{aligned}\tag{*} \right.$$ Applying the inequality to the subcurves $D_1$ and $D_1\cup E$, we get $$-\frac{r}{2}=-\frac{k_{D_1}}{2}\leq \deg_{D_1}L-d\frac{\deg_{D_1}(\omega_C)}{2g-1}=$$ $$=\deg_{D_1\cup E} L-r-d\frac{\deg_{D_1\cup E}(\omega_C)}{2g-2}\leq \frac{k_{D_1\cup E}}{2}-r=-\frac{r}{2}. $$ Therefore one of the inequalities is strict fo the subcurve $D_1$ and this contradicts the fact that $L$ is strictly balanced since $\emptyset \not\neq D_1\not\subseteq C_{\rm exc}$ by construction. Let $\pdb$ be the category whose objects are families of quasistable curves ${\mathcal{C}}\to S$ endowed with a line bundle ${\mathcal{L}}$ of relative degree $d$ whose restriction to each geometric fiber is balanced and whose arrows are Cartesian diagrams of such families. Cleary $\pdb$ is a category fibered in groupoids over the category of schemes. The following theorem summarizes some of the properties of ${\overline{\mathcal Pic}_{d,g}}$ and of its good moduli space ${\overline{P}_{d,g}}$ known thanks to Caporaso and Melo (note that our stacks ${{\mathcal Pic}_{d,g}}$ and ${\overline{\mathcal Pic}_{d,g}}$ are called $\mathcal{G}_{d,g}$ and $\overline{\mathcal G}_{d,g}$ in [@Melo]). *([@Cap], [@Capbis], [@Melo])* 1. ${\overline{\mathcal Pic}_{d,g}}$ is an irreducible, smooth and universally closed Artin stack of finite type over ${\mathbb{C}}$ and of dimension $4g-4$. It contains the stack ${{\mathcal Pic}_{d,g}}$ as a dense open substack. 2. ${\overline{\mathcal Pic}_{d,g}}$ admits a good moduli space ${\overline{P}_{d,g}}$, that is a normal irreducible projective variety of dimension $4g-3$. The geometric points of ${\overline{P}_{d,g}}$ correspond bijectively to the equivalence classes of pairs $(C,L)$ where $C$ is a quasi-stable curve of genus $g$ and $L$ is a strictly balanced line bundle of degree $d$. 3. ${\overline}{P}_{d,g}$ is a coarse moduli scheme for $\pdb$ if and only if $(d+1-g, 2g-2)=1$. In this case ${\overline{P}_{d,g}}$ has only finite quotient singularities. The construction of the scheme ${\overline{P}_{d,g}}$ as a GIT-quotient is due to Caporaso (see [@Cap]); the construction of the stack ${\overline{\mathcal Pic}_{d,g}}$ is due to Caporaso (see [@Capbis]) in the case $(d+1-g,2g-2)=1$ and to Melo (see [@Melo]) in the general case. Note that we have a natural commutative diagram compactifying the diagram : $$\label{diag-maps} \xymatrix{ {\overline{\mathcal Pic}_{d,g}}\ar[r] \ar@{->>}_{\Phi_d}[d] &{\overline{P}_{d,g}}\ar@{->>}^{\phi_d}[d]\\ {\overline{\mathcal{M}}_g}\ar[r] & {\overline{M}_g}. }$$ \[conv-points\] &gt;From now on, for the ease of notation, whenever we write $(C, L)\in {\overline{P}_{d,g}}$ we mean that $L$ is a strictly balanced line bundle on the quasi-stable curve $C$, considered up to the equivalence relation of Definition \[equiv-rel\]. Next we introduce an open subset of ${\overline{P}_{d,g}}$ that will play a special role in the sequel. We denote by ${{\overline}{P}_{d,g}^{\rm st}}$ the open subset of ${\overline{P}_{d,g}}$ consisting of pairs $(C,L)\in {\overline{P}_{d,g}}$ where $L$ is stably balanced. By the above Remark \[GIT-inter\], ${{\overline}{P}_{d,g}^{\rm st}}$ is the open subset of ${\overline{P}_{d,g}}$ where the GIT quotient is geometric. In [@Cap Lemma 2.2], it is proved that the semistable locus (called $H_d$ in loc. cit.) inside the Hilbert scheme whose GIT quotient gives ${\overline{P}_{d,g}}$ is smooth. From this, it follows that ${{\overline}{P}_{d,g}^{\rm st}}$ has finite quotient singularities (see (\[localring2\]) for an explicit local description). Moreover, ${{\overline}{P}_{d,g}^{\rm st}}={\overline{P}_{d,g}}$ if and only if $(d+1-g,2g-2)=1$ by [@Cap Prop. 6.2]. Albeit ${\overline}{P}_{d,g}$ has not necessarily finite quotient singularities, we have the following useful result (see the proof of [@Fon Cor. 1]): *([@Fon])* ${\overline}{P}_{d,g}$ is a ${\mathbb{Q}}$-factorial variety. In view of the above result, we will identify throughout this paper ${\mathbb{Q}}$-Weil divisors and ${\mathbb{Q}}$-Cartier divisors on ${\overline}{P}_{d,g}$. [*The automorphism group ${\mathrm{Aut}}(C,L)$*]{}\[Sec-auto\] For later use, we describe the automorphism group of a pair $(C,L)$ consisting of a quasi-stable curve $C$ and a balanced line bundle $L$ on $C$. An automorphism of $(C, L)$ is given by a pair $(\sigma, \psi)$ such that $\sigma\in {\mathrm{Aut}}(C)$ and $\psi$ is an isomorphism between the line bundles $L$ and $\sigma^*(L)$. The group of automorphisms of $(C,L)$ is denoted by ${\mathrm{Aut}}(C,L)$. We get a natural forgetful homomorphism $$\label{homo-auto} \begin{aligned} F: {\mathrm{Aut}}(C,L)&\to {\mathrm{Aut}}(C)\\ (\sigma,\psi)& \mapsto \sigma\\ \end{aligned}$$ whose kernel is the multiplicative group ${\mathbb{G}_m}$, acting as fiberwise multiplication on $L$, and whose image is the subgroup of $\sigma\in {\mathrm{Aut}}(C)$ such that $\sigma^*(L)\cong L$. The quotient ${\mathrm{Aut}}(C,L)/{\mathbb{G}_m}$ is denoted by ${\overline}{{\mathrm{Aut}}(C,L)}$ and is called the [*reduced*]{} automorphism group of $(C, L)$. Note that ${\mathrm{Aut}}(C,L)$ depends only on the equivalence class $[(L,C)]$ (see Def. \[equiv-rel\]). By composing the above homomorphism $F$ of with the natural homomorphism ${\mathrm{Aut}}(C)\to {\mathrm{Aut}}(C^{\rm st})$ induced by the stabilization map $C\to C^{\rm st}$, we get a homomorphism $$G: {\mathrm{Aut}}(C,L)\to {\mathrm{Aut}}(C^{\rm st}),$$ whose kernel is described in the next Lemma. \[automo\] We have a commutative diagram with exact rows $$\xymatrix{ 0 \ar[r] & {\mathbb{G}_m}^{\gamma({\widetilde}{C})} \ar[r]\ar@{->>}[d]& {\mathrm{Aut}}(C,L) \ar[r]^{G}\ar@{->>}[d]& {\mathrm{Aut}}(C^{\rm st})\ar@{=}[d]\\ 0 \ar[r] & {\mathbb{G}_m}^{\gamma({\widetilde}{C})}/{\mathbb{G}_m}\ar[r] & {\overline}{{\mathrm{Aut}}(C,L)} \ar[r]^{{\overline}{G}} & {\mathrm{Aut}}(C^{\rm st}),\\ }$$ where ${\mathbb{G}_m}\subseteq {\mathbb{G}_m}^{\gamma({\widetilde}{C})}$ is the diagonal embedding. The exactness of the first row is proved using an argument similar to the one used in the proof of [@CCC Lemma 2.3.2]. We sketch the argument for the sake of completeness. Let $E_1,\ldots,E_m$ be the exceptional components of $C$ and let $X_1,\ldots,X_{\gamma({\widetilde}{C})}$ be the connected components of ${\widetilde}{C}$. We identify each $E_i$ to a copy of ${\mathbb{P}}^1$ attached to the rest of the curve at the points $0$ and $\infty$. An element $(\sigma, \psi)\in {\mathrm{Aut}}(C,L)$ belongs to the kernel of the map ${\mathrm{Aut}}(C,L)\to {\mathrm{Aut}}(C^{\rm st})$ if and only if $\sigma_{|{\widetilde}{C}}={\mathrm{id}}_{{\widetilde}{C}}$ and $\sigma$ acts as multiplication by $m_i\in {\mathbb{G}_m}(k)=k^*$ on the exceptional component $E_i$. If we restrict the isomorphism $\psi:L\stackrel{\cong}{\to} \sigma^*(L)$ to ${\widetilde}{C}$, we get that $\psi$ is the fiberwise multiplication by $l_j\in {\mathbb{G}_m}(k)=k^*$ on each line bundle $L_{|X_j}$. The scalars $m_i$ are uniquely determined by the scalars $l_j$: if $0\in E_i$ lies on the component $X_j$ and $\infty \in E_i$ lies on the component $X_h$ (possibly with $j=h$), then by the compatibility between $\sigma$ and $\psi$ we get that $m_i=l_j/l_h$ (see the proof of [@CCC Lemma 2.3.2]). Therefore the element $(\sigma,\psi)$ is uniquely determined by the scalars $l_1,\ldots,l_{\gamma({\widetilde}{C})}$ and we are done. From the above proof, it is clear that the homomorphisms corresponding to the diagonal embedding ${\mathbb{G}_m}\hookrightarrow {\mathbb{G}_m}^{\gamma({\widetilde}{C})}$ are exactly the fiberwise automorphisms on $L$, hence the exactness of the second row follows. \[Cor-auto\] If $(C,L)$ is stably balanced then ${\overline}{{\mathrm{Aut}}(C,L)}$ is a subgroup of ${\mathrm{Aut}}(C^{\rm st})$. In particular, ${\overline}{{\mathrm{Aut}}(C, L)}$ is a finite group. The first assertion follows from the last row of the diagram in Lemma \[automo\] together with the fact that if $L$ is stably balanced on $C$ then ${\widetilde}{C}$ must be connected by Lemma \[compa-bal\]. The last assertion follows from the first one together with the well-known fact that the automorphism group of a stable curve is finite. [*The local structure of ${\overline{P}_{d,g}}$*]{}\[loc-struc\] The complete local ring $\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}$ of ${\overline{P}_{d,g}}$ at a point $(C,L)$ can be described using the deformation theory of pairs $(C,L)$, which can be found in [@ser Sec. 3.3.3] for $C$ smooth and has been extended to the singular case in [@wang]. Let us recall the results of [@wang]. We denote by ${\mathcal{P}}^1_C(L)$ the sheaf of one jets (or sheaf of principle parts) of $L$ on $C$ ([@wang Sec. 2]). The sheaf ${\mathcal{P}}^1_C(L)$ fits into an exact sequence (see [@wang Eq. (2.1)]) $$\label{jets} 0 \to \Omega^1_C \to {\mathcal{P}}^1_C(L)\otimes L^{-1} \to {\mathcal{O}}_C \to 0.$$ Explicitly, the above extension can be described as follows (see [@ser p. 145]). Let ${\mathcal{O}}_C^*\rightarrow \Omega^1_C$ be the homomorphism of sheaves given by sending $u\in \Gamma(U, {\mathcal O}_C^*)$ into $\displaystyle \frac{d u}{u}\in \Gamma(U,\Omega_C^1)$ for any open subset $U\subseteq C$. By passing to cohomology, we get a group homomorphism $\theta_C:{\mathrm{Pic}}(C)=H^1(C,{\mathcal{O}}_C^*) \rightarrow H^1(C, \Omega_C^1)$. By using the identification $H^1(C, \Omega_C^1) \cong {\mathrm{Ext}}^1({\mathcal O}_C, \Omega_C^1)$, the map $\theta_C$ sends the line bundle $L$ to the class of the extension . Wang in [@wang] proves that the sheaf ${\mathcal{P}}^1_C(L)$ controls the tangent and obstruction theory of the pair $(C,L)$. Let us denote by ${\mathrm{Def}}_{(C,L)}$ the functor of infinitesimal deformations of the pair $(C,L)$ (see [@ser p. 146]) and by $T_{{\mathrm{Def}}_{(C,L)}}$ the tangent space to ${\mathrm{Def}}_{(C,L)}$ (in the sense of [@ser Lemma 2.2.1]). *([@wang])* \[wang-thm\] (i) \[wang-thm1\] We have that $T_{{\mathrm{Def}}_{C,L)}}={\mathrm{Ext}}^1({\mathcal{P}}^1_C(L),L)$. (ii) \[wang-thm2\] An obstruction space for ${\mathrm{Def}}_{(C,L)}$ is given by ${\mathrm{Ext}}^2({\mathcal{P}}^1_C(L),L)$. Part (i) is [@wang Thm. 3.1(1)]; Part (ii) is [@wang Thm. 4.6(a)]. Moreover, the infinitesimal automorphisms of the pair $(C,L)$ are governed by ${\mathrm{Ext}}^0({\mathcal{P}}_C^1(L),L)$, as shown by the following Lemma. \[inf-aut\] The tangent space of ${\mathrm{Aut}}(C,L)$ at the identity is equal to ${\mathrm{Ext}}^0({\mathcal{P}}^1_C(L),L)$. The Lemma is certainly well-known to the experts, at least in the case where $C$ is smooth. However, we include a proof for the lack of a suitable reference. According to the discussion in \[Sec-auto\], we have an exact sequence of groups $$\label{seq-aut} 0 \to {\mathbb{G}_m}\to {\mathrm{Aut}}(C,L)\to {\rm Stab}_L({\mathrm{Aut}}(C))\to 0,$$ where ${\rm Stab}_L({\mathrm{Aut}}(C))$ is the stabilizer of $L$ in ${\mathrm{Aut}}(C)$, i.e. the subgroup of ${\mathrm{Aut}}(C)$ consisting of all the automorphisms $\sigma$ of $C$ such that $\sigma^*(L)\cong L$. In other words, ${\rm Stab}_L({\mathrm{Aut}}(C))$ is the image of ${\mathrm{Aut}}(C,L)$ via the map $F$ of . By passing to the tangent spaces at the origin, we get $$\label{sequ1} 0 \to T_0{\mathbb{G}_m}=k \to T_0{\mathrm{Aut}}(C,L)\to T_0{\rm Stab}_L({\mathrm{Aut}}(C))\to 0,$$ where we have denoted by $0$ the identity element in each of the above groups. On the other hand, by dualizing , we get the short exact sequence $$\label{dual-seq} 0\to {\mathcal{O}}_C \to {\mathcal{P}}^1(L)^{\vee}\otimes L\to T_C\to 0.$$ Passing to cohomology, we get the exact sequence $$\label{sequ2} 0\to H^0(C, {\mathcal{O}}_C)=k \to H^0(C, {\mathcal{P}}^1(L)^{\vee}\otimes L)\stackrel{p}{\to} H^0(C, T_C) $$ Compare now the exact sequences and . Since $T_0{\mathrm{Aut}}(C)=H^0(C, T_C)$ by [@ser Prop. 2.6.2] and clearly ${\mathrm{Ext}}^0({\mathcal{P}}^1(L),L)=H^0(C,$ $ {\mathcal{P}}^1(L)^{\vee}\otimes L)$, it is enough to show that $$\label{sequ3} T_0{\rm Stab}_L({\mathrm{Aut}}(C))={\rm Im}(p).$$ Let ${\mathcal U}=\{U_{\alpha}\}$ be an affine open covering of $C$ trivializing $L$ and let $f_{\alpha \beta} \in \Gamma(U_{\alpha \beta}, {\mathcal O}_C^*)$ the transition functions of $L$ with respect to ${\mathcal U}$, where as usual $U_{\alpha\beta}:=U_{\alpha}\cap U_{\beta}$. Then $\theta_C(L)\in H^1(C,\Omega_C^1)$ is represented by the Čech $1$-cocycle $\left(\frac{df_{\alpha \beta}}{f_{\alpha \beta}}\right)\in Z^1({\mathcal U}, \Omega_C^1)$ (see [@ser p. 145]). &gt;From the exact sequence , it follows that the sheaf $({\mathcal{P}}^1(L)^{\vee}\otimes L)_{|U_{\alpha}}$ is isomorphic to $({\mathcal{O}}_C)_{U_{\alpha}}\oplus (T_C)_{|U_{\alpha}}$ and an element of ${\mathrm{Ext}}^0({\mathcal{P}}^1(L), L)=H^0(C, {\mathcal{P}}^1(L)^{\vee}\otimes L)$ is represented by a Čech $0$-cochain $$(k_{\alpha},d_{\alpha}) \in C^0({\mathcal U}, {\mathcal{P}}^1(L)^{\vee}\otimes L)=C^0({\mathcal U}, {\mathcal{O}}_C)\oplus C^0({\mathcal U}, T_C),$$ which satisfies the cocycle conditions: $d_{\alpha}=d_{\beta}$ and $k_{\beta}-k_{\alpha}=\frac{d_{\alpha}(f_{\alpha\beta})}{f_{\alpha\beta}}$ on $U_{\alpha\beta}=U_{\alpha}\cap U_{\beta}$ (see [@ser p. 145]). Since the $f_{\alpha\beta}$’s are the transition functions of $L$, we conclude that the image of $p$ consists of all the Čech $0$-cocycles $(d_{\alpha})\in Z^0({\mathcal U}, T_C)$ corresponding to the infinitesimal automorphisms of $C$ which preserve the line bundle $L$. In other words, is satisfied and we are done. We can now compute the dimension of the vector spaces ${\mathrm{Ext}}^i({\mathcal{P}}^1_C(L),L)$. \[ext-groups\] We have that $$\left\{\begin{aligned} &\dim {\mathrm{Ext}}^i({\mathcal{P}}^1_C(L),L)=0 \text{ for } i\geq 2, \\ &\dim {\mathrm{Ext}}^1({\mathcal{P}}^1_C(L), L)=4g-4+\gamma({\widetilde}{C}), \\ &\dim {\mathrm{Ext}}^0({\mathcal{P}}^1_C(L),L)=\gamma({\widetilde}{C}). \end{aligned}\right.$$ By applying the functor ${\mathrm{Hom}}(-,{\mathcal{O}}_C)$ to the exact sequence and using that ${\mathrm{Ext}}^{\geq 2}({\mathcal{O}}_C,{\mathcal{O}}_C)=H^{\geq 2}(C,{\mathcal{O}}_C)=0$ since $C$ is a curve and that ${\mathrm{Ext}}^{\geq 2}(\Omega_C^1, {\mathcal{O}}_C)=0$ by [@DM Lemma 1.3], we get the vanishing ${\mathrm{Ext}}^{\geq 2}({\mathcal{P}}^1_C(L),L)=0$. The fact that ${\mathrm{Ext}}^0({\mathcal{P}}_C^1(L),L)=\gamma({\widetilde}{C})$ follows from Lemma \[automo\] and \[inf-aut\]. Finally, from the exact sequence (\[jets\]), we get $$\dim {\mathrm{Ext}}^0({\mathcal{P}}^1_C(L),L)- \dim {\mathrm{Ext}}^1({\mathcal{P}}^1_C(L),L)=$$ $$\dim {\mathrm{Ext}}^0({\mathcal{O}}_C,{\mathcal{O}}_C)-\dim {\mathrm{Ext}}^1({\mathcal{O}}_C,{\mathcal{O}}_C)+$$ $$\dim {\mathrm{Ext}}^0(\Omega_C^1,{\mathcal{O}}_C)-\dim {\mathrm{Ext}}^1(\Omega_C^1,{\mathcal{O}}_C)=$$ $$-(g-1)-(3g-3)=-(4g-4),$$ from which we conclude. We can now prove that the functor ${\mathrm{Def}}_{(C,L)}$ has a semiuniversal formal element (in the sense of [@ser Def. 2.2.6]). \[semiuniv\] (i) \[semiuniv1\]�The functor ${\mathrm{Def}}_{(C,L)}$ has a semiuniversal formal element ${\underline}{{\mathrm{Def}}}_{(C,L)}$. (ii) \[semiuniv2\]�${\underline}{{\mathrm{Def}}}_{(C,L)}$ is equal to the formal spectrum of $k[[x_1,\ldots, x_{4g-4+\gamma({\widetilde}{C})}]]$. Part is proved in [@ser Thm 3.3.11(i)] in the case where $C$ is smooth. The proof of loc. cit. consists in showing that Schlessinger’s conditions are satisfied and this extends to the case where $C$ is nodal: the crucial point of the proof is showing that $T_{{\mathrm{Def}}_{(C,L)}}$ is finite dimensional, and this follows in our case from Theorem \[wang-thm\]. &gt;From Theorem \[wang-thm\] and Lemma \[ext-groups\], it follows that ${\underline}{{\mathrm{Def}}}_{(C,L)}$ is formally smooth and that the dimension of the tangent space at its unique closed point is $4g-4+\gamma({\widetilde}{C})$, from which part follows. Now we can describe the complete local ring $\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}$ of ${\overline{P}_{d,g}}$ at a point $(C,L)$. Note that the automorphism group ${\mathrm{Aut}}(C,L)$ acts on ${\underline}{{\mathrm{Def}}}_{(C,L)}$ (hence on ${\mathbb{C}}[[x_1,\ldots,x_{4g-4+\gamma({\widetilde}{C})}]]$) by the semiuniversality of ${\underline}{{\mathrm{Def}}}_{(C,L)}$. By a standard argument based on Luna’s étale slice theorem (see [@Lun p. 97] and also [@Las Sec. II], [@Dre Sec. 7.4]), the formal spectrum (which we denote by ${\mathrm{Spf}}$) of the complete local ring $\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}$ of ${\overline{P}_{d,g}}$ at the point $(C,L)\in {\overline{P}_{d,g}}$ is given by $$\label{localring} {\mathrm{Spf}}\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}={\underline}{{\mathrm{Def}}}_{(C,L)}/{\mathrm{Aut}}(C,L),$$ where ${\underline}{{\mathrm{Def}}}_{(C,L)}/{\mathrm{Aut}}(C,L)$ is the quotient of ${\underline}{{\mathrm{Def}}}_{(C,L)}$ with respect to the natural action of ${\mathrm{Aut}}(C,L)$. In other words, ${\underline}{{\mathrm{Def}}}_{(C,L)}/{\mathrm{Aut}}(C,L)$ is equal to the formal spectrum of the ring of invariants ${\mathbb{C}}[[x_1,\ldots,$ $ x_{4g-4+\gamma({\widetilde}{C})}]]^{{\mathrm{Aut}}(C,L)}$ (see Proposition \[semiuniv\]). Clearly, the scalar automorphisms ${\mathbb{G}_m}\subseteq {\mathrm{Aut}}(C,L)$ act trivially on ${\underline}{{\mathrm{Def}}}_{(C,L)}$ and thus we get the alternative description: $$\label{localring2} {\mathrm{Spf}}\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}={\underline}{{\mathrm{Def}}}_{(C,L)}/{\overline}{{\mathrm{Aut}}(C,L)}.$$ Note that, from the above description and Lemma \[automo\], it follows that ${{\overline}{P}_{d,g}^{\rm st}}$ is the open subset of ${\overline{P}_{d,g}}$ consisting of pairs $(C,L)\in {\overline{P}_{d,g}}$ such that ${\widetilde}{C}$ is connected. We can similarly describe the morphism $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ locally at $(C,L)\in {\overline{P}_{d,g}}$. Denote by ${\underline}{{\mathrm{Def}}}_C$ (resp. ${\underline}{{\mathrm{Def}}}_{C^{\rm st}}$) the semiuniversal formal element associated to the infinitesimal deformation functor ${\mathrm{Def}}_C$ (resp. ${\mathrm{Def}}_{C^{\rm st}}$) of $C$ (resp. $C^{\rm st}$), see [@ser Cor. 2.4.2]. Locally at $(C,L)\in {\overline{P}_{d,g}}$, the morphism $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ is given by $$\label{mor-loc} {\mathrm{Spf}}\widehat{{\mathcal{O}}}_{{\overline{P}_{d,g}},(C,L)}={\underline}{{\mathrm{Def}}}_{(C,L)}/{\overline}{{\mathrm{Aut}}(C,L)}\to {\mathrm{Spf}}\widehat{{\mathcal{O}}}_{{\overline{M}_g},C^{\rm st}}={\underline}{{\mathrm{Def}}}_{C^{\rm st}}/{\mathrm{Aut}}(C^{\rm st}),$$ where the homomorphism of groups ${\overline}{{\mathrm{Aut}}(C,L)}\to {\mathrm{Aut}}(C^{\rm st})$ is the one given by Lemma \[automo\] and the morphism ${\underline}{{\mathrm{Def}}}_{(C,L)}\to {\underline}{{\mathrm{Def}}}_{C^{\rm st}}$ is the composition of the forgetful morphism ${\underline}{{\mathrm{Def}}}_{(C,L)}\to {\underline}{{\mathrm{Def}}}_C$ with the stabilization morphism ${\underline}{{\mathrm{Def}}}_C\to {\underline}{{\mathrm{Def}}}_{C^{\rm st}}$. The induced morphism at the level of tangent spaces $$\label{mor-tang} T_{{\mathrm{Def}}_{(C,L)}}={\mathrm{Ext}}^1({\mathcal{P}}^1_C(L), L)\to T_{{\mathrm{Def}}_{C^{\rm st}}}={\mathrm{Ext}}^1(\Omega^1_{C^{\rm st}}, {\mathcal{O}}_{C^{\rm st}})$$ is given by composing the morphism ${\mathrm{Ext}}^1({\mathcal{P}}^1_C(L), L)\to {\mathrm{Ext}}^1(\Omega_{C}^1\otimes L, L)={\mathrm{Ext}}^1(\Omega_{C}^1, {\mathcal{O}}_C)$ induced by the exact sequence (\[jets\]) with the morphism ${\mathrm{Ext}}^1(\Omega_{C}^1, {\mathcal{O}}_C)\to {\mathrm{Ext}}^1(\Omega_{C^{\rm st}}, {\mathcal{O}}_{C^{\rm st}})$ induced by the stabilization map $C\to C^{\rm st}$. More precisely, let $f:C \rightarrow C^{\rm st}$ be the stabilization morphism and denote by $Lf^*$ (resp. $Rf_*$) the left derived functor of $f^*$ (resp. the right derived functor of $f_*$). We have a natural map $$\label{ext} {\mathrm{Ext}}^1(\Omega_{C}^1,{\mathcal O}_{C}) \to {\mathrm{Ext}}^1(Lf^*\Omega^1_{C^{\rm st}},{\mathcal O}_C)= {\mathrm{Ext}}^1(\Omega^1_{C^{\rm st}},{\mathcal O}_{C^{\rm st}}).$$ The first map in is induced by the composite map $Lf^*\Omega^1_{C^{\rm st}}\to L^0f^*\Omega^1_{C^{\rm st}}= f^*\Omega^1_{C^{\rm st}}\to \Omega^1_C$ in the derived category of coherent sheaves on $C$. The equality in follows from the adjointness of the functors $Lf^*$ and $Rf_*$ between the derived category of coherent sheaves on $C$ and on $C^{\rm st}$, together with the fact that $Rf_*{\mathcal O}_C\cong {\mathcal O}_{C^{\rm st}}$ because $f$ is a sequence of blow-ups with projective spaces as fibers. The fibration $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ {#Fibrat} =============================================================== The aim of this section is to prove Proposition \[kod-fib\] below, which gives the second part of Theorem \[kod-nongeo\]. To this aim, we analize the natural morphism $\phi_d: {\overline{P}_{d,g}}\to {\overline{M}_g}$. Note that $\phi_d$ is a regular fibration (i.e. a proper, surjective morphism with connected fibers), whose general fiber is the degree-$d$ Jacobian ${\mathrm{Pic}}^d(C)$ of a general $[C] \in {\overline{M}_g}$. Following Kawamata (see [@kaw Sec. 1] and [@kaw Cor. 7.3]), we define the variation ${\rm Var}(\phi_d)$ of $\phi_d$ to be $$\label{def-var} {\rm Var}(\phi_d)=\dim {\overline{M}_g}- \dim {\mathrm{Ker}}(\delta_C),$$ where $$\label{Kod-Spe-map} \delta_C:T_{{\overline{M}_g},C}=H^1(C,T_C)\to H^1({\mathrm{Pic}}^d(C),T_{{\mathrm{Pic}}^d(C)})$$ is the Kodaira-Spencer map associated to $\phi_d$ at a general point $C\in {\overline{M}_g}$. \[maxvar\] The algebraic fiber space $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ has maximal variation, i.e. ${\rm Var}(\phi_d)=3g-3$. By , we have to prove the injectivity of the Kodaira-Spencer map $\delta_C$ for a general curve $C\in {\overline{M}_g}$. We will reinterpret the above Kodaira-Spencer map as composition of certain maps that were studied in [@OS Sec. 2], in their analysis of the local Torelli problems for curves. We need to recall their setting, with the simplification that, since we are only interested in the general curve, we can work directly with the coarse moduli spaces $M_g$ and $A_g$ (where, as usual, $A_g$ denotes the coarse moduli space of principally polarized abelian varieties of dimension $g$), without having to pass to their $n$-level covers. Consider the following commutative diagram (see [@OS p. 169]): $$\xymatrix@=.7pc{ T_{M_g,C}\ar^(.4){k_C}_(.4){\cong}[r] \ar_{d t_g}[dd]& H^1(C, T_C) \ar^{d u}[d]\\ & H^1(C, u^*T_{{\mathrm{Pic}}^d(C)}) \\ T_{A_g,{\mathrm{Pic}}^d(C)}\ar^{k_{{\mathrm{Pic}}^d(C)}}[dr] & \\ & H^1({\mathrm{Pic}}^d(C),T_{{\mathrm{Pic}}^d(C)})\ar^{\cong}_{u^*}[uu]. }$$ where $u:C\to {\mathrm{Pic}}^d(C)$ is an Abel-Jacobi map (well-defined only up to translation), $t_g:M_g\to A_g$ is the classical Torelli map, $k_C$ is the Kodaira-Spencer map in $C$ associated to the universal family over an open subset of $M_g$ containing $C$ and $k_{{\mathrm{Pic}}^d(C)}$ is the Kodaira-Spencer map in ${\mathrm{Pic}}^d(C)$ associated to the universal family over an open subset of $A_g$ containing ${\mathrm{Pic}}^d(C)$. The map $k_C$ is an isomorphism (see e. g. [@OS Thm. 2.2]) and the map $u^*$ is an isomorphism since $T_{{\mathrm{Pic}}^d(C)}$ is the trivial bundle of rank $g$ and $$u^*: H^1({\mathrm{Pic}}^d(C), {\mathcal{O}}_{{\mathrm{Pic}}^d(C)}) \stackrel{\cong}{\longrightarrow} H^1(C, {\mathcal{O}}_C).$$ It is easy to see that $$\delta_C=k_{{\mathrm{Pic}}^d(C)}\circ d t_g=(u^*)^{-1}\circ du \circ k_C.$$ Therefore, the injectivity of $\delta_C$ is equivalent to the injectivity of $du$. According to [@OS Thm. 2.6], the map $du$ is the dual of the multiplication map $$\mu: H^0(C, \omega_C)\otimes H^0(C, \omega_C)\to H^0(C, \omega_C^{\otimes 2}),$$ which is well-known (Noether’s theorem) to be surjective if $g=2$ or if $g\geq 3$ and $C$ is not hyperelliptic. Since we assumed $C$ to be generic, we deduce the injectivity of $\delta_C$ and we are done. \[kod-fib\] 1. We have that $\kappa({\overline{P}_{d,g}}) \le 3g-3$. 2. If $\kappa({\overline{M}_g})\geq 0$, then $\kappa({\overline{P}_{d,g}})= 3g-3$ and the map $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ is the Iitaka fibration of ${\overline{P}_{d,g}}$. The subaddivity of the Kodaira dimension (see [@Ueno Thm. 6.12]) applied to the regular fibration $\phi_d$ gives that $$\kappa({\overline{P}_{d,g}})\leq \dim {\overline{M}_g}+\kappa(\phi_d^{-1}(C)),$$ for a general $C\in {\overline{M}_g}$. Since, for a general $C\in {\overline{M}_g}$, the fiber $\phi_d^{-1}(C)={\mathrm{Pic}}^d(C)$ is an abelian variety, we have that $\kappa(\phi_d^{-1}(C))=0$, which proves part (1). Assume now that $\kappa({\overline{M}_g})\geq 0$. Observe that ${\mathrm{Pic}}^d(C)$ is a good minimal model, since it is smooth and the canonical $K_{{\mathrm{Pic}}^d(C)}$ is trivial and thus clearly semi-ample. Therefore, the Iitaka conjecture (in the stronger form of [@kaw p. 1]) does hold true by [@kaw Cor. 1.2] and gives that $$\kappa({\overline{P}_{d,g}})\geq \kappa(\phi_d^{-1}(C))+{\rm max}\{\kappa({\overline{M}_g}),{\rm Var}(\phi_d)\}=3g-3,$$ using the above Lemma \[maxvar\]. This, combined with part (1), proves that $\kappa({\overline{P}_{d,g}})=3g-3$. The last part follows from the birational characterization of the Iitaka fibration (see e. g. [@Ueno Thm. 6.11]) since $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ is an algebraic fiber space such that $\dim {\overline{M}_g}=\kappa({\overline{P}_{d,g}})$ and the generic fiber $\phi_d^{-1}(C)={\mathrm{Pic}}^d(C)$ is smooth and irreducible of Kodaira dimension zero. The singularities of ${{\overline}{P}_{d,g}^{\rm st}}$ {#Sec-sing} ======================================================= The purpose of this section is to study the singularities of ${\overline{P}_{d,g}}$ with the aim of proving Theorem \[conj-sing\]. More generally, we will prove a similar statement (see Theorem \[sing-Pdst\]) for the open subvariety ${{\overline}{P}_{d,g}^{\rm st}}\subseteq {\overline{P}_{d,g}}$ of (\[def-stack-scheme\]) and for any degree $d$. Since ${{\overline}{P}_{d,g}^{\rm st}}={\overline{P}_{d,g}}$ if (and only if) $(d+1-g,2g-2)=1$ (by [@Cap Prop. 6.2]), Theorem \[conj-sing\] is a special case of Theorem \[sing-Pdst\] below. The need of restricting ourselves to the open subset ${{\overline}{P}_{d,g}^{\rm st}}$ is due to the fact that ${{\overline}{P}_{d,g}^{\rm st}}$ has finite quotient singularities and therefore we can apply the Reid–Tai criterion for the canonicity of finite quotient singularities (see e.g. [@HarMum pp. 27-28] or [@Lud Thm. 4.1.11]). For simplicity, we assume throughout this section that $g\geq 4$ in order to avoid problems with the hyperelliptic locus in ${\overline{M}_g}$. In the analysis of the singularities of ${{\overline}{P}_{d,g}^{\rm st}}$, the pairs $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ such that $C$ contains an elliptic tail will play a special role. Let us give some definitions. \[def-ell-tail\] A connected subcurve $E$ of a quasi-stable curve $C$ is called an *elliptic tail* if it has arithmetic genus $1$ and meets the rest of the curve in exactly one node $P$ which is called an elliptic tail node. The following remark is straightforward. \[rmk-ell-tail\] If a quasi-stable curve $C$ has an elliptic tail $E\subseteq C$ then the image $E^{\rm st}:=st(E)\subseteq C^{\rm st}$ of $E$ via the stabilization morphism $st: C\to C^{\rm st}$ is an elliptic tail of $C^{\rm st}$. Conversely, if $C^{\rm st}$ has an elliptic tail $E'\subseteq C^{\rm st}$, then $E:=st^{-1}(E')$ is an elliptic tail of $C$ such that $E'=E^{\rm st}$. In the next Lemma, we describe the pairs $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ such that $C$ has an elliptic tail $E\subseteq C$. \[tail-stable\] Let $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ such that $C$ has an elliptic tail $E\subseteq C$. Then we have: (i) \[el-tail1\] $d\not\equiv g-1 \mod (2g-2)$ and the degree $\deg_E(L)$ of $L$ on $E$ is the unique integer $d_E$ such that $$\label{deg-tail} -\frac{1}{2}< d_E-\frac{d}{2g-2}< \frac{1}{2}.$$ (ii) \[el-tail2\] The elliptic tail node $P$ does not belong to $C_{\rm exc}$. (iii) \[el-tail3\] $E$ is either smooth or it is a rational (irreducible) curve with one node $Q$ or it is formed by two smooth rational curves $R_1$ and $R_2$ meeting in two points $Q_1$ and $Q_2$, as depicted in Figure \[Fig-tails\]. The equation follows from the basic inequality applied to the subcurve $E\subseteq C$ together with the fact that the inequalities must be strict since clearly $E\not\subseteq C_{\rm ex}$ and $L$ is stably balanced by the hypothesis that $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$. The fact that $d\not\equiv g-1 \mod (2g-2)$ follows from the fact that the there exists an integer $d_E$ satisfying the strict inequalities in . This prove part . Next we turn to Part . By contradiction, if $P\in R$ where $R$ is an exceptional component of $C$, then $R^c$ is a disjoint union of two subcurves of $C$ each of which contains some components of ${\widetilde}{C}$. Therefore ${\widetilde}{C}$ is disconnected and this contradicts the hypothesis that $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ by Lemma \[compa-bal\]. Part follows from the fact that $C$ is quasi-stable together with Part and the fact that $E^{\rm st}$ is either a smooth elliptic curve (which occurs for Type I) or a rational irreducible curve with one node $Q$ (which occurs for Types II and III). The elements $(C,L)\in {\overline{P}_{d,g}}$ such that $C$ has an elliptic tail $E$ have special automorphisms that will play a key role in the sequel. \[def-auto-tail\] Given an element $(C,L)\in {\overline{P}_{d,g}}$ such that $C$ has an elliptic tail $E$, an automorphism $\phi=(\sigma,\psi)\in {\mathrm{Aut}}(C,L)$ (or its image in ${\overline}{{\mathrm{Aut}}(C,L)}$) is called an *elliptic tail automorphism* of $(C,L)$ of order $n\geq 1$ (with respect to the elliptic tail $E\subseteq C$) if $\sigma$ is the identity on ${\overline}{C\setminus E}$ and $\sigma_{|E}$ has order $n$. The assumption that $(C,L)$ belongs to ${{\overline}{P}_{d,g}^{\rm st}}$ puts some constraints on the possible elliptic tail automorphisms that can occur. Indeed, under this assumption, using that the map ${\overline}{G}:{\overline}{{\mathrm{Aut}}(C,L)}\to {\mathrm{Aut}}(C^{\rm st})$ is injective (see Corollary \[Cor-auto\]), we deduce immediately the following \[rmk-aut-tail\] Given $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$, an element $\phi\in {\mathrm{Aut}}(C,L)$ is an elliptic tail automorphism of order $n\geq 1$ with respect to the elliptic tail $E\subseteq C$ if and only if $G(\phi)\in {\mathrm{Aut}}(C^{\rm st})$ (see the notation of Lemma \[automo\]) is an elliptic tail automorphism of order $n$ of $C^{\rm st}$ with respect to the elliptic tail $E^{\rm st}\subseteq C^{\rm st} $, i.e. $G(\phi)$ is the identity on ${\overline}{C^{\rm st}\setminus E^{\rm st}}$ and $G(\phi)_{|E^{\rm st}}$ has order $n$. Using this Remark, we can give a complete description of the possible elliptic tail automorphisms of elements $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$. \[auto-ell\] Assume that $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ and that $C$ has an elliptic tail $E\subseteq C$. For an elliptic tail automorphism $\phi=(\sigma,\psi)\in {\mathrm{Aut}}{(C,L)}$ of $(C,L)$ of order $n>1$ with respect to the elliptic tail $E\subseteq C$, the restriction $\sigma_{|E}$ of $\sigma$ to $E$ must satisfy the following conditions (according to whether the elliptic tail $E$ is of Type I, II or III as in Lemma \[tail-stable\]): (i) \[auto-ell1\] Type I: $\sigma_{|E}$ is an automorphism of $E$ fixing $P$ and $n=2$ or $n=4$ (which can occur if and only if $E$ has $j$-invariant equal to $1728$) or $n=3, 6$ (which can occur if and only if $E$ has $j$-invariant equal to $0$). (ii) \[auto-ell2\] Type II: $\sigma_{|E}$ is an automorphism of order $n=2$ fixing $P$ and $Q$. If we call $\nu: E^{\nu}\to E$ the normalization map and identify $E^{\nu}$ with ${\mathbb{P}}^1$ in such a way that $\nu^{-1}(P)=\infty$ and $\nu^{-1}(Q)=\{1, -1\}$, then the automorphism $\sigma_{|E}$ is induced by the automorphism $x\mapsto -x$ on ${\mathbb{P}}^1$. (iii) \[auto-ell3\] Type III: $\sigma_{|E}$ is an automorphism of order $n=2$ such that, if we identify $R_i$ (for $i=1,2$) with ${\mathbb{P}}^1$ in such a way that $Q_1$ and $Q_2$ get identified with $1$ and $-1$ (on both copies of ${\mathbb{P}}^1$) and $P\in R_1$ gets identified with $\infty$, then $\sigma_{|R_i}$ (for $i=1, 2$) is equal to the automorphism $x\mapsto -x$ on ${\mathbb{P}}^1$. In particular, $\sigma_{|E}$ fixes $P$ and exchanges $Q_1$ with $Q_2$. Parts and follow easily from Remark \[rmk-aut-tail\] together with the fact that $E^{\rm st}\cong E$ for Type I and II and the well-known description of the elliptic tail automorphisms of stable curves (see e.g. [@Lud Rmk. 4.2.2]). In order to prove Part , observe that in this case $E^{\rm st}\subseteq C^{\rm st}$ is a rational curve with one node. Therefore, there exists a unique elliptic tail automorphism of $C^{\rm st}$ with respect to $E^{\rm st}$, namely the automorphism $\sigma$ whose restriction $\sigma_{|E}$ is described in Part . We conclude by Remark \[rmk-aut-tail\] together with the fact the elliptic tail automorphism of $(C,L)$ described in Part is the unique (by Corollary \[Cor-auto\]) lift to ${\overline}{{\mathrm{Aut}}(C,L)}$ of the elliptic tail automorphism of $C^{\rm st}$ with respect to $E^{\rm st}$ described in Part . We can now determine the singular locus of ${{\overline}{P}_{d,g}^{\rm st}}$. \[sing\] The singular locus of ${{\overline}{P}_{d,g}^{\rm st}}$ (for $g\geq 4$) is exactly the locus of pairs $(C,L)$ such that ${\overline}{{\mathrm{Aut}}(C,L)}$ is not trivial. Near a point $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$, using the local description (\[localring2\]) and Corollary \[Cor-auto\], the scheme ${{\overline}{P}_{d,g}^{\rm st}}$ is isomorphic to the finite quotient $$T_{{\mathrm{Def}}_{(C,L)}}/{\overline}{{\mathrm{Aut}}(C,L)},$$ where $T_{{\mathrm{Def}}_{(C,L)}}$ is a ${\mathbb{C}}$-vector space of dimension $4g-3$ (by Proposition \[semiuniv\]) and ${\overline}{{\mathrm{Aut}}(C,L)}$ can be naturally identified with a finite subgroup of ${\mathrm{GL}}(T_{{\mathrm{Def}}_{(C,L)}})$ . By a well-known result of Prill (see [@Prill]), it is enough to prove that ${\overline}{{\mathrm{Aut}}(C,L)}\subseteq {\mathrm{GL}}(T_{{\mathrm{Def}}_{(C,L)}})$ does not contain quasi-reflections, i.e. elements $\phi$ such that $1$ is an eigenvalue of $\phi$ with multiplicity equal to $4g-4$ or, equivalently, such that the fixed locus ${\rm Fix}(\phi)$ of $\phi$ is a divisor inside $T_{{\mathrm{Def}}_{(C,L)}}$. Consider the morphism $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ which, according to , locally looks like $$T_{{\mathrm{Def}}_{(C,L)}}/{\overline}{{\mathrm{Aut}}(C,L)}\to T_{{\mathrm{Def}}_{C^{\rm st}}}/{\mathrm{Aut}}(C^{\rm st}),$$ where $T_{{\mathrm{Def}}_{(C,L)}}\twoheadrightarrow T_{{\mathrm{Def}}_{C^{\rm st}}}$ is surjective with kernel $V$ of dimension $g$ and ${\mathrm{Aut}}(C^{\rm st})$ can be naturally identified with a finite subgroup of ${\mathrm{GL}}(T_{{\mathrm{Def}}_{C^{\rm st}}})$. Assume, by contradiction, that $\phi\in {\overline}{{\mathrm{Aut}}(C,L)}\subseteq {\mathrm{GL}}(T_{{\mathrm{Def}}_{(C,L)}})$ is a quasi-reflection. By the above local description of the morphism $\phi_d$, there are two possibilities for the image ${\overline}{G}(\phi)$ of $\phi$ in ${\mathrm{Aut}}(C^{\rm st})\subseteq {\mathrm{GL}}(T_{{\mathrm{Def}}_{C^{\rm st}}})$ via the homomorphism ${\overline}{G}$ of Lemma \[automo\]: (i) $1$ is an eigenvalue of multiplicity $3g-3$ for ${\overline}{G}(\phi)$, i.e. ${\overline}{G}(\phi)={\mathrm{id}}\in {\mathrm{Aut}}(C^{\rm st})$; (ii) $1$ is an eigenvalue of multiplicity $3g-4$ for ${\overline}{G}(\phi)$, i.e. ${\overline}{G}(\phi)$ is a quasi-reflection for ${\mathrm{Aut}}(C^{\rm st})\subseteq {\mathrm{GL}}(T_{{\mathrm{Def}}_{C^{\rm st}}})$. In case (i), we conclude that $\phi={\mathrm{id}}\in {\mathrm{Aut}}(C,L)$ since ${\overline}{G}$ is injective for an element $(C,L)\in {{\overline}{P}_{d,g}^{\rm st}}$ by Corollary \[Cor-auto\]. This contradicts the fact that $\phi$ is a quasi-reflection. In case (ii), it is well-known (see e.g. [@Lud Cor. 4.2.6]) that $C^{\rm st}$ must have an elliptic tail $E$ and ${\overline}{G}(\phi)$ must be equal to the elliptic tail automorphism $i$ of $C^{\rm st}$ of order $2$ with respect to $E$ (see Lemma \[auto-ell\]). Since $i={\overline}{G}(\phi)$ admits a lifting to ${\overline}{{\mathrm{Aut}}(C,L)}$, namely $\phi$, the line bundle $L$ on $C$ must be such that the restriction $L_{|E}$ of $L$ to $E$ is a suitable translate of a $2$-torsion point of ${\mathrm{Pic}}^0(E)$ (using some identification ${\mathrm{Pic}}^{d_E}(E)\cong {\mathrm{Pic}}^0(E)$ and the fact that $i$ acts on ${\mathrm{Pic}}^0(E)$ sending $\eta$ into $\eta^{-1}$). Therefore, the fixed locus ${\rm Fix}(\phi)$ of $\phi$ inside $T_{{\mathrm{Def}}_{(C,L)}}$ has codimension at least two, hence $\phi$ is not a quasi-reflection. By applying the Reid–Tai criterion for canonical singularities, we can prove the following result. \[sing-Pdst\] Assume $g\geq 4$. Then the stable locus ${{\overline}{P}_{d,g}^{\rm st}}$ has canonical singularities. In particular, if ${\widetilde}{{{\overline}{P}_{d,g}^{\rm st}}}$ is a resolution of singularities of ${{\overline}{P}_{d,g}^{\rm st}}$, then every pluricanonical form defined on the smooth locus $({{\overline}{P}_{d,g}^{\rm st}})^{\rm reg}$ of ${{\overline}{P}_{d,g}^{\rm st}}$ extends holomorphically to ${\widetilde}{{{\overline}{P}_{d,g}^{\rm st}}}$, that is, for all integers $m$ we have $$h^0\left(({{\overline}{P}_{d,g}^{\rm st}})^{\rm reg},m K_{({{\overline}{P}_{d,g}^{\rm st}})^{\rm reg}}\right)=h^0\left({\widetilde}{{{\overline}{P}_{d,g}^{\rm st}}}, m K_{{\widetilde}{{{\overline}{P}_{d,g}^{\rm st}}}}\right).$$ We use the notation introduced in the proof of Proposition \[sing\]. Given an element $\phi\in {\overline}{{\mathrm{Aut}}(C,L)}\subseteq {\mathrm{GL}}(T_{{\mathrm{Def}}_{(C,L)}})$ of order $n$, we can choose suitable coordinates of $T_{{\mathrm{Def}}_{(C,L)}}$ and a primitive $n$-th root of unity $\zeta$, such that the action of $\phi$ on $T_{{\mathrm{Def}}_{(C,L)}}$ is given by the sum $M(\phi)\oplus N(\phi)$ of two matrices (with $0\leq a_i< n$ for $1\leq i\leq 4g-3$): $$M(\phi)=\left( \begin{array}{ccc} \zeta^{a_1} & & 0 \\ & \ddots & \\ 0 & & \zeta^{a_{3g-3}} \\ \end{array} \right) \hspace{0,2cm}\text{ and }\hspace{0,2cm} N(\phi)=\left( \begin{array}{ccc} \zeta^{a_{3g-2}} & & 0 \\ & \ddots & \\ 0 & & \zeta^{a_{4g-3}} \\ \end{array} \right)$$ in such a way that the action of $\phi$ on $V$ is given by $N(\phi)$ and the action of ${\overline}{G}(\phi)$ on $T_{{\mathrm{Def}}_{C^{\rm st}}}$ is given by $M(\phi)$. Recall that, according to the Reid–Tai criterion for the canonicity of finite quotient singularities (see e.g. [@HarMum pp. 27–28] or [@Lud Thm. 4.1.11]), a point $(C,L)$ is a canonical singularity if and only if for every $\phi\in {\overline}{{\mathrm{Aut}}(C,L)}$ of some order $n$ and every $n$-th root of unity $\zeta$ we have $$\label{RT-cond} \sum_{i=1}^{4g-3}\frac{a_i}{n}\geq 1.$$ Note that this is true because ${\overline}{{\mathrm{Aut}}(C,L)}$ does not contain quasi-reflections (see the proof of Proposition \[sing\]). Denote, as usual, by $\Delta_1$ the divisor of ${\overline}{M}_g$ consisting of curves having an elliptic tail. If $C^{\rm st}\not\in \Delta_1$ or $C^{\rm st}\in \Delta_1$ but ${\overline}{G}(\phi)$ is not an elliptic tail automorphism (or equivalently, by Remark \[rmk-aut-tail\], $\phi$ is not an elliptic tail automorphism) then by [@HarMum Thm. 2] we get $$\sum_{i=1}^{4g-3}\frac{a_i}{n}\geq \sum_{i=1}^{3g-3}\frac{a_i}{n}\geq 1,$$ and we are done in this case. If $C^{\rm st}\in \Delta_1$ and ${\overline}{G}(\phi)$ is an elliptic tail automorphism with respect to the elliptic tail $E^{\rm st}\subset C^{\rm st}$ (where $E^{\rm st}$ is equal to the image via ${\rm st}:C\to C^{\rm st}$ of the elliptic tail $E\subset C$ as in Remark \[rmk-ell-tail\]) then we choose, as in [@Lud Prop. 4.2.5], the first two coordinates $t_1$ and $t_2$ of $T_{{\mathrm{Def}}_{C^{\rm st}}}$ in such a way that (in the notation of Lemma \[tail-stable\]): $t_1$ corresponds to the elliptic tail node $P$ and $t_2$ correspond to $Q$ if $E^{\rm st}$ is singular and is a coordinate for $T_{(E,P)}(M_{1,1})$ if $E$ is smooth. In [@Lud Prop. 4.2.5], it is proved that the matrix $M(\phi)$ is given by (depending on the choice of the primitive $n$-th root of unity $\zeta$): $$\label{M-matrix} M(\phi)=\begin{cases} \left( \begin{array}{ccc} \zeta^{1} & & \\ & \zeta^0 & \\ & & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=2,\\ \left( \begin{array}{ccc} \zeta^{1} & & \\ & \zeta^2 & \\ & & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{ccc} \zeta^{3} & & \\ & \zeta^2 & \\ & & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=4,\\ \left( \begin{array}{ccc} \zeta^{1} & & \\ & \zeta^2 & \\ & & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{ccc} \zeta^2 & & \\ & \zeta^1 & \\ & & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=3,\\ \left( \begin{array}{ccc} \zeta^5 & & \\ & \zeta^4 & \\ & & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{ccc} \zeta^1 & & \\ & \zeta^2 & \\ & & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=6,\\ \end{cases}$$ where ${\mathbb{I}}$ is the suitable unit matrix. Let us now turn to the matrix $N(\phi)$. We choose the first coordinate $s_1$ on $V$ so that it is a coordinate for $T_{L_{|E}}({\mathrm{Pic}}^{d_E}(E))$, where $d_E$ is defined in Lemma \[tail-stable\]. In order to compute the action of $\phi$ on $s_1$, we distinguish three case according to whether the elliptic tail $E\subset C$ is of Type I, II or III (see Lemma \[tail-stable\] and Figure \[Fig-tails\]). If $E$ is of Type I, i.e. $E$ is smooth, then we can identify $E$ with ${\mathrm{Pic}}^{d_E}(E)$ sending $q\in E$ into ${\mathcal{O}}_{E}(q+(d_E-1)P)\in {\mathrm{Pic}}^{d_E}(E)$. Since $\phi$ acts on ${\mathrm{Pic}}^{d_E}(E)$ via pull-back, if the action of ${\overline}{G}(\phi)$ on $T_P(E)$ is given by the multiplication by a root of unity $\zeta$, then the action of $\phi$ on $T_{L_{|E}}({\mathrm{Pic}}^{d_E}(E))$ is given by the multiplication by $\zeta^{-1}$. Therefore the matrix $N(\phi)$ is equal to (with respect to the same choice of the primitive $n$-th root of unity $\zeta$ as in the above matrix $M(\phi)$): $$\label{N-matrix} N(\phi)=\begin{cases} \left( \begin{array}{cc} \zeta^{1} & \\ & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=2,\\ \left( \begin{array}{cc} \zeta^{3} & \\ & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{cc} \zeta^{1} & \\ & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=4,\\ \left( \begin{array}{cc} \zeta^{2} & \\ & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{cc} \zeta^1 & \\ & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=3,\\ \left( \begin{array}{cc} \zeta^1 & \\ & {\mathbb{I}}\\ \end{array} \right) \text{ or } \left( \begin{array}{cc} \zeta^5 & \\ & {\mathbb{I}}\\ \end{array} \right) & \text{ if } n=6.\\ \end{cases}$$ If $E$ is of type II, i.e. $E$ is an irreducible rational curve with one node $Q$ (as in Figure \[Fig-tails\]), then ${\mathrm{Pic}}^{d_E}(E)\cong {\mathbb{G}_m}$. Explicitly, if we consider the normalization morphism $\nu:E^{\nu}\cong {\mathbb{P}}^1\to E$ and let $\nu^{-1}(Q)=\{u,v\}$, then any $\lambda\in {\mathbb{G}_m}(k)$ determines a unique line bundle $L_{\lambda}\in {\mathrm{Pic}}^{d_E}(E)$ whose local sections are the local sections $s$ of ${\mathcal{O}}_{{\mathbb{P}}^1}(d_E)$ such that $s(u)=\lambda s(v)$. Since $\phi_{|E}$ is induced by an involution of $E^{\nu}$ that exchanges $u$ and $v$ (by Lemma \[auto-ell\]), then clearly $\phi$ will send $L_{\lambda}$ into $L_{\lambda^{-1}}$. This implies that the action of $\phi$ on $T_{L_{|E}}({\mathrm{Pic}}^{d_E}(E))$ is given by multiplication by $-1$, hence the matrix $N(\phi)$ is also in this case given by with $n=2$. If $E$ is of type III, i.e. $E$ is made of two irreducible rational components $R_1$ and $R_2$ meeting in two points $Q_1$ and $Q_2$ (as in Figure \[Fig-tails\]), then again ${\mathrm{Pic}}^{d_E}(E)\cong {\mathbb{G}_m}$. Explicitly, if we consider the normalization morphism $\nu:E^{\nu}=R_1\coprod R_2 \to E$ and let $\nu^{-1}(Q_i)=\{u_i,v_i\}$ with $u_i\in R_1$ and $v_i\in R_2$ (for $i=1,2$), then any $\lambda\in {\mathbb{G}_m}(k)$ determines a unique line bundle $L_{\lambda}\in {\mathrm{Pic}}^{d_E}(E)$ whose local sections are pairs of local sections $(s_1,s_2)$ of $({\mathcal{O}}_{R_1}(d_E-1), {\mathcal{O}}_{R_2}(1))$ such that $$\frac{s_1(u_1)}{s_1(u_2)}=\lambda \frac{s_2(v_1)}{s_2(v_2)}.$$ Since $\phi_{|E}$ is induced by an involution of $E^{\nu}$ that exchanges $u_1$ with $u_2$ and $v_1$ with $v_2$ (by Lemma \[auto-ell\]), then clearly $\phi$ will send $L_{\lambda}$ into $L_{\lambda^{-1}}$. This implies that the action of $\phi$ on $T_{L_{|E}}({\mathrm{Pic}}^{d_E}(E))$ is given by multiplication by $-1$, hence the matrix $N(\phi)$ is also in this case given by with $n=2$. An easy inspection of the matrices $M(\phi)$ in and $N(\phi)$ in reveals that the condition (\[RT-cond\]) is always satisfied, which shows that ${{\overline}{P}_{d,g}^{\rm st}}$ has canonical singularities. The last assertion of the theorem follows from the well-known fact that canonical singularities do not impose adjoint conditions on the pluricanonical forms. The canonical class of ${\overline{\mathcal Pic}_{d,g}}$ and of ${\overline{P}_{d,g}}$. {#canonical} ======================================================================================= The aim of this section is to prove Theorem \[K\]. To achieve that, we first determine the canonical class of the stack ${\overline{\mathcal Pic}_{d,g}}$. \[can-stack\] The canonical class of ${\overline{\mathcal Pic}_{d,g}}$ is equal to $$K_{{\overline{\mathcal Pic}_{d,g}}}=\Phi_d^*(14\lambda-2\delta),$$ where $\lambda$ and $\delta$ are the Hodge and total boundary class on ${\overline{\mathcal{M}}_g}$. Let $\pi: {{\overline}{\mathcal Pic}_{d,g,1}}\rightarrow {\overline{\mathcal Pic}_{d,g}}$ the universal family over ${\overline{\mathcal Pic}_{d,g}}$ and ${\mathcal{L}}_d$ the universal line bundle over ${{\overline}{\mathcal Pic}_{d,g,1}}$ (see [@Melo2] for a modular description of ${{\overline}{\mathcal Pic}_{d,g,1}}$). Denote by $\Omega_{\pi}$ and $\omega_{\pi}$ the sheaf of relative Kähler differentials and the relative dualizing sheaf, respectively. Let $d: {\mathcal O}_{{{\overline}{\mathcal Pic}_{d,g,1}}}^* \rightarrow \Omega_{\pi}$ be the universal derivation and consider the map induced in cohomology $\theta: {\mathrm{Pic}}({{{\overline}{\mathcal Pic}_{d,g,1}}}) \rightarrow H^1({{\overline}{\mathcal Pic}_{d,g,1}}, \Omega_{\pi})$. Since $H^1({{{\overline}{\mathcal Pic}_{d,g,1}}}, \Omega_{\pi}) \cong {\mathrm{Ext}}^1({\mathcal O}_{{{{\overline}{\mathcal Pic}_{d,g,1}}}},\Omega_{\pi}),$ the map $\theta$ sends the line bundle ${{\mathcal{L}}_d}$ on ${{{\overline}{\mathcal Pic}_{d,g,1}}}$ into the class of an extension $$\label{Atiya} 0 \rightarrow \Omega_{\pi} \rightarrow \mathcal{E} \rightarrow {\mathcal O}_{{{{\overline}{\mathcal Pic}_{d,g,1}}}}\rightarrow 0.$$ The restriction of the above extension (\[Atiya\]) to a geometric fiber $(C,L)$ of $\pi$ is the extension (\[jets\]) as it follows from the discussion in Section \[loc-struc\]. &gt;From this and the analysis of the deformation theory of the pair $(C,L)$ carried out in \[loc-struc\], it follows that the tangent space of ${{\overline}{\mathcal Pic}_{d,g}}$ at a geometric point $(C,L)$ is equal to ${\mathrm{Ext}}^1({\mathcal{E}}_{|C},{\mathcal{O}}_{C})-{\mathrm{Ext}}^0({\mathcal{E}}_{|C},{\mathcal{O}}_C)$ Therefore, using relative duality for $\pi$, it follows that the canonical class $K_{{\overline{\mathcal Pic}_{d,g}}}$ of ${\overline{\mathcal Pic}_{d,g}}$ is equal to $$K_{{\overline{\mathcal Pic}_{d,g}}}=c_1\left(\pi_!\left({\mathcal{E}}\otimes \omega_{\pi}\right)\right).$$ To compute this class, we apply the Grothendieck-Riemann-Roch Theorem for quotient stacks ([@EG]) relative to the morphism $\pi$: $$\label{GRR} {\mathrm{ch}}\left(\pi_!\left(\omega_{\pi} \otimes \mathcal{E}\right)\right)=\pi_*\left({\mathrm{ch}}(\omega_{\pi} \otimes \mathcal{E})\cdot {\mathrm{Td}}(\Omega_{\pi})^{-1}\right).$$ Let us compute the degree one part of the right hand side of (\[GRR\]). We set ${\widetilde}{K}:=c_1(\omega_{\pi})$ and ${\widetilde}{\eta}:=c_2(\Omega_{\pi})$. Note that, as remarked in [@HM p. 158], we have ${\widetilde}{K}=c_1(\Omega_{\pi})$. The first three terms of inverse of the Todd class of $\Omega_{\pi}$ are equal to $$\label{Todd} {\mathrm{Td}}(\Omega_{\pi})^{-1}=1-\frac{c_1(\Omega_{\pi})}{2}+\frac{c_1^2(\Omega_{\pi})+c_2(\Omega_{\pi})}{12} + \ldots= 1-\frac{{\widetilde}{K}}{2}+\frac{{\widetilde}{K}^2+{\widetilde}{\eta}}{12}+\ldots$$ Using (\[Atiya\]), we get $${\mathrm{ch}}({\mathcal{E}}\otimes \omega_{\pi})={\mathrm{ch}}({\mathcal{E}})\cdot {\mathrm{ch}}(\omega_{\pi})=({\mathrm{ch}}(\Omega_{\pi})+{\mathrm{ch}}({\mathcal{O}}_{{{\overline}{\mathcal Pic}_{d,g,1}}}))\cdot {\mathrm{ch}}(\omega_{\pi})=$$ $$\label{ch1} =({\mathrm{ch}}(\Omega_{\pi})+1)\cdot {\mathrm{ch}}(\omega_{\pi}).$$ Moreover we have $${\mathrm{ch}}(\omega_{\pi})=1+c_1(\omega_{\pi})+\frac{c_1(\omega_{\pi})}{2}+\ldots=1+{\widetilde}{K}+\frac{{\widetilde}{K}^2}{2}+\ldots$$ $${\mathrm{ch}}(\Omega_{\pi})=1+c_1(\Omega_{\pi})+\frac{c_1(\Omega_{\pi})}{2}-c_2(\Omega_{\pi})+\ldots=1+{\widetilde}{K}+\frac{{\widetilde}{K}^2}{2}-{\widetilde}{\eta}+\ldots$$ Substituting into (\[ch1\]), we arrive at $$\label{ch2} {\mathrm{ch}}({\mathcal{E}}\otimes \omega_{\pi})=2+3{\widetilde}{K}+\frac{5}{2}{\widetilde}{K}^2-{\widetilde}{\eta}+\ldots$$ Combining (\[Todd\]) and (\[ch2\]), we get $$\left[{\mathrm{ch}}(\omega_{\pi} \otimes \mathcal{E})\cdot {\mathrm{Td}}(\Omega_{\pi})^{-1}\right]_2= \frac{{\widetilde}{K}^2+{\widetilde}{\eta}}{6}-\frac{3}{2}{\widetilde}{K}^2+\frac{5}{2}{\widetilde}{K}^2-{\widetilde}{\eta}= \frac{7}{6}{\widetilde}{K}^2-\frac{5}{6}{\widetilde}{\eta},$$ hence, from (\[GRR\]), we deduce $$\label{Kpi} K_{{\overline{\mathcal Pic}_{d,g}}}=\frac76 \pi_*({\widetilde}{K}^2) - \frac56 \pi_*({\widetilde}{\eta}).$$ Let us now apply the Grothendieck-Riemann-Roch theorem to the sheaf $\omega_{\pi}$. Since $R^1\pi_*\omega_{\pi}={\mathcal{O}}_{{\overline{\mathcal Pic}_{d,g}}}$ by relative duality, we get $$c_1(\pi_*\omega_{\pi})=\pi_*\left(\frac{{\widetilde}{K}^2+{\widetilde}{\eta}}{12}-\frac12 {\widetilde}{K}^2 + \frac12{\widetilde}{K}^2\right)=\frac{1}{12} \pi_*({\widetilde}{K}^2)+ \frac{1}{12}\pi_*({\widetilde}{\eta}).$$ If we set ${\widetilde}{\lambda}:=c_1(\pi_*\omega_{\pi})$ and ${\widetilde}{\delta}:=\pi_*({\widetilde}{\eta})$, then the previous relation becomes $12{\widetilde}{\lambda}=\pi_*({\widetilde}{K}^2)+{\widetilde}{\delta}.$ Substituting into (\[Kpi\]), we obtain $$K_{{\overline{\mathcal Pic}_{d,g}}}=14 {\widetilde}{\lambda}-2 {\widetilde}{\delta}.$$ The Lemma below completes the proof. With the notation of Theorem \[can-stack\], we have $${\widetilde}{\lambda}=\Phi_d^*(\lambda) \text{ and } {\widetilde}{\delta}=\Phi_d^*(\delta).$$ Consider the diagram $$\label{univ-fam} \xymatrix{ {{\overline}{\mathcal Pic}_{d,g,1}}\ar[r]^{\Phi_{d,1}} \ar[d]_{\pi} & {\overline{\mathcal M}_{g,1}}\ar[d]^{{\overline}{\pi}}\\ {\overline{\mathcal Pic}_{d,g}}\ar[r]^{\Phi_d} & {\overline{\mathcal{M}}_g}\\ }$$ Recall that the classes $\lambda$ and $\delta$ on ${\overline{\mathcal{M}}_g}$ are defined as $$\lambda:=c_1({\overline}{\pi}_*(\omega_{{\overline}{\pi}})) \text{ and }�\delta:={\overline}{\pi}_*(c_2(\Omega_{{\overline}{\pi}})),$$ where $\Omega_{{\overline}{\pi}}$ and $\omega_{{\overline}{\pi}}$ are the sheaf of relative Kähler differentials and the relative dualizing sheaf of ${\overline}{\pi}$, respectively. The map $\Phi_d$ sends an element $({\mathcal{C}}\to S, {\mathcal{L}})\in {\overline{\mathcal Pic}_{d,g}}(S)$ into the stabilization ${\mathcal{C}}^{\rm st}\to S\in {\overline{\mathcal{M}}_g}(S)$. Recall that for every quasi-stable (or more generally semistable) curve $C$ with stabilization morphism $\psi:C\to C^{\rm st}$, the pull-back via $\psi$ induces a natural isomorphism $\psi^*:H^0(C^{\rm st},\omega_{C^{\rm st}})\stackrel{\cong}{\to} H^0(C,\omega_C)$. Therefore we have $\Phi_d^*({\overline}{\pi}_*(\omega_{{\overline}{\pi}}))=\pi_*(\omega_{\pi})$ and, by taking the first Chern classes, we get ${\widetilde}{\lambda}=\Phi_d^*(\lambda)$. On the other hand, since the class ${\widetilde}{\delta}$ is the total boundary class of ${\overline{\mathcal Pic}_{d,g}}$ and $\delta$ is the total boundary class of ${\overline{\mathcal{M}}_g}$ (see [@HarMum pp. 49–50]), it is clear that ${\widetilde}{\delta}=\Phi_d^*(\delta)$. Let $p:{\overline{\mathcal Pic}_{d,g}}\to {\overline{P}_{d,g}}$ the natural map from the stack ${\overline{\mathcal Pic}_{d,g}}$ to its good moduli space. In view of Theorem \[can-stack\], it is enough to show that $p^*(K_{{\overline{P}_{d,g}}})=K_{{\overline{\mathcal Pic}_{d,g}}}$. Clearly, the two classes agree on the interior ${{\mathcal Pic}_{d,g}}$, since for $C$ varying in an open subset of $M_g$ whose complement has codimension at least two (since $g\geq 4$), by Lemma \[automo\] we have that ${\overline}{{\mathrm{Aut}}(C,L)}\subseteq {\mathrm{Aut}}(C)=\{{\mathrm{id}}\}$. Let us now look at the boundary of ${\overline{P}_{d,g}}$. By [@Fon Prop. 4], the boundary of ${\overline{P}_{d,g}}$ is the union of the irreducible divisors $D_i:=\phi_d^{-1}(\Delta_i)$, for $i=0,\cdots,[g/2]$. Let $k_{d,g}:=(2g-2, d-g+1)$. A general element of $(C,L)\in D_i$ looks as follows (see e.g [@Cap Ex. 7.1, 7.2] and [@Melo Prop. 2]): 1. \[case1\] If $i=0$ then $C$ is a general irreducible nodal curve with one node and $L$ is a general line bundle of degree $d$ on $C$; 2. \[case2\] If $i>0$ and $2g-2$ does not divide $(2i-1)\cdot k_{d,g}$, then $C$ is a stable curve consisting of two general smooth curves $C_1$ and $C_2$ of genera, respectively, $i$ and $g-i$ meeting in one point and $L$ is a general line bundle of multidegree $(\deg_{C_1}(L),\deg_{C_2}(L))=(d_1,d_2=d-d_1)$ where $d_1$ is the unique integer such that $$\left|d_1-\frac{d(2i-1)}{2g-2}\right|< \frac{1}{2}.$$ 3. \[case3\] If $i>0$ and $2g-2$ divides $(2i-1)\cdot k_{d,g}$, then $C$ is a quasi-stable curve consisting of two general smooth curves $C_1$ and $C_2$ of genera, respectively, $i$ and $g-i$ joined by a rational curve $R\cong {\mathbb{P}}^1$ and $L$ is a general line bundle whose multidegree is such that $\deg_R L=1$ and $$\left\{\begin{aligned} & d_1:=\deg_{C_1}L=\frac{d(2i-1)}{(2g-2)}-\frac{1}{2},\\ & d_2:=\deg_{C_2} L= \frac{d(2g-2i-1)}{(2g-2)}-\frac{1}{2}. \end{aligned}\right.$$ We claim that the automorphism group of a general point $(C,L)\in D_i$ is equal to $$\label{auto-boun} {\mathrm{Aut}}(C,L)=\begin{cases} {\mathbb{G}_m}& \text{ in cases } \ref{case1} \text{ and }�\ref{case2}, \\ {\mathbb{G}_m}^2 & \text{ in case } \ref{case3}. \end{cases}$$ Indeed, by the explicit description above, $\gamma({\widetilde}{C})=1$ in cases (\[case1\]) and (\[case2\]), and $\gamma({\widetilde}{C})=2$ in case (\[case3\]). Therefore, the claim will follow from Lemma \[automo\] if we show that the image of ${\mathrm{Aut}}(C,L)\to {\mathrm{Aut}}(C^{\rm st})$ is trivial. This is trivially true if $i\neq 1$ since in this case $C^{\rm st}$ is a general curve in $\Delta_i$, hence ${\mathrm{Aut}}(C^{\rm st})=\{{\mathrm{id}}\}$. If $i=1$ then ${\mathrm{Aut}}(C^{\rm st})={\mathbb{Z}}/2{\mathbb{Z}}$ generated by the elliptic tail involution $\sigma$ with respect to the elliptic tail $C_1$ (see Remark \[rmk-aut-tail\] and the notation there). However, in this case, $\sigma$ comes from an automorphism of the pair $(C,L)$ if and only if $L_{|C_1}(-d_1\cdot P)$ is a $2$-torsion point of ${\mathrm{Pic}}^0(C_1)$, where $P=C_1\cap R$ is the elliptic tail node of $C_1$. Clearly, this is not the case for a general strictly balanced line bundle $L$ on $C$. In cases (\[case1\]) and (\[case2\]), ${\underline}{{\mathrm{Def}}}_{(C,L)}$ has dimension $4g-3$ and ${\mathrm{Aut}}(C,L)={\mathbb{G}_m}$ acts trivially on it (see \[loc-struc\]). Therefore, the morphism $p$ looks locally at $(C,L)$ as ${\widetilde}{p}:[{\underline}{{\mathrm{Def}}}_{(C,L)}/{\mathbb{G}_m}]={\underline}{{\mathrm{Def}}}_{(C,L)}\times B{\mathbb{G}_m}\to {\underline}{{\mathrm{Def}}}_{(C,L)}$. It is clear that in this case ${\widetilde}{p}^*(K_{{\underline}{{\mathrm{Def}}}_{(C,L)}})=K_{{\underline}{{\mathrm{Def}}}_{(C,L)}\times B{\mathbb{G}_m}}$. In case (\[case3\]), ${\underline}{{\mathrm{Def}}}_{(C,L)}$ has dimension $4g-2$ (see \[loc-struc\]). If we choose the first two coordinates $x$ and $y$ of ${\underline}{{\mathrm{Def}}}_{(C,L)}$ in such a way that they correspond to the local deformations of the two nodes $P_1:=C_1\cap R$ and $P_2:=C_2\cap R$, then the action of $(\mu, \nu)\in {\mathrm{Aut}}(C,L)={\mathbb{G}_m}^2$ on the first two coordinates of ${\underline}{{\mathrm{Def}}}_{(C,L)}$ is given by $$\label{exp-act} (\mu, \nu)\cdot (x,y)=(\mu\nu^{-1}x, \mu^{-1}\nu y),$$ while it is trivial on the other coordinates. Therefore, neglecting the trivial coordinates, at $(C,L)$ the morphism $p$ looks locally as $$\X:=[{\mathrm{Spf}}{\mathbb{C}}[[x,y]]/{\mathbb{G}_m}^2]\stackrel{{\widetilde}{p}}{\longrightarrow} {\mathrm{Spf}}{\mathbb{C}}[[x,y]]/{\mathbb{G}_m}^2:=X.$$ Since the ring of invariants for the action (\[exp-act\]) of ${\mathbb{G}_m}^2$ on ${\mathbb{C}}[[x,y]]$ is generated by $xy$, the quotient $X$ is isomorphic to ${\mathrm{Spf}}{\mathbb{C}}[[xy]]$. The quotient map $Y:={\mathrm{Spf}}{\mathbb{C}}[[x,y]]\stackrel{q}{\to} X={\mathrm{Spf}}{\mathbb{C}}[[xy]]$ induces a pull-back map $$\begin{aligned} q^*:\langle d(xy)\rangle= \Omega^1_X& \to \Omega^1_Y=\langle dx, dy\rangle,\\ d(xy) & \mapsto xdy+ydx. \end{aligned}$$ On the other hand, the cotangent complex of $\X$ is equal to the ${\mathbb{G}_m}^2$-equivariant cotangent complex of $Y:={\mathrm{Spf}}{\mathbb{C}}[[x,y]]$ (see e.g. [@Moc p. 37]), which in our case looks like: $${\mathbb{L}}: \langle dx, dy\rangle=\Omega^1_Y \stackrel{f}{\to} {\mathcal{O}}_Y\otimes {\rm Lie}({\mathbb{G}_m}^2)^*={\mathcal{O}}_Y\otimes \left\langle \frac{d\lambda}{\lambda}, \frac{d\mu}{\mu}\right\rangle$$ The map $f$ is the dual of the infinitesimal action of ${\rm Lie}({\mathbb{G}_m}^2)$ on $Y$, hence, by the explicit action (\[exp-act\]), we compute $$\left\{\begin{aligned} f(dx)= & x\frac{d\lambda}{\lambda}-x\frac{d\mu}{\mu},\\ f(dy)= & y\frac{d\mu}{\mu}-y\frac{d\lambda}{\lambda}.\\ \end{aligned}\right.$$ Since the image of $q^*$ is equal to the kernel of $f$, we deduce that ${\widetilde}{p}^*(K_X)=K_{\X}$, which concludes the proof. The Iitaka dimension of $K_{{\overline{P}_{d,g}}}$ {#Iitaka} ================================================== The aim of this section is to prove Theorem \[Iitaka-Mg\]. Since $K_{{\overline{P}_{d,g}}}=\phi_d^*(14\lambda-2\delta)$ (by Theorem \[K\]) and $\phi_d$ has connected fibers, we have $$\label{ineq-Kod} \kappa({K_{{\overline{P}_{d,g}}}})= \kappa(14\lambda-2\delta).$$ Therefore, we are reduced to study the Iitaka dimension of the divisor $14\lambda-2\delta$ on ${\overline{M}_g}$. Notice that the slope of $14\lambda-2\delta$ is equal to $7$. Hence, if the slope $s({\overline{M}_g})$ of ${\overline{M}_g}$ (in the sense of Harris-Morrison [@HM90]) is strictly less than $7$ then we conclude that $14\lambda-2\delta$ is big, i.e. that $\kappa(14\lambda-2\delta)=3g-3$; while if $s({\overline{M}_g})>7$ then $14\lambda-2\delta$ is not pseudo-effective and $\kappa(14\lambda-2\delta)=-\infty$ (see the discussion at the beginning of [@HM90]). \[kod-infty\] If $g \le 9$ then $\kappa(K_{{\overline{P}_{d,g}}})=-\infty$. This follows by the fact that $s({\overline{M}_g})>7$ for $g\leq 9$ (see [@Tan]). By combining the above Proposition \[kod-infty\] with the inequality , we obtain another proof of the fact that $\kappa(P_{d,g})=-\infty$ for $g\leq 9$ and any $d$ (which of course follows from the stronger Theorem \[Ver-thm\]). For $g\geq 12$ we can prove the following \[dim-Kod\] If $g\geq 12$ then $\kappa({K_{{\overline{P}_{d,g}}}})=3g-3$ and the fibration $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ is the Iitaka fibration of ${K_{{\overline{P}_{d,g}}}}$. We have already observed, in the proof of Proposition \[kod-fib\], that if $\kappa({K_{{\overline{P}_{d,g}}}})=3g-3$ then $\phi_d:{\overline{P}_{d,g}}\to {\overline{M}_g}$ is the Iitaka fibration (see [@Laz Def. 2.1.34]) of ${K_{{\overline{P}_{d,g}}}}$. Therefore, it is enough to prove the first assertion. As pointed out before, this will follow if we show that $s({\overline{M}_g})<7$ for $g\geq 12$. By computing the class of the Brill-Noether divisor $D_{\frac{g+1}{2}}^1$, Harris and Mumford proved in [@HarMum] that $$s({\overline{M}_g})\leq 6+\frac{12}{g+1} \text{ if } g \text{ is odd. }$$ Since $6+\frac{12}{g+1}<7$ if and only if $g>11$, we get that $$\label{equ1} s({\overline{M}_g})<7 \text{ if } g \text{ is odd and } g\geq 13.$$ By computing the class of the Petri divisor $E_{\frac{g}{2}+1}^1$, Eisenbud and Harris in [@EH Thm. 2] proved that $$s({\overline{M}_g})\leq 6+\frac{14g+4}{g(g+2)} \text{ if } g \text{ is even. }$$ Since $6+\frac{14g+4}{g(g+2)}<7$ if and only if $g>13$, we get that $$\label{equ2} s({\overline{M}_g})<7 \text{ if } g \text{ is even and } g\geq 14.$$ By computing the slope of some effective divisors on ${\overline{M}_g}$ associated to curves equipped with secant-exceptional linear series, Cotterill in [@Cot §6.2] showed in particular that $$\label{equ3} s(\overline{M}_{12})\leq 6,979...< 7.$$ Equations (\[equ1\]), (\[equ2\]) and (\[equ3\]) together imply the result. The cases $g=10$ and $g=11$ requires a special care since it is known that in this case $s({\overline{M}_g})=7$ (see [@Tan] and [@FarPop Cor. 1.3]). We start with the case $g=10$. \[dim-Kod10\] If $g=10$ then $\kappa({K_{{\overline{P}_{d,g}}}})=0$. Farkas and Popa proved in [@FarPop Thm. 1.6] that the effective irreducible divisor $F$ (which is denoted by ${\overline}{K}$ in loc. cit.) given by the closure of the locus of smooth curves of genus $10$ lying on a $K3$ surface has class equal to $$\label{class-K} F=7\lambda-\delta_0-5 \delta_1-9 \delta_2-12 \delta_3-14 \delta_4- B_5\delta_5,$$ with $B_5\geq 6$. Since it is easily checked that $14\lambda-2\delta$ is the sum of $2 F$ and an effective boundary divisor, we get, using (\[ineq-Kod\]), that $\kappa({K_{{\overline{P}_{d,g}}}})= \kappa(14\lambda-2\delta)\geq \kappa(2F)=\kappa(F)\geq 0$. It remains to prove that $h^0({\overline}{M}_{10}, m(14\lambda-2\delta))=1$ for any $m$ sufficiently divisible. : If $m$ is sufficiently divisible then $2mF$ is contained in the base locus of $ \vert m(14\lambda-2\delta) \vert$. Take $D \in \vert m(14\lambda-2\delta) \vert$ and let $r$ be the multiplicity of $F$ inside $D$. Consider a Lefschetz pencil of curves of genus $10$ lying on a general K3 surface of degree $18$ in ${\mathbb{P}}^{10}$. This gives rise to an irreducible curve $B$ in the moduli space ${\overline}{M}_{10}$. Such pencils $B$ fill the divisor $F$, by results of Mukai ([@Muk1]). Therefore $B$ is not contained in the support of $D-rF$, hence $(D-r F)\cdot B\geq 0$. Using the well-known formulas $\lambda\cdot B=g+1=11$, $\delta_0\cdot B=6(g+3)=78$, $\delta_i\cdot B=0$ for $i\geq 1$ (see e. g. [@FarPop Lemma 2.1]), together with the expression (\[class-K\]), we get that $$0\leq (D-r F)\cdot B=(2m-r)[(7\lambda-\delta_0)\cdot B]= r-2m,$$ which concludes the proof of the Claim. &gt;From the previous Claim, it follows that $$h^0({\overline}{M}_{10}, m(14\lambda-2\delta))=h^0({\overline}{M}_{10}, m(14\lambda-2\delta)-2mF).$$ Note that $m(14\lambda-2\delta)-2mF=m(\sum_{i\geq 1} a_i\delta_i)$ for some $a_i\geq 0$. Therefore the proof of the theorem is concluded by the following : Let $\Delta$ be an effective divisor in ${\overline{M}_g}$ (for $g\geq 3$) whose class in ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$ is equal to $\sum_{i\geq 0} a_i \delta_i$, with $a_i\geq 0$. Then $h^0({\overline{M}_g}, m\Delta)=1$ for any $m$ sufficiently divisible. Take $E\in \vert m\Delta\vert$. We have to show that $E=m\Delta$. If $E$ meets the interior $M_g$ of ${\overline{M}_g}$, then, from the well-known result that ${\mathrm{Pic}}(M_g)_{{\mathbb{Q}}}$ is generated by $\lambda$ and $\lambda$ is ample on $M_g$, we get that the class of $E$ in ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$ is equal to $a\lambda+\sum_{i\geq 0}b_i\delta_i$ with $a>0$ and $b_i\in {\mathbb{Z}}$. However the class of $E$ in ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$ is also equal to the class of $m\Delta$, which is $\sum_{i\geq 0} m a_i \delta_i$. This produces a non-trivial relation between $\lambda$ and the boundary classes $\delta_i$, which contradicts the well-known result that ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$ is freely generated by $\lambda$ and the boundary classes $\delta_i$ for $g\geq 3$ (see [@AC]). Therefore, $E$ must be entirely contained in the boundary ${\overline{M}_g}\setminus M_g=\cup_{i\geq 0}\Delta_i$ of ${\overline{M}_g}$. This implies that $E=\sum_{i\geq 0} b_i\Delta_i$ for some $b_i\geq 0$. Looking at the classes of $\sum_{i\geq 0} b_i\Delta_i$ and $m\Delta$ in ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$ and using the independence of the boundary classes $\delta_i$ in ${\mathrm{Pic}}({\overline{M}_g})_{{\mathbb{Q}}}$, we deduce that $E=m\Delta$, as required. We finally examine the case $g=11$. As usual, denote by ${\mathcal F}_{g}$ ($g\geq 3$) the moduli space of K3 surfaces endowed with a polarization of degree $2g-2$. By work of Mukai ([@Muk]), there exists a fibration $$\psi:{\overline}{M}_{11}\dashrightarrow {\mathcal F}_{11},$$ sending a general curve $C$ of genus $g$ into $(S, {\mathcal{O}}_S(C))$, where $S$ is the unique K3 surface containing $C$. \[dim-Kod11\] If $g=11$ then $\kappa({K_{{\overline{P}_{d,g}}}})=19$ and the Iitaka fibration of ${K_{{\overline{P}_{d,g}}}}$ is the composition $${\overline}{P}_{d,11}\stackrel{\phi_d}{\longrightarrow} {\overline}{M}_{11}\stackrel{\psi}{\dashrightarrow} {\mathcal F}_{11}.$$ Farkas and Popa proved in [@FarPop Prop. 6.2] that the Iitaka dimension of the divisor $$E:=7\lambda-\delta_0-5\delta_1-9\delta_2-8\delta_3-7 \delta_4-7 \delta_5$$ is $19$. Since it is easily checked that $14\lambda-2\delta$ is the sum of $2 E$ and an effective boundary divisor, we get, using (\[ineq-Kod\]), that $\kappa({K_{{\overline{P}_{d,g}}}})=\kappa(14\lambda-2\delta)\geq \kappa(2E)=\kappa(E)=19$. Consider now a general point $(S,L)\in {\mathcal F}_{11}$. The fiber of $\psi$ over $(S,L)$ is the open subset of the complete linear series $|L|\cong {\mathbb{P}}^{11}$ consisting of smooth connected curves. Pick a Lefschetz pencil on $S$ and consider the associated curve $B$ inside ${\overline}{M}_{11}$. It is well-known (see e. g. [@FarPop Lemma 2.1]) that $\lambda\cdot B=12$, $\delta_0\cdot B= 84$ and $\delta_i\cdot B=0$ for every $i>1$. This easily implies that $$\label{inter-B} (14\lambda-2\delta)\cdot B=0.$$ Consider now the Iitaka fibration of the divisor $K_{{\overline}{P}_{d,11}}$, which we denote by $$i_{K_{{\overline}{P}_{d,11}}}:{\overline}{P}_{d,11}\dashrightarrow I(K_{{\overline}{P}_{d,11}}).$$ Since $K_{{\overline}{P}_{d,11}}=\phi_d^*(14\lambda-2\delta)$, the Iitaka fibration $i_{14\lambda-2\delta}$ of $K_{{\overline}{P}_{d,11}}$ is the composition of the Iitaka fibration of $14\lambda -2\delta$ with $\phi_d$, i.e. we have a natural diagram (up to birationality) $$\xymatrix{ {\overline}{P}_{d,11}\ar[rr]^{\phi_d} \ar@{-->}[dr]^{i_{K_{{\overline}{P}_{d,11}}}}& & {\overline}{M}_{11}\ar@{-->}[dl]_{i_{14\lambda-2\delta}} \\ & I(K_{{\overline}{P}_{d,11}})= I(14\lambda-2\delta) }$$ Now, equation (\[inter-B\]) implies that the Iitaka fibration $i_{14\lambda-2\delta}$ contracts the general fiber $\psi^{-1}(S,L)\subset |L|\cong {\mathbb{P}}^{11}$. Therefore the Iitaka fibration $i_{14\lambda-2\delta}$ factors through the fibration $\psi$: $$\xymatrix{ & {\overline}{M}_{11} \ar@{-->}[dr]^{i_{14\lambda-2\delta}} \ar@{-->}[dl]_{\psi} & \\ {\mathcal F}_{11} \ar@{-->}[rr]^{\rho} && I(14\lambda-2\delta) }$$ Recall that $\dim {\mathcal F}_{11}=19$. On other hand, by the usual properties of the Iitaka fibration and what proved before, we have that $\dim I(14\lambda-2\delta)=\kappa(14\lambda-2\delta)\geq 19$. Since $\rho$ is dominant and has connected general fiber, this implies that $\rho$ is a birational isomorphism, hence we are done. Birationalities among different $P_{d,g}$’s {#bir-Pdg} =========================================== In this section, inspired by Lemma 8.1 in [@Cap], we investigate the following \[birat-que\] For what values of $d$ and $d'$ is $P_{d,g}$ birational to $P_{d',g}$? How do the birational maps among them look like? Note that if $d'=d+n(2g-2)$ for some $n\in {\mathbb{Z}}$ then we have the isomorphism $$\label{bira1} \begin{aligned} \psi^1_n:P_{d,g} & \stackrel{\cong}{\longrightarrow} P_{d',g} \\ (C,L) & \mapsto (C,L\otimes \omega_C^n), \end{aligned}$$ while if $d'=-d+n(2g-2)$ for some $n\in {\mathbb{Z}}$ we have the isomorphism $$\label{bira2} \begin{aligned} \psi^2_n:P_{d,g} & \stackrel{\cong}{\longrightarrow} P_{d',g} \\ (C,L) & \mapsto (C,L^{-1}\otimes \omega_C^n). \end{aligned}$$ Clearly the maps $\psi_n^1$ and $\psi_n^2$ commute with the projections $\phi_d$ and $\phi_{d'}$ onto $M_g$. Indeed, the converse is also true, as it follows from an argument of Caporaso (see [@Cap Lemma 8.1] and also, for further details, [@Cap-Lis Prop. 3.2.2]). *([@Cap])*\[Cap-thm\] If $\eta: P_{d,g}\dashrightarrow P_{d',g}$ is a birational map over $M_g$, i.e. a map $\eta$ inducing a commutative diagram $$\xymatrix{ P_{d,g}\ar@{-->}[r]^{\eta}\ar[d]_{\phi_d}& P_{d',g}\ar[d]^{\phi_{d'}} \\ M_g\ar[r]^{{\rm id}} & M_g }$$ then there exists $n\in {\mathbb{Z}}$ such that either $d'=d+n(2g-2)$ and $\eta=\psi^1_n$ or $d'=-d+n(2g-2)$ and $\eta=\psi^2_n$. By using our results on the Kodaira dimension of $P_{d,g}$, we can improve Theorem \[Cap-thm\] at least for genus big enough. \[birationa\] Assume that $g\geq 22$ or $g\geq 12$ and $(d-g+1,2g-2)=1$. Let $\eta: P_{d,g}\dashrightarrow P_{d',g}$ be a birational map. Then there exists $n\in {\mathbb{Z}}$ such that either $d'=d+n(2g-2)$ and $\eta=\psi^1_n$ or $d'=-d+n(2g-2)$ and $\eta=\psi^2_n$. In particular, $\eta$ is an isomorphism. By Theorems \[Kod-geom\] and \[kod-nongeo\], the assumptions of the statement imply that $\kappa(P_{d,g})=3g-3$, hence that $\kappa(P_{d',g})=3g-3$. From the proof of Proposition \[kod-fib\], it follows that $\phi_d:P_{d,g}\to M_g$ is the Iitaka fibration of $P_{d,g}$ and similarly for $P_{d',g}$. Since the Iitaka fibration is a birational invariant, the map $\eta$ induces a birational map $\xi:M_g\dashrightarrow M_g$ such that the following diagram commutes: $$\xymatrix{ P_{d,g} \ar@{-->}[r]^{\eta} \ar[d]_{\phi_d}& P_{d',g}\ar[d]^{\phi_{d'}} \\ M_g \ar@{-->}[r]^{\xi} & M_g. }$$ The map $\xi$ sends a very general curve $C\in M_g$ to a very general curve $C'\in M_g$ so that the restriction of $\eta$ induces a birational map $J(C)\cong{\mathrm{Pic}}^d(C) \dashrightarrow {\mathrm{Pic}}^{d'}(C')\cong J(C')$. By Lemma \[gen-Tor\] below, we get that $C\cong C'$, hence $\xi={\rm id}$. Finally, we conclude by Theorem \[Cap-thm\]. \[gen-Tor\] If $C$ and $C'$ are very general curves in $M_g$ such that $J(C)$ is birational to $J(C')$, then $C\cong C'$. Let $\epsilon:J(C)\dashrightarrow J(C')$ be a birational map. Since $J(C)$ and $J(C')$ are abelian varieties, then it is well-known that $\epsilon$ extends to an isomorphism $\epsilon:J(C)\stackrel{\cong}{\to} J(C')$. Since $C$ (resp. $C'$) are very general curves in $M_g$, we may assume (by [@BL Cor. 17.5.2]) that ${\mathrm{NS}}(J(C))={\mathbb{Z}}$ (resp. ${\mathrm{NS}}(J(C'))={\mathbb{Z}}$) generated by the class of the theta divisor $[\Theta_C]$ (resp. $[\Theta_{C'}]$). Therefore we must have $\epsilon^*([\Theta_{C'}])=\pm [\Theta_C]$. Moreover, since $[\Theta_C]$ is ample and the pull-back morphism $\epsilon^*$ preserves ampleness, we get that actually $\epsilon^*([\Theta_{C'}])=[\Theta_C]$. We conclude that $C\cong C'$ by the classical Torelli theorem. &gt;From the previous result, we can deduce two corollaries. The first one concerns the group of birational self maps ${\rm Bir}(P_{d,g})$ and the group of automorphisms ${\rm Aut}(P_{d,g})$ of $P_{d,g}$. \[Cor-bira1\] With the same assumptions as in Theorem \[birationa\], we have $${\rm Bir}(P_{d,g})={\rm Aut}(P_{d,g})= \begin{cases} {\mathbb{Z}}/2{\mathbb{Z}}& \text{ if } d(g-1) \text{ for some } n\in {\mathbb{Z}},\\ \{{\rm id}\} & \text{ otherwise,} \end{cases}$$ where in the first case the generator of the cyclic group ${\mathbb{Z}}/2{\mathbb{Z}}$ is $\psi_n^2$. The claim follows from Theorem \[birationa\], since the only maps $\psi_n^1$ and $\psi_n^2$ having domain $P_{d,g}$ and codomain $P_{d,g}$ are $\psi_0^1={\rm id}$ and $\psi_n^2$ if $d(g-1)$, and in this case $(\psi_n^2)\circ (\psi_n^2)={\rm id}$. The second corollary is analogous to [@GKM Cor. (0.12)], which states that the boundary $\partial {\overline}{M}_{g}$ of the moduli space ${\overline}{M}_g$ of stable curves of genus $g\geq 2$ is preserved by any automorphism of ${\overline}{M}_g$. \[Cor-bira2\] Same assumptions as in Theorem \[birationa\]. Then any automorphism $\phi:{\overline}{P}_{d,g}\to {\overline}{P}_{d,g}$ preserves the boundary $\partial {\overline}{P}_{d,g}:={\overline}{P}_{d,g}\setminus P_{d,g}$. The restriction $\eta:=\phi_{|P_{d,g}}$ of $\phi$ to $P_{d,g}$ defines a birational self map of $P_{d,g}$. By Corollary \[Cor-bira1\], $\eta$ is an automorphism of $P_{d,g}$. Therefore $\phi$ maps $P_{d,g}$ isomorphically onto $P_{d,g}$, hence it preserves the boundary $\partial {\overline}{P}_{d,g}$. \[Aut-comp\] Under the same assumptions as in Theorem \[birationa\], Corollary \[Cor-bira2\] implies that we have a restriction map $${\rm res}: {\mathrm{Aut}}({\overline{P}_{d,g}})\to {\mathrm{Aut}}({P_{d,g}}).$$ The map ${\rm res}$ is injective since ${\overline{P}_{d,g}}$ is separated. In [@Cap Lemma 8.1], it is claimed that the map $\psi_n^2$ of extends to a map ${\overline{P}_{d,g}}\to {\overline}{P}_{d',g}$ (the analogous statement for $\psi_n^1$ is easy to prove). This fact, together with Corollary \[Cor-bira1\], would imply that ${\rm res}$ is an isomorphism. Finally, note that if one could remove our technical assumption on the degree in Theorem \[Kod-geom\], then Theorem \[birationa\] and Corollaries \[Cor-bira1\] and \[Cor-bira2\] would follow for $g\geq 12$ without any hypothesis on the degree. Acknowledgements {#acknowledgements .unnumbered} ================ We are grateful to Silvia Brannetti, Lucia Caporaso, Eduardo Esteves and Margarida Melo for stimulating discussions on these topics. We thank Gavril Farkas and Sandro Verra for pointing out a gap in the proof of a previous version of Theorem \[birationa\]. We thank the two referees of the paper for their precious remarks that helped in improving the exposition. In particular, we are grateful to one of the referees for pointing out a mistake in a previous version of Theorem \[conj-sing\] and of Theorem \[sing-Pdst\]. [99]{} E. Arbarello, M. Cornalba: The Picard groups of the moduli spaces of curves, Topology 26 (1987), 153–171. C. Birkenhake, H. Lange: Complex abelian varieties. 2nd augmented ed. (English). Grundlehren der Mathematischen Wissenschaften 302, Berlin, Springer (2004). L. Caporaso: A compactification of the universal Picard variety over the moduli space of stable curves, J. Amer. Math. Soc. 7 (1994), 589–660. L. Caporaso: Néron models and compactified Picard schemes over the moduli stack of stable curves. Amer. J. Math. 130 (2008), 1–47. L. Caporaso: Compactified Jacobians of nodal curves. Notes for a minicourse given at the Istituto Superiore Tecnico of Lisbon, 1–4 February 2010 (available at http://www.mat.uniroma3.it/users/caporaso/cjac.pdf). L. Caporaso, C. Casagrande, M. Cornalba: Moduli of roots of line bundles. Transactions of the American Mathematical Society 359 (2007), 3733–3768. G. Casnati, C. Fontanari: On the rationality of the moduli space of pointed curves. Journal of the London Math. Society 75 (2007), 582–596. E. Cotterill: Effective divisors on ${\overline{M}_g}$ associated to curves with exceptional secant planes, arXiv:1004.0327. P. Deligne, D. Mumford: The irreducibility of the space of curves of given genus. Inst. Hautes Études Sci. Publ. Math. 36 (1969), 75–109. J. M. Drézet: Luna’s slice theorem and applications. Algebraic group actions and quotients, Notes of XXIII Autumn School in Algebraic Geometry (Wykno, Poland, 2000), Ed. J. Wisniewski, Hindawi Publ. Corp. (2004), 39-90. (available at http://people.math.jussieu.fr/%7Edrezet/papers/Wykno.pdf) D. Edidin, W. Graham: Riemann-Roch for equivariant Chow groups. Duke Math. J. 102 (2000), 567–594. D. Eisenbud, J. Harris: The Kodaira dimension of the moduli space of curves of genus $\geq 23$. Invent. Math. 90 (1987), 359–387. G. Farkas: The birational type of the moduli space of even spin curves. Advances in Mathematics 223 (2010), 433–443. G. Farkas, M. Popa: Effective divisors on $\overline{\mathcal{M}}_g$, curves on $K3$ surfaces, and the slope conjecture. J. Algebraic Geom. 14 (2005), 241–267. G. Farkas, K. Ludwig: The Kodaira dimension of the moduli space of Prym varieties. Journal of the European Mathematical Society (JEMS) 12 (2010), 755–795. G. Farkas, A. Verra: The geometry of the moduli space of odd spin curves, arXiv:1004.0278. G. Farkas, A. Verra: The classification of universal Jacobians over the moduli space of curves, arXiv:1005.5354. C. Fontanari: On the geometry of moduli of curves and line bundles. Ren. Mat. Acc. Lincei 16 (2005), 45–59. A. Gibney, S. Keel, I. Morrison: Towards the ample cone of $M_{g,n}$. J. Amer. Math. Soc. 15 (2002), 273–294. J. Harris, I. Morrison: Slopes of effective divisors on the moduli space of stable curves. Invent. Math. 99 (1990), 321–355. J. Harris, I. Morrison: Moduli of curves, GTM 187, Springer, New-York, 1998. J. Harris, D. Mumford: On the Kodaira dimension of the moduli space of curves. Invent. Math. 67 (1982), 23–88. Y. Kawamata: Minimal models and the Kodaira dimension of algebraic fiber spaces. J. Reine Angew. Math. 363 (1985), 1–46. Y. Laszlo: Local structure of the moduli space of vector bundles over curves. Comment. Math. Helv. 71 (1996), 373�-401. R. Lazarsfeld: Positivity in algebraic geometry. I. Classical setting: line bundles and linear series. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge 48. Berlin: Springer (2004). A. Logan: The Kodaira dimension of moduli spaces of curves with marked points. American Journal of Math. 125 (2003), 105–138. K. Ludwig: Moduli of spin curves. PhD Thesis, University of Hannover, 2007 (available at http://edok01.tib.uni-hannover.de/edoks/e01dh07/530657929.pdf). K. Ludwig: On the geometry of the moduli space of spin curves. J. Algebraic Geom. 19 (2010), 133–171. D. Luna: Slices étales. Sur les groupes algébriques, pp. 81–105. Bull. Soc. Math. France, Paris, Memoire 33 Soc. Math. France, Paris, 1973. M. Melo: Compactified Picard stacks over $\overline{\mathcal{M}}_g$. Math. Z. 263 (2009), 939–957. M. Melo: Compactified Picard stacks over the moduli stack of stable curves with marked points. Adv. Math. 226 (2011), 727–763. T. Mochizuki: Donaldson type invariants for algebraic surfaces. Lecture Notes in Mathematics 1972, Springer, New York, 2009. S. Mukai: Fano 3-folds, in: Complex Projective Geometry, London Math. Soc. Lecture Notes Ser. 179, Cambridge University Press (1992), 255–263. S. Mukai: Curves and K3 surfaces of genus eleven, in: Moduli of vector bundles, Lecture Notes in Pure and Appl. Math. 179, Dekker (1996), 189–197. F. Oort, J. Steenbrink: The local Torelli problem for algebraic curves. Journées de Géometrie Algébrique d’Angers, Juillet 1979/Algebraic Geometry, Angers, 1979, 157–204. D. Prill: Local classification of quotients of complex manifolds by discontinuous groups. Duke Math. J. 34 (1967), 375–386. E. Sernesi: Deformations of algebraic schemes. Grundlehren der mathematischen Wissenschaften 334, Springer, New York, 2006. S. L. Tan: On the slopes of the moduli spaces of curves. Int. J. Math. 9 (1998), 119–127. K. Ueno: Classification theory of algebraic varieties and compact complex spaces. Lecture Notes in Mathematics 439, Springer, Berlin-New York, 1975. A. Verra: The unirationality of the moduli spaces of curves of genus 14 or lower. Compos. Math. 141 (2005), 1425–1444. J. Wang: Deformations of pairs $(C,L)$ when $C$ is singular, arXiv:1003.6073.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The surface of a three-dimensional topological electron system often hosts symmetry-protected gapless surface states. With the effect of electron interactions, these surface states can be gapped out without symmetry breaking by a surface topological order, in which the anyon excitations carry anomalous symmetry fractionalization that cannot be realized in a genuine two-dimensional system. We show that for a mirror-symmetry-protected topological crystalline insulator with mirror Chern number $n=4$, its surface can be gapped out by an anomalous $\mathbb Z_2$ topological order, where all anyons carry mirror-symmetry fractionalization $M^2=-1$. The identification of such anomalous crystalline symmetry fractionalization implies that in a two-dimensional $\mathbb Z_2$ spin liquid the vison excitation cannot carry $M^2=-1$ if the spinon carries $M^2=-1$ or a half-integer spin.' author: - Yang Qi - Liang Fu bibliography: - 'TCI.bib' - 'additional.bib' title: Anomalous Crystal Symmetry Fractionalization on the Surface of Topological Crystalline Insulators --- The advent of topological insulators (TIs) [@Hasan2010; @QiZhang2011; @Moore2010] and topological superconductors (TSCs) [@Schnyder2008] has greatly broadened our understanding of topological phases in quantum systems. While the concepts of TIs and TSCs originate from topological band theory of noninteracting electrons or quasiparticles, recent theoretical breakthroughs [@FidkowskiKitaev2011; @Ryu2012; @Yao2013; @Qi2DZ82013; @FidkowskiPRX2013; @Wang2014a] have found that interactions can, in principle, change fundamental properties of these topological phases dramatically, thus creating a new dimension to explore. In particular, interactions can drive the gapless Dirac fermion surface states of three-dimensional (3D) TIs and TSCs into topologically ordered phases that are gapped and symmetry preserving. Nonetheless, such a surface manifests the topological property of the bulk in a subtle but unambiguous way: its anyon excitations have anomalous symmetry transformation properties, which cannot be realized in any two-dimensional (2D) system with the same symmetry. Given the profound consequences of interactions in TIs and TSCs, the effect of interactions in topological phases protected by spatial symmetries of crystalline solids, commonly referred to as topological crystalline insulators (TCIs) [@Fu2011a], is now gaining wide attention. A wide array of TCI phases with various crystal symmetries have been found in the framework of topological band theory [@AndoFu2015; @BernevigReview]. One class of TCIs has been predicted and observed in the IV-VI semiconductors SnTe, Pb$_{1-x}$Sn$_x$Se and Pb$_{1-x}$Sn$_x$Te [@HsiehSnTe2012; @AndoSnTe2012; @PolandPbSnSe2012; @HasanPbSnTe2012]. The topological nature of these materials is warranted by a particular mirror symmetry of the underlying rocksalt crystal, and is manifested by the presence of topological surface states on mirror-symmetric crystal faces. Remarkably, there surface states were found to become gapped under structural distortions that break the mirror symmetry [@vidya1; @vidya2], confirming the mechanism of crystalline protection unique to TCIs [@HsiehSnTe2012]. The study of interacting TCIs has just begun. A recent work by Isobe and Fu [@IsobeFu2015] shows that in the presence of interactions, the classification of 3D TCIs protected by mirror symmetry (i.e., the SnTe class) reduces from being characterized by an integer known as the mirror Chern number [@Teo2008] (hereafter denoted by $n$) to its $\mathbb Z_8$ subgroup. This implies that interactions can turn the $n=8$ surface states, which consists of eight copies of 2D massless Dirac fermions with the same chirality, into a completely trivial phase that is gapped, mirror symmetric and without intrinsic topological order. It remains an open question what interactions can do to TCIs with $n \neq 0 \mod 8$. In this work, we take the first step to study strongly interacting TCI surface states for the case $n=4 \mod 8$. Our main result is that the surface of a 3D TCI with mirror Chern number $n=4\mod8$ can become a gapped and mirror-symmetric state with $\mathbb Z_2$ topological order. Remarkably, the mirror symmetry acts on this state in an anomalous way that all three types of anyons carry fractionalized mirror quantum number $\tilde M^2=-1$ (in this Letter we use $\tilde M$ to represent the projective representation of mirror symmetry $M$ acting on an anyon), which cannot be realized in a purely 2D system. Furthermore, the anomalous mirror-symmetry fractionalization protects a twofold degeneracy between two mirror-symmetry-related edges. Such anomalous mirror-symmetry fractionalization cannot be realized in a 2D system, including a 2D $\mathbb Z_2$ spin liquid state [@wen1991; @*wen1991a]. Hence, our finding constrains the possible ways of fractionalizing the mirror symmetry in a 2D $\mathbb Z_2$ spin liquid [@wenpsg; @Essin2013]. Brief reviews of 3D TCIs, 2D $\mathbb Z_2$ spin liquids, and their edge theory are available in the Supplemental Material [^1]. *Noninteracting TCIs.* We begin by considering noninteracting TCIs protected by the mirror symmetry $x \rightarrow -x$. With the mirror symmetry, the extra U(1) symmetry in a TCI does not change the classification in 3D compared to a mirror-protected topological crystalline superconductor. Hence, for convenience, we choose a TCI as our starting point, although the U(1) symmetry plays no role in this work. As we will explain in Sec. \[sec:mz2\] of the Supplemental Material, in order to produce an anomalous $\mathbb Z_2$ surface topological order, the mirror operation must be defined as a $\mathbb Z_2$ symmetry with the property $M^2=1$. In our previous works on spin-orbit coupled systems, the mirror operation $M'$ acts on electron’s spin in addition to its spatial coordinate, which leads to $M'^2= -1$. Nonetheless, one can redefine the mirror operation by combining $M'$ with the U(1) symmetry of charge conservation $c \rightarrow i c$, which restores the property $M^2=1$. We note that without the U(1) symmetry only $M$ satisfying $M^2=+1$ protects nontrivial topological crystalline superconductors. The mirror TCIs are classified by the mirror Chern number $n$, defined for single-particle states on the mirror-symmetric plane $k_x=0$ in the 3D Brillouin zone. The states with mirror eigenvalues $1$ and $-1$ form two different subspaces, each of which has a Chern number denoted by $n_+$ and $n_-$ respectively. This leads to two independent topological invariants for noninteracting systems with mirror symmetry: the total Chern number $n_T = n_+ + n_-$ and the mirror Chern number $n = n_+ - n_-$. The TCI with a nontrivial mirror Chern number $n$ has gapless surface states consisting of $n$ copies of massless Dirac fermions, described by the following surface Hamiltonian $$\label{hs} H_s = v \sum_{A=1}^{n} \psi^\dagger_A (k_x \sigma_y - k_y \sigma_x ) \psi_A,$$ where the two-dimensional fermion fields $\psi_A(x,y)$ transform as the following under mirror operation: $$\label{eq:Mpsi} M: \psi_A (x, y) \rightarrow \sigma_x \psi_A(-x, y).$$ The presence of mirror symmetry  forbids any Dirac mass term $\psi^\dagger_A \sigma_z \psi_B$. As a result, the surface states described by Eq.  cannot be gapped by fermion bilinear terms, for any flavor number $n$. We emphasize that the above Dirac fermions on the surface of a 3D TCI cannot be realized in any 2D system with mirror symmetry, as expected for symmetry-protected topological phases in general. According to the Hamiltonian (\[hs\]), the surface states with $k_x=0$ within a given mirror subspace are chiral as they all move in the same direction [@HsiehSnTe2012]. In contrast, in any 2D system single-particle states within a mirror subspace cannot be chiral (this is demonstrated with a 2D lattice model in Sec. \[sec:piflux\] of the Supplemental Material [@Note1]). *$U(1)$ Higgs phase and $\mathbb Z_2$ topological order.* In this work, we study interacting surface states of TCIs with $n=4$. Starting from four copies of Dirac fermions in the noninteracting limit, we will introduce microscopic interactions and explicitly construct a $\mathbb Z_2$ topologically-ordered phase on the TCI surface, which is gapped and mirror symmetric. Our construction is inspired by the work of @Senthil2006 and @SenthilMotrunich on fractionalized insulators. We construct on the surface of an $n=4$ TCI a Higgs phase with an $xy$-order parameter $\langle b\rangle\neq0$, which is odd under the mirror symmetry and gaps the Dirac fermions. Next, we couple these gapped fermions to additional degrees of freedom $a_\mu$ that are introduced to mimic a $U(1)$ gauge field. This gauge field $a_\mu$ plays three crucial roles: (i) the coupling between matter and $a_\mu$ restores the otherwise broken $U(1)$ symmetry and, thus, the mirror symmetry along with it; (ii) the Goldstone mode is eaten by the gauge boson and becomes massive; (iii) since the $xy$-order parameter carries $U(1)$ charge 2, the $U(1)$ gauge group is broken to the $\mathbb{Z}_2$ subgroup in the Higgs phase. Because of these properties, the Higgs phase thus constructed is a gapped and mirror-symmetric phase with $\mathbb{Z}_2$ topological order. We now elaborate on the construction (details of this construction can be found in Sec. \[sec:intr-u1z2\] of the Supplemental Material [@Note1]). First, we relabel the fermion flavors $A=1,\ldots, 4$ using a spin index $s=\uparrow, \downarrow$ and a $U(1)$-charge index $a=\pm$ (unrelated to the electric charge). We take fermion interactions that are invariant under both the $SU(2)$ spin rotation and the $U(1)$ rotation $$\label{eq:u1psi} U(1): \psi_{as} \rightarrow e^{i a \theta} \psi_{as}, \; a=\pm$$ Moreover, we introduce a boson field $b(x,y)$ that carries $U(1)$-charge 2 and is odd under mirror symmetry, $$\label{eq:u1m} U(1): b\rightarrow e^{i2\theta}b,\quad M: b(x,y) \rightarrow - b(-x,y), $$ and couple this boson to the massless Dirac fermions as follows $$\label{eq:hbf} H_{bf}= V b^\dagger \psi^\dagger_{a s} \tau^-_{ab} \sigma_z \psi_{bs} + \text{H.c.}$$ When these bosons condense, $\langle b\rangle\neq0$ spontaneously breaks both the $U(1)$ and mirror symmetry, and gappes out the fermions. Finally, we introduce another boson vector field $a_\mu(x,y)$, which couples to $b$ and $\psi_{as}$ through minimal coupling. An effective theory of this system has the following form, $$\label{eq:Lsfd} \begin{split} \mathcal L = &-i\psi^\dagger_{s}\alpha^\mu(\partial_\mu+ia_\mu \tau_z)\psi_s + (b \psi^\dagger_s \tau^+ \sigma_z \psi_s + \text{H.c.})\\ &+\frac{1}{2g} |(\partial_\mu - 2i a_\mu)b|^2 + r |b|^2 + u |b|^4 + F_{\mu\nu} F^{\mu\nu}, \end{split}$$ where the matrices $\alpha^0=1$, $\alpha^x=\sigma_y$ and $\alpha^y=\sigma_x$. Furthermore, we add to the effective action an interation term $UN^2$, where $N=\psi^\dagger\tau_z\psi+2b^\dagger b-\nabla\cdot\bm E$ ($E_i=F_{0i}=\partial_0a_i-\partial_ia_0$ is the electric field strength). In the limit of $U\rightarrow\infty$, this enforces the local constraint $N=0$. As a result, the bare fermion $\psi_s$ and boson $b$ are no longer low-energy excitations, since adding them to the ground state violates the constraint $N=0$ and costs an energy $U$. Therefore, in the low-energy effective model $\psi_s$ and $b$ must be screened by the gauge field $a_\mu$ and become quasiparticles $\tilde\psi_{as}=\psi_{as}e^{ia\theta}$ and $\tilde{b}=be^{2i\theta}$, where the operator $e^{in\theta}$ creates $n$ gauge charge of $a_\mu$ and restores the constraint $N=0$. In terms of these quasiparticles, the effective theory becomes $$\label{eq:Lsfd2} \begin{split} \mathcal L = &-i\tilde\psi^\dagger_{s}\alpha^\mu(\partial_\mu+ia_\mu \tau_z)\tilde\psi_s + (\tilde b\tilde\psi^\dagger_s \tau^+ \sigma_z \tilde\psi_s + \text{H.c.})\\ &+\frac{1}{2g} |(\partial_\mu - 2i a_\mu)\tilde b|^2 + r |\tilde b|^2 + u |\tilde b|^4 + F_{\mu\nu} F^{\mu\nu}. \end{split}$$ Furthermore, a $U(1)$ gauge symmetry emerges in the low-energy Hilbert space defined by the local constraint $N=0$ [^2]. Specifically, the constraint is the Gauss law and it restricts the low-energy Hilbert space to states that are invariant under the gauge transformation $$\label{eq:Utheta} U_\phi:\tilde\psi_s\rightarrow e^{i\phi\tau_z}\tilde\psi_s,\quad \tilde b\rightarrow \tilde be^{2i\phi},\quad a_\mu\rightarrow a_\mu-\partial_\mu\phi.$$ In this effective theory with the emergent $U(1)$ gauge field, condensing the boson $\tilde b$ no longer breaks the global $U(1)$ and the mirror symmetries, as it instead breaks the $U(1)$ gauge symmetry to $\mathbb Z_2$. Naively, the mirror symmetry maps $\langle\tilde b\rangle$ to $-\langle\tilde b\rangle$. However these two symmetry breaking vacuua are equivalent because they are related by the gauge symmetry transformation $U_{\pi/2}:\tilde b\rightarrow -\tilde b$. This restoration of mirror symmetry becomes clearly manifested if we assume that $\tilde b$ and $\tilde\psi_s$ transform projectively under mirror symmetry with the additional U(1) gauge transformation $U_{\pi/2}$, $$\label{eq:chisu2} \tilde M: \tilde\psi_s(r)\rightarrow i\tau_z\otimes\sigma_x\tilde\psi_s(r^\prime); \tilde b(r)\rightarrow\tilde b(r^\prime).$$ This Higgs phase obtained by condensing charge-2 $\tilde b$ field indeed has a $\mathbb Z_2$ topological order when the number of Dirac fermions is $n=4$ [@Wang2014a; @Metlitski2014]. This can be understood by identifying the Bogoliubov quasiparticle $\tilde \psi$ and vortices as the anyons $e$, $m$, and $\epsilon$ (see Sec. \[sec:mz2\] of the Supplemental Material [@Note1] for the definition of the notation) in the $\mathbb Z_2$ topological order. $\tilde\psi$ becomes the $\epsilon$ anyon as both are fermions. As the Higgs field gaps out four Dirac fermions, there are four Majorana fermions, or two complex fermion zero modes, in each vortex core. Hence, there are two types of vortices whose core has even or odd fermion parity, respectively. In the case of $n=4$, it can be shown that the vortices carry Bose statistics (see Sec. \[sec:topol-order-higgs\] of the Supplemental Material [@Note1] for details), and they are mapped to the $m$ and $e$ anyons in the $\mathbb Z_2$ topological order, respectively. *Mirror-symmetry fractionalization.* Now we consider how the mirror symmetry acts in the $\mathbb Z_2$ spin liquid phase described by Eq. . In this effective theory, the $\tilde\psi$ field is the fermionic anyon $\epsilon$. Equation  implies that it carries $\tilde M^2=-1$. ![(a) A vortex and an antivortex. The direction of the arrow represents the phase of the Higgs field $\langle\tilde b\rangle$. The dotted line is the mirror axis. (b) Illustration of the fermion spectrum flow from the left to the right as we create a vortex-antivortex pair from the vacuum and move them far apart. In this process, four vortex core states are separated from the bulk spectrum, two from the conducting band and two from the valence band, and become degenerate zero modes. As illustrated in the inset, the core states are twofold degenerate with spin $s=\pm1$.[]{data-label="fig:vortices"}](pair_flow) Next, we consider how the mirror symmetry acts on the $m$ anyon, which is a vortex of the Higgs field where all core states are empty. Since the mirror symmetry preserves the Higgs field $\langle\tilde b\rangle$ but maps $x$ to $-x$, it maps a vortex to an antivortex. Therefore we consider a mirror-symmetric configuration with one vortex and one antivortex, as shown in Fig. \[fig:vortices\](a). The symmetry fractionalization of $\tilde M^2=\pm1$ can be detected from the $M$ parity of the fermion wave function with such a vortice-antivortex pair [@QiCSF]: for two bosonic vortices, the mirror parity is equal to $\tilde M^2$. From Eq. , we get the fermion Hamiltonian $$\label{eq:Hchi} H=v\tilde\psi_s^\dagger\left(k_x\sigma_y-k_y\sigma_x\right)\tilde\psi_s +\tilde\psi_s^\dagger(\langle\tilde b\rangle\tau^+ +\langle\tilde b\rangle^\ast\tau^-)\sigma_z\tilde\psi_s.$$ It has a particle-hole symmetry $\Xi:\tilde\psi\rightarrow\sigma_z\tilde\psi$ which maps $H$ to $\Xi H\Xi=-H$. This implies that its spectrum is symmetric with respect to zero. Assume that the dimension of the whole Hilbert space is $4N$; there are $2N$ states with positive energy and $2N$ states with negative energy. For the vortex configuration in Fig. \[fig:vortices\](a), there are four complex zero modes, two from each vortex core, which are all unoccupied. Therefore, excluding these four states there are $2N-2$ states with negative energy, which are all occupied in the fermion wave function. Next, we consider the mirror eigenvalues of these $2N-2$ occupied fermion states. Since $\tilde\psi$ carries $\tilde M^2=-1$, each state has mirror eigenvalue $\lambda_M=\pm i$. Because both $H$, $\Xi$ and $M$ are diagonal in pseudospin $s=\uparrow,\downarrow$, all occupied states are pseudospin doublets, and two states in each doublet have the same $\lambda_M$. Hence, the mirror eigenvalue of all occupied states, organized as $N-1$ doublets, is $(-1)^{N-1}=-1$. Therefore, the wave function of two empty vortices is odd under mirror symmetry, which implies that the $e$ particle has the symmetry fractionalization $\tilde M^2=-1$. Combining the results that both $e$ and $\epsilon$ carry $\tilde M^2=-1$, we conclude that the $m$ anyon also has $\tilde M^2=-1$ (see the discussion in Sec. \[sec:mz2\] of the Supplemental Material [@Note1]). In summary, the gapped $\mathbb Z_2$ surface state we constructed has an anomalous mirror-symmetry fractionalization that both types of anyons carry $\tilde M^2=-1$, which cannot be realized in a genuine 2D system. For comparison, it is shown in Sec. \[sec:mz2\] of the Supplemental Material [@Note1] that applying the same construction to a 2D lattice model results in a $\mathbb Z_2$ phase with a different mirror-symmetry fractionalization which is not anomalous. *Mirror anomaly.* The anomalous crystal symmetry fractionalization presented in the surface topological order implies a symmetry-protected topological degeneracy associated with the edges of the surface topological ordered region. This mirror anomaly is a remnant of the anomalous surface fermion modes in the free-fermion limit. To see this, we consider the setup presented in Fig. \[fig:subway\], in which the $\mathbb Z_2$ surface topologically ordered state is terminated at two edges symmetric with respect to the mirror plane, by two regions with opposite $\langle b\rangle=\pm1$ on either side of the mirror plane, respectively. ![Two mirror-symmetric edges of a $\mathbb Z_2$ surface topological order. The mirror symmetry maps $x$ to $-x$ with respect to the mirror plane marked by the red disk at $x=0$. The surface topological order marked by the shade terminates at edges against two ordered regions with $\langle b\rangle=\pm1$, respectively.[]{data-label="fig:subway"}](subway) This setup itself does not break the mirror symmetry, and all local excitations can be gapped everywhere on the surface. In particular, since the $\mathbb Z_2$ topological order is not chiral, its edge can be gapped out by condensing either $e$ or $m$ anyons on the edge [@Barkeshli2014a]. The edges next to an ordered phase with $\langle b\rangle\neq0$ are $e$ edges, as condensing $e$ breaks the global U(1) symmetry. A $\mathbb Z_2$ spin liquid state on an infinite cylinder has four degenerate ground states $|\Psi_a\rangle$, each has one type of anyon flux $a=1,e,m,\epsilon$ going through the cylinder. On a finite cylinder with two $e$ edges, only $|\Psi_1\rangle$ and $|\Psi_e\rangle$ remain degenerate, because adding an $m$ or $\epsilon$ anyon on the edge costs a finite energy. In a generic $\mathbb Z_2$ state, this degeneracy can be further lifted by tunneling an $e$ anyon between the two edges, $H_t=\lambda e_L^\dagger e_R^\dagger+\text{H.c.}$, where $e_{L,R}^\dagger$ creates two $e$ anyons on the two edges, respectively. However, the $e$ anyon carries $\tilde M^2=-1$; therefore, the tunneling term $H_t$ is odd under mirror and, thus, forbidden by $M$. As a result, this twofold topological degeneracy is protected by the mirror symmetry even in the limit of $L\rightarrow0$. This argument is formulated using the effective edge Lagrangian in the Supplemental Material [@Note1]. In the limit of $L\rightarrow0$, this topological degeneracy becomes a local degeneracy protected by the mirror symmetry. Therefore if the $\mathbb Z_2$ topological order is killed by collapsing two gapped edges, the ground state is either gapless or mirror-symmetry breaking, and this cannot be avoided regardless of edge types because all types of anyons have $\tilde M^2=-1$. This topological degeneracy reveals the anomalous nature of this mirror-symmetry fractionalization. Furthermore, if we collapse two gapless edges of the $\mathbb Z_2$ state, the edges remain gapless because the anyon tunneling is forbidden by $M$. Hence we get a gapless domain wall with central charge $c=1+1=2$, which recovers the edge with four chiral fermion modes in the aforementioned free-fermion limit. This is explained in more detail in Sec. \[sec:edge-lutt-theory\] of the Supplemental Material [@Note1]. *Conclusion.* In this Letter, we show that the surface of a 3D mirror TCI with mirror Chern number $n=4$, containing four gapless Dirac fermion modes in the free limit, can be gapped out without breaking the mirror symmetry by a $\mathbb Z_2$ topological order. This surface $\mathbb Z_2$ topological order has an anomalous mirror-symmetry fractionalization in which all three types of anyons carry fractionalized mirror-symmetry quantum number $\tilde M^2=-1$, and such a topological order cannot be realized in a purely 2D system. Our finding also puts constraints on possible ways to fractionalize the mirror symmetry in a 2D $\mathbb Z_2$ quantum spin liquid [@Essin2013; @BarkeshliX]. The result of this work indicates that the combination that both the $e$ and $m$ carry the fractionalized $\tilde M^2=-1$ is anomalous and cannot be realized in a 2D $\mathbb Z_2$ spin liquid. Furthermore, our result can be easily generalized to also rule out the combination that the $e$ anyon carries spin-$\frac12$ and $m$ anyon carries $\tilde M^2=-1$ [@HermeleAspen], because if $e$ carries $\tilde M^2=+1$, we can define a new mirror symmetry $M^\prime=Me^{i\pi S^z}$, for which both $e$ and $m$ carry $(\tilde M^\prime)^2=-1$ and, therefore, this combination is also anomalous. In summary, our finding implies that the vison must carry $\tilde M^2=+1$ in a $\mathbb Z_2$ spin liquid where the spinon carries a half-integer spin. We thank Chen Fang and Senthil Todadri for invaluable discussions. Y.Q. is supported by National Basic Research Program of China through Grant No. 2011CBA00108 and by NSFC Grant No. 11104154. L.F. is supported by the DOE Office of Basic Energy Sciences, Division of Materials Sciences and Engineering, under Award No. DE-SC0010526. This research was supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation. [^1]: See the Supplemental Material, which includes Refs. [@LuBFU; @kitaev; @sstri; @Tanaka2005], for reviews of 3D TCIs, 2D $\mathbb Z_2$ spin liquids, and their edge theory. [^2]: Here the constraint $N=0$ can be enforced locally because the U(1) symmetry defined in Eqs.  and is an on-site symmetry operation. Particularly the U(1) rotation of the Dirac fermion $\psi$ is a local rotation between four flavors of Dirac fermions, which arise from the four flavors of Dirac fermions in the bulk of the 3D TCI, as explained in Sec. \[sec:mirr-topol-cryst\] of the Supplemental Material [@Note1].
{ "pile_set_name": "ArXiv" }